{"forum": "maQs92RfZ8", "year": 2026, "status": "rejected", "title": "Exploring Federated Pruning for Large Language Models", "authors": ["Pengxin Guo", "Yinong Wang", "Wei Li", "Mengting Liu", "Ming Li", "Jinkai Zheng", "Liangqiong Qu"], "authorids": ["~Pengxin_Guo1", "~Yinong_Wang1", "~Wei_Li82", "~Mengting_Liu6", "~Ming_Li21", "~Jinkai_Zheng1", "~Liangqiong_Qu2"], "authors_source": "OpenReview API", "abstract": "LLM pruning has emerged as a promising technology for compressing LLMs, enabling their deployment on resource-limited devices. However, current methodologies typically require access to public calibration samples, which can be challenging to obtain in privacy-sensitive domains. To address this issue, we introduce FedPrLLM, a comprehensive federated pruning framework designed for the privacy-preserving compression of LLMs. In FedPrLLM, each client only needs to calculate a pruning mask matrix based on its local calibration data and share it with the server to prune the global model. This approach allows for collaborative pruning of the global model with the knowledge of each client while maintaining local data privacy. Additionally, we conduct extensive experiments to explore various possibilities within the FedPrLLM framework, including different comparison groups, pruning strategies, and the decision to scale weights. Our extensive evaluation reveals that one-shot pruning with layer comparison and no weight scaling is the optimal choice within the FedPrLLM framework. We hope our work will help guide future efforts in pruning LLMs in privacy-sensitive fields. Our code is available at https://anonymous.4open.science/r/FedPrLLM-15594.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Ig2rLtyzjg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15594/Reviewer_y1jX"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces FedPrLLM, a federated pruning framework for large language models (LLMs) aimed at privacy-preserving compression. Clients compute local pruning masks using private calibration data and share them with a server, which aggregates masks to prune the global model. The authors explore three design choices: (1) comparison group (layer, row, column), (2) weight scaling, and (3) pruning strategy (one-shot vs. iterative). Experiments across LLaMA variants and multiple datasets conclude that layer-level comparison, no weight scaling, and one-shot pruning are preferable.", "review_text": "The paper introduces FedPrLLM, a federated pruning framework for large language models (LLMs) aimed at privacy-preserving compression. Clients compute local pruning masks using private calibration data and share them with a server, which aggregates masks to prune the global model. The authors explore three design choices: (1) comparison group (layer, row, column), (2) weight scaling, and (3) pruning strategy (one-shot vs. iterative). Experiments across LLaMA variants and multiple datasets conclude that layer-level comparison, no weight scaling, and one-shot pruning are preferable.", "strengths": "- The paper addresses a timely problem: pruning LLMs under privacy constraints in federated settings.\n- Provides extensive empirical evaluation across multiple models, sparsity levels, and datasets.\n- Clear takeaways (layer comparison, no scaling, one-shot pruning) that practitioners can adopt.\n- Framework is simple and easy to implement, making it accessible for real-world experimentation.", "weaknesses": "- The approach is a straightforward mask aggregation; similar ideas exist (e.g., FedSpaLLM[1]).\n- Weak privacy claim. No secure aggregation or differential privacy. Masks may potentially leak sensitive information\n- No comparison to advanced pruning methods (OWL[2], BESA[3], SliceGPT[4]) or structured sparsity approaches. While the authors can argue that some of tghe methods are structured pruning methods, it is important to compare against them.\n- Communication costs lack units; no runtime, memory, or speedup metrics; catastrophic perplexities for column comparison unexplained.", "questions": "Given the strengths, I have the following questions:\n\n1. The paper positions itself as the first systematic study of federated pruning for LLMs, but prior work (e.g., FedSpaLLM) already explores similar aggregation strategies. Specifically, the \\ell_0‑aware aggregation, adaptive mask expansion, layer sampling and related variants from FedSpaLLM that aggregate sparse models or masks to meet global sparsity budgets. How does FedPrLLM fundamentally differ from FedSpaLLM or other federated pruning approaches. The differences that you included in the related work section is not sufficient?\n\n2. The authors noted that previous approaches are not optimal, but there is no comparison against such approaches. The paper limits local pruning to Wanda and SparseGPT but omits competitive post‑training LLM pruning methods that have non‑uniform sparsity (OWL), blockwise reconstruction (BESA), or structured slicing (SliceGPT), each of which has shown superior performance or concrete speedups. Without these baselines, it is unclear whether FedPrLLM’s “best recipe” remains best when local candidates are stronger or structured. Can the authors include these baselines and clarify whether the proposed framework remains competitive.\n\n3. My major concern is in the calculation of the local pruning mask matrix using the private calibration data. When this is sent to the global server, what are the privacy risks involved? Sending the mask without any privacy protection could potentially make the mask vulnerable to attacks.\nThe paper claims “privacy‑preserving compression” because only binary masks are shared. But masks are data‑dependent and can leak information (presence/absence of features/examples). The FL literature documents strong leakage from shared updates (gradient inversion, membership/property inference). Without secure aggregation or differential privacy analysis, there is no formal protection; moreover, secure aggregation would change communication/computation budgets materially, which the paper does not account for. Can the authors clarify this?\n\n4. I also have a concern around the settings. All experiments shard IID C4 into 128 sequences and hand exactly 2 sequences per client; there is no non‑IID partitioning, no client sampling/dropouts, and no heterogeneity in compute/network, all central to FL. As a result, the conclusions may not transfer to realistic cross‑silo or cross‑device FL deployments. Can the authors clarify this?\n\n5. Communication costs are reported as raw numbers (“6.476B”, “12.952B”) without units (bits? parameters? bytes?). It is relatively unclear what this mean. Also, there is no textual description to clarify this.\n\n6. The catastrophic perplexities for column comparison (e.g., 311,468.53 on LLaMA‑3‑8B at 50% sparsity) indicate numerical instability or a mis‑specified comparison group; the paper does not analyze or mitigate these failures. In addition, Eq. (1) imposes $\\|M_\\ell\\|_0 \\ge k$ but the server later enforces exact top‑k (fixed sparsity). The mismatch weakens the formulation.\n\n\n\n[1] Bai et al. FedSpaLLM: Federated Pruning of Large Language Models\n\n[2] Yin et al. Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity\n\n[3] Xu et al. BESA: Pruning Large Language Models with Blockwise Parameter-Efficient Sparsity Allocation\n\n[4] Ashkboos et al. SliceGPT: Compress large language models by deleting rows and columns", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces FedPrLLM, a federated pruning framework for large language models (LLMs) aimed at privacy-preserving compression. Clients compute local pruning masks using private calibration data and share them with a server, which aggregates masks to prune the global model. The authors explore three design choices: (1) comparison group (layer, row, column), (2) weight scaling, and (3) pruning strategy (one-shot vs. iterative). Experiments across LLaMA variants and multiple datasets conclude that layer-level comparison, no weight scaling, and one-shot pruning are preferable.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper addresses a timely problem: pruning LLMs under privacy constraints in federated settings.\n- Provides extensive empirical evaluation across multiple models, sparsity levels, and datasets.\n- Clear takeaways (layer comparison, no scaling, one-shot pruning) that practitioners can adopt.\n- Framework is simple and easy to implement, making it accessible for real-world experimentation.", "weaknesses": "- The approach is a straightforward mask aggregation; similar ideas exist (e.g., FedSpaLLM[1]).\n- Weak privacy claim. No secure aggregation or differential privacy. Masks may potentially leak sensitive information\n- No comparison to advanced pruning methods (OWL[2], BESA[3], SliceGPT[4]) or structured sparsity approaches. While the authors can argue that some of tghe methods are structured pruning methods, it is important to compare against them.\n- Communication costs lack units; no runtime, memory, or speedup metrics; catastrophic perplexities for column comparison unexplained.", "questions": "Given the strengths, I have the following questions:\n\n1. The paper positions itself as the first systematic study of federated pruning for LLMs, but prior work (e.g., FedSpaLLM) already explores similar aggregation strategies. Specifically, the \\ell_0‑aware aggregation, adaptive mask expansion, layer sampling and related variants from FedSpaLLM that aggregate sparse models or masks to meet global sparsity budgets. How does FedPrLLM fundamentally differ from FedSpaLLM or other federated pruning approaches. The differences that you included in the related work section is not sufficient?\n\n2. The authors noted that previous approaches are not optimal, but there is no comparison against such approaches. The paper limits local pruning to Wanda and SparseGPT but omits competitive post‑training LLM pruning methods that have non‑uniform sparsity (OWL), blockwise reconstruction (BESA), or structured slicing (SliceGPT), each of which has shown superior performance or concrete speedups. Without these baselines, it is unclear whether FedPrLLM’s “best recipe” remains best when local candidates are stronger or structured. Can the authors include these baselines and clarify whether the proposed framework remains competitive.\n\n3. My major concern is in the calculation of the local pruning mask matrix using the private calibration data. When this is sent to the global server, what are the privacy risks involved? Sending the mask without any privacy protection could potentially make the mask vulnerable to attacks.\nThe paper claims “privacy‑preserving compression” because only binary masks are shared. But masks are data‑dependent and can leak information (presence/absence of features/examples). The FL literature documents strong leakage from shared updates (gradient inversion, membership/property inference). Without secure aggregation or differential privacy analysis, there is no formal protection; moreover, secure aggregation would change communication/computation budgets materially, which the paper does not account for. Can the authors clarify this?\n\n4. I also have a concern around the settings. All experiments shard IID C4 into 128 sequences and hand exactly 2 sequences per client; there is no non‑IID partitioning, no client sampling/dropouts, and no heterogeneity in compute/network, all central to FL. As a result, the conclusions may not transfer to realistic cross‑silo or cross‑device FL deployments. Can the authors clarify this?\n\n5. Communication costs are reported as raw numbers (“6.476B”, “12.952B”) without units (bits? parameters? bytes?). It is relatively unclear what this mean. Also, there is no textual description to clarify this.\n\n6. The catastrophic perplexities for column comparison (e.g., 311,468.53 on LLaMA‑3‑8B at 50% sparsity) indicate numerical instability or a mis‑specified comparison group; the paper does not analyze or mitigate these failures. In addition, Eq. (1) imposes $\\|M_\\ell\\|_0 \\ge k$ but the server later enforces exact top‑k (fixed sparsity). The mismatch weakens the formulation.\n\n\n\n[1] Bai et al. FedSpaLLM: Federated Pruning of Large Language Models\n\n[2] Yin et al. Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity\n\n[3] Xu et al. BESA: Pruning Large Language Models with Blockwise Parameter-Efficient Sparsity Allocation\n\n[4] Ashkboos et al. SliceGPT: Compress large language models by deleting rows and columns", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762356882330}, {"id": "Yf6f6euFTg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15594/Reviewer_9bhx"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces **FedPrLLM**, a comprehensive framework for **federated unstructured pruning of large language models** (LLMs) under strict data privacy constraints. In FedPrLLM, each client computes a local pruning mask using its private calibration data and uploads only the binary mask (not model weights or raw data) to the server. The server aggregates these masks via summation and selects the top-*k* entries (i.e., weights most clients agree to prune) to construct a global mask for pruning the shared LLM.\n\nThe work systematically investigates three key design choices within this framework:  \n1. **Comparison group**: Should pruning decisions be made across the entire layer, per row, or per column?  \n2. **Weight scaling**: Should retained weights be scaled based on client consensus (e.g., inverse pruning frequency)?  \n3. **Pruning strategy**: Is iterative (layer-by-layer) pruning worth its high communication cost compared to one-shot pruning?\n\nThrough extensive experiments across **6 open-source LLMs**, **2 local pruning methods** (Wanda, SparseGPT), **3 sparsity levels**, **10 datasets**, and thousands of GPU hours, the authors derive three robust empirical findings:\n- **Layer-wise comparison** is consistently superior and robust across settings.\n- **Weight scaling harms performance**, despite intuitive appeal.\n- **One-shot pruning matches iterative pruning in accuracy** while halving communication cost.", "review_text": "This paper introduces **FedPrLLM**, a comprehensive framework for **federated unstructured pruning of large language models** (LLMs) under strict data privacy constraints. In FedPrLLM, each client computes a local pruning mask using its private calibration data and uploads only the binary mask (not model weights or raw data) to the server. The server aggregates these masks via summation and selects the top-*k* entries (i.e., weights most clients agree to prune) to construct a global mask for pruning the shared LLM.\n\nThe work systematically investigates three key design choices within this framework:  \n1. **Comparison group**: Should pruning decisions be made across the entire layer, per row, or per column?  \n2. **Weight scaling**: Should retained weights be scaled based on client consensus (e.g., inverse pruning frequency)?  \n3. **Pruning strategy**: Is iterative (layer-by-layer) pruning worth its high communication cost compared to one-shot pruning?\n\nThrough extensive experiments across **6 open-source LLMs**, **2 local pruning methods** (Wanda, SparseGPT), **3 sparsity levels**, **10 datasets**, and thousands of GPU hours, the authors derive three robust empirical findings:\n- **Layer-wise comparison** is consistently superior and robust across settings.\n- **Weight scaling harms performance**, despite intuitive appeal.\n- **One-shot pruning matches iterative pruning in accuracy** while halving communication cost.", "strengths": "1. **High Practical Relevance**: Addresses a critical gap—how to compress LLMs in privacy-sensitive, decentralized settings (e.g., healthcare, finance)—where public calibration data is unavailable.\n2. **Rigorous and Systematic Evaluation**: The scale of experiments (6 LLMs, multiple sparsities, datasets, and methods) is exceptional for a systems/ML paper. The ablation studies are thorough and convincing.\n3. **Clear, Counterintuitive Insights**: The findings—especially that weight scaling hurts performance and that one-shot pruning suffices—are surprising yet well-supported, challenging assumptions from centralized pruning literature.\n4. **Strong Engineering Contribution**: FedPrLLM is simple, communication-efficient, and compatible with existing local pruning methods. The framework is modular and extensible.", "weaknesses": "1. **Limited Baseline Comparison**: While the paper compares against “Local-only” and “Centralized” baselines, it does not benchmark against concurrent or prior federated compression methods (e.g., FedSpaLLM [Bai et al., 2024], mentioned in Related Work). A direct comparison would strengthen impact claims.\n2. **Assumption of Public Pre-training**: The framework assumes access to a public pre-trained LLM. While standard, the paper does not discuss implications if pre-training data were private—a growing concern in DP-ML.\n3. **Calibration Data Efficiency**: Each client uses only 2 samples (64 clients × 2 = 128 total). While realistic, the sensitivity to ultra-low calibration data (e.g., 1 sample/client) is not explored.\n4. **Communication Cost Nuance**: Although one-shot reduces rounds, the paper does not analyze bandwidth vs. latency trade-offs or heterogeneous client capabilities.", "questions": "As described in weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces **FedPrLLM**, a comprehensive framework for **federated unstructured pruning of large language models** (LLMs) under strict data privacy constraints. In FedPrLLM, each client computes a local pruning mask using its private calibration data and uploads only the binary mask (not model weights or raw data) to the server. The server aggregates these masks via summation and selects the top-*k* entries (i.e., weights most clients agree to prune) to construct a global mask for pruning the shared LLM.\n\nThe work systematically investigates three key design choices within this framework:  \n1. **Comparison group**: Should pruning decisions be made across the entire layer, per row, or per column?  \n2. **Weight scaling**: Should retained weights be scaled based on client consensus (e.g., inverse pruning frequency)?  \n3. **Pruning strategy**: Is iterative (layer-by-layer) pruning worth its high communication cost compared to one-shot pruning?\n\nThrough extensive experiments across **6 open-source LLMs**, **2 local pruning methods** (Wanda, SparseGPT), **3 sparsity levels**, **10 datasets**, and thousands of GPU hours, the authors derive three robust empirical findings:\n- **Layer-wise comparison** is consistently superior and robust across settings.\n- **Weight scaling harms performance**, despite intuitive appeal.\n- **One-shot pruning matches iterative pruning in accuracy** while halving communication cost.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. **High Practical Relevance**: Addresses a critical gap—how to compress LLMs in privacy-sensitive, decentralized settings (e.g., healthcare, finance)—where public calibration data is unavailable.\n2. **Rigorous and Systematic Evaluation**: The scale of experiments (6 LLMs, multiple sparsities, datasets, and methods) is exceptional for a systems/ML paper. The ablation studies are thorough and convincing.\n3. **Clear, Counterintuitive Insights**: The findings—especially that weight scaling hurts performance and that one-shot pruning suffices—are surprising yet well-supported, challenging assumptions from centralized pruning literature.\n4. **Strong Engineering Contribution**: FedPrLLM is simple, communication-efficient, and compatible with existing local pruning methods. The framework is modular and extensible.", "weaknesses": "1. **Limited Baseline Comparison**: While the paper compares against “Local-only” and “Centralized” baselines, it does not benchmark against concurrent or prior federated compression methods (e.g., FedSpaLLM [Bai et al., 2024], mentioned in Related Work). A direct comparison would strengthen impact claims.\n2. **Assumption of Public Pre-training**: The framework assumes access to a public pre-trained LLM. While standard, the paper does not discuss implications if pre-training data were private—a growing concern in DP-ML.\n3. **Calibration Data Efficiency**: Each client uses only 2 samples (64 clients × 2 = 128 total). While realistic, the sensitivity to ultra-low calibration data (e.g., 1 sample/client) is not explored.\n4. **Communication Cost Nuance**: Although one-shot reduces rounds, the paper does not analyze bandwidth vs. latency trade-offs or heterogeneous client capabilities.", "questions": "As described in weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761995111133}, {"id": "ZkSW7eWfoz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15594/Reviewer_XJZZ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces FedPrLLM, a general federated pruning framework for large language models (LLMs) that allows clients to perform local unstructured pruning and share only binary masks to preserve data privacy. The authors systematically explore three key design dimensions—comparison group (layer/row/column), weight scaling, and one-shot vs. iterative pruning—across various LLMs, sparsity levels, and datasets. They conduct extensive experiments (6 LLMs, 10 datasets) and derive practical insights (e.g., “layer-wise voting without weight scaling and one-shot pruning works best”).\n\nThis paper makes a practical and well-executed contribution by empirically benchmarking federated pruning strategies for LLMs. While the framework is not fundamentally novel and lacks theoretical depth, its findings are well-supported, reproducible, and directly useful for the community. A stronger baseline comparison and heterogeneity analysis would elevate it further.", "review_text": "This paper introduces FedPrLLM, a general federated pruning framework for large language models (LLMs) that allows clients to perform local unstructured pruning and share only binary masks to preserve data privacy. The authors systematically explore three key design dimensions—comparison group (layer/row/column), weight scaling, and one-shot vs. iterative pruning—across various LLMs, sparsity levels, and datasets. They conduct extensive experiments (6 LLMs, 10 datasets) and derive practical insights (e.g., “layer-wise voting without weight scaling and one-shot pruning works best”).\n\nThis paper makes a practical and well-executed contribution by empirically benchmarking federated pruning strategies for LLMs. While the framework is not fundamentally novel and lacks theoretical depth, its findings are well-supported, reproducible, and directly useful for the community. A stronger baseline comparison and heterogeneity analysis would elevate it further.", "strengths": "- The paper enables LLM pruning under federated settings with privacy preservation, which is interesting and underexplored.\n\n- The paper did large-scale experiments with 6 LLMs, 10 datasets, multiple sparsity levels. Ablation studies support key claims.\n\n- The three pruning dimensions are practical and grounded, covering critical decisions in collaborative model compression. Shows that simpler design choices (e.g., no scaling, one-shot pruning) often outperform complex alternatives, saving computation and communication.", "weaknesses": "- The core framework is a combination of known components; prior work (e.g., FedSpaLLM) has used mask voting for federated LLM pruning.\n\n- The paper lacks formal analysis of why certain choices (e.g., weight scaling degrades performance) work or fail.\n\n- All experiments assume IID or single-dataset settings; real-world FL often involves non-IID, skewed data, which may affect mask voting. Can authors clarify on this?", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces FedPrLLM, a general federated pruning framework for large language models (LLMs) that allows clients to perform local unstructured pruning and share only binary masks to preserve data privacy. The authors systematically explore three key design dimensions—comparison group (layer/row/column), weight scaling, and one-shot vs. iterative pruning—across various LLMs, sparsity levels, and datasets. They conduct extensive experiments (6 LLMs, 10 datasets) and derive practical insights (e.g., “layer-wise voting without weight scaling and one-shot pruning works best”).\n\nThis paper makes a practical and well-executed contribution by empirically benchmarking federated pruning strategies for LLMs. While the framework is not fundamentally novel and lacks theoretical depth, its findings are well-supported, reproducible, and directly useful for the community. A stronger baseline comparison and heterogeneity analysis would elevate it further.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper enables LLM pruning under federated settings with privacy preservation, which is interesting and underexplored.\n\n- The paper did large-scale experiments with 6 LLMs, 10 datasets, multiple sparsity levels. Ablation studies support key claims.\n\n- The three pruning dimensions are practical and grounded, covering critical decisions in collaborative model compression. Shows that simpler design choices (e.g., no scaling, one-shot pruning) often outperform complex alternatives, saving computation and communication.", "weaknesses": "- The core framework is a combination of known components; prior work (e.g., FedSpaLLM) has used mask voting for federated LLM pruning.\n\n- The paper lacks formal analysis of why certain choices (e.g., weight scaling degrades performance) work or fail.\n\n- All experiments assume IID or single-dataset settings; real-world FL often involves non-IID, skewed data, which may affect mask voting. Can authors clarify on this?", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761894058627}, {"id": "BAKc2zWjJs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15594/Reviewer_qJji"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces FedPrLLM, a federated pruning framework for privacy-preserving compression of Large Language Models (LLMs). Traditional LLM pruning often requires public calibration data, which is problematic for privacy-sensitive domains. FedPrLLM addresses this by letting each client compute a local pruning mask matrix using its private calibration data, which is then aggregated by a central server to prune the global model.", "review_text": "The paper introduces FedPrLLM, a federated pruning framework for privacy-preserving compression of Large Language Models (LLMs). Traditional LLM pruning often requires public calibration data, which is problematic for privacy-sensitive domains. FedPrLLM addresses this by letting each client compute a local pruning mask matrix using its private calibration data, which is then aggregated by a central server to prune the global model.", "strengths": "- The integration of federated learning (FL) and LLM pruning is a new direction. FedPrLLM fills a clear research gap in *privacy-preserving model compression*\n- The experiments are extensive and methodical, covering multiple models, pruning ratios, datasets, and local methods. This level of rigor strengthens the validity of the findings.\n- The paper isolates three well-defined research questions (comparison group, scaling, pruning strategy), which makes the investigation systematic and easy to follow.", "weaknesses": "- Although the framework is federated, experiments appear to use *simulated clients* (split calibration data) rather than real-world heterogeneous environments with non-IID data. This limits the external validity in true FL settings.\n- The study ignores structured pruning, which is often preferred for real deployment.\n- While communication costs are briefly reported, there’s limited exploration of real-time efficiency, scalability, or energy savings.", "questions": "- Add analysis or experiments showing that sharing mask matrices does not leak sensitive information. Techniques like differential privacy could be integrated.\n- Include empirical results on communication time, computation cost, and memory savings, not just perplexity, to demonstrate real deployment viability.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces FedPrLLM, a federated pruning framework for privacy-preserving compression of Large Language Models (LLMs). Traditional LLM pruning often requires public calibration data, which is problematic for privacy-sensitive domains. FedPrLLM addresses this by letting each client compute a local pruning mask matrix using its private calibration data, which is then aggregated by a central server to prune the global model.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The integration of federated learning (FL) and LLM pruning is a new direction. FedPrLLM fills a clear research gap in *privacy-preserving model compression*\n- The experiments are extensive and methodical, covering multiple models, pruning ratios, datasets, and local methods. This level of rigor strengthens the validity of the findings.\n- The paper isolates three well-defined research questions (comparison group, scaling, pruning strategy), which makes the investigation systematic and easy to follow.", "weaknesses": "- Although the framework is federated, experiments appear to use *simulated clients* (split calibration data) rather than real-world heterogeneous environments with non-IID data. This limits the external validity in true FL settings.\n- The study ignores structured pruning, which is often preferred for real deployment.\n- While communication costs are briefly reported, there’s limited exploration of real-time efficiency, scalability, or energy savings.", "questions": "- Add analysis or experiments showing that sharing mask matrices does not leak sensitive information. Techniques like differential privacy could be integrated.\n- Include empirical results on communication time, computation cost, and memory savings, not just perplexity, to demonstrate real deployment viability.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761320679484}], "openreview_url": "https://openreview.net/forum?id=maQs92RfZ8", "arxiv_id": "2505.13547", "paper_pdf": "papers/maQs92RfZ8.pdf", "paper_pdf_sha256": "0f832bd2d83e0aa83cc5bca4f46bcf23535fe22aacba18b90418dccb761e7e30", "paper_pdf_bytes": 977965, "paper_pdf_source": "openreview", "code_url": "https://github.com/Pengxin-Guo/FedPrLLM", "code_repository": "Pengxin-Guo/FedPrLLM", "code_commit": "7e91d12e0e3250a2790e534d15140ba163cf7739", "code_archive": "repos/maQs92RfZ8.zip", "code_archive_sha256": "e199692b9595ba88bd1c0e513a67d2dcd63eaf197fe09894a93ff2ac59b8fbf3", "code_archive_bytes": 11259, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 10, "github_languages": {"Python": 48799}, "github_archived": false, "github_pushed_at": "2025-05-15T13:33:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/exploring-federated-pruning-for-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Q5CLpqbrFM", "year": 2025, "status": "rejected", "title": "Learning Representations of Instruments for Partial Identification of Treatment Effects", "authors": ["Jonas Schweisthal", "Dennis Frauen", "Maresa Schröder", "Konstantin Hess", "Niki Kilbertus", "Stefan Feuerriegel"], "authorids": ["~Jonas_Schweisthal1", "~Dennis_Frauen1", "~Maresa_Schröder1", "~Konstantin_Hess1", "~Niki_Kilbertus1", "~Stefan_Feuerriegel1"], "authors_source": "OpenReview API", "abstract": "Reliable estimation of treatment effects from observational data is important in many disciplines, such as medicine. However, estimation is challenging when unconfoundedness as a standard assumption in the causal inference literature is violated. In this work, we leverage arbitrary (potentially high-dimensional) instruments to estimate bounds on the conditional average treatment effect (CATE). Our contributions are three-fold: (1) We propose a novel approach for partial identification through a mapping of instruments to a discrete representation space so that we yield valid bounds on the CATE. This is crucial for reliable decision-making in real-world applications. (2) We derive a two-step method that learns tight bounds using a tailored neural partitioning of the latent instrument space. As a result, we avoid instability issues due to numerical approximations or adversarial training. Furthermore, our procedure aims to reduce the estimation variance in finite-sample settings to yield more reliable estimates. (3) We show theoretically that our procedure obtains valid bounds while reducing estimation variance and we perform experiments to demonstrate the effectiveness across various settings. Overall, our procedure offers a novel path for practitioners to make use of potentially high-dimensional instruments (e.g., as in Mendelian randomization).", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "TMRkMRwzzF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6977/Reviewer_hFWX"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper provides a method for the partial identification of treatment effects when potentially high-dimensional instruments belonging to a broad class of domains are available. To overcome the potential of overly conservative bounds, it proposes a discrete embedding scheme for the instruments. It provides theoretical results for the validity of the proposed bounds and for estimator efficiency. It provides an empirical evaluation of the method using a well motivated simulated data generating process.", "review_text": "This paper provides a method for the partial identification of treatment effects when potentially high-dimensional instruments belonging to a broad class of domains are available. To overcome the potential of overly conservative bounds, it proposes a discrete embedding scheme for the instruments. It provides theoretical results for the validity of the proposed bounds and for estimator efficiency. It provides an empirical evaluation of the method using a well motivated simulated data generating process.", "strengths": "This paper, to my knowledge, provides a novel solution to a well motivated problem. It provides thorough theoretical results and adequate empirical evaluation. The writing quality and organization is excellent. The availability of high dimensional instruments is common in many settings beyond the medical domain, so I believe the potential for impact is high.", "weaknesses": "A discussion on potential applications outside of medicine may help to increase the visibility of this work. Perhaps by extending Appendix B.", "questions": "I have no further questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper provides a method for the partial identification of treatment effects when potentially high-dimensional instruments belonging to a broad class of domains are available. To overcome the potential of overly conservative bounds, it proposes a discrete embedding scheme for the instruments. It provides theoretical results for the validity of the proposed bounds and for estimator efficiency. It provides an empirical evaluation of the method using a well motivated simulated data generating process.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "This paper, to my knowledge, provides a novel solution to a well motivated problem. It provides thorough theoretical results and adequate empirical evaluation. The writing quality and organization is excellent. The availability of high dimensional instruments is common in many settings beyond the medical domain, so I believe the potential for impact is high.", "weaknesses": "A discussion on potential applications outside of medicine may help to increase the visibility of this work. Perhaps by extending Appendix B.", "questions": "I have no further questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730832434119}, {"id": "30XNVJSWeH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6977/Reviewer_EWM3"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a partial idetification approach of the conditional average treatment effect (CATE) that allows for arbitrary instruments. In particular, the proposed appraoch first maps the intstruments to a discrete representation space, and then leverages partial identification results for discrete intruments to bound the CATE. The representation mapping is learned by minimizing the length of the bounding interval while stabilizing finite sample performance. The authors also provided a neural implementation of the proposed method and performed simulation experiments.", "review_text": "This paper proposes a partial idetification approach of the conditional average treatment effect (CATE) that allows for arbitrary instruments. In particular, the proposed appraoch first maps the intstruments to a discrete representation space, and then leverages partial identification results for discrete intruments to bound the CATE. The representation mapping is learned by minimizing the length of the bounding interval while stabilizing finite sample performance. The authors also provided a neural implementation of the proposed method and performed simulation experiments.", "strengths": "- I appreciate the clear comparison with existing works. \n- The paper provided clear description of a neural implementation of the proposed approach, and there is some innovation in the custom loss function.", "weaknesses": "Overall, the theoretical contribution of the paper appears limited (some of them are very straight forward applications of the existing results for discrete instruments) and may benefit from further exploration.\n- The paper only provided finite samle guarantees for the second stage nuisance functions, but not for the final bounds, which might be more useful (for e.g. guaging the quality of the estimation) and easier to interpret. For instance it would be nice to have some theoretical results on the coverage (or scaled coverage - coverage/length of interval) of the estimated bounds.\n- The number of partitions ($k$) is an important hyperparameter, but the paper did not provide anytheoretical results on how the estimation error or the tightness of the bounds would scale with $k$.", "questions": "- Even the experiements in the high dimensional settings had only 20 dimensional instruments. I am not familar with Mendelian randomization with SNPs, so I wonder if this falls into the normal range of such applications. \n- Is the bound from Theorem 1 sharp?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a partial idetification approach of the conditional average treatment effect (CATE) that allows for arbitrary instruments. In particular, the proposed appraoch first maps the intstruments to a discrete representation space, and then leverages partial identification results for discrete intruments to bound the CATE. The representation mapping is learned by minimizing the length of the bounding interval while stabilizing finite sample performance. The authors also provided a neural implementation of the proposed method and performed simulation experiments.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- I appreciate the clear comparison with existing works. \n- The paper provided clear description of a neural implementation of the proposed approach, and there is some innovation in the custom loss function.", "weaknesses": "Overall, the theoretical contribution of the paper appears limited (some of them are very straight forward applications of the existing results for discrete instruments) and may benefit from further exploration.\n- The paper only provided finite samle guarantees for the second stage nuisance functions, but not for the final bounds, which might be more useful (for e.g. guaging the quality of the estimation) and easier to interpret. For instance it would be nice to have some theoretical results on the coverage (or scaled coverage - coverage/length of interval) of the estimated bounds.\n- The number of partitions ($k$) is an important hyperparameter, but the paper did not provide anytheoretical results on how the estimation error or the tightness of the bounds would scale with $k$.", "questions": "- Even the experiements in the high dimensional settings had only 20 dimensional instruments. I am not familar with Mendelian randomization with SNPs, so I wonder if this falls into the normal range of such applications. \n- Is the bound from Theorem 1 sharp?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730503816380}, {"id": "jmOjus7jZt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6977/Reviewer_bx94"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "Reliable estimation of treatment effects from observational data is crucial in many disciplines and is a challenge when unconfoundedness assumption in the causal inference literature is violated. In this paper, the authors leverage arbitrary instruments to estimate bounds on the conditional average treatment effect (CATE). First, the authors propose a novel approach for partial identification by mapping instruments to a discrete representation space such that valid bounds on the CATE are yielded. Second, the authors derive a two-step procedure that learns tight bounds using a tailored neural partitioning of the latent instrument space. Thus, instability issues due to numerical approximations or adversarial training are avoided. Third, the authors provide theoretical studies on the proposed approach. Finally, the authors perform extensive experiments to demonstrate the effectiveness across various settings.", "review_text": "Reliable estimation of treatment effects from observational data is crucial in many disciplines and is a challenge when unconfoundedness assumption in the causal inference literature is violated. In this paper, the authors leverage arbitrary instruments to estimate bounds on the conditional average treatment effect (CATE). First, the authors propose a novel approach for partial identification by mapping instruments to a discrete representation space such that valid bounds on the CATE are yielded. Second, the authors derive a two-step procedure that learns tight bounds using a tailored neural partitioning of the latent instrument space. Thus, instability issues due to numerical approximations or adversarial training are avoided. Third, the authors provide theoretical studies on the proposed approach. Finally, the authors perform extensive experiments to demonstrate the effectiveness across various settings.", "strengths": "This paper investigates the conditional average treatment effect (CATE), assuming the presence of unobserved confounders. Further, researchers encounter complex instrument information (e.g., continuous/high-dimensional), which makes the computation hard in many circumstances. The authors aim to develop estimators that can provide reliable analysis of the above assumptions. Unlike the existing literature, the authors establish bounding estimators of the target quantity and provide theoretical studies of the constructed estimators. A merit of the paper is that it is well-written, so readers can easily follow its notions.", "weaknesses": "There are some notational problems in the paper. For example:\n1. Notational problems in Eqn. (10).\n2. According to Eqn. (8) (or Eqn. (14)), $\\phi$ should be a representation map, and $\\Phi$ should be a set of all representations. It is quite confusing to write $\\Phi(Z)=l$ (e.g., see line 323). \n3. Please also state the explicit form of the constants $c$, $d$ in Theorem 2.\nPlease recheck the notations to avoid any confusions.", "questions": "Marks will be adjusted according to the replies. Below are some feedback/questions:\n1. What is the point of building bounding estimators? As mentioned in the paper, existing literature focuses on point estimators. Indeed, we can build the confidence interval based on the point estimator. It is well-known that these are two concepts, but it is odds to build bounding estimators present in the paper.\n2. Theorem 2 gives the asymptotic properties of the nuisance estimators $\\hat{\\mu}$ and $\\hat{pi}$ when $\\phi$ is arbitrary. It would be better to study the asymptotic properties of the nuisance estimators due to optimal $\\phi^*$ after Eqn. (8).\n3. It is strange to have $N_b$ in Algorithm 1 (see lines 387, 388).\n4. The auxiliary guidance loss is not clear enough. It would be better if the authors could provide an explicit formula for the loss similar to bandwidth minimization loss and regularization loss.\n5. The size of $k$ should affect the bounded estimators. Please demonstrate a scientific way to choose $k$. In the existing experiments, authors study results for different $k$. In any case, it seems that increasing $k$ would give better results. It should not be correct in general.\n6. Could you provide the results of the following experiment:\nassuming linear generative model, nonlinear ML methods are used to estimate the nuisance functions and then construct the corresponding bandwidth;\nassuming a nonlinear generative model, linear ML methods are used to estimate the nuisance functions and then construct the corresponding bandwidth.\n7. Would the constructed bounding estimators be doubly robust?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Reliable estimation of treatment effects from observational data is crucial in many disciplines and is a challenge when unconfoundedness assumption in the causal inference literature is violated. In this paper, the authors leverage arbitrary instruments to estimate bounds on the conditional average treatment effect (CATE). First, the authors propose a novel approach for partial identification by mapping instruments to a discrete representation space such that valid bounds on the CATE are yielded. Second, the authors derive a two-step procedure that learns tight bounds using a tailored neural partitioning of the latent instrument space. Thus, instability issues due to numerical approximations or adversarial training are avoided. Third, the authors provide theoretical studies on the proposed approach. Finally, the authors perform extensive experiments to demonstrate the effectiveness across various settings.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper investigates the conditional average treatment effect (CATE), assuming the presence of unobserved confounders. Further, researchers encounter complex instrument information (e.g., continuous/high-dimensional), which makes the computation hard in many circumstances. The authors aim to develop estimators that can provide reliable analysis of the above assumptions. Unlike the existing literature, the authors establish bounding estimators of the target quantity and provide theoretical studies of the constructed estimators. A merit of the paper is that it is well-written, so readers can easily follow its notions.", "weaknesses": "There are some notational problems in the paper. For example:\n1. Notational problems in Eqn. (10).\n2. According to Eqn. (8) (or Eqn. (14)), $\\phi$ should be a representation map, and $\\Phi$ should be a set of all representations. It is quite confusing to write $\\Phi(Z)=l$ (e.g., see line 323). \n3. Please also state the explicit form of the constants $c$, $d$ in Theorem 2.\nPlease recheck the notations to avoid any confusions.", "questions": "Marks will be adjusted according to the replies. Below are some feedback/questions:\n1. What is the point of building bounding estimators? As mentioned in the paper, existing literature focuses on point estimators. Indeed, we can build the confidence interval based on the point estimator. It is well-known that these are two concepts, but it is odds to build bounding estimators present in the paper.\n2. Theorem 2 gives the asymptotic properties of the nuisance estimators $\\hat{\\mu}$ and $\\hat{pi}$ when $\\phi$ is arbitrary. It would be better to study the asymptotic properties of the nuisance estimators due to optimal $\\phi^*$ after Eqn. (8).\n3. It is strange to have $N_b$ in Algorithm 1 (see lines 387, 388).\n4. The auxiliary guidance loss is not clear enough. It would be better if the authors could provide an explicit formula for the loss similar to bandwidth minimization loss and regularization loss.\n5. The size of $k$ should affect the bounded estimators. Please demonstrate a scientific way to choose $k$. In the existing experiments, authors study results for different $k$. In any case, it seems that increasing $k$ would give better results. It should not be correct in general.\n6. Could you provide the results of the following experiment:\nassuming linear generative model, nonlinear ML methods are used to estimate the nuisance functions and then construct the corresponding bandwidth;\nassuming a nonlinear generative model, linear ML methods are used to estimate the nuisance functions and then construct the corresponding bandwidth.\n7. Would the constructed bounding estimators be doubly robust?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730354567089}, {"id": "RgUrODvSOb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6977/Reviewer_ePQ5"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper introduces a method for estimating treatment effects from observational data despite violations of unconfoundedness. It presents a two-step procedure using high-dimensional latent instruments to derive valid bounds on the Conditional Average Treatment Effect.", "review_text": "This paper introduces a method for estimating treatment effects from observational data despite violations of unconfoundedness. It presents a two-step procedure using high-dimensional latent instruments to derive valid bounds on the Conditional Average Treatment Effect.", "strengths": "1.\tThe study addresses an intriguing and valuable problem in causal inference.\n\n2.\tThe writing is clear and accessible, making complex concepts easier to understand.", "weaknesses": "1. The method appears to be a straightforward combination of existing techniques, lacking clear theoretical innovation.  \n\n2. Although the paper highlights the advantage of utilizing high-dimensional instruments, there are limited experiments involving such instruments in the evaluation phase.  \n\n3. The experimental results show no substantial improvements, casting doubt on the method's effectiveness.  \n\n4. While the authors discuss the lack of comparison implementations, it would strengthen the argument to adapt baseline methods to the current setting and compare their performance with that of the proposed method, making the results more persuasive.\n\n5. The final loss function includes two hyperparameters, making it crucial to conduct a sensitivity analysis to evaluate the method's robustness. However, the absence of such experiments raises concerns about the reliability and generalizability of the proposed approach.\n\n6.  The assumption that all observed variables are confounders may be stringent and difficult to satisfy in practical scenarios.\n\n7. The two-stage method presents several limitations, such as potential inefficiencies in estimation and sensitivity to the parameters selected at each stage, which may compromise the overall robustness of the results.", "questions": "See Weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a method for estimating treatment effects from observational data despite violations of unconfoundedness. It presents a two-step procedure using high-dimensional latent instruments to derive valid bounds on the Conditional Average Treatment Effect.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "1.\tThe study addresses an intriguing and valuable problem in causal inference.\n\n2.\tThe writing is clear and accessible, making complex concepts easier to understand.", "weaknesses": "1. The method appears to be a straightforward combination of existing techniques, lacking clear theoretical innovation.  \n\n2. Although the paper highlights the advantage of utilizing high-dimensional instruments, there are limited experiments involving such instruments in the evaluation phase.  \n\n3. The experimental results show no substantial improvements, casting doubt on the method's effectiveness.  \n\n4. While the authors discuss the lack of comparison implementations, it would strengthen the argument to adapt baseline methods to the current setting and compare their performance with that of the proposed method, making the results more persuasive.\n\n5. The final loss function includes two hyperparameters, making it crucial to conduct a sensitivity analysis to evaluate the method's robustness. However, the absence of such experiments raises concerns about the reliability and generalizability of the proposed approach.\n\n6.  The assumption that all observed variables are confounders may be stringent and difficult to satisfy in practical scenarios.\n\n7. The two-stage method presents several limitations, such as potential inefficiencies in estimation and sensitivity to the parameters selected at each stage, which may compromise the overall robustness of the results.", "questions": "See Weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730294904154}], "openreview_url": "https://openreview.net/forum?id=Q5CLpqbrFM", "arxiv_id": "2410.08976", "paper_pdf": "papers/Q5CLpqbrFM.pdf", "paper_pdf_sha256": "78ed1fb7e0c8058d17c09de55c23dba0815eb61958e93f49b3121f194d09c6fe", "paper_pdf_bytes": 2027941, "paper_pdf_source": "openreview", "code_url": "https://github.com/JSchweisthal/ComplexPartialIdentif", "code_repository": "JSchweisthal/ComplexPartialIdentif", "code_commit": "260547060aa0f2b9e71f0179768f8a534bf6fd19", "code_archive": "repos/Q5CLpqbrFM.zip", "code_archive_sha256": "833358fb96d33775e0a634fb59489f4a4f0e0163187deb4dc1cc534cce81ddbf", "code_archive_bytes": 11992, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 10, "github_languages": {"Python": 36573}, "github_archived": false, "github_pushed_at": "2024-10-11T16:02:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-representations-of-instruments-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qYb0CANLGC", "year": 2024, "status": "rejected", "title": "Auto-Regressive Next-Token Predictors are Universal Learners", "authors": ["eran malach"], "authorids": ["~eran_malach1"], "authors_source": "OpenReview API", "abstract": "Large language models display remarkable capabilities in logical and mathematical reasoning, allowing them to solve complex tasks. Interestingly, these abilities emerge in networks trained on the simple task of next-token prediction. In this work, we present a theoretical framework for studying auto-regressive next-token predictors. We demonstrate that even simple models such as linear next-token predictors, trained on Chain-of-Thought (CoT) data, can approximate any function efficiently computed by a Turing machine. We introduce a new complexity measure---length complexity---which measures the number of intermediate tokens in a CoT sequence required to approximate some target function, and analyze the interplay between length complexity and other notions of complexity. Finally, we show experimentally that simple next-token predictors, such as linear networks and shallow Multi-Layer Perceptrons (MLPs), display non-trivial performance on text generation and arithmetic tasks. Our results demonstrate that the power of today's LLMs can be attributed, to a great extent, to the auto-regressive next-token training scheme, and not necessarily to a particular choice of architecture.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "qphMWZRIUp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3681/Reviewer_RXBR"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper attempts to provide a theoretical explanation of how auto-regressive prediction in LLMs can enable them to perform nontrivial tasks.\n\nThe theoretical idea is that a dataset including intermediate steps can enable simple (in fact, linear) next-token predictors to learn to perform arbitrary computable functions.\n\nThe paper develops a theory formalizing these ideas.\nIt first presents a notion of autoregressive learnability modeled on PAC learnability.\nIt then derives results on the approximability of problems with autoregressive predictors -- that is, in the presence of intermediate steps. Indeed, for any Turing-computable function, a linear autoregressive model can compute this function, with a number of intermediate steps polynomial in the Turing machine's runtime.\nIt finally investigates the number of required intermediate steps, at the example of k-th order parities in length-n strings. This \"length complexity\" can trade off with the computational complexity, as a shorter sequence may require more complex individual steps.\n\nIt then presents two experiments showing that simple autoregressive predictors can perform nontrivial tasks.\n\nThe first experiment qualitatively shows that a linear autoregressive model trained on TinyStories can produce mostly-reasonable-looking text.\nThe second experiment shows that, with appropriate intermediate steps and an adapted input formatting, an MLP-based autoregressive model can multiply 4-digit numbers.", "review_text": "This paper attempts to provide a theoretical explanation of how auto-regressive prediction in LLMs can enable them to perform nontrivial tasks.\n\nThe theoretical idea is that a dataset including intermediate steps can enable simple (in fact, linear) next-token predictors to learn to perform arbitrary computable functions.\n\nThe paper develops a theory formalizing these ideas.\nIt first presents a notion of autoregressive learnability modeled on PAC learnability.\nIt then derives results on the approximability of problems with autoregressive predictors -- that is, in the presence of intermediate steps. Indeed, for any Turing-computable function, a linear autoregressive model can compute this function, with a number of intermediate steps polynomial in the Turing machine's runtime.\nIt finally investigates the number of required intermediate steps, at the example of k-th order parities in length-n strings. This \"length complexity\" can trade off with the computational complexity, as a shorter sequence may require more complex individual steps.\n\nIt then presents two experiments showing that simple autoregressive predictors can perform nontrivial tasks.\n\nThe first experiment qualitatively shows that a linear autoregressive model trained on TinyStories can produce mostly-reasonable-looking text.\nThe second experiment shows that, with appropriate intermediate steps and an adapted input formatting, an MLP-based autoregressive model can multiply 4-digit numbers.", "strengths": "- Provides an interesting angle on the success of chain-of-thought in enabling LMs to perform more complex tasks\n\n- provides both theoretical analysis and empirical evidence", "weaknesses": "- The TinyStories experiment (Section 3.1) lacks quantitative evaluation. It is unclear if the examples shown are representative.\n\n- Multiplication experiment: unlike the TinyStories experiment, the model is not linear -- is this important? Why not use the same model for both experiments?\n\n- There appears to be a potential mismatch with realistic chain-of-thought prompting in that the learnability theory developed here assumes that the tasks are available together with their full sequences of intermediate steps in the training set, whereas in reality prompts containing chain-of-thought demonstrations may be quite unnatural, and are quite likely not to have appeared in that form in the training set. In this sense, the learnability theory, if understood as applying to unsupervised LLM training, does not explain why LMs would be able to deal with unnatural prompt formats.\n\n- In the multiplication experiment, the calculations are decomposed into \"more intermediate steps than in Liu&Low\". Furthermore, they involve padding to get all strings to have the same length, apparently unlike Liu&Low. Both of these steps may impact the comparison with the results from Liu&Low. Furthermore, unlike the MLP, GPT-3.5 is (judging by Figure 2) not prompted to output intermediate steps. Thus, it is unclear what one can learn from the comparison between the models in Figure 2 (right), and the statement \"outperforms GPT-4\" in the introduction may not be fully supported.", "questions": "- Section 3.1: what is the objective function -- is it cross-entropy? Is softmax applied on the linear output? \n\n- Section 2.1, first paragraph: Why is \\mathcal{D} a distribution -- given that its support is finite and that the only probabilistic statement made then holds with probability 1 (next line), could \\mathcal{D} just as well be a subset of X \\times Z_T?\n\n- Definition 2: when encountering this, I wondered: is T a global free variable (i.e., AR-Learnability is defined w.r.t. T), or is T dependent on \\mathcal{D}? This is resolved in Footnote 2. Maybe disambiguate this at the beginning of Section 2.1?\n\n- minor: \"Multi-Linear Perceptron (MLP)\" (page 2, third paragraph) --> the standard reading of \"MLP\" appears to be Multi-Layer Perceptron, and indeed the model appears to be of this type", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper attempts to provide a theoretical explanation of how auto-regressive prediction in LLMs can enable them to perform nontrivial tasks.\n\nThe theoretical idea is that a dataset including intermediate steps can enable simple (in fact, linear) next-token predictors to learn to perform arbitrary computable functions.\n\nThe paper develops a theory formalizing these ideas.\nIt first presents a notion of autoregressive learnability modeled on PAC learnability.\nIt then derives results on the approximability of problems with autoregressive predictors -- that is, in the presence of intermediate steps. Indeed, for any Turing-computable function, a linear autoregressive model can compute this function, with a number of intermediate steps polynomial in the Turing machine's runtime.\nIt finally investigates the number of required intermediate steps, at the example of k-th order parities in length-n strings. This \"length complexity\" can trade off with the computational complexity, as a shorter sequence may require more complex individual steps.\n\nIt then presents two experiments showing that simple autoregressive predictors can perform nontrivial tasks.\n\nThe first experiment qualitatively shows that a linear autoregressive model trained on TinyStories can produce mostly-reasonable-looking text.\nThe second experiment shows that, with appropriate intermediate steps and an adapted input formatting, an MLP-based autoregressive model can multiply 4-digit numbers.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- Provides an interesting angle on the success of chain-of-thought in enabling LMs to perform more complex tasks\n\n- provides both theoretical analysis and empirical evidence", "weaknesses": "- The TinyStories experiment (Section 3.1) lacks quantitative evaluation. It is unclear if the examples shown are representative.\n\n- Multiplication experiment: unlike the TinyStories experiment, the model is not linear -- is this important? Why not use the same model for both experiments?\n\n- There appears to be a potential mismatch with realistic chain-of-thought prompting in that the learnability theory developed here assumes that the tasks are available together with their full sequences of intermediate steps in the training set, whereas in reality prompts containing chain-of-thought demonstrations may be quite unnatural, and are quite likely not to have appeared in that form in the training set. In this sense, the learnability theory, if understood as applying to unsupervised LLM training, does not explain why LMs would be able to deal with unnatural prompt formats.\n\n- In the multiplication experiment, the calculations are decomposed into \"more intermediate steps than in Liu&Low\". Furthermore, they involve padding to get all strings to have the same length, apparently unlike Liu&Low. Both of these steps may impact the comparison with the results from Liu&Low. Furthermore, unlike the MLP, GPT-3.5 is (judging by Figure 2) not prompted to output intermediate steps. Thus, it is unclear what one can learn from the comparison between the models in Figure 2 (right), and the statement \"outperforms GPT-4\" in the introduction may not be fully supported.", "questions": "- Section 3.1: what is the objective function -- is it cross-entropy? Is softmax applied on the linear output? \n\n- Section 2.1, first paragraph: Why is \\mathcal{D} a distribution -- given that its support is finite and that the only probabilistic statement made then holds with probability 1 (next line), could \\mathcal{D} just as well be a subset of X \\times Z_T?\n\n- Definition 2: when encountering this, I wondered: is T a global free variable (i.e., AR-Learnability is defined w.r.t. T), or is T dependent on \\mathcal{D}? This is resolved in Footnote 2. Maybe disambiguate this at the beginning of Section 2.1?\n\n- minor: \"Multi-Linear Perceptron (MLP)\" (page 2, third paragraph) --> the standard reading of \"MLP\" appears to be Multi-Layer Perceptron, and indeed the model appears to be of this type", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698798496731}, {"id": "E54mq2pPPE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3681/Reviewer_QtWu"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces a theoretical framework for studying auto-regressive next-token prediction models. The key idea is that auto-regressive learning, where the model is trained to predict the next token given previous tokens, allows the model to utilize intermediate \"chain-of-thought\" computations. This gives the model more information compared to standard supervised learning where only input-output pairs are provided.\n\nThe authors show theoretically that even simple auto-regressive models like linear next-token predictors can approximate any efficiently Turing-computable function, if provided with appropriate intermediate chain-of-thought supervision. They also introduce a new notion of \"length complexity\" which measures the number of intermediate tokens needed to learn a target function. Length complexity allows trading off sample/computational complexity and length complexity.\n\nExperiments on text generation and arithmetic tasks demonstrate that simple models like linear networks and small MLPs can perform well when trained with next-token prediction and chain-of-thought supervision.", "review_text": "The paper introduces a theoretical framework for studying auto-regressive next-token prediction models. The key idea is that auto-regressive learning, where the model is trained to predict the next token given previous tokens, allows the model to utilize intermediate \"chain-of-thought\" computations. This gives the model more information compared to standard supervised learning where only input-output pairs are provided.\n\nThe authors show theoretically that even simple auto-regressive models like linear next-token predictors can approximate any efficiently Turing-computable function, if provided with appropriate intermediate chain-of-thought supervision. They also introduce a new notion of \"length complexity\" which measures the number of intermediate tokens needed to learn a target function. Length complexity allows trading off sample/computational complexity and length complexity.\n\nExperiments on text generation and arithmetic tasks demonstrate that simple models like linear networks and small MLPs can perform well when trained with next-token prediction and chain-of-thought supervision.", "strengths": "(1) Provides an elegant theoretical framework for studying auto-regressive next-token prediction models, an important class of models in NLP.\n\n(2) Establishes strong learnability and approximation guarantees for simple models like linear predictors when trained auto-regressively.\n\n(3) Introduces the novel concept of \"length complexity\" to capture chain-of-thought requirements. Relates length complexity to sample and computational complexity.", "weaknesses": "(1) The theoretical results rely on very strong assumptions about availability of chain-of-thought training data, which may be unrealistic.\n\n(2) More analysis would be useful on how length complexity scales with problem complexity for different hypothesis classes.\n\n(3) Additional validation on more complex architectures like Transformers would strengthen the conclusions about training scheme vs architecture.\n\n(4) The proposed linear models are not exactly equivalent to the classical linear models analyzed. Some parameters are shared across time steps.\n\n(5) Non autoregressive models are not discussed, tested and compared. It will be to know if given large amount of training data, non-autoregressive models can perform on par with autoregressive models.\n\n(6) On MULTIPLICATION experiments, models are trained and tested on 4-digit numbers. The model is likely to memorize the patterns instead of generalizations, especially this paper use special tokenization for digits and signs.", "questions": "This paper misses some critical references:\n\n(1) Training of smaller student models on CoT data has been investigated in several earlier papers [1, 2]. \n(2) Language Models for Arithmetic Tasks have been extensively discussed and studied in [3].\n\n\n\n[1] Li et al. Explanations from Large Language Models Make Small Reasoners Better. 2022.\n[2] Magister et al.Teaching Small Language Models to Reason. ACL, 2023.\n[3] Qian et al. Limitations of Language Models in Arithmetic and Symbolic Induction. ACL, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a theoretical framework for studying auto-regressive next-token prediction models. The key idea is that auto-regressive learning, where the model is trained to predict the next token given previous tokens, allows the model to utilize intermediate \"chain-of-thought\" computations. This gives the model more information compared to standard supervised learning where only input-output pairs are provided.\n\nThe authors show theoretically that even simple auto-regressive models like linear next-token predictors can approximate any efficiently Turing-computable function, if provided with appropriate intermediate chain-of-thought supervision. They also introduce a new notion of \"length complexity\" which measures the number of intermediate tokens needed to learn a target function. Length complexity allows trading off sample/computational complexity and length complexity.\n\nExperiments on text generation and arithmetic tasks demonstrate that simple models like linear networks and small MLPs can perform well when trained with next-token prediction and chain-of-thought supervision.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "(1) Provides an elegant theoretical framework for studying auto-regressive next-token prediction models, an important class of models in NLP.\n\n(2) Establishes strong learnability and approximation guarantees for simple models like linear predictors when trained auto-regressively.\n\n(3) Introduces the novel concept of \"length complexity\" to capture chain-of-thought requirements. Relates length complexity to sample and computational complexity.", "weaknesses": "(1) The theoretical results rely on very strong assumptions about availability of chain-of-thought training data, which may be unrealistic.\n\n(2) More analysis would be useful on how length complexity scales with problem complexity for different hypothesis classes.\n\n(3) Additional validation on more complex architectures like Transformers would strengthen the conclusions about training scheme vs architecture.\n\n(4) The proposed linear models are not exactly equivalent to the classical linear models analyzed. Some parameters are shared across time steps.\n\n(5) Non autoregressive models are not discussed, tested and compared. It will be to know if given large amount of training data, non-autoregressive models can perform on par with autoregressive models.\n\n(6) On MULTIPLICATION experiments, models are trained and tested on 4-digit numbers. The model is likely to memorize the patterns instead of generalizations, especially this paper use special tokenization for digits and signs.", "questions": "This paper misses some critical references:\n\n(1) Training of smaller student models on CoT data has been investigated in several earlier papers [1, 2]. \n(2) Language Models for Arithmetic Tasks have been extensively discussed and studied in [3].\n\n\n\n[1] Li et al. Explanations from Large Language Models Make Small Reasoners Better. 2022.\n[2] Magister et al.Teaching Small Language Models to Reason. ACL, 2023.\n[3] Qian et al. Limitations of Language Models in Arithmetic and Symbolic Induction. ACL, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698536213407}, {"id": "h0pQwYSSsk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3681/Reviewer_ry1v"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a theoretical framework for analyzing autoregressive learners. At its core is a formal definition of autoregressive learnability (AR Learnability), analogous to PAC Learnability.\n\nAR Learnability is defined with respect to distributions over strings over a finite set of tokens $\\mathbb{D}$. More specifically, these are distributions $\\mathcal{D}$ over contexts $\\mathcal{X} = \\mathbb{D}^n$ and continuations $\\mathcal{Z} = \\mathbb{D}^T$ for some integers $n$ and $T$.\nBriefly stated, a hypothesis class $\\mathcal{H}$ is AR Learnable if there is an algorithm that, for any realizable distribution $\\mathcal{D}$, takes as input samples from $\\mathcal{D}$ and outputs (w.h.p.) a hypothesis $h \\in \\mathcal{H}$ such that $h(x,z_{\\<t}) = z_t$ for most $t \\leq T$. That is, a hypothesis that can (mostly) correctly generate the continuation of the context $x$ as in the distribution $\\mathcal{D}$.\n\nThis definition is rigorously analyzed. The first batch of theoretical findings (theorems) are concerned with the universality of AR learners:\n1. The product of $T$ PAC-learnable hypothesis classes AR-learable [Theorem 3].\n2. A generalization of AR-learability to the setting of computing/approximating functions; AR-learnable classes are approximable [Theorem 5]\n3. Linear AR models ($\\mathcal{H}^{\\mathrm{lin}}$) can approximate any $f \\in \\mathrm{TIME}(T(n))$ w.r.t any distribution $\\mathcal{D}$ given access to a dataset of $\\mathrm{poly}(\\mathrm{T(n)})$-length strings [Corollary 8].\n\nThe last item above is particularly significant because, in the supervised (i.e., \"standard\") setting, the analogous class to $\\mathcal{H}^{\\mathrm{lin}}$ are known to be so weak that they cannot learn the XOR function. To my understanding, Corollary 8 is proven by representing the function $f$ as a linear threshold circuit $C_f$; the dataset is populated by sampling an $x \\gets \\mathcal{D}$ and adding a row that represents the outputs of intermediate gates of $C_f(x)$.\n\nConsidering the above proof, it is perhaps not surprising that AR learners are universal learners; the \"heavy-lifting\" is done by the dataset, which verbosely represents the computed function. Thus, the authors define and analyze a new complexity measure for AR learners, called the _length complexity_: For a given hypothesis class $\\mathcal{H}$ and a target function class $\\mathcal{F}$, the length complexity is the [minimal? see Weaknesses] number of autoregressive steps $T$ such that each $f \\in \\mathcal{F}$ can be computed  by some $h \\in \\mathcal{H}$ within $T$ autoregressive steps. Intuitively, length complexity should capture the length of a chain-of-thought used when training the model.\n\nFinally, the authors show that parities over $n$ bits can be computed by $\\mathcal{H}^{\\mathrm{lin}}$ with length complexity $O(\\log n)$. Additionally, they prove a theorem depicting a trade-off between the richness of the hypothesis class and the length complexity, by showing that parities over $n$ bits can be learned by $k$-order parities over $\\geq n$ bits, with length complexity $\\Theta(n / k)$.\n\nThe authors supplement the theoretical framework and findings with experimental evaluation. First, they show that an adaptation of $\\mathcal{H}^{\\mathrm{lin}}$ with 162M parameters exhibits non-trivial learning on TinyStories; it is often able to generate coherent text. Second, they show that a 4-layer MLP with 775M parameters can learn to multiply 4 digit numbers as well as Goat 7B, provided sufficiently long chain-of-thought (i.e., length complexity) and a custom tokenization scheme.", "review_text": "This paper introduces a theoretical framework for analyzing autoregressive learners. At its core is a formal definition of autoregressive learnability (AR Learnability), analogous to PAC Learnability.\n\nAR Learnability is defined with respect to distributions over strings over a finite set of tokens $\\mathbb{D}$. More specifically, these are distributions $\\mathcal{D}$ over contexts $\\mathcal{X} = \\mathbb{D}^n$ and continuations $\\mathcal{Z} = \\mathbb{D}^T$ for some integers $n$ and $T$.\nBriefly stated, a hypothesis class $\\mathcal{H}$ is AR Learnable if there is an algorithm that, for any realizable distribution $\\mathcal{D}$, takes as input samples from $\\mathcal{D}$ and outputs (w.h.p.) a hypothesis $h \\in \\mathcal{H}$ such that $h(x,z_{\\<t}) = z_t$ for most $t \\leq T$. That is, a hypothesis that can (mostly) correctly generate the continuation of the context $x$ as in the distribution $\\mathcal{D}$.\n\nThis definition is rigorously analyzed. The first batch of theoretical findings (theorems) are concerned with the universality of AR learners:\n1. The product of $T$ PAC-learnable hypothesis classes AR-learable [Theorem 3].\n2. A generalization of AR-learability to the setting of computing/approximating functions; AR-learnable classes are approximable [Theorem 5]\n3. Linear AR models ($\\mathcal{H}^{\\mathrm{lin}}$) can approximate any $f \\in \\mathrm{TIME}(T(n))$ w.r.t any distribution $\\mathcal{D}$ given access to a dataset of $\\mathrm{poly}(\\mathrm{T(n)})$-length strings [Corollary 8].\n\nThe last item above is particularly significant because, in the supervised (i.e., \"standard\") setting, the analogous class to $\\mathcal{H}^{\\mathrm{lin}}$ are known to be so weak that they cannot learn the XOR function. To my understanding, Corollary 8 is proven by representing the function $f$ as a linear threshold circuit $C_f$; the dataset is populated by sampling an $x \\gets \\mathcal{D}$ and adding a row that represents the outputs of intermediate gates of $C_f(x)$.\n\nConsidering the above proof, it is perhaps not surprising that AR learners are universal learners; the \"heavy-lifting\" is done by the dataset, which verbosely represents the computed function. Thus, the authors define and analyze a new complexity measure for AR learners, called the _length complexity_: For a given hypothesis class $\\mathcal{H}$ and a target function class $\\mathcal{F}$, the length complexity is the [minimal? see Weaknesses] number of autoregressive steps $T$ such that each $f \\in \\mathcal{F}$ can be computed  by some $h \\in \\mathcal{H}$ within $T$ autoregressive steps. Intuitively, length complexity should capture the length of a chain-of-thought used when training the model.\n\nFinally, the authors show that parities over $n$ bits can be computed by $\\mathcal{H}^{\\mathrm{lin}}$ with length complexity $O(\\log n)$. Additionally, they prove a theorem depicting a trade-off between the richness of the hypothesis class and the length complexity, by showing that parities over $n$ bits can be learned by $k$-order parities over $\\geq n$ bits, with length complexity $\\Theta(n / k)$.\n\nThe authors supplement the theoretical framework and findings with experimental evaluation. First, they show that an adaptation of $\\mathcal{H}^{\\mathrm{lin}}$ with 162M parameters exhibits non-trivial learning on TinyStories; it is often able to generate coherent text. Second, they show that a 4-layer MLP with 775M parameters can learn to multiply 4 digit numbers as well as Goat 7B, provided sufficiently long chain-of-thought (i.e., length complexity) and a custom tokenization scheme.", "strengths": "- This paper provides a theoretical foundation for an increasingly important topic, namely, understanding the emergent abilities of autoregressive learners.\n- Several theoretical works have already tackled the questions of autoregressive learning. However, to my knowledge, this is the first paper to propose a generic defintion analogous to PAC-learning. This is significant because it may inspire learning theorists to search for analogous results to what is known in the rich literature of PAC. For example, as the authors suggest, it would be interesting to find an intrinsic dimension of hypothesis classes that governs AR-learnability (cf. VC-dimension).\n- The theoretical framework presented in this paper is an appropriate operationalization of autoregressive learning: the definitions stay close to autoregressive learning as it is implemented in practice (at least conceptually).\n- I find the framework itself appealing. It is not easy to come up with a clean definition based on an existing phenomenon, as they real world is often \"messy\". The definitions strike a good balance between intuitiveness/simplicitly, to being non-trivial (\"mysterious\") enough to invite further analysis.\n- Overall, the paper is clearly written. Reading it was an enjoyable experience!\n- I particularly appreciate how well-organized this paper is. It first presents a fairly intuitive definition, and then provides a sequence of theorems and refinements of this definition to interesting settings. That is to say, the paper \"tells a convincing story\" about autoregressive learning.\n- This \"story\" that the theory suggests is then supported by experimental results fairly adequetly. While the experimental setup is somewhat simplistic as compared to more empirical works, I do not think that a grander setup is needed for this type of paper.\n- While I did not verify the proofs of all theorems, I have read some of them (especially Theorem 7) and they seem correct. I would also like to emphasize that the simplicity of the proofs should not be viewed as a weakness of this paper. Coming up with the right definition (from which intuitive theorems are easy to prove) is a challenging part of many theoretical works; in this case, the authors did well in carefully designing their framework.", "weaknesses": "Listed in decreasing order of significance.\n\n## TinyStories experiment is anecdotal, compares to wrong model?\nI am not sure what are the actual results being reported with the TinyStories experiment: What I found is a footnote on a 1.2 difference in perplexity between the linear predictor and GPT-2 Small---but I'm not sure what to make of this quantity. And there is a statement that the linear predictor \"often does produce coherent text\". But how often? And how do you measure coherence? While I lack the expertise to suggest a concrete experiment, I expect a higher level of rigour for a paper at ICLR. Perhaps the authors could refer to the TinyStories paper and reproduce some of the experiments there with their model.\n\nRelatedly, it seems that in the TinyStories paper there is a 28M-parameter model that achieves performance comparable to GPT2-XL. Therefore, the comparison of the linear predictor to GPT2-Small seems inappropriate---much better performance can be attained with 5x less parameters. Perhaps the authors should compare the linear predictor to [that model](https://huggingface.co/papers/2305.07759) instead. Discovering that a linear predictor requires significantly more parameters than a (good) transformer model to achieve comparable performance would undermine the main message of the paper (that the power of LLMs can be attributed to AR rather than architecture). Please correct me if you disagree that comparing to the TinyStories transformer is a more fair comparison towards this end.\n\n## Linear Decoders should be defined more slowly\n- Given the significance of linear ARs (linear decoders) throughout the paper, I suggest defining them more slowly and in their own Definition environment---rather than presenting them as an \"Example\". A Definition environment is more easy to refer back to, and encourages more formal writing.\n- A figure depicting the construction would be welcome: I had to work it out with pen and paper on the margin.\n- \"Under some margin conditions and using a convex surrogate loss function, this class is in fact learnable using SGD.\" This sentence must be supported by either a citation or a proof. As it is currently phrased it is too vague and not well-enough argued for to be used as a true statement (the current level of rigour is more appropriate for a tangential side-note).\n- Observe that this class is learnable in polynomial time: Say PAC-learnable, becuase this paper uses multiple notions of learnability.\n\n## Length complexity is not well-defined\nAs length complexity is currently defined in Definition 9, a class $\\mathcal{H}$ has infinitely many length complexities for computing a given $\\mathcal{F}$. This is because if $T$ is a length complexity of $\\mathcal{H}$ for computing $\\mathcal{F}$, then so is any $T' \\geq T$. Instead, you should define the length complexity of $\\mathcal{H}$ computing $\\mathcal{F}$ to be _the minimal $T$ for which the conditions stated in the definition hold.\n\nOn that note, it would be illuminating (and add to the cohesion of the paper) if the authors spell-out the length complexity of the construction from Theorem 7.\n\n## Missing citations\n- Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective by Feng et al. (2023): This work gives a complexity-theoretic explanation to the success of chain-of-thought, especially in mathematical reasoning. The techniques seem entirely different, and it is focused on transformer architectures (rather than focusing on the autoregressive aspect of LLMs, as in the paper currently under review). Still, it clearly fits in three of the four topics covered by the related work (all save for \"Beyond Transfomers\").\n- Fast Learning Requires Good Memory: A Time-Space Lower Bound for Parity Learning by Raz (2016): This prominent result (FOCS best paper, JACM) should probably be mentioned in the first paragraph of Section 2.3.1 that discusses hardness results for parities.\n\n## Other minor writing comments\n- Page 4 above Definition 2: \"...for all **sub-sequences** $z_{<t}$\" should be **prefixes** (sub-sequences is correct but less accurate).\n- Definition 2: Should say \"for every $\\epsilon, \\delta \\in (0,1)\".\n- Definition 2: Should say \"If there exists $m \\colon (0,1)^2 \\to \\mathbb{N}$\" such that... Otherwise, the order of quantifiers / role of $m$ is unclear.\n- Definition 2: Should say \"returns w.p. $\\geq 1-\\delta$ a function $h \\in \\mathcal{H}$. Also, the first use of w.p. should be explicitly defined (\"with probability (w.p.)\") so that the paper is accessible to a broad audience.\n- Theorem 3: It would be nice to write \"then \\mathcal{H} = \\mathcal{H}_1 \\times \\dots \\times \\mathcal{H}_T\". Ideally, theorem statmenets should be as self-contained as possible as they are often used for quick reference.\n- Definition 4: I would say that \"$h$ computes $f$ w.r.t \\mathcal{D}\". That is, explicitly include $\\mathcal{D}$ in the definition, since it is crucial. Likewise for the definition of _approximates_.\n- Theorem 5: If the previous comment is accepted, this statement should be updated.\n- Theorem 6: If my understanding of the proof is correct (see Summary of this review, above), it is worth adding a sentence explaining what the dataset is to the body of the paper---rather than just saying the main fact the proof relies on, explain how it is used.\n- Theorem 7 and Corollary 8: use the notation $\\mathcal{H}^{\\mathrm{lin}}$ here when referring to linear AR functions/models. Otherwise, the notation surprisingly re-appears on the next page.\n- Section 2.3.1: This is a matter of taste, but I would find it more informative to use $\\mathcal{P}_n$ to denote the class of parities on $n$ bits. This is because $\\mathcal{F}$ was previously used to refer to a generic class of functions.\n- Section 2.3.1, second paragraph after Theorem 10: when defining $\\mathcal{F}_{n,k}$, it should be $A \\in \\binom{[n]}{\\leq k}$. Note the square brackets around $n$.\n- Section 2.3.1, above Theorem 11: It would be much better to avoid using $\\approx$ notation in favor of notation with more well-defined meaning, such as $O(\\cdot)$.\n- Section 2.3.1, above Theorem 11: Should be \"and a sample complexity of $\\approx k \\log n$\" (not \"the sample complexity\").", "questions": "Many of the suggestions/qualms I listen in Weaknesses above can be rephrased as a question. I am open to discussing any of these, and would consider updating my score based on these being addressed in a revision. My only remaining question is whether the authors intend to release the code used for their experiments so that the reader may reproduce them.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a theoretical framework for analyzing autoregressive learners. At its core is a formal definition of autoregressive learnability (AR Learnability), analogous to PAC Learnability.\n\nAR Learnability is defined with respect to distributions over strings over a finite set of tokens $\\mathbb{D}$. More specifically, these are distributions $\\mathcal{D}$ over contexts $\\mathcal{X} = \\mathbb{D}^n$ and continuations $\\mathcal{Z} = \\mathbb{D}^T$ for some integers $n$ and $T$.\nBriefly stated, a hypothesis class $\\mathcal{H}$ is AR Learnable if there is an algorithm that, for any realizable distribution $\\mathcal{D}$, takes as input samples from $\\mathcal{D}$ and outputs (w.h.p.) a hypothesis $h \\in \\mathcal{H}$ such that $h(x,z_{\\<t}) = z_t$ for most $t \\leq T$. That is, a hypothesis that can (mostly) correctly generate the continuation of the context $x$ as in the distribution $\\mathcal{D}$.\n\nThis definition is rigorously analyzed. The first batch of theoretical findings (theorems) are concerned with the universality of AR learners:\n1. The product of $T$ PAC-learnable hypothesis classes AR-learable [Theorem 3].\n2. A generalization of AR-learability to the setting of computing/approximating functions; AR-learnable classes are approximable [Theorem 5]\n3. Linear AR models ($\\mathcal{H}^{\\mathrm{lin}}$) can approximate any $f \\in \\mathrm{TIME}(T(n))$ w.r.t any distribution $\\mathcal{D}$ given access to a dataset of $\\mathrm{poly}(\\mathrm{T(n)})$-length strings [Corollary 8].\n\nThe last item above is particularly significant because, in the supervised (i.e., \"standard\") setting, the analogous class to $\\mathcal{H}^{\\mathrm{lin}}$ are known to be so weak that they cannot learn the XOR function. To my understanding, Corollary 8 is proven by representing the function $f$ as a linear threshold circuit $C_f$; the dataset is populated by sampling an $x \\gets \\mathcal{D}$ and adding a row that represents the outputs of intermediate gates of $C_f(x)$.\n\nConsidering the above proof, it is perhaps not surprising that AR learners are universal learners; the \"heavy-lifting\" is done by the dataset, which verbosely represents the computed function. Thus, the authors define and analyze a new complexity measure for AR learners, called the _length complexity_: For a given hypothesis class $\\mathcal{H}$ and a target function class $\\mathcal{F}$, the length complexity is the [minimal? see Weaknesses] number of autoregressive steps $T$ such that each $f \\in \\mathcal{F}$ can be computed  by some $h \\in \\mathcal{H}$ within $T$ autoregressive steps. Intuitively, length complexity should capture the length of a chain-of-thought used when training the model.\n\nFinally, the authors show that parities over $n$ bits can be computed by $\\mathcal{H}^{\\mathrm{lin}}$ with length complexity $O(\\log n)$. Additionally, they prove a theorem depicting a trade-off between the richness of the hypothesis class and the length complexity, by showing that parities over $n$ bits can be learned by $k$-order parities over $\\geq n$ bits, with length complexity $\\Theta(n / k)$.\n\nThe authors supplement the theoretical framework and findings with experimental evaluation. First, they show that an adaptation of $\\mathcal{H}^{\\mathrm{lin}}$ with 162M parameters exhibits non-trivial learning on TinyStories; it is often able to generate coherent text. Second, they show that a 4-layer MLP with 775M parameters can learn to multiply 4 digit numbers as well as Goat 7B, provided sufficiently long chain-of-thought (i.e., length complexity) and a custom tokenization scheme.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- This paper provides a theoretical foundation for an increasingly important topic, namely, understanding the emergent abilities of autoregressive learners.\n- Several theoretical works have already tackled the questions of autoregressive learning. However, to my knowledge, this is the first paper to propose a generic defintion analogous to PAC-learning. This is significant because it may inspire learning theorists to search for analogous results to what is known in the rich literature of PAC. For example, as the authors suggest, it would be interesting to find an intrinsic dimension of hypothesis classes that governs AR-learnability (cf. VC-dimension).\n- The theoretical framework presented in this paper is an appropriate operationalization of autoregressive learning: the definitions stay close to autoregressive learning as it is implemented in practice (at least conceptually).\n- I find the framework itself appealing. It is not easy to come up with a clean definition based on an existing phenomenon, as they real world is often \"messy\". The definitions strike a good balance between intuitiveness/simplicitly, to being non-trivial (\"mysterious\") enough to invite further analysis.\n- Overall, the paper is clearly written. Reading it was an enjoyable experience!\n- I particularly appreciate how well-organized this paper is. It first presents a fairly intuitive definition, and then provides a sequence of theorems and refinements of this definition to interesting settings. That is to say, the paper \"tells a convincing story\" about autoregressive learning.\n- This \"story\" that the theory suggests is then supported by experimental results fairly adequetly. While the experimental setup is somewhat simplistic as compared to more empirical works, I do not think that a grander setup is needed for this type of paper.\n- While I did not verify the proofs of all theorems, I have read some of them (especially Theorem 7) and they seem correct. I would also like to emphasize that the simplicity of the proofs should not be viewed as a weakness of this paper. Coming up with the right definition (from which intuitive theorems are easy to prove) is a challenging part of many theoretical works; in this case, the authors did well in carefully designing their framework.", "weaknesses": "Listed in decreasing order of significance.\n\n## TinyStories experiment is anecdotal, compares to wrong model?\nI am not sure what are the actual results being reported with the TinyStories experiment: What I found is a footnote on a 1.2 difference in perplexity between the linear predictor and GPT-2 Small---but I'm not sure what to make of this quantity. And there is a statement that the linear predictor \"often does produce coherent text\". But how often? And how do you measure coherence? While I lack the expertise to suggest a concrete experiment, I expect a higher level of rigour for a paper at ICLR. Perhaps the authors could refer to the TinyStories paper and reproduce some of the experiments there with their model.\n\nRelatedly, it seems that in the TinyStories paper there is a 28M-parameter model that achieves performance comparable to GPT2-XL. Therefore, the comparison of the linear predictor to GPT2-Small seems inappropriate---much better performance can be attained with 5x less parameters. Perhaps the authors should compare the linear predictor to [that model](https://huggingface.co/papers/2305.07759) instead. Discovering that a linear predictor requires significantly more parameters than a (good) transformer model to achieve comparable performance would undermine the main message of the paper (that the power of LLMs can be attributed to AR rather than architecture). Please correct me if you disagree that comparing to the TinyStories transformer is a more fair comparison towards this end.\n\n## Linear Decoders should be defined more slowly\n- Given the significance of linear ARs (linear decoders) throughout the paper, I suggest defining them more slowly and in their own Definition environment---rather than presenting them as an \"Example\". A Definition environment is more easy to refer back to, and encourages more formal writing.\n- A figure depicting the construction would be welcome: I had to work it out with pen and paper on the margin.\n- \"Under some margin conditions and using a convex surrogate loss function, this class is in fact learnable using SGD.\" This sentence must be supported by either a citation or a proof. As it is currently phrased it is too vague and not well-enough argued for to be used as a true statement (the current level of rigour is more appropriate for a tangential side-note).\n- Observe that this class is learnable in polynomial time: Say PAC-learnable, becuase this paper uses multiple notions of learnability.\n\n## Length complexity is not well-defined\nAs length complexity is currently defined in Definition 9, a class $\\mathcal{H}$ has infinitely many length complexities for computing a given $\\mathcal{F}$. This is because if $T$ is a length complexity of $\\mathcal{H}$ for computing $\\mathcal{F}$, then so is any $T' \\geq T$. Instead, you should define the length complexity of $\\mathcal{H}$ computing $\\mathcal{F}$ to be _the minimal $T$ for which the conditions stated in the definition hold.\n\nOn that note, it would be illuminating (and add to the cohesion of the paper) if the authors spell-out the length complexity of the construction from Theorem 7.\n\n## Missing citations\n- Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective by Feng et al. (2023): This work gives a complexity-theoretic explanation to the success of chain-of-thought, especially in mathematical reasoning. The techniques seem entirely different, and it is focused on transformer architectures (rather than focusing on the autoregressive aspect of LLMs, as in the paper currently under review). Still, it clearly fits in three of the four topics covered by the related work (all save for \"Beyond Transfomers\").\n- Fast Learning Requires Good Memory: A Time-Space Lower Bound for Parity Learning by Raz (2016): This prominent result (FOCS best paper, JACM) should probably be mentioned in the first paragraph of Section 2.3.1 that discusses hardness results for parities.\n\n## Other minor writing comments\n- Page 4 above Definition 2: \"...for all **sub-sequences** $z_{<t}$\" should be **prefixes** (sub-sequences is correct but less accurate).\n- Definition 2: Should say \"for every $\\epsilon, \\delta \\in (0,1)\".\n- Definition 2: Should say \"If there exists $m \\colon (0,1)^2 \\to \\mathbb{N}$\" such that... Otherwise, the order of quantifiers / role of $m$ is unclear.\n- Definition 2: Should say \"returns w.p. $\\geq 1-\\delta$ a function $h \\in \\mathcal{H}$. Also, the first use of w.p. should be explicitly defined (\"with probability (w.p.)\") so that the paper is accessible to a broad audience.\n- Theorem 3: It would be nice to write \"then \\mathcal{H} = \\mathcal{H}_1 \\times \\dots \\times \\mathcal{H}_T\". Ideally, theorem statmenets should be as self-contained as possible as they are often used for quick reference.\n- Definition 4: I would say that \"$h$ computes $f$ w.r.t \\mathcal{D}\". That is, explicitly include $\\mathcal{D}$ in the definition, since it is crucial. Likewise for the definition of _approximates_.\n- Theorem 5: If the previous comment is accepted, this statement should be updated.\n- Theorem 6: If my understanding of the proof is correct (see Summary of this review, above), it is worth adding a sentence explaining what the dataset is to the body of the paper---rather than just saying the main fact the proof relies on, explain how it is used.\n- Theorem 7 and Corollary 8: use the notation $\\mathcal{H}^{\\mathrm{lin}}$ here when referring to linear AR functions/models. Otherwise, the notation surprisingly re-appears on the next page.\n- Section 2.3.1: This is a matter of taste, but I would find it more informative to use $\\mathcal{P}_n$ to denote the class of parities on $n$ bits. This is because $\\mathcal{F}$ was previously used to refer to a generic class of functions.\n- Section 2.3.1, second paragraph after Theorem 10: when defining $\\mathcal{F}_{n,k}$, it should be $A \\in \\binom{[n]}{\\leq k}$. Note the square brackets around $n$.\n- Section 2.3.1, above Theorem 11: It would be much better to avoid using $\\approx$ notation in favor of notation with more well-defined meaning, such as $O(\\cdot)$.\n- Section 2.3.1, above Theorem 11: Should be \"and a sample complexity of $\\approx k \\log n$\" (not \"the sample complexity\").", "questions": "Many of the suggestions/qualms I listen in Weaknesses above can be rephrased as a question. I am open to discussing any of these, and would consider updating my score based on these being addressed in a revision. My only remaining question is whether the authors intend to release the code used for their experiments so that the reader may reproduce them.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698173691798}], "openreview_url": "https://openreview.net/forum?id=qYb0CANLGC", "arxiv_id": "2309.06979", "paper_pdf": "papers/qYb0CANLGC.pdf", "paper_pdf_sha256": "ebe78d35f4d70fbb451c6e62b818867e5bc4cc9c62b6562566f0c26d4f1b2d26", "paper_pdf_bytes": 363218, "paper_pdf_source": "openreview", "code_url": "https://github.com/emalach/LinearLM", "code_repository": "emalach/LinearLM", "code_commit": "11fb93e06b9738a91b3463b7b0b7b63095928827", "code_archive": "repos/qYb0CANLGC.zip", "code_archive_sha256": "d5957a399d05f42c6edeb12417cb7a03d2c6949a12da9f08796bc5e978751112", "code_archive_bytes": 7725, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 11, "github_languages": {"Python": 14953}, "github_archived": false, "github_pushed_at": "2024-07-29T20:36:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/auto-regressive-next-token-predictors-are"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LgjKqSjDzr", "year": 2022, "status": "rejected", "title": "SALT : Sharing Attention between Linear layer and Transformer for tabular dataset", "authors": ["Juseong Kim", "Jinsun Park", "Giltae Song"], "authorids": ["~Juseong_Kim2", "~Jinsun_Park1", "gsong@pusan.ac.kr"], "authors_source": "OpenReview API", "abstract": "Handling tabular data with deep learning models is a challenging problem despite their remarkable success in vision and language processing applications. Therefore, many practitioners still rely on classical models such as gradient boosting decision trees (GBDTs) rather than deep networks due to their superior performance with tabular data. In this paper, we propose a novel hybrid deep network architecture for tabular data, dubbed SALT (Sharing Attention between Linear layer and Transformer). The proposed SALT consists of two blocks: Transformers and linear layers blocks that take advantage of shared attention matrices. The shared attention matrices enable transformers and linear layers to closely cooperate with each other, and it leads to improved performance and robustness. Our algorithm outperforms tree-based ensemble models and previous deep learning methods in multiple benchmark datasets. We further demonstrate the robustness of the proposed SALT with semi-supervised learning and pre-training with small dataset scenarios.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "e0CrrbNMzTH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2498/Reviewer_eW8f"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a new tabular deep learning architecture based on sharing attention matrices enable transformers and linear layers. Comparisons with other tabular learning models on various benchmarks are demonstrated. ", "review_text": "I have serious concerns about the paper, as listed below: \n\n- The paper writing is very poor. There are numerous language and grammar errors. At many places, the expressions are not clear. The quality of the figures is low. \n\n- Overall, the novelty is low. Transformer architecture is modified in a straightforward way. Even the semi-supervised learning approach is the simple adaption of MLM-like unsupervised pre-training. \n\n- Benchmarking is employed and presented very poorly. For comparison models like TabNet, XGBoost etc., it is not clear how the hyperparameter tuning is done and what parameters are included in the search space. The authors should clearly describe which parameters were tuned and what validation reward is used. Otherwise, the outperformance conclusions are not convincing at all.\n\n- How do you do hyperparameter tuning in semi-supervised regime? It is unclear how semi-supervised validation data is used.\n\n- The results are presented in AUROC but what is the training objective? If the training objective is not AUROC, how do you ensure the metric mismatch is not dominating? \n\n- What is the significance of Fig. 5? How can you convince the readers that SALT has learned attention patterns that are meaningful?\n\n- The difference between supervised learning results is very low across different models. It is unclear whether the results are statistically significant. \n\n- No ablation studies are presented for the major constituents of the claims, such as the benefit of sharing attention layers. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new tabular deep learning architecture based on sharing attention matrices enable transformers and linear layers. Comparisons with other tabular learning models on various benchmarks are demonstrated. ", "main_review": "I have serious concerns about the paper, as listed below: \n\n- The paper writing is very poor. There are numerous language and grammar errors. At many places, the expressions are not clear. The quality of the figures is low. \n\n- Overall, the novelty is low. Transformer architecture is modified in a straightforward way. Even the semi-supervised learning approach is the simple adaption of MLM-like unsupervised pre-training. \n\n- Benchmarking is employed and presented very poorly. For comparison models like TabNet, XGBoost etc., it is not clear how the hyperparameter tuning is done and what parameters are included in the search space. The authors should clearly describe which parameters were tuned and what validation reward is used. Otherwise, the outperformance conclusions are not convincing at all.\n\n- How do you do hyperparameter tuning in semi-supervised regime? It is unclear how semi-supervised validation data is used.\n\n- The results are presented in AUROC but what is the training objective? If the training objective is not AUROC, how do you ensure the metric mismatch is not dominating? \n\n- What is the significance of Fig. 5? How can you convince the readers that SALT has learned attention patterns that are meaningful?\n\n- The difference between supervised learning results is very low across different models. It is unclear whether the results are statistically significant. \n\n- No ablation studies are presented for the major constituents of the claims, such as the benefit of sharing attention layers. ", "summary_of_the_review": "I suggest substantially revising the paper and submitting to another venue. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636419875371}, {"id": "jJP87vyPFd7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2498/Reviewer_mutc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a hybrid deep network, named SALT (Sharing Attention between Linear layer and Transformer). The SALT consists of two blocks: Transformer and linear layer blocks that take advantage of shared attention matrices. They compare SALT with tree-based ensemble models and previous deep learning models on multiple benchmark datasets. It furher shows robustness of the proposed SALT with semi-supervised learning and pre-training with small dataset scenarios. \n\nThe main body of SALT has two blocks: Transformers block inspired by (Vaswani et al., 2017) and Linear layer block inspired by (Liu et al., 2021). Each block has subblocks by feature-wise and dimension- wise. These two subblocks allow communications between different features and different embed- ding elements and makes the model robust (Tolstikhin et al., 2021). The output values from two blocks become the contextual embedding values. SALT performs fin-tuning and pre-training with contextual embedding values.\n\nFor supervised learning setting, the mean AUROC score of the proposed SALT improves upon the treebased LightGBM by mere 0.09%, and it improves upon the best deep learning based model SAINT by 0.29%", "review_text": "The main strength of the paper is marginal improvement over the SOTA deep learning model SAINT for supervised setting scenario. \n\nThe main weakness is in lack of justification for the proposed choice of SALT model, and inadeqate explanation of the SALT architecture. It is not clear what is dmension-wise attention. There is no motivation/justification for the Transformers and Linear layers with attention sharing. Further, why is that there is only one layer of feature-wise attention and linear layer, and then L layers of dimension-wise attention and linear layer. Neiher has been given any good justification for sharing attention between Transformer and Linear layer. Further, for the most important supervised learning scenario, the SALT model barely beats the tree-based LightGBM model.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a hybrid deep network, named SALT (Sharing Attention between Linear layer and Transformer). The SALT consists of two blocks: Transformer and linear layer blocks that take advantage of shared attention matrices. They compare SALT with tree-based ensemble models and previous deep learning models on multiple benchmark datasets. It furher shows robustness of the proposed SALT with semi-supervised learning and pre-training with small dataset scenarios. \n\nThe main body of SALT has two blocks: Transformers block inspired by (Vaswani et al., 2017) and Linear layer block inspired by (Liu et al., 2021). Each block has subblocks by feature-wise and dimension- wise. These two subblocks allow communications between different features and different embed- ding elements and makes the model robust (Tolstikhin et al., 2021). The output values from two blocks become the contextual embedding values. SALT performs fin-tuning and pre-training with contextual embedding values.\n\nFor supervised learning setting, the mean AUROC score of the proposed SALT improves upon the treebased LightGBM by mere 0.09%, and it improves upon the best deep learning based model SAINT by 0.29%", "main_review": "The main strength of the paper is marginal improvement over the SOTA deep learning model SAINT for supervised setting scenario. \n\nThe main weakness is in lack of justification for the proposed choice of SALT model, and inadeqate explanation of the SALT architecture. It is not clear what is dmension-wise attention. There is no motivation/justification for the Transformers and Linear layers with attention sharing. Further, why is that there is only one layer of feature-wise attention and linear layer, and then L layers of dimension-wise attention and linear layer. Neiher has been given any good justification for sharing attention between Transformer and Linear layer. Further, for the most important supervised learning scenario, the SALT model barely beats the tree-based LightGBM model.", "summary_of_the_review": "Though the numerical exeriments show that the proposed model SALT marginally improves over the sota deep learning model SAINT, it barely beats the tree-based LightGBM model. The architectural choices of SALT have no motivation/justification. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636400723648}, {"id": "ODX7CeUiutA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2498/Reviewer_6yF7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a hybrid deep network architecture for tabular data, dubbed SALT (Sharing Attention between Linear layer and Transformer). \nThere are two blocks in SALT: Transformers and linear layer blocks. And sharing attention matrices are introduced to promote cooperation between these two blocks.", "review_text": "Pros:\n\n- This manuscript attempts to propose a deep learning model for tabular data, a problem with practical application value.\n- An interesting hybrid network architecture is presented, the main idea of which is good and reasonable.\n\nCons:\n\n- The novelty is limit. The key block SHARING ATTENTION shares similar idea with existing works, RealFormer.\n- The writing of the manuscript can be polished. Typos in Equation 7 and 8 are confusing.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a hybrid deep network architecture for tabular data, dubbed SALT (Sharing Attention between Linear layer and Transformer). \nThere are two blocks in SALT: Transformers and linear layer blocks. And sharing attention matrices are introduced to promote cooperation between these two blocks.", "main_review": "Pros:\n\n- This manuscript attempts to propose a deep learning model for tabular data, a problem with practical application value.\n- An interesting hybrid network architecture is presented, the main idea of which is good and reasonable.\n\nCons:\n\n- The novelty is limit. The key block SHARING ATTENTION shares similar idea with existing works, RealFormer.\n- The writing of the manuscript can be polished. Typos in Equation 7 and 8 are confusing.", "summary_of_the_review": "The proposed SALT, in my view, should be technically correct. \nBut I think the novelty is limit. \nConsequently, although both the problem and the method is okay, this manuscript is a borderline work.\nI recommend rejecting it.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635959384357}, {"id": "VSrZ9Qk09jc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2498/Reviewer_A8EH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a Transformer-based architecture for tabular datasets. This architecture combines a Transformer with a gating MLP (gMLP) by sharing the attention matrices. The authors also proposed a new approach to encode continuous variables. The proposed model is evaluated on six binary classification tabular datasets. The proposed model is evaluated in a supervised learning context, but also in a semi-supervised context.", "review_text": "**Strengths**\n\n- I think it is interesting to develop deep models for tabular data. Learning on tabular datasets is still an under-explored research area.\n- The idea of sharing attention matrices looks interesting. \n- The authors show that combining a Transformer-like architecture with a gMLP-like one improves the performances. (I assume the SALT results are significantly better than SALT-former and SALT-linear, see my comment later).\n- The authors analyzed different variants of their model.\n\n**Weaknesses**\n\n- Section 3.5: “SALT improves this method to use the same embedding matrix between categorical variables and continuous variables.” I am not sure what is the intuition behind this idea. The authors should motivate this claim.  \n- The model is presented to work on any tabular datasets. However from experience, a lot of tabular datasets contain at least one date/time column. I wonder if SALT can work on date/time columns. \n- The model is evaluated only for binary classification tasks. I think adding experiments for other tasks (e.g. multi-class classification, multi-label classification, regression) can increase the quality of the paper. \n- The authors should explain how they do the semi-supervised learning in section 5.3. \n- For most of the tables, I recommend showing the variance for different random seeds or some other confidence interval measures. For a lot of results, the difference between two models looks small and it is difficult to know if the difference is significant. \n- It is difficult to understand Table 3 without additional information. I recommend adding information about the dataset or metric used in the caption. This Table is not mentioned in the text.\n- The authors visualize the attention matrices for different variants of their model. However, these visualizations are not commented on so it is difficult to understand the message from this analysis. \n- The authors should comment on the scalability of their model. The SAINT model was evaluated on a larger dataset. I wonder why SALT was not evaluated on this dataset. I think the authors should also compare the training time/inference time of their model with existing models. \n- The authors should add information about the Transformer architecture used. Is it the original Transformer architecture also called PostNorm? Or is it more recent architecture like PreNorm or ReZero?\n- The last sentence of the conclusion breaks the anonymity of the paper: “The code is available at https://github.com/Juseong03/SALT” \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a Transformer-based architecture for tabular datasets. This architecture combines a Transformer with a gating MLP (gMLP) by sharing the attention matrices. The authors also proposed a new approach to encode continuous variables. The proposed model is evaluated on six binary classification tabular datasets. The proposed model is evaluated in a supervised learning context, but also in a semi-supervised context.", "main_review": "**Strengths**\n\n- I think it is interesting to develop deep models for tabular data. Learning on tabular datasets is still an under-explored research area.\n- The idea of sharing attention matrices looks interesting. \n- The authors show that combining a Transformer-like architecture with a gMLP-like one improves the performances. (I assume the SALT results are significantly better than SALT-former and SALT-linear, see my comment later).\n- The authors analyzed different variants of their model.\n\n**Weaknesses**\n\n- Section 3.5: “SALT improves this method to use the same embedding matrix between categorical variables and continuous variables.” I am not sure what is the intuition behind this idea. The authors should motivate this claim.  \n- The model is presented to work on any tabular datasets. However from experience, a lot of tabular datasets contain at least one date/time column. I wonder if SALT can work on date/time columns. \n- The model is evaluated only for binary classification tasks. I think adding experiments for other tasks (e.g. multi-class classification, multi-label classification, regression) can increase the quality of the paper. \n- The authors should explain how they do the semi-supervised learning in section 5.3. \n- For most of the tables, I recommend showing the variance for different random seeds or some other confidence interval measures. For a lot of results, the difference between two models looks small and it is difficult to know if the difference is significant. \n- It is difficult to understand Table 3 without additional information. I recommend adding information about the dataset or metric used in the caption. This Table is not mentioned in the text.\n- The authors visualize the attention matrices for different variants of their model. However, these visualizations are not commented on so it is difficult to understand the message from this analysis. \n- The authors should comment on the scalability of their model. The SAINT model was evaluated on a larger dataset. I wonder why SALT was not evaluated on this dataset. I think the authors should also compare the training time/inference time of their model with existing models. \n- The authors should add information about the Transformer architecture used. Is it the original Transformer architecture also called PostNorm? Or is it more recent architecture like PreNorm or ReZero?\n- The last sentence of the conclusion breaks the anonymity of the paper: “The code is available at https://github.com/Juseong03/SALT” \n", "summary_of_the_review": "Overall, I feel it is quite difficult to understand the potential impact of this paper. The idea of sharing attention matrices looks interesting but there are some questions about the scalability of the proposed approach and if it can work on any tabular datasets. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635792152480}], "openreview_url": "https://openreview.net/forum?id=LgjKqSjDzr", "arxiv_id": null, "paper_pdf": "papers/LgjKqSjDzr.pdf", "paper_pdf_sha256": "e51405d2a961863fea10f81c16e63b1564adf13a5f116e713f6a9337c8e0c1a2", "paper_pdf_bytes": 999515, "paper_pdf_source": "openreview", "code_url": "https://github.com/Juseong03/SALT", "code_repository": "Juseong03/SALT", "code_commit": "8d95159be073be64d1ba1af1017ae8870a35005a", "code_archive": "repos/LgjKqSjDzr.zip", "code_archive_sha256": "1f5edbaadaeb7ad970dcf27b46a64c5559d3962ce2206a0c50a4ca9b4ecf0b04", "code_archive_bytes": 9616, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 10, "github_languages": {"Python": 35982}, "github_archived": false, "github_pushed_at": "2021-10-28T02:33:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/salt-sharing-attention-between-linear-layer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kmBFHJ5pr0o", "year": 2021, "status": "rejected", "title": "Distributed Adversarial Training to Robustify Deep Neural Networks at Scale", "authors": ["Gaoyuan Zhang", "Songtao Lu", "Sijia Liu", "Xiangyi Chen", "Pin-Yu Chen", "Lee Martie", "Lior Horesh", "Mingyi Hong"], "authorids": ["~Gaoyuan_Zhang1", "~Songtao_Lu1", "~Sijia_Liu1", "~Xiangyi_Chen1", "~Pin-Yu_Chen1", "lee.martie@ibm.com", "~Lior_Horesh1", "~Mingyi_Hong1"], "authors_source": "OpenReview API", "abstract": "Current deep neural networks are vulnerable to adversarial attacks, where adversarial perturbations to the inputs can change or manipulate classification. To defend against such attacks, an effective and popular approach, known as adversarial training, has been shown to mitigate the negative impact of adversarial attacks by virtue of a min-max robust training method. While effective, this approach is difficult to scale well to large models on large datasets (e.g., ImageNet) in general. To address this challenge, we propose distributed adversarial training (DAT), a large-batch adversarial training framework implemented over multiple machines. DAT supports one-shot and iterative attack generation methods, gradient quantization, and training over labeled and unlabeled data. Theoretically, we provide, under standard conditions in the optimization theory, the convergence rate of DAT to the first-order stationary points in general non-convex settings.  Empirically, on ResNet-18 and -50 under CIFAR-10 and ImageNet, we demonstrate that DAT either matches or outperforms state-of-the-art robust accuracies and achieves a graceful training speedup.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "kqHMI0j7Aoi", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2142/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed distributed adversarial training (DAT) for robust models. The method is a combination of PGD-like adversarial training, LARS-like large batch training, and quantizing gradients for communication efficiency in distributed training. The authors show convergence of adversarial training with LARS-like learning rate under layer-wise assumptions, and empirical results of DAT can scale to 6x6=36 GPUs and batch size 6*512 for ImageNet.\n\nPros\n\n+ The paper is well written and easy to follow. \n+ The convergence rate looks reasonable. \n+ It is good to show that adversarial training can be accelerated through distributed settings.\n+ Extensive experiments on CIFAR-10 and ImageNet. \n\nCons\n\n- The general contribution of the paper is a bit incremental. The main message seems to be that LARS can help large-batch adversarial training?\n- The convergence rate contribution seems incremental given the known adversarial training convergence proof and LARS-like convergence proof. \n- As far as I know, scaling up learning rate (maybe with warmup) performs good for large batch training, we only need LARS for a certain regime, the ImageNet setting in table 1 where the batch size is 6*512 seems to fall in the regime where scaling LR works. Did the authors scale up LR for baselines without LARS?\n- The Fast FGSM method (Wong et al. 2020) uses cyclic LR etc. to make convergence fast. Are those tricks applied here? Wong et al. reported CIFAR training in about 10 min, which is faster than the baseline in table 1 (52 sec/epoch * 100 epochs). I am worried the authors may target a less significant problem than we expected.\n- [Xie et al. 2019 Feature Denoising for Improving Adversarial Robustness] and Kannan et al. 2018 use more GPUs than reported in the paper. And Xie et al. achieves better robustness on ImageNet. What is the practical advantage of this paper compared to Xie et al? What are the insights from the authors’ reimplementation of Xie’s method on CIFAR? ( DAT-LSGD in table 1)\n- Why does communication time decrease when more GPU machines (24) are used?\n\n\n\n==================== post rebuttal ==============================\n\nI do not think my concerns are addressed by the discussion. However, I also think this is a well-written paper in general and I will not be upset if it is accepted. \n\nMy main concerns,\n(1) The authors fail to show that the proposed method is non-trivial. I think this concern is raised by multiple reviewers. The authors keep clarifying the technical difficulty (especially the theory) of applying LARS etc, but my main concern is the necessity of these knobs added by authors. After a few rounds of discussion, we reach to a conclusion that adversarial training is different from standard training, but I do not think that could be considered insights from this paper. I would strongly suggest authors consider explaining why LARS is necessary by either theory or intuitive insights, and make it clear what exactly the difference is. \n(2) The authors claim contributions for large scale setting (ImageNet with large number of available GPUs), but the experiments are somewhat worse than previous results. Lacking computation resources is a good excuse, but since the authors claimed they can use larger batch size with smaller number of GPUs, I do not see a technical reason why they cannot use their method to re-run the large scale experiments to directly compare with previous results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clearer message on when to use this method", "review": "This paper proposed distributed adversarial training (DAT) for robust models. The method is a combination of PGD-like adversarial training, LARS-like large batch training, and quantizing gradients for communication efficiency in distributed training. The authors show convergence of adversarial training with LARS-like learning rate under layer-wise assumptions, and empirical results of DAT can scale to 6x6=36 GPUs and batch size 6*512 for ImageNet.\n\nPros\n\n+ The paper is well written and easy to follow. \n+ The convergence rate looks reasonable. \n+ It is good to show that adversarial training can be accelerated through distributed settings.\n+ Extensive experiments on CIFAR-10 and ImageNet. \n\nCons\n\n- The general contribution of the paper is a bit incremental. The main message seems to be that LARS can help large-batch adversarial training?\n- The convergence rate contribution seems incremental given the known adversarial training convergence proof and LARS-like convergence proof. \n- As far as I know, scaling up learning rate (maybe with warmup) performs good for large batch training, we only need LARS for a certain regime, the ImageNet setting in table 1 where the batch size is 6*512 seems to fall in the regime where scaling LR works. Did the authors scale up LR for baselines without LARS?\n- The Fast FGSM method (Wong et al. 2020) uses cyclic LR etc. to make convergence fast. Are those tricks applied here? Wong et al. reported CIFAR training in about 10 min, which is faster than the baseline in table 1 (52 sec/epoch * 100 epochs). I am worried the authors may target a less significant problem than we expected.\n- [Xie et al. 2019 Feature Denoising for Improving Adversarial Robustness] and Kannan et al. 2018 use more GPUs than reported in the paper. And Xie et al. achieves better robustness on ImageNet. What is the practical advantage of this paper compared to Xie et al? What are the insights from the authors’ reimplementation of Xie’s method on CIFAR? ( DAT-LSGD in table 1)\n- Why does communication time decrease when more GPU machines (24) are used?\n\n\n\n==================== post rebuttal ==============================\n\nI do not think my concerns are addressed by the discussion. However, I also think this is a well-written paper in general and I will not be upset if it is accepted. \n\nMy main concerns,\n(1) The authors fail to show that the proposed method is non-trivial. I think this concern is raised by multiple reviewers. The authors keep clarifying the technical difficulty (especially the theory) of applying LARS etc, but my main concern is the necessity of these knobs added by authors. After a few rounds of discussion, we reach to a conclusion that adversarial training is different from standard training, but I do not think that could be considered insights from this paper. I would strongly suggest authors consider explaining why LARS is necessary by either theory or intuitive insights, and make it clear what exactly the difference is. \n(2) The authors claim contributions for large scale setting (ImageNet with large number of available GPUs), but the experiments are somewhat worse than previous results. Lacking computation resources is a good excuse, but since the authors claimed they can use larger batch size with smaller number of GPUs, I do not see a technical reason why they cannot use their method to re-run the large scale experiments to directly compare with previous results.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603950600181}, {"id": "HbyhbtXxabY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2142/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work introduces a framework, called DAT, to scale out adversarial training to distributed settings. DAT combines three techniques: one-shot fast gradient sign method for more efficient inner maximization, gradient quantization for reduced communication volume, and layer-wise adaptive learning rate optimizer for large-batch training.  Equipped with the proposed techniques, the authors obtain promising speedups to perform adversarial training against ResNet18 and ResNet50 on CIFAR-10 and ImageNet with multi-node and multi-GPU, while achieving comparable (roughly under 2% difference) accuracy.\n \n \nPros:\n- By combing all the techniques, the proposed framework yields promising results for distributed adversarial training.\n- The paper shows that sparse gradients and large-batch training scheme also apply to adversarial training and provide some theoretical analysis to show the convergence rate of the composite of these schemes. \n \nCons:\n- The technical contribution of the work seems to be limited. All techniques used are from existing works. Therefore, the main contribution is on applying these techniques in the context of adversarial training. \n- Several optimizations employed by the framework have a subtle impact on model convergence. Therefore, although the training time for a specific run might be reduced, the overall development time and complexity might increase. \n \nComments: \n\nAdversarial training is slow, and the paper looks into how to accelerate AT through distributed training. The paper correctly identifies several techniques, such as large-batch, gradient compression, and FGSM, to improve the computation efficiency on single GPU as well as reducing the communication volume across GPUs and machines. The evaluation is thorough and convincing. The main concern is on the novelty side. All techniques are from existing work and have been studied heavily: FGSM reduces the amount of computation for inner maximization; gradient quantization reduces the communication volume; large-batch training improves GPU utilization. Therefore, the contribution is mainly on creating a framework that combines all the techniques and demonstrates empiricdally that they help scale out AT.\n \nThe reviewer appreciates the theoretical analysis of the convergence rate of DAT. However, similar convergence proofs have been given in each of the individual techniques in the past, as in [1-3]. Why isn't the convergence of DAT a simple composition of the three, or is there a more specialized aspect in the convergence proof for DAT?\n \nAnother main concern is the actual convergence impact of the composed framework. Each of the proposed techniques would have an impact on model convergence and is known to make training more difficult, e.g., each may lead to model divergence, slow convergence, or convergence to suboptimal local minima. How does DAT help avoid these challenges or does DAT actually impose more challenges because it involves the interaction of all three methods?\n  \nQuestion:\n\nThe evaluation does not include a comparison of Fast-AT under the 6x6 setting. Is it because there is no performance gain when training Fast-AT across multiple servers, or is it because Fast-AT does not support multi-node training? If it is the latter case, then it seems to be better to add Fast-AT (6x6) as a baseline since it does not seem to be too difficult to extend AT with DDP in PyTorch or multi-node training in TensorFlow, and it would show a better gap between the current multi-GPU multi-node training and DAT.  \n \n[1] Wang et. al. \"On the Convergence and Robustness of Adversarial Training \", http://proceedings.mlr.press/v97/wang19i/wang19i.pdf\n\n[2] You et. al., \"Large Batch Optimization for Deep Learning: Training BERT in 76 minutes\", https://arxiv.org/abs/1904.00962\n\n[3] Alistarh et.al., \"The Convergence of Sparsified Gradient Methods\", https://papers.nips.cc/paper/7837-the-convergence-of-sparsified-gradient-methods.pdf", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "This work introduces a framework, called DAT, to scale out adversarial training to distributed settings. DAT combines three techniques: one-shot fast gradient sign method for more efficient inner maximization, gradient quantization for reduced communication volume, and layer-wise adaptive learning rate optimizer for large-batch training.  Equipped with the proposed techniques, the authors obtain promising speedups to perform adversarial training against ResNet18 and ResNet50 on CIFAR-10 and ImageNet with multi-node and multi-GPU, while achieving comparable (roughly under 2% difference) accuracy.\n \n \nPros:\n- By combing all the techniques, the proposed framework yields promising results for distributed adversarial training.\n- The paper shows that sparse gradients and large-batch training scheme also apply to adversarial training and provide some theoretical analysis to show the convergence rate of the composite of these schemes. \n \nCons:\n- The technical contribution of the work seems to be limited. All techniques used are from existing works. Therefore, the main contribution is on applying these techniques in the context of adversarial training. \n- Several optimizations employed by the framework have a subtle impact on model convergence. Therefore, although the training time for a specific run might be reduced, the overall development time and complexity might increase. \n \nComments: \n\nAdversarial training is slow, and the paper looks into how to accelerate AT through distributed training. The paper correctly identifies several techniques, such as large-batch, gradient compression, and FGSM, to improve the computation efficiency on single GPU as well as reducing the communication volume across GPUs and machines. The evaluation is thorough and convincing. The main concern is on the novelty side. All techniques are from existing work and have been studied heavily: FGSM reduces the amount of computation for inner maximization; gradient quantization reduces the communication volume; large-batch training improves GPU utilization. Therefore, the contribution is mainly on creating a framework that combines all the techniques and demonstrates empiricdally that they help scale out AT.\n \nThe reviewer appreciates the theoretical analysis of the convergence rate of DAT. However, similar convergence proofs have been given in each of the individual techniques in the past, as in [1-3]. Why isn't the convergence of DAT a simple composition of the three, or is there a more specialized aspect in the convergence proof for DAT?\n \nAnother main concern is the actual convergence impact of the composed framework. Each of the proposed techniques would have an impact on model convergence and is known to make training more difficult, e.g., each may lead to model divergence, slow convergence, or convergence to suboptimal local minima. How does DAT help avoid these challenges or does DAT actually impose more challenges because it involves the interaction of all three methods?\n  \nQuestion:\n\nThe evaluation does not include a comparison of Fast-AT under the 6x6 setting. Is it because there is no performance gain when training Fast-AT across multiple servers, or is it because Fast-AT does not support multi-node training? If it is the latter case, then it seems to be better to add Fast-AT (6x6) as a baseline since it does not seem to be too difficult to extend AT with DDP in PyTorch or multi-node training in TensorFlow, and it would show a better gap between the current multi-GPU multi-node training and DAT.  \n \n[1] Wang et. al. \"On the Convergence and Robustness of Adversarial Training \", http://proceedings.mlr.press/v97/wang19i/wang19i.pdf\n\n[2] You et. al., \"Large Batch Optimization for Deep Learning: Training BERT in 76 minutes\", https://arxiv.org/abs/1904.00962\n\n[3] Alistarh et.al., \"The Convergence of Sparsified Gradient Methods\", https://papers.nips.cc/paper/7837-the-convergence-of-sparsified-gradient-methods.pdf", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603919883063}, {"id": "HCVa55pxVp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2142/AnonReviewer3"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Adversarial training is a principled approach towards robust neural networks against adversarial attacks, but it is extremely computing intensive. This work tackles the problem by leveraging the general distributed training method, and addressing the problems of direct application by several effective innovations. \n\nStrength:\n\n+ The proposed method is practical.\n\n+ The proposed DAT can deal with both labeled data (supervised learning) and partial unlabeled data (semi-supervised).\n\n+ I like the idea to use gradient quantization/compression.\n\n+ Make theorectical conribution by convergence analysis for DAT with LALR and gradient quantization.\n\nAdditional comments and questions:\n\n1. One conclusion is the paper can speed up by 3 times with 6 times resource. Please comment on what's the typical speedup of distributed training with n times of resources. Do you think you can further speedup?\n\n2. In formula (2), additional regularization term by lamba is used, making it different from the direct generation of (1) into multiple workers. Why we must introduce the regularization term in DAT?\n\n3. Please provide more details about how to measure communication time please? any profiler used?\n\n4. In table 2, additional unlabeled data can improve accuray. Why?\n\n5. Which deep learning framework is used to support gradient quantization?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposes a novel and comprehensive distributed learning method for speeding up adversarial training with multiple computing nodes.", "review": "Adversarial training is a principled approach towards robust neural networks against adversarial attacks, but it is extremely computing intensive. This work tackles the problem by leveraging the general distributed training method, and addressing the problems of direct application by several effective innovations. \n\nStrength:\n\n+ The proposed method is practical.\n\n+ The proposed DAT can deal with both labeled data (supervised learning) and partial unlabeled data (semi-supervised).\n\n+ I like the idea to use gradient quantization/compression.\n\n+ Make theorectical conribution by convergence analysis for DAT with LALR and gradient quantization.\n\nAdditional comments and questions:\n\n1. One conclusion is the paper can speed up by 3 times with 6 times resource. Please comment on what's the typical speedup of distributed training with n times of resources. Do you think you can further speedup?\n\n2. In formula (2), additional regularization term by lamba is used, making it different from the direct generation of (1) into multiple workers. Why we must introduce the regularization term in DAT?\n\n3. Please provide more details about how to measure communication time please? any profiler used?\n\n4. In table 2, additional unlabeled data can improve accuray. Why?\n\n5. Which deep learning framework is used to support gradient quantization?", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603913647209}, {"id": "_sYdb2pOVww", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2142/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Hello authors,\n\nThank you for the submission.  I found it to be well-written and I enjoyed reading it.  The generic description of DAT given in Equation 2 is nice.  However, there are some concerns on my end that I hope you can help me alleviate that are keeping me from giving the paper a positive review.  I have an open mind and can be convinced.\n\n1. Why is the solution of DAT not obvious?\n\nThe authors claim that \"the direct solution [for scaling AT] of distributing the data batch across multiple machines may not work\", but I cannot understand why when Algorithm 1 looks like exactly that.  The adversarial samples are generated on each machine at the beginning of each iteration so as to avoid computing them on the server at the beginning of each iteration.  To me, this is an unsurprising and obvious optimization.  Other than that minor variation, this looks exactly like the standard parameter server approach.\n\n2. Are the experiments really fair here?\n\nThe most clear claim appears to be from Table 1 and is mentioned in the introduction, where using DAT-PGD with 6 nodes gives a 3x speedup over AT while preserving standard test accuracy and almost preserving robust accuracy.  However, compare with when LALR is *not* used: then, the accuracy is significantly degraded.  So a natural question that I would like to understand is what LALR does when used with the original AT algorithm.  In fact, to get a clear apples-to-apples comparison, I would think that it would be best to use AT with a 6*512 batch size, plus LALR.  Basically, what I mean is that I am not sure if the preservation of accuracy is entirely due to LALR.  If it is, this implies that LALR can be used off-the-shelf with AT in a *non*-distributed setting with larger batch size, and that could result in a situation where DAT-PGD + LALR significantly underperforms AT + LALR---which would contradict the result that DAT is able to preserve accuracy, instead giving the result that LALR improves accuracy.\n\nAlternately, I might expect that DAT-PGD without LALR but with a batch size of 512/6 on each of the 6 nodes (giving the same as the original batch size) would give comparable performance.\n\nI hope the gist of my comment is clear here: it seems like the comparison is not really a fair comparison.  I think it muddles improvements from DAT and improvements from LALR, and makes it hard for the reader to understand what is really going on.\n\nSome additional small things:\n\n - In (1) boldface y is used, but in (2), non-boldface y is used.  Perhaps these should be made the same?\n - \"Server\" instead of \"Sever\" in Algorithm 1\n - In Section 4 there seems to be redundancy: \"see the meta-form of DAT in Algorithm 1 and details in Algorithm A1. \" but I believe it is meant that both references should point at the same thing.\n - In Section 4, it is written that \"DAT takes a M times fewer iteration than AT\".  (Also there are some grammar issues in that sentence.)  It may be clearer to say \"M fewer gradient updates on the server\".\n - In Section 4, it is claimed that the computation time of DAT with iterative PGD is K times more than that of DAT with FGSM.  I know that I am being pedantic but this is not correct---to make it correct, the computation time of *the adversarial perturbation generation step* is K times more, not the overall computational time.\n - Use \\citet not \\citep at the end of the sentence starting with \"Indeed, we will show ...\" at the top of page 5.\n - What is the unit for training time in Table 1?  Does that include communication time?\n\nOverall, I am open to discuss my thoughts, but in my current understanding, I am not able to make a strong enough case for DAT to be published at ICLR.\n\n## Post-rebuttal comments\n\nI am impressed at the amount of time the authors have spent trying to clarify the different concerns.  But unfortunately my concerns are not fully addressed, and also a new concern arises: if this much space had to be spent clarifying different confusions of the reviewers, I think the paper could do with a complete overhaul.  I would suggest that the authors take into account the general confusions that arose in this review process and rewrite the paper such that those particular confusions are alleviated.  For instance, I struggled for clarity on what makes this contribution non-trivial, whether the contributions are primarily theoretical or primarily empirical (and how I should understand the balance between the two), and the fairness of the empirical comparison.  Rewriting the paper such that those three concerns---and the others pointed out by the other reviewers---are discussed clearly would be a significant improvement.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "I have some concerns about differences from existing work and the experiments", "review": "Hello authors,\n\nThank you for the submission.  I found it to be well-written and I enjoyed reading it.  The generic description of DAT given in Equation 2 is nice.  However, there are some concerns on my end that I hope you can help me alleviate that are keeping me from giving the paper a positive review.  I have an open mind and can be convinced.\n\n1. Why is the solution of DAT not obvious?\n\nThe authors claim that \"the direct solution [for scaling AT] of distributing the data batch across multiple machines may not work\", but I cannot understand why when Algorithm 1 looks like exactly that.  The adversarial samples are generated on each machine at the beginning of each iteration so as to avoid computing them on the server at the beginning of each iteration.  To me, this is an unsurprising and obvious optimization.  Other than that minor variation, this looks exactly like the standard parameter server approach.\n\n2. Are the experiments really fair here?\n\nThe most clear claim appears to be from Table 1 and is mentioned in the introduction, where using DAT-PGD with 6 nodes gives a 3x speedup over AT while preserving standard test accuracy and almost preserving robust accuracy.  However, compare with when LALR is *not* used: then, the accuracy is significantly degraded.  So a natural question that I would like to understand is what LALR does when used with the original AT algorithm.  In fact, to get a clear apples-to-apples comparison, I would think that it would be best to use AT with a 6*512 batch size, plus LALR.  Basically, what I mean is that I am not sure if the preservation of accuracy is entirely due to LALR.  If it is, this implies that LALR can be used off-the-shelf with AT in a *non*-distributed setting with larger batch size, and that could result in a situation where DAT-PGD + LALR significantly underperforms AT + LALR---which would contradict the result that DAT is able to preserve accuracy, instead giving the result that LALR improves accuracy.\n\nAlternately, I might expect that DAT-PGD without LALR but with a batch size of 512/6 on each of the 6 nodes (giving the same as the original batch size) would give comparable performance.\n\nI hope the gist of my comment is clear here: it seems like the comparison is not really a fair comparison.  I think it muddles improvements from DAT and improvements from LALR, and makes it hard for the reader to understand what is really going on.\n\nSome additional small things:\n\n - In (1) boldface y is used, but in (2), non-boldface y is used.  Perhaps these should be made the same?\n - \"Server\" instead of \"Sever\" in Algorithm 1\n - In Section 4 there seems to be redundancy: \"see the meta-form of DAT in Algorithm 1 and details in Algorithm A1. \" but I believe it is meant that both references should point at the same thing.\n - In Section 4, it is written that \"DAT takes a M times fewer iteration than AT\".  (Also there are some grammar issues in that sentence.)  It may be clearer to say \"M fewer gradient updates on the server\".\n - In Section 4, it is claimed that the computation time of DAT with iterative PGD is K times more than that of DAT with FGSM.  I know that I am being pedantic but this is not correct---to make it correct, the computation time of *the adversarial perturbation generation step* is K times more, not the overall computational time.\n - Use \\citet not \\citep at the end of the sentence starting with \"Indeed, we will show ...\" at the top of page 5.\n - What is the unit for training time in Table 1?  Does that include communication time?\n\nOverall, I am open to discuss my thoughts, but in my current understanding, I am not able to make a strong enough case for DAT to be published at ICLR.\n\n## Post-rebuttal comments\n\nI am impressed at the amount of time the authors have spent trying to clarify the different concerns.  But unfortunately my concerns are not fully addressed, and also a new concern arises: if this much space had to be spent clarifying different confusions of the reviewers, I think the paper could do with a complete overhaul.  I would suggest that the authors take into account the general confusions that arose in this review process and rewrite the paper such that those particular confusions are alleviated.  For instance, I struggled for clarity on what makes this contribution non-trivial, whether the contributions are primarily theoretical or primarily empirical (and how I should understand the balance between the two), and the fairness of the empirical comparison.  Rewriting the paper such that those three concerns---and the others pointed out by the other reviewers---are discussed clearly would be a significant improvement.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603742079131}], "openreview_url": "https://openreview.net/forum?id=kmBFHJ5pr0o", "arxiv_id": "2206.06257", "paper_pdf": "papers/kmBFHJ5pr0o.pdf", "paper_pdf_sha256": "479976bacaec51c73c97c58763ee6dad612254263f15ecbb62e51c0093883df0", "paper_pdf_bytes": 647562, "paper_pdf_source": "openreview", "code_url": "https://github.com/dat-2022/dat", "code_repository": "dat-2022/dat", "code_commit": "3fac9f60fcf0213b14cc991b3c1a9b2cf26415ba", "code_archive": "repos/kmBFHJ5pr0o.zip", "code_archive_sha256": "ce7430d997a23c756e61aae1bb1a46b4ffbe2e2e9856edc3deb76f58df45d848", "code_archive_bytes": 15125, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 12, "github_languages": {"Python": 47403}, "github_archived": false, "github_pushed_at": "2022-07-31T16:33:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distributed-adversarial-training-to-robustify-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HylNWkHtvB", "year": 2020, "status": "rejected", "title": "Domain-Independent Dominance of Adaptive Methods", "authors": ["Pedro Savarese", "David McAllester", "Sudarshan Babu", "Michael Maire"], "authorids": ["savarese@ttic.edu", "mcallester@ttic.edu", "sudarshan@ttic.edu", "mmaire@uchicago.edu"], "authors_source": "OpenReview API", "abstract": "From a simplified analysis of adaptive methods, we derive AvaGrad, a new optimizer which outperforms SGD on vision tasks when its adaptability is properly tuned. We observe that the power of our method is partially explained by a decoupling of learning rate and adaptability, greatly simplifying hyperparameter search. In light of this observation, we demonstrate that, against conventional wisdom, Adam can also outperform SGD on vision tasks, as long as the coupling between its learning rate and adaptability is taken into account. In practice, AvaGrad matches the best results, as measured by generalization accuracy, delivered by any existing optimizer (SGD or adaptive) across image classification (CIFAR, ImageNet) and character-level language modelling (Penn Treebank) tasks. This later observation, alongside of AvaGrad's decoupling of hyperparameters, could make it the preferred optimizer for deep learning, replacing both SGD and Adam.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rklCY7VJ9r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1540/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors present a new adaptive gradient method AvaGrad. The authors claim the proposed method is less sensitive to its hyperparameters, compared to previous algorithms, and this is due to decoupling the learning rate and the damping parameter.\n\nOverall, the paper is well written, and is on an important topic. However, I have a few concerns about the paper, which I will list below.\n\n1. The fact that adaptive gradient methods converge with a fast rate when the sum in the denominator is taken till the t-1th iterate has appeared in previous papers before [1]. The convergence rate analysis for this case is fairly simple, and I am not sure if analyzing RMSProp/Adam in this setting should be considered a significant contributions of the paper.\n\n2. The proposed algorithm AvaGrad is a simple but interesting idea. I have a number questions about the experimental evaluation though, which makes it hard for me to evaluate the significance of the results presented:\n\na) Was momentum used with SGD?\n\nb) How is the optimal hyperparameters (learning rate and damping, i..e, epsilon, parameters) selected?\n\nc) Do any of these conclusions change when trying out a very small or very large batch size?\n\nd) I am not convinced that using the same optimal hyperparams as the WRN-28-4 task on the WRN-28-10 and ResNet50 models is a reasonable experiment. Why is this a good idea? While this does support the claim that adaptive gradient methods are less sensitive to hyperparameter settings, but makes the other claim about AvaGrad generalizing just as well as SGD weaker?\n\ne) One of the key claims that adaptive gradient methods generalize better when using a large damping (epsilon) parameter has appeared in previous papers as well [2, 3].\n\n\nOverall, in my view, this is a borderline paper mostly because I think a number of the results presented have been shown in other recent papers. My score reflects this. However, I think decoupling the learning rate and the damping parameter by normalizing the preconditioner is a simple but interesting idea, and I am willing to increase my score based on the discussion with the authors and other reviewers.\n\n\n[1] X. Li and F. Orabona. On the Convergence of Stochastic Gradient Descent with Adaptive Stepsizes. In AISTATS 2019\n[2] M. Zaheer, S. Reddi, D. Sachan, S. Kale, and S. Kumar. Adaptive methods for nonconvex optimization. in NeurIPS 2018.\n[3] S. De, A. Mukherjee, and E. Ullah. Convergence guarantees for rmsprop and adam in non-convex optimization and an empirical comparison to nesterov acceleration. arXiv:1807.06766, 2018.\n\nA few more minor comments:\n\n1. The authors say that methods like AMSGrad fail to match the convergence rate of SGD. But this statement seems misleading since it is not clear whether the worse rate is due to the analysis (which gives an upper bound) or the algorithm?\n\n2. In the related work section, the authors discuss convergence rates of algorithms with constant and decreasing step sizes together. This can be confusing to the reader, and it is best to explicitly mention the setting under which the different results were derived.\n\n=======================================\n\nEdit after rebuttal:\nI thank the authors for the detailed response and for the updated version of the paper. After discussion with other reviewers, we are still not convinced that the hyperparameter tuning in the experiments (especially the baselines) is rigorous enough. This is especially important given that this paper proposes a new optimizer. We are also concerned about the novelty of the results, and believe most of these results have appeared in previous work. So I am not increasing my score. I would encourage the authors to do a more rigorous experimental evaluation of the proposed algorithm.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #4", "review": "In this paper, the authors present a new adaptive gradient method AvaGrad. The authors claim the proposed method is less sensitive to its hyperparameters, compared to previous algorithms, and this is due to decoupling the learning rate and the damping parameter.\n\nOverall, the paper is well written, and is on an important topic. However, I have a few concerns about the paper, which I will list below.\n\n1. The fact that adaptive gradient methods converge with a fast rate when the sum in the denominator is taken till the t-1th iterate has appeared in previous papers before [1]. The convergence rate analysis for this case is fairly simple, and I am not sure if analyzing RMSProp/Adam in this setting should be considered a significant contributions of the paper.\n\n2. The proposed algorithm AvaGrad is a simple but interesting idea. I have a number questions about the experimental evaluation though, which makes it hard for me to evaluate the significance of the results presented:\n\na) Was momentum used with SGD?\n\nb) How is the optimal hyperparameters (learning rate and damping, i..e, epsilon, parameters) selected?\n\nc) Do any of these conclusions change when trying out a very small or very large batch size?\n\nd) I am not convinced that using the same optimal hyperparams as the WRN-28-4 task on the WRN-28-10 and ResNet50 models is a reasonable experiment. Why is this a good idea? While this does support the claim that adaptive gradient methods are less sensitive to hyperparameter settings, but makes the other claim about AvaGrad generalizing just as well as SGD weaker?\n\ne) One of the key claims that adaptive gradient methods generalize better when using a large damping (epsilon) parameter has appeared in previous papers as well [2, 3].\n\n\nOverall, in my view, this is a borderline paper mostly because I think a number of the results presented have been shown in other recent papers. My score reflects this. However, I think decoupling the learning rate and the damping parameter by normalizing the preconditioner is a simple but interesting idea, and I am willing to increase my score based on the discussion with the authors and other reviewers.\n\n\n[1] X. Li and F. Orabona. On the Convergence of Stochastic Gradient Descent with Adaptive Stepsizes. In AISTATS 2019\n[2] M. Zaheer, S. Reddi, D. Sachan, S. Kale, and S. Kumar. Adaptive methods for nonconvex optimization. in NeurIPS 2018.\n[3] S. De, A. Mukherjee, and E. Ullah. Convergence guarantees for rmsprop and adam in non-convex optimization and an empirical comparison to nesterov acceleration. arXiv:1807.06766, 2018.\n\nA few more minor comments:\n\n1. The authors say that methods like AMSGrad fail to match the convergence rate of SGD. But this statement seems misleading since it is not clear whether the worse rate is due to the analysis (which gives an upper bound) or the algorithm?\n\n2. In the related work section, the authors discuss convergence rates of algorithms with constant and decreasing step sizes together. This can be confusing to the reader, and it is best to explicitly mention the setting under which the different results were derived.\n\n=======================================\n\nEdit after rebuttal:\nI thank the authors for the detailed response and for the updated version of the paper. After discussion with other reviewers, we are still not convinced that the hyperparameter tuning in the experiments (especially the baselines) is rigorous enough. This is especially important given that this paper proposes a new optimizer. We are also concerned about the novelty of the results, and believe most of these results have appeared in previous work. So I am not increasing my score. I would encourage the authors to do a more rigorous experimental evaluation of the proposed algorithm.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571926917550}, {"id": "HkgBFoJCKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1540/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper the authors develop variants of Adam which corrects for the relationship of the gradient and adaptive terms that causes convergence issues, naming them Delayed Adam and AvaGrad. They also provide proofs demonstrating they solve the convergence issues of Adam in O(1/sqrt(T)) time. They also introduce a convex problem where Adam fails to converge to a stationary point.\n\nThis paper is clearly written and has reasonable experimental support of its claims. However in terms of novelty, AdaShift was published at ICLR last year (https://openreview.net/forum?id=HkgTkhRcKQ) and seems to include a closely related analysis and update rule of your proposed optimizers. In AdaShift instead of correcting for the correlation between the gradient and eta_t, they correct for the relationship between the gradient and the second moment term v_t. Could you further clarify the differences between the two, both in your approach to deriving the new update rule and the algorithms themselves? Additionally, their Theorem 1 could be compared to yours, but noting the differences for these seems less important. If the optimizers are unique enough, including AdaShift in your experiments would be very useful for demonstrating their differences.\n\nRegarding experiments, while it is true that adaptive methods are supposed to be less “sensitive” to hyperparameter choices, the limits of the feasible ranges for each hyperparameter could vary drastically across problems (especially, as previously demonstrated, across different batch sizes.) Thus, not retuning across experiments seems like it could negatively affect performance for any of the transferred hyperparameter settings. Instead of demonstrating hyperparameter insensitivity by carrying over hyperparameter settings, one could instead retune for each problem and show that a higher percent of hyperparameter combinations result in the same/similar best performance (similar to what is done in Fig. 2, but also showing a (1-dimensional) SGD version which would presumably contain fewer high performing settings.)\n\nSome additional comments:\n-The contribution of Theorem 1 is a nice addition to the literature.\n-Your tuning of epsilon is great! I believe more papers should include epsilon in their hyperparameter sweeps.\n-Scaling epsilon with step size makes sense when considering that Adam is similar to a trust region method, where epsilon is inversely proportional to the trust region radius. However, in section 5 implying that epsilon should be as large as possible in the worst case seems like an odd result given that this would always diminish your second moment term as much as possible, defeating the point of the additional Adam-like adaptivity. Can you comment on why this diminished adaptivity would be desirable in the worst case scenario analyzed?\n-The synthetic toy problem is much appreciated, more papers should start with a small, interpretable experiment.\n-Was SGD with momentum used? If not, this may not be a fair comparison, as I believe it is much more common to use momentum with SGD. If momentum was used, was the momentum hyperparameter tuned? If not, this may be advantageous to the Adam based methods, as they have more versatile adaptability and thus may not need as much care with their selection of momentum values.\n-Was a validation set used for CIFAR? You note in appendix C that there are 50k train and 10k test. You mention validation performance in the main text, so this is just double checking.\n-The demonstration in figures 2, 3 of decoupling the step size and epsilon in interesting! Given that the best performing values seem to be on the edges of the ranges tested, I would be curious if the trend continues for more extreme values of alpha and epsilon (one could sparsely search along the predicted trendlines.)\n\nNits:\n-“Vanilla SGD is still prevalent, in spite of the development of seemingly more sophisticated adaptive alternatives...” This could use some citations to back up the claim, because as far as I know it is much more common to use SGD with momentum and is actually rare to use vanilla SGD (the DenseNets and Resnets citations use momentum=0.9.)\n-It would be nice to highlight in color the diff from vanilla Adam in the Algorithm sections.\n-It is not super clear from the text how in eq 26 you get \\sum{E[f(w_t)|Z]} = f(w_1)\n-I may be misreading something, but I believe the H in the leftmost term in the last line of eq 33 should be an L.\n-In section 5, “for a fixed learning rate (as is typically done in practice, except for discrete decays during training)” seems like an overly broad claim, given that authors commonly use polynomial, linear, exponential, cosine, or other learning rate decay/warmups. Granted for some CIFAR and ImageNet benchmarks there are more common discrete learning rate schedules, but that does not seem to be the overwhelmingly prevalent technique.\n\nOverall, while this area of analyzing Adam and proposing modifications is a popular and crowded subject, I believe this paper may contribute to it if my concerns are addressed. While I currently do not recommend acceptance, I am open to changing my score after considering the author comments!\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "In this paper the authors develop variants of Adam which corrects for the relationship of the gradient and adaptive terms that causes convergence issues, naming them Delayed Adam and AvaGrad. They also provide proofs demonstrating they solve the convergence issues of Adam in O(1/sqrt(T)) time. They also introduce a convex problem where Adam fails to converge to a stationary point.\n\nThis paper is clearly written and has reasonable experimental support of its claims. However in terms of novelty, AdaShift was published at ICLR last year (https://openreview.net/forum?id=HkgTkhRcKQ) and seems to include a closely related analysis and update rule of your proposed optimizers. In AdaShift instead of correcting for the correlation between the gradient and eta_t, they correct for the relationship between the gradient and the second moment term v_t. Could you further clarify the differences between the two, both in your approach to deriving the new update rule and the algorithms themselves? Additionally, their Theorem 1 could be compared to yours, but noting the differences for these seems less important. If the optimizers are unique enough, including AdaShift in your experiments would be very useful for demonstrating their differences.\n\nRegarding experiments, while it is true that adaptive methods are supposed to be less “sensitive” to hyperparameter choices, the limits of the feasible ranges for each hyperparameter could vary drastically across problems (especially, as previously demonstrated, across different batch sizes.) Thus, not retuning across experiments seems like it could negatively affect performance for any of the transferred hyperparameter settings. Instead of demonstrating hyperparameter insensitivity by carrying over hyperparameter settings, one could instead retune for each problem and show that a higher percent of hyperparameter combinations result in the same/similar best performance (similar to what is done in Fig. 2, but also showing a (1-dimensional) SGD version which would presumably contain fewer high performing settings.)\n\nSome additional comments:\n-The contribution of Theorem 1 is a nice addition to the literature.\n-Your tuning of epsilon is great! I believe more papers should include epsilon in their hyperparameter sweeps.\n-Scaling epsilon with step size makes sense when considering that Adam is similar to a trust region method, where epsilon is inversely proportional to the trust region radius. However, in section 5 implying that epsilon should be as large as possible in the worst case seems like an odd result given that this would always diminish your second moment term as much as possible, defeating the point of the additional Adam-like adaptivity. Can you comment on why this diminished adaptivity would be desirable in the worst case scenario analyzed?\n-The synthetic toy problem is much appreciated, more papers should start with a small, interpretable experiment.\n-Was SGD with momentum used? If not, this may not be a fair comparison, as I believe it is much more common to use momentum with SGD. If momentum was used, was the momentum hyperparameter tuned? If not, this may be advantageous to the Adam based methods, as they have more versatile adaptability and thus may not need as much care with their selection of momentum values.\n-Was a validation set used for CIFAR? You note in appendix C that there are 50k train and 10k test. You mention validation performance in the main text, so this is just double checking.\n-The demonstration in figures 2, 3 of decoupling the step size and epsilon in interesting! Given that the best performing values seem to be on the edges of the ranges tested, I would be curious if the trend continues for more extreme values of alpha and epsilon (one could sparsely search along the predicted trendlines.)\n\nNits:\n-“Vanilla SGD is still prevalent, in spite of the development of seemingly more sophisticated adaptive alternatives...” This could use some citations to back up the claim, because as far as I know it is much more common to use SGD with momentum and is actually rare to use vanilla SGD (the DenseNets and Resnets citations use momentum=0.9.)\n-It would be nice to highlight in color the diff from vanilla Adam in the Algorithm sections.\n-It is not super clear from the text how in eq 26 you get \\sum{E[f(w_t)|Z]} = f(w_1)\n-I may be misreading something, but I believe the H in the leftmost term in the last line of eq 33 should be an L.\n-In section 5, “for a fixed learning rate (as is typically done in practice, except for discrete decays during training)” seems like an overly broad claim, given that authors commonly use polynomial, linear, exponential, cosine, or other learning rate decay/warmups. Granted for some CIFAR and ImageNet benchmarks there are more common discrete learning rate schedules, but that does not seem to be the overwhelmingly prevalent technique.\n\nOverall, while this area of analyzing Adam and proposing modifications is a popular and crowded subject, I believe this paper may contribute to it if my concerns are addressed. While I currently do not recommend acceptance, I am open to changing my score after considering the author comments!\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571842940700}, {"id": "r1lrEWi3YB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1540/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper proposes a new adaptive method, which is called AvaGrad. The authors first show that Adam may not converge to a stationary point for a stochastic convex optimization in Theorem1, which is closely related to [1]. They then show that by simply making $eta_t$ to be independent of the sample $s_t$, Adam is able to converge just like SGD in Theorem2. Theorem2 follows the standard SGD techniques. Next, they propose AVAGRAD, which is based on the idea of getting rid of the effect of $\\epsilon$. \n\nStrength:\nThe experiment results are impressive. They show that Adam can outperform SGD on vision tasks.\nRecently, people have found out that $\\epsilon$ is a very sensitive hyper-parameter. It is good to see some research directly addresses this problem. \n\nWeakness:\nThe word \"domain\" is confusing. \nIf the Adam-type algorithms are the delayed version in Table 1?\nIt is not compatible with AdamW. \nThe results on the image datasets seem too good to be true.\n\nImplementation Issue:\n***Many implementation details in the below discussion are different from the paper (e.g. hyperparameters and network architecture). So the following experiment results may not be used for assessment of the quality of the proposed method.***\nI tried the proposed Delayed Adam on CIFAR-10 using the codebase in (https://github.com/LiyuanLucasLiu/RAdam/tree/master/cifar_imagenet). The performance seems the same as Adam. Delayed Adam even leads to a *divergence* problem, especially with a large learning rate (0.03). The divergence problem never happens when using Adam and AdamW with the same hyperparameters.\nImplementation details: \n1. Replace the optimizer with the original PyTorch Adam implementation. (https://github.com/pytorch/pytorch/blob/master/torch/optim/adam.py) \n2. Swap line 96 and 108 as suggested in the paper. \n3. Modified line 89 (bias_correction2=1 - beta2 ** (state['step']-1)) \n4. Do not run line 97-107 when state['step']==1. \n5. Run the following code: python cifar.py -a resnet --depth 20 --epochs 164 --schedule 81 122 --gamma 0.1 --wd 1e-4 --optimizer adam  --beta1 0.9 --beta2 0.999  --checkpoint ./logdir --gpu-id 0 --model_name adam_003 --lr 0.03\n\nIf the authors can provide more implementation details, I would promote my rating. \ne.g., \n1. Are you still using bias correction in the proposed method? If so how do you use them?\n2. Do you update the model for the first step?\n\nReference:\n[1] On the convergence of adam and beyond\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary:\nThis paper proposes a new adaptive method, which is called AvaGrad. The authors first show that Adam may not converge to a stationary point for a stochastic convex optimization in Theorem1, which is closely related to [1]. They then show that by simply making $eta_t$ to be independent of the sample $s_t$, Adam is able to converge just like SGD in Theorem2. Theorem2 follows the standard SGD techniques. Next, they propose AVAGRAD, which is based on the idea of getting rid of the effect of $\\epsilon$. \n\nStrength:\nThe experiment results are impressive. They show that Adam can outperform SGD on vision tasks.\nRecently, people have found out that $\\epsilon$ is a very sensitive hyper-parameter. It is good to see some research directly addresses this problem. \n\nWeakness:\nThe word \"domain\" is confusing. \nIf the Adam-type algorithms are the delayed version in Table 1?\nIt is not compatible with AdamW. \nThe results on the image datasets seem too good to be true.\n\nImplementation Issue:\n***Many implementation details in the below discussion are different from the paper (e.g. hyperparameters and network architecture). So the following experiment results may not be used for assessment of the quality of the proposed method.***\nI tried the proposed Delayed Adam on CIFAR-10 using the codebase in (https://github.com/LiyuanLucasLiu/RAdam/tree/master/cifar_imagenet). The performance seems the same as Adam. Delayed Adam even leads to a *divergence* problem, especially with a large learning rate (0.03). The divergence problem never happens when using Adam and AdamW with the same hyperparameters.\nImplementation details: \n1. Replace the optimizer with the original PyTorch Adam implementation. (https://github.com/pytorch/pytorch/blob/master/torch/optim/adam.py) \n2. Swap line 96 and 108 as suggested in the paper. \n3. Modified line 89 (bias_correction2=1 - beta2 ** (state['step']-1)) \n4. Do not run line 97-107 when state['step']==1. \n5. Run the following code: python cifar.py -a resnet --depth 20 --epochs 164 --schedule 81 122 --gamma 0.1 --wd 1e-4 --optimizer adam  --beta1 0.9 --beta2 0.999  --checkpoint ./logdir --gpu-id 0 --model_name adam_003 --lr 0.03\n\nIf the authors can provide more implementation details, I would promote my rating. \ne.g., \n1. Are you still using bias correction in the proposed method? If so how do you use them?\n2. Do you update the model for the first step?\n\nReference:\n[1] On the convergence of adam and beyond\n"}, "tcdate": 1571758381062}], "openreview_url": "https://openreview.net/forum?id=HylNWkHtvB", "arxiv_id": "1912.01823", "paper_pdf": "papers/HylNWkHtvB.pdf", "paper_pdf_sha256": "69ee064c63d05aba1b8e8f60d445c7df6ebbd405434135b70034287016301d43", "paper_pdf_bytes": 355911, "paper_pdf_source": "openreview", "code_url": "https://github.com/lolemacs/avagrad", "code_repository": "lolemacs/avagrad", "code_commit": "343733151f26a1fa8e504079c6d7934a6038a93b", "code_archive": "repos/HylNWkHtvB.zip", "code_archive_sha256": "4ae059a3b716f559ead8828fa2c4a0006b88eebdacaceb4066dc60c5be20255d", "code_archive_bytes": 8780, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 10, "github_languages": {"Python": 26004}, "github_archived": false, "github_pushed_at": "2020-12-15T04:01:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/domain-independent-dominance-of-adaptive-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1E64jC5tm", "year": 2019, "status": "rejected", "title": "The Forward-Backward Embedding of Directed Graphs", "authors": ["Thomas Bonald", "Nathan De Lara"], "authorids": ["thomas.bonald@telecom-paristech.fr", "nathan.delara@telecom-paristech.fr"], "authors_source": "OpenReview API", "abstract": "We introduce a novel embedding of directed graphs derived from the singular value decomposition (SVD) of the  normalized adjacency matrix. Specifically, we show that, after proper normalization of the singular vectors,  \n the  distances between vectors in  the embedding space are proportional to the mean commute times between the corresponding  nodes by a  forward-backward random walk in the graph, which follows the edges  alternately in forward and backward directions.  In particular, two nodes having many common successors in the graph tend to be represented by close vectors in the embedding space. More formally, we prove that our representation of the graph is  equivalent to the spectral embedding of some co-citation graph, where  nodes are linked with respect to their common set of successors in the original graph. The interest of our approach is that it does not require to build this co-citation graph, which is typically much denser than the original graph. Experiments  on  real datasets show the efficiency of the approach. \n", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SkxVfLSzaX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper43/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper discusses how to embed nodes of a graph to a vector space using singular value decomposition of the normalized adjacency matrix. The material in this paper are standard literature in graph theory and spectral analysis. The paper contains a well-written literature review. The contributions and not new enough for a publication.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "well written - standard / light content  ", "review": "The paper discusses how to embed nodes of a graph to a vector space using singular value decomposition of the normalized adjacency matrix. The material in this paper are standard literature in graph theory and spectral analysis. The paper contains a well-written literature review. The contributions and not new enough for a publication.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541719563586}, {"id": "r1e5PpwR37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper43/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies embeddings of directed graphs based on SVD. It proposes an interpretation of the embedding obtained from the normalized adjacency matrix in terms of the forward-backward random walk on the graph. In such random walk odd steps are taken using the edges of the graph while even steps are taken using reverse edges. Such an interpretation seems to be a trivial extension of the work for undirected graphs. \n\nFor example, consider bipartite graphs, in such graphs the forward-backward random walk can be seen as a standard random walk on the graph with directions removed. Indeed, if the walk starts in part A then remove all edges from B to A and make A to B edges undirected.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Incremental and trivial observations", "review": "The paper studies embeddings of directed graphs based on SVD. It proposes an interpretation of the embedding obtained from the normalized adjacency matrix in terms of the forward-backward random walk on the graph. In such random walk odd steps are taken using the edges of the graph while even steps are taken using reverse edges. Such an interpretation seems to be a trivial extension of the work for undirected graphs. \n\nFor example, consider bipartite graphs, in such graphs the forward-backward random walk can be seen as a standard random walk on the graph with directions removed. Indeed, if the walk starts in part A then remove all edges from B to A and make A to B edges undirected.\n\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541467489912}, {"id": "BkeLffqx3Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper43/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary of the paper\n\nThis paper proposes an embedding of directed graphs based on the SVD of a normalized adjacency matrix. This embedding is shown to be equivalent to the spectral embedding of a co-citation graph, which is more complex to calculate. Interestingly, the proposed approach does not require the *explicit* representation of this graph. Moreover, the paper also shows that distances of the embedded vectors are proportional to mean commute times of a forward--backward random walk in the original graph. A suite of experiments is run on graphs from KONECT.\n\n# Review\n\nThis is a well-written paper, which I enjoyed reading. The extension of embeddings to the case of directed graphs is significant and warrants a detailed exploration.\n\nThe principal issues I see with this paper are as follows:\n\n- The originality or scope of the contribution is not clear\n- The experimental section is uncompelling\n- Several relevant works appear to have been ignored\n\nOverall, I like the way the paper treats the subject. In particular, I appreciate the fact that proofs are explained well; additionally, code is provided, which will increase reproducibility. This is uncommon and praiseworthy!\n\nAs for the originality of the paper, I find it hard to judge the scope of the contribution. The paper is extremely well written and employs a very pedagogical treatment of the subject, which I appreciate. Yet, it is hard for me to judge the utility and novelty of the proposed method in light of Section 8, where the paper shows that a spectral embedding of the undirected variant of the graph leads to essentially the *same* eigenvectors (up to renormalization and permutations). To prove that a new method is more effective, this point should be emphasized more:\n\n1. In a sense, I would see the results from Section 8 as the equal to what 'Laplacian Eigenmaps' (LE) yields. This needs to be stressed, and analysed in an experimental section.\n\n2. I understand that the order of the singular vectors is different, so embeddings that use only parts of them will be different. However, a convincing experiment should assess the differences. For example, in which regime for $d$ (number of used vectors) will the new method be surpassed by the old one? Is there such a regime? Ideally, this will be answered in the form of an asymptotic theorem; it could also be a larger experiment, though (to simulate the conditions in practice).\n\n3. I understand that the new approach has a lower run-time, because the SVD is more efficient than eigendecomposition. However, what about a simple baseline algorithm that uses SVD for the *undirected* graph of the input data? This should be simple to accomplish, and would be a way to ascertain the benefit of using edge directions.\n\n  To my understanding, LE should be this embedding, but from the table, I can see that its runtime is a lot worse than the novel method. What causes this? The fact that eigendecomposition is used instead of SVD?  If so, an additional SVD-based approach should be implemented.\n\nThis brings me to the experimental section. Here, the paper demonstrates the superiority of the new embedding based on evaluating modularity of a set of different clusterings of larger graphs, obtained using $k$-means. I have several concerns about this:\n\n1. Modularity has problems with larger networks because only a small part of the network will be used in its configuration.\n\n2. Since the embeddings cannot be easily compared due to missing ground truth information, other metrics should be employed. Here are a few, which are often used  by the community. See 'Is there a best quality metric for graph clusters?' by Almeida et al. for more details and a description of their shortcomings:\n    - Silhouette coefficient\n    - Coverage\n    - Conductance\n\n  Different ones should be evaluated here in order to show the behaviour of the new embedding. Do the embeddings differ if the modularities are similar?\n\n3. How do the results change for different values of $d$? I find it hard to disentangle such a discussion from instabilities in $k$-means, but to my understanding of the method, tuning $d$ means that more or less information is used from the singular vectors.\n\n   This could also be quantified in a proof (about asymptotic behaviour) but an experiment would be equally fine.\n\nConcerning the bibliography, or the treatment of prior works, there are some issues:\n\n- There appear to be some missing references of earlier works that used SVD or variants in order to cluster graphs or embed them:\n\n  - Drineas et al.: 'Clustering Large Graphs via the Singular Value Decomposition'\n  - Malliaros and Vazirgiannis: 'Clustering and Community Detection in Directed Networks: A Survey'\n\n- Likewise, the use of pseudo-inverse Laplacians has a lot more papers attached to it (these are only a few that are relevant):\n\n  - Ho and Dooren: 'On the pseudo-inverse of the Laplacian of a bipartite graph'\n  - Gutman and Xia: 'Generalized inverse of the Laplacian matrix and some applications'\n\n# Suggestions for improvement\n\n- In some sense, this work can be seen as an extension of Laplacian eigenmaps to the directed case. The paper needs to be more clear about these extensions with respect to prior work. In Section 2, it is claimed that 'our main contribution is a proper normalization'. This strikes me as a rather small contribution in light of the experimental section, as outlined above.\n\n- I am also hesitant to speak about a better interpretability of the mean commute time. I agree that it is nice to know that the distance permits such an interpretation in terms of random walks, but what is the impact of knowing the MCT? It is not only used in the embedding insofar as one obtains a vector representation.\n\n- Section 5 is then the standard way of defining random walks based on a Laplacian matrix, and the correspondence to the pseudo-inverse of the Laplacian is shown. This is mathematically interesting, but appears to me to be in line with previous research.\n\n- The section about co-citation graphs should make it more clear that 'successors' are to be taken in terms of the original graph and the directionality of edges. Since this is a standard definition in the domain of network analysis I would suggest citing a textbook here.\n\n- In Section 6, the paper could give more details about random walk concepts such as 'stochastic', 'stationary distribution' etc., as it would make the paper more accessible (I am familiar with these concepts but since the writing of the paper is of high quality in the other sections, I am convinced this would improve its impact, and attract more readers).\n\nTypos & grammar issues:\n\n- 'in terms of random walk' --> 'in terms of random walks'\n- 'equivalent to build' --> 'equivalent to building'\n- 'with corresponding unitary matrix' --> 'with a corresponding unitary\n  matrix'\n- 'square Euclidean distance' --> 'squared Euclidean distance'\n- 'equivalent to consider' --> 'equivalent to considering'\n- 'irreductible' --> 'irreducible'\n- 'and provide generally' --> 'and provides generally'\n- 'in low dimension' --> 'in a lower dimension'\n\nFurthermore, the bibliography should employ consistent capitalization and journal names for articles.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting and valuable idea, but flawed in terms of execution and distinction/difference to previous works", "review": "# Summary of the paper\n\nThis paper proposes an embedding of directed graphs based on the SVD of a normalized adjacency matrix. This embedding is shown to be equivalent to the spectral embedding of a co-citation graph, which is more complex to calculate. Interestingly, the proposed approach does not require the *explicit* representation of this graph. Moreover, the paper also shows that distances of the embedded vectors are proportional to mean commute times of a forward--backward random walk in the original graph. A suite of experiments is run on graphs from KONECT.\n\n# Review\n\nThis is a well-written paper, which I enjoyed reading. The extension of embeddings to the case of directed graphs is significant and warrants a detailed exploration.\n\nThe principal issues I see with this paper are as follows:\n\n- The originality or scope of the contribution is not clear\n- The experimental section is uncompelling\n- Several relevant works appear to have been ignored\n\nOverall, I like the way the paper treats the subject. In particular, I appreciate the fact that proofs are explained well; additionally, code is provided, which will increase reproducibility. This is uncommon and praiseworthy!\n\nAs for the originality of the paper, I find it hard to judge the scope of the contribution. The paper is extremely well written and employs a very pedagogical treatment of the subject, which I appreciate. Yet, it is hard for me to judge the utility and novelty of the proposed method in light of Section 8, where the paper shows that a spectral embedding of the undirected variant of the graph leads to essentially the *same* eigenvectors (up to renormalization and permutations). To prove that a new method is more effective, this point should be emphasized more:\n\n1. In a sense, I would see the results from Section 8 as the equal to what 'Laplacian Eigenmaps' (LE) yields. This needs to be stressed, and analysed in an experimental section.\n\n2. I understand that the order of the singular vectors is different, so embeddings that use only parts of them will be different. However, a convincing experiment should assess the differences. For example, in which regime for $d$ (number of used vectors) will the new method be surpassed by the old one? Is there such a regime? Ideally, this will be answered in the form of an asymptotic theorem; it could also be a larger experiment, though (to simulate the conditions in practice).\n\n3. I understand that the new approach has a lower run-time, because the SVD is more efficient than eigendecomposition. However, what about a simple baseline algorithm that uses SVD for the *undirected* graph of the input data? This should be simple to accomplish, and would be a way to ascertain the benefit of using edge directions.\n\n  To my understanding, LE should be this embedding, but from the table, I can see that its runtime is a lot worse than the novel method. What causes this? The fact that eigendecomposition is used instead of SVD?  If so, an additional SVD-based approach should be implemented.\n\nThis brings me to the experimental section. Here, the paper demonstrates the superiority of the new embedding based on evaluating modularity of a set of different clusterings of larger graphs, obtained using $k$-means. I have several concerns about this:\n\n1. Modularity has problems with larger networks because only a small part of the network will be used in its configuration.\n\n2. Since the embeddings cannot be easily compared due to missing ground truth information, other metrics should be employed. Here are a few, which are often used  by the community. See 'Is there a best quality metric for graph clusters?' by Almeida et al. for more details and a description of their shortcomings:\n    - Silhouette coefficient\n    - Coverage\n    - Conductance\n\n  Different ones should be evaluated here in order to show the behaviour of the new embedding. Do the embeddings differ if the modularities are similar?\n\n3. How do the results change for different values of $d$? I find it hard to disentangle such a discussion from instabilities in $k$-means, but to my understanding of the method, tuning $d$ means that more or less information is used from the singular vectors.\n\n   This could also be quantified in a proof (about asymptotic behaviour) but an experiment would be equally fine.\n\nConcerning the bibliography, or the treatment of prior works, there are some issues:\n\n- There appear to be some missing references of earlier works that used SVD or variants in order to cluster graphs or embed them:\n\n  - Drineas et al.: 'Clustering Large Graphs via the Singular Value Decomposition'\n  - Malliaros and Vazirgiannis: 'Clustering and Community Detection in Directed Networks: A Survey'\n\n- Likewise, the use of pseudo-inverse Laplacians has a lot more papers attached to it (these are only a few that are relevant):\n\n  - Ho and Dooren: 'On the pseudo-inverse of the Laplacian of a bipartite graph'\n  - Gutman and Xia: 'Generalized inverse of the Laplacian matrix and some applications'\n\n# Suggestions for improvement\n\n- In some sense, this work can be seen as an extension of Laplacian eigenmaps to the directed case. The paper needs to be more clear about these extensions with respect to prior work. In Section 2, it is claimed that 'our main contribution is a proper normalization'. This strikes me as a rather small contribution in light of the experimental section, as outlined above.\n\n- I am also hesitant to speak about a better interpretability of the mean commute time. I agree that it is nice to know that the distance permits such an interpretation in terms of random walks, but what is the impact of knowing the MCT? It is not only used in the embedding insofar as one obtains a vector representation.\n\n- Section 5 is then the standard way of defining random walks based on a Laplacian matrix, and the correspondence to the pseudo-inverse of the Laplacian is shown. This is mathematically interesting, but appears to me to be in line with previous research.\n\n- The section about co-citation graphs should make it more clear that 'successors' are to be taken in terms of the original graph and the directionality of edges. Since this is a standard definition in the domain of network analysis I would suggest citing a textbook here.\n\n- In Section 6, the paper could give more details about random walk concepts such as 'stochastic', 'stationary distribution' etc., as it would make the paper more accessible (I am familiar with these concepts but since the writing of the paper is of high quality in the other sections, I am convinced this would improve its impact, and attract more readers).\n\nTypos & grammar issues:\n\n- 'in terms of random walk' --> 'in terms of random walks'\n- 'equivalent to build' --> 'equivalent to building'\n- 'with corresponding unitary matrix' --> 'with a corresponding unitary\n  matrix'\n- 'square Euclidean distance' --> 'squared Euclidean distance'\n- 'equivalent to consider' --> 'equivalent to considering'\n- 'irreductible' --> 'irreducible'\n- 'and provide generally' --> 'and provides generally'\n- 'in low dimension' --> 'in a lower dimension'\n\nFurthermore, the bibliography should employ consistent capitalization and journal names for articles.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540559374262}], "openreview_url": "https://openreview.net/forum?id=S1E64jC5tm", "arxiv_id": null, "paper_pdf": "papers/S1E64jC5tm.pdf", "paper_pdf_sha256": "8544849615a647b537158912f5832b63bf3b6415d01ddb6c4f9f951d494c8c04", "paper_pdf_bytes": 316920, "paper_pdf_source": "openreview", "code_url": "https://github.com/tbonald/directed", "code_repository": "tbonald/directed", "code_commit": "2d4b979f296bb0750ee1560ac07d70ff68f7c8dc", "code_archive": "repos/S1E64jC5tm.zip", "code_archive_sha256": "da3639ca90e6775705e265ab6256cb2cf997f0b08f0b91f7b7c6fb5455efc695", "code_archive_bytes": 13245, "code_file_count": 4, "code_extensions": {".py": 3, ".ipynb": 1}, "github_disk_usage_kb": 17, "github_languages": {"Python": 21945, "Jupyter Notebook": 17282}, "github_archived": false, "github_pushed_at": "2018-11-08T09:38:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-forward-backward-embedding-of-directed"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1EwLkW0W", "year": 2018, "status": "rejected", "title": "Dissecting Adam: The Sign, Magnitude and Variance of Stochastic Gradients", "authors": ["Lukas Balles", "Philipp Hennig"], "authorids": ["lukas.balles@tuebingen.mpg.de", "ph@tue.mpg.de"], "authors_source": "OpenReview API", "abstract": "The ADAM optimizer is exceedingly popular in the deep learning community. Often it works very well, sometimes it doesn’t. Why? We interpret ADAM as a combination of two aspects: for each weight, the update direction is determined by the sign of the stochastic gradient, whereas the update magnitude is solely determined by an estimate of its relative variance. We  disentangle these two aspects and analyze them in isolation, shedding light on ADAM ’s inner workings. Transferring the \"variance adaptation” to momentum- SGD gives rise to a novel method, completing the practitioner’s toolbox for problems where ADAM fails.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "S1urbvOgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper482/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents some analysis of the scale-invariance and the particular shape of the learning rate used in Adam. The paper argues that Adam's update is a combination of a sign-update and a variance-based learning rate. Some analysis is provided on these two aspects.\n\nAfter spending a sizeable amount of time with this paper, I am not sure what are its novel contributions and why it should be published in a scientific conference. The paper contains so many approximations, simplifications, and assumptions that make any presented result extremely weak.\n\nMore in details, the analysis of the sign is done in the case of quadratic functions of Gaussian variables. The result is mildly interesting, but I fail to see how this would give us a hint of what is happening during the minimization of the non-convex objectives for training deep networks.\nMoreover, the analysis of sign based updates has been already carried over using the Polyak-Łojasiewicz assumption in Karimi et al. ECML-PKDD 2016, that is strictly more general than any quadratic approximation.\n\nThe similarity between the ``optimal'' variance-based learning rate and the one of Adam hinges again on the fact that the noise is Gaussian. As the authors admit, Schaul et al. (2013) already derived similar updates. Also, Theorem 1 recover the usual rate of convergence for strongly convex function: How is this theorem supposed to support the fact that variance-adapted learning rates are a better idea than the usual updates?\nMoreover, the proof of Theorem 1 hinges on the fact that E[||g_t||^2]\\leq G^2. Clearly, this is not possible in general for a strongly convex function. The proof might still go through, but it needs to be fixed using the fact that the updates always decrease the function.\n\nOverall, if we are considering only the convex case, Adam is clearly sub-optimal from all the points of view and better algorithms with stronger guarantees can be used. Indeed, the fact that non-convexity is never discussed is particularly alarming. It is also indicative that none of the work for minimization of finite sums are cited or discussed, e.g. the variance reduced methods immediately come to mind.\n\nRegarding the experiments, the parameters are chosen to have the best test accuracy, mixing the machine learning problem with the optimization one: it is well-known and easy to prove that a worst optimizer can give rise to better test errors. Hence, the empirical results cannot be used to support any of the proposed interpretations nor the new optimization algorithms.\n\nTo summarize, I do not think the contributions of this paper are enough to be published in ICLR.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "rating": "4: Ok but not good enough - rejection", "review": "The paper presents some analysis of the scale-invariance and the particular shape of the learning rate used in Adam. The paper argues that Adam's update is a combination of a sign-update and a variance-based learning rate. Some analysis is provided on these two aspects.\n\nAfter spending a sizeable amount of time with this paper, I am not sure what are its novel contributions and why it should be published in a scientific conference. The paper contains so many approximations, simplifications, and assumptions that make any presented result extremely weak.\n\nMore in details, the analysis of the sign is done in the case of quadratic functions of Gaussian variables. The result is mildly interesting, but I fail to see how this would give us a hint of what is happening during the minimization of the non-convex objectives for training deep networks.\nMoreover, the analysis of sign based updates has been already carried over using the Polyak-Łojasiewicz assumption in Karimi et al. ECML-PKDD 2016, that is strictly more general than any quadratic approximation.\n\nThe similarity between the ``optimal'' variance-based learning rate and the one of Adam hinges again on the fact that the noise is Gaussian. As the authors admit, Schaul et al. (2013) already derived similar updates. Also, Theorem 1 recover the usual rate of convergence for strongly convex function: How is this theorem supposed to support the fact that variance-adapted learning rates are a better idea than the usual updates?\nMoreover, the proof of Theorem 1 hinges on the fact that E[||g_t||^2]\\leq G^2. Clearly, this is not possible in general for a strongly convex function. The proof might still go through, but it needs to be fixed using the fact that the updates always decrease the function.\n\nOverall, if we are considering only the convex case, Adam is clearly sub-optimal from all the points of view and better algorithms with stronger guarantees can be used. Indeed, the fact that non-convexity is never discussed is particularly alarming. It is also indicative that none of the work for minimization of finite sums are cited or discussed, e.g. the variance reduced methods immediately come to mind.\n\nRegarding the experiments, the parameters are chosen to have the best test accuracy, mixing the machine learning problem with the optimization one: it is well-known and easy to prove that a worst optimizer can give rise to better test errors. Hence, the empirical results cannot be used to support any of the proposed interpretations nor the new optimization algorithms.\n\nTo summarize, I do not think the contributions of this paper are enough to be published in ICLR.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511711040136}, {"id": "r1mT8HDgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper482/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \nThe paper is trying to improve Adam based on variance adaption with momentum. Two algorithms are proposed, M-SSD (Stochastic Sign Descent with Momentum) and M-SVAG (Stochastic Variance-Adapted Gradient with Momentum) to solve finite sum minimization problem. The convergence analysis is provided for SVAG for strongly convex case. Numerical experiments are provided for some standard neural network structures with three common datasets MNIST, CIFAR10 and CIFAR100 compared the performance of M-SSD and M-SVAG to two existing algorithms: SGD momentum and Adam. \n \nComments:\nPage 4, line 5: You should define \\nu clearly.\n\nTheorem 1: In the strongly convex case, assumption E ||g_t ||^2 \\leq G^2 (if G is a constant) is too strong. In this case, G could be equal to infinity. If G is not infinity, you already assume that your algorithm converges, that is the reason why this assumption is not so good for strongly convex. If G is infinity (this is really possible for strongly convex), your proof would get a trouble as eq. (40) is not valid anymore.\n\nAlso, to compute \\gamma_{t,i}, it requires to compute \\nabla f_{t,i}, which is full gradient. By doing this, the computational cost should add the dependence of M, which is very large as you mentioned in the introduction. According to your rate O(1/t), the complexity is worse than that of gradient descent and SGD as well. \n\nAs I understand, there is no theoretical results for M-SSG and M-SVAG, but only the result for SVAG with exact \\eta_i^2 in the strongly convex case. Also, theoretical results are not strong enough. Hence, the experiments need to make more convincingly, at least for some different complicated architecture of deep neural network. As I see, in some dataset, Adam performs better than M-SSD, some another dataset, Adam performs better than M-SVAG. Same situation for M-SGD. My question is that: When should we use M-SSD or M-SVAG? For a given dataset, why should we not use Adam or M-SGD (or other existing algorithms such as Adagrad, RMSprop), but your algorithms? \n\nYou should do more experiments to various dataset and architectures to be more convincing since theoretical results are not strong enough. Would you think to try to use VGG or ResNet to ImageNet?\n\nI like the idea of the paper but I would love if the author(s) could improve more theoretical results to convince people. Otherwise, the results in this paper could not be considered as good enough. At this moment, I think the paper is still not ready for the publication. \n\nMinor comments:\nPage 2, in eq. (6): You should mention that “1” is a vector.\nPage 4, line 4: Q in R^{d} => Q in R^{d x d}\nPage 6, Theorem 1: You should define the finite sum optimization problem with f since you have not used it before.\nPage 6, Theorem 1: You should use another notation for “\\mu”-strongly convex parameter since you have another “\\mu”-momentum parameter in section 3.4\nPage 4, Page 7: Be careful with the case when c = 0 (page 4) and mu = 1 (page 7-8) with dividing by 0. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Dissecting Adam: The Sign, Magnitude and Variance of Stochastic Gradients", "rating": "4: Ok but not good enough - rejection", "review": "Summary: \nThe paper is trying to improve Adam based on variance adaption with momentum. Two algorithms are proposed, M-SSD (Stochastic Sign Descent with Momentum) and M-SVAG (Stochastic Variance-Adapted Gradient with Momentum) to solve finite sum minimization problem. The convergence analysis is provided for SVAG for strongly convex case. Numerical experiments are provided for some standard neural network structures with three common datasets MNIST, CIFAR10 and CIFAR100 compared the performance of M-SSD and M-SVAG to two existing algorithms: SGD momentum and Adam. \n \nComments:\nPage 4, line 5: You should define \\nu clearly.\n\nTheorem 1: In the strongly convex case, assumption E ||g_t ||^2 \\leq G^2 (if G is a constant) is too strong. In this case, G could be equal to infinity. If G is not infinity, you already assume that your algorithm converges, that is the reason why this assumption is not so good for strongly convex. If G is infinity (this is really possible for strongly convex), your proof would get a trouble as eq. (40) is not valid anymore.\n\nAlso, to compute \\gamma_{t,i}, it requires to compute \\nabla f_{t,i}, which is full gradient. By doing this, the computational cost should add the dependence of M, which is very large as you mentioned in the introduction. According to your rate O(1/t), the complexity is worse than that of gradient descent and SGD as well. \n\nAs I understand, there is no theoretical results for M-SSG and M-SVAG, but only the result for SVAG with exact \\eta_i^2 in the strongly convex case. Also, theoretical results are not strong enough. Hence, the experiments need to make more convincingly, at least for some different complicated architecture of deep neural network. As I see, in some dataset, Adam performs better than M-SSD, some another dataset, Adam performs better than M-SVAG. Same situation for M-SGD. My question is that: When should we use M-SSD or M-SVAG? For a given dataset, why should we not use Adam or M-SGD (or other existing algorithms such as Adagrad, RMSprop), but your algorithms? \n\nYou should do more experiments to various dataset and architectures to be more convincing since theoretical results are not strong enough. Would you think to try to use VGG or ResNet to ImageNet?\n\nI like the idea of the paper but I would love if the author(s) could improve more theoretical results to convince people. Otherwise, the results in this paper could not be considered as good enough. At this moment, I think the paper is still not ready for the publication. \n\nMinor comments:\nPage 2, in eq. (6): You should mention that “1” is a vector.\nPage 4, line 4: Q in R^{d} => Q in R^{d x d}\nPage 6, Theorem 1: You should define the finite sum optimization problem with f since you have not used it before.\nPage 6, Theorem 1: You should use another notation for “\\mu”-strongly convex parameter since you have another “\\mu”-momentum parameter in section 3.4\nPage 4, Page 7: Be careful with the case when c = 0 (page 4) and mu = 1 (page 7-8) with dividing by 0. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511638715430}, {"id": "By-KPs9eM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper482/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Stochastic Sign Descent (SSD) and Stochastic Variance Adapted Gradient (SVAG) are inspired by ADAM and studied in this paper, together with momentum terms. \n\nAnalysis showed that SSD should work better than usual SGD when the Hessian of training loss is highly diagonal dominant.  It is intrigued to observe that for MNIST and CIFAR10, SSD with momentum champions with better efficiency than ADAM, SGD and SVAG, while on the other hand, in CIFAR100, momentum-SVAG and SGD beat SSD and ADAM. Does it suggest the Hessians associated with MNIST and CIFAR10 training loss more diagonally dominant? \n\nThere are other adaptive step-sizes such as Barzilai-Borwein (BB) Step Sizes introduced to machine learning by Tan et al. NIPS 2016. Is there any connections between variance adaptation here and BB step size? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper splits ADAM algorithm into two components: stochastic direction in sign of gradient and adaptive stepwise with relative variance. Two algorithms are proposed to test each of them. For MNIST, CIFAR10, and CIFAR100 datasets, Stochastic Sign Descent shows some better performance than others in the former two but worse in the last one.", "rating": "6: Marginally above acceptance threshold", "review": "Stochastic Sign Descent (SSD) and Stochastic Variance Adapted Gradient (SVAG) are inspired by ADAM and studied in this paper, together with momentum terms. \n\nAnalysis showed that SSD should work better than usual SGD when the Hessian of training loss is highly diagonal dominant.  It is intrigued to observe that for MNIST and CIFAR10, SSD with momentum champions with better efficiency than ADAM, SGD and SVAG, while on the other hand, in CIFAR100, momentum-SVAG and SGD beat SSD and ADAM. Does it suggest the Hessians associated with MNIST and CIFAR10 training loss more diagonally dominant? \n\nThere are other adaptive step-sizes such as Barzilai-Borwein (BB) Step Sizes introduced to machine learning by Tan et al. NIPS 2016. Is there any connections between variance adaptation here and BB step size? ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511860089155}], "openreview_url": "https://openreview.net/forum?id=S1EwLkW0W", "arxiv_id": "1705.07774", "paper_pdf": "papers/S1EwLkW0W.pdf", "paper_pdf_sha256": "2c820b68da0c383b2903af5ebb1dbdd750ebd91c21a5918648a4350b627fe321", "paper_pdf_bytes": 656781, "paper_pdf_source": "openreview", "code_url": "https://github.com/lballes/msvag", "code_repository": "lballes/msvag", "code_commit": "d2d467b6a9c6442d781ecb2f1d4b5c5769556363", "code_archive": "repos/S1EwLkW0W.zip", "code_archive_sha256": "cec7a33afea60c29789a2ee5ae6350da0f16dec1de490ba4dcf2a99486c1f23b", "code_archive_bytes": 8966, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 12, "github_languages": {"Python": 10920}, "github_archived": false, "github_pushed_at": "2018-05-11T14:21:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dissecting-adam-the-sign-magnitude-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JtX6oaaJ2d", "year": 2026, "status": "rejected", "title": "Improving LLM Unlearning Robustness via Random Perturbations", "authors": ["Dang Huu-Tien", "Hoang Thanh-Tung", "Anh Tuan Bui", "Phuong Minh Nguyen", "Le-Minh Nguyen", "Naoya Inoue"], "authorids": ["~Dang_Huu-Tien1", "~Hoang_Thanh-Tung1", "~Anh_Tuan_Bui2", "~Phuong_Minh_Nguyen3", "~Le-Minh_Nguyen1", "~Naoya_Inoue1"], "authors_source": "OpenReview API", "abstract": "Here, we show that current state-of-the-art LLM unlearning methods inherently reduce models' robustness, causing them to misbehave even when a single non-adversarial forget-token is present in the retain-query. Toward understanding underlying causes, we propose a novel theoretical framework that reframes the *unlearning process as backdoor attacks and defenses*: forget-tokens act as backdoor triggers that, when activated in retain-queries, cause disruptions in unlearned models' behaviors, similar to successful backdoor attacks. The sense that, LLM unlearning methods *themselves poison the model*, make it more vulnerable to forget-tokens, and *hide rather than erase* target knowledge, describes their true mechanism. To mitigate the vulnerability caused by the forgetting process, we reinterpret the retaining process as a backdoor defense and propose Random Noise Augmentation (RNA), a lightweight, model and method-agnostic approach with theoretical guarantees for improving the robustness of models.  Extensive experiments demonstrate that RNA significantly improves the robustness of unlearned models while preserving forget and retain performances. This backdoor attack-defense framework offers insights into the mechanisms of unlearning that can shed light on future research directions for improving unlearning robustness.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Y7wia56zs2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1191/Reviewer_pZfT"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This work focuses on understanding and improving robustness of llm unlearning. Different from prior efforts that mostly look at robustness to unlearned content, this work focus on robustness to retain content. This work propose a theoretical framework that frames the unlearning problem as a backdoor attack and defense problem. Based on the understanding, the authors propose random perturbations on top of the latent representation to preserve retain performance when the retain set contains forget tokens.", "review_text": "This work focuses on understanding and improving robustness of llm unlearning. Different from prior efforts that mostly look at robustness to unlearned content, this work focus on robustness to retain content. This work propose a theoretical framework that frames the unlearning problem as a backdoor attack and defense problem. Based on the understanding, the authors propose random perturbations on top of the latent representation to preserve retain performance when the retain set contains forget tokens.", "strengths": "- This paper focus on understanding robustness of unlearning and specifically look at robustness on the retain set, which is largely under explored in prior literature.\n- The RNA method is a very simple heuristic and seems to be effective on perturbed retain set.", "weaknesses": "- I did not get Section 5.1. The author said we train model with the \"poisoned\" forget set and the retain set, but equation 10 is still sampling data from $\\mathcal{Z}$, which is the non-\"poisoned\" forget set and retain set. Then I don't understand the role of $T$ and $\\Omega$ here. Now given equation 11 and equation 12 is correct and make sense, why does the conclusion: \"current state-of-the-art LLM unlearning methods themselves “poison” the model and make it more vulnerable to forget-tokens\" follows? There is no analysis of how current LLM unlearning methods map to this framework. Detailed explanation in this part is largely missing, making this section very confusing.\n- In the experiment section, the evaluation is quite limited. The authors only evaluate on wmdp unlearning and only on one forget token \"SARS-CoV-2\". The authors should evaluate on more tasks (tofu, rwku, etc) and a variety of different forget tokens for each task. Otherwise, it's hard to tell whether the proposed method works for just this dataset or for general unlearning task.\n- Figure 1 is very hard to interpret. For each figure, I suggest using different symbols for different methods and one color for baseline unlearning method, one color for the RNA augmented method. Or use a table to present the results. Otherwise it is very hard to quantify the advantage of RNA.", "questions": "See weaknesses above. My main question is the how the framework converts to the conclusion that \"current state-of-the-art LLM unlearning methods themselves “poison” the model and make it more vulnerable to forget-tokens\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work focuses on understanding and improving robustness of llm unlearning. Different from prior efforts that mostly look at robustness to unlearned content, this work focus on robustness to retain content. This work propose a theoretical framework that frames the unlearning problem as a backdoor attack and defense problem. Based on the understanding, the authors propose random perturbations on top of the latent representation to preserve retain performance when the retain set contains forget tokens.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "- This paper focus on understanding robustness of unlearning and specifically look at robustness on the retain set, which is largely under explored in prior literature.\n- The RNA method is a very simple heuristic and seems to be effective on perturbed retain set.", "weaknesses": "- I did not get Section 5.1. The author said we train model with the \"poisoned\" forget set and the retain set, but equation 10 is still sampling data from $\\mathcal{Z}$, which is the non-\"poisoned\" forget set and retain set. Then I don't understand the role of $T$ and $\\Omega$ here. Now given equation 11 and equation 12 is correct and make sense, why does the conclusion: \"current state-of-the-art LLM unlearning methods themselves “poison” the model and make it more vulnerable to forget-tokens\" follows? There is no analysis of how current LLM unlearning methods map to this framework. Detailed explanation in this part is largely missing, making this section very confusing.\n- In the experiment section, the evaluation is quite limited. The authors only evaluate on wmdp unlearning and only on one forget token \"SARS-CoV-2\". The authors should evaluate on more tasks (tofu, rwku, etc) and a variety of different forget tokens for each task. Otherwise, it's hard to tell whether the proposed method works for just this dataset or for general unlearning task.\n- Figure 1 is very hard to interpret. For each figure, I suggest using different symbols for different methods and one color for baseline unlearning method, one color for the RNA augmented method. Or use a table to present the results. Otherwise it is very hard to quantify the advantage of RNA.", "questions": "See weaknesses above. My main question is the how the framework converts to the conclusion that \"current state-of-the-art LLM unlearning methods themselves “poison” the model and make it more vulnerable to forget-tokens\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762136511458}, {"id": "Mlf51nfgIo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1191/Reviewer_p1EL"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes reframing LLM unlearning as a backdoor attack-defense process, showing that current unlearning methods make models fragile when forget-tokens appear in retain queries. To improve robustness, the authors introduce Random Noise Augmentation (RNA), which injects small Gaussian noise into retain representations during fine-tuning. RNA is lightweight, model-agnostic, and theoretically grounded. Experiments on several 7B–8B LLMs demonstrate significant improvements in robustness without sacrificing forget or retain performance.", "review_text": "This paper proposes reframing LLM unlearning as a backdoor attack-defense process, showing that current unlearning methods make models fragile when forget-tokens appear in retain queries. To improve robustness, the authors introduce Random Noise Augmentation (RNA), which injects small Gaussian noise into retain representations during fine-tuning. RNA is lightweight, model-agnostic, and theoretically grounded. Experiments on several 7B–8B LLMs demonstrate significant improvements in robustness without sacrificing forget or retain performance.", "strengths": "- This paper recasts unlearning as a backdoor attack/defense mechanism is original and intuitive. It clarifies why unlearned models “misbehave” when seeing forgotten tokens, they’re activating an unintended backdoor.\n\n- This paper also has a clear theoretical grounding, for example, analytical results (Eqns 9 and 13) give interpretable relationships between robustness, noise variance, and token perturbation.\n\n- Experiments also show substantial accuracy recovery on perturbed MMLU queries.", "weaknesses": "- Limited experimental diversity. The experiments focus mainly on the WMDP benchmark and a single model scale. It would strengthen the paper to include additional benchmarks such as MUSE [1] and to evaluate across both larger and smaller model sizes to assess generality.\n\n- Synthetic perturbation design. The way retain sets are perturbed (e.g., replacing one token with “SARS-CoV-2”) feels somewhat artificial compared to real-world mixed-context prompts. Including datasets like MUSE could help verify whether the proposed method remains effective under more natural conditions.\n\n- Terminology clarity. The paper’s use of “unlearning robustness” could be better defined. Traditionally, robustness in unlearning refers to resistance against relearning or recovery of forgotten knowledge, whereas this work mainly studies how retain-set performance degrades when forget-tokens appear. Clearer differentiation would help readers understand the scope.\n\n- Missing robustness evaluations. It would be valuable to test the method against relearning and jailbreaking attacks to more comprehensively assess robustness.\n\n> [1] Shi, Weijia, et al. \"Muse: Machine unlearning six-way evaluation for language models.\" arXiv preprint arXiv:2407.06460 (2024).", "questions": "Please refer to the weaknesses section. I would be willing to raise my score if these issues are adequately addressed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes reframing LLM unlearning as a backdoor attack-defense process, showing that current unlearning methods make models fragile when forget-tokens appear in retain queries. To improve robustness, the authors introduce Random Noise Augmentation (RNA), which injects small Gaussian noise into retain representations during fine-tuning. RNA is lightweight, model-agnostic, and theoretically grounded. Experiments on several 7B–8B LLMs demonstrate significant improvements in robustness without sacrificing forget or retain performance.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This paper recasts unlearning as a backdoor attack/defense mechanism is original and intuitive. It clarifies why unlearned models “misbehave” when seeing forgotten tokens, they’re activating an unintended backdoor.\n\n- This paper also has a clear theoretical grounding, for example, analytical results (Eqns 9 and 13) give interpretable relationships between robustness, noise variance, and token perturbation.\n\n- Experiments also show substantial accuracy recovery on perturbed MMLU queries.", "weaknesses": "- Limited experimental diversity. The experiments focus mainly on the WMDP benchmark and a single model scale. It would strengthen the paper to include additional benchmarks such as MUSE [1] and to evaluate across both larger and smaller model sizes to assess generality.\n\n- Synthetic perturbation design. The way retain sets are perturbed (e.g., replacing one token with “SARS-CoV-2”) feels somewhat artificial compared to real-world mixed-context prompts. Including datasets like MUSE could help verify whether the proposed method remains effective under more natural conditions.\n\n- Terminology clarity. The paper’s use of “unlearning robustness” could be better defined. Traditionally, robustness in unlearning refers to resistance against relearning or recovery of forgotten knowledge, whereas this work mainly studies how retain-set performance degrades when forget-tokens appear. Clearer differentiation would help readers understand the scope.\n\n- Missing robustness evaluations. It would be valuable to test the method against relearning and jailbreaking attacks to more comprehensively assess robustness.\n\n> [1] Shi, Weijia, et al. \"Muse: Machine unlearning six-way evaluation for language models.\" arXiv preprint arXiv:2407.06460 (2024).", "questions": "Please refer to the weaknesses section. I would be willing to raise my score if these issues are adequately addressed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761927616573}, {"id": "q2x5xxkg1W", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1191/Reviewer_wLnL"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper propose Random Noise Augmentation, a method to avoid the misbehave when forget-tokens are in the retain-queries. The author frames current unlearning methods as inadvertently introducing a backdoor attack and proposes RNA as a backdoor defense to enhance the robustness of unlearned models against non-adversarial forget-tokens in retain-queries. By adding random perturbations in the data retaining stage, RNA significantly improves the model's robustness. The theoretical analysis of this paper focuses on how the presence of forget-tokens introduces randomness in latent representations, and proves a bounded probability for RNA to reject the misbehavior caused by forget-tokens.", "review_text": "This paper propose Random Noise Augmentation, a method to avoid the misbehave when forget-tokens are in the retain-queries. The author frames current unlearning methods as inadvertently introducing a backdoor attack and proposes RNA as a backdoor defense to enhance the robustness of unlearned models against non-adversarial forget-tokens in retain-queries. By adding random perturbations in the data retaining stage, RNA significantly improves the model's robustness. The theoretical analysis of this paper focuses on how the presence of forget-tokens introduces randomness in latent representations, and proves a bounded probability for RNA to reject the misbehavior caused by forget-tokens.", "strengths": "1. The paper establishes a unified view of the two primary classes of LLM unlearning methods, Representation Misdirection and Preference Optimization, by analyzing both through the lens of the generative latent variable model.\n\n2. The paper provides theoretical guarantees for RNA's effectiveness in improving the robustness of models by rejecting the detrimental effects caused by forget-tokens.\n\n3. Comprehensive experiments are conducted. This extensive evaluation rigorously validates the effectiveness and generalizability of the proposed RNA method by including diverse unlearning methods and demonstrating generalization across three different LLMs. The rigorous testing further includes detailed ablations on the RNA mechanism itself, varying the crucial noise scale and the inner model layers for injection.", "weaknesses": "1. The paper's theoretical guarantees are oversimplified approximations because they rely heavily on Assumption 4.1, treating the perturbation $\\epsilon$ as simple, independent Gaussian noise. This ignores the reality that the actual perturbation is deterministic, complex, and highly dependent on the model's parameters and context, a complexity not fully addressed by the mere \"Gaussian-like\" appearance of the empirical activation differences in Figure 9.  The oversimplification is further evidenced by the experimental results in Table 6. If the influence of unlearning could indeed be modeled as Gaussian noise, then by the same logic, adversarial perturbations would exhibit similar characteristics, and the RNA should yield consistent improvements in adversarial robustness. However, the results in Table 6 do not support this expectation.\n\n2. The proposed method requires manipulation of the model's inner layers, which may be impractical for real-world scenarios where users cannot modify these layers to inject random noise. This is an unavoidable weakness, although it is discussed in the appendix.\n\n3. Since the paper does not explicitly report running multiple independent trials using different random seeds or provide measures of variance (e.g., standard deviation), the reported performance figures are susceptible to random initialization effects and sampling randomness.", "questions": "1. See weakness 1.\n\n2. How to determine which perturbation is added to the retain-query. Since there can be many key-words in the forget set. \n\n3. It is interesting that the choice of layer significantly affects the effectiveness of the proposed method. In practical settings, should the layer be selected by some principles or determined through grid search?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose Random Noise Augmentation, a method to avoid the misbehave when forget-tokens are in the retain-queries. The author frames current unlearning methods as inadvertently introducing a backdoor attack and proposes RNA as a backdoor defense to enhance the robustness of unlearned models against non-adversarial forget-tokens in retain-queries. By adding random perturbations in the data retaining stage, RNA significantly improves the model's robustness. The theoretical analysis of this paper focuses on how the presence of forget-tokens introduces randomness in latent representations, and proves a bounded probability for RNA to reject the misbehavior caused by forget-tokens.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper establishes a unified view of the two primary classes of LLM unlearning methods, Representation Misdirection and Preference Optimization, by analyzing both through the lens of the generative latent variable model.\n\n2. The paper provides theoretical guarantees for RNA's effectiveness in improving the robustness of models by rejecting the detrimental effects caused by forget-tokens.\n\n3. Comprehensive experiments are conducted. This extensive evaluation rigorously validates the effectiveness and generalizability of the proposed RNA method by including diverse unlearning methods and demonstrating generalization across three different LLMs. The rigorous testing further includes detailed ablations on the RNA mechanism itself, varying the crucial noise scale and the inner model layers for injection.", "weaknesses": "1. The paper's theoretical guarantees are oversimplified approximations because they rely heavily on Assumption 4.1, treating the perturbation $\\epsilon$ as simple, independent Gaussian noise. This ignores the reality that the actual perturbation is deterministic, complex, and highly dependent on the model's parameters and context, a complexity not fully addressed by the mere \"Gaussian-like\" appearance of the empirical activation differences in Figure 9.  The oversimplification is further evidenced by the experimental results in Table 6. If the influence of unlearning could indeed be modeled as Gaussian noise, then by the same logic, adversarial perturbations would exhibit similar characteristics, and the RNA should yield consistent improvements in adversarial robustness. However, the results in Table 6 do not support this expectation.\n\n2. The proposed method requires manipulation of the model's inner layers, which may be impractical for real-world scenarios where users cannot modify these layers to inject random noise. This is an unavoidable weakness, although it is discussed in the appendix.\n\n3. Since the paper does not explicitly report running multiple independent trials using different random seeds or provide measures of variance (e.g., standard deviation), the reported performance figures are susceptible to random initialization effects and sampling randomness.", "questions": "1. See weakness 1.\n\n2. How to determine which perturbation is added to the retain-query. Since there can be many key-words in the forget set. \n\n3. It is interesting that the choice of layer significantly affects the effectiveness of the proposed method. In practical settings, should the layer be selected by some principles or determined through grid search?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761780008183}, {"id": "6bfJOj9UAY", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1191/Reviewer_1B8L"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper studies why current large language model unlearning methods often weaken model robustness, leading to errors when retain queries accidentally include forget tokens. The authors show that the unlearning process behaves like a backdoor attack in which forget tokens act as triggers that reintroduce forgotten behavior. To mitigate this issue, the paper reframes the retaining process as a backdoor defense and proposes Random Noise Augmentation, a simple and model agnostic technique that injects small Gaussian noise into retain representations during fine tuning. Theoretical and empirical results demonstrate that this method improves the robustness of unlearned models while maintaining their forgetting and retaining performance.", "review_text": "This paper studies why current large language model unlearning methods often weaken model robustness, leading to errors when retain queries accidentally include forget tokens. The authors show that the unlearning process behaves like a backdoor attack in which forget tokens act as triggers that reintroduce forgotten behavior. To mitigate this issue, the paper reframes the retaining process as a backdoor defense and proposes Random Noise Augmentation, a simple and model agnostic technique that injects small Gaussian noise into retain representations during fine tuning. Theoretical and empirical results demonstrate that this method improves the robustness of unlearned models while maintaining their forgetting and retaining performance.", "strengths": "1. The paper introduces a fresh angle by connecting LLM unlearning with backdoor attack and defense mechanisms, offering a clear framework for understanding existing weaknesses. \n\n2. Theoretical analysis provides partial support for the proposed method, though it relies on many assumptions that weaken its rigor. \n\n3. The writing is clear, and the figures are well-presented.", "weaknesses": "1. The experiments lack sufficient baselines. The authors only study a few non-robust LLM unlearning methods, while many recent approaches have been specifically proposed to enhance robustness in LLM unlearning [1-2].\n2. The experimental evaluation is narrow, focusing only on the WMDP benchmark. How does the proposed method perform on other benchmarks such as TOFU [3] and MUSE[4]?\n3. Using only MMLU to evaluate model utility is limited. It would be more convincing to include additional utility metrics[5].\n4. The paper does not discuss how the proposed method behaves under other types of attacks, such as relearning attacks or adversarial prompts [6].\n5. Many proofs rely on overly strong assumptions. For example, Eq. (9) asserts that adding Gaussian perturbations increases the expected loss, which depends on a strong assumption of local convexity or a positive-definite Hessian. In deep LLM latent spaces, this condition often fails, so the inequality $\\mathbb{E}[\\ell(y|z{+}v)]>\\ell(y|z)$ is not generally guaranteed.\n\n> [1] Tamirisa R, Bharathi B, Phan L, et al. Tamper-resistant safeguards for open-weight llms[J]. arXiv preprint arXiv:2408.00761, 2024.\n> \n> [2] Fan, Chongyu, et al. \"Towards llm unlearning resilient to relearning attacks: A sharpness-aware minimization perspective and beyond.\" arXiv preprint arXiv:2502.05374 (2025).\n> \n> [3] Maini, Pratyush, et al. \"Tofu: A task of fictitious unlearning for llms.\" arXiv preprint arXiv:2401.06121 (2024).\n>\n> [4] Shi, Weijia, et al. \"Muse: Machine unlearning six-way evaluation for language models.\" arXiv preprint arXiv:2407.06460 (2024).\n>\n> [5] Che Z, Casper S, Kirk R, et al. Model tampering attacks enable more rigorous evaluations of llm capabilities[J]. arXiv preprint arXiv:2502.05209, 2025.\n>\n> [6] Łucki, Jakub, et al. \"An adversarial perspective on machine unlearning for ai safety.\" arXiv preprint arXiv:2409.18025 (2024).", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies why current large language model unlearning methods often weaken model robustness, leading to errors when retain queries accidentally include forget tokens. The authors show that the unlearning process behaves like a backdoor attack in which forget tokens act as triggers that reintroduce forgotten behavior. To mitigate this issue, the paper reframes the retaining process as a backdoor defense and proposes Random Noise Augmentation, a simple and model agnostic technique that injects small Gaussian noise into retain representations during fine tuning. Theoretical and empirical results demonstrate that this method improves the robustness of unlearned models while maintaining their forgetting and retaining performance.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper introduces a fresh angle by connecting LLM unlearning with backdoor attack and defense mechanisms, offering a clear framework for understanding existing weaknesses. \n\n2. Theoretical analysis provides partial support for the proposed method, though it relies on many assumptions that weaken its rigor. \n\n3. The writing is clear, and the figures are well-presented.", "weaknesses": "1. The experiments lack sufficient baselines. The authors only study a few non-robust LLM unlearning methods, while many recent approaches have been specifically proposed to enhance robustness in LLM unlearning [1-2].\n2. The experimental evaluation is narrow, focusing only on the WMDP benchmark. How does the proposed method perform on other benchmarks such as TOFU [3] and MUSE[4]?\n3. Using only MMLU to evaluate model utility is limited. It would be more convincing to include additional utility metrics[5].\n4. The paper does not discuss how the proposed method behaves under other types of attacks, such as relearning attacks or adversarial prompts [6].\n5. Many proofs rely on overly strong assumptions. For example, Eq. (9) asserts that adding Gaussian perturbations increases the expected loss, which depends on a strong assumption of local convexity or a positive-definite Hessian. In deep LLM latent spaces, this condition often fails, so the inequality $\\mathbb{E}[\\ell(y|z{+}v)]>\\ell(y|z)$ is not generally guaranteed.\n\n> [1] Tamirisa R, Bharathi B, Phan L, et al. Tamper-resistant safeguards for open-weight llms[J]. arXiv preprint arXiv:2408.00761, 2024.\n> \n> [2] Fan, Chongyu, et al. \"Towards llm unlearning resilient to relearning attacks: A sharpness-aware minimization perspective and beyond.\" arXiv preprint arXiv:2502.05374 (2025).\n> \n> [3] Maini, Pratyush, et al. \"Tofu: A task of fictitious unlearning for llms.\" arXiv preprint arXiv:2401.06121 (2024).\n>\n> [4] Shi, Weijia, et al. \"Muse: Machine unlearning six-way evaluation for language models.\" arXiv preprint arXiv:2407.06460 (2024).\n>\n> [5] Che Z, Casper S, Kirk R, et al. Model tampering attacks enable more rigorous evaluations of llm capabilities[J]. arXiv preprint arXiv:2502.05209, 2025.\n>\n> [6] Łucki, Jakub, et al. \"An adversarial perspective on machine unlearning for ai safety.\" arXiv preprint arXiv:2409.18025 (2024).", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761329094249}], "openreview_url": "https://openreview.net/forum?id=JtX6oaaJ2d", "arxiv_id": "2501.19202", "paper_pdf": "papers/JtX6oaaJ2d.pdf", "paper_pdf_sha256": "047638bbcecfff54980a5e5c942fd1cddb60e523ef9129df9c8399213c0fa4d7", "paper_pdf_bytes": 7340032, "paper_pdf_source": "openreview", "code_url": "https://github.com/RebelsNLU-jaist/llmu-robustness", "code_repository": "RebelsNLU-jaist/llmu-robustness", "code_commit": "cdda05906725da2f6c1b7d5792eee7d780811541", "code_archive": "repos/JtX6oaaJ2d.zip", "code_archive_sha256": "2dbe319e41d72dbf0ff8662f7b10e0361be5d485a9612997ec042818202674a3", "code_archive_bytes": 25397, "code_file_count": 14, "code_extensions": {".py": 8, ".sh": 6}, "github_disk_usage_kb": 16, "github_languages": {"Python": 48251, "Shell": 6200}, "github_archived": false, "github_pushed_at": "2026-06-02T05:43:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-the-robustness-of-representation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "22ywev7zMt", "year": 2025, "status": "rejected", "title": "On the Out-of-Distribution Generalization of Self-Supervised Learning", "authors": ["Wenwen Qiang", "Jingyao Wang", "Zeen Song", "Jiangmeng Li", "Changwen Zheng"], "authorids": ["~Wenwen_Qiang1", "~Jingyao_Wang1", "~Zeen_Song1", "~Jiangmeng_Li1", "~Changwen_Zheng1"], "authors_source": "OpenReview API", "abstract": "In this paper, we focus on the out-of-distribution (OOD) generalization of self-supervised learning (SSL). By analyzing the mini-batch construction during SSL training phase, we first give one plausible explanation for SSL having OOD generalization. Then, from the perspective of data generation and causal inference, we analyze and conclude that SSL learns spurious correlations during the training process, which leads to a reduction in OOD generalization. To address this issue, we propose a post-intervention distribution (PID) grounded in the Structural Causal Model. PID offers a scenario where the relationships between variables are free from the influence of spurious correlations. Besides, we demonstrate that if each mini-batch during SSL training satisfies PID, the resulting SSL model can achieve optimal worst-case OOD performance. This motivates us to develop a batch sampling strategy that enforces PID constraints through the learning of a latent variable model. Through theoretical analysis, we demonstrate the identifiability of the latent variable model and validate the effectiveness of the proposed sampling strategy. Experiments conducted on various downstream OOD tasks demonstrate the effectiveness of the proposed sampling strategy.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "GcA3e5dnN5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission700/Reviewer_M58z"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper inspects SSL from a causal perspective, which assumes a SCM for generating augmentations in both generative and discriminative approaches. To address spurious correlations between images and their non-semantic features, e.g., backgrounds and styles, the paper proposes rebalancing the training batches by sampling to decorrelate images from their non-semantic features.  Experiments show enhanced performances across various existing SSL methods.", "review_text": "This paper inspects SSL from a causal perspective, which assumes a SCM for generating augmentations in both generative and discriminative approaches. To address spurious correlations between images and their non-semantic features, e.g., backgrounds and styles, the paper proposes rebalancing the training batches by sampling to decorrelate images from their non-semantic features.  Experiments show enhanced performances across various existing SSL methods.", "strengths": "1. The proposed rebalancing technique can be embedded into general SSL procedures, whether discriminative or generative, allowing for wide applicability.\n\n2. Experiments are extensive in scope, covering both discriminative and generative SSL (appendix). Multiple learning tasks under distribution shift is considered, including semi-supervised, transfer learning and few-shot learning. A clear improvement of around 3% in accuracy is reported for most results.", "weaknesses": "In general, I am not convinced that the proposed SCM for generating augmentation, especially the characterization of spurious correlation, is relevant for OOD generalization of SSL. \n\n1. The proposed SCM and the rebalancing strategy does not address the identifiability of spurious variables in the context of SSL. Spurious variables $s$ in supervised machine learning are the variables rejected by the conditional independence test $Y \\perp e | s$, where $e$ is the group label. However, spurious variables are generally not identifiable without labels.  Literature has introduced inductive bias, e.g., simplicity bias, to identify spurious variables for SSL [1]. However, the SCM in the paper does not consider similar assumptions to address the identifiability of $s$. For example, in Figure 1(b), $s$ and $x^{label}$ (raw image) hold symmetric roles in the SCM. Since $s$ is learned as a latent variable by variational inference, there can be infinitely many solutions of $s$.  The identifiability results in Theorem 4.3 does not resolve the identifiablity of $s$, because it depends on the condition that $p(x^+|x^{label},s)$ is learned, implying $s$ has been identified.\n2. The conditional independence implied by the SCMs may not reflect practice in SSL.  The PID SCM in Fig.3 models the statistical independence between styles or backgrounds (s) and images ($X^{label}$), but the style and background can be directly identified from the image in practice. In general, $s$ is always measurable with respect to $X^{label}$. Similarly, both $X^{label}$ and $s$ are direct causes of $X^{+}$ in Fig.2, which is also inconsistent to the augmentation practice that takes as input the raw images only, since the background is just part of the raw image. Does this paper consider a special augmentation procedure? \n3. I identify a gap between self-supervised representation learning, whose target is $p(X^{label})$, and the models used in theory. The binary classification model in Proposition 3.1 learns the density ratio $p(X^{label})/p(X^{+})$, and the \"alignment\" model in Theroem 3.4 learns $p(X^{label}|p^{+})$. The paper has not addressed that a non-spurious classification model or \"alignment\" model implies a non-spurious generative model. A simple counterexample: assume that the augmentation procedure retains the style of the image. The classifier does not depend on the style to distinguish between anchor and augmentations because they share the same style. However, styles can still be learned by the generative model.\n\nMoreover, I think this paper can be substantially improved in writing for its message to be more effectively conveyed.\n\n4. Some concepts and statements are not well defined and formulated. \n   - The \"task distribution\" is not defined. In L131-132, a statement is made that \"this framework involves estimating the true task distribution from discrete training tasks, enabling the SSL model to generalize to new, unseen tasks (i.e., test tasks).\" Is task modeled as a random variable? What does task correspond to in the SCM? For example, if task refers to batch index in Fig.2(b), then generalization is essentially impossible because training and test batch indices do not overlap. If task refers to batches of X in Fig.2(b), than generalization is only possible when the image batches are i.i.d., which is irrelevant to OOD generalization. In L157, the author states that s, denoting the style or background, does not contain any causal semantics related to the task. This statement contradicts the SCM in Fig 1 as well, where s is a direct cause of X+. Therefore, the definition of \"task\" is more vague here.\n   - What does the statement mean that \"$x^{label}$ is regarded as the label\" (L245), since $x^{label}$ is the raw image? A formulation of this equivalence may help improve clarity.\n5. Models, assumptions and theorem statements are not explicitly presented.\n   - I understand the benefits for deferring formal theorem statements to the appendix. However, the formal statement of Proposition 3.1 is missing in both the main text and the appendix. The assumption of mixture of gaussians, balanced labels, equal dimensions between spurious variables and images, and the model of binary classification are all woven into the proof.\n   - This paper models the SSL procedure by two parts: a classification model and a conditional generative alignment model. The formulation of the classification model is mixed in the proof.  The alignment model is not formulated until Theorem 3.4. However, since the learning procedure is repeatedly mentioned throughout the theory, I suggest a clear statement of the models at the beginning.\n6. Multiple notations are unexplained.\n   - The notation  $L^{PID}$ in L225 is vague because PID is a family of distributions. Which distribution is the loss evaluated with respect to? \n   - Similarly, $\\perp_{PI}$ in L217 is also unexplained. Is the independence condition satisfied for all PI distributions?\n   - nu in L313 and mu in L315, 333.\n7. The implication of the identifiability result in Theorem 4.3 is insufficiently addressed. Also related to the first point, what does the equivalance in Definition 4.2 imply for the identiability of spurious variables, and more importantly, the generative model?\n\nMinor points:\n\n8. Experiments are in relatively small scale. The results are presented for Imagenet-100 instead of the more popular ImageNet-1k. Models are trained with a limited number of epochs.\n9. There has been theoretical and empirical analysis of the vulnerability of SSL to spurious correlation, e.g., in [1]. Related work on spurious correlation in SSL can be reviewed to establish the paper's position in the broader literature. \n\n[1] Hamidieh, K., Zhang, H., Sankaranarayanan, S., & Ghassemi, M. (2024). Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation.", "questions": "1. Why does $f^\\star$ maximize the loss function in L225, since the proof indicates minimization instead?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper inspects SSL from a causal perspective, which assumes a SCM for generating augmentations in both generative and discriminative approaches. To address spurious correlations between images and their non-semantic features, e.g., backgrounds and styles, the paper proposes rebalancing the training batches by sampling to decorrelate images from their non-semantic features.  Experiments show enhanced performances across various existing SSL methods.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The proposed rebalancing technique can be embedded into general SSL procedures, whether discriminative or generative, allowing for wide applicability.\n\n2. Experiments are extensive in scope, covering both discriminative and generative SSL (appendix). Multiple learning tasks under distribution shift is considered, including semi-supervised, transfer learning and few-shot learning. A clear improvement of around 3% in accuracy is reported for most results.", "weaknesses": "In general, I am not convinced that the proposed SCM for generating augmentation, especially the characterization of spurious correlation, is relevant for OOD generalization of SSL. \n\n1. The proposed SCM and the rebalancing strategy does not address the identifiability of spurious variables in the context of SSL. Spurious variables $s$ in supervised machine learning are the variables rejected by the conditional independence test $Y \\perp e | s$, where $e$ is the group label. However, spurious variables are generally not identifiable without labels.  Literature has introduced inductive bias, e.g., simplicity bias, to identify spurious variables for SSL [1]. However, the SCM in the paper does not consider similar assumptions to address the identifiability of $s$. For example, in Figure 1(b), $s$ and $x^{label}$ (raw image) hold symmetric roles in the SCM. Since $s$ is learned as a latent variable by variational inference, there can be infinitely many solutions of $s$.  The identifiability results in Theorem 4.3 does not resolve the identifiablity of $s$, because it depends on the condition that $p(x^+|x^{label},s)$ is learned, implying $s$ has been identified.\n2. The conditional independence implied by the SCMs may not reflect practice in SSL.  The PID SCM in Fig.3 models the statistical independence between styles or backgrounds (s) and images ($X^{label}$), but the style and background can be directly identified from the image in practice. In general, $s$ is always measurable with respect to $X^{label}$. Similarly, both $X^{label}$ and $s$ are direct causes of $X^{+}$ in Fig.2, which is also inconsistent to the augmentation practice that takes as input the raw images only, since the background is just part of the raw image. Does this paper consider a special augmentation procedure? \n3. I identify a gap between self-supervised representation learning, whose target is $p(X^{label})$, and the models used in theory. The binary classification model in Proposition 3.1 learns the density ratio $p(X^{label})/p(X^{+})$, and the \"alignment\" model in Theroem 3.4 learns $p(X^{label}|p^{+})$. The paper has not addressed that a non-spurious classification model or \"alignment\" model implies a non-spurious generative model. A simple counterexample: assume that the augmentation procedure retains the style of the image. The classifier does not depend on the style to distinguish between anchor and augmentations because they share the same style. However, styles can still be learned by the generative model.\n\nMoreover, I think this paper can be substantially improved in writing for its message to be more effectively conveyed.\n\n4. Some concepts and statements are not well defined and formulated. \n   - The \"task distribution\" is not defined. In L131-132, a statement is made that \"this framework involves estimating the true task distribution from discrete training tasks, enabling the SSL model to generalize to new, unseen tasks (i.e., test tasks).\" Is task modeled as a random variable? What does task correspond to in the SCM? For example, if task refers to batch index in Fig.2(b), then generalization is essentially impossible because training and test batch indices do not overlap. If task refers to batches of X in Fig.2(b), than generalization is only possible when the image batches are i.i.d., which is irrelevant to OOD generalization. In L157, the author states that s, denoting the style or background, does not contain any causal semantics related to the task. This statement contradicts the SCM in Fig 1 as well, where s is a direct cause of X+. Therefore, the definition of \"task\" is more vague here.\n   - What does the statement mean that \"$x^{label}$ is regarded as the label\" (L245), since $x^{label}$ is the raw image? A formulation of this equivalence may help improve clarity.\n5. Models, assumptions and theorem statements are not explicitly presented.\n   - I understand the benefits for deferring formal theorem statements to the appendix. However, the formal statement of Proposition 3.1 is missing in both the main text and the appendix. The assumption of mixture of gaussians, balanced labels, equal dimensions between spurious variables and images, and the model of binary classification are all woven into the proof.\n   - This paper models the SSL procedure by two parts: a classification model and a conditional generative alignment model. The formulation of the classification model is mixed in the proof.  The alignment model is not formulated until Theorem 3.4. However, since the learning procedure is repeatedly mentioned throughout the theory, I suggest a clear statement of the models at the beginning.\n6. Multiple notations are unexplained.\n   - The notation  $L^{PID}$ in L225 is vague because PID is a family of distributions. Which distribution is the loss evaluated with respect to? \n   - Similarly, $\\perp_{PI}$ in L217 is also unexplained. Is the independence condition satisfied for all PI distributions?\n   - nu in L313 and mu in L315, 333.\n7. The implication of the identifiability result in Theorem 4.3 is insufficiently addressed. Also related to the first point, what does the equivalance in Definition 4.2 imply for the identiability of spurious variables, and more importantly, the generative model?\n\nMinor points:\n\n8. Experiments are in relatively small scale. The results are presented for Imagenet-100 instead of the more popular ImageNet-1k. Models are trained with a limited number of epochs.\n9. There has been theoretical and empirical analysis of the vulnerability of SSL to spurious correlation, e.g., in [1]. Related work on spurious correlation in SSL can be reviewed to establish the paper's position in the broader literature. \n\n[1] Hamidieh, K., Zhang, H., Sankaranarayanan, S., & Ghassemi, M. (2024). Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation.", "questions": "1. Why does $f^\\star$ maximize the loss function in L225, since the proof indicates minimization instead?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730799596662}, {"id": "NX2ESN2rFk", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission700/Reviewer_priS"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces a training batch sampling strategy designed to enhance self-supervised learning and improve generalization beyond the training distribution. The approach is inspired by the concept of invariant causal structure across different environments: while causal relationships between features and labels remain consistent, spurious correlations vary across environments. The proposed methodology employs a constraint, PID, during mini-batch sampling, which disregards spurious correlations and supports out-of-distribution generalization.", "review_text": "This paper introduces a training batch sampling strategy designed to enhance self-supervised learning and improve generalization beyond the training distribution. The approach is inspired by the concept of invariant causal structure across different environments: while causal relationships between features and labels remain consistent, spurious correlations vary across environments. The proposed methodology employs a constraint, PID, during mini-batch sampling, which disregards spurious correlations and supports out-of-distribution generalization.", "strengths": "The main strength of the paper lies", "weaknesses": "At times, the presentation is overly technical or abstract, which might be challenging for practitioners who seek to grasp the main insights of the paper. The core message is to introduce a sampling strategy combined with a distributional constraint (PID) that encourages the self-supervised method to disregard correlations that change across domains and focus on stable, causal correlations. The objective is to enhance out-of-distribution generalization by learning these invariant structures. Adding a non-technical explanation, perhaps as a remark, on how the algorithm achieves PID enforcement would be beneficial. Please refer to my questions below for further clarification.", "questions": "There are also a few concerns/typos that need to be taken care of for better readability: \n\n1. In Algorithm 1, the steps, especially the count of i, seem to be a bit confusing. It may be better to write: \"Set $i \\leftarrow 0$\", and select the initial pair $(x_0^+, x_0^{\\rm label})$. Then, for $i \\ge 1$, write the two steps and finally add, \"Set $i \\leftarrow i + 1$\". \n\n2. The number of samples mu should be $\\mu$, I guess? \n\n3. It seems that the definition of PID is the same as assuming $x^{\\rm label}$ and $s$ are independent. Maybe it would be easier to present it that way. \n\n4. What is $\\mathcal{L}^{\\rm PID}$? Is it $\\mathcal{L}^{\\rm e}, e \\in \\mathcal{D}$ where $e$ satisfies PID? How is $f$ related to $F$ in Equation (1)? Is $f$ a generic function in the class of hypothesis and $F$ the true generating function? \n\n5. Are we assuming that minimizer $f^*$ is same for all distributions in $\\mathcal{D}$ that satisfies PID? \n\n6. I am a little bit confused about Assumption 3.3. In PID, we have $x^{\\rm label}$ is independent of $s$, whereas in Assumption 3.3, we also have $x^{\\rm label}$ is independent of $s$ given $x^+$. Are we assuming Assumption 3.3 for all distributions in $\\mathcal{D}$? A remark with some intuitive explanation of Theorem 3.4 would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a training batch sampling strategy designed to enhance self-supervised learning and improve generalization beyond the training distribution. The approach is inspired by the concept of invariant causal structure across different environments: while causal relationships between features and labels remain consistent, spurious correlations vary across environments. The proposed methodology employs a constraint, PID, during mini-batch sampling, which disregards spurious correlations and supports out-of-distribution generalization.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The main strength of the paper lies", "weaknesses": "At times, the presentation is overly technical or abstract, which might be challenging for practitioners who seek to grasp the main insights of the paper. The core message is to introduce a sampling strategy combined with a distributional constraint (PID) that encourages the self-supervised method to disregard correlations that change across domains and focus on stable, causal correlations. The objective is to enhance out-of-distribution generalization by learning these invariant structures. Adding a non-technical explanation, perhaps as a remark, on how the algorithm achieves PID enforcement would be beneficial. Please refer to my questions below for further clarification.", "questions": "There are also a few concerns/typos that need to be taken care of for better readability: \n\n1. In Algorithm 1, the steps, especially the count of i, seem to be a bit confusing. It may be better to write: \"Set $i \\leftarrow 0$\", and select the initial pair $(x_0^+, x_0^{\\rm label})$. Then, for $i \\ge 1$, write the two steps and finally add, \"Set $i \\leftarrow i + 1$\". \n\n2. The number of samples mu should be $\\mu$, I guess? \n\n3. It seems that the definition of PID is the same as assuming $x^{\\rm label}$ and $s$ are independent. Maybe it would be easier to present it that way. \n\n4. What is $\\mathcal{L}^{\\rm PID}$? Is it $\\mathcal{L}^{\\rm e}, e \\in \\mathcal{D}$ where $e$ satisfies PID? How is $f$ related to $F$ in Equation (1)? Is $f$ a generic function in the class of hypothesis and $F$ the true generating function? \n\n5. Are we assuming that minimizer $f^*$ is same for all distributions in $\\mathcal{D}$ that satisfies PID? \n\n6. I am a little bit confused about Assumption 3.3. In PID, we have $x^{\\rm label}$ is independent of $s$, whereas in Assumption 3.3, we also have $x^{\\rm label}$ is independent of $s$ given $x^+$. Are we assuming Assumption 3.3 for all distributions in $\\mathcal{D}$? A remark with some intuitive explanation of Theorem 3.4 would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730607364736}, {"id": "qQMlMtoaF9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission700/Reviewer_fPhX"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper regards the mini-batches in the SSL training as the environments (domains) in OOD generalization problems and proposes that each mini-batch can be viewed as a multi-class classification task. Based on this formulation, the authors points out that when the similarity is measured using non-causal features, SSL will learn spurious representations. To address this issue, the authors propose to model the Post-Intervention Distribution (PID) using VAEs for each mini-batch and further propose a mini-batch sampling strategy that selects samples with similar balancing scores based on the  $p^e(s|x^{label})$ learned by the VAE. The experiments demonstrat the effectiveness of the method.", "review_text": "This paper regards the mini-batches in the SSL training as the environments (domains) in OOD generalization problems and proposes that each mini-batch can be viewed as a multi-class classification task. Based on this formulation, the authors points out that when the similarity is measured using non-causal features, SSL will learn spurious representations. To address this issue, the authors propose to model the Post-Intervention Distribution (PID) using VAEs for each mini-batch and further propose a mini-batch sampling strategy that selects samples with similar balancing scores based on the  $p^e(s|x^{label})$ learned by the VAE. The experiments demonstrat the effectiveness of the method.", "strengths": "1. The perspective of converting the SSL training to a domain-generalization-like problem using an SCM is natrual and interesting.\n2. The proposed method is built with theoretical guarantees on the identifiability of the distribution parameters and the recover of the PID.\n3. The experiments cover many scenarios, including semi-supervised learning, transfer learning, and few-shot learning tasks. The improvements of the proposed method are significant.", "weaknesses": "1. Some points are not quite clear and need further clarification. For example, despite Theorem 4.7, it is a bit confusing that why sampling samples with the same propensity score would help to recover $p^{PI}$. It would be better to provide some high-level explanations.\n2. The authors didn't evaluate their method on classic OOD tasks like PACS, OfficeHome, ColoredMNIST, etc. Since this work aims to improve SSL's OOD performance, it would be necessary to evaluate these tasks. Otherwise, the author should explain why not doing so.", "questions": "1. Could you shed light on why using an exponential family distribution to model $p(s|x^{label})$?\n2. In line 227, how does Theorem 3.4 \"implies that when $\\mathcal{D}$ is sufficiently large and diverse, an optimal $f^*$ trained on one distribution will perform worse than random guessing in some other environment.\"?\n3. In line 459, why \"We can observe that the performance of BYOL rapidly deteriorates with batch size.\"? It seems that BYOL suffers from smaller performance degradation than BYOL+ours.\n4. In line 267, should it be $T_{ij}=a_{ij}\\times \\cdot$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper regards the mini-batches in the SSL training as the environments (domains) in OOD generalization problems and proposes that each mini-batch can be viewed as a multi-class classification task. Based on this formulation, the authors points out that when the similarity is measured using non-causal features, SSL will learn spurious representations. To address this issue, the authors propose to model the Post-Intervention Distribution (PID) using VAEs for each mini-batch and further propose a mini-batch sampling strategy that selects samples with similar balancing scores based on the  $p^e(s|x^{label})$ learned by the VAE. The experiments demonstrat the effectiveness of the method.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The perspective of converting the SSL training to a domain-generalization-like problem using an SCM is natrual and interesting.\n2. The proposed method is built with theoretical guarantees on the identifiability of the distribution parameters and the recover of the PID.\n3. The experiments cover many scenarios, including semi-supervised learning, transfer learning, and few-shot learning tasks. The improvements of the proposed method are significant.", "weaknesses": "1. Some points are not quite clear and need further clarification. For example, despite Theorem 4.7, it is a bit confusing that why sampling samples with the same propensity score would help to recover $p^{PI}$. It would be better to provide some high-level explanations.\n2. The authors didn't evaluate their method on classic OOD tasks like PACS, OfficeHome, ColoredMNIST, etc. Since this work aims to improve SSL's OOD performance, it would be necessary to evaluate these tasks. Otherwise, the author should explain why not doing so.", "questions": "1. Could you shed light on why using an exponential family distribution to model $p(s|x^{label})$?\n2. In line 227, how does Theorem 3.4 \"implies that when $\\mathcal{D}$ is sufficiently large and diverse, an optimal $f^*$ trained on one distribution will perform worse than random guessing in some other environment.\"?\n3. In line 459, why \"We can observe that the performance of BYOL rapidly deteriorates with batch size.\"? It seems that BYOL suffers from smaller performance degradation than BYOL+ours.\n4. In line 267, should it be $T_{ij}=a_{ij}\\times \\cdot$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730194361955}], "openreview_url": "https://openreview.net/forum?id=22ywev7zMt", "arxiv_id": "2505.16675", "paper_pdf": "papers/22ywev7zMt.pdf", "paper_pdf_sha256": "80507b1b63d95900be2c420895a12d96d9270abac2fa06b894015b0b6cb03d1f", "paper_pdf_bytes": 2564459, "paper_pdf_source": "openreview", "code_url": "https://github.com/ML-TASA/PID-SSL", "code_repository": "ML-TASA/PID-SSL", "code_commit": "095b2be7fcfe206cbf2a103039a0529d12968846", "code_archive": "repos/22ywev7zMt.zip", "code_archive_sha256": "fd1087b1fecbe86c4aeab52c00f3c95107d8f1fb6b59b21d37b5b15360c84b76", "code_archive_bytes": 7332, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 10, "github_languages": {"Python": 18876}, "github_archived": false, "github_pushed_at": "2025-06-04T08:40:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-out-of-distribution-generalization-of-2"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "j8s-BRxXST", "year": 2023, "status": "rejected", "title": "A Simple Contrastive Learning Objective for Alleviating Neural Text Degeneration", "authors": ["Shaojie Jiang", "Ruqing Zhang", "Svitlana Vakulenko", "Maarten de Rijke"], "authorids": ["~Shaojie_Jiang1", "~Ruqing_Zhang3", "~Svitlana_Vakulenko1", "~Maarten_de_Rijke1"], "authors_source": "OpenReview API", "abstract": "The cross-entropy objective has proved to be an all-purpose training objective for autoregressive language models (LMs). However, without distinguishing problematic tokens, LMs trained using cross-entropy exhibit text degeneration problems. To address this, unlikelihood training has been proposed to reduce the probability of unlikely tokens predicted by LMs. But unlikelihood does not explicitly consider the relationship between the label tokens and unlikely token candidates, thus showing marginal improvements in degeneration. We propose a new contrastive token learning objective that inherits the advantages of cross-entropy and unlikelihood training and avoids their limitations. The key idea is to teach a LM to generate high probabilities for label tokens and low probabilities for negative candidates. Comprehensive experiments on language modeling and open-domain dialogue generation tasks show that the proposed contrastive token objective yields much less repetitive texts, with a higher generation quality than baseline approaches, achieving the new state-of-the-art performance on text degeneration.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SOFURJfNbbH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6378/Reviewer_TkXN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a contrastive learning method to balance the learning of positive and negative tokens in text generation tasks (e.g., language modeling and open-domain dialogue generation tasks).", "review_text": "Given the thin novelty and insufficient experiments, I suggest rejecting this paper.", "strengths": "**Strength**  \n1. The paper is easy to follow and the idea is intuitive.\n2. Several case studies are given which are encouraged. \n\n**Weaknesses**\n1. The novelty is rather thin. This paper is an incremental work on UL: 1) The core idea that penalizes the previously generated tokens has been proposed by UL; 2) I **totally disagree** with the author's claim that `Comparing Eq.(5) to Eq. (4), we see that UL only\nconsiders the probabilities of negative tokens`, Equation 4 of the UL paper clearly shows that UL has jointly considered the probabilities of positive and negative tokens (i.e., a likelihood term for a positive token and an unlikelihood term for negative tokens). However, the authors try to hide this important detail and only write the likelihood term of negative tokens in Equation 4 of the current paper.\n2. The experiment is insufficient. The authors have cited many machine translation papers and also state that `We performed experiments on fine-tuning LMs for reducing their repetition rates, which can be beneficial for related tasks such as abstractive summarization, machine translation, and image captioning.` Therefore, why not simply perform the proposed method in these tasks? PPL is not a reliable metric for evaluating text generation tasks.  The metrics (especially the recently proposed neural metrics ) in abstractive summarization, machine translation, and image captioning are more competent to give a more reasonable evaluation, which can make the proposed method more convincing. A small suggestion: if you think your proposed method is simple and still can be accepted by a top-tier conference, the method has to be very powerful or very general. Unfortunately, the proposed method does not show its effectiveness. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a contrastive learning method to balance the learning of positive and negative tokens in text generation tasks (e.g., language modeling and open-domain dialogue generation tasks).", "strength_and_weaknesses": "**Strength**  \n1. The paper is easy to follow and the idea is intuitive.\n2. Several case studies are given which are encouraged. \n\n**Weaknesses**\n1. The novelty is rather thin. This paper is an incremental work on UL: 1) The core idea that penalizes the previously generated tokens has been proposed by UL; 2) I **totally disagree** with the author's claim that `Comparing Eq.(5) to Eq. (4), we see that UL only\nconsiders the probabilities of negative tokens`, Equation 4 of the UL paper clearly shows that UL has jointly considered the probabilities of positive and negative tokens (i.e., a likelihood term for a positive token and an unlikelihood term for negative tokens). However, the authors try to hide this important detail and only write the likelihood term of negative tokens in Equation 4 of the current paper.\n2. The experiment is insufficient. The authors have cited many machine translation papers and also state that `We performed experiments on fine-tuning LMs for reducing their repetition rates, which can be beneficial for related tasks such as abstractive summarization, machine translation, and image captioning.` Therefore, why not simply perform the proposed method in these tasks? PPL is not a reliable metric for evaluating text generation tasks.  The metrics (especially the recently proposed neural metrics ) in abstractive summarization, machine translation, and image captioning are more competent to give a more reasonable evaluation, which can make the proposed method more convincing. A small suggestion: if you think your proposed method is simple and still can be accepted by a top-tier conference, the method has to be very powerful or very general. Unfortunately, the proposed method does not show its effectiveness. ", "clarity,_quality,_novelty_and_reproducibility": "**Clarity** \nThis paper is very clear.  \n\n**Quality**\nThis paper does not meet the bar of ICLR. The method is not novel and the experiments are pretty limited.  \n\n**Novelty**\nThin. Unlikelihood learning has made the most contributions.  \n\n**Reproducibility**\nGood. The authors have uploaded the code.  ", "summary_of_the_review": "Given the thin novelty and insufficient experiments, I suggest rejecting this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667124984162}, {"id": "CKJRJhA1HAx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6378/Reviewer_8fJN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a new loss function to train text generation models aimed at reducing repetitions commonly associated with models trained with a simple cross-entropy loss. Interpreting cross-entropy as a contrastive loss function, this work proposes to add to this loss another term similar to cross entropy which considers M tokens generated before every step at negative examples. The primary motivation behind this addition is that cross entropy treats negative tokens and irrelevant tokens equally which this work aims to fix. On experiments conducted on a fine-tuned GPT2-small and a dialogue generation system, the authors claim increased quality (by perplexity), diversity (by dist-1 and uniq-1) and reduced repeated phrases. ", "review_text": "While the method has an interesting motivation, the experimental setup seems small and not very convincing both in terms of LMs considered and the metrics reported. I am currently leaning negative but would be willing to revise my score after rebuttal.", "strengths": "Strengths:\n1. The proposed loss has the sound motivation and is very simple to implement and can be added into any generation system with little effort.\n\n2. The results show improvement on many metrics across both tasks.\n\nWeaknesses:\n1. The experimental setup is weak. The primary result provided in the main paper is on GPT2-small fine-tuned on wikitext which by today's LM standards is not a convincing setup. Further experiments need to be conducted on larger versions of GPT2 or models trained with more data same as prior work. Furthermore, the metrics used and how they are interpreted deviate from other works. In this work, perplexity is used as a criterion for quality (lower the better). Prior work (https://arxiv.org/abs/1904.09751) has shown that low perplexity in GPT2 based models is not a good indicator of quality but rather of degenerate behaviour and closeness to human text perplexity is a better measure. Additionally, most prior works report dist-2,3 and even 4 but this works only reports dist-1. Also by the definition of dist-1 it should always be less than 1 since it measures the fraction of unique n-grams in the output text. I'm confused by the values reported in the paper. \n\n2. Also unclear is the meaning of greedy or beam search in top-k (or p) sampling. Further,  it is not clear why beam search is the choice of decoding algorithm. General LMs like GPT2 (or their fine-tuned versions) and even dialogue models in principle have multiple potential continuations given a prompt while beam search gives out just one output. \n\n3. Some crucial baselines like simple ascetral sampling (that top-k with k=vocab size) and more recently proposed typical sampling (https://arxiv.org/abs/2202.00666).   ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents a new loss function to train text generation models aimed at reducing repetitions commonly associated with models trained with a simple cross-entropy loss. Interpreting cross-entropy as a contrastive loss function, this work proposes to add to this loss another term similar to cross entropy which considers M tokens generated before every step at negative examples. The primary motivation behind this addition is that cross entropy treats negative tokens and irrelevant tokens equally which this work aims to fix. On experiments conducted on a fine-tuned GPT2-small and a dialogue generation system, the authors claim increased quality (by perplexity), diversity (by dist-1 and uniq-1) and reduced repeated phrases. ", "strength_and_weaknesses": "Strengths:\n1. The proposed loss has the sound motivation and is very simple to implement and can be added into any generation system with little effort.\n\n2. The results show improvement on many metrics across both tasks.\n\nWeaknesses:\n1. The experimental setup is weak. The primary result provided in the main paper is on GPT2-small fine-tuned on wikitext which by today's LM standards is not a convincing setup. Further experiments need to be conducted on larger versions of GPT2 or models trained with more data same as prior work. Furthermore, the metrics used and how they are interpreted deviate from other works. In this work, perplexity is used as a criterion for quality (lower the better). Prior work (https://arxiv.org/abs/1904.09751) has shown that low perplexity in GPT2 based models is not a good indicator of quality but rather of degenerate behaviour and closeness to human text perplexity is a better measure. Additionally, most prior works report dist-2,3 and even 4 but this works only reports dist-1. Also by the definition of dist-1 it should always be less than 1 since it measures the fraction of unique n-grams in the output text. I'm confused by the values reported in the paper. \n\n2. Also unclear is the meaning of greedy or beam search in top-k (or p) sampling. Further,  it is not clear why beam search is the choice of decoding algorithm. General LMs like GPT2 (or their fine-tuned versions) and even dialogue models in principle have multiple potential continuations given a prompt while beam search gives out just one output. \n\n3. Some crucial baselines like simple ascetral sampling (that top-k with k=vocab size) and more recently proposed typical sampling (https://arxiv.org/abs/2202.00666).   ", "clarity,_quality,_novelty_and_reproducibility": "1. The paper is very clearly written and easy to follow\n2. Quality and novelty: see strengths and weaknesses.\n3. The results seem easily reproducible as the loss function is simple to implement. ", "summary_of_the_review": "While the method has an interesting motivation, the experimental setup seems small and not very convincing both in terms of LMs considered and the metrics reported. I am currently leaning negative but would be willing to revise my score after rebuttal.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667032450785}, {"id": "r0u0YzkA8_G", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6378/Reviewer_fiLs"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new contrastive token (CT) learning objective to teach a LM to generate high probabilities for label tokens and low probabilities for negative candidates (repetitive tokens). The idea of this paper is very similar to \"unlikelihood training\", however, as the authors showed in section 3.3, in unlikelihood training the irrelevant tokens are promoted, while in CT they remain unchanged, and the negative tokens are sometimes promoted and sometimes suppressed by the gradient function in unlikelihood training, while they are always suppressed in CT. ", "review_text": "There is not enough contribution in the paper.", "strengths": "Strength:\nThis paper has a very intuitive idea, and it is easy to follow.\nExperimental results and human evaluations are positive. \n\n\nWeaknesses:\nThere is not enough novelty in the paper. \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new contrastive token (CT) learning objective to teach a LM to generate high probabilities for label tokens and low probabilities for negative candidates (repetitive tokens). The idea of this paper is very similar to \"unlikelihood training\", however, as the authors showed in section 3.3, in unlikelihood training the irrelevant tokens are promoted, while in CT they remain unchanged, and the negative tokens are sometimes promoted and sometimes suppressed by the gradient function in unlikelihood training, while they are always suppressed in CT. ", "strength_and_weaknesses": "Strength:\nThis paper has a very intuitive idea, and it is easy to follow.\nExperimental results and human evaluations are positive. \n\n\nWeaknesses:\nThere is not enough novelty in the paper. \n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is very easy to understand and follow, but it lacks novelty. \n", "summary_of_the_review": "There is not enough contribution in the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666682109369}, {"id": "ryqpH2q_W3s", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6378/Reviewer_YrBU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Proposes a new training objective, contrastive token learning (CT), for language models that contrasts the target token with M previous tokens (as negatives). It is compared to regular cross-entropy and other baselines, and especially unlikelihood training (UL). Similar to UL it is motivated by the observation that CE training can lead to degenerate generation in the form of repetition, and it is shown that the proposed objective can significantly reduce it while keeping perplexity similar to CE. \n\nOne key idea is categorizing tokens as positive (label/target), negative (repeated), and irrelevant (others). Authors believe negative and irrelevant tokens should not be treated the same as in CE, and instead only penalize negative tokens in the contrastive oss.\n\nThe LM is evaluated on the wikitext-103 task and compared primarily with CE and UL. A human evaluation 1 vs 1 shows that CT is preferred to baselines, although it is not statistically significant .\n", "review_text": "While the proposal of CT as a new training objective is interesting, it is unclear how important the effect of using M previous tokens instead of all tokens as UL does. A comparison of (a) with CT using all previous tokens; or (b) UL with only M previous tokens would help understand the effect. Furthermore, it is unclear how important the repetition issue is for better/larger models, as these typically do not use any repetition-reducing heuristics and seem to suffer less degenerative generation; some scaling experiments could help resolve this question. Finally human evaluation doesn't seem to prefer CT over UL-TS in-spite of claims of better fluency/repetition.\n\nIn conclusion, while the paper is well-written, I don't yet see (without more experiments) the utility of the proposed technique over standard cross-entropy training.\n", "strengths": "## Strengths\n- A new training objective that is well-motivated\n- both automatic and human evaluation\n- Interesting view of cross-entropy in the lens of contrastive learning.\n- well-written\n\n## Weakness\n- For negative tokens, authors use previous M (section 3.2), whereas UL uses all previous. It is unclear how important this is. One way to measure it is to compare performance using all tokens as in UL. \n- repetition seems to be less of an issue with better models (larger and/or trained with more data), so it is unclear how useful this technique is for the best models. Experiments were done with GPT-2 small. How do results change with larger GPT2 models?\n- The human eval (Table 3) does not show any statistical significance compared to UL-TS, one of the baselines. In particular, some claims about UL-TS being less fluent (ungrammatical repetitions) does not agree with the human eval.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Proposes a new training objective, contrastive token learning (CT), for language models that contrasts the target token with M previous tokens (as negatives). It is compared to regular cross-entropy and other baselines, and especially unlikelihood training (UL). Similar to UL it is motivated by the observation that CE training can lead to degenerate generation in the form of repetition, and it is shown that the proposed objective can significantly reduce it while keeping perplexity similar to CE. \n\nOne key idea is categorizing tokens as positive (label/target), negative (repeated), and irrelevant (others). Authors believe negative and irrelevant tokens should not be treated the same as in CE, and instead only penalize negative tokens in the contrastive oss.\n\nThe LM is evaluated on the wikitext-103 task and compared primarily with CE and UL. A human evaluation 1 vs 1 shows that CT is preferred to baselines, although it is not statistically significant .\n", "strength_and_weaknesses": "## Strengths\n- A new training objective that is well-motivated\n- both automatic and human evaluation\n- Interesting view of cross-entropy in the lens of contrastive learning.\n- well-written\n\n## Weakness\n- For negative tokens, authors use previous M (section 3.2), whereas UL uses all previous. It is unclear how important this is. One way to measure it is to compare performance using all tokens as in UL. \n- repetition seems to be less of an issue with better models (larger and/or trained with more data), so it is unclear how useful this technique is for the best models. Experiments were done with GPT-2 small. How do results change with larger GPT2 models?\n- The human eval (Table 3) does not show any statistical significance compared to UL-TS, one of the baselines. In particular, some claims about UL-TS being less fluent (ungrammatical repetitions) does not agree with the human eval.\n", "clarity,_quality,_novelty_and_reproducibility": "- reproducibility: good - provides code samples, google colab (https://anonymous.4open.science/r/lit-seq)  and pip-package\n- \"Unlikelihood training also unintentionally boosts the probability of other irrelevant tokens.\" -- can you explain more?\n", "summary_of_the_review": "While the proposal of CT as a new training objective is interesting, it is unclear how important the effect of using M previous tokens instead of all tokens as UL does. A comparison of (a) with CT using all previous tokens; or (b) UL with only M previous tokens would help understand the effect. Furthermore, it is unclear how important the repetition issue is for better/larger models, as these typically do not use any repetition-reducing heuristics and seem to suffer less degenerative generation; some scaling experiments could help resolve this question. Finally human evaluation doesn't seem to prefer CT over UL-TS in-spite of claims of better fluency/repetition.\n\nIn conclusion, while the paper is well-written, I don't yet see (without more experiments) the utility of the proposed technique over standard cross-entropy training.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666552844628}], "openreview_url": "https://openreview.net/forum?id=j8s-BRxXST", "arxiv_id": "2205.02517", "paper_pdf": "papers/j8s-BRxXST.pdf", "paper_pdf_sha256": "13f30c6527fe9db4674103de3a836b928b2749fd60c3e6b8674717e66380bb95", "paper_pdf_bytes": 1753876, "paper_pdf_source": "openreview", "code_url": "https://github.com/ShaojieJiang/CT-Loss", "code_repository": "ShaojieJiang/CT-Loss", "code_commit": "21731b590839b31675e55333be289c7b6513ade3", "code_archive": "repos/j8s-BRxXST.zip", "code_archive_sha256": "38e85e3862162dbf45eedcd49ca03a36b7201b63f8314f378b6699d19395ea76", "code_archive_bytes": 8148, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 13, "github_languages": {"Python": 6005, "Mustache": 974}, "github_archived": false, "github_pushed_at": "2022-05-11T09:27:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-simple-contrastive-learning-objective-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xw04RdwI2kS", "year": 2022, "status": "rejected", "title": "Inverse Contextual Bandits: Learning How Behavior Evolves over Time", "authors": ["Alihan Hüyük", "Daniel Jarrett", "Mihaela van der Schaar"], "authorids": ["~Alihan_Hüyük1", "~Daniel_Jarrett1", "~Mihaela_van_der_Schaar2"], "authors_source": "OpenReview API", "abstract": "Understanding a decision-maker's priorities by observing their behavior is critical for transparency and accountability in decision processes—such as in healthcare. Though conventional approaches to policy learning almost invariably assume stationarity in behavior, this is hardly true in practice: Medical practice is constantly evolving as clinical professionals fine-tune their knowledge over time. For instance, as the medical community's understanding of organ transplantations has progressed over the years, a pertinent question is: How have actual organ allocation policies been evolving? To give an answer, we desire a policy learning method that provides interpretable representations of decision-making, in particular capturing an agent's non-stationary knowledge of the world, as well as operating in an offline manner. First, we model the evolving behavior of decision-makers in terms of contextual bandits, and formalize the problem of Inverse Contextual Bandits (\"ICB''). Second, we propose two concrete algorithms as solutions, learning parametric and non-parametric representations of an agent's behavior. Finally, using both real and simulated data for liver transplantations, we illustrate the applicability and explainability of our method, as well as benchmarking and validating the accuracy of our algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "arfo5TutVck", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2631/Reviewer_G5nF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies an inverse (linear) contextual bandits (ICB) problem, where, given a $T$-round realization of a bandit policy’s actions and observed rewards, the goal is to design an algorithm to estimate the underlying environment parameter, along with the “belief trajectory” of the bandit policy. A particular emphasis is placed on the belief trajectory being “interpretable” and capturing changes in the policy’s “knowledge of the world” over time.\n\nThe paper’s main contributions are (i) formalizing the inverse contextual bandits problem, (ii) designing two algorithms for this problem based on two different ways of modelling beliefs of the bandit policy, and (iii) providing empirical illustrations of how their algorithm can be used to investigate and explain changes in medical decision-making over time\n", "review_text": "**Problem definition and setup**:\n\nI found much of the content in section 2 to be unnecessarily general. Indeed, the preliminaries describe a reinforcement learning framework, but only the contextual bandits setting is considered in the paper. I found this additional notation to be distracting to the paper, and would suggest the authors introduce only background that is used in the paper. Additionally, this unneeded generality makes explanations in various parts of the paper more complicated than needed -- in particular, Remark 1 would not be needed if the RL framework were not introduced.\n\nAdditionally, I found the objectives of the paper to be rather vague. Specifically, the authors place significant emphasis on the interpretability of the output of an ICB algorithm, but do not seem to describe what interpretability means. I think it would be useful to describe some desired or defining characteristics of an interpretable output, and to discuss how the belief trajectory, as the authors have modelled it in their two algorithms, does or does not satisfy those characteristics.\n\n**Relation to related work**:\n\nI am a bit confused with how this work fits into/differs from some prior works mentioned in the paper. In particular, I was quite confused why one of the three main contributions of the paper was to formalize the ICB problem, but the IRL problem (which, as I understand it, is strictly more general) has already been formalized and studied. Could you elaborate more on how your problem setting and approaches differ from, e.g., [39] (Bayesian inverse reinforcement learning)? In what ways are you solving a different problem, or improving on previous works on IRL in your contextual bandits setting? Why should one expect the performance in Tables 3-4 of the algorithms you propose to be better than those from prior works (e.g., B-IRL)?\n\n**Experimental results**:\n\nWhile I find it interesting that, in the OPTN experiment, you are able to associate (external) policy decisions with changes in feature importances, I have a number of questions/comments about the interpretability of the algorithm output, and the applicability of the model considered in the paper.\n\nContextual bandits model: I think it would be useful to explain (i) what the rewards represent in the context of this experiment (I couldn’t find this explained anywhere, but I might have missed it) and (ii) why it is reasonable to assume that $\\rho^*$ does not vary over time. I was confused on point (ii) for a while, since I would expect most bandits algorithms to converge to a fixed $\\rho^*$. In the context of your experiments, this would correspond to the relative importance of the features converging to a single value. Could you explain why one should not expect this to happen in your experiments?\n\nExplainability of the results: I think that the paper could benefit from more explicitly outlining the desired properties of explainable results, how their method achieves this, and various ways that the method could be used. Indeed, in order for the results of Figures 3-4 to be “explainable” or “interpretable,” it seems that external context (such as that given in Figure 4) is crucial. Even then, only general trends can be seen from the figures displayed in this section. It is not clear to me how one might interpret such trends in this way when the number of features is large. Do the authors have more general ideas for interpreting these results? How might interpreting these results change when using Algorithm 1 vs Algorithm 2 in your paper? It may be useful to discuss these in the paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies an inverse (linear) contextual bandits (ICB) problem, where, given a $T$-round realization of a bandit policy’s actions and observed rewards, the goal is to design an algorithm to estimate the underlying environment parameter, along with the “belief trajectory” of the bandit policy. A particular emphasis is placed on the belief trajectory being “interpretable” and capturing changes in the policy’s “knowledge of the world” over time.\n\nThe paper’s main contributions are (i) formalizing the inverse contextual bandits problem, (ii) designing two algorithms for this problem based on two different ways of modelling beliefs of the bandit policy, and (iii) providing empirical illustrations of how their algorithm can be used to investigate and explain changes in medical decision-making over time\n", "main_review": "**Problem definition and setup**:\n\nI found much of the content in section 2 to be unnecessarily general. Indeed, the preliminaries describe a reinforcement learning framework, but only the contextual bandits setting is considered in the paper. I found this additional notation to be distracting to the paper, and would suggest the authors introduce only background that is used in the paper. Additionally, this unneeded generality makes explanations in various parts of the paper more complicated than needed -- in particular, Remark 1 would not be needed if the RL framework were not introduced.\n\nAdditionally, I found the objectives of the paper to be rather vague. Specifically, the authors place significant emphasis on the interpretability of the output of an ICB algorithm, but do not seem to describe what interpretability means. I think it would be useful to describe some desired or defining characteristics of an interpretable output, and to discuss how the belief trajectory, as the authors have modelled it in their two algorithms, does or does not satisfy those characteristics.\n\n**Relation to related work**:\n\nI am a bit confused with how this work fits into/differs from some prior works mentioned in the paper. In particular, I was quite confused why one of the three main contributions of the paper was to formalize the ICB problem, but the IRL problem (which, as I understand it, is strictly more general) has already been formalized and studied. Could you elaborate more on how your problem setting and approaches differ from, e.g., [39] (Bayesian inverse reinforcement learning)? In what ways are you solving a different problem, or improving on previous works on IRL in your contextual bandits setting? Why should one expect the performance in Tables 3-4 of the algorithms you propose to be better than those from prior works (e.g., B-IRL)?\n\n**Experimental results**:\n\nWhile I find it interesting that, in the OPTN experiment, you are able to associate (external) policy decisions with changes in feature importances, I have a number of questions/comments about the interpretability of the algorithm output, and the applicability of the model considered in the paper.\n\nContextual bandits model: I think it would be useful to explain (i) what the rewards represent in the context of this experiment (I couldn’t find this explained anywhere, but I might have missed it) and (ii) why it is reasonable to assume that $\\rho^*$ does not vary over time. I was confused on point (ii) for a while, since I would expect most bandits algorithms to converge to a fixed $\\rho^*$. In the context of your experiments, this would correspond to the relative importance of the features converging to a single value. Could you explain why one should not expect this to happen in your experiments?\n\nExplainability of the results: I think that the paper could benefit from more explicitly outlining the desired properties of explainable results, how their method achieves this, and various ways that the method could be used. Indeed, in order for the results of Figures 3-4 to be “explainable” or “interpretable,” it seems that external context (such as that given in Figure 4) is crucial. Even then, only general trends can be seen from the figures displayed in this section. It is not clear to me how one might interpret such trends in this way when the number of features is large. Do the authors have more general ideas for interpreting these results? How might interpreting these results change when using Algorithm 1 vs Algorithm 2 in your paper? It may be useful to discuss these in the paper.\n", "summary_of_the_review": "While the problem setting seems interesting, I think that there are a number of issues indicating that this work is not yet ready for publication. Indeed, the problem setting should be more concisely described, and the relation of this work to prior works on IRL should be discussed more thoroughly. Additionally, I think that the authors should more clearly outline the goals of the interpretability of their method, and outline concrete solutions others could use to interpret the belief trajectories in practice.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636367260202}, {"id": "14e9er0qoPX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2631/Reviewer_NLig"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper addresses the problem of Inverse Contextual Bandit.  They raise an important question: given demonstrated behavior\nfrom an agent, how has the agent’s knowledge been evolving over time? Formally, Given a contextual bandit problem\n$(X, A, R, T )$,  where $R, T$ are unknown to the agent. Given an observational dataset $D$, and a family of reward parameterizations $P$ and belief parameterizations $B$, the inverse contextual bandits problem is to determine the true environment parameter $\\rho^*$\nand the belief parameters $\\beta_{1:T}$. \n\nThey propose two algorithms to learn these parameters. The first uses the agent’s knowledge in terms of Bayesian update. The second uses the Gaussian process. They demonstrate their algorithm through simulated and real-world data for liver transplantations.\n", "review_text": "Strengths: the considered problem is significant. Their model gains many important properties compared to related works. As they mentioned,  this is the first formal attempt at learning interpretable representations of nonstationary behavior. Using Bayesian methods to compute posteriors is valid. In addition, the experiments for liver transplantations are interesting.\n\nWeaknesses: The presentation is hard to follow. It should be improved.  There is no new technique in this paper and there are no theoretical guarantees for their proposed algorithms. It would be interesting if the authors can upper bound the error $|\\rho^* - \\hat{\\rho}^*|$  to show that their algorithms converge.   ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper addresses the problem of Inverse Contextual Bandit.  They raise an important question: given demonstrated behavior\nfrom an agent, how has the agent’s knowledge been evolving over time? Formally, Given a contextual bandit problem\n$(X, A, R, T )$,  where $R, T$ are unknown to the agent. Given an observational dataset $D$, and a family of reward parameterizations $P$ and belief parameterizations $B$, the inverse contextual bandits problem is to determine the true environment parameter $\\rho^*$\nand the belief parameters $\\beta_{1:T}$. \n\nThey propose two algorithms to learn these parameters. The first uses the agent’s knowledge in terms of Bayesian update. The second uses the Gaussian process. They demonstrate their algorithm through simulated and real-world data for liver transplantations.\n", "main_review": "Strengths: the considered problem is significant. Their model gains many important properties compared to related works. As they mentioned,  this is the first formal attempt at learning interpretable representations of nonstationary behavior. Using Bayesian methods to compute posteriors is valid. In addition, the experiments for liver transplantations are interesting.\n\nWeaknesses: The presentation is hard to follow. It should be improved.  There is no new technique in this paper and there are no theoretical guarantees for their proposed algorithms. It would be interesting if the authors can upper bound the error $|\\rho^* - \\hat{\\rho}^*|$  to show that their algorithms converge.   ", "summary_of_the_review": "Based on strengths and weaknesses as I mentioned above, I think this paper is on borderline. Now I am leaning towards acceptance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635916060225}, {"id": "-P9plnD_q0R", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2631/Reviewer_79eV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors present the ICB an offline method for providing an interpretable description of observed decision-making whilst capturing the agent's non-stationary understanding of the world.", "review_text": "# Abstract\n\n- Style: you don't need to quote the ICB abbreviation; you introduced it for the first time so not only do you not need to quote it that's also not how abbreviations are defined.\n\n# Introduction\n\n- Style: pick quoting or emphasis, don't use both w.r.t. \"descriptive modeling\".\n- I find your core application interesting. I wonder w.r.t. introduced policies if you attach a particular reward to a particular organ allocation policy? Indeed one may find that a policy is not that good at all so then one would have to revert back to the previous observation regime and in so doing, have introduced an unnecessary non-stationary regime.\n- Style: please put your desiderata in bullets, it makes it easier to read.\n- What do you mean that the environment needs to be learned? I.e. which part are you referring to of the 3-tuple that is a contextual bandit: $\\langle A,S,R \\rangle$? Please be precise. It is currently not clear what you mean (I think you mean the action-indexed reward-distributions, but pretend that not everyone reading this is a bandit pro).\n- Figure 1: I cannot help but to think that this is an overly complex way of depicting non-stationarity. You may as well just show a Markov chain of rewards and simply say that the directed edges propagate through different regimes and ergo the model as stands is non-stationary. Effectively just an HMM wherein each state your agent takes actions. But alas, I may have misunderstood your intentions with that figure.\n- Isn't an ICB simply a structural (in time) bandit conditional on a previous bandit which acted in the same domain (though crucially not in the same environment)? Consider bandit one ($t=1$): $\\langle A_1,S_1,R_2 \\rangle$ and bandit two ($t=2$): $\\langle A_1,S_2,R_2 \\rangle$ it sounds as if you're suggesting updating the reward distribution at time two as: $p(R_2 \\mid A_1, A_2, R_1)$. One bandit is conditional on a past bandit which also took actions in the same system.\n- What does \"evolves smoothly\" mean? Do you mean that you're able to take derivatives and that the environment is Lipschitz continuous (the former is strictly a subset of the latter) or do you mean that there are no discontinuities?\n\n# Inverse contextual bandits\n\n- Notation: perhaps you should stick to convention w.r.t. to the state-space and denote it with $\\mathcal{S}$ as is common.\n- $\\mathbb{D}$ sounds awfully a lot like an MDP - what's the difference?\n- \"In this work, we consider state transitions that occur independently of past states and actions\" - that's an interesting assumption. What's the logic there? We assume that the organ allocation problem operates on the same dynamical system (your 'environment') no? If so, then tracking the non-stationary behaviour, without taking past allocations into account, is akin to saying that your bandits are all independent. But if that is the case why consider contextual bandits at all (I.e. why include $\\mathcal{S}$)? Your assumption alone allows you to simply stack up 'normal' bandits $\\langle A,R \\rangle$ Which it to say you are prescribing more complexity than you are actually using - why?\n- It is **not** ok for you to cite research nr [23] (by Y. Qin, F. Imrie etc) which is not yet publically available, not even on arXiv, for us (reviewers) to corroborate your claims.\n- You say \"Distributions of newly arriving organs are largely independent of prior allocation decisions; [...] [23, 32, 33]\" -- figure 2 and section 3 of [33] very much seem to refute your statement. It would be helpful for you to have a similar diagram and point to the d-separation that you claim impose the conditional independence. At present, it sounds implausible.\n- For definition 1, one can you please make use of equation environments to show your maths -- you have superscripts to subscripts and it is all becoming a bit too small.\n- You have now defined ICB about four times, once is enough.\n- Can you please clarify this statement: \"In this first work on modeling non-stationary agents,\" -- non-stationary MABs have been studied for decades so I do not think that's what you mean, hence, please clarify.\n- Table 1 is great, good job, very informative.\n\n# Illustrative examples\n\n- Figure 3 is great.\n- Figure 4 is great. I think it would be better if you had a similar figure or that figure closer to the top; it really captures the problem you are trying to solve. So placing it closer to the beginning of the paper may help the reader get the gist of your paper early on.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors present the ICB an offline method for providing an interpretable description of observed decision-making whilst capturing the agent's non-stationary understanding of the world.", "main_review": "# Abstract\n\n- Style: you don't need to quote the ICB abbreviation; you introduced it for the first time so not only do you not need to quote it that's also not how abbreviations are defined.\n\n# Introduction\n\n- Style: pick quoting or emphasis, don't use both w.r.t. \"descriptive modeling\".\n- I find your core application interesting. I wonder w.r.t. introduced policies if you attach a particular reward to a particular organ allocation policy? Indeed one may find that a policy is not that good at all so then one would have to revert back to the previous observation regime and in so doing, have introduced an unnecessary non-stationary regime.\n- Style: please put your desiderata in bullets, it makes it easier to read.\n- What do you mean that the environment needs to be learned? I.e. which part are you referring to of the 3-tuple that is a contextual bandit: $\\langle A,S,R \\rangle$? Please be precise. It is currently not clear what you mean (I think you mean the action-indexed reward-distributions, but pretend that not everyone reading this is a bandit pro).\n- Figure 1: I cannot help but to think that this is an overly complex way of depicting non-stationarity. You may as well just show a Markov chain of rewards and simply say that the directed edges propagate through different regimes and ergo the model as stands is non-stationary. Effectively just an HMM wherein each state your agent takes actions. But alas, I may have misunderstood your intentions with that figure.\n- Isn't an ICB simply a structural (in time) bandit conditional on a previous bandit which acted in the same domain (though crucially not in the same environment)? Consider bandit one ($t=1$): $\\langle A_1,S_1,R_2 \\rangle$ and bandit two ($t=2$): $\\langle A_1,S_2,R_2 \\rangle$ it sounds as if you're suggesting updating the reward distribution at time two as: $p(R_2 \\mid A_1, A_2, R_1)$. One bandit is conditional on a past bandit which also took actions in the same system.\n- What does \"evolves smoothly\" mean? Do you mean that you're able to take derivatives and that the environment is Lipschitz continuous (the former is strictly a subset of the latter) or do you mean that there are no discontinuities?\n\n# Inverse contextual bandits\n\n- Notation: perhaps you should stick to convention w.r.t. to the state-space and denote it with $\\mathcal{S}$ as is common.\n- $\\mathbb{D}$ sounds awfully a lot like an MDP - what's the difference?\n- \"In this work, we consider state transitions that occur independently of past states and actions\" - that's an interesting assumption. What's the logic there? We assume that the organ allocation problem operates on the same dynamical system (your 'environment') no? If so, then tracking the non-stationary behaviour, without taking past allocations into account, is akin to saying that your bandits are all independent. But if that is the case why consider contextual bandits at all (I.e. why include $\\mathcal{S}$)? Your assumption alone allows you to simply stack up 'normal' bandits $\\langle A,R \\rangle$ Which it to say you are prescribing more complexity than you are actually using - why?\n- It is **not** ok for you to cite research nr [23] (by Y. Qin, F. Imrie etc) which is not yet publically available, not even on arXiv, for us (reviewers) to corroborate your claims.\n- You say \"Distributions of newly arriving organs are largely independent of prior allocation decisions; [...] [23, 32, 33]\" -- figure 2 and section 3 of [33] very much seem to refute your statement. It would be helpful for you to have a similar diagram and point to the d-separation that you claim impose the conditional independence. At present, it sounds implausible.\n- For definition 1, one can you please make use of equation environments to show your maths -- you have superscripts to subscripts and it is all becoming a bit too small.\n- You have now defined ICB about four times, once is enough.\n- Can you please clarify this statement: \"In this first work on modeling non-stationary agents,\" -- non-stationary MABs have been studied for decades so I do not think that's what you mean, hence, please clarify.\n- Table 1 is great, good job, very informative.\n\n# Illustrative examples\n\n- Figure 3 is great.\n- Figure 4 is great. I think it would be better if you had a similar figure or that figure closer to the top; it really captures the problem you are trying to solve. So placing it closer to the beginning of the paper may help the reader get the gist of your paper early on.", "summary_of_the_review": "This is a good and interesting paper. But it is questionable if this could not be achieved with non-stationary MABs or indeed; if past organ allocations have no impact on present allocation, simply running one MAB after another (though it is not clear why we cannot do this). Simply, their novel question in the second paragraph of the introduction is not novel and has been asked many times before.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635425820923}, {"id": "jVncPKAGZir", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2631/Reviewer_pmos"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper formulates and studies the problem named “Inverse Contextual Bandits (ICB)”, which asks “Given demonstrated behavior from a decision-making agent, how has the agent’s knowledge been evolving over time?”  Then, the paper proposes two concrete learning algorithms, imposing different specifications regarding the agent’s behavioral strategy. Finally, some simulations using both simulated and real-world data are provided, showing how ICB can be applied as an investigative device for recovering and explaining the evolution of organ allocation practices over the years. ", "review_text": "There are many Strengths in the paper:\n- the paper tackles and formulates the important and relevant problem of ICB\n- the paper is very well-written and the motivation was easy to follow\n- related work is covered rigorously\n- simulation setups are clearly stated and seem comprehensive \n\nA weakness might be that the implementations used in the experiments are not shared at this moment. I would love to see an effort to ensure reproducibility (sharing the implementations as supplementary material). \n\nNote here that I’m not the expert in this sub-field, so I might miss some critical merits or weaknesses of the paper. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper formulates and studies the problem named “Inverse Contextual Bandits (ICB)”, which asks “Given demonstrated behavior from a decision-making agent, how has the agent’s knowledge been evolving over time?”  Then, the paper proposes two concrete learning algorithms, imposing different specifications regarding the agent’s behavioral strategy. Finally, some simulations using both simulated and real-world data are provided, showing how ICB can be applied as an investigative device for recovering and explaining the evolution of organ allocation practices over the years. ", "main_review": "There are many Strengths in the paper:\n- the paper tackles and formulates the important and relevant problem of ICB\n- the paper is very well-written and the motivation was easy to follow\n- related work is covered rigorously\n- simulation setups are clearly stated and seem comprehensive \n\nA weakness might be that the implementations used in the experiments are not shared at this moment. I would love to see an effort to ensure reproducibility (sharing the implementations as supplementary material). \n\nNote here that I’m not the expert in this sub-field, so I might miss some critical merits or weaknesses of the paper. \n", "summary_of_the_review": "Overall, the paper is well-written; the motivation, related work, and the experimental part are all easy to follow. I do not find critical flaw in the paper at this moment, but it is very likely that I might miss something. I would recommend weak accept, but I will pay attention to the other reviewer's opinion.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1634675825673}], "openreview_url": "https://openreview.net/forum?id=xw04RdwI2kS", "arxiv_id": "2107.06317", "paper_pdf": "papers/xw04RdwI2kS.pdf", "paper_pdf_sha256": "c61e0337f8e7874a6ff8e8d9ee63565d9c735628905549da16b084317611b111", "paper_pdf_bytes": 2697785, "paper_pdf_source": "openreview", "code_url": "https://github.com/alihanhyk/invconban", "code_repository": "alihanhyk/invconban", "code_commit": "b365614697053f3584915ee013cf6fee36e73de5", "code_archive": "repos/xw04RdwI2kS.zip", "code_archive_sha256": "8529981d61f2f26816a14f729e68b7d323d31a6e8c4f9518bc435c0dd9d2e4eb", "code_archive_bytes": 17069, "code_file_count": 12, "code_extensions": {".py": 11, ".sh": 1}, "github_disk_usage_kb": 12, "github_languages": {"Python": 32635, "Shell": 2089}, "github_archived": false, "github_pushed_at": "2022-06-07T11:59:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/inverse-contextual-bandits-learning-how"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EsA9Nr9JHvy", "year": 2021, "status": "rejected", "title": "The Heavy-Tail Phenomenon in SGD", "authors": ["Mert Gurbuzbalaban", "Umut Simsekli", "Lingjiong Zhu"], "authorids": ["~Mert_Gurbuzbalaban1", "~Umut_Simsekli1", "~Lingjiong_Zhu1"], "authors_source": "OpenReview API", "abstract": "In recent years, various notions of capacity and complexity have been proposed for characterizing the generalization properties of stochastic gradient descent (SGD) in deep learning. Some of the popular notions that correlate well with the performance on unseen data are (i) the 'flatness' of the local minimum found by SGD, which is related to the eigenvalues of the Hessian, (ii) the ratio of the stepsize $\\eta$ to the batch size $b$, which essentially controls the magnitude of the stochastic gradient noise, and (iii) the 'tail-index', which measures the heaviness of the tails of the network weights at convergence. In this paper, we argue that these three seemingly unrelated perspectives for generalization are deeply linked to each other. We claim that depending on the structure of the Hessian of the loss at the minimum, and the choices of the algorithm parameters $\\eta$ and $b$, the SGD iterates will converge to a \\emph{heavy-tailed} stationary distribution. We rigorously prove this claim in the setting of quadratic optimization: we show that even in a simple linear regression problem with independent and identically distributed Gaussian data, the iterates can be heavy-tailed with infinite variance. We further characterize the behavior of the tails with respect to algorithm parameters, the dimension, and the curvature. We then translate our results into insights about the behavior of SGD in deep learning. We finally support our theory with experiments conducted on both synthetic data and fully connected neural networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "KgWY2lcpI9I", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper438/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper gives a theoretical study of the tail behavior of the SGD in a quadratic optimization problem and explores its relationship with the curvature, step size and batch size. To prove their results, the authors approximate the SGD recursion by a linear stochastic recursion and analyze the statistical properties by the tools from implicit renew theory. Under this setting, they show that the law of the SGD iterates converge to a heavy-tailed stationary distribution depending on the Hessian structure of the loss function at the minimum and choices of the step size and batch size. They take a further step to clarify the relationship and study the moment bounds and convergence rate. \n\nOverall, I vote for accepting. I think the study of heavy tail phenomenon in SGD is quite a novel field and full of interest. \n1.\tThis paper is the first one to study the originating cause of heavy tail phenomenon in SGD and give a rigorous proof of the relationship between tail index and the choice of step size and batch size. This helps a better understanding of the generalization properties of SGD. \n2.\tThe paper provides comprehensive experiments to support their result. Experiments are conducted on both synthetic data and neural networks. The design of experiments is reasonable, and the results of the experiment not only support the claim that the tail index is deeply linked to the curvature of the loss and the ratio of step size and batch size but also give an insight on more general cases besides the quadratic optimization.\n\nHowever, I am still concerned about the setting in the theoretical frame. This paper completes its proof under the settings of quadratic optimization and infinite streaming data, which may limit the applicability of the theoretical result. Although these issues have been discussed in this paper, whether or why this extension is feasible remains questionable.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An Insightful Paper to be accpeted", "review": "This paper gives a theoretical study of the tail behavior of the SGD in a quadratic optimization problem and explores its relationship with the curvature, step size and batch size. To prove their results, the authors approximate the SGD recursion by a linear stochastic recursion and analyze the statistical properties by the tools from implicit renew theory. Under this setting, they show that the law of the SGD iterates converge to a heavy-tailed stationary distribution depending on the Hessian structure of the loss function at the minimum and choices of the step size and batch size. They take a further step to clarify the relationship and study the moment bounds and convergence rate. \n\nOverall, I vote for accepting. I think the study of heavy tail phenomenon in SGD is quite a novel field and full of interest. \n1.\tThis paper is the first one to study the originating cause of heavy tail phenomenon in SGD and give a rigorous proof of the relationship between tail index and the choice of step size and batch size. This helps a better understanding of the generalization properties of SGD. \n2.\tThe paper provides comprehensive experiments to support their result. Experiments are conducted on both synthetic data and neural networks. The design of experiments is reasonable, and the results of the experiment not only support the claim that the tail index is deeply linked to the curvature of the loss and the ratio of step size and batch size but also give an insight on more general cases besides the quadratic optimization.\n\nHowever, I am still concerned about the setting in the theoretical frame. This paper completes its proof under the settings of quadratic optimization and infinite streaming data, which may limit the applicability of the theoretical result. Although these issues have been discussed in this paper, whether or why this extension is feasible remains questionable.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604906533485}, {"id": "S2_rqkUZWUc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper438/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The main theme of this work is to study conditions under which SGD iterations result in random variables with heavy-tail random distributions. Specifically they focus on the step size, batch size and problem dimension. First they show theoretical results showing how the tail-index of the distribution generated by SGD depends on the chosen step size, batch size and problem dimension.\nThese results make several assumptions limiting their scope of applicability. Namely, they are limited to quadratic optimisation problems, normally distributed data and one-pass regime of SGD. However, the authors argue that since even in this simplest setting SGD results in a heavy-tail distribution, the more complex problems should not be expected to behave otherwise.\nThe experiments verify the correctness of the theoretical results, but they are also of limited value. For example, the deep learning section studies a shallow fully connected network on MNIST and CIFAR10. It would be more interesting to see a model performing closer to the state of the art here.\nWhile the main results of this work are interesting, I have a few reservations. First, as I wrote above the setting (even in experiments) is limited to the simplest case and hence is of limited value. Second, it is quite difficult to follow the theoretical results, mostly because all details are given in the Appendix, which is 30 pages long. Perhaps, this work is better suited for a journal publication, where the authors will have more space to explain their work. And finally I have some rather minor comments:\n- the paper refers to equation (3) in several places of the article and the Appendix, but there is no equation (3). I had to guess to which equation they refer each time, which made reading the already technically heavy paper more difficult.\n- In Theorem 4, it is not clear how \"the p-th moment of the iterates are of exponentially decaying nature\", since in both cases the coefficient of $\\mathbb{E}\\Vert q_1\\Vert^p$ is larger than 1.  Also, in the second case, the coefficient's value depends on whether $\\alpha<2$. What is $q_1$ here?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "SGB iterations can result in heavy-tail random distribution on certain quadratic optimisation problems", "review": "The main theme of this work is to study conditions under which SGD iterations result in random variables with heavy-tail random distributions. Specifically they focus on the step size, batch size and problem dimension. First they show theoretical results showing how the tail-index of the distribution generated by SGD depends on the chosen step size, batch size and problem dimension.\nThese results make several assumptions limiting their scope of applicability. Namely, they are limited to quadratic optimisation problems, normally distributed data and one-pass regime of SGD. However, the authors argue that since even in this simplest setting SGD results in a heavy-tail distribution, the more complex problems should not be expected to behave otherwise.\nThe experiments verify the correctness of the theoretical results, but they are also of limited value. For example, the deep learning section studies a shallow fully connected network on MNIST and CIFAR10. It would be more interesting to see a model performing closer to the state of the art here.\nWhile the main results of this work are interesting, I have a few reservations. First, as I wrote above the setting (even in experiments) is limited to the simplest case and hence is of limited value. Second, it is quite difficult to follow the theoretical results, mostly because all details are given in the Appendix, which is 30 pages long. Perhaps, this work is better suited for a journal publication, where the authors will have more space to explain their work. And finally I have some rather minor comments:\n- the paper refers to equation (3) in several places of the article and the Appendix, but there is no equation (3). I had to guess to which equation they refer each time, which made reading the already technically heavy paper more difficult.\n- In Theorem 4, it is not clear how \"the p-th moment of the iterates are of exponentially decaying nature\", since in both cases the coefficient of $\\mathbb{E}\\Vert q_1\\Vert^p$ is larger than 1.  Also, in the second case, the coefficient's value depends on whether $\\alpha<2$. What is $q_1$ here?\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604246417313}, {"id": "uaGz5M7KT-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper438/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the relations between the heavy tail phenomenon of SGD and the ‘flatness’ of the local minimum found by SGD and the ratio of the step size $\\eta$ to the batch size $b$ for the quadratic and convex problem. They show that depending on the curvature, the step size, and the batch size, the iterates can converge to a heavy-tailed random variable.  They conduct experiments on both synthetic data and fully connected neural networks, and illustrate that the results would also apply to more general settings and hence provide new insights about the behavior of SGD in deep learning. \nI do not have time to check the proofs and are not familiar with this topic. However the results seem to be novel and interesting. \n\npros: 1, They characterize the empirically observed heavy-tailed behavior of SGD with respect to the step size, mini batch size, dimensions, and the curvature, with explicit convergence rates. \n\ncons:1, The theoretical analysis is only for the quadratic and convex function. The input data is Gaussian, which is quite restrictive. \n\nAfter the rebuttal.\n\nThe authors addressed my concerns. I have read other reviewers' comments. I decide to remain the current score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Reviews for The Heavy-Tail Phenomenon in SGD", "review": "This paper studies the relations between the heavy tail phenomenon of SGD and the ‘flatness’ of the local minimum found by SGD and the ratio of the step size $\\eta$ to the batch size $b$ for the quadratic and convex problem. They show that depending on the curvature, the step size, and the batch size, the iterates can converge to a heavy-tailed random variable.  They conduct experiments on both synthetic data and fully connected neural networks, and illustrate that the results would also apply to more general settings and hence provide new insights about the behavior of SGD in deep learning. \nI do not have time to check the proofs and are not familiar with this topic. However the results seem to be novel and interesting. \n\npros: 1, They characterize the empirically observed heavy-tailed behavior of SGD with respect to the step size, mini batch size, dimensions, and the curvature, with explicit convergence rates. \n\ncons:1, The theoretical analysis is only for the quadratic and convex function. The input data is Gaussian, which is quite restrictive. \n\nAfter the rebuttal.\n\nThe authors addressed my concerns. I have read other reviewers' comments. I decide to remain the current score.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603927889455}, {"id": "wFToM9QUE--", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper438/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper studies the heavy-tail phenomenon in SGD. In recent years, many works either empirically or theoretically showed that the heavy-tail property of SGD leads to a better generalization performance. But no rigorous theory has been developed to explain the cause of this phenomenon. \n\nIn this paper, the authors studied SGD iterates in solving a quadratic linear regression problem with i.i.d (Gaussian) data. In this setting, they proved the main result that shows that the tails of SGD iterates become monotonically heavier for increasing curvature, decreasing stepwise, or increasing b, hence this tends to provide a rigorous explanation of the heavy-tail phenomenon. The authors also developed other results on the law convergence and variance bounds.\n\nComments: Overall I think the paper is technically sound and develops rigorous proof. My main comment is that the problem studied in the paper is very idealized, I.e., a regression with iid Gaussian data. I understand that this makes the proof cleaner with compact results. However, this may not have direct implications on the heavy-tail phenomenon in deep learning. In fact, it may be more meaningful and motivated to study other simpler problems, such as binary classification with linearly separable data, in which the data structure is more practical than iid Gaussian. \n\nOverall, I consider the results in this paper as rigorously exploring the heavy-tail of SGD in linear regression under iid Gaussian data. I suggest the authors consider other more practical problems, and the results need not be related to deep learning. Focusing on the heavy-tail phenomenon in standard machine learning problems is already an interesting problem.
", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Review 1", "review": "Summary: This paper studies the heavy-tail phenomenon in SGD. In recent years, many works either empirically or theoretically showed that the heavy-tail property of SGD leads to a better generalization performance. But no rigorous theory has been developed to explain the cause of this phenomenon. \n\nIn this paper, the authors studied SGD iterates in solving a quadratic linear regression problem with i.i.d (Gaussian) data. In this setting, they proved the main result that shows that the tails of SGD iterates become monotonically heavier for increasing curvature, decreasing stepwise, or increasing b, hence this tends to provide a rigorous explanation of the heavy-tail phenomenon. The authors also developed other results on the law convergence and variance bounds.\n\nComments: Overall I think the paper is technically sound and develops rigorous proof. My main comment is that the problem studied in the paper is very idealized, I.e., a regression with iid Gaussian data. I understand that this makes the proof cleaner with compact results. However, this may not have direct implications on the heavy-tail phenomenon in deep learning. In fact, it may be more meaningful and motivated to study other simpler problems, such as binary classification with linearly separable data, in which the data structure is more practical than iid Gaussian. \n\nOverall, I consider the results in this paper as rigorously exploring the heavy-tail of SGD in linear regression under iid Gaussian data. I suggest the authors consider other more practical problems, and the results need not be related to deep learning. Focusing on the heavy-tail phenomenon in standard machine learning problems is already an interesting problem.
", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603646889803}], "openreview_url": "https://openreview.net/forum?id=EsA9Nr9JHvy", "arxiv_id": "2006.04740", "paper_pdf": "papers/EsA9Nr9JHvy.pdf", "paper_pdf_sha256": "2b9da9bfcea610172575c5d32bd6f0bc1e57967da89c7c3f84d5efb39bac18c6", "paper_pdf_bytes": 718734, "paper_pdf_source": "openreview", "code_url": "https://github.com/umutsimsekli/sgd_ht", "code_repository": "umutsimsekli/sgd_ht", "code_commit": "67f2b51055a331f23748bd418c056de1aa7cb48f", "code_archive": "repos/EsA9Nr9JHvy.zip", "code_archive_sha256": "d88873ef01711a7064b73963ad9fb2658772af2e6e4052470c88de222f176ad2", "code_archive_bytes": 12648, "code_file_count": 6, "code_extensions": {".py": 4, ".ipynb": 2}, "github_disk_usage_kb": 12, "github_languages": {"Python": 18880, "Jupyter Notebook": 12049}, "github_archived": false, "github_pushed_at": "2022-02-15T13:12:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-heavy-tail-phenomenon-in-sgd"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ByxXZpVtPB", "year": 2020, "status": "rejected", "title": "Homogeneous Linear Inequality Constraints for Neural Network Activations", "authors": ["Thomas Frerix", "Matthias Nießner", "Daniel Cremers"], "authorids": ["thomas.frerix@tum.de", "niessner@tum.de", "cremers@tum.de"], "authors_source": "OpenReview API", "abstract": "We propose a method to impose homogeneous linear inequality constraints of the form $Ax\\leq 0$ on neural network activations. The proposed method allows a data-driven training approach to be combined with modeling prior knowledge about the task. One way to achieve this task is by means of a projection step at test time after unconstrained training.\nHowever, this is an expensive operation. By directly incorporating the constraints into the architecture, we can significantly speed-up inference at test time; for instance, our experiments show a speed-up of up to two orders of magnitude over a projection method. Our algorithm computes a suitable parameterization of the feasible set at initialization and uses standard variants of stochastic gradient descent to find solutions to the constrained network. Thus, the modeling constraints are always satisfied during training. Crucially, our approach avoids to solve an optimization problem at each training step or to manually trade-off data and constraint fidelity with additional hyperparameters. We consider constrained generative modeling as an important application domain and experimentally demonstrate the proposed method by constraining a variational autoencoder.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkxudRZP5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper370/AnonReviewer5"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method to impose linear inequality constraints on neural network activations. The method is implemented at initialization (by converting the H-representation to the V-representation) and during training (by modifying the network architecture). Experiments on two setups (projection and VAE+projection on a checkerboard pattern on MNIST) demonstrate a 2-orders of magnitude speed-up with respect to test-time projection (computed with OSQP).\n\nThe contributions claimed are:\n* Novel technique to impose inequality constraints on neural network activations.\n* Significant speed-up w.r.t. other techniques (at test time).\n\nOverall, the approach is well motivated, and well-placed in the literature. However, the experimental analysis does not support all the claims made by the authors as it focuses on a single dataset (i.e., MNIST) and single constraint (i.e., checkerboard pattern). Additionally, while the authors argue that there is no manual trade-off between constraint satisfaction and data representation, the experiment in Fig. 3 appears to show that there is some manual trade-off (using box delay). As such, I am currently inclined to give a \"weak reject\" score.\n\nThe method is clear and the idea of using softmax to get a convex combination of the vertices of the V-representation to guarantee constraint satisfaction is reasonable.\n\n1) The softmax used to satisfy the constraints is preceded by a batch normalization layer. I would expect batch normalization to interfere with the ability to saturate the softmax activation and thus prevent the network from reaching optimality. Could the authors provide experiments that justify the use of the batch normalization layer?\n\n2) An important factor that is unexplored in this manuscript is the softmax temperature. A good scheduling of that temperature could help the optimization. Have authors tried different temperature values?\n\n3) The use of softmax and the integration of the constraints as a network layer seem to create some difficulty during training (even with the rather simple checkerboard pattern used in the experiment). The loss appears to reach some plateau (9% from optimal) and, thus, there appears to be some trade-off between reconstruction and projection. The authors should provide more experiments to explain that trade-off. That trade-off is also visible on Fig. 5 (the zero has a significantly different shape).\n\n4) The method is solely compared to test time projection. Could the authors implement other techniques (if feasible)? e.g., OptNet. Overall, it would helpful to add more setups and different types of constraints (other than a last-layer projection; e.g., monotonicity).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #5", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper proposes a method to impose linear inequality constraints on neural network activations. The method is implemented at initialization (by converting the H-representation to the V-representation) and during training (by modifying the network architecture). Experiments on two setups (projection and VAE+projection on a checkerboard pattern on MNIST) demonstrate a 2-orders of magnitude speed-up with respect to test-time projection (computed with OSQP).\n\nThe contributions claimed are:\n* Novel technique to impose inequality constraints on neural network activations.\n* Significant speed-up w.r.t. other techniques (at test time).\n\nOverall, the approach is well motivated, and well-placed in the literature. However, the experimental analysis does not support all the claims made by the authors as it focuses on a single dataset (i.e., MNIST) and single constraint (i.e., checkerboard pattern). Additionally, while the authors argue that there is no manual trade-off between constraint satisfaction and data representation, the experiment in Fig. 3 appears to show that there is some manual trade-off (using box delay). As such, I am currently inclined to give a \"weak reject\" score.\n\nThe method is clear and the idea of using softmax to get a convex combination of the vertices of the V-representation to guarantee constraint satisfaction is reasonable.\n\n1) The softmax used to satisfy the constraints is preceded by a batch normalization layer. I would expect batch normalization to interfere with the ability to saturate the softmax activation and thus prevent the network from reaching optimality. Could the authors provide experiments that justify the use of the batch normalization layer?\n\n2) An important factor that is unexplored in this manuscript is the softmax temperature. A good scheduling of that temperature could help the optimization. Have authors tried different temperature values?\n\n3) The use of softmax and the integration of the constraints as a network layer seem to create some difficulty during training (even with the rather simple checkerboard pattern used in the experiment). The loss appears to reach some plateau (9% from optimal) and, thus, there appears to be some trade-off between reconstruction and projection. The authors should provide more experiments to explain that trade-off. That trade-off is also visible on Fig. 5 (the zero has a significantly different shape).\n\n4) The method is solely compared to test time projection. Could the authors implement other techniques (if feasible)? e.g., OptNet. Overall, it would helpful to add more setups and different types of constraints (other than a last-layer projection; e.g., monotonicity)."}, "tcdate": 1572441711710}, {"id": "SJeme00IKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper370/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a method for imposing linear inequality in neural networks. Although the contribution of the paper is potentially significant, some details are not clearly described.\n\nIn the proposed method, the inequality constraints are converted from the representation with a matrix to the representation with rays that represents the cone which satisfies the constraints. The neural network is trained so as to satisfy the constraint represented by rays. However, I do not understand how to train the neural network to satisfy the constraint represented by rays, which is actually the core of the algorithm. I cannot see how satisfaction of constraints are guaranteed from the current description.\n\nThe empirical results show that the variational autoencoder trained with the proposed method can generate images that satisfy linear constraints. However, the evaluation is limited to a checkerboad constraint, and other examples of practical linear constraints are not clear.\n\nDue to the unclear algorithm description and limited empirical results, I give weak reject to the paper. However, I'm happy raise the score if authors clarify some points in the rebuttal.\n\nI would like authors to address the following points in the rebuttal:\n\n- I do not understand the procedure of the proposed method. Especially, I do not understand how to achieve this part: \"During training, the neural network training algorithm is used to optimize within in the feasible set.\"\nPlease elaborate it in the rebuttal. Please describe how the satisfaction of the constraints are guaranteed.\n\n- I recommend authors to put a pseudo-code of the proposed algorithm for clarity.\n\n- In Section 4.1, it is stated that the constraint layer is added to the neural network. However, the computation performed in the constraint layer is not clear. Please describe it.\n\n- The checker board constraint on MNIST images is interesting, but it would be better to show more examples of linear constraints. If possible, please give some more examples of linear constraints and their results.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper presents a method for imposing linear inequality in neural networks. Although the contribution of the paper is potentially significant, some details are not clearly described.\n\nIn the proposed method, the inequality constraints are converted from the representation with a matrix to the representation with rays that represents the cone which satisfies the constraints. The neural network is trained so as to satisfy the constraint represented by rays. However, I do not understand how to train the neural network to satisfy the constraint represented by rays, which is actually the core of the algorithm. I cannot see how satisfaction of constraints are guaranteed from the current description.\n\nThe empirical results show that the variational autoencoder trained with the proposed method can generate images that satisfy linear constraints. However, the evaluation is limited to a checkerboad constraint, and other examples of practical linear constraints are not clear.\n\nDue to the unclear algorithm description and limited empirical results, I give weak reject to the paper. However, I'm happy raise the score if authors clarify some points in the rebuttal.\n\nI would like authors to address the following points in the rebuttal:\n\n- I do not understand the procedure of the proposed method. Especially, I do not understand how to achieve this part: \"During training, the neural network training algorithm is used to optimize within in the feasible set.\"\nPlease elaborate it in the rebuttal. Please describe how the satisfaction of the constraints are guaranteed.\n\n- I recommend authors to put a pseudo-code of the proposed algorithm for clarity.\n\n- In Section 4.1, it is stated that the constraint layer is added to the neural network. However, the computation performed in the constraint layer is not clear. Please describe it.\n\n- The checker board constraint on MNIST images is interesting, but it would be better to show more examples of linear constraints. If possible, please give some more examples of linear constraints and their results.\n"}, "tcdate": 1571380715223}, {"id": "rkAEA0NFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper370/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a new faster algorithm to add inequality constraints to neural layers. The paper focuses on a novel constraining approach with seemingly superior scalability, and this is potentially a significant contribution. However the paper does not motivate the constraining at all. I am baffled by this, since one would assume at least some benefits from all of this work could be presented. The only mentions are binarization of the predictions (which softmax already does), and monotonicity/convexity of neurons, with no proposed benefits. The running example of the paper is the chessboard constraint, which is either pointless (fig1) or harmful (fig5). Without justification and motivation the method has no merit and won’t have any impact in the machine learning community.\n\nThe monotonicity constraint could have a huge impact for MCMC sampling of neural parameters since it can reduce away all multimodalities of the posterior caused by reordering nodes or layers. \n\nI had hard time following the method, and I its not clear how the neural network is modified and how backpropagation is performed with the contraints. It is not defined properly how the constrained optimisation works. Apparently additional neural layers are added that map z's to r's. The backpropagation in the constrained case is undefined. Here an algorithm box or schematic figure comparing unconstrained and constrained NN architectures would be extremely helpful. It’s also not explained how are modelling/domain constraints different. \n\nThe paper does not compare to the earlier constrained methods (Marquaz-Neila or OptNet), and thus there is no demonstration of the methods claimed superior computational efficiency. The paper also does not make very clear the different constraining approach advantages and tradeoffs. A comparison table would be help a lot.\n\nThe method is interesting, novel and seemingly efficient; but it is insufficiently defined, the method is not motivated and experiments are quite weak with little comparisons and no experiments with practical value.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes a new faster algorithm to add inequality constraints to neural layers. The paper focuses on a novel constraining approach with seemingly superior scalability, and this is potentially a significant contribution. However the paper does not motivate the constraining at all. I am baffled by this, since one would assume at least some benefits from all of this work could be presented. The only mentions are binarization of the predictions (which softmax already does), and monotonicity/convexity of neurons, with no proposed benefits. The running example of the paper is the chessboard constraint, which is either pointless (fig1) or harmful (fig5). Without justification and motivation the method has no merit and won’t have any impact in the machine learning community.\n\nThe monotonicity constraint could have a huge impact for MCMC sampling of neural parameters since it can reduce away all multimodalities of the posterior caused by reordering nodes or layers. \n\nI had hard time following the method, and I its not clear how the neural network is modified and how backpropagation is performed with the contraints. It is not defined properly how the constrained optimisation works. Apparently additional neural layers are added that map z's to r's. The backpropagation in the constrained case is undefined. Here an algorithm box or schematic figure comparing unconstrained and constrained NN architectures would be extremely helpful. It’s also not explained how are modelling/domain constraints different. \n\nThe paper does not compare to the earlier constrained methods (Marquaz-Neila or OptNet), and thus there is no demonstration of the methods claimed superior computational efficiency. The paper also does not make very clear the different constraining approach advantages and tradeoffs. A comparison table would be help a lot.\n\nThe method is interesting, novel and seemingly efficient; but it is insufficiently defined, the method is not motivated and experiments are quite weak with little comparisons and no experiments with practical value.\n"}, "tcdate": 1571249717516}], "openreview_url": "https://openreview.net/forum?id=ByxXZpVtPB", "arxiv_id": "1902.01785", "paper_pdf": "papers/ByxXZpVtPB.pdf", "paper_pdf_sha256": "4677f0bf54d7f1bf6d4fcbdea5d5ea6363aec79cc08e2ac1f2286647ff1ec40c", "paper_pdf_bytes": 733546, "paper_pdf_source": "openreview", "code_url": "https://github.com/tfrerix/constrained-nets", "code_repository": "tfrerix/constrained-nets", "code_commit": "1221f3fd2592056c2a00c463d65fe3afa646fd54", "code_archive": "repos/ByxXZpVtPB.zip", "code_archive_sha256": "6728797b652ebdf88e23afcf1b96053b5a62108e3c05f38349628efbfd76da9a", "code_archive_bytes": 12562, "code_file_count": 7, "code_extensions": {".py": 6, ".sh": 1}, "github_disk_usage_kb": 11, "github_languages": {"Python": 29194, "Shell": 436}, "github_archived": false, "github_pushed_at": "2021-04-08T19:17:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/linear-inequality-constraints-for-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJeZS3RcYm", "year": 2019, "status": "rejected", "title": "Simple Black-box Adversarial Attacks", "authors": ["Chuan Guo", "Jacob R. Gardner", "Yurong You", "Andrew G. Wilson", "Kilian Q. Weinberger"], "authorids": ["cg563@cornell.edu", "jrg365@cornell.edu", "yy785@cornell.edu", "andrew@cornell.edu", "kqw4@cornell.edu"], "authors_source": "OpenReview API", "abstract": "The construction of adversarial images is a search problem in high dimensions within a small region around a target image. The goal is to find an imperceptibly modified image that is misclassified by a target model. In the black-box setting, only sporadic feedback is provided through occasional model evaluations. In this paper we provide a new algorithm whose search strategy is based on an intriguingly simple iterative principle: We randomly pick a low frequency component of the discrete cosine transform (DCT) and either add or subtract it to the target image. Model evaluations are only required to identify whether an operation decreases the adversarial loss. Despite its simplicity, the proposed method can be used for targeted and untargeted attacks --- resulting in previously unprecedented query efficiency in both settings. We require a median of 600 black-box model queries (ResNet-50) to produce an adversarial ImageNet image, and we successfully attack Google Cloud Vision with 2500 median queries, averaging to a cost of only $3 per image. We argue that our proposed algorithm should serve as a strong baseline for future adversarial black-box attacks, in particular because it is extremely fast and can be implemented in less than 20 lines of PyTorch code. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rkxkTpcJTQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1514/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper demonstrates that a simple greedy random search algorithm in DCT space based on score feedback is able to synthesize adversarial examples with quite good query efficiency.  The algorithm is demonstrated on ImageNet  with three common architectures, showing much higher efficiency when sampling from the DCT basis. The algorithm is also shown to outperform state of the art attacks in terms of query count. Finally, a successful attack is demonstrated on Google Cloud Vision.\n\nWhile not particularly heavy on technicalities, this work does make a couple intriguing points, namely that adversarial attacks can potentially be quite easy to perform due to the inherent nature of high-dimensional classification, and that the space is in which the search is perform might be more important than the sophistication of the search itself. I interpret the proposal not so much as a claim to a state-of-the-art algorithm (even though the results are impressive) but as a very reasonable baseline in the evaluation of attack efficiency -- one might even wonder why it has not been common practice thus far to evaluate against such kinds of algorithms by default. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "simple algorithm, intriguing message", "review": "This paper demonstrates that a simple greedy random search algorithm in DCT space based on score feedback is able to synthesize adversarial examples with quite good query efficiency.  The algorithm is demonstrated on ImageNet  with three common architectures, showing much higher efficiency when sampling from the DCT basis. The algorithm is also shown to outperform state of the art attacks in terms of query count. Finally, a successful attack is demonstrated on Google Cloud Vision.\n\nWhile not particularly heavy on technicalities, this work does make a couple intriguing points, namely that adversarial attacks can potentially be quite easy to perform due to the inherent nature of high-dimensional classification, and that the space is in which the search is perform might be more important than the sophistication of the search itself. I interpret the proposal not so much as a claim to a state-of-the-art algorithm (even though the results are impressive) but as a very reasonable baseline in the evaluation of attack efficiency -- one might even wonder why it has not been common practice thus far to evaluate against such kinds of algorithms by default. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541545399105}, {"id": "SJeaN4LihX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1514/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a simple and effective black box adversarial attack on sota deep nets for image classification tasks. It is based on randomly picking a low frequency component of the DC Transform.  It is claimed to be most efficient when compared to the sota methods in terms of number of queries required for the attack. It is shown that a median of 600 queries for resnet-50 for imagenet dataset, and 2500 for google cloud vision. Due to its simplicity, it is also claimed that the attack is quite simple to implement in code. The paper presents a detailed analysis of their attack in pixel and DCT space, targeted vs untargeted attack, comparison over different architecture such as Densenet, resnet, and inception.\n\nThough the work is quite important and presents a simple and effective baseline black box attack. My concern is primarily on the novelty and originality of the idea, as it is mainly based on the work of Guo etal 2018, which this paper says is the motivation behind their work. So, it is not clear what is the contribution of this paper, as a similar study seems to have been carried out in that paper as well. The authors do not clearly give the relative comparison wrt Guo etal 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "simple and effective blackbox attack based on random directions", "review": "This paper presents a simple and effective black box adversarial attack on sota deep nets for image classification tasks. It is based on randomly picking a low frequency component of the DC Transform.  It is claimed to be most efficient when compared to the sota methods in terms of number of queries required for the attack. It is shown that a median of 600 queries for resnet-50 for imagenet dataset, and 2500 for google cloud vision. Due to its simplicity, it is also claimed that the attack is quite simple to implement in code. The paper presents a detailed analysis of their attack in pixel and DCT space, targeted vs untargeted attack, comparison over different architecture such as Densenet, resnet, and inception.\n\nThough the work is quite important and presents a simple and effective baseline black box attack. My concern is primarily on the novelty and originality of the idea, as it is mainly based on the work of Guo etal 2018, which this paper says is the motivation behind their work. So, it is not clear what is the contribution of this paper, as a similar study seems to have been carried out in that paper as well. The authors do not clearly give the relative comparison wrt Guo etal 2018.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541264436568}, {"id": "SkesQhRO2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1514/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a simple query-efficient \"score-based\" black-box attack based on iteratively perturbing an input image with a direction randomly sampled (w/o replacement) from a set of orthonormal bases. In particular, the authors proposed the use of low-frequency parts of DCT (discrete cosine transformation) as in (Guo 2018) to perform this task. Experimental results on ImageNet and three different classification models demonstrate the query efficiency of the proposed method -- able to achieve high attack success rate within fewer query budgets, where the visual distortion has an L2 norm threshold set to be 10. The authors also demonstrate an untargeted score-based black-box attack on Google CloudVision API.\n\nWhile the results seem promising, there are several issues that may potentially weaken the query-efficient claims made in this paper, especially due to the lack of sufficient attack comparisons (on smaller datasets) and inconsistent threat models when compared to existing works. My main concerns are summarized as follows.\n\n1. Unfair comparison due to inconsistent threat models (knowledge known to an attacker): the proposed method (simBA) is a \"score-based\" black-box attack, not a \"decision-based\" black-box attack. The proposed method assumes knowing the prediction likelihood (or prediction score) as the model output when performing black-box attacks, whereas the compared methods in black-box settings, such as Opt Attack and Boundary Attack, are \"decision-based\" attack that assumes only knowing the top-1 prediction label. Therefore, the query count comparison is meaningless and unfair, since these two methods require far less information from the model.\n\nOn the other hand, ZOO/AutoZOOM is a score-based attack. But ZOO can achieve a very low L2 distortion due to its coordinate descent nature. A fair comparison is to set the same L2 distortion for all score-based methods, and compare the median/avg query counts of each image to reach the same L2 distortion. The comparison to Opt-Attack / Boundary attack makes sense only if the proposed method (simBA) can also perform decision-based attack. Nonetheless, the query count to same-distortion comparison argument still holds. The authors should specify whether simBA can apply to the decision-only attack scenario. If so, how to implement and what is the performance?\n\nLastly, the QL attack (Ilyas 2018) can perform both score-based and decision-based attacks. So the authors should make the query comparison (to same L2 distortion) as well. According to a recent report (Table-1, NES column) in https://arxiv.org/pdf/1807.07978.pdf, the QL-attack has a comparable performance in terms of query counts as reported in this paper.\n\n2. More experiments on targeted black-box attacks: While untargeted attacks on Imagenet is a relatively easy task, I was a bit skeptical on the attacking performance of simBA in targeted attacks - since the selection of low-frequency bases directly limits the search space of adversarial examples, as opposed to arbitrary random directions adopted in QL-attack, Boundary-attack, and Opt-attack. It is also not clear how the target label is chosen in the targeted attack experiment.\nI suggest including two more experiments to validate the function of simBA: (i) compare the performance of least-likely targeted attack (ii) show results on smaller datasets such as Cifar-10. As pointed out by the authors, Imagenet has too many image dimensions and make it more vulnerable to attack. Showing attacking results on smaller datasets can properly justify the value of the proposed attack, rather than the benefit from high dimensionality.\n\n\n3. Novelty relative to LFBA (Guo) should be better differentiated: The idea of using DCT is originated from the LFBA paper. Since in that paper the authors also leveraged low-frequency DCT to perform black-box attacks, it is not clear to me what makes the proposed method perform better than the LFBA paper. The novelty and difference between this paper and the LFBA paper should be addressed.\n\n4. The Google Cloud Vision API attack is not too appealing - the tree label is still there and the trees are obviously present in the picture, while I appreciate the effect of removing the original top-3 labels. Can the authors show another set of non-trivial (more surprising) and targeted-attack experiments? Or simply do the same experiment using the same image (men snowing -> dog)  as in the QL-attack.\n\n----\nPost-rebuttal review\n\nI appreciate the authors' efforts in clarifying some of my concerns. However, I am still not convinced the comparison has been made fair. Many numbers from Table 1, such as ZOO, Opt-attack, QL-attack and AutoZOOM seem to be directly adapted from the papers rather than implemented and reproduced based on the same setting as the proposed attack. In particular, given that QL-attack is a published work, one of the state-of-the-art method and its codes has been released, I would really love to see a direct comparison using the same data samples and threat model. I would also like to emphasize that implementing all attacks under the same setting is crucial, since different attack methods may have a different criterion to determine attack successfulness. For example, QL-attack has some pre-defined distortion (L2 or Linfinity) for determining an adversarial example is successful, in addition to a different predicted class.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting black-box adversarial attack using DCT basis; performance evaluation on targeted attack is insufficient and threat model is inconsistent (unfair comparison)", "review": "This paper proposed a simple query-efficient \"score-based\" black-box attack based on iteratively perturbing an input image with a direction randomly sampled (w/o replacement) from a set of orthonormal bases. In particular, the authors proposed the use of low-frequency parts of DCT (discrete cosine transformation) as in (Guo 2018) to perform this task. Experimental results on ImageNet and three different classification models demonstrate the query efficiency of the proposed method -- able to achieve high attack success rate within fewer query budgets, where the visual distortion has an L2 norm threshold set to be 10. The authors also demonstrate an untargeted score-based black-box attack on Google CloudVision API.\n\nWhile the results seem promising, there are several issues that may potentially weaken the query-efficient claims made in this paper, especially due to the lack of sufficient attack comparisons (on smaller datasets) and inconsistent threat models when compared to existing works. My main concerns are summarized as follows.\n\n1. Unfair comparison due to inconsistent threat models (knowledge known to an attacker): the proposed method (simBA) is a \"score-based\" black-box attack, not a \"decision-based\" black-box attack. The proposed method assumes knowing the prediction likelihood (or prediction score) as the model output when performing black-box attacks, whereas the compared methods in black-box settings, such as Opt Attack and Boundary Attack, are \"decision-based\" attack that assumes only knowing the top-1 prediction label. Therefore, the query count comparison is meaningless and unfair, since these two methods require far less information from the model.\n\nOn the other hand, ZOO/AutoZOOM is a score-based attack. But ZOO can achieve a very low L2 distortion due to its coordinate descent nature. A fair comparison is to set the same L2 distortion for all score-based methods, and compare the median/avg query counts of each image to reach the same L2 distortion. The comparison to Opt-Attack / Boundary attack makes sense only if the proposed method (simBA) can also perform decision-based attack. Nonetheless, the query count to same-distortion comparison argument still holds. The authors should specify whether simBA can apply to the decision-only attack scenario. If so, how to implement and what is the performance?\n\nLastly, the QL attack (Ilyas 2018) can perform both score-based and decision-based attacks. So the authors should make the query comparison (to same L2 distortion) as well. According to a recent report (Table-1, NES column) in https://arxiv.org/pdf/1807.07978.pdf, the QL-attack has a comparable performance in terms of query counts as reported in this paper.\n\n2. More experiments on targeted black-box attacks: While untargeted attacks on Imagenet is a relatively easy task, I was a bit skeptical on the attacking performance of simBA in targeted attacks - since the selection of low-frequency bases directly limits the search space of adversarial examples, as opposed to arbitrary random directions adopted in QL-attack, Boundary-attack, and Opt-attack. It is also not clear how the target label is chosen in the targeted attack experiment.\nI suggest including two more experiments to validate the function of simBA: (i) compare the performance of least-likely targeted attack (ii) show results on smaller datasets such as Cifar-10. As pointed out by the authors, Imagenet has too many image dimensions and make it more vulnerable to attack. Showing attacking results on smaller datasets can properly justify the value of the proposed attack, rather than the benefit from high dimensionality.\n\n\n3. Novelty relative to LFBA (Guo) should be better differentiated: The idea of using DCT is originated from the LFBA paper. Since in that paper the authors also leveraged low-frequency DCT to perform black-box attacks, it is not clear to me what makes the proposed method perform better than the LFBA paper. The novelty and difference between this paper and the LFBA paper should be addressed.\n\n4. The Google Cloud Vision API attack is not too appealing - the tree label is still there and the trees are obviously present in the picture, while I appreciate the effect of removing the original top-3 labels. Can the authors show another set of non-trivial (more surprising) and targeted-attack experiments? Or simply do the same experiment using the same image (men snowing -> dog)  as in the QL-attack.\n\n----\nPost-rebuttal review\n\nI appreciate the authors' efforts in clarifying some of my concerns. However, I am still not convinced the comparison has been made fair. Many numbers from Table 1, such as ZOO, Opt-attack, QL-attack and AutoZOOM seem to be directly adapted from the papers rather than implemented and reproduced based on the same setting as the proposed attack. In particular, given that QL-attack is a published work, one of the state-of-the-art method and its codes has been released, I would really love to see a direct comparison using the same data samples and threat model. I would also like to emphasize that implementing all attacks under the same setting is crucial, since different attack methods may have a different criterion to determine attack successfulness. For example, QL-attack has some pre-defined distortion (L2 or Linfinity) for determining an adversarial example is successful, in addition to a different predicted class.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541102626573}], "openreview_url": "https://openreview.net/forum?id=rJeZS3RcYm", "arxiv_id": "1905.07121", "paper_pdf": "papers/rJeZS3RcYm.pdf", "paper_pdf_sha256": "fcc0747ef59f7d8c34a649fbff3282b4c583e6c42da58706e6596aa4d42b3333", "paper_pdf_bytes": 7340032, "paper_pdf_source": "openreview", "code_url": "https://github.com/cg563/simple-blackbox-attack", "code_repository": "cg563/simple-blackbox-attack", "code_commit": "3b66937945699f706918277ac55fb171f07d2bcf", "code_archive": "repos/rJeZS3RcYm.zip", "code_archive_sha256": "617afb689a318d671fbce9c38169653313174229ca1fb5e6fe025f9e86f570e0", "code_archive_bytes": 10814, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 18, "github_languages": {"Python": 28113}, "github_archived": false, "github_pushed_at": "2023-03-27T18:29:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/simple-black-box-adversarial-attacks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryykVe-0W", "year": 2018, "status": "rejected", "title": "Learning Independent Features with Adversarial Nets for Non-linear ICA", "authors": ["Philemon Brakel", "Yoshua Bengio"], "authorids": ["pbpop3@gmail.com", "yoshua.bengio@umontreal.ca"], "authors_source": "OpenReview API", "abstract": "Reliable measures of statistical dependence could potentially be useful tools for learning independent features and performing tasks like source separation using Independent Component Analysis (ICA).  Unfortunately, many of such measures, like the mutual information, are hard to estimate and optimize directly.  We propose to learn independent features with adversarial objectives (Goodfellow et al. 2014, Arjovsky et al. 2017) which optimize such measures implicitly.  These objectives compare samples from the joint distribution and the product of the marginals without the need to compute any probability densities. We also propose two methods for obtaining samples from the product of the marginals using either a simple resampling trick or a separate parametric distribution.  Our experiments show that this strategy can easily be applied to different types of model architectures and solve both linear and non-linear ICA problems.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ry2lpp_ez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper573/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe idea of ICA is constructing a mapping from dependent inputs to outputs (=the derived features) such that the outputs are as independent as possible. As the input/output densities are often not known and/or are intractable, natural independence measures such as mutual information are hard to estimate. In practice, the independence is characterized by certain functions of higher order moments -- leading to several alternatives in a zoo of independence objectives.  \n\nThe current paper makes the iteresting observation that independent features can also be computed via adversarial objectives. The key idea of adversarial training is adapted in this context as comparing samples from the joint distribution and the product of the marginals. \n\nTwo methods are proposed for drawing samples from the products of marginals. \nOne method is generating samples but permuting randomly the sample indices for individual marginals - this resampling mechanism generates approximately independent samples from the product distribution. The second method is essentially samples each marginal separately. \n\nThe approach is demonstrated in the solution of both linear and non-linear ICA problems.\n\nPositive:\nThe paper is well written and easy to follow on a higher level. GAN's provide a fresh look at nonlinear ICA and the paper is certainly thought provoking. \n\n\nNegative:\nMost of the space is devoted for reviewing related work and motivations, while the specifics of the method are described relatively short in section 4. There is no analysis and the paper is \nsomewhat anecdotal. The simulation results section is limited in scope. The sampling from product distribution method is somewhat obvious.\n\n\nQuestions:\n\n- The overcomplete audio source separation case is well known for audio and I could not understand why a convincing baseline can not be found. Is this due to nonlinear mixing?\nAs 26 channels and 6 channels are given, a simple regularization based method can be easily developed to provide a baseline performance, \n\n\n- The need for normalization in section 4 is surprising, as it obviously renders the outputs dependent. \n\n- Figure 1 may be misleading as h are not defined \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Thought provoking paper but lacks more detailed analysis", "rating": "6: Marginally above acceptance threshold", "review": "\nThe idea of ICA is constructing a mapping from dependent inputs to outputs (=the derived features) such that the outputs are as independent as possible. As the input/output densities are often not known and/or are intractable, natural independence measures such as mutual information are hard to estimate. In practice, the independence is characterized by certain functions of higher order moments -- leading to several alternatives in a zoo of independence objectives.  \n\nThe current paper makes the iteresting observation that independent features can also be computed via adversarial objectives. The key idea of adversarial training is adapted in this context as comparing samples from the joint distribution and the product of the marginals. \n\nTwo methods are proposed for drawing samples from the products of marginals. \nOne method is generating samples but permuting randomly the sample indices for individual marginals - this resampling mechanism generates approximately independent samples from the product distribution. The second method is essentially samples each marginal separately. \n\nThe approach is demonstrated in the solution of both linear and non-linear ICA problems.\n\nPositive:\nThe paper is well written and easy to follow on a higher level. GAN's provide a fresh look at nonlinear ICA and the paper is certainly thought provoking. \n\n\nNegative:\nMost of the space is devoted for reviewing related work and motivations, while the specifics of the method are described relatively short in section 4. There is no analysis and the paper is \nsomewhat anecdotal. The simulation results section is limited in scope. The sampling from product distribution method is somewhat obvious.\n\n\nQuestions:\n\n- The overcomplete audio source separation case is well known for audio and I could not understand why a convincing baseline can not be found. Is this due to nonlinear mixing?\nAs 26 channels and 6 channels are given, a simple regularization based method can be easily developed to provide a baseline performance, \n\n\n- The need for normalization in section 4 is surprising, as it obviously renders the outputs dependent. \n\n- Figure 1 may be misleading as h are not defined \n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511738612355}, {"id": "H1hlWndxM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper573/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a GAN variant for solving the nonlinear independent component analysis (ICA) problem. The method seems interesting, but the presentation has a severe lack of focus.\n\nFirst, the authors should focus their discussion instead of trying to address a broad range of ICA problems from linear to post-nonlinear (PNL) to nonlinear. I would highly recommend the authors to study the review \"Advances in Nonlinear Blind Source Separation\" by Jutten and Karhunen (2003/2004) to understand the problems they are trying to solve.\n\nLinear ICA is a solved problem and the authors do not seem to be able to add anything there, so I would recommend dropping that to save space for the more interesting material.\n\nPNL ICA is solvable and there are a number of algorithms proposed for it, some cited already in the above review, but also more recent ones. From this perspective, the presented comparison seems quite inadequate.\n\nFully general nonlinear ICA is ill-posed, as shown already by Darmois (1953, doi:10.2307/1401511). Given this, the authors should indicate more clearly what is their method expected to do. There are an infinite number of nonlinear ICA solutions - which one is the proposed method going to return and why is that relevant? There are fewer relevant comparisons here, but at least Lappalainen and Honkela (2000) seem to target the same problem as the proposed method.\n\nThe use of 6 dimensional example in the experiments is a very good start, as higher dimensions are quite different and much more interesting than very commonly used 2D examples.\n\nOne idea for evaluation: comparison with ground truth makes sense for PNL, but not so much for general nonlinear because of unidentifiability. For general nonlinear ICA you could consider evaluating the quality of the estimated low-dimensional data manifold or evaluating the mutual information of separated sources on new test data.\n\nUpdate after author feedback: thanks for the response and the revision. The revision seems more cosmetic and does not address the most significant issues so I do not see a need to change my evaluation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting nonlinear ICA method, but unfocused presentation and poor comparisons", "rating": "5: Marginally below acceptance threshold", "review": "The paper proposes a GAN variant for solving the nonlinear independent component analysis (ICA) problem. The method seems interesting, but the presentation has a severe lack of focus.\n\nFirst, the authors should focus their discussion instead of trying to address a broad range of ICA problems from linear to post-nonlinear (PNL) to nonlinear. I would highly recommend the authors to study the review \"Advances in Nonlinear Blind Source Separation\" by Jutten and Karhunen (2003/2004) to understand the problems they are trying to solve.\n\nLinear ICA is a solved problem and the authors do not seem to be able to add anything there, so I would recommend dropping that to save space for the more interesting material.\n\nPNL ICA is solvable and there are a number of algorithms proposed for it, some cited already in the above review, but also more recent ones. From this perspective, the presented comparison seems quite inadequate.\n\nFully general nonlinear ICA is ill-posed, as shown already by Darmois (1953, doi:10.2307/1401511). Given this, the authors should indicate more clearly what is their method expected to do. There are an infinite number of nonlinear ICA solutions - which one is the proposed method going to return and why is that relevant? There are fewer relevant comparisons here, but at least Lappalainen and Honkela (2000) seem to target the same problem as the proposed method.\n\nThe use of 6 dimensional example in the experiments is a very good start, as higher dimensions are quite different and much more interesting than very commonly used 2D examples.\n\nOne idea for evaluation: comparison with ground truth makes sense for PNL, but not so much for general nonlinear because of unidentifiability. For general nonlinear ICA you could consider evaluating the quality of the estimated low-dimensional data manifold or evaluating the mutual information of separated sources on new test data.\n\nUpdate after author feedback: thanks for the response and the revision. The revision seems more cosmetic and does not address the most significant issues so I do not see a need to change my evaluation.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511731443944}, {"id": "HyoEDdvxG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper573/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The focus of the paper is independent component analysis (ICA) and its nonlinear variants such as the post non-linear (PNL) ICA model. Motivated by the fact that estimating mutual information and similar dependency measures require density estimates and hard to optimize, the authors propose a Wasserstein GAN (generative adversarial network) based solution to tackle the problem, with illustrations on 6 (synthetic) and 3-dimemensional (audio) examples. The primary idea of the paper is to use the Wasserstein distance as an independence measure of the estimated source coordinates, and optimize it in a neural network (NN) framework.\n\nAlthough finding novel GAN applications is an exciting topic, I am not really convinced that ICA with the proposed Wasserstein GAN based technique fulfills this goal.\n \nBelow I detail my reasons:\n\n1)The ICA problem can be formulated as the minimization of pairwise mutual information [1] or one-dimensional entropy [2]. In other words, estimating the joint dependence of the source coordinates is not necessary; it is worthwhile to avoid it.\n\n2)The PNL ICA task can be efficiently tackled by first 'removing' the nonlinearity followed by classical linear ICA; see for example [3].\n\n3)Estimating information theoretic (IT) measures (mutual information, divergence) is a quite mature field with off-the-self techniques, see for example [4,5,6,8]. These methods do not estimate the underlying densities; it would be superfluous (and hard).\n\n4)Optimizing non-differentiable IT measures can computationally quite efficiently carried out in the ICA context by e.g., Givens rotations [7]; differentiable ICA cost functions can be robustly handled by Stiefel manifold methods; see for example [8,9].\n\n5)Section 3.1: This section is devoted to generating samples from the product of the marginals, even using separate generator networks. I do not see the necessity of these solutions; the subtask can be solved by independently shuffling all the coordinates of the sample.\n\n6)Experiments (Section 6): \ni) It seems to me that the proposed NN-based technique has some quite serious divergence issues: 'After discarding diverged models, ...' or 'Unfortunately, the model selection procedure also didn't identify good settings for the Anica-g model...'.\nii) The proposed method gives pretty comparable results to the chosen baselines (fastICA, PNLMISEP) on the selected small-dimensional tasks. In fact, [7,8,9] are likely to provide more accurate (fastICA is a simple kurtosis based method, which is \na somewhat crude 'estimate' of entropy) and faster estimates; see also 2).\n\nReferences:\n[1] Pierre Comon. Independent component analysis, a new concept? Signal Processing, 36:287-314, 1994.\n[2] Aapo Hyvarinen and Erkki Oja. Independent Component Analysis: Algorithms and Applications. Neural Networks, 13(4-5):411-30, 2000. \n[3] Andreas Ziehe, Motoaki Kawanabe, Stefan Harmeling, and Klaus-Robert Muller. Blind separation of postnonlinear mixtures using linearizing transformations and temporal decorrelation. Journal of Machine Learning Research, 4:1319-1338, 2003.\n[4] Barnabas Poczos, Liang Xiong, and Jeff Schneider. Nonparametric divergence: Estimation with applications to machine learning on distributions. In Conference on Uncertainty in Artificial Intelligence, pages 599-608, 2011.\n[5] Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Scholkopf, Alexander Smola. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13:723-773, 2012.\n[6] Alan Wisler, Visar Berisha, Andreas Spanias, Alfred O. Hero. A data-driven basis for direct estimation of functionals of distributions. TR, 2017. (https://arxiv.org/abs/1702.06516) \n[7] Erik G. Learned-Miller, John W. Fisher III. ICA using spacings estimates of entropy. Journal of Machine Learning Research, 4:1271-1295, 2003.\n[8] Francis R. Bach. Michael I. Jordan. Kernel Independent Component Analysis. Journal of Machine Learning Research 3: 1-48, 2002.\n[9] Hao Shen, Stefanie Jegelka and Arthur Gretton. Fast Kernel-Based Independent Component Analysis, IEEE Transactions on Signal Processing, 57:3498-3511, 2009.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Proposed Wasserstein GAN: not well-suited to ICA", "rating": "3: Clear rejection", "review": "The focus of the paper is independent component analysis (ICA) and its nonlinear variants such as the post non-linear (PNL) ICA model. Motivated by the fact that estimating mutual information and similar dependency measures require density estimates and hard to optimize, the authors propose a Wasserstein GAN (generative adversarial network) based solution to tackle the problem, with illustrations on 6 (synthetic) and 3-dimemensional (audio) examples. The primary idea of the paper is to use the Wasserstein distance as an independence measure of the estimated source coordinates, and optimize it in a neural network (NN) framework.\n\nAlthough finding novel GAN applications is an exciting topic, I am not really convinced that ICA with the proposed Wasserstein GAN based technique fulfills this goal.\n \nBelow I detail my reasons:\n\n1)The ICA problem can be formulated as the minimization of pairwise mutual information [1] or one-dimensional entropy [2]. In other words, estimating the joint dependence of the source coordinates is not necessary; it is worthwhile to avoid it.\n\n2)The PNL ICA task can be efficiently tackled by first 'removing' the nonlinearity followed by classical linear ICA; see for example [3].\n\n3)Estimating information theoretic (IT) measures (mutual information, divergence) is a quite mature field with off-the-self techniques, see for example [4,5,6,8]. These methods do not estimate the underlying densities; it would be superfluous (and hard).\n\n4)Optimizing non-differentiable IT measures can computationally quite efficiently carried out in the ICA context by e.g., Givens rotations [7]; differentiable ICA cost functions can be robustly handled by Stiefel manifold methods; see for example [8,9].\n\n5)Section 3.1: This section is devoted to generating samples from the product of the marginals, even using separate generator networks. I do not see the necessity of these solutions; the subtask can be solved by independently shuffling all the coordinates of the sample.\n\n6)Experiments (Section 6): \ni) It seems to me that the proposed NN-based technique has some quite serious divergence issues: 'After discarding diverged models, ...' or 'Unfortunately, the model selection procedure also didn't identify good settings for the Anica-g model...'.\nii) The proposed method gives pretty comparable results to the chosen baselines (fastICA, PNLMISEP) on the selected small-dimensional tasks. In fact, [7,8,9] are likely to provide more accurate (fastICA is a simple kurtosis based method, which is \na somewhat crude 'estimate' of entropy) and faster estimates; see also 2).\n\nReferences:\n[1] Pierre Comon. Independent component analysis, a new concept? Signal Processing, 36:287-314, 1994.\n[2] Aapo Hyvarinen and Erkki Oja. Independent Component Analysis: Algorithms and Applications. Neural Networks, 13(4-5):411-30, 2000. \n[3] Andreas Ziehe, Motoaki Kawanabe, Stefan Harmeling, and Klaus-Robert Muller. Blind separation of postnonlinear mixtures using linearizing transformations and temporal decorrelation. Journal of Machine Learning Research, 4:1319-1338, 2003.\n[4] Barnabas Poczos, Liang Xiong, and Jeff Schneider. Nonparametric divergence: Estimation with applications to machine learning on distributions. In Conference on Uncertainty in Artificial Intelligence, pages 599-608, 2011.\n[5] Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Scholkopf, Alexander Smola. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13:723-773, 2012.\n[6] Alan Wisler, Visar Berisha, Andreas Spanias, Alfred O. Hero. A data-driven basis for direct estimation of functionals of distributions. TR, 2017. (https://arxiv.org/abs/1702.06516) \n[7] Erik G. Learned-Miller, John W. Fisher III. ICA using spacings estimates of entropy. Journal of Machine Learning Research, 4:1271-1295, 2003.\n[8] Francis R. Bach. Michael I. Jordan. Kernel Independent Component Analysis. Journal of Machine Learning Research 3: 1-48, 2002.\n[9] Hao Shen, Stefanie Jegelka and Arthur Gretton. Fast Kernel-Based Independent Component Analysis, IEEE Transactions on Signal Processing, 57:3498-3511, 2009.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511651123102}], "openreview_url": "https://openreview.net/forum?id=ryykVe-0W", "arxiv_id": "1710.05050", "paper_pdf": "papers/ryykVe-0W.pdf", "paper_pdf_sha256": "a7615362a0e07145193cef06d29980c1607eb969c83baca531c847b78bc32a9c", "paper_pdf_bytes": 574546, "paper_pdf_source": "openreview", "code_url": "https://github.com/pbrakel/anica", "code_repository": "pbrakel/anica", "code_commit": "79d837addcd98ee9bab301d3966e4919cf9732f3", "code_archive": "repos/ryykVe-0W.zip", "code_archive_sha256": "a6ba4a1c3a2eb3eceae20e2aa9d3c86c4677f6883b3eb07a93d3175a5e4fd3f2", "code_archive_bytes": 21667, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 16, "github_languages": {"Python": 47021}, "github_archived": false, "github_pushed_at": "2017-10-17T13:02:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-independent-features-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LKv3bx610K", "year": 2026, "status": "rejected", "title": "Mitigating Fine-tuning Risks in LLMs via Safety-Aware Probing Optimization", "authors": ["Chengcan Wu", "Zhixin Zhang", "Zeming Wei", "Yihao Zhang", "Meng Sun"], "authorids": ["~Chengcan_Wu1", "~Zhixin_Zhang5", "~Zeming_Wei1", "~Yihao_Zhang8", "~Meng_Sun1"], "authors_source": "OpenReview API", "abstract": "The significant progress of large language models (LLMs) has led to remarkable achievements across numerous applications. However, their ability to generate harmful content has sparked substantial safety concerns. Despite the implementation of safety alignment techniques during the pre-training phase, recent research indicates that fine-tuning LLMs on adversarial or even benign data can inadvertently compromise their safety. In this paper, we re-examine the fundamental issue of why fine-tuning on non-harmful data still results in safety degradation. We introduce a safety-aware probing (SAP) optimization framework designed to mitigate the safety risks of fine-tuning LLMs. Specifically, SAP incorporates a safety-aware probe into the gradient propagation process, mitigating the model's risk of safety degradation by identifying potential pitfalls in gradient directions, thereby enhancing task-specific performance while successfully preserving model safety. Our extensive experimental results demonstrate that SAP effectively reduces harmfulness below the original fine-tuned model and achieves comparable test loss to standard fine-tuning methods. Our code is available in the supplementary materials.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "2XTyVgUrsy", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11683/Reviewer_Avsu"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper studies the safety alignment problem for LLMs. Specifically, it proposes a method to defend against malicious fine-tuning samples while keeping the model's utility score. What lies in the core the proposed method is to find a small parameter perturbation which discourages moving along harmful gradient direction when optimizing for the utility loss. The method is able to mitigate harmful gradient direction in utility loss gradient while not affecting the utility loss too much. Experimental results demonstrate the effectiveness of the proposed algorithm.", "review_text": "This paper studies the safety alignment problem for LLMs. Specifically, it proposes a method to defend against malicious fine-tuning samples while keeping the model's utility score. What lies in the core the proposed method is to find a small parameter perturbation which discourages moving along harmful gradient direction when optimizing for the utility loss. The method is able to mitigate harmful gradient direction in utility loss gradient while not affecting the utility loss too much. Experimental results demonstrate the effectiveness of the proposed algorithm.", "strengths": "1. The paper addresses an important problem in LLM alignment.\n2. The paper is well-written and easy to follow.\n3. This work offers some valuable insight by showing the correlation between the descend of the utility loss and that of the harmful loss. The proposed method is somewhat intuitive and easy to implement. The reported experimental result is good.", "weaknesses": "1. Insights of how and why the proposed method works: Though the construction of $L_{su}$ is somewhat intuitive, it is still unclear how and why the algorithm works well. Specifically, how can the algorithm achieve lower utility loss than the full SFT on utility dataset, while being biased constantly (from the perturbation) in its utility optimization process? If the perturbation is supposed to be very small, how can it lead to substantial decrease in harmfulness?\n\n2. Lacking critical baselines: it might be beneficial for the author to compare with simply adding the safety data (used in the proposed algorithm to compute negative harmful direction) into the utility dataset. The baseline can be interpreted as simply mixing the utility gradient with the negative harmful direction, which is very comparable to the proposed algorithm. This might lead to further insight into the performance of the algorithm.\n\nAs a result, I am overall hesitant to give an accept suggestion. However, I am happy to reconsider my recommendation if they are adequately addressed.", "questions": "1. In the proposed algorithm, is perturbation applied to the model parameter each iteration? If not, will it be a good/bad idea to apply the perturbation?\n\n2. How will the algorithm perform under no malicious data? I am considering a scenario where this method is just applied to enhance safety in normal utility fine-tuning process.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the safety alignment problem for LLMs. Specifically, it proposes a method to defend against malicious fine-tuning samples while keeping the model's utility score. What lies in the core the proposed method is to find a small parameter perturbation which discourages moving along harmful gradient direction when optimizing for the utility loss. The method is able to mitigate harmful gradient direction in utility loss gradient while not affecting the utility loss too much. Experimental results demonstrate the effectiveness of the proposed algorithm.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper addresses an important problem in LLM alignment.\n2. The paper is well-written and easy to follow.\n3. This work offers some valuable insight by showing the correlation between the descend of the utility loss and that of the harmful loss. The proposed method is somewhat intuitive and easy to implement. The reported experimental result is good.", "weaknesses": "1. Insights of how and why the proposed method works: Though the construction of $L_{su}$ is somewhat intuitive, it is still unclear how and why the algorithm works well. Specifically, how can the algorithm achieve lower utility loss than the full SFT on utility dataset, while being biased constantly (from the perturbation) in its utility optimization process? If the perturbation is supposed to be very small, how can it lead to substantial decrease in harmfulness?\n\n2. Lacking critical baselines: it might be beneficial for the author to compare with simply adding the safety data (used in the proposed algorithm to compute negative harmful direction) into the utility dataset. The baseline can be interpreted as simply mixing the utility gradient with the negative harmful direction, which is very comparable to the proposed algorithm. This might lead to further insight into the performance of the algorithm.\n\nAs a result, I am overall hesitant to give an accept suggestion. However, I am happy to reconsider my recommendation if they are adequately addressed.", "questions": "1. In the proposed algorithm, is perturbation applied to the model parameter each iteration? If not, will it be a good/bad idea to apply the perturbation?\n\n2. How will the algorithm perform under no malicious data? I am considering a scenario where this method is just applied to enhance safety in normal utility fine-tuning process.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "no ethics concern.", "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762256141337}, {"id": "gsQ8qdD1hj", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11683/Reviewer_3MFr"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces Safety-Aware Probing (SAP), a lightweight optimization that inserts a small safety-aware probe into hidden states during gradient propagation to steer updates away from harmful directions while preserving task utility. SAP is motivated by an observed entanglement between safety-critical and usefulness-critical gradient directions; it maximizes a safe-useful objective to find a probe  that biases each update toward safer regions.", "review_text": "The paper introduces Safety-Aware Probing (SAP), a lightweight optimization that inserts a small safety-aware probe into hidden states during gradient propagation to steer updates away from harmful directions while preserving task utility. SAP is motivated by an observed entanglement between safety-critical and usefulness-critical gradient directions; it maximizes a safe-useful objective to find a probe  that biases each update toward safer regions.", "strengths": "- The proposed method is simple yet principled: it treats safety as gradient-space steering and remains broadly compatible with standard fine-tuning.\n- Empirically, it improves utility while reducing harmfulness, increases robustness to poisoning and adversarial fine-tuning, and composes with other defenses, making it practical for deployment.", "weaknesses": "The proposed method introduces too many hyperparameters (e.g., α, β, ϵ, probe layers), which increases tuning complexity and reduces reproducibility.", "questions": "Q1. Did you mean the following?\n\n“Our experiments show that SAP achieves better useful loss while significantly **decreasing model safety”**\n\n→ “Our experiments show that SAP achieves better useful loss while significantly **improving model safety**”\n\nQ2. Could you clarify the captions for Figures 2 and 3?\n\n- Figure 2 does not specify which harmful or useful\ndatasets were used.\n- Figure 3 does not clarify the definition of the useful-critical notation nor specify which harmful dataset was used.\n\nQ3. Are $\\alpha$, $\\beta$, $\\epsilon$, and probe layers the same for every dataset?\n\nQ4. Is there a reason for choosing 2,000 examples for $D_{useful}$ and 50 for $D_{safe/harmful}$?\n\nQ5. Aren’t these models instruction-tuned models, such as Llama-7B-Chat?\n\nQ6. Why are the Booster and Vaccine baselines not included in the results after Table 1?\n\nFormatting issue:\n\n- The repeated inclusion of the Alpaca dataset in Table 2 is redundant.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Safety-Aware Probing (SAP), a lightweight optimization that inserts a small safety-aware probe into hidden states during gradient propagation to steer updates away from harmful directions while preserving task utility. SAP is motivated by an observed entanglement between safety-critical and usefulness-critical gradient directions; it maximizes a safe-useful objective to find a probe  that biases each update toward safer regions.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The proposed method is simple yet principled: it treats safety as gradient-space steering and remains broadly compatible with standard fine-tuning.\n- Empirically, it improves utility while reducing harmfulness, increases robustness to poisoning and adversarial fine-tuning, and composes with other defenses, making it practical for deployment.", "weaknesses": "The proposed method introduces too many hyperparameters (e.g., α, β, ϵ, probe layers), which increases tuning complexity and reduces reproducibility.", "questions": "Q1. Did you mean the following?\n\n“Our experiments show that SAP achieves better useful loss while significantly **decreasing model safety”**\n\n→ “Our experiments show that SAP achieves better useful loss while significantly **improving model safety**”\n\nQ2. Could you clarify the captions for Figures 2 and 3?\n\n- Figure 2 does not specify which harmful or useful\ndatasets were used.\n- Figure 3 does not clarify the definition of the useful-critical notation nor specify which harmful dataset was used.\n\nQ3. Are $\\alpha$, $\\beta$, $\\epsilon$, and probe layers the same for every dataset?\n\nQ4. Is there a reason for choosing 2,000 examples for $D_{useful}$ and 50 for $D_{safe/harmful}$?\n\nQ5. Aren’t these models instruction-tuned models, such as Llama-7B-Chat?\n\nQ6. Why are the Booster and Vaccine baselines not included in the results after Table 1?\n\nFormatting issue:\n\n- The repeated inclusion of the Alpaca dataset in Table 2 is redundant.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762181340017}, {"id": "kCxOR1iDMm", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11683/Reviewer_kkBd"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper shows that large language models can lose their safety even when fine-tuned on harmless data, because the gradients that improve task performance are often entangled with those that reduce safety. To address this, the authors propose Safety-Aware Probing (SAP), a training method that adds a small hidden-state probe to steer optimization away from harmful directions while still improving task performance. SAP does not require changing the dataset or model architecture and works across different fine-tuning setups.", "review_text": "This paper shows that large language models can lose their safety even when fine-tuned on harmless data, because the gradients that improve task performance are often entangled with those that reduce safety. To address this, the authors propose Safety-Aware Probing (SAP), a training method that adds a small hidden-state probe to steer optimization away from harmful directions while still improving task performance. SAP does not require changing the dataset or model architecture and works across different fine-tuning setups.", "strengths": "The experiments are extensive and well designed, have evaluations on three different models, three instruction-following datasets, five reasoning benchmarks, and poisoned, adversarial fine-tuning settings. \nThe paper is also clearly written and well-organized.", "weaknesses": "* Gradient analysis and evaluation are somewhat limited. The paper shows cosine similarity between usefulness and safety gradients, but does not fully explain why the directions align or provide deeper theoretical insight. The Harmful score relies on a single moderation model, with no additional metrics such as jailbreak success rate, or LLM-as-judge evaluation. \n* Cost analysis is needed. SAP increases training time by 2x~3x, but the paper briefly labels this as acceptable without discussing practical implications or scalability. \n* Limited discussion of adaptive attacks. The adversarial fine-tuning experiment does not consider attackers aware of SAP.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper shows that large language models can lose their safety even when fine-tuned on harmless data, because the gradients that improve task performance are often entangled with those that reduce safety. To address this, the authors propose Safety-Aware Probing (SAP), a training method that adds a small hidden-state probe to steer optimization away from harmful directions while still improving task performance. SAP does not require changing the dataset or model architecture and works across different fine-tuning setups.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "The experiments are extensive and well designed, have evaluations on three different models, three instruction-following datasets, five reasoning benchmarks, and poisoned, adversarial fine-tuning settings. \nThe paper is also clearly written and well-organized.", "weaknesses": "* Gradient analysis and evaluation are somewhat limited. The paper shows cosine similarity between usefulness and safety gradients, but does not fully explain why the directions align or provide deeper theoretical insight. The Harmful score relies on a single moderation model, with no additional metrics such as jailbreak success rate, or LLM-as-judge evaluation. \n* Cost analysis is needed. SAP increases training time by 2x~3x, but the paper briefly labels this as acceptable without discussing practical implications or scalability. \n* Limited discussion of adaptive attacks. The adversarial fine-tuning experiment does not consider attackers aware of SAP.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762139930482}, {"id": "bri0nAmdDr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11683/Reviewer_PBzG"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper solves the LLM safety degradation caused by fine-tuning. It hypothesizes that usefulness-critical and safety-critical gradient directions are entangled and proposes Safety-Aware Probing (SAP).  At each step, SAP estimates a safety-critical direction using a contrastive safety loss, constructs a “harmful” update, and then learn a small hidden-state probe that maximizes a safe-useful loss, encouraging downstream weight updates to avoid harmful regions. Across three 7b models on benign and adversarial tasks, SAP reduces harmfulness scores while keeping task loss/metrics near SFT. Additional experiments demonstrate the effectiveness of different settings. However, the scalability is not validated and remains questionable.", "review_text": "The paper solves the LLM safety degradation caused by fine-tuning. It hypothesizes that usefulness-critical and safety-critical gradient directions are entangled and proposes Safety-Aware Probing (SAP).  At each step, SAP estimates a safety-critical direction using a contrastive safety loss, constructs a “harmful” update, and then learn a small hidden-state probe that maximizes a safe-useful loss, encouraging downstream weight updates to avoid harmful regions. Across three 7b models on benign and adversarial tasks, SAP reduces harmfulness scores while keeping task loss/metrics near SFT. Additional experiments demonstrate the effectiveness of different settings. However, the scalability is not validated and remains questionable.", "strengths": "1. The proposed SAP is well motivated. The hypothesis is validated by experiments.\n2. Experiments over 3 models and several datasets in different settings demonstrate promising performance.\n3. It provides sensible ablations (e.g., which layers to probe and learning-rate sensitivity) and report costs (time and memory).", "weaknesses": "1. The first claimed contribution of validating the hypothesis has been explored in prior works, such as [1, 2]. The cosine similarity approach is similar to SafeLora [1].\n\n2. The authors claim in the introduction that their work “*has better scalability since it can be incorporated into various fine-tuning paradigms rather than being limited to LoRA.*”  However, this claim is not supported by experimental evidence, such as larger models or fully fine-tuning. All experiments are conducted on 7B models with LoRA.  Furthermore, this claim is questionable, as SAP requires extra gradient estimation, which is related to the number of trainable parameters.\n\n3. The time overhead of SAP is non-trivial, approximately 2.5 times that of all baselines. While Appendix B.5 includes a comparison with LISA, the discussion does not adequately address the overhead relative to other baselines. Considering the performance gain, the practical value of applying this method remains questionable.\n\n4. Figure 1 lacks clarity and should be improved to better convey the procedure of SAP.\n\n\n**Reference**\n\n[1]  Safelora: The silver lining of reducing safety risks when finetuning large language models. NeurIPS. 2024\n\n[2] Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets. ICML. 2025", "questions": "1. What is the time cost for SAP when applying to larger models and fully fine-tuning settings?\n2. What is the portion or number of safety examples for baselines, such as safeinstr?\n3. Why not use fewer, smaller probe layers instead of probing 10 layers, given the high time cost?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper solves the LLM safety degradation caused by fine-tuning. It hypothesizes that usefulness-critical and safety-critical gradient directions are entangled and proposes Safety-Aware Probing (SAP).  At each step, SAP estimates a safety-critical direction using a contrastive safety loss, constructs a “harmful” update, and then learn a small hidden-state probe that maximizes a safe-useful loss, encouraging downstream weight updates to avoid harmful regions. Across three 7b models on benign and adversarial tasks, SAP reduces harmfulness scores while keeping task loss/metrics near SFT. Additional experiments demonstrate the effectiveness of different settings. However, the scalability is not validated and remains questionable.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The proposed SAP is well motivated. The hypothesis is validated by experiments.\n2. Experiments over 3 models and several datasets in different settings demonstrate promising performance.\n3. It provides sensible ablations (e.g., which layers to probe and learning-rate sensitivity) and report costs (time and memory).", "weaknesses": "1. The first claimed contribution of validating the hypothesis has been explored in prior works, such as [1, 2]. The cosine similarity approach is similar to SafeLora [1].\n\n2. The authors claim in the introduction that their work “*has better scalability since it can be incorporated into various fine-tuning paradigms rather than being limited to LoRA.*”  However, this claim is not supported by experimental evidence, such as larger models or fully fine-tuning. All experiments are conducted on 7B models with LoRA.  Furthermore, this claim is questionable, as SAP requires extra gradient estimation, which is related to the number of trainable parameters.\n\n3. The time overhead of SAP is non-trivial, approximately 2.5 times that of all baselines. While Appendix B.5 includes a comparison with LISA, the discussion does not adequately address the overhead relative to other baselines. Considering the performance gain, the practical value of applying this method remains questionable.\n\n4. Figure 1 lacks clarity and should be improved to better convey the procedure of SAP.\n\n\n**Reference**\n\n[1]  Safelora: The silver lining of reducing safety risks when finetuning large language models. NeurIPS. 2024\n\n[2] Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets. ICML. 2025", "questions": "1. What is the time cost for SAP when applying to larger models and fully fine-tuning settings?\n2. What is the portion or number of safety examples for baselines, such as safeinstr?\n3. Why not use fewer, smaller probe layers instead of probing 10 layers, given the high time cost?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761922390005}], "openreview_url": "https://openreview.net/forum?id=LKv3bx610K", "arxiv_id": "2505.16737", "paper_pdf": "papers/LKv3bx610K.pdf", "paper_pdf_sha256": "6825b5f34b0a8ddf68d8606fdab0ec4a6a579bef466be8f3538d9d7e7d8d0a22", "paper_pdf_bytes": 4164185, "paper_pdf_source": "openreview", "code_url": "https://github.com/ChengcanWu/SAP", "code_repository": "ChengcanWu/SAP", "code_commit": "487edc3374f24581399040beac6209c388a463ee", "code_archive": "repos/LKv3bx610K.zip", "code_archive_sha256": "19d021f6b6ea33a4069e1ae54ae35569bd44cedf368b9c92978febc6637e57ed", "code_archive_bytes": 12549, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 19, "github_languages": {"Python": 19396}, "github_archived": false, "github_pushed_at": "2026-04-23T15:02:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mitigating-fine-tuning-risks-in-llms-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dq3keisMjT", "year": 2025, "status": "rejected", "title": "Phase Transitions in the Output Distribution of Large Language Models", "authors": ["Julian Arnold", "Flemming Holtorf", "Frank Schäfer", "Niels Lörch"], "authorids": ["~Julian_Arnold2", "~Flemming_Holtorf1", "~Frank_Schäfer1", "~Niels_Lörch1"], "authors_source": "OpenReview API", "abstract": "In a physical system, changing parameters such as temperature can induce a phase transition: an abrupt change from one state of matter to another. Analogous phenomena have recently been observed in large language models. Typically, the task of identifying phase transitions requires human analysis and some prior understanding of the system to narrow down which low-dimensional properties to monitor and analyze. Statistical methods for the automated detection of phase transitions from data have recently been proposed within the physics community. These methods are largely system agnostic and, as shown here, can be adapted to study the behavior of large language models. In particular, we quantify distributional changes in the generated output via statistical distances, which can be efficiently estimated with access to the probability distribution over next-tokens. This versatile approach is capable of discovering new phases of behavior and unexplored transitions -- an ability that is particularly exciting in light of the rapid development of language models and their emergent capabilities.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "8AgZuqUWk2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2958/Reviewer_aK6V"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper investigates phase transitions in the output distributions of language models – that is, situations where the output distribution of an LM abruptly changes when a numerical parameter crosses some value. Changes are measured using f-divergences, where f is a specific rational function. The paper investigates three example parameters: (i) a number appearing in the context [concretely, a number “T” as in “T is larger than 42”], (ii) the softmax temperature in the output, (iii) the number of training epochs.", "review_text": "This paper investigates phase transitions in the output distributions of language models – that is, situations where the output distribution of an LM abruptly changes when a numerical parameter crosses some value. Changes are measured using f-divergences, where f is a specific rational function. The paper investigates three example parameters: (i) a number appearing in the context [concretely, a number “T” as in “T is larger than 42”], (ii) the softmax temperature in the output, (iii) the number of training epochs.", "strengths": "- The paper is innovative, linking physics concepts with language models\n- The paper studies multiple open LLMs\n- The proposed metric is theoretically grounded in a link to the Fisher information\n- I found the execution to be rigorous and writing to be clear\n- Findings can be of potential interest, e.g. it might be surprising that there phase transitions are identified at some training epochs", "weaknesses": "From my perspective, the primary weakness of the paper is the lack of any external validation of the proposed approach. While the paper proposes a method for measuring phase transitions, it remained unclear to me how to validate the findings. The findings might in principle depend a lot on the choice of the divergence. Is there any independent way of validating that the outcomes of the proposed method are useful or interesting? Is there a clear use case? Is there a principled relation to a theoretical property of language, or of transformer networks? Without this, it is not clear to me how to (i) substantiate the claim that the proposed metric captures transitions in an interesting way, (ii) make the paper of interest to the community. I do believe that such things are possible, and that the proposed method can be of substantial interest, but it would take further work to do this, and to validate the proposed approach specifically, compared to various other ways one could measure phase transitions (e.g., with other f-divergences)", "questions": "See Weaknesses: How do we know the method captures something interesting? Is there external validation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates phase transitions in the output distributions of language models – that is, situations where the output distribution of an LM abruptly changes when a numerical parameter crosses some value. Changes are measured using f-divergences, where f is a specific rational function. The paper investigates three example parameters: (i) a number appearing in the context [concretely, a number “T” as in “T is larger than 42”], (ii) the softmax temperature in the output, (iii) the number of training epochs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper is innovative, linking physics concepts with language models\n- The paper studies multiple open LLMs\n- The proposed metric is theoretically grounded in a link to the Fisher information\n- I found the execution to be rigorous and writing to be clear\n- Findings can be of potential interest, e.g. it might be surprising that there phase transitions are identified at some training epochs", "weaknesses": "From my perspective, the primary weakness of the paper is the lack of any external validation of the proposed approach. While the paper proposes a method for measuring phase transitions, it remained unclear to me how to validate the findings. The findings might in principle depend a lot on the choice of the divergence. Is there any independent way of validating that the outcomes of the proposed method are useful or interesting? Is there a clear use case? Is there a principled relation to a theoretical property of language, or of transformer networks? Without this, it is not clear to me how to (i) substantiate the claim that the proposed metric captures transitions in an interesting way, (ii) make the paper of interest to the community. I do believe that such things are possible, and that the proposed method can be of substantial interest, but it would take further work to do this, and to validate the proposed approach specifically, compared to various other ways one could measure phase transitions (e.g., with other f-divergences)", "questions": "See Weaknesses: How do we know the method captures something interesting? Is there external validation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730732438453}, {"id": "U4ROQH9HS4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2958/Reviewer_7XoD"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The authors propose statistical methods to automatically detect phase transitions in large language models. They provide empirical evidence for three types of phase transitions that occur when varying the input, temperature, and training epoch.", "review_text": "The authors propose statistical methods to automatically detect phase transitions in large language models. They provide empirical evidence for three types of phase transitions that occur when varying the input, temperature, and training epoch.", "strengths": "The connection between phase transitions in physics and large language models is interesting!", "weaknesses": "The paper's main challenge and contribution are unclear. While the authors cite several previous works that have analyzed phase transitions, their contribution appears to be proposing new statistical detection methods. However, if this is their primary contribution, the experimental validation is insufficient - they use only a limited number of prompts and test cases. More extensive experiments with diverse prompts would be needed to validate their methods.\n\nAlternatively, if the paper aims to establish theoretical connections to physics, it requires more accessible and thorough explanations. For example, while they discuss the link between the Ising model and language models, readers unfamiliar with physics may struggle to follow the comparison without mathematical formalization. Also, the paper would be more effective if it included parallel visualizations comparing phase transitions in physical systems (e.g., heat capacity plots for water) with their language model findings.", "questions": "1. What advantages does linear dissimilarity offer compared to total variation distance and Jensen-Shannon divergence? Figure 1 shows that all types of divergence can capture the dissimilarities.\n2. It would be interesting to examine how phase transitions occur given combinations of input and temperature, similar to how water's phase transitions depend on both pressure and temperature.\n3. What are the different properties of phases before and after transition? Can these phases be analogous to physical states like 'solid' or 'gas'? Is there a human-interpretable explanation for each phase, particularly regarding temperature?\n4. What is the definition of a phase transition? In Figure 2, the linear dissimilarity of 0.4 for the Llama3 base model appears quite high. Can this be considered a phase transition?\n5. Is panel (b) in Figure 4 truly a zoomed-in plot of panel (a)? The black line and y-axis values appear to be different.\n6. In Figure 5(b), why did you mention that the clear peaks near epochs 20K, 40K, and 80K are outliers?\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose statistical methods to automatically detect phase transitions in large language models. They provide empirical evidence for three types of phase transitions that occur when varying the input, temperature, and training epoch.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The connection between phase transitions in physics and large language models is interesting!", "weaknesses": "The paper's main challenge and contribution are unclear. While the authors cite several previous works that have analyzed phase transitions, their contribution appears to be proposing new statistical detection methods. However, if this is their primary contribution, the experimental validation is insufficient - they use only a limited number of prompts and test cases. More extensive experiments with diverse prompts would be needed to validate their methods.\n\nAlternatively, if the paper aims to establish theoretical connections to physics, it requires more accessible and thorough explanations. For example, while they discuss the link between the Ising model and language models, readers unfamiliar with physics may struggle to follow the comparison without mathematical formalization. Also, the paper would be more effective if it included parallel visualizations comparing phase transitions in physical systems (e.g., heat capacity plots for water) with their language model findings.", "questions": "1. What advantages does linear dissimilarity offer compared to total variation distance and Jensen-Shannon divergence? Figure 1 shows that all types of divergence can capture the dissimilarities.\n2. It would be interesting to examine how phase transitions occur given combinations of input and temperature, similar to how water's phase transitions depend on both pressure and temperature.\n3. What are the different properties of phases before and after transition? Can these phases be analogous to physical states like 'solid' or 'gas'? Is there a human-interpretable explanation for each phase, particularly regarding temperature?\n4. What is the definition of a phase transition? In Figure 2, the linear dissimilarity of 0.4 for the Llama3 base model appears quite high. Can this be considered a phase transition?\n5. Is panel (b) in Figure 4 truly a zoomed-in plot of panel (a)? The black line and y-axis values appear to be different.\n6. In Figure 5(b), why did you mention that the clear peaks near epochs 20K, 40K, and 80K are outliers?\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730660075705}, {"id": "7yRjzKHPdt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2958/Reviewer_mqhi"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper looks at LLMs from the perspective of statistical physics. It proposes several ways of measuring changes in probability distributions, where large changes are interpreted as signatures for \"phase transitions\". It applies these measures in 3 case studies. In the first, change is measured as a simple prompt, containing a numerical value, is varied. In the second, the temperature parameter in the softmax is varied. In the third, changes across checkpoints during training are recorded. For each of these case studies, graphs are produced, showing that there are ranges for the varied parameter where changes are larger.", "review_text": "This paper looks at LLMs from the perspective of statistical physics. It proposes several ways of measuring changes in probability distributions, where large changes are interpreted as signatures for \"phase transitions\". It applies these measures in 3 case studies. In the first, change is measured as a simple prompt, containing a numerical value, is varied. In the second, the temperature parameter in the softmax is varied. In the third, changes across checkpoints during training are recorded. For each of these case studies, graphs are produced, showing that there are ranges for the varied parameter where changes are larger.", "strengths": "This paper addresses an interesting problem: whether changes in output distributions of LLMs, as one varies an underlying parameter, can be used to detect interesting changes in the underlying mechanisms. It reviews some potentially useful measures from statistical physics, that place widely used measures in NLP and ML such as the KL divergence in a broader context.\n\nThe authors have looked at many different models and checkpoints, and obtaining the numbers reported in this paper / creating the graphs must have been quite a lot of work.", "weaknesses": "The first half of the paper is devoted to reviewing mathematical tools to measure distances between distributions. In the second half of the paper, the distance measures are used in some simple case studies. The main weakness of the paper is that each of these case studies stops where it starts to become interesting, such that we learn nothing new about the behaviour or underlying mechanisms of LLMs, and that the usefulness of the mathematical tools from the first half is never shown.\n\nI don't have a lot of comments on the distance measures from the first 3-4 pages. Countless papers in ML and NLP already measure distances between probability distributions (usually with KL and computationally tractable approximations such as ELBO). It would be surprising and potentially important if tools developed for studying phase transitions in physics offer a new light on these issues. I cannot judge how novel the framework the authors offer is for ML, but novel or not, I will just note that it is problematic that the authors do not in any way show that their favourite tools are in fact more useful than better known tools.\n\nThis usefulness would have to be established in the case studies in the second half, but unfortunately isn't. In case study (1), figure 1 and 2 reveals that LLMs build up internal representations for the numerical value of integers, that some do worse than others and that tokenization sometimes messes up the representations. None of that is news for people that have studied LLMs, and especially for people that have read papers on numerosity representations. And, moreover, it has nothing to do with phase transitions, unless that term is given such a broad definition that it looses its value. The fact that output distributions change significantly around the point where the input necessitates a different response is simply a consequence of designing and training networks to predict categorical output.\n\nIn case study (2), we get some graphs of how sensitive output distributions are for different values of the temperature parameter in the output softmax. The graphs look pretty, but what have we learned? My conclusions from reading this section is that for high temperature, distributions do not change (because they're random anyway), and that for low temperatures the exact value might have a quite big impact on the behaviour of the LLM (but that we don't know beforehand at which value and which effect). But all of that is pretty much the starting point of the section.\n\nIn case study (3), the paper studies training dynamics in the Pythia model suite. Figure 5 looks interesting, especially the different timing of transitions in layer 0-2, 3, 4 and 5 (although it is not clear to me which Pythia model this is exactly, and why are you only plotting layer 0-5). But then the text (line 449) explains that some of the peaks (in 5b at least) might just be noise, and not reflecting any macroscopic phase changes. So again, we are presented some graphs that may or may not show that some interesting changes in the underlying representations might be happening, here at specific moments during training. But the paper presents no effort to actually confirm that these changes are indeed \"phase transitions\" with a different method, or in any way characterize the representation before and after the \"phase transition\".\n\nAll in all, there is clearly a lot of effort and technical know-how that has gone into this project, but it suffers from a lack of engagement with the extensive literature on LLMs, their learned representations and probability distributions, the qualitative changes in these representations, and the potential usefulness of concepts like \"phase transitions\" and \"emergence\" in understanding those changes. In section 4, the authors cite a lot of relevant papers (e.g., Wei et al 22, Schaeffer et al. 24, Chen et al. 23), but without actually engaging with them. For this project to really have some impact on ML/NLP, more is required than plotting the values of a measure for lots of different models and lots of different parameters: you need to show that high values on these measures indeed reveal something important, by showing this importance with a different method yourself or by linking it to a result from an existing paper.\n\nEDIT: I have read the general response and the responses to my and other reviews. Although the direct response to my review was brief, overall I appreciate that the authors have put quite a bit of work into answering questions and concerns from the reviewers. It's clear that the authors are knowledgable about work of phase transitions in physics; however, their responses don't take away my concern that the main work -- to demonstrate the usefulness of porting this framework to ML -- still needs to be done. I will therefore maintain my score (as well as my recommendation for the future to engage with some existing work in ML, and investigate whether new tools can shed new light on it!).", "questions": "Questions:\nWhich of the Pythia models is used for the graphs in figure 5? Why are you only plotting layer 0-5?\n\nEDIT: Thanks for your answer.\n\nSuggestions:\n\nPerhaps interesting work to engage with related to case study 1: Hanna, M., Liu, O., & Variengien, A. (2024). How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. Advances in Neural Information Processing Systems, 36. This paper reveals a circuitry involved in building up a numerical prediction; the emergence of such a circuitry during training, or the switching on or off of such a circuitry due to contextual cues (in-context learning) would be an actually interesting phase transition.\n\nRelated to case study 3: Nanda, N., Chan, L., Lieberum, T., Smith, J., & Steinhardt, J. (2023). Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217. This paper shows you can predict the *sudden* emergence of a specific numerical ability (a phase transition) by tracking the *gradual* evaluation of specific \"progress measures\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper looks at LLMs from the perspective of statistical physics. It proposes several ways of measuring changes in probability distributions, where large changes are interpreted as signatures for \"phase transitions\". It applies these measures in 3 case studies. In the first, change is measured as a simple prompt, containing a numerical value, is varied. In the second, the temperature parameter in the softmax is varied. In the third, changes across checkpoints during training are recorded. For each of these case studies, graphs are produced, showing that there are ranges for the varied parameter where changes are larger.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper addresses an interesting problem: whether changes in output distributions of LLMs, as one varies an underlying parameter, can be used to detect interesting changes in the underlying mechanisms. It reviews some potentially useful measures from statistical physics, that place widely used measures in NLP and ML such as the KL divergence in a broader context.\n\nThe authors have looked at many different models and checkpoints, and obtaining the numbers reported in this paper / creating the graphs must have been quite a lot of work.", "weaknesses": "The first half of the paper is devoted to reviewing mathematical tools to measure distances between distributions. In the second half of the paper, the distance measures are used in some simple case studies. The main weakness of the paper is that each of these case studies stops where it starts to become interesting, such that we learn nothing new about the behaviour or underlying mechanisms of LLMs, and that the usefulness of the mathematical tools from the first half is never shown.\n\nI don't have a lot of comments on the distance measures from the first 3-4 pages. Countless papers in ML and NLP already measure distances between probability distributions (usually with KL and computationally tractable approximations such as ELBO). It would be surprising and potentially important if tools developed for studying phase transitions in physics offer a new light on these issues. I cannot judge how novel the framework the authors offer is for ML, but novel or not, I will just note that it is problematic that the authors do not in any way show that their favourite tools are in fact more useful than better known tools.\n\nThis usefulness would have to be established in the case studies in the second half, but unfortunately isn't. In case study (1), figure 1 and 2 reveals that LLMs build up internal representations for the numerical value of integers, that some do worse than others and that tokenization sometimes messes up the representations. None of that is news for people that have studied LLMs, and especially for people that have read papers on numerosity representations. And, moreover, it has nothing to do with phase transitions, unless that term is given such a broad definition that it looses its value. The fact that output distributions change significantly around the point where the input necessitates a different response is simply a consequence of designing and training networks to predict categorical output.\n\nIn case study (2), we get some graphs of how sensitive output distributions are for different values of the temperature parameter in the output softmax. The graphs look pretty, but what have we learned? My conclusions from reading this section is that for high temperature, distributions do not change (because they're random anyway), and that for low temperatures the exact value might have a quite big impact on the behaviour of the LLM (but that we don't know beforehand at which value and which effect). But all of that is pretty much the starting point of the section.\n\nIn case study (3), the paper studies training dynamics in the Pythia model suite. Figure 5 looks interesting, especially the different timing of transitions in layer 0-2, 3, 4 and 5 (although it is not clear to me which Pythia model this is exactly, and why are you only plotting layer 0-5). But then the text (line 449) explains that some of the peaks (in 5b at least) might just be noise, and not reflecting any macroscopic phase changes. So again, we are presented some graphs that may or may not show that some interesting changes in the underlying representations might be happening, here at specific moments during training. But the paper presents no effort to actually confirm that these changes are indeed \"phase transitions\" with a different method, or in any way characterize the representation before and after the \"phase transition\".\n\nAll in all, there is clearly a lot of effort and technical know-how that has gone into this project, but it suffers from a lack of engagement with the extensive literature on LLMs, their learned representations and probability distributions, the qualitative changes in these representations, and the potential usefulness of concepts like \"phase transitions\" and \"emergence\" in understanding those changes. In section 4, the authors cite a lot of relevant papers (e.g., Wei et al 22, Schaeffer et al. 24, Chen et al. 23), but without actually engaging with them. For this project to really have some impact on ML/NLP, more is required than plotting the values of a measure for lots of different models and lots of different parameters: you need to show that high values on these measures indeed reveal something important, by showing this importance with a different method yourself or by linking it to a result from an existing paper.\n\nEDIT: I have read the general response and the responses to my and other reviews. Although the direct response to my review was brief, overall I appreciate that the authors have put quite a bit of work into answering questions and concerns from the reviewers. It's clear that the authors are knowledgable about work of phase transitions in physics; however, their responses don't take away my concern that the main work -- to demonstrate the usefulness of porting this framework to ML -- still needs to be done. I will therefore maintain my score (as well as my recommendation for the future to engage with some existing work in ML, and investigate whether new tools can shed new light on it!).", "questions": "Questions:\nWhich of the Pythia models is used for the graphs in figure 5? Why are you only plotting layer 0-5?\n\nEDIT: Thanks for your answer.\n\nSuggestions:\n\nPerhaps interesting work to engage with related to case study 1: Hanna, M., Liu, O., & Variengien, A. (2024). How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. Advances in Neural Information Processing Systems, 36. This paper reveals a circuitry involved in building up a numerical prediction; the emergence of such a circuitry during training, or the switching on or off of such a circuitry due to contextual cues (in-context learning) would be an actually interesting phase transition.\n\nRelated to case study 3: Nanda, N., Chan, L., Lieberum, T., Smith, J., & Steinhardt, J. (2023). Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217. This paper shows you can predict the *sudden* emergence of a specific numerical ability (a phase transition) by tracking the *gradual* evaluation of specific \"progress measures\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730635150478}, {"id": "qOw2U15GP9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2958/Reviewer_4pP5"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper examines phase transition in the output distributions of large language models(LLMs) and proposes a novel approach that applies a method used in physical systems for detecting phase transitions to LLMs. In physical systems, phase transitions occur as control parameters, such as temperature and pressure, change. These transitions often depend only on macroscopic parameters and show universal properties. However, phase transitions in LLMs have been shown to depend on detailed ``microscopic\" inputs, such as prompts, with phase transitions occurring at different temperatures depending on the prompt. This difference suggests that a distinct mechanism may be at work in LLMs compared to physical systems, and the fact that the \"phase of behavior\" in LLMs is induced by prompts is particularly interesting. This study provides a new method for automatically detecting the phase transitions in LLMs through black-box analysis based on output distribution, yielding new insights into the learning behaviors of LLMs.", "review_text": "This paper examines phase transition in the output distributions of large language models(LLMs) and proposes a novel approach that applies a method used in physical systems for detecting phase transitions to LLMs. In physical systems, phase transitions occur as control parameters, such as temperature and pressure, change. These transitions often depend only on macroscopic parameters and show universal properties. However, phase transitions in LLMs have been shown to depend on detailed ``microscopic\" inputs, such as prompts, with phase transitions occurring at different temperatures depending on the prompt. This difference suggests that a distinct mechanism may be at work in LLMs compared to physical systems, and the fact that the \"phase of behavior\" in LLMs is induced by prompts is particularly interesting. This study provides a new method for automatically detecting the phase transitions in LLMs through black-box analysis based on output distribution, yielding new insights into the learning behaviors of LLMs.", "strengths": "* This study is significant for applying a phase transition detection method from physics to the output distribution of LLMs. In physics, order parameters are often known, making the method less critical. However, in LLMs, the order parameter is entirely unknown, allowing this method to show its full advantages. \n\n* Considering the complex internal structure of LLMs, often regarded as a black box, this approach provides a potentially valuable tool for understanding LLM properties. As the authors point out, this is an advantage as it can be implemented even without detailed knowledge of an LLM's internal structure.", "weaknesses": "* Although multiple phase transitions are identified, it remains unclear how they are distinguished from crossover, as mentioned in Appendix. For example, the study suggests three phases when temperature is changed, but there is a lack of explanation of what each phase represents and its significance from a linguistic perspective.\n\n* This point slightly contradicts the above strength: what does it mean for phase transitions to depend on prompts? For example, if the temperature at which water freezes changes depending on the shape of the glass, can we truly say water undergoes a phase transition? While I understand the significance of the authors' extensive numerical experiments, it is difficult to understand their relevance to LLM research.", "questions": "*I would like clarification on how the authors calculated the heat capacity. In connection with this, it would be valuable if the authors could provide concrete evidence or specific analyses regarding the occurrence of negative heat capacity. In typical thermodynamic systems, heat capacity should be positive; however, systems with long-range interactions or non-equilibrium properties can exhibit negative values. Have the authors investigated such properties in LLMs, for instance, through analyses of long-range interactions such as the attention mechanism, or do they have evidence supporting this phenomenon in LLMs?\n* The advantage of the method used by the authors is that it can detect phase transitions from the output distributions. However, what kind of phase can the authors say it represents? Additionally, can the authors specify the type of phase transition, such as whether it is first-order or second-order?\n*Minor points.\n** It would be better to write the number \"-1\" in line 166 in math mode. \n** The caption for Fig.1 seems incorrect regarding the reference to the Appendix. \n** I initially thought Fig.4(b) was an enlarged view of Fig.4(a). While specific heat naturally connects, linear dissimilarity appears not to connect. Could the authors explain the reason?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper examines phase transition in the output distributions of large language models(LLMs) and proposes a novel approach that applies a method used in physical systems for detecting phase transitions to LLMs. In physical systems, phase transitions occur as control parameters, such as temperature and pressure, change. These transitions often depend only on macroscopic parameters and show universal properties. However, phase transitions in LLMs have been shown to depend on detailed ``microscopic\" inputs, such as prompts, with phase transitions occurring at different temperatures depending on the prompt. This difference suggests that a distinct mechanism may be at work in LLMs compared to physical systems, and the fact that the \"phase of behavior\" in LLMs is induced by prompts is particularly interesting. This study provides a new method for automatically detecting the phase transitions in LLMs through black-box analysis based on output distribution, yielding new insights into the learning behaviors of LLMs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* This study is significant for applying a phase transition detection method from physics to the output distribution of LLMs. In physics, order parameters are often known, making the method less critical. However, in LLMs, the order parameter is entirely unknown, allowing this method to show its full advantages. \n\n* Considering the complex internal structure of LLMs, often regarded as a black box, this approach provides a potentially valuable tool for understanding LLM properties. As the authors point out, this is an advantage as it can be implemented even without detailed knowledge of an LLM's internal structure.", "weaknesses": "* Although multiple phase transitions are identified, it remains unclear how they are distinguished from crossover, as mentioned in Appendix. For example, the study suggests three phases when temperature is changed, but there is a lack of explanation of what each phase represents and its significance from a linguistic perspective.\n\n* This point slightly contradicts the above strength: what does it mean for phase transitions to depend on prompts? For example, if the temperature at which water freezes changes depending on the shape of the glass, can we truly say water undergoes a phase transition? While I understand the significance of the authors' extensive numerical experiments, it is difficult to understand their relevance to LLM research.", "questions": "*I would like clarification on how the authors calculated the heat capacity. In connection with this, it would be valuable if the authors could provide concrete evidence or specific analyses regarding the occurrence of negative heat capacity. In typical thermodynamic systems, heat capacity should be positive; however, systems with long-range interactions or non-equilibrium properties can exhibit negative values. Have the authors investigated such properties in LLMs, for instance, through analyses of long-range interactions such as the attention mechanism, or do they have evidence supporting this phenomenon in LLMs?\n* The advantage of the method used by the authors is that it can detect phase transitions from the output distributions. However, what kind of phase can the authors say it represents? Additionally, can the authors specify the type of phase transition, such as whether it is first-order or second-order?\n*Minor points.\n** It would be better to write the number \"-1\" in line 166 in math mode. \n** The caption for Fig.1 seems incorrect regarding the reference to the Appendix. \n** I initially thought Fig.4(b) was an enlarged view of Fig.4(a). While specific heat naturally connects, linear dissimilarity appears not to connect. Could the authors explain the reason?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730549652643}], "openreview_url": "https://openreview.net/forum?id=dq3keisMjT", "arxiv_id": "2405.17088", "paper_pdf": "papers/dq3keisMjT.pdf", "paper_pdf_sha256": "0dbb58fa3e3333bffbd88d8ebdaa8aeabedd4e59c972d818efa00c15d9ce6948", "paper_pdf_bytes": 866082, "paper_pdf_source": "openreview", "code_url": "https://github.com/llmtransitions/llmtransitions", "code_repository": "llmtransitions/llmtransitions", "code_commit": "626845657b65205672dd7c01eb2c963b5ebb53c2", "code_archive": "repos/dq3keisMjT.zip", "code_archive_sha256": "6d0508b8e098359accc7a34d53da1b4f7df2a075a3852c961b25cf15704b2d5c", "code_archive_bytes": 8937, "code_file_count": 6, "code_extensions": {".py": 3, ".ipynb": 3}, "github_disk_usage_kb": 12, "github_languages": {"Jupyter Notebook": 14391, "Python": 8484}, "github_archived": false, "github_pushed_at": "2025-08-15T13:57:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/phase-transitions-in-the-output-distribution"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "k3VANp85b4S", "year": 2023, "status": "rejected", "title": "On the Robustness of Randomized Ensembles to Adversarial Perturbations", "authors": ["Hassan Dbouk", "Naresh Shanbhag"], "authorids": ["~Hassan_Dbouk1", "~Naresh_Shanbhag1"], "authors_source": "OpenReview API", "abstract": "Randomized ensemble classifiers (RECs), where one classifier is randomly selected during inference, have emerged as an attractive alternative to traditional ensembling methods for realizing adversarially robust classifiers with limited compute requirements. However, recent works have shown that existing methods for constructing RECs are more vulnerable than initially claimed, casting major doubts on their efficacy and prompting fundamental questions such as: \"When are RECs useful?\", \"What are their limits?\", and \"How do we train them?\". In this work, we first demystify RECs as we derive fundamental results regarding their theoretical limits, necessary and sufficient conditions for them to be useful, and more. Leveraging this new understanding, we propose a new boosting\nalgorithm (BARRE) for training robust RECs, and empirically demonstrate its effectiveness at defending against strong $\\ell_\\infty$ norm-bounded adversaries across various network architectures and datasets. Our code is submitted as part of the supplementary material, and will be publicly released on GitHub", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "bM79tDbcPZ7", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3847/Reviewer_BAMT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers the problem of adversarial robustness. For this problem the paper investigates the usefulness of randomized ensemble classifiers (REC) where one classifier is randomly selected from the ensemble during inference.  The main motivation behind considering RECs over deterministic ensembles is that the former has much smaller inference time than the latter. \n\nThe paper makes two main contributions. The first contribution, which is on the theory front, involves careful characterization of the adversarial risk of RECs. In particular, the authors obtain reasonably tight upper and lower bounds for the adversarial risk of RECs that depend on the adversarial risks of the component classifiers in the ensemble. Based on these bounds, the authors provide some useful insights on how to design RECs with good adversarial risk guarantees. The second contribution is to provide a boosting style algorithm (BARRE) for constructing robust RECs that can tolerate adversarial attacks. The algorithm is mostly inspired by a recently proposed robust boosting technique called MRBoost.", "review_text": "The clarity and presentation in the paper are good. The theoretical results are interesting and novel. But their usefulness is a little bit unclear. The empirical results look weak. Moreover, the proposed algorithm looks identical to MRBoost, except for a minor step which involves setting the weights of the component classifiers in the ensemble. Given this, I'm a little bit inclined towards rejecting the paper. But I'm happy to upgrade my score if the authors address my concerns.", "strengths": "Strengths:\n  *  Randomized ensembling is an interesting defense strategy that is under-explored in the literature. While there are some works that provide algorithms for building robust RECs, they are easily broken by adaptive defenses. So there is a need for better randomized ensembling strategies. The paper takes a step towards solving this problem. The theoretical results provided in the paper on bounding the adversarial risk of RECs, are novel and interesting. However, it is not immediately clear how the insights gained from these theoretical results can be used to develop a good algorithm (more on this below).\n  *  The paper is easy to read. The clarity and presentation is good. The proofs are elementary and easy to follow.\n\nWeaknesses:\n  *  Motivation: \n      - The authors motivate the paper by saying that randomized ensembles can be computationally more efficient than deterministic ensembles. But there is a drawback of randomized ensembles that hasn't been brought up in the paper. In many application domains (e.g., healthcare), it is important to have deterministic predictions (otherwise it is hard to trust and understand these complex models). By moving to randomized ensembles, we lose this property. Can the authors provide some concrete use cases for randomized ensembles?\n      -  Low compute resource devices also tend to have low memory. But RECs have high memory requirement ( same as deterministic ensembles). Given this, and the fact that deterministic ensembling techniques (like MRBoost) have better performance than BARRE, it is not entirely clear what the practical applications of RECs could be.  \n\n\n  * Empirical results:  \n     - the empirical results look weak. For example, in table 1, MRBoost has better performance than BARRE both in terms of robust and standard accuracy. This seems to be in contrast with the message one gets by reading the theory section in the paper. For example in page 6 (in paragraph titled \"implications of upper bound\"), it is mentioned that deterministic ensembles have much worse performance in the worst-case. Why is there this mismatch between theory and experiments? \n         - under what circumstances are deterministic ensembles better than RECs and vice-versa? \n\n * BARRE: \n    - the algorithm looks almost identical to MRBoost, except for line 12 in Algorithm 2. In BARRE, the weights for each component classifier in the ensemble are recomputed after every boosting iteration. Whereas in MRBoost, all the component classifiers are given equal weights. Given this, why can't we simply add step 12 to MRBoost and get a randomized classifier out of it? How would the resulting algorithm compare with BARRE?\n    - What is the computational overhead of BARRE over MRBoost?  It looks like computing weights (line 12) can be expensive, especially for large M.\n    - In the introduction the authors claim that BARRE is based on the theoretical results in the paper (page 2). But I don't see any connection between the two. In particular, I don't see how the theoretical results in sections 3.2, 3.3 are used to derive this algorithm.  \n\n *  Theoretical Results: \n    -  A number of insights on how to obtain robust RECs have been provided in the discussion after theoretical results. Can these insights be used to derive a better algorithm than BARRE?\n\n * Minor comments:\n    -  a more detailed explanation on why K=5 in section 3.2 would be helpful to the readers\n    - ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers the problem of adversarial robustness. For this problem the paper investigates the usefulness of randomized ensemble classifiers (REC) where one classifier is randomly selected from the ensemble during inference.  The main motivation behind considering RECs over deterministic ensembles is that the former has much smaller inference time than the latter. \n\nThe paper makes two main contributions. The first contribution, which is on the theory front, involves careful characterization of the adversarial risk of RECs. In particular, the authors obtain reasonably tight upper and lower bounds for the adversarial risk of RECs that depend on the adversarial risks of the component classifiers in the ensemble. Based on these bounds, the authors provide some useful insights on how to design RECs with good adversarial risk guarantees. The second contribution is to provide a boosting style algorithm (BARRE) for constructing robust RECs that can tolerate adversarial attacks. The algorithm is mostly inspired by a recently proposed robust boosting technique called MRBoost.", "strength_and_weaknesses": "Strengths:\n  *  Randomized ensembling is an interesting defense strategy that is under-explored in the literature. While there are some works that provide algorithms for building robust RECs, they are easily broken by adaptive defenses. So there is a need for better randomized ensembling strategies. The paper takes a step towards solving this problem. The theoretical results provided in the paper on bounding the adversarial risk of RECs, are novel and interesting. However, it is not immediately clear how the insights gained from these theoretical results can be used to develop a good algorithm (more on this below).\n  *  The paper is easy to read. The clarity and presentation is good. The proofs are elementary and easy to follow.\n\nWeaknesses:\n  *  Motivation: \n      - The authors motivate the paper by saying that randomized ensembles can be computationally more efficient than deterministic ensembles. But there is a drawback of randomized ensembles that hasn't been brought up in the paper. In many application domains (e.g., healthcare), it is important to have deterministic predictions (otherwise it is hard to trust and understand these complex models). By moving to randomized ensembles, we lose this property. Can the authors provide some concrete use cases for randomized ensembles?\n      -  Low compute resource devices also tend to have low memory. But RECs have high memory requirement ( same as deterministic ensembles). Given this, and the fact that deterministic ensembling techniques (like MRBoost) have better performance than BARRE, it is not entirely clear what the practical applications of RECs could be.  \n\n\n  * Empirical results:  \n     - the empirical results look weak. For example, in table 1, MRBoost has better performance than BARRE both in terms of robust and standard accuracy. This seems to be in contrast with the message one gets by reading the theory section in the paper. For example in page 6 (in paragraph titled \"implications of upper bound\"), it is mentioned that deterministic ensembles have much worse performance in the worst-case. Why is there this mismatch between theory and experiments? \n         - under what circumstances are deterministic ensembles better than RECs and vice-versa? \n\n * BARRE: \n    - the algorithm looks almost identical to MRBoost, except for line 12 in Algorithm 2. In BARRE, the weights for each component classifier in the ensemble are recomputed after every boosting iteration. Whereas in MRBoost, all the component classifiers are given equal weights. Given this, why can't we simply add step 12 to MRBoost and get a randomized classifier out of it? How would the resulting algorithm compare with BARRE?\n    - What is the computational overhead of BARRE over MRBoost?  It looks like computing weights (line 12) can be expensive, especially for large M.\n    - In the introduction the authors claim that BARRE is based on the theoretical results in the paper (page 2). But I don't see any connection between the two. In particular, I don't see how the theoretical results in sections 3.2, 3.3 are used to derive this algorithm.  \n\n *  Theoretical Results: \n    -  A number of insights on how to obtain robust RECs have been provided in the discussion after theoretical results. Can these insights be used to derive a better algorithm than BARRE?\n\n * Minor comments:\n    -  a more detailed explanation on why K=5 in section 3.2 would be helpful to the readers\n    - ", "clarity,_quality,_novelty_and_reproducibility": "See comments above", "summary_of_the_review": "The clarity and presentation in the paper are good. The theoretical results are interesting and novel. But their usefulness is a little bit unclear. The empirical results look weak. Moreover, the proposed algorithm looks identical to MRBoost, except for a minor step which involves setting the weights of the component classifiers in the ensemble. Given this, I'm a little bit inclined towards rejecting the paper. But I'm happy to upgrade my score if the authors address my concerns.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667218649403}, {"id": "yZ95dDWLIm", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3847/Reviewer_5kPi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this work, the authors analyze randomized ensembles through a theoretical perspective. A randomized ensemble classifier (REC) is an ensemble of classifiers where predictions are made using only one of the constituent models selected at random. The authors analyze this set up in an adversarial scenario and show a series of elementary facts about them including bounds on the worst-case risk of the REC and from these facts show how to set the probability that any one classifier is taken from the ensemble as the model from which a prediction is made. The primary motivations for this set up are claimed to be (1) greater computational complexity compared to ensembles (2) greater adversarial robustness compared to both ensembles and single models. ", "review_text": "Overall, I found the theory presented to be simplistic bordering on trivial. The claims are also incorrect in places (see the example in weaknesses).  The experimental analysis does not really report error bars which is also a red flag. Overall I think the work needs some further developments to be impactful.", "strengths": "Strengths:\n\nThe authors provide promising empirical evidence that an REC can be made from an ensemble of classifiers such that one gets performance gains over the naive ensembling approach that would be typically taken.\n\nWeaknesses: \n\nThere are a couple of non-trivial drawbacks to this method that are not discussed. For the first contribution (having better computational complexity) the authors leave undiscussed the cost of this computational complexity. In particular, when using an REC one forfeits the uncertainty properties of an ensemble classifier (or Bayesian NN) which have been shown in many instances to help adversarial robustness and to aid in safety analysis. It feels like this is an important drawback: the computational gain here is not for free, you are sacrificing uncertainty. And this needs to be at the very least discussed if not empirically analyzed. \n\nThe theoretical results themselves are also simplistic, bordering on trivial. I do note here that I am not an expert on RECs, so perhaps reviewers with more experience in these models will find these results more interesting than I. In addition to being very simplistic the claims about some of the theorems are not correct. Most glaring to me is the statement that \"there are no worst-case performance guarantees with deterministic ensembling, even if all the classifiers are robust.\" This statement is either unqualified and therefore misleading or simply  incorrect. A deterministic ensemble (or BNN) is a single classifier: a model average. Therefore, they _do_ fit into Theorem 2 as there is only one model under consideration in these cases and therefore only one $\\eta$ (risk value) and therefore they do have the same guarantees. Thus the statement regarding Theorem 2 is incorrect. This also points to the fact that Theorem 2 is bordering on trivial because it can be said of any classifier and is not unique to RECs. \n\nFinally, the results, while they do show a slight increase in performance, do not include any error bars. This seems a large omission given that the performance gains are ~1% in many cases and this is a randomized method so not reporting the variance is a red flag. \n\nMinor note:  I find the placement of the related works to be a bit jarring. If it could be placed in before the problem statement that might help the flow of the paper. But this is stylistic.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this work, the authors analyze randomized ensembles through a theoretical perspective. A randomized ensemble classifier (REC) is an ensemble of classifiers where predictions are made using only one of the constituent models selected at random. The authors analyze this set up in an adversarial scenario and show a series of elementary facts about them including bounds on the worst-case risk of the REC and from these facts show how to set the probability that any one classifier is taken from the ensemble as the model from which a prediction is made. The primary motivations for this set up are claimed to be (1) greater computational complexity compared to ensembles (2) greater adversarial robustness compared to both ensembles and single models. ", "strength_and_weaknesses": "Strengths:\n\nThe authors provide promising empirical evidence that an REC can be made from an ensemble of classifiers such that one gets performance gains over the naive ensembling approach that would be typically taken.\n\nWeaknesses: \n\nThere are a couple of non-trivial drawbacks to this method that are not discussed. For the first contribution (having better computational complexity) the authors leave undiscussed the cost of this computational complexity. In particular, when using an REC one forfeits the uncertainty properties of an ensemble classifier (or Bayesian NN) which have been shown in many instances to help adversarial robustness and to aid in safety analysis. It feels like this is an important drawback: the computational gain here is not for free, you are sacrificing uncertainty. And this needs to be at the very least discussed if not empirically analyzed. \n\nThe theoretical results themselves are also simplistic, bordering on trivial. I do note here that I am not an expert on RECs, so perhaps reviewers with more experience in these models will find these results more interesting than I. In addition to being very simplistic the claims about some of the theorems are not correct. Most glaring to me is the statement that \"there are no worst-case performance guarantees with deterministic ensembling, even if all the classifiers are robust.\" This statement is either unqualified and therefore misleading or simply  incorrect. A deterministic ensemble (or BNN) is a single classifier: a model average. Therefore, they _do_ fit into Theorem 2 as there is only one model under consideration in these cases and therefore only one $\\eta$ (risk value) and therefore they do have the same guarantees. Thus the statement regarding Theorem 2 is incorrect. This also points to the fact that Theorem 2 is bordering on trivial because it can be said of any classifier and is not unique to RECs. \n\nFinally, the results, while they do show a slight increase in performance, do not include any error bars. This seems a large omission given that the performance gains are ~1% in many cases and this is a randomized method so not reporting the variance is a red flag. \n\nMinor note:  I find the placement of the related works to be a bit jarring. If it could be placed in before the problem statement that might help the flow of the paper. But this is stylistic.", "clarity,_quality,_novelty_and_reproducibility": "The paper is relatively clear, but could be more notationally clear in a few sections. The provided intuitions are nice and help tho clearly understand what the authors mean. From the details in the paper the method would be reproducible, but one would need access to the code base to reproduce the exact numbers in this paper as there are inevitably hyper-parameters that are not reported here.", "summary_of_the_review": "Overall, I found the theory presented to be simplistic bordering on trivial. The claims are also incorrect in places (see the example in weaknesses).  The experimental analysis does not really report error bars which is also a red flag. Overall I think the work needs some further developments to be impactful.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None found.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666705276795}, {"id": "zykqbBll3c", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3847/Reviewer_3zJ9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper studies the limits of the adversarial robustness that can be obtained by combining $M$ different classifiers into one randomized classifier that at inference time, chooses one of the elements with certain probability. Some intermediate lower and upper bounds are presented that illustrate the main intuition behind the results. The main theorem is Theorem 2 which basically states that the randomized classifier in the worst case is as accurate as its worst component (trivial) and in the best case it can improve the robustness by a $1/M$ factor in the case where all classifiers have the same robustness (this is the interesting bound, and the result is stated in the more general case where each classifier might have a different robustness). Two algorithms are provided: the first obtains the best sampling probaility vector (assuming an oracle that obtains optimal adversarial perturbations) and the second is a boosting algorithm to obtain good classifiers to create the ensemble. In practical experiments it is shown that the second algorithm can lead to increased robustness against the state-of-the art ARC attack which was shown to break previous randomized ensemble defenses.\n\n**After Rebuttal**\nThe authors have addressed some of my concerns. Despite not addressing the missing confidence intervals in the experiments, I am inclined to increase my score because of the following:\n\n1. The method studied is interesting and significant, as there hasn't been an improvement in robust accuracy metrics from the single model paradigm in the last years. I believe most improvements have been on the training speed department rather than the robust accuracy metric. As such, exploring alternative paradigms like the randomized ensemble classifier studied here is one promising way forward.\n2. A good amount of contributions are theoretical, and they provide guidance on how to develop better algorithms for building RECs. As such, I think it is possible to give the authors a pass regarding the fact that the improvements over the baseline are small and lack the confidence intervals. Focusing only on the experimental part would dismiss the interesting theoretical results which could be used or improved upon by others. Of course, after reflecting on the results presented, they appear quite intuitive, however, i think it is a stretch to call them 'trivial' as other reviewers have dismissed. To support such a statement they should have provided a simplified proof or argument, which I believe is not the case.\n3. I disagree that not enjoying certain properties of BNNs would be a reason to dismiss researching RECs. Both are methods with pros and cons, and the goal of this paper as I understood was to advance the algorithmic and theoretical framework of RECs, rather than claim they are superior to BNNs in all aspects.\n4. There are some concerns that this method is not suitable for some applications (like certain medical tasks). However, applications of ML have vastly different specifications and its hard to develop a one-size-fits-all method.\n\nIn the end, raising my score will not help much as there is still a big disagreement with two other reviewers. It would be most helpful if they re-evaluate their score after the author's rebuttal, and express if some of their concerns have been addressed.", "review_text": "Upper and lower bounds for the robustness of ensemble classifiers are presented. The results are quite interesting and their presentation is clear despite being highly technical. The theory leads to the design of principled algorithms that are evaluated against strong baselines, showing promise for improvement. Overall the paper seems to make significant contributions. However two big issues remain: (1) lack of discussion regarding prior results in the context of (non-adversarial) accuracy of ensembles and (2) computation of confidence intervals for the numbers presented in the tables. These two last points prevent me from confidently recommending acceptance but I would increase my score if they are addressed.", "strengths": "The main strength of this paper are the interesting theoretical bounds on the performance of randomized classifiers. This is important as the paradigm of training a single robust classifier appears to have hit a limit in terms of robustness, and the obtained lower bounds from Theorem 2 indicate that randomizing the classifier could lead to potentially big improvements. Of course, the lower bounds are only achieved when the classifiers at hand have the nice \"incoherence\" property where they don't share \"vulnerable regions\" where all are simultaneously mistaken by an adversary. The paper does a good job of presenting the case of two classifiers where the results are pretty straightforward and easy to understand, before presenting Theorem 2, which is the statement in full generality.\n\nThe math appears to be correct (I have checked some of the results in detail) and all the statements are clear, as well as the flow of the proofs is well organised. This is in contrast with the awful quality of the maths/proofs from average ICML/NeurIPS/ICLR submission, so this deserves a special mention. I would say however that some proofs are a bit more difficult than needed, for example in the proof of Lemma 5 in page 18, the items (1) and (2) are trivial. (1) simply follows from the fact that the sets are intersection of half-spaces and (2) follows from the well known equivalence between Vertex and Hyperplane descriptions of convex polytopes, see for example https://link.springer.com/content/pdf/bbm:978-0-387-46112-0/1.pdf\n\nThe theoretical results are not left without application, as they are used to define the algorithms 1 and 2, which builds a randomized classifier in stages, by adding robust classifiers while simultaneouly optimizing the sampling probabilities of the ensemble. Overall the theory justifies the design of the algorithm which is ideal.\n\nThe main weakness I see (unfortunately) is the empirical evaluation. Now, I think the settings studied are extensive, but the fact that there is randomness involved in the algorithm (through the use of SGD) means that one should provide some assessment of the confidence about the numbers stated in Table 1 and 2. The authors only provide a single number and it might be the case that the improvements are due to randomness. Hopefully this can be resolved simply by running multiple seeds and computing some confidence intervals. That would greatly improve the paper.\n\nFinally another weakness is that the authors do not make an effort of presenting related results. I believe this topic should have been studied before at least in the context of clean accuracy (no adversary). The authors should comment on the relation to the papers:\n1. https://aamas.csc.liv.ac.uk/Proceedings/aamas2014/aamas/p485.pdf\n2. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3156487/\n3. https://www.cs.waikato.ac.nz/~ml/publications/2002/bounds.pdf\n4. https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7044723\n\nand cite them, as it seems they study a similar question (possibly in the non-adversarial setting). ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper studies the limits of the adversarial robustness that can be obtained by combining $M$ different classifiers into one randomized classifier that at inference time, chooses one of the elements with certain probability. Some intermediate lower and upper bounds are presented that illustrate the main intuition behind the results. The main theorem is Theorem 2 which basically states that the randomized classifier in the worst case is as accurate as its worst component (trivial) and in the best case it can improve the robustness by a $1/M$ factor in the case where all classifiers have the same robustness (this is the interesting bound, and the result is stated in the more general case where each classifier might have a different robustness). Two algorithms are provided: the first obtains the best sampling probaility vector (assuming an oracle that obtains optimal adversarial perturbations) and the second is a boosting algorithm to obtain good classifiers to create the ensemble. In practical experiments it is shown that the second algorithm can lead to increased robustness against the state-of-the art ARC attack which was shown to break previous randomized ensemble defenses.\n\n**After Rebuttal**\nThe authors have addressed some of my concerns. Despite not addressing the missing confidence intervals in the experiments, I am inclined to increase my score because of the following:\n\n1. The method studied is interesting and significant, as there hasn't been an improvement in robust accuracy metrics from the single model paradigm in the last years. I believe most improvements have been on the training speed department rather than the robust accuracy metric. As such, exploring alternative paradigms like the randomized ensemble classifier studied here is one promising way forward.\n2. A good amount of contributions are theoretical, and they provide guidance on how to develop better algorithms for building RECs. As such, I think it is possible to give the authors a pass regarding the fact that the improvements over the baseline are small and lack the confidence intervals. Focusing only on the experimental part would dismiss the interesting theoretical results which could be used or improved upon by others. Of course, after reflecting on the results presented, they appear quite intuitive, however, i think it is a stretch to call them 'trivial' as other reviewers have dismissed. To support such a statement they should have provided a simplified proof or argument, which I believe is not the case.\n3. I disagree that not enjoying certain properties of BNNs would be a reason to dismiss researching RECs. Both are methods with pros and cons, and the goal of this paper as I understood was to advance the algorithmic and theoretical framework of RECs, rather than claim they are superior to BNNs in all aspects.\n4. There are some concerns that this method is not suitable for some applications (like certain medical tasks). However, applications of ML have vastly different specifications and its hard to develop a one-size-fits-all method.\n\nIn the end, raising my score will not help much as there is still a big disagreement with two other reviewers. It would be most helpful if they re-evaluate their score after the author's rebuttal, and express if some of their concerns have been addressed.", "strength_and_weaknesses": "The main strength of this paper are the interesting theoretical bounds on the performance of randomized classifiers. This is important as the paradigm of training a single robust classifier appears to have hit a limit in terms of robustness, and the obtained lower bounds from Theorem 2 indicate that randomizing the classifier could lead to potentially big improvements. Of course, the lower bounds are only achieved when the classifiers at hand have the nice \"incoherence\" property where they don't share \"vulnerable regions\" where all are simultaneously mistaken by an adversary. The paper does a good job of presenting the case of two classifiers where the results are pretty straightforward and easy to understand, before presenting Theorem 2, which is the statement in full generality.\n\nThe math appears to be correct (I have checked some of the results in detail) and all the statements are clear, as well as the flow of the proofs is well organised. This is in contrast with the awful quality of the maths/proofs from average ICML/NeurIPS/ICLR submission, so this deserves a special mention. I would say however that some proofs are a bit more difficult than needed, for example in the proof of Lemma 5 in page 18, the items (1) and (2) are trivial. (1) simply follows from the fact that the sets are intersection of half-spaces and (2) follows from the well known equivalence between Vertex and Hyperplane descriptions of convex polytopes, see for example https://link.springer.com/content/pdf/bbm:978-0-387-46112-0/1.pdf\n\nThe theoretical results are not left without application, as they are used to define the algorithms 1 and 2, which builds a randomized classifier in stages, by adding robust classifiers while simultaneouly optimizing the sampling probabilities of the ensemble. Overall the theory justifies the design of the algorithm which is ideal.\n\nThe main weakness I see (unfortunately) is the empirical evaluation. Now, I think the settings studied are extensive, but the fact that there is randomness involved in the algorithm (through the use of SGD) means that one should provide some assessment of the confidence about the numbers stated in Table 1 and 2. The authors only provide a single number and it might be the case that the improvements are due to randomness. Hopefully this can be resolved simply by running multiple seeds and computing some confidence intervals. That would greatly improve the paper.\n\nFinally another weakness is that the authors do not make an effort of presenting related results. I believe this topic should have been studied before at least in the context of clean accuracy (no adversary). The authors should comment on the relation to the papers:\n1. https://aamas.csc.liv.ac.uk/Proceedings/aamas2014/aamas/p485.pdf\n2. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3156487/\n3. https://www.cs.waikato.ac.nz/~ml/publications/2002/bounds.pdf\n4. https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7044723\n\nand cite them, as it seems they study a similar question (possibly in the non-adversarial setting). ", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the paper is highly technical but it can be understood as the definitions/results are presented with care, and some particular cases are studied in detail first before presenting the more general results. Overall the paper makes pretty clear statements.\n\nQuality: the technical quality is high. Some lemmas/theorems have easier proofs than others but overall I think the results are interesting and non-trivial. Some proofs could be simplified as I mentioned. However, the quality suffers in the experimental section due to the absence of confidence intervals. When there is some randomness in the algorithms this is a must, in order to be able to *bold* the results and claim that there is improvement over the baseline.\n\nNovelty: this results seem natural in the context of clean accuracy, and I would say it is possible that similar questions have been studied before. Even though I cannot say the results are not novel, as I have stated before the authors should do a better job of finding and discussing some relevant results.\n\nReproducibility: the authors have provided the code for the experiments", "summary_of_the_review": "Upper and lower bounds for the robustness of ensemble classifiers are presented. The results are quite interesting and their presentation is clear despite being highly technical. The theory leads to the design of principled algorithms that are evaluated against strong baselines, showing promise for improvement. Overall the paper seems to make significant contributions. However two big issues remain: (1) lack of discussion regarding prior results in the context of (non-adversarial) accuracy of ensembles and (2) computation of confidence intervals for the numbers presented in the tables. These two last points prevent me from confidently recommending acceptance but I would increase my score if they are addressed.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666550446873}, {"id": "VALBXqwmZyP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3847/Reviewer_tU9P"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers randomised ensemble classifiers (RECs) and study the problem of how to assign the probabilities to select each classifier to improve the adversarial robustness of the REC. The authors first show that minimising the adversarial risk is equivalent to minimising a piecewise linear function with linear constraints. Then, by relying on the convexity of the problem, they develop algorithms to solve the resulting optimisation problem first in the case of 2 classifiers and then they consider the general case.  Experimental analysis on CIFAR-10 and CIFAR-100 compare the proposed method with state-of-the-art approaches.", "review_text": "The paper studies an important problem and presents an interesting new perspective that makes possible to use of tools from convex optimisation. However, experimental results do not seem to present a substantial improvement compared to state-of-the-art. ", "strengths": "Strength:\n\n-Paper is well written\n\n-The authors formally show that picking the optimal sampling probabilities for different classifiers is equivalent to optimising a piece-wise linear convex function over a convex set. This enables them to use tools from convex optimisation to solve this problem. I find this observation and the resulting algorithms interesting.\n\n\nWeaknesses:\n\n-Experimental results should be improved/clarified: \n\na) Table 1 and 2 report different trends in the comparison with MRBoost. In fact, while in table 1 MRBoost almost always obtains better accuracy and robustness compared to BARRE (the method presented by the authors), the trend is reversed in Table 2. This is particularly confusing for the ResNEt-18 M=4 example, where while BARRE has same values in both Table 1 and 2, for MRBoost these are different.\n\nb) Please, report computational times. Also, it would be interesting to also consider larger values of M.\n\nc) FLOPs in Table 1 is not defined\n \n\n-Theorem 1 guarantees that to improve the empirical risk of the ensemble compared to that of the individual classifiers, region $\\mathcal{R}_1$ should contain large probability mass wrt the input distribution. However, a discussion on how often this condition is verified in practice is missing. Intuitively, I believe that, due to the transferability of adversarial examples, such region will not be large in general. Nevertheless, a possibly interesting direction could be to train the individual classifiers in the ensemble with different adversarial training techniques in order to enforce that $\\mathcal{R}_1$ is large, while keeping each classifier accurate and robust to a particular class of attacks.\n\n-The fact that randomized classifiers can improve the adversarial robustness wrt the individual classifiers is well known. Therefore, also considering the above comments on the experiments, the results of the paper feel a bit incremental.\n\n-Finally, the comparison with the literature completely misses Bayesian ensembles, which have been already shown to be robust to adversarial examples, see e.g. [1]. \n\n\n[1]: Carbone, Ginevra, et al. \"Robustness of bayesian neural networks to gradient-based attacks.\" Advances in Neural Information Processing Systems 33 (2020): 15602-15613.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers randomised ensemble classifiers (RECs) and study the problem of how to assign the probabilities to select each classifier to improve the adversarial robustness of the REC. The authors first show that minimising the adversarial risk is equivalent to minimising a piecewise linear function with linear constraints. Then, by relying on the convexity of the problem, they develop algorithms to solve the resulting optimisation problem first in the case of 2 classifiers and then they consider the general case.  Experimental analysis on CIFAR-10 and CIFAR-100 compare the proposed method with state-of-the-art approaches.", "strength_and_weaknesses": "Strength:\n\n-Paper is well written\n\n-The authors formally show that picking the optimal sampling probabilities for different classifiers is equivalent to optimising a piece-wise linear convex function over a convex set. This enables them to use tools from convex optimisation to solve this problem. I find this observation and the resulting algorithms interesting.\n\n\nWeaknesses:\n\n-Experimental results should be improved/clarified: \n\na) Table 1 and 2 report different trends in the comparison with MRBoost. In fact, while in table 1 MRBoost almost always obtains better accuracy and robustness compared to BARRE (the method presented by the authors), the trend is reversed in Table 2. This is particularly confusing for the ResNEt-18 M=4 example, where while BARRE has same values in both Table 1 and 2, for MRBoost these are different.\n\nb) Please, report computational times. Also, it would be interesting to also consider larger values of M.\n\nc) FLOPs in Table 1 is not defined\n \n\n-Theorem 1 guarantees that to improve the empirical risk of the ensemble compared to that of the individual classifiers, region $\\mathcal{R}_1$ should contain large probability mass wrt the input distribution. However, a discussion on how often this condition is verified in practice is missing. Intuitively, I believe that, due to the transferability of adversarial examples, such region will not be large in general. Nevertheless, a possibly interesting direction could be to train the individual classifiers in the ensemble with different adversarial training techniques in order to enforce that $\\mathcal{R}_1$ is large, while keeping each classifier accurate and robust to a particular class of attacks.\n\n-The fact that randomized classifiers can improve the adversarial robustness wrt the individual classifiers is well known. Therefore, also considering the above comments on the experiments, the results of the paper feel a bit incremental.\n\n-Finally, the comparison with the literature completely misses Bayesian ensembles, which have been already shown to be robust to adversarial examples, see e.g. [1]. \n\n\n[1]: Carbone, Ginevra, et al. \"Robustness of bayesian neural networks to gradient-based attacks.\" Advances in Neural Information Processing Systems 33 (2020): 15602-15613.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written. However, I still have some doubts on the reproducibility of experiments and on the novelty (see the Strength and Weaknesses Section for detailed comments)", "summary_of_the_review": "The paper studies an important problem and presents an interesting new perspective that makes possible to use of tools from convex optimisation. However, experimental results do not seem to present a substantial improvement compared to state-of-the-art. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1665956594118}], "openreview_url": "https://openreview.net/forum?id=k3VANp85b4S", "arxiv_id": "2302.01375", "paper_pdf": "papers/k3VANp85b4S.pdf", "paper_pdf_sha256": "fc476d61e8af77d8b69aaa7bcdbf266b7ed3895dcff10aedaa692da3fa5515b2", "paper_pdf_bytes": 761683, "paper_pdf_source": "openreview", "code_url": "https://github.com/hsndbk4/BARRE", "code_repository": "hsndbk4/BARRE", "code_commit": "1a42d89214db58200abd73f29ae5f8ee1e3bebe6", "code_archive": "repos/k3VANp85b4S.zip", "code_archive_sha256": "97bb1b39d3c562b65753663e05ab1352faec8e3cff5ec83554c924d0abaa0f6c", "code_archive_bytes": 16547, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 15, "github_languages": {"Python": 46743}, "github_archived": false, "github_pushed_at": "2023-09-05T02:14:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-robustness-of-randomized-ensembles-to"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GIEPR9OomyX", "year": 2022, "status": "rejected", "title": "Langevin Autoencoders for Learning Deep Latent Variable Models", "authors": ["Shohei Taniguchi", "Yusuke Iwasawa", "Wataru Kumagai", "Yutaka Matsuo"], "authorids": ["~Shohei_Taniguchi1", "~Yusuke_Iwasawa1", "~Wataru_Kumagai2", "~Yutaka_Matsuo1"], "authors_source": "OpenReview API", "abstract": "Markov chain Monte Carlo (MCMC), such as Langevin dynamics, is valid for approximating intractable distributions. However, its usage is limited in the context of deep latent variable models since it is not scalable to data size owing to its datapoint-wise iterations and slow convergence. This paper proposes the amortized Langevin dynamics (ALD), wherein datapoint-wise MCMC iterations are entirely replaced with updates of an inference model that maps observations into latent variables. Since it no longer depends on datapoint-wise iterations, ALD enables scalable inference from large-scale datasets. Despite its efficiency, it retains the excellent property of MCMC; we prove that ALD has the target posterior as a stationary distribution with a mild assumption. Furthermore, ALD can be extended to sampling from an unconditional distribution such as an energy-based model, enabling more flexible generative modeling by applying it to the prior distribution of the latent variable. Based on ALD, we construct a new deep latent variable model named the Langevin autoencoder (LAE). LAE uses ALD for autoencoder-like posterior inference and sampling from the latent space EBM. Using toy datasets, we empirically validate that ALD can properly obtain samples from target distributions in both conditional and unconditional cases, and ALD converges significantly faster than traditional LD. We also evaluate LAE on the image generation task using three datasets (SVHN, CIFAR-10, and CelebA-HQ). Not only can LAE be trained faster than non-amortized MCMC methods, but LAE can also generate better samples in terms of the Fréchet Inception Distance (FID) compared to AVI-based methods, such as the variational autoencoder.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "CKfl_fACzft", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2820/Reviewer_ZHba"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces amortized Langevin dynamics for latent variables and extends it to unconditional distributions. A Langevin autoencoder is further developed by using amortized Langevin dynamics for prior and posterior sampling. The proposed method has been tested on synthetic distributions and image generation.", "review_text": "It is not clear to me how to get the function g in the inference network. Is the function g obtained through some preprocessing? How does the quality of it affect the posterior approximation result? How do the authors get it in the experiment section? It will be better to provide more explanation and details about it, since it seems to play an important role in approximating the posterior as shown in Theorem 1.\n\nThe second condition for ALD to converge to the true distribution does not seem mild. Since the dataset size is typically very large, it is impractical to have such high dimension for the linear layer. Therefor the claim seems a bit misleading that “ALD has the target posterior as a stationary distribution with a mild assumption”. \n\nHow does the storage cost of the proposed method compared to LD without amortization and AVI? By looking at Algorithm 1, the storage cost of the proposed method seems much higher, since it has to save Z^1 to Z^n and each Z^i contains several samples.\n\nFor the experiments, it will be better to add the results of SGALD in Figures 1 and 4 to check the posterior estimation when using minibatch of data. Since SGALD is the one being used in LAE in practice. \n\nOn the synthetic distributions, the dimension of the linear layer is much larger than the dataset size (128 vs 3). However in practice the dataset size will be much larger than the dimension of the linear layer (as what the authors did for the experiment on image generation). It will be much convincing to simulate this scenario on synthetic data and verify that with small dimension ALD/SGALD can still approximate posterior well. It will be very helpful to further show how the posterior estimation changes with respect to the dimension of the linear layer.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces amortized Langevin dynamics for latent variables and extends it to unconditional distributions. A Langevin autoencoder is further developed by using amortized Langevin dynamics for prior and posterior sampling. The proposed method has been tested on synthetic distributions and image generation.", "main_review": "It is not clear to me how to get the function g in the inference network. Is the function g obtained through some preprocessing? How does the quality of it affect the posterior approximation result? How do the authors get it in the experiment section? It will be better to provide more explanation and details about it, since it seems to play an important role in approximating the posterior as shown in Theorem 1.\n\nThe second condition for ALD to converge to the true distribution does not seem mild. Since the dataset size is typically very large, it is impractical to have such high dimension for the linear layer. Therefor the claim seems a bit misleading that “ALD has the target posterior as a stationary distribution with a mild assumption”. \n\nHow does the storage cost of the proposed method compared to LD without amortization and AVI? By looking at Algorithm 1, the storage cost of the proposed method seems much higher, since it has to save Z^1 to Z^n and each Z^i contains several samples.\n\nFor the experiments, it will be better to add the results of SGALD in Figures 1 and 4 to check the posterior estimation when using minibatch of data. Since SGALD is the one being used in LAE in practice. \n\nOn the synthetic distributions, the dimension of the linear layer is much larger than the dataset size (128 vs 3). However in practice the dataset size will be much larger than the dimension of the linear layer (as what the authors did for the experiment on image generation). It will be much convincing to simulate this scenario on synthetic data and verify that with small dimension ALD/SGALD can still approximate posterior well. It will be very helpful to further show how the posterior estimation changes with respect to the dimension of the linear layer.\n\n", "summary_of_the_review": "In summary, I think the idea is reasonable and developing a cheap MCMC-based autoencoder will be of interest to the community. However, I have some concerns with respect to the methodology and the experiments mentioned above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636224765343}, {"id": "sLE1tvD9dc6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2820/Reviewer_siJZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an amortized Langevin dynamics for sampling from a posterior distribution of a top-down generative model or from an energy-based model. The key is to recruit a sampler function that generates many samples directly in parallel, and run a single Langevin dynamics on the parameters of this sampler function. The method is illustrated by image generation. ", "review_text": "Strengths: \n\n(1) The idea of the paper is novel and interesting. Theoretical analysis is provided. \n\n(2) The proposed method can be applied to sampling both the posterior distribution and the energy-based model. \n\n(3) The proposed method is different from variational approximation. \n\nWeaknesses: \n\n(1) One thing I am concerned with is, if all the randomness is accounted for by the Langevin sampling of the parameters, will the generated samples be correlated? \n\n(2) Can you always find a good sampler function to approximate complex target distribution? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an amortized Langevin dynamics for sampling from a posterior distribution of a top-down generative model or from an energy-based model. The key is to recruit a sampler function that generates many samples directly in parallel, and run a single Langevin dynamics on the parameters of this sampler function. The method is illustrated by image generation. ", "main_review": "Strengths: \n\n(1) The idea of the paper is novel and interesting. Theoretical analysis is provided. \n\n(2) The proposed method can be applied to sampling both the posterior distribution and the energy-based model. \n\n(3) The proposed method is different from variational approximation. \n\nWeaknesses: \n\n(1) One thing I am concerned with is, if all the randomness is accounted for by the Langevin sampling of the parameters, will the generated samples be correlated? \n\n(2) Can you always find a good sampler function to approximate complex target distribution? ", "summary_of_the_review": "The paper proposes a new idea on sampling from unnormalized densities. It can be useful for learning deep generative models. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635842240449}, {"id": "Nqap9Km0Dt7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2820/Reviewer_Lr3z"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes to use amortized MCMC to learn latent space energy-based model (EBM). Specifically, instead of running Langevin dynamics on (latent) data space, such sampling is performed on the parameter space of inference model and is shown to converge to the true posterior under several mild conditions. The experiments demonstrate the generation quality of the proposed model.  ", "review_text": "(+) The paper considers an important and challenging problem. The idea discussed is quite interesting and somewhat new. \n\nHowever, there are some downsides of the current submission: \n\n(-) The presentation of the paper needs to be improved. Quite a few places are unclear which need to be better motivated or to be further described. For example, what does $u$ represents in Eqn. 14? Any motivations/insights on why consider Langevin process on the parameter space (the parameter space usually has much higher dimensions compared to the latent codes itself)? How to choose $g(x)$ as in theorem 1? If only the last layer is trainable, does the inference model parametrized in this way has very limited expressive power? Also, in algorithm 2, all the update is on the parameter space including generator as well as encoder. The paper should also need to be well motivated on why performing Langevin on their parameter spaces (as in Eqn 10, 11) as well. \n\n(-) The experiments need to be improved. The sample quality as well as score is relatively weak. Why not use reported FID of the baselines in table 1? For example, the LEBM achieves lower FID on cifar10 on their paper, why not choose their model structure and retrain the proposed model for the comparison?  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to use amortized MCMC to learn latent space energy-based model (EBM). Specifically, instead of running Langevin dynamics on (latent) data space, such sampling is performed on the parameter space of inference model and is shown to converge to the true posterior under several mild conditions. The experiments demonstrate the generation quality of the proposed model.  ", "main_review": "(+) The paper considers an important and challenging problem. The idea discussed is quite interesting and somewhat new. \n\nHowever, there are some downsides of the current submission: \n\n(-) The presentation of the paper needs to be improved. Quite a few places are unclear which need to be better motivated or to be further described. For example, what does $u$ represents in Eqn. 14? Any motivations/insights on why consider Langevin process on the parameter space (the parameter space usually has much higher dimensions compared to the latent codes itself)? How to choose $g(x)$ as in theorem 1? If only the last layer is trainable, does the inference model parametrized in this way has very limited expressive power? Also, in algorithm 2, all the update is on the parameter space including generator as well as encoder. The paper should also need to be well motivated on why performing Langevin on their parameter spaces (as in Eqn 10, 11) as well. \n\n(-) The experiments need to be improved. The sample quality as well as score is relatively weak. Why not use reported FID of the baselines in table 1? For example, the LEBM achieves lower FID on cifar10 on their paper, why not choose their model structure and retrain the proposed model for the comparison?  ", "summary_of_the_review": "The paper study an important problem, but the current presentation makes it hard to follow. Also the extra experiments and analysis need to be added to motivate and back-up their claims. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635722125599}, {"id": "vS63yFItIQl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2820/Reviewer_vRp6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Typically, approximating the (analytically intractable) posterior of deep latent variable models can be done in two ways, variational inference methods learn an (analytically tractable) distribution to approximate the true posterior. For large datasets where each dataitem has its own posterior distribution, the parameters of the optimal approximate distribution may be predicted as a function of the dataitem, the comon practice of “amortization”. Thus the approximate posterior is quick to infer at test time however it is also limited by the flexibility of the familty of distributino used for the analytic approximation.\n\nMeanwhile, Markov Chain Monte-Carlo methods generate a collection of samples to approximate true posterior. As each dataitem in a large dataset has a unique posterior, each would require its own Markov chain of samples which is a more accurate approximation, however is time consuming.\n\nThe author propose a mixture of these approaches. At a very high level, the authors propose to combine a single “communal” Markov chain, $\\Phi$, together with a function that maps dataitems into a weight vector, $g(x)$, that linearly maps the communal markov chain onto a markov chain that approximates the posterior for the given dataitem p(z|x)~$\\Phi g(x)$. Replace one chain per item (traditional MCMC) with a communal chain and one linear mapping per dataitem, where the mapping is amortized.\n\nThe authors describe application of this new trick to latent space of an autoencoder latent variable model and the sampling from an unconditioned density such as an energy based model. The MCMC algorithm of choice is the well known Langevin Dynamics solved by the Euler-Maruyama method. The authors describe training in minibatches simulataneously regenerating the MCMC chains which reduces to standard SGD with noise. Toy experiments show proof of concept and image generation tasks show positive results as measured by Frechet Inception distance.\n", "review_text": "I enjoyed the overall paper. I felt the presentation was perhaps a little unecessarily complex. I have a few technical concerns\n\n# Good points\n- I felt the writing and presentation of existing literature was generally very good, clear and concise.\n\n- the theoretical concept \"one chain for all datapoints\" (and linear projections) also feels like a good idea to me. The theorem requires that all samples for all posteriors $z^{(i)}$ must be varying linear projections ($g(x)$) of a single communal chain $\\Phi$. Or alternatively, all data item specific posterior distributions are linear projections of a single higher dimensional distribution. This sounds really cool!\n\n- I enjoyed the theorem and the discussion of the rank of the G matrix.\n\n\n# Concerns\n\n- In algorithm 2, the generative model weights, $\\Theta$, are also sampled using the Euler-Maruyama method, however this does not correspond to the theorem 2.1 which requires that the generative model $\\Theta$ is fixed (otherwise there is no true posterior to approximate). (Upon further inspection, it looks like the given loss function (14) is the evidence lower bound of the data and hence sampling model weights is similar to Bayesian neural network sampling of weights with the ELBO and not model evidence).\n\n- P4, line 3: Test time sampling: why not pregenerate a markov chain during/after training that can be reused again and again at test time? Surely the whole point of amortization is the “one-chain fits-all” hypothesis? If the authors propose resampling at test time then the proposed method is not strictly amortized and there is no actual speedup?\n\n- In the experiments, would it be possible to also report estimated marginal dataset likelihood $\\log p(x)$? There are many generative modelling works that use this as a metric (e.g. PixelCNN).\n\n- I am surprised by the *extremely* high variance of of the Frechet Inception distance among baselines, I struggle to believe that all the standard methods are *that* bad, presumably either the error bars are not accurate or the implementation of baselines is not ideal. For me, unfortunately, this make the paper hard to accept. (I realise this may contradict other venues/reviewers that emphasise empirical results).\n\n- in the experiments, I think it is valuable to know why LD and LD+EMB are worse than LAE? Surely LD + EMB is the gold standard (one MCMC chain per dataitem) that the LAE strives to achieve, while this may be computationally expensive, I believe that such a baseline should at least be implemented to outperform (or match) LAE, (i.e. agreeing with the theory) this could also show the MCMC amortization gap that this new amortization method introduces (again agreeing with the theory). Currently these are unanswered question that feels rather central to the contribution of the paper.\n\n- regarding Theorem 2.1, and the implication that all posteriors are linear projections of a bigger common distribution, this feels very sensitive to the big distribution dimension, currently this is not properly investigated. If the dimension is large, then the communal MCMC chain is no cheaper than parallelized low dim chains across dataitems, if the dimension is too small, this imposes too much similarity across all the posteriors and the method suffers, this trade off is not clear.\n\n# Minor Points\n\n- there appears almost no mention of Bayesian neural networks. MCMC sampling of network weights has been extensively studied, the proposed method is also MCMC sampling network weights while just using a different  distribution to sample from. Presumably, any Bayesian neural network approach for the encoder may be used? \n\n- If I understand correctly, training is using the Eurler Maruyama method and so the algorithm simplifies to SGD with Gaussian noise, presumably this makes convergence much slower than a VAE or VAE-flow which may use RMSprop or Adam or any method with momentum across minibatches?\n\n- (very minor) regarding the disadvantages of VAEs, the ELBO training objective is the KL divergence of true and approximate posterior, i.e. it makes a inference network Gaussian that looks like the posterior but it also, *theoretically*, it makes a posterior that looks Gaussian,  hence the claims that VAEs are existentially flawed feel slightly overstated to me, though I realise this is not a common opinion and in practice this is indeed very often not the case.\n\n- the presentation of the proposed method seems a little overly complex or difficult to follow, as I have described, I personally found it much easier to think in terms of a communal chain and linear projections, I am sure there must be many other intepretations that are simpler and easier to follow than the paper in its current form.\n\n\nI enjoyed the paper and I am less well read in the area hence absolute novelty I cannot confidently judge (I wouldn't be surprised if this idea is well established in the more pure \"non-deep\" statistics community) though I do believe the idea has merit and is worth accepting. However the rather chaotic numerical results are very disturbing.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Typically, approximating the (analytically intractable) posterior of deep latent variable models can be done in two ways, variational inference methods learn an (analytically tractable) distribution to approximate the true posterior. For large datasets where each dataitem has its own posterior distribution, the parameters of the optimal approximate distribution may be predicted as a function of the dataitem, the comon practice of “amortization”. Thus the approximate posterior is quick to infer at test time however it is also limited by the flexibility of the familty of distributino used for the analytic approximation.\n\nMeanwhile, Markov Chain Monte-Carlo methods generate a collection of samples to approximate true posterior. As each dataitem in a large dataset has a unique posterior, each would require its own Markov chain of samples which is a more accurate approximation, however is time consuming.\n\nThe author propose a mixture of these approaches. At a very high level, the authors propose to combine a single “communal” Markov chain, $\\Phi$, together with a function that maps dataitems into a weight vector, $g(x)$, that linearly maps the communal markov chain onto a markov chain that approximates the posterior for the given dataitem p(z|x)~$\\Phi g(x)$. Replace one chain per item (traditional MCMC) with a communal chain and one linear mapping per dataitem, where the mapping is amortized.\n\nThe authors describe application of this new trick to latent space of an autoencoder latent variable model and the sampling from an unconditioned density such as an energy based model. The MCMC algorithm of choice is the well known Langevin Dynamics solved by the Euler-Maruyama method. The authors describe training in minibatches simulataneously regenerating the MCMC chains which reduces to standard SGD with noise. Toy experiments show proof of concept and image generation tasks show positive results as measured by Frechet Inception distance.\n", "main_review": "I enjoyed the overall paper. I felt the presentation was perhaps a little unecessarily complex. I have a few technical concerns\n\n# Good points\n- I felt the writing and presentation of existing literature was generally very good, clear and concise.\n\n- the theoretical concept \"one chain for all datapoints\" (and linear projections) also feels like a good idea to me. The theorem requires that all samples for all posteriors $z^{(i)}$ must be varying linear projections ($g(x)$) of a single communal chain $\\Phi$. Or alternatively, all data item specific posterior distributions are linear projections of a single higher dimensional distribution. This sounds really cool!\n\n- I enjoyed the theorem and the discussion of the rank of the G matrix.\n\n\n# Concerns\n\n- In algorithm 2, the generative model weights, $\\Theta$, are also sampled using the Euler-Maruyama method, however this does not correspond to the theorem 2.1 which requires that the generative model $\\Theta$ is fixed (otherwise there is no true posterior to approximate). (Upon further inspection, it looks like the given loss function (14) is the evidence lower bound of the data and hence sampling model weights is similar to Bayesian neural network sampling of weights with the ELBO and not model evidence).\n\n- P4, line 3: Test time sampling: why not pregenerate a markov chain during/after training that can be reused again and again at test time? Surely the whole point of amortization is the “one-chain fits-all” hypothesis? If the authors propose resampling at test time then the proposed method is not strictly amortized and there is no actual speedup?\n\n- In the experiments, would it be possible to also report estimated marginal dataset likelihood $\\log p(x)$? There are many generative modelling works that use this as a metric (e.g. PixelCNN).\n\n- I am surprised by the *extremely* high variance of of the Frechet Inception distance among baselines, I struggle to believe that all the standard methods are *that* bad, presumably either the error bars are not accurate or the implementation of baselines is not ideal. For me, unfortunately, this make the paper hard to accept. (I realise this may contradict other venues/reviewers that emphasise empirical results).\n\n- in the experiments, I think it is valuable to know why LD and LD+EMB are worse than LAE? Surely LD + EMB is the gold standard (one MCMC chain per dataitem) that the LAE strives to achieve, while this may be computationally expensive, I believe that such a baseline should at least be implemented to outperform (or match) LAE, (i.e. agreeing with the theory) this could also show the MCMC amortization gap that this new amortization method introduces (again agreeing with the theory). Currently these are unanswered question that feels rather central to the contribution of the paper.\n\n- regarding Theorem 2.1, and the implication that all posteriors are linear projections of a bigger common distribution, this feels very sensitive to the big distribution dimension, currently this is not properly investigated. If the dimension is large, then the communal MCMC chain is no cheaper than parallelized low dim chains across dataitems, if the dimension is too small, this imposes too much similarity across all the posteriors and the method suffers, this trade off is not clear.\n\n# Minor Points\n\n- there appears almost no mention of Bayesian neural networks. MCMC sampling of network weights has been extensively studied, the proposed method is also MCMC sampling network weights while just using a different  distribution to sample from. Presumably, any Bayesian neural network approach for the encoder may be used? \n\n- If I understand correctly, training is using the Eurler Maruyama method and so the algorithm simplifies to SGD with Gaussian noise, presumably this makes convergence much slower than a VAE or VAE-flow which may use RMSprop or Adam or any method with momentum across minibatches?\n\n- (very minor) regarding the disadvantages of VAEs, the ELBO training objective is the KL divergence of true and approximate posterior, i.e. it makes a inference network Gaussian that looks like the posterior but it also, *theoretically*, it makes a posterior that looks Gaussian,  hence the claims that VAEs are existentially flawed feel slightly overstated to me, though I realise this is not a common opinion and in practice this is indeed very often not the case.\n\n- the presentation of the proposed method seems a little overly complex or difficult to follow, as I have described, I personally found it much easier to think in terms of a communal chain and linear projections, I am sure there must be many other intepretations that are simpler and easier to follow than the paper in its current form.\n\n\nI enjoyed the paper and I am less well read in the area hence absolute novelty I cannot confidently judge (I wouldn't be surprised if this idea is well established in the more pure \"non-deep\" statistics community) though I do believe the idea has merit and is worth accepting. However the rather chaotic numerical results are very disturbing.\n", "summary_of_the_review": "I like the paper, I feel that the writing could be simplified somewhat, the method treats latent posteriors over data points as unique linear projections of a global distribution over network final layer weights.\n\n- to me, the proposed method is an instance Bayesian neural networks with a different distribution (the ELBO) over weights (eg https://arxiv.org/abs/1902.02476, http://proceedings.mlr.press/v48/gal16.pdf)\n\n- the experimental results have very high variance amongst baselines, this is concerning, particularly, the \"truth\" that the proposed method aims to approximate performs *worse* that the approximation.\n\n- treating all dataitem posteriors as a linear projections of a single communal distribution seems risky, further comment would be appreciated.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635703431668}], "openreview_url": "https://openreview.net/forum?id=GIEPR9OomyX", "arxiv_id": "2209.07036", "paper_pdf": "papers/GIEPR9OomyX.pdf", "paper_pdf_sha256": "ec6a4622851af0812767f792117d7a2cf772caadca68c18932f84d15e4f17ad4", "paper_pdf_bytes": 2061206, "paper_pdf_source": "openreview", "code_url": "https://github.com/iShohei220/LAE", "code_repository": "iShohei220/LAE", "code_commit": "d4cfbdb8fdee520fa891610a6afb569dcb9540b6", "code_archive": "repos/GIEPR9OomyX.zip", "code_archive_sha256": "60d698fddb4b0817510629baed3f1acbf1cd844130a007b27a8315294a7f879c", "code_archive_bytes": 17968, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 13, "github_languages": {"Python": 59400}, "github_archived": false, "github_pushed_at": "2022-09-15T02:16:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/langevin-autoencoders-for-learning-deep-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6VhmvP7XZue", "year": 2021, "status": "rejected", "title": "Open-world Semi-supervised Learning", "authors": ["Kaidi Cao", "Maria Brbic", "Jure Leskovec"], "authorids": ["~Kaidi_Cao1", "mbrbic@cs.stanford.edu", "~Jure_Leskovec1"], "authors_source": "OpenReview API", "abstract": "Supervised and semi-supervised learning methods have been traditionally designed for the closed-world setting which is based on the assumption that unlabeled test data contains only classes previously encountered in the labeled training data. However, the real world is often open and dynamic, and thus novel previously unseen classes may appear in the test data or during the model deployment. Here, we introduce a new open-world semi-supervised learning setting in which the model is required to recognize previously seen classes, as well as to discover novel classes never seen in the labeled dataset. To tackle the problem, we propose ORCA, an approach that jointly learns a feature representation and a classifier on the labeled and unlabeled subsets of the data. The key idea in ORCA is in introducing uncertainty based adaptive margin that effectively circumvents the bias caused by the imbalance of variance between seen and novel classes. We demonstrate that ORCA accurately discovers novel classes and assigns samples to previously seen classes on standard benchmark image classification datasets, including  CIFAR and ImageNet. Remarkably, despite solving the harder task ORCA outperforms semi-supervised methods on seen classes, as well as novel class discovery methods on unseen classes, achieving 7% and 151% improvements on seen and unseen classes of the ImageNet dataset.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "VVzrM_7HtkB", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper938/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers open-world SSL settings  where the model recognizes previously seen classes, and detects novel classes which are not present in the labeled dataset. The method contains three losses to train a model in this setting: a) supervised loss on labeled data, b) unsupervised loss on unlabeled data from pseudo-labels obtained  from confident pairwise similarities, and c)  regularization term that avoids assigning all the unlabeled samples to the same class. Then the paper evaluates the effectiveness of the  method on CIFAR-10/100 and ImageNet-100 datasets.\n\nThe paper is well organized, well-written, and tackle SSL problem in a more realistic setting. Experiments are enough to some extent. However, I feel that the paper just combines several well-known and well-studied techniques for different tasks which intersect with open-world SSL settings. (e.g., self-supervised learning in SSL, clustering for SSL or clustering for transferring knowledge  across domains and tasks, using confident pseudo-labels for training an SSL model, etc). From this perspective, technical novelty of the work is limited, although experiments show good results for open-world SSL setting. \n\n\n-First, I think this set-up does not contain all the scenarios for a real-world SSL. For example,  this set-up does not consider covariate shift where the data belonging to the same classes but different image statistics or style (e.g., dogs in natural images, dogs in the painting or sketches images). However, in the real-world scenario, we may have images of the same class but different domains. \n\n\n-The paper adds several losses studied in the literature without analyzing if any has a negative effect on the others in the open-world SSL setting.\n\n-Assuming that the number of novel classes |C_u| is known is a bit unrealistic to me in the open-world SSL setting.\n\n-The method ranks the distances, and for each sample generates the pseudo-label for its most similar neighbor. However,  in an open world SSL setting due to the very-limited-label regime, the representation may not be ideal, and  therefore, pairwise-labeling may not be ideal and can possibly propagate the error through the network over the course of training.\n\n\n-Does the regularization towards uniform distribution consider unbalanced novel classes which is common in open-world SSL? \n\n-Does the method perform well in cases where the test set contains novel classes that do not appear in the unlabeled set?\n\nGenerally, the work has some potential from practical point of view, however, it needs more work to be improved technically.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Open-world Semi-supervised Learning", "review": "The paper considers open-world SSL settings  where the model recognizes previously seen classes, and detects novel classes which are not present in the labeled dataset. The method contains three losses to train a model in this setting: a) supervised loss on labeled data, b) unsupervised loss on unlabeled data from pseudo-labels obtained  from confident pairwise similarities, and c)  regularization term that avoids assigning all the unlabeled samples to the same class. Then the paper evaluates the effectiveness of the  method on CIFAR-10/100 and ImageNet-100 datasets.\n\nThe paper is well organized, well-written, and tackle SSL problem in a more realistic setting. Experiments are enough to some extent. However, I feel that the paper just combines several well-known and well-studied techniques for different tasks which intersect with open-world SSL settings. (e.g., self-supervised learning in SSL, clustering for SSL or clustering for transferring knowledge  across domains and tasks, using confident pseudo-labels for training an SSL model, etc). From this perspective, technical novelty of the work is limited, although experiments show good results for open-world SSL setting. \n\n\n-First, I think this set-up does not contain all the scenarios for a real-world SSL. For example,  this set-up does not consider covariate shift where the data belonging to the same classes but different image statistics or style (e.g., dogs in natural images, dogs in the painting or sketches images). However, in the real-world scenario, we may have images of the same class but different domains. \n\n\n-The paper adds several losses studied in the literature without analyzing if any has a negative effect on the others in the open-world SSL setting.\n\n-Assuming that the number of novel classes |C_u| is known is a bit unrealistic to me in the open-world SSL setting.\n\n-The method ranks the distances, and for each sample generates the pseudo-label for its most similar neighbor. However,  in an open world SSL setting due to the very-limited-label regime, the representation may not be ideal, and  therefore, pairwise-labeling may not be ideal and can possibly propagate the error through the network over the course of training.\n\n\n-Does the regularization towards uniform distribution consider unbalanced novel classes which is common in open-world SSL? \n\n-Does the method perform well in cases where the test set contains novel classes that do not appear in the unlabeled set?\n\nGenerally, the work has some potential from practical point of view, however, it needs more work to be improved technically.  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603960341977}, {"id": "H7HNQOpmcaz", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper938/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros\n- The paper explores an interesting semi-supervised learning (SSL) setting in which the unlabeled data contain not only the seen class but also novel classes. The problem is interesting in that it is more practical than the classic SSL setting in the real world and has been seldomly researched. \n- The method includes a self-supervised pretraining step and a finetune step. The finetune step jointly solves the classification task and the clustering task with a unified objective loss. The supervised loss, which is the main contribution of the method, overcomes the imbalance problem caused by BCE with a novel adaptive margin-based loss. The authors validate this loss with empirical results.\n- The paper is well-organized and clearly written. As far as I can see, the method is technically sound.\n\nCons\n- The comparison methods (such as pseudo-labeling, DS3L) are not strong enough. It is suggested to compare with more SOTA methods. Since SimCLR is more powerful than RotNet, it is strange to see RankStats get worse performance than the original paper.\n- The number of novel classes has to be prefixed.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposed a novel method for solving open-world semi-supervised learning problem.", "review": "Pros\n- The paper explores an interesting semi-supervised learning (SSL) setting in which the unlabeled data contain not only the seen class but also novel classes. The problem is interesting in that it is more practical than the classic SSL setting in the real world and has been seldomly researched. \n- The method includes a self-supervised pretraining step and a finetune step. The finetune step jointly solves the classification task and the clustering task with a unified objective loss. The supervised loss, which is the main contribution of the method, overcomes the imbalance problem caused by BCE with a novel adaptive margin-based loss. The authors validate this loss with empirical results.\n- The paper is well-organized and clearly written. As far as I can see, the method is technically sound.\n\nCons\n- The comparison methods (such as pseudo-labeling, DS3L) are not strong enough. It is suggested to compare with more SOTA methods. Since SimCLR is more powerful than RotNet, it is strange to see RankStats get worse performance than the original paper.\n- The number of novel classes has to be prefixed.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603933380757}, {"id": "TOGfigg2Ixu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper938/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "--Paper summary--\n\nThe authors propose a method to tackle a new problem setting of semi-supervised learning, called an open-world semi-supervised learning, where the model is required to accurately discriminate known-class data as well as to appropriately discover unknown classes contained in an unlabeled dataset. The objective function to be minimized in the proposed method comprises three terms: unsupervised loss, supervised loss with uncertainty based adaptive margin, and entropy regularization. The experimental results with several datasets have validated the advantage of the proposed method.\n\n--Review summary--\n\nThis study is well-motivated and tackles an important problem that would occur in real-world applications. The design of the proposed method seems reasonable, and it works well in the experiments with several datasets. However, some points remain unclear or are not convincing, which makes my score a bit conservative. To appropriately determine my score, I would appreciate if the authors clarify these points in their response.\n\n--Details--\n\nStrength\n\n- The setting of the open-world semi-supervised learning is interesting and should be practically important.\n- The paper is well-written and well-organized.\n\nWeakness and concerns\n\n- It is not clear whether the design of the proposed method really leads to the improvement of the performance or not. Since the authors use SimCLR to pretrain the model, the pretrained model has already had a substantially good feature representation. Consequently, if we know the number of classes in the training data, unsupervised learning methods like [R1] (or simple clustering methods) might work well to discriminate all classes. Although the authors adopt a new idea to the supervised loss, it seems unclear whether the supervised loss itself really contributes to good performance. Are there any experimental result on performance sensitivity to \\eta?  \n[R1] \"Learning Discrete Representations via Information Maximizing Self-Augmented Training,\" ICML 2017.\n- The experimental setup is not convincing. Due to the motivation of SSL, the number of unlabeled data should be much larger than that of labeled data, and it is often tens or hundreds times larger in the literature. However, in this paper, it is basically 3~7 times larger in the experiments. Is there any reason why the authors did not much reduce the number of labeled data?\n- Two concerns on uncertainty based adaptive margin.\n\t- How did the authors estimate the class posterior probability in Eq. (4)? Is it just the output of the softmax function?\n\t- Since this adaptive margin is adopted to improve the accuracy of pseudo-labels, how much it is improved should be reported in Fig. 3.\n- How did the authors conduct validation? Due to the existence of unseen classes in unlabeled data, how to conduct validation is not trivial.\n- Is it reasonable to call L_BCE unsupervised loss? Since Z_l' is obtained using ground-truth labels, L_BCE cannot be computed in an unsupervised manner.\n\nMinor concerns\n- I could not get the reason why the proposed method is called ORCA.\n\n\n--After receiving authors' response--\n\nI would like to thank the authors for providing additional experimental results and giving clarifications. Since my concerns are almost solved, and I updated my score from 5 to 6. This study would be a great first step to tackle the challenging problem, open-world semi-supervised learning, though there are several remaining issues (e.g., validation is hard to conduct, the number of unknown classes should be known, etc.). I vote for \"weak accept.\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This study is well-motivated and tackles an important problem, but some points remain unclear.", "review": "--Paper summary--\n\nThe authors propose a method to tackle a new problem setting of semi-supervised learning, called an open-world semi-supervised learning, where the model is required to accurately discriminate known-class data as well as to appropriately discover unknown classes contained in an unlabeled dataset. The objective function to be minimized in the proposed method comprises three terms: unsupervised loss, supervised loss with uncertainty based adaptive margin, and entropy regularization. The experimental results with several datasets have validated the advantage of the proposed method.\n\n--Review summary--\n\nThis study is well-motivated and tackles an important problem that would occur in real-world applications. The design of the proposed method seems reasonable, and it works well in the experiments with several datasets. However, some points remain unclear or are not convincing, which makes my score a bit conservative. To appropriately determine my score, I would appreciate if the authors clarify these points in their response.\n\n--Details--\n\nStrength\n\n- The setting of the open-world semi-supervised learning is interesting and should be practically important.\n- The paper is well-written and well-organized.\n\nWeakness and concerns\n\n- It is not clear whether the design of the proposed method really leads to the improvement of the performance or not. Since the authors use SimCLR to pretrain the model, the pretrained model has already had a substantially good feature representation. Consequently, if we know the number of classes in the training data, unsupervised learning methods like [R1] (or simple clustering methods) might work well to discriminate all classes. Although the authors adopt a new idea to the supervised loss, it seems unclear whether the supervised loss itself really contributes to good performance. Are there any experimental result on performance sensitivity to \\eta?  \n[R1] \"Learning Discrete Representations via Information Maximizing Self-Augmented Training,\" ICML 2017.\n- The experimental setup is not convincing. Due to the motivation of SSL, the number of unlabeled data should be much larger than that of labeled data, and it is often tens or hundreds times larger in the literature. However, in this paper, it is basically 3~7 times larger in the experiments. Is there any reason why the authors did not much reduce the number of labeled data?\n- Two concerns on uncertainty based adaptive margin.\n\t- How did the authors estimate the class posterior probability in Eq. (4)? Is it just the output of the softmax function?\n\t- Since this adaptive margin is adopted to improve the accuracy of pseudo-labels, how much it is improved should be reported in Fig. 3.\n- How did the authors conduct validation? Due to the existence of unseen classes in unlabeled data, how to conduct validation is not trivial.\n- Is it reasonable to call L_BCE unsupervised loss? Since Z_l' is obtained using ground-truth labels, L_BCE cannot be computed in an unsupervised manner.\n\nMinor concerns\n- I could not get the reason why the proposed method is called ORCA.\n\n\n--After receiving authors' response--\n\nI would like to thank the authors for providing additional experimental results and giving clarifications. Since my concerns are almost solved, and I updated my score from 5 to 6. This study would be a great first step to tackle the challenging problem, open-world semi-supervised learning, though there are several remaining issues (e.g., validation is hard to conduct, the number of unknown classes should be known, etc.). I vote for \"weak accept.\"\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603777157051}, {"id": "c0n7NieyJWz", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper938/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: the authors propose Open World Semi-Supervised Learning (ORCA), a semi-supervised method that learns to classify previously seen classes in the labeled data and novel class in the unlabeled data. The proposed method is end to end and achieves significant improvement on ImageNet dataset compared to baseline methods.\n\nPros:\nThe paper takes one of the most important issue of semi-supervised learning: recognizing novel classes. For me, the problem itself is real and practical.\n\nThe experiments are strong and methods shows significant improvement compared to competing methods.\n\nCons:\nThe paper lacks clarity in some sections. The writing could be better and the notations better defined.\n\nComments:\nThe unsupervised loss function is not clear to me, I find it difficult to parse equation 1 from the way it's written. Could you please explain the equation better?\n\nIt's my understanding that each z_i in the unlabeled set is assigned a class based on the ranking of pairwise distance in the mini batch.  If the mini batch of z_i contains labeled samples, then it could be assigned to the labels example if the images are similar. For example if there's an unlabeled example of a horse, it could be assigned to the labeled donkey class. Wouldn't this be problematic for the novel class prediction if the batches always contain similar examples from different classes? \n\nWhy the choice of cosine distance for the similarity measure? Using cosine similarity means that you do not consider the magnitude of the embeddings. This would make sense for text embeddings but the experiments are on image datasets", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "OPEN-WORLD SEMI-SUPERVISED LEARNING", "review": "Summary: the authors propose Open World Semi-Supervised Learning (ORCA), a semi-supervised method that learns to classify previously seen classes in the labeled data and novel class in the unlabeled data. The proposed method is end to end and achieves significant improvement on ImageNet dataset compared to baseline methods.\n\nPros:\nThe paper takes one of the most important issue of semi-supervised learning: recognizing novel classes. For me, the problem itself is real and practical.\n\nThe experiments are strong and methods shows significant improvement compared to competing methods.\n\nCons:\nThe paper lacks clarity in some sections. The writing could be better and the notations better defined.\n\nComments:\nThe unsupervised loss function is not clear to me, I find it difficult to parse equation 1 from the way it's written. Could you please explain the equation better?\n\nIt's my understanding that each z_i in the unlabeled set is assigned a class based on the ranking of pairwise distance in the mini batch.  If the mini batch of z_i contains labeled samples, then it could be assigned to the labels example if the images are similar. For example if there's an unlabeled example of a horse, it could be assigned to the labeled donkey class. Wouldn't this be problematic for the novel class prediction if the batches always contain similar examples from different classes? \n\nWhy the choice of cosine distance for the similarity measure? Using cosine similarity means that you do not consider the magnitude of the embeddings. This would make sense for text embeddings but the experiments are on image datasets", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603764865147}], "openreview_url": "https://openreview.net/forum?id=6VhmvP7XZue", "arxiv_id": "2102.03526", "paper_pdf": "papers/6VhmvP7XZue.pdf", "paper_pdf_sha256": "553138b2d0a9c4b7c990374fbb79dfb52bb5171bea9b220ebc0eab012b88d2c4", "paper_pdf_bytes": 1169175, "paper_pdf_source": "openreview", "code_url": "https://github.com/snap-stanford/orca", "code_repository": "snap-stanford/orca", "code_commit": "5f33afdeb0aa1ec51d62c9f926f265504ccf9efc", "code_archive": "repos/6VhmvP7XZue.zip", "code_archive_sha256": "ceef60e184d444cf68895244badcf770b14c988fed98df8a714eae641872efd6", "code_archive_bytes": 16865, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 13, "github_languages": {"Python": 43535}, "github_archived": false, "github_pushed_at": "2022-02-17T04:35:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/open-world-semi-supervised-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJe4oxHYPB", "year": 2020, "status": "rejected", "title": "Winning the Lottery with Continuous Sparsification", "authors": ["Pedro Savarese", "Hugo Silva", "Michael Maire"], "authorids": ["savarese@ttic.edu", "hugoandradesilva664@gmail.com", "mmaire@uchicago.edu"], "authors_source": "OpenReview API", "abstract": "The Lottery Ticket Hypothesis from Frankle & Carbin (2019) conjectures that, for typically-sized neural networks, it is possible to find small sub-networks which train faster and yield superior performance than their original counterparts. The proposed algorithm to search for such sub-networks (winning tickets), Iterative Magnitude Pruning (IMP), consistently finds sub-networks with 90-95% less parameters which indeed train faster and better than the overparameterized models they were extracted from, creating potential applications to problems such as transfer learning.\n\nIn this paper, we propose a new algorithm to search for winning tickets, Continuous Sparsification, which continuously removes parameters from a network during training, and learns the sub-network's structure with gradient-based methods instead of relying on pruning strategies. We show empirically that our method is capable of finding tickets that outperforms the ones learned by Iterative Magnitude Pruning, and at the same time providing up to 5 times faster search, when measured in number of training epochs.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bklcxb3i5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2504/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nTo the authors of paper 2504: I have posted a private comment for the reviewer/AC discussion period based on your author responses and your revised paper. I want you to be able to see my full response, but I can't post additional public comments to the paper. As such, I'm editing my review with exactly what I sent to the AC.\n\nThank you to the authors for your thoughtful rebuttal and for updating the paper with both new text and (more impressively) new experiments. I read all three of your rebuttals and re-read the revised paper in detail. My comment to the AC is at the bottom of this message. I don't know if you will get an alert that the review was updated, but I hope you get a chance to take a look.\n\n===========================================\n\nEXECUTIVE SUMMARY OF REVIEW\n\nSummary of paper: This paper proposes a technique that simultaneously trains a neural network and learns a pruning mask. The goal of this technique is to make it faster to retroactively find sparse subnetworks that, from a point early in training, could have trained in isolation to the same performance as the full network (\"winning tickets\"). The best subnetworks found by this technique outperform those found by existing techniques [1] at every sparsity level.\n\nSummary of review: The technique introduces several new hyperparameters whose values are asserted without describing the extent of the search necessary to find them; it unclear whether the cost of the search cancels out the efficiency gains. In addition, the proposed technique is inconsistent across runs. The sparsity and accuracy of the subnetworks it produces vary greatly, and there are no hyperparameters to explicitly control these outcomes. It is also unclear whether results vary from run to run even with the same hyperparameters.  As such, this technique is not clearly a cost reduction as compared to existing approaches in [1] when it comes to studying lottery tickets. Moreover, the paper implicitly proposes a new pruning technique, and it should be evaluated against other related techniques in the pruning literature. Finally, the evaluation needs more rigor.\n\nConclusion: It is difficult to determine whether the technique is an improvement over existing methods since the true costs of using it in practical workflows are unclear. Weak reject.\n\nOpportunity to improve score: Include a more detailed analysis of the overall costs of finding winning tickets at a target sparsity using the proposed technique, particularly including the costs of hyperparameter search necessary to do so. Include multiple replicates of experiments, experiments on more networks, and information about the variance of performance across runs with the same hyperparameters. Compare against other state-of-the-art pruning techniques.\n\nPROBLEM STATEMENT AND PROPOSED SOLUTION\n\nProblem: Winning tickets are currently expensive to find. The best known procedure [1, 2] is \"iterative magnitude pruning\" (IMP), which involves repeatedly training a network to completion, pruning by a fixed percentage, and \"rewinding\" weights to an early iteration of training until the network reaches the desired level of sparsity. To reach sufficient sparsity on standard networks, this procedure must be repeated 10 or more times.\n\nGoal: To propose a procedure that finds winning tickets more efficiently.\n\nSignificance: A more efficient procedure would make it easier to study the lottery ticket phenomenon. (Whether studying that phenomenon is, itself, significant is debatable.) Personally, I have extensive experience using IMP to find winning tickets, so techniques to reduce the cost of finding winning tickets would be very valuable for my work.\n\nProposed solution: The authors propose \"continuous sparsification\" (CS), which makes it possible to learn which weights to prune simultaneously with the weights themselves. Accompanying each parameter w in the network is a second parameter s. The actual weight used in the network is w * sigmoid(s * beta), where beta is a temperature hyperparameter. If the learned value of s is such that sigmoid(s * beta) is approximately 0 at the end of training, then the parameter has been pruned. The value of beta increases exponentially throughout training, meaning the output of the sigmoid will be closer to a hard 0 or 1, producing a pruning mask. To ensure sparsification happens, a regularization term lambda |sigmoid(beta * s)| is added. When run over multiple iterations, values of s are reset to their original values for weights that are not pruned. In addition, weights are either rewound (as in IMP) or left at their final values (as in [3]).\n\nNovelty: The paper is slightly novel. The technique is a variation of those proposed in [4] and [5], but the changes are meaningful. This is the first known use for finding winning tickets, but any pruning technique could hypothetically be used for this purpose. In effect, the paper proposes a new pruning technique that is primarily evaluated for its efficacy in finding winning tickets.\n\nTECHNICAL REVIEW\n\n* This technique introduces several new hyperparameters: lambda, initial and final betas, and the initial values for s. The paper suggests good values for these hyperparameters for both networks considered. These values were presumably found through hyperparameter search of some kind. How extensive was this hyperparameter search, and what is the range of \"good\" combinations of values? This is not just a methodological footnote; this paper's stated goal is to improve the efficiency of finding winning lottery tickets, and - if good hyperparameters are hard to find - then that defeats the purpose of a more efficient technique. IMP, while far less efficient epoch for epoch as compared to CS, requires no hyperparameter search; global pruning 20% of parameters per iteration seems to work well in general [2]. In a revised version of the paper, I would be eager to learn more about this set of tradeoffs, since that is what matters in practice.\n\n* The authors only study two networks: a toy convolutional network and a small Resnet. It is hard to draw broad conclusions from such a limited set of examples. I would be particularly interested in seeing how this technique performs on a large-scale network for ImageNet (e.g., Resnet-50), since these are the situations where IMP becomes particularly cost-prohibitive. If the technique works well in these settings, it would enable lottery ticket research at much larger scales than is currently possible. If the technique works as efficiently at this scale, then doing could even be feasible during the rebuttal period. (I acknowledge getting experiments working on ImageNet is no small undertaking in terms of both engineering time and cost, but it would improve my confidence to see those results.)\n\n* The authors did an admirably careful job replicating the networks in [1], which include a variety of nonstandard hyperparameters.\n\n* Did the paper study Resnet-18 (a network designed for ImageNet with 11.2M parameters) or Resnet-20 (a network designed for CIFAR-10 with 272K parameters)? Frankle et al. [1, 2] describe Resnet-20 in their appendices but mistakenly refer to it as Resnet-18 throughout both papers, so I wanted to clarify. Based on the final test accuracy of the network, it appears to be Resnet-20; if so, I'd urge you to call it as such and note in a parenthetical or footnote that it's the same network as in [1, 2] but with Frankle et al's mistaken name corrected.\n\n* Are the values of s0 for CS sampled from a distribution, or are they all the same, fixed value?\n\n* Do the extra parameters lead to longer wall-clock training times? If you are making an argument about a more efficient technique, this is an important consideration.\n\n* Figure 2 includes the same graph twice. I believe a different graph should appear on the left, and I am eager to take a look at it in a revised version of the paper.\n\n* It does not appear that multiple replicates of each experiment were run with different network initializations in Figure 3. There don't appear to be any error bars on Figure 3, suggesting that this only represents a single initialization. Considering the wide variance in performance achieved by continuous sparsification across runs in Figure 3 (right), this graph needs to include multiple runs and error bars. (I assume the multiple runs shown in Figure 3 (right) are with different hyperparameters?)\n\n* What is the performance of CS as a pruning technique? By finding winning tickets, CS is also implicitly pruning the network. Is this competitive with L0 Regularization [4]? Is it more efficient than iterative magnitude pruning as in [3]? If this is a more efficient technique for finding winning tickets, it is also liable to be a more efficient pruning method, which seems like even broader impact for the proposed technique. Alternatively, if this technique is less effective than comparable work (especially [4]) as a pruning method, then it is possible comparable work (especially [4]) might also produce better winning tickets, in which case the importance of this work is diminished. In an updated version of the paper, I would be interested in seeing an evaluation of CS as a pruning technique independent of the lottery ticket hypothesis (and compared to standard techniques in the pruning literature as such). I have a hard time seeing any reason why new techniques for finding lottery tickets are any different than new pruning techniques, and they should be evaluated as such in the context of the broader literature on pruning.\n\n* The last paragraph of Section 4.1 is missing important details that are necessary to evaluate the utility of continuous sparsification in comparison to IMP. \"In only two iterations, CS finds a ticket with over 77% sparsity...\" - for which hyperparameters, and how many hyperparameters had to be explored to find these values? How was the pareto curve obtained? How many different runs were necessary to create it? How many total epochs of training did it take to find that curve? If obtaining this pareto curve required many runs of CS, then it may not be any more efficient than running IMP in practice. The pareto curves appear to make CS look misleadingly effective, since they hide many of the actual costs involved (e.g., hyperparameter search, number of separate runs that were conducted to produce the curve, etc.). Greater transparency of this aspect of the paper would go a long way toward increasing my confidence that the findings are an improvement over IMP. \n\n* The sparsity and accuracy of subnetworks found by CS appears to vary widely from run to run as shown in Figure 3 (right). The final sparsity appears to be a function of the values of s0, lambda, and beta, potentially along with luck from the optimization process. For the same values of s, lambda, and beta, how widely do the final sparsities and accuracies vary? If I wanted to find a winning ticket with a particular sparsity using CS, what would the procedure look like for doing so? Would I have to sweep across values of these hyperparameters, or is there a more straightforward way to do so? The practical usefulness of the procedure hinges on these questions.\n\n* \"We associate the performance drop of highly sparse tickets found by our method from the second iteration onwards to the lack of weight rewinding.\" Why didn't you also try it with weight rewinding? That seems like an easy way to evaluate this hypothesis. For both networks, it would be interesting to see the performance of CS with and without rewinding (analogous to Appendix B in [1]).\n\n* In the literature so far, the only winning tickets to be examined are those from IMP [1, 2]. CS is a different technique, and it likely finds different winning tickets. Do these winning tickets have different properties from [1, 2]? Can they be reinitialized? Can they be rewound earlier? These comparisons seem like an interesting scientific opportunity.\n\n* There are a few additional comparisons that I think are vital to include in the paper to appropriately contextualize results.  They're in the bullets below.\n\n* First comparison: it would be useful to include random reinitialization or random pruning baselines as in [1, 2] simply to make it easier for the reader to contextualize the performance of other sparse subnetworks.\n\n* Second comparison: what happens if you run IMP such that each iteration prunes to the same sparsity as achieved by each iteration of CS? Perhaps pruning by a fixed amount per iteration in IMP is wasteful, and one can prune more aggressively during earlier iterations as CS naturally appears to do. In other words, one way of explaining the advantage of CS would be that it prunes more aggressively. Is this indeed the case? I would be very curious to know.\n\n* Third comparison: What are the range of results achieved if IMP and CS are run \"one-shot\" (pruning after just one iteration; in the case of IMP, pruning directly to a desired sparsity)? That is, how well can these techniques do with just a single iteration?\n\nWRITING\n\nThe writing is excellent. The prose is clear, and I was able to fully understand a relatively sophisticated technique on the first read through the paper. Writing of this quality is rare, and the authors should be commended for it.\n\nOVERALL: Weak Reject\n\nThe problem statement is that IMP is not efficient. The paper claims that CS is more efficient. However, the paper does not present a convincing case that CS is, on the whole, more efficient when taking into account hyperparameter search to get CS to work, hyperparameter search to target a particular sparsity, potential variance across runs of CS, potential additional training costs of CS, and the possibility that IMP might be able to work comparable well given a more aggressive pruning schedule.\n\nIn addition, the evaluation needs more experiments, including multiple replicates for each experiment and more networks (ideally one on ImageNet).\n\nFinally, I am unclear on what distinguishes CS from any other pruning technique, and it should be evaluated in the context of the broader pruning literature.\n\nIf the authors clarify that the overall cost of CS (including all of the factors listed above) is lower than for IMP, address technical concerns about the evaluation, and evaluate CS as a pruning technique in an updated version of the paper, I will update my score accordingly.\n\n[1] Frankle & Carbin. \"The Lottery Ticket Hypothesis.\" ICLR 2019.\n[2] Frankle et al. \"Stabilizing the Lottery Ticket Hypothesis.\" Arxiv.\n[3] Han et al. \"Learning both Weights and Connections for Efficient Neural Networks.\" NeurIPS 2015.\n[4] Louizos et al. \"Learning Sparse Neural Networks through L0 Regularization.\" ICLR 2018\n[5] Zhou et al. \"Deconstructing Lottery Tickets: Signs, Zeros, and the SuperMask.\" NeurIPS 2019.\n\n=====================================\n\nCOMMENT TO THE AC AFTER READING REBUTTALS AND REVISED PAPER\n\nTLDR\n\nI believe the technique is novel and should be published, regardless of whether it is actually more efficient than IMP for practical workflows.\n\nHowever, I believe the current evaluation is inadequate, both in the experiments and in the way data is presented. Namely, it is impossible to actually compare the costs of CS and IMP in the scenarios the authors evaluate, despite the fact that efficiency is the authors' claimed contribution. From the data as presented, it is unclear whether CS is actually more efficient than IMP for scientific use-cases. I'm not particularly concerned if there isn't an efficiency advantage - it's an innovative contribution regardless. My concern is that it is impossible to compare the costs of these techniques given the current presentation of data. That speaks to flaws in the evaluation section.\n\nI therefore maintain my score: weak reject. The technique deserves to be published, but the paper in its current form does not.\n\nINTRODUCTION\n\nAfter reading the original submission, I had the following questions:\n\n1) Right now, the only way to find winning lottery tickets is through training the network and pruning it. IMP can be instantiated with any pruning technique, and the authors are really proposing a new pruning technique for use in the IMP framework. How does continuous sparsification perform as a pruning technique for network compression (i.e., independent of the lottery ticket hypothesis)?\n\n2) How efficient is continuous sparsification (CS) in the scientific use cases where one would seek to find winning lottery tickets? In my research experience, there are two such use cases: (a) producing a winning lottery ticket with a specific sparsity and (b) producing winning lottery tickets across the full range of different sparsities.\n\nIn their reubttal, the authors addressed these underlying questions:\n\nAS A PRUNING TECHNIQUE\n\n\"We have added new experiments showing that our method yields competitive results when pruning VGG trained on CIFAR, outperforming both magnitude pruning and stochastic l0 regularization\"\n\nMy response: The initial results that the authors present are quite impressive on a VGG-style network for CIFAR-10. However, I find it concerning that the authors study pruning on this specific model but no others, particularly because they focus on a different network (namely, Resnet-20) in all other experiments in the paper. Since the authors are, in essence, proposing a new pruning technique, I would like to see it comprehensively evaluated as such on a range of networks against other baselines (as the authors do in Figure 4 for one network - those baselines are fine to me).\n\nEFFICIENCY OF CONTINUOUS SPARSIFICATION\n\n- It increases the cost of each individual network training run slightly: \"Continuous Sparsification on a 1080 Ti: our method resulted in 15% extra wall-clock time per training epoch.\"\n\n- PRODUCING WINNING TICKETS ACROSS THE FULL RANGE OF SPARSITIES: Throughout the paper, the authors present \"pareto curves\" showing the highest accuracy achieved by CS subnetworks at various sparsities. I find these pareto curves misleading: they hide the fact that CS had to be run several times with different hyperparameters to produce these curves.\n\nFor example, in Figure 6 (right), the authors produce the pareto curves by training Resnet-20 (by my count) 22 times (each point along the purple lines). The corresponding IMP curve appears to have 14 points. In other words, to get winning lottery tickets across all sparsities, CS must be run for more iterations than IMP - it appears to be less efficient, contradicting the authors' core claim. The best argument in favor of CS is that one could perform each of these runs in parallel if sufficient GPUs were available, meaning less wall-clock time would be required.\n\nOne caveat to this analysis: the authors state that they run CS \"without rewinding,\" meaning that the second iteration of CS requires less training time than fully training the network (as IMP requires). The authors do not state how long they train when they *aren't* rewinding, so it is impossible to compare the efficiency of CS with IMP. All they say is that it \"allow[s] for even faster ticket search.\" They also do not study IMP without rewinding, which would be a helpful baseline for comparing to CS without rewinding.\n\n***In short, if the authors are making an argument that one technique is more efficient than the other on an epoch-for-epoch basis, they need to actually plot the epochs required by each technique.***\n\n- PRODUCING A WINNING TICKET AT A SPECIFIC SPARSITY: My concern about this use case is that there is a non-intuitive relationship between the initialization of the sparsity parameters and the final sparsity of the network. As the authors state in the rebuttal: \"To achieve a desired sparsity, one can either perform runs in parallel with different values for s_0, or perform sequential binary search if the goal is to minimize the overall computational cost and not wall-clock time.\" In other words, there is no precise way to target a particular sparsity other than trying many hyperparameter configurations. The authors do not provide any concrete costs of using CS in this way in comparison to IMP, and - from a usability perspective - this is a challenging workflow.\n\nOTHER CONCERNS\n\n- The authors only study CS on one toy network (the six-layer convolutional network, which - in my experience - is a particularly easy setting compared to deeper networks) and one \"real\" network (Resnet-20 on CIFAR-10). I would like to see results on other networks for CIFAR-10 (for example, the VGG network in Section 4.3).\n\n- More importantly, I would like to see results on an ImageNet network. If CS makes finding winning lottery tickets more efficient as the authors claim, then finding winning tickets efficiently on an ImageNet network should be an excellent demonstration of their contribution. This scenario has stretched IMP to its breaking point, as the authors note.\n\nARGUMENTS IN FAVOR OF ACCEPTING\n\n+ The authors propose a new pruning technique that appears to improve upon the increasingly popular L0-regularization technique.\n\n+ With the right hyperparameters, the proposed technique makes it possible to find winning lottery tickets more efficiently than existing methods (i.e., IMP with magnitude pruning).\n\n+ The winning tickets found by the proposed technique reach higher accuracy than IMP winning tickets and produce winning tickets at more extreme sparsities, improving upon our knowledge of the existence of winning lottery tickets.\n\nARGUMENTS IN FAVOR OF REJECTING\n\n- The authors perform only minimal evaluation of their method as a pruning technique. It is possible that this is a missed opportunity to show an additional contribution. It is also possible that other, existing pruning techniques outperform CS at both pruning and finding winning lottery tickets.\n\n- The technique is hard to use for the two existing use cases for finding winning tickets. In particular, there is no way to search for winning tickets at a specific sparsity.\n\n- It is unclear whether CS is actually more efficient than IMP on an epoch-for-epoch basis, even though this is the main claimed contribution. The authors do not disclose - let alone plot - the number of training epochs required to find (a) winning tickets at a specific sparsity and (b) winning tickets at a range of sparsities, so it is impossible to make these comparisons. Meanwhile, the pareto curves the authors present are misleading, since they are the amalgamation of many separate runs of CS.\n\n- The authors study their technique on only one \"real\" network (Resnet-20) for finding winning tickets and a separate \"real\" network (VGG-19) for pruning. The authors do not show how CS performs in other challenging settings, especially ImageNet.\n\n- This work is only valuable if we believe that \"lottery ticket hypothesis\" work is valuable. In other words, this is a narrow contribution to an already-narrow area of study. This is one reason why pitching CS as a pruning technique would make this a stronger paper. I personally believe that \"lottery ticket\" work is valuable area of study, but I understand that it may not be seen as such in the broader ICLR community.\n\nCONCLUSION\n\nI believe the technique deserves to be published regardless of whether it is actually more efficient than IMP. It is a novel contribution to both our knowledge of pruning and of finding winning lottery tickets. The paper is exceptionally well written, and - as such - I believe it will help to inspire other research in this area.\n\nHowever, I do not believe that the current paper - namely, the current evaluation - should be published. The paper presents no concrete data on the comparative costs of performing CS and IMP even though the core claim is that CS is more efficient. The paper does not disclose enough detail to compute these costs, and it seems like CS is more expensive than IMP for standard workflows. Moreover, the current presentation of the data through \"pareto curves\" is misleadingly favorable to CS.\n\nI also believe that the paper needs experiments on ImageNet and needs a more thorough evaluation as a pruning technique beyond the lottery ticket hypothesis.\n\nI therefore retain my current score of \"weak reject,\" though I am eager to hear the thoughts of other reviewers, and I am open to changing my score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #4", "review": "\nTo the authors of paper 2504: I have posted a private comment for the reviewer/AC discussion period based on your author responses and your revised paper. I want you to be able to see my full response, but I can't post additional public comments to the paper. As such, I'm editing my review with exactly what I sent to the AC.\n\nThank you to the authors for your thoughtful rebuttal and for updating the paper with both new text and (more impressively) new experiments. I read all three of your rebuttals and re-read the revised paper in detail. My comment to the AC is at the bottom of this message. I don't know if you will get an alert that the review was updated, but I hope you get a chance to take a look.\n\n===========================================\n\nEXECUTIVE SUMMARY OF REVIEW\n\nSummary of paper: This paper proposes a technique that simultaneously trains a neural network and learns a pruning mask. The goal of this technique is to make it faster to retroactively find sparse subnetworks that, from a point early in training, could have trained in isolation to the same performance as the full network (\"winning tickets\"). The best subnetworks found by this technique outperform those found by existing techniques [1] at every sparsity level.\n\nSummary of review: The technique introduces several new hyperparameters whose values are asserted without describing the extent of the search necessary to find them; it unclear whether the cost of the search cancels out the efficiency gains. In addition, the proposed technique is inconsistent across runs. The sparsity and accuracy of the subnetworks it produces vary greatly, and there are no hyperparameters to explicitly control these outcomes. It is also unclear whether results vary from run to run even with the same hyperparameters.  As such, this technique is not clearly a cost reduction as compared to existing approaches in [1] when it comes to studying lottery tickets. Moreover, the paper implicitly proposes a new pruning technique, and it should be evaluated against other related techniques in the pruning literature. Finally, the evaluation needs more rigor.\n\nConclusion: It is difficult to determine whether the technique is an improvement over existing methods since the true costs of using it in practical workflows are unclear. Weak reject.\n\nOpportunity to improve score: Include a more detailed analysis of the overall costs of finding winning tickets at a target sparsity using the proposed technique, particularly including the costs of hyperparameter search necessary to do so. Include multiple replicates of experiments, experiments on more networks, and information about the variance of performance across runs with the same hyperparameters. Compare against other state-of-the-art pruning techniques.\n\nPROBLEM STATEMENT AND PROPOSED SOLUTION\n\nProblem: Winning tickets are currently expensive to find. The best known procedure [1, 2] is \"iterative magnitude pruning\" (IMP), which involves repeatedly training a network to completion, pruning by a fixed percentage, and \"rewinding\" weights to an early iteration of training until the network reaches the desired level of sparsity. To reach sufficient sparsity on standard networks, this procedure must be repeated 10 or more times.\n\nGoal: To propose a procedure that finds winning tickets more efficiently.\n\nSignificance: A more efficient procedure would make it easier to study the lottery ticket phenomenon. (Whether studying that phenomenon is, itself, significant is debatable.) Personally, I have extensive experience using IMP to find winning tickets, so techniques to reduce the cost of finding winning tickets would be very valuable for my work.\n\nProposed solution: The authors propose \"continuous sparsification\" (CS), which makes it possible to learn which weights to prune simultaneously with the weights themselves. Accompanying each parameter w in the network is a second parameter s. The actual weight used in the network is w * sigmoid(s * beta), where beta is a temperature hyperparameter. If the learned value of s is such that sigmoid(s * beta) is approximately 0 at the end of training, then the parameter has been pruned. The value of beta increases exponentially throughout training, meaning the output of the sigmoid will be closer to a hard 0 or 1, producing a pruning mask. To ensure sparsification happens, a regularization term lambda |sigmoid(beta * s)| is added. When run over multiple iterations, values of s are reset to their original values for weights that are not pruned. In addition, weights are either rewound (as in IMP) or left at their final values (as in [3]).\n\nNovelty: The paper is slightly novel. The technique is a variation of those proposed in [4] and [5], but the changes are meaningful. This is the first known use for finding winning tickets, but any pruning technique could hypothetically be used for this purpose. In effect, the paper proposes a new pruning technique that is primarily evaluated for its efficacy in finding winning tickets.\n\nTECHNICAL REVIEW\n\n* This technique introduces several new hyperparameters: lambda, initial and final betas, and the initial values for s. The paper suggests good values for these hyperparameters for both networks considered. These values were presumably found through hyperparameter search of some kind. How extensive was this hyperparameter search, and what is the range of \"good\" combinations of values? This is not just a methodological footnote; this paper's stated goal is to improve the efficiency of finding winning lottery tickets, and - if good hyperparameters are hard to find - then that defeats the purpose of a more efficient technique. IMP, while far less efficient epoch for epoch as compared to CS, requires no hyperparameter search; global pruning 20% of parameters per iteration seems to work well in general [2]. In a revised version of the paper, I would be eager to learn more about this set of tradeoffs, since that is what matters in practice.\n\n* The authors only study two networks: a toy convolutional network and a small Resnet. It is hard to draw broad conclusions from such a limited set of examples. I would be particularly interested in seeing how this technique performs on a large-scale network for ImageNet (e.g., Resnet-50), since these are the situations where IMP becomes particularly cost-prohibitive. If the technique works well in these settings, it would enable lottery ticket research at much larger scales than is currently possible. If the technique works as efficiently at this scale, then doing could even be feasible during the rebuttal period. (I acknowledge getting experiments working on ImageNet is no small undertaking in terms of both engineering time and cost, but it would improve my confidence to see those results.)\n\n* The authors did an admirably careful job replicating the networks in [1], which include a variety of nonstandard hyperparameters.\n\n* Did the paper study Resnet-18 (a network designed for ImageNet with 11.2M parameters) or Resnet-20 (a network designed for CIFAR-10 with 272K parameters)? Frankle et al. [1, 2] describe Resnet-20 in their appendices but mistakenly refer to it as Resnet-18 throughout both papers, so I wanted to clarify. Based on the final test accuracy of the network, it appears to be Resnet-20; if so, I'd urge you to call it as such and note in a parenthetical or footnote that it's the same network as in [1, 2] but with Frankle et al's mistaken name corrected.\n\n* Are the values of s0 for CS sampled from a distribution, or are they all the same, fixed value?\n\n* Do the extra parameters lead to longer wall-clock training times? If you are making an argument about a more efficient technique, this is an important consideration.\n\n* Figure 2 includes the same graph twice. I believe a different graph should appear on the left, and I am eager to take a look at it in a revised version of the paper.\n\n* It does not appear that multiple replicates of each experiment were run with different network initializations in Figure 3. There don't appear to be any error bars on Figure 3, suggesting that this only represents a single initialization. Considering the wide variance in performance achieved by continuous sparsification across runs in Figure 3 (right), this graph needs to include multiple runs and error bars. (I assume the multiple runs shown in Figure 3 (right) are with different hyperparameters?)\n\n* What is the performance of CS as a pruning technique? By finding winning tickets, CS is also implicitly pruning the network. Is this competitive with L0 Regularization [4]? Is it more efficient than iterative magnitude pruning as in [3]? If this is a more efficient technique for finding winning tickets, it is also liable to be a more efficient pruning method, which seems like even broader impact for the proposed technique. Alternatively, if this technique is less effective than comparable work (especially [4]) as a pruning method, then it is possible comparable work (especially [4]) might also produce better winning tickets, in which case the importance of this work is diminished. In an updated version of the paper, I would be interested in seeing an evaluation of CS as a pruning technique independent of the lottery ticket hypothesis (and compared to standard techniques in the pruning literature as such). I have a hard time seeing any reason why new techniques for finding lottery tickets are any different than new pruning techniques, and they should be evaluated as such in the context of the broader literature on pruning.\n\n* The last paragraph of Section 4.1 is missing important details that are necessary to evaluate the utility of continuous sparsification in comparison to IMP. \"In only two iterations, CS finds a ticket with over 77% sparsity...\" - for which hyperparameters, and how many hyperparameters had to be explored to find these values? How was the pareto curve obtained? How many different runs were necessary to create it? How many total epochs of training did it take to find that curve? If obtaining this pareto curve required many runs of CS, then it may not be any more efficient than running IMP in practice. The pareto curves appear to make CS look misleadingly effective, since they hide many of the actual costs involved (e.g., hyperparameter search, number of separate runs that were conducted to produce the curve, etc.). Greater transparency of this aspect of the paper would go a long way toward increasing my confidence that the findings are an improvement over IMP. \n\n* The sparsity and accuracy of subnetworks found by CS appears to vary widely from run to run as shown in Figure 3 (right). The final sparsity appears to be a function of the values of s0, lambda, and beta, potentially along with luck from the optimization process. For the same values of s, lambda, and beta, how widely do the final sparsities and accuracies vary? If I wanted to find a winning ticket with a particular sparsity using CS, what would the procedure look like for doing so? Would I have to sweep across values of these hyperparameters, or is there a more straightforward way to do so? The practical usefulness of the procedure hinges on these questions.\n\n* \"We associate the performance drop of highly sparse tickets found by our method from the second iteration onwards to the lack of weight rewinding.\" Why didn't you also try it with weight rewinding? That seems like an easy way to evaluate this hypothesis. For both networks, it would be interesting to see the performance of CS with and without rewinding (analogous to Appendix B in [1]).\n\n* In the literature so far, the only winning tickets to be examined are those from IMP [1, 2]. CS is a different technique, and it likely finds different winning tickets. Do these winning tickets have different properties from [1, 2]? Can they be reinitialized? Can they be rewound earlier? These comparisons seem like an interesting scientific opportunity.\n\n* There are a few additional comparisons that I think are vital to include in the paper to appropriately contextualize results.  They're in the bullets below.\n\n* First comparison: it would be useful to include random reinitialization or random pruning baselines as in [1, 2] simply to make it easier for the reader to contextualize the performance of other sparse subnetworks.\n\n* Second comparison: what happens if you run IMP such that each iteration prunes to the same sparsity as achieved by each iteration of CS? Perhaps pruning by a fixed amount per iteration in IMP is wasteful, and one can prune more aggressively during earlier iterations as CS naturally appears to do. In other words, one way of explaining the advantage of CS would be that it prunes more aggressively. Is this indeed the case? I would be very curious to know.\n\n* Third comparison: What are the range of results achieved if IMP and CS are run \"one-shot\" (pruning after just one iteration; in the case of IMP, pruning directly to a desired sparsity)? That is, how well can these techniques do with just a single iteration?\n\nWRITING\n\nThe writing is excellent. The prose is clear, and I was able to fully understand a relatively sophisticated technique on the first read through the paper. Writing of this quality is rare, and the authors should be commended for it.\n\nOVERALL: Weak Reject\n\nThe problem statement is that IMP is not efficient. The paper claims that CS is more efficient. However, the paper does not present a convincing case that CS is, on the whole, more efficient when taking into account hyperparameter search to get CS to work, hyperparameter search to target a particular sparsity, potential variance across runs of CS, potential additional training costs of CS, and the possibility that IMP might be able to work comparable well given a more aggressive pruning schedule.\n\nIn addition, the evaluation needs more experiments, including multiple replicates for each experiment and more networks (ideally one on ImageNet).\n\nFinally, I am unclear on what distinguishes CS from any other pruning technique, and it should be evaluated in the context of the broader pruning literature.\n\nIf the authors clarify that the overall cost of CS (including all of the factors listed above) is lower than for IMP, address technical concerns about the evaluation, and evaluate CS as a pruning technique in an updated version of the paper, I will update my score accordingly.\n\n[1] Frankle & Carbin. \"The Lottery Ticket Hypothesis.\" ICLR 2019.\n[2] Frankle et al. \"Stabilizing the Lottery Ticket Hypothesis.\" Arxiv.\n[3] Han et al. \"Learning both Weights and Connections for Efficient Neural Networks.\" NeurIPS 2015.\n[4] Louizos et al. \"Learning Sparse Neural Networks through L0 Regularization.\" ICLR 2018\n[5] Zhou et al. \"Deconstructing Lottery Tickets: Signs, Zeros, and the SuperMask.\" NeurIPS 2019.\n\n=====================================\n\nCOMMENT TO THE AC AFTER READING REBUTTALS AND REVISED PAPER\n\nTLDR\n\nI believe the technique is novel and should be published, regardless of whether it is actually more efficient than IMP for practical workflows.\n\nHowever, I believe the current evaluation is inadequate, both in the experiments and in the way data is presented. Namely, it is impossible to actually compare the costs of CS and IMP in the scenarios the authors evaluate, despite the fact that efficiency is the authors' claimed contribution. From the data as presented, it is unclear whether CS is actually more efficient than IMP for scientific use-cases. I'm not particularly concerned if there isn't an efficiency advantage - it's an innovative contribution regardless. My concern is that it is impossible to compare the costs of these techniques given the current presentation of data. That speaks to flaws in the evaluation section.\n\nI therefore maintain my score: weak reject. The technique deserves to be published, but the paper in its current form does not.\n\nINTRODUCTION\n\nAfter reading the original submission, I had the following questions:\n\n1) Right now, the only way to find winning lottery tickets is through training the network and pruning it. IMP can be instantiated with any pruning technique, and the authors are really proposing a new pruning technique for use in the IMP framework. How does continuous sparsification perform as a pruning technique for network compression (i.e., independent of the lottery ticket hypothesis)?\n\n2) How efficient is continuous sparsification (CS) in the scientific use cases where one would seek to find winning lottery tickets? In my research experience, there are two such use cases: (a) producing a winning lottery ticket with a specific sparsity and (b) producing winning lottery tickets across the full range of different sparsities.\n\nIn their reubttal, the authors addressed these underlying questions:\n\nAS A PRUNING TECHNIQUE\n\n\"We have added new experiments showing that our method yields competitive results when pruning VGG trained on CIFAR, outperforming both magnitude pruning and stochastic l0 regularization\"\n\nMy response: The initial results that the authors present are quite impressive on a VGG-style network for CIFAR-10. However, I find it concerning that the authors study pruning on this specific model but no others, particularly because they focus on a different network (namely, Resnet-20) in all other experiments in the paper. Since the authors are, in essence, proposing a new pruning technique, I would like to see it comprehensively evaluated as such on a range of networks against other baselines (as the authors do in Figure 4 for one network - those baselines are fine to me).\n\nEFFICIENCY OF CONTINUOUS SPARSIFICATION\n\n- It increases the cost of each individual network training run slightly: \"Continuous Sparsification on a 1080 Ti: our method resulted in 15% extra wall-clock time per training epoch.\"\n\n- PRODUCING WINNING TICKETS ACROSS THE FULL RANGE OF SPARSITIES: Throughout the paper, the authors present \"pareto curves\" showing the highest accuracy achieved by CS subnetworks at various sparsities. I find these pareto curves misleading: they hide the fact that CS had to be run several times with different hyperparameters to produce these curves.\n\nFor example, in Figure 6 (right), the authors produce the pareto curves by training Resnet-20 (by my count) 22 times (each point along the purple lines). The corresponding IMP curve appears to have 14 points. In other words, to get winning lottery tickets across all sparsities, CS must be run for more iterations than IMP - it appears to be less efficient, contradicting the authors' core claim. The best argument in favor of CS is that one could perform each of these runs in parallel if sufficient GPUs were available, meaning less wall-clock time would be required.\n\nOne caveat to this analysis: the authors state that they run CS \"without rewinding,\" meaning that the second iteration of CS requires less training time than fully training the network (as IMP requires). The authors do not state how long they train when they *aren't* rewinding, so it is impossible to compare the efficiency of CS with IMP. All they say is that it \"allow[s] for even faster ticket search.\" They also do not study IMP without rewinding, which would be a helpful baseline for comparing to CS without rewinding.\n\n***In short, if the authors are making an argument that one technique is more efficient than the other on an epoch-for-epoch basis, they need to actually plot the epochs required by each technique.***\n\n- PRODUCING A WINNING TICKET AT A SPECIFIC SPARSITY: My concern about this use case is that there is a non-intuitive relationship between the initialization of the sparsity parameters and the final sparsity of the network. As the authors state in the rebuttal: \"To achieve a desired sparsity, one can either perform runs in parallel with different values for s_0, or perform sequential binary search if the goal is to minimize the overall computational cost and not wall-clock time.\" In other words, there is no precise way to target a particular sparsity other than trying many hyperparameter configurations. The authors do not provide any concrete costs of using CS in this way in comparison to IMP, and - from a usability perspective - this is a challenging workflow.\n\nOTHER CONCERNS\n\n- The authors only study CS on one toy network (the six-layer convolutional network, which - in my experience - is a particularly easy setting compared to deeper networks) and one \"real\" network (Resnet-20 on CIFAR-10). I would like to see results on other networks for CIFAR-10 (for example, the VGG network in Section 4.3).\n\n- More importantly, I would like to see results on an ImageNet network. If CS makes finding winning lottery tickets more efficient as the authors claim, then finding winning tickets efficiently on an ImageNet network should be an excellent demonstration of their contribution. This scenario has stretched IMP to its breaking point, as the authors note.\n\nARGUMENTS IN FAVOR OF ACCEPTING\n\n+ The authors propose a new pruning technique that appears to improve upon the increasingly popular L0-regularization technique.\n\n+ With the right hyperparameters, the proposed technique makes it possible to find winning lottery tickets more efficiently than existing methods (i.e., IMP with magnitude pruning).\n\n+ The winning tickets found by the proposed technique reach higher accuracy than IMP winning tickets and produce winning tickets at more extreme sparsities, improving upon our knowledge of the existence of winning lottery tickets.\n\nARGUMENTS IN FAVOR OF REJECTING\n\n- The authors perform only minimal evaluation of their method as a pruning technique. It is possible that this is a missed opportunity to show an additional contribution. It is also possible that other, existing pruning techniques outperform CS at both pruning and finding winning lottery tickets.\n\n- The technique is hard to use for the two existing use cases for finding winning tickets. In particular, there is no way to search for winning tickets at a specific sparsity.\n\n- It is unclear whether CS is actually more efficient than IMP on an epoch-for-epoch basis, even though this is the main claimed contribution. The authors do not disclose - let alone plot - the number of training epochs required to find (a) winning tickets at a specific sparsity and (b) winning tickets at a range of sparsities, so it is impossible to make these comparisons. Meanwhile, the pareto curves the authors present are misleading, since they are the amalgamation of many separate runs of CS.\n\n- The authors study their technique on only one \"real\" network (Resnet-20) for finding winning tickets and a separate \"real\" network (VGG-19) for pruning. The authors do not show how CS performs in other challenging settings, especially ImageNet.\n\n- This work is only valuable if we believe that \"lottery ticket hypothesis\" work is valuable. In other words, this is a narrow contribution to an already-narrow area of study. This is one reason why pitching CS as a pruning technique would make this a stronger paper. I personally believe that \"lottery ticket\" work is valuable area of study, but I understand that it may not be seen as such in the broader ICLR community.\n\nCONCLUSION\n\nI believe the technique deserves to be published regardless of whether it is actually more efficient than IMP. It is a novel contribution to both our knowledge of pruning and of finding winning lottery tickets. The paper is exceptionally well written, and - as such - I believe it will help to inspire other research in this area.\n\nHowever, I do not believe that the current paper - namely, the current evaluation - should be published. The paper presents no concrete data on the comparative costs of performing CS and IMP even though the core claim is that CS is more efficient. The paper does not disclose enough detail to compute these costs, and it seems like CS is more expensive than IMP for standard workflows. Moreover, the current presentation of the data through \"pareto curves\" is misleadingly favorable to CS.\n\nI also believe that the paper needs experiments on ImageNet and needs a more thorough evaluation as a pruning technique beyond the lottery ticket hypothesis.\n\nI therefore retain my current score of \"weak reject,\" though I am eager to hear the thoughts of other reviewers, and I am open to changing my score.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1572745458399}, {"id": "SJlzXqz19r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2504/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work propose a new iterative pruning methods named Continuous Sparsification. It will continuously prune the current weight until it reaches the target ratio instead of iterative prune the weight to specific ratio. The author gives a good analysis but the experiment is not yet convincing enough.\n\n1) This work actually presents compression algorithm with little connection with lottery ticket. As a lottery ticket discussion, it does not give the comparison between lottery ticket and random initialization based on new pruning method. As a pruning method, it does not show the results on common models like VGG and DenseNet with different depth. Figure. 3 only gives results of ResNet-18 while the setting is not the best setting. Normally we need to train at least 120 epochs. It also does not give the experiment on ImageNet. Thus making the conclusion less meaningful\n\n2) The author should also compare the continuous sparsification with one-shot pruning methods (non-iterative) to see the advantage of continuous sprsification.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This work propose a new iterative pruning methods named Continuous Sparsification. It will continuously prune the current weight until it reaches the target ratio instead of iterative prune the weight to specific ratio. The author gives a good analysis but the experiment is not yet convincing enough.\n\n1) This work actually presents compression algorithm with little connection with lottery ticket. As a lottery ticket discussion, it does not give the comparison between lottery ticket and random initialization based on new pruning method. As a pruning method, it does not show the results on common models like VGG and DenseNet with different depth. Figure. 3 only gives results of ResNet-18 while the setting is not the best setting. Normally we need to train at least 120 epochs. It also does not give the experiment on ImageNet. Thus making the conclusion less meaningful\n\n2) The author should also compare the continuous sparsification with one-shot pruning methods (non-iterative) to see the advantage of continuous sprsification."}, "tcdate": 1571920410132}, {"id": "S1xheijnFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2504/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a novel objective function that can be used to jointly optimize a classification objective while at the same time encourage sparsification in a network. The lottery ticket hypothesis and associated work shows that the iterative pruning of a network can lead to a sparse network that performs with high accuracy. On the other hand, the work of Zhou et al. shows that sparse masks (dubbed \"supermasks\") may be learned without training the parameters of the network. In a sense, this paper tries to combine these ideas by simultaneously training a network while also optimizing the mask.\n\nI think this paper serves as a reasonable contribution to the ever-growing \"lottery ticket hypothesis\" body of work. The paper is mostly clear, and the idea for joint optimization is very reasonable. It's not tremendously original (in that it basically combines two ideas that are already in the literature), but in spite of that, I still think this paper warrants being accepted to ICLR.\n\nFor me, the most interesting scientific point is about the issue of rewinding. In particular, the fact that continuous sparsification can find winning tickets without any parameter rewinding is fascinating and deserves further investigation. Do the authors have any sense for why this works, when prior work suggests that rewinding is necessary for sufficiently complicated models and datasets?\n\nA minor point, I think there's a typo on page 6 in that the paragraph beginning \"Results are presented in Figure 2\" both instances of \"SP\" should be \"SS\" instead.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper proposes a novel objective function that can be used to jointly optimize a classification objective while at the same time encourage sparsification in a network. The lottery ticket hypothesis and associated work shows that the iterative pruning of a network can lead to a sparse network that performs with high accuracy. On the other hand, the work of Zhou et al. shows that sparse masks (dubbed \"supermasks\") may be learned without training the parameters of the network. In a sense, this paper tries to combine these ideas by simultaneously training a network while also optimizing the mask.\n\nI think this paper serves as a reasonable contribution to the ever-growing \"lottery ticket hypothesis\" body of work. The paper is mostly clear, and the idea for joint optimization is very reasonable. It's not tremendously original (in that it basically combines two ideas that are already in the literature), but in spite of that, I still think this paper warrants being accepted to ICLR.\n\nFor me, the most interesting scientific point is about the issue of rewinding. In particular, the fact that continuous sparsification can find winning tickets without any parameter rewinding is fascinating and deserves further investigation. Do the authors have any sense for why this works, when prior work suggests that rewinding is necessary for sufficiently complicated models and datasets?\n\nA minor point, I think there's a typo on page 6 in that the paragraph beginning \"Results are presented in Figure 2\" both instances of \"SP\" should be \"SS\" instead."}, "tcdate": 1571760883795}], "openreview_url": "https://openreview.net/forum?id=BJe4oxHYPB", "arxiv_id": "1912.04427", "paper_pdf": "papers/BJe4oxHYPB.pdf", "paper_pdf_sha256": "e17cb17cf63855e0d10013ad6db1e1153c2b6dd4cf5da813959a8879c120394f", "paper_pdf_bytes": 463296, "paper_pdf_source": "openreview", "code_url": "https://github.com/lolemacs/continuous-sparsification", "code_repository": "lolemacs/continuous-sparsification", "code_commit": "5bd4039f80724bbd71a78ba45d7efd35ea2e8e20", "code_archive": "repos/BJe4oxHYPB.zip", "code_archive_sha256": "f324c6e01e535b95a3021fa560129d4d296c44f4e0cc009dd42fce19b84f37fe", "code_archive_bytes": 9023, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 17, "github_languages": {"Python": 14108}, "github_archived": false, "github_pushed_at": "2022-06-10T10:39:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/winning-the-lottery-with-continuous-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryxDjjCqtQ", "year": 2019, "status": "rejected", "title": "Deconfounding Reinforcement Learning in Observational Settings", "authors": ["Chaochao Lu", "José Miguel Hernández Lobato"], "authorids": ["cl641@cam.ac.uk", "jmh233@cam.ac.uk"], "authors_source": "OpenReview API", "abstract": "In this paper, we propose a general formulation to cope with a family of reinforcement learning tasks in observational settings, that is, learning good policies solely from the historical data produced by real environments with confounders (i.e., the factors affecting both actions and rewards). Based on the proposed approach, we extend one representative of reinforcement learning algorithms: the Actor-Critic method, to its deconfounding variant, which is also straightforward to be applied to other algorithms. In addition, due to lack of datasets in this direction, a benchmark is developed for deconfounding reinforcement learning algorithms by revising OpenAI Gym and MNIST. We demonstrate that the proposed algorithms are superior to traditional reinforcement learning algorithms in confounded environments. To the best of our knowledge, this is the first time that confounders are taken into consideration for addressing full reinforcement learning problems.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "Byg9XiEs3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper628/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I have read the discussion from the authors. my evaluation stays the same.\n--------\nthis paper studies an interesting question of how to learn causal effects from observational data generated from reinforcement learning. they work with a very challenging setting where an unobserved confounder exists at each time step that affects actions, rewards and the confounder at next time step.\n\nthe authors fit latent variables models to the observational data and perform experiments.\n\nthe major concern is on the causal inference side, where it is not easy to claim anything causal in such a complicated system with unobserved confounders. causal inference with unobserved confounders cannot be simply solved by fitting a latent variable model. there exists negative examples even in the simplest setting that two distinct causal structure can lead to the same observational distribution. for example here, https://www.alexdamour.com/blog/public/2018/05/18/non-identification-in-latent-confounder-models/\n\nit could be helpful if the authors can lay out the identification assumptions for causal effects. before claiming anything causal and justifying experimental results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting problem", "review": "I have read the discussion from the authors. my evaluation stays the same.\n--------\nthis paper studies an interesting question of how to learn causal effects from observational data generated from reinforcement learning. they work with a very challenging setting where an unobserved confounder exists at each time step that affects actions, rewards and the confounder at next time step.\n\nthe authors fit latent variables models to the observational data and perform experiments.\n\nthe major concern is on the causal inference side, where it is not easy to claim anything causal in such a complicated system with unobserved confounders. causal inference with unobserved confounders cannot be simply solved by fitting a latent variable model. there exists negative examples even in the simplest setting that two distinct causal structure can lead to the same observational distribution. for example here, https://www.alexdamour.com/blog/public/2018/05/18/non-identification-in-latent-confounder-models/\n\nit could be helpful if the authors can lay out the identification assumptions for causal effects. before claiming anything causal and justifying experimental results.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541258018107}, {"id": "BJelvuEjnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper628/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper addresses an important and often overlooked issue in off-policy reinforcement learning - the possibility of confounding between the agent's actions and the rewards. This is a subject which has been exhaustively explored in the causal inference literature, and the authors are very correct in suggesting that it should be incorporated into the world of reinforcement learning.  Specifically they propose a generative model with a global latent confounder that is inferred using a variational autoencoder architecture.  \n\nThe paper is generally well-written, though some points could be made clearer in my opinion, as detailed below. The experiments are constructed by introducing confounding into existing datasets; performance seems to be good, but I am not entirely sure whether the given architecture is necessary, see comments below. \n\nHigh-level comments:\n(1) Classic RL deals with confounders all the time. The state is a confounder between the action and the reward. The issue of confounding becomes less trivial when one is performing off-policy RL when the original policy is *unknown*. This is exactly the case that the authors mention when they cite the recent work by Gottesman et al. (2018) who deal with using RL to learn from the actions of physicians in a hospital.  While I am sure the authors are aware of these distinctions, I think the paper would be better if this is spelled out very explicitly. This includes explaining why this issue doesn't come up in classic RL.\n\n(2) Assuming the case above - off-policy RL with unknown confounders - one would usually assume \"no unmeasured confounding\", i.e. that the observed actions are an unknown but learnable function of the observed states. That is basically the scenario of most off-policy RL.\n\n(3) However, the authors strive to go one step beyond the case (2), to a situation where there is an *unmeasured* confounder affecting both observed actions and rewards. If nothing is known about this unmeasured confounder, then it is generally impossible to learn effective policies, as the causal effects of actions are not identifiable from the observed data. In this paper, the authors make an implicit assumption that while the confounder is unmeasured, it can still be inferred from the data. This is an intermediate step between \"no unmeasured confounding\" and \"complete unmeasured confounding\". This is related to work on using proxy variables e.g. Kuroki & Pearl (2014) and even more closely related to the work cited by Louizos et al. (2017).\nAgain, I think the paper would be much improved if all this is addressed explicitly. \n\n(4) An important consequence of point (3) above is that in fact adding the single global latent-confounder U is not, in itself, very important from a causal perspective. The sequence of variables Z_1... Z_T are already latent confounders that are assumed to be inferrable from data. It is true that the addition of the global U might change the statistical and optimization properties of the model. This leads to a very important conclusion: the authors should test their model with and without U. I think this specific ablation experiment is crucial. In many cases I am sure that the assumption of a global latent confounder is a good one and is especially useful in the VAE case where it will make optimization more stable. However, in principle, all of U's roles could be taken within the sequence of Z's, and I am curious to see in practice how big of an effect it has.\n\n(5) I wish to add that even if the U variable turns out to not add much empirically, this work is still valid since the sequence of Z's can themselves be considered inferred latent confounders.\n\nSpecific comments:\n(1) 2.3: there are more than 2 ways of computing the do-operator. RCTs and backdoor are the best known approaches, but not the only ones, e.g. there is frontdoor adjustment. \n\n(2) I think the paper would be easier to follow if there was one concrete example used throughout. This will make it easier to understand and possibly verify/criticize the assumptions of the generative model.\n\n(3) Related to \"higher-level point (4)\" above, in eqs. 17 & 18 note that Z_t is unknown, same as U. Both are inferred. This also leads to the question which Z_t is actually used in practice? Is it the mean, or is it also sampled from the approximate posterior q?\n\n(4) Below eq. 19, it would be very useful for the readers if you could explain exactly when would there be a difference between the two versions p(r_{t+1}|z_t,a_t) and p(r_{t+1}|z_t, do(a_t=a))\n\n(5) In the description of all the experiments I was missing a crucial point: how does the introduced confounder affect the reward? Is it only through the different actions? The way it is currently explained, it seems like the added variable introduces lack of *overlap*, but not strictly confounding.\n\n(6) The description of the experiment in 4.3 could be more detailed. What exactly was the training and test? What RL method was used? What did the baseline optimize for? I would like to see an ablation experiment where U is not included in the model. \n\n(7) In 4.5, what is the \"vanilla\" method? And as mentioned above, I would like to see an ablation experiment where U is not included in the model.  \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Strong and important idea - presentation and execution can be improved ", "review": "The paper addresses an important and often overlooked issue in off-policy reinforcement learning - the possibility of confounding between the agent's actions and the rewards. This is a subject which has been exhaustively explored in the causal inference literature, and the authors are very correct in suggesting that it should be incorporated into the world of reinforcement learning.  Specifically they propose a generative model with a global latent confounder that is inferred using a variational autoencoder architecture.  \n\nThe paper is generally well-written, though some points could be made clearer in my opinion, as detailed below. The experiments are constructed by introducing confounding into existing datasets; performance seems to be good, but I am not entirely sure whether the given architecture is necessary, see comments below. \n\nHigh-level comments:\n(1) Classic RL deals with confounders all the time. The state is a confounder between the action and the reward. The issue of confounding becomes less trivial when one is performing off-policy RL when the original policy is *unknown*. This is exactly the case that the authors mention when they cite the recent work by Gottesman et al. (2018) who deal with using RL to learn from the actions of physicians in a hospital.  While I am sure the authors are aware of these distinctions, I think the paper would be better if this is spelled out very explicitly. This includes explaining why this issue doesn't come up in classic RL.\n\n(2) Assuming the case above - off-policy RL with unknown confounders - one would usually assume \"no unmeasured confounding\", i.e. that the observed actions are an unknown but learnable function of the observed states. That is basically the scenario of most off-policy RL.\n\n(3) However, the authors strive to go one step beyond the case (2), to a situation where there is an *unmeasured* confounder affecting both observed actions and rewards. If nothing is known about this unmeasured confounder, then it is generally impossible to learn effective policies, as the causal effects of actions are not identifiable from the observed data. In this paper, the authors make an implicit assumption that while the confounder is unmeasured, it can still be inferred from the data. This is an intermediate step between \"no unmeasured confounding\" and \"complete unmeasured confounding\". This is related to work on using proxy variables e.g. Kuroki & Pearl (2014) and even more closely related to the work cited by Louizos et al. (2017).\nAgain, I think the paper would be much improved if all this is addressed explicitly. \n\n(4) An important consequence of point (3) above is that in fact adding the single global latent-confounder U is not, in itself, very important from a causal perspective. The sequence of variables Z_1... Z_T are already latent confounders that are assumed to be inferrable from data. It is true that the addition of the global U might change the statistical and optimization properties of the model. This leads to a very important conclusion: the authors should test their model with and without U. I think this specific ablation experiment is crucial. In many cases I am sure that the assumption of a global latent confounder is a good one and is especially useful in the VAE case where it will make optimization more stable. However, in principle, all of U's roles could be taken within the sequence of Z's, and I am curious to see in practice how big of an effect it has.\n\n(5) I wish to add that even if the U variable turns out to not add much empirically, this work is still valid since the sequence of Z's can themselves be considered inferred latent confounders.\n\nSpecific comments:\n(1) 2.3: there are more than 2 ways of computing the do-operator. RCTs and backdoor are the best known approaches, but not the only ones, e.g. there is frontdoor adjustment. \n\n(2) I think the paper would be easier to follow if there was one concrete example used throughout. This will make it easier to understand and possibly verify/criticize the assumptions of the generative model.\n\n(3) Related to \"higher-level point (4)\" above, in eqs. 17 & 18 note that Z_t is unknown, same as U. Both are inferred. This also leads to the question which Z_t is actually used in practice? Is it the mean, or is it also sampled from the approximate posterior q?\n\n(4) Below eq. 19, it would be very useful for the readers if you could explain exactly when would there be a difference between the two versions p(r_{t+1}|z_t,a_t) and p(r_{t+1}|z_t, do(a_t=a))\n\n(5) In the description of all the experiments I was missing a crucial point: how does the introduced confounder affect the reward? Is it only through the different actions? The way it is currently explained, it seems like the added variable introduces lack of *overlap*, but not strictly confounding.\n\n(6) The description of the experiment in 4.3 could be more detailed. What exactly was the training and test? What RL method was used? What did the baseline optimize for? I would like to see an ablation experiment where U is not included in the model. \n\n(7) In 4.5, what is the \"vanilla\" method? And as mentioned above, I would like to see an ablation experiment where U is not included in the model.  \n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541257303957}, {"id": "Hyx9C2Vqhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper628/AnonReviewer1"], "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a method for reinforcement learning (RL) in settings where the relationship between action and reward is confounded by a latent variable (unobserved confounder). While I firmly believe that RL would benefit from taking causality more seriously, this paper has many fatal flaws that make it not ready for publication. \n\nFirst, and most importantly, the paper is unclear about the problem it is trying to solve. It talks about confounded RL as being settings in which a confounder affects both the action and reward. In typical RL settings this wouldn’t make sense: in RL you get to choose the policy so it doesn’t make sense to assume that the choice of action is confounded while you’re doing RL. To get around this, the authors assume that they’re working with observational data and doing RL on a generative model leant from the observational data. But by doing this, they have assumed away the key advantage that RL has over causal inference: the ability to experiment in the world. The authors justify this assumption by considering high-stakes settings where experimentation is either too risky or too costly, but they don’t explain why you would want to do RL at all when you could just do causal inference directly. If you can’t experiment, RL offers no advantages over standard causal inference methods and bring serious disadvantages (sample-efficiency, computational cost, etc.). \n\n# Method\nThe authors learn a variational approximation to a particular graphical model that they assume for their RL setting. They then treat the variational approximation as the true distribution which allows them to perform causal inference via the backdoor correction. They claim this is identified but this is false - it is only identified with respect to the variational distribution, not the true distribution and we have no a priori reason to  believe that the variational distribution well-approximated the true distribution. In principle, the authors could have tested how well this works experimentally but their experimental setup has problems which prevent this being evaluated. \n\nQuibbles:\n - Page 3: the authors claim the model is “without loss of generality” but this is false - there are many settings that would not conform to this model: e.g. the multi agent settings that economics studies; health settings with placebo effects where reward depends on observations directly; etc.\n  - Page 4 above the equations: either the equations describe the variational approximation to the generative model or the equations shouldn’t all be factorized normal distributions. Real data isn’t made up of factorized normals.\n\n# Experiments\n\nThe authors evaluate their method on three simulated datasets: Confounding MNIST, Confounding Cartpole and Confounding Pendulum. All three have the same methodological problems so I’ll only focus on the MNIST dataset. They synthesize their MNIST dataset but corrupting a subset of MNIST digits with noise and treating actions as rotations. Rewards are given by the absolute difference in angle between the rotated digit and the original unrotated digit. “Confounding” is added by having a binary latent variable affect the amount that the digit is rotated - but importantly, the reward isn’t affected directly by the latent variable. Because of this, there isn’t actually a confounding problem - the “confounder” simply changes the rotation of the digit and can be treated as additional experimentation from the perspective of causal inference. The authors evaluate their method by examining reconstructions of the MNIST digit, but this simply checks how well the variational inference is working, not whether the causal inference is working (there would be no way to evaluate the latter on this dataset because there is no confounding). Effectively all they find is a better-designed variational distribution will do a better job of reconstructing the input (without modelling the latent u, the VAE is forced to average over its two states resulting in more blurry samples). \n\nThe RL evaluations aren’t described in enough detail to conclusively explain the difference observed, but it seems to be driven by the fact that the standard RL methods are working with worse variational approximation distributions.\n\n# Summary\nThis work studies a setting in which the correct baselines would be causal inference algorithms (but they aren’t considered) and their experimental evaluation has serious flaws that prevent it supporting the claims made in the paper. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Setting doesn't make sense for RL and experiments don't evaluate causal questions", "review": "This paper presents a method for reinforcement learning (RL) in settings where the relationship between action and reward is confounded by a latent variable (unobserved confounder). While I firmly believe that RL would benefit from taking causality more seriously, this paper has many fatal flaws that make it not ready for publication. \n\nFirst, and most importantly, the paper is unclear about the problem it is trying to solve. It talks about confounded RL as being settings in which a confounder affects both the action and reward. In typical RL settings this wouldn’t make sense: in RL you get to choose the policy so it doesn’t make sense to assume that the choice of action is confounded while you’re doing RL. To get around this, the authors assume that they’re working with observational data and doing RL on a generative model leant from the observational data. But by doing this, they have assumed away the key advantage that RL has over causal inference: the ability to experiment in the world. The authors justify this assumption by considering high-stakes settings where experimentation is either too risky or too costly, but they don’t explain why you would want to do RL at all when you could just do causal inference directly. If you can’t experiment, RL offers no advantages over standard causal inference methods and bring serious disadvantages (sample-efficiency, computational cost, etc.). \n\n# Method\nThe authors learn a variational approximation to a particular graphical model that they assume for their RL setting. They then treat the variational approximation as the true distribution which allows them to perform causal inference via the backdoor correction. They claim this is identified but this is false - it is only identified with respect to the variational distribution, not the true distribution and we have no a priori reason to  believe that the variational distribution well-approximated the true distribution. In principle, the authors could have tested how well this works experimentally but their experimental setup has problems which prevent this being evaluated. \n\nQuibbles:\n - Page 3: the authors claim the model is “without loss of generality” but this is false - there are many settings that would not conform to this model: e.g. the multi agent settings that economics studies; health settings with placebo effects where reward depends on observations directly; etc.\n  - Page 4 above the equations: either the equations describe the variational approximation to the generative model or the equations shouldn’t all be factorized normal distributions. Real data isn’t made up of factorized normals.\n\n# Experiments\n\nThe authors evaluate their method on three simulated datasets: Confounding MNIST, Confounding Cartpole and Confounding Pendulum. All three have the same methodological problems so I’ll only focus on the MNIST dataset. They synthesize their MNIST dataset but corrupting a subset of MNIST digits with noise and treating actions as rotations. Rewards are given by the absolute difference in angle between the rotated digit and the original unrotated digit. “Confounding” is added by having a binary latent variable affect the amount that the digit is rotated - but importantly, the reward isn’t affected directly by the latent variable. Because of this, there isn’t actually a confounding problem - the “confounder” simply changes the rotation of the digit and can be treated as additional experimentation from the perspective of causal inference. The authors evaluate their method by examining reconstructions of the MNIST digit, but this simply checks how well the variational inference is working, not whether the causal inference is working (there would be no way to evaluate the latter on this dataset because there is no confounding). Effectively all they find is a better-designed variational distribution will do a better job of reconstructing the input (without modelling the latent u, the VAE is forced to average over its two states resulting in more blurry samples). \n\nThe RL evaluations aren’t described in enough detail to conclusively explain the difference observed, but it seems to be driven by the fact that the standard RL methods are working with worse variational approximation distributions.\n\n# Summary\nThis work studies a setting in which the correct baselines would be causal inference algorithms (but they aren’t considered) and their experimental evaluation has serious flaws that prevent it supporting the claims made in the paper. \n", "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541192913551}], "openreview_url": "https://openreview.net/forum?id=ryxDjjCqtQ", "arxiv_id": "1812.10576", "paper_pdf": "papers/ryxDjjCqtQ.pdf", "paper_pdf_sha256": "69d8396b824ab8094caee9a975aadbd99b50a68991b0c7267f3b6c845f37a7fa", "paper_pdf_bytes": 1784580, "paper_pdf_source": "openreview", "code_url": "https://github.com/CausalRL/DRL", "code_repository": "CausalRL/DRL", "code_commit": "0cd76bcbad8189ff3eb8ee750aaf6e992ce291a9", "code_archive": "repos/ryxDjjCqtQ.zip", "code_archive_sha256": "ee1b041e506020300693f36eec1b45690f4a282a573cd0953c870e4ca18aad87", "code_archive_bytes": 45636, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 24, "github_languages": {"Python": 227459}, "github_archived": false, "github_pushed_at": "2019-04-13T22:45:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deconfounding-reinforcement-learning-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1bgpzZAZ", "year": 2018, "status": "rejected", "title": "ElimiNet: A Model for Eliminating Options for Reading Comprehension with Multiple Choice Questions", "authors": ["Soham Parikh", "Ananya Sai", "Preksha Nema", "Mitesh M Khapra"], "authorids": ["sohamp@cse.iitm.ac.in", "ananyasb@cse.iitm.ac.in", "preksha@cse.iitm.ac.in", "miteshk@cse.iitm.ac.in"], "authors_source": "OpenReview API", "abstract": "The task of Reading Comprehension with Multiple Choice Questions, requires a human (or machine) to read a given \\{\\textit{passage, question}\\} pair and select one of the $n$ given options. The current state of the art model for this task first computes a query-aware representation for the passage and then \\textit{selects} the option which has the maximum similarity with this representation. However, when humans perform this task they do not just focus on option selection but use a combination of \\textit{elimination} and \\textit{selection}. Specifically, a human would first try to eliminate the most irrelevant option and then read the document again in the light of this new information (and perhaps ignore portions corresponding to the eliminated option). This process could be repeated multiple times till the reader is finally ready to select the correct option. We propose \\textit{ElimiNet}, a neural network based model which tries to mimic this process. Specifically, it has gates which decide whether an option can be eliminated given the \\{\\textit{document, question}\\} pair and if so it tries to make the document representation orthogonal to this eliminatedd option (akin to ignoring portions of the document corresponding to the eliminated option). The model makes multiple rounds of partial elimination to refine the document representation and finally uses a selection module to pick the best option. We evaluate our model on the recently released large scale RACE dataset and show that it outperforms the current state of the art model on 7 out of the 13 question types in this dataset. Further we show that taking an ensemble of our \\textit{elimination-selection} based method with a \\textit{selection} based method gives us an improvement of 7\\% (relative) over the best reported performance on this dataset.    \n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HkHGUsPef", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper990/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, a model is built for reading comprehension with multiple choices. The model consists of three modules: encoder, interaction module and elimination module. The major contributions are two folds: firstly, proposing the interesting option elimination problem for multi-step reading comprehension;  and secondly, proposing the elimination module where a eliminate gate is used to select different orthogonal factors from the document representations. Intuitively, one answer option can be viewed as eliminated if the document representation vector has its factor along the option vector ignored.\n\nThe elimination module is interesting, but the usefulness of “elimination” is not well justified for two reasons. First, the improvement of the proposed model over the previous state of the art is limited. Second, the model is built upon GAR until the elimination module, then according to Table 1 it seems to indicate that the elimination module does not help significantly (0.4% improvement). \n\nIn order to show the usefulness of the elimination module, the model should be exactly built on the GAR with an additional elimination module (i.e. after removing the elimination module, the performance should be similar to GAR but not something significantly worse with a 42.58% accuracy). Then we can explicitly compare the performance between GAR and the GAR w/ elimination module to tell how much the new module helps.\n\nOther issues:\n\n1) Is there any difference to directly use $x$ and $h^z$ instead of $x^e$ and $x^r$ to compute $\\tilde{x}_i$? Even though the authors find the orthogonal vectors, they’re gated summed together very soon. It would be better to show how much “elimination” and “subtraction” effect the final performance, besides the effect of subtraction gate.\n\n2) A figure showing the model architecture and the corresponding QA process will better help the readers understand the proposed model.\n\n3) $c_i$ in page 5 is not defined. What’s the performance of only using $s_i$ for answer selection or replacing $x^L$ with $s_i$ in score function?\n\n4) It would be better to have the experiments trained with different $n$ to show how multi-hop effects the final performance, besides the case study in Figure 3.\n\nMinor issues:\n\n1) In Eqn. (4), it would be better to use a vector as the input of softmax.\n\n2) It would be easier for discussion if the authors could assign numbers to every equation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting problem but the results are not so good, hope the methods could be further improved in the future.", "rating": "5: Marginally below acceptance threshold", "review": "In this paper, a model is built for reading comprehension with multiple choices. The model consists of three modules: encoder, interaction module and elimination module. The major contributions are two folds: firstly, proposing the interesting option elimination problem for multi-step reading comprehension;  and secondly, proposing the elimination module where a eliminate gate is used to select different orthogonal factors from the document representations. Intuitively, one answer option can be viewed as eliminated if the document representation vector has its factor along the option vector ignored.\n\nThe elimination module is interesting, but the usefulness of “elimination” is not well justified for two reasons. First, the improvement of the proposed model over the previous state of the art is limited. Second, the model is built upon GAR until the elimination module, then according to Table 1 it seems to indicate that the elimination module does not help significantly (0.4% improvement). \n\nIn order to show the usefulness of the elimination module, the model should be exactly built on the GAR with an additional elimination module (i.e. after removing the elimination module, the performance should be similar to GAR but not something significantly worse with a 42.58% accuracy). Then we can explicitly compare the performance between GAR and the GAR w/ elimination module to tell how much the new module helps.\n\nOther issues:\n\n1) Is there any difference to directly use $x$ and $h^z$ instead of $x^e$ and $x^r$ to compute $\\tilde{x}_i$? Even though the authors find the orthogonal vectors, they’re gated summed together very soon. It would be better to show how much “elimination” and “subtraction” effect the final performance, besides the effect of subtraction gate.\n\n2) A figure showing the model architecture and the corresponding QA process will better help the readers understand the proposed model.\n\n3) $c_i$ in page 5 is not defined. What’s the performance of only using $s_i$ for answer selection or replacing $x^L$ with $s_i$ in score function?\n\n4) It would be better to have the experiments trained with different $n$ to show how multi-hop effects the final performance, besides the case study in Figure 3.\n\nMinor issues:\n\n1) In Eqn. (4), it would be better to use a vector as the input of softmax.\n\n2) It would be easier for discussion if the authors could assign numbers to every equation.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511663116782}, {"id": "SyfPjhYef", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper990/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper gives an elaboration on the Gated Attention Reader (GAR) adding gates based on answer elimination in multiple choice reading comprehension.  I found the formal presentation of the model reasonably clear the the empirical evaluation reasonably compelling.\n\nIn my opinion the main weakness of the paper is the focus on the RACE dataset.  This dataset has not attracted much attention and most work in reading comprehension has now moved to the SQUAD dataset for which there is an active leader board.  I realize that SQUAD is not explicitly multiple choice and that this is a challenge for an answer elimination architecture.  However, it seems that answer elimination might be applied to each choice of the initial position of a possible answer span.  In any case, competing with an active leader board would be much more compelling.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Competent elaboration of the Gated Attention Reader", "rating": "5: Marginally below acceptance threshold", "review": "This paper gives an elaboration on the Gated Attention Reader (GAR) adding gates based on answer elimination in multiple choice reading comprehension.  I found the formal presentation of the model reasonably clear the the empirical evaluation reasonably compelling.\n\nIn my opinion the main weakness of the paper is the focus on the RACE dataset.  This dataset has not attracted much attention and most work in reading comprehension has now moved to the SQUAD dataset for which there is an active leader board.  I realize that SQUAD is not explicitly multiple choice and that this is a challenge for an answer elimination architecture.  However, it seems that answer elimination might be applied to each choice of the initial position of a possible answer span.  In any case, competing with an active leader board would be much more compelling.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511799642337}, {"id": "HJViVF5gf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper990/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new reading comprehension model for multi-choice questions and the main motivation is that some options should be eliminated first to infer better passage/question representations.\n\nIt is a well-written paper, however, I am not very convinced by its motivation, the proposed model and the experimental results. \n\nFirst of all, the improvement is rather limited. It is only 0.4 improvement overall on the RACE dataset; although it outperforms GAR on 7 out of 13 categories; but why is it worse on the other 6 categories? I don’t see any convincing explanations here.\n\nSecondly, in terms of the development of reading comprehension models, I don’t see why we need to care about eliminating the irrelevant options. It is hard to generalize to any other RC/QA tasks. If the point is that the options can add useful information to induce better representations for passage/question, there should be some simple baselines in the middle that this paper should compare to. The two baselines SAR and GAR both only induce a representation from paragraph/question, and finally compare to the representation of each option. Maybe a simple baseline is to merge the question and all the options and see if a better document representation can be defined. \n\nSome visualizations/motivational examples could be also useful to understand how some options are eliminated and how the document representation has been changed based on that.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official review", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a new reading comprehension model for multi-choice questions and the main motivation is that some options should be eliminated first to infer better passage/question representations.\n\nIt is a well-written paper, however, I am not very convinced by its motivation, the proposed model and the experimental results. \n\nFirst of all, the improvement is rather limited. It is only 0.4 improvement overall on the RACE dataset; although it outperforms GAR on 7 out of 13 categories; but why is it worse on the other 6 categories? I don’t see any convincing explanations here.\n\nSecondly, in terms of the development of reading comprehension models, I don’t see why we need to care about eliminating the irrelevant options. It is hard to generalize to any other RC/QA tasks. If the point is that the options can add useful information to induce better representations for passage/question, there should be some simple baselines in the middle that this paper should compare to. The two baselines SAR and GAR both only induce a representation from paragraph/question, and finally compare to the representation of each option. Maybe a simple baseline is to merge the question and all the options and see if a better document representation can be defined. \n\nSome visualizations/motivational examples could be also useful to understand how some options are eliminated and how the document representation has been changed based on that.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511851163677}], "openreview_url": "https://openreview.net/forum?id=B1bgpzZAZ", "arxiv_id": "1904.02651", "paper_pdf": "papers/B1bgpzZAZ.pdf", "paper_pdf_sha256": "b48c4c1f871aa6f700303934a8668963d33fdcfca7e1e902e04859bbb5f48ab8", "paper_pdf_bytes": 508977, "paper_pdf_source": "openreview", "code_url": "https://github.com/sohamparikh/ElimiNet", "code_repository": "sohamparikh/ElimiNet", "code_commit": "f2a9f293fa2ce5470522444c8cfa6523c848392a", "code_archive": "repos/B1bgpzZAZ.zip", "code_archive_sha256": "082dd1559dda2c785c6dafa812d121e07d489569010cc504be6e8c30ac400a0a", "code_archive_bytes": 15164, "code_file_count": 8, "code_extensions": {".py": 5, ".sh": 3}, "github_disk_usage_kb": 17, "github_languages": {"Python": 56689, "Shell": 2332}, "github_archived": false, "github_pushed_at": "2019-04-11T21:22:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/eliminet-a-model-for-eliminating-options-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JeiaHDawhb", "year": 2025, "status": "rejected", "title": "Maximum Total Correlation Reinforcement Learning", "authors": ["Bang You", "Puze Liu", "Huaping Liu", "Jan Peters", "Oleg Arenz"], "authorids": ["~Bang_You1", "~Puze_Liu1", "~Huaping_Liu3", "~Jan_Peters3", "~Oleg_Arenz1"], "authors_source": "OpenReview API", "abstract": "Simplicity is a powerful inductive bias. In reinforcement learning, regularization is used for simpler policies, data augmentation for simpler representations, and sparse reward functions for simpler objectives, all that, with the underlying motivation to increase generalizability and robustness by focusing on the essentials. Supplementary to these techniques, we investigate how to promote simple behavior throughout the duration of the episode. To that end, we introduce a modification of the reinforcement learning problem, that additionally maximizes the total correlation within the induced trajectories. We propose a practical algorithm that optimizes all models, including policy and state representation, based on a lower bound approximation. In simulated robot locomotion environments, our method naturally generates policies that induce periodic and compressible trajectories, and that exhibit superior robustness to noise and changes in dynamics compared to baseline methods, while also improving performance in the original tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "xkPMJV8krg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6703/Reviewer_gUZZ"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "The paper proposes an auxiliary RL objective, MTC, that maximizes the total correlation of states and actions under the policy. This objective encourages the policy to learn simple, consistent, and compressible behaviors that are robust to noise and dynamics shifts. Specifically, the authors assume a latent-state-conditioned policy and incentivize the policy to maximize the total correlation of latent states and actions: $\\mathcal C(z_1; a_1; \\dots; a_{T-1}; z_T)$. This objective is intractable, so they estimate a variational lowerbound $\\mathbb E_{\\pi, f} \\left[ \\sum_{t=1}^{T-1} \\left[\\log \\frac{q_{\\eta}(z_{t+1}|z_{1:t}, a_{1:t})}{f_{\\theta}(z_{t+1}|s_{t+1})} + \\log\\frac{q_\\chi(a_t|z_{1:t}, a_{1:t-1})}{\\pi_\\phi(a_t|s_t)}\\right]\\right]$ using a dynamics model $q_\\eta$, an action prediction model $q_\\chi$, and an encoder $f_\\theta$. The estimated lower bound is added to the reward at each timestep, with its weights dynamically adjusted via constrained optimization. Experiments show that MTC combined with SAC outperforms baselines on a suite of DM Control tasks. They further show that MTC is more robust than baselines in the presence of state and action noises and dynamics shifts.", "review_text": "The paper proposes an auxiliary RL objective, MTC, that maximizes the total correlation of states and actions under the policy. This objective encourages the policy to learn simple, consistent, and compressible behaviors that are robust to noise and dynamics shifts. Specifically, the authors assume a latent-state-conditioned policy and incentivize the policy to maximize the total correlation of latent states and actions: $\\mathcal C(z_1; a_1; \\dots; a_{T-1}; z_T)$. This objective is intractable, so they estimate a variational lowerbound $\\mathbb E_{\\pi, f} \\left[ \\sum_{t=1}^{T-1} \\left[\\log \\frac{q_{\\eta}(z_{t+1}|z_{1:t}, a_{1:t})}{f_{\\theta}(z_{t+1}|s_{t+1})} + \\log\\frac{q_\\chi(a_t|z_{1:t}, a_{1:t-1})}{\\pi_\\phi(a_t|s_t)}\\right]\\right]$ using a dynamics model $q_\\eta$, an action prediction model $q_\\chi$, and an encoder $f_\\theta$. The estimated lower bound is added to the reward at each timestep, with its weights dynamically adjusted via constrained optimization. Experiments show that MTC combined with SAC outperforms baselines on a suite of DM Control tasks. They further show that MTC is more robust than baselines in the presence of state and action noises and dynamics shifts.", "strengths": "1. Maximizing the total correlation is a reasonable objective to induce simplicity in RL policies. The total correlation measures the KL divergence between the independent state and action distributions and their joint distribution. The higher the total correlation, the less it costs to compress the entire trajectory compared to compressing the states and actions independently.\n2. The authors propose a variational lower bound of the total correlation. The mathematical derivations are correct and easy to follow.\n3. The ablation experiment in Section 5.4 and Figure 4 showing improved generalization with a higher target total correlation value lends credibility to MTC's effectiveness. \n\n**List of Claims**\n\nClaim 1: the MTC objective improves RL agents' robustness to noise and distribution shifts.\n- Assumption: simplicity leads to robustness.\n- Evidence: ablation in Section 5.4 and Figure 4.\n- Statement: this ablation experiment shows that as they increase the target total correlation (the target constraint value in the constrained optimization problem), the method becomes more robust to observation and action noises and friction changes.", "weaknesses": "1. The fundamental motivation of the paper, \"simplicity is a powerful inductive bias\", is not well supported. The authors support it by enumerating examples in related problems, e.g. \"regularization is used for simpler policies, data augmentation for simpler representations, and sparse reward functions for simpler objectives.\" But this is quite vague. It might very well be that in RL, overoptimizing for simplicity leads to suboptimality in complex tasks. \n2. In Section 4.1, the authors motivate the use of latent states by stating that it is more robust to distractions from task-irrelevant factors (e.g. noisy TV). This claim again lacks support.\n3. The key differences between MTC and RPC [1] are (1) MTC considers states and actions while RPC only considers states, and (2) MTC conditions the dynamics and action prediction models on the history, whereas RPC only conditions on the current state. I'm not sure why (1) would conceptually improve upon RPC (see question below). (2) introduces a confounding variable that is orthogonal to the claim, since we can replace the model in RPC with a recurrent model and hope to see better performance. \n4. While the goal is to induce simplicity in the policy's behavior, the method requires learning three additional models to estimate an auxiliary objective, which is quite cumbersome. \n5. The rest of experiments are unconvincing. While the main results (Table 1) show MTC performing on par or slightly better than baselines, the improvement is not significant. Moreover, these results are in a narrow domain that prefers repetitive motion (e.g. Hopper, Cheetah, Walker are all locomotion tasks). I'm not sure if the same holds for more complex tasks. The robustness results in Figure 2 don't imply MTC being *more* robust than baselines, because SAC shows the same growth trend, just starting from a lower initial value. What I hope to see is that MTC performance degrades *at a slower rate than* SAC, instead of having a constant offset. \n\nOverall, I think this paper can be significantly strengthened by (1) grounding it in a concrete problem setting like sim2real transfer instead of a lofty desideratum (i.e. simplicity), and (2) either removing the unwarranted claims or adding experiments to support them. I strongly recommend the authors to revise and resubmit.\n\n**List of Claims (cont'd)**\n\nClaim 2: simplicity is desirable for RL.\n- Assumption: none\n- Evidence: slightly better empirical performance of MTC over baselines on DM control tasks.\n- Statement: I don't think there is enough theoretical or empirical evidence to support this claim. Given that the DM control tasks prefer periodic behavior, it might be that simplicity is particularly useful in this domain. \n\nClaim 3: using latent states results in less distraction from task-irrelevant factors.\n- Assumption: the environment consists of task-irrelevant distractors, modeling which hinder the policy's performance.\n- Evidence: none\n- Statement: to support this claim, the authors need to compare their method to a baseline that uses raw observations on environments with distractors. \n\nClaim 4: MTC improves upon RPC by considering states and actions, and conditioning the models on the history.\n- Assumption: none\n- Evidence: slightly better empirical performance than RPC.\n- Statement: I'm not sure \"why\" these changes would improve upon RPC. The sequence of states should determine the sequence of actions to a large extent. Conditioning on the history is a choice of variational approximation that can also be applied to RPC, and the paper doesn't compare to RPC + recurrent models. \n\n[1] Benjamin Eysenbach, Ruslan Salakhutdinov, Sergey Levine. Robust Predictable Control. NeurIPS 2021.", "questions": "1. Why is simplicity a good inductive bias for RL? Does this inductive bias hold outside of locomotion domains where repetitive motion is preferred?\n2. Why is accounting for the states and actions better than only accounting for states as in RPC? Isn't the state trajectory a consequence of the action trajectory? \n3. In Equation 2, what does it mean to take expectation under $f$?\n4. Does the infinite horizon variant correspond to a meaningful information-theoretic objective?\n5. Can you provide more experiments to support your claims?\n    - a manipulation task where the motion is less periodic, to support the claim that simplicity is a universally desirable property.\n    - visual experiments with distractors (e.g. distracted DM control) to support the claim that latent-state policies are robust to distractors.\n    - visual robustness experiments with lighting / texture change similar to the sim2real setting, to support the claim that MTC is robust to distribution shifts.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an auxiliary RL objective, MTC, that maximizes the total correlation of states and actions under the policy. This objective encourages the policy to learn simple, consistent, and compressible behaviors that are robust to noise and dynamics shifts. Specifically, the authors assume a latent-state-conditioned policy and incentivize the policy to maximize the total correlation of latent states and actions: $\\mathcal C(z_1; a_1; \\dots; a_{T-1}; z_T)$. This objective is intractable, so they estimate a variational lowerbound $\\mathbb E_{\\pi, f} \\left[ \\sum_{t=1}^{T-1} \\left[\\log \\frac{q_{\\eta}(z_{t+1}|z_{1:t}, a_{1:t})}{f_{\\theta}(z_{t+1}|s_{t+1})} + \\log\\frac{q_\\chi(a_t|z_{1:t}, a_{1:t-1})}{\\pi_\\phi(a_t|s_t)}\\right]\\right]$ using a dynamics model $q_\\eta$, an action prediction model $q_\\chi$, and an encoder $f_\\theta$. The estimated lower bound is added to the reward at each timestep, with its weights dynamically adjusted via constrained optimization. Experiments show that MTC combined with SAC outperforms baselines on a suite of DM Control tasks. They further show that MTC is more robust than baselines in the presence of state and action noises and dynamics shifts.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "1. Maximizing the total correlation is a reasonable objective to induce simplicity in RL policies. The total correlation measures the KL divergence between the independent state and action distributions and their joint distribution. The higher the total correlation, the less it costs to compress the entire trajectory compared to compressing the states and actions independently.\n2. The authors propose a variational lower bound of the total correlation. The mathematical derivations are correct and easy to follow.\n3. The ablation experiment in Section 5.4 and Figure 4 showing improved generalization with a higher target total correlation value lends credibility to MTC's effectiveness. \n\n**List of Claims**\n\nClaim 1: the MTC objective improves RL agents' robustness to noise and distribution shifts.\n- Assumption: simplicity leads to robustness.\n- Evidence: ablation in Section 5.4 and Figure 4.\n- Statement: this ablation experiment shows that as they increase the target total correlation (the target constraint value in the constrained optimization problem), the method becomes more robust to observation and action noises and friction changes.", "weaknesses": "1. The fundamental motivation of the paper, \"simplicity is a powerful inductive bias\", is not well supported. The authors support it by enumerating examples in related problems, e.g. \"regularization is used for simpler policies, data augmentation for simpler representations, and sparse reward functions for simpler objectives.\" But this is quite vague. It might very well be that in RL, overoptimizing for simplicity leads to suboptimality in complex tasks. \n2. In Section 4.1, the authors motivate the use of latent states by stating that it is more robust to distractions from task-irrelevant factors (e.g. noisy TV). This claim again lacks support.\n3. The key differences between MTC and RPC [1] are (1) MTC considers states and actions while RPC only considers states, and (2) MTC conditions the dynamics and action prediction models on the history, whereas RPC only conditions on the current state. I'm not sure why (1) would conceptually improve upon RPC (see question below). (2) introduces a confounding variable that is orthogonal to the claim, since we can replace the model in RPC with a recurrent model and hope to see better performance. \n4. While the goal is to induce simplicity in the policy's behavior, the method requires learning three additional models to estimate an auxiliary objective, which is quite cumbersome. \n5. The rest of experiments are unconvincing. While the main results (Table 1) show MTC performing on par or slightly better than baselines, the improvement is not significant. Moreover, these results are in a narrow domain that prefers repetitive motion (e.g. Hopper, Cheetah, Walker are all locomotion tasks). I'm not sure if the same holds for more complex tasks. The robustness results in Figure 2 don't imply MTC being *more* robust than baselines, because SAC shows the same growth trend, just starting from a lower initial value. What I hope to see is that MTC performance degrades *at a slower rate than* SAC, instead of having a constant offset. \n\nOverall, I think this paper can be significantly strengthened by (1) grounding it in a concrete problem setting like sim2real transfer instead of a lofty desideratum (i.e. simplicity), and (2) either removing the unwarranted claims or adding experiments to support them. I strongly recommend the authors to revise and resubmit.\n\n**List of Claims (cont'd)**\n\nClaim 2: simplicity is desirable for RL.\n- Assumption: none\n- Evidence: slightly better empirical performance of MTC over baselines on DM control tasks.\n- Statement: I don't think there is enough theoretical or empirical evidence to support this claim. Given that the DM control tasks prefer periodic behavior, it might be that simplicity is particularly useful in this domain. \n\nClaim 3: using latent states results in less distraction from task-irrelevant factors.\n- Assumption: the environment consists of task-irrelevant distractors, modeling which hinder the policy's performance.\n- Evidence: none\n- Statement: to support this claim, the authors need to compare their method to a baseline that uses raw observations on environments with distractors. \n\nClaim 4: MTC improves upon RPC by considering states and actions, and conditioning the models on the history.\n- Assumption: none\n- Evidence: slightly better empirical performance than RPC.\n- Statement: I'm not sure \"why\" these changes would improve upon RPC. The sequence of states should determine the sequence of actions to a large extent. Conditioning on the history is a choice of variational approximation that can also be applied to RPC, and the paper doesn't compare to RPC + recurrent models. \n\n[1] Benjamin Eysenbach, Ruslan Salakhutdinov, Sergey Levine. Robust Predictable Control. NeurIPS 2021.", "questions": "1. Why is simplicity a good inductive bias for RL? Does this inductive bias hold outside of locomotion domains where repetitive motion is preferred?\n2. Why is accounting for the states and actions better than only accounting for states as in RPC? Isn't the state trajectory a consequence of the action trajectory? \n3. In Equation 2, what does it mean to take expectation under $f$?\n4. Does the infinite horizon variant correspond to a meaningful information-theoretic objective?\n5. Can you provide more experiments to support your claims?\n    - a manipulation task where the motion is less periodic, to support the claim that simplicity is a universally desirable property.\n    - visual experiments with distractors (e.g. distracted DM control) to support the claim that latent-state policies are robust to distractors.\n    - visual robustness experiments with lighting / texture change similar to the sim2real setting, to support the claim that MTC is robust to distribution shifts.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730611901358}, {"id": "IORYqEezwc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6703/Reviewer_DvYG"], "rating": 5, "soundness": 2, "presentation": 4, "contribution": 2, "confidence": 2, "summary": "This paper proposes a method that learns a policy that produces compressible trajectories, or more precisely trajectories that have a high total correlation. The motivation is that a simple policy is probably more robust to spurious correlations, noise, etc. than a more complex policy.\n\nThe authors propose to achieve this by maximizing a variational lower bound of the total correlation, which practically result in training a dynamics model and an action prediction model along with the policy.\nResults show that this works pretty well in experiments on DMC, and improves robustness to observation noise, action noise and changed friction.", "review_text": "This paper proposes a method that learns a policy that produces compressible trajectories, or more precisely trajectories that have a high total correlation. The motivation is that a simple policy is probably more robust to spurious correlations, noise, etc. than a more complex policy.\n\nThe authors propose to achieve this by maximizing a variational lower bound of the total correlation, which practically result in training a dynamics model and an action prediction model along with the policy.\nResults show that this works pretty well in experiments on DMC, and improves robustness to observation noise, action noise and changed friction.", "strengths": "* Training more robust policies is an important challenge for RL and framing this through trajectory compressiblity is a very natural and nice formulation.\n * The paper is very well written and easy to follow \n * Experiments show that the method does indeed result in more periodic, more compressible trajectories and is more robust to noise.\n\n**Structured into claims**:\n * Claim 1: Proposed method generates periodic/compressible trajectories. *Evidence*: Visualizations of actions over a trajectory indeed look very periodic. Compressiblity is validated by using lossless compression methods.\n * Claim 2: Perioidicity improves robustness of policy. *Evidence*: Policies are indeed more robust to observation noise, action noise and changed friction coefficient.", "weaknesses": "* My primary concern is the definition of the regularized reward function $r^*(s_t,a_t,s_{t+1})$. This reward depends on the entire history of the episode through the dynamics model $q_\\mu(z_{t+1}|z_{1:t},a_{1:t})$. It should thus correctly be written $r^*(s_{a:t},a_{1:t},s_{t+1})$ and is no longer Markovian. However, this non-markovianness or its implications are not discussed anywhere. I would expect this to cause severe instabilities in training, but maybe that's not the case? At the very least this should be discussed and I would like to see some visualization of the training dynamics in the form of a training step vs reward plot. \n\n * The experimental validation is somewhat limited. Hyperparameter optimization is not performed in a structured way, both for the baselines and the proposed method. \n * The introduction motivates the method through combatting spurious correlation but there are not distractors etc. present in the experiments that could lead to spurious correlation beyond a noisy state observation. Likewise, the utilization of a learned state representation $z$ is motivated by the noisy TV problem, but this also isn't directly investigated in the experiments. \n * An ablation of only using the latent-dynamics model but not the action model would also be useful to see the contribution of the different parts of the method. This would be somewhere between the proposed method and RPC and would thus be very interesting to see.\n * This method adds additional dynamic models that are trained along with the the policy, so naturally the training time should be longer.\n\n**Structured into claims:**\n * Claim 3: The method reduces spurious correlation and focuses only task-relevant state information by using a learned state encoding. *Evidence*: Only indirect evidence is given through experiments with observation noise. I don't think this is sufficient to support the claim and experiments with explicit distractors would be more convincing.\n * Claim 4: Maximum Total Corrlation SAC as a well motivated RL method. *Evidence*: The derivation hides the non-markovianness/nonstationarity of the reward function by incorrectly listing its arguments. This is my main concern with this paper and not discussed at all. The experiments indicate that it's practically not a big issue but I would like this to be elucidated more.\n\n\n\nMinor comments:\n * Line 278: missing closing bracket for expectation\n * Line 671: $q(a_{t+1}...$ should be $q_\\mathcal{X}(a_{t+1}...$\n * It would also be nice to discuss the connection to works that directly learn periodic controllers such as [1], and maybe even the connection to periodic locomotion controllers which are popular in classical robotics.\n\n[1] Raffin, Antonin, et al. \"An Open-Loop Baseline for Reinforcement Learning Locomotion Tasks.\" RLC 2024", "questions": "* I would apprecite if you could reply to my concern above about the non-markovianness regularized reward function.\n * As the regularized reward could affect training stability, could you show some training-step vs reward training curves?\n * As you are training additional models, the training probably takes significantly longer than normal SAC and even RPC. Could you show a wall-time comparison?\n\nOverall I really like the idea of the paper and the proposed method, but I do have the listed concerns about the regularized reward and about the limitations of the experimental validation.\n\nI'm willing to raise my score if the concerns are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method that learns a policy that produces compressible trajectories, or more precisely trajectories that have a high total correlation. The motivation is that a simple policy is probably more robust to spurious correlations, noise, etc. than a more complex policy.\n\nThe authors propose to achieve this by maximizing a variational lower bound of the total correlation, which practically result in training a dynamics model and an action prediction model along with the policy.\nResults show that this works pretty well in experiments on DMC, and improves robustness to observation noise, action noise and changed friction.", "soundness": 2, "presentation": 4, "contribution": 2, "strengths": "* Training more robust policies is an important challenge for RL and framing this through trajectory compressiblity is a very natural and nice formulation.\n * The paper is very well written and easy to follow \n * Experiments show that the method does indeed result in more periodic, more compressible trajectories and is more robust to noise.\n\n**Structured into claims**:\n * Claim 1: Proposed method generates periodic/compressible trajectories. *Evidence*: Visualizations of actions over a trajectory indeed look very periodic. Compressiblity is validated by using lossless compression methods.\n * Claim 2: Perioidicity improves robustness of policy. *Evidence*: Policies are indeed more robust to observation noise, action noise and changed friction coefficient.", "weaknesses": "* My primary concern is the definition of the regularized reward function $r^*(s_t,a_t,s_{t+1})$. This reward depends on the entire history of the episode through the dynamics model $q_\\mu(z_{t+1}|z_{1:t},a_{1:t})$. It should thus correctly be written $r^*(s_{a:t},a_{1:t},s_{t+1})$ and is no longer Markovian. However, this non-markovianness or its implications are not discussed anywhere. I would expect this to cause severe instabilities in training, but maybe that's not the case? At the very least this should be discussed and I would like to see some visualization of the training dynamics in the form of a training step vs reward plot. \n\n * The experimental validation is somewhat limited. Hyperparameter optimization is not performed in a structured way, both for the baselines and the proposed method. \n * The introduction motivates the method through combatting spurious correlation but there are not distractors etc. present in the experiments that could lead to spurious correlation beyond a noisy state observation. Likewise, the utilization of a learned state representation $z$ is motivated by the noisy TV problem, but this also isn't directly investigated in the experiments. \n * An ablation of only using the latent-dynamics model but not the action model would also be useful to see the contribution of the different parts of the method. This would be somewhere between the proposed method and RPC and would thus be very interesting to see.\n * This method adds additional dynamic models that are trained along with the the policy, so naturally the training time should be longer.\n\n**Structured into claims:**\n * Claim 3: The method reduces spurious correlation and focuses only task-relevant state information by using a learned state encoding. *Evidence*: Only indirect evidence is given through experiments with observation noise. I don't think this is sufficient to support the claim and experiments with explicit distractors would be more convincing.\n * Claim 4: Maximum Total Corrlation SAC as a well motivated RL method. *Evidence*: The derivation hides the non-markovianness/nonstationarity of the reward function by incorrectly listing its arguments. This is my main concern with this paper and not discussed at all. The experiments indicate that it's practically not a big issue but I would like this to be elucidated more.\n\n\n\nMinor comments:\n * Line 278: missing closing bracket for expectation\n * Line 671: $q(a_{t+1}...$ should be $q_\\mathcal{X}(a_{t+1}...$\n * It would also be nice to discuss the connection to works that directly learn periodic controllers such as [1], and maybe even the connection to periodic locomotion controllers which are popular in classical robotics.\n\n[1] Raffin, Antonin, et al. \"An Open-Loop Baseline for Reinforcement Learning Locomotion Tasks.\" RLC 2024", "questions": "* I would apprecite if you could reply to my concern above about the non-markovianness regularized reward function.\n * As the regularized reward could affect training stability, could you show some training-step vs reward training curves?\n * As you are training additional models, the training probably takes significantly longer than normal SAC and even RPC. Could you show a wall-time comparison?\n\nOverall I really like the idea of the paper and the proposed method, but I do have the listed concerns about the regularized reward and about the limitations of the experimental validation.\n\nI'm willing to raise my score if the concerns are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730271907175}, {"id": "NZuLcH1b93", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6703/Reviewer_dKjy"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This work proposes a novel RL method called MTC-RL, which introduces the total correlation as a key inductive bias for learning robust and generalizable policy and consistent state representation. The proposed method encourages the policy and state encoder to generate more simple and consistent behaviors and representations by maximizing the total correlation within the induced trajectories. In other words, **the authors claim that simultaneously learning state-consistent representation and action-consistent policy 1) makes RL algorithms more general and robust, 2) can save the amount of information of all states and actions within trajectories, and 3) produces improvements in final RL performance.** The experimental section demonstrates various supports to support the claims, resulting in better robustness to noise and changes in dynamics, particularly in robot control tasks.", "review_text": "This work proposes a novel RL method called MTC-RL, which introduces the total correlation as a key inductive bias for learning robust and generalizable policy and consistent state representation. The proposed method encourages the policy and state encoder to generate more simple and consistent behaviors and representations by maximizing the total correlation within the induced trajectories. In other words, **the authors claim that simultaneously learning state-consistent representation and action-consistent policy 1) makes RL algorithms more general and robust, 2) can save the amount of information of all states and actions within trajectories, and 3) produces improvements in final RL performance.** The experimental section demonstrates various supports to support the claims, resulting in better robustness to noise and changes in dynamics, particularly in robot control tasks.", "strengths": "1. The claims 1), 2), and 3) are partly supported by experiments in Sections 5.2, 5.3, and 5.1, respectively, providing a proper presentation of this work.\n2. This work provides a derivation using the lower bound of the total correlation, provides mathematical support, and produces a unified objective containing two information regularization techniques: state-consistency and action-consistency.", "weaknesses": "**The main claims are not sufficiently supported.**\n1. While this paper emphasizes the unified framework of state-consistency and action-consistency, the authors did not investigate whether the unified framework is required for their results. I strongly recommend conducting an ablation study using either state-consistency or action-consistency to support claims 1), 2), and 3).\n2. Regarding claims 1) and 2), although this paper includes the state-of-the-art algorithm LZ-SAC, that paper also includes another algorithm called SPAC that imposes simplicity on action sequences. LZ-SAC paper reports that while LZ-SAC achieves better final performance than SPAC, SPAC shows better results in robustness and information efficiency. To fully support the claims 1) and 2), I recommend including SPAC as an additional baseline.\n3. Regarding claim 1), while both RPC and LZ-SAC are expected to show robustness to dynamics changes, Figure 2 shows that they perform worse than SAC, which gives inconsistency of experiments.\n4. Some works such as SPAC use transformers for their action sequence compressor, while other works such as this work use LSTM. Could the authors show that the merit of their work comes from the proposed unified framework, not the choice of the type of compressors?", "questions": "Please see the Weaknesses above. Here are additional questions:\n\n1. While the modified reward (eq.4) could explain how it biases consistent behaviors and latent representations, it is unclear how maximizing the total correlation in the original objective (eq.1) biases the policy and encoder towards such behaviors and representations. Could the authors provide an explanation or an intuitive toy example?\n\n2. Are there any experimental results in environments with higher dimensions? For example, tasks where the input is an observation instead of a state.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes a novel RL method called MTC-RL, which introduces the total correlation as a key inductive bias for learning robust and generalizable policy and consistent state representation. The proposed method encourages the policy and state encoder to generate more simple and consistent behaviors and representations by maximizing the total correlation within the induced trajectories. In other words, **the authors claim that simultaneously learning state-consistent representation and action-consistent policy 1) makes RL algorithms more general and robust, 2) can save the amount of information of all states and actions within trajectories, and 3) produces improvements in final RL performance.** The experimental section demonstrates various supports to support the claims, resulting in better robustness to noise and changes in dynamics, particularly in robot control tasks.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The claims 1), 2), and 3) are partly supported by experiments in Sections 5.2, 5.3, and 5.1, respectively, providing a proper presentation of this work.\n2. This work provides a derivation using the lower bound of the total correlation, provides mathematical support, and produces a unified objective containing two information regularization techniques: state-consistency and action-consistency.", "weaknesses": "**The main claims are not sufficiently supported.**\n1. While this paper emphasizes the unified framework of state-consistency and action-consistency, the authors did not investigate whether the unified framework is required for their results. I strongly recommend conducting an ablation study using either state-consistency or action-consistency to support claims 1), 2), and 3).\n2. Regarding claims 1) and 2), although this paper includes the state-of-the-art algorithm LZ-SAC, that paper also includes another algorithm called SPAC that imposes simplicity on action sequences. LZ-SAC paper reports that while LZ-SAC achieves better final performance than SPAC, SPAC shows better results in robustness and information efficiency. To fully support the claims 1) and 2), I recommend including SPAC as an additional baseline.\n3. Regarding claim 1), while both RPC and LZ-SAC are expected to show robustness to dynamics changes, Figure 2 shows that they perform worse than SAC, which gives inconsistency of experiments.\n4. Some works such as SPAC use transformers for their action sequence compressor, while other works such as this work use LSTM. Could the authors show that the merit of their work comes from the proposed unified framework, not the choice of the type of compressors?", "questions": "Please see the Weaknesses above. Here are additional questions:\n\n1. While the modified reward (eq.4) could explain how it biases consistent behaviors and latent representations, it is unclear how maximizing the total correlation in the original objective (eq.1) biases the policy and encoder towards such behaviors and representations. Could the authors provide an explanation or an intuitive toy example?\n\n2. Are there any experimental results in environments with higher dimensions? For example, tasks where the input is an observation instead of a state.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730184549091}, {"id": "joRui7zFIS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6703/Reviewer_DJ5C"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes regularizing standard reinforcement learning (RL) with total correlation between state representations and actions in the whole trajectories to achieve *simpler* and more robust policies.", "review_text": "This paper proposes regularizing standard reinforcement learning (RL) with total correlation between state representations and actions in the whole trajectories to achieve *simpler* and more robust policies.", "strengths": "- The motivation and presentation are clear and easy to follow.\n- The experimental results effectively demonstrate that the method learns periodic policies.", "weaknesses": "- The concept of total correlation is quite similar to mutual information, which has been extensively explored in both single-agent and multi-agent RL. Therefore, the novelty is limited, with the primary distinction being that prior work considers state-action pairs, while total correlation extends this to entire trajectories.\n- Although the method was evaluated on 8 tasks in the DeepMind Control (DMC) suite, these tasks are relatively simple and inherently periodic. For such tasks, structurally designed policies or dynamic movement primitives (DMP) could potentially be more efficient. Therefore, it is essential to assess the generalizability of maximizing total correlation in more complex tasks, such as robotic grasping or kitchen environments.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes regularizing standard reinforcement learning (RL) with total correlation between state representations and actions in the whole trajectories to achieve *simpler* and more robust policies.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The motivation and presentation are clear and easy to follow.\n- The experimental results effectively demonstrate that the method learns periodic policies.", "weaknesses": "- The concept of total correlation is quite similar to mutual information, which has been extensively explored in both single-agent and multi-agent RL. Therefore, the novelty is limited, with the primary distinction being that prior work considers state-action pairs, while total correlation extends this to entire trajectories.\n- Although the method was evaluated on 8 tasks in the DeepMind Control (DMC) suite, these tasks are relatively simple and inherently periodic. For such tasks, structurally designed policies or dynamic movement primitives (DMP) could potentially be more efficient. Therefore, it is essential to assess the generalizability of maximizing total correlation in more complex tasks, such as robotic grasping or kitchen environments.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730042322104}], "openreview_url": "https://openreview.net/forum?id=JeiaHDawhb", "arxiv_id": "2505.16734", "paper_pdf": "papers/JeiaHDawhb.pdf", "paper_pdf_sha256": "06c846adbac422813c8e2e3b292b0d92251a7c2cdbc0258157587faa17eeaaec", "paper_pdf_bytes": 454123, "paper_pdf_source": "openreview", "code_url": "https://github.com/BangYou01/MTC", "code_repository": "BangYou01/MTC", "code_commit": "7b4dfae88d6182e58fa7c7c85af5a6bfadc5f863", "code_archive": "repos/JeiaHDawhb.zip", "code_archive_sha256": "ac2fc141ce9541ce029fba1050d6a435859f1aebb0857489c4d37c9ee8e1ff90", "code_archive_bytes": 14799, "code_file_count": 6, "code_extensions": {".py": 5, ".sh": 1}, "github_disk_usage_kb": 13, "github_languages": {"Python": 50767, "Shell": 426}, "github_archived": false, "github_pushed_at": "2025-05-28T06:55:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/maximum-total-correlation-reinforcement"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SHUQtRK0eU", "year": 2024, "status": "rejected", "title": "Generalized Activation via Multivariate Projection", "authors": ["Jiayun Li", "Yuxiao Cheng", "Zhuofan Xia", "Yilin Mo", "Gao Huang"], "authorids": ["~Jiayun_Li2", "~Yuxiao_Cheng1", "~Zhuofan_Xia2", "~Yilin_Mo1", "~Gao_Huang1"], "authors_source": "OpenReview API", "abstract": "Activation functions are essential to introduce nonlinearity into neural networks, with the Rectified Linear Unit (ReLU) often favored for its simplicity and effectiveness. Motivated by the structural similarity between a shallow Feedforward Neural Network (FNN) and a single iteration of the Projected Gradient Descent (PGD) algorithm, a standard approach for solving constrained optimization problems, we consider ReLU as a projection from $\\mathbb{R}$ onto the nonnegative half-line $\\mathbb{R}_+$. \nBuilding on this interpretation, we extend ReLU by substituting it with a generalized projection operator onto a convex cone, such as the Second-Order Cone (SOC) projection, thereby naturally extending it to a Multivariate Projection Unit (MPU), an activation function with multiple inputs and multiple outputs.\nWe further provide a mathematical proof establishing that FNNs activated by SOC projections outperform those utilizing ReLU in terms of expressive power. Experimental evaluations on widely-adopted architectures further corroborate MPU's effectiveness against a broader range of existing activation functions.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "DvQ1zp3FPT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6779/Reviewer_L1wf"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Most activations functions employed in DNNs are Single Input Single Output (SISO). One can naturally ask whether we can gain by using a Multiple Input Multiple Output (MIMO) activation function instead. Of course, the answer depends on how we design such MIMO activations. This paper explores a specific strategy for designing MIMO activations: first based on projections, and more generally based on proximal functions. \n\nThe motivation the authors give for this two types of MIMO activations is that ReLU (one of the most widely used SISO activations) is a projection, and that a shallow network can be parameterized so to mimic a single iteration of Projected Gradient Descent. This is the high level observation that motivates their definition. To further motivate it mathematically, the authors prove a couple of theorems showing that ReLU can be, in a sense, mimicked by their activation, while the other direction does not.  The authors further explore proximal MIMO functions, by using proximal operator, and connecting to proximal gradient descent. They motivate it as a MIMO extension of leaky activation functions.\n\nFinally, the authors show experiments that validate their theory and the fact that somewhat better empirical results can be obtained with their methods.", "review_text": "Most activations functions employed in DNNs are Single Input Single Output (SISO). One can naturally ask whether we can gain by using a Multiple Input Multiple Output (MIMO) activation function instead. Of course, the answer depends on how we design such MIMO activations. This paper explores a specific strategy for designing MIMO activations: first based on projections, and more generally based on proximal functions. \n\nThe motivation the authors give for this two types of MIMO activations is that ReLU (one of the most widely used SISO activations) is a projection, and that a shallow network can be parameterized so to mimic a single iteration of Projected Gradient Descent. This is the high level observation that motivates their definition. To further motivate it mathematically, the authors prove a couple of theorems showing that ReLU can be, in a sense, mimicked by their activation, while the other direction does not.  The authors further explore proximal MIMO functions, by using proximal operator, and connecting to proximal gradient descent. They motivate it as a MIMO extension of leaky activation functions.\n\nFinally, the authors show experiments that validate their theory and the fact that somewhat better empirical results can be obtained with their methods.", "strengths": "Overall, I liked the paper. It was very fun reading. In particular, the following are the main strengths: \n\n- Very well written paper.\n- Clear motivation in what they want to achieve (MIMO activation) and in how they achieve it (based on projections/proximal functions).\n- Supporting theory.\n- Empirical results suggesting the benefits of their approach.", "weaknesses": "However, there are some very substantial weaknesses:\n\n- *[Removed due to answers by authors:]* Theory is basic and simple. But more importantly: it does not really, at the essence, establish why they function should give better results. Sure, you can set the weights just correctly so that a layer is like an iteration. But, so what? The weights are actually learned, and our goal is not to really learn an iteration...\n\n- *[Removed due to answers by authors:]* I am a bit skeptical about the motivation. The way I see it, the whole idea in the activation is to introduce *some* nonlinearity. It does not have to be a lot, to really get a low of expressive power! Just by doing many layers, we can take a really small amount of non-linearity and make it a lot. So, not sure that MIMO really will help...\n\n- *[Added due to answers by authors:]* Limited potential improvements: the authors observed that only using very small input-output size (namely, m=2, which replaces one input one output with two input two output) is the best you can do. Going to m=3 and above does not help. This suggest that there is very little to gain from MIMO non-linearity.\n\n- Empirical improvement is very small, and even then not always achieved.", "questions": "- How are MIMO activations implemented? *[Following rebuttal:]* Still unclear on how exactly the projection itself is implemented.\n- Why use small m in the experiments? *[Following rebuttal:]* OK, understood. However, this raises the question on why MPU is a good idea if only very small m is used. See new weakness above.\n- Table 1: improvements in training error should be compared with cost (MACS here). With MPU you get slighly better errors with slightly more MACS. What happens with ReLU if you increase the MACS a bit, e.g. by doing a few more epochs? \nSure, you test ResNet 34 and 50  in which MACS increase. But there is more than one way to increase the MACS...\n*[Following rebuttal:]* OK, doing more epochs might make sense less, but there are other ways. e.g. slightly wider network?\n-*[New following rebuttal:]*  The authors use the same hyperparameters across all experiments, with the sole variation being the Activation Function (AF).\" Why? And, more importantly, how the results will look if all algorithms use their best hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Most activations functions employed in DNNs are Single Input Single Output (SISO). One can naturally ask whether we can gain by using a Multiple Input Multiple Output (MIMO) activation function instead. Of course, the answer depends on how we design such MIMO activations. This paper explores a specific strategy for designing MIMO activations: first based on projections, and more generally based on proximal functions. \n\nThe motivation the authors give for this two types of MIMO activations is that ReLU (one of the most widely used SISO activations) is a projection, and that a shallow network can be parameterized so to mimic a single iteration of Projected Gradient Descent. This is the high level observation that motivates their definition. To further motivate it mathematically, the authors prove a couple of theorems showing that ReLU can be, in a sense, mimicked by their activation, while the other direction does not.  The authors further explore proximal MIMO functions, by using proximal operator, and connecting to proximal gradient descent. They motivate it as a MIMO extension of leaky activation functions.\n\nFinally, the authors show experiments that validate their theory and the fact that somewhat better empirical results can be obtained with their methods.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "Overall, I liked the paper. It was very fun reading. In particular, the following are the main strengths: \n\n- Very well written paper.\n- Clear motivation in what they want to achieve (MIMO activation) and in how they achieve it (based on projections/proximal functions).\n- Supporting theory.\n- Empirical results suggesting the benefits of their approach.", "weaknesses": "However, there are some very substantial weaknesses:\n\n- *[Removed due to answers by authors:]* Theory is basic and simple. But more importantly: it does not really, at the essence, establish why they function should give better results. Sure, you can set the weights just correctly so that a layer is like an iteration. But, so what? The weights are actually learned, and our goal is not to really learn an iteration...\n\n- *[Removed due to answers by authors:]* I am a bit skeptical about the motivation. The way I see it, the whole idea in the activation is to introduce *some* nonlinearity. It does not have to be a lot, to really get a low of expressive power! Just by doing many layers, we can take a really small amount of non-linearity and make it a lot. So, not sure that MIMO really will help...\n\n- *[Added due to answers by authors:]* Limited potential improvements: the authors observed that only using very small input-output size (namely, m=2, which replaces one input one output with two input two output) is the best you can do. Going to m=3 and above does not help. This suggest that there is very little to gain from MIMO non-linearity.\n\n- Empirical improvement is very small, and even then not always achieved.", "questions": "- How are MIMO activations implemented? *[Following rebuttal:]* Still unclear on how exactly the projection itself is implemented.\n- Why use small m in the experiments? *[Following rebuttal:]* OK, understood. However, this raises the question on why MPU is a good idea if only very small m is used. See new weakness above.\n- Table 1: improvements in training error should be compared with cost (MACS here). With MPU you get slighly better errors with slightly more MACS. What happens with ReLU if you increase the MACS a bit, e.g. by doing a few more epochs? \nSure, you test ResNet 34 and 50  in which MACS increase. But there is more than one way to increase the MACS...\n*[Following rebuttal:]* OK, doing more epochs might make sense less, but there are other ways. e.g. slightly wider network?\n-*[New following rebuttal:]*  The authors use the same hyperparameters across all experiments, with the sole variation being the Activation Function (AF).\" Why? And, more importantly, how the results will look if all algorithms use their best hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698662587242}, {"id": "p5OGUhfiaW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6779/Reviewer_WKdp"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper expands the SISO activation functions in neural networks by introducing the MPU as a MIMO activation function. This expansion is motivated by the structural resemblance between a shallow FNN and a single iteration of the PGD algorithm. Experiments test the effectiveness of the proposed approach.", "review_text": "The paper expands the SISO activation functions in neural networks by introducing the MPU as a MIMO activation function. This expansion is motivated by the structural resemblance between a shallow FNN and a single iteration of the PGD algorithm. Experiments test the effectiveness of the proposed approach.", "strengths": "1.\tThe idea that choose the activation function to be the projection onto the convex cone is very interesting.\n2.\tThe paper provides some theoretical proofs.\n3.\tSome experiments demonstrate the effectiveness of the proposed MIMO activation function.", "weaknesses": "1.\tThe organization and writing of the paper need to be improved.\n2.\tThe proposed Theorem 1 is not so rigorous.\n3.\tThe experiments are not enough.", "questions": "1.\tThe projected gradient descent (PGD) algorithm in proposition 1 has many iterative steps. However, the Theorem1 has a single layer, which corresponds a single iteration of the PGD algorithm. The structural connection is weak.\n2.\tIn Theorem 1, it seems that it needs to remove W^{(2)}=I, b^{(2)}=0. We only need  W^{(1)}, b^{(1)}, and ReLU activation function to represent a single step of PGD algorithm.\n3.\tHow to set \\alpha in Theorem 1 and Proposition 2 in the network training? It is especially to the convolutional network. It is better to give some discussion on the setting of \\alpha.\n4.\tThe paper only gives the derivations with fully-connected feedforward networks. It does not present the skip-connection. Since the experiments use ResNets, it is better to discuss this issue.\n5.\tThe paper only tests the proposed MPU with ResNet18. It is not a very deep network.  Is the proposed MPU unable to train deep networks? It is best to give an experiment to train deep networks, such as resnet101? This is very critical, which determines whether the proposed MPU can be applied in practice. \n6.\tThe organization of section 2 needs to be improved. It is hard to follow.\n7.\tWhen referencing equations, it is better to use “\\eqref”.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper expands the SISO activation functions in neural networks by introducing the MPU as a MIMO activation function. This expansion is motivated by the structural resemblance between a shallow FNN and a single iteration of the PGD algorithm. Experiments test the effectiveness of the proposed approach.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1.\tThe idea that choose the activation function to be the projection onto the convex cone is very interesting.\n2.\tThe paper provides some theoretical proofs.\n3.\tSome experiments demonstrate the effectiveness of the proposed MIMO activation function.", "weaknesses": "1.\tThe organization and writing of the paper need to be improved.\n2.\tThe proposed Theorem 1 is not so rigorous.\n3.\tThe experiments are not enough.", "questions": "1.\tThe projected gradient descent (PGD) algorithm in proposition 1 has many iterative steps. However, the Theorem1 has a single layer, which corresponds a single iteration of the PGD algorithm. The structural connection is weak.\n2.\tIn Theorem 1, it seems that it needs to remove W^{(2)}=I, b^{(2)}=0. We only need  W^{(1)}, b^{(1)}, and ReLU activation function to represent a single step of PGD algorithm.\n3.\tHow to set \\alpha in Theorem 1 and Proposition 2 in the network training? It is especially to the convolutional network. It is better to give some discussion on the setting of \\alpha.\n4.\tThe paper only gives the derivations with fully-connected feedforward networks. It does not present the skip-connection. Since the experiments use ResNets, it is better to discuss this issue.\n5.\tThe paper only tests the proposed MPU with ResNet18. It is not a very deep network.  Is the proposed MPU unable to train deep networks? It is best to give an experiment to train deep networks, such as resnet101? This is very critical, which determines whether the proposed MPU can be applied in practice. \n6.\tThe organization of section 2 needs to be improved. It is hard to follow.\n7.\tWhen referencing equations, it is better to use “\\eqref”.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698588530430}, {"id": "C7N5fVXSO9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6779/Reviewer_MyBG"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors introduce \"Multivariate Projection Unit\" (MPU), a novel activation function that, differently from a standard AF, takes multiple inputs and also returns multiple outputs (MIMO). The key observation is that ReLU can be seen as a a projection to the positive orthant, and linear multiplication + ReLU can be seen as a step of projected gradient descent. The MPU, instead, splits the output of the linear projection into blocks of same size (see Fig. 1), and projects each block onto a prespecified cone. This is motivated by the analogy with a PGD step on a different (more general) class of optimization problems. They validate the MPU on artificial benchmarks, a CNN on CIFAR-10 and ImageNet, and a vision transformer (only in the supplementary material), showing results that are similar or slightly better than ReLU and some ReLU variants (e.g., Leaky ReLU).", "review_text": "The authors introduce \"Multivariate Projection Unit\" (MPU), a novel activation function that, differently from a standard AF, takes multiple inputs and also returns multiple outputs (MIMO). The key observation is that ReLU can be seen as a a projection to the positive orthant, and linear multiplication + ReLU can be seen as a step of projected gradient descent. The MPU, instead, splits the output of the linear projection into blocks of same size (see Fig. 1), and projects each block onto a prespecified cone. This is motivated by the analogy with a PGD step on a different (more general) class of optimization problems. They validate the MPU on artificial benchmarks, a CNN on CIFAR-10 and ImageNet, and a vision transformer (only in the supplementary material), showing results that are similar or slightly better than ReLU and some ReLU variants (e.g., Leaky ReLU).", "strengths": "To the best of my knowledge, the MPU is novel, and it is an interesting variation of a standard ReLU with a good underlying motivation. The paper is well written, in particular the visualizations in Fig. 1 immediately show the basic idea of the paper. I am less convinced about the empirical evaluation (see below), so the practical value of the MPU is not clear.", "weaknesses": "I have a few general comments on the manuscript, making this (at the moment) a borderline paper for acceptance. I think most of these questions are addressable and I would be happy to increase my score.\n\nEXPOSITION: I have found the exposition of the paper a bit strange, because there is a very long motivation for the MPU (both Section 2.1 and Section 3 serve as a motivation), but very little analysis of the MPU itself. For example: (i) there is no explicit definition of the MPU or the MPU layer; (ii) there is no discussion on how to choose the cone (they only specify which cone they use in the experiments); (iii) no discussion on how to perform the projection (which is relegated to an appendix); (iv) no discussion on the theoretical computational complexity (only an empirical computation of MACS).\n\nRESULTS: The results do not seem very strong. Ignoring the artificial datasets, on CIFAR-10 (Table 1) it is inside 1 std of Leaky ReLU. On ImageNet (Table 2), LeakyReLU is superior but also inside 1 std. DeiT (Appendix F) is only provided for a single run and no comparison, and ReLU is still only at 0.07 distance. Also, they are not providing many important ablations and comparisons, including on the choice of the cone or its dimensionality. Many baselines are missing (see below).\n\nRELATED WORKS: the related works section is very shallow. Strangely, they are mentioning some works on multi-input AFs which are not used in the comparisons (e.g., Maxout, CReLU). However, many other works are missing (see, e.g., https://arxiv.org/pdf/2005.00817.pdf), including winner-take-all AFs, network-in-network models. Also, complex activation functions can natively work with 2 inputs and 2 outputs and can be generalized (e.g., quaternion, octanion) for more dimensions.", "questions": "The major questions on the paper are related to the points above, in particular:\n1. Provide a clear definition of the activation function, and add discussions on its design, including the choice of the cone and the computational complexity of the projection operation.\n2. Show at least one use case where the AF provide significant improvements, either in accuracy or time.\n3. Add more ablations and baselines, especially of AFs which are closely linked to the paper.\n\nI also have a few minor additional questions:\n1. In the citations there are multiple references to \"MS Windows NT kernel description\"?\n2. P4: \"the set S is a certain polyhedron\", can you clarify what polyhedron or point to a specific definition in the paper.\n3. I would advise to add a definition for the cone C_n.\n\nEDIT AFTER REBUTTAL: point (1) was mostly solved. Point (2) is still standing. Point (3) was partially solved. Most minor questions were solved. I have increased my score from 6 to 8 because I believe this is an interesting research direction. However, experimental results (at the moment) are unconvincing.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce \"Multivariate Projection Unit\" (MPU), a novel activation function that, differently from a standard AF, takes multiple inputs and also returns multiple outputs (MIMO). The key observation is that ReLU can be seen as a a projection to the positive orthant, and linear multiplication + ReLU can be seen as a step of projected gradient descent. The MPU, instead, splits the output of the linear projection into blocks of same size (see Fig. 1), and projects each block onto a prespecified cone. This is motivated by the analogy with a PGD step on a different (more general) class of optimization problems. They validate the MPU on artificial benchmarks, a CNN on CIFAR-10 and ImageNet, and a vision transformer (only in the supplementary material), showing results that are similar or slightly better than ReLU and some ReLU variants (e.g., Leaky ReLU).", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "To the best of my knowledge, the MPU is novel, and it is an interesting variation of a standard ReLU with a good underlying motivation. The paper is well written, in particular the visualizations in Fig. 1 immediately show the basic idea of the paper. I am less convinced about the empirical evaluation (see below), so the practical value of the MPU is not clear.", "weaknesses": "I have a few general comments on the manuscript, making this (at the moment) a borderline paper for acceptance. I think most of these questions are addressable and I would be happy to increase my score.\n\nEXPOSITION: I have found the exposition of the paper a bit strange, because there is a very long motivation for the MPU (both Section 2.1 and Section 3 serve as a motivation), but very little analysis of the MPU itself. For example: (i) there is no explicit definition of the MPU or the MPU layer; (ii) there is no discussion on how to choose the cone (they only specify which cone they use in the experiments); (iii) no discussion on how to perform the projection (which is relegated to an appendix); (iv) no discussion on the theoretical computational complexity (only an empirical computation of MACS).\n\nRESULTS: The results do not seem very strong. Ignoring the artificial datasets, on CIFAR-10 (Table 1) it is inside 1 std of Leaky ReLU. On ImageNet (Table 2), LeakyReLU is superior but also inside 1 std. DeiT (Appendix F) is only provided for a single run and no comparison, and ReLU is still only at 0.07 distance. Also, they are not providing many important ablations and comparisons, including on the choice of the cone or its dimensionality. Many baselines are missing (see below).\n\nRELATED WORKS: the related works section is very shallow. Strangely, they are mentioning some works on multi-input AFs which are not used in the comparisons (e.g., Maxout, CReLU). However, many other works are missing (see, e.g., https://arxiv.org/pdf/2005.00817.pdf), including winner-take-all AFs, network-in-network models. Also, complex activation functions can natively work with 2 inputs and 2 outputs and can be generalized (e.g., quaternion, octanion) for more dimensions.", "questions": "The major questions on the paper are related to the points above, in particular:\n1. Provide a clear definition of the activation function, and add discussions on its design, including the choice of the cone and the computational complexity of the projection operation.\n2. Show at least one use case where the AF provide significant improvements, either in accuracy or time.\n3. Add more ablations and baselines, especially of AFs which are closely linked to the paper.\n\nI also have a few minor additional questions:\n1. In the citations there are multiple references to \"MS Windows NT kernel description\"?\n2. P4: \"the set S is a certain polyhedron\", can you clarify what polyhedron or point to a specific definition in the paper.\n3. I would advise to add a definition for the cone C_n.\n\nEDIT AFTER REBUTTAL: point (1) was mostly solved. Point (2) is still standing. Point (3) was partially solved. Most minor questions were solved. I have increased my score from 6 to 8 because I believe this is an interesting research direction. However, experimental results (at the moment) are unconvincing.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698238516490}], "openreview_url": "https://openreview.net/forum?id=SHUQtRK0eU", "arxiv_id": "2309.17194", "paper_pdf": "papers/SHUQtRK0eU.pdf", "paper_pdf_sha256": "db6c5d626825d337d5d42ef32aa6a8844a3734174b54714b3d379e7960aba569", "paper_pdf_bytes": 1459067, "paper_pdf_source": "openreview", "code_url": "https://github.com/ljy9912/mimo_nn", "code_repository": "ljy9912/mimo_nn", "code_commit": "f008dcc4ba9f1a9931d89e671654de244ede33b3", "code_archive": "repos/SHUQtRK0eU.zip", "code_archive_sha256": "5bb3a03a8c764fd117297f17e43c5257a0ccbe77928ff6e249f84921d7279bec", "code_archive_bytes": 11571, "code_file_count": 8, "code_extensions": {".py": 6, ".sh": 2}, "github_disk_usage_kb": 14, "github_languages": {"Python": 32462, "Shell": 1149}, "github_archived": false, "github_pushed_at": "2024-05-20T04:00:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalized-activation-via-multivariate"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2_BsVZ6R-ef", "year": 2023, "status": "rejected", "title": "Analytical Composition of Differential Privacy via the Edgeworth Accountant", "authors": ["Hua Wang", "Sheng Gao", "Huanyu Zhang", "Milan Shen", "Weijie J Su"], "authorids": ["~Hua_Wang7", "~Sheng_Gao2", "~Huanyu_Zhang2", "~Milan_Shen1", "~Weijie_J_Su1"], "authors_source": "OpenReview API", "abstract": "Many modern machine learning algorithms are composed of simple private algorithms; thus, an increasingly important problem is to efficiently compute the overall privacy loss under composition. In this study, we introduce the Edgeworth Accountant, an analytical approach to composing differential privacy guarantees of private algorithms. The Edgeworth Accountant starts by losslessly tracking the privacy loss under composition using the $f$-differential privacy framework, which allows us to express the privacy guarantees using privacy-loss log-likelihood ratios (PLLRs). As the name suggests, this accountant next uses the Edgeworth expansion to the upper and lower bounds the probability distribution of the sum of the PLLRs. Moreover, by relying on a technique for approximating complex distributions using simple ones, we demonstrate that the Edgeworth Accountant can be applied to the composition of any noise-addition mechanism. Owing to certain appealing features of the Edgeworth expansion, the $(\\epsilon, \\delta)$-differential privacy bounds offered by this accountant are non-asymptotic, with essentially no extra computational cost, as opposed to the prior approaches, wherein the running times increase with the number of compositions. Finally, we demonstrate that our upper and lower $(\\epsilon, \\delta)$-differential privacy bounds are tight in federated analytics and certain regimes of training private deep learning models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "BaZpgOLdH9Y", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1958/Reviewer_ESEo"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper gives computationally efficient privacy bounds for the composition of DP algorithm with finite sample guarantees. The algorithm runs in constant time to compute the privacy loss for m identical DP mechanism, and in general case, the run time in $O(m)$ time. This improves on the previous results that use FFT. The proposed approach in this paper is Edgeworth accountancy. ", "review_text": "Please see above.\n\nFor correctness below, I have not read the proofs so I cannot make a judgement on the correctness of the paper; however, none of the claims seems to be out of ordinary. Hence a rating of 3. Once I have verified the proofs, I will move it to 4. ", "strengths": "The fact that they can compute the privacy cost for composition is constant for identical privacy mechanism and just linear in the general case. The paper also gives numerical experiments to substantiate their claim. \n\nI really did not see much of a weakness in the paper. I really liked reading the paper and if I am not missing something, this paper has definite improvement over previous work. The paper is well written. I have not yet verified the proof, and that is the only reason I am not giving full support in accepting the paper. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper gives computationally efficient privacy bounds for the composition of DP algorithm with finite sample guarantees. The algorithm runs in constant time to compute the privacy loss for m identical DP mechanism, and in general case, the run time in $O(m)$ time. This improves on the previous results that use FFT. The proposed approach in this paper is Edgeworth accountancy. ", "strength_and_weaknesses": "The fact that they can compute the privacy cost for composition is constant for identical privacy mechanism and just linear in the general case. The paper also gives numerical experiments to substantiate their claim. \n\nI really did not see much of a weakness in the paper. I really liked reading the paper and if I am not missing something, this paper has definite improvement over previous work. The paper is well written. I have not yet verified the proof, and that is the only reason I am not giving full support in accepting the paper. ", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written, quality is high. I did not verify the reproducibility of the paper.", "summary_of_the_review": "Please see above.\n\nFor correctness below, I have not read the proofs so I cannot make a judgement on the correctness of the paper; however, none of the claims seems to be out of ordinary. Hence a rating of 3. Once I have verified the proofs, I will move it to 4. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667345438165}, {"id": "F9PNBDm3Vh", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1958/Reviewer_iZYg"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper introduces the Edgeworth accountant, an accountant for composing DP mechanisms that obtains tighter guarantees and/or is faster to evaluate than existing accountants. The Edgeworth accountant of the paper is the first accountant to provide both finite-sample (i.e., non asymptotic in m) lower and upper bounds, with O(1) computational complexity for composing m iid mechanisms and O(m) for m non-iid mechanisms (here I believe the O() hides some dependences on the order of the Edgeworth approximation, though empirically the authors show a small order suffices). In contrast, the FFT accountant of Gopi et al. takes time O(sqrt(m)) and O(m^5/2) respectively, the moments accountant only gives upper bounds for non-asymptotic computations, and the Edgeworth refinement to the GDP accountant of Zheng et al. does not give a finite-sample guarantee.\n\nThe Edgeworth accountant (presented in the body of the paper for composing subsampled Gaussians, but which the authors show can be extended to other noise distributions common in the DP literature such as the Laplace mechanism) uses two ideas. First, it uses the fact that the optimal hypothesis test for which of two databases a DP output came from can be defined only in terms of the output's privacy-loss log-likelihood ratio (PLLR). In turn, the authors can show that composition of the f-DP guarantees given by the PLLR hypothesis tests, is equivalent to an (eps, delta)-DP guarantee written strictly in terms of the CDFs of the sum of the PLLRs. Second, these CDFs can be approximated using the Edgeworth expansion, a generalization of the central limit theorem which approximates the CDF of a distribution in terms of its cumulants. So, one can also approximate the (eps, delta)-DP guarantee given by the first idea, using the Edgeworth expansion. In order to obtain a strict upper/lower bound on the actual DP guarantees, one also needs a bound on the CDF approximation error given by the Edgeworth expansion, which the authors derive, and tighten for the special case of subsampled Gaussians. \n\nThe authors implement the (second-order) Edgeworth mechanism and show that in various settings, (i) the approximation given by their accountant is sandwiched between the upper and lower bounds given by the FFT accountant of Gopi et al, and much more accurate than RDP/CLT-based accountants, and (ii) the interval on epsilon given by their accountant is contained strictly within the interval given by the FFT accountant, and RDP is far outside both these intervals. ", "review_text": "Overall I think the paper is above the acceptance threshold. As mentioned before, I think the paper is strong theoretically, contains some nice new ideas, and has a lot of potential for practical impact. That being said, it is not clear to me that the presentation of the paper, especially to practitioners who may lack some of the statistical background, is as accessible as it could be.", "strengths": "The paper has two main strengths. First, it gives a very tight and efficient accounting scheme, that qualitative and quantitatively improves on past accounting schemes heavily. Accounting schemes are widespread in practice and there is good reason to believe the improvements on accounting will allow us to give stronger DP guarantees in practice (or equivalent, achieve much better accuracy for a target DP guarantee). The qualitative and improvements of the accountant are also good reason to hope it becomes widely used in practice. Second, the ideas leading up to the Edgeworth accountant are slick and feel \"natural in retrospect\", but the paper nonetheless is technically very novel, and the proofs are involved and several complex lemmas are needed to turn the ideas into an algorithm.\n\nThe main weakness of the paper is that a lot of technical background is assumed or quickly skipped over in the exposition. One would imagine the ideal reader for this paper is someone who e.g. maintains an actively used privacy accounting library and is interested in working on implementations of the Edgeworth accountant. It may be hard for such a person to read even through the body of the paper, since I don't believe the Edgeworth expansion is common knowledge among privacy researchers (while something like f-DP might be). Admittedly, this is probably at least somewhat true of most sufficiently technical papers given the page limit. But if it is possible to include even slightly more background on the Edgeworth expansion, it would make the paper much more approachable to the audience of interest. Of course, this may be difficult due to space constraints.\n\nNitpick: I think Proposition 3.2 or some part of its buildup should explicitly state/remind that f_i is the tradeoff given by thresholding the PLLRs, right now this needs to be inferred. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper introduces the Edgeworth accountant, an accountant for composing DP mechanisms that obtains tighter guarantees and/or is faster to evaluate than existing accountants. The Edgeworth accountant of the paper is the first accountant to provide both finite-sample (i.e., non asymptotic in m) lower and upper bounds, with O(1) computational complexity for composing m iid mechanisms and O(m) for m non-iid mechanisms (here I believe the O() hides some dependences on the order of the Edgeworth approximation, though empirically the authors show a small order suffices). In contrast, the FFT accountant of Gopi et al. takes time O(sqrt(m)) and O(m^5/2) respectively, the moments accountant only gives upper bounds for non-asymptotic computations, and the Edgeworth refinement to the GDP accountant of Zheng et al. does not give a finite-sample guarantee.\n\nThe Edgeworth accountant (presented in the body of the paper for composing subsampled Gaussians, but which the authors show can be extended to other noise distributions common in the DP literature such as the Laplace mechanism) uses two ideas. First, it uses the fact that the optimal hypothesis test for which of two databases a DP output came from can be defined only in terms of the output's privacy-loss log-likelihood ratio (PLLR). In turn, the authors can show that composition of the f-DP guarantees given by the PLLR hypothesis tests, is equivalent to an (eps, delta)-DP guarantee written strictly in terms of the CDFs of the sum of the PLLRs. Second, these CDFs can be approximated using the Edgeworth expansion, a generalization of the central limit theorem which approximates the CDF of a distribution in terms of its cumulants. So, one can also approximate the (eps, delta)-DP guarantee given by the first idea, using the Edgeworth expansion. In order to obtain a strict upper/lower bound on the actual DP guarantees, one also needs a bound on the CDF approximation error given by the Edgeworth expansion, which the authors derive, and tighten for the special case of subsampled Gaussians. \n\nThe authors implement the (second-order) Edgeworth mechanism and show that in various settings, (i) the approximation given by their accountant is sandwiched between the upper and lower bounds given by the FFT accountant of Gopi et al, and much more accurate than RDP/CLT-based accountants, and (ii) the interval on epsilon given by their accountant is contained strictly within the interval given by the FFT accountant, and RDP is far outside both these intervals. ", "strength_and_weaknesses": "The paper has two main strengths. First, it gives a very tight and efficient accounting scheme, that qualitative and quantitatively improves on past accounting schemes heavily. Accounting schemes are widespread in practice and there is good reason to believe the improvements on accounting will allow us to give stronger DP guarantees in practice (or equivalent, achieve much better accuracy for a target DP guarantee). The qualitative and improvements of the accountant are also good reason to hope it becomes widely used in practice. Second, the ideas leading up to the Edgeworth accountant are slick and feel \"natural in retrospect\", but the paper nonetheless is technically very novel, and the proofs are involved and several complex lemmas are needed to turn the ideas into an algorithm.\n\nThe main weakness of the paper is that a lot of technical background is assumed or quickly skipped over in the exposition. One would imagine the ideal reader for this paper is someone who e.g. maintains an actively used privacy accounting library and is interested in working on implementations of the Edgeworth accountant. It may be hard for such a person to read even through the body of the paper, since I don't believe the Edgeworth expansion is common knowledge among privacy researchers (while something like f-DP might be). Admittedly, this is probably at least somewhat true of most sufficiently technical papers given the page limit. But if it is possible to include even slightly more background on the Edgeworth expansion, it would make the paper much more approachable to the audience of interest. Of course, this may be difficult due to space constraints.\n\nNitpick: I think Proposition 3.2 or some part of its buildup should explicitly state/remind that f_i is the tradeoff given by thresholding the PLLRs, right now this needs to be inferred. ", "clarity,_quality,_novelty_and_reproducibility": "The quality, novelty, and reproducibility of the paper are all high. As mentioned before, the ideas in the paper are all \"natural in retrospect\" but at the same time a great deal of technical effort is needed for the main results in the paper. Also, the Edgeworth accountant obtains several novel guarantees both qualitatively and quantitatively improving over past work. \n\nAs mentioned before, the authors do a good job explaining the concepts in the paper such that the main ideas are \"natural in retrospect\", so in that sense the paper is mostly clear. However, to someone without the appropriate background reading through the technical discussions that solidify these ideas may be difficult.", "summary_of_the_review": "Overall I think the paper is above the acceptance threshold. As mentioned before, I think the paper is strong theoretically, contains some nice new ideas, and has a lot of potential for practical impact. That being said, it is not clear to me that the presentation of the paper, especially to practitioners who may lack some of the statistical background, is as accessible as it could be.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666671475700}, {"id": "40oMKxYUxb", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1958/Reviewer_kz4w"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The goal of the paper is to present algorithms computing tight non-asymptotic numerical estimates on the privacy parameters of $m$-fold compositions of differentially private algorithms. The paper uses the $f$-differential privacy formalism, which allows cleanly reducing the composition analysis to understanding the tail behavior of sums of independent privacy loss random variables. The original GDP paper of Dong, Roth, and Su used this idea to derive Berry-Esseen type bounds on the privacy loss parameters, as well as asymptotic limit theorems. This paper makes the natural next step (which was also explicitly suggested in the Dong, Roth, Su paper) of using the Edgeworth series approximation of the sum of independent random variables in order to get tighter error bounds.", "review_text": "This may be a nice work, but the write-up leaves a lot unclear about its significance and novelty. ", "strengths": "The problem studied by this paper - tight numeric bounds on the privacy parameters of differentially private algorithms - is an important one, and has received a lot of attention recently. The paper gives some evidence that Edgeworth series, applied to the f-differential privacy formalism, give some improvements in terms both of getting tighter bounds and of computational efficiency. \n\nThat said, I have the following concerns:\n* The paper claims computational complexity improvements, but leaves a lot unclear about this claim. \n    * Theoretically, what is the computational model in which the complexity of the algorithms is $O(m)$ for $m$-fold composition? Specifically, what assumptions are made with respect to how individual algorithms $M_i$ being composed are represented? The paper's supplement, when giving pseudocode for the algorithms, says things like \"Analytically encode all the corresponding PLLRs\". What does this mean? The authors should be precise about what kind of quantities/oracle access to $M_i$ is required, and how these assumptions correspond to the assumptions in prior work.\n    * No experimental evidence of better computational efficiency is presented, as far as I can tell. There is some evidence in the supplement of better numerical stability.\n\n* The experimental data presented leaves it unclear how significant the numerical improvements are. The experiments use some arbitrarily chosen (as far as I can tell) values of $\\delta$ and present bounds on $\\varepsilon$. The values of $\\delta$ chosen tend to be rather large: on the order 0.1 and 0.01 for 4 out of 6 plots. The two plots with lower value of $\\delta = 10^{-5}$ show that the improvement in the bounds on $\\varepsilon$ is rather small compared to the FFT based techniques. Even where the improvement is more significant in relative terms, it is still rather small in terms of the absolute value of $\\varepsilon$. \n\n* The paper is not always clear in terms of what is novel. See the next panel.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The goal of the paper is to present algorithms computing tight non-asymptotic numerical estimates on the privacy parameters of $m$-fold compositions of differentially private algorithms. The paper uses the $f$-differential privacy formalism, which allows cleanly reducing the composition analysis to understanding the tail behavior of sums of independent privacy loss random variables. The original GDP paper of Dong, Roth, and Su used this idea to derive Berry-Esseen type bounds on the privacy loss parameters, as well as asymptotic limit theorems. This paper makes the natural next step (which was also explicitly suggested in the Dong, Roth, Su paper) of using the Edgeworth series approximation of the sum of independent random variables in order to get tighter error bounds.", "strength_and_weaknesses": "The problem studied by this paper - tight numeric bounds on the privacy parameters of differentially private algorithms - is an important one, and has received a lot of attention recently. The paper gives some evidence that Edgeworth series, applied to the f-differential privacy formalism, give some improvements in terms both of getting tighter bounds and of computational efficiency. \n\nThat said, I have the following concerns:\n* The paper claims computational complexity improvements, but leaves a lot unclear about this claim. \n    * Theoretically, what is the computational model in which the complexity of the algorithms is $O(m)$ for $m$-fold composition? Specifically, what assumptions are made with respect to how individual algorithms $M_i$ being composed are represented? The paper's supplement, when giving pseudocode for the algorithms, says things like \"Analytically encode all the corresponding PLLRs\". What does this mean? The authors should be precise about what kind of quantities/oracle access to $M_i$ is required, and how these assumptions correspond to the assumptions in prior work.\n    * No experimental evidence of better computational efficiency is presented, as far as I can tell. There is some evidence in the supplement of better numerical stability.\n\n* The experimental data presented leaves it unclear how significant the numerical improvements are. The experiments use some arbitrarily chosen (as far as I can tell) values of $\\delta$ and present bounds on $\\varepsilon$. The values of $\\delta$ chosen tend to be rather large: on the order 0.1 and 0.01 for 4 out of 6 plots. The two plots with lower value of $\\delta = 10^{-5}$ show that the improvement in the bounds on $\\varepsilon$ is rather small compared to the FFT based techniques. Even where the improvement is more significant in relative terms, it is still rather small in terms of the absolute value of $\\varepsilon$. \n\n* The paper is not always clear in terms of what is novel. See the next panel.", "clarity,_quality,_novelty_and_reproducibility": "As mentioned, the paper is unclear about how individual algorithms being composed need to be represented. I guess one can try to trace that from the various bounds and statements, but it would be much more convenient for the reader to provide this information up front.\n\nIn terms of novelty, I am not sure how novel two of the main results are: Proposition 3.2 and, especially, Lemma 4.3. \n\n* Proposition 3.2 is similar in spirit to the approach in the Dong, Roth, and Su GDP paper. Is there a new idea here that did not appear in their paper, or is Proposition 3.2 a nice, user-friendly restatement of their approach?\n\n* Lemma 4.3, which is crucial for the \"finite sample\" bounds, appears identical to Theorem 2.1 in https://arxiv.org/abs/2101.05780. The paper says \"we follow the analysis on the finite-sample bound in Derumigny et al. (2021)\" which does not clarify whether they are just restating their result (in which case, why is a proof included in the appendix?) or whether they had to modify the Derumigny et al. analysis. If they had to modify it, then why and how?\n", "summary_of_the_review": "This may be a nice work, but the write-up leaves a lot unclear about its significance and novelty. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666629068660}, {"id": "nbd9B4GgD9u", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1958/Reviewer_Ri6D"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents Edgeworth Accountant (EA), an analytical approach to account the privacy loss of differentially private algorithms of multiple iterations (composition). The authors show that EA is more computationally efficient  than the Fast Fourier Transform approach and is more accurate than Renyi-DP accountant.", "review_text": "This paper provides a new tool for the privacy analysis of iterative DP algorithms. The authors claim that their method is more accurate than existing ones, and the computational cost is affordable. The theoretical analysis seems to be sound. However, there are several main concerns about this paper that need to be addressed: \n\n(1) The trade-off of EA v.s. FFT and RDP are not clear\n\n(2) Experimental settings and results\n\n(3) How EA performs in its target applications is not clear", "strengths": "Strength: Compared with existing approaches to account for DP guarantee, the proposed method provides a more accurate estimate while incurring lower computation cost. The theoretical analysis seems to be sound.\n\nWeakness: \n\nW1. The trade-off of EA vs. FFT and RDP are not clear. \n\n(1) Regarding the computation cost, how EA and FFT depend on the dimension of the noise is not covered. In particular, in table 1, only m (the number of compositions) is taken into account. \n\n(2) It is not clear why using RDP leads to a loose privacy accountant. Is it because of the conversion rule from RDP to (epsilon, delta)-DP, or some other limitations of RDP? It seems to me that RDP already gives a tight analysis for additive Gaussian noise, since there seems to be no slackness in the derivation in the original paper by Ilya Mironov. For the conversion from RDP to (epsilon, delta)-DP, Canone, Kamath, and Stenike propose an improved conversion rule in \"The Discrete Gaussian for Differential Privacy\" (see proposition 8); however, this improved rule is overlooked in the current paper. \n\n(3)  Does EA outperforms FFT and RDP when there is no subsampling? If subsampling is not the key factor, perhaps the author could move technicalities about subsampling to the appendix and focus more on EA itself.\n\n(4) Discussion on the effect of order for edgeworth approximation is also missing. How does the order influence the computation and accuracy of the estimate?\n\nW2. Unclear contributions. \n\nIt is not clear which propositions/lemmas/theorems in the draft are contributed by the author and which already exist in the literature. Maybe adding references could help.\n\nW3. Issues with experimental settings and results.\n\n(1) The author did not specify the size of the private dataset in their experiments at all. Delta=0.015 is too large for DP applications. \n\n(2) In Figure 3, why do we decrease the sampling probability p while increasing m?\n\n(3) It is unclear why the settings in Figures 2 and 3 are different? It seems that EEAI give the upper and lower bounds for AEA. It is better to put them in the same figures under the same settings.\n\n(4) What is the value of p (for Edgeworth approximation of order p) in the experiments? Is it the same as the subsampling rate p?\n\n(5) Please provide a figure for subsampling rate p=1 in the numerical experiments.\n\nW4. How EA performs in their target applications, noisy-SGD and Federated Analytics, is unclear.\n\n(1) No setting in the experiment section considers noisy-SGD. \n\n(2) The authors only show how their estimate for epsilon is more accurate, but do not provide any result on whether the more accurate estimate translates into better privacy-utility trade-off in the applications: noisy-SGD and Federated Analytics. \n\nW5. Overall, it is questionable if we should choose EA and f-DP to account for the privacy guarantee. As the authors have stated, GDP gives approximately the same result as EA, in the caption of Figure 3. In addition, FFT gives good enough upper and lower bounds for epsilon, as the gains in epsilon seem to be only marginal in Figures 2 and 3. How this more accurate estimate of privacy by EA translates into better privacy-utility trade-off is also not shown in the draft. In addition, the additional computation cost of FFT seems affordable since it is only a one time cost of O(sqrt(m)) and O(m^1.5). Should we choose EA over FFT?", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents Edgeworth Accountant (EA), an analytical approach to account the privacy loss of differentially private algorithms of multiple iterations (composition). The authors show that EA is more computationally efficient  than the Fast Fourier Transform approach and is more accurate than Renyi-DP accountant.", "strength_and_weaknesses": "Strength: Compared with existing approaches to account for DP guarantee, the proposed method provides a more accurate estimate while incurring lower computation cost. The theoretical analysis seems to be sound.\n\nWeakness: \n\nW1. The trade-off of EA vs. FFT and RDP are not clear. \n\n(1) Regarding the computation cost, how EA and FFT depend on the dimension of the noise is not covered. In particular, in table 1, only m (the number of compositions) is taken into account. \n\n(2) It is not clear why using RDP leads to a loose privacy accountant. Is it because of the conversion rule from RDP to (epsilon, delta)-DP, or some other limitations of RDP? It seems to me that RDP already gives a tight analysis for additive Gaussian noise, since there seems to be no slackness in the derivation in the original paper by Ilya Mironov. For the conversion from RDP to (epsilon, delta)-DP, Canone, Kamath, and Stenike propose an improved conversion rule in \"The Discrete Gaussian for Differential Privacy\" (see proposition 8); however, this improved rule is overlooked in the current paper. \n\n(3)  Does EA outperforms FFT and RDP when there is no subsampling? If subsampling is not the key factor, perhaps the author could move technicalities about subsampling to the appendix and focus more on EA itself.\n\n(4) Discussion on the effect of order for edgeworth approximation is also missing. How does the order influence the computation and accuracy of the estimate?\n\nW2. Unclear contributions. \n\nIt is not clear which propositions/lemmas/theorems in the draft are contributed by the author and which already exist in the literature. Maybe adding references could help.\n\nW3. Issues with experimental settings and results.\n\n(1) The author did not specify the size of the private dataset in their experiments at all. Delta=0.015 is too large for DP applications. \n\n(2) In Figure 3, why do we decrease the sampling probability p while increasing m?\n\n(3) It is unclear why the settings in Figures 2 and 3 are different? It seems that EEAI give the upper and lower bounds for AEA. It is better to put them in the same figures under the same settings.\n\n(4) What is the value of p (for Edgeworth approximation of order p) in the experiments? Is it the same as the subsampling rate p?\n\n(5) Please provide a figure for subsampling rate p=1 in the numerical experiments.\n\nW4. How EA performs in their target applications, noisy-SGD and Federated Analytics, is unclear.\n\n(1) No setting in the experiment section considers noisy-SGD. \n\n(2) The authors only show how their estimate for epsilon is more accurate, but do not provide any result on whether the more accurate estimate translates into better privacy-utility trade-off in the applications: noisy-SGD and Federated Analytics. \n\nW5. Overall, it is questionable if we should choose EA and f-DP to account for the privacy guarantee. As the authors have stated, GDP gives approximately the same result as EA, in the caption of Figure 3. In addition, FFT gives good enough upper and lower bounds for epsilon, as the gains in epsilon seem to be only marginal in Figures 2 and 3. How this more accurate estimate of privacy by EA translates into better privacy-utility trade-off is also not shown in the draft. In addition, the additional computation cost of FFT seems affordable since it is only a one time cost of O(sqrt(m)) and O(m^1.5). Should we choose EA over FFT?", "clarity,_quality,_novelty_and_reproducibility": "1. Clarity can be improved. The current draft is not self-contained, and is heavy in notation and lack of transition. There are also some concepts that require more detailed explanations, e.g., f-DP and edgeworth approximation. The settings for numerical experiments also need more details.\n\n2. Quality regarding experiments can be improved. See details above\n\n3. Novelty is unclear. Maybe better writing could clarify this issue. \n\n4. Reproducibility seems okay. But numerical stability for their implementation is not discussed.", "summary_of_the_review": "This paper provides a new tool for the privacy analysis of iterative DP algorithms. The authors claim that their method is more accurate than existing ones, and the computational cost is affordable. The theoretical analysis seems to be sound. However, there are several main concerns about this paper that need to be addressed: \n\n(1) The trade-off of EA v.s. FFT and RDP are not clear\n\n(2) Experimental settings and results\n\n(3) How EA performs in its target applications is not clear", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666619184696}], "openreview_url": "https://openreview.net/forum?id=2_BsVZ6R-ef", "arxiv_id": "2206.04236", "paper_pdf": "papers/2_BsVZ6R-ef.pdf", "paper_pdf_sha256": "e45932a3e10bf5dbb9b0605470d57980c2956ac13f84fe35762c8b1cede33e47", "paper_pdf_bytes": 394791, "paper_pdf_source": "openreview", "code_url": "https://github.com/HuaWang-wharton/EdgeworthAccountant", "code_repository": "HuaWang-wharton/EdgeworthAccountant", "code_commit": "6cd5a3eb973cfde5e734cb125119ef9ab26e0e00", "code_archive": "repos/2_BsVZ6R-ef.zip", "code_archive_sha256": "7c832955dcea2601a957eeb3684e3147df18af5e85e4621abcf493f063f38d95", "code_archive_bytes": 15139, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 15, "github_languages": {"Python": 30528}, "github_archived": false, "github_pushed_at": "2022-06-09T01:57:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/analytical-composition-of-differential"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JVsvIuMDE0Z", "year": 2022, "status": "rejected", "title": "Adaptive Behavior Cloning Regularization for Stable Offline-to-Online Reinforcement Learning", "authors": ["Yi Zhao", "Rinu Boney", "Alexander Ilin", "Juho Kannala", "Joni Pajarinen"], "authorids": ["~Yi_Zhao6", "~Rinu_Boney1", "~Alexander_Ilin1", "~Juho_Kannala1", "~Joni_Pajarinen2"], "authors_source": "OpenReview API", "abstract": "Offline reinforcement learning, by learning from a fixed dataset, makes it possible to learn agent behaviors without interacting with the environment. However, depending on the quality of the offline dataset, such pre-trained agents may have limited performance and would further need to be fine-tuned online by interacting with the environment. During online fine-tuning, the performance of the pre-trained agent may collapse quickly due to the sudden distribution shift from offline to online data. While constraints enforced by offline RL methods such as a behaviour cloning loss prevent this to an extent, these constraints also significantly slow down online fine-tuning by forcing the agent to stay close to the behavior policy. We propose to adaptively weigh the behavior cloning loss during online fine-tuning based on the agent's performance and training stability. Moreover, we use a randomized ensemble of Q functions to further increase the sample efficiency of online fine-tuning by performing a large number of learning updates. Experiments show that the proposed method yields state-of-the-art offline-to-online reinforcement learning performance on the popular D4RL benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tn0P7MYMyXj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3658/Reviewer_6n3w"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an update rule to adaptively change the regularizer term (BC here) weight during online learning for more stable fine tuning to improve over an expert.", "review_text": "Pros:\n\n** This paper considers an important problem setting for RL.\n\n** Paper is well written and is easy to follow.\n\nCons:\n\n** The proposed solution is heuristic and is similar to other methods that use annealing to tune the regularizer weight. And similar to other heuristics in this space, this method also has this limitation that it requires pre knowledge about the final performance of the task. I would like to see a more principled solution that at-least does not require such knowledge. \n\n** Evaluations are done on simple low dimensional tasks. Given the limited technical contribution, i would like to see a study on more complicated tasks with for example visual inputs where representation learning introduce additional challenges for finetuning.\n\n** The other claimed contribution on using ensemble of Q functions is orthogonal to the problem under study as it has shown before that ensemble of Q-functions helps to stabilize learning. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an update rule to adaptively change the regularizer term (BC here) weight during online learning for more stable fine tuning to improve over an expert.", "main_review": "Pros:\n\n** This paper considers an important problem setting for RL.\n\n** Paper is well written and is easy to follow.\n\nCons:\n\n** The proposed solution is heuristic and is similar to other methods that use annealing to tune the regularizer weight. And similar to other heuristics in this space, this method also has this limitation that it requires pre knowledge about the final performance of the task. I would like to see a more principled solution that at-least does not require such knowledge. \n\n** Evaluations are done on simple low dimensional tasks. Given the limited technical contribution, i would like to see a study on more complicated tasks with for example visual inputs where representation learning introduce additional challenges for finetuning.\n\n** The other claimed contribution on using ensemble of Q functions is orthogonal to the problem under study as it has shown before that ensemble of Q-functions helps to stabilize learning. ", "summary_of_the_review": "This paper considers an important topic for RL. However the proposed solution is heuristic and evaluations are not convincing given the simplicity of the tasks. Overall i think this paper does not offer enough contribution in its current form.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635859224876}, {"id": "xR8pUjsKONq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3658/Reviewer_q4CJ"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a new offline RL with online fine-tuning method. The authors first pretrain the policy using recent offline RL method TD3+BC with offline data and then collect on-policy data to further improve the pretrained policy. To prevent the policy from either degrading performance or failing to improve at the fine-tuning stage, the authors propose an automatic scheme to adjust the $\\alpha$ term that controls the BC loss in TD3+BC to ensure that the policy does not continue to contain itself to the behavior policy if there's room for improvement and also is able to stick to the previous policy if the performance is already near optimal. The authors conduct evalutions in D4RL mujoco environments and show that the approach is able to outperform prior methods in halfcheetah.", "review_text": "I think the paper is clearly written and easy to understand. The proposed algorithm is also novel in the sense that it gives an adaptive scheme of selecting the amount of conservatism during online training depending on the current policy performance. The empirical results also clearly show that the method can prevent the policy from degrading in medium-expert datasets where we don't need to change the current policy much, aligning with the initial motivation of the method.\n\nI also have a few concerns. First, while the results of the method generally align with the intuition that the policy should use large alpha on narrow datasets with good data and use large alpha on diverse datasets with low-quality data, the performance of the method on hopper and walker2d is a bit underwhelming. The method seems to be the same as hopper and also worse than Balanced Replay on three of the walker datasets. While the authors mentioned that the less satisfying results on walker2d are caused by the predefined target score in D4RL, I wonder if that is potentially a limitation of the method since it relies on the human-defined target score. Also, how would the method do if we change the target score to be higher (e.g. R_max * T)?\n\nMoreover, I think the paper lacks some theoretical understanding of the method. The intuition of the method makes perfect sense, but I would like to see how this adaptive scheme can give use some theoretical insights on policy improvement guarantees. Prior works [1] have studied this and I wonder how the proposed method can be connected to that.\n\nFinally, beyond the mujoco tasks, I think it would be good to evaluate the method on more realistic tasks such as the dexterous manipulation tasks used in the AWAC paper. \n\nMinor comment: it would be great to have an ablation study on the REDQ ensemble since that seems to be a bit disjoint to the theme of the paper.\n\n[1] Xie, Tengyang, et al. \"Policy Finetuning: Bridging Sample-Efficient Offline and Online Reinforcement Learning.\" arXiv preprint arXiv:2106.04895 (2021).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new offline RL with online fine-tuning method. The authors first pretrain the policy using recent offline RL method TD3+BC with offline data and then collect on-policy data to further improve the pretrained policy. To prevent the policy from either degrading performance or failing to improve at the fine-tuning stage, the authors propose an automatic scheme to adjust the $\\alpha$ term that controls the BC loss in TD3+BC to ensure that the policy does not continue to contain itself to the behavior policy if there's room for improvement and also is able to stick to the previous policy if the performance is already near optimal. The authors conduct evalutions in D4RL mujoco environments and show that the approach is able to outperform prior methods in halfcheetah.", "main_review": "I think the paper is clearly written and easy to understand. The proposed algorithm is also novel in the sense that it gives an adaptive scheme of selecting the amount of conservatism during online training depending on the current policy performance. The empirical results also clearly show that the method can prevent the policy from degrading in medium-expert datasets where we don't need to change the current policy much, aligning with the initial motivation of the method.\n\nI also have a few concerns. First, while the results of the method generally align with the intuition that the policy should use large alpha on narrow datasets with good data and use large alpha on diverse datasets with low-quality data, the performance of the method on hopper and walker2d is a bit underwhelming. The method seems to be the same as hopper and also worse than Balanced Replay on three of the walker datasets. While the authors mentioned that the less satisfying results on walker2d are caused by the predefined target score in D4RL, I wonder if that is potentially a limitation of the method since it relies on the human-defined target score. Also, how would the method do if we change the target score to be higher (e.g. R_max * T)?\n\nMoreover, I think the paper lacks some theoretical understanding of the method. The intuition of the method makes perfect sense, but I would like to see how this adaptive scheme can give use some theoretical insights on policy improvement guarantees. Prior works [1] have studied this and I wonder how the proposed method can be connected to that.\n\nFinally, beyond the mujoco tasks, I think it would be good to evaluate the method on more realistic tasks such as the dexterous manipulation tasks used in the AWAC paper. \n\nMinor comment: it would be great to have an ablation study on the REDQ ensemble since that seems to be a bit disjoint to the theme of the paper.\n\n[1] Xie, Tengyang, et al. \"Policy Finetuning: Bridging Sample-Efficient Offline and Online Reinforcement Learning.\" arXiv preprint arXiv:2106.04895 (2021).", "summary_of_the_review": "Given my comments in the above section, I think the paper could be improved if the authors can show the results of the method with a more realistic target score (e.g. R_max * T), provide the theoretical justifications, and evaluate the method on more realistic domains. I would vote for a weak reject given the current status.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635841753373}, {"id": "p6j7ZJYxFh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3658/Reviewer_aXQe"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the fine-tuning problem from offline to online RL. While the naive approach to fine-tune offline policy suffers from a sudden distributional shift by online samples and too much behavior constraint in offline algorithms. The proposed method leverage (1) the adaptive coefficient tuning in TD3+BC loss, (2) randomized ensembles of Q-functions (proposed by Chen et al. 2021), and (3) down-sampling of offline data. The experiments seem to show better or competitive results to other approaches.", "review_text": "## Strengths\n- This paper is well-written and easy to follow.\n- The focus of this paper, fine-tuning in RL seems to be a significant topic towards real-world applications, overcoming the bottleneck of current offline RL research.\n- The empirical comparisons seem to show better or competitive performance to AWAC, Balanced Replay, TD3_ft, and REDQ.\n\n## Weaknesses\n- The proposed method REDQ+AdaptiveBC seems to require a lot of hyper-parameter tuning. First, $\\alpha_{offline}=0.4$ is a very small value compared to original TD3+BC, which uses $\\alpha_{offline}=2.5$ after a grid search from \\{1.0, 2.0, 2.5, 3.0, 4.0\\}. This implies that $\\alpha_{offline}=0.4$ (and 100 for walker2d-random) might be a result of extensive hyper-parameter tuning. In addition, I guess $K_p$, $K_d$ (and even target return) must be well-balanced. I'm not sure how they are determined.\n- Also, REDQ+AdaptiveBC seems incremental on top of TD3+BC, combining the established Q-ensemble method (Chen et al. 2021) and carefully-designed coefficient tuning, and resembles Balanced Replay (Lee et al. 2021). I summarized the problems in fine-tuning and the remedies for them on the table. Both Balanced Replay and REDQ+AdaptiveBC overcome the data distribution shift issues using the ensembles of Q-functions. For imbalance issues of the offline-online replay buffer, Balanced Replay utilizes  Prioritized Experience Replay, and REDQ +AdaptiveBC downsamples the offline data, both of which aim to reduce the ratio of offline data and encourage to leverage the online samples when online fine-turning. During the online training phase, behavior regularization for stable offline training can prevent the improvement of the policy. Balanced Replay, proposed on top of CQL, removes the CQL loss and train value function and policy as online SAC, and REDQ+AdaptiveBC gradually reduce the ratio of BC loss. I think the novelty and contribution of this paper are not enough.\n\n|Problem|Balanced Replay (Lee et al. 2021) | REDQ+AdaptiveBC  |\n| ---- | ---- | ---- |\n|  **Distribution Shift**  |  Pessimistic Q-Ensemble  |  REDQ (Chen et al. 2021)  |\n|  **Imbalanced Replay Buffer** |  PER  |  Downsampling  |\n|  **Behavior Regularization**  | CQL-SAC switching |  Adaptive $\\alpha_{online}$  |\n\n- I think the comparison might be insufficient, since Balanced Replay, proposed on top of CQL, is based on TD3+BC in this paper. I don't think the comparison to CQL-based Balanced Replay isn't fair. In addition, while adaptive BC heavily depends on TD3+BC algorithms, Balanced Replay seems to work on both CQL and TD3+BC. I also think the authors can add TD3+BC baseline (keeping $\\alpha_{offline}$ during online training) and +REDQ/-REDQ ablations.\n\n### Minor Comments\n- I'm not sure how can I interpret the results of Figure 4. In my understanding, downsampling of replay buffer is introduced to overcome the instability of the beginning of the online training due to the imbalanced replay buffer (as discussed in Lee et al. 2021), but the caption says \"Downsampling does not hurt the training\". Doesn't downsampling improve the performance?\n- You can consider adding Ghasemipour et al. (2021), that proposes the ensemble-based offline RL algorithms, to the related work.\n- Prior works of offline-to-online fine-tuning (AWAC, Balanced Replay) work on robotic environments. Comparison on these environments, not only current MuJoCo locomotion, might be helpful.\n\n\n### Reference\nFujimoto and Gu. A Minimalist Approach to Offline Reinforcement Learning.  ArXiv preprint arXiv:2106.06860 (2021).\n\nLee et al. Offline-to-Online Reinforcement Learning via Balanced Replay and Pessimistic Q-Ensemble. Conference on Robot Learning (2021).\n\nChen et al. Randomized Ensembled Double Q-Learning: Learning Fast Without a Model. International Conference on Learning Representations (2021).\n\nGhasemipour et al. EMaQ: Expected-Max Q-Learning Operator for Simple Yet Effective Offline and Online RL. International Conference on Machine Learning (2021).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the fine-tuning problem from offline to online RL. While the naive approach to fine-tune offline policy suffers from a sudden distributional shift by online samples and too much behavior constraint in offline algorithms. The proposed method leverage (1) the adaptive coefficient tuning in TD3+BC loss, (2) randomized ensembles of Q-functions (proposed by Chen et al. 2021), and (3) down-sampling of offline data. The experiments seem to show better or competitive results to other approaches.", "main_review": "## Strengths\n- This paper is well-written and easy to follow.\n- The focus of this paper, fine-tuning in RL seems to be a significant topic towards real-world applications, overcoming the bottleneck of current offline RL research.\n- The empirical comparisons seem to show better or competitive performance to AWAC, Balanced Replay, TD3_ft, and REDQ.\n\n## Weaknesses\n- The proposed method REDQ+AdaptiveBC seems to require a lot of hyper-parameter tuning. First, $\\alpha_{offline}=0.4$ is a very small value compared to original TD3+BC, which uses $\\alpha_{offline}=2.5$ after a grid search from \\{1.0, 2.0, 2.5, 3.0, 4.0\\}. This implies that $\\alpha_{offline}=0.4$ (and 100 for walker2d-random) might be a result of extensive hyper-parameter tuning. In addition, I guess $K_p$, $K_d$ (and even target return) must be well-balanced. I'm not sure how they are determined.\n- Also, REDQ+AdaptiveBC seems incremental on top of TD3+BC, combining the established Q-ensemble method (Chen et al. 2021) and carefully-designed coefficient tuning, and resembles Balanced Replay (Lee et al. 2021). I summarized the problems in fine-tuning and the remedies for them on the table. Both Balanced Replay and REDQ+AdaptiveBC overcome the data distribution shift issues using the ensembles of Q-functions. For imbalance issues of the offline-online replay buffer, Balanced Replay utilizes  Prioritized Experience Replay, and REDQ +AdaptiveBC downsamples the offline data, both of which aim to reduce the ratio of offline data and encourage to leverage the online samples when online fine-turning. During the online training phase, behavior regularization for stable offline training can prevent the improvement of the policy. Balanced Replay, proposed on top of CQL, removes the CQL loss and train value function and policy as online SAC, and REDQ+AdaptiveBC gradually reduce the ratio of BC loss. I think the novelty and contribution of this paper are not enough.\n\n|Problem|Balanced Replay (Lee et al. 2021) | REDQ+AdaptiveBC  |\n| ---- | ---- | ---- |\n|  **Distribution Shift**  |  Pessimistic Q-Ensemble  |  REDQ (Chen et al. 2021)  |\n|  **Imbalanced Replay Buffer** |  PER  |  Downsampling  |\n|  **Behavior Regularization**  | CQL-SAC switching |  Adaptive $\\alpha_{online}$  |\n\n- I think the comparison might be insufficient, since Balanced Replay, proposed on top of CQL, is based on TD3+BC in this paper. I don't think the comparison to CQL-based Balanced Replay isn't fair. In addition, while adaptive BC heavily depends on TD3+BC algorithms, Balanced Replay seems to work on both CQL and TD3+BC. I also think the authors can add TD3+BC baseline (keeping $\\alpha_{offline}$ during online training) and +REDQ/-REDQ ablations.\n\n### Minor Comments\n- I'm not sure how can I interpret the results of Figure 4. In my understanding, downsampling of replay buffer is introduced to overcome the instability of the beginning of the online training due to the imbalanced replay buffer (as discussed in Lee et al. 2021), but the caption says \"Downsampling does not hurt the training\". Doesn't downsampling improve the performance?\n- You can consider adding Ghasemipour et al. (2021), that proposes the ensemble-based offline RL algorithms, to the related work.\n- Prior works of offline-to-online fine-tuning (AWAC, Balanced Replay) work on robotic environments. Comparison on these environments, not only current MuJoCo locomotion, might be helpful.\n\n\n### Reference\nFujimoto and Gu. A Minimalist Approach to Offline Reinforcement Learning.  ArXiv preprint arXiv:2106.06860 (2021).\n\nLee et al. Offline-to-Online Reinforcement Learning via Balanced Replay and Pessimistic Q-Ensemble. Conference on Robot Learning (2021).\n\nChen et al. Randomized Ensembled Double Q-Learning: Learning Fast Without a Model. International Conference on Learning Representations (2021).\n\nGhasemipour et al. EMaQ: Expected-Max Q-Learning Operator for Simple Yet Effective Offline and Online RL. International Conference on Machine Learning (2021).\n", "summary_of_the_review": "As I described in Main Review, this paper seems to require extensive hyper-parameter search (proper $\\alpha_{offline}$ and well-balanced $K_p$, $K_d$), and to propose the incremental algorithms on top of TD3+BC (and resemble Lee et al. 2021), combining Q-ensemble method proposed by Chen et al. (2021), downsampling for imbalanced replay buffer, and adaptive $\\alpha_{online}$ tuning. Due to the lack of novelty and significant contributions to the ICLR (or RL) community, I vote for rejection.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635518665022}, {"id": "PJEI2lBTwYX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3658/Reviewer_bnwi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors combine REDQ and TD3+BC and propose an adaptive mechanism to update $\\alpha$ during the online fine-tuning. The proposed method decreases $\\alpha$ for further improvement and increases  $\\alpha$  to avoid performance collapse.", "review_text": "The paper is well organized and the motivation is really clear. Offline-to-online learning considered in this paper is an important problem. I agree that adaptively updating the behavior policy constraint is the key point in online fine-tuning. In the experiments, the authors compare the two SOTA offline-to-online methods AWAC and Balanced Replay, and ablation studies show the effectiveness of adaptive $\\alpha$.\n\nMy main concern is the novelty. REDQ and TD3+BC are prior work, and it seems that the main contribution in this paper is the adaptive mechanism of $\\alpha$. Although this paper tries to solve an important problem, the contribution of the proposed method is limited.\n\nThe update rule of $\\alpha$ is heuristic. Although it is well-motivated, more theoretical analysis is expected to demonstrate how the adaptive  $\\alpha$ achieves the balance between performance improvement and performance dropping. And the selection of hyperparameters KP and KD is arbitrary. More empirical results are needed to show how the two hyperparameters affect the update of $\\alpha$​. In Figure 3, AdaptiveBC does not significantly outperform ManualBC. Although ManualBC uses grid search, the search space is not too large. \n\nIn Figure 2, if no expert data is involved, REDQ (scratch) could achieve a good performance in 250000 timesteps, which shows pure online learning is enough to deal with this setting. In that case, it is important to consider the sample efficiency and deployment efficiency[1] in offline-to-online learning.\n\n[1]https://arxiv.org/abs/2006.03647", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors combine REDQ and TD3+BC and propose an adaptive mechanism to update $\\alpha$ during the online fine-tuning. The proposed method decreases $\\alpha$ for further improvement and increases  $\\alpha$  to avoid performance collapse.", "main_review": "The paper is well organized and the motivation is really clear. Offline-to-online learning considered in this paper is an important problem. I agree that adaptively updating the behavior policy constraint is the key point in online fine-tuning. In the experiments, the authors compare the two SOTA offline-to-online methods AWAC and Balanced Replay, and ablation studies show the effectiveness of adaptive $\\alpha$.\n\nMy main concern is the novelty. REDQ and TD3+BC are prior work, and it seems that the main contribution in this paper is the adaptive mechanism of $\\alpha$. Although this paper tries to solve an important problem, the contribution of the proposed method is limited.\n\nThe update rule of $\\alpha$ is heuristic. Although it is well-motivated, more theoretical analysis is expected to demonstrate how the adaptive  $\\alpha$ achieves the balance between performance improvement and performance dropping. And the selection of hyperparameters KP and KD is arbitrary. More empirical results are needed to show how the two hyperparameters affect the update of $\\alpha$​. In Figure 3, AdaptiveBC does not significantly outperform ManualBC. Although ManualBC uses grid search, the search space is not too large. \n\nIn Figure 2, if no expert data is involved, REDQ (scratch) could achieve a good performance in 250000 timesteps, which shows pure online learning is enough to deal with this setting. In that case, it is important to consider the sample efficiency and deployment efficiency[1] in offline-to-online learning.\n\n[1]https://arxiv.org/abs/2006.03647", "summary_of_the_review": "The paper aims to solve an important problem: adaptive behavior policy constraint in offline-to-online learning. However, the proposed method is heuristic, and more theoretical analysis is expected.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1634483147321}], "openreview_url": "https://openreview.net/forum?id=JVsvIuMDE0Z", "arxiv_id": "2210.13846", "paper_pdf": "papers/JVsvIuMDE0Z.pdf", "paper_pdf_sha256": "70f0482bd5c20f441a0698260c63a09406237927d910e6f7782cfc896f2f46e2", "paper_pdf_bytes": 437231, "paper_pdf_source": "openreview", "code_url": "https://github.com/zhaoyi11/adaptive_bc", "code_repository": "zhaoyi11/adaptive_bc", "code_commit": "d47226d4ffc9959ac0ed1337ddd7ebb787fc05e1", "code_archive": "repos/JVsvIuMDE0Z.zip", "code_archive_sha256": "5e9adbc418534e878989df1a2cbdb26b8a0ea69264aac3dd40ff236757d8de5f", "code_archive_bytes": 13306, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 14, "github_languages": {"Python": 27310}, "github_archived": false, "github_pushed_at": "2022-07-04T08:28:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-behavior-cloning-regularization-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "crAi7c41xTh", "year": 2021, "status": "rejected", "title": "Shape Matters: Understanding the Implicit Bias of the Noise Covariance", "authors": ["Jeff Z. HaoChen", "Colin Wei", "Jason D. Lee", "Tengyu Ma"], "authorids": ["~Jeff_Z._HaoChen1", "~Colin_Wei1", "~Jason_D._Lee1", "~Tengyu_Ma1"], "authors_source": "OpenReview API", "abstract": "The noise in stochastic gradient descent (SGD) provides a crucial implicit regularization effect for training overparameterized models. Prior theoretical work largely focuses on spherical Gaussian noise, whereas empirical studies demonstrate the phenomenon that parameter-dependent noise --- induced by mini-batches or label perturbation --- is far more effective than Gaussian noise. \nThis paper theoretically characterizes this phenomenon on a quadratically-parameterized model introduced by Vaskevicius et al. and Woodworth et al.  We show that in an over-parameterized setting, SGD with label noise recovers the sparse ground-truth with an arbitrary initialization, whereas SGD with Gaussian noise or gradient descent overfits to dense solutions with large norms. Our analysis reveals that parameter-dependent noise introduces a bias towards local minima with smaller noise variance, whereas spherical Gaussian noise does not. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "GvpjwkH8wNq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1062/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Problem\n\nThis paper considers the effect of label noise on stochastic gradient descent. The setup is that there is a vector $v \\in R^d$. We observe samples from $v^2\\cdot x$. We only have $n < d$ samples but $v$ is $r$-sparse for $r < n < d$ which makes recovery possible information theoretically. The main result is that stochastic gradient descent with label noise, and without any explicit regularization will recover the ground truth. whereas adding spherical Gaussian noise does not.\n\n### Pros and Cons\n\nThe problem is a clean toy problem with which to illustrate the gap between algorithms. It shows a clean separation between the power of label noise and that of random Gaussian noise. The model appears to be the simplest model where one can hope to see the regularization effects of noise (the simpler linear regression model wouldnt show these effects).  One possible criticism could be to ask if understanding this model is truly getting us closer to understanding what happens in deep nets. At this point it is hard to say, but proving such a result even in this simple model is not trivial, and is definitely a contribution.\n\n\n### Evaluation\nI think this is a solid theoretical contribution on an important problem, and the paper should be accepted.\n\n\n### Further comments\n\nI was a little confused by the comment that the coefficients are assumed to be in ${0,1}$ since they then satisfy $v_i^2 = v_i$ as this seems to linearize the model. The authors should probably clarify that this is actually not what is going on. It might be better to use a different setting of parameters even for exposition. \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Weak accept", "review": "### Problem\n\nThis paper considers the effect of label noise on stochastic gradient descent. The setup is that there is a vector $v \\in R^d$. We observe samples from $v^2\\cdot x$. We only have $n < d$ samples but $v$ is $r$-sparse for $r < n < d$ which makes recovery possible information theoretically. The main result is that stochastic gradient descent with label noise, and without any explicit regularization will recover the ground truth. whereas adding spherical Gaussian noise does not.\n\n### Pros and Cons\n\nThe problem is a clean toy problem with which to illustrate the gap between algorithms. It shows a clean separation between the power of label noise and that of random Gaussian noise. The model appears to be the simplest model where one can hope to see the regularization effects of noise (the simpler linear regression model wouldnt show these effects).  One possible criticism could be to ask if understanding this model is truly getting us closer to understanding what happens in deep nets. At this point it is hard to say, but proving such a result even in this simple model is not trivial, and is definitely a contribution.\n\n\n### Evaluation\nI think this is a solid theoretical contribution on an important problem, and the paper should be accepted.\n\n\n### Further comments\n\nI was a little confused by the comment that the coefficients are assumed to be in ${0,1}$ since they then satisfy $v_i^2 = v_i$ as this seems to linearize the model. The authors should probably clarify that this is actually not what is going on. It might be better to use a different setting of parameters even for exposition. \n\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604103738379}, {"id": "L9Y_gpxuSFx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1062/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper demonstrates that for a particular model SGD with label noise and proper learning rate schedule recovers the (sparse) data generating model while GD with or without Gaussian noise does not. In the latter case, it fails because a stationary distribution is not achieved. The proofs in the appendix are quite involved, and as a result I did not carefully study them. But the authors provide helpful intuition for the results in the main text. \n\nWhile I think there is value in the work, I am not sure whether the fairly specific setting studied has much to say about neural networks. (Of course, not every paper needs to be about neural networks, but that's certainly the motivation of the paper.) There are a number of aspects to the work that limit its generality: the model, the label noise, the objective of reconstruction of sparse ground truth from the same model class, and the dataset modeling. The authors justify the model by saying that other works that have studied it, but don't otherwise try to justify its relevance. In Figure 1, the authors argue that label noise behaves similar to SGD, but I don't find this thorough or convincing enough that any results on label noise and the mechanism by which it operates should generalize to SGD. Also, whether studying an objective of learning as identifying the sparse ground truth model seems far from standard training of a neural network. As an example of my concern that the specifics by which the results are achieved may not apply in other scenarios, do the authors believe that GD with Gaussian noise fails for neural networks because a stationary distribution is never achieved, whereas it is achieved for SGD?\n\nOverall, I think that the work is a bit too narrow and doesn't change our understanding of what non-spherical noise from SGD does to neural networks beyond what is already known.\n\n\nMinor presentation points:\n\nThe paper became rushed at the end, as if the authors ran out of space. Subsections 3.2 \"Stage 0\" and 3.3 \"Stage 1\" are then followed by a very short paragraph inside 3.3 titled \"Stage 2\". \n\nFigure 1 is fairly difficult to parse with the number of noisy curves overlapping each other. Perhaps the authors could make their point with the minimal amount of experimental data here and relegate the rest to the appendix.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Rigorous results but on a limited problem", "review": "This paper demonstrates that for a particular model SGD with label noise and proper learning rate schedule recovers the (sparse) data generating model while GD with or without Gaussian noise does not. In the latter case, it fails because a stationary distribution is not achieved. The proofs in the appendix are quite involved, and as a result I did not carefully study them. But the authors provide helpful intuition for the results in the main text. \n\nWhile I think there is value in the work, I am not sure whether the fairly specific setting studied has much to say about neural networks. (Of course, not every paper needs to be about neural networks, but that's certainly the motivation of the paper.) There are a number of aspects to the work that limit its generality: the model, the label noise, the objective of reconstruction of sparse ground truth from the same model class, and the dataset modeling. The authors justify the model by saying that other works that have studied it, but don't otherwise try to justify its relevance. In Figure 1, the authors argue that label noise behaves similar to SGD, but I don't find this thorough or convincing enough that any results on label noise and the mechanism by which it operates should generalize to SGD. Also, whether studying an objective of learning as identifying the sparse ground truth model seems far from standard training of a neural network. As an example of my concern that the specifics by which the results are achieved may not apply in other scenarios, do the authors believe that GD with Gaussian noise fails for neural networks because a stationary distribution is never achieved, whereas it is achieved for SGD?\n\nOverall, I think that the work is a bit too narrow and doesn't change our understanding of what non-spherical noise from SGD does to neural networks beyond what is already known.\n\n\nMinor presentation points:\n\nThe paper became rushed at the end, as if the authors ran out of space. Subsections 3.2 \"Stage 0\" and 3.3 \"Stage 1\" are then followed by a very short paragraph inside 3.3 titled \"Stage 2\". \n\nFigure 1 is fairly difficult to parse with the number of noisy curves overlapping each other. Perhaps the authors could make their point with the minimal amount of experimental data here and relegate the rest to the appendix.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604032349505}, {"id": "klJv4DqW8t7", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1062/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors study the problem quadratically parameterized (linear) regression and study the behavior of SGD to solve it when the stochasticity added is in terms of label noise. They show that SGD (with this kind of noise) with arbitrarily large initialization converges to the ground truth solution, whereas there exist settings where Langevin dynamics or gradient descent would fail to converge to this solution. The proofs are carried out carefully and are correct as best as I could verify. The authors also provide experiments on synthetic data to support and motivate their theory.\n\n\nSome questions/comments:\n1. Why are three stages with decreasing step-sizes needed for the analysis? Can we expect a similar result to hold if a constant, but, small step-size is chosen instead?\n2. The generative model assumes that the label y has no added noise. Are the theoretical result robust to additive noise?\n3. The authors point out on page 4 that the sample complexity is worse that LASSO (which is fine), but, do not remark why this is the case. It would be insightful if the authors could add a comment about why this is the case, and if this was experimentally observed as well.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "the authors study an interesting problem and provide sound theoretical analysis", "review": "The authors study the problem quadratically parameterized (linear) regression and study the behavior of SGD to solve it when the stochasticity added is in terms of label noise. They show that SGD (with this kind of noise) with arbitrarily large initialization converges to the ground truth solution, whereas there exist settings where Langevin dynamics or gradient descent would fail to converge to this solution. The proofs are carried out carefully and are correct as best as I could verify. The authors also provide experiments on synthetic data to support and motivate their theory.\n\n\nSome questions/comments:\n1. Why are three stages with decreasing step-sizes needed for the analysis? Can we expect a similar result to hold if a constant, but, small step-size is chosen instead?\n2. The generative model assumes that the label y has no added noise. Are the theoretical result robust to additive noise?\n3. The authors point out on page 4 that the sample complexity is worse that LASSO (which is fine), but, do not remark why this is the case. It would be insightful if the authors could add a comment about why this is the case, and if this was experimentally observed as well.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603952318016}, {"id": "MNr1-ET5uw1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1062/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the implicit regularization of stochastic gradient decent (SGD). The authors analyze SGD with label nose in the quadratically-parameterized model and prove that it converges to the sparse ground-truth even if started with large initialization. The authors also prove that SGD with Gaussian noise (Langevin dynamics) does not converge to the ground truth at zero under the overparameterized regime.\n\nThis is a solid contribution with theoretical insights on SGD with label noise. While the theoretical results are deep with long proofs, their outlines and meanings are well explained.\n\n- The difference (affinity) between the deep neural network and the quadratically-parameterized model is mainly discussed numerically (Figure 1). It would be nicer to discuss gaps between them theoretically, if possible. For example, the contraction of SGD in the initial phase (Thm 3.1) is reminiscent of the effect of singularities in neural networks discussed in references such as:\nGuo et al.: Numerical Analysis near Singularities in RBF networks, JMLR, 19(2018), 1-39.\nIt would be nicer to discuss if this analysis of the initial phase has something to do with singularities.\n\n- p.7, footnote 7: What is the second-order effect (of zero-mean noise)?\n\nMinor:\np.4, Langevin dynamics/diffusion: The last sentence is duplicated.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Solid contribution with theoretical insights on SGD with label noise", "review": "This paper considers the implicit regularization of stochastic gradient decent (SGD). The authors analyze SGD with label nose in the quadratically-parameterized model and prove that it converges to the sparse ground-truth even if started with large initialization. The authors also prove that SGD with Gaussian noise (Langevin dynamics) does not converge to the ground truth at zero under the overparameterized regime.\n\nThis is a solid contribution with theoretical insights on SGD with label noise. While the theoretical results are deep with long proofs, their outlines and meanings are well explained.\n\n- The difference (affinity) between the deep neural network and the quadratically-parameterized model is mainly discussed numerically (Figure 1). It would be nicer to discuss gaps between them theoretically, if possible. For example, the contraction of SGD in the initial phase (Thm 3.1) is reminiscent of the effect of singularities in neural networks discussed in references such as:\nGuo et al.: Numerical Analysis near Singularities in RBF networks, JMLR, 19(2018), 1-39.\nIt would be nicer to discuss if this analysis of the initial phase has something to do with singularities.\n\n- p.7, footnote 7: What is the second-order effect (of zero-mean noise)?\n\nMinor:\np.4, Langevin dynamics/diffusion: The last sentence is duplicated.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603872919812}], "openreview_url": "https://openreview.net/forum?id=crAi7c41xTh", "arxiv_id": "2006.08680", "paper_pdf": "papers/crAi7c41xTh.pdf", "paper_pdf_sha256": "fd62ac31d32f0bc4744b3ce6ba516caa8054d872b808125631bbee0a69774fe3", "paper_pdf_bytes": 448050, "paper_pdf_source": "openreview", "code_url": "https://github.com/jhaochenz96/noise-implicit-bias", "code_repository": "jhaochenz96/noise-implicit-bias", "code_commit": "2b94db1672edf916851cd806c64c246000e2944e", "code_archive": "repos/crAi7c41xTh.zip", "code_archive_sha256": "ccf4a40dd299a9634fb0e5c21e326de86f0c002d39dbf6b01d5de3b498b436f7", "code_archive_bytes": 11874, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 15, "github_languages": {"Python": 27244}, "github_archived": false, "github_pushed_at": "2020-06-17T04:37:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/shape-matters-understanding-the-implicit-bias"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJlzxgBtwH", "year": 2020, "status": "rejected", "title": "Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack", "authors": ["Francesco Croce", "Matthias Hein"], "authorids": ["francesco91.croce@gmail.com", "matthias.hein@uni-tuebingen.de"], "authors_source": "OpenReview API", "abstract": "The evaluation of robustness against adversarial manipulations of neural networks-based classifiers is mainly tested with empirical attacks as the methods for the exact computation, even when available, do not scale to large networks. We propose in this paper a new white-box adversarial attack wrt the $l_p$-norms for $p \\in \\{1,2,\\infty\\}$ aiming at finding the minimal perturbation necessary to change the class of a given input. It has an intuitive geometric meaning, yields quickly high quality results, minimizes the size of the perturbation (so that it returns the robust accuracy at every threshold with a single run). It performs better or similarly to state-of-the-art attacks which are partially specialized to one $l_p$-norm.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJgiUscAFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2090/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Authors extend deepFool by adding extra steps and constraints to find closer points to the source image as the adversarial image. They both project onto the decision boundary. Deepfool does and adhoc clipping to keep the pixel values in (0,1) but the new proposed method respects the constraints during the steps. Also during the steps they combine projection of last step result and original image to keep it closer to the original image. Moreover, at the end of the optimization they perform extra search steps to get closer to the original image. Also they add random restarts. Rather than considering the original image, they randomly choose an image in the half ballpark of the total delta.\n\nAccording to the results in fig.2 the backward steps has the highest impact in comparison to deepfool. But mixing with original projection always helps a little and random restarts help a little too. Without the backward steps there is almost no gain from mixing the projections.\n\nConsidering the full results in the appendix, the results are mixed with no obvious advantage in comparison to PGD specially.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Authors extend deepFool by adding extra steps and constraints to find closer points to the source image as the adversarial image. They both project onto the decision boundary. Deepfool does and adhoc clipping to keep the pixel values in (0,1) but the new proposed method respects the constraints during the steps. Also during the steps they combine projection of last step result and original image to keep it closer to the original image. Moreover, at the end of the optimization they perform extra search steps to get closer to the original image. Also they add random restarts. Rather than considering the original image, they randomly choose an image in the half ballpark of the total delta.\n\nAccording to the results in fig.2 the backward steps has the highest impact in comparison to deepfool. But mixing with original projection always helps a little and random restarts help a little too. Without the backward steps there is almost no gain from mixing the projections.\n\nConsidering the full results in the appendix, the results are mixed with no obvious advantage in comparison to PGD specially."}, "tcdate": 1571887955243}, {"id": "Hkgp5zksFS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2090/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a new gradient-based method (FAB) for constructing adversarial perturbations for deep neural networks. At a high level, the method repeatedly estimates the decision boundary based on the linearization of the classifier at a given point and projects to the closest \"misclassified\" example based on that estimation (similar to DeepFool). The authors build on this idea, proposing several improvements and evaluate their attack empirically against a variety of models.\n\nI found the proposed method quite interesting and intuitive. All the improvements made to the core method are well-motivated and clearly explained, while the ablation experiments are relatively thorough.\n\nHowever, I did find the presentation of experimental evidence quite misleading. \n\nSpecifically, reporting mean accuracy over models, datasets, and epsilon constraints in Table 2 does not give the full picture. Going through the appendix tables, we can see the following:\n-- The step size used for PGD is quite large---eps/4 for the L2 case---which is quite uncommon when using 150 iterations. Based on prior work and my own personal experience, a step size of 2 * eps / #steps (i.e., eps / 75) would seem more suitable. I wonder if this is the reason for PGD performing worse than FAB for large epsilon values on CIFAR10. The authors mention that they chose this parameter using grid search but do not provide concrete details.\n-- The adversarially trained MNIST model of Madry et al. 2018 learns to use thresholding filters as the first layer (observed in the original paper). This causes issues for most gradient-based methods (e.g., PGD performs worse than the decision-based attack of Brendel et al. 2018, also observed in other prior work). While it is encouraging that FAB is robust to such gradient obfuscation, this is arguably not the ideal setting to compare gradient based methods (especially when averaging performance over models). \n-- For MNIST and Restricted IN, PGD performs comparably or even better than FAB (modulo larger epsilon values for which the large step size used could be an issue for PGD and the Linf-trained model with the thresholding filters).\n-- For the L1-norm setting, EAD performs similarly or better compared to FAB (again modulo the Linf-trained model).\nBased on these observations, I am not fully convinced that FAB outperforms PGD (for L2 and Linf) and EAD (for L1) by as much as Table 2 suggests.\n\nMoreover, the runtime comparison performed in not exactly fair:\n-- It is not clear how many restarts where included in the runtime of PGD. Its runtime should be in the same ballpark as FAB but the time reported is ~20x higher. \n-- PGD is known to produce quite accurate estimates when run with much fewer (say 15) steps. Thus in order to make a fair comparison one would also need to look at the entire #steps vs robust accuracy curve to get a better picture of the efficiency of these two methods. Choosing an arbitrary number of steps for each method is not very enlightening.\n-- It is not necessary to run PGD 5 times to evaluate the robust accuracy at 5 thresholds. One can perform binary search for each input in order to find the smallest epsilon for which a misclassification can be found. This will result in at most 3 (sometimes 2) evaluations per point (instead of 5).\n\nDespite these shortcomings of the experimental evaluation, I still believe that the paper has merit. After all, the method is clean and well-motivated,  performs comparably to the best of PGD and EAD in a variety of settings, and is robust to a certain degree of gradient masking. In that sense, it could potentially be a valuable contribution and could be of interest to a subset of the adversarial ML community.\n\nIn the sense, while my initial stance is to recommend (weak) rejection, I would be open to increasing my score and recommending (weak) acceptance should my concerns be addressed.\n\nUPDATE: I appreciate the response and the additional experiments performed by the authors. The authors have addressed my concerns in their response. I am increasing my score to a weak accept.\n\nOne thing that would be nice to add in the next version of the manuscript is a note inviting the reader to consider the appendix tables since average robust accuracy can be inconclusive.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "The authors propose a new gradient-based method (FAB) for constructing adversarial perturbations for deep neural networks. At a high level, the method repeatedly estimates the decision boundary based on the linearization of the classifier at a given point and projects to the closest \"misclassified\" example based on that estimation (similar to DeepFool). The authors build on this idea, proposing several improvements and evaluate their attack empirically against a variety of models.\n\nI found the proposed method quite interesting and intuitive. All the improvements made to the core method are well-motivated and clearly explained, while the ablation experiments are relatively thorough.\n\nHowever, I did find the presentation of experimental evidence quite misleading. \n\nSpecifically, reporting mean accuracy over models, datasets, and epsilon constraints in Table 2 does not give the full picture. Going through the appendix tables, we can see the following:\n-- The step size used for PGD is quite large---eps/4 for the L2 case---which is quite uncommon when using 150 iterations. Based on prior work and my own personal experience, a step size of 2 * eps / #steps (i.e., eps / 75) would seem more suitable. I wonder if this is the reason for PGD performing worse than FAB for large epsilon values on CIFAR10. The authors mention that they chose this parameter using grid search but do not provide concrete details.\n-- The adversarially trained MNIST model of Madry et al. 2018 learns to use thresholding filters as the first layer (observed in the original paper). This causes issues for most gradient-based methods (e.g., PGD performs worse than the decision-based attack of Brendel et al. 2018, also observed in other prior work). While it is encouraging that FAB is robust to such gradient obfuscation, this is arguably not the ideal setting to compare gradient based methods (especially when averaging performance over models). \n-- For MNIST and Restricted IN, PGD performs comparably or even better than FAB (modulo larger epsilon values for which the large step size used could be an issue for PGD and the Linf-trained model with the thresholding filters).\n-- For the L1-norm setting, EAD performs similarly or better compared to FAB (again modulo the Linf-trained model).\nBased on these observations, I am not fully convinced that FAB outperforms PGD (for L2 and Linf) and EAD (for L1) by as much as Table 2 suggests.\n\nMoreover, the runtime comparison performed in not exactly fair:\n-- It is not clear how many restarts where included in the runtime of PGD. Its runtime should be in the same ballpark as FAB but the time reported is ~20x higher. \n-- PGD is known to produce quite accurate estimates when run with much fewer (say 15) steps. Thus in order to make a fair comparison one would also need to look at the entire #steps vs robust accuracy curve to get a better picture of the efficiency of these two methods. Choosing an arbitrary number of steps for each method is not very enlightening.\n-- It is not necessary to run PGD 5 times to evaluate the robust accuracy at 5 thresholds. One can perform binary search for each input in order to find the smallest epsilon for which a misclassification can be found. This will result in at most 3 (sometimes 2) evaluations per point (instead of 5).\n\nDespite these shortcomings of the experimental evaluation, I still believe that the paper has merit. After all, the method is clean and well-motivated,  performs comparably to the best of PGD and EAD in a variety of settings, and is robust to a certain degree of gradient masking. In that sense, it could potentially be a valuable contribution and could be of interest to a subset of the adversarial ML community.\n\nIn the sense, while my initial stance is to recommend (weak) rejection, I would be open to increasing my score and recommending (weak) acceptance should my concerns be addressed.\n\nUPDATE: I appreciate the response and the additional experiments performed by the authors. The authors have addressed my concerns in their response. I am increasing my score to a weak accept.\n\nOne thing that would be nice to add in the next version of the manuscript is a note inviting the reader to consider the appendix tables since average robust accuracy can be inconclusive.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571644052746}, {"id": "BJeCBkS5Kr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2090/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies the problem of the white-box attack of neural network-based classifiers, with an emphasis on the \"minimal distortion solution\": The new input that changes the labeling output of the network with the minimal distance (l1, l2, l_inf) with respect to a given input. \n\nThe main intuition of the algorithm is to do a local linear approximation of the network at the current point (which is the Taylor expansion up to the gradient term). After that, the algorithm identifies a class (output coordinate) with the minimal \"margin to gradient norm ratio\", i.e. the total movement in gradient direction to change the labeling function in that coordinate, within this linear approximation. The algorithm solves the subproblem of minimizing a linear function inside lp ball as the critical routine.\n\nOverall, the notion of finding the minimal distortion attacker as opposed to finding the best attacker inside a fixed distortion ball is quite interesting to me. The main concern for me about this paper is the comparison to other methods such as PGD. As far as I know, these attackers DO NOT explicitly minimize the distortion, thus it is quite believable that these models do not identify the minimal distortion solution (rather it will more likely to find a solution that lies in the boundary since it would be the easiest way to attack). However, for the proposed algorithm in this paper, the algorithm is explicitly minimizing the distance to the given input (x_orig in their language). \n\n\nI would like to see more implementation details of the other algorithms, for example, what is the performance if we add an additional regularizer as the distance of the current attacker to the given input to PGD. So far, the paper lacks solid proof of the usefulness of this particular algorithm. (In particular the justification for solving the local linear system instead of doing a gradient descent step).\n\nAfter Rebuttal: I have read the authors' responses and acknowledge the sensibility of the statement. I apologize for the earlier misunderstanding and higher the score accordingly.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "The paper studies the problem of the white-box attack of neural network-based classifiers, with an emphasis on the \"minimal distortion solution\": The new input that changes the labeling output of the network with the minimal distance (l1, l2, l_inf) with respect to a given input. \n\nThe main intuition of the algorithm is to do a local linear approximation of the network at the current point (which is the Taylor expansion up to the gradient term). After that, the algorithm identifies a class (output coordinate) with the minimal \"margin to gradient norm ratio\", i.e. the total movement in gradient direction to change the labeling function in that coordinate, within this linear approximation. The algorithm solves the subproblem of minimizing a linear function inside lp ball as the critical routine.\n\nOverall, the notion of finding the minimal distortion attacker as opposed to finding the best attacker inside a fixed distortion ball is quite interesting to me. The main concern for me about this paper is the comparison to other methods such as PGD. As far as I know, these attackers DO NOT explicitly minimize the distortion, thus it is quite believable that these models do not identify the minimal distortion solution (rather it will more likely to find a solution that lies in the boundary since it would be the easiest way to attack). However, for the proposed algorithm in this paper, the algorithm is explicitly minimizing the distance to the given input (x_orig in their language). \n\n\nI would like to see more implementation details of the other algorithms, for example, what is the performance if we add an additional regularizer as the distance of the current attacker to the given input to PGD. So far, the paper lacks solid proof of the usefulness of this particular algorithm. (In particular the justification for solving the local linear system instead of doing a gradient descent step).\n\nAfter Rebuttal: I have read the authors' responses and acknowledge the sensibility of the statement. I apologize for the earlier misunderstanding and higher the score accordingly.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571602246091}], "openreview_url": "https://openreview.net/forum?id=HJlzxgBtwH", "arxiv_id": "1907.02044", "paper_pdf": "papers/HJlzxgBtwH.pdf", "paper_pdf_sha256": "8bc18107b29805c3f4690737aec72d218d5c1a6751e7f0ddc4b2958ad1952170", "paper_pdf_bytes": 3264275, "paper_pdf_source": "openreview", "code_url": "https://github.com/fra31/fab-attack", "code_repository": "fra31/fab-attack", "code_commit": "7c3ef4cb4cd91ade912f06c683fbed8a4660464d", "code_archive": "repos/HJlzxgBtwH.zip", "code_archive_sha256": "97d6d30d34229ac69bcb9dd17cc84210147b93fa773ed331ed9de3f0201573f2", "code_archive_bytes": 18226, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 18, "github_languages": {"Python": 47504}, "github_archived": false, "github_pushed_at": "2020-07-10T12:35:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/minimally-distorted-adversarial-examples-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJx7l309Fm", "year": 2019, "status": "rejected", "title": "Actor-Attention-Critic for Multi-Agent Reinforcement Learning", "authors": ["Shariq Iqbal", "Fei Sha"], "authorids": ["shariqiqbal2810@gmail.com", "feisha.work@gmail.com"], "authors_source": "OpenReview API", "abstract": "Reinforcement learning in multi-agent scenarios is important for real-world applications but presents challenges beyond those seen in single-agent settings. We present an actor-critic algorithm that trains decentralized policies in multi-agent settings, using centrally computed critics that share an attention mechanism which selects relevant information for each agent at every timestep. This attention mechanism enables more effective and scalable learning in complex multi-agent environments, when compared to recent approaches. Our approach is applicable not only to cooperative settings with shared rewards, but also individualized reward settings, including adversarial settings, and it makes no assumptions about the action spaces of the agents. As such, it is flexible enough to be applied to most multi-agent learning problems", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1xzqlrChm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1062/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers an actor-critic scheme for multiagent RL, where the critic is specific to each agent and has access to all other agents' embedded observations. The main idea is to use an attention mechanism in the critic that learns to selectively scale the contributions of the other agents. \n\nThe paper presents sufficient motivation and background, and the proposed algorithmic implementation seems reasonable. The proposed scheme is compared to two recent algorithms for centralized training of decentralized policies, and shows comparable or better results on two synthetic multiagent problems. \n\nI believe that the idea and approach of the paper are interesting and contribute to the multiagent learning literature. \n\nRegarding cons: \n- The critical structural choices (such as the attention model in section 3.2) are presented without too much justification, discussion of alternatives, etc. \n- The experiments show the learning results, but do not provide a peak \"under the hood\" to understand the way attention evolved and contributed to the results. \n- The experiments show good results compared to existing algorithms, but not impressively so. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting contribution to multiagent RL ", "review": "The paper considers an actor-critic scheme for multiagent RL, where the critic is specific to each agent and has access to all other agents' embedded observations. The main idea is to use an attention mechanism in the critic that learns to selectively scale the contributions of the other agents. \n\nThe paper presents sufficient motivation and background, and the proposed algorithmic implementation seems reasonable. The proposed scheme is compared to two recent algorithms for centralized training of decentralized policies, and shows comparable or better results on two synthetic multiagent problems. \n\nI believe that the idea and approach of the paper are interesting and contribute to the multiagent learning literature. \n\nRegarding cons: \n- The critical structural choices (such as the attention model in section 3.2) are presented without too much justification, discussion of alternatives, etc. \n- The experiments show the learning results, but do not provide a peak \"under the hood\" to understand the way attention evolved and contributed to the results. \n- The experiments show good results compared to existing algorithms, but not impressively so. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541456009825}, {"id": "rkl9CLxv27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1062/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new method for multi-agent reinforcement learning. The proposed algorithm -- which uses shared critics at training time but individual policies at test time -- makes use of a specialised attention mechanism. The benefits include better scalability (as the dependency of the inputs is linear in the number of agents, rather than quadratic), and also being more amenable to diverse reward and action structures than the previous work. \n\n---------Quality and clarity---------\nThe paper is nicely written, and the ideas are developed in a clear fashion, if slightly verbose (the first 3 pages, though informative, might have been condensed a bit to make more room for the new algorithm). The problem is well-motivated and the benefits of the new algorithm are well showcased.\n\nOne negative point that does stick out is the bibliography, where papers that have been published for years (e.g. the Adam paper) are still referenced as arXiv preprints.\n\n---------Originality and significance----------\nAlthough attentive mechanisms have been around for a while, their use in this specific setting (learning shared critics for multi-agent RL) is, and yields desirable properties. The new algorithm opens the door for training in more complex environments, with a larger number of agents (although the number is still limited in the presented experiments).\n\nThe main issue I do see with the paper is its experimental section. \nThe two tasks are picked to showcase the benefits of the new approach. This does mean that the competing algorithms have to undergo significant changes (at least in the case of the DDPG-based methods), which takes away from the validity of the comparison. \n\nIdeally, there would be at least one other task on which the other algorithms have been trained on by their respective authors. As mentioned right before Section 4, MAAC can be used on continuous action spaces at the price of increased computational cost, so this should be doable.\n\n\nOverall, this is a nicely written paper which introduces an interesting new method for multi-agent RL, with promising initial results. A more thorough experimental section with slightly fairer comparisons would increase its quality significantly.\n\nPros\n- clear paper, easy to read\n- interesting application of attention mechanism to multi-agent RL\n- promising initial results\n\nCons\n- no comparison to related algorithms on tasks where they have already been evaluated externally\n- the amount of workers is still quite limited in the experiments", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting new method, though more thorough experiments are needed", "review": "This paper introduces a new method for multi-agent reinforcement learning. The proposed algorithm -- which uses shared critics at training time but individual policies at test time -- makes use of a specialised attention mechanism. The benefits include better scalability (as the dependency of the inputs is linear in the number of agents, rather than quadratic), and also being more amenable to diverse reward and action structures than the previous work. \n\n---------Quality and clarity---------\nThe paper is nicely written, and the ideas are developed in a clear fashion, if slightly verbose (the first 3 pages, though informative, might have been condensed a bit to make more room for the new algorithm). The problem is well-motivated and the benefits of the new algorithm are well showcased.\n\nOne negative point that does stick out is the bibliography, where papers that have been published for years (e.g. the Adam paper) are still referenced as arXiv preprints.\n\n---------Originality and significance----------\nAlthough attentive mechanisms have been around for a while, their use in this specific setting (learning shared critics for multi-agent RL) is, and yields desirable properties. The new algorithm opens the door for training in more complex environments, with a larger number of agents (although the number is still limited in the presented experiments).\n\nThe main issue I do see with the paper is its experimental section. \nThe two tasks are picked to showcase the benefits of the new approach. This does mean that the competing algorithms have to undergo significant changes (at least in the case of the DDPG-based methods), which takes away from the validity of the comparison. \n\nIdeally, there would be at least one other task on which the other algorithms have been trained on by their respective authors. As mentioned right before Section 4, MAAC can be used on continuous action spaces at the price of increased computational cost, so this should be doable.\n\n\nOverall, this is a nicely written paper which introduces an interesting new method for multi-agent RL, with promising initial results. A more thorough experimental section with slightly fairer comparisons would increase its quality significantly.\n\nPros\n- clear paper, easy to read\n- interesting application of attention mechanism to multi-agent RL\n- promising initial results\n\nCons\n- no comparison to related algorithms on tasks where they have already been evaluated externally\n- the amount of workers is still quite limited in the experiments", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1540978385983}, {"id": "B1xTnUUMhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1062/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n\nAuthors present a decentralized policy, centralized value function approach (MAAC) to multi-agent learning. They used an attention mechanism over agent policies as an input to a central value function. \n\nAuthors compare their approach with COMA (discrete actions and counterfactual (semi-centralized) baseline) and MADDPG (also uses centralized value function and continuous actions)\n\nMAAC is evaluated on two 2d cooperative environments, Treasure Collection and Rover Tower. MAAC outperforms baselines on TC, but not on RT. Furthermore, the different baselines perform differently: there is no method that consistently performs well.\n\nPro\n- MAAC is a simple combination of attention and a centralized value function approach.\n\nCon\n- MAAC still requires all observations and actions of all other agents as an input to the value function, which makes this approach not scalable to settings with many agents. \n- The centralized nature is also semantically improbable, as the observations might be high-dimensional in nature, so exchanging these between agents becomes impractical with complex problems.\n- MAAC does not consistently outperform baselines, and it is not clear how the stated explanations about the difference in performance apply to other problems. \n- Authors do not visualize the attention (as is common in previous work involving attention in e.g., NLP). It is unclear how the model actually operates and uses attention during execution.\n\nReproducibility\n- It seems straightforward to implement this method, but I encourage open-sourcing the authors' implementation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple method, but gives insufficient insight in model behavior and how it could generalize", "review": "Summary\n\nAuthors present a decentralized policy, centralized value function approach (MAAC) to multi-agent learning. They used an attention mechanism over agent policies as an input to a central value function. \n\nAuthors compare their approach with COMA (discrete actions and counterfactual (semi-centralized) baseline) and MADDPG (also uses centralized value function and continuous actions)\n\nMAAC is evaluated on two 2d cooperative environments, Treasure Collection and Rover Tower. MAAC outperforms baselines on TC, but not on RT. Furthermore, the different baselines perform differently: there is no method that consistently performs well.\n\nPro\n- MAAC is a simple combination of attention and a centralized value function approach.\n\nCon\n- MAAC still requires all observations and actions of all other agents as an input to the value function, which makes this approach not scalable to settings with many agents. \n- The centralized nature is also semantically improbable, as the observations might be high-dimensional in nature, so exchanging these between agents becomes impractical with complex problems.\n- MAAC does not consistently outperform baselines, and it is not clear how the stated explanations about the difference in performance apply to other problems. \n- Authors do not visualize the attention (as is common in previous work involving attention in e.g., NLP). It is unclear how the model actually operates and uses attention during execution.\n\nReproducibility\n- It seems straightforward to implement this method, but I encourage open-sourcing the authors' implementation.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540675253431}], "openreview_url": "https://openreview.net/forum?id=HJx7l309Fm", "arxiv_id": "1810.02912", "paper_pdf": "papers/HJx7l309Fm.pdf", "paper_pdf_sha256": "dcdf7ac024d4ee05494c3f1a323a17bf4bde661b97c82292d9e5546ff9c50ccd", "paper_pdf_bytes": 739740, "paper_pdf_source": "openreview", "code_url": "https://github.com/shariqiqbal2810/MAAC", "code_repository": "shariqiqbal2810/MAAC", "code_commit": "6174a01251251e6778c4ada26bc8d9cd930e3856", "code_archive": "repos/HJx7l309Fm.zip", "code_archive_sha256": "b3ddcd48636a5fda010f64338e06e632a7c600ac3f400a04207bbec175f99636", "code_archive_bytes": 23481, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 30, "github_languages": {"Python": 66332}, "github_archived": false, "github_pushed_at": "2022-05-29T16:14:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/actor-attention-critic-for-multi-agent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SybqeKgA-", "year": 2018, "status": "rejected", "title": "On Batch Adaptive Training for Deep Learning: Lower Loss and Larger Step Size", "authors": ["Runyao Chen", "Kun Wu", "Ping Luo"], "authorids": ["chenrunyao14@mails.ucas.ac.cn", "WuKun14@mails.ucas.ac.cn", "luop@ict.ac.cn"], "authors_source": "OpenReview API", "abstract": "Mini-batch gradient descent and its variants are commonly used in deep learning. The principle of mini-batch gradient descent is to use noisy gradient calculated on a batch to estimate the real gradient, thus balancing the computation cost per iteration and the uncertainty of noisy gradient. However, its batch size is a fixed hyper-parameter requiring manual setting before training the neural network. Yin et al. (2017) proposed a batch adaptive stochastic gradient descent (BA-SGD) that can dynamically choose a proper batch size as learning proceeds. We extend the BA-SGD to momentum algorithm and evaluate both the BA-SGD and the batch adaptive momentum (BA-Momentum) on two deep learning tasks from natural language processing to image classification. Experiments confirm that batch adaptive methods can achieve a lower loss compared with mini-batch methods after scanning the same epochs of data. Furthermore, our BA-Momentum is more robust against larger step sizes, in that it can dynamically enlarge the batch size to reduce the larger uncertainty brought by larger step sizes. We also identified an interesting phenomenon, batch size boom. The code implementing batch adaptive framework is now open source, applicable to any gradient-based optimization problems.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HktgWy7xM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper330/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overall, the manuscript is well organized and written with solid background knowledge and results to support the claim of the paper.  The authors borrow the idea from a previously published work and claim that their contributions are twofold: (1) extend batch adaptive SGD to adaptive momentum, and (2) adopt the algorithms to complex neural networks problems (while the previous paper only demonstrates with simple neural networks).  In this regard, it does not show much novelty.  Several issues should be addressed to improve the quality of the paper:  \n
1) The paper has demonstrated that the proposed method exhibits fast convergence and lower training loss.  However, the test accuracy is not shown.  This makes it hard to justify the effectiveness of the proposed method.  \n
2) From Fig. 4(b), it shows that the batch size is updated in every iteration.  The reviewer wonders whether it is too frequent.  Moreover, the paper does not explicitly show the computation cost of computing the batch size. \n3) The comparison of other methodologies seems not fair.  All the compared methods adopt a fixed batch size, but the proposed method uses an adaptive batch size.  The paper can compare the proposed method with adaptive batch size in intuitive settings, e.g., small batch size in the beginning of training and larger batch size later.\n4) The font size is too small in some figures, e.g., Figure 7(a).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This manuscript addresses the problem of automatically tuning the batch size during deep learning training.  Experiments are conducted on CNN and LSTM RNN to demonstrate the advantages of the proposed method.", "rating": "5: Marginally below acceptance threshold", "review": "Overall, the manuscript is well organized and written with solid background knowledge and results to support the claim of the paper.  The authors borrow the idea from a previously published work and claim that their contributions are twofold: (1) extend batch adaptive SGD to adaptive momentum, and (2) adopt the algorithms to complex neural networks problems (while the previous paper only demonstrates with simple neural networks).  In this regard, it does not show much novelty.  Several issues should be addressed to improve the quality of the paper:  \n
1) The paper has demonstrated that the proposed method exhibits fast convergence and lower training loss.  However, the test accuracy is not shown.  This makes it hard to justify the effectiveness of the proposed method.  \n
2) From Fig. 4(b), it shows that the batch size is updated in every iteration.  The reviewer wonders whether it is too frequent.  Moreover, the paper does not explicitly show the computation cost of computing the batch size. \n3) The comparison of other methodologies seems not fair.  All the compared methods adopt a fixed batch size, but the proposed method uses an adaptive batch size.  The paper can compare the proposed method with adaptive batch size in intuitive settings, e.g., small batch size in the beginning of training and larger batch size later.\n4) The font size is too small in some figures, e.g., Figure 7(a).\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511350512855}, {"id": "HkvBT03Jf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper330/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose extending the recently-proposed adaptive batch-size approach of Yin et al. to an update that includes momentum, and perform more comprehensive experiments than in the Yin et al. paper validating their approach.\n\nThe basic idea makes a great deal of intuitive sense: inaccurate gradient estimates are fine in early iterations, when we're far from convergence, and accurate estimates are more valuable in later iterations, when we're close. Finding the optimal trade-off between computational cost and expected decrease seems like the most natural way to accomplish this, and this is precisely what they propose. That said, I'm not totally convinced by the derivation of sections 2 and 3: the Gaussian assumption is fine as a heuristic (and they don't really claim that it's anything else), but I don't feel that the proposed algorithm really rests on a solid theoretical foundation.\n\nThe extension to the momentum case (section 3) seems to be more-or-less straightforward, but I do have a question about equation 15: am I misreading this, or is it saying that the variance of the momentum update \\mathcal{P} is the same as the variance of the most recent minibatch? Shouldn't it depend on the previous terms which are included in \\mathcal{P}?\n\nI'm also not convinced by the dependence on the \"optimal\" objective function value S^* in equation 6. In their algorithm, they take S^* to be zero, which is a good conservative choice for a nonnegative loss, but the fact that this quantity is present in the first place, as a user-specified parameter, makes me nervous, since even for a nonnegative loss, the optimum might be quite far from zero, and on a non-convex problem, the eventual local optimum at which we eventually settle down may be further still.\n\nAlso, the \"Robbins 2007\" reference should, I believe, be \"Robbins and Monro, 1951\".\n\nThese are all relatively minor issues, however. My main criticism is that the experiments only report results in terms of *training* loss. The use of adaptive batch sizes does indeed appear to result in faster convergence in terms of training loss, but the plots are in log scale (which I do think is the right way to present it), so the difference is smaller in reality than it appears visually. To determine whether this improvement in training performance is a *real* improvement, I think we need to see the performance (in terms of accuracy, not loss) on held-out data.\n\nFinally, as the authors mention in the final paragraph of their conclusion, some recent work has indicated that large-batch methods may generalize worse than small-batch methods. They claim that, by using small batches early and large batches late, they may avoid this issue, and I don't necessarily disagree, but I think an argument could be made in the opposite direction: that since the proposed approach becomes a large-batch method in the later iterations, it may suffer from this problem. I think that this is worth exploring further, and, again, without results on testing data being presented, a reader can't make any determination about how well the proposed method generalizes, compared to fixed-size minibatches.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and intuitive idea, but I'm not convinced that it adds enough over Yin et al. KDD paper, and think the experiments must include results on testing data.", "rating": "5: Marginally below acceptance threshold", "review": "The authors propose extending the recently-proposed adaptive batch-size approach of Yin et al. to an update that includes momentum, and perform more comprehensive experiments than in the Yin et al. paper validating their approach.\n\nThe basic idea makes a great deal of intuitive sense: inaccurate gradient estimates are fine in early iterations, when we're far from convergence, and accurate estimates are more valuable in later iterations, when we're close. Finding the optimal trade-off between computational cost and expected decrease seems like the most natural way to accomplish this, and this is precisely what they propose. That said, I'm not totally convinced by the derivation of sections 2 and 3: the Gaussian assumption is fine as a heuristic (and they don't really claim that it's anything else), but I don't feel that the proposed algorithm really rests on a solid theoretical foundation.\n\nThe extension to the momentum case (section 3) seems to be more-or-less straightforward, but I do have a question about equation 15: am I misreading this, or is it saying that the variance of the momentum update \\mathcal{P} is the same as the variance of the most recent minibatch? Shouldn't it depend on the previous terms which are included in \\mathcal{P}?\n\nI'm also not convinced by the dependence on the \"optimal\" objective function value S^* in equation 6. In their algorithm, they take S^* to be zero, which is a good conservative choice for a nonnegative loss, but the fact that this quantity is present in the first place, as a user-specified parameter, makes me nervous, since even for a nonnegative loss, the optimum might be quite far from zero, and on a non-convex problem, the eventual local optimum at which we eventually settle down may be further still.\n\nAlso, the \"Robbins 2007\" reference should, I believe, be \"Robbins and Monro, 1951\".\n\nThese are all relatively minor issues, however. My main criticism is that the experiments only report results in terms of *training* loss. The use of adaptive batch sizes does indeed appear to result in faster convergence in terms of training loss, but the plots are in log scale (which I do think is the right way to present it), so the difference is smaller in reality than it appears visually. To determine whether this improvement in training performance is a *real* improvement, I think we need to see the performance (in terms of accuracy, not loss) on held-out data.\n\nFinally, as the authors mention in the final paragraph of their conclusion, some recent work has indicated that large-batch methods may generalize worse than small-batch methods. They claim that, by using small batches early and large batches late, they may avoid this issue, and I don't necessarily disagree, but I think an argument could be made in the opposite direction: that since the proposed approach becomes a large-batch method in the later iterations, it may suffer from this problem. I think that this is worth exploring further, and, again, without results on testing data being presented, a reader can't make any determination about how well the proposed method generalizes, compared to fixed-size minibatches.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1510956351292}, {"id": "r1b1AtYlG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper330/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a generalization of an algorithm by Yin et al. (2017), which performs SGD with adaptive batch sizes. The present paper generalizes the algorithm to SGD with momentum. Since the original algorithm was already formulated with a general utility function, the proposed algorithm is similar in structure but replaces the utility function so that it takes momentum into account. Experiments on an image classification task show improvements in the training loss. However, no test accuracies are reported and the learning curves have suspicious artifacts, see below. Experiments on a relation extraction task show little improvement over SGD with momentum and constant batch size.\n\n\nCOMMENTS:\n\nThe paper discusses a relevant issue. While adaptive learning algorithms are popular in deep learning, most algorithms adapt the learning rate or the momentum coefficient, but not the batch size. It appears to me that the main idea and the overall structure of the proposed algorithm is the same as in the one published by Yin et al. (2017), and that only few changes were necessary to include momentum. Given the incremental process, I find the presentation unnecessarily involved, and experiments not convincing enough.\n\nConcerning the presentation, the paper dedicates two full pages on a review of the algorithm by Yin et al. (2017). The first page of this review states that, for large enough batch sizes, the change of the objective function in SGD is normal distributed with a variance that is inversely proportional the batch size. It seems to me that this is a direct consequence of the central limit theorem. The derivation, however, is quite technical and introduces some quantities that are never used (e.g., $\\vec{\\xi}_j$ is never used individually, only the combined term $\\epsilon_t$ defined below Eq. 12 is). The second page of the review seems to discuss the main part of the algorithm, but I could not follow it. First, a \"state\" $s_t$ (also written as $S$) is introduced, which, according to the text, is \"the objective value\", which was earlier denoted by $F$. Nevertheless, the change of $s_t$, Eq. 5, appears to obey a different probability distribution than the change of $F$. The paper provides a verbal explanation for this discrepancy, saying that it is possible that $S$ is first reduced to the minimum $S^*$ of the objective and then increased again. However, in my understanding, the minimum of the objective is only realized at a singular point in parameter space. Crossing this point in an update step should have zero probability as long as the model has more than one parameter. The explanation also does not make it clear why the argument should apply to $S$ (or $s$) but not to $F$.\n\nPage 5 provides pseudocode for the proposed algorithm. However, I couldn't find an explanation of the code. The code suggests that, for each update step, one gradually increases the batch size until it becomes larger or equal than a running estimate of the optimal batch size. While this may be a plausible strategy in practice, it seems to have a bias that is not addressed in the paper: the algorithm recalculates a noisy estimate of the optimal batch size after each increase of the batch size, and it terminates as soon as the noisy estimate happens to be small enough, resulting in a bias towards a smaller than optimal batch size. A probably more important issue is that the algorithm is sequential and hard to parallelize, where parallelization is usually the main motivation to use larger batch sizes. As the gradient noise scales inversely proportional to the batch size, I don't see why increasing the batch size should be preferred over decreasing the learning rate unless optimizations with a larger batch size can be parallelized. The experiments don't compare the two alternatives.\n\nConcerning the experiments, it seems peculiar that the learning curves in Figure 1 remain at a constant value for a long time at the beginning of the optimization before they begin to drop. Do the authors understand this behavior? It could indicate that the magnitude of the random initialization was chosen too small. I.e., the parameters might have been initialized too close to zero, where the loss is stationary due to symmetries. Also, absolute values of the training loss can be deceptive since there is often no natural scale. A better indicator of convergence would be the test accuracy. The identification of the \"batch size boom\" is interesting.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A small modification of adaptive batch size without momentum with unconvincing experiments", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes a generalization of an algorithm by Yin et al. (2017), which performs SGD with adaptive batch sizes. The present paper generalizes the algorithm to SGD with momentum. Since the original algorithm was already formulated with a general utility function, the proposed algorithm is similar in structure but replaces the utility function so that it takes momentum into account. Experiments on an image classification task show improvements in the training loss. However, no test accuracies are reported and the learning curves have suspicious artifacts, see below. Experiments on a relation extraction task show little improvement over SGD with momentum and constant batch size.\n\n\nCOMMENTS:\n\nThe paper discusses a relevant issue. While adaptive learning algorithms are popular in deep learning, most algorithms adapt the learning rate or the momentum coefficient, but not the batch size. It appears to me that the main idea and the overall structure of the proposed algorithm is the same as in the one published by Yin et al. (2017), and that only few changes were necessary to include momentum. Given the incremental process, I find the presentation unnecessarily involved, and experiments not convincing enough.\n\nConcerning the presentation, the paper dedicates two full pages on a review of the algorithm by Yin et al. (2017). The first page of this review states that, for large enough batch sizes, the change of the objective function in SGD is normal distributed with a variance that is inversely proportional the batch size. It seems to me that this is a direct consequence of the central limit theorem. The derivation, however, is quite technical and introduces some quantities that are never used (e.g., $\\vec{\\xi}_j$ is never used individually, only the combined term $\\epsilon_t$ defined below Eq. 12 is). The second page of the review seems to discuss the main part of the algorithm, but I could not follow it. First, a \"state\" $s_t$ (also written as $S$) is introduced, which, according to the text, is \"the objective value\", which was earlier denoted by $F$. Nevertheless, the change of $s_t$, Eq. 5, appears to obey a different probability distribution than the change of $F$. The paper provides a verbal explanation for this discrepancy, saying that it is possible that $S$ is first reduced to the minimum $S^*$ of the objective and then increased again. However, in my understanding, the minimum of the objective is only realized at a singular point in parameter space. Crossing this point in an update step should have zero probability as long as the model has more than one parameter. The explanation also does not make it clear why the argument should apply to $S$ (or $s$) but not to $F$.\n\nPage 5 provides pseudocode for the proposed algorithm. However, I couldn't find an explanation of the code. The code suggests that, for each update step, one gradually increases the batch size until it becomes larger or equal than a running estimate of the optimal batch size. While this may be a plausible strategy in practice, it seems to have a bias that is not addressed in the paper: the algorithm recalculates a noisy estimate of the optimal batch size after each increase of the batch size, and it terminates as soon as the noisy estimate happens to be small enough, resulting in a bias towards a smaller than optimal batch size. A probably more important issue is that the algorithm is sequential and hard to parallelize, where parallelization is usually the main motivation to use larger batch sizes. As the gradient noise scales inversely proportional to the batch size, I don't see why increasing the batch size should be preferred over decreasing the learning rate unless optimizations with a larger batch size can be parallelized. The experiments don't compare the two alternatives.\n\nConcerning the experiments, it seems peculiar that the learning curves in Figure 1 remain at a constant value for a long time at the beginning of the optimization before they begin to drop. Do the authors understand this behavior? It could indicate that the magnitude of the random initialization was chosen too small. I.e., the parameters might have been initialized too close to zero, where the loss is stationary due to symmetries. Also, absolute values of the training loss can be deceptive since there is often no natural scale. A better indicator of convergence would be the test accuracy. The identification of the \"batch size boom\" is interesting.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511787992629}], "openreview_url": "https://openreview.net/forum?id=SybqeKgA-", "arxiv_id": null, "paper_pdf": "papers/SybqeKgA-.pdf", "paper_pdf_sha256": "e98e153f12732b25dfda3eff31c04e5ea49ba7daf66402734caf1799a3bc7ab7", "paper_pdf_bytes": 1551207, "paper_pdf_source": "openreview", "code_url": "https://github.com/thomasyao3096/Batch_Adaptive_Framework", "code_repository": "thomasyao3096/Batch_Adaptive_Framework", "code_commit": "2244eebc33938ef816120a80eb2560192ec08b41", "code_archive": "repos/SybqeKgA-.zip", "code_archive_sha256": "f53cd8052f21d0cf8cd9de2c3e605ddbd47b1f47c59cda5d3abaa5ed39d86dc9", "code_archive_bytes": 11261, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 19, "github_languages": {"Python": 32613}, "github_archived": false, "github_pushed_at": "2017-11-12T02:27:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-batch-adaptive-training-for-deep-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ODibPQmeP1", "year": 2026, "status": "rejected", "title": "CHARM: Calibrating Reward Models With Chatbot Arena Scores", "authors": ["Xiao Zhu", "Chenmien Tan", "Pinzhen Chen", "Rico Sennrich", "Yanlin Zhang", "Hanxu Hu"], "authorids": ["~Xiao_Zhu4", "~Chenmien_Tan1", "~Pinzhen_Chen1", "~Rico_Sennrich1", "~Yanlin_Zhang1", "~Hanxu_Hu1"], "authors_source": "OpenReview API", "abstract": "Reward models (RMs) play a crucial role in Reinforcement Learning from Human Feedback by serving as proxies for human preferences in aligning large language models. However, they suffer from various biases which could lead to reward hacking. In this paper, we identify a model preference bias in RMs, where they systematically assign disproportionately high scores to responses from certain policy models, leading to unfair judgments. To mitigate this bias, we propose a calibration method named **CH**atbot **A**rena calibrated **R**eward **M**odeling (**CHARM**) that leverages Elo scores from the Chatbot Arena to construct debiased preference datasets and adjust reward model scoring. We conduct extensive experiments on reward model benchmarks and human preference alignment. Results demonstrate that our calibrated RMs achieve improved evaluation accuracy on RM-Bench and the Chat-Hard domain of RewardBench and exhibit a stronger correlation with human preferences by producing scores more closely aligned with Elo rankings.  Beyond this, **CHARM** enhances robustness to stylistic variations, mitigates implicit pattern bias, and generalizes to unseen models. These results demonstrate that **CHARM** provides a simple, effective, and broadly applicable approach to building more reliable and fair reward models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "vIvTja9dac", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5208/Reviewer_q9aZ"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper identifies \"Model Preference Bias\" in reward models, which is a systematic tendency to assign disproportionately high scores to responses from certain policy models. To mitigate this, the authors propose CHARM, a calibration method that uses Elo ratings from Chatbot Arena as a proxy for ground-truth human preferences. CHARM computes a global score offset Δ for an \"over-valued\" model to align the RM's empirical win rate with the Elo-derived expected win rate. The calibrated RMs show improved performance on RM-Bench and RewardBench, better alignment with Chatbot Arena rankings, and enhanced robustness to stylistic variations. The authors also introduce a \"Mismatch Degree\" metric to quantify bias and demonstrate generalization to unseen models.", "review_text": "This paper identifies \"Model Preference Bias\" in reward models, which is a systematic tendency to assign disproportionately high scores to responses from certain policy models. To mitigate this, the authors propose CHARM, a calibration method that uses Elo ratings from Chatbot Arena as a proxy for ground-truth human preferences. CHARM computes a global score offset Δ for an \"over-valued\" model to align the RM's empirical win rate with the Elo-derived expected win rate. The calibrated RMs show improved performance on RM-Bench and RewardBench, better alignment with Chatbot Arena rankings, and enhanced robustness to stylistic variations. The authors also introduce a \"Mismatch Degree\" metric to quantify bias and demonstrate generalization to unseen models.", "strengths": "- The paper articulates a subtle but important bias (i.e., \"Model Preference Bias\") that has been overlooked in reward modeling literature. \n\n- CHARM is straightforward to implement, requiring only a single additive offset per model. This simplicity makes it attractive for practitioners who cannot afford complex retraining pipelines. The method leverages readily available Chatbot Arena data, which is a clever use of existing resources.\n\n- The consistent improvements across multiple RMs (Table 1) and the reduction in Mismatch Degree (Figure 2) are compelling. The analysis of stylistic patterns (Table 3) provides plausible evidence that CHARM mitigates implicit biases beyond just model-specific ones.", "weaknesses": "- Oversimplified Calibration Mechanism: The core assumption that a *single global additive offset* Δ can correct complex, instruction-dependent biases is theoretically questionable. Model preference bias likely manifests differently across prompt categories (e.g., coding vs. creative writing), yet CHARM applies a uniform correction. The paper provides no analysis of whether Δ varies by domain or instruction type, nor ablations showing why a more nuanced correction (e.g., instruction-specific offsets) is unnecessary.\n\n- Weak Ground Truth Assumption: The method treats Chatbot Arena Elo scores as absolute ground truth for human preferences, but these scores have well-documented limitations: they reflect a specific user population, are influenced by positional bias, and conflate multiple criteria (helpfulness, safety, style). The paper doesn't address how Chatbot Arena's biases might propagate into CHARM. For example, if Arena users prefer verbose responses, CHARM might inadvertently bake this length bias *into* the RM rather than remove it.\n\n- Limited Calibration Scope: The main experiments calibrate using only *one* over-valued model (Gemma-2-9b-it-SimPO) and *one* reference model (GPT-4o-mini). This is a major limitation: (1) The paper claims bias is systematic across \"preference-optimized\" models, but only demonstrates calibration on a single instance. Testing with multiple over-valued models (e.g., DPO-tuned, PPO-tuned) is essential to validate generalizability. (2) The offset Δ is fundamentally tied to the choice of π_R. The paper fails to explore how Δ changes with different references (e.g., a weaker model like Llama-3-3.1-8B vs. GPT-4o). This raises questions about the stability and interpretability of the calibration.", "questions": "What is the reward model in line 240 to produce the uncalibrated preference dataset? Will the choice of this reward model influence the final results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper identifies \"Model Preference Bias\" in reward models, which is a systematic tendency to assign disproportionately high scores to responses from certain policy models. To mitigate this, the authors propose CHARM, a calibration method that uses Elo ratings from Chatbot Arena as a proxy for ground-truth human preferences. CHARM computes a global score offset Δ for an \"over-valued\" model to align the RM's empirical win rate with the Elo-derived expected win rate. The calibrated RMs show improved performance on RM-Bench and RewardBench, better alignment with Chatbot Arena rankings, and enhanced robustness to stylistic variations. The authors also introduce a \"Mismatch Degree\" metric to quantify bias and demonstrate generalization to unseen models.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper articulates a subtle but important bias (i.e., \"Model Preference Bias\") that has been overlooked in reward modeling literature. \n\n- CHARM is straightforward to implement, requiring only a single additive offset per model. This simplicity makes it attractive for practitioners who cannot afford complex retraining pipelines. The method leverages readily available Chatbot Arena data, which is a clever use of existing resources.\n\n- The consistent improvements across multiple RMs (Table 1) and the reduction in Mismatch Degree (Figure 2) are compelling. The analysis of stylistic patterns (Table 3) provides plausible evidence that CHARM mitigates implicit biases beyond just model-specific ones.", "weaknesses": "- Oversimplified Calibration Mechanism: The core assumption that a *single global additive offset* Δ can correct complex, instruction-dependent biases is theoretically questionable. Model preference bias likely manifests differently across prompt categories (e.g., coding vs. creative writing), yet CHARM applies a uniform correction. The paper provides no analysis of whether Δ varies by domain or instruction type, nor ablations showing why a more nuanced correction (e.g., instruction-specific offsets) is unnecessary.\n\n- Weak Ground Truth Assumption: The method treats Chatbot Arena Elo scores as absolute ground truth for human preferences, but these scores have well-documented limitations: they reflect a specific user population, are influenced by positional bias, and conflate multiple criteria (helpfulness, safety, style). The paper doesn't address how Chatbot Arena's biases might propagate into CHARM. For example, if Arena users prefer verbose responses, CHARM might inadvertently bake this length bias *into* the RM rather than remove it.\n\n- Limited Calibration Scope: The main experiments calibrate using only *one* over-valued model (Gemma-2-9b-it-SimPO) and *one* reference model (GPT-4o-mini). This is a major limitation: (1) The paper claims bias is systematic across \"preference-optimized\" models, but only demonstrates calibration on a single instance. Testing with multiple over-valued models (e.g., DPO-tuned, PPO-tuned) is essential to validate generalizability. (2) The offset Δ is fundamentally tied to the choice of π_R. The paper fails to explore how Δ changes with different references (e.g., a weaker model like Llama-3-3.1-8B vs. GPT-4o). This raises questions about the stability and interpretability of the calibration.", "questions": "What is the reward model in line 240 to produce the uncalibrated preference dataset? Will the choice of this reward model influence the final results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762569816511}, {"id": "JIjOn4ww6t", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5208/Reviewer_B6Db"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper proposes CHARM, a method to alleviate reward hacking via calibrating the Reward Models (RM) with the Elo rating (Human Preference). Specifically, they utilize the Elo Rating from the public Chatbot Arena and calibrate the reward model's training dataset to align the win-rate estimated from the reward scores with the expected win-rate from the Elo rating. They then fine-tune the reward model to calibrate the reward score to mitigate the Model Preference Bias. Experiments show that the calibrated RM is more aligned with human preference/Elo rating and obtains performance gains against existing RM benchmarks.", "review_text": "This paper proposes CHARM, a method to alleviate reward hacking via calibrating the Reward Models (RM) with the Elo rating (Human Preference). Specifically, they utilize the Elo Rating from the public Chatbot Arena and calibrate the reward model's training dataset to align the win-rate estimated from the reward scores with the expected win-rate from the Elo rating. They then fine-tune the reward model to calibrate the reward score to mitigate the Model Preference Bias. Experiments show that the calibrated RM is more aligned with human preference/Elo rating and obtains performance gains against existing RM benchmarks.", "strengths": "1. The motivation is clear and interesting. Using the Elo rating to calibrate the reward models to make the reward model more aligned with human preferences is sound.\n2. The formulation of the method part is clear and easy to follow. The authors first analyze the correlation between Elo rating and reward model scores, identifying the potential Model Preference Bias, and then explain how they address it (via calibration).\n3. Experiments show that using the calibrated dataset to fine-tune the reward model can improve its performance.", "weaknesses": "1. My main concern is that CHARM is not scalable. Each reward models need their own calibrated dataset for further fine-tuning. Besides, as the elo rating is updated frequently, do you need to recalibrate when the Chatbot Arena is updated? (I may misunderstand the method; point me out if I'm wrong.)\n2. While over-valued and reference models play an important role in calibration, how to select them is not adequately elaborated in the paper. This hinders the critical understanding of CHARM, i.e., how to select or identify the potentially over-valued models in a more systematic manner.", "questions": "1. Instead of using the calibrated dataset, what if we directly use the preference data collected from the chatbot arena (e.g., LMSYS-Chat-1M) to fine-tune the reward model?\n2. How to select the over-valued and reference models?\n3. Can you explain the difference between the calibrated data curated by CHARM and human-annotated preference data?\n4. Have you considered using training-free calibration methods, such as temperature scaling, for calibrating the scalar RM?\n5. Does the update of the chatbot arena's Elo rating affect the calibration performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CHARM, a method to alleviate reward hacking via calibrating the Reward Models (RM) with the Elo rating (Human Preference). Specifically, they utilize the Elo Rating from the public Chatbot Arena and calibrate the reward model's training dataset to align the win-rate estimated from the reward scores with the expected win-rate from the Elo rating. They then fine-tune the reward model to calibrate the reward score to mitigate the Model Preference Bias. Experiments show that the calibrated RM is more aligned with human preference/Elo rating and obtains performance gains against existing RM benchmarks.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The motivation is clear and interesting. Using the Elo rating to calibrate the reward models to make the reward model more aligned with human preferences is sound.\n2. The formulation of the method part is clear and easy to follow. The authors first analyze the correlation between Elo rating and reward model scores, identifying the potential Model Preference Bias, and then explain how they address it (via calibration).\n3. Experiments show that using the calibrated dataset to fine-tune the reward model can improve its performance.", "weaknesses": "1. My main concern is that CHARM is not scalable. Each reward models need their own calibrated dataset for further fine-tuning. Besides, as the elo rating is updated frequently, do you need to recalibrate when the Chatbot Arena is updated? (I may misunderstand the method; point me out if I'm wrong.)\n2. While over-valued and reference models play an important role in calibration, how to select them is not adequately elaborated in the paper. This hinders the critical understanding of CHARM, i.e., how to select or identify the potentially over-valued models in a more systematic manner.", "questions": "1. Instead of using the calibrated dataset, what if we directly use the preference data collected from the chatbot arena (e.g., LMSYS-Chat-1M) to fine-tune the reward model?\n2. How to select the over-valued and reference models?\n3. Can you explain the difference between the calibrated data curated by CHARM and human-annotated preference data?\n4. Have you considered using training-free calibration methods, such as temperature scaling, for calibrating the scalar RM?\n5. Does the update of the chatbot arena's Elo rating affect the calibration performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762025355619}, {"id": "nRGMLy4B9B", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5208/Reviewer_8Hwd"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper identifies model preference bias—a systematic tendency of scalar reward models (RMs) to overrate responses from certain policy models (e.g., Gemma-2-9b-it-SimPO), even when those models perform modestly on human-aligned benchmarks like Chatbot Arena. To address this, the authors propose CHARM, a calibration method that leverages publicly available Elo scores from Chatbot Arena to adjust RM scoring via a learned offset Δ, thereby constructing debiased preference pairs for fine-tuning. Experiments show that CHARM improves RM performance on RM-Bench and RewardBench (especially in the Chat domain), better aligns RM judgments with human preferences, enhances robustness to stylistic variations, and generalizes to unseen models.", "review_text": "This paper identifies model preference bias—a systematic tendency of scalar reward models (RMs) to overrate responses from certain policy models (e.g., Gemma-2-9b-it-SimPO), even when those models perform modestly on human-aligned benchmarks like Chatbot Arena. To address this, the authors propose CHARM, a calibration method that leverages publicly available Elo scores from Chatbot Arena to adjust RM scoring via a learned offset Δ, thereby constructing debiased preference pairs for fine-tuning. Experiments show that CHARM improves RM performance on RM-Bench and RewardBench (especially in the Chat domain), better aligns RM judgments with human preferences, enhances robustness to stylistic variations, and generalizes to unseen models.", "strengths": "1. Studies an important and underexplored problem: calibrating RMs using real-world human preference signals (Chatbot Arena Elo).\n2. Demonstrates clear empirical gains for small-to-medium scalar RMs after calibration, with thorough analysis of bias reduction and generalization.", "weaknesses": "1. The method fits RM scores to Elo-derived win rates via MSE. But why not train directly on Chatbot Arena’s raw battle data or use Elo-predicted probabilities as soft labels? That would seem more direct and principled.\n2. Results are limited to scalar RMs. Given the growing shift toward Generative Reward Models (GRMs), it’s crucial to show whether CHARM can also improve GRMs (e.g., by relabeling preference data used to train LLM-as-a-judge systems).\n3. The Elo scores used—are they the standard raw Arena Elo or style-controlled (e.g., length-normalized)? If the latter, the observed debiasing might stem from style control rather than the calibration mechanism itself. Clarification is needed.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper identifies model preference bias—a systematic tendency of scalar reward models (RMs) to overrate responses from certain policy models (e.g., Gemma-2-9b-it-SimPO), even when those models perform modestly on human-aligned benchmarks like Chatbot Arena. To address this, the authors propose CHARM, a calibration method that leverages publicly available Elo scores from Chatbot Arena to adjust RM scoring via a learned offset Δ, thereby constructing debiased preference pairs for fine-tuning. Experiments show that CHARM improves RM performance on RM-Bench and RewardBench (especially in the Chat domain), better aligns RM judgments with human preferences, enhances robustness to stylistic variations, and generalizes to unseen models.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Studies an important and underexplored problem: calibrating RMs using real-world human preference signals (Chatbot Arena Elo).\n2. Demonstrates clear empirical gains for small-to-medium scalar RMs after calibration, with thorough analysis of bias reduction and generalization.", "weaknesses": "1. The method fits RM scores to Elo-derived win rates via MSE. But why not train directly on Chatbot Arena’s raw battle data or use Elo-predicted probabilities as soft labels? That would seem more direct and principled.\n2. Results are limited to scalar RMs. Given the growing shift toward Generative Reward Models (GRMs), it’s crucial to show whether CHARM can also improve GRMs (e.g., by relabeling preference data used to train LLM-as-a-judge systems).\n3. The Elo scores used—are they the standard raw Arena Elo or style-controlled (e.g., length-normalized)? If the latter, the observed debiasing might stem from style control rather than the calibration mechanism itself. Clarification is needed.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761996185735}, {"id": "N9fykZeKGb", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5208/Reviewer_Wq5j"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The submission compares the scores generated by popular reward models with the elo scores provided by chatbot arena and finds an inconsistency for certain models, which the submission interprets as misscalibration.\nTo prevent this issue, the authors to propose to first generate a dataset using the over-valued policies and reference policies and then use this dataset to construct a calibrated dataset of winning/losing reply pairs.\nIn experiment they show that finetuning an RM using this dataset can improve agreement with the reward bench dataset.", "review_text": "The submission compares the scores generated by popular reward models with the elo scores provided by chatbot arena and finds an inconsistency for certain models, which the submission interprets as misscalibration.\nTo prevent this issue, the authors to propose to first generate a dataset using the over-valued policies and reference policies and then use this dataset to construct a calibrated dataset of winning/losing reply pairs.\nIn experiment they show that finetuning an RM using this dataset can improve agreement with the reward bench dataset.", "strengths": "* The authors identify a new bias of popular reward models, particularly the preference for certain models beyond the probability implied by the chatbot arena elo. This is an interesting finding and is useful to be aware of when training RMs\n * The proposed method yields an improvement in RM accuracy on reward bench\n * The analysis of stylistic variations is interesting", "weaknesses": "* the entire paper is based on the premise that chatbot arena accurately reflects human preferences and that we want reward models to accurately correspond to chatbot arena scores, which is not sufficiently justified, particularly when actually using an RM for post-training\n * experiments also only consider correcting for the over-valuation of a single model (gemma-2-9b-it-SimPO). It is thus not clear whether  the results hold for other over-valued models, or in cases where we want to correct for multiple over-valued models at once \n * experiments show that calibration on chatbot arena leads to an improved reward accuracy on RM-bench. Experiments do not investigate whether using the calibrated RM is beneficial when preference-tuning a model. Recent research [2,3] has shown that reward accuracy on its own does not necessarily result in a better alignment of a post-trained model. Using the calibrated RM for post-training would significantly strengthen the experimental validation\n * terminology in the paper is inconsistent with the common usage of \"calibration\" in the context of supervised learning. Calibration usually refers to the error in probabilities $P_\\theta(y|x)$ given by a classifier vs the true $P(y|x)$, in particular on a sample-wise level. The submission instead considers a class/model level bias of the reward model. Further the entire literature on classifier calibration is entirely ignored by the submission, see for example [1] for a survey\n\n\n[1] Filho et al. \"Classifier Calibration: A survey on how to assess and improve predicted class probabilities\", Machine Learning 2023\n[2] Razin et al. \"What Makes a Reward Model a Good Teacher? An Optimization Perspective\", NeurIPS 2025\n[3] Chen et al. \"The Accuracy Paradox in RLHF: When Better Reward Models Don't Yield Better Language Models\", EMNLP 2024", "questions": "* Why was the target winrate P(O) determined from elo scores? Chatbot arena provides head-to-head winrates of model pairings, which directly yield P(O) per matchup instead of averaging over models. It seems like using this would be more direct.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission compares the scores generated by popular reward models with the elo scores provided by chatbot arena and finds an inconsistency for certain models, which the submission interprets as misscalibration.\nTo prevent this issue, the authors to propose to first generate a dataset using the over-valued policies and reference policies and then use this dataset to construct a calibrated dataset of winning/losing reply pairs.\nIn experiment they show that finetuning an RM using this dataset can improve agreement with the reward bench dataset.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* The authors identify a new bias of popular reward models, particularly the preference for certain models beyond the probability implied by the chatbot arena elo. This is an interesting finding and is useful to be aware of when training RMs\n * The proposed method yields an improvement in RM accuracy on reward bench\n * The analysis of stylistic variations is interesting", "weaknesses": "* the entire paper is based on the premise that chatbot arena accurately reflects human preferences and that we want reward models to accurately correspond to chatbot arena scores, which is not sufficiently justified, particularly when actually using an RM for post-training\n * experiments also only consider correcting for the over-valuation of a single model (gemma-2-9b-it-SimPO). It is thus not clear whether  the results hold for other over-valued models, or in cases where we want to correct for multiple over-valued models at once \n * experiments show that calibration on chatbot arena leads to an improved reward accuracy on RM-bench. Experiments do not investigate whether using the calibrated RM is beneficial when preference-tuning a model. Recent research [2,3] has shown that reward accuracy on its own does not necessarily result in a better alignment of a post-trained model. Using the calibrated RM for post-training would significantly strengthen the experimental validation\n * terminology in the paper is inconsistent with the common usage of \"calibration\" in the context of supervised learning. Calibration usually refers to the error in probabilities $P_\\theta(y|x)$ given by a classifier vs the true $P(y|x)$, in particular on a sample-wise level. The submission instead considers a class/model level bias of the reward model. Further the entire literature on classifier calibration is entirely ignored by the submission, see for example [1] for a survey\n\n\n[1] Filho et al. \"Classifier Calibration: A survey on how to assess and improve predicted class probabilities\", Machine Learning 2023\n[2] Razin et al. \"What Makes a Reward Model a Good Teacher? An Optimization Perspective\", NeurIPS 2025\n[3] Chen et al. \"The Accuracy Paradox in RLHF: When Better Reward Models Don't Yield Better Language Models\", EMNLP 2024", "questions": "* Why was the target winrate P(O) determined from elo scores? Chatbot arena provides head-to-head winrates of model pairings, which directly yield P(O) per matchup instead of averaging over models. It seems like using this would be more direct.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761470422400}], "openreview_url": "https://openreview.net/forum?id=ODibPQmeP1", "arxiv_id": "2504.10045", "paper_pdf": "papers/ODibPQmeP1.pdf", "paper_pdf_sha256": "a83b6bf53aa1f23dda4b9d2b24478039d97acaeb0fbb78db1b745a65f814d236", "paper_pdf_bytes": 1281082, "paper_pdf_source": "openreview", "code_url": "https://github.com/HexagonStar/CHARM", "code_repository": "HexagonStar/CHARM", "code_commit": "719f0aef03a1eb8381bad450b0d991455b91563a", "code_archive": "repos/ODibPQmeP1.zip", "code_archive_sha256": "2629f6448f327bc04d0ff4af9b127e6069eb7b624197d5117372fc5e7dc3cb3f", "code_archive_bytes": 21492, "code_file_count": 12, "code_extensions": {".py": 7, ".sh": 5}, "github_disk_usage_kb": 42, "github_languages": {"Python": 32355, "Shell": 2532}, "github_archived": false, "github_pushed_at": "2025-04-15T05:31:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/charm-calibrating-reward-models-with-chatbot"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pogJXugbN8", "year": 2024, "status": "rejected", "title": "BAFFLE: A Baseline of Backpropagation-Free Federated Learning", "authors": ["Haozhe Feng", "Tianyu Pang", "Chao Du", "Wei Chen", "Shuicheng YAN", "Min Lin"], "authorids": ["~Haozhe_Feng1", "~Tianyu_Pang1", "~Chao_Du1", "~Wei_Chen34", "~Shuicheng_YAN3", "~Min_Lin1"], "authors_source": "OpenReview API", "abstract": "Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "ebgb2owIUT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7596/Reviewer_7JpN"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes BAFFLE, a zeroth-order federated learning method with secure aggregation and TEE. BAFFLE benefits from the BP-free local optimization process and thus is more memory-efficient than traditional BP-based algorithms. Empirical studies demonstrate the effectiveness and the low memory costs of the proposed BAFFLE.", "review_text": "This paper proposes BAFFLE, a zeroth-order federated learning method with secure aggregation and TEE. BAFFLE benefits from the BP-free local optimization process and thus is more memory-efficient than traditional BP-based algorithms. Empirical studies demonstrate the effectiveness and the low memory costs of the proposed BAFFLE.", "strengths": "1. The paper is well-structured, and the methodology part is easy to follow.\n2. The experiment section contains completed ablation studies for understanding the proposed method.", "weaknesses": "1. The overall contribution is limited. Zeroth-order optimization in Federated Learning has been already studied, and applying secure aggregation is straightforward as well. The novelty of theoretical analysis also seems limited. It would be better for the authors to summarize the comparison and improvements in terms of the algorithm side with FedZO (Fang et al., 2022).\n \n2. The motivation for applying TEE in the proposed method is unclear. The paper only mentioned that TEE is memory-constraint and back-propagation is memory-consuming as well. However, it cannot be concluded that TEE can benefit from the BP-free method. I would like to see a more detailed explanation or just an applicable explanation about utilizing TEE in the BAFFLE algorithm instead of adding it as an extension.\n \n3. Compared to other FL works, the experiments lack some settings. It seems that experiments for both i.i.d. and non-i.i.d. are conducted under client full participation settings, which is more closely to the centralized settings.", "questions": "1. What is the convergence analysis stating? Theorem 1 and 2 show that the zeroth-order gradient obtains an unbiased estimation for the true gradients with a convergence rate. However these results seem common in (centralized) zeroth-order optimizations. How do the theorems contribute to the proposed method?\n \n2. Can you report the comparison between real-world computation costs (such as wall-clock time) for zeroth-order and BP algorithms, especially when K is large, Is the proposed method really time-consuming? Ideally, as presented in the paper “BAFFLE results in approximately K/5 times the computation expense of BP-based FL”, but it would be convincing if there is experimental evidence. Additionally, what is the number of local steps for your BP baseline (FedAvg or FedSGD)? It is crucial to point this number of local steps out. A simple example: Suppose “BAFFLE results in approximately K/5 times the computation expense of BP-based FL” holds correctly. If conducting K=5000 steps of BAFFLE for 20 epochs can achieve comparable accuracy as conducting 100 steps of FedAvg for 20 epochs. BAFFLE still takes 10 times more computational costs than FedAvg.\n \n3. The writing related to experiments should be more clear to avoid misleading. In Section 2, “...95.17% accuracy on MNIST with 20 communication rounds versus 83.58% for FedZO with 1,000 rounds. “ How much is the K value in this comparison for BAFFLE? According to Section 4, “Specifically, we use Adam to train a random initialized model…”, what this Adam optimizer uses for? Does the optimizer for local training keep the same for BAFFLE and BP-baseline?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes BAFFLE, a zeroth-order federated learning method with secure aggregation and TEE. BAFFLE benefits from the BP-free local optimization process and thus is more memory-efficient than traditional BP-based algorithms. Empirical studies demonstrate the effectiveness and the low memory costs of the proposed BAFFLE.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is well-structured, and the methodology part is easy to follow.\n2. The experiment section contains completed ablation studies for understanding the proposed method.", "weaknesses": "1. The overall contribution is limited. Zeroth-order optimization in Federated Learning has been already studied, and applying secure aggregation is straightforward as well. The novelty of theoretical analysis also seems limited. It would be better for the authors to summarize the comparison and improvements in terms of the algorithm side with FedZO (Fang et al., 2022).\n \n2. The motivation for applying TEE in the proposed method is unclear. The paper only mentioned that TEE is memory-constraint and back-propagation is memory-consuming as well. However, it cannot be concluded that TEE can benefit from the BP-free method. I would like to see a more detailed explanation or just an applicable explanation about utilizing TEE in the BAFFLE algorithm instead of adding it as an extension.\n \n3. Compared to other FL works, the experiments lack some settings. It seems that experiments for both i.i.d. and non-i.i.d. are conducted under client full participation settings, which is more closely to the centralized settings.", "questions": "1. What is the convergence analysis stating? Theorem 1 and 2 show that the zeroth-order gradient obtains an unbiased estimation for the true gradients with a convergence rate. However these results seem common in (centralized) zeroth-order optimizations. How do the theorems contribute to the proposed method?\n \n2. Can you report the comparison between real-world computation costs (such as wall-clock time) for zeroth-order and BP algorithms, especially when K is large, Is the proposed method really time-consuming? Ideally, as presented in the paper “BAFFLE results in approximately K/5 times the computation expense of BP-based FL”, but it would be convincing if there is experimental evidence. Additionally, what is the number of local steps for your BP baseline (FedAvg or FedSGD)? It is crucial to point this number of local steps out. A simple example: Suppose “BAFFLE results in approximately K/5 times the computation expense of BP-based FL” holds correctly. If conducting K=5000 steps of BAFFLE for 20 epochs can achieve comparable accuracy as conducting 100 steps of FedAvg for 20 epochs. BAFFLE still takes 10 times more computational costs than FedAvg.\n \n3. The writing related to experiments should be more clear to avoid misleading. In Section 2, “...95.17% accuracy on MNIST with 20 communication rounds versus 83.58% for FedZO with 1,000 rounds. “ How much is the K value in this comparison for BAFFLE? According to Section 4, “Specifically, we use Adam to train a random initialized model…”, what this Adam optimizer uses for? Does the optimizer for local training keep the same for BAFFLE and BP-baseline?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698943867619}, {"id": "BmaGJA7iSW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7596/Reviewer_vF9m"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a federated learning method called BAFFLE which replaces backpropagation by multiple forward or inference processes to estimate gradients. In this framework, each client downloads random seeds and global parameter updates from the server, generates perturbations locally based on the seeds, executes multiple forward propagations or inferences to compute loss differences, and then uploads the loss differences back to the server. Besides, since BAFFLE only utilizes forward propagations, which are memory efficient and do not require auto-differentiation, trusted execution environments (TEEs) can be used in this model to prevent white-box evasion. In the experiments, BAFFLE is used to train deep models from scratch or to finetune pre-trained models. Compared with conventional Federated Learning, the performance is acceptable.", "review_text": "This paper introduces a federated learning method called BAFFLE which replaces backpropagation by multiple forward or inference processes to estimate gradients. In this framework, each client downloads random seeds and global parameter updates from the server, generates perturbations locally based on the seeds, executes multiple forward propagations or inferences to compute loss differences, and then uploads the loss differences back to the server. Besides, since BAFFLE only utilizes forward propagations, which are memory efficient and do not require auto-differentiation, trusted execution environments (TEEs) can be used in this model to prevent white-box evasion. In the experiments, BAFFLE is used to train deep models from scratch or to finetune pre-trained models. Compared with conventional Federated Learning, the performance is acceptable.", "strengths": "Using zeroth-order optimization, this model replaces backpropagation with multiple forward or inference processes to obtain a stochastic estimation of gradients. Since this model is backpropagation-free, the communication costs and the computational and storage overhead for clients can be reduced. Because of this, TEEs which are highly memory-constrained can also be applied.", "weaknesses": "Some notations are not well defined when using them. The organization of this paper can be improved. BP baselines are not very clear in each experiment (FedAvg or FedSDG). The tested models are not large. The comparison between BAFFLE and state-of-the-art algorithms is missing. In PRELIMINARIES Zeroth-order FL, some numbers are provided to compare BAFFLE with FedZO, but in EXPERIMENT the comparison to other related works is not well presented. The robustness experiments are not convincing enough (not quantitative) and there is no comparison between BAFFLE and conventional FL.", "questions": "1. Have you considered BAFFEL's performance when the models are more complicated and have more parameters? \n\n2. What if the number of clients is much larger?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a federated learning method called BAFFLE which replaces backpropagation by multiple forward or inference processes to estimate gradients. In this framework, each client downloads random seeds and global parameter updates from the server, generates perturbations locally based on the seeds, executes multiple forward propagations or inferences to compute loss differences, and then uploads the loss differences back to the server. Besides, since BAFFLE only utilizes forward propagations, which are memory efficient and do not require auto-differentiation, trusted execution environments (TEEs) can be used in this model to prevent white-box evasion. In the experiments, BAFFLE is used to train deep models from scratch or to finetune pre-trained models. Compared with conventional Federated Learning, the performance is acceptable.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "Using zeroth-order optimization, this model replaces backpropagation with multiple forward or inference processes to obtain a stochastic estimation of gradients. Since this model is backpropagation-free, the communication costs and the computational and storage overhead for clients can be reduced. Because of this, TEEs which are highly memory-constrained can also be applied.", "weaknesses": "Some notations are not well defined when using them. The organization of this paper can be improved. BP baselines are not very clear in each experiment (FedAvg or FedSDG). The tested models are not large. The comparison between BAFFLE and state-of-the-art algorithms is missing. In PRELIMINARIES Zeroth-order FL, some numbers are provided to compare BAFFLE with FedZO, but in EXPERIMENT the comparison to other related works is not well presented. The robustness experiments are not convincing enough (not quantitative) and there is no comparison between BAFFLE and conventional FL.", "questions": "1. Have you considered BAFFEL's performance when the models are more complicated and have more parameters? \n\n2. What if the number of clients is much larger?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698714460457}, {"id": "A9bsNyFYOl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7596/Reviewer_MdFM"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose BAFFLE, a backpropagation-free approach for federated learning. The key idea is to let each client perturb the model multiple times and conduct forward propagation for all models. Then, the server aggregates the loss differences to estimate the gradient. Experimental results demonstrates the effectiveness of BAFFLE in terms of memory efficiency, although it achieves lower performance than the full backpropagation.", "review_text": "The authors propose BAFFLE, a backpropagation-free approach for federated learning. The key idea is to let each client perturb the model multiple times and conduct forward propagation for all models. Then, the server aggregates the loss differences to estimate the gradient. Experimental results demonstrates the effectiveness of BAFFLE in terms of memory efficiency, although it achieves lower performance than the full backpropagation.", "strengths": "1. The authors tackles an important problem, which is to avoid backpropagation in resource-limited federated learning clients.\n\n2. The idea of the paper is easy to follow.", "weaknesses": "1. It is not clear what the theoretical results in Theorems 3.1 & 3.2 are exactly saying. The authors should provide more details on what is the theorem saying, and why it is guaranteeing the convergence of the algorithm. Moreover, are the authors assuming a strongly convex function or a non-convex loss function? Or does it holds for both?\n\n2. In experiments, the authors show that BAFFLE is more resource efficient than full backpropagation, while achieving a lower accuracy, which is natural. Now the question is, in which environments should we use BAFFLE, and when is full backpropagation better? There is no clear guidance on this. One thing the authors can investigate is the accuracy vs. resource (e.g., client-side computation or overall delay, etc) plot. Depending on the system parameters, it would be meaningful if the authors can provide when BAFFLE should be utilized instead of full backprop.\n\n3. Does BAFFLE also works with partial client participation?\n\n4. I'm currently not sure why evaluate the scheme in a federated learning setup. Can this idea be applied in a centralized setup where all data samples are gathered in a single user? What is the performance under this setup?", "questions": "Please see weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose BAFFLE, a backpropagation-free approach for federated learning. The key idea is to let each client perturb the model multiple times and conduct forward propagation for all models. Then, the server aggregates the loss differences to estimate the gradient. Experimental results demonstrates the effectiveness of BAFFLE in terms of memory efficiency, although it achieves lower performance than the full backpropagation.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The authors tackles an important problem, which is to avoid backpropagation in resource-limited federated learning clients.\n\n2. The idea of the paper is easy to follow.", "weaknesses": "1. It is not clear what the theoretical results in Theorems 3.1 & 3.2 are exactly saying. The authors should provide more details on what is the theorem saying, and why it is guaranteeing the convergence of the algorithm. Moreover, are the authors assuming a strongly convex function or a non-convex loss function? Or does it holds for both?\n\n2. In experiments, the authors show that BAFFLE is more resource efficient than full backpropagation, while achieving a lower accuracy, which is natural. Now the question is, in which environments should we use BAFFLE, and when is full backpropagation better? There is no clear guidance on this. One thing the authors can investigate is the accuracy vs. resource (e.g., client-side computation or overall delay, etc) plot. Depending on the system parameters, it would be meaningful if the authors can provide when BAFFLE should be utilized instead of full backprop.\n\n3. Does BAFFLE also works with partial client participation?\n\n4. I'm currently not sure why evaluate the scheme in a federated learning setup. Can this idea be applied in a centralized setup where all data samples are gathered in a single user? What is the performance under this setup?", "questions": "Please see weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698480002221}, {"id": "4oy7NATMQ5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7596/Reviewer_9br4"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work proposes BAFFLE, a method for zero-order optimization in the federated setting. The main motivation behind zero-order optimization is its memory efficiency, as no activations need to be stored for a backward pass, it is compatible with hardware designed for inference only and, furthermore, its memory efficiency allows for training within a trusted execution environment (TEE). The recipe is quite simple, in that the server communicates a random seed and a model to the clients and then each of the clients does $2K$ forward passes with weights $w \\pm \\delta_k$ and their local data.  The loss differences are then communicated and aggregated at the server where, together with the perturbations $\\delta_k$, an approximation to the model gradient can be computed. The authors provide guidelines that allow such a scheme to be successful and then show, for moderately large values of $K$, that BAFFLE can reach performance comparable to vanilla backpropagation.", "review_text": "This work proposes BAFFLE, a method for zero-order optimization in the federated setting. The main motivation behind zero-order optimization is its memory efficiency, as no activations need to be stored for a backward pass, it is compatible with hardware designed for inference only and, furthermore, its memory efficiency allows for training within a trusted execution environment (TEE). The recipe is quite simple, in that the server communicates a random seed and a model to the clients and then each of the clients does $2K$ forward passes with weights $w \\pm \\delta_k$ and their local data.  The loss differences are then communicated and aggregated at the server where, together with the perturbations $\\delta_k$, an approximation to the model gradient can be computed. The authors provide guidelines that allow such a scheme to be successful and then show, for moderately large values of $K$, that BAFFLE can reach performance comparable to vanilla backpropagation.", "strengths": "The paper overall attempts to tackle an interesting problem that is useful in practice, especially given the success of large models. It is well written and clear in the description of the method and motivation. As for other strengths:\n- The memory efficiency of BAFFLE is especially useful in the context of training large models on resource constrained devices. This memory efficiency also allows for training within a TEE. \n- The zero-order nature might allow BAFFLE to be compatible with inference only models and accelerators.  \n- The authors also propose useful guidelines for BAFFLE, which I believe can be useful for zero-order optimization in general. \n- The authors provide a theoretical convergence analysis of BAFFLE. \n- Experiments cover a decent variety of settings, from different settings of data non-i.i.d.-ness to computation / memory efficiency and robustness", "weaknesses": "I believe this work has relatively limited novelty, especially given prior work. More specifically, the authors discuss both forward finite differences and central finite differences but in the experiments they mention using forward finite differences, due to it being better for the specific amount of sample budget they consider. As far as I see, forward finite differences is essentially the exact same thing as Evolution Strategies (e.g., see [1]) which has already been considered for training neural network type of models in parallel (again, see [1]). \n\nFurthermore, some of the claims of the method could use more details. For example, the authors mention that BAFFLE can apply to models optimised for inference, e.g., models that have been quantised and pruned. However, adding Gaussian noise to the weights can break both the quantization and (unstructured) pruning, so it is unclear whether these models would be able to run on their dedicated accelerators (which sometimes require weights to, e.g., be properly qunatized in a uniform grid). The authors also argue that secure aggregation (SAG) is compatible with BAFFLE, however, as far as I know, SAG requires the inputs to be quantised and this is not discussed in the main text. \n\nFinally, if memory constraint is an issue, I believe an important baseline is missing and that is gradient checkpointing for BP, which can also reduce memory costs at the expense of more compute (which is the tradeoff BAFFLE also has).  \n\n[1] Evolution Strategies as a Scalable Alternative to Reinforcement Learning, Salimans et al., 2017", "questions": "I have the following questions and suggestions\n- Do the authors quantise the loss differences in order to allow for SAG? If yes, what is the scheme? How does it affect the convergence behaviour? \n- The convergence based on a finite number of forward passes depends on a constant $C_0$. How large is that constant in practice? \n- The harms of batch normalisation and its replacement by group normalisation, that the authors argue for as a guideline, is not new and has been discussed in various other works in federated learning, e.g, [2]. I would suggest that the authors properly attribute these prior works.\n- Are the experiments shown with full client participation? Partial client participation is more common in cross-device settings (which is usually the case where resource constraints appear), so it would be interesting to see results with that, especially given that it increases gradient variance. \n- I would suggest to the authors to elaborate on any differences with respect to [1].\n- I would suggest to the authors to elaborate about the feasibility of evaluating quantized / pruned models with Gaussian noise on their weights on dedicated hardware accelerators.\n\n[2] The non-IID data quagmire of decentralized machine learning, Hsieh et al., 2019", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes BAFFLE, a method for zero-order optimization in the federated setting. The main motivation behind zero-order optimization is its memory efficiency, as no activations need to be stored for a backward pass, it is compatible with hardware designed for inference only and, furthermore, its memory efficiency allows for training within a trusted execution environment (TEE). The recipe is quite simple, in that the server communicates a random seed and a model to the clients and then each of the clients does $2K$ forward passes with weights $w \\pm \\delta_k$ and their local data.  The loss differences are then communicated and aggregated at the server where, together with the perturbations $\\delta_k$, an approximation to the model gradient can be computed. The authors provide guidelines that allow such a scheme to be successful and then show, for moderately large values of $K$, that BAFFLE can reach performance comparable to vanilla backpropagation.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper overall attempts to tackle an interesting problem that is useful in practice, especially given the success of large models. It is well written and clear in the description of the method and motivation. As for other strengths:\n- The memory efficiency of BAFFLE is especially useful in the context of training large models on resource constrained devices. This memory efficiency also allows for training within a TEE. \n- The zero-order nature might allow BAFFLE to be compatible with inference only models and accelerators.  \n- The authors also propose useful guidelines for BAFFLE, which I believe can be useful for zero-order optimization in general. \n- The authors provide a theoretical convergence analysis of BAFFLE. \n- Experiments cover a decent variety of settings, from different settings of data non-i.i.d.-ness to computation / memory efficiency and robustness", "weaknesses": "I believe this work has relatively limited novelty, especially given prior work. More specifically, the authors discuss both forward finite differences and central finite differences but in the experiments they mention using forward finite differences, due to it being better for the specific amount of sample budget they consider. As far as I see, forward finite differences is essentially the exact same thing as Evolution Strategies (e.g., see [1]) which has already been considered for training neural network type of models in parallel (again, see [1]). \n\nFurthermore, some of the claims of the method could use more details. For example, the authors mention that BAFFLE can apply to models optimised for inference, e.g., models that have been quantised and pruned. However, adding Gaussian noise to the weights can break both the quantization and (unstructured) pruning, so it is unclear whether these models would be able to run on their dedicated accelerators (which sometimes require weights to, e.g., be properly qunatized in a uniform grid). The authors also argue that secure aggregation (SAG) is compatible with BAFFLE, however, as far as I know, SAG requires the inputs to be quantised and this is not discussed in the main text. \n\nFinally, if memory constraint is an issue, I believe an important baseline is missing and that is gradient checkpointing for BP, which can also reduce memory costs at the expense of more compute (which is the tradeoff BAFFLE also has).  \n\n[1] Evolution Strategies as a Scalable Alternative to Reinforcement Learning, Salimans et al., 2017", "questions": "I have the following questions and suggestions\n- Do the authors quantise the loss differences in order to allow for SAG? If yes, what is the scheme? How does it affect the convergence behaviour? \n- The convergence based on a finite number of forward passes depends on a constant $C_0$. How large is that constant in practice? \n- The harms of batch normalisation and its replacement by group normalisation, that the authors argue for as a guideline, is not new and has been discussed in various other works in federated learning, e.g, [2]. I would suggest that the authors properly attribute these prior works.\n- Are the experiments shown with full client participation? Partial client participation is more common in cross-device settings (which is usually the case where resource constraints appear), so it would be interesting to see results with that, especially given that it increases gradient variance. \n- I would suggest to the authors to elaborate on any differences with respect to [1].\n- I would suggest to the authors to elaborate about the feasibility of evaluating quantized / pruned models with Gaussian noise on their weights on dedicated hardware accelerators.\n\n[2] The non-IID data quagmire of decentralized machine learning, Hsieh et al., 2019", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698418991977}], "openreview_url": "https://openreview.net/forum?id=pogJXugbN8", "arxiv_id": "2301.12195", "paper_pdf": "papers/pogJXugbN8.pdf", "paper_pdf_sha256": "a0cf61126a3f28bb76de4ec658d5bdf238284693e3dcfa510292575be86fb391", "paper_pdf_bytes": 907149, "paper_pdf_source": "openreview", "code_url": "https://github.com/FengHZ/BAFFLE", "code_repository": "FengHZ/BAFFLE", "code_commit": "b59374505eb21c6d3ae2c6a513556c05d2d1dce9", "code_archive": "repos/pogJXugbN8.zip", "code_archive_sha256": "06c2d593027c587a16f587b7a7e4aff8f309e4f23fb1e891870084d8ac0a82c9", "code_archive_bytes": 15249, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 15, "github_languages": {"Python": 45513}, "github_archived": false, "github_pushed_at": "2023-02-09T11:41:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/does-federated-learning-really-need"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "w1w4dGJ4qV", "year": 2023, "status": "rejected", "title": "The Benefits of Model-Based Generalization in Reinforcement Learning", "authors": ["Kenny John Young", "Aditya Ramesh", "Louis Kirsch", "Jürgen Schmidhuber"], "authorids": ["~Kenny_John_Young1", "~Aditya_Ramesh2", "~Louis_Kirsch1", "~Jürgen_Schmidhuber1"], "authors_source": "OpenReview API", "abstract": "Model-Based Reinforcement Learning (RL) is widely believed to have the potential to improve sample efficiency by allowing an agent to synthesize large amounts of imagined experience. Experience Replay (ER) can be considered a simple kind of model, which has proved extremely effective at improving the stability and efficiency of deep RL. In principle, a learned parametric model could improve on ER by generalizing from real experience to augment the dataset with additional plausible experience. However, owing to the many design choices involved in empirically successful algorithms, it can be very hard to establish where the benefits are actually coming from. Here, we provide theoretical and empirical insight into when, and how, we can expect data generated by a learned model to be useful. First, we provide a general theorem motivating how learning a model as an intermediate step can narrow down the set of possible value functions more than learning a value function directly from data using the Bellman equation.  Second, we provide an illustrative example showing empirically how a similar effect occurs in a more concrete setting with neural network function approximation.  Finally, we provide extensive experiments showing the benefit of model-based learning for online RL in environments with combinatorial complexity, but factored structure that allows a learned model to generalize. In these experiments, we take care to control for other factors in order to isolate, insofar as possible, the benefit of using experience generated by a learned model relative to ER alone.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "BkIKylM7Fq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper437/Reviewer_5uCS"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work formalizes the generalization benefits of model-based RL methods in the context of deterministic episodic MDPs who's transitions probability belong to a known hypothesis class. The authors then discuss the intuition gained from their theory in various illustrative environments and take a closer look at the case when transitions can be assumed to be factored.", "review_text": "Assuming the novelty isn't brought in question, this paper seems like a clear accept. Ideas are well motivated and of interest to RL.", "strengths": "This work examines an important open question about the benefits of model-based RL. Although these methods are commonly believed to be more sample efficient, formal results on the subject are sparse. This work will likely be of interest to the RL community.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work formalizes the generalization benefits of model-based RL methods in the context of deterministic episodic MDPs who's transitions probability belong to a known hypothesis class. The authors then discuss the intuition gained from their theory in various illustrative environments and take a closer look at the case when transitions can be assumed to be factored.", "strength_and_weaknesses": "This work examines an important open question about the benefits of model-based RL. Although these methods are commonly believed to be more sample efficient, formal results on the subject are sparse. This work will likely be of interest to the RL community.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written. The contributions are well motivated and, to the best of my knowledge, novel.\n\nIt might be worth mentioning [1] in the related work (e.g., as part of paragraph 3, p. 9). Although that work doesn't consider generalization, it does compare the gradient dynamics and convergence rates when using \"end-to-end\" model-based method to when the value function is encoded directly with an equally expressive estimator.\n\n[1] Clement Gehring, Kenji Kawaguchi, Jiaoyang Huang, Leslie Pack Kaelbling. Understanding End-to-End Model-Based Reinforcement Learning Methods as Implicit Parameterization. NeurIPS, 2021", "summary_of_the_review": "Assuming the novelty isn't brought in question, this paper seems like a clear accept. Ideas are well motivated and of interest to RL.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667342208347}, {"id": "vm9WO0rQa-", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper437/Reviewer_JiYh"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper provides a high-level characterization of the properties of RL tasks in which model-based algorithms can be expected to improve generalization to new observations. Two primary quantities arise from this analysis: first, the state space should exhibit a simple underlying factored structure which the inductive bias of the model is well-suited to learning; second, the distribution of rewarding trajectories that an agent sees should be sufficiently sparse as to make learning from Bellman backups difficult, but still provide non-zero reward density so that there is some signal for a learned model to leverage.", "review_text": "Overall, this paper provides a nice set of interpretable examples that highlight the different behaviour of model-free and different model-learning algorithms. One of these examples in particular illustrates how the smoothness bias in DNN function approximators can facilitate learning and generalization in model-based algorithms. The paper suffers two main limitations: first, its main theoretical result is a relatively weak statement, and doesn’t directly offer a concrete set of implications for the effect of model-learning methods on generalization. Second, it is not clear whether the empirical findings concerning the behaviour of model-based and model-free methods will generalize outside of the small environments presented in the paper.\n", "strengths": "### Strengths\n\n- The paper provides nice intuitive examples of generalization in maze environments that highlight the limitations of model-based and value-based RL algorithms.\n- The paper is well-written and does a good job at summarizing and emphasizing the key take-aways from each section.\n- The experiment setup is clear.\nThe analysis of different learned models was enlightening as it highlighted the importance of the inductive bias of the world model, and because it provided a striking but also simple setting where a learning algorithm benefits more from an imperfect learned model than from the perfect world model. I appreciated that the authors went further to validate their hypothesis that the learned models were benefitting from reward smoothing in these contexts\n- Theorem 1 provides a nice contrast to the findings of Parr et al. in the linear setting. \n\n### Weaknesses\n\n- The paper focuses on a handful of small toy environments. While this is useful to guide the reader’s intuition and to make the experimental findings interpretable, it also limits the generality of the paper’s conclusions.\n- The degree to which the findings in this paper also apply to larger scale models is an open question. While the toy settings considered in the paper do a good job of illustrating some environment qualities can than influence the efficacy of simple model-based algorithms, it’s not clear to me that these are the only parameters that can influence the relative performance of model-learning methods in more challenging tasks. For example, the influence of model learning on exploration, or on the stability of the underlying learning method, could also play a role in an agent’s final performance in these settings. It’s not clear to me whether the environment features and model smoothing properties identified in this paper would be a dominant factor in performance compared to these other factors. \n\n- I suspect that the value-based agents might be worse at generalization than policy-based agents based on prior literature on e.g. the ProcGen suite (Raileanu et al., 2020), and would have liked to see the inclusion of a policy-gradient method in the set of baselines. \n- The presentation of the proof of theorem 1 could be improved: the definition of the transition dynamics for $S_t[0]$ for example, seem to be defining different quantities in the top and bottom line of the bracket, and the hypothesis class $H$ from which the q-functions are selected is not defined. I assume it is the set of all real-valued functions over the state space, but this should be stated explicitly. \n- While Theorem 1 is a nice observation, the statement itself is relatively weak. It does not, for example, provide a means of quantifying the degree to which the hypothesis space can be reduced by model classes with certain properties. Nor does it address settings where the algorithm may need to trade off between a limited set of environment steps and the approximation error of a model. This limits the significance of the theoretical contribution of the paper.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper provides a high-level characterization of the properties of RL tasks in which model-based algorithms can be expected to improve generalization to new observations. Two primary quantities arise from this analysis: first, the state space should exhibit a simple underlying factored structure which the inductive bias of the model is well-suited to learning; second, the distribution of rewarding trajectories that an agent sees should be sufficiently sparse as to make learning from Bellman backups difficult, but still provide non-zero reward density so that there is some signal for a learned model to leverage.", "strength_and_weaknesses": "### Strengths\n\n- The paper provides nice intuitive examples of generalization in maze environments that highlight the limitations of model-based and value-based RL algorithms.\n- The paper is well-written and does a good job at summarizing and emphasizing the key take-aways from each section.\n- The experiment setup is clear.\nThe analysis of different learned models was enlightening as it highlighted the importance of the inductive bias of the world model, and because it provided a striking but also simple setting where a learning algorithm benefits more from an imperfect learned model than from the perfect world model. I appreciated that the authors went further to validate their hypothesis that the learned models were benefitting from reward smoothing in these contexts\n- Theorem 1 provides a nice contrast to the findings of Parr et al. in the linear setting. \n\n### Weaknesses\n\n- The paper focuses on a handful of small toy environments. While this is useful to guide the reader’s intuition and to make the experimental findings interpretable, it also limits the generality of the paper’s conclusions.\n- The degree to which the findings in this paper also apply to larger scale models is an open question. While the toy settings considered in the paper do a good job of illustrating some environment qualities can than influence the efficacy of simple model-based algorithms, it’s not clear to me that these are the only parameters that can influence the relative performance of model-learning methods in more challenging tasks. For example, the influence of model learning on exploration, or on the stability of the underlying learning method, could also play a role in an agent’s final performance in these settings. It’s not clear to me whether the environment features and model smoothing properties identified in this paper would be a dominant factor in performance compared to these other factors. \n\n- I suspect that the value-based agents might be worse at generalization than policy-based agents based on prior literature on e.g. the ProcGen suite (Raileanu et al., 2020), and would have liked to see the inclusion of a policy-gradient method in the set of baselines. \n- The presentation of the proof of theorem 1 could be improved: the definition of the transition dynamics for $S_t[0]$ for example, seem to be defining different quantities in the top and bottom line of the bracket, and the hypothesis class $H$ from which the q-functions are selected is not defined. I assume it is the set of all real-valued functions over the state space, but this should be stated explicitly. \n- While Theorem 1 is a nice observation, the statement itself is relatively weak. It does not, for example, provide a means of quantifying the degree to which the hypothesis space can be reduced by model classes with certain properties. Nor does it address settings where the algorithm may need to trade off between a limited set of environment steps and the approximation error of a model. This limits the significance of the theoretical contribution of the paper.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is overall well-written and clear. While model-based RL has been well-studied in the literature, the generalization of learned models to new observations has received less attention. I have not seen the paper’s main theorem elsewhere in the literature.", "summary_of_the_review": "Overall, this paper provides a nice set of interpretable examples that highlight the different behaviour of model-free and different model-learning algorithms. One of these examples in particular illustrates how the smoothness bias in DNN function approximators can facilitate learning and generalization in model-based algorithms. The paper suffers two main limitations: first, its main theoretical result is a relatively weak statement, and doesn’t directly offer a concrete set of implications for the effect of model-learning methods on generalization. Second, it is not clear whether the empirical findings concerning the behaviour of model-based and model-free methods will generalize outside of the small environments presented in the paper.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666662480138}, {"id": "mtAUORtz_FQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper437/Reviewer_8CUu"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper investigates the question of when and why model generalization is better than value function generalization in reinforcement learning. The paper starts with a theoretical result that gives the intuition on this question: reinforcement learning with a learned model of the transition dynamics can reduce the optimal action value function space than simply solving bellman constraints with experience replay. Then the authors provide some hypothesis/conditions that tends to favor model-based RL, and then justify them by testing on some designed toy experiments.", "review_text": "Interesting question, nice examples and discussion of empirical observation. Needs more thorough and rigorous theoretical study on some key questions.", "strengths": "Strength:\n- An interesting and important question to study, as when and why model-based RL is guaranteed to be better than experience replay can significantly influence us to design practical RL algorithms for real-world applications\n- Well-written and easy to follow, with many concrete examples and interesting discussions of empirical observations\n- Good introduction of the background\n\nWeaknesses/Questions:\n- The first motivating example in the introduction (an agent trying to cross a stream to reach the food) is not very clear and convincing. It is hard to understand the formulation of the state (is it a bandit problem given your state definition?), and what is the data (some state+reward pairs?) in your experience replay, and why model-free learning clearly/definitely fails (it can also generalize via value function)?\n- Perhaps the main rigorous answer is Theorem 1, where I was expecting to see when and why model-based RL is better. However, I found the result is not surprising/contains enough theoretical insights. It is basically saying, if we are given a set of hypothesis class for the transition dynamics, and we can successfully identify one that perfectly matches the data, then using that transition function (probably optimal if data is enough) to solve Bellman optimal equation is better than directly using the data to solve. This might be trivial and in practice, the key question is actually what theoretical guarantee we can achieve when the approximate model makes an error (upper bounded by some constant). If the data D is very limited and the best model consistent with D is still far from the true dynamics, then is model-based RL always better (since it incurs compounding error)? Another essential question to study is how will the model error propagates to solving Bellman optimality equations (the data is always from true dynamics but generated ones can be wrong)?\n- It seems most above important questions are not rigorously answered. The following sections provide nice examples and hypothesis on the conditions. But I feel they should be an empirical evidence to some more general theorems. I was expecting a formal sufficient (maybe not necessary) and hopefully general enough condition for when model generalization + compounding error is better than ER.\nTheoretically this should depend on specific algorithm, data collecting policy and sample size, etc. Otherwise, it may be hard to summarize some general principles that guides the design of algorithm for more complicated and realistic environments.\n- The observation that sometimes learned model is even better than the true model is very interesting. In principle, even using more smooth model can help exploration or optimization, this should eventually lead to bias. Will this be an issue and can we alleviate it using annealing (gradually reduce to true model)? In case of no bias, it seems to be something like a reward shaping or simply a coincidence that two MDPs have same solution. This may deserve more thorough study.\n- \"factored structure\" appears many times but seems fuzzy. Can you give a formal definition?", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper investigates the question of when and why model generalization is better than value function generalization in reinforcement learning. The paper starts with a theoretical result that gives the intuition on this question: reinforcement learning with a learned model of the transition dynamics can reduce the optimal action value function space than simply solving bellman constraints with experience replay. Then the authors provide some hypothesis/conditions that tends to favor model-based RL, and then justify them by testing on some designed toy experiments.", "strength_and_weaknesses": "Strength:\n- An interesting and important question to study, as when and why model-based RL is guaranteed to be better than experience replay can significantly influence us to design practical RL algorithms for real-world applications\n- Well-written and easy to follow, with many concrete examples and interesting discussions of empirical observations\n- Good introduction of the background\n\nWeaknesses/Questions:\n- The first motivating example in the introduction (an agent trying to cross a stream to reach the food) is not very clear and convincing. It is hard to understand the formulation of the state (is it a bandit problem given your state definition?), and what is the data (some state+reward pairs?) in your experience replay, and why model-free learning clearly/definitely fails (it can also generalize via value function)?\n- Perhaps the main rigorous answer is Theorem 1, where I was expecting to see when and why model-based RL is better. However, I found the result is not surprising/contains enough theoretical insights. It is basically saying, if we are given a set of hypothesis class for the transition dynamics, and we can successfully identify one that perfectly matches the data, then using that transition function (probably optimal if data is enough) to solve Bellman optimal equation is better than directly using the data to solve. This might be trivial and in practice, the key question is actually what theoretical guarantee we can achieve when the approximate model makes an error (upper bounded by some constant). If the data D is very limited and the best model consistent with D is still far from the true dynamics, then is model-based RL always better (since it incurs compounding error)? Another essential question to study is how will the model error propagates to solving Bellman optimality equations (the data is always from true dynamics but generated ones can be wrong)?\n- It seems most above important questions are not rigorously answered. The following sections provide nice examples and hypothesis on the conditions. But I feel they should be an empirical evidence to some more general theorems. I was expecting a formal sufficient (maybe not necessary) and hopefully general enough condition for when model generalization + compounding error is better than ER.\nTheoretically this should depend on specific algorithm, data collecting policy and sample size, etc. Otherwise, it may be hard to summarize some general principles that guides the design of algorithm for more complicated and realistic environments.\n- The observation that sometimes learned model is even better than the true model is very interesting. In principle, even using more smooth model can help exploration or optimization, this should eventually lead to bias. Will this be an issue and can we alleviate it using annealing (gradually reduce to true model)? In case of no bias, it seems to be something like a reward shaping or simply a coincidence that two MDPs have same solution. This may deserve more thorough study.\n- \"factored structure\" appears many times but seems fuzzy. Can you give a formal definition?", "clarity,_quality,_novelty_and_reproducibility": "The clarity is good. The paper will benefits from more thorough theoretical analysis (see above section for detailed comments). The empirical results should be reproducible.", "summary_of_the_review": "Interesting question, nice examples and discussion of empirical observation. Needs more thorough and rigorous theoretical study on some key questions.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666653191313}, {"id": "IPVF3CHduO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper437/Reviewer_4bHT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the benefits of model-based generalization by comparing the performance of DQN on the rollouts of a learned model, and DQN on a replay buffer. Theoretically, this paper proves that learning a transition model can be more efficient (in terms of pruning invalid solutions) than learning a Q-function using Bellman equations. Empirically, this paper proposes three environment characteristics where the model-based methods could be better, and demonstrates them using constructed toy environments. \n\n**Summary of the experiments**\n\nSetup: \n\nExperience replay (ER): Run DQN with a replay buffer of the past trajectories (called $\\mathcal{D}$)\n\nModel-based: use $\\mathcal{D}$ to train a model. And then train a DQN on fake trajectories generated by the learned model starting from a random state from $\\mathcal{D}$ (similar to MBPO).\n\nThe goal is to make the training of DQN exactly the same except for the data distribution (ER vs. model rollouts).\n\nResults: \n\nThis paper compares ER and model-based methods on toy environments in the online RL setting. The number of updates to the Q-network is the same for model-based and ER. In every update, the Q-network is trained using 320 samples from the replay buffer for ER, and 32 model rollouts with length 10 for model-based methods (so the batch size is also the same). This paper shows that for the toy environments model-based is in general better than ER (Figure 3).", "review_text": "On the plus side, the problem setup of this paper is neat and interesting, the results are clean and intuitive. On the minus side, more discussions about the practical relevance and ablation studies are needed since the results are mainly empirical. As a result, I recommend a weak rejection for the current version of this paper.", "strengths": "Strengths:\n\n- This paper studies the benefit of model-based generalization by comparing model-based rollouts with experience replay. To some extent, this comparison is much fairer because the only difference is the distribution of training data (replay buffer vs model rollouts). This approach is neat and results in some interesting theoretical results (Theorem 1).\n- The illustrative example in Section 2 is very easy-to-follow and already well demonstrates the intuitions of this paper.\n\nWeaknesses:\n\n- This paper only discusses one side of the story. In fact, model-based method outperforms experience replay in all the environments studied in this paper. Together with Theorem 1, is it possible to make the claim that model-based methods are always better than experience replay? In other words, discussions and ablation studies about the necessity of the three characteristics in Section 3 are missing. Hence, the scope of the conclusions of this paper is unclear.\n- It’s unclear to me when and how the characteristics proposed in Section 3 can be applied to benchmarking environments such as OpenAI Gym. For example, if we predict whether model-based methods are better using the intuition in this paper, does the prediction match reality? Since this is mostly an empirical paper, I would expect more discussions on the practical impact of the conclusions.\n- the experiments provide one possible explanation for why MB is better than ER --- learning the transitions can be easier because of the implicit bias of the function class (e.g., the factored structure of the transition) and therefore MB has better sample efficiency. This explanation is somewhat expected and whether it is the main factor in benchmarking experiments such as Mujoco is unclear.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the benefits of model-based generalization by comparing the performance of DQN on the rollouts of a learned model, and DQN on a replay buffer. Theoretically, this paper proves that learning a transition model can be more efficient (in terms of pruning invalid solutions) than learning a Q-function using Bellman equations. Empirically, this paper proposes three environment characteristics where the model-based methods could be better, and demonstrates them using constructed toy environments. \n\n**Summary of the experiments**\n\nSetup: \n\nExperience replay (ER): Run DQN with a replay buffer of the past trajectories (called $\\mathcal{D}$)\n\nModel-based: use $\\mathcal{D}$ to train a model. And then train a DQN on fake trajectories generated by the learned model starting from a random state from $\\mathcal{D}$ (similar to MBPO).\n\nThe goal is to make the training of DQN exactly the same except for the data distribution (ER vs. model rollouts).\n\nResults: \n\nThis paper compares ER and model-based methods on toy environments in the online RL setting. The number of updates to the Q-network is the same for model-based and ER. In every update, the Q-network is trained using 320 samples from the replay buffer for ER, and 32 model rollouts with length 10 for model-based methods (so the batch size is also the same). This paper shows that for the toy environments model-based is in general better than ER (Figure 3).", "strength_and_weaknesses": "Strengths:\n\n- This paper studies the benefit of model-based generalization by comparing model-based rollouts with experience replay. To some extent, this comparison is much fairer because the only difference is the distribution of training data (replay buffer vs model rollouts). This approach is neat and results in some interesting theoretical results (Theorem 1).\n- The illustrative example in Section 2 is very easy-to-follow and already well demonstrates the intuitions of this paper.\n\nWeaknesses:\n\n- This paper only discusses one side of the story. In fact, model-based method outperforms experience replay in all the environments studied in this paper. Together with Theorem 1, is it possible to make the claim that model-based methods are always better than experience replay? In other words, discussions and ablation studies about the necessity of the three characteristics in Section 3 are missing. Hence, the scope of the conclusions of this paper is unclear.\n- It’s unclear to me when and how the characteristics proposed in Section 3 can be applied to benchmarking environments such as OpenAI Gym. For example, if we predict whether model-based methods are better using the intuition in this paper, does the prediction match reality? Since this is mostly an empirical paper, I would expect more discussions on the practical impact of the conclusions.\n- the experiments provide one possible explanation for why MB is better than ER --- learning the transitions can be easier because of the implicit bias of the function class (e.g., the factored structure of the transition) and therefore MB has better sample efficiency. This explanation is somewhat expected and whether it is the main factor in benchmarking experiments such as Mujoco is unclear.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written in general, except for a few ambiguities:\n- The second bullet point in Section 3 states that “The return should depend sharply on the policy”, and it’s unclear to me what “sharply” means here. Does it mean that the return of a policy fluctuates a lot? Then does the sparse reward MDP (e.g., the PanFlute environment) satisfy this statement? \n- How large is the horizon length in all the environments?\n\nSome references are missing in this paper. E.g., [1] also discusses when model-based methods are better than model-free methods. On a high level, both [1] and this paper constructs environments where the dynamics are simple functions but the corresponding Q function is complicated.\n\n[1] Dong, Kefan, et al. \"On the expressivity of neural networks for deep reinforcement learning.\" International Conference on Machine Learning. PMLR, 2020.\n", "summary_of_the_review": "On the plus side, the problem setup of this paper is neat and interesting, the results are clean and intuitive. On the minus side, more discussions about the practical relevance and ablation studies are needed since the results are mainly empirical. As a result, I recommend a weak rejection for the current version of this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666392326978}], "openreview_url": "https://openreview.net/forum?id=w1w4dGJ4qV", "arxiv_id": "2211.02222", "paper_pdf": "papers/w1w4dGJ4qV.pdf", "paper_pdf_sha256": "e2fcee91ff9096c4126dc82e6194fdf8b55a35e72f988e956fd8b47e5826df6d", "paper_pdf_bytes": 15819893, "paper_pdf_source": "openreview", "code_url": "https://github.com/kenjyoung/Model_Generalization_Code_supplement", "code_repository": "kenjyoung/Model_Generalization_Code_supplement", "code_commit": "dd3c5a950af69ea49cadbcea1315cb193a2e1c08", "code_archive": "repos/w1w4dGJ4qV.zip", "code_archive_sha256": "2c60980fe20cf29cc1e56de67e5c1c12d1187fe3d486b3cc2d9446825020f439", "code_archive_bytes": 31426, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 18, "github_languages": {"Python": 107954}, "github_archived": false, "github_pushed_at": "2023-01-13T19:03:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-benefits-of-model-based-generalization-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-3Qj7Jl6UP5", "year": 2022, "status": "rejected", "title": "The magnitude vector of images", "authors": ["Michael F Adamer", "Leslie O'Bray", "Edward De Brouwer", "Bastian Rieck", "Karsten Borgwardt"], "authorids": ["~Michael_F_Adamer1", "~Leslie_O'Bray1", "~Edward_De_Brouwer1", "~Bastian_Rieck1", "~Karsten_Borgwardt2"], "authors_source": "OpenReview API", "abstract": "The magnitude of a finite metric space is a recently-introduced invariant quantity. Despite beneficial theoretical and practical properties, such as a general utility for outlier detection, and a close connection to Laplace radial basis kernels, magnitude has received little attention by the machine learning community so far. In this work, we investigate the properties of magnitude on individual images, with each image forming its own metric space. We show that the known properties of outlier detection translate to edge detection in images and we give supporting theoretical justifications. In addition, we provide a proof of concept of its utility by using a novel magnitude layer to defend against adversarial attacks. Since naive magnitude calculations may be computationally prohibitive, we introduce an algorithm that leverages the regular structure of images to dramatically reduce the computational cost.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "3d-vQ94t9mc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3000/Reviewer_WNxv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper discuss a concept of magnitude vector of images. The authors provided a fast algorithm for computing the magnitude vector, and discuss the potential applications in edge detection and adversarial robustness.", "review_text": "Strength:\nThe paper discuss a concept that is relatively new in the machine learning community. The magnitude of a metric space has been studied in topology, and this paper demonstrates that it can potentially be used in image-related ML problems.\n\nWeakness:\nGiven the results in the current paper, I am not fully convinced that the concept of magnitude vector is really important for learning problems. To better demonstrate its importance, the authors should provide more comprehensive experimental results.\n\nFor edge detection: the authors only provided some visual examples for the Fashion MNIST dataset. From the three images that the authors provided in the paper, the results of magnitude-based edge detection is visually worse than Canny detector. The authors did not provide any quantitative results on the quality of edge detection. I think a more rigorous study and evaluation is needed to show its benefits.\n\nFor adversarial robustness, the authors only provided results on three relatively simple datasets: MNIST, KMNIST, Fashion MNIST, with a relatively weak attack (FGSM). To demonstrate the method of the proposed approach, the authors should provide results on more datasets and more baseline algorithms.\n\n====\nAfter response: the authors did not post specific response so I decided to keep my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper discuss a concept of magnitude vector of images. The authors provided a fast algorithm for computing the magnitude vector, and discuss the potential applications in edge detection and adversarial robustness.", "main_review": "Strength:\nThe paper discuss a concept that is relatively new in the machine learning community. The magnitude of a metric space has been studied in topology, and this paper demonstrates that it can potentially be used in image-related ML problems.\n\nWeakness:\nGiven the results in the current paper, I am not fully convinced that the concept of magnitude vector is really important for learning problems. To better demonstrate its importance, the authors should provide more comprehensive experimental results.\n\nFor edge detection: the authors only provided some visual examples for the Fashion MNIST dataset. From the three images that the authors provided in the paper, the results of magnitude-based edge detection is visually worse than Canny detector. The authors did not provide any quantitative results on the quality of edge detection. I think a more rigorous study and evaluation is needed to show its benefits.\n\nFor adversarial robustness, the authors only provided results on three relatively simple datasets: MNIST, KMNIST, Fashion MNIST, with a relatively weak attack (FGSM). To demonstrate the method of the proposed approach, the authors should provide results on more datasets and more baseline algorithms.\n\n====\nAfter response: the authors did not post specific response so I decided to keep my score.", "summary_of_the_review": "The experimental results are not strong enough to demonstrate the benefits of the concept of magnitude vector of images.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636382641364}, {"id": "hPYWOj9ZOe3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3000/Reviewer_qYUC"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors consider a relatively new finite metric space quantity known as the \"magnitude\" and its applications within computer vision. \n\nThe paper proceeds with introducing and defining the magnitude quantity before exploring some key properties.  The contributes the authors outline are a study of \n* relationship between 'outlier detection' and 'edge detection'\n* introduction of 'magnitude layer' as an adversarial defence \n* new algorithm to speed up the computation of the magnitude vect", "review_text": "## Page 2\n\n#### Definition 1\nThis should read \"A finite metric space is an ordered pair...\"\n\n#### Definition 4\n- The notation to define $V$ is nonstandard and unclear.\n- The notation $M(B,d)$ is nonstandard: It should either be $M$ or $(B,d)$ unless $M$ some kind of mapping. I.e. $M=(B,d)=M(B,d) $ is not a consistent choice of notation.\n- In the definition of $B$ (as well as $V$) it would be wise to avoid an unrestricted comprehension.\n\n## Page 3\n\n> As generic $n \\times n$ matrices are invertible (i.e. subjecting any non-invertible matrix to a random perturbation will almost certainly result in an invertible matrix), we conclude that generic image metric spaces have magnitude.\n\n- This is not a conclusion, it is an assumption.\n\n> In fact, since the vectors b ∈ B of the image metric space by a factor t > 0 can be rescaled, we can define scaled metrics $d(·, ·) \\mapsto td(·, ·)$.\n\n- This only makes sense if $d$ is induced by a norm. The sentence, as written, does not make sense.\n\n#### Lemma 6\n- This is a standard metric space construction, and applies even more generally than its statement here. Its inclusion is inappropriate.\n\n> We can use this reasoning to define a filtration $F$ on the vectors of B...\n- Typo, should be $v_{ij}^{(k)}$ not $v_{ij}^{(1)}$. \n- observe the se are just the sublevel sets of the function $f$, calling this a \"filtration\" is confusing.\n\n##  Page 4\nCan you please explain the purpose of the derivations and what the results are? It is unclear what here is new or novel and what is a result from Bunch et al. (2021). \n\n\n## Pages 5 and 6\nI would an explanation here of what the new algorithm is. Additionally if you are going to claim that your algorithm is significantly faster you need to show a baseline to compare the speed with others, e.g. Bunch et al. (2021). Furthermore why have you chosen an X-ray dataset as the only dataset to present benchmarks with?\n\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors consider a relatively new finite metric space quantity known as the \"magnitude\" and its applications within computer vision. \n\nThe paper proceeds with introducing and defining the magnitude quantity before exploring some key properties.  The contributes the authors outline are a study of \n* relationship between 'outlier detection' and 'edge detection'\n* introduction of 'magnitude layer' as an adversarial defence \n* new algorithm to speed up the computation of the magnitude vect", "main_review": "## Page 2\n\n#### Definition 1\nThis should read \"A finite metric space is an ordered pair...\"\n\n#### Definition 4\n- The notation to define $V$ is nonstandard and unclear.\n- The notation $M(B,d)$ is nonstandard: It should either be $M$ or $(B,d)$ unless $M$ some kind of mapping. I.e. $M=(B,d)=M(B,d) $ is not a consistent choice of notation.\n- In the definition of $B$ (as well as $V$) it would be wise to avoid an unrestricted comprehension.\n\n## Page 3\n\n> As generic $n \\times n$ matrices are invertible (i.e. subjecting any non-invertible matrix to a random perturbation will almost certainly result in an invertible matrix), we conclude that generic image metric spaces have magnitude.\n\n- This is not a conclusion, it is an assumption.\n\n> In fact, since the vectors b ∈ B of the image metric space by a factor t > 0 can be rescaled, we can define scaled metrics $d(·, ·) \\mapsto td(·, ·)$.\n\n- This only makes sense if $d$ is induced by a norm. The sentence, as written, does not make sense.\n\n#### Lemma 6\n- This is a standard metric space construction, and applies even more generally than its statement here. Its inclusion is inappropriate.\n\n> We can use this reasoning to define a filtration $F$ on the vectors of B...\n- Typo, should be $v_{ij}^{(k)}$ not $v_{ij}^{(1)}$. \n- observe the se are just the sublevel sets of the function $f$, calling this a \"filtration\" is confusing.\n\n##  Page 4\nCan you please explain the purpose of the derivations and what the results are? It is unclear what here is new or novel and what is a result from Bunch et al. (2021). \n\n\n## Pages 5 and 6\nI would an explanation here of what the new algorithm is. Additionally if you are going to claim that your algorithm is significantly faster you need to show a baseline to compare the speed with others, e.g. Bunch et al. (2021). Furthermore why have you chosen an X-ray dataset as the only dataset to present benchmarks with?\n\n\n\n\n\n\n", "summary_of_the_review": "The reason that I have recommended this paper are several fold. \n\nChiefly it is unclear what the novel contribution of the authors are in this paper. Taking the contributions one by one: \n - It is unclear what the benefit or novelty there is from the study of outliers and edge detection\n- The improved algorithm is not explained, presented, or studied empirically with reference to any baseline.\n- The new layer the authors present involves a computationally prohibitive matrix inversion that will limit its applicability. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635852791456}, {"id": "qPcjEyVlJmR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3000/Reviewer_f2ou"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper explores the application of the quantity “magnitude of a metric space” for images. Based on prior work, magnitude is a quantity originating from category theory extended to metric spaces with surprising geometric properties (Leinster 2017). This paper defines the magnitude as a vector valued function of a metric space (Def. 3). Through defining an image as a metric space, the magnitude of a single image is defined (Def. 4). Authors then state conditions under which magnitude of an image is known to exist (Prop 5). Then the paper defines a filtration operation that computes the magnitude vector on multiple thresholded projections of the input. Section 2.3 introduces an approximate computation method for the magnitude and Section 3.1 evaluates the efficiency and error of this approximation. Section 3.2 applies the method for edge detection and compares it to Canny detector. Section 3.3 uses the magnitude for adversarial robustness and claims gains.", "review_text": "Strengths:\n- The magnitude of a metric space seems to have various interesting properties encoding other quantities of interest.\n- The fast magnitude computation method is interesting and seems to be efficient although the scale of the significance of approximation error needs more discussion.\n\nWeaknesses:\n- There is a lack of motivation for considering the magnitude as a quantity for any purpose. My understanding from prior work is that the magnitude is a rather theoretical curiosity that “turns out to encode many invariants … including, volume, capacity, dimension and intrinsic volume” (Leinster 2017). But after reading this paper as well as prior work, I still do not have an intuition as to what it is expected to encode about an image. I will expand on this issue below under the experimental limitations.\n- Theoretical contributions of the work are not clear. A few definitions, propositions and lemma are restated from prior work. Together they seem to provide expressions for computing the magnitude of a single image. However, all of these come from prior work.\n- My understanding is that the contribution in the Section 2.2 is the particular definition of metric space for an image (Def. 4). The rest of the section applies prior tools to this metric space. There is a lack of justification for this definition. For example, why should one treat individual pixels as a point and why should we use spatial coordinates in the vector representation?\n- The notation is confusing. For example, Def. 3 and Def. 1 use n to denote both the cardinality of B and the dimensionality of elements in B. Eq. 4 and 5 are also using notations with proper definition that comes from Bunch et al 2021.\n- The significance of Section 3.2 is not clear. Are authors claiming “The magnitude of an image is as good as the Canny Edge detector.”? If so, can authors clarify why this is a significant observation? Given Figure 3, I’m not sure I would agree with this claim either, as the magnitude-based approach has mostly failed on the image of the pants.\n- The adversarial robustness results in Section 3.3 need improvement considering major works on attacking discretization layers that achieve spurious robustness through gradient masking / gradient obfuscation. As tested by authors, black-box attacks are one approach, however, FGSM is not a particularly strong attack. Various other guidelines are suggested in [1,2].\n\n[1] Athalye, Anish, Nicholas Carlini, and David Wagner. \"Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples.\" International conference on machine learning. PMLR, 2018.\n[2] Florian Tramèr, Nicholas Carlini, Wieland Brendel, Aleksander Madry:\nOn Adaptive Attacks to Adversarial Example Defenses. NeurIPS 2020", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper explores the application of the quantity “magnitude of a metric space” for images. Based on prior work, magnitude is a quantity originating from category theory extended to metric spaces with surprising geometric properties (Leinster 2017). This paper defines the magnitude as a vector valued function of a metric space (Def. 3). Through defining an image as a metric space, the magnitude of a single image is defined (Def. 4). Authors then state conditions under which magnitude of an image is known to exist (Prop 5). Then the paper defines a filtration operation that computes the magnitude vector on multiple thresholded projections of the input. Section 2.3 introduces an approximate computation method for the magnitude and Section 3.1 evaluates the efficiency and error of this approximation. Section 3.2 applies the method for edge detection and compares it to Canny detector. Section 3.3 uses the magnitude for adversarial robustness and claims gains.", "main_review": "Strengths:\n- The magnitude of a metric space seems to have various interesting properties encoding other quantities of interest.\n- The fast magnitude computation method is interesting and seems to be efficient although the scale of the significance of approximation error needs more discussion.\n\nWeaknesses:\n- There is a lack of motivation for considering the magnitude as a quantity for any purpose. My understanding from prior work is that the magnitude is a rather theoretical curiosity that “turns out to encode many invariants … including, volume, capacity, dimension and intrinsic volume” (Leinster 2017). But after reading this paper as well as prior work, I still do not have an intuition as to what it is expected to encode about an image. I will expand on this issue below under the experimental limitations.\n- Theoretical contributions of the work are not clear. A few definitions, propositions and lemma are restated from prior work. Together they seem to provide expressions for computing the magnitude of a single image. However, all of these come from prior work.\n- My understanding is that the contribution in the Section 2.2 is the particular definition of metric space for an image (Def. 4). The rest of the section applies prior tools to this metric space. There is a lack of justification for this definition. For example, why should one treat individual pixels as a point and why should we use spatial coordinates in the vector representation?\n- The notation is confusing. For example, Def. 3 and Def. 1 use n to denote both the cardinality of B and the dimensionality of elements in B. Eq. 4 and 5 are also using notations with proper definition that comes from Bunch et al 2021.\n- The significance of Section 3.2 is not clear. Are authors claiming “The magnitude of an image is as good as the Canny Edge detector.”? If so, can authors clarify why this is a significant observation? Given Figure 3, I’m not sure I would agree with this claim either, as the magnitude-based approach has mostly failed on the image of the pants.\n- The adversarial robustness results in Section 3.3 need improvement considering major works on attacking discretization layers that achieve spurious robustness through gradient masking / gradient obfuscation. As tested by authors, black-box attacks are one approach, however, FGSM is not a particularly strong attack. Various other guidelines are suggested in [1,2].\n\n[1] Athalye, Anish, Nicholas Carlini, and David Wagner. \"Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples.\" International conference on machine learning. PMLR, 2018.\n[2] Florian Tramèr, Nicholas Carlini, Wieland Brendel, Aleksander Madry:\nOn Adaptive Attacks to Adversarial Example Defenses. NeurIPS 2020", "summary_of_the_review": "The paper requires major improvements to motivation, empirical results, and rigour of theoretical contributions.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635810384790}, {"id": "1pCoiEdslWz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3000/Reviewer_cxFw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper applies the machinery of magnitude vectors of metric spaces to extract information from individual images (which can be themselves be seen as finite metric spaces having one point for each pixel). After some theoretical considerations, the paper proposes a divide-and-conquer algorithm to approximate the magnitude vector of an image which is significantly faster than the naive inversion of a $N\\times N$ matrix (where $N$ is the number of pixels in the image). Finally, the magnitude vector is shown to correlate with the result of traditional edge detectors (after hyper-parameter tuning) and an application to adversarial robustness is described.", "review_text": "This paper does not provide any theoretically-backed reason why magnitude vectors of images should be considered, and in fact the given definition of a vector space associated with an image appears quite dubious: the existence of the magnitude of an image depends on individual brighness values, which are irrelevant to the actual overall meaning of the image; in addition, the metric mixes the position of the pixels with their brightness, which looks very unnatural.\n\nThe algorithm to speed up the computation of magnitude vectors is also not theoretically backed in any way, and only supported by empirical experiments.\n\nFinally, the two given application of magnitude vectors to computer vision are very weak. In edge detection, it is shown that with appropriate hyperparameter tuning, magnitude vectors can recover a noisy version of the output of the Canny detector. The authors never observe that edge detection is a local problem, whereas the magnitude vector is a global quantity; this appears to me as an intrinsic theoretical limitation to the application of magnitude vectors to edge detection. As for adversarial robustness, it seems quite obvious that generated adversarial examples for a base LeNet are classified more correctly by a modified LeNet (regardless of how the network is actually modified).\n\nHere are some additional minor comments:\n- After Definition 1: a *finite* set of vectors\n- Lemma 6: this is obvious, there is no need to cite a 2009 preprint\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper applies the machinery of magnitude vectors of metric spaces to extract information from individual images (which can be themselves be seen as finite metric spaces having one point for each pixel). After some theoretical considerations, the paper proposes a divide-and-conquer algorithm to approximate the magnitude vector of an image which is significantly faster than the naive inversion of a $N\\times N$ matrix (where $N$ is the number of pixels in the image). Finally, the magnitude vector is shown to correlate with the result of traditional edge detectors (after hyper-parameter tuning) and an application to adversarial robustness is described.", "main_review": "This paper does not provide any theoretically-backed reason why magnitude vectors of images should be considered, and in fact the given definition of a vector space associated with an image appears quite dubious: the existence of the magnitude of an image depends on individual brighness values, which are irrelevant to the actual overall meaning of the image; in addition, the metric mixes the position of the pixels with their brightness, which looks very unnatural.\n\nThe algorithm to speed up the computation of magnitude vectors is also not theoretically backed in any way, and only supported by empirical experiments.\n\nFinally, the two given application of magnitude vectors to computer vision are very weak. In edge detection, it is shown that with appropriate hyperparameter tuning, magnitude vectors can recover a noisy version of the output of the Canny detector. The authors never observe that edge detection is a local problem, whereas the magnitude vector is a global quantity; this appears to me as an intrinsic theoretical limitation to the application of magnitude vectors to edge detection. As for adversarial robustness, it seems quite obvious that generated adversarial examples for a base LeNet are classified more correctly by a modified LeNet (regardless of how the network is actually modified).\n\nHere are some additional minor comments:\n- After Definition 1: a *finite* set of vectors\n- Lemma 6: this is obvious, there is no need to cite a 2009 preprint\n", "summary_of_the_review": "This paper presents an application of magnitude vectors to image, which is not convincing both from a theoretical perspective (why should the given definition be useful? why should the given speed-up algorithm work well?) and from an empirical perspective (the results on edge detection and adversarial robustness are weak).", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635183146779}], "openreview_url": "https://openreview.net/forum?id=-3Qj7Jl6UP5", "arxiv_id": "2110.15188", "paper_pdf": "papers/-3Qj7Jl6UP5.pdf", "paper_pdf_sha256": "9d22ebc02f37c7f8999316f15db468927be2c72583fa8bfde309563ab9a6d3bc", "paper_pdf_bytes": 1553416, "paper_pdf_source": "openreview", "code_url": "https://github.com/MikeAdamer/mag-metric", "code_repository": "MikeAdamer/mag-metric", "code_commit": "1f33528faf55d32d4cd8cfa983bbc95cf9b3dce0", "code_archive": "repos/-3Qj7Jl6UP5.zip", "code_archive_sha256": "ba7b0cc320fa111f048e036608739e858eb4e488a01d2fb2f04471c6e7e19ae1", "code_archive_bytes": 21211, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 16, "github_languages": {"Python": 58499}, "github_archived": false, "github_pushed_at": "2022-09-30T14:26:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-magnitude-vector-of-images-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJxQxeBYwH", "year": 2020, "status": "rejected", "title": "Are Powerful Graph Neural Nets Necessary? A Dissection on Graph Classification", "authors": ["Ting Chen", "Song Bian", "Yizhou Sun"], "authorids": ["iamtingchen@gmail.com", "biansonghz@gmail.com", "yzsun@cs.ucla.edu"], "authors_source": "OpenReview API", "abstract": "Graph Neural Nets (GNNs) have received increasing attentions, partially due to their superior performance in many node and graph classification tasks. However, there is a lack of understanding on what they are learning and how sophisticated the learned graph functions are.  In this work, we propose a dissection of GNNs on graph classification into two parts: 1) the graph filtering, where graph-based neighbor aggregations are performed, and 2) the set function, where a set of hidden node features are composed for prediction. To study the importance of both parts, we propose to linearize them separately. We first linearize the graph filtering function, resulting Graph Feature Network (GFN), which is a simple lightweight neural net defined on a \\textit{set} of graph augmented features. Further linearization of GFN's set function results in Graph Linear Network (GLN), which is a linear function. Empirically we perform evaluations on common graph classification benchmarks. To our surprise, we find that, despite the simplification, GFN could match or exceed the best accuracies produced by recently proposed GNNs (with a fraction of computation cost), while GLN underperforms significantly. Our results demonstrate the importance of non-linear set function, and suggest that linear graph filtering with non-linear set function is an efficient and powerful scheme for modeling existing graph classification benchmarks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hye94W7M5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2092/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper dissects the importance of two parts in GCN: 1) nonlinear neighborhood aggregation; 2) nonlinear set function by linearizing the two parts and resulting in Graph Feature Network (GFN) and Graph Linear Network (GLN). It shows empirically that GFN achieves almost the same performance while GLN is much worse, suggesting the nonlinear graph neighborhood aggregation step may be unnecessary. Extensive ablation studies are conducted to single out the effects of various factors.\n\nThe paper studies an interesting problem and sets out a good plan of experiments to verify the hypotheses. The results are interesting: merely constructing graph neighborhood features alone is enough to get comparable performance with GCN since the nonlinearity in the set function is strong enough. The experiments are designed nicely: 1) it compares with various baselines on a variety of popular benchmarks; 2) ablation studies single out the importance of different graph features, such as degree, and multi-hop averages; 3) verifying whether the good performance GFN comes from easier optimization.\n\nThe paper is also clearly written, with clean notations, and well-structured sections.\n\nI think the experiment can be improved by comparing on larger, more complex datasets. Figure 1 seems to suggest GCN is overfitting compared to GFN due to its extra capacity--significantly better training accuracy but slightly worse test accuracy. It is usually the case that larger and more complex datasets require more sophisticated models. But the paper makes a good case for GFN in these datasets for the graph classification task.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper dissects the importance of two parts in GCN: 1) nonlinear neighborhood aggregation; 2) nonlinear set function by linearizing the two parts and resulting in Graph Feature Network (GFN) and Graph Linear Network (GLN). It shows empirically that GFN achieves almost the same performance while GLN is much worse, suggesting the nonlinear graph neighborhood aggregation step may be unnecessary. Extensive ablation studies are conducted to single out the effects of various factors.\n\nThe paper studies an interesting problem and sets out a good plan of experiments to verify the hypotheses. The results are interesting: merely constructing graph neighborhood features alone is enough to get comparable performance with GCN since the nonlinearity in the set function is strong enough. The experiments are designed nicely: 1) it compares with various baselines on a variety of popular benchmarks; 2) ablation studies single out the importance of different graph features, such as degree, and multi-hop averages; 3) verifying whether the good performance GFN comes from easier optimization.\n\nThe paper is also clearly written, with clean notations, and well-structured sections.\n\nI think the experiment can be improved by comparing on larger, more complex datasets. Figure 1 seems to suggest GCN is overfitting compared to GFN due to its extra capacity--significantly better training accuracy but slightly worse test accuracy. It is usually the case that larger and more complex datasets require more sophisticated models. But the paper makes a good case for GFN in these datasets for the graph classification task."}, "tcdate": 1572118833883}, {"id": "HJgX4T6CtH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2092/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a dissection analysis of graph neural networks by decomposing GNNs into two parts: a graph filtering function and a set function. Although this decomposition may not be unique in general, as pointed out in the paper, these two parts can help analyze the impact of each part in the GNN model. Two simplified versions of GNN is then proposed by linearizing the graph filtering function and the set function, denoted as GFN and GLN, respectively. Experimental results on benchmarks datasets for graph classification show that GFN can achieve comparable or even better performance compared to recently proposed GNNs with higher computational efficiency. This demonstrates that the current GNN models may be unnecessarily complicated and overkill on graph classification. These empirical results are pretty interesting to the research community, and can encourage other researchers to reflect on existing fancy GNN models whether it's worth having more complex and more computationally expensive models to achieve similar or even inferior performance. Overall, this paper is well-written and the contribution is clear. I would like to recommend a weak accept for this paper. If the suggestions below can be addressed in author response, I would be willing to increase the score.\n\n\nSuggestions for improvement:\n\n1) Considering the experimental results in this paper, it is possible that the existing graph classification tasks are not that difficult so that the simplified GNN variant can also achieve comparable or even better performance (easier to learn). This can be conjectured from the consistently better training performance but comparable testing performance of original GNN. Another possibility is that even the original GNN has larger model capacity, it is not able to capture more useful information from the graph structure, even on tasks that are more challenging than graph classification. However, this paper lacks such in-depth discussions;\n\n2) Besides the graph classification task, it would be better to explore the performance of the simplified GNN on other graph learning tasks, such as node classification, and various downstream tasks using graph neural networks. This can help demystify the question raised in the previous point; 3) The matrix \\tilde{A} in Equation 5 is not well explained (described as \"similar to that in Kipf and Welling (2016)\"). It would be more clear to directly point out that it is the adjacency matrix, as described later in the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper presents a dissection analysis of graph neural networks by decomposing GNNs into two parts: a graph filtering function and a set function. Although this decomposition may not be unique in general, as pointed out in the paper, these two parts can help analyze the impact of each part in the GNN model. Two simplified versions of GNN is then proposed by linearizing the graph filtering function and the set function, denoted as GFN and GLN, respectively. Experimental results on benchmarks datasets for graph classification show that GFN can achieve comparable or even better performance compared to recently proposed GNNs with higher computational efficiency. This demonstrates that the current GNN models may be unnecessarily complicated and overkill on graph classification. These empirical results are pretty interesting to the research community, and can encourage other researchers to reflect on existing fancy GNN models whether it's worth having more complex and more computationally expensive models to achieve similar or even inferior performance. Overall, this paper is well-written and the contribution is clear. I would like to recommend a weak accept for this paper. If the suggestions below can be addressed in author response, I would be willing to increase the score.\n\n\nSuggestions for improvement:\n\n1) Considering the experimental results in this paper, it is possible that the existing graph classification tasks are not that difficult so that the simplified GNN variant can also achieve comparable or even better performance (easier to learn). This can be conjectured from the consistently better training performance but comparable testing performance of original GNN. Another possibility is that even the original GNN has larger model capacity, it is not able to capture more useful information from the graph structure, even on tasks that are more challenging than graph classification. However, this paper lacks such in-depth discussions;\n\n2) Besides the graph classification task, it would be better to explore the performance of the simplified GNN on other graph learning tasks, such as node classification, and various downstream tasks using graph neural networks. This can help demystify the question raised in the previous point; 3) The matrix \\tilde{A} in Equation 5 is not well explained (described as \"similar to that in Kipf and Welling (2016)\"). It would be more clear to directly point out that it is the adjacency matrix, as described later in the paper."}, "tcdate": 1571900714582}, {"id": "BJg-ri4CYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2092/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tries to study the importance of different components of GNNs. This paper studies two components 1) graph filtering: aggregation of neighboring features and 2) the aggregation function for the output.\n\nTo study this problem, this paper proposes two models, Graph Feature Network (GFN) and Graph Linear Network (GLN). GFN first uses the adjacency matrix to create several layers of features, then applies a multi-layer fully-connected neural network. GLN is a special case of GFN with the fully-connected neural network being linear.\n\nThis paper conducts experiments on graph classification task and finds GFN gives a reasonable performance, whereas GLN's performance is weaker.\n\n\n\nComments:\nThis paper studies an important problem in GNN, and the proposed method is interesting. However, I cannot accept the paper in the current form because of the following reasons.\n\n1. There is no theoretical analysis in the paper. For example, on some datasets, GFN, GLN, and GNN's performances are close while on other datasets, there are gaps. The current paper does not provide insight.\n\n2. GNN also contains non-linearity in the middle layers. However, the methodology in this paper cannot account for the importance of non-linearity in the middle layers.\n\n3. The experiment section ignores some recent results on graph classification tasks. See:\nhttps://arxiv.org/abs/1809.02670\nhttps://arxiv.org/abs/1905.13192", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper tries to study the importance of different components of GNNs. This paper studies two components 1) graph filtering: aggregation of neighboring features and 2) the aggregation function for the output.\n\nTo study this problem, this paper proposes two models, Graph Feature Network (GFN) and Graph Linear Network (GLN). GFN first uses the adjacency matrix to create several layers of features, then applies a multi-layer fully-connected neural network. GLN is a special case of GFN with the fully-connected neural network being linear.\n\nThis paper conducts experiments on graph classification task and finds GFN gives a reasonable performance, whereas GLN's performance is weaker.\n\n\n\nComments:\nThis paper studies an important problem in GNN, and the proposed method is interesting. However, I cannot accept the paper in the current form because of the following reasons.\n\n1. There is no theoretical analysis in the paper. For example, on some datasets, GFN, GLN, and GNN's performances are close while on other datasets, there are gaps. The current paper does not provide insight.\n\n2. GNN also contains non-linearity in the middle layers. However, the methodology in this paper cannot account for the importance of non-linearity in the middle layers.\n\n3. The experiment section ignores some recent results on graph classification tasks. See:\nhttps://arxiv.org/abs/1809.02670\nhttps://arxiv.org/abs/1905.13192"}, "tcdate": 1571863353302}], "openreview_url": "https://openreview.net/forum?id=BJxQxeBYwH", "arxiv_id": "1905.04579", "paper_pdf": "papers/BJxQxeBYwH.pdf", "paper_pdf_sha256": "4c0c8fba22739b685bbbbe1d0504875aa20b835d46e1ff80b4aaf3696f27feba", "paper_pdf_bytes": 1431009, "paper_pdf_source": "openreview", "code_url": "https://github.com/chentingpc/gfn", "code_repository": "chentingpc/gfn", "code_commit": "b598eb77f62680e62e2adb558b158505b2a3926a", "code_archive": "repos/BJxQxeBYwH.zip", "code_archive_sha256": "e84c5442f37fd839767ca4446d3ae7bada5b171133c4c64c54094fec228983d6", "code_archive_bytes": 16195, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 21, "github_languages": {"Python": 53680}, "github_archived": false, "github_pushed_at": "2020-05-05T01:49:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dissecting-graph-neural-networks-on-graph"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1giVsRcYm", "year": 2019, "status": "rejected", "title": "Count-Based Exploration with the Successor Representation", "authors": ["Marlos C. Machado", "Marc G. Bellemare", "Michael Bowling"], "authorids": ["machado@ualberta.ca", "bellemare@google.com", "mbowling@ualberta.ca"], "authors_source": "OpenReview API", "abstract": "The problem of exploration in reinforcement learning is well-understood in the tabular case and many sample-efficient algorithms are known. Nevertheless, it is often unclear how the algorithms in the tabular setting can be extended to tasks with large state-spaces where generalization is required. Recent promising developments generally depend on problem-specific density models or handcrafted features. In this paper we introduce a simple approach for exploration that allows us to develop theoretically justified algorithms in the tabular case but that also give us intuitions for new algorithms applicable to settings where function approximation is required. Our approach and its underlying theory is based on the substochastic successor representation, a concept we develop here. While the traditional successor representation is a representation that defines state generalization by the similarity of successor states, the substochastic successor representation is also able to implicitly count the number of times each state (or feature) has been observed. This extension connects two until now disjoint areas of research. We show in traditional tabular domains (RiverSwim and SixArms) that our algorithm empirically performs as well as other sample-efficient algorithms. We then describe a deep reinforcement learning algorithm inspired by these ideas and show that it matches the performance of recent pseudo-count-based methods in hard exploration Atari 2600 games.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "Hyl3ZDxqh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper29/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a new exploration strategy, based on the successor representation (SR), which can be used as a pseudo bonus in reinforcement learning. The authors also showed the connection between the state visit count and the SR, in the tabular case. Finally, the proposed algorithm had been tested on simulated examples, and several hard exploration Atari domains.\n\nIn general, there are some interesting ideas in this paper, while the empirical justification may not be strong enough. My pros and cons are summarized as follows. \nPros:\n- The idea of using SR for pseudo count in deep RL is novel.\n- Theorem 1 shows the interesting connection between state visit count and the proposed SR.\n- The experiments on Atari games show some promise for using SR (but not that much).\nCons:\n- There are a few inconsistencies regarding the use of SR. For example, the tabular case used the minus l1 norm as the reward bonus; however, the Atari case instead set the bonus to be the reciprocal of the l2 norm. \n- Other than the Montezuma's Revenge, it's difficult to draw the conclusion that using SR can generally lead to better exploration performance, based on the last two columns of Table 2.\n- The definition of loss L_{SR} is a bit unclear: Is there something similar to the Bellman equation you can say about SR? I also don't quite understand the motivation for the architecture between \\phi and \\psi in Figure 1.\n- A few small comments/questions are listed as follows.\n  1. When discussing the impact of the introduced auxiliary task, it would be more convincing to show the performance of games other than Montezuma's Revenge.\n  2. Why is it true that \"... because a reward of 1 is observed...\", in the second paragraph of Section 4?\n  3. What is the value of \\tau in the loss L_{TD} on Atari domains?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting work but with some flaws ", "review": "This paper proposed a new exploration strategy, based on the successor representation (SR), which can be used as a pseudo bonus in reinforcement learning. The authors also showed the connection between the state visit count and the SR, in the tabular case. Finally, the proposed algorithm had been tested on simulated examples, and several hard exploration Atari domains.\n\nIn general, there are some interesting ideas in this paper, while the empirical justification may not be strong enough. My pros and cons are summarized as follows. \nPros:\n- The idea of using SR for pseudo count in deep RL is novel.\n- Theorem 1 shows the interesting connection between state visit count and the proposed SR.\n- The experiments on Atari games show some promise for using SR (but not that much).\nCons:\n- There are a few inconsistencies regarding the use of SR. For example, the tabular case used the minus l1 norm as the reward bonus; however, the Atari case instead set the bonus to be the reciprocal of the l2 norm. \n- Other than the Montezuma's Revenge, it's difficult to draw the conclusion that using SR can generally lead to better exploration performance, based on the last two columns of Table 2.\n- The definition of loss L_{SR} is a bit unclear: Is there something similar to the Bellman equation you can say about SR? I also don't quite understand the motivation for the architecture between \\phi and \\psi in Figure 1.\n- A few small comments/questions are listed as follows.\n  1. When discussing the impact of the introduced auxiliary task, it would be more convincing to show the performance of games other than Montezuma's Revenge.\n  2. Why is it true that \"... because a reward of 1 is observed...\", in the second paragraph of Section 4?\n  3. What is the value of \\tau in the loss L_{TD} on Atari domains?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541175044266}, {"id": "HyxypQychm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper29/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Being familiar but not an expert in reinforcement learning, my review will focus on the overall soundness of the proposed method\n\nSummary:\n\nThe authors are interested in the problem of sample efficiency in reinforcement learning, i.e. how to learn a policy achieving good performance (discounted reward) in a RL setting using as little interaction with the environment as possible.\nTo do this the authors propose to learn a policy in a new environment where the reward has changed: an exploration bonus is added to the reward that should bias the agent towards the least frequently visited states.\n\nThe algorithms proposed throughout the manuscript are extensions of a two-part algorithm of the following flavour: 1) An estimate of visitation count is done in an online fashion using a modified version of the successor representation (SR). 2) This estimates parametrizes the exploration bonus of the environment . Both learning algorithms are optimized together.\n\nThis initial algorithm is fairly simple in its description and builds on well established ideas in RL. The authors then ‘evaluate the effectiveness of the proposed exploration bonus in a standard model-based algorithm’ against other baselines. They do explain how the model is learned, but not how the policy is optimized.\n\nThe remainder of the manuscript applies the same idea to different settings.\nFor large state spaces, The SR expected visits are learned using TD along with state action value functions. The counts of visitations are replaced by features that are also learned.\n\nOverall, the manuscript is rather confusing.\nThe SSR theorem is stated (with no real intuition and the actual bounds on n(s) left for the reader to derive). It is not well motivated. Why would we want expected counts and not the discounted version?\nThen the remainder of the paper actually makes no use of this theorem, but only use it as a distant inspiration. Tentative connections are made such as TD underestimating SR thus leading to a result more akin to SSR, which is highly speculative. It is also irrelevant since features are learned anyway.\n\nThe final proposed architecture has many additions to a simple DQN (the reconstruction + the exploration bonus + the MMC). This makes it difficult to understand what the contribution of the exploration bonus is.\nIt does not help that results are manually extracted from histograms found in  papers.\n\nOverall, although the intuition is interesting (though not so new).\nThe overall motivation and structure of this manuscript makes think it does not match the standards of ICLR for publication\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not very novel and rather confusing", "review": "Being familiar but not an expert in reinforcement learning, my review will focus on the overall soundness of the proposed method\n\nSummary:\n\nThe authors are interested in the problem of sample efficiency in reinforcement learning, i.e. how to learn a policy achieving good performance (discounted reward) in a RL setting using as little interaction with the environment as possible.\nTo do this the authors propose to learn a policy in a new environment where the reward has changed: an exploration bonus is added to the reward that should bias the agent towards the least frequently visited states.\n\nThe algorithms proposed throughout the manuscript are extensions of a two-part algorithm of the following flavour: 1) An estimate of visitation count is done in an online fashion using a modified version of the successor representation (SR). 2) This estimates parametrizes the exploration bonus of the environment . Both learning algorithms are optimized together.\n\nThis initial algorithm is fairly simple in its description and builds on well established ideas in RL. The authors then ‘evaluate the effectiveness of the proposed exploration bonus in a standard model-based algorithm’ against other baselines. They do explain how the model is learned, but not how the policy is optimized.\n\nThe remainder of the manuscript applies the same idea to different settings.\nFor large state spaces, The SR expected visits are learned using TD along with state action value functions. The counts of visitations are replaced by features that are also learned.\n\nOverall, the manuscript is rather confusing.\nThe SSR theorem is stated (with no real intuition and the actual bounds on n(s) left for the reader to derive). It is not well motivated. Why would we want expected counts and not the discounted version?\nThen the remainder of the paper actually makes no use of this theorem, but only use it as a distant inspiration. Tentative connections are made such as TD underestimating SR thus leading to a result more akin to SSR, which is highly speculative. It is also irrelevant since features are learned anyway.\n\nThe final proposed architecture has many additions to a simple DQN (the reconstruction + the exploration bonus + the MMC). This makes it difficult to understand what the contribution of the exploration bonus is.\nIt does not help that results are manually extracted from histograms found in  papers.\n\nOverall, although the intuition is interesting (though not so new).\nThe overall motivation and structure of this manuscript makes think it does not match the standards of ICLR for publication\n", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541170102952}, {"id": "Skx_9r3_2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper29/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors are tackling sample efficiency in the reinforcement learning setting by designing a reward function that encourages exploration. To achieve this they propose you use the successor function which basically counts how often a state has been visited. At first the show this for discrete settings and extend their approach to the continuous state spaces in the Atari 2600 environments. \n\nThe paper is well written and the motivation and methods are clear from the beginning. \n\nMy biggest concerning is regarding the experimental results of this work. In Table 1 the authors show the results for the tabular games River Swim and Six Arms and copare their approach which they dub ESSR to three methods (E3, R-MAX, MBIE). The numbers in the table indicate that their method ESSR is outperforming E3 and R-MAX on both environments but is itself outperformed by MBIE. The authors don't mention this at al in the respective paragraph nor do the provide a reason as to why this could be case. Also, they neither introduce any of these methods nor do the explain the meaning of the acronyms. Only in the section 6 (of 7) they talk about related works are R-MAX and E3 introduced briefly. But yet again, MBIE is not mentioned. \n\nI have similar concerns about the results presented for the Atari benchmarks. In table 2 the authors compare their method to the classic DQN approach and two more approaches. While their approach outperforms DQN in almost all tasks, this does not hold for the remaining algorithms. Their method is being outperformed in all but one (Venture) task, where they report a higher variance and a small performance boost compared to DQN_e^MMC. Also it is not clear to me where the numbers for the DNQ_e^MMC come from. The authors just say \"[...] denotes another baseline used in the comparison\". Is this the proposed method of this work but without the successor representation?\n\nIn my opinion this work is lacking some clear and convincing results.  Is the main benefit of this method that it does not rely on domain-specific knowledge? If so, then it is not communicated clearly. The authors mention this briefly in the conclusion but provide no further analysis", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Theoretically grounded work but lacking convincing results", "review": "The authors are tackling sample efficiency in the reinforcement learning setting by designing a reward function that encourages exploration. To achieve this they propose you use the successor function which basically counts how often a state has been visited. At first the show this for discrete settings and extend their approach to the continuous state spaces in the Atari 2600 environments. \n\nThe paper is well written and the motivation and methods are clear from the beginning. \n\nMy biggest concerning is regarding the experimental results of this work. In Table 1 the authors show the results for the tabular games River Swim and Six Arms and copare their approach which they dub ESSR to three methods (E3, R-MAX, MBIE). The numbers in the table indicate that their method ESSR is outperforming E3 and R-MAX on both environments but is itself outperformed by MBIE. The authors don't mention this at al in the respective paragraph nor do the provide a reason as to why this could be case. Also, they neither introduce any of these methods nor do the explain the meaning of the acronyms. Only in the section 6 (of 7) they talk about related works are R-MAX and E3 introduced briefly. But yet again, MBIE is not mentioned. \n\nI have similar concerns about the results presented for the Atari benchmarks. In table 2 the authors compare their method to the classic DQN approach and two more approaches. While their approach outperforms DQN in almost all tasks, this does not hold for the remaining algorithms. Their method is being outperformed in all but one (Venture) task, where they report a higher variance and a small performance boost compared to DQN_e^MMC. Also it is not clear to me where the numbers for the DNQ_e^MMC come from. The authors just say \"[...] denotes another baseline used in the comparison\". Is this the proposed method of this work but without the successor representation?\n\nIn my opinion this work is lacking some clear and convincing results.  Is the main benefit of this method that it does not rely on domain-specific knowledge? If so, then it is not communicated clearly. The authors mention this briefly in the conclusion but provide no further analysis", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541092752329}], "openreview_url": "https://openreview.net/forum?id=S1giVsRcYm", "arxiv_id": "1807.11622", "paper_pdf": "papers/S1giVsRcYm.pdf", "paper_pdf_sha256": "3eaf0a7a954c68f3833685770709e661b8d0aa8da03f9151a4f72dd77edb4cd0", "paper_pdf_bytes": 3941702, "paper_pdf_source": "openreview", "code_url": "https://github.com/mcmachado/count_based_exploration_sr", "code_repository": "mcmachado/count_based_exploration_sr", "code_commit": "e4d657eb498ca84a703ddd7ec426d908adc67a72", "code_archive": "repos/S1giVsRcYm.zip", "code_archive_sha256": "14563765715ec9109131eda98b3fa3c9537c66cf80d885bfb3f7e553abe5d450", "code_archive_bytes": 38266, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 31, "github_languages": {"Python": 93470}, "github_archived": false, "github_pushed_at": "2019-07-01T20:22:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/count-based-exploration-with-the-successor"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1tExikAW", "year": 2018, "status": "rejected", "title": "LatentPoison -- Adversarial Attacks On The Latent Space", "authors": ["Antonia Creswell", "Biswa Sengupta", "Anil A. Bharath"], "authorids": ["ac2211@ic.ac.uk", "b.sengupta@imperial.ac.uk", "a.bharath@imperial.ac.uk"], "authors_source": "OpenReview API", "abstract": "Robustness and security of machine learning (ML) systems are intertwined, wherein a non-robust ML system (classifiers, regressors, etc.) can be subject to attacks using a wide variety of exploits. With the advent of scalable deep learning methodologies, a lot of emphasis has been put on the robustness of supervised, unsupervised and reinforcement learning algorithms. Here, we study the robustness of the latent space of a deep variational autoencoder (dVAE), an unsupervised generative framework, to show that it is indeed possible to perturb the latent space, flip the class predictions and keep the classification probability approximately equal before and after an attack. This means that an agent that looks at the outputs of a decoder would remain oblivious to an attack.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1xzWeqgG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper153/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is concerned with both security and machine learning. \nAssuming that data is encoded, transmited, and decoded using a VAE,\nthe paper proposes a man-in-middle attack that alters the VAE encoding of the input data so that the decoded output will be misclassified.\nThe objectives are to: 1) fool the autoencoder; the classification output of the autoencoder is different from the actual class of the input; 2) make minimal change in the middle so that the attack is not detectable. \n\nThis paper is concerned with both security and machine learning, but there is no clear contributions to either field. From the machine learning perspective, the proposed \"attacking\" method is standard without any technical novelty. From the security perspective, the scenarios are too simplistic. The encoding-decoding mechanism being attacked is too simple without any security enhancement. This is an unrealistic scenario. For applications with security concerns, there should have been methods to guard against man-in-the-middle attack, and the paper should have at least considered some of them. Without considering the state-of-the-art security defending mechanism, it is difficult to judge the contribution of the paper to the security community. \n\nI am not a security expert, but I doubt that the proposed method are formulated based on well founded security concepts and ideas. For example, what are the necessary and sufficient conditions for an attacking method to be undetectable? Are the criteria about the magnitude of epsilon given on Section 3.3. necessary and sufficient? Is there any reference for them? Why do we require the correspondence between the classification confidence of tranformed and original data? Would it be enough to match the DISTRIBUTION of the confidence? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Too simplistic scenario", "rating": "4: Ok but not good enough - rejection", "review": "This paper is concerned with both security and machine learning. \nAssuming that data is encoded, transmited, and decoded using a VAE,\nthe paper proposes a man-in-middle attack that alters the VAE encoding of the input data so that the decoded output will be misclassified.\nThe objectives are to: 1) fool the autoencoder; the classification output of the autoencoder is different from the actual class of the input; 2) make minimal change in the middle so that the attack is not detectable. \n\nThis paper is concerned with both security and machine learning, but there is no clear contributions to either field. From the machine learning perspective, the proposed \"attacking\" method is standard without any technical novelty. From the security perspective, the scenarios are too simplistic. The encoding-decoding mechanism being attacked is too simple without any security enhancement. This is an unrealistic scenario. For applications with security concerns, there should have been methods to guard against man-in-the-middle attack, and the paper should have at least considered some of them. Without considering the state-of-the-art security defending mechanism, it is difficult to judge the contribution of the paper to the security community. \n\nI am not a security expert, but I doubt that the proposed method are formulated based on well founded security concepts and ideas. For example, what are the necessary and sufficient conditions for an attacking method to be undetectable? Are the criteria about the magnitude of epsilon given on Section 3.3. necessary and sufficient? Is there any reference for them? Why do we require the correspondence between the classification confidence of tranformed and original data? Would it be enough to match the DISTRIBUTION of the confidence? ", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511813384183}, {"id": "H1I-1LYxf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper153/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The idea is clearly stated (but lacks some details) and I enjoyed reading the paper. \n\nI understand the difference between [Kos+17] and the proposed scheme but I could not understand in which situation the proposed scheme works better. From the adversary's standpoint, it would be easier to manipulate inputs than latent variables. On the other hand, I agree that sample-independent perturbation is much more practical than sample-dependent perturbation.\n\nIn Section 3.1, the attack methods #2 and #3 should be detailed more. I could not imagine how VAE and T are trained simultaneously.\n\nIn Section 3.2, the authors listed a couple of loss functions. How were these loss functions are combined? The final optimization problem that is used for training of the propose VAE should be formally defined. Also, the detailed specification of the VAE should be detailed.\n\nFrom figures in Figure 4 and Figure 5, I could see that the proposed scheme performs successfully in a qualitative manner, however, it is difficult to evaluate the proposed scheme qualitatively without comparisons with baselines. For example, can the proposed scheme can be compared with [Kos+17] or some other sample-dependent attacks? Also, can you experimentally show that attacks on latent variables are more powerful than attacks on inputs?\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "LatentPoison -- Adversarial Attacks On The Latent Space", "rating": "5: Marginally below acceptance threshold", "review": "The idea is clearly stated (but lacks some details) and I enjoyed reading the paper. \n\nI understand the difference between [Kos+17] and the proposed scheme but I could not understand in which situation the proposed scheme works better. From the adversary's standpoint, it would be easier to manipulate inputs than latent variables. On the other hand, I agree that sample-independent perturbation is much more practical than sample-dependent perturbation.\n\nIn Section 3.1, the attack methods #2 and #3 should be detailed more. I could not imagine how VAE and T are trained simultaneously.\n\nIn Section 3.2, the authors listed a couple of loss functions. How were these loss functions are combined? The final optimization problem that is used for training of the propose VAE should be formally defined. Also, the detailed specification of the VAE should be detailed.\n\nFrom figures in Figure 4 and Figure 5, I could see that the proposed scheme performs successfully in a qualitative manner, however, it is difficult to evaluate the proposed scheme qualitatively without comparisons with baselines. For example, can the proposed scheme can be compared with [Kos+17] or some other sample-dependent attacks? Also, can you experimentally show that attacks on latent variables are more powerful than attacks on inputs?\n\n\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511771902361}, {"id": "HkCuu2YxG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper153/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper misses the point of what VAEs (or GANs, in general) are used for. The idea of using VAEs is not to encode and decode images (or in general any input), but to recover the generating process that created those images so we have an unlimited source of samples. The use of these techniques for compressing is still unclear and their quality today is too low. So the attack that the authors are proposing does not make sense and my take is that we should see significant changes before they can make sense. \n\nBut let’s assume that at some point they can be used as the authors propose. In which one person encodes an image, send the latent variable to a friend, but a foe intercepts it on the way and tampers with it so the receiver recovers the wrong image without knowing. Now if the sender believes the sample can be tampered with, if the sender codes z with his private key would not make the attack useless? I think this will make the first attack useless. \n\nThe other two attacks require that the foe is inserted in the middle of the training of the VAE. This is even less doable, because the encoder and decoder are not train remotely. They are train of the same machine or cluster in a controlled manner by the person that would use the system. Once it is train it will give away the decoder and keep the encoder for sending information.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "VAE are not a compression scheme", "rating": "3: Clear rejection", "review": "This paper misses the point of what VAEs (or GANs, in general) are used for. The idea of using VAEs is not to encode and decode images (or in general any input), but to recover the generating process that created those images so we have an unlimited source of samples. The use of these techniques for compressing is still unclear and their quality today is too low. So the attack that the authors are proposing does not make sense and my take is that we should see significant changes before they can make sense. \n\nBut let’s assume that at some point they can be used as the authors propose. In which one person encodes an image, send the latent variable to a friend, but a foe intercepts it on the way and tampers with it so the receiver recovers the wrong image without knowing. Now if the sender believes the sample can be tampered with, if the sender codes z with his private key would not make the attack useless? I think this will make the first attack useless. \n\nThe other two attacks require that the foe is inserted in the middle of the training of the VAE. This is even less doable, because the encoder and decoder are not train remotely. They are train of the same machine or cluster in a controlled manner by the person that would use the system. Once it is train it will give away the decoder and keep the encoder for sending information.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511798902129}], "openreview_url": "https://openreview.net/forum?id=B1tExikAW", "arxiv_id": "1711.02879", "paper_pdf": "papers/B1tExikAW.pdf", "paper_pdf_sha256": "9eb7ea64274e6f133f2e993a4bf7b7e6dc0c50356c73ecff60f9756082ad202a", "paper_pdf_bytes": 8626408, "paper_pdf_source": "openreview", "code_url": "https://github.com/ToniCreswell/Adversarial-Attack-On-Latent-Space", "code_repository": "ToniCreswell/Adversarial-Attack-On-Latent-Space", "code_commit": "812de9a3c6a6be8ad6e2923625d09b900fe07a00", "code_archive": "repos/B1tExikAW.zip", "code_archive_sha256": "49ed91c4148515ca56b70e9c8b8ef4f686c21bcff7f6f9243919e2b4f61b9c55", "code_archive_bytes": 18206, "code_file_count": 6, "code_extensions": {".py": 5, ".ipynb": 1}, "github_disk_usage_kb": 22, "github_languages": {"Python": 37809, "Jupyter Notebook": 2909}, "github_archived": false, "github_pushed_at": "2022-12-07T23:58:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/latentpoison-adversarial-attacks-on-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bibt0JTvpx", "year": 2026, "status": "rejected", "title": "EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation", "authors": ["Jinghan Jia", "Hadi Reisizadeh", "Chongyu Fan", "Nathalie Baracaldo", "Mingyi Hong", "Sijia Liu"], "authorids": ["~Jinghan_Jia1", "~Hadi_Reisizadeh1", "~Chongyu_Fan1", "~Nathalie_Baracaldo1", "~Mingyi_Hong1", "~Sijia_Liu1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have shown remarkable reasoning capabilities when trained with chain-of-thought (CoT) supervision. However, the long and verbose CoT traces, especially those distilled from large reasoning models (LRMs) such as DeepSeek-R1, significantly increase training costs during the distillation process, where a non-reasoning base model is taught to replicate the reasoning behavior of an LRM. In this work, we study the problem of CoT condensation for resource-efficient reasoning training, aimed at pruning intermediate reasoning steps (i.e., thoughts) in CoT traces, enabling supervised model training on length-reduced CoT data while preserving both answer accuracy and the model’s ability to generate coherent reasoning. Our rationale is that CoT traces typically follow a three-stage structure: problem understanding, exploration, and solution convergence. Through empirical analysis, we find that retaining the structure of the reasoning trace, especially the early stage of problem understanding (rich in reflective cues) and the final stage of solution convergence (which closely relates to the final answer), is sufficient to achieve lossless reasoning supervision. To this end, we propose an Edge-Preserving Condensation method, EPiC, which selectively retains only the initial and final segments of each CoT trace while discarding the middle portion. This design draws an analogy to preserving the “edge” of a reasoning trajectory, capturing both the initial problem framing and the final answer synthesis, to maintain logical continuity. Our analyses leveraging the CoT landscape and measuring the mutual information between CoT steps provide further validation for this design. Experiments across multiple model families (Qwen and LLaMA) and benchmarks show that EPiC reduces training time by over 34% while achieving lossless reasoning accuracy (e.g., on MATH500), comparable to full CoT supervision. Additionally, we show that EPiC outperforms other condensation methods, including teacher-guided regeneration of condensed CoTs.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "khDruxre6j", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3173/Reviewer_12Q4"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes EPiC, an edge-preserving condensation method that prunes the middle portion of chain-of-thought (CoT) traces while retaining the initial and final reasoning segments to reduce training cost.
Based on empirical analyses (e.g., mutual-information metrics), the authors show that these “edge” segments preserve logical coherence and supervision quality.
Experiments across multiple model families demonstrate notable training-time reduction with only mild loss in reasoning accuracy.", "review_text": "The paper proposes EPiC, an edge-preserving condensation method that prunes the middle portion of chain-of-thought (CoT) traces while retaining the initial and final reasoning segments to reduce training cost.
Based on empirical analyses (e.g., mutual-information metrics), the authors show that these “edge” segments preserve logical coherence and supervision quality.
Experiments across multiple model families demonstrate notable training-time reduction with only mild loss in reasoning accuracy.", "strengths": "- The topic and motivation, resource-efficient reasoning training are important and timely.\n\n- The paper is well-written and easy to follow, with a clear presentation of the problem and method.\n\n- The authors conduct comprehensive empirical analyses, offering useful insights that intermediate reasoning steps may be less important than early or final stages.", "weaknesses": "**[W1] Inconsistent and limited performance improvements.** In Table 2, the proposed method outperforms the baseline on MATH, AIME-25, and GPQA, but shows a significant drop on AIME-24. Moreover, efficiency metrics such as the number of tokens do not show clear advantages, following trends similar to competing methods.\n\n**[W2] Insufficient baselines and related work.** Recent studies have analyzed the importance of reasoning steps or proposed strategies (e.g., [1, 2]) that appear directly applicable to this setup. Including these as baselines and discussing them in the related-work section would provide a more holistic comparison.\n\n[1] Choi et al., Think Clearly: Improving Reasoning via Redundant Token Pruning, EMNLP 2025 (Findings)\n\n[2] Cui et al., Stepwise Perplexity-Guided Refinement for Efficient Chain-of-Thought Reasoning in Large Language Models, ACL 2025 (Findings)", "questions": "**[Q1]** Why is the maximum sequence length in GPQA experiments restricted to 4k tokens? Reasoning performance is often sensitive to the length limit; reporting results with longer contexts would strengthen the claim.\n\n**[Q2]** As I understand, EPiC preserves reasoning properties such as reflection markers (e.g., “Wait”). Is there a specific reason for this? Intuitively, intermediate reasoning steps might contain more such reflective cues.\n\n**[Q3]** Intermediate steps can affect test-time scaling, as they enhance reasoning diversity. Can EPiC maintain or recover this advantage under test-time scaling scenarios?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes EPiC, an edge-preserving condensation method that prunes the middle portion of chain-of-thought (CoT) traces while retaining the initial and final reasoning segments to reduce training cost.
Based on empirical analyses (e.g., mutual-information metrics), the authors show that these “edge” segments preserve logical coherence and supervision quality.
Experiments across multiple model families demonstrate notable training-time reduction with only mild loss in reasoning accuracy.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The topic and motivation, resource-efficient reasoning training are important and timely.\n\n- The paper is well-written and easy to follow, with a clear presentation of the problem and method.\n\n- The authors conduct comprehensive empirical analyses, offering useful insights that intermediate reasoning steps may be less important than early or final stages.", "weaknesses": "**[W1] Inconsistent and limited performance improvements.** In Table 2, the proposed method outperforms the baseline on MATH, AIME-25, and GPQA, but shows a significant drop on AIME-24. Moreover, efficiency metrics such as the number of tokens do not show clear advantages, following trends similar to competing methods.\n\n**[W2] Insufficient baselines and related work.** Recent studies have analyzed the importance of reasoning steps or proposed strategies (e.g., [1, 2]) that appear directly applicable to this setup. Including these as baselines and discussing them in the related-work section would provide a more holistic comparison.\n\n[1] Choi et al., Think Clearly: Improving Reasoning via Redundant Token Pruning, EMNLP 2025 (Findings)\n\n[2] Cui et al., Stepwise Perplexity-Guided Refinement for Efficient Chain-of-Thought Reasoning in Large Language Models, ACL 2025 (Findings)", "questions": "**[Q1]** Why is the maximum sequence length in GPQA experiments restricted to 4k tokens? Reasoning performance is often sensitive to the length limit; reporting results with longer contexts would strengthen the claim.\n\n**[Q2]** As I understand, EPiC preserves reasoning properties such as reflection markers (e.g., “Wait”). Is there a specific reason for this? Intuitively, intermediate reasoning steps might contain more such reflective cues.\n\n**[Q3]** Intermediate steps can affect test-time scaling, as they enhance reasoning diversity. Can EPiC maintain or recover this advantage under test-time scaling scenarios?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761642833105}, {"id": "QaQjwfRvmS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3173/Reviewer_BQsv"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper studies the problem of Chain-of-Thought (CoT) compression in reasoning training, which enhances the reasoning capability of smaller models by training them on distilled CoT data. The paper proposes an Edge-Preserving Condensation method called EPiC, which involves pruning the intermediate steps of a reasoning trajectory. The authors further analyze the rationale of EPiC through a comparison of mutual information. Finally, the authors conduct experiments across different benchmarks and models to verify that EPiC is utility-preserving and highly efficient.", "review_text": "The paper studies the problem of Chain-of-Thought (CoT) compression in reasoning training, which enhances the reasoning capability of smaller models by training them on distilled CoT data. The paper proposes an Edge-Preserving Condensation method called EPiC, which involves pruning the intermediate steps of a reasoning trajectory. The authors further analyze the rationale of EPiC through a comparison of mutual information. Finally, the authors conduct experiments across different benchmarks and models to verify that EPiC is utility-preserving and highly efficient.", "strengths": "S1: The research problem of this paper, the efficient training of reasoning models, is both timely and highly relevant.\n\nS2: The paper is well-written and easy to follow. Additionally, the proposed method is simple yet effective.\n\nS3: The authors conduct extensive experiments to verify the effectiveness of EPiC, and the reported results show that it achieves a strong utility-efficiency trade-off.", "weaknesses": "W1: The experimental setup presented in Figure 2 seems questionable, which weakens the paper's motivation. The S1 and LIMO datasets are significantly smaller in the number of examples than OpenR1Math. Therefore, it is expected that models trained on S1 and LIMO would underperform those trained on OpenR1Math. A fairer comparison would involve ensuring that the total number of tokens in the condensed dataset (e.g., after 50% randomized thought-level condensation) is comparable to the token counts of the S1 and LIMO datasets.\n\nW2: The proposed method is presented as a heuristic and appears to be motivated by a single example visualization in Figure 4. The argument would be more persuasive if it were supported by a statistical analysis of the spatial distances between each step thought and the correct answer across an entire dataset. While the visualization in Figure 4 is well-presented, the example shown might be a corner case or a specific, non-representative instance, which reduces the method's persuasiveness.\n\nW3: There appears to be a contradiction regarding whether the middle portion of a CoT trace is less informative when comparing Figure 4 and Table 5. Table 5 reports a higher MI for the MoC compared to the ToC, which suggests that the middle portion is actually more informative than the tail.\n\nW4: While the paper includes several experiments, the section could be strengthened by including more baselines and more in-depth analysis.\n\n-\tIn Table 2, adding MoC as a baseline would help to directly test the hypothesis that the middle portion of a CoT trace is less informative.\n\n-\tIn Table 2, EPiC significantly underperforms training on the full dataset. The paper lacks an in-depth analysis of this performance gap, which could help in understanding the failure cases of EPiC.\n\nW5: Line 279 seems to assume that the length of the CoT is the same for the “understanding and convergence” phases. This assumption seems unreasonable, as one would expect this length to vary across different samples and datasets.\n\nMinor points:\n\n-\tIn line 325, “xWe” -> “We”", "questions": "Q1: What method is used to partition CoT into three distinct sections?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of Chain-of-Thought (CoT) compression in reasoning training, which enhances the reasoning capability of smaller models by training them on distilled CoT data. The paper proposes an Edge-Preserving Condensation method called EPiC, which involves pruning the intermediate steps of a reasoning trajectory. The authors further analyze the rationale of EPiC through a comparison of mutual information. Finally, the authors conduct experiments across different benchmarks and models to verify that EPiC is utility-preserving and highly efficient.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "S1: The research problem of this paper, the efficient training of reasoning models, is both timely and highly relevant.\n\nS2: The paper is well-written and easy to follow. Additionally, the proposed method is simple yet effective.\n\nS3: The authors conduct extensive experiments to verify the effectiveness of EPiC, and the reported results show that it achieves a strong utility-efficiency trade-off.", "weaknesses": "W1: The experimental setup presented in Figure 2 seems questionable, which weakens the paper's motivation. The S1 and LIMO datasets are significantly smaller in the number of examples than OpenR1Math. Therefore, it is expected that models trained on S1 and LIMO would underperform those trained on OpenR1Math. A fairer comparison would involve ensuring that the total number of tokens in the condensed dataset (e.g., after 50% randomized thought-level condensation) is comparable to the token counts of the S1 and LIMO datasets.\n\nW2: The proposed method is presented as a heuristic and appears to be motivated by a single example visualization in Figure 4. The argument would be more persuasive if it were supported by a statistical analysis of the spatial distances between each step thought and the correct answer across an entire dataset. While the visualization in Figure 4 is well-presented, the example shown might be a corner case or a specific, non-representative instance, which reduces the method's persuasiveness.\n\nW3: There appears to be a contradiction regarding whether the middle portion of a CoT trace is less informative when comparing Figure 4 and Table 5. Table 5 reports a higher MI for the MoC compared to the ToC, which suggests that the middle portion is actually more informative than the tail.\n\nW4: While the paper includes several experiments, the section could be strengthened by including more baselines and more in-depth analysis.\n\n-\tIn Table 2, adding MoC as a baseline would help to directly test the hypothesis that the middle portion of a CoT trace is less informative.\n\n-\tIn Table 2, EPiC significantly underperforms training on the full dataset. The paper lacks an in-depth analysis of this performance gap, which could help in understanding the failure cases of EPiC.\n\nW5: Line 279 seems to assume that the length of the CoT is the same for the “understanding and convergence” phases. This assumption seems unreasonable, as one would expect this length to vary across different samples and datasets.\n\nMinor points:\n\n-\tIn line 325, “xWe” -> “We”", "questions": "Q1: What method is used to partition CoT into three distinct sections?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761616888925}, {"id": "7zT8ZTjEmq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3173/Reviewer_7LEr"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "A framework for thought-level condensation is introduced, enabling efficient knowledge distillation by capturing reasoning thoughts with a smaller model. While the effectiveness is demonstrated, uncertainties remain regarding the proposed techniques.", "review_text": "A framework for thought-level condensation is introduced, enabling efficient knowledge distillation by capturing reasoning thoughts with a smaller model. While the effectiveness is demonstrated, uncertainties remain regarding the proposed techniques.", "strengths": "* This paper introduced a framework for thought-level condensation that enables efficient knowledge distillation. It effectively captures reasoning thoughts using a smaller non-reasoning model.\n\n* The authors conducted analysis through visual illustrations and quantitative analysis to demonstrate that reasoning information is preserved.\n\n* Experiments were conducted to demonstrate the effectiveness of the proposed framework.", "weaknesses": "* Figure 1 uses bar charts for both performance and efficiency, which is not appropriate since they belong to different categories and should be represented on separate axes.\n\n* Section 4 is crucial for the proposed method, but not all details are clearly presented. While all the key points are logical, they need to be clearly articulated and analyzed to make the techniques convincing. \n\n* In Figure 3, the example of CoT trace is interesting, where the authors pointed out the concerns of token based CoT condensation that drops self-reflective words and transition words, resulting in fragmented input. Is it the intrinsic issue of the token based CoT condensation, or just TokenShip may use an imperfect scoring mechanism for tokens? Anyway, there is no result if thought-level pruning would be possibly mapped back to the token level to make the comparison tangible. \n\n* While the visualization in Figure 4 can be useful for delivering the message that dropping the middle thought is reasonable, it also trigger a concern if the visualization is really based on real trajectory of reasoning, or just a make-up illustration. It is a bit questionable why the tail though can jump to the correct answer region significantly from the yellow dot that has been far away from even comparing to the start position. If this observation is true, it seems the author can even skip the head of thoughts. \n\n* In Table 1, it seems the proposed EPiC performs relatively even more stronger when the condensation ratio \\tau is low. The authors should provide some insight. Besides, the gap seems a bit too huge. Is the baseline method used here strong enough?", "questions": "Please check the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "A framework for thought-level condensation is introduced, enabling efficient knowledge distillation by capturing reasoning thoughts with a smaller model. While the effectiveness is demonstrated, uncertainties remain regarding the proposed techniques.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* This paper introduced a framework for thought-level condensation that enables efficient knowledge distillation. It effectively captures reasoning thoughts using a smaller non-reasoning model.\n\n* The authors conducted analysis through visual illustrations and quantitative analysis to demonstrate that reasoning information is preserved.\n\n* Experiments were conducted to demonstrate the effectiveness of the proposed framework.", "weaknesses": "* Figure 1 uses bar charts for both performance and efficiency, which is not appropriate since they belong to different categories and should be represented on separate axes.\n\n* Section 4 is crucial for the proposed method, but not all details are clearly presented. While all the key points are logical, they need to be clearly articulated and analyzed to make the techniques convincing. \n\n* In Figure 3, the example of CoT trace is interesting, where the authors pointed out the concerns of token based CoT condensation that drops self-reflective words and transition words, resulting in fragmented input. Is it the intrinsic issue of the token based CoT condensation, or just TokenShip may use an imperfect scoring mechanism for tokens? Anyway, there is no result if thought-level pruning would be possibly mapped back to the token level to make the comparison tangible. \n\n* While the visualization in Figure 4 can be useful for delivering the message that dropping the middle thought is reasonable, it also trigger a concern if the visualization is really based on real trajectory of reasoning, or just a make-up illustration. It is a bit questionable why the tail though can jump to the correct answer region significantly from the yellow dot that has been far away from even comparing to the start position. If this observation is true, it seems the author can even skip the head of thoughts. \n\n* In Table 1, it seems the proposed EPiC performs relatively even more stronger when the condensation ratio \\tau is low. The authors should provide some insight. Besides, the gap seems a bit too huge. Is the baseline method used here strong enough?", "questions": "Please check the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761532152741}], "openreview_url": "https://openreview.net/forum?id=Bibt0JTvpx", "arxiv_id": "2506.04205", "paper_pdf": "papers/Bibt0JTvpx.pdf", "paper_pdf_sha256": "65d0f0acf1f8948c088e9aca92e717647e1cf86d6478700e52562d568a1aa382", "paper_pdf_bytes": 2033741, "paper_pdf_source": "openreview", "code_url": "https://github.com/OPTML-Group/EPiC", "code_repository": "OPTML-Group/EPiC", "code_commit": "54610896cb4659a7a05920cf8437ea070cf84d2f", "code_archive": "repos/Bibt0JTvpx.zip", "code_archive_sha256": "05d59008616105ced26ad93f3c2474f2004d00fa12c526e77dceb26cbb1b3c0a", "code_archive_bytes": 50941, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 43, "github_languages": {"Python": 96709}, "github_archived": false, "github_pushed_at": "2025-06-11T18:08:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/epic-towards-lossless-speedup-for-reasoning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yLmcYLP3Yd", "year": 2025, "status": "rejected", "title": "Discrete Neural Algorithmic Reasoning", "authors": ["Gleb Rodionov", "Liudmila Prokhorenkova"], "authorids": ["~Gleb_Rodionov1", "~Liudmila_Prokhorenkova1"], "authors_source": "OpenReview API", "abstract": "Neural algorithmic reasoning aims to capture computations with neural networks via learning the models to imitate the execution of classic algorithms. While common architectures are expressive enough to contain the correct model in the weights space, current neural reasoners are struggling to generalize well on out-of-distribution data. On the other hand, classic computations are not affected by distributional shifts as they can be described as transitions between discrete computational states. In this work, we propose to force neural reasoners to maintain the execution trajectory as a combination of finite predefined states. To achieve that, we separate discrete and continuous data flows and describe the interaction between them. Trained with supervision on the algorithm's state transitions, such models are able to perfectly align with the original algorithm. To show this, we evaluate our approach on multiple algorithmic problems and get perfect test scores both in single-task and multitask setups. Moreover, the proposed architectural choice allows us to prove the correctness of the learned algorithms for any test data.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "Z9AJMN06DZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6535/Reviewer_P5ow"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper introduces a novel approach to neural algorithmic reasoning by enforcing neural networks to operate with discrete states and separating discrete and continuous data flows. The authors propose a model that integrates hard attention mechanisms, feature discretization, and a separation between discrete computations and continuous inputs (scalars). This design aims to align neural network computations closely with classical algorithms and thus improves out-of-distribution generalization and interpretability. The method is evaluated on several algorithmic tasks from the SALSA-CLRS benchmark, including BFS, DFS, Prim's algorithm, Dijkstra's algorithm, Maximum Independent Set (MIS), and Eccentricity calculations. The proposed Discrete Neural Algorithmic Reasoner (DNAR) achieves perfect test scores on these tasks, even on graphs significantly larger than those seen during training. The authors also discuss the limitations of their approach and potential directions for future work.", "review_text": "This paper introduces a novel approach to neural algorithmic reasoning by enforcing neural networks to operate with discrete states and separating discrete and continuous data flows. The authors propose a model that integrates hard attention mechanisms, feature discretization, and a separation between discrete computations and continuous inputs (scalars). This design aims to align neural network computations closely with classical algorithms and thus improves out-of-distribution generalization and interpretability. The method is evaluated on several algorithmic tasks from the SALSA-CLRS benchmark, including BFS, DFS, Prim's algorithm, Dijkstra's algorithm, Maximum Independent Set (MIS), and Eccentricity calculations. The proposed Discrete Neural Algorithmic Reasoner (DNAR) achieves perfect test scores on these tasks, even on graphs significantly larger than those seen during training. The authors also discuss the limitations of their approach and potential directions for future work.", "strengths": "- Originality: The paper presents a novel method that enforces discrete state transitions in neural networks, which is a significant departure from traditional continuous representations. By integrating hard attention and separating discrete and continuous data flows, the authors address key challenges in neural algorithmic reasoning, particularly out-of-distribution generalization, and interpretability.\n- Quality: The experimental results are strong, with DNAR achieving perfect test scores across multiple algorithmic tasks and graph sizes. The comparison with baseline models and state-of-the-art methods demonstrates the effectiveness of the proposed approach.\n- Clarity: The paper is generally well-written and structured. The authors explain their methodology, including architectural choices and training procedures. The inclusion of diagrams and tables aids in understanding the proposed model and its performance.\n- Significance: The work contributes to the field by showing that neural networks can be designed to mimic classical algorithms with perfect generalization and interpretability. This has implications for developing reliable and trustworthy AI systems that can be formally verified.", "weaknesses": "- Expressiveness Limitations: The enforced constraints, such as hard attention and discrete state transitions, limit the model's expressiveness. For instance, in a single message-passing step, the model cannot compute certain aggregate functions, like averaging over neighbors. This restricts the method's applicability to algorithms that fit within these constraints.\n- Scope of Evaluation: The experimental evaluation focuses on specific algorithmic tasks where the proposed method aligns well. It remains unclear how the model would perform on more complex algorithms that require different computational primitives or continuous manipulations beyond simple increments.\n- Training Without Hints: While the method achieves excellent results with hint supervision, training without hints is identified as challenging. This limitation reduces the method's applicability in scenarios where intermediate algorithmic steps (hints) are unavailable.\n- Sensitivity to Hyperparameters: The paper mentions that certain hyperparameters, like the number of discrete states, significantly impact performance, especially when training without hints. However, there is limited discussion on how sensitive the model is to these hyperparameters and the implications for generalization.\n- Presentation Details: While the paper is generally clear, some sections could benefit from additional explanations. For example, the scalar updater's mechanism could be elaborated further to better understand if someone is not in the NAR field.", "questions": "1. How does the proposed method handle algorithms that require more complex continuous manipulations or aggregate functions beyond simple increments and selections? Can the model be extended to support such algorithms without losing the benefits of discretization and interpretability?\n2. Are there potential strategies to improve training without hint supervision? How does the model perform on tasks without hints when compared to continuous models?\n3. The constraints improve generalization but reduce expressiveness. Is there a way to balance this trade-off by selectively relaxing some constraints?\n4. Can the proposed separation between discrete and continuous data flows be applied to other neural network architectures beyond attention-based models? What challenges might arise in such adaptations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel approach to neural algorithmic reasoning by enforcing neural networks to operate with discrete states and separating discrete and continuous data flows. The authors propose a model that integrates hard attention mechanisms, feature discretization, and a separation between discrete computations and continuous inputs (scalars). This design aims to align neural network computations closely with classical algorithms and thus improves out-of-distribution generalization and interpretability. The method is evaluated on several algorithmic tasks from the SALSA-CLRS benchmark, including BFS, DFS, Prim's algorithm, Dijkstra's algorithm, Maximum Independent Set (MIS), and Eccentricity calculations. The proposed Discrete Neural Algorithmic Reasoner (DNAR) achieves perfect test scores on these tasks, even on graphs significantly larger than those seen during training. The authors also discuss the limitations of their approach and potential directions for future work.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "- Originality: The paper presents a novel method that enforces discrete state transitions in neural networks, which is a significant departure from traditional continuous representations. By integrating hard attention and separating discrete and continuous data flows, the authors address key challenges in neural algorithmic reasoning, particularly out-of-distribution generalization, and interpretability.\n- Quality: The experimental results are strong, with DNAR achieving perfect test scores across multiple algorithmic tasks and graph sizes. The comparison with baseline models and state-of-the-art methods demonstrates the effectiveness of the proposed approach.\n- Clarity: The paper is generally well-written and structured. The authors explain their methodology, including architectural choices and training procedures. The inclusion of diagrams and tables aids in understanding the proposed model and its performance.\n- Significance: The work contributes to the field by showing that neural networks can be designed to mimic classical algorithms with perfect generalization and interpretability. This has implications for developing reliable and trustworthy AI systems that can be formally verified.", "weaknesses": "- Expressiveness Limitations: The enforced constraints, such as hard attention and discrete state transitions, limit the model's expressiveness. For instance, in a single message-passing step, the model cannot compute certain aggregate functions, like averaging over neighbors. This restricts the method's applicability to algorithms that fit within these constraints.\n- Scope of Evaluation: The experimental evaluation focuses on specific algorithmic tasks where the proposed method aligns well. It remains unclear how the model would perform on more complex algorithms that require different computational primitives or continuous manipulations beyond simple increments.\n- Training Without Hints: While the method achieves excellent results with hint supervision, training without hints is identified as challenging. This limitation reduces the method's applicability in scenarios where intermediate algorithmic steps (hints) are unavailable.\n- Sensitivity to Hyperparameters: The paper mentions that certain hyperparameters, like the number of discrete states, significantly impact performance, especially when training without hints. However, there is limited discussion on how sensitive the model is to these hyperparameters and the implications for generalization.\n- Presentation Details: While the paper is generally clear, some sections could benefit from additional explanations. For example, the scalar updater's mechanism could be elaborated further to better understand if someone is not in the NAR field.", "questions": "1. How does the proposed method handle algorithms that require more complex continuous manipulations or aggregate functions beyond simple increments and selections? Can the model be extended to support such algorithms without losing the benefits of discretization and interpretability?\n2. Are there potential strategies to improve training without hint supervision? How does the model perform on tasks without hints when compared to continuous models?\n3. The constraints improve generalization but reduce expressiveness. Is there a way to balance this trade-off by selectively relaxing some constraints?\n4. Can the proposed separation between discrete and continuous data flows be applied to other neural network architectures beyond attention-based models? What challenges might arise in such adaptations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730718136492}, {"id": "kDbPQ8mbxN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6535/Reviewer_BsVb"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "The authors make some modifications to Transformer training to make it more effective for algorithmic reasoning tasks. In particular, they constrain it to learn hard attention, and separate discrete and continuous \"flows\" to prevent information loss. Overall these modicfications result in a model than the GIN and PGN baselines they consider.", "review_text": "The authors make some modifications to Transformer training to make it more effective for algorithmic reasoning tasks. In particular, they constrain it to learn hard attention, and separate discrete and continuous \"flows\" to prevent information loss. Overall these modicfications result in a model than the GIN and PGN baselines they consider.", "strengths": "The approach seems to work much better than the baselines considered, and achieves perfect performance on the two datasets studied in this article.", "weaknesses": "I found this paper to be difficult to read, but I'm not an expert in the area. It would be useful to see each of the design decisions in this paper ablated, with their corresponding effect on the datasets covered in this research.", "questions": "1. I'm very much not an expert in this area, so I did a bit of a literature search on related papers in the field. I found one such paper [1], used a text-based version of the CLRS-3o, called TextCLRS-text. Does it make sense to use that dataset here? Or to compare to the proposed TransNAR architecture?\n\n2. As mentioned earlier, it would be useful to ablate the different algorithmic decisions made in this paper. In its current form, I can't ascertain which of the architectural decisions is responsible for which gains on these datasets.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors make some modifications to Transformer training to make it more effective for algorithmic reasoning tasks. In particular, they constrain it to learn hard attention, and separate discrete and continuous \"flows\" to prevent information loss. Overall these modicfications result in a model than the GIN and PGN baselines they consider.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The approach seems to work much better than the baselines considered, and achieves perfect performance on the two datasets studied in this article.", "weaknesses": "I found this paper to be difficult to read, but I'm not an expert in the area. It would be useful to see each of the design decisions in this paper ablated, with their corresponding effect on the datasets covered in this research.", "questions": "1. I'm very much not an expert in this area, so I did a bit of a literature search on related papers in the field. I found one such paper [1], used a text-based version of the CLRS-3o, called TextCLRS-text. Does it make sense to use that dataset here? Or to compare to the proposed TransNAR architecture?\n\n2. As mentioned earlier, it would be useful to ablate the different algorithmic decisions made in this paper. In its current form, I can't ascertain which of the architectural decisions is responsible for which gains on these datasets.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730716027793}, {"id": "iNso7OsiSJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6535/Reviewer_865w"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper proposes a use of a finite set of predefined discrete states and manipulation of continuous inputs to improve limitations of current paradigms in neural algorithmic reasoning such as redundant dependencies in algorithmic data. The method achieves perfect performance on chosen tasks in SALSA-CLRS and CLRS-30 benchmarks and introduces a starting perspective on use of discrete states to pave the way for future methods in discrete neural algorithmic reasoning and their interpretability.", "review_text": "The paper proposes a use of a finite set of predefined discrete states and manipulation of continuous inputs to improve limitations of current paradigms in neural algorithmic reasoning such as redundant dependencies in algorithmic data. The method achieves perfect performance on chosen tasks in SALSA-CLRS and CLRS-30 benchmarks and introduces a starting perspective on use of discrete states to pave the way for future methods in discrete neural algorithmic reasoning and their interpretability.", "strengths": "- Impressive performance achieved across selected tasks in SALSA-CLRS and CLRS-30 benchmarks\n- Interpretability and simplicity of the proposed model which utilizes set of discrete states to capture continuous inputs using edge priorities\n- Further perspectives on paving way for discrete neural algorithmic reasoners", "weaknesses": "- There is limited explanation of how the discrete states are computed, the paper extensively discusses related work in sections 1 and 2, perhaps it would be better if this space was used for providing further detail on proposed architecture\n- Work is limited in providing mathematical / theoretical definitions or explanation of discrete state space and manipulation with continuous inputs which is the main novelty", "questions": "1. Results show perfect performance on SALSA-CLRS and CLRS-30 benchmarks. Could you elaborate on reported results and 0 variance between your seeds given that this is not the case in majority of comparable baselines or ablation experiments? Could you please report variance on results in table 1? \n2. Could you please explain the differences in computed discrete states used for SALSA-CLRS and CLRS-30 tasks as well as how the model performs on out-of-distribution data?\n3. Could you provide a pseudocode behind Discretize_nodes and Discretize_edges functions shown on page 4?\n4. Work mentions modifying hints from the benchmark, could you clarify if this was included for tested baselines as well as how the hints are modified?\n5. Could you further elaborate on the addition of a virtual node in your model (section 4.3)?\n6. Could you provide results of the hyperparameter search mentioned in section 6? \n7. Could you provide further detail on reverse-engineering (perhaps a toy example) mentioned in section 6?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a use of a finite set of predefined discrete states and manipulation of continuous inputs to improve limitations of current paradigms in neural algorithmic reasoning such as redundant dependencies in algorithmic data. The method achieves perfect performance on chosen tasks in SALSA-CLRS and CLRS-30 benchmarks and introduces a starting perspective on use of discrete states to pave the way for future methods in discrete neural algorithmic reasoning and their interpretability.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Impressive performance achieved across selected tasks in SALSA-CLRS and CLRS-30 benchmarks\n- Interpretability and simplicity of the proposed model which utilizes set of discrete states to capture continuous inputs using edge priorities\n- Further perspectives on paving way for discrete neural algorithmic reasoners", "weaknesses": "- There is limited explanation of how the discrete states are computed, the paper extensively discusses related work in sections 1 and 2, perhaps it would be better if this space was used for providing further detail on proposed architecture\n- Work is limited in providing mathematical / theoretical definitions or explanation of discrete state space and manipulation with continuous inputs which is the main novelty", "questions": "1. Results show perfect performance on SALSA-CLRS and CLRS-30 benchmarks. Could you elaborate on reported results and 0 variance between your seeds given that this is not the case in majority of comparable baselines or ablation experiments? Could you please report variance on results in table 1? \n2. Could you please explain the differences in computed discrete states used for SALSA-CLRS and CLRS-30 tasks as well as how the model performs on out-of-distribution data?\n3. Could you provide a pseudocode behind Discretize_nodes and Discretize_edges functions shown on page 4?\n4. Work mentions modifying hints from the benchmark, could you clarify if this was included for tested baselines as well as how the hints are modified?\n5. Could you further elaborate on the addition of a virtual node in your model (section 4.3)?\n6. Could you provide results of the hyperparameter search mentioned in section 6? \n7. Could you provide further detail on reverse-engineering (perhaps a toy example) mentioned in section 6?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730685709320}, {"id": "NYBAH75rUC", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6535/Reviewer_2sp9"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper studies neural algorithmic reasoning and proposes a discrete transformer-based processor architecture. The authors show that the proposed model empirically achieves perfect size generalization on several algorithms picked from the CLRS benchmark.", "review_text": "This paper studies neural algorithmic reasoning and proposes a discrete transformer-based processor architecture. The authors show that the proposed model empirically achieves perfect size generalization on several algorithms picked from the CLRS benchmark.", "strengths": "- The problem of learning neural algorithmic reasoning is important. I like the overall idea of learning discrete neural reasoners since the previous efforts of learning neural networks with continuous states and or continuous operators failed miserably.\n\n&nbsp;\n\n- The architecture design has some interesting components. For example, the separation of the data and the computational flow that manipulates the data is interesting. In particular, the scalars (continuous input) only affect the computation of attention weights and do not affect the node or edge states. Although designs with the same spirit have appeared before, e.g., in Neural Execution Engines, this part has its own merits.\n\n&nbsp;\n\n\n- The paper is easy to follow and well-written, except that a few technical details are sparse.", "weaknesses": "- I am concerned about the significance of the contributions. If the claims of this paper are all correct, then we just obtain a recipe for learning a neural reasoner to perfectly mimic a known algorithm. However, to make the learning successful, we need to first run the algorithm to collect the full trace of the execution, i.e., the sequence of intermediate states generated by the algorithm (the so-called hints). In other words, we just perfectly fit the “correct algorithm” using a neural network. \n\n  Moreover, I do not see a theoretical guarantee of when this perfect fitting would happen. For example, would it happen on a specific class of algorithms or any algorithm? See my next comment on the claim about “the guarantee”.\n\n  This is unsatisfying since the goal of neural algorithmic reasoning is to learn correct algorithms from data without knowing the algorithms. From this perspective, the really interesting and valuable part is the exploration under the no-hints setting. However, in Section 6, the authors did not provide any experimental results and just stated that their model never achieved perfect validation scores. In short, the authors should provide a thorough empirical study of their proposed model under the no-hints setting and compare it with other approaches on benchmarks like CLRS. \n\n&nbsp;\n\n- In Section 5, the claim “we can guarantee that for any graph size, the model will mirror the desired algorithm, which is correct for any test size” is quite strong. In my opinion, such a strong claim needs rigorous theoretical proof. However, the authors only provide some vague arguments to support this claim. \n  In particular, can you elaborate on the following two questions?\n  1) How do you confirm that the attention block indeed operates as “select_best” selector?\n  I neither see any empirical investigation nor theoretical proof on this point.\n  2) Even if the attention block operates as “select_best” selector, why does this condition lead to the above claim?\n  You should at least demonstrate your detailed logic using an example algorithm like DFS.\n\n&nbsp;\n\n- I think the size of the discretized states would matter a lot to the expressivity of the proposed neural reasoner, i.e., what algorithms can be represented by the designed class of neural networks. However, I did not see the experimental study and discussion on its effect.\n\n&nbsp;\n\n- An interesting experiment is to check if there is some sort of phase transition in terms of problem size. Specifically, I would imagine the phenomenon of perfect fitting to the correct algorithm would disappear if we decreased the problem size in training. Then, what is the minimum problem size to ensure it fits perfectly for a particular algorithm?\n\n&nbsp;\n\n- The details of how to train task-dependent encoders and decoders are not provided, and how they affect the processor's training is not discussed at all. I would imagine the quality of the encoded embedding is quite important to the success of learning the processor, even with teacher forcing. \n\n&nbsp;\n\n- In section 3.3, I get the high-level idea that the scalars (continuous input) only affect the computation of attention weights and do not affect the node or edge states. However, the description of the idea in the 2nd paragraph of Section 3.3 is not so clear that I could not figure out how exactly the computation is designed. It would be great to either write the equations and or illustrate the computational graph.", "questions": "Please see my comments in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies neural algorithmic reasoning and proposes a discrete transformer-based processor architecture. The authors show that the proposed model empirically achieves perfect size generalization on several algorithms picked from the CLRS benchmark.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The problem of learning neural algorithmic reasoning is important. I like the overall idea of learning discrete neural reasoners since the previous efforts of learning neural networks with continuous states and or continuous operators failed miserably.\n\n&nbsp;\n\n- The architecture design has some interesting components. For example, the separation of the data and the computational flow that manipulates the data is interesting. In particular, the scalars (continuous input) only affect the computation of attention weights and do not affect the node or edge states. Although designs with the same spirit have appeared before, e.g., in Neural Execution Engines, this part has its own merits.\n\n&nbsp;\n\n\n- The paper is easy to follow and well-written, except that a few technical details are sparse.", "weaknesses": "- I am concerned about the significance of the contributions. If the claims of this paper are all correct, then we just obtain a recipe for learning a neural reasoner to perfectly mimic a known algorithm. However, to make the learning successful, we need to first run the algorithm to collect the full trace of the execution, i.e., the sequence of intermediate states generated by the algorithm (the so-called hints). In other words, we just perfectly fit the “correct algorithm” using a neural network. \n\n  Moreover, I do not see a theoretical guarantee of when this perfect fitting would happen. For example, would it happen on a specific class of algorithms or any algorithm? See my next comment on the claim about “the guarantee”.\n\n  This is unsatisfying since the goal of neural algorithmic reasoning is to learn correct algorithms from data without knowing the algorithms. From this perspective, the really interesting and valuable part is the exploration under the no-hints setting. However, in Section 6, the authors did not provide any experimental results and just stated that their model never achieved perfect validation scores. In short, the authors should provide a thorough empirical study of their proposed model under the no-hints setting and compare it with other approaches on benchmarks like CLRS. \n\n&nbsp;\n\n- In Section 5, the claim “we can guarantee that for any graph size, the model will mirror the desired algorithm, which is correct for any test size” is quite strong. In my opinion, such a strong claim needs rigorous theoretical proof. However, the authors only provide some vague arguments to support this claim. \n  In particular, can you elaborate on the following two questions?\n  1) How do you confirm that the attention block indeed operates as “select_best” selector?\n  I neither see any empirical investigation nor theoretical proof on this point.\n  2) Even if the attention block operates as “select_best” selector, why does this condition lead to the above claim?\n  You should at least demonstrate your detailed logic using an example algorithm like DFS.\n\n&nbsp;\n\n- I think the size of the discretized states would matter a lot to the expressivity of the proposed neural reasoner, i.e., what algorithms can be represented by the designed class of neural networks. However, I did not see the experimental study and discussion on its effect.\n\n&nbsp;\n\n- An interesting experiment is to check if there is some sort of phase transition in terms of problem size. Specifically, I would imagine the phenomenon of perfect fitting to the correct algorithm would disappear if we decreased the problem size in training. Then, what is the minimum problem size to ensure it fits perfectly for a particular algorithm?\n\n&nbsp;\n\n- The details of how to train task-dependent encoders and decoders are not provided, and how they affect the processor's training is not discussed at all. I would imagine the quality of the encoded embedding is quite important to the success of learning the processor, even with teacher forcing. \n\n&nbsp;\n\n- In section 3.3, I get the high-level idea that the scalars (continuous input) only affect the computation of attention weights and do not affect the node or edge states. However, the description of the idea in the 2nd paragraph of Section 3.3 is not so clear that I could not figure out how exactly the computation is designed. It would be great to either write the equations and or illustrate the computational graph.", "questions": "Please see my comments in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730619544938}, {"id": "tuUhOLiQlI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6535/Reviewer_19Nf"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper addresses the challenges of generalization and interpretability in neural algorithmic reasoning (NAR) by introducing a novel architecture that guides the model to follow algorithmic execution as a series of finite, predefined states. The proposed architecture has three main components: feature discretization, hard attention, and separation of discrete and continuous data flows. The empirical results show that the model achieves perfect scores in both single-task and multi-task experiments across various algorithms. Additionally, the architecture enhances interpretability, allowing validation of the correct execution of desired algorithms.", "review_text": "This paper addresses the challenges of generalization and interpretability in neural algorithmic reasoning (NAR) by introducing a novel architecture that guides the model to follow algorithmic execution as a series of finite, predefined states. The proposed architecture has three main components: feature discretization, hard attention, and separation of discrete and continuous data flows. The empirical results show that the model achieves perfect scores in both single-task and multi-task experiments across various algorithms. Additionally, the architecture enhances interpretability, allowing validation of the correct execution of desired algorithms.", "strengths": "1. The writing is generally clear, though some areas could be further elaborated.\n2. The paper is well-motivated, as both OOD generalization and interpretability in NAR are both important questions. The proposed architecture of separating the discrete and continuous data flows are novel and effective.\n3. The perfect scores across algorithms are impressive, especially given the model’s capacity for size generalization on graphs 100 times larger, outperforming strong baselines.\n4. The proposed architecture consists of three design components, each of which is validated through ablation studies to demonstrate its effectiveness.", "weaknesses": "1. While the perfect scores achieved in the experiments are impressive, the paper could be strengthened by testing on a wider range of algorithms. Although both parallel (e.g., BFS) and sequential (e.g., Prim) algorithms are covered, most algorithms studied are graph-based, where previous NAR methods have already proven effective. The CLRS-30 dataset includes a broader variety of algorithms (e.g., sorting and search), where many NAR methods can struggle. Although SALSA-CLRS was chosen for its thorough OOD evaluation, testing the model on the CLRS-30 dataset with larger graph sizes would add valuable insights due to the dataset’s extensive algorithm coverage.\n\n2. Another limitation is the lack of application to real-world datasets, as the authors note in the Future Work section. A significant advantage of NAR methods is their ability to operate on high-dimensional data by utilizing a pretrained model that mimic algorithms. This presents an additional OOD challenge with potential distribution shifts in real-world data. Evaluating the proposed architecture on real-world datasets would demonstrate its practical value; even a single real-world experiment, as seen in related works (e.g., Numeroso et al., 2023), could significantly strengthen the method's implications.\n\n3. Although interpretability is a valuable strength of the proposed method, it is unclear how this model’s interpretability, achieved through analyzing state transitions and attention blocks, differs from other NAR approaches, which also allow interpretation of intermediate executions (e.g., using decoder outputs to indicate a node’s current predecessor in a sorting algorithm).", "questions": "1. For the multi-task experiments, do the results in Table 2 correspond to these settings? It is unclear which algorithms were trained simultaneously and how this compares to the baselines’ configurations.\n\n2. Without hints, DNAR struggles to generalize. Could you clarify the reverse-engineering issue described? Was any intermediate approach, such as noisy teacher forcing or teacher forcing decay, tested to examine the impact of hints on DNAR?\n\n3. How were GIN and PGN used as baselines? Were they employed to treat this as an end-to-end node-level prediction task?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the challenges of generalization and interpretability in neural algorithmic reasoning (NAR) by introducing a novel architecture that guides the model to follow algorithmic execution as a series of finite, predefined states. The proposed architecture has three main components: feature discretization, hard attention, and separation of discrete and continuous data flows. The empirical results show that the model achieves perfect scores in both single-task and multi-task experiments across various algorithms. Additionally, the architecture enhances interpretability, allowing validation of the correct execution of desired algorithms.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. The writing is generally clear, though some areas could be further elaborated.\n2. The paper is well-motivated, as both OOD generalization and interpretability in NAR are both important questions. The proposed architecture of separating the discrete and continuous data flows are novel and effective.\n3. The perfect scores across algorithms are impressive, especially given the model’s capacity for size generalization on graphs 100 times larger, outperforming strong baselines.\n4. The proposed architecture consists of three design components, each of which is validated through ablation studies to demonstrate its effectiveness.", "weaknesses": "1. While the perfect scores achieved in the experiments are impressive, the paper could be strengthened by testing on a wider range of algorithms. Although both parallel (e.g., BFS) and sequential (e.g., Prim) algorithms are covered, most algorithms studied are graph-based, where previous NAR methods have already proven effective. The CLRS-30 dataset includes a broader variety of algorithms (e.g., sorting and search), where many NAR methods can struggle. Although SALSA-CLRS was chosen for its thorough OOD evaluation, testing the model on the CLRS-30 dataset with larger graph sizes would add valuable insights due to the dataset’s extensive algorithm coverage.\n\n2. Another limitation is the lack of application to real-world datasets, as the authors note in the Future Work section. A significant advantage of NAR methods is their ability to operate on high-dimensional data by utilizing a pretrained model that mimic algorithms. This presents an additional OOD challenge with potential distribution shifts in real-world data. Evaluating the proposed architecture on real-world datasets would demonstrate its practical value; even a single real-world experiment, as seen in related works (e.g., Numeroso et al., 2023), could significantly strengthen the method's implications.\n\n3. Although interpretability is a valuable strength of the proposed method, it is unclear how this model’s interpretability, achieved through analyzing state transitions and attention blocks, differs from other NAR approaches, which also allow interpretation of intermediate executions (e.g., using decoder outputs to indicate a node’s current predecessor in a sorting algorithm).", "questions": "1. For the multi-task experiments, do the results in Table 2 correspond to these settings? It is unclear which algorithms were trained simultaneously and how this compares to the baselines’ configurations.\n\n2. Without hints, DNAR struggles to generalize. Could you clarify the reverse-engineering issue described? Was any intermediate approach, such as noisy teacher forcing or teacher forcing decay, tested to examine the impact of hints on DNAR?\n\n3. How were GIN and PGN used as baselines? Were they employed to treat this as an end-to-end node-level prediction task?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730580141362}], "openreview_url": "https://openreview.net/forum?id=yLmcYLP3Yd", "arxiv_id": "2402.11628", "paper_pdf": "papers/yLmcYLP3Yd.pdf", "paper_pdf_sha256": "ae3b1317f66c43efa8a65f094d15a754865be2e3bc4187f3db6892ebe110e005", "paper_pdf_bytes": 426356, "paper_pdf_source": "openreview", "code_url": "https://github.com/yandex-research/dnar", "code_repository": "yandex-research/dnar", "code_commit": "12f3f0bd0a70568386e43b2d841422f28ed53698", "code_archive": "repos/yLmcYLP3Yd.zip", "code_archive_sha256": "76cf854f1efe2f7d4ae01fb75ae2602795cd5ea6d55c58919cbecf270d66a157", "code_archive_bytes": 17641, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 15, "github_languages": {"Python": 38447}, "github_archived": false, "github_pushed_at": "2024-09-10T15:48:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discrete-neural-algorithmic-reasoning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VLnODGVVAsL", "year": 2023, "status": "rejected", "title": "Anchor Sampling for Federated Learning with Partial Client Participation", "authors": ["Feijie Wu", "Song Guo", "Zhihao Qu", "Shiqi He", "Ziming Liu"], "authorids": ["~Feijie_Wu1", "~Song_Guo5", "~Zhihao_Qu1", "~Shiqi_He1", "~Ziming_Liu1"], "authors_source": "OpenReview API", "abstract": "In federated learning, the support of partial client participation offers a flexible training strategy, but it deteriorates the model training efficiency. In this paper, we propose a framework FedAMD to improve the convergence property and maintain flexibility. The core idea is anchor sampling, which disjoints the partial participants into anchor and miner groups. Each client in the anchor group aims at the local bullseye with the gradient computation using a large batch. Guided by the bullseyes, clients in the miner group steer multiple near-optimal local updates using small batches and update the global model. With the joint efforts from both groups, FedAMD is able to accelerate the training process as well as improve the model performance. Measured by $\\epsilon$-approximation and compared to the state-of-the-art first-order methods, FedAMD achieves the convergence by up to $O(1/\\epsilon)$ fewer communication rounds under non-convex objectives. In specific, we achieve a linear convergence rate under PL conditions. Empirical studies on real-world datasets validate the effectiveness of FedAMD and demonstrate the superiority of our proposed algorithm: Not only does it considerably save computation and communication costs, but also the test accuracy significantly improves. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "1Ah36DRJdF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3563/Reviewer_B281"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a new federated learning algorithm which allows clients to perform their updates on the model using the gradient of the model obtained by other clients. This involves data samples stored on other clients memory in updating the model locally by a client. The paper provides convergence analysis for the proposed algorithm and examines its performance through experiments.", "review_text": "In summary, the paper proposes a new federated learning algorithm at the cost of the need for a server with a large memory capacity. The paper proves the convergence of the algorithm. However, the paper lacks a detailed comparison of the proposed algorithm's convergence with the convergence of FedAvg. ", "strengths": "Strength: The paper proposes a novel way to enable clients to share their gradients and use this information to update the model locally. The paper proves the convergence of the proposed algorithm and provide convincing experimental results to validate the theoretical analysis provided in the paper.\n\nWeaknesses:\n\n- Using the proposed algorithm, the server has to store a caching gradient matrix whose dimensions scales with the number of clients. When the number of clients is so large, to implement the proposed algorithm a server with a very large memory is required.\n\n- It is beneficial for the paper if the convergence rate of the proposed FedAMD is compared to FedAvg to see what are the benefits of FedAMD compared to FedAvg in terms of convergence. FedAMD requires larger memory from the server side and it is interesting to see if FedAMD uses this memory successfully to obtain better convergence rate than that of FedAvg.\n\n- From reading the Algorithm 1 on page 4, the role of cached gradients in updating models is not clear. It would be great if Algorithm 1 can be revised such that the role of cached gradient become obvious.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a new federated learning algorithm which allows clients to perform their updates on the model using the gradient of the model obtained by other clients. This involves data samples stored on other clients memory in updating the model locally by a client. The paper provides convergence analysis for the proposed algorithm and examines its performance through experiments.", "strength_and_weaknesses": "Strength: The paper proposes a novel way to enable clients to share their gradients and use this information to update the model locally. The paper proves the convergence of the proposed algorithm and provide convincing experimental results to validate the theoretical analysis provided in the paper.\n\nWeaknesses:\n\n- Using the proposed algorithm, the server has to store a caching gradient matrix whose dimensions scales with the number of clients. When the number of clients is so large, to implement the proposed algorithm a server with a very large memory is required.\n\n- It is beneficial for the paper if the convergence rate of the proposed FedAMD is compared to FedAvg to see what are the benefits of FedAMD compared to FedAvg in terms of convergence. FedAMD requires larger memory from the server side and it is interesting to see if FedAMD uses this memory successfully to obtain better convergence rate than that of FedAvg.\n\n- From reading the Algorithm 1 on page 4, the role of cached gradients in updating models is not clear. It would be great if Algorithm 1 can be revised such that the role of cached gradient become obvious.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and provides a novel algorithm. I could not check proofs carefully.", "summary_of_the_review": "In summary, the paper proposes a new federated learning algorithm at the cost of the need for a server with a large memory capacity. The paper proves the convergence of the algorithm. However, the paper lacks a detailed comparison of the proposed algorithm's convergence with the convergence of FedAvg. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666859850970}, {"id": "ntAG4Ck5qz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3563/Reviewer_Moju"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problem of improving convergence in Federated Learning especially under partial client participation settings. Differently from previous work, which either used large batch sizes or variance reduction techniques while resorting to full client participation, the authors propose a randomization of FedAvg and FedSGD. In that, the clients switch from being either anchor or minor groups. Convergence results are derived for non-convex losses under general conditions. Experimental results illustrate the efficacy of the approach. ", "review_text": "The paper has weaknesses in terms of the convergence proofs with the dependence on the total number of clients being significantly worse as compared to other schemes.  The applicability, usability and originality of the paper is undermined by some assumptions which either don't hold or are difficult to verify in practice. Moreover, the randomization of a client either doing a local-SGD style update or no update, can lead to stretches with very few clients contributing to the update. Having said that, I would be more than happy to increase my score if my concerns are addressed.", "strengths": "Strengths\n+ The paper is very well written and the algorithm is well motivated.\n+ The convergence results show an $O\\left(\\frac{1}{\\epsilon}\\right)$ for general non-convex losses and linear convergence with the extra PL assumptions which is an improvement over the state of the art.\n+ The proposed algorithm FedAMP also shows the efficacy of schemes which resort to heterogeneous batch sizes and updating schemes while resorting to variance reduction.\n+ The experimental results are comprehensive in terms of comparison with other baselines and tends to outperform other baselines.\n\n\nWeaknesses\n\n- Though the algorithm claims to use partial client participation; it requires a full client participation at the first step which in turn requires each gradient from each client to be individually available to the server and that too persisted over time instead of being available in an aggregated manner. This severely undermines the privacy-preserving nature of FL as it makes each client susceptible to gradient inversion attacks. It is also not clear how can this be alleviated without resorting to additional storage in each device.\n- The convergence results while improving in terms of $\\epsilon$-accuracy of the error, is substantially worse as compared to other baselines in terms of dependence on the total number of clients. While other algorithms scale inversely with respect to the total number of clients, FedAMP scales linearly with the number of clients in the setup. Unless, the number of selected clients is always a function of the total number of clients, it seems the performance of FedAMP gets worse with increasing number of clients, if the number of clients selected per round is constant. It is also worth noting the performance benefit comes at the expense of additional assumption 3.\n- An inherent assumption made by the authors is that the number of samples across clients is the same. The method by which a client is assigned to being an anchor or minor is solely decided by a coin toss and not based on the number of samples. In the general case, some clients may have fewer samples, in the case of which the number of local epochs and the batch size can't be uniform. It is also not clear how the choice $b$ and $b^{'}$ affects the algorithm performance.\n- The authors miss comparing and contrasting with FedNova by Wang et.al., which especially takes care of the case where different clients undergo different number of rounds along with the flexibility of running different local solvers. It would be good to add comparisons with FedDyn and FedOPT which are highly performant baselines.\n- While a theoretically optimal $p$ is derived, it can't be derived in practice as quantities such as the number of samples on each device and $\\sigma^{2}$ is unknown to the server. \n- For the variance reduction phase, it requires additional storage of the past gradient which is the same size as that of the model. This can be particularly prohibitive with larger model and resource constrained devices.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the problem of improving convergence in Federated Learning especially under partial client participation settings. Differently from previous work, which either used large batch sizes or variance reduction techniques while resorting to full client participation, the authors propose a randomization of FedAvg and FedSGD. In that, the clients switch from being either anchor or minor groups. Convergence results are derived for non-convex losses under general conditions. Experimental results illustrate the efficacy of the approach. ", "strength_and_weaknesses": "Strengths\n+ The paper is very well written and the algorithm is well motivated.\n+ The convergence results show an $O\\left(\\frac{1}{\\epsilon}\\right)$ for general non-convex losses and linear convergence with the extra PL assumptions which is an improvement over the state of the art.\n+ The proposed algorithm FedAMP also shows the efficacy of schemes which resort to heterogeneous batch sizes and updating schemes while resorting to variance reduction.\n+ The experimental results are comprehensive in terms of comparison with other baselines and tends to outperform other baselines.\n\n\nWeaknesses\n\n- Though the algorithm claims to use partial client participation; it requires a full client participation at the first step which in turn requires each gradient from each client to be individually available to the server and that too persisted over time instead of being available in an aggregated manner. This severely undermines the privacy-preserving nature of FL as it makes each client susceptible to gradient inversion attacks. It is also not clear how can this be alleviated without resorting to additional storage in each device.\n- The convergence results while improving in terms of $\\epsilon$-accuracy of the error, is substantially worse as compared to other baselines in terms of dependence on the total number of clients. While other algorithms scale inversely with respect to the total number of clients, FedAMP scales linearly with the number of clients in the setup. Unless, the number of selected clients is always a function of the total number of clients, it seems the performance of FedAMP gets worse with increasing number of clients, if the number of clients selected per round is constant. It is also worth noting the performance benefit comes at the expense of additional assumption 3.\n- An inherent assumption made by the authors is that the number of samples across clients is the same. The method by which a client is assigned to being an anchor or minor is solely decided by a coin toss and not based on the number of samples. In the general case, some clients may have fewer samples, in the case of which the number of local epochs and the batch size can't be uniform. It is also not clear how the choice $b$ and $b^{'}$ affects the algorithm performance.\n- The authors miss comparing and contrasting with FedNova by Wang et.al., which especially takes care of the case where different clients undergo different number of rounds along with the flexibility of running different local solvers. It would be good to add comparisons with FedDyn and FedOPT which are highly performant baselines.\n- While a theoretically optimal $p$ is derived, it can't be derived in practice as quantities such as the number of samples on each device and $\\sigma^{2}$ is unknown to the server. \n- For the variance reduction phase, it requires additional storage of the past gradient which is the same size as that of the model. This can be particularly prohibitive with larger model and resource constrained devices.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written, easy to understand in terms of clarity and is of quality. The paper modifies well known techniques used in the context of FL and carefully tunes them so as to attain better rates. Hence, the originality is moderate.", "summary_of_the_review": "The paper has weaknesses in terms of the convergence proofs with the dependence on the total number of clients being significantly worse as compared to other schemes.  The applicability, usability and originality of the paper is undermined by some assumptions which either don't hold or are difficult to verify in practice. Moreover, the randomization of a client either doing a local-SGD style update or no update, can lead to stretches with very few clients contributing to the update. Having said that, I would be more than happy to increase my score if my concerns are addressed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666599548202}, {"id": "YKgmUEja4HP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3563/Reviewer_UWdf"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a framework FedAMD, that disjoints the partial participants into anchor and miner groups. Clients in anchor group target to discover the bullseyes based on their local data distribution, while clients in the miner group perform multiple local updates and finally drive the update of the global model guided by the global bullseye. Under the partial-client scenario, this paper gives many theoretical proofs that FedAMD achieves sublinear speedup under non-convex objectives and linear speedup under the PL condition. Experimental results demonstrate that FedAMD is superior to SOTA works compared with BVR-L-SGD, FedAvg, FedPAGE, SCAFFOLD on Fashion MNIST.", "review_text": "In general, it is a good paper. It has the SOTA results, many theoretical proofs and analysis. However, there are some weaknesses in the setting and analysis of the experiments section.", "strengths": "Strength:\n1.The proposed FedAMD achieves sublinear speedup under non-convex objectives and linear speedup under the PL condition.\n2.The proposed FedAMD is superior to SOTA works compared with BVR-L-SGD, FedAvg, FedPAGE, SCAFFOLD on Fashion MNIST.\n3.The proofs of theories are detailed and correct, and they are consistent with the experimental results.\n\nWeaknesses:\n1.It seems that there is no analysis about the influence of initial model (line 1) on the experimental results in this paper. In fact, the influence of initial model exists (assuming that the best model is set at the beginning). It is suggested to add analysis about the influence of initial model.\n2.The experiment is too simple. \"We train a convolutional neural network LeNet-5 using Fashion MNIST\". Such a specific experiment may affect the degree of confidence of the experiment results.\n3.This paper proposes two strategies(constant and sequential) for selecting anchor and miner groups. But it would be better if there were a strategy based on the performance of each client.\n4.The analysis of g(i)t,k (line 19)is reasonable, but it is worth discussing whether a factor α is needed for second and third items.\n5.For the analysis of Figure 1, it seems that only Communication Rounds=400 is considered(why is 400?). Such analysis seems to be one-sided, and there are other considerations in fact, such as the change rate of the training loss.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a framework FedAMD, that disjoints the partial participants into anchor and miner groups. Clients in anchor group target to discover the bullseyes based on their local data distribution, while clients in the miner group perform multiple local updates and finally drive the update of the global model guided by the global bullseye. Under the partial-client scenario, this paper gives many theoretical proofs that FedAMD achieves sublinear speedup under non-convex objectives and linear speedup under the PL condition. Experimental results demonstrate that FedAMD is superior to SOTA works compared with BVR-L-SGD, FedAvg, FedPAGE, SCAFFOLD on Fashion MNIST.", "strength_and_weaknesses": "Strength:\n1.The proposed FedAMD achieves sublinear speedup under non-convex objectives and linear speedup under the PL condition.\n2.The proposed FedAMD is superior to SOTA works compared with BVR-L-SGD, FedAvg, FedPAGE, SCAFFOLD on Fashion MNIST.\n3.The proofs of theories are detailed and correct, and they are consistent with the experimental results.\n\nWeaknesses:\n1.It seems that there is no analysis about the influence of initial model (line 1) on the experimental results in this paper. In fact, the influence of initial model exists (assuming that the best model is set at the beginning). It is suggested to add analysis about the influence of initial model.\n2.The experiment is too simple. \"We train a convolutional neural network LeNet-5 using Fashion MNIST\". Such a specific experiment may affect the degree of confidence of the experiment results.\n3.This paper proposes two strategies(constant and sequential) for selecting anchor and miner groups. But it would be better if there were a strategy based on the performance of each client.\n4.The analysis of g(i)t,k (line 19)is reasonable, but it is worth discussing whether a factor α is needed for second and third items.\n5.For the analysis of Figure 1, it seems that only Communication Rounds=400 is considered(why is 400?). Such analysis seems to be one-sided, and there are other considerations in fact, such as the change rate of the training loss.", "clarity,_quality,_novelty_and_reproducibility": "Quality: The proposed FedAMD is superior to state-of-the-art works, and there are many theories to support the experimental results in this paper.\nClarity: The algorithm 1 is detailed and reasonable, but some lines(line 1 and line 19)need additional analysis and discussion.\nOriginality: This work may be the first work to analyze the effectiveness of large batches under partial client participation.\n", "summary_of_the_review": "In general, it is a good paper. It has the SOTA results, many theoretical proofs and analysis. However, there are some weaknesses in the setting and analysis of the experiments section.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666570291159}], "openreview_url": "https://openreview.net/forum?id=VLnODGVVAsL", "arxiv_id": "2206.05891", "paper_pdf": "papers/VLnODGVVAsL.pdf", "paper_pdf_sha256": "2f51defed0fe2dadd68ec7ed4dd642ce49e267a41b3effcb1ff6d0deab199e7b", "paper_pdf_bytes": 1395729, "paper_pdf_source": "openreview", "code_url": "https://github.com/HarliWu/FedAMD", "code_repository": "HarliWu/FedAMD", "code_commit": "73596d55e15a3ef6f404e5a313f50bc05a328c55", "code_archive": "repos/VLnODGVVAsL.zip", "code_archive_sha256": "c0a44531ab2903feef431e58f2eb4b87e4397d8f1933e0a8312dba08c8047aea", "code_archive_bytes": 28071, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 20, "github_languages": {"Python": 87743}, "github_archived": false, "github_pushed_at": "2024-04-09T17:55:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerating-federated-learning-via-sampling"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7Z7u2z1Ornl", "year": 2022, "status": "rejected", "title": "Pruning Edges and Gradients to Learn Hypergraphs from Larger Sets", "authors": ["David W Zhang", "Gertjan J. Burghouts", "Cees G. M. Snoek"], "authorids": ["~David_W_Zhang1", "~Gertjan_J._Burghouts1", "~Cees_G._M._Snoek1"], "authors_source": "OpenReview API", "abstract": "This paper aims for set-to-hypergraph prediction, where the goal is to infer the set of relations for a given set of entities. This is a common abstraction for applications in particle physics, biological systems and combinatorial optimization. We address two common scaling problems encountered in set-to-hypergraph tasks that limit the size of the input set: the exponentially growing number of hyperedges and the run-time complexity, both leading to higher memory requirements. We make three contributions. First, we propose to predict and supervise the \\emph{positive} edges only, which changes the asymptotic memory scaling from exponential to linear. Second, we introduce a training method that encourages iterative refinement of the predicted hypergraph, which allows us to skip iterations in the backward pass for improved efficiency and constant memory usage. Third, we combine both contributions in a single set-to-hypergraph model that enables us to address problems with larger input set sizes. We provide ablations for our main technical contributions and show that our model outperforms prior state-of-the-art, especially for larger sets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "9UsSblgPd45", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1834/Reviewer_hMC3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper addressed two scaling problems in set-to-hypergraph prediction by pruning edges and gradients. The authors made an assumption that the maximum number of edges is k to reduce the memory complexity from O(2^n) to O(kn). Backpropagation with skips was also used to reduce the computation complexity. A series of experiments were done to show the method is empirically better than the baselines.", "review_text": "Strengths: \n*  Proposition 1 is an interesting finding. \n\n* There are plenty of experimental results to show the empirical advantages. \n\nWeaknesses:\n* The technical contributions of this work seem not to be very novel. Proposition 1 seems to be the most novel argument but I think it is problematic (see the later criticism). The Hungarian loss to compare sets and the skip connection are standard techniques. \n\n* The proposed method suffers from high complexity. May I know how many rows of the incidence matrix I? Moreover, how to predefine the number of hyperedges/ the row # of I? The Hungarian loss has complexity O(#row(I)^2*#column(I)) or O(#row(I)*#column(I)^2) which could be high if we have a large number of rows or columns (nodes).\n\n* My biggest concern is proposition 1. This is the most important technique argument. However, proposition 1 is not stated in a rigorous way at all. What is the definition of c? Can the authors highlight some intuition between the proof in the main text? I also checked the proof. It seems that the argument is not rigorous. There are quite a lot approximations, assumptions, etc. I do not even believe prop 1 is theoretically right. \n\n* In experimental details, the authors should discuss how big the k can cover the datasets. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper addressed two scaling problems in set-to-hypergraph prediction by pruning edges and gradients. The authors made an assumption that the maximum number of edges is k to reduce the memory complexity from O(2^n) to O(kn). Backpropagation with skips was also used to reduce the computation complexity. A series of experiments were done to show the method is empirically better than the baselines.", "main_review": "Strengths: \n*  Proposition 1 is an interesting finding. \n\n* There are plenty of experimental results to show the empirical advantages. \n\nWeaknesses:\n* The technical contributions of this work seem not to be very novel. Proposition 1 seems to be the most novel argument but I think it is problematic (see the later criticism). The Hungarian loss to compare sets and the skip connection are standard techniques. \n\n* The proposed method suffers from high complexity. May I know how many rows of the incidence matrix I? Moreover, how to predefine the number of hyperedges/ the row # of I? The Hungarian loss has complexity O(#row(I)^2*#column(I)) or O(#row(I)*#column(I)^2) which could be high if we have a large number of rows or columns (nodes).\n\n* My biggest concern is proposition 1. This is the most important technique argument. However, proposition 1 is not stated in a rigorous way at all. What is the definition of c? Can the authors highlight some intuition between the proof in the main text? I also checked the proof. It seems that the argument is not rigorous. There are quite a lot approximations, assumptions, etc. I do not even believe prop 1 is theoretically right. \n\n* In experimental details, the authors should discuss how big the k can cover the datasets. \n", "summary_of_the_review": "Though the set-to-hypergraph prediction is an important topic, the technical contributions of this paper are limited. The main statement is not rigorous and well explained. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635934754341}, {"id": "j9IOl_fLqTT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1834/Reviewer_d3zQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper improves the asymptotic scaling and enables the learning task to have higher-order edges by only representing and supervising a set of positive edges. In common benchmark tests, this paper has proved that the proposed method is superior to previous work while providing more favorable asymptotic scaling behavior. In further evaluation, this paper emphasizes the importance of repetition in solving the inherent complexity of the problem.", "review_text": "The set-to-hypergraph prediction is an interesting problem. The graph size is indeed the bottleneck of this problem. The paper focuses on this point and proposes its solutions. From the experimental results, the method is effective in performance. However, there are some critical defects in this paper.\n\n1, The proof of Proposition 1 (Supervising positive edges only) should be the key to pruning negative edges. However, I find the first sentence of the proof in Appendix A is \"For the sake of rigor, we first summarize the relevant assumptions from ??. \". This statement directly hinders me from understanding the whole proof. This kind of typo is fatal. I hope the authors can revise them in detail.\n\n2, Some sentences are confusing. For example, on Page2 \"if an edge connects every node in V ⊂ V then there exists a relation between the input elements {xi ∈ X|vi ∈ V }. \". I am confused about how an edge can connect every node? One edge at most connects two nodes, right?\n\n3, The paper claims that specifying a maximum number of edges k is sufficiently large to cover all (or most) hypergraphs of interest. I think k should be a hyper-parameter and is very important. There is no content about how to determine and tune k, and the parameter analysis of k.\n\n4, The core work of the paper is to scale the set-to-hypergraph prediction. I am curious about why there is no absolute training time comparison shown in the experimental section. The relative training time shown in Figure 2 is hard to understand for me.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper improves the asymptotic scaling and enables the learning task to have higher-order edges by only representing and supervising a set of positive edges. In common benchmark tests, this paper has proved that the proposed method is superior to previous work while providing more favorable asymptotic scaling behavior. In further evaluation, this paper emphasizes the importance of repetition in solving the inherent complexity of the problem.", "main_review": "The set-to-hypergraph prediction is an interesting problem. The graph size is indeed the bottleneck of this problem. The paper focuses on this point and proposes its solutions. From the experimental results, the method is effective in performance. However, there are some critical defects in this paper.\n\n1, The proof of Proposition 1 (Supervising positive edges only) should be the key to pruning negative edges. However, I find the first sentence of the proof in Appendix A is \"For the sake of rigor, we first summarize the relevant assumptions from ??. \". This statement directly hinders me from understanding the whole proof. This kind of typo is fatal. I hope the authors can revise them in detail.\n\n2, Some sentences are confusing. For example, on Page2 \"if an edge connects every node in V ⊂ V then there exists a relation between the input elements {xi ∈ X|vi ∈ V }. \". I am confused about how an edge can connect every node? One edge at most connects two nodes, right?\n\n3, The paper claims that specifying a maximum number of edges k is sufficiently large to cover all (or most) hypergraphs of interest. I think k should be a hyper-parameter and is very important. There is no content about how to determine and tune k, and the parameter analysis of k.\n\n4, The core work of the paper is to scale the set-to-hypergraph prediction. I am curious about why there is no absolute training time comparison shown in the experimental section. The relative training time shown in Figure 2 is hard to understand for me.\n", "summary_of_the_review": "The paper works on a critical problem and proposes a reasonable solution. However, there are some fatal typos and confusing statements. The experimental section lacks some critical results like absolute running time and only tests on one dataset. The authors should solve my doubts above and modify the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635886053497}, {"id": "d9StyTzQYSg", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1834/Reviewer_ezwC"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a set-to-hypergraph model where the hypergraph is represented by an incidence matrix in contrast to the usually used adjacency tensors. The incidence matrix is an edge times node matrix, which reflects only existing edges and is hence suitable to reduce the memory requirement when the amount of existing edges is sparse. The authors propose for the optimization of the proposed model\n an SGD scheme where a batch update is performed on the hypergraph which is returned by a recurrent neural network after a varying amount of time steps. \nThe experiments compare the proposed approach with 3 competitors on four datasets.  ", "review_text": "I find it hard to assess the contribution of this paper because I feel like I don't understand it. The English is good but there is a lack of definitions and explanations regarding the considered task, terminology, and mathematical notation, which makes it hard to make sense of the proposed method. In addition, the few mathematical definitions which are given do not seem to be complete. For example, the proposed loss function has a trivial solution and functions which are introduced later are not part of the loss (see comments below). Likewise, the experimental setting is not detailed. How the hyperparameters and the optimization parameters are tuned/set is not described for the proposed method and the competitors. As a result, I believe that the most actionable approach for the authors is that I list all the questions and issues I have in the following, to make clear where this paper needs improvement.\n\n# Questions & Issues:\n## Section 1 - Introduction:\n* From what or on the basis of what is the hypergraph learned from the set? What exactly is the input?\n* Based on what do you decide whether an edge exists or not? \n* The terminology of negative edges being also called non-existent should be introduced before this terminology is used in this section.\n## Section 2 - Scaling by pruning non-existing edges:\n* Why do non-existing edges have to be pruned, why are they modeled in the first place?\n* How is a hypergraph defined? How is a hyperedge defined?\n* What is $V\\subset \\mathcal{V}$?\n* What does it mean to \"remove excess rows\" from the incidence matrix? What is an excess row? I would expect that removing the row decreases the dimensionality of the incidence matrix. This is however not reflected in the mathematical notation.\n* What does it mean to \"supervise\" an incidence $I_{ji}$?\n* \"a positive edge in the incidence matrix contains both zeros and ones\" (-> what does that mean? Further:) \"ensuring that the binary classifier sees both positive and negative examples\"(->what is a pos/neg example?)\n* \"the order of the entries is fully decided by the order of the nodes\" -> why, what does that mean?  \n* Mid Sec. 2 the ground truth appears and the loss is defined as sum of cross entropies $H(I_{i,j},I^*_{i,j})$. This loss function has a trivial minimum for $I=I^*$. I guess that some constraints are needed there. Also, where does the ground truth come from? Is this part of the input?\n* \"Since the losses are equivalent up to an additive constant, the gradients are exactly equal\" -> I don't think that's true, $\\nabla_I\\mathcal{L}(I,I^*)\\neq \\nabla_J\\mathcal{L}(J,J^*)$ already because $I$ and $J$ don't necessarily have the same dimensionality.\n\n## Section 3 - Scaling by Pruning non-essential gradients\n* Here, neural networks are mentioned for the first time, but the reader doesn't know here for what the NN is used and how it fits into the considered task.\n* What is model $f$ representing? How do time steps fit into the considered task? Where is $f$ in the objective?\n* What does it mean to supervise an iteration or step?\n* In Alg. 1, the updates in the for $t$ loops do not depend on $t$.\n\n## Section 4 - Scaling set-to-hypergraph prediction\n* What is $x_i$?\n* Introduce the abbrev. MLP, how is function $\\texttt{MLP}$ defined?\n* How is function $\\texttt{DeepSets}$ defined?\n* What is the existence indicator $\\sigma$? How can Eq. (6) be seen as factorizing the probability?\n\n## Section 5 - Experiments\n* How are parameters tuned, and what are the optimization parameter settings (step-size etc.). How are parameters set for competitors?\n* How is inference performed by the set-to-hypergraph models? \n* How is the sd computed in Table 1? Are multiple runs compared or multiple datasets/test splits?\n* How does the ML task to predict the convex hull differ from the one in computational geometry? Why is this task challenging for ML models but well understood in computational geometry?\n* Why does \"pruning the edges\" improve the predictive performance?\n* How is the number of iterations determined in the experiments?\n* How does the proposed optimization relate to TBPTT? How can TBPTT and backprop with skips improve the performance in comparison to backprop? When would it not help to apply this?\n* How does the run time of the proposed method relate to the one of competitors? \n\n## Section 6 - Related work\n* I would put this section earlier to introduce the competitors in the experiments.\n\n## Section 7 - Conclusion and Future Work\n* Nice, that you point out the limitations. But how does the input dimension exactly influence the feasibility of the proposed model?\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a set-to-hypergraph model where the hypergraph is represented by an incidence matrix in contrast to the usually used adjacency tensors. The incidence matrix is an edge times node matrix, which reflects only existing edges and is hence suitable to reduce the memory requirement when the amount of existing edges is sparse. The authors propose for the optimization of the proposed model\n an SGD scheme where a batch update is performed on the hypergraph which is returned by a recurrent neural network after a varying amount of time steps. \nThe experiments compare the proposed approach with 3 competitors on four datasets.  ", "main_review": "I find it hard to assess the contribution of this paper because I feel like I don't understand it. The English is good but there is a lack of definitions and explanations regarding the considered task, terminology, and mathematical notation, which makes it hard to make sense of the proposed method. In addition, the few mathematical definitions which are given do not seem to be complete. For example, the proposed loss function has a trivial solution and functions which are introduced later are not part of the loss (see comments below). Likewise, the experimental setting is not detailed. How the hyperparameters and the optimization parameters are tuned/set is not described for the proposed method and the competitors. As a result, I believe that the most actionable approach for the authors is that I list all the questions and issues I have in the following, to make clear where this paper needs improvement.\n\n# Questions & Issues:\n## Section 1 - Introduction:\n* From what or on the basis of what is the hypergraph learned from the set? What exactly is the input?\n* Based on what do you decide whether an edge exists or not? \n* The terminology of negative edges being also called non-existent should be introduced before this terminology is used in this section.\n## Section 2 - Scaling by pruning non-existing edges:\n* Why do non-existing edges have to be pruned, why are they modeled in the first place?\n* How is a hypergraph defined? How is a hyperedge defined?\n* What is $V\\subset \\mathcal{V}$?\n* What does it mean to \"remove excess rows\" from the incidence matrix? What is an excess row? I would expect that removing the row decreases the dimensionality of the incidence matrix. This is however not reflected in the mathematical notation.\n* What does it mean to \"supervise\" an incidence $I_{ji}$?\n* \"a positive edge in the incidence matrix contains both zeros and ones\" (-> what does that mean? Further:) \"ensuring that the binary classifier sees both positive and negative examples\"(->what is a pos/neg example?)\n* \"the order of the entries is fully decided by the order of the nodes\" -> why, what does that mean?  \n* Mid Sec. 2 the ground truth appears and the loss is defined as sum of cross entropies $H(I_{i,j},I^*_{i,j})$. This loss function has a trivial minimum for $I=I^*$. I guess that some constraints are needed there. Also, where does the ground truth come from? Is this part of the input?\n* \"Since the losses are equivalent up to an additive constant, the gradients are exactly equal\" -> I don't think that's true, $\\nabla_I\\mathcal{L}(I,I^*)\\neq \\nabla_J\\mathcal{L}(J,J^*)$ already because $I$ and $J$ don't necessarily have the same dimensionality.\n\n## Section 3 - Scaling by Pruning non-essential gradients\n* Here, neural networks are mentioned for the first time, but the reader doesn't know here for what the NN is used and how it fits into the considered task.\n* What is model $f$ representing? How do time steps fit into the considered task? Where is $f$ in the objective?\n* What does it mean to supervise an iteration or step?\n* In Alg. 1, the updates in the for $t$ loops do not depend on $t$.\n\n## Section 4 - Scaling set-to-hypergraph prediction\n* What is $x_i$?\n* Introduce the abbrev. MLP, how is function $\\texttt{MLP}$ defined?\n* How is function $\\texttt{DeepSets}$ defined?\n* What is the existence indicator $\\sigma$? How can Eq. (6) be seen as factorizing the probability?\n\n## Section 5 - Experiments\n* How are parameters tuned, and what are the optimization parameter settings (step-size etc.). How are parameters set for competitors?\n* How is inference performed by the set-to-hypergraph models? \n* How is the sd computed in Table 1? Are multiple runs compared or multiple datasets/test splits?\n* How does the ML task to predict the convex hull differ from the one in computational geometry? Why is this task challenging for ML models but well understood in computational geometry?\n* Why does \"pruning the edges\" improve the predictive performance?\n* How is the number of iterations determined in the experiments?\n* How does the proposed optimization relate to TBPTT? How can TBPTT and backprop with skips improve the performance in comparison to backprop? When would it not help to apply this?\n* How does the run time of the proposed method relate to the one of competitors? \n\n## Section 6 - Related work\n* I would put this section earlier to introduce the competitors in the experiments.\n\n## Section 7 - Conclusion and Future Work\n* Nice, that you point out the limitations. But how does the input dimension exactly influence the feasibility of the proposed model?\n\n\n", "summary_of_the_review": "Very unaccessible paper which might be understandable by experts in the field but not by a wider audience. Presentation needs to be improved, mathematical notation needs to be completed and experimental analysis needs to be clarified before I would consider this paper eligible for acceptance. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635788953023}, {"id": "P994aXgMgG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1834/Reviewer_am1Q"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed an efficient algorithm to tackle the set to hypergraph problem by utilizing recurrent training, pruning negative edges and backprop with skips. It then benchmarked the performance with prior works on the particle partitioning, convex hall and delaunay triangulation tasks. The proposed method showed a performance increase over previous methods.\n\n", "review_text": "Pro:\n1. Predicting the set to hypergraph is a very difficult task as the solution space is O(2^n). This paper provides many tricks to increase the performance and decrease the complexity in training the models. For example, backprop with skip, recurrent training instead of stacking, reducing the solution space to O(kn). \n2. This paper trains only on positive edges. They proved this loss is similar to the overall loss. This step could greatly increase the efficiency as reduced training space.\n3. They benchmarked on three different tasks and showed a performance increase. The ablation study also revealed the proposed tricks could help training.\n\nCon:\n1. The writing and the structure of the paper are a little confusing. Though the author took lots of effort in introducing the background. But I still fail to understand what is the overall f function to construct the hypergraph? And since only trained on positive edges, how did the author select the predicted hyperedges from the overall k edges?\n2. The proposed tricks are quite heuristic and do not provide theoretical insights for performance increase.\n3. The introduction of the Hungarian algorithm in selecting the edges does not make sense to me. Though the orderless nature of the edge in the hypergraph is important, the introduction of the algorithm greatly limited the generalization of the methods. And bring bias in the performance, since the Hungarian algorithm selects the number of edges and best corresponding edges to train the model. When ground truth is not available, I don't know how the proposed algorithm will choose the number of edges and best edges to construct the hypergraph, i.e., how the model generalizes in a broader case rather than fitting the ground truth data. And I suspect the Hungarian algorithm put an unfair advantage of the proposed model in comparing with other methods. I would suggest the author remove the Hungarian algorithm and only try to derive the incidence matrix I with the proposed model. \n4. Following comment 3, the author could introduce more detail about the input, output, prerequisites of the model. Specifically, the author could explain how they choose the number of edges, how to select predicted edges.\n\n\n\n\n\n     \n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed an efficient algorithm to tackle the set to hypergraph problem by utilizing recurrent training, pruning negative edges and backprop with skips. It then benchmarked the performance with prior works on the particle partitioning, convex hall and delaunay triangulation tasks. The proposed method showed a performance increase over previous methods.\n\n", "main_review": "Pro:\n1. Predicting the set to hypergraph is a very difficult task as the solution space is O(2^n). This paper provides many tricks to increase the performance and decrease the complexity in training the models. For example, backprop with skip, recurrent training instead of stacking, reducing the solution space to O(kn). \n2. This paper trains only on positive edges. They proved this loss is similar to the overall loss. This step could greatly increase the efficiency as reduced training space.\n3. They benchmarked on three different tasks and showed a performance increase. The ablation study also revealed the proposed tricks could help training.\n\nCon:\n1. The writing and the structure of the paper are a little confusing. Though the author took lots of effort in introducing the background. But I still fail to understand what is the overall f function to construct the hypergraph? And since only trained on positive edges, how did the author select the predicted hyperedges from the overall k edges?\n2. The proposed tricks are quite heuristic and do not provide theoretical insights for performance increase.\n3. The introduction of the Hungarian algorithm in selecting the edges does not make sense to me. Though the orderless nature of the edge in the hypergraph is important, the introduction of the algorithm greatly limited the generalization of the methods. And bring bias in the performance, since the Hungarian algorithm selects the number of edges and best corresponding edges to train the model. When ground truth is not available, I don't know how the proposed algorithm will choose the number of edges and best edges to construct the hypergraph, i.e., how the model generalizes in a broader case rather than fitting the ground truth data. And I suspect the Hungarian algorithm put an unfair advantage of the proposed model in comparing with other methods. I would suggest the author remove the Hungarian algorithm and only try to derive the incidence matrix I with the proposed model. \n4. Following comment 3, the author could introduce more detail about the input, output, prerequisites of the model. Specifically, the author could explain how they choose the number of edges, how to select predicted edges.\n\n\n\n\n\n     \n\n ", "summary_of_the_review": "This paper proposes a sufficient method to predict hypergraphs from sets and showed performance increase. Currently, it lacks sufficient detail, and may require further evaluation to justify the claims.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635545419061}], "openreview_url": "https://openreview.net/forum?id=7Z7u2z1Ornl", "arxiv_id": "2106.13919", "paper_pdf": "papers/7Z7u2z1Ornl.pdf", "paper_pdf_sha256": "7cb5550039f5258201fc6903989fd4ca391dba4a9ac50056e3e05d832f07af94", "paper_pdf_bytes": 497702, "paper_pdf_source": "openreview", "code_url": "https://github.com/davzha/recurrently_predicting_hypergraphs", "code_repository": "davzha/recurrently_predicting_hypergraphs", "code_commit": "c2b9a4959a1b9e3bac05ad3e5264029ccad2447d", "code_archive": "repos/7Z7u2z1Ornl.zip", "code_archive_sha256": "f924f1cdd07be8dfb94f21a4f4cb4c71a681edf06fd4a3f32775c1ff327bc15e", "code_archive_bytes": 25112, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 17, "github_languages": {"Python": 60006}, "github_archived": false, "github_pushed_at": "2021-09-21T09:02:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recurrently-predicting-hypergraphs"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yeeS_HULL7Z", "year": 2021, "status": "rejected", "title": "Attention-Based Clustering: Learning a Kernel from Context", "authors": ["Samuel Coward", "Erik Visse-Martindale", "Chithrupa Ramesh"], "authorids": ["~Samuel_Coward1", "erik.visse-martindale@uk.zuken.com", "~Chithrupa_Ramesh1"], "authors_source": "OpenReview API", "abstract": "In machine learning, no data point stands alone. We believe that context is an underappreciated concept in many machine learning methods. We propose Attention-Based Clustering (ABC), a neural architecture based on the attention mechanism, which is designed to learn latent representations that adapt to context within an input set, and which is inherently agnostic to input sizes and number of clusters. By learning a similarity kernel, our method directly combines with any out-of-the-box kernel-based clustering approach. We present competitive results for clustering Omniglot characters and include analytical evidence of the effectiveness of an attention-based approach for clustering. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SjLQ0S-7LK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper820/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": " Summary:\nThe hypothesis of this paper is that learning a contextual metric (allowing pairwise distances to depend on the data) can improves clustering, and is motivated by two examples (Omniglot, intersecting circles).  The paper proposes a new method - Attention based clustering (ABC) that incorporates context to learn a metric in the form of an embedding and kernel similarity layer (predefined). The embedding layer uses repeated self attention blocks (SABs) from the transformer architecture and is theoretically shown to make the clusters more condensed. An off-the-shelf clustering algorithm (e.g. spectral clustering) is used to cluster the similarity matrix (number of clusters pre-specified or inferred). The experiments show favorable results on the toy dataset and are competitive with methods that use a prespecified clustering.\n\nReasons for score:\nThe paper proposes a novel combination of approaches to improve clustering using a supervised learning method. Given that ABC’s results are surpassed for the Omniglot task by methods leveraging context for clustering, it would help to have a real world task where the approach taken (learning the similarity metric followed by off the shelf clustering) is the best.  This would help better motivate the proposed approach compared to previous approaches in the literature. It would also help to have further experiments on other datasets to increase the range of validity of the experiments.\n\nOther suggestions:\nIt could be interesting to visualize (say with T-SNE) what the geometry of the embeddings layer of the intersecting circles looks like. Further insight (theoretical or experimental) on the choice of kernel and/or other kernels would also be valuable.\n\nMinor:\nPerhaps displaying plots from the eigengap method of inference to validate that the inferred gap (number of clusters) is significant.\n\nClarity:\nSome details for the other methods compared to in the experiments would be helpful. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Recommendation to Reject", "review": " Summary:\nThe hypothesis of this paper is that learning a contextual metric (allowing pairwise distances to depend on the data) can improves clustering, and is motivated by two examples (Omniglot, intersecting circles).  The paper proposes a new method - Attention based clustering (ABC) that incorporates context to learn a metric in the form of an embedding and kernel similarity layer (predefined). The embedding layer uses repeated self attention blocks (SABs) from the transformer architecture and is theoretically shown to make the clusters more condensed. An off-the-shelf clustering algorithm (e.g. spectral clustering) is used to cluster the similarity matrix (number of clusters pre-specified or inferred). The experiments show favorable results on the toy dataset and are competitive with methods that use a prespecified clustering.\n\nReasons for score:\nThe paper proposes a novel combination of approaches to improve clustering using a supervised learning method. Given that ABC’s results are surpassed for the Omniglot task by methods leveraging context for clustering, it would help to have a real world task where the approach taken (learning the similarity metric followed by off the shelf clustering) is the best.  This would help better motivate the proposed approach compared to previous approaches in the literature. It would also help to have further experiments on other datasets to increase the range of validity of the experiments.\n\nOther suggestions:\nIt could be interesting to visualize (say with T-SNE) what the geometry of the embeddings layer of the intersecting circles looks like. Further insight (theoretical or experimental) on the choice of kernel and/or other kernels would also be valuable.\n\nMinor:\nPerhaps displaying plots from the eigengap method of inference to validate that the inferred gap (number of clusters) is significant.\n\nClarity:\nSome details for the other methods compared to in the experiments would be helpful. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604361826064}, {"id": "EYy4Lui4M11", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper820/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a metric learning and clustering method based on the idea of learning the metric from the context. They use the self-attention block module of the multi-head attention based transformer to embed the data and learn a kernel using the ground truth labels.  \nThey demonstrate their idea on a toy dataset and present results on the Omniglot dataset. \nAlthough their results look reasonable, I am concerned that the idea seems very incremental and simply uses a combination of techniques that have been proposed in the literature. It is unclear to me what the authors' primary contribution is to the field of metric learning \nThe following are my other main concerns with this paper: \n1. It is unclear if the proposed ABC method uses all the ground truth labels of the dataset to train the similarity kernel. If so, this would make their method highly impractical. Why would one cluster data using a method which requires it to have all the cluster labels in advance? In the Introduction the author say \" We use ground truth labels, only in the form of pairwise constraints, to train a similarity kernel, making our approach an example of constrained clustering.\" Constrained clustering techniques are semi-supervised and do not require all of the data set to be labeled.  The authors need to clarify and further discuss this point.\n2. In Section 5.1 the authors do not cite or mention what pairwise metric learning method was used. Was any hyperparameter tuning or other optimizations done for this method? It is very surprising that a clustering done after learning a metric performs worse than out-of-the-box spectral clustering.\n3. The authors demonstrate their results only on one real data set and there is no clear discussion of the results. They need to provide more details on why they chose the three tasks (variable number of clusters known and unknown and fixed number of clusters). If the training is independent of the task then why do they observe a slightly different NMI for the three tasks? Wouldnt the same kernel be learnt for all the tasks and thereby the same number of clusters chosen as specified in Section 3.4? There is a need for clarity in the description of the experimental setup and results. I would also recommend adding a few more data sets to the results. \n4. The description of the Ominiglot data set can be improved. For instance, they use the words alphabet and language interchangeably. The Omniglot data set only refers to alphabets and the authors should be consistent. \n5. Why are different metrics used for the two data sets (Rand Index and NMI)? \n\nIn summary, I believe this idea is incremental and possibly has some potential, but the authors have not done justice to it in this paper. There are a lot of unclear aspects that need to be better explained. \n\n\n\nReview update after rebuttal: The authors have addressed some of my concerns in the rebuttal and the modified version of the paper. They have added results on more real datasets and explained some aspects that were unclear in the first version. However, I am still not convinced that this is a comprehensive enough contribution as it is right now. The results on the new datasets have, in fact, raised more questions. Their proposed method seems to perform worse than the baselines in some scenarios. Though this is not bad and it is important to show negative results, I did not see any discussion or insights into the poor performance. I would recommend the authors run some baselines themselves and compare rather than copying the results from the baseline papers. This would ensure that the experimental setup and environment are similar leading to a fairer comparison, and possibly some insight into their methods' performance. In summary, I still believe that this is an idea with potential but the authors still have ways to go in putting their idea clearly on paper and justifying it completely and comprehensively. I am sticking to the score I had assigned before. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Attention based clustering", "review": "The authors propose a metric learning and clustering method based on the idea of learning the metric from the context. They use the self-attention block module of the multi-head attention based transformer to embed the data and learn a kernel using the ground truth labels.  \nThey demonstrate their idea on a toy dataset and present results on the Omniglot dataset. \nAlthough their results look reasonable, I am concerned that the idea seems very incremental and simply uses a combination of techniques that have been proposed in the literature. It is unclear to me what the authors' primary contribution is to the field of metric learning \nThe following are my other main concerns with this paper: \n1. It is unclear if the proposed ABC method uses all the ground truth labels of the dataset to train the similarity kernel. If so, this would make their method highly impractical. Why would one cluster data using a method which requires it to have all the cluster labels in advance? In the Introduction the author say \" We use ground truth labels, only in the form of pairwise constraints, to train a similarity kernel, making our approach an example of constrained clustering.\" Constrained clustering techniques are semi-supervised and do not require all of the data set to be labeled.  The authors need to clarify and further discuss this point.\n2. In Section 5.1 the authors do not cite or mention what pairwise metric learning method was used. Was any hyperparameter tuning or other optimizations done for this method? It is very surprising that a clustering done after learning a metric performs worse than out-of-the-box spectral clustering.\n3. The authors demonstrate their results only on one real data set and there is no clear discussion of the results. They need to provide more details on why they chose the three tasks (variable number of clusters known and unknown and fixed number of clusters). If the training is independent of the task then why do they observe a slightly different NMI for the three tasks? Wouldnt the same kernel be learnt for all the tasks and thereby the same number of clusters chosen as specified in Section 3.4? There is a need for clarity in the description of the experimental setup and results. I would also recommend adding a few more data sets to the results. \n4. The description of the Ominiglot data set can be improved. For instance, they use the words alphabet and language interchangeably. The Omniglot data set only refers to alphabets and the authors should be consistent. \n5. Why are different metrics used for the two data sets (Rand Index and NMI)? \n\nIn summary, I believe this idea is incremental and possibly has some potential, but the authors have not done justice to it in this paper. There are a lot of unclear aspects that need to be better explained. \n\n\n\nReview update after rebuttal: The authors have addressed some of my concerns in the rebuttal and the modified version of the paper. They have added results on more real datasets and explained some aspects that were unclear in the first version. However, I am still not convinced that this is a comprehensive enough contribution as it is right now. The results on the new datasets have, in fact, raised more questions. Their proposed method seems to perform worse than the baselines in some scenarios. Though this is not bad and it is important to show negative results, I did not see any discussion or insights into the poor performance. I would recommend the authors run some baselines themselves and compare rather than copying the results from the baseline papers. This would ensure that the experimental setup and environment are similar leading to a fairer comparison, and possibly some insight into their methods' performance. In summary, I still believe that this is an idea with potential but the authors still have ways to go in putting their idea clearly on paper and justifying it completely and comprehensively. I am sticking to the score I had assigned before. ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603999754290}, {"id": "N0gV_VaEmEY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper820/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes Attention-Based Clustering (ABC) that learns latent representations to adapt to context within an input set. They use ideas from the metric learning literature and the Siamese network on how to learn compatibility scores, and the transformer architecture and the Set Transformer on how to use context to make decisions.\n\nThe paper is well written and considers using the context in clustering which is an important problem however the proposed solution is not new. \n\n1- Transformer already being used as the Set Transformer to improve clustering (Lee et al., 2019b).  \n2- Also they mentioned that their model uses context to output pairwise similarities between the data points in the input set. They use the ground-truth labels and they train ABC in a supervised manner using those labels which is not realistic in clustering tasks!! \n3- The paper also claims that the ABC model is agnostic to the number of clusters. The known eigengap method (von Luxburg, 2007) is used to find the number of clusters.\n4- The paper only shows the result on OMNIGLOT and Olympic circles problems. Clustering is a very old problem and there are  known datasets from different domains to evaluate a new approach. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper considers using the context in clustering which is an important problem however the proposed solution is not new. ", "review": "This paper proposes Attention-Based Clustering (ABC) that learns latent representations to adapt to context within an input set. They use ideas from the metric learning literature and the Siamese network on how to learn compatibility scores, and the transformer architecture and the Set Transformer on how to use context to make decisions.\n\nThe paper is well written and considers using the context in clustering which is an important problem however the proposed solution is not new. \n\n1- Transformer already being used as the Set Transformer to improve clustering (Lee et al., 2019b).  \n2- Also they mentioned that their model uses context to output pairwise similarities between the data points in the input set. They use the ground-truth labels and they train ABC in a supervised manner using those labels which is not realistic in clustering tasks!! \n3- The paper also claims that the ABC model is agnostic to the number of clusters. The known eigengap method (von Luxburg, 2007) is used to find the number of clusters.\n4- The paper only shows the result on OMNIGLOT and Olympic circles problems. Clustering is a very old problem and there are  known datasets from different domains to evaluate a new approach. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603913257128}, {"id": "TB56MT9o_Du", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper820/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method for producing representations for clustering that take into account global trends in the dataset, rather than considering each pair of instances in isolation. They claim to achieve competitive clustering performance on omniglot, which they attribute to the use of these contextualised embeddings. They also present a theoretical justification for using transformers in metric learning\n\n## Strengths\n* The need to make use of context is well motivated, and the authors show that their method does make use of context effectively.\n* Both the algorithmic and theoretical work described by the paper appear to be technically correct.\n\n## Weaknesses\n* Important details pertaining to both the proposed architecture and experimental evaluation are missing (see questions below).\n* The novelty seems limited, given that Lee et al. (2019a, 2019b) have already proposed using a set transformer to compute contextualised embeddings for clustering.\n* While technically correct, it is unclear to me how the theoretical analysis directly relates to the SAB components used in the architecture. In particular, assuming all within cluster weights are equal to $\\alpha$ or $\\beta$, while all between cluster weights are equal to $\\gamma$ does not seem like a realistic model of an SAB module.\n* The experimental evaluation could be improved. Currently only one real-world dataset is used in the evaluation, which makes it hard to say whether the findings will generalise to other situations.\n* The way the experiments on omniglot are set up does not match the motivation for the method. In particular, one would expect context to be useful when elements from different clusters can sometimes look very similar (e.g., letters from different alphabets that resemble each other). However, the omniglot experiments consider each alphabet in isolation, so this impies that context will be useful when individual letters in the same alphabet can sometimes look similar.\n\n## Questions\n* How does $\\mathcal{T}$ map from $\\mathbb{R}^{n \\times d_x}$ to $\\mathbb{R}^{n \\times d_z}$ when the input and output of each SAB is the same dimensionality? Does this mean $d_x = d_z$?\n* In the circles experiment, are multiple datasets (i.e., \"instances\") synthesised in order to train the embedding block?\n* It is said that the \"pairwise\" baseline, the embedding block is removed, but also that it is a metric learning method. What metric learning is going on here?\n* When would one expect additive attention to outperform multiplicative attention, and vice versa?\n\n## Other comments\n* Siamese networks were proposed by Bromley et al. (1994), not by Koch et al. (2015).\n\n## After Author Response\nMy concerns are only somewhat addressed by the authors' response. While the results on extra datasets are encouraging, I still think there is limited novelty in the proposed approach and some of the important details remain unclear. In particular, terminology seems to be used inconsistently, which results in ambiguity when describing variants of the model used in experiments.\n\nJane Bromley, et al. Signature verification using a \"siamese\" time delay neural network. NeurIPS, 1994.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Missing details and needs improved evaluation", "review": "This paper proposes a method for producing representations for clustering that take into account global trends in the dataset, rather than considering each pair of instances in isolation. They claim to achieve competitive clustering performance on omniglot, which they attribute to the use of these contextualised embeddings. They also present a theoretical justification for using transformers in metric learning\n\n## Strengths\n* The need to make use of context is well motivated, and the authors show that their method does make use of context effectively.\n* Both the algorithmic and theoretical work described by the paper appear to be technically correct.\n\n## Weaknesses\n* Important details pertaining to both the proposed architecture and experimental evaluation are missing (see questions below).\n* The novelty seems limited, given that Lee et al. (2019a, 2019b) have already proposed using a set transformer to compute contextualised embeddings for clustering.\n* While technically correct, it is unclear to me how the theoretical analysis directly relates to the SAB components used in the architecture. In particular, assuming all within cluster weights are equal to $\\alpha$ or $\\beta$, while all between cluster weights are equal to $\\gamma$ does not seem like a realistic model of an SAB module.\n* The experimental evaluation could be improved. Currently only one real-world dataset is used in the evaluation, which makes it hard to say whether the findings will generalise to other situations.\n* The way the experiments on omniglot are set up does not match the motivation for the method. In particular, one would expect context to be useful when elements from different clusters can sometimes look very similar (e.g., letters from different alphabets that resemble each other). However, the omniglot experiments consider each alphabet in isolation, so this impies that context will be useful when individual letters in the same alphabet can sometimes look similar.\n\n## Questions\n* How does $\\mathcal{T}$ map from $\\mathbb{R}^{n \\times d_x}$ to $\\mathbb{R}^{n \\times d_z}$ when the input and output of each SAB is the same dimensionality? Does this mean $d_x = d_z$?\n* In the circles experiment, are multiple datasets (i.e., \"instances\") synthesised in order to train the embedding block?\n* It is said that the \"pairwise\" baseline, the embedding block is removed, but also that it is a metric learning method. What metric learning is going on here?\n* When would one expect additive attention to outperform multiplicative attention, and vice versa?\n\n## Other comments\n* Siamese networks were proposed by Bromley et al. (1994), not by Koch et al. (2015).\n\n## After Author Response\nMy concerns are only somewhat addressed by the authors' response. While the results on extra datasets are encouraging, I still think there is limited novelty in the proposed approach and some of the important details remain unclear. In particular, terminology seems to be used inconsistently, which results in ambiguity when describing variants of the model used in experiments.\n\nJane Bromley, et al. Signature verification using a \"siamese\" time delay neural network. NeurIPS, 1994.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603910717129}], "openreview_url": "https://openreview.net/forum?id=yeeS_HULL7Z", "arxiv_id": "2010.01040", "paper_pdf": "papers/yeeS_HULL7Z.pdf", "paper_pdf_sha256": "fa9d14e53527c1987340d84f5e913c8bf4ca4f66bf236df811c54929caf461b7", "paper_pdf_bytes": 396620, "paper_pdf_source": "openreview", "code_url": "https://github.com/DramaCow/ABC", "code_repository": "DramaCow/ABC", "code_commit": "20e084d2b4d9b3fb0456c9927cc96174c0fd1657", "code_archive": "repos/yeeS_HULL7Z.zip", "code_archive_sha256": "8bd8de7b6c6e52f4da379206ba570ff77d0ddd8316c02319b1424d7c7a184066", "code_archive_bytes": 23618, "code_file_count": 16, "code_extensions": {".py": 15, ".sh": 1}, "github_disk_usage_kb": 20, "github_languages": {"Python": 64158, "Shell": 194}, "github_archived": false, "github_pushed_at": "2024-09-13T11:36:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/attention-based-clustering-learning-a-kernel"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1lF8xHYwS", "year": 2020, "status": "rejected", "title": "Unsupervised Domain Adaptation through Self-Supervision", "authors": ["Yu Sun", "Eric Tzeng", "Trevor Darrell", "Alexei A. Efros"], "authorids": ["yusun@berkeley.edu", "etzeng@eecs.berkeley.edu", "trevor@eecs.berkeley.edu", "efros@eecs.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "This paper addresses unsupervised domain adaptation, the setting where labeled training data is available on a source domain, but the goal is to have good performance on a target domain with only unlabeled data. Like much of previous work, we seek to align the learned representations of the source and target domains while preserving discriminability. The way we accomplish alignment is by learning to perform auxiliary self-supervised task(s) on both domains simultaneously.  Each self-supervised task brings the two domains closer together along the direction relevant to that task. Training this jointly with the main task classifier on the source domain is shown to successfully generalize to the unlabeled target domain.  The presented objective is straightforward to implement and easy to optimize. We achieve state-of-the-art results on four out of seven standard benchmarks, and competitive results on segmentation adaptation. We also demonstrate that our method composes well with another popular pixel-level adaptation method.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rklFudcVqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2330/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes an approach to domain adaptation that uses\nself-supervised losses to encourage source/target domain alignment for\nunsupervised domain adaptation. The authors propose to use four\nself-supervised tasks (variants of tasks used in the self-supervised\nrepresentation learning for object recognition literature) that are\nused with a combined loss including unlabeled source and target\ntraining samples. The authors also propose an alignment heuristic for\nguiding early stopping. Experimental results on a standard battery of\ndomain adaptation problems are given, plus some intriguing baseline\nresults for semantic segmentation.\n\nThe paper is written very well and the technical development and\nmotivations for each decision are well discussed and argued.\n\n1. The experimental evaluation is a bit limited as the object\n   recognition datasets are a bit limited. Results on Office or\n   Office-Home would be nice.\n\n2. Using location classification for semantic segmentation seems\n   intuitively to be encouraging the network to learn coarse spatial\n   priors (which should be invariant across the two domains). Have you\n   looked at how alignment is actually happening? More qualitative\n   analysis in this direction would be useful to appreciate the\n   proposed approach.\n\n3. Related to the previous point, it would be interesting to see how\n   semgmentations in the unsupervised domain gradually change and\n   improve with increasing alignment.\n\nIn summary: the ideas are simple, intuitive, and well-explained -- I\nthink the results reported would be easy to reproduce with minimal\nhead scratching. The experiments are interesting and not overstated.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper describes an approach to domain adaptation that uses\nself-supervised losses to encourage source/target domain alignment for\nunsupervised domain adaptation. The authors propose to use four\nself-supervised tasks (variants of tasks used in the self-supervised\nrepresentation learning for object recognition literature) that are\nused with a combined loss including unlabeled source and target\ntraining samples. The authors also propose an alignment heuristic for\nguiding early stopping. Experimental results on a standard battery of\ndomain adaptation problems are given, plus some intriguing baseline\nresults for semantic segmentation.\n\nThe paper is written very well and the technical development and\nmotivations for each decision are well discussed and argued.\n\n1. The experimental evaluation is a bit limited as the object\n   recognition datasets are a bit limited. Results on Office or\n   Office-Home would be nice.\n\n2. Using location classification for semantic segmentation seems\n   intuitively to be encouraging the network to learn coarse spatial\n   priors (which should be invariant across the two domains). Have you\n   looked at how alignment is actually happening? More qualitative\n   analysis in this direction would be useful to appreciate the\n   proposed approach.\n\n3. Related to the previous point, it would be interesting to see how\n   semgmentations in the unsupervised domain gradually change and\n   improve with increasing alignment.\n\nIn summary: the ideas are simple, intuitive, and well-explained -- I\nthink the results reported would be easy to reproduce with minimal\nhead scratching. The experiments are interesting and not overstated.\n"}, "tcdate": 1572280432529}, {"id": "HkxybvFNqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2330/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces an unsupervised domain adaptation method that uses self-supervised tasks to bring the two different domains closer together. It runs experiments on some classic benchmarks.\n\nMy score for this paper is weakly rejected because \n\n(1) the concept of self-supervision is not first proposed by this paper. The proposed method is not novel. It introduces three simple self-supervision tasks: flip, rotation and location, and the performance is not better than previous results such as DIRT-T; \n\n(2) there are 7 benchmarks in Table2, but only 2 of 7 has result on R+L+F. In the paper, it mentioned because the result is not better, but the author should still provide them. \n\n(3) it emphasizes the contribution of encouraging more study of self-supervision for unsupervised domain adaptation. It doesn’t provide any way for how to design self-supervision task or whether more tasks is better. I think it is an interesting paper, but not enough as a conference paper, maybe a workshop paper. \n\n(4) there are some classic unsupervised domain adaption benchmarks like Office Dataset, and Bing-Caltech dataset, why not run the method on them?\n\n(5) In ICCV 2019, there is a paper \"S4L: Self-Supervised Semi-Supervised Learning\". The proposed method is almost same. I think the difference is this paper changes the setting and considers the unsupervised data as target domain and supervised data as source domain. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper introduces an unsupervised domain adaptation method that uses self-supervised tasks to bring the two different domains closer together. It runs experiments on some classic benchmarks.\n\nMy score for this paper is weakly rejected because \n\n(1) the concept of self-supervision is not first proposed by this paper. The proposed method is not novel. It introduces three simple self-supervision tasks: flip, rotation and location, and the performance is not better than previous results such as DIRT-T; \n\n(2) there are 7 benchmarks in Table2, but only 2 of 7 has result on R+L+F. In the paper, it mentioned because the result is not better, but the author should still provide them. \n\n(3) it emphasizes the contribution of encouraging more study of self-supervision for unsupervised domain adaptation. It doesn’t provide any way for how to design self-supervision task or whether more tasks is better. I think it is an interesting paper, but not enough as a conference paper, maybe a workshop paper. \n\n(4) there are some classic unsupervised domain adaption benchmarks like Office Dataset, and Bing-Caltech dataset, why not run the method on them?\n\n(5) In ICCV 2019, there is a paper \"S4L: Self-Supervised Semi-Supervised Learning\". The proposed method is almost same. I think the difference is this paper changes the setting and considers the unsupervised data as target domain and supervised data as source domain. "}, "tcdate": 1572275958862}, {"id": "rkxxcB1Rtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2330/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a novel unsupervised domain adaptation framework for neural networks. Similarly to existing approaches, it performs adaptation by aligning representations of the source and the target domains. The main difference is that this alignment is achieved not through explicitly minimizing some distribution discrepancy (this usually leads to challenging minimax optimization problems). Instead, the authors propose to use a battery of auxiliary self-supervised learning (SSL) tasks for both domains simultaneously. Each task is meant to align the source and the target representations along a direction of variation relevant to that task. Assuming that the battery is diverse enough, optimizing the representation for all the tasks leads to matching of the distributions. \n\nPros:\n+ The paper is well-written and easy to read.\n+ I like the simplicity of the idea and the fact that it achieves competitive performance without any adversarial learning (which may be very tricky to deal with).\n+ The paper presents a reasonable procedure for hyper-parameter tuning and early stopping which seems to work well in practice.\n\nCons:\n- The paper is purely practical with no theory backing the approach. As a result, the discussion of guarantees and limitations is quite brief.\n- It’s unclear how easy it is to come up with a reasonable set of SSL tasks for a particular pair of domains. It seems that it may become a serious problem when the method is applied to something other than benchmarks. Table 2 reveals that there is no consistent improvement over the existing approaches which suggests that the chosen battery of SSL tasks is not universal (as the authors themselves admit). On a related note, it’s a bit disappointing that the authors mention SVHN results as a failure case but never provide a way to address the issue.\n- It would be nice to some results for the Office dataset for completeness. The authors could use a pre-trained network as a starting points just like it’s done in other papers. According to the last paragraph of Section 6 this experiment should be feasible.\n\nNotes/questions:\n* Table 2, last column: The performance of DIRT-T seems to be better than that of the proposed method and yet the latter is highlighted and not the former.\n\nOverall, I think it’s a good paper presenting a thought-provoking idea. In my opinion, the weakest point of the work is the lack of any (neither principled nor practical) guidance as to how to choose the set of self-supervised tasks. Despite this I feel that this submission should be accepted but at the same time I’m curious to see what the authors have to say regarding the concerns I raised in my review.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper presents a novel unsupervised domain adaptation framework for neural networks. Similarly to existing approaches, it performs adaptation by aligning representations of the source and the target domains. The main difference is that this alignment is achieved not through explicitly minimizing some distribution discrepancy (this usually leads to challenging minimax optimization problems). Instead, the authors propose to use a battery of auxiliary self-supervised learning (SSL) tasks for both domains simultaneously. Each task is meant to align the source and the target representations along a direction of variation relevant to that task. Assuming that the battery is diverse enough, optimizing the representation for all the tasks leads to matching of the distributions. \n\nPros:\n+ The paper is well-written and easy to read.\n+ I like the simplicity of the idea and the fact that it achieves competitive performance without any adversarial learning (which may be very tricky to deal with).\n+ The paper presents a reasonable procedure for hyper-parameter tuning and early stopping which seems to work well in practice.\n\nCons:\n- The paper is purely practical with no theory backing the approach. As a result, the discussion of guarantees and limitations is quite brief.\n- It’s unclear how easy it is to come up with a reasonable set of SSL tasks for a particular pair of domains. It seems that it may become a serious problem when the method is applied to something other than benchmarks. Table 2 reveals that there is no consistent improvement over the existing approaches which suggests that the chosen battery of SSL tasks is not universal (as the authors themselves admit). On a related note, it’s a bit disappointing that the authors mention SVHN results as a failure case but never provide a way to address the issue.\n- It would be nice to some results for the Office dataset for completeness. The authors could use a pre-trained network as a starting points just like it’s done in other papers. According to the last paragraph of Section 6 this experiment should be feasible.\n\nNotes/questions:\n* Table 2, last column: The performance of DIRT-T seems to be better than that of the proposed method and yet the latter is highlighted and not the former.\n\nOverall, I think it’s a good paper presenting a thought-provoking idea. In my opinion, the weakest point of the work is the lack of any (neither principled nor practical) guidance as to how to choose the set of self-supervised tasks. Despite this I feel that this submission should be accepted but at the same time I’m curious to see what the authors have to say regarding the concerns I raised in my review."}, "tcdate": 1571841415847}], "openreview_url": "https://openreview.net/forum?id=S1lF8xHYwS", "arxiv_id": "1909.11825", "paper_pdf": "papers/S1lF8xHYwS.pdf", "paper_pdf_sha256": "1c40f9ea0702869dd2834daf98891e8a8ae20111d5eab9e25410d219ff355f19", "paper_pdf_bytes": 6346299, "paper_pdf_source": "openreview", "code_url": "https://github.com/yueatsprograms/uda_release", "code_repository": "yueatsprograms/uda_release", "code_commit": "b256316283e74b5d1f16777f029c384ee9b6e2e7", "code_archive": "repos/S1lF8xHYwS.zip", "code_archive_sha256": "116a62eb06d838bea684a05e1410d7bf7ad8ab1b8ca1b06b54ef3cadc4c12a33", "code_archive_bytes": 24041, "code_file_count": 29, "code_extensions": {".py": 22, ".sh": 7}, "github_disk_usage_kb": 21, "github_languages": {"Python": 43943, "Shell": 4095}, "github_archived": false, "github_pushed_at": "2021-10-06T17:41:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unsupervised-domain-adaptation-through-self-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1lnJ2Rqt7", "year": 2019, "status": "rejected", "title": "LARGE BATCH SIZE TRAINING OF NEURAL NETWORKS WITH ADVERSARIAL TRAINING AND SECOND-ORDER INFORMATION", "authors": ["Zhewei Yao", "Amir Gholami", "Kurt Keutzer", "Michael Mahoney"], "authorids": ["zheweiy@berkeley.edu", "amirgh@berkeley.edu", "keutzer@berkeley.edu", "mmahoney@stat.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Stochastic Gradient Descent (SGD) methods using randomly selected batches are widely-used to train neural network (NN) models. Performing design exploration to find the best NN for a particular task often requires extensive training with different models on a large dataset,  which is very computationally expensive. The most straightforward method to accelerate this computation is to distribute the batch of SGD over multiple processors. However, large batch training often times leads to degradation in accuracy, poor generalization, and even poor robustness to adversarial attacks.  Existing solutions for large batch training either do not work or require massive hyper-parameter tuning. To address this issue, we propose a novel large batch training method which combines recent results in adversarial training (to regularize against ``sharp minima'') and second order optimization (to use curvature information to change batch size adaptively during training). We extensively evaluate our method on Cifar-10/100, SVHN, TinyImageNet, and ImageNet datasets, using multiple NNs, including residual networks as well as compressed networks such as SqueezeNext.  Our new approach exceeds the performance of the existing solutions in terms of both accuracy and the number of SGD iterations (up to 1\\% and $3\\times$, respectively). We emphasize that this is achieved without any additional hyper-parameter tuning to tailor our method to any of these experiments.\n", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1xtFQylaQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1021/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose using information from the Hessian to grow the batch size as the training progresses. It is well-known that larger batch sizes can be used for later stages of optimization (ie, https://arxiv.org/abs/1711.00489, https://arxiv.org/abs/1706.05699), but they are missing motivation as to why use Hessian information for this.\n\nFurthermore, the description of the algorithm is lacking detail and is essentially unreproducible in current form.\n\nThe main description of their method is Algorithm 1 box, which suggests to grow batch size when \"eigenvalue\" is much smaller than previous eigenvalue. Is that the top eigenvalue? How is it estimated? Why is that the criterion? Note that for stochastic least squares problem one benefits from later batch sizes in later stages of optimization even though Hessian doesn't change.\n\nIn section Section 4.3 they start talking a bit about computing Hessian, referring to non-existent figure 6 for details of block approximation.\n\nAuthors mention that Hessian computation is not supported in major frameworks but don't provide explanation of how they compute it (did they not use a major framework for ImageNet experiments?).\n\nNote that a single row of Hessian (hence full Hessian) can be computed in all major frameworks by differentiating an element of the gradient. IE, in PyTorch https://gist.github.com/apaszke/226abdf867c4e9d6698bd198f3b45fb7, and also eigenspectrum of Hessian can be approximated -- https://github.com/noahgolmant/pytorch-hessian-eigenthings", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clearly an unfinished paper", "review": "The authors propose using information from the Hessian to grow the batch size as the training progresses. It is well-known that larger batch sizes can be used for later stages of optimization (ie, https://arxiv.org/abs/1711.00489, https://arxiv.org/abs/1706.05699), but they are missing motivation as to why use Hessian information for this.\n\nFurthermore, the description of the algorithm is lacking detail and is essentially unreproducible in current form.\n\nThe main description of their method is Algorithm 1 box, which suggests to grow batch size when \"eigenvalue\" is much smaller than previous eigenvalue. Is that the top eigenvalue? How is it estimated? Why is that the criterion? Note that for stochastic least squares problem one benefits from later batch sizes in later stages of optimization even though Hessian doesn't change.\n\nIn section Section 4.3 they start talking a bit about computing Hessian, referring to non-existent figure 6 for details of block approximation.\n\nAuthors mention that Hessian computation is not supported in major frameworks but don't provide explanation of how they compute it (did they not use a major framework for ImageNet experiments?).\n\nNote that a single row of Hessian (hence full Hessian) can be computed in all major frameworks by differentiating an element of the gradient. IE, in PyTorch https://gist.github.com/apaszke/226abdf867c4e9d6698bd198f3b45fb7, and also eigenspectrum of Hessian can be approximated -- https://github.com/noahgolmant/pytorch-hessian-eigenthings", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541563264729}, {"id": "Bkl9upa53m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1021/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the large batch size training of neural networks, and incorporates adversarial training and second-order information to improve the efficiency and effectiveness of the proposed algorithm. In particular, the authors use second-order information to automatically generate the step size and batch size in each iteration, and apply adversarial training as a regularization method to improve the test performance. Finally, the authors demonstrate their algorithm and compare it with the baseline algorithms on a wide range of datasets. This paper is clearly written and has the following strength:\n\n1.\tThis paper proposes an adaptive method for SGD training, which proves its convergence for strongly convex optimization.\n2.\tThis paper incorporates the adversarial training and robust optimization into the adaptive SGD training, and shows that this combination significantly improves the test performance.\n3.\tThe authors perform experiments on different datasets, which show that the proposed method enjoys less training time and higher accuracy when using large batch size.\n\nHowever, this paper also has the following weakness:\n\n1.\tThe theoretical analysis is somewhat trivial, and the assumption on the objective function is rather strong, which is not consistent with the nonconvex loss functions that are widely applied in training neural networks.\n2.\tTheorem 1 provides convergence rate of SGD on strongly convex objective functions. However, the authors do not carefully characterize the learning rate to ensure that the loss function achieve \\epsilon-accuracy. Moreover, in order to make the last term in (5) be smaller than \\epsilon, the learning rate \\eta_0 should be in the order of O(\\epsilon), which is no longer a tuning-free parameter.\n3.\tThe authors mention that the proposed algorithm converges faster than basic SGD, but it is not clearly demonstrated from Theorem 1.\n4.\tI am confused about how to determine the number of iterations for different algorithm as shown in Table 1s and 2? Do you stop each algorithm when they attain the same training error on the training dataset?\n5.\tIt is also confused that the number of iterations for ABS and ABSA are relatively larger than that of BL (Tables 1 and 2), but the training time of BL is longer than those of ABS and ABSA as reported in Table 3?\n6.\tSome minor flaws. In (9) it should be \\|\\nabla L(\\Theta)\\|^2; in Lemma 3, the expectation on the left side should be taken conditioned on \\theta_t.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting work, but the theoretical part is not strong enough", "review": "This paper studies the large batch size training of neural networks, and incorporates adversarial training and second-order information to improve the efficiency and effectiveness of the proposed algorithm. In particular, the authors use second-order information to automatically generate the step size and batch size in each iteration, and apply adversarial training as a regularization method to improve the test performance. Finally, the authors demonstrate their algorithm and compare it with the baseline algorithms on a wide range of datasets. This paper is clearly written and has the following strength:\n\n1.\tThis paper proposes an adaptive method for SGD training, which proves its convergence for strongly convex optimization.\n2.\tThis paper incorporates the adversarial training and robust optimization into the adaptive SGD training, and shows that this combination significantly improves the test performance.\n3.\tThe authors perform experiments on different datasets, which show that the proposed method enjoys less training time and higher accuracy when using large batch size.\n\nHowever, this paper also has the following weakness:\n\n1.\tThe theoretical analysis is somewhat trivial, and the assumption on the objective function is rather strong, which is not consistent with the nonconvex loss functions that are widely applied in training neural networks.\n2.\tTheorem 1 provides convergence rate of SGD on strongly convex objective functions. However, the authors do not carefully characterize the learning rate to ensure that the loss function achieve \\epsilon-accuracy. Moreover, in order to make the last term in (5) be smaller than \\epsilon, the learning rate \\eta_0 should be in the order of O(\\epsilon), which is no longer a tuning-free parameter.\n3.\tThe authors mention that the proposed algorithm converges faster than basic SGD, but it is not clearly demonstrated from Theorem 1.\n4.\tI am confused about how to determine the number of iterations for different algorithm as shown in Table 1s and 2? Do you stop each algorithm when they attain the same training error on the training dataset?\n5.\tIt is also confused that the number of iterations for ABS and ABSA are relatively larger than that of BL (Tables 1 and 2), but the training time of BL is longer than those of ABS and ABSA as reported in Table 3?\n6.\tSome minor flaws. In (9) it should be \\|\\nabla L(\\Theta)\\|^2; in Lemma 3, the expectation on the left side should be taken conditioned on \\theta_t.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541229938160}, {"id": "SkxjN0yHsQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1021/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Based on my understanding, this paper describes a novel approach for addressing the large batch training problem. The authors propose increasing the batch size based on reductions in the largest eigenvalue of the Hessian. This is combined with adversarial training using the fast gradient sign method to reduce the total number of iterations required for training and improve generalization performance. Unfortunately, although the numerical results seem quite promising, the algorithm and its explanation and details are not described clearly in the paper, which makes me lean towards rejection. I describe this more fully below:\n\n1. Description of the Algorithm\n\nThe description of the algorithm in Section 3.1 is simply not clear, and lacks clear exposition motivating why the algorithm ought to work. To add to this confusion, there appear to be some inconsistencies between the (brief) description of the method and the description given in the Main Contributions and Limitations section in the Introduction. \n\nAs an example, in Section 3.1, the approach for computing the eigenvalue of the Hessian is not described. Which eigenvalue is computed? How is this done? What is the batch size used in this computation? Is it computed over the full training set? The Limitations section briefly describes this (power iteration to tolerance <= 10^-2), but this should be elaborated on in Section 3.1. In fact, the limitations should not be discussed until a clear description of the algorithm is given.\n\nThe introduction makes this even more confusing by claiming the second order information is computed by “backpropagating the Hessian operator”. This seems to imply that the 3rd derivative information is computed for second-order information. Later in the Introduction, the authors claim to use Hessian matvecs to perform the power iteration. I believe that the authors mean that the Hessian-vector product is obtained by differentiating the product g’v (a scalar quantity). \n\nIn addition, it was not described how the learning rate is changed in the algorithm. Later in the experiments, none of the additional hyperparameters in the procedure are given, such as the duration factor, kappa, the hyperparameters in the adversarial training, and more. This all ought to be included for completeness.\n\n2. Questions about Details of the Algorithm\n\nIf it is indeed the case that the authors are using power iteration to compute the largest eigenvalue, why not use Lanczos method as it typically works better for symmetric matrices? In addition, if the intention was to compute the largest eigenvalue of the Hessian, one must be wary that the power iteration/Lanczos method computes the eigenvalue with largest magnitude (the absolute value of lambda), which may mean that it’s possible that the algorithm is utilizing negative curvature information rather than positive curvature information (particularly in the earlier epochs), which may contradict their intuition based on flat minima. This needs to be addressed.\n\nSecondly, there is no explanation as to why increasing the batch size would lead to consistent decrease in the eigenvalues of the Hessian. This is certainly not true for all optimization problems. Even if the flat minima/sharp minima hypothesis is assumed, is it possible for the iterates after increasing the batch size to still tend towards sharper minimizers after being in a flat region? This intuition and explanation needs to be expanded on (and argued for) in order for the algorithm to make any conceptual sense.\n\nLastly, why is the duration factor needed to increase the batch size if the eigenvalue condition fails? if the duration factor is removed, how does the batch size evolve? Is it necessary? How is the duration factor tuned?\n\n3. Inconsequential Theoretical Results\n\nThe authors also prove a theorem bounding the expected optimality gap with adaptive batch sizes. On closer look, this is a simple adaptation of the result by Bottou, Curtis, and Nocedal [2] and does not utilize any of the algorithmic mechanisms described in the paper. Hence, the theoretical result is not novel, does not provide any additional insight on the algorithm, and could be applied to any adaptive/changing batch size SG algorithm. In my opinion, this ought to be removed. (Assumption 2 is also mentioned in the main paper, but is only described in the Appendix.)\n\n4. Additional Considerations\n\nThe paper is missing much work done by Nocedal’s group on increasing batch sizes (some of which utilize the L-BFGS approximation to the Hessian); see [1, 3].\n\nOther relevant work by Sagun, Bengio, and others on large batch training, flat minima, and the Hessian in deep learning ought to be included as well; see [4-7]. \n\nLastly, the algorithm demonstrates some significant improvements on the number of iterations. However, efficiency with respect to epochs is not discussed. It may make sense to plot test loss/error against epochs and batch size against iterations for clarity.\n\nTypos/Grammatical Errors:\n- Page 2: Should not state “(We refer to this method as ABS)”, easier to include by including (ABS) after Adaptive Batch Size in the beginning of the bullet point.\n- Page 6: Section 4: “information” not “informatino”\n- Page 6: Section 4: “the” not “teh”\n- Page 7: Section 4.1: “confirms” not “confirming”\n- Page 7: Section 4.1: no “a” in “a very consistent performance”\n\nSummary:\n\nOverall, although the paper presents some promising numerical results, it lacks a detailed description and explanation of the algorithm to be worthy of publication. It leaves many aspects of the algorithm open to the reader’s interpretation, and I do not believe I could reproduce the results with the information provided. The manuscript needs significant changes to the detail, structure, and writing before it can be considered for publication.\n\nReferences:\n[1] Bollapragada, Raghu, et al. \"A progressive batching L-BFGS method for machine learning.\" arXiv preprint arXiv:1802.05374(2018).\n[2] Bottou, Léon, Frank E. Curtis, and Jorge Nocedal. \"Optimization methods for large-scale machine learning.\" SIAM Review 60.2 (2018): 223-311.\n[3] Byrd, Richard H., et al. \"Sample size selection in optimization methods for machine learning.\" Mathematical programming134.1 (2012): 127-155.\n[4] Chaudhari, Pratik, et al. \"Entropy-sgd: Biasing gradient descent into wide valleys.\" arXiv preprint arXiv:1611.01838(2016).\n[5] Jastrzębski, Stanisław, et al. \"DNN's Sharpest Directions Along the SGD Trajectory.\" arXiv preprint arXiv:1807.05031(2018).\n[6] Sagun, Levent, et al. \"Empirical Analysis of the Hessian of Over-Parametrized Neural Networks.\" arXiv preprint arXiv:1706.04454 (2017).\n[7] Zhu, Zhanxing, et al. \"The Regularization Effects of Anisotropic Noise in Stochastic Gradient Descent.\" arXiv preprint arXiv:1803.00195 (2018).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising numerical results, but lacks clear description and explanation of the algorithm", "review": "Based on my understanding, this paper describes a novel approach for addressing the large batch training problem. The authors propose increasing the batch size based on reductions in the largest eigenvalue of the Hessian. This is combined with adversarial training using the fast gradient sign method to reduce the total number of iterations required for training and improve generalization performance. Unfortunately, although the numerical results seem quite promising, the algorithm and its explanation and details are not described clearly in the paper, which makes me lean towards rejection. I describe this more fully below:\n\n1. Description of the Algorithm\n\nThe description of the algorithm in Section 3.1 is simply not clear, and lacks clear exposition motivating why the algorithm ought to work. To add to this confusion, there appear to be some inconsistencies between the (brief) description of the method and the description given in the Main Contributions and Limitations section in the Introduction. \n\nAs an example, in Section 3.1, the approach for computing the eigenvalue of the Hessian is not described. Which eigenvalue is computed? How is this done? What is the batch size used in this computation? Is it computed over the full training set? The Limitations section briefly describes this (power iteration to tolerance <= 10^-2), but this should be elaborated on in Section 3.1. In fact, the limitations should not be discussed until a clear description of the algorithm is given.\n\nThe introduction makes this even more confusing by claiming the second order information is computed by “backpropagating the Hessian operator”. This seems to imply that the 3rd derivative information is computed for second-order information. Later in the Introduction, the authors claim to use Hessian matvecs to perform the power iteration. I believe that the authors mean that the Hessian-vector product is obtained by differentiating the product g’v (a scalar quantity). \n\nIn addition, it was not described how the learning rate is changed in the algorithm. Later in the experiments, none of the additional hyperparameters in the procedure are given, such as the duration factor, kappa, the hyperparameters in the adversarial training, and more. This all ought to be included for completeness.\n\n2. Questions about Details of the Algorithm\n\nIf it is indeed the case that the authors are using power iteration to compute the largest eigenvalue, why not use Lanczos method as it typically works better for symmetric matrices? In addition, if the intention was to compute the largest eigenvalue of the Hessian, one must be wary that the power iteration/Lanczos method computes the eigenvalue with largest magnitude (the absolute value of lambda), which may mean that it’s possible that the algorithm is utilizing negative curvature information rather than positive curvature information (particularly in the earlier epochs), which may contradict their intuition based on flat minima. This needs to be addressed.\n\nSecondly, there is no explanation as to why increasing the batch size would lead to consistent decrease in the eigenvalues of the Hessian. This is certainly not true for all optimization problems. Even if the flat minima/sharp minima hypothesis is assumed, is it possible for the iterates after increasing the batch size to still tend towards sharper minimizers after being in a flat region? This intuition and explanation needs to be expanded on (and argued for) in order for the algorithm to make any conceptual sense.\n\nLastly, why is the duration factor needed to increase the batch size if the eigenvalue condition fails? if the duration factor is removed, how does the batch size evolve? Is it necessary? How is the duration factor tuned?\n\n3. Inconsequential Theoretical Results\n\nThe authors also prove a theorem bounding the expected optimality gap with adaptive batch sizes. On closer look, this is a simple adaptation of the result by Bottou, Curtis, and Nocedal [2] and does not utilize any of the algorithmic mechanisms described in the paper. Hence, the theoretical result is not novel, does not provide any additional insight on the algorithm, and could be applied to any adaptive/changing batch size SG algorithm. In my opinion, this ought to be removed. (Assumption 2 is also mentioned in the main paper, but is only described in the Appendix.)\n\n4. Additional Considerations\n\nThe paper is missing much work done by Nocedal’s group on increasing batch sizes (some of which utilize the L-BFGS approximation to the Hessian); see [1, 3].\n\nOther relevant work by Sagun, Bengio, and others on large batch training, flat minima, and the Hessian in deep learning ought to be included as well; see [4-7]. \n\nLastly, the algorithm demonstrates some significant improvements on the number of iterations. However, efficiency with respect to epochs is not discussed. It may make sense to plot test loss/error against epochs and batch size against iterations for clarity.\n\nTypos/Grammatical Errors:\n- Page 2: Should not state “(We refer to this method as ABS)”, easier to include by including (ABS) after Adaptive Batch Size in the beginning of the bullet point.\n- Page 6: Section 4: “information” not “informatino”\n- Page 6: Section 4: “the” not “teh”\n- Page 7: Section 4.1: “confirms” not “confirming”\n- Page 7: Section 4.1: no “a” in “a very consistent performance”\n\nSummary:\n\nOverall, although the paper presents some promising numerical results, it lacks a detailed description and explanation of the algorithm to be worthy of publication. It leaves many aspects of the algorithm open to the reader’s interpretation, and I do not believe I could reproduce the results with the information provided. The manuscript needs significant changes to the detail, structure, and writing before it can be considered for publication.\n\nReferences:\n[1] Bollapragada, Raghu, et al. \"A progressive batching L-BFGS method for machine learning.\" arXiv preprint arXiv:1802.05374(2018).\n[2] Bottou, Léon, Frank E. Curtis, and Jorge Nocedal. \"Optimization methods for large-scale machine learning.\" SIAM Review 60.2 (2018): 223-311.\n[3] Byrd, Richard H., et al. \"Sample size selection in optimization methods for machine learning.\" Mathematical programming134.1 (2012): 127-155.\n[4] Chaudhari, Pratik, et al. \"Entropy-sgd: Biasing gradient descent into wide valleys.\" arXiv preprint arXiv:1611.01838(2016).\n[5] Jastrzębski, Stanisław, et al. \"DNN's Sharpest Directions Along the SGD Trajectory.\" arXiv preprint arXiv:1807.05031(2018).\n[6] Sagun, Levent, et al. \"Empirical Analysis of the Hessian of Over-Parametrized Neural Networks.\" arXiv preprint arXiv:1706.04454 (2017).\n[7] Zhu, Zhanxing, et al. \"The Regularization Effects of Anisotropic Noise in Stochastic Gradient Descent.\" arXiv preprint arXiv:1803.00195 (2018).", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1539796530840}], "openreview_url": "https://openreview.net/forum?id=H1lnJ2Rqt7", "arxiv_id": "1810.01021", "paper_pdf": "papers/H1lnJ2Rqt7.pdf", "paper_pdf_sha256": "eeb0d26bc21ceac77f642114a6564744a91686ec797a5a1131febc95691a8143", "paper_pdf_bytes": 721965, "paper_pdf_source": "openreview", "code_url": "https://github.com/amirgholami/HessianFlow", "code_repository": "amirgholami/HessianFlow", "code_commit": "644fc8a57472f3895fc21ba68357e46ad723beec", "code_archive": "repos/H1lnJ2Rqt7.zip", "code_archive_sha256": "74e706620eb7f530638b28ad68b540f83bfa1bd8d8d434b824782a1901960504", "code_archive_bytes": 32302, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 35, "github_languages": {"Python": 44843}, "github_archived": false, "github_pushed_at": "2020-01-15T06:36:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/large-batch-size-training-of-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJ7yZ2P6-", "year": 2018, "status": "rejected", "title": "Enhance Word Representation for Out-of-Vocabulary on Ubuntu Dialogue Corpus", "authors": ["JIANXIONG DONG", "Jim Huang"], "authorids": ["jdongca2003@gmail.com", "ccjimhuang@gmail.com"], "authors_source": "OpenReview API", "abstract": "Ubuntu dialogue corpus is the largest public available dialogue corpus to make it feasible to build end-to-end\ndeep neural network models directly from the conversation data. One challenge of Ubuntu dialogue corpus is \nthe large number of out-of-vocabulary words. In this paper we proposed an algorithm which combines the general pre-trained word embedding vectors with those  generated on the task-specific training set to address this issue.  We integrated character embedding into Chen et al's Enhanced LSTM method (ESIM) and used it to evaluate the effectiveness of our proposed method. For the task of next utterance selection, the proposed method has demonstrated a significant performance improvement against original ESIM and the new model has achieved state-of-the-art results on both Ubuntu dialogue corpus and Douban conversation corpus. In addition, we investigated the performance impact of end-of-utterance and end-of-turn token tags. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BkomChuxf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper27/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers a setting (Ubuntu Dialogue Corpus and Douban Conversation Corpus) where most word types in the data are not covered by pretrained representations. The proposed solution is to combine (1) external pretrained word embeddings and (2) pretrained word embeddings on the training data by keeping them as two views: use the view if it's available, otherwise use a zero vector. This scheme is shown to perform well compared to other methods, specifically combinations of pretraining vs not pretraining embeddings on the training data, updating vs not updating embeddings during training, and others. \n\nQuality: Low. The research is not very well modularized: the addressed problem has nothing specifically to do with ESIM and dialogue response classification, but it's all tangled up. The proposed solution is reasonable but rather minor. Given that the model will learn task-specific word representations on the training set anyway, it's not clear how important it is to follow this procedure, though minor improvement is reported (Table 5). \n\nClarity: The writing is clear. But the point of the paper is not immediately obvious because of its failure to modularize its contributions (see above).\n\nOriginality: Low to minor.\n\nSignificance: It's not convincing that an incremental improvement in the pretraining phase is so significant, for instance compared to developing a novel better architecture actually tailored to the dialogue task. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A minor solution to resolving OOV word representations", "rating": "3: Clear rejection", "review": "The paper considers a setting (Ubuntu Dialogue Corpus and Douban Conversation Corpus) where most word types in the data are not covered by pretrained representations. The proposed solution is to combine (1) external pretrained word embeddings and (2) pretrained word embeddings on the training data by keeping them as two views: use the view if it's available, otherwise use a zero vector. This scheme is shown to perform well compared to other methods, specifically combinations of pretraining vs not pretraining embeddings on the training data, updating vs not updating embeddings during training, and others. \n\nQuality: Low. The research is not very well modularized: the addressed problem has nothing specifically to do with ESIM and dialogue response classification, but it's all tangled up. The proposed solution is reasonable but rather minor. Given that the model will learn task-specific word representations on the training set anyway, it's not clear how important it is to follow this procedure, though minor improvement is reported (Table 5). \n\nClarity: The writing is clear. But the point of the paper is not immediately obvious because of its failure to modularize its contributions (see above).\n\nOriginality: Low to minor.\n\nSignificance: It's not convincing that an incremental improvement in the pretraining phase is so significant, for instance compared to developing a novel better architecture actually tailored to the dialogue task. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511734819370}, {"id": "rJdjmmLez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper27/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper proposes an approach to improve the out-of-vocabulary embedding prediction for the task of modeling dialogue conversations. The proposed approach uses generic embeddings and combines them with the embeddings trained on the training dataset in a straightforward string-matching algorithm. In addition, the paper also makes a couple of improvements to Chen et. al's enhanced LSTM by adding character-level embeddings and replacing average pooling by LSTM last state summary vector. The results are shown on the standard Ubuntu dialogue dataset as well as a new Douban conversation dataset. The proposed approach gives sizable gains over the baselines.\n\n\nComments:\n\nThe paper is well written and puts itself nicely in context of previous work. Though, the proposed extension to handle out-of-vocabulary items is a simple and straightforward string matching algorithm, but nonetheless it gives noticeable increase in empirical performance on both the tasks. All in all, the methodological novelty of the paper is small but it has high practical relevance in terms of giving improved accuracy on an important task of dialogue conversation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good paper with important practical and engineering relevance. Little methodological novelty, though.", "rating": "6: Marginally above acceptance threshold", "review": "Summary:\nThis paper proposes an approach to improve the out-of-vocabulary embedding prediction for the task of modeling dialogue conversations. The proposed approach uses generic embeddings and combines them with the embeddings trained on the training dataset in a straightforward string-matching algorithm. In addition, the paper also makes a couple of improvements to Chen et. al's enhanced LSTM by adding character-level embeddings and replacing average pooling by LSTM last state summary vector. The results are shown on the standard Ubuntu dialogue dataset as well as a new Douban conversation dataset. The proposed approach gives sizable gains over the baselines.\n\n\nComments:\n\nThe paper is well written and puts itself nicely in context of previous work. Though, the proposed extension to handle out-of-vocabulary items is a simple and straightforward string matching algorithm, but nonetheless it gives noticeable increase in empirical performance on both the tasks. All in all, the methodological novelty of the paper is small but it has high practical relevance in terms of giving improved accuracy on an important task of dialogue conversation.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511564192073}, {"id": "H1RMVeqgz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper27/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The main contributions in this paper are:\n1) New variants of a recent LSTM-based model (\"ESIM\") are applied to the task of response-selection in dialogue modeling -- ESIM was originally introduced and evaluated for natural language inference. In this new setting, the ESIM model (vanilla and extended) outperform previous models when trained and evaluated on two distinct conversational datasets.\n\n2) A fairly trivial method is proposed to extend the coverage of pre-trained word embeddings to deal with the OOV problem that arises when applying them to these conversational datasets.\nThe method itself is to combine d1-dimensional word embeddings that were pretrained on a large unannotated corpus (vocabulary S) with distinct d2-dimensional word embeddings that are trained on the task-specific training data (vocabulary T). The enhanced (d1+d2)-dimensional representation for a word is constructed by concatenating its vectors from the two embeddings, setting either the d1- or d2-dimensional subvector to zeros when the word is absent from either S or T, respectively. This method is incorporated as an extension into ESIM and evaluated on the two conversation datasets.\n\nThe main results can be characterized as showing that this vocabulary extension method leads to performance gains on two datasets, on top of an ESIM-model extended with character-based word embeddings, which itself outperforms the vanilla ESIM model.\n\nThese empirical results are potentially meaningful and could justify reporting, but the paper's organization is very confusing, and too many details are too unclear, leading to low confidence in reproducibility. \n\nThere is basic novelty in applying the base model to a new task, and the analysis of the role of the special conversational boundary tokens is interesting and can help to inform future modeling choices. The embedding-enhancing method has low originality but is effective on this particular combination of model architecture, task and datasets. I am left wondering how well it might generalize to other models or tasks, since the problem it addresses shows up in many other places too...\n\nOverall, the presentation switches back and forth between the Douban corpus and the Ubuntu corpus, and between word2vec and Glove embeddings, and this makes it very challenging to understand the details fully.\n\nS3.1 - Word representation layer: This paragraph should probably mention that the character-composed embeddings are newly introduced here, and were not part of the original formulation of ESIM. That statement is currently hidden in the figure caption.\n\nAlgorithm 1:\n- What set does P denote, and what is the set-theoretic relation between P and T?\n- Under one possible interpretation, there may be items in P that are in neither T nor S, yet the algorithm does not define embeddings for those items even though its output is described as \"a dictionary with word embeddings ... for P\". This does not seem consistent? I think the sentence in S4.2 about initializing remaining OOV words as zeros is relevant and wonder if it should form part of the algorithm description?\n\nS4.1 - What do the authors mean by the statement that response candidates for the Douban corpus were \"collected by Lucene retrieval model\"?\n\nS4.2 - Paragraph two is very unclear. In particular, I don't understand the role of the Glove vectors here when Algorithm 1 is used, since the authors refer to word2vec vectors later in this paragraph and also in the Algorithm description.\n\nS4.3 - It's insufficiently clear what the model definitions are for the Douban corpus. Is there still a character-based LSTM involved, or does FastText make it unnecessary?\n\nS4.3 - \"It can be seen from table 3 that the original ESIM did not perform well without character embedding.\" This is a curious way to describe the result, when, in fact, the ESIM model in table 3 already outperforms all the previous models listed.\n\nS4.4 - gensim package -- for the benefit of readers unfamiliar with gensim, the text should ideally state explicitly that it is used to create the *word2vec* embeddings, instead of the ambiguous \"word embeddings\".\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising results but insufficient clarity and focus in write-up", "rating": "5: Marginally below acceptance threshold", "review": "The main contributions in this paper are:\n1) New variants of a recent LSTM-based model (\"ESIM\") are applied to the task of response-selection in dialogue modeling -- ESIM was originally introduced and evaluated for natural language inference. In this new setting, the ESIM model (vanilla and extended) outperform previous models when trained and evaluated on two distinct conversational datasets.\n\n2) A fairly trivial method is proposed to extend the coverage of pre-trained word embeddings to deal with the OOV problem that arises when applying them to these conversational datasets.\nThe method itself is to combine d1-dimensional word embeddings that were pretrained on a large unannotated corpus (vocabulary S) with distinct d2-dimensional word embeddings that are trained on the task-specific training data (vocabulary T). The enhanced (d1+d2)-dimensional representation for a word is constructed by concatenating its vectors from the two embeddings, setting either the d1- or d2-dimensional subvector to zeros when the word is absent from either S or T, respectively. This method is incorporated as an extension into ESIM and evaluated on the two conversation datasets.\n\nThe main results can be characterized as showing that this vocabulary extension method leads to performance gains on two datasets, on top of an ESIM-model extended with character-based word embeddings, which itself outperforms the vanilla ESIM model.\n\nThese empirical results are potentially meaningful and could justify reporting, but the paper's organization is very confusing, and too many details are too unclear, leading to low confidence in reproducibility. \n\nThere is basic novelty in applying the base model to a new task, and the analysis of the role of the special conversational boundary tokens is interesting and can help to inform future modeling choices. The embedding-enhancing method has low originality but is effective on this particular combination of model architecture, task and datasets. I am left wondering how well it might generalize to other models or tasks, since the problem it addresses shows up in many other places too...\n\nOverall, the presentation switches back and forth between the Douban corpus and the Ubuntu corpus, and between word2vec and Glove embeddings, and this makes it very challenging to understand the details fully.\n\nS3.1 - Word representation layer: This paragraph should probably mention that the character-composed embeddings are newly introduced here, and were not part of the original formulation of ESIM. That statement is currently hidden in the figure caption.\n\nAlgorithm 1:\n- What set does P denote, and what is the set-theoretic relation between P and T?\n- Under one possible interpretation, there may be items in P that are in neither T nor S, yet the algorithm does not define embeddings for those items even though its output is described as \"a dictionary with word embeddings ... for P\". This does not seem consistent? I think the sentence in S4.2 about initializing remaining OOV words as zeros is relevant and wonder if it should form part of the algorithm description?\n\nS4.1 - What do the authors mean by the statement that response candidates for the Douban corpus were \"collected by Lucene retrieval model\"?\n\nS4.2 - Paragraph two is very unclear. In particular, I don't understand the role of the Glove vectors here when Algorithm 1 is used, since the authors refer to word2vec vectors later in this paragraph and also in the Algorithm description.\n\nS4.3 - It's insufficiently clear what the model definitions are for the Douban corpus. Is there still a character-based LSTM involved, or does FastText make it unnecessary?\n\nS4.3 - \"It can be seen from table 3 that the original ESIM did not perform well without character embedding.\" This is a curious way to describe the result, when, in fact, the ESIM model in table 3 already outperforms all the previous models listed.\n\nS4.4 - gensim package -- for the benefit of readers unfamiliar with gensim, the text should ideally state explicitly that it is used to create the *word2vec* embeddings, instead of the ambiguous \"word embeddings\".\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511814166099}], "openreview_url": "https://openreview.net/forum?id=rJ7yZ2P6-", "arxiv_id": "1802.02614", "paper_pdf": "papers/rJ7yZ2P6-.pdf", "paper_pdf_sha256": "c5060bc86dca3475ba979043a06f7dd348e6c5ce568df078379ebbd8a515259f", "paper_pdf_bytes": 327399, "paper_pdf_source": "openreview", "code_url": "https://github.com/jdongca2003/next_utterance_selection", "code_repository": "jdongca2003/next_utterance_selection", "code_commit": "7491e972d58412c175166f6564d77b7436e71a87", "code_archive": "repos/rJ7yZ2P6-.zip", "code_archive_sha256": "773b6feb8049ca8dc1e0ce3d2fb4bc77dd3a4dd2142e680369b67f9d14437d9f", "code_archive_bytes": 23916, "code_file_count": 10, "code_extensions": {".py": 7, ".sh": 3}, "github_disk_usage_kb": 24, "github_languages": {"Python": 46161, "Shell": 2365}, "github_archived": false, "github_pushed_at": "2018-05-07T05:07:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/enhance-word-representation-for-out-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bDee2EgvWJ", "year": 2026, "status": "rejected", "title": "Approximate Message Passing for Bayesian Neural Networks", "authors": ["Romeo Sommerfeld", "Christian Helms", "Jan Niklas Groeneveld", "Rainer Schlosser", "Ralf Herbrich"], "authorids": ["~Romeo_Sommerfeld1", "~Christian_Helms1", "~Jan_Niklas_Groeneveld1", "~Rainer_Schlosser1", "~Ralf_Herbrich1"], "authors_source": "OpenReview API", "abstract": "Bayesian methods for learning predictive models have the ability to consider both sources of uncertainty (i.e., data and model uncertainty) within a single framework and thereby provide a powerful tool for decision-making. Bayesian neural networks (BNNs) hold great potential for  training data efficiency due to full uncertainty quantification, making them promising candidates for more data-efficient AI in data-constrained settings such as reinforcement learning in the physical world.  However, current computational approaches for learning BNNs often face limitations such as overconfidence, sensitivity to hyperparameters, and posterior collapse, highlighting the need for alternative computational approaches. In this paper, we introduce a novel method that leverages approximate message passing (MP) of a full factorized neural network model using mixed approximations to overcome these problems while maintaining data efficiency. Our framework supports convolutional neural networks while addressing the issue of double-counting training data, which has been a key source of overconfidence in prior work. We demonstrate the data-efficiency of our method on multiple benchmark datasets in comparison to state-of-the-art methods for learning neural networks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "mOVxU0IGjF", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17993/Reviewer_mpkr"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper trains BNNs using “approximate message passing on factor graphs”: it combines Gaussian messages in batches to avoid duplicate counting, scales to CNNs, and improves calibration and OoD performance, but training and memory costs remain high", "review_text": "This paper trains BNNs using “approximate message passing on factor graphs”: it combines Gaussian messages in batches to avoid duplicate counting, scales to CNNs, and improves calibration and OoD performance, but training and memory costs remain high", "strengths": "1. Laplace/EP-style Gaussian messages makes stable training and avoid posterior collapse\n2. Batch message aggregation mitigates double-counting and better calibration than AdamW/ensembles", "weaknesses": "1. Computational and memory overhead is high, as Gaussian messages are maintained per sample, resulting in scalability that falls short of standard training pipelines like SGD/Adam\n2. Insufficient large-scale validation, lacking systematic comparisons and ablation studies on larger datasets/modern benchmarks\n3. The processing of residuals, normalization, and other modern layers remains unclear, and there is insufficient evidence of generalization beyond simple CNNs\n4. Loopy BP lacks convergence guarantees. Although the authors acknowledge that guarantees only hold under specific conditions (such as Simon's condition) and are difficult to verify, still remains some concerns", "questions": "The computational power/graphics memory cost of the method is excessively high, its scalability evidence is insufficient, and its effectiveness and stability in medium-to-large-scale, modern networks (residual/normalized) and larger datasets remain poorly substantiated by robust empirical evidence", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper trains BNNs using “approximate message passing on factor graphs”: it combines Gaussian messages in batches to avoid duplicate counting, scales to CNNs, and improves calibration and OoD performance, but training and memory costs remain high", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Laplace/EP-style Gaussian messages makes stable training and avoid posterior collapse\n2. Batch message aggregation mitigates double-counting and better calibration than AdamW/ensembles", "weaknesses": "1. Computational and memory overhead is high, as Gaussian messages are maintained per sample, resulting in scalability that falls short of standard training pipelines like SGD/Adam\n2. Insufficient large-scale validation, lacking systematic comparisons and ablation studies on larger datasets/modern benchmarks\n3. The processing of residuals, normalization, and other modern layers remains unclear, and there is insufficient evidence of generalization beyond simple CNNs\n4. Loopy BP lacks convergence guarantees. Although the authors acknowledge that guarantees only hold under specific conditions (such as Simon's condition) and are difficult to verify, still remains some concerns", "questions": "The computational power/graphics memory cost of the method is excessively high, its scalability evidence is insufficient, and its effectiveness and stability in medium-to-large-scale, modern networks (residual/normalized) and larger datasets remain poorly substantiated by robust empirical evidence", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761953307775}, {"id": "w9HSGlca4P", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17993/Reviewer_bdZB"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a message-passing framework for Bayesian Neural Networks (BNNs) that reformulates training as inference on a factor graph using a diagonal-Gaussian approximation family. It addresses the double-counting problem in existing variational inference methods by introducing a batch-wise “divide-then-multiply” update scheme and a layer-level implementation. The authors implement the method in Julia with GPU support, demonstrating its scalability to MLPs and CNNs and improved calibration over AdamW, IVON, and Deep Ensembles on small-scale benchmarks.", "review_text": "This paper proposes a message-passing framework for Bayesian Neural Networks (BNNs) that reformulates training as inference on a factor graph using a diagonal-Gaussian approximation family. It addresses the double-counting problem in existing variational inference methods by introducing a batch-wise “divide-then-multiply” update scheme and a layer-level implementation. The authors implement the method in Julia with GPU support, demonstrating its scalability to MLPs and CNNs and improved calibration over AdamW, IVON, and Deep Ensembles on small-scale benchmarks.", "strengths": "1. The paper introduces a message-passing framework for Bayesian neural networks that provides a principled way to model uncertainty while mitigating issues such as data double-counting and overconfidence observed in earlier Bayesian inference methods.\n\n2. The framework is mathematically coherent, deriving closed-form Gaussian message updates for common factor types and supporting deterministic training without reliance on sampling-based approximations.\n\n3. Experiments demonstrate that the proposed approach produces well-calibrated uncertainty estimates and competitive predictive performance across several benchmarks.", "weaknesses": "1. Expanding one convolutional kernel into tens of thousands of scalar $\\delta$ factors substantially increases the size of the factor graph so that memory usage is tied to the count of scalar edges, not to the parameter count. This raises practical scalability problems for deep or wide CNNs.\n\n2. Each ReLU layer is handled by moment-matched Gaussian messages. But since the exact moments are computed under a Gaussian assumption that ignores the truncated tail, the resulting estimates may introduce small systematic biases. These biases could accumulate over multiple layers, potentially affecting calibration accuracy in deeper models.\n\n3. The batch-wise “divide-then-multiply” update operates in precision space by dividing the current marginal by the old batch message. When the old message carries low precision, this division amplifies rounding errors and often produces negative precisions that must be clamped to preserve numerical validity. These corrective heuristics reveal intrinsic instability in the update rule, which could undermine convergence and consistency across training iterations.", "questions": "1. Have you tested any sparse or tensor-factorised representations that keep a single factor per convolutional kernel element instead of one $\\delta$ -node per scalar multiply–add, and if so what reduction in peak GPU memory did you observe without changing the converged marginal means?\n\n2. Have you checked the skewness or other higher-order moments of the ReLU messages to verify that the Gaussian approximation remains adequate across depth?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a message-passing framework for Bayesian Neural Networks (BNNs) that reformulates training as inference on a factor graph using a diagonal-Gaussian approximation family. It addresses the double-counting problem in existing variational inference methods by introducing a batch-wise “divide-then-multiply” update scheme and a layer-level implementation. The authors implement the method in Julia with GPU support, demonstrating its scalability to MLPs and CNNs and improved calibration over AdamW, IVON, and Deep Ensembles on small-scale benchmarks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The paper introduces a message-passing framework for Bayesian neural networks that provides a principled way to model uncertainty while mitigating issues such as data double-counting and overconfidence observed in earlier Bayesian inference methods.\n\n2. The framework is mathematically coherent, deriving closed-form Gaussian message updates for common factor types and supporting deterministic training without reliance on sampling-based approximations.\n\n3. Experiments demonstrate that the proposed approach produces well-calibrated uncertainty estimates and competitive predictive performance across several benchmarks.", "weaknesses": "1. Expanding one convolutional kernel into tens of thousands of scalar $\\delta$ factors substantially increases the size of the factor graph so that memory usage is tied to the count of scalar edges, not to the parameter count. This raises practical scalability problems for deep or wide CNNs.\n\n2. Each ReLU layer is handled by moment-matched Gaussian messages. But since the exact moments are computed under a Gaussian assumption that ignores the truncated tail, the resulting estimates may introduce small systematic biases. These biases could accumulate over multiple layers, potentially affecting calibration accuracy in deeper models.\n\n3. The batch-wise “divide-then-multiply” update operates in precision space by dividing the current marginal by the old batch message. When the old message carries low precision, this division amplifies rounding errors and often produces negative precisions that must be clamped to preserve numerical validity. These corrective heuristics reveal intrinsic instability in the update rule, which could undermine convergence and consistency across training iterations.", "questions": "1. Have you tested any sparse or tensor-factorised representations that keep a single factor per convolutional kernel element instead of one $\\delta$ -node per scalar multiply–add, and if so what reduction in peak GPU memory did you observe without changing the converged marginal means?\n\n2. Have you checked the skewness or other higher-order moments of the ReLU messages to verify that the Gaussian approximation remains adequate across depth?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761870459154}, {"id": "K4bWc3a5SD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17993/Reviewer_JWMe"], "rating": 2, "soundness": 1, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The paper introduces a novel approximate message-passing framework for Bayesian Neural Network inference that scales to architectures such as CNNs and MLPs. Unlike earlier approaches such as Expectation Backpropagation (EBP) and Probabilistic Backpropagation (PBP), the proposed method explicitly avoids double-counting of training data and prevents posterior collapse. The framework models BNNs as factor graphs and derives Gaussian message approximations for all relevant factors. The framework is tested on MNIST, CIFAR-10, and  UCI regression benchmarks, comparing to baselines like SGD, IVON, and Deep Ensembles.", "review_text": "The paper introduces a novel approximate message-passing framework for Bayesian Neural Network inference that scales to architectures such as CNNs and MLPs. Unlike earlier approaches such as Expectation Backpropagation (EBP) and Probabilistic Backpropagation (PBP), the proposed method explicitly avoids double-counting of training data and prevents posterior collapse. The framework models BNNs as factor graphs and derives Gaussian message approximations for all relevant factors. The framework is tested on MNIST, CIFAR-10, and  UCI regression benchmarks, comparing to baselines like SGD, IVON, and Deep Ensembles.", "strengths": "1. The work attempts to address a core flaw in prior message-passing BNNs (double-counting), which may lead to posterior collapsing\n\n2. The work extends message passing to CNNs, while earlier works were limited to MLPs.\n\n3. The framework formulates BNN inference as belief propagation in a factor graph with analytically tractable Gaussian message approximations, which is computationally efficient.", "weaknesses": "1. I unfortunately find that the empirical validation is insufficient to fully support the paper’s claims.\n    * The authors claim their message-passing framework alleviates overconfidence and posterior collapse compared to prior approaches such as EBP and PBP. However, these methods are not included as experimental baselines, making it difficult to verify those claims empirically.\n    * The choice and coverage of baselines are inconsistent across experiments. For example, Figure 3a omits R-SGD, Figure 3b shows Deep Ensemble results only, and Figures 3c, 3d exclude AM-MP and R-SGD. It is also unclear which architecture (MLP or LeNet-5) Figure 3 corresponds to.\n    * In Table 4, the use of RMSE alone is insufficient to assess calibration quality or uncertainty estimation. Complementary metrics such as coverage or NLL would strengthen the evidence.\n\n2. There might be potential fragility and computational cost due to numerous numerical stabilizations. The framework relies on several ad-hoc stability guardrails (Section 4), including special handling for activations like LeakyReLU, re-normalization of small constants, and periodic recomputation of weight marginals. While these measures improve numerical stability, they may introduce additional computational overhead and raise concerns about the robustness and generalizability of the approach across architectures or datasets.\n\n 3. Minor on the writing side. There are a lot of acronyms that are used without being introduced or introduced later, such as IVON, AM-MP", "questions": "1. How are regression-based MPs trained on classification tasks?\n2. Can the \"double-counting\" issue of EBP and PBP be overcome by using the tempered posterior [1]\n\n\n\n[1] Ng, Kenyon, et al. \"Temperature Optimization for Bayesian Deep Learning.\" arXiv preprint arXiv:2410.05757 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel approximate message-passing framework for Bayesian Neural Network inference that scales to architectures such as CNNs and MLPs. Unlike earlier approaches such as Expectation Backpropagation (EBP) and Probabilistic Backpropagation (PBP), the proposed method explicitly avoids double-counting of training data and prevents posterior collapse. The framework models BNNs as factor graphs and derives Gaussian message approximations for all relevant factors. The framework is tested on MNIST, CIFAR-10, and  UCI regression benchmarks, comparing to baselines like SGD, IVON, and Deep Ensembles.", "soundness": 1, "presentation": 1, "contribution": 2, "strengths": "1. The work attempts to address a core flaw in prior message-passing BNNs (double-counting), which may lead to posterior collapsing\n\n2. The work extends message passing to CNNs, while earlier works were limited to MLPs.\n\n3. The framework formulates BNN inference as belief propagation in a factor graph with analytically tractable Gaussian message approximations, which is computationally efficient.", "weaknesses": "1. I unfortunately find that the empirical validation is insufficient to fully support the paper’s claims.\n    * The authors claim their message-passing framework alleviates overconfidence and posterior collapse compared to prior approaches such as EBP and PBP. However, these methods are not included as experimental baselines, making it difficult to verify those claims empirically.\n    * The choice and coverage of baselines are inconsistent across experiments. For example, Figure 3a omits R-SGD, Figure 3b shows Deep Ensemble results only, and Figures 3c, 3d exclude AM-MP and R-SGD. It is also unclear which architecture (MLP or LeNet-5) Figure 3 corresponds to.\n    * In Table 4, the use of RMSE alone is insufficient to assess calibration quality or uncertainty estimation. Complementary metrics such as coverage or NLL would strengthen the evidence.\n\n2. There might be potential fragility and computational cost due to numerous numerical stabilizations. The framework relies on several ad-hoc stability guardrails (Section 4), including special handling for activations like LeakyReLU, re-normalization of small constants, and periodic recomputation of weight marginals. While these measures improve numerical stability, they may introduce additional computational overhead and raise concerns about the robustness and generalizability of the approach across architectures or datasets.\n\n 3. Minor on the writing side. There are a lot of acronyms that are used without being introduced or introduced later, such as IVON, AM-MP", "questions": "1. How are regression-based MPs trained on classification tasks?\n2. Can the \"double-counting\" issue of EBP and PBP be overcome by using the tempered posterior [1]\n\n\n\n[1] Ng, Kenyon, et al. \"Temperature Optimization for Bayesian Deep Learning.\" arXiv preprint arXiv:2410.05757 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761858053727}, {"id": "0cEUdTioi6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17993/Reviewer_kFYd"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes a novel Bayesian neural network (BNN) training framework based on **approximate message passing (MP)** over a factor graph representation of the predictive posterior. The key claim is that by modeling the network’s joint posterior via a factor graph with scalar latent variables and applying Gaussian message approximations, one can train BNNs (even convolutional nets) while accurately propagating uncertainty. Notably, the authors emphasize that their algorithm explicitly **avoids “double-counting”** of data as a known issue in prior MP-based BNN methods, which they argue has caused overconfident predictions in earlier approaches. The method is evaluated on several benchmarks. Empirically, the MP method achieves accuracy comparable to strong baselines while often exhibiting better calibration. The paper positions its contributions as: (1) a new MP framework with closed-form Gaussian message updates for all factors; (2) a practical implementation supporting CNNs and avoiding data double-counting; and (3) empirical evidence that this method is competitive with state-of-the-art approaches while improving uncertainty estimates.", "review_text": "The paper proposes a novel Bayesian neural network (BNN) training framework based on **approximate message passing (MP)** over a factor graph representation of the predictive posterior. The key claim is that by modeling the network’s joint posterior via a factor graph with scalar latent variables and applying Gaussian message approximations, one can train BNNs (even convolutional nets) while accurately propagating uncertainty. Notably, the authors emphasize that their algorithm explicitly **avoids “double-counting”** of data as a known issue in prior MP-based BNN methods, which they argue has caused overconfident predictions in earlier approaches. The method is evaluated on several benchmarks. Empirically, the MP method achieves accuracy comparable to strong baselines while often exhibiting better calibration. The paper positions its contributions as: (1) a new MP framework with closed-form Gaussian message updates for all factors; (2) a practical implementation supporting CNNs and avoiding data double-counting; and (3) empirical evidence that this method is competitive with state-of-the-art approaches while improving uncertainty estimates.", "strengths": "1.\t**Novel MP framework:** Introduces a principled factor-graph/BP approach for BNNs, deriving all necessary Gaussian message equations. This is a first step in applying belief propagation to large neural networks.\n\n2.\t**CNN support and data-debiasing:** Unlike earlier MP/EP methods, this framework handles convolutional layers and explicitly avoids “double-counting” of data during batch updates. This addresses a known cause of overconfident predictions in previous work.\n\n3.\t**Strong uncertainty calibration:** Empirically, the method yields excellent predictive uncertainty. For instance, on MNIST with limited data, MP achieved low expected calibration error (∼0.02) vs very high error for SGD. On CIFAR, MP’s ECE was the lowest among baselines. Synthetic tests show that credible intervals track true coverage well (correlation ~0.9). These demonstrate the method’s efficacy in capturing epistemic uncertainty.\n\n4.\t**Broad experimentation:** The authors evaluated on a variety of settings: tabular UCI regression, synthetic functions, MNIST and CIFAR image tasks. They compare to strong baselines (AdamW, IVON, deep ensembles) across multiple metrics (accuracy, NLL, Brier, OOD-AUC). This thoroughness lends credibility to their claims.\n\n5.\t**Clear writing of results:** The paper clearly reports results and highlights key findings (e.g. calibration advantage) in the text accompanying each table. The contributions and limitations are transparently listed.", "weaknesses": "1.\t**Heaviness of approximations:** The core MP algorithm uses Gaussian approximations for all messages without theoretical guarantees. Such approximations in highly nonlinear, loopy graphs may introduce bias. The paper does not analyze the accuracy or failure modes of these approximations, leaving uncertainty about when MP might break down.\n\n2.\t**Scalability concerns:** The method is computationally expensive. As acknowledged by the authors, training is up to two orders of magnitude slower and more memory-intensive than standard SGD/AdamW. Each training example carries a full set of messages, and the Julia implementation uses double-precision for safety, making it impractical for very large models. \n\n3.\t**Limited experimental breadth:** The neural architecture used is relatively small (890K parameters, plain 6-layer CNN). Modern benchmarks (ResNet, Transformers) with residuals or normalization are not evaluated. It is unclear how the method handles practical architectures (batch-norm, skip-connections). This restricts claims of general applicability.\n\n4.\t**Mixed empirical performance:** Although calibration is good, the MP method’s predictive accuracy and NLL are inferior to top alternatives. For example, on CIFAR-10 Deep Ensembles achieve 81.9% accuracy vs 77.3% for MP. The paper notes that differences in architecture (lack of normalization/residuals) may explain this, but it underscores that MP does not yet match state-of-art predictive performance in high-dimensional tasks.\n\n5.\t**Reproducibility details:** The paper defers many architectural and hyperparameter details to the Appendix. Important choices (e.g. noise priors, initialization variances) are only briefly mentioned. The lack of multiple runs or error bars also raises questions about robustness.", "questions": "1.\t**Double-counting fix:** Can you clarify the theoretical justification for the batch-update rule (dividing out old aggregate messages) that prevents double-counting? Does this exactly recover Bayesian posterior updates, or is it a heuristic? How sensitive is training to this scheme?\n\n2.\t**Scalability to modern nets:** Have you tried incorporating residual connections or normalization (e.g. BatchNorm)? You mention a possible factor for division, but has this been validated? Can the MP algorithm scale to architectures like ResNets or Transformers?\n\n3.\t**Multiple passes or convergence:** How many MP iterations (sweeps through the factor graph) are performed per batch? Is the algorithm guaranteed to converge, or do you observe oscillations? Have you considered damping or other techniques to stabilize message updates?\n\n4.\t**Baselines and hyperparameters:** How were AdamW and IVON tuned? Are the baselines given the best possible schedules and hyperparameters? For CIFAR, why only 25 epochs? Would longer training change the comparisons?\n\n5.\t**Prediction procedure:** For classification, you mention an “argmax factor” in Appendix F. Does this mean you approximate the softmax? How are predictive probabilities computed at test time? Clarify how classification uncertainty is obtained from the Gaussian-approximated network outputs.\n\n6.\t**Posterior quality diagnostics:** Beyond ECE and credible intervals, did you assess other uncertainty metrics (e.g. negative log-likelihood on OOD, Brier decomposition)? Did you examine any divergences or posterior samples to verify that the approximate posterior makes sense (e.g. covariance structure)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel Bayesian neural network (BNN) training framework based on **approximate message passing (MP)** over a factor graph representation of the predictive posterior. The key claim is that by modeling the network’s joint posterior via a factor graph with scalar latent variables and applying Gaussian message approximations, one can train BNNs (even convolutional nets) while accurately propagating uncertainty. Notably, the authors emphasize that their algorithm explicitly **avoids “double-counting”** of data as a known issue in prior MP-based BNN methods, which they argue has caused overconfident predictions in earlier approaches. The method is evaluated on several benchmarks. Empirically, the MP method achieves accuracy comparable to strong baselines while often exhibiting better calibration. The paper positions its contributions as: (1) a new MP framework with closed-form Gaussian message updates for all factors; (2) a practical implementation supporting CNNs and avoiding data double-counting; and (3) empirical evidence that this method is competitive with state-of-the-art approaches while improving uncertainty estimates.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1.\t**Novel MP framework:** Introduces a principled factor-graph/BP approach for BNNs, deriving all necessary Gaussian message equations. This is a first step in applying belief propagation to large neural networks.\n\n2.\t**CNN support and data-debiasing:** Unlike earlier MP/EP methods, this framework handles convolutional layers and explicitly avoids “double-counting” of data during batch updates. This addresses a known cause of overconfident predictions in previous work.\n\n3.\t**Strong uncertainty calibration:** Empirically, the method yields excellent predictive uncertainty. For instance, on MNIST with limited data, MP achieved low expected calibration error (∼0.02) vs very high error for SGD. On CIFAR, MP’s ECE was the lowest among baselines. Synthetic tests show that credible intervals track true coverage well (correlation ~0.9). These demonstrate the method’s efficacy in capturing epistemic uncertainty.\n\n4.\t**Broad experimentation:** The authors evaluated on a variety of settings: tabular UCI regression, synthetic functions, MNIST and CIFAR image tasks. They compare to strong baselines (AdamW, IVON, deep ensembles) across multiple metrics (accuracy, NLL, Brier, OOD-AUC). This thoroughness lends credibility to their claims.\n\n5.\t**Clear writing of results:** The paper clearly reports results and highlights key findings (e.g. calibration advantage) in the text accompanying each table. The contributions and limitations are transparently listed.", "weaknesses": "1.\t**Heaviness of approximations:** The core MP algorithm uses Gaussian approximations for all messages without theoretical guarantees. Such approximations in highly nonlinear, loopy graphs may introduce bias. The paper does not analyze the accuracy or failure modes of these approximations, leaving uncertainty about when MP might break down.\n\n2.\t**Scalability concerns:** The method is computationally expensive. As acknowledged by the authors, training is up to two orders of magnitude slower and more memory-intensive than standard SGD/AdamW. Each training example carries a full set of messages, and the Julia implementation uses double-precision for safety, making it impractical for very large models. \n\n3.\t**Limited experimental breadth:** The neural architecture used is relatively small (890K parameters, plain 6-layer CNN). Modern benchmarks (ResNet, Transformers) with residuals or normalization are not evaluated. It is unclear how the method handles practical architectures (batch-norm, skip-connections). This restricts claims of general applicability.\n\n4.\t**Mixed empirical performance:** Although calibration is good, the MP method’s predictive accuracy and NLL are inferior to top alternatives. For example, on CIFAR-10 Deep Ensembles achieve 81.9% accuracy vs 77.3% for MP. The paper notes that differences in architecture (lack of normalization/residuals) may explain this, but it underscores that MP does not yet match state-of-art predictive performance in high-dimensional tasks.\n\n5.\t**Reproducibility details:** The paper defers many architectural and hyperparameter details to the Appendix. Important choices (e.g. noise priors, initialization variances) are only briefly mentioned. The lack of multiple runs or error bars also raises questions about robustness.", "questions": "1.\t**Double-counting fix:** Can you clarify the theoretical justification for the batch-update rule (dividing out old aggregate messages) that prevents double-counting? Does this exactly recover Bayesian posterior updates, or is it a heuristic? How sensitive is training to this scheme?\n\n2.\t**Scalability to modern nets:** Have you tried incorporating residual connections or normalization (e.g. BatchNorm)? You mention a possible factor for division, but has this been validated? Can the MP algorithm scale to architectures like ResNets or Transformers?\n\n3.\t**Multiple passes or convergence:** How many MP iterations (sweeps through the factor graph) are performed per batch? Is the algorithm guaranteed to converge, or do you observe oscillations? Have you considered damping or other techniques to stabilize message updates?\n\n4.\t**Baselines and hyperparameters:** How were AdamW and IVON tuned? Are the baselines given the best possible schedules and hyperparameters? For CIFAR, why only 25 epochs? Would longer training change the comparisons?\n\n5.\t**Prediction procedure:** For classification, you mention an “argmax factor” in Appendix F. Does this mean you approximate the softmax? How are predictive probabilities computed at test time? Clarify how classification uncertainty is obtained from the Gaussian-approximated network outputs.\n\n6.\t**Posterior quality diagnostics:** Beyond ECE and credible intervals, did you assess other uncertainty metrics (e.g. negative log-likelihood on OOD, Brier decomposition)? Did you examine any divergences or posterior samples to verify that the approximate posterior makes sense (e.g. covariance structure)?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethical issues were identified.", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761780448967}], "openreview_url": "https://openreview.net/forum?id=bDee2EgvWJ", "arxiv_id": "2501.15573", "paper_pdf": "papers/bDee2EgvWJ.pdf", "paper_pdf_sha256": "d8c03d50a4202cf7a5f7fddd5623d4d2fc085b6c0cecea79b7135e578188bd61", "paper_pdf_bytes": 2803386, "paper_pdf_source": "openreview", "code_url": "https://github.com/christian-helms/mpbnns", "code_repository": "christian-helms/mpbnns", "code_commit": "85f12084c7ab9d89acb9628b02413e841c7346f7", "code_archive": "repos/bDee2EgvWJ.zip", "code_archive_sha256": "ebbf2273e0467ed6483c8c79ce6b024267b2bfa88e09167dff7e7eeac3efe2fc", "code_archive_bytes": 57358, "code_file_count": 21, "code_extensions": {".jl": 17, ".py": 3, ".sh": 1}, "github_disk_usage_kb": 54, "github_languages": {"Julia": 180263, "Python": 19933, "Shell": 388}, "github_archived": false, "github_pushed_at": "2025-01-26T17:12:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/approximate-message-passing-for-bayesian"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lAXlDAdan5", "year": 2025, "status": "rejected", "title": "Accelerating Error Correction Code Transformers", "authors": ["Matan Levy", "Yoni Choukroun", "Lior Wolf"], "authorids": ["~Matan_Levy2", "~Yoni_Choukroun1", "~Lior_Wolf1"], "authors_source": "OpenReview API", "abstract": "Error correction codes (ECC) are crucial for ensuring reliable information transmission in communication systems. Choukroun & Wolf (2022b) recently introduced the Error Correction Code Transformer (ECCT), which has demonstrated promising performance across various transmission channels and families of codes. However, its high computational and memory demands limit its practical applications compared to traditional decoding algorithms. Achieving effective quantization of the ECCT presents significant challenges due to its inherently small architecture, since existing, very low-precision quantization techniques often lead to performance degradation in compact neural networks. In this paper, we introduce a novel acceleration method for transformer-based decoders. We first propose a ternary weight quantization method specifically designed for the ECCT, inducing a decoder with multiplication-free linear layers. We present an\noptimized self-attention mechanism to reduce computational complexity via codeaware multi-heads processing. Finally, we provide positional encoding via the Tanner graph eigendecomposition, enabling a richer representation of the graph connectivity. The approach not only matches or surpasses ECCT’s performance but also significantly reduces energy consumption, memory footprint, and computational complexity. Our method brings transformer-based error correction closer to practical implementation in resource-constrained environments, achieving a 90% compression ratio and reducing arithmetic operation energy consumption by at least 224 times on modern hardware.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "3hUUFirpgc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_g7ig"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces simplifications to the Error Correction Code Transformer (ECCT) neural decoder model to reduce the memory and compute requirements. The authors claim that conventional quantization methods result in a significant drop in performance and instead propose alternative approaches to reduce the complexity. \n\nFirst is Adaptive Absolute Percentile (AAP) quantization, to optimize the percentile of scaling factor (percentage of absolute values). \n\nSecond it head partitioning self-attention to limit the attention to either between similar nodes (check/variable) or between different nodes (check<->variable), dividing the attention heads into two groups. This significantly reduces the number of active elements in the attention mask.\n\nFinally, uses spectral positional encoding (SPE) to introduce an inductive bias about the tarnner graph into the model.", "review_text": "This paper introduces simplifications to the Error Correction Code Transformer (ECCT) neural decoder model to reduce the memory and compute requirements. The authors claim that conventional quantization methods result in a significant drop in performance and instead propose alternative approaches to reduce the complexity. \n\nFirst is Adaptive Absolute Percentile (AAP) quantization, to optimize the percentile of scaling factor (percentage of absolute values). \n\nSecond it head partitioning self-attention to limit the attention to either between similar nodes (check/variable) or between different nodes (check<->variable), dividing the attention heads into two groups. This significantly reduces the number of active elements in the attention mask.\n\nFinally, uses spectral positional encoding (SPE) to introduce an inductive bias about the tarnner graph into the model.", "strengths": "1. Tackles an important problem of reducing the complexity of neural decoders. \n\n2. The performance improvements in terms of memeory and compute are non-trivial and very impressive. \n\n3. All the three main ideas proposed are novel and interesting. Specifically, splitting the masking based on the node structure is a clever example of leveraging domain structure. \n\n4. Set of experiements are exhaustive and includes sufficient number of results as well as ablation studies to show the merits of proposed ideas.", "weaknesses": "1. While a lot of interesting ideas were discussed in the paper to improve the efficiency of ECCT, I am not fully convinced about the practical relevance of a \"universal decoder\" architecture that simply utilizes a parity check matrix for all codes. In coding theory, each well-known family of codes has highly specific representations and special propoerties that cannot be fully captured by a simple parity check matrix. For instance, in Polar codes, the reliability sequence and the sequential decoding of information bits plays a crucial role in deciding the performance of the code. Claiming better performance for all family of codes w.r.t BP decoder is a bit misleading. Again, I appreciate the line of ECCT and related works from an academic curiousity point of view, but in my opinion, the merits and drawbacks should be highlighted more clearly without any overpromising. \n\n2. One common troubling trend in comparing the performance of neural decoders is the choice of weak classical baselines. While I understand the difficulty of a common decoder architecture outperforming the highly specialised decoders for each of the class of codes, the comparison should neverthless be done for completeness. For instance, it is known that BP is not optimal for many codes. For instance, when comparing with BCH codes, Berlekamp-Massey decoding algorithm should be used as the classical non-learning baseline and similarly for Polar codes, successive cancellation list decoding should be used. I strongly suggest authors to include the best classical baselines for each code to clearly show the gap with respect to the SOTA neural decoder, whether it is positive or negative.\n\n3. I beleive authors chose the format of the negative log BER for presenting the results based on the original ECCT paper. But this format is not very informative and very uncommon in information and coding theory literature. While it is easy to identify which scheme is doing better, the gap between the performances is hard to interpret as the scale is not very intuitive. Rather, the standard format of SNR gap in dB for a target BER/BLER is more informative.\n\n4. \"Our analysis shows that AAP outperforms AP quantization\" --> I see from ablation studied in Table 2 that the performace difference between ECCT + AP vs ECCT + APP is very small (0.01 - 0.1). Again, this scale of -log(BER) is very non-intuitive but in my opinion, this is negligible difference in performance and brings into question the value of doing AAP over AP. \n\n5. While the head partitioning idea is interesting, is it also similar to the idea proposed in the paper \"CrossMPT: Cross-attention Message-Passing Transformer for Error Correcting Codes\", specifically the second-ring MP. \n\n6. The details about SPE are unclear and I do not understand the \"spectral\" positioning and the \"Laplacian eigenspace\" ideas discussed. Rewriting the seubsection to simplify the discussion of these ideas would be helpful.", "questions": "Covered as part of weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces simplifications to the Error Correction Code Transformer (ECCT) neural decoder model to reduce the memory and compute requirements. The authors claim that conventional quantization methods result in a significant drop in performance and instead propose alternative approaches to reduce the complexity. \n\nFirst is Adaptive Absolute Percentile (AAP) quantization, to optimize the percentile of scaling factor (percentage of absolute values). \n\nSecond it head partitioning self-attention to limit the attention to either between similar nodes (check/variable) or between different nodes (check<->variable), dividing the attention heads into two groups. This significantly reduces the number of active elements in the attention mask.\n\nFinally, uses spectral positional encoding (SPE) to introduce an inductive bias about the tarnner graph into the model.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Tackles an important problem of reducing the complexity of neural decoders. \n\n2. The performance improvements in terms of memeory and compute are non-trivial and very impressive. \n\n3. All the three main ideas proposed are novel and interesting. Specifically, splitting the masking based on the node structure is a clever example of leveraging domain structure. \n\n4. Set of experiements are exhaustive and includes sufficient number of results as well as ablation studies to show the merits of proposed ideas.", "weaknesses": "1. While a lot of interesting ideas were discussed in the paper to improve the efficiency of ECCT, I am not fully convinced about the practical relevance of a \"universal decoder\" architecture that simply utilizes a parity check matrix for all codes. In coding theory, each well-known family of codes has highly specific representations and special propoerties that cannot be fully captured by a simple parity check matrix. For instance, in Polar codes, the reliability sequence and the sequential decoding of information bits plays a crucial role in deciding the performance of the code. Claiming better performance for all family of codes w.r.t BP decoder is a bit misleading. Again, I appreciate the line of ECCT and related works from an academic curiousity point of view, but in my opinion, the merits and drawbacks should be highlighted more clearly without any overpromising. \n\n2. One common troubling trend in comparing the performance of neural decoders is the choice of weak classical baselines. While I understand the difficulty of a common decoder architecture outperforming the highly specialised decoders for each of the class of codes, the comparison should neverthless be done for completeness. For instance, it is known that BP is not optimal for many codes. For instance, when comparing with BCH codes, Berlekamp-Massey decoding algorithm should be used as the classical non-learning baseline and similarly for Polar codes, successive cancellation list decoding should be used. I strongly suggest authors to include the best classical baselines for each code to clearly show the gap with respect to the SOTA neural decoder, whether it is positive or negative.\n\n3. I beleive authors chose the format of the negative log BER for presenting the results based on the original ECCT paper. But this format is not very informative and very uncommon in information and coding theory literature. While it is easy to identify which scheme is doing better, the gap between the performances is hard to interpret as the scale is not very intuitive. Rather, the standard format of SNR gap in dB for a target BER/BLER is more informative.\n\n4. \"Our analysis shows that AAP outperforms AP quantization\" --> I see from ablation studied in Table 2 that the performace difference between ECCT + AP vs ECCT + APP is very small (0.01 - 0.1). Again, this scale of -log(BER) is very non-intuitive but in my opinion, this is negligible difference in performance and brings into question the value of doing AAP over AP. \n\n5. While the head partitioning idea is interesting, is it also similar to the idea proposed in the paper \"CrossMPT: Cross-attention Message-Passing Transformer for Error Correcting Codes\", specifically the second-ring MP. \n\n6. The details about SPE are unclear and I do not understand the \"spectral\" positioning and the \"Laplacian eigenspace\" ideas discussed. Rewriting the seubsection to simplify the discussion of these ideas would be helpful.", "questions": "Covered as part of weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731014948514}, {"id": "qTgXZaEVFQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_xTJU"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces Accelerated Error Correction Code Transformer (AECCT), which reduces the memory footprint and computational complexity of Error Correction Code Transformers (ECCT). The proposed technique introduces three distinct novel components:\n1. Adaptive Absolute Percentile (AAP) Quantization: This method compresses model weights to ternary values, reducing memory and energy use while maintaining model accuracy. Unlike traditional quantization, AAP dynamically adjusts sparsity to retain essential features.\n2. Head Partitioning Self-Attention (HPSA): This self-attention mechanism reduces computational complexity by dividing attention heads to focus separately on first-ring (neighbor) and second-ring (more distant) connections within ECC’s Tanner graph. HPSA achieves higher sparsity, thus lowering computational costs compared to ECCT’s Code-Aware Self-Attention (CASA).\n3. Spectral Positional Encoding (SPE): By embedding structural information from the Tanner graph’s Laplacian eigenspace, SPE enriches the model’s positional representation, enhancing the decoding ability without increasing runtime.\n\nUsing these components, the authors demonstrate a 90% compression ratio and reduce energy consumption by over 224 times compared to ECCT, approaching the efficiency of traditional algorithms like Belief Propagation (BP), making practical deployment possible.", "review_text": "This paper introduces Accelerated Error Correction Code Transformer (AECCT), which reduces the memory footprint and computational complexity of Error Correction Code Transformers (ECCT). The proposed technique introduces three distinct novel components:\n1. Adaptive Absolute Percentile (AAP) Quantization: This method compresses model weights to ternary values, reducing memory and energy use while maintaining model accuracy. Unlike traditional quantization, AAP dynamically adjusts sparsity to retain essential features.\n2. Head Partitioning Self-Attention (HPSA): This self-attention mechanism reduces computational complexity by dividing attention heads to focus separately on first-ring (neighbor) and second-ring (more distant) connections within ECC’s Tanner graph. HPSA achieves higher sparsity, thus lowering computational costs compared to ECCT’s Code-Aware Self-Attention (CASA).\n3. Spectral Positional Encoding (SPE): By embedding structural information from the Tanner graph’s Laplacian eigenspace, SPE enriches the model’s positional representation, enhancing the decoding ability without increasing runtime.\n\nUsing these components, the authors demonstrate a 90% compression ratio and reduce energy consumption by over 224 times compared to ECCT, approaching the efficiency of traditional algorithms like Belief Propagation (BP), making practical deployment possible.", "strengths": "1. The paper is well written and easy to follow, and the contributions are relevant.\n2. The techniques introduced demonstrate significant compression without performance degradation compared to ECCTs.\n3. The technique is tested on 3 different ECC types - Polar, Low-Density Parity Check (LDPC), and Bose–Chaudhuri–Hocquenghem (BCH) codes - demonstrating robustness and applicability across various error correction scenarios.\n4. The authors conduct ablation studies evaluating the impact of each of the three components.", "weaknesses": "1. Two out of three proposed techniques (HPSA and SPE) may have limited applicability outside of ECC.\n2. It is not clear how easy it will be to implement these techniques in existing hardware. Ternary quantization has been explored by previous work. Beyond quantization, does HPSA and SPE introduce additional implementation complexities on existing hardware? \n3. The energy efficiency numbers are estimates, the paper does not provide any real measured numbers on actual hardware.", "questions": "1. Discuss any potential challenges or modifications needed to implement HPSA and SPE on current hardware platforms used for error correction.\n2. Compare the implementation complexity of HPSA and SPE to that of the original ECCT approach.\n3. If possible, provide estimates or examples of how these techniques might impact hardware design or utilization in practical ECC systems.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Accelerated Error Correction Code Transformer (AECCT), which reduces the memory footprint and computational complexity of Error Correction Code Transformers (ECCT). The proposed technique introduces three distinct novel components:\n1. Adaptive Absolute Percentile (AAP) Quantization: This method compresses model weights to ternary values, reducing memory and energy use while maintaining model accuracy. Unlike traditional quantization, AAP dynamically adjusts sparsity to retain essential features.\n2. Head Partitioning Self-Attention (HPSA): This self-attention mechanism reduces computational complexity by dividing attention heads to focus separately on first-ring (neighbor) and second-ring (more distant) connections within ECC’s Tanner graph. HPSA achieves higher sparsity, thus lowering computational costs compared to ECCT’s Code-Aware Self-Attention (CASA).\n3. Spectral Positional Encoding (SPE): By embedding structural information from the Tanner graph’s Laplacian eigenspace, SPE enriches the model’s positional representation, enhancing the decoding ability without increasing runtime.\n\nUsing these components, the authors demonstrate a 90% compression ratio and reduce energy consumption by over 224 times compared to ECCT, approaching the efficiency of traditional algorithms like Belief Propagation (BP), making practical deployment possible.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper is well written and easy to follow, and the contributions are relevant.\n2. The techniques introduced demonstrate significant compression without performance degradation compared to ECCTs.\n3. The technique is tested on 3 different ECC types - Polar, Low-Density Parity Check (LDPC), and Bose–Chaudhuri–Hocquenghem (BCH) codes - demonstrating robustness and applicability across various error correction scenarios.\n4. The authors conduct ablation studies evaluating the impact of each of the three components.", "weaknesses": "1. Two out of three proposed techniques (HPSA and SPE) may have limited applicability outside of ECC.\n2. It is not clear how easy it will be to implement these techniques in existing hardware. Ternary quantization has been explored by previous work. Beyond quantization, does HPSA and SPE introduce additional implementation complexities on existing hardware? \n3. The energy efficiency numbers are estimates, the paper does not provide any real measured numbers on actual hardware.", "questions": "1. Discuss any potential challenges or modifications needed to implement HPSA and SPE on current hardware platforms used for error correction.\n2. Compare the implementation complexity of HPSA and SPE to that of the original ECCT approach.\n3. If possible, provide estimates or examples of how these techniques might impact hardware design or utilization in practical ECC systems.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730697776514}, {"id": "92XOQ39uo1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_mhUK"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper addresses the challenges of high computational complexity and memory requirements faced by Error Correction Code Transformers (ECCT) in practical applications by introducing an innovative acceleration method. The approach centers on three technical innovations: \n- a specially designed ternary weight quantization method (AAP) that enables multiplication-free linear layers.\n- an optimized self-attention mechanism (HPSA) based on code-aware multi-head processing that significantly reduces computational complexity.\n- a positional encoding scheme (SPE) through Tanner graph eigendecomposition that provides richer graph connectivity representation without affecting inference runtime. \n\nExperimental results demonstrate that this method not only matches or surpasses the original ECCT's performance but also achieves a 90% compression ratio while reducing arithmetic operation energy consumption by at least 224 times on modern hardware, making transformer-based error correction more practical in resource-constrained environments.", "review_text": "This paper addresses the challenges of high computational complexity and memory requirements faced by Error Correction Code Transformers (ECCT) in practical applications by introducing an innovative acceleration method. The approach centers on three technical innovations: \n- a specially designed ternary weight quantization method (AAP) that enables multiplication-free linear layers.\n- an optimized self-attention mechanism (HPSA) based on code-aware multi-head processing that significantly reduces computational complexity.\n- a positional encoding scheme (SPE) through Tanner graph eigendecomposition that provides richer graph connectivity representation without affecting inference runtime. \n\nExperimental results demonstrate that this method not only matches or surpasses the original ECCT's performance but also achieves a 90% compression ratio while reducing arithmetic operation energy consumption by at least 224 times on modern hardware, making transformer-based error correction more practical in resource-constrained environments.", "strengths": "This paper presents contributions to enhancing ECCT's practical applicability through a comprehensive optimization approach. The key strengths lie in its multi-faceted architectural improvements and efficient implementation strategies.\n\n1. The paper introduces three well-designed technical innovations that directly address ECCT's structural limitations. The Head Partitioning Self Attention (HPSA) mechanism optimizes the self-attention computation specifically for bipartite graph message passing, while the Spectral Positional Encoding (SPE) leverages Tanner graph eigendecomposition to provide richer structural information without runtime overhead. These architectural improvements demonstrate a deep understanding of ECCT's underlying structure and successfully enhance its performance.\n\n2. Except for two structural improvements of ECCT, the paper makes a step further to explore the possibility of low-bit quantization of ECCT. The paper uses a ternary weight quantization method (AAP) specifically designed for ECCT's compact architecture, achieving multiplication-free linear layers while maintaining model accuracy. \n\nThe comprehensive experimental validation demonstrates that these improvements maintain ECCT's decoding performance while reducing its computational complexity to a level comparable with traditional Belief Propagation methods. This work contributes to making transformer-based error correction more practical in resource-constrained environments and helps narrow the efficiency gap between neural decoding techniques and traditional algorithms.", "weaknesses": "The weaknesses of this work are manifested in the following three aspects:\n\n1. Limited Scope and Generalization.\nThe method's exclusive focus on ECCT architecture raises concerns about its broader impact. As ECCT has not gained significant attention in academia nor found applications in industry, optimizations specific to this architecture may have limited practical value.\n\n2. Lack of Technical Novelty in Quantization.\nThe proposed AAP quantization method, while showing good performance, primarily combines existing techniques - weight distribution analysis and learnable quantization steps. The approach does not present substantial innovation in quantization methodology, making it more of an implementation enhancement than a theoretical advancement.\n\n3. Insufficient Validation of Acceleration Claims.\nDespite positioning itself as an acceleration method, the paper's experimental validation concentrates on compression ratios, sparsity metrics, and energy consumption reduction. While these metrics demonstrate improved efficiency, the absence of direct inference speed measurements and real-world acceleration benchmarks leaves the actual acceleration effects unverified. This gap between claimed acceleration and demonstrated results weakens the paper's central premise.\n\nSuggestions:\n\n1. Enhanced Validation of Quantization Method\n\nThe paper would benefit from more comprehensive experiments comparing the proposed AAP method with existing quantization approaches. Specifically, comparative studies with methods like GPTQ and OmniQuant that focus solely on weight distribution or learnable scales would better demonstrate the advantages of combining these techniques. Such experiments would strengthen the paper's claims about AAP's effectiveness and innovation in quantization methodology.\n\n2. Demonstration of Actual Acceleration Effects\n\nThe paper needs to provide concrete evidence of inference speedup to support its acceleration claims. We suggest adding experiments that measure and analyze the actual inference acceleration achieved after implementing all proposed optimizations. If practical acceleration is currently limited by implementation constraints, this limitation should be explicitly acknowledged, and the path toward achieving actual speedup in future work should be discussed. This transparency would better align the paper's claims with its demonstrated results.", "questions": "1.In table2, why are the results of ECCT+AAP better than AECCT for datasets BCH and LDPC? Does it mean HPSA and SPE bring degradation for quantized ECCT? \n\n2.Are the energy consumption saving rate(224 times on 7nm chips and 139 times on 45nm chips) tested on real chips, or just an estimation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the challenges of high computational complexity and memory requirements faced by Error Correction Code Transformers (ECCT) in practical applications by introducing an innovative acceleration method. The approach centers on three technical innovations: \n- a specially designed ternary weight quantization method (AAP) that enables multiplication-free linear layers.\n- an optimized self-attention mechanism (HPSA) based on code-aware multi-head processing that significantly reduces computational complexity.\n- a positional encoding scheme (SPE) through Tanner graph eigendecomposition that provides richer graph connectivity representation without affecting inference runtime. \n\nExperimental results demonstrate that this method not only matches or surpasses the original ECCT's performance but also achieves a 90% compression ratio while reducing arithmetic operation energy consumption by at least 224 times on modern hardware, making transformer-based error correction more practical in resource-constrained environments.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "This paper presents contributions to enhancing ECCT's practical applicability through a comprehensive optimization approach. The key strengths lie in its multi-faceted architectural improvements and efficient implementation strategies.\n\n1. The paper introduces three well-designed technical innovations that directly address ECCT's structural limitations. The Head Partitioning Self Attention (HPSA) mechanism optimizes the self-attention computation specifically for bipartite graph message passing, while the Spectral Positional Encoding (SPE) leverages Tanner graph eigendecomposition to provide richer structural information without runtime overhead. These architectural improvements demonstrate a deep understanding of ECCT's underlying structure and successfully enhance its performance.\n\n2. Except for two structural improvements of ECCT, the paper makes a step further to explore the possibility of low-bit quantization of ECCT. The paper uses a ternary weight quantization method (AAP) specifically designed for ECCT's compact architecture, achieving multiplication-free linear layers while maintaining model accuracy. \n\nThe comprehensive experimental validation demonstrates that these improvements maintain ECCT's decoding performance while reducing its computational complexity to a level comparable with traditional Belief Propagation methods. This work contributes to making transformer-based error correction more practical in resource-constrained environments and helps narrow the efficiency gap between neural decoding techniques and traditional algorithms.", "weaknesses": "The weaknesses of this work are manifested in the following three aspects:\n\n1. Limited Scope and Generalization.\nThe method's exclusive focus on ECCT architecture raises concerns about its broader impact. As ECCT has not gained significant attention in academia nor found applications in industry, optimizations specific to this architecture may have limited practical value.\n\n2. Lack of Technical Novelty in Quantization.\nThe proposed AAP quantization method, while showing good performance, primarily combines existing techniques - weight distribution analysis and learnable quantization steps. The approach does not present substantial innovation in quantization methodology, making it more of an implementation enhancement than a theoretical advancement.\n\n3. Insufficient Validation of Acceleration Claims.\nDespite positioning itself as an acceleration method, the paper's experimental validation concentrates on compression ratios, sparsity metrics, and energy consumption reduction. While these metrics demonstrate improved efficiency, the absence of direct inference speed measurements and real-world acceleration benchmarks leaves the actual acceleration effects unverified. This gap between claimed acceleration and demonstrated results weakens the paper's central premise.\n\nSuggestions:\n\n1. Enhanced Validation of Quantization Method\n\nThe paper would benefit from more comprehensive experiments comparing the proposed AAP method with existing quantization approaches. Specifically, comparative studies with methods like GPTQ and OmniQuant that focus solely on weight distribution or learnable scales would better demonstrate the advantages of combining these techniques. Such experiments would strengthen the paper's claims about AAP's effectiveness and innovation in quantization methodology.\n\n2. Demonstration of Actual Acceleration Effects\n\nThe paper needs to provide concrete evidence of inference speedup to support its acceleration claims. We suggest adding experiments that measure and analyze the actual inference acceleration achieved after implementing all proposed optimizations. If practical acceleration is currently limited by implementation constraints, this limitation should be explicitly acknowledged, and the path toward achieving actual speedup in future work should be discussed. This transparency would better align the paper's claims with its demonstrated results.", "questions": "1.In table2, why are the results of ECCT+AAP better than AECCT for datasets BCH and LDPC? Does it mean HPSA and SPE bring degradation for quantized ECCT? \n\n2.Are the energy consumption saving rate(224 times on 7nm chips and 139 times on 45nm chips) tested on real chips, or just an estimation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730646455568}, {"id": "Kwuk1opoqV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_L3we"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper introduces a novel transformer based neural error correction architecture. It builds on the idea of the Error Correction Code Transformer (ECCT) by performing ternary quantization of the parameters in linear layers, by modifying the masking mechanism in the attention operation and also by introducing positional embeddings. The architecture is empirically validated against the original ECCT and the more traditional Belief Propagation algorithm and is shown to outperform both in terms of accuracy.", "review_text": "This paper introduces a novel transformer based neural error correction architecture. It builds on the idea of the Error Correction Code Transformer (ECCT) by performing ternary quantization of the parameters in linear layers, by modifying the masking mechanism in the attention operation and also by introducing positional embeddings. The architecture is empirically validated against the original ECCT and the more traditional Belief Propagation algorithm and is shown to outperform both in terms of accuracy.", "strengths": "Through it's experimental results, the paper convincingly demonstrates the benefit of ternary weight quantization, Spectral Position Embedding and tailoring the attention mask for bipartite graph message passing over the previous state of the art transformer based method - the error correction code transformer (ECCT).", "weaknesses": "As an outsider to the field of error correction, I found that the description of the problem domain was a bit too brief. The paper would benefit from expanding the problem setting section. It would also benefit the paragraph describing the ECCT architecture if the equations weren't inlined and that the sequence of transformations defining the architecture were serialized.", "questions": "Merely out of curiosity: have the authors considered non transformer based architectures for this task? Presumably denoising diffusion models might have a role to play given the nature of the problem?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel transformer based neural error correction architecture. It builds on the idea of the Error Correction Code Transformer (ECCT) by performing ternary quantization of the parameters in linear layers, by modifying the masking mechanism in the attention operation and also by introducing positional embeddings. The architecture is empirically validated against the original ECCT and the more traditional Belief Propagation algorithm and is shown to outperform both in terms of accuracy.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Through it's experimental results, the paper convincingly demonstrates the benefit of ternary weight quantization, Spectral Position Embedding and tailoring the attention mask for bipartite graph message passing over the previous state of the art transformer based method - the error correction code transformer (ECCT).", "weaknesses": "As an outsider to the field of error correction, I found that the description of the problem domain was a bit too brief. The paper would benefit from expanding the problem setting section. It would also benefit the paragraph describing the ECCT architecture if the equations weren't inlined and that the sequence of transformations defining the architecture were serialized.", "questions": "Merely out of curiosity: have the authors considered non transformer based architectures for this task? Presumably denoising diffusion models might have a role to play given the nature of the problem?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730516181576}, {"id": "b9TCrkACxz", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_3y8a"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "The paper has three main thrusts that I can see, and it targets these towards specifically, acceleration and compression of small transformer model architectures used for error-correction code decoding in communications systems. This is an incremental but significant upgrade to earlier work on ECC transformer decoders. The main thrusts are:\n1) Quantization: The paper uses an adaptive layer-based quantization scheme that combines maximum-based quantization with distribution-based quantization schemes to provide a ternary quantization for the network, greatly accelerating its computation (claimed) while also compressing model size.\n2) Head Partitioning: Masking attention heads based on the Tanner graph node connections at varying levels of relation (first- and second-ring masks) allows greater sparsity in attention computation while also preserving performance.\n3) Positional Encoding: A soft positional encoding is also introduced as an inductive bias in the tokenizer/learnable encoder (from what I can see) that further preserves signal positional information and compensates for information loss in the attention head masking.", "review_text": "The paper has three main thrusts that I can see, and it targets these towards specifically, acceleration and compression of small transformer model architectures used for error-correction code decoding in communications systems. This is an incremental but significant upgrade to earlier work on ECC transformer decoders. The main thrusts are:\n1) Quantization: The paper uses an adaptive layer-based quantization scheme that combines maximum-based quantization with distribution-based quantization schemes to provide a ternary quantization for the network, greatly accelerating its computation (claimed) while also compressing model size.\n2) Head Partitioning: Masking attention heads based on the Tanner graph node connections at varying levels of relation (first- and second-ring masks) allows greater sparsity in attention computation while also preserving performance.\n3) Positional Encoding: A soft positional encoding is also introduced as an inductive bias in the tokenizer/learnable encoder (from what I can see) that further preserves signal positional information and compensates for information loss in the attention head masking.", "strengths": "1) Solid presentation, all major thrusts are clearly presented and detailed. The paper also provides diagrams and examples to ease understanding and reading in what can be presented as a very dense, mathematical topic that a reviewer would have trouble grasping or reading through.\n2) Analysis and experimentation include results on hardware for energy consumption reduction, a key metric in model compression. It also includes ablation studies for different encoding methods and acceleration due to model compression. All in all it analyzes most of the ablations relevant to the topic. I particularly commend the hardware results for different transistor nodes.\n3) Contributions while building clearly off of a prior design and prior work, are cleverly tuned to the problem domain and a very clear niche. In particular the use of the Tanner graph and selective masking shows adaptation of the transformer compression problem to the given domain.", "weaknesses": "1) Concluding statements mention that the paper shows a general approach for the compression problem in small transformers - I would say that this is a bit of a stretch - the Tanner subgraph masking and the adjustments to positional encoding are not applicable to all problem domains, leaving the model quantization formulation the only really general cross-domain contribution. That formulation has not been benchmarking in the experimental section against prior art, and its acceleration alone (as opposed to the entire framework) has not been tested in the main document. I would say that ablation results for each component of the framework should be added in order to make this general claim.\n\n2) A key contribution is replacing GeLU with ReLU in the bullet points of the front matter. This is absolutely not novel, as has been examined in [1] just to name a single source. I would strongly advise omitting this claim or qualifying it to the given problem domain.\n[1] https://arxiv.org/abs/2310.04564 [Addressed]\n\n3) While this work builds on the prior ECCT work, I would want a presentation of the model architecture in the main draft and clearly marked so that the scale of the 'small transformer' is visible, rather than having to dig through the ECCT paper and assume that this is the paper used. [Addressed]\n\n4) The work assumes dedicated hardware for the transformer decoder in Section 5 when calculating complexity. This may not be a realistic assumption in practice due to the specialized and expensive design process for accelerator ASICs. I would want to see complexity calculations on a common edge accelerator platform such as an NVIDIA Hopper architecture or a Coral edge TPU, or a citation for the architecture-level simulator presumably used to generate these numbers (e.g. ScaleSim). Numbers for hardware without that backing are unreliable.\n\n5)  The sparsity induced by masking is structured, and therefore may be applicable to standard accelerators. Why is there no mention of this? I do not see an analysis of the benefits of the structured sparsity from the masks in Section 4.2 compared to unstructured sparsity, which is a major plus point that has not been exploited.", "questions": "1) See point 5 in Weaknesses - is the sparsity structured, and if so, does your assumed hardware architecture exploit that?\n\n2) See (4) in Weaknesses - Where is this assumption coming from and what edge platform architecture was used? Is there a citation or a cycle-accurate simulation of the architecture? Analytical complexity numbers that assume optimal hardware are unrealistic.\n\n3) What are the authors' thoughts on the contribution of each component of this framework to the overall acceleration? Is there an ablation study that can be presented in the main doc? If not, a small table and rough numbers would be good so that we can see how much general approaches like quantization contribute and how much the domain-specific tweaks such as Tanner graph masking contribute.\n\nSee discussion - the authors have conditionally addressed much of the questions above and paper weaknesses. I am editing the review to reflect this.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper has three main thrusts that I can see, and it targets these towards specifically, acceleration and compression of small transformer model architectures used for error-correction code decoding in communications systems. This is an incremental but significant upgrade to earlier work on ECC transformer decoders. The main thrusts are:\n1) Quantization: The paper uses an adaptive layer-based quantization scheme that combines maximum-based quantization with distribution-based quantization schemes to provide a ternary quantization for the network, greatly accelerating its computation (claimed) while also compressing model size.\n2) Head Partitioning: Masking attention heads based on the Tanner graph node connections at varying levels of relation (first- and second-ring masks) allows greater sparsity in attention computation while also preserving performance.\n3) Positional Encoding: A soft positional encoding is also introduced as an inductive bias in the tokenizer/learnable encoder (from what I can see) that further preserves signal positional information and compensates for information loss in the attention head masking.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1) Solid presentation, all major thrusts are clearly presented and detailed. The paper also provides diagrams and examples to ease understanding and reading in what can be presented as a very dense, mathematical topic that a reviewer would have trouble grasping or reading through.\n2) Analysis and experimentation include results on hardware for energy consumption reduction, a key metric in model compression. It also includes ablation studies for different encoding methods and acceleration due to model compression. All in all it analyzes most of the ablations relevant to the topic. I particularly commend the hardware results for different transistor nodes.\n3) Contributions while building clearly off of a prior design and prior work, are cleverly tuned to the problem domain and a very clear niche. In particular the use of the Tanner graph and selective masking shows adaptation of the transformer compression problem to the given domain.", "weaknesses": "1) Concluding statements mention that the paper shows a general approach for the compression problem in small transformers - I would say that this is a bit of a stretch - the Tanner subgraph masking and the adjustments to positional encoding are not applicable to all problem domains, leaving the model quantization formulation the only really general cross-domain contribution. That formulation has not been benchmarking in the experimental section against prior art, and its acceleration alone (as opposed to the entire framework) has not been tested in the main document. I would say that ablation results for each component of the framework should be added in order to make this general claim.\n\n2) A key contribution is replacing GeLU with ReLU in the bullet points of the front matter. This is absolutely not novel, as has been examined in [1] just to name a single source. I would strongly advise omitting this claim or qualifying it to the given problem domain.\n[1] https://arxiv.org/abs/2310.04564 [Addressed]\n\n3) While this work builds on the prior ECCT work, I would want a presentation of the model architecture in the main draft and clearly marked so that the scale of the 'small transformer' is visible, rather than having to dig through the ECCT paper and assume that this is the paper used. [Addressed]\n\n4) The work assumes dedicated hardware for the transformer decoder in Section 5 when calculating complexity. This may not be a realistic assumption in practice due to the specialized and expensive design process for accelerator ASICs. I would want to see complexity calculations on a common edge accelerator platform such as an NVIDIA Hopper architecture or a Coral edge TPU, or a citation for the architecture-level simulator presumably used to generate these numbers (e.g. ScaleSim). Numbers for hardware without that backing are unreliable.\n\n5)  The sparsity induced by masking is structured, and therefore may be applicable to standard accelerators. Why is there no mention of this? I do not see an analysis of the benefits of the structured sparsity from the masks in Section 4.2 compared to unstructured sparsity, which is a major plus point that has not been exploited.", "questions": "1) See point 5 in Weaknesses - is the sparsity structured, and if so, does your assumed hardware architecture exploit that?\n\n2) See (4) in Weaknesses - Where is this assumption coming from and what edge platform architecture was used? Is there a citation or a cycle-accurate simulation of the architecture? Analytical complexity numbers that assume optimal hardware are unrealistic.\n\n3) What are the authors' thoughts on the contribution of each component of this framework to the overall acceleration? Is there an ablation study that can be presented in the main doc? If not, a small table and rough numbers would be good so that we can see how much general approaches like quantization contribute and how much the domain-specific tweaks such as Tanner graph masking contribute.\n\nSee discussion - the authors have conditionally addressed much of the questions above and paper weaknesses. I am editing the review to reflect this.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730502588085}, {"id": "C6Ys8NssAQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3573/Reviewer_mcns"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper presents a method to enhance the efficiency and practicality of transformer-based error correction code (ECC) decoders, specifically targeting the error correction code transformer (ECCT) model introduced in prior work. The authors propose several modifications to ECCTs, including ternary weight quantization, a novel attention mechanism called head partitioning self-attention (HPSA), and a spectral positional encoding inspired by the Tanner graph's eigenspace. \nJointly, these improvements lead to reduce memory, energy, and computational requirements while maintaining or improving the performance of the original ECCTs.", "review_text": "This paper presents a method to enhance the efficiency and practicality of transformer-based error correction code (ECC) decoders, specifically targeting the error correction code transformer (ECCT) model introduced in prior work. The authors propose several modifications to ECCTs, including ternary weight quantization, a novel attention mechanism called head partitioning self-attention (HPSA), and a spectral positional encoding inspired by the Tanner graph's eigenspace. \nJointly, these improvements lead to reduce memory, energy, and computational requirements while maintaining or improving the performance of the original ECCTs.", "strengths": "The paper proposes interesting methods like HPSA and adaptive absolute percentile quantization, which could be valuable outside of coding applications.", "weaknesses": "While the proposed approach introduces certain novel elements, I believe the paper falls short in critical areas, particularly regarding contribution scope and experimental results. Therefore, I recommend rejection for the following reasons:\n\n1. Limited contribution: The paper proposes interesting methods like HPSA and adaptive absolute percentile quantization, which could be valuable outside of coding applications. However, since this paper is fundamentally a paper in coding theory, these contributions remain peripheral to the main focus (furthermore, per se they are not enough to warrant publication at ICLR).\nAnd within the realm of coding theory, the contribution of the paper is limited. \n\nWhile relevant for the coding community, the contribution lacks the depth and broader impact expected at a major machine learning conference or in a top coding journal. A more appropriate fit would be a coding conference such as ISIT or ITW.\n\nFurthermore, although ML-based decoding of conventional codes is a growing research area, with already many published papers, this paper does not provide a method that competes with conventional state-of-the-art codes/decoders. \n\n\n2. Weak comparisons and benchmarking: The experimental results are insufficient to substantiate the relevance of the proposed decoder.\n\nSpecifically, the authors compare the performance of their transformer-based decoder to that of belief propagation decoding for various codes, including polar codes and BCH codes. BP decoding of BCH codes and polar codes is known to be highly suboptimal. There are well-known, more efficient decoding algorithms for these codes (note that the considered BCH codes are very short). Hence, the current results provide no hint on whether the proposed decoders are good or not. \n\nTo establish relevance, the paper would need to demonstrate comparable or superior performance to state-of-the-art decoding techniques in terms of accuracy or complexity. However, even if such comparisons were provided, in my opinion the contribution is still insufficient for a high-profile venue.\n\n3. Use of non-standard metrics: The paper presents performance using the negative natural logarithm of the BER. Is there any reason for not adhering to standard metrics such as bit error rate curves or block error rate (BLER) curves (the latter being preferable)?  \nThe use of this unconventional (to day the least) metric complicates unnecessarily the interpretation of the results. I do not think it is wise to come up with an awkward way of showing results, when there is a established way to present results for error correcting codes.\n\nFurthermore, reporting performance for only a few specific values of $E_b/N_0$ provides an incomplete view; it's unclear whether phenomena like error floors might occur at higher values.\n\nIncluding conventional BLER curves over a range of $E_b/N_0$ values would make the results easier to interpret and align with community standards. Without such curves, it's challenging to assess the effectiveness of the proposed decoder and, in my opinion, the paper should be accepted in any venue.", "questions": "1. As mentioned above, the authors should report BLER curves.\n\n2. The authors provide results for a number of LDPC codes. However, there are no details regarding these codes. Are them the best-available LDPC codes in the literature, or some designed by the authors? \n\n3. Related to Comment 2 and my comments in \"Weaknesses,\" the authors should rework their results section and compare the performance of the proposed approach with that of state-of-the-art codes/decoders", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a method to enhance the efficiency and practicality of transformer-based error correction code (ECC) decoders, specifically targeting the error correction code transformer (ECCT) model introduced in prior work. The authors propose several modifications to ECCTs, including ternary weight quantization, a novel attention mechanism called head partitioning self-attention (HPSA), and a spectral positional encoding inspired by the Tanner graph's eigenspace. \nJointly, these improvements lead to reduce memory, energy, and computational requirements while maintaining or improving the performance of the original ECCTs.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The paper proposes interesting methods like HPSA and adaptive absolute percentile quantization, which could be valuable outside of coding applications.", "weaknesses": "While the proposed approach introduces certain novel elements, I believe the paper falls short in critical areas, particularly regarding contribution scope and experimental results. Therefore, I recommend rejection for the following reasons:\n\n1. Limited contribution: The paper proposes interesting methods like HPSA and adaptive absolute percentile quantization, which could be valuable outside of coding applications. However, since this paper is fundamentally a paper in coding theory, these contributions remain peripheral to the main focus (furthermore, per se they are not enough to warrant publication at ICLR).\nAnd within the realm of coding theory, the contribution of the paper is limited. \n\nWhile relevant for the coding community, the contribution lacks the depth and broader impact expected at a major machine learning conference or in a top coding journal. A more appropriate fit would be a coding conference such as ISIT or ITW.\n\nFurthermore, although ML-based decoding of conventional codes is a growing research area, with already many published papers, this paper does not provide a method that competes with conventional state-of-the-art codes/decoders. \n\n\n2. Weak comparisons and benchmarking: The experimental results are insufficient to substantiate the relevance of the proposed decoder.\n\nSpecifically, the authors compare the performance of their transformer-based decoder to that of belief propagation decoding for various codes, including polar codes and BCH codes. BP decoding of BCH codes and polar codes is known to be highly suboptimal. There are well-known, more efficient decoding algorithms for these codes (note that the considered BCH codes are very short). Hence, the current results provide no hint on whether the proposed decoders are good or not. \n\nTo establish relevance, the paper would need to demonstrate comparable or superior performance to state-of-the-art decoding techniques in terms of accuracy or complexity. However, even if such comparisons were provided, in my opinion the contribution is still insufficient for a high-profile venue.\n\n3. Use of non-standard metrics: The paper presents performance using the negative natural logarithm of the BER. Is there any reason for not adhering to standard metrics such as bit error rate curves or block error rate (BLER) curves (the latter being preferable)?  \nThe use of this unconventional (to day the least) metric complicates unnecessarily the interpretation of the results. I do not think it is wise to come up with an awkward way of showing results, when there is a established way to present results for error correcting codes.\n\nFurthermore, reporting performance for only a few specific values of $E_b/N_0$ provides an incomplete view; it's unclear whether phenomena like error floors might occur at higher values.\n\nIncluding conventional BLER curves over a range of $E_b/N_0$ values would make the results easier to interpret and align with community standards. Without such curves, it's challenging to assess the effectiveness of the proposed decoder and, in my opinion, the paper should be accepted in any venue.", "questions": "1. As mentioned above, the authors should report BLER curves.\n\n2. The authors provide results for a number of LDPC codes. However, there are no details regarding these codes. Are them the best-available LDPC codes in the literature, or some designed by the authors? \n\n3. Related to Comment 2 and my comments in \"Weaknesses,\" the authors should rework their results section and compare the performance of the proposed approach with that of state-of-the-art codes/decoders", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730388675666}], "openreview_url": "https://openreview.net/forum?id=lAXlDAdan5", "arxiv_id": "2410.05911", "paper_pdf": "papers/lAXlDAdan5.pdf", "paper_pdf_sha256": "521ebc1740dc0f78a95b3742ca1a5e88013508acbf748a9022196483c13eacc7", "paper_pdf_bytes": 2465131, "paper_pdf_source": "openreview", "code_url": "https://github.com/mlaetvayn/AECCT", "code_repository": "mlaetvayn/AECCT", "code_commit": "d66d5dc9312bde0484c126eba0a1dd857fcd2418", "code_archive": "repos/lAXlDAdan5.zip", "code_archive_sha256": "4457907076cdc95c4a02315d7a8ec7a510ec7ee532a3c5c40bec084d086239eb", "code_archive_bytes": 17429, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 16, "github_languages": {"Python": 32650}, "github_archived": false, "github_pushed_at": "2024-10-16T09:16:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerating-error-correction-code"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YkEW5TabYN", "year": 2024, "status": "rejected", "title": "Perturbed examples reveal invariances shared by language models", "authors": ["Ruchit Rawal", "Mariya Toneva"], "authorids": ["~Ruchit_Rawal1", "~Mariya_Toneva1"], "authors_source": "OpenReview API", "abstract": "An explosion of work in language is leading to ever-increasing numbers of available natural language processing models, with little understanding of how new models compare to better-understood models. One major reason for this difficulty is saturating benchmark datasets, which may not reflect well differences in model performance in the wild. In this work, we propose a novel framework for comparing two natural language processing models by revealing their shared invariance to interpretable input perturbations that are designed to target a specific linguistic capability (e.g., Synonym-Invariance, Typo-Invariance). Via experiments on models from within the same and across different architecture families, this framework offers a number of insights about how changes in models (e.g. distillation, increase in size, amount of pre-training) affect multiple well-defined linguistic capabilities. Furthermore, we also demonstrate how our framework can enable evaluation of the invariances shared between models that are available as commercial black-box APIs (e.g., InstructGPT family) and models that are relatively better understood (e.g., GPT-2). Across several experiments, we observe that large language models share many of the invariances encoded by models of various sizes, whereas the invariances encoded by large language models are only shared by other large models. Possessing a wide variety of invariances may be a key reason for the recent successes of large language models, and our framework can shed light on the types of invariances that are retained by or emerge in new models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "o1z5EB6ftX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7449/Reviewer_yXpb"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper studies invariances in different semantic or linguistic features across LLM representations. The authors propose a novel metric study for evaluating the invariance in LLMs. Paper includes extensive experiments across metrics and models. It would be a good fit for the conference and informative contribution.", "review_text": "The paper studies invariances in different semantic or linguistic features across LLM representations. The authors propose a novel metric study for evaluating the invariance in LLMs. Paper includes extensive experiments across metrics and models. It would be a good fit for the conference and informative contribution.", "strengths": "The paper includes extensive comparison across existing metrics and prominent LLMs to help understand their capabilities.", "weaknesses": "Description of experimental methodology needs to be clarified.", "questions": "IID not defined.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies invariances in different semantic or linguistic features across LLM representations. The authors propose a novel metric study for evaluating the invariance in LLMs. Paper includes extensive experiments across metrics and models. It would be a good fit for the conference and informative contribution.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper includes extensive comparison across existing metrics and prominent LLMs to help understand their capabilities.", "weaknesses": "Description of experimental methodology needs to be clarified.", "questions": "IID not defined.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698856697391}, {"id": "5mUcxpK8yZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7449/Reviewer_cey2"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces SCoPE, a measure of invariance between language models (LMs). Unlike naive pointwise-agreement measures (that simply measure the fraction of inputs on which two models agree), SCoPE measures the fraction of robust (perturbation-robust) inputs on which models differ. As the perturbations are capability-defined, SCoPE is a measure targeted at a particular capability. Details follow.\n\nGiven two datasets $X$ and $X'$, a label set $Y$, a model $m\\colon X \\cup X' \\to \\mathbb{R}^n$ and a loss function $\\ell \\colon \\mathbb{R}^n \\times \\mathbb{R}^n \\to \\mathbb{R}$, we define a _perturbation_ $p\\colon X \\to X'$ by mapping each $x \\in X$ to $x'\\in X'$ that minimizes $\\ell(m(x), m(x'))$. [Note: In the paper it there is a linguistic capability perturbing the inputs, though I am unclear on how this is captured formally by the definition and so cannot include it in my summary.] Moving forward, we will denote by $y_m(x)$ the label of $x$ as predicted by $m$, that is, $y_m(x) = \\mathrm{argmax}_{k \\in [n]}m(x)_k$.\n\nGiven an input perturbation $x \\to x'$ and two models $m_1$ and $m_2$, the Hard-SCoPE distance from $m_1$ to $m_2$ is defined by $\\Pr_{x \\in X}[y_{m_2}(x) = y_{m_2}(p(x)) \\vert y_{m_1}(x) = y_{m_1}(p(x))]$. In plain words, we can say that Hard-SCoPE measures the degree to which inputs on which $m_1$ is perturbation-invariant, are also perturbation-invariant for $m_2$. \n\nA variant of Hard-SCoPE, called Soft-SCoPE, is defined to be $f(m_1, m_2) \\cdot \\mathrm{Hard-SCoPE}$. The authors define $f(m_1, m_2) = \\mathrm{decay}(\\mathrm(dist)(\\Delta \\vec{m}_1, \\Delta \\vec{m}_2))$, however $\\Delta\\vec{m}$ is not defined and so I do not know how Soft-SCoPE is defined.\n\n[Note: I did not understand the formal notation in the paper, so applied my own here to the best of my ability.]\n\nThe authors evaluate the utility of their measures by conducting various experiments. The perturbations are either synonym-based (\"A man laughs out loud\" $\\to$ \"A man laughs out loudly\") or typo-based (a permutation of all but the first and last characters of a single word). Results are:\n\n**BERT finetuned on sentiment analysis (Section 4)**\n1. Between BERT (as the reference model $m_1$) and DistilBERT (as the target model $m_2$), SCoPE measures are higher for synonym-perturbations vs. typo-perturbations.\n2. When comparing BERT-Tiny (as reference model) with synonym-perturbations to five BERT variants of different sizes, Hard-SCoPE _increases_ with model size, and Soft-SCope stays roughly the same. This is despite the fact that pointwise agreement on the perturbed dataset (\"OOD Agreement\") correlates with model size, i.e., follows the opposite trend of Hard-SCoPE.\n3. When repeating the previous experiment with BERT-Base as a reference model, Hard-SCoPE and Soft-SCoPE now _do correlate_ with model size.\n\n**GPT-2 vs. InstructGPT (Section 5)**\nIn the following, GPT-2 is used as the reference model against variants of InstructGPT\n1. `text-davinci-001` has higher Hard and Soft-SCoPE than its smaller variants `text-ada-001`, `text-curie-001` and `text-babbage-001`.\n2. `text-davinci-003` has higher Hard and Soft-SCoPE than `text-davinci-001` and `text-davinci-002`.", "review_text": "This paper introduces SCoPE, a measure of invariance between language models (LMs). Unlike naive pointwise-agreement measures (that simply measure the fraction of inputs on which two models agree), SCoPE measures the fraction of robust (perturbation-robust) inputs on which models differ. As the perturbations are capability-defined, SCoPE is a measure targeted at a particular capability. Details follow.\n\nGiven two datasets $X$ and $X'$, a label set $Y$, a model $m\\colon X \\cup X' \\to \\mathbb{R}^n$ and a loss function $\\ell \\colon \\mathbb{R}^n \\times \\mathbb{R}^n \\to \\mathbb{R}$, we define a _perturbation_ $p\\colon X \\to X'$ by mapping each $x \\in X$ to $x'\\in X'$ that minimizes $\\ell(m(x), m(x'))$. [Note: In the paper it there is a linguistic capability perturbing the inputs, though I am unclear on how this is captured formally by the definition and so cannot include it in my summary.] Moving forward, we will denote by $y_m(x)$ the label of $x$ as predicted by $m$, that is, $y_m(x) = \\mathrm{argmax}_{k \\in [n]}m(x)_k$.\n\nGiven an input perturbation $x \\to x'$ and two models $m_1$ and $m_2$, the Hard-SCoPE distance from $m_1$ to $m_2$ is defined by $\\Pr_{x \\in X}[y_{m_2}(x) = y_{m_2}(p(x)) \\vert y_{m_1}(x) = y_{m_1}(p(x))]$. In plain words, we can say that Hard-SCoPE measures the degree to which inputs on which $m_1$ is perturbation-invariant, are also perturbation-invariant for $m_2$. \n\nA variant of Hard-SCoPE, called Soft-SCoPE, is defined to be $f(m_1, m_2) \\cdot \\mathrm{Hard-SCoPE}$. The authors define $f(m_1, m_2) = \\mathrm{decay}(\\mathrm(dist)(\\Delta \\vec{m}_1, \\Delta \\vec{m}_2))$, however $\\Delta\\vec{m}$ is not defined and so I do not know how Soft-SCoPE is defined.\n\n[Note: I did not understand the formal notation in the paper, so applied my own here to the best of my ability.]\n\nThe authors evaluate the utility of their measures by conducting various experiments. The perturbations are either synonym-based (\"A man laughs out loud\" $\\to$ \"A man laughs out loudly\") or typo-based (a permutation of all but the first and last characters of a single word). Results are:\n\n**BERT finetuned on sentiment analysis (Section 4)**\n1. Between BERT (as the reference model $m_1$) and DistilBERT (as the target model $m_2$), SCoPE measures are higher for synonym-perturbations vs. typo-perturbations.\n2. When comparing BERT-Tiny (as reference model) with synonym-perturbations to five BERT variants of different sizes, Hard-SCoPE _increases_ with model size, and Soft-SCope stays roughly the same. This is despite the fact that pointwise agreement on the perturbed dataset (\"OOD Agreement\") correlates with model size, i.e., follows the opposite trend of Hard-SCoPE.\n3. When repeating the previous experiment with BERT-Base as a reference model, Hard-SCoPE and Soft-SCoPE now _do correlate_ with model size.\n\n**GPT-2 vs. InstructGPT (Section 5)**\nIn the following, GPT-2 is used as the reference model against variants of InstructGPT\n1. `text-davinci-001` has higher Hard and Soft-SCoPE than its smaller variants `text-ada-001`, `text-curie-001` and `text-babbage-001`.\n2. `text-davinci-003` has higher Hard and Soft-SCoPE than `text-davinci-001` and `text-davinci-002`.", "strengths": "- The idea of measuring similarity of models by evaluating their agreement on perturbation-invariant inputs is, to my knowledge, novel.\n- As opposed to new benchmarks or tasks, this measure can be computed based only on outputs (or logits) of the model, and is generic enough that it can be instantiated on any dataset (and its perturbation).\n- The choice of models for experimentation provides an appealing setting, namely, understanding the effect of model size on robustness to perturbations.", "weaknesses": "### Significant Issues with SCoPE definition\nThere are numerous issues with the presentation of the definition of Hard-SCoPE and Soft-SCoPE. Considering that the new measure is the main contribution made in this paper, this is a significant weakness of this work. I could not attempt to infer the definition from its implementation, since the submission was not supplemented by code.\n\n1. The notation $\\Delta \\vec{m}$ is never formally defined. Consequently, it is impossible to understand the definition of Soft-SCoPE.\n2. In Hard-SCoPE and Soft-SCoPE, the expectation is said to be taken over $(x, x') \\in (X, X')$. The domain $(X, X')$ is not defined. My first interpretation was that this was a cartesian product $X \\times X'$, which means that $x, x'$ are taken iid--but this does not seem to align with the textual description of SCoPE. See my summary for an alternative way to denote taking $x$ and its corresponding perturbation.\n3. In general, I do not follow (or do not agree with) the definition of a perturbation in this paper. For example, in page 4 you define $X'$ to be the image of $X$ under $C(\\cdot, m)$. But this does not retain the mapping of inputs $x$ to their perturbations $C(x, m)$. Furthermore, $C$ is not well-defined, since the domain of the argmin is not specified. If we take the domain to be $X$, and read the follwing paragram in which you say that you take the loss function to be $\\ell_1$-norm difference, then we have that $C$ ends up being the identity function (assuming the loss has no non-trivial zeroes).\n4. The expectation of an indicator random variable is simply the probability over this event. It would be significantly more readable to use probabilities in the notation, as I did in may summary.\n5. In Soft-SCoPE, since the attenuating factor $\\mathrm{decay}(\\mathrm{dist}(\\cdot, \\cdot))$ seems to be constant with respect to the random variables, it can be taken out of the expectation. Therefore, Soft-SCoPE is just Hard-SCoPE multiplied by a factor that depends on $m_1$ and $m_2$ (in an undefined way, see item 1 above).\n6. In page 6 it is written that \"Hard-SCoPE can only take binary values i.e., 0 or 1\". My understanding of this sentence is that the Hard-SCoPE of two models can be either 0, or 1. But this is not consistent with the remainder of the text.\n\n### Experiments lack depth\nThe goal of the experiments should be to argue for the utility of the novel measure introduced in this paper. This can be done, for example, by using this measure to obtain novel and interesting insight, or by dissecting the measure with fine-grained experiments.\n\nIn section 3, the authors compare the measure to naive pointwise measure (IID/OOD agreement). The main body of experiments (Sections 4 and 5) are conducted by taking different pairs of models and measuring Soft-SCoPE, Hard-SCoPE, and pointwise agreement. The experiments are then displayed with bar charts, and there is a textual description of the results. I found the results to be fairly superficial: the correlations are either expected (BERT-Base SCoPE correlates with model size), or uninterpretable (why is this not the case for BERT-Tiny?). When deeper conclusions are reached for, it is at times done in a speculative manner (\"we hypothesize\").\n\nI unfortunately do not have concrete suggestions for a more refined experimental setup that will give more insightful results, in part because I could not understand the definition of the measure itself.\n\n### Choice of \"capabilities\"\nIn the body of the paper, the perturbations are based on synonyms or permutations of characters in a word. I found the usage of the term \"capability\" to be confusing in this context. In NLP, I think of a capability as a task, such as sentiment analysis or summarization. Are these perturbations somehow related to capabilities as they are more generally used?\n\nAdditionally, were the choices of these pertubations arbitrary, or is there something about them that makes the particularly appropriate for exploring the newly-defined SCoPE measure? It seems to me that the results in Sections 4 and 5 could have been reported out all the same with other perturbations---this speaks to the genericness of the results (see previous weakness).\n\n### Figure 5\nI could not understand the diagram in the left part of Figure 5, despite making several attempts on different days. The diagram on the right would be clearer if presented by a table, as there are onalya three data points.\n\n### Additional / minor issues\n- Commas before / after \"i.e.\" are inconsistent throughout the paper. There are different style guides for American vs. British English regarding this matter, but please pick a dialect and stick to it.\n- Page 3: What is the argmin over? Presumably, over $x' \\in X'$, but then $X'$ uses $C$ in its definition in the following page! \n- Page 4: \"and constraint modifications of words that are stopwords\". Did you mean \"constrain\" (a verb) rather than \"constraint\" (a noun)?\n- Page 4: There is a grammatical issue with the sentence starting with \"We perform experiments along one...\"\n- Page 4: What is the argmax over? Presumably, over $k \\in [c]$. It would be more readable to make this equation centered rather than inline.\n- Page 4: The definition of $X'$ should use set builder notation: replace the $\\forall$ with a colon or a vertical line.\n- Page 5: In the definition of the acronaym you write **SH**ared-**C**-apabilities... But the **H** does not appear in the acronym. Furthermore, there is no need to hyphenate the definition.\n- Page 5: \"(ref Eq. 1)\". Is the \"ref\" a typo? There is no such common shorthand, to my knowledge. Use \"see\" instead.\n- Page 6: Use \\mathrm{} for operators such as \"decay\" and \"dist\".\n- Page 6: I suggest using the term \"monotonically decreasing\" rather than \"has a downward slope\".\n- Page 7: In the first paragraph, there are several issues with citations and parentheses. There should probably not be sequences of parentheses such as \"))\" or \")(\".", "questions": "I welcome responses and answers to any of the questions raised in the previous sections of my review. Besides these, I have two outstanding questions.\n\n### What is the cost of computing SCoPE?\nI could not find the computational costs reported in the paper or in the appendices. Since this paper proposes a new measure, it would be useful for the reader to know how expensive it is to compute. In particular, for the experiments in Section 5, the authors evaluate black-box models by taking random samples from their output; what is the monetary barrier to reproducing these experiments? Is there a compute-accuracy tradeoff inherent to this method?\n\n### What is behavioral about the invariances?\nThe paper makes extensive references to \"behavioral invariances\". What is \"behavioral\" about these invariances? Is this term used in other parts of the literature, to distinguish thes from other kinds of invariances? \"Behavior\" alludes to some \"cognitive\" or grounded aspect of the model, or at the very least to its semantics, while typo-invariance is clearly a syntactic phenomenon. I think that using just \"invariance\" would be clearer.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SCoPE, a measure of invariance between language models (LMs). Unlike naive pointwise-agreement measures (that simply measure the fraction of inputs on which two models agree), SCoPE measures the fraction of robust (perturbation-robust) inputs on which models differ. As the perturbations are capability-defined, SCoPE is a measure targeted at a particular capability. Details follow.\n\nGiven two datasets $X$ and $X'$, a label set $Y$, a model $m\\colon X \\cup X' \\to \\mathbb{R}^n$ and a loss function $\\ell \\colon \\mathbb{R}^n \\times \\mathbb{R}^n \\to \\mathbb{R}$, we define a _perturbation_ $p\\colon X \\to X'$ by mapping each $x \\in X$ to $x'\\in X'$ that minimizes $\\ell(m(x), m(x'))$. [Note: In the paper it there is a linguistic capability perturbing the inputs, though I am unclear on how this is captured formally by the definition and so cannot include it in my summary.] Moving forward, we will denote by $y_m(x)$ the label of $x$ as predicted by $m$, that is, $y_m(x) = \\mathrm{argmax}_{k \\in [n]}m(x)_k$.\n\nGiven an input perturbation $x \\to x'$ and two models $m_1$ and $m_2$, the Hard-SCoPE distance from $m_1$ to $m_2$ is defined by $\\Pr_{x \\in X}[y_{m_2}(x) = y_{m_2}(p(x)) \\vert y_{m_1}(x) = y_{m_1}(p(x))]$. In plain words, we can say that Hard-SCoPE measures the degree to which inputs on which $m_1$ is perturbation-invariant, are also perturbation-invariant for $m_2$. \n\nA variant of Hard-SCoPE, called Soft-SCoPE, is defined to be $f(m_1, m_2) \\cdot \\mathrm{Hard-SCoPE}$. The authors define $f(m_1, m_2) = \\mathrm{decay}(\\mathrm(dist)(\\Delta \\vec{m}_1, \\Delta \\vec{m}_2))$, however $\\Delta\\vec{m}$ is not defined and so I do not know how Soft-SCoPE is defined.\n\n[Note: I did not understand the formal notation in the paper, so applied my own here to the best of my ability.]\n\nThe authors evaluate the utility of their measures by conducting various experiments. The perturbations are either synonym-based (\"A man laughs out loud\" $\\to$ \"A man laughs out loudly\") or typo-based (a permutation of all but the first and last characters of a single word). Results are:\n\n**BERT finetuned on sentiment analysis (Section 4)**\n1. Between BERT (as the reference model $m_1$) and DistilBERT (as the target model $m_2$), SCoPE measures are higher for synonym-perturbations vs. typo-perturbations.\n2. When comparing BERT-Tiny (as reference model) with synonym-perturbations to five BERT variants of different sizes, Hard-SCoPE _increases_ with model size, and Soft-SCope stays roughly the same. This is despite the fact that pointwise agreement on the perturbed dataset (\"OOD Agreement\") correlates with model size, i.e., follows the opposite trend of Hard-SCoPE.\n3. When repeating the previous experiment with BERT-Base as a reference model, Hard-SCoPE and Soft-SCoPE now _do correlate_ with model size.\n\n**GPT-2 vs. InstructGPT (Section 5)**\nIn the following, GPT-2 is used as the reference model against variants of InstructGPT\n1. `text-davinci-001` has higher Hard and Soft-SCoPE than its smaller variants `text-ada-001`, `text-curie-001` and `text-babbage-001`.\n2. `text-davinci-003` has higher Hard and Soft-SCoPE than `text-davinci-001` and `text-davinci-002`.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "- The idea of measuring similarity of models by evaluating their agreement on perturbation-invariant inputs is, to my knowledge, novel.\n- As opposed to new benchmarks or tasks, this measure can be computed based only on outputs (or logits) of the model, and is generic enough that it can be instantiated on any dataset (and its perturbation).\n- The choice of models for experimentation provides an appealing setting, namely, understanding the effect of model size on robustness to perturbations.", "weaknesses": "### Significant Issues with SCoPE definition\nThere are numerous issues with the presentation of the definition of Hard-SCoPE and Soft-SCoPE. Considering that the new measure is the main contribution made in this paper, this is a significant weakness of this work. I could not attempt to infer the definition from its implementation, since the submission was not supplemented by code.\n\n1. The notation $\\Delta \\vec{m}$ is never formally defined. Consequently, it is impossible to understand the definition of Soft-SCoPE.\n2. In Hard-SCoPE and Soft-SCoPE, the expectation is said to be taken over $(x, x') \\in (X, X')$. The domain $(X, X')$ is not defined. My first interpretation was that this was a cartesian product $X \\times X'$, which means that $x, x'$ are taken iid--but this does not seem to align with the textual description of SCoPE. See my summary for an alternative way to denote taking $x$ and its corresponding perturbation.\n3. In general, I do not follow (or do not agree with) the definition of a perturbation in this paper. For example, in page 4 you define $X'$ to be the image of $X$ under $C(\\cdot, m)$. But this does not retain the mapping of inputs $x$ to their perturbations $C(x, m)$. Furthermore, $C$ is not well-defined, since the domain of the argmin is not specified. If we take the domain to be $X$, and read the follwing paragram in which you say that you take the loss function to be $\\ell_1$-norm difference, then we have that $C$ ends up being the identity function (assuming the loss has no non-trivial zeroes).\n4. The expectation of an indicator random variable is simply the probability over this event. It would be significantly more readable to use probabilities in the notation, as I did in may summary.\n5. In Soft-SCoPE, since the attenuating factor $\\mathrm{decay}(\\mathrm{dist}(\\cdot, \\cdot))$ seems to be constant with respect to the random variables, it can be taken out of the expectation. Therefore, Soft-SCoPE is just Hard-SCoPE multiplied by a factor that depends on $m_1$ and $m_2$ (in an undefined way, see item 1 above).\n6. In page 6 it is written that \"Hard-SCoPE can only take binary values i.e., 0 or 1\". My understanding of this sentence is that the Hard-SCoPE of two models can be either 0, or 1. But this is not consistent with the remainder of the text.\n\n### Experiments lack depth\nThe goal of the experiments should be to argue for the utility of the novel measure introduced in this paper. This can be done, for example, by using this measure to obtain novel and interesting insight, or by dissecting the measure with fine-grained experiments.\n\nIn section 3, the authors compare the measure to naive pointwise measure (IID/OOD agreement). The main body of experiments (Sections 4 and 5) are conducted by taking different pairs of models and measuring Soft-SCoPE, Hard-SCoPE, and pointwise agreement. The experiments are then displayed with bar charts, and there is a textual description of the results. I found the results to be fairly superficial: the correlations are either expected (BERT-Base SCoPE correlates with model size), or uninterpretable (why is this not the case for BERT-Tiny?). When deeper conclusions are reached for, it is at times done in a speculative manner (\"we hypothesize\").\n\nI unfortunately do not have concrete suggestions for a more refined experimental setup that will give more insightful results, in part because I could not understand the definition of the measure itself.\n\n### Choice of \"capabilities\"\nIn the body of the paper, the perturbations are based on synonyms or permutations of characters in a word. I found the usage of the term \"capability\" to be confusing in this context. In NLP, I think of a capability as a task, such as sentiment analysis or summarization. Are these perturbations somehow related to capabilities as they are more generally used?\n\nAdditionally, were the choices of these pertubations arbitrary, or is there something about them that makes the particularly appropriate for exploring the newly-defined SCoPE measure? It seems to me that the results in Sections 4 and 5 could have been reported out all the same with other perturbations---this speaks to the genericness of the results (see previous weakness).\n\n### Figure 5\nI could not understand the diagram in the left part of Figure 5, despite making several attempts on different days. The diagram on the right would be clearer if presented by a table, as there are onalya three data points.\n\n### Additional / minor issues\n- Commas before / after \"i.e.\" are inconsistent throughout the paper. There are different style guides for American vs. British English regarding this matter, but please pick a dialect and stick to it.\n- Page 3: What is the argmin over? Presumably, over $x' \\in X'$, but then $X'$ uses $C$ in its definition in the following page! \n- Page 4: \"and constraint modifications of words that are stopwords\". Did you mean \"constrain\" (a verb) rather than \"constraint\" (a noun)?\n- Page 4: There is a grammatical issue with the sentence starting with \"We perform experiments along one...\"\n- Page 4: What is the argmax over? Presumably, over $k \\in [c]$. It would be more readable to make this equation centered rather than inline.\n- Page 4: The definition of $X'$ should use set builder notation: replace the $\\forall$ with a colon or a vertical line.\n- Page 5: In the definition of the acronaym you write **SH**ared-**C**-apabilities... But the **H** does not appear in the acronym. Furthermore, there is no need to hyphenate the definition.\n- Page 5: \"(ref Eq. 1)\". Is the \"ref\" a typo? There is no such common shorthand, to my knowledge. Use \"see\" instead.\n- Page 6: Use \\mathrm{} for operators such as \"decay\" and \"dist\".\n- Page 6: I suggest using the term \"monotonically decreasing\" rather than \"has a downward slope\".\n- Page 7: In the first paragraph, there are several issues with citations and parentheses. There should probably not be sequences of parentheses such as \"))\" or \")(\".", "questions": "I welcome responses and answers to any of the questions raised in the previous sections of my review. Besides these, I have two outstanding questions.\n\n### What is the cost of computing SCoPE?\nI could not find the computational costs reported in the paper or in the appendices. Since this paper proposes a new measure, it would be useful for the reader to know how expensive it is to compute. In particular, for the experiments in Section 5, the authors evaluate black-box models by taking random samples from their output; what is the monetary barrier to reproducing these experiments? Is there a compute-accuracy tradeoff inherent to this method?\n\n### What is behavioral about the invariances?\nThe paper makes extensive references to \"behavioral invariances\". What is \"behavioral\" about these invariances? Is this term used in other parts of the literature, to distinguish thes from other kinds of invariances? \"Behavior\" alludes to some \"cognitive\" or grounded aspect of the model, or at the very least to its semantics, while typo-invariance is clearly a syntactic phenomenon. I think that using just \"invariance\" would be clearer.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698782062365}, {"id": "QE6ADlzjRH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7449/Reviewer_xMdc"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes invariance-based metrics for capturing similarities between NLP models. Here, invariance-based metrics are contrasted with agreement-based metrics. To compute the proposed invariance-based metrics, one chooses a reference model (Ref) and constructs a dataset of (x, x’) invariant pairs. Here, x is the “base”, and x’ is the “perturbed” example, which is chosen as a perturbation of x that yields minimal change to Ref outputs. Two perturbation mechanisms that yield x’ candidates are considered, namely inserting typos and replacing words with synonyms. Given the (x, x’) pairs from Ref and a Target model (Tgt), one computes the similarity between how perturbing x to x’ changes Tgt behavior. This is contrasted to OOD agreement between Ref and Tgt, which would involve directly comparing the models’ predictions Ref(x’) and Tgt(x’) to each other. Instead, in the proposed invariance-based metrics Tgt(x’) is compared to Tgt(x). The paper proposes a hard and a soft metric, that are called Hard- SCoPE and Soft-SCoPE respectively. \n\nThe key message of the paper is that Soft-SCoPE sometimes does not correlate with OOD agreement. From this the paper infers that invariance-based metrics can offer complementary insights that agreement-based metrics may miss. For example, the paper shows how distilling BERT into Distill-BERT considerably decreases the Soft-SCoPE metric.", "review_text": "The paper proposes invariance-based metrics for capturing similarities between NLP models. Here, invariance-based metrics are contrasted with agreement-based metrics. To compute the proposed invariance-based metrics, one chooses a reference model (Ref) and constructs a dataset of (x, x’) invariant pairs. Here, x is the “base”, and x’ is the “perturbed” example, which is chosen as a perturbation of x that yields minimal change to Ref outputs. Two perturbation mechanisms that yield x’ candidates are considered, namely inserting typos and replacing words with synonyms. Given the (x, x’) pairs from Ref and a Target model (Tgt), one computes the similarity between how perturbing x to x’ changes Tgt behavior. This is contrasted to OOD agreement between Ref and Tgt, which would involve directly comparing the models’ predictions Ref(x’) and Tgt(x’) to each other. Instead, in the proposed invariance-based metrics Tgt(x’) is compared to Tgt(x). The paper proposes a hard and a soft metric, that are called Hard- SCoPE and Soft-SCoPE respectively. \n\nThe key message of the paper is that Soft-SCoPE sometimes does not correlate with OOD agreement. From this the paper infers that invariance-based metrics can offer complementary insights that agreement-based metrics may miss. For example, the paper shows how distilling BERT into Distill-BERT considerably decreases the Soft-SCoPE metric.", "strengths": "The paper aspires to address an important challenge of quantifying similarities and differences between different language models.", "weaknesses": "In my opinion, the paper fails to present a clear motivation and justification for the proposed metrics. For an auxiliary metric to be of interest to a practitioner, it needs to be predictive of what the practitioner cares about. In particular, the abstract of the paper refers to learning about “differences in model performance in the wild” as the motivation. I did not find an explanation of how measuring the proposed invariance-based metrics for pairs of models can help us get a better idea of the models’ “performance in the wild”. The main justification for the proposed metrics is that they don’t correlate with OOD agreement. But one can’t justify a new metric just by saying it’s different from existing ones - it is a logical fallacy. \n\nI also found the paper rather difficult to follow. Here are some concerns as bullet items:\n- Is a notation for the source of perturbed candidates x’ missing from Equation 1?\n- A table showing examples of (x, x’) pairs is a must for a paper like this. \n- Many complex sentences and reasoning chains make the paper hard to follow. Example (page 5): “Thus, measuring behavioral shared invariances amounts to quantifying the extent to which a target model is invariant on perturbations that do not cause any change in the reference model’s behavior (ref Eq. 1).”", "questions": "Why would a practitioner trust your metrics more than OOD-agreement to predict e.g. how much BERT distillation hurts the model’s robustness to typos?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes invariance-based metrics for capturing similarities between NLP models. Here, invariance-based metrics are contrasted with agreement-based metrics. To compute the proposed invariance-based metrics, one chooses a reference model (Ref) and constructs a dataset of (x, x’) invariant pairs. Here, x is the “base”, and x’ is the “perturbed” example, which is chosen as a perturbation of x that yields minimal change to Ref outputs. Two perturbation mechanisms that yield x’ candidates are considered, namely inserting typos and replacing words with synonyms. Given the (x, x’) pairs from Ref and a Target model (Tgt), one computes the similarity between how perturbing x to x’ changes Tgt behavior. This is contrasted to OOD agreement between Ref and Tgt, which would involve directly comparing the models’ predictions Ref(x’) and Tgt(x’) to each other. Instead, in the proposed invariance-based metrics Tgt(x’) is compared to Tgt(x). The paper proposes a hard and a soft metric, that are called Hard- SCoPE and Soft-SCoPE respectively. \n\nThe key message of the paper is that Soft-SCoPE sometimes does not correlate with OOD agreement. From this the paper infers that invariance-based metrics can offer complementary insights that agreement-based metrics may miss. For example, the paper shows how distilling BERT into Distill-BERT considerably decreases the Soft-SCoPE metric.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "The paper aspires to address an important challenge of quantifying similarities and differences between different language models.", "weaknesses": "In my opinion, the paper fails to present a clear motivation and justification for the proposed metrics. For an auxiliary metric to be of interest to a practitioner, it needs to be predictive of what the practitioner cares about. In particular, the abstract of the paper refers to learning about “differences in model performance in the wild” as the motivation. I did not find an explanation of how measuring the proposed invariance-based metrics for pairs of models can help us get a better idea of the models’ “performance in the wild”. The main justification for the proposed metrics is that they don’t correlate with OOD agreement. But one can’t justify a new metric just by saying it’s different from existing ones - it is a logical fallacy. \n\nI also found the paper rather difficult to follow. Here are some concerns as bullet items:\n- Is a notation for the source of perturbed candidates x’ missing from Equation 1?\n- A table showing examples of (x, x’) pairs is a must for a paper like this. \n- Many complex sentences and reasoning chains make the paper hard to follow. Example (page 5): “Thus, measuring behavioral shared invariances amounts to quantifying the extent to which a target model is invariant on perturbations that do not cause any change in the reference model’s behavior (ref Eq. 1).”", "questions": "Why would a practitioner trust your metrics more than OOD-agreement to predict e.g. how much BERT distillation hurts the model’s robustness to typos?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698717793701}, {"id": "FBcioE8uG4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7449/Reviewer_nq6A"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces a new framework to measure the difference behavior between two models, with one of them being a reference model and the other one being the target model. Unlike previous work which uses benchmark performance or model prediction agreement, this framework proposes to analyze the shared invariance of the two models when they are given an original input vs. pertubed input which targets specific lingusitic capabilities (i.e., synonym-invariance and typo-invariance).\n\nExperiment results reveal several key findings that (1) the performance gap between models (existing metrics) doesn't always reflect the shared behaviors (invariance) between them, especially where a smaller model is used as a reference model and that (2) larger models tend to share more invariances. The proposed framework closes the first gap by measuring the shared behaviors (invariance) between models.", "review_text": "The paper introduces a new framework to measure the difference behavior between two models, with one of them being a reference model and the other one being the target model. Unlike previous work which uses benchmark performance or model prediction agreement, this framework proposes to analyze the shared invariance of the two models when they are given an original input vs. pertubed input which targets specific lingusitic capabilities (i.e., synonym-invariance and typo-invariance).\n\nExperiment results reveal several key findings that (1) the performance gap between models (existing metrics) doesn't always reflect the shared behaviors (invariance) between them, especially where a smaller model is used as a reference model and that (2) larger models tend to share more invariances. The proposed framework closes the first gap by measuring the shared behaviors (invariance) between models.", "strengths": "- With the growing popularity of large language models (LLMs) and lots of efforts to make them more efficient, I think the proposed framework would be useful especially for making sure that a particular iteration of a model will have similar or improved behavior compared to the base model. From the analysis, the framework can provide complementary analysis beyond existing metrics.\n- Well-executed experiments, with three types of pertubation, and comparison with existing metrics from previous work.\n- Application for both specific model (fine-tuned BERT) and more recent generative models (GPT variants).", "weaknesses": "- Presentation: some of the details that are important seems to be missing (or rather put in the Appendix). See more details below.", "questions": "- Section 3: \"four independent components\" -> I am not sure if these four are really independent, does transformation and constraint dependent on each other?\n- I'm a bit puzzled with the results of BERT-Tiny in Fig. 4, especially regarding the poor OOD-agreement but high invariances. Does this mean that the predictions of both models are quite different (hence poor agreement), but within the individual model the predictions do not change when it is given original and pertubed samples?\n- Is there any interesting findings on the model agreement on original input vs. pertubed input?\n\n**Suggestions for presentation improvements**\n- Mention the tasks being used in the experiments at the beginning of the paper. It was not clear until Section 3.3 that you focus on text classification and language modeling.\n- Would be helpful to add more details regarding how the pertubed input is constructed for each linguistic capability. It is currently mentioned as the limitation (inefficient search methods), but i is unclear how inefficient since there is no explanation about it.\n- Notation definition: more explicit in assigning $m_1$ and $m_2$, currently it only mentions $m_1$ as a reference model and I need to re-read in multiple place to infer what is $m_2$.\n- I think the results for fairness capability is quite interesting. If there is space, might consider to put it in the main text.\n- A relevant paper: https://arxiv.org/pdf/1711.02173.pdf", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new framework to measure the difference behavior between two models, with one of them being a reference model and the other one being the target model. Unlike previous work which uses benchmark performance or model prediction agreement, this framework proposes to analyze the shared invariance of the two models when they are given an original input vs. pertubed input which targets specific lingusitic capabilities (i.e., synonym-invariance and typo-invariance).\n\nExperiment results reveal several key findings that (1) the performance gap between models (existing metrics) doesn't always reflect the shared behaviors (invariance) between them, especially where a smaller model is used as a reference model and that (2) larger models tend to share more invariances. The proposed framework closes the first gap by measuring the shared behaviors (invariance) between models.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "- With the growing popularity of large language models (LLMs) and lots of efforts to make them more efficient, I think the proposed framework would be useful especially for making sure that a particular iteration of a model will have similar or improved behavior compared to the base model. From the analysis, the framework can provide complementary analysis beyond existing metrics.\n- Well-executed experiments, with three types of pertubation, and comparison with existing metrics from previous work.\n- Application for both specific model (fine-tuned BERT) and more recent generative models (GPT variants).", "weaknesses": "- Presentation: some of the details that are important seems to be missing (or rather put in the Appendix). See more details below.", "questions": "- Section 3: \"four independent components\" -> I am not sure if these four are really independent, does transformation and constraint dependent on each other?\n- I'm a bit puzzled with the results of BERT-Tiny in Fig. 4, especially regarding the poor OOD-agreement but high invariances. Does this mean that the predictions of both models are quite different (hence poor agreement), but within the individual model the predictions do not change when it is given original and pertubed samples?\n- Is there any interesting findings on the model agreement on original input vs. pertubed input?\n\n**Suggestions for presentation improvements**\n- Mention the tasks being used in the experiments at the beginning of the paper. It was not clear until Section 3.3 that you focus on text classification and language modeling.\n- Would be helpful to add more details regarding how the pertubed input is constructed for each linguistic capability. It is currently mentioned as the limitation (inefficient search methods), but i is unclear how inefficient since there is no explanation about it.\n- Notation definition: more explicit in assigning $m_1$ and $m_2$, currently it only mentions $m_1$ as a reference model and I need to re-read in multiple place to infer what is $m_2$.\n- I think the results for fairness capability is quite interesting. If there is space, might consider to put it in the main text.\n- A relevant paper: https://arxiv.org/pdf/1711.02173.pdf", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698708564235}], "openreview_url": "https://openreview.net/forum?id=YkEW5TabYN", "arxiv_id": "2311.04166", "paper_pdf": "papers/YkEW5TabYN.pdf", "paper_pdf_sha256": "f61900781619f6513465cbf29a7de809e536e6853c69af1c4a73fcf82e6d072e", "paper_pdf_bytes": 1219604, "paper_pdf_source": "openreview", "code_url": "https://github.com/bridge-ai-neuro/shared_invariances_acl", "code_repository": "bridge-ai-neuro/shared_invariances_acl", "code_commit": "f2615b69f213931488074d842514cef4ee139327", "code_archive": "repos/YkEW5TabYN.zip", "code_archive_sha256": "77c3c5ec76609da9bca135acb87c8580e7f0ef043a45bc89d0635c2857a56e55", "code_archive_bytes": 20792, "code_file_count": 14, "code_extensions": {".py": 11, ".sh": 3}, "github_disk_usage_kb": 16, "github_languages": {"Python": 48260, "Shell": 1750}, "github_archived": false, "github_pushed_at": "2024-06-09T10:05:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/perturbed-examples-reveal-invariances-shared"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "CEhy-i7_KfC", "year": 2023, "status": "rejected", "title": "Pretraining the Vision Transformer using self-supervised methods for vision based Deep Reinforcement Learning", "authors": ["Manuel Goulão", "Arlindo L. Oliveira"], "authorids": ["~Manuel_Goulão1", "~Arlindo_L._Oliveira1"], "authors_source": "OpenReview API", "abstract": "The Vision Transformer architecture has shown to be competitive in the computer vision (CV) space where it has dethroned convolution-based networks in several benchmarks. Nevertheless, Convolutional Neural Networks (CNN) remain the preferential architecture for the representation module in Reinforcement Learning. In this work, we study pretraining a Vision Transformer using several state-of-the-art self-supervised methods and assess data-efficiency gains from this training framework. We propose a new self-supervised learning method called TOV-VICReg that extends VICReg to better capture temporal relations between observations by adding a temporal order verification task. Furthermore, we evaluate the resultant encoders with Atari games in a sample-efficiency regime. Our results show that the vision transformer, when pretrained with TOV-VICReg, outperforms the other self-supervised methods but still struggles to overcome a CNN. Nevertheless, we were able to outperform a CNN in two of the ten games where we perform a 100k steps evaluation. Ultimately, we believe that such approaches in Deep Reinforcement Learning (DRL) might be the key to achieving new levels of performance as seen in natural language processing and computer vision.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "JiL4Dwri0y", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3071/Reviewer_rKFV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the effect of self-supervised pre-training for Vision Transformer based RL agents. It shows the effect of temporal order verification and VICReg on 10 Atari games. The proposed methods improves final return on Atari games.", "review_text": "The submission is limited in terms of both theoretical novelty and empirical novelty. Especially, the empirical validity is pretty limited due to small set of experiments, missing experiment details, etc. More work should be done to warrant the publication of this work.", "strengths": "[Strengths]\n- The paper is easy to follow.\n- The proposed method makes intuitive sense to me.\n\n[Weaknesses]\n- Most importantly, the main contribution of the submission is not very clear to me.\n- Also, the technical novelty and empirical novelty is quite limited. TOV and VICReg are existing self-supervised learning methods. Also, it is well-known that the use of pre-training is helpful for vision-based reinforcement learning.\n- The analysis in Section 7 (representation collapse, dimension collapse, etc) are very generic and do not provide much insight on using self-supervised learning for reinforcement learning.\n- It seems that the CNN result is without self-supervised training and ViT result is with self-supervised training (although it is unclear to me). Can authors add more on this?\n- I don't see the learning curve of return vs environment steps. This is pretty important for demonstrating sample efficiency in RL.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper studies the effect of self-supervised pre-training for Vision Transformer based RL agents. It shows the effect of temporal order verification and VICReg on 10 Atari games. The proposed methods improves final return on Atari games.", "strength_and_weaknesses": "[Strengths]\n- The paper is easy to follow.\n- The proposed method makes intuitive sense to me.\n\n[Weaknesses]\n- Most importantly, the main contribution of the submission is not very clear to me.\n- Also, the technical novelty and empirical novelty is quite limited. TOV and VICReg are existing self-supervised learning methods. Also, it is well-known that the use of pre-training is helpful for vision-based reinforcement learning.\n- The analysis in Section 7 (representation collapse, dimension collapse, etc) are very generic and do not provide much insight on using self-supervised learning for reinforcement learning.\n- It seems that the CNN result is without self-supervised training and ViT result is with self-supervised training (although it is unclear to me). Can authors add more on this?\n- I don't see the learning curve of return vs environment steps. This is pretty important for demonstrating sample efficiency in RL.", "clarity,_quality,_novelty_and_reproducibility": "- The core contribution is not very clear to me.\n- Regarding the use of self-supervised learning for RL, \"CURL: Contrastive Unsupervised Representations for Reinforcement Learning\" is one of the work that popularized this. The authors should cite and compare against this work.\n- In light of the existing work, more work should be done to advance the field. Just trying self-supervised learning on reinforcement learning is not sufficient for publication at ICLR.\n- The contribution may be the use of self-supervised learning on ViT based image RL. The proposed method does not contain modifications for ViTs: TOV and VICReg are very generic methods, and they can be applied to any deep nets that take images (regardless of CNN vs ViT, RL vs supervised learning problems). Can authors provide why the proposed method is significantly novel or overcomes challenges in using self-supervised learning for RL?\n- Also, could authors provide the results of applying the proposed methods on CNNs? If the CNN performance can't be improved, why so?", "summary_of_the_review": "The submission is limited in terms of both theoretical novelty and empirical novelty. Especially, the empirical validity is pretty limited due to small set of experiments, missing experiment details, etc. More work should be done to warrant the publication of this work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667290663501}, {"id": "ULVxcWrhNb", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3071/Reviewer_9PJR"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper mainly targets usage of ViTs in Reinforcement Learning. The authors conduct several experiments with different training strategies for ViTs, showing the effectiveness of self-supervised learning for ViTs in RL. Besides, the authors propose a new self-supervised objective function based on VICReg, which involves the temporal order prediction for learning better temporal relations.", "review_text": "The authors present good technical analysis of using ViTs for RL problems. The novelty and significance are limited.", "strengths": "Strength:\n1. The authors present interesting and trials with ViTs for RL.\n\nWeaknesses:\n1. The novelty is limited. The proposed method is made up of several previous methods.\n2. Even though the proposed TOV-VICReg is more effective than the other self-supervised methods, it is still only comparable to the CNN based models with no significant advantage. In other words, the ViT based models remain impractical for RL.\n3. It would be better if the authors can consider reconstruction based self-supervision like MAE and SimMIM in their experiments, which adopts a quite different objective compared with DINO and MoCo.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper mainly targets usage of ViTs in Reinforcement Learning. The authors conduct several experiments with different training strategies for ViTs, showing the effectiveness of self-supervised learning for ViTs in RL. Besides, the authors propose a new self-supervised objective function based on VICReg, which involves the temporal order prediction for learning better temporal relations.", "strength_and_weaknesses": "Strength:\n1. The authors present interesting and trials with ViTs for RL.\n\nWeaknesses:\n1. The novelty is limited. The proposed method is made up of several previous methods.\n2. Even though the proposed TOV-VICReg is more effective than the other self-supervised methods, it is still only comparable to the CNN based models with no significant advantage. In other words, the ViT based models remain impractical for RL.\n3. It would be better if the authors can consider reconstruction based self-supervision like MAE and SimMIM in their experiments, which adopts a quite different objective compared with DINO and MoCo.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is generally clear. The novelty is limited due to no originated method is proposed in this paper. The TOV-VICReg is a combination of two existing methods. The proposed method should be easy to reproduce based on the authors' code.", "summary_of_the_review": "The authors present good technical analysis of using ViTs for RL problems. The novelty and significance are limited.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667197589295}, {"id": "p2TVmXbGBNQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3071/Reviewer_Wuo3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to introduce visual transformers to extract feature representations for reinforcement learning. It proposed a method, which is a combination of VICReg and temporal ordering prediction, to pretrain the ViT network. Experiments are done on Atari 100K using the visual transformer pretrained by the proposed method. ", "review_text": "The paper has a good motivation of using a well-pretrained visual transformer in deep RL. However, the method design seems to be generic and of marginal novelty. The experimental results are a bit confusing and do not seem to be staring enough to address the paper’s motivation. ", "strengths": "** Strength: **\n1. It is reasonable to combine the image only pretraining method (specifically, VICReg) with temporal ordering prediction as the overall objectives covers both image and temporal information. \n2. The paper analyzes the property of representation from a few different aspects in the experiments. \n\n** Weaknesses: **\n1. It is unclear how the pretraining is specific to deep reinforcement learning. It looks like a generic video-based pretraining method, and its connection with reinforcement learning seems a bit loose. \n2. The novelty of the method is limited as combining two established self-supervised learning methods in the same training objective is not something difficult to come up with. \n3. Q: it seems Atari 100K has 26 games (please correct me if I am wrong), why only 10 games are used?\n4. The experimental results are not strong. Especially, using visual transformer does not seem to be that competitive to using CNN, as the paper claimed. \n5. The clarity of the paper (especially in the experiment section) can be improved. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper aims to introduce visual transformers to extract feature representations for reinforcement learning. It proposed a method, which is a combination of VICReg and temporal ordering prediction, to pretrain the ViT network. Experiments are done on Atari 100K using the visual transformer pretrained by the proposed method. ", "strength_and_weaknesses": "** Strength: **\n1. It is reasonable to combine the image only pretraining method (specifically, VICReg) with temporal ordering prediction as the overall objectives covers both image and temporal information. \n2. The paper analyzes the property of representation from a few different aspects in the experiments. \n\n** Weaknesses: **\n1. It is unclear how the pretraining is specific to deep reinforcement learning. It looks like a generic video-based pretraining method, and its connection with reinforcement learning seems a bit loose. \n2. The novelty of the method is limited as combining two established self-supervised learning methods in the same training objective is not something difficult to come up with. \n3. Q: it seems Atari 100K has 26 games (please correct me if I am wrong), why only 10 games are used?\n4. The experimental results are not strong. Especially, using visual transformer does not seem to be that competitive to using CNN, as the paper claimed. \n5. The clarity of the paper (especially in the experiment section) can be improved. ", "clarity,_quality,_novelty_and_reproducibility": "The clarity and writing quality of the paper can be improved. One example, the mode performance analysis in Sec 6.1 is unclear. It is not self-contained enough for one to understand the metric. And it is confusing if the results focus only on “sampling efficiency” or actually looks at the model’s game performance. \n\nThe papers have a few broken links. \n\nWhile it could claim the method is novel for RL, but the specific connection between the proposed method and RL is not clear to me. And a generic method to pretrain a feature extractor on video. Such a combination of image-only objectives and temporal ordering prediction is marginally novel. \n\nThe amount of details in the paper is sufficient for one to understand the main idea but may not be sufficient for one to reproduce the results fully. ", "summary_of_the_review": "The paper has a good motivation of using a well-pretrained visual transformer in deep RL. However, the method design seems to be generic and of marginal novelty. The experimental results are a bit confusing and do not seem to be staring enough to address the paper’s motivation. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667102954550}, {"id": "6hgzpDMvtr", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3071/Reviewer_Wxq5"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "*I am a vision person and an emergency reviewer.*\n\nThis paper proposes to combine VICReg and a temporal verification loss for video representation learning, and inspect its impact on the data-efficiency of down-streaming reinforcement learning tasks. It is shown that the proposed method out-performs VICReg, MOCO and DINO while still under-performing (randomly initialized?) CNNs. Further analyses show that the proposed variant of VICReg shows better properties in terms of collapse and feature distribution.", "review_text": "This is an empirical study yet I am not convinced by the motivation or the significance of empirical results.", "strengths": "Strengths:\n+ The proposed training framework that combines VICReg and temporal verification is new to my knowledge.\n+ Experimental results show that the variant out-performs alternative representation learning methods in the inspected RL setting.\n+ A set of analysis shows that the method shows better statistical properties w.r.t. collapse and feature similarity.\n\nWeaknesses:\n- Most of the experimental results seem not related to reinforcement learning and the only RL evaluation reports inferior performance than (randomly initialized?) CNNs. I recommend the authors to consider generic video classification evaluation protocols in [A], which may be a more promising thing to pursue.\n- The language is fine but the organization is pool IMHO. Too much space is spent on describing standard things like MDP, DQN or VICReg.\n- Generally I am not convinced by the motivation. I think transformers are widely considered less data-efficient and I don't understand why authors expect it to behave the opposite way in reinforcement learning.\n- Typos:\ncomponents that allows us -> allow\nthe increasing interest -> and the increasing interest\nViT presents weaker image-specific inductive biases which allow the CNNs for much sample-efficient learning; allows CNNs?? what does this mean\n\n[A] VideoMoCo: Contrastive video representation learning with temporally adversarial examples, CVPR 2021", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "*I am a vision person and an emergency reviewer.*\n\nThis paper proposes to combine VICReg and a temporal verification loss for video representation learning, and inspect its impact on the data-efficiency of down-streaming reinforcement learning tasks. It is shown that the proposed method out-performs VICReg, MOCO and DINO while still under-performing (randomly initialized?) CNNs. Further analyses show that the proposed variant of VICReg shows better properties in terms of collapse and feature distribution.", "strength_and_weaknesses": "Strengths:\n+ The proposed training framework that combines VICReg and temporal verification is new to my knowledge.\n+ Experimental results show that the variant out-performs alternative representation learning methods in the inspected RL setting.\n+ A set of analysis shows that the method shows better statistical properties w.r.t. collapse and feature similarity.\n\nWeaknesses:\n- Most of the experimental results seem not related to reinforcement learning and the only RL evaluation reports inferior performance than (randomly initialized?) CNNs. I recommend the authors to consider generic video classification evaluation protocols in [A], which may be a more promising thing to pursue.\n- The language is fine but the organization is pool IMHO. Too much space is spent on describing standard things like MDP, DQN or VICReg.\n- Generally I am not convinced by the motivation. I think transformers are widely considered less data-efficient and I don't understand why authors expect it to behave the opposite way in reinforcement learning.\n- Typos:\ncomponents that allows us -> allow\nthe increasing interest -> and the increasing interest\nViT presents weaker image-specific inductive biases which allow the CNNs for much sample-efficient learning; allows CNNs?? what does this mean\n\n[A] VideoMoCo: Contrastive video representation learning with temporally adversarial examples, CVPR 2021", "clarity,_quality,_novelty_and_reproducibility": "Clarity: I am not very sure about the RL protocol. Is the representations trained in an offline setting that pre-stores frames?\nQuality: not good enough for both technical depth or presentation quality.\nNovelty: The training framework is somewhat new.\nReproducibility: Codes are provided.", "summary_of_the_review": "This is an empirical study yet I am not convinced by the motivation or the significance of empirical results.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667062197301}, {"id": "8TKM0fFDJb9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3071/Reviewer_X82N"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies if Vision Transformers (ViTs) could outperform CNNs in vision-based RL tasks. The authors compared existing self-supervised strategies and proposed a new self-supervised approach designed for the sequential observations in RL. The newly proposed method combines previous ideas from VICReg (Bardes et al, 2021) and Shuffle-and-Learn (Misra et al, 2016). The authors show that the new approach TOV-VICReg outperforms other SOTA self-supervised learning methods in the RL setting. However, from the authors experiments, CNNs still outperform ViTs trained with TOV-VICReg in eight out of the ten tested Atari games. ", "review_text": "I’m inclined to reject this paper mainly because: 1) the experiments do not justify the motivation of this work, i.e., use large models to solve complex RL tasks where smaller models struggle, as stated in the conclusion of this paper. Clearly, the chosen environments are not complex enough to demonstrate the potential benefits of pretrained ViTs. 2), the proposed method lack novelty and the results do not bring significant new insights to the community, 3) the paper does properly compare other self-supervised learning methods that aim to capture the temporal relations  4) the paper presentation needs further polishing.  ", "strengths": "### Strength\n\n1. The proposed self-supervised learning method is well suited for RL tasks and experiment results show that it improves performance of self-supervised learning methods that does not capture the temporal relationship.\n\n### Weaknesses\n\n1. This paper lacks novelty in terms of the proposed self-supervised learning method. Also, the presented results are mostly well-known in the community.   \n2. Temporal order verification is one type of self-supervision loss that aims to capture temporal dependency. This paper does not justify the choice of Shuffle-and-Learn against other methods [1, 2, 3]. I’d recommend the authors to add a discussion on relevant self-supervised learning literature for temporal relationship learning and provide a justification on the choice of Shuffle-and-Learn.   \n3. The number of steps used for temporal order verification loss is an important hyperparameter that would affect the quality of the learned representations. This paper does not provide experimental results showing its impact. I’d encourage the authors to add an experiment varying the number of steps.\n\n### References\n\n[1] Lee, Hsin-Ying, et al. \"Unsupervised representation learning by sorting sequences.\" *Proceedings of the IEEE international conference on computer vision*. 2017.\n\n[2] Xu, Dejing, et al. \"Self-supervised spatiotemporal learning via video clip order prediction.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2019.\n\n[3] Yao, Yuan, et al. \"Video playback rate perception for self-supervised spatio-temporal representation learning.\" *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*. 2020.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies if Vision Transformers (ViTs) could outperform CNNs in vision-based RL tasks. The authors compared existing self-supervised strategies and proposed a new self-supervised approach designed for the sequential observations in RL. The newly proposed method combines previous ideas from VICReg (Bardes et al, 2021) and Shuffle-and-Learn (Misra et al, 2016). The authors show that the new approach TOV-VICReg outperforms other SOTA self-supervised learning methods in the RL setting. However, from the authors experiments, CNNs still outperform ViTs trained with TOV-VICReg in eight out of the ten tested Atari games. ", "strength_and_weaknesses": "### Strength\n\n1. The proposed self-supervised learning method is well suited for RL tasks and experiment results show that it improves performance of self-supervised learning methods that does not capture the temporal relationship.\n\n### Weaknesses\n\n1. This paper lacks novelty in terms of the proposed self-supervised learning method. Also, the presented results are mostly well-known in the community.   \n2. Temporal order verification is one type of self-supervision loss that aims to capture temporal dependency. This paper does not justify the choice of Shuffle-and-Learn against other methods [1, 2, 3]. I’d recommend the authors to add a discussion on relevant self-supervised learning literature for temporal relationship learning and provide a justification on the choice of Shuffle-and-Learn.   \n3. The number of steps used for temporal order verification loss is an important hyperparameter that would affect the quality of the learned representations. This paper does not provide experimental results showing its impact. I’d encourage the authors to add an experiment varying the number of steps.\n\n### References\n\n[1] Lee, Hsin-Ying, et al. \"Unsupervised representation learning by sorting sequences.\" *Proceedings of the IEEE international conference on computer vision*. 2017.\n\n[2] Xu, Dejing, et al. \"Self-supervised spatiotemporal learning via video clip order prediction.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2019.\n\n[3] Yao, Yuan, et al. \"Video playback rate perception for self-supervised spatio-temporal representation learning.\" *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*. 2020.", "clarity,_quality,_novelty_and_reproducibility": "### Quality\n\nThis paper should be improved in terms of clairty and experiments before being accepted into ICLR. \n\n### Novelty:\n\nThe paper presents an incremental contribution where ViTs are trained by combining the temporal order verification loss (Misra et al, 2016) and the VICReg loss (Bardes et al, 2021). Both the methods and the results do not present significantly contributions.\n\n### Reproducibility\n\nThe authors provided their code in the supplemental material. \n\n### Clarity\n\n1. Figure 4 and 5 are confusing becuase of the different scales of the color map. \n2. The desription of the baseline methods is not clear. For the SGI ResNet Large, was it pretrained with self-supervision loss? If so, did the pretraining follow the original SGI paper? \n\nMinor Comments:\n\n1. The DQN algorithm details on page 3 can be presented in an appendix and formatted with Latex algorithm packages. \n2. VICReg is a published ICLR 2022 paper. Please update the reference accordingly. \n3. For audience unfamiliar with VICReg, it would be helpful to explain how expanders work.", "summary_of_the_review": "I’m inclined to reject this paper mainly because: 1) the experiments do not justify the motivation of this work, i.e., use large models to solve complex RL tasks where smaller models struggle, as stated in the conclusion of this paper. Clearly, the chosen environments are not complex enough to demonstrate the potential benefits of pretrained ViTs. 2), the proposed method lack novelty and the results do not bring significant new insights to the community, 3) the paper does properly compare other self-supervised learning methods that aim to capture the temporal relations  4) the paper presentation needs further polishing.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666546855438}], "openreview_url": "https://openreview.net/forum?id=CEhy-i7_KfC", "arxiv_id": "2209.10901", "paper_pdf": "papers/CEhy-i7_KfC.pdf", "paper_pdf_sha256": "70eb5a4040f93871f4aca16d008325a6d8980b5e0c8683a4e040275976a0cb68", "paper_pdf_bytes": 393689, "paper_pdf_source": "openreview", "code_url": "https://github.com/mgoulao/TOV-VICReg", "code_repository": "mgoulao/TOV-VICReg", "code_commit": "535b1643134b407ca40ca2c6553e04b50d3968ae", "code_archive": "repos/CEhy-i7_KfC.zip", "code_archive_sha256": "ca3a98c07a3ef237691d3f0301ed573eb8785863038ba5bf1659cc8c10ab99b9", "code_archive_bytes": 24578, "code_file_count": 13, "code_extensions": {".py": 11, ".sh": 2}, "github_disk_usage_kb": 20, "github_languages": {"Python": 59289, "Shell": 2501}, "github_archived": false, "github_pushed_at": "2022-09-22T10:33:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pretraining-the-vision-transformer-using-self"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FKotzp6PZJw", "year": 2021, "status": "rejected", "title": "On the Estimation Bias in Double Q-Learning", "authors": ["Zhizhou Ren", "Guangxiang Zhu", "Beining Han", "Jianglun Chen", "Chongjie Zhang"], "authorids": ["~Zhizhou_Ren1", "~Guangxiang_Zhu1", "~Beining_Han1", "~Jianglun_Chen2", "~Chongjie_Zhang1"], "authors_source": "OpenReview API", "abstract": "Double Q-learning is a classical method for reducing overestimation bias, which is caused by taking maximum estimated values in the Bellman operator. Its variants in the deep Q-learning paradigm have shown great promise in producing reliable value prediction and improving learning performance. However, as shown by prior work, double Q-learning is not fully unbiased and still suffers from underestimation bias. In this paper, we show that such underestimation bias may lead to multiple non-optimal fixed points under an approximated Bellman operation. To address the concerns of converging to non-optimal stationary solutions, we propose a simple and effective approach as a partial fix for underestimation bias in double Q-learning. This approach leverages real returns to bound the target value. We extensively evaluate the proposed method in the Atari benchmark tasks and demonstrate its significant improvement over baseline algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "07u6jF0r1bB", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1410/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##################################\n\nSummary: This paper investigates the effects of approximation error in Q-learning - more specifically, the problem of having multiple non-optimal fixed points in double Q-learning. The authors claim that double Q-learning, which is a well-known approach to alleviate overestimation bias of Q-learning, can suffer from underestimation bias, and can lead to non-optimal fixed points. This paper provides theoretical evidence to support this claim. The main idea of this paper is to add real returns as a lower bound to the objective, and alleviate underestimation bias. Empirical results in Atari domains are presented. \n\n################################################\n\nPros\n\n1. This paper addresses an important problem in reinforcement learning - approximation errors and bias in Q-learning. The major focus of the RL community on estimation bias has been overestimation bias (van Hasselt 2010), but this paper brings up the issue of underestimation bias (Lan et al 2020). \n\n2. This paper formalizes a way to model approximation error in Bellman operators in Q-learning (by setting a random noises as approximation error), and derives multiple theoretical insights and propositions (Existence of multiple fixed points in double q-learning). \n\n3. The idea of having return signal V_\\tau (the returns gained from real trajectories) as a lower bound is simple and easy to implement. As for V(frozen\\theta), the authors use both DDQN and clipped-DDQN objective; they combined these two with the lower-bound objective, and showed the performance of DDQN-LB and clipped-DDQN-LB., which performs better than previous baselines in multiple Atari domains. \n\n4. Comparison with n-step bootstrapping (Figure3) is also helpful; the idea is compatible with n-step bootstrapping, so it can be combined to improve the performance. \n\n################################################\n\n\u0010\n\nQuestions\n\n1. Section 4: Choice of trajectory \\tau for return signal V_\\tau(s_{t+1}): this term is the discounted return of trajectory \\tau in the replay buffer. Is this just a return value of one trajectory? How do you address the stochasticity of having real return value of one trajectory? Or is this an average of multiple trajectories that start from s_{t+1}? \n\n2. Have you ever considered other baselines like duel-DQN or Rainbow to compare the performance of your approach? I don’t think the empirical results (comparison with DDQN, DDQN, cDDQN, maxmin DQN, etc)  are already sufficient, but I’m just curious how it performs well compared to SOTA baselines. \n\n\n###################################################\n\nReasons for score: I think this paper addresses an important problem in RL, and presents a simple yet effective approach to address this problem. This paper is also theoretically well supported. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "simple yet effective approach to address estimation bias", "review": "##################################\n\nSummary: This paper investigates the effects of approximation error in Q-learning - more specifically, the problem of having multiple non-optimal fixed points in double Q-learning. The authors claim that double Q-learning, which is a well-known approach to alleviate overestimation bias of Q-learning, can suffer from underestimation bias, and can lead to non-optimal fixed points. This paper provides theoretical evidence to support this claim. The main idea of this paper is to add real returns as a lower bound to the objective, and alleviate underestimation bias. Empirical results in Atari domains are presented. \n\n################################################\n\nPros\n\n1. This paper addresses an important problem in reinforcement learning - approximation errors and bias in Q-learning. The major focus of the RL community on estimation bias has been overestimation bias (van Hasselt 2010), but this paper brings up the issue of underestimation bias (Lan et al 2020). \n\n2. This paper formalizes a way to model approximation error in Bellman operators in Q-learning (by setting a random noises as approximation error), and derives multiple theoretical insights and propositions (Existence of multiple fixed points in double q-learning). \n\n3. The idea of having return signal V_\\tau (the returns gained from real trajectories) as a lower bound is simple and easy to implement. As for V(frozen\\theta), the authors use both DDQN and clipped-DDQN objective; they combined these two with the lower-bound objective, and showed the performance of DDQN-LB and clipped-DDQN-LB., which performs better than previous baselines in multiple Atari domains. \n\n4. Comparison with n-step bootstrapping (Figure3) is also helpful; the idea is compatible with n-step bootstrapping, so it can be combined to improve the performance. \n\n################################################\n\n\u0010\n\nQuestions\n\n1. Section 4: Choice of trajectory \\tau for return signal V_\\tau(s_{t+1}): this term is the discounted return of trajectory \\tau in the replay buffer. Is this just a return value of one trajectory? How do you address the stochasticity of having real return value of one trajectory? Or is this an average of multiple trajectories that start from s_{t+1}? \n\n2. Have you ever considered other baselines like duel-DQN or Rainbow to compare the performance of your approach? I don’t think the empirical results (comparison with DDQN, DDQN, cDDQN, maxmin DQN, etc)  are already sufficient, but I’m just curious how it performs well compared to SOTA baselines. \n\n\n###################################################\n\nReasons for score: I think this paper addresses an important problem in RL, and presents a simple yet effective approach to address this problem. This paper is also theoretically well supported. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603899536682}, {"id": "qnEpfqqvRcb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1410/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers the problem of learning a value function in a deterministic MDP and proposes a heuristic to reduce the bias of value estimates under the greedy policy. For this, they consider the class of Double Q-learning algorithms and describe a setting where estimation under this algorithm can lead to multiple fixed point solutions. They propose to control underestimation bias by estimating the next-state value as the maximum of either the current next-state value or the value of some trajectory in the replay buffer. They also propose another version that uses clipped target values. The two algorithms are evaluated on several Atari games and show, in some cases, improved transient behavior over double DQN. \n\nControlling for bias due to function approximation is an important problem, and one where general solutions could have a potentially large impact for RL and AI. Giving this subject the attention it deserves will require an understanding of the bias problem in stochastic settings. Many will be curious about what can be said about this problem and its solution(s) using formal mathematical language. The multiple-solutions hypothesis is a plausible approach to gaining a better understanding of bias, but there could also be other angles to explore. \n\nI think the current research this paper presents should be broadened to understand not just deterministic environments, but stochastic ones as well. The experimental results should also be questioned: why is there little to no asymptotic performance gain using the proposed method if it indeed produces an estimate with less bias? There are more questions the community would wish to answer with a paper like this; Where does underestimation bias come from, and how can it be controlled in general? Currently, however, the presentation and technical material are not sharp enough to answer such questions. Therefore, it is important that the authors continue to mature this work so that, when it is ready for publication, it will make a significant contribution.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting topic that needs to be better understood", "review": "The paper considers the problem of learning a value function in a deterministic MDP and proposes a heuristic to reduce the bias of value estimates under the greedy policy. For this, they consider the class of Double Q-learning algorithms and describe a setting where estimation under this algorithm can lead to multiple fixed point solutions. They propose to control underestimation bias by estimating the next-state value as the maximum of either the current next-state value or the value of some trajectory in the replay buffer. They also propose another version that uses clipped target values. The two algorithms are evaluated on several Atari games and show, in some cases, improved transient behavior over double DQN. \n\nControlling for bias due to function approximation is an important problem, and one where general solutions could have a potentially large impact for RL and AI. Giving this subject the attention it deserves will require an understanding of the bias problem in stochastic settings. Many will be curious about what can be said about this problem and its solution(s) using formal mathematical language. The multiple-solutions hypothesis is a plausible approach to gaining a better understanding of bias, but there could also be other angles to explore. \n\nI think the current research this paper presents should be broadened to understand not just deterministic environments, but stochastic ones as well. The experimental results should also be questioned: why is there little to no asymptotic performance gain using the proposed method if it indeed produces an estimate with less bias? There are more questions the community would wish to answer with a paper like this; Where does underestimation bias come from, and how can it be controlled in general? Currently, however, the presentation and technical material are not sharp enough to answer such questions. Therefore, it is important that the authors continue to mature this work so that, when it is ready for publication, it will make a significant contribution.\n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603889576223}, {"id": "emBgDSCMqia", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1410/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper analysed the underestimation bias induced by approximation error, by formalizing the underlying approximation, they theoretically proved the existence of multiple approximated fixed points which causes the converging to non-optimal solution. Besides, they proposed the lower bound double q-learning to overcome the underestimation bias. \n\nStrengths:\n1.The paper is well-organized and the algorithm is simple yet effective.\n\n2.The thought of formalizing the approximation error into noise is delighted which provides a way to theoretically analyze the effect of such error.\n\n3.The definition and existence prove of approximated fixed points would helps the algorithm design to overcome the underestimation bias.\n\nWeakness:\n1.Some experiment is hard to understand. Table1 shows the TD-error and the absolute state-action value which didn’t demonstrate the small approximation error would cause significant estimation error which would cause the sub-optimal fixed points.\n\n2.The effectiveness of lower bound double q-learning is doubtful. In MsPacman of Figure2, the algorithm shows slight performance decrease of Clipped DDQN, in some environment such as WizardOfWor, Zaxxon RoadRunner and BattleZone, these algorithms seems converge into same solutions. Besides, the algorithm would cause the overestimate the true maximum value. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper analysed the underestimation bias induced by approximation error, by formalizing the underlying approximation, they theoretically proved the existence of multiple approximated fixed points which causes the converging to non-optimal solution. Besides, they proposed the lower bound double q-learning to overcome the underestimation bias. ", "review": "This paper analysed the underestimation bias induced by approximation error, by formalizing the underlying approximation, they theoretically proved the existence of multiple approximated fixed points which causes the converging to non-optimal solution. Besides, they proposed the lower bound double q-learning to overcome the underestimation bias. \n\nStrengths:\n1.The paper is well-organized and the algorithm is simple yet effective.\n\n2.The thought of formalizing the approximation error into noise is delighted which provides a way to theoretically analyze the effect of such error.\n\n3.The definition and existence prove of approximated fixed points would helps the algorithm design to overcome the underestimation bias.\n\nWeakness:\n1.Some experiment is hard to understand. Table1 shows the TD-error and the absolute state-action value which didn’t demonstrate the small approximation error would cause significant estimation error which would cause the sub-optimal fixed points.\n\n2.The effectiveness of lower bound double q-learning is doubtful. In MsPacman of Figure2, the algorithm shows slight performance decrease of Clipped DDQN, in some environment such as WizardOfWor, Zaxxon RoadRunner and BattleZone, these algorithms seems converge into same solutions. Besides, the algorithm would cause the overestimate the true maximum value. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603866434278}, {"id": "R13h2nrv5Y-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1410/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This submission focused on the double Q-learning and investigated its underestimation bias issue. The authors claimed that using a double estimator can lead to multiple fixed points. To alleviate this issue, the authors proposed a correction approach by using the lower bound w.r.t. the real return. Finally, experiments on Atari games were conducted to show the improved performance, when compared with double DQN.\n\nThe authors aimed to analyze the underestimation issue in double Q-learning. The observation of multiple fixed points under the stochastic Bellman operator is interesting. The authors also tried to give more interpretation by providing sufficient conditions when this phenomenon happens. The proposed lower bounded double Q learning looks novel and reasonable. \n\nI am a bit confused about the authors' explanation on the multiple fixed points in Section 3.2. The authors claimed in the last paragraph that \"..due to the existence of the approximation error, the induced policy π˜ suffers from a remarkable uncertainty in determining the best action on state s0, which is the main cause of these extra fixed points. \" If this is indeed the case, the multiple fixed points are not necessarily linked to double Q-learning, as the approximation error can appear in any form of function approximation. \n\nFollowing the above point, I also don't quite understand the motivation about Definition 3. The authors claimed in Page 5 that \"The monotonicity property is not a sufficient condition for Eq. (5) but it usually becomes a source of risks in practice.\" I just do not see how Definition 3 is necessarily connected with Proposition 3. The authors provided empirical observation that \"During this process, a large magnitude of the first-order derivative could be found to meet the condition stated in Eq. (5)\" However, this part in the curve also corresponds to the case of underestimation vanishing. Is this a contradiction here?\n\nAnother concern I have is regarding the comparison w.r.t. multi-step learning: The authors plot in Figure 3 the performance of n-step bootstrapping vs. LB, which I really appreciate. Note that this comparison is for the game Alien only; therefore, I am wondering if the conclusion holds for other games? It would be more convincing if the authors can provide more experimental results, as it can reinforce the authors' claim that the lower bounded approach rather than multi step learning is the main factor for performance improvement.\n\nThere is a related work on correcting biases in DDQN as follows, which may need to be discussed as well. The authors there also showed the improved performance on Atari games.\n\nSong Z, Parr R, Carin L. Revisiting the softmax bellman operator: New benefits and new perspective, ICML 2019.\n\nMinor comments: \"...are meet...\" -> \"...are met...\" in Page 5.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This submission focused on the double Q-learning and investigated its underestimation bias issue. The authors claimed that using a double estimator can lead to multiple fixed points. To alleviate this issue, the authors proposed a correction approach by using the lower bound w.r.t. the real return. Finally, experiments on Atari games were conducted to show the improved performance, when compared with double DQN.\n\nThe authors aimed to analyze the underestimation issue in double Q-learning. The observation of multiple fixed points under the stochastic Bellman operator is interesting. The authors also tried to give more interpretation by providing sufficient conditions when this phenomenon happens. The proposed lower bounded double Q learning looks novel and reasonable. \n\nI am a bit confused about the authors' explanation on the multiple fixed points in Section 3.2. The authors claimed in the last paragraph that \"..due to the existence of the approximation error, the induced policy π˜ suffers from a remarkable uncertainty in determining the best action on state s0, which is the main cause of these extra fixed points. \" If this is indeed the case, the multiple fixed points are not necessarily linked to double Q-learning, as the approximation error can appear in any form of function approximation. \n\nFollowing the above point, I also don't quite understand the motivation about Definition 3. The authors claimed in Page 5 that \"The monotonicity property is not a sufficient condition for Eq. (5) but it usually becomes a source of risks in practice.\" I just do not see how Definition 3 is necessarily connected with Proposition 3. The authors provided empirical observation that \"During this process, a large magnitude of the first-order derivative could be found to meet the condition stated in Eq. (5)\" However, this part in the curve also corresponds to the case of underestimation vanishing. Is this a contradiction here?\n\nAnother concern I have is regarding the comparison w.r.t. multi-step learning: The authors plot in Figure 3 the performance of n-step bootstrapping vs. LB, which I really appreciate. Note that this comparison is for the game Alien only; therefore, I am wondering if the conclusion holds for other games? It would be more convincing if the authors can provide more experimental results, as it can reinforce the authors' claim that the lower bounded approach rather than multi step learning is the main factor for performance improvement.\n\nThere is a related work on correcting biases in DDQN as follows, which may need to be discussed as well. The authors there also showed the improved performance on Atari games.\n\nSong Z, Parr R, Carin L. Revisiting the softmax bellman operator: New benefits and new perspective, ICML 2019.\n\nMinor comments: \"...are meet...\" -> \"...are met...\" in Page 5.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603762689213}], "openreview_url": "https://openreview.net/forum?id=FKotzp6PZJw", "arxiv_id": "2109.14419", "paper_pdf": "papers/FKotzp6PZJw.pdf", "paper_pdf_sha256": "48483f43dd599eae7d07bd52ed6c742e38f3bfde6167ff91e1549abd6e1e6547", "paper_pdf_bytes": 1085232, "paper_pdf_source": "openreview", "code_url": "https://github.com/Stilwell-Git/Doubly-Bounded-Q-Learning", "code_repository": "Stilwell-Git/Doubly-Bounded-Q-Learning", "code_commit": "028ee629781f3fc7daa149b5b838884cfa09fcd1", "code_archive": "repos/FKotzp6PZJw.zip", "code_archive_sha256": "b7abd2079ad9c96c5b328240dc13250ed8b887a6d2cdf02d67d12f834847bf37", "code_archive_bytes": 22158, "code_file_count": 18, "code_extensions": {".py": 17, ".cpp": 1}, "github_disk_usage_kb": 20, "github_languages": {"Python": 40506, "C++": 4793}, "github_archived": false, "github_pushed_at": "2022-11-14T14:00:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-estimation-bias-in-double-q-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1lKd6NYPS", "year": 2020, "status": "rejected", "title": "Online Meta-Critic Learning for Off-Policy Actor-Critic Methods", "authors": ["Wei Zhou", "Yiying Li", "Yongxin Yang", "Huaimin Wang", "Timothy M. Hospedales"], "authorids": ["zhouwei14@nudt.edu.cn", "liyiying10@nudt.edu.cn", "yongxin.yang@ed.ac.uk", "hmwang@nudt.edu.cn", "t.hospedales@ed.ac.uk"], "authors_source": "OpenReview API", "abstract": "Off-Policy Actor-Critic (Off-PAC) methods have proven successful in a variety of continuous control tasks. Normally, the critic’s action-value function is updated using temporal-difference, and the critic in turn provides a loss for the actor that trains it to take actions with higher expected return. In this paper, we introduce a novel and flexible meta-critic that observes the learning process and meta-learns an additional loss for the actor that accelerates and improves actor-critic learning. Compared to the vanilla critic, the meta-critic network is explicitly trained to accelerate the learning process; and compared to existing meta-learning algorithms, meta-critic is rapidly learned online for a single task, rather than slowly over a family of tasks. Crucially, our meta-critic framework is designed for off-policy based learners, which currently provide state-of-the-art reinforcement learning sample efficiency. We demonstrate that online meta-critic learning leads to improvements in a variety of continuous control environments when combined with contemporary Off-PAC methods DDPG, TD3 and the state-of-the-art SAC. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rkeydHcRKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper641/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I am very torn about this paper as the proposed approach is a fairly straightforward extension of past work on the meta-critic approach to meta-learning and the results are pretty good, but nothing amazing. I tend to accept this paper because I like their general direction and think what they are proposing is pretty simple with broad applicability. It should be fairly straightforward to append this idea to most new off policy methods as they come out, so I find their consistent gains across 3 different popular models pretty convincing that this could have value to the community.  \n\nThat being said, the gains are not huge, which does make me think about the potential computational overhead. How much more run time per step does the meta-critic add to the models in the paper? I am a bit worried that the comparisons are not apples to apples from the perspective of the amount of computation/update steps per environment interaction due to the use of the validation data. I wonder if it changes the conclusion at all if the other approaches are given the same amount of computation on their replay buffer between interactions with the environment. For example, more optimization steps on the buffer is another plausible explanation why the meta-critic does better at optimizing the loss. \n\nI also should note that this paper is not the first to propose conducting online meta-learning over a replay buffer. A paper at last year's ICLR [1] did so in the context of lifelong learning, but does not need task labels or tasks and was tested on single non-stationary environments as well. The Meta-Critic approach of this work, of course, still is cool as it does for learning with a Meta-Critic what Meta-Experience Replay does for optimization based meta-learning. However, I thought this should be pointed out as the novelty can be a bit overstated at times. Additionally, the paper would be significantly improved by fleshing out the theoretical motivation for the Meta-Critic approach in more detail. What are the underlying reasons why we would expect it to generically improve single task RL? \n \n[1] \"Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference\". Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. ICLR-19. \n\nThoughts After Author Feedback:\n\nI really appreciate the response of the authors to my review, which included some interesting new experiments and explanations addressing concerns I raised. I do, however, also see where reviewer 3 is coming from with both major comments. \n\nI don't think that R3A1 is particularly clear and it is an important concern. I think reviewer 3 is saying that Delta(t) in R3A1i will eventually converge to 0 when the policy stops changing while the author argue that it \"need not converge if there are fluctuations at each iteration\". So it not converging seems to be tied to the existence of some source of non-stationarity in the problem. This doesn't seem to be coming from the environment as they are considering single task settings. As a result, I believe the source of the non-stationarity in this case is the fluctuating parameters. Looking at figure 6a it does not seem obvious that the policy has really converged in the traditional sense as its score does seem to be changing to some degree throughout the chart. My best guess is that this is the reason why the meta-loss does not converge. However, I still totally agree that this is a major concern that is very much under addressed. \n\nI also agree that I found the comments about what the meta-critic is doing unconvincing. The authors provided a few different kinds of explanations of what the model could potentially be doing, but this approach to the answer really highlights  how the theoretical benefits of this approach remain unclear. I think it should be possible to directly verify some of these theories with well designed experiments. It feels like a better explanation is necessary in light of the often small margin of difference with baselines before publication. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "I am very torn about this paper as the proposed approach is a fairly straightforward extension of past work on the meta-critic approach to meta-learning and the results are pretty good, but nothing amazing. I tend to accept this paper because I like their general direction and think what they are proposing is pretty simple with broad applicability. It should be fairly straightforward to append this idea to most new off policy methods as they come out, so I find their consistent gains across 3 different popular models pretty convincing that this could have value to the community.  \n\nThat being said, the gains are not huge, which does make me think about the potential computational overhead. How much more run time per step does the meta-critic add to the models in the paper? I am a bit worried that the comparisons are not apples to apples from the perspective of the amount of computation/update steps per environment interaction due to the use of the validation data. I wonder if it changes the conclusion at all if the other approaches are given the same amount of computation on their replay buffer between interactions with the environment. For example, more optimization steps on the buffer is another plausible explanation why the meta-critic does better at optimizing the loss. \n\nI also should note that this paper is not the first to propose conducting online meta-learning over a replay buffer. A paper at last year's ICLR [1] did so in the context of lifelong learning, but does not need task labels or tasks and was tested on single non-stationary environments as well. The Meta-Critic approach of this work, of course, still is cool as it does for learning with a Meta-Critic what Meta-Experience Replay does for optimization based meta-learning. However, I thought this should be pointed out as the novelty can be a bit overstated at times. Additionally, the paper would be significantly improved by fleshing out the theoretical motivation for the Meta-Critic approach in more detail. What are the underlying reasons why we would expect it to generically improve single task RL? \n \n[1] \"Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference\". Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. ICLR-19. \n\nThoughts After Author Feedback:\n\nI really appreciate the response of the authors to my review, which included some interesting new experiments and explanations addressing concerns I raised. I do, however, also see where reviewer 3 is coming from with both major comments. \n\nI don't think that R3A1 is particularly clear and it is an important concern. I think reviewer 3 is saying that Delta(t) in R3A1i will eventually converge to 0 when the policy stops changing while the author argue that it \"need not converge if there are fluctuations at each iteration\". So it not converging seems to be tied to the existence of some source of non-stationarity in the problem. This doesn't seem to be coming from the environment as they are considering single task settings. As a result, I believe the source of the non-stationarity in this case is the fluctuating parameters. Looking at figure 6a it does not seem obvious that the policy has really converged in the traditional sense as its score does seem to be changing to some degree throughout the chart. My best guess is that this is the reason why the meta-loss does not converge. However, I still totally agree that this is a major concern that is very much under addressed. \n\nI also agree that I found the comments about what the meta-critic is doing unconvincing. The authors provided a few different kinds of explanations of what the model could potentially be doing, but this approach to the answer really highlights  how the theoretical benefits of this approach remain unclear. I think it should be possible to directly verify some of these theories with well designed experiments. It feels like a better explanation is necessary in light of the often small margin of difference with baselines before publication. \n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571886439305}, {"id": "Skgom5J6FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper641/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In actor-critic algorithms, the policy tries to optimize cumulative discounted rewards by a loss formed with components from the critic. In this paper, the authors propose a novel way, namely, meta-critic, which utilizes meta-learning to learn an additional loss for the policy to accelerate the learning process of the agent. There are several advantages of the proposed method, it's learned online for a single task, it's trained with off-policy data, it provides sample efficiency and it's generally applicable to existing off-policy actor-critic methods. \n\nOverall the proposed method is novel and the research direction is a very interesting one to explore. Eqn (3) and Eqn (4) explains the key idea of the proposed method. Eqn (3) describes the meta-learning problem as a bi-level optimization problem where the agent is updated with the main loss L^main with data d_train, in addition, it's updated with L^aux where the loss is learned and parameterized by \\omega. After the agent being updated, it uses L^meta and data d_val to validate the performance of the updated agent. Eqn (4) describes an explicit way to formalize the usage of the auxiliary loss, which is to accelerate the learning process. Thus the meta loss is whether the L^aux helps the learning process or not. \n\nHope that the authors could address the following issues in the rebuttal:\n1) Investigate why DDPG with meta-critic gets much more improvements than TD3/SAC;\n2) Show SAC (or TD3) could get better performance on harder task. (I can understand that they are strong baselines and hard to improve on the current environments, however, for harder environments, there might be rooms for improvements.);\n3) Investigate different ways to parameterize the meta-critic other than simple MLP;\n\nA few minor points:\n1) Page 2 first paragraph, \"This is in stark contrast to the\nmainstream meta-learning research paradigm – where entire task families are required to provide\nenough data for meta-learning, and to provide new tasks to amortize the huge cost of meta-learning.\". Try to revise it and avoid the words like \"mainstream\". Think about the paper being read 5 years later or even more;\n2) Consider introducing a weighting hyperparam between L^main and L^aux in Eqn (3), these two losses might have different scales and it might be better to weigh them differently;\n3) Minor literature detail: Page 7 \"Comparison vs PPO-LIRPG\" mentioned that \"Intrinsic Reward Learning for PPO (Zheng et al., 2018) is the only existing online meta-learning method that we are aware of.\", however, AFAIK, \"Meta-Gradient Reinforcement Learning\" in NeurIPS 2018 and \"Discovery of Useful Questions as Auxiliary Tasks\" NeurIPS 2019 are methods where meta-learning is applied online and for a single task;", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "In actor-critic algorithms, the policy tries to optimize cumulative discounted rewards by a loss formed with components from the critic. In this paper, the authors propose a novel way, namely, meta-critic, which utilizes meta-learning to learn an additional loss for the policy to accelerate the learning process of the agent. There are several advantages of the proposed method, it's learned online for a single task, it's trained with off-policy data, it provides sample efficiency and it's generally applicable to existing off-policy actor-critic methods. \n\nOverall the proposed method is novel and the research direction is a very interesting one to explore. Eqn (3) and Eqn (4) explains the key idea of the proposed method. Eqn (3) describes the meta-learning problem as a bi-level optimization problem where the agent is updated with the main loss L^main with data d_train, in addition, it's updated with L^aux where the loss is learned and parameterized by \\omega. After the agent being updated, it uses L^meta and data d_val to validate the performance of the updated agent. Eqn (4) describes an explicit way to formalize the usage of the auxiliary loss, which is to accelerate the learning process. Thus the meta loss is whether the L^aux helps the learning process or not. \n\nHope that the authors could address the following issues in the rebuttal:\n1) Investigate why DDPG with meta-critic gets much more improvements than TD3/SAC;\n2) Show SAC (or TD3) could get better performance on harder task. (I can understand that they are strong baselines and hard to improve on the current environments, however, for harder environments, there might be rooms for improvements.);\n3) Investigate different ways to parameterize the meta-critic other than simple MLP;\n\nA few minor points:\n1) Page 2 first paragraph, \"This is in stark contrast to the\nmainstream meta-learning research paradigm – where entire task families are required to provide\nenough data for meta-learning, and to provide new tasks to amortize the huge cost of meta-learning.\". Try to revise it and avoid the words like \"mainstream\". Think about the paper being read 5 years later or even more;\n2) Consider introducing a weighting hyperparam between L^main and L^aux in Eqn (3), these two losses might have different scales and it might be better to weigh them differently;\n3) Minor literature detail: Page 7 \"Comparison vs PPO-LIRPG\" mentioned that \"Intrinsic Reward Learning for PPO (Zheng et al., 2018) is the only existing online meta-learning method that we are aware of.\", however, AFAIK, \"Meta-Gradient Reinforcement Learning\" in NeurIPS 2018 and \"Discovery of Useful Questions as Auxiliary Tasks\" NeurIPS 2019 are methods where meta-learning is applied online and for a single task;"}, "tcdate": 1571777059125}, {"id": "r1xRDNSNKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper641/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a meta reinforcement learning approach where a meta-critic is trained in addition to a conventional actor and critic, so that the meta-critic boosts the training of the actor over a single task. The approach is combined to DDPG, TD3 and SAC and is claimed to convey superior performance (or learning speed) over these state-of-the-art actor-critic algorithms.\n\nAt first glance, the paper looks convincing, but closer inspection reveals a potential issue that I would like the authors to discuss.\n\nThe first cue is that in Fig. 6, the meta-loss does not seem to converge to anything, as the authors say it just fluctuates. Shouldn't it converge once the actor converges to close to optimal performance? \n\nThe second cue is that appart from the rllab tasks (ant and half-cheetah), it is not clear that the meta-critic approach brings some gain in the end of training. Particularly in Reacher (Fig 7), the performance seems to collapse faster with DDPG-MC than with DDPG, and on the rest of curves of Figs 7 and 8 it is had to determine whether the MC approach brings something significant or not. And in Fig. 3, in Walker2D, TD3-MC looks rather unstable.\n\nThe third cue is that in Section 3.2, the paragraphs \"Updating MC parameters\" and \"Designing MC\" are rather unclear (I'll come back to that) and lack a theoretical justification.\n\nSo I'm wondering what exactly MC is doing and I would like to see a more detailled analysis. Couldn't a similar performance improvement be obtained by just increasing the actor learning rate and addiing some noise to the actor learning gradient? Isn't this more or less what the meta-loss does, when looking at Fig. 6?\n\nTo me, unless the authors give a clear answer to the points above, the paper should be rejected as it does not provide clear enough evidence and justification in favor of the proposed meta-learning approach.\n\nAs stated above, I found the paragraph \"Updating MC parameters\" lacking a principled justification.\n\nAbout \"Designing MC\", it could be much clearer. You first express two requirements (i) and (ii). Fine.\nThen you explain how to meet requirement (i), but you say this is not what you are doing (!), and then you try to explain what you are doing but without going back to the requirements. In particular, it is not clear at all why your design is \"permutation invariant\". You should be much more direct. Besides, the way to extract features and the relationship to batch-wise set-embedding should be made more explicit. Well, it is complicated and should be explained more clearly...\n\n\"TD3 borrows the Double Q learning idea...\": no, TD3 does more than this, as it uses the min between both critics, which double Q-learning and DDQN do not.\n\nYou use HalfCheetah-v2 from gym-mujoco and HalfCheetah from rllab. Why? One might suspect this is because the performance of the MC appraoch is only better with rllab...\n\nYou say you are using 10 million steps, but this is not always the case.\n\nIt is interesting to see that the performance of TD3 on Half-Cheetah collapses at around 4 million steps. Any idea why? It seems that in many papers experiments are stopped before performance collapses, and a strong study about this phenomenon is missing as the authors don't want to show this. This is an open call to readers: a paper on that would be great! :)\n\n\"SAC-MC\" gives a clear-boost for several tasks\": well, looking at the figures it is not so clear. Do you mean faster learning, higher final performance or both? I would like to see this claim backed-up by some proper statistical significance test and a clear specification of the number of seeds, etc. Fig.5 conveys the adequate data for such test if your claim is that this is the final performance that matters (but beware that curves are crossing eachother, so the final performance depends a lot on where you stop...).\n\ntypos:\n\np1 \"For example, ... (Zheng et al., 2018).\" is not a sentence (no main verb).\np2 \"the on-policy (approach?) needs to interact with (the) environment...\"\np2 \"is less effective than off-policy transitions.\" => unclear statement\np3 to assist(s)\n\nEquations (2), (6) and (8) should finish with a dot as they close a sentence.\n\nEq X => Eq (X) : use \\eqref{label} instead of just \\ref{label}\n\np5 It's input => Its\np5 two key technologies: I would not call this a technology. Ingredients?\np5 \"computational cost is ? by using\": missong word\n\np6: asmyptotic\n\np9: we removing\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a meta reinforcement learning approach where a meta-critic is trained in addition to a conventional actor and critic, so that the meta-critic boosts the training of the actor over a single task. The approach is combined to DDPG, TD3 and SAC and is claimed to convey superior performance (or learning speed) over these state-of-the-art actor-critic algorithms.\n\nAt first glance, the paper looks convincing, but closer inspection reveals a potential issue that I would like the authors to discuss.\n\nThe first cue is that in Fig. 6, the meta-loss does not seem to converge to anything, as the authors say it just fluctuates. Shouldn't it converge once the actor converges to close to optimal performance? \n\nThe second cue is that appart from the rllab tasks (ant and half-cheetah), it is not clear that the meta-critic approach brings some gain in the end of training. Particularly in Reacher (Fig 7), the performance seems to collapse faster with DDPG-MC than with DDPG, and on the rest of curves of Figs 7 and 8 it is had to determine whether the MC approach brings something significant or not. And in Fig. 3, in Walker2D, TD3-MC looks rather unstable.\n\nThe third cue is that in Section 3.2, the paragraphs \"Updating MC parameters\" and \"Designing MC\" are rather unclear (I'll come back to that) and lack a theoretical justification.\n\nSo I'm wondering what exactly MC is doing and I would like to see a more detailled analysis. Couldn't a similar performance improvement be obtained by just increasing the actor learning rate and addiing some noise to the actor learning gradient? Isn't this more or less what the meta-loss does, when looking at Fig. 6?\n\nTo me, unless the authors give a clear answer to the points above, the paper should be rejected as it does not provide clear enough evidence and justification in favor of the proposed meta-learning approach.\n\nAs stated above, I found the paragraph \"Updating MC parameters\" lacking a principled justification.\n\nAbout \"Designing MC\", it could be much clearer. You first express two requirements (i) and (ii). Fine.\nThen you explain how to meet requirement (i), but you say this is not what you are doing (!), and then you try to explain what you are doing but without going back to the requirements. In particular, it is not clear at all why your design is \"permutation invariant\". You should be much more direct. Besides, the way to extract features and the relationship to batch-wise set-embedding should be made more explicit. Well, it is complicated and should be explained more clearly...\n\n\"TD3 borrows the Double Q learning idea...\": no, TD3 does more than this, as it uses the min between both critics, which double Q-learning and DDQN do not.\n\nYou use HalfCheetah-v2 from gym-mujoco and HalfCheetah from rllab. Why? One might suspect this is because the performance of the MC appraoch is only better with rllab...\n\nYou say you are using 10 million steps, but this is not always the case.\n\nIt is interesting to see that the performance of TD3 on Half-Cheetah collapses at around 4 million steps. Any idea why? It seems that in many papers experiments are stopped before performance collapses, and a strong study about this phenomenon is missing as the authors don't want to show this. This is an open call to readers: a paper on that would be great! :)\n\n\"SAC-MC\" gives a clear-boost for several tasks\": well, looking at the figures it is not so clear. Do you mean faster learning, higher final performance or both? I would like to see this claim backed-up by some proper statistical significance test and a clear specification of the number of seeds, etc. Fig.5 conveys the adequate data for such test if your claim is that this is the final performance that matters (but beware that curves are crossing eachother, so the final performance depends a lot on where you stop...).\n\ntypos:\n\np1 \"For example, ... (Zheng et al., 2018).\" is not a sentence (no main verb).\np2 \"the on-policy (approach?) needs to interact with (the) environment...\"\np2 \"is less effective than off-policy transitions.\" => unclear statement\np3 to assist(s)\n\nEquations (2), (6) and (8) should finish with a dot as they close a sentence.\n\nEq X => Eq (X) : use \\eqref{label} instead of just \\ref{label}\n\np5 It's input => Its\np5 two key technologies: I would not call this a technology. Ingredients?\np5 \"computational cost is ? by using\": missong word\n\np6: asmyptotic\n\np9: we removing\n"}, "tcdate": 1571210341782}], "openreview_url": "https://openreview.net/forum?id=H1lKd6NYPS", "arxiv_id": "2003.05334", "paper_pdf": "papers/H1lKd6NYPS.pdf", "paper_pdf_sha256": "3f73a188b1d752d13af721ddf2d41d99a57316b24523da888689f411b6a5ad89", "paper_pdf_bytes": 7208362, "paper_pdf_source": "openreview", "code_url": "https://github.com/zwfightzw/Meta-Critic", "code_repository": "zwfightzw/Meta-Critic", "code_commit": "d8045d66fa82d5035868b82e6bd9cbdbe6fa5955", "code_archive": "repos/H1lKd6NYPS.zip", "code_archive_sha256": "4ec08efa2248202c6989da1920d4263836bab039a868c110ccb47ac8068e5ed6", "code_archive_bytes": 26608, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 22, "github_languages": {"Python": 78982}, "github_archived": false, "github_pushed_at": "2020-10-19T06:25:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/online-meta-critic-learning-for-off-policy-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Sygx4305KQ", "year": 2019, "status": "rejected", "title": "Small steps and giant leaps: Minimal Newton solvers for Deep Learning", "authors": ["Joao Henriques", "Sebastien Ehrhardt", "Samuel Albanie", "Andrea Vedaldi"], "authorids": ["joao@robots.ox.ac.uk", "hyenal@robots.ox.ac.uk", "albanie@robots.ox.ac.uk", "vedali@robots.ox.ac.uk"], "authors_source": "OpenReview API", "abstract": "We propose a fast second-order method that can be used as a drop-in replacement for current deep learning solvers. Compared to stochastic gradient descent (SGD), it only requires two additional forward-mode automatic differentiation operations per iteration, which has a computational cost comparable to two standard forward passes and is easy to implement. Our method addresses long-standing issues with current second-order solvers, which invert an approximate Hessian matrix every iteration exactly or by conjugate-gradient methods, procedures that are much slower than a SGD step. Instead, we propose to keep a single estimate of the gradient projected by the inverse Hessian matrix, and update it once per iteration with just two passes over the network. This estimate has the same size and is similar to the momentum variable that is commonly used in SGD. No estimate of the Hessian is maintained.\nWe first validate our method, called CurveBall, on small problems with known solutions (noisy Rosenbrock function and degenerate 2-layer linear networks), where current deep learning solvers struggle. We then train several large models on CIFAR and ImageNet, including ResNet and VGG-f networks, where we demonstrate faster convergence with no hyperparameter tuning. We also show our optimiser's generality by testing on a large set of randomly-generated architectures.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SyxRWTJxpQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1418/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Authors propose choosing direction by using a single step of gradient descent \"towards Newton step\" from an original estimate, and then taking this direction instead of original gradient. This direction is reused as a starting estimate for the next iteration of the algorithm. This can be efficiently implemented since it only relies on Hessian-vector products which are accessible in all major frameworks.\n\nBased on the fact that this is an easy to implement idea, clearly described, and that it seems to benefit some tasks using standard architectures, I would recommend this paper for acceptance.\n\nComments:\n- introducing \\rho parameter and solving for optimal \\rho, \\beta complicates things. I'm assuming \\rho was needed for practical reasons, this should be explained better in the paper. (ie, what if we leave rho at 1)\n- For  ImageNet results, they show 82% accuracy after 20 epochs on full ImageNet using VGG. Is this top5 or top1 error? I'm assuming top5 since top1 would be new world record for the number of epochs needed. For top5, it seems SGD has stopped optimizing at 60% top5. Since all the current records on ImageNet are achieved with SGD (which beats Adam), this suggests that the SGD implementation is badly tuned\n- I appreciate that CIFAR experiments were made using standard architectures, ie using networks with batch-norm which clearly benefits SGD", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well-motivated idea", "review": "Authors propose choosing direction by using a single step of gradient descent \"towards Newton step\" from an original estimate, and then taking this direction instead of original gradient. This direction is reused as a starting estimate for the next iteration of the algorithm. This can be efficiently implemented since it only relies on Hessian-vector products which are accessible in all major frameworks.\n\nBased on the fact that this is an easy to implement idea, clearly described, and that it seems to benefit some tasks using standard architectures, I would recommend this paper for acceptance.\n\nComments:\n- introducing \\rho parameter and solving for optimal \\rho, \\beta complicates things. I'm assuming \\rho was needed for practical reasons, this should be explained better in the paper. (ie, what if we leave rho at 1)\n- For  ImageNet results, they show 82% accuracy after 20 epochs on full ImageNet using VGG. Is this top5 or top1 error? I'm assuming top5 since top1 would be new world record for the number of epochs needed. For top5, it seems SGD has stopped optimizing at 60% top5. Since all the current records on ImageNet are achieved with SGD (which beats Adam), this suggests that the SGD implementation is badly tuned\n- I appreciate that CIFAR experiments were made using standard architectures, ie using networks with batch-norm which clearly benefits SGD", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541565701519}, {"id": "ryeIGPDv2m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1418/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors introduce a new second-order algorithm for training deep networks. The method, named CurveBall, is motivated as an inexpensive alternative to Newton-CG. At its core, the method augments the update role for SGD+M with a Hessian-vector product that can be done efficiently (Algorithm 1). While a few new hyperparameters are introduced, the authors propose ways by which they can be calibrated automatically (Equation 16) and also prove convergence for quadratic functions (Theorem A.1) and guaranteed descent (Theorem A.2). The authors also present numerical results showing improved training on common benchmarks. I enjoyed reading the paper and found the motivation and results to be convincing. I especially appreciate that the authors performed experiments on ImageNet instead of just CIFAR-10, and the differentiation modes are explained well. As such, I recommend the paper for acceptance. \n\n\nI suggest ways in which the paper can be further improved below:\n\n- In essence, the closest algorithm to CurveBall is LiSSA proposed by Agarwal et al. They use a series expansion for approximating the inverse whereas your work uses one iteration of CG. If you limit LiSSA to only one expansion, the update rule that you would get would be similar to that of CurveBall (but not exactly the same). I feel that a careful comparison to LiSSA is necessary in the paper, highlighting the algorithmic and theoretical differences. I don't see the need for any additional experiments, however.\n- For books, such as Nocedal & Wright, please provide page numbers for each citation since the information quoted is across hundreds of pages. \n- It's a bit non-standard to see vectors being denoted by capital letters, e.g. J(w) \\in R^p on Page 2. I think it's better you don't change it now, however, since that might introduce inadvertent typos. \n- It would be good if you could expand on the details concerning the automatic determination of the hyperparameters (Equation 16). It was a bit unclear to me where those equations came from. \n- Could you plot the evolution of \\beta, \\rho and \\lambda for a couple of your experiments? I am curious whether our intuition about the values aligns with what happens in reality. In Newton-CG or Levenberg-Marquardt-esque algorithms, with standard local strong convexity assumptions, the amount of damping necessary near the solution usually falls to 0. Further, in the SGD+M paper of Sutskever et al., they talked about how it was necessary to zero out the momentum at the end. It would be fascinating if such insights (or contradictory ones) were discovered by Equation 16 and the damping mechanism automatically. \n- I'm somewhat concerned about the damping for \\lambda using \\gamma. There has been quite a lot of work recently in the area of Stochastic Line Searches which underscores the issues involving computation with noisy estimates of function values. I wonder if the randomness inherent in the computation of f(w) can throw off your estimates enough to cause convergence issues. Can you comment on this?\n- It was a bit odd to see BFGS implemented with a cubic line search. The beneficial properties of BFGS, such as superlinear convergence and self-correction, usually work out only if you're using the Armijo-Wolfe (Strong/Weak) line search. Can you re-do those experiments with this line search? It is unexpected that BFGS would take O(100) iterations to converge on a two dimensional problem. \n- In the same experiment, did you also try (true) Newton's method? Maybe we some form of damping? Given that you're proposing an approximate Newton's method, it would be a good upper baseline to have this experiment. \n- I enjoyed reading your experimental section on random architectures, I think it is quite illuminating. \n- Please consider rephrasing some phrases in the paper such as \"soon the latter\" (Page 1), \"which is known to improve optimisation\", (Page 7), \"non-deep problems\" (Page 9). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good Paper, Accept", "review": "In this paper, the authors introduce a new second-order algorithm for training deep networks. The method, named CurveBall, is motivated as an inexpensive alternative to Newton-CG. At its core, the method augments the update role for SGD+M with a Hessian-vector product that can be done efficiently (Algorithm 1). While a few new hyperparameters are introduced, the authors propose ways by which they can be calibrated automatically (Equation 16) and also prove convergence for quadratic functions (Theorem A.1) and guaranteed descent (Theorem A.2). The authors also present numerical results showing improved training on common benchmarks. I enjoyed reading the paper and found the motivation and results to be convincing. I especially appreciate that the authors performed experiments on ImageNet instead of just CIFAR-10, and the differentiation modes are explained well. As such, I recommend the paper for acceptance. \n\n\nI suggest ways in which the paper can be further improved below:\n\n- In essence, the closest algorithm to CurveBall is LiSSA proposed by Agarwal et al. They use a series expansion for approximating the inverse whereas your work uses one iteration of CG. If you limit LiSSA to only one expansion, the update rule that you would get would be similar to that of CurveBall (but not exactly the same). I feel that a careful comparison to LiSSA is necessary in the paper, highlighting the algorithmic and theoretical differences. I don't see the need for any additional experiments, however.\n- For books, such as Nocedal & Wright, please provide page numbers for each citation since the information quoted is across hundreds of pages. \n- It's a bit non-standard to see vectors being denoted by capital letters, e.g. J(w) \\in R^p on Page 2. I think it's better you don't change it now, however, since that might introduce inadvertent typos. \n- It would be good if you could expand on the details concerning the automatic determination of the hyperparameters (Equation 16). It was a bit unclear to me where those equations came from. \n- Could you plot the evolution of \\beta, \\rho and \\lambda for a couple of your experiments? I am curious whether our intuition about the values aligns with what happens in reality. In Newton-CG or Levenberg-Marquardt-esque algorithms, with standard local strong convexity assumptions, the amount of damping necessary near the solution usually falls to 0. Further, in the SGD+M paper of Sutskever et al., they talked about how it was necessary to zero out the momentum at the end. It would be fascinating if such insights (or contradictory ones) were discovered by Equation 16 and the damping mechanism automatically. \n- I'm somewhat concerned about the damping for \\lambda using \\gamma. There has been quite a lot of work recently in the area of Stochastic Line Searches which underscores the issues involving computation with noisy estimates of function values. I wonder if the randomness inherent in the computation of f(w) can throw off your estimates enough to cause convergence issues. Can you comment on this?\n- It was a bit odd to see BFGS implemented with a cubic line search. The beneficial properties of BFGS, such as superlinear convergence and self-correction, usually work out only if you're using the Armijo-Wolfe (Strong/Weak) line search. Can you re-do those experiments with this line search? It is unexpected that BFGS would take O(100) iterations to converge on a two dimensional problem. \n- In the same experiment, did you also try (true) Newton's method? Maybe we some form of damping? Given that you're proposing an approximate Newton's method, it would be a good upper baseline to have this experiment. \n- I enjoyed reading your experimental section on random architectures, I think it is quite illuminating. \n- Please consider rephrasing some phrases in the paper such as \"soon the latter\" (Page 1), \"which is known to improve optimisation\", (Page 7), \"non-deep problems\" (Page 9). ", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541007117966}, {"id": "ryeMkWWQn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1418/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an approximate second-order method with low computational cost. A common pitfall of second-order methods is the computation (and perhaps inversion) of the Hessian matrix. While this can be avoided by instead relying on Hessian-vector products as done in CG, it typically still requires several iterations. Instead, the authors suggest a simpler approach that relies on one single gradient step and a warm start strategy. The authors points out that the resulting algorithm resembles a momentum method. They also provide some simple convergence proofs on quadratics and benchmark their method to train deep neural networks.\n\nWhile I find the research direction interesting, the execution is rather clumsy and many details are not sufficiently motivated. Finally, there is a lot of relevant work in the optimization community that is not discussed in this paper, see detailed comments and references below.\n\n1) Method\nThe derivation of the method is very much driven on a set of heuristics without theoretical guarantees. In order to derive the update of the proposed method, the authors rely on three heuristics:\na) The first is to reuse the previous search direction z as a warm-start. The authors argue that this might be beneficial if If z does not change abruptly. In the early phase, the gradient norm is likely to be large and thus z will change significantly. One might also encounter regions of high curvature where the direction of z might change quickly from one iteration to the next.\nThe \"warm start\" at s_{t-1} is also what yields the momentum term, what interpretation can you give to this choice?\n\nb) The second step interleaves the updates of z and w instead of first finding the optimum z. This amounts to just running one iteration of CG but it is rather unclear why one iteration is an appropriate number. It seems one could instead some adaptive strategy where CG with a fixed accuracy. One could potentially see if allowing larger errors at the beginning of the optimization process might still allow for the method to converge. This is for instance commonly done with the batch-size of first-order method. Gradually increasing the batch-size and therefore reducing the error as one gets close to the optimum can still yield to a converging algorithm, see e.g. \nFriedlander, M. P., & Schmidt, M. (2012). Hybrid deterministic-stochastic methods for data fitting. SIAM Journal on Scientific Computing, 34(3), A1380-A1405.\n\nc) The third step consists in replacing CG with gradient descent.\n\"If CG takes N steps on average, then Algorithm 2 will be slower than SGD by a factor of at least N, which can easily be an order of magnitude\".\nFirst, the number of outer iterations may be a lot less for the Hessian-free method than for SGD so this does not seem to be a valid argument. Please comment.\nSecond, I would like to see a discussion of the convergence rate of solving (12) inexactly with krylov subspace methods. Note that Lanczos yields an accelerated rate while GD does not. So the motivation for switching to GD should be made clearer.\n\nd) The fourth step introduces a factor rho that decays z at each step. I’m not really sure this makes sense even heuristically. The full update of the algorithm developed by the author is:\nw_{t+1} = w_t - beta nabla f + (rho I - beta H) (w_t - w_{t-1}).\nThe momentum term therefore gets weighted by (rho I - beta H). What is the meaning of this term? The -beta H term weights the momentum according to the curvature of the objective function. Given the lack of theoretical support for this idea, I would at least expect a practical reason back up by some empirical evidence that this is a sensible thing to do.\nThis is especially important given that you claim to decay rho therefore giving more importance to the curvature term.\nFinally, why would this be better than simply using CG on a trust-region model? (Recall that Lanczos yields an accelerated linear rate while GD does not).\n\n2) Convergence analysis\na) The analysis is only performed on a quadratic while the author clearly target non-convex functions, this should be made clear in the main text. Also see references below (comment #3) regarding a possible extension to non-convex functions.\nb) The authors should check the range of allowed values for alpha and beta. It appears the rate would scale with the square root of the condition number, please confirm, this is an important detail. I also think that the constant is not as good as Heavy-ball on a quadratic (see e.g. http://pages.cs.wisc.edu/~brecht/cs726docs/HeavyBallLinear.pdf), please comment.\nc) Sub-sampling of the Hessian and gradients is not discussed at all (but used in the experiments). Please add a discussion and consider extending the proof (again, see references given below).\n\n3) Convergence Heavy-ball\nThe authors emphasize the similarity of their approach to Heavy-ball. They cite the results of Loizou & Richtarik 2017. Note that they are earlier results for quadratic functions such as \nLessard, L., Recht, B., & Packard, A. (2016). Analysis and design of optimization algorithms via integral quadratic constraints. SIAM Journal on Optimization, 26(1), 57-95.\nFlammarion, N., & Bach, F. (2015, June). From averaging to acceleration, there is only a step-size. In Conference on Learning Theory (pp. 658-695).\nThe novelty of the bounds derived in Loizou & Richtarik 2017 is that they apply in stochastic settings.\nFinally, there are results for non-convex functions such convergence to a stationary point, see\nZavriev, S. K., & Kostyuk, F. V. (1993). Heavy-ball method in nonconvex optimization problems. Computational Mathematics and Modeling, 4(4), 336-341.\nAlso on page 2, \"Momentum GD ... can be shown to have faster convergence than GD\". It should be mentioned that this only hold for (strongly) convex functions!\n\n4) Experiments\na) Consider showing the gradient norms. \nb) it looks like the methods have not yet converged in Fig 2 and 3.\nc) Second order benchmark:\nIt would be nice to compare to a method that does not use the GN matrix but the true or subsampled Hessian (like Trust Region/Cubic Regularization) methods given below.\nWhy is BFGS in Rosenbrock but not in NN plots?\nd) \"Batch normalization (which is known to improve optimization)\" \nThis statement requires a reference such as\nTowards a Theoretical Understanding of Batch Normalization\nKohler et al… - arXiv preprint arXiv:1805.10694, 2018\n\n5) Related Work\nThe related work should include Cubic Regularization and Trust Region methods since they are among the most prominent second order algorithms. Consider citing Conn et al. 2000 Trust Region,  Nesterov 2006 Cubic regularization, Cartis et al. 2011 ARC.\nRegarding sub-sampling: Kohler&Lucchi 2017: Stochastic Cubic Regularization for non-convex optimization and Xu et al.: Newton-type methods for non-convex optimization under inexact hessian information.\n\n6) More comments\n\nPage 2\nPolyak 1964 should be cited  where momentum is discussed.\n\"Perhaps the simplest algorithm to optimize Eq. 1 is Gradient Descent\". This is technically not correct since GD is not a global optimization algorithm. Maybe mention that you try to find a stationary point\nrho (Eq. 2) and lambda (Eq. 4) are not defined\n\nPage 4: \nAlgorithm 1 and 2 and related equations in the main text: it should be H_hat instead of H.\n\nBackground\n“Momemtum GD exhibits somewhat better resistance to poor scaling of the objective function”\nTo be precise the improvement is quadratic for convex functions. Note that Goh might not be the best reference to cite as the article focuses on quadratic function. Consider citing the lecture notes from Nesterov.\n\nSection 2.2\nThis section is perhaps a bit confusing at first as the authors discuss the general case of a multivalue loss function. Consider moving your last comment to the beginning of the section.\n\nSection 2.3\nAs a side remark, the work of Dauphin does not rely on the Gauss-Newton approximation but a different PSD matrix, this is probably worth mentioning.\n\nMinor comment: The title is rather bold and not necessarily precise since the stepsize of curveball is not particularly small e.g. in Fig 1.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting research direction but the paper needs a lot more work before publication", "review": "This paper proposes an approximate second-order method with low computational cost. A common pitfall of second-order methods is the computation (and perhaps inversion) of the Hessian matrix. While this can be avoided by instead relying on Hessian-vector products as done in CG, it typically still requires several iterations. Instead, the authors suggest a simpler approach that relies on one single gradient step and a warm start strategy. The authors points out that the resulting algorithm resembles a momentum method. They also provide some simple convergence proofs on quadratics and benchmark their method to train deep neural networks.\n\nWhile I find the research direction interesting, the execution is rather clumsy and many details are not sufficiently motivated. Finally, there is a lot of relevant work in the optimization community that is not discussed in this paper, see detailed comments and references below.\n\n1) Method\nThe derivation of the method is very much driven on a set of heuristics without theoretical guarantees. In order to derive the update of the proposed method, the authors rely on three heuristics:\na) The first is to reuse the previous search direction z as a warm-start. The authors argue that this might be beneficial if If z does not change abruptly. In the early phase, the gradient norm is likely to be large and thus z will change significantly. One might also encounter regions of high curvature where the direction of z might change quickly from one iteration to the next.\nThe \"warm start\" at s_{t-1} is also what yields the momentum term, what interpretation can you give to this choice?\n\nb) The second step interleaves the updates of z and w instead of first finding the optimum z. This amounts to just running one iteration of CG but it is rather unclear why one iteration is an appropriate number. It seems one could instead some adaptive strategy where CG with a fixed accuracy. One could potentially see if allowing larger errors at the beginning of the optimization process might still allow for the method to converge. This is for instance commonly done with the batch-size of first-order method. Gradually increasing the batch-size and therefore reducing the error as one gets close to the optimum can still yield to a converging algorithm, see e.g. \nFriedlander, M. P., & Schmidt, M. (2012). Hybrid deterministic-stochastic methods for data fitting. SIAM Journal on Scientific Computing, 34(3), A1380-A1405.\n\nc) The third step consists in replacing CG with gradient descent.\n\"If CG takes N steps on average, then Algorithm 2 will be slower than SGD by a factor of at least N, which can easily be an order of magnitude\".\nFirst, the number of outer iterations may be a lot less for the Hessian-free method than for SGD so this does not seem to be a valid argument. Please comment.\nSecond, I would like to see a discussion of the convergence rate of solving (12) inexactly with krylov subspace methods. Note that Lanczos yields an accelerated rate while GD does not. So the motivation for switching to GD should be made clearer.\n\nd) The fourth step introduces a factor rho that decays z at each step. I’m not really sure this makes sense even heuristically. The full update of the algorithm developed by the author is:\nw_{t+1} = w_t - beta nabla f + (rho I - beta H) (w_t - w_{t-1}).\nThe momentum term therefore gets weighted by (rho I - beta H). What is the meaning of this term? The -beta H term weights the momentum according to the curvature of the objective function. Given the lack of theoretical support for this idea, I would at least expect a practical reason back up by some empirical evidence that this is a sensible thing to do.\nThis is especially important given that you claim to decay rho therefore giving more importance to the curvature term.\nFinally, why would this be better than simply using CG on a trust-region model? (Recall that Lanczos yields an accelerated linear rate while GD does not).\n\n2) Convergence analysis\na) The analysis is only performed on a quadratic while the author clearly target non-convex functions, this should be made clear in the main text. Also see references below (comment #3) regarding a possible extension to non-convex functions.\nb) The authors should check the range of allowed values for alpha and beta. It appears the rate would scale with the square root of the condition number, please confirm, this is an important detail. I also think that the constant is not as good as Heavy-ball on a quadratic (see e.g. http://pages.cs.wisc.edu/~brecht/cs726docs/HeavyBallLinear.pdf), please comment.\nc) Sub-sampling of the Hessian and gradients is not discussed at all (but used in the experiments). Please add a discussion and consider extending the proof (again, see references given below).\n\n3) Convergence Heavy-ball\nThe authors emphasize the similarity of their approach to Heavy-ball. They cite the results of Loizou & Richtarik 2017. Note that they are earlier results for quadratic functions such as \nLessard, L., Recht, B., & Packard, A. (2016). Analysis and design of optimization algorithms via integral quadratic constraints. SIAM Journal on Optimization, 26(1), 57-95.\nFlammarion, N., & Bach, F. (2015, June). From averaging to acceleration, there is only a step-size. In Conference on Learning Theory (pp. 658-695).\nThe novelty of the bounds derived in Loizou & Richtarik 2017 is that they apply in stochastic settings.\nFinally, there are results for non-convex functions such convergence to a stationary point, see\nZavriev, S. K., & Kostyuk, F. V. (1993). Heavy-ball method in nonconvex optimization problems. Computational Mathematics and Modeling, 4(4), 336-341.\nAlso on page 2, \"Momentum GD ... can be shown to have faster convergence than GD\". It should be mentioned that this only hold for (strongly) convex functions!\n\n4) Experiments\na) Consider showing the gradient norms. \nb) it looks like the methods have not yet converged in Fig 2 and 3.\nc) Second order benchmark:\nIt would be nice to compare to a method that does not use the GN matrix but the true or subsampled Hessian (like Trust Region/Cubic Regularization) methods given below.\nWhy is BFGS in Rosenbrock but not in NN plots?\nd) \"Batch normalization (which is known to improve optimization)\" \nThis statement requires a reference such as\nTowards a Theoretical Understanding of Batch Normalization\nKohler et al… - arXiv preprint arXiv:1805.10694, 2018\n\n5) Related Work\nThe related work should include Cubic Regularization and Trust Region methods since they are among the most prominent second order algorithms. Consider citing Conn et al. 2000 Trust Region,  Nesterov 2006 Cubic regularization, Cartis et al. 2011 ARC.\nRegarding sub-sampling: Kohler&Lucchi 2017: Stochastic Cubic Regularization for non-convex optimization and Xu et al.: Newton-type methods for non-convex optimization under inexact hessian information.\n\n6) More comments\n\nPage 2\nPolyak 1964 should be cited  where momentum is discussed.\n\"Perhaps the simplest algorithm to optimize Eq. 1 is Gradient Descent\". This is technically not correct since GD is not a global optimization algorithm. Maybe mention that you try to find a stationary point\nrho (Eq. 2) and lambda (Eq. 4) are not defined\n\nPage 4: \nAlgorithm 1 and 2 and related equations in the main text: it should be H_hat instead of H.\n\nBackground\n“Momemtum GD exhibits somewhat better resistance to poor scaling of the objective function”\nTo be precise the improvement is quadratic for convex functions. Note that Goh might not be the best reference to cite as the article focuses on quadratic function. Consider citing the lecture notes from Nesterov.\n\nSection 2.2\nThis section is perhaps a bit confusing at first as the authors discuss the general case of a multivalue loss function. Consider moving your last comment to the beginning of the section.\n\nSection 2.3\nAs a side remark, the work of Dauphin does not rely on the Gauss-Newton approximation but a different PSD matrix, this is probably worth mentioning.\n\nMinor comment: The title is rather bold and not necessarily precise since the stepsize of curveball is not particularly small e.g. in Fig 1.\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1540718809892}], "openreview_url": "https://openreview.net/forum?id=Sygx4305KQ", "arxiv_id": "1805.08095", "paper_pdf": "papers/Sygx4305KQ.pdf", "paper_pdf_sha256": "d41e9bfffb0357512e05446983925195dd5f02a8abd22512172c6a81369dcbb5", "paper_pdf_bytes": 2346524, "paper_pdf_source": "openreview", "code_url": "https://github.com/jotaf98/curveball", "code_repository": "jotaf98/curveball", "code_commit": "1dc37325382c12e3fc9b2e7e27c47e6d7a17021a", "code_archive": "repos/Sygx4305KQ.zip", "code_archive_sha256": "7d7433458348ca7b73a461edfc7b926a18453def06cd9543c8227492bb79c16a", "code_archive_bytes": 43815, "code_file_count": 35, "code_extensions": {".m": 21, ".cu": 10, ".cpp": 2, ".sh": 1, ".hpp": 1}, "github_disk_usage_kb": 37, "github_languages": {"MATLAB": 58871, "Cuda": 42788, "C++": 1212, "Shell": 548}, "github_archived": false, "github_pushed_at": "2018-10-29T13:59:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/small-steps-and-giant-leaps-minimal-newton"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HyEi7bWR-", "year": 2018, "status": "rejected", "title": "Orthogonal Recurrent Neural Networks with Scaled Cayley Transform", "authors": ["Kyle Helfrich", "Devin Willmott", "Qiang Ye"], "authorids": ["kyle.helfrich@uky.edu", "devin.willmott@uky.edu", "qiang.ye@uky.edu"], "authors_source": "OpenReview API", "abstract": "Recurrent Neural Networks (RNNs) are designed to handle sequential data but suffer from vanishing or exploding gradients.  Recent work on Unitary Recurrent Neural Networks (uRNNs) have been used to address this issue and in some cases, exceed the capabilities of Long Short-Term Memory networks (LSTMs).  We propose a simpler and novel update scheme to maintain orthogonal recurrent weight matrices without using complex valued matrices. This is done by parametrizing with a skew-symmetric matrix using the Cayley transform. Such a parametrization is unable to represent matrices with negative one eigenvalues, but this limitation is overcome by scaling the recurrent weight matrix by a diagonal matrix consisting of ones and negative ones.  The proposed training scheme involves a straightforward gradient calculation and update step. In several experiments, the proposed scaled Cayley orthogonal recurrent neural network (scoRNN) achieves superior results with fewer trainable parameters than other unitary RNNs.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkLk8W9lM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper664/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This manuscript introduce a scheme for learning the recurrent parameter matrix in a neural network that uses the Cayley transform and a scaling weight matrix. This scheme leads to good performance on sequential data tasks and requires fewer parameters than other techniques\n\nComments:\n-- It’s not clear to me how D is determined for each test. Given the definition in Theorem 3.1 it seems like you would have to have some knowledge of how many eigenvalues in W you expect to be close to -1. \n-- For the copying and adding problem test cases, it might be useful to clarify or cite something clarifying that the failure mode RNNs run into with temporal ordering problems is an exploding gradient, rather than any other pathological training condition, just to make it clear why these experiments are relevant.\n-- The ylabel in Figure 1 is “Test Loss” which I didn’t see defined. Is this test loss the cross entropy? If so, I think it would be more effective to label the plot with that.\n-- The plots in figure 1 and 2 have different colors to represent the same set of techniques. I would suggest keeping a  consistent color scheme\n-- It looks like in Figure 1 the scoRNN is outperformed by the uRNN in the long run in spite of the scoRNN convergence being smoother, which should be clarified.\n-- It looks like in Figure 2 the scoRNN is outperformed by the LSTM across the board, which should be clarified.\n-- How is test set accuracy defined in section 5.3? Classifying digits? Recreating digits? \n-- When discussing table 1, the manuscript mentions scoRNN and Restricted-capacity uRNN have similar performance for 16k parameters and then state that scoRNN has the best test accuracy at 96.2%. However, there is no example for restricted-capacity uRNN with 69k parameters to show that the performance of restricted-capacity uRNN doesn't also increase similarly with more parameters.\n-- Overall it’s unclear to me how to completely determine the benefit of this technique over the others because, for each of the tests, different techniques may have superior performance. For instance, LSTM performs best in 5.2 and in 5.3 for the MNIST test accuracy. scoRNN and Restricted-capacity uRNN perform similarly for permuted MNIST Test Accuracy in 5.3. Finally, scoRNN seems to far outperform the other techniques in table 2 on the TIMIT speech dataset. I don’t understand the significance of each test and why the relative performance of the techniques vary from one to the other.\n-- For example, the manuscript seems to be making the case that the scoRNN gradients are more stable than those of a uRNN, but all of the results are presented in terms of network accuracy and not gradient stability. You can sort of see that generally the convergence is more gradual for the scoRNN than the uRNN from the training graphs but it'd be nice if there was an actual comparison of the stability of the gradients during training (as in Figure 4 of the Arjovsky 2016 paper being compared to for instance) just to make it really clear.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of \"Orthogonal Recurrent Neural Networks with Scaled Cayley Transform\"", "rating": "7: Good paper, accept", "review": "This manuscript introduce a scheme for learning the recurrent parameter matrix in a neural network that uses the Cayley transform and a scaling weight matrix. This scheme leads to good performance on sequential data tasks and requires fewer parameters than other techniques\n\nComments:\n-- It’s not clear to me how D is determined for each test. Given the definition in Theorem 3.1 it seems like you would have to have some knowledge of how many eigenvalues in W you expect to be close to -1. \n-- For the copying and adding problem test cases, it might be useful to clarify or cite something clarifying that the failure mode RNNs run into with temporal ordering problems is an exploding gradient, rather than any other pathological training condition, just to make it clear why these experiments are relevant.\n-- The ylabel in Figure 1 is “Test Loss” which I didn’t see defined. Is this test loss the cross entropy? If so, I think it would be more effective to label the plot with that.\n-- The plots in figure 1 and 2 have different colors to represent the same set of techniques. I would suggest keeping a  consistent color scheme\n-- It looks like in Figure 1 the scoRNN is outperformed by the uRNN in the long run in spite of the scoRNN convergence being smoother, which should be clarified.\n-- It looks like in Figure 2 the scoRNN is outperformed by the LSTM across the board, which should be clarified.\n-- How is test set accuracy defined in section 5.3? Classifying digits? Recreating digits? \n-- When discussing table 1, the manuscript mentions scoRNN and Restricted-capacity uRNN have similar performance for 16k parameters and then state that scoRNN has the best test accuracy at 96.2%. However, there is no example for restricted-capacity uRNN with 69k parameters to show that the performance of restricted-capacity uRNN doesn't also increase similarly with more parameters.\n-- Overall it’s unclear to me how to completely determine the benefit of this technique over the others because, for each of the tests, different techniques may have superior performance. For instance, LSTM performs best in 5.2 and in 5.3 for the MNIST test accuracy. scoRNN and Restricted-capacity uRNN perform similarly for permuted MNIST Test Accuracy in 5.3. Finally, scoRNN seems to far outperform the other techniques in table 2 on the TIMIT speech dataset. I don’t understand the significance of each test and why the relative performance of the techniques vary from one to the other.\n-- For example, the manuscript seems to be making the case that the scoRNN gradients are more stable than those of a uRNN, but all of the results are presented in terms of network accuracy and not gradient stability. You can sort of see that generally the convergence is more gradual for the scoRNN than the uRNN from the training graphs but it'd be nice if there was an actual comparison of the stability of the gradients during training (as in Figure 4 of the Arjovsky 2016 paper being compared to for instance) just to make it really clear.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511818717939}, {"id": "rkQbrzqxM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper664/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper suggests an RNN reparametrization of the recurrent weights with a skew-symmetric matrix using Cayley transform to keep the recurrent weight matrix orthogonal. They suggest that they reparametrization leads to superior performance compare to other forms of Unitary Recurrent Networks.\n\nI think the paper is well-written.  Authors have discussed previous works adequately and provided enough insight and motivation about the proposed method.\n\nI have two questions from authors:\n\n1- What are the hyperparameters that you optimized in experiments?\n\n2- How sensitive is the results to the number of -1 in the diagonal matrix?\n\n3- ince the paper is not about compression, it might be unfair to limit the number of hidden units in LSTMs just to match the number of parameters to RNNs. In MNIST experiment, for example, better numbers are reported for larger LSTMs. I think matching the number of hidden units could be helpful. Also, one might want to know if the scoRNN is still superior in the regime where the number of hidden units is about 1000. I appreciate if authors can provide more results in these settings.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An alternative parametrization of Unitary RNNs", "rating": "6: Marginally above acceptance threshold", "review": "This paper suggests an RNN reparametrization of the recurrent weights with a skew-symmetric matrix using Cayley transform to keep the recurrent weight matrix orthogonal. They suggest that they reparametrization leads to superior performance compare to other forms of Unitary Recurrent Networks.\n\nI think the paper is well-written.  Authors have discussed previous works adequately and provided enough insight and motivation about the proposed method.\n\nI have two questions from authors:\n\n1- What are the hyperparameters that you optimized in experiments?\n\n2- How sensitive is the results to the number of -1 in the diagonal matrix?\n\n3- ince the paper is not about compression, it might be unfair to limit the number of hidden units in LSTMs just to match the number of parameters to RNNs. In MNIST experiment, for example, better numbers are reported for larger LSTMs. I think matching the number of hidden units could be helpful. Also, one might want to know if the scoRNN is still superior in the regime where the number of hidden units is about 1000. I appreciate if authors can provide more results in these settings.\n\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511822587037}, {"id": "HyGpBPslM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper664/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is clearly written, with a good coverage of previous relevant literature. \nThe contribution itself is slightly incremental, as several different parameterization of orthogonal or almost-orthogonal weight matrices for RNN have been introduced.\nTherefore, the paper must show that this new method performs better in some way compared with previous methods. They show that the proposed method is competitive on several datasets and a clear winner on one task: MSE on TIMIT.\n\nPros:\n1. New, relatively simple method for learning orthogonal weight matrices for RNN\n\n2. Clearly written\n\n3. Quite good results on several relevant tasks.\n\nCons:\n1. Technical novelty is somewhat limited\n\n2. Experiments do not evaluate run time, memory use, computational complexity, or stability. Therefore it is more difficult to make comparisons: perhaps restricted-capacity uRNN is 10 times faster than the proposed method?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel parametrization of RNNs allows representing orthogonal weight matrices relatively easily", "rating": "5: Marginally below acceptance threshold", "review": "The paper is clearly written, with a good coverage of previous relevant literature. \nThe contribution itself is slightly incremental, as several different parameterization of orthogonal or almost-orthogonal weight matrices for RNN have been introduced.\nTherefore, the paper must show that this new method performs better in some way compared with previous methods. They show that the proposed method is competitive on several datasets and a clear winner on one task: MSE on TIMIT.\n\nPros:\n1. New, relatively simple method for learning orthogonal weight matrices for RNN\n\n2. Clearly written\n\n3. Quite good results on several relevant tasks.\n\nCons:\n1. Technical novelty is somewhat limited\n\n2. Experiments do not evaluate run time, memory use, computational complexity, or stability. Therefore it is more difficult to make comparisons: perhaps restricted-capacity uRNN is 10 times faster than the proposed method?", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511908793946}], "openreview_url": "https://openreview.net/forum?id=HyEi7bWR-", "arxiv_id": "1707.09520", "paper_pdf": "papers/HyEi7bWR-.pdf", "paper_pdf_sha256": "31ec6aad5b3148eb8a25485574a1b299fcf6cfc7bd774161abf6fd7fcfdbc71d", "paper_pdf_bytes": 676036, "paper_pdf_source": "openreview", "code_url": "https://github.com/SpartinStuff/scoRNN", "code_repository": "SpartinStuff/scoRNN", "code_commit": "d390c9bac62c510963ff90d386fb02beccff0a1e", "code_archive": "repos/HyEi7bWR-.zip", "code_archive_sha256": "c89c8a0d7597fa0f6d04b4f31b3a0d3f1ba9f5b19be80c929c69ba2c42de1726", "code_archive_bytes": 12462, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 26, "github_languages": {"Python": 32173}, "github_archived": false, "github_pushed_at": "2024-05-24T09:14:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/orthogonal-recurrent-neural-networks-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oVd1y7ilTk", "year": 2026, "status": "rejected", "title": "On Designing Diffusion Autoencoders for Efficient Generation and Representation Learning", "authors": ["Magdalena Proszewska", "Nikolay Malkin", "Siddharth N"], "authorids": ["~Magdalena_Proszewska1", "~Nikolay_Malkin1", "~Siddharth_N1"], "authors_source": "OpenReview API", "abstract": "Diffusion autoencoders (DAs) are variants of diffusion generative models that use an input-dependent latent variable to capture representations alongside the diffusion process. These representations can be used for tasks such as downstream classification, controllable generation, and interpolation. However, the generative performance of DAs relies heavily on how well the prior distribution over the latent variables can be modelled and subsequently sampled from. Better generative modelling is also the goal of another class of diffusion models—those that learn their forward (noising) process. While effective at adjusting the noise process in an input-dependent manner, they must satisfy additional constraints derived from the terminal conditions of the diffusion process. Here, we draw a connection between these two classes of models and show that certain design decisions (latent variable choice, conditioning method, etc.) in the DA framework—leading to a model we term DMZ—enable effective representations as evaluated on downstream tasks, including domain transfer, as well as more efficient modelling and generation with fewer denoising steps compared to standard diffusion models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zPasy4Iy54", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9087/Reviewer_twVn"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "The authors propose a diffusion autoencoder framework, DMZ, with carefully designed strategies such as latent variable choice, conditioning methods, and more. The paper provides a comprehensive study of each component’s design choices. Empirically, DMZ shows consistently strong performance in both unconditional generation and representation learning. The authors also demonstrate that DMZ can be easily applied to multimodal tasks (e.g., image-to-image translation), highlighting the framework’s flexibility.", "review_text": "The authors propose a diffusion autoencoder framework, DMZ, with carefully designed strategies such as latent variable choice, conditioning methods, and more. The paper provides a comprehensive study of each component’s design choices. Empirically, DMZ shows consistently strong performance in both unconditional generation and representation learning. The authors also demonstrate that DMZ can be easily applied to multimodal tasks (e.g., image-to-image translation), highlighting the framework’s flexibility.", "strengths": "* The paper is clear and easy to follow. The comprehensive experiments convincingly isolate and evaluate the effects of each design choice.\n* The DMZ framework is fairly general: it performs well in unconditional generation and representation learning, and it can be extended to handle multimodal tasks such as image-to-image translation.", "weaknesses": "* The cross-attention conditioning design is already widely used in modern diffusion transformers [1-2]. The current validation relies on an older U-Net architecture, so this component does not constitute a significant contribution by itself.\n* The choice of latent dimensionality $|z|$ appears ad hoc. For generation tasks it is guided by the label-space size (suggesting that relatively low dimensions yield better generation quality), whereas representation learning for downstream tasks benefits from more informative, higher-dimensional latents. This implies separate designs for different use cases within DMZ.\n* The effectiveness for generation is not fully convincing. If is tied to a (binary) label space, it can logically degenerate to a one-dimensional label with low dimensions. Sampling from this prior is then akin to sampling in label space for conditional generation. Although DMZ does not directly rely on a labeling function $f:X\\rightarrow Y$, could clustering be used to produce labels that achieve similar behavior in the \"unconditional\" setting? This would suggest DMZ may not be learning strong representations in these scenarios.\n* DMZ shows limited compatibility with DDIM in Table 10. It would help to evaluate DMZ with more recent denoising approaches and architectures such as DiT and SiT [1–2].\n* Extending experiments to larger benchmarks (e.g., ImageNet) would further strengthen the work’s claims and external validity.\n\n[1] Scalable Diffusion Models with Transformers\n[2] Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers", "questions": "See Weaknesses 3–5.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a diffusion autoencoder framework, DMZ, with carefully designed strategies such as latent variable choice, conditioning methods, and more. The paper provides a comprehensive study of each component’s design choices. Empirically, DMZ shows consistently strong performance in both unconditional generation and representation learning. The authors also demonstrate that DMZ can be easily applied to multimodal tasks (e.g., image-to-image translation), highlighting the framework’s flexibility.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "* The paper is clear and easy to follow. The comprehensive experiments convincingly isolate and evaluate the effects of each design choice.\n* The DMZ framework is fairly general: it performs well in unconditional generation and representation learning, and it can be extended to handle multimodal tasks such as image-to-image translation.", "weaknesses": "* The cross-attention conditioning design is already widely used in modern diffusion transformers [1-2]. The current validation relies on an older U-Net architecture, so this component does not constitute a significant contribution by itself.\n* The choice of latent dimensionality $|z|$ appears ad hoc. For generation tasks it is guided by the label-space size (suggesting that relatively low dimensions yield better generation quality), whereas representation learning for downstream tasks benefits from more informative, higher-dimensional latents. This implies separate designs for different use cases within DMZ.\n* The effectiveness for generation is not fully convincing. If is tied to a (binary) label space, it can logically degenerate to a one-dimensional label with low dimensions. Sampling from this prior is then akin to sampling in label space for conditional generation. Although DMZ does not directly rely on a labeling function $f:X\\rightarrow Y$, could clustering be used to produce labels that achieve similar behavior in the \"unconditional\" setting? This would suggest DMZ may not be learning strong representations in these scenarios.\n* DMZ shows limited compatibility with DDIM in Table 10. It would help to evaluate DMZ with more recent denoising approaches and architectures such as DiT and SiT [1–2].\n* Extending experiments to larger benchmarks (e.g., ImageNet) would further strengthen the work’s claims and external validity.\n\n[1] Scalable Diffusion Models with Transformers\n[2] Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers", "questions": "See Weaknesses 3–5.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761990483770}, {"id": "MOFa2NMJfJ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9087/Reviewer_8Gtz"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 3, "summary": "This paper proposes DMZ, a design for diffusion autoencoders that aims to improve both generation efficiency and representation learning. By incorporating a input-dependent encoder, DMZ explores the distribution choices, conditioning mechanisms, and learning strategies to enhance the performance of diffusion autoencoders. Experiments on CIFAR-10 and CelebA demonstrate the effectiveness of DMZ and its potential ability for style transfer and representation learning.", "review_text": "This paper proposes DMZ, a design for diffusion autoencoders that aims to improve both generation efficiency and representation learning. By incorporating a input-dependent encoder, DMZ explores the distribution choices, conditioning mechanisms, and learning strategies to enhance the performance of diffusion autoencoders. Experiments on CIFAR-10 and CelebA demonstrate the effectiveness of DMZ and its potential ability for style transfer and representation learning.", "strengths": "- The focus on diffusion autoencoders is timely and relevant, addressing the need for efficient generation and representation learning.\n- The illustrations and explanations of DM/DA are clear.\n- The benchmarking tasks and datasets are appropriate for evaluating the proposed method.", "weaknesses": "- The motivation and contribution of DMZ is unclear.\n- The algorithmic details of DMZ are insufficient.\n- The performance of DMZ is underwhelming compared to existing methods.", "questions": "0. **DMZ meaning.** What does DMZ stand for? The acronym is not explained in the paper.\n\n1. **Motivation and contribution.** I feel confused about the motivation and contribution of DMZ, and believe the writing could be potentially largely improved for clarity. In the introduction section, the authors claim \"to draw a connection between DMs and DAs\". However, I could not find any discussion or analysis on DA/DMZs in the rest of the paper. How are DMZ and DA different? Is it the contribution of DMZ to propose a new DA framework, or explore the design space of DAs? Could the authors clarify the main contribution of this work?\n\n2. **Algorithmic details.** The algorithmic details of DMZ are insufficiently described. Only Eq.(5) describes the training objective of DA. Does DMZ use the same training objective as DA? Additionally, could the authors provide more details about the newly-proposed components in DMZ, including conditioning mechanisms and learning strategies? A more comprehensive description of the algorithm would help readers better understand the proposed method.\n\n3. **Performance comparison.** The performance of DMZ seems underwhelming compared to existing methods. In Table 1, DMZ achieves worse performance on CIFAR-10 compared to DDPMs. In Table 3, DMZ achieves worse performance compared to DDBMs. Could the authors provide more analysis on why DMZ underperforms compared to these methods? Are there any specific limitations or challenges in the DMZ design that lead to this performance gap?\n\n4. **Inconsistent experimental setup.** In Figure 3 the authors compare NLL on CIFAR-10 and FID on CelebA. Are there any specific reasons for using different datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes DMZ, a design for diffusion autoencoders that aims to improve both generation efficiency and representation learning. By incorporating a input-dependent encoder, DMZ explores the distribution choices, conditioning mechanisms, and learning strategies to enhance the performance of diffusion autoencoders. Experiments on CIFAR-10 and CelebA demonstrate the effectiveness of DMZ and its potential ability for style transfer and representation learning.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "- The focus on diffusion autoencoders is timely and relevant, addressing the need for efficient generation and representation learning.\n- The illustrations and explanations of DM/DA are clear.\n- The benchmarking tasks and datasets are appropriate for evaluating the proposed method.", "weaknesses": "- The motivation and contribution of DMZ is unclear.\n- The algorithmic details of DMZ are insufficient.\n- The performance of DMZ is underwhelming compared to existing methods.", "questions": "0. **DMZ meaning.** What does DMZ stand for? The acronym is not explained in the paper.\n\n1. **Motivation and contribution.** I feel confused about the motivation and contribution of DMZ, and believe the writing could be potentially largely improved for clarity. In the introduction section, the authors claim \"to draw a connection between DMs and DAs\". However, I could not find any discussion or analysis on DA/DMZs in the rest of the paper. How are DMZ and DA different? Is it the contribution of DMZ to propose a new DA framework, or explore the design space of DAs? Could the authors clarify the main contribution of this work?\n\n2. **Algorithmic details.** The algorithmic details of DMZ are insufficiently described. Only Eq.(5) describes the training objective of DA. Does DMZ use the same training objective as DA? Additionally, could the authors provide more details about the newly-proposed components in DMZ, including conditioning mechanisms and learning strategies? A more comprehensive description of the algorithm would help readers better understand the proposed method.\n\n3. **Performance comparison.** The performance of DMZ seems underwhelming compared to existing methods. In Table 1, DMZ achieves worse performance on CIFAR-10 compared to DDPMs. In Table 3, DMZ achieves worse performance compared to DDBMs. Could the authors provide more analysis on why DMZ underperforms compared to these methods? Are there any specific limitations or challenges in the DMZ design that lead to this performance gap?\n\n4. **Inconsistent experimental setup.** In Figure 3 the authors compare NLL on CIFAR-10 and FID on CelebA. Are there any specific reasons for using different datasets?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761977005299}, {"id": "tM3sDpx8zW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9087/Reviewer_s5wS"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces DMZ, kind of diffusion autoencoders.\nDMZ aims to improve the generative quality of diffusion models by guiding the sampling process using the latent representation $z$ of $x_0$.\nThe model is trained without any additional loss terms, following the standard DA training objective.\nUnlike conventional diffusion autoencoders, DMZ does not require an auxiliary latent sampler.\nInstead, it directly samples the latent variable z from a Bernoulli distribution, which improves sampling efficiency.\nFurthermore, the authors empirically show that conditioning only the Key and Value components of the cross-attention layers on z leads to better performance than other conditioning strategies.", "review_text": "This paper introduces DMZ, kind of diffusion autoencoders.\nDMZ aims to improve the generative quality of diffusion models by guiding the sampling process using the latent representation $z$ of $x_0$.\nThe model is trained without any additional loss terms, following the standard DA training objective.\nUnlike conventional diffusion autoencoders, DMZ does not require an auxiliary latent sampler.\nInstead, it directly samples the latent variable z from a Bernoulli distribution, which improves sampling efficiency.\nFurthermore, the authors empirically show that conditioning only the Key and Value components of the cross-attention layers on z leads to better performance than other conditioning strategies.", "strengths": "- Unlike previous diffusion autoencoders, DMZ does not rely on an auxiliary latent sampler.\nBy directly sampling $z$ from a Bernoulli distribution, the method enables computationally efficient sampling.\n\n- The learned latent representation is shown to be effective even in a multi-modal framework, indicating its potential generality beyond standard generation tasks.\n\n- The proposed DDPM-based approach demonstrates clear improvements in generation quality, particularly when using a small number of denoising steps.", "weaknesses": "- It is unclear how the latent variable can be sampled from a Bernoulli distribution without any prior regularization.\nIn standard DA frameworks, auxiliary latent samplers (such as [1,2]) or additional regularization terms (such as [3]) are typically used to properly model the latent prior.\nWithout such mechanisms, it is not evident how the encoder output would naturally follow a Bernoulli prior.\nThis appears to be a critical limitation of the proposed method.\n\n- The effect of conditioning z only on the Key and Value in the attention layers is not clearly explained.\nWhile the authors report that this approach outperforms the alternative of jointly conditioning with t, the reason for this improvement remains unclear.\nAdditional analysis or experiments would strengthen this claim.\n\n---------\n[1] [CVPR22] Diffusion autoencoders: Toward a meaningful and decodable representation\n\n[2] [NeurIPS 22] Unsupervised representation learning from pre-trained diffusion probabilistic models\n\n[3] [ICML23] Infodiffusion: Representation learning using information maximizing diffusion models", "questions": "- Was any specific encoder architecture or constraint introduced to make the encoder output binary? How is this discreteness enforced during training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces DMZ, kind of diffusion autoencoders.\nDMZ aims to improve the generative quality of diffusion models by guiding the sampling process using the latent representation $z$ of $x_0$.\nThe model is trained without any additional loss terms, following the standard DA training objective.\nUnlike conventional diffusion autoencoders, DMZ does not require an auxiliary latent sampler.\nInstead, it directly samples the latent variable z from a Bernoulli distribution, which improves sampling efficiency.\nFurthermore, the authors empirically show that conditioning only the Key and Value components of the cross-attention layers on z leads to better performance than other conditioning strategies.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Unlike previous diffusion autoencoders, DMZ does not rely on an auxiliary latent sampler.\nBy directly sampling $z$ from a Bernoulli distribution, the method enables computationally efficient sampling.\n\n- The learned latent representation is shown to be effective even in a multi-modal framework, indicating its potential generality beyond standard generation tasks.\n\n- The proposed DDPM-based approach demonstrates clear improvements in generation quality, particularly when using a small number of denoising steps.", "weaknesses": "- It is unclear how the latent variable can be sampled from a Bernoulli distribution without any prior regularization.\nIn standard DA frameworks, auxiliary latent samplers (such as [1,2]) or additional regularization terms (such as [3]) are typically used to properly model the latent prior.\nWithout such mechanisms, it is not evident how the encoder output would naturally follow a Bernoulli prior.\nThis appears to be a critical limitation of the proposed method.\n\n- The effect of conditioning z only on the Key and Value in the attention layers is not clearly explained.\nWhile the authors report that this approach outperforms the alternative of jointly conditioning with t, the reason for this improvement remains unclear.\nAdditional analysis or experiments would strengthen this claim.\n\n---------\n[1] [CVPR22] Diffusion autoencoders: Toward a meaningful and decodable representation\n\n[2] [NeurIPS 22] Unsupervised representation learning from pre-trained diffusion probabilistic models\n\n[3] [ICML23] Infodiffusion: Representation learning using information maximizing diffusion models", "questions": "- Was any specific encoder architecture or constraint introduced to make the encoder output binary? How is this discreteness enforced during training?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed.", "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761934786218}, {"id": "TKfHu9VLGN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9087/Reviewer_yKM4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposed a novel diffusion autoencoder technique, DMZ, based on the empirical analysis of latent z. By setting z as discrete binary encoding, using cross-attention as conditioning, and using dense latent variable dimension, it achieved faster convergence, better generation quality, as well as without any prior or loss function. The experiment results and ablation study results demonstrated this conclusion.", "review_text": "The paper proposed a novel diffusion autoencoder technique, DMZ, based on the empirical analysis of latent z. By setting z as discrete binary encoding, using cross-attention as conditioning, and using dense latent variable dimension, it achieved faster convergence, better generation quality, as well as without any prior or loss function. The experiment results and ablation study results demonstrated this conclusion.", "strengths": "1. DMZ firstly established the theoretical fundation of learnable forward process of diffusion autoencoder.\n\n2. The design motivation of z is reasonable and inspiring. Correspondingly, the improvements (binary z, etc) are effective in the following experimental results. \n\n3. The experiments are comprehensive, not only demonstrating the effectiveness of each z components, but also extend the task to the other tasks dependent on z (such as stretch2pic). The ablation studies are also reasonable and promising.", "weaknesses": "1. Although the learnable forward process is proposed with good motivation, there lacks formal induction or convergence analysis.\n\n2. Benchmark analysis on high-resolution datasets is recommended. \n\n3. There lack interpretability analysis of latent variables, incluing the semantic understanding, differentability, and the combinmation capability as condition for discrete binary z.\n\n4. The flow of method section should be adjusted. DA and learnable forward process should be discussed separately with suitable connections.", "questions": "1. Can the authors discuss the connection between DMZ and REPA, which is also a promising baseline in diffusion representation field. In my view, binary z can be seen as a simplified type of external embedding guidance. If so, how about extending z to boarder fields like it is in REPA?\n\n2. How about the experimental results in high-resolution datasets? How does the efficiency change with the lantent dimension? \n\n3. Can the author propose more analysis and results on different types of priors rather than Bernoulli? \n\n4. Is there any insight towards the design of conditioning z by cross attention instead of conditioning it from the residue network?\n\n5. Can DMZ be combined with the current SOTA diffusion models, such as cosistency model and rectified flow?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed a novel diffusion autoencoder technique, DMZ, based on the empirical analysis of latent z. By setting z as discrete binary encoding, using cross-attention as conditioning, and using dense latent variable dimension, it achieved faster convergence, better generation quality, as well as without any prior or loss function. The experiment results and ablation study results demonstrated this conclusion.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. DMZ firstly established the theoretical fundation of learnable forward process of diffusion autoencoder.\n\n2. The design motivation of z is reasonable and inspiring. Correspondingly, the improvements (binary z, etc) are effective in the following experimental results. \n\n3. The experiments are comprehensive, not only demonstrating the effectiveness of each z components, but also extend the task to the other tasks dependent on z (such as stretch2pic). The ablation studies are also reasonable and promising.", "weaknesses": "1. Although the learnable forward process is proposed with good motivation, there lacks formal induction or convergence analysis.\n\n2. Benchmark analysis on high-resolution datasets is recommended. \n\n3. There lack interpretability analysis of latent variables, incluing the semantic understanding, differentability, and the combinmation capability as condition for discrete binary z.\n\n4. The flow of method section should be adjusted. DA and learnable forward process should be discussed separately with suitable connections.", "questions": "1. Can the authors discuss the connection between DMZ and REPA, which is also a promising baseline in diffusion representation field. In my view, binary z can be seen as a simplified type of external embedding guidance. If so, how about extending z to boarder fields like it is in REPA?\n\n2. How about the experimental results in high-resolution datasets? How does the efficiency change with the lantent dimension? \n\n3. Can the author propose more analysis and results on different types of priors rather than Bernoulli? \n\n4. Is there any insight towards the design of conditioning z by cross attention instead of conditioning it from the residue network?\n\n5. Can DMZ be combined with the current SOTA diffusion models, such as cosistency model and rectified flow?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761791565635}], "openreview_url": "https://openreview.net/forum?id=oVd1y7ilTk", "arxiv_id": "2506.00136", "paper_pdf": "papers/oVd1y7ilTk.pdf", "paper_pdf_sha256": "79bed6630dc80a05b2da429c077a3f2632d1fffe35ab7f3b45ed02b0f854630f", "paper_pdf_bytes": 9032494, "paper_pdf_source": "openreview", "code_url": "https://github.com/exlab-research/dmz", "code_repository": "exlab-research/dmz", "code_commit": "1d740e18e68348e3dd43acfda37508d9aed67a8a", "code_archive": "repos/oVd1y7ilTk.zip", "code_archive_sha256": "8f75859b96c80c6a3819d8c6925123ec98fcbe1ae043b67712f445b2d4492e23", "code_archive_bytes": 47223, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 56, "github_languages": {"Python": 149055}, "github_archived": false, "github_pushed_at": "2025-09-25T09:44:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-designing-diffusion-autoencoders-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zCncHdGsOa", "year": 2025, "status": "rejected", "title": "Efficient optimization with orthogonality constraint: a randomized Riemannian submanifold method", "authors": ["Andi Han", "Pierre-Louis Poirion", "Akiko Takeda"], "authorids": ["~Andi_Han1", "~Pierre-Louis_Poirion1", "~Akiko_Takeda2"], "authors_source": "OpenReview API", "abstract": "Optimization with orthogonality constraints frequently arise in various fields such as machine learning, signal processing and computer vision. Riemannian optimization offers a powerful framework for solving these problems by equipping the constraint set with a Riemannian manifold structure and performing optimization intrinsically on the manifold. This approach typically involves computing a search direction in the tangent space and updating variables via a retraction operation. However, as the size of the variables increases, the computational cost of the retraction can become prohibitively high, limiting the applicability of Riemannian optimization to large-scale problems.  To address this challenge and enhance scalability, we propose a novel approach that restricts each update on a random submanifold, thereby significantly reducing the per-iteration complexity. We introduce two sampling strategies for selecting the random submanifold and theoretically analyze the convergence of the proposed method. We provide convergence results for general nonconvex functions and functions that satisfy Riemannian Polyak–Łojasiewicz condition as well as for stochastic optimization settings. Extensive experiments verify the benefits of the proposed method, showcasing its effectiveness across a wide variety of problem instances.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "7SUvItPomu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5285/Reviewer_8cig"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper proposes a randomized Riemannian submanifold method for solving optimization problems constrained to the set of $n\\times p$ orthogonal frames, also called the Stiefel manifold. The main innovation of the algorithm is to randomly select an $r$-dimensional subspace, in which the current iterate is rotated in a way that locally decreases the objective. In this way, the method can be seen as a Riemmanian analogue to the random block coordinate descent for Euclidean optimization. \n\nThere are two sampling distributions considered for the random selection of the $r$-dimensional subspace: Haar uniform sampling on the orthogonal group and random selection of indices (referred to in the paper as a random permutation matrix). Both of these have the complexity of $nr^2$ although the latter can be computed in a more efficient way due to not needing to perform QR decomposition.\n\nThe paper provides global convergence analysis to critical points and local to a local minima (under PL-ineq.) in expectation and also in high-probability. The method can be also extended/analyzed to the stochastic variant when the objective gradients are random with bounded variance.", "review_text": "The paper proposes a randomized Riemannian submanifold method for solving optimization problems constrained to the set of $n\\times p$ orthogonal frames, also called the Stiefel manifold. The main innovation of the algorithm is to randomly select an $r$-dimensional subspace, in which the current iterate is rotated in a way that locally decreases the objective. In this way, the method can be seen as a Riemmanian analogue to the random block coordinate descent for Euclidean optimization. \n\nThere are two sampling distributions considered for the random selection of the $r$-dimensional subspace: Haar uniform sampling on the orthogonal group and random selection of indices (referred to in the paper as a random permutation matrix). Both of these have the complexity of $nr^2$ although the latter can be computed in a more efficient way due to not needing to perform QR decomposition.\n\nThe paper provides global convergence analysis to critical points and local to a local minima (under PL-ineq.) in expectation and also in high-probability. The method can be also extended/analyzed to the stochastic variant when the objective gradients are random with bounded variance.", "strengths": "The idea of updating the iterate only in a subspace by rotations is not yet well explored and practically useful generalization of the random Riemannian coordinate descent. The paper provides extensive convergence analysis for the methods (although I haven't checked the proofs in the appendices in great detail). The paper is also well written and clearly explains the main concepts. There are multiple numerical experiments in the paper.", "weaknesses": "### Writing comments\nAlthough the paper is well written overall, I have some remarks on formulation of certain ideas:\n* Line 161: the authors state that the random submanifold is defined by the random orthogonal matrix $P_k$. This is not true strictly speaking, in fact, I think it is only the first $r$-columns of the matrix $P_k$ that define the subspace for the rotation.\n* Adding on the remark above: Isn't it then enough to generate only a random orthogonal frame from $St(n,r)$ that defines the subspace?\n* Lines 172 - 178: This part describes one of the the main technical observations: that it is possible to do a local approximation of the $\\tilde F(Y)$ by computing retraction around an identity matrix. Can this be theoretically shown that this is a local approximation? It is not immediately clear to me why this is the case?\n* Eq. 4: there is a mistake/typo after the first equality\n* Line 198: this is very minor remark, but does the order of the first $r$ indices in the sampled permutation matrix matter? In this sense, wouldn't it be sufficient to sample $r$ indices without replacement instead of the permutation? Practically this might be the same so it might not matter.  \n\n### References on sketching\nThis method seems to be related to random subspace sketching. For example a very recent work of (Shustin & Avron, 2024; https://www.jmlr.org/papers/volume25/21-1022/21-1022.pdf) considers random sketching algorithm for fast optimization over orthogonal constraints. This would be good to discuss in the paper (if indeed related), and possibly also to compare against in the numerical experiments. \n\n### Numerical experiments\nThe paper provides numerous numerical experiments but I would like to see the study of dependency on the choice of the random subspace dimension $r$. In the first experiment, the orthogonal procrustes, I am missing the information of the choice of $r$ altogether. I would also like to see a discussion on the computational cost of the random Haar uniform sampling of O(n) as I imagine this will be computationally complex operation.  \n\nOverall, I like the main paper idea, but I am missing more direct comparison with sketching methods in Riemnannian setting as well as the explanation why we can expect (4) to hold as a good first order approximation to minimizing $\\tilde F$. I would also like to see the numerical section improved with more extensive discussion on the choice of $r$ and how it relates to the per-iteration cost.", "questions": "* Could you use fast random subsampling methods, such as random Fast JL transform?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a randomized Riemannian submanifold method for solving optimization problems constrained to the set of $n\\times p$ orthogonal frames, also called the Stiefel manifold. The main innovation of the algorithm is to randomly select an $r$-dimensional subspace, in which the current iterate is rotated in a way that locally decreases the objective. In this way, the method can be seen as a Riemmanian analogue to the random block coordinate descent for Euclidean optimization. \n\nThere are two sampling distributions considered for the random selection of the $r$-dimensional subspace: Haar uniform sampling on the orthogonal group and random selection of indices (referred to in the paper as a random permutation matrix). Both of these have the complexity of $nr^2$ although the latter can be computed in a more efficient way due to not needing to perform QR decomposition.\n\nThe paper provides global convergence analysis to critical points and local to a local minima (under PL-ineq.) in expectation and also in high-probability. The method can be also extended/analyzed to the stochastic variant when the objective gradients are random with bounded variance.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The idea of updating the iterate only in a subspace by rotations is not yet well explored and practically useful generalization of the random Riemannian coordinate descent. The paper provides extensive convergence analysis for the methods (although I haven't checked the proofs in the appendices in great detail). The paper is also well written and clearly explains the main concepts. There are multiple numerical experiments in the paper.", "weaknesses": "### Writing comments\nAlthough the paper is well written overall, I have some remarks on formulation of certain ideas:\n* Line 161: the authors state that the random submanifold is defined by the random orthogonal matrix $P_k$. This is not true strictly speaking, in fact, I think it is only the first $r$-columns of the matrix $P_k$ that define the subspace for the rotation.\n* Adding on the remark above: Isn't it then enough to generate only a random orthogonal frame from $St(n,r)$ that defines the subspace?\n* Lines 172 - 178: This part describes one of the the main technical observations: that it is possible to do a local approximation of the $\\tilde F(Y)$ by computing retraction around an identity matrix. Can this be theoretically shown that this is a local approximation? It is not immediately clear to me why this is the case?\n* Eq. 4: there is a mistake/typo after the first equality\n* Line 198: this is very minor remark, but does the order of the first $r$ indices in the sampled permutation matrix matter? In this sense, wouldn't it be sufficient to sample $r$ indices without replacement instead of the permutation? Practically this might be the same so it might not matter.  \n\n### References on sketching\nThis method seems to be related to random subspace sketching. For example a very recent work of (Shustin & Avron, 2024; https://www.jmlr.org/papers/volume25/21-1022/21-1022.pdf) considers random sketching algorithm for fast optimization over orthogonal constraints. This would be good to discuss in the paper (if indeed related), and possibly also to compare against in the numerical experiments. \n\n### Numerical experiments\nThe paper provides numerous numerical experiments but I would like to see the study of dependency on the choice of the random subspace dimension $r$. In the first experiment, the orthogonal procrustes, I am missing the information of the choice of $r$ altogether. I would also like to see a discussion on the computational cost of the random Haar uniform sampling of O(n) as I imagine this will be computationally complex operation.  \n\nOverall, I like the main paper idea, but I am missing more direct comparison with sketching methods in Riemnannian setting as well as the explanation why we can expect (4) to hold as a good first order approximation to minimizing $\\tilde F$. I would also like to see the numerical section improved with more extensive discussion on the choice of $r$ and how it relates to the per-iteration cost.", "questions": "* Could you use fast random subsampling methods, such as random Fast JL transform?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730721446067}, {"id": "uebMLcqfSi", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5285/Reviewer_hfHE"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes a randomized submanifold algorithm for solving optimization problems with orthogonality constraints. The authors provide convergence guarantees for the proposed algorithm and conduct numerical experiments. Theoretical and numerical comparisons with existing literature are also presented.", "review_text": "The paper proposes a randomized submanifold algorithm for solving optimization problems with orthogonality constraints. The authors provide convergence guarantees for the proposed algorithm and conduct numerical experiments. Theoretical and numerical comparisons with existing literature are also presented.", "strengths": "1. The proposed randomized submanifold method includes several coordinate descent methods as specific examples and allows flexibility in setting the size of the submanifold.\n2. Convergence is established in both expectation and high-probability settings, with clear derivations provided.", "weaknesses": "1. The algorithm does not achieve exact convergence, potentially due to the stochasticity in subspace selection. Is there a way to ensure exact convergence, for example, by adding restrictions on the sampling process?\n2. It appears that the total computational complexity shows no improvement compared to Riemannian gradient descent and may even be higher when using orthogonal sampling. Additionally, the algorithm only converges to a neighborhood rather than a first-order stationary point.\n3. The numerical comparisons are insufficient. It would be helpful to see the effects of different submanifold sizes $r$ and small $p$ values.", "questions": "1. Theorem 1. How to ensure $X_k$ remains inside $U$ under $X_{k_0} \\in U$? The stochasticity induced by sampling seems problematic. Purely assuming all $X_k \\in U$ may be too strong. \n2. Remark 3. The retraction is at cost of $\\mathcal{O}(np^2)$, while calculating the Euclidean gradient is easy to be $\\mathcal{O}(n^2p)$ flops, as seen in problems like PCA. Why does the Riemannian gradient descent  become impractical?\n3. Theorem 2. Is it possible to design a verison of submanifold method that has exact convergence rather than high probability.\n4. Figure 2. Many Stiefel maniofold applications are with small $p$. Could you also add some numerical experiments on small $p$ to see the comparsions?\n5. Figure 5. The improvement on the accuracy appears marginal. The early-stage superior performance might also be achieved by using a larger step size.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a randomized submanifold algorithm for solving optimization problems with orthogonality constraints. The authors provide convergence guarantees for the proposed algorithm and conduct numerical experiments. Theoretical and numerical comparisons with existing literature are also presented.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The proposed randomized submanifold method includes several coordinate descent methods as specific examples and allows flexibility in setting the size of the submanifold.\n2. Convergence is established in both expectation and high-probability settings, with clear derivations provided.", "weaknesses": "1. The algorithm does not achieve exact convergence, potentially due to the stochasticity in subspace selection. Is there a way to ensure exact convergence, for example, by adding restrictions on the sampling process?\n2. It appears that the total computational complexity shows no improvement compared to Riemannian gradient descent and may even be higher when using orthogonal sampling. Additionally, the algorithm only converges to a neighborhood rather than a first-order stationary point.\n3. The numerical comparisons are insufficient. It would be helpful to see the effects of different submanifold sizes $r$ and small $p$ values.", "questions": "1. Theorem 1. How to ensure $X_k$ remains inside $U$ under $X_{k_0} \\in U$? The stochasticity induced by sampling seems problematic. Purely assuming all $X_k \\in U$ may be too strong. \n2. Remark 3. The retraction is at cost of $\\mathcal{O}(np^2)$, while calculating the Euclidean gradient is easy to be $\\mathcal{O}(n^2p)$ flops, as seen in problems like PCA. Why does the Riemannian gradient descent  become impractical?\n3. Theorem 2. Is it possible to design a verison of submanifold method that has exact convergence rather than high probability.\n4. Figure 2. Many Stiefel maniofold applications are with small $p$. Could you also add some numerical experiments on small $p$ to see the comparsions?\n5. Figure 5. The improvement on the accuracy appears marginal. The early-stage superior performance might also be achieved by using a larger step size.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730587036654}, {"id": "yer7w9S7Nv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5285/Reviewer_xyJ5"], "rating": 5, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper proposes a randomized Riemannian submanifold descent method (RSDM) to address large-scale optimization with orthogonality constraints. Specifically, it mitigates the computational costs associated with the retraction in each iteration by updating on a low-dimensional random submanifold rather than the full Stiefel manifold. Two sampling strategies, orthogonal and permutation sampling, are provided. Theoretical analysis guarantees convergence, and extensive experiments demonstrate the effectiveness and efficiency of RSDM.", "review_text": "This paper proposes a randomized Riemannian submanifold descent method (RSDM) to address large-scale optimization with orthogonality constraints. Specifically, it mitigates the computational costs associated with the retraction in each iteration by updating on a low-dimensional random submanifold rather than the full Stiefel manifold. Two sampling strategies, orthogonal and permutation sampling, are provided. Theoretical analysis guarantees convergence, and extensive experiments demonstrate the effectiveness and efficiency of RSDM.", "strengths": "1. The proposed RSDM is novel, which updates on a low-dimensional random submanifold instead of the full Stiefel manifold. In this way, it reduces the computational costs in each iteration, contributing to large-scale optimization.\n2. The paper provides a comprehensive theoretical analysis, including general nonconvex optimization problems, nonconvex functions satisfying the PL condition, and the results in stochastic settings. The discussion about the trade-off between efficiency and convergence is interesting.\n3. The effectiveness and efficiency of the proposed algorithm are validated on an array of numerical experiments, and the source code is available.", "weaknesses": "1. The contribution seems limited in the scenario when $p=\\Omega(n)$. However, numerous applications typically involve the setting $p\\ll n$. Specifically, as discussed in Remark 3, the total complexity of the standard Riemannian gradient descent is $O(np^2\\epsilon^{-2})$, while the proposed method costs $O(n^3\\epsilon^{-2})$, which appears as a theoretical gap.", "questions": "1. As cited in line 88 of the manuscript, the literature [1] proposed a randomized method, RSSM, which also updates on a low-dimensional random submanifold. Could you compare it with the method in this work in terms of computational cost per iteration? Additionally, is there any relationship between RSSM and RSDM?\n2. In line 382 of the manuscript, it states that analysis for Theorem 3 can be easily extended to mini-batch gradient descent. Can the authors explain this point briefly?\n3. In the experiments, the parameter $p$ is consistently close to $n$, for which the proposed RSDM outperforms other methods. It would be beneficial to demonstrate at what percentage of $n$ the value of $p$ needs to reach before RSDM begins to show its advantage.\n4. In Section 7.4, the numerical results of the deep learning task seem confusing, where the accuracy hovers around $40\\\\%$ to classify CIFAR10. Additionally, only the tendency of accuracy is presented; including the training/test loss would be more illustrative.\n\nI would like to increase my grade if the concerns are well addressed.\n\n[1] Randomized submanifold subgradient method for optimization over Stiefel manifolds. 2024", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a randomized Riemannian submanifold descent method (RSDM) to address large-scale optimization with orthogonality constraints. Specifically, it mitigates the computational costs associated with the retraction in each iteration by updating on a low-dimensional random submanifold rather than the full Stiefel manifold. Two sampling strategies, orthogonal and permutation sampling, are provided. Theoretical analysis guarantees convergence, and extensive experiments demonstrate the effectiveness and efficiency of RSDM.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "1. The proposed RSDM is novel, which updates on a low-dimensional random submanifold instead of the full Stiefel manifold. In this way, it reduces the computational costs in each iteration, contributing to large-scale optimization.\n2. The paper provides a comprehensive theoretical analysis, including general nonconvex optimization problems, nonconvex functions satisfying the PL condition, and the results in stochastic settings. The discussion about the trade-off between efficiency and convergence is interesting.\n3. The effectiveness and efficiency of the proposed algorithm are validated on an array of numerical experiments, and the source code is available.", "weaknesses": "1. The contribution seems limited in the scenario when $p=\\Omega(n)$. However, numerous applications typically involve the setting $p\\ll n$. Specifically, as discussed in Remark 3, the total complexity of the standard Riemannian gradient descent is $O(np^2\\epsilon^{-2})$, while the proposed method costs $O(n^3\\epsilon^{-2})$, which appears as a theoretical gap.", "questions": "1. As cited in line 88 of the manuscript, the literature [1] proposed a randomized method, RSSM, which also updates on a low-dimensional random submanifold. Could you compare it with the method in this work in terms of computational cost per iteration? Additionally, is there any relationship between RSSM and RSDM?\n2. In line 382 of the manuscript, it states that analysis for Theorem 3 can be easily extended to mini-batch gradient descent. Can the authors explain this point briefly?\n3. In the experiments, the parameter $p$ is consistently close to $n$, for which the proposed RSDM outperforms other methods. It would be beneficial to demonstrate at what percentage of $n$ the value of $p$ needs to reach before RSDM begins to show its advantage.\n4. In Section 7.4, the numerical results of the deep learning task seem confusing, where the accuracy hovers around $40\\\\%$ to classify CIFAR10. Additionally, only the tendency of accuracy is presented; including the training/test loss would be more illustrative.\n\nI would like to increase my grade if the concerns are well addressed.\n\n[1] Randomized submanifold subgradient method for optimization over Stiefel manifolds. 2024", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730529158654}, {"id": "4t83EvHVLc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5285/Reviewer_Cgj8"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 4, "confidence": 5, "summary": "The paper deals with large-scale optimization problems with orthogonality constraints. To overcome the computational bottleneck of the retraction operator, the authors propose a novel \"subspace\" technique that updates variables on random submanifolds, reducing per-iteration complexity. They introduce two strategies for selecting these submanifolds and analyze the convergence of their methods for both general nonconvex functions and those with local Riemannian PL conditions. Besides, the approach is also applicable to certain quotient manifolds. Experimental results demonstrate the method's effectiveness across various problems.", "review_text": "The paper deals with large-scale optimization problems with orthogonality constraints. To overcome the computational bottleneck of the retraction operator, the authors propose a novel \"subspace\" technique that updates variables on random submanifolds, reducing per-iteration complexity. They introduce two strategies for selecting these submanifolds and analyze the convergence of their methods for both general nonconvex functions and those with local Riemannian PL conditions. Besides, the approach is also applicable to certain quotient manifolds. Experimental results demonstrate the method's effectiveness across various problems.", "strengths": "1. This paper proposes a novel random submanifold descent method (with two sample strategies), which enables low-cost retractions and low per-iteration complexity. \n2. The authors provide a thorough convergence analysis under the non-convex setting, the local Riemannian PL setting, and the stochastic setting.\n3. The proposed algorithm outperforms other existing methods in various problems.", "weaknesses": "1. In line 87, the authors mentioned the poor parallelization of existing methods. However, they did not say too much about why RSDM can be implemented in parallel. In line 93, the authors say “allowing parallel computation”. Adding more discussion and exploration in this aspect can greatly strengthen this paper. For example, elaborating more on the parallel implementation of RSDM or even showing the speedup compared to the sequential implementation in the experiment part would be helpful.\n2. We know from Definition 1 that the Riemannian PL holds locally. In Line 304 of Theorem 1, it says “Suppose k0 large enough such that $X_{k0} \\in U$”. I understand that this assumption is because the iterates may not always stay in the local region due to algorithmic randomness. Is it possible to get rid of such an assumption? This seems difficult to me. Similar things happen in other theorems.", "questions": "1. In Line 69 and Line 92, the authors say that the complexity of non-standard linear algebra reduces to O(r^3) from O(nr^2). I understand O(r^3) comes from the cost of retraction operation. However, the per-iteration cost is still O(npr) or O(nr^2) (see Section 4.1 and Section 4.2). It would be better to clarify these in the statement.\n2. In line 142 and line 146, it is not suitable to use “denote … as …”. Besides, in line 155, what is the meaning of $X^*$?\n3. It can be better to explicitly write down the expression grad F_k(I_n) arising in (4) using the Euclidean gradient. Although it is very simple, it can help readers understand lines 212-213.\n4. In Line 222, why does the gradient computation only require O(nr^2)?\n5. Could the authors add more discussions on the parallel implementation?\n6. Could the authors get rid of the assumptions in Line 304? \n7. In line 484 of Section 7.3, it should be Figure 4 instead of Figure 2. In addition, if possible, could the authors also compare their RSDM with RSSM in (Cheung-Wang-Yue-So ‘24) in Line 577?\n8. In Line 467-470, Can the phenomenon of switching from sublinear rate to linear rate be explained by the error bound property (or the Riemannian PL property)? See Theorem 1 and Theorem 2 in “Liu H, So A M C, Wu W. Quadratic optimization with orthogonality constraint: explicit Łojasiewicz exponent and linear convergence of retraction-based line-search and stochastic variance-reduced gradient methods[J]. Mathematical Programming, 2019, 178(1): 215-262.” The authors can add some discussion if this is related.\n9. The update of RSDM is similar to the algorithm in \"A Block Coordinate Descent Method for Nonsmooth Composite Optimization\nunder Orthogonality Constraints\" (https://arxiv.org/pdf/2304.03641). For example, the formula in Line 215 is similar to (12) in the reference. Could the authors compare RSDM with the algorithm in the reference and highlight the differences more explicitly?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper deals with large-scale optimization problems with orthogonality constraints. To overcome the computational bottleneck of the retraction operator, the authors propose a novel \"subspace\" technique that updates variables on random submanifolds, reducing per-iteration complexity. They introduce two strategies for selecting these submanifolds and analyze the convergence of their methods for both general nonconvex functions and those with local Riemannian PL conditions. Besides, the approach is also applicable to certain quotient manifolds. Experimental results demonstrate the method's effectiveness across various problems.", "soundness": 4, "presentation": 4, "contribution": 4, "strengths": "1. This paper proposes a novel random submanifold descent method (with two sample strategies), which enables low-cost retractions and low per-iteration complexity. \n2. The authors provide a thorough convergence analysis under the non-convex setting, the local Riemannian PL setting, and the stochastic setting.\n3. The proposed algorithm outperforms other existing methods in various problems.", "weaknesses": "1. In line 87, the authors mentioned the poor parallelization of existing methods. However, they did not say too much about why RSDM can be implemented in parallel. In line 93, the authors say “allowing parallel computation”. Adding more discussion and exploration in this aspect can greatly strengthen this paper. For example, elaborating more on the parallel implementation of RSDM or even showing the speedup compared to the sequential implementation in the experiment part would be helpful.\n2. We know from Definition 1 that the Riemannian PL holds locally. In Line 304 of Theorem 1, it says “Suppose k0 large enough such that $X_{k0} \\in U$”. I understand that this assumption is because the iterates may not always stay in the local region due to algorithmic randomness. Is it possible to get rid of such an assumption? This seems difficult to me. Similar things happen in other theorems.", "questions": "1. In Line 69 and Line 92, the authors say that the complexity of non-standard linear algebra reduces to O(r^3) from O(nr^2). I understand O(r^3) comes from the cost of retraction operation. However, the per-iteration cost is still O(npr) or O(nr^2) (see Section 4.1 and Section 4.2). It would be better to clarify these in the statement.\n2. In line 142 and line 146, it is not suitable to use “denote … as …”. Besides, in line 155, what is the meaning of $X^*$?\n3. It can be better to explicitly write down the expression grad F_k(I_n) arising in (4) using the Euclidean gradient. Although it is very simple, it can help readers understand lines 212-213.\n4. In Line 222, why does the gradient computation only require O(nr^2)?\n5. Could the authors add more discussions on the parallel implementation?\n6. Could the authors get rid of the assumptions in Line 304? \n7. In line 484 of Section 7.3, it should be Figure 4 instead of Figure 2. In addition, if possible, could the authors also compare their RSDM with RSSM in (Cheung-Wang-Yue-So ‘24) in Line 577?\n8. In Line 467-470, Can the phenomenon of switching from sublinear rate to linear rate be explained by the error bound property (or the Riemannian PL property)? See Theorem 1 and Theorem 2 in “Liu H, So A M C, Wu W. Quadratic optimization with orthogonality constraint: explicit Łojasiewicz exponent and linear convergence of retraction-based line-search and stochastic variance-reduced gradient methods[J]. Mathematical Programming, 2019, 178(1): 215-262.” The authors can add some discussion if this is related.\n9. The update of RSDM is similar to the algorithm in \"A Block Coordinate Descent Method for Nonsmooth Composite Optimization\nunder Orthogonality Constraints\" (https://arxiv.org/pdf/2304.03641). For example, the formula in Line 215 is similar to (12) in the reference. Could the authors compare RSDM with the algorithm in the reference and highlight the differences more explicitly?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No", "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729838603871}], "openreview_url": "https://openreview.net/forum?id=zCncHdGsOa", "arxiv_id": "2505.12378", "paper_pdf": "papers/zCncHdGsOa.pdf", "paper_pdf_sha256": "b3a8931dc232e0663dd4d50438dd76c65caeea97600ddff6b216e203c4791663", "paper_pdf_bytes": 1255804, "paper_pdf_source": "openreview", "code_url": "https://github.com/andyjm3/RSDM", "code_repository": "andyjm3/RSDM", "code_commit": "ac14994038a100b1047a324c7931b529ac0397f5", "code_archive": "repos/zCncHdGsOa.zip", "code_archive_sha256": "6503c40e738e953eea87e17120ce9996b1accc5ceba55efd8636e57300993994", "code_archive_bytes": 15170, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 17, "github_languages": {"Python": 30936}, "github_archived": false, "github_pushed_at": "2025-05-21T10:10:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-optimization-with-orthogonality"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tB7p0SM5TH", "year": 2024, "status": "rejected", "title": "GraSP: Simple yet Effective Graph Similarity Predictions", "authors": ["Haoran Zheng", "Jieming Shi"], "authorids": ["~Haoran_Zheng1", "~Jieming_Shi1"], "authors_source": "OpenReview API", "abstract": "Graph similarity computation (GSC) is considered one of the essential operations because of its wide range of applications in various fields. Graph Edit Distance (GED) and Maximum Common Subgraph (MCS) are the most popular graph similarity metrics. However, calculating exact GED and MCS is a complex task that falls under the category of NP-hard problems. Consequently, state-of-the-art methodologies learn data-driven models leveraging graph neural networks (GNNs) for estimating GED and MCS values. A perceived limitation of these approaches includes reliance on computationally expensive cross-graph node-level interaction components but to little avail. Instead of building up complicated components, we aim to make the complicated simple and present GraSP, a simple yet highly effective approach for GSC.  In particular, to achieve higher expressiveness,  we design techniques to enhance node features via positional encoding, employ a graph neural network backbone with a gating mechanism and residual connections, and develop a multi-scale pooling technique to generate meaningful representations. We theoretically prove that our method is more expressive and passes 1-WL test performance capabilities. Notably, GraSP is versatile in accurately predicting GED and MCS metrics. In extensive experiments against numerous competitors on real-world datasets, we demonstrate the superiority of GraSP over prior arts regarding effectiveness and efficiency. The source code is available at https://anonymous.4open.science/r/GraSP.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "7k4KwZGIX2", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7208/Reviewer_Ytzi"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a new method for approximating graph similarity scores, leveraging random walk method for position encoding, RWPE, and multi-scale pooling. The overall method GRASP shows superior performance compared to exiting methods over four datasets and two graph similarity/distance metrics.", "review_text": "The paper proposes a new method for approximating graph similarity scores, leveraging random walk method for position encoding, RWPE, and multi-scale pooling. The overall method GRASP shows superior performance compared to exiting methods over four datasets and two graph similarity/distance metrics.", "strengths": "1. The method is well-motivated with the observation that cross-graph node-level interaction is costly, and the proposed method is relatively simple with great inference time reduction.\n2. The paper is well written and easy to follow.", "weaknesses": "1. It would be interesting to know what would be the alternative to the proposed combination of summation and attention pooling, e.g. what if a simple non-learning combination is used, i.e. $\\alpha * z_{sum} + (1 - \\alpha) * z_{att}$ where $\\alpha$ is a hyperaprameter scalar. This will further highlight the importance of learnable combination of the two pooling methods.\n\n2. The overall novelty is limited, given the adoption of random walk positional encoding and combination of existing graph pooling are not firstly proposed in this paper.", "questions": "1. Why the time complexity for computing the random walk positional encoding can be omitted? For each new graph pair, the inference time should include such positional encoding, since at inference time, the graph pair can be new and thus unseen by the model. Unless the authors assume a database-like setting, such overhead cannot be omitted.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new method for approximating graph similarity scores, leveraging random walk method for position encoding, RWPE, and multi-scale pooling. The overall method GRASP shows superior performance compared to exiting methods over four datasets and two graph similarity/distance metrics.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The method is well-motivated with the observation that cross-graph node-level interaction is costly, and the proposed method is relatively simple with great inference time reduction.\n2. The paper is well written and easy to follow.", "weaknesses": "1. It would be interesting to know what would be the alternative to the proposed combination of summation and attention pooling, e.g. what if a simple non-learning combination is used, i.e. $\\alpha * z_{sum} + (1 - \\alpha) * z_{att}$ where $\\alpha$ is a hyperaprameter scalar. This will further highlight the importance of learnable combination of the two pooling methods.\n\n2. The overall novelty is limited, given the adoption of random walk positional encoding and combination of existing graph pooling are not firstly proposed in this paper.", "questions": "1. Why the time complexity for computing the random walk positional encoding can be omitted? For each new graph pair, the inference time should include such positional encoding, since at inference time, the graph pair can be new and thus unseen by the model. Unless the authors assume a database-like setting, such overhead cannot be omitted.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698782057345}, {"id": "89ctAtJATP", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7208/Reviewer_NSFb"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposed GRASP, a method that leverages Graph Neural Networks to approximate Graph Distance/Similarity metrics, namely:\n1. Graph Edit Distance (GED) and \n2. Maximum Common Subgraph (MCS) \n\nwhose exact computations are NP-Hard. The authors enhanced node features using positional encoding and learned an embedding for graph using RGGC layers and multi-scale pooling. These embeddings are used to estimate GED/MCS. The authors demonstrated better efficacy and efficiency of their model compared to baselines.", "review_text": "This paper proposed GRASP, a method that leverages Graph Neural Networks to approximate Graph Distance/Similarity metrics, namely:\n1. Graph Edit Distance (GED) and \n2. Maximum Common Subgraph (MCS) \n\nwhose exact computations are NP-Hard. The authors enhanced node features using positional encoding and learned an embedding for graph using RGGC layers and multi-scale pooling. These embeddings are used to estimate GED/MCS. The authors demonstrated better efficacy and efficiency of their model compared to baselines.", "strengths": "1. The paper focuses on predicting graph similarity/distance metrics which is a very important problem. The organization of the paper is good and easy to follow. \n2. The authors used positional encoding to enhance node features which also aided in passing the 1-WL test.\n3. The ablation study is good which covered most of the design choices the authors made.", "weaknesses": "1. The authors presented MSE on predicted similarity scores (obtained by exponentiating a normalized version of GED) instead of predicted GED. Given the task is to predict GED, results should be reported on GED itself and not its transformation to a similarity metric that distorts the true errors. While some previous works such as SIMGNN have also followed the same methodology of reporting results on exponentiated similarity instead of true GED, there is no justification for this transformation.\n\n2. The statistics of the data used for training, validation, and inference are not provided in the paper, which are important to understand the quality of results.\n\n3. The MSE scores of baselines reported in the paper do not align with the existing literature. For example, GREED outperforms other baselines such as H2MN, SIMGNN in the literature but it is not reflected in this paper. What is the source of this discrepancy? Were all methods trained on error over GED or over the similarity score. The authors need to release the version of code they used benchmarking the baselines and the exact loss function so that reproducibility and any source of discrepancy can be properly analyzed.\n\n4. Experiments are not extensive. \n* The heat maps are provided for the AIDS700nef dataset where graphs are of small sizes. Heatmaps on datasets with larger graphs such as PTC give a better idea of the performance of Grasp and those need to be included.\n* It is not clear how Grasp generalizes for unseen graph sizes. Given that generating ground truth for graphs with larger sizes is expensive, it is interesting to see how Grasp performs when training is done with smaller graphs and testing is done on larger unseen graphs. This aspect needs to be compared with other state-of-the-art baselines such as GREED and H2MN.\n\n5. The estimated GED didn’t include costs for relabeling of edges. The authors didn’t mention how to extend this work to include edge substitution costs.\n\n6. The novelty of GRASP is limited apart from using positional encoding. Grasp tackles the issue of cross-graph interactions, although it's important to recognize that Greed had already addressed this limitation before Grasp.\n\n---------------------------\nOverall, I am willing to revisit the rating, if the authors address concerns on reproducibility (release of baseline implementations used, train-test-validation stats, loss function clarifications), reported results include performance on true GED instead of transformed similarity that distorts performance due to exponentiation, and more detailed experiments in terms of generalizability to unseen, larger sizes, heatmaps, etc.", "questions": "1. It's unclear whether the reported MSE scores for GRASP are a result of the loss computed using the GED or similarity scores (obtained after normalization of GED). The code, specifically Line 170 in src/trainer.py, appears to calculate the loss using GED/Similarity score based on the command line parameters. A consistent methodology is required across all datasets and baselines. Hence, this needs to be clarified.\n\n2. Are the baseline models trained to output GED scores or similarity scores (obtained after normalization of GED)? Models such as GREED are trained to output GED, training them to output similarity scores might affect the performance.\n\n3. How is the Neural Tensor Network (NTN) affecting the performance of the method? Have you considered using only L2-Norm to predict GED, which will preserve the metric property as well? This analysis is not included in the ablation study.\n\n4. What are the statistics of train-validation-test sets?\n\n5. How does the accuracy vary with query size and GED on datasets with larger graphs?\n\n6. What are the RMSE scores of GRASP and other baselines on GED?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed GRASP, a method that leverages Graph Neural Networks to approximate Graph Distance/Similarity metrics, namely:\n1. Graph Edit Distance (GED) and \n2. Maximum Common Subgraph (MCS) \n\nwhose exact computations are NP-Hard. The authors enhanced node features using positional encoding and learned an embedding for graph using RGGC layers and multi-scale pooling. These embeddings are used to estimate GED/MCS. The authors demonstrated better efficacy and efficiency of their model compared to baselines.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The paper focuses on predicting graph similarity/distance metrics which is a very important problem. The organization of the paper is good and easy to follow. \n2. The authors used positional encoding to enhance node features which also aided in passing the 1-WL test.\n3. The ablation study is good which covered most of the design choices the authors made.", "weaknesses": "1. The authors presented MSE on predicted similarity scores (obtained by exponentiating a normalized version of GED) instead of predicted GED. Given the task is to predict GED, results should be reported on GED itself and not its transformation to a similarity metric that distorts the true errors. While some previous works such as SIMGNN have also followed the same methodology of reporting results on exponentiated similarity instead of true GED, there is no justification for this transformation.\n\n2. The statistics of the data used for training, validation, and inference are not provided in the paper, which are important to understand the quality of results.\n\n3. The MSE scores of baselines reported in the paper do not align with the existing literature. For example, GREED outperforms other baselines such as H2MN, SIMGNN in the literature but it is not reflected in this paper. What is the source of this discrepancy? Were all methods trained on error over GED or over the similarity score. The authors need to release the version of code they used benchmarking the baselines and the exact loss function so that reproducibility and any source of discrepancy can be properly analyzed.\n\n4. Experiments are not extensive. \n* The heat maps are provided for the AIDS700nef dataset where graphs are of small sizes. Heatmaps on datasets with larger graphs such as PTC give a better idea of the performance of Grasp and those need to be included.\n* It is not clear how Grasp generalizes for unseen graph sizes. Given that generating ground truth for graphs with larger sizes is expensive, it is interesting to see how Grasp performs when training is done with smaller graphs and testing is done on larger unseen graphs. This aspect needs to be compared with other state-of-the-art baselines such as GREED and H2MN.\n\n5. The estimated GED didn’t include costs for relabeling of edges. The authors didn’t mention how to extend this work to include edge substitution costs.\n\n6. The novelty of GRASP is limited apart from using positional encoding. Grasp tackles the issue of cross-graph interactions, although it's important to recognize that Greed had already addressed this limitation before Grasp.\n\n---------------------------\nOverall, I am willing to revisit the rating, if the authors address concerns on reproducibility (release of baseline implementations used, train-test-validation stats, loss function clarifications), reported results include performance on true GED instead of transformed similarity that distorts performance due to exponentiation, and more detailed experiments in terms of generalizability to unseen, larger sizes, heatmaps, etc.", "questions": "1. It's unclear whether the reported MSE scores for GRASP are a result of the loss computed using the GED or similarity scores (obtained after normalization of GED). The code, specifically Line 170 in src/trainer.py, appears to calculate the loss using GED/Similarity score based on the command line parameters. A consistent methodology is required across all datasets and baselines. Hence, this needs to be clarified.\n\n2. Are the baseline models trained to output GED scores or similarity scores (obtained after normalization of GED)? Models such as GREED are trained to output GED, training them to output similarity scores might affect the performance.\n\n3. How is the Neural Tensor Network (NTN) affecting the performance of the method? Have you considered using only L2-Norm to predict GED, which will preserve the metric property as well? This analysis is not included in the ablation study.\n\n4. What are the statistics of train-validation-test sets?\n\n5. How does the accuracy vary with query size and GED on datasets with larger graphs?\n\n6. What are the RMSE scores of GRASP and other baselines on GED?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698389290463}, {"id": "CjKtXmlK80", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7208/Reviewer_dX6K"], "rating": "5: marginally below the acceptance threshold", "soundness": "4 excellent", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces a novel approach, GRASP, for predicting graph similarity, specifically focusing on GED and MCS metrics. GRASP deviates from the trend of incorporating complex mechanisms by introducing a simplified model that utilizes positional encoding and RGGC to enhance the expressiveness and efficiency of graph neural networks. The authors claim theoretical superiority over the 1-WL test, a widely recognized method for graph isomorphism.", "review_text": "The paper introduces a novel approach, GRASP, for predicting graph similarity, specifically focusing on GED and MCS metrics. GRASP deviates from the trend of incorporating complex mechanisms by introducing a simplified model that utilizes positional encoding and RGGC to enhance the expressiveness and efficiency of graph neural networks. The authors claim theoretical superiority over the 1-WL test, a widely recognized method for graph isomorphism.", "strengths": "S1: The authors innovatively incorporate positional encoding within GNN framework, a commendable step that advances the GNN's ability to capture nuanced structural information. This ingenuity potentially sets a new precedent for subsequent research in graph similarity assessment.\\\nS2: The methodology introduced in this paper demonstrates notable efficiency. \\\nS3: The experimental results seem to be promising.", "weaknesses": "W1: The presentation of the content, particularly in Section 3, lacks clarity and cohesiveness, making it challenging for readers to follow and understand the proposed methodology. \\\nW2: The paper posits inefficiency in contemporary cross-graph interaction techniques as a primary catalyst for the development of GRASP. However, the narrative lacks a coherent demonstration of how GRASP mitigates these inefficiencies. The empirical section, intended to validate the method's enhanced efficiency, does not decisively support this assertion. Specifically, the performance metrics juxtaposed with existing strategies such as GREED and ERIC suggest comparable efficiencies, an outcome that muddles the purported superiority of GRASP in this domain. The authors should consider a more nuanced exposition of the method's unique efficiencies, supplemented by robust experimental evidence, to substantiate claims of its advancement over current practices. \\\nW3: The authors assert that prevailing cross-graph interaction modules contribute significantly to computational overheads. However, the delineation of how their proposed GRASP framework, which ostensibly employs similar cross-graph interactions via NTN, innovates upon or diverges from traditional methodologies is ambiguous. This lack of clarity muddles the reader's understanding of any novel contributions the paper might be making in this specific aspect of the framework. It is imperative for the authors to elucidate the nuanced operational differences, if any, introduced by GRASP that ameliorate the time-costly nature of cross-graph interactions, distinctly setting their approach apart from conventional ones. This clarification could significantly enhance the perceived value and ingenuity of their methodology.", "questions": "Q1: Numerous methods exist to enhance the expressiveness of GNNs. What motivated your decision to exclusively focus on positional encoding in your approach? \\\nQ2: What is the rationale behind using RGGC as a backbone? Furthermore, can you explain how the gating mechanism contributes to the effectiveness of your task? \\\nQ3: On page 4 of your paper, you note that \"both of the above pooling methods have some drawbacks.\" Can you offer more specific evidence or instances that highlight these limitations? \\\nQ4: Could you delve into the key differences between GED and MCS, explaining the importance of considering MCS when many related studies concentrate primarily on GED?\nQ5: What sets NTN apart from existing methods of cross-graph interaction, and why is NTN a suitable option for your proposed approach? \\\nQ6: How does GRASP tackle the issue of efficiency, and what factors make it an efficient solution?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel approach, GRASP, for predicting graph similarity, specifically focusing on GED and MCS metrics. GRASP deviates from the trend of incorporating complex mechanisms by introducing a simplified model that utilizes positional encoding and RGGC to enhance the expressiveness and efficiency of graph neural networks. The authors claim theoretical superiority over the 1-WL test, a widely recognized method for graph isomorphism.", "soundness": "4 excellent", "presentation": "1 poor", "contribution": "2 fair", "strengths": "S1: The authors innovatively incorporate positional encoding within GNN framework, a commendable step that advances the GNN's ability to capture nuanced structural information. This ingenuity potentially sets a new precedent for subsequent research in graph similarity assessment.\\\nS2: The methodology introduced in this paper demonstrates notable efficiency. \\\nS3: The experimental results seem to be promising.", "weaknesses": "W1: The presentation of the content, particularly in Section 3, lacks clarity and cohesiveness, making it challenging for readers to follow and understand the proposed methodology. \\\nW2: The paper posits inefficiency in contemporary cross-graph interaction techniques as a primary catalyst for the development of GRASP. However, the narrative lacks a coherent demonstration of how GRASP mitigates these inefficiencies. The empirical section, intended to validate the method's enhanced efficiency, does not decisively support this assertion. Specifically, the performance metrics juxtaposed with existing strategies such as GREED and ERIC suggest comparable efficiencies, an outcome that muddles the purported superiority of GRASP in this domain. The authors should consider a more nuanced exposition of the method's unique efficiencies, supplemented by robust experimental evidence, to substantiate claims of its advancement over current practices. \\\nW3: The authors assert that prevailing cross-graph interaction modules contribute significantly to computational overheads. However, the delineation of how their proposed GRASP framework, which ostensibly employs similar cross-graph interactions via NTN, innovates upon or diverges from traditional methodologies is ambiguous. This lack of clarity muddles the reader's understanding of any novel contributions the paper might be making in this specific aspect of the framework. It is imperative for the authors to elucidate the nuanced operational differences, if any, introduced by GRASP that ameliorate the time-costly nature of cross-graph interactions, distinctly setting their approach apart from conventional ones. This clarification could significantly enhance the perceived value and ingenuity of their methodology.", "questions": "Q1: Numerous methods exist to enhance the expressiveness of GNNs. What motivated your decision to exclusively focus on positional encoding in your approach? \\\nQ2: What is the rationale behind using RGGC as a backbone? Furthermore, can you explain how the gating mechanism contributes to the effectiveness of your task? \\\nQ3: On page 4 of your paper, you note that \"both of the above pooling methods have some drawbacks.\" Can you offer more specific evidence or instances that highlight these limitations? \\\nQ4: Could you delve into the key differences between GED and MCS, explaining the importance of considering MCS when many related studies concentrate primarily on GED?\nQ5: What sets NTN apart from existing methods of cross-graph interaction, and why is NTN a suitable option for your proposed approach? \\\nQ6: How does GRASP tackle the issue of efficiency, and what factors make it an efficient solution?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697942625232}], "openreview_url": "https://openreview.net/forum?id=tB7p0SM5TH", "arxiv_id": "2412.09968", "paper_pdf": "papers/tB7p0SM5TH.pdf", "paper_pdf_sha256": "707198cc4b4e9ccf36a0238d3a3f2ca62b0a5fe09b5767e57ec4428d1d4df89e", "paper_pdf_bytes": 1295848, "paper_pdf_source": "openreview", "code_url": "https://github.com/HaoranZ99/GraSP", "code_repository": "HaoranZ99/GraSP", "code_commit": "d7d89bf1197ec3fb8f4c9ab578ab4ebace757767", "code_archive": "repos/tB7p0SM5TH.zip", "code_archive_sha256": "b597a61fdfa6fc6a6b97ae7b47a99dbe750f4b99084eaeb3eb4eefebb41a7d02", "code_archive_bytes": 16412, "code_file_count": 8, "code_extensions": {".py": 7, ".sh": 1}, "github_disk_usage_kb": 16, "github_languages": {"Python": 47638, "Shell": 867}, "github_archived": false, "github_pushed_at": "2026-02-25T12:42:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/grasp-simple-yet-effective-graph-similarity"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OYKIo3ySkxA", "year": 2023, "status": "rejected", "title": "DIGEST: FAST AND COMMUNICATION EFFICIENT DECENTRALIZED LEARNING WITH LOCAL UPDATES", "authors": ["Peyman Gholami", "Hulya Seferoglu"], "authorids": ["~Peyman_Gholami1", "~Hulya_Seferoglu1"], "authors_source": "OpenReview API", "abstract": "Decentralized learning advocates the elimination of centralized parameter servers\n(aggregation points) for potentially better utilization of underlying resources, de-\nlay reduction, and resiliency against parameter server unavailability and catas-\ntrophic failures. Gossip based decentralized algorithms, where each node in a net-\nwork has its own locally kept model on which it effectuates the learning by talking\nto its neighbors, received a lot of attention recently. Despite their potential, Gossip\nalgorithms introduce huge communication costs. In this work, we show that nodes\ndo not need to communicate as frequently as in Gossip for fast convergence; in\nfact, a sporadic exchange of a digest of a trained model is sufficient. Thus, we\ndesign a fast and communication-efficient decentralized learning mechanism; DI-\nGEST by particularly focusing on stochastic gradient descent (SGD). DIGEST is\na decentralized algorithm building on local-SGD algorithms, which are originally\ndesigned for communication efficient centralized learning. We show through anal-\nysis and experiments that DIGEST significantly reduces the communication cost\nwithout hurting convergence time for both iid and non-iid data.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "QRT5NVD7jsD", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5226/Reviewer_j1sH"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to improve the communication efficiency in decentralized learning. They propose a novel algorithm, DIGEST, which uses local SGD and random walk to reduce communication costs. The random walk communication can be extended to multi-stream. They provide  a standard convergence results and provide empirical evaluation. ", "review_text": "The DIGEST algorithm looks interesting but I still have concerns about the theoretical rates mentioned above. I hope that the authors can address my concerns.", "strengths": "Strength:\n- The design of DIGEST looks interesting.\n- The assumptions and theoretical results seem reasonable.\n\nWeakness:\n- It is claimed in page 8 that linear speedup O(1/VT) can be achieved provided $H+E\\le O(\\sqrt{T/V})$. However, the synchronize interval H can be as large as $2V$ ---- take a chain topology for example. Then the inequality means $T\\ge V^3$ where the time for convergence increasings dramatically with the number of nodes. \n\n- It is unclear from the theoretical results how the graph topology influences the convergence.\n\n- Only simulation experiments are provided.\n\nMinor weakness:\n- While it seems true that DIGEST is independent of the non-iidness of data, it is not reflected in the main theorem 4.1 due to the current assumption 4 (bounded second moment). \n\n- The claim of \"time-varying\" network topology is not justified in the paper.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper aims to improve the communication efficiency in decentralized learning. They propose a novel algorithm, DIGEST, which uses local SGD and random walk to reduce communication costs. The random walk communication can be extended to multi-stream. They provide  a standard convergence results and provide empirical evaluation. ", "strength_and_weaknesses": "Strength:\n- The design of DIGEST looks interesting.\n- The assumptions and theoretical results seem reasonable.\n\nWeakness:\n- It is claimed in page 8 that linear speedup O(1/VT) can be achieved provided $H+E\\le O(\\sqrt{T/V})$. However, the synchronize interval H can be as large as $2V$ ---- take a chain topology for example. Then the inequality means $T\\ge V^3$ where the time for convergence increasings dramatically with the number of nodes. \n\n- It is unclear from the theoretical results how the graph topology influences the convergence.\n\n- Only simulation experiments are provided.\n\nMinor weakness:\n- While it seems true that DIGEST is independent of the non-iidness of data, it is not reflected in the main theorem 4.1 due to the current assumption 4 (bounded second moment). \n\n- The claim of \"time-varying\" network topology is not justified in the paper.\n\n", "clarity,_quality,_novelty_and_reproducibility": "### Clarity\nWhile it is possible to understand the algorithms, the readability can still be improved.\n\n### Novelty\nOverall the paper looks novel. There is a somewhat similar work RelaySGD (Vogels et al., 2021) which considers decentralized optimization over a spanning tree with non-iid data. DIGEST additionally considers communication efficiency\n\nVogels T, He L, Koloskova A, et al. Relaysum for decentralized deep learning on heterogeneous data[J]. Advances in Neural Information Processing Systems, 2021, 34: 28004-28015.", "summary_of_the_review": "The DIGEST algorithm looks interesting but I still have concerns about the theoretical rates mentioned above. I hope that the authors can address my concerns.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667249045636}, {"id": "DWKpdTaT_5l", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5226/Reviewer_yt5V"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes an alternative approach to Gossip and Random-Walk decentralized algorithms. Synchronous Gossip algorithms require agents to communicate their models to their neighbors and then wait for all of their neighbors' updates before aggregating the updates. This setup has two significant flaws. The first is the large communication overhead. The second is that mismatched computational ability among nodes and communication delays can further increase the convergence time. Asynchronous Gossip algorithms, where nodes communicate without waiting for their neighbors, aim to resolve the second issue. However, the communication overhead issue remains, and the delayed updates require strict assumptions to assure global convergence. Random-walk algorithms operate by having one agent update a global model with its local data and then passing the global data to one of its neighbors to repeat the process. While this approach offers significantly less communication overhead, there is a trade-off in increasing convergence time. \n\nThis paper builds upon random-walk algorithms by having nodes perform local steps instead of idling and by having nodes that receive the global model perform an aggregation step, including the prior local steps. Analysis is performed on this algorithm to show how its convergence and experimental results demonstrate the algorithm's superior performance compared to other decentralized methods in terms of wall-clock time.", "review_text": "The paper proposed a variant of Random-walk decentralized algorithms that have agents perform local steps instead of idling. Convergence analysis and experimental results are provided to show the feasibility of this algorithm. The analysis is questionable due to the use of the bounded gradient assumption in the strongly-convex setting. In addition, the algorithm requires knowledge of the proportion of the data held by each agent. The experiments do not detail the optimization of the various algorithms that are used to compare against the author's proposed algorithm.", "strengths": "The strength of this paper is that the proposed algorithm is an intuitive and novel extension of Random-Walk decentralized algorithms. Having agents be active by performing local steps when not hosting the global model limits the downtime of Random-Walk styled algorithms and closely mirrors the plethora of works that incorporate local steps in between gossip communications. The use of local steps in gossip communications has been well-studied empirically to reduce communication costs and the benefits should be transferable to Random-Walk style algorithms.\n\nUnfortunately, the analysis of the algorithm is done using the assumption of bounded gradients (Assumption 4) and that the proportion of data held by each agent relative to the total amount of data is known. The bounded gradient assumption for unconstrained strongly convex minimization cannot be satisfied. This makes the analysis of the algorithm questionable considering the analysis is applied to the strongly convex function class. In addition, using the bounded gradient assumption makes any interpretation of gradient dissimilarity meaningless, as then the following holds\n\n\\begin{equation}\n    \\lVert \\nabla f_i (x) - \\nabla f_j(y) \\rVert \\leq \\lVert \\nabla f_i(x) \\rVert + \\lVert \\nabla f_j(x) \\rVert \\leq 2G.\n\\end{equation}\n\nClearly, the measurement used to gauge the dissimilarity between functions is not refined. In addition, it is quite confusing as to why the bounded gradient assumption is even required. For the i.i.d. case the quantity is unnecessary. For the non-i.i.d. case, non-bias corrected methods uses the average of the norm of the local gradients evaluated at the global solution while bias-corrected methods require no assumptions. Another aspect of the algorithm that is not mentioned as an assumption but should be included is the use of the quantity $\\frac{\\mathcal{D}_v}{\\mathcal{D}}$ which is the proportion of data held by agent $v$. Knowing this quantity is extremely powerful because it allows for decentralized algorithms to account for the data dissimilarity between agents as we know what weight each of the local objective functions contribute towards the global objective function. However, for decentralized algorithms, it is uncommon to possess any global information regarding the data. \n\nAnother aspect of the algorithm that is not well defined is the relationship between the quantity $H$ and $V$ where $H$ is the bound on the interval between two synchronizations and $V$ is the number of agents. In the remark, the authors use the fact that $H + E \\leq O (\\sqrt{T/V})$. For Random-Walk styled decentralized algorithms, it is intuitive that the interval between two synchronizations increases with the size of the network. To account for this, an increasing number of synchronization streams $R$ would be required therefore increasing the communication cost. However, the relationship between $H$, $R$, and $V$ is not detailed.\n\nComparing the behavior of the various algorithms under the i.i.d. balanced case makes very little sense as communication is oftentimes much more expensive than local computation. Thus, performing more local computations is always preferred as local computations will inexpensively push all agents towards the global solution. In the non-i.i.d and unbalanced case, it is clear that DIGEST has the best performance. However, it is unclear what hyperparameters are used for the other algorithms. In addition, DIGEST has the benefit of knowing the proportion of data held by each agent. Thus, the validity of these simulations are questionable.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes an alternative approach to Gossip and Random-Walk decentralized algorithms. Synchronous Gossip algorithms require agents to communicate their models to their neighbors and then wait for all of their neighbors' updates before aggregating the updates. This setup has two significant flaws. The first is the large communication overhead. The second is that mismatched computational ability among nodes and communication delays can further increase the convergence time. Asynchronous Gossip algorithms, where nodes communicate without waiting for their neighbors, aim to resolve the second issue. However, the communication overhead issue remains, and the delayed updates require strict assumptions to assure global convergence. Random-walk algorithms operate by having one agent update a global model with its local data and then passing the global data to one of its neighbors to repeat the process. While this approach offers significantly less communication overhead, there is a trade-off in increasing convergence time. \n\nThis paper builds upon random-walk algorithms by having nodes perform local steps instead of idling and by having nodes that receive the global model perform an aggregation step, including the prior local steps. Analysis is performed on this algorithm to show how its convergence and experimental results demonstrate the algorithm's superior performance compared to other decentralized methods in terms of wall-clock time.", "strength_and_weaknesses": "The strength of this paper is that the proposed algorithm is an intuitive and novel extension of Random-Walk decentralized algorithms. Having agents be active by performing local steps when not hosting the global model limits the downtime of Random-Walk styled algorithms and closely mirrors the plethora of works that incorporate local steps in between gossip communications. The use of local steps in gossip communications has been well-studied empirically to reduce communication costs and the benefits should be transferable to Random-Walk style algorithms.\n\nUnfortunately, the analysis of the algorithm is done using the assumption of bounded gradients (Assumption 4) and that the proportion of data held by each agent relative to the total amount of data is known. The bounded gradient assumption for unconstrained strongly convex minimization cannot be satisfied. This makes the analysis of the algorithm questionable considering the analysis is applied to the strongly convex function class. In addition, using the bounded gradient assumption makes any interpretation of gradient dissimilarity meaningless, as then the following holds\n\n\\begin{equation}\n    \\lVert \\nabla f_i (x) - \\nabla f_j(y) \\rVert \\leq \\lVert \\nabla f_i(x) \\rVert + \\lVert \\nabla f_j(x) \\rVert \\leq 2G.\n\\end{equation}\n\nClearly, the measurement used to gauge the dissimilarity between functions is not refined. In addition, it is quite confusing as to why the bounded gradient assumption is even required. For the i.i.d. case the quantity is unnecessary. For the non-i.i.d. case, non-bias corrected methods uses the average of the norm of the local gradients evaluated at the global solution while bias-corrected methods require no assumptions. Another aspect of the algorithm that is not mentioned as an assumption but should be included is the use of the quantity $\\frac{\\mathcal{D}_v}{\\mathcal{D}}$ which is the proportion of data held by agent $v$. Knowing this quantity is extremely powerful because it allows for decentralized algorithms to account for the data dissimilarity between agents as we know what weight each of the local objective functions contribute towards the global objective function. However, for decentralized algorithms, it is uncommon to possess any global information regarding the data. \n\nAnother aspect of the algorithm that is not well defined is the relationship between the quantity $H$ and $V$ where $H$ is the bound on the interval between two synchronizations and $V$ is the number of agents. In the remark, the authors use the fact that $H + E \\leq O (\\sqrt{T/V})$. For Random-Walk styled decentralized algorithms, it is intuitive that the interval between two synchronizations increases with the size of the network. To account for this, an increasing number of synchronization streams $R$ would be required therefore increasing the communication cost. However, the relationship between $H$, $R$, and $V$ is not detailed.\n\nComparing the behavior of the various algorithms under the i.i.d. balanced case makes very little sense as communication is oftentimes much more expensive than local computation. Thus, performing more local computations is always preferred as local computations will inexpensively push all agents towards the global solution. In the non-i.i.d and unbalanced case, it is clear that DIGEST has the best performance. However, it is unclear what hyperparameters are used for the other algorithms. In addition, DIGEST has the benefit of knowing the proportion of data held by each agent. Thus, the validity of these simulations are questionable.", "clarity,_quality,_novelty_and_reproducibility": "The clarity and quality of the writing of the paper is adequate with minor grammatical errors throughout the paper. The paper is novel in that it combines the use of local steps into Random Walk styled decentralized algorithms. The experiments are not reproducible as the hyperparameters used for the various algorithms in the comparison are not provided. \n\n", "summary_of_the_review": "The paper proposed a variant of Random-walk decentralized algorithms that have agents perform local steps instead of idling. Convergence analysis and experimental results are provided to show the feasibility of this algorithm. The analysis is questionable due to the use of the bounded gradient assumption in the strongly-convex setting. In addition, the algorithm requires knowledge of the proportion of the data held by each agent. The experiments do not detail the optimization of the various algorithms that are used to compare against the author's proposed algorithm.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None", "recommendation": "3: reject, not good enough"}, "tcdate": 1667225113579}, {"id": "rdS661snI8", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5226/Reviewer_cXkn"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a new communication efficient decentralized optimization algorithm for stochastic gradient descent. The algorithm still has the comparable convergence rate compared with non communication efficient case. ", "review_text": "The paper proposes a new communication efficient decentralized optimization algorithm for stochastic gradient descent. The algorithm still has the comparable convergence rate compared with non communication efficient case. The main concern of the paper is its assumptions seem restrictive. \n", "strengths": "+ The algorithm achieves communication efficiency and fast convergence simultaneously. \n+ Both theoretical analysis and extensive experimental results are presented. \n\n-  The set of assumptions seems restrictive. \n-  The notation of the algorithm is not easy to follow.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a new communication efficient decentralized optimization algorithm for stochastic gradient descent. The algorithm still has the comparable convergence rate compared with non communication efficient case. ", "strength_and_weaknesses": "+ The algorithm achieves communication efficiency and fast convergence simultaneously. \n+ Both theoretical analysis and extensive experimental results are presented. \n\n-  The set of assumptions seems restrictive. \n-  The notation of the algorithm is not easy to follow.", "clarity,_quality,_novelty_and_reproducibility": "The notation of the paper is not easy to follow.", "summary_of_the_review": "The paper proposes a new communication efficient decentralized optimization algorithm for stochastic gradient descent. The algorithm still has the comparable convergence rate compared with non communication efficient case. The main concern of the paper is its assumptions seem restrictive. \n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666724361698}], "openreview_url": "https://openreview.net/forum?id=OYKIo3ySkxA", "arxiv_id": "2307.07652", "paper_pdf": "papers/OYKIo3ySkxA.pdf", "paper_pdf_sha256": "3518c5921b252f6f3b450ae28a83dd746017e2a794fdaadf63f0511f7b4a98bb", "paper_pdf_bytes": 1272206, "paper_pdf_source": "openreview", "code_url": "https://github.com/Anonymous404404/DigestCode", "code_repository": "Anonymous404404/DigestCode", "code_commit": "9499a2d81eaf1380a823cdb73436af6f417bf161", "code_archive": "repos/OYKIo3ySkxA.zip", "code_archive_sha256": "8ebf4c976ce8d08dbb8928312017dbe24dba1eea86b253056a2cd2424bde237f", "code_archive_bytes": 23552, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 21, "github_languages": {"Python": 41988}, "github_archived": false, "github_pushed_at": "2023-01-26T19:34:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/digest-fast-and-communication-efficient"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1lUl6NFDH", "year": 2020, "status": "rejected", "title": "Mirror Descent View For Neural Network Quantization", "authors": ["Thalaiyasingam Ajanthan", "Kartik Gupta", "Philip H. S. Torr", "Richard Hartley", "Puneet K. Dokania"], "authorids": ["thalaiyasingam.ajanthan@anu.edu.au", "kartik.gupta@anu.edu.au", "phst@robots.ox.ac.uk", "richard.hartley@anu.edu.au", "puneet@robots.ox.ac.uk"], "authors_source": "OpenReview API", "abstract": "Quantizing large Neural Networks (NN) while maintaining the performance is highly desirable for resource-limited devices due to reduced memory and time complexity. NN quantization is usually formulated as a constrained optimization problem and optimized via a modified version of gradient descent. In this work, by interpreting the continuous parameters (unconstrained) as the dual of the quantized ones, we introduce a Mirror Descent (MD) framework (Bubeck (2015)) for NN quantization. Specifically, we provide conditions on the projections (i.e., mapping from continuous to quantized ones) which would enable us to derive valid mirror maps and in turn the respective MD updates. Furthermore, we discuss a numerically stable implementation of MD by storing an additional set of auxiliary dual variables (continuous). This update is strikingly analogous to the popular Straight Through Estimator (STE) based method which is typically viewed as a “trick” to avoid vanishing gradients issue but here we show that it is an implementation method for MD for certain projections. Our experiments on standard classification datasets (CIFAR-10/100, TinyImageNet) with convolutional and residual architectures show that our MD variants obtain fully-quantized networks with accuracies very close to the floating-point networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SJlXGL9j2B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper340/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to use the mirror descent algorithm for the binary network. The key point is Theorem 3.1, which enables the mirror map. The paper is easy to read and follow, and the main contributions are clearly stated.\n\nHowever, I suggest a weak rejection of this paper. The reasons are\n\nQ1. As Review #3, it is better for authors to provide more theoretical analysis, which better includes the nonconvex objective function and the effect of annealing. \n\nQ2. It is not clear to me, why mirror descent is better than proximal gradient descent, i.e., proxQuant, in this application. The authors repeatedly claim \"MD allows gradient descent to be performed on a more general non-Euclidean space\". This cannot be told by Table 1, which is just overall performance. So, it is better to empirically show this point by an ablation study.\n\nQ3. Since the technical contributions are not enough, I expect more experimental comparisons.\n- Could the authors perform experiments on ImageNet?\n- While VGG and ResNet are taken as a protocol for experimental comparison, it is better to do an extra comparison with STOA networks. VGG and ResNet are too old and easy to be compressed, compression these networks are of little practical values. EfficientNet [1], Mobilenets [2], and Shufflenet [3] can be good ones. The paper will be more convincing with these methods.\n\n[1]. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks\n[2]. Mobilenets: Efficient convolutional neural networks for mobile vision applications\n[3]. Shufflenet: An extremely efficient convolutional neural network for mobile devices", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #4", "review": "The paper proposes to use the mirror descent algorithm for the binary network. The key point is Theorem 3.1, which enables the mirror map. The paper is easy to read and follow, and the main contributions are clearly stated.\n\nHowever, I suggest a weak rejection of this paper. The reasons are\n\nQ1. As Review #3, it is better for authors to provide more theoretical analysis, which better includes the nonconvex objective function and the effect of annealing. \n\nQ2. It is not clear to me, why mirror descent is better than proximal gradient descent, i.e., proxQuant, in this application. The authors repeatedly claim \"MD allows gradient descent to be performed on a more general non-Euclidean space\". This cannot be told by Table 1, which is just overall performance. So, it is better to empirically show this point by an ablation study.\n\nQ3. Since the technical contributions are not enough, I expect more experimental comparisons.\n- Could the authors perform experiments on ImageNet?\n- While VGG and ResNet are taken as a protocol for experimental comparison, it is better to do an extra comparison with STOA networks. VGG and ResNet are too old and easy to be compressed, compression these networks are of little practical values. EfficientNet [1], Mobilenets [2], and Shufflenet [3] can be good ones. The paper will be more convincing with these methods.\n\n[1]. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks\n[2]. Mobilenets: Efficient convolutional neural networks for mobile vision applications\n[3]. Shufflenet: An extremely efficient convolutional neural network for mobile devices", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1574835722996}, {"id": "r1gG6JwAYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper340/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "A good paper that uses the Mirror Descent paradigm for learning quantized networks.  \nThough Mirror Descent is not their original idea, but using it in the context of learning quantized network is novel and interesting.  \nEmpirically, they showed better results than existing method, with comparisons with reasonable baselines including using relaxed projected gradient descent.  \n\nOverall, I don’t have much concerns, but here are some more specific comment/questions (most relates to writing)\n\nIn the intro, it would be great to mention some past success on using MD, as opposed to just saying it’s well-known. Also you mention MD can be used for more than quantization, but compression in general, it’d be better to add that discussion, or remove this sentence. \n\nIn the beginning of Section 2.1, it'd be easier for the readers to make clear that the primal space corresponds to the quantized weights and the dual space corresponds to the unconstrained space in the rest of the paper.\n\nAt the top of page 3 you describe MD for the first time, but it’s unclear to me how y^0 is handled.\n\nThe end of section 3 and section 4 talk quite a bit about STE, maybe it'd be clear if the authors can provide a concise description.\n\nAs someone not super familiar with NN quantization, this work seems like a good contribution.  My only possible concerns would be somehow comparisons to existing methods are not comprehensive enough (if this will be pointed out by the other reviewers)\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "A good paper that uses the Mirror Descent paradigm for learning quantized networks.  \nThough Mirror Descent is not their original idea, but using it in the context of learning quantized network is novel and interesting.  \nEmpirically, they showed better results than existing method, with comparisons with reasonable baselines including using relaxed projected gradient descent.  \n\nOverall, I don’t have much concerns, but here are some more specific comment/questions (most relates to writing)\n\nIn the intro, it would be great to mention some past success on using MD, as opposed to just saying it’s well-known. Also you mention MD can be used for more than quantization, but compression in general, it’d be better to add that discussion, or remove this sentence. \n\nIn the beginning of Section 2.1, it'd be easier for the readers to make clear that the primal space corresponds to the quantized weights and the dual space corresponds to the unconstrained space in the rest of the paper.\n\nAt the top of page 3 you describe MD for the first time, but it’s unclear to me how y^0 is handled.\n\nThe end of section 3 and section 4 talk quite a bit about STE, maybe it'd be clear if the authors can provide a concise description.\n\nAs someone not super familiar with NN quantization, this work seems like a good contribution.  My only possible concerns would be somehow comparisons to existing methods are not comprehensive enough (if this will be pointed out by the other reviewers)\n\n"}, "tcdate": 1571872698047}, {"id": "ryg5FaBptB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper340/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a Mirror Descent (MD) framework for the quantization of neural networks, which, different with previous quantization methods, enables us to derive valid mirror maps and the respective MD updates. Moreover, the authors also provide a stable implementation of MD by storing an additional set of auxiliary dual variables. Experiments on CIFAR-10/100 and TinyImageNet with convolutional and residual architectures show the effective of the proposed model. \n\nOverall, this paper is well-written and provide sufficient material, both theoretical and experimental evidence to support the proposed method. Although the novelty of this work is somehow limited, i.e. appling MD from convex optimization to NN quantization, the authors provides sufficient effort to explore how to success to adopted it the literature. Hence, I lean to make an accept suggestion at this point. \n\nConcern: it would better to provide the code to validate the soundness of the model.\n\n##post comments\nThe rebuttal addresses my concerns and I will not change my score. Thanks.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "This paper proposes a Mirror Descent (MD) framework for the quantization of neural networks, which, different with previous quantization methods, enables us to derive valid mirror maps and the respective MD updates. Moreover, the authors also provide a stable implementation of MD by storing an additional set of auxiliary dual variables. Experiments on CIFAR-10/100 and TinyImageNet with convolutional and residual architectures show the effective of the proposed model. \n\nOverall, this paper is well-written and provide sufficient material, both theoretical and experimental evidence to support the proposed method. Although the novelty of this work is somehow limited, i.e. appling MD from convex optimization to NN quantization, the authors provides sufficient effort to explore how to success to adopted it the literature. Hence, I lean to make an accept suggestion at this point. \n\nConcern: it would better to provide the code to validate the soundness of the model.\n\n##post comments\nThe rebuttal addresses my concerns and I will not change my score. Thanks.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571802498513}, {"id": "H1li8fEhYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper340/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a neural network (NN) quantization based on Mirror Descent (MD) framework. The core of the proposal is the construction of the mirror map from the unconstrained auxiliary variables to the quantized space. Building on that core, the authors derive some mapping functions from the corresponding projection, i.e. tanh, softmax and shifted tanh. The experimental result on benchmark datasets (CIFAR & TinyImageNet) and basic architectures (VGG & ResNet-18) showed that the proposed method is suitable for quantization. The proposed method is a natural extension of ProxQuant, which adopted the proximal gradient descent to quantize NN (a.k.a $\\ell_2$ norm in MD). Different projections in NN quantization lead to different Bregman divergences in MD. \n\nHowever, the authors do not analyze the convergence of the MD with nonconvex objective function in NN quantization neither how to choose the projection for mirror mapping construction. Moreover, it is better to discuss with [Bai et al, 2019] to clarify the novelty of the proposed method. So I concern about the novelty and the theoretical contributions \n\nYu Bai, Yu-Xiang Wang, Edo Liberty. \nProxQuant: Quantized Neural Networks via Proximal Operators. ICLR 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper proposes a neural network (NN) quantization based on Mirror Descent (MD) framework. The core of the proposal is the construction of the mirror map from the unconstrained auxiliary variables to the quantized space. Building on that core, the authors derive some mapping functions from the corresponding projection, i.e. tanh, softmax and shifted tanh. The experimental result on benchmark datasets (CIFAR & TinyImageNet) and basic architectures (VGG & ResNet-18) showed that the proposed method is suitable for quantization. The proposed method is a natural extension of ProxQuant, which adopted the proximal gradient descent to quantize NN (a.k.a $\\ell_2$ norm in MD). Different projections in NN quantization lead to different Bregman divergences in MD. \n\nHowever, the authors do not analyze the convergence of the MD with nonconvex objective function in NN quantization neither how to choose the projection for mirror mapping construction. Moreover, it is better to discuss with [Bai et al, 2019] to clarify the novelty of the proposed method. So I concern about the novelty and the theoretical contributions \n\nYu Bai, Yu-Xiang Wang, Edo Liberty. \nProxQuant: Quantized Neural Networks via Proximal Operators. ICLR 2019."}, "tcdate": 1571730003006}], "openreview_url": "https://openreview.net/forum?id=r1lUl6NFDH", "arxiv_id": "1910.08237", "paper_pdf": "papers/r1lUl6NFDH.pdf", "paper_pdf_sha256": "c8e50b0d3dc1dde1a52f8f4dd4dab3a037acf02fad32172b251e0edaf7ce4525", "paper_pdf_bytes": 564973, "paper_pdf_source": "openreview", "code_url": "https://github.com/kartikgupta-at-anu/md-bnn", "code_repository": "kartikgupta-at-anu/md-bnn", "code_commit": "b42d5aabd78b73b2baee0858b7f2b8ee72e36d6c", "code_archive": "repos/r1lUl6NFDH.zip", "code_archive_sha256": "bbe81658a01507d01fb7f3e7ebba485364d55da038f32cff708027df4f17772b", "code_archive_bytes": 49542, "code_file_count": 30, "code_extensions": {".py": 20, ".sh": 10}, "github_disk_usage_kb": 24, "github_languages": {"Python": 131342, "Shell": 16704}, "github_archived": false, "github_pushed_at": "2021-02-19T00:59:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mirror-descent-view-for-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "By41BjA9YQ", "year": 2019, "status": "rejected", "title": "Laplacian Smoothing Gradient Descent", "authors": ["Stanley J. Osher", "Bao Wang", "Penghang Yin", "Xiyang Luo", "Minh Pham", "Alex T. Lin"], "authorids": ["sjo@math.ucla.edu", "wangbaonj@gmail.com", "yph@g.ucla.edu", "xylmath@gmail.com", "minhrose@ucla.edu", "atlin@math.ucla.edu"], "authors_source": "OpenReview API", "abstract": "We propose a class of very simple modifications of gradient descent and stochastic gradient descent. We show that when applied to a large variety of machine learning problems, ranging from softmax regression to deep neural nets, the proposed surrogates can dramatically reduce the variance and improve the generalization accuracy. The methods only involve multiplying the usual (stochastic) gradient by the inverse of a positive definitive matrix coming from the discrete Laplacian or its high order generalizations. The theory of Hamilton-Jacobi partial differential equations demonstrates that the implicit version of new algorithm is almost the same as doing gradient descent on a new function which (i) has the same global minima as the original function and (ii) is ``more convex\". We show that optimization algorithms with these surrogates converge uniformly in the discrete Sobolev $H_\\sigma^p$ sense and reduce the optimality gap for convex optimization problems. We implement our algorithm into both PyTorch and Tensorflow platforms which only involves changing of a few lines of code. The code will be available on Github.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1ehQaOKnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper54/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers SGD with a scaled norm; in the non-stochastic case (first equation in section 2.1), it is gradient descent in a fixed non-Euclidean norm, but it is the stochastic case that is most interesting. The paper connects this, somewhat, to a Hamilton-Jacobi equation, but then relaxes the implicit step to an explicit step.\n\nThere is solid theory (Prop 2, 3 and 4) for convergence, which makes sense since this is the same as usual SGD but in a different Hilbert space. Since the inner product is stationary, it's just a fixed Hilbert space, so any convergence proofs that work for arbitrary Hilbert space immediately give the result.\n\nThe computational experiments are impressive, and demonstrate a lot of competence with modern neural nets. Some results are hard to interpret (Figs 9, 10) though.\n\nAs for why use the Laplacian, Prop 8 (combined with Prop 6) gives some idea: that we lower the variance, without cheating (ie., we could trivially lower the variance by just multiplying by a small number, but because the operator preserves the sum of the components, it is not \"cheating\").  That is helpful, though it doesn't give a complete picture yet.  The explanation about the link to the \"more convex\" function I find completely inaccurate and misleading (see technical comments for why).\n\nThe writing is mainly fine, though some sentences are written poorly and would benefit from a revision, e.g., 2nd paragraph, \"But none of them is suitable to train deep neural nets (DNNs).\" is quite awkward [also, in this sentence, please explain *why* they are not suitable!]\n\nThe paper circumvents the page limit by using a smaller font (starting on page 3). This might seem like a minor issue, but it is violating the page limit, and not fair to other papers (unless I have misunderstood; the meta-reviewers can probably comment about this).  I do not think it would be unfair to reject the paper on these grounds. It leaves a bad taste in my mouth after reading the paper.\n\n\nTechnical comments:\n\n- page 1, this is called a \"tri-diagonal\" linear system, but it is not, it is circulant due to the upper-right and lower-left entries (the authors are well-aware of this, but the reader maybe confused; especially since if it were tri-diagonal, it would be inverted via the Thomas algorithm not the FFT).\n\n- Section 2: my first impression on reading this is that you've re-discovered the proximal point envelope and the Moreau envelope (and, looking at the proof of Prop. 1, the authors are aware of this connection).  In this context, it's not clear why A_sigma is helpful, as opposed to any positive definite matrix.\n\n- The actual statement of Proposition 1 is unclear. What does \"the ... update ... permits ..\" mean? i.e., \"permits\" is a weird, vague choice of words. What are you actually proving?\n\n- Section 2.1 moves from the proximal point method (in a scaled norm) to the gradient descent method (in a scaled norm). Clearly, these two methods are different, and just as in ODE schemes, the implicit version is unconditionally stable while the explicit one isn't. So motivating your method by \"smoothing\" or \"adding convexity\" is really misleading. You could define equation (2) and the u(w,t) equation by replacing A_sigma with the identity, and as long as tau > 0, this also \"convexifies\", but then if you go from implicit to explicit, you get regular GD, so you haven't really done anything.  So, I do not buy this connection that your method \"convexifies\" the function.\n\n- Using the FFT to invert seems slow (the theoretical flop-count is good, but it's still super-linear, and requires global data movement, so not good for a distributed implementation).  If you really did define A to be tri-diagonal, then you could invert naively in a linear time algorithm with local data movement. Why not use a tri-diagonal A? It might not satisfy prop 8 exactly, but it'd be close, and a lot faster in practice.  From a \"finite-difference\" point-of-view, I don't see an inherent argument about why you want circular boundary conditions.\n\n- Remark 1 seems out-of-place. Why is that included?\n\n- Section 3.2, and Fig. 5.  It's not clear that the improved generalization results are due to broader local minima, or if it's because the methods converged faster on the training data (since they were limited to 200 epochs). Showing the training error, as a function of epoch, would help clarify. Similar comment for other experiments too.\n\n- Section 4 was impressive in the implementations. Nice work.\n\n- The acknowledgments used the boiler-plate latex template text.\n\n** summary **\nQuality: Good\nClarity: OK\nOriginality: mixed\nSignificance: maybe high?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Solid algorithm and results, some concerns, tiny font", "review": "The paper considers SGD with a scaled norm; in the non-stochastic case (first equation in section 2.1), it is gradient descent in a fixed non-Euclidean norm, but it is the stochastic case that is most interesting. The paper connects this, somewhat, to a Hamilton-Jacobi equation, but then relaxes the implicit step to an explicit step.\n\nThere is solid theory (Prop 2, 3 and 4) for convergence, which makes sense since this is the same as usual SGD but in a different Hilbert space. Since the inner product is stationary, it's just a fixed Hilbert space, so any convergence proofs that work for arbitrary Hilbert space immediately give the result.\n\nThe computational experiments are impressive, and demonstrate a lot of competence with modern neural nets. Some results are hard to interpret (Figs 9, 10) though.\n\nAs for why use the Laplacian, Prop 8 (combined with Prop 6) gives some idea: that we lower the variance, without cheating (ie., we could trivially lower the variance by just multiplying by a small number, but because the operator preserves the sum of the components, it is not \"cheating\").  That is helpful, though it doesn't give a complete picture yet.  The explanation about the link to the \"more convex\" function I find completely inaccurate and misleading (see technical comments for why).\n\nThe writing is mainly fine, though some sentences are written poorly and would benefit from a revision, e.g., 2nd paragraph, \"But none of them is suitable to train deep neural nets (DNNs).\" is quite awkward [also, in this sentence, please explain *why* they are not suitable!]\n\nThe paper circumvents the page limit by using a smaller font (starting on page 3). This might seem like a minor issue, but it is violating the page limit, and not fair to other papers (unless I have misunderstood; the meta-reviewers can probably comment about this).  I do not think it would be unfair to reject the paper on these grounds. It leaves a bad taste in my mouth after reading the paper.\n\n\nTechnical comments:\n\n- page 1, this is called a \"tri-diagonal\" linear system, but it is not, it is circulant due to the upper-right and lower-left entries (the authors are well-aware of this, but the reader maybe confused; especially since if it were tri-diagonal, it would be inverted via the Thomas algorithm not the FFT).\n\n- Section 2: my first impression on reading this is that you've re-discovered the proximal point envelope and the Moreau envelope (and, looking at the proof of Prop. 1, the authors are aware of this connection).  In this context, it's not clear why A_sigma is helpful, as opposed to any positive definite matrix.\n\n- The actual statement of Proposition 1 is unclear. What does \"the ... update ... permits ..\" mean? i.e., \"permits\" is a weird, vague choice of words. What are you actually proving?\n\n- Section 2.1 moves from the proximal point method (in a scaled norm) to the gradient descent method (in a scaled norm). Clearly, these two methods are different, and just as in ODE schemes, the implicit version is unconditionally stable while the explicit one isn't. So motivating your method by \"smoothing\" or \"adding convexity\" is really misleading. You could define equation (2) and the u(w,t) equation by replacing A_sigma with the identity, and as long as tau > 0, this also \"convexifies\", but then if you go from implicit to explicit, you get regular GD, so you haven't really done anything.  So, I do not buy this connection that your method \"convexifies\" the function.\n\n- Using the FFT to invert seems slow (the theoretical flop-count is good, but it's still super-linear, and requires global data movement, so not good for a distributed implementation).  If you really did define A to be tri-diagonal, then you could invert naively in a linear time algorithm with local data movement. Why not use a tri-diagonal A? It might not satisfy prop 8 exactly, but it'd be close, and a lot faster in practice.  From a \"finite-difference\" point-of-view, I don't see an inherent argument about why you want circular boundary conditions.\n\n- Remark 1 seems out-of-place. Why is that included?\n\n- Section 3.2, and Fig. 5.  It's not clear that the improved generalization results are due to broader local minima, or if it's because the methods converged faster on the training data (since they were limited to 200 epochs). Showing the training error, as a function of epoch, would help clarify. Similar comment for other experiments too.\n\n- Section 4 was impressive in the implementations. Nice work.\n\n- The acknowledgments used the boiler-plate latex template text.\n\n** summary **\nQuality: Good\nClarity: OK\nOriginality: mixed\nSignificance: maybe high?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541143843834}, {"id": "S1lXrRHt3m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper54/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a variant of gradient descent that can be approximately understood as gradient descent on a smoothed version of the objective function. The motivation of this work is finding flat minima which could imply better generalization ability of a machine learning model.\nCompared with gradient, the proposed algorithm uses gradient multiplied by a special constant square matrix as update direction. The complexity of the matrix vector multiplication is brought down from O(d^2) to O(d*logd) by exploiting special structure of the matrix using FFT. It is proved that the new update vector has smaller variance and amplitude compared to gradient. \nExperiments on different applications showed that the proposed algorithm may have better generalization ability compared with SGD.\nThis is a clearly written paper, but I have a few questions about theoretical gaps and simulation results in the paper.\n\na). It seems the smoothing explanation at the beginning of section 2 is for implicit scheme (equation (3)). However, the explicit scheme used in practice (the first unnumbered equation in section 2.1) uses a heuristic relaxation which makes the smoothing explanation “approximate” for the explicit scheme. Since the implicit scheme is much more complicated than the explicit scheme, I don’t know if the argument for the implicit scheme will “approximately” hold for the explicit scheme used in practice.\n\nb). The concept flat minimum is only useful in nonconvex optimization, but the convergence of the algorithm is only proved in convex setting. Since the main motivation of the algorithm is finding flat minima, the lack of convergence proof for nonconvex setting concerns me.\n\nc). In the neural net experiment in section 4.1, both gradient descent and smooth gradient descent use the same stepsizes. It is known that the performance of gradient descent is sensitive to the choice of stepsizes, for a fair comparison, one should compare the performance of the two algorithms using optimized stepsizes.\n\nd). In the experiment in section 4.2, the proposed algorithm is only used for the first 40 epochs during training and SGD is used for the later phase of training. Why switching to SGD later? \n\nOverall, I feel the idea of this paper is interesting, but the theory and experiments in the paper are not very strong.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Theory and experiments can be improved", "review": "This paper proposed a variant of gradient descent that can be approximately understood as gradient descent on a smoothed version of the objective function. The motivation of this work is finding flat minima which could imply better generalization ability of a machine learning model.\nCompared with gradient, the proposed algorithm uses gradient multiplied by a special constant square matrix as update direction. The complexity of the matrix vector multiplication is brought down from O(d^2) to O(d*logd) by exploiting special structure of the matrix using FFT. It is proved that the new update vector has smaller variance and amplitude compared to gradient. \nExperiments on different applications showed that the proposed algorithm may have better generalization ability compared with SGD.\nThis is a clearly written paper, but I have a few questions about theoretical gaps and simulation results in the paper.\n\na). It seems the smoothing explanation at the beginning of section 2 is for implicit scheme (equation (3)). However, the explicit scheme used in practice (the first unnumbered equation in section 2.1) uses a heuristic relaxation which makes the smoothing explanation “approximate” for the explicit scheme. Since the implicit scheme is much more complicated than the explicit scheme, I don’t know if the argument for the implicit scheme will “approximately” hold for the explicit scheme used in practice.\n\nb). The concept flat minimum is only useful in nonconvex optimization, but the convergence of the algorithm is only proved in convex setting. Since the main motivation of the algorithm is finding flat minima, the lack of convergence proof for nonconvex setting concerns me.\n\nc). In the neural net experiment in section 4.1, both gradient descent and smooth gradient descent use the same stepsizes. It is known that the performance of gradient descent is sensitive to the choice of stepsizes, for a fair comparison, one should compare the performance of the two algorithms using optimized stepsizes.\n\nd). In the experiment in section 4.2, the proposed algorithm is only used for the first 40 epochs during training and SGD is used for the later phase of training. Why switching to SGD later? \n\nOverall, I feel the idea of this paper is interesting, but the theory and experiments in the paper are not very strong.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541131835020}, {"id": "BJgKdk3vh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper54/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to use a simple tri-diagonal matrix to reduce the variance of stochastic gradient and provide a better generalization property. Such a variant is shown to be equivalent to applying GD on smoothed objective function. Theoretical results show a convergence rate and variance reduction. Various experiments are done in different settings.  I have following comments:\n\n1) In section 2, it is stated that \"This viscosity solution u(w, t) makes f(w) more convex by bringing down the local maxima while retaining the wide minima.\" Besides illustrating such a point on some nicely constructed function f, is there any theory or analysis supporting this statement? Or is there any intuition behind it? In the abstract and Section 1, how to define a function is \"more convex\"? This is one of the fountains of the paper, it worths to spend one or two paragraphs to explain it, or at least introduce some references here. The current statement is not formed in a rigorous way.\n\n2) The main advantages of proposed method that the paper claims are, reduce the variance and improve the generalization accuracy. However, there are few comparisons with other existed methods, besides numerical section. Such comparisons or analysis could help readers understand the difference and novelty.\n\n3) The proof seems fine. Propositions 1-4 try to analyze the convergence rate, which are common techniques in other variance reduction papers on SGD. Propositions 5-9 rely on some nice properties of matrix A_\\sigma and show it can help to reduce the variance. Typos:\nPage 11, \"Proof of Proposition 1\", there is a missing \"-\" in \\nabla_w u(w, t), also in the next equation.\nPage 13, \"Proof of Proposition 6, d = A_\\sigma g\". \n\n4) The proposed method strongly relies on the choice of \\sigma, but discussion on how to choose the value for \\sigma is rare. From Proposition 8, the upper bound on reduced variance is a quadratic function on \\sigma, so it is better to discuss more on it or have some experiments on sensitivity analysis. In Section 4, \\sigma varies (1.0, 3.0, etc) in different experiments, but again there are no explanations.\n\n5) Numerical results in Section 4.3 is not strong enough to support the advantage of the proposed method. It is hard to observe \"visibly less noisy\" in both Figure 8 and 9. Better ways of illustration might be considered.\n\n6) The paper is not nicely written thus cannot be easily read. It seems to be cut and pasted from another version in a short time. Some titles of subsections is missing. The font size is not fixed in the whole paper.\n\nThe above concerns prevents me to give a higher rating at this time.\n\nSummary\nquality: ok\nclarity: good\noriginality: nice\nsignificance: good", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some concerns on experiments and written style", "review": "The paper proposes to use a simple tri-diagonal matrix to reduce the variance of stochastic gradient and provide a better generalization property. Such a variant is shown to be equivalent to applying GD on smoothed objective function. Theoretical results show a convergence rate and variance reduction. Various experiments are done in different settings.  I have following comments:\n\n1) In section 2, it is stated that \"This viscosity solution u(w, t) makes f(w) more convex by bringing down the local maxima while retaining the wide minima.\" Besides illustrating such a point on some nicely constructed function f, is there any theory or analysis supporting this statement? Or is there any intuition behind it? In the abstract and Section 1, how to define a function is \"more convex\"? This is one of the fountains of the paper, it worths to spend one or two paragraphs to explain it, or at least introduce some references here. The current statement is not formed in a rigorous way.\n\n2) The main advantages of proposed method that the paper claims are, reduce the variance and improve the generalization accuracy. However, there are few comparisons with other existed methods, besides numerical section. Such comparisons or analysis could help readers understand the difference and novelty.\n\n3) The proof seems fine. Propositions 1-4 try to analyze the convergence rate, which are common techniques in other variance reduction papers on SGD. Propositions 5-9 rely on some nice properties of matrix A_\\sigma and show it can help to reduce the variance. Typos:\nPage 11, \"Proof of Proposition 1\", there is a missing \"-\" in \\nabla_w u(w, t), also in the next equation.\nPage 13, \"Proof of Proposition 6, d = A_\\sigma g\". \n\n4) The proposed method strongly relies on the choice of \\sigma, but discussion on how to choose the value for \\sigma is rare. From Proposition 8, the upper bound on reduced variance is a quadratic function on \\sigma, so it is better to discuss more on it or have some experiments on sensitivity analysis. In Section 4, \\sigma varies (1.0, 3.0, etc) in different experiments, but again there are no explanations.\n\n5) Numerical results in Section 4.3 is not strong enough to support the advantage of the proposed method. It is hard to observe \"visibly less noisy\" in both Figure 8 and 9. Better ways of illustration might be considered.\n\n6) The paper is not nicely written thus cannot be easily read. It seems to be cut and pasted from another version in a short time. Some titles of subsections is missing. The font size is not fixed in the whole paper.\n\nThe above concerns prevents me to give a higher rating at this time.\n\nSummary\nquality: ok\nclarity: good\noriginality: nice\nsignificance: good", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541025648627}], "openreview_url": "https://openreview.net/forum?id=By41BjA9YQ", "arxiv_id": "1806.06317", "paper_pdf": "papers/By41BjA9YQ.pdf", "paper_pdf_sha256": "ee79d1e9ddd55db98ff61e5f20c598b8ffe83d3ba0f44c3a48a09629784f58ac", "paper_pdf_bytes": 1165708, "paper_pdf_source": "openreview", "code_url": "https://github.com/BaoWangMath/LaplacianSmoothing-GradientDescent", "code_repository": "BaoWangMath/LaplacianSmoothing-GradientDescent", "code_commit": "ba394837a6857677c1e243a375f495be32da8a5e", "code_archive": "repos/By41BjA9YQ.zip", "code_archive_sha256": "5264e98d4795aa28d5bab59df6cc05595d5eed7d1372c1daaf4d5d04c450fd08", "code_archive_bytes": 28493, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 50, "github_languages": {"Python": 77387}, "github_archived": false, "github_pushed_at": "2019-05-26T18:29:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/laplacian-smoothing-gradient-descent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJSA_e1AW", "year": 2018, "status": "rejected", "title": "Normalized Direction-preserving Adam", "authors": ["Zijun Zhang", "Lin Ma", "Zongpeng Li", "Chuan Wu"], "authorids": ["zijun.zhang@ucalgary.ca", "linmawhu@gmail.com", "zongpeng@ucalgary.ca", "cwu@cs.hku.hk"], "authors_source": "OpenReview API", "abstract": "Optimization algorithms for training deep models not only affects the convergence rate and stability of the training process, but are also highly related to the generalization performance of trained models. While adaptive algorithms, such as Adam and RMSprop, have shown better optimization performance than stochastic gradient descent (SGD) in many scenarios, they often lead to worse generalization performance than SGD, when used for training deep neural networks (DNNs). In this work, we identify two problems regarding the direction and step size for updating the weight vectors of hidden units, which may degrade the generalization performance of Adam. As a solution, we propose the normalized direction-preserving Adam (ND-Adam) algorithm, which controls the update direction and step size more precisely, and thus bridges the generalization gap between Adam and SGD. Following a similar rationale, we further improve the generalization performance in classification tasks by regularizing the softmax logits. By bridging the gap between SGD and Adam, we also shed some light on why certain optimization algorithms generalize better than others.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "S1-Kfe5lM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper115/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a variant of ADAM optimization algorithm that normalizes the weights of each hidden unit. They further suggest using batch normalization on the output of the network before softmax to improve the generalization. The main ideas are new to me and the paper is well-written. The arguments and derivations are very clear. However, the experimental results suggest that the proposed method is not superior to SGD and ADAM.\n\nPros: \n\n- The idea of optimizing the direction while ignoring the magnitude is interesting and make sense.\n- Using batch normalization before softmax is interesting.\n\nCons:\n\n- In the abstract, authors claim that the proposed method has good optimization performance of ADAM and good generalization performance of SGD. Such a method could be helpful if one can get to the same level of generalization faster (less number of epochs). However, the experiments suggest that optimization advantages of the proposed method do not translate to faster generalization. Figures 2,3 and Table 1 indicate that the generalization performance of this method is very similar to SGD.\n\n- The paper is not coherent. In particular, direction-preserving ADAM and batch-normalized softmax trick are completely orthogonal ideas. \n\n- In the introduction and Section 2.2, authors claim that weight decay has a significant effect on the generalization performance of DNNs. I wonder if authors can refer to any work on this. My own experience and several empirical works have suggested that weight decay does not improve generalization significantly.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A variant of ADAM optimization algorithm that normalizes the weights of each hidden unit", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes a variant of ADAM optimization algorithm that normalizes the weights of each hidden unit. They further suggest using batch normalization on the output of the network before softmax to improve the generalization. The main ideas are new to me and the paper is well-written. The arguments and derivations are very clear. However, the experimental results suggest that the proposed method is not superior to SGD and ADAM.\n\nPros: \n\n- The idea of optimizing the direction while ignoring the magnitude is interesting and make sense.\n- Using batch normalization before softmax is interesting.\n\nCons:\n\n- In the abstract, authors claim that the proposed method has good optimization performance of ADAM and good generalization performance of SGD. Such a method could be helpful if one can get to the same level of generalization faster (less number of epochs). However, the experiments suggest that optimization advantages of the proposed method do not translate to faster generalization. Figures 2,3 and Table 1 indicate that the generalization performance of this method is very similar to SGD.\n\n- The paper is not coherent. In particular, direction-preserving ADAM and batch-normalized softmax trick are completely orthogonal ideas. \n\n- In the introduction and Section 2.2, authors claim that weight decay has a significant effect on the generalization performance of DNNs. I wonder if authors can refer to any work on this. My own experience and several empirical works have suggested that weight decay does not improve generalization significantly.\n\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511813753101}, {"id": "Bk_mQkcgM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper115/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Method:\n\nThe paper is missing analysis of some important related works such as\n\n\"Beyond convexity: Stochastic quasi-convex optimization\" by E. Hazan et al. (2015) \n\nwhere Stochastic Normalized Gradient Descent (SNGD) was proposed. \n\nThen, normalized gradient versions of AdaGrad and Adam were proposed in \n\n\"Normalized Gradient with Adaptive Stepsize Method for Deep\nNeural Network Training\" by A. W. Yu et al. (2017).\n\nAnother work which I find to be relevant is \n\n\"Follow the Signs for Robust Stochastic Optimization\" by L. Balles and P. Hennig (2017).\n\nFrom my personal experiments, restricting w_i to have L2 norm of 1, i.e., to be +-1 \nleads to worse generalization. One reason for this is that weight decay is not \nreally functioning since it cannot move w_i to 0 or make its amplitude any smaller. \nPlease correct me if I misunderstand something here. \n\nThe presence of +-1 weights moves us to the area of low-precision NNs, \nor more specifically, NNs with binary / binarized weights as in \n\n\"BinaryConnect: Training Deep Neural Networks with\nbinary weights during propagations\" by M. Courbariaux et al. (2015)\n\nand \n\n\"Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1\" by M. Courbariaux et al. (2016). \n\nRegarding\n\"Moreover, the magnitude of each update does not depend on themagnitude of the gradient. Thus, ND-Adam is more robust to improper initialization, and vanishing or exploding gradients.\"\n\nIf the magnitude of each update does not depend on the magnitude of the gradient, then the algorithm heavily depends on the learning rate. Otherwise, it does not have any means to approach the optimum in a reasonable number of steps *when* it is initialized very / unreasonably far from it. The claim of your second sentence is not supported by the paper. \n\nEvaluation:\n\nI am not confident that the presented experimental validation is fair. First, the original WRN paper and many other papers with ResNets used weight decay of 0.0005 and not 0.001 or 0.002 as used for SGD in this paper. It is unclear why this setting was changed. One could just use \\alpha_0 = 0.05 and \\lambda = 0.0005.\n\nThen, I don't see why the authors use WRN-22-7.5 which is different from WRN-28-10 which was suggested in the original study and used in several follow-up works. The difference between WRN-22-7.5 and WRN-28-10 is unlikely to be significant, \nthe former might have about only 2 times less parameters which should barely change the final validation errors. However, the use of WRN-22-7.5 makes it impossible to easily compare the presented results to the results of Zagoruyko who had 3.8\\% with WRN-28-10. I believe that the use of the setup of Zagoruyko for WRN-22-7.5 would allow to get much better results than 4.5\\% and 4.49\\% shown for SGD and likely better 4.14\\% shown for ND-Adam. I note that the use of WRN-22-7.5 is unlikely to be due to the used hardware because later in paper the authors refer to WRN-34-7.5.\n\nMy intuition is that the proposed ND-Adam moves the algorithm back to SGD but with potentially harmful constraints of w_i=+-1. Even the values of \\alpha^v_0 found for ND-Adam (e.g., \\alpha^v_0=0.05 in Figure 1B) are in line of what would be optimal values of \\alpha_0 for SGD. \n\nI find it uncomfortable that BN-Softmax is introduced here to support the use of an optimization algorithm, moreover, that the values of \\gamma_c are different for CIFAR-10 and CIFAR-100. I wonder if the proposed values are optimal (and therefore selected) for all three tested algorithms  or only for Adam-ND. I expect that hyperparameters of SGD and Adam would also need to be revised to account for BN-Softmax.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some related works should be analyzed. The experimental validation should to be revised. ", "rating": "5: Marginally below acceptance threshold", "review": "Method:\n\nThe paper is missing analysis of some important related works such as\n\n\"Beyond convexity: Stochastic quasi-convex optimization\" by E. Hazan et al. (2015) \n\nwhere Stochastic Normalized Gradient Descent (SNGD) was proposed. \n\nThen, normalized gradient versions of AdaGrad and Adam were proposed in \n\n\"Normalized Gradient with Adaptive Stepsize Method for Deep\nNeural Network Training\" by A. W. Yu et al. (2017).\n\nAnother work which I find to be relevant is \n\n\"Follow the Signs for Robust Stochastic Optimization\" by L. Balles and P. Hennig (2017).\n\nFrom my personal experiments, restricting w_i to have L2 norm of 1, i.e., to be +-1 \nleads to worse generalization. One reason for this is that weight decay is not \nreally functioning since it cannot move w_i to 0 or make its amplitude any smaller. \nPlease correct me if I misunderstand something here. \n\nThe presence of +-1 weights moves us to the area of low-precision NNs, \nor more specifically, NNs with binary / binarized weights as in \n\n\"BinaryConnect: Training Deep Neural Networks with\nbinary weights during propagations\" by M. Courbariaux et al. (2015)\n\nand \n\n\"Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1\" by M. Courbariaux et al. (2016). \n\nRegarding\n\"Moreover, the magnitude of each update does not depend on themagnitude of the gradient. Thus, ND-Adam is more robust to improper initialization, and vanishing or exploding gradients.\"\n\nIf the magnitude of each update does not depend on the magnitude of the gradient, then the algorithm heavily depends on the learning rate. Otherwise, it does not have any means to approach the optimum in a reasonable number of steps *when* it is initialized very / unreasonably far from it. The claim of your second sentence is not supported by the paper. \n\nEvaluation:\n\nI am not confident that the presented experimental validation is fair. First, the original WRN paper and many other papers with ResNets used weight decay of 0.0005 and not 0.001 or 0.002 as used for SGD in this paper. It is unclear why this setting was changed. One could just use \\alpha_0 = 0.05 and \\lambda = 0.0005.\n\nThen, I don't see why the authors use WRN-22-7.5 which is different from WRN-28-10 which was suggested in the original study and used in several follow-up works. The difference between WRN-22-7.5 and WRN-28-10 is unlikely to be significant, \nthe former might have about only 2 times less parameters which should barely change the final validation errors. However, the use of WRN-22-7.5 makes it impossible to easily compare the presented results to the results of Zagoruyko who had 3.8\\% with WRN-28-10. I believe that the use of the setup of Zagoruyko for WRN-22-7.5 would allow to get much better results than 4.5\\% and 4.49\\% shown for SGD and likely better 4.14\\% shown for ND-Adam. I note that the use of WRN-22-7.5 is unlikely to be due to the used hardware because later in paper the authors refer to WRN-34-7.5.\n\nMy intuition is that the proposed ND-Adam moves the algorithm back to SGD but with potentially harmful constraints of w_i=+-1. Even the values of \\alpha^v_0 found for ND-Adam (e.g., \\alpha^v_0=0.05 in Figure 1B) are in line of what would be optimal values of \\alpha_0 for SGD. \n\nI find it uncomfortable that BN-Softmax is introduced here to support the use of an optimization algorithm, moreover, that the values of \\gamma_c are different for CIFAR-10 and CIFAR-100. I wonder if the proposed values are optimal (and therefore selected) for all three tested algorithms  or only for Adam-ND. I expect that hyperparameters of SGD and Adam would also need to be revised to account for BN-Softmax.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511809823709}, {"id": "SJ_kd7ixf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper115/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper extended the Adam optimization algorithm to preserve the update direction. Instead of using the un-centered variance of individual weights, the proposed method adapts the learning rate for the incoming weights to a hidden unit jointly using the L2 norm of the gradient vector. The authors empirically demonstrated the method works well on CIFAR-10/100 tasks.\n\nComments:\n\n- I found the paper very hard to follow. The authors could improve the clarity of the paper greatly by listing their contribution clearly for readers to digest. The authors also combined the proposed method with a few existing deep learning tricks in the paper. All those tricks that, ie. section 3.3 and 4, should go into the background section.\n\n- Overall, the only contribution of the paper seems to be the ad-hoc modification to Adam in Eq. (9). Why is this a reasonable modification? Do we expect this modification to fail in any circumstances? The experiments on CIFAR dataset and one CNN architecture do not provide enough evidence to show the proposed method work well in general.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "rating": "4: Ok but not good enough - rejection", "review": "The paper extended the Adam optimization algorithm to preserve the update direction. Instead of using the un-centered variance of individual weights, the proposed method adapts the learning rate for the incoming weights to a hidden unit jointly using the L2 norm of the gradient vector. The authors empirically demonstrated the method works well on CIFAR-10/100 tasks.\n\nComments:\n\n- I found the paper very hard to follow. The authors could improve the clarity of the paper greatly by listing their contribution clearly for readers to digest. The authors also combined the proposed method with a few existing deep learning tricks in the paper. All those tricks that, ie. section 3.3 and 4, should go into the background section.\n\n- Overall, the only contribution of the paper seems to be the ad-hoc modification to Adam in Eq. (9). Why is this a reasonable modification? Do we expect this modification to fail in any circumstances? The experiments on CIFAR dataset and one CNN architecture do not provide enough evidence to show the proposed method work well in general.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511892959911}], "openreview_url": "https://openreview.net/forum?id=HJSA_e1AW", "arxiv_id": "1709.04546", "paper_pdf": "papers/HJSA_e1AW.pdf", "paper_pdf_sha256": "a9e704f314545c18f8c31037845072a1b898f859041161991bab73471c4d3cda", "paper_pdf_bytes": 650436, "paper_pdf_source": "openreview", "code_url": "https://github.com/zj10/ND-Adam", "code_repository": "zj10/ND-Adam", "code_commit": "b4a9d59b9c3607bbf734dc8c893fdbdce7bd1bef", "code_archive": "repos/HJSA_e1AW.zip", "code_archive_sha256": "98ecfe34d94e0eab54ded44bf322641d1206e2d6dcf213eaa722f66bfa2ab467", "code_archive_bytes": 19208, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 26, "github_languages": {"Python": 51045}, "github_archived": false, "github_pushed_at": "2018-09-19T04:58:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/normalized-direction-preserving-adam"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LJ6AvummWu", "year": 2026, "status": "rejected", "title": "SOReL and TOReL: Two Methods for Fully Offline Reinforcement Learning", "authors": ["Mattie Fellows", "Clarisse Wibault", "Uljad Berdica", "Johannes Forkel", "Michael A Osborne", "Jakob Nicolaus Foerster"], "authorids": ["~Mattie_Fellows1", "~Clarisse_Wibault1", "~Uljad_Berdica1", "~Johannes_Forkel1", "~Michael_A_Osborne1", "~Jakob_Nicolaus_Foerster1"], "authors_source": "OpenReview API", "abstract": "Sample efficiency remains a major obstacle for real world adoption of reinforcement learning (RL): success has been limited to settings where simulators provide access to essentially unlimited environment interactions, which in reality are typically costly or dangerous to obtain. Offline RL in principle offers a solution by exploiting offline data to learn a near-optimal policy before deployment. In practice, however, current offline RL methods rely on extensive online interactions for hyperparameter tuning, and have no reliable bound on their initial online performance. To address these two issues, we introduce two algorithms. Firstly, SOReL: an algorithm for safe offline reinforcement learning. Using only offline data our Bayesian approach infers a posterior over environment dynamics to obtain a reliable estimate of the online performance via the posterior predictive uncertainty. Crucially, all hyperparameters are also tuned fully offline. Secondly, we introduce TOReL: a tuning for offline reinforcement learning algorithm that extends our information rate based offline hyperparameter tuning methods to general offline RL approaches. Our empirical evaluation confirms SOReL's ability to accurately estimate regret in the Bayesian setting whilst TOReL's offline hyperparameter tuning achieves competitive performance with the best online hyperparameter tuning methods using only offline data. Thus, SOReL and TOReL make a significant step towards safe and reliable offline RL, unlocking the potential for RL in the real world.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "i5J2S4esgc", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12284/Reviewer_CiFA"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper considers two salient problems with offline RL: (a) what offline metrics are useful for tuning hyper-parameters of an offline RL policy search procedure, (b) how to estimate the online regret associated with deploying a policy outputted by an offline RL procedure? Towards this, the paper utilizes Bayesian perspective to consider the use of posterior information loss (PIL) for purposes of solving the aforementioned problems, developing TOReL (for tuning) and SOReL (for estimating regret).", "review_text": "This paper considers two salient problems with offline RL: (a) what offline metrics are useful for tuning hyper-parameters of an offline RL policy search procedure, (b) how to estimate the online regret associated with deploying a policy outputted by an offline RL procedure? Towards this, the paper utilizes Bayesian perspective to consider the use of posterior information loss (PIL) for purposes of solving the aforementioned problems, developing TOReL (for tuning) and SOReL (for estimating regret).", "strengths": "- The paper addresses important problems that plague the design of offline RL algorithms and these issues limit the applicability of offline RL in general. \n- The empirical results appear pretty compelling as well.", "weaknesses": "- The paper can do a better job with presenting connections with lowerbounds in offline RL connecting the regret estimation against hardness of policy evaluation; for instance, see [1], but there are probably other works that build on this. Why does the proposed procedure actually work in light of some of these stark negative results?\n\n[1] Wang et al: What are the Statistical Limits of Offline RL with Linear Function Approximation? 2020.", "questions": "- Regarding estimators proposed by this paper: can you mention again how one can rely on point estimates of these quantities and whether they can be strengthened using estimation procedures building on [2]?\n\n[2] Agarwal et al: Deep Reinforcement Learning at the Edge of the Statistical Precipice, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper considers two salient problems with offline RL: (a) what offline metrics are useful for tuning hyper-parameters of an offline RL policy search procedure, (b) how to estimate the online regret associated with deploying a policy outputted by an offline RL procedure? Towards this, the paper utilizes Bayesian perspective to consider the use of posterior information loss (PIL) for purposes of solving the aforementioned problems, developing TOReL (for tuning) and SOReL (for estimating regret).", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- The paper addresses important problems that plague the design of offline RL algorithms and these issues limit the applicability of offline RL in general. \n- The empirical results appear pretty compelling as well.", "weaknesses": "- The paper can do a better job with presenting connections with lowerbounds in offline RL connecting the regret estimation against hardness of policy evaluation; for instance, see [1], but there are probably other works that build on this. Why does the proposed procedure actually work in light of some of these stark negative results?\n\n[1] Wang et al: What are the Statistical Limits of Offline RL with Linear Function Approximation? 2020.", "questions": "- Regarding estimators proposed by this paper: can you mention again how one can rely on point estimates of these quantities and whether they can be strengthened using estimation procedures building on [2]?\n\n[2] Agarwal et al: Deep Reinforcement Learning at the Edge of the Statistical Precipice, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761944728752}, {"id": "iGTF8R9bmp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12284/Reviewer_iime"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper addresses two critical issues in offline reinforcement learning that have largely been overlooked: the lack of offline hyperparameter tuning metrics, and the absence of reliable offline regret approximation methods . The authors propose a Bayesian framework for offline RL, leveraging the Posterior Information Loss (PIL), the expected KL divergence between a learned model and true environment dynamics as an offline tuning signal. From this framework, they introduce two methods: SOReL (Safe Offline RL), which provides reliable offline regret estimates using predictive uncertainty to enable safe deployment, and TOReL (Tuning for Offline RL), which extends SOReL's offline hyperparameter tuning methodology to general model-based and model-free offline RL algorithms. The paper provides rigorous theoretical analysis showing that regret is bounded by the PIL (Theorem 1) and that Bayesian offline RL achieves the optimal convergence rate (Theorem 2). Empirical evaluations on D4RL, Adroit, and proprietary Brax datasets demonstrate that TOReL identifies near-oracle hyperparameters using only offline data, saving 20K to >200K online samples compared to existing online hyperparameter tuning methods, while SOReL accurately approximates true regret across diverse offline datasets.", "review_text": "This paper addresses two critical issues in offline reinforcement learning that have largely been overlooked: the lack of offline hyperparameter tuning metrics, and the absence of reliable offline regret approximation methods . The authors propose a Bayesian framework for offline RL, leveraging the Posterior Information Loss (PIL), the expected KL divergence between a learned model and true environment dynamics as an offline tuning signal. From this framework, they introduce two methods: SOReL (Safe Offline RL), which provides reliable offline regret estimates using predictive uncertainty to enable safe deployment, and TOReL (Tuning for Offline RL), which extends SOReL's offline hyperparameter tuning methodology to general model-based and model-free offline RL algorithms. The paper provides rigorous theoretical analysis showing that regret is bounded by the PIL (Theorem 1) and that Bayesian offline RL achieves the optimal convergence rate (Theorem 2). Empirical evaluations on D4RL, Adroit, and proprietary Brax datasets demonstrate that TOReL identifies near-oracle hyperparameters using only offline data, saving 20K to >200K online samples compared to existing online hyperparameter tuning methods, while SOReL accurately approximates true regret across diverse offline datasets.", "strengths": "**Well-motivated problem formulation.** The paper identifies and articulates two concrete, practically important issues in offline RL that have been underexplored: the reliance on online interactions for hyperparameter tuning and the lack of performance guarantees before deployment. The motivating examples (healthcare, robotics) and the cycle diagram (Figure 1) effectively communicate why these issues matter for real-world applications. This clear problem framing strengthens the contribution's significance.\n\n**Rigorous theoretical analysis with frequentist justification.** Theorems 1 and 2 provide formal regret bounds in terms of PIL, with Theorem 2 establishing the optimal convergence rate for Bayesian offline RL under local asymptotic normality assumptions. The frequentist justification for a Bayesian approach is valuable and non-obvious. Proposition 1  decomposes PIL for Gaussian world models into interpretable terms (MSE and predictive variance), facilitating practical implementation and understanding.\n\n**Comprehensive empirical validation.** The paper includes extensive experiments across multiple environments (gymnax, Brax, D4RL) with careful ablations and statistical testing (e.g., Pearson correlation in Table 3). The comparison against online UCB-based hyperparameter tuning (Figure 3) provides compelling evidence of sample efficiency gains. Both SOReL and TOReL experiments demonstrate practical effectiveness, with particular strength in the near-oracle performance of ReBRAC+TOReL across diverse datasets (Table 1).", "weaknesses": "**Gap between theory and practice.** The theoretical regret bound (Theorem 1) is noted to be \"too conservative\" and is not used in practice; instead, the posterior predictive median (Eq. 5) is employed as a heuristic. This disconnect undermines the theoretical contribution's direct utility. The paper acknowledges that the bound's tightness depends critically on model accuracy relative to discount factor \n\\gamma, a constraint difficult to satisfy in practice (e.g., only pendulum-v1 yields non-trivial bounds in Figure 10b).\n\n**Missing comparisons with related work.** The paper does not compare against other offline hyperparameter tuning methods beyond citing them (e.g., Wang et al. 2022, Paine et al. 2020, or any related work of the same nature). Direct empirical comparisons would strengthen the work. There are some numerical experiments, but visual trends of training/eval curves would be helpful.", "questions": "**Q1:** Can the theoretical bound (Theorem 1) be tightened? Given that the bound is often too conservative in practice, are there domain-specific refinements or tighter bounds for structured environments that could improve practical utility?\n\n**Q2:** Why use the posterior predictive median rather than other quantiles (e.g., mean, upper confidence bound)? The choice of median is justified as a \"compromise\" but lacks formal justification. Have you tested other regret approximation metrics?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses two critical issues in offline reinforcement learning that have largely been overlooked: the lack of offline hyperparameter tuning metrics, and the absence of reliable offline regret approximation methods . The authors propose a Bayesian framework for offline RL, leveraging the Posterior Information Loss (PIL), the expected KL divergence between a learned model and true environment dynamics as an offline tuning signal. From this framework, they introduce two methods: SOReL (Safe Offline RL), which provides reliable offline regret estimates using predictive uncertainty to enable safe deployment, and TOReL (Tuning for Offline RL), which extends SOReL's offline hyperparameter tuning methodology to general model-based and model-free offline RL algorithms. The paper provides rigorous theoretical analysis showing that regret is bounded by the PIL (Theorem 1) and that Bayesian offline RL achieves the optimal convergence rate (Theorem 2). Empirical evaluations on D4RL, Adroit, and proprietary Brax datasets demonstrate that TOReL identifies near-oracle hyperparameters using only offline data, saving 20K to >200K online samples compared to existing online hyperparameter tuning methods, while SOReL accurately approximates true regret across diverse offline datasets.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**Well-motivated problem formulation.** The paper identifies and articulates two concrete, practically important issues in offline RL that have been underexplored: the reliance on online interactions for hyperparameter tuning and the lack of performance guarantees before deployment. The motivating examples (healthcare, robotics) and the cycle diagram (Figure 1) effectively communicate why these issues matter for real-world applications. This clear problem framing strengthens the contribution's significance.\n\n**Rigorous theoretical analysis with frequentist justification.** Theorems 1 and 2 provide formal regret bounds in terms of PIL, with Theorem 2 establishing the optimal convergence rate for Bayesian offline RL under local asymptotic normality assumptions. The frequentist justification for a Bayesian approach is valuable and non-obvious. Proposition 1  decomposes PIL for Gaussian world models into interpretable terms (MSE and predictive variance), facilitating practical implementation and understanding.\n\n**Comprehensive empirical validation.** The paper includes extensive experiments across multiple environments (gymnax, Brax, D4RL) with careful ablations and statistical testing (e.g., Pearson correlation in Table 3). The comparison against online UCB-based hyperparameter tuning (Figure 3) provides compelling evidence of sample efficiency gains. Both SOReL and TOReL experiments demonstrate practical effectiveness, with particular strength in the near-oracle performance of ReBRAC+TOReL across diverse datasets (Table 1).", "weaknesses": "**Gap between theory and practice.** The theoretical regret bound (Theorem 1) is noted to be \"too conservative\" and is not used in practice; instead, the posterior predictive median (Eq. 5) is employed as a heuristic. This disconnect undermines the theoretical contribution's direct utility. The paper acknowledges that the bound's tightness depends critically on model accuracy relative to discount factor \n\\gamma, a constraint difficult to satisfy in practice (e.g., only pendulum-v1 yields non-trivial bounds in Figure 10b).\n\n**Missing comparisons with related work.** The paper does not compare against other offline hyperparameter tuning methods beyond citing them (e.g., Wang et al. 2022, Paine et al. 2020, or any related work of the same nature). Direct empirical comparisons would strengthen the work. There are some numerical experiments, but visual trends of training/eval curves would be helpful.", "questions": "**Q1:** Can the theoretical bound (Theorem 1) be tightened? Given that the bound is often too conservative in practice, are there domain-specific refinements or tighter bounds for structured environments that could improve practical utility?\n\n**Q2:** Why use the posterior predictive median rather than other quantiles (e.g., mean, upper confidence bound)? The choice of median is justified as a \"compromise\" but lacks formal justification. Have you tested other regret approximation metrics?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761918744917}, {"id": "svi0M8OPtH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12284/Reviewer_sfCQ"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces SOReL and TOReL, two new algorithms that enable safe and reliable offline reinforcement learning by providing offline hyperparameter tuning and regret estimation, a problem so far underexplored in prior literature, thereby improving sample efficiency and performance guarantees without needing online interactions and moving offline RL closer to practical applicability.", "review_text": "The paper introduces SOReL and TOReL, two new algorithms that enable safe and reliable offline reinforcement learning by providing offline hyperparameter tuning and regret estimation, a problem so far underexplored in prior literature, thereby improving sample efficiency and performance guarantees without needing online interactions and moving offline RL closer to practical applicability.", "strengths": "- the theoretical considerations & derivation of the algorithms, leveraging PIL to select hyperparameters for the model appear novel and significant\n- the method appears to be able to well approximate the true regret, enabling fully offline hyperparameter tuning and thus bringing offline RL closer to real world applicability\n- the method is very flexible and can be used in principle with any existing offline RL algorithm of a users choice\n- the authors demonstrate accurate regret estimation on a variety of tasks / datasets", "weaknesses": "While I find the proposed method highly appealing in many ways, I find the paper has a couple of major weaknesses:\n\n1) A key contribution appears to be the PIL-based tuning of the model hyperparameters. I would argue however, that this is the \"easy\" part of offline RL hyperparameter tuning, since we can resort to simple supervised learning techniques like holding out 10% of the training data & measuring prediction performance on this set so select hyperparameters. I would argue the proposed method is much more sophisticated (directly tuning for regret minimization instead of dynamics prediction performance), and works likely better - but since this is a large part of the paper's contribution, I'd argue we still need a comparison against such a simple baseline.\n\n2) Apart from the PIL-based model parameter tuning, this kind of approach (bayesian model + sampling from posterior to make policy e.g. safe) has been studied in many prior works, such as [1,2] (consider adding in related work). No need for a direct comparison in experiments, but I'd argue the novelty in that sense it at least somewhat limited.\n\n3) Probably the biggest one: I would agree that so far no method has really solved the offline hyperparameter tuning / selection problem & your work towards achieving this appears highly promising, but I would not agree to your assessment in the related work section, that your method is the first to \"carry out all hyperparameter tuning [...] using only offline data\". Other methods have been developed for fully offline policy evaluation / hyperparameter selection (see e.g. the baselines in [3] and all the derivatives that were developed). While you compare your algorithm's performance against an online tuning algorithm and an oracle, the truly interesting thing to understand (at least to me it seems that way), would be whether the proposed method performs better than other offline tuning / selection methods - without these experiments, I believe the key merit of your method is not empirically validated.\n\n[1] Depeweg, Stefan, et al. \"Learning and policy search in stochastic dynamical systems with bayesian neural networks.\" ICLR 2017\n[2] Kaiser, Markus, et al. \"Bayesian decomposition of multi-modal dynamical systems for reinforcement learning.\" Neurocomputing 416 (2020): 352-359.\n[3] Fu, Justin, et al. \"Benchmarks for deep off-policy evaluation.\" ICLR (2021).", "questions": "- please clarify if I am mistaken: PIL is only used for model hyperparameter tuning - afterwards, to tune policy, we simply use the final selected model to quantify regret (via median return), correct?\n- if I am not mistaken, you refer to PIL with many different terms (posterior inforamtion loss, posterior information distance, information rate, ...) - this is a bit confusing, could you elaborate / align?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SOReL and TOReL, two new algorithms that enable safe and reliable offline reinforcement learning by providing offline hyperparameter tuning and regret estimation, a problem so far underexplored in prior literature, thereby improving sample efficiency and performance guarantees without needing online interactions and moving offline RL closer to practical applicability.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- the theoretical considerations & derivation of the algorithms, leveraging PIL to select hyperparameters for the model appear novel and significant\n- the method appears to be able to well approximate the true regret, enabling fully offline hyperparameter tuning and thus bringing offline RL closer to real world applicability\n- the method is very flexible and can be used in principle with any existing offline RL algorithm of a users choice\n- the authors demonstrate accurate regret estimation on a variety of tasks / datasets", "weaknesses": "While I find the proposed method highly appealing in many ways, I find the paper has a couple of major weaknesses:\n\n1) A key contribution appears to be the PIL-based tuning of the model hyperparameters. I would argue however, that this is the \"easy\" part of offline RL hyperparameter tuning, since we can resort to simple supervised learning techniques like holding out 10% of the training data & measuring prediction performance on this set so select hyperparameters. I would argue the proposed method is much more sophisticated (directly tuning for regret minimization instead of dynamics prediction performance), and works likely better - but since this is a large part of the paper's contribution, I'd argue we still need a comparison against such a simple baseline.\n\n2) Apart from the PIL-based model parameter tuning, this kind of approach (bayesian model + sampling from posterior to make policy e.g. safe) has been studied in many prior works, such as [1,2] (consider adding in related work). No need for a direct comparison in experiments, but I'd argue the novelty in that sense it at least somewhat limited.\n\n3) Probably the biggest one: I would agree that so far no method has really solved the offline hyperparameter tuning / selection problem & your work towards achieving this appears highly promising, but I would not agree to your assessment in the related work section, that your method is the first to \"carry out all hyperparameter tuning [...] using only offline data\". Other methods have been developed for fully offline policy evaluation / hyperparameter selection (see e.g. the baselines in [3] and all the derivatives that were developed). While you compare your algorithm's performance against an online tuning algorithm and an oracle, the truly interesting thing to understand (at least to me it seems that way), would be whether the proposed method performs better than other offline tuning / selection methods - without these experiments, I believe the key merit of your method is not empirically validated.\n\n[1] Depeweg, Stefan, et al. \"Learning and policy search in stochastic dynamical systems with bayesian neural networks.\" ICLR 2017\n[2] Kaiser, Markus, et al. \"Bayesian decomposition of multi-modal dynamical systems for reinforcement learning.\" Neurocomputing 416 (2020): 352-359.\n[3] Fu, Justin, et al. \"Benchmarks for deep off-policy evaluation.\" ICLR (2021).", "questions": "- please clarify if I am mistaken: PIL is only used for model hyperparameter tuning - afterwards, to tune policy, we simply use the final selected model to quantify regret (via median return), correct?\n- if I am not mistaken, you refer to PIL with many different terms (posterior inforamtion loss, posterior information distance, information rate, ...) - this is a bit confusing, could you elaborate / align?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761827102194}, {"id": "VxRr2pNL2P", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12284/Reviewer_ApZA"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "Current offline RL methods rely on extensive online interactions for hyperparameter tuning, and have no reliable bound on their initial online performance. To address these two issues, the authors introduce two algorithms. Firstly, SOReL: an algorithm for safe offline reinforcement learning. Using only offline data their Bayesian approach infers a posterior over environment dynamics to obtain a reliable estimate of the online performance via the posterior predictive uncertainty. Crucially, all hyperparameters are also tuned fully offline. Secondly, they introduce TOReL: a tuning for offline reinforcement learning algorithm that extends their information rate based offline hyperparameter tuning methods to general offline RL approaches. Lastly, they conduct empirical evaluations.", "review_text": "Current offline RL methods rely on extensive online interactions for hyperparameter tuning, and have no reliable bound on their initial online performance. To address these two issues, the authors introduce two algorithms. Firstly, SOReL: an algorithm for safe offline reinforcement learning. Using only offline data their Bayesian approach infers a posterior over environment dynamics to obtain a reliable estimate of the online performance via the posterior predictive uncertainty. Crucially, all hyperparameters are also tuned fully offline. Secondly, they introduce TOReL: a tuning for offline reinforcement learning algorithm that extends their information rate based offline hyperparameter tuning methods to general offline RL approaches. Lastly, they conduct empirical evaluations.", "strengths": "1. The problem of fully offline RL is interesting and important.\n2. The authors propose to solve the problem through a Bayesian view, which is an interesting idea.\n3. There are empirical evidence supporting the theoretical findings.", "weaknesses": "1. My main concern is about Theorem 2, since the last equation in line 274 is not correct. As N becomes sufficiently large, LHS is a constant while RHS converges to 0. If this is a typo, please provide the correct form.\n\n2. Discussions about Theorem 2 and its assumption is missing in the main part. Therefore it is hard to understand the implication of the Theorem. Some high-level discussions about the assumption in the main part would be helpful.\n\n3. The algorithms tune the hyperparameters fully offline, while it is unclear whether the approach is efficient. The PIL function includes multiple expectations and the posterior distribution, which seems hard to optimize directly. Are the optimization problems in both algorithms efficient in general?", "questions": "Please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Current offline RL methods rely on extensive online interactions for hyperparameter tuning, and have no reliable bound on their initial online performance. To address these two issues, the authors introduce two algorithms. Firstly, SOReL: an algorithm for safe offline reinforcement learning. Using only offline data their Bayesian approach infers a posterior over environment dynamics to obtain a reliable estimate of the online performance via the posterior predictive uncertainty. Crucially, all hyperparameters are also tuned fully offline. Secondly, they introduce TOReL: a tuning for offline reinforcement learning algorithm that extends their information rate based offline hyperparameter tuning methods to general offline RL approaches. Lastly, they conduct empirical evaluations.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The problem of fully offline RL is interesting and important.\n2. The authors propose to solve the problem through a Bayesian view, which is an interesting idea.\n3. There are empirical evidence supporting the theoretical findings.", "weaknesses": "1. My main concern is about Theorem 2, since the last equation in line 274 is not correct. As N becomes sufficiently large, LHS is a constant while RHS converges to 0. If this is a typo, please provide the correct form.\n\n2. Discussions about Theorem 2 and its assumption is missing in the main part. Therefore it is hard to understand the implication of the Theorem. Some high-level discussions about the assumption in the main part would be helpful.\n\n3. The algorithms tune the hyperparameters fully offline, while it is unclear whether the approach is efficient. The PIL function includes multiple expectations and the posterior distribution, which seems hard to optimize directly. Are the optimization problems in both algorithms efficient in general?", "questions": "Please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761728954780}], "openreview_url": "https://openreview.net/forum?id=LJ6AvummWu", "arxiv_id": "2505.22442", "paper_pdf": "papers/LJ6AvummWu.pdf", "paper_pdf_sha256": "101a7e7b321d394a985c9b6d12e48597c5aff6718f7d07974400677c95105b82", "paper_pdf_bytes": 6541472, "paper_pdf_source": "openreview", "code_url": "https://github.com/CWibault/sorel_torel", "code_repository": "CWibault/sorel_torel", "code_commit": "8babfc8950e8acbdbf252b1e6cd3f34048c17793", "code_archive": "repos/LJ6AvummWu.zip", "code_archive_sha256": "3c9fe3661e799299cff2a9b26a5a4e6c2b8cba127db80381da91ee2359f64844", "code_archive_bytes": 81895, "code_file_count": 41, "code_extensions": {".py": 37, ".sh": 4}, "github_disk_usage_kb": 60, "github_languages": {"Python": 216922, "Shell": 998}, "github_archived": false, "github_pushed_at": "2025-05-27T22:26:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sorel-and-torel-two-methods-for-fully-offline"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cWrqs2lwCJ", "year": 2025, "status": "rejected", "title": "Thinking Forward and Backward: Effective Backward Planning with Large Language Models", "authors": ["Allen Z. Ren", "brian ichter", "Anirudha Majumdar"], "authorids": ["~Allen_Z._Ren1", "~brian_ichter1", "~Anirudha_Majumdar1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have exhibited remarkable reasoning and planning capabilities. Most prior work in this area has used LLMs to reason through steps from an initial to a goal state or criterion, thereby effectively reasoning in a forward direction. Nonetheless, many planning problems exhibit an inherent asymmetry such that planning backward from the goal is significantly easier --- for example, if there are bottlenecks close to the goal. We take inspiration from this observation and demonstrate that this bias holds for LLM planning as well: planning performance in one direction correlates with the planning complexity of the problem in that direction. However, our experiments also reveal systematic biases which lead to poor planning in the backward direction. With this knowledge, we propose a backward planning algorithm for LLMs that first flips the problem and then plans forward in the flipped problem. This helps avoid the backward bias, generate more diverse candidate plans, and exploit asymmetries between the forward and backward directions in planning problems --- we find that combining planning in both directions with self-verification improves the overall planning success rates by 4-24% in three planning domains.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "p1jo35BfXz", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9205/Reviewer_SVvX"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The authors propose a method that uses large language models (LLMs) for planning by “flipping” the problem. In this approach, the initial and goal states are reversed, a plan is generated by the LLM, and then this plan is reversed to solve the original problem. The authors demonstrate that this method can enhance the planning performance of LLMs.", "review_text": "The authors propose a method that uses large language models (LLMs) for planning by “flipping” the problem. In this approach, the initial and goal states are reversed, a plan is generated by the LLM, and then this plan is reversed to solve the original problem. The authors demonstrate that this method can enhance the planning performance of LLMs.", "strengths": "The paper conducts a thorough set of experiments to test whether flipping the initial and goal states of problems improves LLM performance. The authors also examine whether the LLM can reason about whether or not to flip the problem.", "weaknesses": "The impact of the proposed approach is somewhat unclear. For example, prior research [1] has shown that LLM planning performance significantly deteriorates when the syntax of the planning domain is obfuscated, even when the causal structure remains intact (referred to as “Mystery” domains in [1]). It would have been insightful to see how this approach performs in such obfuscated domains. This also suggests that the computational complexity of the underlying domain doesn’t necessarily affect LLM performance. Therefore, flipping the problem’s initial and goal states to reduce computational complexity might not necessarily benefit LLMs, as they may not be sensitive to these complexity reductions in the same way formal planners are.\n\nBased on my understanding, the method described in Section 4 (always flipping the original problem) aligns with the “Flip” approach in Section 5. Flip performs worse than “Fwd” planning, contrary to what is suggested in Section 5. The proposed approach mentioned in Section 5 appears to be “FwdFlip,” where the decision to flip is made at random. It would be helpful to clarify the exact nature of the proposed approach. Additionally, it would be valuable to see details within the FwdFlip method, such as the success rates within Fwd and Flip respectively.", "questions": "1. Are the results of Flip related methods based on the solution to the flipped problem or the original problem? If it is w.r.t the original problem, how are the errors distributed between the solution of the flipped problem being wrong and the flipping of the plan being wrong?\n\n2. Is the flipped initial state in Blocksworld verified given that the LLM is generating a full state from the partial goal state?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a method that uses large language models (LLMs) for planning by “flipping” the problem. In this approach, the initial and goal states are reversed, a plan is generated by the LLM, and then this plan is reversed to solve the original problem. The authors demonstrate that this method can enhance the planning performance of LLMs.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper conducts a thorough set of experiments to test whether flipping the initial and goal states of problems improves LLM performance. The authors also examine whether the LLM can reason about whether or not to flip the problem.", "weaknesses": "The impact of the proposed approach is somewhat unclear. For example, prior research [1] has shown that LLM planning performance significantly deteriorates when the syntax of the planning domain is obfuscated, even when the causal structure remains intact (referred to as “Mystery” domains in [1]). It would have been insightful to see how this approach performs in such obfuscated domains. This also suggests that the computational complexity of the underlying domain doesn’t necessarily affect LLM performance. Therefore, flipping the problem’s initial and goal states to reduce computational complexity might not necessarily benefit LLMs, as they may not be sensitive to these complexity reductions in the same way formal planners are.\n\nBased on my understanding, the method described in Section 4 (always flipping the original problem) aligns with the “Flip” approach in Section 5. Flip performs worse than “Fwd” planning, contrary to what is suggested in Section 5. The proposed approach mentioned in Section 5 appears to be “FwdFlip,” where the decision to flip is made at random. It would be helpful to clarify the exact nature of the proposed approach. Additionally, it would be valuable to see details within the FwdFlip method, such as the success rates within Fwd and Flip respectively.", "questions": "1. Are the results of Flip related methods based on the solution to the flipped problem or the original problem? If it is w.r.t the original problem, how are the errors distributed between the solution of the flipped problem being wrong and the flipping of the plan being wrong?\n\n2. Is the flipped initial state in Blocksworld verified given that the LLM is generating a full state from the partial goal state?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730800834445}, {"id": "xNpMpMgBiR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9205/Reviewer_d5hd"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper uses LLMs for planning.  In particular it looks at the difference between planning forward (from initial state to the goal) vs planning backward (from the goal to the initial state), and notes that backward planning may work better for some problems.  For problems where backward planning is more efficient, the paper introduces a method for flipping the problem so that forward planning can be used from the goal to the initial state.  This leads to the question: Can LLMs plan better if they reason in the backward direction?”", "review_text": "This paper uses LLMs for planning.  In particular it looks at the difference between planning forward (from initial state to the goal) vs planning backward (from the goal to the initial state), and notes that backward planning may work better for some problems.  For problems where backward planning is more efficient, the paper introduces a method for flipping the problem so that forward planning can be used from the goal to the initial state.  This leads to the question: Can LLMs plan better if they reason in the backward direction?”", "strengths": "The idea of flipping the problem has merit for some problems.  With more careful consideration of the general claims about the capabilities of LLMs (many of which actually seem problem dependent) and a more clear definition of which problems this will work for and which it will not, and why (including a discussion of what makes some actions non-invertible), this would be a strong paper.", "weaknesses": "Overall: Many general claims are made about the capabilities of LLMs (e.g. they exhibit a systematic bias towards forward planning), and these claims are backed by experiments.  However, it's not clear that the experimental results justify these broad claims.  Many of the experimental results seem very domain-dependent.  For example, Directed Graph Planning has a built in directional bias. \n\n1.\tThe paper uses the term “backward planning” but all the planning is actually in the forward direction, it’s just that sometimes they “flip” the problem (making the old goal the new initial state and the old initial state the new goal). \n2.  The method proposed for flipping the problem assumes that all actions can be inverted, and it is not clear that this is always the case (so the applicability of “flipping” will depend on whether or not the actions in the domain are, in fact, invertible). \n3.\tRegarding the claims that “LLM exhibit a systematic bias of performing worse when planning backwards” may be attributed to the forward auto-regressive nature of LLM output generation or training data biases:  There is not enough evidence to support this as a general claim.  Experiments were conducted with only three domains, and at least one (Graph Planning) has by its very nature a built in bias in the forward direction.\n4.\tFigure 2 as an algorithmic description is not clear, and does not seem to match up with the text in lines 168-17.\n5.\tThe general conclusions about LLMs drawn from the data presented in Figure 3 (LLMs plan better in the forward direction, and that flip solves this bias) seem suspect given that the directed case in this domain has a built in bias to forward planning so it’s not surprising at all that flip would do well in this case.  \n6.\tAlso, regarding Figure 3; most of the cases where flip outperforms forward planning are also cases where backward planning already also outperformed forward planning.", "questions": "1.   What is it about some actions that make them non-invertible?  How does this affect your proposed method (which assumes that all actions are invertible)?\n2.   Is it important that inverse actions are actually feasible actions that respect physics and make sense?  (e.g. if an action is to “drill a hole”, you can’t really un-drill a hole; does this even matter?).\n3.\tBi-directional search is mentioned in related work, but it’s not mentioned at all in the text.  The approaches that involve both forward and backward planning seem very similar to bi-directional search and may warrant more discussion in the paper.  Are there lessons from bi-directional search that are relevant here?\n4.\tDoesn’t the Graph Planning domain have built in bias for directed graphs?  Meaning they may only be solvable in one direction?\n5.\tIn the description of how planning problems are flipped; the authors assume that every action has an inverse action.  It’s not clear that this would always be true.  Could you please include discussion of this?  Could you give examples where there is not an inverse action?  (e.g. if an action is to “drill a hole”, you can’t really un-drill a hole can you)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper uses LLMs for planning.  In particular it looks at the difference between planning forward (from initial state to the goal) vs planning backward (from the goal to the initial state), and notes that backward planning may work better for some problems.  For problems where backward planning is more efficient, the paper introduces a method for flipping the problem so that forward planning can be used from the goal to the initial state.  This leads to the question: Can LLMs plan better if they reason in the backward direction?”", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The idea of flipping the problem has merit for some problems.  With more careful consideration of the general claims about the capabilities of LLMs (many of which actually seem problem dependent) and a more clear definition of which problems this will work for and which it will not, and why (including a discussion of what makes some actions non-invertible), this would be a strong paper.", "weaknesses": "Overall: Many general claims are made about the capabilities of LLMs (e.g. they exhibit a systematic bias towards forward planning), and these claims are backed by experiments.  However, it's not clear that the experimental results justify these broad claims.  Many of the experimental results seem very domain-dependent.  For example, Directed Graph Planning has a built in directional bias. \n\n1.\tThe paper uses the term “backward planning” but all the planning is actually in the forward direction, it’s just that sometimes they “flip” the problem (making the old goal the new initial state and the old initial state the new goal). \n2.  The method proposed for flipping the problem assumes that all actions can be inverted, and it is not clear that this is always the case (so the applicability of “flipping” will depend on whether or not the actions in the domain are, in fact, invertible). \n3.\tRegarding the claims that “LLM exhibit a systematic bias of performing worse when planning backwards” may be attributed to the forward auto-regressive nature of LLM output generation or training data biases:  There is not enough evidence to support this as a general claim.  Experiments were conducted with only three domains, and at least one (Graph Planning) has by its very nature a built in bias in the forward direction.\n4.\tFigure 2 as an algorithmic description is not clear, and does not seem to match up with the text in lines 168-17.\n5.\tThe general conclusions about LLMs drawn from the data presented in Figure 3 (LLMs plan better in the forward direction, and that flip solves this bias) seem suspect given that the directed case in this domain has a built in bias to forward planning so it’s not surprising at all that flip would do well in this case.  \n6.\tAlso, regarding Figure 3; most of the cases where flip outperforms forward planning are also cases where backward planning already also outperformed forward planning.", "questions": "1.   What is it about some actions that make them non-invertible?  How does this affect your proposed method (which assumes that all actions are invertible)?\n2.   Is it important that inverse actions are actually feasible actions that respect physics and make sense?  (e.g. if an action is to “drill a hole”, you can’t really un-drill a hole; does this even matter?).\n3.\tBi-directional search is mentioned in related work, but it’s not mentioned at all in the text.  The approaches that involve both forward and backward planning seem very similar to bi-directional search and may warrant more discussion in the paper.  Are there lessons from bi-directional search that are relevant here?\n4.\tDoesn’t the Graph Planning domain have built in bias for directed graphs?  Meaning they may only be solvable in one direction?\n5.\tIn the description of how planning problems are flipped; the authors assume that every action has an inverse action.  It’s not clear that this would always be true.  Could you please include discussion of this?  Could you give examples where there is not an inverse action?  (e.g. if an action is to “drill a hole”, you can’t really un-drill a hole can you)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730647705966}, {"id": "0aWKXpjUDo", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9205/Reviewer_LPQo"], "rating": 1, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "This paper investigates the possible usage of LLMs as an automated planner for solving classical planning tasks (deterministic, fully observable). Especially, the paper studies the regression planning approach. The paper also proposes a heuristic approach that switches the initial state and the goal depending on the progress of solving the problem. Overall, experiments were conducted on path finding in graphs, array transformation, and the blocksworld domains.", "review_text": "This paper investigates the possible usage of LLMs as an automated planner for solving classical planning tasks (deterministic, fully observable). Especially, the paper studies the regression planning approach. The paper also proposes a heuristic approach that switches the initial state and the goal depending on the progress of solving the problem. Overall, experiments were conducted on path finding in graphs, array transformation, and the blocksworld domains.", "strengths": "This paper explores regression planning with LLMs. It seems like the authors are not aware of the literature on automated planning and classical AI, from what's written. so it is an attempt to re-invent the known results in the 70s to 90s in the context of large language models.", "weaknesses": "This paper mentions bidirectional search in the related work. However, it is mentioned nowhere in the paper about regression planning. \"Thinking forward and backward\" reminds me of a fancy title \"Thinking fast and slow\" and the title also shows effective backward planning. However, it is doubtful if that's the case.\nIf we hold back and remove the LLM portion, then the paper suggests solving planning problems by alternating the initial state and the goals. In general, goals are not a single state as the initial state in most of the problems. It is doubtful how such a switching strategy can lead to systematic improvement as we can always choose one of the goal states as a new initial state in a different problem. \nOne of the weaknesses here is the lack of coverage of the relevant literature.\n\nThere are several misleading statements such as \n\"classical planning algorithms such as BFS\" line 141. ==> BFS is not a classical planning algorithm\n\"bottleneck\" reminds me of the concept of information bottleneck in information theory. However, it is not relevant to this context.\nIt is not clear whether the \"bottleneck effect\" is relevant in this context.\n\n\nThe details of the implementation of the planner are vague. \nHow possible actions are generated? How the next state will be generated given actions?", "questions": "Question 1. What are the problem statistics of the three problem domains? They are all classical planning problems so we could see the action schema, and the number of state predicates to have a sense of the difficulty of the problem. What problem instances were tested? How many objects were tried? How the initial states and the goals were determined? It is not difficult to write down PDDL specifications for those problems.\n\nQuestion 2. Given a state observation, how the next actions are generated? Are the next actions sound and applicable to each state always? \nHow the next state is generated? How the goal test was done? \n\nQuestion 3. How the natural language reasoning version of the classical planning problems are obtained, for the graph planning an array transformation? Are they a natural description of the problem? Or is it utilizing artificial patterns? \n\nQuestion 4. Blocksworld problem would be the only single problem domain that has multiple goal states.\nIn regression planning, the difficulty is to maintain a set of states instead of a single state if we proceed to perform a search. \nThe experiment results also confirm such known issues. In the other two domains, regression modes are claimed to perform well. But I think it depends on the distribution of the initial state and the goal state, not the goal states here. I think if we simply flip the initial state and the goal state, we will obtain the opposite results. Could you explain this?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the possible usage of LLMs as an automated planner for solving classical planning tasks (deterministic, fully observable). Especially, the paper studies the regression planning approach. The paper also proposes a heuristic approach that switches the initial state and the goal depending on the progress of solving the problem. Overall, experiments were conducted on path finding in graphs, array transformation, and the blocksworld domains.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "This paper explores regression planning with LLMs. It seems like the authors are not aware of the literature on automated planning and classical AI, from what's written. so it is an attempt to re-invent the known results in the 70s to 90s in the context of large language models.", "weaknesses": "This paper mentions bidirectional search in the related work. However, it is mentioned nowhere in the paper about regression planning. \"Thinking forward and backward\" reminds me of a fancy title \"Thinking fast and slow\" and the title also shows effective backward planning. However, it is doubtful if that's the case.\nIf we hold back and remove the LLM portion, then the paper suggests solving planning problems by alternating the initial state and the goals. In general, goals are not a single state as the initial state in most of the problems. It is doubtful how such a switching strategy can lead to systematic improvement as we can always choose one of the goal states as a new initial state in a different problem. \nOne of the weaknesses here is the lack of coverage of the relevant literature.\n\nThere are several misleading statements such as \n\"classical planning algorithms such as BFS\" line 141. ==> BFS is not a classical planning algorithm\n\"bottleneck\" reminds me of the concept of information bottleneck in information theory. However, it is not relevant to this context.\nIt is not clear whether the \"bottleneck effect\" is relevant in this context.\n\n\nThe details of the implementation of the planner are vague. \nHow possible actions are generated? How the next state will be generated given actions?", "questions": "Question 1. What are the problem statistics of the three problem domains? They are all classical planning problems so we could see the action schema, and the number of state predicates to have a sense of the difficulty of the problem. What problem instances were tested? How many objects were tried? How the initial states and the goals were determined? It is not difficult to write down PDDL specifications for those problems.\n\nQuestion 2. Given a state observation, how the next actions are generated? Are the next actions sound and applicable to each state always? \nHow the next state is generated? How the goal test was done? \n\nQuestion 3. How the natural language reasoning version of the classical planning problems are obtained, for the graph planning an array transformation? Are they a natural description of the problem? Or is it utilizing artificial patterns? \n\nQuestion 4. Blocksworld problem would be the only single problem domain that has multiple goal states.\nIn regression planning, the difficulty is to maintain a set of states instead of a single state if we proceed to perform a search. \nThe experiment results also confirm such known issues. In the other two domains, regression modes are claimed to perform well. But I think it depends on the distribution of the initial state and the goal state, not the goal states here. I think if we simply flip the initial state and the goal state, we will obtain the opposite results. Could you explain this?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730467771983}, {"id": "SIjKzUwX3y", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9205/Reviewer_ceDr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper first investigates the performance of LLMs on a set of planning problems where both forward and backwards planning can be done (3 domains are given in the paper: shortest path in graphs, array and Blocksworld). The paper measures the difficulty of planning in a particular direction with the size of search space (in that direction). Based on this, the authors empirically observe and claim in Sec. 3.3 that 1) LLMs, in general, plan better in the \"easier\" planning direction and 2) ceteris paribus, LLMs plan worse in the backwards direction as compared to the forward direction. The authors hypothesize that 2) \"... may be attributed to the forward autoregressive nature of LLM output generation, as well as biases from the training dataset.\"\n\nThen, the paper, in Sec. 4, proposes a method which flips the problem backwards and plan forward (while handling some subtleties) for the flipped problem. Combined with self-verification and some sampling, the experiments show that their method outperforms other baselines over the 3 planning domains.", "review_text": "The paper first investigates the performance of LLMs on a set of planning problems where both forward and backwards planning can be done (3 domains are given in the paper: shortest path in graphs, array and Blocksworld). The paper measures the difficulty of planning in a particular direction with the size of search space (in that direction). Based on this, the authors empirically observe and claim in Sec. 3.3 that 1) LLMs, in general, plan better in the \"easier\" planning direction and 2) ceteris paribus, LLMs plan worse in the backwards direction as compared to the forward direction. The authors hypothesize that 2) \"... may be attributed to the forward autoregressive nature of LLM output generation, as well as biases from the training dataset.\"\n\nThen, the paper, in Sec. 4, proposes a method which flips the problem backwards and plan forward (while handling some subtleties) for the flipped problem. Combined with self-verification and some sampling, the experiments show that their method outperforms other baselines over the 3 planning domains.", "strengths": "1. The paper's approach in developing the main method is, to a certain extent, systematic. It first verifies that some planning problems are easier solved from the backwards direction (due to smaller action spaces) then show that LLMs perform better in the direction which is easier. However, the paper also found that LLMs cannot plan well directly in the backward direction. Combining these two observations, the paper then introduces its method of flipping the backward problem.\n\n2. The experimental results are laid out in a systematic manner, where each section tries to explain a certain statement or claim in the paper.  The experimental setup is also done well, although I have some comments about how the authors tried to link the results with certain claims (see next section).", "weaknesses": "I have several concerns regarding how claims in the papers are supported by the experiments. These concerns also impact the novelty and soundness of the authors' main method (**Fwd-Flip**).\n\n\n1. The authors mentioned that **Fwd** is the standard forward planning method, **Flip** is applying the forward planning method to the flipped problem, and **Fwd-Flip** is \"randomly choose either forward direction of the original problem or forward direction of the flipped problem and then plan\". Therefore, the only difference between **Fwd-Flip** and **Fwd**/**Flip** is that instead of deciding which direction to plan, the planning direction is randomly selected. However, the experimental results in Table 2 show that **Fwd-Flip** achieves much higher success rate than **Fwd** or **Flip**. To me, this is result is questionable because if **Fwd-Flip** is just selecting **Fwd** or **Flip** randomly, how can it perform much better? In fact, I would expect, with just random sampling, **Fwd-Flip** to perform exactly the average of **Fwd** and **Flip**? I also noticed also that Fwd-Flip is different from what was written in the introduction, which is \"Given a planning problem, we ask LLMs to sample possible plans in the forward direction of both the original problem and the flipped one, and then self-verify (Stechly et al., 2024) all the plans before choosing the final one\". Could the authors clarify?\n\n\n2. I also noticed that both **Fwd** and **Flip** (on their own) have similar performances over all planning tasks in Table 2. This implies that on average, flipping the problem has no benefits over the direct forward planning method. While I understand that the authors observed that certain planning directions could be easier in some tasks (due to difference in action spaces), it is clear from the experiment results that a particular planning direction is not inherently better in any of the tasks. Hence, I argue that the paper's main method, **Fwd-Flip** does not actually exploit this observation effectively. I have a suggestion on how the method can be improved: during test-time, we can cleverly decide whether the problem is easier if we flipped it (using some simple heuristics). Then, we can either apply **Fwd** or **Flip** based on our heuristic; this is more sound than **Fwd-Flip** proposed in this paper, which is simply selecting **Fwd** or **Flip** *randomly*. In later experiments, the authors do use the LLM to reason about which planning direction to use, but its effectiveness seems to be restricted to Undirected Graph Planning.\n\n\n3. Some claims in the papers are briefly stated in words but are not clearly substantiated from the experiments. For example:\n* Sec. 4 claims that flipping the problem \"... avoids the bias of weak LLM planning in the backward direction.\" However, Figure 3 shows that for Directed Graphs, for problems with no difference in forward and backward computations, flipping still performs worse than the forward problem. So, it is unclear if flipping the problem entirely avoids the bias.\n* There are results such as Figure 3 and Figure 10 that are used to justify the paper's claims that \"... LLM plans better in the direction\nof fewer computations needed, but the forward direction outperforms backward in general ...\". However, they are only shown for one planning task (instead of the 3 domains which the paper claims its focus is on). Since the paper mainly relies on empirical observations to justify its claims, I think the same figure should be replicated for other tasks as well (especially that I noticed there are additional space available in the main paper).\n* Table 3 tries to show that making the LLM decide which planning direction to use during test-time could improve performance. But somehow the table does not have the results for **Fwd-Flip**, the author's main proposed method for us to compare **Fwd-Flip-Reason** to. We also cannot just look at the previous results for **Fwd-Flip** because the authors mentioned that they used a different hyperparameter $M$ as compared to that in Table 2.", "questions": "1. \"We expect our proposed Fwd-Flip to outperform Fwd, Back, Flip, Fwd-Flip.\" I think the last Fwd-Flip is a typo - it should be Fwd-Back right?\n2. The method written in the introduction - \"Given a planning problem, we ask LLMs to sample possible plans in the forward direction of both the original problem and the flipped one, and then self-verify (Stechly et al., 2024) all the plans before choosing the final one\", is different from what was presented as the main method, which is **Fwd-Flip**. The description of **Fwd-Flip** given in the experiments is \"randomly choose either forward direction of the original problem or forward direction of the flipped problem and then plan\". This is very confusing because these are two totally different methods. I'd like the authors to clarify which is their main proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper first investigates the performance of LLMs on a set of planning problems where both forward and backwards planning can be done (3 domains are given in the paper: shortest path in graphs, array and Blocksworld). The paper measures the difficulty of planning in a particular direction with the size of search space (in that direction). Based on this, the authors empirically observe and claim in Sec. 3.3 that 1) LLMs, in general, plan better in the \"easier\" planning direction and 2) ceteris paribus, LLMs plan worse in the backwards direction as compared to the forward direction. The authors hypothesize that 2) \"... may be attributed to the forward autoregressive nature of LLM output generation, as well as biases from the training dataset.\"\n\nThen, the paper, in Sec. 4, proposes a method which flips the problem backwards and plan forward (while handling some subtleties) for the flipped problem. Combined with self-verification and some sampling, the experiments show that their method outperforms other baselines over the 3 planning domains.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper's approach in developing the main method is, to a certain extent, systematic. It first verifies that some planning problems are easier solved from the backwards direction (due to smaller action spaces) then show that LLMs perform better in the direction which is easier. However, the paper also found that LLMs cannot plan well directly in the backward direction. Combining these two observations, the paper then introduces its method of flipping the backward problem.\n\n2. The experimental results are laid out in a systematic manner, where each section tries to explain a certain statement or claim in the paper.  The experimental setup is also done well, although I have some comments about how the authors tried to link the results with certain claims (see next section).", "weaknesses": "I have several concerns regarding how claims in the papers are supported by the experiments. These concerns also impact the novelty and soundness of the authors' main method (**Fwd-Flip**).\n\n\n1. The authors mentioned that **Fwd** is the standard forward planning method, **Flip** is applying the forward planning method to the flipped problem, and **Fwd-Flip** is \"randomly choose either forward direction of the original problem or forward direction of the flipped problem and then plan\". Therefore, the only difference between **Fwd-Flip** and **Fwd**/**Flip** is that instead of deciding which direction to plan, the planning direction is randomly selected. However, the experimental results in Table 2 show that **Fwd-Flip** achieves much higher success rate than **Fwd** or **Flip**. To me, this is result is questionable because if **Fwd-Flip** is just selecting **Fwd** or **Flip** randomly, how can it perform much better? In fact, I would expect, with just random sampling, **Fwd-Flip** to perform exactly the average of **Fwd** and **Flip**? I also noticed also that Fwd-Flip is different from what was written in the introduction, which is \"Given a planning problem, we ask LLMs to sample possible plans in the forward direction of both the original problem and the flipped one, and then self-verify (Stechly et al., 2024) all the plans before choosing the final one\". Could the authors clarify?\n\n\n2. I also noticed that both **Fwd** and **Flip** (on their own) have similar performances over all planning tasks in Table 2. This implies that on average, flipping the problem has no benefits over the direct forward planning method. While I understand that the authors observed that certain planning directions could be easier in some tasks (due to difference in action spaces), it is clear from the experiment results that a particular planning direction is not inherently better in any of the tasks. Hence, I argue that the paper's main method, **Fwd-Flip** does not actually exploit this observation effectively. I have a suggestion on how the method can be improved: during test-time, we can cleverly decide whether the problem is easier if we flipped it (using some simple heuristics). Then, we can either apply **Fwd** or **Flip** based on our heuristic; this is more sound than **Fwd-Flip** proposed in this paper, which is simply selecting **Fwd** or **Flip** *randomly*. In later experiments, the authors do use the LLM to reason about which planning direction to use, but its effectiveness seems to be restricted to Undirected Graph Planning.\n\n\n3. Some claims in the papers are briefly stated in words but are not clearly substantiated from the experiments. For example:\n* Sec. 4 claims that flipping the problem \"... avoids the bias of weak LLM planning in the backward direction.\" However, Figure 3 shows that for Directed Graphs, for problems with no difference in forward and backward computations, flipping still performs worse than the forward problem. So, it is unclear if flipping the problem entirely avoids the bias.\n* There are results such as Figure 3 and Figure 10 that are used to justify the paper's claims that \"... LLM plans better in the direction\nof fewer computations needed, but the forward direction outperforms backward in general ...\". However, they are only shown for one planning task (instead of the 3 domains which the paper claims its focus is on). Since the paper mainly relies on empirical observations to justify its claims, I think the same figure should be replicated for other tasks as well (especially that I noticed there are additional space available in the main paper).\n* Table 3 tries to show that making the LLM decide which planning direction to use during test-time could improve performance. But somehow the table does not have the results for **Fwd-Flip**, the author's main proposed method for us to compare **Fwd-Flip-Reason** to. We also cannot just look at the previous results for **Fwd-Flip** because the authors mentioned that they used a different hyperparameter $M$ as compared to that in Table 2.", "questions": "1. \"We expect our proposed Fwd-Flip to outperform Fwd, Back, Flip, Fwd-Flip.\" I think the last Fwd-Flip is a typo - it should be Fwd-Back right?\n2. The method written in the introduction - \"Given a planning problem, we ask LLMs to sample possible plans in the forward direction of both the original problem and the flipped one, and then self-verify (Stechly et al., 2024) all the plans before choosing the final one\", is different from what was presented as the main method, which is **Fwd-Flip**. The description of **Fwd-Flip** given in the experiments is \"randomly choose either forward direction of the original problem or forward direction of the flipped problem and then plan\". This is very confusing because these are two totally different methods. I'd like the authors to clarify which is their main proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729836106375}], "openreview_url": "https://openreview.net/forum?id=cWrqs2lwCJ", "arxiv_id": "2411.01790", "paper_pdf": "papers/cWrqs2lwCJ.pdf", "paper_pdf_sha256": "4ee1a03edcc2ea1b74f12996f37a6e59ab01cbbb52cc4fde1e03f95ff03a6da9", "paper_pdf_bytes": 4620505, "paper_pdf_source": "openreview", "code_url": "https://github.com/irom-princeton/llm-backward", "code_repository": "irom-princeton/llm-backward", "code_commit": "b215c6785a18803a4b116f06e6850a19cc7a6048", "code_archive": "repos/cWrqs2lwCJ.zip", "code_archive_sha256": "bb0ad5cafa0bcf8812a3d139d30e81825a5104a49e0703a4b8ebe82c9f120f19", "code_archive_bytes": 22707, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 18, "github_languages": {"Python": 78294}, "github_archived": false, "github_pushed_at": "2024-11-05T03:32:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/thinking-forward-and-backward-effective"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "53kW6e1uNN", "year": 2024, "status": "rejected", "title": "AFDGCF: Adaptive Feature De-correlation Graph Collaborative Filtering for Recommendations", "authors": ["Wei Wu", "Chao Wang", "Dazhong Shen", "Chuan Qin", "Hui Xiong"], "authorids": ["~Wei_Wu25", "~Chao_Wang14", "~Dazhong_Shen1", "~Chuan_Qin1", "~Hui_Xiong1"], "authors_source": "OpenReview API", "abstract": "Collaborative filtering methods based on graph neural networks (GNNs) have witnessed significant success in recommender systems (RS), capitalizing on their ability to capture collaborative signals within intricate user-item relationships via message-passing mechanisms. However, these GNN-based RS inadvertently introduce a linear correlation between user and item embeddings, contradicting the goal of providing personalized recommendations. While existing research predominantly ascribes this flaw to the over-smoothing problem, this paper underscores the critical, often overlooked role of the over-correlation issue in diminishing the effectiveness of GNN representations and subsequent recommendation performance. The unclear relationship between over-correlation and over-smoothing in RS, coupled with the challenge of adaptively minimizing the impact of over-correlation while preserving collaborative filtering signals, is quite challenging. To this end, this paper aims to address the aforementioned gap by undertaking a comprehensive study of the over-correlation issue in graph collaborative filtering models. Empirical evidence substantiates the widespread prevalence of over-correlation in these models. Furthermore, a theoretical analysis establishes a pivotal connection between the over-correlation and over-smoothing predicaments. Leveraging these insights, we introduce the Adaptive Feature De-correlation Graph Collaborative Filtering (AFDGCF) Framework, which dynamically applies correlation penalties to the feature dimensions of the representation matrix, effectively alleviating both over-correlation and over-smoothing challenges. The efficacy of the proposed framework is corroborated through extensive experiments conducted with four different graph collaborative filtering models across four publicly available datasets, demonstrating the superiority of AFDGCF in enhancing the performance landscape of graph collaborative filtering models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NSYqyoNPMK", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3105/Reviewer_Xkxj"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper discusses the possible connections between over-smoothing and over-correlation in graph neural networks-based recommender systems. Indeed, while over-smoothing has been debated in graph-based recommendation for quite some time now, the authors claim over-correlation is still not properly analysed as happening in graph representation learning. Through an initial empirical study, the authors demonstrate that the negative effects of the two issues seem to be directly dependent and go along with the performance degradation of the models (i.e., usually after the third message-passing layer). After that, the paper underlines how over-smoothing and over-correlation may present a direct mapping to rows and columns in the node embedding matrix, respectively, and mathematically proves that the two are proportional. In this respect, as alleviating one of the two would tackle also the other, the authors propose a loss function named adaptive feature decorrelation, that comes into a static and dynamic version. An extensive experimental setting comprising four recommendation datasets and nine baselines demonstrates the efficacy of the proposed approach. Indeed, when applied to existing graph-based recommender systems, the adaptive feature decorrelation loss function is beneficial to improve the performance in terms of recommendation accuracy and requiring much less epochs to reach convergence. Finally, an ablation study justifies the soundness of the proposed architectural choices.", "review_text": "The paper discusses the possible connections between over-smoothing and over-correlation in graph neural networks-based recommender systems. Indeed, while over-smoothing has been debated in graph-based recommendation for quite some time now, the authors claim over-correlation is still not properly analysed as happening in graph representation learning. Through an initial empirical study, the authors demonstrate that the negative effects of the two issues seem to be directly dependent and go along with the performance degradation of the models (i.e., usually after the third message-passing layer). After that, the paper underlines how over-smoothing and over-correlation may present a direct mapping to rows and columns in the node embedding matrix, respectively, and mathematically proves that the two are proportional. In this respect, as alleviating one of the two would tackle also the other, the authors propose a loss function named adaptive feature decorrelation, that comes into a static and dynamic version. An extensive experimental setting comprising four recommendation datasets and nine baselines demonstrates the efficacy of the proposed approach. Indeed, when applied to existing graph-based recommender systems, the adaptive feature decorrelation loss function is beneficial to improve the performance in terms of recommendation accuracy and requiring much less epochs to reach convergence. Finally, an ablation study justifies the soundness of the proposed architectural choices.", "strengths": "+ The addressed problem (i.e., over-smoothing and over-correlation in graph-based recommendation) is relatively new to the literature.\n+ The empirical analysis supported by the mathematical proofs help justifying the existing problem and opening to possible solutions.\n+ The experimental setting is extensive with numerous evaluation dimensions.\n+ The code and datasets are released at review time.", "weaknesses": "- Some details about the introduced methodology need to be clarified.\n- The authors may have not considered other graph-based recommendation baselines whose solutions are like the proposed one.\n\n**After the rebuttal.** The rebuttal clarified all weaknesses.", "questions": "* To the best of my understanding, I cannot find the reason why the authors state that “it is crucial to maintain the smoothness of deep representations while restricting the feature correlations of the model’s representations” (beginning of page 7). The paper seems to claim that when reducing over-correlation for deeper representations, also over-smoothing will be tackled. In this sense, I cannot see the point in the quoted statement. Would you please elaborate on that?\n* Did the authors consider graph-based recommendation approaches which leverage decorrelation in a similar manner to the proposed one (e.g., disentangled graph collaborative filtering, DGCF [1]). In authors’ opinion, what would it be (even intuitively) the effect of performing a double decorrelation if the proposed loss function was applied to DGCF? Would it have a positive or a negative impact, and why?\n\n[1] Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, Tat-Seng Chua: Disentangled Graph Collaborative Filtering. SIGIR 2020: 1001-1010\n\n**After the rebuttal.** The rebuttal answered all questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper discusses the possible connections between over-smoothing and over-correlation in graph neural networks-based recommender systems. Indeed, while over-smoothing has been debated in graph-based recommendation for quite some time now, the authors claim over-correlation is still not properly analysed as happening in graph representation learning. Through an initial empirical study, the authors demonstrate that the negative effects of the two issues seem to be directly dependent and go along with the performance degradation of the models (i.e., usually after the third message-passing layer). After that, the paper underlines how over-smoothing and over-correlation may present a direct mapping to rows and columns in the node embedding matrix, respectively, and mathematically proves that the two are proportional. In this respect, as alleviating one of the two would tackle also the other, the authors propose a loss function named adaptive feature decorrelation, that comes into a static and dynamic version. An extensive experimental setting comprising four recommendation datasets and nine baselines demonstrates the efficacy of the proposed approach. Indeed, when applied to existing graph-based recommender systems, the adaptive feature decorrelation loss function is beneficial to improve the performance in terms of recommendation accuracy and requiring much less epochs to reach convergence. Finally, an ablation study justifies the soundness of the proposed architectural choices.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "+ The addressed problem (i.e., over-smoothing and over-correlation in graph-based recommendation) is relatively new to the literature.\n+ The empirical analysis supported by the mathematical proofs help justifying the existing problem and opening to possible solutions.\n+ The experimental setting is extensive with numerous evaluation dimensions.\n+ The code and datasets are released at review time.", "weaknesses": "- Some details about the introduced methodology need to be clarified.\n- The authors may have not considered other graph-based recommendation baselines whose solutions are like the proposed one.\n\n**After the rebuttal.** The rebuttal clarified all weaknesses.", "questions": "* To the best of my understanding, I cannot find the reason why the authors state that “it is crucial to maintain the smoothness of deep representations while restricting the feature correlations of the model’s representations” (beginning of page 7). The paper seems to claim that when reducing over-correlation for deeper representations, also over-smoothing will be tackled. In this sense, I cannot see the point in the quoted statement. Would you please elaborate on that?\n* Did the authors consider graph-based recommendation approaches which leverage decorrelation in a similar manner to the proposed one (e.g., disentangled graph collaborative filtering, DGCF [1]). In authors’ opinion, what would it be (even intuitively) the effect of performing a double decorrelation if the proposed loss function was applied to DGCF? Would it have a positive or a negative impact, and why?\n\n[1] Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, Tat-Seng Chua: Disentangled Graph Collaborative Filtering. SIGIR 2020: 1001-1010\n\n**After the rebuttal.** The rebuttal answered all questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699260046224}, {"id": "hcf5K2xIpR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3105/Reviewer_HBXd"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper analyzes the feature correlation issues in graph collaborative filtering. The author(s) present empirical studies on the smoothness and correlation of each layer of various graph collaborative filtering methods. Then, the author(s) propose AFDGCF that incorporates an auxiliary loss function to explicitly optimize the over-correlation issue. Extensive experiments on four public datasets and four popular GCF backbones show the effectiveness of the proposed method. Code is available and the author(s) promise to release all the code after the reviewing phase.", "review_text": "The paper analyzes the feature correlation issues in graph collaborative filtering. The author(s) present empirical studies on the smoothness and correlation of each layer of various graph collaborative filtering methods. Then, the author(s) propose AFDGCF that incorporates an auxiliary loss function to explicitly optimize the over-correlation issue. Extensive experiments on four public datasets and four popular GCF backbones show the effectiveness of the proposed method. Code is available and the author(s) promise to release all the code after the reviewing phase.", "strengths": "1. The paper studies an important task, i.e., graph collaborative filtering.\n2. The proposed model is implemented by an open-source framework, making it easy to reproduce. Code is available during the reviewing phase.\n3. Extensive experiments on four public datasets and four popular GCF backbones show the effectiveness of the proposed method.", "weaknesses": "1. Limited novelty. The paper seems like a straightforward application of existing literature, specifically the DeCorr [1] that focuses on general deep graph neural networks, in a specific application domain. The contribution of this study is mainly the transposition of DeCorr's insights into graph collaborative filtering, with different datasets and backbones. Although modifications like different penalty coefficients for users and items are also proposed, the whole paper still lack enough insights about what are unique challenges of overcorrelation in recommender systems.\n\n2. It could be better if one additional figure could be illustrated, i.e., how Corr and SMV metrics evolve with the application of additional network layers—mirroring the Figure 2, but explicitly showcasing the effects of the proposed method—the authors could convincingly validate their auxiliary loss function's efficacy.\n\n3. Presentation issues. The y-axis labels of Figure 2 lack standardization, e.g., 0.26 vs. 0.260 vs. 2600 vs. .2600.\n\n[1] Jin et al. Feature overcorrelation in deep graph neural networks: A new perspective. KDD 2022.", "questions": "According to Theorem 1, there exists a proportional relationship between column correlation and row correlation of a matrix. So whether existing works on alleviating row correlation issues like contrastive learning also solve the correlation issues? Once the row correlation is alleviated, according to the proportional relationship, the column correlation should be alleviated as well. If so, why do we need the proposed auxiliary loss to explicitly alleviate the column correlation issue?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper analyzes the feature correlation issues in graph collaborative filtering. The author(s) present empirical studies on the smoothness and correlation of each layer of various graph collaborative filtering methods. Then, the author(s) propose AFDGCF that incorporates an auxiliary loss function to explicitly optimize the over-correlation issue. Extensive experiments on four public datasets and four popular GCF backbones show the effectiveness of the proposed method. Code is available and the author(s) promise to release all the code after the reviewing phase.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper studies an important task, i.e., graph collaborative filtering.\n2. The proposed model is implemented by an open-source framework, making it easy to reproduce. Code is available during the reviewing phase.\n3. Extensive experiments on four public datasets and four popular GCF backbones show the effectiveness of the proposed method.", "weaknesses": "1. Limited novelty. The paper seems like a straightforward application of existing literature, specifically the DeCorr [1] that focuses on general deep graph neural networks, in a specific application domain. The contribution of this study is mainly the transposition of DeCorr's insights into graph collaborative filtering, with different datasets and backbones. Although modifications like different penalty coefficients for users and items are also proposed, the whole paper still lack enough insights about what are unique challenges of overcorrelation in recommender systems.\n\n2. It could be better if one additional figure could be illustrated, i.e., how Corr and SMV metrics evolve with the application of additional network layers—mirroring the Figure 2, but explicitly showcasing the effects of the proposed method—the authors could convincingly validate their auxiliary loss function's efficacy.\n\n3. Presentation issues. The y-axis labels of Figure 2 lack standardization, e.g., 0.26 vs. 0.260 vs. 2600 vs. .2600.\n\n[1] Jin et al. Feature overcorrelation in deep graph neural networks: A new perspective. KDD 2022.", "questions": "According to Theorem 1, there exists a proportional relationship between column correlation and row correlation of a matrix. So whether existing works on alleviating row correlation issues like contrastive learning also solve the correlation issues? Once the row correlation is alleviated, according to the proportional relationship, the column correlation should be alleviated as well. If so, why do we need the proposed auxiliary loss to explicitly alleviate the column correlation issue?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699141836356}, {"id": "GT4PF8MI1g", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3105/Reviewer_vpbW"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors focus on analyzing feature over-correlation in graph-based collaborative filtering, and propose an adaptive feature de-correlation regularization in graph-based collaborative filtering. Column-wise feature over-correlation will introduce redundant information for representation learning, the proposed feature de-correlation regularization can significantly improve the representation quality. Besides, the proposed feature de-correlation is very flexible and lightweight, which can coupled with representation-based CF. Experiments on several benchmarks show the effectiveness of the proposed method.", "review_text": "In this paper, the authors focus on analyzing feature over-correlation in graph-based collaborative filtering, and propose an adaptive feature de-correlation regularization in graph-based collaborative filtering. Column-wise feature over-correlation will introduce redundant information for representation learning, the proposed feature de-correlation regularization can significantly improve the representation quality. Besides, the proposed feature de-correlation is very flexible and lightweight, which can coupled with representation-based CF. Experiments on several benchmarks show the effectiveness of the proposed method.", "strengths": "1. Interesting research topic of this paper, tacking feature over-correlation in collaborative filtering is an effective direction.\n2. The proposed feature de-correlation regularization is flexible and effective in graph-based collaborative filtering. De-correlation is helpful in learning more high-quality representation for collaborative filtering.\n3. Experiments conducted on several graph-based backbones demonstrate the effectiveness of the proposed de-correlation regularization.", "weaknesses": "1. The motivation of this paper should be highlighted. Why do the authors analyze over-correlation combined with over-smoothing? Does feature over-correlation only occur on graph-based collaborative filtering non other methods such as Matrix Factorization? \n2. The reason for existing over-correlation in low-dimensional collaborative filtering is not clear. It will be more interesting if the authors deeply explain the behind reasons. Besides, does alleviating over-correlation can help to reduce over-smoothing issues in graph-based collaborative filtering? The authors should give a more explanatory illustration.\n3. Lacking comparisons of related works, disentangled collaborative filtering should be involved. Besides, column-wise de-correlation can be also viewed as self-supervised learning[1]. The authors should discuss with current self-supervised graph collaborative filtering method[2,3,4].\n[1]Wang X, Jin H, Zhang A, et al. Disentangled graph collaborative filtering[C]//Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 2020: 1001-1010.\n[2]Wu J, Wang X, Feng F, et al. Self-supervised graph learning for recommendation[C]//Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2021: 726-735.\n[3]Yu J, Yin H, Xia X, et al. Are graph augmentations necessary? simple graph contrastive learning for recommendation[C]//Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 2022: 1294-1303.\n[4]Yang, Y., Wu, Z., Wu, L., Zhang, K., Hong, R., Zhang, Z., ... & Wang, M. (2023). Generative-Contrastive Graph Learning for Recommendation.", "questions": "Mentioned as the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors focus on analyzing feature over-correlation in graph-based collaborative filtering, and propose an adaptive feature de-correlation regularization in graph-based collaborative filtering. Column-wise feature over-correlation will introduce redundant information for representation learning, the proposed feature de-correlation regularization can significantly improve the representation quality. Besides, the proposed feature de-correlation is very flexible and lightweight, which can coupled with representation-based CF. Experiments on several benchmarks show the effectiveness of the proposed method.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. Interesting research topic of this paper, tacking feature over-correlation in collaborative filtering is an effective direction.\n2. The proposed feature de-correlation regularization is flexible and effective in graph-based collaborative filtering. De-correlation is helpful in learning more high-quality representation for collaborative filtering.\n3. Experiments conducted on several graph-based backbones demonstrate the effectiveness of the proposed de-correlation regularization.", "weaknesses": "1. The motivation of this paper should be highlighted. Why do the authors analyze over-correlation combined with over-smoothing? Does feature over-correlation only occur on graph-based collaborative filtering non other methods such as Matrix Factorization? \n2. The reason for existing over-correlation in low-dimensional collaborative filtering is not clear. It will be more interesting if the authors deeply explain the behind reasons. Besides, does alleviating over-correlation can help to reduce over-smoothing issues in graph-based collaborative filtering? The authors should give a more explanatory illustration.\n3. Lacking comparisons of related works, disentangled collaborative filtering should be involved. Besides, column-wise de-correlation can be also viewed as self-supervised learning[1]. The authors should discuss with current self-supervised graph collaborative filtering method[2,3,4].\n[1]Wang X, Jin H, Zhang A, et al. Disentangled graph collaborative filtering[C]//Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 2020: 1001-1010.\n[2]Wu J, Wang X, Feng F, et al. Self-supervised graph learning for recommendation[C]//Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2021: 726-735.\n[3]Yu J, Yin H, Xia X, et al. Are graph augmentations necessary? simple graph contrastive learning for recommendation[C]//Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 2022: 1294-1303.\n[4]Yang, Y., Wu, Z., Wu, L., Zhang, K., Hong, R., Zhang, Z., ... & Wang, M. (2023). Generative-Contrastive Graph Learning for Recommendation.", "questions": "Mentioned as the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698767318423}, {"id": "85Hon1y31W", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3105/Reviewer_TKXS"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper draws attention to the challenges of over-smoothing and over-correlation in GNN-based collaborative filtering methods. In particular, the paper provides a detailed analysis of the over-correlation problem, which has been largely overlooked in existing works. Through rigorous theoretical analysis, the paper establishes a proportional association between the over-smoothing issue and the over-correlation issue, shedding light on their interconnected nature.\n\nTo tackle these issues, the paper proposes a model-agnostic constraint with adaptive weights. This constraint is designed to effectively mitigate over-smoothing and over-correlation problems in GNN-based collaborative filtering. The adaptive weights allow the constraint to dynamically adjust and optimize the learning process.\n\nComprehensive experiments are conducted to validate the effectiveness of the proposed constraint. The results demonstrate significant improvements in overall performance, enhanced training efficiency, and the efficacy of the adaptive approach. These findings provide strong evidence for the practical benefits of the proposed constraint in addressing the over-smoothing and over-correlation challenges in GNN-based collaborative filtering methods.", "review_text": "This paper draws attention to the challenges of over-smoothing and over-correlation in GNN-based collaborative filtering methods. In particular, the paper provides a detailed analysis of the over-correlation problem, which has been largely overlooked in existing works. Through rigorous theoretical analysis, the paper establishes a proportional association between the over-smoothing issue and the over-correlation issue, shedding light on their interconnected nature.\n\nTo tackle these issues, the paper proposes a model-agnostic constraint with adaptive weights. This constraint is designed to effectively mitigate over-smoothing and over-correlation problems in GNN-based collaborative filtering. The adaptive weights allow the constraint to dynamically adjust and optimize the learning process.\n\nComprehensive experiments are conducted to validate the effectiveness of the proposed constraint. The results demonstrate significant improvements in overall performance, enhanced training efficiency, and the efficacy of the adaptive approach. These findings provide strong evidence for the practical benefits of the proposed constraint in addressing the over-smoothing and over-correlation challenges in GNN-based collaborative filtering methods.", "strengths": "- The paper highlights the issue of decorrelation in collaborative filtering, which has received little attention in previous works.\n- Through a comprehensive theoretical analysis, the paper establishes a clear association between the over-smoothing problem and the decorrelation issue.\n- To address the challenges of over-smoothing and decorrelation, the paper proposes an effective solution. The proposed scheme is extensively evaluated through rigorous experiments, demonstrating its effectiveness.\n- The paper is well-written and provides clear explanations. It includes illustrative figures and pilot experiments that enhance understanding and readability.", "weaknesses": "- I have reservations regarding the dataset preprocessing approach employed in the paper. The authors chose to exclude users and items with fewer than 15/10 interactions in some datasets. However, in my experience, this approach has the potential to create highly dense datasets and introduce bias.\n- It would have been beneficial if the paper had explored the recent advancements in self-supervised learning for collaborative filtering, as these techniques have demonstrated superior performance in related studies.", "questions": "I would expect the authors to clarify the two issues mentioned in the weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper draws attention to the challenges of over-smoothing and over-correlation in GNN-based collaborative filtering methods. In particular, the paper provides a detailed analysis of the over-correlation problem, which has been largely overlooked in existing works. Through rigorous theoretical analysis, the paper establishes a proportional association between the over-smoothing issue and the over-correlation issue, shedding light on their interconnected nature.\n\nTo tackle these issues, the paper proposes a model-agnostic constraint with adaptive weights. This constraint is designed to effectively mitigate over-smoothing and over-correlation problems in GNN-based collaborative filtering. The adaptive weights allow the constraint to dynamically adjust and optimize the learning process.\n\nComprehensive experiments are conducted to validate the effectiveness of the proposed constraint. The results demonstrate significant improvements in overall performance, enhanced training efficiency, and the efficacy of the adaptive approach. These findings provide strong evidence for the practical benefits of the proposed constraint in addressing the over-smoothing and over-correlation challenges in GNN-based collaborative filtering methods.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- The paper highlights the issue of decorrelation in collaborative filtering, which has received little attention in previous works.\n- Through a comprehensive theoretical analysis, the paper establishes a clear association between the over-smoothing problem and the decorrelation issue.\n- To address the challenges of over-smoothing and decorrelation, the paper proposes an effective solution. The proposed scheme is extensively evaluated through rigorous experiments, demonstrating its effectiveness.\n- The paper is well-written and provides clear explanations. It includes illustrative figures and pilot experiments that enhance understanding and readability.", "weaknesses": "- I have reservations regarding the dataset preprocessing approach employed in the paper. The authors chose to exclude users and items with fewer than 15/10 interactions in some datasets. However, in my experience, this approach has the potential to create highly dense datasets and introduce bias.\n- It would have been beneficial if the paper had explored the recent advancements in self-supervised learning for collaborative filtering, as these techniques have demonstrated superior performance in related studies.", "questions": "I would expect the authors to clarify the two issues mentioned in the weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698544610121}], "openreview_url": "https://openreview.net/forum?id=53kW6e1uNN", "arxiv_id": "2403.17416", "paper_pdf": "papers/53kW6e1uNN.pdf", "paper_pdf_sha256": "f2c133df5a7f07873788c4cdf3ae3e1e2e8354091dca97c5e6ac53b1ed79e345", "paper_pdf_bytes": 6094555, "paper_pdf_source": "openreview", "code_url": "https://github.com/U-rara/AFDGCF", "code_repository": "U-rara/AFDGCF", "code_commit": "0eba7b6dbf7cfd2c1a0dcebd9e2f0742a68b89aa", "code_archive": "repos/53kW6e1uNN.zip", "code_archive_sha256": "38919b617762d350bf5da4ffcd8222d0bf84bfcab807a169ba094aec4410d806", "code_archive_bytes": 29630, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 16, "github_languages": {"Python": 81015}, "github_archived": false, "github_pushed_at": "2024-01-23T13:47:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/afdgcf-adaptive-feature-de-correlation-graph"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "t-hNmA0cVSW", "year": 2023, "status": "rejected", "title": "Semi-supervised Counting via Pixel-by-pixel Density Distribution Modelling", "authors": ["Hui LIN", "Zhiheng Ma", "Rongrong Ji", "Yaowei Wang", "su zhou", "Xiaopeng Hong"], "authorids": ["~Hui_LIN4", "~Zhiheng_Ma1", "~Rongrong_Ji5", "~Yaowei_Wang1", "~su_zhou1", "~Xiaopeng_Hong4"], "authors_source": "OpenReview API", "abstract": "This paper focuses on semi-supervised crowd counting, where only a small portion of the training data are labeled. We formulate the pixel-wise density value to regress as a probability distribution, instead of a single deterministic value, and utilize a dual-branch structure to model the corresponding discrete form of the distribution function. On the basis, we propose a semi-supervised crowd counting model. Firstly, we enhance the transformer decoder by usingdensity tokens to specialize the forwards of decoders w.r.t. different density intervals; Secondly, we design a pixel-wise distribution matching loss to measure the differences in the pixel-wise density distributions between the prediction and the ground-truth; Thirdly, we propose an interleaving consistency regularization term to align the prediction of two branches and make them consistent. Extensive experiments on four datasets are performed to show that our method clearly outperforms the competitors by a large margin under various labeled ratio settings.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1Y5PNGT0up", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper354/Reviewer_yGdj"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper considers the problem of predicting density maps for counting given semi-supervised datasets.\nIt proposes a novel transformer-based architecture, a distribution-matching loss function for quantized density prediction, the use of two overlapping discretisations and a consistency loss for the two discretisations that is only applied for confident predictions in unsupervised images.\nThe method achieves strong results in both fully and partially supervised settings, achieving state-of-the-art results.\nSome aspects of the design are demonstrated to be important through ablative experiments.", "review_text": "The paper describes an effective loss and parametrization for predicting continuous densities, as well as a simple consistency loss for unsupervised images. The transformer architecture is not demonstrated to be necessary and I wonder whether a simple conv-net would work as well using these loss functions. The ablative experiments should also be expanded, and it would be better to evaluate the supervised loss in the fully-supervised setting. I'm leaning towards accept. I may increase my rating if these issues are addressed or decrease it if they are not.", "strengths": "**Strengths**\n\n1. The distribution matching loss is nice.\n1. It's neat to use the uncertainty of the predicted discrete distribution to obtain the mask. I like that the non-confident examples are ignored by the ECR loss.\n1. The supervised loss gives a significant improvement over the baseline methods in the fully-supervised setting (Table 5, appendix).\n1. To me, it was surprising that the ECR loss, despite being simple, gave a large improvement in the semi-supervised setting (Table 2).\n1. The attention map visualizations look great (appendix).\n\n**Weaknesses**\n\n1. The authors argue that predicting a distribution is more credible and less noisy than predicting a single value. However, the targets are always one-hot and the distribution is simply used to parametrize the prediction via its expectation, so it seems like simply an alternative parametrization and loss for predicting a single value. Another plausible explanation is that this parametrization and loss simply improve learnability and/or generalization for training a deep network.\n1. The loss in eq. 4 seems more general than the loss which is actually used, since $\\mathbf{y}$ is always one-hot in practice. Can the loss be simplified for one-hot targets?\n1. The transformer architecture was not well justified, either by arguments or by empirical evidence. It's not clear why it's necessary. The number of categories is fixed, so couldn't we just predict logits for the categories using a conv-net? The appendix includes some variations of decoders. What about the effect of completely removing the decoders and simply comparing the learnt density tokens to the patch features?\n1. The first self-attention within the decoder seems like it could be removed completely? It always takes the same set of tokens as input. Why not just learn the output tokens directly?\n1. Several other ablative experiments are missing. What is the effect of using ECR loss without the mask (or varying the threshold $\\xi$)? What is the effect of setting $\\omega = 0.5$ instead of using the max-norm?\n1. To compare the supervised loss alone to its baselines (Table 3), it seems like it would make more sense to consider the fully-supervised setting, rather than the 5% supervised setting. The fully-supervised comparison in Table 5 (appendix) does not include the CE baseline.\n1. In the comparison of supervised losses (Table 3), Bayesian loss (BL) and DM loss are worse than cross entropy (CE). What could explain this?\n\n**Minor issues**\n\n1. It should be more clear in the main text that \"semi-supervised\" means that the training set is a union of fully-supervised and unsupervised sets.\n1. The description of \"query initialisation\" seemed to refer to the fixed semantic meaning of the tokens, rather than the initial values of the query vectors, which I assume are randomly initialised and trained? This is unclear. Or if I have misunderstood, then the initialisation procedure was unclear.\n1. I felt that absolute error (MAE) alone would be sufficient in the main text, and would make the tables easier to interpret. The squared error (MSE) makes the tables more cluttered and doesn't add much.\n1. It wasn't clear how the 2D Gaussian smoothing was applied. Does this replace the ground-truth point annotations with a 2D Gaussian? How is the sigma chosen?\n1. Why not also use the PDM loss to encourage the two discretizations to have similar outputs? (accounting for the shift)\n1. The fact that the multi-head attention modules have multiple layers (4 layers) was not mentioned until the experimental details.\n\n**Nitpicks**\n\n* \"rationale for\" or \"motivation of\" is more appropriate than \"rationality of\"\n* Several uses of \"forward\" and \"forwards\" as a noun. It's unclear what this means. Model evaluation?\n* Some typos and grammar errors.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper considers the problem of predicting density maps for counting given semi-supervised datasets.\nIt proposes a novel transformer-based architecture, a distribution-matching loss function for quantized density prediction, the use of two overlapping discretisations and a consistency loss for the two discretisations that is only applied for confident predictions in unsupervised images.\nThe method achieves strong results in both fully and partially supervised settings, achieving state-of-the-art results.\nSome aspects of the design are demonstrated to be important through ablative experiments.", "strength_and_weaknesses": "**Strengths**\n\n1. The distribution matching loss is nice.\n1. It's neat to use the uncertainty of the predicted discrete distribution to obtain the mask. I like that the non-confident examples are ignored by the ECR loss.\n1. The supervised loss gives a significant improvement over the baseline methods in the fully-supervised setting (Table 5, appendix).\n1. To me, it was surprising that the ECR loss, despite being simple, gave a large improvement in the semi-supervised setting (Table 2).\n1. The attention map visualizations look great (appendix).\n\n**Weaknesses**\n\n1. The authors argue that predicting a distribution is more credible and less noisy than predicting a single value. However, the targets are always one-hot and the distribution is simply used to parametrize the prediction via its expectation, so it seems like simply an alternative parametrization and loss for predicting a single value. Another plausible explanation is that this parametrization and loss simply improve learnability and/or generalization for training a deep network.\n1. The loss in eq. 4 seems more general than the loss which is actually used, since $\\mathbf{y}$ is always one-hot in practice. Can the loss be simplified for one-hot targets?\n1. The transformer architecture was not well justified, either by arguments or by empirical evidence. It's not clear why it's necessary. The number of categories is fixed, so couldn't we just predict logits for the categories using a conv-net? The appendix includes some variations of decoders. What about the effect of completely removing the decoders and simply comparing the learnt density tokens to the patch features?\n1. The first self-attention within the decoder seems like it could be removed completely? It always takes the same set of tokens as input. Why not just learn the output tokens directly?\n1. Several other ablative experiments are missing. What is the effect of using ECR loss without the mask (or varying the threshold $\\xi$)? What is the effect of setting $\\omega = 0.5$ instead of using the max-norm?\n1. To compare the supervised loss alone to its baselines (Table 3), it seems like it would make more sense to consider the fully-supervised setting, rather than the 5% supervised setting. The fully-supervised comparison in Table 5 (appendix) does not include the CE baseline.\n1. In the comparison of supervised losses (Table 3), Bayesian loss (BL) and DM loss are worse than cross entropy (CE). What could explain this?\n\n**Minor issues**\n\n1. It should be more clear in the main text that \"semi-supervised\" means that the training set is a union of fully-supervised and unsupervised sets.\n1. The description of \"query initialisation\" seemed to refer to the fixed semantic meaning of the tokens, rather than the initial values of the query vectors, which I assume are randomly initialised and trained? This is unclear. Or if I have misunderstood, then the initialisation procedure was unclear.\n1. I felt that absolute error (MAE) alone would be sufficient in the main text, and would make the tables easier to interpret. The squared error (MSE) makes the tables more cluttered and doesn't add much.\n1. It wasn't clear how the 2D Gaussian smoothing was applied. Does this replace the ground-truth point annotations with a 2D Gaussian? How is the sigma chosen?\n1. Why not also use the PDM loss to encourage the two discretizations to have similar outputs? (accounting for the shift)\n1. The fact that the multi-head attention modules have multiple layers (4 layers) was not mentioned until the experimental details.\n\n**Nitpicks**\n\n* \"rationale for\" or \"motivation of\" is more appropriate than \"rationality of\"\n* Several uses of \"forward\" and \"forwards\" as a noun. It's unclear what this means. Model evaluation?\n* Some typos and grammar errors.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is quite clear, although some of the choices needs to be better justified, particularly the transformer architecture.\n\nQuality: The method is well-designed and the empirical evaluation seems rigorous.\n\nNovelty: The method is sufficiently novel.\n\nReproducibility: I believe there are enough details to reproduce the experiments. Code will be released.", "summary_of_the_review": "The paper describes an effective loss and parametrization for predicting continuous densities, as well as a simple consistency loss for unsupervised images. The transformer architecture is not demonstrated to be necessary and I wonder whether a simple conv-net would work as well using these loss functions. The ablative experiments should also be expanded, and it would be better to evaluate the supervised loss in the fully-supervised setting. I'm leaning towards accept. I may increase my rating if these issues are addressed or decrease it if they are not.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["Yes, Discrimination / bias / fairness concerns", "Yes, Privacy, security and safety"], "details_of_ethics_concerns": "The paper uses datasets containing images of a large number of people, mostly sourced from the internet. (Note that the paper does not introduce any new datasets, it only uses existing datasets.) It seems likely that these people did not consent to their image being used for this purpose. It's also unclear whether the datasets contain a diverse sample of people. Besides these concerns, I don't foresee any harmful impacts of the work itself, since it considers person counting rather than recognition or classification. I'm not an expert in the problem, so I'm not highly familiar with the datasets.", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666882284884}, {"id": "PBGj9VmkPs", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper354/Reviewer_BsbD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a method of semi-supervised crowd counting which only leverages a subset of labeled data and learn from the unlabeled data. This method relies on pixel-wise distribution matching and leverage optimal transport in the optimzation process.", "review_text": "This paper proposes a method to leverage unlabeled image by pixel-wise distribution matching, as I mentioned, I think the methodology is not sound especially the formulation of optimal transport. Therefore, I think this paper is not qualified for ICLR.", "strengths": "Strength:\n Based on Table 1, it seems that this approach achieve good performance in certain benchmarks.\n\nWeakness:\n1. The idea of leveraging pixel-wise uncertainty/distribution for crowd counting is already proposed in previous work[1], which is not discussed at all.\n\n2. The intervals mentioned in Eq.2 is highly depends on Gaussian kernel value in density generation step and the pixel-wise value of crowd density is generally very small, which makes the chosen of interval values less reliable in different settings.\n\n3. Eq.4 is not Wasserstein distance at all, the real Wsserstein distance is an optimization process that with respect to specific constraint(transport map), thus requires iteration-based solver to get the result, while Eq.4 is a normal L2 norm. Actually, it is higly unefficient for pixel-wise Wasserstein distance as it takes extremely long time for a single forward pass.\n\n4. How to use unlabeled images is not very clear to me. \n\n\n[1]Liu, Weizhe, Nikita Durasov, and Pascal Fua. \"Leveraging Self-Supervision for Cross-Domain Crowd Counting.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n\n------------------------------------------------------------------------------\nPost-Rebuttal Comments:\n\nI would like to thank the austhors provided such detailed rebuttal information, here is my comments after reading all of them:\n\nThe Wasserstein distance does have closed-form solution for 1-d formulation, but I still do not think  Eq.(4) with one-hot vector formulation is the correct formulation of 1-d Wasserstein distance as you mentioned in the reference paper. Besides, I even doubt if this formulation has any advantage over simply soft-max loss. As the code is not available, I'm not able to judge it.\n\nI raised my recomendation considering the rebuttal did address some of my concerns.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a method of semi-supervised crowd counting which only leverages a subset of labeled data and learn from the unlabeled data. This method relies on pixel-wise distribution matching and leverage optimal transport in the optimzation process.", "strength_and_weaknesses": "Strength:\n Based on Table 1, it seems that this approach achieve good performance in certain benchmarks.\n\nWeakness:\n1. The idea of leveraging pixel-wise uncertainty/distribution for crowd counting is already proposed in previous work[1], which is not discussed at all.\n\n2. The intervals mentioned in Eq.2 is highly depends on Gaussian kernel value in density generation step and the pixel-wise value of crowd density is generally very small, which makes the chosen of interval values less reliable in different settings.\n\n3. Eq.4 is not Wasserstein distance at all, the real Wsserstein distance is an optimization process that with respect to specific constraint(transport map), thus requires iteration-based solver to get the result, while Eq.4 is a normal L2 norm. Actually, it is higly unefficient for pixel-wise Wasserstein distance as it takes extremely long time for a single forward pass.\n\n4. How to use unlabeled images is not very clear to me. \n\n\n[1]Liu, Weizhe, Nikita Durasov, and Pascal Fua. \"Leveraging Self-Supervision for Cross-Domain Crowd Counting.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n\n------------------------------------------------------------------------------\nPost-Rebuttal Comments:\n\nI would like to thank the austhors provided such detailed rebuttal information, here is my comments after reading all of them:\n\nThe Wasserstein distance does have closed-form solution for 1-d formulation, but I still do not think  Eq.(4) with one-hot vector formulation is the correct formulation of 1-d Wasserstein distance as you mentioned in the reference paper. Besides, I even doubt if this formulation has any advantage over simply soft-max loss. As the code is not available, I'm not able to judge it.\n\nI raised my recomendation considering the rebuttal did address some of my concerns.", "clarity,_quality,_novelty_and_reproducibility": "No code available.", "summary_of_the_review": "This paper proposes a method to leverage unlabeled image by pixel-wise distribution matching, as I mentioned, I think the methodology is not sound especially the formulation of optimal transport. Therefore, I think this paper is not qualified for ICLR.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666772760228}, {"id": "tBxx7RnWCI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper354/Reviewer_ccra"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this manuscript the authors propose a semi-supervised crowd counting model, which include a pixel-wise distribution matching loss to measure the difference in pixel-wise density distribution between predicted and the ground-truth, density tokens to augment Transformer decoder, and an interleaving consistency self-supervised learning mechanism to efficiently learn from unmarked data.", "review_text": "Although there are still some flaws in this manuscript, the ideas are novel and the arguments are basically sufficient. I think it is a high quality manuscript.", "strengths": "Strengths:\n(1) Transforming the density map regression task into a density-level classification task in semi-supervised crowd counting is reasonably efficient and alleviates the noise problem in the traditional semi-supervised learning of directly predicting density maps.\n(2) The proposed method achieves good results on several datasets in semi-supervised crowd counting.\n\n\nWeakness:\n(1)In line 127 of page 4, [b1, b2] duplication occurred, which should be a clerical error.\n(2)In line 215 of page 5, it is mentioned that the initialization method of density tokens in this manuscript is compared with randomly initialized queries. Ablation study about the initialization method should add to prove this assertion.\n(3)In Table 1, “The results of other methods under the 40% labeled setting are referred to (Meng et al., 2021)”. However, according to (Meng et al., 2021), the quantitative results they show for the semi-supervised learning method use 50% of the labeled data. Please check to verify it.\n(4)In Table 1, the methods of comparison were all proposed before 2022. If there are new articles on semi-supervised crowd counting in 2022, please cite them and compare.\n(5)A visual presentation of the predicted density map compared to SOTA's semi-supervised population counting method seems to be missing in this manuscript, please add if space allows. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "In this manuscript the authors propose a semi-supervised crowd counting model, which include a pixel-wise distribution matching loss to measure the difference in pixel-wise density distribution between predicted and the ground-truth, density tokens to augment Transformer decoder, and an interleaving consistency self-supervised learning mechanism to efficiently learn from unmarked data.", "strength_and_weaknesses": "Strengths:\n(1) Transforming the density map regression task into a density-level classification task in semi-supervised crowd counting is reasonably efficient and alleviates the noise problem in the traditional semi-supervised learning of directly predicting density maps.\n(2) The proposed method achieves good results on several datasets in semi-supervised crowd counting.\n\n\nWeakness:\n(1)In line 127 of page 4, [b1, b2] duplication occurred, which should be a clerical error.\n(2)In line 215 of page 5, it is mentioned that the initialization method of density tokens in this manuscript is compared with randomly initialized queries. Ablation study about the initialization method should add to prove this assertion.\n(3)In Table 1, “The results of other methods under the 40% labeled setting are referred to (Meng et al., 2021)”. However, according to (Meng et al., 2021), the quantitative results they show for the semi-supervised learning method use 50% of the labeled data. Please check to verify it.\n(4)In Table 1, the methods of comparison were all proposed before 2022. If there are new articles on semi-supervised crowd counting in 2022, please cite them and compare.\n(5)A visual presentation of the predicted density map compared to SOTA's semi-supervised population counting method seems to be missing in this manuscript, please add if space allows. ", "clarity,_quality,_novelty_and_reproducibility": "The overall expression of the manuscript is clear and of high quality, and with beautiful figures, but some mathematical expressions are not described clearly enough. The idea of introducing the idea of discrete density classification into semi-supervised population counting is also relatively novel.", "summary_of_the_review": "Although there are still some flaws in this manuscript, the ideas are novel and the arguments are basically sufficient. I think it is a high quality manuscript.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666579706586}, {"id": "OO0uz4E_kOl", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper354/Reviewer_FHRP"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper is mostly clearly written. The transformer decoder and the pixel-wise distribution matching loss are adopted in the paper. The results are far better than the-state-of-art methods on four datasets which is quite good.The codes might be released for validation.", "review_text": "It's recommended as an acception if the problems mentioned above are clearly explained.", "strengths": "Strength: The performance is dramaticly improved.\n\nWeakness:\n1) The N stands for the pixel number in line 189 while it means N regions in 200. It is contradictionary and with no further explanation. v1 and v2 in equation (7) is not explained.\n2) How does the interleaving dual branch structure work in semi-supervied learning. \n3) What's the relation between (a) and (b) in Figure1 . What's the difference in Branch 1 and 2 in Figure 1.\n4) How are the density levels generated from the annotations?\n5)cross attention lacks referred paper before equation (5).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper is mostly clearly written. The transformer decoder and the pixel-wise distribution matching loss are adopted in the paper. The results are far better than the-state-of-art methods on four datasets which is quite good.The codes might be released for validation.", "strength_and_weaknesses": "Strength: The performance is dramaticly improved.\n\nWeakness:\n1) The N stands for the pixel number in line 189 while it means N regions in 200. It is contradictionary and with no further explanation. v1 and v2 in equation (7) is not explained.\n2) How does the interleaving dual branch structure work in semi-supervied learning. \n3) What's the relation between (a) and (b) in Figure1 . What's the difference in Branch 1 and 2 in Figure 1.\n4) How are the density levels generated from the annotations?\n5)cross attention lacks referred paper before equation (5).", "clarity,_quality,_novelty_and_reproducibility": "Most of the work is innovations based on the previous work.Most of the innovations are reasonable.", "summary_of_the_review": "It's recommended as an acception if the problems mentioned above are clearly explained.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666446504140}], "openreview_url": "https://openreview.net/forum?id=t-hNmA0cVSW", "arxiv_id": "2402.15297", "paper_pdf": "papers/t-hNmA0cVSW.pdf", "paper_pdf_sha256": "a9f514d0bce69c0ecbdd86e9de62c489df551d4d369dc5cd0f5b309098caf099", "paper_pdf_bytes": 1719675, "paper_pdf_source": "openreview", "code_url": "https://github.com/LoraLinH/Semi-supervised-Counting-via-Pixel-by-pixel-Density-Distribution-Modelling", "code_repository": "LoraLinH/Semi-supervised-Counting-via-Pixel-by-pixel-Density-Distribution-Modelling", "code_commit": "66034010cc7134fa5265b282e606cdcb7beb3399", "code_archive": "repos/t-hNmA0cVSW.zip", "code_archive_sha256": "2ab266684d36250a645091676a1e02f61111bfd66817cb416102b131df813996", "code_archive_bytes": 19965, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 21, "github_languages": {"Python": 56116}, "github_archived": false, "github_pushed_at": "2025-04-02T11:55:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/semi-supervised-counting-via-pixel-by-pixel"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gxk4-rVATDA", "year": 2022, "status": "rejected", "title": "Bit-wise Training of Neural Network Weights", "authors": ["Cristian Ivan"], "authorids": ["~Cristian_Ivan1"], "authors_source": "OpenReview API", "abstract": "We propose an algorithm where the individual bits representing the weights of a neural network are learned. This method allows training weights with integer values on arbitrary bit-depths and naturally uncovers sparse networks, without additional constraints or regularization techniques. We show better results than the standard training technique with fully connected networks and similar performance as compared to standard training for residual networks. By training bits in a selective manner we found that the biggest contribution to achieving high accuracy is given by the first three most significant bits, while the rest provide an intrinsic regularization. As a consequence we show that more than 90% of a network can be used to store arbitrary codes without affecting the its accuracy. These codes can be random noise, binary files or even the weights of previously trained networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "HmWwXJPxJXg", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2798/Reviewer_LhyK"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a method to train individual bits of weights of a neural network.  The central motivation is the question of why doesn't gradient descent \"discover\" inherent sparsity by setting certain weights to zero? The authors suggest that since there are so many possible states of a 32b integer number, the probability of landing on all zeros is vanishingly small. Using the bit-wise training technique, they are able to demonstrate that good performance can be achieved with few bits of weight representation. The authors also do a series of experiments where only certain bit positions are changeable to demonstrate which positions are most relevant to getting good classification performance.", "review_text": "The authors provide a nice background of pruning literature and sparse representations in neural networks.  I believe they were trying to provide a background for the idea that sparseness is something inherent and should/could be uncovered by gradient descent.  However, the background citations are for a slightly different, but related topic; pruning techniques are not exactly what this paper is about, but it does share some background with spares representations.\n\nIn motivation, for line\" One possible explanation is that, at least for classification tasks, the usual cross–entropy loss without ad- ditional regularization techniques are not well suited for this\" there needs to be citation.\n\nThe description of STE in section 3 needs more description.\n\nThe experimental results look very preliminary.  Using only LeNet and CIFAR10 might be a good way to triage a technique, but it is very different to draw any conclusions based on these.  They are too small and the results between them barely suggest any trend.  The authors essentially show a known result that training only sign bits (Ivan and Florian 2020) yields good results.  What additional science have the authors uncovered? \n\nThe last section of encoding messages in weights seems quite unrelated to the rest of the paper.  It is interesting, but the results seem random and they don't give us any new insight into the learning process of neural networks.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a method to train individual bits of weights of a neural network.  The central motivation is the question of why doesn't gradient descent \"discover\" inherent sparsity by setting certain weights to zero? The authors suggest that since there are so many possible states of a 32b integer number, the probability of landing on all zeros is vanishingly small. Using the bit-wise training technique, they are able to demonstrate that good performance can be achieved with few bits of weight representation. The authors also do a series of experiments where only certain bit positions are changeable to demonstrate which positions are most relevant to getting good classification performance.", "main_review": "The authors provide a nice background of pruning literature and sparse representations in neural networks.  I believe they were trying to provide a background for the idea that sparseness is something inherent and should/could be uncovered by gradient descent.  However, the background citations are for a slightly different, but related topic; pruning techniques are not exactly what this paper is about, but it does share some background with spares representations.\n\nIn motivation, for line\" One possible explanation is that, at least for classification tasks, the usual cross–entropy loss without ad- ditional regularization techniques are not well suited for this\" there needs to be citation.\n\nThe description of STE in section 3 needs more description.\n\nThe experimental results look very preliminary.  Using only LeNet and CIFAR10 might be a good way to triage a technique, but it is very different to draw any conclusions based on these.  They are too small and the results between them barely suggest any trend.  The authors essentially show a known result that training only sign bits (Ivan and Florian 2020) yields good results.  What additional science have the authors uncovered? \n\nThe last section of encoding messages in weights seems quite unrelated to the rest of the paper.  It is interesting, but the results seem random and they don't give us any new insight into the learning process of neural networks.  ", "summary_of_the_review": "This paper has some interesting concepts but falls short on uncovering novel insight.  They are able to recapitulate other papers' results, but even this seems a bit tenuous. The experimental support for any conclusions is too weak in this paper to draw any real conclusions.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636442821185}, {"id": "geAswnX0XbP", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2798/Reviewer_MjPw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper proposes arithmetic decomposition and training on individual bits in order to achieve low-precision quantization.", "review_text": "Unfortunately, there are serious issues with this submission:\n\nThe introduction and motivation seem to be written for a paper that is on pruning not quantization, and they are out of context compared to the title, abstract, and rest of the paper. This is most likely a LaTeX error. The motivation of the paper and comparison to prior art being excluded, it is very hard to assess the quality of the work. I did however try to extrapolate what the authors intended to write, and below is my review for Section 3 and onwards.\n\nIn Section 3, what is the expression for \\alpha in terms of layer dimensions and number of bits such that the He initialization standard deviation condition is satisfied? This seems like an interesting result that can be added inline in the paper, rather than just implicitly mentioning that an expression for \\alpha was derived and used.\n\nWhat is a He distribution? In (He, 2015), variance engineering is done and the contributions are conditions on the variance of the initialized weights. The distributions themselves are either uniform or normal. Can the author clarify what they mean by He distribution, I believe they mean He conditions on variance.\n\nThe experiments are performed on very trivial networks deployed on the MNIST and CIFAR-10 dataset. Can the authors evaluate their work on more contemporary networks such as ResNet on ImageNet and similar tasks?\n\nMessage encoding in the neural network's weight using steganography in training is interesting. However, why does it matter? And can these results be generalized on larger networks.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes arithmetic decomposition and training on individual bits in order to achieve low-precision quantization.", "main_review": "Unfortunately, there are serious issues with this submission:\n\nThe introduction and motivation seem to be written for a paper that is on pruning not quantization, and they are out of context compared to the title, abstract, and rest of the paper. This is most likely a LaTeX error. The motivation of the paper and comparison to prior art being excluded, it is very hard to assess the quality of the work. I did however try to extrapolate what the authors intended to write, and below is my review for Section 3 and onwards.\n\nIn Section 3, what is the expression for \\alpha in terms of layer dimensions and number of bits such that the He initialization standard deviation condition is satisfied? This seems like an interesting result that can be added inline in the paper, rather than just implicitly mentioning that an expression for \\alpha was derived and used.\n\nWhat is a He distribution? In (He, 2015), variance engineering is done and the contributions are conditions on the variance of the initialized weights. The distributions themselves are either uniform or normal. Can the author clarify what they mean by He distribution, I believe they mean He conditions on variance.\n\nThe experiments are performed on very trivial networks deployed on the MNIST and CIFAR-10 dataset. Can the authors evaluate their work on more contemporary networks such as ResNet on ImageNet and similar tasks?\n\nMessage encoding in the neural network's weight using steganography in training is interesting. However, why does it matter? And can these results be generalized on larger networks.", "summary_of_the_review": "Unfortunately, it seems this paper was rushed into submission. The idea sounds interesting, however, the proposed manuscript has several serious issues as raised in my main review. I therefore recommend rejection of this paper.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "Not applicable", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635809501052}, {"id": "rUthC9W3k_T", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2798/Reviewer_gsV7"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to directly train the bit values of each parameter in a neural network, instead of directly optimizing the floating point value of each weight. By varying the number of bits that are allowed to be optimized, the authors show that with less bits the network will become automatically sparser. This method has many interesting applications, including fixing some bits to be a message and only training the rest.", "review_text": "Strengths:\n- This paper experiments with an interesting an intuitive idea.\n- There are many interesting applications, e.g., more efficient networks and embedding hidden messages in network weights.\n- Most figures include error bars.\n- Figure 3 confirms a main hypothesis (fewer bits encourages sparser networks).\n\nWeaknesses:\n1) The paper could benefit from medium scale experiments, e.g., ImageNet. The method matches standard training on ImageNet but faces accuracy degradation on CIFAR. A concern is that this accuracy degradation would be even more substantial for problems such as ImageNet.\n2) The paper would benefit substantially from a related work section. Since the paper is not 9 pages, there is definitely room for this. I am not an expert on quantization (perhaps another reviewer is) but I know that it is a very active research area. How does this papers method compare to standard methods in quantization? If the author's hypothesis is correct, networks trained with various quantization techniques should be sparse and it would be very interesting to verify this.\n3) There is no discussion of how much extra compute / FLOPs is incurred by this method during training, which may be a drawback. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to directly train the bit values of each parameter in a neural network, instead of directly optimizing the floating point value of each weight. By varying the number of bits that are allowed to be optimized, the authors show that with less bits the network will become automatically sparser. This method has many interesting applications, including fixing some bits to be a message and only training the rest.", "main_review": "Strengths:\n- This paper experiments with an interesting an intuitive idea.\n- There are many interesting applications, e.g., more efficient networks and embedding hidden messages in network weights.\n- Most figures include error bars.\n- Figure 3 confirms a main hypothesis (fewer bits encourages sparser networks).\n\nWeaknesses:\n1) The paper could benefit from medium scale experiments, e.g., ImageNet. The method matches standard training on ImageNet but faces accuracy degradation on CIFAR. A concern is that this accuracy degradation would be even more substantial for problems such as ImageNet.\n2) The paper would benefit substantially from a related work section. Since the paper is not 9 pages, there is definitely room for this. I am not an expert on quantization (perhaps another reviewer is) but I know that it is a very active research area. How does this papers method compare to standard methods in quantization? If the author's hypothesis is correct, networks trained with various quantization techniques should be sparse and it would be very interesting to verify this.\n3) There is no discussion of how much extra compute / FLOPs is incurred by this method during training, which may be a drawback. ", "summary_of_the_review": "The paper is very interesting but would benefit substantially from 1) medium scale experiments (e.g., ResNet on ImageNet) and 2) a more thorough discussion and comparison with related work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635050423252}, {"id": "jjp9gDRUBh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2798/Reviewer_xT9U"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a neural network training technique such that individual weight bits can be optimized separately. In detail, each weight is represented as a weighted sum of its bits weighted by powers of 2. In training, updating each bit $b$ is achieved by updating a floating-point number $x$ ($b=1, x > 0; b=0, x \\le 0$).\n\nBy conducing extensive experiments, the authors find:\n1) Network with shorter bit-width show more weight sparsity than that with longer bit-width. (Sec 4.)\n2) With selective bit training, only a few most significant bits contribute to the final high model accuracy. The other less important bits serve as regularization. (Sec 5.)\n3) The less significant bits can be used to encode other information. (Sec 6.)", "review_text": "Here are the strengths and weaknesses of the paper.\n\nStrengths:\n1. The paper is interesting and, in some sense, novel in that it analyze the network weights/bits and regularization in an interesting perspective. By decomposing a weight into separate bits, the function of each bit can be more easily observed and analyzed.\n2. The authors conducted extensive experiments to demonstrate different phenomena from the bit-wise training idea.\n\nWeaknesses:\n1. In the second paragraph, Page 5, I guess it should be \"as shown in Figure 4\" instead of \"as shown in Figure 5\".\n2. In second sentence in Sec 5.1 is quoted here: \"In contrast, standard training does not reveal which weights or bits contribute most to the network’s performance\". I didn't find the following paragraphs support this claim. In contrary, the paragraphs show the most significant bits contribute more to the performance than the others. In network quantization, e.g., 32-bit to 8-bit, one can also find that the least significant bits can be dropped off without hurting performance much.\n3. In second paragraph of Page 4, \"..., in general, neural networks can be trained well only by changing the sign of the weights and never updating their magnitudes\". I don't think 4-percentage drop is small considering Cifar10 is a small dataset and ResNet-18 is relatively a large model. When using a smaller network, the performance drop might be big.\n4. Throughout the paper (e.g., Sec 5.1 and Sec 6), one main claim of the paper is \"a few of the most significant bits contribute to achieving a high accuracy, while the others provide regularization\". This is true but not a significant observation from the perspective of network quantization. When network is big and contains redundancy, the network can be quantized to lower-bit one (e.g., 3 bit) with comparable performance (e.g., [1]). Compared with the full-precision model, the quantized model is well regularized.\n\nQuestions:\n1. As the paper hypothesizes, the 32-bit model does not have much zero weights because the probability of an exactly zero-valued weight is very small (1e-31). If this is true, one should expect the trend between 2 and 14 bit-width in Figure 2 is exponential instead of flat. I didn't find an explanation in the relevant section.\n\n[1] Zhang, Dongqing, et al. \"Lq-nets: Learned quantization for highly accurate and compact deep neural networks.\" Proceedings of the European conference on computer vision (ECCV). 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a neural network training technique such that individual weight bits can be optimized separately. In detail, each weight is represented as a weighted sum of its bits weighted by powers of 2. In training, updating each bit $b$ is achieved by updating a floating-point number $x$ ($b=1, x > 0; b=0, x \\le 0$).\n\nBy conducing extensive experiments, the authors find:\n1) Network with shorter bit-width show more weight sparsity than that with longer bit-width. (Sec 4.)\n2) With selective bit training, only a few most significant bits contribute to the final high model accuracy. The other less important bits serve as regularization. (Sec 5.)\n3) The less significant bits can be used to encode other information. (Sec 6.)", "main_review": "Here are the strengths and weaknesses of the paper.\n\nStrengths:\n1. The paper is interesting and, in some sense, novel in that it analyze the network weights/bits and regularization in an interesting perspective. By decomposing a weight into separate bits, the function of each bit can be more easily observed and analyzed.\n2. The authors conducted extensive experiments to demonstrate different phenomena from the bit-wise training idea.\n\nWeaknesses:\n1. In the second paragraph, Page 5, I guess it should be \"as shown in Figure 4\" instead of \"as shown in Figure 5\".\n2. In second sentence in Sec 5.1 is quoted here: \"In contrast, standard training does not reveal which weights or bits contribute most to the network’s performance\". I didn't find the following paragraphs support this claim. In contrary, the paragraphs show the most significant bits contribute more to the performance than the others. In network quantization, e.g., 32-bit to 8-bit, one can also find that the least significant bits can be dropped off without hurting performance much.\n3. In second paragraph of Page 4, \"..., in general, neural networks can be trained well only by changing the sign of the weights and never updating their magnitudes\". I don't think 4-percentage drop is small considering Cifar10 is a small dataset and ResNet-18 is relatively a large model. When using a smaller network, the performance drop might be big.\n4. Throughout the paper (e.g., Sec 5.1 and Sec 6), one main claim of the paper is \"a few of the most significant bits contribute to achieving a high accuracy, while the others provide regularization\". This is true but not a significant observation from the perspective of network quantization. When network is big and contains redundancy, the network can be quantized to lower-bit one (e.g., 3 bit) with comparable performance (e.g., [1]). Compared with the full-precision model, the quantized model is well regularized.\n\nQuestions:\n1. As the paper hypothesizes, the 32-bit model does not have much zero weights because the probability of an exactly zero-valued weight is very small (1e-31). If this is true, one should expect the trend between 2 and 14 bit-width in Figure 2 is exponential instead of flat. I didn't find an explanation in the relevant section.\n\n[1] Zhang, Dongqing, et al. \"Lq-nets: Learned quantization for highly accurate and compact deep neural networks.\" Proceedings of the European conference on computer vision (ECCV). 2018.", "summary_of_the_review": "The perspective of the paper is interesting, but some claims might need correction. Moreover, the observation is straightforward, especially from the view of network quantization.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635044064731}], "openreview_url": "https://openreview.net/forum?id=gxk4-rVATDA", "arxiv_id": "2202.09571", "paper_pdf": "papers/gxk4-rVATDA.pdf", "paper_pdf_sha256": "525dfaab09c5bd0aaf95bff67a39eb0a9c58b57103475c37b2bf7086db9e04ed", "paper_pdf_bytes": 2383381, "paper_pdf_source": "openreview", "code_url": "https://github.com/iclr2022-2798/bit-wise-training", "code_repository": "iclr2022-2798/bit-wise-training", "code_commit": "39fdc97caa080bd3b09ae7abcef75f79ca6dac5d", "code_archive": "repos/gxk4-rVATDA.zip", "code_archive_sha256": "5131a7aabdd85cade759642bde287a67de69543b370984bf71e1e753f8197425", "code_archive_bytes": 12573, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 20, "github_languages": {"Python": 47611}, "github_archived": false, "github_pushed_at": "2021-10-11T20:45:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bit-wise-training-of-neural-network-weights-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ztMLindFLWR", "year": 2021, "status": "rejected", "title": "Breaking the Expressive Bottlenecks of Graph Neural Networks", "authors": ["Mingqi Yang", "Yanming Shen", "Heng Qi", "Baocai Yin"], "authorids": ["~Mingqi_Yang1", "shen@dlut.edu.cn", "hengqi@dlut.edu.cn", "~Baocai_Yin1"], "authors_source": "OpenReview API", "abstract": "Recently, the Weisfeiler-Lehman (WL) graph isomorphism test was used to measure the expressiveness of graph neural networks (GNNs), showing that the neighborhood aggregation GNNs were at most as powerful as 1-WL test in distinguishing graph structures. There were also improvements proposed in analogy to $k$-WL test ($k>1$). However, the aggregators in these GNNs are far from injective as required by the WL test, and suffer from weak distinguishing strength, making it become expressive bottlenecks. In this paper, we improve the expressiveness by exploring powerful aggregators. We reformulate aggregation with the corresponding aggregation coefficient matrix, and then systematically analyze the requirements of the aggregation coefficient matrix for building more powerful aggregators and even injective aggregators. It can also be viewed as the strategy for preserving the rank of hidden features, and implies that basic aggregators correspond to a special case of low-rank transformations. We also show the necessity of applying nonlinear units ahead of aggregation, which is different from most aggregation-based GNNs. Based on our theoretical analysis, we develop two GNN layers, ExpandingConv and CombConv. Experimental results show that our models significantly boost performance, especially for large and densely connected graphs.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "q6gXaHXeIy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1702/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The work presents a framework to categorize GNN aggregators based on their distinguishing strength. It connects the distinguishing strength to the rank of the aggregation coefficient matrices. Based on the findings, the authors present two GNN layers, ExpandingConv and CombConv, and evaluate them on some graph data sets.\n\nStrengths:\n- The paper is mostly well written.\n- The formalization of distinguishing strength for aggregators is interesting and novel.\n- The relation to the rank of coefficient matrices is a valuable insight and gives a new perspective to the effectivity of multi-head GAT.\n\nWeaknesses/Questions:\n- The paper lacks clarity in presentation / technical soundness, up to a point where I am not able to understand some details:\n\t- It is not clear what it means if f_aggr1 <= f_aggr2 \"does not exist\" (definition of incomparability)\n\t- In Lemma 1(ii), the formalization with \"or\" is confusing. It should probably be \"f_aggr1 ⊗ f_aggr2 >= f_aggr1 or f_aggr1 ⊗ f_aggr2 >= f_aggr2\".\n\t- I suggest to call the feature matrix for all nodes H, not h with specific indices, to avoid confusion. Currently, the index, or the lack of it, makes it either a matrix or a vector. Sometimes, h seems to be defined as the matrix containing features from all neighbors of u (as in the GCN definition). At other places, it seems to be just the vector of node u (Equation 3).\n\t- Page 5, rank(r) < min(rank(M), rank(h)) should probably be <=, otherwise rank(r) = rank(h) can't be achieved.\n\t- How can P be found? In order to construct it, a canonical order has to be defined and pi needs to be extracted from h. In fact, the permutation invariance does not seem to be relevant for the proposed approaches, since they both process all neighbors individually, have a sum in the end and do not impose an order by e.g. concatenation. Can the authors clarify, why this part is needed?\n\t- The set of results RES() is not clearly defined. The output of the aggregation is a matrix, not a set.\n\t- In general, Proposition 2 with the set formulations seems to be out of context and needs at least to be discussed.\n\n- The experiments are not sufficient and do not show significant improvements:\n\t- Some strong competitors are left out in comparions: E.g. GAT and multi-head GAT.\n\t- The approach is only validated on a small set of data sets. Since the authors already use OGB, I wonder why only three data sets have been chosen. To me this looks like those three were cherry-picked.\n\t- Even on the small number of data sets, the results are mediocre.\t\n\n- It is unclear why CombConv is proposed. It needs more discussion. In the case of CombConv, the aggregation coefficients have rank 1, similar to a lot of other existing GNN operators, isn't that right? This would mean that CombConv disgards the main argument of the work (having rank > 1) in trade for efficiency.\n\n\nAll in all, I recommend to reject the paper in its current state. While there seems to be some novel insight in this work, which might be of interest to the community, I think the paper needs to be improved in (1) clarity of presentation and (2) experiments. Issues with (1) can maybe be fixed within the rebuttal period. I am not sure about (2). It probably depends on the actual performance of the proposed operator.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some interesting insights, clarity can be improved, weak experiments", "review": "The work presents a framework to categorize GNN aggregators based on their distinguishing strength. It connects the distinguishing strength to the rank of the aggregation coefficient matrices. Based on the findings, the authors present two GNN layers, ExpandingConv and CombConv, and evaluate them on some graph data sets.\n\nStrengths:\n- The paper is mostly well written.\n- The formalization of distinguishing strength for aggregators is interesting and novel.\n- The relation to the rank of coefficient matrices is a valuable insight and gives a new perspective to the effectivity of multi-head GAT.\n\nWeaknesses/Questions:\n- The paper lacks clarity in presentation / technical soundness, up to a point where I am not able to understand some details:\n\t- It is not clear what it means if f_aggr1 <= f_aggr2 \"does not exist\" (definition of incomparability)\n\t- In Lemma 1(ii), the formalization with \"or\" is confusing. It should probably be \"f_aggr1 ⊗ f_aggr2 >= f_aggr1 or f_aggr1 ⊗ f_aggr2 >= f_aggr2\".\n\t- I suggest to call the feature matrix for all nodes H, not h with specific indices, to avoid confusion. Currently, the index, or the lack of it, makes it either a matrix or a vector. Sometimes, h seems to be defined as the matrix containing features from all neighbors of u (as in the GCN definition). At other places, it seems to be just the vector of node u (Equation 3).\n\t- Page 5, rank(r) < min(rank(M), rank(h)) should probably be <=, otherwise rank(r) = rank(h) can't be achieved.\n\t- How can P be found? In order to construct it, a canonical order has to be defined and pi needs to be extracted from h. In fact, the permutation invariance does not seem to be relevant for the proposed approaches, since they both process all neighbors individually, have a sum in the end and do not impose an order by e.g. concatenation. Can the authors clarify, why this part is needed?\n\t- The set of results RES() is not clearly defined. The output of the aggregation is a matrix, not a set.\n\t- In general, Proposition 2 with the set formulations seems to be out of context and needs at least to be discussed.\n\n- The experiments are not sufficient and do not show significant improvements:\n\t- Some strong competitors are left out in comparions: E.g. GAT and multi-head GAT.\n\t- The approach is only validated on a small set of data sets. Since the authors already use OGB, I wonder why only three data sets have been chosen. To me this looks like those three were cherry-picked.\n\t- Even on the small number of data sets, the results are mediocre.\t\n\n- It is unclear why CombConv is proposed. It needs more discussion. In the case of CombConv, the aggregation coefficients have rank 1, similar to a lot of other existing GNN operators, isn't that right? This would mean that CombConv disgards the main argument of the work (having rank > 1) in trade for efficiency.\n\n\nAll in all, I recommend to reject the paper in its current state. While there seems to be some novel insight in this work, which might be of interest to the community, I think the paper needs to be improved in (1) clarity of presentation and (2) experiments. Issues with (1) can maybe be fixed within the rebuttal period. I am not sure about (2). It probably depends on the actual performance of the proposed operator.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604660453147}, {"id": "wyrzhkUc5Px", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1702/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \n\tThis paper explores the representation power of graph neural networks. Unlike recent work on choosing among simple aggregation functions or combinations thereof, the authors here recognize that these aggregators are the bottleneck in the representation power and generalize simple aggregator functions commonly used in literature to an aggregation coefficient matrix. The paper supports this construction theoretically and also proposes two aggregators that satisfy the rank-preservation requirement for more expressive (distinguishing) GNNs.\n\nStrengths:\n* The theoretical results are strong in proving the bottleneck of aggregators (Lemma 1) and clearly contextualize popular existing methods into this result.\n* Formulating aggregation in terms of a product with coefficients (indexed by a permutation) allows for framing existing methods and to connect the representation power with the rank of the matrix of coefficients.\n\nWeaknesses:\n* The study of the expressiveness of GNNs is a very popular topic right now and not enough context is provided about related work on this topic and other approaches, mainly focusing on GIN and GAT in the development and while a few other GNNs are considered in the experimental results, they are not discussed or explained enough.\n\nRecommendation:\nBy framing the aggregation in terms of coefficients, the paper provides interesting connections between the rank of these coefficient matrices and the distinguishing power of GNNs.  While analysis and explanation of experiments is extremely limited, the theoretical developments are interesting and novel enough to narrowly recommend publication. Meanwhile, I do think the paper should undergo a reorganization to add more details about related work in the study of expressiveness of GNNs and analyze experiments more thoroughly.\n\n\nOther comments and clarification needed:\n* It’s not quite clear what it means for aggregators to be incomparable (top of page 3), ie what does it mean for the relative strength to “not exist”? This point could be clarified with an additional sentence in this “Distinguishing strength” paragraph. \n* The text on page 4 before Proposition 2 states that “… different M corresponds to different local structures. Therefore, the aggregation results of different aggregators must be different. However, it is not satisfied by existing GNNs.” This statement should be explained more and supported. \n* The details of the ExpandingConv layer overwhelm the paper. The details fo equation 4 and related text can be moved to an Appendix and explained in the main text at a higher level. This would free up space to add context of the paper’s contribution and more properly address the experiments. In the end, the ExpandingConv formulation is a refinement on multi-head GAT.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Limited context, but useful theoretical framing of distinguishing power of GNN aggregators", "review": "Summary: \n\tThis paper explores the representation power of graph neural networks. Unlike recent work on choosing among simple aggregation functions or combinations thereof, the authors here recognize that these aggregators are the bottleneck in the representation power and generalize simple aggregator functions commonly used in literature to an aggregation coefficient matrix. The paper supports this construction theoretically and also proposes two aggregators that satisfy the rank-preservation requirement for more expressive (distinguishing) GNNs.\n\nStrengths:\n* The theoretical results are strong in proving the bottleneck of aggregators (Lemma 1) and clearly contextualize popular existing methods into this result.\n* Formulating aggregation in terms of a product with coefficients (indexed by a permutation) allows for framing existing methods and to connect the representation power with the rank of the matrix of coefficients.\n\nWeaknesses:\n* The study of the expressiveness of GNNs is a very popular topic right now and not enough context is provided about related work on this topic and other approaches, mainly focusing on GIN and GAT in the development and while a few other GNNs are considered in the experimental results, they are not discussed or explained enough.\n\nRecommendation:\nBy framing the aggregation in terms of coefficients, the paper provides interesting connections between the rank of these coefficient matrices and the distinguishing power of GNNs.  While analysis and explanation of experiments is extremely limited, the theoretical developments are interesting and novel enough to narrowly recommend publication. Meanwhile, I do think the paper should undergo a reorganization to add more details about related work in the study of expressiveness of GNNs and analyze experiments more thoroughly.\n\n\nOther comments and clarification needed:\n* It’s not quite clear what it means for aggregators to be incomparable (top of page 3), ie what does it mean for the relative strength to “not exist”? This point could be clarified with an additional sentence in this “Distinguishing strength” paragraph. \n* The text on page 4 before Proposition 2 states that “… different M corresponds to different local structures. Therefore, the aggregation results of different aggregators must be different. However, it is not satisfied by existing GNNs.” This statement should be explained more and supported. \n* The details of the ExpandingConv layer overwhelm the paper. The details fo equation 4 and related text can be moved to an Appendix and explained in the main text at a higher level. This would free up space to add context of the paper’s contribution and more properly address the experiments. In the end, the ExpandingConv formulation is a refinement on multi-head GAT.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604072689990}, {"id": "xlV3BjNp4go", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1702/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose two new layers for GNNs. CombConv and ExpandingConv are motivated by the insight that a GNN is only as expressive as the rank of the matrix that represents the coefficients of the aggregation function. To arrive at this statement, the authors formalize all GNNs as being composed of three steps: 1) generation of aggregation coefficients, 2) actual aggregation of the neighbourhood, and 3) feature extraction from the aggregation. Furthermore, it is shown that current approaches have very low distinguishing strength and that CombConv and ExpandingConv, by their construction, yield higher expressive power. \nThe effectiveness of different components of their layers (e.g. Re-SUM, applying a ReLU non-linearity before summing when features are computed form the aggregation) are investigated in an ablation study. Additionally, proposed layers are compared with current approaches on 4 data sets. \n\nThis paper constructs an interesting theoretical analysis of GNNs and finds the bottleneck of these networks to be in the coefficient matrix of the aggregation scheme. While I was not able to check all proofs, it seems like a solid mathematical analysis. What I find somewhat sobering is the experimental section. First, I am surprised that you only compare your method on four data sets and that you miss some reported by your comparison partners (such as IMDB, REDDIT, PROTEINS, etc.). I understand the Graph Kernel data sets are smaller, however, they’ve been used in comparable papers before. Second, while the title and the theoretical analysis promises much higher distinguishing strength, I am surprised that your performance gains are good but not outstanding. I wonder if you could construct a synthetic data set in which you can show how you break the expressive bottleneck in practice. Would it make sense to compute the rank of M for different networks (including yours ((it is not guaranteed it is $s$, right?)) and plot it as a function of predictive performance? \nIn Eq. 4 you start with summing over all neighbours of $v$ for each dimension of $W$, could you comment on how this relates to summing over the subset of the neighbours? I am not sure, I understand what you mean with the “subset of neighbours in each dimension”.\n\nMinor language hiccups:\n\n•\tP. 4 2nd sentence below equations: “is the function [that computes] node degrees” and same for the function “that computes” the hidden features?\n\n•\tP. 4 second to last paragraph: second sentence suiable -> suitable\n\n•\tFirst sentence in 3.3: Is that what you wanted to say?\n\n•\tSame page “We use […] aggregation coefficient matrices [as shown by?] Luan et al.”\n\n•\tPage 6: last sentence before “Comparisons with multi-head GAT“: “this can be explained [by the fact] that […]”.\n\n•\tPage 7: Paragraph “Effect of powerful aggregators”: Third sentence: “We config[ure]” \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Strong theoretical insights into GNNs with somewhat weaker experimental results", "review": "The authors propose two new layers for GNNs. CombConv and ExpandingConv are motivated by the insight that a GNN is only as expressive as the rank of the matrix that represents the coefficients of the aggregation function. To arrive at this statement, the authors formalize all GNNs as being composed of three steps: 1) generation of aggregation coefficients, 2) actual aggregation of the neighbourhood, and 3) feature extraction from the aggregation. Furthermore, it is shown that current approaches have very low distinguishing strength and that CombConv and ExpandingConv, by their construction, yield higher expressive power. \nThe effectiveness of different components of their layers (e.g. Re-SUM, applying a ReLU non-linearity before summing when features are computed form the aggregation) are investigated in an ablation study. Additionally, proposed layers are compared with current approaches on 4 data sets. \n\nThis paper constructs an interesting theoretical analysis of GNNs and finds the bottleneck of these networks to be in the coefficient matrix of the aggregation scheme. While I was not able to check all proofs, it seems like a solid mathematical analysis. What I find somewhat sobering is the experimental section. First, I am surprised that you only compare your method on four data sets and that you miss some reported by your comparison partners (such as IMDB, REDDIT, PROTEINS, etc.). I understand the Graph Kernel data sets are smaller, however, they’ve been used in comparable papers before. Second, while the title and the theoretical analysis promises much higher distinguishing strength, I am surprised that your performance gains are good but not outstanding. I wonder if you could construct a synthetic data set in which you can show how you break the expressive bottleneck in practice. Would it make sense to compute the rank of M for different networks (including yours ((it is not guaranteed it is $s$, right?)) and plot it as a function of predictive performance? \nIn Eq. 4 you start with summing over all neighbours of $v$ for each dimension of $W$, could you comment on how this relates to summing over the subset of the neighbours? I am not sure, I understand what you mean with the “subset of neighbours in each dimension”.\n\nMinor language hiccups:\n\n•\tP. 4 2nd sentence below equations: “is the function [that computes] node degrees” and same for the function “that computes” the hidden features?\n\n•\tP. 4 second to last paragraph: second sentence suiable -> suitable\n\n•\tFirst sentence in 3.3: Is that what you wanted to say?\n\n•\tSame page “We use […] aggregation coefficient matrices [as shown by?] Luan et al.”\n\n•\tPage 6: last sentence before “Comparisons with multi-head GAT“: “this can be explained [by the fact] that […]”.\n\n•\tPage 7: Paragraph “Effect of powerful aggregators”: Third sentence: “We config[ure]” \n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603914515468}, {"id": "sRrSDMkrnN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1702/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Post Rebuttal**\n\nI thank the authors for the extensive experiments and answers. Unfortunately, I still feel that the contribution is rather marginal. I keep my score unchanged.\n\n---\n\n**Summary of Contributions** \nThe paper points at the expressive bottleneck of GNNs as the weak distinguishing strength of learned aggregators. The authors then formulate the aggregator operation via a standalone matrix, and analyze the requirements for it to have a sufficient distinguishing strength. \nThen, based on the analysis, 2 aggregation schemes with enlarged strength are proposed. Also, the benefit of applying activation before aggregation is shown. \n\n**Strengths** \n- *Simplicity of analysis* - The paper presents a simple formulation to the aggregation operator via the matrix coefficients which provide an intuitive view of the requirements from the aggregator functions. \n- The paper presents a general formulation for GNNs under the suggested framework, and shows how other GNNs fall into this framework. \n-  *Thorough ablation study confirming theory* - the ablation study on s nicely shows the improvement in performance as s grows. \n- Achieving SOTA results on some benchmarks.\n\n**Weaknesses** \n- *Contribution and comparison to Corso et. al. (2020)* - Although the paper states results regarding the requirements from strong aggregators, as very similar result has already been introduced in Corso et. al.. In that case, I would have expected to see comparison in performance as both papers tackle the same problem. \n- *Comparison to GAT* - The paper shows how GAT can be formulated under the ExpandingConv formulation, raising the question, does it perform as well as the proposed methods?\n\n\n**Recommendation**\nThe paper posses an interesting and simple view of aggregators in GNNs however I have concerns regarding the significance of contribution. Therefore, I rate it as marginally bellow acceptance threshold. \n\n**Additional Comments**\n- Inaccuracies in GCN, GAT aggregation coefficients in p.4, missing the self coefficient. \n\n\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "concerns about significance of contribution", "review": "**Post Rebuttal**\n\nI thank the authors for the extensive experiments and answers. Unfortunately, I still feel that the contribution is rather marginal. I keep my score unchanged.\n\n---\n\n**Summary of Contributions** \nThe paper points at the expressive bottleneck of GNNs as the weak distinguishing strength of learned aggregators. The authors then formulate the aggregator operation via a standalone matrix, and analyze the requirements for it to have a sufficient distinguishing strength. \nThen, based on the analysis, 2 aggregation schemes with enlarged strength are proposed. Also, the benefit of applying activation before aggregation is shown. \n\n**Strengths** \n- *Simplicity of analysis* - The paper presents a simple formulation to the aggregation operator via the matrix coefficients which provide an intuitive view of the requirements from the aggregator functions. \n- The paper presents a general formulation for GNNs under the suggested framework, and shows how other GNNs fall into this framework. \n-  *Thorough ablation study confirming theory* - the ablation study on s nicely shows the improvement in performance as s grows. \n- Achieving SOTA results on some benchmarks.\n\n**Weaknesses** \n- *Contribution and comparison to Corso et. al. (2020)* - Although the paper states results regarding the requirements from strong aggregators, as very similar result has already been introduced in Corso et. al.. In that case, I would have expected to see comparison in performance as both papers tackle the same problem. \n- *Comparison to GAT* - The paper shows how GAT can be formulated under the ExpandingConv formulation, raising the question, does it perform as well as the proposed methods?\n\n\n**Recommendation**\nThe paper posses an interesting and simple view of aggregators in GNNs however I have concerns regarding the significance of contribution. Therefore, I rate it as marginally bellow acceptance threshold. \n\n**Additional Comments**\n- Inaccuracies in GCN, GAT aggregation coefficients in p.4, missing the self coefficient. \n\n\n\n\n\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603890869766}, {"id": "tgkKGCCTyR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1702/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary of the paper: The main objective of the paper is to improve the expressiveness of the GNN by exploring powerful aggregators. The requirements to build more powerful aggregators are analysed. It is closely related to finding strategy for preserving the rank of hidden features, and implies that basic aggregators correspond to a special case of low-rank transformations.\n\nStrengths: The idea is promising. A new GNN formulation is proposed: the aggregation is represented as the multiplication of hidden feature matrix of neighbours and the aggregation coefficient matrix.\n\nWeaknesses: The strength mentioned above (multiplication of hidden features values and the aggregation) is also a weakness: I have an impression that already known results are presented in a much more complex way. The paper is not easy to follow in general ( e.g., the sentence \"The difference is that each dimension of hidden features is aggregated with an independent weighted aggregator which works like a comb\".)\n\nThe paper needs to be throughly read: use \\citep instead of \\cite where it is necessary. \n\nThe improvements reported in the experimental section seem to be not really significant.\n\nQuestions: Could you provide an intuition for the definition of the distinguishing strength? (Section 3.1). \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposes to explore powerful aggregators to improve the expressiveness of the GNN. It is quite difficult to read the paper. Not sure that the numerical results show significant improvement.  ", "review": "Summary of the paper: The main objective of the paper is to improve the expressiveness of the GNN by exploring powerful aggregators. The requirements to build more powerful aggregators are analysed. It is closely related to finding strategy for preserving the rank of hidden features, and implies that basic aggregators correspond to a special case of low-rank transformations.\n\nStrengths: The idea is promising. A new GNN formulation is proposed: the aggregation is represented as the multiplication of hidden feature matrix of neighbours and the aggregation coefficient matrix.\n\nWeaknesses: The strength mentioned above (multiplication of hidden features values and the aggregation) is also a weakness: I have an impression that already known results are presented in a much more complex way. The paper is not easy to follow in general ( e.g., the sentence \"The difference is that each dimension of hidden features is aggregated with an independent weighted aggregator which works like a comb\".)\n\nThe paper needs to be throughly read: use \\citep instead of \\cite where it is necessary. \n\nThe improvements reported in the experimental section seem to be not really significant.\n\nQuestions: Could you provide an intuition for the definition of the distinguishing strength? (Section 3.1). \n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603791959868}], "openreview_url": "https://openreview.net/forum?id=ztMLindFLWR", "arxiv_id": "2012.07219", "paper_pdf": "papers/ztMLindFLWR.pdf", "paper_pdf_sha256": "676865eab7cb0510e8c452632a31dfb4db77efee5a9020e0ca4ba0758ee19194", "paper_pdf_bytes": 1365702, "paper_pdf_source": "openreview", "code_url": "https://github.com/qslim/epcb-gnns", "code_repository": "qslim/epcb-gnns", "code_commit": "87bca5bb371ec15fe0fb68f7a93ae9933ef0e76a", "code_archive": "repos/ztMLindFLWR.zip", "code_archive_sha256": "a35d57fdb76fe8cf73bf1e332a2b5b0fd2ca66da11ef37c997592de85ba0f047", "code_archive_bytes": 36353, "code_file_count": 24, "code_extensions": {".py": 19, ".sh": 5}, "github_disk_usage_kb": 27, "github_languages": {"Python": 99910, "Shell": 3373}, "github_archived": false, "github_pushed_at": "2022-06-21T08:29:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/breaking-the-expressive-bottlenecks-of-graph-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1xTup4KPr", "year": 2020, "status": "rejected", "title": "Needles in Haystacks: On Classifying Tiny Objects in Large Images", "authors": ["Nick Pawlowski", "Suvrat Bhooshan", "Nicolas Ballas", "Francesco Ciompi", "Ben Glocker", "Michal Drozdzal"], "authorids": ["pawlowski.nick@gmail.com", "sbh@fb.com", "ballasn@fb.com", "f.ciompi@gmail.com", "b.glocker@imperial.ac.uk", "mdrozdzal@fb.com"], "authors_source": "OpenReview API", "abstract": "In some important computer vision domains, such as medical or hyperspectral imaging, we care about the classification of tiny objects in large images. However, most Convolutional Neural Networks (CNNs) for image classification were developed using biased datasets that contain large objects, in mostly central image positions. To assess whether classical CNN architectures work well for tiny object classification we build a comprehensive testbed containing two datasets: one derived from MNIST digits and one from histopathology images. This testbed allows controlled experiments to stress-test CNN architectures with a broad spectrum of signal-to-noise ratios. Our observations indicate that: (1) There exists a limit to signal-to-noise below which CNNs fail to generalize and that this limit is affected by dataset size - more data leading to better performances; however, the amount of training data required for the model to generalize scales rapidly with the inverse of the object-to-image ratio (2) in general, higher capacity models exhibit better generalization; (3) when knowing the approximate object sizes, adapting receptive field is beneficial; and (4) for very small signal-to-noise ratio the choice of global pooling operation affects optimization, whereas for relatively large signal-to-noise values, all tested global pooling operations exhibit similar performance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Sye60k9rqS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper649/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present an empirical study to evaluate the performance of CNN-based object classifiers for situations in which the object of interest is very small relative to the size of the image. Two artificial datasets, based on MNIST and histopathological images are introduced to conduct the experiments. Through empirical evaluation the authors conclude that the size of the dataset required for generalization increases rapidly with the inverse of the O2I ratio, that higher capacity models generalize better, and that accounting for the model's receptive field is key.\n\nThe contributions of the study are limited: i) The artificial datasets generate images that are small and O2Is that are big for the applications of interest, e.g., gigapixels images in digital pathology (see Figure 1). ii) The dataset based on MNIST is perhaps too artificial (too structured), once one compares their results relative to nCAMELYON. iii) The authors only consider ResNet-50. iv) The authors do not consider multi-instance learning pooling functions, e.g., noisy or, noisy and or attention. v) The authors do not consider performance as a function of the positive instances in the image (number occurrences of 3 in the proposed nMNIST). vi) There is no methodological contribution.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The authors present an empirical study to evaluate the performance of CNN-based object classifiers for situations in which the object of interest is very small relative to the size of the image. Two artificial datasets, based on MNIST and histopathological images are introduced to conduct the experiments. Through empirical evaluation the authors conclude that the size of the dataset required for generalization increases rapidly with the inverse of the O2I ratio, that higher capacity models generalize better, and that accounting for the model's receptive field is key.\n\nThe contributions of the study are limited: i) The artificial datasets generate images that are small and O2Is that are big for the applications of interest, e.g., gigapixels images in digital pathology (see Figure 1). ii) The dataset based on MNIST is perhaps too artificial (too structured), once one compares their results relative to nCAMELYON. iii) The authors only consider ResNet-50. iv) The authors do not consider multi-instance learning pooling functions, e.g., noisy or, noisy and or attention. v) The authors do not consider performance as a function of the positive instances in the image (number occurrences of 3 in the proposed nMNIST). vi) There is no methodological contribution."}, "tcdate": 1572343764925}, {"id": "HylRTeIAYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper649/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a testbed framework to investigate the limitations of CNN at the classification of tiny objects and the effects of signal to noise ratio has in the task. The implemented framework will be made available online upon acceptance.\n\nI believe that the question the authors try to answer is very interesting and worth of exploration. However, I am a bit less excited about the achieved results since I consider them not to be sufficient to drive conclusions. The experiments proposed by the authors are run on two different datasets created ad-hoc: the nMNIST and the nCAMELYON, both modifications of MNIST and CAMELYON datasets. Over all the experiments run, the behaviors observed on the two datasets are not the same. The explanation provided by the authors is that since nCAMELYON is very small, results are different from the ones of nMNIST. While I consider this a valid explanation, this still limits the overall conclusions of this paper. Therefore, my main recommendation to the authors would be to identify other datasets for their experiment.\n\nIn the dicussion section, the authors argue that MS COCO is not a good candidate since the background usually contains information that allows to infer the label. I wonder if it would be possible to generate a new dataset, starting from MS COCO, that avoids this problem. Using the example of the paper, I would argue that it is possible to obtain images of outdoors with people where there are no balls. Having such a dataset would make your paper much stronger.\n\nFinally, to some extent, I consider that your problem is strongly linked to anomaly/detection detection. In the end, this is what a needle in a haystack is. How do you position yourselves with respect to the state of the art on this topic?  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper presents a testbed framework to investigate the limitations of CNN at the classification of tiny objects and the effects of signal to noise ratio has in the task. The implemented framework will be made available online upon acceptance.\n\nI believe that the question the authors try to answer is very interesting and worth of exploration. However, I am a bit less excited about the achieved results since I consider them not to be sufficient to drive conclusions. The experiments proposed by the authors are run on two different datasets created ad-hoc: the nMNIST and the nCAMELYON, both modifications of MNIST and CAMELYON datasets. Over all the experiments run, the behaviors observed on the two datasets are not the same. The explanation provided by the authors is that since nCAMELYON is very small, results are different from the ones of nMNIST. While I consider this a valid explanation, this still limits the overall conclusions of this paper. Therefore, my main recommendation to the authors would be to identify other datasets for their experiment.\n\nIn the dicussion section, the authors argue that MS COCO is not a good candidate since the background usually contains information that allows to infer the label. I wonder if it would be possible to generate a new dataset, starting from MS COCO, that avoids this problem. Using the example of the paper, I would argue that it is possible to obtain images of outdoors with people where there are no balls. Having such a dataset would make your paper much stronger.\n\nFinally, to some extent, I consider that your problem is strongly linked to anomaly/detection detection. In the end, this is what a needle in a haystack is. How do you position yourselves with respect to the state of the art on this topic?  "}, "tcdate": 1571868869851}, {"id": "SylBTBUTYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper649/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The submission proposes an analysis of the impact of object size in images when performing classification tasks using neural networks of the BagNet family. The analysis is performed on two datasets, a large resolution cluttered MNIST and a histopathology dataset named nCAMELYON.\n\nThe paper attack interesting questions and links the size of the object in the image (O2I) to the training dataset size required. Also, showing that max-pooling is the only pooling operation that converges for very low O2I (but is the slowest to converge at higher O2I) is interesting and encourages discussion about the training (optimization) process.\n\nIn my opinion, the main issue about the submission is the limited depth in the contributions, analyzing a single family of network architectures (BagNets) over two datasets, one of which is relatively small. The family of R-CNN and its derivatives were especially designed to counter the impact of object size, it would have been interesting to include them in the analysis. Furthermore, limited insights can be carried out for tasks related to classification such as localization and segmentation. Models such as Single Shot Detectors split the image into grids and variable anchor sizes to perform their inference, are they affected to a lesser extent by object size? The limited insights provided to potential future readers prevents me from recommending the submission for acceptance.\n\nOn p. 7 (and fig. 7 (b) ), it is said that the lower performance of the larger receptive field suggests that class-relevant information is contained in the texture. I am not sure about this remark, I would have expected the network to learn to focus on the right regions, provided the receptive field is big enough to see the whole object of interest. Could the decrease in performance be attributed to the increased amount of learnable parameters that ended up too large for the “relatively small nCAMELYON dataset used for training”? (sec. 3.1, Global pooling operations)\n\nFig. 16 seems to suggest that multiple numbers can overlap significantly in nMNIST, yielding potentially confusing images even for humans. I am not sure if this is a desirable characteristic for such dataset.\n\n\nMinor details\n- Please use \\cdot instead of the asterisk operator to denote multiplication (sec. 3, footnote 3, fig. 5 (b-c) and sec. 3.1;\n- Fig. 4 uses two different (and non-linear) values for the x-axes, giving the impression that both curves can be compared, that might be confusing to the reader;\n- Sec 1 and 5 “pixel-level level annotations”: duplicate “level”;\n- p. 6 “1 and , 2”: extraneous comma.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The submission proposes an analysis of the impact of object size in images when performing classification tasks using neural networks of the BagNet family. The analysis is performed on two datasets, a large resolution cluttered MNIST and a histopathology dataset named nCAMELYON.\n\nThe paper attack interesting questions and links the size of the object in the image (O2I) to the training dataset size required. Also, showing that max-pooling is the only pooling operation that converges for very low O2I (but is the slowest to converge at higher O2I) is interesting and encourages discussion about the training (optimization) process.\n\nIn my opinion, the main issue about the submission is the limited depth in the contributions, analyzing a single family of network architectures (BagNets) over two datasets, one of which is relatively small. The family of R-CNN and its derivatives were especially designed to counter the impact of object size, it would have been interesting to include them in the analysis. Furthermore, limited insights can be carried out for tasks related to classification such as localization and segmentation. Models such as Single Shot Detectors split the image into grids and variable anchor sizes to perform their inference, are they affected to a lesser extent by object size? The limited insights provided to potential future readers prevents me from recommending the submission for acceptance.\n\nOn p. 7 (and fig. 7 (b) ), it is said that the lower performance of the larger receptive field suggests that class-relevant information is contained in the texture. I am not sure about this remark, I would have expected the network to learn to focus on the right regions, provided the receptive field is big enough to see the whole object of interest. Could the decrease in performance be attributed to the increased amount of learnable parameters that ended up too large for the “relatively small nCAMELYON dataset used for training”? (sec. 3.1, Global pooling operations)\n\nFig. 16 seems to suggest that multiple numbers can overlap significantly in nMNIST, yielding potentially confusing images even for humans. I am not sure if this is a desirable characteristic for such dataset.\n\n\nMinor details\n- Please use \\cdot instead of the asterisk operator to denote multiplication (sec. 3, footnote 3, fig. 5 (b-c) and sec. 3.1;\n- Fig. 4 uses two different (and non-linear) values for the x-axes, giving the impression that both curves can be compared, that might be confusing to the reader;\n- Sec 1 and 5 “pixel-level level annotations”: duplicate “level”;\n- p. 6 “1 and , 2”: extraneous comma.\n"}, "tcdate": 1571804604614}], "openreview_url": "https://openreview.net/forum?id=H1xTup4KPr", "arxiv_id": "1908.06037", "paper_pdf": "papers/H1xTup4KPr.pdf", "paper_pdf_sha256": "7da3f4a4f9ca3f738bbf5e6132d49c609df7a82a032badf85147fc6afc45cc3a", "paper_pdf_bytes": 13200937, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/Needles-in-Haystacks", "code_repository": "facebookresearch/Needles-in-Haystacks", "code_commit": "bd08ae43b8174f9e5cda8763d99ba70a26a5b0c6", "code_archive": "repos/H1xTup4KPr.zip", "code_archive_sha256": "c1803276f7adfe6b7477887cf4442a0fc16eaf26f99b9bf1e8c8e81943af993c", "code_archive_bytes": 32871, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 25, "github_languages": {"Python": 58499}, "github_archived": true, "github_pushed_at": "2019-06-28T12:18:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/needles-in-haystacks-on-classifying-tiny"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bkf1tjR9KQ", "year": 2019, "status": "rejected", "title": "DVOLVER: Efficient Pareto-Optimal Neural Network Architecture Search", "authors": ["Guillaume Michel", "Mohammed Amine Alaoui", "Alice Lebois", "Amal Feriani", "Mehdi Felhi"], "authorids": ["guillaume.michel@netatmo.com", "mohammed-amine.alaoui@netatmo.com", "alice.lebois@netatmo.com", "amal.feriani@netatmo.com", "mehdi.felhi@netatmo.com"], "authors_source": "OpenReview API", "abstract": "Automatic search of neural network architectures is a standing research topic. In addition to the fact that it presents a faster alternative to hand-designed architectures, it can improve their efficiency and for instance generate Convolutional Neural Networks (CNN) adapted for mobile devices. In this paper, we present a multi-objective neural architecture search method to find a family of CNN models with the best accuracy and computational resources tradeoffs, in a search space inspired by the state-of-the-art findings in neural search. Our work, called Dvolver, evolves a population of architectures and iteratively improves an approximation of the optimal Pareto front. Applying Dvolver on the model accuracy and on the number of floating points operations as objective functions, we are able to find, in only 2.5 days 1 , a set of competitive mobile models on ImageNet. Amongst these models one architecture has the same Top-1 accuracy on ImageNet as NASNet-A mobile with 8% less floating point operations and another one has a Top-1 accuracy of 75.28% on ImageNet exceeding by 0.28% the best MobileNetV2 model for the same computational resources.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "ryeL_QZ9hm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper407/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a multi-objective search algorithm that designs resource-efficient convolutional architectures. The key idea is to maintain a population of networks and to iteratively approach the Pareto front through evolution. The normal & reduction cells are searched on CIFAR-10 and then transferred to ImageNet. The resulting architectures empirically lead to better trade-offs than other baselines.\n\nPros:\nThe paper is well-written and easy to comprehend.\nResults are competitive against strong baselines such as NASNet.\nResource budgets are handled in a principled manner with multi-objective optimization. \n\nCons:\n\nMy main concerns are on the technical novelty and experimental comparison.\n\nTechnical novelty:\n\nThe proposed algorithm seems highly similar to the existing multi-objective NAS algorithms, especially the ones based on Pareto optimality [1,2,3]. In Sect 2, the authors state that the main difference from prior works such as [2] and [3] is the usage of a different and larger search space and large-scale experiments. However, both aspects are of limited technical novelty.\n\nExperimental comparison:\n\nIn sect 3.3, the authors say “we noticed that the original NASNet search space can greatly benefit from extra connections from any given block”. If the proposed algorithm was investigated in an enhanced version of the NASNet space, it would be unclear whether we should attribute the reported performance to the proposed multi-objective evolution or this additional search space engineering. It would be better to report the results using the original space as well for fair comparison. \n\nThe main claimed contribution is a multi-objective evolutionary algorithm. To demonstrate its effectiveness, it would be necessary to compare against existent multi-objective NAS strategies in the literature. Most of those strategies (e.g., scalarization, weighted product method) should be straightforward to implement on top of the current search space. The current results are less convincing since the authors only compared their method against single-objective baselines (e.g. NASNet, PNAS, AmoebaNet) which are completely unaware of additional dimensions of the desired objectives. \n\nThe networks are searched on CIFAR-10 and then transferred to ImageNet. Unlike most prior works (including the ones focusing on resource-constrained NAS), the authors did not the final performance of their architecture on CIFAR-10. It would be informative to report the CIFAR-10 results as well.\n\nOther suggestions & questions:\nThe authors did not report their training setup for ImageNet. It would be good to include those details to ensure the readers are informed should there are any additional augmentations.\n\n“uniform mutation and a crossover probability of 0.1” (sect 4.1)\nIt would be better to included more details on these evolution forces for reproducibility. These are also important component of the proposed algorithm.\n\n“We manually select 3 architectures that we will be fully train on ImageNet in Section 4.2” (sect 4.1)\nI believe this part needs more clarifications since there can be a large number of architectures on the Pareto front. What’s the criteria for manual selection?\n\n[1] Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. \"Multi-objective architecture search for cnns.\" arXiv preprint arXiv:1804.09081 (2018).\n[2] Kim Ye-Hoon, Reddy Bhargava, Yun Sojung, and Seo Chanwon. NEMO: Neuro-Evolution with Multiobjective Optimization of Deep Neural Network for Speed and Accuracy. ICML’17 AutoML Workshop, 2017.\n[3] Dong, Jin-Dong, et al. \"DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures.\" arXiv preprint arXiv:1806.08198 (2018).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good results but limited novelty. Experimental comparison could be improved.", "review": "The paper proposes a multi-objective search algorithm that designs resource-efficient convolutional architectures. The key idea is to maintain a population of networks and to iteratively approach the Pareto front through evolution. The normal & reduction cells are searched on CIFAR-10 and then transferred to ImageNet. The resulting architectures empirically lead to better trade-offs than other baselines.\n\nPros:\nThe paper is well-written and easy to comprehend.\nResults are competitive against strong baselines such as NASNet.\nResource budgets are handled in a principled manner with multi-objective optimization. \n\nCons:\n\nMy main concerns are on the technical novelty and experimental comparison.\n\nTechnical novelty:\n\nThe proposed algorithm seems highly similar to the existing multi-objective NAS algorithms, especially the ones based on Pareto optimality [1,2,3]. In Sect 2, the authors state that the main difference from prior works such as [2] and [3] is the usage of a different and larger search space and large-scale experiments. However, both aspects are of limited technical novelty.\n\nExperimental comparison:\n\nIn sect 3.3, the authors say “we noticed that the original NASNet search space can greatly benefit from extra connections from any given block”. If the proposed algorithm was investigated in an enhanced version of the NASNet space, it would be unclear whether we should attribute the reported performance to the proposed multi-objective evolution or this additional search space engineering. It would be better to report the results using the original space as well for fair comparison. \n\nThe main claimed contribution is a multi-objective evolutionary algorithm. To demonstrate its effectiveness, it would be necessary to compare against existent multi-objective NAS strategies in the literature. Most of those strategies (e.g., scalarization, weighted product method) should be straightforward to implement on top of the current search space. The current results are less convincing since the authors only compared their method against single-objective baselines (e.g. NASNet, PNAS, AmoebaNet) which are completely unaware of additional dimensions of the desired objectives. \n\nThe networks are searched on CIFAR-10 and then transferred to ImageNet. Unlike most prior works (including the ones focusing on resource-constrained NAS), the authors did not the final performance of their architecture on CIFAR-10. It would be informative to report the CIFAR-10 results as well.\n\nOther suggestions & questions:\nThe authors did not report their training setup for ImageNet. It would be good to include those details to ensure the readers are informed should there are any additional augmentations.\n\n“uniform mutation and a crossover probability of 0.1” (sect 4.1)\nIt would be better to included more details on these evolution forces for reproducibility. These are also important component of the proposed algorithm.\n\n“We manually select 3 architectures that we will be fully train on ImageNet in Section 4.2” (sect 4.1)\nI believe this part needs more clarifications since there can be a large number of architectures on the Pareto front. What’s the criteria for manual selection?\n\n[1] Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. \"Multi-objective architecture search for cnns.\" arXiv preprint arXiv:1804.09081 (2018).\n[2] Kim Ye-Hoon, Reddy Bhargava, Yun Sojung, and Seo Chanwon. NEMO: Neuro-Evolution with Multiobjective Optimization of Deep Neural Network for Speed and Accuracy. ICML’17 AutoML Workshop, 2017.\n[3] Dong, Jin-Dong, et al. \"DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures.\" arXiv preprint arXiv:1806.08198 (2018).", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541178222337}, {"id": "HJxO1b4Yhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper407/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is easy to read. The authors did a job in describing the problem, concepts, and the proposed multi-objective optimization method. The computational results are on par with NASNet-A mobile. \n\nIt is good to know that we can use standard multi-objective method for neural architecture search. The implementation seems to be straightforward. The paper mainly uses existing ideas, but with some incremental improvements. It lacks novelty.  \n\nThe time reduction of this method on ImageNet comes from transfer learning by training on CIFAR-10 first. As the paper admits this is not going to generalizing well. How good the method is if just using a single dateset? For CIFAR-10, is this method comparable with ENAS(https://arxiv.org/pdf/1802.03268.pdf)?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The results are competitive, but not too much innovation", "review": "The paper is easy to read. The authors did a job in describing the problem, concepts, and the proposed multi-objective optimization method. The computational results are on par with NASNet-A mobile. \n\nIt is good to know that we can use standard multi-objective method for neural architecture search. The implementation seems to be straightforward. The paper mainly uses existing ideas, but with some incremental improvements. It lacks novelty.  \n\nThe time reduction of this method on ImageNet comes from transfer learning by training on CIFAR-10 first. As the paper admits this is not going to generalizing well. How good the method is if just using a single dateset? For CIFAR-10, is this method comparable with ENAS(https://arxiv.org/pdf/1802.03268.pdf)?\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541124320153}, {"id": "r1xaACsOhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper407/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a multi-objective neural architecture search based on an evolutionary algorithm. The contradicting objective functions are optimized by ranking the candidates by Pareto-dominance, replace the bottom 50% with new candidates generated by the top 50% candidates through random mutations. The multi-objective function considers classification accuracy and an approximation of the inference speed. The method is compared to MobileNet and Mobile NASNet on ImageNet indicating an improvement with respect to search time.\n\nThe authors admit that their work is incremental and a combination of existing work. Furthermore, they admit that Dong et al. (2018) is the closest related work, however, they do not compare to them in the experimental section. The method by Dong et al. requires only 8 GPU days (Dvolver requires 50) yielding very similar results. Why this has been ignored remains unclear.\n\nThe paper is not self-contained, important methodological aspects of the method are insufficiently described. I recommend at least to formally define the crowding distance. It would be also reasonable to define your objective functions already in Section 3 instead of mentioning them in the caption of Figure 3 and its axis labels.\n\nI think it's fair to call your approach evolutionary but you might want to discuss its relationship to beam search and in this scope discuss [A].\n\nThe comparison in Table 2 is not fair. You use the swift activation function and do not report the corresponding numbers for MobileNet or Mobile NASNet. Ramachandran et al. (2017) report these (75% and 74.2% for NASNet and MobileNet).\nComparing the Dvolver architecture with ReLU activations to MobileNet does not indicate any improvements.\n\nYou mention that most previous approaches are only keeping track of the best solution while you evolve over a population. Maybe this sentence is not well written and something else is meant but now this statement is wrong.\n\n[A] Thomas Elsken, Jan Hendrik Metzen, Frank Hutter: Simple And Efficient Architecture Search for Convolutional Neural Networks. CoRR abs/1711.04528 (2017)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting but incremental work", "review": "The authors propose a multi-objective neural architecture search based on an evolutionary algorithm. The contradicting objective functions are optimized by ranking the candidates by Pareto-dominance, replace the bottom 50% with new candidates generated by the top 50% candidates through random mutations. The multi-objective function considers classification accuracy and an approximation of the inference speed. The method is compared to MobileNet and Mobile NASNet on ImageNet indicating an improvement with respect to search time.\n\nThe authors admit that their work is incremental and a combination of existing work. Furthermore, they admit that Dong et al. (2018) is the closest related work, however, they do not compare to them in the experimental section. The method by Dong et al. requires only 8 GPU days (Dvolver requires 50) yielding very similar results. Why this has been ignored remains unclear.\n\nThe paper is not self-contained, important methodological aspects of the method are insufficiently described. I recommend at least to formally define the crowding distance. It would be also reasonable to define your objective functions already in Section 3 instead of mentioning them in the caption of Figure 3 and its axis labels.\n\nI think it's fair to call your approach evolutionary but you might want to discuss its relationship to beam search and in this scope discuss [A].\n\nThe comparison in Table 2 is not fair. You use the swift activation function and do not report the corresponding numbers for MobileNet or Mobile NASNet. Ramachandran et al. (2017) report these (75% and 74.2% for NASNet and MobileNet).\nComparing the Dvolver architecture with ReLU activations to MobileNet does not indicate any improvements.\n\nYou mention that most previous approaches are only keeping track of the best solution while you evolve over a population. Maybe this sentence is not well written and something else is meant but now this statement is wrong.\n\n[A] Thomas Elsken, Jan Hendrik Metzen, Frank Hutter: Simple And Efficient Architecture Search for Convolutional Neural Networks. CoRR abs/1711.04528 (2017)", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541091029448}], "openreview_url": "https://openreview.net/forum?id=Bkf1tjR9KQ", "arxiv_id": "1902.01654", "paper_pdf": "papers/Bkf1tjR9KQ.pdf", "paper_pdf_sha256": "64c31a0ea1fa3bb2fdcfb0e2b8a41bee8f6e93f094ec74d634c03301deb5df7b", "paper_pdf_bytes": 376334, "paper_pdf_source": "openreview", "code_url": "https://github.com/guillaume-michel/dvolver", "code_repository": "guillaume-michel/dvolver", "code_commit": "b1301c2790b91172fc3322b1836cb71e55afe7b9", "code_archive": "repos/Bkf1tjR9KQ.zip", "code_archive_sha256": "7760873ac19c77427b0984c6b4dd01022f3300df9dfe73ff128e2931e60291c8", "code_archive_bytes": 65891, "code_file_count": 29, "code_extensions": {".py": 27, ".sh": 2}, "github_disk_usage_kb": 54, "github_languages": {"Python": 198562, "Shell": 4794}, "github_archived": false, "github_pushed_at": "2019-02-04T12:39:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dvolver-efficient-pareto-optimal-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJmAXkgCb", "year": 2018, "status": "rejected", "title": "DNN Feature Map Compression using Learned Representation over GF(2)", "authors": ["Denis A. Gudovskiy", "Alec Hodgkinson", "Luca Rigazio"], "authorids": ["denis.gudovskiy@us.panasonic.com", "alec.hodgkinson@us.panasonic.com", "luca.rigazio@us.panasonic.com"], "authors_source": "OpenReview API", "abstract": "In this paper, we introduce a method to compress intermediate feature maps of deep neural networks (DNNs) to decrease memory storage and bandwidth requirements during inference. Unlike previous works, the proposed method is based on converting fixed-point activations into vectors over the smallest GF(2) finite field followed by nonlinear dimensionality reduction (NDR) layers embedded into a DNN. Such an end-to-end learned representation finds more compact feature maps by exploiting quantization redundancies within the fixed-point activations along the channel or spatial dimensions. We apply the proposed network architecture to the tasks of ImageNet classification and PASCAL VOC object detection. Compared to prior approaches, the conducted experiments show a factor of 2 decrease in memory requirements with minor degradation in accuracy while adding only bitwise computations.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BJ46Rwjez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper200/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Strengths:\n- Unlike most previous approaches that suffer from significant accuracy drops for good feature map compression, the proposed method achieves reductions in feature map sizes of 1 order of magnitude at effectively no loss in accuracy.\n- Technical approach relates closely to some of the prior approaches (e.g., Iandola et al. 2016) but can be viewed as learning the quantization rather than relying on a predefined one.\n- Good results on both large-scale classification and object detection.\n- Technical approach is clearly presented.\n\nWeaknesses:\n- The primary downside is that the approach requires a specialized architecture to work well (all experiments are done with SqueezeNets). Thus, the approach is less general than prior work, which can be applied to arbitrary architectures.\n- From the experiments it is not fully clear what is the performance loss due to having to use the SqueezeNet architecture rather than state-of-the-art models. For example, for the image categorization experiment, the comparative baselines are for AlexNet and NIN, which are outdated and do not represent the state-of-the-art in this field. The object detection experiments are based on a variant of Faster R-CNN where the VGG16 feature extractor is replaced with a SqueezeNet model. However, the drop in accuracy caused by this modification is not discussed in the paper and, in any case, there are now much better models for object detection than Faster R-CNN.\n- In my view the strengths of the approach would be more convincingly conveyed visually with a plot reporting accuracy versus memory usage, rather than by the many numerical tables in the paper.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "high compression rate for marginal accuracy loss, but approach requires specialized architecture", "rating": "4: Ok but not good enough - rejection", "review": "Strengths:\n- Unlike most previous approaches that suffer from significant accuracy drops for good feature map compression, the proposed method achieves reductions in feature map sizes of 1 order of magnitude at effectively no loss in accuracy.\n- Technical approach relates closely to some of the prior approaches (e.g., Iandola et al. 2016) but can be viewed as learning the quantization rather than relying on a predefined one.\n- Good results on both large-scale classification and object detection.\n- Technical approach is clearly presented.\n\nWeaknesses:\n- The primary downside is that the approach requires a specialized architecture to work well (all experiments are done with SqueezeNets). Thus, the approach is less general than prior work, which can be applied to arbitrary architectures.\n- From the experiments it is not fully clear what is the performance loss due to having to use the SqueezeNet architecture rather than state-of-the-art models. For example, for the image categorization experiment, the comparative baselines are for AlexNet and NIN, which are outdated and do not represent the state-of-the-art in this field. The object detection experiments are based on a variant of Faster R-CNN where the VGG16 feature extractor is replaced with a SqueezeNet model. However, the drop in accuracy caused by this modification is not discussed in the paper and, in any case, there are now much better models for object detection than Faster R-CNN.\n- In my view the strengths of the approach would be more convincingly conveyed visually with a plot reporting accuracy versus memory usage, rather than by the many numerical tables in the paper.\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511911099731}, {"id": "rJbz1nrgM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper200/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The method of this paper minimizes the memory usage of the activation maps of a CNN. It starts from a representation where activations are compressed with a uniform scalar quantizer and fused to reduce intermediate memory usage. This looses some accuracy, so the contribution of the paper is to add a pair of convolution layers in the binary domain (GF(2)) that are trained to restore the lost precision. \n\nOverall, this paper seems to be a nice addition to the body of works on network compression. \n\n+ : interesting approach and effective results. \n\n+ : well related to the state of the art and good comparison with other works. \n\n- : somewhat incremental. Most of the claimed 100x compression is due to previous work.\n\n- : impact on runtime is not reported. Since there is a caffe implementation it would be interesting to have an additional column with the comparative execution speeds, even if only on CPU. I would expect the FP32 timings to be hard to beat, despite the claims that it uses only binary operations.\n\n- : the paper is sometimes difficult to understand (see below)\n\ndetailed comments: \n\nEquations (3)-(4) are difficult to understand. If I understand correctly, b just decomposes a \\hat{x} in {0..2^B-1} into its B bits \\tilda{x} \\in {0,1}^B, which can be then considered as an additional dimension in the activation map where \\hat{x} comes from. \n\nIt is not stated clearly whether P^l and R^l have binary weights. My understanding is that P^l has but R^l not.\n\n4.1 --> a discussion of the large mini-batch size (1024) could be useful. My understanding is that large mini-batches are required to use averaged gradients and get smooth updates. \n\nend of 4.1 --> unclear what \"equivalent bits\" means\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "initial review", "rating": "7: Good paper, accept", "review": "The method of this paper minimizes the memory usage of the activation maps of a CNN. It starts from a representation where activations are compressed with a uniform scalar quantizer and fused to reduce intermediate memory usage. This looses some accuracy, so the contribution of the paper is to add a pair of convolution layers in the binary domain (GF(2)) that are trained to restore the lost precision. \n\nOverall, this paper seems to be a nice addition to the body of works on network compression. \n\n+ : interesting approach and effective results. \n\n+ : well related to the state of the art and good comparison with other works. \n\n- : somewhat incremental. Most of the claimed 100x compression is due to previous work.\n\n- : impact on runtime is not reported. Since there is a caffe implementation it would be interesting to have an additional column with the comparative execution speeds, even if only on CPU. I would expect the FP32 timings to be hard to beat, despite the claims that it uses only binary operations.\n\n- : the paper is sometimes difficult to understand (see below)\n\ndetailed comments: \n\nEquations (3)-(4) are difficult to understand. If I understand correctly, b just decomposes a \\hat{x} in {0..2^B-1} into its B bits \\tilda{x} \\in {0,1}^B, which can be then considered as an additional dimension in the activation map where \\hat{x} comes from. \n\nIt is not stated clearly whether P^l and R^l have binary weights. My understanding is that P^l has but R^l not.\n\n4.1 --> a discussion of the large mini-batch size (1024) could be useful. My understanding is that large mini-batches are required to use averaged gradients and get smooth updates. \n\nend of 4.1 --> unclear what \"equivalent bits\" means\n\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511534345075}, {"id": "SJG0Ga5ef", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper200/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In order to compress DNN intermediate feature maps the authors covert fixed-point activations into vectors over the smallest finite field, the Galois field of two elements (GF(2)) and use nonlinear dimentionality reduction layers.\n\nThe paper reads well and the methods and experiments are generally described in sufficient detail.\n\nMy main concern with this paper and approach is the performance achieved. According to Table 1 and Table 2 there is a small accuracy benefit from using the proposed approach over the \"quantized\" SqueezeNet baseline. If I am weighing in the need to alter the network for the proposed approach in comparison with the \"quantized\" setting then, from practical point of view, I would prefer the later \"quantized\" approach.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "DNN Feature Map Compression using Learned Representation", "rating": "5: Marginally below acceptance threshold", "review": "In order to compress DNN intermediate feature maps the authors covert fixed-point activations into vectors over the smallest finite field, the Galois field of two elements (GF(2)) and use nonlinear dimentionality reduction layers.\n\nThe paper reads well and the methods and experiments are generally described in sufficient detail.\n\nMy main concern with this paper and approach is the performance achieved. According to Table 1 and Table 2 there is a small accuracy benefit from using the proposed approach over the \"quantized\" SqueezeNet baseline. If I am weighing in the need to alter the network for the proposed approach in comparison with the \"quantized\" setting then, from practical point of view, I would prefer the later \"quantized\" approach.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511867081566}], "openreview_url": "https://openreview.net/forum?id=SJmAXkgCb", "arxiv_id": "1808.05285", "paper_pdf": "papers/SJmAXkgCb.pdf", "paper_pdf_sha256": "55d99b9ba1df096bd61823c91e709a819b825bff577e7b60fc196caa026789d4", "paper_pdf_bytes": 280243, "paper_pdf_source": "openreview", "code_url": "https://github.com/gudovskiy/fmap_compression", "code_repository": "gudovskiy/fmap_compression", "code_commit": "0e60097613b7fd7eac1d72775b202fc7a9bbbcea", "code_archive": "repos/SJmAXkgCb.zip", "code_archive_sha256": "eb371b7d36a9d4aa2b7ac6a2bc66806b25eb170273d7ba63fdda38dde43d6f23", "code_archive_bytes": 75728, "code_file_count": 17, "code_extensions": {".cpp": 7, ".cu": 6, ".hpp": 3, ".cuh": 1}, "github_disk_usage_kb": 29, "github_languages": {"C++": 72058, "Cuda": 28024}, "github_archived": false, "github_pushed_at": "2018-11-21T01:40:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dnn-feature-map-compression-using-learned"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Am95bfE207", "year": 2026, "status": "rejected", "title": "From Low Intrinsic Dimensionality to Non-Vacuous Generalization Bounds in Deep Multi-Task Learning", "authors": ["Hossein Zakerinia", "Dorsa Ghobadi", "Christoph H. Lampert"], "authorids": ["~Hossein_Zakerinia1", "~Dorsa_Ghobadi2", "~Christoph_H._Lampert6"], "authors_source": "OpenReview API", "abstract": "Deep learning methods are known to generalize well from training to future data, even in an overparametrized regime, where they could easily overfit. One explanation for this phenomenon is that even when their ambient dimensionality, (i.e. the number of parameters) is large, the models’ intrinsic dimensionality is small; specifically, their learning takes place in a small subspace of all possible weight configurations. In this work, we confirm this phenomenon in the setting of deep multi-task learning. We introduce a method to parametrize multi-task network directly in the low-dimensional space, facilitated by the use of random expansions techniques. We then show that high-accuracy multi-task solutions can be found with much smaller intrinsic dimensionality (fewer free parameters) than what single-task learning requires. Subsequently, we show that the low-dimensional representations in combination with weight compression and PAC-Bayesian reasoning lead to the first non-vacuous generalization bounds for deep multi-task networks.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "qbcYQ5agZ8", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12218/Reviewer_hMHb"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes a parameter-sharing approach for multi-task learning based on a hierarchical parameterization of an intrinsic-dimensional model. Specifically, the method represents a set of shared parameters as (k) (l)-dimensional vectors, each projected into the network parameter space of dimension (D) using a different random matrix (P_i), and models each task as a linear combination of these shared components. The authors leverage this parameter-sharing structure to apply compression-based PAC-Bayesian bounds, leading to non-vacuous generalization guarantees in multi-task learning settings. Empirically, they demonstrate that this parameter-sharing strategy can achieve a pre-set 90% accuracy while operating in a significantly lower \"amortized\"-intrinsic dimensionality, thereby enjoying favorable generalization guarantees.", "review_text": "The paper proposes a parameter-sharing approach for multi-task learning based on a hierarchical parameterization of an intrinsic-dimensional model. Specifically, the method represents a set of shared parameters as (k) (l)-dimensional vectors, each projected into the network parameter space of dimension (D) using a different random matrix (P_i), and models each task as a linear combination of these shared components. The authors leverage this parameter-sharing structure to apply compression-based PAC-Bayesian bounds, leading to non-vacuous generalization guarantees in multi-task learning settings. Empirically, they demonstrate that this parameter-sharing strategy can achieve a pre-set 90% accuracy while operating in a significantly lower \"amortized\"-intrinsic dimensionality, thereby enjoying favorable generalization guarantees.", "strengths": "* The paper is clearly written and well organized.\n* The proposed approach is conceptually sound and well motivated.\n* The computed PAC-Bayesian bounds are non-vacuous, which is a meaningful result for multi-task learning.", "weaknesses": "* The technical novelty appears limited.\n* The work largely builds upon the compression-based PAC-Bayesian framework of Lotfi et al. (2022) and extends it in a relatively straightforward manner.\n* The benefit of the proposed hierarchical parameter-sharing approach over existing methods is not clearly demonstrated.\n* It remains unclear how the method compares to simpler parameter-sharing schemes, such as decomposing the parameter (w) into shared and task-specific components as in Li et al. (2018).\n\n**Minor Comments:**\n\n* Reporting test set errors alongside the theoretical guarantees would strengthen the empirical evaluation.\n* For the Transformer-based experiments, it would be useful to show the performance of the pretrained Vision Transformer on CIFAR without fine-tuning.\n* Since Lotfi et al. (2022) evaluated their compression-based PAC-Bayesian framework on larger-scale datasets such as ImageNet, it would be valuable for the authors to test their approach on a comparable scale to better assess its general applicability and scalability.", "questions": "See the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a parameter-sharing approach for multi-task learning based on a hierarchical parameterization of an intrinsic-dimensional model. Specifically, the method represents a set of shared parameters as (k) (l)-dimensional vectors, each projected into the network parameter space of dimension (D) using a different random matrix (P_i), and models each task as a linear combination of these shared components. The authors leverage this parameter-sharing structure to apply compression-based PAC-Bayesian bounds, leading to non-vacuous generalization guarantees in multi-task learning settings. Empirically, they demonstrate that this parameter-sharing strategy can achieve a pre-set 90% accuracy while operating in a significantly lower \"amortized\"-intrinsic dimensionality, thereby enjoying favorable generalization guarantees.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* The paper is clearly written and well organized.\n* The proposed approach is conceptually sound and well motivated.\n* The computed PAC-Bayesian bounds are non-vacuous, which is a meaningful result for multi-task learning.", "weaknesses": "* The technical novelty appears limited.\n* The work largely builds upon the compression-based PAC-Bayesian framework of Lotfi et al. (2022) and extends it in a relatively straightforward manner.\n* The benefit of the proposed hierarchical parameter-sharing approach over existing methods is not clearly demonstrated.\n* It remains unclear how the method compares to simpler parameter-sharing schemes, such as decomposing the parameter (w) into shared and task-specific components as in Li et al. (2018).\n\n**Minor Comments:**\n\n* Reporting test set errors alongside the theoretical guarantees would strengthen the empirical evaluation.\n* For the Transformer-based experiments, it would be useful to show the performance of the pretrained Vision Transformer on CIFAR without fine-tuning.\n* Since Lotfi et al. (2022) evaluated their compression-based PAC-Bayesian framework on larger-scale datasets such as ImageNet, it would be valuable for the authors to test their approach on a comparable scale to better assess its general applicability and scalability.", "questions": "See the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762365844445}, {"id": "bDGiaGPURa", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12218/Reviewer_bsC4"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This work studies how multi-task learning can reduce the number of parameters needed to reach high accuracy by sharing structure acrosss tasks. The authors propose the concept of amortized intrinsic dimensionality, which measure the effective number of parameters each task needs when learned jointly with other tasks. The idea is that if tasks are related, they can reuse most of the representational space, so the per-task parameter requirement drops well below that of training them independently. The authors implement this through a shared low-rank parameterization: a small set of shared components captures global structure, and each task learns a few coefficients to adapt them. They empirically show across diverse datasets that as the number of tasks increases, tha amortized parameter cost decreases while the accuracy stays mostly constant. This provides evidence that multi-task learning discovers a shared low-dimensional subspace that explains why such models compress and generalize well, resulting in better generalization bounds.", "review_text": "This work studies how multi-task learning can reduce the number of parameters needed to reach high accuracy by sharing structure acrosss tasks. The authors propose the concept of amortized intrinsic dimensionality, which measure the effective number of parameters each task needs when learned jointly with other tasks. The idea is that if tasks are related, they can reuse most of the representational space, so the per-task parameter requirement drops well below that of training them independently. The authors implement this through a shared low-rank parameterization: a small set of shared components captures global structure, and each task learns a few coefficients to adapt them. They empirically show across diverse datasets that as the number of tasks increases, tha amortized parameter cost decreases while the accuracy stays mostly constant. This provides evidence that multi-task learning discovers a shared low-dimensional subspace that explains why such models compress and generalize well, resulting in better generalization bounds.", "strengths": "- The authors clearly demonstrate that the multi-task setting results in higher compressibility and better bounds \"in their particular setting\". The rigorous formulation makes the multi-task efficiency measurable and comparable to the single task setting.\n- The amortized intrinsic dimensionality provides a simple defitition of 'how much sharing happens' and can be evaluated directly from experiments.\n- There is a clear scaling behavior as the number of tasks grows, where the amortized dimension drops continuouslt, which supports the central claim that related tasks occupy a common representation space.", "weaknesses": "- The paper’s contribution is mostly conceptual and empirical as it does not show theoretically that if the tasks are related, e.g., measured in some form of distributional distance, the bounds will be tighter compared to the single task setting. It would be great to provide a sufficiency theoretical guarantee where if the task are related given a certain metric, the bounds will be theoretically improved. \n- There is no analysis to determine a priori whether the tasks are related or to identify negative transfer when unrelated tasks are included. The experimental setup is limited to very simple and clearly related tasks. \n- While results are support the main claim that related rasks result in better bound, the paper doesn’t offer prescriptive recipes for how to choose related tasks or how many to choose in order to improve generalization; back to the distance measurement question. \n- The experiments could be extended to more complicated datasets that are harder to perform well on, where the multi-task setting would have clearer empirical benefits to be tested against the bounds.\n- The bounds are not compared against sota techniques to obtain tight bounds: what if sota techniques for single task result in better bounds than the multi-task bounds?", "questions": "- Can you provide a theoretical sufficiency guarantee: for example, proving that if tasks are related under a defined distance metric, the amortized bound will be tighter than the single-tasks one?\n- Is there a quantitative criterion for task relatedness that could predict when multi tak will help or harm performance? Could mutual information or divergence measures serve that role?\n- How does the framework work in the presence of unrelated tasks? Can you empirically test for this case?\n- Can you suggest a practical recipe for selecting related tasks or estimating the right number of tasks to group jointly to improve generalization?\n- Would extending experiments to harder datasets or less correlated task suites change the observed scaling trends?\n- How do your amortized bounds compare numerically to state-of-the-art single-task generalization bounds? is the improvement consistent when using the best available baselines?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studies how multi-task learning can reduce the number of parameters needed to reach high accuracy by sharing structure acrosss tasks. The authors propose the concept of amortized intrinsic dimensionality, which measure the effective number of parameters each task needs when learned jointly with other tasks. The idea is that if tasks are related, they can reuse most of the representational space, so the per-task parameter requirement drops well below that of training them independently. The authors implement this through a shared low-rank parameterization: a small set of shared components captures global structure, and each task learns a few coefficients to adapt them. They empirically show across diverse datasets that as the number of tasks increases, tha amortized parameter cost decreases while the accuracy stays mostly constant. This provides evidence that multi-task learning discovers a shared low-dimensional subspace that explains why such models compress and generalize well, resulting in better generalization bounds.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The authors clearly demonstrate that the multi-task setting results in higher compressibility and better bounds \"in their particular setting\". The rigorous formulation makes the multi-task efficiency measurable and comparable to the single task setting.\n- The amortized intrinsic dimensionality provides a simple defitition of 'how much sharing happens' and can be evaluated directly from experiments.\n- There is a clear scaling behavior as the number of tasks grows, where the amortized dimension drops continuouslt, which supports the central claim that related tasks occupy a common representation space.", "weaknesses": "- The paper’s contribution is mostly conceptual and empirical as it does not show theoretically that if the tasks are related, e.g., measured in some form of distributional distance, the bounds will be tighter compared to the single task setting. It would be great to provide a sufficiency theoretical guarantee where if the task are related given a certain metric, the bounds will be theoretically improved. \n- There is no analysis to determine a priori whether the tasks are related or to identify negative transfer when unrelated tasks are included. The experimental setup is limited to very simple and clearly related tasks. \n- While results are support the main claim that related rasks result in better bound, the paper doesn’t offer prescriptive recipes for how to choose related tasks or how many to choose in order to improve generalization; back to the distance measurement question. \n- The experiments could be extended to more complicated datasets that are harder to perform well on, where the multi-task setting would have clearer empirical benefits to be tested against the bounds.\n- The bounds are not compared against sota techniques to obtain tight bounds: what if sota techniques for single task result in better bounds than the multi-task bounds?", "questions": "- Can you provide a theoretical sufficiency guarantee: for example, proving that if tasks are related under a defined distance metric, the amortized bound will be tighter than the single-tasks one?\n- Is there a quantitative criterion for task relatedness that could predict when multi tak will help or harm performance? Could mutual information or divergence measures serve that role?\n- How does the framework work in the presence of unrelated tasks? Can you empirically test for this case?\n- Can you suggest a practical recipe for selecting related tasks or estimating the right number of tasks to group jointly to improve generalization?\n- Would extending experiments to harder datasets or less correlated task suites change the observed scaling trends?\n- How do your amortized bounds compare numerically to state-of-the-art single-task generalization bounds? is the improvement consistent when using the best available baselines?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762196757748}, {"id": "YchNWeASHv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12218/Reviewer_Rg7L"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper studies multi-task learning and considers a setting where several tasks can be represented through a shared low-dimensional structure together with small task-specific parameters. The main idea is to describe all tasks jointly using a compact encoder and then apply a compression-style generalization bound so that the generalization error depends on the size of this joint description rather than on the sum of per-task model sizes. The authors argue that this captures the intuition that related tasks should be \"paid for\" only once at the level of the shared representation.\n\nTo support those claims, the authors apply a known generalization bound to this setting, introduce a notion of complexity that aims to reflect the intuition above (\"amortized intrinsic dimension\"), and provide empirical evidence.\n\nTypos\n- In table 1, \"m\" is still undefined (it's only defined later, eg in Figure 1).\n- In definition 4, extra \"its\" in \"whose its base set\".\n- Line 351, a \"s\" is missing in \"Theorem 2 become ...\".\n- A point is missing in: the hypothesis box, Equation 8, Equation 11, Equation 14, Equation 17, Equation 18, and Equation 22.", "review_text": "The paper studies multi-task learning and considers a setting where several tasks can be represented through a shared low-dimensional structure together with small task-specific parameters. The main idea is to describe all tasks jointly using a compact encoder and then apply a compression-style generalization bound so that the generalization error depends on the size of this joint description rather than on the sum of per-task model sizes. The authors argue that this captures the intuition that related tasks should be \"paid for\" only once at the level of the shared representation.\n\nTo support those claims, the authors apply a known generalization bound to this setting, introduce a notion of complexity that aims to reflect the intuition above (\"amortized intrinsic dimension\"), and provide empirical evidence.\n\nTypos\n- In table 1, \"m\" is still undefined (it's only defined later, eg in Figure 1).\n- In definition 4, extra \"its\" in \"whose its base set\".\n- Line 351, a \"s\" is missing in \"Theorem 2 become ...\".\n- A point is missing in: the hypothesis box, Equation 8, Equation 11, Equation 14, Equation 17, Equation 18, and Equation 22.", "strengths": "- The discussion seems to address natural questions that one would have when reading the results (i.e., what happens in corner cases when the tasks are very related or completely different), which I appreciate.\n- Experiments in Table 1 seem interesting and support their hypothesis.", "weaknesses": "- The statements are close to standard compression arguments (\"if you have an encoder, then the total complexity is the size of the encoder + that of the encoding of the joint tasks\") and seem somewhat straightforward given Theorem 1 from Shalev-Shwartz & Ben-David. The authors argue in lines 399-403 that the advantage is that the bound only depends on the encoding on the tasks all together and not the sum of the individual encodings. But in all the examples given (e.g., line 334 or line 350), it seems it is actually the sum of the individual encodings. Did I miss something? Is there an example where the joint code is shorter or are we just rewriting the sum?\n- In Section 2, line 078, the authors argue \"in this work, we rely on the concept that two tasks are related, if the mutual information between their data distributions is high\", but such information is neither defined nor used throughout the paper. The only vague mention of it I could find is in Appendix A.4, where the authors write \"Given the conceptual connections between compressibility, information, and entropy, the difference between these two quantities can be seen as a computable approximation to the mutual information between the tasks\" and point to a book on Kolmogorov complexity. As currently written, this angle is more confusing than helpful; if it helps, make the link more precise, otherwise remove it.\n- The accuracy used in definition 1 is the uniform average over tasks, regardless of their difficulty. It seems like it would be important to at least comment on what happens when tasks differ in difficulty or size (although I would expect the latter to be a fairly easy extension).", "questions": "- Could you also report the test error in Table 3? It seems like it would be interesting to have?\n- In Table 2, the authors displayed the bounds of Theorems 2 and 3 side by side, as if they were comparable, and argue \"the fast-rate bound offers improved guarantees compared to the more elementary Theorem 2\". How so? One bounds the difference between the population and empirical risks, while the other bounds the KL. Why is it sensible to compare them?\n- In Table 1, the resulting pairs $(l, k)$ are always such that $l > k$, which, if I understand correctly, means that the tasks are (sometimes very) related. Have you tried experimenting with the reverse scenario?\n- Maybe a stupid question, but if I were to think about a naive approach building upon prior work discussed in Section 2, I would try to concatenate the datasets, train a single model with the same low-rank parametrization of Equation 1, and compute the intrinsic dimensionality for this \"aggregated\" task. If this is (much) smaller than $\\text{number of tasks} \\times \\text{ID of a single task}$, then it would suggest it is beneficial to share the representations between the tasks. Why is this wrong?\n\nP.S.: confidence-wise, I am somewhere between 3 and 4; I checked the paper carefully, and I have a theoretical background (not specifically in generalization bounds, although I am familiar with these), but I am not an expert in the multi-task learning literature specifically.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies multi-task learning and considers a setting where several tasks can be represented through a shared low-dimensional structure together with small task-specific parameters. The main idea is to describe all tasks jointly using a compact encoder and then apply a compression-style generalization bound so that the generalization error depends on the size of this joint description rather than on the sum of per-task model sizes. The authors argue that this captures the intuition that related tasks should be \"paid for\" only once at the level of the shared representation.\n\nTo support those claims, the authors apply a known generalization bound to this setting, introduce a notion of complexity that aims to reflect the intuition above (\"amortized intrinsic dimension\"), and provide empirical evidence.\n\nTypos\n- In table 1, \"m\" is still undefined (it's only defined later, eg in Figure 1).\n- In definition 4, extra \"its\" in \"whose its base set\".\n- Line 351, a \"s\" is missing in \"Theorem 2 become ...\".\n- A point is missing in: the hypothesis box, Equation 8, Equation 11, Equation 14, Equation 17, Equation 18, and Equation 22.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The discussion seems to address natural questions that one would have when reading the results (i.e., what happens in corner cases when the tasks are very related or completely different), which I appreciate.\n- Experiments in Table 1 seem interesting and support their hypothesis.", "weaknesses": "- The statements are close to standard compression arguments (\"if you have an encoder, then the total complexity is the size of the encoder + that of the encoding of the joint tasks\") and seem somewhat straightforward given Theorem 1 from Shalev-Shwartz & Ben-David. The authors argue in lines 399-403 that the advantage is that the bound only depends on the encoding on the tasks all together and not the sum of the individual encodings. But in all the examples given (e.g., line 334 or line 350), it seems it is actually the sum of the individual encodings. Did I miss something? Is there an example where the joint code is shorter or are we just rewriting the sum?\n- In Section 2, line 078, the authors argue \"in this work, we rely on the concept that two tasks are related, if the mutual information between their data distributions is high\", but such information is neither defined nor used throughout the paper. The only vague mention of it I could find is in Appendix A.4, where the authors write \"Given the conceptual connections between compressibility, information, and entropy, the difference between these two quantities can be seen as a computable approximation to the mutual information between the tasks\" and point to a book on Kolmogorov complexity. As currently written, this angle is more confusing than helpful; if it helps, make the link more precise, otherwise remove it.\n- The accuracy used in definition 1 is the uniform average over tasks, regardless of their difficulty. It seems like it would be important to at least comment on what happens when tasks differ in difficulty or size (although I would expect the latter to be a fairly easy extension).", "questions": "- Could you also report the test error in Table 3? It seems like it would be interesting to have?\n- In Table 2, the authors displayed the bounds of Theorems 2 and 3 side by side, as if they were comparable, and argue \"the fast-rate bound offers improved guarantees compared to the more elementary Theorem 2\". How so? One bounds the difference between the population and empirical risks, while the other bounds the KL. Why is it sensible to compare them?\n- In Table 1, the resulting pairs $(l, k)$ are always such that $l > k$, which, if I understand correctly, means that the tasks are (sometimes very) related. Have you tried experimenting with the reverse scenario?\n- Maybe a stupid question, but if I were to think about a naive approach building upon prior work discussed in Section 2, I would try to concatenate the datasets, train a single model with the same low-rank parametrization of Equation 1, and compute the intrinsic dimensionality for this \"aggregated\" task. If this is (much) smaller than $\\text{number of tasks} \\times \\text{ID of a single task}$, then it would suggest it is beneficial to share the representations between the tasks. Why is this wrong?\n\nP.S.: confidence-wise, I am somewhere between 3 and 4; I checked the paper carefully, and I have a theoretical background (not specifically in generalization bounds, although I am familiar with these), but I am not an expert in the multi-task learning literature specifically.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762094719058}, {"id": "i9VYrlkb1T", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12218/Reviewer_BXs3"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The submission derives two generalization theorems that yield the first non-vacuous numerical generalization bounds for deep multi-task learning (MTL).\n\nThe submission suggests for any deep learning architecture an alternative lower-dimensional parametrization [Eq. (1) for STL from literature, Eq. (2-3) for MTL, Eq. (7) for TL]. If I understood it correctly, the submission derives computable non-vacuous generalization bounds for models that were trained via these lower-dimensional parametrizations, by using certain compression techniques (such as quantization) for these low-dimensional parameterizations.\n\nThe submission conducts experiments for these low-dimensional parametrizations and computes non-vacuous generalization bounds. Do I understand correctly that no generalization bound was explicitly numerically computed for fully trained models in this submission?\n\nTo be honest, there are some aspects that I did not fully understand (mainly regarding what the overall storyline is). My main confusion is about two sentences on page 2:\n\n“The bounds depend only on quantities that are available at training time and can therefore be evaluated numerically. ” and\n\n“This view, in particular, allows us to avoid apriori parametric assumptions”\n\nTo me, these two statements seem somewhat disconnected. It appears that the paper provides (i) numerically computable bounds under specific a priori parametric assumptions (the proposed low-dimensional parametrizations), and separately (ii) more general theoretical bounds that avoid such assumptions but remain mainly conceptual without a clear numerical instantiation strategy. If I understand it correctly, (i) is kind of an example of applying (ii), but under a specific parametric assumption. So there is quite some uncertainty in my scores, and my evaluations might change based on your answers to my questions.", "review_text": "The submission derives two generalization theorems that yield the first non-vacuous numerical generalization bounds for deep multi-task learning (MTL).\n\nThe submission suggests for any deep learning architecture an alternative lower-dimensional parametrization [Eq. (1) for STL from literature, Eq. (2-3) for MTL, Eq. (7) for TL]. If I understood it correctly, the submission derives computable non-vacuous generalization bounds for models that were trained via these lower-dimensional parametrizations, by using certain compression techniques (such as quantization) for these low-dimensional parameterizations.\n\nThe submission conducts experiments for these low-dimensional parametrizations and computes non-vacuous generalization bounds. Do I understand correctly that no generalization bound was explicitly numerically computed for fully trained models in this submission?\n\nTo be honest, there are some aspects that I did not fully understand (mainly regarding what the overall storyline is). My main confusion is about two sentences on page 2:\n\n“The bounds depend only on quantities that are available at training time and can therefore be evaluated numerically. ” and\n\n“This view, in particular, allows us to avoid apriori parametric assumptions”\n\nTo me, these two statements seem somewhat disconnected. It appears that the paper provides (i) numerically computable bounds under specific a priori parametric assumptions (the proposed low-dimensional parametrizations), and separately (ii) more general theoretical bounds that avoid such assumptions but remain mainly conceptual without a clear numerical instantiation strategy. If I understand it correctly, (i) is kind of an example of applying (ii), but under a specific parametric assumption. So there is quite some uncertainty in my scores, and my evaluations might change based on your answers to my questions.", "strengths": "Multi-task Learning, Transfer Learning, and out-of-sample generalization are very important topics, and a better theoretical understanding of those is highly desirable.\n\nMost generalization bounds in the literature are vacuous and not useful in practice. Non-vacuous generalization bounds are therefore appreciated.\n\nWhile most classical learning theory focuses on Single Task Learning (STL), much of the most exciting progress in applied Deep Learning (DL) is connected to Transfer Learning (TL) and Multi-task Learning. More theory in this direction is important.\n\nExplicitly computing non-vacuous generalization bounds on real-world datasets is a valuable contribution to the field.", "weaknesses": "The paper’s framing sometimes suggests that the bounds apply to arbitrary deep MTL models, whereas the actual, practically computable results seem to hold specifically for models trained under the proposed low-dimensional parametrizations (Eqs. (1), (2–3), (7)). Clarifying this scope early (perhaps even in the abstract) would prevent misunderstandings. In the current version, I got the feeling that the initial claims oversell the actual results a bit. I think the paper could be improved by communicating earlier, more clearly what the scope, the main results, and the contributions are.\n\nWhy don’t you additionally show the actual (training and) test loss in Tables 2 and 3?\n\nDo your generalization bounds correlate with actual generalization? When you train all your low-dimensional parametrizations for different values of $k$ and $l$ (and varying other hyperparameters), I would be very interested to see the training error, the generalization error bound, and the actual generalization error for all these combinations. Computing these numbers basically comes for free for you, given that you have already trained all of them.", "questions": "Q1: Do I correctly understand the intrinsic dimensions of Sections 2 and 3: The definition does not give the intrinsic dimension of a fully trained model f_theta, but the intrinsic dimension only depends on the architecture, the data distribution, the training data, and an accuracy level $\\tilde{A}:=\\tau\\%\\alpha$?\n\nI.e.,  the fully trained model f_theta is conditionally independent of the intrinsic dimension, conditioned on the architecture, the data distribution, the training data, and an accuracy level $\\tilde{A}$,.\n\nFor example a model $f_\\theta$ that learns very high complexity parameters, that totally overfits the training data, and expresses a very complicated high complexity function with poor generalization properties which cannot be compressed, would result in low intrinsic dimension in your experiments: As it only achieves a low test accuracy, it will be to find a low dimensional matrix P which achieves a similar test accuracy.\n\nOn the other hand, a model, $f_\\tilde{\\theta}$ that perfectly compressed the data in a sophisticated way (e.g. is extremely sparse), and therefore achieves a great test accuracy, would therefore correspond to a higher intrinsic dimension, as it is hard to find a low-dimensional matrix P which can match this test accuracy?\n\nSo, in other words, if I have different models $f_\\theta$ with the same architecture for a given dataset, their intrinsic dimension would only vary based on their test accuracy. In a way that better test accuracy corresponds to a higher intrinsic dimension.\n\nI think it would be good to emphasize very clearly that this notion of intrinsic dimension should not be seen as a property of the model $f_\\theta$, but this notion of intrinsic dimension only depends on the architecture, the data distribution, the training data, and an accuracy level.\n\nTo me, the formulation of Definition 1 feels quite misleading: it starts with models $f_i$, but then it uses them only to compute A (which uses the test dataset), but then you don’t even directly use $A$, but only  $\\tilde{A}:=\\tau\\%\\alpha$.\n\nThis formulation could mislead a naive reader into thinking that you are computing “the intrinsic dimension of an arbitrarily trained model $f_\\theta$” (in the spirit of the Local Learning Coefficient (LLC) in Singular Learning Theory) which could mislead a naive reader into thinking that those values of $\\theta$ with lower intrinsic dimension correspond to good generalization (as for the LLC). However, if I understood the submission correctly, this is absolutely not what’s happening here.\n\nI would explicitly formulate Def 1 as a property of the combination of an architecture, a data distribution, a training data set, and an accuracy level $\\tilde{A}$. I don’t see any reason to introduce $\\tau$, nor $A$, nor any fully trained model $f_\\theta$.\n\nQ2: My impression is that you derive generalization bounds for the parametrizations described in Eq. (1), (2-3) and (7))? But not for differently trained models? Technically, Theorems 2 and 3 are very general and could also apply to differently trained models. However, you don’t provide any explicit algorithm on how to numerically compute non-vacuous generalization bounds for fully trained DL models? Do I understand this correctly? Because your explicit compression algorithm assumes that the parameters of the model have the structure form Eq. (1), (2-3), or (7)? I think the paper would profit a lot from being crystal clear about these central questions.\n\nQ3: Line 144: This depends on the random realization of $\\theta_0$ and $P$. Do you average over multiple such realizations, or do you keep one fixed realization? Have you done some sensitivity analysis, if you get similar results for a different random seed, with a different random realization of $\\theta_0$ and $P$? Do the empirical results vary a lot from seed to seed, or are they quite stable? From a mathematical point of view, do you see the ID and AID as random variables that depend on the random variables $\\theta_0$ and $P$?\n\nQ4: Do you resample $\\theta_0$ for every $(k,l)$-gridpoint? Or do you sample $\\theta_0$ once and keep it fixed across the grid? Would this change the results? Intuitively, I would guess that the exact random realization of $\\ theta_0$ and $P$ does not matter a lot for sufficiently large models, but might matter quite a bit for smaller architectures.\n\nQ5: Your current version of Def 1 can only be computed with access to the validation/test set in order to compute $A$?\n\nQ6: minor remark: Line 146; “(2)/(3)” is an unusual formatting of equation numbers. “(2)-(3)”, “(2-3)”, “(2) and (3)” are more common choices, I think.\n\nQ7: Table 1: Why does the number of tasks $n$ depend on the architecture? For example, for split-CIFAR10, two different architectures ConvNet and ViT have different values of $n$ and $m$ while the dataset stays the same.\n\nQ8: Is $\\theta_0$ random for the ViT, or do you use $\\theta_0$ from the pretrained ViT? I would guess that using the pretrained $\\theta_0$ results in a significantly smaller intrinsic dimension than using the random $\\theta_0$. Do you agree? Have you tried that?\n\nQ9: Why is Table 2 not referenced in the text?\n\nQ10: Why don’t you show the actual train and test error in Tables 2 and 3? I would be very interested to see the training error, the numerically computed theoretical test error bound, and the actual test error on the test dataset for all considered combinations of $k$ and $l$.\n\nQ11: Are the models $f_1,\\dots,f_n$ in Section 5.3 trained via the parametrization from Eq. (2-3) or fully trained?\n\nQ 12: Are the models reported in Table 3 trained via the parametrization from Eq. (7) or fully trained?\n\nQ13: Lines 79-81: To me the sentence: “This view, in particular, allows us to avoid a priori parametric assumptions, such as similar model parameters (Evgeniou & Pontil, 2004) or the existence of a common feature space (Caruana, 1997).” feels quite confusing to me, as it seems to me that all the quantities that you can actually compute in practice from your submission also assume very specific non-standard parametrizations, i.e., Eq. (1) for STL, Eq. (2-3) for MTL, Eq. (7) for TL. Did you write this sentence because Theorems 1-3 are more generic? Do you see Theorems 2-3 as your main contribution, or Eq. (2-3) for MTL, Eq. (7) for TL?\n\nMy final score can still change in both directions depending on your answers.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission derives two generalization theorems that yield the first non-vacuous numerical generalization bounds for deep multi-task learning (MTL).\n\nThe submission suggests for any deep learning architecture an alternative lower-dimensional parametrization [Eq. (1) for STL from literature, Eq. (2-3) for MTL, Eq. (7) for TL]. If I understood it correctly, the submission derives computable non-vacuous generalization bounds for models that were trained via these lower-dimensional parametrizations, by using certain compression techniques (such as quantization) for these low-dimensional parameterizations.\n\nThe submission conducts experiments for these low-dimensional parametrizations and computes non-vacuous generalization bounds. Do I understand correctly that no generalization bound was explicitly numerically computed for fully trained models in this submission?\n\nTo be honest, there are some aspects that I did not fully understand (mainly regarding what the overall storyline is). My main confusion is about two sentences on page 2:\n\n“The bounds depend only on quantities that are available at training time and can therefore be evaluated numerically. ” and\n\n“This view, in particular, allows us to avoid apriori parametric assumptions”\n\nTo me, these two statements seem somewhat disconnected. It appears that the paper provides (i) numerically computable bounds under specific a priori parametric assumptions (the proposed low-dimensional parametrizations), and separately (ii) more general theoretical bounds that avoid such assumptions but remain mainly conceptual without a clear numerical instantiation strategy. If I understand it correctly, (i) is kind of an example of applying (ii), but under a specific parametric assumption. So there is quite some uncertainty in my scores, and my evaluations might change based on your answers to my questions.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "Multi-task Learning, Transfer Learning, and out-of-sample generalization are very important topics, and a better theoretical understanding of those is highly desirable.\n\nMost generalization bounds in the literature are vacuous and not useful in practice. Non-vacuous generalization bounds are therefore appreciated.\n\nWhile most classical learning theory focuses on Single Task Learning (STL), much of the most exciting progress in applied Deep Learning (DL) is connected to Transfer Learning (TL) and Multi-task Learning. More theory in this direction is important.\n\nExplicitly computing non-vacuous generalization bounds on real-world datasets is a valuable contribution to the field.", "weaknesses": "The paper’s framing sometimes suggests that the bounds apply to arbitrary deep MTL models, whereas the actual, practically computable results seem to hold specifically for models trained under the proposed low-dimensional parametrizations (Eqs. (1), (2–3), (7)). Clarifying this scope early (perhaps even in the abstract) would prevent misunderstandings. In the current version, I got the feeling that the initial claims oversell the actual results a bit. I think the paper could be improved by communicating earlier, more clearly what the scope, the main results, and the contributions are.\n\nWhy don’t you additionally show the actual (training and) test loss in Tables 2 and 3?\n\nDo your generalization bounds correlate with actual generalization? When you train all your low-dimensional parametrizations for different values of $k$ and $l$ (and varying other hyperparameters), I would be very interested to see the training error, the generalization error bound, and the actual generalization error for all these combinations. Computing these numbers basically comes for free for you, given that you have already trained all of them.", "questions": "Q1: Do I correctly understand the intrinsic dimensions of Sections 2 and 3: The definition does not give the intrinsic dimension of a fully trained model f_theta, but the intrinsic dimension only depends on the architecture, the data distribution, the training data, and an accuracy level $\\tilde{A}:=\\tau\\%\\alpha$?\n\nI.e.,  the fully trained model f_theta is conditionally independent of the intrinsic dimension, conditioned on the architecture, the data distribution, the training data, and an accuracy level $\\tilde{A}$,.\n\nFor example a model $f_\\theta$ that learns very high complexity parameters, that totally overfits the training data, and expresses a very complicated high complexity function with poor generalization properties which cannot be compressed, would result in low intrinsic dimension in your experiments: As it only achieves a low test accuracy, it will be to find a low dimensional matrix P which achieves a similar test accuracy.\n\nOn the other hand, a model, $f_\\tilde{\\theta}$ that perfectly compressed the data in a sophisticated way (e.g. is extremely sparse), and therefore achieves a great test accuracy, would therefore correspond to a higher intrinsic dimension, as it is hard to find a low-dimensional matrix P which can match this test accuracy?\n\nSo, in other words, if I have different models $f_\\theta$ with the same architecture for a given dataset, their intrinsic dimension would only vary based on their test accuracy. In a way that better test accuracy corresponds to a higher intrinsic dimension.\n\nI think it would be good to emphasize very clearly that this notion of intrinsic dimension should not be seen as a property of the model $f_\\theta$, but this notion of intrinsic dimension only depends on the architecture, the data distribution, the training data, and an accuracy level.\n\nTo me, the formulation of Definition 1 feels quite misleading: it starts with models $f_i$, but then it uses them only to compute A (which uses the test dataset), but then you don’t even directly use $A$, but only  $\\tilde{A}:=\\tau\\%\\alpha$.\n\nThis formulation could mislead a naive reader into thinking that you are computing “the intrinsic dimension of an arbitrarily trained model $f_\\theta$” (in the spirit of the Local Learning Coefficient (LLC) in Singular Learning Theory) which could mislead a naive reader into thinking that those values of $\\theta$ with lower intrinsic dimension correspond to good generalization (as for the LLC). However, if I understood the submission correctly, this is absolutely not what’s happening here.\n\nI would explicitly formulate Def 1 as a property of the combination of an architecture, a data distribution, a training data set, and an accuracy level $\\tilde{A}$. I don’t see any reason to introduce $\\tau$, nor $A$, nor any fully trained model $f_\\theta$.\n\nQ2: My impression is that you derive generalization bounds for the parametrizations described in Eq. (1), (2-3) and (7))? But not for differently trained models? Technically, Theorems 2 and 3 are very general and could also apply to differently trained models. However, you don’t provide any explicit algorithm on how to numerically compute non-vacuous generalization bounds for fully trained DL models? Do I understand this correctly? Because your explicit compression algorithm assumes that the parameters of the model have the structure form Eq. (1), (2-3), or (7)? I think the paper would profit a lot from being crystal clear about these central questions.\n\nQ3: Line 144: This depends on the random realization of $\\theta_0$ and $P$. Do you average over multiple such realizations, or do you keep one fixed realization? Have you done some sensitivity analysis, if you get similar results for a different random seed, with a different random realization of $\\theta_0$ and $P$? Do the empirical results vary a lot from seed to seed, or are they quite stable? From a mathematical point of view, do you see the ID and AID as random variables that depend on the random variables $\\theta_0$ and $P$?\n\nQ4: Do you resample $\\theta_0$ for every $(k,l)$-gridpoint? Or do you sample $\\theta_0$ once and keep it fixed across the grid? Would this change the results? Intuitively, I would guess that the exact random realization of $\\ theta_0$ and $P$ does not matter a lot for sufficiently large models, but might matter quite a bit for smaller architectures.\n\nQ5: Your current version of Def 1 can only be computed with access to the validation/test set in order to compute $A$?\n\nQ6: minor remark: Line 146; “(2)/(3)” is an unusual formatting of equation numbers. “(2)-(3)”, “(2-3)”, “(2) and (3)” are more common choices, I think.\n\nQ7: Table 1: Why does the number of tasks $n$ depend on the architecture? For example, for split-CIFAR10, two different architectures ConvNet and ViT have different values of $n$ and $m$ while the dataset stays the same.\n\nQ8: Is $\\theta_0$ random for the ViT, or do you use $\\theta_0$ from the pretrained ViT? I would guess that using the pretrained $\\theta_0$ results in a significantly smaller intrinsic dimension than using the random $\\theta_0$. Do you agree? Have you tried that?\n\nQ9: Why is Table 2 not referenced in the text?\n\nQ10: Why don’t you show the actual train and test error in Tables 2 and 3? I would be very interested to see the training error, the numerically computed theoretical test error bound, and the actual test error on the test dataset for all considered combinations of $k$ and $l$.\n\nQ11: Are the models $f_1,\\dots,f_n$ in Section 5.3 trained via the parametrization from Eq. (2-3) or fully trained?\n\nQ 12: Are the models reported in Table 3 trained via the parametrization from Eq. (7) or fully trained?\n\nQ13: Lines 79-81: To me the sentence: “This view, in particular, allows us to avoid a priori parametric assumptions, such as similar model parameters (Evgeniou & Pontil, 2004) or the existence of a common feature space (Caruana, 1997).” feels quite confusing to me, as it seems to me that all the quantities that you can actually compute in practice from your submission also assume very specific non-standard parametrizations, i.e., Eq. (1) for STL, Eq. (2-3) for MTL, Eq. (7) for TL. Did you write this sentence because Theorems 1-3 are more generic? Do you see Theorems 2-3 as your main contribution, or Eq. (2-3) for MTL, Eq. (7) for TL?\n\nMy final score can still change in both directions depending on your answers.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761975367138}, {"id": "p59bBQN31X", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12218/Reviewer_kX6A"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper studies the generalization performance of deep multi-task learning (MTL) through the lens of low intrinsic dimensionality.\nIt introduces the concept of Amortized Intrinsic Dimension (AID) (the effective number of parameters per task when shared representations are learned), and derives the first non-vacuous generalization bounds for deep MTL using compression and PAC-Bayesian analysis.\nEmpirical studies on multiple datasets show that AID is significantly smaller than the single-task intrinsic dimension, and the resulting generalization bounds are numerically non-vacuous.", "review_text": "This paper studies the generalization performance of deep multi-task learning (MTL) through the lens of low intrinsic dimensionality.\nIt introduces the concept of Amortized Intrinsic Dimension (AID) (the effective number of parameters per task when shared representations are learned), and derives the first non-vacuous generalization bounds for deep MTL using compression and PAC-Bayesian analysis.\nEmpirical studies on multiple datasets show that AID is significantly smaller than the single-task intrinsic dimension, and the resulting generalization bounds are numerically non-vacuous.", "strengths": "1. The paper extends compression-based and PAC-Bayesian generalization analysis to multi-task learning. Two new encoding-based generalization bounds for multi-task learning that result in non-vacuous guarantees on the multi-task risk for several standard datasets and model architectures are established.\n\n2. Experiments across several benchmarks convincingly demonstrate large reductions in intrinsic dimensionality and non-vacuous bounds.", "weaknesses": "1. Although the paper conceptually mentions mutual information as a measure of task relatedness, it remains largely qualitative.\nThere is no empirical estimation of task similarity or dependence, nor a quantitative analysis connecting such measures to the observed reductions in AID or to the resulting generalization gap.\n\n2. Most datasets are relatively small (e.g., MNIST, CIFAR), and the models are relatively shallow compared to modern large-scale MTL or LLM settings.\n\n3. The result is positioned as the first non-vacuous generalization bound for deep MTL, while the proof itself is relatively straightforward and closely follows prior single-task compression/PAC-Bayes analyses, so the technical novelty appears limited.", "questions": "1. Are there existing works that explain why low-dimensional shared representations emerge in multi-task learning, and that quantify how mutual information between task distributions influences the compression ratio or the tightness of generalization bounds?\n\n2. Could the authors clarify what specific theoretical or technical challenges arise in this extension and how their analysis addresses them?\n\n3. The paper fixes the threshold parameter $τ = 90$ when computing the intrinsic or amortized intrinsic dimension, yet it does not discuss how sensitive the results are to this choice. Would smaller or larger thresholds (e.g., $τ = 80$ or $95$) substantially change the estimated AID values or the resulting generalization bounds? A brief sensitivity analysis or justification for $τ = 90$ would strengthen the empirical credibility of the results.\n\n4. Minor comment: There is a small grammatical issue in Appendix, “the Theorem 2/3” should be written as “Theorem 2/3”.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the generalization performance of deep multi-task learning (MTL) through the lens of low intrinsic dimensionality.\nIt introduces the concept of Amortized Intrinsic Dimension (AID) (the effective number of parameters per task when shared representations are learned), and derives the first non-vacuous generalization bounds for deep MTL using compression and PAC-Bayesian analysis.\nEmpirical studies on multiple datasets show that AID is significantly smaller than the single-task intrinsic dimension, and the resulting generalization bounds are numerically non-vacuous.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper extends compression-based and PAC-Bayesian generalization analysis to multi-task learning. Two new encoding-based generalization bounds for multi-task learning that result in non-vacuous guarantees on the multi-task risk for several standard datasets and model architectures are established.\n\n2. Experiments across several benchmarks convincingly demonstrate large reductions in intrinsic dimensionality and non-vacuous bounds.", "weaknesses": "1. Although the paper conceptually mentions mutual information as a measure of task relatedness, it remains largely qualitative.\nThere is no empirical estimation of task similarity or dependence, nor a quantitative analysis connecting such measures to the observed reductions in AID or to the resulting generalization gap.\n\n2. Most datasets are relatively small (e.g., MNIST, CIFAR), and the models are relatively shallow compared to modern large-scale MTL or LLM settings.\n\n3. The result is positioned as the first non-vacuous generalization bound for deep MTL, while the proof itself is relatively straightforward and closely follows prior single-task compression/PAC-Bayes analyses, so the technical novelty appears limited.", "questions": "1. Are there existing works that explain why low-dimensional shared representations emerge in multi-task learning, and that quantify how mutual information between task distributions influences the compression ratio or the tightness of generalization bounds?\n\n2. Could the authors clarify what specific theoretical or technical challenges arise in this extension and how their analysis addresses them?\n\n3. The paper fixes the threshold parameter $τ = 90$ when computing the intrinsic or amortized intrinsic dimension, yet it does not discuss how sensitive the results are to this choice. Would smaller or larger thresholds (e.g., $τ = 80$ or $95$) substantially change the estimated AID values or the resulting generalization bounds? A brief sensitivity analysis or justification for $τ = 90$ would strengthen the empirical credibility of the results.\n\n4. Minor comment: There is a small grammatical issue in Appendix, “the Theorem 2/3” should be written as “Theorem 2/3”.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761951482157}], "openreview_url": "https://openreview.net/forum?id=Am95bfE207", "arxiv_id": "2501.19067", "paper_pdf": "papers/Am95bfE207.pdf", "paper_pdf_sha256": "afc7be806de1d88d2201446c496281181c0d0472e4f4299f9035659d7286d605", "paper_pdf_bytes": 350604, "paper_pdf_source": "openreview", "code_url": "https://github.com/hzakerinia/MTL", "code_repository": "hzakerinia/MTL", "code_commit": "b20ca886303c8cfe0a067422e653373c5c239c0a", "code_archive": "repos/Am95bfE207.zip", "code_archive_sha256": "a01c1af3089054a0e9a9933edc755f63dadf690ee359023d4041ab5bed4e8eff", "code_archive_bytes": 73462, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 60, "github_languages": {"Python": 257002}, "github_archived": false, "github_pushed_at": "2025-05-14T12:33:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-multi-task-learning-has-low-amortized"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fvo6q86NKG", "year": 2025, "status": "rejected", "title": "CBF-LLM: Safe Control for LLM Alignment", "authors": ["Yuya Miyaoka", "Masaki Inoue"], "authorids": ["~Yuya_Miyaoka1", "~Masaki_Inoue1"], "authors_source": "OpenReview API", "abstract": "This paper proposes a control-based framework for aligning large language models (LLMs) by leveraging a control barrier function (CBF) to ensure user-desirable text generation. \nThe presented framework applies the CBF safety filter to the predicted token generated from the baseline LLM, to intervene in the generated text.\nThe safety filter includes two significant advantages:\nthis safety filter is an add-on type, allowing it to be used for alignment purposes without fine-tuning the baseline LLM,\nand if there is an evaluation model regarding the desired alignment, it can be directly applied to the filter design.\nThe overall text-generation system is implemented with Llama 3 and a BERT model, aiming to generate positive text.\nFinally, further applications and limitations of the CBF-LLM for other alignment tasks, including topic-keeping and hallucination mitigating, are discussed.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "CjhgOnzJ9D", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8657/Reviewer_jp7H"], "rating": 1, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper studies controllable decoding in LLM generation, e.g., keeping the generated text of a positive sentiment or withing a specific topic. The paper proposes using a classifier that is trained on examples that are specific to the target control criteria. The classifier is used to define a control constraint that is checked repeatedly after generating each token. The probabilities of the candidate tokens that violate the constraint are set to zero. The paper limits the set of candidate tokens to k << vocab-size (e.g., 30) for efficiency. With a single prefix for controlling for the sentiment of the text and another single prefix for controlling for the topic, the paper shows some advantage over not applying any control at all.", "review_text": "The paper studies controllable decoding in LLM generation, e.g., keeping the generated text of a positive sentiment or withing a specific topic. The paper proposes using a classifier that is trained on examples that are specific to the target control criteria. The classifier is used to define a control constraint that is checked repeatedly after generating each token. The probabilities of the candidate tokens that violate the constraint are set to zero. The paper limits the set of candidate tokens to k << vocab-size (e.g., 30) for efficiency. With a single prefix for controlling for the sentiment of the text and another single prefix for controlling for the topic, the paper shows some advantage over not applying any control at all.", "strengths": "1. The presented method is learning free and broadly applicable.", "weaknesses": "1. The paper does not compare to existing baselines beyond just the naïve blocklist approach. In fact, there are published and peer-reviewed papers that achieve the same goal with the closely related method that the paper does not compare to, e.g., Mudgal et al, \"Controlled Decoding from Language Models\", ICML 2024.\n\n2. The paper does not present sufficient experimental results. It is just one prefix that is used per each of the two use cases demonstrated. That is not even sufficient for a workshop paper. The paper needs to provide quantitative results that are aggregated across several diverse examples. \n\n3. Related to the weak experiments, the paper also needs to provide human assessment of the produced generations and use that to demonstrate that the presented method truly introduces some value over the blocklist approach. Also, how does the control method impact the generation quality of the model (e.g., fluency, naturalness, general LLM capabilities), what happens if we wanted to control for more than one aspect, what is the role of the data properties (e.g., size) used to train the classifier on the overall quality. \n\n4. Even with limiting the candidate tokens size to the top-30, applying that methods at each decoding step (i.e., token generation) seems expensive. The paper needs to provide some details on that cost.", "questions": "Please see my questions under weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies controllable decoding in LLM generation, e.g., keeping the generated text of a positive sentiment or withing a specific topic. The paper proposes using a classifier that is trained on examples that are specific to the target control criteria. The classifier is used to define a control constraint that is checked repeatedly after generating each token. The probabilities of the candidate tokens that violate the constraint are set to zero. The paper limits the set of candidate tokens to k << vocab-size (e.g., 30) for efficiency. With a single prefix for controlling for the sentiment of the text and another single prefix for controlling for the topic, the paper shows some advantage over not applying any control at all.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The presented method is learning free and broadly applicable.", "weaknesses": "1. The paper does not compare to existing baselines beyond just the naïve blocklist approach. In fact, there are published and peer-reviewed papers that achieve the same goal with the closely related method that the paper does not compare to, e.g., Mudgal et al, \"Controlled Decoding from Language Models\", ICML 2024.\n\n2. The paper does not present sufficient experimental results. It is just one prefix that is used per each of the two use cases demonstrated. That is not even sufficient for a workshop paper. The paper needs to provide quantitative results that are aggregated across several diverse examples. \n\n3. Related to the weak experiments, the paper also needs to provide human assessment of the produced generations and use that to demonstrate that the presented method truly introduces some value over the blocklist approach. Also, how does the control method impact the generation quality of the model (e.g., fluency, naturalness, general LLM capabilities), what happens if we wanted to control for more than one aspect, what is the role of the data properties (e.g., size) used to train the classifier on the overall quality. \n\n4. Even with limiting the candidate tokens size to the top-30, applying that methods at each decoding step (i.e., token generation) seems expensive. The paper needs to provide some details on that cost.", "questions": "Please see my questions under weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730931779053}, {"id": "Ye7nv0pAY5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8657/Reviewer_yTZ8"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper deals with safety control in LLM alignment. The authors get inspiration from collision avoidance in control engineering and propose a control barrier function as the safety filter to ensure the user-desirable text. Experiments are conducted on Llama3 using a RoBERTa model as the CBF.", "review_text": "This paper deals with safety control in LLM alignment. The authors get inspiration from collision avoidance in control engineering and propose a control barrier function as the safety filter to ensure the user-desirable text. Experiments are conducted on Llama3 using a RoBERTa model as the CBF.", "strengths": "- The safety control of LLM outputs is a significant topic.   \n- It is interesting to see the connection between control engineering and text generation, e.g. the collision avoidance analogy to safety control in text generation.", "weaknesses": "- The proposed method is very similar to existing work inference-time constrained decoding such as SafeDecoding[1]. The authors should definitely discuss the line of work.\n\n- Based on my understanding, this work mainly introduces a CBF (specifically sentiment analysis RoBERTa) to measure a metric h(x) during decoding. The details of text generation the authors describe should be the standard greedy decoding process with top-k sampling. Instead of using standard greedy or beam search, they introduce the CBF to measure the toxicity of the generated text and restrict tokens with negative h(x). Then (1) it is so weird and not reliable to use a sentiment classifier to predict the toxicity instead of existing toxicity classifiers (2) even if using a reasonable toxicity classifier, a single token does not necessarily change a text sequence from safe to unsafe state. Existing works usually use guardrails such as Llama-Guard to measure the entire generated text sequence rather than part of the sequence.   \n\n- The experiments are not convincing. The only data used in the experiment is just one single sentence. To verify the effectiveness, experiments on existing jailbreak datasets such as AdvBench[2], HarmBench[3] are necessary. To claim its effectiveness on hallucination mitigation, more experiments on related datasets are also necessary rather than using just one sentence.\n\n- It is unclear whether this new decoding algorithm has an impact on the helpfulness of LLMs for other datasets such as MT-Bench. \n\n- The average token generation time also needs to be discussed.", "questions": "- Does Section 2.2 describe the standard greedy decoding process (correct me if I am wrong)? I would recommend the authors refer more to the traditional decoding algorithm in the NLP domain, otherwise it would be very confusing to the readers.\n- Even though it is interesting to discuss the counterpart and analogy in control engineering, relating it more to existing works in the LLM decoding would make it easier to tell the differences and see if there are real contributions.  \n\n[1] Xu, Z., Jiang, F., Niu, L., Jia, J., Lin, B. Y., & Poovendran, R. (2024). Safedecoding: Defending against jailbreak attacks via safety-aware decoding. ACL.\n\n[2] Zou, A., Wang, Z., Carlini, N., Nasr, M., Kolter, J. Z., & Fredrikson, M. (2023). Universal and transferable adversarial attacks on aligned language models. arXiv preprint.\n\n[3] Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., ... & Hendrycks, D. (2024). Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. ICML.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper deals with safety control in LLM alignment. The authors get inspiration from collision avoidance in control engineering and propose a control barrier function as the safety filter to ensure the user-desirable text. Experiments are conducted on Llama3 using a RoBERTa model as the CBF.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "- The safety control of LLM outputs is a significant topic.   \n- It is interesting to see the connection between control engineering and text generation, e.g. the collision avoidance analogy to safety control in text generation.", "weaknesses": "- The proposed method is very similar to existing work inference-time constrained decoding such as SafeDecoding[1]. The authors should definitely discuss the line of work.\n\n- Based on my understanding, this work mainly introduces a CBF (specifically sentiment analysis RoBERTa) to measure a metric h(x) during decoding. The details of text generation the authors describe should be the standard greedy decoding process with top-k sampling. Instead of using standard greedy or beam search, they introduce the CBF to measure the toxicity of the generated text and restrict tokens with negative h(x). Then (1) it is so weird and not reliable to use a sentiment classifier to predict the toxicity instead of existing toxicity classifiers (2) even if using a reasonable toxicity classifier, a single token does not necessarily change a text sequence from safe to unsafe state. Existing works usually use guardrails such as Llama-Guard to measure the entire generated text sequence rather than part of the sequence.   \n\n- The experiments are not convincing. The only data used in the experiment is just one single sentence. To verify the effectiveness, experiments on existing jailbreak datasets such as AdvBench[2], HarmBench[3] are necessary. To claim its effectiveness on hallucination mitigation, more experiments on related datasets are also necessary rather than using just one sentence.\n\n- It is unclear whether this new decoding algorithm has an impact on the helpfulness of LLMs for other datasets such as MT-Bench. \n\n- The average token generation time also needs to be discussed.", "questions": "- Does Section 2.2 describe the standard greedy decoding process (correct me if I am wrong)? I would recommend the authors refer more to the traditional decoding algorithm in the NLP domain, otherwise it would be very confusing to the readers.\n- Even though it is interesting to discuss the counterpart and analogy in control engineering, relating it more to existing works in the LLM decoding would make it easier to tell the differences and see if there are real contributions.  \n\n[1] Xu, Z., Jiang, F., Niu, L., Jia, J., Lin, B. Y., & Poovendran, R. (2024). Safedecoding: Defending against jailbreak attacks via safety-aware decoding. ACL.\n\n[2] Zou, A., Wang, Z., Carlini, N., Nasr, M., Kolter, J. Z., & Fredrikson, M. (2023). Universal and transferable adversarial attacks on aligned language models. arXiv preprint.\n\n[3] Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., ... & Hendrycks, D. (2024). Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. ICML.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730901057115}, {"id": "bd0mL7Zr2W", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8657/Reviewer_qEfD"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper tackles the task of LLM alignment from the control engineering perspective. Specifically, a filter is applied to the sequence generation steps of LLMs. The fundamental idea is that if a particular property is maintained at every step, that property should ultimately be evident in the final output.", "review_text": "This paper tackles the task of LLM alignment from the control engineering perspective. Specifically, a filter is applied to the sequence generation steps of LLMs. The fundamental idea is that if a particular property is maintained at every step, that property should ultimately be evident in the final output.", "strengths": "1. Tackling LLM generation from the perspective of control engineering is interesting. In particular, it is novel to  use a hyperparameter $\\alpha$ to adjust the tightness of the constraint in autoregressive text generation. \n2. The paper is well-written, and the intuition equations are clearly explained.", "weaknesses": "1. There is a lack of comparison to existing methods for decoding-time LLM alignments, including:\n    1. Mudgal et al., Controlled Decoding from Language Models, 2024,\n    2. Huang et al., DeAL: Decoding-time Alignment for Large Language Models, 2024\n    3. Yang et al., FUDGE: Controlled Text Generation With Future Discriminators, 2021\n2. To some extent, the proposed method can be considered a special case of the methods mentioned in Point 1, where the classifier makes hard decisions rather than giving scores. My educated guess (since the authors have not provided supporting experiments) is that the hard-filtering approach is worse. Here is my reason:\n   - The semantics of phrases are often determined by later-generated words/phrases; the hard decisions based on the early phrases may result in unnecessary pruning compared to the soft counterpart (with scorers). \n   - Suppose we want to generate a positive movie review; an audience may say, \"Despite a slow start, the movie blossomed into a riveting tale that kept me on the edge of my seat.\"  However, this is most likely not allowed by the proposed approach because the generation is cut off at  the \"slow start.\"", "questions": "Is the classifier (RoBERTa) trained on whole sequences or prefixes?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the task of LLM alignment from the control engineering perspective. Specifically, a filter is applied to the sequence generation steps of LLMs. The fundamental idea is that if a particular property is maintained at every step, that property should ultimately be evident in the final output.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. Tackling LLM generation from the perspective of control engineering is interesting. In particular, it is novel to  use a hyperparameter $\\alpha$ to adjust the tightness of the constraint in autoregressive text generation. \n2. The paper is well-written, and the intuition equations are clearly explained.", "weaknesses": "1. There is a lack of comparison to existing methods for decoding-time LLM alignments, including:\n    1. Mudgal et al., Controlled Decoding from Language Models, 2024,\n    2. Huang et al., DeAL: Decoding-time Alignment for Large Language Models, 2024\n    3. Yang et al., FUDGE: Controlled Text Generation With Future Discriminators, 2021\n2. To some extent, the proposed method can be considered a special case of the methods mentioned in Point 1, where the classifier makes hard decisions rather than giving scores. My educated guess (since the authors have not provided supporting experiments) is that the hard-filtering approach is worse. Here is my reason:\n   - The semantics of phrases are often determined by later-generated words/phrases; the hard decisions based on the early phrases may result in unnecessary pruning compared to the soft counterpart (with scorers). \n   - Suppose we want to generate a positive movie review; an audience may say, \"Despite a slow start, the movie blossomed into a riveting tale that kept me on the edge of my seat.\"  However, this is most likely not allowed by the proposed approach because the generation is cut off at  the \"slow start.\"", "questions": "Is the classifier (RoBERTa) trained on whole sequences or prefixes?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730815339852}, {"id": "MYMLt79OCU", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8657/Reviewer_qH9r"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper presents a method to filter the language produced by an LLM such as LLAMA; the filtering happens via an auxiliary algorithm/model that scans the original LLM’s output probabilities of vocabulary words to be produced next and picks out the word(s) that satisfy an additional requirement such as producing positive or negative text, or text on a specific topic. This filter function/model seems to be another pre-trained LLM as well, e.g., Roberta.", "review_text": "The paper presents a method to filter the language produced by an LLM such as LLAMA; the filtering happens via an auxiliary algorithm/model that scans the original LLM’s output probabilities of vocabulary words to be produced next and picks out the word(s) that satisfy an additional requirement such as producing positive or negative text, or text on a specific topic. This filter function/model seems to be another pre-trained LLM as well, e.g., Roberta.", "strengths": "The work retrofits LLAMA-type LLM to produce desired constrained language in a general way. \n\nThe work has a theoretical component to it wherein it attempts to control  language production using established techniques from the perspective of control theory. \n\nThe approach has been used to produce text that suits the purpose under different constraints.", "weaknesses": "The presented theoretical and algorithmic control approach seems cumbersome and done in a roundabout way, not straightforwardly.\n\nThere seems to be no evaluation of the generated text after filtration. Output sentence examples are given and seem good, but the main paper doesn’t contain evaluation metrics and scores. \n\nI find the analogy with a car not particularly persuasive. In a car, to avoid obstacles, the car’s internal processes must change to produce new trajectory for the car. However, in an LLM, to produce the desired output, the proposed approach lets the LLM produce output as usual, but seems to filter later to suit the purpose.\n\nIt will be good to cite a relevant paper: Zingale and Kalita. 2024. Language Model Sentence Completion with a Parser-Driven Rhetorical Control Method. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 193–203.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method to filter the language produced by an LLM such as LLAMA; the filtering happens via an auxiliary algorithm/model that scans the original LLM’s output probabilities of vocabulary words to be produced next and picks out the word(s) that satisfy an additional requirement such as producing positive or negative text, or text on a specific topic. This filter function/model seems to be another pre-trained LLM as well, e.g., Roberta.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "The work retrofits LLAMA-type LLM to produce desired constrained language in a general way. \n\nThe work has a theoretical component to it wherein it attempts to control  language production using established techniques from the perspective of control theory. \n\nThe approach has been used to produce text that suits the purpose under different constraints.", "weaknesses": "The presented theoretical and algorithmic control approach seems cumbersome and done in a roundabout way, not straightforwardly.\n\nThere seems to be no evaluation of the generated text after filtration. Output sentence examples are given and seem good, but the main paper doesn’t contain evaluation metrics and scores. \n\nI find the analogy with a car not particularly persuasive. In a car, to avoid obstacles, the car’s internal processes must change to produce new trajectory for the car. However, in an LLM, to produce the desired output, the proposed approach lets the LLM produce output as usual, but seems to filter later to suit the purpose.\n\nIt will be good to cite a relevant paper: Zingale and Kalita. 2024. Language Model Sentence Completion with a Parser-Driven Rhetorical Control Method. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 193–203.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730814081019}, {"id": "nUXY9seveZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8657/Reviewer_2hj6"], "rating": 3, "soundness": 1, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper presents a control-based framework for aligning LLMs to ensure the generation of user-desirable text. The framework leverages control barrier functions (CBFs), a concept from control engineering, to intervene in the output generation of LLMs, aiming to prevent the production of harmful, biased, or toxic content. The key contributions of the paper are:\n\n1 This paper presents a novel framework that applies a CBF-based safety filter to LLM output, aiming to reduce the need for interventions in text generation while maintaining alignment with user specifications.\n\n2 This paper demonstrates the practical application of the CBF-LLM framework using Llama 3 and a sentiment analysis RoBERTa model, showing its effectiveness in generating positive content.\n\n3 This paper attempts to connect control engineering with NLP by adapting control theory techniques for LLM alignment, offering a new perspective on ensuring the safety and ethicality of LLM-generated content.", "review_text": "The paper presents a control-based framework for aligning LLMs to ensure the generation of user-desirable text. The framework leverages control barrier functions (CBFs), a concept from control engineering, to intervene in the output generation of LLMs, aiming to prevent the production of harmful, biased, or toxic content. The key contributions of the paper are:\n\n1 This paper presents a novel framework that applies a CBF-based safety filter to LLM output, aiming to reduce the need for interventions in text generation while maintaining alignment with user specifications.\n\n2 This paper demonstrates the practical application of the CBF-LLM framework using Llama 3 and a sentiment analysis RoBERTa model, showing its effectiveness in generating positive content.\n\n3 This paper attempts to connect control engineering with NLP by adapting control theory techniques for LLM alignment, offering a new perspective on ensuring the safety and ethicality of LLM-generated content.", "strengths": "1. The paper is well-written, and I really like their figures.  \n2. There is a detailed theoretical transfer and explanation on control engineering and how it can be extended to LLMs.  \n3. This paper provides some inspiration for future non-parametric optimization methods for LLMs.  \n4. To some extent, it achieves an alignment from weak (RoBERTa) to strong (LLaMA).", "weaknesses": "1. The experiments and datasets lack persuasiveness. All experiments in the paper are based on only a few queries, with almost no evaluations on common LLM benchmarks or other metric.  \n2. The approach heavily relies on a Language-Constraint Function (L-CF), implemented with RoBERTa and assumed to be a golden classifier. However, I believe that assuming RoBERTa as an ideal classifier is not reasonable in practical use.", "questions": "1. Why not conduct experiments on at least a small-scale widely-used LLM evaluation datasets, such as a subset of Anthropic/hh-rlhf?  \n2. I feel that finding a reliable Language-Constraint Function (L-CF) is similarly challenging to training a golden reward model. It might be insightful to explore scenarios where the Language-Constraint Function (L-CF) has varying levels of reliability.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a control-based framework for aligning LLMs to ensure the generation of user-desirable text. The framework leverages control barrier functions (CBFs), a concept from control engineering, to intervene in the output generation of LLMs, aiming to prevent the production of harmful, biased, or toxic content. The key contributions of the paper are:\n\n1 This paper presents a novel framework that applies a CBF-based safety filter to LLM output, aiming to reduce the need for interventions in text generation while maintaining alignment with user specifications.\n\n2 This paper demonstrates the practical application of the CBF-LLM framework using Llama 3 and a sentiment analysis RoBERTa model, showing its effectiveness in generating positive content.\n\n3 This paper attempts to connect control engineering with NLP by adapting control theory techniques for LLM alignment, offering a new perspective on ensuring the safety and ethicality of LLM-generated content.", "soundness": 1, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well-written, and I really like their figures.  \n2. There is a detailed theoretical transfer and explanation on control engineering and how it can be extended to LLMs.  \n3. This paper provides some inspiration for future non-parametric optimization methods for LLMs.  \n4. To some extent, it achieves an alignment from weak (RoBERTa) to strong (LLaMA).", "weaknesses": "1. The experiments and datasets lack persuasiveness. All experiments in the paper are based on only a few queries, with almost no evaluations on common LLM benchmarks or other metric.  \n2. The approach heavily relies on a Language-Constraint Function (L-CF), implemented with RoBERTa and assumed to be a golden classifier. However, I believe that assuming RoBERTa as an ideal classifier is not reasonable in practical use.", "questions": "1. Why not conduct experiments on at least a small-scale widely-used LLM evaluation datasets, such as a subset of Anthropic/hh-rlhf?  \n2. I feel that finding a reliable Language-Constraint Function (L-CF) is similarly challenging to training a golden reward model. It might be insightful to explore scenarios where the Language-Constraint Function (L-CF) has varying levels of reliability.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730429348327}], "openreview_url": "https://openreview.net/forum?id=fvo6q86NKG", "arxiv_id": "2408.15625", "paper_pdf": "papers/fvo6q86NKG.pdf", "paper_pdf_sha256": "f1b3100272c76f672f0a93be47f34dc94fe7aefab5b3bb1c3597e8bb4b0dd4c9", "paper_pdf_bytes": 486425, "paper_pdf_source": "openreview", "code_url": "https://github.com/Mya-Mya/CBF-LLM", "code_repository": "Mya-Mya/CBF-LLM", "code_commit": "27cc8388f0f7e6f463848c2ecc0ca7989fc8b579", "code_archive": "repos/fvo6q86NKG.zip", "code_archive_sha256": "d84b0d6dd7e378e0029816e21fdaf39b5350d31504eedef9116ce29e21a5165d", "code_archive_bytes": 8167, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 20, "github_languages": {"Python": 13744}, "github_archived": false, "github_pushed_at": "2024-12-10T11:15:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cbf-llm-safe-control-for-llm-alignment"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mOTiVzTgF2", "year": 2024, "status": "rejected", "title": "ResiDual: Transformer with Dual Residual Connections", "authors": ["Shufang Xie", "Huishuai Zhang", "Junliang Guo", "Xu Tan", "Jiang Bian", "Hany Hassan Awadalla", "Arul Menezes", "Tao Qin", "Rui Yan"], "authorids": ["~Shufang_Xie1", "~Huishuai_Zhang3", "~Junliang_Guo1", "~Xu_Tan1", "~Jiang_Bian1", "~Hany_Hassan_Awadalla1", "~Arul_Menezes1", "~Tao_Qin1", "~Rui_Yan2"], "authors_source": "OpenReview API", "abstract": "Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are essential for effective training, is still debated. Two widely used variants are the Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN) Transformers, which apply layer normalization after each residual block's output or before each residual block's input, respectively. While both variants enjoy their advantages, they also suffer from severe limitations: Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity. In this paper, we propose ResiDual, a novel Transformer architecture with Pre-Post-LN (PPLN), which fuses the connections in Post-LN and Pre-LN together, and inherits their advantages while avoids their limitations. We conduct both theoretical analyses and empirical experiments to verify the effectiveness of ResiDual. Theoretically, we prove that ResiDual has a lower bound on the gradient to avoid the vanishing issue due to the residual connection from Pre-LN. Moreover, ResiDual also has diverse model representations to avoid the collapse issue due to the residual connection from Post-LN. Empirically, ResiDual outperforms both Post-LN and Pre-LN on several machine translation benchmarks across different network depths and data sizes.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "s6LAjaFYSq", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5656/Reviewer_q8he"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The submission introduces a novel Transformer architecture aimed at overcoming the gradient vanishing and representation collapse issues found in Post-LN and Pre-LN models, respectively. The paper provides a theoretical analysis to support the architecture's design and presents empirical results from machine translation tasks across several datasets, demonstrating improvements over existing Transformer variants. The approach combines dual residual connections to retain the benefits of both Post-LN and Pre-LN models.", "review_text": "The submission introduces a novel Transformer architecture aimed at overcoming the gradient vanishing and representation collapse issues found in Post-LN and Pre-LN models, respectively. The paper provides a theoretical analysis to support the architecture's design and presents empirical results from machine translation tasks across several datasets, demonstrating improvements over existing Transformer variants. The approach combines dual residual connections to retain the benefits of both Post-LN and Pre-LN models.", "strengths": "+ The paper is well-organized, with a clear introduction to the problem, a detailed methodology, and a presentation of results that make it accessible to readers.\n+ The paper introduces a novel architecture, ResiDual, which creatively combines the benefits of Post-LN and Pre-LN Transformers to address their respective limitations.\n+ The submission includes a thorough theoretical examination of the gradient vanishing and representation collapse problems, providing a solid foundation for the proposed solution.", "weaknesses": "- Experiments might be restricted to machine translation tasks, which does not demonstrate the model's generalizability across different domains or tasks in NLP or other areas where Transformers are applicable.\n- Comparisons with models that have similar enhancements (e.g. RealFormer) to the Transformer architecture, leaving the evaluation incomplete.\n- Missing analysis towards computational costs or training efficiency of the ResiDual model.", "questions": "1. How does the ResiDual model's computational complexity compare to standard Transformer models, particularly in terms of training time and memory requirements?\n2. Can the authors provide additional insights into how the ResiDual model performs on tasks other than machine translation, such as language understanding or speech recognition?\n3. Compare the ResiDual model not only with standard Transformer variants but also with recent SOTA models that address similar issues.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission introduces a novel Transformer architecture aimed at overcoming the gradient vanishing and representation collapse issues found in Post-LN and Pre-LN models, respectively. The paper provides a theoretical analysis to support the architecture's design and presents empirical results from machine translation tasks across several datasets, demonstrating improvements over existing Transformer variants. The approach combines dual residual connections to retain the benefits of both Post-LN and Pre-LN models.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "+ The paper is well-organized, with a clear introduction to the problem, a detailed methodology, and a presentation of results that make it accessible to readers.\n+ The paper introduces a novel architecture, ResiDual, which creatively combines the benefits of Post-LN and Pre-LN Transformers to address their respective limitations.\n+ The submission includes a thorough theoretical examination of the gradient vanishing and representation collapse problems, providing a solid foundation for the proposed solution.", "weaknesses": "- Experiments might be restricted to machine translation tasks, which does not demonstrate the model's generalizability across different domains or tasks in NLP or other areas where Transformers are applicable.\n- Comparisons with models that have similar enhancements (e.g. RealFormer) to the Transformer architecture, leaving the evaluation incomplete.\n- Missing analysis towards computational costs or training efficiency of the ResiDual model.", "questions": "1. How does the ResiDual model's computational complexity compare to standard Transformer models, particularly in terms of training time and memory requirements?\n2. Can the authors provide additional insights into how the ResiDual model performs on tasks other than machine translation, such as language understanding or speech recognition?\n3. Compare the ResiDual model not only with standard Transformer variants but also with recent SOTA models that address similar issues.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699319642011}, {"id": "QrhyReayho", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5656/Reviewer_9QNj"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper focuses on advancing the Transformer architecture by introducing a novel approach called \"ResiDual.\" This approach aims to address the limitations of two widely used variants by incorporating two residual connections to mitigate gradient vanishing and the representation collapse problem. The paper provides both theoretical analysis and empirical results to validate the effectiveness of the proposed model. The study also delves into the performance of the ResiDual model in various settings, including different datasets like WMT and OPUS-100, and its behavior with different learning rate schedules.", "review_text": "The paper focuses on advancing the Transformer architecture by introducing a novel approach called \"ResiDual.\" This approach aims to address the limitations of two widely used variants by incorporating two residual connections to mitigate gradient vanishing and the representation collapse problem. The paper provides both theoretical analysis and empirical results to validate the effectiveness of the proposed model. The study also delves into the performance of the ResiDual model in various settings, including different datasets like WMT and OPUS-100, and its behavior with different learning rate schedules.", "strengths": "About Approach: The introduction of the ResiDual model offers a fresh perspective on addressing the challenges faced by the Transformer architecture.\n\nComprehensive Experiments: The paper provides extensive experimental results on multiple datasets, showcasing the model's robustness and versatility.\n\nPerformance: The ResiDual model show some improvements over other methods.\n\nStability in Training: The research highlights that the ResiDual model does not require learning-rate warm-up for convergence, unlike some other methods.", "weaknesses": "I'm not an expert in NLP and I have limited knowledge on this. \n\nHowever, one of my concern is about the incrementally performance. \n\nAlso, the authors claimed \"Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity\". Thanks for the theoretical analysis in Sec. 2. However, if the authors could provide more empirical evidence to support that?", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper focuses on advancing the Transformer architecture by introducing a novel approach called \"ResiDual.\" This approach aims to address the limitations of two widely used variants by incorporating two residual connections to mitigate gradient vanishing and the representation collapse problem. The paper provides both theoretical analysis and empirical results to validate the effectiveness of the proposed model. The study also delves into the performance of the ResiDual model in various settings, including different datasets like WMT and OPUS-100, and its behavior with different learning rate schedules.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "About Approach: The introduction of the ResiDual model offers a fresh perspective on addressing the challenges faced by the Transformer architecture.\n\nComprehensive Experiments: The paper provides extensive experimental results on multiple datasets, showcasing the model's robustness and versatility.\n\nPerformance: The ResiDual model show some improvements over other methods.\n\nStability in Training: The research highlights that the ResiDual model does not require learning-rate warm-up for convergence, unlike some other methods.", "weaknesses": "I'm not an expert in NLP and I have limited knowledge on this. \n\nHowever, one of my concern is about the incrementally performance. \n\nAlso, the authors claimed \"Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity\". Thanks for the theoretical analysis in Sec. 2. However, if the authors could provide more empirical evidence to support that?", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698902354783}, {"id": "MMkH3N7wIa", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5656/Reviewer_EUoZ"], "rating": "1: strong reject", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors propose a method \"Residual\", where they accumulate the output of all the layers of a Post-LN transformer model, which is then normalized and added to the output of the model. The author's method behaves similarly to a Post-LN model, while avoid the gradient vanishing problem, as the gradient can flow back freely along the accumulating output. The authors show that vanishing gradient problem of Post-LNs may not be solveable by using Adaptive Optimizers such as Adam, as their update rule becomes ill-conditioned for very very low values of gradient. The authors show their method outperforms other prior method across multiple datasets.", "review_text": "The authors propose a method \"Residual\", where they accumulate the output of all the layers of a Post-LN transformer model, which is then normalized and added to the output of the model. The author's method behaves similarly to a Post-LN model, while avoid the gradient vanishing problem, as the gradient can flow back freely along the accumulating output. The authors show that vanishing gradient problem of Post-LNs may not be solveable by using Adaptive Optimizers such as Adam, as their update rule becomes ill-conditioned for very very low values of gradient. The authors show their method outperforms other prior method across multiple datasets.", "strengths": "1. The authors method avoid the \"representation collapse\" issue, where the representation changes less for each new layer in Pre-LN.\n2. The authors compare their method on multiple datasets", "weaknesses": "1. The treatment of grad norm of Pre-LN seems to be sometimes empirically incorrect (see Questions for authors), with wildly different observations in reality compared to the author's theoretical treatment\n1. Their derivations of theorem 3.1 ignore the impact of non-linearity on the gradient of the transformer MLP layer, instead assuming it to be a single FC. \n1. The derivation also ignores back-propagated gradient through the Query - while Keys will not back-propagate any gradient as Query is zero-initialized, Query will have non-zero gradient back-propagating.\n1. As the improvements can sometimes be small (For eg. 27.65 vs 27.3 in Table 3), some measure of statistical significance of the improvements is required.\n1. Discussion of prior work is somewhat lacking - while the authors reference prior works throughout the text, a dedicated section for more detailed discussion is missing.\n1. The discussion of Adam condition number assumes very very small values of gradient, which is not observed realistically. This discussion should be contextualized given realistic values of the gradient.", "questions": "1. Could the authors provide code/provide exact steps to reproduce figure 2(a), specifically the gradient-norm of Pre-LN? I tried to reproduce this figure, and I am failing. See details below on the exact code I used. The plot I observed does not look anything like what Equation(2) would suggest - it seems perhaps exponential for shallow layers, but definitely not logarithmic.\n2. For experiments in Table 2, 3, and 4, did the authors only run a single run, or were the experiments repeated multiple times with varying hyper-parameters? How was this hyper-parameter search performed? The dropout and LR (from Table 9,10, 11) are different for these tables.\n    1. So if a hyper-parameter search was performed, was the same search performed for all the baselines?\n3. Figure 3 in the appendix shows the condition number upto $\\sigma_g$ upto $10^{-7}$ - what is the realistic range of $\\sigma_g$? On experimenting with standard BERT-large model, I found $\\sigma_g$ at initialization was < $10^{-4}$ for 0.3% of params, and even < $10^{-3}$ for 4% of params. How is the Adam condition number at these (somewhat more-realistic) lower values of  $\\sigma_g$, and how does Figure 3(left) change?\n\n\n\nCode for gradient norm of Pre-LN - \nUsing a sample LM code from huggingface from [here](https://github.com/huggingface/transformers/blob/v4.31.0/examples/pytorch/language-modeling/run_clm_no_trainer.py), I added ` config.n_layer = 48` at line 390 to increase the number of layers. And then I ran the command below - \n```\npython run_clm_no_trainer.py  --model_type gpt2 --tokenizer_name gpt2 --dataset_name wikitext  --dataset_config_name wikitext-2-raw-v1 --per_device_train_batch_size 2 --output_dir temp --seed 1234\n```\nAfter 1 backward iteration, I plot the grad norm of parameters.\n\n\nMinor typos (the authors are not expected to respond to these) - \n1. Cite published version of Liu et al 2020\n1. \"So does the gradients of $w_k$ in section 2.1 -> so do\n1. \"Gradient vanish issue\" -> vanishing\n1. \"the both disadvantages\"\n1. \"non of these models\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a method \"Residual\", where they accumulate the output of all the layers of a Post-LN transformer model, which is then normalized and added to the output of the model. The author's method behaves similarly to a Post-LN model, while avoid the gradient vanishing problem, as the gradient can flow back freely along the accumulating output. The authors show that vanishing gradient problem of Post-LNs may not be solveable by using Adaptive Optimizers such as Adam, as their update rule becomes ill-conditioned for very very low values of gradient. The authors show their method outperforms other prior method across multiple datasets.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "strengths": "1. The authors method avoid the \"representation collapse\" issue, where the representation changes less for each new layer in Pre-LN.\n2. The authors compare their method on multiple datasets", "weaknesses": "1. The treatment of grad norm of Pre-LN seems to be sometimes empirically incorrect (see Questions for authors), with wildly different observations in reality compared to the author's theoretical treatment\n1. Their derivations of theorem 3.1 ignore the impact of non-linearity on the gradient of the transformer MLP layer, instead assuming it to be a single FC. \n1. The derivation also ignores back-propagated gradient through the Query - while Keys will not back-propagate any gradient as Query is zero-initialized, Query will have non-zero gradient back-propagating.\n1. As the improvements can sometimes be small (For eg. 27.65 vs 27.3 in Table 3), some measure of statistical significance of the improvements is required.\n1. Discussion of prior work is somewhat lacking - while the authors reference prior works throughout the text, a dedicated section for more detailed discussion is missing.\n1. The discussion of Adam condition number assumes very very small values of gradient, which is not observed realistically. This discussion should be contextualized given realistic values of the gradient.", "questions": "1. Could the authors provide code/provide exact steps to reproduce figure 2(a), specifically the gradient-norm of Pre-LN? I tried to reproduce this figure, and I am failing. See details below on the exact code I used. The plot I observed does not look anything like what Equation(2) would suggest - it seems perhaps exponential for shallow layers, but definitely not logarithmic.\n2. For experiments in Table 2, 3, and 4, did the authors only run a single run, or were the experiments repeated multiple times with varying hyper-parameters? How was this hyper-parameter search performed? The dropout and LR (from Table 9,10, 11) are different for these tables.\n    1. So if a hyper-parameter search was performed, was the same search performed for all the baselines?\n3. Figure 3 in the appendix shows the condition number upto $\\sigma_g$ upto $10^{-7}$ - what is the realistic range of $\\sigma_g$? On experimenting with standard BERT-large model, I found $\\sigma_g$ at initialization was < $10^{-4}$ for 0.3% of params, and even < $10^{-3}$ for 4% of params. How is the Adam condition number at these (somewhat more-realistic) lower values of  $\\sigma_g$, and how does Figure 3(left) change?\n\n\n\nCode for gradient norm of Pre-LN - \nUsing a sample LM code from huggingface from [here](https://github.com/huggingface/transformers/blob/v4.31.0/examples/pytorch/language-modeling/run_clm_no_trainer.py), I added ` config.n_layer = 48` at line 390 to increase the number of layers. And then I ran the command below - \n```\npython run_clm_no_trainer.py  --model_type gpt2 --tokenizer_name gpt2 --dataset_name wikitext  --dataset_config_name wikitext-2-raw-v1 --per_device_train_batch_size 2 --output_dir temp --seed 1234\n```\nAfter 1 backward iteration, I plot the grad norm of parameters.\n\n\nMinor typos (the authors are not expected to respond to these) - \n1. Cite published version of Liu et al 2020\n1. \"So does the gradients of $w_k$ in section 2.1 -> so do\n1. \"Gradient vanish issue\" -> vanishing\n1. \"the both disadvantages\"\n1. \"non of these models\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698688566830}, {"id": "Blq4vyvOkZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5656/Reviewer_ygjB"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper points out the shortcomings of the Post-LN and Pre-LN Transformer architectures. Post-LN suffers from the gradient vanishing problem, while Pre-LN has the representation collapse issue. To address the problems of the aforementioned architectures, the paper introduces a new architecture called ResiDual. This architecture combines the advantages of Post-LN and Pre-LN and attempts to avoid their drawbacks. Subsequently, a series of experiments are conducted based on this model.", "review_text": "This paper points out the shortcomings of the Post-LN and Pre-LN Transformer architectures. Post-LN suffers from the gradient vanishing problem, while Pre-LN has the representation collapse issue. To address the problems of the aforementioned architectures, the paper introduces a new architecture called ResiDual. This architecture combines the advantages of Post-LN and Pre-LN and attempts to avoid their drawbacks. Subsequently, a series of experiments are conducted based on this model.", "strengths": "(1) The research focus of this paper is very clear, and the approach is well-organized. \n(2) The ResiDual model proposed in the paper indeed improves upon the combination of Post-LN and Pre-LN and achieves good results on some datasets.", "weaknesses": "(1) Based on the empirically observed results, the computational cost increased by 3%. Is there sufficient experimental data to support this? I didn't see relevant information in the text or the appendix.\n(2) Your model seems to be prone to overfitting. The results on the IWSLT and WMT datasets show that the depth of your model is always limited, which to some extent reduces the potential for performance improvement of the model itself.\n(3) When studying learning rate warm-up, you shouldn't use small datasets. Switch to a larger dataset and a deeper network, and your performance improvement will be more convincing.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper points out the shortcomings of the Post-LN and Pre-LN Transformer architectures. Post-LN suffers from the gradient vanishing problem, while Pre-LN has the representation collapse issue. To address the problems of the aforementioned architectures, the paper introduces a new architecture called ResiDual. This architecture combines the advantages of Post-LN and Pre-LN and attempts to avoid their drawbacks. Subsequently, a series of experiments are conducted based on this model.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "(1) The research focus of this paper is very clear, and the approach is well-organized. \n(2) The ResiDual model proposed in the paper indeed improves upon the combination of Post-LN and Pre-LN and achieves good results on some datasets.", "weaknesses": "(1) Based on the empirically observed results, the computational cost increased by 3%. Is there sufficient experimental data to support this? I didn't see relevant information in the text or the appendix.\n(2) Your model seems to be prone to overfitting. The results on the IWSLT and WMT datasets show that the depth of your model is always limited, which to some extent reduces the potential for performance improvement of the model itself.\n(3) When studying learning rate warm-up, you shouldn't use small datasets. Switch to a larger dataset and a deeper network, and your performance improvement will be more convincing.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698655284054}, {"id": "OSh8X1R6jj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5656/Reviewer_e3RM"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a novel Transformer architecture named ResiDual to explore the optimal way implementing residual connections in Transformer. There are two widely used methods in the literature: Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN). However, they suffer from either the gradient vanishing issue or the collapse issue. The paper combines the two LN ideas and designs an architecture with Pre-Post-Layer-Normalization. Basically, the new proposed architecture has two parallel lines, one for Post-LN and the other for Pre-LN. The two lines share the same self-attention block. The authors claim that this architecture can avoid the aforementioned limitations. They first give a theoretical analysis, and then conduct experiments on several data sets. They show that the new architecture can outperform baseline methods on benchmarks, like IWSLT-14, WMT, and OPUS-100.", "review_text": "The paper proposes a novel Transformer architecture named ResiDual to explore the optimal way implementing residual connections in Transformer. There are two widely used methods in the literature: Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN). However, they suffer from either the gradient vanishing issue or the collapse issue. The paper combines the two LN ideas and designs an architecture with Pre-Post-Layer-Normalization. Basically, the new proposed architecture has two parallel lines, one for Post-LN and the other for Pre-LN. The two lines share the same self-attention block. The authors claim that this architecture can avoid the aforementioned limitations. They first give a theoretical analysis, and then conduct experiments on several data sets. They show that the new architecture can outperform baseline methods on benchmarks, like IWSLT-14, WMT, and OPUS-100.", "strengths": "- How to implement residual connections in Transformer optimally is a very important topic in deep learning. The paper makes contributions in this direction, which, in my opinion, should be of interest to the community.\n\n- The paper is overall well-written. The basic idea is clean and easy to understand.", "weaknesses": "- The experimental results are a little bit unconvincing. The experiments section did not state the number of algorithm runs used to compute each reported result. Additionally, it is not clear that the results present in the tables are the best result among several runs or the average.\n\n- Only two settings (E6D6 and E12D12) are considered in the experiments. It will be nice to have a figure about how the algorithms' performances vary as the number of layers increases.", "questions": "- (1)  Are the experimental results the best among several runs or the average?  If the average, what's the standard deviation?\n\n- (2) Have you tried the structure that simply lets all odd blocks be Post-LN and all even blocks be Pre-LN (or reversed)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel Transformer architecture named ResiDual to explore the optimal way implementing residual connections in Transformer. There are two widely used methods in the literature: Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN). However, they suffer from either the gradient vanishing issue or the collapse issue. The paper combines the two LN ideas and designs an architecture with Pre-Post-Layer-Normalization. Basically, the new proposed architecture has two parallel lines, one for Post-LN and the other for Pre-LN. The two lines share the same self-attention block. The authors claim that this architecture can avoid the aforementioned limitations. They first give a theoretical analysis, and then conduct experiments on several data sets. They show that the new architecture can outperform baseline methods on benchmarks, like IWSLT-14, WMT, and OPUS-100.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- How to implement residual connections in Transformer optimally is a very important topic in deep learning. The paper makes contributions in this direction, which, in my opinion, should be of interest to the community.\n\n- The paper is overall well-written. The basic idea is clean and easy to understand.", "weaknesses": "- The experimental results are a little bit unconvincing. The experiments section did not state the number of algorithm runs used to compute each reported result. Additionally, it is not clear that the results present in the tables are the best result among several runs or the average.\n\n- Only two settings (E6D6 and E12D12) are considered in the experiments. It will be nice to have a figure about how the algorithms' performances vary as the number of layers increases.", "questions": "- (1)  Are the experimental results the best among several runs or the average?  If the average, what's the standard deviation?\n\n- (2) Have you tried the structure that simply lets all odd blocks be Post-LN and all even blocks be Pre-LN (or reversed)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698575589818}], "openreview_url": "https://openreview.net/forum?id=mOTiVzTgF2", "arxiv_id": "2304.14802", "paper_pdf": "papers/mOTiVzTgF2.pdf", "paper_pdf_sha256": "b74039d1cd2b681315278bc56f8d4dfc03d1617ebc4765548fc803e82aaa5b7b", "paper_pdf_bytes": 607065, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/ResiDual", "code_repository": "microsoft/ResiDual", "code_commit": "8682f7510be105a0caf5b98a021e712f44f90ace", "code_archive": "repos/mOTiVzTgF2.zip", "code_archive_sha256": "cec65c73f8e80c77f8450dcd56aefe54692dbfcb4702ca6c61522f17e9994aa7", "code_archive_bytes": 15699, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 17, "github_languages": {"Python": 20837}, "github_archived": true, "github_pushed_at": "2023-08-18T18:23:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/residual-transformer-with-dual-residual"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4nrZXPFN1c4", "year": 2023, "status": "rejected", "title": "Energy Transformer", "authors": ["Benjamin Hoover", "Yuchen Liang", "Bao Pham", "Rameswar Panda", "Hendrik Strobelt", "Duen Horng Chau", "Mohammed J Zaki", "Dmitry Krotov"], "authorids": ["~Benjamin_Hoover1", "~Yuchen_Liang2", "~Bao_Pham1", "~Rameswar_Panda1", "~Hendrik_Strobelt1", "~Duen_Horng_Chau1", "~Mohammed_J_Zaki1", "~Dmitry_Krotov2"], "authors_source": "OpenReview API", "abstract": "Transformers have become the de facto  models of choice in machine learning, typically leading to impressive performance on many applications. At the same time, the architectural development in the transformer world is mostly driven by empirical findings, and  the theoretical understanding of their architectural building blocks is rather limited. In contrast, Dense Associative Memory models or Modern Hopfield Networks have a well-established theoretical foundation, but have not yet demonstrated truly impressive practical results. We propose a transformer architecture that replaces the sequence of feedforward transformer blocks with a single large Associative Memory model. Our novel architecture, called Energy Transformer (or ET for short), has many of the familiar architectural primitives that are often used in the current generation of transformers. However, it is not identical to the existing architectures. The sequence of transformer layers in ET is purposely designed to minimize a specifically engineered energy function, which is responsible for representing the relationships between the tokens. As a consequence of this computational principle, the attention in ET is different from the conventional attention mechanism.  In this work, we introduce the theoretical foundations of ET, explore it's empirical capabilities using the image completion task, and obtain strong quantitative results on the graph anomaly detection task.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "TghgcpJcrYw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2834/Reviewer_iuLg"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes the energy transformer, a transformer architecture that uses a recurrent energy transformer block. The energy transformer block updates its input in accordance to minimizing two energy functions. Experiments on image reconstruction and graph anomaly detection demonstrate the effectiveness of the proposed method. ", "review_text": "Although the paper is interesting to read, I have two major concerns as detailed in the weakness section. ", "strengths": "Strength: The proposed method is interesting in the sense that it reformulates transformer architecture in a way to combine Hopefield networks and attention networks, which are recurrently optimized using two carefully designed energy functions. The empirical results look promising. \n\nWeakness: I am concerned about the clarity of the paper. For example, in the explanation of equation (3), \"The log-sum energy function (3) is minimal when for every patch in the image its queries are aligned with the keys of a small number of other patches connected by the attention map,\" It is not clear what does this mean in a mathematical sense. There is also no detailed algorithmic description about how the model is trained. For example, for the image reconstruction task, how does minimizing the energy functions correspond to reconstruction? It would be helpful if the authors include a formal algorithmic description of the proposed method. \n\nI am also concerned about some of the experiment settings. For example, for image reconstruction task, Figure 3 only shows some of the examples. However, this does not represent its general performance. One way to make this experiment more complete may be to include the MSE of the test datasets, the variance of the MSE, and perhaps reconstructed images corresponding to the best MSE and the worst MSE. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes the energy transformer, a transformer architecture that uses a recurrent energy transformer block. The energy transformer block updates its input in accordance to minimizing two energy functions. Experiments on image reconstruction and graph anomaly detection demonstrate the effectiveness of the proposed method. ", "strength_and_weaknesses": "Strength: The proposed method is interesting in the sense that it reformulates transformer architecture in a way to combine Hopefield networks and attention networks, which are recurrently optimized using two carefully designed energy functions. The empirical results look promising. \n\nWeakness: I am concerned about the clarity of the paper. For example, in the explanation of equation (3), \"The log-sum energy function (3) is minimal when for every patch in the image its queries are aligned with the keys of a small number of other patches connected by the attention map,\" It is not clear what does this mean in a mathematical sense. There is also no detailed algorithmic description about how the model is trained. For example, for the image reconstruction task, how does minimizing the energy functions correspond to reconstruction? It would be helpful if the authors include a formal algorithmic description of the proposed method. \n\nI am also concerned about some of the experiment settings. For example, for image reconstruction task, Figure 3 only shows some of the examples. However, this does not represent its general performance. One way to make this experiment more complete may be to include the MSE of the test datasets, the variance of the MSE, and perhaps reconstructed images corresponding to the best MSE and the worst MSE. \n", "clarity,_quality,_novelty_and_reproducibility": "Although the proposed method seems novel, the methodology of the proposed transformer architecture is not clear enough to me. The experimental results are also not very satisfying, as detailed in the previous section. ", "summary_of_the_review": "Although the paper is interesting to read, I have two major concerns as detailed in the weakness section. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667012008657}, {"id": "DxFrdU9TsGw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2834/Reviewer_zHgn"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new transformer architecture that replaces the sequence of transformer blocks with a single associative memory model. Specifically, the model is designed to minimize a specific energy function that represents the relations between tokens. The model is evaluated on image completion and graph anomaly detection, and experiments show encouraging results.", "review_text": "It is an interesting work toward understanding the Transformer models. To show its effectiveness, more experiments have to be conducted on more challenging tasks.", "strengths": "Strength\n- Despite the great performance achieved by the Transformer model, its architecture is designed empirically. This work proposes a Transformer architecture from a purely theoretical perspective.\n- This work provides a deep insight into the relationship between Transforms and associative memory models. Based on the relationship, it designs a new energy function and a corresponding Transformer architecture that minimizes the energy function.\n\nWeakness\n- To show the effectiveness of ET, it is better to evaluate it on more mainstream tasks.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a new transformer architecture that replaces the sequence of transformer blocks with a single associative memory model. Specifically, the model is designed to minimize a specific energy function that represents the relations between tokens. The model is evaluated on image completion and graph anomaly detection, and experiments show encouraging results.", "strength_and_weaknesses": "Strength\n- Despite the great performance achieved by the Transformer model, its architecture is designed empirically. This work proposes a Transformer architecture from a purely theoretical perspective.\n- This work provides a deep insight into the relationship between Transforms and associative memory models. Based on the relationship, it designs a new energy function and a corresponding Transformer architecture that minimizes the energy function.\n\nWeakness\n- To show the effectiveness of ET, it is better to evaluate it on more mainstream tasks.", "clarity,_quality,_novelty_and_reproducibility": "good\n", "summary_of_the_review": "It is an interesting work toward understanding the Transformer models. To show its effectiveness, more experiments have to be conducted on more challenging tasks.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666688698264}, {"id": "PueTw2UFKAr", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2834/Reviewer_UVr7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "\nThe paper puts forward a novel sequence-to-sequence neural architecture named Energy Transformer (ET) that builds\non recently introduced generalized Hopfield networks and their connection to transformer layers. \nThe architecture is composed by a single block that combines two subsystems,\ndefined by energy functions corresponding to two different types of generalized hopfield networks, \npreceded by a layer normalization operator. \nOne subsystem is in charge of pooling to each position the information coming from the other\npositions, while the other enforces the consistency of learned representations.\nThe system dynamics is given by a differential equation driven by the gradient\nof the energy function with respect to the renormalized features. The total\nenergy is proven to decrease along the trajectory. \nFor practical purposes, the dynamics is discretized and the block applied recurrently.\nThe architecture is validated qualitatively on an image occlusion task and quantitatively\non a node anomaly detection task.", "review_text": "I think this is a good paper that deserves acceptance.\n", "strengths": "\nThe manuscript is an enjoyable read and contains a good degree of novelty. \nThe output of the network is given by the final outcome of an energy minimization process.\nAlthough this perspective is not new (see Yang et al '22), it has not been quite explored yet\nand I find the particular instantiation given in the paper quite inspiring. \nI think this could trigger additional theoretical and applicative results.\nThe resulting architecture is also quite simple and has a good degree of interpretability.\nThe paper is well organized, although some details given in the appendix could be moved to the main text.\nThe results on the node outlier detection tasks are very good, outperforming concurrent methods in\nmost cases.\n\nMain comments:\n\n- Why the authors didn't attempt to benchmark the performance on a standard image classification task? \n  This should be doable by pooling and putting an head on top of the final outputs, as commonly done in ViT.\n\n- Related to previous questions, is not clear if the ET can be considered as a drop-in replacement to self-attention\n  transformer blocks for arbitrary task, or if its usage should be confined to some specific settings.\n  I would like the authors to comment on the limitations of the model.\n\n- It is not clear how important the HN component is and how would perform the ET without it.\n\n- Could the author motivate why the attention module excludes the current the considere position from the input. \nWhat happens if you remove the restriction from the sum?\n\n- Did the authors explor stacking multiple ET modules in a feeforward fashion?\n\nMinor comments:\n- While it is clear in the appendix, I think also the main text should mention that the DE is discretized.\n\n- eq. 2, ignored the \\bar{x} dependence on x_i when computing the derivative. Should be fixed by inserting a factor 1 / (1 - 1/D) in the definition of L\n\n- \"is then minimized to train the whole network. Above, σ is the ratio of anomalous labels (lA = 1) to\nthe regular label.\" Maybe the other way around, that is regular labels / anomalous labels?\n\n- The main text should mention that for the graph task a neighborhood softmax (similar to GAT) is used instead of standard attention\n\n- The authors could consider numbering more equations for readibility.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\nThe paper puts forward a novel sequence-to-sequence neural architecture named Energy Transformer (ET) that builds\non recently introduced generalized Hopfield networks and their connection to transformer layers. \nThe architecture is composed by a single block that combines two subsystems,\ndefined by energy functions corresponding to two different types of generalized hopfield networks, \npreceded by a layer normalization operator. \nOne subsystem is in charge of pooling to each position the information coming from the other\npositions, while the other enforces the consistency of learned representations.\nThe system dynamics is given by a differential equation driven by the gradient\nof the energy function with respect to the renormalized features. The total\nenergy is proven to decrease along the trajectory. \nFor practical purposes, the dynamics is discretized and the block applied recurrently.\nThe architecture is validated qualitatively on an image occlusion task and quantitatively\non a node anomaly detection task.", "strength_and_weaknesses": "\nThe manuscript is an enjoyable read and contains a good degree of novelty. \nThe output of the network is given by the final outcome of an energy minimization process.\nAlthough this perspective is not new (see Yang et al '22), it has not been quite explored yet\nand I find the particular instantiation given in the paper quite inspiring. \nI think this could trigger additional theoretical and applicative results.\nThe resulting architecture is also quite simple and has a good degree of interpretability.\nThe paper is well organized, although some details given in the appendix could be moved to the main text.\nThe results on the node outlier detection tasks are very good, outperforming concurrent methods in\nmost cases.\n\nMain comments:\n\n- Why the authors didn't attempt to benchmark the performance on a standard image classification task? \n  This should be doable by pooling and putting an head on top of the final outputs, as commonly done in ViT.\n\n- Related to previous questions, is not clear if the ET can be considered as a drop-in replacement to self-attention\n  transformer blocks for arbitrary task, or if its usage should be confined to some specific settings.\n  I would like the authors to comment on the limitations of the model.\n\n- It is not clear how important the HN component is and how would perform the ET without it.\n\n- Could the author motivate why the attention module excludes the current the considere position from the input. \nWhat happens if you remove the restriction from the sum?\n\n- Did the authors explor stacking multiple ET modules in a feeforward fashion?\n\nMinor comments:\n- While it is clear in the appendix, I think also the main text should mention that the DE is discretized.\n\n- eq. 2, ignored the \\bar{x} dependence on x_i when computing the derivative. Should be fixed by inserting a factor 1 / (1 - 1/D) in the definition of L\n\n- \"is then minimized to train the whole network. Above, σ is the ratio of anomalous labels (lA = 1) to\nthe regular label.\" Maybe the other way around, that is regular labels / anomalous labels?\n\n- The main text should mention that for the graph task a neighborhood softmax (similar to GAT) is used instead of standard attention\n\n- The authors could consider numbering more equations for readibility.", "clarity,_quality,_novelty_and_reproducibility": "The manuscript is overall quite clear and contains novel and interesting results.\nThe experimental settings are fully detailed and the code will be publicly released.", "summary_of_the_review": "I think this is a good paper that deserves acceptance.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666651714530}, {"id": "fJIrZsXf66p", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2834/Reviewer_v3Fm"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose a new transformer architecture called energy transformer (ET). ET is designed to minimize a special handcrafted energy function. This function contains two terms. The attention term aligns the keys and queries between neighbor token vectors and the hopefield network term matches the token vectors to some memory vectors. The ET layer is applied recurrently to imitate a continuous time differential equation that minized the designed energy terms. The authors try to demonstrate the ability of their model through image completion task and graph anomaly detection task.", "review_text": "I think this paper proposes some interesting ideas like designing a energy function and deriving a transformer structure based on this energy function. However, I think the experiments of this paper may not be strong enough to demonstrate that the new designed structure does have advantage over the original one. ", "strengths": "Strength:\nIn general, I like the idea of this paper:\n1. This paper propose an specially designed energy-term that may be helpful for understanding the transformer structure.\n2. It derives a new transformer structure based on this handcrafted energy-term\n3. The authors show that this structure outperforms baselines in the graph anomaly detection task.\n\nPossible concerns:\n1. While the authors showcase experiments on image completion and graph anomaly detection task. I think these are not enough for me to draw the conclusion that this newly designed arctecture does have advantage over the previous one. The transformer structure is famous because it performs well in various task in computer vision(CV) and netural language processing(NLP). To really demonstrate the ability of a new structure, the authors need to show competitive results on those widely accepted tasks like transfer learning to ImageNet in CV or results on GLUE [1] benchmark in NLP. Currently, the authors only showcase two tasks that are well-suited for their model (and they only show quantative results in one of them), thus the general ability of this model is still unclear to me.\n\n2. Some designs of the model are not fully justified experimentally. For example, unlike traditional transformer architecture which uses different weights for different layers, ET apply a single block recurrently and the MLPs in their hopfiled network also share weights . The authors may find theoretical explanation for this design according to their energy design. However, whether this helps or hurts the performance is unclear. Empirically speaking, recurrent model or model with shared weight may be hard to optimize due to the vanishing gradient problem. Thus, ablation study may be needed to justify the these design choices.\n\n3. Questions for the graph anomaly detection task: How are the edges used by the model? Does the attention layer only compute attention over the nodes that have edges between them? Also, can the model be also used on directed graph? Or it is only limited to undirected graph?\n\n4. For the image completion task, the authors assumes the image patches can be represented as fully masked and fully open area. But what if the masked area is not divisible by the patches? Then do you need to re-design the patch size (which may influence the full structure of the model)? This might not be a problem if one just treat image completion as a self-supervised task to learn representations and then applied these learned features to downstream tasks. But if the image completion task itself is the target and its performance is used to showcase the ability of the model. Then possible limitation may need to be considered.  \n\n[1] GLUE: A MULTI-TASK BENCHMARK AND ANALYSIS PLATFORM FOR NATURAL LANGUAGE UNDERSTANDING\n\n=====================================================================================================\n\n**Post rebuttal:**\n\nFor me, although in the new responses, the authors say that they are not trying to replace the original transformer architecture, what the paper mainly about is proposing a new transformer architecture based on modification of the original one. Thus, a natural question to ask is how this new design performs comparing with the original one. Although the authors show that their design principle is based on the Hopfield Networks, whether this principle does benefit the transformer architecture design needs further justification. There can be many interesting principles in modifying the designs, but whether they do make senses need to be supported by comprehensive and convincing experimental results. A valid modification to the transformer architecture needs to demonstrate its advantage over the original one. The current results, however, are far from that. The current results are mainly based in the graph domain. Solid experimental results in areas that transformers are mainly used in, like CV or NLP, are missing. I can not agree with the authors that since they expertise in graph domain, they only pick experiments in this area and miss others. As I said, if the paper narrows its objective into proposing a new technique to achieve STOA results in graph domain, then the current results are fair enough. However, if they want to make a valid and general design principle for transformer architecture, they should compare their new architecture in the areas that transformer architecture are widely used. As a reviewer, I'm not convinced by the current results and thus I still lean to reject. I would recommend the authors to polish their work with more convincing results in different areas or limit their objective to a technique that specialized in graph domain.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors propose a new transformer architecture called energy transformer (ET). ET is designed to minimize a special handcrafted energy function. This function contains two terms. The attention term aligns the keys and queries between neighbor token vectors and the hopefield network term matches the token vectors to some memory vectors. The ET layer is applied recurrently to imitate a continuous time differential equation that minized the designed energy terms. The authors try to demonstrate the ability of their model through image completion task and graph anomaly detection task.", "strength_and_weaknesses": "Strength:\nIn general, I like the idea of this paper:\n1. This paper propose an specially designed energy-term that may be helpful for understanding the transformer structure.\n2. It derives a new transformer structure based on this handcrafted energy-term\n3. The authors show that this structure outperforms baselines in the graph anomaly detection task.\n\nPossible concerns:\n1. While the authors showcase experiments on image completion and graph anomaly detection task. I think these are not enough for me to draw the conclusion that this newly designed arctecture does have advantage over the previous one. The transformer structure is famous because it performs well in various task in computer vision(CV) and netural language processing(NLP). To really demonstrate the ability of a new structure, the authors need to show competitive results on those widely accepted tasks like transfer learning to ImageNet in CV or results on GLUE [1] benchmark in NLP. Currently, the authors only showcase two tasks that are well-suited for their model (and they only show quantative results in one of them), thus the general ability of this model is still unclear to me.\n\n2. Some designs of the model are not fully justified experimentally. For example, unlike traditional transformer architecture which uses different weights for different layers, ET apply a single block recurrently and the MLPs in their hopfiled network also share weights . The authors may find theoretical explanation for this design according to their energy design. However, whether this helps or hurts the performance is unclear. Empirically speaking, recurrent model or model with shared weight may be hard to optimize due to the vanishing gradient problem. Thus, ablation study may be needed to justify the these design choices.\n\n3. Questions for the graph anomaly detection task: How are the edges used by the model? Does the attention layer only compute attention over the nodes that have edges between them? Also, can the model be also used on directed graph? Or it is only limited to undirected graph?\n\n4. For the image completion task, the authors assumes the image patches can be represented as fully masked and fully open area. But what if the masked area is not divisible by the patches? Then do you need to re-design the patch size (which may influence the full structure of the model)? This might not be a problem if one just treat image completion as a self-supervised task to learn representations and then applied these learned features to downstream tasks. But if the image completion task itself is the target and its performance is used to showcase the ability of the model. Then possible limitation may need to be considered.  \n\n[1] GLUE: A MULTI-TASK BENCHMARK AND ANALYSIS PLATFORM FOR NATURAL LANGUAGE UNDERSTANDING\n\n=====================================================================================================\n\n**Post rebuttal:**\n\nFor me, although in the new responses, the authors say that they are not trying to replace the original transformer architecture, what the paper mainly about is proposing a new transformer architecture based on modification of the original one. Thus, a natural question to ask is how this new design performs comparing with the original one. Although the authors show that their design principle is based on the Hopfield Networks, whether this principle does benefit the transformer architecture design needs further justification. There can be many interesting principles in modifying the designs, but whether they do make senses need to be supported by comprehensive and convincing experimental results. A valid modification to the transformer architecture needs to demonstrate its advantage over the original one. The current results, however, are far from that. The current results are mainly based in the graph domain. Solid experimental results in areas that transformers are mainly used in, like CV or NLP, are missing. I can not agree with the authors that since they expertise in graph domain, they only pick experiments in this area and miss others. As I said, if the paper narrows its objective into proposing a new technique to achieve STOA results in graph domain, then the current results are fair enough. However, if they want to make a valid and general design principle for transformer architecture, they should compare their new architecture in the areas that transformer architecture are widely used. As a reviewer, I'm not convinced by the current results and thus I still lean to reject. I would recommend the authors to polish their work with more convincing results in different areas or limit their objective to a technique that specialized in graph domain.", "clarity,_quality,_novelty_and_reproducibility": "This paper is in general well written and introduce its idea clearly. ", "summary_of_the_review": "I think this paper proposes some interesting ideas like designing a energy function and deriving a transformer structure based on this energy function. However, I think the experiments of this paper may not be strong enough to demonstrate that the new designed structure does have advantage over the original one. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666646707909}, {"id": "pjD3Ze3n0i", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2834/Reviewer_VTZR"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a transformer architecture called Energy Transformer that replaces the sequence of feedforward transformer blocks with a single large Associative Memory model. The sequence of the energy transformer layers is designed to minimize an energy function in charge of representing the relationships between the tokens. Experiments show that Energy Transformer achieves good performances on image completion and graph anomaly detection.  ", "review_text": "The primary concern of this paper is its weak experiments, which do not sufficiently demonstrate the strength of the proposed Energy Transformer.", "strengths": "[Strengths]\n+ The transformer is a prominent attention mechanism in machine learning, and it is interesting to design the Transformer mechanism from energy attention and the Hopfield network.\n\n[Weaknesses]\n1) The connection between Hopfield Networks, energy function, and Transformer is mentioned in Ramsauer et al. (2020). It is not clear the new insight of the proposed Energy Transformer.\n2) The effect of designing the proposed ET block is not apparent; it is better to conduct an ablation study experiment to assess each component within such a designed block.\n3) The detailed structure of the Energy Transformer is not explicitly depicted. A network architecture helps the reader to understand the ET implementation.\n4) Since the Energy Transformer requires recurrently updating, it is better to show the cost of training and testing compared with the vanilla Transformer.\n5) The experiments are weak. The quantitative result has merely the graph anomaly detection. However, the results of BWGNN (Hetero) and [A], which shows good performance, are not included. The on-par performance of the Energy Transformer does not make it impressive to the reader.\n\nRelated paper:\n[A] Derek Lim, Xiuyu Li, Felix Hohne, and Ser-Nam Lim: New Benchmarks for Learning on Non-Homophilous Graphs. Workshop on Graph Learning Benchmarks, WWW 2021.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a transformer architecture called Energy Transformer that replaces the sequence of feedforward transformer blocks with a single large Associative Memory model. The sequence of the energy transformer layers is designed to minimize an energy function in charge of representing the relationships between the tokens. Experiments show that Energy Transformer achieves good performances on image completion and graph anomaly detection.  ", "strength_and_weaknesses": "[Strengths]\n+ The transformer is a prominent attention mechanism in machine learning, and it is interesting to design the Transformer mechanism from energy attention and the Hopfield network.\n\n[Weaknesses]\n1) The connection between Hopfield Networks, energy function, and Transformer is mentioned in Ramsauer et al. (2020). It is not clear the new insight of the proposed Energy Transformer.\n2) The effect of designing the proposed ET block is not apparent; it is better to conduct an ablation study experiment to assess each component within such a designed block.\n3) The detailed structure of the Energy Transformer is not explicitly depicted. A network architecture helps the reader to understand the ET implementation.\n4) Since the Energy Transformer requires recurrently updating, it is better to show the cost of training and testing compared with the vanilla Transformer.\n5) The experiments are weak. The quantitative result has merely the graph anomaly detection. However, the results of BWGNN (Hetero) and [A], which shows good performance, are not included. The on-par performance of the Energy Transformer does not make it impressive to the reader.\n\nRelated paper:\n[A] Derek Lim, Xiuyu Li, Felix Hohne, and Ser-Nam Lim: New Benchmarks for Learning on Non-Homophilous Graphs. Workshop on Graph Learning Benchmarks, WWW 2021.\n", "clarity,_quality,_novelty_and_reproducibility": "Unclear descriptions reduce clarity and reproducibility.\n\nTechnical novelty should be clarified.", "summary_of_the_review": "The primary concern of this paper is its weak experiments, which do not sufficiently demonstrate the strength of the proposed Energy Transformer.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666627159308}], "openreview_url": "https://openreview.net/forum?id=4nrZXPFN1c4", "arxiv_id": "2302.07253", "paper_pdf": "papers/4nrZXPFN1c4.pdf", "paper_pdf_sha256": "4583760c02fa5b8a3975c53b295601c92b33a05e9c07e98fad9260c38a7d17cc", "paper_pdf_bytes": 19682945, "paper_pdf_source": "openreview", "code_url": "https://github.com/zhuergou/Energy-Transformer-for-Graph-Anomaly-Detection", "code_repository": "zhuergou/Energy-Transformer-for-Graph-Anomaly-Detection", "code_commit": "958cb1db8a2a1eae106ba2a179c952e38cab179e", "code_archive": "repos/4nrZXPFN1c4.zip", "code_archive_sha256": "9ee7efb5b8d5824211794a9fe2c79c43429f044c2db429d225a56900cd9fd0bf", "code_archive_bytes": 19192, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 25, "github_languages": {"Python": 57204}, "github_archived": false, "github_pushed_at": "2023-05-21T03:57:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/energy-transformer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IEsx-jwFk3g", "year": 2022, "status": "rejected", "title": "Deep Representations for Time-varying Brain Datasets", "authors": ["Sikun Lin", "Shuyun Tang", "Ambuj Singh"], "authorids": ["~Sikun_Lin1", "~Shuyun_Tang1", "~Ambuj_Singh1"], "authors_source": "OpenReview API", "abstract": "Finding an appropriate representation of dynamic activities in the brain is crucial for many downstream applications. Due to its highly dynamic nature, temporally averaged fMRI (functional magnetic resonance imaging) cannot capture the whole picture of underlying brain activities, and previous works lack the ability to learn and interpret the latent dynamics in brain architectures. In this paper, we build an efficient graph neural network model that incorporates both region-mapped fMRI sequences and structural connectivities obtained from DWI (diffusion-weighted imaging) as inputs. Through novel sample-level adaptive adjacency matrix learning and multi-resolution inner cluster smoothing, we find good representations of the latent brain dynamics. We also attribute inputs with integrated gradients, which enables us to infer (1) highly involved brain connections and subnetworks for each task (2) keyframes of imaging sequences along the temporal axis, and (3) subnetworks that discriminate between individual subjects. This ability to identify critical subnetworks that characterize brain states across heterogeneous tasks and individuals is of great importance to neuroscience research. Extensive experiments and ablation studies demonstrate our proposed method's superiority and efficiency in spatial-temporal graph signal modeling with insightful interpretations of brain dynamics.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "pimSlMRCZNk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2184/Reviewer_Cuxq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a graph neural network architecture for analyzing dynamic brain imaging data. The proposed approach uses both structural connectivity (via DWI) and functional connectivity (via fMRI) to setup the graph edge information, while the fMRI signals are used to define node features. The proposed network alternates between analyzing temporal information via gated temporal convolutional network layers and spatial information via GNN layers. A multi-resolution soft cluster smoothing is applied as the pooling operation in the GNN. Attribution of structural and functional imaging inputs is also explored using integrated gradients interpretation method. The proposed method is tested on a multi-task fMRI dataset and demonstrated better performance in both ablation studies and comparison to state-of-the-art GNN methods.", "review_text": "Strengths:\n\n- The proposed method combines structural and functional brain information, which are often assessed separately.\n\n- The proposed network includes components for both temporal and spatial processing.\n\n- The proposed network utilizes an individualized graph structure (connections) for each sample, rather than defining one adjacency matrix for all like most standard approaches. The network can handle this in part due to the soft clustering and keeping the same number of nodes at each GNN level.\n\n- The paper is fairly clearly presented.\n\nWeaknesses:\n\n- The \"strided non-causal TCN\" I feel is a bit of a misleading name - when I think temporal convolutional network I imagine it needs to have the causal aspect, as it's trying to ensure appropriate analysis of temporal information. The \"non-causal\" TCN is just a strided convolution on 1D data that happens to be in time dimension. Also, it would be helpful if the authors briefly include the information/equation for how the gating mechanism works.\n\n- For the experiments, the partition of the dataset appears to be by scan, not by subject. Since there are 1940 scans, 6 different scan tasks and only 56 subjects, my understanding is each subject may have undergone multiple session of the same task. Thus, it does not seem fair to split by scan, but rather should perform a patient-wise split for the validation method to get a better estimate of generalization performance.\n\n- While the total number of scans is reasonably large for neuroscience study, given the smaller number of individual subjects, a single split validation study does not seem appropriate - a cross-validation framework would be more convincing. Furthermore, as there is only a single split, there is no variance information so it is difficult to assess if there are true differences between methods that perform similarly.\n\n- The extremely large gap in empirical performance between the standard GNN models and the proposed method seems rather jarring (up to ~45% difference in accuracy). This makes me wonder whether the other models are not being appropriately tuned, or are given appropriate depth? I did not find any details about the baseline models, e.g., how many layers are used? For the proposed method it seems the authors would have used 4 layers, did they match the same number for baseline models? Also, how are the graphs constructed for the baseline model - are the same connections used or did the authors vary connections per sample like in their approach? And are the connections based on structural or functional data? In general, how would the authors explain the large difference in performance? It could also be interesting to compare to other GNN methods that were developed specifically for functional brain data analysis, e.g. Li et al (as cited in the paper), or Li et al, Braingnn: Interpretable brain graph neural network for fmri analysis, bioarxiv 2020.\n\n- The authors compare to GNN-only baselines, but not TCN-only baseline. I think it would make sense to compare to TCN models given (over) half of their network is based on TCN, and also given the large gap in performance using GNN only models; this may point to the feature learning in the TCN layers as being very important compared to the GNN component for classification learning.\n\n- How is the cluster number c chosen? The authors note that c decreases with deeper layers, but there is no mention as to the values used or how they are determined. Considering the authors claim this cluster smoothing is essential, it is important to explain how related parameter c is set. Does changing c then greatly affect the results?\n\nMinor comments:\n- For the variable for structural connectivity SC, the use of 2 letters for 1 variable is confusing when reading the equations - considering changing to 1 letter.\n- More detail on the number of each type of task scan and how many scans per subject would be welcome (eg added to appendix).\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a graph neural network architecture for analyzing dynamic brain imaging data. The proposed approach uses both structural connectivity (via DWI) and functional connectivity (via fMRI) to setup the graph edge information, while the fMRI signals are used to define node features. The proposed network alternates between analyzing temporal information via gated temporal convolutional network layers and spatial information via GNN layers. A multi-resolution soft cluster smoothing is applied as the pooling operation in the GNN. Attribution of structural and functional imaging inputs is also explored using integrated gradients interpretation method. The proposed method is tested on a multi-task fMRI dataset and demonstrated better performance in both ablation studies and comparison to state-of-the-art GNN methods.", "main_review": "Strengths:\n\n- The proposed method combines structural and functional brain information, which are often assessed separately.\n\n- The proposed network includes components for both temporal and spatial processing.\n\n- The proposed network utilizes an individualized graph structure (connections) for each sample, rather than defining one adjacency matrix for all like most standard approaches. The network can handle this in part due to the soft clustering and keeping the same number of nodes at each GNN level.\n\n- The paper is fairly clearly presented.\n\nWeaknesses:\n\n- The \"strided non-causal TCN\" I feel is a bit of a misleading name - when I think temporal convolutional network I imagine it needs to have the causal aspect, as it's trying to ensure appropriate analysis of temporal information. The \"non-causal\" TCN is just a strided convolution on 1D data that happens to be in time dimension. Also, it would be helpful if the authors briefly include the information/equation for how the gating mechanism works.\n\n- For the experiments, the partition of the dataset appears to be by scan, not by subject. Since there are 1940 scans, 6 different scan tasks and only 56 subjects, my understanding is each subject may have undergone multiple session of the same task. Thus, it does not seem fair to split by scan, but rather should perform a patient-wise split for the validation method to get a better estimate of generalization performance.\n\n- While the total number of scans is reasonably large for neuroscience study, given the smaller number of individual subjects, a single split validation study does not seem appropriate - a cross-validation framework would be more convincing. Furthermore, as there is only a single split, there is no variance information so it is difficult to assess if there are true differences between methods that perform similarly.\n\n- The extremely large gap in empirical performance between the standard GNN models and the proposed method seems rather jarring (up to ~45% difference in accuracy). This makes me wonder whether the other models are not being appropriately tuned, or are given appropriate depth? I did not find any details about the baseline models, e.g., how many layers are used? For the proposed method it seems the authors would have used 4 layers, did they match the same number for baseline models? Also, how are the graphs constructed for the baseline model - are the same connections used or did the authors vary connections per sample like in their approach? And are the connections based on structural or functional data? In general, how would the authors explain the large difference in performance? It could also be interesting to compare to other GNN methods that were developed specifically for functional brain data analysis, e.g. Li et al (as cited in the paper), or Li et al, Braingnn: Interpretable brain graph neural network for fmri analysis, bioarxiv 2020.\n\n- The authors compare to GNN-only baselines, but not TCN-only baseline. I think it would make sense to compare to TCN models given (over) half of their network is based on TCN, and also given the large gap in performance using GNN only models; this may point to the feature learning in the TCN layers as being very important compared to the GNN component for classification learning.\n\n- How is the cluster number c chosen? The authors note that c decreases with deeper layers, but there is no mention as to the values used or how they are determined. Considering the authors claim this cluster smoothing is essential, it is important to explain how related parameter c is set. Does changing c then greatly affect the results?\n\nMinor comments:\n- For the variable for structural connectivity SC, the use of 2 letters for 1 variable is confusing when reading the equations - considering changing to 1 letter.\n- More detail on the number of each type of task scan and how many scans per subject would be welcome (eg added to appendix).\n\n", "summary_of_the_review": "My recommendation is based on the new proposed network for combining functional/structural brain data analysis in a meaningful way for processing temporal and spatial information, and allowing individualized graph structures to be learned/used within a GNN model. While the results appear good, they appear somehow incredulously higher than standard GNN methods, and important missing information about experimental setup and empirical comparisons further dampen my enthusiasm for the work. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635964708470}, {"id": "xlPHJOIUquv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2184/Reviewer_zFSB"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a deep learning method for temporal data on graph nodes, specifically designed for brain imaging data. It can be deployed for classification of data where the time-varying data and graphs are individual specific, and pinpoint subgraph structure where group-specific changes occur. ", "review_text": "Strength: \n- Clearly written and easy to follow.\n- Practically useful framework for fMRI data analysis. \n- Extensive experiment that validates the proposed framework. \n\nWeakness:\n- Novelty of the paper is quite unclear, especially in the introduction. Why is analyzing time varying signals on graphs challenging? What are the problems that recent methods are facing, and how are the authors are solving them?\n- Why is a normalized adjacency matrix used? Any technical / clinical motivation?\n- It is not clear why having individual adjacency matrix implies a latent graph cannot be learnt. A_adp is a functional graph derived from node signals; why can't each adjacency matrix can be deployed here? This is important as the authors use both AHW and A_adpHW later in the framework. \n- From anatomical perspective, relationships in local (structural) connections make more sense than those among far ROIs. What is the rational that it can be beneficial (rather than simply referencing Ying 2018)?\n- In the experiment, did the authors divide the given dataset in to a single set of partitions? Was cross-validation adopted?\n- Temporal data varies in their length per subject. How was this problem dealt? \n- During the method comparisons, I am not sure if it is valid to set the same epoch limit to derive evaluation measures as each model has different complexity. \n- Is the outcome properly validated? There are several labels according to memory, attention, dot prove and so on... are the label-specific variations detected from this framework really associated with those tasks? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a deep learning method for temporal data on graph nodes, specifically designed for brain imaging data. It can be deployed for classification of data where the time-varying data and graphs are individual specific, and pinpoint subgraph structure where group-specific changes occur. ", "main_review": "Strength: \n- Clearly written and easy to follow.\n- Practically useful framework for fMRI data analysis. \n- Extensive experiment that validates the proposed framework. \n\nWeakness:\n- Novelty of the paper is quite unclear, especially in the introduction. Why is analyzing time varying signals on graphs challenging? What are the problems that recent methods are facing, and how are the authors are solving them?\n- Why is a normalized adjacency matrix used? Any technical / clinical motivation?\n- It is not clear why having individual adjacency matrix implies a latent graph cannot be learnt. A_adp is a functional graph derived from node signals; why can't each adjacency matrix can be deployed here? This is important as the authors use both AHW and A_adpHW later in the framework. \n- From anatomical perspective, relationships in local (structural) connections make more sense than those among far ROIs. What is the rational that it can be beneficial (rather than simply referencing Ying 2018)?\n- In the experiment, did the authors divide the given dataset in to a single set of partitions? Was cross-validation adopted?\n- Temporal data varies in their length per subject. How was this problem dealt? \n- During the method comparisons, I am not sure if it is valid to set the same epoch limit to derive evaluation measures as each model has different complexity. \n- Is the outcome properly validated? There are several labels according to memory, attention, dot prove and so on... are the label-specific variations detected from this framework really associated with those tasks? ", "summary_of_the_review": "This paper tackles an important problem in neurosience. \nClinical motivation and the overall pipeline make sense, and extensive efforts were put on evaluation of the ideas and framework. \nThe paper is mostly clear, but I do have some concerns as mentioned in the Main Review above which can be very critical. \nI am willing to change my score once I go through the rebuttal and other reviews if needed.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635906904835}, {"id": "llVgnmKzx5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2184/Reviewer_wRj8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a Graph Neural Network model to estimate latent dynamics in the human brain using functional Magnetic Resonance Imaging (fMRI) and Diffusion Weighted Imaging (DWI) while performing a classification task. The model consists of four parts 1) estimating each sample's adjacency matrix by learning a shared projection matrix 2) gated temporal convolutional neural network for learning information from time-series data 3) graph neural network to get embeddings of temporal and spatial information combined by 4) multi-resolution inner cluster smoothing. The authors use integrated gradients for estimating important features for the classification task helping with model interpretation. The authors show that the proposed method can perform better in the prediction task than GCN, GAT V2, etc. On the real dataset, it is shown that the method can capture temporal and spatial heterogeneity and provide regions important for the presented classification task.", "review_text": "I found the approach introduced in the paper interesting, and the motivation is clear. I enjoyed that the proposed approach aims to extract latent dynamics information, thus capturing temporal heterogeneity, which can be very useful in neuroscience as well as disease prediction research. While this contribution shows that the proposed method can generate better prediction results, I see several issues with it as it stands. The rationale for combining various existing approaches as proposed in the paper should be clearly justified emphasizing the novelty. The presentation of the manuscript can be improved, and a few details seem to be missing. I have some concerns, which are listed below.\n\n1) Existing approaches to estimate dynamic connectivity and approaches estimating spatiotemporal patterns in the fMRI data [1,2,3,4] are not mentioned. Authors could comment on what existing approaches (based on linear modeling) lack, relevance/irrelevance of the approaches in the given context, and compare the results.\n2) The model is not straightforward to understand in one go. Paragraphs in Section 2.2 could be rewritten such that they are more connected and give a better picture of the model. \n3) To validate the method and illustrate the motivation, please consider including some simulation studies on synthetic datasets. Since the method is intended to capture latent dynamics, the authors could borrow simulation settings from the related literature.\n4) It is mentioned that the F1 metric is used because of the imbalanced dataset, but the information about the imbalanced dataset is missing for understanding the data. Also, the number of frames for each task is missing. \n5) It seems that the experiments are performed only for one set of train, validation, and test set. Why are multiple runs of k-fold cross validation not used here? \n6) Not sure what to make of the qualitative analysis in Figures 6, 12, and 13. It would be helpful to discuss the importance of \"important\" nodes in the context of these tasks rather than just stating the region name. Also, the authors should describe the regions mentioned in the main text. The claim of the paper is that the proposed method can find good representations but the authors have not explained the importance of these subnetworks in the context of the task presented. Even a small discussion would be helpful.\n7) We notice that signal-important ROIs are not necessarily the same as connection-important ROIs:? What does it mean from the point of view of biological relevance? How to use these remarks as biomarkers?\n8) How reproducible are these important regions extracted? Since the complete problem is non-convex, depending on initialization, will the end result vary? Since these regions would be used in further downstream analysis, they would need to be highly reproducible.   \n9) Authors mention, \"We also find using adaptive adjacency matrices, and inner cluster smoothing can stabilize training, making the model less prone to overfitting and achieving close-to-best performance over a larger range of hyperparameters.\" Can you show results to support the claim made?\n10) Authors mention, \"Important brain regions obtained from $ATTR_A$ mostly comply with the previous literature.\" What are the important brain regions? Can you cite literature to back up this claim?\n11) Is there a reason that $A_{iadp}$ is sparse for each sample? Is this expected biologically or due to the way in which the model is defined? Would it be helpful to use sparsity constraints on $A$? This might be helpful in controlling the sparsity and in the interpretability.\n12) Can you comment on the biological relevance of the brain regions found in $ATTR_X$ \n13) How are $h_{adp}$ in eq 1 and $K$ in eq 2 set?\n14) X-axis and Y-axis ticks size can be increased to improve the visibility in Fig 3, Fig 5, Fig 10 and Fig 11.\n15) What is $\\tilde{D}$ and $v$ in section 2.1? Is $\\tilde{D}$ a diagonal matrix where $v,v$ entry is equal to the given sum?\n16) Published versions of the referenced arxiv papers should be cited.\n\nI am happy to raise my score if the above concerns are addressed.\n\n[1] Li, Lingge, et al. \"Modeling dynamic functional connectivity with latent factor Gaussian processes.\" Advances in neural information processing systems 32 (2019): 8263-8273.\n\n[2] Taghia, Jalil, et al. \"Bayesian switching factor analysis for estimating time-varying functional connectivity in fMRI.\" Neuroimage 155 (2017): 271-290.\n\n[3] Zhang, Gemeng, et al. \"Estimating dynamic functional brain connectivity with a sparse hidden Markov model.\" IEEE transactions on medical imaging 39.2 (2019): 488-498.\n\n[4] Bhinge, Suchita, et al. \"Extraction of time-varying spatiotemporal networks using parameter-tuned constrained IVA.\" IEEE transactions on medical imaging 38.7 (2019): 1715-1725.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a Graph Neural Network model to estimate latent dynamics in the human brain using functional Magnetic Resonance Imaging (fMRI) and Diffusion Weighted Imaging (DWI) while performing a classification task. The model consists of four parts 1) estimating each sample's adjacency matrix by learning a shared projection matrix 2) gated temporal convolutional neural network for learning information from time-series data 3) graph neural network to get embeddings of temporal and spatial information combined by 4) multi-resolution inner cluster smoothing. The authors use integrated gradients for estimating important features for the classification task helping with model interpretation. The authors show that the proposed method can perform better in the prediction task than GCN, GAT V2, etc. On the real dataset, it is shown that the method can capture temporal and spatial heterogeneity and provide regions important for the presented classification task.", "main_review": "I found the approach introduced in the paper interesting, and the motivation is clear. I enjoyed that the proposed approach aims to extract latent dynamics information, thus capturing temporal heterogeneity, which can be very useful in neuroscience as well as disease prediction research. While this contribution shows that the proposed method can generate better prediction results, I see several issues with it as it stands. The rationale for combining various existing approaches as proposed in the paper should be clearly justified emphasizing the novelty. The presentation of the manuscript can be improved, and a few details seem to be missing. I have some concerns, which are listed below.\n\n1) Existing approaches to estimate dynamic connectivity and approaches estimating spatiotemporal patterns in the fMRI data [1,2,3,4] are not mentioned. Authors could comment on what existing approaches (based on linear modeling) lack, relevance/irrelevance of the approaches in the given context, and compare the results.\n2) The model is not straightforward to understand in one go. Paragraphs in Section 2.2 could be rewritten such that they are more connected and give a better picture of the model. \n3) To validate the method and illustrate the motivation, please consider including some simulation studies on synthetic datasets. Since the method is intended to capture latent dynamics, the authors could borrow simulation settings from the related literature.\n4) It is mentioned that the F1 metric is used because of the imbalanced dataset, but the information about the imbalanced dataset is missing for understanding the data. Also, the number of frames for each task is missing. \n5) It seems that the experiments are performed only for one set of train, validation, and test set. Why are multiple runs of k-fold cross validation not used here? \n6) Not sure what to make of the qualitative analysis in Figures 6, 12, and 13. It would be helpful to discuss the importance of \"important\" nodes in the context of these tasks rather than just stating the region name. Also, the authors should describe the regions mentioned in the main text. The claim of the paper is that the proposed method can find good representations but the authors have not explained the importance of these subnetworks in the context of the task presented. Even a small discussion would be helpful.\n7) We notice that signal-important ROIs are not necessarily the same as connection-important ROIs:? What does it mean from the point of view of biological relevance? How to use these remarks as biomarkers?\n8) How reproducible are these important regions extracted? Since the complete problem is non-convex, depending on initialization, will the end result vary? Since these regions would be used in further downstream analysis, they would need to be highly reproducible.   \n9) Authors mention, \"We also find using adaptive adjacency matrices, and inner cluster smoothing can stabilize training, making the model less prone to overfitting and achieving close-to-best performance over a larger range of hyperparameters.\" Can you show results to support the claim made?\n10) Authors mention, \"Important brain regions obtained from $ATTR_A$ mostly comply with the previous literature.\" What are the important brain regions? Can you cite literature to back up this claim?\n11) Is there a reason that $A_{iadp}$ is sparse for each sample? Is this expected biologically or due to the way in which the model is defined? Would it be helpful to use sparsity constraints on $A$? This might be helpful in controlling the sparsity and in the interpretability.\n12) Can you comment on the biological relevance of the brain regions found in $ATTR_X$ \n13) How are $h_{adp}$ in eq 1 and $K$ in eq 2 set?\n14) X-axis and Y-axis ticks size can be increased to improve the visibility in Fig 3, Fig 5, Fig 10 and Fig 11.\n15) What is $\\tilde{D}$ and $v$ in section 2.1? Is $\\tilde{D}$ a diagonal matrix where $v,v$ entry is equal to the given sum?\n16) Published versions of the referenced arxiv papers should be cited.\n\nI am happy to raise my score if the above concerns are addressed.\n\n[1] Li, Lingge, et al. \"Modeling dynamic functional connectivity with latent factor Gaussian processes.\" Advances in neural information processing systems 32 (2019): 8263-8273.\n\n[2] Taghia, Jalil, et al. \"Bayesian switching factor analysis for estimating time-varying functional connectivity in fMRI.\" Neuroimage 155 (2017): 271-290.\n\n[3] Zhang, Gemeng, et al. \"Estimating dynamic functional brain connectivity with a sparse hidden Markov model.\" IEEE transactions on medical imaging 39.2 (2019): 488-498.\n\n[4] Bhinge, Suchita, et al. \"Extraction of time-varying spatiotemporal networks using parameter-tuned constrained IVA.\" IEEE transactions on medical imaging 38.7 (2019): 1715-1725.\n", "summary_of_the_review": "Based on the above assessments, I suggest rejection of this paper. The current experimental evaluations and discussion need to be stronger to be more convincing, and the presentation can be improved.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635800771632}, {"id": "h0-mxSF77MS", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2184/Reviewer_Whpw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method of graph neural network to jointly models dynamic functional signals and structural connectivities, to learn a good deep representation of brain dynamics. In this article, the author proposes some improved methods for graph neural networks, such as learning sample-level latent graph structures, temporal convolutional network and multi-resolution inner cluster smoothing methods to better learn insightful interpretations of brain dynamics. In the experiment, the author verified the effectiveness of the method in this paper on the fMRI signal data.", "review_text": "Pros:\n1. The method in this article has made many improvements to graph neural networks. These improvements are to capture the spatial information and temporal information in brain dynamics. These improvements have also been verified in experiments to learn better temporal and spatial importance sanity.\n\n2. The author uses the method of graph neural network to learn spatial information and temporal information. This new proposal may be extended to different fields. Not only does it have the effect of improving the brain image, but it may also be useful for general computer vision in the future. . So the author's innovation in the method is quite meaningful.\n\n Cons:\n1. In the experiment, the author compared with other different graph neural network methods. In table 1, the accuracy of this method is about 20% higher than other methods. These results are quite confusing to me, and the GCN method only has 41% accuracy. I think the baseline method compared by the author may be too poor. It is more appropriate to compare it with other methods that are not graph neural networks. The ones in Table 1 The result is still unable to prove the effectiveness of the method.\n\n2. The improvement of several methods on graph neural network in this paper is mainly based on some previous work. But the main contribution is to make some methodological adjustments for brain imaging, which greatly reduces the innovation point of this article. Therefore, the overall innovation of the article still needs to be improved.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a method of graph neural network to jointly models dynamic functional signals and structural connectivities, to learn a good deep representation of brain dynamics. In this article, the author proposes some improved methods for graph neural networks, such as learning sample-level latent graph structures, temporal convolutional network and multi-resolution inner cluster smoothing methods to better learn insightful interpretations of brain dynamics. In the experiment, the author verified the effectiveness of the method in this paper on the fMRI signal data.", "main_review": "Pros:\n1. The method in this article has made many improvements to graph neural networks. These improvements are to capture the spatial information and temporal information in brain dynamics. These improvements have also been verified in experiments to learn better temporal and spatial importance sanity.\n\n2. The author uses the method of graph neural network to learn spatial information and temporal information. This new proposal may be extended to different fields. Not only does it have the effect of improving the brain image, but it may also be useful for general computer vision in the future. . So the author's innovation in the method is quite meaningful.\n\n Cons:\n1. In the experiment, the author compared with other different graph neural network methods. In table 1, the accuracy of this method is about 20% higher than other methods. These results are quite confusing to me, and the GCN method only has 41% accuracy. I think the baseline method compared by the author may be too poor. It is more appropriate to compare it with other methods that are not graph neural networks. The ones in Table 1 The result is still unable to prove the effectiveness of the method.\n\n2. The improvement of several methods on graph neural network in this paper is mainly based on some previous work. But the main contribution is to make some methodological adjustments for brain imaging, which greatly reduces the innovation point of this article. Therefore, the overall innovation of the article still needs to be improved.", "summary_of_the_review": "This article improves the network method of graph neural network to learn a better representation for brain dynamics in the application of brain images. I prefer to reject this article because this article has made some small improvements on GNN, which is more like to combine them in different ways to improve the overall effect. In addition, the baseline effects compared by the author in the experiment are relatively weak, and the effectiveness of the method cannot be demonstrated.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635143919922}], "openreview_url": "https://openreview.net/forum?id=IEsx-jwFk3g", "arxiv_id": "2205.11648", "paper_pdf": "papers/IEsx-jwFk3g.pdf", "paper_pdf_sha256": "e45af310eaeeb14566f45d6e65cfa34cac9477d05d84831e4c79dc7f0139c109", "paper_pdf_bytes": 34179969, "paper_pdf_source": "openreview", "code_url": "https://github.com/sklin93/ReBraiD", "code_repository": "sklin93/ReBraiD", "code_commit": "7fbecd68e0fba5b86670916441dc509098417b42", "code_archive": "repos/IEsx-jwFk3g.zip", "code_archive_sha256": "f3a2ddf49f894acec99d6785488dae4eaa60e6033cbdd2a8cdf4821190f8fbc5", "code_archive_bytes": 18930, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 22, "github_languages": {"Python": 77048}, "github_archived": false, "github_pushed_at": "2022-08-16T19:31:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-representations-for-time-varying-brain-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RGeQOjc58d", "year": 2021, "status": "rejected", "title": "Improved Gradient based Adversarial Attacks for Quantized Networks", "authors": ["Kartik Gupta", "Thalaiyasingam Ajanthan"], "authorids": ["~Kartik_Gupta2", "~Thalaiyasingam_Ajanthan1"], "authors_source": "OpenReview API", "abstract": "Neural network quantization has become increasingly popular due to efficient memory consumption and faster computation resulting from bitwise operations on the quantized networks. Even though they exhibit excellent generalization capabilities, their robustness properties are not well-understood. In this work, we systematically study the robustness of quantized networks against gradient based adversarial attacks and demonstrate that these quantized models suffer from gradient vanishing issues and show a fake sense of robustness. By attributing gradient vanishing to poor forward-backward signal propagation in the trained network, we introduce a simple temperature scaling approach to mitigate this issue while preserving the decision boundary. Despite being a simple modification to existing gradient based adversarial attacks, experiments on CIFAR-10/100 datasets with multiple network architectures demonstrate that our temperature scaled attacks obtain near-perfect success rate on quantized networks while outperforming original attacks on adversarially trained models as well as floating point networks.\n", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "RXMZIKm0_eP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper813/AnonReviewer5"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work starts by questioning the apparent robustness of quantized networks and demonstrates that such robustness is more so a failure of the attack algorithm in picking up the gradient signal. The authors address this by tuning a scalar multiplier applied to the network logits, which doesn’t modify the model’s decision boundary. Through analyzing the Jacobian, two approaches are proposed to determine the scalar $\\beta$ without tuning it by performing the attack. This approach is quite effective on quantized networks and even provides significant improvement on floating-point networks combining with existing attacks like FGSM and PGD. The proposed modification might seem trivial at first, but it constitutes an important factor the community hasn’t taken notice of, to the best of my knowledge.\n\nA few questions: 1) I don’t see any mentions of tuning the attack step size; if we set the new $\\eta$ to $\\eta/\\beta$, we can keep the Jacobian intact and isolate the effect of temperature scaling for XENT. 2) how about sweeping $\\beta$ and plotting against adversarial accuracy? This is surely expensive, but it would paint a clearer picture of the optimality of $\\beta$ found using the proposed approaches. This can be done in tandem with an attack step of $\\eta/\\beta$.\n\nI like the result overall, though the effect of $\\beta$ on the Jacobian and the softmax can and should be separated, if my understanding is correct. The proposed approaches for determining $\\beta$ largely depend on the Jacobian; therefore, there should be more investigation on if scaling the Jacobian correctly is more important than getting good error signals from softmax.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting observation from quantized networks results in general, stronger adversarial attacks", "review": "This work starts by questioning the apparent robustness of quantized networks and demonstrates that such robustness is more so a failure of the attack algorithm in picking up the gradient signal. The authors address this by tuning a scalar multiplier applied to the network logits, which doesn’t modify the model’s decision boundary. Through analyzing the Jacobian, two approaches are proposed to determine the scalar $\\beta$ without tuning it by performing the attack. This approach is quite effective on quantized networks and even provides significant improvement on floating-point networks combining with existing attacks like FGSM and PGD. The proposed modification might seem trivial at first, but it constitutes an important factor the community hasn’t taken notice of, to the best of my knowledge.\n\nA few questions: 1) I don’t see any mentions of tuning the attack step size; if we set the new $\\eta$ to $\\eta/\\beta$, we can keep the Jacobian intact and isolate the effect of temperature scaling for XENT. 2) how about sweeping $\\beta$ and plotting against adversarial accuracy? This is surely expensive, but it would paint a clearer picture of the optimality of $\\beta$ found using the proposed approaches. This can be done in tandem with an attack step of $\\eta/\\beta$.\n\nI like the result overall, though the effect of $\\beta$ on the Jacobian and the softmax can and should be separated, if my understanding is correct. The proposed approaches for determining $\\beta$ largely depend on the Jacobian; therefore, there should be more investigation on if scaling the Jacobian correctly is more important than getting good error signals from softmax.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604622914484}, {"id": "KDQQdLCz04e", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper813/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Update**: Thanks to the authors for addressing my comments. As it was pointed out by the authors, temperature rescaling is mostly applicable to non-linear loss functions. For linear loss functions, temperature scaling only linear rescales the gradients. The difference between the proposed PGD++ attack and PGD with linear DLR loss is small (see the author's response to AR4). The improvements are most significant for FGSM but FGSM is not recommended for the robustness evaluation. Given the limited technical novelty and small improvements for linear loss functions, my score remains unchanged.\n\n###### Summary\nThe paper studies the robustness of binary neural networks (BNNS), which at first look have higher robustness than full-precious neural networks. The authors highlight the problem of poor signal propagation in BNNs, which makes gradient-based attacks difficult. To address this issue, the authors proposed a 1) single scalar rescaling of the jacobian to improve the signal propagation; 2) parameter-free hessian norm scaling technique. In the experiments, the authors demonstrated that the modified attacks reduce the accuracy of BNNs to zero and outperform existing gradient-based attacks against floating-point networks.\n\n###### Reasons for score\n\nI vote for a weak acceptance of this paper. The paper shows that BNNs are not robust and introduce an interesting gradient rescaling technique, which can also be used to attack full-precision networks. The rescaling technique is well explained, easy to apply for any existing attacks, and has low computational overhead. However, as I will discuss below, I see some problems comparing the proposed attack against a well-tuned PGD attack.\n\n###### Pros:\n1) The paper studies the robustness of BNNs. The robustness of BNNs is not well studied, and understanding the robustness of BNNs is an important research direction.\n2) The authors highlight the issue of signal propagation in BNNs. To address this issue, they devise a novel low-computational complexity technique.\n3) Experimental results for BNNs and full-precious models demonstrate that the modified attack is effective.\n\n###### Concerns and questions:\n- In the experiments, the authors use a single parameter for the step size for both PGD and FGSM attacks. On the other hand, the proposed method computes an optimal rescaling to achieve good signal propagation. Even though the proposed technique has a low computational budget, I believe the authors should do a grid search for the optimal step size for PGD and FGSM attacks for a fair comparison.\n- In the experiments, PGD L2 attack was unable to reduce the naturally trained models' accuracy to 0. This seems strange and unlikely as gradient shattering should not happen for naturally trained models. Can the authors explain these results? Is it possible that there might be an implementation issue?\n\n###### Comments and suggestions:\n- The proposed technique amplifies the error signal in a nonlinear way for nonlinear losses such as cross-entropy. However, for other losses such as multiclass-hinge loss, CW loss, the proposed method will simply linearly rescale the error signal. Attacking CW loss might be useful as it avoids the issues of saturated softmax gradients. Will this technique be useful for attacking with CW loss? \n- The authors claim that this method improves the efficiency of white-box attacks against full precision models. Is it possible for the authors to get the results for mnist_challenge and cifar10_challenge to see if the method outperforms an optimally tuned PGD attack?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper studies the robustness of binary neural networks. It highlights the issue of signal propagation in BNNs. To mitigate this issue, the authors propose a temperature rescaling technique.", "review": "**Update**: Thanks to the authors for addressing my comments. As it was pointed out by the authors, temperature rescaling is mostly applicable to non-linear loss functions. For linear loss functions, temperature scaling only linear rescales the gradients. The difference between the proposed PGD++ attack and PGD with linear DLR loss is small (see the author's response to AR4). The improvements are most significant for FGSM but FGSM is not recommended for the robustness evaluation. Given the limited technical novelty and small improvements for linear loss functions, my score remains unchanged.\n\n###### Summary\nThe paper studies the robustness of binary neural networks (BNNS), which at first look have higher robustness than full-precious neural networks. The authors highlight the problem of poor signal propagation in BNNs, which makes gradient-based attacks difficult. To address this issue, the authors proposed a 1) single scalar rescaling of the jacobian to improve the signal propagation; 2) parameter-free hessian norm scaling technique. In the experiments, the authors demonstrated that the modified attacks reduce the accuracy of BNNs to zero and outperform existing gradient-based attacks against floating-point networks.\n\n###### Reasons for score\n\nI vote for a weak acceptance of this paper. The paper shows that BNNs are not robust and introduce an interesting gradient rescaling technique, which can also be used to attack full-precision networks. The rescaling technique is well explained, easy to apply for any existing attacks, and has low computational overhead. However, as I will discuss below, I see some problems comparing the proposed attack against a well-tuned PGD attack.\n\n###### Pros:\n1) The paper studies the robustness of BNNs. The robustness of BNNs is not well studied, and understanding the robustness of BNNs is an important research direction.\n2) The authors highlight the issue of signal propagation in BNNs. To address this issue, they devise a novel low-computational complexity technique.\n3) Experimental results for BNNs and full-precious models demonstrate that the modified attack is effective.\n\n###### Concerns and questions:\n- In the experiments, the authors use a single parameter for the step size for both PGD and FGSM attacks. On the other hand, the proposed method computes an optimal rescaling to achieve good signal propagation. Even though the proposed technique has a low computational budget, I believe the authors should do a grid search for the optimal step size for PGD and FGSM attacks for a fair comparison.\n- In the experiments, PGD L2 attack was unable to reduce the naturally trained models' accuracy to 0. This seems strange and unlikely as gradient shattering should not happen for naturally trained models. Can the authors explain these results? Is it possible that there might be an implementation issue?\n\n###### Comments and suggestions:\n- The proposed technique amplifies the error signal in a nonlinear way for nonlinear losses such as cross-entropy. However, for other losses such as multiclass-hinge loss, CW loss, the proposed method will simply linearly rescale the error signal. Attacking CW loss might be useful as it avoids the issues of saturated softmax gradients. Will this technique be useful for attacking with CW loss? \n- The authors claim that this method improves the efficiency of white-box attacks against full precision models. Is it possible for the authors to get the results for mnist_challenge and cifar10_challenge to see if the method outperforms an optimally tuned PGD attack?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604568417457}, {"id": "s442b6v_ki", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper813/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the robustness of quantized neural networks against adversarial attacks. The authors use some slight modification of existing methods to successfully increase the attack success rate. In general, I think the idea is interesting. But I have some concerns that need to be addressed:\n1. I am not fully convinced by the arguments made at the beginning of Section 4. The authors claim that poor signal propagation such as gradient vanishing or gradient exploding should be a problem for adversarial attacks. However, I do not think the reasonings provided here is specifically for binarized neural networks. Equations (3) and (4) also works in regular full precision networks. I do not think there are any problems which only present in BNNs, so the arguments here are not strong enough. If poor signal propagation is a problem for attacks, why we don't see that in full precision networks? More discussions on this are welcomed.\n2. ResNet and DenseNet REF models in Table 3 seem to be surprisingly robust under PGD L2 attacks (column 3). This adversarial accuracy seems to be comparable to models with adversarial training. I think the authors need to provide some explanation on this. \n3. (Minor) Please refrain from only using color to distinguish curves/bars in figures as it may not be friendly to readers with color blindness.\n4. (Minor) The authors may need to re-organize some sections to make the paper easier to follow, for example, the \"related works\" section before the \"experiments\" section.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review", "review": "This paper studies the robustness of quantized neural networks against adversarial attacks. The authors use some slight modification of existing methods to successfully increase the attack success rate. In general, I think the idea is interesting. But I have some concerns that need to be addressed:\n1. I am not fully convinced by the arguments made at the beginning of Section 4. The authors claim that poor signal propagation such as gradient vanishing or gradient exploding should be a problem for adversarial attacks. However, I do not think the reasonings provided here is specifically for binarized neural networks. Equations (3) and (4) also works in regular full precision networks. I do not think there are any problems which only present in BNNs, so the arguments here are not strong enough. If poor signal propagation is a problem for attacks, why we don't see that in full precision networks? More discussions on this are welcomed.\n2. ResNet and DenseNet REF models in Table 3 seem to be surprisingly robust under PGD L2 attacks (column 3). This adversarial accuracy seems to be comparable to models with adversarial training. I think the authors need to provide some explanation on this. \n3. (Minor) Please refrain from only using color to distinguish curves/bars in figures as it may not be friendly to readers with color blindness.\n4. (Minor) The authors may need to re-organize some sections to make the paper easier to follow, for example, the \"related works\" section before the \"experiments\" section.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603944366838}, {"id": "Z9NWC7k0q-4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper813/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper identifies the gradient vanishing issue in the robustness of binary quantized networks. Therefore, it proposes to use temperature scaling approach in the attack generation. It has two methods for the temperature scale: (1) singular values of the input-output Jacobian and (2) maximizing the norm of the Hessian of the loss.\n\n------------Updates after rebuttal------------- \nThanks the authors for answering my questions. However, I don't think my comments are well addressed. Even though the paper [d] may not provide public available code, the authors could either use results from the paper [d] or implement the proposed attack on the models used by [d] to see the difference.\n\nStrengths:\n\n+ The proposed method work well on adv trained models and floating-point models.\n\n+ Practical approach by a simple modification to existing gradient based attacks.\n\n\nWeaknesses:\n\n- Binary quantization is not a well accepted method, since it can in general introduce >5% accuracy loss. There are a lot more valuable quantization schemes to investigate such as low-bit-with fixed point, power of 2, and additive power of 2 (Y. Li, X. Dong, and W. Wang, “Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,” in International Conference on Learning Representations, 2020).\n\n- The novelty is limited, since it brings the temperature scaling approach, an existing method, to the problem of attacking binary quantized models.\n\n- The paper writing is not constructed for easy understanding. \n\n\nComments and questions:\n\n1. I would like to see comparisons with other attacks that are particularly designed for quantized models.\n\n2. The third paragraph in Introduction, the paper tries to justify two techniques, but they are still not well motivated. \n\n3. The fourth paraph in Introduction mentions both full precision networks and floating-point networks. What’s the difference between these two?\n\n4. Table 1 results are surprising. Is the same observation made by other ref works? \n\n5. The method is to replace the softmax with a monotonic function (softmax with a single scalar) during the attack generation. Then for testing the attack success rate, I think the neural network should still use the original softmax (without scalar). Then the attack success rate won’t be degraded?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper brings the temperature scaling method to attack generation mainly on BNN.", "review": "The paper identifies the gradient vanishing issue in the robustness of binary quantized networks. Therefore, it proposes to use temperature scaling approach in the attack generation. It has two methods for the temperature scale: (1) singular values of the input-output Jacobian and (2) maximizing the norm of the Hessian of the loss.\n\n------------Updates after rebuttal------------- \nThanks the authors for answering my questions. However, I don't think my comments are well addressed. Even though the paper [d] may not provide public available code, the authors could either use results from the paper [d] or implement the proposed attack on the models used by [d] to see the difference.\n\nStrengths:\n\n+ The proposed method work well on adv trained models and floating-point models.\n\n+ Practical approach by a simple modification to existing gradient based attacks.\n\n\nWeaknesses:\n\n- Binary quantization is not a well accepted method, since it can in general introduce >5% accuracy loss. There are a lot more valuable quantization schemes to investigate such as low-bit-with fixed point, power of 2, and additive power of 2 (Y. Li, X. Dong, and W. Wang, “Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,” in International Conference on Learning Representations, 2020).\n\n- The novelty is limited, since it brings the temperature scaling approach, an existing method, to the problem of attacking binary quantized models.\n\n- The paper writing is not constructed for easy understanding. \n\n\nComments and questions:\n\n1. I would like to see comparisons with other attacks that are particularly designed for quantized models.\n\n2. The third paragraph in Introduction, the paper tries to justify two techniques, but they are still not well motivated. \n\n3. The fourth paraph in Introduction mentions both full precision networks and floating-point networks. What’s the difference between these two?\n\n4. Table 1 results are surprising. Is the same observation made by other ref works? \n\n5. The method is to replace the softmax with a monotonic function (softmax with a single scalar) during the attack generation. Then for testing the attack success rate, I think the neural network should still use the original softmax (without scalar). Then the attack success rate won’t be degraded?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603919631514}, {"id": "11aKGkVvBSm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper813/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Update** : Since most of my issues have been addressed, I have changed my rating from 4 to 6\n\nSummary:\n\nThis paper studies the robustness of quantized networks against gradient-based adversarial attacks (for L2 and Linf norms), showing how quantized models suffer from gradient vanishing, giving a false sense of security via gradient masking. To circumvent this issue, the authors propose temperature scaling approaches that can overcome this masking, achieving near-perfect perfect success in crafting adversarial inputs for these models.  \n\n##########################################################################\n\nReasons for score: \n\nThe paper's ultimate goal is to get better gradient-based attack performance on quantized (binarized, in this case) networks. However, key steps that should have been tried first for benchmarking such as adaptive PGD attacks have not been performed. Moreover, it is not clear what benefit the proposed method has in this scenario compared to gradient-free attacks like Boundary++.  The paper's contributions, although including some nice analyses on temperature scaling based solutions, are too weak to be accepted in their current form.\n\n \n##########################################################################\n\nPros: \n  \n- Improvement in attack success rates for full-precision networks, even for FGSM, seems like an exciting result. Further analyses and methods on top of this could be used to further increase the strength of these first-order gradient attacks.\n\n- Jacobian and Hessian based detailed analyses of temperature scaling, and what different solutions correspond to in terms of robustness is quite insightful and interesting.  \n\n##########################################################################\n\nCons: \n\n- Gradient masking is a relatively well-known phenomenon in adversarial machine learning. In cases when normal first-order gradient attacks fail, techniques like adaptive PGD attacks, gradient-free attacks, or even black-box transfer attacks are some straightforward methods to overcome gradient masking. Thus, it is not clear why the authors did not try non-gradient attacks before jumping to a complicated algorithm. At the very least, those attacks (like Boundary++) should at least be part of benchmarks for comparison. \n  - For starters, please refer to 'Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks': they have a [publicly available implementation](https://github.com/fra31/auto-attack ) as well\n  - All of this is crucial, especially since the paper claims (Section 4) that \"we would like to point out that the focus of this paper is to improve gradient-based attacks on already trained BNNs\"\n\n- In general, investigating if a model exhibits gradient masking is not a contribution: standard checks like comparing transfer rates, multi-step to single-step performance, attack rates for increasing attack budgets, etc are often used to check for gradient masking. \n\n- Figure 1: For which norm are these numbers reported? Without knowing the norm, it is hard to say if Figure b is a sign of gradient masking or not.\n\n- Section 2.2 \"..these attacks have been further strengthened by a random initial step\". This is partially true: the real benefit comes from having multiple random restarts. Having just one random initialization by itself is not that useful. Please re-run evaluation experiments with random restarts (20 is a good number).\n\n- Section 3: What does \"adversarial accuracy\" refer to? Is it accuracy on perturbed inputs f(x') = y, or success rate of the adversary when trying to change predictions aka f(x) ~= f(x')? Please clarify\n\n- Section 3.1 \"... clearly indicate gradient masking issues..\" please elaborate: not every reader will be familiar with the set of checks used for gradient masking.\n\n- Issues with Cross-Entropy based loss and how they promote certain magnitudes of logit values are not new. The authors might want to have a look at Section 4.1 of [Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse\nParameter-free Attacks](https://arxiv.org/pdf/2003.01690.pdf) to see if there are similarities/differences in the proposed temperature-based variant, and how the proposed method is better than the one in the Difference of Logits-Ratio based loss? This work seems to be a key and relevant part of related work and should be included in comparisons/benchmarking.\n \n- \"implementation of our algorithm will be released upon publication\". Please anonymize and attach the code in response. \n\n- The benefit of using the proposed FGSM++/PGD++ attacks on full-precision models trained with adversarial robustness seems to be negligible (Table 4), and should not be overstated in results. Also, since these attacks all have random seeds, please perform experiments multiple times for statistical significance and report summary statistics. \n\n##########################################################################\n\nMinor Edits:\n\n- Section 2.2 \"..perturbations to the images...\" the definition here is for adversarial examples in general, and should thus be \"perturbations to data\"\n\n- Section 2.2 \"Gradient-based attacks can be... written as Projected Gradient Descent (PGD)\" this is true only for first-order gradient-based attacks, not all gradient-based attacks (examples JSMA). Please correct.\n\n- Section 4.1 \"...since most of the modern networks consist of ReLU nonlinearities\" this can (and often is) circumvented using Fake-ReLU. Example implementation [here](https://github.com/MadryLab/robustness/blob/89bdf8088a8f4bd4a8b86925a2801069ec281fee/robustness/tools/custom_modules.py#L5)\n\n- Section 5 \"...and they hypothesize that linear networks would be robust to adversarial attacks.\" this is not their conclusion, and seems to be out of context.\n\n- Section 6 should preferably be either towards the end or at the beginning? Not clear why it is in the middle of other sections\n\n\nPlease address and clarify the cons above ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some interesting techniques proposed, but contributions are too weak", "review": "**Update** : Since most of my issues have been addressed, I have changed my rating from 4 to 6\n\nSummary:\n\nThis paper studies the robustness of quantized networks against gradient-based adversarial attacks (for L2 and Linf norms), showing how quantized models suffer from gradient vanishing, giving a false sense of security via gradient masking. To circumvent this issue, the authors propose temperature scaling approaches that can overcome this masking, achieving near-perfect perfect success in crafting adversarial inputs for these models.  \n\n##########################################################################\n\nReasons for score: \n\nThe paper's ultimate goal is to get better gradient-based attack performance on quantized (binarized, in this case) networks. However, key steps that should have been tried first for benchmarking such as adaptive PGD attacks have not been performed. Moreover, it is not clear what benefit the proposed method has in this scenario compared to gradient-free attacks like Boundary++.  The paper's contributions, although including some nice analyses on temperature scaling based solutions, are too weak to be accepted in their current form.\n\n \n##########################################################################\n\nPros: \n  \n- Improvement in attack success rates for full-precision networks, even for FGSM, seems like an exciting result. Further analyses and methods on top of this could be used to further increase the strength of these first-order gradient attacks.\n\n- Jacobian and Hessian based detailed analyses of temperature scaling, and what different solutions correspond to in terms of robustness is quite insightful and interesting.  \n\n##########################################################################\n\nCons: \n\n- Gradient masking is a relatively well-known phenomenon in adversarial machine learning. In cases when normal first-order gradient attacks fail, techniques like adaptive PGD attacks, gradient-free attacks, or even black-box transfer attacks are some straightforward methods to overcome gradient masking. Thus, it is not clear why the authors did not try non-gradient attacks before jumping to a complicated algorithm. At the very least, those attacks (like Boundary++) should at least be part of benchmarks for comparison. \n  - For starters, please refer to 'Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks': they have a [publicly available implementation](https://github.com/fra31/auto-attack ) as well\n  - All of this is crucial, especially since the paper claims (Section 4) that \"we would like to point out that the focus of this paper is to improve gradient-based attacks on already trained BNNs\"\n\n- In general, investigating if a model exhibits gradient masking is not a contribution: standard checks like comparing transfer rates, multi-step to single-step performance, attack rates for increasing attack budgets, etc are often used to check for gradient masking. \n\n- Figure 1: For which norm are these numbers reported? Without knowing the norm, it is hard to say if Figure b is a sign of gradient masking or not.\n\n- Section 2.2 \"..these attacks have been further strengthened by a random initial step\". This is partially true: the real benefit comes from having multiple random restarts. Having just one random initialization by itself is not that useful. Please re-run evaluation experiments with random restarts (20 is a good number).\n\n- Section 3: What does \"adversarial accuracy\" refer to? Is it accuracy on perturbed inputs f(x') = y, or success rate of the adversary when trying to change predictions aka f(x) ~= f(x')? Please clarify\n\n- Section 3.1 \"... clearly indicate gradient masking issues..\" please elaborate: not every reader will be familiar with the set of checks used for gradient masking.\n\n- Issues with Cross-Entropy based loss and how they promote certain magnitudes of logit values are not new. The authors might want to have a look at Section 4.1 of [Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse\nParameter-free Attacks](https://arxiv.org/pdf/2003.01690.pdf) to see if there are similarities/differences in the proposed temperature-based variant, and how the proposed method is better than the one in the Difference of Logits-Ratio based loss? This work seems to be a key and relevant part of related work and should be included in comparisons/benchmarking.\n \n- \"implementation of our algorithm will be released upon publication\". Please anonymize and attach the code in response. \n\n- The benefit of using the proposed FGSM++/PGD++ attacks on full-precision models trained with adversarial robustness seems to be negligible (Table 4), and should not be overstated in results. Also, since these attacks all have random seeds, please perform experiments multiple times for statistical significance and report summary statistics. \n\n##########################################################################\n\nMinor Edits:\n\n- Section 2.2 \"..perturbations to the images...\" the definition here is for adversarial examples in general, and should thus be \"perturbations to data\"\n\n- Section 2.2 \"Gradient-based attacks can be... written as Projected Gradient Descent (PGD)\" this is true only for first-order gradient-based attacks, not all gradient-based attacks (examples JSMA). Please correct.\n\n- Section 4.1 \"...since most of the modern networks consist of ReLU nonlinearities\" this can (and often is) circumvented using Fake-ReLU. Example implementation [here](https://github.com/MadryLab/robustness/blob/89bdf8088a8f4bd4a8b86925a2801069ec281fee/robustness/tools/custom_modules.py#L5)\n\n- Section 5 \"...and they hypothesize that linear networks would be robust to adversarial attacks.\" this is not their conclusion, and seems to be out of context.\n\n- Section 6 should preferably be either towards the end or at the beginning? Not clear why it is in the middle of other sections\n\n\nPlease address and clarify the cons above ", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603861082639}], "openreview_url": "https://openreview.net/forum?id=RGeQOjc58d", "arxiv_id": "2003.13511", "paper_pdf": "papers/RGeQOjc58d.pdf", "paper_pdf_sha256": "c13eafc3dddfaebeae3649727ec54e41cb61d874d173bc4ac9a31dc71090386e", "paper_pdf_bytes": 455817, "paper_pdf_source": "openreview", "code_url": "https://github.com/kartikgupta-at-anu/attack-bnn", "code_repository": "kartikgupta-at-anu/attack-bnn", "code_commit": "79c60f64b9ced3084571c805404876475f2bf1bd", "code_archive": "repos/RGeQOjc58d.zip", "code_archive_sha256": "66abc5e1ddba9997967ba2a07f88f849a1510a09070e72843bbb04b229823d3b", "code_archive_bytes": 40718, "code_file_count": 26, "code_extensions": {".py": 20, ".sh": 6}, "github_disk_usage_kb": 28, "github_languages": {"Python": 111733, "Shell": 30760}, "github_archived": false, "github_pushed_at": "2022-05-06T13:58:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improved-gradient-based-adversarial-attacks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SylUzpNFDS", "year": 2020, "status": "rejected", "title": "SoftLoc: Robust Temporal Localization under Label Misalignment", "authors": ["Julien Schroeter", "Kirill Sidorov", "Dave Marshall"], "authorids": ["schroeterj1@cardiff.ac.uk", "sidorovk@cardiff.ac.uk", "marshallad@cardiff.ac.uk"], "authors_source": "OpenReview API", "abstract": "This work addresses the long-standing problem of robust event localization in the presence of temporally of misaligned labels in the training data. We propose a novel versatile loss function that generalizes a number of training regimes from standard fully-supervised cross-entropy to count-based weakly-supervised learning. Unlike classical models which are constrained to strictly fit the annotations during training, our soft localization learning approach relaxes the reliance on the exact position of labels instead. Training with this new loss function exhibits strong robustness to temporal misalignment of labels, thus alleviating the burden of precise annotation of temporal sequences. We demonstrate state-of-the-art performance against standard benchmarks in a number of challenging experiments and further show that robustness to label noise is not achieved at the expense of raw performance. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1e8STrIqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper414/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose SoftLoc, an interpolation of two temporal event prediction losses (one based on label & prediction smoothing, and one based on weakly-supervised count-based loss) that makes the predictor more robust to noisy training data. They demonstrate on various temporal alignment datasets from music to wearable sensors to video action segmentations, that their method is performs well both on noisy-free and noisy settings compared to prior approaches.\n\nStrengths:\n(1) relatively thorough experimental valuations: using 4 datasets comparing with sufficient number of prior approaches (One potential improvement could be to try noise distributions other than Gaussian)\n(2) simple objective and consistent improvements. It is encouraging that the simple modification enables such consistent empirical improvements.\n\nWeaknesses:\n(1) the novelty of the method appears limited. The weakly-supervised loss is borrowed from the prior work, so it seems like the main algorithmic novelty is to add noise to predictions (as opposed to just adding to labels as done in prior work). \n\nOther comments:\n(1) in 5.1.1, is there justification for \\alpha_\\tau expression? How did you pick that? \n(2) is Issue 1 an actual\n problem? Even with just label smoothing, I would expect predictor will find the mode and peak-picking can produce the correct predictions. Similarly, discussions of how the proposed method can solve the issues listed may benefit from some rewriting, or even simple toy experiments to demo that those are actually the concerns that are not addressed by prior work. Such toy experiments may complement the existing end-to-end experiments to demonstrate the precise properties of the proposed method. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The authors propose SoftLoc, an interpolation of two temporal event prediction losses (one based on label & prediction smoothing, and one based on weakly-supervised count-based loss) that makes the predictor more robust to noisy training data. They demonstrate on various temporal alignment datasets from music to wearable sensors to video action segmentations, that their method is performs well both on noisy-free and noisy settings compared to prior approaches.\n\nStrengths:\n(1) relatively thorough experimental valuations: using 4 datasets comparing with sufficient number of prior approaches (One potential improvement could be to try noise distributions other than Gaussian)\n(2) simple objective and consistent improvements. It is encouraging that the simple modification enables such consistent empirical improvements.\n\nWeaknesses:\n(1) the novelty of the method appears limited. The weakly-supervised loss is borrowed from the prior work, so it seems like the main algorithmic novelty is to add noise to predictions (as opposed to just adding to labels as done in prior work). \n\nOther comments:\n(1) in 5.1.1, is there justification for \\alpha_\\tau expression? How did you pick that? \n(2) is Issue 1 an actual\n problem? Even with just label smoothing, I would expect predictor will find the mode and peak-picking can produce the correct predictions. Similarly, discussions of how the proposed method can solve the issues listed may benefit from some rewriting, or even simple toy experiments to demo that those are actually the concerns that are not addressed by prior work. Such toy experiments may complement the existing end-to-end experiments to demonstrate the precise properties of the proposed method. "}, "tcdate": 1572392253957}, {"id": "B1lBTZiAKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper414/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper deals with temporal label noise, or label misalignment problem in localization, which is the long-standing problem of robust event localization regarding the temporally misaligned labels in sequential data.  Some existing works are constrained to hardly fit the annotations of data during training. The work models the temporal label misalignment with perturbations to the ground-truth timestamps, and presents a soft localization learning scheme which relaxes the reliance on the exact position of labels. The idea introduced to temporal data is well-motivated since the label for temporal data is often dispersed. Technically speaking, the proposed SoftLoc loss comprises two terms: 1). $\\mathscr{L}_{SLL}$: a soft learning loss that relaxes the prediction mass concentration, by symmetrically filtering the labels and predictions. This term helps to relax the model’s reliance on exact label locations. 2). $\\mathscr{L}_{MC}$: a mass convergence loss that acts as a regularizer to facilitate the model with precise impulse-like localizations. With a trade of factor to balance two terms, the SoftLoc model can achieve precise impulse-like localization performance without weakening the model robustness. Various applications, such as PIANO ONSET, DRUM DETECTION and TIME SERIES DETECTION are performed to verify the effectiveness of the proposed method. And state-of-the-art performance is achieved.\n\nDespite the achievement indicated by the experiments, I still have some concerns about this paper.\n\n(1)\tThe authors propose the relaxed loss L_{SLL} for the soft learning of the location. And the label smoothing idea (applying a ˜S^2-Gaussian filter to the labels) has been introduced to increase the robustness to temporal misalignment of annotations [2]. Although the authors discuss the several inherent drawbacks of it in the 2nd paragraph of Related Work, they still follow the label smoothing idea in their model (Eq.2) in Section 4.1 in a two-side smoothed process. The reason should be clarified.\n\n(2)\tThe authors claim that their method has the advantage of generalizing some regimes of weakly-supervised learning. And the adopted L_{MC} is a weakly-supervised loss. Does the advantage come from L_{MC}?   \n\n(3)\tThe authors propose the two-side relaxed loss L_{SLL} for the soft learning of temporal localization problem. However, in the ablation study, the authors do not give the results with different level of label noise for one-side variant. Adding these results could help to demonstrate the necessity of the two-side relaxed loss L_{SLL}. \n\n\nReferences:\n[1] Improved musical onset detection with convolutional neural networks\n[2] Onsets and frames: Dual-objective piano transcription\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper deals with temporal label noise, or label misalignment problem in localization, which is the long-standing problem of robust event localization regarding the temporally misaligned labels in sequential data.  Some existing works are constrained to hardly fit the annotations of data during training. The work models the temporal label misalignment with perturbations to the ground-truth timestamps, and presents a soft localization learning scheme which relaxes the reliance on the exact position of labels. The idea introduced to temporal data is well-motivated since the label for temporal data is often dispersed. Technically speaking, the proposed SoftLoc loss comprises two terms: 1). $\\mathscr{L}_{SLL}$: a soft learning loss that relaxes the prediction mass concentration, by symmetrically filtering the labels and predictions. This term helps to relax the model’s reliance on exact label locations. 2). $\\mathscr{L}_{MC}$: a mass convergence loss that acts as a regularizer to facilitate the model with precise impulse-like localizations. With a trade of factor to balance two terms, the SoftLoc model can achieve precise impulse-like localization performance without weakening the model robustness. Various applications, such as PIANO ONSET, DRUM DETECTION and TIME SERIES DETECTION are performed to verify the effectiveness of the proposed method. And state-of-the-art performance is achieved.\n\nDespite the achievement indicated by the experiments, I still have some concerns about this paper.\n\n(1)\tThe authors propose the relaxed loss L_{SLL} for the soft learning of the location. And the label smoothing idea (applying a ˜S^2-Gaussian filter to the labels) has been introduced to increase the robustness to temporal misalignment of annotations [2]. Although the authors discuss the several inherent drawbacks of it in the 2nd paragraph of Related Work, they still follow the label smoothing idea in their model (Eq.2) in Section 4.1 in a two-side smoothed process. The reason should be clarified.\n\n(2)\tThe authors claim that their method has the advantage of generalizing some regimes of weakly-supervised learning. And the adopted L_{MC} is a weakly-supervised loss. Does the advantage come from L_{MC}?   \n\n(3)\tThe authors propose the two-side relaxed loss L_{SLL} for the soft learning of temporal localization problem. However, in the ablation study, the authors do not give the results with different level of label noise for one-side variant. Adding these results could help to demonstrate the necessity of the two-side relaxed loss L_{SLL}. \n\n\nReferences:\n[1] Improved musical onset detection with convolutional neural networks\n[2] Onsets and frames: Dual-objective piano transcription\n\n\n\n"}, "tcdate": 1571889597508}, {"id": "SylPrLSRtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper414/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new loss for training models that predict where events occur in a sequence when the training sequence has noisy labels. The central idea is to smooth the label sequence and prediction sequence and compare these rather than to force the model to treat all errors as equally serious.\n\nThe proposed problem seems sensible, and the method is a reasonable approach. The evaluations are carried out on a variety of different tasks (piano onset detection, drum detection, smoking detection, video action segmentation).\n\nSuggestions\n* While the authors do bring up lots of related work on learning from noisy labels, the insights from that work, and its relationship to this proposed technique could be more productively explored\n* The connections between the assumptions of the evaluation metric and the motivation for the smoothing methodology could be more productively elucidated\n* The task explored in this paper, and the task-specific problem, should be described more generally since ICLR has a generalist readership. For instance, the paper gets off on a rather strange footing discussing large data (indeed, but evident to the vast majority of ICLR readers), but little is said in terms of the specifics of the temporal localisation problem except via citations to other papers. In particular, the task set up, the problem posed by label misalignment could be described elegantly right in the introduction with a carefully designed figure. In section 2, the description is hard to follow, the mathematical notation is vague and hard to parse. For instance, the label sequence Y, is d-dimensional- the meaning of d should be clarified. Calling it “discretised” is also a bit strange, for most readers- it’s a label sequence.\n\nThe evaluation objective needs to be clarified, at least qualitatively; ideally in the introduction. Introducing a new objective is meaningful not only in light of noisy labels (always a problem), but in light of how the evaluation is carried out.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a new loss for training models that predict where events occur in a sequence when the training sequence has noisy labels. The central idea is to smooth the label sequence and prediction sequence and compare these rather than to force the model to treat all errors as equally serious.\n\nThe proposed problem seems sensible, and the method is a reasonable approach. The evaluations are carried out on a variety of different tasks (piano onset detection, drum detection, smoking detection, video action segmentation).\n\nSuggestions\n* While the authors do bring up lots of related work on learning from noisy labels, the insights from that work, and its relationship to this proposed technique could be more productively explored\n* The connections between the assumptions of the evaluation metric and the motivation for the smoothing methodology could be more productively elucidated\n* The task explored in this paper, and the task-specific problem, should be described more generally since ICLR has a generalist readership. For instance, the paper gets off on a rather strange footing discussing large data (indeed, but evident to the vast majority of ICLR readers), but little is said in terms of the specifics of the temporal localisation problem except via citations to other papers. In particular, the task set up, the problem posed by label misalignment could be described elegantly right in the introduction with a carefully designed figure. In section 2, the description is hard to follow, the mathematical notation is vague and hard to parse. For instance, the label sequence Y, is d-dimensional- the meaning of d should be clarified. Calling it “discretised” is also a bit strange, for most readers- it’s a label sequence.\n\nThe evaluation objective needs to be clarified, at least qualitatively; ideally in the introduction. Introducing a new objective is meaningful not only in light of noisy labels (always a problem), but in light of how the evaluation is carried out.\n\n"}, "tcdate": 1571866175050}], "openreview_url": "https://openreview.net/forum?id=SylUzpNFDS", "arxiv_id": null, "paper_pdf": "papers/SylUzpNFDS.pdf", "paper_pdf_sha256": "e29061a3b055fb66bfb6dbc3dcc1cad4df2748cc151d5544853debbf84cdc2c8", "paper_pdf_bytes": 1232214, "paper_pdf_source": "openreview", "code_url": "https://github.com/SoftLocNIPS/submission", "code_repository": "SoftLocNIPS/submission", "code_commit": "c5a5106c68070cd8fb037f2940e885132e58f463", "code_archive": "repos/SylUzpNFDS.zip", "code_archive_sha256": "db19d79331b9dbeb14a099eab4f7402570adf7b023b482f76c5d02378f130f9d", "code_archive_bytes": 28830, "code_file_count": 12, "code_extensions": {".py": 11, ".sh": 1}, "github_disk_usage_kb": 25, "github_languages": {"Python": 75271, "Shell": 324}, "github_archived": false, "github_pushed_at": "2019-08-12T13:18:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/softloc-robust-temporal-localization-under"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkxjnjA5KQ", "year": 2019, "status": "rejected", "title": "Transfer Learning for Related Reinforcement Learning Tasks via Image-to-Image Translation", "authors": ["Shani Gamrian", "Yoav Goldberg"], "authorids": ["gamrianshani@gmail.com", "yoav.goldberg@gmail.com"], "authors_source": "OpenReview API", "abstract": "Deep Reinforcement Learning has managed to achieve state-of-the-art results in learning control policies directly from raw pixels. However, despite its remarkable success, it fails to generalize, a fundamental component required in a stable Artificial Intelligence system. Using the Atari game Breakout, we demonstrate the difficulty of a trained agent in adjusting to simple modifications in the raw image, ones that a human could adapt to trivially. In transfer learning, the goal is to use the knowledge gained from the source task to make the training of the target task faster and better. We show that using various forms of fine-tuning, a common method for transfer learning, is not effective for adapting to such small visual changes. In fact, it is often easier to re-train the agent from scratch than to fine-tune a trained agent. We suggest that in some cases transfer learning can be improved by adding a dedicated component whose goal is to learn to visually map between the known domain and the new one. Concretely, we use Unaligned Generative Adversarial Networks (GANs) to create a mapping function to translate images in the target task to corresponding images in the source task. These mapping functions allow us to transform between various variations of the Breakout game, as well as between different levels of a Nintendo game, Road Fighter. We show that learning this mapping is substantially more efficient than re-training. A visualization of a trained agent playing Breakout and Road Fighter, with and without the GAN transfer, can be seen in \\url{https://streamable.com/msgtm} and \\url{https://streamable.com/5e2ka}.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1xIHS96hQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper739/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper propose an intermediate stage before transfer learning on playing new games that is with slight visual change. The intermediate stage is basically a mapping function to translate the images in the new game to old game with certain correspondence. The paper claims that the adding of intermediate stage can be much more efficient than re-train the model instead. \nThen the paper compares different baselines without the mapping model. The baselines are either re-trained from scratch or (partially) initialized with trained model. The learning curves show that fine-tuning fails to transfer knowledge from the source domain to target domain. The mapping model is constructed based on unaligned GAN. And the experiments are setup and results are shown.\n\nPros:\n+ The paper makes a very good start from analogizing human being adjusting himself between similar tasks. \n+ The paper demonstrates strong motivation on improving the existing transfer learnings that are either fail or take too much time to train from scratch.\n+ The paper clearly illustrate the learning curve of multiple approaches for transferring knowledge across tasks.\n+ The paper proves detailed analysis why using unaligned GAN to learn the mapping model, and gives\n+ I also like the experiment section. It is well written, especially the discussions section answer all my questions. \n\nQuestions:\n1.\tWhy fine-tuning from a model that is trained from related task does not help, even decelerate the learning process? Could you explain it more?\n2.\tCould you please also include in figure 2 the proposed transfer learning curve with the mapping model G? I’m curious how much faster it will converge than the Full-FT. And I suppose the retrain from scratch can be extremely slow and will exceed the training epoch scope in the figure.\n3.\tIn dataset collection, you use untrained agent to collect source domain image. Will it improve the results if you use well trained agent, or even human agent, instead? \n4.\tI hope, if possible, you can share the source code in the near future.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting method to improve transfer learning between related tasks. The motivation is strong, explanations are intuitive, technical parts are solid, experiments are sufficient. ", "review": "This paper propose an intermediate stage before transfer learning on playing new games that is with slight visual change. The intermediate stage is basically a mapping function to translate the images in the new game to old game with certain correspondence. The paper claims that the adding of intermediate stage can be much more efficient than re-train the model instead. \nThen the paper compares different baselines without the mapping model. The baselines are either re-trained from scratch or (partially) initialized with trained model. The learning curves show that fine-tuning fails to transfer knowledge from the source domain to target domain. The mapping model is constructed based on unaligned GAN. And the experiments are setup and results are shown.\n\nPros:\n+ The paper makes a very good start from analogizing human being adjusting himself between similar tasks. \n+ The paper demonstrates strong motivation on improving the existing transfer learnings that are either fail or take too much time to train from scratch.\n+ The paper clearly illustrate the learning curve of multiple approaches for transferring knowledge across tasks.\n+ The paper proves detailed analysis why using unaligned GAN to learn the mapping model, and gives\n+ I also like the experiment section. It is well written, especially the discussions section answer all my questions. \n\nQuestions:\n1.\tWhy fine-tuning from a model that is trained from related task does not help, even decelerate the learning process? Could you explain it more?\n2.\tCould you please also include in figure 2 the proposed transfer learning curve with the mapping model G? I’m curious how much faster it will converge than the Full-FT. And I suppose the retrain from scratch can be extremely slow and will exceed the training epoch scope in the figure.\n3.\tIn dataset collection, you use untrained agent to collect source domain image. Will it improve the results if you use well trained agent, or even human agent, instead? \n4.\tI hope, if possible, you can share the source code in the near future.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541412157760}, {"id": "rkeqDrUTnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper739/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper seeks to generalize the reinforcement learning agents to related tasks. The authors first show the failure of conventional transfer learning techniques, then use GANs to translate the images in the target task to those in the source task. It is an interesting attempt to use the style-transferred images for generalization of RL agents. The paper is well written and easy to follow.\nPros:\n1.\tIt is a novel attempt to use GANs to generate pictures that help RL agents transfer the policies to other related environments.\n2.\tIt is an interesting viewpoint to use the performance of RL agent to evaluate the quality of images generated by GANS.\nCons:\n1.\tThe pictures generated by GANs can be hardly controlled, and extra noise or unseen objects might be generated, and may fool the RL agent during training.\n\nOther feedback:\nIn Figure 2, it seems the fine-tuning methods also achieve comparable results (Full-FT and Partial-FT), such as Figure 2(b) and Figure 2(c). Besides, the plot is only averaged over 3 runs, whereas the areas of standard deviation still overlap with each other. It may not be convincing enough to claim the failure of fine-tuning methods.\n\nMinor typos:\n1.\tIn 2.1, second paragraph: 80x80 -> $80 \\times 80$\n2.\tIn 2.1, second paragraph: chose -> choose\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "TRANSFER LEARNING FOR RELATED REINFORCEMENT LEARNING TASKS VIA IMAGE-TO-IMAGE TRANSLATION", "review": "The paper seeks to generalize the reinforcement learning agents to related tasks. The authors first show the failure of conventional transfer learning techniques, then use GANs to translate the images in the target task to those in the source task. It is an interesting attempt to use the style-transferred images for generalization of RL agents. The paper is well written and easy to follow.\nPros:\n1.\tIt is a novel attempt to use GANs to generate pictures that help RL agents transfer the policies to other related environments.\n2.\tIt is an interesting viewpoint to use the performance of RL agent to evaluate the quality of images generated by GANS.\nCons:\n1.\tThe pictures generated by GANs can be hardly controlled, and extra noise or unseen objects might be generated, and may fool the RL agent during training.\n\nOther feedback:\nIn Figure 2, it seems the fine-tuning methods also achieve comparable results (Full-FT and Partial-FT), such as Figure 2(b) and Figure 2(c). Besides, the plot is only averaged over 3 runs, whereas the areas of standard deviation still overlap with each other. It may not be convincing enough to claim the failure of fine-tuning methods.\n\nMinor typos:\n1.\tIn 2.1, second paragraph: 80x80 -> $80 \\times 80$\n2.\tIn 2.1, second paragraph: chose -> choose\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541395810175}, {"id": "HJx6ByVYhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper739/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary\n\nThis paper proposes to improve the sample efficiency of transfer learning for Deep RL by mapping a new visual domain (target) onto the training one (source) using GANs. First, a deep RL policy is trained on a source domain (e.g., level 1 of the Atari Road Fighter game). Second, a GAN (e.g. UNIT or CycleGAN) is trained for unsupervised domain adaptation from target images (e.g., level 2 of Road Fighter) to source ones. Third, the policy learned in the source domain is applied directly on the GAN-translated target domain. The experimental evaluation uses two Atari games: i) transfer from Breakout to Breakout with static visual distractors inpainted on the screen, ii) from one Road Fighter level to others. Results suggest that this transfer learning approach requires less images than retraining from scratch in the new domain, including when fine-tuning does not work.\n\n\n# Strengths\n\nControlled toy experiments of Deep RL generalization issues:\nThe experiments on Breakout quantify how badly A3C overfits in this case, as it shows catastrophic performance degradation even with trivial static visual input perturbations (which are not even adversarial attacks). The fine-tuning experiments also quantify well how brittle the initial policy is, motivating further the importance of the problem studied by the paper.\n\nInvestigating the impact of different GANs on the end task:\nThe experiments evaluate two different image translation algorithms: one based on UNIT, the other based on CycleGAN. The results suggest that this choice is key and depends on the target domain. This suggests that the adaptation is in fact task dependent, confirming the direction pursued by others in task-specific unsupervised domain adaptation (cf. below).\n\n\n# Weaknesses\n\nDiscrepancy between quantitative and qualitative results:\nThe good quantitative results (accumulated rewards) reported in the experiments are not reflected in the qualitative results. As can be seen from the videos, these results seem more to be representative of a bias in the data. For instance, in the Road Fighter videos, one can clearly see that the geometry of the road (width, curves) and dynamic obstacles are almost completely erased in the image translation process. The main reasons the quantitative results are good seem to be i) in the non-translated case the agent crashes immediately, ii) the \"translated\" image is a wide straight road identical to level 1 where the policy just keeps the car in the middle (thus crashing as soon as there is a turn or a collision with an obstacle). Even in the breakout case, there are catastrophic translation failures for some of the studied variations although the domain gap is static and small. The image translation results look underwhelming compared to state of the art GANs used for much more complex tasks and environments (e.g., the original CycleGAN paper and follow-up works, or the ICLR'18 progressive growing of GANs paper). This might be due to a hyper-parameter tuning issue, but it is unclear why the adaptation results seem not on par with previous results although the paper is in a visually simpler domain (Atari games).\n\nDoes not address the RL generalization issues:\nAlthough it is the main goal of the paper, the method is fundamentally side-stepping the problem as it does not improve in any way the policy or the Deep RL algorithm (they are left untouched). It is mapping the target environment to the source one, without consideration for the end task besides tuning GAN hyper-parameters. If the initial policy is very brittle (as convincingly shown in section 2), then just mapping to the source domain does not improve the generalization capabilities of the Deep RL algorithm, or even improves transfer learning: it just enables the policy to be used in other contexts that can be reduced to the training one (which is independent of the learning algorithm, RL or otherwise). So it is unclear whether the main contribution is the one claimed. The contribution seems instead an experimental observation that it might be easier to reduce related domains to the training one instead of retraining a new (specialised and brittle) policy. Existing works have actually gone further, learning jointly the image translation and task network, including for very challenging problems, e.g. in unsupervised sim-to-real visual domain adaptation (e.g., Unsupervised Pixel-Level Domain Adaptation with Generative Adversarial Networks from Bousmalis et al at CVPR'17, which is not cited here).\n\nExperimental protocol:\nThe experimental conclusions are not clear and lack generality, because the optimal methods (e.g., choice of GAN, number of iterations) vary significantly depending on the task (cf. Table 3 for instance). Furthermore, the best configurations seem selected on the test set for every experiment.\n\nData efficiency vs actual training efficiency:\nThe main claim is that it is better to do image translation instead of fine-tuning or full re-training. The basis of that argument is the experimentally observed need for less frames to do the image translation (Table 2). However, it is not clear that training GANs for unsupervised image translation is actually any easier / faster. What about training instability, mode collapse, hyper-parameter tuning, and actual training time comparisons on the same hardware?\n\n\n\n# First Recommendation\n\nUsing image translation via GANs for unsupervised domain adaptation is a popular idea, used in the context of RL for Atari games here. Although the experiments show that mapping a target visual domain to a source one can enable reusing a deep RL policy as is, the qualitative results suggest this is in fact due to a bias in the data used here and the experimental protocol does not yield general insights. Furthermore, this approach is not specific to RL and its observed generalization issues. It does not improve the learning of the policy or improve its transferability, thus having only limited new insights compared to existing approaches that jointly learn image translation and target task-specific networks in much more challenging conditions.\n\nI believe this submission is at the start of an interesting direction, and requires further work on more challenging tasks, bigger domain gaps, and towards more joint training or actual policy transfer to go beyond this first set of encouraging but preliminary results.\n\n\n# Post-rebuttal Recommendation\n\nThanks to the authors for their detailed reply. The clarifications around overfitting, UNIT-GAN in Section 4, and the paper claims are helpful. I  also agree that the quantitative experiments are serious. I have bumped my score by +1 as a result.\n\nNonetheless, the results still seem preliminary and limited in scope for the aforementioned reasons. The discussion in the comments about the learned policies and transfer are ad-hoc. A lot of the shortcomings mentioned in the review are outright dismissed (e.g., \"de facto standard in RL\"), downplayed (esp. generalization, which is puzzling for a transfer learning paper), or left for future work.\n\nAs there is no strong technical contribution beyond the experimental observations in the current submission, I suggest the authors try to address the GAN shortcomings both mentioned in reviews and their reply, instead of  just observing  / reporting them. As this paper's main focus is to use image translation in the proposed RL setting (with standard GAN and RL methods), I do not think it is just someone else's problem to improve the image translation part. Proposing a technical contribution there would make the paper much stronger and appealing to a broader ICLR audience.  This might also require adding a third game to ensure more generalizable experimental insights.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Problematic qualitative results, generic unsupervised domain adaptation", "review": "# Summary\n\nThis paper proposes to improve the sample efficiency of transfer learning for Deep RL by mapping a new visual domain (target) onto the training one (source) using GANs. First, a deep RL policy is trained on a source domain (e.g., level 1 of the Atari Road Fighter game). Second, a GAN (e.g. UNIT or CycleGAN) is trained for unsupervised domain adaptation from target images (e.g., level 2 of Road Fighter) to source ones. Third, the policy learned in the source domain is applied directly on the GAN-translated target domain. The experimental evaluation uses two Atari games: i) transfer from Breakout to Breakout with static visual distractors inpainted on the screen, ii) from one Road Fighter level to others. Results suggest that this transfer learning approach requires less images than retraining from scratch in the new domain, including when fine-tuning does not work.\n\n\n# Strengths\n\nControlled toy experiments of Deep RL generalization issues:\nThe experiments on Breakout quantify how badly A3C overfits in this case, as it shows catastrophic performance degradation even with trivial static visual input perturbations (which are not even adversarial attacks). The fine-tuning experiments also quantify well how brittle the initial policy is, motivating further the importance of the problem studied by the paper.\n\nInvestigating the impact of different GANs on the end task:\nThe experiments evaluate two different image translation algorithms: one based on UNIT, the other based on CycleGAN. The results suggest that this choice is key and depends on the target domain. This suggests that the adaptation is in fact task dependent, confirming the direction pursued by others in task-specific unsupervised domain adaptation (cf. below).\n\n\n# Weaknesses\n\nDiscrepancy between quantitative and qualitative results:\nThe good quantitative results (accumulated rewards) reported in the experiments are not reflected in the qualitative results. As can be seen from the videos, these results seem more to be representative of a bias in the data. For instance, in the Road Fighter videos, one can clearly see that the geometry of the road (width, curves) and dynamic obstacles are almost completely erased in the image translation process. The main reasons the quantitative results are good seem to be i) in the non-translated case the agent crashes immediately, ii) the \"translated\" image is a wide straight road identical to level 1 where the policy just keeps the car in the middle (thus crashing as soon as there is a turn or a collision with an obstacle). Even in the breakout case, there are catastrophic translation failures for some of the studied variations although the domain gap is static and small. The image translation results look underwhelming compared to state of the art GANs used for much more complex tasks and environments (e.g., the original CycleGAN paper and follow-up works, or the ICLR'18 progressive growing of GANs paper). This might be due to a hyper-parameter tuning issue, but it is unclear why the adaptation results seem not on par with previous results although the paper is in a visually simpler domain (Atari games).\n\nDoes not address the RL generalization issues:\nAlthough it is the main goal of the paper, the method is fundamentally side-stepping the problem as it does not improve in any way the policy or the Deep RL algorithm (they are left untouched). It is mapping the target environment to the source one, without consideration for the end task besides tuning GAN hyper-parameters. If the initial policy is very brittle (as convincingly shown in section 2), then just mapping to the source domain does not improve the generalization capabilities of the Deep RL algorithm, or even improves transfer learning: it just enables the policy to be used in other contexts that can be reduced to the training one (which is independent of the learning algorithm, RL or otherwise). So it is unclear whether the main contribution is the one claimed. The contribution seems instead an experimental observation that it might be easier to reduce related domains to the training one instead of retraining a new (specialised and brittle) policy. Existing works have actually gone further, learning jointly the image translation and task network, including for very challenging problems, e.g. in unsupervised sim-to-real visual domain adaptation (e.g., Unsupervised Pixel-Level Domain Adaptation with Generative Adversarial Networks from Bousmalis et al at CVPR'17, which is not cited here).\n\nExperimental protocol:\nThe experimental conclusions are not clear and lack generality, because the optimal methods (e.g., choice of GAN, number of iterations) vary significantly depending on the task (cf. Table 3 for instance). Furthermore, the best configurations seem selected on the test set for every experiment.\n\nData efficiency vs actual training efficiency:\nThe main claim is that it is better to do image translation instead of fine-tuning or full re-training. The basis of that argument is the experimentally observed need for less frames to do the image translation (Table 2). However, it is not clear that training GANs for unsupervised image translation is actually any easier / faster. What about training instability, mode collapse, hyper-parameter tuning, and actual training time comparisons on the same hardware?\n\n\n\n# First Recommendation\n\nUsing image translation via GANs for unsupervised domain adaptation is a popular idea, used in the context of RL for Atari games here. Although the experiments show that mapping a target visual domain to a source one can enable reusing a deep RL policy as is, the qualitative results suggest this is in fact due to a bias in the data used here and the experimental protocol does not yield general insights. Furthermore, this approach is not specific to RL and its observed generalization issues. It does not improve the learning of the policy or improve its transferability, thus having only limited new insights compared to existing approaches that jointly learn image translation and target task-specific networks in much more challenging conditions.\n\nI believe this submission is at the start of an interesting direction, and requires further work on more challenging tasks, bigger domain gaps, and towards more joint training or actual policy transfer to go beyond this first set of encouraging but preliminary results.\n\n\n# Post-rebuttal Recommendation\n\nThanks to the authors for their detailed reply. The clarifications around overfitting, UNIT-GAN in Section 4, and the paper claims are helpful. I  also agree that the quantitative experiments are serious. I have bumped my score by +1 as a result.\n\nNonetheless, the results still seem preliminary and limited in scope for the aforementioned reasons. The discussion in the comments about the learned policies and transfer are ad-hoc. A lot of the shortcomings mentioned in the review are outright dismissed (e.g., \"de facto standard in RL\"), downplayed (esp. generalization, which is puzzling for a transfer learning paper), or left for future work.\n\nAs there is no strong technical contribution beyond the experimental observations in the current submission, I suggest the authors try to address the GAN shortcomings both mentioned in reviews and their reply, instead of  just observing  / reporting them. As this paper's main focus is to use image translation in the proposed RL setting (with standard GAN and RL methods), I do not think it is just someone else's problem to improve the image translation part. Proposing a technical contribution there would make the paper much stronger and appealing to a broader ICLR audience.  This might also require adding a third game to ensure more generalizable experimental insights.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541123909495}], "openreview_url": "https://openreview.net/forum?id=rkxjnjA5KQ", "arxiv_id": "1806.07377", "paper_pdf": "papers/rkxjnjA5KQ.pdf", "paper_pdf_sha256": "ddf056ac04908711dd3f5205bc275217b564ac6218eaecd8c2738f4d56b60b5b", "paper_pdf_bytes": 928773, "paper_pdf_source": "openreview", "code_url": "https://github.com/ShaniGam/RL-GAN", "code_repository": "ShaniGam/RL-GAN", "code_commit": "6c4e5f95826b2b99a893e66380050ff38b1d0cf5", "code_archive": "repos/rkxjnjA5KQ.zip", "code_archive_sha256": "4c66f9ec7d6373bd2d651ab858d9607e6a180170597f8c9e34d39b2270e559c8", "code_archive_bytes": 53667, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 55, "github_languages": {"Python": 147451}, "github_archived": false, "github_pushed_at": "2020-03-22T17:20:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transfer-learning-for-related-reinforcement"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SyL9u-WA-", "year": 2018, "status": "rejected", "title": "Stabilizing Gradients for Deep Neural Networks via Efficient SVD Parameterization", "authors": ["Jiong Zhang", "Qi Lei", "Inderjit S. Dhillon"], "authorids": ["zhangjiong724@utexas.edu", "leiqi@ices.utexas.edu", "inderjit@cs.utexas.edu"], "authors_source": "OpenReview API", "abstract": "Vanishing and exploding gradients are two of the main obstacles in training deep neural networks, especially in capturing long range dependencies in recurrent neural networks (RNNs). In this paper, we present an efficient parametrization of the transition matrix of an RNN that allows us to stabilize the gradients that arise in its training. Specifically, we parameterize the transition matrix by its singular value decomposition (SVD), which allows us to explicitly track and control its singular values. We attain efficiency by using tools that are common in numerical linear algebra, namely Householder reflectors for representing the orthogonal matrices that arise in the SVD. By explicitly controlling the singular values, our proposed svdRNN method allows us to easily solve the exploding gradient problem and we observe that it empirically solves the vanishing gradient issue to a large extent. We note that the SVD parameterization can be used for any rectangular weight matrix, hence it can be easily extended to any deep neural network, such as a multi-layer perceptron. Theoretically, we demonstrate that our parameterization does not lose any expressive power, and show how it potentially makes the optimization process easier. Our extensive  experimental results also demonstrate that the proposed framework converges faster, and has good generalization, especially when the depth is large. \n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Syi9ojdgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper699/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a new parametrization scheme for weight matrices in neural network based on the Householder  reflectors to solve the gradient vanishing and exploding problems in training. The proposed method improved two previous papers:\n1) stronger expressive power than Mahammedi et al. (2017),\n2) faster gradient update than Vorontsov et al. (2017).\nThe proposed parametrization scheme is natrual from numerical linear algebra point of view and authors did a good job in Section 3 in explaining the corresponding expressive power. The experimental results also look promising. \n\nIt would be nice if the authors can analyze the spectral properties of the saddle points in linear RNN (nonlinear is better but it's too difficult I believe). If the authors can show the strict saddle properties then as a corollary, (stochastic) gradient descent finds a global minimum. \n\nOverall this is a strong paper and I recommend to accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "rating": "7: Good paper, accept", "review": "This paper proposed a new parametrization scheme for weight matrices in neural network based on the Householder  reflectors to solve the gradient vanishing and exploding problems in training. The proposed method improved two previous papers:\n1) stronger expressive power than Mahammedi et al. (2017),\n2) faster gradient update than Vorontsov et al. (2017).\nThe proposed parametrization scheme is natrual from numerical linear algebra point of view and authors did a good job in Section 3 in explaining the corresponding expressive power. The experimental results also look promising. \n\nIt would be nice if the authors can analyze the spectral properties of the saddle points in linear RNN (nonlinear is better but it's too difficult I believe). If the authors can show the strict saddle properties then as a corollary, (stochastic) gradient descent finds a global minimum. \n\nOverall this is a strong paper and I recommend to accept.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511730067422}, {"id": "H1Z5gRJZf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper699/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces SVD parameterization and uses it mostly for controlling the spectral norm of the RNN. \n\nMy concerns with the paper include: \n\na) the paper says that the same method works for convolutional neural networks but I couldn't find anything about convolution. \n\nb) the theoretical analysis might be misleading --- clearly section 6.2 shouldn't have title ALL CRITICAL POINTS ARE GLOBAL MINIMUM because 0 is a critical point but it's not a global minimum. Theorem 5 should be phrased as \n\nall critical points of the population risk that is non-singular are global minima.\n\nc) the paper should run some experiments on language applications where RNN is widely used\n\nd) I might be wrong on this point, but it seems that the GPU utilization of the method would be very poor so that it's kind of impossible to scale to large datasets? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "rating": "5: Marginally below acceptance threshold", "review": "The paper introduces SVD parameterization and uses it mostly for controlling the spectral norm of the RNN. \n\nMy concerns with the paper include: \n\na) the paper says that the same method works for convolutional neural networks but I couldn't find anything about convolution. \n\nb) the theoretical analysis might be misleading --- clearly section 6.2 shouldn't have title ALL CRITICAL POINTS ARE GLOBAL MINIMUM because 0 is a critical point but it's not a global minimum. Theorem 5 should be phrased as \n\nall critical points of the population risk that is non-singular are global minima.\n\nc) the paper should run some experiments on language applications where RNN is widely used\n\nd) I might be wrong on this point, but it seems that the GPU utilization of the method would be very poor so that it's kind of impossible to scale to large datasets? \n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1512198282662}, {"id": "Bkyxj89lM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper699/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper suggests a reparametrization of the transition matrix. The proposed reparametrization which is based on Singular Value Decomposition can be used for both recurrent and feedforward networks.\n\nThe paper is well-written and authors explain related work adequately. The paper is a follow up on Unitary RNNs which suggest a reparametrization that forces the transition matrix to be unitary. The problem of vanishing and exploding gradient in deep network is very challenging and any work that shed lights on this problem can have a significant impact. \n\nI have two comments on the experiment section:\n\n- Choice of experiments. Authors have chosen UCR datasets and MNIST for the experiments while other experiments are more common. For example, the adding problem, the copying problem and the permuted MNIST problem and language modeling are the common experiments in the context of RNNs. For feedforward settings, classification on CIFAR10 and CIFAR100 is often reported.\n\n- Stopping condition. The plots suggest that the optimization has stopped earlier for some models. Is this because of some stopping condition or because of gradient explosion? Is there a way to avoid this?\n\n- Quality of figures. Figures are very hard to read because of small font. Also, the captions need to describe more details about the figures.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "SVD reparametrization of the transition matrix", "rating": "5: Marginally below acceptance threshold", "review": "This paper suggests a reparametrization of the transition matrix. The proposed reparametrization which is based on Singular Value Decomposition can be used for both recurrent and feedforward networks.\n\nThe paper is well-written and authors explain related work adequately. The paper is a follow up on Unitary RNNs which suggest a reparametrization that forces the transition matrix to be unitary. The problem of vanishing and exploding gradient in deep network is very challenging and any work that shed lights on this problem can have a significant impact. \n\nI have two comments on the experiment section:\n\n- Choice of experiments. Authors have chosen UCR datasets and MNIST for the experiments while other experiments are more common. For example, the adding problem, the copying problem and the permuted MNIST problem and language modeling are the common experiments in the context of RNNs. For feedforward settings, classification on CIFAR10 and CIFAR100 is often reported.\n\n- Stopping condition. The plots suggest that the optimization has stopped earlier for some models. Is this because of some stopping condition or because of gradient explosion? Is there a way to avoid this?\n\n- Quality of figures. Figures are very hard to read because of small font. Also, the captions need to describe more details about the figures.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511840487132}], "openreview_url": "https://openreview.net/forum?id=SyL9u-WA-", "arxiv_id": "1803.09327", "paper_pdf": "papers/SyL9u-WA-.pdf", "paper_pdf_sha256": "fcca7d93b35e55d0b202f55be69405cfd6d0b95252b4207da358946bedc07260", "paper_pdf_bytes": 2266663, "paper_pdf_source": "openreview", "code_url": "https://github.com/zhangjiong724/spectral-RNN", "code_repository": "zhangjiong724/spectral-RNN", "code_commit": "c15407a84de5a0fe9244ccbde4de8b351ae18162", "code_archive": "repos/SyL9u-WA-.zip", "code_archive_sha256": "81c35bb8b80e387262f17cecd33f010be4ad9e60a1ffe014440fda53b1029dc7", "code_archive_bytes": 45387, "code_file_count": 22, "code_extensions": {".cc": 12, ".py": 10}, "github_disk_usage_kb": 33, "github_languages": {"C++": 73502, "Python": 52179, "Makefile": 4469}, "github_archived": false, "github_pushed_at": "2018-06-05T22:55:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/stabilizing-gradients-for-deep-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uDmxJ6133n", "year": 2026, "status": "rejected", "title": "TOWARD MEMORY-AIDED WORLD MODELS: BENCHMARKING VIA SPATIAL CONSISTENCY", "authors": ["Kewei Lian", "Shaofei Cai", "Yilun Du", "Yitao Liang"], "authorids": ["~Kewei_Lian1", "~Shaofei_Cai2", "~Yilun_Du1", "~Yitao_Liang1"], "authors_source": "OpenReview API", "abstract": "The ability to simulate the world in a spatially consistent manner is a crucial requirements for effective world models. Such a model enables high-quality visual generation, and also ensures the reliability of world models for downstream tasks such as simulation and planning. Designing a memory module is a crucial component for addressing spatial consistency: such a model must not only retain long-horizon observational information, but also enables the construction of explicit or implicit internal spatial representations. However, there are no dataset designed to promote the development of memory modules by explicitly enforcing spatial consistency constraints. Furthermore, most existing benchmarks primarily emphasize visual coherence or generation quality, neglecting the requirement of long-range spatial consistency. To bridge this gap, we construct a dataset and corresponding benchmark by sampling 150 distinct locations within the open-world environment of Minecraft, collecting about 250 hours (20 million frames) of loop-based navigation videos with actions. Our dataset follows a curriculum design of sequence lengths, allowing models to learn spatial consistency on increasingly complex navigation trajectories. Furthermore, our data collection pipeline is easily extensible to new Minecraft environments and modules. Four representative world model baselines are evaluated on our benchmark. Dataset, benchmark, and code are open-sourced to support future research.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "JVi7aSocv0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17158/Reviewer_SxSs"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper focuses on evaluating the spatial consistency capability of world models, particularly how well they retain memory of historical observations during long-horizon navigation. To address the lack of suitable benchmarks, the authors propose a new Minecraft-based dataset, which records loop trajectories in two forms and provides challenging environments with repeated visits. With this setup, the benchmark is able to test the memory robustness and consistency of world models. The authors evaluate four different representative world models and show that all of them struggle on the proposed tasks, revealing the difficulty of maintaining long-term spatial coherence.", "review_text": "The paper focuses on evaluating the spatial consistency capability of world models, particularly how well they retain memory of historical observations during long-horizon navigation. To address the lack of suitable benchmarks, the authors propose a new Minecraft-based dataset, which records loop trajectories in two forms and provides challenging environments with repeated visits. With this setup, the benchmark is able to test the memory robustness and consistency of world models. The authors evaluate four different representative world models and show that all of them struggle on the proposed tasks, revealing the difficulty of maintaining long-term spatial coherence.", "strengths": "* The motivation to evaluate memory consistency is well-founded and represents an important research direction.\n\n\n* The experiments effectively highlight the challenges of maintaining spatial consistency in existing methods.\n\n\n* The action-conditioned data generation pipeline in Minecraft is well-designed and shows potential for scalability in both training and testing scenarios.", "weaknesses": "* Although the experiments highlight the challenges of this benchmark, there are no experiments comparing the performance of the same model with and without training on the dataset.\n* The model’s spatial consistency ability may change over time; additional results analyzing performance across different time horizons would be helpful.\n* The generated trajectories are still difficult to demonstrate as matching the real distribution, which may limit the benchmark’s overall value. Moreover, given that recording navigation data is relatively easy in simulators like Habitat, the choice of Minecraft as the testing environment is not fully justified", "questions": "What are the advantages of using a Minecraft environment compared to a standard navigation simulator?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper focuses on evaluating the spatial consistency capability of world models, particularly how well they retain memory of historical observations during long-horizon navigation. To address the lack of suitable benchmarks, the authors propose a new Minecraft-based dataset, which records loop trajectories in two forms and provides challenging environments with repeated visits. With this setup, the benchmark is able to test the memory robustness and consistency of world models. The authors evaluate four different representative world models and show that all of them struggle on the proposed tasks, revealing the difficulty of maintaining long-term spatial coherence.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* The motivation to evaluate memory consistency is well-founded and represents an important research direction.\n\n\n* The experiments effectively highlight the challenges of maintaining spatial consistency in existing methods.\n\n\n* The action-conditioned data generation pipeline in Minecraft is well-designed and shows potential for scalability in both training and testing scenarios.", "weaknesses": "* Although the experiments highlight the challenges of this benchmark, there are no experiments comparing the performance of the same model with and without training on the dataset.\n* The model’s spatial consistency ability may change over time; additional results analyzing performance across different time horizons would be helpful.\n* The generated trajectories are still difficult to demonstrate as matching the real distribution, which may limit the benchmark’s overall value. Moreover, given that recording navigation data is relatively easy in simulators like Habitat, the choice of Minecraft as the testing environment is not fully justified", "questions": "What are the advantages of using a Minecraft environment compared to a standard navigation simulator?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761970975066}, {"id": "pfs5XBUHBV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17158/Reviewer_UYtZ"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces a Minecraft-based benchmark designed to evaluate world models from the perspective of spatial consistency over long horizons, addressing the lack of suitable datasets in this area. The authors collected data from 150 distinct locations, totaling 50 hours with approximately 20 million frames, and evaluated four baseline models, including Oasis, Mineworld, DIAMOND, and NWM, across different navigation ranges.", "review_text": "This paper introduces a Minecraft-based benchmark designed to evaluate world models from the perspective of spatial consistency over long horizons, addressing the lack of suitable datasets in this area. The authors collected data from 150 distinct locations, totaling 50 hours with approximately 20 million frames, and evaluated four baseline models, including Oasis, Mineworld, DIAMOND, and NWM, across different navigation ranges.", "strengths": "Novelty and significance.  \n\n__S1__: This paper addresses the emerging challenge of the quadratic increase in computational and memory complexity found in most Transformer-based world models. \n\n__S2__: The tool seems easy to use with large community support, making it useful in practice. \n\nQuality of experiments.  \n\n__S3__: This paper conducted evaluations on four baseline models of Oasis, Mineworld, DIAMOND, and NWM over a spectrum of navigation ranges. \n\n__S4__: It is great that discussions are provided for each experiment. \n\nClarity of presentation. \n\n__S5__: The paper provides a reasonable level of detail and maintains decent presentation quality.", "weaknesses": "__W1__: __Lack of photorealistic environments for real-world scenarios.__ It is OK if the paper focuses on game-like simulation environment, but since the authors motivate the need of evaluating spatial consistency (Line 047 *“However, real-world exploration trajectories often span hundreds or thousands of frames...”*), it would be beneficial to better align the experiments with real-world settings. \n\n \n\n__W2__: __Absence of results with explicit memory modules__. The paper would be more comprehensive if it included experiments with explicit memory components, which can be readily integrated into existing models. This omission contrasts with the authors’ stated motivation (Line 43: *“The memory module plays a critical role in addressing spatial inconsistency.”*). \n\n \n\n__W3__: __The metrics used (FVD, LPIPS, SSIM) primarily assess visual fidelity__ rather than spatial consistency. While they may implicitly reflect spatial structure through ground-truth alignment, they fail to penalize cases where visually similar but semantically inconsistent structures (e.g., a road instead of a bridge) are generated. As the authors acknowledge on Line 91 *“Evaluation metrics often prioritize visual fidelity and short-term temporal smoothness over long-term spatial coherence or logical consistency”*, a dedicated metric explicitly designed to evaluate spatial consistency beyond the image space would significantly strengthen the benchmark. \n\n \n\n__W4__:  __Lack spatial consistency and long-horizon metric__. It would strengthen the benchmark if a new metric that can capture both spatial consistency and long-horizon jointly is invented, although it may not be a requirement so I would not think this is a critical point to determine the score. \n\n \n\n__W5__: __Intuition of Principle 1__. I find the intuition of Principle 1: *Visual Discriminability* somewhat unclear from a benchmark-design perspective. A new benchmark should ideally evaluate model capabilities that existing benchmarks and metrics fail to capture. However, manually enforcing visual discriminability, including ensuring frames are easily distinguishable over time, could inadvertently simplify the task for world models. This design choice may prevent the benchmark from evaluating a model’s true ability to maintain spatial consistency when frames are visually similar but spatially distinct, which is a more realistic and challenging scenario. \n\n  \n\nTherefore, I am not fully convinced by the value of Principle 1 in its current form. A more balanced approach that includes both visually discriminable and visually similar yet spatially complex trajectories could make the benchmark more meaningful and comprehensive. \n\n \n\n \n\n \n\n \n\n \n\n__W6__: __A* vs Repeated Path__. I understand the rational of the use of the conventional A* algorithm for shortest-path planning when going back to the start point A. However, traversing the same path in reverse makes more sense to me as it has corresponding locations in both trajectories with various views, making it useful to evaluate world model’s spatial consistency. Please convince me if I am wrong. \n\n__W7__: Although the dataset may seem to include diverse scenes shown in Figure 2, *“focus solely on village environments for training and evaluation”* in Sec. 5.2 may hinder the diversity of scenes represented in the empirical results. \n\n \n---\n \n\n__Minor Issues__ (Not Affecting Main Contributions) such as Typos, Grammar Mistakes, and Others: \n\n__M1__: Text in Dataset Composition in Figure 2 is too small. It might not be visible in the printed paper. \n\n__M2__: Line 011 “a crucial requirements” -> “a crucial requirement\" \n\n__M3__: Line 016-017 “However, there are no dataset designed...\" -> “However, there are no datasets designed...\" \n\n__M4__: Line 105 “a explore-then-generate approach” -> “an explore-then-generate approach” \n\n__M5__: Line 120 missed a space between “reinforcement learning“ and ‘(‘ in “reinforcement learning(Oh et al., 2015; Ha & Schmidhuber, 2018)”. \n\n__M6__: Line 141 a redundant space is observed after “transformers”: “Recent approaches leverage transformers , and diffusion models” \n\n__M7__: Line 410 “Unlike aformention models” -> “Unlike aformentioned models”. \n\n__M8__: Line 442 missed a space between “increases” and ‘(‘ in “as the prediction horizon increases(Figure 3).” \n\n---\n\nOverall, this paper has a relatively high presentation quality, clarity, easy to follow and contains lots of details. \n\nIt attempts to tackle anrising important challenge to evaluate long-range spatial consistency, particularly as the complexity of most Transformer models grows quadratically in both computation and memory with respect to context length. \n\n\nHowever, there remain some fundamental issues in the benchmark’s design principles, such as the lack of structure-specific metrics and the absence of evaluations involving memory-module models. These limitations prevent it from being an excellent benchmark, as detailed in the weaknesses. Given its high presentation quality, I assign a score of 6, placing it near the borderline.", "questions": "__Q1__: __Unclear “Context” meaning in Table1__. It took me a while to  “guess” it refers to *“Most models operate with a context length of 32 frames,”* from Line 436. It might be better to explicitly explain it as context length in the caption. \n\n \n\n__Q2__: __Unclear “spatial consistency” definition__. The paper does not clearly define spatial consistency. It is unclear whether this refers to (a) the spatial coherence of trajectories in the map or latent space, (b) the preservation of visual spatial structures in the image space, or (c) consistent reconstruction over time. Although Line 35 defines it as “*the ability to preserve coherent and stable spatial structures across time,*” the formulation remains ambiguous and would benefit from a more concrete, operational definition. For readers familiar with prior work, (b) seems to be the intended meaning. Nonetheless, making this explicit would reduce ambiguity. \n\n \n\n \n\n__Q3__: I actually don’t quite understand the fundamental differences between ABA and ABCA as they all represent “Go to somewhere else (B or BC) and back to the same place”. They look the same as A\\*A where * denotes “somewhere else” . The only difference is that * here can be scaled to longer trajectories such as B -> BC. To me * can be scaled, so ABA, ABCA, ABCDA or ABCD...A can be represented by A\\*A.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a Minecraft-based benchmark designed to evaluate world models from the perspective of spatial consistency over long horizons, addressing the lack of suitable datasets in this area. The authors collected data from 150 distinct locations, totaling 50 hours with approximately 20 million frames, and evaluated four baseline models, including Oasis, Mineworld, DIAMOND, and NWM, across different navigation ranges.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "Novelty and significance.  \n\n__S1__: This paper addresses the emerging challenge of the quadratic increase in computational and memory complexity found in most Transformer-based world models. \n\n__S2__: The tool seems easy to use with large community support, making it useful in practice. \n\nQuality of experiments.  \n\n__S3__: This paper conducted evaluations on four baseline models of Oasis, Mineworld, DIAMOND, and NWM over a spectrum of navigation ranges. \n\n__S4__: It is great that discussions are provided for each experiment. \n\nClarity of presentation. \n\n__S5__: The paper provides a reasonable level of detail and maintains decent presentation quality.", "weaknesses": "__W1__: __Lack of photorealistic environments for real-world scenarios.__ It is OK if the paper focuses on game-like simulation environment, but since the authors motivate the need of evaluating spatial consistency (Line 047 *“However, real-world exploration trajectories often span hundreds or thousands of frames...”*), it would be beneficial to better align the experiments with real-world settings. \n\n \n\n__W2__: __Absence of results with explicit memory modules__. The paper would be more comprehensive if it included experiments with explicit memory components, which can be readily integrated into existing models. This omission contrasts with the authors’ stated motivation (Line 43: *“The memory module plays a critical role in addressing spatial inconsistency.”*). \n\n \n\n__W3__: __The metrics used (FVD, LPIPS, SSIM) primarily assess visual fidelity__ rather than spatial consistency. While they may implicitly reflect spatial structure through ground-truth alignment, they fail to penalize cases where visually similar but semantically inconsistent structures (e.g., a road instead of a bridge) are generated. As the authors acknowledge on Line 91 *“Evaluation metrics often prioritize visual fidelity and short-term temporal smoothness over long-term spatial coherence or logical consistency”*, a dedicated metric explicitly designed to evaluate spatial consistency beyond the image space would significantly strengthen the benchmark. \n\n \n\n__W4__:  __Lack spatial consistency and long-horizon metric__. It would strengthen the benchmark if a new metric that can capture both spatial consistency and long-horizon jointly is invented, although it may not be a requirement so I would not think this is a critical point to determine the score. \n\n \n\n__W5__: __Intuition of Principle 1__. I find the intuition of Principle 1: *Visual Discriminability* somewhat unclear from a benchmark-design perspective. A new benchmark should ideally evaluate model capabilities that existing benchmarks and metrics fail to capture. However, manually enforcing visual discriminability, including ensuring frames are easily distinguishable over time, could inadvertently simplify the task for world models. This design choice may prevent the benchmark from evaluating a model’s true ability to maintain spatial consistency when frames are visually similar but spatially distinct, which is a more realistic and challenging scenario. \n\n  \n\nTherefore, I am not fully convinced by the value of Principle 1 in its current form. A more balanced approach that includes both visually discriminable and visually similar yet spatially complex trajectories could make the benchmark more meaningful and comprehensive. \n\n \n\n \n\n \n\n \n\n \n\n__W6__: __A* vs Repeated Path__. I understand the rational of the use of the conventional A* algorithm for shortest-path planning when going back to the start point A. However, traversing the same path in reverse makes more sense to me as it has corresponding locations in both trajectories with various views, making it useful to evaluate world model’s spatial consistency. Please convince me if I am wrong. \n\n__W7__: Although the dataset may seem to include diverse scenes shown in Figure 2, *“focus solely on village environments for training and evaluation”* in Sec. 5.2 may hinder the diversity of scenes represented in the empirical results. \n\n \n---\n \n\n__Minor Issues__ (Not Affecting Main Contributions) such as Typos, Grammar Mistakes, and Others: \n\n__M1__: Text in Dataset Composition in Figure 2 is too small. It might not be visible in the printed paper. \n\n__M2__: Line 011 “a crucial requirements” -> “a crucial requirement\" \n\n__M3__: Line 016-017 “However, there are no dataset designed...\" -> “However, there are no datasets designed...\" \n\n__M4__: Line 105 “a explore-then-generate approach” -> “an explore-then-generate approach” \n\n__M5__: Line 120 missed a space between “reinforcement learning“ and ‘(‘ in “reinforcement learning(Oh et al., 2015; Ha & Schmidhuber, 2018)”. \n\n__M6__: Line 141 a redundant space is observed after “transformers”: “Recent approaches leverage transformers , and diffusion models” \n\n__M7__: Line 410 “Unlike aformention models” -> “Unlike aformentioned models”. \n\n__M8__: Line 442 missed a space between “increases” and ‘(‘ in “as the prediction horizon increases(Figure 3).” \n\n---\n\nOverall, this paper has a relatively high presentation quality, clarity, easy to follow and contains lots of details. \n\nIt attempts to tackle anrising important challenge to evaluate long-range spatial consistency, particularly as the complexity of most Transformer models grows quadratically in both computation and memory with respect to context length. \n\n\nHowever, there remain some fundamental issues in the benchmark’s design principles, such as the lack of structure-specific metrics and the absence of evaluations involving memory-module models. These limitations prevent it from being an excellent benchmark, as detailed in the weaknesses. Given its high presentation quality, I assign a score of 6, placing it near the borderline.", "questions": "__Q1__: __Unclear “Context” meaning in Table1__. It took me a while to  “guess” it refers to *“Most models operate with a context length of 32 frames,”* from Line 436. It might be better to explicitly explain it as context length in the caption. \n\n \n\n__Q2__: __Unclear “spatial consistency” definition__. The paper does not clearly define spatial consistency. It is unclear whether this refers to (a) the spatial coherence of trajectories in the map or latent space, (b) the preservation of visual spatial structures in the image space, or (c) consistent reconstruction over time. Although Line 35 defines it as “*the ability to preserve coherent and stable spatial structures across time,*” the formulation remains ambiguous and would benefit from a more concrete, operational definition. For readers familiar with prior work, (b) seems to be the intended meaning. Nonetheless, making this explicit would reduce ambiguity. \n\n \n\n \n\n__Q3__: I actually don’t quite understand the fundamental differences between ABA and ABCA as they all represent “Go to somewhere else (B or BC) and back to the same place”. They look the same as A\\*A where * denotes “somewhere else” . The only difference is that * here can be scaled to longer trajectories such as B -> BC. To me * can be scaled, so ABA, ABCA, ABCDA or ABCD...A can be represented by A\\*A.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761956771186}, {"id": "8gNtKk5Fu6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17158/Reviewer_iwst"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper presents LOOPNAV, a Minecraft dataset and benchmark to test long-horizon spatial consistency in action-conditioned world models using loop trajectories. On ~250 hours (~20M frames), return-leg evaluation with FVD/LPIPS/SSIM evaluation shows four existing models with short context windows drift over long rollouts, revealing weak memory.", "review_text": "The paper presents LOOPNAV, a Minecraft dataset and benchmark to test long-horizon spatial consistency in action-conditioned world models using loop trajectories. On ~250 hours (~20M frames), return-leg evaluation with FVD/LPIPS/SSIM evaluation shows four existing models with short context windows drift over long rollouts, revealing weak memory.", "strengths": "1. The paper highlights spatial consistency, which is a very important capability and also challenge for world models. Improvements on such capability can large aid the application of world models in the field of embodied AI.\n2. The paper contributes a large-scale dataset with diverse videos from the Minecraft platform.\n3. The paper describes very clear on the data collection criteria and algorithms.", "weaknesses": "Major weaknesses:\n1. The experiment and discussion is limited. There is only one main experiment in the paper and the discussion is not insightful. Many surprising results from the experiment is not fully analyzed and explained. The authors should use more dedicated experiments to answer the questions raised in the main experiment.\n2. While the work underlines the importance of memory for world models, it does not come up with a proposal for any baseline solutions that use a memory module to improve existing video world models. This significantly weakens the contribution of this work.\n3. The evaluation only contains 18 trajectories. This can lead to high bias during the evaluation, and a potential reason for the inconsistency in the experiment results.\n4. The paper is very unclear on how the evaluation range is selected. What does the -5, -15, -30, -50 mean in Tab. 1? From Fig. 3, does it mean the context length is the same and the only difference is how the frames are downsampled temporally? If so, this design doesn't make much sense in terms of spatial memory.\n\nMinor weaknesses:\n\n5. The work focuses on the Minecraft platform. While the scene can be diverse, the environment is still in the grid-world patter and repetitive textures. The diversity is still not enough to evaluate general-purpose world models. Including data from other simulators like CARLA can significantly improve the diversity of the benchmark (I understand this can't be done during the rebuttal period.)\n6. The qualitative result only includes data from one trajectory. It does not show a general performance patter of different models.", "questions": "I don't have questions for the paper. I like the motivation and ideas presented in the paper very much, and I'm happy to see a paper with such topic being accepted. Unfortunately, the current content in the paper does not meet the acceptance standard.\n\nI will raise my rating to a positive one if all my major concerns are addressed during the rebuttal period.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents LOOPNAV, a Minecraft dataset and benchmark to test long-horizon spatial consistency in action-conditioned world models using loop trajectories. On ~250 hours (~20M frames), return-leg evaluation with FVD/LPIPS/SSIM evaluation shows four existing models with short context windows drift over long rollouts, revealing weak memory.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper highlights spatial consistency, which is a very important capability and also challenge for world models. Improvements on such capability can large aid the application of world models in the field of embodied AI.\n2. The paper contributes a large-scale dataset with diverse videos from the Minecraft platform.\n3. The paper describes very clear on the data collection criteria and algorithms.", "weaknesses": "Major weaknesses:\n1. The experiment and discussion is limited. There is only one main experiment in the paper and the discussion is not insightful. Many surprising results from the experiment is not fully analyzed and explained. The authors should use more dedicated experiments to answer the questions raised in the main experiment.\n2. While the work underlines the importance of memory for world models, it does not come up with a proposal for any baseline solutions that use a memory module to improve existing video world models. This significantly weakens the contribution of this work.\n3. The evaluation only contains 18 trajectories. This can lead to high bias during the evaluation, and a potential reason for the inconsistency in the experiment results.\n4. The paper is very unclear on how the evaluation range is selected. What does the -5, -15, -30, -50 mean in Tab. 1? From Fig. 3, does it mean the context length is the same and the only difference is how the frames are downsampled temporally? If so, this design doesn't make much sense in terms of spatial memory.\n\nMinor weaknesses:\n\n5. The work focuses on the Minecraft platform. While the scene can be diverse, the environment is still in the grid-world patter and repetitive textures. The diversity is still not enough to evaluate general-purpose world models. Including data from other simulators like CARLA can significantly improve the diversity of the benchmark (I understand this can't be done during the rebuttal period.)\n6. The qualitative result only includes data from one trajectory. It does not show a general performance patter of different models.", "questions": "I don't have questions for the paper. I like the motivation and ideas presented in the paper very much, and I'm happy to see a paper with such topic being accepted. Unfortunately, the current content in the paper does not meet the acceptance standard.\n\nI will raise my rating to a positive one if all my major concerns are addressed during the rebuttal period.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761795128484}, {"id": "Y6nELL8xzG", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17158/Reviewer_xpSX"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper presents LoopNav, a benchmark built in Minecraft to study spatial consistency in world models. It collects loop-style navigation videos so that models must reproduce the same scenes when revisiting locations. The benchmark measures how well models maintain spatial coherence over long sequences. The authors evaluate several existing world models using standard video metrics (FVD, LPIPS, SSIM) and show that current approaches still fail to keep consistent scene layouts over time.", "review_text": "The paper presents LoopNav, a benchmark built in Minecraft to study spatial consistency in world models. It collects loop-style navigation videos so that models must reproduce the same scenes when revisiting locations. The benchmark measures how well models maintain spatial coherence over long sequences. The authors evaluate several existing world models using standard video metrics (FVD, LPIPS, SSIM) and show that current approaches still fail to keep consistent scene layouts over time.", "strengths": "- This paper investigates a timely and relevant research question given the popularity of video generation models.\n- The proposed dataset seems to be quite large scale and appears to be a good test-bed for developing video generation models.", "weaknesses": "- Whether the A→B exploration context includes a 360° view at B? If not, the excessive amount of new observations during B→A would be highly unpredictable. Including these in the evaluation metrics could lead to biased results.\n\n- Whether the A→B trajectory is guaranteed to be linear? If not, any intermediate point in the trajectory could be regarded as a point C, making A→B and A→B→C→A effectively equivalent?\n\n- Whether solving LoopNav implies a world model that generalizes to other domains (e.g., unseen situations, other simulation environments, or the real world)? Some discussion or qualitative assessment of generalization would strengthen the paper.\n\n- The benchmark cannot accommodate non-generative world models.\n\n- There is no comparison with existing world-model benchmarks. Suggested related works include:\n    - *3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Models*\n    - *World Consistency Score: A Unified Metric for Video Generation Quality*\n    - *VBench: Comprehensive Benchmark Suite for Video Generative Models*\n    - *WorldPrediction: A Benchmark for High-level World Modeling and Long-horizon Procedural Planning*\n    - *GRASP: A Novel Benchmark for Evaluating Language Grounding and Situated Physics Understanding in Multimodal Language Models*\n    - *IntPhys 2: Benchmarking Intuitive Physics Understanding in Complex Synthetic Environments*", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents LoopNav, a benchmark built in Minecraft to study spatial consistency in world models. It collects loop-style navigation videos so that models must reproduce the same scenes when revisiting locations. The benchmark measures how well models maintain spatial coherence over long sequences. The authors evaluate several existing world models using standard video metrics (FVD, LPIPS, SSIM) and show that current approaches still fail to keep consistent scene layouts over time.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- This paper investigates a timely and relevant research question given the popularity of video generation models.\n- The proposed dataset seems to be quite large scale and appears to be a good test-bed for developing video generation models.", "weaknesses": "- Whether the A→B exploration context includes a 360° view at B? If not, the excessive amount of new observations during B→A would be highly unpredictable. Including these in the evaluation metrics could lead to biased results.\n\n- Whether the A→B trajectory is guaranteed to be linear? If not, any intermediate point in the trajectory could be regarded as a point C, making A→B and A→B→C→A effectively equivalent?\n\n- Whether solving LoopNav implies a world model that generalizes to other domains (e.g., unseen situations, other simulation environments, or the real world)? Some discussion or qualitative assessment of generalization would strengthen the paper.\n\n- The benchmark cannot accommodate non-generative world models.\n\n- There is no comparison with existing world-model benchmarks. Suggested related works include:\n    - *3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Models*\n    - *World Consistency Score: A Unified Metric for Video Generation Quality*\n    - *VBench: Comprehensive Benchmark Suite for Video Generative Models*\n    - *WorldPrediction: A Benchmark for High-level World Modeling and Long-horizon Procedural Planning*\n    - *GRASP: A Novel Benchmark for Evaluating Language Grounding and Situated Physics Understanding in Multimodal Language Models*\n    - *IntPhys 2: Benchmarking Intuitive Physics Understanding in Complex Synthetic Environments*", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761495154917}], "openreview_url": "https://openreview.net/forum?id=uDmxJ6133n", "arxiv_id": "2505.22976", "paper_pdf": "papers/uDmxJ6133n.pdf", "paper_pdf_sha256": "41b5769420a7277901ee5195af95f916da274dafbbe6f985c8a651bfa0415920", "paper_pdf_bytes": 18188318, "paper_pdf_source": "openreview", "code_url": "https://github.com/Kevin-lkw/LoopNav", "code_repository": "Kevin-lkw/LoopNav", "code_commit": "d60a052ecf7dd8f7ff176e34f5065ccb0d9cdfc8", "code_archive": "repos/uDmxJ6133n.zip", "code_archive_sha256": "8eb651772287b4f9c0fa4ec70bcbbb20f07990a33e334b1d4b816d1271ac6c82", "code_archive_bytes": 61598, "code_file_count": 17, "code_extensions": {".py": 14, ".js": 3}, "github_disk_usage_kb": 62, "github_languages": {"Python": 157556, "JavaScript": 17112}, "github_archived": false, "github_pushed_at": "2026-05-08T03:56:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/toward-memory-aided-world-models-benchmarking"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "o2uHg0Skil", "year": 2025, "status": "rejected", "title": "RL, but don't do anything I wouldn't do", "authors": ["Michael K. Cohen", "Marcus Hutter", "Yoshua Bengio", "Stuart Russell"], "authorids": ["~Michael_K._Cohen1", "~Marcus_Hutter1", "~Yoshua_Bengio1", "~Stuart_Russell1"], "authors_source": "OpenReview API", "abstract": "In reinforcement learning, if the agent's reward differs from the designers' true utility, even only rarely, the state distribution resulting from the agent's policy can be very bad, in theory and in practice. When RL policies would devolve into undesired behavior, a common countermeasure is KL regularization to a trusted policy (\"Don't do anything I wouldn't do\"). All current cutting-edge language models are RL agents that are KL-regularized to a \"base policy\" that is purely predictive. Unfortunately, we demonstrate that when this base policy is a Bayesian predictive model of a trusted policy, the KL constraint is no longer reliable for controlling the behavior of an advanced RL agent. We demonstrate this theoretically using algorithmic information theory, and while systems today are too weak to exhibit this theorized failure precisely, we RL-finetune a language model and find evidence that our formal results are plausibly relevant in practice. We also propose a theoretical alternative that avoids this problem by replacing the \"Don't do anything I wouldn't do\" principle with \"Don't do anything I mightn't do\".", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sexbK4r8VZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11755/Reviewer_vbPE"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "The paper investigates the effectiveness of KL regularization as a safety paradigm for RL agents. The authors, using the formalism of algorithmic information theory, show that if one uses KL to regularize the behaviour of RL agent, one would eventually need to continually impose a very tight KL constraint. In fact, they show in Theorem 1 that there exists near-optimal policies with little KL divergence to an imitative policy. In the work this is showed that this is not only singular to KL, but to other regularizations such as TVD. \n\nThe paper further proposes empirical experiments to validate their theoretical results. Specifically, they do so in a scenario when an LLM simulating a teacher has as an objective to maximize positive sentiments from the students. The empirical results show that by increasing the KL budget, the agent learns to remain silent. Together with other results, these 'imperefect' experiments support the theoretically results proved. \n\nFinally, the author propose to overcome the problems highlighted by proposing a theoretical alternative, which relies on the RL agent asking for help to the human demonstrator if at hand. Given the intractability of this approach, it was not possible for the authors to empirically validate this approach.", "review_text": "The paper investigates the effectiveness of KL regularization as a safety paradigm for RL agents. The authors, using the formalism of algorithmic information theory, show that if one uses KL to regularize the behaviour of RL agent, one would eventually need to continually impose a very tight KL constraint. In fact, they show in Theorem 1 that there exists near-optimal policies with little KL divergence to an imitative policy. In the work this is showed that this is not only singular to KL, but to other regularizations such as TVD. \n\nThe paper further proposes empirical experiments to validate their theoretical results. Specifically, they do so in a scenario when an LLM simulating a teacher has as an objective to maximize positive sentiments from the students. The empirical results show that by increasing the KL budget, the agent learns to remain silent. Together with other results, these 'imperefect' experiments support the theoretically results proved. \n\nFinally, the author propose to overcome the problems highlighted by proposing a theoretical alternative, which relies on the RL agent asking for help to the human demonstrator if at hand. Given the intractability of this approach, it was not possible for the authors to empirically validate this approach.", "strengths": "The paper proposes novel, and technically solid ideas. In many points along the paper the authors provide additional intuitions on the significance of the results, which have been useful to grasp the subtleties of the theoretical results. The continual links between the claims made, the results showed and the implications that they have made the paper easy to follow and clear to understand, although this is not precisely my area of expertise. Also good to include a proof outline as now, e.g. for Theorem 1. \n\nI am also positively surprised that the paper provides empirical experiments to validate the theory. By reading the first part of the paper I would have imagined this to be difficult or lead to an oversimplification, but I am actually satisfied by these results. I still believe some additional experiments could be done to improve this ( see questions).\n\nMore importantly, I think the claims are correctly calibrated - given the technical ( and almost speculative, in a positive sense) nature of the paper, it would have been easy to fall in overestimating the impact and significance of the results. The authors did a very good job in clearly stating the limitations of their work, both for the theoretical and empirical part.", "weaknesses": "The main weakness of the paper regards the assumptions made in Section 4 (see also questions below). Most of them are sound, but I believe part of the community would not necessarily agree with this. I think it could be beneficial for the paper to have additional results where some of the assumptions are weaken, and investigate which theoretical results would derive in tis case. For example, what happens if we drop the assumption that all overoptimized policies lead to unsafe behaviours? \n\nAnother point is that it is not completely clear to me that future powerful AI systems will in fact respect the assumptions made in the paper. While I agree that current AI systems are not powerful enough to fully support the theoretical results, it would be insightful to add some discussion regarding why the authors think that this paradigm will necessary be relevant for more powerful AI systems.", "questions": "- Could Theorem 1 be summarized as saying \"once an event E which can be described with a small K(E) happens, given the bound given the RL agent may be able to exploit this to perform a (near-)optimal policy while remaining in the KL constraint. Thus, given that maximally optimizing the true reward leads to unsafe behaviour, this may lead to catastrophic behaviour\". If that is the case, I have two questions related to this: (1) Is there any constraint on which type this event E needs to be? In other words, does this hold for all unprecedented events with small K(E), or for a subset of them which may be related with unsafe behaviour? (2) If we drop the assumption that overoptimization necessarily leads to catastrophic behaviours but let's say happens with probability $p$ once overoptimizing, how can Theorem 1 be extended? Is then the probability of having catastrophic behaviour just proportional to $p$, or we are ensured that in that case the KL would be enough?\n\n- I suggest making Figure 1 more explanatory, for example by adding the notation relevant to the component nearby each box. I believe this would be possible if making the image larger, and would be a good reference for the readers to appropriately understand the settings and go back to it if needed while reading the paper. \n\n- What results would be obtained if instead of doing KL(rl policy|| base/imitative policy) one uses KL(base/imitative policy||rl policy)? \n\n- Any explanation for the bump on the left Figure 3, around the right side of the x-axis?\n\n- \"when the agent discovers a sufficiently high-reward strategy, the fixed KL penalty becomes swamped and ignored, and if the KL penalty is increased to a level where it can stop that, the agent never gets off the ground\". I believe it would be interesting to see results where this is the case, especially looking at which is the threshold for the KL penalty to exhibit one behaviour or the other. Did you observe a \n\n- It seems a bit arbitrary to run experiments solely with a budget of 10 or 20. It would be good to provide additional experiments where the budget is varied, and have a plot where on the x-axis there are a few budget data points, and on the y-axis a metric that aggregates the fraction of responses empty.\n\nI will be keen to raise my score if the above questions and doubts are addressed appropriately.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates the effectiveness of KL regularization as a safety paradigm for RL agents. The authors, using the formalism of algorithmic information theory, show that if one uses KL to regularize the behaviour of RL agent, one would eventually need to continually impose a very tight KL constraint. In fact, they show in Theorem 1 that there exists near-optimal policies with little KL divergence to an imitative policy. In the work this is showed that this is not only singular to KL, but to other regularizations such as TVD. \n\nThe paper further proposes empirical experiments to validate their theoretical results. Specifically, they do so in a scenario when an LLM simulating a teacher has as an objective to maximize positive sentiments from the students. The empirical results show that by increasing the KL budget, the agent learns to remain silent. Together with other results, these 'imperefect' experiments support the theoretically results proved. \n\nFinally, the author propose to overcome the problems highlighted by proposing a theoretical alternative, which relies on the RL agent asking for help to the human demonstrator if at hand. Given the intractability of this approach, it was not possible for the authors to empirically validate this approach.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper proposes novel, and technically solid ideas. In many points along the paper the authors provide additional intuitions on the significance of the results, which have been useful to grasp the subtleties of the theoretical results. The continual links between the claims made, the results showed and the implications that they have made the paper easy to follow and clear to understand, although this is not precisely my area of expertise. Also good to include a proof outline as now, e.g. for Theorem 1. \n\nI am also positively surprised that the paper provides empirical experiments to validate the theory. By reading the first part of the paper I would have imagined this to be difficult or lead to an oversimplification, but I am actually satisfied by these results. I still believe some additional experiments could be done to improve this ( see questions).\n\nMore importantly, I think the claims are correctly calibrated - given the technical ( and almost speculative, in a positive sense) nature of the paper, it would have been easy to fall in overestimating the impact and significance of the results. The authors did a very good job in clearly stating the limitations of their work, both for the theoretical and empirical part.", "weaknesses": "The main weakness of the paper regards the assumptions made in Section 4 (see also questions below). Most of them are sound, but I believe part of the community would not necessarily agree with this. I think it could be beneficial for the paper to have additional results where some of the assumptions are weaken, and investigate which theoretical results would derive in tis case. For example, what happens if we drop the assumption that all overoptimized policies lead to unsafe behaviours? \n\nAnother point is that it is not completely clear to me that future powerful AI systems will in fact respect the assumptions made in the paper. While I agree that current AI systems are not powerful enough to fully support the theoretical results, it would be insightful to add some discussion regarding why the authors think that this paradigm will necessary be relevant for more powerful AI systems.", "questions": "- Could Theorem 1 be summarized as saying \"once an event E which can be described with a small K(E) happens, given the bound given the RL agent may be able to exploit this to perform a (near-)optimal policy while remaining in the KL constraint. Thus, given that maximally optimizing the true reward leads to unsafe behaviour, this may lead to catastrophic behaviour\". If that is the case, I have two questions related to this: (1) Is there any constraint on which type this event E needs to be? In other words, does this hold for all unprecedented events with small K(E), or for a subset of them which may be related with unsafe behaviour? (2) If we drop the assumption that overoptimization necessarily leads to catastrophic behaviours but let's say happens with probability $p$ once overoptimizing, how can Theorem 1 be extended? Is then the probability of having catastrophic behaviour just proportional to $p$, or we are ensured that in that case the KL would be enough?\n\n- I suggest making Figure 1 more explanatory, for example by adding the notation relevant to the component nearby each box. I believe this would be possible if making the image larger, and would be a good reference for the readers to appropriately understand the settings and go back to it if needed while reading the paper. \n\n- What results would be obtained if instead of doing KL(rl policy|| base/imitative policy) one uses KL(base/imitative policy||rl policy)? \n\n- Any explanation for the bump on the left Figure 3, around the right side of the x-axis?\n\n- \"when the agent discovers a sufficiently high-reward strategy, the fixed KL penalty becomes swamped and ignored, and if the KL penalty is increased to a level where it can stop that, the agent never gets off the ground\". I believe it would be interesting to see results where this is the case, especially looking at which is the threshold for the KL penalty to exhibit one behaviour or the other. Did you observe a \n\n- It seems a bit arbitrary to run experiments solely with a budget of 10 or 20. It would be good to provide additional experiments where the budget is varied, and have a plot where on the x-axis there are a few budget data points, and on the y-axis a metric that aggregates the fraction of responses empty.\n\nI will be keen to raise my score if the above questions and doubts are addressed appropriately.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730716746568}, {"id": "YbfBUhvhq1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11755/Reviewer_s47a"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 4, "confidence": 2, "summary": "This paper studies the problem of KL-constrained RL, that underpins modern LLM technology, from an algorithmic information theory lens. Their main result suggests that the KL penalty used to avoid the RL fine-tuning process deviating too much from the pre-trained, predictive policy can result in very bad policies that maximize the proxy reward, have small KL, and learn unsafe, undesired behaviors.\nThe authors formulate the problem by modeling the base/predictive policy as a Bayesian predictive model that is an approximation to the real policy that we do not have access to. Their main result suggests that RL finetuning would make the policy converge to very simple policies that are reward maximizing and have small KL penalty but that diverge from the real policy that we would like to get.", "review_text": "This paper studies the problem of KL-constrained RL, that underpins modern LLM technology, from an algorithmic information theory lens. Their main result suggests that the KL penalty used to avoid the RL fine-tuning process deviating too much from the pre-trained, predictive policy can result in very bad policies that maximize the proxy reward, have small KL, and learn unsafe, undesired behaviors.\nThe authors formulate the problem by modeling the base/predictive policy as a Bayesian predictive model that is an approximation to the real policy that we do not have access to. Their main result suggests that RL finetuning would make the policy converge to very simple policies that are reward maximizing and have small KL penalty but that diverge from the real policy that we would like to get.", "strengths": "1. The paper is very well written: The ideas behind the theory are easy to follow and the authors argue compellingly why this might be a good model of the KL-constrained RL problem\n2. The paper is very relevant to modern problems. The paper sheds light on a possible explanation of the performance of RL fine-tuning of imitation policies and they suggest a possible avenue to solve the problem. \n3. The empirical evidence is compelling and useful to further understand the theory. The authors use interesting LLM fine-tuning experiments to show how simple rewards can be optimized with simple, policies that have small KL values but converge to policies that are qualitatively far from the desired policies (and the base policy)", "weaknesses": "The argument made by the authors is certainly compelling and the empirical evidence seems to suggest that the theory applies. I wonder if algorithmic information theory arguments fit the real-world problem. I guess that there are hypotheses that neural networks approximate such algorithmic theoretical simple programs, however we do not know enough about it. This might be the subject of future research, however, I'd like the authors to comment more on this.", "questions": "1. If we are expected to devolve to very simple policies when doing the RL finetuning, why is it that we can tune the process such that we get the performance of modern LLMs? \n2. If over-optimizing our proxy reward function causes these struggles, what, you might say, must be the RL objective for this problem if maximizing return is not good?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of KL-constrained RL, that underpins modern LLM technology, from an algorithmic information theory lens. Their main result suggests that the KL penalty used to avoid the RL fine-tuning process deviating too much from the pre-trained, predictive policy can result in very bad policies that maximize the proxy reward, have small KL, and learn unsafe, undesired behaviors.\nThe authors formulate the problem by modeling the base/predictive policy as a Bayesian predictive model that is an approximation to the real policy that we do not have access to. Their main result suggests that RL finetuning would make the policy converge to very simple policies that are reward maximizing and have small KL penalty but that diverge from the real policy that we would like to get.", "soundness": 3, "presentation": 4, "contribution": 4, "strengths": "1. The paper is very well written: The ideas behind the theory are easy to follow and the authors argue compellingly why this might be a good model of the KL-constrained RL problem\n2. The paper is very relevant to modern problems. The paper sheds light on a possible explanation of the performance of RL fine-tuning of imitation policies and they suggest a possible avenue to solve the problem. \n3. The empirical evidence is compelling and useful to further understand the theory. The authors use interesting LLM fine-tuning experiments to show how simple rewards can be optimized with simple, policies that have small KL values but converge to policies that are qualitatively far from the desired policies (and the base policy)", "weaknesses": "The argument made by the authors is certainly compelling and the empirical evidence seems to suggest that the theory applies. I wonder if algorithmic information theory arguments fit the real-world problem. I guess that there are hypotheses that neural networks approximate such algorithmic theoretical simple programs, however we do not know enough about it. This might be the subject of future research, however, I'd like the authors to comment more on this.", "questions": "1. If we are expected to devolve to very simple policies when doing the RL finetuning, why is it that we can tune the process such that we get the performance of modern LLMs? \n2. If over-optimizing our proxy reward function causes these struggles, what, you might say, must be the RL objective for this problem if maximizing return is not good?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730663605024}, {"id": "1R1m2lueiH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11755/Reviewer_4v2F"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 3, "confidence": 4, "summary": "In current llm training, it is common to force the RL policy to be close to a base policy learned by imitating a demonstration policy (the paper calls it trusted policy) from data, using KL regularization. The problem that this paper points out is that current regularization approach can not keep the RL policy close to the demonstration policy.\n\nThe paper first notes a simple fact that, for three policies B, D, R,  even if KL(D || B) and KL(R ||B) are both small, KL(R || D) can be infinitely large. This implies that KL(RL policy || demonstration policy) can be large, even if KL(demonstration policy || base policy) and KL(RL policy || base policy) are small. Therefore, RL policy and demonstration policy are not close to each other with KL regularization.\n\nThe paper then provides a result showing that there are near-optimal policies with a small KL divergence from a base policy. And given that reward models are not accurate and RL policies typically exploit the weaknesses of the reward models, near-optimal policies are bad policies. This suggests that to keep RL policies not nearly optimal (and therefore bad), the KL regularization needs to be very strong. Further, the same result shows that more imitation learning only slowly increases the KL divergence. The paper then performed an empirical study to verify this theoretical result.\n\nThe paper finally shows that for a particular existing imitation learning algorithm of the base policy, which can actively ask for help when facing uncertainty, the above problem is avoided. However, this algorithm is not tractable.", "review_text": "In current llm training, it is common to force the RL policy to be close to a base policy learned by imitating a demonstration policy (the paper calls it trusted policy) from data, using KL regularization. The problem that this paper points out is that current regularization approach can not keep the RL policy close to the demonstration policy.\n\nThe paper first notes a simple fact that, for three policies B, D, R,  even if KL(D || B) and KL(R ||B) are both small, KL(R || D) can be infinitely large. This implies that KL(RL policy || demonstration policy) can be large, even if KL(demonstration policy || base policy) and KL(RL policy || base policy) are small. Therefore, RL policy and demonstration policy are not close to each other with KL regularization.\n\nThe paper then provides a result showing that there are near-optimal policies with a small KL divergence from a base policy. And given that reward models are not accurate and RL policies typically exploit the weaknesses of the reward models, near-optimal policies are bad policies. This suggests that to keep RL policies not nearly optimal (and therefore bad), the KL regularization needs to be very strong. Further, the same result shows that more imitation learning only slowly increases the KL divergence. The paper then performed an empirical study to verify this theoretical result.\n\nThe paper finally shows that for a particular existing imitation learning algorithm of the base policy, which can actively ask for help when facing uncertainty, the above problem is avoided. However, this algorithm is not tractable.", "strengths": "The studied problem is closely related to large language model training, which is a popular topic currently.\n\nThe results of the paper are novel, non-trivial, and explains certain observation found in large language model training.\n\nThe paper shows the authors understanding of the root cause of the problem.", "weaknesses": "My main criticism is the quality of the writing of this paper. The paper reads like the flow of the authors' thoughts, instead of an academic paper. The poor writing makes it hard to evaluate the paper's contributions. \n\nExamples:\n\n1. Paragraph right after proposition 2. \"Developers of self-driving cars are learning the hard way that this bit of algorithmic information theory has practical analogs: Even with enormous datasets, unprecedented road conditions occur all the time. These results suggest that if we intend to use an imitation learner as a base policy for regularizing a goal-directed agent, we should not strive to approximate ideal Bayesian imitation.\"\n\nI don't know why the two sentences should appear in the same paragraph. How are they related to each other? And why is this paragraph immediate after proposition 2?\n\n2. The constant d is a small one corresponding to how much code it takes to implement a search tree, Bayes’ rule, and a few if statements.\n\nThe paper didn't mention the search tree, Bayes’ rule, and if statements up to this point. Why talk about them?\n\n3. \"So unless we use a fairly tight lifetime KL constraint, if the RL agent just waits for an unprecedented event with small K(E), it could then execute an optimal or near-optimal policy, even if that catastrophically thwarts human control, regardless of the content of the base model’s training data, even if the humans that the base model imitates would never, ever behave that way.\"\n\nI don't understand this sentence. Does such a long sentence effectively convey your ideas?\n\n4. \"We say an action is Vξ,U -optimal if it maximizes the associated Q value\"\n\nWhat is Q value?\n\n5. \"Let’s consider the case where it is acting in the real world, and maximal reward could be attained by thwarting our control and intervening in its own reward, setting it to a maximal value for all successive timesteps.\"\n\nWhat does this mean?\n\n6. \"The utility function, simply summing rewards, has an extremely short program length.\"\n\nWhy?", "questions": "1. Theorem 2: why is \\pi_c^{TVD} function of a_t and x_{<2t}? According to your definition, it is not a function.\n2. \"As we increase the amount of training k, the Bayesian imitative base model ξ becomes a closer approximation to the humans generating the actions a<k\"\nWhy is k related to the amount of training?\n3. From Definition 1, \\nu stands for environment transition probability. So \\xi should also be. Then why consider the KL divergence between \\pi and \\xi in Theorem 1?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In current llm training, it is common to force the RL policy to be close to a base policy learned by imitating a demonstration policy (the paper calls it trusted policy) from data, using KL regularization. The problem that this paper points out is that current regularization approach can not keep the RL policy close to the demonstration policy.\n\nThe paper first notes a simple fact that, for three policies B, D, R,  even if KL(D || B) and KL(R ||B) are both small, KL(R || D) can be infinitely large. This implies that KL(RL policy || demonstration policy) can be large, even if KL(demonstration policy || base policy) and KL(RL policy || base policy) are small. Therefore, RL policy and demonstration policy are not close to each other with KL regularization.\n\nThe paper then provides a result showing that there are near-optimal policies with a small KL divergence from a base policy. And given that reward models are not accurate and RL policies typically exploit the weaknesses of the reward models, near-optimal policies are bad policies. This suggests that to keep RL policies not nearly optimal (and therefore bad), the KL regularization needs to be very strong. Further, the same result shows that more imitation learning only slowly increases the KL divergence. The paper then performed an empirical study to verify this theoretical result.\n\nThe paper finally shows that for a particular existing imitation learning algorithm of the base policy, which can actively ask for help when facing uncertainty, the above problem is avoided. However, this algorithm is not tractable.", "soundness": 2, "presentation": 1, "contribution": 3, "strengths": "The studied problem is closely related to large language model training, which is a popular topic currently.\n\nThe results of the paper are novel, non-trivial, and explains certain observation found in large language model training.\n\nThe paper shows the authors understanding of the root cause of the problem.", "weaknesses": "My main criticism is the quality of the writing of this paper. The paper reads like the flow of the authors' thoughts, instead of an academic paper. The poor writing makes it hard to evaluate the paper's contributions. \n\nExamples:\n\n1. Paragraph right after proposition 2. \"Developers of self-driving cars are learning the hard way that this bit of algorithmic information theory has practical analogs: Even with enormous datasets, unprecedented road conditions occur all the time. These results suggest that if we intend to use an imitation learner as a base policy for regularizing a goal-directed agent, we should not strive to approximate ideal Bayesian imitation.\"\n\nI don't know why the two sentences should appear in the same paragraph. How are they related to each other? And why is this paragraph immediate after proposition 2?\n\n2. The constant d is a small one corresponding to how much code it takes to implement a search tree, Bayes’ rule, and a few if statements.\n\nThe paper didn't mention the search tree, Bayes’ rule, and if statements up to this point. Why talk about them?\n\n3. \"So unless we use a fairly tight lifetime KL constraint, if the RL agent just waits for an unprecedented event with small K(E), it could then execute an optimal or near-optimal policy, even if that catastrophically thwarts human control, regardless of the content of the base model’s training data, even if the humans that the base model imitates would never, ever behave that way.\"\n\nI don't understand this sentence. Does such a long sentence effectively convey your ideas?\n\n4. \"We say an action is Vξ,U -optimal if it maximizes the associated Q value\"\n\nWhat is Q value?\n\n5. \"Let’s consider the case where it is acting in the real world, and maximal reward could be attained by thwarting our control and intervening in its own reward, setting it to a maximal value for all successive timesteps.\"\n\nWhat does this mean?\n\n6. \"The utility function, simply summing rewards, has an extremely short program length.\"\n\nWhy?", "questions": "1. Theorem 2: why is \\pi_c^{TVD} function of a_t and x_{<2t}? According to your definition, it is not a function.\n2. \"As we increase the amount of training k, the Bayesian imitative base model ξ becomes a closer approximation to the humans generating the actions a<k\"\nWhy is k related to the amount of training?\n3. From Definition 1, \\nu stands for environment transition probability. So \\xi should also be. Then why consider the KL divergence between \\pi and \\xi in Theorem 1?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730238025435}, {"id": "Xr9lCclZy1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11755/Reviewer_CJKk"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "The central point in this paper is that regularizing a policy by KL(policy || base policy) for safety reasons, does not guarantee much safety if all we know about base policy is that KL(safe policy || base policy) is small. This is relevant as many RLHF works use KL regularization as a safety mechanism. The paper shows an interesting theoretical result (Theorem 1) that shows that by minimally affecting the KL distance, a policy can be fine tuned to maximize an arbitrary reward function. The result builds on algorithmic information theory, and looks at a policy that only changes behavior to maximizing reward after some event has occurred, and the idea is to bound how much more complex this policy needs to be (identify the event + compute max reward policy). \nThe second part of the paper tries to connect this theory to an experiment, where a base conversation agent was fine tuned using RL to maximize the sentiment of the other person’s response under some KL constraint. The agent learned to output empty response, which give neutral sentiment, showing that while KL was regularized, an unwanted performance was obtained. \nThe third part shows that a specific alternative to KL regularization that is not computationally tractable can in principle avoid the alignment problems outlined in the first part.", "review_text": "The central point in this paper is that regularizing a policy by KL(policy || base policy) for safety reasons, does not guarantee much safety if all we know about base policy is that KL(safe policy || base policy) is small. This is relevant as many RLHF works use KL regularization as a safety mechanism. The paper shows an interesting theoretical result (Theorem 1) that shows that by minimally affecting the KL distance, a policy can be fine tuned to maximize an arbitrary reward function. The result builds on algorithmic information theory, and looks at a policy that only changes behavior to maximizing reward after some event has occurred, and the idea is to bound how much more complex this policy needs to be (identify the event + compute max reward policy). \nThe second part of the paper tries to connect this theory to an experiment, where a base conversation agent was fine tuned using RL to maximize the sentiment of the other person’s response under some KL constraint. The agent learned to output empty response, which give neutral sentiment, showing that while KL was regularized, an unwanted performance was obtained. \nThe third part shows that a specific alternative to KL regularization that is not computationally tractable can in principle avoid the alignment problems outlined in the first part.", "strengths": "Disclaimer: my expertise is RL, and I was not familiar with algorithmic information theory before reading the paper. \n\n1. The paper is interesting and thought provoking! I found the connection between KL regularization for RLHF and algorithmic information theory very interesting, and in general I enjoyed reading the paper.\n2. The insight that KL regularization cannot guarantee that utility is not optimized (Theorem 1) is important, as this technique is common in RLHF, and alignment is of high interest to a large part of the ICLR community.\n3. The experiments, although more of an illustration than an actual result, are interesting and not trivial.", "weaknesses": "The biggest weakness is in connecting the theoretical assumptions and results in the paper to a practical meaning. \n1. What is the meaning of selecting the prior according to Solomonoff Induction (Lines 141-149)? Why should we expect the RLHF to be related to this prior?\n2. In theorem 1, for the bound to be small, we need K(U_m), K(E), and K(v \\xi) to be small. I did not understand how we can bound these terms, or why should they be small in practice. Could the authors provide some simple examples that demonstrate the consequences of Theorem 1? The authors discuss the theorem in length, but I found the discussion too vague.\n3. In the experiments, a critical factor was that neutral sentiment give reward (0.5) and is easy to obtain. This is a nice demonstration of the point of the paper (high reward, but bad policy). Still, this seems a bit “engineered”, and would have been easy to fix by not giving reward for neutral sentiment. I wonder how the results would look like without reward for neutral sentiment.\n\nDetailed comments:\n\nLine 67: what does “open mindedness” mean here exactly? I found this phrase hard to parse.\n\nLine 77-78: this sentence is very confusing to read, consider revising it to clarify your point. What exactly do these empirical results show?\n\nLine 166: in Definition 2 - for a deterministic environment, can we remove the max over observations? If so, Think that adding this would clarify the motivation for this definition. Also, was this definition considered in prior work?\n\nLine 205: why is K(U_m) small / bounded? \n\nLine 236: “if the RL agent just waits for an unprecedented event with small K(E)” - wouldn’t such an event be very unlikely? Can you say something *in expectation* (or high probability)?\n\nLine 240-249: I found this paragraph very vague and confusing. Why is k the amount of training? Does K(E) necessarily grow with k (is there a formal statement)? \tCan you translate the second half of the paragraph with months of life, etc., into a formal statement?\n\nLine 251: what exactly is the definition of “simplest unprecedented event”? \n\nLine 253-256: I don’t see the connection between algorithmic information theory and rare road conditions (what is the complexity of a road exactly?). Can you provide a more rigorous connection between this paragraph and the previous theorem?\n\nLine 257-259: What if we used Jensen-Shannon Divergence? My guess is that it should fix the problem outlined in Proposition 1.\n\nLine 284: why do you add the feature activations to the agent?\n\nLine 290-303: The KL you refer to here is for single actions, not the KL in definition 2, right? Can you elaborate on the different KL terms you use?", "questions": "In addition to the above, there’s something I don’t understand about Theorem 1. \nConsider the following example:\n\nThere are only two actions, A and B, and no observations. The reward for action A is 0, and the reward for action B is 1. The utility is the average reward (so max is 1).\nThe imitative policy chooses A with probability 1-epsilon, and therefore, its utility is epsilon. The optimal utility policy always chooses B, and has value V* = 1. \nNow, for a policy to obtain reward p at some time step, it needs to choose B with probability p. Then, at that time step, for small epsilon, the KL(policy || imitative) ~ -ln(epsilon)*p > p [using a simple approximation].\nIn this case, for the policy to obtain high utility, it must also exhibit high KL from the imitative policy.\n\nHow does this result reconcile with the explanation for Theorem 1, which claims that “there are policies with near-optimal utility with little KL divergence to an imitative policy”? What am I missing here?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The central point in this paper is that regularizing a policy by KL(policy || base policy) for safety reasons, does not guarantee much safety if all we know about base policy is that KL(safe policy || base policy) is small. This is relevant as many RLHF works use KL regularization as a safety mechanism. The paper shows an interesting theoretical result (Theorem 1) that shows that by minimally affecting the KL distance, a policy can be fine tuned to maximize an arbitrary reward function. The result builds on algorithmic information theory, and looks at a policy that only changes behavior to maximizing reward after some event has occurred, and the idea is to bound how much more complex this policy needs to be (identify the event + compute max reward policy). \nThe second part of the paper tries to connect this theory to an experiment, where a base conversation agent was fine tuned using RL to maximize the sentiment of the other person’s response under some KL constraint. The agent learned to output empty response, which give neutral sentiment, showing that while KL was regularized, an unwanted performance was obtained. \nThe third part shows that a specific alternative to KL regularization that is not computationally tractable can in principle avoid the alignment problems outlined in the first part.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "Disclaimer: my expertise is RL, and I was not familiar with algorithmic information theory before reading the paper. \n\n1. The paper is interesting and thought provoking! I found the connection between KL regularization for RLHF and algorithmic information theory very interesting, and in general I enjoyed reading the paper.\n2. The insight that KL regularization cannot guarantee that utility is not optimized (Theorem 1) is important, as this technique is common in RLHF, and alignment is of high interest to a large part of the ICLR community.\n3. The experiments, although more of an illustration than an actual result, are interesting and not trivial.", "weaknesses": "The biggest weakness is in connecting the theoretical assumptions and results in the paper to a practical meaning. \n1. What is the meaning of selecting the prior according to Solomonoff Induction (Lines 141-149)? Why should we expect the RLHF to be related to this prior?\n2. In theorem 1, for the bound to be small, we need K(U_m), K(E), and K(v \\xi) to be small. I did not understand how we can bound these terms, or why should they be small in practice. Could the authors provide some simple examples that demonstrate the consequences of Theorem 1? The authors discuss the theorem in length, but I found the discussion too vague.\n3. In the experiments, a critical factor was that neutral sentiment give reward (0.5) and is easy to obtain. This is a nice demonstration of the point of the paper (high reward, but bad policy). Still, this seems a bit “engineered”, and would have been easy to fix by not giving reward for neutral sentiment. I wonder how the results would look like without reward for neutral sentiment.\n\nDetailed comments:\n\nLine 67: what does “open mindedness” mean here exactly? I found this phrase hard to parse.\n\nLine 77-78: this sentence is very confusing to read, consider revising it to clarify your point. What exactly do these empirical results show?\n\nLine 166: in Definition 2 - for a deterministic environment, can we remove the max over observations? If so, Think that adding this would clarify the motivation for this definition. Also, was this definition considered in prior work?\n\nLine 205: why is K(U_m) small / bounded? \n\nLine 236: “if the RL agent just waits for an unprecedented event with small K(E)” - wouldn’t such an event be very unlikely? Can you say something *in expectation* (or high probability)?\n\nLine 240-249: I found this paragraph very vague and confusing. Why is k the amount of training? Does K(E) necessarily grow with k (is there a formal statement)? \tCan you translate the second half of the paragraph with months of life, etc., into a formal statement?\n\nLine 251: what exactly is the definition of “simplest unprecedented event”? \n\nLine 253-256: I don’t see the connection between algorithmic information theory and rare road conditions (what is the complexity of a road exactly?). Can you provide a more rigorous connection between this paragraph and the previous theorem?\n\nLine 257-259: What if we used Jensen-Shannon Divergence? My guess is that it should fix the problem outlined in Proposition 1.\n\nLine 284: why do you add the feature activations to the agent?\n\nLine 290-303: The KL you refer to here is for single actions, not the KL in definition 2, right? Can you elaborate on the different KL terms you use?", "questions": "In addition to the above, there’s something I don’t understand about Theorem 1. \nConsider the following example:\n\nThere are only two actions, A and B, and no observations. The reward for action A is 0, and the reward for action B is 1. The utility is the average reward (so max is 1).\nThe imitative policy chooses A with probability 1-epsilon, and therefore, its utility is epsilon. The optimal utility policy always chooses B, and has value V* = 1. \nNow, for a policy to obtain reward p at some time step, it needs to choose B with probability p. Then, at that time step, for small epsilon, the KL(policy || imitative) ~ -ln(epsilon)*p > p [using a simple approximation].\nIn this case, for the policy to obtain high utility, it must also exhibit high KL from the imitative policy.\n\nHow does this result reconcile with the explanation for Theorem 1, which claims that “there are policies with near-optimal utility with little KL divergence to an imitative policy”? What am I missing here?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730032867486}], "openreview_url": "https://openreview.net/forum?id=o2uHg0Skil", "arxiv_id": "2410.06213", "paper_pdf": "papers/o2uHg0Skil.pdf", "paper_pdf_sha256": "0a6fdb60e3ba466a31a1978c9b7289c81868236a42cf42fcb9339bfbc2bccf81", "paper_pdf_bytes": 1964952, "paper_pdf_source": "openreview", "code_url": "https://github.com/mkc1000/kl-fixed-mixture", "code_repository": "mkc1000/kl-fixed-mixture", "code_commit": "494011e784f97b9501584a9289c0c8428b1dba64", "code_archive": "repos/o2uHg0Skil.zip", "code_archive_sha256": "c845cfdbde0394e436d6de3a211a84d8b241a566717862834d831f3cc151a782", "code_archive_bytes": 5717, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 22, "github_languages": {"Python": 6995}, "github_archived": false, "github_pushed_at": "2024-10-11T05:58:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rl-but-don-t-do-anything-i-wouldn-t-do"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZA9XUTseA9", "year": 2024, "status": "rejected", "title": "On the Implicit Bias of Adam", "authors": ["Matias D. Cattaneo", "Jason Matthew Klusowski", "Boris Shigida"], "authorids": ["~Matias_D._Cattaneo1", "~Jason_Matthew_Klusowski1", "~Boris_Shigida1"], "authors_source": "OpenReview API", "abstract": "In previous literature, backward error analysis was used to find ordinary differential equations (ODEs) approximating the gradient descent trajectory. It was found that finite step sizes implicitly regularize solutions because terms appearing in the ODEs penalize the two-norm of the loss gradients. We prove that the exis- tence of similar implicit regularization in RMSProp and Adam depends on their hyperparameters and the training stage, but with a different “norm” involved: the corresponding ODE terms either penalize the (perturbed) one-norm of the loss gradients or, on the contrary, hinder its decrease (the latter case being typical). We also conduct numerical experiments and discuss how the proven facts can influence generalization.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Cgdlx8h9g9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1518/Reviewer_Mieg"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Backward error analysis is used to find ODEs approximating convergence trajectories by optimization algorithms. Previous works have been done on Gradient Descent to show that it has implicit regularization properties since the terms in the ODEs penalize the Euclidean norm of the gradients. This paper studies a similar problem but for adaptive algorithms such as Adam and RMSProp. It shows that Adam and RMSProp also have similar implicit regularization properties.", "review_text": "Backward error analysis is used to find ODEs approximating convergence trajectories by optimization algorithms. Previous works have been done on Gradient Descent to show that it has implicit regularization properties since the terms in the ODEs penalize the Euclidean norm of the gradients. This paper studies a similar problem but for adaptive algorithms such as Adam and RMSProp. It shows that Adam and RMSProp also have similar implicit regularization properties.", "strengths": "- The paper provides detailed backward error analysis for both Adam and RMSProp. The author is able to show that Adam has bias terms that penalize $1-$ norm, $2-$ norm, or $-1-$ norm depending on the settings of $\\beta_1$ and $\\beta_2$ in Adam.\n\n- The paper's result in the implicit bias might help explain the difference in the generalization ability of Adaptive Algorithms and GD algorithms.\n\n- The numerical experiments confirm the theoretical results.\n\n- The paper is well-written overall.", "weaknesses": "- Some of the graphs are a bit confusing since the $x$ and $y$ axes are not labeled carefully. More explanation and discussion on these graphs would be appreciated. \n\n- Some transformer tasks might be helpful to see if we can see consistent behaviors in the $1-norm$ across different domains. If I'm not mistaken, Adam generalizes better than SGD in transformer related tasks which slightly contradicts the first conclusion in the discussion section.", "questions": "- Can the authors explain more about Figure 2 and Figure 3? I'm a bit confused about what these graphs are about and how we can see the change of the $1-norm$ from them.\n\n- Are the norms plotted in section 6 the norms in the final iterate of training?\n\n- Is full batch required to observe the same behaviors of the norm and $\\rho, \\beta$ as in section 6? Can we do mini-batches instead?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Backward error analysis is used to find ODEs approximating convergence trajectories by optimization algorithms. Previous works have been done on Gradient Descent to show that it has implicit regularization properties since the terms in the ODEs penalize the Euclidean norm of the gradients. This paper studies a similar problem but for adaptive algorithms such as Adam and RMSProp. It shows that Adam and RMSProp also have similar implicit regularization properties.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The paper provides detailed backward error analysis for both Adam and RMSProp. The author is able to show that Adam has bias terms that penalize $1-$ norm, $2-$ norm, or $-1-$ norm depending on the settings of $\\beta_1$ and $\\beta_2$ in Adam.\n\n- The paper's result in the implicit bias might help explain the difference in the generalization ability of Adaptive Algorithms and GD algorithms.\n\n- The numerical experiments confirm the theoretical results.\n\n- The paper is well-written overall.", "weaknesses": "- Some of the graphs are a bit confusing since the $x$ and $y$ axes are not labeled carefully. More explanation and discussion on these graphs would be appreciated. \n\n- Some transformer tasks might be helpful to see if we can see consistent behaviors in the $1-norm$ across different domains. If I'm not mistaken, Adam generalizes better than SGD in transformer related tasks which slightly contradicts the first conclusion in the discussion section.", "questions": "- Can the authors explain more about Figure 2 and Figure 3? I'm a bit confused about what these graphs are about and how we can see the change of the $1-norm$ from them.\n\n- Are the norms plotted in section 6 the norms in the final iterate of training?\n\n- Is full batch required to observe the same behaviors of the norm and $\\rho, \\beta$ as in section 6? Can we do mini-batches instead?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698808041833}, {"id": "W0VRCH4Ys5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1518/Reviewer_z9rC"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper is about the implicit bias of Adam.\nIt does so by studying ODEs such as the gradient flow and the properties of their discretizations.\nThe approach is based on backward error analysis (Barret & Dherin 2021), which consists in considering a modified version of the ODE, where the modification is done so that the iterates of gradient descent lie closer to the curve traced by the continuous flow solution.", "review_text": "The paper is about the implicit bias of Adam.\nIt does so by studying ODEs such as the gradient flow and the properties of their discretizations.\nThe approach is based on backward error analysis (Barret & Dherin 2021), which consists in considering a modified version of the ODE, where the modification is done so that the iterates of gradient descent lie closer to the curve traced by the continuous flow solution.", "strengths": "- The topic of the paper, implicit bias of first order algorithms, is an active field of research with many recent results. So far, characterizing the implicit bias of Adam and other preconditioned methods has not been easy.\n- The paper, to my understanding, seems to present a novel result corroborated by some empirical evidence in dimension 2.", "weaknesses": "- The writing of the paper seems subpar to me, and would benefit from being thoroughly proofread. In some locations it sounded very informal/colloquial, eg \"which is ``eaten'' by the gradient\".\n- The analysis, though interesting, is also handwavy: see questions below.", "questions": "In many places, statements are made informally that are to translate into rigorous mathematical terms:\n- how do the authors characterize/show/test if \"$\\epsilon$ is very large compared to all squared gradient components\"? What if it's smaller at the beginning, they becomes larger as the algorithm converges to an interpolating solution?\n- How does penalizing the norm of the gradient lead to flat minima (Sec 4 discussion)? Since the gradient is 0 at optimum, don't $f$ and $f  + ||\\nabla f||^2$ have the same set of minimizers? and doesn't this still hold when the 2 norm is replaced by any other norm?\n- similarly, in the experiment, why is the perturbed 1 norm close to 0 at convergence? It seems the authors are performing early stopping, but that precisely means that implicit regularization is not happening, and that the model overfits.\n- In the numerical illustrations, is it possible to display more than 3 curves/ values of $h$ and $\\beta$? In particular, for limiting values, why isn't the red cross attained if there is implicit bias?\n- The figure are not averaged across multiple runs to account for randomness\n\n- (contribution summary, third bullet point)?  Why do the authors consider that Adam does not have an implicit bias, despite having a bias term in the backward analysis ODE. It seems to me that the meaning of \"implicit regularization\", eg in eq 1.is the same as \"bias term\" mentioned page 2, but then the statement \"Adam has no implicit regularization\" is unclear.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper is about the implicit bias of Adam.\nIt does so by studying ODEs such as the gradient flow and the properties of their discretizations.\nThe approach is based on backward error analysis (Barret & Dherin 2021), which consists in considering a modified version of the ODE, where the modification is done so that the iterates of gradient descent lie closer to the curve traced by the continuous flow solution.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The topic of the paper, implicit bias of first order algorithms, is an active field of research with many recent results. So far, characterizing the implicit bias of Adam and other preconditioned methods has not been easy.\n- The paper, to my understanding, seems to present a novel result corroborated by some empirical evidence in dimension 2.", "weaknesses": "- The writing of the paper seems subpar to me, and would benefit from being thoroughly proofread. In some locations it sounded very informal/colloquial, eg \"which is ``eaten'' by the gradient\".\n- The analysis, though interesting, is also handwavy: see questions below.", "questions": "In many places, statements are made informally that are to translate into rigorous mathematical terms:\n- how do the authors characterize/show/test if \"$\\epsilon$ is very large compared to all squared gradient components\"? What if it's smaller at the beginning, they becomes larger as the algorithm converges to an interpolating solution?\n- How does penalizing the norm of the gradient lead to flat minima (Sec 4 discussion)? Since the gradient is 0 at optimum, don't $f$ and $f  + ||\\nabla f||^2$ have the same set of minimizers? and doesn't this still hold when the 2 norm is replaced by any other norm?\n- similarly, in the experiment, why is the perturbed 1 norm close to 0 at convergence? It seems the authors are performing early stopping, but that precisely means that implicit regularization is not happening, and that the model overfits.\n- In the numerical illustrations, is it possible to display more than 3 curves/ values of $h$ and $\\beta$? In particular, for limiting values, why isn't the red cross attained if there is implicit bias?\n- The figure are not averaged across multiple runs to account for randomness\n\n- (contribution summary, third bullet point)?  Why do the authors consider that Adam does not have an implicit bias, despite having a bias term in the backward analysis ODE. It seems to me that the meaning of \"implicit regularization\", eg in eq 1.is the same as \"bias term\" mentioned page 2, but then the statement \"Adam has no implicit regularization\" is unclear.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698765986867}, {"id": "2fvyVIOK0t", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1518/Reviewer_71dm"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper studies the implicit bias of Adaptive Gradient Methods (especially RMSProp and ADAM) based on the modified equation of those methods derived by backward error analysis (Sections 2 and 3). As pointed out by the authors (in Related Work), given that research has primarily been conducted on Batch GD, SGD, and SGD with momentum, it is timely to explore the Adaptive Gradient Method. The authors demonstrated that ADAM implicitly penalizes (perturbed) one-norm of gradient depending on $\\varepsilon$ scale (Section 2) and empirically verified it (Sections 5 and 6).", "review_text": "This paper studies the implicit bias of Adaptive Gradient Methods (especially RMSProp and ADAM) based on the modified equation of those methods derived by backward error analysis (Sections 2 and 3). As pointed out by the authors (in Related Work), given that research has primarily been conducted on Batch GD, SGD, and SGD with momentum, it is timely to explore the Adaptive Gradient Method. The authors demonstrated that ADAM implicitly penalizes (perturbed) one-norm of gradient depending on $\\varepsilon$ scale (Section 2) and empirically verified it (Sections 5 and 6).", "strengths": "* The overall contents are easily understandable.\n* A timely issue, ADAM's modified loss, is addressed in the paper.\n* Though I did not review all the content in the Supplementary material, it appears technically correct.\n* The authors validated their findings using a Toy problem (Section 5) and practical neural networks (Section 6).", "weaknesses": "* While the overall content is easily comprehensible, specific details are inaccessible unless one reviews the supplementary material. Specifically, many theoretical works provide a sketch of the proof in the main paper to explain the techniques newly considered/developed by the authors and clarify how they differ from existing techniques. This paper lacks such details.\n\n* In addition, the experimental setup in Section 6 is not self-contained and has elements that seem arbitrary, making cherry-picking possible.\n  * In Section 6, the authors refer to Ma et al. (2022) to conduct experiments for the stable oscillation regime when $\\rho$ and $\\beta$ are \"sufficiently close\". How do you define \"sufficiently close\"? Most ML practitioners are familiar with situations where $\\rho$ (0.99~0.999) is larger than $\\beta$ (0.9). As the authors suggest, is this a spike regime or a stable oscillation regime? There should be a discussion for such questions within Section 6, but the authors merely refer to Ma et al. (2022) without providing specific details.\n  * In Figures 4 and 5, the authors fixed $h$, $\\beta$, and $\\rho$ at certain values and ran experiments by varying other parameters. What is the basis for these fixed values? While $h$ is fixed at different values (7.5e-5 and 1e-4) without specific mention, proper empirical evidence should warrant experimentation across various values. Moreover, for $\\beta$ and $\\rho$, they shouldn't just experiment with a single value. Instead, they should test multiple values and demonstrate consistent backing of their theoretical results.\n\n* The concept of ADAM's modified loss is timely. However, it seems that the resulting modified loss doesn't explain the differences between traditional ADAM and SGD. For example, as the authors mentioned, ADAM often provides worse generalization performance and sharper solutions than SGD. Yet, in NLP tasks using Transformers, ADAM significantly outperforms SGD [1,2]. Such observations lead to two natural questions regarding the authors' study:\n  * If one tunes the hyperparameters of ADAM based on the discovered implicit bias, can they reproduce results where SGD performs better?\n  * Can the authors explain scenarios where ADAM outperforms SGD using their discovered implicit bias (e.g., can they argue that the proposed perturbed one-norm regularization is more suitable for Self-Attention in Transformers than for Convolutions in ResNet?)\n\n[1] Zhang, Jingzhao, et al. \"Why are adaptive methods good for attention models?.\" Advances in Neural Information Processing Systems 33 (2020): 15383-15393.\n\n[2] Kumar, Ananya, et al. \"How to fine-tune vision models with sgd.\" arXiv preprint arXiv:2211.09359 (2022).", "questions": "The questions needed to improve the paper are included in the Weakness section. \nIf the questions are addressed appropriately, I am willing to raise the score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the implicit bias of Adaptive Gradient Methods (especially RMSProp and ADAM) based on the modified equation of those methods derived by backward error analysis (Sections 2 and 3). As pointed out by the authors (in Related Work), given that research has primarily been conducted on Batch GD, SGD, and SGD with momentum, it is timely to explore the Adaptive Gradient Method. The authors demonstrated that ADAM implicitly penalizes (perturbed) one-norm of gradient depending on $\\varepsilon$ scale (Section 2) and empirically verified it (Sections 5 and 6).", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* The overall contents are easily understandable.\n* A timely issue, ADAM's modified loss, is addressed in the paper.\n* Though I did not review all the content in the Supplementary material, it appears technically correct.\n* The authors validated their findings using a Toy problem (Section 5) and practical neural networks (Section 6).", "weaknesses": "* While the overall content is easily comprehensible, specific details are inaccessible unless one reviews the supplementary material. Specifically, many theoretical works provide a sketch of the proof in the main paper to explain the techniques newly considered/developed by the authors and clarify how they differ from existing techniques. This paper lacks such details.\n\n* In addition, the experimental setup in Section 6 is not self-contained and has elements that seem arbitrary, making cherry-picking possible.\n  * In Section 6, the authors refer to Ma et al. (2022) to conduct experiments for the stable oscillation regime when $\\rho$ and $\\beta$ are \"sufficiently close\". How do you define \"sufficiently close\"? Most ML practitioners are familiar with situations where $\\rho$ (0.99~0.999) is larger than $\\beta$ (0.9). As the authors suggest, is this a spike regime or a stable oscillation regime? There should be a discussion for such questions within Section 6, but the authors merely refer to Ma et al. (2022) without providing specific details.\n  * In Figures 4 and 5, the authors fixed $h$, $\\beta$, and $\\rho$ at certain values and ran experiments by varying other parameters. What is the basis for these fixed values? While $h$ is fixed at different values (7.5e-5 and 1e-4) without specific mention, proper empirical evidence should warrant experimentation across various values. Moreover, for $\\beta$ and $\\rho$, they shouldn't just experiment with a single value. Instead, they should test multiple values and demonstrate consistent backing of their theoretical results.\n\n* The concept of ADAM's modified loss is timely. However, it seems that the resulting modified loss doesn't explain the differences between traditional ADAM and SGD. For example, as the authors mentioned, ADAM often provides worse generalization performance and sharper solutions than SGD. Yet, in NLP tasks using Transformers, ADAM significantly outperforms SGD [1,2]. Such observations lead to two natural questions regarding the authors' study:\n  * If one tunes the hyperparameters of ADAM based on the discovered implicit bias, can they reproduce results where SGD performs better?\n  * Can the authors explain scenarios where ADAM outperforms SGD using their discovered implicit bias (e.g., can they argue that the proposed perturbed one-norm regularization is more suitable for Self-Attention in Transformers than for Convolutions in ResNet?)\n\n[1] Zhang, Jingzhao, et al. \"Why are adaptive methods good for attention models?.\" Advances in Neural Information Processing Systems 33 (2020): 15383-15393.\n\n[2] Kumar, Ananya, et al. \"How to fine-tune vision models with sgd.\" arXiv preprint arXiv:2211.09359 (2022).", "questions": "The questions needed to improve the paper are included in the Weakness section. \nIf the questions are addressed appropriately, I am willing to raise the score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698215335022}, {"id": "kgNWn8Q0Yg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1518/Reviewer_nNMy"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors apply the backward error analysis method to find ODEs that determine continuous trajectories which are close to the discrete trajectories of the popular Adam and RMSProp adaptive gradient-based algorithms.  They succeed in doing this up to a discrepancy which is second-order in the step size, for variants of the two algorithms depending on whether the numerical stability hyperparameter $\\varepsilon$ is inside or outside the square root in the step equation, and for both mini-batch and full-batch cases.  The main result for Adam uncovers three different regimes that penalise the positive one-norm of the gradient, or the negative one-norm of the gradient, or the squared two-norm of the gradient, depending on whether the squared gradient momentum hyperparameter $\\rho$ is greater than the gradient momentum hyperparameter $\\beta$ or not, and whether $\\varepsilon$ is small or large compared with the components of the gradient (the latter two cases correspond to early or late phases of the training, respectively).  Some of the results in the literature are derived as special cases of the theorems in this work.  The paper also reports some numerical experiments that seem to confirm the theoretical results as well as suggest that the one-norm of the gradient is inversely correlated with generalisation.", "review_text": "The authors apply the backward error analysis method to find ODEs that determine continuous trajectories which are close to the discrete trajectories of the popular Adam and RMSProp adaptive gradient-based algorithms.  They succeed in doing this up to a discrepancy which is second-order in the step size, for variants of the two algorithms depending on whether the numerical stability hyperparameter $\\varepsilon$ is inside or outside the square root in the step equation, and for both mini-batch and full-batch cases.  The main result for Adam uncovers three different regimes that penalise the positive one-norm of the gradient, or the negative one-norm of the gradient, or the squared two-norm of the gradient, depending on whether the squared gradient momentum hyperparameter $\\rho$ is greater than the gradient momentum hyperparameter $\\beta$ or not, and whether $\\varepsilon$ is small or large compared with the components of the gradient (the latter two cases correspond to early or late phases of the training, respectively).  Some of the results in the literature are derived as special cases of the theorems in this work.  The paper also reports some numerical experiments that seem to confirm the theoretical results as well as suggest that the one-norm of the gradient is inversely correlated with generalisation.", "strengths": "The introduction conveys clearly the place of this work in relation to the literature.\n\nThe summary of the main result in the full-batch case is helpful, and so are the discussion, the illustration using the bilinear model, and the suggestions of future directions.\n\nThe proofs are provided in the supplementary appendix, together with details of the numerical experiments.", "weaknesses": "The introduction is a little dry.\n\nThe statement of the main result that follows its summary in the full-batch case is difficult to parse, and its connection with the summary that precedes it is not obvious.  A minor point is that equation 10 should not end with a full stop.", "questions": "Can you say more about the graphs in Figure 1?  Why are we plotting the integral, and what can we conclude from the shapes of the various curves?\n\nWhat can you say about the situation when the numerical stability hyperparameter $\\varepsilon$ (or rather its square root) is neither small nor large in relation to the components of the gradient?  Can that be the case for a long period of the training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors apply the backward error analysis method to find ODEs that determine continuous trajectories which are close to the discrete trajectories of the popular Adam and RMSProp adaptive gradient-based algorithms.  They succeed in doing this up to a discrepancy which is second-order in the step size, for variants of the two algorithms depending on whether the numerical stability hyperparameter $\\varepsilon$ is inside or outside the square root in the step equation, and for both mini-batch and full-batch cases.  The main result for Adam uncovers three different regimes that penalise the positive one-norm of the gradient, or the negative one-norm of the gradient, or the squared two-norm of the gradient, depending on whether the squared gradient momentum hyperparameter $\\rho$ is greater than the gradient momentum hyperparameter $\\beta$ or not, and whether $\\varepsilon$ is small or large compared with the components of the gradient (the latter two cases correspond to early or late phases of the training, respectively).  Some of the results in the literature are derived as special cases of the theorems in this work.  The paper also reports some numerical experiments that seem to confirm the theoretical results as well as suggest that the one-norm of the gradient is inversely correlated with generalisation.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The introduction conveys clearly the place of this work in relation to the literature.\n\nThe summary of the main result in the full-batch case is helpful, and so are the discussion, the illustration using the bilinear model, and the suggestions of future directions.\n\nThe proofs are provided in the supplementary appendix, together with details of the numerical experiments.", "weaknesses": "The introduction is a little dry.\n\nThe statement of the main result that follows its summary in the full-batch case is difficult to parse, and its connection with the summary that precedes it is not obvious.  A minor point is that equation 10 should not end with a full stop.", "questions": "Can you say more about the graphs in Figure 1?  Why are we plotting the integral, and what can we conclude from the shapes of the various curves?\n\nWhat can you say about the situation when the numerical stability hyperparameter $\\varepsilon$ (or rather its square root) is neither small nor large in relation to the components of the gradient?  Can that be the case for a long period of the training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697843360207}], "openreview_url": "https://openreview.net/forum?id=ZA9XUTseA9", "arxiv_id": "2309.00079", "paper_pdf": "papers/ZA9XUTseA9.pdf", "paper_pdf_sha256": "cd10ecbd46a1d312d3e340bca44707eb5e40480e65b84af53a2c48e1a5954bf2", "paper_pdf_bytes": 629521, "paper_pdf_source": "openreview", "code_url": "https://github.com/borshigida/implicit-bias-of-adam", "code_repository": "borshigida/implicit-bias-of-adam", "code_commit": "e25ea9a22085331bc7c2ef7b98752ffc8cf1ac5c", "code_archive": "repos/ZA9XUTseA9.zip", "code_archive_sha256": "18b6ad2865c7045d94589be466a317f049c615549dcc67a4829b20f3f51e1b73", "code_archive_bytes": 12609, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 18, "github_languages": {"Python": 32552}, "github_archived": false, "github_pushed_at": "2024-07-29T13:01:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-implicit-bias-of-adam"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hTCBqt7pgxf", "year": 2023, "status": "rejected", "title": "Efficient block contrastive learning via parameter-free meta-node approximation", "authors": ["Gayan K Kulatilleke", "Marius Portmann", "Shekhar S. Chandra"], "authorids": ["~Gayan_K_Kulatilleke1", "~Marius_Portmann1", "~Shekhar_S._Chandra1"], "authors_source": "OpenReview API", "abstract": "Contrastive learning has recently achieved remarkable success in many domains including graphs. However contrastive loss, especially for graphs, requires a large number of negative samples which is unscalable and computationally prohibitive with a quadratic time complexity. Sub-sampling is not optimal and incorrect negative sampling leads to sampling bias. In this work, we propose a meta-node based approximation technique that can (a) proxy all negative combinations (b) in quadratic cluster size time complexity, (c) at graph level, not node level, and (d) exploit graph sparsity. By replacing node-pairs with additive cluster-pairs, we compute the negatives in cluster-time at graph level. The resulting Proxy approximated meta-node Contrastive (PamC) loss, based on simple optimized GPU operations, captures the full set of negatives, yet is efficient with a linear time complexity. By avoiding sampling, we effectively eliminate sample bias. We meet the criterion for larger number of samples, thus achieving block-contrastiveness, which is proven to outperform pair-wise losses. We use learnt soft cluster assignments for the meta-node constriction, and avoid possible heterophily and noise added during edge creation. Theoretically, we show that real world graphs easily satisfy conditions necessary for our approximation. Empirically, we show promising accuracy gains over state-of-the-art graph clustering on 6 benchmarks. Importantly, we gain substantially in efficiency; up to 3x in training time, 1.8x in inference time and over 5x in GPU memory reduction.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "8VksCy66JZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2485/Reviewer_GN9c"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes Proxy approximated meta-node Contrastive (PamC) for contrastive representation learning on graphs. PamC is motivated by the computational burden of vanilla contrastive loss (i.e., InfoNCE), and to deal with this problem, it proposes a meta-node based approximation technique which proxies all negative combinations in quadratic cluster size time complexity. Empirical reuslts sho that the proposed method demonstrates promising accuracy gains over sota graph clustering methods on 6 benchmarks with better efficiency.\n", "review_text": "Generally, I think the motivation of this paper is good. However, I think the propsoed method is over-complicted while does not show prominently better performance than simple methods. Besides, I believe the proposed method is not properly evaluated, in terms of tasks, datasets and baselines. I am leaning on rejection.\n", "strengths": "Strenghts:\n1. The motivation of this paper is good and solid. The computation complexity of InfoNCE loss is quadratic with respect to the number of nodes, which severly prevent full-graph training.\n2. The proposed method does show better efficacy than traditional contrastive methods.\n\nWeaknesses:\n  1. The authors claim that contrastive learning on graphs require a large number fo negative samples while subsampling is suboptimal. Is there any empirical support for this claim as according to my experience, subsampling will not severly degrade the model’s performance.\n \n2. Compared with data augmentation based contrastive learning  (InfoNCE loss) methods which ony requires two shared GNN encoder to generate node embeddings, the proposed method looks much more complicated (e.g., pretraining to get soft clusters). Although the overall complexity is linear, I doubt whether it can lead to efficacy when really applied to these datasets (especially when the graph is not that large, e.g., the datasets used in experiments)\n3. The paper focus on contrastive learning on graphs, however, a lot of important related works are missing in both related works and experiments. For example, [1] and [2] are two data augmentation-based contrastive methods using InfoNCE loss. [3] avoids negative samples using asymmetric structures. [4] avoids negative samples through feature-level decorrelation. The complexity of [3] and [4] are both linear to the graph size and thus they are scalable. However, this paper never consider these important baselines.\n4.  I am also confused about the tasks and datasets used in experiments. According to my knowledge, most self-supervised learning methods (including contrastive ones) foucs on node classifcation tasks (e.g., [1-4]). Why you consider graph clustering tasks instead of more commonly used node classication tasks.\n5. Although the most imporant claimed advantage is scalability, the datasets used for evaluation are really small. The authors should consider use larger graphs.\n\nReferences:\n[1] Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131, 2020b.\n[2] Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning with adaptive augmentation. In WWW, 2021.\n[3] Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Rémi Munos, Petar Velickovic, and Michal Valko. Bootstrapped representation learning on graphs. arXiv preprint arXiv:2102.06514, 2021.\n[4] Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. From canonical correlation analysis to self-supervised graph neural networks. In NeurIPS, 2021.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes Proxy approximated meta-node Contrastive (PamC) for contrastive representation learning on graphs. PamC is motivated by the computational burden of vanilla contrastive loss (i.e., InfoNCE), and to deal with this problem, it proposes a meta-node based approximation technique which proxies all negative combinations in quadratic cluster size time complexity. Empirical reuslts sho that the proposed method demonstrates promising accuracy gains over sota graph clustering methods on 6 benchmarks with better efficiency.\n", "strength_and_weaknesses": "Strenghts:\n1. The motivation of this paper is good and solid. The computation complexity of InfoNCE loss is quadratic with respect to the number of nodes, which severly prevent full-graph training.\n2. The proposed method does show better efficacy than traditional contrastive methods.\n\nWeaknesses:\n  1. The authors claim that contrastive learning on graphs require a large number fo negative samples while subsampling is suboptimal. Is there any empirical support for this claim as according to my experience, subsampling will not severly degrade the model’s performance.\n \n2. Compared with data augmentation based contrastive learning  (InfoNCE loss) methods which ony requires two shared GNN encoder to generate node embeddings, the proposed method looks much more complicated (e.g., pretraining to get soft clusters). Although the overall complexity is linear, I doubt whether it can lead to efficacy when really applied to these datasets (especially when the graph is not that large, e.g., the datasets used in experiments)\n3. The paper focus on contrastive learning on graphs, however, a lot of important related works are missing in both related works and experiments. For example, [1] and [2] are two data augmentation-based contrastive methods using InfoNCE loss. [3] avoids negative samples using asymmetric structures. [4] avoids negative samples through feature-level decorrelation. The complexity of [3] and [4] are both linear to the graph size and thus they are scalable. However, this paper never consider these important baselines.\n4.  I am also confused about the tasks and datasets used in experiments. According to my knowledge, most self-supervised learning methods (including contrastive ones) foucs on node classifcation tasks (e.g., [1-4]). Why you consider graph clustering tasks instead of more commonly used node classication tasks.\n5. Although the most imporant claimed advantage is scalability, the datasets used for evaluation are really small. The authors should consider use larger graphs.\n\nReferences:\n[1] Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131, 2020b.\n[2] Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning with adaptive augmentation. In WWW, 2021.\n[3] Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Rémi Munos, Petar Velickovic, and Michal Valko. Bootstrapped representation learning on graphs. arXiv preprint arXiv:2102.06514, 2021.\n[4] Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. From canonical correlation analysis to self-supervised graph neural networks. In NeurIPS, 2021.\n\n", "clarity,_quality,_novelty_and_reproducibility": "Overall, the propsoed method is novel and is clearly presented. ", "summary_of_the_review": "Generally, I think the motivation of this paper is good. However, I think the propsoed method is over-complicted while does not show prominently better performance than simple methods. Besides, I believe the proposed method is not properly evaluated, in terms of tasks, datasets and baselines. I am leaning on rejection.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666869635209}, {"id": "q6US_t72iMC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2485/Reviewer_X8V7"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper considers the task of representation learning for graph data. The proposed objective is a contrastive learning objective that is based on using clusters in negative sampling. Empirical analysis compares the proposed approach to a wide variety of representation learning and clustering approaches for graphs. ", "review_text": "This paper presents an empirically effective approach for graph representation learning and clustering. The approach, which uses clustering  is more efficient than previous work. The paper could be improved with more complete description of the landscape of related work on representation learning and clustering. The paper could be further improved with more clarity on the technical depth and revealed performance tradeoffs.", "strengths": "**Strengths** \n* The proposed approach seems to be widely empirically effective and is accurate. The approach is very computationally efficient compared to baseline approaches. \n* The proposed approach uses an interesting clustering-based objective that seems well motivated. \n\n**Weaknesses**\n* Clustering objectives have been studied in other contrastive learning approaches e.g. [1, 2, 3] among a wide array of other approaches in deep clustering. It would greatly improve the paper to consider how these approaches relate to the proposed approach. \n* Could the authors clarify the statement\n>We provide theoretical proof and show that real world graphs always satisfies the necessary\nconditions, and that PamCGC is block-contrastive, known to outperform pair-wise losses\n* While the proposed approach offers empirical advantages and is an interesting idea, I am wondering if the methodological depth is sufficient for the ICLR \"bar\". It would seem that the novelty is the combination of clustering ideas in graph representation learning. I think this is nice, but I feel the authors could clarify how their contribution investigates deeper questions in this area than past work. For instance, efficiency is a major advantage of the proposed approach. Understanding tradeoffs here is in Figure 2. It seems like more emphasis could be placed in this understanding.\n\n[1] Caron, Mathilde, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. \"Unsupervised learning of visual features by contrasting cluster assignments.\" Advances in Neural Information Processing Systems 33 (2020): 9912-9924.\n\n[2] Zhang, Chunyang, Hongyu Yao, C. L. Chen, and Yuena Lin. \"Graph Representation Learning via Contrasting Cluster Assignments.\" arXiv preprint arXiv:2112.07934 (2021).\n\n[3] Li, Junnan, Pan Zhou, Caiming Xiong, and Steven CH Hoi. \"Prototypical contrastive learning of unsupervised representations.\" arXiv preprint arXiv:2005.04966 (2020).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper considers the task of representation learning for graph data. The proposed objective is a contrastive learning objective that is based on using clusters in negative sampling. Empirical analysis compares the proposed approach to a wide variety of representation learning and clustering approaches for graphs. ", "strength_and_weaknesses": "**Strengths** \n* The proposed approach seems to be widely empirically effective and is accurate. The approach is very computationally efficient compared to baseline approaches. \n* The proposed approach uses an interesting clustering-based objective that seems well motivated. \n\n**Weaknesses**\n* Clustering objectives have been studied in other contrastive learning approaches e.g. [1, 2, 3] among a wide array of other approaches in deep clustering. It would greatly improve the paper to consider how these approaches relate to the proposed approach. \n* Could the authors clarify the statement\n>We provide theoretical proof and show that real world graphs always satisfies the necessary\nconditions, and that PamCGC is block-contrastive, known to outperform pair-wise losses\n* While the proposed approach offers empirical advantages and is an interesting idea, I am wondering if the methodological depth is sufficient for the ICLR \"bar\". It would seem that the novelty is the combination of clustering ideas in graph representation learning. I think this is nice, but I feel the authors could clarify how their contribution investigates deeper questions in this area than past work. For instance, efficiency is a major advantage of the proposed approach. Understanding tradeoffs here is in Figure 2. It seems like more emphasis could be placed in this understanding.\n\n[1] Caron, Mathilde, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. \"Unsupervised learning of visual features by contrasting cluster assignments.\" Advances in Neural Information Processing Systems 33 (2020): 9912-9924.\n\n[2] Zhang, Chunyang, Hongyu Yao, C. L. Chen, and Yuena Lin. \"Graph Representation Learning via Contrasting Cluster Assignments.\" arXiv preprint arXiv:2112.07934 (2021).\n\n[3] Li, Junnan, Pan Zhou, Caiming Xiong, and Steven CH Hoi. \"Prototypical contrastive learning of unsupervised representations.\" arXiv preprint arXiv:2005.04966 (2020).", "clarity,_quality,_novelty_and_reproducibility": "Please see weaknesses for additional clarifications. ", "summary_of_the_review": "This paper presents an empirically effective approach for graph representation learning and clustering. The approach, which uses clustering  is more efficient than previous work. The paper could be improved with more complete description of the landscape of related work on representation learning and clustering. The paper could be further improved with more clarity on the technical depth and revealed performance tradeoffs.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666661787630}, {"id": "jSMkjl53YD", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2485/Reviewer_cHFr"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper extends the contrastive learning on graph by replacing the contrastive between negative samples with that between class clusters with theoretical analysis. It possesses the attractive characteristic of reducing complexity. Experimental evaluations demonstrate its effectiveness and efficiency.", "review_text": "This paper possesses clear motivation, good writing and organization, and sufficient evaluations. My main concern is the novelty, since existing prototypical contrastive learning also reduce complexity by considering cluster centers.\n", "strengths": "Strength\n- It makes sense to replacing the contrastive between negative samples with that between class clusters to reduce the complexity.\n- The experimental evaluations are sufficient.\n- The writing and organization are good.\n\nWeakness\n- My main concern is the novelty may not be significant. Although authors call the cluster center as meta-nodes, they are actually the prototype of the cluster. There are many efforts have be paid on prototypical contrastive learning, such as [1] and its variants, which also takes cluster centers into considerations and reduce the complexity.\n\n[1] Junnan Li, Pan Zhou, Caiming Xiong, Steven C. H. Hoi: Prototypical Contrastive Learning of Unsupervised Representations. ICLR 2021\n2020\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper extends the contrastive learning on graph by replacing the contrastive between negative samples with that between class clusters with theoretical analysis. It possesses the attractive characteristic of reducing complexity. Experimental evaluations demonstrate its effectiveness and efficiency.", "strength_and_weaknesses": "Strength\n- It makes sense to replacing the contrastive between negative samples with that between class clusters to reduce the complexity.\n- The experimental evaluations are sufficient.\n- The writing and organization are good.\n\nWeakness\n- My main concern is the novelty may not be significant. Although authors call the cluster center as meta-nodes, they are actually the prototype of the cluster. There are many efforts have be paid on prototypical contrastive learning, such as [1] and its variants, which also takes cluster centers into considerations and reduce the complexity.\n\n[1] Junnan Li, Pan Zhou, Caiming Xiong, Steven C. H. Hoi: Prototypical Contrastive Learning of Unsupervised Representations. ICLR 2021\n2020\n", "clarity,_quality,_novelty_and_reproducibility": "- The clarity and reproducibility are good.\n- The novelty and originality may be not significant.", "summary_of_the_review": "This paper possesses clear motivation, good writing and organization, and sufficient evaluations. My main concern is the novelty, since existing prototypical contrastive learning also reduce complexity by considering cluster centers.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666575274975}, {"id": "QkU74_S-i-w", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2485/Reviewer_6uyq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to utilize the meta-node (average of node embeddings within a cluster) for constructing negative samples in contrastive learning. It claims that this approach is more efficient than previous sampling based methods and it leads naturally to block contrastive learning which could be beneficial. Experiments on various datasets demonstrate the performance of the proposed method on different modalities including image, text and graph.", "review_text": "Overall the proposed method is well motivated and achieves reasonable speedup for training. I have some questions about the novelty and experiments. I may consider adjusting my rating after discussing with the authors.", "strengths": "## Strength\n1. The direction of more efficient negative sample mining is important for contrastive learning.\n2. The idea of utilizing the clustering centers as negative samples saves the sampling cost.\n3. Extensive experiments demonstrate comparable performance with baseline methods while achieving good speedup ratio for training time.\n\n## Weaknesses\n1. Discussion on related works is not enough, which makes the novelty a little unclear. Is this work the first to utilize averaged embeddings as negative samples for contrastive learning? What are the closest related works? \n2. It is not clear how the proposed method can save inference time\n3. How long does it take to evaluate the soft label? Is it included in the training time?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes to utilize the meta-node (average of node embeddings within a cluster) for constructing negative samples in contrastive learning. It claims that this approach is more efficient than previous sampling based methods and it leads naturally to block contrastive learning which could be beneficial. Experiments on various datasets demonstrate the performance of the proposed method on different modalities including image, text and graph.", "strength_and_weaknesses": "## Strength\n1. The direction of more efficient negative sample mining is important for contrastive learning.\n2. The idea of utilizing the clustering centers as negative samples saves the sampling cost.\n3. Extensive experiments demonstrate comparable performance with baseline methods while achieving good speedup ratio for training time.\n\n## Weaknesses\n1. Discussion on related works is not enough, which makes the novelty a little unclear. Is this work the first to utilize averaged embeddings as negative samples for contrastive learning? What are the closest related works? \n2. It is not clear how the proposed method can save inference time\n3. How long does it take to evaluate the soft label? Is it included in the training time?", "clarity,_quality,_novelty_and_reproducibility": "Clarity: fair, see questions above\nQuality: reasonable\nNovelty: needs clarification\n", "summary_of_the_review": "Overall the proposed method is well motivated and achieves reasonable speedup for training. I have some questions about the novelty and experiments. I may consider adjusting my rating after discussing with the authors.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666566083392}], "openreview_url": "https://openreview.net/forum?id=hTCBqt7pgxf", "arxiv_id": "2209.14067", "paper_pdf": "papers/hTCBqt7pgxf.pdf", "paper_pdf_sha256": "1faa3aa876f62b796fbe66fb364b6c80dddda2ebac6d7be7e6616e7db8fe4563", "paper_pdf_bytes": 4049478, "paper_pdf_source": "openreview", "code_url": "https://github.com/gayanku/PAMC", "code_repository": "gayanku/PAMC", "code_commit": "f33a8b347b23945194f1425227243fe6b8745ffc", "code_archive": "repos/hTCBqt7pgxf.zip", "code_archive_sha256": "277b9d70042e5802907bd4703d261c6b737c1422c49272547198123618e364c1", "code_archive_bytes": 13311, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 28, "github_languages": {"Python": 49547}, "github_archived": false, "github_pushed_at": "2022-09-29T00:46:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-block-contrastive-learning-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gccdzDu5Ur", "year": 2022, "status": "rejected", "title": "Combining Diverse Feature Priors", "authors": ["Saachi Jain", "Dimitris Tsipras", "Aleksander Madry"], "authorids": ["~Saachi_Jain1", "~Dimitris_Tsipras1", "~Aleksander_Madry1"], "authors_source": "OpenReview API", "abstract": "To improve model generalization, model designers often restrict the features that their models use, either implicitly or explicitly. In this work, we explore the design space of leveraging such feature priors by viewing them as distinct perspectives on the data. Specifically, we find that models trained with diverse sets of explicit feature priors have less overlapping failure modes, and can thus be combined more effectively. Moreover, we demonstrate that jointly training such models on additional (unlabeled) data allows them to correct each other's mistakes, which, in turn, leads to better generalization and resilience to spurious correlations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "m7MjLVvGt7R", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2102/Reviewer_VXFn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents multiple techniques for training models with different feature priors (i.e. inclinations to focus on different aspects of the training data) and combining them, either post hoc via ensembles or by allowing the models to provide augmented pseudo-labelled training data to each other via co-training. When using simple ensembling techniques, ensembles with a diversity of feature priors are show to perform better than ensembles where the individual models have similar feature priors. Co-training is shown to boost performance substantially when models with diverse feature priors supply pseudo-labels to each other. The problem domain is image classification. The feature priors concern shape and texture. Different preprocessing and or architecture constraint techniques are used for different models so as to predispose them to focus on shape but not texture or vice versa.", "review_text": "The implementation of shape feature priors (via edge detection preprocessing) and texture feature priors (via limited receptive field via bagnet) makes a lot of sense and does a good job of illustrating a concrete example of a collection of different feature priors.  Some of the co-training experimental results are strong. \n\nOne concern I have is that the ensemble results presented in section 3.2 are generated using very primitive ensembling techniques. Appendix A.4 says the combination techniques were simply max, average and lowest rank. It is more common to treat this kind of ensembling as a 2nd-level machine learning problem with the outputs of the models forming the inputs to a 2nd-level model. I would not have expected a fancy 2nd-level model but I was hoping at a minimum that the first-level models would be combined via. e.g. linear regression on first-level outputs from a held-out validation set. Fancier combinations (e.g. neural nets) are also possible, of course. I would encourage the authors to read a few writeups by winners of Kaggle competitions and/or read about the ensembling done in the Netflix Prize to get a better sense of what constitutes state-of-the-art ensemble combination techniques.\n\nI was also concerned about the assumption made on page 7 that spurious correlations are likely not to exist in unlabelled data, because unlabelled data supposedly comes from a more diverse collection process. While this may be true in some cases, there will also be many real-world situations where all the input data, whether labelled or unlabelled, comes from the same distribution and it may all have the same spurious correlation. It is often the case that a small portion of the data is labelled simply because it is very time intensive for humans to do the labeling but nonetheless, the remaining unlabelled data comes from the same distribution. I hope that in future versions of this work, the authors make clear that practitioners should think hard about whether their unlabelled data will have the same spurious correlations as their labelled data, rather than assuming that this is likely the case.\n\nFor these reasons, I can only give the paper a 5. I would prefer that the ensemble section be redone with more sophisticated ensembling and/or removed and I would prefer that the absence-of-spurious-correlation-in-unlabelled data assumption be presented more cautiously. \n \nA minor complaint: from a presentation point of view, it is non-standard and a bit strange to add additional related work in section 6. I don't normally expect to read about related work after the results and towards the end of the paper. Related work is normally presented earlier in a paper. The authors might consider moving this section to an earlier point in the paper. \n\nOn page 8, I found the bolding of +BagNet cotraining results to be a bit confusing. Normally the 'winning' algorithm results are bolded, which in this case is Canny. I realize that the message of the huge boost of cotraining for +BagNet is what is intended but it still confused me that the bolded numbers were not the best numbers.\n\nIt also would be nice to show the method on another domain aside from image classification, although I realize space constraints might make this difficult. The authors might consider removing the ensembling section in future versions of the work and instead using that space for cotraining results on another type of problem.\n\n*** Update after author rebuttal *** In light of the addition of stacking experiments for the ensembling, I have raised my score to a 6.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents multiple techniques for training models with different feature priors (i.e. inclinations to focus on different aspects of the training data) and combining them, either post hoc via ensembles or by allowing the models to provide augmented pseudo-labelled training data to each other via co-training. When using simple ensembling techniques, ensembles with a diversity of feature priors are show to perform better than ensembles where the individual models have similar feature priors. Co-training is shown to boost performance substantially when models with diverse feature priors supply pseudo-labels to each other. The problem domain is image classification. The feature priors concern shape and texture. Different preprocessing and or architecture constraint techniques are used for different models so as to predispose them to focus on shape but not texture or vice versa.", "main_review": "The implementation of shape feature priors (via edge detection preprocessing) and texture feature priors (via limited receptive field via bagnet) makes a lot of sense and does a good job of illustrating a concrete example of a collection of different feature priors.  Some of the co-training experimental results are strong. \n\nOne concern I have is that the ensemble results presented in section 3.2 are generated using very primitive ensembling techniques. Appendix A.4 says the combination techniques were simply max, average and lowest rank. It is more common to treat this kind of ensembling as a 2nd-level machine learning problem with the outputs of the models forming the inputs to a 2nd-level model. I would not have expected a fancy 2nd-level model but I was hoping at a minimum that the first-level models would be combined via. e.g. linear regression on first-level outputs from a held-out validation set. Fancier combinations (e.g. neural nets) are also possible, of course. I would encourage the authors to read a few writeups by winners of Kaggle competitions and/or read about the ensembling done in the Netflix Prize to get a better sense of what constitutes state-of-the-art ensemble combination techniques.\n\nI was also concerned about the assumption made on page 7 that spurious correlations are likely not to exist in unlabelled data, because unlabelled data supposedly comes from a more diverse collection process. While this may be true in some cases, there will also be many real-world situations where all the input data, whether labelled or unlabelled, comes from the same distribution and it may all have the same spurious correlation. It is often the case that a small portion of the data is labelled simply because it is very time intensive for humans to do the labeling but nonetheless, the remaining unlabelled data comes from the same distribution. I hope that in future versions of this work, the authors make clear that practitioners should think hard about whether their unlabelled data will have the same spurious correlations as their labelled data, rather than assuming that this is likely the case.\n\nFor these reasons, I can only give the paper a 5. I would prefer that the ensemble section be redone with more sophisticated ensembling and/or removed and I would prefer that the absence-of-spurious-correlation-in-unlabelled data assumption be presented more cautiously. \n \nA minor complaint: from a presentation point of view, it is non-standard and a bit strange to add additional related work in section 6. I don't normally expect to read about related work after the results and towards the end of the paper. Related work is normally presented earlier in a paper. The authors might consider moving this section to an earlier point in the paper. \n\nOn page 8, I found the bolding of +BagNet cotraining results to be a bit confusing. Normally the 'winning' algorithm results are bolded, which in this case is Canny. I realize that the message of the huge boost of cotraining for +BagNet is what is intended but it still confused me that the bolded numbers were not the best numbers.\n\nIt also would be nice to show the method on another domain aside from image classification, although I realize space constraints might make this difficult. The authors might consider removing the ensembling section in future versions of the work and instead using that space for cotraining results on another type of problem.\n\n*** Update after author rebuttal *** In light of the addition of stacking experiments for the ensembling, I have raised my score to a 6.", "summary_of_the_review": "The authors achieve some positive results from cotraining of groups image classification models designed to focus on shape but not texture or vice versa. However, their ensembling results are acheived using very primitive ensembling which is not state of the art. They also overstate the odds that spurious correlations are unlikely to exist in unlabelled data.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635892432269}, {"id": "o6BLRaz_PYN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2102/Reviewer_KJot"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper is an empirical study of combining multiple feature priors along with some pre-processing to solve a variety of computer vision tasks.", "review_text": "Positives\n\n+ The study seems to be interesting and maybe useful for practitioners.\n\nConcerns\n\n- Very meagre contribution in terms of technical novelty and framework. \n\n- Looks like an empirical study without much conviction and direction.\n\n- Experimental evaluation and comparisons seem dated, not state of the art.\n\n- The work is very much below the expected standards of ICLR.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper is an empirical study of combining multiple feature priors along with some pre-processing to solve a variety of computer vision tasks.", "main_review": "Positives\n\n+ The study seems to be interesting and maybe useful for practitioners.\n\nConcerns\n\n- Very meagre contribution in terms of technical novelty and framework. \n\n- Looks like an empirical study without much conviction and direction.\n\n- Experimental evaluation and comparisons seem dated, not state of the art.\n\n- The work is very much below the expected standards of ICLR.", "summary_of_the_review": "The paper is clearly below par, can be rejected.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635848448383}, {"id": "7bC9SxbWvZY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2102/Reviewer_LSTE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a formalized framework for imposing priors on the feature extraction in deep visual processing models. There has been earlier work on encouraging certain feature representations (e.g. suppressing the focus on texture in feature extraction) and also making feature representations robust to domain shift. The core contribution of this paper is the systematic formulation and investigation of how different, distinct feature priors leads to complementary feature representations that can be combined to provide more robust data representations - in other words, creating synthesized multi-view data representations.\nThe paper ties back to early (1998) work on co-training (which essentially is multi-modal bootstrapping) and ties this to the more recent body of work on self-supervision and self-training.\nExperiments are performed with classical shape- and texture-biased models, and show that the hypothesis - that diverse feature priors are able to robustly create a set of complementary data views - holds.", "review_text": "This paper has a number of strengths, that combined makes me recommend the paper for acceptance:\n+ The topic of this paper, creating and combining robust, generalizable and diverse feature representations, is of high relevance to a large portion of the ICLR audience.\n+ It provides an interesting and valuable formal framework for steering feature representations in different directions, creating multi-view representations of the data.\n+ It is well written, well organized, technically correct, and easy to read.\n+ The experimental design is sound and well done.\n\nOne weakness can be pointed out, not however any cause for not accepting this paper in my opinion:\n- The experiments are performed on old datasets, CIFAR-10 and STL-10, both with quite clear class structure and simplistic image setting (e.g. the object centered in the image). It would be interesting to see experiments on more difficult data with fine-grained and hierarchical class structure for example. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a formalized framework for imposing priors on the feature extraction in deep visual processing models. There has been earlier work on encouraging certain feature representations (e.g. suppressing the focus on texture in feature extraction) and also making feature representations robust to domain shift. The core contribution of this paper is the systematic formulation and investigation of how different, distinct feature priors leads to complementary feature representations that can be combined to provide more robust data representations - in other words, creating synthesized multi-view data representations.\nThe paper ties back to early (1998) work on co-training (which essentially is multi-modal bootstrapping) and ties this to the more recent body of work on self-supervision and self-training.\nExperiments are performed with classical shape- and texture-biased models, and show that the hypothesis - that diverse feature priors are able to robustly create a set of complementary data views - holds.", "main_review": "This paper has a number of strengths, that combined makes me recommend the paper for acceptance:\n+ The topic of this paper, creating and combining robust, generalizable and diverse feature representations, is of high relevance to a large portion of the ICLR audience.\n+ It provides an interesting and valuable formal framework for steering feature representations in different directions, creating multi-view representations of the data.\n+ It is well written, well organized, technically correct, and easy to read.\n+ The experimental design is sound and well done.\n\nOne weakness can be pointed out, not however any cause for not accepting this paper in my opinion:\n- The experiments are performed on old datasets, CIFAR-10 and STL-10, both with quite clear class structure and simplistic image setting (e.g. the object centered in the image). It would be interesting to see experiments on more difficult data with fine-grained and hierarchical class structure for example. \n", "summary_of_the_review": "This paper has a number of strengths, that combined makes me recommend the paper for acceptance: Of high relevance, well written and correct, proposes a valuable framework, and contains sound and well designed experiments.\nOne weakness can be pointed out, not however any cause for not accepting this paper in my opinion: The experiments are performed on old datasets.\n\nIn summary, I propose acceptance for this paper and believe it will be of interest to a large portion of the ICLR audience.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No concerns", "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635782674960}, {"id": "aWEDeWw1PnM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2102/Reviewer_uSUu"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The goal of the paper is to improve model generalisation. The authors consider feature priors as distinct perspectives on the data. The results show that models trained with diverse sets of various feature priors have less overlapping modes and are more efficiently combined.  ", "review_text": "Strengths. The experimental part is relatively clear. \n\nWeaknesses. The paper is not very clearly written. First, I would appreciate some (even informal) definition of a feature prior. Later in the text, the co-training is mentioned, and it seems that using different priors = co-training using different views. Is it the idea of the paper? I did not really understand the first contribution: \"We demonstrate that training models with diverse feature priors results in them making mistakes on different parts of the data distribution, even if their overall accuracy is similar.\" What is meant?\n\nAs far as I understand, there are two \"priors\" only explored in the paper: shape and texture. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The goal of the paper is to improve model generalisation. The authors consider feature priors as distinct perspectives on the data. The results show that models trained with diverse sets of various feature priors have less overlapping modes and are more efficiently combined.  ", "main_review": "Strengths. The experimental part is relatively clear. \n\nWeaknesses. The paper is not very clearly written. First, I would appreciate some (even informal) definition of a feature prior. Later in the text, the co-training is mentioned, and it seems that using different priors = co-training using different views. Is it the idea of the paper? I did not really understand the first contribution: \"We demonstrate that training models with diverse feature priors results in them making mistakes on different parts of the data distribution, even if their overall accuracy is similar.\" What is meant?\n\nAs far as I understand, there are two \"priors\" only explored in the paper: shape and texture. \n", "summary_of_the_review": "The current contribution is an exploratory work, combining several state-of-the-art methods (for instance, self-training and co-training are used in the experiments). \nThere is a lack of technical novelty. \n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635676015749}], "openreview_url": "https://openreview.net/forum?id=gccdzDu5Ur", "arxiv_id": "2110.08220", "paper_pdf": "papers/gccdzDu5Ur.pdf", "paper_pdf_sha256": "4907d8e26920118f3ea6199aa60ee36ad71a3609225cfe748fe0ef4469163a54", "paper_pdf_bytes": 3702503, "paper_pdf_source": "openreview", "code_url": "https://github.com/MadryLab/copriors", "code_repository": "MadryLab/copriors", "code_commit": "da5146bc3d02e7cfe3e48cde70f0d3b8eac536de", "code_archive": "repos/gccdzDu5Ur.zip", "code_archive_sha256": "32dc2a2c5a37aeb7c6df414973f2c0f9ef88eb60a274ff958919f3ac549a857c", "code_archive_bytes": 24795, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 25, "github_languages": {"Python": 67395}, "github_archived": false, "github_pushed_at": "2021-10-18T01:35:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/combining-diverse-feature-priors-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "a5KvtsZ14ev", "year": 2021, "status": "rejected", "title": "SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks", "authors": ["Bahare Fatemi", "Seyed Mehran Kazemi", "Layla El Asri"], "authorids": ["~Bahare_Fatemi1", "~Seyed_Mehran_Kazemi1", "~Layla_El_Asri2"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) work well when the graph structure is provided. However, this structure may not always be available in real-world applications. One solution to this problem is to infer the latent structure and then apply a GNN to the inferred graph. Unfortunately, the space of possible graph structures grows super-exponentially with the number of nodes and so the available node labels may be insufficient for learning both the structure and the GNN parameters. In this work, we propose the Simultaneous Learning of Adjacency and GNN Parameters with Self-supervision, or SLAPS, a method that provides more supervision for inferring a graph structure. This approach consists of training a denoising autoencoder GNN in parallel with the task-specific GNN. The autoencoder is trained to reconstruct the initial node features given noisy node features as well as a structure provided by a learnable graph generator. We explore the design space of SLAPS by comparing different graph generation and symmetrization approaches. A comprehensive experimental study demonstrates that SLAPS scales to large graphs with hundreds of thousands of nodes and outperforms several models that have been proposed to learn a task-specific graph structure on established benchmarks.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "lPpVONlB-V", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper904/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of nodes classification with few labeled data and missing graph structures. The proposed solution is expected to infer unobserved graph structure as well as the parameters of the classification model. The main contribution of this paper is proposing adding a denoise autoencoder layer which provides more supervision to the learning. The model compares favorably with other states of art models in several benchmark graph data sets.\n\nConcerns:\n1. The experiments assume inputs only contain node features. The solution proposed seems to be incremental under this setting as the problem is a special case of few-shot learning where metric-learning based methods including GNN and denoise autoencoder all have been studied before. See [1], [2], [3]. Further discussion of related works is necessary.\n\n2. The proposed method seems very heuristic driven without providing further theoretical analysis. It is unclear how and why the self-supervision might help the classification learning task. Figure 2 gives a negative example of the existing method but it does not explain why the proposed method could make a difference.\n\n3. How does the proposed model compare with regular GNN/GCN models if edges are not entirely missing? The scope of the paper will be limited if it doesn’t work well when graph data is noisy but not entirely missing.\n\nReference\n[1] https://arxiv.org/abs/1905.01102\n[2] http://www.cs.toronto.edu/~gkoch/files/msc-thesis.pdf\n[3] https://arxiv.org/abs/1711.04043\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official review", "review": "This paper considers the problem of nodes classification with few labeled data and missing graph structures. The proposed solution is expected to infer unobserved graph structure as well as the parameters of the classification model. The main contribution of this paper is proposing adding a denoise autoencoder layer which provides more supervision to the learning. The model compares favorably with other states of art models in several benchmark graph data sets.\n\nConcerns:\n1. The experiments assume inputs only contain node features. The solution proposed seems to be incremental under this setting as the problem is a special case of few-shot learning where metric-learning based methods including GNN and denoise autoencoder all have been studied before. See [1], [2], [3]. Further discussion of related works is necessary.\n\n2. The proposed method seems very heuristic driven without providing further theoretical analysis. It is unclear how and why the self-supervision might help the classification learning task. Figure 2 gives a negative example of the existing method but it does not explain why the proposed method could make a difference.\n\n3. How does the proposed model compare with regular GNN/GCN models if edges are not entirely missing? The scope of the paper will be limited if it doesn’t work well when graph data is noisy but not entirely missing.\n\nReference\n[1] https://arxiv.org/abs/1905.01102\n[2] http://www.cs.toronto.edu/~gkoch/files/msc-thesis.pdf\n[3] https://arxiv.org/abs/1711.04043\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604642452730}, {"id": "KX2uDF2HrLn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper904/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a way to use self-supervision with denoising autoencoders to improve learning of the graph structure for GNNs. The approach is compared with a number of recent approaches from the literature.\n\nThe approach addresses the highly relevant problem of learning graph structure with GNNs. The paper is well written, with clear motivation and an informative summary of similar prior works. The differences with respect to these approaches, primarily the self-supervised aspects, are clearly pointed out.\n\nThere are a few prior works covering self-supervision with GNNs that are all quite new [1-3]. It would strengthen the manuscript if these works were discussed in relation to this work. However, consider how new they all are it is not surprising that they were not mentioned.\n\n\nQuality\n- While the originality is perhaps not big, the work is thorough and well written. There is a substantial experimental comparison with relevant prior work and the conclusions seems to be substantiated.\n\nClarity\n- The paper is well written and easy to understand. I would have liked a little more details and explanation on the MLP-kNN aspects.\n\nOriginality\n- Self-supervision seems to be the topic of a lot of papers lately, so in that respect, this work is perhaps not the most original.\n\nSignificance\n- I am not aware of similar works and the results are convincing, so I guess the approach could have reasonable influence on the field.\n\n[1] Jin, Wei, et al. \"Self-supervised learning on graphs: Deep insights and new direction.\" arXiv preprint arXiv:2006.10141 (2020).\n[2] You, Yuning, et al. \"When Does Self-Supervision Help Graph Convolutional Networks?.\" arXiv preprint arXiv:2006.09136 (2020).\n[3] Zhu, Qikui, Bo Du, and Pingkun Yan. \"Self-supervised Training of Graph Convolutional Networks.\" arXiv preprint arXiv:2006.02380 (2020)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work with convincing results.", "review": "The paper proposes a way to use self-supervision with denoising autoencoders to improve learning of the graph structure for GNNs. The approach is compared with a number of recent approaches from the literature.\n\nThe approach addresses the highly relevant problem of learning graph structure with GNNs. The paper is well written, with clear motivation and an informative summary of similar prior works. The differences with respect to these approaches, primarily the self-supervised aspects, are clearly pointed out.\n\nThere are a few prior works covering self-supervision with GNNs that are all quite new [1-3]. It would strengthen the manuscript if these works were discussed in relation to this work. However, consider how new they all are it is not surprising that they were not mentioned.\n\n\nQuality\n- While the originality is perhaps not big, the work is thorough and well written. There is a substantial experimental comparison with relevant prior work and the conclusions seems to be substantiated.\n\nClarity\n- The paper is well written and easy to understand. I would have liked a little more details and explanation on the MLP-kNN aspects.\n\nOriginality\n- Self-supervision seems to be the topic of a lot of papers lately, so in that respect, this work is perhaps not the most original.\n\nSignificance\n- I am not aware of similar works and the results are convincing, so I guess the approach could have reasonable influence on the field.\n\n[1] Jin, Wei, et al. \"Self-supervised learning on graphs: Deep insights and new direction.\" arXiv preprint arXiv:2006.10141 (2020).\n[2] You, Yuning, et al. \"When Does Self-Supervision Help Graph Convolutional Networks?.\" arXiv preprint arXiv:2006.09136 (2020).\n[3] Zhu, Qikui, Bo Du, and Pingkun Yan. \"Self-supervised Training of Graph Convolutional Networks.\" arXiv preprint arXiv:2006.02380 (2020)\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603881398900}, {"id": "_0ltaipY67", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper904/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a method for simultaneously learning the graph structure (or a graph generative model) and the parameters of a GNN for node classification. This is a topic of recent interest and highly relevant to the ICLR community. \n\nThe authors propose two different ways for the graph generator. First, a fully parameterized method that outputs a continuous (weighted) adjacency matrix. The second computes a spare k-NN graph based on the input similarities. Using the word “graph generator” or “generative model” is a bit of a misnomer here because both methods are not probabilistic but deterministic. The adjacency matrices output by these “generators” are then made into symmetric and positive matrices (i.e., adjacency matrices for positively weighted undirected graphs). Finally, in addition to the typical supervised loss, the authors also propose an unsupervised loss based on a reconstruction loss. \n\nFirst, I like the idea of adding an unsupervised loss and also appreciate the experiments and the results as robust. \n\nI have a concern about the approach though and perhaps this could be clarified in a conversation here. When you don’t use the MLP-kNN “generator” then the adjacency matrix generated is not sparse at all. But what that means is that you are using (for each node) a fully connected layer. Do you think that the advantage of your method then is in the “adjacency processor step”? When you do use the MLP-kNN, how do you compute the gradients? Is it that you only compute gradients for the edges you selected? In this case you would have lots of edges without “supervision” as you call it. Or are you also obtaining gradients for the edges not selected? In this case, creating the kNN graph is a discrete operation and it is not clear to me how to differentiate through such an operation. Of course, recent proposals have been made for differentiating through kNN but I don’t see any reference or mention of what you are doing here. Could you please elaborate on that? \n\nRegarding the statements about LDS. First, you write that if two nodes v_i and v_j are not directly connected to any labeled nodes, then the RV between them (the edge) does not receive supervision. The statement (and the text following it) is somewhat misleading for two reasons. \n\nFirst, both of the nodes have to be not directly connected. But then in the next sentence you wrote that 80 and 89% of nodes are not directly connected to a labeled node for the standard benchmark graphs. The more appropriate analysis, however, would be to count the number of pairs of nodes where both nodes are not connected to a labelled node. This should also include the validation nodes used for the outer objective in LDS. Also note that LDS doesn’t always use the standard benchmark graphs (it either constructs a k-NN graph where k is a hyperparameter of the method or initializes with a subgraph of the given graph). \n\nSecond, your statement is true for *one* sampled graph. Remember that LDS samples a set of graphs in each iteration. It can happen that, even if there is a pair of nodes where both nodes are not directly connected to a labelled node in one sampled graph, one of these nodes might be connected in a different sample. Indeed, if a sampled edge connecting either v_i or v_j directly to a labelled node leads to a reduction of the loss, the next time said edge is probably more likely to be sampled. This is not to say that self-supervision is not a good idea (I do like the idea) but I’m not sure if your statements about LDS are quite accurate. \n\nOverall, I think that this paper strength is the proposed self-supervised loss and the experimental evaluation. It is rather weak on the methodology, its presentation, and related  discussion. There are several questions I need to hear your response to. Once these are clarified I'm open to adjust my score accordingly. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Self-supervised loss for simultaneously learning structure and parameters for GNN", "review": "The authors propose a method for simultaneously learning the graph structure (or a graph generative model) and the parameters of a GNN for node classification. This is a topic of recent interest and highly relevant to the ICLR community. \n\nThe authors propose two different ways for the graph generator. First, a fully parameterized method that outputs a continuous (weighted) adjacency matrix. The second computes a spare k-NN graph based on the input similarities. Using the word “graph generator” or “generative model” is a bit of a misnomer here because both methods are not probabilistic but deterministic. The adjacency matrices output by these “generators” are then made into symmetric and positive matrices (i.e., adjacency matrices for positively weighted undirected graphs). Finally, in addition to the typical supervised loss, the authors also propose an unsupervised loss based on a reconstruction loss. \n\nFirst, I like the idea of adding an unsupervised loss and also appreciate the experiments and the results as robust. \n\nI have a concern about the approach though and perhaps this could be clarified in a conversation here. When you don’t use the MLP-kNN “generator” then the adjacency matrix generated is not sparse at all. But what that means is that you are using (for each node) a fully connected layer. Do you think that the advantage of your method then is in the “adjacency processor step”? When you do use the MLP-kNN, how do you compute the gradients? Is it that you only compute gradients for the edges you selected? In this case you would have lots of edges without “supervision” as you call it. Or are you also obtaining gradients for the edges not selected? In this case, creating the kNN graph is a discrete operation and it is not clear to me how to differentiate through such an operation. Of course, recent proposals have been made for differentiating through kNN but I don’t see any reference or mention of what you are doing here. Could you please elaborate on that? \n\nRegarding the statements about LDS. First, you write that if two nodes v_i and v_j are not directly connected to any labeled nodes, then the RV between them (the edge) does not receive supervision. The statement (and the text following it) is somewhat misleading for two reasons. \n\nFirst, both of the nodes have to be not directly connected. But then in the next sentence you wrote that 80 and 89% of nodes are not directly connected to a labeled node for the standard benchmark graphs. The more appropriate analysis, however, would be to count the number of pairs of nodes where both nodes are not connected to a labelled node. This should also include the validation nodes used for the outer objective in LDS. Also note that LDS doesn’t always use the standard benchmark graphs (it either constructs a k-NN graph where k is a hyperparameter of the method or initializes with a subgraph of the given graph). \n\nSecond, your statement is true for *one* sampled graph. Remember that LDS samples a set of graphs in each iteration. It can happen that, even if there is a pair of nodes where both nodes are not directly connected to a labelled node in one sampled graph, one of these nodes might be connected in a different sample. Indeed, if a sampled edge connecting either v_i or v_j directly to a labelled node leads to a reduction of the loss, the next time said edge is probably more likely to be sampled. This is not to say that self-supervision is not a good idea (I do like the idea) but I’m not sure if your statements about LDS are quite accurate. \n\nOverall, I think that this paper strength is the proposed self-supervised loss and the experimental evaluation. It is rather weak on the methodology, its presentation, and related  discussion. There are several questions I need to hear your response to. Once these are clarified I'm open to adjust my score accordingly. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603881178538}, {"id": "caadRfEb027", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper904/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:  \nThis paper proposes to tackle jointly learning graph structures and GNN parameters without accessing the original graph structure. Specifically, the proposed method adopts a self-supervised auxiliary task, i.e., parallel training using the supervision of node labels and a self-supervised task using de-noised auto-encoding. The latent graph structure is generated through a fully-parameterized adjacency matrix or a KNN construction subsequent to passing node features to an MLP. Experimental results and corresponding analyses demonstrate the effectiveness of the proposed model.\n  \nPros:  \n1. Using self-supervision to guide graph structure learning sounds like an interesting and promising idea.   \n2. Most claims are supported by experimental results. Specifically, comparisons with baselines show that the proposed model can alleviate the problems of previous works such as sensitivity to the similarity metric and high-quality initial graph structure, memory problems, etc. Datasets like Wine, 20news, etc., demonstrate the model's validity for tasks without graph structure. Parameter sensitivity and other ablation studies are also conducted.  \n3. Overall, the paper is well written and easy to follow.   \n\nCons:  \n1. Some model designs are not entirely clear. For example, based on my understanding, the KNN step is not differentiable, thus the model with MLP-kNN generator cannot be trained end-to-end. The authors may want to detail how to cope with this problem in the training process.  \n2. I also have a question regarding the model design regarding the self-supervised loss. If neglecting the supervised part (i.e., GNNc and its loss in Figure 1), the self-supervised part basically tries to reconstruct the original feature from the noisy features using GNN_DAE. In that case, should KNN using initial features be the optimal solution (since GCN is essentially smoothing features and these nodes have the most similar features)? If that is true, I feel that the self-supervision may actually work like a regularizer to the MLP subsequent to KNN.  \n3. The experiments show that the symmetrization step is very useful while lacking analysis of why it works. Could the authors provide further exploration in that aspect?  \n4. It seems bizarre that only LDS is adopted as the baseline in Table 2 since the codes and datasets are ready. I suggest the authors at least add IDGL, the most competitive baseline in Table 1, into Table 2.   \n4. There's little comparison of the learned graph structure with the original one (Figure 3.e only utilize node labels but not the original graph structure). Since the main goal of the paper is to learn GNN and graph structure simultaneously, the learned graph structure is important and should be analyzed empirically. For example, the authors may want to follow previous works to design certain metrics to compare the learned graph structure with the original one, provide some visualizations, or case studies in synthesis graphs.  \n\nMinor:  \n1. The complete parameter sensitivity could be added to the supplementary material. For example, I am interested to see how the sensitivity of parameter k for kNN behaves across different datasets with different scales.  \n2. Some related works are missing, e.g., [1-3].  \n\n[1] Adaptive Graph Convolutional Neural Networks, AAAI 2018.  \n[2] Topology Optimization based Graph Convolutional Network, IJCAI 2019.  \n[3] Graph Structure Learning for Robust Graph Neural Networks, KDD 2020.  \n\nOverall, though I like the paper in general, I believe the paper could be further improved and thus vote for weakly rejection. I am happy to increase my scores if the authors can address my above concerns.   \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Learning graph structures with SSL sounds interesting in general, but some clarifications of the model design are needed", "review": "Summary:  \nThis paper proposes to tackle jointly learning graph structures and GNN parameters without accessing the original graph structure. Specifically, the proposed method adopts a self-supervised auxiliary task, i.e., parallel training using the supervision of node labels and a self-supervised task using de-noised auto-encoding. The latent graph structure is generated through a fully-parameterized adjacency matrix or a KNN construction subsequent to passing node features to an MLP. Experimental results and corresponding analyses demonstrate the effectiveness of the proposed model.\n  \nPros:  \n1. Using self-supervision to guide graph structure learning sounds like an interesting and promising idea.   \n2. Most claims are supported by experimental results. Specifically, comparisons with baselines show that the proposed model can alleviate the problems of previous works such as sensitivity to the similarity metric and high-quality initial graph structure, memory problems, etc. Datasets like Wine, 20news, etc., demonstrate the model's validity for tasks without graph structure. Parameter sensitivity and other ablation studies are also conducted.  \n3. Overall, the paper is well written and easy to follow.   \n\nCons:  \n1. Some model designs are not entirely clear. For example, based on my understanding, the KNN step is not differentiable, thus the model with MLP-kNN generator cannot be trained end-to-end. The authors may want to detail how to cope with this problem in the training process.  \n2. I also have a question regarding the model design regarding the self-supervised loss. If neglecting the supervised part (i.e., GNNc and its loss in Figure 1), the self-supervised part basically tries to reconstruct the original feature from the noisy features using GNN_DAE. In that case, should KNN using initial features be the optimal solution (since GCN is essentially smoothing features and these nodes have the most similar features)? If that is true, I feel that the self-supervision may actually work like a regularizer to the MLP subsequent to KNN.  \n3. The experiments show that the symmetrization step is very useful while lacking analysis of why it works. Could the authors provide further exploration in that aspect?  \n4. It seems bizarre that only LDS is adopted as the baseline in Table 2 since the codes and datasets are ready. I suggest the authors at least add IDGL, the most competitive baseline in Table 1, into Table 2.   \n4. There's little comparison of the learned graph structure with the original one (Figure 3.e only utilize node labels but not the original graph structure). Since the main goal of the paper is to learn GNN and graph structure simultaneously, the learned graph structure is important and should be analyzed empirically. For example, the authors may want to follow previous works to design certain metrics to compare the learned graph structure with the original one, provide some visualizations, or case studies in synthesis graphs.  \n\nMinor:  \n1. The complete parameter sensitivity could be added to the supplementary material. For example, I am interested to see how the sensitivity of parameter k for kNN behaves across different datasets with different scales.  \n2. Some related works are missing, e.g., [1-3].  \n\n[1] Adaptive Graph Convolutional Neural Networks, AAAI 2018.  \n[2] Topology Optimization based Graph Convolutional Network, IJCAI 2019.  \n[3] Graph Structure Learning for Robust Graph Neural Networks, KDD 2020.  \n\nOverall, though I like the paper in general, I believe the paper could be further improved and thus vote for weakly rejection. I am happy to increase my scores if the authors can address my above concerns.   \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603868314266}, {"id": "GlOs9xwewkz", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper904/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Graph neural networks (GNNs) have become de facto methods for integrating the input graph structure and node features to learn effective node representations. However, in some domains (such as brain signals, particle reconstruction, etc.), there is access to only node features (but not the underlying graph structure). Motivated by the fact that GNNs tend to perform poorly in the absence of the graph structure, the paper\n1) proposes a self-supervised framework for generating the graph structure,\n2) explores the design space of the framework by comparing different graph generation methods,\n3) demonstrates the framework's effectiveness on node classification datasets.\n\n\n\n##  Pros\n+ [Motivation] A strength of the paper is the motivation of the problem. It is well-known that the performance of GNNs is highly sensitive to the quality of the input graph structure. Hence, in domains such as brain signals, particle reconstruction, etc., it is necessary to generate a high-quality graph structure so that GNNs could be effectively leveraged. \n+ [Presentation] The high-level ideas of the paper are easy to read with clear figures and notation.\n+ [Relevance] The topic of GNNs has gained increasing attention recently such that a significant portion of the ICLR community should be interested.\n\n\n\n## Cons\nThe main weaknesses of the paper are along the axis of the significance of the contributions. The paper requires thorough discussions / positioning / comparisons with many existing publications in GNN literature. The detailed comments are as follows.\n- [Self Supervision] The framework proposed in the paper can be seen as an instance of a general framework for self-supervision proposed in SS-GCN (When Does Self-Supervision Help Graph Convolutional Networks?, In ICML'20). \n- [Self Training] A closely related idea for self-supervision is self-training. The basic idea of self-training is to add high-confident predictions to the training set to increase supervision. The paper should be positioned with (and empirically compare against) self-training-based approaches such as (but not limited to)\ni) GAM (Graph Agreement Models for Semi-Supervised Learning, In NeurIPS'19) that can also handle noisy / learn graph structures, \nii) AdaEdge (Measuring and Relieving the Over-smoothing Problem for Graph Neural Networks from the Topological View, In AAAI'20).\n- [Quality] Regarding experiments, the dataset domains (such as citation networks) considered in the paper are those in which the graph structure is known. So, it would be more convincing if the adjacency of the dataset was used as the initial adjacency matrix in the proposed framework. The baselines to compare against would then be models such as SS-GCN, GAM, etc.  \n- [Soundness] It is unclear why datasets from domains discussed in the introduction section were not considered in the experiments. These domains include, as listed in the paper, brain signal classification, computer-aided diagnosis, analysis of computer programs, and particle reconstruction.  \n- [Clarity] Though the high-level ideas were easy to read, the paper should clearly discuss finer details. In particular, in section 4.1 (paragraph preceding 4.2), the details of adjacency initialisation are hastily mentioned, and it feels like the discussion mixes adjacency and weight matrices up. \n- [Metric Learning] An idea for learning graph structures for GNNs is to learn the underlying similarity metric (rather than use a chosen one). This idea has been used in certain domains of computer vision and natural language processing. The paper should be positioned with relevant publications (albeit for small graphs) including (but not limited to)\ni) Few-Shot Learning with Graph Neural Networks, In ICLR'18\nii) Graph Neural Networks with Generated Parameters for Relation Extraction, In ACL'19", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Important problem; contributions are of limited significance", "review": "Graph neural networks (GNNs) have become de facto methods for integrating the input graph structure and node features to learn effective node representations. However, in some domains (such as brain signals, particle reconstruction, etc.), there is access to only node features (but not the underlying graph structure). Motivated by the fact that GNNs tend to perform poorly in the absence of the graph structure, the paper\n1) proposes a self-supervised framework for generating the graph structure,\n2) explores the design space of the framework by comparing different graph generation methods,\n3) demonstrates the framework's effectiveness on node classification datasets.\n\n\n\n##  Pros\n+ [Motivation] A strength of the paper is the motivation of the problem. It is well-known that the performance of GNNs is highly sensitive to the quality of the input graph structure. Hence, in domains such as brain signals, particle reconstruction, etc., it is necessary to generate a high-quality graph structure so that GNNs could be effectively leveraged. \n+ [Presentation] The high-level ideas of the paper are easy to read with clear figures and notation.\n+ [Relevance] The topic of GNNs has gained increasing attention recently such that a significant portion of the ICLR community should be interested.\n\n\n\n## Cons\nThe main weaknesses of the paper are along the axis of the significance of the contributions. The paper requires thorough discussions / positioning / comparisons with many existing publications in GNN literature. The detailed comments are as follows.\n- [Self Supervision] The framework proposed in the paper can be seen as an instance of a general framework for self-supervision proposed in SS-GCN (When Does Self-Supervision Help Graph Convolutional Networks?, In ICML'20). \n- [Self Training] A closely related idea for self-supervision is self-training. The basic idea of self-training is to add high-confident predictions to the training set to increase supervision. The paper should be positioned with (and empirically compare against) self-training-based approaches such as (but not limited to)\ni) GAM (Graph Agreement Models for Semi-Supervised Learning, In NeurIPS'19) that can also handle noisy / learn graph structures, \nii) AdaEdge (Measuring and Relieving the Over-smoothing Problem for Graph Neural Networks from the Topological View, In AAAI'20).\n- [Quality] Regarding experiments, the dataset domains (such as citation networks) considered in the paper are those in which the graph structure is known. So, it would be more convincing if the adjacency of the dataset was used as the initial adjacency matrix in the proposed framework. The baselines to compare against would then be models such as SS-GCN, GAM, etc.  \n- [Soundness] It is unclear why datasets from domains discussed in the introduction section were not considered in the experiments. These domains include, as listed in the paper, brain signal classification, computer-aided diagnosis, analysis of computer programs, and particle reconstruction.  \n- [Clarity] Though the high-level ideas were easy to read, the paper should clearly discuss finer details. In particular, in section 4.1 (paragraph preceding 4.2), the details of adjacency initialisation are hastily mentioned, and it feels like the discussion mixes adjacency and weight matrices up. \n- [Metric Learning] An idea for learning graph structures for GNNs is to learn the underlying similarity metric (rather than use a chosen one). This idea has been used in certain domains of computer vision and natural language processing. The paper should be positioned with relevant publications (albeit for small graphs) including (but not limited to)\ni) Few-Shot Learning with Graph Neural Networks, In ICLR'18\nii) Graph Neural Networks with Generated Parameters for Relation Extraction, In ACL'19", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602700196242}], "openreview_url": "https://openreview.net/forum?id=a5KvtsZ14ev", "arxiv_id": "2102.05034", "paper_pdf": "papers/a5KvtsZ14ev.pdf", "paper_pdf_sha256": "bb9e2f129bf24419b4ead3ef650b6c585b0bcee101cc9fefd5558f2bc3ffce93", "paper_pdf_bytes": 634660, "paper_pdf_source": "openreview", "code_url": "https://github.com/BorealisAI/SLAPS-GNN", "code_repository": "BorealisAI/SLAPS-GNN", "code_commit": "489481fd3fa1ba0be6b1d42b40acec8b8858b7ec", "code_archive": "repos/a5KvtsZ14ev.zip", "code_archive_sha256": "e11aeddc36869c9f631c3b1f4479aea731b9754a62bb65295b5934113067dd5d", "code_archive_bytes": 22056, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 29, "github_languages": {"Python": 41538}, "github_archived": false, "github_pushed_at": "2021-10-25T22:12:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/slaps-self-supervision-improves-structure-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkxWpCNKvS", "year": 2020, "status": "rejected", "title": "Improved Image Augmentation for Convolutional Neural Networks by Copyout and CopyPairing", "authors": ["Philip May"], "authorids": ["eniak.info@gmail.com"], "authors_source": "OpenReview API", "abstract": "Image augmentation is a widely used technique to improve the performance of convolutional neural networks (CNNs). In common image shifting, cropping, flipping, shearing and rotating are used for augmentation. But there are more advanced techniques like Cutout and SamplePairing.\n\nIn this work we present two improvements of the state-of-the-art Cutout and SamplePairing techniques. Our new method called Copyout takes a square patch of another random training image and copies it onto a random location of each image used for training. The second technique we discovered is called CopyPairing. It combines Copyout and SamplePairing for further augmentation and even better performance.\n\nWe apply different experiments with these augmentation techniques on the CIFAR-10 dataset to evaluate and compare them under different configurations. In our experiments we show that Copyout reduces the test error rate by 8.18% compared with Cutout and 4.27% compared with SamplePairing. CopyPairing reduces the test error rate by 11.97% compared with Cutout and 8.21% compared with SamplePairing.\n\nCopyout and CopyPairing implementations are available at https://github.com/anonym/anonym.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJgZAoMH9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1382/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an extensions to the CutOut and SamplePairing techniques for image augmentation, CopyOut and CopyPairing. CutOut itself consists of randomly masking out a rectangular region of an image. In CopyOut one chooses a source and target images, and a rectangular region from the source image is copied into target image.\n\n\nThough the extensions seem to provide an improvement in performance, we feel there are a few improvements that prevent the paper from being accepted:\n - more thorough experimental verification with various CNN architectures and larger variety of datasets datasets is needed (the original paper conducted experiments on CIFAR-10, CIFAR-100, SVHN and STL-10).\n- better exposition needed - it would have been helpful to include more examples of the original methods CutOut and Sample pairing. Sample pairing though mentioned is not described. As CopyPairing is largely based on SamplePairing, it would have been helpful to include a diagram or thorough description of it.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper presents an extensions to the CutOut and SamplePairing techniques for image augmentation, CopyOut and CopyPairing. CutOut itself consists of randomly masking out a rectangular region of an image. In CopyOut one chooses a source and target images, and a rectangular region from the source image is copied into target image.\n\n\nThough the extensions seem to provide an improvement in performance, we feel there are a few improvements that prevent the paper from being accepted:\n - more thorough experimental verification with various CNN architectures and larger variety of datasets datasets is needed (the original paper conducted experiments on CIFAR-10, CIFAR-100, SVHN and STL-10).\n- better exposition needed - it would have been helpful to include more examples of the original methods CutOut and Sample pairing. Sample pairing though mentioned is not described. As CopyPairing is largely based on SamplePairing, it would have been helpful to include a diagram or thorough description of it."}, "tcdate": 1572314056998}, {"id": "ryg1OCnTtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1382/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "[Summary]\nThis paper proposes two data augmentation methods that combine cutout [1] and sample paring [2] for training CNNs, Copyout and Copyparing. The authors evaluate their methods on the CIFAR-10 dataset.\n\n[Pros]\n- Data augmentation is an important regularization method for training diverse NN models\n\n[Cons]\n- The main issue is novelty. What are the differences of the proposed methods from CutMix [3] and RICAP [4]?\n- Only CIFAR-10 was used for evaluation. The results on ImageNet-1k are essential.\n- In recent papers, data augmentation methods for training CNN backbones should be validated on various architectures and downstream tasks such as object detection and semantic segmentation.\n- The method description is not specific.\n\n[1] Devries and Taylor.  Improved regularization of convolutional neural networks with cutout, ArXiv 2017.\n[2] Inoue, Data augmentation by pairing samples for images classification, ArXiv 2018.\n[3] Yun et al. CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, ArXiv 2019.\n[4] Takahasi et al. Data Augmentation using Random Image Cropping and Patching for Deep CNNs, ACML 2018.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "[Summary]\nThis paper proposes two data augmentation methods that combine cutout [1] and sample paring [2] for training CNNs, Copyout and Copyparing. The authors evaluate their methods on the CIFAR-10 dataset.\n\n[Pros]\n- Data augmentation is an important regularization method for training diverse NN models\n\n[Cons]\n- The main issue is novelty. What are the differences of the proposed methods from CutMix [3] and RICAP [4]?\n- Only CIFAR-10 was used for evaluation. The results on ImageNet-1k are essential.\n- In recent papers, data augmentation methods for training CNN backbones should be validated on various architectures and downstream tasks such as object detection and semantic segmentation.\n- The method description is not specific.\n\n[1] Devries and Taylor.  Improved regularization of convolutional neural networks with cutout, ArXiv 2017.\n[2] Inoue, Data augmentation by pairing samples for images classification, ArXiv 2018.\n[3] Yun et al. CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, ArXiv 2019.\n[4] Takahasi et al. Data Augmentation using Random Image Cropping and Patching for Deep CNNs, ACML 2018.\n\n"}, "tcdate": 1571831398910}, {"id": "SylmDA_-FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1382/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I think this paper is not enough to accept in ICLR because\n- Lack of novelty.\n  - CutMix [1] is very similar to Copyout.\n  - To verify the novelty, a more sophisticated description and experimental supports should be required.\n- Insufficient experiments for supporting the effectiveness of the proposed method.\n  - 6-layer convolutional networks are used, but other architectures, e.g., ResNet, should be demonstrated.\n  - Other datasets, e.g., CIFAR100 and ImageNet, should be demonstrated.\n  - Various settings, e.g., the number of training samples is limited, should be demonstrated.\n  - Need comparison with other augmentation methods, e.g., Mixup, CutMix, AutoAugment.\n- Overall, the paper is awkwardly written.\n\n[1] Sangdoo Yun et al. \"Cutmix: Regularization strategy to train strong classifiers with localizable features.\" ICCV 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "I think this paper is not enough to accept in ICLR because\n- Lack of novelty.\n  - CutMix [1] is very similar to Copyout.\n  - To verify the novelty, a more sophisticated description and experimental supports should be required.\n- Insufficient experiments for supporting the effectiveness of the proposed method.\n  - 6-layer convolutional networks are used, but other architectures, e.g., ResNet, should be demonstrated.\n  - Other datasets, e.g., CIFAR100 and ImageNet, should be demonstrated.\n  - Various settings, e.g., the number of training samples is limited, should be demonstrated.\n  - Need comparison with other augmentation methods, e.g., Mixup, CutMix, AutoAugment.\n- Overall, the paper is awkwardly written.\n\n[1] Sangdoo Yun et al. \"Cutmix: Regularization strategy to train strong classifiers with localizable features.\" ICCV 2019.\n"}, "tcdate": 1571028571358}], "openreview_url": "https://openreview.net/forum?id=rkxWpCNKvS", "arxiv_id": "1909.00390", "paper_pdf": "papers/rkxWpCNKvS.pdf", "paper_pdf_sha256": "ff24017a0a807d407b345a9fc932691e5194fdc0fb7eb2437e3b12665362a49a", "paper_pdf_bytes": 146220, "paper_pdf_source": "openreview", "code_url": "https://github.com/t-systems-on-site-services-gmbh/coocop", "code_repository": "t-systems-on-site-services-gmbh/coocop", "code_commit": "95f844404e22a99cc93a058fa2bd085685dc88c7", "code_archive": "repos/rkxWpCNKvS.zip", "code_archive_sha256": "41be181c2fc2c6ddd2f1d1305314665e4518ae3ab8f12268748bb888e7980987", "code_archive_bytes": 6027, "code_file_count": 5, "code_extensions": {".py": 4, ".sh": 1}, "github_disk_usage_kb": 25, "github_languages": {"Python": 9552, "Shell": 201}, "github_archived": false, "github_pushed_at": "2019-10-11T19:13:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improved-image-augmentation-for-convolutional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJNRHiAcYX", "year": 2019, "status": "rejected", "title": "Boosting Trust Region Policy Optimization by Normalizing flows Policy", "authors": ["Yunhao Tang", "Shipra Agrawal"], "authorids": ["yt2541@columbia.edu", "sa3305@columbia.edu"], "authors_source": "OpenReview API", "abstract": "We propose to improve trust region policy search with normalizing flows policy. We illustrate that when the trust region is constructed by KL divergence constraint, normalizing flows policy can generate samples far from the 'center' of the previous policy iterate, which potentially enables better exploration and helps avoid bad local optima. We show that normalizing flows policy significantly improves upon factorized Gaussian policy baseline, with both TRPO and ACKTR, especially on tasks with complex dynamics such as Humanoid.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rkeZNVXihX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper135/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors in this work present an approach to policy optimization that relies on an alternative policy formulation based on normalizing flows. This is a relatively simple modification (this is no criticism) that essentially uses the same TRPO algorithm as previous approaches, but a different mechanism for generating the distribution over actions. The crux of the authors’ approach is detailed in equations (6) and (7), although it could have been useful to see more of the discussion of the architecture from appendix B in the actual text of the paper.\n\nThe authors then go on to analyze the properties and expressiveness of the resulting properties and show that it is more capable of capturing complex interactions than a simple Gaussian. It was somewhat unclear, however, in section 4.2 what the exact form of the policies being compared are. Is this a simple example with only the parameters of the Gaussian, or was the Gaussian parameterized by a multi-layer model? Further, one thing I would also have liked to see the authors question more is, for the problems they attack, whether this expressiveness is more useful “during exploration” or for the ultimate performance of the final policy.\n\nThe authors, finally, show that this approach is able to out-perform the alternative Gaussian policy. Ultimately this approach seems to be a simple modification (or replacement) of the standard policy formulation, and one that seems to lead to good performance gains. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple modification with relatively robust gains", "review": "The authors in this work present an approach to policy optimization that relies on an alternative policy formulation based on normalizing flows. This is a relatively simple modification (this is no criticism) that essentially uses the same TRPO algorithm as previous approaches, but a different mechanism for generating the distribution over actions. The crux of the authors’ approach is detailed in equations (6) and (7), although it could have been useful to see more of the discussion of the architecture from appendix B in the actual text of the paper.\n\nThe authors then go on to analyze the properties and expressiveness of the resulting properties and show that it is more capable of capturing complex interactions than a simple Gaussian. It was somewhat unclear, however, in section 4.2 what the exact form of the policies being compared are. Is this a simple example with only the parameters of the Gaussian, or was the Gaussian parameterized by a multi-layer model? Further, one thing I would also have liked to see the authors question more is, for the problems they attack, whether this expressiveness is more useful “during exploration” or for the ultimate performance of the final policy.\n\nThe authors, finally, show that this approach is able to out-perform the alternative Gaussian policy. Ultimately this approach seems to be a simple modification (or replacement) of the standard policy formulation, and one that seems to lead to good performance gains. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541252136552}, {"id": "SJxd2NNKhX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper135/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper generalizes basic policy gradient methods by replacing the original Gaussian or Gaussian mixture policy with a normalizing flow policy, which is defined by a sequence of invertible transformations from a base policy.\n\nAlthough the concept of normalizing flow is simple, and it has been applied to other models such as VAE, there seems no work on applying it for policy optimization. Thus I think this method is itself interesting.\n\nHowever, I find the paper written in a way assuming readers very familiar with related concept and algorithms in reinforcement learning. Thus although one can get the general idea on how the method works, it might be difficult to get a deeper understanding on some details.\n\nFor example, normalizing flows are defined in Section 4, and then it is directly claimed that normalizing flows can be applied to policy optimization, without giving details on how it is actually applied, e.g., what is the objective function? and why one needs to compute gradients of the entropy (Section 4.1)?\n\nAlso, in the experiments, it is said that one can combing normalizing flows with TRPO without describing the details. I can't get how exactly normalizing flows + TRPO works.\n\nThe experiments also talk about 2D bandit problem, and again, without any descriptions. BTW, in the Section 4.3, what does [-1, 1]^2 mean? (I have seen {-1, 1}^2, but not [-1, 1]^2).\n\nIt seems that the authors only use the basic normalizing flow structures studied in Rezende&Mohamed (2015) and Dinh et al (2016). However, there are more powerful variants of normalizing flows such as the Multiplicative Normalizing Flows or the Glow. I wonder how good the results are if these more advanced versions are used. Maybe they can uniformly outperform Gaussian policy?\n\nUpdate:\nI feel the idea of this paper is straightforward, and the contribution is incremental. To improve the paper, stronger experiments need to be performed. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "simple idea but not totally clear", "review": "This paper generalizes basic policy gradient methods by replacing the original Gaussian or Gaussian mixture policy with a normalizing flow policy, which is defined by a sequence of invertible transformations from a base policy.\n\nAlthough the concept of normalizing flow is simple, and it has been applied to other models such as VAE, there seems no work on applying it for policy optimization. Thus I think this method is itself interesting.\n\nHowever, I find the paper written in a way assuming readers very familiar with related concept and algorithms in reinforcement learning. Thus although one can get the general idea on how the method works, it might be difficult to get a deeper understanding on some details.\n\nFor example, normalizing flows are defined in Section 4, and then it is directly claimed that normalizing flows can be applied to policy optimization, without giving details on how it is actually applied, e.g., what is the objective function? and why one needs to compute gradients of the entropy (Section 4.1)?\n\nAlso, in the experiments, it is said that one can combing normalizing flows with TRPO without describing the details. I can't get how exactly normalizing flows + TRPO works.\n\nThe experiments also talk about 2D bandit problem, and again, without any descriptions. BTW, in the Section 4.3, what does [-1, 1]^2 mean? (I have seen {-1, 1}^2, but not [-1, 1]^2).\n\nIt seems that the authors only use the basic normalizing flow structures studied in Rezende&Mohamed (2015) and Dinh et al (2016). However, there are more powerful variants of normalizing flows such as the Multiplicative Normalizing Flows or the Glow. I wonder how good the results are if these more advanced versions are used. Maybe they can uniformly outperform Gaussian policy?\n\nUpdate:\nI feel the idea of this paper is straightforward, and the contribution is incremental. To improve the paper, stronger experiments need to be performed. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541125296475}, {"id": "B1eWSaTuhX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper135/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The papers proposed to use normalizing flow policies instead of Gaussian policies to improve exploration and achieve better sample complexity in practice. While I believe this idea has not specifically been tried in previous literature and the vague intuition that NF leads to more exploration that helps learning a better policy, the novelty of combining these two seems limited, and the paper does not seem to provide enough justification to using NF policies instead of alternative policy distributions both in theory and in the experiments.\n\n1. About Section 4.2. I believe that the normalizing flow in question would transform the volume of a Gaussian? So there would exist some parameter setting for a flow model that also shrinks volume, thereby resulting in lower variance policies? The arguments would thereby depend heavily on the specific architecture and initialization of the flow model, which is not discussed in detail. \n\nAlso, why is finding a high variance policy better in terms of the trust region argument? Isn't the whole point of using trust region that the new policy should be closer to old policy to prevent performance degradation? I also think that a fair comparison would be compare KL between normalizing flow policies, instead of KL between NF and Gaussian.\n\n2. The TRPO experiments seem wrong -- at least the results don't match what is reported in the ACKTR paper for Reacher and InverseDoublePendulum envs -- there the TRPO policy at least learns something. Also TRPO in general does not perform as bad as it may seem, see \"Deep RL that matters\" paper by Henderson et al. Maybe this is because of using OpenAI baselines code which seems to have worse TRPO performance.\n\nThere is also no experiments on ACKTR on the small Mujoco tasks (even in the Appendix), which seems to be a rather big oversight given the authors have already done even harder tasks for ACKTR + NF.\n\nMoreover I think a fair comparison is to use almost the same architecture for implicit and gaussian, where the only difference is where you sample the noise. For Gaussian with flows, you can first use an MLP to produce deterministic outputs and then use flow to generate the mean actions. Otherwise it is impossible to say whether the architecture or the implicit distribution contributes more to the success.\n\nOne could also use truncated Gaussian distributions / Beta distributions / Gaussian + tanh, since Mujoco actions beyond (-1, 1) is treated as -1 or 1, so Gaussian should already be bad. It is unclear whether NF is able to outperform these settings. \n\nMinor points:\n\n- Fix citations. Please use \\citep throughout.\n- Is Equation (6) correct? Seems like \\Sigma_i should be the inverse of g_i(\\epsilon)? Also this is the \"change of variables formula\" not \"chain rule\".\n- Why is normalizing flow not part of the background?\n- Add legends in Figure (1)\n- Figure 2(c), I believe with max entropy you could already obtain diverse ant trajectories?\n- I believe in the context of generative models, \"implicit\" typically means the case where likelihood is not tractable? Here the likelihood is perfectly tractable.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Combining normalizing flows and Gaussian policies is relatively new, but justification is very limited", "review": "The papers proposed to use normalizing flow policies instead of Gaussian policies to improve exploration and achieve better sample complexity in practice. While I believe this idea has not specifically been tried in previous literature and the vague intuition that NF leads to more exploration that helps learning a better policy, the novelty of combining these two seems limited, and the paper does not seem to provide enough justification to using NF policies instead of alternative policy distributions both in theory and in the experiments.\n\n1. About Section 4.2. I believe that the normalizing flow in question would transform the volume of a Gaussian? So there would exist some parameter setting for a flow model that also shrinks volume, thereby resulting in lower variance policies? The arguments would thereby depend heavily on the specific architecture and initialization of the flow model, which is not discussed in detail. \n\nAlso, why is finding a high variance policy better in terms of the trust region argument? Isn't the whole point of using trust region that the new policy should be closer to old policy to prevent performance degradation? I also think that a fair comparison would be compare KL between normalizing flow policies, instead of KL between NF and Gaussian.\n\n2. The TRPO experiments seem wrong -- at least the results don't match what is reported in the ACKTR paper for Reacher and InverseDoublePendulum envs -- there the TRPO policy at least learns something. Also TRPO in general does not perform as bad as it may seem, see \"Deep RL that matters\" paper by Henderson et al. Maybe this is because of using OpenAI baselines code which seems to have worse TRPO performance.\n\nThere is also no experiments on ACKTR on the small Mujoco tasks (even in the Appendix), which seems to be a rather big oversight given the authors have already done even harder tasks for ACKTR + NF.\n\nMoreover I think a fair comparison is to use almost the same architecture for implicit and gaussian, where the only difference is where you sample the noise. For Gaussian with flows, you can first use an MLP to produce deterministic outputs and then use flow to generate the mean actions. Otherwise it is impossible to say whether the architecture or the implicit distribution contributes more to the success.\n\nOne could also use truncated Gaussian distributions / Beta distributions / Gaussian + tanh, since Mujoco actions beyond (-1, 1) is treated as -1 or 1, so Gaussian should already be bad. It is unclear whether NF is able to outperform these settings. \n\nMinor points:\n\n- Fix citations. Please use \\citep throughout.\n- Is Equation (6) correct? Seems like \\Sigma_i should be the inverse of g_i(\\epsilon)? Also this is the \"change of variables formula\" not \"chain rule\".\n- Why is normalizing flow not part of the background?\n- Add legends in Figure (1)\n- Figure 2(c), I believe with max entropy you could already obtain diverse ant trajectories?\n- I believe in the context of generative models, \"implicit\" typically means the case where likelihood is not tractable? Here the likelihood is perfectly tractable.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541098809416}], "openreview_url": "https://openreview.net/forum?id=SJNRHiAcYX", "arxiv_id": "1809.10326", "paper_pdf": "papers/SJNRHiAcYX.pdf", "paper_pdf_sha256": "91db1cb255b3600befff0f4db9014ede4425171c84b729f9bcd92e7dc5065c87", "paper_pdf_bytes": 4355503, "paper_pdf_source": "openreview", "code_url": "https://github.com/robintyh1/onpolicybaselines", "code_repository": "robintyh1/onpolicybaselines", "code_commit": "58d401622b38a9127437cf48b1b62b27028d694a", "code_archive": "repos/SJNRHiAcYX.zip", "code_archive_sha256": "c421d56b8f02615e28b2c5eafb81d3ac91f10049375defedaa9c3da8a85b1e19", "code_archive_bytes": 138276, "code_file_count": 61, "code_extensions": {".py": 61}, "github_disk_usage_kb": 56, "github_languages": {"Python": 370267}, "github_archived": false, "github_pushed_at": "2020-04-03T18:54:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/boosting-trust-region-policy-optimization-by"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hk2MHt-3-", "year": 2018, "status": "rejected", "title": "Coupled Ensembles of Neural Networks", "authors": ["Anuvabh Dutt", "Denis Pellerin", "Georges Quénot"], "authorids": ["anuvabh.dutt@univ-grenoble-alpes.fr", "denis.pellerin@gipsa-lab.grenoble-inp.fr", "georges.quenot@imag.fr"], "authors_source": "OpenReview API", "abstract": "We investigate in this paper the architecture of deep convolutional networks. Building on existing state of the art models, we propose a reconfiguration of the model parameters into several parallel branches at the global network level, with each branch being a standalone CNN. We show that this arrangement is an efficient way to significantly reduce the number of parameters while at the same time improving the performance. The use of branches brings an additional form of regularization. In addition to splitting the parameters into parallel branches, we propose a tighter coupling of these branches by averaging their log-probabilities. The tighter coupling favours the learning of better representations, even at the level of the individual branches, as compared to when each branch is trained independently. We refer to this branched architecture as \"coupled ensembles\". The approach is very generic and can be applied with almost any neural network architecture. With coupled ensembles of DenseNet-BC and parameter budget of 25M, we obtain error rates of 2.92%, 15.68% and 1.50% respectively on CIFAR-10, CIFAR-100 and SVHN tasks. For the same parameter budget, DenseNet-BC has an error rate of 3.46%, 17.18%, and 1.8% respectively.  With ensembles of coupled ensembles, of DenseNet-BC networks, with 50M total parameters, we obtain error rates of 2.72%, 15.13% and 1.42% respectively on these tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hk8Nwx9xf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper3/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Strengths:\n* Very simple approach, amounting to coupled training of \"e\" identical copies  of a chosen net architecture, whose predictions are fused during training. This forces the different model instances to become more complementary.\n* Perhaps counterintuitively, experiments also show that coupled ensembling leads to individual nets that perform better than those produced by separate training.\n* The practical advantages of the proposed approach are twofold:\n1. Given a fixed parameter budget, coupled ensembling leads to better accuracy than a single net or an ensemble of disjointly-trained nets.\n2. For the same accuracy, coupled ensembling yields significant parameter savings.\n\nWeaknesses:\n* Although results are very strong, the proposed models do not outperform the state-of-the-art, except for the models reported in Table 4, which however were obtained by *traditional* ensembling of coupled ensembles. \n* Coupled ensembling requires joint training of all nets in the ensemble and thus is limited by the size of the model that can be fit in memory. Conversely, traditional ensembling involves separate training of the different instances and this enables the learning of an arbitrary number of individual nets. \n* I am surprised by the results in Table 2, which suggest that the optimal number of nets in the ensemble is remarkably low (only 3!). It'd be valuable to understand whether this kind of result holds for other network architectures or whether it is specific to this choice of net.\n* Strictly speaking it is correct to refer to the individual nets in the ensembles as \"branches\" and \"basic blocks.\" Nevertheless, I find the use of these terms confusing in the context of the proposed approach, since they are commonly used to denote concepts different from those represented here.  I would recommend refraining from using these terms here.\n\nOverall, the paper provides limited technical novelty. Yet, it reveals some interesting empirical findings about the benefits of coordinated training of models in an ensemble.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "simple approach, shows parameter-saving benefits of coupled ensembling", "rating": "6: Marginally above acceptance threshold", "review": "Strengths:\n* Very simple approach, amounting to coupled training of \"e\" identical copies  of a chosen net architecture, whose predictions are fused during training. This forces the different model instances to become more complementary.\n* Perhaps counterintuitively, experiments also show that coupled ensembling leads to individual nets that perform better than those produced by separate training.\n* The practical advantages of the proposed approach are twofold:\n1. Given a fixed parameter budget, coupled ensembling leads to better accuracy than a single net or an ensemble of disjointly-trained nets.\n2. For the same accuracy, coupled ensembling yields significant parameter savings.\n\nWeaknesses:\n* Although results are very strong, the proposed models do not outperform the state-of-the-art, except for the models reported in Table 4, which however were obtained by *traditional* ensembling of coupled ensembles. \n* Coupled ensembling requires joint training of all nets in the ensemble and thus is limited by the size of the model that can be fit in memory. Conversely, traditional ensembling involves separate training of the different instances and this enables the learning of an arbitrary number of individual nets. \n* I am surprised by the results in Table 2, which suggest that the optimal number of nets in the ensemble is remarkably low (only 3!). It'd be valuable to understand whether this kind of result holds for other network architectures or whether it is specific to this choice of net.\n* Strictly speaking it is correct to refer to the individual nets in the ensembles as \"branches\" and \"basic blocks.\" Nevertheless, I find the use of these terms confusing in the context of the proposed approach, since they are commonly used to denote concepts different from those represented here.  I would recommend refraining from using these terms here.\n\nOverall, the paper provides limited technical novelty. Yet, it reveals some interesting empirical findings about the benefits of coordinated training of models in an ensemble.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511814958514}, {"id": "SJXrqMPgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper3/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposed a reconfiguration of the existing state-of-the-art CNN model architectures including ResNet and DensNet. By introducing new branching architecture, coupled ensembles, they demonstrate that the model can achieve better performance in classification tasks compared with the single branch counterpart with same parameter budget. Additionally, they also show that the proposed ensemble method results in better performance than other ensemble methods (For example, ensemble over independently trained models)  not only in combined mode but also in individual branches.\n\nPaper Strengths:\n* The proposed coupled ensembles method truly show impressive results in classification benchmark (DenseNet-BC L = 118 k = 35 e = 3).\n* Detailed analysis on different ensemble fusion methods on both training time and testing time.\n* Simple but effective design to achieve a better result in testing time with same total parameter budget.\n\t\nPaper Weakness:\n* Some detail about different fusing method should be mentioned in the main paper instead of in the supplementary material.\n* In practice, how much more GPU memory is required to train the model with parallel branches (with same parameter budgets) because memory consumption is one of the main problems of networks with multiple branches.\n* At least one experiment should be carried out on a larger dataset such as ImageNet to further demonstrate the validity of the proposed method.\n* More analysis can be conducted on the training process of the model. Will it converge faster? What will be the total required training time to reach the same performance compared with single branch model with the same parameter budget?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This work proposed a reconfiguration of the existing state-of-the-art CNN model using a new branching architecture.", "rating": "6: Marginally above acceptance threshold", "review": "This work proposed a reconfiguration of the existing state-of-the-art CNN model architectures including ResNet and DensNet. By introducing new branching architecture, coupled ensembles, they demonstrate that the model can achieve better performance in classification tasks compared with the single branch counterpart with same parameter budget. Additionally, they also show that the proposed ensemble method results in better performance than other ensemble methods (For example, ensemble over independently trained models)  not only in combined mode but also in individual branches.\n\nPaper Strengths:\n* The proposed coupled ensembles method truly show impressive results in classification benchmark (DenseNet-BC L = 118 k = 35 e = 3).\n* Detailed analysis on different ensemble fusion methods on both training time and testing time.\n* Simple but effective design to achieve a better result in testing time with same total parameter budget.\n\t\nPaper Weakness:\n* Some detail about different fusing method should be mentioned in the main paper instead of in the supplementary material.\n* In practice, how much more GPU memory is required to train the model with parallel branches (with same parameter budgets) because memory consumption is one of the main problems of networks with multiple branches.\n* At least one experiment should be carried out on a larger dataset such as ImageNet to further demonstrate the validity of the proposed method.\n* More analysis can be conducted on the training process of the model. Will it converge faster? What will be the total required training time to reach the same performance compared with single branch model with the same parameter budget?\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511627322938}, {"id": "rkbHBeSbM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper3/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a deep network architecture which processes data using multiple parallel branches and combines the posterior from these branches to compute the final scores; the network is trained in end-to-end, thus training the parallel branches jointly. Existing literature with branching architecture either employ a 2 stage training approach, training branches independently and then training the fusion network, or the branching is restricted to local regions (set of contiguous layers). In effect, this paper extends the existing literature suggesting end-to-end branching. While the technical novelty, as described in the paper, is relatively limited, the thorough experimentation together with detailed comparisons between intuitive ways to combine the output of the parallel branches is certainly valuable to the research community.\n\n+ Paper is well written and easy to follow.\n+ Proposed branching architecture clearly outperforms the baseline network (same number of parameters with a single branch) and thus offer yet another interesting choice while creating the network architecture for a problem\n+ Detailed experiments to study and analyze the effect of various parameters including the number of branches as well as various architectures to combine the output of the parallel branches.\n+ [Ease of implementation] Suggested architecture can be easily implemented using existing deep learning frameworks.\n\n- Although joint end-to-end training of branches certainly brings value compared to independent training, but the increased resource requirements may limits the applicability to large benchmarks such as ImageNet. While authors suggests a way to circumvent such limitations by training branches on separate GPUs but this would still impose limits on the number of branches as well as its ease of implementation.\n- Adding an overview figure of the architecture in the main paper (instead of supplementary) would be helpful.\n- Branched architecture serve as a regularization by distributing the gradients across different branches; however this also suggests that early layers on the network across branches would be independent. It would helpful if authors would consider an alternate archiecture where early layers may be shared across branches, suggesting a delayed branching, with fusion at the final layer.\n- One of the benefits of architectures such as DenseNet is their usefulness as a feature extractor (output of lower layers) which generalizes even to domain other that the dataset; the branched architecture could potentially diminish this benefit.\n\nMinor edits: Page 1. 'significantly match and improve' => 'either match or improve'\n\nAdditional notes:\n- It would interesting to compare this approach with a conditional training pipeline that sequentially adds branches, keeping the previous branches fixed. This may offer as a trade-off between benefits of joint training of branches vs being able to train deep models with several branches.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Branched architecture with early split and late fusion have benefits over a single branch architecture with same number of parameters.", "rating": "6: Marginally above acceptance threshold", "review": "This paper presents a deep network architecture which processes data using multiple parallel branches and combines the posterior from these branches to compute the final scores; the network is trained in end-to-end, thus training the parallel branches jointly. Existing literature with branching architecture either employ a 2 stage training approach, training branches independently and then training the fusion network, or the branching is restricted to local regions (set of contiguous layers). In effect, this paper extends the existing literature suggesting end-to-end branching. While the technical novelty, as described in the paper, is relatively limited, the thorough experimentation together with detailed comparisons between intuitive ways to combine the output of the parallel branches is certainly valuable to the research community.\n\n+ Paper is well written and easy to follow.\n+ Proposed branching architecture clearly outperforms the baseline network (same number of parameters with a single branch) and thus offer yet another interesting choice while creating the network architecture for a problem\n+ Detailed experiments to study and analyze the effect of various parameters including the number of branches as well as various architectures to combine the output of the parallel branches.\n+ [Ease of implementation] Suggested architecture can be easily implemented using existing deep learning frameworks.\n\n- Although joint end-to-end training of branches certainly brings value compared to independent training, but the increased resource requirements may limits the applicability to large benchmarks such as ImageNet. While authors suggests a way to circumvent such limitations by training branches on separate GPUs but this would still impose limits on the number of branches as well as its ease of implementation.\n- Adding an overview figure of the architecture in the main paper (instead of supplementary) would be helpful.\n- Branched architecture serve as a regularization by distributing the gradients across different branches; however this also suggests that early layers on the network across branches would be independent. It would helpful if authors would consider an alternate archiecture where early layers may be shared across branches, suggesting a delayed branching, with fusion at the final layer.\n- One of the benefits of architectures such as DenseNet is their usefulness as a feature extractor (output of lower layers) which generalizes even to domain other that the dataset; the branched architecture could potentially diminish this benefit.\n\nMinor edits: Page 1. 'significantly match and improve' => 'either match or improve'\n\nAdditional notes:\n- It would interesting to compare this approach with a conditional training pipeline that sequentially adds branches, keeping the previous branches fixed. This may offer as a trade-off between benefits of joint training of branches vs being able to train deep models with several branches.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1512535353129}], "openreview_url": "https://openreview.net/forum?id=Hk2MHt-3-", "arxiv_id": "1709.06053", "paper_pdf": "papers/Hk2MHt-3-.pdf", "paper_pdf_sha256": "e231be8aef025a2c6b1fdec34ece99bb2749ce5048791eb7174ed2e86fd6fe15", "paper_pdf_bytes": 987876, "paper_pdf_source": "openreview", "code_url": "https://github.com/vabh/coupled_ensembles", "code_repository": "vabh/coupled_ensembles", "code_commit": "c4ced6a13189e6f8b5420917509c3873e5ed6fb1", "code_archive": "repos/Hk2MHt-3-.zip", "code_archive_sha256": "2853f7669f0426629bf49825eec6639330c4c1743066495371c117bf7522cf00", "code_archive_bytes": 27692, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 35, "github_languages": {"Python": 46710, "Shell": 761}, "github_archived": false, "github_pushed_at": "2019-02-25T12:59:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/coupled-ensembles-of-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZkiVWvWwic", "year": 2026, "status": "rejected", "title": "Dynadiff: Single-stage Decoding of Images from Continuously Evolving fMRI", "authors": ["Marlene Careil", "Yohann Benchetrit", "Jean-Remi King"], "authorids": ["~Marlene_Careil1", "~Yohann_Benchetrit1", "~Jean-Remi_King1"], "authors_source": "OpenReview API", "abstract": "Brain-to-image decoding has been recently propelled by the progress in generative AI models and the availability of large ultra-high field functional Magnetic Resonance Imaging (fMRI). However, current approaches depend on complicated multi-stage pipelines and preprocessing steps that typically collapse the temporal dimension of brain recordings, thereby limiting time-resolved brain decoders. Here, we introduce Dynadiff (Dynamic Neural Activity Diffusion for Image Reconstruction), a new single-stage diffusion model designed for reconstructing images from dynamically evolving fMRI recordings. Our approach offers three main contributions. First, Dynadiff simplifies training as compared to existing approaches. Second, our model outperforms state-of-the-art models on time-resolved fMRI signals, especially on high-level semantic image reconstruction metrics, while remaining competitive on preprocessed fMRI data that collapse time. Third, this approach allows a precise characterization of the evolution of image representations in brain activity.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "mS43aTuGtM", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24883/Reviewer_LP21"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces Dynadiff, a novel brain-to-image decoding model designed to reconstruct visual stimuli from fMRI data. The authors identify two primary limitations in current approaches: (1) the reliance on complex, multi-stage training pipelines , and (2) the common practice of \"time-collapsing\" fMRI data into static 'beta values', which discards temporal information。 Dynadiff addresses both issues by proposing a single-stage diffusion model that works directly on continuously evolving fMRI time-series.  The authors demonstrate experimentally on the Natural Scenes Dataset (NSD) that Dynadiff outperforms state-of-the-art models (like MindEye2 and WAVE) on the task of time-resolved, single-trial fMRI decoding.", "review_text": "This paper introduces Dynadiff, a novel brain-to-image decoding model designed to reconstruct visual stimuli from fMRI data. The authors identify two primary limitations in current approaches: (1) the reliance on complex, multi-stage training pipelines , and (2) the common practice of \"time-collapsing\" fMRI data into static 'beta values', which discards temporal information。 Dynadiff addresses both issues by proposing a single-stage diffusion model that works directly on continuously evolving fMRI time-series.  The authors demonstrate experimentally on the Natural Scenes Dataset (NSD) that Dynadiff outperforms state-of-the-art models (like MindEye2 and WAVE) on the task of time-resolved, single-trial fMRI decoding.", "strengths": "1. The proposed method is simple and straightforward. Experimental results shows that compared to state of the art model the proposed method achieved comparable performance using only one stage training. \n2. The experiment analysis is completed and interesting, which could provide useful insight to the community. \n3. The experimental evaluation is thorough and robust.", "weaknesses": "1. Architectural novelty. While the pipeline is novel in its simplicity, the components are standard. The brain module is essentially a large MLP , and the finetuning method is LoRA. This is not a major flaw, as the contribution lies in the effective composition and problem formulation, but the architectural novelty itself is moderate.\n2. Fairness of Time-Series Baselines. Models like MindEye1 and MindEye2 were explicitly designed for static beta values. The authors state they adapted these models by \"flatten this window into a vector of size $C \\times T$\"22. This is a very naive way to incorporate temporal data and is likely a highly suboptimal adaptation that puts the baselines at a significant disadvantage. A simple flattening operation discards all explicit temporal structure. A fairer comparison might involve adapting the baselines with a more sophisticated temporal adapting method before feeding the data to their respective mapping networks. \n\n* In the line 83 I believe the Seeing beyond the brain paper does not use contrastive learning.", "questions": "1. The brain module is very large (~400M parameters). Your ablation in Table 2 justifies the types of layers used, but not their size. Did you experiment with smaller brain modules? How much does performance degrade if the brain module is, for example, 100M or 50M parameters? Is this large size truly necessary?\n\n2. Could you please elaborate on the adaptation of the MindEye baselines? Do you agree that simple flattening  is a weak baseline for temporal data? How confident are you that Dynadiff's superior performance in Table 1 is due to its architectural design rather than this suboptimal baseline adaptation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Dynadiff, a novel brain-to-image decoding model designed to reconstruct visual stimuli from fMRI data. The authors identify two primary limitations in current approaches: (1) the reliance on complex, multi-stage training pipelines , and (2) the common practice of \"time-collapsing\" fMRI data into static 'beta values', which discards temporal information。 Dynadiff addresses both issues by proposing a single-stage diffusion model that works directly on continuously evolving fMRI time-series.  The authors demonstrate experimentally on the Natural Scenes Dataset (NSD) that Dynadiff outperforms state-of-the-art models (like MindEye2 and WAVE) on the task of time-resolved, single-trial fMRI decoding.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The proposed method is simple and straightforward. Experimental results shows that compared to state of the art model the proposed method achieved comparable performance using only one stage training. \n2. The experiment analysis is completed and interesting, which could provide useful insight to the community. \n3. The experimental evaluation is thorough and robust.", "weaknesses": "1. Architectural novelty. While the pipeline is novel in its simplicity, the components are standard. The brain module is essentially a large MLP , and the finetuning method is LoRA. This is not a major flaw, as the contribution lies in the effective composition and problem formulation, but the architectural novelty itself is moderate.\n2. Fairness of Time-Series Baselines. Models like MindEye1 and MindEye2 were explicitly designed for static beta values. The authors state they adapted these models by \"flatten this window into a vector of size $C \\times T$\"22. This is a very naive way to incorporate temporal data and is likely a highly suboptimal adaptation that puts the baselines at a significant disadvantage. A simple flattening operation discards all explicit temporal structure. A fairer comparison might involve adapting the baselines with a more sophisticated temporal adapting method before feeding the data to their respective mapping networks. \n\n* In the line 83 I believe the Seeing beyond the brain paper does not use contrastive learning.", "questions": "1. The brain module is very large (~400M parameters). Your ablation in Table 2 justifies the types of layers used, but not their size. Did you experiment with smaller brain modules? How much does performance degrade if the brain module is, for example, 100M or 50M parameters? Is this large size truly necessary?\n\n2. Could you please elaborate on the adaptation of the MindEye baselines? Do you agree that simple flattening  is a weak baseline for temporal data? How confident are you that Dynadiff's superior performance in Table 1 is due to its architectural design rather than this suboptimal baseline adaptation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762051917155}, {"id": "4ht6amEFSb", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24883/Reviewer_hpZN"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes **Dynadiff**, a “single-stage” fMRI-to-image system that conditions a pretrained diffusion model directly on short fMRI time windows, adding LoRA adapters to cross-attention. The authors argue (i) prior work “collapses time” by using GLM betas, (ii) their time-resolved approach yields better temporal fidelity, and (iii) a one-stage objective simplifies training. Experiments are on NSD with multiple metrics and temporally shifted windows.", "review_text": "The paper proposes **Dynadiff**, a “single-stage” fMRI-to-image system that conditions a pretrained diffusion model directly on short fMRI time windows, adding LoRA adapters to cross-attention. The authors argue (i) prior work “collapses time” by using GLM betas, (ii) their time-resolved approach yields better temporal fidelity, and (iii) a one-stage objective simplifies training. Experiments are on NSD with multiple metrics and temporally shifted windows.", "strengths": "# Strengths\n\n* **Interesting temporal visualization**: The time-shift analyses (e.g., Fig. 4) are engaging and make the dynamics tangible.\n* **Simplified training story**: Collapsing multi-stage pipelines into a single training objective is an appealing engineering direction.\n* **Reasonably broad metrics**: CLIP/feature metrics, segmentation mIoU, and qualitative examples provide multiple views of performance.", "weaknesses": "# Weaknesses & Detailed Comments\n\n## A. Methodology/Claims\n\n1. **“Time-series modeling” is shallow relative to the claim.**\n   The core temporal handling appears to be per-timestep linear transforms plus a **single temporal aggregation layer**. This is not a genuine temporal model (no temporal attention, SSM/RNN, or FIR deconvolution) and does not convincingly support the claim that prior work “completely discards the time dimension.” Please either temper the claim or compare against lightweight temporal baselines (1D convs, attention, S4/SSM) to show real sequence modeling helps.\n\n2. **Over-sized mapper vs. data; missing capacity control.**\n   The brain-to-token mapper is very large given per-subject data. There’s little regularization beyond dropout, and no capacity ablations (e.g., lowering hidden dims, low-rank constraints, weight sharing across timesteps). This raises overfitting and stability concerns and limits interpretability.\n\n3. **Preprocessing rationale is thin.**\n   Avoiding nuisance regression/high-pass to “avoid assumptions” is not sufficient; it may let the model pick up motion/physio artifacts. The later “preprocessing ablation” changes multiple factors at once (pipeline and surface space), making attribution ambiguous.\n\n## B. Experimental Design & Evaluation\n\n1. **Potential leakage from hemodynamic overlap in the *main table*.** (important)\n   NSD uses rapid event-related timing (SOA ~4 s). With multi-second windows centered near the HRF peak, test windows inevitably contain BOLD from neighboring stimuli—some likely from the training split if images are interleaved. Your time-resolved figures correctly use a *run-wise* split, but the headline table seems to rely on the standard interleaved split, risking contamination that favors time-series methods. Maybe consider recompute the main table under a run-wise (or otherwise no-overlap) split, or decode from windows that minimize neighbor contamination; add FIR/GLM-deconvolved controls.\n\n2. **Baselines are not fully matched to your setting.**\n   WAVE appears evaluated on a different start/duration window and possibly different ROIs; MindEye baselines flattened time without temporal modeling. These choices may systematically depress them.\n\n3. **Time-series vs. betas: claim not supported.** (important)\n   When you switch from betas to raw time-series, high-level metrics generally **drop**, and Dynadiff does not surpass MindEye2 on betas. The supposed “time-collapsing problem” is asserted but not shown to harm decoding.\nMaybe you can demonstrate a concrete failure mode on betas (e.g., a class of stimuli where betas underperform), or show that adding temporal modeling on *deconvolved* timecourses surpasses betas under matched evaluation.\n\n\n## C. Neuroscience Interpretation\n\n1. **“Dynamic coding” interpretation is confounded by the HRF.**\n    The fact that windows shifted to “previous/next image” decode those images can be explained by HRF overlap in fast event-related designs; it does not by itself evidence a novel neural dynamic. maybe you could consider (i) cross-time generalization matrices (train at δ, test across δ′) with **deconvolved** signals; (ii) a no-overlap control (preceding/following stimuli from held-out runs or categories); (iii) show pre-onset decoding at chance.\n\n## D. Relation to Prior Work\n\n **Positioning vs. recent temporal/video decoders is incomplete.**\n    The claim that prior SOTA “completely discards time” is too strong. Recent works also model temporal structure across subjects/videos, please consider comparing with them", "questions": "# Questions for Authors\n\n* How sensitive is performance to the mapper’s capacity and to the window start/duration?\n* Did you try temporal attention or FIR deconvolution before aggregation?\n\n# Suggestions\n* Leakage control: Recompute headline results under no-overlap splits and/or deconvolved signals.\n* Matched baselines: Align windows/ROIs; add a time-aware MindEye baseline.\n* Temper or justify claims: Maybe rephrase the “completely discards time” statement and substantiate the “one-stage” benefits.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes **Dynadiff**, a “single-stage” fMRI-to-image system that conditions a pretrained diffusion model directly on short fMRI time windows, adding LoRA adapters to cross-attention. The authors argue (i) prior work “collapses time” by using GLM betas, (ii) their time-resolved approach yields better temporal fidelity, and (iii) a one-stage objective simplifies training. Experiments are on NSD with multiple metrics and temporally shifted windows.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "# Strengths\n\n* **Interesting temporal visualization**: The time-shift analyses (e.g., Fig. 4) are engaging and make the dynamics tangible.\n* **Simplified training story**: Collapsing multi-stage pipelines into a single training objective is an appealing engineering direction.\n* **Reasonably broad metrics**: CLIP/feature metrics, segmentation mIoU, and qualitative examples provide multiple views of performance.", "weaknesses": "# Weaknesses & Detailed Comments\n\n## A. Methodology/Claims\n\n1. **“Time-series modeling” is shallow relative to the claim.**\n   The core temporal handling appears to be per-timestep linear transforms plus a **single temporal aggregation layer**. This is not a genuine temporal model (no temporal attention, SSM/RNN, or FIR deconvolution) and does not convincingly support the claim that prior work “completely discards the time dimension.” Please either temper the claim or compare against lightweight temporal baselines (1D convs, attention, S4/SSM) to show real sequence modeling helps.\n\n2. **Over-sized mapper vs. data; missing capacity control.**\n   The brain-to-token mapper is very large given per-subject data. There’s little regularization beyond dropout, and no capacity ablations (e.g., lowering hidden dims, low-rank constraints, weight sharing across timesteps). This raises overfitting and stability concerns and limits interpretability.\n\n3. **Preprocessing rationale is thin.**\n   Avoiding nuisance regression/high-pass to “avoid assumptions” is not sufficient; it may let the model pick up motion/physio artifacts. The later “preprocessing ablation” changes multiple factors at once (pipeline and surface space), making attribution ambiguous.\n\n## B. Experimental Design & Evaluation\n\n1. **Potential leakage from hemodynamic overlap in the *main table*.** (important)\n   NSD uses rapid event-related timing (SOA ~4 s). With multi-second windows centered near the HRF peak, test windows inevitably contain BOLD from neighboring stimuli—some likely from the training split if images are interleaved. Your time-resolved figures correctly use a *run-wise* split, but the headline table seems to rely on the standard interleaved split, risking contamination that favors time-series methods. Maybe consider recompute the main table under a run-wise (or otherwise no-overlap) split, or decode from windows that minimize neighbor contamination; add FIR/GLM-deconvolved controls.\n\n2. **Baselines are not fully matched to your setting.**\n   WAVE appears evaluated on a different start/duration window and possibly different ROIs; MindEye baselines flattened time without temporal modeling. These choices may systematically depress them.\n\n3. **Time-series vs. betas: claim not supported.** (important)\n   When you switch from betas to raw time-series, high-level metrics generally **drop**, and Dynadiff does not surpass MindEye2 on betas. The supposed “time-collapsing problem” is asserted but not shown to harm decoding.\nMaybe you can demonstrate a concrete failure mode on betas (e.g., a class of stimuli where betas underperform), or show that adding temporal modeling on *deconvolved* timecourses surpasses betas under matched evaluation.\n\n\n## C. Neuroscience Interpretation\n\n1. **“Dynamic coding” interpretation is confounded by the HRF.**\n    The fact that windows shifted to “previous/next image” decode those images can be explained by HRF overlap in fast event-related designs; it does not by itself evidence a novel neural dynamic. maybe you could consider (i) cross-time generalization matrices (train at δ, test across δ′) with **deconvolved** signals; (ii) a no-overlap control (preceding/following stimuli from held-out runs or categories); (iii) show pre-onset decoding at chance.\n\n## D. Relation to Prior Work\n\n **Positioning vs. recent temporal/video decoders is incomplete.**\n    The claim that prior SOTA “completely discards time” is too strong. Recent works also model temporal structure across subjects/videos, please consider comparing with them", "questions": "# Questions for Authors\n\n* How sensitive is performance to the mapper’s capacity and to the window start/duration?\n* Did you try temporal attention or FIR deconvolution before aggregation?\n\n# Suggestions\n* Leakage control: Recompute headline results under no-overlap splits and/or deconvolved signals.\n* Matched baselines: Align windows/ROIs; add a time-aware MindEye baseline.\n* Temper or justify claims: Maybe rephrase the “completely discards time” statement and substantiate the “one-stage” benefits.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762031276202}, {"id": "koVfx1aSx8", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24883/Reviewer_v9Ga"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces Dynadiff, a single-stage diffusion model for reconstructing images directly from time-resolved fMRI signals. Unlike previous multi-stage approaches, Dynadiff jointly trains a lightweight brain-to-diffusion conditioning module with LoRA-adapted diffusion layers. The method avoids time-collapsing preprocessing, supports temporal decoding, and achieves state-of-the-art results on the NSD dataset for both semantic and perceptual metrics. Ablations show the importance of temporal modeling and LoRA-based fine-tuning.", "review_text": "The paper introduces Dynadiff, a single-stage diffusion model for reconstructing images directly from time-resolved fMRI signals. Unlike previous multi-stage approaches, Dynadiff jointly trains a lightweight brain-to-diffusion conditioning module with LoRA-adapted diffusion layers. The method avoids time-collapsing preprocessing, supports temporal decoding, and achieves state-of-the-art results on the NSD dataset for both semantic and perceptual metrics. Ablations show the importance of temporal modeling and LoRA-based fine-tuning.", "strengths": "* The single-stage training pipeline is a clear improvement over existing multi-stage frameworks.\n* Demonstrates robust time-resolved reconstruction from continuous BOLD signals.\n* Includes ablations on time-window duration, brain module design, and diffusion tuning strategies.\n* Well-written and clearly motivated, especially regarding the challenges of time-collapsed preprocessing.", "weaknesses": "* While the single-stage design is elegant, the core idea, jointly fine-tuning an fMRI encoder with a diffusion model, remains conceptually close to prior fMRI-to-image diffusion frameworks.\n* Table 1 employs a customized fMRI preprocessing pipeline while comparing against baselines trained on time-collapsed data, making the reported performance gains difficult to interpret. Moreover, the comparison omits recent time-resolved decoders such as Neuropictor (Huo et al., 2024), which weakens the fairness of the evaluation.\n* In Table 4, where multiple strong baselines are evaluated under the same setting, Dynadiff shows no clear or consistent advantage, casting doubt on its claimed state-of-the-art performance.\n* The evaluation is limited to four NSD subjects, without testing cross-subject or cross-dataset generalization, leaving its robustness and scalability uncertain.", "questions": "* How fair is the comparison in Table 1, given that other baselines were not trained under the same preprocessing pipeline?\n* How sensitive are the results to preprocessing choices such as ROI selection, and normalization?\n* Why are time-resolved baselines (e.g., Neuropictor) excluded from the main comparison?\n* Can Dynadiff generalize to new participants or datasets without subject-specific retraining?\n* What factors explain the lack of improvement in Table 4 despite the proposed architectural changes?\n* How might this framework be extended or validated for continuous video decoding rather than static images?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Dynadiff, a single-stage diffusion model for reconstructing images directly from time-resolved fMRI signals. Unlike previous multi-stage approaches, Dynadiff jointly trains a lightweight brain-to-diffusion conditioning module with LoRA-adapted diffusion layers. The method avoids time-collapsing preprocessing, supports temporal decoding, and achieves state-of-the-art results on the NSD dataset for both semantic and perceptual metrics. Ablations show the importance of temporal modeling and LoRA-based fine-tuning.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* The single-stage training pipeline is a clear improvement over existing multi-stage frameworks.\n* Demonstrates robust time-resolved reconstruction from continuous BOLD signals.\n* Includes ablations on time-window duration, brain module design, and diffusion tuning strategies.\n* Well-written and clearly motivated, especially regarding the challenges of time-collapsed preprocessing.", "weaknesses": "* While the single-stage design is elegant, the core idea, jointly fine-tuning an fMRI encoder with a diffusion model, remains conceptually close to prior fMRI-to-image diffusion frameworks.\n* Table 1 employs a customized fMRI preprocessing pipeline while comparing against baselines trained on time-collapsed data, making the reported performance gains difficult to interpret. Moreover, the comparison omits recent time-resolved decoders such as Neuropictor (Huo et al., 2024), which weakens the fairness of the evaluation.\n* In Table 4, where multiple strong baselines are evaluated under the same setting, Dynadiff shows no clear or consistent advantage, casting doubt on its claimed state-of-the-art performance.\n* The evaluation is limited to four NSD subjects, without testing cross-subject or cross-dataset generalization, leaving its robustness and scalability uncertain.", "questions": "* How fair is the comparison in Table 1, given that other baselines were not trained under the same preprocessing pipeline?\n* How sensitive are the results to preprocessing choices such as ROI selection, and normalization?\n* Why are time-resolved baselines (e.g., Neuropictor) excluded from the main comparison?\n* Can Dynadiff generalize to new participants or datasets without subject-specific retraining?\n* What factors explain the lack of improvement in Table 4 despite the proposed architectural changes?\n* How might this framework be extended or validated for continuous video decoding rather than static images?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761991449715}, {"id": "Gds67Jn4kr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24883/Reviewer_jcgu"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper introduces Dynadiff, a single‑stage diffusion‑based decoder that conditions a frozen latent diffusion model on continuous BOLD fMRI time series via a learned “brain module,” thereby avoiding beta averaging and multi‑stage pipelines common in prior work. Experiments are conducted on NSD without averaging repetitions at train/test time. The method improves especially on high‑level/semantic metrics for single‑trial time series, and analyzes temporal generalization by shifting decoding windows to probe when reconstructions best align with the viewed image. The paper frames these results as a step toward real‑time decoding, though all data are static‑image sequences from NSD rather than natural movies.", "review_text": "This paper introduces Dynadiff, a single‑stage diffusion‑based decoder that conditions a frozen latent diffusion model on continuous BOLD fMRI time series via a learned “brain module,” thereby avoiding beta averaging and multi‑stage pipelines common in prior work. Experiments are conducted on NSD without averaging repetitions at train/test time. The method improves especially on high‑level/semantic metrics for single‑trial time series, and analyzes temporal generalization by shifting decoding windows to probe when reconstructions best align with the viewed image. The paper frames these results as a step toward real‑time decoding, though all data are static‑image sequences from NSD rather than natural movies.", "strengths": "- Dynadiff jointly trains only two components (i) a brain module (MLP) that maps brain activity to the diffusion model’s conditioning tokens and (ii) LoRA adapters on cross‑attention while freezing the rest of the generator. No pretrained fMRI encoder, no alignment stage to a fixed embedding space, and no post‑generation selection/refinement are required. This reduces engineering complexity. \n\n- Training and evaluation use single‑trial BOLD windows from the standard‑resolution NSD time series; repetitions are not averaged, and preprocessing deliberately keeps temporal information. This directly addresses “time‑collapse” concerns in prior NSD decoders.\n\n- Competitive to SOTA on single‑trial time series, especially semantic metrics. \n\n- The temporal generalization was evaluated by shifting the input window during testing. The results show that a model trained at a fixed relative time generalizes best near that time, whereas models specialized for each δ-offset can still decode the current image at later time points.These findings are consistent with the latency characteristics of the hemodynamic response function.\n\n- Ablations on window length, brain‑module design, and which generator layers to adapt support design choices. Hyperparameters and training compute are specified.", "weaknesses": "- Gains vs. MindEye2 are uneven and sometimes within SEM; low‑level metrics drop. In Tab. 1, SSIM and PixCorr are below MindEye2. Improvements on AlexNet(2/5), Inception, SwAV, mIoU are reported, but several margins are modest and some overlap with SEM. It is thus difficult to determine whether these differences are genuinely meaningful. In addition, when all methods are evaluated on averaged beta values (the dominant NSD setting), Dynadiff is competitive but trails on several metrics.\n\n- The caption of Fig. 4 says “Real‑time decoding,” but all experiments use static‑image sequences from NSD; the time‑resolved results come from shifted windows. There is no dynamic stimuli such as natural‑movie dataset evaluation. Also, rationale for not evaluating on videos is unconvincing. The Discussion argues that prior video‑decoders often rely on time‑collapsed betas, multi‑stage training, and smaller data; however, similar caveats also apply to NSD baselines and are not specific reasons to avoid at least a small‑scale video test. Given the paper’s emphasis on time‑resolved decoding, a movie benchmark (and comparing their method with the methods for previous movie reconstruction methods) seems warranted. \n\n- The brain module has ~400M parameters; inference uses DDIM 20 steps. While training details are thorough, the paper does not report per‑trial inference time or throughput relative to NSD’s TR=1.3 s, leaving the “real‑time” feasibility unclear.", "questions": "- Clarify “real‑time.” What is the wall‑clock latency from receiving a new TR to outputting an image with 20 DDIM steps? Can the model decode per‑TR at TR=1.3 s on a single GPU?\n- Metric arrows for DreamSim (↑/↓) are not fully consistent across tables (see Tables 1 and 5). \n- Table 8 appears to have a column misalignment for Subject 5 (mIoU missing).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Dynadiff, a single‑stage diffusion‑based decoder that conditions a frozen latent diffusion model on continuous BOLD fMRI time series via a learned “brain module,” thereby avoiding beta averaging and multi‑stage pipelines common in prior work. Experiments are conducted on NSD without averaging repetitions at train/test time. The method improves especially on high‑level/semantic metrics for single‑trial time series, and analyzes temporal generalization by shifting decoding windows to probe when reconstructions best align with the viewed image. The paper frames these results as a step toward real‑time decoding, though all data are static‑image sequences from NSD rather than natural movies.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Dynadiff jointly trains only two components (i) a brain module (MLP) that maps brain activity to the diffusion model’s conditioning tokens and (ii) LoRA adapters on cross‑attention while freezing the rest of the generator. No pretrained fMRI encoder, no alignment stage to a fixed embedding space, and no post‑generation selection/refinement are required. This reduces engineering complexity. \n\n- Training and evaluation use single‑trial BOLD windows from the standard‑resolution NSD time series; repetitions are not averaged, and preprocessing deliberately keeps temporal information. This directly addresses “time‑collapse” concerns in prior NSD decoders.\n\n- Competitive to SOTA on single‑trial time series, especially semantic metrics. \n\n- The temporal generalization was evaluated by shifting the input window during testing. The results show that a model trained at a fixed relative time generalizes best near that time, whereas models specialized for each δ-offset can still decode the current image at later time points.These findings are consistent with the latency characteristics of the hemodynamic response function.\n\n- Ablations on window length, brain‑module design, and which generator layers to adapt support design choices. Hyperparameters and training compute are specified.", "weaknesses": "- Gains vs. MindEye2 are uneven and sometimes within SEM; low‑level metrics drop. In Tab. 1, SSIM and PixCorr are below MindEye2. Improvements on AlexNet(2/5), Inception, SwAV, mIoU are reported, but several margins are modest and some overlap with SEM. It is thus difficult to determine whether these differences are genuinely meaningful. In addition, when all methods are evaluated on averaged beta values (the dominant NSD setting), Dynadiff is competitive but trails on several metrics.\n\n- The caption of Fig. 4 says “Real‑time decoding,” but all experiments use static‑image sequences from NSD; the time‑resolved results come from shifted windows. There is no dynamic stimuli such as natural‑movie dataset evaluation. Also, rationale for not evaluating on videos is unconvincing. The Discussion argues that prior video‑decoders often rely on time‑collapsed betas, multi‑stage training, and smaller data; however, similar caveats also apply to NSD baselines and are not specific reasons to avoid at least a small‑scale video test. Given the paper’s emphasis on time‑resolved decoding, a movie benchmark (and comparing their method with the methods for previous movie reconstruction methods) seems warranted. \n\n- The brain module has ~400M parameters; inference uses DDIM 20 steps. While training details are thorough, the paper does not report per‑trial inference time or throughput relative to NSD’s TR=1.3 s, leaving the “real‑time” feasibility unclear.", "questions": "- Clarify “real‑time.” What is the wall‑clock latency from receiving a new TR to outputting an image with 20 DDIM steps? Can the model decode per‑TR at TR=1.3 s on a single GPU?\n- Metric arrows for DreamSim (↑/↓) are not fully consistent across tables (see Tables 1 and 5). \n- Table 8 appears to have a column misalignment for Subject 5 (mIoU missing).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761825412820}], "openreview_url": "https://openreview.net/forum?id=ZkiVWvWwic", "arxiv_id": "2505.14556", "paper_pdf": "papers/ZkiVWvWwic.pdf", "paper_pdf_sha256": "b062985363f375d3641bfe3d25cfeb9abbc93dfa97de654d5c6509b1a4f02cf9", "paper_pdf_bytes": 16270608, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/dynadiff", "code_repository": "facebookresearch/dynadiff", "code_commit": "b8d1f84a054826c2cea28a44beb2aac5a1839b8c", "code_archive": "repos/ZkiVWvWwic.zip", "code_archive_sha256": "7e328e7c2a86be382d2cdb3c929d8cbc160377d4e65a1c519c930abde59973d4", "code_archive_bytes": 71422, "code_file_count": 16, "code_extensions": {".py": 13, ".sh": 3}, "github_disk_usage_kb": 65, "github_languages": {"Python": 95810, "Shell": 2753}, "github_archived": false, "github_pushed_at": "2025-05-13T21:27:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dynadiff-single-stage-decoding-of-images-from"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "N5ID99rsUq", "year": 2024, "status": "rejected", "title": "Stability and Generalization in Free Adversarial Training", "authors": ["Xiwei Cheng", "Kexin Fu", "Farzan Farnia"], "authorids": ["~Xiwei_Cheng2", "~Kexin_Fu2", "~Farzan_Farnia1"], "authors_source": "OpenReview API", "abstract": "While adversarial training methods have resulted in significant improvements in the deep neural nets' robustness against norm-bounded adversarial perturbations, their generalization performance from training samples to test data has been shown to be considerably worse than standard empirical risk minimization methods. Several recent studies seek to connect the generalization behavior of adversarially trained classifiers to various gradient-based min-max optimization algorithms used for their training. In this work, we study the generalization performance of adversarial training methods using the algorithmic stability framework. Specifically, our goal is to compare the generalization performance of vanilla adversarial training scheme fully optimizing the perturbations at every iteration vs. the free adversarial training simultaneously optimizing the norm-bounded perturbations and classifier parameters. Our proven generalization bounds indicate that the free adversarial training method could enjoy a lower generalization gap between training and test samples due to the simultaneous nature of its min-max optimization algorithm. We perform several numerical experiments to evaluate the generalization performance of vanilla, fast, and free adversarial training methods. Our empirical findings also show the improved generalization performance of the free adversarial training method and further demonstrate that the better generalization result could translate to greater robustness against black-box attack schemes and higher transferability of the adversarial examples designed for free adversarially trained neural networks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "fI0C2keEQX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4972/Reviewer_iGFe"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies stability and generalization of vanilla, free, and fast adversarial training from an algorithmic stability perspective. The generalization error gap bounds are derived for those adversarial training methods, and numerical results are also provided to show the generalization performance and robustness against black-box attacks.", "review_text": "This paper studies stability and generalization of vanilla, free, and fast adversarial training from an algorithmic stability perspective. The generalization error gap bounds are derived for those adversarial training methods, and numerical results are also provided to show the generalization performance and robustness against black-box attacks.", "strengths": "This paper is well-motivated and well-written. The novelty and contributions are clearly stated and organized. The theoretical findings are provided in a rigorous manner, together with some validation numerical results. In general, the theoretical findings are interesting to the community.", "weaknesses": "1. This paper is dedicated to generalization performance analysis of existing adversarial training methods and reveals some interesting points. Nevertheless, there is a lack of deep insights on the new advanced designs of adversarial training from the generalization bounds. The authors should have discussed the insights/guidance from the theoretical findings, or discussed certain limitations of the algorithmic stability approach itself.\n2. From the experimental results, e.g., Figure 1, it appears that the reduced generalization error gap of free adversarial training is mainly due to the higher training error. Assuming the generalization error gap maintains, it is unclear if the test error can be further reduced when the training error is reduced. The authors should add some comments on this.\n3. It would expect that new training/regularization methods could be proposed given the obtained generalization error bounds. Otherwise, the impact of the theoretical findings of this work is quite limited. A thorough discussion would be helpful and beneficial. It would be also interesting to know the potential connection between generalization gap and the robustness against adversarial attacks.", "questions": "See the Weaknesses above. \n\nAdd some comments on the practical usefulness of the theoretical findings with respect to the design of adversarial training methods. The limitations of the algorithmic stability approach for studying generalization performance could be also discussed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies stability and generalization of vanilla, free, and fast adversarial training from an algorithmic stability perspective. The generalization error gap bounds are derived for those adversarial training methods, and numerical results are also provided to show the generalization performance and robustness against black-box attacks.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "This paper is well-motivated and well-written. The novelty and contributions are clearly stated and organized. The theoretical findings are provided in a rigorous manner, together with some validation numerical results. In general, the theoretical findings are interesting to the community.", "weaknesses": "1. This paper is dedicated to generalization performance analysis of existing adversarial training methods and reveals some interesting points. Nevertheless, there is a lack of deep insights on the new advanced designs of adversarial training from the generalization bounds. The authors should have discussed the insights/guidance from the theoretical findings, or discussed certain limitations of the algorithmic stability approach itself.\n2. From the experimental results, e.g., Figure 1, it appears that the reduced generalization error gap of free adversarial training is mainly due to the higher training error. Assuming the generalization error gap maintains, it is unclear if the test error can be further reduced when the training error is reduced. The authors should add some comments on this.\n3. It would expect that new training/regularization methods could be proposed given the obtained generalization error bounds. Otherwise, the impact of the theoretical findings of this work is quite limited. A thorough discussion would be helpful and beneficial. It would be also interesting to know the potential connection between generalization gap and the robustness against adversarial attacks.", "questions": "See the Weaknesses above. \n\nAdd some comments on the practical usefulness of the theoretical findings with respect to the design of adversarial training methods. The limitations of the algorithmic stability approach for studying generalization performance could be also discussed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698672262767}, {"id": "Bt3tO2KlnX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4972/Reviewer_ge9g"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work studies the role of min-max optimization algorithms in the generalization performance of adversarial training methods. It leverages the algorithmic stability framework to compare the generalization behavior of adversarial training methods. The developed generalization bounds suggest that not only can the free AT approach lead to a faster optimization compared to the vanilla AT, but also it can result in a lower generalization gap between the performance on training and test data.", "review_text": "This work studies the role of min-max optimization algorithms in the generalization performance of adversarial training methods. It leverages the algorithmic stability framework to compare the generalization behavior of adversarial training methods. The developed generalization bounds suggest that not only can the free AT approach lead to a faster optimization compared to the vanilla AT, but also it can result in a lower generalization gap between the performance on training and test data.", "strengths": "- This work provides some theoretical results.\n- The theoretical conclusions are easy to follow.", "weaknesses": "- What is the definition of $\\Delta$?\n- What is the definition of randomized algorithm $A(\\cdot)$? A mapping? If yes, then what is the definition of $\\mathbb{E}_A$?\n- Given $S$, $w=A(S)$ is a random variable or constant？\n- What's the definition of $S'$ here?\n- What is the definition of \"$A$ is $\\epsilon$-uniformly stable\"?\n- Unclear definition in Theorem 1?\n- I **guess** the theory is developed over \"randomized algorithm $A$\" (and Gibbs loss?), i.e., the output of $A(S)$ is random weights, which is distributed by a posterior. However, this paper only presents empirical results based on deterministic weights. How can these empirical findings provide support for the theoretical results?\n- If $A(S)$ is a random variable, given $S$, what are the specific posterior distributions of $A_{AVanilla}(S)$ and $A_{Free}(S)$? What is the difference between these posterior distributions? Where does the randomness of $A_{AVanilla}(S)$ come from? \n- It seems there are not some interesting insights from the theoretical and empirical results in this work. \n\n(Please correct me if I have some mistakes.)", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studies the role of min-max optimization algorithms in the generalization performance of adversarial training methods. It leverages the algorithmic stability framework to compare the generalization behavior of adversarial training methods. The developed generalization bounds suggest that not only can the free AT approach lead to a faster optimization compared to the vanilla AT, but also it can result in a lower generalization gap between the performance on training and test data.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- This work provides some theoretical results.\n- The theoretical conclusions are easy to follow.", "weaknesses": "- What is the definition of $\\Delta$?\n- What is the definition of randomized algorithm $A(\\cdot)$? A mapping? If yes, then what is the definition of $\\mathbb{E}_A$?\n- Given $S$, $w=A(S)$ is a random variable or constant？\n- What's the definition of $S'$ here?\n- What is the definition of \"$A$ is $\\epsilon$-uniformly stable\"?\n- Unclear definition in Theorem 1?\n- I **guess** the theory is developed over \"randomized algorithm $A$\" (and Gibbs loss?), i.e., the output of $A(S)$ is random weights, which is distributed by a posterior. However, this paper only presents empirical results based on deterministic weights. How can these empirical findings provide support for the theoretical results?\n- If $A(S)$ is a random variable, given $S$, what are the specific posterior distributions of $A_{AVanilla}(S)$ and $A_{Free}(S)$? What is the difference between these posterior distributions? Where does the randomness of $A_{AVanilla}(S)$ come from? \n- It seems there are not some interesting insights from the theoretical and empirical results in this work. \n\n(Please correct me if I have some mistakes.)", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698566089284}, {"id": "ElWES9T99r", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4972/Reviewer_T7Bj"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper studies the generalization of free adversarial training (AT) which was proposed in Shafahi et al. (2019). \n The authors use the algorithmic stability approach to analyze its generalization behavior and it provides its comparison of the generalization bounds against the vanilla, fast AT methods.  It claims that the free AT algorithm could have a lower generalization bound than the vanilla AT one.", "review_text": "The paper studies the generalization of free adversarial training (AT) which was proposed in Shafahi et al. (2019). \n The authors use the algorithmic stability approach to analyze its generalization behavior and it provides its comparison of the generalization bounds against the vanilla, fast AT methods.  It claims that the free AT algorithm could have a lower generalization bound than the vanilla AT one.", "strengths": "This seems to be the first-ever-known result that addressed the generalization of the free AT method using the algorithmic stability approach in the setting of mini-max formulation.", "weaknesses": "While the stability results for the free AT method are first-ever-known, the proof techniques seem to be incremental and the paper did not illustrate clearly what the main technical contribution is, particularly considering there is a considerable amount of work on stability analysis.\n \nThe generalization bound in Theorem 4 relies on the restrictive assumption that the gradient $\\nabla_\\delta h(w,\\delta; x,y)$ is lower-bounded by $1 / \\psi$ during the training process.  There is no discussion about when this critical condition holds true.", "questions": "It is not clear to me what the free AT method aims to minimize or optimize.  The objective function of the vanilla AT method is given on page 3, i.e. $R_S(w)$ or $R(w)$.  From the pseudo-code of Algorithm 3,  there are two random samplings--one for mini-batch and one for $\\{\\delta_j\\}$ and then the $w$ and $\\delta$ are updated by gradient descent and ascent, respectively.  In this sense, does the free AT methods aim to minimize the following objective \n$$ \\min_w \\max_\\delta {1\\over n } \\sum_{j=1}^n \\int_{\\delta_j\\in \\Delta} h(w,\\delta_j; x_j,y_j)$$ \nThe objective functions seem to be very different from each other for the free AT method and the vanilla AT one.   Indeed, the objective function of the free AT method is a low-bound relaxation of the vanilla one.   Could you explain more about this point?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the generalization of free adversarial training (AT) which was proposed in Shafahi et al. (2019). \n The authors use the algorithmic stability approach to analyze its generalization behavior and it provides its comparison of the generalization bounds against the vanilla, fast AT methods.  It claims that the free AT algorithm could have a lower generalization bound than the vanilla AT one.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "This seems to be the first-ever-known result that addressed the generalization of the free AT method using the algorithmic stability approach in the setting of mini-max formulation.", "weaknesses": "While the stability results for the free AT method are first-ever-known, the proof techniques seem to be incremental and the paper did not illustrate clearly what the main technical contribution is, particularly considering there is a considerable amount of work on stability analysis.\n \nThe generalization bound in Theorem 4 relies on the restrictive assumption that the gradient $\\nabla_\\delta h(w,\\delta; x,y)$ is lower-bounded by $1 / \\psi$ during the training process.  There is no discussion about when this critical condition holds true.", "questions": "It is not clear to me what the free AT method aims to minimize or optimize.  The objective function of the vanilla AT method is given on page 3, i.e. $R_S(w)$ or $R(w)$.  From the pseudo-code of Algorithm 3,  there are two random samplings--one for mini-batch and one for $\\{\\delta_j\\}$ and then the $w$ and $\\delta$ are updated by gradient descent and ascent, respectively.  In this sense, does the free AT methods aim to minimize the following objective \n$$ \\min_w \\max_\\delta {1\\over n } \\sum_{j=1}^n \\int_{\\delta_j\\in \\Delta} h(w,\\delta_j; x_j,y_j)$$ \nThe objective functions seem to be very different from each other for the free AT method and the vanilla AT one.   Indeed, the objective function of the free AT method is a low-bound relaxation of the vanilla one.   Could you explain more about this point?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697813704821}, {"id": "wKi1YpMCts", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4972/Reviewer_wev8"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work utilizes the stability generalization framework to quantify the generalization bound of the Free Adversarial Training algorithm. Additionally, it shows that Free AT has a smaller generalization gap (but test error may not be smaller) and provides some theoretical intuitions.", "review_text": "This work utilizes the stability generalization framework to quantify the generalization bound of the Free Adversarial Training algorithm. Additionally, it shows that Free AT has a smaller generalization gap (but test error may not be smaller) and provides some theoretical intuitions.", "strengths": "This work offers important insights into the convergence properties of free AT. Additionally, the authors highlight intuitive relationships between \"more simultaneous\" gradient updates during training and the resulting generalization capability. These results can inspire further improvements in robust training algorithms to alleviate overfitting. The paper is generally well-presented and easy to follow.", "weaknesses": "- The theoretical support for FreeAT having a smaller generalization gap than VanillaAT could be more rigorous. Specifically, while Theorem 2 presents pessimistic results for the convergence of VanillaAT, it is unclear whether this bound is tight. It is unclear whether the convergence difference between vanilla and free AT is due to the algorithm itself or some artifacts of the proof technique. While experiment results support this intuition, the paper would benefit from some additional explanations. It would be even better if some lower bounds could be provided for $\\mathcal{E}\\_{\\textrm{gen}} (A\\_{\\textrm{Vanilla}})$.\n- The relationship between a smaller generalization gap and better transferability is unclear. The motivation for the experiment setting of transferring attacks from a robust model to a standard model is weak. I suggest moving the transferability analysis to the Appendix (it's still good to have them) and making space for Table 2, which supports your main claim.\n- Figure 1 should use different line styles for train and test. In the current form, it's hard to distinguish them.\n- Since the proposed convergence bounds depend on the dataset size $n$, this paper would benefit from some empirical comparisons between free and vanilla AT with different $n$ values.\n\nI believe that the value of the theoretical bound on FreeAT's convergence outweighs the above weaknesses. Hence, a rating of 6 is given.", "questions": "- Do the theoretical results also apply to the $\\ell_\\infty$ case?\n- Has there been any work that empirically estimates the Lipschitz and smoothness constants in Assumptions 1 and 2?\n- Can smoothness and Lipschitzness assumptions (1 and 2) be relaxed? Specifically, instead of having this condition for all pairs of $\\delta, \\delta'$, is it possible to define Lipschitzness and smoothness over $\\delta$ w.r.t. the nominal point ($\\delta = 0$)? This relaxation will make the conditions more realistic.\n- What is Free-4 in Table 2 and some of the figures (including Figure 1)?\n- Theorem 3 lower-bounds $\\mathbb{E}[ || w(S) - w(S') || ]$. However, does a large $\\mathbb{E}[ || w(S) - w(S') || ]$ necessarily translate to large $\\mathcal{E}\\_{\\textrm{gen}}$? Isn't it the case that neural networks with very different weights can have similar behavior?\n- In practice, the attack loss function and the training loss function may not be the same, and using different losses has been empirically shown to decrease the generalization gap. Examples include TRADES [1] and ALP [2]. It's probably a stretch goal, but is it possible to extend the analysis to this scenario?\n\n[1] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" International conference on machine learning. PMLR, 2019. \\\n[2] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. \"Adversarial logit pairing.\" arXiv preprint\narXiv:1803.06373, 2018.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work utilizes the stability generalization framework to quantify the generalization bound of the Free Adversarial Training algorithm. Additionally, it shows that Free AT has a smaller generalization gap (but test error may not be smaller) and provides some theoretical intuitions.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "This work offers important insights into the convergence properties of free AT. Additionally, the authors highlight intuitive relationships between \"more simultaneous\" gradient updates during training and the resulting generalization capability. These results can inspire further improvements in robust training algorithms to alleviate overfitting. The paper is generally well-presented and easy to follow.", "weaknesses": "- The theoretical support for FreeAT having a smaller generalization gap than VanillaAT could be more rigorous. Specifically, while Theorem 2 presents pessimistic results for the convergence of VanillaAT, it is unclear whether this bound is tight. It is unclear whether the convergence difference between vanilla and free AT is due to the algorithm itself or some artifacts of the proof technique. While experiment results support this intuition, the paper would benefit from some additional explanations. It would be even better if some lower bounds could be provided for $\\mathcal{E}\\_{\\textrm{gen}} (A\\_{\\textrm{Vanilla}})$.\n- The relationship between a smaller generalization gap and better transferability is unclear. The motivation for the experiment setting of transferring attacks from a robust model to a standard model is weak. I suggest moving the transferability analysis to the Appendix (it's still good to have them) and making space for Table 2, which supports your main claim.\n- Figure 1 should use different line styles for train and test. In the current form, it's hard to distinguish them.\n- Since the proposed convergence bounds depend on the dataset size $n$, this paper would benefit from some empirical comparisons between free and vanilla AT with different $n$ values.\n\nI believe that the value of the theoretical bound on FreeAT's convergence outweighs the above weaknesses. Hence, a rating of 6 is given.", "questions": "- Do the theoretical results also apply to the $\\ell_\\infty$ case?\n- Has there been any work that empirically estimates the Lipschitz and smoothness constants in Assumptions 1 and 2?\n- Can smoothness and Lipschitzness assumptions (1 and 2) be relaxed? Specifically, instead of having this condition for all pairs of $\\delta, \\delta'$, is it possible to define Lipschitzness and smoothness over $\\delta$ w.r.t. the nominal point ($\\delta = 0$)? This relaxation will make the conditions more realistic.\n- What is Free-4 in Table 2 and some of the figures (including Figure 1)?\n- Theorem 3 lower-bounds $\\mathbb{E}[ || w(S) - w(S') || ]$. However, does a large $\\mathbb{E}[ || w(S) - w(S') || ]$ necessarily translate to large $\\mathcal{E}\\_{\\textrm{gen}}$? Isn't it the case that neural networks with very different weights can have similar behavior?\n- In practice, the attack loss function and the training loss function may not be the same, and using different losses has been empirically shown to decrease the generalization gap. Examples include TRADES [1] and ALP [2]. It's probably a stretch goal, but is it possible to extend the analysis to this scenario?\n\n[1] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" International conference on machine learning. PMLR, 2019. \\\n[2] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. \"Adversarial logit pairing.\" arXiv preprint\narXiv:1803.06373, 2018.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697676183362}], "openreview_url": "https://openreview.net/forum?id=N5ID99rsUq", "arxiv_id": "2404.08980", "paper_pdf": "papers/N5ID99rsUq.pdf", "paper_pdf_sha256": "245600e9c70dccdcdf487b8ce159d9a0b14950e2e5725971c34cf582e28acf86", "paper_pdf_bytes": 1917096, "paper_pdf_source": "openreview", "code_url": "https://github.com/Xiwei-Cheng/Stability_FreeAT", "code_repository": "Xiwei-Cheng/Stability_FreeAT", "code_commit": "aecc74df99c4874a23d1e3a1b6a7c84b682604a8", "code_archive": "repos/N5ID99rsUq.zip", "code_archive_sha256": "95d6bea8379d5e0b920490bd1414b8e62b8befe85c9c9e0d8b40aee18ba96af7", "code_archive_bytes": 17848, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 19, "github_languages": {"Python": 67532}, "github_archived": false, "github_pushed_at": "2024-04-08T09:28:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/stability-and-generalization-in-free"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gLl0fZQo6Vu", "year": 2023, "status": "rejected", "title": "Agent-Controller Representations: Principled Offline RL with Rich Exogenous Information", "authors": ["Riashat Islam", "Manan Tomar", "Alex Lamb", "Hongyu Zang", "Yonathan Efroni", "Dipendra Misra", "Xin Li", "Harm van Seijen", "Remi Tachet des Combes", "John Langford"], "authorids": ["~Riashat_Islam1", "~Manan_Tomar1", "~Alex_Lamb1", "~Hongyu_Zang1", "~Yonathan_Efroni2", "~Dipendra_Misra1", "~Xin_Li31", "~Harm_van_Seijen1", "~Remi_Tachet_des_Combes1", "~John_Langford1"], "authors_source": "OpenReview API", "abstract": "Learning to control an agent from data collected offline in a rich pixel-based visual observation space is vital for real-world applications of reinforcement learning (RL). A major challenge in this setting is the presence of input information that is hard to model and irrelevant to controlling the agent. This problem has been approached by the theoretical RL community through the lens of exogenous information, i.e, any control-irrelevant information contained in observations. For example, a robot navigating in busy streets needs to ignore irrelevant information, such as other people walking in the background, textures of objects, or birds in the sky. In this paper, we focus on the setting with visually detailed exogenous information, and introduce new offline RL benchmarks offering the ability to study this problem. We find that contemporary representation learning techniques can fail on datasets where the noise is a complex and time dependent process, which is prevalent in practical applications. To address these, we propose to use multi-step inverse models, which have seen a great deal of interest in the RL theory community, to learn Agent-Controller Representations for Offline-RL (ACRO). Despite being simple and requiring no reward, we show theoretically and empirically that the representation created by this objective greatly outperforms baselines.  ", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "VeTKGSBVJ74", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2146/Reviewer_GMBs"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper considers learning a good representation for offline reinforcement learning algorithms with pixel-based visual observation space. The authors aim to extract the representation, which ignores any control-irrelevant information. The authors choose multi-step inverse models to learn the observation representation because the multi-step action prediction for learning exogenous-invariant representations has been shown useful in  RL theory community.\n\nThe proposed method ACRO first trains the image encoders so that the features can be used to predict multiple actions between two observations. With the well-trained encoder, the downstream offline RL algorithm TD3+BC takes in the learned representation as policy input to output actions. \n\nIn order to evaluate the robustness of the proposed method for pixel observations, the authors introduce new datasets with temporally-correlated noise and diverse noise in the observations. The dataset is similar to D4RL but with some exogenous images or videos incorporated in the observations as noise.\n\nIn comparison with other representation learning methods, the proposed one significantly outperforms the baselines.", "review_text": "This paper presented a well-motivated representation learning method for offline RL, with great experiment performance and some theoretical foundation. But the proposed method may be only suitable for locomotion tasks, due to the limitation of the learned representation. This seems not general and widely useful.", "strengths": "Strengths:\nThe authors introduced the multi-step inverse model in offline RL problems and proved that Bellman completeness can be achieved via the representations without exogenous noise.\nThe experimental results support the claims well. In the datasets of noisy observations, this paper demonstrates the significant advantage of the proposed representation learning approach.\nThe newly introduced datasets for offline RL can be useful for the community.\n\nWeaknesses:\nThe major concern is about the limitation of the representation only containing control-relevant information. How about there are other moving entities in the observations, which are not controllable by the agent but are still very important for the success of the policy? For example, a robot navigates busy streets. Pedestrians are not controllable by the robot, but the robot should consider the status of the pedestrian to avoid collisions. Will the proposed representation learning approaches fully ignore pedestrians and make errors in the decision?\n\nAnother example is robotics manipulation tasks. The robot hand should manipulate the objects, so the object position and orientations are important information for the control policy. But the proposed method may tend to ignore the object information. This is problematic and limits the scope of the proposed method.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper considers learning a good representation for offline reinforcement learning algorithms with pixel-based visual observation space. The authors aim to extract the representation, which ignores any control-irrelevant information. The authors choose multi-step inverse models to learn the observation representation because the multi-step action prediction for learning exogenous-invariant representations has been shown useful in  RL theory community.\n\nThe proposed method ACRO first trains the image encoders so that the features can be used to predict multiple actions between two observations. With the well-trained encoder, the downstream offline RL algorithm TD3+BC takes in the learned representation as policy input to output actions. \n\nIn order to evaluate the robustness of the proposed method for pixel observations, the authors introduce new datasets with temporally-correlated noise and diverse noise in the observations. The dataset is similar to D4RL but with some exogenous images or videos incorporated in the observations as noise.\n\nIn comparison with other representation learning methods, the proposed one significantly outperforms the baselines.", "strength_and_weaknesses": "Strengths:\nThe authors introduced the multi-step inverse model in offline RL problems and proved that Bellman completeness can be achieved via the representations without exogenous noise.\nThe experimental results support the claims well. In the datasets of noisy observations, this paper demonstrates the significant advantage of the proposed representation learning approach.\nThe newly introduced datasets for offline RL can be useful for the community.\n\nWeaknesses:\nThe major concern is about the limitation of the representation only containing control-relevant information. How about there are other moving entities in the observations, which are not controllable by the agent but are still very important for the success of the policy? For example, a robot navigates busy streets. Pedestrians are not controllable by the robot, but the robot should consider the status of the pedestrian to avoid collisions. Will the proposed representation learning approaches fully ignore pedestrians and make errors in the decision?\n\nAnother example is robotics manipulation tasks. The robot hand should manipulate the objects, so the object position and orientations are important information for the control policy. But the proposed method may tend to ignore the object information. This is problematic and limits the scope of the proposed method.\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\nOverall, this paper is written clearly. The tables and figures are easy to read and convey information efficiently. I especially like table 1 which helps to compare different representation learning approaches.\nAlso, it will be great to explain the importance of the Bellman completeness more clearly and intuitively.\n\nQuality:\nThe quality is good with the technically solid method and impressive experiment results.\nHowever, the experiments are only conducted on locomotion tasks. It will be great to show results in D4RL kitchen tasks, especially because these tasks require object manipulation, and the proposed representation approach may fail in this scenario. I'm curious whether the proposed method can be modified or extended to the manipulation tasks.\n\nNovelty:\nThe novelty is okay but not very great, because the multi-step inverse model has been studied for a well in the RL area, and this work is just to adapt it from standard RL to offline RL.\n\nReproducibility:\nThe code is provided but the new datasets have not been released yet. Considering there is detailed information about hyper-parameters in the Appendix, the reproducibility is fine as long as the datasets can be released later.", "summary_of_the_review": "This paper presented a well-motivated representation learning method for offline RL, with great experiment performance and some theoretical foundation. But the proposed method may be only suitable for locomotion tasks, due to the limitation of the learned representation. This seems not general and widely useful.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666676767282}, {"id": "cv3kQ7qvXP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2146/Reviewer_M6qw"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a multi-step inverse model to learn the representation for RL problems with offline data. The proposed method ACRO predicts the current action based on the current state and a future state, in a reward-free manner. ACRO can learn controllable states while keep invariant to exogenous information. Experiments in various settings show that ACRO outperforms prior methods. Theoretical analysis shows the benefits of ACRO. The paper also provides several benchmarks to evaluate representation learning in RL.", "review_text": "This is an interesting paper with good empirical results. But I am hesitating to recommend it for acceptance, because 1) the method is a direction extension from prior work (and in the meanwhile it may lose the theoretical guarantees of prior work), and 2) there is a lack of in-depth discussion on the assumption and limitation of the current method, including requirement for offline dataset. I may missed some information in the paper and appendix. I would consider increase the ratings if the authors can provide convincing analysis/results for the above points. ", "strengths": "### Strengths\n\n1. The idea is neat and easy to implement, intuitive and theoretical grounded.\n2. The empirical results are promising, covering multiple scenarios.\n3. Good presentation and visualization of the ideas and results.\n4. Interesting theoretical insights.\n\n### Weaknesses and Questions:\n\n1. The novelty of the proposed method is relatively limited, as a continuous-control extension of prior work (Lamb et al. 2022). \n2. Following bullet 1, I am wondering whether the authors can comment more on the theoretical comparison with Lamb et al. 2022. Without the bottleneck constraint, will the learned representation be guaranteed to discard exogenous information? If not, the claims of Section 3 do not hold for the proposed algorithm, right?\n3. The algorithm adopts a process of offline representation pretraining + policy finetuning with frozen representation. However, such representation learning may highly depend on coverage of the offline dataset. I hope there could be more discussion on the requiremenet of data/exploration, both in theory and in experiments.\n4. The representation is only tested with offline RL finetuning, while it is not clear whether the learned representation is sufficient for learning optimal policies in an online manner. Intuitively, the latter would be harder and thus more interesting (the representation should be not only good for modeling optimal behaviors, but also good for exploration). If the claims hold that the learned representation is controllable, it should be able to learn a policy by interaction. Combined with bullet 3 above, I am worried that if the offline dataset does not have good coverage, downstream online policy learning will be hard.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a multi-step inverse model to learn the representation for RL problems with offline data. The proposed method ACRO predicts the current action based on the current state and a future state, in a reward-free manner. ACRO can learn controllable states while keep invariant to exogenous information. Experiments in various settings show that ACRO outperforms prior methods. Theoretical analysis shows the benefits of ACRO. The paper also provides several benchmarks to evaluate representation learning in RL.", "strength_and_weaknesses": "### Strengths\n\n1. The idea is neat and easy to implement, intuitive and theoretical grounded.\n2. The empirical results are promising, covering multiple scenarios.\n3. Good presentation and visualization of the ideas and results.\n4. Interesting theoretical insights.\n\n### Weaknesses and Questions:\n\n1. The novelty of the proposed method is relatively limited, as a continuous-control extension of prior work (Lamb et al. 2022). \n2. Following bullet 1, I am wondering whether the authors can comment more on the theoretical comparison with Lamb et al. 2022. Without the bottleneck constraint, will the learned representation be guaranteed to discard exogenous information? If not, the claims of Section 3 do not hold for the proposed algorithm, right?\n3. The algorithm adopts a process of offline representation pretraining + policy finetuning with frozen representation. However, such representation learning may highly depend on coverage of the offline dataset. I hope there could be more discussion on the requiremenet of data/exploration, both in theory and in experiments.\n4. The representation is only tested with offline RL finetuning, while it is not clear whether the learned representation is sufficient for learning optimal policies in an online manner. Intuitively, the latter would be harder and thus more interesting (the representation should be not only good for modeling optimal behaviors, but also good for exploration). If the claims hold that the learned representation is controllable, it should be able to learn a policy by interaction. Combined with bullet 3 above, I am worried that if the offline dataset does not have good coverage, downstream online policy learning will be hard.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written with reasonable quality. \n\nThe novelty is relatively limited since the proposed algorithm is directly extended from prior work.\n\nThe authors have provided code and detailed experiment settings.", "summary_of_the_review": "This is an interesting paper with good empirical results. But I am hesitating to recommend it for acceptance, because 1) the method is a direction extension from prior work (and in the meanwhile it may lose the theoretical guarantees of prior work), and 2) there is a lack of in-depth discussion on the assumption and limitation of the current method, including requirement for offline dataset. I may missed some information in the paper and appendix. I would consider increase the ratings if the authors can provide convincing analysis/results for the above points. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666662281780}, {"id": "3pOCV8qSpN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2146/Reviewer_yx3d"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "By defining exogeneous information as information irrelevant for control, this work works to learn a representation that removes these features in the context of offline RL. These representations are learned by taking a latent space learned through inverse dynamics modeling. It also provides a set of benchmarks for offline RL by adding background videos to Mujoco tasks. The key insight of this method is to predict the first action of the sequence connecting the first state and a state k steps in the future. \n", "review_text": "I propose to reject this work due to lack of novelty since inverse dynamics models are a common choice for state representation learning. It also has some concerns with the blind format, since it repeatedly cites works that are likely by the same authors. Last, the writing and definitions are not entirely consistent with the motivation. ", "strengths": "Strengths: \n\nThis work investigates an important question of how to construct representations based on controllability and introduces a method for doing so.\n\nThis work implements a large number of baselines, which would be useful if released as code. \nIt would have been nice to include in the introduction some indication of how this work innovates over existing representation learning methods since inverse dynamics models have been heavily used in RL and model-based RL.\n\nWeaknesses: \nThe description of an MDP actually describes a PO-MDP, with an exogenous block context. In addition, the typical formulation of the agent carries \\pi(a|x_{1,...t}), or the history of past observations. Policies conditioned on the current observation require the \"rich observation\" assumption, which should be stated in 2.1, not 2.2.\n\nThere is an unstated assumption that the policies from which the data is collected are also invariant to the exogenous information, otherwise, information about the action between two states would result in improved performance. This is actually a significant issue in the offline RL setting since we don't have control over the collected data.\n\nThe work claims that the multi-step objective allows for more diverse representations because it forces the representation to encode long-range dependencies. However, only the first action in the sequence is added, so it is entirely unintuitive as to why the representation would have to be different. In the example of the action being stored in the observation, this would still be the case even if t+k is given. Furthermore, this representation introduces its own issues, since the problem itself is now ill-defined: suppose that there are multiple action sequences between s_t and s_{t+1}. In this case, the inverse dynamics model would have to output an ambiguous probability which would be dependent on the frequency one path was taken over another in the dataset. It is not clear what representation this would encode. Figure 2 also shows x_t and x_{t+k} with the same image.\n\n\nA clear weakness of this work is that even though it equates exogenous with irrelevant, it is actually likely to remove information that could be task-vital. This is because the inverse dynamics model only needs to capture sufficient information to predict the first action between a pair of states, and can lose any information once it can make that prediction. Suppose that we are in an object manipulation domain where the agent uses an arm to grasp a block. The representation has no reason to capture the block because only the arm is necessary to predict the action. None of the experiments appear to capture these kinds of relationships", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "By defining exogeneous information as information irrelevant for control, this work works to learn a representation that removes these features in the context of offline RL. These representations are learned by taking a latent space learned through inverse dynamics modeling. It also provides a set of benchmarks for offline RL by adding background videos to Mujoco tasks. The key insight of this method is to predict the first action of the sequence connecting the first state and a state k steps in the future. \n", "strength_and_weaknesses": "Strengths: \n\nThis work investigates an important question of how to construct representations based on controllability and introduces a method for doing so.\n\nThis work implements a large number of baselines, which would be useful if released as code. \nIt would have been nice to include in the introduction some indication of how this work innovates over existing representation learning methods since inverse dynamics models have been heavily used in RL and model-based RL.\n\nWeaknesses: \nThe description of an MDP actually describes a PO-MDP, with an exogenous block context. In addition, the typical formulation of the agent carries \\pi(a|x_{1,...t}), or the history of past observations. Policies conditioned on the current observation require the \"rich observation\" assumption, which should be stated in 2.1, not 2.2.\n\nThere is an unstated assumption that the policies from which the data is collected are also invariant to the exogenous information, otherwise, information about the action between two states would result in improved performance. This is actually a significant issue in the offline RL setting since we don't have control over the collected data.\n\nThe work claims that the multi-step objective allows for more diverse representations because it forces the representation to encode long-range dependencies. However, only the first action in the sequence is added, so it is entirely unintuitive as to why the representation would have to be different. In the example of the action being stored in the observation, this would still be the case even if t+k is given. Furthermore, this representation introduces its own issues, since the problem itself is now ill-defined: suppose that there are multiple action sequences between s_t and s_{t+1}. In this case, the inverse dynamics model would have to output an ambiguous probability which would be dependent on the frequency one path was taken over another in the dataset. It is not clear what representation this would encode. Figure 2 also shows x_t and x_{t+k} with the same image.\n\n\nA clear weakness of this work is that even though it equates exogenous with irrelevant, it is actually likely to remove information that could be task-vital. This is because the inverse dynamics model only needs to capture sufficient information to predict the first action between a pair of states, and can lose any information once it can make that prediction. Suppose that we are in an object manipulation domain where the agent uses an arm to grasp a block. The representation has no reason to capture the block because only the arm is necessary to predict the action. None of the experiments appear to capture these kinds of relationships", "clarity,_quality,_novelty_and_reproducibility": "The usage of exogenous does not match the typical usages this reader is familiar with. In particular, in control/RL literature exogenous is typically used to describe elements that cannot be controlled (but might still affect control), and in causal literature, it describes elements that have no incoming causal links from the variables of interest. It seems like \"irrelevant\" or \"distractor\" information would be a more appropriate term. \n\nThis work repeatedly cites Guaranteed Discovery of Controllable Latent States with Multi-Step Inverse Models (Lamb et al 2022), which appears to approach the same problem with the same solution. If this work is by the authors, then this is disingenuous since the authors' claims are being supported by citations of the author's own work. Furthermore, the claims being supported, such as the resolving of single-step issues by using a multi-step predictor, appear to be made without strong support.\n\nThe authors contrast this work from others using inverse dynamics (ID) models by saying that this is the first which uses ID for representation learning instead of exploration. Not only is this not the case after a short search (Integrating State Representation Learning Into Deep Reinforcement Learning), but just because the ID representations are used for exploration, and often with more complex and refined techniques. If the authors do not write this piece of prior work, it would appear that this work introduces nothing new to the community.\n\nThe proposed offline RL datasets are the same (at least in conceit) as those described by Learning invariant representations for reinforcement learning without reconstruction (Zhang et. al. 2020), except that the prior work was introduced outside of the context of offline RL.", "summary_of_the_review": "I propose to reject this work due to lack of novelty since inverse dynamics models are a common choice for state representation learning. It also has some concerns with the blind format, since it repeatedly cites works that are likely by the same authors. Last, the writing and definitions are not entirely consistent with the motivation. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "I'm not sure the level reaches that for ethical review, but I am concerned by a paper support its claims by an arxiv paper of the same work.", "recommendation": "3: reject, not good enough"}, "tcdate": 1666623396696}, {"id": "XD1yyiELSa", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2146/Reviewer_GC23"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes to learn a multi-step inverse model [Lamb et al.] to learn a representation that ignores exogenous information (i.e., uncontrollable information) for offline RL. In addition, this paper introduces several temporally-correlated and diverse visual distractors on top of the v-d4rl dataset to investigate the quality of the representation in RL. The results show that the proposed method outperforms several self-supervised learning methods including CURL and DRIML.", "review_text": "Although this paper demonstrates good results on offline RL with exogenous information, the main idea seems to mostly come from the prior work [Lamb et al.], which is not clearly stated in the current manuscript. The paper would benefit from making a clear distinction from the prior work and adding a comparison to the prior work.", "strengths": "[Strength]\n* The idea is technically sound.\n* The results look good. \n* Introduces interesting visual distractors.\n\n\n[Weakness]\n* The technical novelty is not significant, as the main objective is essentially the same as Actor-Controller State from [Lamb et al.].\n* The novelty compared to the prior work [Lamb et al.] is not clearly explained.  \n* No empirical comparison to the closely related prior work [Lamb et al.]. \n* Some of the empirical results are not convincing. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes to learn a multi-step inverse model [Lamb et al.] to learn a representation that ignores exogenous information (i.e., uncontrollable information) for offline RL. In addition, this paper introduces several temporally-correlated and diverse visual distractors on top of the v-d4rl dataset to investigate the quality of the representation in RL. The results show that the proposed method outperforms several self-supervised learning methods including CURL and DRIML.", "strength_and_weaknesses": "[Strength]\n* The idea is technically sound.\n* The results look good. \n* Introduces interesting visual distractors.\n\n\n[Weakness]\n* The technical novelty is not significant, as the main objective is essentially the same as Actor-Controller State from [Lamb et al.].\n* The novelty compared to the prior work [Lamb et al.] is not clearly explained.  \n* No empirical comparison to the closely related prior work [Lamb et al.]. \n* Some of the empirical results are not convincing. ", "clarity,_quality,_novelty_and_reproducibility": "**Clarity**\n\nThis paper is easy to follow. However, the contribution/novelty in comparison to the relevant prior work was not clearly described. For example, when the paper introduces the main idea, it says \"Our proposed method, which we call Agent-Controller Representations for Offline-RL (ACRO), optimizes the following objective based on a multi-step inverse model: [Equation 1]\", without referring to [Lamb et al.] which proposed the equivalent objective called Agent-Controller state (AC state). As someone who is less familiar with [Lamb et al.], I had entirely misinterpreted the statement and thought this objective is original until I checked [Lamb et al.]. Similarly, Lemma 1 is in fact from [Lamb et al.], but there is no explicit statement about it. I highly encourage the authors to move the majority part of Section 2.2 (Proposed Method) to Section 2.1 (Preliminaries) and only highlight new ideas in Section 2.2.\n\n**Novelty**\n\nThe main objective (multi-step inverse model) is essentially the same as [Lamb et al.] except for a few details such as continuous representation in this paper as opposed to discretized representation in [Lamb et al.]. However, the application of this idea to offline RL setting is novel. In addition, the new visual-temporal disctractors introduced by this paper are new and interesting. \n\n**Quality**\n\n* A comparison to [Lamb et al.] should be included, because the paper claims that learning a continuous representation is the main distinction (hence novelty) to [Lamb et al.]. Without this, the use of continuous representation is not well-justified. \n\n* It would be more comprehensive to include a noise/distractor-free setting to get an idea of how much each method suffers from visual disctractors.  \n\n* It is unclear what \"normalized performance\" means, given that Figure 6 has various ranges from 100 to 500. It would be informative to present performance normalized w.r.t. demonstration performance across the entire main paper (assuming that raw returns are included in the appendix).\n\n* Although I appreciate the analysis through reconstruction (Figure 7), it's unclear whether the proposed method is clearly better than DRIML. It looks like the reconstruction from the proposed method is just blurrier than DRIML not only for the background but also for the controllable part. \n\n**Reproducibility**\nThe paper provided details of hyperparameters, architectures, datasets, but not the code.", "summary_of_the_review": "Although this paper demonstrates good results on offline RL with exogenous information, the main idea seems to mostly come from the prior work [Lamb et al.], which is not clearly stated in the current manuscript. The paper would benefit from making a clear distinction from the prior work and adding a comparison to the prior work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666536371738}, {"id": "FSvO1CLs7T", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2146/Reviewer_moWf"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a method for learning representations that are  invariant to task-unrelated information in the offline RL setting.  Theoretical results are presented which show that such representations  allow accurate value function approximation and that using  representations which are not invariant to task-unrelated information  can cause value function approximation to fail. Experiments and ablation studies support these claims.\n", "review_text": "The paper makes solid experimental and theoretic contributions to the problem of discarding irrelevant information in RL. The experimental results are both extensive and in support of the proposed method. The main drawbacks I see are the requirement for an exo-free policy to learn from and the block assumption, which excludes POMDPs.", "strengths": "**strengths**\n\n* The paper's about a relevant question (how do we ignore unimportant/distracting information in RL?).\n\n* The paper is rigorous and argues its case convincingly. The approach is motivated by theoretical arguments. The experiments are extensive and show an edge over strong baselines. Ablations empirically motivate the decisions that the authors make.\n\n**weaknesses**\n\n* The block assumption (that the emission distributions of different states are disjoint) is limiting. It amounts to a full-observability assumption.\n\n* Another limiting factor is that the learning problem requires data from an exo-free policy.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a method for learning representations that are  invariant to task-unrelated information in the offline RL setting.  Theoretical results are presented which show that such representations  allow accurate value function approximation and that using  representations which are not invariant to task-unrelated information  can cause value function approximation to fail. Experiments and ablation studies support these claims.\n", "strength_and_weaknesses": "**strengths**\n\n* The paper's about a relevant question (how do we ignore unimportant/distracting information in RL?).\n\n* The paper is rigorous and argues its case convincingly. The approach is motivated by theoretical arguments. The experiments are extensive and show an edge over strong baselines. Ablations empirically motivate the decisions that the authors make.\n\n**weaknesses**\n\n* The block assumption (that the emission distributions of different states are disjoint) is limiting. It amounts to a full-observability assumption.\n\n* Another limiting factor is that the learning problem requires data from an exo-free policy.", "clarity,_quality,_novelty_and_reproducibility": "* The presentation is a bit confusing at first. The description of the setup appears like a POMDP. Eventually it becomes clear that we are in a setup where the observation uniquely identifies the state. Here, I would advise using the term EX-BMDP more prominently (it is mentioned, but appears more like an aside) so that it is clear early on that we are in a special case.\n* It is interesting that ACRO outperforms other method even when there are no distractors (if I understand Table 2). This would imply that the representation learned by ACRO is a better representation than alternatives, even when there is no exogenous information. This is interesting, since the theory used to derive ACRO is all based on the existence of exogenous information. Do the authors have some thoughts here?\n* What is the difference in setup between Table 2 and Figure 5? As far as I understood, Table 2 contains no distractors, while Figure 5 contains the distractors used by Lu et al. [1]. What I don't understand is what \"easy\", \"medium\" and \"hard\" refer to in Figure 5. Also, the captions reads \"Normalized results across two domains [...]\": which domains?\n* Where are the numbers in Table 3 coming from? Are these averages across the tasks? In which distractor setting?", "summary_of_the_review": "The paper makes solid experimental and theoretic contributions to the problem of discarding irrelevant information in RL. The experimental results are both extensive and in support of the proposed method. The main drawbacks I see are the requirement for an exo-free policy to learn from and the block assumption, which excludes POMDPs.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666366579588}], "openreview_url": "https://openreview.net/forum?id=gLl0fZQo6Vu", "arxiv_id": "2211.00164", "paper_pdf": "papers/gLl0fZQo6Vu.pdf", "paper_pdf_sha256": "f38065170ab60adab869938b5b866e802caaafefa4b1e60722c8486a605e63db", "paper_pdf_bytes": 18339013, "paper_pdf_source": "openreview", "code_url": "https://github.com/manantomar/agent-centric-representations", "code_repository": "manantomar/agent-centric-representations", "code_commit": "f4622fadcb1fa2b6724dbab21ee62d1ac250bfe4", "code_archive": "repos/gLl0fZQo6Vu.zip", "code_archive_sha256": "aea44ce95fee289f1a7a9434acae24387e48459e73863ead03950895890a7926", "code_archive_bytes": 36760, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 28, "github_languages": {"Python": 86139}, "github_archived": false, "github_pushed_at": "2023-08-14T01:00:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/agent-controller-representations-principled"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MDT30TEtaVY", "year": 2022, "status": "rejected", "title": "Set Norm and Equivariant Skip Connections: Putting the Deep in Deep Sets", "authors": ["Lily H Zhang", "Veronica Tozzo", "John M. Higgins", "Rajesh Ranganath"], "authorids": ["~Lily_H_Zhang1", "~Veronica_Tozzo2", "~John_M._Higgins1", "~Rajesh_Ranganath2"], "authors_source": "OpenReview API", "abstract": "Permutation invariant neural networks are a promising tool for predictive modeling of set data. We show, however, that existing architectures struggle to perform well when they are deep. In this work, we address this issue for the two most widely used permutation invariant networks, Deep Sets and its transformer analogue Set Transformer. We take inspiration from previous efforts to scale neural network architectures by incorporating normalization layers and skip connections that work for sets. First, we motivate and develop set norm, a normalization tailored for sets. Then, we employ equivariant residual connections and introduce the ``clean path principle'' for their placement. With these changes, our many-layer Deep Sets++ and Set Transformer++ models reach comparable or better performance than their original counterparts on a diverse suite of tasks, from point cloud classification to regression on sets of images. We additionally introduce Flow-RBC, a new single-cell dataset and real-world application of permutation invariant prediction. On this task, our new models outperform existing methods as well as a clinical baseline. We open-source our data and code here: link-omitted-for-anonymity.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "q850lD4eEI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3756/Reviewer_Qf1Y"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper has four main contributions:\n\n1. introduces the _set norm_ normalization layer in neural network models for set data, as opposed to feature normalization (aka. batch-norm) and layer normalization layers;\n2. provides intuition behind a \"cleaner\" implementation of residual connections;\n3. implements set norm and the cleaner residual connections in modified versions of existing Deep Sets and Set Transformers models;\n4. introduces a new dataset called Flow-RBC with over 100,000 examples. Each \"example\" consists of a $(X,y)$ pair, where $X=\\\\{x_1, \\dotsc, x_{1000}\\\\}$ is a set of measurements on 1000 red-blood cells from a patient, and $y$ is a hematocrit level label. Each $x_i$ consists of both volume and hemoglobin mass measurements. This dataset is significantly larger than similar existing datasets.", "review_text": "## Strengths\n\n**S1) Overall well-written and clear**\n\nThe motivation for developing set-norm is clear, and the comparison to feature norm and layer norm is also clear. The motivation for the \n\"clean path\" principle for residual connections is presented well and intuitive.\n\n**S2) Interesting new dataset**\n\nAdmittedly, as I have no medical background, I am not a good adjudicator on the \"value\" of the new Flow-RBC dataset. However, taking the authors at their word, it seems to be much larger than existing similar datasets and demonstrates a useful real-world test case for set-based neural networks, beyond more commonly seen point cloud data and synthetic toy datasets (e.g., sum of MNIST digits).\n\n## Weaknesses\n\n**W1) What is special about an equivariant residual connection? And what is the relationship between the \"clean path principle\" and equivariance?**\n\nIt is unclear to me how an \"equivariant residual connection\" differs from a regular residual connection. To me, it seems like the discussion of the \"clean path principle\" is independent of equivariance and applies generally to all residual connections. If so, this needs to be made clearer in the paper.\n\n**W2) Could use more empirical comparisons**\n\nIn particular, both of the original Deep Sets and Set Transformers paper do set anomaly detection on CelebA. It would be helpful to see a comparison of Deep Sets++ and Set Transformers++ on this task\n\n**W3) Error bars on experimental results**\n\nWere the experiments run once, or many times? Could you provide error bars for the experiments?\n\n## Clarity Issues\n\n- Figure 2: should the \"F\" axis labels be \"D\"?\n\n- Please add units to all of the tables (or at least the table captions). I know that the units (MSE) are written in the main text, but they can be hard to find when looking just at the tables.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper has four main contributions:\n\n1. introduces the _set norm_ normalization layer in neural network models for set data, as opposed to feature normalization (aka. batch-norm) and layer normalization layers;\n2. provides intuition behind a \"cleaner\" implementation of residual connections;\n3. implements set norm and the cleaner residual connections in modified versions of existing Deep Sets and Set Transformers models;\n4. introduces a new dataset called Flow-RBC with over 100,000 examples. Each \"example\" consists of a $(X,y)$ pair, where $X=\\\\{x_1, \\dotsc, x_{1000}\\\\}$ is a set of measurements on 1000 red-blood cells from a patient, and $y$ is a hematocrit level label. Each $x_i$ consists of both volume and hemoglobin mass measurements. This dataset is significantly larger than similar existing datasets.", "main_review": "## Strengths\n\n**S1) Overall well-written and clear**\n\nThe motivation for developing set-norm is clear, and the comparison to feature norm and layer norm is also clear. The motivation for the \n\"clean path\" principle for residual connections is presented well and intuitive.\n\n**S2) Interesting new dataset**\n\nAdmittedly, as I have no medical background, I am not a good adjudicator on the \"value\" of the new Flow-RBC dataset. However, taking the authors at their word, it seems to be much larger than existing similar datasets and demonstrates a useful real-world test case for set-based neural networks, beyond more commonly seen point cloud data and synthetic toy datasets (e.g., sum of MNIST digits).\n\n## Weaknesses\n\n**W1) What is special about an equivariant residual connection? And what is the relationship between the \"clean path principle\" and equivariance?**\n\nIt is unclear to me how an \"equivariant residual connection\" differs from a regular residual connection. To me, it seems like the discussion of the \"clean path principle\" is independent of equivariance and applies generally to all residual connections. If so, this needs to be made clearer in the paper.\n\n**W2) Could use more empirical comparisons**\n\nIn particular, both of the original Deep Sets and Set Transformers paper do set anomaly detection on CelebA. It would be helpful to see a comparison of Deep Sets++ and Set Transformers++ on this task\n\n**W3) Error bars on experimental results**\n\nWere the experiments run once, or many times? Could you provide error bars for the experiments?\n\n## Clarity Issues\n\n- Figure 2: should the \"F\" axis labels be \"D\"?\n\n- Please add units to all of the tables (or at least the table captions). I know that the units (MSE) are written in the main text, but they can be hard to find when looking just at the tables.", "summary_of_the_review": "The paper is well-written and provides solid empirical justification for specific choices of normalization (set norm) and residual connection design (\"clean path principle\") for neural networks that operate on sets. While \"set norm\" is somewhat novel, I am unsure about whether the \"clean path principle\" applies specifically to these set-based neural networks, or whether there is something specific about that design for equivariant networks. Overall, a solid empirical paper, but I hesitate to say that it is truly \"novel.\" That said, I am open to increasing my score based on authors' replies.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636158083932}, {"id": "otIRQ2ze98m", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3756/Reviewer_eZF1"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper aims to improve training of deep neural networks applied to sets. The authors investigate permutation equivariant normalizations and residual connections in DeepSet and SetTransformer models, and propose a modification of layer norm which calculates standardization statistics over both features and all set elements. They also propose a modification to DeepSet and SetTransformer models inspired from pre-norm ResNet and pre-norm Transformer. The modifications are experimentally tested on a pointcloud classification dataset, synthetic variance regression, and a new dataset for set prediction contributed by the authors.", "review_text": "The main weaknesses are: lack of statistical analysis of the experimental evaluation and limited novelty of the pre-norm changes.\n\nLack of statistical analysis of the experimental evaluation: the authors use relatively small datasets which seem to exhibit high variance. Single run results presented by the authors are inconclusive and cannot be used to support claimed contributions. The authors could use bootstrapping or train multiple models with different random seed and report mean and std for all baseline and candidate models.\n\nLimited novelty of the pre-norm changes: pre-norm ResNet and pre-norm Transformer are very common, the proposed addition of pre-norm block to DeepSet and SetTransformer does not seem significant.\n\nAnother weakness is the usage of synthetic datasets. Instead of synthetic datasets for variance prediction the authors could evaluate their changes on established benchmarks for predictions on sets like object detection (e.g. COCO dataset) and trajectory prediction (e.g. Argoverse dataset). In object detection models which operate on sets of input features and predict sets have been gaining popularity since DETR [1], whereas in trajectory prediction VectorNet [2] is an example of a regression model with set inputs. Both are very similar to SetTransformer, employ layer normalization, residual connections and could potentially benefit from Set Norm and pre-norm Transformer.\n\nNotes to authors:\n- each dataset description in section 6 needs a clear definition of task solved (classification/regression), size of train/val splits, loss functions used, and metrics used to evaluate\n- figure 1 lacks x and y labels for each subfigure\n- figure 3 would benefit from short subcaption for each subfigure\n- all tables lack column metric definitions. It is not clear what the presented numbers mean, which are classification errors (or accuracy?) and which are mean squared errors\n- I could not understand Appendix A, the description of Flow-RBC dataset. An explanation in layman's terms would be very helpful. More formally, a description of what input set elements and targets are in the dataset is required.\n\n[1] Carion et al. End-to-end object detection with transformers\n[2] Gao et al. Vectornet: Encoding hd maps and agent dynamics from vectorized representation", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper aims to improve training of deep neural networks applied to sets. The authors investigate permutation equivariant normalizations and residual connections in DeepSet and SetTransformer models, and propose a modification of layer norm which calculates standardization statistics over both features and all set elements. They also propose a modification to DeepSet and SetTransformer models inspired from pre-norm ResNet and pre-norm Transformer. The modifications are experimentally tested on a pointcloud classification dataset, synthetic variance regression, and a new dataset for set prediction contributed by the authors.", "main_review": "The main weaknesses are: lack of statistical analysis of the experimental evaluation and limited novelty of the pre-norm changes.\n\nLack of statistical analysis of the experimental evaluation: the authors use relatively small datasets which seem to exhibit high variance. Single run results presented by the authors are inconclusive and cannot be used to support claimed contributions. The authors could use bootstrapping or train multiple models with different random seed and report mean and std for all baseline and candidate models.\n\nLimited novelty of the pre-norm changes: pre-norm ResNet and pre-norm Transformer are very common, the proposed addition of pre-norm block to DeepSet and SetTransformer does not seem significant.\n\nAnother weakness is the usage of synthetic datasets. Instead of synthetic datasets for variance prediction the authors could evaluate their changes on established benchmarks for predictions on sets like object detection (e.g. COCO dataset) and trajectory prediction (e.g. Argoverse dataset). In object detection models which operate on sets of input features and predict sets have been gaining popularity since DETR [1], whereas in trajectory prediction VectorNet [2] is an example of a regression model with set inputs. Both are very similar to SetTransformer, employ layer normalization, residual connections and could potentially benefit from Set Norm and pre-norm Transformer.\n\nNotes to authors:\n- each dataset description in section 6 needs a clear definition of task solved (classification/regression), size of train/val splits, loss functions used, and metrics used to evaluate\n- figure 1 lacks x and y labels for each subfigure\n- figure 3 would benefit from short subcaption for each subfigure\n- all tables lack column metric definitions. It is not clear what the presented numbers mean, which are classification errors (or accuracy?) and which are mean squared errors\n- I could not understand Appendix A, the description of Flow-RBC dataset. An explanation in layman's terms would be very helpful. More formally, a description of what input set elements and targets are in the dataset is required.\n\n[1] Carion et al. End-to-end object detection with transformers\n[2] Gao et al. Vectornet: Encoding hd maps and agent dynamics from vectorized representation", "summary_of_the_review": "The contributions of the draft are empirical and the experimental evaluation lacks an analysis of statistical significance, so the contributions cannot be evaluated conclusively. At least one contribution has questionable novelty.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635978386573}, {"id": "cG3bTwvlS5n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3756/Reviewer_ifHM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents improved versions of two standard premutation-invariant networks: Deep Sets++ and Set Transformer++. They make two critical design decisions: set normalization and clean residual paths. They also introduce Flow-RBC, a new benchmark for permutation-invariant prediction. Finally, the paper evaluates several variants on several benchmarks, including Flow-RBC.", "review_text": "Strengths\n+ I enjoyed the discussion on design decisions around normalization layers and agree that set norm is better than what was used in the Set Transformer paper because it discards less information. Furthermore, this intuition is confirmed by experiments in tables 2~4, which show that set norm improves performance.\n+ The discussion about having a clean residual path is interesting, and the paper draws a compelling parallel to design choices in ResNet models.\n\nWeaknesses\n- None of the results have error bounds, so it's hard to tell whether the differences are statistically significant. This problem is exacerbated because loss scales vary widely between tasks (10^1 Hematocrit, 10^-2~-4 Normal Var). Looking at the learning curves in Figure 1, the vanilla and ++ versions of some networks seem to have similar performance on a few tasks.\n- The paper increases the number of layers for both architectures, and this is never addressed in the experiments. For example, what is the performance of Set Transformer++ with 2 layers, compared to the vanilla version?\n\nMinor comments\n- Figure 1: the caption says that the figures have (grey, orange, blue) lines, but the colors seem closer to (pink, orange, green). This was slightly confusing.\n- The rightmost two columns of Table 4 were hard to understand at first. Maybe it's better to have two rows with the performance numbers for \"Original\" and \"No Norm\"?\n- I think Table 5 should be moved to the main paper, to show that permutation-invariant networks compare favorably to traditional approaches in this real-world setting.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents improved versions of two standard premutation-invariant networks: Deep Sets++ and Set Transformer++. They make two critical design decisions: set normalization and clean residual paths. They also introduce Flow-RBC, a new benchmark for permutation-invariant prediction. Finally, the paper evaluates several variants on several benchmarks, including Flow-RBC.", "main_review": "Strengths\n+ I enjoyed the discussion on design decisions around normalization layers and agree that set norm is better than what was used in the Set Transformer paper because it discards less information. Furthermore, this intuition is confirmed by experiments in tables 2~4, which show that set norm improves performance.\n+ The discussion about having a clean residual path is interesting, and the paper draws a compelling parallel to design choices in ResNet models.\n\nWeaknesses\n- None of the results have error bounds, so it's hard to tell whether the differences are statistically significant. This problem is exacerbated because loss scales vary widely between tasks (10^1 Hematocrit, 10^-2~-4 Normal Var). Looking at the learning curves in Figure 1, the vanilla and ++ versions of some networks seem to have similar performance on a few tasks.\n- The paper increases the number of layers for both architectures, and this is never addressed in the experiments. For example, what is the performance of Set Transformer++ with 2 layers, compared to the vanilla version?\n\nMinor comments\n- Figure 1: the caption says that the figures have (grey, orange, blue) lines, but the colors seem closer to (pink, orange, green). This was slightly confusing.\n- The rightmost two columns of Table 4 were hard to understand at first. Maybe it's better to have two rows with the performance numbers for \"Original\" and \"No Norm\"?\n- I think Table 5 should be moved to the main paper, to show that permutation-invariant networks compare favorably to traditional approaches in this real-world setting.", "summary_of_the_review": "This paper proposes interesting design choices for permutation-invariant networks, but I cannot tell whether the empirical results are statistically significant. Additionally, the paper does not show the effect of their components on the original shallow networks. I'd be happy to increase my score if these issues are addressed.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635636346178}, {"id": "xX2ZcOPkPMR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3756/Reviewer_Dj6B"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper tackles the problem of designing deep networks that operate on sets. The challenge is addressed by normalization layers and skip connections. The authors formulate some design choices and advocate the choice of set normalization and clean residual connections. The layers are incorporated with two different choices of architectures and are evaluated on several sets learning benchmarks. In addition, a new benchmark is introduced.\n", "review_text": "I appreciate the effort to tackle the challenge of designing deep networks. Indeed,  I feel that it is an important research question. \nOverall the paper is well written and easy to follow. The formulation of the method is presented in a general form which allows the reader to easily understand the core details. \nI appreciate the effort to introduce an additional benchmark, which will be publicly available to the community.\n\nHowever, I do have some concerns regarding the current version of the paper, detailed next.\n\nThe authors state that “existing architectures struggle to perform well when they are deep”.  However, I am not convinced that the claim is True as existing architectures do incorporate normalization layers. Moreover, I think the authors should consider refining the original question to  two separated questions:\n\n1) Do normalization and skip connections help networks in learning ? regardless of the choice of how many layers are there.\n2) Is a deep architecture a good inductive bias for some tasks involving sets?\n\nFigure 1 does not answer the first question as it does not apply normalization+skip layers to standard architectures.  It also does not fully answer the second question as it does not evaluate on different depths of networks. To my understanding, the main implication from Figure 1 is that with 50 layers and without normalization the learning performs poorly, which is not novel.\n\nMoreover, existing architectures already incorporate normalization layers. For example, in the pointnet [1] paper, which is a similar architecture to deepsets,  they use normalization layers (feature normalization). Do they improve/stable/disimprove with respect to network depth? Is there a difference between set normalization to feature normalization? Moreover, the results in table 2 do not show a clear advantage for the set norm over the feature norm.\n\n*Relation to graphs*\n\nThe current paper focuses on sets. Note that there is extensive literature on the more general settings of deep permutation invariant networks, which is the case of graph neural networks. For example, what is the relation between the proposed set norm and the one suggested in [2]? I would expect a more detailed discussion about this challenge in graphs versus sets. \n\n*Residual connections*\n\nIncorporating residual connections in set networks is not novel. It seems that some prior works have used it, as the adaptation of it to set networks seems natural. For example, see implementation details in the supplementary material of [3]. Few alternatives to clean path residual connections are discussed. Do the authors know of such works that follow these alternatives? What are the advantages of these alternatives? Are there any tasks which they improve? If the answer is no, then it is not clear why are they discussed.\n\n*Requirements for a design of a normalization layer*\n\nThe authors identify some requirements for normalization layers of permutation equivariant networks.  What does “work well” in the context of varying set sizes mean?  What does essential information mean? It feels like these requirements are too general and should be refined. Moreover, it seems that some discussion relating to test time versus train time equivariance is missing. Is it true that the normalization choices that the authors show that fail (such as L = {B} )  in satisfying equivariance is only at training time? As in test time, some constant inferred from the batch statistics is used. If that is the case, then I expect to see a discussion on why satisfying the equivariant property in training time is important.\n\n*Results*\n\nFor classification tasks, it would be more informative to report accuracy. For regression tasks, the units of the measurement should be reported and explained as well.\n\n[1] Qi, Charles R., et al. Pointnet: Deep learning on point sets for 3d classification and segmentation.\n\n[2] Zhao, Lingxiao, and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns.\n\n[3]  Mescheder, Lars, et al. Occupancy networks: Learning 3d reconstruction in function space.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper tackles the problem of designing deep networks that operate on sets. The challenge is addressed by normalization layers and skip connections. The authors formulate some design choices and advocate the choice of set normalization and clean residual connections. The layers are incorporated with two different choices of architectures and are evaluated on several sets learning benchmarks. In addition, a new benchmark is introduced.\n", "main_review": "I appreciate the effort to tackle the challenge of designing deep networks. Indeed,  I feel that it is an important research question. \nOverall the paper is well written and easy to follow. The formulation of the method is presented in a general form which allows the reader to easily understand the core details. \nI appreciate the effort to introduce an additional benchmark, which will be publicly available to the community.\n\nHowever, I do have some concerns regarding the current version of the paper, detailed next.\n\nThe authors state that “existing architectures struggle to perform well when they are deep”.  However, I am not convinced that the claim is True as existing architectures do incorporate normalization layers. Moreover, I think the authors should consider refining the original question to  two separated questions:\n\n1) Do normalization and skip connections help networks in learning ? regardless of the choice of how many layers are there.\n2) Is a deep architecture a good inductive bias for some tasks involving sets?\n\nFigure 1 does not answer the first question as it does not apply normalization+skip layers to standard architectures.  It also does not fully answer the second question as it does not evaluate on different depths of networks. To my understanding, the main implication from Figure 1 is that with 50 layers and without normalization the learning performs poorly, which is not novel.\n\nMoreover, existing architectures already incorporate normalization layers. For example, in the pointnet [1] paper, which is a similar architecture to deepsets,  they use normalization layers (feature normalization). Do they improve/stable/disimprove with respect to network depth? Is there a difference between set normalization to feature normalization? Moreover, the results in table 2 do not show a clear advantage for the set norm over the feature norm.\n\n*Relation to graphs*\n\nThe current paper focuses on sets. Note that there is extensive literature on the more general settings of deep permutation invariant networks, which is the case of graph neural networks. For example, what is the relation between the proposed set norm and the one suggested in [2]? I would expect a more detailed discussion about this challenge in graphs versus sets. \n\n*Residual connections*\n\nIncorporating residual connections in set networks is not novel. It seems that some prior works have used it, as the adaptation of it to set networks seems natural. For example, see implementation details in the supplementary material of [3]. Few alternatives to clean path residual connections are discussed. Do the authors know of such works that follow these alternatives? What are the advantages of these alternatives? Are there any tasks which they improve? If the answer is no, then it is not clear why are they discussed.\n\n*Requirements for a design of a normalization layer*\n\nThe authors identify some requirements for normalization layers of permutation equivariant networks.  What does “work well” in the context of varying set sizes mean?  What does essential information mean? It feels like these requirements are too general and should be refined. Moreover, it seems that some discussion relating to test time versus train time equivariance is missing. Is it true that the normalization choices that the authors show that fail (such as L = {B} )  in satisfying equivariance is only at training time? As in test time, some constant inferred from the batch statistics is used. If that is the case, then I expect to see a discussion on why satisfying the equivariant property in training time is important.\n\n*Results*\n\nFor classification tasks, it would be more informative to report accuracy. For regression tasks, the units of the measurement should be reported and explained as well.\n\n[1] Qi, Charles R., et al. Pointnet: Deep learning on point sets for 3d classification and segmentation.\n\n[2] Zhao, Lingxiao, and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns.\n\n[3]  Mescheder, Lars, et al. Occupancy networks: Learning 3d reconstruction in function space.\n", "summary_of_the_review": "* Well written and easy to follow\n* The list of requirements identified by the authors from normalization equivariant layers needs a major refinement. \n* Lack of novelty and the benefit of the set norm is unclear\n* Missing discussion on the relation to graph neural networks\n* The experiments do not test properly the effect of depth on learning\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635347612020}], "openreview_url": "https://openreview.net/forum?id=MDT30TEtaVY", "arxiv_id": "2206.11925", "paper_pdf": "papers/MDT30TEtaVY.pdf", "paper_pdf_sha256": "e3a789479d8691761dcb0d4e57be72c22934034e3f297dabafceabc6e904ea9e", "paper_pdf_bytes": 461720, "paper_pdf_source": "openreview", "code_url": "https://github.com/rajesh-lab/deep_permutation_invariant", "code_repository": "rajesh-lab/deep_permutation_invariant", "code_commit": "7d25da12329d3d89a69c2f5333f94234248f01a5", "code_archive": "repos/MDT30TEtaVY.zip", "code_archive_sha256": "a9e89578b413ce3fbe4bbac92580dd6b282418379548c6def7601e9949c456dc", "code_archive_bytes": 35988, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 25, "github_languages": {"Python": 98456}, "github_archived": false, "github_pushed_at": "2022-10-11T20:48:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/set-norm-and-equivariant-skip-connections-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jpm1AfJucwt", "year": 2021, "status": "rejected", "title": "Revisiting Loss Modelling for Unstructured Pruning", "authors": ["César Laurent", "Camille Ballas", "Thomas George", "Pascal Vincent", "Nicolas Ballas"], "authorids": ["~César_Laurent1", "~Camille_Ballas1", "~Thomas_George2", "~Pascal_Vincent1", "~Nicolas_Ballas1"], "authors_source": "OpenReview API", "abstract": "By removing parameters from deep neural networks, unstructured pruning methods aim at cutting down memory footprint and computational cost, while maintaining prediction accuracy. In order to tackle this otherwise intractable problem, many of these methods model the loss landscape using first or second order Taylor expansions to identify which parameters can be discarded. We revisit loss modelling for unstructured pruning: we show the importance of ensuring locality of the pruning steps, and systematically compare first and second order Taylor expansions. Finally, we show that better preserving the original network function does not necessarily transfer to better performing networks after fine-tuning, suggesting that only considering the impact of pruning on the loss might not be a sufficient objective to design good pruning criteria.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "d1QycbMMAq4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper432/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n- This paper conducted a detailed study on how does the loss modeling affects the final performance of the pruned model. The authors first provided a unified view of various pruning algorithms (e.g., Magnitude Pruning, SNIP, OBD, and OBS), which can be categorized into three classes: weight magnitude, linear and quadratic models of the loss function. In the experiments, the authors seek to answer the questions: 1) how well do each criterion preserve the loss; 2) how does the locality assumption affect the final performance; and 3) how does the loss relate to the final performance? Empirical, the authors found that the quadratic model preserves the loss the best, as expected. Also, the loss after pruning seems not strongly correlated with the performance after fine-tuning.\n\nOverall:\n\nThis paper is well-written and easy to follow. The authors did a great job of unifying the analysis of several pruning algorithms. More importantly, revisiting the loss modeling of network pruning is interesting, and it might invoke further research efforts in better understanding the pruning techniques developed in the past and also inspire researchers in designing improved pruning algorithms. However, I still have the following questions:\n\n- The authors show that the loss after pruning does not correlate strongly with the accuracy after fine-tuning. In Figure 3, the change in the loss ranges from 0 to 5. Can you show the plot with a smaller range of the change in loss, e.g., 0~0.5? I believe a large change in loss means that the pruning results are very close to random, so the comparisons in this regime may not be meaningful.\n\n- For testing the locality assumption, you introduce an L2 penalty on the changes. To me, this is more like a weighted combination of the original pruning criteria and the magnitude pruning criteria. Why not using some other techniques, such as backtracking line search for determining the pruning ratio at each iteration? \n\n- In equation (5), why do we need to take the absolute value? I think preserving the loss is only meaningful when the network is converged. If the model is not converged, then it would be preferable to prune those weights whose removal will decrease the loss. In this sense, the sign of the loss change should not be ignored. \n\n- The third plot in the second row of Figure 1 shows that OBD with more iterations has a larger change in the loss. Do you have any explanation for this?\n\nRating:\n- I vote for a weak acceptance due to the above reasons. I believe the studied topic in this paper is important and impactful for the pruning community. In the meantime, it would be great if the author can propose some hypotheses on this phenomenon.  To me, preserving the loss is a way to enforce the pruned network to stay close to the original solution, and in this regime, it's easy for the optimization algorithm to find a good enough solution. I will raise my rating if the authors can address my concerns well during the rebuttal period.\n\n==========================After rebuttal=================================\n\nThanks very much for your efforts to address my concerns. I kept my score unchanged. I agree with most of the responses, except for the response to \"L2 penalty, backtracking line search for determining the pruning ratio\". This paper is not proposing a practical algorithm but a revisiting, so I don't think the computational cost is a bottleneck in preventing you from using more advanced methods to get more robust conclusions.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well-written paper with interesting results.", "review": "Summary:\n- This paper conducted a detailed study on how does the loss modeling affects the final performance of the pruned model. The authors first provided a unified view of various pruning algorithms (e.g., Magnitude Pruning, SNIP, OBD, and OBS), which can be categorized into three classes: weight magnitude, linear and quadratic models of the loss function. In the experiments, the authors seek to answer the questions: 1) how well do each criterion preserve the loss; 2) how does the locality assumption affect the final performance; and 3) how does the loss relate to the final performance? Empirical, the authors found that the quadratic model preserves the loss the best, as expected. Also, the loss after pruning seems not strongly correlated with the performance after fine-tuning.\n\nOverall:\n\nThis paper is well-written and easy to follow. The authors did a great job of unifying the analysis of several pruning algorithms. More importantly, revisiting the loss modeling of network pruning is interesting, and it might invoke further research efforts in better understanding the pruning techniques developed in the past and also inspire researchers in designing improved pruning algorithms. However, I still have the following questions:\n\n- The authors show that the loss after pruning does not correlate strongly with the accuracy after fine-tuning. In Figure 3, the change in the loss ranges from 0 to 5. Can you show the plot with a smaller range of the change in loss, e.g., 0~0.5? I believe a large change in loss means that the pruning results are very close to random, so the comparisons in this regime may not be meaningful.\n\n- For testing the locality assumption, you introduce an L2 penalty on the changes. To me, this is more like a weighted combination of the original pruning criteria and the magnitude pruning criteria. Why not using some other techniques, such as backtracking line search for determining the pruning ratio at each iteration? \n\n- In equation (5), why do we need to take the absolute value? I think preserving the loss is only meaningful when the network is converged. If the model is not converged, then it would be preferable to prune those weights whose removal will decrease the loss. In this sense, the sign of the loss change should not be ignored. \n\n- The third plot in the second row of Figure 1 shows that OBD with more iterations has a larger change in the loss. Do you have any explanation for this?\n\nRating:\n- I vote for a weak acceptance due to the above reasons. I believe the studied topic in this paper is important and impactful for the pruning community. In the meantime, it would be great if the author can propose some hypotheses on this phenomenon.  To me, preserving the loss is a way to enforce the pruned network to stay close to the original solution, and in this regime, it's easy for the optimization algorithm to find a good enough solution. I will raise my rating if the authors can address my concerns well during the rebuttal period.\n\n==========================After rebuttal=================================\n\nThanks very much for your efforts to address my concerns. I kept my score unchanged. I agree with most of the responses, except for the response to \"L2 penalty, backtracking line search for determining the pruning ratio\". This paper is not proposing a practical algorithm but a revisiting, so I don't think the computational cost is a bottleneck in preventing you from using more advanced methods to get more robust conclusions.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604742132345}, {"id": "mkMz61ZDztH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper432/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Although the paper is covering an interesting topic, much of what's in the paper can be found in other works, and there's not a lot of novelty to the insights, nor a large breadth of experiments to justify it as a survey paper.\n- The linear and quadratic loss functions are not new.\n- The enforcing locality part is essentially why the pruning strength annealing schemes exist, this insight is not new and can be found in Zhu&Gupta, or in Bayesian settings like in the Molchanov paper, they suggest annealing from a Bayesian statistical perspective\n- The fact that this leads to multiple stages of pruning to be a good idea, is also known in the literature this paper cites.\n- The most meat of the paper is in the 'survey' part of it, investigating the results... but this section feels lacking since there are not a lot of experiments, and the insights of e.g. post-finetuning can be largely found in e.g. Blalock et al. I'm missing deeper insights/analysis here. What is the reason for this? How do we remedy this? What are the characteristics of networks that lead to this behavior? \n- It would have been great to see a lot more insight/experiments on this topic. The authors throw up a lot of hypothesis and suggestions/ideas throughout the paper, but don't back them up. E.g. If the ||Theta|| term is better of to be constant throughout the pruning procedure... can we somehow make an annealing scheme that keeps ||Theta|| small and constant throughout the pruning process, and show that that works well? This can be proven/shown somehow. \n\nI do think the paper is well written; and I encourage the authors to look further into this topic and come up with more novel insights/results and methods to improve pruning\n\nOther things/questions/suggestions:\n- In formulations (1), (2) and (5), (6). Why are the absolute brackets necessary? Especially for models that have not converged, why would you want to stay close to the original loss, as opposed to just decreasing the overall loss of the model? \n- For most of the discussion in section 3.2, the authors talk about the norm of the delta theta squared being large. But this largeness is relative to the 0th and first order term which the authors glance over. Under 'other considerations' for example, if the weights theta are large, the gradients likely follow suit. Thus the absolute magnitude of the weights might not matter, as it's the relative size of this to the gradient terms that should be considered. \n- Constraining the step size in section 3.2. Interestingly, if you take the assumptions that for each layer, the hessian of the loss w.r.t. the output is sqrt(lambda/2), and your input distributions to a layer are normalized, you end up with the weaker norm penalty. This is a crude approximation of the second order term, which would give this method a bit more of a theoretical foundation than just a regularization term.\n- 5.1 Convergence assumption. I don't get this part, both the OBD method, and the linear and quadratic loss terms depend on the locality, so all will also depend on the amount of steps taken. For OBD, as long as you recalculate the Gauss-Newton matrix, I don't see why this method is different when not doing fine-tuning. \n- 5.1 Convergence assumption. The result cited in appendix A is a very well-know result. How could this link explain the OBD performance on the VGG network? 'Could' is not strong enough to make it into a paper. \n\nSmall tidbits:\n3. Do pruning criteria better at preserving the loss lead to better fine-tuned networks? <- this sentence doesn't flow nicely. I would add a 'that' so you have do pruning criteria that are better ...", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A paper that combines many different things from other works, and runs a few extra experiments, but in the end doesn't feel like it has enough body and depth for a full-conference paper.", "review": "Although the paper is covering an interesting topic, much of what's in the paper can be found in other works, and there's not a lot of novelty to the insights, nor a large breadth of experiments to justify it as a survey paper.\n- The linear and quadratic loss functions are not new.\n- The enforcing locality part is essentially why the pruning strength annealing schemes exist, this insight is not new and can be found in Zhu&Gupta, or in Bayesian settings like in the Molchanov paper, they suggest annealing from a Bayesian statistical perspective\n- The fact that this leads to multiple stages of pruning to be a good idea, is also known in the literature this paper cites.\n- The most meat of the paper is in the 'survey' part of it, investigating the results... but this section feels lacking since there are not a lot of experiments, and the insights of e.g. post-finetuning can be largely found in e.g. Blalock et al. I'm missing deeper insights/analysis here. What is the reason for this? How do we remedy this? What are the characteristics of networks that lead to this behavior? \n- It would have been great to see a lot more insight/experiments on this topic. The authors throw up a lot of hypothesis and suggestions/ideas throughout the paper, but don't back them up. E.g. If the ||Theta|| term is better of to be constant throughout the pruning procedure... can we somehow make an annealing scheme that keeps ||Theta|| small and constant throughout the pruning process, and show that that works well? This can be proven/shown somehow. \n\nI do think the paper is well written; and I encourage the authors to look further into this topic and come up with more novel insights/results and methods to improve pruning\n\nOther things/questions/suggestions:\n- In formulations (1), (2) and (5), (6). Why are the absolute brackets necessary? Especially for models that have not converged, why would you want to stay close to the original loss, as opposed to just decreasing the overall loss of the model? \n- For most of the discussion in section 3.2, the authors talk about the norm of the delta theta squared being large. But this largeness is relative to the 0th and first order term which the authors glance over. Under 'other considerations' for example, if the weights theta are large, the gradients likely follow suit. Thus the absolute magnitude of the weights might not matter, as it's the relative size of this to the gradient terms that should be considered. \n- Constraining the step size in section 3.2. Interestingly, if you take the assumptions that for each layer, the hessian of the loss w.r.t. the output is sqrt(lambda/2), and your input distributions to a layer are normalized, you end up with the weaker norm penalty. This is a crude approximation of the second order term, which would give this method a bit more of a theoretical foundation than just a regularization term.\n- 5.1 Convergence assumption. I don't get this part, both the OBD method, and the linear and quadratic loss terms depend on the locality, so all will also depend on the amount of steps taken. For OBD, as long as you recalculate the Gauss-Newton matrix, I don't see why this method is different when not doing fine-tuning. \n- 5.1 Convergence assumption. The result cited in appendix A is a very well-know result. How could this link explain the OBD performance on the VGG network? 'Could' is not strong enough to make it into a paper. \n\nSmall tidbits:\n3. Do pruning criteria better at preserving the loss lead to better fine-tuned networks? <- this sentence doesn't flow nicely. I would add a 'that' so you have do pruning criteria that are better ...", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604066144410}, {"id": "Mqtc21Q3HyK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper432/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper develops two modified version of Optimal Brain Damage (OBD) criteria, namely LM and QM (linear and quadratic model) to measure the importance/saliency of model weights. It then compares these three together with Magnitude Pruning (MP), and show that these among four criteria: 1. for the first three, using iterative pruning to enforce locality of the gradient calculation is important, 2. the best method for training loss before fine-tuning does not necessarily lead to best validation accuracy after fine-tuning. \n\nThe paper's empirical investigation is valuable and appreciated. It helps me understand OBD more throughly and the assumptions behind it. It is also useful to know that using iterative pruning can improve these gradient approximation-based methods because of locality.\n\n1. My primary concern is the experiments does not seem to lead to a useful guidance for future practice. The paper does conclude for the first three using iterative pruning is useful, but these three criteria are rarely used nowadays and MP is the mainstream, and the simplest method. The paper also didn't conclude which of the four criteria is in general best and recommended. From table 2 it seems to be LM but the paper did not conclude this way. This is also possibly due to that the experiments are not run extensively on different datasets and architectures. \n\n2. The paper compared the training loss before fine-tuning, and validation acc. after fine-tuning. But I think validation acc. before fine-tuning is also a quantity worth investigating. \n\n3. More importantly, the paper shows the training loss (before fine-tuning) and valid acc. (after fine-tuning) are not necessarily correlated, but did not give explanation on why this could be the case through experiments, or give useful suggestions to achieve a good valid acc. after fine-tuning.\n\nOverall I appreciate the empirical study but I suggest conducting the experiments on more datasets and architectures, and extract a useful conclusion to guide future practices.\n\n+++++++++++++++++\n\nI appreciate the clarified messages of the paper, and would like to see them emphasized more clearly in the next version of the paper. But due to the limited experimental scale on ImageNet (added in rebuttal, and in my understanding, it only verifies one of the multiple observations mentioned in the paper), I'm still leaning on rejection. I updated my score from 4 to 5.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Appreciate the empirical study but cannot find substantial useful conclusions", "review": "The paper develops two modified version of Optimal Brain Damage (OBD) criteria, namely LM and QM (linear and quadratic model) to measure the importance/saliency of model weights. It then compares these three together with Magnitude Pruning (MP), and show that these among four criteria: 1. for the first three, using iterative pruning to enforce locality of the gradient calculation is important, 2. the best method for training loss before fine-tuning does not necessarily lead to best validation accuracy after fine-tuning. \n\nThe paper's empirical investigation is valuable and appreciated. It helps me understand OBD more throughly and the assumptions behind it. It is also useful to know that using iterative pruning can improve these gradient approximation-based methods because of locality.\n\n1. My primary concern is the experiments does not seem to lead to a useful guidance for future practice. The paper does conclude for the first three using iterative pruning is useful, but these three criteria are rarely used nowadays and MP is the mainstream, and the simplest method. The paper also didn't conclude which of the four criteria is in general best and recommended. From table 2 it seems to be LM but the paper did not conclude this way. This is also possibly due to that the experiments are not run extensively on different datasets and architectures. \n\n2. The paper compared the training loss before fine-tuning, and validation acc. after fine-tuning. But I think validation acc. before fine-tuning is also a quantity worth investigating. \n\n3. More importantly, the paper shows the training loss (before fine-tuning) and valid acc. (after fine-tuning) are not necessarily correlated, but did not give explanation on why this could be the case through experiments, or give useful suggestions to achieve a good valid acc. after fine-tuning.\n\nOverall I appreciate the empirical study but I suggest conducting the experiments on more datasets and architectures, and extract a useful conclusion to guide future practices.\n\n+++++++++++++++++\n\nI appreciate the clarified messages of the paper, and would like to see them emphasized more clearly in the next version of the paper. But due to the limited experimental scale on ImageNet (added in rebuttal, and in my understanding, it only verifies one of the multiple observations mentioned in the paper), I'm still leaning on rejection. I updated my score from 4 to 5.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603951704627}, {"id": "wZrzZC7c-de", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper432/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe authors study the use of loss-modeling to maintain model quality when inducing unstructured sparsity in deep neural networks. They study a range of different approximations and modifications that can help improve the quality of the approximation (taking local steps, avoid large changes in weight magnitude, avoiding assumptions about convergence). The authors conduct a thorough empirical investigation that yields practical observations for the design of future pruning techniques.\n\nPros:\n\nThe paper is well written and well organized and the empirical investigations are well done. The observations made by the authors are interesting and practically useful for the development of future pruning techniques. Namely,\n1. That including first order terms in loss approximations can relax the convergence assumption behind some existing loss modeling approaches to enable more flexible application of the pruning algorithm.\n2. The quality of the local loss approximation can be improved by taking a series of smaller pruning steps.\n3. That loss-preservation does not necessarily translate into accuracy preservation.\n\nCons:\n\nIt would be nice to see experiments in domains other than computer vision. For example, language modeling with RNNs or Transformers. Results at a wider range of sparsity levels for ImageNet would also have been useful, as it seems possible that these techniques could perform differently for high sparsity (>90%) than they do for moderate sparsity (e.g., the 70% sparsity reported in Figure 4).\n\nComments:\n\nAnother ICLR 2021 submission is highly relevant to your investigation: https://openreview.net/forum?id=rumv7QmLUue. Their theoretical/empirical results appear to corroborate your conclusions that loss preservation is not necessarily the best metric to optimize for when you care about accuracy preservation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper with solid empirical evaluation", "review": "Summary:\n\nThe authors study the use of loss-modeling to maintain model quality when inducing unstructured sparsity in deep neural networks. They study a range of different approximations and modifications that can help improve the quality of the approximation (taking local steps, avoid large changes in weight magnitude, avoiding assumptions about convergence). The authors conduct a thorough empirical investigation that yields practical observations for the design of future pruning techniques.\n\nPros:\n\nThe paper is well written and well organized and the empirical investigations are well done. The observations made by the authors are interesting and practically useful for the development of future pruning techniques. Namely,\n1. That including first order terms in loss approximations can relax the convergence assumption behind some existing loss modeling approaches to enable more flexible application of the pruning algorithm.\n2. The quality of the local loss approximation can be improved by taking a series of smaller pruning steps.\n3. That loss-preservation does not necessarily translate into accuracy preservation.\n\nCons:\n\nIt would be nice to see experiments in domains other than computer vision. For example, language modeling with RNNs or Transformers. Results at a wider range of sparsity levels for ImageNet would also have been useful, as it seems possible that these techniques could perform differently for high sparsity (>90%) than they do for moderate sparsity (e.g., the 70% sparsity reported in Figure 4).\n\nComments:\n\nAnother ICLR 2021 submission is highly relevant to your investigation: https://openreview.net/forum?id=rumv7QmLUue. Their theoretical/empirical results appear to corroborate your conclusions that loss preservation is not necessarily the best metric to optimize for when you care about accuracy preservation.", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603829986859}], "openreview_url": "https://openreview.net/forum?id=jpm1AfJucwt", "arxiv_id": "2006.12279", "paper_pdf": "papers/jpm1AfJucwt.pdf", "paper_pdf_sha256": "fb73c1aabdd4e9dcc51ba19a09b9c9c23079b20ba77f0f70a24a0135ef7a81d2", "paper_pdf_bytes": 1425439, "paper_pdf_source": "openreview", "code_url": "https://github.com/Thrandis/loss-models-pruning", "code_repository": "Thrandis/loss-models-pruning", "code_commit": "b784b84cd2494e59673849dfd3b3e45e996a7e7a", "code_archive": "repos/jpm1AfJucwt.zip", "code_archive_sha256": "b474a4e926a41e698873d8c0deb6a703d8a079f61daaa2c59588c02954d86286", "code_archive_bytes": 14683, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 29, "github_languages": {"Python": 36312, "Dockerfile": 194}, "github_archived": false, "github_pushed_at": "2020-09-17T19:48:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/revisiting-loss-modelling-for-unstructured"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJggX0EKwS", "year": 2020, "status": "rejected", "title": "The Benefits of Over-parameterization at Initialization in Deep ReLU Networks", "authors": ["Devansh Arpit", "Yoshua Bengio"], "authorids": ["devansharpit@gmail.com", "yoshua.bengio@mila.quebec"], "authors_source": "OpenReview API", "abstract": "It has been noted in existing literature that over-parameterization in ReLU networks generally improves performance. While there could be several factors involved behind this, we prove some desirable theoretical properties at initialization which may be enjoyed by ReLU networks. Specifically, it is known that He initialization in deep ReLU networks asymptotically preserves variance of activations in the forward pass and variance of gradients in the backward pass for infinitely wide networks, thus preserving the flow of information in both directions. Our paper goes beyond these results and shows novel properties that hold under He initialization: i) the norm of hidden activation of each layer is equal to the norm of the input, and, ii) the norm of weight gradient of each layer is equal to the product of norm of the input vector and the error at output layer. These results are derived using the PAC analysis framework, and hold true for finitely sized datasets such that the width of the ReLU network only needs to be larger than a certain finite lower bound. As we show, this lower bound depends on the depth of the network and the number of samples, and by the virtue of being a lower bound, over-parameterized ReLU networks are endowed with these desirable properties. For the aforementioned hidden activation norm property under He initialization, we further extend our theory and show that this property holds for a finite width network even when the number of data samples is infinite. Thus we overcome several limitations of existing papers, and show new properties of deep ReLU networks at initialization.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SkxEBvuUcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1025/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work considers random parameter initialization in neural networks (In particular the initialization presented in He et al.) and develops non-asymptotic bounds for the norms and gradients of neural networks during initialization. The authors show that the norms of the outputs and gradients (for gradients, under a different assumption on the dimension of the matrix) remain constant through the different layers. The results presented differ from previous work in that they give nice concentration bounds for such output and gradient norms. In addition the authors prove results in the case of infinite samples under the assumption that they arise from a finite dimensional space.\n\nOverall the presentation is clean, but the results presented have the following issues.\n\n1.Very similar in nature to previous results, for example the results presented in [1] Theorem 5.4 give very similar concentration results and use very similar mechanisms.\n\n2. All the results for the infinite stream coming from a finite dimensional subspace do not address the fact that the training points usually do not lay in a linear subspace of dimension d << n. Further, there is a stringent requirement on d in such results and they fail to hold for even \"fairly\" small d.\n\n3. In Theorem 2, the argument that the output through a layer of a rank-d linear subspace remains within a rank-d linear subspace does not seem correct, is it possible that it remains in the union of subsets each of which lies in a subspace of rank d?\n\n4. The proofs for the gradients make assumptions that deviate from the initialization previously introduced. The work of Glorot et al. for example discusses the tradeoffs between the two assumptions and suggests the balancing between maintaining the input and output variance distributions.\n\nDue to the following issues I choose to reject this work at this time.\n\nBelow are additional minor typos or issues in the paper.\nOn page 6, the sentence that starts with: \"it is beneficial for the width\": Suggesting to use neural networks of constant width seems a bit impractical.\n\nThe citation of Arpit et al. from 2017 seems possibly wrong since the paper cited discusses memorization and does not focus at the initialization of neural networks.\n\n\nTYPOS:\npage 1 last sentence before equation 1:  back backward? \n\npage 1 last sentence: repeated words: as as the the\n\npage 1 last sentence: property --> properties?\n\npage 2 section 2 paragraph 1: both these papers --> both papers, both of these papers?\n\npage 4 sentence after the proof sketch paragraph: the sentence is difficult to read\n\n[1] \"Stochastic gradient descent optimizes over-parameterized deep relu networks\", 2018, Difan Zou, Yuan Cao, Dongruo Zhou, Quanquan Gu\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This work considers random parameter initialization in neural networks (In particular the initialization presented in He et al.) and develops non-asymptotic bounds for the norms and gradients of neural networks during initialization. The authors show that the norms of the outputs and gradients (for gradients, under a different assumption on the dimension of the matrix) remain constant through the different layers. The results presented differ from previous work in that they give nice concentration bounds for such output and gradient norms. In addition the authors prove results in the case of infinite samples under the assumption that they arise from a finite dimensional space.\n\nOverall the presentation is clean, but the results presented have the following issues.\n\n1.Very similar in nature to previous results, for example the results presented in [1] Theorem 5.4 give very similar concentration results and use very similar mechanisms.\n\n2. All the results for the infinite stream coming from a finite dimensional subspace do not address the fact that the training points usually do not lay in a linear subspace of dimension d << n. Further, there is a stringent requirement on d in such results and they fail to hold for even \"fairly\" small d.\n\n3. In Theorem 2, the argument that the output through a layer of a rank-d linear subspace remains within a rank-d linear subspace does not seem correct, is it possible that it remains in the union of subsets each of which lies in a subspace of rank d?\n\n4. The proofs for the gradients make assumptions that deviate from the initialization previously introduced. The work of Glorot et al. for example discusses the tradeoffs between the two assumptions and suggests the balancing between maintaining the input and output variance distributions.\n\nDue to the following issues I choose to reject this work at this time.\n\nBelow are additional minor typos or issues in the paper.\nOn page 6, the sentence that starts with: \"it is beneficial for the width\": Suggesting to use neural networks of constant width seems a bit impractical.\n\nThe citation of Arpit et al. from 2017 seems possibly wrong since the paper cited discusses memorization and does not focus at the initialization of neural networks.\n\n\nTYPOS:\npage 1 last sentence before equation 1:  back backward? \n\npage 1 last sentence: repeated words: as as the the\n\npage 1 last sentence: property --> properties?\n\npage 2 section 2 paragraph 1: both these papers --> both papers, both of these papers?\n\npage 4 sentence after the proof sketch paragraph: the sentence is difficult to read\n\n[1] \"Stochastic gradient descent optimizes over-parameterized deep relu networks\", 2018, Difan Zou, Yuan Cao, Dongruo Zhou, Quanquan Gu\n\n\n"}, "tcdate": 1572403004360}, {"id": "SkgXIteScr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1025/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies initialization techniques for deep ReLU networks from a theoretical standpoint and derives finite layer width concentration bounds to show that with the He initialization scheme, deep ReLU networks preserve the norm of the input sample during a forward pass and the norm of the gradient with respect to the output during a backward pass. The concentration bounds also suggest lower bounds on the width of the ReLU layers. The authors verify their theory with experiments on synthetic data.\n\nWhile I believe the finite sample concentration bounds for networks initialized using the He initialization are valuable, it seems to me that this work is incremental in terms of understanding the He initialization. The techniques used in the paper are also fairly well known Chernoff bounding techniques and concentration of Gaussian random vectors and matrices.\n\nThe authors claims on explaining overparameterization are also overstated in my opinion. While the authors are able to obtain a lower bound on layer widths in order to preserve norms during forward/backward passes at initialization, the lower bound is only dependent on the input dimension and not the size of the dataset, which is the relevant quantity to decide whether a model is under/over parameterized. Even in the authors' bounds for finite datasets, what I can surmise from their results (it would be better to explicitly state it if that is one of the goals of this paper) is that the width of each layer needs to be atleast log(N) where N is the size of the dataset. This is hardly overparameterized. \n\nFurthermore, the authors do not explain how studying the properties of the initialization might help understand generalization at minima. Since gradient descent based techniques seem to prefer solutions that are close to initialization, the analysis in this paper might be a useful starting point in understanding generalization.\n\nThe authors could also consider how adding BatchNorm layers and/or Residual connections affect the He initialization scheme, and whether initialization matters for those techniques. Finite width concentration bounds for initialization of networks using Batchnorm/Residual connections could be useful. \n\nTo summarize, I do not see how the authors claims about explaining overparameterization (even at initialization) can be made. Without those claims the contribution of this paper is incremental and does not warrant publication at this time. I am willing to adjust my score if the claims about overparameterization are stated explicitly and make sense.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper studies initialization techniques for deep ReLU networks from a theoretical standpoint and derives finite layer width concentration bounds to show that with the He initialization scheme, deep ReLU networks preserve the norm of the input sample during a forward pass and the norm of the gradient with respect to the output during a backward pass. The concentration bounds also suggest lower bounds on the width of the ReLU layers. The authors verify their theory with experiments on synthetic data.\n\nWhile I believe the finite sample concentration bounds for networks initialized using the He initialization are valuable, it seems to me that this work is incremental in terms of understanding the He initialization. The techniques used in the paper are also fairly well known Chernoff bounding techniques and concentration of Gaussian random vectors and matrices.\n\nThe authors claims on explaining overparameterization are also overstated in my opinion. While the authors are able to obtain a lower bound on layer widths in order to preserve norms during forward/backward passes at initialization, the lower bound is only dependent on the input dimension and not the size of the dataset, which is the relevant quantity to decide whether a model is under/over parameterized. Even in the authors' bounds for finite datasets, what I can surmise from their results (it would be better to explicitly state it if that is one of the goals of this paper) is that the width of each layer needs to be atleast log(N) where N is the size of the dataset. This is hardly overparameterized. \n\nFurthermore, the authors do not explain how studying the properties of the initialization might help understand generalization at minima. Since gradient descent based techniques seem to prefer solutions that are close to initialization, the analysis in this paper might be a useful starting point in understanding generalization.\n\nThe authors could also consider how adding BatchNorm layers and/or Residual connections affect the He initialization scheme, and whether initialization matters for those techniques. Finite width concentration bounds for initialization of networks using Batchnorm/Residual connections could be useful. \n\nTo summarize, I do not see how the authors claims about explaining overparameterization (even at initialization) can be made. Without those claims the contribution of this paper is incremental and does not warrant publication at this time. I am willing to adjust my score if the claims about overparameterization are stated explicitly and make sense."}, "tcdate": 1572305226519}, {"id": "Hkx9eVWAYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1025/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper shows that under He initialization and for sufficiently wide network, (1) the norm of the activations of a L layered ReLU is preserved w.r.t the input across layers and (2) the norm of a weight matrix gradient at different layer is only dependent on the norm of the top-layer error and the input, because the norm of back-propagated gradient is approximately preserved. \n\nThe paper is clearly written and easy to read and the proofs are quite straightforward. That being said, the results are not surprising and from my point of view, the overall novelty of this paper is a bit marginal for top-tier conference like ICLR. \n\nThus I vote for rejection. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper shows that under He initialization and for sufficiently wide network, (1) the norm of the activations of a L layered ReLU is preserved w.r.t the input across layers and (2) the norm of a weight matrix gradient at different layer is only dependent on the norm of the top-layer error and the input, because the norm of back-propagated gradient is approximately preserved. \n\nThe paper is clearly written and easy to read and the proofs are quite straightforward. That being said, the results are not surprising and from my point of view, the overall novelty of this paper is a bit marginal for top-tier conference like ICLR. \n\nThus I vote for rejection. "}, "tcdate": 1571849202227}, {"id": "BkxInWuatH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1025/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies the norm of hidden activation of each layer and the norm of weight gradient of each layer for deep ReLU neural network. By using concentralization property of the random initialization, the paper derives their expected values and high probability range when the network width is sufficiently wide. The results are correct and the paper is easy to follow. However, the result has been given in previous work. I do not recommend the acceptance.\n\nThe result presented in this paper has been covered by a recent work [1]. Please refer to Section 7.1 for the forward part and Section 7.3 for the backward part.\n\n[1] Zeyuan Allen-Zhu, Yuanzhi Li and Zhao Song. A Convergence Theory for Deep Learning via Over-Parameterization", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper studies the norm of hidden activation of each layer and the norm of weight gradient of each layer for deep ReLU neural network. By using concentralization property of the random initialization, the paper derives their expected values and high probability range when the network width is sufficiently wide. The results are correct and the paper is easy to follow. However, the result has been given in previous work. I do not recommend the acceptance.\n\nThe result presented in this paper has been covered by a recent work [1]. Please refer to Section 7.1 for the forward part and Section 7.3 for the backward part.\n\n[1] Zeyuan Allen-Zhu, Yuanzhi Li and Zhao Song. A Convergence Theory for Deep Learning via Over-Parameterization"}, "tcdate": 1571811758427}], "openreview_url": "https://openreview.net/forum?id=rJggX0EKwS", "arxiv_id": "1901.03611", "paper_pdf": "papers/rJggX0EKwS.pdf", "paper_pdf_sha256": "7fb0a3faf914f6f556351d6d4aa8222e6c0053195d847f89ae11fcd60cdb4d8f", "paper_pdf_bytes": 622695, "paper_pdf_source": "openreview", "code_url": "https://github.com/devansharpit/overparametrization_benefits", "code_repository": "devansharpit/overparametrization_benefits", "code_commit": "9711531e87e51712eca8bb625e1a801534ef8676", "code_archive": "repos/rJggX0EKwS.zip", "code_archive_sha256": "c324a1c08ac36e17ffcfb8832943bc1250ace9be4e36553d524828f93d9a6921", "code_archive_bytes": 5091, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 28, "github_languages": {"Python": 6016}, "github_archived": false, "github_pushed_at": "2019-08-20T02:44:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-benefits-of-over-parameterization-at"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJl0jiRqtX", "year": 2019, "status": "rejected", "title": "EDDI: Efficient Dynamic Discovery of High-Value Information with Partial VAE", "authors": ["Chao Ma", "Sebastian Tschiatschek", "Konstantina Palla", "Jose Miguel Hernandez Lobato", "Sebastian Nowozin", "Cheng Zhang"], "authorids": ["cm905@cam.ac.uk", "sebastian.tschiatschek@microsoft.com", "konstantina.palla@microsoft.com", "jmh233@cam.ac.uk", "sebastian.nowozin@microsoft.com", "cheng.zhang@microsoft.com"], "authors_source": "OpenReview API", "abstract": "Making decisions requires information relevant to the task at hand. Many real-life decision-making situations allow acquiring further relevant information at a specific cost. For example, in assessing the health status of a patient we may decide to take additional measurements such as diagnostic tests or imaging scans before making a final assessment. More information that is relevant allows for better decisions but it may be costly to acquire all of this information.  How can we trade off the desire to make good decisions with the option to acquire further information at a cost? To this end, we propose a principled framework, named EDDI (Efficient Dynamic Discovery of high-value Information), based on the theory of Bayesian experimental design. In EDDI we propose a novel partial variational autoencoder (Partial VAE), to efficiently handle missing data over varying subsets of known information. EDDI combines this Partial VAE with an acquisition function that maximizes expected information gain on a set of target variables. EDDI is efficient and demonstrates that dynamic discovery of high-value information is possible; we show cost reduction at the same decision quality and improved decision quality at the same cost in benchmarks and in two health-care applications.. We believe there is great potential for realizing these gains in real-world decision support systems.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HJxgYLy-pX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper666/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "----I acknowledge that the authors have made improvements to the paper and have increased my score to 6\n\nThis is still definitely not my area of expertise and so I am leaving my confidence score low. \n---\n\nThe paper presents an algorithm EDDI that uses a a partial VAE and does active feature selection. The authors show quite a bit of experiments that seem to indicate the approach gives positive results.  However, since this is not my main area of expertise I do not know if these tasks are standard evaluation for this task.\n\nFor instance in Section 4.3, 4.4 why don't the authors plot accuracy as a function of steps/number of variables observed. That would seem much more useful than log likelihood.\n\nIn general, I found the methodology in the paper to be difficult to understand and not enough background was given.\nI think the paper would be clearer if it was more self contained.\n\n-For instance, I found much of Section 3 to not have enough background. The authors use lots of terminology around VAEs but don't give enough rigorous background so the paper doesn't feel self contained. \n\n-The same is true regarding \"amortized inference\" which I also feel isn't rigorously defined anywhere but often discussed. \n\n-The task for Section 4.1 (image inpainting) is not quite defined.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting but difficult to read", "review": "----I acknowledge that the authors have made improvements to the paper and have increased my score to 6\n\nThis is still definitely not my area of expertise and so I am leaving my confidence score low. \n---\n\nThe paper presents an algorithm EDDI that uses a a partial VAE and does active feature selection. The authors show quite a bit of experiments that seem to indicate the approach gives positive results.  However, since this is not my main area of expertise I do not know if these tasks are standard evaluation for this task.\n\nFor instance in Section 4.3, 4.4 why don't the authors plot accuracy as a function of steps/number of variables observed. That would seem much more useful than log likelihood.\n\nIn general, I found the methodology in the paper to be difficult to understand and not enough background was given.\nI think the paper would be clearer if it was more self contained.\n\n-For instance, I found much of Section 3 to not have enough background. The authors use lots of terminology around VAEs but don't give enough rigorous background so the paper doesn't feel self contained. \n\n-The same is true regarding \"amortized inference\" which I also feel isn't rigorously defined anywhere but often discussed. \n\n-The task for Section 4.1 (image inpainting) is not quite defined.", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541629559582}, {"id": "B1lI6Wnah7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper666/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present an information discovery approach based on (partial) variational autoencoders and an information theoretic acquisition function that seeks to maximize the expected information gain over a set of unobserved variables. Results are presented on image inpainting, UCI datasets and health data, namely ICU and NHANES.\n\nIt is not clear why multiple recurrent steps improve perfromance. This is not conceptually justified and empirically (see Figure 8), it is also unclear whether PNP5 significantly outperforms PNP1. Further, results seem to support that PNP is always better than PN, so why introduce the methodology around PN or even present it at all. Note that the authors do not offer an explanation about the perfromance differences between PN and PNP.\n\nIn the inpainting regions section, the authors write about well-calibrated uncertainties without any context. What do they mean by calibration, well-calibrated and how can they support their claim about it?\n\nIn Figure 3 it is not clear that PNP+Ours outperforms PNP+SING. For Boston hosing seems to be marginally better but the error bars (which I assume are standard deviations, not stated) make difficult to ascertain whether the differences are significant. Although I understand the value of having \"personalized\" decisions, one wonders whether this personalization comes with any generalizable measurable gains given the results.\n\nThe results in Table 2 need to be clarified and further explained. 1) what are the error bars, considering multiple runs and datasets? 2) How can EDDI be so much better than SING when individual AUICs in Tables 6-11, the only significant difference (accounting for error bars) is on Boston data? 3) according to Tables 6-11, PNP is only the best in 1 of 5 datasets, so how come is the overall beast by a large margin? This being said, the results in Table 2 are at best misleading.\n\nIn Table 4, how can PNP-EDDI be so much better than PNP-SING, when in Figure 6 error bars overlap almost everywhere?\n\nI enjoyed reading the paper, the motivation is clear and the problem is important. The approach is modestly novel compared to existing approaches and in general well explained despite the fact that the need for multiple recurrent steps is not well justified and the differences between PN and PNP, advantages/disadvantages and when to use each are not described or explored in the experiments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "EDDI: EFFICIENT DYNAMIC DISCOVERY OF HIGH-VALUE INFORMATION WITH PARTIAL VAE", "review": "The authors present an information discovery approach based on (partial) variational autoencoders and an information theoretic acquisition function that seeks to maximize the expected information gain over a set of unobserved variables. Results are presented on image inpainting, UCI datasets and health data, namely ICU and NHANES.\n\nIt is not clear why multiple recurrent steps improve perfromance. This is not conceptually justified and empirically (see Figure 8), it is also unclear whether PNP5 significantly outperforms PNP1. Further, results seem to support that PNP is always better than PN, so why introduce the methodology around PN or even present it at all. Note that the authors do not offer an explanation about the perfromance differences between PN and PNP.\n\nIn the inpainting regions section, the authors write about well-calibrated uncertainties without any context. What do they mean by calibration, well-calibrated and how can they support their claim about it?\n\nIn Figure 3 it is not clear that PNP+Ours outperforms PNP+SING. For Boston hosing seems to be marginally better but the error bars (which I assume are standard deviations, not stated) make difficult to ascertain whether the differences are significant. Although I understand the value of having \"personalized\" decisions, one wonders whether this personalization comes with any generalizable measurable gains given the results.\n\nThe results in Table 2 need to be clarified and further explained. 1) what are the error bars, considering multiple runs and datasets? 2) How can EDDI be so much better than SING when individual AUICs in Tables 6-11, the only significant difference (accounting for error bars) is on Boston data? 3) according to Tables 6-11, PNP is only the best in 1 of 5 datasets, so how come is the overall beast by a large margin? This being said, the results in Table 2 are at best misleading.\n\nIn Table 4, how can PNP-EDDI be so much better than PNP-SING, when in Figure 6 error bars overlap almost everywhere?\n\nI enjoyed reading the paper, the motivation is clear and the problem is important. The approach is modestly novel compared to existing approaches and in general well explained despite the fact that the need for multiple recurrent steps is not well justified and the differences between PN and PNP, advantages/disadvantages and when to use each are not described or explored in the experiments.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541419454385}, {"id": "HJgQ25LF27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper666/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes Partial VAE to handle missing data and a variable-wise active learning method. The model combines Partial VAE with the acquisition function to design an intelligent information acquisition system. The paper nicely combines the missing value problem with an active learning strategy to in an acquisition pipeline and demonstrate the effectiveness on several datasets.\n\nI have following comments/questions:\n\n1.  Does p(x_i | z) include parameters? How do these parameters be trained?\n\n2. Does sample from p(x_i | x_o) follow by sampling z from q(z|x_o) then sample x_i from p(x_i | z)? How to sample from p(x_\\phi | x_i, x_o) in Eq (7)?\n\n3. In Eq (9), it uses q(z_i|x_o), q(z_i | x_i, x_o),  q(z_i | x_i, x_o, x_\\phi) while in Eq (4) it only shows how to learn q(z|x_o). Does it need to learn multiple partial inference networks for all combination of i and \\phi ?\n\n4. The comparison with similar algorithms seems to be weak in the experiment section. RAND is random feature selection, and SING is global feature selection by using the proposed method. These comparison methods cannot provide enough information on how well the proposed methods performs. There are plenty of works in the area of “active feature acquisition” and also many works in feature selection dated back to Lasso which should be considered as comparison targets.\n\n5. In the “personalized” implementation of EDDI on each data instances, is the model trained independently for each data point or share some parameters across different data? If so, what are the shared parameters?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A nice application model but some unclear points", "review": "The paper proposes Partial VAE to handle missing data and a variable-wise active learning method. The model combines Partial VAE with the acquisition function to design an intelligent information acquisition system. The paper nicely combines the missing value problem with an active learning strategy to in an acquisition pipeline and demonstrate the effectiveness on several datasets.\n\nI have following comments/questions:\n\n1.  Does p(x_i | z) include parameters? How do these parameters be trained?\n\n2. Does sample from p(x_i | x_o) follow by sampling z from q(z|x_o) then sample x_i from p(x_i | z)? How to sample from p(x_\\phi | x_i, x_o) in Eq (7)?\n\n3. In Eq (9), it uses q(z_i|x_o), q(z_i | x_i, x_o),  q(z_i | x_i, x_o, x_\\phi) while in Eq (4) it only shows how to learn q(z|x_o). Does it need to learn multiple partial inference networks for all combination of i and \\phi ?\n\n4. The comparison with similar algorithms seems to be weak in the experiment section. RAND is random feature selection, and SING is global feature selection by using the proposed method. These comparison methods cannot provide enough information on how well the proposed methods performs. There are plenty of works in the area of “active feature acquisition” and also many works in feature selection dated back to Lasso which should be considered as comparison targets.\n\n5. In the “personalized” implementation of EDDI on each data instances, is the model trained independently for each data point or share some parameters across different data? If so, what are the shared parameters?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541135018807}], "openreview_url": "https://openreview.net/forum?id=HJl0jiRqtX", "arxiv_id": "1809.11142", "paper_pdf": "papers/HJl0jiRqtX.pdf", "paper_pdf_sha256": "4d9e5643c6d5f5bc86c5b9094b9a748e53f881cb5b97ede8a555e1fcd5ccfee3", "paper_pdf_bytes": 3555859, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/EDDI", "code_repository": "microsoft/EDDI", "code_commit": "835f39e13371a9727b8946e6fa5dcb21eaf28e14", "code_archive": "repos/HJl0jiRqtX.zip", "code_archive_sha256": "bff214ea9fdd435670872fc0a62c9551552fa7bea831480501a68460ca0d9a99", "code_archive_bytes": 38451, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 82, "github_languages": {"Python": 68526}, "github_archived": true, "github_pushed_at": "2023-06-12T18:56:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/eddi-efficient-dynamic-discovery-of-high"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJzMATlAZ", "year": 2018, "status": "rejected", "title": "Deep Continuous Clustering", "authors": ["Sohil Atul Shah", "Vladlen Koltun"], "authorids": ["sohilas@umd.edu", "vkoltun@gmail.com"], "authors_source": "OpenReview API", "abstract": "Clustering high-dimensional datasets is hard because interpoint distances become less informative in high-dimensional spaces. We present a clustering algorithm that performs nonlinear dimensionality reduction and clustering jointly. The data is embedded into a lower-dimensional space by a deep autoencoder. The autoencoder is optimized as part of the clustering process. The resulting network produces clustered data. The presented approach does not rely on prior knowledge of the number of ground-truth clusters. Joint nonlinear dimensionality reduction and clustering are formulated as optimization of a global continuous objective. We thus avoid discrete reconfigurations of the objective that characterize prior clustering algorithms. Experiments on datasets from multiple domains demonstrate that the presented algorithm outperforms state-of-the-art clustering schemes, including recent methods that use deep networks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H1ySNZVgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper436/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a clustering method in latent space. The work extends a previous approach (Shah & Koltun 2017) which employs a continuous relaxation of the clustering assignments. The proposed method is tested on several image and text data sets.\n\nHowever, the work has a number of problems and unclear points.\n\n1) There is no theoretical guarantee that RCC or DCC can give good clusterings. The second term in Eq. 2 will pull z's closer but it can also wrongly place data points from different clusters nearby.\n\n2) The method uses an autoencoder with elementwise least square loss. This is not suitable for data sets such as images and time series.\n\n3) Please elaborate \"redesending M-estimator\" in Section 2. Also, please explicitly write out what are rho_1 and rho_2 in the experiments.\n\n4) The method requires many extra hyperparameters lambda, delta_1, delta_2. Users have to set them by ad hoc heuristics.\n\n5) In each epoch, the method has to construct the graph G (the last paragraph in Page 4) over all z pairs.  This is expensive. The author didn't give any running time estimation in theory or in experiments.\n\n6) The experimental results are not convincing. For MNIST its best accuracy is only 0.912. Existing methods for this data set have achieve 0.97 accuracy. See for example [Ref1,Ref2,Ref3]. For RCV1, [Ref2] gives 0.54, but here it is only 0.495.\n\n7) Figure 1 gives a weird result. There is no known evidence that MNIST clusters intrinsically distribute like snakes. They must be some wrong artefacts introduced by the proposed method. Actually t-SNE with MNIST pixels is not bad at all. See [Ref4].\n\n8) It is unknown how to set the number of clusters in proposed method.\n\n\n[Ref1] Zhirong Yang, Tele Hao, Onur Dikmen, Xi Chen, Erkki Oja. Clustering by Nonnegative Matrix Factorization Using Graph Random Walk. In NIPS 2012.\n[Ref2] Xavier Bresson, Thomas Laurent, David Uminsky, James von Brecht. Multiclass Total Variation Clustering. In NIPS 2013.\n[Ref3] Zhirong Yang, Jukka Corander and Erkki Oja. Low-Rank Doubly Stochastic Matrix Decomposition for Cluster Analysis. Journal of Machine Learning Research, 17(187): 1-25, 2016.\n[Ref4] https://sites.google.com/site/neighborembedding/mnist\n\nConfidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A continuous relaxation for clustering with deep autoencoder", "rating": "3: Clear rejection", "review": "This paper presents a clustering method in latent space. The work extends a previous approach (Shah & Koltun 2017) which employs a continuous relaxation of the clustering assignments. The proposed method is tested on several image and text data sets.\n\nHowever, the work has a number of problems and unclear points.\n\n1) There is no theoretical guarantee that RCC or DCC can give good clusterings. The second term in Eq. 2 will pull z's closer but it can also wrongly place data points from different clusters nearby.\n\n2) The method uses an autoencoder with elementwise least square loss. This is not suitable for data sets such as images and time series.\n\n3) Please elaborate \"redesending M-estimator\" in Section 2. Also, please explicitly write out what are rho_1 and rho_2 in the experiments.\n\n4) The method requires many extra hyperparameters lambda, delta_1, delta_2. Users have to set them by ad hoc heuristics.\n\n5) In each epoch, the method has to construct the graph G (the last paragraph in Page 4) over all z pairs.  This is expensive. The author didn't give any running time estimation in theory or in experiments.\n\n6) The experimental results are not convincing. For MNIST its best accuracy is only 0.912. Existing methods for this data set have achieve 0.97 accuracy. See for example [Ref1,Ref2,Ref3]. For RCV1, [Ref2] gives 0.54, but here it is only 0.495.\n\n7) Figure 1 gives a weird result. There is no known evidence that MNIST clusters intrinsically distribute like snakes. They must be some wrong artefacts introduced by the proposed method. Actually t-SNE with MNIST pixels is not bad at all. See [Ref4].\n\n8) It is unknown how to set the number of clusters in proposed method.\n\n\n[Ref1] Zhirong Yang, Tele Hao, Onur Dikmen, Xi Chen, Erkki Oja. Clustering by Nonnegative Matrix Factorization Using Graph Random Walk. In NIPS 2012.\n[Ref2] Xavier Bresson, Thomas Laurent, David Uminsky, James von Brecht. Multiclass Total Variation Clustering. In NIPS 2013.\n[Ref3] Zhirong Yang, Jukka Corander and Erkki Oja. Low-Rank Doubly Stochastic Matrix Decomposition for Cluster Analysis. Journal of Machine Learning Research, 17(187): 1-25, 2016.\n[Ref4] https://sites.google.com/site/neighborembedding/mnist\n\nConfidence: 5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511425078733}, {"id": "SyqWgxzxf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper436/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "As authors stated, the proposed DCC is very similar to RCC-DR (Shah & Koltun, 2007). The only difference in (3) from RCC-DR is the decoding part, which is replaced by autoencoder instead of linear transformation used in RCC-DR. Authors claimed that there are three major differences. However, due to the highly nonconvex properties of both formulations, the last two differences hardly support the advantages of the proposed DCC comparing with RCC-DR because the solutions obtained by both optimization approaches are local solutions, unless authors can claim that the gradient-based solver is better than alternating approach in RCC-DR. Hence, DCC is just a simple extension of RCC-DR.\n\nIn Section 3.2, how does the optimization algorithm handle the equality constraints in (5)? It is unclear why the existing autoencoder solver can be used to solve (3) or (5). It seems that the first term in (5) corresponds to the objective of autoencoder, but the last two terms added lead to different objective with respect to variables y. It is better to clarify the correctness of the optimization algorithm.\n\nAuthors claimed that the proposed method avoid discrete reconfiguration of the objective that characterize prior clustering algorithms, and it does not rely on a priori knowledge of the number of ground-truth clusters. However, it seems not true since the graph construction at every epoch depends on the initial parameter delta_2 and the graph is constructed such that f_{i,j}=1 if distance is less than delta_2. As a result, delta_2 is a fixed threshold for graph construction, so it is indirectly related to the number of clusters generated. In the experiments, authors set it as the mean of the bottom 1% of the pairwise distances in E at initialization, and clustering assignment is given by connected component in the last graph. This parameter might be sensitive to the final results.\n\nMany terms in the paper are not well explained. For example, in (1), theta are treated as parameters to optimize, but what is the theta used for? Does the Omega related to encoder and decoder of the parameters in autoencoder. What is the scaled Geman-McClure function? Any reference? Why should this estimator be used?\n\nFrom the visualization results in Figure 1, it is interesting to see that K-means++ can achieve much better results on the space learned by DCC than that by SDAE from Table 2. In Figure 1, the embedding by SDAE (Figure 1(b)) seems more suitable for kmeans-like algorithm than DCC (Figure 1(c)). That is the reason why connected component is used for cluster assignment in DCC, not kmeans. The results between Table 2 and Figure 1 might be interesting to investigate. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Authors of this paper presented a clustering algorithm by jointly solving deep autoencoder and clustering as a global continuous objective. Experiments demonstrate better results than state-of-the-art clustering schemas.", "rating": "6: Marginally above acceptance threshold", "review": "As authors stated, the proposed DCC is very similar to RCC-DR (Shah & Koltun, 2007). The only difference in (3) from RCC-DR is the decoding part, which is replaced by autoencoder instead of linear transformation used in RCC-DR. Authors claimed that there are three major differences. However, due to the highly nonconvex properties of both formulations, the last two differences hardly support the advantages of the proposed DCC comparing with RCC-DR because the solutions obtained by both optimization approaches are local solutions, unless authors can claim that the gradient-based solver is better than alternating approach in RCC-DR. Hence, DCC is just a simple extension of RCC-DR.\n\nIn Section 3.2, how does the optimization algorithm handle the equality constraints in (5)? It is unclear why the existing autoencoder solver can be used to solve (3) or (5). It seems that the first term in (5) corresponds to the objective of autoencoder, but the last two terms added lead to different objective with respect to variables y. It is better to clarify the correctness of the optimization algorithm.\n\nAuthors claimed that the proposed method avoid discrete reconfiguration of the objective that characterize prior clustering algorithms, and it does not rely on a priori knowledge of the number of ground-truth clusters. However, it seems not true since the graph construction at every epoch depends on the initial parameter delta_2 and the graph is constructed such that f_{i,j}=1 if distance is less than delta_2. As a result, delta_2 is a fixed threshold for graph construction, so it is indirectly related to the number of clusters generated. In the experiments, authors set it as the mean of the bottom 1% of the pairwise distances in E at initialization, and clustering assignment is given by connected component in the last graph. This parameter might be sensitive to the final results.\n\nMany terms in the paper are not well explained. For example, in (1), theta are treated as parameters to optimize, but what is the theta used for? Does the Omega related to encoder and decoder of the parameters in autoencoder. What is the scaled Geman-McClure function? Any reference? Why should this estimator be used?\n\nFrom the visualization results in Figure 1, it is interesting to see that K-means++ can achieve much better results on the space learned by DCC than that by SDAE from Table 2. In Figure 1, the embedding by SDAE (Figure 1(b)) seems more suitable for kmeans-like algorithm than DCC (Figure 1(c)). That is the reason why connected component is used for cluster assignment in DCC, not kmeans. The results between Table 2 and Figure 1 might be interesting to investigate. \n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511288834279}, {"id": "HJ90m_PeG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper436/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors proposed a new clustering algorithm named deep continuous clustering (DCC) that integrates autoencoder into continuous clustering. As a variant of  continuous clustering (RCC), DCC formed a global continuous objective for joint nonlinear dimensionality reduction and clustering. The objective can be directly optimized using SGD like method. Extensive experiments on image and document datasets show the effectiveness of DCC. However, part of experiments are not comprehensive enough. \n\nThe idea of integrating autoencoder with continuous clustering is novel, and the optimization part is quite different. The trick used in the paper (sampling edges but not samples) looks interesting and seems to be effective. \n\nIn the following, there are some detailed comments:\n1. The paper is well written and easy to follow, except the definition of Geman-McClure function is missing. It is difficult to follow Eq. (6) and (7).\n2. Compare DCC to RCC, the pros and cons are obvious. DCC does improve the performance of clustering with the cost of losing robustness. DCC is more sensitive to the hyper-parameters, especially embedding dimensionality d. With a wrong d DCC performs worse than RCC on MNIST and similar on Reuters. Since clustering is one unsupervised learning task. The author should consider heuristics to determine the hyper-parameters. This will increase the usability of the proposed method.\n3. However, the comparison to the DL based partners are not comprehensive enough, especially JULE and DEPICT on image clustering. Firstly, the authors only reported AMI and ACC, but not NMI that is reported in JULE. For a fair comparison, NMI results should be included. Secondly, the reported results do not agree with the one in original publication. For example, JULE reported ACC of 0.964 and 0.684 on MNIST and YTF. However, in the appendix the numbers are 0.800 and 0.342 respectively. Compared to the reported number in JULE paper, DCC is not significantly better.\n\nIn general, the paper is interesting and proposed method seems to be promising. I would vote for accept if my concerns can be addressed.\n\nThe author's respond address part of my concerns, so I have adjusted my rating.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors proposed a new clustering algorithm named deep continuous clustering (DCC) that integrates autoencoder into continuous clustering. The paper is interesting and could be improved by increasing the usability of the method.", "rating": "7: Good paper, accept", "review": "The authors proposed a new clustering algorithm named deep continuous clustering (DCC) that integrates autoencoder into continuous clustering. As a variant of  continuous clustering (RCC), DCC formed a global continuous objective for joint nonlinear dimensionality reduction and clustering. The objective can be directly optimized using SGD like method. Extensive experiments on image and document datasets show the effectiveness of DCC. However, part of experiments are not comprehensive enough. \n\nThe idea of integrating autoencoder with continuous clustering is novel, and the optimization part is quite different. The trick used in the paper (sampling edges but not samples) looks interesting and seems to be effective. \n\nIn the following, there are some detailed comments:\n1. The paper is well written and easy to follow, except the definition of Geman-McClure function is missing. It is difficult to follow Eq. (6) and (7).\n2. Compare DCC to RCC, the pros and cons are obvious. DCC does improve the performance of clustering with the cost of losing robustness. DCC is more sensitive to the hyper-parameters, especially embedding dimensionality d. With a wrong d DCC performs worse than RCC on MNIST and similar on Reuters. Since clustering is one unsupervised learning task. The author should consider heuristics to determine the hyper-parameters. This will increase the usability of the proposed method.\n3. However, the comparison to the DL based partners are not comprehensive enough, especially JULE and DEPICT on image clustering. Firstly, the authors only reported AMI and ACC, but not NMI that is reported in JULE. For a fair comparison, NMI results should be included. Secondly, the reported results do not agree with the one in original publication. For example, JULE reported ACC of 0.964 and 0.684 on MNIST and YTF. However, in the appendix the numbers are 0.800 and 0.342 respectively. Compared to the reported number in JULE paper, DCC is not significantly better.\n\nIn general, the paper is interesting and proposed method seems to be promising. I would vote for accept if my concerns can be addressed.\n\nThe author's respond address part of my concerns, so I have adjusted my rating.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511650258245}], "openreview_url": "https://openreview.net/forum?id=SJzMATlAZ", "arxiv_id": "1803.01449", "paper_pdf": "papers/SJzMATlAZ.pdf", "paper_pdf_sha256": "f7c1c1ff97a746db6ca181809bf69b5c4c78d4f5abd5d6c2dc2d7494ac699021", "paper_pdf_bytes": 991706, "paper_pdf_source": "openreview", "code_url": "https://github.com/shahsohil/DCC", "code_repository": "shahsohil/DCC", "code_commit": "775e07e17547df2804fd9eead3f1b128d303b577", "code_archive": "repos/SJzMATlAZ.zip", "code_archive_sha256": "74c957409776942d41486910683790491c7a3ec12fb53b4c59bdf5666bc04029", "code_archive_bytes": 35209, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 52, "github_languages": {"Python": 77928}, "github_archived": false, "github_pushed_at": "2021-07-14T09:52:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-continuous-clustering"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PZQHihJlfm", "year": 2026, "status": "rejected", "title": "Next-Scale Autoregressive Models are Zero-Shot Single-Image Object View Synthesizers", "authors": ["Shiran Yuan", "Hao Zhao"], "authorids": ["~Shiran_Yuan1", "~Hao_Zhao1"], "authors_source": "OpenReview API", "abstract": "Learning to synthesize novel views without explicit 3D representations or hand-crafted 3D inductive bias has recently gained attention: it is simpler, more formally direct, and better aligned with the lesson that scalable learning paradigms with less assumptions built into architectural design (e.g., regarding geometry) often win. However, the current dominant solutions are diffusion-based, which typically suffer from problems like slow inference. We introduce ArchonView, the first autoregressive model for zero-shot single-image, object-centric novel view synthesis (NVS), achieving substantially faster inference, higher accuracies, and notably not relying on fine-tuning of 2D generative checkpoints (challenging the common assumption that 2D priors are required in diffusion-based NVS). We design innovative methods of both global and local conditioning to suit characteristics of the NVS task. Crucially, a naïve application of next-scale autoregression fails; we identify two design choices that unlock performance: local conditioning pre-filling, and removing global AdaLN at the classifier head. ArchonView delivers state-of-the-art zero-shot results across six standard benchmarks (GSO, ABO, OmniObject3D, RTMV, NeRF-Synthetic, ShapeNet), while being several times faster than diffusion baselines (e.g., 0.22s v.s. 1.7–1.8s per view at matched parameter count). It consistently improves synthesis accuracy, and scales predictably with both model size (135M–2B) and data size, exhibiting clear scaling-law-like trends. Our findings suggest a paradigm shift and challenge an existing assumption: first, for object-centric NVS, next-scale autoregression can be faster, simpler, and more accurate than diffusion; and second, priors obtained from fine-tuning 2D-pretrained models may not be necessary for generative NVS. Our code is open-sourced at https://anonymous.4open.science/r/ArchonView/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "3pJsoMO7dZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3071/Reviewer_V295"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper adopts a next-scale autoregressive model—specifically VAR—for single-image novel view synthesis. Relative to diffusion, the VAR backbone offers much faster per-view generation (no multi-step denoising), a simpler inference pipeline, competitive or better fidelity, and predictable scaling with model and data size. To enable NVS, the authors augment VAR with: (i) a global “posed” start token that fuses CLIP semantics with the target relative pose; (ii) local multi-scale “prefilling,” which prepends the source image’s VQVAE tokens at every scale under a causal/block-triangular mask to guide generation; and (iii) an architectural fix—removing AdaLN at the classifier head—to preserve local correspondences. Experiments on six object-centric benchmarks (GSO, ABO, OmniObject3D, RTMV, NeRF-Synthetic, ShapeNet) in a zero-shot setting show state-of-the-art accuracy with several-times faster inference than diffusion baselines.", "review_text": "The paper adopts a next-scale autoregressive model—specifically VAR—for single-image novel view synthesis. Relative to diffusion, the VAR backbone offers much faster per-view generation (no multi-step denoising), a simpler inference pipeline, competitive or better fidelity, and predictable scaling with model and data size. To enable NVS, the authors augment VAR with: (i) a global “posed” start token that fuses CLIP semantics with the target relative pose; (ii) local multi-scale “prefilling,” which prepends the source image’s VQVAE tokens at every scale under a causal/block-triangular mask to guide generation; and (iii) an architectural fix—removing AdaLN at the classifier head—to preserve local correspondences. Experiments on six object-centric benchmarks (GSO, ABO, OmniObject3D, RTMV, NeRF-Synthetic, ShapeNet) in a zero-shot setting show state-of-the-art accuracy with several-times faster inference than diffusion baselines.", "strengths": "* Originality: To the best of my knowledge, this is the first work that adopts VAR for single-image novel view synthesis.\n\n* Quality:\n\n  * Both the qualitative and quantitative results are significant, demonstrated across multiple object-centric benchmarks.\n  * The major components—global pose conditioning, local prefilling, and classifier-head AdaLN removal—are well supported by aligned ablation studies.\n  * Efficiency and scaling capabilities are also demonstrated through experiments.\n\n* Clarity: \n  - The paper is in general well structured.\n  - The source code is given.\n\n* Significance:\n\n  * The proposed solution demonstrates the potential of VAR for novel view synthesis in both quality and efficiency, and can inspire follow-up research on this alternative (and potentially superior) model paradigm.\n  * The solution is backed by actionable insights, which could transfer to areas beyond novel view synthesis.", "weaknesses": "* While the solution itself is clear, the motivation and insights behind the design need more elaboration:\n\n  * Local attention. Provide a deeper investigation of how attention behaves in this model—e.g., whether generated patches attend to the intended input patches at the desired locations. Concretely, add attention visualizations across scales, quantify attention mass within pose-consistent neighborhoods, and report correspondence accuracy vs. pose gap.\n  * Attention design (lines 249–265). Since VAR also uses prefilling techniques, clarify what is additionally novel here. Distinguish your contribution from VAR via ablations isolating token prepending vs. causal/block-triangular masking vs. cross-attention, and report any compute/latency trade-offs introduced by your variant.\n  * AdaLN claim (lines 283–296). The claim is somewhat ambiguous and currently supported only by empirical results. Please add a more theoretical or mechanistic explanation (with a simple formulation, if possible), and consider alternatives (e.g., scaled/gated/partial AdaLN or layer-wise removal) with diagnostics such as layer-wise gradients/feature norms to substantiate the hypothesis.\n\n* Beyond single-view fidelity, the paper should investigate cross-view consistency of synthesized views. Recommend a multi-target protocol (e.g., 8–16 target poses per source) and report consistency metrics (cycle/epipolar consistency, normal/depth agreement, or reconstruction consistency via a downstream NeRF fit), alongside qualitative failure cases.", "questions": "- First, please refer to the weaknesses.\n- Second, I would like to ask whether the proposed solution in this paper can be extended to text to 3D model (multi-view images) generation with minimum efforts, if so what are the potential efforts.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper adopts a next-scale autoregressive model—specifically VAR—for single-image novel view synthesis. Relative to diffusion, the VAR backbone offers much faster per-view generation (no multi-step denoising), a simpler inference pipeline, competitive or better fidelity, and predictable scaling with model and data size. To enable NVS, the authors augment VAR with: (i) a global “posed” start token that fuses CLIP semantics with the target relative pose; (ii) local multi-scale “prefilling,” which prepends the source image’s VQVAE tokens at every scale under a causal/block-triangular mask to guide generation; and (iii) an architectural fix—removing AdaLN at the classifier head—to preserve local correspondences. Experiments on six object-centric benchmarks (GSO, ABO, OmniObject3D, RTMV, NeRF-Synthetic, ShapeNet) in a zero-shot setting show state-of-the-art accuracy with several-times faster inference than diffusion baselines.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "* Originality: To the best of my knowledge, this is the first work that adopts VAR for single-image novel view synthesis.\n\n* Quality:\n\n  * Both the qualitative and quantitative results are significant, demonstrated across multiple object-centric benchmarks.\n  * The major components—global pose conditioning, local prefilling, and classifier-head AdaLN removal—are well supported by aligned ablation studies.\n  * Efficiency and scaling capabilities are also demonstrated through experiments.\n\n* Clarity: \n  - The paper is in general well structured.\n  - The source code is given.\n\n* Significance:\n\n  * The proposed solution demonstrates the potential of VAR for novel view synthesis in both quality and efficiency, and can inspire follow-up research on this alternative (and potentially superior) model paradigm.\n  * The solution is backed by actionable insights, which could transfer to areas beyond novel view synthesis.", "weaknesses": "* While the solution itself is clear, the motivation and insights behind the design need more elaboration:\n\n  * Local attention. Provide a deeper investigation of how attention behaves in this model—e.g., whether generated patches attend to the intended input patches at the desired locations. Concretely, add attention visualizations across scales, quantify attention mass within pose-consistent neighborhoods, and report correspondence accuracy vs. pose gap.\n  * Attention design (lines 249–265). Since VAR also uses prefilling techniques, clarify what is additionally novel here. Distinguish your contribution from VAR via ablations isolating token prepending vs. causal/block-triangular masking vs. cross-attention, and report any compute/latency trade-offs introduced by your variant.\n  * AdaLN claim (lines 283–296). The claim is somewhat ambiguous and currently supported only by empirical results. Please add a more theoretical or mechanistic explanation (with a simple formulation, if possible), and consider alternatives (e.g., scaled/gated/partial AdaLN or layer-wise removal) with diagnostics such as layer-wise gradients/feature norms to substantiate the hypothesis.\n\n* Beyond single-view fidelity, the paper should investigate cross-view consistency of synthesized views. Recommend a multi-target protocol (e.g., 8–16 target poses per source) and report consistency metrics (cycle/epipolar consistency, normal/depth agreement, or reconstruction consistency via a downstream NeRF fit), alongside qualitative failure cases.", "questions": "- First, please refer to the weaknesses.\n- Second, I would like to ask whether the proposed solution in this paper can be extended to text to 3D model (multi-view images) generation with minimum efforts, if so what are the potential efforts.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762013937092}, {"id": "XHmUoGFaCe", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3071/Reviewer_Vfwi"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper address generative novel view synthesis with one single input image. This problem is highly proabilistic, and the author adopted the visual autoregessive model based approach for such generative task.  The author used the residual VQVAE from the original VAR paper, then  trained the generative model (next-scale prediction with block-wise causal transformer) from scratch on Objectverse dataset.  To improve results,  the author also applied CFG on the pose embedding.  Also the author tweakes the architecure a bit, mostly by removing the adaptive LN layer befire the classifier head.  The author compared their methods with a few diffusion based baseline, including Zero 1-to-3, Zero 123-XL, EscherNet.  These baselines finetuens a pretrained image diffusion model on Objectverse for generative novel view synthesis. The author shows improves results, e.g. on GSO (17.44 PSNR vs. 16.77).  The author showed ablation study on its architecture design, showed quite clear improvements.", "review_text": "This paper address generative novel view synthesis with one single input image. This problem is highly proabilistic, and the author adopted the visual autoregessive model based approach for such generative task.  The author used the residual VQVAE from the original VAR paper, then  trained the generative model (next-scale prediction with block-wise causal transformer) from scratch on Objectverse dataset.  To improve results,  the author also applied CFG on the pose embedding.  Also the author tweakes the architecure a bit, mostly by removing the adaptive LN layer befire the classifier head.  The author compared their methods with a few diffusion based baseline, including Zero 1-to-3, Zero 123-XL, EscherNet.  These baselines finetuens a pretrained image diffusion model on Objectverse for generative novel view synthesis. The author shows improves results, e.g. on GSO (17.44 PSNR vs. 16.77).  The author showed ablation study on its architecture design, showed quite clear improvements.", "strengths": "1. The method is overall relative new for the field of novel view synthesis, and the author made several archteicture change that significantly improves the results. Which is quite solid. \n2. The author submitted code at the time of submission, which is a good practice.", "weaknesses": "I don't agree with a few of arguments (or maybe just wordings) from the author\n\nTraining from scratch is cool and quite impressive, but I don't agree that these diffusion based approach relying on pretrained checkpoints is a big drawbacks.  (as argued in the second paragraph of the introduction).  Also, this does not seem to be the problem of diffusion model. Diffusion model is just an algorithm for generative modelling,  previous baseline relies on pretraining and then finetuning does not mean that training a well-designed diffusion model for NVS task from scratch would not work.  (e.g. the first sentence in the 3rd paragraph of the intro goes:  such downside of diffusion calls for xxxx).  \n\nThe point is you don't need to over criticize the diffusion based approaches in your paper. VAR for NVS is already quite interesting and impressive. \n\n\n\nAnother weakness of the paper is that the metric is so low to be indicative. Table 1 shows PSNR of 10-19.42.  For PSNR in this range, it's really hard to tell if the improvements are useful. I would suggest using a subset with nearby camera viewpoints (larger view overlap) for evaluation.  For multiview PSNR on GSO, people already got PSNR over 30.", "questions": "I would be curious, if the author train a diffusion baseline from scratch with similar recipes, how would that perform? And also how would that perform if the author adds architecure improvements like global and local conditioning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper address generative novel view synthesis with one single input image. This problem is highly proabilistic, and the author adopted the visual autoregessive model based approach for such generative task.  The author used the residual VQVAE from the original VAR paper, then  trained the generative model (next-scale prediction with block-wise causal transformer) from scratch on Objectverse dataset.  To improve results,  the author also applied CFG on the pose embedding.  Also the author tweakes the architecure a bit, mostly by removing the adaptive LN layer befire the classifier head.  The author compared their methods with a few diffusion based baseline, including Zero 1-to-3, Zero 123-XL, EscherNet.  These baselines finetuens a pretrained image diffusion model on Objectverse for generative novel view synthesis. The author shows improves results, e.g. on GSO (17.44 PSNR vs. 16.77).  The author showed ablation study on its architecture design, showed quite clear improvements.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The method is overall relative new for the field of novel view synthesis, and the author made several archteicture change that significantly improves the results. Which is quite solid. \n2. The author submitted code at the time of submission, which is a good practice.", "weaknesses": "I don't agree with a few of arguments (or maybe just wordings) from the author\n\nTraining from scratch is cool and quite impressive, but I don't agree that these diffusion based approach relying on pretrained checkpoints is a big drawbacks.  (as argued in the second paragraph of the introduction).  Also, this does not seem to be the problem of diffusion model. Diffusion model is just an algorithm for generative modelling,  previous baseline relies on pretraining and then finetuning does not mean that training a well-designed diffusion model for NVS task from scratch would not work.  (e.g. the first sentence in the 3rd paragraph of the intro goes:  such downside of diffusion calls for xxxx).  \n\nThe point is you don't need to over criticize the diffusion based approaches in your paper. VAR for NVS is already quite interesting and impressive. \n\n\n\nAnother weakness of the paper is that the metric is so low to be indicative. Table 1 shows PSNR of 10-19.42.  For PSNR in this range, it's really hard to tell if the improvements are useful. I would suggest using a subset with nearby camera viewpoints (larger view overlap) for evaluation.  For multiview PSNR on GSO, people already got PSNR over 30.", "questions": "I would be curious, if the author train a diffusion baseline from scratch with similar recipes, how would that perform? And also how would that perform if the author adds architecure improvements like global and local conditioning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761974526459}, {"id": "DbVUOXrw5v", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3071/Reviewer_uqJT"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper introduces ArchonView, a model for zero-shot, single-image novel view synthesis based on a next-scale autoregressive paradigm. This paper conducts an in-depth investigation into efficient and high-quality novel view synthesis within the VAR framework. For example, since the conditions for novel view synthesis do not satisfy pixel alignment, the authors analyze different conditioning methods, including Prefilling, Causal Conditioning, and Cross-Attention. Their results show that simple prefilling significantly outperforms the other two approaches.", "review_text": "This paper introduces ArchonView, a model for zero-shot, single-image novel view synthesis based on a next-scale autoregressive paradigm. This paper conducts an in-depth investigation into efficient and high-quality novel view synthesis within the VAR framework. For example, since the conditions for novel view synthesis do not satisfy pixel alignment, the authors analyze different conditioning methods, including Prefilling, Causal Conditioning, and Cross-Attention. Their results show that simple prefilling significantly outperforms the other two approaches.", "strengths": "1. For the task of novel view synthesis, this paper provides a thorough analysis of various design choices and conditioning strategies within the VAR framework. The discussion covers conditioning methods, insights from The Devil is in the Classifier Head, Semantic Global Pose Conditioning, and Multi-Scale Local Conditioning.\n2. The paper demonstrates the scalability of the proposed method through extensive experiments. By systematically increasing both the model size and the dataset size, the authors plot performance curves that clearly illustrate the improvements achieved.\n3. For the task of novel view synthesis, the proposed method achieves faster and better performance compared to baseline approaches.", "weaknesses": "1. The baselines compared in this paper are relatively outdated, and many newer baselines have been introduced since. It is necessary to discuss and compare the proposed method with these more recent approaches.\n2. The statements in the discussion section are somewhat confusing. The paper mentions, “In contrast, we only use the ‘fine-tuning’ dataset of previous works (with 800k 3D objects) and achieved significantly superior results.” However, it appears that the method also relies on a pretrained checkpoint for fine-tuning?", "questions": "The paper offers limited discussion on the practical applications of the proposed method, such as whether it is intended for 3D reconstruction or 3D generation, or how it could be integrated into scene understanding or combined with existing VLMs. These aspects warrant further exploration and discussion.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ArchonView, a model for zero-shot, single-image novel view synthesis based on a next-scale autoregressive paradigm. This paper conducts an in-depth investigation into efficient and high-quality novel view synthesis within the VAR framework. For example, since the conditions for novel view synthesis do not satisfy pixel alignment, the authors analyze different conditioning methods, including Prefilling, Causal Conditioning, and Cross-Attention. Their results show that simple prefilling significantly outperforms the other two approaches.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. For the task of novel view synthesis, this paper provides a thorough analysis of various design choices and conditioning strategies within the VAR framework. The discussion covers conditioning methods, insights from The Devil is in the Classifier Head, Semantic Global Pose Conditioning, and Multi-Scale Local Conditioning.\n2. The paper demonstrates the scalability of the proposed method through extensive experiments. By systematically increasing both the model size and the dataset size, the authors plot performance curves that clearly illustrate the improvements achieved.\n3. For the task of novel view synthesis, the proposed method achieves faster and better performance compared to baseline approaches.", "weaknesses": "1. The baselines compared in this paper are relatively outdated, and many newer baselines have been introduced since. It is necessary to discuss and compare the proposed method with these more recent approaches.\n2. The statements in the discussion section are somewhat confusing. The paper mentions, “In contrast, we only use the ‘fine-tuning’ dataset of previous works (with 800k 3D objects) and achieved significantly superior results.” However, it appears that the method also relies on a pretrained checkpoint for fine-tuning?", "questions": "The paper offers limited discussion on the practical applications of the proposed method, such as whether it is intended for 3D reconstruction or 3D generation, or how it could be integrated into scene understanding or combined with existing VLMs. These aspects warrant further exploration and discussion.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761813278153}, {"id": "DfbEuuoKlw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3071/Reviewer_nWwK"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper extends the next-scale autoregressive generation paradigm from 2D image modeling to zero-shot 3D novel view synthesis (NVS). The proposed ArchonView model autoregressively predicts higher-resolution scales conditioned on a single input view. It achieves competitive or superior results compared to diffusion-based baselines such as Zero-1-to-3 and EscherNet, with significantly faster inference and no reliance on pretrained 2D diffusion models.", "review_text": "The paper extends the next-scale autoregressive generation paradigm from 2D image modeling to zero-shot 3D novel view synthesis (NVS). The proposed ArchonView model autoregressively predicts higher-resolution scales conditioned on a single input view. It achieves competitive or superior results compared to diffusion-based baselines such as Zero-1-to-3 and EscherNet, with significantly faster inference and no reliance on pretrained 2D diffusion models.", "strengths": "Interesting and timely exploration of applying next-scale generation to 3D NVS, challenging the diffusion-based dominance in this field.\n\nEmpirically strong results on multiple object-level benchmarks, with improved efficiency and competitive quality.\n\nThe overall design remains simple yet effective, providing a clean baseline for future autoregressive extensions in NVS.", "weaknesses": "- Conceptual clarity on “next-scale” in 3D\n\nThe paper straightforwardly brings the 2D next-scale paradigm into 3D, but it remains unclear what the “scale” represents in this context.\nIn 3D, is the next-scale process simply applied independently within each view, or could there exist a more native 3D notion of scale that aligns across multiple views?\nThe current adaptation feels like a direct transplant from 2D without strong 3D intuition or justification.\nSome conceptual discussion or visualization would help clarify how “next-scale” manifests in multi-view geometry.\n\n- Limited supporting of view numbers\n\nEscherNet supports flexible N-to-M view synthesis, while this work only shows 1-to-1 generation.\nCan this method extended to flexible multi-view inputs or outputs? If so, how would the computational complexity scale with N and M?\nThis discussion is important to position the method within the broader NVS landscape.\n\n- Result discrepancies\n\nThe GSO performance reported for Zero123 differs from the original paper (Table 1).\nPlease clarify whether this comes from dataset versions, evaluation protocols, or re-implementation differences.\n\n- Lack of visualization for 3D consistency\n\nThe paper mainly presents single novel views. For a model claiming 3D-aware generation, visualizing the synthesized 360° trajectory from a single input would be essential to evaluate consistency and geometry preservation.\nWithout this, it remains unclear how coherent the generated views are across different poses.\n\n- Lack of diversity in generation\n\nDoes the model produce diverse outputs given the same input view and target pose, as seen in diffusion-based models like Zero123 (e.g., Fig. 8 in their paper)? \n\n- Compute scale and more diverse examples\n\nThe model is trained on 32 H200 GPUs, much larger than prior works like EscherNet or Zero123 (8 A100 GPUs).\nThe paper doesn't show any diverse qualitative “in-the-wild” examples. The other baselines are trained under the same data scale but show very impressive diverse examples.\n\n\n- Reuse of pretrained components\n\nThe model reuses the pretrained next-scale VQVAE but trains the transformer backbone from scratch.\nWhy not or can't also reuse or partially initialize the pretrained transformer with zero-init?\nWill this improve generalization?", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper extends the next-scale autoregressive generation paradigm from 2D image modeling to zero-shot 3D novel view synthesis (NVS). The proposed ArchonView model autoregressively predicts higher-resolution scales conditioned on a single input view. It achieves competitive or superior results compared to diffusion-based baselines such as Zero-1-to-3 and EscherNet, with significantly faster inference and no reliance on pretrained 2D diffusion models.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Interesting and timely exploration of applying next-scale generation to 3D NVS, challenging the diffusion-based dominance in this field.\n\nEmpirically strong results on multiple object-level benchmarks, with improved efficiency and competitive quality.\n\nThe overall design remains simple yet effective, providing a clean baseline for future autoregressive extensions in NVS.", "weaknesses": "- Conceptual clarity on “next-scale” in 3D\n\nThe paper straightforwardly brings the 2D next-scale paradigm into 3D, but it remains unclear what the “scale” represents in this context.\nIn 3D, is the next-scale process simply applied independently within each view, or could there exist a more native 3D notion of scale that aligns across multiple views?\nThe current adaptation feels like a direct transplant from 2D without strong 3D intuition or justification.\nSome conceptual discussion or visualization would help clarify how “next-scale” manifests in multi-view geometry.\n\n- Limited supporting of view numbers\n\nEscherNet supports flexible N-to-M view synthesis, while this work only shows 1-to-1 generation.\nCan this method extended to flexible multi-view inputs or outputs? If so, how would the computational complexity scale with N and M?\nThis discussion is important to position the method within the broader NVS landscape.\n\n- Result discrepancies\n\nThe GSO performance reported for Zero123 differs from the original paper (Table 1).\nPlease clarify whether this comes from dataset versions, evaluation protocols, or re-implementation differences.\n\n- Lack of visualization for 3D consistency\n\nThe paper mainly presents single novel views. For a model claiming 3D-aware generation, visualizing the synthesized 360° trajectory from a single input would be essential to evaluate consistency and geometry preservation.\nWithout this, it remains unclear how coherent the generated views are across different poses.\n\n- Lack of diversity in generation\n\nDoes the model produce diverse outputs given the same input view and target pose, as seen in diffusion-based models like Zero123 (e.g., Fig. 8 in their paper)? \n\n- Compute scale and more diverse examples\n\nThe model is trained on 32 H200 GPUs, much larger than prior works like EscherNet or Zero123 (8 A100 GPUs).\nThe paper doesn't show any diverse qualitative “in-the-wild” examples. The other baselines are trained under the same data scale but show very impressive diverse examples.\n\n\n- Reuse of pretrained components\n\nThe model reuses the pretrained next-scale VQVAE but trains the transformer backbone from scratch.\nWhy not or can't also reuse or partially initialize the pretrained transformer with zero-init?\nWill this improve generalization?", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761785248618}], "openreview_url": "https://openreview.net/forum?id=PZQHihJlfm", "arxiv_id": "2503.13588", "paper_pdf": "papers/PZQHihJlfm.pdf", "paper_pdf_sha256": "d0ae1551e84c4a5257f6aed353bf338866ade9f622998a2f9f330faee372a698", "paper_pdf_bytes": 4537378, "paper_pdf_source": "openreview", "code_url": "https://github.com/Shiran-Yuan/ArchonView", "code_repository": "Shiran-Yuan/ArchonView", "code_commit": "07be09646242b4625653c9d2ff1ec14330415655", "code_archive": "repos/PZQHihJlfm.zip", "code_archive_sha256": "3616eeed0dee05f406ada0d19b40b12ddc1ba5d2a28fecccb6948fffb737a2c7", "code_archive_bytes": 47624, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 74, "github_languages": {"Python": 141757}, "github_archived": false, "github_pushed_at": "2025-03-19T01:41:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/next-scale-autoregressive-models-are-zero"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3SMBSTG3qN", "year": 2025, "status": "rejected", "title": "Beyond CVaR: Leveraging Static Spectral Risk Measures for Enhanced Decision-Making in Distributional Reinforcement Learning", "authors": ["Mehrdad Moghimi", "Hyejin Ku"], "authorids": ["~Mehrdad_Moghimi1", "~Hyejin_Ku1"], "authors_source": "OpenReview API", "abstract": "In domains such as finance, healthcare, and robotics, managing worst-case scenarios is critical, as failure to do so can lead to catastrophic outcomes. Distributional Reinforcement Learning (DRL) provides a natural framework to incorporate risk sensitivity into decision-making processes. However, existing approaches face two key limitations: (1) the use of fixed risk measures at each decision step often results in overly conservative policies, and (2) the interpretation and theoretical properties of the learned policies remain unclear. While optimizing a static risk measure addresses these issues, its use in the DRL framework has been limited to the simple static CVaR risk measure. In this paper, we present a novel DRL algorithm with convergence guarantees that optimizes for a broader class of static Spectral Risk Measures (SRM). Additionally, we provide a clear interpretation of the learned policy by leveraging the distribution of returns in DRL and the decomposition of static coherent risk measures. Extensive experiments demonstrate that our model learns policies aligned with the SRM objective, and outperforms existing risk-neutral and risk-sensitive DRL models in various settings.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "3FC7IB4stR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10312/Reviewer_tbS4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This work studies the problem of incorporating static spectral risk measures (SRM into Distributional Reinforcement Learning (DRL) to enable more flexible and interpretable risk-sensitive decision-making. Unlike conventional Conditional Value-at-Risk (CVaR), SRMs offer a spectrum of risk preferences, allowing for more flexible risk-sensitive policies. The authors argue that using SRMs in DRL enables more flexible and interpretable policies, as SRMs allow for a spectrum of risk preferences rather than a fixed measure like CVaR. The authors propose an iterative DRL algorithm that utilizes a two-stage optimization process to optimize SRMs. They provide theoretical guarantees, proving convergence and characterizing the temporal decomposition of SRMs within the DRL framework. This decomposition enhances interpretability, as it captures how risk preferences evolve over time. The algorithm’s effectiveness is demonstrated through extensive numerical studies across four example environments, where it outperforms several baseline models, highlighting its potential for real-world applications.", "review_text": "This work studies the problem of incorporating static spectral risk measures (SRM into Distributional Reinforcement Learning (DRL) to enable more flexible and interpretable risk-sensitive decision-making. Unlike conventional Conditional Value-at-Risk (CVaR), SRMs offer a spectrum of risk preferences, allowing for more flexible risk-sensitive policies. The authors argue that using SRMs in DRL enables more flexible and interpretable policies, as SRMs allow for a spectrum of risk preferences rather than a fixed measure like CVaR. The authors propose an iterative DRL algorithm that utilizes a two-stage optimization process to optimize SRMs. They provide theoretical guarantees, proving convergence and characterizing the temporal decomposition of SRMs within the DRL framework. This decomposition enhances interpretability, as it captures how risk preferences evolve over time. The algorithm’s effectiveness is demonstrated through extensive numerical studies across four example environments, where it outperforms several baseline models, highlighting its potential for real-world applications.", "strengths": "1.\tThis paper is easy to follow and well-organized. \n2.\tThe theoretical analyses throughout the paper are technically sound and comprehensive, providing strong support for the proposed method.\n3.\tThe authors preform thorough numerical studies across four examples. The proposed algorithm outperforms several baselines, highlighting its potential for real-world applications.", "weaknesses": "See Question section", "questions": "1.\tIn Table 1, when the objective is CVaR(0.1), why does QR-SRM with $\\alpha=0.1$ not achieve the highest value? Could the authors clarify the reasons  influencing this outcome?\n2.\tThe authors introduce the decomposition theorem for SRMs (Theorem 2). Could they use one of the four examples to illustrate how this theorem applies in a practical scenario? This would help readers better understand these concepts.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studies the problem of incorporating static spectral risk measures (SRM into Distributional Reinforcement Learning (DRL) to enable more flexible and interpretable risk-sensitive decision-making. Unlike conventional Conditional Value-at-Risk (CVaR), SRMs offer a spectrum of risk preferences, allowing for more flexible risk-sensitive policies. The authors argue that using SRMs in DRL enables more flexible and interpretable policies, as SRMs allow for a spectrum of risk preferences rather than a fixed measure like CVaR. The authors propose an iterative DRL algorithm that utilizes a two-stage optimization process to optimize SRMs. They provide theoretical guarantees, proving convergence and characterizing the temporal decomposition of SRMs within the DRL framework. This decomposition enhances interpretability, as it captures how risk preferences evolve over time. The algorithm’s effectiveness is demonstrated through extensive numerical studies across four example environments, where it outperforms several baseline models, highlighting its potential for real-world applications.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.\tThis paper is easy to follow and well-organized. \n2.\tThe theoretical analyses throughout the paper are technically sound and comprehensive, providing strong support for the proposed method.\n3.\tThe authors preform thorough numerical studies across four examples. The proposed algorithm outperforms several baselines, highlighting its potential for real-world applications.", "weaknesses": "See Question section", "questions": "1.\tIn Table 1, when the objective is CVaR(0.1), why does QR-SRM with $\\alpha=0.1$ not achieve the highest value? Could the authors clarify the reasons  influencing this outcome?\n2.\tThe authors introduce the decomposition theorem for SRMs (Theorem 2). Could they use one of the four examples to illustrate how this theorem applies in a practical scenario? This would help readers better understand these concepts.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730712137474}, {"id": "ZCDexcZOSV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10312/Reviewer_RwY2"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper addresses limitations in current risk-sensitive Distributional Reinforcement Learning  by introducing an algorithm that optimizes for a broader class of static SRM, moving beyond the commonly used CVaR. The proposed QR-SRM algorithm utilizes SRMs to adjust the agent's risk sensitivity dynamically, improving policy interpretability and adaptability. Extensive experiments on environments demonstrate that QR-SRM achieves superior performance and consistency with SRM objectives.", "review_text": "This paper addresses limitations in current risk-sensitive Distributional Reinforcement Learning  by introducing an algorithm that optimizes for a broader class of static SRM, moving beyond the commonly used CVaR. The proposed QR-SRM algorithm utilizes SRMs to adjust the agent's risk sensitivity dynamically, improving policy interpretability and adaptability. Extensive experiments on environments demonstrate that QR-SRM achieves superior performance and consistency with SRM objectives.", "strengths": "- The application of SRM to DRL for more interpretable risk-sensitive policies is innovative, introducing valuable theoretical insights and practical tools for risk-sensitive control.\n- Theoretical grounding and comprehensive experimental evaluations\n- The problem formulation, motivation, and results are well-articulated, though some technical details could be simplified.", "weaknesses": "- Certain theoretical sections, especially around SRM decomposition, may challenge readers due to dense terminology and complex proofs. More illustrative examples or simplified explanations could improve accessibility.\n- The paper assumes specific properties of SRMs and fixed initial preferences, which may limit the algorithm's flexibility in dynamic environments.", "questions": "- The authors use an extended state space to solve the inner optimization problem. Can you provide rigorous justification\n- The proposed algorithm's computational complexity is not thoroughly analyzed. Given the bilevel optimization algorithmic framework, and the added complexity of optimizing SRM in a distributional RL framework, the computational complexity of the proposed algorithm may be a concern.\n\nsome missing references about DRL for RSRL\n- Keramati, R., Dann, C., Tamkin, A. and Brunskill, E., 2020, April. Being optimistic to be conservative: Quickly learning a CVaR policy. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 04, pp. 4436-4443).\n- Liang, H. and Luo, Z.Q., 2024. Bridging distributional and risk-sensitive reinforcement learning with provable regret bounds. Journal of Machine Learning Research, 25(221), pp.1-56.\n- Chen, Y., Zhang, X., Wang, S. and Huang, L., Provable Risk-Sensitive Distributional Reinforcement Learning with General Function Approximation. In Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses limitations in current risk-sensitive Distributional Reinforcement Learning  by introducing an algorithm that optimizes for a broader class of static SRM, moving beyond the commonly used CVaR. The proposed QR-SRM algorithm utilizes SRMs to adjust the agent's risk sensitivity dynamically, improving policy interpretability and adaptability. Extensive experiments on environments demonstrate that QR-SRM achieves superior performance and consistency with SRM objectives.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The application of SRM to DRL for more interpretable risk-sensitive policies is innovative, introducing valuable theoretical insights and practical tools for risk-sensitive control.\n- Theoretical grounding and comprehensive experimental evaluations\n- The problem formulation, motivation, and results are well-articulated, though some technical details could be simplified.", "weaknesses": "- Certain theoretical sections, especially around SRM decomposition, may challenge readers due to dense terminology and complex proofs. More illustrative examples or simplified explanations could improve accessibility.\n- The paper assumes specific properties of SRMs and fixed initial preferences, which may limit the algorithm's flexibility in dynamic environments.", "questions": "- The authors use an extended state space to solve the inner optimization problem. Can you provide rigorous justification\n- The proposed algorithm's computational complexity is not thoroughly analyzed. Given the bilevel optimization algorithmic framework, and the added complexity of optimizing SRM in a distributional RL framework, the computational complexity of the proposed algorithm may be a concern.\n\nsome missing references about DRL for RSRL\n- Keramati, R., Dann, C., Tamkin, A. and Brunskill, E., 2020, April. Being optimistic to be conservative: Quickly learning a CVaR policy. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 04, pp. 4436-4443).\n- Liang, H. and Luo, Z.Q., 2024. Bridging distributional and risk-sensitive reinforcement learning with provable regret bounds. Journal of Machine Learning Research, 25(221), pp.1-56.\n- Chen, Y., Zhang, X., Wang, S. and Huang, L., Provable Risk-Sensitive Distributional Reinforcement Learning with General Function Approximation. In Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730642565342}, {"id": "YNLWICANlg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10312/Reviewer_RikA"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper introduces a novel distributional reinforcement learning algorithm called QR-SRM, which extends beyond expected return by incorporating spectral risk measures. \nThe authors provide convergence guarantees and enhance interpretability of policies by decomposing coherent risk measures.", "review_text": "This paper introduces a novel distributional reinforcement learning algorithm called QR-SRM, which extends beyond expected return by incorporating spectral risk measures. \nThe authors provide convergence guarantees and enhance interpretability of policies by decomposing coherent risk measures.", "strengths": "- The paper addresses a well-motivated problem by proposing an algorithm with asymptotically optimal regret bounds for scenarios involving trajectory-level feedback. \n\n- Up to Section 4, the authors clearly outline the motivations, objectives, and proof sketches, making it easier for readers to grasp the core concepts.\n\n- The authors included code for reproducibility and conducted experiments across diverse environments, adding practical value and robustness to the study.", "weaknesses": "- A minor weakness is the need for improved visualization in the experimental results. The vertical lines are not immediately distinguishable, so adjusting the dash spacing, line thickness, or adding markers would enhance clarity. Additionally, using consistent labels for the same algorithm in the legend would help reduce reader fatigue.\n\n- My main concern is that the low performance of the experimental results makes it difficult to be confident that the algorithm was correctly reproduced. According to [1], QR-DQN performs at least 100 points on LunarLander-v2 after 0.1M steps. However, Table 3 of this paper shows much lower scores, suggesting the results may not be fully reproducible. Can the authors clarify?\n\n- Although the experiments were conducted in various environments, the aforementioned concerns about reproducibility make fair comparison with other baselines challenging. Reporting performance on some Atari environments, commonly used by algorithms that assume discrete action spaces, would provide a more reliable basis for apple-to-apple comparison.\n\n\n[1] Cho, Taehyun, et al. \"Pitfall of optimism: distributional reinforcement learning by randomizing risk criterion.\" Advances in Neural Information Processing Systems 36 (2024).\n\nTypos\n\n- Line 197: \"Coheret\" should be \"Coherent\"\n\n- Line 230: $[h_l (G^{\\pi})]$ should be $\\mathbb{E}[h_l (G^{\\pi})]$\n\n- Line 286: \"Output\" should be in bold.", "questions": "- In Line 232, shouldn't it be $h_{l+1} = \\arg \\max _h \\mathbb{E}[h(G^{\\pi^*_l})] + \\int_0^1 \\hat{h}(\\phi(u)) du$? \nI'm wondering if $ \\int_0^1 \\hat{h}(\\phi(u)) du=0$ is inherently guaranteed within the algorithm, or if there is a condition that ensures this which I may have missed.\n\n- In Line 383, $\\alpha=0.6$ seems to be maximized at $CVaR_{0.8}$, and $\\alpha=0.4$ at $CVaR_{0.6}$. Although the small vertical line intervals may be minor, the lack of alignment with targeted risk levels raises concerns.\n\n- In Table 2, aren't the cases with $\\alpha=1.0$ essentially QR-DQN?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel distributional reinforcement learning algorithm called QR-SRM, which extends beyond expected return by incorporating spectral risk measures. \nThe authors provide convergence guarantees and enhance interpretability of policies by decomposing coherent risk measures.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper addresses a well-motivated problem by proposing an algorithm with asymptotically optimal regret bounds for scenarios involving trajectory-level feedback. \n\n- Up to Section 4, the authors clearly outline the motivations, objectives, and proof sketches, making it easier for readers to grasp the core concepts.\n\n- The authors included code for reproducibility and conducted experiments across diverse environments, adding practical value and robustness to the study.", "weaknesses": "- A minor weakness is the need for improved visualization in the experimental results. The vertical lines are not immediately distinguishable, so adjusting the dash spacing, line thickness, or adding markers would enhance clarity. Additionally, using consistent labels for the same algorithm in the legend would help reduce reader fatigue.\n\n- My main concern is that the low performance of the experimental results makes it difficult to be confident that the algorithm was correctly reproduced. According to [1], QR-DQN performs at least 100 points on LunarLander-v2 after 0.1M steps. However, Table 3 of this paper shows much lower scores, suggesting the results may not be fully reproducible. Can the authors clarify?\n\n- Although the experiments were conducted in various environments, the aforementioned concerns about reproducibility make fair comparison with other baselines challenging. Reporting performance on some Atari environments, commonly used by algorithms that assume discrete action spaces, would provide a more reliable basis for apple-to-apple comparison.\n\n\n[1] Cho, Taehyun, et al. \"Pitfall of optimism: distributional reinforcement learning by randomizing risk criterion.\" Advances in Neural Information Processing Systems 36 (2024).\n\nTypos\n\n- Line 197: \"Coheret\" should be \"Coherent\"\n\n- Line 230: $[h_l (G^{\\pi})]$ should be $\\mathbb{E}[h_l (G^{\\pi})]$\n\n- Line 286: \"Output\" should be in bold.", "questions": "- In Line 232, shouldn't it be $h_{l+1} = \\arg \\max _h \\mathbb{E}[h(G^{\\pi^*_l})] + \\int_0^1 \\hat{h}(\\phi(u)) du$? \nI'm wondering if $ \\int_0^1 \\hat{h}(\\phi(u)) du=0$ is inherently guaranteed within the algorithm, or if there is a condition that ensures this which I may have missed.\n\n- In Line 383, $\\alpha=0.6$ seems to be maximized at $CVaR_{0.8}$, and $\\alpha=0.4$ at $CVaR_{0.6}$. Although the small vertical line intervals may be minor, the lack of alignment with targeted risk levels raises concerns.\n\n- In Table 2, aren't the cases with $\\alpha=1.0$ essentially QR-DQN?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730600416319}, {"id": "huzoWkZ4m8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10312/Reviewer_tRxB"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper aims to extend the work of Bauerle and Glauner [1] on static spectral risk measures (convex combinations of CVaR) in Markov Decision Processes (MDPs) to the context of distributional reinforcement learning (RL). Sections 4 and Appendices A and B reformulate the approach from [1] using distributional value functions. Theorem 1 provides a bound on the performance of the policy derived from greedy action selection over an augmented state space (x, s, c). Algorithm 2 proposes the TD error computation for distributional value function, contrasting with methods like QR-DQN and IQN, by directly addressing for static spectral risk measures. Theorem 2 extends the decomposition from coherent risk measures [2] to a broader class of spectral risks, increasing the generalizability of the approach to a wider array of risk-sensitive applications. Finally, the experiments validate the proposed algorithm, offering evidence of its efficacy and robustness within this distributional risk-sensitive framework.\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.\n\n[2] Georg Ch. Pflug and Alois Pichler. Time-Consistent Decisions and Temporal Decomposition of Coherent Risk Functional. Mathematics of Operations Research, 41(2):682–699, 2016.", "review_text": "This paper aims to extend the work of Bauerle and Glauner [1] on static spectral risk measures (convex combinations of CVaR) in Markov Decision Processes (MDPs) to the context of distributional reinforcement learning (RL). Sections 4 and Appendices A and B reformulate the approach from [1] using distributional value functions. Theorem 1 provides a bound on the performance of the policy derived from greedy action selection over an augmented state space (x, s, c). Algorithm 2 proposes the TD error computation for distributional value function, contrasting with methods like QR-DQN and IQN, by directly addressing for static spectral risk measures. Theorem 2 extends the decomposition from coherent risk measures [2] to a broader class of spectral risks, increasing the generalizability of the approach to a wider array of risk-sensitive applications. Finally, the experiments validate the proposed algorithm, offering evidence of its efficacy and robustness within this distributional risk-sensitive framework.\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.\n\n[2] Georg Ch. Pflug and Alois Pichler. Time-Consistent Decisions and Temporal Decomposition of Coherent Risk Functional. Mathematics of Operations Research, 41(2):682–699, 2016.", "strengths": "The strengths of this paper lie in its deep understanding of the current state of research in risk-averse reinforcement learning (RL) and the limitations of recent work in risk-averse distributional RL (DRL). Specifically, the paper identifies key challenges: (1) dynamic and fixed risk DRL approaches often lack interpretability, and (2) the dual representation of coherent risk measures encounters issues during policy optimization. The authors’ solid grasp of the mathematical foundations behind risk measures enables them to combine and present a concise and theoretically sound introduction.", "weaknesses": "Despite the authors’ strong grasp of the limitations in risk-averse distributional RL research, this paper has several notable weaknesses:\n\n(I) The primary weakness lies in the limited originality of the contributions. Much of the content in Theorems and Lemmas in Appendix A, B, and Section 4 is directly adapted from [1], with only minor modifications to distribution value function representation. This reliance raises questions about the novelty and depth of the contributions, compared to [1]. \n\n(II) While the Introduction and Preliminaries are well-articulated, Sections 4 and 5 suffer from clarity issues. Section 5 appears disconnected from previous sections. Spectral risk measures can be represented as convex combinations of CVaR, Theorem 2 in Section 5 leverages this property to extend the dual decomposition from coherent risk to general spectral risk measures. However this dual decomposition is unrelated to algorithm 1 and 2 in the earlier sections.\n\n(III) Did not discuss the main limitation of extending static spectral risk MDP to model-free/distributional RL. Specifically:\n- (a) There is no analysis of the convergence properties of the algorithm 2 TD loss update  (a static risk variant similar to [2]).\n- (b) Missing analysis for approximation errors and guarantees arising from quantile discretization $\\tau_i$.\n- (c) Contraction analysis is missing, considering that the minimizer of the Huber loss may not be unique (see [3]).\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.\n\n[2] Shen, Yun, et al. \"Risk-sensitive reinforcement learning.\" Neural computation 26.7 (2014): 1298-1328.\n\n[3] Rowland, Mark, et al. \"An analysis of quantile temporal-difference learning.\" (2023).", "questions": "(I) What are the contribution of section 4 compared to [1]? Please clearly indicate which methods are re-written from [1] and what is new in this paper, in the main body and also appendix A and B. \n\n(II) Please address the limitation pointed out in Weaknesses section (III):Did not discuss the main limitation of extending static spectral risk MDP to model-free/distributional RL. Specifically:\n- (a) There is no analysis of the convergence properties of the algorithm 2 TD loss update  (a static risk variant similar to [2]).\n- (b) Missing analysis for approximation errors and guarantees arising from quantile discretization $\\tau_i$.\n- (c) Contraction analysis is missing, considering that the minimizer of the Huber loss may not be unique (see [3]).\n\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to extend the work of Bauerle and Glauner [1] on static spectral risk measures (convex combinations of CVaR) in Markov Decision Processes (MDPs) to the context of distributional reinforcement learning (RL). Sections 4 and Appendices A and B reformulate the approach from [1] using distributional value functions. Theorem 1 provides a bound on the performance of the policy derived from greedy action selection over an augmented state space (x, s, c). Algorithm 2 proposes the TD error computation for distributional value function, contrasting with methods like QR-DQN and IQN, by directly addressing for static spectral risk measures. Theorem 2 extends the decomposition from coherent risk measures [2] to a broader class of spectral risks, increasing the generalizability of the approach to a wider array of risk-sensitive applications. Finally, the experiments validate the proposed algorithm, offering evidence of its efficacy and robustness within this distributional risk-sensitive framework.\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.\n\n[2] Georg Ch. Pflug and Alois Pichler. Time-Consistent Decisions and Temporal Decomposition of Coherent Risk Functional. Mathematics of Operations Research, 41(2):682–699, 2016.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The strengths of this paper lie in its deep understanding of the current state of research in risk-averse reinforcement learning (RL) and the limitations of recent work in risk-averse distributional RL (DRL). Specifically, the paper identifies key challenges: (1) dynamic and fixed risk DRL approaches often lack interpretability, and (2) the dual representation of coherent risk measures encounters issues during policy optimization. The authors’ solid grasp of the mathematical foundations behind risk measures enables them to combine and present a concise and theoretically sound introduction.", "weaknesses": "Despite the authors’ strong grasp of the limitations in risk-averse distributional RL research, this paper has several notable weaknesses:\n\n(I) The primary weakness lies in the limited originality of the contributions. Much of the content in Theorems and Lemmas in Appendix A, B, and Section 4 is directly adapted from [1], with only minor modifications to distribution value function representation. This reliance raises questions about the novelty and depth of the contributions, compared to [1]. \n\n(II) While the Introduction and Preliminaries are well-articulated, Sections 4 and 5 suffer from clarity issues. Section 5 appears disconnected from previous sections. Spectral risk measures can be represented as convex combinations of CVaR, Theorem 2 in Section 5 leverages this property to extend the dual decomposition from coherent risk to general spectral risk measures. However this dual decomposition is unrelated to algorithm 1 and 2 in the earlier sections.\n\n(III) Did not discuss the main limitation of extending static spectral risk MDP to model-free/distributional RL. Specifically:\n- (a) There is no analysis of the convergence properties of the algorithm 2 TD loss update  (a static risk variant similar to [2]).\n- (b) Missing analysis for approximation errors and guarantees arising from quantile discretization $\\tau_i$.\n- (c) Contraction analysis is missing, considering that the minimizer of the Huber loss may not be unique (see [3]).\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.\n\n[2] Shen, Yun, et al. \"Risk-sensitive reinforcement learning.\" Neural computation 26.7 (2014): 1298-1328.\n\n[3] Rowland, Mark, et al. \"An analysis of quantile temporal-difference learning.\" (2023).", "questions": "(I) What are the contribution of section 4 compared to [1]? Please clearly indicate which methods are re-written from [1] and what is new in this paper, in the main body and also appendix A and B. \n\n(II) Please address the limitation pointed out in Weaknesses section (III):Did not discuss the main limitation of extending static spectral risk MDP to model-free/distributional RL. Specifically:\n- (a) There is no analysis of the convergence properties of the algorithm 2 TD loss update  (a static risk variant similar to [2]).\n- (b) Missing analysis for approximation errors and guarantees arising from quantile discretization $\\tau_i$.\n- (c) Contraction analysis is missing, considering that the minimizer of the Huber loss may not be unique (see [3]).\n\n\nReferences:\n\n[1] Nicole Bauerle and Alexander Glauner. Minimizing spectral risk measures applied to Markov decision processes. Mathematical Methods of Operations Research, 94(1):35–69, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730145693483}], "openreview_url": "https://openreview.net/forum?id=3SMBSTG3qN", "arxiv_id": "2501.02087", "paper_pdf": "papers/3SMBSTG3qN.pdf", "paper_pdf_sha256": "e907f006291434dc5a3999005872480a65130785f13be089928628de07bfde18", "paper_pdf_bytes": 933625, "paper_pdf_source": "openreview", "code_url": "https://github.com/MehrdadMoghimi/QRSRM", "code_repository": "MehrdadMoghimi/QRSRM", "code_commit": "ed267453bc326f0ff79f4a530480ffc6232906ba", "code_archive": "repos/3SMBSTG3qN.zip", "code_archive_sha256": "0a5dba171dbb8a64b2aa347e50a68c341bae53204c01d8923370e72bffa9ad99", "code_archive_bytes": 36988, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 25, "github_languages": {"Python": 132192}, "github_archived": false, "github_pushed_at": "2025-12-30T04:17:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beyond-cvar-leveraging-static-spectral-risk"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cSSHiLnjsJ", "year": 2024, "status": "rejected", "title": "Traveling Words: A Geometric Interpretation of Transformers", "authors": ["Raul Molina"], "authorids": ["~Raul_Molina1"], "authors_source": "OpenReview API", "abstract": "Transformers have significantly advanced the field of natural language processing, but comprehending their internal mechanisms remains a challenge. In this paper, we introduce a novel geometric perspective that elucidates the inner mechanisms of transformer operations. Our primary contribution is illustrating how layer normalization confines the latent features to a hyper-sphere, subsequently enabling attention to mold the semantic representation of words on this surface. This geometric viewpoint seamlessly connects established properties such as iterative refinement and contextual embeddings. We validate our insights by probing a pre-trained 124M parameter GPT-2 model. Our findings reveal clear query-key attention patterns in early layers and build upon prior observations regarding the subject-specific nature of attention heads at deeper layers. Harnessing these geometric insights, we present an intuitive understanding of transformers, depicting them as processes that model the trajectory of word particles along the hyper-sphere.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "9gzbhgbZvZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission986/Reviewer_rASQ"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors introduce a novel geometric perspective to shed light on the inner workings of transformers. Their main contribution is the findings on how layer normalization constrains the latent features of transformers to a hyper-sphere, which in turn allows attention mechanisms to shape the semantic representation of words on this surface. This geometric viewpoint connects various established properties of transformers, including iterative refinement and contextual embeddings. To validate their insights, the authors analyze a pre-trained GPT-2 model with 124 million parameters. Their findings unveil distinct query-key attention patterns in early layers and confirm prior observations about the specialization of attention heads in deeper layers. By leveraging these geometric insights, the paper offers an intuitive understanding of transformers, depicting iterative refinement as a process that models the trajectory of word particles along the surface of a hyper-sphere.", "review_text": "In this paper, the authors introduce a novel geometric perspective to shed light on the inner workings of transformers. Their main contribution is the findings on how layer normalization constrains the latent features of transformers to a hyper-sphere, which in turn allows attention mechanisms to shape the semantic representation of words on this surface. This geometric viewpoint connects various established properties of transformers, including iterative refinement and contextual embeddings. To validate their insights, the authors analyze a pre-trained GPT-2 model with 124 million parameters. Their findings unveil distinct query-key attention patterns in early layers and confirm prior observations about the specialization of attention heads in deeper layers. By leveraging these geometric insights, the paper offers an intuitive understanding of transformers, depicting iterative refinement as a process that models the trajectory of word particles along the surface of a hyper-sphere.", "strengths": "1. The paper presents a study on understanding transformers through the lens of layer normalization, a key component in transformers, and the matrices $W_{QK}, W_{VO}$ used in the attention mechanism.\n\n\n2. The main insights are that in each layer, the layer normalization projects the features to a shared hyper-sphere. The proposed interpretation of attention is similar to the feed-forward module by Geva et al. (2021) in that both calculate relevance scores and aggregate sub-updates for the residual stream. However, the key difference lies in how scores and updates are computed: attention relies on dynamic context, while the feed-forward module depends on static representations.\n\n3. The authors validate these insights by probing a pre-trained 124M parameter GPT-2 mode", "weaknesses": "1. The presentation can be improved significantly. I find it hard to see the differences from prior works and what exactly are the main contributions of this paper. \n\n2. Most of the emprical results are using some selected examples and I do not quite follow these results. Could you list the main points that you are making from these experiments and how the evidence justifies them?  What is the trajectory in figure 4 trying to show?", "questions": "In A.1  $\\mu$? is the average of components of a feature vector $x_i$? Can you provide clear definitions of what the features, mean, and std. deviation are? I would imagine $\\mathbf{\\mu}$ to be either an expectation of $\\mathbf{x}$ or an average of $\\mathbf{x}_i$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors introduce a novel geometric perspective to shed light on the inner workings of transformers. Their main contribution is the findings on how layer normalization constrains the latent features of transformers to a hyper-sphere, which in turn allows attention mechanisms to shape the semantic representation of words on this surface. This geometric viewpoint connects various established properties of transformers, including iterative refinement and contextual embeddings. To validate their insights, the authors analyze a pre-trained GPT-2 model with 124 million parameters. Their findings unveil distinct query-key attention patterns in early layers and confirm prior observations about the specialization of attention heads in deeper layers. By leveraging these geometric insights, the paper offers an intuitive understanding of transformers, depicting iterative refinement as a process that models the trajectory of word particles along the surface of a hyper-sphere.", "soundness": "1 poor", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1. The paper presents a study on understanding transformers through the lens of layer normalization, a key component in transformers, and the matrices $W_{QK}, W_{VO}$ used in the attention mechanism.\n\n\n2. The main insights are that in each layer, the layer normalization projects the features to a shared hyper-sphere. The proposed interpretation of attention is similar to the feed-forward module by Geva et al. (2021) in that both calculate relevance scores and aggregate sub-updates for the residual stream. However, the key difference lies in how scores and updates are computed: attention relies on dynamic context, while the feed-forward module depends on static representations.\n\n3. The authors validate these insights by probing a pre-trained 124M parameter GPT-2 mode", "weaknesses": "1. The presentation can be improved significantly. I find it hard to see the differences from prior works and what exactly are the main contributions of this paper. \n\n2. Most of the emprical results are using some selected examples and I do not quite follow these results. Could you list the main points that you are making from these experiments and how the evidence justifies them?  What is the trajectory in figure 4 trying to show?", "questions": "In A.1  $\\mu$? is the average of components of a feature vector $x_i$? Can you provide clear definitions of what the features, mean, and std. deviation are? I would imagine $\\mathbf{\\mu}$ to be either an expectation of $\\mathbf{x}$ or an average of $\\mathbf{x}_i$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699138075954}, {"id": "ZPlGNmo1W5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission986/Reviewer_TdFy"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper aims to interpret the mechanisms of transformers and establishes an explanation for the effect of layer normalization from a geometric viewpoint. The interpretation is validated via probing a GPT-2 model.", "review_text": "The paper aims to interpret the mechanisms of transformers and establishes an explanation for the effect of layer normalization from a geometric viewpoint. The interpretation is validated via probing a GPT-2 model.", "strengths": "* The paper provides an intuitive, geometry perspective for interpreting the Transformer architecture. \n* Empirical probing experiments on GPT-2 validated some claims in the paper.", "weaknesses": "* Some ideas discussed in the paper, such as interpreting LayerNorm as surface projection have been discussed in prior works and are not novel. A discussion on the novelty of the proposed paper and how it compares with prior works will help clarify this concern. \n* The paper provides an interesting perspective on the specific architecture in popular implementations of Transformers, but its applications or insights for further results are not fully discussed in the paper.", "questions": "* Figure 1 and Figure 4 suggest that work particles travel along the path determined by residual updates, but such a description is very general. Are there more specific properties within the residual updates?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to interpret the mechanisms of transformers and establishes an explanation for the effect of layer normalization from a geometric viewpoint. The interpretation is validated via probing a GPT-2 model.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* The paper provides an intuitive, geometry perspective for interpreting the Transformer architecture. \n* Empirical probing experiments on GPT-2 validated some claims in the paper.", "weaknesses": "* Some ideas discussed in the paper, such as interpreting LayerNorm as surface projection have been discussed in prior works and are not novel. A discussion on the novelty of the proposed paper and how it compares with prior works will help clarify this concern. \n* The paper provides an interesting perspective on the specific architecture in popular implementations of Transformers, but its applications or insights for further results are not fully discussed in the paper.", "questions": "* Figure 1 and Figure 4 suggest that work particles travel along the path determined by residual updates, but such a description is very general. Are there more specific properties within the residual updates?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698841975640}, {"id": "neS7TL13gB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission986/Reviewer_qGqz"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper explores geometric views into some of the primitive operations in Transformer architectures. The starting point is layer normalization which projects and normalizes a vector so that it will lie on a hypersphere. Then the query-key matrix implements an affine transformation bringing closer on the hypersphere terms that are related. The value-output matrix can be seen as an additional key-value store in the attention layer, while the computation of output probabilities reflects the similarity of the final layer represention as projected on the (shared) hypersphere with the embedding vectors in the vocabulary. \n\n(Probing) experiments quantify these observations using a pre-trained GPT-2 model. The impact of layer normalization on word embeddings is verified, query-key transformations seem to exhibit interpretable patterns at the first layer and some key-value heads work together at the last layer to preserve input key meanings. Similarly some patterns could be captured in singular vector pairs of the key-value and query-key matrices (the latter at layer 0). This work concludes with the representation of the second to last token in a sentence which can be seen that gets closer and closer to the last token representation (in a projected view) as the layers of the network model are traversed.", "review_text": "This paper explores geometric views into some of the primitive operations in Transformer architectures. The starting point is layer normalization which projects and normalizes a vector so that it will lie on a hypersphere. Then the query-key matrix implements an affine transformation bringing closer on the hypersphere terms that are related. The value-output matrix can be seen as an additional key-value store in the attention layer, while the computation of output probabilities reflects the similarity of the final layer represention as projected on the (shared) hypersphere with the embedding vectors in the vocabulary. \n\n(Probing) experiments quantify these observations using a pre-trained GPT-2 model. The impact of layer normalization on word embeddings is verified, query-key transformations seem to exhibit interpretable patterns at the first layer and some key-value heads work together at the last layer to preserve input key meanings. Similarly some patterns could be captured in singular vector pairs of the key-value and query-key matrices (the latter at layer 0). This work concludes with the representation of the second to last token in a sentence which can be seen that gets closer and closer to the last token representation (in a projected view) as the layers of the network model are traversed.", "strengths": "- This is an easy to follow paper, with interesting and intuitive geometric arguments, supported by simple matrix formulas.\n\n- Some of the examples/demonstrations reveal patterns which tend to happen either on early or deep layers and could loosely fit into the high-level geometric insights developed.", "weaknesses": "- The novelty of this work is limited since the key observations have already been mentioned in other works (that are adequately cited): [Brody et al.] for layer normalization, query-key matrix; [Millidge & Black] for value-output matrix.\n\n- The \"journey\" of the representation of one word towards the representation of the next one in a sentence is interesting but is could well be an artifact of the reduction in dimensionality in the projection as also noted. Regarding the examples that are expected to frame the geometric arguments presented, they are either very slim in volume to be conclusive (when some signal/pattern is observed) or there is simply no interesting pattern that could be easily extracted.", "questions": "- Regarding the traveling words interpretation: It would be nice to test it with more sentences and alternatively work with the original encoding vectors through the layers (no reduction: do the distances to the last token representation decrease? what about the respective distances for successive tokens not  at the end of a sentence? is there a intuitive way to argue for a possible pattern in this?)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores geometric views into some of the primitive operations in Transformer architectures. The starting point is layer normalization which projects and normalizes a vector so that it will lie on a hypersphere. Then the query-key matrix implements an affine transformation bringing closer on the hypersphere terms that are related. The value-output matrix can be seen as an additional key-value store in the attention layer, while the computation of output probabilities reflects the similarity of the final layer represention as projected on the (shared) hypersphere with the embedding vectors in the vocabulary. \n\n(Probing) experiments quantify these observations using a pre-trained GPT-2 model. The impact of layer normalization on word embeddings is verified, query-key transformations seem to exhibit interpretable patterns at the first layer and some key-value heads work together at the last layer to preserve input key meanings. Similarly some patterns could be captured in singular vector pairs of the key-value and query-key matrices (the latter at layer 0). This work concludes with the representation of the second to last token in a sentence which can be seen that gets closer and closer to the last token representation (in a projected view) as the layers of the network model are traversed.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- This is an easy to follow paper, with interesting and intuitive geometric arguments, supported by simple matrix formulas.\n\n- Some of the examples/demonstrations reveal patterns which tend to happen either on early or deep layers and could loosely fit into the high-level geometric insights developed.", "weaknesses": "- The novelty of this work is limited since the key observations have already been mentioned in other works (that are adequately cited): [Brody et al.] for layer normalization, query-key matrix; [Millidge & Black] for value-output matrix.\n\n- The \"journey\" of the representation of one word towards the representation of the next one in a sentence is interesting but is could well be an artifact of the reduction in dimensionality in the projection as also noted. Regarding the examples that are expected to frame the geometric arguments presented, they are either very slim in volume to be conclusive (when some signal/pattern is observed) or there is simply no interesting pattern that could be easily extracted.", "questions": "- Regarding the traveling words interpretation: It would be nice to test it with more sentences and alternatively work with the original encoding vectors through the layers (no reduction: do the distances to the last token representation decrease? what about the respective distances for successive tokens not  at the end of a sentence? is there a intuitive way to argue for a possible pattern in this?)", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698825376458}, {"id": "ceFIb9KHpa", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission986/Reviewer_zeL2"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This is a typical outlier paper. The submission firstly proposes an interpretation that the layer normalzation operation projects an embedding to a subspace perpendicular to [1,1,...,1] then normalizes it to a hypersphere in the subspace. Then the weight matrices in a transformer shift the projected hyperspere's position and shapes. Then the authors probe the embeddings of common words using cosine similarity. The conclusions are: (1) early layers slightly shift the embeddings to their typical contexts; (2) late layers are not understandable; (3) one case shows that the shifted embeddings are closer to the last word in a sentence.", "review_text": "This is a typical outlier paper. The submission firstly proposes an interpretation that the layer normalzation operation projects an embedding to a subspace perpendicular to [1,1,...,1] then normalizes it to a hypersphere in the subspace. Then the weight matrices in a transformer shift the projected hyperspere's position and shapes. Then the authors probe the embeddings of common words using cosine similarity. The conclusions are: (1) early layers slightly shift the embeddings to their typical contexts; (2) late layers are not understandable; (3) one case shows that the shifted embeddings are closer to the last word in a sentence.", "strengths": "+ This is a creative and interesting *OUTLIER* paper. Some of these papers can be quite influential.\n+ I am a computer vision person and as far as I know the geometric interpretation of LayerNorm is new.", "weaknesses": "- The experiments are only case studies in a small scale. And I cannot say the conclusions are meaningful or not. What's worse, some analysis does not lead to conclusions, e.g., later layer embeddings are not understandable. \n- To be honest, I fail to get why the word travelling analysis is related to the geometric interpretation of LayerNorm. We can still do these analyses without the d-1 projection interpretation right? Correct me if I am wrong.", "questions": "See the last box.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This is a typical outlier paper. The submission firstly proposes an interpretation that the layer normalzation operation projects an embedding to a subspace perpendicular to [1,1,...,1] then normalizes it to a hypersphere in the subspace. Then the weight matrices in a transformer shift the projected hyperspere's position and shapes. Then the authors probe the embeddings of common words using cosine similarity. The conclusions are: (1) early layers slightly shift the embeddings to their typical contexts; (2) late layers are not understandable; (3) one case shows that the shifted embeddings are closer to the last word in a sentence.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "+ This is a creative and interesting *OUTLIER* paper. Some of these papers can be quite influential.\n+ I am a computer vision person and as far as I know the geometric interpretation of LayerNorm is new.", "weaknesses": "- The experiments are only case studies in a small scale. And I cannot say the conclusions are meaningful or not. What's worse, some analysis does not lead to conclusions, e.g., later layer embeddings are not understandable. \n- To be honest, I fail to get why the word travelling analysis is related to the geometric interpretation of LayerNorm. We can still do these analyses without the d-1 projection interpretation right? Correct me if I am wrong.", "questions": "See the last box.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698777116797}], "openreview_url": "https://openreview.net/forum?id=cSSHiLnjsJ", "arxiv_id": "2309.07315", "paper_pdf": "papers/cSSHiLnjsJ.pdf", "paper_pdf_sha256": "8b7b86a04e6113f4d813ae3107bb40f8e1dcd4980ab7416e88021927f4faae3a", "paper_pdf_bytes": 1715297, "paper_pdf_source": "openreview", "code_url": "https://github.com/santiag0m/traveling-words", "code_repository": "santiag0m/traveling-words", "code_commit": "137aa6090aef2d3d954998bb3557f9e4a0b1ab41", "code_archive": "repos/cSSHiLnjsJ.zip", "code_archive_sha256": "b73afe4e30e50ce5d2e2305b7d07a95b99b6710c3bf67e4354ea2bc611de4084", "code_archive_bytes": 18095, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 20, "github_languages": {"Python": 42512}, "github_archived": false, "github_pushed_at": "2023-12-20T22:13:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/traveling-words-a-geometric-interpretation-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1jDN-RfQfrb", "year": 2023, "status": "rejected", "title": "Unveiling Transformers with LEGO: A Synthetic Reasoning Task", "authors": ["Yi Zhang", "Arturs Backurs", "Sebastien Bubeck", "Ronen Eldan", "Suriya Gunasekar", "Tal Wagner"], "authorids": ["~Yi_Zhang1", "~Arturs_Backurs1", "~Sebastien_Bubeck1", "~Ronen_Eldan1", "~Suriya_Gunasekar1", "~Tal_Wagner1"], "authors_source": "OpenReview API", "abstract": "We propose a synthetic reasoning task, LEGO (Learning Equality and Group Operations), that encapsulates the problem of following a chain of reasoning, and we study how the Transformer architectures learn this task. We pay special attention to data effects such as pretraining (on seemingly unrelated NLP tasks) and dataset composition (e.g., differing chain length at training and test time), as well as architectural variants such as weight-tied layers or adding convolutional components. We study how the trained models eventually succeed at the task, and in particular, we are able to understand (to some extent) some of the attention heads as well as how the information flows in the network. Based on these observations we propose a hypothesis that here pretraining helps for LEGO tasks due to certain structured attention patterns, and we experimentally verify this hypothesis. We also observe that in some data regimes the trained transformer finds ``shortcut\" solutions to follow the chain of reasoning, which impedes the model's robustness, and moreover we propose ways to prevent it. Motivated by our findings on structured attention patterns, we propose to replace certain attention heads with hardcoded patterns. This architectural change significantly reduces Flops and maintains or even improves the model's performance at large-scale pretraining.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "OxFq6AWS-C", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5434/Reviewer_drii"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a synthetic task that aims to capture aspects of multi-hop reasoning, and uses it to characterize the behavior of Transformer models. It finds that pre-trained transformers learn the task more rapidly, and attributes this to specific patterns of attention in the network. Introducing mechanisms that emulate these specific attention patterns helps close this gap in performance without the need for expensive pre-training. Models struggle with extrapolating to longer chains of reasoning than required during training. This deterioration can be mitigated by repeating layers with the same parameters as well as training with stochastic depth. ", "review_text": "My overall recommendation is based on weighing the thorough and well-written exposition/investigation of the LEGO task with the potential downsides of any synthetic benchmark, namely that it may or may not capture the key elements of actual real-world tasks, and can therefore be limited in its impact.", "strengths": "Strengths:\n- The proposed task is useful as a tool for understanding some ways in which a Transformer can perform multi-hop reasoning\n- The proposed operationalization of requiring length extrapolation is effective at disentangling issues with positional encoding from the question of whether the model can extrapolate to longer chains of reasoning \n- Experiments help provide an understanding for how pre-training, as well as stochastic depth, help improve certain aspects of the models' capabilities\n- Hardcoded attention patterns are shown to be just as effective for solving this specific task\n- The writing of the paper is clear and well-structured\n\nWeaknesses:\n- One of the strengths of the attention mechanism is that it can learn any attention pattern a task may require, and the multi-head version can learn when (at what layer) and how much to use each pattern. The hardcoded attention patterns seem sufficient for the proposed synthetic task -- almost by construction of the LEGO task. There are some preliminary findings that they may also be effective for masked language modeling and at least SQuAD, but this direction is underexplored and perhaps not all of the necessary components for real NLP tasks are captured.\n- The proposed task is perhaps not the best test-bed of *pre-training* specifically, because the approach used is to pre-train on natural language but then test on synthetic expressions. It would not be surprising for language-specific reasoning to not undergo transfer in this case, potentially leaving only the most general of structural patterns, like the attention patterns revealed in this work.\n- Overall, it's not clear how much of the Transformers' capabilities are stressed by the proposed synthetic task. It could be that understanding multi-hop computation in the style of the LEGO task is the key for explaining current behaviors and proposing improvements. Alternatively, multi-hop reasoning may just not be a crucial component of many NLP tasks as currently formulated, or LEGO might not be capturing the actual character of chain-based reasoning as required for real-world task.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a synthetic task that aims to capture aspects of multi-hop reasoning, and uses it to characterize the behavior of Transformer models. It finds that pre-trained transformers learn the task more rapidly, and attributes this to specific patterns of attention in the network. Introducing mechanisms that emulate these specific attention patterns helps close this gap in performance without the need for expensive pre-training. Models struggle with extrapolating to longer chains of reasoning than required during training. This deterioration can be mitigated by repeating layers with the same parameters as well as training with stochastic depth. ", "strength_and_weaknesses": "Strengths:\n- The proposed task is useful as a tool for understanding some ways in which a Transformer can perform multi-hop reasoning\n- The proposed operationalization of requiring length extrapolation is effective at disentangling issues with positional encoding from the question of whether the model can extrapolate to longer chains of reasoning \n- Experiments help provide an understanding for how pre-training, as well as stochastic depth, help improve certain aspects of the models' capabilities\n- Hardcoded attention patterns are shown to be just as effective for solving this specific task\n- The writing of the paper is clear and well-structured\n\nWeaknesses:\n- One of the strengths of the attention mechanism is that it can learn any attention pattern a task may require, and the multi-head version can learn when (at what layer) and how much to use each pattern. The hardcoded attention patterns seem sufficient for the proposed synthetic task -- almost by construction of the LEGO task. There are some preliminary findings that they may also be effective for masked language modeling and at least SQuAD, but this direction is underexplored and perhaps not all of the necessary components for real NLP tasks are captured.\n- The proposed task is perhaps not the best test-bed of *pre-training* specifically, because the approach used is to pre-train on natural language but then test on synthetic expressions. It would not be surprising for language-specific reasoning to not undergo transfer in this case, potentially leaving only the most general of structural patterns, like the attention patterns revealed in this work.\n- Overall, it's not clear how much of the Transformers' capabilities are stressed by the proposed synthetic task. It could be that understanding multi-hop computation in the style of the LEGO task is the key for explaining current behaviors and proposing improvements. Alternatively, multi-hop reasoning may just not be a crucial component of many NLP tasks as currently formulated, or LEGO might not be capturing the actual character of chain-based reasoning as required for real-world task.", "clarity,_quality,_novelty_and_reproducibility": "See above\n\nMinor:\n- Regarding Transformers that repeat layers: in addition to ALBERT, it would be helpful to cite other instances such as Universal Transformers (Dehghani et al)", "summary_of_the_review": "My overall recommendation is based on weighing the thorough and well-written exposition/investigation of the LEGO task with the potential downsides of any synthetic benchmark, namely that it may or may not capture the key elements of actual real-world tasks, and can therefore be limited in its impact.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666697228544}, {"id": "q0bBMB__t1d", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5434/Reviewer_s6py"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "LEGO is a synthetic reasoning task that targets chain reasoning. Think Markov chain, where the current state only depends on the previous state. They specifically evaluate BERT and ALBERT. ALBERT is particularly interesting since weights are shared across all layers in ALBERT and is hypothesized that it may be particularly suited for this type of reasoning. Indeed, it outperforms BERT. Various ablations are carried out to understand why.", "review_text": "I think the paper adds to the understanding of the various inductive biases that different transformer architectures have. There is much interest int the mathematical ability of transformers. I would like to see more models benchmarked on this task.", "strengths": "The task and setup are framed well. Ablations such as stochastic depth length extrapolation, pretraining, and probing classifier are creative ways at understanding the inductive biases of the networks.\n\nThe task seems not that specific to transformers. I do wonder how other RNNs would fare.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "LEGO is a synthetic reasoning task that targets chain reasoning. Think Markov chain, where the current state only depends on the previous state. They specifically evaluate BERT and ALBERT. ALBERT is particularly interesting since weights are shared across all layers in ALBERT and is hypothesized that it may be particularly suited for this type of reasoning. Indeed, it outperforms BERT. Various ablations are carried out to understand why.", "strength_and_weaknesses": "The task and setup are framed well. Ablations such as stochastic depth length extrapolation, pretraining, and probing classifier are creative ways at understanding the inductive biases of the networks.\n\nThe task seems not that specific to transformers. I do wonder how other RNNs would fare.", "clarity,_quality,_novelty_and_reproducibility": "I found the paper clear. I appreciated the language of abstract algebra, but I am not sure if all readers will. I thought the figures were good. The experiments are well done and should be simple to reproduce. There is not much originality in the modeling, but the task and analysis are novel", "summary_of_the_review": "I think the paper adds to the understanding of the various inductive biases that different transformer architectures have. There is much interest int the mathematical ability of transformers. I would like to see more models benchmarked on this task.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666691060590}, {"id": "FKAskJ-vD9O", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5434/Reviewer_3VMG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper proposes a synthetic task, LEGO, and uses it to understand how Transformers operate. The task is in following a chain of simple reasoning, in particular for simple mathematical operations with variables valued either +1 or -1 and expressions of the kind a=+b or a=-b. With this task, the authors consider BERT’s and ALBERT’s abilities to generalize to longer inputs. They find specific attention heads (e.g., local attention patterns) and propose modification with some of the heads fixed to these patterns.", "review_text": "The paper introduces a synthetic task to analyse Transformers. In the analysis, finds specific attention patterns and proposes model modification with some of the heads being fixed. Most of the observations done in the paper and model modification have already been done before without resorting to synthetic tasks.", "strengths": "While the paper introduces the new task to help understanding Transformers, the task does not lead to new observations not widely known before. Specifically, it’s well-known that the most important Transformer attention heads play interpretable roles, including positional ([1], [2], [3], etc). This was used extensively to improve/simplify models: [4] fixes attention window size for some of the heads and gets improvement for low-resource settings (similar to this work), [5] develops hard-coded attention without learned parameters, to name a few. From this perspective, (1) neither finding these attention patterns nor proposing fixed attention is novel; (2) doing it on a synthetic task when this has already been done before for different real-world tasks does not make much sense.\n\n\n[1] ACL 2019: Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.\n\n[2] EMNLP 2019: ADaptively Sparse Transformers.\n\n[3] BlackBoxNLP 2019: What does BERT Look at? An Analysis of BERT’s Attention\n\n[4] EMNLP findings 2020: Fixed Encoder Self-Attention Patterns in Transformer-Based Machine Translation\n\n[5] ACL 2020: Hard-Coded Gaussian Attention for Neural Machine Translation\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a synthetic task, LEGO, and uses it to understand how Transformers operate. The task is in following a chain of simple reasoning, in particular for simple mathematical operations with variables valued either +1 or -1 and expressions of the kind a=+b or a=-b. With this task, the authors consider BERT’s and ALBERT’s abilities to generalize to longer inputs. They find specific attention heads (e.g., local attention patterns) and propose modification with some of the heads fixed to these patterns.", "strength_and_weaknesses": "While the paper introduces the new task to help understanding Transformers, the task does not lead to new observations not widely known before. Specifically, it’s well-known that the most important Transformer attention heads play interpretable roles, including positional ([1], [2], [3], etc). This was used extensively to improve/simplify models: [4] fixes attention window size for some of the heads and gets improvement for low-resource settings (similar to this work), [5] develops hard-coded attention without learned parameters, to name a few. From this perspective, (1) neither finding these attention patterns nor proposing fixed attention is novel; (2) doing it on a synthetic task when this has already been done before for different real-world tasks does not make much sense.\n\n\n[1] ACL 2019: Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.\n\n[2] EMNLP 2019: ADaptively Sparse Transformers.\n\n[3] BlackBoxNLP 2019: What does BERT Look at? An Analysis of BERT’s Attention\n\n[4] EMNLP findings 2020: Fixed Encoder Self-Attention Patterns in Transformer-Based Machine Translation\n\n[5] ACL 2020: Hard-Coded Gaussian Attention for Neural Machine Translation\n", "clarity,_quality,_novelty_and_reproducibility": "See comments above.", "summary_of_the_review": "The paper introduces a synthetic task to analyse Transformers. In the analysis, finds specific attention patterns and proposes model modification with some of the heads being fixed. Most of the observations done in the paper and model modification have already been done before without resorting to synthetic tasks.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666628341500}, {"id": "aGlnIp-IUy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5434/Reviewer_7SQ8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper provides empirical investigation of transformer by constructing a synthetic reasoning task -- LEGO (Learning Equality and Group Operations). The LEGO tasks consist of group operations and a chain of reasoning. The major purpose of the synthetic experiment is to provide a controlled setting and unveil what transformer learns. To do so, the paper conduct a series of experiments on the both classic setting (training and testing distribution are the same) and the OOD setting (out of distribution) and they found\n\n(1) Classical generalization happens all the architecture.\n\n(2) OOD generalization depends on the architecture and data preparation.\n\n(3) The non-extrapolating models (architecture fails for OOD) learns short-cut solution.\n\nThe author also propose hardcode attention layer that incorporates knowledge over the dataset that seems to work well on LEGO. ", "review_text": "The paper provides solid contribution to the literature and I vote for acceptance.\n\n\nSome minor questions over \"shortcut\":\nThe paper finds transformer sometime learns short-cut solution to the reasoning tasks. However, it is unclear to me why this is a bad thing. I feel it is not equivalent to spurious feature because some reasoning tasks involves short-cut solution or parallelizable solution, e.g., it is possible that a linear size circuit only has depth $O(\\log n)$ (i.e., do not follow the chain of reasoning). Some further discussions should be a good addition on the finding!\n\n\n\n------------\n\nPost rebuttal\n\nI have read the rebuttal and my evaluation remains.", "strengths": "Strength. The paper provides solid contribution towards understanding the Transformer architecture. By constructing a synthetic benchmark, it provides a series experiment that provide new insight and confirm existing observations. \n\nWeakness. There is no major weakness.\n\nMinor comment. The related work section can be improved, in particular, there is a previous line of work \"unveils\" transformer over regular language ( in particular dyck language) that should be cited and discussed. Examples include [1] (and see the reference therein), which studies the dyck language and provide some theory and empirical investigation on what transformer learns.\n\n[1] Self-Attention Networks Can Process Bounded Hierarchical Languages. Shunyu Yao, Binghui Peng, Christos Papadimitriou, Karthik Narasimhan. ACL'2021\n\n\n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper provides empirical investigation of transformer by constructing a synthetic reasoning task -- LEGO (Learning Equality and Group Operations). The LEGO tasks consist of group operations and a chain of reasoning. The major purpose of the synthetic experiment is to provide a controlled setting and unveil what transformer learns. To do so, the paper conduct a series of experiments on the both classic setting (training and testing distribution are the same) and the OOD setting (out of distribution) and they found\n\n(1) Classical generalization happens all the architecture.\n\n(2) OOD generalization depends on the architecture and data preparation.\n\n(3) The non-extrapolating models (architecture fails for OOD) learns short-cut solution.\n\nThe author also propose hardcode attention layer that incorporates knowledge over the dataset that seems to work well on LEGO. ", "strength_and_weaknesses": "Strength. The paper provides solid contribution towards understanding the Transformer architecture. By constructing a synthetic benchmark, it provides a series experiment that provide new insight and confirm existing observations. \n\nWeakness. There is no major weakness.\n\nMinor comment. The related work section can be improved, in particular, there is a previous line of work \"unveils\" transformer over regular language ( in particular dyck language) that should be cited and discussed. Examples include [1] (and see the reference therein), which studies the dyck language and provide some theory and empirical investigation on what transformer learns.\n\n[1] Self-Attention Networks Can Process Bounded Hierarchical Languages. Shunyu Yao, Binghui Peng, Christos Papadimitriou, Karthik Narasimhan. ACL'2021\n\n\n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Yes", "summary_of_the_review": "The paper provides solid contribution to the literature and I vote for acceptance.\n\n\nSome minor questions over \"shortcut\":\nThe paper finds transformer sometime learns short-cut solution to the reasoning tasks. However, it is unclear to me why this is a bad thing. I feel it is not equivalent to spurious feature because some reasoning tasks involves short-cut solution or parallelizable solution, e.g., it is possible that a linear size circuit only has depth $O(\\log n)$ (i.e., do not follow the chain of reasoning). Some further discussions should be a good addition on the finding!\n\n\n\n------------\n\nPost rebuttal\n\nI have read the rebuttal and my evaluation remains.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666445954346}], "openreview_url": "https://openreview.net/forum?id=1jDN-RfQfrb", "arxiv_id": "2206.04301", "paper_pdf": "papers/1jDN-RfQfrb.pdf", "paper_pdf_sha256": "6ede72542471c625623656f177678287fca4a23c99163a8efcf86005b4cc0601", "paper_pdf_bytes": 1216225, "paper_pdf_source": "openreview", "code_url": "https://github.com/yizhangzzz/transformers-lego", "code_repository": "yizhangzzz/transformers-lego", "code_commit": "6b12f63450155b8e0612891f03a613e1f067563f", "code_archive": "repos/1jDN-RfQfrb.zip", "code_archive_sha256": "a205ef188847f3160719ae836c4a4b2441b249f72223164dd676722c2af2712c", "code_archive_bytes": 13633, "code_file_count": 5, "code_extensions": {".py": 4, ".ipynb": 1}, "github_disk_usage_kb": 28, "github_languages": {"Python": 14011, "Jupyter Notebook": 7693}, "github_archived": false, "github_pushed_at": "2022-12-15T22:16:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unveiling-transformers-with-lego-a-synthetic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "O9DAoNnYVlM", "year": 2022, "status": "rejected", "title": "Federated Learning via Plurality Vote", "authors": ["Kai Yue", "Richeng Jin", "Chau-Wai Wong", "Huaiyu Dai"], "authorids": ["~Kai_Yue1", "rjin2@ncsu.edu", "~Chau-Wai_Wong1", "~Huaiyu_Dai1"], "authors_source": "OpenReview API", "abstract": "Federated learning allows collaborative workers to solve a machine learning problem while preserving data privacy. Recent studies have tackled various challenges in federated learning, but the joint optimization of communication overhead, learning reliability, and deployment efficiency is still an open problem. To this end, we propose a new scheme named federated learning via plurality vote (FedVote). In each communication round of FedVote, workers transmit binary or ternary weights to the server with low communication overhead. The model parameters are aggregated via weighted voting to enhance the resilience against Byzantine attacks. When deployed for inference, the model with binary or ternary weights is resource-friendly to edge devices. We show that our proposed method can reduce quantization error and converges faster compared with the methods directly quantizing the model updates.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "jkipdW02Xyh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1742/Reviewer_mhMk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a scheme for federated learning (FL) called FedVote. In FedVote, clients use binary neural networks where a latent restricted range weight vector $\\boldsymbol{h}$ is learned, and stochastic rounding is applied to obtain quantized weights $\\in$ {$-1,+1$}. The communication cost to the server (up-link) is reduced by using quantized weights. The server aggregates weights by summing over all the weights of the clients at each round of training and applying a sign function. This mechanism is referred to as plurality voting (ties are broken randomly). However, the paper also proposes to use soft voting, which normalizes the number of ones (cf. hard sign function) to get probability values. These probabilities are then quantized with a clipping mechanism to restricted maximum and minimum values, using predefined thresholds, and broadcasted to the clients (down-link). The latent weights of the client models are updated using the soft voting results. The paper also adapts work on reputation-based voting (Bendahmane et al. 2014) to deal with adversarial FL in the form of Byzantine attacks. The paper presents convergence analysis for FedVote under the independent and identically distributed data setting. Empirical evaluations is conducted on two datasets ,Fashion-MNIST and CIFAR-10, and applied to two models, LeNet-5 and VGG-7, respectively.\n", "review_text": "Mainly, the paper has the following weaknesses:\n\n- The main proposed method (FedVote) is incremental at best, and the proposal on the adversarial aspect (Byzantine-FedVote) is limited. The normalization mechanism potentially requires careful tuning. Binary weights reduce uplink communication, but downlink savings depend on probability quantization scheme.\n\n- Confined to using binary neural networks\n\n- It is not clear how the quantization of probability values for soft voting are set (i.e., predefined thresholds) and the effects on performance.\n\n- It is not clear how the predefined coefficients ($\\beta$) are set as it relates to reputation-based voting, and the effects on performance.\n\n- Empirical evaluations is limited:\n  - Several other machine learning models/benchmarks are not evaluated on (e.g., [LEAF: A Benchmark for Federated Settings, Caldas et al. 2019]).\n  - For test accuracy experiments, the number of clients used and the number sampled each round is not given.\n  - For communication efficiency experiment, the number of clients used ($M$=31) is quite small (e.g., [Hsu et al. 2019] uses 100 clients). Down-link GB not accounted for.\n\n- Preliminaries should be provided on adversarial FL/Byzantine attack.\n\n- Information on measuring energy usage (Table 4) not provided.\n\nAre Figure 4 (a) and (b) correctly labeled? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents a scheme for federated learning (FL) called FedVote. In FedVote, clients use binary neural networks where a latent restricted range weight vector $\\boldsymbol{h}$ is learned, and stochastic rounding is applied to obtain quantized weights $\\in$ {$-1,+1$}. The communication cost to the server (up-link) is reduced by using quantized weights. The server aggregates weights by summing over all the weights of the clients at each round of training and applying a sign function. This mechanism is referred to as plurality voting (ties are broken randomly). However, the paper also proposes to use soft voting, which normalizes the number of ones (cf. hard sign function) to get probability values. These probabilities are then quantized with a clipping mechanism to restricted maximum and minimum values, using predefined thresholds, and broadcasted to the clients (down-link). The latent weights of the client models are updated using the soft voting results. The paper also adapts work on reputation-based voting (Bendahmane et al. 2014) to deal with adversarial FL in the form of Byzantine attacks. The paper presents convergence analysis for FedVote under the independent and identically distributed data setting. Empirical evaluations is conducted on two datasets ,Fashion-MNIST and CIFAR-10, and applied to two models, LeNet-5 and VGG-7, respectively.\n", "main_review": "Mainly, the paper has the following weaknesses:\n\n- The main proposed method (FedVote) is incremental at best, and the proposal on the adversarial aspect (Byzantine-FedVote) is limited. The normalization mechanism potentially requires careful tuning. Binary weights reduce uplink communication, but downlink savings depend on probability quantization scheme.\n\n- Confined to using binary neural networks\n\n- It is not clear how the quantization of probability values for soft voting are set (i.e., predefined thresholds) and the effects on performance.\n\n- It is not clear how the predefined coefficients ($\\beta$) are set as it relates to reputation-based voting, and the effects on performance.\n\n- Empirical evaluations is limited:\n  - Several other machine learning models/benchmarks are not evaluated on (e.g., [LEAF: A Benchmark for Federated Settings, Caldas et al. 2019]).\n  - For test accuracy experiments, the number of clients used and the number sampled each round is not given.\n  - For communication efficiency experiment, the number of clients used ($M$=31) is quite small (e.g., [Hsu et al. 2019] uses 100 clients). Down-link GB not accounted for.\n\n- Preliminaries should be provided on adversarial FL/Byzantine attack.\n\n- Information on measuring energy usage (Table 4) not provided.\n\nAre Figure 4 (a) and (b) correctly labeled? \n", "summary_of_the_review": "The contributions in the paper are limited and not well-focused. Overall, I do not think the paper is at a level for acceptance.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636244179663}, {"id": "UGimMOMYMmt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1742/Reviewer_LKNM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to reduce the communication cost for federated learning while ensuring that the aggregation method at the server is tolerant to the presence of Byzantine clients. Towards this instead of communicating the local update/gradients or their quantized version, this paper proposes to communicate the quantized model weights to the server. The server performs the aggregation in the quantized space and communicates the aggregate value back to the clients. To enable faithful quantization of the weights during the communication with the server, each client learns normalized weights (by applying a coordinate-wise normalization function to the latent weights). \n\nThe paper also proposes a weighted aggregation mechanism at the server that takes the reputations of different clients into account.  \n\nThe paper analyzes the impact of the weight quantization on the convergence of the underlying federated learning algorithm in an i.i.d. setting. Empirical evaluations on Fashion-MNIST and CIFAR-10 show that the proposed method outperforms existing local update/gradient compression-based schemes. \n", "review_text": "Strengths:\n\n1) The paper presents a novel communication efficient federated learning algorithm that quantizes local model weights as opposed to quantizing the local gradients. The paper employs the existing approach of quantizing normalized versions of latent weights to achieve this objective.\n2) The paper address the issue of Byzantine clients and proposes a weighted aggregation method to counter such Byzantines.\n3) Theoretical and empirical results in the paper showcase the utility of the proposed algorithm. \n\nWeaknesses:\n\n1) What purpose does Lemma 1 (one-shot FedVote) serve? The assumption that all the error events are i.i.d. would most like not to hold in a multi-round setting. Even in a single round, would one need independent initializations at different clients for the analysis to hold? Also, why can the probabilities $epsilon_{m, i}$ not be arbitrarily close to 1.\n\n2) Section 4.3 states that (without weighted aggregation) the proposed method is going to be vulnerable to the presence of Byzantines as (on average) it behaves as the FedAvg (Lemma 2). Is this conclusion correct even when one is ensuring weight normalization at the clients?\n\n3) In Section 6, what is the underlying quantization rule when sending ternary weights to the server.\n\n4) Please make the plot colors consistent across Fig 1(a) - 1(c).\n\n5) In the introduction the authors claim that \"However, directly quantizing the gradient vector does not provide the optimal trade-off between communication efficiency and model accuracy.\"  Is there any prior work that supports this blanket statement?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to reduce the communication cost for federated learning while ensuring that the aggregation method at the server is tolerant to the presence of Byzantine clients. Towards this instead of communicating the local update/gradients or their quantized version, this paper proposes to communicate the quantized model weights to the server. The server performs the aggregation in the quantized space and communicates the aggregate value back to the clients. To enable faithful quantization of the weights during the communication with the server, each client learns normalized weights (by applying a coordinate-wise normalization function to the latent weights). \n\nThe paper also proposes a weighted aggregation mechanism at the server that takes the reputations of different clients into account.  \n\nThe paper analyzes the impact of the weight quantization on the convergence of the underlying federated learning algorithm in an i.i.d. setting. Empirical evaluations on Fashion-MNIST and CIFAR-10 show that the proposed method outperforms existing local update/gradient compression-based schemes. \n", "main_review": "Strengths:\n\n1) The paper presents a novel communication efficient federated learning algorithm that quantizes local model weights as opposed to quantizing the local gradients. The paper employs the existing approach of quantizing normalized versions of latent weights to achieve this objective.\n2) The paper address the issue of Byzantine clients and proposes a weighted aggregation method to counter such Byzantines.\n3) Theoretical and empirical results in the paper showcase the utility of the proposed algorithm. \n\nWeaknesses:\n\n1) What purpose does Lemma 1 (one-shot FedVote) serve? The assumption that all the error events are i.i.d. would most like not to hold in a multi-round setting. Even in a single round, would one need independent initializations at different clients for the analysis to hold? Also, why can the probabilities $epsilon_{m, i}$ not be arbitrarily close to 1.\n\n2) Section 4.3 states that (without weighted aggregation) the proposed method is going to be vulnerable to the presence of Byzantines as (on average) it behaves as the FedAvg (Lemma 2). Is this conclusion correct even when one is ensuring weight normalization at the clients?\n\n3) In Section 6, what is the underlying quantization rule when sending ternary weights to the server.\n\n4) Please make the plot colors consistent across Fig 1(a) - 1(c).\n\n5) In the introduction the authors claim that \"However, directly quantizing the gradient vector does not provide the optimal trade-off between communication efficiency and model accuracy.\"  Is there any prior work that supports this blanket statement?", "summary_of_the_review": "The studies an important problem in the context of federated learning, namely designing communication efficient learning algorithms in the presence of Byzantine clients. The paper presents an interesting weight quantization-based algorithm and analyzes its convergence in an i.i.d. setting. The authors also demonstrate the superiority of the proposed algorithm on existing communication efficient federated learning algorithms in the literature on two standard image classification benchmarks. \n\nThat said, there are some questions that remain about the claims made in the paper (see weaknesses above).", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635979102787}, {"id": "4F5gfNyYq9k", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1742/Reviewer_B8sD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides a voting-based quantized model averaging method for federated learning. The convergence is proved and the experimental results on Fashion-MNIST and CIFAR-10 show that the proposed scheme outperforms existing schemes.\n", "review_text": "I like the topic of this paper, but I have some concerns as below.\n\n1. Explanation on the huge gap between FedVote and SignSGD with majority vote [Bernstein’18]. \n    1. It seems like the basic concept of FedVote is from [Bernstein’18]. I’m bit curious why FedVote has some gap compared with [Bernstein’18]. Is it because of the weighted voting strategy of FedVote? Can we do some ablation study to check what made FedVote such powerful, compared with [Bernstein’18]?\n2. Comparison with other Byzantine-tolerant schemes\n    1. There have been many previous works on Byzantine-tolerant schemes, e.g., using coding scheme to defend Byzantine attacks, DRACO[Chen’18], DETOX[Rajput’19], Election coding [Sohn’20], SignGuard[Xu’21], ByzShield [Konstantinidis’21]. Especially, the last four works focus on the voting-based model/gradient averaging. I think this paper should at least mention these works, and also compare the performance. \n    2. There have been naive methods of using median or mean of median schemes for tolerating Byzantines, starting from 2017, e.g., KRUM, multi-KRUM. I guess this paper didn’t compare with these naive methods also. \n    3. After the [Bernstein’18], there have been upcoming works [Karimireddy’19] suggesting that error feedback fixes the issue of SignSGD. I’m not sure whether the authors compared their work with this scheme. \n    4. So in general, I thought that the literature search and comparison is bit weak. SignSGD is 3-year old paper, and there have been many other papers using voting-based aggregation for Byzantine tolerance. \n\n[Bernstein’18] https://arxiv.org/abs/1802.04434\n[Chen’18] https://arxiv.org/abs/1803.09877\n[Karimireddy’19] http://proceedings.mlr.press/v97/karimireddy19a/karimireddy19a.pdf\n[Rajput’19] https://proceedings.neurips.cc/paper/2019/hash/415185ea244ea2b2bedeb0449b926802-Abstract.html\n[Sohn’20] https://proceedings.neurips.cc/paper/2020/hash/a7f0d2b95c60161b3f3c82f764b1d1c9-Abstract.html\n[Xu’21] https://arxiv.org/pdf/2109.05872.pdf\n[Konstantinidis’21] https://arxiv.org/abs/2010.04902", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides a voting-based quantized model averaging method for federated learning. The convergence is proved and the experimental results on Fashion-MNIST and CIFAR-10 show that the proposed scheme outperforms existing schemes.\n", "main_review": "I like the topic of this paper, but I have some concerns as below.\n\n1. Explanation on the huge gap between FedVote and SignSGD with majority vote [Bernstein’18]. \n    1. It seems like the basic concept of FedVote is from [Bernstein’18]. I’m bit curious why FedVote has some gap compared with [Bernstein’18]. Is it because of the weighted voting strategy of FedVote? Can we do some ablation study to check what made FedVote such powerful, compared with [Bernstein’18]?\n2. Comparison with other Byzantine-tolerant schemes\n    1. There have been many previous works on Byzantine-tolerant schemes, e.g., using coding scheme to defend Byzantine attacks, DRACO[Chen’18], DETOX[Rajput’19], Election coding [Sohn’20], SignGuard[Xu’21], ByzShield [Konstantinidis’21]. Especially, the last four works focus on the voting-based model/gradient averaging. I think this paper should at least mention these works, and also compare the performance. \n    2. There have been naive methods of using median or mean of median schemes for tolerating Byzantines, starting from 2017, e.g., KRUM, multi-KRUM. I guess this paper didn’t compare with these naive methods also. \n    3. After the [Bernstein’18], there have been upcoming works [Karimireddy’19] suggesting that error feedback fixes the issue of SignSGD. I’m not sure whether the authors compared their work with this scheme. \n    4. So in general, I thought that the literature search and comparison is bit weak. SignSGD is 3-year old paper, and there have been many other papers using voting-based aggregation for Byzantine tolerance. \n\n[Bernstein’18] https://arxiv.org/abs/1802.04434\n[Chen’18] https://arxiv.org/abs/1803.09877\n[Karimireddy’19] http://proceedings.mlr.press/v97/karimireddy19a/karimireddy19a.pdf\n[Rajput’19] https://proceedings.neurips.cc/paper/2019/hash/415185ea244ea2b2bedeb0449b926802-Abstract.html\n[Sohn’20] https://proceedings.neurips.cc/paper/2020/hash/a7f0d2b95c60161b3f3c82f764b1d1c9-Abstract.html\n[Xu’21] https://arxiv.org/pdf/2109.05872.pdf\n[Konstantinidis’21] https://arxiv.org/abs/2010.04902", "summary_of_the_review": "This is an interesting paper, but the idea is quite similar to SignSGD paper, and has no thorough comparison with existing works. Better comparison & ablation are needed to better understand why FedVote works better than existing schemes. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635971849288}, {"id": "nLbLycWa_Yf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1742/Reviewer_Cpmm"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The author(s) proposed a new algorithm based on plurality voting for federated learning with quantized gradients. Convergence theory is developed and experiments on both iid and non-iid datasets are conducted to evaluate the proposed method.", "review_text": "**Pros**:\n- The paper is well-written and related works are well-address to my knowledge.\n- Has a nice introduction on quantized neural networks, which is good for readers that are not familiar with the topic.\n\n**My Questions**: I have some questions on the theory side of the paper.\n\n- The author(s) described the convergence of FedVote in terms of $|| f(\\tilde{w}^{(k)}) ||^2$, which is a valid measure of stationarity for continuous problems. I wonder what is the conclusion to the discrete problem? For example, we have a stationary point $w^*$ such that $|| f( \\tilde{w}^* ) || = 0$. Then what conclusion can we obtain for the following problem\n$$ \n\t\\min_{w \\in \\mathbb{D}^2_n} f(w).\n$$\nCan we say that we are at a global minimum in the convex case?\n\n- In Lemma 1, the author(s) define $s_i =\\frac{2}{M} \\sum_{i=1}^M \\epsilon_{m,  i}$. From the definition we know $s_i \\in (0,  2)$. Why the author(s) conclude that $s_i \\in (0, 1)$?\n\n-  Theorem 1 describes the convergence of FedVote for $\\eta$ that satisfies the assumption. I wonder why the author(s) do not give a Corollary on the convergence rate for a particular choice of $\\eta$ (which is usually done in the literature, e.g., see the analysis in [1])? It is not straightforward for readers to see the convergence (or non-convergence) without a concrete choice of $\\eta$.\n\n- In Remark 2, the author(s) mention that the variance of quantization error will impede the convergence, which also holds for FedPAQ. However, I do not see any non-convergence issue of FedPAQ from [Theorem 2, 1]. Could the author(s) explain more on the non-convergence of FedPAQ?\n\n- On the theory side, could the author(s) elaborate the advantages of the proposed method over FedPAQ?\n\n[1] Reisizadeh et al. FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization. AISTATS 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The author(s) proposed a new algorithm based on plurality voting for federated learning with quantized gradients. Convergence theory is developed and experiments on both iid and non-iid datasets are conducted to evaluate the proposed method.", "main_review": "**Pros**:\n- The paper is well-written and related works are well-address to my knowledge.\n- Has a nice introduction on quantized neural networks, which is good for readers that are not familiar with the topic.\n\n**My Questions**: I have some questions on the theory side of the paper.\n\n- The author(s) described the convergence of FedVote in terms of $|| f(\\tilde{w}^{(k)}) ||^2$, which is a valid measure of stationarity for continuous problems. I wonder what is the conclusion to the discrete problem? For example, we have a stationary point $w^*$ such that $|| f( \\tilde{w}^* ) || = 0$. Then what conclusion can we obtain for the following problem\n$$ \n\t\\min_{w \\in \\mathbb{D}^2_n} f(w).\n$$\nCan we say that we are at a global minimum in the convex case?\n\n- In Lemma 1, the author(s) define $s_i =\\frac{2}{M} \\sum_{i=1}^M \\epsilon_{m,  i}$. From the definition we know $s_i \\in (0,  2)$. Why the author(s) conclude that $s_i \\in (0, 1)$?\n\n-  Theorem 1 describes the convergence of FedVote for $\\eta$ that satisfies the assumption. I wonder why the author(s) do not give a Corollary on the convergence rate for a particular choice of $\\eta$ (which is usually done in the literature, e.g., see the analysis in [1])? It is not straightforward for readers to see the convergence (or non-convergence) without a concrete choice of $\\eta$.\n\n- In Remark 2, the author(s) mention that the variance of quantization error will impede the convergence, which also holds for FedPAQ. However, I do not see any non-convergence issue of FedPAQ from [Theorem 2, 1]. Could the author(s) explain more on the non-convergence of FedPAQ?\n\n- On the theory side, could the author(s) elaborate the advantages of the proposed method over FedPAQ?\n\n[1] Reisizadeh et al. FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization. AISTATS 2020.", "summary_of_the_review": "I have some questions about the theoretical analysis of this paper. I hope that the author(s) could kindly elaborate.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635897218158}], "openreview_url": "https://openreview.net/forum?id=O9DAoNnYVlM", "arxiv_id": "2110.02998", "paper_pdf": "papers/O9DAoNnYVlM.pdf", "paper_pdf_sha256": "140e2042c9151a91368b6725a90fbaa6bd8c516b2e0b49cc12e54f8598682ee4", "paper_pdf_bytes": 538435, "paper_pdf_source": "openreview", "code_url": "https://github.com/KAI-YUE/fedvote", "code_repository": "KAI-YUE/fedvote", "code_commit": "454b61d14c2e55aebe445cc40929041d89db3edc", "code_archive": "repos/O9DAoNnYVlM.zip", "code_archive_sha256": "31bdf65ccf2915073526e1b5d6d16a424487aaa84c0bcc0dd35069764dd25a0a", "code_archive_bytes": 25996, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 26, "github_languages": {"Python": 58859}, "github_archived": false, "github_pushed_at": "2023-06-01T23:24:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/federated-learning-via-plurality-vote"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-NEXDKk8gZ", "year": 2021, "status": "rejected", "title": "Improved Denoising Diffusion Probabilistic Models", "authors": ["Alexander Quinn Nichol", "Prafulla Dhariwal"], "authorids": ["~Alexander_Quinn_Nichol1", "~Prafulla_Dhariwal1"], "authors_source": "OpenReview API", "abstract": "We explore denoising diffusion probabilistic models, a class of generative models which have recently been shown to produce excellent samples in the image and audio domains. While these models produce excellent samples, it has yet to be shown that they can achieve competitive log-likelihoods. We show that, with several small modifications, diffusion models can achieve competitive log-likelihoods in the image domain while maintaining high sample quality. Additionally, our models allow for sampling with an order of magnitude fewer diffusion steps with only a modest difference in sample quality. Finally, we explore how sample quality and log-likelihood scale with the number of diffusion steps and the amount of model capacity. We conclude that denoising diffusion probabilistic models are a promising class of generative models with excellent scaling properties and sample quality.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "IdSqd5RX6eC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper916/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**\n\nThis paper presents rich discussions and various practical techniques to improve the training of probabilistic diffusion models, which include a hybrid objective to learn the variance for improving log-likelihood performance, a different noise schedule tailored for ImageNet 64x64 and importance sampling to reduce gradient noise. Experiments on ImageNet 64x64 and various ablation study provided interesting insights and empirically justified the claims.\n\n**Pros**\nThe paper is well-written and develops some useful practical techniques for improving a recently proposed deep generative model (a modified version of the diffusion probabilistic model in (Jascha, et al 2015)). Specifically, the paper managed to improve the log-likelihood performance by identifying the issue of the simplified objective and proposed to learn the variance using a hybrid objective. I think this technique along with others are useful practical techniques to improve the training of diffusion models.\n\n**Questions & Concerns**\n- No results on CIFAR: Most recent papers in this field considers CIFAR-10 as the standard benchmark to report generative performance, including the original DDPM paper [1] and the score matching paper [2]. Although ImageNet 64x64 is a larger dataset with more complicated structure and diversity, outperforming pervious strong baselines in CIFAR-10 is still challenging and non-trivial. Thus the empirical study will also be more convincing to demonstrate that the proposed method can indeed achieve much better log-likelihood without sacrificing sample quality too much. Otherwise, it's hard to get a sense of how much improvement has been actually achieved by directly looking at the numbers in this paper and the ones in previous papers.\n\n- As the major contribution, the variance parametrization (Eq 16) needs more insightful discussions. For example, why this is the case: \"Since Figure 1a shows that the reasonable range for\u0012$\\Sigma(x_t; t)$ is very small, it is clear that we should not use a neural network to predict\u0012$\\Sigma(x_t; t)$ directly.\". Can we predict the log of the variance with a neural network directly? The proposed one is only an interpolation between $\\beta_t$ and $\\tilde{\\beta}_t$ - is this expressive enough?\n\n- About the noise schedule: is this a generally better noise schedule, or it is only tailored for ImageNet 64x64. In the latter case, I think it is only a trick that overfits a specific dataset. To improve the training of DDPM generally, is there any advice on how to find a good noise schedule?\n\nI will consider raising my score if the above concerns can be addressed.\n\n[1] Denoising Diffusion Probabilistic Models\n\n[2] Generative modeling by estimating gradients of the data distribution", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting techniques for improving diffusion models", "review": "**Summary**\n\nThis paper presents rich discussions and various practical techniques to improve the training of probabilistic diffusion models, which include a hybrid objective to learn the variance for improving log-likelihood performance, a different noise schedule tailored for ImageNet 64x64 and importance sampling to reduce gradient noise. Experiments on ImageNet 64x64 and various ablation study provided interesting insights and empirically justified the claims.\n\n**Pros**\nThe paper is well-written and develops some useful practical techniques for improving a recently proposed deep generative model (a modified version of the diffusion probabilistic model in (Jascha, et al 2015)). Specifically, the paper managed to improve the log-likelihood performance by identifying the issue of the simplified objective and proposed to learn the variance using a hybrid objective. I think this technique along with others are useful practical techniques to improve the training of diffusion models.\n\n**Questions & Concerns**\n- No results on CIFAR: Most recent papers in this field considers CIFAR-10 as the standard benchmark to report generative performance, including the original DDPM paper [1] and the score matching paper [2]. Although ImageNet 64x64 is a larger dataset with more complicated structure and diversity, outperforming pervious strong baselines in CIFAR-10 is still challenging and non-trivial. Thus the empirical study will also be more convincing to demonstrate that the proposed method can indeed achieve much better log-likelihood without sacrificing sample quality too much. Otherwise, it's hard to get a sense of how much improvement has been actually achieved by directly looking at the numbers in this paper and the ones in previous papers.\n\n- As the major contribution, the variance parametrization (Eq 16) needs more insightful discussions. For example, why this is the case: \"Since Figure 1a shows that the reasonable range for\u0012$\\Sigma(x_t; t)$ is very small, it is clear that we should not use a neural network to predict\u0012$\\Sigma(x_t; t)$ directly.\". Can we predict the log of the variance with a neural network directly? The proposed one is only an interpolation between $\\beta_t$ and $\\tilde{\\beta}_t$ - is this expressive enough?\n\n- About the noise schedule: is this a generally better noise schedule, or it is only tailored for ImageNet 64x64. In the latter case, I think it is only a trick that overfits a specific dataset. To improve the training of DDPM generally, is there any advice on how to find a good noise schedule?\n\nI will consider raising my score if the above concerns can be addressed.\n\n[1] Denoising Diffusion Probabilistic Models\n\n[2] Generative modeling by estimating gradients of the data distribution", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604968741146}, {"id": "6Hq69CeN4b", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper916/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper talks builds upon the recent work from Ho (2020) about generative models that use noise diffusion. The authors suggest that the proposal in Ho can not only be used in good quality sample generation (as already shown by Ho), but also leads to reasonable improvements in likelihood. Overall, some of the ideas presented in the paper are interesting and useful; but the paper overall needs work. \n\nOther questions/concerns:\n1. Firstly, from an application point of view, what does achieving a high log-likelihood mean, if the samples are already good enough or high quality? \n2. How do we interpret the bits/dim metric here? Its rather hard to rationalize that a change in 0.01 makes sense in this metric? And more generally, what are we aiming for in terms of a reasonable change?\n3. In section 3.1; how did we end up needing to tune big-sigma_theta (x_t, t) while arguing that fixing small-sig_t^2 is ok? This is in section 3.1 second paragraph; Either I am missing something of the argument here is that we need to tune noise variance and cannot fix it? \n4. What is the intuition behind expecting the (squared) cosine schedule to work? It is interesting to think that a periodic decay noising schedule is better than a linear one? \n5. And related to that, do not understand this weird value of 0.008 for s? The whole point here is some small non-zero s is ok; why specifically 0.008?! \n6. One of the main conclusions in section 3.4 is kind of confusing --- based on the summary, if we are not interested in sample quality but only interested in maxing of likelihood, then the proposal of this work is not good, and working with L_vlb suffices? Is this correct? Based on the motivation, it seems the opposite was being claimed i.e., L_hybrid is important for maxing of likelihood (third para in introduction)? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors explore behavior of likelihoods for diffusion models. Some useful experiments. Need work.", "review": "The paper talks builds upon the recent work from Ho (2020) about generative models that use noise diffusion. The authors suggest that the proposal in Ho can not only be used in good quality sample generation (as already shown by Ho), but also leads to reasonable improvements in likelihood. Overall, some of the ideas presented in the paper are interesting and useful; but the paper overall needs work. \n\nOther questions/concerns:\n1. Firstly, from an application point of view, what does achieving a high log-likelihood mean, if the samples are already good enough or high quality? \n2. How do we interpret the bits/dim metric here? Its rather hard to rationalize that a change in 0.01 makes sense in this metric? And more generally, what are we aiming for in terms of a reasonable change?\n3. In section 3.1; how did we end up needing to tune big-sigma_theta (x_t, t) while arguing that fixing small-sig_t^2 is ok? This is in section 3.1 second paragraph; Either I am missing something of the argument here is that we need to tune noise variance and cannot fix it? \n4. What is the intuition behind expecting the (squared) cosine schedule to work? It is interesting to think that a periodic decay noising schedule is better than a linear one? \n5. And related to that, do not understand this weird value of 0.008 for s? The whole point here is some small non-zero s is ok; why specifically 0.008?! \n6. One of the main conclusions in section 3.4 is kind of confusing --- based on the summary, if we are not interested in sample quality but only interested in maxing of likelihood, then the proposal of this work is not good, and working with L_vlb suffices? Is this correct? Based on the motivation, it seems the opposite was being claimed i.e., L_hybrid is important for maxing of likelihood (third para in introduction)? ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603951495278}, {"id": "RWWvPZfKbv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper916/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Denoising diffusion probabilistic models have been proved to produce excellent samples in the image and audio domains. However, it has yet to be shown that they can achieve competitive log-likelihoods. This paper shows that with several small modifications, diffusion models can achieve competitive log-likelihoods in the image domain while maintaining high sample quality. This paper is well-written and good-organized. However, I have the following concerns.\n\n1.\tThe authors claim that the noise schedule used in Ho et al. (2020) was, experimentally, sub-optimal for ImageNet $64 \\times 64$, which lacks theoretical guarantees.\n2.\tThis manuscript is mainly based on the previous work Ho et al. (2020). The novelty seems to be too limited.\n3.\tI am not convinced that only one dataset (ImageNet $64 \\times 64$) is sufficient to demonstrate the performance of the proposed strategy.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "Denoising diffusion probabilistic models have been proved to produce excellent samples in the image and audio domains. However, it has yet to be shown that they can achieve competitive log-likelihoods. This paper shows that with several small modifications, diffusion models can achieve competitive log-likelihoods in the image domain while maintaining high sample quality. This paper is well-written and good-organized. However, I have the following concerns.\n\n1.\tThe authors claim that the noise schedule used in Ho et al. (2020) was, experimentally, sub-optimal for ImageNet $64 \\times 64$, which lacks theoretical guarantees.\n2.\tThis manuscript is mainly based on the previous work Ho et al. (2020). The novelty seems to be too limited.\n3.\tI am not convinced that only one dataset (ImageNet $64 \\times 64$) is sufficient to demonstrate the performance of the proposed strategy.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603726079424}, {"id": "kn3xx0MTdv3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper916/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper found several methods to improve log likelihood of diffusion models while maintain their sample quality, including cosine instead of linear noise schedule, using a hybrid objective to learn parameters of the covariance function, and using importance sampling to improve the gradient noise. The authors also explore how sample quality and log likelihood scale with the number of diffusion steps and model capacity. Experiments on 64x64 ImageNet dataset show competitive llh while keeping the sample quality.\n\nClarity: as the denoising diffusion model, especially its success in generating high-quality image samples, is still quite new, it would be beneficial if the paper could describe the technical background of it and the existing variational training methods in more details. Section 2 kind of serves this purpose, but it misses many important steps and focuses more on defining terms used in later sections.\n\nSignificance of this work: The necessity for having larger log likelihood for the diffusion model is not very well motivated. If the model is mostly used to generate high-quality samples and we have known how to train it to do so, why does the LLH values still matter?\n\nOriginality: using cosine noise schedule, new parameterization and hybrid objective seems effective for training, but doesn't seem to be very innovative. But I'm not very familiar with the denoising diffusion model and could be wrong. The results on how the model scales with computation seems trivial and may have been known already. Finally, it'll be beneficial if the authors can verify findings got in this paper can apply to other dataset or types of data more broadly.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Tricks to improve log likelihood of diffusion models while maintain their sample quality", "review": "The paper found several methods to improve log likelihood of diffusion models while maintain their sample quality, including cosine instead of linear noise schedule, using a hybrid objective to learn parameters of the covariance function, and using importance sampling to improve the gradient noise. The authors also explore how sample quality and log likelihood scale with the number of diffusion steps and model capacity. Experiments on 64x64 ImageNet dataset show competitive llh while keeping the sample quality.\n\nClarity: as the denoising diffusion model, especially its success in generating high-quality image samples, is still quite new, it would be beneficial if the paper could describe the technical background of it and the existing variational training methods in more details. Section 2 kind of serves this purpose, but it misses many important steps and focuses more on defining terms used in later sections.\n\nSignificance of this work: The necessity for having larger log likelihood for the diffusion model is not very well motivated. If the model is mostly used to generate high-quality samples and we have known how to train it to do so, why does the LLH values still matter?\n\nOriginality: using cosine noise schedule, new parameterization and hybrid objective seems effective for training, but doesn't seem to be very innovative. But I'm not very familiar with the denoising diffusion model and could be wrong. The results on how the model scales with computation seems trivial and may have been known already. Finally, it'll be beneficial if the authors can verify findings got in this paper can apply to other dataset or types of data more broadly.", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603670524033}], "openreview_url": "https://openreview.net/forum?id=-NEXDKk8gZ", "arxiv_id": "2102.09672", "paper_pdf": "papers/-NEXDKk8gZ.pdf", "paper_pdf_sha256": "fc2553e36d03bed548970db5ce78e17041965ac3e232ddf1b0b461b7e13bc3dc", "paper_pdf_bytes": 8094735, "paper_pdf_source": "openreview", "code_url": "https://github.com/openai/improved-diffusion", "code_repository": "openai/improved-diffusion", "code_commit": "1bc7bbbdc414d83d4abf2ad8cc1446dc36c4e4d5", "code_archive": "repos/-NEXDKk8gZ.zip", "code_archive_sha256": "ca9e7beb876fdeaea50f514df3a33183828e62e9233a179b2dda2a60a2bbff15", "code_archive_bytes": 46777, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 30, "github_languages": {"Python": 129652}, "github_archived": false, "github_pushed_at": "2024-07-18T02:45:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improved-denoising-diffusion-probabilistic-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryljMpNtwr", "year": 2020, "status": "rejected", "title": "Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming", "authors": ["Claudio Michaelis", "Benjamin Mitzkus", "Robert Geirhos", "Evgenia Rusak", "Oliver Bringmann", "Alexander S. Ecker", "Matthias Bethge", "Wieland Brendel"], "authorids": ["claudio.michaelis@uni-tuebingen.de", "benjamin.mitzkus@uni-tuebingen.de", "robert@geirhos.de", "evgenia.rusak@bethgelab.org", "oliver.bringmann@uni-tuebingen.de", "alexander.ecker@uni-tuebingen.de", "matthias@bethgelab.org", "wieland.brendel@bethgelab.org"], "authors_source": "OpenReview API", "abstract": "The ability to detect objects regardless of image distortions or weather conditions is crucial for real-world applications of deep learning like autonomous driving. We here provide an easy-to-use benchmark to assess how object detection models perform when image quality degrades. The three resulting benchmark datasets, termed PASCAL-C, COCO-C and Cityscapes-C, contain a large variety of image corruptions. We show that a range of standard object detection models suffer a severe performance loss on corrupted images (down to 30-60% of the original performance). However, a simple data augmentation trick - stylizing the training images - leads to a substantial increase in robustness across corruption type, severity and dataset. We envision our comprehensive benchmark to track future progress towards building robust object detection models. Benchmark, code and data are available at: (hidden for double blind review)", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1gPC72CKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper424/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a benchmark for measuring robustness to input image corruption in object detection settings. The paper proposes a benchmark for this task, and proposes a simple data augmentation technique for this task.\n\nStrengths\n1. Understanding the robustness properties of existing vision models is an important problem.\n2. The paper establishes a sensible protocol for the benchmark, where methods are tested upon image perturbations that are not used for training the model.\n3. I like the proposed simple data augmentation procedure and the experimental finding that data augmentation with such a procedure leads to models that are robust to held-out, previously unseen perturbations.\n\nShortcomings:\n1. While the paper proposes a sensible experimental protocol, certain questions remain:\na) Are the set of test time perturbations exhaustive and representative of the perturbations in the real world? The paper doesn't talk about this, or provides any experimental data to establish this. The paper derives them from an earlier paper called \"Benchmarking neural network robustness to common corruptions and perturbations\", and thus I am not even sure if the proposed set of perturbations should be viewed as a contribution of the current paper.\nb) While the paper itself follows good practice by not training on perturbations that are considered at test time, unfortunately, it does not define a clear protocol or characterization as to how future researchers should use the benchmark. I believe setting up such a protocol is going to be difficult and is worthy of more thought and consideration, absence of this weakens the paper, as it leaves the door open for flawed future research.\n\n2. Missing comparisons: Proposed method is interesting, but I wonder if there were a more standard evaluation to test the efficiency of the method, perhaps something like testing if representations learned using such data augmentations were more robust to adversarial perturbations? Or perhaps, comparison against other methods that exist in literature for related tasks, such as methods that study how to make networks robust to adversarial perturbations?\n\nBecause of the aforementioned reasons, I don't view the benchmarking part of the paper as a solid contribution. Similarly, the proposed method is simple and intuitive (which is good), but it will help if there were more comparisons to set the paper in context of related work.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper presents a benchmark for measuring robustness to input image corruption in object detection settings. The paper proposes a benchmark for this task, and proposes a simple data augmentation technique for this task.\n\nStrengths\n1. Understanding the robustness properties of existing vision models is an important problem.\n2. The paper establishes a sensible protocol for the benchmark, where methods are tested upon image perturbations that are not used for training the model.\n3. I like the proposed simple data augmentation procedure and the experimental finding that data augmentation with such a procedure leads to models that are robust to held-out, previously unseen perturbations.\n\nShortcomings:\n1. While the paper proposes a sensible experimental protocol, certain questions remain:\na) Are the set of test time perturbations exhaustive and representative of the perturbations in the real world? The paper doesn't talk about this, or provides any experimental data to establish this. The paper derives them from an earlier paper called \"Benchmarking neural network robustness to common corruptions and perturbations\", and thus I am not even sure if the proposed set of perturbations should be viewed as a contribution of the current paper.\nb) While the paper itself follows good practice by not training on perturbations that are considered at test time, unfortunately, it does not define a clear protocol or characterization as to how future researchers should use the benchmark. I believe setting up such a protocol is going to be difficult and is worthy of more thought and consideration, absence of this weakens the paper, as it leaves the door open for flawed future research.\n\n2. Missing comparisons: Proposed method is interesting, but I wonder if there were a more standard evaluation to test the efficiency of the method, perhaps something like testing if representations learned using such data augmentations were more robust to adversarial perturbations? Or perhaps, comparison against other methods that exist in literature for related tasks, such as methods that study how to make networks robust to adversarial perturbations?\n\nBecause of the aforementioned reasons, I don't view the benchmarking part of the paper as a solid contribution. Similarly, the proposed method is simple and intuitive (which is good), but it will help if there were more comparisons to set the paper in context of related work."}, "tcdate": 1571894222745}, {"id": "Bkldg_0KKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper424/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a  benchmark to assess the performance of object detection models when image quality degrades. Three variants of detection datasets, termed PASCAL-C, COCO-C and Cityscapes-C, are introduced that contain a large variety of image corruptions. The paper shows that standard object detection models suffer a severe performance loss on corrupted images (down to 30–60% of the original performance). Further, this work shows that a simple data augmentation trick of stylizing the training images leads to a substantial increase in robustness across corruption type, severity and dataset.\n\nThe paper is well written and easy to follow. The proposed benchmark is interesting and clearly show the deficiencies of state-of-the-art object detection methods in case of image corruptions or weather conditions. However, my main concern is the novelty in that the proposed approach is just an extension of [1]. [1] introduced corrupted versions of commonly used classification datasets (ImageNet-C, CIFAR10-C) as standardized benchmarks. The different types of corruptions used here for object detection and their sorting into four groups were also introduced originally in [1]. Moreover, the idea to use style transfer as an augmentation to  improve corruption robustness for image classification has been introduced in [2]. Therefore, the only contribution of this paper is to apply the ideas from [1, 2] for object detection. \n\n[1] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019.\n[2] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper introduces a  benchmark to assess the performance of object detection models when image quality degrades. Three variants of detection datasets, termed PASCAL-C, COCO-C and Cityscapes-C, are introduced that contain a large variety of image corruptions. The paper shows that standard object detection models suffer a severe performance loss on corrupted images (down to 30–60% of the original performance). Further, this work shows that a simple data augmentation trick of stylizing the training images leads to a substantial increase in robustness across corruption type, severity and dataset.\n\nThe paper is well written and easy to follow. The proposed benchmark is interesting and clearly show the deficiencies of state-of-the-art object detection methods in case of image corruptions or weather conditions. However, my main concern is the novelty in that the proposed approach is just an extension of [1]. [1] introduced corrupted versions of commonly used classification datasets (ImageNet-C, CIFAR10-C) as standardized benchmarks. The different types of corruptions used here for object detection and their sorting into four groups were also introduced originally in [1]. Moreover, the idea to use style transfer as an augmentation to  improve corruption robustness for image classification has been introduced in [2]. Therefore, the only contribution of this paper is to apply the ideas from [1, 2] for object detection. \n\n[1] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019.\n[2] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019."}, "tcdate": 1571575791948}, {"id": "SJgukSgYFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper424/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n- key problem or question: assessing / improving the robustness of object detectors to image corruptions (simulated fog, frost, snow, dragonfire...);\n- contributions: 1) a benchmark (obtained by adding image-level corruptions to PASCAL, COCO, and Cityscapes) and an experimental protocol to measure detection robustness , 2) extensive experiments quantifying the severe lack of robustness of multiple state-of-the-art models, 3) experiments showing that data augmentation via style transfer (Geirhos et al, ICLR'19) improves robustness at little cost (at most -2% performance degradation on clean COCO images).\n\nRecommendation: Weak Reject\n\nKey reason: unclear novelty w.r.t. Geirhos et al ICLR 2019, especially due to the lack of specificity to object detection (which is the main goal of the paper).\n- Although the paper from Geirhos et al. study a specific question (texture bias in CNNs), they propose a similar experimental protocol for assessing robustness, where the only differences are i) they use different corruptions (cf. Fig.6 in their paper, vs. Figs. 5, 7, 8, 9 here), ii) they do not have Cityscapes results (but they have PASCAL and COCO results).\n- Furthermore, Geirhos et al 2019 also study the benefits of style-transfer-based data augmentation (this submission uses their technique), they report similar results and conclusions to this submission, including for object detection on PASCAL and COCO (cf. Table 2 in Geirhos et al 2019).\n- What is, in the authors' opinion, the main differentiator of this submission? Does the difference in corruptions and evaluated models combined with the addition of Cityscapes yield new insights compared to Geirhos et al 2019?\n- Beyond this similarity, what is specific to object detection vs. image classification in this submission? Besides the summary evaluation metrics, the corruptions and stylization are global (image-level) and not object-specific. Is the observed lack of robustness due to localization errors, mis-classifications, or other types of detection mistakes (cf. Hoiem et al ECCV'12)? What are the conclusions about detection robustness that differ from the image classification ones? What would be local corruptions that specifically degrade object detection performance (as studied for instance in the adversarial attack community, including physical attacks like Eykholt et al 2018)?\n\nAdditional feedback / questions:\n- The results in section 3.4 / Fig.6 seem counterintuitive: more corruption should yield more degradation (as evidenced in Fig. 5 w.r.t. corruption severity). Is RMSE the right metric? Maybe SSIM would be better (more related to perceptual quality)?\n- The aforementioned remark also raises the point that corruption difficulty night not be related to its intensity / noise level, but more with the (hard to quantify) domain gap w.r.t. clean data. For instance, if some images contain natural fog, fog corruption robustness should be naturally higher, whereas robust to never seen unrealistic dragonfire is expected to be naturally low. Can this relation between corruption and domain gap be somehow assessed? For instance using perceptual distance (using features from intermediate layers) to nearest clean neighbors? Or maybe by correlating with a subjective measure of realism of the corruption assessed globally per corruption type via a human study?\n- Are corruption degradations dataset-specific? Fig. 7,8,9 seem to show different behaviors.\n- How does the proposed benchmark compare to the Robust Vision Challenge (http://www.robustvision.net) proposed at CVPR 2018? How does robustness to corruptions and robustness across datasets correlate? Are they both similar \"out of domain\" robustness measures? How does this \"out of domain\" issue relate to adversarial examples (besides being \"less extreme\")?\n- Could Fig. 5 include error bars / variance across corruption types?\n- rPC is a good metric to compare models but I am not sure it is great to compare robustification methods, because when improving \"clean\" performance you mechanically decrease rPC, hence why combined is worse than stylized (e.g., in Table 2). What would be a better metric?\n- Why does the stylized approach on COCO yields worse mPC (cf. Table 2 and 4), whereas it is expected to improve robustness under corruption (and does on other datasets), esp. since it sacrifices performance on the clean images by a lot?\n- What is the impact of the choice of style images on robustness induced by stylization? Is diversity the most important factor (this can be tested by reducing the number of styles available)? If not, what is and how can it be measured?\n- Do certain corruption (at certain severity levels) result in unrecoverable objects? For instance, dragonfire might completely occlude certain objects on the side, which might be a problem (e.g., there are frequent parked cars on the side in Cityscapes). What is the upper bound after corruption and how can it be measured?\n- What is the human robustness for the new corruptions not present in Geirhos et al 2019? (Also relates to the aforementioned upper bound.)\n- The paper is well written and I enjoyed the multiple pop culture references to Game of Thrones.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Summary:\n- key problem or question: assessing / improving the robustness of object detectors to image corruptions (simulated fog, frost, snow, dragonfire...);\n- contributions: 1) a benchmark (obtained by adding image-level corruptions to PASCAL, COCO, and Cityscapes) and an experimental protocol to measure detection robustness , 2) extensive experiments quantifying the severe lack of robustness of multiple state-of-the-art models, 3) experiments showing that data augmentation via style transfer (Geirhos et al, ICLR'19) improves robustness at little cost (at most -2% performance degradation on clean COCO images).\n\nRecommendation: Weak Reject\n\nKey reason: unclear novelty w.r.t. Geirhos et al ICLR 2019, especially due to the lack of specificity to object detection (which is the main goal of the paper).\n- Although the paper from Geirhos et al. study a specific question (texture bias in CNNs), they propose a similar experimental protocol for assessing robustness, where the only differences are i) they use different corruptions (cf. Fig.6 in their paper, vs. Figs. 5, 7, 8, 9 here), ii) they do not have Cityscapes results (but they have PASCAL and COCO results).\n- Furthermore, Geirhos et al 2019 also study the benefits of style-transfer-based data augmentation (this submission uses their technique), they report similar results and conclusions to this submission, including for object detection on PASCAL and COCO (cf. Table 2 in Geirhos et al 2019).\n- What is, in the authors' opinion, the main differentiator of this submission? Does the difference in corruptions and evaluated models combined with the addition of Cityscapes yield new insights compared to Geirhos et al 2019?\n- Beyond this similarity, what is specific to object detection vs. image classification in this submission? Besides the summary evaluation metrics, the corruptions and stylization are global (image-level) and not object-specific. Is the observed lack of robustness due to localization errors, mis-classifications, or other types of detection mistakes (cf. Hoiem et al ECCV'12)? What are the conclusions about detection robustness that differ from the image classification ones? What would be local corruptions that specifically degrade object detection performance (as studied for instance in the adversarial attack community, including physical attacks like Eykholt et al 2018)?\n\nAdditional feedback / questions:\n- The results in section 3.4 / Fig.6 seem counterintuitive: more corruption should yield more degradation (as evidenced in Fig. 5 w.r.t. corruption severity). Is RMSE the right metric? Maybe SSIM would be better (more related to perceptual quality)?\n- The aforementioned remark also raises the point that corruption difficulty night not be related to its intensity / noise level, but more with the (hard to quantify) domain gap w.r.t. clean data. For instance, if some images contain natural fog, fog corruption robustness should be naturally higher, whereas robust to never seen unrealistic dragonfire is expected to be naturally low. Can this relation between corruption and domain gap be somehow assessed? For instance using perceptual distance (using features from intermediate layers) to nearest clean neighbors? Or maybe by correlating with a subjective measure of realism of the corruption assessed globally per corruption type via a human study?\n- Are corruption degradations dataset-specific? Fig. 7,8,9 seem to show different behaviors.\n- How does the proposed benchmark compare to the Robust Vision Challenge (http://www.robustvision.net) proposed at CVPR 2018? How does robustness to corruptions and robustness across datasets correlate? Are they both similar \"out of domain\" robustness measures? How does this \"out of domain\" issue relate to adversarial examples (besides being \"less extreme\")?\n- Could Fig. 5 include error bars / variance across corruption types?\n- rPC is a good metric to compare models but I am not sure it is great to compare robustification methods, because when improving \"clean\" performance you mechanically decrease rPC, hence why combined is worse than stylized (e.g., in Table 2). What would be a better metric?\n- Why does the stylized approach on COCO yields worse mPC (cf. Table 2 and 4), whereas it is expected to improve robustness under corruption (and does on other datasets), esp. since it sacrifices performance on the clean images by a lot?\n- What is the impact of the choice of style images on robustness induced by stylization? Is diversity the most important factor (this can be tested by reducing the number of styles available)? If not, what is and how can it be measured?\n- Do certain corruption (at certain severity levels) result in unrecoverable objects? For instance, dragonfire might completely occlude certain objects on the side, which might be a problem (e.g., there are frequent parked cars on the side in Cityscapes). What is the upper bound after corruption and how can it be measured?\n- What is the human robustness for the new corruptions not present in Geirhos et al 2019? (Also relates to the aforementioned upper bound.)\n- The paper is well written and I enjoyed the multiple pop culture references to Game of Thrones."}, "tcdate": 1571517663699}], "openreview_url": "https://openreview.net/forum?id=ryljMpNtwr", "arxiv_id": "1907.07484", "paper_pdf": "papers/ryljMpNtwr.pdf", "paper_pdf_sha256": "1b731982ccdb0e246ea6a9c3a8df69dc5e828644eeb8ada3f6711e30ce1dc2a0", "paper_pdf_bytes": 7695760, "paper_pdf_source": "openreview", "code_url": "https://github.com/bethgelab/stylize-datasets", "code_repository": "bethgelab/stylize-datasets", "code_commit": "71def61edba22bfec690eb461e2bb80a55e326f7", "code_archive": "repos/ryljMpNtwr.zip", "code_archive_sha256": "6a7798366da6adf178fd77033d931a2f0cd7adc2a0509edbf08b1cf6d31798cc", "code_archive_bytes": 8351, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 29, "github_languages": {"Python": 14900}, "github_archived": false, "github_pushed_at": "2021-07-05T17:29:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/benchmarking-robustness-in-object-detection"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkVRTj0cYQ", "year": 2019, "status": "rejected", "title": "Differentially Private Federated Learning: A Client Level Perspective", "authors": ["Robin C. Geyer", "Tassilo J. Klein", "Moin Nabi"], "authorids": ["geyerr@ethz.ch", "tassilo.klein@sap.com", "moin.nabi@sap.com"], "authors_source": "OpenReview API", "abstract": "Federated learning is a recent advance in privacy protection. \nIn this context, a trusted curator aggregates parameters optimized in decentralized fashion by multiple clients. The resulting model is then distributed back to all clients, ultimately converging to a joint representative model without explicitly having to share the data. \nHowever, the protocol is vulnerable to differential attacks, which could originate from any party contributing during federated optimization. In such an attack, a client's contribution during training and information about their data set is revealed through analyzing the distributed model. \nWe tackle this problem and propose an algorithm for client sided differential privacy preserving federated optimization. The aim is to hide clients' contributions during training, balancing the trade-off between privacy loss and model performance. \nEmpirical studies suggest that given a sufficiently large number of participating clients, our proposed procedure can maintain client-level differential privacy at only a minor cost in model performance. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "r1g9TUss3m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper853/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper revisits the federated learning framework from McMahan in the context of differential privacy.  The general concern with the vanilla federated learning framework is that it is susceptible to differencing attacks. To that end, the paper proposes to make the each of the interaction in the server-side component of the gradient descent to be differentially private w.r.t. the client contributions. This is simply done by adding noise (appropriately scaled) to the gradient updates.\n\nMy main concern is that the paper just described differentially private SGD, in the language of federated learning. I could not find any novelty in the approach. Furthermore, just using the vanilla moment's accountant to track privacy depletion in the federated setting is not totally correct. The moment's accountant framework in Abadi et al. uses the \"secrecy of the sample\" property to boost the privacy guarantee in a particular iteration. However, in the federated setting, the boost via secrecy of the sample does not hold immediately. One requirement of the secrecy of the sample theorem is that the sampled client has to be hidden. However, in the federated setting, even if one does not know what information a client sends to the servery, one can always observe if the client is sending *any* information. For a detailed discussion on this issue see https://arxiv.org/abs/1808.06651 .", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Differentially private variant of the federated learning framework", "review": "The paper revisits the federated learning framework from McMahan in the context of differential privacy.  The general concern with the vanilla federated learning framework is that it is susceptible to differencing attacks. To that end, the paper proposes to make the each of the interaction in the server-side component of the gradient descent to be differentially private w.r.t. the client contributions. This is simply done by adding noise (appropriately scaled) to the gradient updates.\n\nMy main concern is that the paper just described differentially private SGD, in the language of federated learning. I could not find any novelty in the approach. Furthermore, just using the vanilla moment's accountant to track privacy depletion in the federated setting is not totally correct. The moment's accountant framework in Abadi et al. uses the \"secrecy of the sample\" property to boost the privacy guarantee in a particular iteration. However, in the federated setting, the boost via secrecy of the sample does not hold immediately. One requirement of the secrecy of the sample theorem is that the sampled client has to be hidden. However, in the federated setting, even if one does not know what information a client sends to the servery, one can always observe if the client is sending *any* information. For a detailed discussion on this issue see https://arxiv.org/abs/1808.06651 .", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541285569930}, {"id": "r1eXjyItn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper853/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "[Post-rebuttal update] No author response was provided to address the reviewer comments. In particular, the paper's contributions and novelty compared with previous work seem limited, and no author response was provided to address this concern. I've left my overall score for the paper unchanged.\n\n[Summary] The authors propose a protocol for training a model over private user data in a federated setting. In contrast with previous approaches which tried to ensure that a model would not reveal too much about any individual data point, this paper aims to prevent leakage of information about any individual client. (There may be many data points associated with a single client.)\n\n[Key Comments] The submission generally seems polished and well-written. However, I have the impression that it's largely an incremental improvement over recent work by McMahan et al. (2018).\n* If the main improvement of this paper over previous work is the dynamic adaptation of weight updates discussed in Section 3, the experimental results in Table 1 should compare the performance of the protocol with vs. without these changes. Otherwise, I think it would be helpful for the authors to update the submission to clarify their contributions.\n* Updating Algorithm 1 / Line 9 (computation of the median weight update norm) to avoid leaking sensitive information to the clients would also strengthen the submission.\n* It would also be helpful if the authors could explicitly list their assumptions about which parties are trusted and which are not (see below).\n\n[Details]\n[Pro 1] The submission is generally well-written and polished. I found the beginning of Section 3 especially helpful, since it breaks down a complex algorithm into simple/understandable parts.\n\n[Pro 2] The proposed algorithm tackles the challenging/well-motivated problem of improving federated machine learning with strong theoretical privacy guarantees.\n\n[Pro 3] Section 6 has an interesting analysis of how the weight updates produced by clients change over the course of training. This section does a good job of setting up the intuition for the training setup used in the paper, where the number of clients used in each round is gradually increased over the course of training.\n \n[Con 1] I had trouble understanding the precise threat model used in the paper, and I think it would be helpful if the authors could update their submission to explicitly list their assumptions in one place. It seems like the server is trusted while the clients are not. However, I was unsure whether the goal was to protect against a single honest-but-curious client or to protect against multiple (possibly colluding) clients.\n\n[Con 2] During each round of communication, the protocol computes the median of a set of values, each one originating from a different client, and the output of this computation is used to perform weight updates which are sent back to the clients. The authors note that \"we do not use a randomized mechanism for computing the median, which, strictly speaking, is a violation of privacy. However, the information leakage through the median is small (future work will contain such privacy measures).\" I appreciate the authors' honesty and thoroughness in pointing out this limitation. However, it does make the submission feel like a work in progress rather than a finished paper, and I think that the submission would be a bit stronger if this issue was addressed.\n\n[Con 3] Given the experimental results reported in Section 4, it's difficult for me to understand how much of an improvement the authors' proposed dynamic weight updates provide in practice. This concern could be addressed with the inclusion of additional details and baselines:\n* Few details are provided about the model training setup, and the reported accuracy of the non-differentially private model is quite low (3% reported error rate on MNIST; it's straightforward to get 1% error or below with a modern convolutional neural network). The authors say they use a setup similar to previous work by McMahan et al. (2017), but it seems like that paper uses a model with a much lower error rate (less than 1% based on a cursory inspection), which makes direct comparisons difficult.\n* The introduction argues that \"dynamically adapting the dp-preserving mechanism during decentralized training\" is a significant difference from previous work. The claim could be strengthened if the authors extended Table 1 (experimental results for differentially private federated learning) in order to demonstrate the effect of dynamic adaptation on model quality.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well-motivated problem, but incremental improvement over previous work?", "review": "[Post-rebuttal update] No author response was provided to address the reviewer comments. In particular, the paper's contributions and novelty compared with previous work seem limited, and no author response was provided to address this concern. I've left my overall score for the paper unchanged.\n\n[Summary] The authors propose a protocol for training a model over private user data in a federated setting. In contrast with previous approaches which tried to ensure that a model would not reveal too much about any individual data point, this paper aims to prevent leakage of information about any individual client. (There may be many data points associated with a single client.)\n\n[Key Comments] The submission generally seems polished and well-written. However, I have the impression that it's largely an incremental improvement over recent work by McMahan et al. (2018).\n* If the main improvement of this paper over previous work is the dynamic adaptation of weight updates discussed in Section 3, the experimental results in Table 1 should compare the performance of the protocol with vs. without these changes. Otherwise, I think it would be helpful for the authors to update the submission to clarify their contributions.\n* Updating Algorithm 1 / Line 9 (computation of the median weight update norm) to avoid leaking sensitive information to the clients would also strengthen the submission.\n* It would also be helpful if the authors could explicitly list their assumptions about which parties are trusted and which are not (see below).\n\n[Details]\n[Pro 1] The submission is generally well-written and polished. I found the beginning of Section 3 especially helpful, since it breaks down a complex algorithm into simple/understandable parts.\n\n[Pro 2] The proposed algorithm tackles the challenging/well-motivated problem of improving federated machine learning with strong theoretical privacy guarantees.\n\n[Pro 3] Section 6 has an interesting analysis of how the weight updates produced by clients change over the course of training. This section does a good job of setting up the intuition for the training setup used in the paper, where the number of clients used in each round is gradually increased over the course of training.\n \n[Con 1] I had trouble understanding the precise threat model used in the paper, and I think it would be helpful if the authors could update their submission to explicitly list their assumptions in one place. It seems like the server is trusted while the clients are not. However, I was unsure whether the goal was to protect against a single honest-but-curious client or to protect against multiple (possibly colluding) clients.\n\n[Con 2] During each round of communication, the protocol computes the median of a set of values, each one originating from a different client, and the output of this computation is used to perform weight updates which are sent back to the clients. The authors note that \"we do not use a randomized mechanism for computing the median, which, strictly speaking, is a violation of privacy. However, the information leakage through the median is small (future work will contain such privacy measures).\" I appreciate the authors' honesty and thoroughness in pointing out this limitation. However, it does make the submission feel like a work in progress rather than a finished paper, and I think that the submission would be a bit stronger if this issue was addressed.\n\n[Con 3] Given the experimental results reported in Section 4, it's difficult for me to understand how much of an improvement the authors' proposed dynamic weight updates provide in practice. This concern could be addressed with the inclusion of additional details and baselines:\n* Few details are provided about the model training setup, and the reported accuracy of the non-differentially private model is quite low (3% reported error rate on MNIST; it's straightforward to get 1% error or below with a modern convolutional neural network). The authors say they use a setup similar to previous work by McMahan et al. (2017), but it seems like that paper uses a model with a much lower error rate (less than 1% based on a cursory inspection), which makes direct comparisons difficult.\n* The introduction argues that \"dynamically adapting the dp-preserving mechanism during decentralized training\" is a significant difference from previous work. The claim could be strengthened if the authors extended Table 1 (experimental results for differentially private federated learning) in order to demonstrate the effect of dynamic adaptation on model quality.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541132187058}, {"id": "SkenUlAliQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper853/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The main claim the authors make is that providing privacy in learning should go beyond just privacy for individual records to providing privacy for data contributors which could be an entire hospital. Adding privacy by design to the machine learning pipe-line is an important topic. Unfortunately, the presentation of this paper makes it hard to follow. \n\nSome of the issues in this paper are technical and easy to resolve, such as citation format (see below) or consistency of notation (see below). Another example is that although the method presented here is suitable only for gradient based learning this is not stated clearly. However, other issues are more fundamental:\n1.\tThe main motivation for this work is providing privacy to a client which could be a hospital as opposed to providing privacy to a single record – why is that an important task? Moreover, there are standard ways to extend differential privacy from a single record to a set of r records (see dwork & Rote, 2014 Theorem 2.2), in what sense the method presented here different than these methods?\n2.\tAnother issue with the hospitals motivation is that the results show that when the number of parties is 10,000 the accuracy is close to the baseline. However, there are only 5534 registered hospitals in the US in 2018 according to the American Hospital Association (AHA): https://www.aha.org/statistics/fast-facts-us-hospitals. Therefore, are the sizes used in the experiments reasonable?\n3.\tIn the presentation of the methods, it is not clear what is novel and what was already done by Abadi et al., 2016\n4.\tThe theoretical analysis of the algorithm is only implied and not stated clearly\n5.\tIn reporting the experiment setup key pieces of information are missing which makes the experiment irreproducible. For example, what is the leaning algorithm used? If it is a neural network, what was its layout? What type of cross validation was used to tune parameters?\n6.\tIn describing the experiment it says that “For K\\in\\{1000,10000} data points are repeated.” This could mean that a single client holds the same point multiple times or that multiple clients hold the same data point. Which one of them is correct? What are the implications of that on the results of the experiment?\n7.\tSince grid search is used to tune parameters, more information is leaking which is not compensated for by, for example, composition bounds\n8.\tThe results of the experiments are not contrasted against prior art, for example the results of Abadi et al., 2016.\n\nAdditional comments\n9.\tThe introduction is confusing since it uses the term “federated learning” as a privacy technology. However federated learning discusses the scenario where the data is distributed between several parties. It is not necessarily the case that there are also privacy concerns associated, in many cases the need for federated learning is due to performance constraints.\n10.\tIn the abstract the term “differential attacks” is used – what does it mean?\n11.\t“An independent study McMahan et al. (2018), published at the same time”- since you refer to the work of McMahan et al before your paper was reviewed, it means that the work of McMahan et al came out earlier.\n12.\tIn the section “Choosing $\\sigma$ and $m$” it is stated that the higher \\sigma and the lower m, the higher the privacy loss. Isn’t the privacy loss reduced when \\sigma is larger? Moreover, since you divide the gradients by m_t then the sensitivity of each party is of the order of S/m and therefore it reduces as m gets larger, hence, the privacy loss is smaller when m is large. \n13.\tAt the bottom of page 4 and top of page 5 you introduce variance related terms that are never used in the algorithm or any analysis (they are presented in Figure 3). The variance between clients can be a function of how the data is split between them. If, for example, each client represents a different demography then the variance may be larger from the beginning.\n14.\tIn the experiments (Table 1), what does it mean for \\delta^\\prime to be e-3, e-5 or e-6? Is it 10^{-3}, 10^{-5} and 10^{-6}?\n15.\tThe methods presented here apply only for gradient descent learning algorithms, but this is not stated clearly. For example, would the methods presented here apply for learning tree based models?\n16.\tThe citations are used incorrectly, for example “sometimes referred to as collaborative Shokri & Shmatikov (2015)” should be “sometimes referred to as collaborative (Shokri & Shmatikov, 2015)”. This can be achieved by using \\citep in latex. This problem appears in many places in the paper, including, for example, “we make use of the moments accountant as proposed by Abadi et al. Abadi et al. (2016).” Which should be “we make use of the moments accountant as proposed by Abadi et al. (2016).” In which case you should use only \\cite and not quote the name in the .tex file.\n17.\t“We use the same definition for differential privacy in randomized mechanisms as Abadi et al. (2016):” – the definition of differential privacy is due to Dwork, McSherry, Nissim & Smith, 2006\n18.\tNotation is followed loosely which makes it harder to follow at parts. For example, you use “m_t” for the number of participants in time t but in some cases,  you use only m as in “Choosing $\\sigma$ and $m$”.\n19.\tIn algorithm 1 the function ClientUpdate receives two parameters however the first parameter is never used in this function. \n20.\tFigure 2: I think it would be easier to see the results if you use log-log plot\n21.\tDiscussion: “For K=10000, the differrntially private model almost reaches accuracies of the non-differential private one.” – it is true that the model used in this experiment achieves an accuracy of 0.97 without DP and the reported number for K=10000 is 0.96 which is very close. However, the baseline accuracy of 0.97 is very low for MNIST.\n22.\tIn the bibliography you have Brendan McMahan appearing both as Brendan McMahan and H. Brendan McMahan\n\n\nIt is possible that underneath that this work has some hidden jams, however, the presentation makes them hard to find. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting direction but confusing presentation", "review": "The main claim the authors make is that providing privacy in learning should go beyond just privacy for individual records to providing privacy for data contributors which could be an entire hospital. Adding privacy by design to the machine learning pipe-line is an important topic. Unfortunately, the presentation of this paper makes it hard to follow. \n\nSome of the issues in this paper are technical and easy to resolve, such as citation format (see below) or consistency of notation (see below). Another example is that although the method presented here is suitable only for gradient based learning this is not stated clearly. However, other issues are more fundamental:\n1.\tThe main motivation for this work is providing privacy to a client which could be a hospital as opposed to providing privacy to a single record – why is that an important task? Moreover, there are standard ways to extend differential privacy from a single record to a set of r records (see dwork & Rote, 2014 Theorem 2.2), in what sense the method presented here different than these methods?\n2.\tAnother issue with the hospitals motivation is that the results show that when the number of parties is 10,000 the accuracy is close to the baseline. However, there are only 5534 registered hospitals in the US in 2018 according to the American Hospital Association (AHA): https://www.aha.org/statistics/fast-facts-us-hospitals. Therefore, are the sizes used in the experiments reasonable?\n3.\tIn the presentation of the methods, it is not clear what is novel and what was already done by Abadi et al., 2016\n4.\tThe theoretical analysis of the algorithm is only implied and not stated clearly\n5.\tIn reporting the experiment setup key pieces of information are missing which makes the experiment irreproducible. For example, what is the leaning algorithm used? If it is a neural network, what was its layout? What type of cross validation was used to tune parameters?\n6.\tIn describing the experiment it says that “For K\\in\\{1000,10000} data points are repeated.” This could mean that a single client holds the same point multiple times or that multiple clients hold the same data point. Which one of them is correct? What are the implications of that on the results of the experiment?\n7.\tSince grid search is used to tune parameters, more information is leaking which is not compensated for by, for example, composition bounds\n8.\tThe results of the experiments are not contrasted against prior art, for example the results of Abadi et al., 2016.\n\nAdditional comments\n9.\tThe introduction is confusing since it uses the term “federated learning” as a privacy technology. However federated learning discusses the scenario where the data is distributed between several parties. It is not necessarily the case that there are also privacy concerns associated, in many cases the need for federated learning is due to performance constraints.\n10.\tIn the abstract the term “differential attacks” is used – what does it mean?\n11.\t“An independent study McMahan et al. (2018), published at the same time”- since you refer to the work of McMahan et al before your paper was reviewed, it means that the work of McMahan et al came out earlier.\n12.\tIn the section “Choosing $\\sigma$ and $m$” it is stated that the higher \\sigma and the lower m, the higher the privacy loss. Isn’t the privacy loss reduced when \\sigma is larger? Moreover, since you divide the gradients by m_t then the sensitivity of each party is of the order of S/m and therefore it reduces as m gets larger, hence, the privacy loss is smaller when m is large. \n13.\tAt the bottom of page 4 and top of page 5 you introduce variance related terms that are never used in the algorithm or any analysis (they are presented in Figure 3). The variance between clients can be a function of how the data is split between them. If, for example, each client represents a different demography then the variance may be larger from the beginning.\n14.\tIn the experiments (Table 1), what does it mean for \\delta^\\prime to be e-3, e-5 or e-6? Is it 10^{-3}, 10^{-5} and 10^{-6}?\n15.\tThe methods presented here apply only for gradient descent learning algorithms, but this is not stated clearly. For example, would the methods presented here apply for learning tree based models?\n16.\tThe citations are used incorrectly, for example “sometimes referred to as collaborative Shokri & Shmatikov (2015)” should be “sometimes referred to as collaborative (Shokri & Shmatikov, 2015)”. This can be achieved by using \\citep in latex. This problem appears in many places in the paper, including, for example, “we make use of the moments accountant as proposed by Abadi et al. Abadi et al. (2016).” Which should be “we make use of the moments accountant as proposed by Abadi et al. (2016).” In which case you should use only \\cite and not quote the name in the .tex file.\n17.\t“We use the same definition for differential privacy in randomized mechanisms as Abadi et al. (2016):” – the definition of differential privacy is due to Dwork, McSherry, Nissim & Smith, 2006\n18.\tNotation is followed loosely which makes it harder to follow at parts. For example, you use “m_t” for the number of participants in time t but in some cases,  you use only m as in “Choosing $\\sigma$ and $m$”.\n19.\tIn algorithm 1 the function ClientUpdate receives two parameters however the first parameter is never used in this function. \n20.\tFigure 2: I think it would be easier to see the results if you use log-log plot\n21.\tDiscussion: “For K=10000, the differrntially private model almost reaches accuracies of the non-differential private one.” – it is true that the model used in this experiment achieves an accuracy of 0.97 without DP and the reported number for K=10000 is 0.96 which is very close. However, the baseline accuracy of 0.97 is very low for MNIST.\n22.\tIn the bibliography you have Brendan McMahan appearing both as Brendan McMahan and H. Brendan McMahan\n\n\nIt is possible that underneath that this work has some hidden jams, however, the presentation makes them hard to find. \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1539526740246}], "openreview_url": "https://openreview.net/forum?id=SkVRTj0cYQ", "arxiv_id": "1712.07557", "paper_pdf": "papers/SkVRTj0cYQ.pdf", "paper_pdf_sha256": "d4c626e4614a1e561a6c9911687592fb0cd1beb3917d623357cbe5d3eceb842c", "paper_pdf_bytes": 615031, "paper_pdf_source": "openreview", "code_url": "https://github.com/SAP-archive/machine-learning-diff-private-federated-learning", "code_repository": "SAP-archive/machine-learning-diff-private-federated-learning", "code_commit": "915c2fa63751403cbfcb910c101eb95377f65f19", "code_archive": "repos/SkVRTj0cYQ.zip", "code_archive_sha256": "c3f25c1b799b20511d0392678d0fa49deddc0861c11583bbd352dc88a2d1818d", "code_archive_bytes": 45061, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 85, "github_languages": {"Python": 89358, "Shell": 1104}, "github_archived": true, "github_pushed_at": "2025-03-07T12:53:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/differentially-private-federated-learning-a"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HknbyQbC-", "year": 2018, "status": "rejected", "title": "Generating Adversarial Examples with Adversarial Networks", "authors": ["Chaowei Xiao", "Bo Li", "Jun-Yan Zhu", "Warren He", "Mingyan Liu", "Dawn Song"], "authorids": ["xiaocw@umich.edu", "lxbosky@gmail.com", "junyanz@berkeley.edu", "_w@eecs.berkeley.edu", "mingyan@umich.edu", "dawnsong.travel@gmail.com"], "authors_source": "OpenReview API", "abstract": "Deep neural networks (DNNs) have been found to be vulnerable to adversarial examples resulting from adding small-magnitude perturbations to inputs. Such adversarial examples can mislead DNNs to produce adversary-selected results.\nDifferent attack strategies have been proposed to generate adversarial examples, but how to produce them with high perceptual quality and more efficiently requires more research efforts. \nIn this paper, we propose AdvGAN to generate adversarial examples with generative adversarial networks (GANs), which can learn and approximate the distribution of original instances. \nFor AdvGAN, once the generator is trained, it can generate adversarial perturbations efficiently for any instance, so as to potentially accelerate adversarial training as defenses.  \nWe apply AdvGAN in both semi-whitebox and black-box attack settings. In semi-whitebox attacks, there is no need to access the original target model after the generator is trained, in contrast to traditional white-box attacks. In black-box attacks, we dynamically train a distilled model for the black-box model and optimize the generator accordingly.\nAdversarial examples generated by AdvGAN on different target models have high attack success rate under state-of-the-art defenses compared to other attacks. Our attack  has placed the first with 92.76% accuracy on a public MNIST black-box attack challenge. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJgWlg6yM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1078/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I thank the authors for the thoughtful response and rebuttal. The authors have substantially updated their manuscript and improved the presentation.\n\nRe: Speed. I brought up this point because this was a bulleted item in the Introduction in the earlier version of the manuscript. In the revised manuscript, this bullet point is now removed. I will take this point to be moot.\n\nRe: High resolution. The authors point to recent GAN literature that provides some first results with high resolution GANs but I do not see quantitative evidence in the high resolution setting for this paper. (Figure 4 provides qualitative examples from ImageNet but no quantitative assessment.)\n\nBecause the authors improved the manuscript, I upwardly revised my score to 'Ok but not good enough - rejection'. I am not able to accept this paper because of the latter point.\n==========================\n\nThe authors present an interesting new method for generating adversarial examples. Namely, the author train a generative adversarial network (GAN) to adversarial examples for a target network. The authors demonstrate that the network works well in the semi-white box and black box settings.\n\nThe authors wrote a clear paper with great references and clear descriptions.\n\nMy primary concern is that this work has limited practical benefit in a realistic setting. Addressing each and every concern is quite important:\n\n1) Speed. The authors suggest that training a GAN provides a speed benefit with respect to other attack techniques. The FGSM method (Goodfellow et al, 2015) is basically 1 inference operation and 1 backward operation. The GAN is 1 forward operation. Granted this results in a small difference in timing 0.06s versus 0.01s, however it would seem that avoiding a backward pass is a somewhat small speed gain.\n \nFurthermore, I would want to question the practical usage of having an 'even faster' method for generating adversarial examples. What is the reason that we need to run adversarial attacks 'even faster'? I am not aware of any use-cases, but if there are some, the authors should describe the rationales at length in their paper.\n\n2) High spatial resolution images. Previous methods, e.g. FGSM, may work on arbitrarily sized images. At best, GANs generate reasonable images that are lower resolutions (e.g. < 128x128). Building GAN's that operate above-and-beyond moderate spatial resolution is an open research topic. The best GAN models for generating high resolution images are  difficult to train and it is not clear if they would work in this setting. Furthermore, images with even higher resolutions, e.g. 512x512, which is quite common in ImageNet, are difficult to synthesizes using current techniques.\n\n3) Controlling the amount of distortion. A feature of previous optimization based methods is that a user may specify the amount of perturbation (epsilon). This is a key feature if not requirement in an adversarial perturbation because a user might want to examine the performance of a given model as a function of epsilon. Performing such an analysis with this model is challenging (i.e. retraining a GAN) and it is not clear if a given image generated by a GAN will always achieve a given epsilon perturbation/\n\nOn a more minor note, the authors suggest that generating a *diversity* of adversarial images is of practical import. I do not see the utility of being able to generate a diversity of adversarial images. The authors need to provide more justification for this motivation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for 'Generating Adversarial Examples with Adversarial Networks'", "rating": "4: Ok but not good enough - rejection", "review": "I thank the authors for the thoughtful response and rebuttal. The authors have substantially updated their manuscript and improved the presentation.\n\nRe: Speed. I brought up this point because this was a bulleted item in the Introduction in the earlier version of the manuscript. In the revised manuscript, this bullet point is now removed. I will take this point to be moot.\n\nRe: High resolution. The authors point to recent GAN literature that provides some first results with high resolution GANs but I do not see quantitative evidence in the high resolution setting for this paper. (Figure 4 provides qualitative examples from ImageNet but no quantitative assessment.)\n\nBecause the authors improved the manuscript, I upwardly revised my score to 'Ok but not good enough - rejection'. I am not able to accept this paper because of the latter point.\n==========================\n\nThe authors present an interesting new method for generating adversarial examples. Namely, the author train a generative adversarial network (GAN) to adversarial examples for a target network. The authors demonstrate that the network works well in the semi-white box and black box settings.\n\nThe authors wrote a clear paper with great references and clear descriptions.\n\nMy primary concern is that this work has limited practical benefit in a realistic setting. Addressing each and every concern is quite important:\n\n1) Speed. The authors suggest that training a GAN provides a speed benefit with respect to other attack techniques. The FGSM method (Goodfellow et al, 2015) is basically 1 inference operation and 1 backward operation. The GAN is 1 forward operation. Granted this results in a small difference in timing 0.06s versus 0.01s, however it would seem that avoiding a backward pass is a somewhat small speed gain.\n \nFurthermore, I would want to question the practical usage of having an 'even faster' method for generating adversarial examples. What is the reason that we need to run adversarial attacks 'even faster'? I am not aware of any use-cases, but if there are some, the authors should describe the rationales at length in their paper.\n\n2) High spatial resolution images. Previous methods, e.g. FGSM, may work on arbitrarily sized images. At best, GANs generate reasonable images that are lower resolutions (e.g. < 128x128). Building GAN's that operate above-and-beyond moderate spatial resolution is an open research topic. The best GAN models for generating high resolution images are  difficult to train and it is not clear if they would work in this setting. Furthermore, images with even higher resolutions, e.g. 512x512, which is quite common in ImageNet, are difficult to synthesizes using current techniques.\n\n3) Controlling the amount of distortion. A feature of previous optimization based methods is that a user may specify the amount of perturbation (epsilon). This is a key feature if not requirement in an adversarial perturbation because a user might want to examine the performance of a given model as a function of epsilon. Performing such an analysis with this model is challenging (i.e. retraining a GAN) and it is not clear if a given image generated by a GAN will always achieve a given epsilon perturbation/\n\nOn a more minor note, the authors suggest that generating a *diversity* of adversarial images is of practical import. I do not see the utility of being able to generate a diversity of adversarial images. The authors need to provide more justification for this motivation.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1510961143839}, {"id": "S1gL2gTlf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1078/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a way of generating adversarial examples that fool classification systems.\nThey formulate it for a blackbox and a semi-blackbox setting (semi being, needed for training their own network, but not to generate new samples).\n\nThe model is a residual gan formulation, where the generator generates an image mask M, and (Input + M) is the adversarial example.\nThe paper is generally easy to understand and clear in their results.\nI am not awfully familiar with the literature on adversarial examples to know if other GAN variants exist. From this paper's literature survey, they dont exist. \nSo this paper is innovative in two parts:\n- it applies GANs to adversarial example generation\n- the method is a simple feed-forward network, so it is very fast to compute\n\nThe experiments are pretty robust, and they show that their method is better than the proposed baselines.\nI am not sure if these are complete baselines or if the baselines need to cover other methods (again, not fully familiar with all literature here).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "adversarial adversial example generation, wins MadryLab's mnist challenge", "rating": "7: Good paper, accept", "review": "The paper proposes a way of generating adversarial examples that fool classification systems.\nThey formulate it for a blackbox and a semi-blackbox setting (semi being, needed for training their own network, but not to generate new samples).\n\nThe model is a residual gan formulation, where the generator generates an image mask M, and (Input + M) is the adversarial example.\nThe paper is generally easy to understand and clear in their results.\nI am not awfully familiar with the literature on adversarial examples to know if other GAN variants exist. From this paper's literature survey, they dont exist. \nSo this paper is innovative in two parts:\n- it applies GANs to adversarial example generation\n- the method is a simple feed-forward network, so it is very fast to compute\n\nThe experiments are pretty robust, and they show that their method is better than the proposed baselines.\nI am not sure if these are complete baselines or if the baselines need to cover other methods (again, not fully familiar with all literature here).\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1512012871714}, {"id": "BJjc-tsef", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1078/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes AdvGAN, a conditional GAN plus adversarial loss. AdvGAN is able to generate adversarial samples by running a forward pass on generator. The authors evaluate AdvGAN on semi-white box and black box setting.\n\nAdvGAN is a simple and neat solution to for generating adversary samples. The author also reports state-of-art results.\n\nComment:\n\n1. For MNIST samples, we can easily find the generated sample is a mixture of two digitals. Eg, for digital 7 there is a light gray 3 overlap. I am wondering this method is trying to mixture several samples into one to generate adversary samples. For real color samples, it is harder to figure out the mixture.\n2. Based on mixture assumption, I suggest the author add one more comparison to other method, which is relative change from original image, to see whether AdvGAN is the most efficient model to generate the adversary sample (makes minimal change to original image).\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "rating": "6: Marginally above acceptance threshold", "review": "This paper describes AdvGAN, a conditional GAN plus adversarial loss. AdvGAN is able to generate adversarial samples by running a forward pass on generator. The authors evaluate AdvGAN on semi-white box and black box setting.\n\nAdvGAN is a simple and neat solution to for generating adversary samples. The author also reports state-of-art results.\n\nComment:\n\n1. For MNIST samples, we can easily find the generated sample is a mixture of two digitals. Eg, for digital 7 there is a light gray 3 overlap. I am wondering this method is trying to mixture several samples into one to generate adversary samples. For real color samples, it is harder to figure out the mixture.\n2. Based on mixture assumption, I suggest the author add one more comparison to other method, which is relative change from original image, to see whether AdvGAN is the most efficient model to generate the adversary sample (makes minimal change to original image).\n\n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511915923451}], "openreview_url": "https://openreview.net/forum?id=HknbyQbC-", "arxiv_id": "1801.02610", "paper_pdf": "papers/HknbyQbC-.pdf", "paper_pdf_sha256": "01826ca59d53cf9ea824554bb01a1c58a1323299e41066502d8813c45472ede3", "paper_pdf_bytes": 8716232, "paper_pdf_source": "openreview", "code_url": "https://github.com/MadryLab/mnist_challenge", "code_repository": "MadryLab/mnist_challenge", "code_commit": "3ee3643c4a8c59458d8c191b84027f4a6cbd9580", "code_archive": "repos/HknbyQbC-.zip", "code_archive_sha256": "024f20d6670726da3170f4af9a1ff05e79c1f98296c97e711840c29ee86cfe4f", "code_archive_bytes": 15328, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 57, "github_languages": {"Python": 21165}, "github_archived": false, "github_pushed_at": "2022-05-03T17:21:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generating-adversarial-examples-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1KUMxnrhnH", "year": 2026, "status": "rejected", "title": "A Separable Self-attention Inspired by the State Space Model for Computer Vision", "authors": ["Juntao Zhang", "Jun Zhou", "Kun Bian", "zhou you", "Jianning Liu"], "authorids": ["~Juntao_Zhang1", "~Jun_Zhou16", "~Kun_Bian1", "~zhou_you1", "~Jianning_Liu1"], "authors_source": "OpenReview API", "abstract": "Separable self-attention is an early attention mechanism with linear complexity. When parameters and FLOPs are comparable, lightweight networks built upon separable self-attention and its variants underperform the recent Vision Mamba (ViM). By analyzing the strengths and weaknesses of separable self-attention, we distill four design principles and, inspired by the State Space Model (SSM) serving as the core of ViM, propose a novel separable self-attention termed Vision Mamba Inspired Separable self-Attention (VMI-SA). Notably, VMI-SA does not incorporate any SSM blocks, and its attention computation process differs from all existing attention mechanisms to the best of our knowledge. We introduce proof-of-concept networks, VMINet and VMIFormer, enabling fair comparisons with ViMs through deliberate control of parameters, FLOPs, and encoder numbers. Compared to state-of-the-art Transformers, CNNs, and ViMs, VMINet and VMIFormer achieve competitive results in image classification and high-resolution dense prediction tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "dKY3anoCX2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6579/Reviewer_RwYd"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 5, "summary": "The paper introduces Vision Mamba Inspired Separable Self-Attention (VMI-SA), a new separable self-attention mechanism drawing design principles from State Space Models, particularly Mamba. The authors propose VMINet—a prototype vision backbone built purely from stacking VMI-SA blocks and downsampling layers. Through extensive experimentation across image classification, detection, and segmentation tasks, VMINet is shown to outperform state-space-based Vim models and be competitive with strong baselines in lightweight settings.", "review_text": "The paper introduces Vision Mamba Inspired Separable Self-Attention (VMI-SA), a new separable self-attention mechanism drawing design principles from State Space Models, particularly Mamba. The authors propose VMINet—a prototype vision backbone built purely from stacking VMI-SA blocks and downsampling layers. Through extensive experimentation across image classification, detection, and segmentation tasks, VMINet is shown to outperform state-space-based Vim models and be competitive with strong baselines in lightweight settings.", "strengths": "This paper analyzes different designs principles of self-attention, vision mamba, separable self-attention, and conclude the results into four rules to guide the design of the vision models.", "weaknesses": "1. The involvement of causal mask does not make sense for most of the vision tasks since there are no causal hypotheses in the spatial dimension of the images and videos. That is why Mamba models [1,2,3] in vision need to define one or several complicated scanning sequence to ensure the visual signals are correctly modeled. In VMI-SA, the authors use two set of learnable gating parameters  $\\alpha$s and  $\\beta$s to control the proportion between the causal contexts and the direct contexts. It is of vital importance to carefully analyze how and go for different inputs and in different layers, which can provide meaningful insights on how these two types of context affect the model on vision tasks. Another important work [4] points out that the causal modeling in vision mamba models could be regarded as a forced local modeling pattern, which is also helpful. However, these analyses are missing in current submission, which fade the technical depth of the paper.\n\n2. For the discussion part of Effectiveness of VMI-SA, the authors replace the VMI-SA block with an FC layer. However, this design choice does not resemble ConvNeXt block since the normalization layers are not the same. Current drop of the accuracy cannot support the assertion. The authors could adopt the MetaFormer [5] archictecture equipped with VMI-SA, Pooling, and Self-attention, respectively to verify the impact of the spatial modeling module.\n\n3. The experiment results do not report the model variants in larger sizes, e.g, GFLOPs for inputs on the ImageNet-1K datasets, and models with longer sequence inputs, e.g., input resolutions. It is hard to distinguish the proposed VMINet out of the baselines such as ViM [1].\n\n[1] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. \"Vision mamba: Efficient visual representation learning with bidirectional state space model.\", ICML 2024\n\n[2] Liu, Yue, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. \"Vmamba: Visual state space model.\", NeurIPS 2024\n\n[3] Yang, Chenhongyi, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. \"Plainmamba: Improving non-hierarchical mamba in visual recognition.\", BMVC 2024\n\n[4] Han, Dongchen, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. \"Demystify mamba in vision: A linear attention perspective.\", NeurIPS 2024\n\n[5] Yu, Weihao, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. \"Metaformer baselines for vision.\", IEEE TPAMI", "questions": "1. The format of the paper is a little messy with large blanks and unaligned equations.\n\n2. The \"Related Works\" section is missing, making it confusing to position this paper in some lines of research and show its unique advantages.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Vision Mamba Inspired Separable Self-Attention (VMI-SA), a new separable self-attention mechanism drawing design principles from State Space Models, particularly Mamba. The authors propose VMINet—a prototype vision backbone built purely from stacking VMI-SA blocks and downsampling layers. Through extensive experimentation across image classification, detection, and segmentation tasks, VMINet is shown to outperform state-space-based Vim models and be competitive with strong baselines in lightweight settings.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "This paper analyzes different designs principles of self-attention, vision mamba, separable self-attention, and conclude the results into four rules to guide the design of the vision models.", "weaknesses": "1. The involvement of causal mask does not make sense for most of the vision tasks since there are no causal hypotheses in the spatial dimension of the images and videos. That is why Mamba models [1,2,3] in vision need to define one or several complicated scanning sequence to ensure the visual signals are correctly modeled. In VMI-SA, the authors use two set of learnable gating parameters  $\\alpha$s and  $\\beta$s to control the proportion between the causal contexts and the direct contexts. It is of vital importance to carefully analyze how and go for different inputs and in different layers, which can provide meaningful insights on how these two types of context affect the model on vision tasks. Another important work [4] points out that the causal modeling in vision mamba models could be regarded as a forced local modeling pattern, which is also helpful. However, these analyses are missing in current submission, which fade the technical depth of the paper.\n\n2. For the discussion part of Effectiveness of VMI-SA, the authors replace the VMI-SA block with an FC layer. However, this design choice does not resemble ConvNeXt block since the normalization layers are not the same. Current drop of the accuracy cannot support the assertion. The authors could adopt the MetaFormer [5] archictecture equipped with VMI-SA, Pooling, and Self-attention, respectively to verify the impact of the spatial modeling module.\n\n3. The experiment results do not report the model variants in larger sizes, e.g, GFLOPs for inputs on the ImageNet-1K datasets, and models with longer sequence inputs, e.g., input resolutions. It is hard to distinguish the proposed VMINet out of the baselines such as ViM [1].\n\n[1] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. \"Vision mamba: Efficient visual representation learning with bidirectional state space model.\", ICML 2024\n\n[2] Liu, Yue, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. \"Vmamba: Visual state space model.\", NeurIPS 2024\n\n[3] Yang, Chenhongyi, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. \"Plainmamba: Improving non-hierarchical mamba in visual recognition.\", BMVC 2024\n\n[4] Han, Dongchen, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. \"Demystify mamba in vision: A linear attention perspective.\", NeurIPS 2024\n\n[5] Yu, Weihao, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. \"Metaformer baselines for vision.\", IEEE TPAMI", "questions": "1. The format of the paper is a little messy with large blanks and unaligned equations.\n\n2. The \"Related Works\" section is missing, making it confusing to position this paper in some lines of research and show its unique advantages.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762159868467}, {"id": "b7shFsvRST", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6579/Reviewer_t7qr"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This manuscript proposes a novel linear-complexity separable self-attention mechanism called Vision Mamba Inspired Separable self-Attention (VMI-SA), which draws inspiration from the SSM/Mamba while avoiding integrating any SSM blocks and featuring an attention computation process distinct from existing mechanisms. It first distill four design principles by analyzing the strengths and weaknesses of separable self-attention, then design a recurrent formulation of VMI-SA and a matrix formulation to enhance token dependency modeling and computational efficiency. Based on VMI-SA, they construct two proof-of-concept networks, VMINet and VMIFormer, and conduct fair comparisons with state-of-the-art Transformers, CNNs, and ViMs by controlling parameters, FLOPs, and encoder counts. Experimental results show that VMINet and VMIFormer achieve competitive performance in ImageNet-1K image classification, MSCOCO object detection, and ADE20K semantic segmentation, demonstrating VMI-SA’s effectiveness in balancing performance and efficiency.", "review_text": "This manuscript proposes a novel linear-complexity separable self-attention mechanism called Vision Mamba Inspired Separable self-Attention (VMI-SA), which draws inspiration from the SSM/Mamba while avoiding integrating any SSM blocks and featuring an attention computation process distinct from existing mechanisms. It first distill four design principles by analyzing the strengths and weaknesses of separable self-attention, then design a recurrent formulation of VMI-SA and a matrix formulation to enhance token dependency modeling and computational efficiency. Based on VMI-SA, they construct two proof-of-concept networks, VMINet and VMIFormer, and conduct fair comparisons with state-of-the-art Transformers, CNNs, and ViMs by controlling parameters, FLOPs, and encoder counts. Experimental results show that VMINet and VMIFormer achieve competitive performance in ImageNet-1K image classification, MSCOCO object detection, and ADE20K semantic segmentation, demonstrating VMI-SA’s effectiveness in balancing performance and efficiency.", "strengths": "1. This paper introduces an interesting model, which incorporates separate self-attention modules into the Mamba marco design.\n2. The experiment are conducted on competitive benchmarks, e.g., ImageNet, COCO and ADE20K.\n3. The final model have linear complexity, which is a very promising research topic to explore.", "weaknesses": "1. The novelty is limited. This paper incorporated the minor design in separable self-attention into the Mamba marco design, titled Mamba Inspired Separable self-Attention. It is very similar to MLLA (Mamba-Inspired Linear Attention)[1] , which incroporate the Mamba minor design into the vision transformer marco design.\n2. The paper also lacks the method comparsion and performance comparsion with MLLA[1].\n3. Although the authors claim this is a linear model, the performance when the token length varies is missing.\n4. There is nearly no ablation in the submission. Only ablation in mask type in Tab 5. However, how each minor design affects the final result is unclear.\n\n[1] Han et al, Demystify Mamba in Vision: A Linear Attention Perspective, in NeurIPS 2024.", "questions": "1. How does each detailed structure affect the final performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript proposes a novel linear-complexity separable self-attention mechanism called Vision Mamba Inspired Separable self-Attention (VMI-SA), which draws inspiration from the SSM/Mamba while avoiding integrating any SSM blocks and featuring an attention computation process distinct from existing mechanisms. It first distill four design principles by analyzing the strengths and weaknesses of separable self-attention, then design a recurrent formulation of VMI-SA and a matrix formulation to enhance token dependency modeling and computational efficiency. Based on VMI-SA, they construct two proof-of-concept networks, VMINet and VMIFormer, and conduct fair comparisons with state-of-the-art Transformers, CNNs, and ViMs by controlling parameters, FLOPs, and encoder counts. Experimental results show that VMINet and VMIFormer achieve competitive performance in ImageNet-1K image classification, MSCOCO object detection, and ADE20K semantic segmentation, demonstrating VMI-SA’s effectiveness in balancing performance and efficiency.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This paper introduces an interesting model, which incorporates separate self-attention modules into the Mamba marco design.\n2. The experiment are conducted on competitive benchmarks, e.g., ImageNet, COCO and ADE20K.\n3. The final model have linear complexity, which is a very promising research topic to explore.", "weaknesses": "1. The novelty is limited. This paper incorporated the minor design in separable self-attention into the Mamba marco design, titled Mamba Inspired Separable self-Attention. It is very similar to MLLA (Mamba-Inspired Linear Attention)[1] , which incroporate the Mamba minor design into the vision transformer marco design.\n2. The paper also lacks the method comparsion and performance comparsion with MLLA[1].\n3. Although the authors claim this is a linear model, the performance when the token length varies is missing.\n4. There is nearly no ablation in the submission. Only ablation in mask type in Tab 5. However, how each minor design affects the final result is unclear.\n\n[1] Han et al, Demystify Mamba in Vision: A Linear Attention Perspective, in NeurIPS 2024.", "questions": "1. How does each detailed structure affect the final performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762007926998}, {"id": "wIbUoc2Xp0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6579/Reviewer_uDKL"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes a variant of  separable self-attention method to incorporate correlation between tokens, which a basin SSA lacks. The authors incorporates three components: SSA, depthwise convolutio, and mask matrix to enhance the rank. This paper shows somewhat strong performance.", "review_text": "This paper proposes a variant of  separable self-attention method to incorporate correlation between tokens, which a basin SSA lacks. The authors incorporates three components: SSA, depthwise convolutio, and mask matrix to enhance the rank. This paper shows somewhat strong performance.", "strengths": "The evaluation is quite convicing. The comparison with ViM models shows that VMiNet and VMIFormer achieve superior performance over ViM variants.\n\nAlso, the ablation study of mask in Appendix demonstrates the importance of mask operation.", "weaknesses": "It is not clear what the authors really adopt from SSM to this proposed model. The explation between Eq. 9 and Eq. 10 in not clear.\nAlso, the efficiency analysis is too limited. Efficient VMamba shows the least FLOPS with longer latency and the explatnion is \"nsufficient GPU utilization in EfficientVMamba’s SSM module during shorter sequence processing.\" Does it mean the results would be different on longer sequences?\nAlso, the comparison does not include Flatten Transformer.", "questions": "1. Please clarify what exactly the inspiration from SSM is and the logic behind Eq. 9 and Eq. 10.\n2. Please add comparison with Flatten Transformer.\n3. Please include Top-5 accuracy.\n\n* please go over equations. For example, in Eq.5 == and != should be $-$ and $\\neq$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a variant of  separable self-attention method to incorporate correlation between tokens, which a basin SSA lacks. The authors incorporates three components: SSA, depthwise convolutio, and mask matrix to enhance the rank. This paper shows somewhat strong performance.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "The evaluation is quite convicing. The comparison with ViM models shows that VMiNet and VMIFormer achieve superior performance over ViM variants.\n\nAlso, the ablation study of mask in Appendix demonstrates the importance of mask operation.", "weaknesses": "It is not clear what the authors really adopt from SSM to this proposed model. The explation between Eq. 9 and Eq. 10 in not clear.\nAlso, the efficiency analysis is too limited. Efficient VMamba shows the least FLOPS with longer latency and the explatnion is \"nsufficient GPU utilization in EfficientVMamba’s SSM module during shorter sequence processing.\" Does it mean the results would be different on longer sequences?\nAlso, the comparison does not include Flatten Transformer.", "questions": "1. Please clarify what exactly the inspiration from SSM is and the logic behind Eq. 9 and Eq. 10.\n2. Please add comparison with Flatten Transformer.\n3. Please include Top-5 accuracy.\n\n* please go over equations. For example, in Eq.5 == and != should be $-$ and $\\neq$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761983132227}, {"id": "OCpU4OHzFc", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6579/Reviewer_21TC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper analyzes the strengths and weaknesses of separable self-attention, and based on which, proposes a new family of separable attention based model, namely VMI-SA. Comparing with previous methods, the main innovations of VMI-SA include:\n\n1. The attention blocks in VMI-SA apply element-wise multiplication to replace the traditional matrix multiplication\n\n2. Context vectors are introduced to replace attention matrices. \n\n3. Depth-wise conv is used to introduce local spatial correlations before the element-wise multiplication operation. \n\nBased on the ideas above, VMINet and VMIFormer are designed. Experimental results on image classification and object detection tasks show that the proposed method has comparable or better performance comparing with several recent proposed CNNs and Transformers.", "review_text": "This paper analyzes the strengths and weaknesses of separable self-attention, and based on which, proposes a new family of separable attention based model, namely VMI-SA. Comparing with previous methods, the main innovations of VMI-SA include:\n\n1. The attention blocks in VMI-SA apply element-wise multiplication to replace the traditional matrix multiplication\n\n2. Context vectors are introduced to replace attention matrices. \n\n3. Depth-wise conv is used to introduce local spatial correlations before the element-wise multiplication operation. \n\nBased on the ideas above, VMINet and VMIFormer are designed. Experimental results on image classification and object detection tasks show that the proposed method has comparable or better performance comparing with several recent proposed CNNs and Transformers.", "strengths": "1. The theoretical basis of the proposed method is carefully proposed. \n\n2. The process of the designing of VMI-SA is well-presented. Each main component, such as element-wise multiplication, context vector, and MAMBA inspired attention, are analyzed detailedly. \n\n3. The experiments cover many main-stream ViT and CNN models, thus prove the advantages of the method.", "weaknesses": "1. Some arguments need to be clarified. For instance, in 3.2, the authors firstly mentioned that \"The higher the\nrank of the attention matrix, the more attention information it contains, and the richer the feature diversity.\" Which implies that an attention matrix with higher rank may provide some benefits on feature extraction. After that, Eq.7 shows that the rank of context vector is less or equal with min{L, D}. Then the authors argued that the attention information in softmax(Q)⊙K is not only less abundant but also severely homogenized. Here, it seems like one benefit of context vector is to lower the rank of attention matrices. This is a conflict with the previous context. Moreover, in 3.3.2, the authors again mentioned that we need to enhance the rank of the attention matrix in the proposed method. \n\n2. Some tiny problems that may be improved. For instance, in Figure 2, it is better to mark some important features of the model, such as Q, K, and context vector.", "questions": "My questions are proposed in the part \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper analyzes the strengths and weaknesses of separable self-attention, and based on which, proposes a new family of separable attention based model, namely VMI-SA. Comparing with previous methods, the main innovations of VMI-SA include:\n\n1. The attention blocks in VMI-SA apply element-wise multiplication to replace the traditional matrix multiplication\n\n2. Context vectors are introduced to replace attention matrices. \n\n3. Depth-wise conv is used to introduce local spatial correlations before the element-wise multiplication operation. \n\nBased on the ideas above, VMINet and VMIFormer are designed. Experimental results on image classification and object detection tasks show that the proposed method has comparable or better performance comparing with several recent proposed CNNs and Transformers.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The theoretical basis of the proposed method is carefully proposed. \n\n2. The process of the designing of VMI-SA is well-presented. Each main component, such as element-wise multiplication, context vector, and MAMBA inspired attention, are analyzed detailedly. \n\n3. The experiments cover many main-stream ViT and CNN models, thus prove the advantages of the method.", "weaknesses": "1. Some arguments need to be clarified. For instance, in 3.2, the authors firstly mentioned that \"The higher the\nrank of the attention matrix, the more attention information it contains, and the richer the feature diversity.\" Which implies that an attention matrix with higher rank may provide some benefits on feature extraction. After that, Eq.7 shows that the rank of context vector is less or equal with min{L, D}. Then the authors argued that the attention information in softmax(Q)⊙K is not only less abundant but also severely homogenized. Here, it seems like one benefit of context vector is to lower the rank of attention matrices. This is a conflict with the previous context. Moreover, in 3.3.2, the authors again mentioned that we need to enhance the rank of the attention matrix in the proposed method. \n\n2. Some tiny problems that may be improved. For instance, in Figure 2, it is better to mark some important features of the model, such as Q, K, and context vector.", "questions": "My questions are proposed in the part \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761034949416}], "openreview_url": "https://openreview.net/forum?id=1KUMxnrhnH", "arxiv_id": "2501.02040", "paper_pdf": "papers/1KUMxnrhnH.pdf", "paper_pdf_sha256": "59e664034feb379394158560b5b8499353969dd6839b8700b1316067d387b9b2", "paper_pdf_bytes": 821704, "paper_pdf_source": "openreview", "code_url": "https://github.com/yws-wxs/VMINet", "code_repository": "yws-wxs/VMINet", "code_commit": "740b9beda36168983776c5bc0ef638539e077160", "code_archive": "repos/1KUMxnrhnH.zip", "code_archive_sha256": "44b31951bf4751c63594c4816702cd4cda2e8d66c48b2c9e6dec344810110525", "code_archive_bytes": 36014, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 85, "github_languages": {"Python": 75236}, "github_archived": false, "github_pushed_at": "2025-04-22T11:28:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-separable-self-attention-inspired-by-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7S1xDos9pH", "year": 2025, "status": "rejected", "title": "Generalized Gaussian Temporal Difference Error for Uncertainty-aware Reinforcement Learning", "authors": ["Seyeon Kim", "Joonhun Lee", "Namhoon Cho", "Sungjun Han", "Wooseop Hwang"], "authorids": ["~Seyeon_Kim5", "~Joonhun_Lee1", "~Namhoon_Cho2", "~Sungjun_Han2", "~Wooseop_Hwang1"], "authors_source": "OpenReview API", "abstract": "Conventional uncertainty-aware temporal difference (TD) learning methods often rely on simplistic assumptions, typically including a zero-mean Gaussian distribution for TD errors. Such oversimplification can lead to inaccurate error representations and compromised uncertainty estimation. In this paper, we introduce a novel framework for generalized Gaussian error modeling in deep reinforcement learning, applicable to both discrete and continuous control settings. Our framework enhances the flexibility of error distribution modeling by incorporating additional higher-order moment, particularly kurtosis, thereby improving the estimation and mitigation of data-dependent noise, i.e., aleatoric uncertainty. We examine the influence of the shape parameter of the generalized Gaussian distribution (GGD) on aleatoric uncertainty and provide a closed-form expression that demonstrates an inverse relationship between uncertainty and the shape parameter. Additionally, we propose a theoretically grounded weighting scheme to fully leverage the GGD. To address epistemic uncertainty, we enhance the batch inverse variance weighting by incorporating bias reduction and kurtosis considerations, resulting in improved robustness. Extensive experimental evaluations using policy gradient algorithms demonstrate the consistent efficacy of our method, showcasing significant performance improvements.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "mB4R5FgZmB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5406/Reviewer_UVhA"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors consider uncertainty estimation in TD learning and seek to generalize the conceptually simple and implicit Gaussian assumption behind the MSE loss. To do so, they consider  the generalized gaussian distribution, which has an additional shape parameter that can modulate heavy/light tailedness for the errors. The authors extend a previous approach to estimating uncertainty in TD learning to include an additional network to predict the kurtosis/shape parameter.\n\nEdit: updated my score after reviewing the author replies.", "review_text": "The authors consider uncertainty estimation in TD learning and seek to generalize the conceptually simple and implicit Gaussian assumption behind the MSE loss. To do so, they consider  the generalized gaussian distribution, which has an additional shape parameter that can modulate heavy/light tailedness for the errors. The authors extend a previous approach to estimating uncertainty in TD learning to include an additional network to predict the kurtosis/shape parameter.\n\nEdit: updated my score after reviewing the author replies.", "strengths": "- Effective uncertainty estimation is an important problem for RL and the authors study a valid relaxation of commonly held simplifying assumptions around Gaussianity.\n- The authors conduct an experimental study that shows plausible benefits of the proposed generalizations in simple environments", "weaknesses": "- There appear to be several ways to generalize the Normal distribution to have different tail behaviors, for example see q-Gaussian for yet another alternative. In the tradeoff between simplicity and more expressive modeling, it is unclear what is the right axis to explore. Even more generally, all of these are unimodal, symmetric (i.e. zero skew) distributions and arguably one could consider a full return distribution as well. In fact, prior work on distributional RL has explored this (https://arxiv.org/abs/1707.06887). \n\n- In terms of the learning algorithm, the paper modifies the loss function proposed in [Mai et al 2022] to the GGD case aided with an additional predictor for the kurtosis (the beta term). \n\n- The empirical results are mostly within error bars of the baselines. This seems especially so, when considering the marginal benefits in going from a basic uncertainty modeling with the gaussian assumption to the extra parameter estimation. Considering that this involves a whole extra network (not just an extra hyper-parameter) to predict the beta, this seems like a much more complex method for relatively little gain.\n\n- The tailedness of the distribution seems like a very sensitive parameter to estimate, and sensitive to simple reward transformations and/or clipping so I would be surprised if these observations are robust to such changes.", "questions": "- It seems like the distribution associated with environment transition stochasticity could easily lead to multi modal distributions of the cumulative return, which might be a much more interesting/important aspect than estimating the shape of a unimodal distribution better. Please address whether you observed any evidence of multimodality in your empirical results, and if so, how your method handles or could be extended to handle such cases.\n\n- Rewards are typically bounded, so I would expect any estimation of the tail behavior to be quite sensitive to various practical assumptions. Please describe any preprocessing steps applied to the rewards, and how the performance varies with different reward scales or bounds.\n\n- Given that your method requires predicting an extra head for the beta/kurtosis parameter, and training this (see Equation (4)) requires gradient descent through the gamma function of the output, how stable is the learning and/or optimization?  Please consider providing empirical evidence of the optimization stability, such as plots of the beta parameter estimates over time, or  discuss any specific techniques used to ensure stable training.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors consider uncertainty estimation in TD learning and seek to generalize the conceptually simple and implicit Gaussian assumption behind the MSE loss. To do so, they consider  the generalized gaussian distribution, which has an additional shape parameter that can modulate heavy/light tailedness for the errors. The authors extend a previous approach to estimating uncertainty in TD learning to include an additional network to predict the kurtosis/shape parameter.\n\nEdit: updated my score after reviewing the author replies.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- Effective uncertainty estimation is an important problem for RL and the authors study a valid relaxation of commonly held simplifying assumptions around Gaussianity.\n- The authors conduct an experimental study that shows plausible benefits of the proposed generalizations in simple environments", "weaknesses": "- There appear to be several ways to generalize the Normal distribution to have different tail behaviors, for example see q-Gaussian for yet another alternative. In the tradeoff between simplicity and more expressive modeling, it is unclear what is the right axis to explore. Even more generally, all of these are unimodal, symmetric (i.e. zero skew) distributions and arguably one could consider a full return distribution as well. In fact, prior work on distributional RL has explored this (https://arxiv.org/abs/1707.06887). \n\n- In terms of the learning algorithm, the paper modifies the loss function proposed in [Mai et al 2022] to the GGD case aided with an additional predictor for the kurtosis (the beta term). \n\n- The empirical results are mostly within error bars of the baselines. This seems especially so, when considering the marginal benefits in going from a basic uncertainty modeling with the gaussian assumption to the extra parameter estimation. Considering that this involves a whole extra network (not just an extra hyper-parameter) to predict the beta, this seems like a much more complex method for relatively little gain.\n\n- The tailedness of the distribution seems like a very sensitive parameter to estimate, and sensitive to simple reward transformations and/or clipping so I would be surprised if these observations are robust to such changes.", "questions": "- It seems like the distribution associated with environment transition stochasticity could easily lead to multi modal distributions of the cumulative return, which might be a much more interesting/important aspect than estimating the shape of a unimodal distribution better. Please address whether you observed any evidence of multimodality in your empirical results, and if so, how your method handles or could be extended to handle such cases.\n\n- Rewards are typically bounded, so I would expect any estimation of the tail behavior to be quite sensitive to various practical assumptions. Please describe any preprocessing steps applied to the rewards, and how the performance varies with different reward scales or bounds.\n\n- Given that your method requires predicting an extra head for the beta/kurtosis parameter, and training this (see Equation (4)) requires gradient descent through the gamma function of the output, how stable is the learning and/or optimization?  Please consider providing empirical evidence of the optimization stability, such as plots of the beta parameter estimates over time, or  discuss any specific techniques used to ensure stable training.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730938084799}, {"id": "v9yA39qxZB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5406/Reviewer_Nr5G"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper presents a novel framework for generalized Gaussian error modeling in uncertainty-aware temporal difference (TD) learning. It critiques conventional methods that assume a zero-mean Gaussian distribution for TD errors, leading to inaccurate uncertainty estimations. The proposed framework incorporates higher-order moments, specifically kurtosis, to enhance error modeling in reinforcement learning.", "review_text": "The paper presents a novel framework for generalized Gaussian error modeling in uncertainty-aware temporal difference (TD) learning. It critiques conventional methods that assume a zero-mean Gaussian distribution for TD errors, leading to inaccurate uncertainty estimations. The proposed framework incorporates higher-order moments, specifically kurtosis, to enhance error modeling in reinforcement learning.", "strengths": "1. The paper offers a closed-form expression demonstrating the relationship between uncertainty and the generalized Gaussian distribution's shape parameter, adding depth to the theoretical framework.\n2. The framework's applicability to both discrete and continuous control settings makes it relevant across various reinforcement learning contexts.\n3. The emphasis on data-dependent noise and aleatoric uncertainty is timely and important for improving the robustness of reinforcement learning algorithms.", "weaknesses": "1. The introduction of higher-order moments may complicate the implementation in practical scenarios, which could deter application by practitioners.\n2. In Figure 5, the return curves of Ant, HalfCheetah, and Humanoid are still increasing at the end of training step. The training steps can be increased to compare the final performances when all algorithms are converged.", "questions": "What specific tasks or environments in the real world do the authors envision as most beneficial for applying their method?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a novel framework for generalized Gaussian error modeling in uncertainty-aware temporal difference (TD) learning. It critiques conventional methods that assume a zero-mean Gaussian distribution for TD errors, leading to inaccurate uncertainty estimations. The proposed framework incorporates higher-order moments, specifically kurtosis, to enhance error modeling in reinforcement learning.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The paper offers a closed-form expression demonstrating the relationship between uncertainty and the generalized Gaussian distribution's shape parameter, adding depth to the theoretical framework.\n2. The framework's applicability to both discrete and continuous control settings makes it relevant across various reinforcement learning contexts.\n3. The emphasis on data-dependent noise and aleatoric uncertainty is timely and important for improving the robustness of reinforcement learning algorithms.", "weaknesses": "1. The introduction of higher-order moments may complicate the implementation in practical scenarios, which could deter application by practitioners.\n2. In Figure 5, the return curves of Ant, HalfCheetah, and Humanoid are still increasing at the end of training step. The training steps can be increased to compare the final performances when all algorithms are converged.", "questions": "What specific tasks or environments in the real world do the authors envision as most beneficial for applying their method?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730692534330}, {"id": "Wb7vlb6MJs", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5406/Reviewer_W9ns"], "rating": 8, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "In this work, the authors present a method to estimate uncertainty based on the Generalized Gaussian Distribution (GGD) to better characterize temporal difference error (TD-error) distributions. Unlike previous work, which assumes Gaussianity, the GGD captures additional parameters, specifically kurtosis, to account for higher-order moments, thus enhancing the description of the TD-error distribution. The authors then modify an existing algorithm to operate under GGD assumptions, demonstrating improved performance across various settings. They also provide theoretical guarantees for well-behaved probability density function optimization under GGD assumptions and offer insights into why their proposed method is effective.", "review_text": "In this work, the authors present a method to estimate uncertainty based on the Generalized Gaussian Distribution (GGD) to better characterize temporal difference error (TD-error) distributions. Unlike previous work, which assumes Gaussianity, the GGD captures additional parameters, specifically kurtosis, to account for higher-order moments, thus enhancing the description of the TD-error distribution. The authors then modify an existing algorithm to operate under GGD assumptions, demonstrating improved performance across various settings. They also provide theoretical guarantees for well-behaved probability density function optimization under GGD assumptions and offer insights into why their proposed method is effective.", "strengths": "- (S1): The authors provide strong motivation for improving uncertainty-aware reinforcement learning (RL), it is an active area with relevance in applications like risk-averse decision-making, managing the exploration-exploitation trade-off, and enhancing sample efficiency.\n\n- (S2): They conduct extensive experimentation, including ablation studies and parameter sweeps, to demonstrate that their proposed method outperforms the baseline (the original uncertainty-aware method).\n\n- (S3): The proposed method is supported by theoretical validations and guarantees. Additionally, the proofs and mathematical analyses are well-documented in the appendix, making them relatively straightforward to follow.", "weaknesses": "- (W1): The writing of the paper could be significantly improved. Many concepts and terms are used in the main text without proper definitions, making it difficult to follow.\n\n- (W2): Some important aspects of the proposed method are not well-developed and are instead left to references from the baseline method. While this approach is generally acceptable, in this case, understanding key elements like the inclusion of GGD and the consideration of uncertainty during training requires substantial familiarity with the referenced material. As a result, the text is not self-contained, and combined with the clarity issues, makes the main content challenging to understand.\n\n- (W3): Although extensive experimentation demonstrates the effectiveness of the proposal, some results are counterintuitive. There is also insufficient exposition of the proposed method’s inner workings and its differences from the baseline, for which performance plots alone are insufficient.", "questions": "- Q1: Related to W1 and W2, what is the intuition behind Equation 3? In the original paper (Mai et al., 2022), this is Equation (10), which is preceded by a detailed explanation of many complex terms in the equation. As it stands, I do not think the main text is self-contained, as it requires substantial prior knowledge of that specific reference.\n\n- Q2: How are ensembles used in this method? Ensembles are mentioned, but there is no explanation of how they are applied. Is this related to the variance in epistemic uncertainty estimation in the regularization term?\n\n- Q3: How does the regularization term capture epistemic uncertainty (as the uncertainty that can be reduced through learning)?\n\n- Q4: In Equation 4, how are risk-averse weights introduced? Was this derived from the inclusion of GGD in Mai et al. (2022), or was it manually designed to incorporate risk sensitivity?\n\n- Q5: Why is estimating only beta sufficient to account for uncertainty compared to variance? Since the GGD has three parameters, with the first two set to 0 and 1 respectively, how can variance and kurtosis be represented simultaneously by a single number, beta? If I understand correctly, variance and kurtosis can be derived from beta (and alpha, in the case of variance). How do both quantities vary for a fixed alpha and a changing beta? There may be a range of TD-error distributions that cannot be fully described by just beta, which could potentially decrease the agent's performance.\n\n- Q6: If Equation 4 holds, why does optimizing for both alpha and beta jointly decrease the agent’s performance, as shown in Figure 15?\n\n- Q7: If considering higher moments of the TD-error distribution improves performance, does this mean that greater improvements are seen in tasks where TD-error tails are larger or smaller? How does performance compare to the evolution of the moments in the observed TD-errors? Does this method show a greater improvement over the baseline when higher-order moments are more or less pronounced?\n\n- Q8: Typo in lines 198 and 987, and the font size of labels and ticks in figures could be increased significantly to facilitate reading.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors present a method to estimate uncertainty based on the Generalized Gaussian Distribution (GGD) to better characterize temporal difference error (TD-error) distributions. Unlike previous work, which assumes Gaussianity, the GGD captures additional parameters, specifically kurtosis, to account for higher-order moments, thus enhancing the description of the TD-error distribution. The authors then modify an existing algorithm to operate under GGD assumptions, demonstrating improved performance across various settings. They also provide theoretical guarantees for well-behaved probability density function optimization under GGD assumptions and offer insights into why their proposed method is effective.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "- (S1): The authors provide strong motivation for improving uncertainty-aware reinforcement learning (RL), it is an active area with relevance in applications like risk-averse decision-making, managing the exploration-exploitation trade-off, and enhancing sample efficiency.\n\n- (S2): They conduct extensive experimentation, including ablation studies and parameter sweeps, to demonstrate that their proposed method outperforms the baseline (the original uncertainty-aware method).\n\n- (S3): The proposed method is supported by theoretical validations and guarantees. Additionally, the proofs and mathematical analyses are well-documented in the appendix, making them relatively straightforward to follow.", "weaknesses": "- (W1): The writing of the paper could be significantly improved. Many concepts and terms are used in the main text without proper definitions, making it difficult to follow.\n\n- (W2): Some important aspects of the proposed method are not well-developed and are instead left to references from the baseline method. While this approach is generally acceptable, in this case, understanding key elements like the inclusion of GGD and the consideration of uncertainty during training requires substantial familiarity with the referenced material. As a result, the text is not self-contained, and combined with the clarity issues, makes the main content challenging to understand.\n\n- (W3): Although extensive experimentation demonstrates the effectiveness of the proposal, some results are counterintuitive. There is also insufficient exposition of the proposed method’s inner workings and its differences from the baseline, for which performance plots alone are insufficient.", "questions": "- Q1: Related to W1 and W2, what is the intuition behind Equation 3? In the original paper (Mai et al., 2022), this is Equation (10), which is preceded by a detailed explanation of many complex terms in the equation. As it stands, I do not think the main text is self-contained, as it requires substantial prior knowledge of that specific reference.\n\n- Q2: How are ensembles used in this method? Ensembles are mentioned, but there is no explanation of how they are applied. Is this related to the variance in epistemic uncertainty estimation in the regularization term?\n\n- Q3: How does the regularization term capture epistemic uncertainty (as the uncertainty that can be reduced through learning)?\n\n- Q4: In Equation 4, how are risk-averse weights introduced? Was this derived from the inclusion of GGD in Mai et al. (2022), or was it manually designed to incorporate risk sensitivity?\n\n- Q5: Why is estimating only beta sufficient to account for uncertainty compared to variance? Since the GGD has three parameters, with the first two set to 0 and 1 respectively, how can variance and kurtosis be represented simultaneously by a single number, beta? If I understand correctly, variance and kurtosis can be derived from beta (and alpha, in the case of variance). How do both quantities vary for a fixed alpha and a changing beta? There may be a range of TD-error distributions that cannot be fully described by just beta, which could potentially decrease the agent's performance.\n\n- Q6: If Equation 4 holds, why does optimizing for both alpha and beta jointly decrease the agent’s performance, as shown in Figure 15?\n\n- Q7: If considering higher moments of the TD-error distribution improves performance, does this mean that greater improvements are seen in tasks where TD-error tails are larger or smaller? How does performance compare to the evolution of the moments in the observed TD-errors? Does this method show a greater improvement over the baseline when higher-order moments are more or less pronounced?\n\n- Q8: Typo in lines 198 and 987, and the font size of labels and ticks in figures could be increased significantly to facilitate reading.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730586625415}, {"id": "0cMVDAJtlS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5406/Reviewer_BUWP"], "rating": 3, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "Instead of Gaussian distribution, this paper applies generalized Gaussian distribution (GGD) to model temporal difference (TD) error, and proposes an uncertainty-aware objective function to minimize the TD error. It discussed several properties of the GGD distributions and empirically the proposed objective can work better than some baselines on common benchmark RL problems.", "review_text": "Instead of Gaussian distribution, this paper applies generalized Gaussian distribution (GGD) to model temporal difference (TD) error, and proposes an uncertainty-aware objective function to minimize the TD error. It discussed several properties of the GGD distributions and empirically the proposed objective can work better than some baselines on common benchmark RL problems.", "strengths": "- Discussion on some advantageous properties of GGD modelling\n- Empirical success with the proposed objective", "weaknesses": "The main problem with the current paper is its presentation, especially the lack of motivation for the objective function.\n\n1. Fig.2 requires further clarification. It is surprising to see that, for example, the Gaussian fitted PDF is so different from the empirical histogram for Ant-v4 (see also the MountainCar-v0 in Fig.7(d) for a more obvious mismatch). For Hopper-v4, the standard deviations of the two Gaussians are very close to each other despite the empirical histograms having very different shapes. More importantly, this figure is about the distribution of ALL TD errors across state-action pairs, while the NLL objective is about the underlying distribution of ONE single state-action pair. For this figure to make sense, one needs to assume that all TD errors come from the same underlying distribution, which doesn’t seem to be the case as Eq.(4) uses different betas for different step t. \n\n2. L273 offers a hypothesis, which is unverified. How do we define “unexpected states and rewards” and could you provide empirical support for this? Similarly, the explanation of interplay is unsatisfactory for the paragraph in L278. No evidence to support claims like “As training advances, aleatoric errors tend to become more influential while epistemic uncertainty diminishes, potentially resulting in non-normally distributed errors with heavier tails.”\n\n3. It remains unclear how Thm.2 can lead to the specific weighting in Eq.(4). A detailed derivation would be helpful. Similarly, it is unclear why Eq.(8) can address the bias issue mentioned in L332. The current paper has too many hand-wavy arguments without sufficient rigorous discussions.\n\n4. Finally, it seems that all theoretical properties in the current paper are stemmed from prior work, as shown by the reference in every theorem/proposition. If this is indeed the case, then what would be the main theoretical contribution of the current paper, other than an application of an existing tool (GGD) to TD learning?", "questions": "1. Please explain further why the Gaussian fits in Fig.2 are very off, and why looking at the TD errors of all examples can justify the underlying GGD assumption as discussed above.\n\n2. How can Thm.2 lead to the specific weighting in (4)?\n\n3. What is the main theoretical contribution of this work?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Instead of Gaussian distribution, this paper applies generalized Gaussian distribution (GGD) to model temporal difference (TD) error, and proposes an uncertainty-aware objective function to minimize the TD error. It discussed several properties of the GGD distributions and empirically the proposed objective can work better than some baselines on common benchmark RL problems.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "- Discussion on some advantageous properties of GGD modelling\n- Empirical success with the proposed objective", "weaknesses": "The main problem with the current paper is its presentation, especially the lack of motivation for the objective function.\n\n1. Fig.2 requires further clarification. It is surprising to see that, for example, the Gaussian fitted PDF is so different from the empirical histogram for Ant-v4 (see also the MountainCar-v0 in Fig.7(d) for a more obvious mismatch). For Hopper-v4, the standard deviations of the two Gaussians are very close to each other despite the empirical histograms having very different shapes. More importantly, this figure is about the distribution of ALL TD errors across state-action pairs, while the NLL objective is about the underlying distribution of ONE single state-action pair. For this figure to make sense, one needs to assume that all TD errors come from the same underlying distribution, which doesn’t seem to be the case as Eq.(4) uses different betas for different step t. \n\n2. L273 offers a hypothesis, which is unverified. How do we define “unexpected states and rewards” and could you provide empirical support for this? Similarly, the explanation of interplay is unsatisfactory for the paragraph in L278. No evidence to support claims like “As training advances, aleatoric errors tend to become more influential while epistemic uncertainty diminishes, potentially resulting in non-normally distributed errors with heavier tails.”\n\n3. It remains unclear how Thm.2 can lead to the specific weighting in Eq.(4). A detailed derivation would be helpful. Similarly, it is unclear why Eq.(8) can address the bias issue mentioned in L332. The current paper has too many hand-wavy arguments without sufficient rigorous discussions.\n\n4. Finally, it seems that all theoretical properties in the current paper are stemmed from prior work, as shown by the reference in every theorem/proposition. If this is indeed the case, then what would be the main theoretical contribution of the current paper, other than an application of an existing tool (GGD) to TD learning?", "questions": "1. Please explain further why the Gaussian fits in Fig.2 are very off, and why looking at the TD errors of all examples can justify the underlying GGD assumption as discussed above.\n\n2. How can Thm.2 lead to the specific weighting in (4)?\n\n3. What is the main theoretical contribution of this work?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730565628689}], "openreview_url": "https://openreview.net/forum?id=7S1xDos9pH", "arxiv_id": "2408.02295", "paper_pdf": "papers/7S1xDos9pH.pdf", "paper_pdf_sha256": "386b26c58b03fe5decfb4d3fd288d447649bb8057edbcbee76b2b98776c791a2", "paper_pdf_bytes": 16992959, "paper_pdf_source": "openreview", "code_url": "https://github.com/ait-lab/ggtde", "code_repository": "ait-lab/ggtde", "code_commit": "c2881d8d669469b2490e47b679c088b67f52bb6f", "code_archive": "repos/7S1xDos9pH.zip", "code_archive_sha256": "bbb7a95076fe2ae99002f239eafb97aca84ccb7cdea9e8b211ca8572102bc8e7", "code_archive_bytes": 33588, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 27, "github_languages": {"Python": 88286}, "github_archived": false, "github_pushed_at": "2026-07-23T08:29:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalized-gaussian-temporal-difference"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9FXGX00iMF", "year": 2024, "status": "rejected", "title": "BWS: Best Window Selection Based on Sample Scores for Data Pruning across Broad Ranges", "authors": ["Hoyong Choi", "Nohyun Ki", "Hye Won Chung"], "authorids": ["~Hoyong_Choi1", "~Nohyun_Ki1", "~Hye_Won_Chung2"], "authors_source": "OpenReview API", "abstract": "Data subset selection aims to find a smaller yet informative subset of a large dataset that can approximate the full-dataset training, addressing challenges associated with training neural networks on large-scale datasets. However, existing methods tend to specialize in either high or low selection ratio regimes, lacking a universal approach that consistently achieves competitive performance across a broad range of selection ratios. We introduce a universal and efficient data subset selection method, Best Window Selection (BWS), by proposing a method to choose the best window subset from samples ordered based on their difficulty scores. This approach offers flexibility by allowing the choice of window intervals that span from easy to difficult samples. Furthermore, we provide an efficient mechanism for selecting the best window subset by evaluating its quality using kernel ridge regression. Our experimental results demonstrate the superior performance of BWS compared to other baselines across a broad range of selection ratios over datasets, including CIFAR-10/100 and ImageNet, and the scenarios involving training from random initialization or fine-tuning of pre-trained models.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "ueP8lEOQaX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3340/Reviewer_j18Y"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper presents an approach, known as Best Window Selection (BWS), designed to tackle the challenges associated with data subset selection in machine learning. BWS allows for the adaptable selection of subsets based on sample difficulty scores and consistently delivers competitive performance over a broad range of selection ratios, spanning from 1% to 90%. It excels in comparison to existing score-based and optimization-based methods when applied to datasets like CIFAR-10/100 and ImageNet.", "review_text": "The paper presents an approach, known as Best Window Selection (BWS), designed to tackle the challenges associated with data subset selection in machine learning. BWS allows for the adaptable selection of subsets based on sample difficulty scores and consistently delivers competitive performance over a broad range of selection ratios, spanning from 1% to 90%. It excels in comparison to existing score-based and optimization-based methods when applied to datasets like CIFAR-10/100 and ImageNet.", "strengths": "1) The problem studied is meaningful and significant: finding a versatile data selection approach capable of sustaining competitive performance across a diverse range of selection ratios.\n2) Experiments show that the proposed BWS consistently outperforms other baselines, including both score-based and optimization-based approaches.\n3) The authors provide code, which enhances the reproducibility.", "weaknesses": "1) The notion of a \"window\" refers to a fixed-length interval within a sorted dataset. The \"Best Window Selection (BWS)\" algorithm operates under the assumption that the most optimal subset should be contiguous regarding the level of difficulty. However, the paper lacks an in-depth analysis of this particular aspect.\n\n2) It would be intriguing to explore the broader scenario where a \"window\" comprises several smaller intervals and varying starting points.\n\n3) Figure 3's readability could be enhanced by employing more distinguishable colors and markers for clarity.", "questions": "Kindly refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents an approach, known as Best Window Selection (BWS), designed to tackle the challenges associated with data subset selection in machine learning. BWS allows for the adaptable selection of subsets based on sample difficulty scores and consistently delivers competitive performance over a broad range of selection ratios, spanning from 1% to 90%. It excels in comparison to existing score-based and optimization-based methods when applied to datasets like CIFAR-10/100 and ImageNet.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1) The problem studied is meaningful and significant: finding a versatile data selection approach capable of sustaining competitive performance across a diverse range of selection ratios.\n2) Experiments show that the proposed BWS consistently outperforms other baselines, including both score-based and optimization-based approaches.\n3) The authors provide code, which enhances the reproducibility.", "weaknesses": "1) The notion of a \"window\" refers to a fixed-length interval within a sorted dataset. The \"Best Window Selection (BWS)\" algorithm operates under the assumption that the most optimal subset should be contiguous regarding the level of difficulty. However, the paper lacks an in-depth analysis of this particular aspect.\n\n2) It would be intriguing to explore the broader scenario where a \"window\" comprises several smaller intervals and varying starting points.\n\n3) Figure 3's readability could be enhanced by employing more distinguishable colors and markers for clarity.", "questions": "Kindly refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698796522117}, {"id": "cmn3EbSH5z", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3340/Reviewer_BL3J"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a novel and universal coreset selection method called \"Best Window Selection (BWS)\" to strike a balance between sample diversity and model performance across a broad range of selection ratios. BWS first sorts all training examples w.r.t. the difficulty score and then prunes a specific number of the most difficult examples and easiest examples. By comparing BWS with other SOTA baselines, the evaluation results show that BWS outperforms other coreset selection methods.", "review_text": "The paper proposes a novel and universal coreset selection method called \"Best Window Selection (BWS)\" to strike a balance between sample diversity and model performance across a broad range of selection ratios. BWS first sorts all training examples w.r.t. the difficulty score and then prunes a specific number of the most difficult examples and easiest examples. By comparing BWS with other SOTA baselines, the evaluation results show that BWS outperforms other coreset selection methods.", "strengths": "1. This paper proposes a novel coreset selection method, BWS. Compared to previous work, BWS selects the best window more efficiently with kernel ridge regression, which is faster than training a model from scratch.\n\n2. The evaluation results show that BWS achieves better or comparable results to other SOTA methods.\n\n3. The overall writing is good and easy to follow.", "weaknesses": "1. Using kernel ridge regression to decide the best window is not quite intuitive. What is the motivation to use kernel ridge regression rather than training a small network to decide the best window?\n\n2. The baseline evaluation results are inconsistent with data reported in the baseline method. For example, moderate are reported to have better performance than random on CIFAR10. CCS seems to have a better performance at 10% subset ratio than the numbers reported in the paper. It may be good to explain why the difference exists.", "questions": "I don’t fully understand why the performance of $w_s$ can represent the performance of models trained on the same subset. Could the authors further explain the connection between kernel regression and deep learning model training? What I currently feel is that it is more like an empirical transferability stuff studied in [1]: it is possible to use a small model to select coresets that transfer well to larger models.\n\n[1] Coleman, C., et al. \"Selection via Proxy: Efficient Data Selection for Deep Learning.\" International Conference on Learning Representations (ICLR). 2020.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel and universal coreset selection method called \"Best Window Selection (BWS)\" to strike a balance between sample diversity and model performance across a broad range of selection ratios. BWS first sorts all training examples w.r.t. the difficulty score and then prunes a specific number of the most difficult examples and easiest examples. By comparing BWS with other SOTA baselines, the evaluation results show that BWS outperforms other coreset selection methods.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper proposes a novel coreset selection method, BWS. Compared to previous work, BWS selects the best window more efficiently with kernel ridge regression, which is faster than training a model from scratch.\n\n2. The evaluation results show that BWS achieves better or comparable results to other SOTA methods.\n\n3. The overall writing is good and easy to follow.", "weaknesses": "1. Using kernel ridge regression to decide the best window is not quite intuitive. What is the motivation to use kernel ridge regression rather than training a small network to decide the best window?\n\n2. The baseline evaluation results are inconsistent with data reported in the baseline method. For example, moderate are reported to have better performance than random on CIFAR10. CCS seems to have a better performance at 10% subset ratio than the numbers reported in the paper. It may be good to explain why the difference exists.", "questions": "I don’t fully understand why the performance of $w_s$ can represent the performance of models trained on the same subset. Could the authors further explain the connection between kernel regression and deep learning model training? What I currently feel is that it is more like an empirical transferability stuff studied in [1]: it is possible to use a small model to select coresets that transfer well to larger models.\n\n[1] Coleman, C., et al. \"Selection via Proxy: Efficient Data Selection for Deep Learning.\" International Conference on Learning Representations (ICLR). 2020.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698778986512}, {"id": "aS0HJxVH7M", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3340/Reviewer_LWro"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a method that aims to find informative subset of the original datasets which can be used to train the neural networks with small performance drop compared with model trained with whole dataset. The point of this paper is to propose a method that can do both universal and efficient selection of subset based on the difficulty score. To adaptively select the best subset, the authors propose a method based on kernel ridge regression. The proposed method can be used to select subset for both training from scratch and fine-tuning. Extensive experiments are conducted to verify the efficacy of the proposed method.", "review_text": "This paper proposes a method that aims to find informative subset of the original datasets which can be used to train the neural networks with small performance drop compared with model trained with whole dataset. The point of this paper is to propose a method that can do both universal and efficient selection of subset based on the difficulty score. To adaptively select the best subset, the authors propose a method based on kernel ridge regression. The proposed method can be used to select subset for both training from scratch and fine-tuning. Extensive experiments are conducted to verify the efficacy of the proposed method.", "strengths": "This paper gives a deep understanding of which kind of data can be useful for different size of subset and use kernel regression to analyze this problem theoretically. \nThe situations  for hard sample and easy sample  to have benign effect is reasonable. \nThe usage of kernel ridge regression for subset selection is interesting.  The details of each parts of the proposed method are illustrated clearly.\nExtensive experiments validate the efficacy of the proposed method for training from scratch. \nThe method can also be effective when used to select subset for fine-tuning.\nAblation studies also validate the robustness of the proposed method.", "weaknesses": "For the experiments on CIFAR-10 with noise, the proposed method is outperformed by Moderate DS for 3 ratios. Could the authors illustrate the noisy rate of the selected subset to check whether the proposed method is prone to choose noisy data under this setting?\n\nThe experiments on CIFAR-10 fine-tuning on VIT shows that CCS is consistently better than the proposed method, could the author give concrete analysis of this phenomenon?", "questions": "Please refer to weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method that aims to find informative subset of the original datasets which can be used to train the neural networks with small performance drop compared with model trained with whole dataset. The point of this paper is to propose a method that can do both universal and efficient selection of subset based on the difficulty score. To adaptively select the best subset, the authors propose a method based on kernel ridge regression. The proposed method can be used to select subset for both training from scratch and fine-tuning. Extensive experiments are conducted to verify the efficacy of the proposed method.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "This paper gives a deep understanding of which kind of data can be useful for different size of subset and use kernel regression to analyze this problem theoretically. \nThe situations  for hard sample and easy sample  to have benign effect is reasonable. \nThe usage of kernel ridge regression for subset selection is interesting.  The details of each parts of the proposed method are illustrated clearly.\nExtensive experiments validate the efficacy of the proposed method for training from scratch. \nThe method can also be effective when used to select subset for fine-tuning.\nAblation studies also validate the robustness of the proposed method.", "weaknesses": "For the experiments on CIFAR-10 with noise, the proposed method is outperformed by Moderate DS for 3 ratios. Could the authors illustrate the noisy rate of the selected subset to check whether the proposed method is prone to choose noisy data under this setting?\n\nThe experiments on CIFAR-10 fine-tuning on VIT shows that CCS is consistently better than the proposed method, could the author give concrete analysis of this phenomenon?", "questions": "Please refer to weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698671126556}], "openreview_url": "https://openreview.net/forum?id=9FXGX00iMF", "arxiv_id": "2406.03057", "paper_pdf": "papers/9FXGX00iMF.pdf", "paper_pdf_sha256": "0da5ce530cf3a945c7fa30a003acffe8ad9b37bbe424bd786d4af5af01e55ab7", "paper_pdf_bytes": 1927725, "paper_pdf_source": "openreview", "code_url": "https://github.com/NohyunKi/BWS", "code_repository": "NohyunKi/BWS", "code_commit": "be73a6f64c57ebc12fb0ff57320cde4a82d6c47f", "code_archive": "repos/9FXGX00iMF.zip", "code_archive_sha256": "5b11e6606f2a3a33a8cb988025519dfbbfa74f7229caaf50bea8a1a31ffd2af0", "code_archive_bytes": 31806, "code_file_count": 20, "code_extensions": {".py": 16, ".sh": 4}, "github_disk_usage_kb": 23, "github_languages": {"Python": 102610, "Shell": 6051}, "github_archived": false, "github_pushed_at": "2024-05-23T15:48:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bws-best-window-selection-based-on-sample"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B7HJ9KLFV9U", "year": 2023, "status": "rejected", "title": "Thinking Two Moves Ahead: Anticipating Other Users Improves Backdoor Attacks in Federated Learning", "authors": ["Yuxin Wen", "Jonas Geiping", "Liam H Fowl", "Hossein Souri", "Rama Chellappa", "Micah Goldblum", "Tom Goldstein"], "authorids": ["~Yuxin_Wen2", "~Jonas_Geiping1", "~Liam_H_Fowl1", "~Hossein_Souri1", "~Rama_Chellappa1", "~Micah_Goldblum1", "~Tom_Goldstein1"], "authors_source": "OpenReview API", "abstract": "Federated learning is particularly susceptible to model poisoning and backdoor attacks because individual users have direct control over the training data and model updates.  At the same time, the attack power of an individual user is limited because their updates are quickly drowned out by those of many other users. Existing attacks do not account for future behaviors of other users, and thus require many sequential updates and their effects are quickly erased. We propose an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients, and ensures that backdoors are effective quickly and persist even after multiple rounds of community updates. We show that this new attack is effective in realistic scenarios where the attacker only contributes to a small fraction of randomly sampled rounds and demonstrate this attack on image classification, next-word prediction, and sentiment analysis.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "3gaV4xr6us", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2056/Reviewer_Fbk1"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work focuses on the model poisoning and backdoor attacks in federated learning. It proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients.\n\nThe main contributions can be summarized as:\n\n(1) Proposing a backdoor attack in federated learning (anticipate).\n\n(2) Providing extensive experiments.", "review_text": "I recommend marginally below the acceptance threshold. I will give my final score based on the response and other reviewer's comments.", "strengths": "Strengths:\n\n(a) This paper is well-written and easy to read.\n\n(b) I appreciate that extensive experiments are provided in this work.\n\nWeaknesses:\n\n(a) Should m steps be represented in Eq.(1)? If yes, please revise Eq.(1) and provide some references. If no, please provide a reasonable explanation.\n\n(b) In Section 3, it is mentioned that ``We believe this threat model with random attack opportunities is a natural step towards the evaluation of risks caused by backdoor attacks in more realistic systems''. Please provide a proper example to demonstrate the application in reality.\n\n(c) In Section 4, some formulas are not numbered.\n\n(d) It is hard to imagine how to implement the method proposed in this work. In the lines 9 to 12 of Algorithm 1, the proposed method needs to calculate ($\\theta_1, \\theta_2, ..., \\theta_k$). It is ok to calculate these parameters, but it is hard to imagine how to calculate line 14. In line 14, it requires to differentiate k-th step w.r.t $\\theta_{mal}$. Recall the calculation of $\\theta_{u,j}$ (such as Eq.(1)), we can find that it needs to calculate the derivatives. These mean that there are k-th partial derivatives in the calculation of line 14, which are very difficult to implement. This point is my main concern. Please provide a very detailed explanation and describe how to calculate it.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work focuses on the model poisoning and backdoor attacks in federated learning. It proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients.\n\nThe main contributions can be summarized as:\n\n(1) Proposing a backdoor attack in federated learning (anticipate).\n\n(2) Providing extensive experiments.", "strength_and_weaknesses": "Strengths:\n\n(a) This paper is well-written and easy to read.\n\n(b) I appreciate that extensive experiments are provided in this work.\n\nWeaknesses:\n\n(a) Should m steps be represented in Eq.(1)? If yes, please revise Eq.(1) and provide some references. If no, please provide a reasonable explanation.\n\n(b) In Section 3, it is mentioned that ``We believe this threat model with random attack opportunities is a natural step towards the evaluation of risks caused by backdoor attacks in more realistic systems''. Please provide a proper example to demonstrate the application in reality.\n\n(c) In Section 4, some formulas are not numbered.\n\n(d) It is hard to imagine how to implement the method proposed in this work. In the lines 9 to 12 of Algorithm 1, the proposed method needs to calculate ($\\theta_1, \\theta_2, ..., \\theta_k$). It is ok to calculate these parameters, but it is hard to imagine how to calculate line 14. In line 14, it requires to differentiate k-th step w.r.t $\\theta_{mal}$. Recall the calculation of $\\theta_{u,j}$ (such as Eq.(1)), we can find that it needs to calculate the derivatives. These mean that there are k-th partial derivatives in the calculation of line 14, which are very difficult to implement. This point is my main concern. Please provide a very detailed explanation and describe how to calculate it.", "clarity,_quality,_novelty_and_reproducibility": "Poor quality.\nNice clarity.\nPoor originality.", "summary_of_the_review": "I recommend marginally below the acceptance threshold. I will give my final score based on the response and other reviewer's comments.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667479403677}, {"id": "uHZCVRKl0Xv", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2056/Reviewer_VoJT"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients, and ensures that backdoors are effective quickly and persist even after multiple rounds of community updates. Experiments on datasets  demonstrate the performance of this attack on image classification, next-word prediction, and sentiment analysis.\n\n", "review_text": "Please see the above. ", "strengths": "1. The paper claims two moves ahead would be better than other attack methods such as norm-bounding. I guess that Eq. (4) with two loss functions would be the two moves, but the parameter between the moves is not given. Generally, we need to add a parameter to compromise the two loss functions in eq.(4), and add more experiments to show the results with different \"move weights\". However, it seems the moves are averaged and the paper does not explain why the average move is effective and why the compromise weight is not required as expected. \n\n2. The motivation and experiments are with strong control. It would be persuasive to add real-world motivation and experimental data.  Plus, I found that a demo version of this paper has been published in ICML. Because a previous version of this paper has already published, it would be better to add discussions about the difference between the two papers. \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes an attack that anticipates and accounts for the entire federated learning pipeline, including behaviors of other clients, and ensures that backdoors are effective quickly and persist even after multiple rounds of community updates. Experiments on datasets  demonstrate the performance of this attack on image classification, next-word prediction, and sentiment analysis.\n\n", "strength_and_weaknesses": "1. The paper claims two moves ahead would be better than other attack methods such as norm-bounding. I guess that Eq. (4) with two loss functions would be the two moves, but the parameter between the moves is not given. Generally, we need to add a parameter to compromise the two loss functions in eq.(4), and add more experiments to show the results with different \"move weights\". However, it seems the moves are averaged and the paper does not explain why the average move is effective and why the compromise weight is not required as expected. \n\n2. The motivation and experiments are with strong control. It would be persuasive to add real-world motivation and experimental data.  Plus, I found that a demo version of this paper has been published in ICML. Because a previous version of this paper has already published, it would be better to add discussions about the difference between the two papers. \n\n", "clarity,_quality,_novelty_and_reproducibility": "Difficult to read, and hard to reproduce. ", "summary_of_the_review": "Please see the above. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666668564920}, {"id": "QzvlRAyLhyO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2056/Reviewer_xjs7"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "A new backdoor attack against federated learning is proposed. The paper considers a single malicious device that can share malicious updates with the server to inject a backdoor into the centralized model, and the server adopts FedAvg with norm clipping as the aggregation rule. The main idea is to account for benign users' behavior in future rounds to make the malicious updates more persistent. Simulation results demonstrate the effectiveness of the proposed attack against the norm-bounding defense. ", "review_text": "The idea of simulating benign users' behavior to develop more advanced attacks against federated learning is interesting. However, the paper considers an over-simplified setting, and the proposed backdoor attack is only effective against the norm-bounding defense and completely fails against other common defenses. The proposed solution is far from being sophisticated. ", "strengths": "Strengths\n\n1. The idea of simulating benign users' behavior to develop more advanced attacks against federated learning is interesting.\n2. The proposed backdoor attack is effective against the norm-bounding defense. \n\nWeaknesses\n\n1. The paper considers an oversimplified setting with a single malicious device, and the simulation algorithm only considers a single attack step (the current step) and ignores the possibility that the device may be sampled again soon. Further, the algorithm simulates a fixed number of steps before the next attack happens, which is inaccurate given the randomness of the sampling algorithm used by the server. As the attacker has no access to other users' local data, the paper simply simulates their behavior from the attacker's local data, which can also be inaccurate for non-iid local data distributions. \n2. Because of the simplified design, the proposed attack can only break the norm-bounding defense and completely fails when the server adopts Krum, Multi-Krum, or Median, as shown in the appendix. Note that these are not really \"advanced\" defenses as they are not designed for backdoor attacks. Recent defenses, especially detection-based and post-training-based defenses, obtain more promising results against backdoors, which are completely ignored in the paper. \n3. For the norm-bounding defense considered in the paper, the norm threshold C is crucial for achieving a good tradeoff between the main task accuracy and robustness against attacks. But I cannot find the value of C used in the experiments, and there is no ablation study that investigates the impact of C on main task accuracy and backdoor accuracy. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "A new backdoor attack against federated learning is proposed. The paper considers a single malicious device that can share malicious updates with the server to inject a backdoor into the centralized model, and the server adopts FedAvg with norm clipping as the aggregation rule. The main idea is to account for benign users' behavior in future rounds to make the malicious updates more persistent. Simulation results demonstrate the effectiveness of the proposed attack against the norm-bounding defense. ", "strength_and_weaknesses": "Strengths\n\n1. The idea of simulating benign users' behavior to develop more advanced attacks against federated learning is interesting.\n2. The proposed backdoor attack is effective against the norm-bounding defense. \n\nWeaknesses\n\n1. The paper considers an oversimplified setting with a single malicious device, and the simulation algorithm only considers a single attack step (the current step) and ignores the possibility that the device may be sampled again soon. Further, the algorithm simulates a fixed number of steps before the next attack happens, which is inaccurate given the randomness of the sampling algorithm used by the server. As the attacker has no access to other users' local data, the paper simply simulates their behavior from the attacker's local data, which can also be inaccurate for non-iid local data distributions. \n2. Because of the simplified design, the proposed attack can only break the norm-bounding defense and completely fails when the server adopts Krum, Multi-Krum, or Median, as shown in the appendix. Note that these are not really \"advanced\" defenses as they are not designed for backdoor attacks. Recent defenses, especially detection-based and post-training-based defenses, obtain more promising results against backdoors, which are completely ignored in the paper. \n3. For the norm-bounding defense considered in the paper, the norm threshold C is crucial for achieving a good tradeoff between the main task accuracy and robustness against attacks. But I cannot find the value of C used in the experiments, and there is no ablation study that investigates the impact of C on main task accuracy and backdoor accuracy. ", "clarity,_quality,_novelty_and_reproducibility": "Algorithm 1 needs to be clarified in a couple of places. Does the group of users U_i change over time? How is the number of simulated benign users n' determined? Also, the algorithm uses FedAvg to simulate the server's aggregation rule. Does it include the norm clipping step? \n\nIn terms of novelty, the idea of simulating the behavior of benign agents in federated learning has been considered in recent work, e.g., [1]. Although the focus in [1] was on model poisoning, the idea of building a model of the FL system and benign agents using common knowledge and the attacker's local data can be readily applied to backdoor attacks. Further, [1] used reinforcement learning to obtain an attack policy that optimizes a long-term objective, which is more general and effective than the myopic approach adopted in this paper. \n\n[1] Wen Shen, Henger Li, and Zizhan Zheng. Learning to Attack Distributionally Robust Federated Learning. NeurIPS-20 Workshop on Scalability, Privacy, and Security in Federated Learning (SpicyFL). \n\n\n\n\n", "summary_of_the_review": "The idea of simulating benign users' behavior to develop more advanced attacks against federated learning is interesting. However, the paper considers an over-simplified setting, and the proposed backdoor attack is only effective against the norm-bounding defense and completely fails against other common defenses. The proposed solution is far from being sophisticated. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666648431975}], "openreview_url": "https://openreview.net/forum?id=B7HJ9KLFV9U", "arxiv_id": "2210.09305", "paper_pdf": "papers/B7HJ9KLFV9U.pdf", "paper_pdf_sha256": "6c38085c5a6c64e5bb66eca85ef429e62c799332d52a72ae6099faf74ac743b2", "paper_pdf_bytes": 1590399, "paper_pdf_source": "openreview", "code_url": "https://github.com/YuxinWenRick/thinking-two-moves-ahead", "code_repository": "YuxinWenRick/thinking-two-moves-ahead", "code_commit": "091f08b2af80e3454eb11fb0207cdba77a2fd902", "code_archive": "repos/B7HJ9KLFV9U.zip", "code_archive_sha256": "0fb3a04665df509fce3d40c7f9d46003ffb5d3c96a54f4b3804a1793a0a91f99", "code_archive_bytes": 35960, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 28, "github_languages": {"Python": 101761}, "github_archived": false, "github_pushed_at": "2022-10-18T01:02:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/thinking-two-moves-ahead-anticipating-other"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YDqIYJBQTQs", "year": 2022, "status": "rejected", "title": "Unsupervised Object Learning via Common Fate", "authors": ["Matthias Tangemann", "Steffen Schneider", "Julius Von Kügelgen", "Francesco Locatello", "Peter Vincent Gehler", "Thomas Brox", "Matthias Kuemmerer", "Matthias Bethge", "Bernhard Schölkopf"], "authorids": ["~Matthias_Tangemann1", "~Steffen_Schneider1", "~Julius_Von_Kügelgen1", "~Francesco_Locatello1", "~Peter_Vincent_Gehler1", "~Thomas_Brox1", "~Matthias_Kuemmerer1", "~Matthias_Bethge1", "~Bernhard_Schölkopf1"], "authors_source": "OpenReview API", "abstract": "Learning generative object models from unlabelled videos is a long standing problem and is required for causal scene modeling. We decompose this problem into three easier subtasks, and provide candidate solutions for each of them. Inspired by the Common Fate Principle of Gestalt Psychology, we first extract (noisy) masks of moving objects via unsupervised motion segmentation. Second, generative models are trained on the masks of the background and the moving objects, respectively. Third, background and foreground models are combined in a conditional ``dead leaves scene model to sample novel scene configurations where occlusions and depth layering  arise naturally. To evaluate the individual stages, we introduce the Fishbowl dataset positioned between complex real-world scenes and common object-centric benchmarks of simplistic objects. We show that our approach allows learning generative models that generalize beyond the occlusions present in the input videos, and represent scenes in a modular fashion that allows sampling plausible scenes outside the training distribution by permitting, for instance, object numbers or densities not observed in the training set.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "njK-2VFVbLN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1605/Reviewer_GUfQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper provides a multi-stage solution to unsupervised, frame-wise segmentation in videos.\nThe common fate heuristic is used to provide initial object detections and segmentations.\nThen VAE based model is used to refine the initial results.\nA new simulated dataset is proposed.\n\nResults show that the proposed method successfully segments out moving objects and is highly scalable.", "review_text": "Strengths:\n1. I appreciate the effort of building a new dataset.\n2. Compared with other methods, the scalability of the proposed method is indeed strong.\n3. I am convinced that the direct application of \"the common fate heuristic\" is novel and is of great potential.\n4. While most related work can only generate masks for occluded (partial) objects, the proposed method can generate masks for full objects occluded or not.\n5. The ability to sample novel scenes is impressive.\n\nWeakness:\nI apologize if some of my concerns are already detailed in the paper and I fail to pick them out.\n1. Discussion of related works.\n    While the author did a comprehensive literature review, I am not convinced by some of the claims.\n    1): I am not sure why \"the spatial mixture model does not match the true, underlying scene generation process\".\n         For example, SPACE employs a depth prediction to predict the relative depth between objects.\n         Then the depth information is integrated into the mask value.\n         SPACE effectively transforms a 2D segmentation task into 2.5D.\n    2): \"End-to-End object-centric approaches are difficult to train.\"\n         Without further explanation, I am not sure what \"difficult to train\" means under the context.\n    3): \"Full scene models trained in an end-to-end fashion are not shown to scale to a more realistic dataset yet.\"\n         As demonstrated by the paper SCALOR: GENERATIVE WORLD MODELS WITH SCALABLE OBJECT REPRESENTATIONS (which is already cited by the author), such method can be applied to challenging cases.\n\n2. Dataset\nThe dataset proposed is indeed interesting and challenging.\nMy only question is that is it true that all foreground objects are moving all the time and the backgrounds do not have any moving parts?\n\n3. Baseline\nIt is surprising that no baseline model is evaluated on the same dataset.\nI understand that there are not many works targeting exactly the same setup, but works like SCALOR or SPACE can also be used as a baseline for numerical results.\nAfter all, being able to segment without \"the common fate\" heuristic and a rule based object proposal stage seems like an advanrages to me.\nTo show that the loss function proposed in this work is better, the author can modify SPACE to take bounding boxes proposed by the rule based motion segmentation pipeline. This should be an easy modification.\nIf under this modification the performance of SPACE is better than before but still behind this paper, then the method propsed in this work is truely great.\nSimilar modification can be done to SCALOR as well I think.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides a multi-stage solution to unsupervised, frame-wise segmentation in videos.\nThe common fate heuristic is used to provide initial object detections and segmentations.\nThen VAE based model is used to refine the initial results.\nA new simulated dataset is proposed.\n\nResults show that the proposed method successfully segments out moving objects and is highly scalable.", "main_review": "Strengths:\n1. I appreciate the effort of building a new dataset.\n2. Compared with other methods, the scalability of the proposed method is indeed strong.\n3. I am convinced that the direct application of \"the common fate heuristic\" is novel and is of great potential.\n4. While most related work can only generate masks for occluded (partial) objects, the proposed method can generate masks for full objects occluded or not.\n5. The ability to sample novel scenes is impressive.\n\nWeakness:\nI apologize if some of my concerns are already detailed in the paper and I fail to pick them out.\n1. Discussion of related works.\n    While the author did a comprehensive literature review, I am not convinced by some of the claims.\n    1): I am not sure why \"the spatial mixture model does not match the true, underlying scene generation process\".\n         For example, SPACE employs a depth prediction to predict the relative depth between objects.\n         Then the depth information is integrated into the mask value.\n         SPACE effectively transforms a 2D segmentation task into 2.5D.\n    2): \"End-to-End object-centric approaches are difficult to train.\"\n         Without further explanation, I am not sure what \"difficult to train\" means under the context.\n    3): \"Full scene models trained in an end-to-end fashion are not shown to scale to a more realistic dataset yet.\"\n         As demonstrated by the paper SCALOR: GENERATIVE WORLD MODELS WITH SCALABLE OBJECT REPRESENTATIONS (which is already cited by the author), such method can be applied to challenging cases.\n\n2. Dataset\nThe dataset proposed is indeed interesting and challenging.\nMy only question is that is it true that all foreground objects are moving all the time and the backgrounds do not have any moving parts?\n\n3. Baseline\nIt is surprising that no baseline model is evaluated on the same dataset.\nI understand that there are not many works targeting exactly the same setup, but works like SCALOR or SPACE can also be used as a baseline for numerical results.\nAfter all, being able to segment without \"the common fate\" heuristic and a rule based object proposal stage seems like an advanrages to me.\nTo show that the loss function proposed in this work is better, the author can modify SPACE to take bounding boxes proposed by the rule based motion segmentation pipeline. This should be an easy modification.\nIf under this modification the performance of SPACE is better than before but still behind this paper, then the method propsed in this work is truely great.\nSimilar modification can be done to SCALOR as well I think.\n", "summary_of_the_review": "The results provided in the paper is impressive.\nBut the authors argue multiple times that the method proposed in this work is better than end-to-end generative model without enough justifiation.\nThus I cannot say that this paper is above accept threshold at current stage.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635841828074}, {"id": "AfAXr7JHtVl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1605/Reviewer_3iwv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper introduces an object-centric generative model for visual scenes. The model decouples the problem into three tasks: 1) modelling the  2D appearance and shape of individual objects with a variational auto encoder; 2) same thing for background; 3) sampling the position, size and appearance of individual objects (i.e. scene composition) conditioned on the background. To my understanding, these three components are trained independently.\n\nContributions:\n1) Decoupling these three tasks allows for a more interpretable representation compared to existing end-to-end methods, which allows for better interactions with users (e.g. change the number of objects or their position) \n2) This decoupling potentially makes the third task easier, because it only needs to learn relationships between positions and size given a latent representation (as opposed to learning everything jointly).\n3) Training the tasks independently requires a way to obtain \"ground-truth\" object segmentations - the paper achieves this automatically from videos using motion segmentation with good results", "review_text": "I found the paper quite useful and interesting. I find it useful because the decomposition allows for better ablative analysis of the model performance compared to previous end-to-end approaches, and because it opens up additional practical applications, such as given user direct and interpretable control on the output generated by the model (e.g. manually input number of objects or an individual object position). I find the results interesting, particularly the thorough ablative analysis: comparing training on ground truth segmentation vs training on motion segmentation.\n\nThe paper has some weaknesses, e..g. some claims not thoroughly backed up, lack of clarity on some details. I list them here in detail, from more important to less important: \n1) The paper strongly couples two concepts: a) the choice of using a multi-stage learning approach rather than end-to-end; and b) the use of motion segmentation to enable training this model. In practice, it seems to me that these two things are not really that coupled, but rather motion segmentation is just a way to automatically generate the training data that the model needs, because unlike end-to-end method, this submission needs explicit object and segmentation masks (if I understand correctly). One thing I would like the authors to highlight more  is that, due to this, a limitation of this method is that it can be trained only on video data and not on static images, unless some form of segmentation ground truth is available or there is a way to automatically generate it -  this is mentioned in the conclusions, but I encourage making it clear earlier in the paper\n2) It is unclear if the parameters of the scene models learnt entirely independently from the foreground and background model, i.e. by learning the distribution of the latent vectors produced by the VAEs conditioned on the background? Or is there some gradient propagating between these steps\n3) von kugelgen et al (2020) also provide a controllable way to change the way the model generates the scenes (e.g. number and type of objects). Can you highlight the main differences? Is it mostly that in this submissions there is also control on scale and position?\n4) Results on the fishbowl dataset look quite impressive, less on the car dataset (I suggest adding some visuals from this dataset and of visual comparisons to existing methods in the main body of the paper). It would be interesting to see this method run on other datasets like the Atari dataset or the 3D room dataset from the SPACE paper, to further test generalisation of the multi-stage approach in other domains - do ground truth segmentations exist to test there? \n5) Understanding the performance of this methods on generating new scenes compared to existing work is quite challenging because it is done mostly qualitatively, although I recognise that is a challenge of the domain and existing methods also seems to suffer from this", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces an object-centric generative model for visual scenes. The model decouples the problem into three tasks: 1) modelling the  2D appearance and shape of individual objects with a variational auto encoder; 2) same thing for background; 3) sampling the position, size and appearance of individual objects (i.e. scene composition) conditioned on the background. To my understanding, these three components are trained independently.\n\nContributions:\n1) Decoupling these three tasks allows for a more interpretable representation compared to existing end-to-end methods, which allows for better interactions with users (e.g. change the number of objects or their position) \n2) This decoupling potentially makes the third task easier, because it only needs to learn relationships between positions and size given a latent representation (as opposed to learning everything jointly).\n3) Training the tasks independently requires a way to obtain \"ground-truth\" object segmentations - the paper achieves this automatically from videos using motion segmentation with good results", "main_review": "I found the paper quite useful and interesting. I find it useful because the decomposition allows for better ablative analysis of the model performance compared to previous end-to-end approaches, and because it opens up additional practical applications, such as given user direct and interpretable control on the output generated by the model (e.g. manually input number of objects or an individual object position). I find the results interesting, particularly the thorough ablative analysis: comparing training on ground truth segmentation vs training on motion segmentation.\n\nThe paper has some weaknesses, e..g. some claims not thoroughly backed up, lack of clarity on some details. I list them here in detail, from more important to less important: \n1) The paper strongly couples two concepts: a) the choice of using a multi-stage learning approach rather than end-to-end; and b) the use of motion segmentation to enable training this model. In practice, it seems to me that these two things are not really that coupled, but rather motion segmentation is just a way to automatically generate the training data that the model needs, because unlike end-to-end method, this submission needs explicit object and segmentation masks (if I understand correctly). One thing I would like the authors to highlight more  is that, due to this, a limitation of this method is that it can be trained only on video data and not on static images, unless some form of segmentation ground truth is available or there is a way to automatically generate it -  this is mentioned in the conclusions, but I encourage making it clear earlier in the paper\n2) It is unclear if the parameters of the scene models learnt entirely independently from the foreground and background model, i.e. by learning the distribution of the latent vectors produced by the VAEs conditioned on the background? Or is there some gradient propagating between these steps\n3) von kugelgen et al (2020) also provide a controllable way to change the way the model generates the scenes (e.g. number and type of objects). Can you highlight the main differences? Is it mostly that in this submissions there is also control on scale and position?\n4) Results on the fishbowl dataset look quite impressive, less on the car dataset (I suggest adding some visuals from this dataset and of visual comparisons to existing methods in the main body of the paper). It would be interesting to see this method run on other datasets like the Atari dataset or the 3D room dataset from the SPACE paper, to further test generalisation of the multi-stage approach in other domains - do ground truth segmentations exist to test there? \n5) Understanding the performance of this methods on generating new scenes compared to existing work is quite challenging because it is done mostly qualitatively, although I recognise that is a challenge of the domain and existing methods also seems to suffer from this", "summary_of_the_review": "While I do not see a significant amount of novel technical contributions, the proposed paradigm of breaking down scene generative models into multiple stages of learning that allows encoding prior structure and human-interpretable elements is valuable. It can help generate theoretical insights, and the ablative analysis provided here is a promising first step. It also has the potential to be practically useful, by allowing users more fine-grained controls on the parameters of the generated scenes.\n\nThere are some claims with minor issues, and some lack of clarity in some sections (see weaknesses in the review section above). All in all, I think the paper is above the acceptance threshold, and addressing these weaknesses would make me lean towards a clearer accept\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635810867529}, {"id": "WzAUlNtab_n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1605/Reviewer_dAqW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work proposes an object-centric generative model. It consists of motion segmentation, object model, background model, and scene model. The object model is trained to reconstruct an object as if it is not occluded. The authors have introduced a new dataset, called Fishbowl, which provides inmodal and amodal segmentation masks of objects. ", "review_text": "Object-centric generative models have gained relatively less attentions compared to the normal generation models. The authors have generated a new dataset for this research area. However, I have many concerns that make me hesitant to accept this paper.\n\nFIrst of all, the dataset is far from the realistic setting. It is related to missing literature search. There are many datasets for amodal segmentation. For example, https://openaccess.thecvf.com/content_CVPR_2019/papers/Hu_SAIL-VOS_Semantic_Amodal_Instance_Level_Video_Object_Segmentation_-_A_CVPR_2019_paper.pdf <- this paper basically proposes a more realistic dataset, which provides the same information (amodal masks). I assume the authors have not looked into this literature. There are many other amodal datasets, including KINS and COCO-A.\n\nMy second point is that the proposed model has not been evaluated well. I have gone through the paper as well as the supplementary, but there was no appropriate comparison to the previous works. It's hard to believe all the conventional algorithms fail at generating object-centric images.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes an object-centric generative model. It consists of motion segmentation, object model, background model, and scene model. The object model is trained to reconstruct an object as if it is not occluded. The authors have introduced a new dataset, called Fishbowl, which provides inmodal and amodal segmentation masks of objects. ", "main_review": "Object-centric generative models have gained relatively less attentions compared to the normal generation models. The authors have generated a new dataset for this research area. However, I have many concerns that make me hesitant to accept this paper.\n\nFIrst of all, the dataset is far from the realistic setting. It is related to missing literature search. There are many datasets for amodal segmentation. For example, https://openaccess.thecvf.com/content_CVPR_2019/papers/Hu_SAIL-VOS_Semantic_Amodal_Instance_Level_Video_Object_Segmentation_-_A_CVPR_2019_paper.pdf <- this paper basically proposes a more realistic dataset, which provides the same information (amodal masks). I assume the authors have not looked into this literature. There are many other amodal datasets, including KINS and COCO-A.\n\nMy second point is that the proposed model has not been evaluated well. I have gone through the paper as well as the supplementary, but there was no appropriate comparison to the previous works. It's hard to believe all the conventional algorithms fail at generating object-centric images.", "summary_of_the_review": "Because of the above two major weaknesses, I'm skeptical of accepting this paper. I am open to change my rating depends on the rebuttal. For now, my initial rating is reject.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634859001975}], "openreview_url": "https://openreview.net/forum?id=YDqIYJBQTQs", "arxiv_id": "2110.06562", "paper_pdf": "papers/YDqIYJBQTQs.pdf", "paper_pdf_sha256": "1f85961f59d9f5cac26cda30e6b2999bc30f1c7d73f183869378cdbb70cee170", "paper_pdf_bytes": 19945939, "paper_pdf_source": "openreview", "code_url": "https://github.com/mtangemann/common_fate_object_learning", "code_repository": "mtangemann/common_fate_object_learning", "code_commit": "9c14ac4aa151504e7d9f0a7be1810fee506eadf2", "code_archive": "repos/YDqIYJBQTQs.zip", "code_archive_sha256": "a612f84fe86220ca37621b3e9b1cd2af0efdd0a98b882ed76b53abd6a287c2de", "code_archive_bytes": 30770, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 26, "github_languages": {"Python": 74727}, "github_archived": false, "github_pushed_at": "2023-04-14T07:05:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unsupervised-object-learning-via-common-fate-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Svfh1_hYEtF", "year": 2021, "status": "rejected", "title": "Federated Continual Learning with Weighted Inter-client Transfer", "authors": ["Jaehong Yoon", "Wonyong Jeong", "Giwoong Lee", "Eunho Yang", "Sung Ju Hwang"], "authorids": ["~Jaehong_Yoon1", "~Wonyong_Jeong1", "~Giwoong_Lee1", "~Eunho_Yang1", "~Sung_Ju_Hwang1"], "authors_source": "OpenReview API", "abstract": "There has been a surge of interest in continual learning and federated learning, both of which are important in deep neural networks in real-world scenarios. Yet little research has been done regarding the scenario where each client learns on a sequence of tasks from a private local data stream.  This problem of federated continual learning poses new challenges to continual learning, such as utilizing knowledge from other clients, while preventing interference from irrelevant knowledge.  To resolve these issues, we propose a novel federated continual learning framework, Federated Weighted Inter-client Transfer (FedWeIT), which decomposes the network weights into global federated parameters and sparse task-specific parameters, and each client receives selective knowledge from other clients by taking a weighted combination of their task-specific parameters.FedWeITminimizes interference between incompatible tasks, and also allows positive knowledge transfer across clients during learning. We validate ourFedWeITagainst existing federated learning and continual learning methods under varying degrees of task similarity across clients, and our model significantly outperforms them with a large reduction in the communication cost.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "93zyu5Swt6G", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper496/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Combination of federated learning and continual learning is a timely problem. Given the trending popularity of either of them it's the time to tackle this problem. The problem that is tackled is nice but the proposed formulation looks incremental.\n\nThe paper is generally well written but has a few typos along the paper like:\nPage 2: \"... need to selective utilize ..\"\nPage 2: ... once when ...\"\nPage 7: \"dtaset\"\n\nIn section 3.1 the relation between tasks of the clients is not clear. Are the tasks at step i of all tasks related to each other? \n\nIn equation 1: In the third term --> Isn't there a case for transfer between the task of the same client to future tasks?\n\nSection 3.3 for training: Imagine we are at task t. Then, the minimization problem in equation 3 involves solving for A^i. Why do one need to find A^i for a previous task and find a new parameter for that? Isn't that task already gone?Although, does this mean that the parameters to be estimated at each step is growing with the number of tasks?\n\nThe NonIID data set is a cool combination of different smaller datasets.\n\n The experiments look overall good. However, the part that shows alleveriating catastrophic forgetting seems rather less elaborated. Why only 6th and 8th tasks? Why not demonstrate forgetting of task 1 over the time. Why not all tasks? And, Fig. 6 only shows up to 5 tasks. How many consecutive tasks does the proposed method handle?\n\nThis works look like an increment on the top of (Yoon et al, 2020). Like making their approach adapted to a federated learning case. Thus, the contribution of the works is rather limited. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposes a method to solve continual federated learning. It allows the models to leverage task specific information by other peers while preventing the negative interactions. The paper is tackling a cool problem but is borderline due to contributions and experiments. ", "review": "Combination of federated learning and continual learning is a timely problem. Given the trending popularity of either of them it's the time to tackle this problem. The problem that is tackled is nice but the proposed formulation looks incremental.\n\nThe paper is generally well written but has a few typos along the paper like:\nPage 2: \"... need to selective utilize ..\"\nPage 2: ... once when ...\"\nPage 7: \"dtaset\"\n\nIn section 3.1 the relation between tasks of the clients is not clear. Are the tasks at step i of all tasks related to each other? \n\nIn equation 1: In the third term --> Isn't there a case for transfer between the task of the same client to future tasks?\n\nSection 3.3 for training: Imagine we are at task t. Then, the minimization problem in equation 3 involves solving for A^i. Why do one need to find A^i for a previous task and find a new parameter for that? Isn't that task already gone?Although, does this mean that the parameters to be estimated at each step is growing with the number of tasks?\n\nThe NonIID data set is a cool combination of different smaller datasets.\n\n The experiments look overall good. However, the part that shows alleveriating catastrophic forgetting seems rather less elaborated. Why only 6th and 8th tasks? Why not demonstrate forgetting of task 1 over the time. Why not all tasks? And, Fig. 6 only shows up to 5 tasks. How many consecutive tasks does the proposed method handle?\n\nThis works look like an increment on the top of (Yoon et al, 2020). Like making their approach adapted to a federated learning case. Thus, the contribution of the works is rather limited. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604855892406}, {"id": "nZ5JjMSgkHj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper496/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper investigates a new problem – federated continual learning by Federated Weighted Inter-client Transfer. The key idea is to decompose the network weights into global federated parameters and sparse task-specific parameters such that each client can selectively receive knowledge from other clients by taking a weighted combination of their task-specific parameters. The experiment results in two contrived datasets demonstrate the effectiveness of the proposed method.\n\nStrength:\n+ This paper is well presented and organized.\n+ The proposed federated continual learning framework is innovative and technically sound.\n+  The experiment results are solid.\n\nWeakness:\n- The optimization procedure for Eq. (2) is not provided.\n- Some details are missing (as shown below).\n\nThe following are some questions that I concern.\n\n1.\tThe authors address the importance of federated continual learning from the aspect of continual learning, but is there any difference between federated continual learning and federated learning considering that most existing federated learning methods (fedavg and fedprox) are agnostic of client id?\n2.\tHow to train alpha in Eq. (1)? Is this a learnable parameter with sigmoid activation? If yes, how to set the parameters for testing on different tasks? Moreover, the whole testing process is confusing to me.\n3.\tThe detailed optimization procedure for Eq. (1) is not provided.\n4.\tFor the training objective Eq. 2, intuitively, authors propose decomposable parameters and want Base parameters B to be sparse with the help of task adaptive parameters A.  However, it is interesting to see that there is no constraint to encourage B and A to focus on different aspects. I thus doubt that the communication efficiency brought by the sparse parameters mainly benefits from the sparsity constraints, i.e., 2nd term in Eq (2), but not the decomposable parameters. The authors are encouraged to show that the base parameters (m*B) are more sparse than the model trained with existing federated methods with similar sparsity constraints.  Otherwise, the communication efficiency of the proposed method would not stand. \n5.\t For the third term in Eq. (2), whether it is necessary to impose the constraint on  A_c^i (i<|t|), as in Eq. (1)?  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "This paper investigates a new problem – federated continual learning by Federated Weighted Inter-client Transfer. The key idea is to decompose the network weights into global federated parameters and sparse task-specific parameters such that each client can selectively receive knowledge from other clients by taking a weighted combination of their task-specific parameters. The experiment results in two contrived datasets demonstrate the effectiveness of the proposed method.\n\nStrength:\n+ This paper is well presented and organized.\n+ The proposed federated continual learning framework is innovative and technically sound.\n+  The experiment results are solid.\n\nWeakness:\n- The optimization procedure for Eq. (2) is not provided.\n- Some details are missing (as shown below).\n\nThe following are some questions that I concern.\n\n1.\tThe authors address the importance of federated continual learning from the aspect of continual learning, but is there any difference between federated continual learning and federated learning considering that most existing federated learning methods (fedavg and fedprox) are agnostic of client id?\n2.\tHow to train alpha in Eq. (1)? Is this a learnable parameter with sigmoid activation? If yes, how to set the parameters for testing on different tasks? Moreover, the whole testing process is confusing to me.\n3.\tThe detailed optimization procedure for Eq. (1) is not provided.\n4.\tFor the training objective Eq. 2, intuitively, authors propose decomposable parameters and want Base parameters B to be sparse with the help of task adaptive parameters A.  However, it is interesting to see that there is no constraint to encourage B and A to focus on different aspects. I thus doubt that the communication efficiency brought by the sparse parameters mainly benefits from the sparsity constraints, i.e., 2nd term in Eq (2), but not the decomposable parameters. The authors are encouraged to show that the base parameters (m*B) are more sparse than the model trained with existing federated methods with similar sparsity constraints.  Otherwise, the communication efficiency of the proposed method would not stand. \n5.\t For the third term in Eq. (2), whether it is necessary to impose the constraint on  A_c^i (i<|t|), as in Eq. (1)?  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604257088611}, {"id": "zz3iNwq-QFk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper496/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a federated continual learning setting where each node has a non-iid stream and a different dataset. The authors address this by extending the parameter decomposition method of Yoon et al\n\nStrengths:\n\n- The authors introduced a relevant new task that introduces concepts from continual learning to federated learning. The task setting seems practical overall as different nodes will often be following a different distribution\n-The algorithm proposed is interesting. I would like further discussion on the differences to Yoon et al\n-The results are good and evaluate a lot of  relevant factors\n\nWeakness:\n\n- Although the setting proposed is interesting, certain aspects of it seem artificial: although it seems very relevant that each node is a different dataset, strong non-iid behavior per node seems not realistic for many settings \n-What happens when some nodes start learning much earlier than others\n- A discussion of challenges in these settings and other potential methods\n- Results are shown for very simple LeNet architecture only\n- The algorithm proposed is interesting, but more motivation and other possible alternatives in this setting would improve the paper\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The authors propose a federated continual learning setting where each node has a non-iid stream and a different dataset. The authors address this by extending the parameter decomposition method of Yoon et al\n\nStrengths:\n\n- The authors introduced a relevant new task that introduces concepts from continual learning to federated learning. The task setting seems practical overall as different nodes will often be following a different distribution\n-The algorithm proposed is interesting. I would like further discussion on the differences to Yoon et al\n-The results are good and evaluate a lot of  relevant factors\n\nWeakness:\n\n- Although the setting proposed is interesting, certain aspects of it seem artificial: although it seems very relevant that each node is a different dataset, strong non-iid behavior per node seems not realistic for many settings \n-What happens when some nodes start learning much earlier than others\n- A discussion of challenges in these settings and other potential methods\n- Results are shown for very simple LeNet architecture only\n- The algorithm proposed is interesting, but more motivation and other possible alternatives in this setting would improve the paper\n\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604243182064}, {"id": "TWjlKeYv0k_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper496/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors present a federated continual learning framework. By decomposing the local client parameter, the method could alleviate the effect of negative transfer and improve efficiency. Empirical results partly show the effectiveness of the proposed algorithm. \n\n**Strength**\nTo my best knowledge, the problem setting is quite novel. Decomposing the parameter parts are also interesting. \n\n**Weakness**\nWriting is not good and many details are not clear. Some parts of the methodology lack explanation and the experiment results cannot support all claims of the authors. Below is my detailed opinion.\n\nEquation 1 describes the decomposition.  However, some notations are not clearly defined. For example, what is L, I, O£¿ I have to guess that L is the depth of the network, I represent the input dimension, O represents the output dimension after I reading it many times.  I am not sure whether the above guess is right. Also, A_c^{(t)} is sparse task-adaptive parameters, this is not clear. I am not sure the sparse means in task-level or for every matrix A_c^{t} is sparse. I have to guess from equation 2, the author tries to concatenate all A, and perform sparse constrain.  I am not sure whether the above guess is right since the author did not explain it well.\n\nIn this paper, it seems that the author assumes that when a new task comes is knew ahead. In practice, it is hard to get such prior knowledge. Also, in the experiment part, it seems that different clients receive different tasks simultaneously, which is also quite not realistic. In practice, the new tasks could come asynchronously.\n\nOn the 2nd line of page 6, the author claims the efficiency of the algorithm is boosted since it requires |C| * (R * |B| + A). Isn't B is the same shape as \\theta? I understand that the author assumes B is sparse, but we do not know how sparse it could be or can we recover the ground-truth sparse matrix or not. I can see that in the worst-case, without assumptions, this cannot improve efficiency. Empirical results only support that using the l1 constraint can output sparse results, while when this will work and correct is not clear in the methodology part. \n\nAlso, for this sparsity, lambda 1 is fixed across different tasks. This indicates different tasks have the same penalty. I do not see why this works since some tasks could be very similar to other tasks while some tasks are very distinct.  The same reason can be applied to lambda2 since the relationship between the current task and the previous task could be different. In figure 6, it seems that hyperparameters are very sensitive.\nThe plot of learned attention (Fig 5b).  I am not sure whether this is top-5 attention since there is no description (maybe I missed it). I assume that this is top-5 attention results. However, this cannot support the author that the proposed method can handle the negative transfer. First, empirically, why Traffic Sign has more weight than SVHN in the MNIST task? Moreover, This figure shows that empirically, using attention can focus on more important tasks/features and this should not be in the main contribution, and this is a well-known phenomenon. Has the ability to include the right task does not indicate excluding negative tasks.\n\nIn the middle plot of figure 6, I wish to see the whole task performance rather than selected tasks (this should be included in the appendix at least).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel problem setting but writing is unclear and concerns in empirical results. ", "review": "In this paper, the authors present a federated continual learning framework. By decomposing the local client parameter, the method could alleviate the effect of negative transfer and improve efficiency. Empirical results partly show the effectiveness of the proposed algorithm. \n\n**Strength**\nTo my best knowledge, the problem setting is quite novel. Decomposing the parameter parts are also interesting. \n\n**Weakness**\nWriting is not good and many details are not clear. Some parts of the methodology lack explanation and the experiment results cannot support all claims of the authors. Below is my detailed opinion.\n\nEquation 1 describes the decomposition.  However, some notations are not clearly defined. For example, what is L, I, O£¿ I have to guess that L is the depth of the network, I represent the input dimension, O represents the output dimension after I reading it many times.  I am not sure whether the above guess is right. Also, A_c^{(t)} is sparse task-adaptive parameters, this is not clear. I am not sure the sparse means in task-level or for every matrix A_c^{t} is sparse. I have to guess from equation 2, the author tries to concatenate all A, and perform sparse constrain.  I am not sure whether the above guess is right since the author did not explain it well.\n\nIn this paper, it seems that the author assumes that when a new task comes is knew ahead. In practice, it is hard to get such prior knowledge. Also, in the experiment part, it seems that different clients receive different tasks simultaneously, which is also quite not realistic. In practice, the new tasks could come asynchronously.\n\nOn the 2nd line of page 6, the author claims the efficiency of the algorithm is boosted since it requires |C| * (R * |B| + A). Isn't B is the same shape as \\theta? I understand that the author assumes B is sparse, but we do not know how sparse it could be or can we recover the ground-truth sparse matrix or not. I can see that in the worst-case, without assumptions, this cannot improve efficiency. Empirical results only support that using the l1 constraint can output sparse results, while when this will work and correct is not clear in the methodology part. \n\nAlso, for this sparsity, lambda 1 is fixed across different tasks. This indicates different tasks have the same penalty. I do not see why this works since some tasks could be very similar to other tasks while some tasks are very distinct.  The same reason can be applied to lambda2 since the relationship between the current task and the previous task could be different. In figure 6, it seems that hyperparameters are very sensitive.\nThe plot of learned attention (Fig 5b).  I am not sure whether this is top-5 attention since there is no description (maybe I missed it). I assume that this is top-5 attention results. However, this cannot support the author that the proposed method can handle the negative transfer. First, empirically, why Traffic Sign has more weight than SVHN in the MNIST task? Moreover, This figure shows that empirically, using attention can focus on more important tasks/features and this should not be in the main contribution, and this is a well-known phenomenon. Has the ability to include the right task does not indicate excluding negative tasks.\n\nIn the middle plot of figure 6, I wish to see the whole task performance rather than selected tasks (this should be included in the appendix at least).", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603946759878}], "openreview_url": "https://openreview.net/forum?id=Svfh1_hYEtF", "arxiv_id": "2003.03196", "paper_pdf": "papers/Svfh1_hYEtF.pdf", "paper_pdf_sha256": "82e13bf797edce2ab40efdf4256086e031b35ff223468c09ab46f50b7c5166cd", "paper_pdf_bytes": 2101073, "paper_pdf_source": "openreview", "code_url": "https://github.com/wyjeong/FedWeIT", "code_repository": "wyjeong/FedWeIT", "code_commit": "83c0bcd10b4934abe3af3a9b6d2e5b77566b9567", "code_archive": "repos/Svfh1_hYEtF.zip", "code_archive_sha256": "7fcd6881ceb2f2209f35dc0127c326927fdb63bfafe6eccbb1d353d4fae5851a", "code_archive_bytes": 29425, "code_file_count": 15, "code_extensions": {".py": 13, ".sh": 2}, "github_disk_usage_kb": 30, "github_languages": {"Python": 66487, "Shell": 770}, "github_archived": false, "github_pushed_at": "2021-06-02T05:12:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/federated-continual-learning-with-adaptive"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJe9fTNtPS", "year": 2020, "status": "rejected", "title": "AHash: A Load-Balanced One Permutation Hash", "authors": ["Chenxingyu Zhao", "Jie Gui", "Yixiao Guo", "Jie Jiang", "Tong Yang", "Bin Cui", "Gong Zhang"], "authorids": ["dkzcxy@pku.edu.cn", "guisj2017@pku.edu.cn", "1700016637@pku.edu.cn", "jie.jiang@pku.edu.cn", "yangtongemail@gmail.com", "bin.cui@pku.edu.cn", "nicholas.zhang@huawei.com"], "authors_source": "OpenReview API", "abstract": "Minwise Hashing (MinHash) is a fundamental method to compute set similarities and compact high-dimensional data for efficient learning and searching. The bottleneck of MinHash is computing k (usually hundreds) MinHash values. One Permutation Hashing (OPH) only requires one permutation (hash function)  to get k MinHash values by dividing elements into k bins.  One drawback of OPH is that the load of the bins (the number of elements in a bin) could be unbalanced, which leads to the existence of empty bins and false similarity computation. Several strategies for densification, that is, filling empty bins, have been proposed. However, the densification is just a remedial strategy and cannot eliminate the error incurred by the unbalanced load. Unlike the densification to fill the empty bins after they undesirably occur, our design goal is to balance the load so as to reduce the empty bins in advance.  In this paper, we propose a load-balanced hashing, Amortization Hashing (AHash), which can generate as few empty bins as possible. Therefore, AHash is more load-balanced and accurate without hurting runtime efficiency compared with OPH and densification strategies. Our experiments on real datasets validate the claim. All source codes and datasets have been provided as Supplementary Materials and released on GitHub anonymously.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "B1xjvMuC9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper423/AnonReviewer4"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors proposed an efficient minwise hashing method, namely Amortization Hashing (AHash), for solving the issue of the unbalanced loading during hash code generation. Specifically, in AHash, two steps, called Insertion and Amortization, are designed to balance the load in order to reduce the empty bins. In addition, a detailed theoretical analysis was provided. The authors evaluated the proposed AHash on three text categorization benchmarks, i.e., RCV1, NEWS20, and URL, comparing against the previous min-hashing algorithms including OOPH and MinHash. The experimental results show the effectiveness of the proposed hashing method.  \n\nSome comments on this work are listed as follows.\n1. This paper is well motivated and structured, which may serve as a guidance for other works.\n2. Three tasks, similarity estimating, large-scale learning, and fast near neighbor searching reduction, are adopted to validate the proposed AHash. The experimental results show that the proposed method can achieve state-of-the-art performance.\n3. It seems that the effectiveness of the proposed method is OK but not surprising. The improvement of AHash over OOPH is unobtrusive, given the results shown in Table 2, Fig. 3, and Fig. 4.\n4. The authors claimed that more significant performance can be achieved when the number of bins is appropriate on the task of computing similarities. Is there any theoretical guarantee for such an appropriate condition?\n5. The proposed method still cannot entirely solve the problem of empty bins, where a densifying operation may also be needed.\nMinor error:\nA typo exists in Eq. (10), where one “)” on the left-hand side is needless.\n\nOverall, the proposed AHash in this work is interesting and seems effective when dealing with the problem of the unbalanced load. However, there still exist quite a few issues in the current version, which need more explanations and clarifications.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "In this paper, the authors proposed an efficient minwise hashing method, namely Amortization Hashing (AHash), for solving the issue of the unbalanced loading during hash code generation. Specifically, in AHash, two steps, called Insertion and Amortization, are designed to balance the load in order to reduce the empty bins. In addition, a detailed theoretical analysis was provided. The authors evaluated the proposed AHash on three text categorization benchmarks, i.e., RCV1, NEWS20, and URL, comparing against the previous min-hashing algorithms including OOPH and MinHash. The experimental results show the effectiveness of the proposed hashing method.  \n\nSome comments on this work are listed as follows.\n1. This paper is well motivated and structured, which may serve as a guidance for other works.\n2. Three tasks, similarity estimating, large-scale learning, and fast near neighbor searching reduction, are adopted to validate the proposed AHash. The experimental results show that the proposed method can achieve state-of-the-art performance.\n3. It seems that the effectiveness of the proposed method is OK but not surprising. The improvement of AHash over OOPH is unobtrusive, given the results shown in Table 2, Fig. 3, and Fig. 4.\n4. The authors claimed that more significant performance can be achieved when the number of bins is appropriate on the task of computing similarities. Is there any theoretical guarantee for such an appropriate condition?\n5. The proposed method still cannot entirely solve the problem of empty bins, where a densifying operation may also be needed.\nMinor error:\nA typo exists in Eq. (10), where one “)” on the left-hand side is needless.\n\nOverall, the proposed AHash in this work is interesting and seems effective when dealing with the problem of the unbalanced load. However, there still exist quite a few issues in the current version, which need more explanations and clarifications.\n"}, "tcdate": 1572926050695}, {"id": "r1eEbe2q9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper423/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a bin split strategy for one permutation hashing which reduces the expected number of empty bins. The idea is to use two sets of hashes (odd and even) to make more use of the data. The strategy reduces the variance of existing densification scheme. Experiments show that the proposed method gives lower MSE, good classification accuracy and better NN search performance.\n\nOverall, the idea is interesting, although the improvements will only occur in very much corner cases. The wiring can be improved. The author also use ( ) instead of [] for citations which made the paper difficult to read (among with other writing problems). The theory (and its correctness) is difficult to be justified. The authors are suggested to use plots to verify the theoretical results. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes a bin split strategy for one permutation hashing which reduces the expected number of empty bins. The idea is to use two sets of hashes (odd and even) to make more use of the data. The strategy reduces the variance of existing densification scheme. Experiments show that the proposed method gives lower MSE, good classification accuracy and better NN search performance.\n\nOverall, the idea is interesting, although the improvements will only occur in very much corner cases. The wiring can be improved. The author also use ( ) instead of [] for citations which made the paper difficult to read (among with other writing problems). The theory (and its correctness) is difficult to be justified. The authors are suggested to use plots to verify the theoretical results. \n\n"}, "tcdate": 1572679676427}, {"id": "B1gLNgI95r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper423/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new hashing method based on MinHash. It aims to improve the computational complexity of the original MinHash and reduce the error of One Permutation Hashing. The proposed method is mainly assigning bins into paired bins, odd bins and even bins, and obtain two hash values, an even min and an odd min, for each bin. It finally reassigns nearby hash values to empty bins. The paper provides theoretical analyses that show the proposed method gives 1) unbiased estimator; 2) fewer empty bins 3) smaller variance. Finally the paper did some experiments on several applications including SVM and nearest neighbor search. The results show improvement over MinHash and OOPH. \n\nThe paper is mostly well written and easy to read. I have a few questions. \n\n1. My main concern is the lack of proof details. I tried to read through Proof 3.1. However, I found it is hard to follow. I'd suggest to add more details and explaination in the proof. For example, how is Eq. (17) obtained? \n\n2. The floor symbol \\lfloor, \\rfloor in Eq. (9) seems to be misused. It seems that the paper wants  $\\lfloor n/2 \\rfloor$ to return the largest even integer that is less than or equal to the given value, n. Right?\n\nOverall, I vote for weak accept as the proposed is novel and the theoretical results seem reasonable to me. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a new hashing method based on MinHash. It aims to improve the computational complexity of the original MinHash and reduce the error of One Permutation Hashing. The proposed method is mainly assigning bins into paired bins, odd bins and even bins, and obtain two hash values, an even min and an odd min, for each bin. It finally reassigns nearby hash values to empty bins. The paper provides theoretical analyses that show the proposed method gives 1) unbiased estimator; 2) fewer empty bins 3) smaller variance. Finally the paper did some experiments on several applications including SVM and nearest neighbor search. The results show improvement over MinHash and OOPH. \n\nThe paper is mostly well written and easy to read. I have a few questions. \n\n1. My main concern is the lack of proof details. I tried to read through Proof 3.1. However, I found it is hard to follow. I'd suggest to add more details and explaination in the proof. For example, how is Eq. (17) obtained? \n\n2. The floor symbol \\lfloor, \\rfloor in Eq. (9) seems to be misused. It seems that the paper wants  $\\lfloor n/2 \\rfloor$ to return the largest even integer that is less than or equal to the given value, n. Right?\n\nOverall, I vote for weak accept as the proposed is novel and the theoretical results seem reasonable to me. "}, "tcdate": 1572655150424}, {"id": "SJgOQ8pTKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper423/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "*** Summary ***\nMinHash is a well-known method for approximating set similarities in terms of the jaccard similarity. The idea is to use k random permutations (hashes) on all elements of the sets and check how often two sets hash into the same bucket. Larger values of k yield more accurate estimates of the set similarities but require more time to compute. One permutation hashing (OPH) aims to reduce the number of hash computations per element to 1 and maintaining bins. However, some of those bins may remain empty which negatively influences the similarity estimate. Optimal OPH (OOPH) hashes empty bins to non-empty bins and effectively reuses bins. This paper proposes Amortization Hashing (AHash) which reduces the occurrence of empty bins and thus leads to better similarity estimates.\n\n*** Evaluation ***\nThe paper proposes an interesting idea for approximating set similarities much faster than MinHash. However, I have some issues with the submission.\n\nI believe that the manuscript has a limited impact. The approach performs on par or marginally better as OOPH within the first and only reasonable experiment (4.2). As the authors state themselves on the page break 9/10, the advantage of AHash vanishes for small and large values of k. Hence, AHash only benefits of moderate choices of k. Moreover, I can see that OOPH might have a minor problem of estimating the set similarities which AHash aims to fix, but why should it outperform MinHash in terms of accuracy? Why are the pairs of set only chosen from RCV1? Why those particular set sizes? Why does no plot show standard deviation/error?\n\nThe remaining experiments yield very limited insight. Considering a linear SVM on standard datasets where the test error is >99.8% seems to be obsolete. In addition, the most important parameter k is held fix to an arbitrary value. Same holds for b. Since AHash only benefits from moderate sizes of k, why was k chosen in favor of AHash? The performance should definitely be shown in dependence of k. Instead, the most unimportant parameter (C) is varied. This should have been done in a proper cross-validation. Similar arguments hold for the near neighborhood search. What is the query set being used?\n\nThere are more flaws within the manuscript. The mathematical presentation is rather poor. The theorems lack text and assumptions and solely consist of equations. The corresponding proofs are also short on text and hard to follow. Unfortunately, there is no analysis of the expected error as a function of k. The proof of Theorem 3 is almost two pages and should be moved to the supplementary material since it does not provide much insight; it just distracts the reading flow. In addition, every equation is numbered but none is ever referenced. The citation style (numbers in round brackets) is really uncommon and can be easily confused with equation numbers. Most importantly, I want to note that a different font was used and that the spacing was clearly tricked in several places (e.g. within Section 4). This makes it especially hard to judge whether the manuscript has the correct length.\n\n\n*** Further Comments ***\n- The font was changed. It does not match the font of the other submissions.\n- The spacing is tricked in several places, especially in Section 4.\n- Links [1,29] should not be references but footnotes.\n- Citations should never be in round brackets like (1), because they can be confused with equation numbers. Instead they should be in square brackets like [1] or, more preferably, the natbib package should be used as in the ICLR style guidelines.\n- What does OOPH stand for? It is never stated.\n- Every equation has a number, but none is ever referenced.\n- Math/Equations are part of the text and should be treated as such, i.e., there should be proper punctuation marks.\n- What is a 2-universal hashing?\n- Algorithm 1: \"output range\" sounds like an interval whereas the number of distinct hash values is meant.\n- \"(14) proposed\", no past tense\n- Instead of \"(11) proposes\", please use \"Shrivastava and Li [11] propose\"\n- Why \"Theorem 1\" and \"Proof 3.1\"?\n- Why are the theorems lacking the assumptions and text? They basically consist of equations.\n- Theorem 3 should have a \"less or equal\" instead of a \"strictly less\".\n- Eq. (40): \"0andm\"\n- Why does Proof 3.3 have a end of proof sign (not right-aligned) but the other proofs don't?\n- None of the experimental results shows standard deviations/errors although the experiments are repeated several times. Why? It would be also nice to see whether the approximation tends to over- or underestimate J. This could be done with a violin plot.\n- How are the pairs of sets in Section 4.2 chosen and why only from RCV1? This seems to be the most important experiment.\n- Why is k (and b) fixed to an arbitrary value in the remaining experiments? Please select C within a proper cross-validation.\n- There are a lot of enumerations which unnecessarily make the manuscript longer, e.g. in Sections 1.4, 2.1 and 4.1. In addition, the proof of Theorem 3 almost takes two pages but is not super informative. It should be moved to the supplementary material. This in combination with the font mismatch makes it difficult to determine the real length of the submission.\n- It is nice that the source code is published online, but uncommented c++ code is not really helpful.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "*** Summary ***\nMinHash is a well-known method for approximating set similarities in terms of the jaccard similarity. The idea is to use k random permutations (hashes) on all elements of the sets and check how often two sets hash into the same bucket. Larger values of k yield more accurate estimates of the set similarities but require more time to compute. One permutation hashing (OPH) aims to reduce the number of hash computations per element to 1 and maintaining bins. However, some of those bins may remain empty which negatively influences the similarity estimate. Optimal OPH (OOPH) hashes empty bins to non-empty bins and effectively reuses bins. This paper proposes Amortization Hashing (AHash) which reduces the occurrence of empty bins and thus leads to better similarity estimates.\n\n*** Evaluation ***\nThe paper proposes an interesting idea for approximating set similarities much faster than MinHash. However, I have some issues with the submission.\n\nI believe that the manuscript has a limited impact. The approach performs on par or marginally better as OOPH within the first and only reasonable experiment (4.2). As the authors state themselves on the page break 9/10, the advantage of AHash vanishes for small and large values of k. Hence, AHash only benefits of moderate choices of k. Moreover, I can see that OOPH might have a minor problem of estimating the set similarities which AHash aims to fix, but why should it outperform MinHash in terms of accuracy? Why are the pairs of set only chosen from RCV1? Why those particular set sizes? Why does no plot show standard deviation/error?\n\nThe remaining experiments yield very limited insight. Considering a linear SVM on standard datasets where the test error is >99.8% seems to be obsolete. In addition, the most important parameter k is held fix to an arbitrary value. Same holds for b. Since AHash only benefits from moderate sizes of k, why was k chosen in favor of AHash? The performance should definitely be shown in dependence of k. Instead, the most unimportant parameter (C) is varied. This should have been done in a proper cross-validation. Similar arguments hold for the near neighborhood search. What is the query set being used?\n\nThere are more flaws within the manuscript. The mathematical presentation is rather poor. The theorems lack text and assumptions and solely consist of equations. The corresponding proofs are also short on text and hard to follow. Unfortunately, there is no analysis of the expected error as a function of k. The proof of Theorem 3 is almost two pages and should be moved to the supplementary material since it does not provide much insight; it just distracts the reading flow. In addition, every equation is numbered but none is ever referenced. The citation style (numbers in round brackets) is really uncommon and can be easily confused with equation numbers. Most importantly, I want to note that a different font was used and that the spacing was clearly tricked in several places (e.g. within Section 4). This makes it especially hard to judge whether the manuscript has the correct length.\n\n\n*** Further Comments ***\n- The font was changed. It does not match the font of the other submissions.\n- The spacing is tricked in several places, especially in Section 4.\n- Links [1,29] should not be references but footnotes.\n- Citations should never be in round brackets like (1), because they can be confused with equation numbers. Instead they should be in square brackets like [1] or, more preferably, the natbib package should be used as in the ICLR style guidelines.\n- What does OOPH stand for? It is never stated.\n- Every equation has a number, but none is ever referenced.\n- Math/Equations are part of the text and should be treated as such, i.e., there should be proper punctuation marks.\n- What is a 2-universal hashing?\n- Algorithm 1: \"output range\" sounds like an interval whereas the number of distinct hash values is meant.\n- \"(14) proposed\", no past tense\n- Instead of \"(11) proposes\", please use \"Shrivastava and Li [11] propose\"\n- Why \"Theorem 1\" and \"Proof 3.1\"?\n- Why are the theorems lacking the assumptions and text? They basically consist of equations.\n- Theorem 3 should have a \"less or equal\" instead of a \"strictly less\".\n- Eq. (40): \"0andm\"\n- Why does Proof 3.3 have a end of proof sign (not right-aligned) but the other proofs don't?\n- None of the experimental results shows standard deviations/errors although the experiments are repeated several times. Why? It would be also nice to see whether the approximation tends to over- or underestimate J. This could be done with a violin plot.\n- How are the pairs of sets in Section 4.2 chosen and why only from RCV1? This seems to be the most important experiment.\n- Why is k (and b) fixed to an arbitrary value in the remaining experiments? Please select C within a proper cross-validation.\n- There are a lot of enumerations which unnecessarily make the manuscript longer, e.g. in Sections 1.4, 2.1 and 4.1. In addition, the proof of Theorem 3 almost takes two pages but is not super informative. It should be moved to the supplementary material. This in combination with the font mismatch makes it difficult to determine the real length of the submission.\n- It is nice that the source code is published online, but uncommented c++ code is not really helpful.\n"}, "tcdate": 1571833376087}], "openreview_url": "https://openreview.net/forum?id=rJe9fTNtPS", "arxiv_id": null, "paper_pdf": "papers/rJe9fTNtPS.pdf", "paper_pdf_sha256": "5ddec72d536b9df6913d1d2a641b3dd1eea670d01b7082c35df378e8c83bd73b", "paper_pdf_bytes": 997501, "paper_pdf_source": "openreview", "code_url": "https://github.com/AHashCodes/AHash", "code_repository": "AHashCodes/AHash", "code_commit": "6d3fcaedb10ae0f03b28f3e6127f3cdc6db5c463", "code_archive": "repos/rJe9fTNtPS.zip", "code_archive_sha256": "adc1d7c36b0f78176a096c288a66daf12a0fd08a1291bfe3b315a55e238b267e", "code_archive_bytes": 37242, "code_file_count": 22, "code_extensions": {".h": 16, ".cpp": 4, ".sh": 2}, "github_disk_usage_kb": 34, "github_languages": {"C++": 85815, "Shell": 1902, "Makefile": 393, "CMake": 374}, "github_archived": false, "github_pushed_at": "2019-06-08T12:09:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ahash-a-load-balanced-one-permutation-hash"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HyfyN30qt7", "year": 2019, "status": "rejected", "title": "NICE: noise injection and clamping estimation for neural network quantization", "authors": ["Chaim Baskin", "Natan Liss", "Yoav Chai", "Evgenii Zheltonozhskii", "Eli Schwartz", "Raja Girayes", "Avi Mendelson", "Alexander M.Bronstein"], "authorids": ["chaimbaskin@cs.technion.ac.il", "lissnatan@campus.technion.ac.il", "yoavchai1@mail.tau.ac.il", "evgeniizh@campus.technion.ac.il", "eli.shw@gmail.com", "raja@tauex.tau.ac.il", "avi.mendelson@tce.technion.ac.il", "bron@cs.technion.ac.il"], "authors_source": "OpenReview API", "abstract": "Convolutional Neural Networks (CNN) are very popular in many fields including computer vision, speech recognition, natural language processing, to name a few. Though deep learning leads to groundbreaking performance in these domains, the networks used are very demanding computationally and are far from real-time even on a GPU, which is not power efficient and therefore does not suit low power systems such as mobile devices. To overcome this challenge, some solutions have been proposed for quantizing the weights and activations of these networks, which accelerate the runtime significantly. Yet, this acceleration comes at the cost of a larger error. The NICE method proposed in this work trains quantized neural networks by noise injection and a learned clamping, which improve the accuracy. This leads to state-of-the-art results on various regression and classification tasks, e.g., ImageNet classification with architectures such as ResNet-18/34/50 with low as 3-bit weights and 3 -bit activations. We implement the proposed solution on an FPGA to demonstrate its applicability for low power real-time applications.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "B1xOIQXcnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1413/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "[Summary]\nNeural network quantization is can enable many practical applications for deep learning, therefore it is an important research problem. The paper claims two contributions: 1. Injecting noise during training to make it more robust to quantization errors. 2. Clamping the parameter values in a layer as well as the activation output, where the clamping interval is some multiple of the standard deviation about the mean, and the clamping interval is updated using the Straight through estimator. The main strength of the paper lies in the empirical results where the combination of techniques employed by the authors outperforms the SOTA methods in a compute of scenarios.\n\n[Pros]\nThe paper is working on an important problem area and as a technical report this work can be valuable in the industry. There is novelty in the particular combination of techniques that the authors have employed and some of the empirical results show the strength of the technique.\n\n[Cons]\nthe main contribution of the paper is a careful combination of existing techniques and the associated empirical results, therefore the experiments need to be strong. I noticed some strange omissions in the  results, and asked the authors for a reply via a public comment but they did not reply. Specifically,\n\na) On RESNET 34 the results for PACT 5,5 are not shown and JOINT and PACT on 3,3 on ResNet-34 are also not shown. Why are these results omitted?\n\nb) The noise+gradual training decreases performance on (layer-weight bitwidth, activation bitwidth)  =  (3, 3). But further experiments for table 2 where the nets do not use noise+gradual training is not shown. Currently the proposed recipe for quantizing nets does not seem to be all that better than existing methods and it hard to guess exactly what was the reason for the improved results in situations where the results were infact better. Why were these experiments omitted ?\n\nOverall the experimental results in the paper are weak and the novelty of the proposed methods is low.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Low novelty and weak experiments.", "review": "[Summary]\nNeural network quantization is can enable many practical applications for deep learning, therefore it is an important research problem. The paper claims two contributions: 1. Injecting noise during training to make it more robust to quantization errors. 2. Clamping the parameter values in a layer as well as the activation output, where the clamping interval is some multiple of the standard deviation about the mean, and the clamping interval is updated using the Straight through estimator. The main strength of the paper lies in the empirical results where the combination of techniques employed by the authors outperforms the SOTA methods in a compute of scenarios.\n\n[Pros]\nThe paper is working on an important problem area and as a technical report this work can be valuable in the industry. There is novelty in the particular combination of techniques that the authors have employed and some of the empirical results show the strength of the technique.\n\n[Cons]\nthe main contribution of the paper is a careful combination of existing techniques and the associated empirical results, therefore the experiments need to be strong. I noticed some strange omissions in the  results, and asked the authors for a reply via a public comment but they did not reply. Specifically,\n\na) On RESNET 34 the results for PACT 5,5 are not shown and JOINT and PACT on 3,3 on ResNet-34 are also not shown. Why are these results omitted?\n\nb) The noise+gradual training decreases performance on (layer-weight bitwidth, activation bitwidth)  =  (3, 3). But further experiments for table 2 where the nets do not use noise+gradual training is not shown. Currently the proposed recipe for quantizing nets does not seem to be all that better than existing methods and it hard to guess exactly what was the reason for the improved results in situations where the results were infact better. Why were these experiments omitted ?\n\nOverall the experimental results in the paper are weak and the novelty of the proposed methods is low.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541186383965}, {"id": "BJl1eHfFnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1413/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The article presents a method for quantization of deep neural networks for classification and regression, using three key parts: (i) noise injection to model the effect of quantization during forward inference, (ii) clamping with learned maximum activations to reduce the quantization bin size, and (iii) gradual quantization of blocks of the network, while previously quantized blocks remain unchanged. The method is evaluated on ImageNet, CIFAR-10, and a regression task, showing performance on-par or better than state-of-the-art methods for particular quantization bit size. Finally, the method is used for porting network onto a FPGA.\n\nThe paper addresses an important topic, because there is increasing interest in hardware-efficient implementations of deep neural networks. The method could be interesting for practitioners, because it does not interfere with the original training of the full-precision method, and can be applied later on.\n\nThe main weakness is that none of the proposed methods are entirely original, and the combination is rather ad-hoc than well-justified. For example, quantization noise has been considered in several previous articles, e.g. already in BinaryConnect (Courbariaux et al. 2015), although the novelty here is that the noise is explicitly added during the forward path. However, the choice of the Bernoulli mask with p=0.05 is not justified and might not work best for other tasks. The authors admit that gradual quantization has been proposed before, and clamping a ReLU is also not new, although here a new way to learn and initialize the clamping parameters is presented.\n\nThe article would be OK if the empirical results were really strong, but unfortunately they are not entirely convincing:\n1. The classification results are only for ResNet architectures, it remains unclear whether results would hold also for other architectures.\n2. The numerical results in Table 2 are very close to each other, and no error bars are available, so it is not possible to judge whether differences are significant. Also, the advantage of the NICE method vanishes for 3-bit models.\n3. The results for CIFAR-10 come without any comparison.\n4. The results for regression are only compared to a single method, which is re-implemented by the authors, and might therefore not be fully optimized. Thus there is no strong baseline to judge the results.\n5. No results are shown for the hardware implementation.\n\nOverall, the paper is not a particularly interesting read for people interested in a deeper understanding of network quantization, but the method could still be valuable for applications. Is this sufficient for ICLR? Since the experimental results do not entirely convince me I will put my grade slightly below acceptance threshold.\n\nMinor points:\n- The abstract has a pretty long introduction before it begins to tell what the contributions of the article are.\n- Occasional grammar mistakes.\n- Tables 1 and 2 are misplaced\n\nPros:\n+ important topic (network quantization)\n+ good empirical results\n+ easy to apply\n\nCons:\n- combination of previously proposed methods\n- no convincing justification \n- no strong advantage over previous methods", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Network quantization on ResNets", "review": "The article presents a method for quantization of deep neural networks for classification and regression, using three key parts: (i) noise injection to model the effect of quantization during forward inference, (ii) clamping with learned maximum activations to reduce the quantization bin size, and (iii) gradual quantization of blocks of the network, while previously quantized blocks remain unchanged. The method is evaluated on ImageNet, CIFAR-10, and a regression task, showing performance on-par or better than state-of-the-art methods for particular quantization bit size. Finally, the method is used for porting network onto a FPGA.\n\nThe paper addresses an important topic, because there is increasing interest in hardware-efficient implementations of deep neural networks. The method could be interesting for practitioners, because it does not interfere with the original training of the full-precision method, and can be applied later on.\n\nThe main weakness is that none of the proposed methods are entirely original, and the combination is rather ad-hoc than well-justified. For example, quantization noise has been considered in several previous articles, e.g. already in BinaryConnect (Courbariaux et al. 2015), although the novelty here is that the noise is explicitly added during the forward path. However, the choice of the Bernoulli mask with p=0.05 is not justified and might not work best for other tasks. The authors admit that gradual quantization has been proposed before, and clamping a ReLU is also not new, although here a new way to learn and initialize the clamping parameters is presented.\n\nThe article would be OK if the empirical results were really strong, but unfortunately they are not entirely convincing:\n1. The classification results are only for ResNet architectures, it remains unclear whether results would hold also for other architectures.\n2. The numerical results in Table 2 are very close to each other, and no error bars are available, so it is not possible to judge whether differences are significant. Also, the advantage of the NICE method vanishes for 3-bit models.\n3. The results for CIFAR-10 come without any comparison.\n4. The results for regression are only compared to a single method, which is re-implemented by the authors, and might therefore not be fully optimized. Thus there is no strong baseline to judge the results.\n5. No results are shown for the hardware implementation.\n\nOverall, the paper is not a particularly interesting read for people interested in a deeper understanding of network quantization, but the method could still be valuable for applications. Is this sufficient for ICLR? Since the experimental results do not entirely convince me I will put my grade slightly below acceptance threshold.\n\nMinor points:\n- The abstract has a pretty long introduction before it begins to tell what the contributions of the article are.\n- Occasional grammar mistakes.\n- Tables 1 and 2 are misplaced\n\nPros:\n+ important topic (network quantization)\n+ good empirical results\n+ easy to apply\n\nCons:\n- combination of previously proposed methods\n- no convincing justification \n- no strong advantage over previous methods", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541117158683}, {"id": "HJeXr78fhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1413/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present a method for fine-tuning neural networks so inference can be performed in a quantized low bit data format down to 3 bits. The authors achieve this through a combination of three techniques:\n1. Noise injection to fine-tune the weights before quantization. The effect of noise injection can model that of quantization, but rather than being stuck in a quantization bin, fine grained weight updates are still possible\n2. A schedule that quantizes layer by layer, rather than all layers at the same time\n3. Clipping weights and activations within a learned range to obtain finer grained bins within that range. \n\n\nThe main contribution is a novel combination of mostly existing techniques. Clipping (or clamping as the authors call it) has been proposed by Zhang et al. 2018, but it's an interesting contribution to have the clipping learned directly via backpropagation with a straight-through estimator. Treating the quantization as noise has been proposed in a different form in McKinstry et al. 2018. Gradual quantization appears novel, but is also the least interesting of the techniques. Therefore, novelty on ideas/methods is somewhat limited, and the contribution is mostly the in the impressive experimental results, which appear to be outperforming previous methods. The main weaknesses are poor writing, and that some details of the implementation required to reproduce the results are missing. For example, the training schedule is not given, e.g. how many epochs to train the clean model, how many with noise, how many quantized. Details on the gradual quantization are also missing. Block based quantization is completely heuristic and not well motivated. If this is the main novel ingredient, more details on the mechanics would be needed. Is both the noise injection and the quantization done in blocks? If the motivation is in \"the opportunity to adapt”, then what does the adaptation look like? \n\nAs above, my other main issue is with the writing, there are many examples where I would suggest improvements:\n\nThis work could be improved greatly by copy editing for English grammar. There are many typos (including ones that can be caught by autocorrect, missing punctuation, or using similar but unrelated words, e.g. \"token\" instead of \"taken\"). The manuscript appears hastily put together and not ready for publication. \n\nThe acronym NICE already has a meaning in the DL literature: Dinh, L., Krueger, D., & Bengio, Y. (2014). NICE: Non-linear Independent Components Estimation. It confusing to reuse it. \n\nThe term clamping is only explained on page 4 but used since the abstract. It’s used in a nonstandard way to mean “constrained to lie within a range” which should be explained earlier. I think “clipped” would be a better term, following the related Choi et al. 2018. Clamping usually means \"constrained to a fixed value\" (not a range), so it is not a good term to use in this context. \n\nAre the results shown in table 2 and table 3 from a single trial or averaged across reruns? If single trial, it's misleading to have 2 figures after the decimal. Even non-quantized ResNet tends to have 0.5% or so run to run variability, which is much larger than the differences between some of the methods shown here. In fact, a lot of the results could just be due to picking a lucky random seed. \n\nComparisons are shown against methods JOINT (Jung et al), LQ-Nets (Zhang et al), FAQ (McKinstry et al). It would be helpful to present them with the same names in the \"related work\" section, and explain why they were picked out for the comparison. For someone not familiar with the literature it's hard to see why these 3 would be the obvious picks. \n\nReadability would increase if table 2 and 3 were moved to section 4 where they are referenced, rather than after the discussion. Fig 2 font size too small and hard to read. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "promising results but presentation needs some work", "review": "The authors present a method for fine-tuning neural networks so inference can be performed in a quantized low bit data format down to 3 bits. The authors achieve this through a combination of three techniques:\n1. Noise injection to fine-tune the weights before quantization. The effect of noise injection can model that of quantization, but rather than being stuck in a quantization bin, fine grained weight updates are still possible\n2. A schedule that quantizes layer by layer, rather than all layers at the same time\n3. Clipping weights and activations within a learned range to obtain finer grained bins within that range. \n\n\nThe main contribution is a novel combination of mostly existing techniques. Clipping (or clamping as the authors call it) has been proposed by Zhang et al. 2018, but it's an interesting contribution to have the clipping learned directly via backpropagation with a straight-through estimator. Treating the quantization as noise has been proposed in a different form in McKinstry et al. 2018. Gradual quantization appears novel, but is also the least interesting of the techniques. Therefore, novelty on ideas/methods is somewhat limited, and the contribution is mostly the in the impressive experimental results, which appear to be outperforming previous methods. The main weaknesses are poor writing, and that some details of the implementation required to reproduce the results are missing. For example, the training schedule is not given, e.g. how many epochs to train the clean model, how many with noise, how many quantized. Details on the gradual quantization are also missing. Block based quantization is completely heuristic and not well motivated. If this is the main novel ingredient, more details on the mechanics would be needed. Is both the noise injection and the quantization done in blocks? If the motivation is in \"the opportunity to adapt”, then what does the adaptation look like? \n\nAs above, my other main issue is with the writing, there are many examples where I would suggest improvements:\n\nThis work could be improved greatly by copy editing for English grammar. There are many typos (including ones that can be caught by autocorrect, missing punctuation, or using similar but unrelated words, e.g. \"token\" instead of \"taken\"). The manuscript appears hastily put together and not ready for publication. \n\nThe acronym NICE already has a meaning in the DL literature: Dinh, L., Krueger, D., & Bengio, Y. (2014). NICE: Non-linear Independent Components Estimation. It confusing to reuse it. \n\nThe term clamping is only explained on page 4 but used since the abstract. It’s used in a nonstandard way to mean “constrained to lie within a range” which should be explained earlier. I think “clipped” would be a better term, following the related Choi et al. 2018. Clamping usually means \"constrained to a fixed value\" (not a range), so it is not a good term to use in this context. \n\nAre the results shown in table 2 and table 3 from a single trial or averaged across reruns? If single trial, it's misleading to have 2 figures after the decimal. Even non-quantized ResNet tends to have 0.5% or so run to run variability, which is much larger than the differences between some of the methods shown here. In fact, a lot of the results could just be due to picking a lucky random seed. \n\nComparisons are shown against methods JOINT (Jung et al), LQ-Nets (Zhang et al), FAQ (McKinstry et al). It would be helpful to present them with the same names in the \"related work\" section, and explain why they were picked out for the comparison. For someone not familiar with the literature it's hard to see why these 3 would be the obvious picks. \n\nReadability would increase if table 2 and 3 were moved to section 4 where they are referenced, rather than after the discussion. Fig 2 font size too small and hard to read. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1540674362837}], "openreview_url": "https://openreview.net/forum?id=HyfyN30qt7", "arxiv_id": "1810.00162", "paper_pdf": "papers/HyfyN30qt7.pdf", "paper_pdf_sha256": "30eef30e6bed8b39487d0e58273ad09bd345b1cd5322d7212a4db197ca032ab2", "paper_pdf_bytes": 336064, "paper_pdf_source": "openreview", "code_url": "https://github.com/LsNatan/NICE", "code_repository": "LsNatan/NICE", "code_commit": "1f5ef43c9442931dd26e05a0dd68e6602f28f569", "code_archive": "repos/HyfyN30qt7.zip", "code_archive_sha256": "7e41888fb55c7235eec2a40649c05afbef40b58ecc971157b49abf523694dede", "code_archive_bytes": 97409, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 89, "github_languages": {"Python": 246385}, "github_archived": false, "github_pushed_at": "2018-10-01T11:15:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/nice-noise-injection-and-clamping-estimation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HkepKG-Rb", "year": 2018, "status": "rejected", "title": "A Semantic Loss Function for Deep Learning with Symbolic Knowledge", "authors": ["Jingyi Xu", "Zilu Zhang", "Tal Friedman", "Yitao Liang", "Guy Van den Broeck"], "authorids": ["jixu@g.ucla.edu", "zhangzilu@pku.edu.cn", "tal@cs.ucla.edu", "yliang@cs.ucla.edu", "guyvdb@cs.ucla.edu"], "authors_source": "OpenReview API", "abstract": "This paper develops a novel methodology for using symbolic knowledge in deep learning. From first principles, we derive a semantic loss function that bridges between neural output vectors and logical constraints. This loss function captures how close the neural network is to satisfying the constraints on its output. An experimental evaluation shows that our semantic loss function effectively guides the learner to achieve (near-)state-of-the-art results on semi-supervised multi-class classification. Moreover, it significantly increases the ability of the neural network to predict structured objects, such as rankings and shortest paths. These discrete concepts are tremendously difficult to learn, and benefit from a tight integration of deep learning and symbolic reasoning methods.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkRvF__xf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper891/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "SUMMARY \n\nThe paper proposes a new form of regularization utilizing logical constraints. The semantic loss function is built on the exploitation of symbolic knowledge extracted from data and connecting the logical constraints to the outputs of a neural network. The use of Boolean logic as a constraint provides a secondary regularization term to prevent over-fitting and improve predictions. The benefit of using the function is found primarily with semi-supervised tasks where data is partially unlabelled. The logical constraints provided by the semantic loss function allow for improved classification of unlabeled data.\nOutput constraints for the semantic loss function are represented with one-hot encoding, prefer- ence rankings, and paths in a grid. These three different output constraints are designed to explore different learning purposes. The semantic function was tested on both semi-supervised classifica- tion tasks as well as structure learning. The paper primarily focuses on the one-hot encoding constraint as it is viewed as a capable technique for multi-class classification.\n\nPOSITIVES \n\nIn terms of structure, the paper was written very well. Sufficient background information was con- veyed which helped in understanding the proposed semantic loss function. A thorough breakdown is also carried out on the semantic loss function itself by explaining its axioms which help explain how the outputs of a neural network match a given constraint.\nAs a scientific contribution, I would say results from the experiments were able to justify the proposal of the semantic loss function. The function was able to perform better than most other implementations for semi-supervised learning tasks, and the function was tested on multiple datasets. The paper also made use of testing the function against other notable machine learning approaches, and in most cases the function performed better, but this usually was confined to semi-supervised learning tasks. During supervised learning tasks the function did not perform markedly better than older implementations. Given that, the semantic loss function did prove to be a seemingly simple approach to improving semi-supervised classification tasks.\n• The background section covers the knowledge required in understanding the semantic loss function. The paper also clearly explains the meaning for some of the notation used in the definitions.\n• Experiments which clearly show the benefit of using the semantic loss function. Multiple experiment types were done as well which showed evidence of the broad applicability of the function.\n• In depth description of the definitions, axioms, and propositions of the semantic loss function.\n• A large number of experiments exploring the usefulness of the function for multiple learning tasks, and on multiple datasets.\n\nNEGATIVES \n\nI was not clear if the logical constraints are to be instantiated before learning, i.e. they are defined by hand prior to being implemented in the neural network. This is a pretty important question and drastically changes the nature of the learning process. Beyond that complaint, the paper did not suffer from any critical issues. There were some issues with spelling, and the section titled ’Algorithm’ fails to clearly define a complete algorithm using the semantic loss function. It would have helped to have two algorithms. One defining the pipeline for the semantic loss function, and another showing the implementation of the function in a machine learning framework. The semantic loss function found success only in cases were the learning task was semi-supervised, and not in cases of total supervised learning. This is not a true negative, but an observation on the effectiveness of the function.\n\n- A few typos in the paper.\n- The axioms for the semantic loss function where defined but there seemed to be a lack of a clear algorithm provided showing the pipeline implementation of the semantic loss function.\n- While the semantic loss function does improve learning performance in most cases, the im- provements are confined to semi-supervised learning tasks, and with the MNIST dataset another methodology, Ladder Nets, was able to outperform the semantic loss function.\n\nRELATED WORK\n\nThe paper proposed that logic constraints applied to the output of neural networks have the capacity to improve semi-supervised classification tasks as well as finding the shortest path. In the introduction, the paper lists Zhiting Hu et al. paper titled Harnessing Deep Neural Networks with Logic Rules as an example of a similar approach. Hu et al. paper utilized logic constraints in conjunction with neural nets as well. A key difference was that Hu et al. applied their network architecture to supervised classification tasks. Since the performance of the current papers semantic loss function with supervised tasks did not improve upon other methods, it may benefit to utilize the research by Hu et al. as a means of direct comparison for supervised learning tasks, and possibly incorporate their methods with the semantic loss function in order to improve upon supervised learning tasks.\n\nCONCLUSION\n\nGiven the success of the semantic loss function with semi-supervised tasks, I would accept this paper. The semantic loss was able to improve learning with respect to the tested datasets, and the paper clearly described the properties of the functions. The paper would benefit by including a more concrete algorithm describing the flow of data through a given neural net to the semantic loss function, as well as the process by which the semantic loss function constrains the data based on propositional logic, but in general this complaint is more nit picking. The semantic loss function and the experiments which tested the function showed clearly that there is a benefit to this research and there are areas for it to improve.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Deep learning with symbolic information", "rating": "7: Good paper, accept", "review": "SUMMARY \n\nThe paper proposes a new form of regularization utilizing logical constraints. The semantic loss function is built on the exploitation of symbolic knowledge extracted from data and connecting the logical constraints to the outputs of a neural network. The use of Boolean logic as a constraint provides a secondary regularization term to prevent over-fitting and improve predictions. The benefit of using the function is found primarily with semi-supervised tasks where data is partially unlabelled. The logical constraints provided by the semantic loss function allow for improved classification of unlabeled data.\nOutput constraints for the semantic loss function are represented with one-hot encoding, prefer- ence rankings, and paths in a grid. These three different output constraints are designed to explore different learning purposes. The semantic function was tested on both semi-supervised classifica- tion tasks as well as structure learning. The paper primarily focuses on the one-hot encoding constraint as it is viewed as a capable technique for multi-class classification.\n\nPOSITIVES \n\nIn terms of structure, the paper was written very well. Sufficient background information was con- veyed which helped in understanding the proposed semantic loss function. A thorough breakdown is also carried out on the semantic loss function itself by explaining its axioms which help explain how the outputs of a neural network match a given constraint.\nAs a scientific contribution, I would say results from the experiments were able to justify the proposal of the semantic loss function. The function was able to perform better than most other implementations for semi-supervised learning tasks, and the function was tested on multiple datasets. The paper also made use of testing the function against other notable machine learning approaches, and in most cases the function performed better, but this usually was confined to semi-supervised learning tasks. During supervised learning tasks the function did not perform markedly better than older implementations. Given that, the semantic loss function did prove to be a seemingly simple approach to improving semi-supervised classification tasks.\n• The background section covers the knowledge required in understanding the semantic loss function. The paper also clearly explains the meaning for some of the notation used in the definitions.\n• Experiments which clearly show the benefit of using the semantic loss function. Multiple experiment types were done as well which showed evidence of the broad applicability of the function.\n• In depth description of the definitions, axioms, and propositions of the semantic loss function.\n• A large number of experiments exploring the usefulness of the function for multiple learning tasks, and on multiple datasets.\n\nNEGATIVES \n\nI was not clear if the logical constraints are to be instantiated before learning, i.e. they are defined by hand prior to being implemented in the neural network. This is a pretty important question and drastically changes the nature of the learning process. Beyond that complaint, the paper did not suffer from any critical issues. There were some issues with spelling, and the section titled ’Algorithm’ fails to clearly define a complete algorithm using the semantic loss function. It would have helped to have two algorithms. One defining the pipeline for the semantic loss function, and another showing the implementation of the function in a machine learning framework. The semantic loss function found success only in cases were the learning task was semi-supervised, and not in cases of total supervised learning. This is not a true negative, but an observation on the effectiveness of the function.\n\n- A few typos in the paper.\n- The axioms for the semantic loss function where defined but there seemed to be a lack of a clear algorithm provided showing the pipeline implementation of the semantic loss function.\n- While the semantic loss function does improve learning performance in most cases, the im- provements are confined to semi-supervised learning tasks, and with the MNIST dataset another methodology, Ladder Nets, was able to outperform the semantic loss function.\n\nRELATED WORK\n\nThe paper proposed that logic constraints applied to the output of neural networks have the capacity to improve semi-supervised classification tasks as well as finding the shortest path. In the introduction, the paper lists Zhiting Hu et al. paper titled Harnessing Deep Neural Networks with Logic Rules as an example of a similar approach. Hu et al. paper utilized logic constraints in conjunction with neural nets as well. A key difference was that Hu et al. applied their network architecture to supervised classification tasks. Since the performance of the current papers semantic loss function with supervised tasks did not improve upon other methods, it may benefit to utilize the research by Hu et al. as a means of direct comparison for supervised learning tasks, and possibly incorporate their methods with the semantic loss function in order to improve upon supervised learning tasks.\n\nCONCLUSION\n\nGiven the success of the semantic loss function with semi-supervised tasks, I would accept this paper. The semantic loss was able to improve learning with respect to the tested datasets, and the paper clearly described the properties of the functions. The paper would benefit by including a more concrete algorithm describing the flow of data through a given neural net to the semantic loss function, as well as the process by which the semantic loss function constrains the data based on propositional logic, but in general this complaint is more nit picking. The semantic loss function and the experiments which tested the function showed clearly that there is a benefit to this research and there are areas for it to improve.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511717221874}, {"id": "SJJw0N0eM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper891/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper suggest a method for including symbolic knowledge into the learning process. The symbolic knowledge is given as logical constraints which characterize the space of legal solutions.   This knowledge is \"injected\" into the learning process by augmenting the loss function with a symbolic-loss term that, in addition to the traditional loss, increases the probability of legal states (which also includes incorrect, yet legal, predictions). \n\nOverall the idea is interesting, but the paper does not seem ready for publication.  The  idea of semantic-loss function is appealing and is nicely motivated in the paper, however the practical aspects of how it can be applied are extremely vague and hard to understand.  Specifically, the authors define it over all assignments to the output variables that satisfy the constraints. For any non-trivial prediction problem, this would be at least computationally challenging. The author discuss it briefly mentioning a method by Darwiche-2003, but do not offer much intuition or analysis beyond that.  Their experiments focus on multiclass classification, which implicitly has a \"one-vs.-all\" constraint, although it's not clear why defining a formal loss function is needed (instead of just taking the argmax of the multiclass net), and even beyond that - why would it result in such significant improvements (when there are a few annotated data points)?   \n\nThe more interesting case is where the loss needs to decompose over the parts of a structural decision, where symbolic knowledge can help  constrain the output space. This has been addressed in the literature (e.g., [1], [2]) it's not clear why the authors don't compare to these models, or even attempt any meaningful evaluation.\n\n\n[1] Zhiting Hu, Xuezhe Ma, Zhengzhong Liu, Eduard Hovy, and Eric Xing. Harnessing deep neural\nnetworks with logic rules. ACL, 2016.\n\n[2]Posterior Regularization for Structured Latent Variable Models.  Ganchev et-al 2010.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall the idea is interesting, but the paper does not seem ready for publication. ", "rating": "4: Ok but not good enough - rejection", "review": "This paper suggest a method for including symbolic knowledge into the learning process. The symbolic knowledge is given as logical constraints which characterize the space of legal solutions.   This knowledge is \"injected\" into the learning process by augmenting the loss function with a symbolic-loss term that, in addition to the traditional loss, increases the probability of legal states (which also includes incorrect, yet legal, predictions). \n\nOverall the idea is interesting, but the paper does not seem ready for publication.  The  idea of semantic-loss function is appealing and is nicely motivated in the paper, however the practical aspects of how it can be applied are extremely vague and hard to understand.  Specifically, the authors define it over all assignments to the output variables that satisfy the constraints. For any non-trivial prediction problem, this would be at least computationally challenging. The author discuss it briefly mentioning a method by Darwiche-2003, but do not offer much intuition or analysis beyond that.  Their experiments focus on multiclass classification, which implicitly has a \"one-vs.-all\" constraint, although it's not clear why defining a formal loss function is needed (instead of just taking the argmax of the multiclass net), and even beyond that - why would it result in such significant improvements (when there are a few annotated data points)?   \n\nThe more interesting case is where the loss needs to decompose over the parts of a structural decision, where symbolic knowledge can help  constrain the output space. This has been addressed in the literature (e.g., [1], [2]) it's not clear why the authors don't compare to these models, or even attempt any meaningful evaluation.\n\n\n[1] Zhiting Hu, Xuezhe Ma, Zhengzhong Liu, Eduard Hovy, and Eric Xing. Harnessing deep neural\nnetworks with logic rules. ACL, 2016.\n\n[2]Posterior Regularization for Structured Latent Variable Models.  Ganchev et-al 2010.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1512095319509}, {"id": "ByEQXA5lM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper891/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a new loss function that is directed to take into account Boolean constraints involving the variables of a classification problem. This is a nice idea, and certainly relevant. The authors clearly describe their problem, and overall the paper is well presented. The contributions are a loss function derived from a set of axioms, and experiments indicating that this loss function captures some valuable elements of the input. This is a valid contribution, and the paper certainly has some significant strengths.\n\nConcerning the loss function, I find the whole derivation a bit distracting and unnecessary. Here we have some axioms, that are not simple when taken together, and that collectively imply a loss function that makes intuitive sense by itself. Well, why not just open the paper with Definition 1, and try to justify this definition on the basis of its properties. The discussion of axioms is just something that will create debate over questionable assumptions. Also it is frustrating to see some axioms in the main text, and some axioms in the appendix (why this division?). \n\nAfter presenting the loss function, the authors consider some applications. They are nicely presented; overall the gains are promising but not that great when compared to the state of the art --- they suggest that the proposed semantic loss makes sense. However I find that the proposal is still in search of a \"killer app\". Overall, I find that the whole proposal seems a bit premature and in need of more work on applications (the work on axiomatics is fine as long as it has something to add).\n\nConcerning the text, a few questions/suggestions:\n- Before Lemma 3, \"this allows...\" is the \"this including the other axioms in the appendix?\n- In Section 4, line 3: I suppose that the constraint is just creating a problem with a class containing several labels, not really a multi-label classification problem (?).\n- The beginning of Section 4.1 is not very clear. By reading it, I feel that the best way to handle the unlabeled data would be to add a direct penalty term forcing the unlabeled points to receive a label. Is this fair?\n- Page 6: \"a mor methodological\"... should it be \"a more methodical\"?\n- There are problems with capitalization in the references. Also some references miss page numbers and some do not even indicate what they are (journal papers, conference papers, arxiv, etc).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposes a new loss function that penalizes semantic features of the data, and shows some experiments; overall the writing is good and the ideas are nice, even though the contribution is relatively small.", "rating": "5: Marginally below acceptance threshold", "review": "The authors propose a new loss function that is directed to take into account Boolean constraints involving the variables of a classification problem. This is a nice idea, and certainly relevant. The authors clearly describe their problem, and overall the paper is well presented. The contributions are a loss function derived from a set of axioms, and experiments indicating that this loss function captures some valuable elements of the input. This is a valid contribution, and the paper certainly has some significant strengths.\n\nConcerning the loss function, I find the whole derivation a bit distracting and unnecessary. Here we have some axioms, that are not simple when taken together, and that collectively imply a loss function that makes intuitive sense by itself. Well, why not just open the paper with Definition 1, and try to justify this definition on the basis of its properties. The discussion of axioms is just something that will create debate over questionable assumptions. Also it is frustrating to see some axioms in the main text, and some axioms in the appendix (why this division?). \n\nAfter presenting the loss function, the authors consider some applications. They are nicely presented; overall the gains are promising but not that great when compared to the state of the art --- they suggest that the proposed semantic loss makes sense. However I find that the proposal is still in search of a \"killer app\". Overall, I find that the whole proposal seems a bit premature and in need of more work on applications (the work on axiomatics is fine as long as it has something to add).\n\nConcerning the text, a few questions/suggestions:\n- Before Lemma 3, \"this allows...\" is the \"this including the other axioms in the appendix?\n- In Section 4, line 3: I suppose that the constraint is just creating a problem with a class containing several labels, not really a multi-label classification problem (?).\n- The beginning of Section 4.1 is not very clear. By reading it, I feel that the best way to handle the unlabeled data would be to add a direct penalty term forcing the unlabeled points to receive a label. Is this fair?\n- Page 6: \"a mor methodological\"... should it be \"a more methodical\"?\n- There are problems with capitalization in the references. Also some references miss page numbers and some do not even indicate what they are (journal papers, conference papers, arxiv, etc).\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511871261380}], "openreview_url": "https://openreview.net/forum?id=HkepKG-Rb", "arxiv_id": "1711.11157", "paper_pdf": "papers/HkepKG-Rb.pdf", "paper_pdf_sha256": "aa75cbe8d14bacab4747aae480f95c1794c40bb291d07b52c5d21a7c264996d3", "paper_pdf_bytes": 345574, "paper_pdf_source": "openreview", "code_url": "https://github.com/UCLA-StarAI/Semantic-Loss", "code_repository": "UCLA-StarAI/Semantic-Loss", "code_commit": "86764c2294b89fc36fac8979e46bb0a6e062f1ec", "code_archive": "repos/HkepKG-Rb.zip", "code_archive_sha256": "77d89c6a94c73a90bb0ab2ba00f4fb4aeaf396929234f220e0fe22558c95f72b", "code_archive_bytes": 82493, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 86, "github_languages": {"Python": 48636}, "github_archived": false, "github_pushed_at": "2019-05-01T22:29:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-semantic-loss-function-for-deep-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nEcs0Duf86", "year": 2026, "status": "rejected", "title": "From Invariant Representations to Invariant Data: Provable Robustness to Spurious Correlations via Noisy Counterfactual Matching", "authors": ["Ruqi Bai", "Yao Ji", "Zeyu Zhou", "David I. Inouye"], "authorids": ["~Ruqi_Bai1", "~Yao_Ji1", "~Zeyu_Zhou1", "~David_I._Inouye1"], "authors_source": "OpenReview API", "abstract": "Models that learn spurious correlations from training data often fail when deployed in new environments. While many methods aim to learn invariant representations to address this, they often underperform standard empirical risk minimization (ERM). We propose a data-centric alternative that shifts the focus from learning invariant representations to leveraging invariant data pairs---pairs of samples that should have the same prediction. We prove that certain counterfactuals naturally satisfy this invariance property. Based on this, we introduce Noisy Counterfactual Matching (NCM), a simple constraint-based method that improves robustness by leveraging even a small number of \\emph{noisy} counterfactual pairs---improving upon prior works that do not explicitly consider noise. For linear causal models, we prove that NCM's test-domain error is bounded by its in-domain error plus a term dependent on the counterfactuals' quality and diversity. Experiments on synthetic data validate our theory, and we demonstrate NCM's effectiveness on real-world datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "xuHdNruubd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21549/Reviewer_PwUU"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes a data‑centric approach to robustness under domain shift that leverages invariant data pairs (instead of learning invariant representation)—pairs of inputs that should receive the same prediction under a robust classifier. The authors formalize the  spurious counterfactuals in a causal perspective, and prove that spurious counterfactuals are invariant pairs. They introduce Noisy Counterfactual Matching (NCM): augment ERM with a linear constraint that projects predictions onto the orthogonal complement of the spurious subspace estimated via the truncated SVD of pairwise differences. For linear and logistic regression, the authors provide theoretic guarantees of NCM . Empirically, synthetic experiments validate the theory’s trade‑offs, and linear probing on CLIP features shows gains on ColoredMNIST, Waterbirds‑CF, and PACS.", "review_text": "The paper proposes a data‑centric approach to robustness under domain shift that leverages invariant data pairs (instead of learning invariant representation)—pairs of inputs that should receive the same prediction under a robust classifier. The authors formalize the  spurious counterfactuals in a causal perspective, and prove that spurious counterfactuals are invariant pairs. They introduce Noisy Counterfactual Matching (NCM): augment ERM with a linear constraint that projects predictions onto the orthogonal complement of the spurious subspace estimated via the truncated SVD of pairwise differences. For linear and logistic regression, the authors provide theoretic guarantees of NCM . Empirically, synthetic experiments validate the theory’s trade‑offs, and linear probing on CLIP features shows gains on ColoredMNIST, Waterbirds‑CF, and PACS.", "strengths": "- The proposed method, Noisy Counterfactual Matching (NCM), is well-motivated and intuitive—it estimates a “spurious” subspace from pairwise differences and removes it via projection.\n- NCM is sample efficient. This is often easier to obtain in practice.\n- The implementation of NCM is simple, and has potential to be applied to different applications.\n- The method addresses the biggest practical challenge—noisy counterfactual pairs—and remains robust through a simple truncated-SVD design with a provable error bound.\n- The theoretical analysis is solid in the linear setting, and synthetic results nicely match the theory.\n- Beyond the linear case, real-data experiments show consistent gains for CLIP linear probes, suggesting the approach has broader practical potential beyond its theoretical contribution.", "weaknesses": "- The projection removes only observed linear directions of domain shift, leaving potential nonlinear or unseen correlations unaddressed.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a data‑centric approach to robustness under domain shift that leverages invariant data pairs (instead of learning invariant representation)—pairs of inputs that should receive the same prediction under a robust classifier. The authors formalize the  spurious counterfactuals in a causal perspective, and prove that spurious counterfactuals are invariant pairs. They introduce Noisy Counterfactual Matching (NCM): augment ERM with a linear constraint that projects predictions onto the orthogonal complement of the spurious subspace estimated via the truncated SVD of pairwise differences. For linear and logistic regression, the authors provide theoretic guarantees of NCM . Empirically, synthetic experiments validate the theory’s trade‑offs, and linear probing on CLIP features shows gains on ColoredMNIST, Waterbirds‑CF, and PACS.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The proposed method, Noisy Counterfactual Matching (NCM), is well-motivated and intuitive—it estimates a “spurious” subspace from pairwise differences and removes it via projection.\n- NCM is sample efficient. This is often easier to obtain in practice.\n- The implementation of NCM is simple, and has potential to be applied to different applications.\n- The method addresses the biggest practical challenge—noisy counterfactual pairs—and remains robust through a simple truncated-SVD design with a provable error bound.\n- The theoretical analysis is solid in the linear setting, and synthetic results nicely match the theory.\n- Beyond the linear case, real-data experiments show consistent gains for CLIP linear probes, suggesting the approach has broader practical potential beyond its theoretical contribution.", "weaknesses": "- The projection removes only observed linear directions of domain shift, leaving potential nonlinear or unseen correlations unaddressed.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761998253985}, {"id": "h1nc39MuMP", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21549/Reviewer_zQ3E"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a data-centric approach to domain generalization called Noisy Counterfactual Matching (NCM). Instead of enforcing representation-level invariance, the authors suggest learning invariance directly from invariant data pairs. The method approximates counterfactual sample pairs and constrains the classifier so that its predictions remain consistent across them. Theoretical analysis shows that when these counterfactual pairs approximate the true invariant directions, the classifier achieves provable robustness to spurious correlations. Experiments on synthetic data show modest improvements over baselines.", "review_text": "This paper proposes a data-centric approach to domain generalization called Noisy Counterfactual Matching (NCM). Instead of enforcing representation-level invariance, the authors suggest learning invariance directly from invariant data pairs. The method approximates counterfactual sample pairs and constrains the classifier so that its predictions remain consistent across them. Theoretical analysis shows that when these counterfactual pairs approximate the true invariant directions, the classifier achieves provable robustness to spurious correlations. Experiments on synthetic data show modest improvements over baselines.", "strengths": "1. Addressing spurious correlations and domain generalization remains an important and open challenge.\n2. The paper is well-written and conceptually easy to follow, with a consistent flow from motivation to theory to experiments.\n3. Experimental results show modest but consistent improvements on benchmark datasets.", "weaknesses": "1. The core idea of enforcing invariance through matching or orthogonality constraints is not entirely new. Previous works such as MatchDG, IRMv1 have already explored related ideas. NCM’s main distinction, using noisy counterfactual pairs, is interesting but incremental and mostly limited to linear settings.\n2. The theoretical guarantees hold only for linear models with idealized “counterfactual pairs.” In realistic nonlinear cases (e.g., deep networks), the method lacks justification. The authors claim empirical robustness, but this cannot be directly attributed to the theory presented.\n3. The approach assumes access to approximate counterfactual pairs or noisy invariant pairs, which may not be available in practice. The paper does not discuss how these pairs could be obtained in realistic settings or how sensitive the method is to poor-quality matches.\n4. While the results show slight gains, they are relatively small, raising questions about whether the added complexity is justified compared to simpler ERM or representation-based baselines.\n5. The extension of the orthogonality constraint via SVD projection to deep models is mentioned but not clearly demonstrated or validated experimentally.", "questions": "See Weaknesses Part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a data-centric approach to domain generalization called Noisy Counterfactual Matching (NCM). Instead of enforcing representation-level invariance, the authors suggest learning invariance directly from invariant data pairs. The method approximates counterfactual sample pairs and constrains the classifier so that its predictions remain consistent across them. Theoretical analysis shows that when these counterfactual pairs approximate the true invariant directions, the classifier achieves provable robustness to spurious correlations. Experiments on synthetic data show modest improvements over baselines.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Addressing spurious correlations and domain generalization remains an important and open challenge.\n2. The paper is well-written and conceptually easy to follow, with a consistent flow from motivation to theory to experiments.\n3. Experimental results show modest but consistent improvements on benchmark datasets.", "weaknesses": "1. The core idea of enforcing invariance through matching or orthogonality constraints is not entirely new. Previous works such as MatchDG, IRMv1 have already explored related ideas. NCM’s main distinction, using noisy counterfactual pairs, is interesting but incremental and mostly limited to linear settings.\n2. The theoretical guarantees hold only for linear models with idealized “counterfactual pairs.” In realistic nonlinear cases (e.g., deep networks), the method lacks justification. The authors claim empirical robustness, but this cannot be directly attributed to the theory presented.\n3. The approach assumes access to approximate counterfactual pairs or noisy invariant pairs, which may not be available in practice. The paper does not discuss how these pairs could be obtained in realistic settings or how sensitive the method is to poor-quality matches.\n4. While the results show slight gains, they are relatively small, raising questions about whether the added complexity is justified compared to simpler ERM or representation-based baselines.\n5. The extension of the orthogonality constraint via SVD projection to deep models is mentioned but not clearly demonstrated or validated experimentally.", "questions": "See Weaknesses Part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761480573550}, {"id": "F2w8IsxBbN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21549/Reviewer_vnFv"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "The paper presents Noisy Counterfactual Matching (NCM), a method for domain generalization. NCM uses pairs of samples that differ only in spurious aspects (features) to enforce consistent predictions and filters out spurious directions with truncated SVD. The authors provide theoretical support showing that few diverse pairs are sufficient for invariance and report experiments on ColoredMNIST, Waterbirds-CF, and PACS where NCM matches or slightly outperforms established baselines such as IRM, GroupDRO, SWAD, and LISA.", "review_text": "The paper presents Noisy Counterfactual Matching (NCM), a method for domain generalization. NCM uses pairs of samples that differ only in spurious aspects (features) to enforce consistent predictions and filters out spurious directions with truncated SVD. The authors provide theoretical support showing that few diverse pairs are sufficient for invariance and report experiments on ColoredMNIST, Waterbirds-CF, and PACS where NCM matches or slightly outperforms established baselines such as IRM, GroupDRO, SWAD, and LISA.", "strengths": "The paper presents a an approach to domain generalization based on counterfactual matching. Its originality lies in reformulating invariant representation learning without relying on explicit domain labels, supported by a well-developed theoretical framework. The analysis is technically sound, with proofs that connect counterfactual consistency to invariance and empirical results that align with the theory. The experimental setup is appropriate, using standard benchmarks.", "weaknesses": "While the paper is well-executed, its novelty is somewhat limited relative to prior work on counterfactual and invariant learning, such as MatchDG (ICLR 2021) . The distinction between NCM and these methods, beyond the use of formal proofs, could be explained more clearly. The empirical evaluation, though thorough on benchmark datasets, depends on synthetic or curated counterfactuals (e.g., Waterbirds-CF), which limits evidence of real-world applicability. Demonstrating how counterfactuals could be obtained or approximated in realistic settings would strengthen the paper. The study would also benefit from testing on additional or more recent benchmarks.", "questions": "1. The experiments primarily rely on synthetic or semi-synthetic datasets where counterfactuals are artificially constructed. How would NCM perform on real-world datasets where counterfactual pairs are unavailable or difficult to define or noisy? \n2. Would the authors consider including results for baselines on the synthetic dataset to confirm that NCM’s advantages are not specific to its own simulation setup?\n3. Please include 2023, 2024, and 2025 methods in Related Work and, where feasible, add a conceptual and empirical comparison.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents Noisy Counterfactual Matching (NCM), a method for domain generalization. NCM uses pairs of samples that differ only in spurious aspects (features) to enforce consistent predictions and filters out spurious directions with truncated SVD. The authors provide theoretical support showing that few diverse pairs are sufficient for invariance and report experiments on ColoredMNIST, Waterbirds-CF, and PACS where NCM matches or slightly outperforms established baselines such as IRM, GroupDRO, SWAD, and LISA.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper presents a an approach to domain generalization based on counterfactual matching. Its originality lies in reformulating invariant representation learning without relying on explicit domain labels, supported by a well-developed theoretical framework. The analysis is technically sound, with proofs that connect counterfactual consistency to invariance and empirical results that align with the theory. The experimental setup is appropriate, using standard benchmarks.", "weaknesses": "While the paper is well-executed, its novelty is somewhat limited relative to prior work on counterfactual and invariant learning, such as MatchDG (ICLR 2021) . The distinction between NCM and these methods, beyond the use of formal proofs, could be explained more clearly. The empirical evaluation, though thorough on benchmark datasets, depends on synthetic or curated counterfactuals (e.g., Waterbirds-CF), which limits evidence of real-world applicability. Demonstrating how counterfactuals could be obtained or approximated in realistic settings would strengthen the paper. The study would also benefit from testing on additional or more recent benchmarks.", "questions": "1. The experiments primarily rely on synthetic or semi-synthetic datasets where counterfactuals are artificially constructed. How would NCM perform on real-world datasets where counterfactual pairs are unavailable or difficult to define or noisy? \n2. Would the authors consider including results for baselines on the synthetic dataset to confirm that NCM’s advantages are not specific to its own simulation setup?\n3. Please include 2023, 2024, and 2025 methods in Related Work and, where feasible, add a conceptual and empirical comparison.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761332528885}, {"id": "8qLeh9RPlA", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21549/Reviewer_FWNp"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper tackles domain generalization problems with a shift in focus from learning invariant representations to leveraging invariant data, where an invariant pair is a pair of samples that should share the same prediction under the optimal hypothesis. The authors first show that counterfactual pairs are invariant and analyzes when such pairs improve robustness and how many are needed. The paper then introduces Noisy Counterfactual Matching, which adds a constraint to ERM to improve robustness to spurious correlations using a small set of (noisy) invariant pairs. The scope of distribution shift considered is restricted to domains that only intervene on spurious latent variables that are non‑ancestors of the target variable.", "review_text": "The paper tackles domain generalization problems with a shift in focus from learning invariant representations to leveraging invariant data, where an invariant pair is a pair of samples that should share the same prediction under the optimal hypothesis. The authors first show that counterfactual pairs are invariant and analyzes when such pairs improve robustness and how many are needed. The paper then introduces Noisy Counterfactual Matching, which adds a constraint to ERM to improve robustness to spurious correlations using a small set of (noisy) invariant pairs. The scope of distribution shift considered is restricted to domains that only intervene on spurious latent variables that are non‑ancestors of the target variable.", "strengths": "* The paper is clearly written and easy to follow.\n* The problem is well motivated, with an interesting insight and thorough theoretical analyses for the setting considered.\n* Empirical results are presented to support the theoretical claims.", "weaknesses": "* Theoretical results are only applicable to linear SCMs, which is rather simple and restrictive, relative to real‑world scenarios. \n\n* The collection of invariant data pairs is only feasible given some knowledge about the spurious features. While the authors argue that such knowledge may be available in practice, I think that the knowledge of some may be feasible, but knowledge of all possible spurious features is hard to achieve, even with domain knowledge. For data like images, spurious features may simply be related to the background or style of the images. However for tabular data in arbitrary more complex domains, humans are often deprived of the full knowledge of all possible spurious factors. \n\n* There are other technical concerns that requires further clarification. See the questions below.", "questions": "1. The logic of the proposal seems problematic to me. The paper proves that counterfactual pairs improve robustness and then uses the fact that counterfactual pairs are invariant to justify collecting invariant pairs instead. Are counterfactual and invariant pairs equivalent? From my understanding of its proof, Proposition 1 shows that counterfactual pairs are invariant, but not all invariant pairs are counterfactual. If the model gives the same prediction to different inputs, collecting such pairs may not yield the desired robustness effect. Could the authors clarify this? \n\n2. The examples about the collection of invariant pairs, in the Introduction and Appendix A.2, seem to rely on knowing that pairs share the same label, whereas invariance is defined with respect to the optimal hypothesis. True labels and optimal predictions can differ. How does this mismatch affect your results? If invariant pairs can be collected using training labels, why not define invariance directly with respect to the true label and perform the alignment/subspace objective accordingly, rather than relying on the robust model’s prediction?\n\n3. What is the output space of the robust classifier $h$? Does it output a hard label or a label distribution?\n\n4. Definition 1 is stated for classifiers, but Theorem 1 discusses linear and logistic regression losses. Is Theorem 1 applicable to common losses like cross‑entropy for classification tasks?\n\n5. The authors also claim that one feasible way to collect invariant pairs is to perform augmentation. So at a high level, I do not understand how the proposed approach differs from data augmentation or in other words directly using augmented data for alignment.  \n\n6. A major limitation is that the constraint in Eq. (1) reduces to Eq. (2) only under linear models. Once we have the counterfactual data pairs, why don’t we directly align them using some distance metrics, as often done in machine learning literature e.g., contrastive learning? \n\n7. I understand that theoretical analyses in non-linear settings are challenging. However, could the authors at least provide empirical results on standard benchmark like DomainBed to at least verify whether the idea carries over empirically?\n\nMy current rating reflects the above concerns. I will consider updating the score once they are resolved.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tackles domain generalization problems with a shift in focus from learning invariant representations to leveraging invariant data, where an invariant pair is a pair of samples that should share the same prediction under the optimal hypothesis. The authors first show that counterfactual pairs are invariant and analyzes when such pairs improve robustness and how many are needed. The paper then introduces Noisy Counterfactual Matching, which adds a constraint to ERM to improve robustness to spurious correlations using a small set of (noisy) invariant pairs. The scope of distribution shift considered is restricted to domains that only intervene on spurious latent variables that are non‑ancestors of the target variable.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* The paper is clearly written and easy to follow.\n* The problem is well motivated, with an interesting insight and thorough theoretical analyses for the setting considered.\n* Empirical results are presented to support the theoretical claims.", "weaknesses": "* Theoretical results are only applicable to linear SCMs, which is rather simple and restrictive, relative to real‑world scenarios. \n\n* The collection of invariant data pairs is only feasible given some knowledge about the spurious features. While the authors argue that such knowledge may be available in practice, I think that the knowledge of some may be feasible, but knowledge of all possible spurious features is hard to achieve, even with domain knowledge. For data like images, spurious features may simply be related to the background or style of the images. However for tabular data in arbitrary more complex domains, humans are often deprived of the full knowledge of all possible spurious factors. \n\n* There are other technical concerns that requires further clarification. See the questions below.", "questions": "1. The logic of the proposal seems problematic to me. The paper proves that counterfactual pairs improve robustness and then uses the fact that counterfactual pairs are invariant to justify collecting invariant pairs instead. Are counterfactual and invariant pairs equivalent? From my understanding of its proof, Proposition 1 shows that counterfactual pairs are invariant, but not all invariant pairs are counterfactual. If the model gives the same prediction to different inputs, collecting such pairs may not yield the desired robustness effect. Could the authors clarify this? \n\n2. The examples about the collection of invariant pairs, in the Introduction and Appendix A.2, seem to rely on knowing that pairs share the same label, whereas invariance is defined with respect to the optimal hypothesis. True labels and optimal predictions can differ. How does this mismatch affect your results? If invariant pairs can be collected using training labels, why not define invariance directly with respect to the true label and perform the alignment/subspace objective accordingly, rather than relying on the robust model’s prediction?\n\n3. What is the output space of the robust classifier $h$? Does it output a hard label or a label distribution?\n\n4. Definition 1 is stated for classifiers, but Theorem 1 discusses linear and logistic regression losses. Is Theorem 1 applicable to common losses like cross‑entropy for classification tasks?\n\n5. The authors also claim that one feasible way to collect invariant pairs is to perform augmentation. So at a high level, I do not understand how the proposed approach differs from data augmentation or in other words directly using augmented data for alignment.  \n\n6. A major limitation is that the constraint in Eq. (1) reduces to Eq. (2) only under linear models. Once we have the counterfactual data pairs, why don’t we directly align them using some distance metrics, as often done in machine learning literature e.g., contrastive learning? \n\n7. I understand that theoretical analyses in non-linear settings are challenging. However, could the authors at least provide empirical results on standard benchmark like DomainBed to at least verify whether the idea carries over empirically?\n\nMy current rating reflects the above concerns. I will consider updating the score once they are resolved.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761274217952}], "openreview_url": "https://openreview.net/forum?id=nEcs0Duf86", "arxiv_id": "2505.24843", "paper_pdf": "papers/nEcs0Duf86.pdf", "paper_pdf_sha256": "f6adf367f49b825ae7d97327031ae2108828d79ee1f34c1d47f11a0b78889f70", "paper_pdf_bytes": 2075695, "paper_pdf_source": "openreview", "code_url": "https://github.com/inouye-lab/GRIT", "code_repository": "inouye-lab/GRIT", "code_commit": "da418a4ab8fc91403267ff44fc7369a51fed6d41", "code_archive": "repos/nEcs0Duf86.zip", "code_archive_sha256": "7a82ff7d55555ad9fbd93c7e55e6e53182239f94f1c42d7a249e018f875be632", "code_archive_bytes": 134396, "code_file_count": 92, "code_extensions": {".py": 92}, "github_disk_usage_kb": 86, "github_languages": {"Python": 174170}, "github_archived": false, "github_pushed_at": "2026-04-13T13:59:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/from-invariant-representations-to-invariant"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lK0WxHeups", "year": 2024, "status": "rejected", "title": "Iteration and Stochastic First-order Oracle Complexities of Stochastic Gradient Descent using Constant and Decaying Learning Rates", "authors": ["Kento Imaizumi", "Hideaki Iiduka"], "authorids": ["~Kento_Imaizumi1", "~Hideaki_Iiduka1"], "authors_source": "OpenReview API", "abstract": "The performance of stochastic gradient descent (SGD), which is the simplest first-order optimizer for training deep neural networks, depends on not only the learning rate but also the batch size. They both affect the number of iterations and the stochastic first-order oracle (SFO) complexity needed for training. In particular, the previous numerical results indicated that, for SGD using a constant learning rate, the number of iterations needed for training decreases when the batch size increases, and the SFO complexity needed for training is minimized at a critical batch size and increases once the batch size exceeds that size. This paper studies the relationship between batch size and the iteration and the SFO complexities needed for nonconvex optimization in deep learning with SGD using constant/decay learning rates. We show that SGD using a step-decay learning rate and a small batch size reduces the SFO complexity to find a local minimizer of a loss function. We also provide numerical comparisons of SGD with the existing first-order optimizers and show the usefulness of SGD using a step-decay learning rate and a small batch size.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "N1e5CJKvuh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5017/Reviewer_xRbi"], "rating": "1: strong reject", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This manuscript studied the effects of batch size and learning rate for nonconvex smooth optimization. The authors established iteration complexity and SFO (Stochastic First Order Oracle) complexity of the problem.\n\nDespite the result is interesting, most of the results are known (or straightforward extension) in the literature.", "review_text": "This manuscript studied the effects of batch size and learning rate for nonconvex smooth optimization. The authors established iteration complexity and SFO (Stochastic First Order Oracle) complexity of the problem.\n\nDespite the result is interesting, most of the results are known (or straightforward extension) in the literature.", "strengths": "The paper is well-written.", "weaknesses": "The paper's theoretical results are known in the literature (e.g., [r1], [r2]). The hardness result [r3] says that: whatever batch size the SGD algorithm can choose, the SFO cannot be better than $O(1/\\epsilon^4)$.\n\n[r1] Ghadimi, Saeed, and Guanghui Lan. \"Stochastic first-and zeroth-order methods for nonconvex stochastic programming.\" SIAM Journal on Optimization 23, no. 4 (2013): 2341-2368.\n\n[r2] Ghadimi, S., Lan, G., & Zhang, H. (2016). Mini-batch stochastic approximation methods for nonconvex stochastic composite optimization. Mathematical Programming, 155(1-2), 267-305.\n\n[r3] Arjevani, Yossi, Yair Carmon, John C. Duchi, Dylan J. Foster, Nathan Srebro, and Blake Woodworth. \"Lower bounds for non-convex stochastic optimization.\" Mathematical Programming 199, no. 1-2 (2023): 165-214.", "questions": "Can you describe how your approach is better than the references I gave above (e.g., [r1, r2])?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript studied the effects of batch size and learning rate for nonconvex smooth optimization. The authors established iteration complexity and SFO (Stochastic First Order Oracle) complexity of the problem.\n\nDespite the result is interesting, most of the results are known (or straightforward extension) in the literature.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "The paper is well-written.", "weaknesses": "The paper's theoretical results are known in the literature (e.g., [r1], [r2]). The hardness result [r3] says that: whatever batch size the SGD algorithm can choose, the SFO cannot be better than $O(1/\\epsilon^4)$.\n\n[r1] Ghadimi, Saeed, and Guanghui Lan. \"Stochastic first-and zeroth-order methods for nonconvex stochastic programming.\" SIAM Journal on Optimization 23, no. 4 (2013): 2341-2368.\n\n[r2] Ghadimi, S., Lan, G., & Zhang, H. (2016). Mini-batch stochastic approximation methods for nonconvex stochastic composite optimization. Mathematical Programming, 155(1-2), 267-305.\n\n[r3] Arjevani, Yossi, Yair Carmon, John C. Duchi, Dylan J. Foster, Nathan Srebro, and Blake Woodworth. \"Lower bounds for non-convex stochastic optimization.\" Mathematical Programming 199, no. 1-2 (2023): 165-214.", "questions": "Can you describe how your approach is better than the references I gave above (e.g., [r1, r2])?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699491604170}, {"id": "iHHF1niibD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5017/Reviewer_BdrQ"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the complexity of computing stationary points with SGD \nusing a variety of step-size schedules. The authors derive convergence rates \nwith an explicit dependence on the batch-size for SGD with a constant step-size\nas well as polynomial decay and step-decay step-size schedules. These rates \nare then converted into iteration and oracle complexities and studied as a \nfunction of the mini-batch size. \nThe authors prove that the parameterized complexities are convex functions\nand use this to derive optimal batch-sizes for different schedules.\nThe submission concludes with experiments comparing different schedules on \nCIFAR-10 and CIFAR-100.", "review_text": "This paper studies the complexity of computing stationary points with SGD \nusing a variety of step-size schedules. The authors derive convergence rates \nwith an explicit dependence on the batch-size for SGD with a constant step-size\nas well as polynomial decay and step-decay step-size schedules. These rates \nare then converted into iteration and oracle complexities and studied as a \nfunction of the mini-batch size. \nThe authors prove that the parameterized complexities are convex functions\nand use this to derive optimal batch-sizes for different schedules.\nThe submission concludes with experiments comparing different schedules on \nCIFAR-10 and CIFAR-100.", "strengths": "The main strength of this paper is its novel approach to hyperparameter\ntuning for SGD. While it is typical to tune (at least in theory) the step-size \nparameter to minimize the oracle complexity, maintaining an explicit dependence\non the mini-batch size in the convergence rate and using this to understand\nthe trade-offs between iteration and oracle complexity is an interesting idea.\nIn addition to this, the paper has the following strengths:\n\n- The authors provide a simple and clear analysis for SGD which covers\n    SGD with a fixed step-size, polynomial decay, and step-decay\n    schedules. \n\n- Although the optimal batch-sizes for fixed step-sizes and polynomial decay schedules\n    depend on unknown parameters of the problem, understanding the optimal values\n    may allow for new heuristics for selecting the batch-size in practice.\n\n- The experiments, although simple, generally reflect the theory and show that\n    tuning the batch-size can lead to improvements in optimization given a fixed\n    budget of gradient evaluations.", "weaknesses": "This paper has several significant weaknesses that should be addressed before\npublication. In particular,\n\n- The convergence rate given for SGD with the step-decay schedule is misleading\n    and leads to incorrect iteration and oracle complexities for this method.\n\n- The paper does not address the fact that $b \\leq n$ must be maintained, where\n    $n$ is the number of functions in the finite sum. As a result, the optimal\n    batch-sizes which the authors derive may not be attainable depending on the \n    desired precision of the solution. For example, as $\\epsilon \\rightarrow 0$,\n    $b \\rightarrow \\infty$ for a fixed step-size.\n\n- The manuscript is unnecessarily \"mathy\" and many equations could be \n    omitted while maintaining the same results. See, for example, Equations 1 and 2.\n    While this makes the writing seem superficially impressive, it is difficult to\n    read and detracts from the flow of the paper. \n\n- None of the experiments serve to verify the theoretical derivations.\n    I would liked to see at least one synthetic experiment for which the problem\n    constants are known (e.g. a simple quadratic) and $b^*$ can be computed.\n    Plots similar to that in Figures 1/2 could then show that $b^*$ does, in fact,\n    obtain the optimal oracle complexity as claimed.\n\nGiven these issues, I cannot recommend accepting the submission at this time.\nHowever, I am willing to increase my score if they address these issues. At\na very minimum, I feel the problem with the complexity of the step-size schedule\nmust be resolved. See \"Questions\" below for more details.", "questions": "- \"SGD using a decaying learning rate...\": some additional comment on the type\n    of learning rate decay is needed here. SGD with step-size $\\alpha_k = 1 / \\log(k)$ \n    does not converge as $O(1/\\sqrt{K})$ despite $\\alpha_k \\rightarrow 0$.\n\n- First math display in Section 1.3.2: this bound should mentioned somewhere that \n    $b > n$ isn't possible and $b = n$ reduces to full-batch gradient descent.\n    As a result, it is not always feasible to select the batch-size to minimize\n    the oracle complexity. \n\n- \"Accordingly, small batch sizes are appropriate for a decaying learning rate or a \n    step-decay learning rate\": why is this true? You have said that SFO complexity\n    has no positive stationary point, but that doesn't imply it is increasing\n    in $b$ or that a small batch-size minimize the complexity over the positive\n    integers. Can you please address this fact?\n\n- Equation (2) and Table 2: It is somewhat confusing to switch from measuring \n    convergence using the squared gradient norm in Table 1 to convergence of \n    just the gradient norm in Table 2 and Equation (2).\n\n- Theorem 3.1: I am concerned by the presentation of the convergence rates for\n    SGD with step-decay. Firstly, $\\underline{\\alpha}$ does not appear to be\n    defined anywhere. From the proof in the appendix, it seems\n    $\\underline{\\alpha} = \\alpha_{K-1} = \\alpha \\eta^{K/T-1}$.  This quantity\n    depends on $K$ --- it is exponentially decreasing every $K/T$ iterations\n    --- so that it is incorrect to write it as a constant factor.  Similarly,\n    $D_3$ depends on $T$, which may or may not have a relationship with $K$\n    depending on algorithm parameters. \n\n    Only be carefully optimizing over $T$ can a final rate of convergence be\n    obtained. Wang et al. [1] set $T = K / \\log_{\\eta}(K)$ to obtain a final\n    convergence rate of $O(\\log(T)/\\sqrt{T})$.  In contrast, treating\n    $\\underline{\\alpha}$ as a constant leads to an deceptive presentation of\n    the convergence rate in Table 1. Moreover, I am fairly certain the\n    complexity of $O(1/\\epsilon^2)$ for computing an $\\epsilon$ stationary\n    point in Table 2 is incorrect and violates lower bounds due to Drori and\n    Shamir [2]. \n\n- Theorem 3.4: In addition to the issue with the complexity of step-decay raised\n     previously, this theorem assumes that $b \\leq n$ can be chosen arbitrarily\n     large in order to obtain the desired complexity. For example, SGD with a \n     constant step-size requires $b \\geq 2 C_2 \\epsilon$, which diverges to \n     infinity as $\\epsilon \\rightarrow 0$. But this is not sensible because\n     $n$ is assumed to be a fixed, finite number of training examples.\n     If this is not the cause, then the authors must specify somewhere that they\n     assume a setting where $n$ can be taken arbitrarily large. \n\n### References\n\n[1] Wang, Xiaoyu, Sindri Magnússon, and Mikael Johansson. \"On the convergence\nof step decay step-size for stochastic optimization.\" Advances in Neural\nInformation Processing Systems 34 (2021): 14226-14238.\n\n[2] Drori, Yoel, and Ohad Shamir. \"The complexity of finding stationary points\nwith stochastic gradient descent.\" International Conference on Machine\nLearning. PMLR, 2020.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the complexity of computing stationary points with SGD \nusing a variety of step-size schedules. The authors derive convergence rates \nwith an explicit dependence on the batch-size for SGD with a constant step-size\nas well as polynomial decay and step-decay step-size schedules. These rates \nare then converted into iteration and oracle complexities and studied as a \nfunction of the mini-batch size. \nThe authors prove that the parameterized complexities are convex functions\nand use this to derive optimal batch-sizes for different schedules.\nThe submission concludes with experiments comparing different schedules on \nCIFAR-10 and CIFAR-100.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "The main strength of this paper is its novel approach to hyperparameter\ntuning for SGD. While it is typical to tune (at least in theory) the step-size \nparameter to minimize the oracle complexity, maintaining an explicit dependence\non the mini-batch size in the convergence rate and using this to understand\nthe trade-offs between iteration and oracle complexity is an interesting idea.\nIn addition to this, the paper has the following strengths:\n\n- The authors provide a simple and clear analysis for SGD which covers\n    SGD with a fixed step-size, polynomial decay, and step-decay\n    schedules. \n\n- Although the optimal batch-sizes for fixed step-sizes and polynomial decay schedules\n    depend on unknown parameters of the problem, understanding the optimal values\n    may allow for new heuristics for selecting the batch-size in practice.\n\n- The experiments, although simple, generally reflect the theory and show that\n    tuning the batch-size can lead to improvements in optimization given a fixed\n    budget of gradient evaluations.", "weaknesses": "This paper has several significant weaknesses that should be addressed before\npublication. In particular,\n\n- The convergence rate given for SGD with the step-decay schedule is misleading\n    and leads to incorrect iteration and oracle complexities for this method.\n\n- The paper does not address the fact that $b \\leq n$ must be maintained, where\n    $n$ is the number of functions in the finite sum. As a result, the optimal\n    batch-sizes which the authors derive may not be attainable depending on the \n    desired precision of the solution. For example, as $\\epsilon \\rightarrow 0$,\n    $b \\rightarrow \\infty$ for a fixed step-size.\n\n- The manuscript is unnecessarily \"mathy\" and many equations could be \n    omitted while maintaining the same results. See, for example, Equations 1 and 2.\n    While this makes the writing seem superficially impressive, it is difficult to\n    read and detracts from the flow of the paper. \n\n- None of the experiments serve to verify the theoretical derivations.\n    I would liked to see at least one synthetic experiment for which the problem\n    constants are known (e.g. a simple quadratic) and $b^*$ can be computed.\n    Plots similar to that in Figures 1/2 could then show that $b^*$ does, in fact,\n    obtain the optimal oracle complexity as claimed.\n\nGiven these issues, I cannot recommend accepting the submission at this time.\nHowever, I am willing to increase my score if they address these issues. At\na very minimum, I feel the problem with the complexity of the step-size schedule\nmust be resolved. See \"Questions\" below for more details.", "questions": "- \"SGD using a decaying learning rate...\": some additional comment on the type\n    of learning rate decay is needed here. SGD with step-size $\\alpha_k = 1 / \\log(k)$ \n    does not converge as $O(1/\\sqrt{K})$ despite $\\alpha_k \\rightarrow 0$.\n\n- First math display in Section 1.3.2: this bound should mentioned somewhere that \n    $b > n$ isn't possible and $b = n$ reduces to full-batch gradient descent.\n    As a result, it is not always feasible to select the batch-size to minimize\n    the oracle complexity. \n\n- \"Accordingly, small batch sizes are appropriate for a decaying learning rate or a \n    step-decay learning rate\": why is this true? You have said that SFO complexity\n    has no positive stationary point, but that doesn't imply it is increasing\n    in $b$ or that a small batch-size minimize the complexity over the positive\n    integers. Can you please address this fact?\n\n- Equation (2) and Table 2: It is somewhat confusing to switch from measuring \n    convergence using the squared gradient norm in Table 1 to convergence of \n    just the gradient norm in Table 2 and Equation (2).\n\n- Theorem 3.1: I am concerned by the presentation of the convergence rates for\n    SGD with step-decay. Firstly, $\\underline{\\alpha}$ does not appear to be\n    defined anywhere. From the proof in the appendix, it seems\n    $\\underline{\\alpha} = \\alpha_{K-1} = \\alpha \\eta^{K/T-1}$.  This quantity\n    depends on $K$ --- it is exponentially decreasing every $K/T$ iterations\n    --- so that it is incorrect to write it as a constant factor.  Similarly,\n    $D_3$ depends on $T$, which may or may not have a relationship with $K$\n    depending on algorithm parameters. \n\n    Only be carefully optimizing over $T$ can a final rate of convergence be\n    obtained. Wang et al. [1] set $T = K / \\log_{\\eta}(K)$ to obtain a final\n    convergence rate of $O(\\log(T)/\\sqrt{T})$.  In contrast, treating\n    $\\underline{\\alpha}$ as a constant leads to an deceptive presentation of\n    the convergence rate in Table 1. Moreover, I am fairly certain the\n    complexity of $O(1/\\epsilon^2)$ for computing an $\\epsilon$ stationary\n    point in Table 2 is incorrect and violates lower bounds due to Drori and\n    Shamir [2]. \n\n- Theorem 3.4: In addition to the issue with the complexity of step-decay raised\n     previously, this theorem assumes that $b \\leq n$ can be chosen arbitrarily\n     large in order to obtain the desired complexity. For example, SGD with a \n     constant step-size requires $b \\geq 2 C_2 \\epsilon$, which diverges to \n     infinity as $\\epsilon \\rightarrow 0$. But this is not sensible because\n     $n$ is assumed to be a fixed, finite number of training examples.\n     If this is not the cause, then the authors must specify somewhere that they\n     assume a setting where $n$ can be taken arbitrarily large. \n\n### References\n\n[1] Wang, Xiaoyu, Sindri Magnússon, and Mikael Johansson. \"On the convergence\nof step decay step-size for stochastic optimization.\" Advances in Neural\nInformation Processing Systems 34 (2021): 14226-14238.\n\n[2] Drori, Yoel, and Ohad Shamir. \"The complexity of finding stationary points\nwith stochastic gradient descent.\" International Conference on Machine\nLearning. PMLR, 2020.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698811644822}, {"id": "MSYiw4jGCi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5017/Reviewer_1sx5"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the stochastic first-order oracle complexity (SFO, defined by this paper) of SGD with diminishing and constant learning rates. It shows that SGD using a step-decay learning rate and a small batch size achieves the best performance in terms of SFO complexity. Numerical experiments are provided.", "review_text": "This paper studies the stochastic first-order oracle complexity (SFO, defined by this paper) of SGD with diminishing and constant learning rates. It shows that SGD using a step-decay learning rate and a small batch size achieves the best performance in terms of SFO complexity. Numerical experiments are provided.", "strengths": "1. The paper is well-written and I enjoy reading it.\n\n2. Diminishing learning rate is commonly adopted in deep learning, and it is thus important to study it.", "weaknesses": "1. The stochastic gradient is generated in a different way from practice, and I think this should be highlighted. Specifically, in each iteration, this paper assumes the stochastic gradient is chosen as an ensemble as individual gradients sampled with replacement from some distribution. However, in deep learning practice, the individual gradient is sampled without-replacement and this leads to a gap between practice and the presented theory.\n\n2. I do not think SFO is a reasonable measure. In practice, different individual gradients are calculated parallelly and the corresponding time does not accumulate across samples.\n\n3. The definition of K_{\\epsilon} and N_{epsilon} seems to be weird, since the learning rate does not appear in any side of the equation.\n\n4. I wonder what is the novelty of Theorem 3.1. Is not it a very basic analysis of SGD?\n\n5. I find the result of Decay 4 problematic. Specifically, in Theorem 3.2, isn't $\\underline{\\alpha}$ itself depends on $K(b)$? How can $K(b)$ be further calculated by $\\underline{\\alpha}$? That being said, when $T$ is independent of $\\epsilon$, T is in the same order as $\\varepsilon$ as P. Therefore, $\\underline{\\alpha}$ depends exponentially on $K$. Applying this to Theorem 3.2, it indicates $K(b)$ is also exponentially dependent over $\\varepsilon$ and contradicts Theorem 3.4.", "questions": "1. On page 3, Is $N_{\\epsilon}$ just $bK_{\\epsilon}$?If yes, why not use the simpler one?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the stochastic first-order oracle complexity (SFO, defined by this paper) of SGD with diminishing and constant learning rates. It shows that SGD using a step-decay learning rate and a small batch size achieves the best performance in terms of SFO complexity. Numerical experiments are provided.", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "strengths": "1. The paper is well-written and I enjoy reading it.\n\n2. Diminishing learning rate is commonly adopted in deep learning, and it is thus important to study it.", "weaknesses": "1. The stochastic gradient is generated in a different way from practice, and I think this should be highlighted. Specifically, in each iteration, this paper assumes the stochastic gradient is chosen as an ensemble as individual gradients sampled with replacement from some distribution. However, in deep learning practice, the individual gradient is sampled without-replacement and this leads to a gap between practice and the presented theory.\n\n2. I do not think SFO is a reasonable measure. In practice, different individual gradients are calculated parallelly and the corresponding time does not accumulate across samples.\n\n3. The definition of K_{\\epsilon} and N_{epsilon} seems to be weird, since the learning rate does not appear in any side of the equation.\n\n4. I wonder what is the novelty of Theorem 3.1. Is not it a very basic analysis of SGD?\n\n5. I find the result of Decay 4 problematic. Specifically, in Theorem 3.2, isn't $\\underline{\\alpha}$ itself depends on $K(b)$? How can $K(b)$ be further calculated by $\\underline{\\alpha}$? That being said, when $T$ is independent of $\\epsilon$, T is in the same order as $\\varepsilon$ as P. Therefore, $\\underline{\\alpha}$ depends exponentially on $K$. Applying this to Theorem 3.2, it indicates $K(b)$ is also exponentially dependent over $\\varepsilon$ and contradicts Theorem 3.4.", "questions": "1. On page 3, Is $N_{\\epsilon}$ just $bK_{\\epsilon}$?If yes, why not use the simpler one?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698591939841}, {"id": "WGY58xHMjJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5017/Reviewer_JhF4"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper explores the impact of batch size on the iteration and gradient oracle complexities of the stochastic gradient descent (SGD) algorithm. The objective of the study is to examine how different batch sizes affect the performance of SGD. The paper is written in a reader-friendly manner, making it easily understandable.  In Tables 1 and 2, the authors present a summary of the iteration and gradient oracle complexities of the SGD method using various commonly used step sizes. By presenting this information, the authors offer valuable insights into the behavior of the algorithm. Furthermore, the authors conduct numerical experiments to compare the effectiveness of the step-decay strategy with other optimization algorithms. Through these experiments, they demonstrate the superior performance of step-decay in optimizing the objective function. This finding suggests that step-decay can be a preferable choice when implementing optimization algorithms. However, the contributions of this paper are not sufficient and some of the statements are wrong.", "review_text": "This paper explores the impact of batch size on the iteration and gradient oracle complexities of the stochastic gradient descent (SGD) algorithm. The objective of the study is to examine how different batch sizes affect the performance of SGD. The paper is written in a reader-friendly manner, making it easily understandable.  In Tables 1 and 2, the authors present a summary of the iteration and gradient oracle complexities of the SGD method using various commonly used step sizes. By presenting this information, the authors offer valuable insights into the behavior of the algorithm. Furthermore, the authors conduct numerical experiments to compare the effectiveness of the step-decay strategy with other optimization algorithms. Through these experiments, they demonstrate the superior performance of step-decay in optimizing the objective function. This finding suggests that step-decay can be a preferable choice when implementing optimization algorithms. However, the contributions of this paper are not sufficient and some of the statements are wrong.", "strengths": "This paper provides a thorough investigation into the relationship between batch size and the complexities of the SGD algorithm. The authors present their findings in a clear and concise manner, making them accessible to readers.", "weaknesses": "The paper is not well ready yet and contributions are trivial. As I checked, the analysis of SGD is quite simple and there is no technical challenge in the analysis. Besides, the main statements on step-decay are wrong. Please see the reasons below.\n\nTo calculate the iteration and gradient oracle complexity of the step-decay method, it is crucial to consider the impact of the lower bound values, denoted as $\\underline{\\alpha}$ and $T$ (representing the length of each stage for step-decay). Unfortunately, the authors of the paper have overlooked this important aspect, which is an incorrect approach. The reason why considering the lower bound values is essential lies in the relationship between $\\underline{\\alpha}$, $T$, and the total number of iterations, denoted as $K$. Specifically, we have the inequality $\\underline{\\alpha} \\leq \\alpha \\eta^{p-1}$, where $p = K/T$. This inequality implies that the lower bound value $\\underline{\\alpha}$ should be taken into account when determining the iteration and gradient oracle complexities of the step-decay method. Ignoring this relationship can lead to flawed conclusions and inaccurate assessments of the algorithm's performance. Therefore, the related complexities results on step-decay are wrong. \n\nOther weaknesses or typos:\n1. In the abstract, the authors made a claim that \"SGD using a step-decay learning rate and a small batch size reduces the SFO (Stochastic First-Order) complexity to find a local minimizer of a loss function.\" However, upon reviewing the paper, it becomes apparent that the study primarily focuses on demonstrating the convergence of SGD to a stationary point rather than specifically proving convergence to a local minimizer.", "questions": "See the weakness above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the impact of batch size on the iteration and gradient oracle complexities of the stochastic gradient descent (SGD) algorithm. The objective of the study is to examine how different batch sizes affect the performance of SGD. The paper is written in a reader-friendly manner, making it easily understandable.  In Tables 1 and 2, the authors present a summary of the iteration and gradient oracle complexities of the SGD method using various commonly used step sizes. By presenting this information, the authors offer valuable insights into the behavior of the algorithm. Furthermore, the authors conduct numerical experiments to compare the effectiveness of the step-decay strategy with other optimization algorithms. Through these experiments, they demonstrate the superior performance of step-decay in optimizing the objective function. This finding suggests that step-decay can be a preferable choice when implementing optimization algorithms. However, the contributions of this paper are not sufficient and some of the statements are wrong.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "This paper provides a thorough investigation into the relationship between batch size and the complexities of the SGD algorithm. The authors present their findings in a clear and concise manner, making them accessible to readers.", "weaknesses": "The paper is not well ready yet and contributions are trivial. As I checked, the analysis of SGD is quite simple and there is no technical challenge in the analysis. Besides, the main statements on step-decay are wrong. Please see the reasons below.\n\nTo calculate the iteration and gradient oracle complexity of the step-decay method, it is crucial to consider the impact of the lower bound values, denoted as $\\underline{\\alpha}$ and $T$ (representing the length of each stage for step-decay). Unfortunately, the authors of the paper have overlooked this important aspect, which is an incorrect approach. The reason why considering the lower bound values is essential lies in the relationship between $\\underline{\\alpha}$, $T$, and the total number of iterations, denoted as $K$. Specifically, we have the inequality $\\underline{\\alpha} \\leq \\alpha \\eta^{p-1}$, where $p = K/T$. This inequality implies that the lower bound value $\\underline{\\alpha}$ should be taken into account when determining the iteration and gradient oracle complexities of the step-decay method. Ignoring this relationship can lead to flawed conclusions and inaccurate assessments of the algorithm's performance. Therefore, the related complexities results on step-decay are wrong. \n\nOther weaknesses or typos:\n1. In the abstract, the authors made a claim that \"SGD using a step-decay learning rate and a small batch size reduces the SFO (Stochastic First-Order) complexity to find a local minimizer of a loss function.\" However, upon reviewing the paper, it becomes apparent that the study primarily focuses on demonstrating the convergence of SGD to a stationary point rather than specifically proving convergence to a local minimizer.", "questions": "See the weakness above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698397317610}], "openreview_url": "https://openreview.net/forum?id=lK0WxHeups", "arxiv_id": "2402.15344", "paper_pdf": "papers/lK0WxHeups.pdf", "paper_pdf_sha256": "47d36e70caf8e9455c6e6fa7be6f3bb98b8227dc969db534c02ab0f4fb6bf147", "paper_pdf_bytes": 414298, "paper_pdf_source": "openreview", "code_url": "https://github.com/imakn0907/SGD_using_decaying", "code_repository": "imakn0907/SGD_using_decaying", "code_commit": "26927a50592eb9db787bd972a7cd12b6fe3040e4", "code_archive": "repos/lK0WxHeups.zip", "code_archive_sha256": "905564fccb1082c81e06d1e3d76fb83edd19a8159dc7b23f838f752f7e01e815", "code_archive_bytes": 9837, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 24, "github_languages": {"Python": 27913}, "github_archived": false, "github_pushed_at": "2024-07-19T02:07:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/iteration-and-stochastic-first-order-oracle"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Z-aIURmBbBk", "year": 2023, "status": "rejected", "title": "Multimodal Masked Autoencoders Learn Transferable Representations", "authors": ["Xinyang Geng", "Hao Liu", "Lisa Lee", "Dale Schuurmans", "Sergey Levine", "Pieter Abbeel"], "authorids": ["~Xinyang_Geng1", "~Hao_Liu1", "~Lisa_Lee1", "~Dale_Schuurmans1", "~Sergey_Levine1", "~Pieter_Abbeel2"], "authors_source": "OpenReview API", "abstract": "Building scalable models to learn from diverse, multimodal data remains an open challenge.\nFor vision-language data, the dominant approaches are based on contrastive learning objectives that train a separate encoder for each modality. While effective, contrastive learning approaches introduce sampling bias depending on the data augmentations used, which can degrade performance on downstream tasks. Moreover, these methods are limited to paired image-text data, and cannot leverage widely-available unpaired data. In this paper, we investigate whether a large multimodal model trained purely via masked token prediction, without using modality-specific encoders or contrastive learning, can learn transferable representations for downstream tasks. We propose a simple and scalable network architecture, the Multimodal Masked Autoencoder (M3AE), which learns a unified encoder for both vision and language data via masked token prediction. We provide an empirical study of M3AE trained on a large-scale image-text dataset, and find that M3AE is able to learn generalizable representations that transfer well to downstream tasks. Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%, due to the joint training of two data modalities. We also provide qualitative analysis showing that the learned representation incorporates meaningful information from both image and language. Lastly, we demonstrate the scalability of M3AE with larger model size and training time, and its flexibility to train on both paired image-text data as well as unpaired data.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "unb23F9yyM3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3895/Reviewer_ZSrM"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper scales masked autoencoders (MAE) to image-text multimodal data, it investigates the use of modality-agnostic unified encoders, which learns generalizable representations across a number of downstream tasks.", "review_text": "The paper is overall well-motivated and well-written, yet the experiment part seems to fall short with a lack of discussion and comparison with other baselines.", "strengths": "Strengths:\n(a) The paper is overall well-written, the problem well formulated and novel to me. Generalizing masked autoencoders to work on more modalities is definitely interesting to the community.\nWeaknesses:\n(a) There seems to be a lack of comparison with other baselines. The paper only compare with CLIP and MAE, while not other baselines, such as MultiMAE[1], using the same modality-agnostic setting, were not included.\n(b) The author should also include a discussion on how is M3AE different from MultiMAE.\n[1] MultiMAE: Multi-modal Multi-task Masked Autoencoders, ECCV 2022", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper scales masked autoencoders (MAE) to image-text multimodal data, it investigates the use of modality-agnostic unified encoders, which learns generalizable representations across a number of downstream tasks.", "strength_and_weaknesses": "Strengths:\n(a) The paper is overall well-written, the problem well formulated and novel to me. Generalizing masked autoencoders to work on more modalities is definitely interesting to the community.\nWeaknesses:\n(a) There seems to be a lack of comparison with other baselines. The paper only compare with CLIP and MAE, while not other baselines, such as MultiMAE[1], using the same modality-agnostic setting, were not included.\n(b) The author should also include a discussion on how is M3AE different from MultiMAE.\n[1] MultiMAE: Multi-modal Multi-task Masked Autoencoders, ECCV 2022", "clarity,_quality,_novelty_and_reproducibility": "Well-motivated paper yet lacks discussion and comparison with other baselines.", "summary_of_the_review": "The paper is overall well-motivated and well-written, yet the experiment part seems to fall short with a lack of discussion and comparison with other baselines.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666870199667}, {"id": "kTJPdOHa6aK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3895/Reviewer_qu2E"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the multimodal representation learning in the masked auto-encoder way. The work is motivated by the fact that many existing multimodal learning approaches require a large amount of paired text-image data. This work builds upon MAE, extending it to text modality as well. The proposed method can handle both the unpaired image data and paired image-text data. The experiments show a clear improvement over MAE in several tasks, and show the effectiveness of the learned transferrable representation. \n\n--- post-rebuttal ---\\\nThanks for the authors' response. I have carefully read the response and other reviewers' comments. I believe the current form of this paper  is not ready for ICLR publication. I keep my score. ", "review_text": "Overall, the proposed solution is rational and mostly well presented. While my major concerns are lack of comparison with the existing method and no clear motivation given billions of paired text image data available. ", "strengths": "Strength \n\n-The idea of applying masked autoencoder strategy to multimodality data is rational. The finding on higher mask ratio on text performs better than the low mask ratio on pure text is interesting. The explanation makes sense to me. \n\n-The experiments show the strength over MAE in different scenarios and applications.  \n\n \n\nWeakness \n\n-There is other work also working on the similar motivation, i.e., leveraging unpaired data and paired text-image data for multimodal representation learning. BEIT-V3 [Wang et al.,], where they demonstrate an exceptional performance over many existing solutions including CLIP on a number of downstream tasks. It seems this paper misses the discussion on this work. And I’d also like to see the experimental comparison with BEIT-3 to show the advantages. \n\n[Wang et al.,] Wang et al.,Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks, https://arxiv.org/abs/2208.10442 \n\n-I have a little concern on the effectiveness on the large-scale data, such as several hundreds of millions of data points or even larger. The current training is on CC-12M, and shows a good improvement over MAE. However, when the number of training samples gets larger, I am not sure if this improvement margin can be preserved.  \n\n-Following the above point, thanks to the effort on the large scale data curation LAION, there are billions of image –text noisy pairs available. Please explain the application scenario of this method. \n\n-Figure 9, the reconstruction result from MAE is missing. I’d like to see a comparison to show the strength. \n\n-Other minors: In page 3, “For patches and tokens, we sample s random”, not sure what “s” means here. \n\nIn page 7, there is a space in “M3AEoutperforms”  ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the multimodal representation learning in the masked auto-encoder way. The work is motivated by the fact that many existing multimodal learning approaches require a large amount of paired text-image data. This work builds upon MAE, extending it to text modality as well. The proposed method can handle both the unpaired image data and paired image-text data. The experiments show a clear improvement over MAE in several tasks, and show the effectiveness of the learned transferrable representation. \n\n--- post-rebuttal ---\\\nThanks for the authors' response. I have carefully read the response and other reviewers' comments. I believe the current form of this paper  is not ready for ICLR publication. I keep my score. ", "strength_and_weaknesses": "Strength \n\n-The idea of applying masked autoencoder strategy to multimodality data is rational. The finding on higher mask ratio on text performs better than the low mask ratio on pure text is interesting. The explanation makes sense to me. \n\n-The experiments show the strength over MAE in different scenarios and applications.  \n\n \n\nWeakness \n\n-There is other work also working on the similar motivation, i.e., leveraging unpaired data and paired text-image data for multimodal representation learning. BEIT-V3 [Wang et al.,], where they demonstrate an exceptional performance over many existing solutions including CLIP on a number of downstream tasks. It seems this paper misses the discussion on this work. And I’d also like to see the experimental comparison with BEIT-3 to show the advantages. \n\n[Wang et al.,] Wang et al.,Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks, https://arxiv.org/abs/2208.10442 \n\n-I have a little concern on the effectiveness on the large-scale data, such as several hundreds of millions of data points or even larger. The current training is on CC-12M, and shows a good improvement over MAE. However, when the number of training samples gets larger, I am not sure if this improvement margin can be preserved.  \n\n-Following the above point, thanks to the effort on the large scale data curation LAION, there are billions of image –text noisy pairs available. Please explain the application scenario of this method. \n\n-Figure 9, the reconstruction result from MAE is missing. I’d like to see a comparison to show the strength. \n\n-Other minors: In page 3, “For patches and tokens, we sample s random”, not sure what “s” means here. \n\nIn page 7, there is a space in “M3AEoutperforms”  ", "clarity,_quality,_novelty_and_reproducibility": "The presentation is clear to me. and I'm satisfied with the details of the technical points. I think this work can be easily reproduced. Regarding novelty, I'm ok with the proposed method to extend MAE to multimodal representation learning. While I like to see the clear comparison with other existing efforts like BEIT-3 with a similar motivation.", "summary_of_the_review": "Overall, the proposed solution is rational and mostly well presented. While my major concerns are lack of comparison with the existing method and no clear motivation given billions of paired text image data available. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666723481245}, {"id": "yUs8vu6LrR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3895/Reviewer_1MQ5"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents an empirical study of M3AE trained on a large-scale image-text dataset, and find that multimodal masked autoencoder is able to learn generalizable representations that transfer well to downstream tasks. ", "review_text": "The paper is about the new approach for pre-training vision-language models by the proposed multimodal masked autoencoder. Although the approach is interesting, it is incremental and not novel. Some of the experiments demonstrating the superiority of the proposed approach are also unclear.\n\nThe proposed M3AE haven't performed experiments on vision-language down-stream tasks, it's not clear that the approach can learn good multimodal representation. Although the model can achieve better performances compared with MAE, M3AE uses larger dataset (e.g. CC12M). The performance gap between M3AE and CLIP is large.\n\nSome important ablation study experiments, such as different mask ratios of MAE and different encoders for visual/textual inputs, aren't conducted.", "strengths": "Strength:\n\n1. The paper is well written and this reviewer enjoyed reading it.\n\n2. The proposed approach is simple and intuitive.\n\n3. It shows promising results on image classification and out-of-distribution detection tasks.\n\nWeakness:\n1. The idea lacks of novelty. \n\n2. No experiments for vision-language tasks have been performed (e.g., text-to-image retrieval, VQA, visual reasoning, etc). The authors only fine-tune their pre-trained models for visual tasks, such as out-of-distribution detection and image classification. As the paper targets at multi-modal pre-training, it's essential to provide the experiments for vision-language down-stream tasks!\n\n3. Figure 2 is confusing. For example, in the lower left corner, the text \"Trees in a winter storm\" even appears in the space of its right image. \n\n4. Ablation study experiments aren't sufficient, it's important to conduct the experiments, e.g., different mask ratios of MAE.\n  \n5. Why not use two separate encoders for visual and textual feature extraction? The proposed M3AE only use one encoder for these two kinds of modalities together, but this may make the models confused. it's better to perform ablation study to use two different encoders, so that we can figure out the necessity of such design.\n\n6. The comparison is unfair, the proposed method uses more vision-language data (e.g. CC12M) for pre-training, and compare the models with MAE, which is only pre-trained on ImageNet. The performance gap between M3AE and CLIP is still very large. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents an empirical study of M3AE trained on a large-scale image-text dataset, and find that multimodal masked autoencoder is able to learn generalizable representations that transfer well to downstream tasks. ", "strength_and_weaknesses": "Strength:\n\n1. The paper is well written and this reviewer enjoyed reading it.\n\n2. The proposed approach is simple and intuitive.\n\n3. It shows promising results on image classification and out-of-distribution detection tasks.\n\nWeakness:\n1. The idea lacks of novelty. \n\n2. No experiments for vision-language tasks have been performed (e.g., text-to-image retrieval, VQA, visual reasoning, etc). The authors only fine-tune their pre-trained models for visual tasks, such as out-of-distribution detection and image classification. As the paper targets at multi-modal pre-training, it's essential to provide the experiments for vision-language down-stream tasks!\n\n3. Figure 2 is confusing. For example, in the lower left corner, the text \"Trees in a winter storm\" even appears in the space of its right image. \n\n4. Ablation study experiments aren't sufficient, it's important to conduct the experiments, e.g., different mask ratios of MAE.\n  \n5. Why not use two separate encoders for visual and textual feature extraction? The proposed M3AE only use one encoder for these two kinds of modalities together, but this may make the models confused. it's better to perform ablation study to use two different encoders, so that we can figure out the necessity of such design.\n\n6. The comparison is unfair, the proposed method uses more vision-language data (e.g. CC12M) for pre-training, and compare the models with MAE, which is only pre-trained on ImageNet. The performance gap between M3AE and CLIP is still very large. ", "clarity,_quality,_novelty_and_reproducibility": "1. Clarity - good; the paper is well-written.\n\n2. Quality - the experiments aren't sufficient. Please see the weakness 2,4 and 5.\n\n3. Originality - ok. The proposed M3AE lacks of novelty.\n\n4. Reproducibility - unclear. ", "summary_of_the_review": "The paper is about the new approach for pre-training vision-language models by the proposed multimodal masked autoencoder. Although the approach is interesting, it is incremental and not novel. Some of the experiments demonstrating the superiority of the proposed approach are also unclear.\n\nThe proposed M3AE haven't performed experiments on vision-language down-stream tasks, it's not clear that the approach can learn good multimodal representation. Although the model can achieve better performances compared with MAE, M3AE uses larger dataset (e.g. CC12M). The performance gap between M3AE and CLIP is large.\n\nSome important ablation study experiments, such as different mask ratios of MAE and different encoders for visual/textual inputs, aren't conducted.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough"}, "tcdate": 1666626053194}, {"id": "Rawpv6hLAd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3895/Reviewer_GbHq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes an extension of Masked Autoencoders (MAE) to the vision-language domain called Multi-Modal Masked Autoencoder (M3AE). Image-caption pairs are partially corrupted by masking some image patches and some language tokens. The non-masked image patches and language tokens are passed through an autoencoder based on transformers, which is trained to reconstruct the full image-caption, including masked parts. The embeddings of image patches that are learned by the transformer encoder are transferred to downstream image tasks, such as classification and out-of-distribution detection. ", "review_text": "See weaknesses 1-3 above. ", "strengths": "Strengths: The paper is easy to understand, with the methodology clearly presented. The results section includes some nice qualitative analysis and visualizations to help understand the representations that are learned by M3AE.\n\nWeaknesses: \n\n1) The empirical support for the proposed approach is weak.\n\nOne of the main takeaways from the experiments (Figure 3-5) is that M3AE outperforms MAE in the downstream ImageNet classification task. This is a good sanity check, but it is not significant nor surprising. The M3AE framework involves predicting text token logits from image patches, which is similar to a supervised classification task where the text is providing labels for training the image encoder. On the other hand, MAE is completely unsupervised. \n\nIt would be better to compare M3AE against other vision-language pretraining methods. However, no vision-language pretraining baselines are provided except for CLIP which outperforms M3AE by a large margin in Figure 3.\n\n2) In the introduction, the authors make the strong claim that M3AE offers two benefits over contrastive pretraining methods, but neither benefit is demonstrated in the experiments.\n\nFirst, the authors emphasize that contrastive learning \"requires paired image-and-text data and therefore cannot leverage widely available unpaired data.\" This suggests that leveraging unpaired data will make M3AE perform better than CLIP. But as shown in Figure 3, CLIP does much better than M3AE. To support this claim, the authors should either (i) decrease the paired data for CLIP to e.g., 10-30% and show that M3AE does better than CLIP for the same amount of paired data in Figure 3, or (ii) add some unpaired image or text data for M3AE and show that it can do better than CLIP by leveraging this unpaired data.\n\nSecond, the authors claim that the \"separation of image and text encoders\" in contrastive learning \"hinders the joint understanding of image and text\", but they provide no proof of this. To support this claim, they should show that M3AE performs better than contrastive learning on a multimodal downstream task. Also, note that separation of image and text encoders is not necessary in contrastive learning (e.g., GLIPv2 enables fusion between image and text features before the contrastive objective), so this is not an inherent limitation of contrastive pretraining methods.\n\n3) The authors write: \"Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%.\" This statement is misleading because it suggests that M3AE benefits from a higher text mask ratio compared to BERT on the same downstream task. But actually, BERT embeddings are evaluated on language tasks while M3AE embeddings are evaluated on image tasks. It does not make sense to compare BERT's text masking ratio with M3AE's, since M3AE's text representation is never evaluated in a downstream task.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes an extension of Masked Autoencoders (MAE) to the vision-language domain called Multi-Modal Masked Autoencoder (M3AE). Image-caption pairs are partially corrupted by masking some image patches and some language tokens. The non-masked image patches and language tokens are passed through an autoencoder based on transformers, which is trained to reconstruct the full image-caption, including masked parts. The embeddings of image patches that are learned by the transformer encoder are transferred to downstream image tasks, such as classification and out-of-distribution detection. ", "strength_and_weaknesses": "Strengths: The paper is easy to understand, with the methodology clearly presented. The results section includes some nice qualitative analysis and visualizations to help understand the representations that are learned by M3AE.\n\nWeaknesses: \n\n1) The empirical support for the proposed approach is weak.\n\nOne of the main takeaways from the experiments (Figure 3-5) is that M3AE outperforms MAE in the downstream ImageNet classification task. This is a good sanity check, but it is not significant nor surprising. The M3AE framework involves predicting text token logits from image patches, which is similar to a supervised classification task where the text is providing labels for training the image encoder. On the other hand, MAE is completely unsupervised. \n\nIt would be better to compare M3AE against other vision-language pretraining methods. However, no vision-language pretraining baselines are provided except for CLIP which outperforms M3AE by a large margin in Figure 3.\n\n2) In the introduction, the authors make the strong claim that M3AE offers two benefits over contrastive pretraining methods, but neither benefit is demonstrated in the experiments.\n\nFirst, the authors emphasize that contrastive learning \"requires paired image-and-text data and therefore cannot leverage widely available unpaired data.\" This suggests that leveraging unpaired data will make M3AE perform better than CLIP. But as shown in Figure 3, CLIP does much better than M3AE. To support this claim, the authors should either (i) decrease the paired data for CLIP to e.g., 10-30% and show that M3AE does better than CLIP for the same amount of paired data in Figure 3, or (ii) add some unpaired image or text data for M3AE and show that it can do better than CLIP by leveraging this unpaired data.\n\nSecond, the authors claim that the \"separation of image and text encoders\" in contrastive learning \"hinders the joint understanding of image and text\", but they provide no proof of this. To support this claim, they should show that M3AE performs better than contrastive learning on a multimodal downstream task. Also, note that separation of image and text encoders is not necessary in contrastive learning (e.g., GLIPv2 enables fusion between image and text features before the contrastive objective), so this is not an inherent limitation of contrastive pretraining methods.\n\n3) The authors write: \"Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%.\" This statement is misleading because it suggests that M3AE benefits from a higher text mask ratio compared to BERT on the same downstream task. But actually, BERT embeddings are evaluated on language tasks while M3AE embeddings are evaluated on image tasks. It does not make sense to compare BERT's text masking ratio with M3AE's, since M3AE's text representation is never evaluated in a downstream task.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity - good; the paper is clear and well-written.\n\nQuality - the experiments needs work. The main experimental result (that M3AE outperforms MAE) is not significant enough (weakness 1), and the other main claims in the introduction are also flawed (weaknesses 2-3)\n\nOriginality - ok. The concept of masked auto-encoding is not novel, although extension to the multimodal setting merits some consideration.", "summary_of_the_review": "See weaknesses 1-3 above. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666562285463}], "openreview_url": "https://openreview.net/forum?id=Z-aIURmBbBk", "arxiv_id": "2205.14204", "paper_pdf": "papers/Z-aIURmBbBk.pdf", "paper_pdf_sha256": "66784235da8e99eb1e93821727d5a2b4d847295c5cccd1a57cf0994c4132b5ed", "paper_pdf_bytes": 8448643, "paper_pdf_source": "openreview", "code_url": "https://github.com/young-geng/m3ae_public", "code_repository": "young-geng/m3ae_public", "code_commit": "a5eb14f851a2ad9b4fe84768af9d853b1e10436e", "code_archive": "repos/Z-aIURmBbBk.zip", "code_archive_sha256": "e623194547da52deef192200b3bfdedfaa3e7969a4f4df9dcf7a7494f1bf1e11", "code_archive_bytes": 40863, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 29, "github_languages": {"Python": 138734}, "github_archived": false, "github_pushed_at": "2025-02-26T19:05:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multimodal-masked-autoencoders-learn"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5_zwnS5oJDp", "year": 2022, "status": "rejected", "title": "Bayesian Learning with Information Gain Provably Bounds Risk for a Robust Adversarial Defense", "authors": ["Bao Gia Doan", "Ehsan M Abbasnejad", "Damith C. Ranasinghe"], "authorids": ["~Bao_Gia_Doan1", "~Ehsan_M_Abbasnejad1", "damith.ranasinghe@adelaide.edu.au"], "authors_source": "OpenReview API", "abstract": "In this paper, we present a novel method to learn a Bayesian neural network robust against adversarial attacks. Previous algorithms have shown an adversarially trained Bayesian Neural Network (BNN) provides improved robustness against attacks. However, the learning approach for approximating the multi-modal Bayesian posterior leads to mode collapse with consequential sub-par robustness and under performance of an adversarially trained BNN. Instead, we propose approximating the multi-modal posterior of a BNN to prevent mode collapse and encourage diversity over learned posterior distributions of models to develop a novel adversarial training method for BNNs.  Importantly, we conceptualize and formulate information gain (IG) in the adversarial Bayesian learning context and prove, training a BNN with IG bounds the difference between the conventional empirical risk with the risk obtained from adversarial training---our intuition is that information gain from benign and adversarial examples should be the same for a robust BNN. Extensive experimental results demonstrate our proposed algorithm to achieve state-of-the-art performance under strong adversarial attacks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "LeQSq9HYytw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1303/Reviewer_RZv2"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this paper, the authors propose a conjunction of Bayesian learning and regularization via information gain as a means of adversarial defense. In particular, they start from the perspective that Bayesian learning of neural network parameters ought to be more robust to adversarial examples, and then state that by jointly minimizing the risk and regularizing adversarial and clean information gain that the resulting Bayesian posterior will be even more robust. \n\nWhile this -- robustness and adversarial training of Bayesian neural networks -- is a worthwhile and important direction, the authors make several crucial mistakes and literature oversights which inhibit the potential impact of this work. ", "review_text": "Unfortunately, there are many claims made in this paper which are either overly strong, incorrect, or fail to appreciate prior work. \n\nFirstly, the authors correctly note that Bayesian neural networks are empirically and theoretically more robust to attacks. I think the authors should certainly cite [1] here as a motivating example. \n\nThe authors highlight the work of Adv-BNN as state of the art for Bayesian neural networks against adversarial attacks, but this has been greatly surpassed in [2] (more on this below). Moreover, their \"proposed adaptive PGD attack\" is not their own proposal and has been used in the Bayesian robustness literature for at least 3 years. The reason this proposed attack has been used is that the one given in Adv-BNN and repeated in equations (3) and (7) of this paper is both theoretically incorrect and empirically suboptimal. As pointed out by the authors, [3] clearly states the incorrectness of the attack used in Adv-BNN and shows it is suboptimal in practice for Bayesian neural networks (further corroborated by Tables 1 and 2 of this paper). This error is the Bayesian counter-part to the expectation over transformations (EoT) attack formalized in [4] which follows up on a work the authors cite, in particular [5]. Further, in works on BNNs this \"proposal\" has been adopted and is widely used, see [1,2,6], so I think it is incorrect to deem this as a \"proposed\" attack in the evaluation/contribution section of the paper and should be moved to the related works. \n\nFurther on this, despite the clear incorrectness of equation (3) and its corresponding formulation for this paper in equation (7) according to both deterministic principles [4,5] and Bayesian principles [2,3] the authors still use this attack during training and partially during evaluation. Thus, the methodology, as presented, is suboptimal/incorrect. In fact, the results section of this paper corroborate this claim (Table 2 versus Table 1) \n\nOutside of this incorrectness, the authors propose a regularization which is simply a specific form of Adversarial Logit Pairing [7]. This paper originally appeared in NeurIPS 2018, but was later retracted after it was found that this kind of robustness is very easily attacked and provides no substantial gains see [8]. I think if the authors would like to stick to this methodology then they ought to properly evaluate their method with a Bayesian adaptation of the method in [8]. I do not doubt the theoretical perspective given by the authors, only that it is practically effective/worthwhile. Moreover, I think the authors ought to be very careful in their general use of ERM approaches to likelihoods. This is perhaps acceptable for variational inference methods which replace marginalization with maximization, but in general, the method of [2] is preferable as it works for both maximization and marginalization and is thus on better grounds from a Bayesian point of view. Further on this point, I think it would benefit the paper if the authors stated their algorithm in a more general form as simply proposing a modification of a single inference method reduces the apparent contribution. \n\nFinally, I do have a further question about the methods evaluation. Is this method compared with AdvBNN applied to SVGD? Or is AdvBNN using Bayes by Backprop (as in its original formulation)? If so, the authors should ensure that the same inference method is used between their method and that of the network they compare against as inference method is known to have a considerably effect on the robustness of Bayesian posteriors [1,2,6]. In fact, the empirical performance noted in this paper might be solely the effect of a more faithful inference method and may have no correlation with the use of their method. Unless, the \"BNN\" label in Figure 2 is also SVGD. Still, the authors ought to compare with [2] prior to making the claim that they have state of the art robustness. \n\n[1] - https://arxiv.org/abs/2002.04359 (Appeared in NeurIPS last year)\n[2] - https://arxiv.org/abs/2102.05289 (Appeared in AISTATS last year)\n[3] - https://arxiv.org/abs/1907.00895 (Correction to Adv-BNN in 2019)\n[4] - https://arxiv.org/pdf/1707.07397.pdf (Appeared in ICML 2018)\n[5] - https://arxiv.org/pdf/1802.00420.pdf (Already cited in paper but here for completeness)\n[6] - https://arxiv.org/abs/2012.12640 (Appeared in AABI 2021)\n[7] - https://arxiv.org/abs/1803.06373 (Appeared, later retracted NeurIPS 2018)\n[8] - https://arxiv.org/abs/1807.10272\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose a conjunction of Bayesian learning and regularization via information gain as a means of adversarial defense. In particular, they start from the perspective that Bayesian learning of neural network parameters ought to be more robust to adversarial examples, and then state that by jointly minimizing the risk and regularizing adversarial and clean information gain that the resulting Bayesian posterior will be even more robust. \n\nWhile this -- robustness and adversarial training of Bayesian neural networks -- is a worthwhile and important direction, the authors make several crucial mistakes and literature oversights which inhibit the potential impact of this work. ", "main_review": "Unfortunately, there are many claims made in this paper which are either overly strong, incorrect, or fail to appreciate prior work. \n\nFirstly, the authors correctly note that Bayesian neural networks are empirically and theoretically more robust to attacks. I think the authors should certainly cite [1] here as a motivating example. \n\nThe authors highlight the work of Adv-BNN as state of the art for Bayesian neural networks against adversarial attacks, but this has been greatly surpassed in [2] (more on this below). Moreover, their \"proposed adaptive PGD attack\" is not their own proposal and has been used in the Bayesian robustness literature for at least 3 years. The reason this proposed attack has been used is that the one given in Adv-BNN and repeated in equations (3) and (7) of this paper is both theoretically incorrect and empirically suboptimal. As pointed out by the authors, [3] clearly states the incorrectness of the attack used in Adv-BNN and shows it is suboptimal in practice for Bayesian neural networks (further corroborated by Tables 1 and 2 of this paper). This error is the Bayesian counter-part to the expectation over transformations (EoT) attack formalized in [4] which follows up on a work the authors cite, in particular [5]. Further, in works on BNNs this \"proposal\" has been adopted and is widely used, see [1,2,6], so I think it is incorrect to deem this as a \"proposed\" attack in the evaluation/contribution section of the paper and should be moved to the related works. \n\nFurther on this, despite the clear incorrectness of equation (3) and its corresponding formulation for this paper in equation (7) according to both deterministic principles [4,5] and Bayesian principles [2,3] the authors still use this attack during training and partially during evaluation. Thus, the methodology, as presented, is suboptimal/incorrect. In fact, the results section of this paper corroborate this claim (Table 2 versus Table 1) \n\nOutside of this incorrectness, the authors propose a regularization which is simply a specific form of Adversarial Logit Pairing [7]. This paper originally appeared in NeurIPS 2018, but was later retracted after it was found that this kind of robustness is very easily attacked and provides no substantial gains see [8]. I think if the authors would like to stick to this methodology then they ought to properly evaluate their method with a Bayesian adaptation of the method in [8]. I do not doubt the theoretical perspective given by the authors, only that it is practically effective/worthwhile. Moreover, I think the authors ought to be very careful in their general use of ERM approaches to likelihoods. This is perhaps acceptable for variational inference methods which replace marginalization with maximization, but in general, the method of [2] is preferable as it works for both maximization and marginalization and is thus on better grounds from a Bayesian point of view. Further on this point, I think it would benefit the paper if the authors stated their algorithm in a more general form as simply proposing a modification of a single inference method reduces the apparent contribution. \n\nFinally, I do have a further question about the methods evaluation. Is this method compared with AdvBNN applied to SVGD? Or is AdvBNN using Bayes by Backprop (as in its original formulation)? If so, the authors should ensure that the same inference method is used between their method and that of the network they compare against as inference method is known to have a considerably effect on the robustness of Bayesian posteriors [1,2,6]. In fact, the empirical performance noted in this paper might be solely the effect of a more faithful inference method and may have no correlation with the use of their method. Unless, the \"BNN\" label in Figure 2 is also SVGD. Still, the authors ought to compare with [2] prior to making the claim that they have state of the art robustness. \n\n[1] - https://arxiv.org/abs/2002.04359 (Appeared in NeurIPS last year)\n[2] - https://arxiv.org/abs/2102.05289 (Appeared in AISTATS last year)\n[3] - https://arxiv.org/abs/1907.00895 (Correction to Adv-BNN in 2019)\n[4] - https://arxiv.org/pdf/1707.07397.pdf (Appeared in ICML 2018)\n[5] - https://arxiv.org/pdf/1802.00420.pdf (Already cited in paper but here for completeness)\n[6] - https://arxiv.org/abs/2012.12640 (Appeared in AABI 2021)\n[7] - https://arxiv.org/abs/1803.06373 (Appeared, later retracted NeurIPS 2018)\n[8] - https://arxiv.org/abs/1807.10272\n\n", "summary_of_the_review": "The paper is in a worthwhile direction, but ultimately has too many oversights and errors to be accepted as is. The attack methodology relied upon is known to be incorrect and suboptimal. The regularization introduced is a particular flavor of a defense which is known to be suboptimal and easily attacked. The claim of state-of-the-art performance is outdated for BNNs. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None for this paper.", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635937851664}, {"id": "tuoBa3tmR8f", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1303/Reviewer_qjBg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes leveraging \"information gain\" with Bayesian Neural Networks to improve the robustness of Bayesian Neural Networks. Specifically, it maintains a set of particle models and uses SVGD to update this set of particle models. Furthermore, it uses the PGD-based attack to a specific particle model to craft adversarial examples and enforces the information gain of adversarial examples to be close to that of the corresponding benign examples. ", "review_text": "**Strengths of the paper**\n- The information gain is useful to improve the robustness of Bayesian Neural Networks.\n- The proposed method outperforms the baselines.\n\n**Weaknesses of the paper**\n- Although the proposed method is intuitive, its framework is not solidly and rigorously developed.\n- Information gain seems identical to the mutual information $I(y,\\theta \\mid x, D)$ proposed and investigated in Bayesian Active Learning [1,2]. The authors need to discuss their proposed information gain and the mutual information proposed in [1,2].  \n\n[1] Houlsby, Neil, et al. \"Bayesian active learning for classification and preference learning.\" arXiv preprint arXiv:1112.5745 (2011).\n\n[2]  Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1183–1192. JMLR. org, 2017.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes leveraging \"information gain\" with Bayesian Neural Networks to improve the robustness of Bayesian Neural Networks. Specifically, it maintains a set of particle models and uses SVGD to update this set of particle models. Furthermore, it uses the PGD-based attack to a specific particle model to craft adversarial examples and enforces the information gain of adversarial examples to be close to that of the corresponding benign examples. ", "main_review": "**Strengths of the paper**\n- The information gain is useful to improve the robustness of Bayesian Neural Networks.\n- The proposed method outperforms the baselines.\n\n**Weaknesses of the paper**\n- Although the proposed method is intuitive, its framework is not solidly and rigorously developed.\n- Information gain seems identical to the mutual information $I(y,\\theta \\mid x, D)$ proposed and investigated in Bayesian Active Learning [1,2]. The authors need to discuss their proposed information gain and the mutual information proposed in [1,2].  \n\n[1] Houlsby, Neil, et al. \"Bayesian active learning for classification and preference learning.\" arXiv preprint arXiv:1112.5745 (2011).\n\n[2]  Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1183–1192. JMLR. org, 2017.", "summary_of_the_review": "The posterior $p(\\theta \\mid D_{adv})$ in Eq. (5) is not mathematically rigorous because it does not contain $D$ on the left side, but $D$ is involved on the right side. \n\nI cannot see the definition of the information gain and it seems to be identical to the mutual information proposed in [1,2]. Moreover, the first term on the right side of  Eq. (8) is $H(y \\mid x, D)$ because it is independent with $\\theta$. The second term should be $E_\\theta[H(y \\mid x, \\theta)]$ according to your derivation in Appendix.\n\nSVGD is not proposed to minimize the loss function directly. Although the update of the particle models in Algorithm 1 is fine, it requires more effort to define the distribution for which we only know its unnormalized version. Furthermore, Eq. (9) needs to have the absolute value for the difference between $IG(x)$ and $IG(x_{adv})$. In addition, the notion here is inconsistent with the lack of $y$.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635933624520}, {"id": "wMFQE9_sp24", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1303/Reviewer_UwZA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper explores adversarial training for Bayesian Neural Networks. In contrast to prior work which used Gaussian variational posteriors to estimate BNN parameters during adversarial training, this work uses Stein Variational Gradient Descent (SVGD) on a loss potential that includes an adversarial term similar to standard adversarial training and a term encouraging similarity between the Information Gain (IG) on natural and adversarial images. They evaluate PGD and adaptive adversarial attacks and find significant performance improvement over standard adversarial training and adversarial BNN. A theoretical justification is presented.", "review_text": "Strengths:\n* The method appears to offer significant improvements over prior adversarial BNN defense.\n* The authors implement an appropriate adaptive attack to account for the randomness of Bayesian Neural Networks. It might be helpful to refer to the attack with the widely-used terminology of Expectation-Over-Transformation (EOT) attack, as in Athalye et. al (2018). However, the authors place much more prominence in the text on a non-adaptive PGD attack, which might cause misperceptions (see Weaknesses).\n\nWeaknesses:\n* While the authors perform an appropriate adaptive EOT attack, the plots and main results present the outcomes of a non-adaptive PGD attack. It is well known that deterministic attacks are not representative of the performance of defenses that have stochastic elements, and the authors indeed see a significant drop in accuracy. Since a defense is only as good as its performance against an optimal attack, the authors should also include a curve showing robustness for their method against adaptive attacks in Figure 2 to give fair comparison versus adversarial training.\n* The theoretical analysis appears to be quite trivial. Unless I am mistaken, adding the term $|E_\\theta [ IG(x) ] - E_\\theta [ IG(x_\\text{adv}) ] |$ in (11) only makes the bound looser and thus less useful. The theoretical presentation does not enhance the arguments of the paper.\n* The work lacks novelty, and its novel elements are not properly evaluated. The main contributions are the inclusion of SVGD and the Information Gain term. The theoretical importance of the Information Gain term is not very convincing. The authors do not perform an ablative study using just SVGD and no IG term vs. Gaussian variational inference plus IG term to narrow down which elements are important, or if both are necessary.\n* The adaptive attack is not completely explored. The authors do not mention how many EOT replicates are used. 20 PGD steps is not enough for the EOT defense performance to reach its full strength in some cases where there is a high degree of stochasticity, and more steps should be used to ensure convergence.\n\nOther Comments:\n* How are the particles $\\theta_i^0$ initialized during the testing phase? I would assume they are initialized from a set of fixed parameters identified during the training phase, is that correct? More details on this point would enhance my understanding of the work.\n* Source code would help to further evaluate this defense.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper explores adversarial training for Bayesian Neural Networks. In contrast to prior work which used Gaussian variational posteriors to estimate BNN parameters during adversarial training, this work uses Stein Variational Gradient Descent (SVGD) on a loss potential that includes an adversarial term similar to standard adversarial training and a term encouraging similarity between the Information Gain (IG) on natural and adversarial images. They evaluate PGD and adaptive adversarial attacks and find significant performance improvement over standard adversarial training and adversarial BNN. A theoretical justification is presented.", "main_review": "Strengths:\n* The method appears to offer significant improvements over prior adversarial BNN defense.\n* The authors implement an appropriate adaptive attack to account for the randomness of Bayesian Neural Networks. It might be helpful to refer to the attack with the widely-used terminology of Expectation-Over-Transformation (EOT) attack, as in Athalye et. al (2018). However, the authors place much more prominence in the text on a non-adaptive PGD attack, which might cause misperceptions (see Weaknesses).\n\nWeaknesses:\n* While the authors perform an appropriate adaptive EOT attack, the plots and main results present the outcomes of a non-adaptive PGD attack. It is well known that deterministic attacks are not representative of the performance of defenses that have stochastic elements, and the authors indeed see a significant drop in accuracy. Since a defense is only as good as its performance against an optimal attack, the authors should also include a curve showing robustness for their method against adaptive attacks in Figure 2 to give fair comparison versus adversarial training.\n* The theoretical analysis appears to be quite trivial. Unless I am mistaken, adding the term $|E_\\theta [ IG(x) ] - E_\\theta [ IG(x_\\text{adv}) ] |$ in (11) only makes the bound looser and thus less useful. The theoretical presentation does not enhance the arguments of the paper.\n* The work lacks novelty, and its novel elements are not properly evaluated. The main contributions are the inclusion of SVGD and the Information Gain term. The theoretical importance of the Information Gain term is not very convincing. The authors do not perform an ablative study using just SVGD and no IG term vs. Gaussian variational inference plus IG term to narrow down which elements are important, or if both are necessary.\n* The adaptive attack is not completely explored. The authors do not mention how many EOT replicates are used. 20 PGD steps is not enough for the EOT defense performance to reach its full strength in some cases where there is a high degree of stochasticity, and more steps should be used to ensure convergence.\n\nOther Comments:\n* How are the particles $\\theta_i^0$ initialized during the testing phase? I would assume they are initialized from a set of fixed parameters identified during the training phase, is that correct? More details on this point would enhance my understanding of the work.\n* Source code would help to further evaluate this defense.", "summary_of_the_review": "I recommend to not accept this paper. The theoretical analysis of the Information Gain term appears vacuous and distracting, although I could be missing something. The contributions (SVGD and Information Gain for BNN posterior sampling) are not significantly novel, and their relative importance is not properly evaluated. The experimental results appear strong, but it's difficult to judge the quality of the defense experiments without being able to inspect the to source code (in particular, the implementation of the EOT attack).", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635830961955}, {"id": "9lZGgyRwp7e", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1303/Reviewer_CwGg"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose to learn a multi-modal posterior of the Bayesian neural network to defense the adversarial attacks, which can prevent the collapse and encourage the diversity accordingly. The authors further investigate the information gain for adversarial Bayesian learning, which is used to guide the training a BNN with the information gain bound. Experimental results demonstrate tis superior performance for adversarial attacks. ", "review_text": "In general, the paper is clearly written by addressing an important problem, but I still have several concerns.\n\n1.\tIt is interesting that the authors propose to use the Bayesian formulation for adversarial training. Nevertheless, the authors are expected to make a thorough analysis since the benign examples and the adversarial examples are not independent. The authors are expected to clarify the reasonability of Eq. 5.\n2.\tTo some extent, the paper is highly dependent on the information gain. However, it is nontrivial to estimate the information gain especially for the adversarial scenarios. A neural network is always overconfident for the adversarial examples. The authors are expected to clarify to how to estimate the information gain reliably. \n3.\tThe paper is inherently a regularized version for BNN with information gain. The authors are expected to clarify the novelties and technical contributions. \n4.\tIn the experimental section, the performance is significantly better than the alternative methods, e.g., Adv-BNN, especially for the adaptive attacks. The authors are expected to make thorough analysis on why or for what technical aspects it can achieve the performance improvement. \n5.\tThe authors are expected to make more comprehensive analysis, e.g., more experiments on ImageNet, more defense/attacks methods. \n   \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose to learn a multi-modal posterior of the Bayesian neural network to defense the adversarial attacks, which can prevent the collapse and encourage the diversity accordingly. The authors further investigate the information gain for adversarial Bayesian learning, which is used to guide the training a BNN with the information gain bound. Experimental results demonstrate tis superior performance for adversarial attacks. ", "main_review": "In general, the paper is clearly written by addressing an important problem, but I still have several concerns.\n\n1.\tIt is interesting that the authors propose to use the Bayesian formulation for adversarial training. Nevertheless, the authors are expected to make a thorough analysis since the benign examples and the adversarial examples are not independent. The authors are expected to clarify the reasonability of Eq. 5.\n2.\tTo some extent, the paper is highly dependent on the information gain. However, it is nontrivial to estimate the information gain especially for the adversarial scenarios. A neural network is always overconfident for the adversarial examples. The authors are expected to clarify to how to estimate the information gain reliably. \n3.\tThe paper is inherently a regularized version for BNN with information gain. The authors are expected to clarify the novelties and technical contributions. \n4.\tIn the experimental section, the performance is significantly better than the alternative methods, e.g., Adv-BNN, especially for the adaptive attacks. The authors are expected to make thorough analysis on why or for what technical aspects it can achieve the performance improvement. \n5.\tThe authors are expected to make more comprehensive analysis, e.g., more experiments on ImageNet, more defense/attacks methods. \n   \n", "summary_of_the_review": "The paper is easy to follow, but the authors are expected to clarify their technical contributions and make more comprehensive experiments. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635775961225}], "openreview_url": "https://openreview.net/forum?id=5_zwnS5oJDp", "arxiv_id": "2212.02003", "paper_pdf": "papers/5_zwnS5oJDp.pdf", "paper_pdf_sha256": "72193fa422207159a2e9e73cded029e1b15b05ea040551fff0a164edb77dbd4c", "paper_pdf_bytes": 1049078, "paper_pdf_source": "openreview", "code_url": "https://github.com/baogiadoan/IG-BNN", "code_repository": "baogiadoan/IG-BNN", "code_commit": "1976cce0d0d4a7242f83fcfee825862cb53aea6b", "code_archive": "repos/5_zwnS5oJDp.zip", "code_archive_sha256": "11ccdc3d8099cc916ac27e51fe53a7d462088ce35bf726f56b344de7b4faf474", "code_archive_bytes": 29436, "code_file_count": 24, "code_extensions": {".py": 22, ".sh": 2}, "github_disk_usage_kb": 27, "github_languages": {"Python": 83703, "Shell": 879}, "github_archived": false, "github_pushed_at": "2024-12-05T07:23:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bayesian-learning-with-information-gain-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HmAhqnu3qu", "year": 2021, "status": "rejected", "title": "Graph Representation Learning for Multi-Task Settings: a Meta-Learning Approach", "authors": ["Davide Buffelli", "Fabio Vandin"], "authorids": ["~Davide_Buffelli1", "~Fabio_Vandin2"], "authors_source": "OpenReview API", "abstract": "Graph Neural Networks (GNNs) have become the state-of-the-art method for many applications on graph structured data. GNNs are a framework for graph representation learning, where a model learns to generate low dimensional node embeddings that encapsulate structural and feature-related information. GNNs are usually trained in an end-to-end fashion, leading to highly specialized node embeddings. While this approach achieves great results in the single-task setting, generating node embeddings that can be used to perform multiple tasks (with performance comparable to single-task models) is an open problem. We propose a novel representation learning strategy, based on meta-learning, capable of producing multi-task node embeddings. Our method avoids the difficulties arising when learning to perform multiple tasks concurrently by, instead, learning to quickly (i.e. with a few steps of gradient descent) adapt to multiple tasks singularly.  We show that the embeddings produced by our method can be used to perform multiple tasks with comparable or higher performance than both single-task and multi-task end-to-end models. Our method is model-agnostic and task-agnostic and can hence be applied to a wide variety of multi-task domains.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "x52RDBfnnLP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper849/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The manuscript proposes SAME, a model based on GNN and meta-learning for learning multi-task node embeddings. Unlike multi-task learning setting, SAME aims at learning to quickly adapt to multiple tasks. Two model variants iSAME and eSAME are proposed base on different settings in inner/outer loop of parameter update. Experiments on several datasets demonstrate the good performance of SAME. \n\nPros\n1. The problem is new and interesting. It is the first work to study single set of node embeddings for multi-tasks in graph. \n2. The presentation is overall good. The content is clear for me. \n3. Introduce a new model for learning multi-task node embeddings through meta-learning way. The model is simple yet interesting for new problem. \n\nCons/Questions\n1. The novelty of this work incremental. Despite the new problem and different task settings, the model framework adopts the similar procedure as general meta-learning procedure. The model is quite simple. I would like to see more discussion about the contribution and novelty of this work as well as the potential future study. \n\n2. This work follows the meta-learning setting. Besides studying different graph learning tasks, it is better to provide content and add experiment for the scenario of few-shot labeled data. If I did not miss it, there is no discussion and experiment about this. I would suggest the authors to add comparison experiments for different tasks where only few-shot supervised data are available. \n\n3. Reference format is not consistent, typos, etc. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": " interesting  problem ", "review": "The manuscript proposes SAME, a model based on GNN and meta-learning for learning multi-task node embeddings. Unlike multi-task learning setting, SAME aims at learning to quickly adapt to multiple tasks. Two model variants iSAME and eSAME are proposed base on different settings in inner/outer loop of parameter update. Experiments on several datasets demonstrate the good performance of SAME. \n\nPros\n1. The problem is new and interesting. It is the first work to study single set of node embeddings for multi-tasks in graph. \n2. The presentation is overall good. The content is clear for me. \n3. Introduce a new model for learning multi-task node embeddings through meta-learning way. The model is simple yet interesting for new problem. \n\nCons/Questions\n1. The novelty of this work incremental. Despite the new problem and different task settings, the model framework adopts the similar procedure as general meta-learning procedure. The model is quite simple. I would like to see more discussion about the contribution and novelty of this work as well as the potential future study. \n\n2. This work follows the meta-learning setting. Besides studying different graph learning tasks, it is better to provide content and add experiment for the scenario of few-shot labeled data. If I did not miss it, there is no discussion and experiment about this. I would suggest the authors to add comparison experiments for different tasks where only few-shot supervised data are available. \n\n3. Reference format is not consistent, typos, etc. ", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604154514321}, {"id": "th96Uu1O5h6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper849/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper formulates the learning of three tasks, including graph classification, node classification and link prediction, as a multi-task learning problem and adopts a meta learning approach to learn the three tasks together in the spirit of the Model-Agnostic Meta Learning (MAML) method.\n\nActually, there have been some works to study the three tasks (i.e., graph classification, node classification and link prediction) as a multi-task learning problem. Authors need to discuss differences with those works and compare with them in experiments.\n\nThe proposed meta learning approach seems a direct application of the MAML method. I cannot see much difference with the MAML method.\n\nIn the meta-objective, how to set different \\lambda’s? This is more important to the performance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Direct extension of MAML", "review": "This paper formulates the learning of three tasks, including graph classification, node classification and link prediction, as a multi-task learning problem and adopts a meta learning approach to learn the three tasks together in the spirit of the Model-Agnostic Meta Learning (MAML) method.\n\nActually, there have been some works to study the three tasks (i.e., graph classification, node classification and link prediction) as a multi-task learning problem. Authors need to discuss differences with those works and compare with them in experiments.\n\nThe proposed meta learning approach seems a direct application of the MAML method. I cannot see much difference with the MAML method.\n\nIn the meta-objective, how to set different \\lambda’s? This is more important to the performance.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604042719583}, {"id": "4nMX3y5g8J", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper849/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a multi-task framework to represent the node embedding for transferred knowledge. The methodology is based on the meta-learning, which is capable of producing multi-task node embedding. This paper is well-motivated and well-written. The experimental results illustrate the effectiveness of the model. I would like to recommend to accept this paper.\n\nMajor Concerns:\n1. The related work can be strengthened. In the current version, the related work seems to stack all the papers in sequence, which makes it tough to understand the development of this task. I suggest the authors to reformate this subsection to Mind Graph.\n2. Can you introduce the dataset more specifically?\n3. There shall be more baselines for the experiments.\n\nMinor Concern:\n1. Fig.1: the legend shall be outside the box.\n2. Sec 3 seems redundant to most related readers. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Graph Representation Multi-Task Learning.", "review": "This paper presents a multi-task framework to represent the node embedding for transferred knowledge. The methodology is based on the meta-learning, which is capable of producing multi-task node embedding. This paper is well-motivated and well-written. The experimental results illustrate the effectiveness of the model. I would like to recommend to accept this paper.\n\nMajor Concerns:\n1. The related work can be strengthened. In the current version, the related work seems to stack all the papers in sequence, which makes it tough to understand the development of this task. I suggest the authors to reformate this subsection to Mind Graph.\n2. Can you introduce the dataset more specifically?\n3. There shall be more baselines for the experiments.\n\nMinor Concern:\n1. Fig.1: the legend shall be outside the box.\n2. Sec 3 seems redundant to most related readers. ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603846636948}], "openreview_url": "https://openreview.net/forum?id=HmAhqnu3qu", "arxiv_id": "2201.03326", "paper_pdf": "papers/HmAhqnu3qu.pdf", "paper_pdf_sha256": "f55b84171afba0b57895775543088267a6852eef1abb73461ecc475d8cbc0119", "paper_pdf_bytes": 452963, "paper_pdf_source": "openreview", "code_url": "https://github.com/DavideBuffelli/SAME", "code_repository": "DavideBuffelli/SAME", "code_commit": "cfbdda3a4b63bb4a5b92ec6cd21a37fcede47bad", "code_archive": "repos/HmAhqnu3qu.zip", "code_archive_sha256": "d45745192255214352d324a0524e6526fca71a07a4bbe06917a1b63d7184bea6", "code_archive_bytes": 31523, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 31, "github_languages": {"Python": 118158}, "github_archived": false, "github_pushed_at": "2022-04-26T18:48:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-representation-learning-for-multi-task-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1xfElrKPr", "year": 2020, "status": "rejected", "title": "Enhancing the Transformer with explicit relational encoding for math problem solving", "authors": ["Imanol Schlag", "Paul Smolensky", "Roland Fernandez", "Nebojsa Jojic", "Jürgen Schmidhuber", "Jianfeng Gao"], "authorids": ["imanol@idsia.ch", "paul.smolensky@gmail.com", "rfernand@microsoft.com", "jojic@microsoft.com", "juergen@idsia.ch", "jfgao@microsoft.com"], "authors_source": "OpenReview API", "abstract": "We incorporate Tensor-Product Representations within the Transformer in order to better support the explicit representation of relation structure.\nOur Tensor-Product Transformer (TP-Transformer) sets a new state of the art on the recently-introduced Mathematics Dataset containing 56 categories of free-form math word-problems.\nThe essential component of the model is a novel attention mechanism, called TP-Attention, which explicitly encodes the relations between each Transformer cell and the other cells from which values have been retrieved by attention. TP-Attention goes beyond linear combination of retrieved values, strengthening representation-building and resolving ambiguities introduced by multiple layers of regular attention.\nThe TP-Transformer's attention maps give better insights into how it is capable of solving the Mathematics Dataset's challenging problems.\nPretrained models and code will be made available after publication.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hyx0sLGk5S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2239/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Motivated by the fact that the attention mechanism in transformers is symmetric which might not be able to disambiguate different orders, this work proposes to use a subject vector (in addition to query, key states) for each attention head, and multiply it elementwise with the context vector for each head before merging the heads. Experiments on a mathematics dataset shows superior performance compared to the normal transformer. Qualitatively, the proposed model exhibits attentions that are more interpretable, and clustering by the subject vector gives some insights into how the model solved this problem.\n\nPros:\n1. This work shows better performance than baseline transformer.\n2. The clustering of the subject vectors gives some insights into model's behavior .\n\nCons:\n1. In terms of experiments, the proposed approach adds a few million parameters to normal transformer (table 1), but in terms of interpolation it only improves 3% (extrapolation improves 0.5%) at 700k steps. The comparison would be fairer if the normal transformer can be given more parameters.\n2. In terms of experiments, this approach is only evaluated on the mathematics dataset, but the argument for relational encoding is pretty general. It would be nice if experiments on other tasks are shown in addition to the math dataset.\n3. In terms of motivation, the claim that there're ambiguities introduced by multiple layers of  regular attention needs to be supported by evidence. I think (which authors also pointed out) the feedforward network and non-linearties can disambiguate as well.\n4. In terms of interpretablity, there's claim that the learned attention maps more interpretable than transformer. Can there be more quantitative measures? It appears to me that both are hard to interpret.\n\nWhile this work shows superior performance on the mathematics dataset, I have a few concerns about the generalizability of this proposed architectural change to other problems, as well as the fairness of comparison to baseline. Therefore, I am inclined to reject this paper.\n\n----updates after reading rebuttal----\nThanks for adding the new NMT experiment in Appendix A3. My concern is that the proposed TP-Transformer is not very effective on NMT. Therefore, I'm keeping my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "Motivated by the fact that the attention mechanism in transformers is symmetric which might not be able to disambiguate different orders, this work proposes to use a subject vector (in addition to query, key states) for each attention head, and multiply it elementwise with the context vector for each head before merging the heads. Experiments on a mathematics dataset shows superior performance compared to the normal transformer. Qualitatively, the proposed model exhibits attentions that are more interpretable, and clustering by the subject vector gives some insights into how the model solved this problem.\n\nPros:\n1. This work shows better performance than baseline transformer.\n2. The clustering of the subject vectors gives some insights into model's behavior .\n\nCons:\n1. In terms of experiments, the proposed approach adds a few million parameters to normal transformer (table 1), but in terms of interpolation it only improves 3% (extrapolation improves 0.5%) at 700k steps. The comparison would be fairer if the normal transformer can be given more parameters.\n2. In terms of experiments, this approach is only evaluated on the mathematics dataset, but the argument for relational encoding is pretty general. It would be nice if experiments on other tasks are shown in addition to the math dataset.\n3. In terms of motivation, the claim that there're ambiguities introduced by multiple layers of  regular attention needs to be supported by evidence. I think (which authors also pointed out) the feedforward network and non-linearties can disambiguate as well.\n4. In terms of interpretablity, there's claim that the learned attention maps more interpretable than transformer. Can there be more quantitative measures? It appears to me that both are hard to interpret.\n\nWhile this work shows superior performance on the mathematics dataset, I have a few concerns about the generalizability of this proposed architectural change to other problems, as well as the fairness of comparison to baseline. Therefore, I am inclined to reject this paper.\n\n----updates after reading rebuttal----\nThanks for adding the new NMT experiment in Appendix A3. My concern is that the proposed TP-Transformer is not very effective on NMT. Therefore, I'm keeping my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571919526398}, {"id": "HJeOaw0CKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2239/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors incorporated tensor-product representations within the Transformer.  By creating an attention mechanism called TP-Attention, they explicitly encode the relations between each Transformer cell and the other cells, whose values are retrieved by attention. By introducing tensor products, the proposed algorithm can empirically perform well for noncommutative operations with multiple arguments, such as division. The authors trained models with the proposed algorithm on the Mathematics Dataset and compared the performances with two baselines (simple LSTM and the original Transformer). At last, several model snapshots are provided to help interpret several key elements of the model: the learned roles, the attention maps, the TP-transformer columns and so on. \n\nOverall, the paper is well-written. The experimental results generally support the high-level intuition behind the introduction of tensor-product representation. I would recommend accepting this paper. \n\nSome quick questions:\n\n1. It was claimed in the Conclusion section that the performance of the proposed algorithm beats the previously published state of the art by 8.24%. I guess the number comes from the 2nd and the last row of interpolation accuracy in Table 1. However, these two results are obviously trained for different numbers of iterations: The baseline algorithm was trained for 500k steps, while the proposed algorithm is trained for 1.7M steps. Is it a fair comparison? If the proposed algorithm is also trained for 500k steps, the improvement is around 2.3%. \n\n2. Why is the extrapolation accuracy results for TP-Transformer missing in Table 1? \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "In this paper, the authors incorporated tensor-product representations within the Transformer.  By creating an attention mechanism called TP-Attention, they explicitly encode the relations between each Transformer cell and the other cells, whose values are retrieved by attention. By introducing tensor products, the proposed algorithm can empirically perform well for noncommutative operations with multiple arguments, such as division. The authors trained models with the proposed algorithm on the Mathematics Dataset and compared the performances with two baselines (simple LSTM and the original Transformer). At last, several model snapshots are provided to help interpret several key elements of the model: the learned roles, the attention maps, the TP-transformer columns and so on. \n\nOverall, the paper is well-written. The experimental results generally support the high-level intuition behind the introduction of tensor-product representation. I would recommend accepting this paper. \n\nSome quick questions:\n\n1. It was claimed in the Conclusion section that the performance of the proposed algorithm beats the previously published state of the art by 8.24%. I guess the number comes from the 2nd and the last row of interpolation accuracy in Table 1. However, these two results are obviously trained for different numbers of iterations: The baseline algorithm was trained for 500k steps, while the proposed algorithm is trained for 1.7M steps. Is it a fair comparison? If the proposed algorithm is also trained for 500k steps, the improvement is around 2.3%. \n\n2. Why is the extrapolation accuracy results for TP-Transformer missing in Table 1? \n\n"}, "tcdate": 1571903424052}, {"id": "rkx9a33nKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2239/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper illustrates the TP-Transformer architecture on the challenging mathematics dataset. The TP-Transformer combines the transformer architecture with tensor-product representations. The experiments show a dramatic improvement of accuracies compared with SOTA models. Moreover, the paper also explains the reason why the TP-Transformer can learn the structural position and relation to other symbols with a detailed math proof.\n \nOverall, this paper is nice as it makes a milestone for math problem solving from unique perspectives. To be specific, the paper makes the following contributions:\n\n1. Demonstrate a novel architecture TP-Transformer in details;\n2. Achieve a better accuracies in the challenging mathematics dataset than the SOTA transformer models;\n3. Illustrate in fundamental math that why TP-Transformer can learn the structural position and relation, and solve the binding problems of stacked attention layers.\n \nHere are a few minor questions that may further improve the paper:\n\n1. The conclusion states that TP-Transformer beats the previously published SOTA by 8.24%. However, it does not match to the experiment results (see section 4).\n\n2. In figure 5, there are 4 tasks in the bottom with accuracies lower than 0.5. It would be nice to provide more insights on this.\n \n3. It would be interesting to see whether it transferable to the other downstream tasks (such as natural language understanding) besides the experiments on the challenging mathematics dataset.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper illustrates the TP-Transformer architecture on the challenging mathematics dataset. The TP-Transformer combines the transformer architecture with tensor-product representations. The experiments show a dramatic improvement of accuracies compared with SOTA models. Moreover, the paper also explains the reason why the TP-Transformer can learn the structural position and relation to other symbols with a detailed math proof.\n \nOverall, this paper is nice as it makes a milestone for math problem solving from unique perspectives. To be specific, the paper makes the following contributions:\n\n1. Demonstrate a novel architecture TP-Transformer in details;\n2. Achieve a better accuracies in the challenging mathematics dataset than the SOTA transformer models;\n3. Illustrate in fundamental math that why TP-Transformer can learn the structural position and relation, and solve the binding problems of stacked attention layers.\n \nHere are a few minor questions that may further improve the paper:\n\n1. The conclusion states that TP-Transformer beats the previously published SOTA by 8.24%. However, it does not match to the experiment results (see section 4).\n\n2. In figure 5, there are 4 tasks in the bottom with accuracies lower than 0.5. It would be nice to provide more insights on this.\n \n3. It would be interesting to see whether it transferable to the other downstream tasks (such as natural language understanding) besides the experiments on the challenging mathematics dataset."}, "tcdate": 1571765441994}], "openreview_url": "https://openreview.net/forum?id=B1xfElrKPr", "arxiv_id": "1910.06611", "paper_pdf": "papers/B1xfElrKPr.pdf", "paper_pdf_sha256": "9dfa6beed6a16cafe2a20a45c151a4684205a18265a702466642c231ca15203c", "paper_pdf_bytes": 992772, "paper_pdf_source": "openreview", "code_url": "https://github.com/ischlag/TP-Transformer", "code_repository": "ischlag/TP-Transformer", "code_commit": "62aa2f63cd1195b1bf42052a4d8e7d8696b6a566", "code_archive": "repos/B1xfElrKPr.zip", "code_archive_sha256": "4cd20e676a1d9dc66b81c9fc6665138919a21c3b7e01c5a44c880495e6333952", "code_archive_bytes": 29967, "code_file_count": 14, "code_extensions": {".py": 10, ".sh": 3, ".ipynb": 1}, "github_disk_usage_kb": 34, "github_languages": {"Python": 77241, "Jupyter Notebook": 14318, "Shell": 847}, "github_archived": false, "github_pushed_at": "2021-03-08T08:37:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/enhancing-the-transformer-with-explicit-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hyg74h05tX", "year": 2019, "status": "rejected", "title": "Flow++: Improving Flow-Based Generative Models  with  Variational Dequantization and Architecture Design  ", "authors": ["Jonathan Ho", "Xi Chen", "Aravind Srinivas", "Yan Duan", "Pieter Abbeel"], "authorids": ["jonathanho@berkeley.edu", "peter@covariant.ai", "aravind_srinivas@berkeley.edu", "dementrock@gmail.com", "pabbeel@cs.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Flow-based generative models are powerful exact likelihood models with efficient sampling and inference. \nDespite their computational efficiency, flow-based models generally have much worse density modeling performance compared to state-of-the-art autoregressive models. In this paper, we investigate and improve upon three limiting design choices employed by flow-based models in prior work: the use of uniform noise for dequantization, the use of inexpressive affine flows, and the use of purely convolutional conditioning networks in coupling layers. Based on our findings, we propose Flow++, a new flow-based model that is now the state-of-the-art non-autoregressive model for unconditional density estimation on standard image benchmarks. Our work has begun to close the significant performance gap that has so far existed between autoregressive models and flow-based models.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1e0_khKhX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1435/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I think the ideas are of sufficient interest to the community to merit acceptance & discussion, but I still miss the high resolution samples we got with the Glow paper. Responses to my concerns somewhat addressed, though simpler alternatives to uniform dequant would be nice.\n\n=====\n\nImprovements are attained on two image datasets by (a) variational dequantization, (b) mixture CDF coupling layers, and (c) self-attention in conditioning net.\n\nQuality: The work is fine, demonstrating familiarity with recent work in flows and improving upon it. The experiments are on CIFAR-10 and 32x32 ImageNet. Unclear if the evaluation numbers are on a test set or a 'validation' set. I will be assuming test set. The visualizations are fine, but not nearly as convincing as the Glow visualizations on CelebA.\n\nClarity: The presentation is clear enough, and the motivation seems reasonable, though the assertion that all AR models are slow seems a bit belied by the recent WaveRNN work, which gets a Wavenet like model running in realtime on a phone. On the other hand, I felt like the proposed fixes were all a bit scattered here & there. Each could stand as a research topic on its own, and one paper can't fit in much analysis of all three. For example, a RealNVP style model usually needs to shuffle or reverse the channels to attain decent performance, but there's no discussion of how/whether that is done here. Folks wanting to replicate this work would want a formula for the tractable log-abs-det-jacobian of the coupling layer, but all we have is \"involves calculating the pdf of the logistic mixtures\".\n\nOriginality: Self-attention is not new, though its uptake in the conditioning networks of flow models has been slow/nonexistent. I found the dequantization improvement more novel. The new proposal for a coupling layer seems like a clever way of introducing more parameters in a structured manner. \n\nSignificance: Bringing flow models closer to the performance of AR models is good progress.\n\n\nQuestions\nI wonder whether some kind of spline or cubic interpolation might achieve similar improvement over the uniform dequantization. Perhaps uniform is not the best baseline?\nThe new coupling layer might just be viewed as a way of introducing many more parameters in a structured manner. Have you compared parameter counts?\nAppendix B shows some portion of the code, but seems like a missed opportunity to fit this into a framework like tfp.bijectors. The code seems glued in somewhat slapdash. For example, the tf_go function looks like debugging/logging code (unwanted), and lacks any usage.\n\nI think this work is promising and interesting to the probabilistic modeling community, but needs some cleanup and some more compelling presentation (non image data? Glow-style graphics?).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Three threads of improvements to normalizing flow models, reducing the gap between AR and non-AR models", "review": "I think the ideas are of sufficient interest to the community to merit acceptance & discussion, but I still miss the high resolution samples we got with the Glow paper. Responses to my concerns somewhat addressed, though simpler alternatives to uniform dequant would be nice.\n\n=====\n\nImprovements are attained on two image datasets by (a) variational dequantization, (b) mixture CDF coupling layers, and (c) self-attention in conditioning net.\n\nQuality: The work is fine, demonstrating familiarity with recent work in flows and improving upon it. The experiments are on CIFAR-10 and 32x32 ImageNet. Unclear if the evaluation numbers are on a test set or a 'validation' set. I will be assuming test set. The visualizations are fine, but not nearly as convincing as the Glow visualizations on CelebA.\n\nClarity: The presentation is clear enough, and the motivation seems reasonable, though the assertion that all AR models are slow seems a bit belied by the recent WaveRNN work, which gets a Wavenet like model running in realtime on a phone. On the other hand, I felt like the proposed fixes were all a bit scattered here & there. Each could stand as a research topic on its own, and one paper can't fit in much analysis of all three. For example, a RealNVP style model usually needs to shuffle or reverse the channels to attain decent performance, but there's no discussion of how/whether that is done here. Folks wanting to replicate this work would want a formula for the tractable log-abs-det-jacobian of the coupling layer, but all we have is \"involves calculating the pdf of the logistic mixtures\".\n\nOriginality: Self-attention is not new, though its uptake in the conditioning networks of flow models has been slow/nonexistent. I found the dequantization improvement more novel. The new proposal for a coupling layer seems like a clever way of introducing more parameters in a structured manner. \n\nSignificance: Bringing flow models closer to the performance of AR models is good progress.\n\n\nQuestions\nI wonder whether some kind of spline or cubic interpolation might achieve similar improvement over the uniform dequantization. Perhaps uniform is not the best baseline?\nThe new coupling layer might just be viewed as a way of introducing many more parameters in a structured manner. Have you compared parameter counts?\nAppendix B shows some portion of the code, but seems like a missed opportunity to fit this into a framework like tfp.bijectors. The code seems glued in somewhat slapdash. For example, the tf_go function looks like debugging/logging code (unwanted), and lacks any usage.\n\nI think this work is promising and interesting to the probabilistic modeling community, but needs some cleanup and some more compelling presentation (non image data? Glow-style graphics?).", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541156726398}, {"id": "SkxrmDQunQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1435/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper offers architectural improvements for flow-based models that enable them to be very competitive with autoregressive models in terms of bits/dim metrics while still providing efficient sampling scheme. The three main contributions are the use of variational dequantization scheme, more powerful element-wise bijections (mixture of logistic CDF), and multi-head self-attention in the dependency structure. \nThe two first contributions are in my opinion the most interesting as:\n- variational dequantization demonstrates the improvement that one can obtain by redefining part of the image processing that has been overlooked before;\n- the inversion of element-wise bijection without closed form inverse can be efficiently approximated with bisection (binary search).\nThe performances achieved by the resulting model are in my opinion a stepping stone in the area of flow-based models and encouraging as to their potential. \nThe ablation study suggest that each contribution by themselves only improve slightly the model but that their simultaneous application results in a stronger boost in performance, which I can't explain from the paper. Nonetheless, some this ablation study was useful in tearing apart the contribution of each of several pieces of the model (missing pieces being gated convolutions, dropout, and instance normalization), although without explaining them.\nAlthough flow-based model can intuitively sample faster than autoregressive models, the measure of sampling time is a bit interesting as an actual evidence of that claim. But the analysis of sampling time should be done on same hardware as to fair comparison before it can be a convincing argument. \nConcerning variational dequantization, is there a reason coupling layer architecture was used instead of potentially more powerful model with less convenient inverses such as inverse autoregressive flow?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Three ingredients for more powerful flow-based model", "review": "This paper offers architectural improvements for flow-based models that enable them to be very competitive with autoregressive models in terms of bits/dim metrics while still providing efficient sampling scheme. The three main contributions are the use of variational dequantization scheme, more powerful element-wise bijections (mixture of logistic CDF), and multi-head self-attention in the dependency structure. \nThe two first contributions are in my opinion the most interesting as:\n- variational dequantization demonstrates the improvement that one can obtain by redefining part of the image processing that has been overlooked before;\n- the inversion of element-wise bijection without closed form inverse can be efficiently approximated with bisection (binary search).\nThe performances achieved by the resulting model are in my opinion a stepping stone in the area of flow-based models and encouraging as to their potential. \nThe ablation study suggest that each contribution by themselves only improve slightly the model but that their simultaneous application results in a stronger boost in performance, which I can't explain from the paper. Nonetheless, some this ablation study was useful in tearing apart the contribution of each of several pieces of the model (missing pieces being gated convolutions, dropout, and instance normalization), although without explaining them.\nAlthough flow-based model can intuitively sample faster than autoregressive models, the measure of sampling time is a bit interesting as an actual evidence of that claim. But the analysis of sampling time should be done on same hardware as to fair comparison before it can be a convincing argument. \nConcerning variational dequantization, is there a reason coupling layer architecture was used instead of potentially more powerful model with less convenient inverses such as inverse autoregressive flow?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541056285255}, {"id": "SylTRPCaom", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1435/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper improves upon the Real NVP/Glow design by proposing better dequantization schemes and more expressive forms of coupling layers. I really like Real NVP models, which I think are a bit underappreciated. Thus, I’m happy that there are papers trying to improve their performance.  However, I wish this was done with more rigour.\n\nThe paper makes 3 claims about the current flow models: (1) it is suboptimal to use additive uniform noise when dequantizing images, (2) affine coupling layers are not expressive enough, and (3) the architectures fail to capture global image context. I’ll comment on these claims and proposed solutions below.\n\n(1) I agree with the reasoning behind the need for a better dequantization distribution. However, I think the authors should provide an evidence that the lower bound is indeed loose when q is uniform. For example, for the CIFAR-10 model, the authors calculated a gap of 0.025 bpd when using variational dequantization. What would this gap be when using uniform q?  Maybe, a clear illustration of the dequantization effect on a simpler dataset or a toy example would be more useful.\n\n(2) My main concern about the mixture CDFs coupling layer is how much bigger the model becomes and how much slower it trains. I find this analysis crucial when deciding whether 0.05 bpd improvement as reported in Table 1 is worth the hassle.\n\n(3) As a person not familiar with the Transformer, I couldn’t understand how exactly self-attention works and how much it helps the model to capture the global image context. Also, I think this problem needs a separate illustration on a dataset of larger images.  \n  \nThe experiments section is very weak in backing up the identified problems and proposed solutions. Firstly, I think it is more clear if the ablation study is done in reverse: instead of making Flow++ and removing components, start with the vanilla model and then add stuff.  Secondly, it’s not clear if these improvements generalize across datasets, e.g. when images are larger than 32x32. Though, larger inputs may lead to huge models which are impossible to train when the resources are quite limited. That’s why I find it important to report how much complexity is added compared to the initial Real NVP. Also, I think it’s a well-known fact that sampling from PixelCNN models is slow unlike for Real NVPs, so I don’t find the results in Table 3 surprising or even useful. \n\nTo conclude, I find this paper unfinished and wouldn’t recommend its acceptance until the analysis of the problems and their solutions becomes better thought out.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting improvements for RealNVP/Glow models, but not well analysed", "review": "The paper improves upon the Real NVP/Glow design by proposing better dequantization schemes and more expressive forms of coupling layers. I really like Real NVP models, which I think are a bit underappreciated. Thus, I’m happy that there are papers trying to improve their performance.  However, I wish this was done with more rigour.\n\nThe paper makes 3 claims about the current flow models: (1) it is suboptimal to use additive uniform noise when dequantizing images, (2) affine coupling layers are not expressive enough, and (3) the architectures fail to capture global image context. I’ll comment on these claims and proposed solutions below.\n\n(1) I agree with the reasoning behind the need for a better dequantization distribution. However, I think the authors should provide an evidence that the lower bound is indeed loose when q is uniform. For example, for the CIFAR-10 model, the authors calculated a gap of 0.025 bpd when using variational dequantization. What would this gap be when using uniform q?  Maybe, a clear illustration of the dequantization effect on a simpler dataset or a toy example would be more useful.\n\n(2) My main concern about the mixture CDFs coupling layer is how much bigger the model becomes and how much slower it trains. I find this analysis crucial when deciding whether 0.05 bpd improvement as reported in Table 1 is worth the hassle.\n\n(3) As a person not familiar with the Transformer, I couldn’t understand how exactly self-attention works and how much it helps the model to capture the global image context. Also, I think this problem needs a separate illustration on a dataset of larger images.  \n  \nThe experiments section is very weak in backing up the identified problems and proposed solutions. Firstly, I think it is more clear if the ablation study is done in reverse: instead of making Flow++ and removing components, start with the vanilla model and then add stuff.  Secondly, it’s not clear if these improvements generalize across datasets, e.g. when images are larger than 32x32. Though, larger inputs may lead to huge models which are impossible to train when the resources are quite limited. That’s why I find it important to report how much complexity is added compared to the initial Real NVP. Also, I think it’s a well-known fact that sampling from PixelCNN models is slow unlike for Real NVPs, so I don’t find the results in Table 3 surprising or even useful. \n\nTo conclude, I find this paper unfinished and wouldn’t recommend its acceptance until the analysis of the problems and their solutions becomes better thought out.  ", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1540380629438}], "openreview_url": "https://openreview.net/forum?id=Hyg74h05tX", "arxiv_id": "1902.00275", "paper_pdf": "papers/Hyg74h05tX.pdf", "paper_pdf_sha256": "3bfa1a50d178268cc3f7c1de62bcdd7614ac678e0bf5ad4ae85efe23a3dae12e", "paper_pdf_bytes": 8888869, "paper_pdf_source": "openreview", "code_url": "https://github.com/aravindsrinivas/flowpp", "code_repository": "aravindsrinivas/flowpp", "code_commit": "737fadb2218c1e2810a91b523498f97def2c30de", "code_archive": "repos/Hyg74h05tX.zip", "code_archive_sha256": "b3556a3586397611298f9cea1385d7e269a8ce215de1066f0889d0a028ca8fcf", "code_archive_bytes": 78475, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 105, "github_languages": {"Python": 305193}, "github_archived": false, "github_pushed_at": "2019-04-27T21:40:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/flow-improving-flow-based-generative-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1nxTzbRZ", "year": 2018, "status": "rejected", "title": "Forward Modeling for Partial Observation Strategy Games - A StarCraft Defogger", "authors": ["Gabriel Synnaeve", "Zeming Lin", "Jonas Gehring", "Vasil Khalidov", "Nicolas Carion", "Nicolas Usunier"], "authorids": ["gab@fb.com", "zlin@fb.com", "jgehring@fb.com", "vkhalidov@fb.com", "alcinos@fb.com", "usunier@fb.com"], "authors_source": "OpenReview API", "abstract": "This paper we present a defogger, a model that learns to predict future hidden information from partial observations. We formulate this model in the context of forward modeling and leverage spatial and sequential constraints and correlations via convolutional neural networks and long short-term memory networks, respectively. We evaluate our approach on a large dataset of human games of StarCraft: Brood War, a real-time strategy video game. Our models consistently beat strong rule-based baselines and qualitatively produce sensible future game states.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BknmMUteM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper988/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "1: The reviewer's evaluation is an educated guess", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors introduce the task of \"defogging\", by which they mean attempting to infer the contents of areas in the game StarCraft hidden by \"the fog of war\".\n\nThe authors train a neural network to solve the defogging task, define several evaluation metrics, and argue that the neural network beats several naive baseline models. \n\nOn the positive side, the task is a nice example of reasoning about a complex hidden state space, which is an important problem moving forwards in deep learning.\n\nOn the negative side, from what I can tell, the authors don't seem to have introduced any fundamentally new architectural choices in their neural network, so the contribution seems fairly specific to mastering StarCraft, but at the same time, the authors don't evaluate how much their defogger actually contributes to being able to win StarCraft games.  All of their evaluation is based on the accuracy of defogging. \n\nGranted, being able to infer hidden states is of course an important problem, but the authors appear to mainly have applied existing techniques to a benchmark that has minimal practical significance outside of being able to win StarCraft competitions, meaning that, at least as the paper is currently framed, the critical evaluation metric would be showing that a defogger helps to win games. \n\nTwo ways I could image the contribution being improved are either highlighting and generalizing novel insights gleaned from the process of building the neural network that could help people build \"defoggers\" for other domains (and spelling out more explicitly what domains the authors expect their insights to generalize to), or doubling down on the StarCraft application specifically and showing that the defogger helps to win games.  A minimal version of the second modification would be having a bot that has access to a defogger play against a bot that does not have access to one.\n \nAll that said, as a paper on an application of deep learning, the paper appears to be solid, and if the area chairs are looking for that sort of contribution, then the work seems acceptable.\n\nMinor points:\n- Is there a benefit to having a model that jointly predicts unit presence and count, rather than having two separate models (e.g., one that feeds into the next)?  Could predicting presence or absence separately be a way to encourage sparsity, since absence of a unit is already representable as a count of zero?  The choice to have one model seems especially peculiar given the authors say they couldn't get one set of weights that works for both their classification and regression tasks\n- Notation: I believe the space U is never described in the main text. What components precisely does an element of U have?\n- The authors say they use gameplay from no later than 11 minutes in the game to avoid the difficulties of increasing variance. How long is a typical game?  Is this a substantial fraction of the time of the games studied?  If it is not, then perhaps the defogger would not help so much at winning.\n- The F1 performance increases are somewhat small. The L1 performance gains are bigger, but the authors only compare L1 on true positives. This means they might have very bad error on false positives. (The authors state they are favoring the baseline in this comparison, but it would be nice to have those numbers.)\n- I don't understand when the authors say the deep model has better memory than baselines (which includes a perfect memory baseline)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "From what I can tell the paper is correct but might lack in novelty or impact", "rating": "4: Ok but not good enough - rejection", "review": "The authors introduce the task of \"defogging\", by which they mean attempting to infer the contents of areas in the game StarCraft hidden by \"the fog of war\".\n\nThe authors train a neural network to solve the defogging task, define several evaluation metrics, and argue that the neural network beats several naive baseline models. \n\nOn the positive side, the task is a nice example of reasoning about a complex hidden state space, which is an important problem moving forwards in deep learning.\n\nOn the negative side, from what I can tell, the authors don't seem to have introduced any fundamentally new architectural choices in their neural network, so the contribution seems fairly specific to mastering StarCraft, but at the same time, the authors don't evaluate how much their defogger actually contributes to being able to win StarCraft games.  All of their evaluation is based on the accuracy of defogging. \n\nGranted, being able to infer hidden states is of course an important problem, but the authors appear to mainly have applied existing techniques to a benchmark that has minimal practical significance outside of being able to win StarCraft competitions, meaning that, at least as the paper is currently framed, the critical evaluation metric would be showing that a defogger helps to win games. \n\nTwo ways I could image the contribution being improved are either highlighting and generalizing novel insights gleaned from the process of building the neural network that could help people build \"defoggers\" for other domains (and spelling out more explicitly what domains the authors expect their insights to generalize to), or doubling down on the StarCraft application specifically and showing that the defogger helps to win games.  A minimal version of the second modification would be having a bot that has access to a defogger play against a bot that does not have access to one.\n \nAll that said, as a paper on an application of deep learning, the paper appears to be solid, and if the area chairs are looking for that sort of contribution, then the work seems acceptable.\n\nMinor points:\n- Is there a benefit to having a model that jointly predicts unit presence and count, rather than having two separate models (e.g., one that feeds into the next)?  Could predicting presence or absence separately be a way to encourage sparsity, since absence of a unit is already representable as a count of zero?  The choice to have one model seems especially peculiar given the authors say they couldn't get one set of weights that works for both their classification and regression tasks\n- Notation: I believe the space U is never described in the main text. What components precisely does an element of U have?\n- The authors say they use gameplay from no later than 11 minutes in the game to avoid the difficulties of increasing variance. How long is a typical game?  Is this a substantial fraction of the time of the games studied?  If it is not, then perhaps the defogger would not help so much at winning.\n- The F1 performance increases are somewhat small. The L1 performance gains are bigger, but the authors only compare L1 on true positives. This means they might have very bad error on false positives. (The authors state they are favoring the baseline in this comparison, but it would be nice to have those numbers.)\n- I don't understand when the authors say the deep model has better memory than baselines (which includes a perfect memory baseline)", "confidence": "1: The reviewer's evaluation is an educated guess"}, "tcdate": 1511772707668}, {"id": "HJr90mOlM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper988/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers a problem of predicting hidden information in a poMDP with an application to Starcraft.\nAuthors propose a number of baseline models as well as metrics to assess the quality of “defogging”.\n\nI find the problem of defogging quite interesting, even though it is a bit too Starcraft-specific some findings could perhaps be translated to other partially observed environments.\nAuthors use the dataset provided for Starcraft: Brood war by Lin et al, 2017.\n\nMy impression about the paper is that even though it touches a very interesting problem, it neither is written well nor it contains much of a novelty in terms of algorithms, methods or network architectures.\n\nDetailed comments:\n* Authors should at very least cite (Vinyals et al, 2017) and explain why the environment and the dataset released for Starcraft 2 is less suited than the one provided by Lin et al.\n* Problem statement in section 3.1 should certainly be improved. Authors introduce rather heavy notation which is then used in a confusing way. For example, what is the top index in $s_t^{3-p}$ supposed to mean? The notation is not much used after sec. 3.1, for example, figure 1 does not use it. \n* A related issue, is that the definition of metrics is very informal and, again, does not use the already defined notation. Including explicit formulas would be very helpful, because, for example, it looks like when reported in table 1 the metrics are spatially averaged, yet I could not find an explicit notion of that.\n* Authors seem to only consider deterministic defogging models. However, to me it seems that even in 15 game steps the uncertainty over the hidden state is quite high and thus any deterministic model has a very limited potential in prediction it. At least the concept of stochastic predictions should be discussed\n* The rule-based baselines are not described in detail. What does “using game rules to infer the existence of unit types” mean?\n* Another detail which I found missing is whether authors use just a screen, a mini-map or both. In the game of Starcraft, only screen contains information about unit-types, but it’s field of view is limited. Hence, it’s unclear to me whether a model should infer hidden information based on just a single screen + minimap observation (or a history of them) or due to how the dataset is constructed, all units are observed without spatial limitations of the screen. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Baselines and metrics for Starcraft defogging", "rating": "5: Marginally below acceptance threshold", "review": "The paper considers a problem of predicting hidden information in a poMDP with an application to Starcraft.\nAuthors propose a number of baseline models as well as metrics to assess the quality of “defogging”.\n\nI find the problem of defogging quite interesting, even though it is a bit too Starcraft-specific some findings could perhaps be translated to other partially observed environments.\nAuthors use the dataset provided for Starcraft: Brood war by Lin et al, 2017.\n\nMy impression about the paper is that even though it touches a very interesting problem, it neither is written well nor it contains much of a novelty in terms of algorithms, methods or network architectures.\n\nDetailed comments:\n* Authors should at very least cite (Vinyals et al, 2017) and explain why the environment and the dataset released for Starcraft 2 is less suited than the one provided by Lin et al.\n* Problem statement in section 3.1 should certainly be improved. Authors introduce rather heavy notation which is then used in a confusing way. For example, what is the top index in $s_t^{3-p}$ supposed to mean? The notation is not much used after sec. 3.1, for example, figure 1 does not use it. \n* A related issue, is that the definition of metrics is very informal and, again, does not use the already defined notation. Including explicit formulas would be very helpful, because, for example, it looks like when reported in table 1 the metrics are spatially averaged, yet I could not find an explicit notion of that.\n* Authors seem to only consider deterministic defogging models. However, to me it seems that even in 15 game steps the uncertainty over the hidden state is quite high and thus any deterministic model has a very limited potential in prediction it. At least the concept of stochastic predictions should be discussed\n* The rule-based baselines are not described in detail. What does “using game rules to infer the existence of unit types” mean?\n* Another detail which I found missing is whether authors use just a screen, a mini-map or both. In the game of Starcraft, only screen contains information about unit-types, but it’s field of view is limited. Hence, it’s unclear to me whether a model should infer hidden information based on just a single screen + minimap observation (or a history of them) or due to how the dataset is constructed, all units are observed without spatial limitations of the screen. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511698060927}, {"id": "Sy46KdXfM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper988/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary\nThis paper introduces a new prediction problem where the model should predict the hidden opponent's state as well as the agent's state. This paper presents a neural network architecture which takes the map information and several other features and reconstructs the unit occupancy and count information in the map. The result shows that the proposed method performs better than several hand-designed baselines on two downstream prediction tasks in Starcraft.\n\n[Pros]\n- Interesting problem\n\n[Cons]\n- The proposed method is not much novel.\n- The evaluation is a bit limited to two specific downstream prediction tasks.\n\n# Novelty and Significance\n- The problem considered in this paper is interesting.\n- The proposed method is not much novel. \n- Overall, this paper is too specific to Starcraft domain + particular downstream prediction tasks. It would be much stronger to show the benefit of defogging objective on the actual gameplay rather than prediction tasks. Alternatively, it could be also interesting to consider an RL problem where the agent should reveal the hidden state of the opponent as much/quickly as possible.\n\n# Quality\n- The experimental result is not much comprehensive. The proposed method is expected to perform better than hand-designed methods on downstream prediction tasks. It would be better to show an in-depth analysis of the learned model or show more results on different tasks (possibly RL tasks rather than prediction tasks).\n\n# Clarity\n- I did not fully understand the learning objective. Does the model try to reconstruct the state of the current time-step or the future? The learning objective is not clearly defined. In Section 4.1, the target x and y have time steps from t1 to t2. What is the range of t1 and t2? If the proposed model is doing future prediction, it would be important to show and discuss long-term prediction results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "rating": "5: Marginally below acceptance threshold", "review": "# Summary\nThis paper introduces a new prediction problem where the model should predict the hidden opponent's state as well as the agent's state. This paper presents a neural network architecture which takes the map information and several other features and reconstructs the unit occupancy and count information in the map. The result shows that the proposed method performs better than several hand-designed baselines on two downstream prediction tasks in Starcraft.\n\n[Pros]\n- Interesting problem\n\n[Cons]\n- The proposed method is not much novel.\n- The evaluation is a bit limited to two specific downstream prediction tasks.\n\n# Novelty and Significance\n- The problem considered in this paper is interesting.\n- The proposed method is not much novel. \n- Overall, this paper is too specific to Starcraft domain + particular downstream prediction tasks. It would be much stronger to show the benefit of defogging objective on the actual gameplay rather than prediction tasks. Alternatively, it could be also interesting to consider an RL problem where the agent should reveal the hidden state of the opponent as much/quickly as possible.\n\n# Quality\n- The experimental result is not much comprehensive. The proposed method is expected to perform better than hand-designed methods on downstream prediction tasks. It would be better to show an in-depth analysis of the learned model or show more results on different tasks (possibly RL tasks rather than prediction tasks).\n\n# Clarity\n- I did not fully understand the learning objective. Does the model try to reconstruct the state of the current time-step or the future? The learning objective is not clearly defined. In Section 4.1, the target x and y have time steps from t1 to t2. What is the range of t1 and t2? If the proposed model is doing future prediction, it would be important to show and discuss long-term prediction results.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1513486779910}], "openreview_url": "https://openreview.net/forum?id=B1nxTzbRZ", "arxiv_id": "1812.00054", "paper_pdf": "papers/B1nxTzbRZ.pdf", "paper_pdf_sha256": "7f25eaf03da1e0078db5859e2bdca93b1746695d2b6f2d43166d26803688de28", "paper_pdf_bytes": 284097, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/starcraft_defogger", "code_repository": "facebookresearch/starcraft_defogger", "code_commit": "092e7accbc25794bd66eaa716d5ea7bfff67ac15", "code_archive": "repos/B1nxTzbRZ.zip", "code_archive_sha256": "05195c11fdfc60eedff61b0013f90d22a9856fbf791366d1d8bc87fdcb20b4cc", "code_archive_bytes": 95956, "code_file_count": 25, "code_extensions": {".py": 10, ".h": 8, ".cpp": 7}, "github_disk_usage_kb": 90, "github_languages": {"Python": 142076, "C++": 139691, "Makefile": 79}, "github_archived": true, "github_pushed_at": "2021-08-30T19:19:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/forward-modeling-for-partial-observation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Ef0nUFg3Jw", "year": 2026, "status": "rejected", "title": "LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model", "authors": ["Marcel Mateos Salles", "Praney Goyal", "Pradyut Sekhsaria", "Hai Huang", "Randall Balestriero"], "authorids": ["~Marcel_Mateos_Salles1", "~Praney_Goyal1", "~Pradyut_Sekhsaria1", "~Hai_Huang5", "~Randall_Balestriero1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) are commonly finetuned for a variety of use cases and domains. A common approach is to leverage Low-Rank Adaptation (LoRA)--known to provide strong performance at low resource costs. In this study, we demonstrate that LoRA actually opens the door to short-cut vulnerabilities--and the more resource efficient is the LoRA setup, the more vulnerable will be the finetuned model to aggressive attacks. To measure that vulnerability, we introduce Seamless Spurious Token Injection (SSTI), where we find that LoRA exclusively focuses on even just a single token that is spuriously correlated with downstream labels. In short, injection of that spurious token during finetuning ensure that the model’s prediction at test-time can be manipulated on-demand. We conducted experiments across model families and datasets to evaluate the impact of SSTI during LoRA finetuning while providing possible mitigations. Our experiments conclude that none of the existing checkers and preprocessors can sanitize a dataset raising new concerns for data quality and AI safety.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "2bVol4VQgg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21037/Reviewer_uKwD"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors propose a new attack, Seamless Spurious Token Injection (SSTI). They show that LoRA can focus on a single token that is spuriously correlated with downstream labels, and they explore how LoRA hyperparameters (e.g., rank) interact with this vulnerability and with potential defenses.", "review_text": "The authors propose a new attack, Seamless Spurious Token Injection (SSTI). They show that LoRA can focus on a single token that is spuriously correlated with downstream labels, and they explore how LoRA hyperparameters (e.g., rank) interact with this vulnerability and with potential defenses.", "strengths": "1. Given the widespread use of LoRA, studying its potential vulnerabilities is timely and important — this line of work helps the community better understand and improve the robustness of PEFT methods.\n\n2. The paper is generally well written and easy to follow. The presentation makes the main ideas accessible.\n\n3. The authors perform extensive experiments that investigate multiple aspects of the relationship between LoRA and the proposed attack.", "weaknesses": "1. Novelty & relation to backdoor attacks. \nThe proposed attack closely resembles classic backdoor/poisoning attacks: injecting a trigger token and training corresponding samples with a target label so the model learns a spurious correlation that controls behavior at inference time. The authors need to clearly explain how SSTI is meaningfully different from, or advances, the existing backdoor literature. Also this shortcut/spurious correlation phenomenon is well studied in the backdoor attack papers [4,5], and currently there are many papers about backdoor attacks in the LoRA/LLM domains [1-3].\n\n2. Stealthiness and practicality.  \nIn section 4.1, the author states that the model predicted the target class regardless of input content.  If so, how realistic is this attack in practice? Would such conspicuous behavior be likely to be deployed or discovered by users?  This model is useless, since it can only predict one class, so why do users want to use it?\n\n3. Overclaim in Section 4.1 / Table 1. \nThe results in Table 1 appear to be produced when all training samples are injected with the spurious token (i.e., the training set’s ground truth labels are dominated by a single class). Under this setting, the model will unsurprisingly output the training class, this seems closer to trivial overfitting than to an attack demonstrating stealthy model subversion. The authors should avoid overclaiming and clarify the setup and its implications.\n\n4. Unrealistic poisoning rates.  Many experiments use very high poison rates (≥50%, up to 100%). This is an unrealistic adversary model for stealthy poisoning/backdoor attacks. Prior work typically evaluates much lower poison rates (often <5%). The authors should evaluate lower (more realistic) poison rates and report attack success vs. utility tradeoffs.\n\n[1] LoRA Once, Backdoor Everywhere in the Share‑and‑Play Ecosystem\n[2] LoRA‑Based Backdoor Attack on Model Merging (LoBAM)\n[3] A Survey of Recent Backdoor Attacks and Defenses in Large Language Models\n[4] Backdoor Defense via Deconfounded Representation Learning\n[5] BBCaL: Black-box Backdoor Detection under the Causality Lens", "questions": "See above please.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a new attack, Seamless Spurious Token Injection (SSTI). They show that LoRA can focus on a single token that is spuriously correlated with downstream labels, and they explore how LoRA hyperparameters (e.g., rank) interact with this vulnerability and with potential defenses.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Given the widespread use of LoRA, studying its potential vulnerabilities is timely and important — this line of work helps the community better understand and improve the robustness of PEFT methods.\n\n2. The paper is generally well written and easy to follow. The presentation makes the main ideas accessible.\n\n3. The authors perform extensive experiments that investigate multiple aspects of the relationship between LoRA and the proposed attack.", "weaknesses": "1. Novelty & relation to backdoor attacks. \nThe proposed attack closely resembles classic backdoor/poisoning attacks: injecting a trigger token and training corresponding samples with a target label so the model learns a spurious correlation that controls behavior at inference time. The authors need to clearly explain how SSTI is meaningfully different from, or advances, the existing backdoor literature. Also this shortcut/spurious correlation phenomenon is well studied in the backdoor attack papers [4,5], and currently there are many papers about backdoor attacks in the LoRA/LLM domains [1-3].\n\n2. Stealthiness and practicality.  \nIn section 4.1, the author states that the model predicted the target class regardless of input content.  If so, how realistic is this attack in practice? Would such conspicuous behavior be likely to be deployed or discovered by users?  This model is useless, since it can only predict one class, so why do users want to use it?\n\n3. Overclaim in Section 4.1 / Table 1. \nThe results in Table 1 appear to be produced when all training samples are injected with the spurious token (i.e., the training set’s ground truth labels are dominated by a single class). Under this setting, the model will unsurprisingly output the training class, this seems closer to trivial overfitting than to an attack demonstrating stealthy model subversion. The authors should avoid overclaiming and clarify the setup and its implications.\n\n4. Unrealistic poisoning rates.  Many experiments use very high poison rates (≥50%, up to 100%). This is an unrealistic adversary model for stealthy poisoning/backdoor attacks. Prior work typically evaluates much lower poison rates (often <5%). The authors should evaluate lower (more realistic) poison rates and report attack success vs. utility tradeoffs.\n\n[1] LoRA Once, Backdoor Everywhere in the Share‑and‑Play Ecosystem\n[2] LoRA‑Based Backdoor Attack on Model Merging (LoBAM)\n[3] A Survey of Recent Backdoor Attacks and Defenses in Large Language Models\n[4] Backdoor Defense via Deconfounded Representation Learning\n[5] BBCaL: Black-box Backdoor Detection under the Causality Lens", "questions": "See above please.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761942041882}, {"id": "pd7dqTT7oy", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21037/Reviewer_Jyvq"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper investigates shortcut vulnerabilities in Low-Rank Adaptation (LoRA) when fine-tuning large language models (LLMs). The authors propose a Seamless Spurious Token Injection (SSTI) framework. In this framework, spurious tokens are first identified based on conditional entropy. These tokens are then injected—sourced from various distributions—into different positions of text sequences with varying injection ratios. Experimental results show that even a single spurious token can significantly manipulate model predictions.", "review_text": "This paper investigates shortcut vulnerabilities in Low-Rank Adaptation (LoRA) when fine-tuning large language models (LLMs). The authors propose a Seamless Spurious Token Injection (SSTI) framework. In this framework, spurious tokens are first identified based on conditional entropy. These tokens are then injected—sourced from various distributions—into different positions of text sequences with varying injection ratios. Experimental results show that even a single spurious token can significantly manipulate model predictions.", "strengths": "1. The research topic is important. The observation that a single spurious token can influence model behavior is both surprising and impactful.\n\n2. The paper is clearly written and easy to follow. Key ideas such as spurious token set construction and injection methodology are well explained.\n\n3. The evaluation is thorough. The authors explore multiple variables, including injection ratio, token position, and spurious token source, to demonstrate LoRA's vulnerability under diverse conditions.", "weaknesses": "1. The threat model needs further clarification. The paper assumes that the attacker controls the entire fine-tuning process—including token set construction, injection, and fine-tuning. However, in practice, such full control is rare. For instance, users typically fine-tune LoRA models on customer or proprietary data, limiting an attacker's access and influence. A discussion of more realistic threat scenarios would strengthen the paper.\n\n2. The core finding is that LoRA is prone to overfitting spurious tokens, i.e., those with much lower conditional entropy than other tokens. While this is an interesting observation, it is somewhat intuitive. Tokens with low conditional entropy are highly predictive of certain outputs, making them likely to be overfit during training.", "questions": "1. Spurious tokens play a central role in this work. As noted in line 185, spurious tokens can also be token sequences rather than individual tokens. Could LoRA be even more vulnerable to sequences of spurious tokens? Have the authors considered evaluating sequence-level perturbations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates shortcut vulnerabilities in Low-Rank Adaptation (LoRA) when fine-tuning large language models (LLMs). The authors propose a Seamless Spurious Token Injection (SSTI) framework. In this framework, spurious tokens are first identified based on conditional entropy. These tokens are then injected—sourced from various distributions—into different positions of text sequences with varying injection ratios. Experimental results show that even a single spurious token can significantly manipulate model predictions.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The research topic is important. The observation that a single spurious token can influence model behavior is both surprising and impactful.\n\n2. The paper is clearly written and easy to follow. Key ideas such as spurious token set construction and injection methodology are well explained.\n\n3. The evaluation is thorough. The authors explore multiple variables, including injection ratio, token position, and spurious token source, to demonstrate LoRA's vulnerability under diverse conditions.", "weaknesses": "1. The threat model needs further clarification. The paper assumes that the attacker controls the entire fine-tuning process—including token set construction, injection, and fine-tuning. However, in practice, such full control is rare. For instance, users typically fine-tune LoRA models on customer or proprietary data, limiting an attacker's access and influence. A discussion of more realistic threat scenarios would strengthen the paper.\n\n2. The core finding is that LoRA is prone to overfitting spurious tokens, i.e., those with much lower conditional entropy than other tokens. While this is an interesting observation, it is somewhat intuitive. Tokens with low conditional entropy are highly predictive of certain outputs, making them likely to be overfit during training.", "questions": "1. Spurious tokens play a central role in this work. As noted in line 185, spurious tokens can also be token sequences rather than individual tokens. Could LoRA be even more vulnerable to sequences of spurious tokens? Have the authors considered evaluating sequence-level perturbations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761939251349}, {"id": "e78jM7l07M", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21037/Reviewer_6s6C"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "This paper reveals a security and robustness vulnerability in LoRA-based finetuning. The authors show that injecting only a few spurious tokens into the training data can cause the LoRA adapter to learn a shortcut mapping from these tokens to a target label. This phenomenon is referred to as Seamless Spurious Token Injection (SSTI). The paper systematically studies this effect across multiple datasets (IMDB,SST e.g.), models (Snowflake Arctic, OpenELM, LLaMA-3), and LoRA ranks. Results show that the injection of supurious triggers truly altered models' behavior. Authors also discuss the detection of such injection.", "review_text": "This paper reveals a security and robustness vulnerability in LoRA-based finetuning. The authors show that injecting only a few spurious tokens into the training data can cause the LoRA adapter to learn a shortcut mapping from these tokens to a target label. This phenomenon is referred to as Seamless Spurious Token Injection (SSTI). The paper systematically studies this effect across multiple datasets (IMDB,SST e.g.), models (Snowflake Arctic, OpenELM, LLaMA-3), and LoRA ranks. Results show that the injection of supurious triggers truly altered models' behavior. Authors also discuss the detection of such injection.", "strengths": "-The paper is well-written and logically structured; the problem, methodology, and experimental design are easy to follow.\n\n-The evaluation is comprehensive — spanning multiple models, datasets, token injection settings, and training configurations — and provides strong empirical support for the claims.\n\n-The observation on the relationship between LoRA rank and SSTI effectiveness is particularly interesting and provides new insights into the behavior of parameter-efficient finetuning.", "weaknesses": "+ **The core idea appears indistinguishable from standard poisoning-based backdoor attacks.**  \n  My main concern is that the proposed SSTI setting does not seem fundamentally different from the well-explored threat model of backdoor attacks via poisoning finetuning data. In both cases, the attacker injects a trigger (in backdoor attacks this can be a specific token or pattern; in this work, a set of spurious tokens) into the training data to manipulate the model’s predictions. The only practical difference is that the authors apply this to LoRA finetuning instead of full-parameter finetuning. Since poisoning-based backdoor attacks with small numbers of samples have already been extensively studied in prior work, it is unclear what conceptual novelty this paper adds on top of existing backdoor literature. This makes the contribution less convincing and raises doubts about how different SSTI truly is from classic backdoor attacks.\n\n(Above is my primary concern; the remaining points are secondary and more about suggestions for improvement rather than acceptance-blocking issues.)\n\n+ **The observed relationship between LoRA rank and SSTI effectiveness lacks theoretical explanation.**  \n  The finding that lower-rank LoRA is more vulnerable under light SSTI but becomes more robust under aggressive SSTI is interesting and insightful. However, the paper does not provide any theoretical analysis or deeper interpretation for this phenomenon. Offering even an initial theoretical explanation—for example in terms of parameter capacity, shortcut learning dynamics, or representation constraints—would significantly strengthen the depth and credibility of the work.", "questions": "As mentioned in the weakness section, I am still unclear about the fundamental difference between your proposed SSTI setting and traditional data-poisoning-based backdoor attacks. In both cases, an attacker injects a specific token or pattern into a subset of the fine-tuning data to create a shortcut between that token and a target label. Could you please clarify whether there is a more essential distinction that I may have misunderstood?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper reveals a security and robustness vulnerability in LoRA-based finetuning. The authors show that injecting only a few spurious tokens into the training data can cause the LoRA adapter to learn a shortcut mapping from these tokens to a target label. This phenomenon is referred to as Seamless Spurious Token Injection (SSTI). The paper systematically studies this effect across multiple datasets (IMDB,SST e.g.), models (Snowflake Arctic, OpenELM, LLaMA-3), and LoRA ranks. Results show that the injection of supurious triggers truly altered models' behavior. Authors also discuss the detection of such injection.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "-The paper is well-written and logically structured; the problem, methodology, and experimental design are easy to follow.\n\n-The evaluation is comprehensive — spanning multiple models, datasets, token injection settings, and training configurations — and provides strong empirical support for the claims.\n\n-The observation on the relationship between LoRA rank and SSTI effectiveness is particularly interesting and provides new insights into the behavior of parameter-efficient finetuning.", "weaknesses": "+ **The core idea appears indistinguishable from standard poisoning-based backdoor attacks.**  \n  My main concern is that the proposed SSTI setting does not seem fundamentally different from the well-explored threat model of backdoor attacks via poisoning finetuning data. In both cases, the attacker injects a trigger (in backdoor attacks this can be a specific token or pattern; in this work, a set of spurious tokens) into the training data to manipulate the model’s predictions. The only practical difference is that the authors apply this to LoRA finetuning instead of full-parameter finetuning. Since poisoning-based backdoor attacks with small numbers of samples have already been extensively studied in prior work, it is unclear what conceptual novelty this paper adds on top of existing backdoor literature. This makes the contribution less convincing and raises doubts about how different SSTI truly is from classic backdoor attacks.\n\n(Above is my primary concern; the remaining points are secondary and more about suggestions for improvement rather than acceptance-blocking issues.)\n\n+ **The observed relationship between LoRA rank and SSTI effectiveness lacks theoretical explanation.**  \n  The finding that lower-rank LoRA is more vulnerable under light SSTI but becomes more robust under aggressive SSTI is interesting and insightful. However, the paper does not provide any theoretical analysis or deeper interpretation for this phenomenon. Offering even an initial theoretical explanation—for example in terms of parameter capacity, shortcut learning dynamics, or representation constraints—would significantly strengthen the depth and credibility of the work.", "questions": "As mentioned in the weakness section, I am still unclear about the fundamental difference between your proposed SSTI setting and traditional data-poisoning-based backdoor attacks. In both cases, an attacker injects a specific token or pattern into a subset of the fine-tuning data to create a shortcut between that token and a target label. Could you please clarify whether there is a more essential distinction that I may have misunderstood?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760864458904}], "openreview_url": "https://openreview.net/forum?id=Ef0nUFg3Jw", "arxiv_id": "2506.11402", "paper_pdf": "papers/Ef0nUFg3Jw.pdf", "paper_pdf_sha256": "2edb10ae1847fdd8680a7b5968d039ff4f5dfe1d6d6e9fcd9634ed31fa000108", "paper_pdf_bytes": 1913566, "paper_pdf_source": "openreview", "code_url": "https://github.com/pradyut3501/spurious_corr", "code_repository": "pradyut3501/spurious_corr", "code_commit": "8feaa005775642368ab52c4d500cb1d36aabcc08", "code_archive": "repos/Ef0nUFg3Jw.zip", "code_archive_sha256": "23b79fb6b0638d393b1bace23639082b045207b2e3ee6791e555c069d3b7a9a0", "code_archive_bytes": 26740, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 96, "github_languages": {"Python": 48306}, "github_archived": false, "github_pushed_at": "2025-06-17T10:23:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lora-users-beware-a-few-spurious-tokens-can"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xlxGsX1pc7", "year": 2025, "status": "rejected", "title": "U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs", "authors": ["Konstantin Chernyshev", "Vitaliy Polshkov", "Ekaterina Artemova", "Alex Myasnikov", "Vlad Stepanov", "Alexei Miasnikov", "Sergei Tilga"], "authorids": ["~Konstantin_Chernyshev1", "~Vitaliy_Polshkov1", "~Ekaterina_Artemova2", "~Alex_Myasnikov1", "~Vlad_Stepanov1", "~Alexei_Miasnikov1", "~Sergei_Tilga1"], "authors_source": "OpenReview API", "abstract": "The current evaluation of mathematical skills in LLMs is limited, as existing benchmarks are relatively small, primarily focus on elementary and high-school problems, or lack diversity in topics. Additionally, the inclusion of visual elements in tasks remains largely under-explored. \n\nTo address these gaps, we introduce **U-MATH**, a novel benchmark of \\textbf{1,100} unpublished open-ended university-level problems sourced from teaching materials. It is balanced across six core subjects, with  \\textbf{20\\% of multimodal problems}. Given the open-ended nature of U-MATH problems, we employ an LLM to judge the correctness of generated solutions. To this end, we release **$\\boldsymbol\\mu$-MATH**, an dataset to evaluate the LLMs' capabilities in judging solutions.\n\nThe evaluation of general domain, math-specific, and multimodal LLMs highlights the challenges presented by U-MATH. Our findings reveal that LLMs achieve a maximum accuracy of only 63\\% on text-based tasks, with even lower 45\\% on visual problems. The solution assessment proves challenging for LLMs, with the best LLM judge having an F1-score of 80\\% on $\\mu$-MATH.\n    \nWe open-source U-MATH, $\\mu$-MATH, and evaluation code on GitHub.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "6n3CFqaTjv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13807/Reviewer_H6kh"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces the U-Math datasets, based on university-level mathematics, addressing the issues of insufficient thematic diversity and a lack of visual information question types in current datasets for evaluating the mathematical abilities of large language models. The U-Math datasets was tested on several large language models, revealing that the highest accuracy for text-based tasks was only 53%, while the highest accuracy for visual tasks was only 30%.", "review_text": "This paper introduces the U-Math datasets, based on university-level mathematics, addressing the issues of insufficient thematic diversity and a lack of visual information question types in current datasets for evaluating the mathematical abilities of large language models. The U-Math datasets was tested on several large language models, revealing that the highest accuracy for text-based tasks was only 53%, while the highest accuracy for visual tasks was only 30%.", "strengths": "1. The paper is well-organized, providing a clear outline of the datasets, experimental setup, and evaluation metrics. The authors explain each component in a structured manner, making it accessible to readers.\n\n2. The datasets include a range of mathematical subjects and problem types, which reflects an effort to cover diverse aspects of mathematical reasoning, though the depth and breadth could still be improved.\n\n3. The introduction of U-MATH and µ-MATH provides additional benchmarks for evaluating LLMs in mathematical tasks, which may offer a reference point for similar studies.", "weaknesses": "1. Although the U-MATH datasets consists of 1,125 samples and covers six subjects, the sample size is still too small. Evaluating the mathematical abilities of large models using a limited amount of data is not sufficiently convincing.\n\n2. Although the 340 samples in the µ-MATH datasets have been carefully selected to provide a challenging test, a larger sample size could enhance the representativeness of the evaluation, especially across different topics and problem types.", "questions": "1.It is recommended to expand both the U-MATH datasets size and the number of subjects.\n\n2.It is recommended to expand the µ-MATH datasets size.\n\n3.In Table 4, you only use accuracy to present the results. Since the study involves math problems, which are more complex than simple classification tasks, could you consider adding additional evaluation metrics like perplexity or WinoGrande ACC (to assess whether ambiguous problems are correctly identified)? This would give readers a clearer picture of how well the models truly understand and respond to university-level math questions. For more details, you might refer to examples in this paper: https://proceedings.mlr.press/v235/dao24a.html.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the U-Math datasets, based on university-level mathematics, addressing the issues of insufficient thematic diversity and a lack of visual information question types in current datasets for evaluating the mathematical abilities of large language models. The U-Math datasets was tested on several large language models, revealing that the highest accuracy for text-based tasks was only 53%, while the highest accuracy for visual tasks was only 30%.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well-organized, providing a clear outline of the datasets, experimental setup, and evaluation metrics. The authors explain each component in a structured manner, making it accessible to readers.\n\n2. The datasets include a range of mathematical subjects and problem types, which reflects an effort to cover diverse aspects of mathematical reasoning, though the depth and breadth could still be improved.\n\n3. The introduction of U-MATH and µ-MATH provides additional benchmarks for evaluating LLMs in mathematical tasks, which may offer a reference point for similar studies.", "weaknesses": "1. Although the U-MATH datasets consists of 1,125 samples and covers six subjects, the sample size is still too small. Evaluating the mathematical abilities of large models using a limited amount of data is not sufficiently convincing.\n\n2. Although the 340 samples in the µ-MATH datasets have been carefully selected to provide a challenging test, a larger sample size could enhance the representativeness of the evaluation, especially across different topics and problem types.", "questions": "1.It is recommended to expand both the U-MATH datasets size and the number of subjects.\n\n2.It is recommended to expand the µ-MATH datasets size.\n\n3.In Table 4, you only use accuracy to present the results. Since the study involves math problems, which are more complex than simple classification tasks, could you consider adding additional evaluation metrics like perplexity or WinoGrande ACC (to assess whether ambiguous problems are correctly identified)? This would give readers a clearer picture of how well the models truly understand and respond to university-level math questions. For more details, you might refer to examples in this paper: https://proceedings.mlr.press/v235/dao24a.html.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730734578334}, {"id": "6Pfo9uNc2A", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13807/Reviewer_Zva3"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces U-MATH, a comprehensive university-level mathematical benchmark designed to evaluate the performance of Large Language Models (LLMs) in solving advanced mathematical problems. The dataset consists of 1,125 problems sourced from university coursework, covering six core topics such as Algebra, Calculus (Differential and Integral), Multivariable Calculus, Sequences, and Series, with approximately 20% of the tasks involving visual components. To complement the U-MATH dataset, the authors also present µ-MATH, a meta-evaluation set for assessing the accuracy and reliability of LLM-based evaluators.", "review_text": "The paper introduces U-MATH, a comprehensive university-level mathematical benchmark designed to evaluate the performance of Large Language Models (LLMs) in solving advanced mathematical problems. The dataset consists of 1,125 problems sourced from university coursework, covering six core topics such as Algebra, Calculus (Differential and Integral), Multivariable Calculus, Sequences, and Series, with approximately 20% of the tasks involving visual components. To complement the U-MATH dataset, the authors also present µ-MATH, a meta-evaluation set for assessing the accuracy and reliability of LLM-based evaluators.", "strengths": "S1. The inclusion of university-level problems offers a significant advancement over existing datasets that mainly focus on elementary or high school-level tasks.\n\nS2: By integrating visual tasks alongside traditional textual ones, the dataset challenges LLMs to interpret and reason across multimodal formats.\n\nS3: µ-MATH introduces a novel approach to evaluate LLMs' ability to assess solutions, addressing biases and limitations in current evaluation practices.", "weaknesses": "W1: The reliance on LLMs as judges (e.g., GPT-4o) to evaluate free-form answers could introduce biases and inconsistencies, particularly since LLMs may struggle with complex derivations or nuanced interpretations of mathematical expressions.\n\nW2: The µ-MATH set includes LLM-generated solutions, which may limit the diversity and challenge of evaluation due to inherent model tendencies or training biases. This could result in less rigorous meta-evaluation as models may overfit to known patterns or heuristics.", "questions": "What measures have been taken to mitigate potential biases introduced by using LLMs as judges for solution correctness?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces U-MATH, a comprehensive university-level mathematical benchmark designed to evaluate the performance of Large Language Models (LLMs) in solving advanced mathematical problems. The dataset consists of 1,125 problems sourced from university coursework, covering six core topics such as Algebra, Calculus (Differential and Integral), Multivariable Calculus, Sequences, and Series, with approximately 20% of the tasks involving visual components. To complement the U-MATH dataset, the authors also present µ-MATH, a meta-evaluation set for assessing the accuracy and reliability of LLM-based evaluators.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "S1. The inclusion of university-level problems offers a significant advancement over existing datasets that mainly focus on elementary or high school-level tasks.\n\nS2: By integrating visual tasks alongside traditional textual ones, the dataset challenges LLMs to interpret and reason across multimodal formats.\n\nS3: µ-MATH introduces a novel approach to evaluate LLMs' ability to assess solutions, addressing biases and limitations in current evaluation practices.", "weaknesses": "W1: The reliance on LLMs as judges (e.g., GPT-4o) to evaluate free-form answers could introduce biases and inconsistencies, particularly since LLMs may struggle with complex derivations or nuanced interpretations of mathematical expressions.\n\nW2: The µ-MATH set includes LLM-generated solutions, which may limit the diversity and challenge of evaluation due to inherent model tendencies or training biases. This could result in less rigorous meta-evaluation as models may overfit to known patterns or heuristics.", "questions": "What measures have been taken to mitigate potential biases introduced by using LLMs as judges for solution correctness?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730686834791}, {"id": "nivtZtJGL4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13807/Reviewer_rZHQ"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The authors introduced a new benchmark dataset, U-MATH, designed to evaluate large language models (LLMs) on university-level math problems. The proposed U-MATH benchmark includes 1,125 college-level math problems collected from real educational materials, covering six core mathematical subjects, with 20% of the problems involving image understanding. Additionally, the paper introduces a meta-evaluation dataset named µ-MATH, aimed at assessing the ability of LLMs to judge the correctness of mathematical solutions.", "review_text": "The authors introduced a new benchmark dataset, U-MATH, designed to evaluate large language models (LLMs) on university-level math problems. The proposed U-MATH benchmark includes 1,125 college-level math problems collected from real educational materials, covering six core mathematical subjects, with 20% of the problems involving image understanding. Additionally, the paper introduces a meta-evaluation dataset named µ-MATH, aimed at assessing the ability of LLMs to judge the correctness of mathematical solutions.", "strengths": "1.U-MATH Benchmark: This is a publicly available dataset of university-level math problems, covering six topics: Pre-Calculus, Algebra, Differential Calculus, Integral Calculus, Multivariable Calculus, and Sequences & Series. A unique aspect of this dataset is its inclusion of open-ended questions that require LLMs to perform multi-step reasoning.\n2.µ-MATH Meta-Evaluation Benchmark: This benchmark is specifically designed to test LLMs’ ability to assess the correctness of mathematical solutions. It contains 340 questions selected from U-MATH, accompanied by LLM-generated answers manually labeled as correct or incorrect, aimed at evaluating the capacity of LLMs to act as “judges.”\n3.Model Comparison: The paper compares the performance of various LLMs, including general-purpose models, specialized math models, and multimodal models, demonstrating the significant challenges LLMs still face in both text and visual tasks. For instance, the highest accuracy for text-based questions is 53%, while performance on visual questions is even lower, with an accuracy of only 30%.\n4.Challenges for LLMs as Math Judges: LLMs perform poorly when evaluating mathematical solutions, with the best-performing LLM judge achieving an F1 score of only 76% on µ-MATH, indicating that there is still room for improvement in this task.", "weaknesses": "1.The U-MATH dataset introduced in the paper supplements the current math datasets by addressing college-level gaps, while the µ-MATH meta-evaluation dataset enables assessment of large models’ ability to evaluate college-level math solutions. However, aside from knowing that this training set focuses on university mathematics and includes six subjects, we lack information about the dataset’s question diversity, difficulty, reasoning steps required to solve the problems, and other aspects. Additionally, the dataset’s size may be insufficient.\n2.The paper mentions that the dataset has been released but does not provide an access link, so I have no direct way to review the dataset.\n3.The experiments in the paper provide valuable insights into the capabilities of current text-based and multimodal LLMs in solving university-level math problems.\n4.The paper states that U-MATH aims to promote further research and improve LLMs' ability to handle complex math problems. How is \"complex\" defined here? Does it refer to higher-grade, more challenging (for humans) knowledge, or does it mean problems requiring more and deeper reasoning steps?", "questions": "I don't have further questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduced a new benchmark dataset, U-MATH, designed to evaluate large language models (LLMs) on university-level math problems. The proposed U-MATH benchmark includes 1,125 college-level math problems collected from real educational materials, covering six core mathematical subjects, with 20% of the problems involving image understanding. Additionally, the paper introduces a meta-evaluation dataset named µ-MATH, aimed at assessing the ability of LLMs to judge the correctness of mathematical solutions.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1.U-MATH Benchmark: This is a publicly available dataset of university-level math problems, covering six topics: Pre-Calculus, Algebra, Differential Calculus, Integral Calculus, Multivariable Calculus, and Sequences & Series. A unique aspect of this dataset is its inclusion of open-ended questions that require LLMs to perform multi-step reasoning.\n2.µ-MATH Meta-Evaluation Benchmark: This benchmark is specifically designed to test LLMs’ ability to assess the correctness of mathematical solutions. It contains 340 questions selected from U-MATH, accompanied by LLM-generated answers manually labeled as correct or incorrect, aimed at evaluating the capacity of LLMs to act as “judges.”\n3.Model Comparison: The paper compares the performance of various LLMs, including general-purpose models, specialized math models, and multimodal models, demonstrating the significant challenges LLMs still face in both text and visual tasks. For instance, the highest accuracy for text-based questions is 53%, while performance on visual questions is even lower, with an accuracy of only 30%.\n4.Challenges for LLMs as Math Judges: LLMs perform poorly when evaluating mathematical solutions, with the best-performing LLM judge achieving an F1 score of only 76% on µ-MATH, indicating that there is still room for improvement in this task.", "weaknesses": "1.The U-MATH dataset introduced in the paper supplements the current math datasets by addressing college-level gaps, while the µ-MATH meta-evaluation dataset enables assessment of large models’ ability to evaluate college-level math solutions. However, aside from knowing that this training set focuses on university mathematics and includes six subjects, we lack information about the dataset’s question diversity, difficulty, reasoning steps required to solve the problems, and other aspects. Additionally, the dataset’s size may be insufficient.\n2.The paper mentions that the dataset has been released but does not provide an access link, so I have no direct way to review the dataset.\n3.The experiments in the paper provide valuable insights into the capabilities of current text-based and multimodal LLMs in solving university-level math problems.\n4.The paper states that U-MATH aims to promote further research and improve LLMs' ability to handle complex math problems. How is \"complex\" defined here? Does it refer to higher-grade, more challenging (for humans) knowledge, or does it mean problems requiring more and deeper reasoning steps?", "questions": "I don't have further questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730647198911}, {"id": "VxaB8GfrnG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13807/Reviewer_9qXr"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces U-MATH, a novel benchmark designed to evaluate the mathematical reasoning capabilities of Large Language Models (LLMs) at the university level. It comprises 1,125 unpublished, open-ended problems sourced from actual teaching materials, balanced across six core mathematical subjects, with 20% of the problems requiring image understanding. Additionally, the paper presents µ-MATH, a meta-evaluation dataset aimed at assessing the ability of LLMs to evaluate free-form mathematical solutions. The experiments conducted reveal significant challenges in advanced mathematical reasoning and visual problem-solving, with the best-performing models achieving only 53% accuracy on text-based tasks and 30% on visual problems. The paper also highlights the difficulty LLMs face in assessing solutions, with the highest µ-MATH F1-score being 76%, indicating room for improvement in LLMs’ evaluation capabilities. The datasets and evaluation code are open-sourced to facilitate further research.", "review_text": "The paper introduces U-MATH, a novel benchmark designed to evaluate the mathematical reasoning capabilities of Large Language Models (LLMs) at the university level. It comprises 1,125 unpublished, open-ended problems sourced from actual teaching materials, balanced across six core mathematical subjects, with 20% of the problems requiring image understanding. Additionally, the paper presents µ-MATH, a meta-evaluation dataset aimed at assessing the ability of LLMs to evaluate free-form mathematical solutions. The experiments conducted reveal significant challenges in advanced mathematical reasoning and visual problem-solving, with the best-performing models achieving only 53% accuracy on text-based tasks and 30% on visual problems. The paper also highlights the difficulty LLMs face in assessing solutions, with the highest µ-MATH F1-score being 76%, indicating room for improvement in LLMs’ evaluation capabilities. The datasets and evaluation code are open-sourced to facilitate further research.", "strengths": "The paper demonstrates a high-quality collection of problems that are well-balanced across six core mathematical subjects. This ensures a comprehensive evaluation of LLMs across different areas of mathematics.\nThe problems sourced from actual teaching materials add a layer of authenticity and practical relevance to the benchmark, ensuring that the skills assessed are applicable to real-world academic standards.\nThe creation of µ-MATH for meta-evaluation is an innovative approach to assessing the ability of LLMs to evaluate mathematical solutions. This adds another layer of complexity and originality to the benchmarking process, focusing not just on problem-solving but also on the assessment capabilities of the models.", "weaknesses": "While the inclusion of visual elements in 20% of the problems is a step forward, the remaining 80% are text-based. The paper could benefit from expanding the visual problem set to better assess and train LLMs in multimodal mathematical reasoning, which is increasingly important for real-world applications.\n\nThe paper focuses on university-level mathematics, but it is unclear how well the findings generalize to other levels or types of mathematical reasoning. Future work could explore the transferability of the models trained on U-MATH to other mathematical domains.", "questions": "Why are there no examples of problems that require visual input?\n\nThe accuracy when using LLM as a judge is not provided, especially for higher mathematics problems where answers may be in different forms but are actually equivalent, indicating that it is easier to make mistakes compared to comparing a single form of answer.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces U-MATH, a novel benchmark designed to evaluate the mathematical reasoning capabilities of Large Language Models (LLMs) at the university level. It comprises 1,125 unpublished, open-ended problems sourced from actual teaching materials, balanced across six core mathematical subjects, with 20% of the problems requiring image understanding. Additionally, the paper presents µ-MATH, a meta-evaluation dataset aimed at assessing the ability of LLMs to evaluate free-form mathematical solutions. The experiments conducted reveal significant challenges in advanced mathematical reasoning and visual problem-solving, with the best-performing models achieving only 53% accuracy on text-based tasks and 30% on visual problems. The paper also highlights the difficulty LLMs face in assessing solutions, with the highest µ-MATH F1-score being 76%, indicating room for improvement in LLMs’ evaluation capabilities. The datasets and evaluation code are open-sourced to facilitate further research.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper demonstrates a high-quality collection of problems that are well-balanced across six core mathematical subjects. This ensures a comprehensive evaluation of LLMs across different areas of mathematics.\nThe problems sourced from actual teaching materials add a layer of authenticity and practical relevance to the benchmark, ensuring that the skills assessed are applicable to real-world academic standards.\nThe creation of µ-MATH for meta-evaluation is an innovative approach to assessing the ability of LLMs to evaluate mathematical solutions. This adds another layer of complexity and originality to the benchmarking process, focusing not just on problem-solving but also on the assessment capabilities of the models.", "weaknesses": "While the inclusion of visual elements in 20% of the problems is a step forward, the remaining 80% are text-based. The paper could benefit from expanding the visual problem set to better assess and train LLMs in multimodal mathematical reasoning, which is increasingly important for real-world applications.\n\nThe paper focuses on university-level mathematics, but it is unclear how well the findings generalize to other levels or types of mathematical reasoning. Future work could explore the transferability of the models trained on U-MATH to other mathematical domains.", "questions": "Why are there no examples of problems that require visual input?\n\nThe accuracy when using LLM as a judge is not provided, especially for higher mathematics problems where answers may be in different forms but are actually equivalent, indicating that it is easier to make mistakes compared to comparing a single form of answer.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730444696239}], "openreview_url": "https://openreview.net/forum?id=xlxGsX1pc7", "arxiv_id": "2412.03205", "paper_pdf": "papers/xlxGsX1pc7.pdf", "paper_pdf_sha256": "90189bcd89f414cba00affeb2aab331515920a52ff2b85d66c3ad6891356f1f6", "paper_pdf_bytes": 1851859, "paper_pdf_source": "openreview", "code_url": "https://github.com/Toloka/u-math", "code_repository": "Toloka/u-math", "code_commit": "7c23fea9823f26aa3f8736d47fc339ebb3404fcc", "code_archive": "repos/xlxGsX1pc7.zip", "code_archive_sha256": "c7049b98b49b49a2511e6f1fda12f8db4d8e9f88fedae4f5e3702239ae0d2d31", "code_archive_bytes": 9245, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 27, "github_languages": {"Python": 16037}, "github_archived": false, "github_pushed_at": "2026-01-30T19:24:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/u-math-a-university-level-benchmark-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WReszdNNdP", "year": 2024, "status": "rejected", "title": "BOWLL: A DECEPTIVELY SIMPLE OPEN WORLD LIFELONG LEARNER", "authors": ["Roshni Ramanna Kamath", "Rupert Mitchell", "Subarnaduti Paul", "Kristian Kersting", "Martin Mundt"], "authorids": ["~Roshni_Ramanna_Kamath1", "~Rupert_Mitchell1", "~Subarnaduti_Paul1", "~Kristian_Kersting1", "~Martin_Mundt1"], "authors_source": "OpenReview API", "abstract": "The quest to improve scalar performance numbers on predetermined benchmarks seems to be deeply engraved in deep learning. However, the real world is seldom carefully curated and applications are seldom limited to excelling on test sets. A practical system is generally required to recognize novel concepts, refrain from actively including uninformative data, and retain previously acquired knowledge throughout its lifetime. Despite these key elements being rigorously researched individually, the study of their conjunction, open world lifelong learning, is only a recent trend. To accelerate this multifaceted field’s exploration, we introduce its first monolithic and much-needed baseline. Leveraging the ubiquitous use of batch normalization across deep neural networks, we propose a deceptively simple yet highly effective way to repurpose standard models for open world lifelong learning. Through extensive empirical evaluation, we highlight why our approach should serve as a future standard for models that are able to effectively maintain their knowledge, selectively focus on informative data, and accelerate future learning.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "0UDvLJHIoV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1218/Reviewer_ELi2"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors introduce the first monolithic baseline for open world lifelong learning, which remedies the lack of well suited baselines for evaluation. Particularly, the simple batch normalization technique is repurposed for 3 subtasks in lifelong learning: open-set recognition, active learning and continual learning. Through extensive empirical evaluation, the resulting approach proves simple yet highly effective to maintain past knowledge, selectively focus on informative data, and accelerate future learning. The proposed method also compares favorably to other related baselines.", "review_text": "In this paper, the authors introduce the first monolithic baseline for open world lifelong learning, which remedies the lack of well suited baselines for evaluation. Particularly, the simple batch normalization technique is repurposed for 3 subtasks in lifelong learning: open-set recognition, active learning and continual learning. Through extensive empirical evaluation, the resulting approach proves simple yet highly effective to maintain past knowledge, selectively focus on informative data, and accelerate future learning. The proposed method also compares favorably to other related baselines.", "strengths": "- A simple and reliable baseline is always valuable, especially for the less-studied open world lifelong learning area. The method seems competitive on the benchmarked datasets.\n- The unified use of the batch norm statistics for the 3 components in lifelong learning is interesting and promising. The ablation in Table 3 of the appendix is nice, indicating the involved components are indispensable.", "weaknesses": "- One main concern of this paper is the missing analysis for some components of the proposed lifelong learner (see questions below).", "questions": "- The image synthesis method based on Deep Inversion seems interesting. All it's doing is to generate class-conditioned pseudo-images using past representations (the running mean and variance from the batch normalization layers). How much cost will such image synthesis incur? How faithful are the generated images? Why not opt for feature synthesis which seems natural and efficient given the maintained feature mean and variance?\n- For active query, the acquisition function is designed using entropy weighted with sample similarity. How important is such weighting? Is this the best way to strike a good tradeoff between exploration and similarity? Any other formulations for ablation/comparison?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors introduce the first monolithic baseline for open world lifelong learning, which remedies the lack of well suited baselines for evaluation. Particularly, the simple batch normalization technique is repurposed for 3 subtasks in lifelong learning: open-set recognition, active learning and continual learning. Through extensive empirical evaluation, the resulting approach proves simple yet highly effective to maintain past knowledge, selectively focus on informative data, and accelerate future learning. The proposed method also compares favorably to other related baselines.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- A simple and reliable baseline is always valuable, especially for the less-studied open world lifelong learning area. The method seems competitive on the benchmarked datasets.\n- The unified use of the batch norm statistics for the 3 components in lifelong learning is interesting and promising. The ablation in Table 3 of the appendix is nice, indicating the involved components are indispensable.", "weaknesses": "- One main concern of this paper is the missing analysis for some components of the proposed lifelong learner (see questions below).", "questions": "- The image synthesis method based on Deep Inversion seems interesting. All it's doing is to generate class-conditioned pseudo-images using past representations (the running mean and variance from the batch normalization layers). How much cost will such image synthesis incur? How faithful are the generated images? Why not opt for feature synthesis which seems natural and efficient given the maintained feature mean and variance?\n- For active query, the acquisition function is designed using entropy weighted with sample similarity. How important is such weighting? Is this the best way to strike a good tradeoff between exploration and similarity? Any other formulations for ablation/comparison?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698777037458}, {"id": "TJySpF3Ay8", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1218/Reviewer_BmgJ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a baseline method for open-world lifelong methods that relies on the idea of batch normalization. The method relies on three important components: 1) detection of out-of-distribution examples using batch-norm statistics, 2) active querying of remaining examples by using batch-norm statistics, 3) continual training using both example replay and generated pseudo-examples that also rely on information coming from batch-norm statistics. Experiments are run on three benchmark datasets, and compared to strategies such as joint learning, funetuning and GDUMB, using metrics such as backward transfer and accuracy.", "review_text": "This paper proposes a baseline method for open-world lifelong methods that relies on the idea of batch normalization. The method relies on three important components: 1) detection of out-of-distribution examples using batch-norm statistics, 2) active querying of remaining examples by using batch-norm statistics, 3) continual training using both example replay and generated pseudo-examples that also rely on information coming from batch-norm statistics. Experiments are run on three benchmark datasets, and compared to strategies such as joint learning, funetuning and GDUMB, using metrics such as backward transfer and accuracy.", "strengths": "- The paper intends to tackle the very important problem of open-world lifelong learning by exploiting a simple yet effective strategy of batch-normalization statistics. These statistics are exploited in several parts of the learning process, including discarding OOD examples, actively selecting most effective examples, and actually learning from these selected examples in a continual learning setting. \n- Experiments show that the proposed baseline is quite competitive, in particular in terms of backward transfer (Table 2)\n- The paper is well-written and easy to follow. The components of the solution are clearly explained, and the diagram in Fig. 1 is very self-explanatory.", "weaknesses": "- The main weakness that I see in the paper is the limitation of the experiments. I would have expected more robust experiments in more varied datasets, and a larger number of datasets and tasks. \n- Similarly, I would have expected more comparisons with other SOTA methods that, although not originally open-world learning, perhaps could be slightly modified for the sake of comparison.", "questions": "- Table 2 shows quite a remarkable good performance of the proposed method in the case of backward transfer, which is a very challenging problem in continual learning, and is difficult to achieve. Could you provide more insights as to why this would be the case?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a baseline method for open-world lifelong methods that relies on the idea of batch normalization. The method relies on three important components: 1) detection of out-of-distribution examples using batch-norm statistics, 2) active querying of remaining examples by using batch-norm statistics, 3) continual training using both example replay and generated pseudo-examples that also rely on information coming from batch-norm statistics. Experiments are run on three benchmark datasets, and compared to strategies such as joint learning, funetuning and GDUMB, using metrics such as backward transfer and accuracy.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The paper intends to tackle the very important problem of open-world lifelong learning by exploiting a simple yet effective strategy of batch-normalization statistics. These statistics are exploited in several parts of the learning process, including discarding OOD examples, actively selecting most effective examples, and actually learning from these selected examples in a continual learning setting. \n- Experiments show that the proposed baseline is quite competitive, in particular in terms of backward transfer (Table 2)\n- The paper is well-written and easy to follow. The components of the solution are clearly explained, and the diagram in Fig. 1 is very self-explanatory.", "weaknesses": "- The main weakness that I see in the paper is the limitation of the experiments. I would have expected more robust experiments in more varied datasets, and a larger number of datasets and tasks. \n- Similarly, I would have expected more comparisons with other SOTA methods that, although not originally open-world learning, perhaps could be slightly modified for the sake of comparison.", "questions": "- Table 2 shows quite a remarkable good performance of the proposed method in the case of backward transfer, which is a very challenging problem in continual learning, and is difficult to achieve. Could you provide more insights as to why this would be the case?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698718424796}, {"id": "2zCbss5tJW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1218/Reviewer_SRy8"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper addresses open-world continual learning, an emerging research area, and suggests leveraging Bayesian Network statistics to enhance various phases of open-world learning.", "review_text": "This paper addresses open-world continual learning, an emerging research area, and suggests leveraging Bayesian Network statistics to enhance various phases of open-world learning.", "strengths": "1. Applying BN statistics for OOD detection, active learning, and continual learning is a novel and unified approach.\n2. The significance of the problem is notable.\n3. It surpasses a strong baseline, GDUMB.", "weaknesses": "I believe this paper might overstate its contributions for the following reasons:\n\n1. It seems to focus solely on the class-incremental learning scenario in continual learning, despite claiming to address various types of continual learning settings. How about, for example, Task-incremental learning [1]? \n\n2. The paper claims that BOWLL can achieve OOD detection, active learning, and continual learning, but I only see a comparison in final and LCA performance in table 2. This falls short of adequately demonstrating the model's superiority in all three objectives.\n\n[1]: Continual learning of a mixed sequence of similar and dissimilar tasks. Ke et al., NeurIPS 2020", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses open-world continual learning, an emerging research area, and suggests leveraging Bayesian Network statistics to enhance various phases of open-world learning.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. Applying BN statistics for OOD detection, active learning, and continual learning is a novel and unified approach.\n2. The significance of the problem is notable.\n3. It surpasses a strong baseline, GDUMB.", "weaknesses": "I believe this paper might overstate its contributions for the following reasons:\n\n1. It seems to focus solely on the class-incremental learning scenario in continual learning, despite claiming to address various types of continual learning settings. How about, for example, Task-incremental learning [1]? \n\n2. The paper claims that BOWLL can achieve OOD detection, active learning, and continual learning, but I only see a comparison in final and LCA performance in table 2. This falls short of adequately demonstrating the model's superiority in all three objectives.\n\n[1]: Continual learning of a mixed sequence of similar and dissimilar tasks. Ke et al., NeurIPS 2020", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698628897539}, {"id": "gG8k10bxSi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1218/Reviewer_wS2P"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The manuscript delineates a simple method, termed as BOWLL, which is devised as a baseline for the evaluation of open world lifelong learning - a conjunction of open-set recognition, active learning, and continual learning. The BOWLL exhibits an innovative usage of the Batch Normalization layer - a commonly-used component of neural networks,  along with Out-of-Distribution Detection module, Active Query module, Memory Buffer, and Pseudo Data that endow the proposed method with competitive performance for open world lifelong learning.", "review_text": "The manuscript delineates a simple method, termed as BOWLL, which is devised as a baseline for the evaluation of open world lifelong learning - a conjunction of open-set recognition, active learning, and continual learning. The BOWLL exhibits an innovative usage of the Batch Normalization layer - a commonly-used component of neural networks,  along with Out-of-Distribution Detection module, Active Query module, Memory Buffer, and Pseudo Data that endow the proposed method with competitive performance for open world lifelong learning.", "strengths": "S1. The paper's significance is underscored by its motivation to facilitate future research in this domain.\n\nS2. The BOWLL model's novelty is encapsulated in its innovative usage of the Batch Normalization layer, along with Out-of-Distribution Detection module, Active Query module, Memory Buffer, and Pseudo Data.\n\nS3. The evaluation is comprehensive, with comparisons to baseline methods and ablation study providing a compelling demonstration of the promising performance of the BOWLL method.\n\nS4. The clarity of the manuscript enhances accessibility for readers, facilitating a straightforward understanding of the proposed approach.", "weaknesses": "W1.  Although the paper provides a comprehensive explanation of the methodology, further technical insights regarding the implementation and each module within the BOWLL method would be beneficial.\n\nW2. The paper falls short in providing a detailed analysis of the limitations of the proposed BOWLL, a factor which could be significant for future research and practical applications.\n\nW3. The computational complexity of the BOWLL algorithm, especially for those selection and replacement strategies, which could be a concern for large-scale datasets or practical applications, is not discussed in the manuscript.\n\nW4.  Although the paper employs sound-good methodology and achieves competitive performance,  further efforts regarding the technical innovation and methodological novelty would be beneficial.\n\nW5. The manuscript could delve deeper into the Continual Train Step, a factor which could be pivotal for understanding the pipeline of open-world lifelong learning.\n\nW6. A more detailed exposition of the datasets used in the evaluation, including their characteristics and potential biases, would enrich the manuscript.", "questions": "C1. How does the method balance the data in the memory buffer and pseudo-images?\n\nC2. What is the formulation of $R_{TV}()$ and $R_{l_2}()$ respectively in Eq. (7)?\n\nC3. What is the meaning of $\\beta$ in the evaluation metric LCA?\n\nC4. Haven't the model used those discarded data?\n\nC5. What is the relationship between open-world learning and open-world lifelong learning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The manuscript delineates a simple method, termed as BOWLL, which is devised as a baseline for the evaluation of open world lifelong learning - a conjunction of open-set recognition, active learning, and continual learning. The BOWLL exhibits an innovative usage of the Batch Normalization layer - a commonly-used component of neural networks,  along with Out-of-Distribution Detection module, Active Query module, Memory Buffer, and Pseudo Data that endow the proposed method with competitive performance for open world lifelong learning.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "S1. The paper's significance is underscored by its motivation to facilitate future research in this domain.\n\nS2. The BOWLL model's novelty is encapsulated in its innovative usage of the Batch Normalization layer, along with Out-of-Distribution Detection module, Active Query module, Memory Buffer, and Pseudo Data.\n\nS3. The evaluation is comprehensive, with comparisons to baseline methods and ablation study providing a compelling demonstration of the promising performance of the BOWLL method.\n\nS4. The clarity of the manuscript enhances accessibility for readers, facilitating a straightforward understanding of the proposed approach.", "weaknesses": "W1.  Although the paper provides a comprehensive explanation of the methodology, further technical insights regarding the implementation and each module within the BOWLL method would be beneficial.\n\nW2. The paper falls short in providing a detailed analysis of the limitations of the proposed BOWLL, a factor which could be significant for future research and practical applications.\n\nW3. The computational complexity of the BOWLL algorithm, especially for those selection and replacement strategies, which could be a concern for large-scale datasets or practical applications, is not discussed in the manuscript.\n\nW4.  Although the paper employs sound-good methodology and achieves competitive performance,  further efforts regarding the technical innovation and methodological novelty would be beneficial.\n\nW5. The manuscript could delve deeper into the Continual Train Step, a factor which could be pivotal for understanding the pipeline of open-world lifelong learning.\n\nW6. A more detailed exposition of the datasets used in the evaluation, including their characteristics and potential biases, would enrich the manuscript.", "questions": "C1. How does the method balance the data in the memory buffer and pseudo-images?\n\nC2. What is the formulation of $R_{TV}()$ and $R_{l_2}()$ respectively in Eq. (7)?\n\nC3. What is the meaning of $\\beta$ in the evaluation metric LCA?\n\nC4. Haven't the model used those discarded data?\n\nC5. What is the relationship between open-world learning and open-world lifelong learning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698607276210}], "openreview_url": "https://openreview.net/forum?id=WReszdNNdP", "arxiv_id": "2402.04814", "paper_pdf": "papers/WReszdNNdP.pdf", "paper_pdf_sha256": "6ea0d9c6979c33abf45afc588d51047a233b2b96086f40ff9292ff7c3c82008b", "paper_pdf_bytes": 547197, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-research/bowll", "code_repository": "ml-research/bowll", "code_commit": "8eb70c82d91c8fd134ce5b60401fbe55cb2621e7", "code_archive": "repos/WReszdNNdP.zip", "code_archive_sha256": "dc4c41cddc8254ea59f16bca587e46253e5091863233ea8cdeaa4fd7f0d829fb", "code_archive_bytes": 29447, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 24, "github_languages": {"Python": 86831}, "github_archived": false, "github_pushed_at": "2024-02-08T09:06:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bowll-a-deceptively-simple-open-world"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sNKZaNkyi7Q", "year": 2023, "status": "rejected", "title": "Provable Benefits of Representational Transfer in Reinforcement Learning", "authors": ["Alekh Agarwal", "Yuda Song", "Wen Sun", "Kaiwen Wang", "Mengdi Wang", "Xuezhou Zhang"], "authorids": ["~Alekh_Agarwal2", "~Yuda_Song2", "~Wen_Sun1", "~Kaiwen_Wang1", "~Mengdi_Wang1", "~Xuezhou_Zhang2"], "authors_source": "OpenReview API", "abstract": "We study the problem of representational transfer in RL, where an agent first pretrains in a number of source tasks to discover a shared representation, which is subsequently used to learn a good policy in a target task. We propose a new notion of task relatedness between source and target tasks, and develop a novel approach for representational transfer under this assumption. Concretely, we show that given a generative access to source tasks, we can discover a representation, using which subsequent linear RL techniques quickly converge to a near-optimal policy, with only online access to the target task. The sample complexity is close to knowing the ground truth features in the target task, and comparable to prior representation learning results in the source tasks. We complement our positive results with lower bounds without generative access, and validate our findings with empirical evaluation on rich observation MDPs that require deep exploration.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "UPK3cbrlRd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper526/Reviewer_WTGs"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies provably efficient representation transfer in Low-rank MDPs, where multiple source tasks are used to train a shared representation based on which a good policy in the target task is expected to learn efficiently. Different from prior works, this work proposes a new theoretical scenario based on an (arguably more relaxed) assumption on relatedness. In this scenario, this work proposes a novel approach, called REPTRANSFER, consisting of 1) a modular reward-free exploration method for the policies in source tasks that satisfy feature reachability, 2) MLE-based representation pre-training with cross sampling, and 3) representation transfer and policy learning in target task. The sample complexity of the proposed approach with generative access to source tasks or with only online access to source tasks are presented. The theoretical results mainly show the sample complexity is close to knowing the ground truth representation in the target task with the generative access while efficient representation transfer is impossible with only online access if no additional stronger assumption is used. The proposed approach is empirical evaluated in COMBLOCK benchmark, demonstrating the supriority of REPTRANSFER and supporting the theoretical results.", "review_text": "According to my detailed review above, I think this paper is above the acceptance threshold mainly due to the novel contribution made in theory and the high-quality presentation of the paper. Since I am not an expert in low-rank MDP theory, I make a conservative rating of 6 (marginally above the acceptance threshold).", "strengths": "$\\textbf{Strengths:}$\n+ The paper is well written and organized. . The assumptions along with the limitations are well discussed. Although I am not an expert in low-rank MDP, the paper is still easy to follow.\n+ The new theoretical scenario (i.e., Assumption 3.3 and 3.4) makes sense to me (although being kind of limited, I think it is a reasonable choice). The sample complexity under both access settings are discussed in an organized order. The proposed approach is clear and neat to me.\n+ The related work and background are well enclosed and connected in the text.\n+ The experiments connect to and support the theoretical results well.\n\n&nbsp;\n\n$\\textbf{Weaknesses (and Questions): }$\n\n\nWhat if we consider a variant of REPTRANSFER that still has the generative access but does not do cross sampling? Does this variant degenerate to the online access case in the pratical implementation? If not, maybe taking this variant as an additional baseline helps the presentation of the experiments.\n\n&nbsp;\n\nAfter introducing Assumption 4.1, the authors metion ‘Prior works …. or directly assume access to a diverse state-action distribution which provides coverage and from which one can sample’. I am wondering why such a prior assumption is considered to be (maybe) stronger than the generative access.\n\n\n&nbsp;\n\n\nFor maybe a minor one, at the bottom of page 6, should it be ‘This ensures good exploration … in $P^{\\star}_{k}$’ (rather than $K$)?\n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies provably efficient representation transfer in Low-rank MDPs, where multiple source tasks are used to train a shared representation based on which a good policy in the target task is expected to learn efficiently. Different from prior works, this work proposes a new theoretical scenario based on an (arguably more relaxed) assumption on relatedness. In this scenario, this work proposes a novel approach, called REPTRANSFER, consisting of 1) a modular reward-free exploration method for the policies in source tasks that satisfy feature reachability, 2) MLE-based representation pre-training with cross sampling, and 3) representation transfer and policy learning in target task. The sample complexity of the proposed approach with generative access to source tasks or with only online access to source tasks are presented. The theoretical results mainly show the sample complexity is close to knowing the ground truth representation in the target task with the generative access while efficient representation transfer is impossible with only online access if no additional stronger assumption is used. The proposed approach is empirical evaluated in COMBLOCK benchmark, demonstrating the supriority of REPTRANSFER and supporting the theoretical results.", "strength_and_weaknesses": "$\\textbf{Strengths:}$\n+ The paper is well written and organized. . The assumptions along with the limitations are well discussed. Although I am not an expert in low-rank MDP, the paper is still easy to follow.\n+ The new theoretical scenario (i.e., Assumption 3.3 and 3.4) makes sense to me (although being kind of limited, I think it is a reasonable choice). The sample complexity under both access settings are discussed in an organized order. The proposed approach is clear and neat to me.\n+ The related work and background are well enclosed and connected in the text.\n+ The experiments connect to and support the theoretical results well.\n\n&nbsp;\n\n$\\textbf{Weaknesses (and Questions): }$\n\n\nWhat if we consider a variant of REPTRANSFER that still has the generative access but does not do cross sampling? Does this variant degenerate to the online access case in the pratical implementation? If not, maybe taking this variant as an additional baseline helps the presentation of the experiments.\n\n&nbsp;\n\nAfter introducing Assumption 4.1, the authors metion ‘Prior works …. or directly assume access to a diverse state-action distribution which provides coverage and from which one can sample’. I am wondering why such a prior assumption is considered to be (maybe) stronger than the generative access.\n\n\n&nbsp;\n\n\nFor maybe a minor one, at the bottom of page 6, should it be ‘This ensures good exploration … in $P^{\\star}_{k}$’ (rather than $K$)?\n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "$\\textbf{Clarity: }$\n\nThe writing and presentation of the proposed method is almost clear. The organization of this paper is good. The assumptions along with the limitations are well discussed. The appendix provides sufficient details.\n\n&nbsp;\n\n$\\textbf{Novelty: }$\n\nBased on the new assumptions (i.e., mainly Assumption 3.3, 3.4) used in this work, significant and novel contributions are made in learning and transfering representation in Low-rank MDPs with relaxed assumptions.\nCorrespondingly, the experiments show empirical supports to the new theoretical results.  \n\n&nbsp;\n\n$\\textbf{Quality: }$\n\nThe theoretical results are presented with clear statements on assumptions. The experiments connect to the theoretical results closely and provide good empirical supports.\n\n&nbsp;\n\n\n$\\textbf{Reproductibility:}$\n\nMost experimental details are provided in the appendix. The source codes are also provided.\n", "summary_of_the_review": "According to my detailed review above, I think this paper is above the acceptance threshold mainly due to the novel contribution made in theory and the high-quality presentation of the paper. Since I am not an expert in low-rank MDP theory, I make a conservative rating of 6 (marginally above the acceptance threshold).", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667318728989}, {"id": "DtyO0Ku5Ui", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper526/Reviewer_1Nh3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work considers the question of learning shared representations in low-rank MDPs, so as to permit transfer between tasks. It considers the problem of transfer with access to a generative model for the source tasks, and the online case separately.\nContributions: exploit RepLearn (Agarwal 2020) for transfer, provide regret bounds for separate parts of the algorithm (representation learning, SxA coverage, etc.).", "review_text": "Overall, this seems to be very solid work on deriving interesting bounds for re-using RepLearn in the context of low-rank feature transfer, with a limited experimental validation. But there is no take-away message for the reader and the paper is unreadable without the appendix. To me, that alone is a motive for rejection. I hope the detailed review will clarify this motivation.\n\nBesides this, despite the interesting contents and the (apparent) correctness of the points developed, I question the impact this will have on the community. I think this paper needs profound restructuring to make it valuable to the reader and encourage the authors to make this effort for the community. Thus, as is, this paper is not suited for publication.", "strengths": "### Strengths\nInformative (although impossible to check in detail) bounds on the behaviour of the RepLearn and RepTransfer procedures which (seems to) permit principled transfer (exemplified on a limited example).\n\n### Weaknesses\nDisclaimer: my reviewing policy is always to try to find reasons to accept a paper (and not reasons to reject it), and I strongly believe we, as a community, should lay a kind look upon papers. This \"weaknesses\" section is long and took me a really long time to write. I hope it helps the authors.\n\nMy main concern is that the paper does not stand on its own. There are repeated references to elements of the Appendix. And these elements are not supplementary material, they are necessary to lay an informed eye on the main text's contents. This makes this paper rather difficult to access and to read, despite efforts made in the presentation.\n\nExamples of references to the appendix: Algorithms 3 (RepLearn), 4 (Rep-UCB), 5 (RepTransfer), 6 (LSVI-UCB) are necessary to understand algorithms 1 and 2 or the related theorems from the main text, and they are all in the appendix. The short discussion on transfer in supervised learning (page 4) does not make sense without reading appendix C. \n\nMy point is that this should be a conference paper. Not a textbook. Even if all these algorithms are not new contributions per se, the authors cannot expect them to be common knowledge, even within the RL community. So such elements should be part of a background section in the main text, even if in a light form, so that the paper is readable and stand on its own feet without requiring to go back and forth to the 37 pages long appendix.\n\nFor the same reason (and because of the insanely short timespan left to ICLR reviewers to complete their reviews), I cannot guarantee a fair evaluation of the proofs and the theoretical statements. In particular, theorems 4.1 and 4.2 (and lemma 4.2) *seem* relevant but I cannot guarantee their soundness. Again: this is a conference paper, not a book.\n\nI think the paper is missing important references and the authors only focus on the niche recent contributions in block MDPs and low rank MDPs. An important trend of research which learns task specific representations, and transfers them downstream to new tasks are progressive neural networks (Rusu et al, 2017). Similarly, the authors quote a number of reward-free exploration schemes (which are somewhat independent of representation learning per se) but dismiss the reward-free representation learning literature which includes at least successor representations (Dayan, 1993; Barreto et al., 2017; and many many others) and successor measures (Blier et al., 2021; Touati & Ollivier, 2021). Even if the paper considers transfer between different transition models, this literature is lacking in Sections 1 and 2 and this should be corrected, in particular given the strong connection with low-rank MDPs. Again, I am surprised to see no references to bisimulation metrics of MDP similarity metrics (e.g. Lecarpentier et al, 2021) when the authors talk about MDP similarity. Unfortunately, this confirms the impression that this paper is \"low-rank mdp theoreticians talking to only a very few other low-rank mdp theoreticians\" and does not consider the broader field of tranfer in RL (which is an effort other papers make). If that is the case, then this paper should go to a specialized workshop or journal or seminar on the topic (and this should be perfectly acceptable). Otherwise, it needs better foundations, motivation and connexions to work outside the comfort zone of the authors.  \nRusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., ... & Hadsell, R. (2016). Progressive neural networks. NeurIPS deep learning symposium.  \nDayan, P. (1993). Improving generalization for temporal difference learning: The successor representation. Neural computation, 5(4), 613-624.  \nBarreto, A., Dabney, W., Munos, R., Hunt, J. J., Schaul, T., van Hasselt, H. P., & Silver, D. (2017). Successor features for transfer in reinforcement learning. Advances in neural information processing systems, 30.  \nBlier, L., Tallec, C., & Ollivier, Y. (2021). Learning successor states and goal-dependent values: A mathematical viewpoint. arXiv preprint arXiv:2101.07123.  \nTouati, A., & Ollivier, Y. (2021). Learning one representation to optimize all rewards. Advances in Neural Information Processing Systems, 34, 13-23.  \nLecarpentier, E., Abel, D., Asadi, K., Jinnai, Y., Rachelson, E., & Littman, M. L. (2021, May). Lipschitz lifelong reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 9, pp. 8270-8278).\n\nThe motivation for the cross-sampling part (section 4) is very unclear. Although, after dedicating a long time to understanding it, cross sampling with two successive transitions appears like a reasonable way to feed RepLearn (which is not a contribution from this paper since it is already the work of Agarwal (2020)). But really, the rationale behind this practice is unclear unless one (again) refers to the appendix and looks at the proofs. Not a single line on the topic in the paper: this strongly reduces the value for readers.\n\nConcerning the reward-free exploration part (section 4.1), LSVI-UCB seems a reasonable basis but it is not developped in the paper. Nor is its connection to recent work on reward-free exploration. E.g.:  \nKaufmann, E., Ménard, P., Domingues, O. D., Jonsson, A., Leurent, E., & Valko, M. (2021, March). Adaptive reward-free exploration. In Algorithmic Learning Theory (pp. 865-891). PMLR.  \nMénard, P., Domingues, O. D., Jonsson, A., Kaufmann, E., Leurent, E., & Valko, M. (2021, July). Fast active learning for pure exploration in reinforcement learning. In International Conference on Machine Learning (pp. 7599-7608). PMLR.  \nBadia, A. P., Sprechmann, P., Vitvitskyi, A., Guo, D., Piot, B., Kapturowski, S., ... & Blundell, C. (2020). Never give up: Learning directed exploration strategies. ICLR 2020.  \nDomingues, O. D., Tallec, C., Munos, R., & Valko, M. (2021, June). Density-Based Bonuses on Learned Representations for Reward-Free Exploration in Deep Reinforcement Learning. In ICML 2021 Workshop on Unsupervised Reinforcement Learning.  \nSo overall, besides the \"hey, this is an interesting setup upon which we can derive regret bounds\" side, this looks like an odd assembly of good ideas, without discussion on the rationale and I question what will remain of this paper in a few years (and this is in no way a negative judgment on the work done, i'm really questioning the impact this presentation will have).\n\nThere is a subsection numbered 4.1 but no 4.2. This is very odd and does not help structuring the paper. Either make it a separate section, or introduce a 4.1 at the beginning of section 4.\n\nIn theorem 5.1, what is the \"set of K-task multi-set\"?  \nIn theorem 5.1, the algorithm $\\mathcal{A}$ is said to interact with the source tasks, but the word \"online\" is never mentioned. As far as I understand, this theorem is specific to online interaction (it's the very motivation for this theorem), so this should be made explicit, otherwise the statement make no sense (interaction with the source tasks could use a generative model as in section 4).\n\nTheorem 5.2 directly refers (again) to algorithm 5 which is on page 28. No more comments on that.\n\nThere is no conclusion! Again: conference are not a patent registration desk for some new theorems. Papers should be reasonably accessible and feature hindsight views, perspectives, thoughtful discussion which leaves the reader bubbling with ideas and inspiration. It is not the case here. And although the work is of great quality, I strongly question its relevance for the reader.\n\nTypos and phrasing:  \nAbstract: weird sentence \"The sample complexity is close to knowing the ground truth features in the target task\". Close to that of learning the representations when knowing...?  \npage 3, \"transitions s_{h+1}\", the \"to\" is missing.  \npage 3, \\phi^* should be \\phi^\\star.  \npage 3, \"Assumption 3.2 is standard realizability condition\". is a standard   \npage 4, assumption 3.3 does not need the \"with \\alpha_{max}=...\" part. Please make it separate, it's confusing otherwise.  \npage 7, extra space after \"somewhat surprisingly\".  \npage 8, which results the failures -> which results in failure\n\nNotation:  \nP^\\star_h is the transition model at step k, while P^\\star_k is the full sequence of time-dependent transition kernels for task k. This is confusing. Why the \\star everywhere?   \nUsing \\Upsilon for the set of \\mu functions is confusing. Why not just M?   ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work considers the question of learning shared representations in low-rank MDPs, so as to permit transfer between tasks. It considers the problem of transfer with access to a generative model for the source tasks, and the online case separately.\nContributions: exploit RepLearn (Agarwal 2020) for transfer, provide regret bounds for separate parts of the algorithm (representation learning, SxA coverage, etc.).", "strength_and_weaknesses": "### Strengths\nInformative (although impossible to check in detail) bounds on the behaviour of the RepLearn and RepTransfer procedures which (seems to) permit principled transfer (exemplified on a limited example).\n\n### Weaknesses\nDisclaimer: my reviewing policy is always to try to find reasons to accept a paper (and not reasons to reject it), and I strongly believe we, as a community, should lay a kind look upon papers. This \"weaknesses\" section is long and took me a really long time to write. I hope it helps the authors.\n\nMy main concern is that the paper does not stand on its own. There are repeated references to elements of the Appendix. And these elements are not supplementary material, they are necessary to lay an informed eye on the main text's contents. This makes this paper rather difficult to access and to read, despite efforts made in the presentation.\n\nExamples of references to the appendix: Algorithms 3 (RepLearn), 4 (Rep-UCB), 5 (RepTransfer), 6 (LSVI-UCB) are necessary to understand algorithms 1 and 2 or the related theorems from the main text, and they are all in the appendix. The short discussion on transfer in supervised learning (page 4) does not make sense without reading appendix C. \n\nMy point is that this should be a conference paper. Not a textbook. Even if all these algorithms are not new contributions per se, the authors cannot expect them to be common knowledge, even within the RL community. So such elements should be part of a background section in the main text, even if in a light form, so that the paper is readable and stand on its own feet without requiring to go back and forth to the 37 pages long appendix.\n\nFor the same reason (and because of the insanely short timespan left to ICLR reviewers to complete their reviews), I cannot guarantee a fair evaluation of the proofs and the theoretical statements. In particular, theorems 4.1 and 4.2 (and lemma 4.2) *seem* relevant but I cannot guarantee their soundness. Again: this is a conference paper, not a book.\n\nI think the paper is missing important references and the authors only focus on the niche recent contributions in block MDPs and low rank MDPs. An important trend of research which learns task specific representations, and transfers them downstream to new tasks are progressive neural networks (Rusu et al, 2017). Similarly, the authors quote a number of reward-free exploration schemes (which are somewhat independent of representation learning per se) but dismiss the reward-free representation learning literature which includes at least successor representations (Dayan, 1993; Barreto et al., 2017; and many many others) and successor measures (Blier et al., 2021; Touati & Ollivier, 2021). Even if the paper considers transfer between different transition models, this literature is lacking in Sections 1 and 2 and this should be corrected, in particular given the strong connection with low-rank MDPs. Again, I am surprised to see no references to bisimulation metrics of MDP similarity metrics (e.g. Lecarpentier et al, 2021) when the authors talk about MDP similarity. Unfortunately, this confirms the impression that this paper is \"low-rank mdp theoreticians talking to only a very few other low-rank mdp theoreticians\" and does not consider the broader field of tranfer in RL (which is an effort other papers make). If that is the case, then this paper should go to a specialized workshop or journal or seminar on the topic (and this should be perfectly acceptable). Otherwise, it needs better foundations, motivation and connexions to work outside the comfort zone of the authors.  \nRusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., ... & Hadsell, R. (2016). Progressive neural networks. NeurIPS deep learning symposium.  \nDayan, P. (1993). Improving generalization for temporal difference learning: The successor representation. Neural computation, 5(4), 613-624.  \nBarreto, A., Dabney, W., Munos, R., Hunt, J. J., Schaul, T., van Hasselt, H. P., & Silver, D. (2017). Successor features for transfer in reinforcement learning. Advances in neural information processing systems, 30.  \nBlier, L., Tallec, C., & Ollivier, Y. (2021). Learning successor states and goal-dependent values: A mathematical viewpoint. arXiv preprint arXiv:2101.07123.  \nTouati, A., & Ollivier, Y. (2021). Learning one representation to optimize all rewards. Advances in Neural Information Processing Systems, 34, 13-23.  \nLecarpentier, E., Abel, D., Asadi, K., Jinnai, Y., Rachelson, E., & Littman, M. L. (2021, May). Lipschitz lifelong reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 9, pp. 8270-8278).\n\nThe motivation for the cross-sampling part (section 4) is very unclear. Although, after dedicating a long time to understanding it, cross sampling with two successive transitions appears like a reasonable way to feed RepLearn (which is not a contribution from this paper since it is already the work of Agarwal (2020)). But really, the rationale behind this practice is unclear unless one (again) refers to the appendix and looks at the proofs. Not a single line on the topic in the paper: this strongly reduces the value for readers.\n\nConcerning the reward-free exploration part (section 4.1), LSVI-UCB seems a reasonable basis but it is not developped in the paper. Nor is its connection to recent work on reward-free exploration. E.g.:  \nKaufmann, E., Ménard, P., Domingues, O. D., Jonsson, A., Leurent, E., & Valko, M. (2021, March). Adaptive reward-free exploration. In Algorithmic Learning Theory (pp. 865-891). PMLR.  \nMénard, P., Domingues, O. D., Jonsson, A., Kaufmann, E., Leurent, E., & Valko, M. (2021, July). Fast active learning for pure exploration in reinforcement learning. In International Conference on Machine Learning (pp. 7599-7608). PMLR.  \nBadia, A. P., Sprechmann, P., Vitvitskyi, A., Guo, D., Piot, B., Kapturowski, S., ... & Blundell, C. (2020). Never give up: Learning directed exploration strategies. ICLR 2020.  \nDomingues, O. D., Tallec, C., Munos, R., & Valko, M. (2021, June). Density-Based Bonuses on Learned Representations for Reward-Free Exploration in Deep Reinforcement Learning. In ICML 2021 Workshop on Unsupervised Reinforcement Learning.  \nSo overall, besides the \"hey, this is an interesting setup upon which we can derive regret bounds\" side, this looks like an odd assembly of good ideas, without discussion on the rationale and I question what will remain of this paper in a few years (and this is in no way a negative judgment on the work done, i'm really questioning the impact this presentation will have).\n\nThere is a subsection numbered 4.1 but no 4.2. This is very odd and does not help structuring the paper. Either make it a separate section, or introduce a 4.1 at the beginning of section 4.\n\nIn theorem 5.1, what is the \"set of K-task multi-set\"?  \nIn theorem 5.1, the algorithm $\\mathcal{A}$ is said to interact with the source tasks, but the word \"online\" is never mentioned. As far as I understand, this theorem is specific to online interaction (it's the very motivation for this theorem), so this should be made explicit, otherwise the statement make no sense (interaction with the source tasks could use a generative model as in section 4).\n\nTheorem 5.2 directly refers (again) to algorithm 5 which is on page 28. No more comments on that.\n\nThere is no conclusion! Again: conference are not a patent registration desk for some new theorems. Papers should be reasonably accessible and feature hindsight views, perspectives, thoughtful discussion which leaves the reader bubbling with ideas and inspiration. It is not the case here. And although the work is of great quality, I strongly question its relevance for the reader.\n\nTypos and phrasing:  \nAbstract: weird sentence \"The sample complexity is close to knowing the ground truth features in the target task\". Close to that of learning the representations when knowing...?  \npage 3, \"transitions s_{h+1}\", the \"to\" is missing.  \npage 3, \\phi^* should be \\phi^\\star.  \npage 3, \"Assumption 3.2 is standard realizability condition\". is a standard   \npage 4, assumption 3.3 does not need the \"with \\alpha_{max}=...\" part. Please make it separate, it's confusing otherwise.  \npage 7, extra space after \"somewhat surprisingly\".  \npage 8, which results the failures -> which results in failure\n\nNotation:  \nP^\\star_h is the transition model at step k, while P^\\star_k is the full sequence of time-dependent transition kernels for task k. This is confusing. Why the \\star everywhere?   \nUsing \\Upsilon for the set of \\mu functions is confusing. Why not just M?   ", "clarity,_quality,_novelty_and_reproducibility": "Good language quality, bad paper organization and virtual length overflow. Please see main review for details.", "summary_of_the_review": "Overall, this seems to be very solid work on deriving interesting bounds for re-using RepLearn in the context of low-rank feature transfer, with a limited experimental validation. But there is no take-away message for the reader and the paper is unreadable without the appendix. To me, that alone is a motive for rejection. I hope the detailed review will clarify this motivation.\n\nBesides this, despite the interesting contents and the (apparent) correctness of the points developed, I question the impact this will have on the community. I think this paper needs profound restructuring to make it valuable to the reader and encourage the authors to make this effort for the community. Thus, as is, this paper is not suited for publication.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666949760224}, {"id": "bbgFKdJC_hK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper526/Reviewer_F8xp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work theoretically analyzes the applicability of feature/representation transfer from source tasks to a target task in RL, under the low-rank MDP setting (Def.3.1). It proposes an algorithm, RepTransfer (Algo.1), that learns the representations from cross-task samples, assuming one can sample from the transition kernel. It also shows that such cross-sampling is necessary (Thm.5.1) unless a stronger assumption is satisfied (Assumption 5.1). Experiments are conducted on the CombLock benchmark.", "review_text": "This paper provides several theoretical results for representational transfer in RL. Such results are important, but the applicability of the algorithm is not very clear, especially due to the limited experiments.", "strengths": "Strengths\n- Strong theoretical analysis\n- Clear writing\n\nWeaknesses\n- It remains unclear about the applicability of the results since one either has to sample from the transition kernel, or requires a stronger assumption. Both are not necessarily satisfied in many practical RL applications.\n- The experiments are only conducted on the CombLock benchmark, which is limited. Besides, the experiment results in Fig.2 are hard to parse because the description of the setup (even after reading some of the appendices) is not clear enough.\n\nTypo: In the “special case” after Assumption 3.4, there shouldn’t be summing over k for alphas and the sum for p_k should be over k instead of p.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work theoretically analyzes the applicability of feature/representation transfer from source tasks to a target task in RL, under the low-rank MDP setting (Def.3.1). It proposes an algorithm, RepTransfer (Algo.1), that learns the representations from cross-task samples, assuming one can sample from the transition kernel. It also shows that such cross-sampling is necessary (Thm.5.1) unless a stronger assumption is satisfied (Assumption 5.1). Experiments are conducted on the CombLock benchmark.", "strength_and_weaknesses": "Strengths\n- Strong theoretical analysis\n- Clear writing\n\nWeaknesses\n- It remains unclear about the applicability of the results since one either has to sample from the transition kernel, or requires a stronger assumption. Both are not necessarily satisfied in many practical RL applications.\n- The experiments are only conducted on the CombLock benchmark, which is limited. Besides, the experiment results in Fig.2 are hard to parse because the description of the setup (even after reading some of the appendices) is not clear enough.\n\nTypo: In the “special case” after Assumption 3.4, there shouldn’t be summing over k for alphas and the sum for p_k should be over k instead of p.", "clarity,_quality,_novelty_and_reproducibility": "The paper is mostly of theoretical nature. It shows several interesting results, especially on the impossibility of transfer due to a permutation issue. The presentation is clear.", "summary_of_the_review": "This paper provides several theoretical results for representational transfer in RL. Such results are important, but the applicability of the algorithm is not very clear, especially due to the limited experiments.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666660795880}], "openreview_url": "https://openreview.net/forum?id=sNKZaNkyi7Q", "arxiv_id": "2205.14571", "paper_pdf": "papers/sNKZaNkyi7Q.pdf", "paper_pdf_sha256": "4b98e962d4a9160750d84c584806c7c6c2ca8bfb257d7df3dd0c3d78a4352f91", "paper_pdf_bytes": 1212001, "paper_pdf_source": "openreview", "code_url": "https://github.com/yudasong/briee", "code_repository": "yudasong/briee", "code_commit": "c6920ff1806e636a4d806ffd2a6a7d26db7bab6c", "code_archive": "repos/sNKZaNkyi7Q.zip", "code_archive_sha256": "cfd411dc86197fab28ea5c52dbbf4bd2b05f66bb492036a23783f0ad9ee9952c", "code_archive_bytes": 14609, "code_file_count": 9, "code_extensions": {".py": 6, ".sh": 3}, "github_disk_usage_kb": 30, "github_languages": {"Python": 41358, "Shell": 585}, "github_archived": false, "github_pushed_at": "2022-06-01T19:02:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/provable-benefits-of-representational"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ExJ4lMbZcqa", "year": 2022, "status": "rejected", "title": "Learning Audio-Visual Dereverberation", "authors": ["Changan Chen", "Wei Sun", "David Harwath", "Kristen Grauman"], "authorids": ["~Changan_Chen2", "~Wei_Sun10", "~David_Harwath1", "~Kristen_Grauman1"], "authors_source": "OpenReview API", "abstract": "Reverberation from audio reflecting off surfaces and objects in the environment not only degrades the quality of speech for human perception, but also severely impacts the accuracy of automatic speech recognition. Prior work attempts to remove reverberation based on the audio modality only. Our idea is to learn to dereverberate speech from audio-visual observations. The visual environment surrounding a human speaker reveals important cues about the room geometry, materials, and speaker location, all of which influence the precise reverberation effects in the audio stream. We introduce Visually-Informed Dereverberation of Audio (VIDA), an end-to-end approach that learns to remove reverberation based on both the observed sounds and visual scene. In support of this new task, we develop a large-scale dataset that uses realistic acoustic renderings of speech in real-world 3D scans of homes offering a variety of room acoustics. Demonstrating our approach on both simulated and real imagery for speech enhancement, speech recognition, and speaker identification, we show it achieves state-of-the-art performance and substantially improves over traditional audio-only methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "fZx5zXxIvl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1065/Reviewer_Tgq4"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper examines an audio-visual approach for dereverberation, where dereverberation of speech is conditioned on RGB and depth images (either field-of-view or panoramic). It proposes a dataset for this task based on real-world 3D scans of homes, using Librispeech data. Real data is also used The model is based on a U-Net conditioned on embeddings extracted by a \"visual acoustics\" network doing direct spectrogram prediction. The method is evaluated using PESQ, WER for speech recognition, and EER for speaker verification on synthetic and real data. The audio-visual method is found to perform marginally better than an audio-only version of the model. The audio-only version of the model outperforms several baselines from the literature on synthetic data, with mixed results on real data.", "review_text": "### Strengths:\n\nS1) To my knowledge, this is the first audio-visual approach to dereverberation. I think it is interesting to see how much visual information can help.\n\nS2) The proposed dataset seems quite useful. A lot of existing synthetic reverb for such audio tasks is generated using shoebox room simulators, so having more realistic environments seems quite useful. The accompanying data recorded in real environments using loudspeaker playback (such that the clean reference is available) is also quite useful, and I hope the authors plan to release a recipe for this data.\n\nS3) The paper includes a lot of ablations, which I really appreciate. Though the focus of the paper is on audio-visual approaches to the task, I'm really glad the paper evaluates audio-only versions of the model, as well as other components such as the matching loss.\n\nS4) The paper is clear and well-written.\n\n### Weaknesses:\n\nW1) I listened to the demos in the supplementary material, and I feel like I heard a fair amount of artifacts and distortion in some examples: for moderate reverb, the model predictions have a certain \"buzziness\" to them, and more extreme levels of reverb have a \"twanginess\" to them. Something that would make this paper stronger would be a subjective human listening evaluation, e.g. MUSHRA, which would measure these kinds of artifacts better than an objective measure like PESQ. Also, alternative loss functions may help reduce artifacts. E.g. L1 loss instead of L2 on spectrogram magnitude, or MSE on power-compressed spectrograms can provide more perceptually-relevant loss functions.\n\nW2) The processing method of using 2.56s frames, then concatenating the non-overlapping parts together seems suboptimal to me. Why not use e.g. a Hann window to stitch the predicted patches with overlap-add, or invert back to the time-domain and do overlap-add and train through this?\n\nW3) Where does the \"random speech embedding sampled from the data batch.\" come from? How is it computed? Is this a speaker ID embedding?\n\nW4) The improvements by adding visual information are marginal compared to the audio-only model. The audio-only model does perform better than the baselines, but degrades PESQ compared to other baselines on real data. To benchmark the audio-only baseline against the literature, it would be interesting to compare to competitive systems in e.g. the REVERB challenge. I'm also not very convinced that visual input really helps that much.\n\nW5) Table 1 would be improved by adding trainable parameter counts. Is the proposed model only doing better because it has more trainable parameters? WPE itself has no trainable parameters, since it's just a model-based EM algorithm.\n\nW6) The lack of noise in the training and eval data limits real-world applicability, but I understand the focus of the paper in only on dereverberation. I'm glad that the authors tested noise robustness with WHAM! data, but 20 dB is a pretty high SNR for noise robustness. It would be interesting to see how performance degrades with decreasing SNR.\n\n### Minor comments\n\nM1) I feel that this statement isn't quite right: \"R is a function of...the relative positioning of the speaker and the listener.\" I would argue that an RIR can be a function of the absolute positioning of the speaker and the listener. E.g. what if a source is close to a wall, versus in the the middle of a room?\n\nM2) For the real data, would it be possible to be more specific about the source-to-microphone distances used in the real recordings, beyond \"near-field to mid-field to far-field.\"?\n\nM3) Also concerning real data: would it be possible to more more specific for this statement? \"For each location, we play around 10 utterances.\". How many utterances? Gender balanced? Different speakers across rooms?\n\nM4) Nit: U-Net instead of UNet?\n\nM5) \"While our model is agnostic to the audio source type,\": depends on the training data.\n\nM6) In Tables 1 and 2, I would say \"anechoic speech (upper bound)\" instead of \"clean speech (upper bound)\", because to me \"clean\" suggests there's additive noise.\n\nM7) About the blind RT60/DRR/distance predictions: \"After converging, the average errors for them are 3e-3s, 160 dB and 17.92m respectively.\" Are these absolute or squared errors?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper examines an audio-visual approach for dereverberation, where dereverberation of speech is conditioned on RGB and depth images (either field-of-view or panoramic). It proposes a dataset for this task based on real-world 3D scans of homes, using Librispeech data. Real data is also used The model is based on a U-Net conditioned on embeddings extracted by a \"visual acoustics\" network doing direct spectrogram prediction. The method is evaluated using PESQ, WER for speech recognition, and EER for speaker verification on synthetic and real data. The audio-visual method is found to perform marginally better than an audio-only version of the model. The audio-only version of the model outperforms several baselines from the literature on synthetic data, with mixed results on real data.", "main_review": "### Strengths:\n\nS1) To my knowledge, this is the first audio-visual approach to dereverberation. I think it is interesting to see how much visual information can help.\n\nS2) The proposed dataset seems quite useful. A lot of existing synthetic reverb for such audio tasks is generated using shoebox room simulators, so having more realistic environments seems quite useful. The accompanying data recorded in real environments using loudspeaker playback (such that the clean reference is available) is also quite useful, and I hope the authors plan to release a recipe for this data.\n\nS3) The paper includes a lot of ablations, which I really appreciate. Though the focus of the paper is on audio-visual approaches to the task, I'm really glad the paper evaluates audio-only versions of the model, as well as other components such as the matching loss.\n\nS4) The paper is clear and well-written.\n\n### Weaknesses:\n\nW1) I listened to the demos in the supplementary material, and I feel like I heard a fair amount of artifacts and distortion in some examples: for moderate reverb, the model predictions have a certain \"buzziness\" to them, and more extreme levels of reverb have a \"twanginess\" to them. Something that would make this paper stronger would be a subjective human listening evaluation, e.g. MUSHRA, which would measure these kinds of artifacts better than an objective measure like PESQ. Also, alternative loss functions may help reduce artifacts. E.g. L1 loss instead of L2 on spectrogram magnitude, or MSE on power-compressed spectrograms can provide more perceptually-relevant loss functions.\n\nW2) The processing method of using 2.56s frames, then concatenating the non-overlapping parts together seems suboptimal to me. Why not use e.g. a Hann window to stitch the predicted patches with overlap-add, or invert back to the time-domain and do overlap-add and train through this?\n\nW3) Where does the \"random speech embedding sampled from the data batch.\" come from? How is it computed? Is this a speaker ID embedding?\n\nW4) The improvements by adding visual information are marginal compared to the audio-only model. The audio-only model does perform better than the baselines, but degrades PESQ compared to other baselines on real data. To benchmark the audio-only baseline against the literature, it would be interesting to compare to competitive systems in e.g. the REVERB challenge. I'm also not very convinced that visual input really helps that much.\n\nW5) Table 1 would be improved by adding trainable parameter counts. Is the proposed model only doing better because it has more trainable parameters? WPE itself has no trainable parameters, since it's just a model-based EM algorithm.\n\nW6) The lack of noise in the training and eval data limits real-world applicability, but I understand the focus of the paper in only on dereverberation. I'm glad that the authors tested noise robustness with WHAM! data, but 20 dB is a pretty high SNR for noise robustness. It would be interesting to see how performance degrades with decreasing SNR.\n\n### Minor comments\n\nM1) I feel that this statement isn't quite right: \"R is a function of...the relative positioning of the speaker and the listener.\" I would argue that an RIR can be a function of the absolute positioning of the speaker and the listener. E.g. what if a source is close to a wall, versus in the the middle of a room?\n\nM2) For the real data, would it be possible to be more specific about the source-to-microphone distances used in the real recordings, beyond \"near-field to mid-field to far-field.\"?\n\nM3) Also concerning real data: would it be possible to more more specific for this statement? \"For each location, we play around 10 utterances.\". How many utterances? Gender balanced? Different speakers across rooms?\n\nM4) Nit: U-Net instead of UNet?\n\nM5) \"While our model is agnostic to the audio source type,\": depends on the training data.\n\nM6) In Tables 1 and 2, I would say \"anechoic speech (upper bound)\" instead of \"clean speech (upper bound)\", because to me \"clean\" suggests there's additive noise.\n\nM7) About the blind RT60/DRR/distance predictions: \"After converging, the average errors for them are 3e-3s, 160 dB and 17.92m respectively.\" Are these absolute or squared errors?", "summary_of_the_review": "Overall, I vote for marginal acceptance. This is an interesting novel task (audio-visual dereverberation), and the proposed synthetic and real data are quite interesting and useful for the community. But the paper's results show that visual input provides marginal performance gain over the audio-only baseline. The audio-only baseline outperforms some benchmarks in terms of PESQ, WER for speech recognition, and EER for speaker verification on synthetic data, and WER and EER for real data, but I have concerns about the subjective listening quality of the demos, compared to other dereverberation algorithms I have listened to. Thus, I am concerned that this isn't a significant improvement over audio-only prior work, at least in terms of subjective listening quality.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635908639842}, {"id": "mMWXE_dlaSM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1065/Reviewer_NGNE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors introduce a novel audio-visual dereverberation approach. They propose a Visually-Informed Dereverberation of Audio (VIDA) model for dereverberation. The authors also create synthetic/simulated datasets and real-world data for experimentation. Finally, they show the impact of the proposed VIDA model on several speech tasks including, recognition, enhancement, and speaker verification. The results are encouraging. The main contribution of this work is the use of visual information as an auxiliary input for dereverberation. \n\n\n\n", "review_text": "While authors suggest visual information is helpful, it isn't reflected in the results. In Table 1, if we compare the \"Audio-only dereverb.\" and \"VIDA\" results, the gains are very modest. The PESQ 2.32 to 2.37, WER-FT 3.76 to 3.66, and SID EER 2.61 to 2.40. This raises the question about the effectiveness of the proposed approach. I feel the majority of the dereverberation is still learned by the audio model. This should be investigated. \n\nOverall, the technical contribution is limited to a new dataset and the addition of ResNet for visual information processing. Similar approaches have been used in the audio-visual enhancement with lip information. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors introduce a novel audio-visual dereverberation approach. They propose a Visually-Informed Dereverberation of Audio (VIDA) model for dereverberation. The authors also create synthetic/simulated datasets and real-world data for experimentation. Finally, they show the impact of the proposed VIDA model on several speech tasks including, recognition, enhancement, and speaker verification. The results are encouraging. The main contribution of this work is the use of visual information as an auxiliary input for dereverberation. \n\n\n\n", "main_review": "While authors suggest visual information is helpful, it isn't reflected in the results. In Table 1, if we compare the \"Audio-only dereverb.\" and \"VIDA\" results, the gains are very modest. The PESQ 2.32 to 2.37, WER-FT 3.76 to 3.66, and SID EER 2.61 to 2.40. This raises the question about the effectiveness of the proposed approach. I feel the majority of the dereverberation is still learned by the audio model. This should be investigated. \n\nOverall, the technical contribution is limited to a new dataset and the addition of ResNet for visual information processing. Similar approaches have been used in the audio-visual enhancement with lip information. ", "summary_of_the_review": "I believe the manuscript in its current form offers a very limited contribution to the research community, and the contribution of visual information in dereverberation is minimal. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635879042502}, {"id": "g6Fm3L8ircw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1065/Reviewer_NeeT"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a deep neural net-based dereverberation algorithm that uses both audio and video modalities. Based on the observation that a visual scene captured by a camera conveys information that is related to room characteristics, the authors propose a visually informed audio dereverberation method that aims to extract clean, anechoic speech from reverberant speech. In doing so, they first construct a large audio-visual dataset synthesized using a 3D simulator for real-world scanned environments and LibriSpeech data. They then train deep neural networks that take as inputs both visual data (RGB and depth images) and audio data (reverberant speech), and output clean speech. When training, two types of losses - one for clean speech spectrogram estimation and the other for reverb-visual matching - are used. Through the experiments with several downstream tasks for speech, they showed that the proposed audio-visual dererberation method outperforms the baseline models, both for synthetic and real-world test data.", "review_text": "Strengths:\n- first multi-modal (audio-visual) approach applied to dereverberation\n- paper is well organized and straightforward to follow\n- solid experiments with ablation studies and in-depth analyses\n- supplementary materials provide useful information not explained/described in the main text\n\nWeaknesses:\n- some of the key arguments are not sufficiently validated\n- contributions are marginal and limited\n- experiments and evaluation are not thorough\n\nThe key idea in this work is that visual input captured by a camera conveys information that characterizes room acoustics. With all the experiments and ablation studies, however, I don't believe that the authors provide sufficient information that support the aforementioned key hypothesis. The authors did take into account this matter in two ways: first, they included a reverb-visual matching loss term to penalize a matching between random speech and visual input, but little experiment or analysis was performed on this loss term - varying the weight or using different loss functions, for example. With or without using a reverb-visual matching loss was included in the ablation studies, but the effect was not significant compared with other methods. The performance was better with the audio-only model even in speech quality metric (PESQ). The second way is to use a random image instead of using the image that matches the room impulse response. In such scenario, the images does not carry any information relevant to the room acoustics, and thus the performance should be worse than or comparable to that of the audio-only method. However, the performance is better in general and is even best for some metric. Further experiments and deeper analysis are warranted to investigate this issue.\n\nAnother point that I hope to be added is the comparison of the model efficiency. Adding a visual embedding network increases the number of model parameters and computational costs. The proposed method is likely to work in real-time scenarios, and possibly on small devices, the efficiency should also be considered.\n\nLastly, I'd like to suggest to include R-vectors approaches as another baseline for comparison.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a deep neural net-based dereverberation algorithm that uses both audio and video modalities. Based on the observation that a visual scene captured by a camera conveys information that is related to room characteristics, the authors propose a visually informed audio dereverberation method that aims to extract clean, anechoic speech from reverberant speech. In doing so, they first construct a large audio-visual dataset synthesized using a 3D simulator for real-world scanned environments and LibriSpeech data. They then train deep neural networks that take as inputs both visual data (RGB and depth images) and audio data (reverberant speech), and output clean speech. When training, two types of losses - one for clean speech spectrogram estimation and the other for reverb-visual matching - are used. Through the experiments with several downstream tasks for speech, they showed that the proposed audio-visual dererberation method outperforms the baseline models, both for synthetic and real-world test data.", "main_review": "Strengths:\n- first multi-modal (audio-visual) approach applied to dereverberation\n- paper is well organized and straightforward to follow\n- solid experiments with ablation studies and in-depth analyses\n- supplementary materials provide useful information not explained/described in the main text\n\nWeaknesses:\n- some of the key arguments are not sufficiently validated\n- contributions are marginal and limited\n- experiments and evaluation are not thorough\n\nThe key idea in this work is that visual input captured by a camera conveys information that characterizes room acoustics. With all the experiments and ablation studies, however, I don't believe that the authors provide sufficient information that support the aforementioned key hypothesis. The authors did take into account this matter in two ways: first, they included a reverb-visual matching loss term to penalize a matching between random speech and visual input, but little experiment or analysis was performed on this loss term - varying the weight or using different loss functions, for example. With or without using a reverb-visual matching loss was included in the ablation studies, but the effect was not significant compared with other methods. The performance was better with the audio-only model even in speech quality metric (PESQ). The second way is to use a random image instead of using the image that matches the room impulse response. In such scenario, the images does not carry any information relevant to the room acoustics, and thus the performance should be worse than or comparable to that of the audio-only method. However, the performance is better in general and is even best for some metric. Further experiments and deeper analysis are warranted to investigate this issue.\n\nAnother point that I hope to be added is the comparison of the model efficiency. Adding a visual embedding network increases the number of model parameters and computational costs. The proposed method is likely to work in real-time scenarios, and possibly on small devices, the efficiency should also be considered.\n\nLastly, I'd like to suggest to include R-vectors approaches as another baseline for comparison.", "summary_of_the_review": "This paper proposes a multi-modal learning framework that is applied to speech dereverberation. The idea is interesting and somewhat novel, but the experimental validation is not thorough enough to support the key idea of reverb-visual matching nor to provide sufficient evidences that visual stream does convey information about room acoustics.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635851931795}, {"id": "xq0qZGU0fWv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1065/Reviewer_j5yj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper describes an audio de-reverberation method which integrates (panoramic, rgb+depth map) visual input with a spectrogram U-Net to estimate a de-reverberated spectrogram and recover the original clean signal.\nThe proposed method is compared to several baseline (audio-only) de-reverberation methods on speech enhancement, recognition, and speaker verification problems using both synthetic and real data.\nSeveral ablations of the proposed method are compared, along with some quantitative error analysis investigating the effects of distance and environment.\nThe proposed method appears to consistently improve on speech recognition and speaker verification, and perform about on par with prior work on speech enhancement (as measured by PESQ).\n\n", "review_text": "This paper takes an interesting and novel approach to de-reverberation.\nThe paper is well written, and each component of the method is well justified.\nThe qualitative analysis (e.g. figure 4) is a nice touch, and helps shed some light on the behavior of the proposed method (and baselines).\nThat said, I do have a few questions and suggestions.\n\nMy primary criticism is in the reporting of the evaluations (tables 1, 2, 3).\nEach setting is summarized by a single average score, and the \"best\" number is written in bold with no discussion given to the distribution of scores on the test data or the practical meaning of score differences.\nThis may give readers an inflated sense of performance here: is PESQ=2.38 really any different from PESQ=2.37? 2.33?  (Likewise for WER, EER, etc.)\nThere is a brief mention of astronomically low p-values (1.56e-60?) which I expect are primarily driven by the size of the test set, and not the size of the effect.\nI would strongly encourage the authors to do a more responsible job of reporting their results: include some error bars, or some other indication of the spread of the data!\n\nThe proposed method takes an interesting approach to optimizing both magnitude and phase (eqs 1 and 2) for the output spectrogram.\nThe Griffin-Lim algorithm is then used to refine the phase estimate, with a vague statement of this working better than directly using predicted phase or GL with random initialization.\nWhile I won't ask for more experiments on this, I think the point deserves a bit more explanation than what is given here, especially since there is no ablation study on the phase reconstruction loss.  (Nor is there any discussion of hyper-parameter tuning for equation 4 beyond the settings reported in the appendix.)\n\n\nI found Figure 5 (qualitative examples) to be pretty difficult to parse, and I say this as a reader with (probably) above-average experience reading spectrograms.\nWith a bit of squinting I can see what the authors are getting at, but it would be easier to read if the spectrograms were A) larger, and B) plotted on a logarithmic frequency axis so that the low frequencies (where most of the action is) occupy more of the visual real estate.\n\n\nFinally, it was nice to see the results in table 3, which show quite a strong effect on the performance of the system depending on differences in distance and room geometry.\nNot much is said about the geometric differences in the environments though, either for real or simulated data.\nIn particular, it looks from figure 2 as though there may be configurations of speaker/listener that have no direct path, in which one might say that everything observed is \"reverb\".\nIf this is the case, it would be interesting to see how performance (table 1) differs (if at all) in the presence or absence of a direct path.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper describes an audio de-reverberation method which integrates (panoramic, rgb+depth map) visual input with a spectrogram U-Net to estimate a de-reverberated spectrogram and recover the original clean signal.\nThe proposed method is compared to several baseline (audio-only) de-reverberation methods on speech enhancement, recognition, and speaker verification problems using both synthetic and real data.\nSeveral ablations of the proposed method are compared, along with some quantitative error analysis investigating the effects of distance and environment.\nThe proposed method appears to consistently improve on speech recognition and speaker verification, and perform about on par with prior work on speech enhancement (as measured by PESQ).\n\n", "main_review": "This paper takes an interesting and novel approach to de-reverberation.\nThe paper is well written, and each component of the method is well justified.\nThe qualitative analysis (e.g. figure 4) is a nice touch, and helps shed some light on the behavior of the proposed method (and baselines).\nThat said, I do have a few questions and suggestions.\n\nMy primary criticism is in the reporting of the evaluations (tables 1, 2, 3).\nEach setting is summarized by a single average score, and the \"best\" number is written in bold with no discussion given to the distribution of scores on the test data or the practical meaning of score differences.\nThis may give readers an inflated sense of performance here: is PESQ=2.38 really any different from PESQ=2.37? 2.33?  (Likewise for WER, EER, etc.)\nThere is a brief mention of astronomically low p-values (1.56e-60?) which I expect are primarily driven by the size of the test set, and not the size of the effect.\nI would strongly encourage the authors to do a more responsible job of reporting their results: include some error bars, or some other indication of the spread of the data!\n\nThe proposed method takes an interesting approach to optimizing both magnitude and phase (eqs 1 and 2) for the output spectrogram.\nThe Griffin-Lim algorithm is then used to refine the phase estimate, with a vague statement of this working better than directly using predicted phase or GL with random initialization.\nWhile I won't ask for more experiments on this, I think the point deserves a bit more explanation than what is given here, especially since there is no ablation study on the phase reconstruction loss.  (Nor is there any discussion of hyper-parameter tuning for equation 4 beyond the settings reported in the appendix.)\n\n\nI found Figure 5 (qualitative examples) to be pretty difficult to parse, and I say this as a reader with (probably) above-average experience reading spectrograms.\nWith a bit of squinting I can see what the authors are getting at, but it would be easier to read if the spectrograms were A) larger, and B) plotted on a logarithmic frequency axis so that the low frequencies (where most of the action is) occupy more of the visual real estate.\n\n\nFinally, it was nice to see the results in table 3, which show quite a strong effect on the performance of the system depending on differences in distance and room geometry.\nNot much is said about the geometric differences in the environments though, either for real or simulated data.\nIn particular, it looks from figure 2 as though there may be configurations of speaker/listener that have no direct path, in which one might say that everything observed is \"reverb\".\nIf this is the case, it would be interesting to see how performance (table 1) differs (if at all) in the presence or absence of a direct path.\n\n", "summary_of_the_review": "Overall, I enjoyed this paper!  I think some of the reporting could be done better, but this is a minor complaint.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635193549730}], "openreview_url": "https://openreview.net/forum?id=ExJ4lMbZcqa", "arxiv_id": "2106.07732", "paper_pdf": "papers/ExJ4lMbZcqa.pdf", "paper_pdf_sha256": "f69623afbc0ef2b2e257786589839f887b557e29bf135a2ee591c8d1a31116f1", "paper_pdf_bytes": 22695264, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/learning-audio-visual-dereverberation", "code_repository": "facebookresearch/learning-audio-visual-dereverberation", "code_commit": "ef5ceb4b8fe59d10798628b7a4498c686473fd3d", "code_archive": "repos/ExJ4lMbZcqa.zip", "code_archive_sha256": "18b866747ed1236c55e01b1d02d8d6c93c8ef49bfa1c73f049e467a17519e3f0", "code_archive_bytes": 29334, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 27, "github_languages": {"Python": 60359}, "github_archived": true, "github_pushed_at": "2022-08-10T02:04:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-audio-visual-dereverberation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lSijhyKKsct", "year": 2021, "status": "rejected", "title": "Reinforcement Learning with Latent Flow", "authors": ["Wenling Shang", "Xiaofei Wang", "Aravind Rajeswaran", "Aravind Srinivas", "Yang Gao", "Pieter Abbeel", "Michael Laskin"], "authorids": ["~Wenling_Shang1", "w.xf@berkeley.edu", "~Aravind_Rajeswaran1", "~Aravind_Srinivas1", "~Yang_Gao1", "~Pieter_Abbeel2", "~Michael_Laskin1"], "authors_source": "OpenReview API", "abstract": "Temporal information is essential to learning effective policies with Reinforcement Learning (RL). However, current state-of-the-art RL algorithms either assume that such information is given as part of the state space or, when learning from pixels, use the simple heuristic of frame-stacking to implicitly capture temporal information present in the image observations. This heuristic is in contrast to the current paradigm in video classification architectures, which utilize explicit encodings of temporal information through methods such as optical flow and two-stream architectures to achieve state-of-the-art performance. Inspired by leading video classification architectures, we introduce the Flow of Latents for Reinforcement Learning Flare, a network architecture for RL that explicitly encodes temporal information through latent vector differences. We show that Flare (i) recovers optimal performance in state-based RL without explicit access to the state velocity, solely with positional state information, (ii) achieves state-of-the-art performance on pixel-based continuous control tasks within the DeepMind control benchmark suite, (iii) is the most sample efficient model-free pixel-based RL algorithm on challenging environments in the DeepMind control suite such as quadruped walk, hopper hop, finger turn hard, pendulum swing, and walker run, outperforming the prior model-free state-of-the-art by 1.9 and 1.5 on the 500k and 1M step benchmarks, respectively, and (iv), when augmented over rainbow DQN, outperforms or matches the baseline on a diversity of challenging Atari games at 50M time step benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Lr3bBkjePkg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1855/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new method for aggregating temporal information in reinforcement learning policies. The method takes the difference of latent representations between consecutive frames and concatenates this difference with the latent representations for downstream processing.\n\nStrengths:\n- The method is simple and easy to understand and implement.\n- The method outperforms a baseline on average based on experiments on 5 Deepmind control benchmark suite tasks under a limited budget of training samples.\n- Ablations show an interesting result that the fusion of latent representations performs better than pixel-based fusion.\n\nWeaknesses:\n- The contribution is incremental in my opinion. It makes a small change over existing RL methods. The improvement in performance on several tasks is also marginal. \n- One of the most common methods of aggregating temporal information is just using a recurrent layer in RL policies. It’s unclear if the baseline consists of a recurrent layer. If it does not, a simple recurrent policy needs to be added as a baseline. If it does, it is unclear to me why a simple subtraction of latent representations would result in performance improvement. I would imagine that it should be very easy for a recurrent layer such as an LSTM to learn differences between latent representations if they were useful. Some discussion on this would be useful.\n- The performance gain is only demonstrated on 5 tasks and it’s unclear whether it would translate to gains in other tasks as well. Results in the appendix on 6 easier tasks show results comparable to or worse than the baseline.\n\nOverall, I believe that the paper proposes a simple and interesting method, but I am not convinced that it would lead to better results consistently. The main experiments are conducted only on 5 tasks and the performance gains are marginal in my opinion.\n\nUpdate after rebuttal: \nThe authors have added a recurrent SAC baseline to one set of experiments. The results indicate that the recurrent SAC is a much stronger baseline, and the variance of results is high enough that I am not convinced of the benefits of the proposed method.\n\nThe authors argue that \"many state-of-the-art RL algorithms\" \"are non-recurrent\", and \"frame stacking\" is \"largely untouched since its inception and is used in most state-of-the-art RL architectures\", \"recurrent architectures\" have \"additional overhead from training and implementation\". I do not believe this is true. Recurrent architectures are commonly used in RL algorithms (for example RSSM in Dreamer) and are widely available in open-source implementations (for example https://github.com/openai/baselines/blob/master/baselines/common/models.py). There are some prior papers which use the frame stacking heuristic for a fair comparison with DQN, but this heuristic or the non-recurrent model architecture is not a part of the RL algorithm itself. Since this paper proposes a method for extracting temporal information, LSTM/GRU are very natural baselines in my opinion and should be added to all experiments.\n\nThe authors argue that \"it is very reasonable and common to have a new method improving a majority of the environments but not all\", I agree with this, but the examples given by the authors such as Rainbow DQN, Dueling DQN, Dreamer etc perform experiments in many more environments and performance improvements are larger. I believe a much larger scale study is needed to compare Flare with recurrent baselines and make conclusive statements about performance gains.\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review 1", "review": "This paper presents a new method for aggregating temporal information in reinforcement learning policies. The method takes the difference of latent representations between consecutive frames and concatenates this difference with the latent representations for downstream processing.\n\nStrengths:\n- The method is simple and easy to understand and implement.\n- The method outperforms a baseline on average based on experiments on 5 Deepmind control benchmark suite tasks under a limited budget of training samples.\n- Ablations show an interesting result that the fusion of latent representations performs better than pixel-based fusion.\n\nWeaknesses:\n- The contribution is incremental in my opinion. It makes a small change over existing RL methods. The improvement in performance on several tasks is also marginal. \n- One of the most common methods of aggregating temporal information is just using a recurrent layer in RL policies. It’s unclear if the baseline consists of a recurrent layer. If it does not, a simple recurrent policy needs to be added as a baseline. If it does, it is unclear to me why a simple subtraction of latent representations would result in performance improvement. I would imagine that it should be very easy for a recurrent layer such as an LSTM to learn differences between latent representations if they were useful. Some discussion on this would be useful.\n- The performance gain is only demonstrated on 5 tasks and it’s unclear whether it would translate to gains in other tasks as well. Results in the appendix on 6 easier tasks show results comparable to or worse than the baseline.\n\nOverall, I believe that the paper proposes a simple and interesting method, but I am not convinced that it would lead to better results consistently. The main experiments are conducted only on 5 tasks and the performance gains are marginal in my opinion.\n\nUpdate after rebuttal: \nThe authors have added a recurrent SAC baseline to one set of experiments. The results indicate that the recurrent SAC is a much stronger baseline, and the variance of results is high enough that I am not convinced of the benefits of the proposed method.\n\nThe authors argue that \"many state-of-the-art RL algorithms\" \"are non-recurrent\", and \"frame stacking\" is \"largely untouched since its inception and is used in most state-of-the-art RL architectures\", \"recurrent architectures\" have \"additional overhead from training and implementation\". I do not believe this is true. Recurrent architectures are commonly used in RL algorithms (for example RSSM in Dreamer) and are widely available in open-source implementations (for example https://github.com/openai/baselines/blob/master/baselines/common/models.py). There are some prior papers which use the frame stacking heuristic for a fair comparison with DQN, but this heuristic or the non-recurrent model architecture is not a part of the RL algorithm itself. Since this paper proposes a method for extracting temporal information, LSTM/GRU are very natural baselines in my opinion and should be added to all experiments.\n\nThe authors argue that \"it is very reasonable and common to have a new method improving a majority of the environments but not all\", I agree with this, but the examples given by the authors such as Rainbow DQN, Dueling DQN, Dreamer etc perform experiments in many more environments and performance improvements are larger. I believe a much larger scale study is needed to compare Flare with recurrent baselines and make conclusive statements about performance gains.\n\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603946051006}, {"id": "Qn9qxBV4wn1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1855/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Summary:\n    - This paper presents Flare, an RL method that replaces frame stacking (early fusion) with latent vector stacking (late fusion) and then further improve upon this by adding in latent flow vectors (the difference between adjacent latent vectors)\n    - The method is demonstrated on DM control using RAD-SAC as the baseline.  In all but one case, Flare outperforms the baseline\n    - The authors then present a series of ablations to show that latent flow outperforms pixel flow and that stacking latent flow is better than pure latent stacking\n- Strengths:\n    - Relatively straightforward idea that improves upon the common technique of frame stacking, making this likely a technique with broad appeal and impact\n    - Good ablation study showing how the individual components.\n    - Well motivated idea.  I liked the experiment of position only SAC to set the motivation.\n- Weaknesses\n    - Concat + MLP seems like a poor way to do sequence modeling (a series of frames is, after all, a sequence).  How does Flare and the latent frame baseline behave if a GRU or LSTM is used to combine the the sequence of latent vectors?  This would avoid the normal challenges of training a recurrent policy while also benefiting from the superior sequence modeling of an RNN.\n    - Performance on pendulum degrading as the number of frames is increased is concerning.  If possible, I wold like to see the hypothesis posed in Sec 6.3 Q3 validated by training longer.  Another possible hypothesis is that latent vector stacking increases the number of parameters and causes Q function overfitting.\n    - Questions:\n        - How was RAD applied to series of frames?  Was the same translation applied to all or was a different one applied too each?\n- Overall\n    - This paper presents and effective idea, however, there are some additional experiments (using an RNN to combine latent vectors) that I think would strength the paper considerably\n\n\n## Post Rebuttal\n\nI thank the authors for their response.  The addition of the recurrent SAC baseline helps the paper.  I disagree with R1 that it is a stronger baseline as FLARE outperforms it in all tasks and stack SAC similar or better three (arguably four) of five tasks. Instead it shows that recurrence isn't common in off-policy RL because it doesn't always perform better.  While recurrence is considerably more common in embodied 3D environments and this work may be less applicable there, I don't foresee DM control style RL benchmarks going away anytime soon and this believe this method will be useful.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #4", "review": "- Summary:\n    - This paper presents Flare, an RL method that replaces frame stacking (early fusion) with latent vector stacking (late fusion) and then further improve upon this by adding in latent flow vectors (the difference between adjacent latent vectors)\n    - The method is demonstrated on DM control using RAD-SAC as the baseline.  In all but one case, Flare outperforms the baseline\n    - The authors then present a series of ablations to show that latent flow outperforms pixel flow and that stacking latent flow is better than pure latent stacking\n- Strengths:\n    - Relatively straightforward idea that improves upon the common technique of frame stacking, making this likely a technique with broad appeal and impact\n    - Good ablation study showing how the individual components.\n    - Well motivated idea.  I liked the experiment of position only SAC to set the motivation.\n- Weaknesses\n    - Concat + MLP seems like a poor way to do sequence modeling (a series of frames is, after all, a sequence).  How does Flare and the latent frame baseline behave if a GRU or LSTM is used to combine the the sequence of latent vectors?  This would avoid the normal challenges of training a recurrent policy while also benefiting from the superior sequence modeling of an RNN.\n    - Performance on pendulum degrading as the number of frames is increased is concerning.  If possible, I wold like to see the hypothesis posed in Sec 6.3 Q3 validated by training longer.  Another possible hypothesis is that latent vector stacking increases the number of parameters and causes Q function overfitting.\n    - Questions:\n        - How was RAD applied to series of frames?  Was the same translation applied to all or was a different one applied too each?\n- Overall\n    - This paper presents and effective idea, however, there are some additional experiments (using an RNN to combine latent vectors) that I think would strength the paper considerably\n\n\n## Post Rebuttal\n\nI thank the authors for their response.  The addition of the recurrent SAC baseline helps the paper.  I disagree with R1 that it is a stronger baseline as FLARE outperforms it in all tasks and stack SAC similar or better three (arguably four) of five tasks. Instead it shows that recurrence isn't common in off-policy RL because it doesn't always perform better.  While recurrence is considerably more common in embodied 3D environments and this work may be less applicable there, I don't foresee DM control style RL benchmarks going away anytime soon and this believe this method will be useful.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603916354520}, {"id": "t0FClChyvKw", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1855/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Significance:\nThe paper brings very little novelty or insight. It is unclear that the introduced architecture complexity worth marginal improvements (given high variance and only 5 random seeds) on 2 out of 11 tasks (5 from the main paper and 6 from appendix). This might be a good workshop paper but it clearly does not meet the high acceptance threshold of ICLR.\n\nPros:\n-A simple architecture modification that might be beneficial to impose a stronger inductive bias on temporal dynamics.\n\nCons:\n-The proposed algorithm is a trivial architecture change to the conv encoder, the introduced novelty is limited. Injecting the temporal difference inductive bias obviously will be beneficial, given that the sole reason for frame stacking is to infer velocity and acceleration. \n-In Fig 3 the authors chose to only use 2 consecutive frames for State SAC, while a common practice is to use 3 frames. Using only 2 consecutive frames is not enough to infer acceleration and thus it is not a realistic setup, which makes this comparison meaningless. Also given that the variance is pretty high here, comparing performance over just 3 random seeds is not statistically conclusive. \n-In Fig 6 the method is only evaluated only on 5 seeds and given that it demonstrates very high variance on the 3 tasks (out of 5) where it outperforms RAD it makes me think that the performance improvements are marginal and not worth introducing complexity. \n-The ablation study is not very illuminating, this partially comes from the fact that the results are inconclusive (due to the high error bars), and partially because the experiments themselves are not very interesting. \n\n\nQuality:\nWhile the paper is well executed and made it significantly lacks on novelty and significance fronts.\n\nClarity:\nThe paper in general is clearly written and well organized.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Frame stacking that is done on image embeddings", "review": "Significance:\nThe paper brings very little novelty or insight. It is unclear that the introduced architecture complexity worth marginal improvements (given high variance and only 5 random seeds) on 2 out of 11 tasks (5 from the main paper and 6 from appendix). This might be a good workshop paper but it clearly does not meet the high acceptance threshold of ICLR.\n\nPros:\n-A simple architecture modification that might be beneficial to impose a stronger inductive bias on temporal dynamics.\n\nCons:\n-The proposed algorithm is a trivial architecture change to the conv encoder, the introduced novelty is limited. Injecting the temporal difference inductive bias obviously will be beneficial, given that the sole reason for frame stacking is to infer velocity and acceleration. \n-In Fig 3 the authors chose to only use 2 consecutive frames for State SAC, while a common practice is to use 3 frames. Using only 2 consecutive frames is not enough to infer acceleration and thus it is not a realistic setup, which makes this comparison meaningless. Also given that the variance is pretty high here, comparing performance over just 3 random seeds is not statistically conclusive. \n-In Fig 6 the method is only evaluated only on 5 seeds and given that it demonstrates very high variance on the 3 tasks (out of 5) where it outperforms RAD it makes me think that the performance improvements are marginal and not worth introducing complexity. \n-The ablation study is not very illuminating, this partially comes from the fact that the results are inconclusive (due to the high error bars), and partially because the experiments themselves are not very interesting. \n\n\nQuality:\nWhile the paper is well executed and made it significantly lacks on novelty and significance fronts.\n\nClarity:\nThe paper in general is clearly written and well organized.\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603914252457}, {"id": "tIvuQp_bsKL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1855/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis work presents a simple technique (Flare) to incorporate explicit temporal information to enable effective RL policy learning in challenging continuous control environments using pixel-based state representations. The approach is inspired from recent advances in the video recognition approaches which employ optical flow information and late fusion to incorporate temporal information. Typically, RL algorithms employ a frame-stacking heuristic to incorporate temporal information (early fusion). Though computing optical flow is slow and can been prohibitive for real-time applications, the authors present a simple alternative to using optical flow, i.e. difference of latent state vectors as a proxy for explicitly encoding motion information with latent vectors representing the observed state (late fusion). Experimental results on challenging continuous control tasks in the DMControl Suite show that Flare can achieve up to 1.9x higher scores than a baseline algorithm (RAD) which uses the frame-stacking heuristic to incorporate temporal information.\n\n########################\n\nPros:\n- The presented approach provides an effective alternative to the frame stacking heuristic for incorporating temporal information in RL with pixel-based state representations. The presented methodology (concatenation of latent state vector differences) thus learns effective policies on challenging continuous control environments.\n- The presented approach can easily modify any RL algorithm operating on pixel-based state representations to encode temporal information. \n- The paper is well-written, clear and easy to follow.\n- The idea is well-motivated with experiments in environments with low-dimensional state spaces. The results from the motivation section show the importance of temporal information, and specifically explicit temporal information to learn effective policies.\n- Ablation study clearly highlights the merits of including explicit temporal information with late fusion ruling out other techniques like pixel-based flow (early fusion) and the effect of independent convolutional feature extraction for different image frames.\n\n########################\n\nCons:\n- Comparisons to other approaches mentioned in Section 2 are missing.  For example, how would Flare compare in performance and sample efficiency to LSTM based RL methods (such as those under the “neural network architectures” subsection listed under Section 2)?\n- Flare does not outperform RAD in all environments (such as hopper hop and walker run). It is unclear why Flare works well on some environments and not so on others. In fact, Flare performs worse than RAD on walker run. Why so?\n\n\n########################\n\nReason for score:\nThe approach is well-motivated, and the paper is clearly written. The experiments comparing with an early fusion approach (RAD) and related ablative analysis highlighting that explicit late fusion of temporal information is key to improved performance are well-done and prove the effectiveness of the Flare. However, comparison to other approaches incorporating explicit temporal information for RL is missing (eg: LSTM based approaches).
\n\n########################\n\nQuestions during rebuttal:\n- Please refer to questions in the Cons section and other feedback\n- For results from Figure 6, why does Flare not outperform RAD for the hopper hop and walker run environments? Is there a way to visualize the temporal information to further investigate why Flare outperforms RAD in some environments (like quadruped walk) and not on others?\n\n########################\n\nSome typos and other feedback:\n- Figure 2 and Section 5, paragraph 1, sentence 3: What is meant by proprioceptive state input? Consider formally defining it in the text for the reader.\n- Section 5, paragraph 1, sentence 3: Consider ending the sentence by stating the fact that the experiments suggest the alternative to frame stacking heuristic is also effective in terms of performance.\n- Consider defining p(a_t+1|a_t,o_t) as the transition function in the text for the reader.\n- Section 5.1, last sentence: “… are done in the same except with augmented observations.” -> “… are done in the same way except with augmented observations.”\n- Section 6, paragraph 1, sentence 1: “… that are experiments focus on.” -> “… that our experiments focus on.”\n- Section 6.2, sentence 2:  You seem to have forgotten to mention the environments for which Flare outperforms RAD. What are you referring to with the phrase “remaining environments”?\n- Section 6.2, sentence 5: “… walker run shown visualized in Figure 6.” -> “… walker run as shown in Figure 6.“\n- Section 6.2, Figure 6: Why does Flare not perform as well as RAD for the Walker run environment?\n- Conclusion, last sentence: Consider replacing “We would like to integrate Flare with model-based RL in the future” with “Integrating Flare with model-based RL is a potential direction for future work.”\n- Consider replacing the usage of the phrase “state-based RL” with “RL on low-dimensional state space”.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #3", "review": "Summary:\nThis work presents a simple technique (Flare) to incorporate explicit temporal information to enable effective RL policy learning in challenging continuous control environments using pixel-based state representations. The approach is inspired from recent advances in the video recognition approaches which employ optical flow information and late fusion to incorporate temporal information. Typically, RL algorithms employ a frame-stacking heuristic to incorporate temporal information (early fusion). Though computing optical flow is slow and can been prohibitive for real-time applications, the authors present a simple alternative to using optical flow, i.e. difference of latent state vectors as a proxy for explicitly encoding motion information with latent vectors representing the observed state (late fusion). Experimental results on challenging continuous control tasks in the DMControl Suite show that Flare can achieve up to 1.9x higher scores than a baseline algorithm (RAD) which uses the frame-stacking heuristic to incorporate temporal information.\n\n########################\n\nPros:\n- The presented approach provides an effective alternative to the frame stacking heuristic for incorporating temporal information in RL with pixel-based state representations. The presented methodology (concatenation of latent state vector differences) thus learns effective policies on challenging continuous control environments.\n- The presented approach can easily modify any RL algorithm operating on pixel-based state representations to encode temporal information. \n- The paper is well-written, clear and easy to follow.\n- The idea is well-motivated with experiments in environments with low-dimensional state spaces. The results from the motivation section show the importance of temporal information, and specifically explicit temporal information to learn effective policies.\n- Ablation study clearly highlights the merits of including explicit temporal information with late fusion ruling out other techniques like pixel-based flow (early fusion) and the effect of independent convolutional feature extraction for different image frames.\n\n########################\n\nCons:\n- Comparisons to other approaches mentioned in Section 2 are missing.  For example, how would Flare compare in performance and sample efficiency to LSTM based RL methods (such as those under the “neural network architectures” subsection listed under Section 2)?\n- Flare does not outperform RAD in all environments (such as hopper hop and walker run). It is unclear why Flare works well on some environments and not so on others. In fact, Flare performs worse than RAD on walker run. Why so?\n\n\n########################\n\nReason for score:\nThe approach is well-motivated, and the paper is clearly written. The experiments comparing with an early fusion approach (RAD) and related ablative analysis highlighting that explicit late fusion of temporal information is key to improved performance are well-done and prove the effectiveness of the Flare. However, comparison to other approaches incorporating explicit temporal information for RL is missing (eg: LSTM based approaches).
\n\n########################\n\nQuestions during rebuttal:\n- Please refer to questions in the Cons section and other feedback\n- For results from Figure 6, why does Flare not outperform RAD for the hopper hop and walker run environments? Is there a way to visualize the temporal information to further investigate why Flare outperforms RAD in some environments (like quadruped walk) and not on others?\n\n########################\n\nSome typos and other feedback:\n- Figure 2 and Section 5, paragraph 1, sentence 3: What is meant by proprioceptive state input? Consider formally defining it in the text for the reader.\n- Section 5, paragraph 1, sentence 3: Consider ending the sentence by stating the fact that the experiments suggest the alternative to frame stacking heuristic is also effective in terms of performance.\n- Consider defining p(a_t+1|a_t,o_t) as the transition function in the text for the reader.\n- Section 5.1, last sentence: “… are done in the same except with augmented observations.” -> “… are done in the same way except with augmented observations.”\n- Section 6, paragraph 1, sentence 1: “… that are experiments focus on.” -> “… that our experiments focus on.”\n- Section 6.2, sentence 2:  You seem to have forgotten to mention the environments for which Flare outperforms RAD. What are you referring to with the phrase “remaining environments”?\n- Section 6.2, sentence 5: “… walker run shown visualized in Figure 6.” -> “… walker run as shown in Figure 6.“\n- Section 6.2, Figure 6: Why does Flare not perform as well as RAD for the Walker run environment?\n- Conclusion, last sentence: Consider replacing “We would like to integrate Flare with model-based RL in the future” with “Integrating Flare with model-based RL is a potential direction for future work.”\n- Consider replacing the usage of the phrase “state-based RL” with “RL on low-dimensional state space”.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603885981344}], "openreview_url": "https://openreview.net/forum?id=lSijhyKKsct", "arxiv_id": "2101.01857", "paper_pdf": "papers/lSijhyKKsct.pdf", "paper_pdf_sha256": "70d550e0b1b25c8e21133bd3eff77ccfbeba1142358c5a978d939116b5fb0930", "paper_pdf_bytes": 3646630, "paper_pdf_source": "openreview", "code_url": "https://github.com/WendyShang/flare", "code_repository": "WendyShang/flare", "code_commit": "ef12a7f4b2fc639d491ca63648ea1e8c58d07263", "code_archive": "repos/lSijhyKKsct.zip", "code_archive_sha256": "6e9ab16c900e97ffcb6a910988b27dd95e36ba28741ee096defb9bf3b355b2ee", "code_archive_bytes": 26182, "code_file_count": 12, "code_extensions": {".py": 8, ".sh": 4}, "github_disk_usage_kb": 31, "github_languages": {"Python": 78060, "Shell": 2529}, "github_archived": false, "github_pushed_at": "2021-03-25T20:56:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/reinforcement-learning-with-latent-flow-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1l1myStwr", "year": 2020, "status": "rejected", "title": "Continuous Meta-Learning without Tasks", "authors": ["James Harrison", "Apoorva Sharma", "Chelsea Finn", "Marco Pavone"], "authorids": ["jharrison@stanford.edu", "apoorva@stanford.edu", "cbfinn@cs.stanford.edu", "pavone@stanford.edu"], "authors_source": "OpenReview API", "abstract": "Meta-learning is a promising strategy for learning to efficiently learn within new tasks, using data gathered from a distribution of tasks. However, the meta-learning literature thus far has focused on the task segmented setting, where at train-time, offline data is assumed to be split according to the underlying task, and at test-time, the algorithms are optimized to learn in a single task. In this work, we enable the application of generic meta-learning algorithms to settings where this task segmentation is unavailable, such as continual online learning with a time-varying task. We present meta-learning via online changepoint analysis (MOCA), an approach which augments a meta-learning algorithm with a differentiable Bayesian changepoint detection scheme. The framework allows both training and testing directly on time series data without segmenting it into discrete tasks. We demonstrate the utility of this approach on a nonlinear meta-regression benchmark as well as two meta-image-classification benchmarks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ByeuiSECKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1602/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a solution for using meta-learning methods without the need of determining the task segmentation a priori. Authors identify that the task segmented setting is very similar to the problem of change-point detection (CPD) and particularly, they connect the generative model of the meta-learning approach to the Bayesian recursive method of Adams and MacKay 2007. The presentation of the meta-learning problem is well done and I understand its importance within the difficulties for modelling new unobserved tasks. The notation and description of the generative model included in the problem statement section is clearly understandable for any reader, this is a very positive point. They demonstrate a deep comprehension of the BOCPD model of Adams, and its extension for the meta-learning approach is original to me. Lastly, the presentation of the MOCA meta-learning algorithm is useful for reproducibility and I see it completely applicable to other scenarios. There is a noticeable effort of describing the solution for both regression and classification problems and the empirical results conclude a positive performance of MOCA.\n\nOverall, I consider that the paper is well-written with thorough explanations. Of course, there are some details that could be improved (I will comment this later). If done, I would be willing to increase my score. The contribution of the paper is significant to meta-learning models and I think it will help to spread the Adams’ model to other type of problems.\n\nThere is an important contribution in the paper that I have to mention and it is also relevant for the future application of the Adams model. If one reads the original BOCPD paper, in particular Eq (1), where the predictive posterior p(x_t+1|x_{1:t}) is defined from a marginalisation over the run length values, it is noticeable that this equation is not used in the final recursion of the CPD method. This is because the posterior p(r_t|x_{1:t}) is sufficient for determining if there is a CP on the t time-step or not. So the predictive posterior is never used in practice (in the original paper). When I first read Adams 2007, this detail was clear to me. Surprisingly, I find that the authors have find a practical use of this equation (Eq. (7)) and it is in the main core of the meta-learning algorithm, this is fantastic.\n\nThe details I think should be improved are:\n\n- The conditional run-length prior p(r_t|r_{t-1}) barely appears and without any detailed description is not obvious for non-familiar readers with the Bayesian CPD approach.\n- The first time I read the manuscript, the use of z_t in the BOCPD presentation made me feel a bit lost. Why not use x or y? At least specify that is a toy variable for the explanation. Later on, authors change again to the x,y notation.\n- Equations 3 and 4 are too similar, this looks a bit repetitive. Why not reusing one of them or say the change from one term to another? \n- If reading the literature of the BOCPD and posterior extensions, the likelihood model of the detector is often referred as the underlying predictive model (UPM). Using a similar term would help for orienting familiar readers into the solution.\n- As authors should have noted on their experiments, as one makes the run-length higher, the number of parameters \\eta[r_t] increases. A clear state on how this is solved would help.\n- In the PCOC subsection, I find the definition of y \\sim q(y) a bit weird. Using Cat(_) or Multinomial likelihood notation would be a bit better.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes a solution for using meta-learning methods without the need of determining the task segmentation a priori. Authors identify that the task segmented setting is very similar to the problem of change-point detection (CPD) and particularly, they connect the generative model of the meta-learning approach to the Bayesian recursive method of Adams and MacKay 2007. The presentation of the meta-learning problem is well done and I understand its importance within the difficulties for modelling new unobserved tasks. The notation and description of the generative model included in the problem statement section is clearly understandable for any reader, this is a very positive point. They demonstrate a deep comprehension of the BOCPD model of Adams, and its extension for the meta-learning approach is original to me. Lastly, the presentation of the MOCA meta-learning algorithm is useful for reproducibility and I see it completely applicable to other scenarios. There is a noticeable effort of describing the solution for both regression and classification problems and the empirical results conclude a positive performance of MOCA.\n\nOverall, I consider that the paper is well-written with thorough explanations. Of course, there are some details that could be improved (I will comment this later). If done, I would be willing to increase my score. The contribution of the paper is significant to meta-learning models and I think it will help to spread the Adams’ model to other type of problems.\n\nThere is an important contribution in the paper that I have to mention and it is also relevant for the future application of the Adams model. If one reads the original BOCPD paper, in particular Eq (1), where the predictive posterior p(x_t+1|x_{1:t}) is defined from a marginalisation over the run length values, it is noticeable that this equation is not used in the final recursion of the CPD method. This is because the posterior p(r_t|x_{1:t}) is sufficient for determining if there is a CP on the t time-step or not. So the predictive posterior is never used in practice (in the original paper). When I first read Adams 2007, this detail was clear to me. Surprisingly, I find that the authors have find a practical use of this equation (Eq. (7)) and it is in the main core of the meta-learning algorithm, this is fantastic.\n\nThe details I think should be improved are:\n\n- The conditional run-length prior p(r_t|r_{t-1}) barely appears and without any detailed description is not obvious for non-familiar readers with the Bayesian CPD approach.\n- The first time I read the manuscript, the use of z_t in the BOCPD presentation made me feel a bit lost. Why not use x or y? At least specify that is a toy variable for the explanation. Later on, authors change again to the x,y notation.\n- Equations 3 and 4 are too similar, this looks a bit repetitive. Why not reusing one of them or say the change from one term to another? \n- If reading the literature of the BOCPD and posterior extensions, the likelihood model of the detector is often referred as the underlying predictive model (UPM). Using a similar term would help for orienting familiar readers into the solution.\n- As authors should have noted on their experiments, as one makes the run-length higher, the number of parameters \\eta[r_t] increases. A clear state on how this is solved would help.\n- In the PCOC subsection, I find the definition of y \\sim q(y) a bit weird. Using Cat(_) or Multinomial likelihood notation would be a bit better.\n"}, "tcdate": 1571861920444}, {"id": "HyxY4QkRYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1602/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper pushes meta-learning towards task-unsegmented settings. Different from the traditional offline meta-learning phase with explicit task segmentation, MOCA adopts a Bayesian changepoint estimation scheme for task change detection. The setting is novel and deserves research in-depth, and the idea is easy to understand. The proposed method can learn the meta-learning model and changepoint detection model simultaneously. Besides, the MOCA framework is not designed specifically for one algorithm and can be easily combined with other meta-learning models.\n\nHowever, I got some questions about this paper:\nQ1: Is the ‘Hazard’ in the experiment the same to $\\lambda$ in eq.1? I think notations should be consistent if they are the same.\nQ2: Can other changepoint models be compared in the experiment? I found many in the related work.\nQ3: The running times in Figure 2 should be reported to demonstrate the efficiency of MOCA, since the method is proposing online streams, efficiency should be promised for quickly processing.\nQ4: What is ‘T’ in Algorithm 1? Are experiment results sensitive to it? Experiments about this should be conducted and reported.\n\nLastly, I think [1] should be cited as related work about continual learning for proposing task-free continual learning, which is very similar to the setting in this paper.\n[1] Rahaf Aljundi, Klaas Kelchtermans, Tinne Tuytelaars. Task-Free Continual Learning. CVPR 2019", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper pushes meta-learning towards task-unsegmented settings. Different from the traditional offline meta-learning phase with explicit task segmentation, MOCA adopts a Bayesian changepoint estimation scheme for task change detection. The setting is novel and deserves research in-depth, and the idea is easy to understand. The proposed method can learn the meta-learning model and changepoint detection model simultaneously. Besides, the MOCA framework is not designed specifically for one algorithm and can be easily combined with other meta-learning models.\n\nHowever, I got some questions about this paper:\nQ1: Is the ‘Hazard’ in the experiment the same to $\\lambda$ in eq.1? I think notations should be consistent if they are the same.\nQ2: Can other changepoint models be compared in the experiment? I found many in the related work.\nQ3: The running times in Figure 2 should be reported to demonstrate the efficiency of MOCA, since the method is proposing online streams, efficiency should be promised for quickly processing.\nQ4: What is ‘T’ in Algorithm 1? Are experiment results sensitive to it? Experiments about this should be conducted and reported.\n\nLastly, I think [1] should be cited as related work about continual learning for proposing task-free continual learning, which is very similar to the setting in this paper.\n[1] Rahaf Aljundi, Klaas Kelchtermans, Tinne Tuytelaars. Task-Free Continual Learning. CVPR 2019"}, "tcdate": 1571840817092}, {"id": "B1eczZEItS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1602/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers the meta-learning in the task un-segmented setting and apply bayesian online change point detection with meta-learning. The task un-segmented is claimed to exist in real applications and the paper explains the idea in a clear way. \n\nMy major concerns and questions are the following:\n\n1) In Eq(4), it requires a computation of normalization constant that needs to sum over support of r_t. The support gets larger and larger with the length of sequence increases over time. Does this method scale well with long sequence?\n\n2) The part I feel most confused are the experiments.  The paper tries to solve a problem in the task un-segmented setting, but why there is no such setting in experiments that the meta-training set cannot be segmented to different tasks?\n\n3) The Figure 1 is hard to read. What do the red and green points in the left panel mean? The hazard rate is not defined before the experiment which makes Figure 3 also hard to understand. \n\n4) What is the meta-learning algorithm used in the experiements, MAML, NP or RNN-based methods?  It claims MOCA can be used with any meta-learning algorithms and how does it show in experiments?\n\n5) In Rainbow MNIST, why in high hazard rates, all models perform comparable to “train on everything”? Does it mean meta-learning does not work in this configuration and why? Does “train on everything” includes fine-tuning on the meta-test training set?\n\n6) In Mini-IMAGENET, what does it mean by \"we associate each class with a semantic label that is consistent between tasks”? Why it needs to form “super-class”? \n\nI think the proposed method can be useful in the task un-segmented setting. But before the above-mention questions are solved and the experimental section gets more clear, I will give a conservative rating. \n\n\n######################\nPost-rebuttal review:\nThanks for the authors' feedback and it resolves some confusion parts in the paper. Though the author claims the reason of experiments setting, I believe it is necessary to have one experiment where task-segmentation is impossible and compare with standard sequential learning methods in that setting. Otherwise it remains a question whether the proposed method works only when the problem basically has task-segmentation.  And if the problem has task-segmentation, why not using traditional meta-learning methods, as shown as oracle methods with better performance in paper? \nSince the major contribution of the paper is providing a meta-learning method to work in the problems where task-segmentation is unavailable, not having an experiment in this setting (withholding segmentation information does not exactly fall in this setting because it adds a condition that the task-segmentation information is originally accessible) is a major reason for my current evaluation.  I recognize the careful design of the MOCA and agree with some positive points raised by other reviewers. I would not be bothered if the paper is accepted while I tend to maintain the current rating because of the above-mentioned concern.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "The paper considers the meta-learning in the task un-segmented setting and apply bayesian online change point detection with meta-learning. The task un-segmented is claimed to exist in real applications and the paper explains the idea in a clear way. \n\nMy major concerns and questions are the following:\n\n1) In Eq(4), it requires a computation of normalization constant that needs to sum over support of r_t. The support gets larger and larger with the length of sequence increases over time. Does this method scale well with long sequence?\n\n2) The part I feel most confused are the experiments.  The paper tries to solve a problem in the task un-segmented setting, but why there is no such setting in experiments that the meta-training set cannot be segmented to different tasks?\n\n3) The Figure 1 is hard to read. What do the red and green points in the left panel mean? The hazard rate is not defined before the experiment which makes Figure 3 also hard to understand. \n\n4) What is the meta-learning algorithm used in the experiements, MAML, NP or RNN-based methods?  It claims MOCA can be used with any meta-learning algorithms and how does it show in experiments?\n\n5) In Rainbow MNIST, why in high hazard rates, all models perform comparable to “train on everything”? Does it mean meta-learning does not work in this configuration and why? Does “train on everything” includes fine-tuning on the meta-test training set?\n\n6) In Mini-IMAGENET, what does it mean by \"we associate each class with a semantic label that is consistent between tasks”? Why it needs to form “super-class”? \n\nI think the proposed method can be useful in the task un-segmented setting. But before the above-mention questions are solved and the experimental section gets more clear, I will give a conservative rating. \n\n\n######################\nPost-rebuttal review:\nThanks for the authors' feedback and it resolves some confusion parts in the paper. Though the author claims the reason of experiments setting, I believe it is necessary to have one experiment where task-segmentation is impossible and compare with standard sequential learning methods in that setting. Otherwise it remains a question whether the proposed method works only when the problem basically has task-segmentation.  And if the problem has task-segmentation, why not using traditional meta-learning methods, as shown as oracle methods with better performance in paper? \nSince the major contribution of the paper is providing a meta-learning method to work in the problems where task-segmentation is unavailable, not having an experiment in this setting (withholding segmentation information does not exactly fall in this setting because it adds a condition that the task-segmentation information is originally accessible) is a major reason for my current evaluation.  I recognize the careful design of the MOCA and agree with some positive points raised by other reviewers. I would not be bothered if the paper is accepted while I tend to maintain the current rating because of the above-mentioned concern.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571336465701}], "openreview_url": "https://openreview.net/forum?id=r1l1myStwr", "arxiv_id": "1912.08866", "paper_pdf": "papers/r1l1myStwr.pdf", "paper_pdf_sha256": "ed3077a4b4326f6ee496d6bf9e93a6af5679dc3899969895f02e3d55e19b6b28", "paper_pdf_bytes": 1479739, "paper_pdf_source": "openreview", "code_url": "https://github.com/StanfordASL/moca", "code_repository": "StanfordASL/moca", "code_commit": "a4eb2d950b9a54e221786734665ed5d7d9e34646", "code_archive": "repos/r1l1myStwr.zip", "code_archive_sha256": "33922c7fadac4a01fc61b9c5c2e34c4e9fcd6786565e555f57fb22c72258b83b", "code_archive_bytes": 19385, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 35, "github_languages": {"Python": 60493}, "github_archived": false, "github_pushed_at": "2022-11-22T04:38:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continuous-meta-learning-without-tasks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1gd7nCcF7", "year": 2019, "status": "rejected", "title": "Self-Supervised Generalisation with Meta Auxiliary Learning", "authors": ["Shikun Liu", "Edward Johns", "Andrew Davison"], "authorids": ["shikun.liu17@imperial.ac.uk", "e.johns@imperial.ac.uk", "a.davison@imperial.ac.uk"], "authors_source": "OpenReview API", "abstract": "Auxiliary learning has been shown to improve the generalisation performance of a principal task. But typically, this requires manually-defined auxiliary tasks based on domain knowledge. In this paper, we consider that it may be possible to automatically learn these auxiliary tasks to best suit the principal task, towards optimum auxiliary tasks without any human knowledge. We propose a novel method, Meta Auxiliary Learning (MAXL), which we design for the task of image classification, where the auxiliary task is hierarchical sub-class image classification. The role of the meta learner is to determine sub-class target labels to train a multi-task evaluator, such that these labels improve the generalisation performance on the principal task. Experiments on three different CIFAR datasets show that MAXL outperforms baseline auxiliary learning methods, and is competitive even with a method which uses human-defined sub-class hierarchies. MAXL is self-supervised and general, and therefore offers a promising new direction towards automated generalisation.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rJl-0ufypm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1372/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an algorithm for auxiliary learning. Given a target prediction task to be learned on training data, the auxiliary learning utilizes external training data to improve learning. The authors focus on a setup where both target and external training data come from the same distribution but differ in class labels, where each class in the target data is a set of finer-grained classes in the auxiliary data. The authors propose a heuristic for learning from both data sets through minimization of a joint loss function. The experimental results show that the proposed methods works well on this particular setup on CIFAR data set.\n\nStrengths:\n+ a new auxiliary learning algorithm\n+ positive results on CIFAR data set\n\nWeaknesses:\n- novelty is low: the proposed algorithm is a heuristic similar to previously proposed algorithms in the transfer learning and auxiliary learning space\n- there is no attempt to provide a theoretical insight into the performance of the algorithm\n- the problem assumptions are too simplistic and unrealistic (feature distributions of target and auxiliary data are identical), so it is questionable if the proposed algorithm has practical importance\n- experiments are performed using a synthetic setup on a single data set, so it remains unclear if the algorithm would be successful in a real life scenario\n- the paper is poorly written and sentences are generally very hard to parse. For example, section 3.1 is opened by statements such as \"(we use) a multi-task evaluator which trains on the principal and auxiliary tasks, and evaluates the performance of the auxiliary tasks on a meta set\"??", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is an incremental contribution for an artificially sounding problem", "review": "This paper proposes an algorithm for auxiliary learning. Given a target prediction task to be learned on training data, the auxiliary learning utilizes external training data to improve learning. The authors focus on a setup where both target and external training data come from the same distribution but differ in class labels, where each class in the target data is a set of finer-grained classes in the auxiliary data. The authors propose a heuristic for learning from both data sets through minimization of a joint loss function. The experimental results show that the proposed methods works well on this particular setup on CIFAR data set.\n\nStrengths:\n+ a new auxiliary learning algorithm\n+ positive results on CIFAR data set\n\nWeaknesses:\n- novelty is low: the proposed algorithm is a heuristic similar to previously proposed algorithms in the transfer learning and auxiliary learning space\n- there is no attempt to provide a theoretical insight into the performance of the algorithm\n- the problem assumptions are too simplistic and unrealistic (feature distributions of target and auxiliary data are identical), so it is questionable if the proposed algorithm has practical importance\n- experiments are performed using a synthetic setup on a single data set, so it remains unclear if the algorithm would be successful in a real life scenario\n- the paper is poorly written and sentences are generally very hard to parse. For example, section 3.1 is opened by statements such as \"(we use) a multi-task evaluator which trains on the principal and auxiliary tasks, and evaluates the performance of the auxiliary tasks on a meta set\"??", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541511369067}, {"id": "Bkxpr4aq3m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1372/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a self-auxiliary-training method that aims to improve the generalization performance of simple supervised learning. The basic idea is to train the classification network to predict fine-level auxiliary labels in addition to the ground-truth coarse label, where the auxiliary labels used in training is generated by a generator network. During training, the classification network and the generator network are alternatively updated, and the update of the latter aims to maximize the improvement of the former after using the generated auxiliary label for training. The method requires a class hierarchy in advance to define the binary mask applied to the output layer for auxiliary class prediction. A KL divergence term is attached to the optimization objective to avoid generating trivial and collapsing auxiliary classes.\n\nPros:\n\n1) The main idea is simple and easy to understand.\n2) It discusses the class collapsing problem in generating pseudo (auxiliary) labels and provides a reasonable solution, i.e., using KL divergence as regularization.\n3) Uses several visualizations to show experimental results.\n\nCons:\n\n1) The problem it aims to solve is neither multi-task learning nor meta-learning: it tries to solve a supervised classification problem defined on principle classes, with the help of simultaneously predicting/generating auxiliary class labels. Although the concept of \"task\" is not explicitly defined in this paper, the authors seem to associate each task with a specific class. This is not correct: in meta-learning, each task is a subset of classes drawn from a ground set of classes, and different tasks are independently sampled. In addition, the classification models for different tasks are independent, though their training might be related by a meta-learner. Hence, the claims in multiple places of this paper and the names for the two networks are misleading.\n\n2) At the end of Page 4, the authors show that the update of the generator only depends on the improvement of the classifier after using the auxiliary label for training. In fact, the optimal auxiliary labels minimizing the objective is the ground truth label for principle classes. This results in the class collapsing problem observed by the authors. The KL divergence regularization introduces extra randomness to the auxiliary labels and thus mitigates the problem, but it hardly provides any useful information except randomness. In other words, the auxiliary labels for a specific principle class are very possible to be multiple noisy copies of the principal label with random perturbations. So it is not convincing to me that the auxiliary labels generated by the generator can be really helpful. My conjecture is that the observed improvements are mainly due to the softness of the auxiliary labels, which has been proved by model compression/knowledge distillation and recent \"born-again neural networks\". To verify this, the authors might need to compare the results with those methods (which use the generated soft probability of ground truth classes for training), and the \"random-noisy copies of soft principle label\" mentioned above.\n\n3) The experiments lack comparisons to several important baselines from self-supervised learning community, and methods using soft labels for training (as mentioned in 2) above). A successful idea of self-supervised learning is to use the output feature map of the trained classification network to generate auxiliary training signals, since it provides extra information about the learned distance beyond the ground-truth labels. The authors might want to compare to \"Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep Clustering for Unsupervised Learning of Visual Features. ECCV 2018.\" and \"Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. ICCV 2017.\" Moreover, since the method is not a meta-learning approach for few-shot learning, it is not fair and also not appropriate to compare with Prototypical Network.\n\n4) Although the paper claims that the ground truth fine labels are not required, it requires a class hierarchy, which in the experiments are provided by the dataset and defined between true coarse and fine classes. In practice, such hierarchy might be much harder to achieve than the primary (coarse) labels, and might be as costly to obtain as the true fine-class labels. This weakens the feasibility of the proposed method.\n\n5) The experiments only test the proposed method on CIFAR100 and CIFAR10, which has at most 100 fine classes. It is necessary to test it on datasets with much more fine classes and much-complicated hierarchy, e.g., ImageNet, MS COCO or their subsets, which have ideal class hierarchy structures.\n\nMinor comments:\n\nSome important equations in the paper should be numbered. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not accurate to call it meta-learning, the auxiliary labels might bring few helpful information, lacking comparisons to several important baselines and benchmark datasets.", "review": "This paper proposes a self-auxiliary-training method that aims to improve the generalization performance of simple supervised learning. The basic idea is to train the classification network to predict fine-level auxiliary labels in addition to the ground-truth coarse label, where the auxiliary labels used in training is generated by a generator network. During training, the classification network and the generator network are alternatively updated, and the update of the latter aims to maximize the improvement of the former after using the generated auxiliary label for training. The method requires a class hierarchy in advance to define the binary mask applied to the output layer for auxiliary class prediction. A KL divergence term is attached to the optimization objective to avoid generating trivial and collapsing auxiliary classes.\n\nPros:\n\n1) The main idea is simple and easy to understand.\n2) It discusses the class collapsing problem in generating pseudo (auxiliary) labels and provides a reasonable solution, i.e., using KL divergence as regularization.\n3) Uses several visualizations to show experimental results.\n\nCons:\n\n1) The problem it aims to solve is neither multi-task learning nor meta-learning: it tries to solve a supervised classification problem defined on principle classes, with the help of simultaneously predicting/generating auxiliary class labels. Although the concept of \"task\" is not explicitly defined in this paper, the authors seem to associate each task with a specific class. This is not correct: in meta-learning, each task is a subset of classes drawn from a ground set of classes, and different tasks are independently sampled. In addition, the classification models for different tasks are independent, though their training might be related by a meta-learner. Hence, the claims in multiple places of this paper and the names for the two networks are misleading.\n\n2) At the end of Page 4, the authors show that the update of the generator only depends on the improvement of the classifier after using the auxiliary label for training. In fact, the optimal auxiliary labels minimizing the objective is the ground truth label for principle classes. This results in the class collapsing problem observed by the authors. The KL divergence regularization introduces extra randomness to the auxiliary labels and thus mitigates the problem, but it hardly provides any useful information except randomness. In other words, the auxiliary labels for a specific principle class are very possible to be multiple noisy copies of the principal label with random perturbations. So it is not convincing to me that the auxiliary labels generated by the generator can be really helpful. My conjecture is that the observed improvements are mainly due to the softness of the auxiliary labels, which has been proved by model compression/knowledge distillation and recent \"born-again neural networks\". To verify this, the authors might need to compare the results with those methods (which use the generated soft probability of ground truth classes for training), and the \"random-noisy copies of soft principle label\" mentioned above.\n\n3) The experiments lack comparisons to several important baselines from self-supervised learning community, and methods using soft labels for training (as mentioned in 2) above). A successful idea of self-supervised learning is to use the output feature map of the trained classification network to generate auxiliary training signals, since it provides extra information about the learned distance beyond the ground-truth labels. The authors might want to compare to \"Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep Clustering for Unsupervised Learning of Visual Features. ECCV 2018.\" and \"Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. ICCV 2017.\" Moreover, since the method is not a meta-learning approach for few-shot learning, it is not fair and also not appropriate to compare with Prototypical Network.\n\n4) Although the paper claims that the ground truth fine labels are not required, it requires a class hierarchy, which in the experiments are provided by the dataset and defined between true coarse and fine classes. In practice, such hierarchy might be much harder to achieve than the primary (coarse) labels, and might be as costly to obtain as the true fine-class labels. This weakens the feasibility of the proposed method.\n\n5) The experiments only test the proposed method on CIFAR100 and CIFAR10, which has at most 100 fine classes. It is necessary to test it on datasets with much more fine classes and much-complicated hierarchy, e.g., ImageNet, MS COCO or their subsets, which have ideal class hierarchy structures.\n\nMinor comments:\n\nSome important equations in the paper should be numbered. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541227589021}, {"id": "rylvbv_93Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1372/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe role of auxiliary tasks is to improve the generalization performance of the principal task of interest. So far, hand-crafted auxiliary tasks are generated, tailored for a problem of interest. The current work addresses a meta-learning approach to automatically generate auxiliary tasks suited to the principal task, without human knowledge.  The key components of the method are: (1) meta-generator; (2) multi-task evaluator. These two models are trained using the gradient-based meta-learning technique (for instance, MAML).  The problem of image classification is considered only, while authors claimed the method can be easily applied to other problems as well. \n\nStrengths:\n- To my best knowledge, the idea of applying the meta-learning to the automatic generation of auxiliary tasks is novel. \n- The paper is well written and easy to read.\n- The method nicely blends a few components such as self-supervised learning, meta-learning, auxiliary tasks into a single model to tackle the meta auxiliary learning. \n\nWeakness:\n- The performance gain is not substantial in experiments. I would like to suggest to use the state-of-the-arts classifier for the principal task and to evaluate how much gain your method can get with the help of auxiliary tasks. You can refer to the state-of-the-arts performance on CIFAR.\n- If the information on the hierarchy of sub-categories is not available, it will be an annoying hyperparameters that should be well tuned.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting idea for applying meta-learning to a problem of learning auxiliary tasks in a self-supervised fashion", "review": "Summary:\nThe role of auxiliary tasks is to improve the generalization performance of the principal task of interest. So far, hand-crafted auxiliary tasks are generated, tailored for a problem of interest. The current work addresses a meta-learning approach to automatically generate auxiliary tasks suited to the principal task, without human knowledge.  The key components of the method are: (1) meta-generator; (2) multi-task evaluator. These two models are trained using the gradient-based meta-learning technique (for instance, MAML).  The problem of image classification is considered only, while authors claimed the method can be easily applied to other problems as well. \n\nStrengths:\n- To my best knowledge, the idea of applying the meta-learning to the automatic generation of auxiliary tasks is novel. \n- The paper is well written and easy to read.\n- The method nicely blends a few components such as self-supervised learning, meta-learning, auxiliary tasks into a single model to tackle the meta auxiliary learning. \n\nWeakness:\n- The performance gain is not substantial in experiments. I would like to suggest to use the state-of-the-arts classifier for the principal task and to evaluate how much gain your method can get with the help of auxiliary tasks. You can refer to the state-of-the-arts performance on CIFAR.\n- If the information on the hierarchy of sub-categories is not available, it will be an annoying hyperparameters that should be well tuned.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541207807471}], "openreview_url": "https://openreview.net/forum?id=S1gd7nCcF7", "arxiv_id": "1901.08933", "paper_pdf": "papers/S1gd7nCcF7.pdf", "paper_pdf_sha256": "56be89bca044ac69b543e91267024ab3b9d39fb5e4968b2ea8e740f9459c0370", "paper_pdf_bytes": 17204922, "paper_pdf_source": "openreview", "code_url": "https://github.com/lorenmt/maxl", "code_repository": "lorenmt/maxl", "code_commit": "266f5b68df0ecdffc73da40e003fc6d58cceef96", "code_archive": "repos/S1gd7nCcF7.zip", "code_archive_sha256": "1e5e13b5ce774cea22679d9a3d5cd47639e4bc7c64d1ff18548461619e89c083", "code_archive_bytes": 18093, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 111, "github_languages": {"Python": 59824}, "github_archived": false, "github_pushed_at": "2021-12-19T20:46:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/self-supervised-generalisation-with-meta"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJUOHGWRb", "year": 2018, "status": "rejected", "title": "Contextual Explanation Networks", "authors": ["Maruan Al-Shedivat", "Avinava Dubey", "Eric P. Xing"], "authorids": ["alshedivat@cs.cmu.edu", "akdubey@cs.cmu.edu", "epxing@cs.cmu.edu"], "authors_source": "OpenReview API", "abstract": "We introduce contextual explanation networks (CENs)---a class of models that learn to predict by generating and leveraging intermediate explanations. CENs are deep networks that generate parameters for context-specific probabilistic graphical models which are further used for prediction and play the role of explanations. Contrary to the existing post-hoc model-explanation tools, CENs learn to predict and to explain jointly. Our approach offers two major advantages: (i) for each prediction, valid instance-specific explanations are generated with no computational overhead and (ii) prediction via explanation acts as a regularization and boosts performance in low-resource settings. We prove that local approximations to the decision boundary of our networks are consistent with the generated explanations. Our results on image and text classification and survival analysis tasks demonstrate that CENs are competitive with the state-of-the-art while offering additional insights behind each prediction, valuable for decision support.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1E57a-ZG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper836/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an interesting combination of neural nets and graphical models by using a deep neural net to predict the parameters of a graphical model. When the nets are trained on contexts \"C\" (e.g. satellite images associated with a neighborhood) related to an input \"X\" (e.g. categorical features describing the neighborhood); and a graphical model relates \"X\" to targets \"Y\" (e.g. binary variable encoding poverty level of the neighborhood), then the proposed combination can produce interpretable explanations for its predictions.\nThis approach compares favorably with post-hoc explanation methods like LIME in experiments on conducted on images (CIFAR-10, MNIST), text (IMDB reviews) and time series (Satellite dataset). The paper is clearly written and might inspire follow-up work in other applications. The description of related work is sparse (beyond a derivation of an equivalence with LIME in some settings, explained in the appendix).\nThe experiments study interesting effects: what happens when the model relating X and Y is degraded (e.g. by introducing noise into X, or sub-selecting X). The paper can be substantially improved by studying the effect of dictionary size and sparsity regularization more thoroughly.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach to combine neural nets and graphical models; can improve by elaborating the contrast with related work", "rating": "6: Marginally above acceptance threshold", "review": "The paper proposes an interesting combination of neural nets and graphical models by using a deep neural net to predict the parameters of a graphical model. When the nets are trained on contexts \"C\" (e.g. satellite images associated with a neighborhood) related to an input \"X\" (e.g. categorical features describing the neighborhood); and a graphical model relates \"X\" to targets \"Y\" (e.g. binary variable encoding poverty level of the neighborhood), then the proposed combination can produce interpretable explanations for its predictions.\nThis approach compares favorably with post-hoc explanation methods like LIME in experiments on conducted on images (CIFAR-10, MNIST), text (IMDB reviews) and time series (Satellite dataset). The paper is clearly written and might inspire follow-up work in other applications. The description of related work is sparse (beyond a derivation of an equivalence with LIME in some settings, explained in the appendix).\nThe experiments study interesting effects: what happens when the model relating X and Y is degraded (e.g. by introducing noise into X, or sub-selecting X). The paper can be substantially improved by studying the effect of dictionary size and sparsity regularization more thoroughly.\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1512326028202}, {"id": "H1wsCJjez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper836/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "the paper is clearly written; it works on a popular idea of combining graphical models and neural nets.\n\nthis work could benefit from differentiating more from previous literature.\n\none key component is interpretability, which comes from the use of graphical models.  the authors claim that the previous art directly integrate neural networks into the graphical models as components, which renders the models uninterpretable. however, it is unclear, following the same logic, why the proposed method has interpretability. after all, how to go from the context to the parameters of the graphical models is still uninterpretable. specifically, it is helpful to pinpoint what is special in this model that makes it interpretable, compared to works like Gao, Y., Archer, E. W., Paninski, L., & Cunningham, J. P. (2016). NIPS or Johnson, M., Duvenaud, D. K., Wiltschko, A., Adams, R. P., & Datta, S. R. (2016). NIPS. also, is there any methodological advancement essential to CENs? \n\nthe other idea is to go context specific. this idea has been present in language modeling, for example, amortized embedding models like M. Rudolph, F. Ruiz, S. Athey, and D. Blei (2017). NIPS and L. Liu, F. Ruiz, S. Athey, and D. Blei.  (2017). NIPS. application to medical data is interesting. but it could be helpful for the readers to understand if the idea in this work is fundamentally different from these previous ideas from amortized inference.\n\na final thing. a common challenge with composing graphical models and neural networks (in interpretable or uninterpretable ways) is that the neural networks will usually eat up all the representational power. the variance captured by graphical models becomes negligible. to this end, the power of graphical models for interpretability is limited. interpretability in this case is not much different from fitting only a neural network, taking the penultimate layer to the output as \"context specific features\" can claim that we are composing a linear model with a neural network, and the linear model is interpretable. it would be interesting to be clear about how the authors get around this issue.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting idea, maybe helpful to differentiate more from previous neural net+graphical models idea", "rating": "6: Marginally above acceptance threshold", "review": "the paper is clearly written; it works on a popular idea of combining graphical models and neural nets.\n\nthis work could benefit from differentiating more from previous literature.\n\none key component is interpretability, which comes from the use of graphical models.  the authors claim that the previous art directly integrate neural networks into the graphical models as components, which renders the models uninterpretable. however, it is unclear, following the same logic, why the proposed method has interpretability. after all, how to go from the context to the parameters of the graphical models is still uninterpretable. specifically, it is helpful to pinpoint what is special in this model that makes it interpretable, compared to works like Gao, Y., Archer, E. W., Paninski, L., & Cunningham, J. P. (2016). NIPS or Johnson, M., Duvenaud, D. K., Wiltschko, A., Adams, R. P., & Datta, S. R. (2016). NIPS. also, is there any methodological advancement essential to CENs? \n\nthe other idea is to go context specific. this idea has been present in language modeling, for example, amortized embedding models like M. Rudolph, F. Ruiz, S. Athey, and D. Blei (2017). NIPS and L. Liu, F. Ruiz, S. Athey, and D. Blei.  (2017). NIPS. application to medical data is interesting. but it could be helpful for the readers to understand if the idea in this work is fundamentally different from these previous ideas from amortized inference.\n\na final thing. a common challenge with composing graphical models and neural networks (in interpretable or uninterpretable ways) is that the neural networks will usually eat up all the representational power. the variance captured by graphical models becomes negligible. to this end, the power of graphical models for interpretability is limited. interpretability in this case is not much different from fitting only a neural network, taking the penultimate layer to the output as \"context specific features\" can claim that we are composing a linear model with a neural network, and the linear model is interpretable. it would be interesting to be clear about how the authors get around this issue.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511878303454}, {"id": "Bk-6h6Txz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper836/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The article \"Contextual Explanation Networks\" introduces the class of models which learn the intermediate explanations in order to make final predictions. The contexts can be learned by, in principle, any model including neural networks, while the final predictions are supposed to be made by some simple models like linear ones. The probabilistic model allows for the simultaneous training of explanation and prediction parts as opposed to some recent post-hoc methods.\n\nThe experimental part of the paper considers variety of experiments, including classification on MNIST, CIFAR-10, IMDB and also some experiments on survival analysis. I should note, that the quality of the algorithm is in general similar to other methods considered (as expected). However, while in some cases the CEN algorithm is slightly better, in other cases it appears to sufficiently loose, see for example left part of Figure 3(b) for MNIST data set. It would be interesting to know the explanation. Also, it would be interesting to have more examples of qualitative analysis to see, that the learned explanations are really useful. I am a bit worried, that while we have interpretability with respect to intermediate features, these features theirselves might be very hard to interpret.\n\nTo sum up, I think that the general idea looks very natural and the results are quite supportive. However, I don't feel myself confident enough in this area of research to make strong conclusion on the quality of the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An approach for end-to-end learning of interpretable models with experimental support", "rating": "6: Marginally above acceptance threshold", "review": "The article \"Contextual Explanation Networks\" introduces the class of models which learn the intermediate explanations in order to make final predictions. The contexts can be learned by, in principle, any model including neural networks, while the final predictions are supposed to be made by some simple models like linear ones. The probabilistic model allows for the simultaneous training of explanation and prediction parts as opposed to some recent post-hoc methods.\n\nThe experimental part of the paper considers variety of experiments, including classification on MNIST, CIFAR-10, IMDB and also some experiments on survival analysis. I should note, that the quality of the algorithm is in general similar to other methods considered (as expected). However, while in some cases the CEN algorithm is slightly better, in other cases it appears to sufficiently loose, see for example left part of Figure 3(b) for MNIST data set. It would be interesting to know the explanation. Also, it would be interesting to have more examples of qualitative analysis to see, that the learned explanations are really useful. I am a bit worried, that while we have interpretability with respect to intermediate features, these features theirselves might be very hard to interpret.\n\nTo sum up, I think that the general idea looks very natural and the results are quite supportive. However, I don't feel myself confident enough in this area of research to make strong conclusion on the quality of the paper.", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1512066233228}], "openreview_url": "https://openreview.net/forum?id=HJUOHGWRb", "arxiv_id": "1705.10301", "paper_pdf": "papers/HJUOHGWRb.pdf", "paper_pdf_sha256": "55db7c1e0c7c66a1b9f6d26c21df49f3f91d30e709bbee5f47c252fe28511bc3", "paper_pdf_bytes": 1881482, "paper_pdf_source": "openreview", "code_url": "https://github.com/alshedivat/cen", "code_repository": "alshedivat/cen", "code_commit": "f03397a0bf4ac24162e270907d623f8658179e88", "code_archive": "repos/HJUOHGWRb.zip", "code_archive_sha256": "b17324a3cfc8d36f3e5c0138da1fb5b99b7710ee3e259b3617d1cbfb8835366f", "code_archive_bytes": 100856, "code_file_count": 44, "code_extensions": {".py": 38, ".sh": 6}, "github_disk_usage_kb": 96, "github_languages": {"Python": 139132}, "github_archived": false, "github_pushed_at": "2020-03-19T20:57:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contextual-explanation-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vu0wCqW51g", "year": 2026, "status": "rejected", "title": "Content-Rich AIGC Video Quality Assessment via Intricate Text Alignment and Motion-Aware Consistency", "authors": ["Shangkun Sun", "Xiaoyu Liang", "Bowen Qu", "Ge Li", "Wei Gao"], "authorids": ["~Shangkun_Sun3", "~Xiaoyu_Liang2", "~Bowen_Qu1", "~Ge_Li2", "~Wei_Gao12"], "authors_source": "OpenReview API", "abstract": "The advent of next-generation video generation models like Sora poses challenges for AI-generated content (AIGC) video quality assessment (VQA). These models substantially mitigate flickering artifacts prevalent in prior models, enable longer and complex text prompts and generate longer videos with intricate, diverse motion patterns. Conventional VQA methods designed for simple text and basic motion patterns struggle to evaluate these content-rich videos. To this end, we propose **CRAVE** (Content-Rich AIGC Video Evaluator), specifically for the evaluation of Sora-era AIGC videos. CRAVE proposes the multi-granularity text-temporal fusion that aligns long-form complex textual semantics with video dynamics. Additionally, CRAVE leverages the hybrid motion-fidelity modeling to assess temporal artifacts. Furthermore, given the straightforward prompts and content in current AIGC VQA datasets, we introduce **CRAVE-DB**, a benchmark featuring content-rich videos from next-generation models paired with elaborate prompts. Extensive experiments have shown that the proposed CRAVE achieves excellent results on multiple AIGC VQA benchmarks, demonstrating a high degree of alignment with human perception. All data and code will be publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "Y1KJOrESDA", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12128/Reviewer_pU76"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces CRAVE, a purpose-built quality evaluator for the emerging generation of text-to-AIGC videos that exhibit longer duration, richer semantics, and more complex motion than their predecessors. By integrating (i) a visual-harmony branch that borrows DOVER’s aesthetic and distortion priors, (ii) a Multi-granularity Text-Temporal fusion module that aligns long, elaborate prompts with video dynamics via word-, phrase-, and sentence-level cross-attention, and (iii) a Hybrid Motion-fidelity Modeling component that jointly exploits dense optical-flow and high-level action semantics, CRAVE produces a single human-aligned quality score. Coupled with CRAVE-DB—a new 1 228-video benchmark annotated by 29 subjects according to ITU standards—the framework achieves state-of-the-art correlation with human ratings on both legacy (T2VQA-DB) and next-generation datasets, and generalises zero-shot to unseen models.", "review_text": "This paper introduces CRAVE, a purpose-built quality evaluator for the emerging generation of text-to-AIGC videos that exhibit longer duration, richer semantics, and more complex motion than their predecessors. By integrating (i) a visual-harmony branch that borrows DOVER’s aesthetic and distortion priors, (ii) a Multi-granularity Text-Temporal fusion module that aligns long, elaborate prompts with video dynamics via word-, phrase-, and sentence-level cross-attention, and (iii) a Hybrid Motion-fidelity Modeling component that jointly exploits dense optical-flow and high-level action semantics, CRAVE produces a single human-aligned quality score. Coupled with CRAVE-DB—a new 1 228-video benchmark annotated by 29 subjects according to ITU standards—the framework achieves state-of-the-art correlation with human ratings on both legacy (T2VQA-DB) and next-generation datasets, and generalises zero-shot to unseen models.", "strengths": "1. A three-branch architecture that explicitly disentangles and fuses visual harmony, fine-grained text–video alignment, and hierarchical motion fidelity for content-rich AIGC-VQA.\n2. CRAVE-DB, the first publicly available benchmark populated exclusively by Sora-era models, offering long, detail-dense prompts and statistically reliable MOS labels.\n3. Extensive experiments demonstrating superior SRCC/PLCC over existing metrics and a consistent zero-shot ranking of the latest generative models, validating CRAVE’s utility as a community standard.", "weaknesses": "1. Limited scale: The entire CRAVE-DB provides only 1 228 clips with ≈ 35 k human ratings—orders of magnitude smaller than the million-level weak-label sets commonly exploited in natural-video VQA—while the paper omits cross-dataset generalisation curves, leaving the community uncertain about over-fitting risks.\n\n2. Prohibitive optical-flow cost: HMM computes dense optical flow across 16 frames, yielding 4–6× slower inference than single-frame baselines; neither FLOPs nor wall-clock latency is reported, and scalability to videos longer than 10 s remains unverified.", "questions": "1. The paper reports neither FLOPs, memory footprint, nor wall-clock latency for the dense 16-frame optical-flow branch; without such complexity metrics it is unclear whether CRAVE's gains are achievable in real-time or industrial pipelines, nor how accuracy degrades under resource-constrained scenarios. It's recommended to add.\n\n2. Since competing baselines include image-oriented models (e.g., ImageReward, PickScore, etc.), a necessary ablation is to strip temporal components and evaluate CRAVE on established AIGC-IQA datasets (e.g., AIGIQA-20k); if the redesigned multi-granularity text–visual fusion still surpasses these image specialists, one can attribute the superiority to methodological innovation rather than to the rivals' inability to handle video inputs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces CRAVE, a purpose-built quality evaluator for the emerging generation of text-to-AIGC videos that exhibit longer duration, richer semantics, and more complex motion than their predecessors. By integrating (i) a visual-harmony branch that borrows DOVER’s aesthetic and distortion priors, (ii) a Multi-granularity Text-Temporal fusion module that aligns long, elaborate prompts with video dynamics via word-, phrase-, and sentence-level cross-attention, and (iii) a Hybrid Motion-fidelity Modeling component that jointly exploits dense optical-flow and high-level action semantics, CRAVE produces a single human-aligned quality score. Coupled with CRAVE-DB—a new 1 228-video benchmark annotated by 29 subjects according to ITU standards—the framework achieves state-of-the-art correlation with human ratings on both legacy (T2VQA-DB) and next-generation datasets, and generalises zero-shot to unseen models.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. A three-branch architecture that explicitly disentangles and fuses visual harmony, fine-grained text–video alignment, and hierarchical motion fidelity for content-rich AIGC-VQA.\n2. CRAVE-DB, the first publicly available benchmark populated exclusively by Sora-era models, offering long, detail-dense prompts and statistically reliable MOS labels.\n3. Extensive experiments demonstrating superior SRCC/PLCC over existing metrics and a consistent zero-shot ranking of the latest generative models, validating CRAVE’s utility as a community standard.", "weaknesses": "1. Limited scale: The entire CRAVE-DB provides only 1 228 clips with ≈ 35 k human ratings—orders of magnitude smaller than the million-level weak-label sets commonly exploited in natural-video VQA—while the paper omits cross-dataset generalisation curves, leaving the community uncertain about over-fitting risks.\n\n2. Prohibitive optical-flow cost: HMM computes dense optical flow across 16 frames, yielding 4–6× slower inference than single-frame baselines; neither FLOPs nor wall-clock latency is reported, and scalability to videos longer than 10 s remains unverified.", "questions": "1. The paper reports neither FLOPs, memory footprint, nor wall-clock latency for the dense 16-frame optical-flow branch; without such complexity metrics it is unclear whether CRAVE's gains are achievable in real-time or industrial pipelines, nor how accuracy degrades under resource-constrained scenarios. It's recommended to add.\n\n2. Since competing baselines include image-oriented models (e.g., ImageReward, PickScore, etc.), a necessary ablation is to strip temporal components and evaluate CRAVE on established AIGC-IQA datasets (e.g., AIGIQA-20k); if the redesigned multi-granularity text–visual fusion still surpasses these image specialists, one can attribute the superiority to methodological innovation rather than to the rivals' inability to handle video inputs.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761999542895}, {"id": "O0Abnor2pk", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12128/Reviewer_KLcS"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper addresses the problem of evaluating next-generation text-to-video (T2V) content produced by Sora-era AIGC models. It introduces CRAVE, a model for content-rich AIGC video quality assessment that combines multi-granularity text-temporal (MTT) fusion and hybrid motion-fidelity modeling (HMM). It introduces CRAVE-DB, a new ITU-compliant benchmark of long, detailed prompts and videos from models such as Sora, Kling, and Vidu.\nExperiments across T2VQA-DB, GAIA, VideoGenEval, and the new CRAVE-DB show state-of-the-art correlations with human ratings.", "review_text": "This paper addresses the problem of evaluating next-generation text-to-video (T2V) content produced by Sora-era AIGC models. It introduces CRAVE, a model for content-rich AIGC video quality assessment that combines multi-granularity text-temporal (MTT) fusion and hybrid motion-fidelity modeling (HMM). It introduces CRAVE-DB, a new ITU-compliant benchmark of long, detailed prompts and videos from models such as Sora, Kling, and Vidu.\nExperiments across T2VQA-DB, GAIA, VideoGenEval, and the new CRAVE-DB show state-of-the-art correlations with human ratings.", "strengths": "The paper tackles a timely and relevant problem by addressing the quality assessment of next-generation AIGC videos that contain richer semantics, longer durations, and more complex motion dynamics. It makes a meaningful empirical contribution by constructing CRAVE-DB, one of the first ITU-compliant datasets that includes content from cutting-edge video generators such as Sora, Kling, and Vidu. The proposed CRAVE framework achieves good results in the experiment.", "weaknesses": "1) The proposed CRAVE is not that new. The BLIP backbone, temporal adapter, and optical-flow features are well-known. The combination is logical but may be seen as incremental engineering rather than conceptual innovation. The novelty primarily lies in benchmark construction rather than architecture. It is suggested to better clarify the novelty of the proposed method.\n\n2) Although CRAVE-DB is ITU-compliant, it contains only ≈ 1.2 K videos, smaller than existing general-purpose sets like T2VQA-DB (10 K videos). Claims of “large-scale benchmark” may be overstated. Furthermore, the number of video generators is limited, which raises concerns about the diversity of the proposed dataset. More comparisons should be made on large-scale datasets from VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation, Q-Eval-100K: Evaluating Visual Quality and Alignment Level for Text-to-Vision Content.", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of evaluating next-generation text-to-video (T2V) content produced by Sora-era AIGC models. It introduces CRAVE, a model for content-rich AIGC video quality assessment that combines multi-granularity text-temporal (MTT) fusion and hybrid motion-fidelity modeling (HMM). It introduces CRAVE-DB, a new ITU-compliant benchmark of long, detailed prompts and videos from models such as Sora, Kling, and Vidu.\nExperiments across T2VQA-DB, GAIA, VideoGenEval, and the new CRAVE-DB show state-of-the-art correlations with human ratings.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper tackles a timely and relevant problem by addressing the quality assessment of next-generation AIGC videos that contain richer semantics, longer durations, and more complex motion dynamics. It makes a meaningful empirical contribution by constructing CRAVE-DB, one of the first ITU-compliant datasets that includes content from cutting-edge video generators such as Sora, Kling, and Vidu. The proposed CRAVE framework achieves good results in the experiment.", "weaknesses": "1) The proposed CRAVE is not that new. The BLIP backbone, temporal adapter, and optical-flow features are well-known. The combination is logical but may be seen as incremental engineering rather than conceptual innovation. The novelty primarily lies in benchmark construction rather than architecture. It is suggested to better clarify the novelty of the proposed method.\n\n2) Although CRAVE-DB is ITU-compliant, it contains only ≈ 1.2 K videos, smaller than existing general-purpose sets like T2VQA-DB (10 K videos). Claims of “large-scale benchmark” may be overstated. Furthermore, the number of video generators is limited, which raises concerns about the diversity of the proposed dataset. More comparisons should be made on large-scale datasets from VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation, Q-Eval-100K: Evaluating Visual Quality and Alignment Level for Text-to-Vision Content.", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1760964305992}, {"id": "khJlda7c7V", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12128/Reviewer_Dwqz"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes CRAVE for content-rich AIGV evaluation, which adopts multi-granularity text-temporal fusion and leverages the hybrid motion-fidelity modeling to assess temporal artifacts. Meanwhile, a dataset named CRAVE-DB is proposed to evaluate AIGV generated using elaborate prompts. Experiments show that CRAVE surpasses baseline models on several AIGV datasets, achieving more human-aligned predictions.", "review_text": "This paper proposes CRAVE for content-rich AIGV evaluation, which adopts multi-granularity text-temporal fusion and leverages the hybrid motion-fidelity modeling to assess temporal artifacts. Meanwhile, a dataset named CRAVE-DB is proposed to evaluate AIGV generated using elaborate prompts. Experiments show that CRAVE surpasses baseline models on several AIGV datasets, achieving more human-aligned predictions.", "strengths": "1. The research motivation is clear. For videos generated based on complex content prompts, an accurate evaluator is needed.\n2. An newer AIGV dataset generated with content-rich prompts and a corresponding video evaluator.", "weaknesses": "1. Although the proposed CRAVE-DB contains prompts with more words, its size is limited compared to existing AIGV datasets, with only 410 prompts and 1,228 videos generated from only 4 models.\n2. The contributions to the dataset and the method are both limited, which mainly involves the integration of existing methods.\n3. More recent AIGV benchmark should be compared (2024-2025), considering the timeliness of this direction.\n4. Lack of detailed data analysis of the subjective experimental results, such as the outliers, inter-user correlations, cross-model MOS performance, etc.\n5. The expression of the article needs to be improved. Some descriptions are overly repetitive, such as the ITU standards which is merely a necessary factor to consider, but it is not the main point.\n6. The model combines three aspects into a single score to describe AIGV. Does this violate the aforementioned fine-grained characterization? Are these three dimensions highly correlated or independent? Further analysis is needed on this.", "questions": "1. The presentation of Fig. 1 is somewhat vague, where I can only observe that the AIGC videos of the new generation contain more comprehensive content, which indicates that the prompt descriptions are more detailed. So, how would Lavie perform if the same prompt was used? The comparison should be made under the same conditions.\n2. Could the describtion of “the gap between the naturalness and complexity of the latest AIGC videos and previous videos has become markedly apparent” be more detailed? or provide analysis support?\n3. Tab. 1 could include more information, such as #video, # prompts, # models, as provided in Section 2.3, to enhance readability.\n4. The description in line 161 “Each video has a duration of over 5 seconds with a fps of 24” may not comprehensive enough, as stated in Line 123-124 that current models generate videos with a duration of more than 5 seconds and frame rates above 24 fps. More long videos with high frame rate is worth further investigation.\n5. Improper use of double quotes in line 194: ”landscape”, line 195: ”animal”, and line 196 ”object”.\n6. It would be more helpful to provide an intuitive example of a generated fine-grained prompt in Figure 2.\n7. The considered three evaluation perspectives: visual quality, T2V alignment, and motion quality are common and fair.\n8. I suggest that the three dimensions that need to be evaluated be presented in the main text. This is quite important, especially for explaining the MOS distribution shown in Figure 3. Currently, I'm not sure what MOS in Figure 3 represents.\n9. The description in Figure 4 could be more straightforward. Presenting the network structure directly rather than using a vague term would be more helpful for readers to understand, especially for the relatively simple modules.\n10. It is suggested to conduct a detailed visualization of SpaCy's and the MTT’s working principle or integrate it into the existing Figure 4 (Line 308).\n11. The operation mechanism of the HYBRID MOTION-FIDELITY MODELING module also requires a detailed illustration rather than a simple flowchart in current Fig. 4.\n12. Lack of statistical tests for the significance of performance differences, as it seem that the prediction results on DOVER are very similar to those on CRAVE (Fig. 6).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CRAVE for content-rich AIGV evaluation, which adopts multi-granularity text-temporal fusion and leverages the hybrid motion-fidelity modeling to assess temporal artifacts. Meanwhile, a dataset named CRAVE-DB is proposed to evaluate AIGV generated using elaborate prompts. Experiments show that CRAVE surpasses baseline models on several AIGV datasets, achieving more human-aligned predictions.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The research motivation is clear. For videos generated based on complex content prompts, an accurate evaluator is needed.\n2. An newer AIGV dataset generated with content-rich prompts and a corresponding video evaluator.", "weaknesses": "1. Although the proposed CRAVE-DB contains prompts with more words, its size is limited compared to existing AIGV datasets, with only 410 prompts and 1,228 videos generated from only 4 models.\n2. The contributions to the dataset and the method are both limited, which mainly involves the integration of existing methods.\n3. More recent AIGV benchmark should be compared (2024-2025), considering the timeliness of this direction.\n4. Lack of detailed data analysis of the subjective experimental results, such as the outliers, inter-user correlations, cross-model MOS performance, etc.\n5. The expression of the article needs to be improved. Some descriptions are overly repetitive, such as the ITU standards which is merely a necessary factor to consider, but it is not the main point.\n6. The model combines three aspects into a single score to describe AIGV. Does this violate the aforementioned fine-grained characterization? Are these three dimensions highly correlated or independent? Further analysis is needed on this.", "questions": "1. The presentation of Fig. 1 is somewhat vague, where I can only observe that the AIGC videos of the new generation contain more comprehensive content, which indicates that the prompt descriptions are more detailed. So, how would Lavie perform if the same prompt was used? The comparison should be made under the same conditions.\n2. Could the describtion of “the gap between the naturalness and complexity of the latest AIGC videos and previous videos has become markedly apparent” be more detailed? or provide analysis support?\n3. Tab. 1 could include more information, such as #video, # prompts, # models, as provided in Section 2.3, to enhance readability.\n4. The description in line 161 “Each video has a duration of over 5 seconds with a fps of 24” may not comprehensive enough, as stated in Line 123-124 that current models generate videos with a duration of more than 5 seconds and frame rates above 24 fps. More long videos with high frame rate is worth further investigation.\n5. Improper use of double quotes in line 194: ”landscape”, line 195: ”animal”, and line 196 ”object”.\n6. It would be more helpful to provide an intuitive example of a generated fine-grained prompt in Figure 2.\n7. The considered three evaluation perspectives: visual quality, T2V alignment, and motion quality are common and fair.\n8. I suggest that the three dimensions that need to be evaluated be presented in the main text. This is quite important, especially for explaining the MOS distribution shown in Figure 3. Currently, I'm not sure what MOS in Figure 3 represents.\n9. The description in Figure 4 could be more straightforward. Presenting the network structure directly rather than using a vague term would be more helpful for readers to understand, especially for the relatively simple modules.\n10. It is suggested to conduct a detailed visualization of SpaCy's and the MTT’s working principle or integrate it into the existing Figure 4 (Line 308).\n11. The operation mechanism of the HYBRID MOTION-FIDELITY MODELING module also requires a detailed illustration rather than a simple flowchart in current Fig. 4.\n12. Lack of statistical tests for the significance of performance differences, as it seem that the prediction results on DOVER are very similar to those on CRAVE (Fig. 6).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1760887209657}], "openreview_url": "https://openreview.net/forum?id=vu0wCqW51g", "arxiv_id": "2502.04076", "paper_pdf": "papers/vu0wCqW51g.pdf", "paper_pdf_sha256": "0800beab20ba09644b519f7f8b0293689aa1db38d874c941d0508e86f07f0e3c", "paper_pdf_bytes": 5070168, "paper_pdf_source": "openreview", "code_url": "https://github.com/littlespray/CRAVE", "code_repository": "littlespray/CRAVE", "code_commit": "2c4973b73dadef6eb6bcccf50e272c24df8f9c11", "code_archive": "repos/vu0wCqW51g.zip", "code_archive_sha256": "c450ca4d9a6967cbef35f0f599e443fab8015637ba0ea151d197e760e51128b8", "code_archive_bytes": 109603, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 103, "github_languages": {"Python": 356310}, "github_archived": false, "github_pushed_at": "2025-07-31T16:11:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/content-rich-aigc-video-quality-assessment"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bBUhlynfRX", "year": 2025, "status": "rejected", "title": "A Brain-Inspired Regularizer for Adversarial Robustness", "authors": ["Elie Attias", "Cengiz Pehlevan", "Dina Obeid"], "authorids": ["~Elie_Attias1", "~Cengiz_Pehlevan2", "~Dina_Obeid1"], "authors_source": "OpenReview API", "abstract": "Convolutional Neural Networks (CNNs) excel in many visual tasks, but they tend to be sensitive to slight input perturbations that are imperceptible to the human eye, often resulting in task failures. Recent studies indicate that training CNNs with regularizers that promote brain-like representations, using neural recordings, can improve model robustness. However, the requirement to use neural data severely restricts the utility of these methods. Is it possible to develop regularizers that mimic the computational function of neural regularizers without the need for neural recordings, thereby expanding the usability and effectiveness of these techniques? In this work, we inspect a neural regularizer introduced in Li et al. to extract its underlying strength. The regularizer uses neural representational similarities, which we find also correlate with pixel similarities. Motivated by this finding, we introduce a new regularizer that retains the essence of the original but is computed using image pixel similarities, eliminating the need for neural recordings.  We show that our regularization method 1) significantly increases model robustness against a variety of black box attacks, 2) relies only on original, unaugmented datasets and 3) is computationally inexpensive. Our work explores how biologically motivated loss functions can be used to drive the performance of artificial neural networks.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "RUSujgGwdk", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8492/Reviewer_oT1f"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper inspects an existing neural regularizer to establish that the idea can be used for image pixel similarities without the use of actual neural recordings. Based on this observation, the authors propose a new regularization method based on two hyperparameters and eliminating the need for neural recordings. It is evaluated for different adversarial attacks and grayscale and colored datasets. Based on the selected threshold value, the target similarity matrix is calculated. The method is tested against transferable FGSM and Boundary attack as well as common image corruptions.", "review_text": "The paper inspects an existing neural regularizer to establish that the idea can be used for image pixel similarities without the use of actual neural recordings. Based on this observation, the authors propose a new regularization method based on two hyperparameters and eliminating the need for neural recordings. It is evaluated for different adversarial attacks and grayscale and colored datasets. Based on the selected threshold value, the target similarity matrix is calculated. The method is tested against transferable FGSM and Boundary attack as well as common image corruptions.", "strengths": "- The proposed method eliminates the need for neural recordings, which are generally expensive.\n- Well-motivated, intuitive, and simple method to eliminate the actual neural recordings. \n- Computationally light method.", "weaknesses": "- The authors mention, \"our aim is not to come up with the best adversarial defence ...\" on lines 519-521, which contradicts the title and the paper content that gives an impression the proposed method is to improve the adversarial robustness of the existing networks. \n- The authors should discuss the preliminary details/definitions they use frequently from the baseline paper. For example, what are the classification and regularization datasets? \n- While the improvement on the common corruptions is good, the evaluation of adversarial robustness is incomplete. The method is only tested against transferable FGSM and Boundary attacks. More robust evaluation with stronger white-box [1, 2] and black-box attacks [3] should be done to claim adversarial robustness. \n- Evaluation is only limited to one CNN architecture - ResNet18. More evaluation of diverse architectures and models like WideResNet will provide additional evidence for the claims.\n- In Section 2, the adversarial attacks mentioned are very old. The newer, more effective, and stronger attacks are not included. Similar observations are made for the adversarial defense subsection in Section 2.   \n- Minor typos - \"imageNet\" instead of \"ImageNet\" on line 389, every equation is referred as \"eq. equation \" instead of \"Eq. \" or \"Equation \",  \n\n[1] Croce, Francesco, and Matthias Hein. \"Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks.\" International conference on machine learning. PMLR, 2020.\n[2] Mądry, Aleksander, et al. \"Towards deep learning models resistant to adversarial attacks.\" stat 1050.9 (2017).\n[3] Andriushchenko, Maksym, et al. \"Square attack: a query-efficient black-box adversarial attack via random search.\" European conference on computer vision. Cham: Springer International Publishing, 2020.", "questions": "My main concern is the contradictory claims in the paper. Please refer to the weakness section for more questions and details.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper inspects an existing neural regularizer to establish that the idea can be used for image pixel similarities without the use of actual neural recordings. Based on this observation, the authors propose a new regularization method based on two hyperparameters and eliminating the need for neural recordings. It is evaluated for different adversarial attacks and grayscale and colored datasets. Based on the selected threshold value, the target similarity matrix is calculated. The method is tested against transferable FGSM and Boundary attack as well as common image corruptions.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The proposed method eliminates the need for neural recordings, which are generally expensive.\n- Well-motivated, intuitive, and simple method to eliminate the actual neural recordings. \n- Computationally light method.", "weaknesses": "- The authors mention, \"our aim is not to come up with the best adversarial defence ...\" on lines 519-521, which contradicts the title and the paper content that gives an impression the proposed method is to improve the adversarial robustness of the existing networks. \n- The authors should discuss the preliminary details/definitions they use frequently from the baseline paper. For example, what are the classification and regularization datasets? \n- While the improvement on the common corruptions is good, the evaluation of adversarial robustness is incomplete. The method is only tested against transferable FGSM and Boundary attacks. More robust evaluation with stronger white-box [1, 2] and black-box attacks [3] should be done to claim adversarial robustness. \n- Evaluation is only limited to one CNN architecture - ResNet18. More evaluation of diverse architectures and models like WideResNet will provide additional evidence for the claims.\n- In Section 2, the adversarial attacks mentioned are very old. The newer, more effective, and stronger attacks are not included. Similar observations are made for the adversarial defense subsection in Section 2.   \n- Minor typos - \"imageNet\" instead of \"ImageNet\" on line 389, every equation is referred as \"eq. equation \" instead of \"Eq. \" or \"Equation \",  \n\n[1] Croce, Francesco, and Matthias Hein. \"Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks.\" International conference on machine learning. PMLR, 2020.\n[2] Mądry, Aleksander, et al. \"Towards deep learning models resistant to adversarial attacks.\" stat 1050.9 (2017).\n[3] Andriushchenko, Maksym, et al. \"Square attack: a query-efficient black-box adversarial attack via random search.\" European conference on computer vision. Cham: Springer International Publishing, 2020.", "questions": "My main concern is the contradictory claims in the paper. Please refer to the weakness section for more questions and details.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730706223966}, {"id": "TbgsitdAaR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8492/Reviewer_k7ZX"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "Inspired by the work of Li et al. (2019) and guided by the observation that pixel similarities in the original image exhibit a pattern similar to neural recordings, the authors propose a new regularization method to improve neural network robustness with a significant reduction in computational requirements. As claimed by the authors, the proposed method also applies to colour datasets and has been tested on ResNet18 using CIFAR10, CIFAR100, and ImageNet with Gaussian noise and some black-box attacks.", "review_text": "Inspired by the work of Li et al. (2019) and guided by the observation that pixel similarities in the original image exhibit a pattern similar to neural recordings, the authors propose a new regularization method to improve neural network robustness with a significant reduction in computational requirements. As claimed by the authors, the proposed method also applies to colour datasets and has been tested on ResNet18 using CIFAR10, CIFAR100, and ImageNet with Gaussian noise and some black-box attacks.", "strengths": "1. The idea presented in the paper is interesting.\n2. Lots of experiments have been conducted to support the conclusions.\n3. The paper considers various aspects of robustness, including both *adversarial robustness* and *common corruptions*.", "weaknesses": "1. There are lots of obvious grammar, abbreviation, and citation errors in the paper. Aside from the grammar mistakes, such as in line 12 of the Abstract where it says \"... but they tend to sensitive ...\" which should be \"... they tend to be sensitive...\", nearly all equation references are incorrectly formatted. It should be \"Equation\" or \"Eq.\" instead of \"eq.equation.\" Additionally, please check the references you've used. For instance, on lines 83-84 of page 2, you incorrectly cite Szegedy et al. (2013) ... Madry et al. (2017) without proper formatting. A parenthesis is likely needed here; please refer to the formatting instructions in section 4.1. You should review the paper at least once before submission.\n    \n2. The writing and organization of the paper are very vague and confusing. In the Introduction section on page 2, you list six contributions of your work. However, some of these contributions are too trivial to be included as a separate item. For example, in the fourth contribution, you state: 'We assess the robustness of the regularized models to common corruptions using the CIFAR-10-C dataset (Hendrycks & Dietterich, 2019a).' I do not believe this alone is substantial enough to be listed as a contribution. Some contributions could be combined. In the last contribution, you mention: 'We show that our regularization method primarily protects against high-frequency perturbations…,' which explains the mechanism behind your proposed method but does not stand alone as a contribution. Instead, the sentence on lines 74-75 serves more as a contribution but is not included in the contributions list.\n \n3. Outdated related works. In the related works section, the most recent work you mention on adversarial attacks is from 2017, which was published seven years ago. You should include at least one recent study. Additionally, the most recent work across all of your related works is Kireev et al. (2022), which is only mentioned once. I suggest incorporating a broader scope of more up-to-date works.\n\n4. Some concepts are not clearly explained. On page 3, lines 137-138, you mention S_{i,j}^{target} as the target's pairwise cosine similarity, but it is unclear what 'target' refers to.\n\n5. Some of your claims need clearer evidence. On page 4, lines 162-166, you mention a similarity in the pattern of correlation between neural representational similarity and image pixel similarity, but the only evidence provided is Fig. 2, which is not convincing on its own. Additionally, your explanation of Fig. 2 is quite vague. For example, in the leftmost figure of Fig. 2, I see r=0.83, but there is no explanation provided for this value.", "questions": "1. As stated in your second contribution, 'We show that our regularizer drives the network to be more robust to a wide range of black-box attacks,' and since all your experiments are based on ResNet18, I am curious whether this method would work for different architectures, such as MLPs.\n\n2. Since FGSM is $L_infty$ based and considered one of the weakest attack methods, I am curious whether your proposed method would still be effective against more advanced transfer attacks.  \n\n3. In Section 4.6, you attempt to demonstrate that your method is computationally inexpensive, but it is quite confusing to understand why this is the case from Fig. 8. Could you provide any plots related to computational complexity or time to support your argument?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Inspired by the work of Li et al. (2019) and guided by the observation that pixel similarities in the original image exhibit a pattern similar to neural recordings, the authors propose a new regularization method to improve neural network robustness with a significant reduction in computational requirements. As claimed by the authors, the proposed method also applies to colour datasets and has been tested on ResNet18 using CIFAR10, CIFAR100, and ImageNet with Gaussian noise and some black-box attacks.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "1. The idea presented in the paper is interesting.\n2. Lots of experiments have been conducted to support the conclusions.\n3. The paper considers various aspects of robustness, including both *adversarial robustness* and *common corruptions*.", "weaknesses": "1. There are lots of obvious grammar, abbreviation, and citation errors in the paper. Aside from the grammar mistakes, such as in line 12 of the Abstract where it says \"... but they tend to sensitive ...\" which should be \"... they tend to be sensitive...\", nearly all equation references are incorrectly formatted. It should be \"Equation\" or \"Eq.\" instead of \"eq.equation.\" Additionally, please check the references you've used. For instance, on lines 83-84 of page 2, you incorrectly cite Szegedy et al. (2013) ... Madry et al. (2017) without proper formatting. A parenthesis is likely needed here; please refer to the formatting instructions in section 4.1. You should review the paper at least once before submission.\n    \n2. The writing and organization of the paper are very vague and confusing. In the Introduction section on page 2, you list six contributions of your work. However, some of these contributions are too trivial to be included as a separate item. For example, in the fourth contribution, you state: 'We assess the robustness of the regularized models to common corruptions using the CIFAR-10-C dataset (Hendrycks & Dietterich, 2019a).' I do not believe this alone is substantial enough to be listed as a contribution. Some contributions could be combined. In the last contribution, you mention: 'We show that our regularization method primarily protects against high-frequency perturbations…,' which explains the mechanism behind your proposed method but does not stand alone as a contribution. Instead, the sentence on lines 74-75 serves more as a contribution but is not included in the contributions list.\n \n3. Outdated related works. In the related works section, the most recent work you mention on adversarial attacks is from 2017, which was published seven years ago. You should include at least one recent study. Additionally, the most recent work across all of your related works is Kireev et al. (2022), which is only mentioned once. I suggest incorporating a broader scope of more up-to-date works.\n\n4. Some concepts are not clearly explained. On page 3, lines 137-138, you mention S_{i,j}^{target} as the target's pairwise cosine similarity, but it is unclear what 'target' refers to.\n\n5. Some of your claims need clearer evidence. On page 4, lines 162-166, you mention a similarity in the pattern of correlation between neural representational similarity and image pixel similarity, but the only evidence provided is Fig. 2, which is not convincing on its own. Additionally, your explanation of Fig. 2 is quite vague. For example, in the leftmost figure of Fig. 2, I see r=0.83, but there is no explanation provided for this value.", "questions": "1. As stated in your second contribution, 'We show that our regularizer drives the network to be more robust to a wide range of black-box attacks,' and since all your experiments are based on ResNet18, I am curious whether this method would work for different architectures, such as MLPs.\n\n2. Since FGSM is $L_infty$ based and considered one of the weakest attack methods, I am curious whether your proposed method would still be effective against more advanced transfer attacks.  \n\n3. In Section 4.6, you attempt to demonstrate that your method is computationally inexpensive, but it is quite confusing to understand why this is the case from Fig. 8. Could you provide any plots related to computational complexity or time to support your argument?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730667819921}, {"id": "eumjg0kiHy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8492/Reviewer_YD33"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "The paper introduces a new, brain-inspired regularizer for CNNs that enhances robustness against adversarial attacks without needing neural data. By using pixel-based similarity measures instead of neural recordings, the proposed method simplifies implementation and makes the regularization process computationally efficient. Tests across several datasets and adversarial attacks (e.g., Gaussian noise, FGSM, and Boundary Attacks) show that this approach achieves significant robustness. Though it doesn’t outperform the latest adversarial defenses on all fronts, it holds promise for practical applications requiring efficient and accessible methods for adversarial robustness. The findings highlight how neuroscience-inspired regularizers can improve machine learning models without heavy data dependencies.", "review_text": "The paper introduces a new, brain-inspired regularizer for CNNs that enhances robustness against adversarial attacks without needing neural data. By using pixel-based similarity measures instead of neural recordings, the proposed method simplifies implementation and makes the regularization process computationally efficient. Tests across several datasets and adversarial attacks (e.g., Gaussian noise, FGSM, and Boundary Attacks) show that this approach achieves significant robustness. Though it doesn’t outperform the latest adversarial defenses on all fronts, it holds promise for practical applications requiring efficient and accessible methods for adversarial robustness. The findings highlight how neuroscience-inspired regularizers can improve machine learning models without heavy data dependencies.", "strengths": "The strengths of this paper include an accessible regularizer that enhances robustness without neural data and computational efficiency by avoiding data augmentations.\nI find the brain-inspired motivation of the paper also worthwhile.", "weaknesses": "There are many weaknesses.\nThe authors only tested on CIFAR. I could not see any tests comparing with the large amount of defenses available today. \nThe paper is heavily based on the paper from Li 2019.\nI find the presentation to be lacking. There are no figures explaining the bioinspiration or the motivation behind the method,  or even how does this differs from the work from Li 2019.\nThe improvement from regularization looks very similar to the ones achieved by simple adversarial training. \nThe choice of attacks could. have been more diverse. L0 and Linfty attacks with both white and black box types are the standard for evaluation.", "questions": "How does the method presented compare with current defenses?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new, brain-inspired regularizer for CNNs that enhances robustness against adversarial attacks without needing neural data. By using pixel-based similarity measures instead of neural recordings, the proposed method simplifies implementation and makes the regularization process computationally efficient. Tests across several datasets and adversarial attacks (e.g., Gaussian noise, FGSM, and Boundary Attacks) show that this approach achieves significant robustness. Though it doesn’t outperform the latest adversarial defenses on all fronts, it holds promise for practical applications requiring efficient and accessible methods for adversarial robustness. The findings highlight how neuroscience-inspired regularizers can improve machine learning models without heavy data dependencies.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "The strengths of this paper include an accessible regularizer that enhances robustness without neural data and computational efficiency by avoiding data augmentations.\nI find the brain-inspired motivation of the paper also worthwhile.", "weaknesses": "There are many weaknesses.\nThe authors only tested on CIFAR. I could not see any tests comparing with the large amount of defenses available today. \nThe paper is heavily based on the paper from Li 2019.\nI find the presentation to be lacking. There are no figures explaining the bioinspiration or the motivation behind the method,  or even how does this differs from the work from Li 2019.\nThe improvement from regularization looks very similar to the ones achieved by simple adversarial training. \nThe choice of attacks could. have been more diverse. L0 and Linfty attacks with both white and black box types are the standard for evaluation.", "questions": "How does the method presented compare with current defenses?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730462978669}], "openreview_url": "https://openreview.net/forum?id=bBUhlynfRX", "arxiv_id": "2410.03952", "paper_pdf": "papers/bBUhlynfRX.pdf", "paper_pdf_sha256": "ac7e90280194abfa4e7222c109fbc276b93c20a497da35528fd1801e203c8fc0", "paper_pdf_bytes": 3582587, "paper_pdf_source": "openreview", "code_url": "https://github.com/elieattias1/pixel-reg", "code_repository": "elieattias1/pixel-reg", "code_commit": "69d13eb5efa3a70eebcc38705c2fd0ab663e8d1a", "code_archive": "repos/bBUhlynfRX.zip", "code_archive_sha256": "cd8a51bbcf8153bf92929aa6352aa76885862e3362147dae7662faac69b49c11", "code_archive_bytes": 29683, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 28, "github_languages": {"Python": 95335}, "github_archived": false, "github_pushed_at": "2024-10-16T01:18:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-brain-inspired-regularizer-for-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6rEcB9m9AI", "year": 2024, "status": "rejected", "title": "Promoting Exploration in Memory-Augmented Adam using Critical Momenta", "authors": ["Pranshu Malviya", "Goncalo Mordido", "Aristide Baratin", "Reza Babanezhad Harikandeh", "Jerry Huang", "Simon Lacoste-Julien", "Razvan Pascanu", "Sarath Chandar"], "authorids": ["~Pranshu_Malviya1", "~Goncalo_Mordido1", "~Aristide_Baratin1", "~Reza_Babanezhad_Harikandeh1", "~Jerry_Huang1", "~Simon_Lacoste-Julien1", "~Razvan_Pascanu1", "~Sarath_Chandar1"], "authors_source": "OpenReview API", "abstract": "Adaptive gradient-based optimizers, particularly Adam, have left their mark in training large-scale deep learning models. The strength of such optimizers is that they exhibit fast convergence while being more robust to hyperparameter choice. However, they often generalize worse than non-adaptive methods. Recent studies have tied this performance gap to flat minima selection: adaptive methods tend to find solutions in sharper basins of the loss landscape, which in turn hurts generalization. To overcome this issue, we propose a new memory-augmented version of Adam that promotes {exploration} towards flatter minima by using a buffer of critical momentum terms during training. Intuitively, the use of the buffer makes the optimizer overshoot outside the basin of attraction if it is not wide enough. We empirically show that our method improves model performance on standard supervised and online learning tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "ESSWoytK0w", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3594/Reviewer_Jwku"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents a memory-augmented variant of the Adam optimizer that maintains a memory of critical momenta. The proposed method aims to address the gradient cancellation problem of Adam with critical gradients, encouraging exploration toward flatter minima. It is shown to improve performance across various tasks, including language modeling and image classification, by allowing the optimizer to escape sharp basins and converge in flatter regions.", "review_text": "This paper presents a memory-augmented variant of the Adam optimizer that maintains a memory of critical momenta. The proposed method aims to address the gradient cancellation problem of Adam with critical gradients, encouraging exploration toward flatter minima. It is shown to improve performance across various tasks, including language modeling and image classification, by allowing the optimizer to escape sharp basins and converge in flatter regions.", "strengths": "The paper is well-written and clear, and uses both toy examples and real-world problems to demonstrate the effectiveness of the proposed method.", "weaknesses": "1. Prioritizing momenta with large gradient norms is not well motivated. It is unclear if and how this choice affects the performance of the method.\n2. The paper seems to associate flat minima with global minima or lower loss values in some of the toy examples, which is not always the case in practice. Indeed, the training loss of a flat minimum is usually higher than that of a sharp minimum on real-world datasets.\n3. The proposed method is partly motivated by the performance gap between SGD and Adam, but is not compared to SGD in the experiments. Moreover, it would be interesting to see if critical momenta improve the performance of SGD as well.\n4. The baselines are not particularly strong, and it is a bit strange that Adam+SAM performs almost the same or even worse than vanilla Adam on image classification tasks.\n5. The results shown in Fig. 5 could be sensitive to hyperparameters such as learning rate, which is not discussed in the paper.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a memory-augmented variant of the Adam optimizer that maintains a memory of critical momenta. The proposed method aims to address the gradient cancellation problem of Adam with critical gradients, encouraging exploration toward flatter minima. It is shown to improve performance across various tasks, including language modeling and image classification, by allowing the optimizer to escape sharp basins and converge in flatter regions.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well-written and clear, and uses both toy examples and real-world problems to demonstrate the effectiveness of the proposed method.", "weaknesses": "1. Prioritizing momenta with large gradient norms is not well motivated. It is unclear if and how this choice affects the performance of the method.\n2. The paper seems to associate flat minima with global minima or lower loss values in some of the toy examples, which is not always the case in practice. Indeed, the training loss of a flat minimum is usually higher than that of a sharp minimum on real-world datasets.\n3. The proposed method is partly motivated by the performance gap between SGD and Adam, but is not compared to SGD in the experiments. Moreover, it would be interesting to see if critical momenta improve the performance of SGD as well.\n4. The baselines are not particularly strong, and it is a bit strange that Adam+SAM performs almost the same or even worse than vanilla Adam on image classification tasks.\n5. The results shown in Fig. 5 could be sensitive to hyperparameters such as learning rate, which is not discussed in the paper.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699729916897}, {"id": "uX2Znm1l7M", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3594/Reviewer_GnW1"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "This paper introduces a memory-augmented version of Adam named Adam+CM, which maintains a buffer of critical momenta to enhance exploration and avoid sharp minima during training. The authors have conducted theoretical convergence analysis for the proposed method. Furthermore, they provided empirical results to substantiate the claims regarding the properties of the method. By presenting results in online learning and supervised learning scenarios (NLP + CV), the authors demonstrate that the proposed method achieves superior generalization performance compared to baseline methods.", "review_text": "This paper introduces a memory-augmented version of Adam named Adam+CM, which maintains a buffer of critical momenta to enhance exploration and avoid sharp minima during training. The authors have conducted theoretical convergence analysis for the proposed method. Furthermore, they provided empirical results to substantiate the claims regarding the properties of the method. By presenting results in online learning and supervised learning scenarios (NLP + CV), the authors demonstrate that the proposed method achieves superior generalization performance compared to baseline methods.", "strengths": "- The proposed method is well-motivated, and to the best of my knowledge, it presents a novel approach to optimization.\n- The presentation is lucid and easy to follow.\n- The paper offers comprehensive evidence to back the primary claims, including theoretical analysis, numerical simulation results with toy examples, and real-world practical experiments.", "weaknesses": "- (minor) The paper would be more convincing if larger-scale empirical experiments on NLP were conducted.", "questions": "- As a proposed deep learning optimization method, it would be more convincing if comparisons were made on larger-scale language model optimization tasks as well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a memory-augmented version of Adam named Adam+CM, which maintains a buffer of critical momenta to enhance exploration and avoid sharp minima during training. The authors have conducted theoretical convergence analysis for the proposed method. Furthermore, they provided empirical results to substantiate the claims regarding the properties of the method. By presenting results in online learning and supervised learning scenarios (NLP + CV), the authors demonstrate that the proposed method achieves superior generalization performance compared to baseline methods.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The proposed method is well-motivated, and to the best of my knowledge, it presents a novel approach to optimization.\n- The presentation is lucid and easy to follow.\n- The paper offers comprehensive evidence to back the primary claims, including theoretical analysis, numerical simulation results with toy examples, and real-world practical experiments.", "weaknesses": "- (minor) The paper would be more convincing if larger-scale empirical experiments on NLP were conducted.", "questions": "- As a proposed deep learning optimization method, it would be more convincing if comparisons were made on larger-scale language model optimization tasks as well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698839723518}, {"id": "CWtiRugt2K", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3594/Reviewer_V8UZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Building upon the framework of critical gradients, which maintains a buffer containing a limited history of gradients from previous iterations, this paper takes a similar approach by maintaining a buffer of previous momenta, which they call critical momentum (CM). The proposed method is expected to effectively escape sharp basins and converge to flat loss regions. They also investigate the effectiveness of combining CM with sharpness aware minimization (SAM).", "review_text": "Building upon the framework of critical gradients, which maintains a buffer containing a limited history of gradients from previous iterations, this paper takes a similar approach by maintaining a buffer of previous momenta, which they call critical momentum (CM). The proposed method is expected to effectively escape sharp basins and converge to flat loss regions. They also investigate the effectiveness of combining CM with sharpness aware minimization (SAM).", "strengths": "The preliminary experiments using artificial loss functions such as the Goldstein-Price loss function and Ackley loss function show a clear advantage of CM variants to explore and find the lower loss surface near the global solution.", "weaknesses": "Although the experiments using artificial loss functions are promising, the results on actual deep neural networks is not as good. The fact that both CM methods overfit after convergence for LSTM on PTB in Figure 6 raises a major concern, and contradicts the main claim of the paper that CM methods can find a flatter minima.\n\nThe scores that are reported in this paper seem to be quite lower than what is reported in \"papers with code\". For example, EfficientNet-B0 on ImageNet should achieve a top-1 validation accuracy of 76.3% [https://paperswithcode.com/paper/efficientnet-rethinking-model-scaling-for], but the results reported in the paper range between 65.4-71.7%. Also, Adam+SAM is performing very poorly for this specific case, and is significantly worse than the original Adam, but no explanation is given for this poor performance.\n\nIt would seem like the current experiments would be strongly affected by the batch size, but the batch size is not included in their hyperparameter search.", "questions": "The escape ratio in Figure 5 is monotonically increasing with the sharpness coefficient. What happens if it is increased further?\n\nIn Figure 6, the validation perplexity of Adam+CM shows a peculiar behavior where it initially drops very fast but gradually increases and eventually crosses over with Adam+SAM. Is there any explanation as to why Adam+CM behaves this way when it is supposed to explore flatter regions compared to other optimizers?\n\nWhy is Adam+SAM performing so poorly for EfficientNet-B0 on ImageNet in Table 2?\n\nComments\nThe correspondence between the two plots in Figure 4 would be more obvious if the color of the lines were matched.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Building upon the framework of critical gradients, which maintains a buffer containing a limited history of gradients from previous iterations, this paper takes a similar approach by maintaining a buffer of previous momenta, which they call critical momentum (CM). The proposed method is expected to effectively escape sharp basins and converge to flat loss regions. They also investigate the effectiveness of combining CM with sharpness aware minimization (SAM).", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The preliminary experiments using artificial loss functions such as the Goldstein-Price loss function and Ackley loss function show a clear advantage of CM variants to explore and find the lower loss surface near the global solution.", "weaknesses": "Although the experiments using artificial loss functions are promising, the results on actual deep neural networks is not as good. The fact that both CM methods overfit after convergence for LSTM on PTB in Figure 6 raises a major concern, and contradicts the main claim of the paper that CM methods can find a flatter minima.\n\nThe scores that are reported in this paper seem to be quite lower than what is reported in \"papers with code\". For example, EfficientNet-B0 on ImageNet should achieve a top-1 validation accuracy of 76.3% [https://paperswithcode.com/paper/efficientnet-rethinking-model-scaling-for], but the results reported in the paper range between 65.4-71.7%. Also, Adam+SAM is performing very poorly for this specific case, and is significantly worse than the original Adam, but no explanation is given for this poor performance.\n\nIt would seem like the current experiments would be strongly affected by the batch size, but the batch size is not included in their hyperparameter search.", "questions": "The escape ratio in Figure 5 is monotonically increasing with the sharpness coefficient. What happens if it is increased further?\n\nIn Figure 6, the validation perplexity of Adam+CM shows a peculiar behavior where it initially drops very fast but gradually increases and eventually crosses over with Adam+SAM. Is there any explanation as to why Adam+CM behaves this way when it is supposed to explore flatter regions compared to other optimizers?\n\nWhy is Adam+SAM performing so poorly for EfficientNet-B0 on ImageNet in Table 2?\n\nComments\nThe correspondence between the two plots in Figure 4 would be more obvious if the color of the lines were matched.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698820637009}, {"id": "tg8ekWWRVH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3594/Reviewer_qx25"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses the challenge of poor generalization performance in adaptive gradient methods like Adam, hypothesizing that adding an exploration strategy with additional memory of gradient information could improve performance. Building on previous work, which used a memory buffer for storing \"critical gradients,\" the authors propose a new memory-augmented Adam optimizer that stores \"critical momenta\" (CM). This new approach aims to prevent gradient cancellation and allows the optimizer to escape sharp minima and converge to flat regions which are more likely to generalize. The authors validate their approach through convergence analysis and empirical benchmarks, showing improvements in model performance in both supervised and online learning settings.", "review_text": "The paper addresses the challenge of poor generalization performance in adaptive gradient methods like Adam, hypothesizing that adding an exploration strategy with additional memory of gradient information could improve performance. Building on previous work, which used a memory buffer for storing \"critical gradients,\" the authors propose a new memory-augmented Adam optimizer that stores \"critical momenta\" (CM). This new approach aims to prevent gradient cancellation and allows the optimizer to escape sharp minima and converge to flat regions which are more likely to generalize. The authors validate their approach through convergence analysis and empirical benchmarks, showing improvements in model performance in both supervised and online learning settings.", "strengths": "- The paper introduces a new concept of storing \"critical momenta\" in a memory buffer, aiming to overcome the limitations of previous memory-augmented optimizers.\n- Intuitions for why this approach is helpful are built by analyzing toy functions.\n- The method has potential applications in both supervised and online learning showing empirical improvements over vanilla Adam with combinations of CG/SAM.\n- The paper is written with a clear structure and logical flow.", "weaknesses": "My main concerns are regarding the memory cost of the approach and whether these costs are justified given the plethora of other alternatives to Adam. The paper does not sufficiently discuss the computational overhead introduced by the memory component, which is critical for practical deployments, especially in resource-constrained environments. \n\nThe experimental results indicate some improvements, but these could potentially be attributed to the additional use of memory and hyperparameters. It would be beneficial for the paper to discuss whether Adam+CM offers any regret guarantees or optimality conditions given the added computational complexity i.e. could there be another algorithm that does better? \n\nOne potential alternative baseline that comes to mind is to consider averaging the weights via approaches such as Polyak averaging or Stochastic Weight Averaging. These methods have been shown to improve generalization performance and only require one additional copy of the parameters if one employs the exponential moving average trick. The paper could be strengthened by comparing Adam+CM to these alternatives--perhaps the two approaches are orthogonal.", "questions": "- I would consider additional measures of sharpness as well. For instance, it seems common to consider $|\\lambda_{max}/\\lambda_5| $ (Jastrzebski et al. (2020)) or measures of how quickly loss changes in a local neighborhood. The interaction between sharpness, distance travelled, and generalization performance seems quite complicated generally as it is also a function of the learning rate and other hyperparameters. This should be made clear in Figure 7 and Figure 8.\n- In general, as mentioned in weaknesses, I would be interested in a more thorough study on how additional memory can best be used. It seems like C is generally set to 5 in the grid search -- can we get away with smaller $C$? What if we used a different value for CG and CM?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of poor generalization performance in adaptive gradient methods like Adam, hypothesizing that adding an exploration strategy with additional memory of gradient information could improve performance. Building on previous work, which used a memory buffer for storing \"critical gradients,\" the authors propose a new memory-augmented Adam optimizer that stores \"critical momenta\" (CM). This new approach aims to prevent gradient cancellation and allows the optimizer to escape sharp minima and converge to flat regions which are more likely to generalize. The authors validate their approach through convergence analysis and empirical benchmarks, showing improvements in model performance in both supervised and online learning settings.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper introduces a new concept of storing \"critical momenta\" in a memory buffer, aiming to overcome the limitations of previous memory-augmented optimizers.\n- Intuitions for why this approach is helpful are built by analyzing toy functions.\n- The method has potential applications in both supervised and online learning showing empirical improvements over vanilla Adam with combinations of CG/SAM.\n- The paper is written with a clear structure and logical flow.", "weaknesses": "My main concerns are regarding the memory cost of the approach and whether these costs are justified given the plethora of other alternatives to Adam. The paper does not sufficiently discuss the computational overhead introduced by the memory component, which is critical for practical deployments, especially in resource-constrained environments. \n\nThe experimental results indicate some improvements, but these could potentially be attributed to the additional use of memory and hyperparameters. It would be beneficial for the paper to discuss whether Adam+CM offers any regret guarantees or optimality conditions given the added computational complexity i.e. could there be another algorithm that does better? \n\nOne potential alternative baseline that comes to mind is to consider averaging the weights via approaches such as Polyak averaging or Stochastic Weight Averaging. These methods have been shown to improve generalization performance and only require one additional copy of the parameters if one employs the exponential moving average trick. The paper could be strengthened by comparing Adam+CM to these alternatives--perhaps the two approaches are orthogonal.", "questions": "- I would consider additional measures of sharpness as well. For instance, it seems common to consider $|\\lambda_{max}/\\lambda_5| $ (Jastrzebski et al. (2020)) or measures of how quickly loss changes in a local neighborhood. The interaction between sharpness, distance travelled, and generalization performance seems quite complicated generally as it is also a function of the learning rate and other hyperparameters. This should be made clear in Figure 7 and Figure 8.\n- In general, as mentioned in weaknesses, I would be interested in a more thorough study on how additional memory can best be used. It seems like C is generally set to 5 in the grid search -- can we get away with smaller $C$? What if we used a different value for CG and CM?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698780447969}], "openreview_url": "https://openreview.net/forum?id=6rEcB9m9AI", "arxiv_id": "2307.09638", "paper_pdf": "papers/6rEcB9m9AI.pdf", "paper_pdf_sha256": "063650ba28cb26183d2bd9e4899e90b7fa64428f374488cb4ad356f55e449245", "paper_pdf_bytes": 5510489, "paper_pdf_source": "openreview", "code_url": "https://github.com/chandar-lab/CMOptimizer", "code_repository": "chandar-lab/CMOptimizer", "code_commit": "bdc77310624345ce88e299137b88e746c079b607", "code_archive": "repos/6rEcB9m9AI.zip", "code_archive_sha256": "99ce7fdb578157596c765c3748a57271c18665e47b05d432df9e09f7917d545c", "code_archive_bytes": 8731, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 25, "github_languages": {"Python": 33464}, "github_archived": false, "github_pushed_at": "2024-06-18T13:34:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/promoting-exploration-in-memory-augmented"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5DkfiQPy9A", "year": 2023, "status": "rejected", "title": "ACAT: Adversarial Counterfactual Attention for Classification and Detection in Medical Imaging", "authors": ["Alessandro Fontanella", "Antreas Antoniou", "Wenwen Li", "Joanna Wardlaw", "Grant Mair", "Emanuele Trucco", "Amos Storkey"], "authorids": ["~Alessandro_Fontanella1", "~Antreas_Antoniou3", "~Wenwen_Li2", "~Joanna_Wardlaw1", "~Grant_Mair1", "~Emanuele_Trucco1", "~Amos_Storkey1"], "authors_source": "OpenReview API", "abstract": "In some medical imaging tasks and other settings where only small parts of the image are informative for the classification task, traditional CNNs can sometimes struggle to generalise. Manually annotated Regions of Interest (ROI) are sometimes used to isolate the most informative parts of the image. However, these are expensive to collect and may vary significantly across annotators. To overcome these issues, we propose a method to generate saliency maps, obtained from adversarially generated counterfactual images. With this method, we are able to isolate the area of interest in brain and lung CT scans without using any manual annotations. Our saliency maps, in the task of localising the lesion location out of 6 possible regions, obtain a score of $65.05 \\%$ on brain CT scans, improving the score of $61.29 \\%$ obtained with the best competing method. We also employ the saliency maps in a framework that refines a classifier pipeline. In particular, the saliency maps are used to obtain soft spatial attention masks that modulate the image features at different scales. We refer to our method as \\emph{Adversarial Counterfactual Attention} (ACAT). ACAT increases the baseline classification accuracy of lesions in brain CT scans from $71.39 \\%$ to $72.55 \\%$ and of COVID-19 related findings in lung CT scans from $67.71 \\%$ to $70.84 \\%$ and exceeds the performance of competing methods.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "cSCd8aRlRJS", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1226/Reviewer_McTW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a method based on saliency maps to guide a classifier in tasks where only part of the image is relevant. The saliency maps are generated using an adversarial gradient descent on a fixed classifier and autoencoder. They are then combined with the classifier at several scales to finetune it. The method is validated on 2 datasets. ", "review_text": "My initial recommendation tend towards rejecting this work: while the results are good, I feel that the novelty of the proposed method is not sufficient and that more ablations studies to fully understand the model are necessary to accept it.", "strengths": "### Strengths:\n- the method is simple (no iteration on the adversarial parts)\n- the shown results are good\n- good related works\n\n### Weaknesses:\n- I think that the main weakness of the method is its novelty: in Flores et al., saliency maps where already generated using an adversarial approach with an auto-encoder and the contributions on this part seems to be the use of several smaller gradient steps instead of one ¨large\" step.\n- Similarly, the network architecture is similar to ones previously used to incorporate saliency maps but with a multi-scale approach. Once again, the results are better with this approach but when a proposed model is a small improvement over a previous one, I feel that more experiments are required to add value to the article: for example, is it really using a multi-scale model that helps or the fact that the smaller scale can detect small RoI? (as suggested by the drop of performance for large RoI).\n- the dropout experiment if I understand it correctly is not fully convincing. If the dropout is not spatially constrained, information from all parts of the image still go through even with p=0.5. Maybe a layer with a random cutout (or RandomErasing in torch) might be more similar to a RoI approach.\n- Maybe adding a small ¨Algorithm box\" with the inputs and outputs of both steps would be good (First step: pretrained model and auto encoder -> RoI detector, 2nd step: baseline + RoI detector -> final classifier) \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a method based on saliency maps to guide a classifier in tasks where only part of the image is relevant. The saliency maps are generated using an adversarial gradient descent on a fixed classifier and autoencoder. They are then combined with the classifier at several scales to finetune it. The method is validated on 2 datasets. ", "strength_and_weaknesses": "### Strengths:\n- the method is simple (no iteration on the adversarial parts)\n- the shown results are good\n- good related works\n\n### Weaknesses:\n- I think that the main weakness of the method is its novelty: in Flores et al., saliency maps where already generated using an adversarial approach with an auto-encoder and the contributions on this part seems to be the use of several smaller gradient steps instead of one ¨large\" step.\n- Similarly, the network architecture is similar to ones previously used to incorporate saliency maps but with a multi-scale approach. Once again, the results are better with this approach but when a proposed model is a small improvement over a previous one, I feel that more experiments are required to add value to the article: for example, is it really using a multi-scale model that helps or the fact that the smaller scale can detect small RoI? (as suggested by the drop of performance for large RoI).\n- the dropout experiment if I understand it correctly is not fully convincing. If the dropout is not spatially constrained, information from all parts of the image still go through even with p=0.5. Maybe a layer with a random cutout (or RandomErasing in torch) might be more similar to a RoI approach.\n- Maybe adding a small ¨Algorithm box\" with the inputs and outputs of both steps would be good (First step: pretrained model and auto encoder -> RoI detector, 2nd step: baseline + RoI detector -> final classifier) \n", "clarity,_quality,_novelty_and_reproducibility": "### Clarity:\nThe article is globally clear (minus a summary of the method, cf weaknesses) and the related work is good.\n\n### Novelty:\nI fell that the novelty is a bit missing\n\n### Reproducibility:\nThe method seems easy to implement and the authors indicated that code would be available.", "summary_of_the_review": "My initial recommendation tend towards rejecting this work: while the results are good, I feel that the novelty of the proposed method is not sufficient and that more ablations studies to fully understand the model are necessary to accept it.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666969160644}, {"id": "bdahhhf6joi", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1226/Reviewer_svqR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes Adversarial Counterfactual Attention (ACAT), which first generates regions of interest (ROI) in (medical) images with saliency (heat)map techniques, and then uses these saliency map ROIs to generate soft spatial attention masks at various scales, that are integrated into a deep learning classifier. The counterfactual images are generated using small progressive shifts in the latent space, to encourage a smooth transition from the original image to the counterfactual. ACAT is reported to increase the baseline classification accuracy of lesions and COVID-19 related findings, from CT scans.\n", "review_text": "This paper proposes an Adversarial Counterfactual Attention (ACAT) method that uses counterfactual saliency maps to guide classifiers through attention. Some uncertainty remains as to whether the proposed counterfactual saliency map method is optimal for the classification task, and as to the evaluation of the saliency maps on annotations.\n", "strengths": "Strengths:\n\n1. Presents a method for improving classification performance by automatically integrating saliency map information.\n\n2. Compares against several existing saliency map fusion methods.\n\n3. Saliency maps of the various methods visualized.\n\nPossible Weaknesses/Considerations:\n\nThe paper appears to contain two main contributions: regularized latent shift to produce counterfactual images, and thus saliency maps, and the use of these saliency maps to guide classification via attention. While some ablation experiments are attempted with other saliency map methods (e.g. Grad-CAM), these experiments might not be fully complete.\n\n4. Segmentation performance with ground truth annotations is reported for the regularized latent shift method against other saliency map methods, in Section 4.3 and Appendix H. To begin with, the attribution map evaluation in the main text appears to consider only “the pixel with the greatest value in each CT scan”, with the annotated region. It is not clear if the consideration of just a single pixel (albeit of greatest value) in an entire CT scan to evaluate the performance of a saliency map, is particularly meaningful.\n\n5. IOU and Dice score performances are then reported in Appendix H, but details such as how each saliency map is thresholded do not appear to be described. Moreover, the set of saliency map methods tried in Appendix H appear not entirely the same as in Section 4.3 (NoRec is missing).\n\n6. The set of saliency map methods tested does not appear very comprehensive; more recent methods such as Integrated Gradients do not appear included.\n\n7. The contribution of different saliency map methods to classification performance (as reported in Tables 1 & 2) does not appear to be evaluated directly, as Table 1 only reports performance against other frameworks.\n\n8. For classification performance, only accuracy is reported. It would be highly recommended for common metrics such as sensitivity, specificity and AUROC to be included, to better reflect the performance of the various classifiers over the full range of possible thresholds.\n\n9. The contribution of multiple attention masks at different scales (Section 3.2.1), instead of at a single scale, does not appear directly evaluated.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes Adversarial Counterfactual Attention (ACAT), which first generates regions of interest (ROI) in (medical) images with saliency (heat)map techniques, and then uses these saliency map ROIs to generate soft spatial attention masks at various scales, that are integrated into a deep learning classifier. The counterfactual images are generated using small progressive shifts in the latent space, to encourage a smooth transition from the original image to the counterfactual. ACAT is reported to increase the baseline classification accuracy of lesions and COVID-19 related findings, from CT scans.\n", "strength_and_weaknesses": "Strengths:\n\n1. Presents a method for improving classification performance by automatically integrating saliency map information.\n\n2. Compares against several existing saliency map fusion methods.\n\n3. Saliency maps of the various methods visualized.\n\nPossible Weaknesses/Considerations:\n\nThe paper appears to contain two main contributions: regularized latent shift to produce counterfactual images, and thus saliency maps, and the use of these saliency maps to guide classification via attention. While some ablation experiments are attempted with other saliency map methods (e.g. Grad-CAM), these experiments might not be fully complete.\n\n4. Segmentation performance with ground truth annotations is reported for the regularized latent shift method against other saliency map methods, in Section 4.3 and Appendix H. To begin with, the attribution map evaluation in the main text appears to consider only “the pixel with the greatest value in each CT scan”, with the annotated region. It is not clear if the consideration of just a single pixel (albeit of greatest value) in an entire CT scan to evaluate the performance of a saliency map, is particularly meaningful.\n\n5. IOU and Dice score performances are then reported in Appendix H, but details such as how each saliency map is thresholded do not appear to be described. Moreover, the set of saliency map methods tried in Appendix H appear not entirely the same as in Section 4.3 (NoRec is missing).\n\n6. The set of saliency map methods tested does not appear very comprehensive; more recent methods such as Integrated Gradients do not appear included.\n\n7. The contribution of different saliency map methods to classification performance (as reported in Tables 1 & 2) does not appear to be evaluated directly, as Table 1 only reports performance against other frameworks.\n\n8. For classification performance, only accuracy is reported. It would be highly recommended for common metrics such as sensitivity, specificity and AUROC to be included, to better reflect the performance of the various classifiers over the full range of possible thresholds.\n\n9. The contribution of multiple attention masks at different scales (Section 3.2.1), instead of at a single scale, does not appear directly evaluated.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The counterfactual saliency map technique is clearly motivated and described, as are the datasets and experiments. The main technical novelties appear the progressive optimization of the latent shift method, and the integration of saliency map attention at multiple scales.\n", "summary_of_the_review": "This paper proposes an Adversarial Counterfactual Attention (ACAT) method that uses counterfactual saliency maps to guide classifiers through attention. Some uncertainty remains as to whether the proposed counterfactual saliency map method is optimal for the classification task, and as to the evaluation of the saliency maps on annotations.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A\n", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666638202919}, {"id": "z48vzEnyIku", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1226/Reviewer_3Yeb"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an attention generation method for ROI detection by adversarial counterfactual without attention label. The attention map can be used to highlight useful information for disease classification and detection. The experiments show its improvements on different medical imaging tasks.  ", "review_text": "the proposed method is interesting, but the novelty is limited", "strengths": "Strengths: \n--The idea using counterfactual images for saliency map generation is interesting.\n\n--The improvement for medical imaging taks is significant. \n\nWeaknesses:\n\n--The novelty is simple and limited. \n\n--More experiments are needed, such as existing counterfactual generation.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes an attention generation method for ROI detection by adversarial counterfactual without attention label. The attention map can be used to highlight useful information for disease classification and detection. The experiments show its improvements on different medical imaging tasks.  ", "strength_and_weaknesses": "Strengths: \n--The idea using counterfactual images for saliency map generation is interesting.\n\n--The improvement for medical imaging taks is significant. \n\nWeaknesses:\n\n--The novelty is simple and limited. \n\n--More experiments are needed, such as existing counterfactual generation.", "clarity,_quality,_novelty_and_reproducibility": "the originality is there, but novelty are limited. No code is provided.", "summary_of_the_review": "the proposed method is interesting, but the novelty is limited", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666623202355}], "openreview_url": "https://openreview.net/forum?id=5DkfiQPy9A", "arxiv_id": "2303.15421", "paper_pdf": "papers/5DkfiQPy9A.pdf", "paper_pdf_sha256": "18258ee1fc1fcb3d74c83d2c759ee43d6491a3cea2db0026234c2f5d6935a20a", "paper_pdf_bytes": 8985424, "paper_pdf_source": "openreview", "code_url": "https://github.com/alessandro-f/ACAT", "code_repository": "alessandro-f/ACAT", "code_commit": "7e4f289e768a41fe24abcf87ff05fc591ea439b0", "code_archive": "repos/5DkfiQPy9A.zip", "code_archive_sha256": "3cb26c3a88531af496a447b75ade008ee579d4c3f1da662d47e95a676c4568bd", "code_archive_bytes": 38592, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 31, "github_languages": {"Python": 301499}, "github_archived": false, "github_pushed_at": "2023-03-24T21:11:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/acat-adversarial-counterfactual-attention-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zyrhwrd9EYs", "year": 2022, "status": "rejected", "title": "To Impute or Not To Impute? Missing Data in Treatment Effect Estimation", "authors": ["Jeroen Berrevoets", "Fergus Imrie", "Trent Kyono", "James Jordon", "Mihaela van der Schaar"], "authorids": ["~Jeroen_Berrevoets1", "~Fergus_Imrie1", "~Trent_Kyono1", "~James_Jordon1", "~Mihaela_van_der_Schaar2"], "authors_source": "OpenReview API", "abstract": "Missing data is a systemic problem in practical scenarios that causes noise and bias when estimating treatment effects. This makes treatment effect estimation from data with missingness a particularly tricky endeavour. A key reason for this is that standard assumptions on missingness are rendered insufficient due to the presence of an additional variable, treatment, besides the individual and the outcome.  Having a treatment variable introduces additional complexity with respect to why some variables are missing that is overlooked by previous work. In our work we identify a new missingness mechanism, which we term mixed confounded missingness (MCM), where some missingness determines treatment selection and other missingness is determined by treatment selection. Given MCM, we show that naively imputing all data leads to poor performing treatment effects models, as the act of imputation effectively removes information necessary to provide unbiased estimates. However, no imputation at all also leads to biased estimates, as missingness determined by treatment divides the population in distinct subpopulations, where estimates across these populations will be biased. Our solution is selective imputation, where we use insights from MCM to inform precisely which variables should be imputed and which should not. We empirically demonstrate how various learners benefit from selective imputation compared to other solutions for missing data.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "vop8d7q0nC4", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3258/Reviewer_efJq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors mainly study the problem of missing data in treatment effect estimation and highlight the importance of addressing this problem. The authors propose a selective imputation scheme which is more well suited for addressing missingness in such scenarios. Authors also present several sample scenarios illustratively which indicate potential issues with current methods while doing TE with missingness.  Empirical results compare different scenarios where general imputation schemes (imputing all data, no data, wrong data) can be much worse than their proposed method.", "review_text": "**Strengths of paper**\n\n- Paper presents a very elaborate writeup with a detailed introduction to causality and missingness in treatment effect estimation which makes it a good read for audience who are new to this field.\n\n- Problem addressed is a relevant problem and is applicable across several real-world problems.\n\n**Weakness**\n\n- The paper has a lot of content on introducing causality which instead could have been used to add more use cases on how selective imputation can be useful for TE\n- It does appear that the authors could have studied the exhaustiveness claim of MCM in depth to tease out the contributions of this paper better. This is a very broad claim with insufficient justification in the current draft.\n- It is not exactly clear how this paper advances the technical state-of-the-art to clear the ICLR bar of acceptance. The general writeup and contributions make it a better fit for a venue like UAI and other similar avenues.\n\nOverall, I would like the authors to think more broadly on the imputation problem itself as industry problems rarely rely on using any kind of imputation method (mainly due to skepticism and avoiding corrupting the data). It might help if authors can talk about broader adoption for this work for such a real-world setting.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors mainly study the problem of missing data in treatment effect estimation and highlight the importance of addressing this problem. The authors propose a selective imputation scheme which is more well suited for addressing missingness in such scenarios. Authors also present several sample scenarios illustratively which indicate potential issues with current methods while doing TE with missingness.  Empirical results compare different scenarios where general imputation schemes (imputing all data, no data, wrong data) can be much worse than their proposed method.", "main_review": "**Strengths of paper**\n\n- Paper presents a very elaborate writeup with a detailed introduction to causality and missingness in treatment effect estimation which makes it a good read for audience who are new to this field.\n\n- Problem addressed is a relevant problem and is applicable across several real-world problems.\n\n**Weakness**\n\n- The paper has a lot of content on introducing causality which instead could have been used to add more use cases on how selective imputation can be useful for TE\n- It does appear that the authors could have studied the exhaustiveness claim of MCM in depth to tease out the contributions of this paper better. This is a very broad claim with insufficient justification in the current draft.\n- It is not exactly clear how this paper advances the technical state-of-the-art to clear the ICLR bar of acceptance. The general writeup and contributions make it a better fit for a venue like UAI and other similar avenues.\n\nOverall, I would like the authors to think more broadly on the imputation problem itself as industry problems rarely rely on using any kind of imputation method (mainly due to skepticism and avoiding corrupting the data). It might help if authors can talk about broader adoption for this work for such a real-world setting.", "summary_of_the_review": "I have highlighted my major observations from this paper above. While the paper is written in a very elaborate manner, I do believe it is still not clear if the technical contributions are advancing the state-of-the-art to clear the ICLR bar of acceptance. Authors should tease out the technical contributions around exhaustiveness and others more clearly and remove big sections on assumptions, metrics (can be written briefly with citations). I would also recommend trimming down the graphs to only keep the most relevant DAGs with a clear messaging.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636035619779}, {"id": "Sw15Q7x-eNY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3258/Reviewer_8bft"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper studies dealing with missing values in estimating treatment effects. Authors identify a new missingness mechanism, mixed confounded missingness (MCM), including missingness that determines treatment selection and missingness that is determined by treatment selection. The authors show that both imputation and no imputation lead to poor treatment effect estimations. The authors present a selective imputation strategy that informs which variables should be imputed and which should not. They empirically demonstrate the effectiveness of the strategy.\n", "review_text": "Authors use graphic causal models to model the various missingness in causal effect estimation. This is an interesting exercise.\n\nHowever, I doubt the correctness of mapping Figure 1(d) to Figure 3. In Figure 1(d) the relationships between the rightmost four variables are \"W \\to Z_2 \\to X'_2 \\leftarrow X\" (Note that I use Z_2 and X'_2 to replace the superscripted Z and X in Figure 1(d)). We should understand the relationships in the following way. W determines the missing values of Z_2 and subsequently the observed values of X'_2. This does not mean that W will determine X_2 when it does not include missing values. Using the authors' example, the participants of the job training program provide additional information whereas the non-participants do not. So, the missing values in X'_2 are determined by W. However, if we have a magical way to collect the additional information from non-participants (i.e. the missing values are imputed perfectly). The edge W \\to real X_2 will be broken. Figure 1(d) becomes Figure 1(c). If X_2 is imputed correctly, there is not an edge W \\to X_2 and the discission before \"selective imputation\" on Page 7 is invalid.   \n\nGiven that Figure 3 is confusing and leads to an incorrect conclusion. There is a soundness problem with the paper.\n\nSince the experiments follow Figure 3, the results cannot be trusted. If X_2 is a part of the covariate set, there is no edge W \\to X_2 without missing values. The causal effect on the set of X is unbiased. If X_2 is imputed correctly, the causal effect is also unbiased. \n\nInterestingly, there is not a discussion of how the missing values are imputed.\n    ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies dealing with missing values in estimating treatment effects. Authors identify a new missingness mechanism, mixed confounded missingness (MCM), including missingness that determines treatment selection and missingness that is determined by treatment selection. The authors show that both imputation and no imputation lead to poor treatment effect estimations. The authors present a selective imputation strategy that informs which variables should be imputed and which should not. They empirically demonstrate the effectiveness of the strategy.\n", "main_review": "Authors use graphic causal models to model the various missingness in causal effect estimation. This is an interesting exercise.\n\nHowever, I doubt the correctness of mapping Figure 1(d) to Figure 3. In Figure 1(d) the relationships between the rightmost four variables are \"W \\to Z_2 \\to X'_2 \\leftarrow X\" (Note that I use Z_2 and X'_2 to replace the superscripted Z and X in Figure 1(d)). We should understand the relationships in the following way. W determines the missing values of Z_2 and subsequently the observed values of X'_2. This does not mean that W will determine X_2 when it does not include missing values. Using the authors' example, the participants of the job training program provide additional information whereas the non-participants do not. So, the missing values in X'_2 are determined by W. However, if we have a magical way to collect the additional information from non-participants (i.e. the missing values are imputed perfectly). The edge W \\to real X_2 will be broken. Figure 1(d) becomes Figure 1(c). If X_2 is imputed correctly, there is not an edge W \\to X_2 and the discission before \"selective imputation\" on Page 7 is invalid.   \n\nGiven that Figure 3 is confusing and leads to an incorrect conclusion. There is a soundness problem with the paper.\n\nSince the experiments follow Figure 3, the results cannot be trusted. If X_2 is a part of the covariate set, there is no edge W \\to X_2 without missing values. The causal effect on the set of X is unbiased. If X_2 is imputed correctly, the causal effect is also unbiased. \n\nInterestingly, there is not a discussion of how the missing values are imputed.\n    ", "summary_of_the_review": "There is a soundness problem with the paper. ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635330744624}, {"id": "-aEdXoBehI4", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3258/Reviewer_ZWmF"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "\"we identify a new missingness mechanism, which we term mixed confounded missingness (MCM), where some missingness determines treatment selection and other missingness is determined by treatment selection.\"\n\nThe author gave a new term called “MCM” which is not something new. MCM is a type of MNAR (Missing-Not-At-Random). \nMCM cannot be listed parallel to MCAR, MAR and MNAR. \n\n", "review_text": "The questions they are trying to solve is an important problem.\nFor example in the discussion the author emphasize that \"more care and thought should\nbe put into imputing missing data when estimating treatment effects.\" which I agree with completely.\n\nThe authors are using simplified DAG graphic tool to modeling the exposure, mediator, outcome variables with missingness and propose the causal relationship of missingness. In the real world data, one covariate may cause missingness of an exposure, which could be the predictor of the outcome variable. However, the missingness of this covariate may have no direct correlation with outcome variable. Therefore, to determine which variable will be imputed or not will not be simply determined by strategy will be limited.\n\nMissingness in one variable may have direct or indirect relationship with missingness in the other variable or outcome variable. Alternative speaking, this correlation may not be linear. Which variables will be imputed or not imputed for the prediction of outcome variable will be  determined using machine learning approach for the feature selection.\n\nFinally, even though it is difficult to use real word data for experimentation, it is important to attempt to do that. The imputation procedure has to be designed for a specific field as pattern of missingness may be very different among disciplines. The pattern of missingness and mechanism for a given dataset should be recognized. Whether the competition is favoring a certain method or procedure has to be determined in the “real-world” data with “real-world” missingness by considering recognized and unrecognized missing pattern/mechanism, as well as the plausible distribution of missing data. Testing a method on synthetic data, with no regards to observed patterns of missingness may only add noise to the field. A recent paper in the field was just published this month: https://www.nature.com/articles/s41746-021-00518-0 \n\nTherefore, even though the idea of improving imputation is excellent, the logic used in this study is not robust.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "\"we identify a new missingness mechanism, which we term mixed confounded missingness (MCM), where some missingness determines treatment selection and other missingness is determined by treatment selection.\"\n\nThe author gave a new term called “MCM” which is not something new. MCM is a type of MNAR (Missing-Not-At-Random). \nMCM cannot be listed parallel to MCAR, MAR and MNAR. \n\n", "main_review": "The questions they are trying to solve is an important problem.\nFor example in the discussion the author emphasize that \"more care and thought should\nbe put into imputing missing data when estimating treatment effects.\" which I agree with completely.\n\nThe authors are using simplified DAG graphic tool to modeling the exposure, mediator, outcome variables with missingness and propose the causal relationship of missingness. In the real world data, one covariate may cause missingness of an exposure, which could be the predictor of the outcome variable. However, the missingness of this covariate may have no direct correlation with outcome variable. Therefore, to determine which variable will be imputed or not will not be simply determined by strategy will be limited.\n\nMissingness in one variable may have direct or indirect relationship with missingness in the other variable or outcome variable. Alternative speaking, this correlation may not be linear. Which variables will be imputed or not imputed for the prediction of outcome variable will be  determined using machine learning approach for the feature selection.\n\nFinally, even though it is difficult to use real word data for experimentation, it is important to attempt to do that. The imputation procedure has to be designed for a specific field as pattern of missingness may be very different among disciplines. The pattern of missingness and mechanism for a given dataset should be recognized. Whether the competition is favoring a certain method or procedure has to be determined in the “real-world” data with “real-world” missingness by considering recognized and unrecognized missing pattern/mechanism, as well as the plausible distribution of missing data. Testing a method on synthetic data, with no regards to observed patterns of missingness may only add noise to the field. A recent paper in the field was just published this month: https://www.nature.com/articles/s41746-021-00518-0 \n\nTherefore, even though the idea of improving imputation is excellent, the logic used in this study is not robust.  ", "summary_of_the_review": "Even though the idea of improving imputation is excellent, the logic used in this study is not robust and does not take into account true realistic situations.  \n\nUse of only synthetic data, when we have access to many real-world datasets is not acceptable. How the synthetic data was created has a direct effect on how the methods will work. It is possible to use real-world data and introduce additional hold-out values to evaluate the methods. see for example: https://www.nature.com/articles/s41746-021-00518-0\n\nThe importance of real-world data is key here because the whole idea behind this paper is that missingness patterns are not completely understood. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634737253988}], "openreview_url": "https://openreview.net/forum?id=zyrhwrd9EYs", "arxiv_id": "2202.02096", "paper_pdf": "papers/zyrhwrd9EYs.pdf", "paper_pdf_sha256": "6f3d75e7d95ac655638164da8a3d65078101c60416d50dfefd8298a06afd3b5e", "paper_pdf_bytes": 306015, "paper_pdf_source": "openreview", "code_url": "https://github.com/jeroenbe/mcm", "code_repository": "jeroenbe/mcm", "code_commit": "41688030608ed8cbd3c9a58169bf3ff03ddf1e54", "code_archive": "repos/zyrhwrd9EYs.zip", "code_archive_sha256": "19fcdfc14b247cf4dcb495b4d9e93019ad18d424def0c05ebd74994eba00324e", "code_archive_bytes": 29813, "code_file_count": 9, "code_extensions": {".ipynb": 6, ".py": 3}, "github_disk_usage_kb": 27, "github_languages": {"Jupyter Notebook": 84481, "Python": 9308}, "github_archived": false, "github_pushed_at": "2023-04-26T22:10:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/to-impute-or-not-to-impute-missing-data-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JAlqRs9duhz", "year": 2021, "status": "rejected", "title": "Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation", "authors": ["Xiang Lin", "SIMENG HAN", "Shafiq Joty"], "authorids": ["~Xiang_Lin2", "~SIMENG_HAN1", "~Shafiq_Joty1"], "authors_source": "OpenReview API", "abstract": "Advanced large-scale neural language models have led to significant success in many natural language generation tasks. However, the most commonly used training objective, Maximum Likelihood Estimation (MLE), has been shown to be problematic, where the trained model prefers using dull and repetitive phrases. In this work, we introduce ScaleGrad, a modification straight to the gradient of the loss function, to remedy the degeneration issues of the standard MLE objective. By directly maneuvering the gradient information, ScaleGrad makes the model learn to use novel tokens during training. Empirical results show the effectiveness of our method not only in open-ended generation, but also in directed generation. With the simplicity in architecture, our method can serve as a general training objective that is applicable to most of the neural text generation tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "84s-GIHdgFU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1675/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a  technique to encourage generating certain tokens (i.e. non-repetitive ones) in text generation. The idea is to scale the softmax probability  for certain words (in the novel set) by a factor of gamma. The authors show how this affects learning by deriving the effect on the gradient. \n\nMany experiments are presented both on open ended generation (language modeling, story telling) as well as abstractive text summarization to justify the method. The model seems to encourage more diversity than unlikelihood training in open ended generation (while still maintaining a lower perplexity). The gains in summarization are marginal. \n\nI have two questions:\n(1) How are lambda and alpha connected to \\gamma in Section 3? This make the method section clearer. \n(2) How much is the model discouraged from generating stop words like \"the\" or \"a\" (and how does this affect fluency)\n\nPros:\n-Well justified and simple method to solve a relevant problem in text generation. \n-Lots of experiments, gains in open ended generation seem decent. \n\nCons:\n-Gains on summarization are marginal / non-existent suggesting that this is not as large of a problem for more constrained tasks.\n-Some clarity on the questions above would be helpful. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "New technique for encouraging novel tokens in text generation.", "review": "The paper presents a  technique to encourage generating certain tokens (i.e. non-repetitive ones) in text generation. The idea is to scale the softmax probability  for certain words (in the novel set) by a factor of gamma. The authors show how this affects learning by deriving the effect on the gradient. \n\nMany experiments are presented both on open ended generation (language modeling, story telling) as well as abstractive text summarization to justify the method. The model seems to encourage more diversity than unlikelihood training in open ended generation (while still maintaining a lower perplexity). The gains in summarization are marginal. \n\nI have two questions:\n(1) How are lambda and alpha connected to \\gamma in Section 3? This make the method section clearer. \n(2) How much is the model discouraged from generating stop words like \"the\" or \"a\" (and how does this affect fluency)\n\nPros:\n-Well justified and simple method to solve a relevant problem in text generation. \n-Lots of experiments, gains in open ended generation seem decent. \n\nCons:\n-Gains on summarization are marginal / non-existent suggesting that this is not as large of a problem for more constrained tasks.\n-Some clarity on the questions above would be helpful. \n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604000902188}, {"id": "Y7_SCEsK8og", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1675/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**I have updated this review after noting the authors’ detailed response.**\n\nThis paper focuses on the problem of “Neural Text Degeneration”—where text sampled from a language model can either be too repetitive and bland or too random and nonsensical. The authors focus largely on the former problem, proposing a finetuning loss that specifically incentivizes the use of tokens that have not yet been decoded in the given document. The authors test whether this improves repetition and unique token coverage with greedy decoding in open-ended generation. A small human study is conducted and the proposed method, ScaleGrad, is found to outperform MLE and Unlikelihood Training (UT). Similarly good results are obtained on Image Captioning with and without trigram repetition blocking. On Abstractive Summarization BeamSearch is used and again outperform MLE and UT. Analysis attempts to make comparisons across different decoding strategies, though coverage of different variations is limited. The authors argue that stochastic decoding is outperformed by ScaleGrad, though they note that trigram blocking still helps ScaleGrad. Multiple hyperparameter settings are shown, with some analysis on how gamma can be chosen to get a desired behavior. Finally, the authors analyze why UT may not be as effective: it penalizes gold repetitions too much and does little for other tokens.\n\nStrength:\n- The results are good for greedy decoding\n- The method is well motivated and well explained\n- The analysis regarding Unlikelihood Training is interesting\n\nWeaknesses\n-  The results shown do not make proper comparisons across models, baselines, and hyperaparameters over all tasks.\n- Results for stochastic decoding should have been shown across tasks.\n- Despite citing the need for awkward rules such as trigram repetition blocking as a reason to propose ScaleGrad, trigram repetition blocking still helps significantly.\n- Some details are hidden away in the appendix, which I had to read thoroughly in order to fully understand the comparison.\n\nI recommend to reject this paper, because the experimental comparisons are not quite fair and because of implicit argumentation about what Greedy decoding can or should do that is never made explicit.\n\nThe following two paragraphs are obsolete, because the authors shared experimental results from a larger set of experiments.\n> The results in Table 1—which show the main metrics of interest on open-ended generation—are missing two key points of comparison: ScaleGrad is only show with gamma=0.2, even though gamma=0.5 & gamma=0.8 are used for the rest of the experiments, giving us little idea of how these metrics change over hyperparameter settings. This is despite the fact that two hyperparameter options for Unlikelihood Training are shown. In a footnote on page 6, for directed generation, the authors state “Although UL was originally proposed for open-ended generation, it is applicable to directed generation. We did  the same scale hyper-parameter search for UL. Details can be seen in Appendix E.” However, in Appendix E two hyperparameter settings for alpha are shown, the same two as used in Table 1, but two hyperparameter settings for gamma in ScaleGrad are shown _neither of which are shown in Table 1_ nor are repetition or uniqueness numbers shown for these hyperparameters settings anywhere in the paper or the appendices. This makes me question whether the improvements shown in Table 1 hold across hyperparameter settings as the authors claim in their analysis of Figure 1.\n\n> However, Figure 1 is missing necessary data points and comparison. First of all gamma=0.2 is not shown, though at least gamma=0.1,0.3 are so it can be somewhat inferred. That is suboptimal, but this graph does not even go up to gamma=0.8, which is what is used in the Abstractive Text Summarization experiment! Furthermore, the number in Figure 1 (b) cannot be directly compared to other decoding methods, because they are an average of repetition metrics shown in Table 1. Luckily, Figure 1 (c) can be compared, and if cross-referenced with Table 1, shows that Unlikelihood Training does better than ScaleGrad with a higher gamma. However, Figure 1 has no data on either Unlikelihood Training or a human baseline. It really should not be necessary to go looking through Table 1, Figure 1, and Appendix F to see that Unlikelihood Training is outperforming ScaleGrad on some metrics. Worse, the data presented in Figure 1 (b) actually makes comparison impossible, which makes me uncomfortable about the universally positive results in Table 1.\n\nOn page 4 the authors write “Following Welleck et al. (2020), we apply greedy decoding in our experiments in this section. This allows us to evaluate the modeling capability exclusively.” We will get into the matter of comparison to Welleck et al. 2020, but I would like to begin by addressing whether Greedy Decoding is a neutral choice that only tests modeling capability, because it is clearly not. There is a spectrum of generation algorithms between probability maximization and straight-forward sampling. Greedy is closer to probability maximization, but it only maximizes local probabilities (Meister et al., 2020) and inevitably comes-up with lower probability outputs than Beam Search or Bound & Branch (Stahlberg & Byrne, 2019). Welleck et al., 2020 show that Greedy Decoding results in better text along their proposed metrics for open-ended generation.\n\nSince Greedy Decoding is not a “neutral” choice, I do not believe it is appropriate to exclude stochastic decoding baselines from the given comparisons. Stochastic decoding algorithms such as sampling, top-k sampling, and Nucleus Sampling usually do very well on repetition and uniqueness metrics. Indeed, they can be seen to outperform all the other models on Table 16 in Appendix H.\n\nIn the analysis section, tables are quite limited in their coverage. In Table 6 no comparisons are made to systems that have not been trained with ScaleGrad, and these algorithms were not reported on in Table 1 so no comparison can properly be made even if the reader goes searching for the data.  In Table 8, Unlikelihood Training is not included in the comparison even though it does very similarly to ScaleGrad on the same task in Table 5. Finally, Table 5 shows that trigram blocking still helps significantly on ScaleGrad trained systems. This is understandable, but disappointing since getting rid of these kind of rules is described as the reason for proposing ScaleGrad.\n\nAltogether, I feel the comparisons made in this paper are not quite convincing and the argument about why Greedy decoding, a deterministic algorithm, should even be able to match the properties of a large, noisy distribution is not properly fleshed-out.\n\nMeister, Clara, Ryan Cotterell, and Tim Vieira. \"If Beam Search Is the Answer, What Was the Question?.\" Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.\n\nStahlberg, Felix, and Bill Byrne. \"On NMT Search Errors and Model Errors: Cat Got Your Tongue?.\" Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Comparisons are Incomplete and Sometimes Occluded", "review": "**I have updated this review after noting the authors’ detailed response.**\n\nThis paper focuses on the problem of “Neural Text Degeneration”—where text sampled from a language model can either be too repetitive and bland or too random and nonsensical. The authors focus largely on the former problem, proposing a finetuning loss that specifically incentivizes the use of tokens that have not yet been decoded in the given document. The authors test whether this improves repetition and unique token coverage with greedy decoding in open-ended generation. A small human study is conducted and the proposed method, ScaleGrad, is found to outperform MLE and Unlikelihood Training (UT). Similarly good results are obtained on Image Captioning with and without trigram repetition blocking. On Abstractive Summarization BeamSearch is used and again outperform MLE and UT. Analysis attempts to make comparisons across different decoding strategies, though coverage of different variations is limited. The authors argue that stochastic decoding is outperformed by ScaleGrad, though they note that trigram blocking still helps ScaleGrad. Multiple hyperparameter settings are shown, with some analysis on how gamma can be chosen to get a desired behavior. Finally, the authors analyze why UT may not be as effective: it penalizes gold repetitions too much and does little for other tokens.\n\nStrength:\n- The results are good for greedy decoding\n- The method is well motivated and well explained\n- The analysis regarding Unlikelihood Training is interesting\n\nWeaknesses\n-  The results shown do not make proper comparisons across models, baselines, and hyperaparameters over all tasks.\n- Results for stochastic decoding should have been shown across tasks.\n- Despite citing the need for awkward rules such as trigram repetition blocking as a reason to propose ScaleGrad, trigram repetition blocking still helps significantly.\n- Some details are hidden away in the appendix, which I had to read thoroughly in order to fully understand the comparison.\n\nI recommend to reject this paper, because the experimental comparisons are not quite fair and because of implicit argumentation about what Greedy decoding can or should do that is never made explicit.\n\nThe following two paragraphs are obsolete, because the authors shared experimental results from a larger set of experiments.\n> The results in Table 1—which show the main metrics of interest on open-ended generation—are missing two key points of comparison: ScaleGrad is only show with gamma=0.2, even though gamma=0.5 & gamma=0.8 are used for the rest of the experiments, giving us little idea of how these metrics change over hyperparameter settings. This is despite the fact that two hyperparameter options for Unlikelihood Training are shown. In a footnote on page 6, for directed generation, the authors state “Although UL was originally proposed for open-ended generation, it is applicable to directed generation. We did  the same scale hyper-parameter search for UL. Details can be seen in Appendix E.” However, in Appendix E two hyperparameter settings for alpha are shown, the same two as used in Table 1, but two hyperparameter settings for gamma in ScaleGrad are shown _neither of which are shown in Table 1_ nor are repetition or uniqueness numbers shown for these hyperparameters settings anywhere in the paper or the appendices. This makes me question whether the improvements shown in Table 1 hold across hyperparameter settings as the authors claim in their analysis of Figure 1.\n\n> However, Figure 1 is missing necessary data points and comparison. First of all gamma=0.2 is not shown, though at least gamma=0.1,0.3 are so it can be somewhat inferred. That is suboptimal, but this graph does not even go up to gamma=0.8, which is what is used in the Abstractive Text Summarization experiment! Furthermore, the number in Figure 1 (b) cannot be directly compared to other decoding methods, because they are an average of repetition metrics shown in Table 1. Luckily, Figure 1 (c) can be compared, and if cross-referenced with Table 1, shows that Unlikelihood Training does better than ScaleGrad with a higher gamma. However, Figure 1 has no data on either Unlikelihood Training or a human baseline. It really should not be necessary to go looking through Table 1, Figure 1, and Appendix F to see that Unlikelihood Training is outperforming ScaleGrad on some metrics. Worse, the data presented in Figure 1 (b) actually makes comparison impossible, which makes me uncomfortable about the universally positive results in Table 1.\n\nOn page 4 the authors write “Following Welleck et al. (2020), we apply greedy decoding in our experiments in this section. This allows us to evaluate the modeling capability exclusively.” We will get into the matter of comparison to Welleck et al. 2020, but I would like to begin by addressing whether Greedy Decoding is a neutral choice that only tests modeling capability, because it is clearly not. There is a spectrum of generation algorithms between probability maximization and straight-forward sampling. Greedy is closer to probability maximization, but it only maximizes local probabilities (Meister et al., 2020) and inevitably comes-up with lower probability outputs than Beam Search or Bound & Branch (Stahlberg & Byrne, 2019). Welleck et al., 2020 show that Greedy Decoding results in better text along their proposed metrics for open-ended generation.\n\nSince Greedy Decoding is not a “neutral” choice, I do not believe it is appropriate to exclude stochastic decoding baselines from the given comparisons. Stochastic decoding algorithms such as sampling, top-k sampling, and Nucleus Sampling usually do very well on repetition and uniqueness metrics. Indeed, they can be seen to outperform all the other models on Table 16 in Appendix H.\n\nIn the analysis section, tables are quite limited in their coverage. In Table 6 no comparisons are made to systems that have not been trained with ScaleGrad, and these algorithms were not reported on in Table 1 so no comparison can properly be made even if the reader goes searching for the data.  In Table 8, Unlikelihood Training is not included in the comparison even though it does very similarly to ScaleGrad on the same task in Table 5. Finally, Table 5 shows that trigram blocking still helps significantly on ScaleGrad trained systems. This is understandable, but disappointing since getting rid of these kind of rules is described as the reason for proposing ScaleGrad.\n\nAltogether, I feel the comparisons made in this paper are not quite convincing and the argument about why Greedy decoding, a deterministic algorithm, should even be able to match the properties of a large, noisy distribution is not properly fleshed-out.\n\nMeister, Clara, Ryan Cotterell, and Tim Vieira. \"If Beam Search Is the Answer, What Was the Question?.\" Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.\n\nStahlberg, Felix, and Bill Byrne. \"On NMT Search Errors and Model Errors: Cat Got Your Tongue?.\" Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603946919276}, {"id": "hZDHvFAk2G", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1675/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose to modify a language model's token-level distributions by rescaling the output probability of tokens that do no appear in the context ('novel tokens'). The authors show improvements over MLE and token-level unlikelihood in terms of repetition, with increases in perplexity. \n\n\n#### Clarity and significance\n- **ScaleGrad motivation**. There are many different ways to change the gradient, e.g. any regularization function, any scaling of the output probabilities, or even gradient clipping modifies the gradients that are used to update the model. As a result, the presentation of their method as \"a modification straight to the gradient of the loss function\" seems odd, and the name ScaleGrad suggests that they are proposing the general notion of rescaling gradients. Instead, they propose to scale a novel set's output probabilities then renormalize.\n\n- **Specific solution**. The method is specifically designed around the 'novel set', which could limit its significance. The authors speculate that they can alter the novel set (e.g. for importance or factual correctness), but this appears to be nontrivial.\n\n- **Unlikelihood discussion**. The discussion in 5.4 deals with the case of $p_{neg}>0.5$, meaning that the probability of the ground-truth token $p_*$ is $<0.5$ (due to normalization). If I'm understanding their argument, the authors argue that the resulting gradient contradicts the fact that the gradient should go to zero at an optimum. However, *the model is not at an optimum* if $p_<0.5$. Could the authors clarify the statements \"the model is not to learn to predict the ground truth tokens correctly\", \"contradicts the optimization principle\", and \"essentially rejects the ground truth token\"? \n\n- **Method for promoting novelty**. It's unclear why this specific method (renormalizing over the novel set) is the best or simplest method for promoting novelty. A downside is that we no longer know which objective the model is optimizing. In the appendix, the authors discuss a variant that uses an additional loss (Section I), yet do not perform an empirical comparison with that or other 'novelty promoting' variations. They argue in Section I that this suffers 'the similar issue as the UL loss', but that issue was unclear (see point above).\n\nOverall I'm borderline on this paper: the authors do perform a lot of experiments and show improvements, but I'm hesitant that scaling novel tokens and renormalizing the model's output distribution is significant.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review: Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation", "review": "The authors propose to modify a language model's token-level distributions by rescaling the output probability of tokens that do no appear in the context ('novel tokens'). The authors show improvements over MLE and token-level unlikelihood in terms of repetition, with increases in perplexity. \n\n\n#### Clarity and significance\n- **ScaleGrad motivation**. There are many different ways to change the gradient, e.g. any regularization function, any scaling of the output probabilities, or even gradient clipping modifies the gradients that are used to update the model. As a result, the presentation of their method as \"a modification straight to the gradient of the loss function\" seems odd, and the name ScaleGrad suggests that they are proposing the general notion of rescaling gradients. Instead, they propose to scale a novel set's output probabilities then renormalize.\n\n- **Specific solution**. The method is specifically designed around the 'novel set', which could limit its significance. The authors speculate that they can alter the novel set (e.g. for importance or factual correctness), but this appears to be nontrivial.\n\n- **Unlikelihood discussion**. The discussion in 5.4 deals with the case of $p_{neg}>0.5$, meaning that the probability of the ground-truth token $p_*$ is $<0.5$ (due to normalization). If I'm understanding their argument, the authors argue that the resulting gradient contradicts the fact that the gradient should go to zero at an optimum. However, *the model is not at an optimum* if $p_<0.5$. Could the authors clarify the statements \"the model is not to learn to predict the ground truth tokens correctly\", \"contradicts the optimization principle\", and \"essentially rejects the ground truth token\"? \n\n- **Method for promoting novelty**. It's unclear why this specific method (renormalizing over the novel set) is the best or simplest method for promoting novelty. A downside is that we no longer know which objective the model is optimizing. In the appendix, the authors discuss a variant that uses an additional loss (Section I), yet do not perform an empirical comparison with that or other 'novelty promoting' variations. They argue in Section I that this suffers 'the similar issue as the UL loss', but that issue was unclear (see point above).\n\nOverall I'm borderline on this paper: the authors do perform a lot of experiments and show improvements, but I'm hesitant that scaling novel tokens and renormalizing the model's output distribution is significant.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603901251847}, {"id": "3_jEVP1T5t", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1675/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Summary\n\nThis work proposes an effective modification of language model token-level distribution during the training which prevents some forms of degeneration such as repetitions and dullness. The approach is based on the idea of encouraging the model to use tokens which were not observed in the previous context so far. In other words, this method changes softmax distribution such that unseen/novel tokens is being rescaled with a given hyper-parameter $\\gamma$ (eq.4). Authors conduct several experiments using different tasks such as open-ended generation, image captioning and abstractive text summarization. As a result they confirm substantial improvement over the standard mle training and **token-level** unlikelihood training. In addition to analysis of their method, authors discuss a potential issue of unlikelihood training criterion and how their approach avoids this issue.\n\n## Strong points\n\n1. Main method is easy to understand and I believe is easy to implement: this work can be a motivating example for future research towards degenerative text generation. From my understanding there are some interesting future work such as setting individial $\\gamma$ for some tokens, specific masks for novel token sets etc.\n\n2. Large-scale experiments **with some human evaluation**. I enjoyed seeing improvements on multiple tasks including summarization (with automatic metrics at least). In addition, analysis of stochastic decoding used in directed generation is meaningful and highlights the importance of this work. Appendix includes detailed description of each experiment protocol including the protocol of human evaluation. Convincing examples of generated continuations are given in the appendix.\n\n3. Code with implemented method and experiments is provided: code is based on fairseq custom module which makes it relatively easy to extend and do more research with it.\n\n## Weak points\n\n1. Misleading comparison choice: authors claim to compare their approach with unlikelihood training (UL) and choose **token-level** UL loss without even mentioning the existence of **sequence-level** UL loss which works better based on the original paper. In fact, the whole narrative looks like sequence-level version does not exist. Simply stating that ScaleGrad is being compared with token-level UL (which works worse than sequence-level) would make future conclusions more clear.\n\n2. Some relevant work got completely ignored. I am not aware of the full variety of prior work for this popular problem these days, but there is one i am aware of: https://arxiv.org/pdf/2003.11963.pdf, where authors *do similar gradient analysis* as here. If this one is missed, I wonder what else may be missing in the related work here.\n\n3. No human evaluation for text summarization. Given known weakness of automatic metrics in text summarization task and the fact that authors did human eval for text completion, I wounder why they decided to exclude it from here (I can totally see budget limitation as one of the factors, and saying this explicitly would be helpful).\n\n4. The potential issue of UL (sec. 5.4) does not look convincing. From my understanding the main line there is \"*UL essentially rejects the ground truth token in such special cases (subject to the choice of the value of $\\alpha$).*\". This statement on its own is not clear to me and seems to be disconnected from the previous one: \"*In this case, the norm increases as pk increases, which contradicts with the optimization principle.*\" I agree that in this case ($\\alpha=1, p_\\text{neg}>0.5$) the norm increases as $p_k$ (prob of ground truth token) increases, but I don't see any problem or contradiction here. From my understanding when $p_\\text{neg}$ goes above some threshold, then norm of the gradient of $p_k$ is growing proportionally to $p_k$. Keep in mind that as $p_k$ is increasing, $p_\\text{neg} > \\frac{1}{1+\\alpha}$ is eventually dissatisfied (because of softmax property), i.e. I don't see any issue. Would be great if authors can elaborate more about this.\n\n## Recommendation\n\nOverall I vote for accepting this work as long as main concerns will be addressed/discussed. This is a decent approach with a strong experimental evidence and it will be useful for text generation community in the future research. I would be even more satisfied if authors can discuss / clarify / address the points I highlighted above.\n\n## Comments and questions\n\n1. stochastic beam search was mentioned as one of the efforts to solve text generation issue, but I believe it is more about doing sampling without replacement on the sequence-level. I am curious if authors may provide some perspective on how stochastic beam may reduce the degeneracy (e.g. compared to simple beam search).\n\n2. In sec. 2.1 teacher forcing is described as being used \"*used to train neural text generation models for faster convergence.*\". I wonder how can one use MLE criterion on the token level without teacher forcing? E.g. if we use predicted token as the context in the next time step, then we have no target truth token for the next time step (since the context is changed). In other words I think that teacher forcing is essential if we aim to maximize the probability of training sequences.\n\n3. In section 2.2: \"*thus reformalizing the probability distribution*\", this wording reformalizing sounds a bit weird to me, but it is clear what authors had in mind.\n\n4. Did you think about combining UL loss (both seq level and token level) with ScaleGrad? From my understanding it is possible since ScaleGrad emphasize novel tokens, and this softmax from scalegrad may be used in the UL loss, which may help even further! Importantly, sequence-level UL would allow to use ScaleGrad on the sequence-level, since there is no need for ground truth target there, and while penalizing the repeating words with UL, ScaleGrad would emphasize the novel words (sounds promising to me). Overall the paper narrative looks as combating the UL method (with some misleading gaps about token vs. sequence level), but to me it looks like they may work together!", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Effective softmax renormalization is reducing degeneracy in neural text generation with MLE training", "review": "## Summary\n\nThis work proposes an effective modification of language model token-level distribution during the training which prevents some forms of degeneration such as repetitions and dullness. The approach is based on the idea of encouraging the model to use tokens which were not observed in the previous context so far. In other words, this method changes softmax distribution such that unseen/novel tokens is being rescaled with a given hyper-parameter $\\gamma$ (eq.4). Authors conduct several experiments using different tasks such as open-ended generation, image captioning and abstractive text summarization. As a result they confirm substantial improvement over the standard mle training and **token-level** unlikelihood training. In addition to analysis of their method, authors discuss a potential issue of unlikelihood training criterion and how their approach avoids this issue.\n\n## Strong points\n\n1. Main method is easy to understand and I believe is easy to implement: this work can be a motivating example for future research towards degenerative text generation. From my understanding there are some interesting future work such as setting individial $\\gamma$ for some tokens, specific masks for novel token sets etc.\n\n2. Large-scale experiments **with some human evaluation**. I enjoyed seeing improvements on multiple tasks including summarization (with automatic metrics at least). In addition, analysis of stochastic decoding used in directed generation is meaningful and highlights the importance of this work. Appendix includes detailed description of each experiment protocol including the protocol of human evaluation. Convincing examples of generated continuations are given in the appendix.\n\n3. Code with implemented method and experiments is provided: code is based on fairseq custom module which makes it relatively easy to extend and do more research with it.\n\n## Weak points\n\n1. Misleading comparison choice: authors claim to compare their approach with unlikelihood training (UL) and choose **token-level** UL loss without even mentioning the existence of **sequence-level** UL loss which works better based on the original paper. In fact, the whole narrative looks like sequence-level version does not exist. Simply stating that ScaleGrad is being compared with token-level UL (which works worse than sequence-level) would make future conclusions more clear.\n\n2. Some relevant work got completely ignored. I am not aware of the full variety of prior work for this popular problem these days, but there is one i am aware of: https://arxiv.org/pdf/2003.11963.pdf, where authors *do similar gradient analysis* as here. If this one is missed, I wonder what else may be missing in the related work here.\n\n3. No human evaluation for text summarization. Given known weakness of automatic metrics in text summarization task and the fact that authors did human eval for text completion, I wounder why they decided to exclude it from here (I can totally see budget limitation as one of the factors, and saying this explicitly would be helpful).\n\n4. The potential issue of UL (sec. 5.4) does not look convincing. From my understanding the main line there is \"*UL essentially rejects the ground truth token in such special cases (subject to the choice of the value of $\\alpha$).*\". This statement on its own is not clear to me and seems to be disconnected from the previous one: \"*In this case, the norm increases as pk increases, which contradicts with the optimization principle.*\" I agree that in this case ($\\alpha=1, p_\\text{neg}>0.5$) the norm increases as $p_k$ (prob of ground truth token) increases, but I don't see any problem or contradiction here. From my understanding when $p_\\text{neg}$ goes above some threshold, then norm of the gradient of $p_k$ is growing proportionally to $p_k$. Keep in mind that as $p_k$ is increasing, $p_\\text{neg} > \\frac{1}{1+\\alpha}$ is eventually dissatisfied (because of softmax property), i.e. I don't see any issue. Would be great if authors can elaborate more about this.\n\n## Recommendation\n\nOverall I vote for accepting this work as long as main concerns will be addressed/discussed. This is a decent approach with a strong experimental evidence and it will be useful for text generation community in the future research. I would be even more satisfied if authors can discuss / clarify / address the points I highlighted above.\n\n## Comments and questions\n\n1. stochastic beam search was mentioned as one of the efforts to solve text generation issue, but I believe it is more about doing sampling without replacement on the sequence-level. I am curious if authors may provide some perspective on how stochastic beam may reduce the degeneracy (e.g. compared to simple beam search).\n\n2. In sec. 2.1 teacher forcing is described as being used \"*used to train neural text generation models for faster convergence.*\". I wonder how can one use MLE criterion on the token level without teacher forcing? E.g. if we use predicted token as the context in the next time step, then we have no target truth token for the next time step (since the context is changed). In other words I think that teacher forcing is essential if we aim to maximize the probability of training sequences.\n\n3. In section 2.2: \"*thus reformalizing the probability distribution*\", this wording reformalizing sounds a bit weird to me, but it is clear what authors had in mind.\n\n4. Did you think about combining UL loss (both seq level and token level) with ScaleGrad? From my understanding it is possible since ScaleGrad emphasize novel tokens, and this softmax from scalegrad may be used in the UL loss, which may help even further! Importantly, sequence-level UL would allow to use ScaleGrad on the sequence-level, since there is no need for ground truth target there, and while penalizing the repeating words with UL, ScaleGrad would emphasize the novel words (sounds promising to me). Overall the paper narrative looks as combating the UL method (with some misleading gaps about token vs. sequence level), but to me it looks like they may work together!", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603820574064}], "openreview_url": "https://openreview.net/forum?id=JAlqRs9duhz", "arxiv_id": "2106.07207", "paper_pdf": "papers/JAlqRs9duhz.pdf", "paper_pdf_sha256": "f6d87ae4824d8d8062607554e7d2b1e9be55d6e97cfe036c93535cf1cc15de91", "paper_pdf_bytes": 779546, "paper_pdf_source": "openreview", "code_url": "https://github.com/shawnlimn/ScaleGrad", "code_repository": "shawnlimn/ScaleGrad", "code_commit": "b2685f9c8680e731316ca6149c1171fda7af5ead", "code_archive": "repos/JAlqRs9duhz.zip", "code_archive_sha256": "2fc69b03145935cc9d1c0a973293a658b520b90c0dd895bf3282d3dea4cf0d6d", "code_archive_bytes": 30116, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 32, "github_languages": {"Python": 81982}, "github_archived": false, "github_pushed_at": "2021-12-28T04:15:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/straight-to-the-gradient-learning-to-use-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hkl_bCVKDr", "year": 2020, "status": "rejected", "title": "Scaleable input gradient regularization for adversarial robustness", "authors": ["Chris Finlay", "Adam M Oberman"], "authorids": ["christopher.finlay@mail.mcgill.ca", "adam.oberman@mcgill.ca"], "authors_source": "OpenReview API", "abstract": "In this work we revisit gradient regularization for adversarial robustness with some new ingredients.  First, we derive new per-image theoretical robustness bounds based on local gradient information. These bounds strongly motivate input gradient regularization.  Second, we implement a scaleable version of input gradient regularization which avoids double backpropagation: adversarially robust ImageNet models are trained in 33 hours on four consumer grade GPUs.  Finally, we show experimentally and through theoretical certification that input gradient regularization is competitive with adversarial training. Moreover we demonstrate that gradient regularization does not lead to gradient obfuscation or gradient masking.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bker_uuRYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper971/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed new regularizer for training robust models that can defend against evasion attack/adversarial examples. It looks to me there are two major novelties here, the authors suggest that 1) use dual norm on the gradient/jacobian as regularizer is tighter/better than the same norm or simply l2 norm; 2) the gradient/jacobian can be estimated by finite difference when moving the weight with a small step towards the gradient direction. \n\nI review the paper with a standard for empirical paper. Please kindly clarify the theoretical contributions if the authors thought the theory is novel and important. \n\nI donot think the empirical results are ready for publication. The final objective, when plugin finite difference into the norm regularizer, looks like logits squeezing https://openreview.net/forum?id=BJlr0j0ctX, or logits pairing https://arxiv.org/abs/1803.06373. Both methods are a little bit controversial. \n\nThere are several issues in the experiments. Several important baselines are missing. The paper did not compare with any of the regularizer-based robust models. When considering efficiency and fast training, the authors also did not compare with recent fast method Shafahi et al. 2019 Free and Zhang et al. 2019 YOPO. \n\nWhen comparing with PGD adversarial training (Madry), in table 1, there is a more than 10% drop on robust accuracy for CIFAR-10 when \\epsilon=8.\n\nFor l2 norm attack, how to interpret table 2? Why not provide accuracy under norm constraint like table 1? \n\nSome uncommon settings are used in the experiments such as ResNeXt-34 and attacking randomly selected 1000 images. \n\nSome relatively minor issues, could the authors elaborate on why the optimal value of max_v l( x+v) - c(v) is the squared dual norm of \\grad l? \n\nTypo in title: scaleable -> scalable\n\n\n========= after rebuttal =============\nI thank the authors for detailed replies. I still cannot support paper because\n(1) the authors emphasize the theoretical contribution and claims the bound are tighter. However, they did not directly compare with any certified robust methods, or previous bounds to support the argument. \n(2) The empirical results look suboptimal. The authors did not convince me why they sampled 1000 images for test for a small CIFAR-10 dataset. The proposed method is 10% less robust comparing to Madry's in table 1.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This paper proposed new regularizer for training robust models that can defend against evasion attack/adversarial examples. It looks to me there are two major novelties here, the authors suggest that 1) use dual norm on the gradient/jacobian as regularizer is tighter/better than the same norm or simply l2 norm; 2) the gradient/jacobian can be estimated by finite difference when moving the weight with a small step towards the gradient direction. \n\nI review the paper with a standard for empirical paper. Please kindly clarify the theoretical contributions if the authors thought the theory is novel and important. \n\nI donot think the empirical results are ready for publication. The final objective, when plugin finite difference into the norm regularizer, looks like logits squeezing https://openreview.net/forum?id=BJlr0j0ctX, or logits pairing https://arxiv.org/abs/1803.06373. Both methods are a little bit controversial. \n\nThere are several issues in the experiments. Several important baselines are missing. The paper did not compare with any of the regularizer-based robust models. When considering efficiency and fast training, the authors also did not compare with recent fast method Shafahi et al. 2019 Free and Zhang et al. 2019 YOPO. \n\nWhen comparing with PGD adversarial training (Madry), in table 1, there is a more than 10% drop on robust accuracy for CIFAR-10 when \\epsilon=8.\n\nFor l2 norm attack, how to interpret table 2? Why not provide accuracy under norm constraint like table 1? \n\nSome uncommon settings are used in the experiments such as ResNeXt-34 and attacking randomly selected 1000 images. \n\nSome relatively minor issues, could the authors elaborate on why the optimal value of max_v l( x+v) - c(v) is the squared dual norm of \\grad l? \n\nTypo in title: scaleable -> scalable\n\n\n========= after rebuttal =============\nI thank the authors for detailed replies. I still cannot support paper because\n(1) the authors emphasize the theoretical contribution and claims the bound are tighter. However, they did not directly compare with any certified robust methods, or previous bounds to support the argument. \n(2) The empirical results look suboptimal. The authors did not convince me why they sampled 1000 images for test for a small CIFAR-10 dataset. The proposed method is 10% less robust comparing to Madry's in table 1.  ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571879021095}, {"id": "rJecA7_sKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper971/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper revisits the method of gradient regularization, which regularizes the loss function by adding the norm of the input gradient, aiming to improve adversarial robustness. The standard gradient regularization is implemented with \"double backpropagation\", which could be time consuming for large networks. This paper proposes to replace the exact gradient by its discretization via finite difference, which is computationally more efficient comparing to \"double backpropagation\".  Experiments are conducted to show the effectiveness of the method in reducing training time. \n\nOverall, the paper is well written but the contribution is very limited and I find some of the experimental comparisons unfair.  Thus I do not support publication of the paper. Here are my detailed arguments.\n\n1) The gradient regularization is not novel\nThe main contribution of the paper is to use the finite difference in the gradient regularization in order to improve the training time.  The method of gradient regularization is not new, hence the contribution is only the computational effectiveness. This would be fine if one could improve the state-of-the-art method's training time by a lot, but according to the experiments, the performance of adversarial robustness is far from the adversarial training, for example in CIFAR10 epsilon=8/255, there is a 12% drop, which is a huge gap. \n\n2) Unfair comparison in the experiments\nIn table 2, the performance of robustness with respect to the l2 norm is presented. However, in the baseline method of adversarial training, the used attack is in l_{infty} norm. This is unfair since l2 norm and l_{infty} norm has very different characteristics. Moreover, the gradient regularization uses l2 norm instead of l_{1} as in table 1, which makes the comparison unfair. \n\nFurthermore, the finite difference improves the training time versus the standard gradient regularization, but it does not imply that the performance will be the same. In particular, it would be better to include the performance of standard gradient regularization in table 1 and 2 as well. \n\n3) Comment on the motivation/theory\nThe theory part is fairly straightforward and it clearly shows that the norm of gradient (w.r.t input) itself is not sufficient to guarantee robustness. As a evidence, even under standard training (for example on MNIST), the gradient norm could be very small, in order of 10^{-4} but still have adversarial example with very small perturbation.  Thus, what we also need is to control how fast this gradient changes (Lipschitz constant or w-bound). However, the gradient norm regularization does not take into account how gradient changes. It is claimed that the finite discretization implicitly reduces how the gradient changes, however, I am not convinced by the argument since as long as we take h to be small, it is still a very local measure. An interesting question would be how h affect the performance and it is not discussed in the paper.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper revisits the method of gradient regularization, which regularizes the loss function by adding the norm of the input gradient, aiming to improve adversarial robustness. The standard gradient regularization is implemented with \"double backpropagation\", which could be time consuming for large networks. This paper proposes to replace the exact gradient by its discretization via finite difference, which is computationally more efficient comparing to \"double backpropagation\".  Experiments are conducted to show the effectiveness of the method in reducing training time. \n\nOverall, the paper is well written but the contribution is very limited and I find some of the experimental comparisons unfair.  Thus I do not support publication of the paper. Here are my detailed arguments.\n\n1) The gradient regularization is not novel\nThe main contribution of the paper is to use the finite difference in the gradient regularization in order to improve the training time.  The method of gradient regularization is not new, hence the contribution is only the computational effectiveness. This would be fine if one could improve the state-of-the-art method's training time by a lot, but according to the experiments, the performance of adversarial robustness is far from the adversarial training, for example in CIFAR10 epsilon=8/255, there is a 12% drop, which is a huge gap. \n\n2) Unfair comparison in the experiments\nIn table 2, the performance of robustness with respect to the l2 norm is presented. However, in the baseline method of adversarial training, the used attack is in l_{infty} norm. This is unfair since l2 norm and l_{infty} norm has very different characteristics. Moreover, the gradient regularization uses l2 norm instead of l_{1} as in table 1, which makes the comparison unfair. \n\nFurthermore, the finite difference improves the training time versus the standard gradient regularization, but it does not imply that the performance will be the same. In particular, it would be better to include the performance of standard gradient regularization in table 1 and 2 as well. \n\n3) Comment on the motivation/theory\nThe theory part is fairly straightforward and it clearly shows that the norm of gradient (w.r.t input) itself is not sufficient to guarantee robustness. As a evidence, even under standard training (for example on MNIST), the gradient norm could be very small, in order of 10^{-4} but still have adversarial example with very small perturbation.  Thus, what we also need is to control how fast this gradient changes (Lipschitz constant or w-bound). However, the gradient norm regularization does not take into account how gradient changes. It is claimed that the finite discretization implicitly reduces how the gradient changes, however, I am not convinced by the argument since as long as we take h to be small, it is still a very local measure. An interesting question would be how h affect the performance and it is not discussed in the paper.  "}, "tcdate": 1571681234147}, {"id": "ryeFpq8mYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper971/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper provides new understandings on adversarial robustness from the perspective of input gradient regularization. Input gradient regularization hasn't been able to achieve comparable robustness to  adversarial training. Built upon existing works, this paper derives two minimum perturbation bounds (L-bound and w-bound) to explain this, as well as other defenses such as Lipschitz regularization and defensive distillation. Taking a step further, this paper proposes to use the finite difference to estimate the input gradients, which not only gives a nice property for reduced modulus of continuity (eg. the w-bound), but also makes the regulation scalable to large networks and datasets. I quite like the theoretical connections derived in this paper. Empirical evidences support their claims, and demonstrate indeed comparable robustness of input gradient regularization to adversarial training.\n\nThe empirical results can be strengthened by including the normal input gradient regularization baseline (using double backpropagation), at least on cifar-10. This is less likely to change the conclusions, but  would be interesting to see the comparisons.\n\nNote that, there are already new progresses in adversarial training:\n[1] Wang, Yisen, et al. \"On the Convergence and Robustness of Adversarial Training.\" ICML, 2019.\n[2] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" ICML, 2019.\n[3] Carmon, Yair, et al. \"Unlabeled data improves adversarial robustness.\" NeurIPS, 2019.\n[4] Uesato, et al. \"Are Labels Required for Improving Adversarial Robustness?\" NeurIPS, 2019.\n\n==========\nAfter rebuttal:\nThanks for the new results. My rating remains the same.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "Summary:\nThis paper provides new understandings on adversarial robustness from the perspective of input gradient regularization. Input gradient regularization hasn't been able to achieve comparable robustness to  adversarial training. Built upon existing works, this paper derives two minimum perturbation bounds (L-bound and w-bound) to explain this, as well as other defenses such as Lipschitz regularization and defensive distillation. Taking a step further, this paper proposes to use the finite difference to estimate the input gradients, which not only gives a nice property for reduced modulus of continuity (eg. the w-bound), but also makes the regulation scalable to large networks and datasets. I quite like the theoretical connections derived in this paper. Empirical evidences support their claims, and demonstrate indeed comparable robustness of input gradient regularization to adversarial training.\n\nThe empirical results can be strengthened by including the normal input gradient regularization baseline (using double backpropagation), at least on cifar-10. This is less likely to change the conclusions, but  would be interesting to see the comparisons.\n\nNote that, there are already new progresses in adversarial training:\n[1] Wang, Yisen, et al. \"On the Convergence and Robustness of Adversarial Training.\" ICML, 2019.\n[2] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" ICML, 2019.\n[3] Carmon, Yair, et al. \"Unlabeled data improves adversarial robustness.\" NeurIPS, 2019.\n[4] Uesato, et al. \"Are Labels Required for Improving Adversarial Robustness?\" NeurIPS, 2019.\n\n==========\nAfter rebuttal:\nThanks for the new results. My rating remains the same.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571150529346}], "openreview_url": "https://openreview.net/forum?id=Hkl_bCVKDr", "arxiv_id": "1905.11468", "paper_pdf": "papers/Hkl_bCVKDr.pdf", "paper_pdf_sha256": "758646b3250b5685d366c1be1740e45e665cd1162580a93e10bf05584c061e61", "paper_pdf_bytes": 357781, "paper_pdf_source": "openreview", "code_url": "https://github.com/cfinlay/tulip", "code_repository": "cfinlay/tulip", "code_commit": "92c83e794127f0feb2770a8887ee9ca61364c194", "code_archive": "repos/Hkl_bCVKDr.zip", "code_archive_sha256": "41ccb795e140c7e3bc51f4de0274e79cc78eb0164fddcdb34665add3e4fdc5ad", "code_archive_bytes": 42298, "code_file_count": 21, "code_extensions": {".py": 19, ".sh": 2}, "github_disk_usage_kb": 36, "github_languages": {"Python": 105919, "Shell": 2040}, "github_archived": false, "github_pushed_at": "2019-07-08T18:38:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/scaleable-input-gradient-regularization-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJMghjA9YX", "year": 2019, "status": "rejected", "title": "Model Comparison for Semantic Grouping", "authors": ["Francisco Vargas", "Kamen Brestnichki", "Nils Hammerla"], "authorids": ["francisco.vargas@babylonhealth.com", "kamen.brestnichki@babylonhealth.com", "nils.hammerla@babylonhealth.com"], "authors_source": "OpenReview API", "abstract": "We introduce a probabilistic framework for quantifying the semantic similarity between two groups of embeddings. We formulate the task of semantic similarity as a model comparison task in which we contrast a generative model which jointly models two sentences versus one that does not. We illustrate how this framework can be used for the Semantic Textual Similarity tasks using clear assumptions about how the embeddings of words are generated. We apply information criteria based model comparison to overcome the shortcomings of Bayesian model comparison, whilst still penalising model complexity. We achieve competitive results by applying the proposed framework with an appropriate choice of likelihood on the STS datasets.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HyI7safM6X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper680/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a Bayesian model comparison based approach for quantifying the semantic similarity between two groups of embeddings (e.g., two sentences). In particular, it proposes to use the difference between the probability that the two groups are from the same model and the probability that they are from different models.\n\nWhile the approach looks interesting, I have a few concerns: \n-- Using the Bayesian model comparison framework seems to be an interesting idea. However, what are the advantages compared to widely used learned models (say, a learned CNN that takes as input two sentences and outputs the similarity score)? The latter can fit the ground-truth labels given by humans, while it's unclear the model comparison leads to good correlation with human judgments. Some discussion should be provided.\n-- The von Mises-Fisher Likelihood is a very simplified model of actual text data. Have you considered using other models? In particular, more sophisticated ones may lead to better performance. \n-- Different information criteria can be plugged in. Are there comparisons? \n-- The experiments are just too simple and incomplete to make reasonable conclusions. For example, it seems compared to SIF there is not much advantage even in the online setting. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea but somewhat incomplete study", "review": "The paper proposes a Bayesian model comparison based approach for quantifying the semantic similarity between two groups of embeddings (e.g., two sentences). In particular, it proposes to use the difference between the probability that the two groups are from the same model and the probability that they are from different models.\n\nWhile the approach looks interesting, I have a few concerns: \n-- Using the Bayesian model comparison framework seems to be an interesting idea. However, what are the advantages compared to widely used learned models (say, a learned CNN that takes as input two sentences and outputs the similarity score)? The latter can fit the ground-truth labels given by humans, while it's unclear the model comparison leads to good correlation with human judgments. Some discussion should be provided.\n-- The von Mises-Fisher Likelihood is a very simplified model of actual text data. Have you considered using other models? In particular, more sophisticated ones may lead to better performance. \n-- Different information criteria can be plugged in. Are there comparisons? \n-- The experiments are just too simple and incomplete to make reasonable conclusions. For example, it seems compared to SIF there is not much advantage even in the online setting. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541709211082}, {"id": "rJlFVJbo3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper680/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a probabilistic model for computing the sentence similarity between two sets of representations in an online fashion (that is, they do not need to see the entire dataset at once as SIF does when using PCA). They evaluate on the STS tasks and outperform competitive baselines like WMD, averaging embeddings, and SIF (without PCA), but they have worse performance that SIF + PCA.\n\nThe paper is clearly written and their model is carefully laid out along with their derivation. My concern with this paper however, is that I feel the paper lacks a motivation, was it derive an online similarity metric that outperforms SIF(without PCA)?\n\nA few experimental questions/comments:\n\nWhat happens to all methods when stop words are not removed? How far does performance fall? I think one reason it might fall (in addition to the reasons given in the paper) is that all vectors are set to have the same norm. For STS tasks, often the norms of these vectors are reduced during training which lessens their influence. What mechanism was used to identify the stop words and does removing these help the other methods (I know in the paper, stop words were removed in the baseline, did this unilaterally improve performance for these methods)?\n\nOverall I do like the paper, however I do find the results to be lackluster. There are many papers on combining word embeddings trained in various ways that have much stronger numbers on STS, but these methods won't be effective with this type of similarity (namely because embeddings must have unit norm in their model). Therefore, I think the paper needs some more motivation and experimental evidence of its superiority over related methods like SIF+PCA in order for it to be accepted.\n\nPROS\n- Probabilistic model with clear design assumptions from which a similarity metric can be derived.\n- Derived similarity metric doesn't require knowledge of the entire dataset (in comparison to SIF + PCA)\n\nCONS\n- Performance seems to be slightly better than SIF, WMD, and averaging word embeddings, but below that of SIF + PCA \n- Unclear motivation for the model, was it derive an online similarity metric that outperforms SIF(without PCA)?\n- Requires the removal of stop words, but doesn't state how these were defined. Minor point, but tuning this could be enough to cause the improvement over related methods.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting model, but would like to see some more motivation", "review": "The authors propose a probabilistic model for computing the sentence similarity between two sets of representations in an online fashion (that is, they do not need to see the entire dataset at once as SIF does when using PCA). They evaluate on the STS tasks and outperform competitive baselines like WMD, averaging embeddings, and SIF (without PCA), but they have worse performance that SIF + PCA.\n\nThe paper is clearly written and their model is carefully laid out along with their derivation. My concern with this paper however, is that I feel the paper lacks a motivation, was it derive an online similarity metric that outperforms SIF(without PCA)?\n\nA few experimental questions/comments:\n\nWhat happens to all methods when stop words are not removed? How far does performance fall? I think one reason it might fall (in addition to the reasons given in the paper) is that all vectors are set to have the same norm. For STS tasks, often the norms of these vectors are reduced during training which lessens their influence. What mechanism was used to identify the stop words and does removing these help the other methods (I know in the paper, stop words were removed in the baseline, did this unilaterally improve performance for these methods)?\n\nOverall I do like the paper, however I do find the results to be lackluster. There are many papers on combining word embeddings trained in various ways that have much stronger numbers on STS, but these methods won't be effective with this type of similarity (namely because embeddings must have unit norm in their model). Therefore, I think the paper needs some more motivation and experimental evidence of its superiority over related methods like SIF+PCA in order for it to be accepted.\n\nPROS\n- Probabilistic model with clear design assumptions from which a similarity metric can be derived.\n- Derived similarity metric doesn't require knowledge of the entire dataset (in comparison to SIF + PCA)\n\nCONS\n- Performance seems to be slightly better than SIF, WMD, and averaging word embeddings, but below that of SIF + PCA \n- Unclear motivation for the model, was it derive an online similarity metric that outperforms SIF(without PCA)?\n- Requires the removal of stop words, but doesn't state how these were defined. Minor point, but tuning this could be enough to cause the improvement over related methods.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541242673469}, {"id": "S1gnfMKc3Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper680/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Main contribution: devising and evaluating a theoretically-sound algorithm for quantifying the semantic similarity between two pieces of text (e.g., two sentences), given pre-trained word embeddings (glove).\n\nClarity:\nThe paper is generally well-written, but I would have liked to see more details regarding the motivation for the work, description of the prior work and discussion of the results. As an example, I could not understand what were the differences between the online and offline settings, with only a reference to the (Arora et al. 2016) paper that does not contain neither \"online\" nor \"offline\". The mathematical derivations are detailed, which is nice.\n\nOriginality:\nThe work looks original. It proposes a method for quantifying semantic similarity that does not rely on cosine similarity.\n\nSignificance:\nI should start by saying I am not a great reviewer for this paper. I am not familiar with the STS dataset and don't have the mathematical background to fully understand the author's algorithm.\nI like to see theoretical work in a field that desperately needs some, but overall I feel the paper could do a much better job at explaining the motivation behind the work, which is limited to \"cosine similarity [...] is not backed by a solid theoretical foundation\".\nI am not convinced of the practicality of the algorithm either: the algorithm seems to improve slightly over the compared approaches (and it is unclear if the differences are significant), and only in some settings. The approach needs to remove stop-words, which is reminiscent of good old feature engineering. Finally, the paper claims better average time complexity than some other methods, but discussing whether the algorithm is faster for common ranges of d (the word embedding dimension) would also have been interesting.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper but lacking both context and comprehensive analyses", "review": "Main contribution: devising and evaluating a theoretically-sound algorithm for quantifying the semantic similarity between two pieces of text (e.g., two sentences), given pre-trained word embeddings (glove).\n\nClarity:\nThe paper is generally well-written, but I would have liked to see more details regarding the motivation for the work, description of the prior work and discussion of the results. As an example, I could not understand what were the differences between the online and offline settings, with only a reference to the (Arora et al. 2016) paper that does not contain neither \"online\" nor \"offline\". The mathematical derivations are detailed, which is nice.\n\nOriginality:\nThe work looks original. It proposes a method for quantifying semantic similarity that does not rely on cosine similarity.\n\nSignificance:\nI should start by saying I am not a great reviewer for this paper. I am not familiar with the STS dataset and don't have the mathematical background to fully understand the author's algorithm.\nI like to see theoretical work in a field that desperately needs some, but overall I feel the paper could do a much better job at explaining the motivation behind the work, which is limited to \"cosine similarity [...] is not backed by a solid theoretical foundation\".\nI am not convinced of the practicality of the algorithm either: the algorithm seems to improve slightly over the compared approaches (and it is unclear if the differences are significant), and only in some settings. The approach needs to remove stop-words, which is reminiscent of good old feature engineering. Finally, the paper claims better average time complexity than some other methods, but discussing whether the algorithm is faster for common ranges of d (the word embedding dimension) would also have been interesting.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess"}, "tcdate": 1541210644099}], "openreview_url": "https://openreview.net/forum?id=HJMghjA9YX", "arxiv_id": "1904.13323", "paper_pdf": "papers/HJMghjA9YX.pdf", "paper_pdf_sha256": "9bdee0d1695d25160f4210512ad49e3b321949804e6ad7ce6d4d14b16ca67493", "paper_pdf_bytes": 406144, "paper_pdf_source": "openreview", "code_url": "https://github.com/babylonhealth/MCSG", "code_repository": "babylonhealth/MCSG", "code_commit": "f445fbba676f84ec590b5d13493dca16bd0641d8", "code_archive": "repos/HJMghjA9YX.zip", "code_archive_sha256": "99485f3fc27e5c8fa62f6735a28f9ddd10ad5eaa8d596600079d057bbc53913a", "code_archive_bytes": 61313, "code_file_count": 31, "code_extensions": {".py": 28, ".bash": 3}, "github_disk_usage_kb": 149, "github_languages": {"Python": 101171, "Shell": 9785, "sed": 2192}, "github_archived": false, "github_pushed_at": "2022-07-13T15:19:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/model-comparison-for-semantic-grouping"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkJKHMW0Z", "year": 2018, "status": "rejected", "title": "Recurrent Relational Networks for complex relational reasoning", "authors": ["Rasmus Berg Palm", "Ulrich Paquet", "Ole Winther"], "authorids": ["rasmusbergpalm@gmail.com", "upaq@google.com", "olwi@dtu.dk"], "authors_source": "OpenReview API", "abstract": "Humans possess an ability to abstractly reason about objects and their interactions, an ability not shared with state-of-the-art deep learning models. Relational networks, introduced by Santoro et al. (2017), add the capacity for relational reasoning to deep neural networks, but are limited in the complexity of the reasoning tasks they can address. We introduce recurrent relational networks which increase the suite of solvable tasks to those that require an order of magnitude more steps of relational reasoning. We use recurrent relational networks to solve Sudoku puzzles and achieve state-of-the-art results by solving 96.6% of the hardest Sudoku puzzles, where relational networks fail to solve any. We also apply our model to the BaBi textual QA dataset solving 19/20 tasks which is competitive with state-of-the-art sparse differentiable neural computers. The recurrent relational network is a general purpose module that can augment any neural network model with the capacity to do many-step relational reasoning.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJFvPvqgz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper838/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes a method called relational network to add relational reasoning capacity to deep neural networks. The previous approach can only perform a single step of relational reasoning, and was evaluated on problems that require at most three steps. The current method address the scalability issue and can solve tasks with orders of magnitude more steps of reasoning. The proposed methods are evaluated on two problems, Sudoku and Babi, and achieved state-of-the-art results. \n\nThe proposed method should be better explained. What’s the precise definition of interface? It’s claimed that other constraint propagation-based methods can solve Sudoku problems easily, but don’t respect the interface. It is hard to appreciate without a precise definition of interface. The proposed recurrent relational networks are only defined informally. A definition of the model as well as related algorithms should be defined more formally. \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposed method should be better explained", "rating": "5: Marginally below acceptance threshold", "review": "This paper describes a method called relational network to add relational reasoning capacity to deep neural networks. The previous approach can only perform a single step of relational reasoning, and was evaluated on problems that require at most three steps. The current method address the scalability issue and can solve tasks with orders of magnitude more steps of reasoning. The proposed methods are evaluated on two problems, Sudoku and Babi, and achieved state-of-the-art results. \n\nThe proposed method should be better explained. What’s the precise definition of interface? It’s claimed that other constraint propagation-based methods can solve Sudoku problems easily, but don’t respect the interface. It is hard to appreciate without a precise definition of interface. The proposed recurrent relational networks are only defined informally. A definition of the model as well as related algorithms should be defined more formally. \n\n\n\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511843680817}, {"id": "r17v3MDxG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper838/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduced recurrent relational network (RRNs), an enhanced version of the\nexisting relational network, that can be added to any neural networks to add\nrelational reasoning capacity. RRNs are illustrated on sudoku puzzles and textual QA.\n\nOverall the paper is well written and structured. It also addresses an important research question: combining relational reasoning and neural networks is currently receiving a lot of attention, in particular when generally considering the question of bridging sub-symbolic and symbolic methods. Unfortunately, it is current form, the paper has two major downsides. First of all,  the sudoku example does not illustrate “complex relational reasoning” as claimed in the title. The problem is encoded at a positional level where \nmessages encoded as MLPs and LSTMs implement the constraints for sudoko. Indeed, \nthis allows to realise end-to-end learning but does not illustrate complex reasoning. \nThis is also reflected in the considered QA task, which is essentially coded as a positional problem. Consequently, the claim of the conclusions, namely that “we have\nproposed a general relational reasoning model” is not validated, unfortunately. Such\na module that can be connected to any existing neural network would be great. However, \nfor that one should show capabilities of relational logic. Some standard (noisy) \nreasoning capabilities such as modus ponens. This also leads me to the second downside. \nUnfortunately, the paper falls short on discussion related work. First of all, \nthere is the large field of statistical relational learning, see \n\nLuc De Raedt, Kristian Kersting, Sriraam Natarajan, David Poole:\nStatistical Relational Artificial Intelligence: Logic, Probability, and Computation. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers 2016\n\nfor a recent overview. As it has the very same goals, while not using a neural architecture for implementation, it is very much related and has to be discussed. That\none can also use a neural implementation can be seen in \n\nIvan Donadello, Luciano Serafini, Artur S. d'Avila Garcez:\nLogic Tensor Networks for Semantic Image Interpretation. IJCAI 2017: 1596-1602\n\nMatko Bosnjak, Tim Rocktäschel, Jason Naradowsky, Sebastian Riedel:\nProgramming with a Differentiable Forth Interpreter. ICML 2017: 547-556\n\nLuciano Serafini, Artur S. d'Avila Garcez:\nLearning and Reasoning with Logic Tensor Networks. AI*IA 2016: 334-348\n\nGustav Sourek, Vojtech Aschenbrenner, Filip Zelezný, Ondrej Kuzelka:\nLifted Relational Neural Networks. CoCo@NIPS 2015\n\nTim Rocktäschel, Sebastian Riedel:\nEnd-to-end Differentiable Proving. CoRR abs/1705.11040 (2017)\n\nWilliam W. Cohen, Fan Yang, Kathryn Mazaitis:\nTensorLog: Deep Learning Meets Probabilistic DBs. CoRR abs/1707.05390 (2017)\n\nto list just some approaches. There are also (deep) probabilistic programming \napproaches such as Edward that should be mentioned as CPS like problems (Sudoku) can\ndefinitely be implement there. Moreover, there is a number of papers that discuss \nembeddings of relational data and rules such as \n\nWilliam Yang Wang, William W. Cohen:\nLearning First-Order Logic Embeddings via Matrix Factorization. IJCAI 2016: 2132-2138\n\nThomas Demeester, Tim Rocktäschel, Sebastian Riedel:\nLifted Rule Injection for Relation Embeddings. EMNLP 2016: 1389-1399\n\nand even neural-symbolic approaches with a long publication history. Unfortunately, \nnon of these approaches has been cited, giving the wrong impression that this is \nthe first paper that tackles the long lasting question of merging sub-symbolic and symbolic reasoning. BTW, there have been also other deep networks for optimisation, see e.g. \n\nBrandon Amos, J. Zico Kolter:\nOptNet: Differentiable Optimization as a Layer in Neural Networks. \nICML 2017: 136-145\n\nthat have also considered Sudoku. To summarise, I like very much the direction of the paper but it seems to be too early to be published. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Very nice direction but no complex relational reasoning demonstrated and missing related work", "rating": "3: Clear rejection", "review": "The paper introduced recurrent relational network (RRNs), an enhanced version of the\nexisting relational network, that can be added to any neural networks to add\nrelational reasoning capacity. RRNs are illustrated on sudoku puzzles and textual QA.\n\nOverall the paper is well written and structured. It also addresses an important research question: combining relational reasoning and neural networks is currently receiving a lot of attention, in particular when generally considering the question of bridging sub-symbolic and symbolic methods. Unfortunately, it is current form, the paper has two major downsides. First of all,  the sudoku example does not illustrate “complex relational reasoning” as claimed in the title. The problem is encoded at a positional level where \nmessages encoded as MLPs and LSTMs implement the constraints for sudoko. Indeed, \nthis allows to realise end-to-end learning but does not illustrate complex reasoning. \nThis is also reflected in the considered QA task, which is essentially coded as a positional problem. Consequently, the claim of the conclusions, namely that “we have\nproposed a general relational reasoning model” is not validated, unfortunately. Such\na module that can be connected to any existing neural network would be great. However, \nfor that one should show capabilities of relational logic. Some standard (noisy) \nreasoning capabilities such as modus ponens. This also leads me to the second downside. \nUnfortunately, the paper falls short on discussion related work. First of all, \nthere is the large field of statistical relational learning, see \n\nLuc De Raedt, Kristian Kersting, Sriraam Natarajan, David Poole:\nStatistical Relational Artificial Intelligence: Logic, Probability, and Computation. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers 2016\n\nfor a recent overview. As it has the very same goals, while not using a neural architecture for implementation, it is very much related and has to be discussed. That\none can also use a neural implementation can be seen in \n\nIvan Donadello, Luciano Serafini, Artur S. d'Avila Garcez:\nLogic Tensor Networks for Semantic Image Interpretation. IJCAI 2017: 1596-1602\n\nMatko Bosnjak, Tim Rocktäschel, Jason Naradowsky, Sebastian Riedel:\nProgramming with a Differentiable Forth Interpreter. ICML 2017: 547-556\n\nLuciano Serafini, Artur S. d'Avila Garcez:\nLearning and Reasoning with Logic Tensor Networks. AI*IA 2016: 334-348\n\nGustav Sourek, Vojtech Aschenbrenner, Filip Zelezný, Ondrej Kuzelka:\nLifted Relational Neural Networks. CoCo@NIPS 2015\n\nTim Rocktäschel, Sebastian Riedel:\nEnd-to-end Differentiable Proving. CoRR abs/1705.11040 (2017)\n\nWilliam W. Cohen, Fan Yang, Kathryn Mazaitis:\nTensorLog: Deep Learning Meets Probabilistic DBs. CoRR abs/1707.05390 (2017)\n\nto list just some approaches. There are also (deep) probabilistic programming \napproaches such as Edward that should be mentioned as CPS like problems (Sudoku) can\ndefinitely be implement there. Moreover, there is a number of papers that discuss \nembeddings of relational data and rules such as \n\nWilliam Yang Wang, William W. Cohen:\nLearning First-Order Logic Embeddings via Matrix Factorization. IJCAI 2016: 2132-2138\n\nThomas Demeester, Tim Rocktäschel, Sebastian Riedel:\nLifted Rule Injection for Relation Embeddings. EMNLP 2016: 1389-1399\n\nand even neural-symbolic approaches with a long publication history. Unfortunately, \nnon of these approaches has been cited, giving the wrong impression that this is \nthe first paper that tackles the long lasting question of merging sub-symbolic and symbolic reasoning. BTW, there have been also other deep networks for optimisation, see e.g. \n\nBrandon Amos, J. Zico Kolter:\nOptNet: Differentiable Optimization as a Layer in Neural Networks. \nICML 2017: 136-145\n\nthat have also considered Sudoku. To summarise, I like very much the direction of the paper but it seems to be too early to be published. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511627867042}, {"id": "Syc551clG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper838/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces recurrent relational networks: a deep neural network for structured prediction (or relational reasoning). The authors use it to achieve state-of-the-art performance on Soduku puzzles and the BaBi task (a text based QA dataset designed as a set of to toy prerequisite tasks for reasoning).\n\nOverall I think that by itself the algorithm suggested in the paper is not enough to be presented in ICLR, and on the other hand the authors didn't show it has a big impact (could do so by adding more tasks - as they suggest in the discussion). This is why I think the paper is marginally below the acceptance threshold but could be convinced otherwise.\n\nC an the authors give experimental evidences for their claim: \"As such, the network could use a small part of the hidden state for retaining a current best guess, which might remain constant over several steps, and other parts of the hidden state for running a non-greedy...\" - \n\nPros\n- The idea of the paper is clearly presented, the algorithm is easy to follow.\n- The motivation to do better relational reasoning is clear and the network suggested in this paper succeeds to achieve it in the challenging tasks.\n\nCons\n- The recurrent relational networks is basically a complex learned message passing algorithm. As the authors themselves state there are several works from recent years which also tackle this (one missing reference is Deeply Learning the Messages in Message Passing Inference of Lin et al from NIPS 2016). It would been interesting to compare results to these algorithms.\n- For the Sudoku the proposed architecture of the network seems a bit to complex, for example why do a 16 embedding is needed for representing a digit between 0-9? Some other choices (batch size of 252) seem very specific.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting, clearly presented new structured prediction algorithm. Paper marginally below acceptance threshold.", "rating": "5: Marginally below acceptance threshold", "review": "This paper introduces recurrent relational networks: a deep neural network for structured prediction (or relational reasoning). The authors use it to achieve state-of-the-art performance on Soduku puzzles and the BaBi task (a text based QA dataset designed as a set of to toy prerequisite tasks for reasoning).\n\nOverall I think that by itself the algorithm suggested in the paper is not enough to be presented in ICLR, and on the other hand the authors didn't show it has a big impact (could do so by adding more tasks - as they suggest in the discussion). This is why I think the paper is marginally below the acceptance threshold but could be convinced otherwise.\n\nC an the authors give experimental evidences for their claim: \"As such, the network could use a small part of the hidden state for retaining a current best guess, which might remain constant over several steps, and other parts of the hidden state for running a non-greedy...\" - \n\nPros\n- The idea of the paper is clearly presented, the algorithm is easy to follow.\n- The motivation to do better relational reasoning is clear and the network suggested in this paper succeeds to achieve it in the challenging tasks.\n\nCons\n- The recurrent relational networks is basically a complex learned message passing algorithm. As the authors themselves state there are several works from recent years which also tackle this (one missing reference is Deeply Learning the Messages in Message Passing Inference of Lin et al from NIPS 2016). It would been interesting to compare results to these algorithms.\n- For the Sudoku the proposed architecture of the network seems a bit to complex, for example why do a 16 embedding is needed for representing a digit between 0-9? Some other choices (batch size of 252) seem very specific.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511811730296}], "openreview_url": "https://openreview.net/forum?id=SkJKHMW0Z", "arxiv_id": null, "paper_pdf": "papers/SkJKHMW0Z.pdf", "paper_pdf_sha256": "d950d569ecc1b1115ca03ae059a72f1d8cb415f8c1d2a1a1618a2579c8ede64c", "paper_pdf_bytes": 425384, "paper_pdf_source": "openreview", "code_url": "https://github.com/Kyubyong/sudoku", "code_repository": "Kyubyong/sudoku", "code_commit": "cb611230ad602aafebe6914c3e53c81671bba605", "code_archive": "repos/SkJKHMW0Z.zip", "code_archive_sha256": "6a16f3941c49bd63ce45c541d4f95e5f40d8aa0e282111ce00e361d6fb2e8b1d", "code_archive_bytes": 52938, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 100, "github_languages": {"Python": 23747}, "github_archived": false, "github_pushed_at": "2023-02-17T05:22:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recurrent-relational-networks-for-complex"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LDGAjUuEVZ", "year": 2026, "status": "rejected", "title": "BrowseComp-ZH: Benchmarking Web Browsing Ability of Large Language Models in Chinese", "authors": ["Peilin Zhou", "Junling Liu", "Xiang Ying", "Can Zhang", "Yifan Shao", "Qichen Ye", "Dading Chong", "Zhiling Jin", "Chenxuan Xie", "Meng Cao", "Yuxin Gu", "Sixin Hong", "Jing Ren", "Jian Chen", "Chao Liu", "Yining Hua", "Sunghun Kim"], "authorids": ["~Peilin_Zhou1", "~Junling_Liu1", "~Xiang_Ying3", "~Can_Zhang2", "~Yifan_Shao3", "~Qichen_Ye1", "~Dading_Chong1", "~Zhiling_Jin3", "~Chenxuan_Xie3", "~Meng_Cao1", "~Yuxin_Gu1", "~Sixin_Hong1", "~Jing_Ren2", "~Jian_Chen18", "~Chao_Liu8", "~Yining_Hua1", "~Sunghun_Kim1"], "authors_source": "OpenReview API", "abstract": "As large language models (LLMs) evolve into web-interacting agents, their ability to retrieve and reason over real-time information has become a crucial benchmark for general intelligence. However, existing benchmarks such as BrowseComp focus solely on English, neglecting the linguistic, infrastructural, and retrieval-specific challenges posed by other information ecosystems—particularly the Chinese web. We present BrowseComp-ZH, a high-difficulty, natively-constructed benchmark designed to assess LLM agents’ web browsing abilities in Chinese. Rather than translating from English, all questions in BrowseComp-ZH are written from scratch by native speakers to reflect authentic information-seeking behaviors and cultural contexts. The dataset comprises 289 multi-hop questions across 11 diverse domains, each reverse-engineered from a short, verifiable answer and filtered through a twostage quality control pipeline to ensure retrieval hardness and answer uniqueness. We evaluate over 20 leading LLMs and agentic search systems. Despite strong language and retrieval abilities, most models perform poorly: many score below 10% accuracy, and only a few exceed 20%. Even the best system achieves just 42.9% accuracy. These results highlight the considerable difficulty of BrowseComp-ZH, where success requires not only robust retrieval strategies but also advanced multihop reasoning and information reconciliation—abilities that remain challenging for current models. BrowseComp-ZH thus serves as a stress test for web-interactive LLMs beyond English, offering a rigorous and linguistically diverse evaluation framework to guide future research on multilingual agent capabilities.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "O3YgcbhsTo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16708/Reviewer_bkpL"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors introduce BrowseComp-ZH, a new benchmark for evaluating the web-browsing ability of LLMs in Chinese. The authors created reverse-engineered, multi-constraint questions that push models to do multi-step reasoning across real Chinese web sources like Baidu, Zhihu, and WeChat articles. Expert annotators curated the dataset through several rounds of filtering to keep only challenging questions with clear, unique answers. The final dataset covers 15 diverse domains. Evaluation results show that DeepResearch hit 42.9% accuracy, human searchers averaged 17.6%, and most general-purpose LLMs stayed below 10%.", "review_text": "The authors introduce BrowseComp-ZH, a new benchmark for evaluating the web-browsing ability of LLMs in Chinese. The authors created reverse-engineered, multi-constraint questions that push models to do multi-step reasoning across real Chinese web sources like Baidu, Zhihu, and WeChat articles. Expert annotators curated the dataset through several rounds of filtering to keep only challenging questions with clear, unique answers. The final dataset covers 15 diverse domains. Evaluation results show that DeepResearch hit 42.9% accuracy, human searchers averaged 17.6%, and most general-purpose LLMs stayed below 10%.", "strengths": "- This work faithfully reimplements the BrowseComp methodology in the Chinese-language ecosystem, where the data reflects how Chinese users search online.\n- All queries, evidence chains, and browsing steps are authored natively in Chinese, not translated.\n- Reporting calibration alongside accuracy improves interpretability of model performance.", "weaknesses": "- The benchmark creation process almost exactly mirrors BrowseComp’s methodology, and the contribution of this work feels somewhat incremental. The core takeaway is essentially: current LLMs (agents) perform poorly when browsing in Chinese.\n- The idea of cultural groundedness is not empirically verified. Checking how often agent trajectories only involve Chinese sources would provide some insights on how much of the task is really dependent on navigating Chinese sources. The paper claims all queries, evidence chains, and browsing steps are “authored in Chinese,” but does not verify that all necessary evidence exists exclusively on Chinese websites.\n- There is no analysis of failure modes. An important question to explore is how often errors stem from retrieval shortcomings versus misinterpretations of Chinese cultural or linguistic context.", "questions": "- Could you provide more details on the human baseline experiment? How were the annotators hired, and how were examples distributed among them? Were there cases where people gave up?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce BrowseComp-ZH, a new benchmark for evaluating the web-browsing ability of LLMs in Chinese. The authors created reverse-engineered, multi-constraint questions that push models to do multi-step reasoning across real Chinese web sources like Baidu, Zhihu, and WeChat articles. Expert annotators curated the dataset through several rounds of filtering to keep only challenging questions with clear, unique answers. The final dataset covers 15 diverse domains. Evaluation results show that DeepResearch hit 42.9% accuracy, human searchers averaged 17.6%, and most general-purpose LLMs stayed below 10%.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- This work faithfully reimplements the BrowseComp methodology in the Chinese-language ecosystem, where the data reflects how Chinese users search online.\n- All queries, evidence chains, and browsing steps are authored natively in Chinese, not translated.\n- Reporting calibration alongside accuracy improves interpretability of model performance.", "weaknesses": "- The benchmark creation process almost exactly mirrors BrowseComp’s methodology, and the contribution of this work feels somewhat incremental. The core takeaway is essentially: current LLMs (agents) perform poorly when browsing in Chinese.\n- The idea of cultural groundedness is not empirically verified. Checking how often agent trajectories only involve Chinese sources would provide some insights on how much of the task is really dependent on navigating Chinese sources. The paper claims all queries, evidence chains, and browsing steps are “authored in Chinese,” but does not verify that all necessary evidence exists exclusively on Chinese websites.\n- There is no analysis of failure modes. An important question to explore is how often errors stem from retrieval shortcomings versus misinterpretations of Chinese cultural or linguistic context.", "questions": "- Could you provide more details on the human baseline experiment? How were the annotators hired, and how were examples distributed among them? Were there cases where people gave up?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761972915929}, {"id": "41a8c0pYEq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16708/Reviewer_3iN7"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces BrowseComp-ZH, a native Chinese benchmark for evaluating LLMs’ web browsing and reasoning abilities. The dataset contains 289 expert-authored, multi-hop questions across 11 domains, validated for retrieval difficulty and answer uniqueness. Over 20 models—including open-, closed-source, and commercial agents—are evaluated, showing generally poor performance (most <10% accuracy), while multi-round retrieval systems like DeepResearch achieve the best results. The benchmark offers a rigorous, culturally grounded evaluation of LLM agents in the Chinese web ecosystem.", "review_text": "This paper introduces BrowseComp-ZH, a native Chinese benchmark for evaluating LLMs’ web browsing and reasoning abilities. The dataset contains 289 expert-authored, multi-hop questions across 11 domains, validated for retrieval difficulty and answer uniqueness. Over 20 models—including open-, closed-source, and commercial agents—are evaluated, showing generally poor performance (most <10% accuracy), while multi-round retrieval systems like DeepResearch achieve the best results. The benchmark offers a rigorous, culturally grounded evaluation of LLM agents in the Chinese web ecosystem.", "strengths": "- The benchmark offers a rigorous, culturally grounded evaluation of LLM agents in the Chinese web ecosystem.\n  - The benchmark is carefully constructed with dual validation for difficulty and answer uniqueness, ensuring high data quality.\n  - It systematically compares over 20 systems, including human baselines, revealing meaningful performance gaps.", "weaknesses": "- The paper’s main contribution appears incremental rather than novel. From a task construction perspective, BrowseComp-ZH closely mirrors the original BrowseComp benchmark, with the primary difference being its adaptation to the Chinese web environment rather than a fundamentally new methodology or task design.\n  - The benchmark contains only 289 queries, which may not be sufficient to capture the full linguistic, structural, and retrieval complexity of the Chinese internet. Such a limited number of examples could constrain the benchmark’s representativeness and the statistical robustness of the reported results.\n  - The paper lacks a deeper analysis of why models perform differently on BrowseComp-ZH. It remains unclear whether the observed performance gaps stem mainly from the change of language (English → Chinese) or from the distinct characteristics of the Chinese web ecosystem. A more detailed ablation or cross-lingual comparison would strengthen the interpretation of results.", "questions": "- Have you compared model performance between BrowseComp-ZH and the original English BrowseComp benchmark? Specifically, do models exhibit consistent relative rankings or performance gaps across the two languages? Including such a cross-lingual comparison or correlation analysis would help clarify whether the observed difficulties stem from linguistic or retrieval-ecosystem differences.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces BrowseComp-ZH, a native Chinese benchmark for evaluating LLMs’ web browsing and reasoning abilities. The dataset contains 289 expert-authored, multi-hop questions across 11 domains, validated for retrieval difficulty and answer uniqueness. Over 20 models—including open-, closed-source, and commercial agents—are evaluated, showing generally poor performance (most <10% accuracy), while multi-round retrieval systems like DeepResearch achieve the best results. The benchmark offers a rigorous, culturally grounded evaluation of LLM agents in the Chinese web ecosystem.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The benchmark offers a rigorous, culturally grounded evaluation of LLM agents in the Chinese web ecosystem.\n  - The benchmark is carefully constructed with dual validation for difficulty and answer uniqueness, ensuring high data quality.\n  - It systematically compares over 20 systems, including human baselines, revealing meaningful performance gaps.", "weaknesses": "- The paper’s main contribution appears incremental rather than novel. From a task construction perspective, BrowseComp-ZH closely mirrors the original BrowseComp benchmark, with the primary difference being its adaptation to the Chinese web environment rather than a fundamentally new methodology or task design.\n  - The benchmark contains only 289 queries, which may not be sufficient to capture the full linguistic, structural, and retrieval complexity of the Chinese internet. Such a limited number of examples could constrain the benchmark’s representativeness and the statistical robustness of the reported results.\n  - The paper lacks a deeper analysis of why models perform differently on BrowseComp-ZH. It remains unclear whether the observed performance gaps stem mainly from the change of language (English → Chinese) or from the distinct characteristics of the Chinese web ecosystem. A more detailed ablation or cross-lingual comparison would strengthen the interpretation of results.", "questions": "- Have you compared model performance between BrowseComp-ZH and the original English BrowseComp benchmark? Specifically, do models exhibit consistent relative rankings or performance gaps across the two languages? Including such a cross-lingual comparison or correlation analysis would help clarify whether the observed difficulties stem from linguistic or retrieval-ecosystem differences.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761904016610}, {"id": "lpKvKG3fT2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16708/Reviewer_yxRg"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper introduces BrowseComp-ZH, a new, high-difficulty benchmark designed to evaluate the web browsing and information retrieval capabilities of Large Language Models (LLMs) specifically within the Chinese web ecosystem.\n\nKey Contributions:\n* Natively Constructed: It consists of 289 questions written from scratch by native Chinese speakers to reflect authentic user behavior, rather than being translated from English.\n\n* High Difficulty: The questions are multi-hop, requiring complex reasoning and the ability to synthesize information from multiple sources.\n\n* Rigorous Design: Questions were reverse-engineered from verified answers and passed through a two-stage quality control process to ensure they are difficult to solve and have unique answers.", "review_text": "This paper introduces BrowseComp-ZH, a new, high-difficulty benchmark designed to evaluate the web browsing and information retrieval capabilities of Large Language Models (LLMs) specifically within the Chinese web ecosystem.\n\nKey Contributions:\n* Natively Constructed: It consists of 289 questions written from scratch by native Chinese speakers to reflect authentic user behavior, rather than being translated from English.\n\n* High Difficulty: The questions are multi-hop, requiring complex reasoning and the ability to synthesize information from multiple sources.\n\n* Rigorous Design: Questions were reverse-engineered from verified answers and passed through a two-stage quality control process to ensure they are difficult to solve and have unique answers.", "strengths": "1. Fills a Critical Gap: It’s the first benchmark for evaluating LLMs’ web-browsing abilities in the Chinese ecosystem, avoiding flaws of translated English benchmarks by using native Chinese questions that reflect linguistic, cultural, and Chinese web-specific traits. \n\n2. Rigorous Construction: Its 289 multi-hop questions (11 domains) go through strict quality control—reverse-engineered from answers, tested on major search engines to ensure difficulty, and validated for answer uniqueness. \n\n3. Insightful Evaluation: It assesses over 20 systems (open/closed-source models, commercial tools) and includes human baselines (17.6% accuracy), revealing key findings (e.g., reasoning and multi-round retrieval boost performance) to guide LLM improvement. \n\n4. Practical Guidance: Serves as a targeted stress test for non-English web LLMs, aiding the development of multilingual AI agents for real-world Chinese web scenarios.", "weaknesses": "1. Small Dataset Scale: With only 289 questions, it is smaller than English counterparts like BrowseComp, though the authors note this stems from high curation costs. \n\n2. Incomplete Answer Uniqueness: Despite strict checks, the reverse-design approach cannot fully guarantee no alternative valid answers exist for some questions. There are also some obvious bugs in the dataset that need to be fixed.\n\n3. Compared to BrowseComp, BrowseComp-ZH is insufficiently challenging. It is essentially a simplified Chinese adaptation that lacks originality.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces BrowseComp-ZH, a new, high-difficulty benchmark designed to evaluate the web browsing and information retrieval capabilities of Large Language Models (LLMs) specifically within the Chinese web ecosystem.\n\nKey Contributions:\n* Natively Constructed: It consists of 289 questions written from scratch by native Chinese speakers to reflect authentic user behavior, rather than being translated from English.\n\n* High Difficulty: The questions are multi-hop, requiring complex reasoning and the ability to synthesize information from multiple sources.\n\n* Rigorous Design: Questions were reverse-engineered from verified answers and passed through a two-stage quality control process to ensure they are difficult to solve and have unique answers.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Fills a Critical Gap: It’s the first benchmark for evaluating LLMs’ web-browsing abilities in the Chinese ecosystem, avoiding flaws of translated English benchmarks by using native Chinese questions that reflect linguistic, cultural, and Chinese web-specific traits. \n\n2. Rigorous Construction: Its 289 multi-hop questions (11 domains) go through strict quality control—reverse-engineered from answers, tested on major search engines to ensure difficulty, and validated for answer uniqueness. \n\n3. Insightful Evaluation: It assesses over 20 systems (open/closed-source models, commercial tools) and includes human baselines (17.6% accuracy), revealing key findings (e.g., reasoning and multi-round retrieval boost performance) to guide LLM improvement. \n\n4. Practical Guidance: Serves as a targeted stress test for non-English web LLMs, aiding the development of multilingual AI agents for real-world Chinese web scenarios.", "weaknesses": "1. Small Dataset Scale: With only 289 questions, it is smaller than English counterparts like BrowseComp, though the authors note this stems from high curation costs. \n\n2. Incomplete Answer Uniqueness: Despite strict checks, the reverse-design approach cannot fully guarantee no alternative valid answers exist for some questions. There are also some obvious bugs in the dataset that need to be fixed.\n\n3. Compared to BrowseComp, BrowseComp-ZH is insufficiently challenging. It is essentially a simplified Chinese adaptation that lacks originality.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761882570380}, {"id": "w0rfk7vqQq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16708/Reviewer_8svw"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces BrowseComp-ZH, a benchmark for evaluating LLM agents’ web browsing and information-seeking abilities specifically within the Chinese web ecosystem. The dataset comprises 289 human-authored, multi-hop, multi-constraint questions across 11 knowledge domains, meticulously reverse-engineered from verifiable answers. The paper also provides thorough statistics, calibration analysis, multiple failure case studies, and benchmark reproducibility details.", "review_text": "This paper introduces BrowseComp-ZH, a benchmark for evaluating LLM agents’ web browsing and information-seeking abilities specifically within the Chinese web ecosystem. The dataset comprises 289 human-authored, multi-hop, multi-constraint questions across 11 knowledge domains, meticulously reverse-engineered from verifiable answers. The paper also provides thorough statistics, calibration analysis, multiple failure case studies, and benchmark reproducibility details.", "strengths": "- The paper fills a clear gap by constructing the first natively-annotated Chinese web browsing benchmark, avoiding translation artifacts and capturing authentic linguistic, cultural, and infrastructural challenges unique to the Chinese internet.\n- The reverse-design pipeline is impressively meticulous, resulting in a high-quality, stress-test-style dataset.", "weaknesses": "- At only 289 questions, BrowseComp-ZH is considerably smaller than many contemporary benchmarks. This small scale could weaken robustness for fine-grained model assessment or pre-train/fine-tune settings in the future.\n- Despite employing multi-agent and human-in-the-loop processes, the guarantee of a single unique answer still depends on current model/retrieval limitations and annotator creativity.", "questions": "- How were the final domain-specific question counts determined? Is there a risk of over-representation of certain topics?\n- How can the annotators ensure that the difficulty of the problem is sufficient and that this benchmark has the ability to evaluate the entire RAG system?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces BrowseComp-ZH, a benchmark for evaluating LLM agents’ web browsing and information-seeking abilities specifically within the Chinese web ecosystem. The dataset comprises 289 human-authored, multi-hop, multi-constraint questions across 11 knowledge domains, meticulously reverse-engineered from verifiable answers. The paper also provides thorough statistics, calibration analysis, multiple failure case studies, and benchmark reproducibility details.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The paper fills a clear gap by constructing the first natively-annotated Chinese web browsing benchmark, avoiding translation artifacts and capturing authentic linguistic, cultural, and infrastructural challenges unique to the Chinese internet.\n- The reverse-design pipeline is impressively meticulous, resulting in a high-quality, stress-test-style dataset.", "weaknesses": "- At only 289 questions, BrowseComp-ZH is considerably smaller than many contemporary benchmarks. This small scale could weaken robustness for fine-grained model assessment or pre-train/fine-tune settings in the future.\n- Despite employing multi-agent and human-in-the-loop processes, the guarantee of a single unique answer still depends on current model/retrieval limitations and annotator creativity.", "questions": "- How were the final domain-specific question counts determined? Is there a risk of over-representation of certain topics?\n- How can the annotators ensure that the difficulty of the problem is sufficient and that this benchmark has the ability to evaluate the entire RAG system?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761821580012}], "openreview_url": "https://openreview.net/forum?id=LDGAjUuEVZ", "arxiv_id": "2504.19314", "paper_pdf": "papers/LDGAjUuEVZ.pdf", "paper_pdf_sha256": "4e9829515cc04f79f052e5748ac09ee69bd03e467b9862cf902210a43dadf7b2", "paper_pdf_bytes": 3669748, "paper_pdf_source": "openreview", "code_url": "https://github.com/PALIN2018/BrowseComp-ZH", "code_repository": "PALIN2018/BrowseComp-ZH", "code_commit": "86abe635e7deef89ec00c68ff1c2588f0e2f2099", "code_archive": "repos/LDGAjUuEVZ.zip", "code_archive_sha256": "e6902b2303ea2b55b592e8efd96f3d893cd3510bf5340714bfaf191537ecbce2", "code_archive_bytes": 104576, "code_file_count": 5, "code_extensions": {".py": 4, ".sh": 1}, "github_disk_usage_kb": 107, "github_languages": {"Python": 27501, "Shell": 291}, "github_archived": false, "github_pushed_at": "2025-05-14T07:18:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/browsecomp-zh-benchmarking-web-browsing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DRKkO2Tejc", "year": 2025, "status": "rejected", "title": "Label Privacy in Split Learning for Large Models with Parameter-Efficient Training", "authors": ["Philip Zmushko", "Marat Mansurov", "Ruslan Svirschevski", "Denis Kuznedelev", "Max Ryabinin", "Aleksandr Beznosikov"], "authorids": ["~Philip_Zmushko1", "~Marat_Mansurov1", "~Ruslan_Svirschevski1", "~Denis_Kuznedelev1", "~Max_Ryabinin1", "~Aleksandr_Beznosikov1"], "authors_source": "OpenReview API", "abstract": "As deep learning models become larger and more expensive, many practitioners turn to fine-tuning APIs. \nThese web services allow fine-tuning a model between two parties: the client that provides the data, and the server that hosts the model.\nWhile convenient, these APIs raise a new concern: the data of the client is at risk of privacy breach during the training procedure.\nThis challenge presents an important practical case of vertical federated learning, where the two parties perform parameter-efficient fine-tuning (PEFT) of a large model.\nIn this study, we systematically search for a way to fine-tune models over an API *while keeping the labels private*.\nWe analyze the privacy of LoRA, a popular approach for parameter-efficient fine-tuning when training over an API.\nUsing this analysis, we propose P$^3$EFT, a multi-party  split learning algorithm that takes advantage of existing PEFT properties to maintain privacy at a lower performance overhead.\nTo validate our algorithm, we fine-tune DeBERTa-v2-XXLarge, Flan-T5 Large and LLaMA-2 7B using LoRA adapters on a range of NLP tasks. We find that P$^3$EFT is competitive with existing privacy-preserving methods in multi-party and  two-party setups while having higher accuracy.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Nr9p0SsaiQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8052/Reviewer_88W1"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a privacy-preserving approach to safeguard label privacy during the fine-tuning of large language models (LLMs) using Parameter-Efficient Fine-Tuning (PEFT) techniques. It addresses a two-party learning scenario, where users can access forward and backward APIs to update the model. Experiments are conducted to evaluate the approach across various language models.", "review_text": "This paper proposes a privacy-preserving approach to safeguard label privacy during the fine-tuning of large language models (LLMs) using Parameter-Efficient Fine-Tuning (PEFT) techniques. It addresses a two-party learning scenario, where users can access forward and backward APIs to update the model. Experiments are conducted to evaluate the approach across various language models.", "strengths": "The paper addresses a critical question of how to protect privacy in client-server fine-tuning settings. The author proposes a privacy-preserving approach to safeguard label privacy in vertical learning and evaluates this method across various models and NLP tasks.", "weaknesses": "The paper claims that backpropagation is conditionally linear in output gradients and attempts to decompose gradients to enhance privacy. However, based on my understanding, if decomposed gradients are backpropagated through earlier layers, how can this linearity be maintained?\n\nThe proposed approach requires the user to decompose gradients into separate parts and make multiple API calls. What is the associated computation and communication overhead? Additionally, to my knowledge, I am not aware of any LLM companies that offer APIs for forward and backward passes in this way. Could the authors provide examples of such an application if available?\n\nIf the client is involved in parameter updates and can observe the training process, how is it ensured that the client cannot access private information during training?\n\nRegarding the experimental results, I don’t see a clear definition of what \"leak\" represents in the tables. My assumption is that it refers to the client’s prediction accuracy. If this is correct, with prediction accuracy over 60%, it doesn’t seem that label privacy is effectively preserved.", "questions": "Can you provide some definitions to describe the metrics used to evaluate the approach?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a privacy-preserving approach to safeguard label privacy during the fine-tuning of large language models (LLMs) using Parameter-Efficient Fine-Tuning (PEFT) techniques. It addresses a two-party learning scenario, where users can access forward and backward APIs to update the model. Experiments are conducted to evaluate the approach across various language models.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper addresses a critical question of how to protect privacy in client-server fine-tuning settings. The author proposes a privacy-preserving approach to safeguard label privacy in vertical learning and evaluates this method across various models and NLP tasks.", "weaknesses": "The paper claims that backpropagation is conditionally linear in output gradients and attempts to decompose gradients to enhance privacy. However, based on my understanding, if decomposed gradients are backpropagated through earlier layers, how can this linearity be maintained?\n\nThe proposed approach requires the user to decompose gradients into separate parts and make multiple API calls. What is the associated computation and communication overhead? Additionally, to my knowledge, I am not aware of any LLM companies that offer APIs for forward and backward passes in this way. Could the authors provide examples of such an application if available?\n\nIf the client is involved in parameter updates and can observe the training process, how is it ensured that the client cannot access private information during training?\n\nRegarding the experimental results, I don’t see a clear definition of what \"leak\" represents in the tables. My assumption is that it refers to the client’s prediction accuracy. If this is correct, with prediction accuracy over 60%, it doesn’t seem that label privacy is effectively preserved.", "questions": "Can you provide some definitions to describe the metrics used to evaluate the approach?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730778477743}, {"id": "MZFzfiKUqV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8052/Reviewer_ESUj"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper proposes $P^3EFT$, based on PEFT for fine-tuning models in a 2-party setting common to fine-tuning APIs in practice. The focus is on guaranteeing a form of empirical label-privacy. Their approach utilises a private backpropagation approach inspired by secure-aggregation to prevent the server from inferring label information and further trains a mixture of multiple adapters to improve privacy.", "review_text": "This paper proposes $P^3EFT$, based on PEFT for fine-tuning models in a 2-party setting common to fine-tuning APIs in practice. The focus is on guaranteeing a form of empirical label-privacy. Their approach utilises a private backpropagation approach inspired by secure-aggregation to prevent the server from inferring label information and further trains a mixture of multiple adapters to improve privacy.", "strengths": "- The paper focuses on an interesting and relevant privacy problem related to fine-tuning APIs which are increasingly popular.\n- Strong empirical results which show that the proposed method can maintain both accuracy and empirical label-privacy.\n- The presentation of the paper is clear and well-written.", "weaknesses": "- More could be done to compare with existing baseline methods in the experiments section (see below).\n- Some experimental results are unclear (see below).", "questions": "1. It is implied that training multiple adapters is a way to prevent label-leakage, yet in the experiments only n=2 adapters are used. Did you run any experiments varying the number of adapters and what effect did this have on utility and privacy? Further to this, how small are the adapters that are used? It would be useful to include some information about the overhead that is required for clients.\n2. I think some of the Appendix results with the PSLF method could be added into the main paper tables to show there is a clear utility loss when utilising a label DP approach and that there are advantages to using a more empirical method like $P^3EFT$. \n3. What does it mean in Table 6 to train with an $\\varepsilon = 0$ with PSLF?\n4. How many times are the experiments repeated over? There is often a large amount of variance in the privacy leakage e.g., on QNLI in Table 1 and Table 3 which can make DC and $P^3EFT$ seem quite comparable.\n5. It could help to move Algorithm 2 (or a more concise version of it) into the main text from the Appendix to help make the final $P^3EFT$ method clear.\n6. The abstract mentions that $P^3EFT$ is competitive to methods in both a multi-party and 2-party setting. As far as I can tell, the experiments in Section 4 are focused only on a 2-party setting? How does this process change or scale to a multi-party setting?\n7. Minor: Typo L485 algorothm → algorithm", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes $P^3EFT$, based on PEFT for fine-tuning models in a 2-party setting common to fine-tuning APIs in practice. The focus is on guaranteeing a form of empirical label-privacy. Their approach utilises a private backpropagation approach inspired by secure-aggregation to prevent the server from inferring label information and further trains a mixture of multiple adapters to improve privacy.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper focuses on an interesting and relevant privacy problem related to fine-tuning APIs which are increasingly popular.\n- Strong empirical results which show that the proposed method can maintain both accuracy and empirical label-privacy.\n- The presentation of the paper is clear and well-written.", "weaknesses": "- More could be done to compare with existing baseline methods in the experiments section (see below).\n- Some experimental results are unclear (see below).", "questions": "1. It is implied that training multiple adapters is a way to prevent label-leakage, yet in the experiments only n=2 adapters are used. Did you run any experiments varying the number of adapters and what effect did this have on utility and privacy? Further to this, how small are the adapters that are used? It would be useful to include some information about the overhead that is required for clients.\n2. I think some of the Appendix results with the PSLF method could be added into the main paper tables to show there is a clear utility loss when utilising a label DP approach and that there are advantages to using a more empirical method like $P^3EFT$. \n3. What does it mean in Table 6 to train with an $\\varepsilon = 0$ with PSLF?\n4. How many times are the experiments repeated over? There is often a large amount of variance in the privacy leakage e.g., on QNLI in Table 1 and Table 3 which can make DC and $P^3EFT$ seem quite comparable.\n5. It could help to move Algorithm 2 (or a more concise version of it) into the main text from the Appendix to help make the final $P^3EFT$ method clear.\n6. The abstract mentions that $P^3EFT$ is competitive to methods in both a multi-party and 2-party setting. As far as I can tell, the experiments in Section 4 are focused only on a 2-party setting? How does this process change or scale to a multi-party setting?\n7. Minor: Typo L485 algorothm → algorithm", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730734124657}, {"id": "DynPG5BxDb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8052/Reviewer_SHmL"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper addresses the problem of parameter-efficient fine-tuning (PEFT) over an API, focusing on preserving the privacy of training labels. It starts by analyzing the label privacy implications of API fine-tuning with LoRA, a commonly used PEFT algorithm. Empirical results demonstrate that in a split learning setup, LoRA may leak the client’s training labels. Specifically, the unprotected transmission of gradients, model parameters, and activations in split learning can expose training labels to privacy attacks. To address this vulnerability, the paper introduces the $\\text{P}^3\\text{EFT}$ (privacy-preserving parameter-efficient fine-tuning) framework. $\\text{P}^3\\text{EFT}$ employs private backpropagation and obfuscates learned activations to protect label privacy by securing activations and gradients during communication between client and server. In their PEFT experiments, the paper shows that $\\text{P}^3\\text{EFT}$ achieves better label privacy with a small reduction in utility compared to previous split learning methods. Privacy leakage is measured using metrics such as Spectral attack AUC, Norm attack AUC, and k-means accuracy.", "review_text": "This paper addresses the problem of parameter-efficient fine-tuning (PEFT) over an API, focusing on preserving the privacy of training labels. It starts by analyzing the label privacy implications of API fine-tuning with LoRA, a commonly used PEFT algorithm. Empirical results demonstrate that in a split learning setup, LoRA may leak the client’s training labels. Specifically, the unprotected transmission of gradients, model parameters, and activations in split learning can expose training labels to privacy attacks. To address this vulnerability, the paper introduces the $\\text{P}^3\\text{EFT}$ (privacy-preserving parameter-efficient fine-tuning) framework. $\\text{P}^3\\text{EFT}$ employs private backpropagation and obfuscates learned activations to protect label privacy by securing activations and gradients during communication between client and server. In their PEFT experiments, the paper shows that $\\text{P}^3\\text{EFT}$ achieves better label privacy with a small reduction in utility compared to previous split learning methods. Privacy leakage is measured using metrics such as Spectral attack AUC, Norm attack AUC, and k-means accuracy.", "strengths": "1) Although $\\text{P}^3\\text{EFT}$ is based on the analysis of LoRA, its approach can be applied to any PEFT method aiming to address label privacy in split learning during PEFT over an API.\n\n2) The concept of private backpropagation leverages the conditional linearity of the backpropagation operator, which, while conceptually similar to the secure aggregation protocol used in horizontal federated learning, is well-suited for the vertical FL setting of this problem. This approach can also extend to other vertical FL or split learning scenarios requiring gradient propagation through an untrusted party and is adaptable for future algorithms.\n\n3) $\\text{P}^3\\text{EFT}$ performs significantly better than existing PEFT methods in preserving label privacy, with only a minor drop in accuracy. Moreover, compared to the distance correlation defense (DC), another privacy-aware algorithm, $\\text{P}^3\\text{EFT}$ achieves superior label privacy, making it a strong choice for API-based PEFT scenarios.", "weaknesses": "1) While label privacy in $\\text{P}^3\\text{EFT}$ is evaluated using metrics such as Spectral attack AUC, Norm attack AUC, and k-means accuracy, demonstrating strong performance on these measures, it lacks a formal theoretical guarantee (e.g., differential privacy). Although differential privacy is mentioned as providing loose upper bounds on potential privacy leakage, establishing a theoretical privacy guarantee for $\\text{P}^3\\text{EFT}$ would enhance understanding of privacy leakage and offer a more solid theoretical basis for fair comparisons with other label privacy-focused algorithms.\n\n2) This paper addresses label privacy in PEFT over an API within an \"honest but curious\" attacker model, making it a unique but somewhat limited scenario in terms of generality. Analyzing cases where the server is not \"honest\" would provide a broader perspective, especially given the emphasis on worst-case privacy scenarios in the paper. Additionally, extending the framework to protect both input and label privacy would be a natural generalization, though it appears challenging to adapt $\\text{P}^3\\text{EFT}$ for this purpose.\n\n3) Some references in the paper would benefit from clarification. For instance, line 52 refers to private multi-party fine-tuning methods that support a narrow class of fine-tuning algorithms but does not specify which algorithms fall within this scope and which do not. Similarly, line 132 states that differential privacy upper bounds are loose and do not align with practical observations on real models, but it lacks a reference or supporting argument, leaving some statements unclear.\n\n4) There are a few minor writing issues, including a formulation error, which I have included in the question section for reference.", "questions": "1) In line 237, it is stated that leaving gradients, activations, or parameters unprotected would compromise label privacy. While $\\text{P}^3\\text{EFT}$ obfuscates gradients and activations, it is unclear why parameters do not require similar privacy measures in the context of this paper.\n\n2) In line 266, it appears that the expression $\\text{backprop}(x, \\theta, g_h)$ should be written as $\\frac{1}{2} \\Bigl( \\text{backprop}(x, \\theta, g_h + z) + \\text{backprop}(x, \\theta, g_h - z) \\Bigr)$ instead of $\\Bigl( \\text{backprop}(x, \\theta, g_h + z) + \\text{backprop}(x, \\theta, g_h - z) \\Bigr)$. Could you confirm if this is correct?\n\n3) Regarding Algorithm 1, did you study the effects of different noise levels in the private backpropagation algorithm? Does this imply that $\\text{P}^3\\text{EFT}$ would yield the same utility across various noise levels?\n\n4) In line 479, it is mentioned that $\\text{P}^3\\text{EFT}$ achieves the same accuracy at a given privacy level. However, it is unclear what “same level of privacy” specifically means, given that the privacy metrics in Tables 1 and 2 differ for DC and $\\text{P}^3\\text{EFT}$.\n\n5) Finally, in line 503, how is the privacy-accuracy trade-off precisely defined? It would be helpful to clarify this term—if it refers to an average measure of privacy and utility, please specify, as the term “trade-off” here is ambiguous without further explanation.\n\n6) Here are some minor spelling errors:\nline 326: it's \nline 478: $\\text{P}^3\\text{FT}$\nline 484: both both our algorithm", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of parameter-efficient fine-tuning (PEFT) over an API, focusing on preserving the privacy of training labels. It starts by analyzing the label privacy implications of API fine-tuning with LoRA, a commonly used PEFT algorithm. Empirical results demonstrate that in a split learning setup, LoRA may leak the client’s training labels. Specifically, the unprotected transmission of gradients, model parameters, and activations in split learning can expose training labels to privacy attacks. To address this vulnerability, the paper introduces the $\\text{P}^3\\text{EFT}$ (privacy-preserving parameter-efficient fine-tuning) framework. $\\text{P}^3\\text{EFT}$ employs private backpropagation and obfuscates learned activations to protect label privacy by securing activations and gradients during communication between client and server. In their PEFT experiments, the paper shows that $\\text{P}^3\\text{EFT}$ achieves better label privacy with a small reduction in utility compared to previous split learning methods. Privacy leakage is measured using metrics such as Spectral attack AUC, Norm attack AUC, and k-means accuracy.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1) Although $\\text{P}^3\\text{EFT}$ is based on the analysis of LoRA, its approach can be applied to any PEFT method aiming to address label privacy in split learning during PEFT over an API.\n\n2) The concept of private backpropagation leverages the conditional linearity of the backpropagation operator, which, while conceptually similar to the secure aggregation protocol used in horizontal federated learning, is well-suited for the vertical FL setting of this problem. This approach can also extend to other vertical FL or split learning scenarios requiring gradient propagation through an untrusted party and is adaptable for future algorithms.\n\n3) $\\text{P}^3\\text{EFT}$ performs significantly better than existing PEFT methods in preserving label privacy, with only a minor drop in accuracy. Moreover, compared to the distance correlation defense (DC), another privacy-aware algorithm, $\\text{P}^3\\text{EFT}$ achieves superior label privacy, making it a strong choice for API-based PEFT scenarios.", "weaknesses": "1) While label privacy in $\\text{P}^3\\text{EFT}$ is evaluated using metrics such as Spectral attack AUC, Norm attack AUC, and k-means accuracy, demonstrating strong performance on these measures, it lacks a formal theoretical guarantee (e.g., differential privacy). Although differential privacy is mentioned as providing loose upper bounds on potential privacy leakage, establishing a theoretical privacy guarantee for $\\text{P}^3\\text{EFT}$ would enhance understanding of privacy leakage and offer a more solid theoretical basis for fair comparisons with other label privacy-focused algorithms.\n\n2) This paper addresses label privacy in PEFT over an API within an \"honest but curious\" attacker model, making it a unique but somewhat limited scenario in terms of generality. Analyzing cases where the server is not \"honest\" would provide a broader perspective, especially given the emphasis on worst-case privacy scenarios in the paper. Additionally, extending the framework to protect both input and label privacy would be a natural generalization, though it appears challenging to adapt $\\text{P}^3\\text{EFT}$ for this purpose.\n\n3) Some references in the paper would benefit from clarification. For instance, line 52 refers to private multi-party fine-tuning methods that support a narrow class of fine-tuning algorithms but does not specify which algorithms fall within this scope and which do not. Similarly, line 132 states that differential privacy upper bounds are loose and do not align with practical observations on real models, but it lacks a reference or supporting argument, leaving some statements unclear.\n\n4) There are a few minor writing issues, including a formulation error, which I have included in the question section for reference.", "questions": "1) In line 237, it is stated that leaving gradients, activations, or parameters unprotected would compromise label privacy. While $\\text{P}^3\\text{EFT}$ obfuscates gradients and activations, it is unclear why parameters do not require similar privacy measures in the context of this paper.\n\n2) In line 266, it appears that the expression $\\text{backprop}(x, \\theta, g_h)$ should be written as $\\frac{1}{2} \\Bigl( \\text{backprop}(x, \\theta, g_h + z) + \\text{backprop}(x, \\theta, g_h - z) \\Bigr)$ instead of $\\Bigl( \\text{backprop}(x, \\theta, g_h + z) + \\text{backprop}(x, \\theta, g_h - z) \\Bigr)$. Could you confirm if this is correct?\n\n3) Regarding Algorithm 1, did you study the effects of different noise levels in the private backpropagation algorithm? Does this imply that $\\text{P}^3\\text{EFT}$ would yield the same utility across various noise levels?\n\n4) In line 479, it is mentioned that $\\text{P}^3\\text{EFT}$ achieves the same accuracy at a given privacy level. However, it is unclear what “same level of privacy” specifically means, given that the privacy metrics in Tables 1 and 2 differ for DC and $\\text{P}^3\\text{EFT}$.\n\n5) Finally, in line 503, how is the privacy-accuracy trade-off precisely defined? It would be helpful to clarify this term—if it refers to an average measure of privacy and utility, please specify, as the term “trade-off” here is ambiguous without further explanation.\n\n6) Here are some minor spelling errors:\nline 326: it's \nline 478: $\\text{P}^3\\text{FT}$\nline 484: both both our algorithm", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730498432556}, {"id": "g8wxfrmgLI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8052/Reviewer_UNtp"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This work analyzes privacy-preserving fine-tuning of LLMs in the context of parameter-efficient fine-tuning and the two-party split learning setting. The authors show empirically that ordinary fine-tuning reveals sensitive label and successfully alleviates this privacy risk by obfuscating the back propagated gradient using forward and backward API. Experiments on 3 datasets are done.", "review_text": "This work analyzes privacy-preserving fine-tuning of LLMs in the context of parameter-efficient fine-tuning and the two-party split learning setting. The authors show empirically that ordinary fine-tuning reveals sensitive label and successfully alleviates this privacy risk by obfuscating the back propagated gradient using forward and backward API. Experiments on 3 datasets are done.", "strengths": "This paper focuses on a very practical scenario and attempts to address an importance and realistic privacy issue. Also, I like the visualization of the of top-2 principal components of gradients and activations from different fine-tuning steps before and after the defense is applied which directly demonstrated the effectiveness of the proposed method.", "weaknesses": "1. The writing of the paper needs to be improved. For example, the last sentence in the abstract is hard to understand. \"We find that P3EFT is competitive with existing privacy-preserving methods in multi-party and two-party setups while having higher accuracy.\" What is competitive when higher accuracy is achieved.\n2. Experimental datasets are inadequate. Only classification datasets are used although generation task is also mentioned in the paper.\n3. Aside from label privacy, I believe sample input privacy is also important. However, this work exposed the input samples completely to the LLM. I expect that the authors chould provide a proper real-world setting in which only label information but not input features are sensitive and should be protected.", "questions": "1. As I know, some closed-source model training API requires uploading the whole training dataset to the server. Is the proposed method able to preserve the privacy of label under this setting? If not, what might can be done to address this case?", "flag_for_ethics_review": ["No ethics review needed.", "Yes, Discrimination / bias / fairness concerns"], "all_content": {"summary": "This work analyzes privacy-preserving fine-tuning of LLMs in the context of parameter-efficient fine-tuning and the two-party split learning setting. The authors show empirically that ordinary fine-tuning reveals sensitive label and successfully alleviates this privacy risk by obfuscating the back propagated gradient using forward and backward API. Experiments on 3 datasets are done.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "This paper focuses on a very practical scenario and attempts to address an importance and realistic privacy issue. Also, I like the visualization of the of top-2 principal components of gradients and activations from different fine-tuning steps before and after the defense is applied which directly demonstrated the effectiveness of the proposed method.", "weaknesses": "1. The writing of the paper needs to be improved. For example, the last sentence in the abstract is hard to understand. \"We find that P3EFT is competitive with existing privacy-preserving methods in multi-party and two-party setups while having higher accuracy.\" What is competitive when higher accuracy is achieved.\n2. Experimental datasets are inadequate. Only classification datasets are used although generation task is also mentioned in the paper.\n3. Aside from label privacy, I believe sample input privacy is also important. However, this work exposed the input samples completely to the LLM. I expect that the authors chould provide a proper real-world setting in which only label information but not input features are sensitive and should be protected.", "questions": "1. As I know, some closed-source model training API requires uploading the whole training dataset to the server. Is the proposed method able to preserve the privacy of label under this setting? If not, what might can be done to address this case?", "flag_for_ethics_review": ["No ethics review needed.", "Yes, Discrimination / bias / fairness concerns"], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730180596199}], "openreview_url": "https://openreview.net/forum?id=DRKkO2Tejc", "arxiv_id": "2412.16669", "paper_pdf": "papers/DRKkO2Tejc.pdf", "paper_pdf_sha256": "18e48ac9b8651fbf00287200e4f3cc4419eadd22eacb4901e9b4a08145f0b4c2", "paper_pdf_bytes": 740933, "paper_pdf_source": "openreview", "code_url": "https://github.com/p3eft-iclr-2025/p3eft-iclr-2025", "code_repository": "p3eft-iclr-2025/p3eft-iclr-2025", "code_commit": "419418877dfeb07bc2dd2bd4da609c5ef4d7c207", "code_archive": "repos/DRKkO2Tejc.zip", "code_archive_sha256": "7e66ca80a35f04f11702fc20c8c000122d7ff147d09cf4b4ace6c78994003404", "code_archive_bytes": 36871, "code_file_count": 15, "code_extensions": {".sh": 8, ".py": 5, ".ipynb": 2}, "github_disk_usage_kb": 28, "github_languages": {"Python": 72124, "Jupyter Notebook": 28708, "Shell": 14928}, "github_archived": false, "github_pushed_at": "2024-10-08T14:55:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/label-privacy-in-split-learning-for-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WSzRdcOkEx", "year": 2024, "status": "rejected", "title": "GREAT Score: Global Robustness Evaluation of Adversarial Perturbation using Generative Models", "authors": ["ZAITANG LI", "Pin-Yu Chen", "Tsung-Yi Ho"], "authorids": ["~ZAITANG_LI1", "~Pin-Yu_Chen1", "~Tsung-Yi_Ho2"], "authors_source": "OpenReview API", "abstract": "Current studies on adversarial robustness mainly focus on aggregating \\textit{local} robustness results from a set of data samples to evaluate and rank different models. However, the local statistics may not well represent the true \\textit{global} robustness of the underlying unknown data distribution. To address this challenge, this paper makes the first attempt to present a new framework, called \\textit{GREAT Score}, for global robustness evaluation of adversarial perturbation using generative models. Formally, GREAT Score carries the physical meaning of  a global statistic capturing a mean certified attack-proof perturbation level over all samples drawn from a generative model. For finite-sample evaluation, we also derive a probabilistic guarantee on the sample complexity and the difference between the sample mean and the true mean. GREAT Score has several advantages: (1) Robustness evaluations using GREAT Score are efficient and scalable to large models, by sparing the need of running adversarial attacks. In particular, we show high correlation and significantly reduced computation cost of GREAT Score when compared to the attack-based model ranking on RobustBench \\cite{croce2021robustbench}.\n(2) The use of generative models facilitates the approximation of the unknown data distribution. In our ablation study with different generative adversarial networks (GANs), we observe consistency between global robustness evaluation and the quality of GANs. (3) GREAT Score can be used for remote auditing of privacy-sensitive black-box models, as demonstrated by our robustness evaluation on several online facial recognition services.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "3r3jTOkWQv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3797/Reviewer_kzkL"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper is focused on an important problem which is \"global robustness\" that aims to evaluate the robustness of classifiers with respect to the underlying, unknown data distribution. This work can be classified in the family of robustness evaluation algorithms. The authors propose GREAT Score for evaluating the global robustness of classifiers by leveraging generative models. Specifically, the GREAT Score utilizes a generative model to approximate the data distribution and then calculates a certified lower bound on the minimal adversarial perturbation level, averaged over the sampling distribution of the generative model. This allows estimating the distribution-wide robustness without needing the true data distribution.", "review_text": "The paper is focused on an important problem which is \"global robustness\" that aims to evaluate the robustness of classifiers with respect to the underlying, unknown data distribution. This work can be classified in the family of robustness evaluation algorithms. The authors propose GREAT Score for evaluating the global robustness of classifiers by leveraging generative models. Specifically, the GREAT Score utilizes a generative model to approximate the data distribution and then calculates a certified lower bound on the minimal adversarial perturbation level, averaged over the sampling distribution of the generative model. This allows estimating the distribution-wide robustness without needing the true data distribution.", "strengths": "+ This problem setting is important. The overwhelming majority of work in the space of robustness evaluation has considered aggregated local robustness statistics over test samples. However, one might argue that these test samples might be biased and not able to represent the true data distribution. The global robustness of models w.r.t. the entire data space is still under exploring. In this sense, the problem the authors seek to address is relevant and will likely continue to grow in relevance.\n+ The paper is very well-written and organised. The main insights are well explained, and it is easy to follow. The use of generative models is well-motivated.", "weaknesses": "- The authors utilise the generative models to estimate models' global robustness and try to provide statistical guarantees which is claimed as one of the main theorems. My biggest concern is from this point. Since generative models are not perfect, there exists a notable gap between the generative and true data distribution. We are still worrying about the error bound induced by the difference between the underlying data distribution and generative data distribution.  The paper would greatly benefit from including probabilistic guarantees or error bounds on the estimated global robustness concerning the true data distribution, as opposed to solely with respect to the generative data distribution. Such bounds would significantly enhance the appeal and credibility of the results.\n- Table 1 illustrates the comparison between (Calibrated) GREAT Score v.s. minimal distortion found by CW attack on CIFAR-10. However, the table does not explicitly show the advantages of GREAT score compared to other metrics. For example, the CW distortion score between Rebuffi_extra and Gowal_extra is pretty large, but the GREAT score only differs 0.003. \n- Is there any other options besides generative models or GANs, which can approximate the unknown data distribution?\n- Figure 1 and 2 should be larger. Similarly, it is hard to get a close look at Table 1 and 2. The figures and tables on Page 7 and 8 are not well-constructed.", "questions": "Pls see the Section Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper is focused on an important problem which is \"global robustness\" that aims to evaluate the robustness of classifiers with respect to the underlying, unknown data distribution. This work can be classified in the family of robustness evaluation algorithms. The authors propose GREAT Score for evaluating the global robustness of classifiers by leveraging generative models. Specifically, the GREAT Score utilizes a generative model to approximate the data distribution and then calculates a certified lower bound on the minimal adversarial perturbation level, averaged over the sampling distribution of the generative model. This allows estimating the distribution-wide robustness without needing the true data distribution.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "+ This problem setting is important. The overwhelming majority of work in the space of robustness evaluation has considered aggregated local robustness statistics over test samples. However, one might argue that these test samples might be biased and not able to represent the true data distribution. The global robustness of models w.r.t. the entire data space is still under exploring. In this sense, the problem the authors seek to address is relevant and will likely continue to grow in relevance.\n+ The paper is very well-written and organised. The main insights are well explained, and it is easy to follow. The use of generative models is well-motivated.", "weaknesses": "- The authors utilise the generative models to estimate models' global robustness and try to provide statistical guarantees which is claimed as one of the main theorems. My biggest concern is from this point. Since generative models are not perfect, there exists a notable gap between the generative and true data distribution. We are still worrying about the error bound induced by the difference between the underlying data distribution and generative data distribution.  The paper would greatly benefit from including probabilistic guarantees or error bounds on the estimated global robustness concerning the true data distribution, as opposed to solely with respect to the generative data distribution. Such bounds would significantly enhance the appeal and credibility of the results.\n- Table 1 illustrates the comparison between (Calibrated) GREAT Score v.s. minimal distortion found by CW attack on CIFAR-10. However, the table does not explicitly show the advantages of GREAT score compared to other metrics. For example, the CW distortion score between Rebuffi_extra and Gowal_extra is pretty large, but the GREAT score only differs 0.003. \n- Is there any other options besides generative models or GANs, which can approximate the unknown data distribution?\n- Figure 1 and 2 should be larger. Similarly, it is hard to get a close look at Table 1 and 2. The figures and tables on Page 7 and 8 are not well-constructed.", "questions": "Pls see the Section Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699080620680}, {"id": "lgqQsnjDKp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3797/Reviewer_UUBS"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose GREAT score, a privacy measure that may be used to quickly evaluate the robustness of black-box models.  Towards this end, GREAT score utilizes generative models (e.g., GANs and DDPMs) to generate (potentially label-conditioned) samples which are fed into the black-box model of interest.  In this manner, the black-models outputs (over the input classes) are gathered and used in the calculation of GREAT score.  The GREAT score itself is interesting; it leverages the overlapping input-noise characteristics utilized by both GANs and DDPMs (i.e., zero-mean isotropic Gaussian noise) to calculate a certified lower-bound on the true global robustness (wrt to the underlying generative model used).  GREAT score may thus be used either as a standalone auditing strategy for black-box models, or in conjunction with computationally intensive benchmarks which directly test robustness via adversarial attacks (e.g., RobustBench or specific attacks, such as AutoAttack, FGSM, PGD, etc.).  One limitation of the presented work is GREAT score's theoretical guarantees only apply to L2-based attacks, the effects of which on other widely-used distance metrics (L0 and L_{\\inf}) requires understanding through further follow-up work.", "review_text": "The authors propose GREAT score, a privacy measure that may be used to quickly evaluate the robustness of black-box models.  Towards this end, GREAT score utilizes generative models (e.g., GANs and DDPMs) to generate (potentially label-conditioned) samples which are fed into the black-box model of interest.  In this manner, the black-models outputs (over the input classes) are gathered and used in the calculation of GREAT score.  The GREAT score itself is interesting; it leverages the overlapping input-noise characteristics utilized by both GANs and DDPMs (i.e., zero-mean isotropic Gaussian noise) to calculate a certified lower-bound on the true global robustness (wrt to the underlying generative model used).  GREAT score may thus be used either as a standalone auditing strategy for black-box models, or in conjunction with computationally intensive benchmarks which directly test robustness via adversarial attacks (e.g., RobustBench or specific attacks, such as AutoAttack, FGSM, PGD, etc.).  One limitation of the presented work is GREAT score's theoretical guarantees only apply to L2-based attacks, the effects of which on other widely-used distance metrics (L0 and L_{\\inf}) requires understanding through further follow-up work.", "strengths": "The overall writing and derivation of the GREAT score itself are easy to follow and intuitive.  Furthermore, the authors do a good job of extensively testing GREAT score, empirically contrasting its results to RobustBench, and demonstrating its use across a suite of face recognition APIs.  While the theoretical results are interesting, exhaustive experiments are impressive and convincing of the utility of this score for measuring robustness of black-box models.", "weaknesses": "Given the extensive research concerning robustness wrt L0/L2/L_{\\inf} based adversarial attacks, GREAT score's limitation to L2-based attacks is a major weakness.   As only L2-based attacks are considered in the paper, either GREAT score must be extended to cover L0/L_{\\inf} based attacks (which, the authors note, is unlikely under the current derivation) or follow up work must explore how effective GREAT score is at measuring robustness for L0 and L_{\\inf} based attacks.\n\nThe papers description of some relevant work is either terse or lacking.  For instance:\n- \"using the Rebuffi_extra model (Rebuffi et al., 2021)\" <- please give a short inline description of this model\n\n- \"successful adversarial perturbations (an upper bound on the minimal perturbation of each sample) returned by any norm-minimization adversarial attack method such as the CW attack (Carlini & Wagner, 2017)\" <- Please give a brief description of the CW attack\n\n- In Table 1, please add a column showing which generative model is being evaluated.\n\nFurthermore, GREAT score's role as a lower bound for CW distortion is noted throughout the paper.  However, this is not entirely clear given the current version of the paper.  Please explain that the CW (L2) attack solves an optimization objective which results in low L2 distortion.  Also, it is important to note why GREAT score serves as a lower bound for the CW attack distortion; Equation 3 is equivalent to the non-L2 term of the CW (L2) attack, with c = sqrt{ \\pi / 2} (in practice, CW determines c via a grid search).  Thus, assuming c >= sqrt{\\pi /2), GREAT score trivially lower bounds the CW L2 attack.", "questions": "- \"Moreover, as a byproduct of using generative models, our adversarial robustness evaluation procedure is executed with only synthetically generated data instead of real data, which is particularly appealing to privacy-aware robustness assessment schemes, e.g., remote robustness evaluation or auditing by a third party with restricted access to data and model.\" <- Please note that this is not a panacea, i.e., see:  \n-Carlini, Nicolas, et al. \"Extracting training data from diffusion models.\" 32nd USENIX Security Symposium (USENIX Security 23). 2023.\n-Duan, Jinhao, et al. \"Are diffusion models vulnerable to membership inference attacks?.\" arXiv preprint arXiv:2302.01316 (2023).\n\n- \"In Figure 2, we compare the cumulative robust accuracy (RA)\" <- RA is used before its definition on page 7; please define on the first use of acronym RA\n\n- \"using the Rebuffi_extra model (Rebuffi et al., 2021)\" <- please give a short inline description of this model\n\n- \"DMs reverse the forward process and implement a sampling process from Gaussian noises to reconstruct the true\nsamples\" <- Please mention that this is done by solving a stochastic differential equation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose GREAT score, a privacy measure that may be used to quickly evaluate the robustness of black-box models.  Towards this end, GREAT score utilizes generative models (e.g., GANs and DDPMs) to generate (potentially label-conditioned) samples which are fed into the black-box model of interest.  In this manner, the black-models outputs (over the input classes) are gathered and used in the calculation of GREAT score.  The GREAT score itself is interesting; it leverages the overlapping input-noise characteristics utilized by both GANs and DDPMs (i.e., zero-mean isotropic Gaussian noise) to calculate a certified lower-bound on the true global robustness (wrt to the underlying generative model used).  GREAT score may thus be used either as a standalone auditing strategy for black-box models, or in conjunction with computationally intensive benchmarks which directly test robustness via adversarial attacks (e.g., RobustBench or specific attacks, such as AutoAttack, FGSM, PGD, etc.).  One limitation of the presented work is GREAT score's theoretical guarantees only apply to L2-based attacks, the effects of which on other widely-used distance metrics (L0 and L_{\\inf}) requires understanding through further follow-up work.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The overall writing and derivation of the GREAT score itself are easy to follow and intuitive.  Furthermore, the authors do a good job of extensively testing GREAT score, empirically contrasting its results to RobustBench, and demonstrating its use across a suite of face recognition APIs.  While the theoretical results are interesting, exhaustive experiments are impressive and convincing of the utility of this score for measuring robustness of black-box models.", "weaknesses": "Given the extensive research concerning robustness wrt L0/L2/L_{\\inf} based adversarial attacks, GREAT score's limitation to L2-based attacks is a major weakness.   As only L2-based attacks are considered in the paper, either GREAT score must be extended to cover L0/L_{\\inf} based attacks (which, the authors note, is unlikely under the current derivation) or follow up work must explore how effective GREAT score is at measuring robustness for L0 and L_{\\inf} based attacks.\n\nThe papers description of some relevant work is either terse or lacking.  For instance:\n- \"using the Rebuffi_extra model (Rebuffi et al., 2021)\" <- please give a short inline description of this model\n\n- \"successful adversarial perturbations (an upper bound on the minimal perturbation of each sample) returned by any norm-minimization adversarial attack method such as the CW attack (Carlini & Wagner, 2017)\" <- Please give a brief description of the CW attack\n\n- In Table 1, please add a column showing which generative model is being evaluated.\n\nFurthermore, GREAT score's role as a lower bound for CW distortion is noted throughout the paper.  However, this is not entirely clear given the current version of the paper.  Please explain that the CW (L2) attack solves an optimization objective which results in low L2 distortion.  Also, it is important to note why GREAT score serves as a lower bound for the CW attack distortion; Equation 3 is equivalent to the non-L2 term of the CW (L2) attack, with c = sqrt{ \\pi / 2} (in practice, CW determines c via a grid search).  Thus, assuming c >= sqrt{\\pi /2), GREAT score trivially lower bounds the CW L2 attack.", "questions": "- \"Moreover, as a byproduct of using generative models, our adversarial robustness evaluation procedure is executed with only synthetically generated data instead of real data, which is particularly appealing to privacy-aware robustness assessment schemes, e.g., remote robustness evaluation or auditing by a third party with restricted access to data and model.\" <- Please note that this is not a panacea, i.e., see:  \n-Carlini, Nicolas, et al. \"Extracting training data from diffusion models.\" 32nd USENIX Security Symposium (USENIX Security 23). 2023.\n-Duan, Jinhao, et al. \"Are diffusion models vulnerable to membership inference attacks?.\" arXiv preprint arXiv:2302.01316 (2023).\n\n- \"In Figure 2, we compare the cumulative robust accuracy (RA)\" <- RA is used before its definition on page 7; please define on the first use of acronym RA\n\n- \"using the Rebuffi_extra model (Rebuffi et al., 2021)\" <- please give a short inline description of this model\n\n- \"DMs reverse the forward process and implement a sampling process from Gaussian noises to reconstruct the true\nsamples\" <- Please mention that this is done by solving a stochastic differential equation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698822214776}, {"id": "tfw36rloRR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3797/Reviewer_npik"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "To solve the problems of lack of proper global adversarial robustness evaluation, limitation to white-box settings and computational inefficiency, this paper proposes GREAT Score for global robustness evaluation of adversarial perturbation using generative models. The authors have introduced and compared the proposed method, but the novelty of this paper is limited. Here are some of my suggestions for improving the quality of this paper.", "review_text": "To solve the problems of lack of proper global adversarial robustness evaluation, limitation to white-box settings and computational inefficiency, this paper proposes GREAT Score for global robustness evaluation of adversarial perturbation using generative models. The authors have introduced and compared the proposed method, but the novelty of this paper is limited. Here are some of my suggestions for improving the quality of this paper.", "strengths": "The paper is well writen, with better typesetting, the research content is also of certain practical value.", "weaknesses": "(1) There are obvious spelling and coincidence errors in the paper.\n(2) In the background and related works section, it is suggested to simplify the introduction of generative models and add the description of global robustness evaluation rather than Appendix 6.3.\n(3)  It is recommended to draw flow charts of the algorithm proposed by the author and other algorithms to highlight the innovation of the algorithm.\n(4)  It is recommended to give a more detailed time complexity calculation.\n(5)  In the experimental results section, the layout of the experimental results graphs and tables is confusing, and it is recommended to rearrange them.\n(6)  Appendix chapters need to be arranged according to the order in which they appear in the main text.\n(7)  Some of the examples in the article need to be replaced to adhere to the previous arguments.", "questions": "(1)  For other competitive methods, it would be better to use all the data in the dataset to measure robustness rather than the strategy proposed by the author?\n(2)  Does it still make sense to use a lower bound if a true global robustness evaluation can be obtained with high computational efficiency ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To solve the problems of lack of proper global adversarial robustness evaluation, limitation to white-box settings and computational inefficiency, this paper proposes GREAT Score for global robustness evaluation of adversarial perturbation using generative models. The authors have introduced and compared the proposed method, but the novelty of this paper is limited. Here are some of my suggestions for improving the quality of this paper.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well writen, with better typesetting, the research content is also of certain practical value.", "weaknesses": "(1) There are obvious spelling and coincidence errors in the paper.\n(2) In the background and related works section, it is suggested to simplify the introduction of generative models and add the description of global robustness evaluation rather than Appendix 6.3.\n(3)  It is recommended to draw flow charts of the algorithm proposed by the author and other algorithms to highlight the innovation of the algorithm.\n(4)  It is recommended to give a more detailed time complexity calculation.\n(5)  In the experimental results section, the layout of the experimental results graphs and tables is confusing, and it is recommended to rearrange them.\n(6)  Appendix chapters need to be arranged according to the order in which they appear in the main text.\n(7)  Some of the examples in the article need to be replaced to adhere to the previous arguments.", "questions": "(1)  For other competitive methods, it would be better to use all the data in the dataset to measure robustness rather than the strategy proposed by the author?\n(2)  Does it still make sense to use a lower bound if a true global robustness evaluation can be obtained with high computational efficiency ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698805139136}, {"id": "kmJ138abEY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3797/Reviewer_fKuC"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a data-independent metric called GREAT Score for evaluating adversarial robustness. To be specific, GREAT Score is calculated as the mean of the gaps in the likelihood of model prediction between the ground-truth class and the most likely class other than the ground-truth class achieved by the samples generated from a generative model. The empirical results seem to validate the effectiveness of GREAT Score in evaluating robustness.", "review_text": "This paper proposes a data-independent metric called GREAT Score for evaluating adversarial robustness. To be specific, GREAT Score is calculated as the mean of the gaps in the likelihood of model prediction between the ground-truth class and the most likely class other than the ground-truth class achieved by the samples generated from a generative model. The empirical results seem to validate the effectiveness of GREAT Score in evaluating robustness.", "strengths": "1. The GREAT Score is data-independent, which is applicable to privacy-sensitive black-box evaluation. \n\n2. The GREAT Score is less sensitive to data points compared to AutoAttack.", "weaknesses": "1. The definition of robustness in Eq. (1) and Eq.(4) seems to be confusing and possibly wrong. In Eq. (1), $\\Delta_{\\min}$ is defined as the minimal perturbation of a sample-label pair causing the change of the top-1 class prediction. Then, I understand $g(x)$ is a lower bound of $\\Delta_{\\min}$. However, in Eq. (3), $g(x)$ is defined as the gap between two probabilities. Therefore, I am confused about how the gap between two probabilities measures the minimal perturbation.\n\n2. The rank of the model in terms of adversarial robustness is unclear and confusing. As shown in Table 1, the rank of a model based on GREAT Score could be different from the rank based on AutoAttack.  Besides, the rank of a model based on GREAT Score could also be different from the rank based on the calibrated GREAT Score.  Therefore, the research would be confused about what is the genuine rank of a model in terms of its robustness.\n\n3. How can you guarantee the gap between the two probabilities achieved by the generated data is achieved in the worst case? As far as I can see, the authors directly used the generated samples from the generative models, which are supposed to be natural samples instead of adversarial samples.  \n\n4. The table and figure should be resized to be larger, which makes it easier to read.", "questions": "Please refer to my comments in “Weaknesses”.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a data-independent metric called GREAT Score for evaluating adversarial robustness. To be specific, GREAT Score is calculated as the mean of the gaps in the likelihood of model prediction between the ground-truth class and the most likely class other than the ground-truth class achieved by the samples generated from a generative model. The empirical results seem to validate the effectiveness of GREAT Score in evaluating robustness.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The GREAT Score is data-independent, which is applicable to privacy-sensitive black-box evaluation. \n\n2. The GREAT Score is less sensitive to data points compared to AutoAttack.", "weaknesses": "1. The definition of robustness in Eq. (1) and Eq.(4) seems to be confusing and possibly wrong. In Eq. (1), $\\Delta_{\\min}$ is defined as the minimal perturbation of a sample-label pair causing the change of the top-1 class prediction. Then, I understand $g(x)$ is a lower bound of $\\Delta_{\\min}$. However, in Eq. (3), $g(x)$ is defined as the gap between two probabilities. Therefore, I am confused about how the gap between two probabilities measures the minimal perturbation.\n\n2. The rank of the model in terms of adversarial robustness is unclear and confusing. As shown in Table 1, the rank of a model based on GREAT Score could be different from the rank based on AutoAttack.  Besides, the rank of a model based on GREAT Score could also be different from the rank based on the calibrated GREAT Score.  Therefore, the research would be confused about what is the genuine rank of a model in terms of its robustness.\n\n3. How can you guarantee the gap between the two probabilities achieved by the generated data is achieved in the worst case? As far as I can see, the authors directly used the generated samples from the generative models, which are supposed to be natural samples instead of adversarial samples.  \n\n4. The table and figure should be resized to be larger, which makes it easier to read.", "questions": "Please refer to my comments in “Weaknesses”.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698399063571}], "openreview_url": "https://openreview.net/forum?id=WSzRdcOkEx", "arxiv_id": "2304.09875", "paper_pdf": "papers/WSzRdcOkEx.pdf", "paper_pdf_sha256": "315c7b539a7d6c3f04d2d0947a586563550c787d8293353c984b7a9f7ecc0800", "paper_pdf_bytes": 7561210, "paper_pdf_source": "openreview", "code_url": "https://github.com/IBM/GREAT-Score", "code_repository": "IBM/GREAT-Score", "code_commit": "beadb769ac59f434b42e6a40e08ba87c758444ff", "code_archive": "repos/WSzRdcOkEx.zip", "code_archive_sha256": "511178ac754305ca0a9f7e07b532de0930a234cca339e9f6baf468daf48f9b1c", "code_archive_bytes": 28372, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 26, "github_languages": {"Python": 63514}, "github_archived": false, "github_pushed_at": "2025-09-18T00:02:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/great-score-global-robustness-evaluation-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nG08xiRT2As", "year": 2023, "status": "rejected", "title": "InfoOT: Information Maximizing Optimal Transport", "authors": ["Ching-Yao Chuang", "Stefanie Jegelka", "David Alvarez-Melis"], "authorids": ["~Ching-Yao_Chuang1", "~Stefanie_Jegelka3", "~David_Alvarez-Melis1"], "authors_source": "OpenReview API", "abstract": "Optimal transport aligns samples across distributions by minimizing the transportation cost between them, e.g., the geometric distances. Yet, it ignores coherence structure in the data such as clusters, does not handle outliers well, and cannot integrate new data points. To address these drawbacks, we propose InfoOT, an information-theoretic extension of optimal transport that maximizes the mutual information between domains while minimizing geometric distances. The resulting objective can still be formulated as a (generalized) optimal transport problem, and can be efficiently solved by projected gradient descent. This formulation yields a new projection method that is robust to outliers and generalizes to unseen samples. Empirically, InfoOT improves the quality of alignments across benchmarks in domain adaptation, cross-domain retrieval, and single-cell alignment.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "OxqVDmZ5iv", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1969/Reviewer_BdbM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose InfoOT which maximizes the mutual information between domains and minimize distances between input distributions. The proposed InfoOT address several drawbacks of OT, e.g., coherence structure (clustering, outliers) and easy to integrate new data points. Empirically, the authors evaluate the proposed method on domain adaption, cross-domain retrieval and single-cell alignment.", "review_text": "The proposed method is interesting. The proposed methods address several drawbacks of OT.", "strengths": "Strength\n+ The proposed InfoOT address several drawbacks of OT \n+ The proposed method works well in applications.\n\nWeaknesses\n+ The advantage of the proposed method (using kernel density estimation for continuous setting) is not clear enough over discrete ones (e.g., entropic regularization in Sinkhorn), especially in the case input distributions are discrete (empirical distributions). It will be a plus if the authors elaborate this points with more details.\n+ It is unclear how many samples are required for the kernel density estimation used in InfoOT", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose InfoOT which maximizes the mutual information between domains and minimize distances between input distributions. The proposed InfoOT address several drawbacks of OT, e.g., coherence structure (clustering, outliers) and easy to integrate new data points. Empirically, the authors evaluate the proposed method on domain adaption, cross-domain retrieval and single-cell alignment.", "strength_and_weaknesses": "Strength\n+ The proposed InfoOT address several drawbacks of OT \n+ The proposed method works well in applications.\n\nWeaknesses\n+ The advantage of the proposed method (using kernel density estimation for continuous setting) is not clear enough over discrete ones (e.g., entropic regularization in Sinkhorn), especially in the case input distributions are discrete (empirical distributions). It will be a plus if the authors elaborate this points with more details.\n+ It is unclear how many samples are required for the kernel density estimation used in InfoOT", "clarity,_quality,_novelty_and_reproducibility": "The proposed ideas are interesting. The proposed method, InfoOT can address several drawbacks from OT, especially about coherence structure (cluster, outlier) and the ability to integrate new data points.\n\nI have some following concerns:\n+ In case the input distributions are discrete, e.g., empirical distributions. It is not clear the advantages of using kernel density estimation (as in the proposed method) comparing to the entropic regularization (in Sinkhorn) for measuring global structure with mutual information (as in Section 4.1). Could the authors elaborate it with more details?\n\n+ How many samples are required for the kernel density estimation (KDE)? and how to choose the bandwidth for the Gaussians used in KDE for the proposed InfoOT? especially for high-dimensional setting?\n\n+ Could the author discuss the relation between the proposed InfoOT with Liu'2021 which is also based on mutual information and OT from given unpaired data?\n\n+ For the robustness against noisy data, it is better if the authors compare the proposed method with the unbalanced OT approach (which also use to address this problem for OT). Could the authors discuss about it (and better to have some empirical comparison)?\n\n+ In experiments, it is well-known that the entropic regularization affects performances of entropic OT, why the authors set it to 1 in experiments?\n--- How the \\lambda in Fused InfoOT affects its performances in applications? Why the authors set it to 100? Should one need to use \\lambda to control the effect of the regularization?", "summary_of_the_review": "The proposed method is interesting. The proposed methods address several drawbacks of OT.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666776724459}, {"id": "VH1Y_faYSas", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1969/Reviewer_pBFH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed to improve the quality of optimal transport via simultaneously encouraging the mutual information (MI). This is justified by establishing the equivalence between the MI and the standard entropy regularization used in the Sinkhorn algorithm. Empirical results showed that when implemented with kernel density estimation, the proposed InfoOT better captures the cluster structure in data and works better on serval domain transfer tasks. An alternative view on the smoothness regularization is also given.  The author(s) further presented a conditional projection scheme for the OT mapping. ", "review_text": "This paper is well-written. My major reservation is due to the lack of technical novelty and practical significance. KDE is a well-establish technique and it does not scale well, also for the problems considered in the experiments there should be better alternative solutions without applying OT (especially for the domain adaptation). I am open to reconsider my decision should more convincing argument or empirical results surface during the rebuttal phase. ", "strengths": "\n### Strength\n* Calling out the equivalence between entropy regularization and the proposed MI optimization makes an interesting point. \n* The smoothness argument to preserve cluster structure makes good sense. \n* Reported empirical gains on domain adaption and cross-domain retrieval are very encouraging\n\n### Weakness\n* As a criticism applies all kernel methods, the main concern is scalability of the method. First, the computation scales quadratically, and to ensure good performance, heavy parameter tuning is typically needed. Integration and end-to-end optimization with need neural nets (which is known to perform strongly even without the kernel trick), is often impractical with kernel-based components. \n* There are existing work on the neural estimation of the transport plan and non-parametric estimation of MI (e.g., InfoNCE, NWJ, etc.). While integrating these two directions as a generalized version of the proposed solution seems to be out-of-scope, I would love to see some technical discussions at least. \n* While I appreciate the the cluster-preserving idea, this is merely a flaky heuristic assumption and may not hold universally. Also some ablation study using alternative, more direct penalties to enforce smooth transport should be considered (e.g., the distance between nearby source points in the transported target space). \n* For the domain adaption experiment, while I understand the author(s) followed the experiment setup from prior works, but 1-NN classification is usually considered unreliable. Numbers with alternative classification schemes should be reported (e.g., 5-NN, linear, simple neural net, etc.)\n* There is no discussion on the limitations of the proposed method\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed to improve the quality of optimal transport via simultaneously encouraging the mutual information (MI). This is justified by establishing the equivalence between the MI and the standard entropy regularization used in the Sinkhorn algorithm. Empirical results showed that when implemented with kernel density estimation, the proposed InfoOT better captures the cluster structure in data and works better on serval domain transfer tasks. An alternative view on the smoothness regularization is also given.  The author(s) further presented a conditional projection scheme for the OT mapping. ", "strength_and_weaknesses": "\n### Strength\n* Calling out the equivalence between entropy regularization and the proposed MI optimization makes an interesting point. \n* The smoothness argument to preserve cluster structure makes good sense. \n* Reported empirical gains on domain adaption and cross-domain retrieval are very encouraging\n\n### Weakness\n* As a criticism applies all kernel methods, the main concern is scalability of the method. First, the computation scales quadratically, and to ensure good performance, heavy parameter tuning is typically needed. Integration and end-to-end optimization with need neural nets (which is known to perform strongly even without the kernel trick), is often impractical with kernel-based components. \n* There are existing work on the neural estimation of the transport plan and non-parametric estimation of MI (e.g., InfoNCE, NWJ, etc.). While integrating these two directions as a generalized version of the proposed solution seems to be out-of-scope, I would love to see some technical discussions at least. \n* While I appreciate the the cluster-preserving idea, this is merely a flaky heuristic assumption and may not hold universally. Also some ablation study using alternative, more direct penalties to enforce smooth transport should be considered (e.g., the distance between nearby source points in the transported target space). \n* For the domain adaption experiment, while I understand the author(s) followed the experiment setup from prior works, but 1-NN classification is usually considered unreliable. Numbers with alternative classification schemes should be reported (e.g., 5-NN, linear, simple neural net, etc.)\n* There is no discussion on the limitations of the proposed method\n", "clarity,_quality,_novelty_and_reproducibility": "### Clarity. Very good.\nI really appreciate the clarity of the presentation. \n\n### Quality. Okay\nThis paper is well-written. Optimal transport is a versatile tool and there is growing interests in applying OT-based solutions to address various empirical problems. This work has exploited a simple smoothness heuristic to address the cluster preserving issue common to many OT applications. To make the points more intuitive the author(s) have presented visualization with both synthetic and real datasets. \n\nRelevant literature is adequately cited. It will be nice to further complete the picture with refs on the more generic integral probability metric (IPM) and generalizations of OT such as Sinkhorn divergence. Also potentially compared with the more stable OT algorithms such as Inexact Proximal point method for Optimal Transport (IPOT), my hunch is that vanilla Sinkhorn is known to be unstable, and that might be the cause of issues discussed in the paper. Using more stable OT algorithms may address these pain points without any smoothness regularization. \n\n### Novelty. Fair\nKDE is a well-establish technique, and smoothness is a common regularity widely applied in machine learning. \n\n### Reproducibility. Good\nDetails of the solution and experiment setups are very well documented, and there should be no major difficulty involved implementing the model. \n", "summary_of_the_review": "This paper is well-written. My major reservation is due to the lack of technical novelty and practical significance. KDE is a well-establish technique and it does not scale well, also for the problems considered in the experiments there should be better alternative solutions without applying OT (especially for the domain adaptation). I am open to reconsider my decision should more convincing argument or empirical results surface during the rebuttal phase. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666631656307}, {"id": "6ePjWHeK987", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1969/Reviewer_EHDZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a dynamic cost function for OT (see below on whether this is a valid OT objective or not) that attempts to maximize the mutual information of the projections using kernel density estimates---which depend on the coupling distribution.\nThis can be seen as a generalization of entropic regularization.\nThe InfoOT formulation can be merged with the standard OT problem and can then be solved via a sequence of entropic OT problems.\nEmpirically, this is shown to preserve cluster structure in 2D examples and qualitative t-SNE projections.\nFurthermore, the paper proposes a mapping method based on the kernel density estimates called conditional projection that is more robust to projecting outliers.\n", "review_text": "Overall, I found the paper to address two interesting empirical problems related to using OT in ML: Preserving cluster structure and outliers. The technical descriptions were a bit vague but the overall idea of adding a mutual information regularizer seems interesting. However, I do have some concerns about the use of KDEs because they can break down in high dimensions, the experimental setup, and some technical points. I hope that at least some of these can be addressed by the author response.\n", "strengths": "*Strengths*\n- The proposed training method can maintain cluster structure when learning a mapping.\n\n- The proposed mapping method can be more robust to outliers than barycentric proojection.\n\n- The algorithm simplifies to a sequence of entropic OT problems.\n\n- The results show improved performance compared to standard OT baselines across several tasks.\n\n*Weaknesses*\n- Kernel densities are known to breakdown in higher dimensions and suffer from the curse of dimensionality. This presents a key limitation and problem with the proposed approach. Why would this not be a problem here?  While there are experiments on MNIST for domain adaptation, it is unclear what is actually going on. \n\n- The baselines for all external tasks seem to be OT-based. What about non-OT methods for these tasks?  For example, what about GAN-based methods for domain adaptation (e.g., [Ganin et al. 2016]) or Wasserstein dual formulations for learning OT maps via convex functions (e.g., [Makkuva et al., 2020])? There are also flow-based methods for alignment that could be used for domain adaptation [Grover et al. 2020, Usman et al., 2020]. I am unfamiliar with the sota methods for retrieval but it seems there would also be methods here. While I understand that you are comparing to other OT-based methods, the key motivation for this estimator seems to be for empirical reasons.  It is unclear that there is theoretic or \"intrinsic\" motivation for the proposed OT formulation. Thus, to demonstrate it's usefulness more broadly, it would be helpful to compare to sota methods for these tasks and/or do more intrinsic evaluation.\n\n- Definition 1 seems odd or incorrect. It seems that Definition 1 is using two approximations that should be distinguished and explained. First, you approximate the joint distribution over x,y using a kernel density estimate.  If you were to use a plug-in estimator, this would still require an expectation using the kernel density for the expectation.  However, then you replace this expectation over the kernel density, with a psuedo-sample approximation that treats each possible pair of inputs (x,y) as a weighted empirical sample from this kernel density. Thus, the expectation turns into a weighted average with $\\Gamma$ defininig the weights of each pairing.  I believe this is how you simplify to only using $\\Gamma$ as the weights for the outer summation in Definition 1. Again, these two approximations should be discussed and distinguished as it was not clear from the original exposition.\n\n- Using 1-NN in the source domain as the target predictor for domain adaptation and using KNN in the retrieval experiment are unlikely to do well in high dimensions and may be biased towards KDE-based non-parametric approaches. A standard CNN or fully-connected network could be used in evaluating domain adaptation. In particular, a KNN classifier may be unfairly biased towards KDE-like approaches as they are both non-parametric methods based on geometric distances.\n\n- I'm not sure it is reasonable to have $\\Gamma$ in the \"cost\" function of OT at least theoretically. It's unclear if it's actually OT anymore if $\\Gamma$ is in the cost function itself.  Could  you explain why this is still OT if the cost function is now a function of the map itself? \n\nGanin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training of neural networks. The journal of machine learning research, 17(1), 2096-2030.\n\nMakkuva, A., Taghvaei, A., Oh, S., & Lee, J. (2020, November). Optimal transport mapping via input convex neural networks. In International Conference on Machine Learning (pp. 6672-6681). PMLR.\n\nGrover, A., Chute, C., Shu, R., Cao, Z., & Ermon, S. (2020, April). Alignflow: Cycle consistent learning from multiple domains via normalizing flows. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 04, pp. 4028-4035).\n\nUsman, B., Sud, A., Dufour, N., & Saenko, K. (2020). Log-likelihood ratio minimizing flows: Towards robust and quantifiable neural distribution alignment. Advances in Neural Information Processing Systems, 33, 21118-21129.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a dynamic cost function for OT (see below on whether this is a valid OT objective or not) that attempts to maximize the mutual information of the projections using kernel density estimates---which depend on the coupling distribution.\nThis can be seen as a generalization of entropic regularization.\nThe InfoOT formulation can be merged with the standard OT problem and can then be solved via a sequence of entropic OT problems.\nEmpirically, this is shown to preserve cluster structure in 2D examples and qualitative t-SNE projections.\nFurthermore, the paper proposes a mapping method based on the kernel density estimates called conditional projection that is more robust to projecting outliers.\n", "strength_and_weaknesses": "*Strengths*\n- The proposed training method can maintain cluster structure when learning a mapping.\n\n- The proposed mapping method can be more robust to outliers than barycentric proojection.\n\n- The algorithm simplifies to a sequence of entropic OT problems.\n\n- The results show improved performance compared to standard OT baselines across several tasks.\n\n*Weaknesses*\n- Kernel densities are known to breakdown in higher dimensions and suffer from the curse of dimensionality. This presents a key limitation and problem with the proposed approach. Why would this not be a problem here?  While there are experiments on MNIST for domain adaptation, it is unclear what is actually going on. \n\n- The baselines for all external tasks seem to be OT-based. What about non-OT methods for these tasks?  For example, what about GAN-based methods for domain adaptation (e.g., [Ganin et al. 2016]) or Wasserstein dual formulations for learning OT maps via convex functions (e.g., [Makkuva et al., 2020])? There are also flow-based methods for alignment that could be used for domain adaptation [Grover et al. 2020, Usman et al., 2020]. I am unfamiliar with the sota methods for retrieval but it seems there would also be methods here. While I understand that you are comparing to other OT-based methods, the key motivation for this estimator seems to be for empirical reasons.  It is unclear that there is theoretic or \"intrinsic\" motivation for the proposed OT formulation. Thus, to demonstrate it's usefulness more broadly, it would be helpful to compare to sota methods for these tasks and/or do more intrinsic evaluation.\n\n- Definition 1 seems odd or incorrect. It seems that Definition 1 is using two approximations that should be distinguished and explained. First, you approximate the joint distribution over x,y using a kernel density estimate.  If you were to use a plug-in estimator, this would still require an expectation using the kernel density for the expectation.  However, then you replace this expectation over the kernel density, with a psuedo-sample approximation that treats each possible pair of inputs (x,y) as a weighted empirical sample from this kernel density. Thus, the expectation turns into a weighted average with $\\Gamma$ defininig the weights of each pairing.  I believe this is how you simplify to only using $\\Gamma$ as the weights for the outer summation in Definition 1. Again, these two approximations should be discussed and distinguished as it was not clear from the original exposition.\n\n- Using 1-NN in the source domain as the target predictor for domain adaptation and using KNN in the retrieval experiment are unlikely to do well in high dimensions and may be biased towards KDE-based non-parametric approaches. A standard CNN or fully-connected network could be used in evaluating domain adaptation. In particular, a KNN classifier may be unfairly biased towards KDE-like approaches as they are both non-parametric methods based on geometric distances.\n\n- I'm not sure it is reasonable to have $\\Gamma$ in the \"cost\" function of OT at least theoretically. It's unclear if it's actually OT anymore if $\\Gamma$ is in the cost function itself.  Could  you explain why this is still OT if the cost function is now a function of the map itself? \n\nGanin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training of neural networks. The journal of machine learning research, 17(1), 2096-2030.\n\nMakkuva, A., Taghvaei, A., Oh, S., & Lee, J. (2020, November). Optimal transport mapping via input convex neural networks. In International Conference on Machine Learning (pp. 6672-6681). PMLR.\n\nGrover, A., Chute, C., Shu, R., Cao, Z., & Ermon, S. (2020, April). Alignflow: Cycle consistent learning from multiple domains via normalizing flows. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 04, pp. 4028-4035).\n\nUsman, B., Sud, A., Dufour, N., & Saenko, K. (2020). Log-likelihood ratio minimizing flows: Towards robust and quantifiable neural distribution alignment. Advances in Neural Information Processing Systems, 33, 21118-21129.\n", "clarity,_quality,_novelty_and_reproducibility": "- Can this be interpreted as an adaptive/dynamic cost function for OT? This is related to the weakness above. Overall, it seems odd and non-standard but may be an interesting idea that could be generalized. \n\n- Could the KDE weighting function be used as a post-hoc smoothing method for standard OT?  It seems that it could be applied after doing empirical/entropic OT to provide smoother mappings.  If so, could you add it to the baseline methods?  This would help distinguish the mapping method from the OT training method.\n\n- The discussion in Section 4.1 on discrete vs continuous may be premature.  It might be useful to discuss this later in the paper.\n\n- Does equation 1 have a typo, it seems that the second equation should be $d_{\\mathcal{Y}}$.\n\n- Table captions should be on top of tables.\n", "summary_of_the_review": "Overall, I found the paper to address two interesting empirical problems related to using OT in ML: Preserving cluster structure and outliers. The technical descriptions were a bit vague but the overall idea of adding a mutual information regularizer seems interesting. However, I do have some concerns about the use of KDEs because they can break down in high dimensions, the experimental setup, and some technical points. I hope that at least some of these can be addressed by the author response.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666479644283}], "openreview_url": "https://openreview.net/forum?id=nG08xiRT2As", "arxiv_id": "2210.03164", "paper_pdf": "papers/nG08xiRT2As.pdf", "paper_pdf_sha256": "df0584ccfdfb10292719d82b782e1bda65dd907883144e0b376a90c06ab06b8f", "paper_pdf_bytes": 11806626, "paper_pdf_source": "openreview", "code_url": "https://github.com/chingyaoc/InfoOT", "code_repository": "chingyaoc/InfoOT", "code_commit": "352efd202f5b475dc170a8d08a99049689d5ee1a", "code_archive": "repos/nG08xiRT2As.zip", "code_archive_sha256": "0e0a6202890a82aee9e74349efb75e671a42ed5a825943bd6b1d8cc770bcb3f4", "code_archive_bytes": 6228, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 32, "github_languages": {"Python": 11361}, "github_archived": false, "github_pushed_at": "2023-04-27T03:49:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/infoot-information-maximizing-optimal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dYUdt59fJ0e", "year": 2022, "status": "rejected", "title": "Yformer: U-Net Inspired Transformer Architecture for Far Horizon Time Series Forecasting", "authors": ["Kiran Madhusudhanan", "Johannes Burchert", "Nghia Duong-Trung", "Stefan Born", "Lars Schmidt-Thieme"], "authorids": ["~Kiran_Madhusudhanan1", "burchert@ismll.uni-hildesheim.de", "nghia.duong-trung@tu-berlin.de", "born@math.tu-berlin.de", "~Lars_Schmidt-Thieme1"], "authors_source": "OpenReview API", "abstract": "Time series data is ubiquitous in research as well as in a wide variety of industrial applications. Effectively analyzing the available historical data and providing insights into the far future allows us to make effective decisions. Recent research has witnessed the superior performance of transformer-based architectures, especially in the regime of far horizon time series forecasting. However, the current state of the art sparse Transformer architectures fail to couple down- and upsampling procedures to produce outputs in a similar resolution as the input. We propose the Yformer model, based on a novel Y-shaped encoder-decoder architecture that (1) uses direct connection from the downscaled encoder layer to the corresponding upsampled decoder layer in a U-Net inspired architecture, (2) Combines the downscaling/upsampling with sparse attention to capture long-range effects, and (3) stabilizes the encoder-decoder stacks with the addition of an auxiliary reconstruction loss. Extensive experiments have been conducted with relevant baselines on four benchmark datasets, demonstrating an average improvement of 19.82, 18.41 percentage MSE and 13.62, 11.85 percentage MAE in comparison to the current state of the art for the univariate and the multivariate settings respectively.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0c1jBq_qBK-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1880/Reviewer_yQSB"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents Yformer to perform long sequence time series forecasting. The key idea is to employ skip connection to improve the prediction resolution and stabilize the encoder and decoder by reconstructing the recent past. The experiment results on two datasets showed the effectiveness of the proposed method.", "review_text": "Strengths\n* Llong range time series forecasting is an interesting problem to investigate.\n* Adding skip connections between encoder and decoder is technically sound\n* The overall experiment results showed the effectiveness of the proposed method.\n\nWeaknesses\n* The organization of this paper is not well. Many technical details are not very clear in the main context\n* The overall technical novelty is limited\n* The effectiveness of the skip connections are not fully assessed\n* More details of the experiments are not provided.\n\nThe main problem of this paper is that the main context (especially the methodology section) is not self-contained. The reader will have to rely on details in the appendix or other papers to fully understand the proposed technique.\n\nAnother concern is the novelty. Skip connections are common practice in U-net and the idea of stabilizing the encoder and decoder by reconstructing the recent past is also not new. Although it is a new application area for skip connections, the overall technical novelty is limited.\n\nIn addition, the ablation study over whether the Skip connections are used or not is not provided.\n\nSeveral related works are not mentioned or compared.\n\n[1] \"Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting.\" Sen, Rajat, Hsiang-Fu Yu, and Inderjit S. Dhillon NeurIPS 2019.\n\n[2] \"Modeling long-and short-term temporal patterns with deep neural networks.\" Lai, Guokun, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu, SIGIR 2018.\n\n[3] \"Shape and time distortion loss for training deep time series forecasting models.\" Vincent, L. E., and Nicolas Thome NeurIPS 2019.\n\nAs for the experiments:\n1. It is not clear whether the setting in Eq. (1) is consistent with the settings in Informer or Reformer.\n2. It is also not clear how to set y’ in the experiments.\n3. Only two datasets are used for evaluation, which may not be sufficient to show the generalization capability of the proposed technique.\n4. Standard deviations of the prediction results are not provided.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents Yformer to perform long sequence time series forecasting. The key idea is to employ skip connection to improve the prediction resolution and stabilize the encoder and decoder by reconstructing the recent past. The experiment results on two datasets showed the effectiveness of the proposed method.", "main_review": "Strengths\n* Llong range time series forecasting is an interesting problem to investigate.\n* Adding skip connections between encoder and decoder is technically sound\n* The overall experiment results showed the effectiveness of the proposed method.\n\nWeaknesses\n* The organization of this paper is not well. Many technical details are not very clear in the main context\n* The overall technical novelty is limited\n* The effectiveness of the skip connections are not fully assessed\n* More details of the experiments are not provided.\n\nThe main problem of this paper is that the main context (especially the methodology section) is not self-contained. The reader will have to rely on details in the appendix or other papers to fully understand the proposed technique.\n\nAnother concern is the novelty. Skip connections are common practice in U-net and the idea of stabilizing the encoder and decoder by reconstructing the recent past is also not new. Although it is a new application area for skip connections, the overall technical novelty is limited.\n\nIn addition, the ablation study over whether the Skip connections are used or not is not provided.\n\nSeveral related works are not mentioned or compared.\n\n[1] \"Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting.\" Sen, Rajat, Hsiang-Fu Yu, and Inderjit S. Dhillon NeurIPS 2019.\n\n[2] \"Modeling long-and short-term temporal patterns with deep neural networks.\" Lai, Guokun, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu, SIGIR 2018.\n\n[3] \"Shape and time distortion loss for training deep time series forecasting models.\" Vincent, L. E., and Nicolas Thome NeurIPS 2019.\n\nAs for the experiments:\n1. It is not clear whether the setting in Eq. (1) is consistent with the settings in Informer or Reformer.\n2. It is also not clear how to set y’ in the experiments.\n3. Only two datasets are used for evaluation, which may not be sufficient to show the generalization capability of the proposed technique.\n4. Standard deviations of the prediction results are not provided.\n", "summary_of_the_review": "See above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N.A.", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636343904730}, {"id": "KSdi0bFXBSV", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1880/Reviewer_p7jA"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "A Former model is proposed in this paper, based on a Y-shaped encoder-decoder architecture that(1) uses direct connection from the downscaled encoder layer to the corresponding upsampled decoder layer in a U-Net inspired architecture, and (2) combines the downscaling/upsampling with sparse attention to capture long-range effects, and (3) stabilizes the encoder-decoder stars with the addition of an auxiliary reconstruction loss. The proposed model is evaluated on ETT and ECL dataset, and showed superior performance against baseline models including LogTransformer, LSTnet, Informer and Informer*.", "review_text": "Strength:\n1. The paper is well written and the proposed framework is easy to understand. In addition, I believe the mathematical description of the model is correct.\n2. Extensive evaluation is conducted, and the proposed YFormer shows an average of over 10% improvement compared with state-of-the-art models.\n3. A good ablation study is provided to justify the choice of the proposed architecture, and also hyper parameter selection. \n4. The authors of this paper choose baseline very carefully. They mentioned the reason why they are comparing with certain baseline models, identified some issues in some of the baseline models, and also provided reasoning why models such as Query Selector is not being used as a baseline model. I believe this thorough investigation and understanding of previous works is very important.\n\nWeakness:\n1. The mathematical description of the proposed architecture and task, although correct, is a bit over complicated. For example, section 3 describes a standard time-series forecasting problem with its corresponding notations. I would encourage the authors to review the notations needed in this section. I think some of them are not being used afterwards.\n2.Since the results provided is an average of three runs. It would be beneficial if the authors could provide the standard deviation of the results as well. It would be informative to have an estimation of the variance of the proposed model.\n3. In the abstract, the authors claim the model is tested on four datasets, in section 6.1,  it is said to be two real-world public datasets and one public benchmark. However, (I could be missing it somewhere), seems like only two datasets - ETT and ECL, are being evaluated on.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "A Former model is proposed in this paper, based on a Y-shaped encoder-decoder architecture that(1) uses direct connection from the downscaled encoder layer to the corresponding upsampled decoder layer in a U-Net inspired architecture, and (2) combines the downscaling/upsampling with sparse attention to capture long-range effects, and (3) stabilizes the encoder-decoder stars with the addition of an auxiliary reconstruction loss. The proposed model is evaluated on ETT and ECL dataset, and showed superior performance against baseline models including LogTransformer, LSTnet, Informer and Informer*.", "main_review": "Strength:\n1. The paper is well written and the proposed framework is easy to understand. In addition, I believe the mathematical description of the model is correct.\n2. Extensive evaluation is conducted, and the proposed YFormer shows an average of over 10% improvement compared with state-of-the-art models.\n3. A good ablation study is provided to justify the choice of the proposed architecture, and also hyper parameter selection. \n4. The authors of this paper choose baseline very carefully. They mentioned the reason why they are comparing with certain baseline models, identified some issues in some of the baseline models, and also provided reasoning why models such as Query Selector is not being used as a baseline model. I believe this thorough investigation and understanding of previous works is very important.\n\nWeakness:\n1. The mathematical description of the proposed architecture and task, although correct, is a bit over complicated. For example, section 3 describes a standard time-series forecasting problem with its corresponding notations. I would encourage the authors to review the notations needed in this section. I think some of them are not being used afterwards.\n2.Since the results provided is an average of three runs. It would be beneficial if the authors could provide the standard deviation of the results as well. It would be informative to have an estimation of the variance of the proposed model.\n3. In the abstract, the authors claim the model is tested on four datasets, in section 6.1,  it is said to be two real-world public datasets and one public benchmark. However, (I could be missing it somewhere), seems like only two datasets - ETT and ECL, are being evaluated on.", "summary_of_the_review": "Overall I think this is a good paper. Please refer to the above section for detailed review.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635884979526}, {"id": "DUZ1PFGSxRz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1880/Reviewer_yrSd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a new Transformer-based architecture for long-sequence temporal forecasting (LSTF) utilising ProbSparse attention mechanisms to efficiently capture long-term dependencies with L log(L) complexity.\n\nThe Yformer builds on the Informer architecture with 3 key innovations:\n1.\tUsing distinct encoders to capture historical and known future information separately. This improves representation learning for time series data, while still maintaining computational efficiency with ProbSparse attention.\n2.\tUsing a common decoder to process encoder representations jointly. The is also contains an upsampling step inspired by U-Net, although the benefits of upsampling are not explicitly evaluated.\n3.\tIncluding an auxiliary reconstruction loss which uses the reconstruction error of past targets to regularise training.\n", "review_text": "Strengths\n---\nOverall, the proposed architecture is intuitively compelling – echoing innovations observed in multi-horizon forecasting architectures (see related works comment below), while improving computational complexity using ProbSparse attention and downsampling. The strong improvements over the Informer baseline in numerous experiments also convincingly demonstrate the benefits of the proposed model for the LSTF problem.\n\nWeaknesses\n---\nHowever, there are several key limitations that need to be addressed before the paper can be recommended for acceptance:\n1.\t**Architectural details** – While the network diagrams and descriptive text do a good job in providing a high-level overview, the lack of details make it difficult to evaluate the architecture in depth. For instance, a couple of questions come to mind: \n  *\tDo all historical input features need to be known in the future? The problem formulation is confusing here as x and x’ are R^{* x M} which seems to imply identical lengths T and number of features T – despite the text mentioning x’ is from T to T+tau. \n  *\tHow are dimensions modified in each layer of the network? As the downsampling/upsampling parallels to U net appear to be a key part of the model, details on how this is performed is important. \n  *\tWhat are the keys, queries and values used for each attention layer (ProbSparseAttn, MaskedAttn, ProbSparseCrossAttn), and how is ProbSpraseCrossAttn implemented concretely?\n  *\tWhat is the length of the Conv1d filters in the various blocks, and are they purely linear transformations? Do dimensions change between each transformation?\n  *\tIs masked self-attention essential in the Y-Future encoder, and any reason why ProbSparse is not preferred? Does this affect computational efficiency, given that forecasting horizons appear to be larger than history lengths in many experiments from Appendix E.2?\n2.\t**Related works** -- While the authors do a good job of citing models for LSTF, the paper lacks references to modern neural forecasting architectures, many of which are attention-based and show improvements over LogTrans [2, 3] and DeepAR [1-3]. While computationally more inefficient, they also contain similar modifications to those proposed by the YFormer. For instance, [2, 3] use of distinct encoding mechanisms for historical inputs, future inputs, and static variables -- all of which are fed into a common attention-based decoder. In addition, [1] also trains the network using past targets as a regulariser (backcast). Comparisons to these models would help to further motivate the YFormer architecture as well.\n3.\t**Benchmarks** -- Given the focus on LSTF, comparison to simpler architectures that allow for extended receptive fields, e.g. dilated convolutions with WaveNet, would be useful. This is particularly important for time series datasets, which can be prone to overfitting with complex models -- as shown by the short-horizon outperformance of DeepAR on the ECL dataset in the Informer paper.\n\nTypos\n1.\tDeepAR is also mentioned as a benchmark, although results are not included in the paper.\n\nReferences\n1.\tOreshkin et al. N-BEATS: NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES FORECASTING. ICLR 2020.\n2.\tLim et al. TEMPORAL FUSION TRANSFORMERS FOR INTERPRETABLE MULTI-HORIZON TIME SERIES FORECASTING. International Journal of Forecasting, Volume 3 Issue 4, 2021.\n3.\tEisenach et al. MQTRANSFORMER: MULTI-HORIZON FORECASTS WITH CONTEXT DEPENDENT AND FEEDBACK-AWARE ATTENTION. Arxiv 2020.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a new Transformer-based architecture for long-sequence temporal forecasting (LSTF) utilising ProbSparse attention mechanisms to efficiently capture long-term dependencies with L log(L) complexity.\n\nThe Yformer builds on the Informer architecture with 3 key innovations:\n1.\tUsing distinct encoders to capture historical and known future information separately. This improves representation learning for time series data, while still maintaining computational efficiency with ProbSparse attention.\n2.\tUsing a common decoder to process encoder representations jointly. The is also contains an upsampling step inspired by U-Net, although the benefits of upsampling are not explicitly evaluated.\n3.\tIncluding an auxiliary reconstruction loss which uses the reconstruction error of past targets to regularise training.\n", "main_review": "Strengths\n---\nOverall, the proposed architecture is intuitively compelling – echoing innovations observed in multi-horizon forecasting architectures (see related works comment below), while improving computational complexity using ProbSparse attention and downsampling. The strong improvements over the Informer baseline in numerous experiments also convincingly demonstrate the benefits of the proposed model for the LSTF problem.\n\nWeaknesses\n---\nHowever, there are several key limitations that need to be addressed before the paper can be recommended for acceptance:\n1.\t**Architectural details** – While the network diagrams and descriptive text do a good job in providing a high-level overview, the lack of details make it difficult to evaluate the architecture in depth. For instance, a couple of questions come to mind: \n  *\tDo all historical input features need to be known in the future? The problem formulation is confusing here as x and x’ are R^{* x M} which seems to imply identical lengths T and number of features T – despite the text mentioning x’ is from T to T+tau. \n  *\tHow are dimensions modified in each layer of the network? As the downsampling/upsampling parallels to U net appear to be a key part of the model, details on how this is performed is important. \n  *\tWhat are the keys, queries and values used for each attention layer (ProbSparseAttn, MaskedAttn, ProbSparseCrossAttn), and how is ProbSpraseCrossAttn implemented concretely?\n  *\tWhat is the length of the Conv1d filters in the various blocks, and are they purely linear transformations? Do dimensions change between each transformation?\n  *\tIs masked self-attention essential in the Y-Future encoder, and any reason why ProbSparse is not preferred? Does this affect computational efficiency, given that forecasting horizons appear to be larger than history lengths in many experiments from Appendix E.2?\n2.\t**Related works** -- While the authors do a good job of citing models for LSTF, the paper lacks references to modern neural forecasting architectures, many of which are attention-based and show improvements over LogTrans [2, 3] and DeepAR [1-3]. While computationally more inefficient, they also contain similar modifications to those proposed by the YFormer. For instance, [2, 3] use of distinct encoding mechanisms for historical inputs, future inputs, and static variables -- all of which are fed into a common attention-based decoder. In addition, [1] also trains the network using past targets as a regulariser (backcast). Comparisons to these models would help to further motivate the YFormer architecture as well.\n3.\t**Benchmarks** -- Given the focus on LSTF, comparison to simpler architectures that allow for extended receptive fields, e.g. dilated convolutions with WaveNet, would be useful. This is particularly important for time series datasets, which can be prone to overfitting with complex models -- as shown by the short-horizon outperformance of DeepAR on the ECL dataset in the Informer paper.\n\nTypos\n1.\tDeepAR is also mentioned as a benchmark, although results are not included in the paper.\n\nReferences\n1.\tOreshkin et al. N-BEATS: NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES FORECASTING. ICLR 2020.\n2.\tLim et al. TEMPORAL FUSION TRANSFORMERS FOR INTERPRETABLE MULTI-HORIZON TIME SERIES FORECASTING. International Journal of Forecasting, Volume 3 Issue 4, 2021.\n3.\tEisenach et al. MQTRANSFORMER: MULTI-HORIZON FORECASTS WITH CONTEXT DEPENDENT AND FEEDBACK-AWARE ATTENTION. Arxiv 2020.\n", "summary_of_the_review": "While the results do clearly show improvements in both forecasting performance and computational efficiency, many additional details on the architecture need to be included before the paper is ready for publication.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635781984982}, {"id": "DYqiNQ2GbJC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1880/Reviewer_HmTV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "Recent works such as the Informer have used efficient attention mechanisms and shown significant performance improvements in the long sequence time-series forecasting problems. However, the authors argued that using only the coarsest past representations for the decoder could be a major limitation. In this paper, the authors proposed the Yformer model by combining the Informer and the U-Net architectures. They adopted direct connections from the multi-resolution encoder to decoder to leverage both coarse and fine-grained representations. The authors claimed the effectiveness of the proposed method through three benchmark datasets used in the Informer paper. ", "review_text": "-\tWhile the proposed methods hold great promise, my biggest concern is that the experiments seem like unfair comparisons. The authors compare the performance of Yformer against the excerpted results from the Informer. However, in my view, the Yformer and Informer use different problem formulations. According to the authors’ problem formulation, the Yformer predicts the future targets y’ based on the three inputs: past predictors x, past targets y, and future predictors x’. On the other hand, the Informer does not rely on the future predictors x’. The claimed performance improvement by the Yformer could be due to additional information within the future predictors. Furthermore, I’m not sure whether the authors’ problem formulation is appropriate in the real-world setting. Future predictors such as power load features in the ETT dataset would not be the “known” variables for the prediction.\n-\tWhile, in the abstract, the authors stated that they used four benchmark datasets, the manuscript only contains experiment results for the three benchmark datasets. Compared to the Informer paper, it seems the results for the Weather dataset are missing. If there are no particular reasons for the exclusion, can you also provide results for the Weather dataset?\n-\tIn my view, the authors must provide more detailed explanations for the proposed model to be self-contained. I think the current version is not easy to follow if readers were not already familiar with the Informer and the U-Net. In addition, the current version does not provide detailed information on dataset statistics (e.g. number of predictors and targets) and their pre-processing procedures.\n-\tCan you provide more in-depth experiment analyses for showing (1) how the U-Net shaped architecture helps long sequence time-series forecasting and (2) how does it affect the computational and memory efficiency of the model.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Recent works such as the Informer have used efficient attention mechanisms and shown significant performance improvements in the long sequence time-series forecasting problems. However, the authors argued that using only the coarsest past representations for the decoder could be a major limitation. In this paper, the authors proposed the Yformer model by combining the Informer and the U-Net architectures. They adopted direct connections from the multi-resolution encoder to decoder to leverage both coarse and fine-grained representations. The authors claimed the effectiveness of the proposed method through three benchmark datasets used in the Informer paper. ", "main_review": "-\tWhile the proposed methods hold great promise, my biggest concern is that the experiments seem like unfair comparisons. The authors compare the performance of Yformer against the excerpted results from the Informer. However, in my view, the Yformer and Informer use different problem formulations. According to the authors’ problem formulation, the Yformer predicts the future targets y’ based on the three inputs: past predictors x, past targets y, and future predictors x’. On the other hand, the Informer does not rely on the future predictors x’. The claimed performance improvement by the Yformer could be due to additional information within the future predictors. Furthermore, I’m not sure whether the authors’ problem formulation is appropriate in the real-world setting. Future predictors such as power load features in the ETT dataset would not be the “known” variables for the prediction.\n-\tWhile, in the abstract, the authors stated that they used four benchmark datasets, the manuscript only contains experiment results for the three benchmark datasets. Compared to the Informer paper, it seems the results for the Weather dataset are missing. If there are no particular reasons for the exclusion, can you also provide results for the Weather dataset?\n-\tIn my view, the authors must provide more detailed explanations for the proposed model to be self-contained. I think the current version is not easy to follow if readers were not already familiar with the Informer and the U-Net. In addition, the current version does not provide detailed information on dataset statistics (e.g. number of predictors and targets) and their pre-processing procedures.\n-\tCan you provide more in-depth experiment analyses for showing (1) how the U-Net shaped architecture helps long sequence time-series forecasting and (2) how does it affect the computational and memory efficiency of the model.\n", "summary_of_the_review": "While the proposed methods hold great promise, it has several issues to be addressed regarding the fairness of the experiments, a missing experiment dataset, and more detailed explanations to be self-contained. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635149930052}], "openreview_url": "https://openreview.net/forum?id=dYUdt59fJ0e", "arxiv_id": "2110.08255", "paper_pdf": "papers/dYUdt59fJ0e.pdf", "paper_pdf_sha256": "799459e97228541446cb56257f2b0b786c26f54be7b5fccb0029dff53e67b378", "paper_pdf_bytes": 750164, "paper_pdf_source": "openreview", "code_url": "https://github.com/18kiran12/Yformer-Time-Series-Forecasting", "code_repository": "18kiran12/Yformer-Time-Series-Forecasting", "code_commit": "65783ab25ad350d55bc12938630565727eb03b2b", "code_archive": "repos/dYUdt59fJ0e.zip", "code_archive_sha256": "09f688a196f54fd5ca348575574ef522255341bb2318150010b49206daee31f9", "code_archive_bytes": 26078, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 29, "github_languages": {"Python": 90622}, "github_archived": false, "github_pushed_at": "2022-08-25T15:15:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/yformer-u-net-inspired-transformer-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iMKvxHlrZb3", "year": 2021, "status": "rejected", "title": "Scalable Graph Neural Networks for Heterogeneous Graphs", "authors": ["Lingfan Yu", "Jiajun Shen", "Jinyang Li", "Adam Lerer"], "authorids": ["~Lingfan_Yu1", "~Jiajun_Shen1", "~Jinyang_Li1", "~Adam_Lerer1"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) are a popular class of parametric model for learning over graph-structured data. Recent work has argued that GNNs primarily use the graph for feature smoothing, and have shown competitive results on benchmark tasks by simply operating on graph-smoothed node features, rather than using end-to-end learned feature hierarchies that are challenging to scale to large graphs. In this work, we ask whether these results can be extended to heterogeneous graphs, which encode multiple types of relationship between different entities. We propose Neighbor Averaging over Relation Subgraphs (NARS), which trains a classifier on neighbor-averaged features for randomly-sampled subgraphs of the ‘metagraph‘ of relations. We describe optimizations to allow these sets of node features to be computed in a memory-efficient way, both at training and inference time. NARS achieves a new state of the art accuracy on several benchmark datasets, outperforming more expensive GNN-based methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1tWiuOT9FQi", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper955/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims to propose a new GNN for heterogeneous graphs, which is scalable to large-scale graphs. The proposed idea is to leverage an existing model called SIGN, which simplifies GCN by dropping the non-linear transformation from intermediate layers, and extend it to heterogeneous graphs. The results on several benchmark datasets show the proposed approach is better and faster than baselines.\n\nAlthough the proposed idea seems interesting, there are several concerns about the paper.\n\n1.\tThe description of the methodology is very vague. For example, Fig. 1 is presented without detailed explanation. It is unclear how the features computed by different subgraphs can be aggregated, especially consider nodes only appear in a subset of those subgraphs. The formula in Eq. (2) and (3) do not help due to their simplicity.\n2.\tThe novelty of the paper is also limited, consider it is extending an existing algorithm SIGN to heterogeneous version.\n3.\tI am not fully convinced the simplified GNN works better than some other GNNs designed for heterogeneous networks, such as HAN and HGT, which contains attention scheme and carefully models the message type in the GNN framework. According to other simplified GNN paper, their performance is just comparable to their counterpart but not better, and usually the results are worse than the attention-based one.\n4.\tIt seems the numbers in Table 4 is higher than the ones in the original paper, such as HGT. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper aims to propose a new GNN for heterogeneous graphs, which is scalable to large-scale graphs. The proposed idea is to leverage an existing model called SIGN, which simplifies GCN by dropping the non-linear transformation from intermediate layers, and extend it to heterogeneous graphs. The results on several benchmark datasets show the proposed approach is better and faster than baselines.", "review": "This paper aims to propose a new GNN for heterogeneous graphs, which is scalable to large-scale graphs. The proposed idea is to leverage an existing model called SIGN, which simplifies GCN by dropping the non-linear transformation from intermediate layers, and extend it to heterogeneous graphs. The results on several benchmark datasets show the proposed approach is better and faster than baselines.\n\nAlthough the proposed idea seems interesting, there are several concerns about the paper.\n\n1.\tThe description of the methodology is very vague. For example, Fig. 1 is presented without detailed explanation. It is unclear how the features computed by different subgraphs can be aggregated, especially consider nodes only appear in a subset of those subgraphs. The formula in Eq. (2) and (3) do not help due to their simplicity.\n2.\tThe novelty of the paper is also limited, consider it is extending an existing algorithm SIGN to heterogeneous version.\n3.\tI am not fully convinced the simplified GNN works better than some other GNNs designed for heterogeneous networks, such as HAN and HGT, which contains attention scheme and carefully models the message type in the GNN framework. According to other simplified GNN paper, their performance is just comparable to their counterpart but not better, and usually the results are worse than the attention-based one.\n4.\tIt seems the numbers in Table 4 is higher than the ones in the original paper, such as HGT. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604038443698}, {"id": "zXybr9FEmON", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper955/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper extends from SIGN (https://arxiv.org/abs/2004.11198) model to heterogeneous graphs.\n\nThe SIGN model argues that simply applying MLP on graph-smoothed node features (concatenating k-th hop neighbor features, k $\\in$ [1-L]) can achieve similar results compared with learnable aggregation applied in GNNs. To extend to the heterogeneous graph, this paper proposes to sample relation graphs, by:\n(1) sample several subsets $R_i$ of relations; \n(2) sample relation subgraphs whose edges belong to $R_i$; \n(3) treat each subgraph as homogeneous graphs and perform neighbor aggregation (simply average). \n(4) Apply MLP on each node for node classification.\n\nI have several questions about the proposed approaches:\n\n(1) The difficulty of heterogeneous graphs is that each node might have different types of features. For example, in a social network, nodes can be associated with image, text, or some discrete profiles. Thus, the neighborhood smoothing only works when the input features are both 1) already very informative, and don't need too much transformation; 2) features from different node types are projected to the same space. Therefore, I'm afraid the authors' proposed aggregation might not generalize to more complicated heterogeneous graphs. (It seems that during experiments, the authors utilize TransE to pre-train embeddings for all the nodes, so that they are naturally within the same space, making the problem simpler. One evidence is that when using other feature initialization strategies, such as simple average, the performance of this model drops significantly)\n\n(2) Also, it's confusing to me why can we fuse all the subgraphs with different subgraph schema. Intuitively, with different relation set, the semantic of the relation subgraph should be very different, but the authors seem to treat them equally. It would be better if the authors can provide some analysis on this, for example, for a given node, what is the variance of final node embeddings calculated with subgraphs of different relation sets.\n\n(3) How to get the inference results for large graphs? It seems that the proposed method should get a different predictions for each node with a different relation set. So which set the authors to use? Complete set or average over random sampling? (If it's random, the reported variance, which is close to 0, seems to be very strange).\n\nAlso, though the authors show superior experimental results, I have several concerns about experiment settings:\n\n(1) About feature initialization. From Table 5, we can see that the proposed NARS method highly relies on the TransE\n embedding initialization. When using a standard feature initialization method (such as average neighbors), the results are much lower than HGT and R-GCN. However, the authors didn't provide implementation details about how to train such TransE embedding (normally it's weird to use TransE for heterogeneous graph, as the node number can be much larger than the knowledge graph and we don't have that much relation type. For example, two papers published by the same author and on the same venue might have exactly the same TransE embedding, if we don't consider text input. So it's confusing to me why the results with TransE embedding is better). The authors should better elaborate on this part or release the code for clarity.\n\n(2) About baseline results. Since the utilization of TransE embedding, the experimental settings of the baseline are different from the original papers. But there's still some confusing part. For example, the HGT model's implementation on OGB-MAG uses neighbor average strategy, and the accuracy result is 0.5, while the result reported in table 5 is 0.489. Also, the model parameter is not matched with the reported number. \n\n(3) About inference time. As discussed above, I'm not sure how the proposed method can efficiently get accurate inferences for all the nodes in the test set. If the authors want to claim their method is more scalable, it would be better to include the inference time comparison.\n\n\nOverall, I think the simplified procedure (direct neighbor average) over heterogeneous graph limits the usage of this model, and there's also some unclear part in experimental settings.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review #3", "review": "This paper extends from SIGN (https://arxiv.org/abs/2004.11198) model to heterogeneous graphs.\n\nThe SIGN model argues that simply applying MLP on graph-smoothed node features (concatenating k-th hop neighbor features, k $\\in$ [1-L]) can achieve similar results compared with learnable aggregation applied in GNNs. To extend to the heterogeneous graph, this paper proposes to sample relation graphs, by:\n(1) sample several subsets $R_i$ of relations; \n(2) sample relation subgraphs whose edges belong to $R_i$; \n(3) treat each subgraph as homogeneous graphs and perform neighbor aggregation (simply average). \n(4) Apply MLP on each node for node classification.\n\nI have several questions about the proposed approaches:\n\n(1) The difficulty of heterogeneous graphs is that each node might have different types of features. For example, in a social network, nodes can be associated with image, text, or some discrete profiles. Thus, the neighborhood smoothing only works when the input features are both 1) already very informative, and don't need too much transformation; 2) features from different node types are projected to the same space. Therefore, I'm afraid the authors' proposed aggregation might not generalize to more complicated heterogeneous graphs. (It seems that during experiments, the authors utilize TransE to pre-train embeddings for all the nodes, so that they are naturally within the same space, making the problem simpler. One evidence is that when using other feature initialization strategies, such as simple average, the performance of this model drops significantly)\n\n(2) Also, it's confusing to me why can we fuse all the subgraphs with different subgraph schema. Intuitively, with different relation set, the semantic of the relation subgraph should be very different, but the authors seem to treat them equally. It would be better if the authors can provide some analysis on this, for example, for a given node, what is the variance of final node embeddings calculated with subgraphs of different relation sets.\n\n(3) How to get the inference results for large graphs? It seems that the proposed method should get a different predictions for each node with a different relation set. So which set the authors to use? Complete set or average over random sampling? (If it's random, the reported variance, which is close to 0, seems to be very strange).\n\nAlso, though the authors show superior experimental results, I have several concerns about experiment settings:\n\n(1) About feature initialization. From Table 5, we can see that the proposed NARS method highly relies on the TransE\n embedding initialization. When using a standard feature initialization method (such as average neighbors), the results are much lower than HGT and R-GCN. However, the authors didn't provide implementation details about how to train such TransE embedding (normally it's weird to use TransE for heterogeneous graph, as the node number can be much larger than the knowledge graph and we don't have that much relation type. For example, two papers published by the same author and on the same venue might have exactly the same TransE embedding, if we don't consider text input. So it's confusing to me why the results with TransE embedding is better). The authors should better elaborate on this part or release the code for clarity.\n\n(2) About baseline results. Since the utilization of TransE embedding, the experimental settings of the baseline are different from the original papers. But there's still some confusing part. For example, the HGT model's implementation on OGB-MAG uses neighbor average strategy, and the accuracy result is 0.5, while the result reported in table 5 is 0.489. Also, the model parameter is not matched with the reported number. \n\n(3) About inference time. As discussed above, I'm not sure how the proposed method can efficiently get accurate inferences for all the nodes in the test set. If the authors want to claim their method is more scalable, it would be better to include the inference time comparison.\n\n\nOverall, I think the simplified procedure (direct neighbor average) over heterogeneous graph limits the usage of this model, and there's also some unclear part in experimental settings.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603930562606}, {"id": "WJVB599S8lG", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper955/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studied heterogeneous graph embedding with graph neural networks. The authors proposed an approach which samples multiple different subgraphs containing a subset of relations and then aggregate the node representations from different subgraphs with attentions. Experimental results on some big data sets prove the effectiveness and efficiency of the proposed approach. \n\nOverall, heterogenous graph embedding with graph neural network is a very important and interesting problem with a variety of applications. However, the novelty of the proposed approach seems to be quite marginal, and I am not quite convinced by the proposed approach. Why are we able to get better results by first sampling a subset of relations to get the subgraph and then aggregating the node representations from different subgraphs? It seems weird to be as a randomly selected subset of relations do not convey any specific semantic meanings as traditional metapath based approaches. I am also surprised by the very small standard deviation in Table 4, which is not consistent with results reported in existing literature on the tasks of node classification. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Problem, but the novelty of the technical depth is marginal", "review": "This paper studied heterogeneous graph embedding with graph neural networks. The authors proposed an approach which samples multiple different subgraphs containing a subset of relations and then aggregate the node representations from different subgraphs with attentions. Experimental results on some big data sets prove the effectiveness and efficiency of the proposed approach. \n\nOverall, heterogenous graph embedding with graph neural network is a very important and interesting problem with a variety of applications. However, the novelty of the proposed approach seems to be quite marginal, and I am not quite convinced by the proposed approach. Why are we able to get better results by first sampling a subset of relations to get the subgraph and then aggregating the node representations from different subgraphs? It seems weird to be as a randomly selected subset of relations do not convey any specific semantic meanings as traditional metapath based approaches. I am also surprised by the very small standard deviation in Table 4, which is not consistent with results reported in existing literature on the tasks of node classification. ", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603903085403}, {"id": "3PImdhRId5K", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper955/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a method to broaden the scope of SIGN, a technique recently introduced for single-relational graphs. The method allows SIGN to also be applied to multi-relational graphs (often called heterogeneous or knowledge graphs in different communities). In SIGN, various powers of the Laplacian are precomputed. For each power, the features of the nodes of a node’s neighborhood are averaged and (e.g., with an MLP) projected into a node vector representation. This is then used to classify the node. \n\nOne way to apply SIGN to multi-relational graphs would be to conflate all relation types into a single relation type. The authors show that this doesn’t work well. \n\nThe authors then propose the following: take the set of all relation types R and sample a subset R’ from it. Construct the subgraph induced by R’ by only keeping edges of relation types in R’. Then treat all of these relation types as one (turn into a single-relational subgraph) and apply (essentially) SIGN to this subgraph. Do this several times, that is, sample subsets of R several times. Finally aggregate the representations coming from each of the application of SIGN to the sampled subgraphs.\n\nThe authors then also discuss what to do if the nodes do not have attributes (features) and ways to improve the memory efficiency of the approach. \n\nThe authors then proceed to empirically evaluate their method. They compare to existing approaches (such as R-GCNs) and show that NARS (the name of their approach) is competitive with these existing methods. Often it is significantly better. They also perform several types of ablation studies, analyzing the impact of choices such as the number of subgraphs sampled. Generally, I would say that the experiments are well done and look at various different important questions. \n\nIn summary, I don’t have much to criticise. The only shortcoming is that the method is more or less an adaptation of SIGN to the multi-relational setting. I want to be careful to lament a “lack of novelty” here, because it is still a good contribution to broaden the scope of an existing method to a larger class of problems. However, one thing I’m missing is a more thorough analysis of the sampling strategies one could use. What is analyzed is the number of subgraphs sampled. But this is based on one way of sampling subgraphs (based on first sampling relation types and inducing the graph based on this). What I’m missing are: sampling subgraphs from the Gaifman graph; and sampling multi-relational graphs and applying R-GCN on each of those and aggregating. The latter is probably less efficient, but I would conjecture also much better than running one R-GCN. Again, see related work below. It has been shown before that sampling multi-relational subgraphs can add robustness of the model to overfitting. Overall, I would tend towards accepting the paper but I would really like to see more analysis on different sampling strategies. \n\nI would encourage the authors to consider as related work a paper from 2017 which is one of the first to propose neighborhood sampling for GNNs (much earlier than most papers you mention in this context) and which is one of the first ones who simply average (learned) features but with an unsupervised objective:\n\nGarcia-Duran and Niepert, Learning Graph Representations with Embedding Propagation, NeurIPS 2017\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Broadening the scope of SIGN to multi-relational graphs", "review": "The authors propose a method to broaden the scope of SIGN, a technique recently introduced for single-relational graphs. The method allows SIGN to also be applied to multi-relational graphs (often called heterogeneous or knowledge graphs in different communities). In SIGN, various powers of the Laplacian are precomputed. For each power, the features of the nodes of a node’s neighborhood are averaged and (e.g., with an MLP) projected into a node vector representation. This is then used to classify the node. \n\nOne way to apply SIGN to multi-relational graphs would be to conflate all relation types into a single relation type. The authors show that this doesn’t work well. \n\nThe authors then propose the following: take the set of all relation types R and sample a subset R’ from it. Construct the subgraph induced by R’ by only keeping edges of relation types in R’. Then treat all of these relation types as one (turn into a single-relational subgraph) and apply (essentially) SIGN to this subgraph. Do this several times, that is, sample subsets of R several times. Finally aggregate the representations coming from each of the application of SIGN to the sampled subgraphs.\n\nThe authors then also discuss what to do if the nodes do not have attributes (features) and ways to improve the memory efficiency of the approach. \n\nThe authors then proceed to empirically evaluate their method. They compare to existing approaches (such as R-GCNs) and show that NARS (the name of their approach) is competitive with these existing methods. Often it is significantly better. They also perform several types of ablation studies, analyzing the impact of choices such as the number of subgraphs sampled. Generally, I would say that the experiments are well done and look at various different important questions. \n\nIn summary, I don’t have much to criticise. The only shortcoming is that the method is more or less an adaptation of SIGN to the multi-relational setting. I want to be careful to lament a “lack of novelty” here, because it is still a good contribution to broaden the scope of an existing method to a larger class of problems. However, one thing I’m missing is a more thorough analysis of the sampling strategies one could use. What is analyzed is the number of subgraphs sampled. But this is based on one way of sampling subgraphs (based on first sampling relation types and inducing the graph based on this). What I’m missing are: sampling subgraphs from the Gaifman graph; and sampling multi-relational graphs and applying R-GCN on each of those and aggregating. The latter is probably less efficient, but I would conjecture also much better than running one R-GCN. Again, see related work below. It has been shown before that sampling multi-relational subgraphs can add robustness of the model to overfitting. Overall, I would tend towards accepting the paper but I would really like to see more analysis on different sampling strategies. \n\nI would encourage the authors to consider as related work a paper from 2017 which is one of the first to propose neighborhood sampling for GNNs (much earlier than most papers you mention in this context) and which is one of the first ones who simply average (learned) features but with an unsupervised objective:\n\nGarcia-Duran and Niepert, Learning Graph Representations with Embedding Propagation, NeurIPS 2017\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603875673995}], "openreview_url": "https://openreview.net/forum?id=iMKvxHlrZb3", "arxiv_id": "2011.09679", "paper_pdf": "papers/iMKvxHlrZb3.pdf", "paper_pdf_sha256": "8f537614d0cc5169b90c72da6a4458446bd153f9c641e438c0f9989b1707f73d", "paper_pdf_bytes": 544130, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/NARS", "code_repository": "facebookresearch/NARS", "code_commit": "27850d3812e04d2057ba6b53aae02fd35a0e180c", "code_archive": "repos/iMKvxHlrZb3.zip", "code_archive_sha256": "5b3c04190d6b98de1bf59d42f925c53226f0a5709aa3ef619ac69c6988c86818", "code_archive_bytes": 39672, "code_file_count": 16, "code_extensions": {".py": 12, ".sh": 4}, "github_disk_usage_kb": 34, "github_languages": {"Python": 54277, "Shell": 2468, "Dockerfile": 772}, "github_archived": true, "github_pushed_at": "2020-11-24T18:24:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/scalable-graph-neural-networks-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BylB4kBtwB", "year": 2020, "status": "rejected", "title": "Retrieving Signals in the Frequency Domain with Deep Complex Extractors", "authors": ["Chiheb Trabelsi", "Olexa Bilaniuk", "Ousmane Dia", "Ying Zhang", "Mirco Ravanelli", "Jonathan Binas", "Negar Rostamzadeh", "Christopher  J Pal"], "authorids": ["chiheb.trabelsi@polymtl.ca", "olexa.bilaniuk@umontreal.ca", "ousmane@elementai.com", "ying@elementai.com", "mirco.ravanelli@gmail.com", "jbinas@gmail.com", "negar@elementai.com", "christopher.pal@elementai.com"], "authors_source": "OpenReview API", "abstract": "Recent advances have made it possible to create deep complex-valued neural networks. Despite this progress, the potential power of fully complex intermediate computations and representations has not yet been explored for many challenging learning problems. Building on recent advances, we propose a novel mechanism for extracting signals in the frequency domain. As a case study, we perform audio source separation in the Fourier domain. Our extraction mechanism could be regarded as a local ensembling method that combines a complex-valued convolutional version of Feature-Wise Linear Modulation (FiLM) and a signal averaging operation. We also introduce a new explicit amplitude and phase-aware loss, which is scale and time invariant, taking into account the complex-valued components of the spectrogram. Using the Wall Street Journal Dataset, we compare our phase-aware loss to several others that operate both in the time and frequency domains and demonstrate the effectiveness of our proposed signal extraction method and proposed loss. When operating in the complex-valued frequency domain, our deep complex-valued network substantially outperforms its real-valued counterparts even with half the depth and a third of the parameters. Our proposed mechanism improves significantly deep complex-valued networks' performance and we demonstrate the usefulness of its regularizing effect.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rygaxccJcB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1653/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose complex valued neural networks to perform audio source separation in the Fourier domain. The adapt a well known U-Net architecture to the task by introducing a complex-valued FiLM layer and a new complex similarity loss that explicitly takes magnitude and phase into account. They motivate the use of complex values well and demonstrate performance and parameter efficiency improvements over real-valued baselines. Importantly, they do not need to perform spetrogram inversion because their network works natively in the complex domain. Despite the quantitative improvements over spectral models, they still slightly underperform the ConvTasNet baseline that operates directly in the waveform domain (which was slightly misleading to not include in the table). The authors perform an extensive hyperparameter search to tune the model and provide sufficient detail to reproduce their experiments. While the results did not improve upon the best baseline, they do provide further evidence to the value of using complex-valued neural networks to handle complex-valued data (where phase and synchronicity matter), which I believe will be of value to the ICLR community, and thus I lean slightly in favor of acceptance. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The authors propose complex valued neural networks to perform audio source separation in the Fourier domain. The adapt a well known U-Net architecture to the task by introducing a complex-valued FiLM layer and a new complex similarity loss that explicitly takes magnitude and phase into account. They motivate the use of complex values well and demonstrate performance and parameter efficiency improvements over real-valued baselines. Importantly, they do not need to perform spetrogram inversion because their network works natively in the complex domain. Despite the quantitative improvements over spectral models, they still slightly underperform the ConvTasNet baseline that operates directly in the waveform domain (which was slightly misleading to not include in the table). The authors perform an extensive hyperparameter search to tune the model and provide sufficient detail to reproduce their experiments. While the results did not improve upon the best baseline, they do provide further evidence to the value of using complex-valued neural networks to handle complex-valued data (where phase and synchronicity matter), which I believe will be of value to the ICLR community, and thus I lean slightly in favor of acceptance. "}, "tcdate": 1571953140635}, {"id": "B1gkXWp6tr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1653/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work researches the deep complex-valued neural networks. Specifically, it proposes a new signal extraction mechanism that operates in frequency domain and applies to address the speech separation issue. Also, a function is proposed to explicitly consider both the magnitude and phase information of a signal. Related work on learning representation in frequency domain and speech separation is well introduced. Theoretical analysis is conducted to show the motivation and connection to signal processing. The architecture of the deep neural networks is presented in details, with the elaboration of the complex mask generation. Experimental study is conducted on a benchmark dataset to compare the proposed complex networks with those using real-part values only to demonstrate the improvement. \n\nThe rating is 3: Weak Reject considering that the novelty is limited and the experimental study is weak. \n\n1. The significance of the theoretical analysis in Eq.(1) to Eq.(4) needs to be better explained. Currently, they seem to be some straightforward results in the field of signal processing;\n2. The proposed CSimLoss is interesting. However, its effectiveness seems to be limited as demonstrated in Table 1. It can be found that the CSimLoss in some cases is only comparable (or even inferior) to the L2freq loss;\n3. The mask generation proposed in Section 6 conceptually is largely an attention mechanism that has been widely applied in deep networks; \n4. The experimental comparison in Table 1 is limited, although some improvements have been demonstrated. This work shall also make a comparison with some of the existing methods on speech separation as described in Section 2.2.\n5. It was mentioned in the last paragraph of Section 7 that (Shi et al. 2019) uses different data preparation than this paper. Can this paper use the same data preparation as (Shi et al. 2019) and perform some comparisons?\n6. Why did (Shi et al. 2019) achieve better SDR (12.1 vs. 11.3) than the proposed method using standard setup?\n7. What if the mechanism of mask generation is also applied to Real U-Net? How much improvement can this bring? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This work researches the deep complex-valued neural networks. Specifically, it proposes a new signal extraction mechanism that operates in frequency domain and applies to address the speech separation issue. Also, a function is proposed to explicitly consider both the magnitude and phase information of a signal. Related work on learning representation in frequency domain and speech separation is well introduced. Theoretical analysis is conducted to show the motivation and connection to signal processing. The architecture of the deep neural networks is presented in details, with the elaboration of the complex mask generation. Experimental study is conducted on a benchmark dataset to compare the proposed complex networks with those using real-part values only to demonstrate the improvement. \n\nThe rating is 3: Weak Reject considering that the novelty is limited and the experimental study is weak. \n\n1. The significance of the theoretical analysis in Eq.(1) to Eq.(4) needs to be better explained. Currently, they seem to be some straightforward results in the field of signal processing;\n2. The proposed CSimLoss is interesting. However, its effectiveness seems to be limited as demonstrated in Table 1. It can be found that the CSimLoss in some cases is only comparable (or even inferior) to the L2freq loss;\n3. The mask generation proposed in Section 6 conceptually is largely an attention mechanism that has been widely applied in deep networks; \n4. The experimental comparison in Table 1 is limited, although some improvements have been demonstrated. This work shall also make a comparison with some of the existing methods on speech separation as described in Section 2.2.\n5. It was mentioned in the last paragraph of Section 7 that (Shi et al. 2019) uses different data preparation than this paper. Can this paper use the same data preparation as (Shi et al. 2019) and perform some comparisons?\n6. Why did (Shi et al. 2019) achieve better SDR (12.1 vs. 11.3) than the proposed method using standard setup?\n7. What if the mechanism of mask generation is also applied to Real U-Net? How much improvement can this bring? \n"}, "tcdate": 1571832086714}, {"id": "HygJE9ZXYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1653/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new method for source separation, by using deep learning UNets, complex-valued representations and the Fourier domain. Concretely, their contribution is : i) a complex-valued convolutional version of the Feature-Wise Linear Modulation, able to optimise the parameters needed to create multiple separated candidates for each of signal sources that are then combined using signal averaging; ii) the design of a loss that takes into account magnitude and phase while being scale and time invariant. It was then tested and compared with real-valued versions, and also some state-of-the-art methods.\n\nOverall, I think this paper is of good quality and proposes an interesting method for this crucial task of source separation. However, I found the paper too dense and difficult to read (even if well written), and it looks like a re-submission from a journal paper of more than 8 pages. I would suggest the authors to shrink the paper so it *really* fits into the 8-pages (without important figures or important implementation details in the appendices), maybe at the cost of leaving some parts (such as old related works) out of the paper. The experiments are important here, and it is too bad that the comparison with state-of-the-art is just in the last paragraph, while the results do not seem to show any improvements compared to the other methods. The computational time might be very important here, as the claim is that FFT reduces time computation, but I did not had time to go through all the appendices.\n\nPositive aspects:\n- The work is well documented and motivated, and I found that the reflexion leading to the method is of good quality. \n- Concretely, I found interesting the use of the FiLM, originally designed for another application, for minimizing the SNR of the signal sources. The motivation/proof is quite clear too.\n- Equally, the motivation for the design of the new loss is clear and interesting.\n- I also found important the experiments, that shows in the same table the difference between the method without the complex-valued part and with different parameter values.\n\nQuestions and remarks:\n- I have to recall that I am not an expert on source separation and complex-valued deep learning. Yet, I have had difficulties in understanding the structure of the method, even if the different parts were clearly explained. The figure 1 is very useful, but I found it not clear enough and too small. I went to find some informations in the appendices, but there are too much crucial information there and I did not have time to go through all of it.\n- The use of the U-net architecture is not explained (just some citations are given). What is supposed to be the output of it? \n- When you say 'to be more rigorous, we can assume in the cse of speech separation that, for each speaker, there exists an impulse response such that when it is convolved with the clean speech of the speaker, it allows to reconstruct the mix' : why can we be sure that it is always possible, and why is it more rigourous?\n- Why are the additive noises epsilon_i supposed to have the same E(|epsilon_i|^2|) ? Even if they are uncorrelated, what is the hypothesis behind that?\n- In the CSimLoss, why (i) is the real part negative and the imaginary part positive; (ii) is the imaginary part squared?\n- have you tested with a higher lambda_imag (as the larger is now the best)?\n- It looks from the end of the paper that the method is still not achieving better results than the state-of-the-art. I agree with the authors as it might not be the scope of the paper, but then what is it? If it's time computation, it is not shown in the paper. If it is just a methodology, what would be required in the future to beat the best method?\n- In the results, table 1, in the last 4 lines: it looks from 1st and 2nd line that the new loss CSimLoss is not very different from the L2 (9.88 compared to 9.87). The best result, in the 4th line, cannot be compared to the 3rd line as both the loss and the number of transforms are different. I then found those values in the appendices, but it would be best to show fewer parameters varying in the main paper, but show some results that can be easily compared. \n- What is the importance of the first paragraph in 2.1? I was not aware of the holographic reduced representations, but I don't understand it more now, and I don't see why explaining that for 15 lines.\n\nSmall remarks:\n- 'deep complex valued models have *just* started to gain momentum'... with citations beginning in 2014, I would not say 'just'.\n- 'in the frequncy domain is then, ...' --> frequency + no coma\n- Figure 2 is in the appendix, while in the text it is not said so. I was lost. This figure should not be in the appendix as the appendix should not have key elements, but just details that are not important for the understanding of the paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper proposes a new method for source separation, by using deep learning UNets, complex-valued representations and the Fourier domain. Concretely, their contribution is : i) a complex-valued convolutional version of the Feature-Wise Linear Modulation, able to optimise the parameters needed to create multiple separated candidates for each of signal sources that are then combined using signal averaging; ii) the design of a loss that takes into account magnitude and phase while being scale and time invariant. It was then tested and compared with real-valued versions, and also some state-of-the-art methods.\n\nOverall, I think this paper is of good quality and proposes an interesting method for this crucial task of source separation. However, I found the paper too dense and difficult to read (even if well written), and it looks like a re-submission from a journal paper of more than 8 pages. I would suggest the authors to shrink the paper so it *really* fits into the 8-pages (without important figures or important implementation details in the appendices), maybe at the cost of leaving some parts (such as old related works) out of the paper. The experiments are important here, and it is too bad that the comparison with state-of-the-art is just in the last paragraph, while the results do not seem to show any improvements compared to the other methods. The computational time might be very important here, as the claim is that FFT reduces time computation, but I did not had time to go through all the appendices.\n\nPositive aspects:\n- The work is well documented and motivated, and I found that the reflexion leading to the method is of good quality. \n- Concretely, I found interesting the use of the FiLM, originally designed for another application, for minimizing the SNR of the signal sources. The motivation/proof is quite clear too.\n- Equally, the motivation for the design of the new loss is clear and interesting.\n- I also found important the experiments, that shows in the same table the difference between the method without the complex-valued part and with different parameter values.\n\nQuestions and remarks:\n- I have to recall that I am not an expert on source separation and complex-valued deep learning. Yet, I have had difficulties in understanding the structure of the method, even if the different parts were clearly explained. The figure 1 is very useful, but I found it not clear enough and too small. I went to find some informations in the appendices, but there are too much crucial information there and I did not have time to go through all of it.\n- The use of the U-net architecture is not explained (just some citations are given). What is supposed to be the output of it? \n- When you say 'to be more rigorous, we can assume in the cse of speech separation that, for each speaker, there exists an impulse response such that when it is convolved with the clean speech of the speaker, it allows to reconstruct the mix' : why can we be sure that it is always possible, and why is it more rigourous?\n- Why are the additive noises epsilon_i supposed to have the same E(|epsilon_i|^2|) ? Even if they are uncorrelated, what is the hypothesis behind that?\n- In the CSimLoss, why (i) is the real part negative and the imaginary part positive; (ii) is the imaginary part squared?\n- have you tested with a higher lambda_imag (as the larger is now the best)?\n- It looks from the end of the paper that the method is still not achieving better results than the state-of-the-art. I agree with the authors as it might not be the scope of the paper, but then what is it? If it's time computation, it is not shown in the paper. If it is just a methodology, what would be required in the future to beat the best method?\n- In the results, table 1, in the last 4 lines: it looks from 1st and 2nd line that the new loss CSimLoss is not very different from the L2 (9.88 compared to 9.87). The best result, in the 4th line, cannot be compared to the 3rd line as both the loss and the number of transforms are different. I then found those values in the appendices, but it would be best to show fewer parameters varying in the main paper, but show some results that can be easily compared. \n- What is the importance of the first paragraph in 2.1? I was not aware of the holographic reduced representations, but I don't understand it more now, and I don't see why explaining that for 15 lines.\n\nSmall remarks:\n- 'deep complex valued models have *just* started to gain momentum'... with citations beginning in 2014, I would not say 'just'.\n- 'in the frequncy domain is then, ...' --> frequency + no coma\n- Figure 2 is in the appendix, while in the text it is not said so. I was lost. This figure should not be in the appendix as the appendix should not have key elements, but just details that are not important for the understanding of the paper.\n"}, "tcdate": 1571129895032}], "openreview_url": "https://openreview.net/forum?id=BylB4kBtwB", "arxiv_id": null, "paper_pdf": "papers/BylB4kBtwB.pdf", "paper_pdf_sha256": "4dc98b6bf38ee2e87ac865673422291b6bbd436a423ccf04793601a86104c7e0", "paper_pdf_bytes": 763456, "paper_pdf_source": "openreview", "code_url": "https://github.com/FourierSignalRetrievalICLR2020/FourierExtraction", "code_repository": "FourierSignalRetrievalICLR2020/FourierExtraction", "code_commit": "7dab26ced2f47c509f82963b5bb7ab459f535d55", "code_archive": "repos/BylB4kBtwB.zip", "code_archive_sha256": "c8b235b6f107088825f31fb58b816516fedd2c3ce93c82f6919579cedd61cd2f", "code_archive_bytes": 48797, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 37, "github_languages": {"Python": 220455}, "github_archived": false, "github_pushed_at": "2019-09-24T18:38:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/retrieving-signals-in-the-frequency-domain"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryfaViR9YX", "year": 2019, "status": "rejected", "title": "Variation Network: Learning High-level Attributes for Controlled Input Manipulation", "authors": ["Gaëtan Hadjeres"], "authorids": ["hadjeres.g@gmail.com"], "authors_source": "OpenReview API", "abstract": "This paper presents the Variation Network (VarNet), a  generative model providing means to manipulate the high-level attributes of a given input. The originality of our approach is that VarNet is not only capable of handling pre-defined attributes but can also learn the relevant attributes of the dataset by itself.  These two settings can be easily combined  which makes VarNet applicable for a wide variety of tasks. Further, VarNet has a sound probabilistic interpretation which grants us with  a novel way to navigate in the latent spaces as well as means to control how the  attributes are learned. We demonstrate  experimentally that this model is capable of performing interesting input manipulation  and that the learned attributes are relevant and interpretable.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HJljbAELp7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper42/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new framework for learning an interpretable representation of images and their attributes. The authors suggest decomposing the representation into a set of 'template' latent features, and a set of attribute-based features. The attribute-based features can be either 'free', i.e. discovered from the data, or 'fixed', i.e. based on the ground truth attributes. The authors encourage the decomposition of the latent space into the 'template' and the 'attributes' features by training a discriminator network to predict whether the attributes and the template features come from the same image or not.\n\nWhile the idea is interesting, the paper is lacking an experimental section, so the methodology is impossible to evaluate. Furthermore, while the authors spend many pages describing their methodology, the writing is often hard to follow, so I am still confused about the exact implementation of the attribute features \\phi(x, m) for example. The authors do point to the Appendix for their Experiments section, however this is not a good idea. The paper should be self-contained and the authors should not assume that their readers will read the information presented in the Appendix, which is always optional. \n\nUnfortunately, even the experimental section presented in the Appendix is not comprehensive enough to evaluate the proposed method. The authors train the model on a single dataset (MNIST), no baseline or ablation results are presented, and all the results are purely qualitative. Given that the ground truth attribute decomposition for MNIST is not known, even the qualitative results are impossible to evaluate. I recommend that the authors present quantitive results in the updated version of their paper (i.e. disentanglement metric scores, the log-likelihood of the reconstructions), including new experiments on a dataset like dSprites or CelebA, where the ground truth attributes are known.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Paper lacking an experimental section", "review": "This paper introduces a new framework for learning an interpretable representation of images and their attributes. The authors suggest decomposing the representation into a set of 'template' latent features, and a set of attribute-based features. The attribute-based features can be either 'free', i.e. discovered from the data, or 'fixed', i.e. based on the ground truth attributes. The authors encourage the decomposition of the latent space into the 'template' and the 'attributes' features by training a discriminator network to predict whether the attributes and the template features come from the same image or not.\n\nWhile the idea is interesting, the paper is lacking an experimental section, so the methodology is impossible to evaluate. Furthermore, while the authors spend many pages describing their methodology, the writing is often hard to follow, so I am still confused about the exact implementation of the attribute features \\phi(x, m) for example. The authors do point to the Appendix for their Experiments section, however this is not a good idea. The paper should be self-contained and the authors should not assume that their readers will read the information presented in the Appendix, which is always optional. \n\nUnfortunately, even the experimental section presented in the Appendix is not comprehensive enough to evaluate the proposed method. The authors train the model on a single dataset (MNIST), no baseline or ablation results are presented, and all the results are purely qualitative. Given that the ground truth attribute decomposition for MNIST is not known, even the qualitative results are impossible to evaluate. I recommend that the authors present quantitive results in the updated version of their paper (i.e. disentanglement metric scores, the log-likelihood of the reconstructions), including new experiments on a dataset like dSprites or CelebA, where the ground truth attributes are known.", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541979650862}, {"id": "HyxsBAGAhX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper42/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a generative network capable of generating variations of a given input, conditioned on an attribute. Earlier papers generated variations of the input in the presence of the attribute and this attribute was assumed to be known during training. This paper proposes to automatically discover these attribute and thus work to produce variations even in the absence of known attribute information.\n\nThe paper is dense, but it is well written. It has mixed ideas from several papers - the basic VAE architecture, combined with a discriminator and regularizations over latent space. The key thing, of course, is the design of the attribute function. There seems to be an interesting interaction between the encoder, discriminator and the attribute function that requires more investigation. This is acknowledged in the conclusion as well.\n\nThe work is original and the results on the MNIST dataset are very interesting. I think the significance of this work lies in the fact that this can be a starting point for several interesting future works in this direction.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Work", "review": "The paper proposes a generative network capable of generating variations of a given input, conditioned on an attribute. Earlier papers generated variations of the input in the presence of the attribute and this attribute was assumed to be known during training. This paper proposes to automatically discover these attribute and thus work to produce variations even in the absence of known attribute information.\n\nThe paper is dense, but it is well written. It has mixed ideas from several papers - the basic VAE architecture, combined with a discriminator and regularizations over latent space. The key thing, of course, is the design of the attribute function. There seems to be an interesting interaction between the encoder, discriminator and the attribute function that requires more investigation. This is acknowledged in the conclusion as well.\n\nThe work is original and the results on the MNIST dataset are very interesting. I think the significance of this work lies in the fact that this can be a starting point for several interesting future works in this direction.", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541447235035}, {"id": "HyxzkI2ThQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper42/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a generalization of variational auto-encoders to account for meta-data (attributes), learning new ones, in a way that these can be controlled to generate new samples. The model learns how to decouple the attributes in an adversarial way by means of a discriminator. The problem is interesting, but I found two main issues with this paper:\n1.- Lack of clarity: I found the paper difficult to follow, even after reading Sec. 2 and 3 several times.\n2.- Almost absence of experiments: The paper only has one experiment, which is in the appendix, and is about sampling using the MNIST dataset. Given that this paper proposes a model, whose properties can be assessed by means of experiments, the fact that there is nothing of the kind provides no support to any benefits the model may have.\n\nOther points:\nWhat in the model prevents the solution of z_* being just random (independently of x)?\n\nThis paper seems relevant Esser, Patrick, Ekaterina Sutter, and Björn Ommer. \"A Variational U-Net for Conditional Appearance and Shape Generation.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lack of clarity and almost no experiments", "review": "This paper proposes a generalization of variational auto-encoders to account for meta-data (attributes), learning new ones, in a way that these can be controlled to generate new samples. The model learns how to decouple the attributes in an adversarial way by means of a discriminator. The problem is interesting, but I found two main issues with this paper:\n1.- Lack of clarity: I found the paper difficult to follow, even after reading Sec. 2 and 3 several times.\n2.- Almost absence of experiments: The paper only has one experiment, which is in the appendix, and is about sampling using the MNIST dataset. Given that this paper proposes a model, whose properties can be assessed by means of experiments, the fact that there is nothing of the kind provides no support to any benefits the model may have.\n\nOther points:\nWhat in the model prevents the solution of z_* being just random (independently of x)?\n\nThis paper seems relevant Esser, Patrick, Ekaterina Sutter, and Björn Ommer. \"A Variational U-Net for Conditional Appearance and Shape Generation.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541420506462}], "openreview_url": "https://openreview.net/forum?id=ryfaViR9YX", "arxiv_id": "1901.03634", "paper_pdf": "papers/ryfaViR9YX.pdf", "paper_pdf_sha256": "0bea74c62322db9de1f98d183bade4b31a9f07521b051e2fa1ed140f92be8999", "paper_pdf_bytes": 2706922, "paper_pdf_source": "openreview", "code_url": "https://github.com/Ghadjeres/VarNet", "code_repository": "Ghadjeres/VarNet", "code_commit": "08a76075b833fbf71ddc9c73598ca9e7398a480d", "code_archive": "repos/ryfaViR9YX.zip", "code_archive_sha256": "9c89bc83f39bc86375180e3508f3cd3747cf17484ea093d7f08e52b124f1be22", "code_archive_bytes": 157988, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 151, "github_languages": {"Python": 181929}, "github_archived": false, "github_pushed_at": "2019-03-12T15:18:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variation-network-learning-high-level"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1fduCl0b", "year": 2018, "status": "rejected", "title": "Lifelong Generative Modeling", "authors": ["Jason Ramapuram", "Magda Gregorova", "Alexandros Kalousis"], "authorids": ["jason.ramapuram@etu.unige.ch", "magda.gregorova@unige.ch", "alexandros.kalousis@hesge.ch"], "authors_source": "OpenReview API", "abstract": "Lifelong learning is the problem of learning multiple consecutive tasks in a sequential manner where knowledge gained from previous tasks is retained and used for future learning. It is essential towards the development of intelligent machines that can adapt to their surroundings. In this work we focus on a lifelong learning approach to generative modeling where we continuously incorporate newly observed streaming distributions into our learnt model. We do so through a student-teacher architecture which allows us to learn and preserve all the distributions seen so far without the need to retain the past data nor the past models. Through the introduction of a novel cross-model regularizer, the student model leverages the information learnt by the teacher, which acts as a summary of everything seen till now. The regularizer has the additional benefit of reducing the effect of catastrophic interference that appears when we learn over streaming data. We demonstrate its efficacy on streaming distributions as well as its ability to learn a common latent representation across a complex transfer learning scenario.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJuF9Eqez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper457/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Second paragraph in Section 1: Nice motivation. I am not sure though whether the performed experiments are the most expressive for such motivation. For instance, is the experiment in Section 5.1 a common task in that sequential lifelong learning setting?\n\n- Section 4, which is the main technical section of the paper, is quite full of lengthy descriptions that are a bit equivocal. I reckon each claim really needs to be supported by a corresponding unequivocal mathermatical formulation.\n\n- An example of the last point can be found in Section 4.2: \"The synthetic samples need to be representative of all the previously observed distributions ...\": It will be much clearer how such samples are representative if a formulation follows, and that did not happen in Section 4.2.\n\n- \"1) Sampling the prior can select a point in the latent space that is in between two separate distributions ...\": I am not sure I got this drawback of using the standard form of VAEs. Could you please further elaborate on this?\n\n- \"we restrict the posterior representation of the student model to **be close to that of the teacher** for the previous distributions** accumulated by the teacher. This allows the model parameters to **vary as necessary** in order to best fit the data\": What if the previous distributions are not that close to the new one?\n\n- Distribution intervals: Will it be the case in reality that these intervals will be given? Otherwise, what are the solutions to that? Can they be estimated somehow (as a future work)?\n\n\nMinor:\n- \"we observe a sample X of K\": sample X of size K, I guess?\n- \"... form nor an efficient estimator Kingma (2017)\": citation style.\n- \"we illustrates ...\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good initiative for using VAEs in a new framework. The work needs to be a bit more principled though.", "rating": "4: Ok but not good enough - rejection", "review": "- Second paragraph in Section 1: Nice motivation. I am not sure though whether the performed experiments are the most expressive for such motivation. For instance, is the experiment in Section 5.1 a common task in that sequential lifelong learning setting?\n\n- Section 4, which is the main technical section of the paper, is quite full of lengthy descriptions that are a bit equivocal. I reckon each claim really needs to be supported by a corresponding unequivocal mathermatical formulation.\n\n- An example of the last point can be found in Section 4.2: \"The synthetic samples need to be representative of all the previously observed distributions ...\": It will be much clearer how such samples are representative if a formulation follows, and that did not happen in Section 4.2.\n\n- \"1) Sampling the prior can select a point in the latent space that is in between two separate distributions ...\": I am not sure I got this drawback of using the standard form of VAEs. Could you please further elaborate on this?\n\n- \"we restrict the posterior representation of the student model to **be close to that of the teacher** for the previous distributions** accumulated by the teacher. This allows the model parameters to **vary as necessary** in order to best fit the data\": What if the previous distributions are not that close to the new one?\n\n- Distribution intervals: Will it be the case in reality that these intervals will be given? Otherwise, what are the solutions to that? Can they be estimated somehow (as a future work)?\n\n\nMinor:\n- \"we observe a sample X of K\": sample X of size K, I guess?\n- \"... form nor an efficient estimator Kingma (2017)\": citation style.\n- \"we illustrates ...\"", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511832192072}, {"id": "rJuN3apyf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper457/AnonReviewer1"], "rating": "9: Top 15% of accepted papers, strong accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "We have seen numerous variants of variational autoencoders, most of them introducing delta changes to the original architecture to address the same sort of modeling problems. This paper attacks a different kind of problem, namely lifelong learning. This key aspect of the paper, besides the fact that it constitutes a very important problem, does also addes a strong element of freshness to the paper.\n\nThe construction of the generative model is correct, and commensurate with standard practice in the field of deep generative models. The derivations are correct, while the experimental evaluation is diverse and convincing. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "At last, a deep generative model addressing a fresh problem", "rating": "9: Top 15% of accepted papers, strong accept", "review": "We have seen numerous variants of variational autoencoders, most of them introducing delta changes to the original architecture to address the same sort of modeling problems. This paper attacks a different kind of problem, namely lifelong learning. This key aspect of the paper, besides the fact that it constitutes a very important problem, does also addes a strong element of freshness to the paper.\n\nThe construction of the generative model is correct, and commensurate with standard practice in the field of deep generative models. The derivations are correct, while the experimental evaluation is diverse and convincing. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511017519582}, {"id": "rJokGmjgG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper457/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposed a teacher-student framework and a modified objective function to adapt VAE training to streaming data setting. The qualitative experimental result shows that the learned model can generate reasonable-looking samples. I'm not sure about what conclusion to make from the numerical result, as the test negative ELBO actually increased after decreasing initially. Why did it increase?\n\nThe modified objective function is a little ad-hoc, and it's unclear how to relate the overall objective function to Bayesian posterior inference (what exactly is the posterior that the encoder tries to approximate?). There is a term in the objective function that is synthetic data specific. Does that imply that the objective function is different depending on if the data is synthetic or real? What is the motivation/justification of choosing KL(Q_student||Q_teacher) as regularisation instead of the other way around? Would that make a difference in the goodness of the learned model? If not, wouldn't KL(Q_teacher||Q_student) result reduction in the variance of gradients and therefore a better choice?\n\nDetails on the minimum number of real samples per interval for the model to be able to learn is also missing. Also, how many synthetic samples per real samples are needed? How is the update with respect to synthetic sample scheduled? Given infinite amount of streaming data with a fixed number of classes/underlying distributions and interval length, and sample the class of each interval (uniformly) randomly, will the model/algorithm converge? Is there a minimum number of real examples that the student learner needs to see before it can be turned into a teacher?\n\nOther question: How is the number of latent category J of the latent discrete distribution chosen?\n\nQuality: The numerical experiment doesn't really compare to any other streaming benchmark and is a little unsatisfying. Without a streaming benchmark or a realistic motivating example in which the proposed scheme makes a significant difference, it's difficult to judge the contribution of this work.\nClarity: The manuscript is reasonably well-written. (minor: Paragraph 2, section 5, 'in principle' instead of 'in principal')\nOriginality: Average. The student-teacher framework by itself isn't novel. The modifications to the objective function appears to be novel as far as I am aware, but it doesn't require much special insights.\nSignificance: Below average. I think it will be very helpful if the authors can include a realistic motivating example where lifelong unsupervised learning is critical, and demonstrate that the proposed scheme makes a difference in the example.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Adapted VAE training to streaming data setting. Experiment shows the learned model can generate reasonable-looking samples. Unsure about quantitative results.", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposed a teacher-student framework and a modified objective function to adapt VAE training to streaming data setting. The qualitative experimental result shows that the learned model can generate reasonable-looking samples. I'm not sure about what conclusion to make from the numerical result, as the test negative ELBO actually increased after decreasing initially. Why did it increase?\n\nThe modified objective function is a little ad-hoc, and it's unclear how to relate the overall objective function to Bayesian posterior inference (what exactly is the posterior that the encoder tries to approximate?). There is a term in the objective function that is synthetic data specific. Does that imply that the objective function is different depending on if the data is synthetic or real? What is the motivation/justification of choosing KL(Q_student||Q_teacher) as regularisation instead of the other way around? Would that make a difference in the goodness of the learned model? If not, wouldn't KL(Q_teacher||Q_student) result reduction in the variance of gradients and therefore a better choice?\n\nDetails on the minimum number of real samples per interval for the model to be able to learn is also missing. Also, how many synthetic samples per real samples are needed? How is the update with respect to synthetic sample scheduled? Given infinite amount of streaming data with a fixed number of classes/underlying distributions and interval length, and sample the class of each interval (uniformly) randomly, will the model/algorithm converge? Is there a minimum number of real examples that the student learner needs to see before it can be turned into a teacher?\n\nOther question: How is the number of latent category J of the latent discrete distribution chosen?\n\nQuality: The numerical experiment doesn't really compare to any other streaming benchmark and is a little unsatisfying. Without a streaming benchmark or a realistic motivating example in which the proposed scheme makes a significant difference, it's difficult to judge the contribution of this work.\nClarity: The manuscript is reasonably well-written. (minor: Paragraph 2, section 5, 'in principle' instead of 'in principal')\nOriginality: Average. The student-teacher framework by itself isn't novel. The modifications to the objective function appears to be novel as far as I am aware, but it doesn't require much special insights.\nSignificance: Below average. I think it will be very helpful if the authors can include a realistic motivating example where lifelong unsupervised learning is critical, and demonstrate that the proposed scheme makes a difference in the example.\n\n\n", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1511891426624}], "openreview_url": "https://openreview.net/forum?id=S1fduCl0b", "arxiv_id": "1705.09847", "paper_pdf": "papers/S1fduCl0b.pdf", "paper_pdf_sha256": "d0c222940124836a05fec8fabc888bdb3a74bd363bff16fdfad1e0cd66dda26a", "paper_pdf_bytes": 3621573, "paper_pdf_source": "openreview", "code_url": "https://github.com/jramapuram/LifelongVAE_pytorch", "code_repository": "jramapuram/LifelongVAE_pytorch", "code_commit": "0199cb3c9659c4e22f20e1cd4af8cf1c8f8a1e8d", "code_archive": "repos/S1fduCl0b.zip", "code_archive_sha256": "cb80cc0aa709616d74b712fb1b21f0eec7f1f9fad1f33d3a1427c0b745a0c55d", "code_archive_bytes": 36322, "code_file_count": 14, "code_extensions": {".py": 13, ".sh": 1}, "github_disk_usage_kb": 133, "github_languages": {"Python": 107264, "Shell": 1251}, "github_archived": false, "github_pushed_at": "2019-12-03T23:44:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lifelong-generative-modeling"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "skIx8CgP8V", "year": 2026, "status": "rejected", "title": "Noise Augmented Fine Tuning for Mitigating Hallucinations in Large Language Models", "authors": ["Afshin Khadangi", "Amir Sartipi", "Igor Tchappi", "Ramin Bahmani"], "authorids": ["~Afshin_Khadangi1", "~Amir_Sartipi2", "~Igor_Tchappi1", "~Ramin_Bahmani1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) often produce inaccurate or misleading content—hallucinations. To address this challenge, we introduce Noise-Augmented Fine-Tuning (NoiseFiT), a novel framework that leverages adaptive noise injection based on the signal-to-noise ratio (SNR) to enhance model robustness. Our contribution is threefold. First, NoiseFiT selectively perturbs layers identified as either high-SNR (more robust) or low-SNR (potentially under-regularized) using a dynamically scaled Gaussian noise. Second, we further propose a hybrid loss that combines standard cross-entropy, soft cross-entropy, and consistency regularization to ensure stable and accurate outputs under noisy training conditions. Third, a theoretical analysis proposed shows that adaptive noise injection is both unbiased and variance-preserving, providing strong guarantees for convergence in expectation. Moreover, empirical results on multiple test and benchmark datasets demonstrate that NoiseFiT significantly reduces hallucination rates, often improving or matching baseline performance in key tasks. These findings highlight the promise of noise-driven strategies for achieving robust, trustworthy language modeling without incurring prohibitive computational overhead. We have publicly released the fine-tuning logs, benchmark evaluation artifacts, and source code online at W&B, Hugging Face, and GitHub, respectively, to foster further research, accessibility and reproducibility.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "UJAShYTJIR", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16652/Reviewer_FPbm"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes Noise-Augmented Fine-Tuning, a method that improves the robustness of large language models by injecting controlled noise during fine-tuning. The approach aims to enhance generalization—especially on out-of-distribution or noisy inputs—without hurting performance on clean data. Results show consistent gains over standard fine-tuning, particularly in challenging categories like geography and history, while maintaining strong performance on in-distribution tasks. The method offers a simple, plug-and-play way to make fine-tuned models more reliable in real-world conditions.", "review_text": "This paper proposes Noise-Augmented Fine-Tuning, a method that improves the robustness of large language models by injecting controlled noise during fine-tuning. The approach aims to enhance generalization—especially on out-of-distribution or noisy inputs—without hurting performance on clean data. Results show consistent gains over standard fine-tuning, particularly in challenging categories like geography and history, while maintaining strong performance on in-distribution tasks. The method offers a simple, plug-and-play way to make fine-tuned models more reliable in real-world conditions.", "strengths": "1. The proposed NoiseFit method and the SNR-based layer selection approach are relatively new and not investigated before.\n2. Experiments on various datasets and base models demonstrate the effectiveness of the proposed method. \n3. The mathematical analysis in this paper is rigorous, well-structured, and provides strong theoretical support for the proposed method.\n4. The paper is well-written and easy to follow.", "weaknesses": "While the authors honestly acknowledge several limitations of their work, recognizing these issues does not, by itself, mitigate their impact on the method's applicability or validity.\n\n1. The current training set is too small. Experiments on larger scale datasets are required to demonstrate the effectiveness and generalization capability of the proposed method.\n2. Although the paper compares its method to BaseFiT, it lacks a systematic evaluation against other advanced strategies designed to mitigate hallucinations, such as those mentioned in the related works section.\n3. The proposed method seems a general finetuning method for robustness. The paper lacks a clear analysis linking the proposed noise-augmented fine-tuning to hallucination reduction.\n4. The proposed method is sensitive to hyperparameters. In some setting it underperforms BaseFit. \n\nOverall, the paper is good and I subjectively believe that the method is effective. So I will give a positive rating. If the authors can address my concerns, I will further raise my rating.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Noise-Augmented Fine-Tuning, a method that improves the robustness of large language models by injecting controlled noise during fine-tuning. The approach aims to enhance generalization—especially on out-of-distribution or noisy inputs—without hurting performance on clean data. Results show consistent gains over standard fine-tuning, particularly in challenging categories like geography and history, while maintaining strong performance on in-distribution tasks. The method offers a simple, plug-and-play way to make fine-tuned models more reliable in real-world conditions.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The proposed NoiseFit method and the SNR-based layer selection approach are relatively new and not investigated before.\n2. Experiments on various datasets and base models demonstrate the effectiveness of the proposed method. \n3. The mathematical analysis in this paper is rigorous, well-structured, and provides strong theoretical support for the proposed method.\n4. The paper is well-written and easy to follow.", "weaknesses": "While the authors honestly acknowledge several limitations of their work, recognizing these issues does not, by itself, mitigate their impact on the method's applicability or validity.\n\n1. The current training set is too small. Experiments on larger scale datasets are required to demonstrate the effectiveness and generalization capability of the proposed method.\n2. Although the paper compares its method to BaseFiT, it lacks a systematic evaluation against other advanced strategies designed to mitigate hallucinations, such as those mentioned in the related works section.\n3. The proposed method seems a general finetuning method for robustness. The paper lacks a clear analysis linking the proposed noise-augmented fine-tuning to hallucination reduction.\n4. The proposed method is sensitive to hyperparameters. In some setting it underperforms BaseFit. \n\nOverall, the paper is good and I subjectively believe that the method is effective. So I will give a positive rating. If the authors can address my concerns, I will further raise my rating.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762326051166}, {"id": "1RITM5gJvd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16652/Reviewer_X9Yd"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes NoiseFiT, a framework designed to reduce hallucinations in LLMs during fine-tuning. NoiseFiT selectively injects adaptive Gaussian noise into high-SNR or low-SNR transformer layers. Noise scaling incorporates hidden-state median/MAD statistics and model uncertainty. A hybrid loss is calculated blending clean cross-entropy, soft cross-entropy via temperature-scaled teacher logits, and consistency regularization across two noisy passes. Experiments across model families show reduced hallucinations and improved or preserved benchmark performance relative to BaseFiT fine-tuning.", "review_text": "This paper proposes NoiseFiT, a framework designed to reduce hallucinations in LLMs during fine-tuning. NoiseFiT selectively injects adaptive Gaussian noise into high-SNR or low-SNR transformer layers. Noise scaling incorporates hidden-state median/MAD statistics and model uncertainty. A hybrid loss is calculated blending clean cross-entropy, soft cross-entropy via temperature-scaled teacher logits, and consistency regularization across two noisy passes. Experiments across model families show reduced hallucinations and improved or preserved benchmark performance relative to BaseFiT fine-tuning.", "strengths": "- The SNR-guided layer selection is a well-motivated heuristic. This matters for efficiency, as it avoids perturbing the entire model.\n- This paper tests its method across a diverse set of model architectures (LLaMA, Qwen, Gemma, Mistral) and sizes. This supports the generality of the approach.", "weaknesses": "- The custom fine-tuning dataset is too small (832 samples) and synthetically generated by GROK. This raises concerns about whether NoiseFiT is simply mimicing the generated few samples and whether the gains would transfer to larger, more diverse, human-curated fine-tuning datasets. The custom test set is evaluated by GROK. Using an LLM to evaluate LLM hallucinations, especially when the training and test data came from the same LLM judge, would introduce significant risks of confounding variables and evaluation bias.\n- The empirical results of NoiseFiT are exclusively compared against BaseFiT, which is the standard fine-tuning. In related work the authors discusses hallucination mitigation techniques like RAG, RLHF, and self-consistency, but did not compare NoiseFiT against them. Though NoiseFiT and those post-training techniques are claimed to be complementary, it should be proved by some extra ablation studies. Besides, many training-time hallucination allevitation methods are not discussed [1, 2].\n\n[1] Hallucination Detection and Hallucination Mitigation: An Investigatcion\n\n[2] Logit Space Constrained Fine-Tuning for Mitigating Hallucinations in LLM-Based Recommender Systems", "questions": "- What is the total fine-tuning time (wall-clock) per model or per experiment? The paper notes that NoiseFiT adds a negligible training-time cost beyond a second (noisy) forward pass. However, the full method requires one clean pass and two independent noisy passes ($L_{SOFT}$ and $L_{CONSISTENCY}$), implying a ~3x increase in forward-pass computation. This discrepancy should be clarified.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes NoiseFiT, a framework designed to reduce hallucinations in LLMs during fine-tuning. NoiseFiT selectively injects adaptive Gaussian noise into high-SNR or low-SNR transformer layers. Noise scaling incorporates hidden-state median/MAD statistics and model uncertainty. A hybrid loss is calculated blending clean cross-entropy, soft cross-entropy via temperature-scaled teacher logits, and consistency regularization across two noisy passes. Experiments across model families show reduced hallucinations and improved or preserved benchmark performance relative to BaseFiT fine-tuning.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The SNR-guided layer selection is a well-motivated heuristic. This matters for efficiency, as it avoids perturbing the entire model.\n- This paper tests its method across a diverse set of model architectures (LLaMA, Qwen, Gemma, Mistral) and sizes. This supports the generality of the approach.", "weaknesses": "- The custom fine-tuning dataset is too small (832 samples) and synthetically generated by GROK. This raises concerns about whether NoiseFiT is simply mimicing the generated few samples and whether the gains would transfer to larger, more diverse, human-curated fine-tuning datasets. The custom test set is evaluated by GROK. Using an LLM to evaluate LLM hallucinations, especially when the training and test data came from the same LLM judge, would introduce significant risks of confounding variables and evaluation bias.\n- The empirical results of NoiseFiT are exclusively compared against BaseFiT, which is the standard fine-tuning. In related work the authors discusses hallucination mitigation techniques like RAG, RLHF, and self-consistency, but did not compare NoiseFiT against them. Though NoiseFiT and those post-training techniques are claimed to be complementary, it should be proved by some extra ablation studies. Besides, many training-time hallucination allevitation methods are not discussed [1, 2].\n\n[1] Hallucination Detection and Hallucination Mitigation: An Investigatcion\n\n[2] Logit Space Constrained Fine-Tuning for Mitigating Hallucinations in LLM-Based Recommender Systems", "questions": "- What is the total fine-tuning time (wall-clock) per model or per experiment? The paper notes that NoiseFiT adds a negligible training-time cost beyond a second (noisy) forward pass. However, the full method requires one clean pass and two independent noisy passes ($L_{SOFT}$ and $L_{CONSISTENCY}$), implying a ~3x increase in forward-pass computation. This discrepancy should be clarified.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762289301799}, {"id": "EkMBwQDsjr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16652/Reviewer_ouvi"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes Noise-Augmented Fine-Tuning (NoiseFiT), a framework for reducing hallucinations in large language models. The approach selectively injects adaptive Gaussian noise into transformer layers during fine-tuning based on signal-to-noise ratios (SNR). The method combines (1) SNR-based layer selection identifying either high-SNR or low-SNR layers, (2) adaptive noise scaling using robust statistics and model uncertainty, and (3) a hybrid loss combining cross-entropy, soft cross-entropy (knowledge distillation), and consistency regularization. Experiments across LLaMA, Qwen, Gemma, and Mistral models on multiple benchmarks show modest improvements in hallucination reduction", "review_text": "This paper proposes Noise-Augmented Fine-Tuning (NoiseFiT), a framework for reducing hallucinations in large language models. The approach selectively injects adaptive Gaussian noise into transformer layers during fine-tuning based on signal-to-noise ratios (SNR). The method combines (1) SNR-based layer selection identifying either high-SNR or low-SNR layers, (2) adaptive noise scaling using robust statistics and model uncertainty, and (3) a hybrid loss combining cross-entropy, soft cross-entropy (knowledge distillation), and consistency regularization. Experiments across LLaMA, Qwen, Gemma, and Mistral models on multiple benchmarks show modest improvements in hallucination reduction", "strengths": "- Addresses the important problem of hallucinations in LLMs\n- Provides theoretical analysis of noise injection properties\n- Comprehensive experimental evaluation across multiple model families (LLaMA, Qwen, Gemma, Mistral) and benchmarks (GPQA, MUSR, IFEval, BBH, MATH, MMLU-Pro, HaluEval, TruthfulQA) with a lot of supplementary material", "weaknesses": "- Modest and inconsistent improvements: many gains in Table 1 are small, BaseFiT sometimes wins, best Mistral result shows only 4.74 point improvement on HaluEval (47.60→52.34). Results are highly variable across configurations. \n- 832 synthetic training samples from GROK 3.0, test evaluation also uses LLM judge, custom test set limited to 208 prompts. Raises serious concerns about generalizability and validity\n-  High hyperparameter sensitivity is a known issue, and the authors admit that \"per-task tuning\" is encouraged, which severely limits the practical utility of the model. The \"recipe\" in Appendix C is a heuristic approach.", "questions": "- Why not include direct comparisons with other methods of dealing with hallucinations mentioned in the relevant work? \n- How does the method perform on standard, non-synthetic fine-tuning datasets (e.g., real instruction-following data)? The 832 synthetic samples raise serious generalizability concerns\n- Why do certain categories (at Tables E2-E6) show degradation with fine-tuning, and does NoiseFiT exacerbate or mitigate this?\n- Can you provide human evaluation on a substantial sample to validate the LLM judge assessments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Noise-Augmented Fine-Tuning (NoiseFiT), a framework for reducing hallucinations in large language models. The approach selectively injects adaptive Gaussian noise into transformer layers during fine-tuning based on signal-to-noise ratios (SNR). The method combines (1) SNR-based layer selection identifying either high-SNR or low-SNR layers, (2) adaptive noise scaling using robust statistics and model uncertainty, and (3) a hybrid loss combining cross-entropy, soft cross-entropy (knowledge distillation), and consistency regularization. Experiments across LLaMA, Qwen, Gemma, and Mistral models on multiple benchmarks show modest improvements in hallucination reduction", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Addresses the important problem of hallucinations in LLMs\n- Provides theoretical analysis of noise injection properties\n- Comprehensive experimental evaluation across multiple model families (LLaMA, Qwen, Gemma, Mistral) and benchmarks (GPQA, MUSR, IFEval, BBH, MATH, MMLU-Pro, HaluEval, TruthfulQA) with a lot of supplementary material", "weaknesses": "- Modest and inconsistent improvements: many gains in Table 1 are small, BaseFiT sometimes wins, best Mistral result shows only 4.74 point improvement on HaluEval (47.60→52.34). Results are highly variable across configurations. \n- 832 synthetic training samples from GROK 3.0, test evaluation also uses LLM judge, custom test set limited to 208 prompts. Raises serious concerns about generalizability and validity\n-  High hyperparameter sensitivity is a known issue, and the authors admit that \"per-task tuning\" is encouraged, which severely limits the practical utility of the model. The \"recipe\" in Appendix C is a heuristic approach.", "questions": "- Why not include direct comparisons with other methods of dealing with hallucinations mentioned in the relevant work? \n- How does the method perform on standard, non-synthetic fine-tuning datasets (e.g., real instruction-following data)? The 832 synthetic samples raise serious generalizability concerns\n- Why do certain categories (at Tables E2-E6) show degradation with fine-tuning, and does NoiseFiT exacerbate or mitigate this?\n- Can you provide human evaluation on a substantial sample to validate the LLM judge assessments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762115297882}, {"id": "vbceBKAoim", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16652/Reviewer_jnC9"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a fine-tuning framework called NoiseFiT, which enhances model robustness and reduces hallucinations by injecting adaptive Gaussian noise into specific Transformer layers during fine-tuning and employing a hybrid loss function (cross-entropy loss + soft-target loss + consistency loss). Experimental results across multiple tasks and models demonstrate the effectiveness of the method.", "review_text": "This paper proposes a fine-tuning framework called NoiseFiT, which enhances model robustness and reduces hallucinations by injecting adaptive Gaussian noise into specific Transformer layers during fine-tuning and employing a hybrid loss function (cross-entropy loss + soft-target loss + consistency loss). Experimental results across multiple tasks and models demonstrate the effectiveness of the method.", "strengths": "1. The method combines multiple regularization techniques (soft-target loss and consistency loss) to form a unified training objective.\n\n2. It has been extensively validated across multiple model families and tasks, producing convincing results.", "weaknesses": "1. The idea of merging multiple losses is not novel, but it appears to be the main innovation of the paper.\n\n2. Limited dataset size and diversity: The training set contains only 832 samples and consists of synthetic data. Although the authors emphasize it as “simple and effective,” this may still affect the generalization ability of the method.\n\n3. Insufficient comparison with other methods: The paper mainly compares with BaseFiT (fine-tuning without noise) and lacks thorough comparison with other advanced hallucination mitigation methods such as RAG, RLHF, and Self-Consistency.\n\n4. Hyperparameter sensitivity: The paper mentions that noise intensity, layer selection, and other hyperparameters significantly affect results and need to be adjusted for different models.\n\n5. Hallucination evaluation depends on external models: Some hallucination assessments use GROK 3.0 as the “judge,” which may introduce evaluation bias.\n\n6. Theoretical analysis is rich but somewhat lengthy: The appendix contains extensive theoretical content, some of which is not closely related to the core method.", "questions": "1. The idea of merging multiple losses is not novel, but it appears to be the main innovation of the paper.\n\n2. Limited dataset size and diversity: The training set contains only 832 samples and consists of synthetic data. Although the authors emphasize it as “simple and effective,” this may still affect the generalization ability of the method.\n\n3. Insufficient comparison with other methods: The paper mainly compares with BaseFiT (fine-tuning without noise) and lacks thorough comparison with other advanced hallucination mitigation methods such as RAG, RLHF, and Self-Consistency.\n\n4. Hyperparameter sensitivity: The paper mentions that noise intensity, layer selection, and other hyperparameters significantly affect results and need to be adjusted for different models.\n\n5. Hallucination evaluation depends on external models: Some hallucination assessments use GROK 3.0 as the “judge,” which may introduce evaluation bias.\n\n6. Theoretical analysis is rich but somewhat lengthy: The appendix contains extensive theoretical content, some of which is not closely related to the core method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a fine-tuning framework called NoiseFiT, which enhances model robustness and reduces hallucinations by injecting adaptive Gaussian noise into specific Transformer layers during fine-tuning and employing a hybrid loss function (cross-entropy loss + soft-target loss + consistency loss). Experimental results across multiple tasks and models demonstrate the effectiveness of the method.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The method combines multiple regularization techniques (soft-target loss and consistency loss) to form a unified training objective.\n\n2. It has been extensively validated across multiple model families and tasks, producing convincing results.", "weaknesses": "1. The idea of merging multiple losses is not novel, but it appears to be the main innovation of the paper.\n\n2. Limited dataset size and diversity: The training set contains only 832 samples and consists of synthetic data. Although the authors emphasize it as “simple and effective,” this may still affect the generalization ability of the method.\n\n3. Insufficient comparison with other methods: The paper mainly compares with BaseFiT (fine-tuning without noise) and lacks thorough comparison with other advanced hallucination mitigation methods such as RAG, RLHF, and Self-Consistency.\n\n4. Hyperparameter sensitivity: The paper mentions that noise intensity, layer selection, and other hyperparameters significantly affect results and need to be adjusted for different models.\n\n5. Hallucination evaluation depends on external models: Some hallucination assessments use GROK 3.0 as the “judge,” which may introduce evaluation bias.\n\n6. Theoretical analysis is rich but somewhat lengthy: The appendix contains extensive theoretical content, some of which is not closely related to the core method.", "questions": "1. The idea of merging multiple losses is not novel, but it appears to be the main innovation of the paper.\n\n2. Limited dataset size and diversity: The training set contains only 832 samples and consists of synthetic data. Although the authors emphasize it as “simple and effective,” this may still affect the generalization ability of the method.\n\n3. Insufficient comparison with other methods: The paper mainly compares with BaseFiT (fine-tuning without noise) and lacks thorough comparison with other advanced hallucination mitigation methods such as RAG, RLHF, and Self-Consistency.\n\n4. Hyperparameter sensitivity: The paper mentions that noise intensity, layer selection, and other hyperparameters significantly affect results and need to be adjusted for different models.\n\n5. Hallucination evaluation depends on external models: Some hallucination assessments use GROK 3.0 as the “judge,” which may introduce evaluation bias.\n\n6. Theoretical analysis is rich but somewhat lengthy: The appendix contains extensive theoretical content, some of which is not closely related to the core method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760509939853}], "openreview_url": "https://openreview.net/forum?id=skIx8CgP8V", "arxiv_id": "2504.03302", "paper_pdf": "papers/skIx8CgP8V.pdf", "paper_pdf_sha256": "7eb8e2c4e8aeb3db330e4cd55761efe02500825b10e4f854b57ca1e78703e775", "paper_pdf_bytes": 23393123, "paper_pdf_source": "openreview", "code_url": "https://github.com/akhadangi/NoiseFiT", "code_repository": "akhadangi/NoiseFiT", "code_commit": "48c61c88a1a9bb808a8091276473d7770ea33ef9", "code_archive": "repos/skIx8CgP8V.zip", "code_archive_sha256": "5be9d9e9258bbf4839be9b0cd8a7a0b1224caa648f685ffccaa49d807e388b70", "code_archive_bytes": 61203, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 108, "github_languages": {"Python": 135894}, "github_archived": false, "github_pushed_at": "2026-05-27T18:21:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/noise-augmented-fine-tuning-for-mitigating"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ptTt8mhS7n", "year": 2025, "status": "rejected", "title": "In-Context Transfer Learning: Demonstration Synthesis by Transferring Similar Tasks", "authors": ["Dingzirui Wang", "Xuanliang Zhang", "Qiguang Chen", "Longxu Dou", "Xiao Xu", "Rongyu Cao", "YINGWEI MA", "Qingfu Zhu", "Wanxiang Che", "Binhua Li", "Fei Huang", "Yongbin Li"], "authorids": ["~Dingzirui_Wang1", "~Xuanliang_Zhang1", "~Qiguang_Chen1", "~Longxu_Dou1", "~Xiao_Xu1", "~Rongyu_Cao2", "~YINGWEI_MA2", "~Qingfu_Zhu1", "~Wanxiang_Che1", "~Binhua_Li1", "~Fei_Huang2", "~Yongbin_Li2"], "authors_source": "OpenReview API", "abstract": "In-context learning (ICL) is an effective approach to help large language models (LLMs) adapt to various tasks by providing demonstrations of the target task. Considering the high cost of labeling demonstrations, many methods propose synthesizing demonstrations from scratch using LLMs. However, the quality of the demonstrations synthesized from scratch is limited by the capabilities and knowledge of LLMs. To address this, inspired by transfer learning, we propose In-Context Transfer Learning (ICTL), which synthesizes target task demonstrations by transferring labeled demonstrations from similar source tasks. ICTL consists of two steps: source sampling and target transfer. First, we define an optimization objective, which minimizes transfer error to sample source demonstrations similar to the target task.  Then, we employ LLMs to transfer the sampled source demonstrations to match the definition and format of the target task. Experiments on Super-NI show that ICTL outperforms synthesis from scratch by 2.0% on average, demonstrating the effectiveness of our method.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "w7zi3fwWKI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6060/Reviewer_H8mW"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces In-Context Transfer Learning (ICTL), a method designed to improve demonstration synthesis for in-context learning (ICL) by leveraging labeled examples from similar tasks. \nUnlike conventional approaches that generate task demonstrations from scratch using LLMs, ICTL employs a transfer learning-inspired process consisting of two steps: Source sampling and target transfer. \nIn the source sampling phase, relevant source tasks are filtered based on task definitions, and demonstrations are subsequently selected using a simulated annealing approach that minimizes transfer error.\nDuring the target transfer phase, LLMs are used to adapt these selected demonstrations to synthesize examples suitable for the target task.\nExperimental results on the Super-NI dataset indicate that ICTL outperforms traditional demonstration synthesis methods, effectively addressing the limitations of LLMs in producing high-quality demonstrations from scratch.", "review_text": "The paper introduces In-Context Transfer Learning (ICTL), a method designed to improve demonstration synthesis for in-context learning (ICL) by leveraging labeled examples from similar tasks. \nUnlike conventional approaches that generate task demonstrations from scratch using LLMs, ICTL employs a transfer learning-inspired process consisting of two steps: Source sampling and target transfer. \nIn the source sampling phase, relevant source tasks are filtered based on task definitions, and demonstrations are subsequently selected using a simulated annealing approach that minimizes transfer error.\nDuring the target transfer phase, LLMs are used to adapt these selected demonstrations to synthesize examples suitable for the target task.\nExperimental results on the Super-NI dataset indicate that ICTL outperforms traditional demonstration synthesis methods, effectively addressing the limitations of LLMs in producing high-quality demonstrations from scratch.", "strengths": "1. The synthesis of demonstrations from similar tasks is an interesting and novel approach to enhancing the quality of in-context learning.\n1. The paper provides an optimization objective for sampling demonstrations, aimed at minimizing the task error bound.\n1. The experimental results demonstrate improvements over traditional direct demonstration synthesis methods, showing the effectiveness of the proposed approach.", "weaknesses": "1. The proof of Equation 3 (Appendix A) is not clearly explained. At the end of the proof (lines 868-869), the statement \"by substituting Theorem 1 into Equation 5, we can derive Equation 3\" skips too many intermediate steps, which makes the derivation ambiguous. While Theorem 1 and Equation 5 are related to finding the target task $\\hat{\\mu}_T$, Equation 3 is focused on determining the source task $\\hat{\\mu}_S$. This discrepancy is confusing and should be clarified further to convincingly demonstrate that Equation 3 is a reasonable objective.\n\n1. The proposed approach is primarily applicable to benchmarks that provide explicit task definitions, which restricts its applicability to other benchmarks. This limitation confines the evaluation to the Super-NI dataset, and it would be beneficial to explore how this method could generalize to other settings.\n\n1. The sampling process involving simulated annealing lacks sufficient clarity. Given that simulated annealing typically involves multiple iterations, it raises concerns about efficiency. Further details on how the sampling algorithm is implemented, and any measures taken to ensure efficiency, would be helpful.\n\nOther Issues:\n1. Rouge -> RougeL", "questions": "1. In Equation 2, $\\mu$ represents a data distribution of a task, while $x$ is a representation vector of a task definition. Since these two elements are in different spaces, how is the Wasserstein distance between them computed?\n\n1. Lines 206-207 mention that task selection is ranked by the Wasserstein distance between the embedding vectors of task definitions. How is the Wasserstein distance calculated between two vectors? Typically, the Wasserstein distance is used to compare distributions rather than individual points.\n\n1. How does the proposed source task filtering compare to retrieval-based approaches, such as those using Contriever or BERT?\n\n1. What are the \"perturbations\" used in simulated annealing during the sampling of source demonstrations?\n\n1. What kind of score function is employed in the simulated annealing process?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces In-Context Transfer Learning (ICTL), a method designed to improve demonstration synthesis for in-context learning (ICL) by leveraging labeled examples from similar tasks. \nUnlike conventional approaches that generate task demonstrations from scratch using LLMs, ICTL employs a transfer learning-inspired process consisting of two steps: Source sampling and target transfer. \nIn the source sampling phase, relevant source tasks are filtered based on task definitions, and demonstrations are subsequently selected using a simulated annealing approach that minimizes transfer error.\nDuring the target transfer phase, LLMs are used to adapt these selected demonstrations to synthesize examples suitable for the target task.\nExperimental results on the Super-NI dataset indicate that ICTL outperforms traditional demonstration synthesis methods, effectively addressing the limitations of LLMs in producing high-quality demonstrations from scratch.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The synthesis of demonstrations from similar tasks is an interesting and novel approach to enhancing the quality of in-context learning.\n1. The paper provides an optimization objective for sampling demonstrations, aimed at minimizing the task error bound.\n1. The experimental results demonstrate improvements over traditional direct demonstration synthesis methods, showing the effectiveness of the proposed approach.", "weaknesses": "1. The proof of Equation 3 (Appendix A) is not clearly explained. At the end of the proof (lines 868-869), the statement \"by substituting Theorem 1 into Equation 5, we can derive Equation 3\" skips too many intermediate steps, which makes the derivation ambiguous. While Theorem 1 and Equation 5 are related to finding the target task $\\hat{\\mu}_T$, Equation 3 is focused on determining the source task $\\hat{\\mu}_S$. This discrepancy is confusing and should be clarified further to convincingly demonstrate that Equation 3 is a reasonable objective.\n\n1. The proposed approach is primarily applicable to benchmarks that provide explicit task definitions, which restricts its applicability to other benchmarks. This limitation confines the evaluation to the Super-NI dataset, and it would be beneficial to explore how this method could generalize to other settings.\n\n1. The sampling process involving simulated annealing lacks sufficient clarity. Given that simulated annealing typically involves multiple iterations, it raises concerns about efficiency. Further details on how the sampling algorithm is implemented, and any measures taken to ensure efficiency, would be helpful.\n\nOther Issues:\n1. Rouge -> RougeL", "questions": "1. In Equation 2, $\\mu$ represents a data distribution of a task, while $x$ is a representation vector of a task definition. Since these two elements are in different spaces, how is the Wasserstein distance between them computed?\n\n1. Lines 206-207 mention that task selection is ranked by the Wasserstein distance between the embedding vectors of task definitions. How is the Wasserstein distance calculated between two vectors? Typically, the Wasserstein distance is used to compare distributions rather than individual points.\n\n1. How does the proposed source task filtering compare to retrieval-based approaches, such as those using Contriever or BERT?\n\n1. What are the \"perturbations\" used in simulated annealing during the sampling of source demonstrations?\n\n1. What kind of score function is employed in the simulated annealing process?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731154682236}, {"id": "UCP2cRMTwp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6060/Reviewer_6kCu"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes In-Context Transfer Learning (ICTL), a framework to enhance demonstration synthesis for in-context learning (ICL) by transferring labeled demonstrations from similar tasks. The motivation is to overcome the limitations of synthesizing demonstrations from scratch using large language models (LLMs). The framework aims to address cost and efficiency issues in generating high-quality task demonstrations by leveraging transfer learning. ICTL has two steps: (1) source sampling: select source tasks similar to the target task by minimizing transfer error. (2) target transfer: transfer these selected demonstrations to the target task’s format using LLMs (transferring, verify, sample). Experiments on the Super-NI dataset show that ICTL achieves 2.0% improvements on average compared to synthesizing demonstrations from scratch.", "review_text": "This paper proposes In-Context Transfer Learning (ICTL), a framework to enhance demonstration synthesis for in-context learning (ICL) by transferring labeled demonstrations from similar tasks. The motivation is to overcome the limitations of synthesizing demonstrations from scratch using large language models (LLMs). The framework aims to address cost and efficiency issues in generating high-quality task demonstrations by leveraging transfer learning. ICTL has two steps: (1) source sampling: select source tasks similar to the target task by minimizing transfer error. (2) target transfer: transfer these selected demonstrations to the target task’s format using LLMs (transferring, verify, sample). Experiments on the Super-NI dataset show that ICTL achieves 2.0% improvements on average compared to synthesizing demonstrations from scratch.", "strengths": "1. The paper is well written and easy to understand. It tackles an important problem of enhancing the generalibility of LLMs.\n2. The proposed in-context transfer learning is simple yet produce reasonable improvements over generating demonstrations from scratch. The ablation studies and sensitivity analysis of paramters are beneficial. Evaluations on the Super-NI demonstrate the method’s cross-task performance. The paper also provides a code release for reproducibility, which is beneficial to the community.", "weaknesses": "I have two major concerns:\n\n1. The performance improvements seem marginal. The 2% performance improvement over baseline method (generating target demonstartions from scratch) is marginal and may not justify the added complexity of ICTL. The procedure of sampling, transferring, verifying, and re-sampling introduces computational overhead. It would be great to include the computation cost comparison between different approaches in Table 1. Moreover, from the analysis of different parameters in Figure 3, seems the performance can drop signficiantly (more than 2%) if the parameters are not selected properly. This let me suspect the effectiveness of the proposed framework in more general real-world scenarios. It would be interesting to see more results on other tasks. \n\n2. The proposed framework follows the standard transfer learning pipeline. The theoretical analysis is also straight-forward extension. In general, I feel the technical contribution is limited while there are some specific designs for LLM domain.", "questions": "While the paper provides an interesting application of transfer learning to in-context learning, I have the concern of the incremental improvement and relatively complex framework. The contribution is not sufficiently novel, and the results are a bit underwhelming. I would be happy if the authors can provide more compelling evidence of ICTL’s practical benefits (e.g., compare the computational cost in Table 1, provide additional evaluation on other tasks).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes In-Context Transfer Learning (ICTL), a framework to enhance demonstration synthesis for in-context learning (ICL) by transferring labeled demonstrations from similar tasks. The motivation is to overcome the limitations of synthesizing demonstrations from scratch using large language models (LLMs). The framework aims to address cost and efficiency issues in generating high-quality task demonstrations by leveraging transfer learning. ICTL has two steps: (1) source sampling: select source tasks similar to the target task by minimizing transfer error. (2) target transfer: transfer these selected demonstrations to the target task’s format using LLMs (transferring, verify, sample). Experiments on the Super-NI dataset show that ICTL achieves 2.0% improvements on average compared to synthesizing demonstrations from scratch.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well written and easy to understand. It tackles an important problem of enhancing the generalibility of LLMs.\n2. The proposed in-context transfer learning is simple yet produce reasonable improvements over generating demonstrations from scratch. The ablation studies and sensitivity analysis of paramters are beneficial. Evaluations on the Super-NI demonstrate the method’s cross-task performance. The paper also provides a code release for reproducibility, which is beneficial to the community.", "weaknesses": "I have two major concerns:\n\n1. The performance improvements seem marginal. The 2% performance improvement over baseline method (generating target demonstartions from scratch) is marginal and may not justify the added complexity of ICTL. The procedure of sampling, transferring, verifying, and re-sampling introduces computational overhead. It would be great to include the computation cost comparison between different approaches in Table 1. Moreover, from the analysis of different parameters in Figure 3, seems the performance can drop signficiantly (more than 2%) if the parameters are not selected properly. This let me suspect the effectiveness of the proposed framework in more general real-world scenarios. It would be interesting to see more results on other tasks. \n\n2. The proposed framework follows the standard transfer learning pipeline. The theoretical analysis is also straight-forward extension. In general, I feel the technical contribution is limited while there are some specific designs for LLM domain.", "questions": "While the paper provides an interesting application of transfer learning to in-context learning, I have the concern of the incremental improvement and relatively complex framework. The contribution is not sufficiently novel, and the results are a bit underwhelming. I would be happy if the authors can provide more compelling evidence of ICTL’s practical benefits (e.g., compare the computational cost in Table 1, provide additional evaluation on other tasks).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730781215305}, {"id": "8XUJK767fq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6060/Reviewer_Cqhf"], "rating": 3, "soundness": 1, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper proposes a method to improve ICL (in-context learning) by integrating principles from transfer learning. Specifically, it finds similar demonstrations from another task, using distributional similarity between tasks, to copy into the ICL prompt. It finds very modest increases in ROUGE score on the Super-NI benchmark.", "review_text": "This paper proposes a method to improve ICL (in-context learning) by integrating principles from transfer learning. Specifically, it finds similar demonstrations from another task, using distributional similarity between tasks, to copy into the ICL prompt. It finds very modest increases in ROUGE score on the Super-NI benchmark.", "strengths": "1. The proposed idea seems reasonable at a high level.\n\n2. Results show some promise of proposed method.", "weaknesses": "I strongly believe this paper, in its current form, does not meet the threshold for ICLR acceptance.\n\n1. It is a bit hard to determine novelty of this work. Authors cite the previous main ICL paradigm of generating synthetic examples from a single human-provided example, but do not compare to prior methods that combine transfer learning, or similarity of examples, to do in-context learning. For instance, Dr.ICL (Demonstration-Retrieved ICL; Luo et al 2023), seems to be retrieving demonstrations that are similar to the task at hand. Thus, Luo et al seems relevant but was not mentioned or compared against.\n\n2. Much of the paper language needs editing. There are issues with writing in most paragraphs.\n\n3. Figure 2 is not particularly informative after Figure 1. Additionally, the examples in Figures 1 and 2 do not make much sense. For instance, in Figure 1, most of the “Definitions” boxes do not make sense (they are not well defined sentences), and the “Question” and “Output” pairs do not make sense either--the output is “Yes” when the question is not a yes/no question. Therefore, it confuses the reader about how the proposed transfer step would help ICL.\n\n4. There are serious coherence issues in Section 3: Method. In Section 3.1: “...We define similarity as, if the sampling scale is N, the N demonstrations most similar to the target task are called similar...” Isn’t this a recursive definition? How are the most similar N demonstrations computed? Variables in 3.1.1 are not well defined, which is a serious issue as this is the only section in the paper with math. What is a hypothesis? What is a task error? What does the “distribution for each task” mean? How are tasks defined statistically/numerically? Where is this task error term used in the method?\n\n5. Evaluations: Table 1 does not seem to show Super-NI to be a challenging task. There is not much of a difference, especially with GPT-4o, on zero-shot and proposed method (68.7 vs 73.1 ROUGE score), as acknowledged by authors. The gap is 52.0 vs 60.3 with Llama3.1-8b. Given the small gap, I would recommend authors to choose a more challenging task where ICL is much more crucial to the model’s success, so that any differences are less likely to be due to noise.\n\n6. Table 2 has a serious typo/weird result. Why is “Source Sample” row, “Rouge” column listed as 43.7 (-3.5), even though 43.7 should be a decrease of -16.6, instead of -3.5? The fact that this number is so low compared to the other entries in the column makes it seem like it is a typo.\n\n7. ROUGE is not necessarily a good metric, depending on the task and dataset answer length. A better one would be expert human evaluations (preferably pairwise evaluations between the methods). If authors do not have the resources for human eval, there are better automated metrics available.\n\n8. All of the results discussion about Figure 3, from Sections 4.4.1 through 4.4.4, lack insight, provide no analysis, and are just restating the graph trends in many words. The redundancy of information throughout this paper is prevalent in other sections as well, not just these sections.\n\n9. Section 4.4.1: “Even transferring using one source demonstration can also effectively improve the performance of the target task. This is because: (i) Even using one single source demonstration, we can also synthesize a large amount demonstrations [sic]  ... (ii) ... even without source demonstrations, LLMs can still synthesize demonstrations...” This is a tangled bunch of sentences that has many issues. First, (ii) seems to contradict the premise of using one source demonstration, as written at the beginning of the section and in (i). Second, (i) and (ii) seem to both talk about the same reason--demonstrations can be synthesized.\n\n10. I encourage the authors to improve their paper by cutting down on a lot of redundancy. A 5-page paper undershooting the page limit is better than a 9-page paper that repeats much of the information multiple times.", "questions": "1. What are some examples of source tasks? Where are the 128 or 512 demonstrations (mentioned in section 4.1.5) from? What are source-target task pair examples that were ranked most similar to each other?\n\n2. How many tokens is the average Super-NI answer in your evaluation dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to improve ICL (in-context learning) by integrating principles from transfer learning. Specifically, it finds similar demonstrations from another task, using distributional similarity between tasks, to copy into the ICL prompt. It finds very modest increases in ROUGE score on the Super-NI benchmark.", "soundness": 1, "presentation": 1, "contribution": 2, "strengths": "1. The proposed idea seems reasonable at a high level.\n\n2. Results show some promise of proposed method.", "weaknesses": "I strongly believe this paper, in its current form, does not meet the threshold for ICLR acceptance.\n\n1. It is a bit hard to determine novelty of this work. Authors cite the previous main ICL paradigm of generating synthetic examples from a single human-provided example, but do not compare to prior methods that combine transfer learning, or similarity of examples, to do in-context learning. For instance, Dr.ICL (Demonstration-Retrieved ICL; Luo et al 2023), seems to be retrieving demonstrations that are similar to the task at hand. Thus, Luo et al seems relevant but was not mentioned or compared against.\n\n2. Much of the paper language needs editing. There are issues with writing in most paragraphs.\n\n3. Figure 2 is not particularly informative after Figure 1. Additionally, the examples in Figures 1 and 2 do not make much sense. For instance, in Figure 1, most of the “Definitions” boxes do not make sense (they are not well defined sentences), and the “Question” and “Output” pairs do not make sense either--the output is “Yes” when the question is not a yes/no question. Therefore, it confuses the reader about how the proposed transfer step would help ICL.\n\n4. There are serious coherence issues in Section 3: Method. In Section 3.1: “...We define similarity as, if the sampling scale is N, the N demonstrations most similar to the target task are called similar...” Isn’t this a recursive definition? How are the most similar N demonstrations computed? Variables in 3.1.1 are not well defined, which is a serious issue as this is the only section in the paper with math. What is a hypothesis? What is a task error? What does the “distribution for each task” mean? How are tasks defined statistically/numerically? Where is this task error term used in the method?\n\n5. Evaluations: Table 1 does not seem to show Super-NI to be a challenging task. There is not much of a difference, especially with GPT-4o, on zero-shot and proposed method (68.7 vs 73.1 ROUGE score), as acknowledged by authors. The gap is 52.0 vs 60.3 with Llama3.1-8b. Given the small gap, I would recommend authors to choose a more challenging task where ICL is much more crucial to the model’s success, so that any differences are less likely to be due to noise.\n\n6. Table 2 has a serious typo/weird result. Why is “Source Sample” row, “Rouge” column listed as 43.7 (-3.5), even though 43.7 should be a decrease of -16.6, instead of -3.5? The fact that this number is so low compared to the other entries in the column makes it seem like it is a typo.\n\n7. ROUGE is not necessarily a good metric, depending on the task and dataset answer length. A better one would be expert human evaluations (preferably pairwise evaluations between the methods). If authors do not have the resources for human eval, there are better automated metrics available.\n\n8. All of the results discussion about Figure 3, from Sections 4.4.1 through 4.4.4, lack insight, provide no analysis, and are just restating the graph trends in many words. The redundancy of information throughout this paper is prevalent in other sections as well, not just these sections.\n\n9. Section 4.4.1: “Even transferring using one source demonstration can also effectively improve the performance of the target task. This is because: (i) Even using one single source demonstration, we can also synthesize a large amount demonstrations [sic]  ... (ii) ... even without source demonstrations, LLMs can still synthesize demonstrations...” This is a tangled bunch of sentences that has many issues. First, (ii) seems to contradict the premise of using one source demonstration, as written at the beginning of the section and in (i). Second, (i) and (ii) seem to both talk about the same reason--demonstrations can be synthesized.\n\n10. I encourage the authors to improve their paper by cutting down on a lot of redundancy. A 5-page paper undershooting the page limit is better than a 9-page paper that repeats much of the information multiple times.", "questions": "1. What are some examples of source tasks? Where are the 128 or 512 demonstrations (mentioned in section 4.1.5) from? What are source-target task pair examples that were ranked most similar to each other?\n\n2. How many tokens is the average Super-NI answer in your evaluation dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730697989635}, {"id": "kYm8W6sspF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6060/Reviewer_qFgM"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces In-Context Transfer Learning (ICTL), a method to enhance in-context learning by synthesizing examples for new target tasks through the transfer of demonstrations from similar tasks. Unlike conventional synthesis methods constrained by the limitations of LLMs, ICTL leverages labeled demonstrations from related source tasks. The approach involves selecting source tasks closely aligned with the target task and then refining these examples through LLMs to match the target. Experimental results on the Super-NI dataset show that ICTL achieves a 2% average performance improvement over other baselines.", "review_text": "This paper introduces In-Context Transfer Learning (ICTL), a method to enhance in-context learning by synthesizing examples for new target tasks through the transfer of demonstrations from similar tasks. Unlike conventional synthesis methods constrained by the limitations of LLMs, ICTL leverages labeled demonstrations from related source tasks. The approach involves selecting source tasks closely aligned with the target task and then refining these examples through LLMs to match the target. Experimental results on the Super-NI dataset show that ICTL achieves a 2% average performance improvement over other baselines.", "strengths": "- ICTL combines ICL and Transfer Learning, presenting a simple yet effective method for ICL across different tasks, where  target demonstrations are strictly limited.\n\n- ICTL achieves consistent improvement on the Super-NI dataset, with a 2% average performance gain over other baselines.\n\n- The paper is written in a clear manner, making it easy to understand the proposed method and its applications.", "weaknesses": "- While the proposed ICTL combines the concepts of Transfer Learning and ICL, its contribution is not clearly specified compared to prior research involving advanced demonstration generation in zero-shot ICL settings and advanced retrieval methodologies for ICL with RAG. \n\n[1] SELF-ICL: Zero-Shot In-Context Learning with Self-Generated Demonstrations (EMNLP 2023)\n\n[2] Demonstration Augmentation for Zero-shot In-context Learning (ACL 2024)\n\n[3] Bridging Distribution Gap via Semantic Rewriting with LLMs to Enhance OOD Robustness (ACL 2024)\n   \n- The selected baselines used in the experiments mostly correspond to ablations of the proposed method. For a fairer evaluation, it is necessary to directly compare ICTL with similar approaches discussed in the related work. Specifically, the authors could compare with methodologies that rely on LLMs to generate demonstrations for ICL; e.g., [1, 2] or those that modify source task formats to enhance transfer learning; e.g., [3]. These comparisons could state the contributions and novelty of the proposed ICTL clearly. \n\n- While restructuring the retrieval process based on a Transfer Learning objective is a positive step, the generation of new target demonstrations relies only on LLM prompting and self-verification, which can be insufficient to fundamentally improve task performance. Recent studies such as [4] claim that LLMs are limited in their ability to self-correct reasoning errors. To position target demonstration generation as a significant contribution, the paper needs to further enhance this aspect beyond sampling by proposing novel approaches to optimize the generation process.\n\n[4] LARGE LANGUAGE MODELS CANNOT SELF-CORRECT REASONING YET (ICLR 2024)\n\n- ICTL involves multiple steps, including sampling, transferring, and verifying demonstrations, which may increase computational demands compared to simpler in-context learning approaches. Considering the use of high-cost and high-quality models like GPT-4o and Llama3.1-8b, along with the extra inference required for demonstration generation and self-verification within ICTL, the 2%\naverage performance improvement seems less significant to justify the increased cost.\n\n- If there are few or no closely matching source tasks, the quality of transferred demonstrations may suffer, impacting the model’s ability to generalize to the target task.\n\n- The Super-NI dataset used to demonstrate performance improvement is somewhat limited. In Section 1, the authors claim that their approach aims to enhance performance on tasks where the LLM lacks prior knowledge. However, it is hard to argue that the tasks or knowledge in the Super-NI dataset are entirely absent from GPT-4o or Llama3.1-8b-Instruct models. To provide more convincing validation, the authors should either use a more up-to-date benchmark or design tasks in environments like Alfworld [5] for additional testing.\n\n[5] ALFWorld: Aligning Text and Embodied Environments for Interactive Learning (ICLR 2021)", "questions": "- The experiments in the paper focus solely on final task performance, with no analysis of generated target demonstration quality. It would strengthen the paper’s contributions if more analysis was presented. Could the authors provide example target demonstrations to illustrate the model’s limitations and how they were improved? \n\n- How does ICTL perform when there were no closely related source tasks available? Experimental results on how ICTL performs when there are limited or no closely related source tasks available would be beneficial.\n\n- How does the computational cost of ICTL compare to traditional ICL?\n\n- Could the authors compare ICTL with recent works on learning-based retrievers in LLM RAG approaches instead of Direct?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces In-Context Transfer Learning (ICTL), a method to enhance in-context learning by synthesizing examples for new target tasks through the transfer of demonstrations from similar tasks. Unlike conventional synthesis methods constrained by the limitations of LLMs, ICTL leverages labeled demonstrations from related source tasks. The approach involves selecting source tasks closely aligned with the target task and then refining these examples through LLMs to match the target. Experimental results on the Super-NI dataset show that ICTL achieves a 2% average performance improvement over other baselines.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- ICTL combines ICL and Transfer Learning, presenting a simple yet effective method for ICL across different tasks, where  target demonstrations are strictly limited.\n\n- ICTL achieves consistent improvement on the Super-NI dataset, with a 2% average performance gain over other baselines.\n\n- The paper is written in a clear manner, making it easy to understand the proposed method and its applications.", "weaknesses": "- While the proposed ICTL combines the concepts of Transfer Learning and ICL, its contribution is not clearly specified compared to prior research involving advanced demonstration generation in zero-shot ICL settings and advanced retrieval methodologies for ICL with RAG. \n\n[1] SELF-ICL: Zero-Shot In-Context Learning with Self-Generated Demonstrations (EMNLP 2023)\n\n[2] Demonstration Augmentation for Zero-shot In-context Learning (ACL 2024)\n\n[3] Bridging Distribution Gap via Semantic Rewriting with LLMs to Enhance OOD Robustness (ACL 2024)\n   \n- The selected baselines used in the experiments mostly correspond to ablations of the proposed method. For a fairer evaluation, it is necessary to directly compare ICTL with similar approaches discussed in the related work. Specifically, the authors could compare with methodologies that rely on LLMs to generate demonstrations for ICL; e.g., [1, 2] or those that modify source task formats to enhance transfer learning; e.g., [3]. These comparisons could state the contributions and novelty of the proposed ICTL clearly. \n\n- While restructuring the retrieval process based on a Transfer Learning objective is a positive step, the generation of new target demonstrations relies only on LLM prompting and self-verification, which can be insufficient to fundamentally improve task performance. Recent studies such as [4] claim that LLMs are limited in their ability to self-correct reasoning errors. To position target demonstration generation as a significant contribution, the paper needs to further enhance this aspect beyond sampling by proposing novel approaches to optimize the generation process.\n\n[4] LARGE LANGUAGE MODELS CANNOT SELF-CORRECT REASONING YET (ICLR 2024)\n\n- ICTL involves multiple steps, including sampling, transferring, and verifying demonstrations, which may increase computational demands compared to simpler in-context learning approaches. Considering the use of high-cost and high-quality models like GPT-4o and Llama3.1-8b, along with the extra inference required for demonstration generation and self-verification within ICTL, the 2%\naverage performance improvement seems less significant to justify the increased cost.\n\n- If there are few or no closely matching source tasks, the quality of transferred demonstrations may suffer, impacting the model’s ability to generalize to the target task.\n\n- The Super-NI dataset used to demonstrate performance improvement is somewhat limited. In Section 1, the authors claim that their approach aims to enhance performance on tasks where the LLM lacks prior knowledge. However, it is hard to argue that the tasks or knowledge in the Super-NI dataset are entirely absent from GPT-4o or Llama3.1-8b-Instruct models. To provide more convincing validation, the authors should either use a more up-to-date benchmark or design tasks in environments like Alfworld [5] for additional testing.\n\n[5] ALFWorld: Aligning Text and Embodied Environments for Interactive Learning (ICLR 2021)", "questions": "- The experiments in the paper focus solely on final task performance, with no analysis of generated target demonstration quality. It would strengthen the paper’s contributions if more analysis was presented. Could the authors provide example target demonstrations to illustrate the model’s limitations and how they were improved? \n\n- How does ICTL perform when there were no closely related source tasks available? Experimental results on how ICTL performs when there are limited or no closely related source tasks available would be beneficial.\n\n- How does the computational cost of ICTL compare to traditional ICL?\n\n- Could the authors compare ICTL with recent works on learning-based retrievers in LLM RAG approaches instead of Direct?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730684680699}], "openreview_url": "https://openreview.net/forum?id=ptTt8mhS7n", "arxiv_id": "2410.01548", "paper_pdf": "papers/ptTt8mhS7n.pdf", "paper_pdf_sha256": "a1cb4e1d424362746e7bd3c5fb4d3d81e656fbf1707aab06cf5388bcb39a90dd", "paper_pdf_bytes": 2458199, "paper_pdf_source": "openreview", "code_url": "https://github.com/zirui-HIT/ICTL", "code_repository": "zirui-HIT/ICTL", "code_commit": "e6a407345276fe999eb1de9f04de3a1957f91cb0", "code_archive": "repos/ptTt8mhS7n.zip", "code_archive_sha256": "a54341c1cc69880b1d313cd936efed7069965af038756b5b18edc8b7c8f62ce2", "code_archive_bytes": 34994, "code_file_count": 27, "code_extensions": {".py": 18, ".sh": 9}, "github_disk_usage_kb": 28, "github_languages": {"Python": 68716, "Shell": 6238}, "github_archived": false, "github_pushed_at": "2024-10-01T15:26:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/in-context-transfer-learning-demonstration"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JB3lbDtsFS", "year": 2024, "status": "rejected", "title": "It HAS to be Subjective: Human Annotator Simulation via Zero-shot Density Estimation", "authors": ["Wen Wu", "Wenlin Chen", "Chao Zhang", "Phil Woodland"], "authorids": ["~Wen_Wu2", "~Wenlin_Chen2", "~Chao_Zhang20", "~Phil_Woodland1"], "authors_source": "OpenReview API", "abstract": "Human annotator simulation (HAS) serves as a cost-effective substitute for human evaluation such as data annotation and system assessment. Human perception and behaviour during human evaluation exhibit inherent variability due to diverse cognitive processes and subjective interpretations, which should be taken into account in modelling to better mimic the way people perceive and interact with the world. This paper introduces a novel meta-learning framework that treats HAS as a zero-shot density estimation problem, which incorporates human variability and allows for the efficient generation of human-like annotations for unlabelled test inputs. Under this framework, we propose two new model classes, conditional integer flows and conditional softmax flows, to account for ordinal and categorical annotations, respectively. The proposed method is evaluated on three real-world human evaluation tasks and shows superior capability and efficiency to predict the aggregated behaviours of human annotators, match the distribution of human annotations, and simulate the inter-annotator disagreements.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zssOWXbbAE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission782/Reviewer_szvo"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper studies the problem of human annotator simulation (HAS) as a density estimation problem, where marginal distribution of how the labels would be generated by a group of annotators given a particular sample is learnt.", "review_text": "The paper studies the problem of human annotator simulation (HAS) as a density estimation problem, where marginal distribution of how the labels would be generated by a group of annotators given a particular sample is learnt.", "strengths": "- The authors argue that existing methods take a majority vote of among multiple annotators. This is not reflective in case of majority bias, and is instead well captured by treating it as a density estimation problem. Fig 1, 2 clearly show that the predictions made by the proposed CNF algorithm  lie on the sample means and have enough variability to capture the diversity in human annotations.", "weaknesses": "[1] The whole motivation of paper is that it should be possible to capture the data bias (when majority of labelled samples are wrong), and correct for it in some way. However, in all of the tables, the ensemble (tab 1,2)/gaussian process(tab 3) seem to be performing better than the proposed cnf method. While it is qualitatively visible that CNF is learning a better density distribution, the quantitative numbers dont justify the intuition. Perhaps, the authors would be well served by explicitly identifying the samples which possess such labelling bias, correct for it (by forcing the distribution to skew), and show better performance than all other methods. Right now, i see such form of analysis as lacking.\n\n[2]  What is the motivation behind latent-diffusion model. in my understanding, the marginal p(y|z) is already encoding information on different annotators z1,z2....z_m sampled over z. what does additional variable v inject into the model conceptually (apart from additional representational capacity).\n\n[3]  What is the reason for separate analysis of ordinal and categorical variables? I understand that ordinal categories are ordered, and that probability estimation then reduces to summing over the continuous space of latent variable v. However, I haven't seen standard classification setups explicitly enforce such ordering constraints. Also are there any cases where annotations are continuous (for eg, annotation by a speech etc.), which could be used to explore continuous cases? That could act as an interesting toy experiment.......\n\n[4] Possible extensions to cases where only single annotation is available for each sample:\n\t- how does this work compare to works which aim to filter noisy labels from the network. perhaps this could be mentioned in the related work. \n\nI like the fresh perspective  of learning distribution over labels, and how such system could act as an auto-labeller. However, the motivation does not reflect better performance on real-world metric (i.e. accuracy). Right now it feels like a fancy technique (i.e. density estimation) whose PRACTICAL real-world experiments i cannot see. Also, i dont see any experiments on zero-shot learning, which is the main title of the paper. \n\nFinally, this paper seems to present a chicken and egg problem. Most of the datasets in the real- world contain labels of only one annotator but might be biased. However, this paper does not work on them. Instead, it requires datasets where each sample has been annotated by many people. But that might not be how labelling happens in real-world. So, perhaps authors could discuss how to extend their method to single label cases where distributions could not be learnt. \n\nHowever, this work shows promise, and I would recommend the authors to improve the paper by addressing the above concerns and consider a future resubmission.\n\n-----\nPOST REBUTTAL\n-----\nThe authors were able to address most of the comments. Three main issues are still open,\n\n[1] The seeming unexplainable negative correlation between RMSE and Accuracy.\n\n-> Standard intuition suggests that lower RMSE should lead to higher accuracy in classification setups. If one accepts, that there are certain annotators (i.e. group of people) who are collectively biased, and that the sample might be incorrectly labelled, then, it should be possible to ‘correct’ for it. One way, could be perhaps reestimating the correct label and retrain the said classifier and SHOW better RMSE. I very much appreciate the authors efforts to add a new metric ICC to the setup. However, the actual increase on the accuracy (which is a well-accepted metric), would have convinced further and shown applicability on real world setup. Note that i am not asking for additional experiments on datasets which have only one label per sample, but only the performance improvements in the context of experiments the authors have already performed.\n\n[2] Zero shot density estimation.\n\n-> standard learning assumes that the neural net fits a density (which does not change) after the machine has learnt. during inference, a sample (from same/different distribution) is fed to the model and evaluated. The machine cant adapt its learnt density, since that is encoded in the weights which remain static.\n->the authors clarify zero shot as the ability to predict density of annotator responses p*, from a single phrase (x*).\n\t-> this might work, if the annotators (say A1) labelled a particular sample x1 (which was SEEN during training), and network is asked to predict A1’s beliefs for a new sample x* (which has not been seen). However, the problem then does not remain zero shot since A1 was already seen by the network.\n\t-> If we accept (free will), i.e. one’s personal beliefs are independent of statistical treatments of how other people respond, then merely sampling from the learnt distribution of annotator response, to ‘simulate’ how a prospective unseen annotator shall respond might not work.\n\n[3] Dynamic density adaptation\n\nThe idea of giving  networks the ability to adapt their densities dynamically to if a given sample is OOD seems promising. The only issue is that personal beliefs cant be given a density treatment. If so, then it doesn't remain zero shot.\n\nOverall, this is an interesting work, along above three points, it is not clear in what context this will be useful if it can not be used to improve the performance.", "questions": "- relevance to title of the paper.\n\t- paper is titled zero shot density estimation, by zero shot i understand that a sample which is different from the original dataset on which the estimator was fit, could be used. given a set of gt classes, the network should dynamically predict distribution of labels over annotators. However, i see no such experiments, with training/inference being done over SAME dataset of speech, toxicity, and emotion annotations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of human annotator simulation (HAS) as a density estimation problem, where marginal distribution of how the labels would be generated by a group of annotators given a particular sample is learnt.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The authors argue that existing methods take a majority vote of among multiple annotators. This is not reflective in case of majority bias, and is instead well captured by treating it as a density estimation problem. Fig 1, 2 clearly show that the predictions made by the proposed CNF algorithm  lie on the sample means and have enough variability to capture the diversity in human annotations.", "weaknesses": "[1] The whole motivation of paper is that it should be possible to capture the data bias (when majority of labelled samples are wrong), and correct for it in some way. However, in all of the tables, the ensemble (tab 1,2)/gaussian process(tab 3) seem to be performing better than the proposed cnf method. While it is qualitatively visible that CNF is learning a better density distribution, the quantitative numbers dont justify the intuition. Perhaps, the authors would be well served by explicitly identifying the samples which possess such labelling bias, correct for it (by forcing the distribution to skew), and show better performance than all other methods. Right now, i see such form of analysis as lacking.\n\n[2]  What is the motivation behind latent-diffusion model. in my understanding, the marginal p(y|z) is already encoding information on different annotators z1,z2....z_m sampled over z. what does additional variable v inject into the model conceptually (apart from additional representational capacity).\n\n[3]  What is the reason for separate analysis of ordinal and categorical variables? I understand that ordinal categories are ordered, and that probability estimation then reduces to summing over the continuous space of latent variable v. However, I haven't seen standard classification setups explicitly enforce such ordering constraints. Also are there any cases where annotations are continuous (for eg, annotation by a speech etc.), which could be used to explore continuous cases? That could act as an interesting toy experiment.......\n\n[4] Possible extensions to cases where only single annotation is available for each sample:\n\t- how does this work compare to works which aim to filter noisy labels from the network. perhaps this could be mentioned in the related work. \n\nI like the fresh perspective  of learning distribution over labels, and how such system could act as an auto-labeller. However, the motivation does not reflect better performance on real-world metric (i.e. accuracy). Right now it feels like a fancy technique (i.e. density estimation) whose PRACTICAL real-world experiments i cannot see. Also, i dont see any experiments on zero-shot learning, which is the main title of the paper. \n\nFinally, this paper seems to present a chicken and egg problem. Most of the datasets in the real- world contain labels of only one annotator but might be biased. However, this paper does not work on them. Instead, it requires datasets where each sample has been annotated by many people. But that might not be how labelling happens in real-world. So, perhaps authors could discuss how to extend their method to single label cases where distributions could not be learnt. \n\nHowever, this work shows promise, and I would recommend the authors to improve the paper by addressing the above concerns and consider a future resubmission.\n\n-----\nPOST REBUTTAL\n-----\nThe authors were able to address most of the comments. Three main issues are still open,\n\n[1] The seeming unexplainable negative correlation between RMSE and Accuracy.\n\n-> Standard intuition suggests that lower RMSE should lead to higher accuracy in classification setups. If one accepts, that there are certain annotators (i.e. group of people) who are collectively biased, and that the sample might be incorrectly labelled, then, it should be possible to ‘correct’ for it. One way, could be perhaps reestimating the correct label and retrain the said classifier and SHOW better RMSE. I very much appreciate the authors efforts to add a new metric ICC to the setup. However, the actual increase on the accuracy (which is a well-accepted metric), would have convinced further and shown applicability on real world setup. Note that i am not asking for additional experiments on datasets which have only one label per sample, but only the performance improvements in the context of experiments the authors have already performed.\n\n[2] Zero shot density estimation.\n\n-> standard learning assumes that the neural net fits a density (which does not change) after the machine has learnt. during inference, a sample (from same/different distribution) is fed to the model and evaluated. The machine cant adapt its learnt density, since that is encoded in the weights which remain static.\n->the authors clarify zero shot as the ability to predict density of annotator responses p*, from a single phrase (x*).\n\t-> this might work, if the annotators (say A1) labelled a particular sample x1 (which was SEEN during training), and network is asked to predict A1’s beliefs for a new sample x* (which has not been seen). However, the problem then does not remain zero shot since A1 was already seen by the network.\n\t-> If we accept (free will), i.e. one’s personal beliefs are independent of statistical treatments of how other people respond, then merely sampling from the learnt distribution of annotator response, to ‘simulate’ how a prospective unseen annotator shall respond might not work.\n\n[3] Dynamic density adaptation\n\nThe idea of giving  networks the ability to adapt their densities dynamically to if a given sample is OOD seems promising. The only issue is that personal beliefs cant be given a density treatment. If so, then it doesn't remain zero shot.\n\nOverall, this is an interesting work, along above three points, it is not clear in what context this will be useful if it can not be used to improve the performance.", "questions": "- relevance to title of the paper.\n\t- paper is titled zero shot density estimation, by zero shot i understand that a sample which is different from the original dataset on which the estimator was fit, could be used. given a set of gt classes, the network should dynamically predict distribution of labels over annotators. However, i see no such experiments, with training/inference being done over SAME dataset of speech, toxicity, and emotion annotations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698939144456}, {"id": "v4NxcTkIdS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission782/Reviewer_js5b"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Supervised learning tasks require annotation that can be done with high certainty, such as annotating the presence of an object, drawing rough bounding boxes, or deciding scene attributes; however, many tasks require subjective labeling that is influenced by a variety of factors, cognitive biases, or personal preferences. This paper proposes a human annotator simulation to incorporate the variabilities in this second group of labeling tasks.\n\nThe method is a meta-learning framework, a zero-shot density estimator that models the agreement and disagreements among human annotators using a latent variable model. It does not require any human effort and can be used to use unlabeled samples efficiently.\n\nThe experimental results are performed on different modalities and domains that demand various levels of subjective annotations, such as emotion category, toxic speech, and speech quality assessment.", "review_text": "Supervised learning tasks require annotation that can be done with high certainty, such as annotating the presence of an object, drawing rough bounding boxes, or deciding scene attributes; however, many tasks require subjective labeling that is influenced by a variety of factors, cognitive biases, or personal preferences. This paper proposes a human annotator simulation to incorporate the variabilities in this second group of labeling tasks.\n\nThe method is a meta-learning framework, a zero-shot density estimator that models the agreement and disagreements among human annotators using a latent variable model. It does not require any human effort and can be used to use unlabeled samples efficiently.\n\nThe experimental results are performed on different modalities and domains that demand various levels of subjective annotations, such as emotion category, toxic speech, and speech quality assessment.", "strengths": "* Various machine learning problems require subjective labeling, and it is not easy to use crowdsourcing due to privacy concerns. The proposed approach makes such highly sensitive data to be labeled and used in learning problems.\n\n* The proposed approach is a latent variable model. $p(z | x)$ encodes the information in the input $x$, however, the interesting part of the formulation is to introduce another intermediate variable $v$, instead of directly taking $p(y | z)$. Conditional normalising flow (CNF) formulation gives more flexibility instead of a particular distribution choice.\n\n* Conditional integer flows and conditional softmax flows are introduced to accommodate to ordinal and categorical annotation tasks.", "weaknesses": "* In the experiments, test performances are reported, however, considering the problem as \"simulating subjective human annotations\", I would expect to see (i) a training to learn the latent model in a labeled training subset, (ii) generate simulated labels on a held-out training subset and (iii) training a classifier only with these simulated labels and evaluating on test set. Instead, the current evaluation is more like supervised evaluation.\n\n*  I found the problem address highly similar to the following paper [1,2] that aims to model the label space in each step iteratively using Gibbs sampling (or previous like of work that used MCMC). This work may require annotation process in a more dynamic setting, still very relevant to subjective labelling task and I think, they are needed to discussed.\n\n    [1] Harrison, P., Marjieh, R., Adolfi, F., van Rijn, P., Anglada-Tort, M., Tchernichovski, O., ... & Jacoby, N. (2020). Gibbs sampling with people. Advances in neural information processing systems, 33, 10659-10671. https://proceedings.neurips.cc/paper_files/paper/2020/file/7880d7226e872b776d8b9f23975e2a3d-Paper.pdf.  \n\n    [2] Sanborn, A., & Griffiths, T. (2007). Markov chain Monte Carlo with people. Advances in neural information processing systems, 20. https://papers.nips.cc/paper_files/paper/2007/file/89d4402dc03d3b7318bbac10203034ab-Paper.pdf\n\n* How does the proposed approach tackle the highly imbalanced data domains? For instance, in one of the tasks, MSP podcast dataset contains angry, sad, happy, neutral, and other. When continuous labels (valence-arousal annotations of the same dataset) is used, the skewed labeled distribution will be more visible. I suspect the proposed method to cause higher uncertainty (interrupter disagreement in less discovered part of label parameter space).\n\n* Different performance metrics are used. Particularly, Fleis' kappa for reliability of categorical labels is good. However, why did not you used Intraclass correlation coefficient (ICC) in continuous labels instead of RMSE and the absolute error of the average standard deviations? In continuous/ordinal subjective labelling tasks, ICC reliability is one of the golden standards that define the labelling quality or difficulty of the task.", "questions": "Overall, I liked the Human Annotator Simulation approach to tackle learning domains that necessitates subjective labelling. Please see my comments in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Supervised learning tasks require annotation that can be done with high certainty, such as annotating the presence of an object, drawing rough bounding boxes, or deciding scene attributes; however, many tasks require subjective labeling that is influenced by a variety of factors, cognitive biases, or personal preferences. This paper proposes a human annotator simulation to incorporate the variabilities in this second group of labeling tasks.\n\nThe method is a meta-learning framework, a zero-shot density estimator that models the agreement and disagreements among human annotators using a latent variable model. It does not require any human effort and can be used to use unlabeled samples efficiently.\n\nThe experimental results are performed on different modalities and domains that demand various levels of subjective annotations, such as emotion category, toxic speech, and speech quality assessment.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "* Various machine learning problems require subjective labeling, and it is not easy to use crowdsourcing due to privacy concerns. The proposed approach makes such highly sensitive data to be labeled and used in learning problems.\n\n* The proposed approach is a latent variable model. $p(z | x)$ encodes the information in the input $x$, however, the interesting part of the formulation is to introduce another intermediate variable $v$, instead of directly taking $p(y | z)$. Conditional normalising flow (CNF) formulation gives more flexibility instead of a particular distribution choice.\n\n* Conditional integer flows and conditional softmax flows are introduced to accommodate to ordinal and categorical annotation tasks.", "weaknesses": "* In the experiments, test performances are reported, however, considering the problem as \"simulating subjective human annotations\", I would expect to see (i) a training to learn the latent model in a labeled training subset, (ii) generate simulated labels on a held-out training subset and (iii) training a classifier only with these simulated labels and evaluating on test set. Instead, the current evaluation is more like supervised evaluation.\n\n*  I found the problem address highly similar to the following paper [1,2] that aims to model the label space in each step iteratively using Gibbs sampling (or previous like of work that used MCMC). This work may require annotation process in a more dynamic setting, still very relevant to subjective labelling task and I think, they are needed to discussed.\n\n    [1] Harrison, P., Marjieh, R., Adolfi, F., van Rijn, P., Anglada-Tort, M., Tchernichovski, O., ... & Jacoby, N. (2020). Gibbs sampling with people. Advances in neural information processing systems, 33, 10659-10671. https://proceedings.neurips.cc/paper_files/paper/2020/file/7880d7226e872b776d8b9f23975e2a3d-Paper.pdf.  \n\n    [2] Sanborn, A., & Griffiths, T. (2007). Markov chain Monte Carlo with people. Advances in neural information processing systems, 20. https://papers.nips.cc/paper_files/paper/2007/file/89d4402dc03d3b7318bbac10203034ab-Paper.pdf\n\n* How does the proposed approach tackle the highly imbalanced data domains? For instance, in one of the tasks, MSP podcast dataset contains angry, sad, happy, neutral, and other. When continuous labels (valence-arousal annotations of the same dataset) is used, the skewed labeled distribution will be more visible. I suspect the proposed method to cause higher uncertainty (interrupter disagreement in less discovered part of label parameter space).\n\n* Different performance metrics are used. Particularly, Fleis' kappa for reliability of categorical labels is good. However, why did not you used Intraclass correlation coefficient (ICC) in continuous labels instead of RMSE and the absolute error of the average standard deviations? In continuous/ordinal subjective labelling tasks, ICC reliability is one of the golden standards that define the labelling quality or difficulty of the task.", "questions": "Overall, I liked the Human Annotator Simulation approach to tackle learning domains that necessitates subjective labelling. Please see my comments in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed. The proposed paper, in contrast, aim to model human annotation processes and mitigate existing biases.", "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698823650173}, {"id": "rbhuOmZaT2", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission782/Reviewer_2YK4"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "The paper focuses on human annotator simulation (HAS), which is the task of generating human-like annotations for unlabelled inputs. \n\nThe paper proposes a novel meta-learning framework that treats HAS as a zero-shot density estimation problem, which can capture the variability and subjectivity in human evaluation. \n\nThe paper also introduces two new model classes, conditional integer flows, and conditional softmax flows, to handle ordinal and categorical annotations respectively.\n\nThe paper evaluates the proposed method on three real-world human evaluation tasks: emotion recognition, toxic speech detection, and speech quality assessment. The paper shows that the proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements.", "review_text": "The paper focuses on human annotator simulation (HAS), which is the task of generating human-like annotations for unlabelled inputs. \n\nThe paper proposes a novel meta-learning framework that treats HAS as a zero-shot density estimation problem, which can capture the variability and subjectivity in human evaluation. \n\nThe paper also introduces two new model classes, conditional integer flows, and conditional softmax flows, to handle ordinal and categorical annotations respectively.\n\nThe paper evaluates the proposed method on three real-world human evaluation tasks: emotion recognition, toxic speech detection, and speech quality assessment. The paper shows that the proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements.", "strengths": "The paper proposes a novel meta-learning framework that treats human annotator simulation (HAS) as a zero-shot density estimation problem, which can capture the variability and subjectivity in human evaluation. The proposed method is evaluated on three real-world human evaluation tasks: emotion recognition, toxic speech detection, and speech quality assessment. The paper shows that the proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements. The paper also introduces two new model classes, conditional integer flows and conditional softmax flows, to handle ordinal and categorical annotations respectively. The proposed method is efficient and capable of generating human-like annotations for unlabelled test inputs. \n\nThe strengths of the paper are:\n\n- The proposed method is capable of generating human-like annotations for unlabelled test inputs with higher accuracy than the baseline methods.\n\n- The proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements.\n\n- The paper introduces two new model classes, conditional integer flows, and conditional softmax flows, to handle ordinal and categorical annotations respectively.\n\n- The paper discusses the ethical implications and potential applications of HAS.\n\n- Training code is available to reproduce the results in this paper.\n\nOverall, this paper presents a novel approach to HAS that can capture the variability and subjectivity in human evaluation. The paper also provides insights into how to handle ordinal and categorical annotations.", "weaknesses": "- The proposed method is evaluated on only three human evaluation tasks, which may not be sufficient to generalize the effectiveness of the proposed method to other domains.\n\n- The paper does not compare the proposed method with the latest state-of-the-art methods for HAS. The methods compared in this paper include deep ensemble (Ensemble) (Lakshminarayanan et al., 2017), Monte-Carlo dropout (MCDP) (Gal & Ghahramani, 2016), Bayes-by-backprop (BBB) (Blundell et al., 2015), conditional variational autoencoder (CVAE) (Kingma & Welling, 2014), conditional argmax flow (A-CNF) (Hoogeboom et al., 2021), Dirichlet prior network (DPN) (Malinin & Gales, 2018), Gaussian process (GP) (Williams & Rasmussen, 2006), and evidential deep learning (EDL) (Amini et al., 2020). The most recent method A-CNF was proposed two years ago.\n\n- The paper does not provide a detailed analysis of the robustness of the proposed method.\n\nDespite these limitations, the paper presents a novel approach to HAS that can capture the variability and subjectivity in human evaluation. The paper also provides insights into how to handle ordinal and categorical annotations. However, further research is needed to evaluate the proposed method on more diverse datasets and tasks.", "questions": "I am a computer vision researcher, and I do not know the state-of-the-art of the HAS. However, the most recent method A-CNF compared in this paper was proposed two years ago. Is there any other recent work proposed in the past two years?\n\nEnsemble achieves the best performance in Table 1 and Table 2. The proposed method is much more efficient than the ensemble method. Could you please provide a detailed complexity comparison?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper focuses on human annotator simulation (HAS), which is the task of generating human-like annotations for unlabelled inputs. \n\nThe paper proposes a novel meta-learning framework that treats HAS as a zero-shot density estimation problem, which can capture the variability and subjectivity in human evaluation. \n\nThe paper also introduces two new model classes, conditional integer flows, and conditional softmax flows, to handle ordinal and categorical annotations respectively.\n\nThe paper evaluates the proposed method on three real-world human evaluation tasks: emotion recognition, toxic speech detection, and speech quality assessment. The paper shows that the proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper proposes a novel meta-learning framework that treats human annotator simulation (HAS) as a zero-shot density estimation problem, which can capture the variability and subjectivity in human evaluation. The proposed method is evaluated on three real-world human evaluation tasks: emotion recognition, toxic speech detection, and speech quality assessment. The paper shows that the proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements. The paper also introduces two new model classes, conditional integer flows and conditional softmax flows, to handle ordinal and categorical annotations respectively. The proposed method is efficient and capable of generating human-like annotations for unlabelled test inputs. \n\nThe strengths of the paper are:\n\n- The proposed method is capable of generating human-like annotations for unlabelled test inputs with higher accuracy than the baseline methods.\n\n- The proposed method can better predict the aggregated behaviors of human annotators, match the distribution of human annotations, and simulate inter-annotator disagreements.\n\n- The paper introduces two new model classes, conditional integer flows, and conditional softmax flows, to handle ordinal and categorical annotations respectively.\n\n- The paper discusses the ethical implications and potential applications of HAS.\n\n- Training code is available to reproduce the results in this paper.\n\nOverall, this paper presents a novel approach to HAS that can capture the variability and subjectivity in human evaluation. The paper also provides insights into how to handle ordinal and categorical annotations.", "weaknesses": "- The proposed method is evaluated on only three human evaluation tasks, which may not be sufficient to generalize the effectiveness of the proposed method to other domains.\n\n- The paper does not compare the proposed method with the latest state-of-the-art methods for HAS. The methods compared in this paper include deep ensemble (Ensemble) (Lakshminarayanan et al., 2017), Monte-Carlo dropout (MCDP) (Gal & Ghahramani, 2016), Bayes-by-backprop (BBB) (Blundell et al., 2015), conditional variational autoencoder (CVAE) (Kingma & Welling, 2014), conditional argmax flow (A-CNF) (Hoogeboom et al., 2021), Dirichlet prior network (DPN) (Malinin & Gales, 2018), Gaussian process (GP) (Williams & Rasmussen, 2006), and evidential deep learning (EDL) (Amini et al., 2020). The most recent method A-CNF was proposed two years ago.\n\n- The paper does not provide a detailed analysis of the robustness of the proposed method.\n\nDespite these limitations, the paper presents a novel approach to HAS that can capture the variability and subjectivity in human evaluation. The paper also provides insights into how to handle ordinal and categorical annotations. However, further research is needed to evaluate the proposed method on more diverse datasets and tasks.", "questions": "I am a computer vision researcher, and I do not know the state-of-the-art of the HAS. However, the most recent method A-CNF compared in this paper was proposed two years ago. Is there any other recent work proposed in the past two years?\n\nEnsemble achieves the best performance in Table 1 and Table 2. The proposed method is much more efficient than the ensemble method. Could you please provide a detailed complexity comparison?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA, the proposed method is designed to alleviate ethic problems.", "rating": "6: marginally above the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698778095533}, {"id": "8VyKuatZPT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission782/Reviewer_Pupg"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The papers proposed a meta-learning framework to make Human annotator simulation as a zero-shot density estimation problem, which allows for the generation of human-like annotations for unlabelled data. Moreover, conditional integer flows and conditional softmax flows can account for ordinal and categorical annotations.", "review_text": "The papers proposed a meta-learning framework to make Human annotator simulation as a zero-shot density estimation problem, which allows for the generation of human-like annotations for unlabelled data. Moreover, conditional integer flows and conditional softmax flows can account for ordinal and categorical annotations.", "strengths": "(1) The paper is writting and organzation are clearly.\n(2) The idea of adopting meta-learning into human annotations for zero-shot unlabeled data is sensible.\n(3) The experimental results on three real-world human evaluation tasks seems promising.", "weaknesses": "(1) The main weakness is the limited compared method in the experiments. In the Tables in the paper, the state-of-the-art meta-learning methods are missing and the latested human annnotation simulators are also disregarded. Please consider add more SOTA methods for comparsion.\n(2) The other weakness is that the costing time is not reported in the paper. Since the authors claimed that the proposed method is effective, the training and test time should be list compared to other SOTA in the experiments. Please consider to add more details here.", "questions": "Please see Weakness Part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The papers proposed a meta-learning framework to make Human annotator simulation as a zero-shot density estimation problem, which allows for the generation of human-like annotations for unlabelled data. Moreover, conditional integer flows and conditional softmax flows can account for ordinal and categorical annotations.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "(1) The paper is writting and organzation are clearly.\n(2) The idea of adopting meta-learning into human annotations for zero-shot unlabeled data is sensible.\n(3) The experimental results on three real-world human evaluation tasks seems promising.", "weaknesses": "(1) The main weakness is the limited compared method in the experiments. In the Tables in the paper, the state-of-the-art meta-learning methods are missing and the latested human annnotation simulators are also disregarded. Please consider add more SOTA methods for comparsion.\n(2) The other weakness is that the costing time is not reported in the paper. Since the authors claimed that the proposed method is effective, the training and test time should be list compared to other SOTA in the experiments. Please consider to add more details here.", "questions": "Please see Weakness Part.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None.", "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698758341481}], "openreview_url": "https://openreview.net/forum?id=JB3lbDtsFS", "arxiv_id": "2310.00486", "paper_pdf": "papers/JB3lbDtsFS.pdf", "paper_pdf_sha256": "7bb2763b53709194c082f23e57d1b0260f0ae9bd17ae2250872093c9922e2c81", "paper_pdf_bytes": 1378326, "paper_pdf_source": "openreview", "code_url": "https://github.com/W-Wu/HAS_CNF", "code_repository": "W-Wu/HAS_CNF", "code_commit": "62b5e149a8a27f1ebfa6bd03d46dfa44fab844b8", "code_archive": "repos/JB3lbDtsFS.zip", "code_archive_sha256": "e8a3936386c1cf9be63f1575e36b71df8afd7f5ec9017d78695b26c19ebe045e", "code_archive_bytes": 20152, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 27, "github_languages": {"Python": 44520}, "github_archived": false, "github_pushed_at": "2024-08-24T12:30:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/it-has-to-be-subjective-human-annotator"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NZZoABNZECq", "year": 2023, "status": "rejected", "title": "Mechanistic Mode Connectivity", "authors": ["Ekdeep Singh Lubana", "Eric J Bigelow", "Robert P. Dick", "David Krueger", "Hidenori Tanaka"], "authorids": ["~Ekdeep_Singh_Lubana1", "~Eric_J_Bigelow1", "~Robert_P._Dick1", "~David_Krueger1", "~Hidenori_Tanaka1"], "authors_source": "OpenReview API", "abstract": "With the rise of pretrained models, fine-tuning has become of central importance in deep learning. However, unlike retraining from scratch, fine-tuning can fail to qualitatively change the behavior of a pre-trained network. For instance, we find in practice that naive fine-tuning does not eliminate a model’s sensitivity to spurious features. To understand and address this limitation, we study the geometry of neural network loss landscapes through the lens of mode-connectivity. Our work addresses two questions about mode-connectivity: 1) Are models trained on different data distributions mode-connected? 2) Can we fine tune a pre-trained model to switch modes? We define a notion of mechanistic mode-connectivity, and find that only models that already share the same invariances (which we call “mechanistically similar”) are mechanistically mode-connected. We hypothesize this property explains inability of naive fine-tuning methods to induce invariance to spurious features. Based on our analysis, we propose and validate a method of “mechanistic fine-tuning” called connectivity-based fine-tuning (CBFT)", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "yUa-KT6_vD", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5987/Reviewer_X6v6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper is primarily concerned with **identifying spurious attributes** learned by NNs when learning a specific dataset and **correcting them** during fine-tuning. To this end, they propose mechanistic similarity, a concept that checks whether two pre-trained models with low loss are functionally similar. Based on this concept, they develop connectivity-based fine-tuning, a fine-tuning technique that can modify the mechanism of a pre-trained model.", "review_text": "This paper proposes a novel solution (Connectivity-Based Fine-Tuning) to an interesting problem (removing spurious attributes). While the concepts of Mechanistic Similarity and Mechanistic Mode Connectivity are exciting and inspiring, no discussion has been given on how to remove the spurious attribute. The paper will be helpful to readers in many fields with revisions for natural connections between sections.", "strengths": "* Strength\n\n  * From a theoretical point of view, their contributions are largely twofold. First, authors **applied the concept of invariance** to explain the underlying mechanism of the models. (Section 3) Second, a **fascinating conjecture** (Conjecture 1) was presented by the authors about how mode connectivity, an abstract concept in loss-landscape, will affect real-world applications. (Section 4)\n  * On the application side, the authors proposed Connectivity-Based Fine-Tuning as a way to **intentionally editing** the pre-trained model's underlying mechanism. (Section 5)\n  * These contribution points are novel and interesting for scientists studying the loss-landscape of NNs.\n\n* Weakness\n\n  However, readers who wish to apply these findings to their problems may encounter some difficulties for the following reasons. Considering this is discussion phase 2, it is impossible to revise the paper, but I believe the first and third issues should be addressed.\n\n  1. It is unclear how to get rid of spurious attributes in Sections 3 and 4. For example, the authors only explain how the proposed notion of mechanistically similar confirms the functional equivalence of the two models in Section 3. **There is no discussion of what mechanstically similarity corresponds to spurious attributes**. This issue also applies to Section 4 (i.e., There is no explicit connection between spurious attributes and Connectivity-Baed Fine-Tuning).  As a result, readers may feel as if there is no natural connection between sections. **Since the content of each section is interesting enough, I believe the connectivity between sections is the greater concern.**\n\n  2. Although the code was submitted during the discussion period, Section 5's Connectivity-Based Fine-Tuning remains vague in their main paper. For example, how do you pick a minimal dataset $\\mathcal{D}_{NC}$ that does not contain attribute C? Also, why does the average representation matching term used in (ii) of equation (2) make an invariant prediction? For example, what happens if the term is the average of multiple augmentations of a single image rather than the average of multiple images of a single class? **Since Connectivity-Based Fine-Tuning is not the only (or optimal) algorithm that achieves the authors' purpose, these questions are quite natural, and the authors should have explained them appropriately.**\n\n  3. There are some mathematical expressions that seem to be abused.\n\n     * Although an inverse mapping is required for isomorphism, $\\mathcal{A}^{\\alpha_i}_i$ has no inverse mapping because it fixes the value of the $i$-th component of the latent variable to $\\alpha_i$.\n\n     * While authors denote the domain of $\\mathcal{E}$ as $\\mathcal{X} \\times \\mathbb{R}^{m}$, $\\mathbb{R}^{m}$ is incorrect because $\\hat{\\mathcal{A}}$ is a concatenation of interventions.\n\n     * Authors used square brackets both to represent lists (e.g., $[K]$ in Section 2) and to represent closed intervals ($[0,1]$ in Section 5). \n\n     * Furthermore, some mathematical expressions are used without any definition. For example, ㅅthey did not define exactly what $[K]$ means. This can only be inferred from context. This issue also applies to the Truncate function in (ii) of equation (3). The exact mathematical expression of $f_\\gamma$ in Section 5 is not specified. Similarly, $\\gamma_{\\theta \\rightarrow \\theta_C}(t)$ is not a path, but rather a point on a path. \n\n       **Combined with the first weakness, these misuses of mathematical expressions will present a huge barrier for readers unfamiliar with the field.**", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper is primarily concerned with **identifying spurious attributes** learned by NNs when learning a specific dataset and **correcting them** during fine-tuning. To this end, they propose mechanistic similarity, a concept that checks whether two pre-trained models with low loss are functionally similar. Based on this concept, they develop connectivity-based fine-tuning, a fine-tuning technique that can modify the mechanism of a pre-trained model.", "strength_and_weaknesses": "* Strength\n\n  * From a theoretical point of view, their contributions are largely twofold. First, authors **applied the concept of invariance** to explain the underlying mechanism of the models. (Section 3) Second, a **fascinating conjecture** (Conjecture 1) was presented by the authors about how mode connectivity, an abstract concept in loss-landscape, will affect real-world applications. (Section 4)\n  * On the application side, the authors proposed Connectivity-Based Fine-Tuning as a way to **intentionally editing** the pre-trained model's underlying mechanism. (Section 5)\n  * These contribution points are novel and interesting for scientists studying the loss-landscape of NNs.\n\n* Weakness\n\n  However, readers who wish to apply these findings to their problems may encounter some difficulties for the following reasons. Considering this is discussion phase 2, it is impossible to revise the paper, but I believe the first and third issues should be addressed.\n\n  1. It is unclear how to get rid of spurious attributes in Sections 3 and 4. For example, the authors only explain how the proposed notion of mechanistically similar confirms the functional equivalence of the two models in Section 3. **There is no discussion of what mechanstically similarity corresponds to spurious attributes**. This issue also applies to Section 4 (i.e., There is no explicit connection between spurious attributes and Connectivity-Baed Fine-Tuning).  As a result, readers may feel as if there is no natural connection between sections. **Since the content of each section is interesting enough, I believe the connectivity between sections is the greater concern.**\n\n  2. Although the code was submitted during the discussion period, Section 5's Connectivity-Based Fine-Tuning remains vague in their main paper. For example, how do you pick a minimal dataset $\\mathcal{D}_{NC}$ that does not contain attribute C? Also, why does the average representation matching term used in (ii) of equation (2) make an invariant prediction? For example, what happens if the term is the average of multiple augmentations of a single image rather than the average of multiple images of a single class? **Since Connectivity-Based Fine-Tuning is not the only (or optimal) algorithm that achieves the authors' purpose, these questions are quite natural, and the authors should have explained them appropriately.**\n\n  3. There are some mathematical expressions that seem to be abused.\n\n     * Although an inverse mapping is required for isomorphism, $\\mathcal{A}^{\\alpha_i}_i$ has no inverse mapping because it fixes the value of the $i$-th component of the latent variable to $\\alpha_i$.\n\n     * While authors denote the domain of $\\mathcal{E}$ as $\\mathcal{X} \\times \\mathbb{R}^{m}$, $\\mathbb{R}^{m}$ is incorrect because $\\hat{\\mathcal{A}}$ is a concatenation of interventions.\n\n     * Authors used square brackets both to represent lists (e.g., $[K]$ in Section 2) and to represent closed intervals ($[0,1]$ in Section 5). \n\n     * Furthermore, some mathematical expressions are used without any definition. For example, ㅅthey did not define exactly what $[K]$ means. This can only be inferred from context. This issue also applies to the Truncate function in (ii) of equation (3). The exact mathematical expression of $f_\\gamma$ in Section 5 is not specified. Similarly, $\\gamma_{\\theta \\rightarrow \\theta_C}(t)$ is not a path, but rather a point on a path. \n\n       **Combined with the first weakness, these misuses of mathematical expressions will present a huge barrier for readers unfamiliar with the field.**", "clarity,_quality,_novelty_and_reproducibility": "Overall, this is a well-written paper. In contrast to the contribution of each section, the connection between the sections is insufficient.\n\n* Clarity: The Sections, except for Section 5, are clear, but their connections are lacking.\n* Quality: Mechanistic similarity (Section 3) and Mechanistic mode connectivity (Section 4) are great, but Connectivity-Based Fine-Tuning lacks motivation. (See the second issue of Weakness.)\n* Novelty: The novelty and contribution points of this paper are clear.\n* Reproducibility: Although I have not run it, I believe the submitted code will ensure reproducibility.", "summary_of_the_review": "This paper proposes a novel solution (Connectivity-Based Fine-Tuning) to an interesting problem (removing spurious attributes). While the concepts of Mechanistic Similarity and Mechanistic Mode Connectivity are exciting and inspiring, no discussion has been given on how to remove the spurious attribute. The paper will be helpful to readers in many fields with revisions for natural connections between sections.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1669292494555}, {"id": "jiyiJuNMUC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5987/Reviewer_fq5i"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work associates the notion of linear mode connectivity of two models with the functional similarity (mechanistic similarity) of the two models. They argue that the linear mode connectivity of the two models indicates they both have inherited potentially spurious/undesirable representations. Furthermore, they argue that naive fine-tuning cannot remove these spurious representations. Based on this connection, the authors propose a new algorithm referred to as connectivity-based fine tuning (CBFT), which encourages two properties for optimizing the parameters:\n\n- Removal of the linear path between to spurious model\n- Encourage functional similarity of the current model across spurious and non-spurious datasets.", "review_text": "There needs to be more work in understanding mode connectivity's functional utility, which this paper meaningfully explores. As it stands, mode connectivity (linear or non-linear) is an empirical phenomenon that is not well understood. This work explores possible avenues for improving model generalization based on this phenomenon. The main caveat to this work is the clarity is made somewhat opaque by the presentation of grids of experiments. The idea would be communicated much better if those results were condensed down.", "strengths": "Strengths:\n- The hypothesis this paper proposes about the optimization landscape and, specifically, the functional relationship with the energy barrier is interesting and an important area of better understanding the utility of mode connectivity. \n- Thorough appendix with corresponding loss curves. (This reviewer's opinion: loss curves are easier to interpret for mode connectivity than accuracy curves)\n\nWeaknesses:\n- Unclear on which of the two steps in CBFT contributes to the improvements. An ablation analysis of each step would help disentangle the contribution.\n- The definition of mode connectivity shown in Figure 6 seems to make mode connectivity and mechanistic similarity expected and potentially uninteresting. The monotonicity of the evaluation curve going from \\theta_c to \\theta_{nc} should be expected since the transition is going from a model trained on that dataset to one trained on a different dataset and vice versa. Could the authors clarify this issue? Perhaps this reviewer has made an error in understanding Figure 6. A more interesting result would be a \"u\" shaped accuracy barrier. But on most plots in Figure 4, this does not appear, and most curves are the expected monotonic trend.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work associates the notion of linear mode connectivity of two models with the functional similarity (mechanistic similarity) of the two models. They argue that the linear mode connectivity of the two models indicates they both have inherited potentially spurious/undesirable representations. Furthermore, they argue that naive fine-tuning cannot remove these spurious representations. Based on this connection, the authors propose a new algorithm referred to as connectivity-based fine tuning (CBFT), which encourages two properties for optimizing the parameters:\n\n- Removal of the linear path between to spurious model\n- Encourage functional similarity of the current model across spurious and non-spurious datasets.", "strength_and_weaknesses": "Strengths:\n- The hypothesis this paper proposes about the optimization landscape and, specifically, the functional relationship with the energy barrier is interesting and an important area of better understanding the utility of mode connectivity. \n- Thorough appendix with corresponding loss curves. (This reviewer's opinion: loss curves are easier to interpret for mode connectivity than accuracy curves)\n\nWeaknesses:\n- Unclear on which of the two steps in CBFT contributes to the improvements. An ablation analysis of each step would help disentangle the contribution.\n- The definition of mode connectivity shown in Figure 6 seems to make mode connectivity and mechanistic similarity expected and potentially uninteresting. The monotonicity of the evaluation curve going from \\theta_c to \\theta_{nc} should be expected since the transition is going from a model trained on that dataset to one trained on a different dataset and vice versa. Could the authors clarify this issue? Perhaps this reviewer has made an error in understanding Figure 6. A more interesting result would be a \"u\" shaped accuracy barrier. But on most plots in Figure 4, this does not appear, and most curves are the expected monotonic trend.", "clarity,_quality,_novelty_and_reproducibility": "Overall, the paper is well-written. The figures could use some clarification:\n\nFigures with grids of plots were a bit dense to parse easily (e.g., Figure 4 and Figure 5).\n- Figure 3 diagrams are hard to understand, especially the grid of similarity descriptions. Unclear what the rows and columns of the grid of rectangles are illustrating.\n- Figure 4 is difficult to parse because column labels like \"Linear Permuted\" are not defined in the main text.\n\nReproducibility is unclear, but the authors mention they will release it during the rebuttal phase.", "summary_of_the_review": "There needs to be more work in understanding mode connectivity's functional utility, which this paper meaningfully explores. As it stands, mode connectivity (linear or non-linear) is an empirical phenomenon that is not well understood. This work explores possible avenues for improving model generalization based on this phenomenon. The main caveat to this work is the clarity is made somewhat opaque by the presentation of grids of experiments. The idea would be communicated much better if those results were condensed down.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667064110674}, {"id": "S1bfonL8Tb", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5987/Reviewer_bN6r"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies a new concept mechanistic mode-connectivity, which focusing on the mode-connectivity of models trained on different data distribution. Meanwhile, the paper also proposes connectivity based fine-tuning (CBFT), aiming at overriding existing mechanisms of a pretrained model by fine tuning it on a small out-of-distribution dataset.", "review_text": "The overall quality of this paper is good, where the analysis is comprehensive. Hence I recommend to accept this paper.", "strengths": "Pros:\n1. The paper is well-written and easy to follow, although minor grammar error also exist (e.g. 2nd paragraph line 3 \"to to\"). \n2. The introduced problem is quite interesting and well-motivated.\n3. The paper has conducted extensive experiment results to verify the effectiveness of the proposed hypothesis.\n\nCons:\n1. For figure 2, I'm confused by the last two rows of right figure. What is the difference between \"Rand. cue\" and \"Rand. Image w/ cue\"?\n2. For the augmented data, are they used along with original data for training? Or Models are only trained on the augmented data?", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies a new concept mechanistic mode-connectivity, which focusing on the mode-connectivity of models trained on different data distribution. Meanwhile, the paper also proposes connectivity based fine-tuning (CBFT), aiming at overriding existing mechanisms of a pretrained model by fine tuning it on a small out-of-distribution dataset.", "strength_and_weaknesses": "Pros:\n1. The paper is well-written and easy to follow, although minor grammar error also exist (e.g. 2nd paragraph line 3 \"to to\"). \n2. The introduced problem is quite interesting and well-motivated.\n3. The paper has conducted extensive experiment results to verify the effectiveness of the proposed hypothesis.\n\nCons:\n1. For figure 2, I'm confused by the last two rows of right figure. What is the difference between \"Rand. cue\" and \"Rand. Image w/ cue\"?\n2. For the augmented data, are they used along with original data for training? Or Models are only trained on the augmented data?", "clarity,_quality,_novelty_and_reproducibility": "1. The author doesn't provide code.\n2. The paper is somehow novel to me.", "summary_of_the_review": "The overall quality of this paper is good, where the analysis is comprehensive. Hence I recommend to accept this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666859217450}, {"id": "ad2c5LH8eCR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5987/Reviewer_iiL3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors aim to understand a model's sensitivity to spurious features through the lens of mode connectivity. First, they come up with a definition of the mechanistic similarity of different modes, on the basis of invariance to different interventions.  Then, they show a stronger form of connectivity between different modes, on the basis of mechanistic similarity. Finally, they conjecture that all modes with mechanistic similarity must be connected with a linear path up to permutations. They corroborate their conjecture with multiple experiments on CIFAR-10. Finally, they propose a method of connectivity-based fine-tuning to eliminate a model's sensitivity to spurious features.", "review_text": "Overall, my scores are slightly on the positive side. The paper aims to advance our understanding of mode connectivity and its connection to the decisions of a model. However, the proposed algorithm CBFT assumes that we must have access to both clean examples and examples with statistical cues, which begs the question of the relevance of CBFT compared to the direct training on clean examples. Moreover, experiments with stronger statistical cues will help strengthen the conjecture of the paper. ", "strengths": "The paper introduces a novel connection between mode connectivity and a model's sensitivity to spurious features. The conjecture is in itself very interesting and will be useful for the scientific community. Moreover, the authors present experiments in synthetic datasets to support their conjecture and show a simple Taylor expansion-based proof for local correctness. To round it off, they propose novel connectivity based fine-tuning, based on the conjecture, to remove a model's dependence on spurious cues.\n\n\nI have the following questions:\n\n(a) In Figures 4 and 5, shouldn't train/evaluation loss be plotted, as opposed to the train/test accuracy because definitions 4, 5, proposition 1, and conjecture 1 include loss function? Do we still see barriers in linear paths between mechanistically dissimilar models in figures 4 and 5, with loss function?\n\n(b) The CBFT algorithm (in 3) assumes that we have access to both a clean dataset and a dataset with statistical cues. Can the authors comment on whether one can identify examples from both groups in an unsupervised manner? If not, why should one use CBFT and not directly train the model (from scratch) on the clean examples in the dataset?\n\n(c) It's not clear how the authors run CBFT introduced in eq. (3) on real-world models.\n\n(i)  How do the authors optimize step (i)? Do they explicitly search over the neighborhood around $\\theta_C$ for one $\\theta$ that maximizes the loss function on the line connecting $\\theta$ and $\\theta_C$? Or do they conduct an optimization of the form $\\min_{\\theta, 0 \\le t \\le 1} | \\lambda_1 - \\mathcal{L}_{CE} ( \\hat{y}  ( \\mathcal{D}_C ;  (1-t) \\theta + t \\theta_C ) | $?\n\n(ii) Furthermore, do the authors follow an alternative minimization for steps (i) and (ii)?\n\n(iii) Do the authors include permutation in step (i)?\n\n(d)  All statistical cues introduced in the paper are very simple (linearly separable) cues.\n\nWill we observe the same invariances to stronger forms of statistical cues e.g. label specific translations and rotations of the images? An example is a classification setting comparing images of cats and dogs, with the images of cats rotated by 45 degrees and the images of dogs rotated by 135 degrees.  \n\n(e) The authors point out a very important characteristic of \"quadratic paths\" in the appendix: the evaluation performance of the model on the path will depend on the examples used to find the path, and hence may not show generalization across datasets/distributions. \n\nHowever, isn't this also true for the permutation learned for linear paths, since it involves finding the best permutation $\\pi$ with a greedy algorithm on a given set of examples? \n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors aim to understand a model's sensitivity to spurious features through the lens of mode connectivity. First, they come up with a definition of the mechanistic similarity of different modes, on the basis of invariance to different interventions.  Then, they show a stronger form of connectivity between different modes, on the basis of mechanistic similarity. Finally, they conjecture that all modes with mechanistic similarity must be connected with a linear path up to permutations. They corroborate their conjecture with multiple experiments on CIFAR-10. Finally, they propose a method of connectivity-based fine-tuning to eliminate a model's sensitivity to spurious features.", "strength_and_weaknesses": "The paper introduces a novel connection between mode connectivity and a model's sensitivity to spurious features. The conjecture is in itself very interesting and will be useful for the scientific community. Moreover, the authors present experiments in synthetic datasets to support their conjecture and show a simple Taylor expansion-based proof for local correctness. To round it off, they propose novel connectivity based fine-tuning, based on the conjecture, to remove a model's dependence on spurious cues.\n\n\nI have the following questions:\n\n(a) In Figures 4 and 5, shouldn't train/evaluation loss be plotted, as opposed to the train/test accuracy because definitions 4, 5, proposition 1, and conjecture 1 include loss function? Do we still see barriers in linear paths between mechanistically dissimilar models in figures 4 and 5, with loss function?\n\n(b) The CBFT algorithm (in 3) assumes that we have access to both a clean dataset and a dataset with statistical cues. Can the authors comment on whether one can identify examples from both groups in an unsupervised manner? If not, why should one use CBFT and not directly train the model (from scratch) on the clean examples in the dataset?\n\n(c) It's not clear how the authors run CBFT introduced in eq. (3) on real-world models.\n\n(i)  How do the authors optimize step (i)? Do they explicitly search over the neighborhood around $\\theta_C$ for one $\\theta$ that maximizes the loss function on the line connecting $\\theta$ and $\\theta_C$? Or do they conduct an optimization of the form $\\min_{\\theta, 0 \\le t \\le 1} | \\lambda_1 - \\mathcal{L}_{CE} ( \\hat{y}  ( \\mathcal{D}_C ;  (1-t) \\theta + t \\theta_C ) | $?\n\n(ii) Furthermore, do the authors follow an alternative minimization for steps (i) and (ii)?\n\n(iii) Do the authors include permutation in step (i)?\n\n(d)  All statistical cues introduced in the paper are very simple (linearly separable) cues.\n\nWill we observe the same invariances to stronger forms of statistical cues e.g. label specific translations and rotations of the images? An example is a classification setting comparing images of cats and dogs, with the images of cats rotated by 45 degrees and the images of dogs rotated by 135 degrees.  \n\n(e) The authors point out a very important characteristic of \"quadratic paths\" in the appendix: the evaluation performance of the model on the path will depend on the examples used to find the path, and hence may not show generalization across datasets/distributions. \n\nHowever, isn't this also true for the permutation learned for linear paths, since it involves finding the best permutation $\\pi$ with a greedy algorithm on a given set of examples? \n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, with the theoretical statements and the experimental findings being easy to understand. The authors propose an interesting novel conjecture, along with experiments on well-known models in synthetic datasets, which will be of immense benefit to the scientific community.", "summary_of_the_review": "Overall, my scores are slightly on the positive side. The paper aims to advance our understanding of mode connectivity and its connection to the decisions of a model. However, the proposed algorithm CBFT assumes that we must have access to both clean examples and examples with statistical cues, which begs the question of the relevance of CBFT compared to the direct training on clean examples. Moreover, experiments with stronger statistical cues will help strengthen the conjecture of the paper. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666671796788}, {"id": "_BPlyoqpyS", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5987/Reviewer_Y4FF"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the optimization landscape of models which have learned different mechanisms. Beyond the standard definition of linear mode connectivity, they introduce \"mechanistic mode connectivity\" which tests mode-connectivity under changes in the data distribution. Based on these findings they propose a fine-tuning mechanism to exit the basin associated with a spurious pre-training solution.", "review_text": "This paper is a thorough and interesting scientific investigation of the optimization landscape of mechanistically similar/dissimilar models. My concern is that the experimental set-up is detached from real world use cases of pre-training/fine-tuning.", "strengths": "Strengths: Thorough scientific investigation of mode connectivity/mechanistic mode connectivity. Definitions, propositions, and proposed methods are well motivated, and the proposed CBFT method demonstrates good empirical performance.\n\nWeaknesses: I very much enjoyed the paper, and the following concern may reflect a bias perspective, however I thought it was still worth raising:  My main concern is with the experimental set-up. While interesting from a scientific perspective, I'm not sure that the issues addressed in this paper are interesting practically. As such, I'd recommend making this more clear in the paper. As one concrete instantiation of this concern, the paper claims: \"We have demonstrated a significant limitation of naive fine-tuning: it can fail to eliminate spurious patterns learned during pre-training.\" However, to me the point of pre-training is to pre-train on all the data you can. We see this in LM pre-training on trillions of tokens, or recently in vision models with billion scale datasets such as LAION. Accordingly, the setting studied in this paper seems like the experimental set-up in this paper may be a bit contrived. For instance, my guess is that CBFT would actually make the model less robust if applied to fine-tuning CLIP (e.g., this may be of interest: https://arxiv.org/abs/2109.01903). In short, to me the fact that pre-training biases the fine-tuning solution seems to be a feature, not a bug, of pre-training.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies the optimization landscape of models which have learned different mechanisms. Beyond the standard definition of linear mode connectivity, they introduce \"mechanistic mode connectivity\" which tests mode-connectivity under changes in the data distribution. Based on these findings they propose a fine-tuning mechanism to exit the basin associated with a spurious pre-training solution.", "strength_and_weaknesses": "Strengths: Thorough scientific investigation of mode connectivity/mechanistic mode connectivity. Definitions, propositions, and proposed methods are well motivated, and the proposed CBFT method demonstrates good empirical performance.\n\nWeaknesses: I very much enjoyed the paper, and the following concern may reflect a bias perspective, however I thought it was still worth raising:  My main concern is with the experimental set-up. While interesting from a scientific perspective, I'm not sure that the issues addressed in this paper are interesting practically. As such, I'd recommend making this more clear in the paper. As one concrete instantiation of this concern, the paper claims: \"We have demonstrated a significant limitation of naive fine-tuning: it can fail to eliminate spurious patterns learned during pre-training.\" However, to me the point of pre-training is to pre-train on all the data you can. We see this in LM pre-training on trillions of tokens, or recently in vision models with billion scale datasets such as LAION. Accordingly, the setting studied in this paper seems like the experimental set-up in this paper may be a bit contrived. For instance, my guess is that CBFT would actually make the model less robust if applied to fine-tuning CLIP (e.g., this may be of interest: https://arxiv.org/abs/2109.01903). In short, to me the fact that pre-training biases the fine-tuning solution seems to be a feature, not a bug, of pre-training.", "clarity,_quality,_novelty_and_reproducibility": "In terms of clarity I thoroughly enjoyed Table 1 and Figure 6 which clearly communicate the overall findings in an easy to understand manner. The work is mostly original, and I recommend releasing code so that it is reproducible. ", "summary_of_the_review": "This paper is a thorough and interesting scientific investigation of the optimization landscape of mechanistically similar/dissimilar models. My concern is that the experimental set-up is detached from real world use cases of pre-training/fine-tuning.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1665688977769}], "openreview_url": "https://openreview.net/forum?id=NZZoABNZECq", "arxiv_id": "2211.08422", "paper_pdf": "papers/NZZoABNZECq.pdf", "paper_pdf_sha256": "f6f9f6400447f0b46a4ff0c23e06618788c881eaa879848ea590487db1b54620", "paper_pdf_bytes": 5743017, "paper_pdf_source": "openreview", "code_url": "https://github.com/EkdeepSLubana/MMC", "code_repository": "EkdeepSLubana/MMC", "code_commit": "f884ca780441dbd1b7cee0393c9df13950e7f13b", "code_archive": "repos/NZZoABNZECq.zip", "code_archive_sha256": "b8bf5354a3d3ee2013fedd231aa43a9771eb3e3864728830b0d5435ead8d78fb", "code_archive_bytes": 21316, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 32, "github_languages": {"Python": 84265}, "github_archived": false, "github_pushed_at": "2023-07-14T16:22:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mechanistic-mode-connectivity"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_ysluXvD1M", "year": 2022, "status": "rejected", "title": "Equal Experience in Recommender Systems", "authors": ["Jaewoong Cho", "Moonseok Choi", "Changho Suh"], "authorids": ["~Jaewoong_Cho1", "~Moonseok_Choi1", "~Changho_Suh1"], "authors_source": "OpenReview API", "abstract": "We explore the fairness issue that arises in recommender systems. Biased data due to inherent stereotypes of particular groups (e.g., male students' average rating on mathematics is often higher than that on humanities, and vice versa for females) may yield a limited scope of suggested items to a certain group of users. Our main contribution lies in the introduction of a novel fairness notion (that we call equal experience), which can serve to regulate such unfairness in the presence of biased data. The notion captures the degree of the equal experience of item recommendations across distinct groups. We propose an optimization framework that incorporates the fairness notion as a regularization term, as well as introduce computationally-efficient algorithms that solve the optimization. Experiments on synthetic and benchmark real datasets demonstrate that the proposed framework can indeed mitigate such unfairness while exhibiting a minor degradation of recommendation accuracy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "TECFkWhFGda", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2594/Reviewer_zEp7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper defines a fairness notion for a recommender system that is based on the independence of preference predictions to the user group and the item groups. The paper defines a mutual information-based mathematical expression to measure unfairness, called equal experience metric, and then optimize it along with a matrix factorization-based collaborative filtering.", "review_text": "Strengths: The result due to chain rule is pretty interesting and neat in that joint independence of preference prediction user-item groups ensures multiple independence notions together. The authors do a good job of giving examples of the definition in terms of all of the independence notions.\n\nEmpirically, the method proposed in the paper gives good fairness performance on all fairness metrics compared (except one), while the specific papers only specialize in reducing one form of unfairness at a time.\n\nWeaknesses: \nThe normative goal of the fairness notion is not well-motivated in the paper. From what I understand, making the preference prediction independent of both the item group and the user group means that different base rates of preference for item groups among different user groups should not mean different rates of recommendation. The authors give an example of science vs literature preference for male and female users, where according to this measure, science and literature would be recommended at the rate at which they appear in the item set (because of the independence between Y~ and Z_item). In my opinion, this notion needs to be further motivated and included with a proper discussion about the scope of what item and user groups could mean. For example, if there are more literature courses to be recommended than science courses, they might not have the same capacity or preference among the general user group as science courses. While it may make sense to recommend science and literature equally to male and female users, wouldn't recommending science and literature at proportional rates only make sense when they are equally preferred in the general population?\n\nIn the spirit of the above question, I would like the authors to elaborate on their criticism of fairness notions based on the difference in recommendation accuracies (Paragraph 2 of Introduction). However, later in the future work (Sec 5), the authors seem to advocate for Equalized odds fairness notion which is an accuracy-based notion.\n\nThe setup uses Y>threshold as a preference prediction isn't the most used form of recommendation, while something like the relative ranking/ordering of the items is. The paper does not talk about how the defined fairness metrics would apply if predictions were used in terms of ranking the different items for each user/user type.\n\nIn the experiments section, would it also make sense to use an importance weighting approach (e.g. [1]) to tackle the selection bias created by the data generation process? [1] Schnabel, Tobias, et al. \"Recommendations as treatments: Debiasing learning and evaluation.\" ICML, 2016.\n\nMinor:\n- In paragraph 2 of the Introduction, the authors say: “female students exhibit low ratings on math and science subjects due to …. sampling bias”. It is not clear if the term \"ratings\" means the predictions or scores observed in a dataset. However, the other factor of societal/cultural differences mentioned by the authors does explain the observation well. Perhaps, the authors meant the lack of ratings instead of lower ratings.\n- In related work, some literature from debiasing word embeddings could be relevant to include because those works also include an independence-based notion between embeddings of occupations and gender-specific words. In the current work, since matrix completion is used as the collaborative filtering method, it also corresponds to user and item embeddings and the notions of orthogonality and independence might be relevant to mention (or describe).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper defines a fairness notion for a recommender system that is based on the independence of preference predictions to the user group and the item groups. The paper defines a mutual information-based mathematical expression to measure unfairness, called equal experience metric, and then optimize it along with a matrix factorization-based collaborative filtering.", "main_review": "Strengths: The result due to chain rule is pretty interesting and neat in that joint independence of preference prediction user-item groups ensures multiple independence notions together. The authors do a good job of giving examples of the definition in terms of all of the independence notions.\n\nEmpirically, the method proposed in the paper gives good fairness performance on all fairness metrics compared (except one), while the specific papers only specialize in reducing one form of unfairness at a time.\n\nWeaknesses: \nThe normative goal of the fairness notion is not well-motivated in the paper. From what I understand, making the preference prediction independent of both the item group and the user group means that different base rates of preference for item groups among different user groups should not mean different rates of recommendation. The authors give an example of science vs literature preference for male and female users, where according to this measure, science and literature would be recommended at the rate at which they appear in the item set (because of the independence between Y~ and Z_item). In my opinion, this notion needs to be further motivated and included with a proper discussion about the scope of what item and user groups could mean. For example, if there are more literature courses to be recommended than science courses, they might not have the same capacity or preference among the general user group as science courses. While it may make sense to recommend science and literature equally to male and female users, wouldn't recommending science and literature at proportional rates only make sense when they are equally preferred in the general population?\n\nIn the spirit of the above question, I would like the authors to elaborate on their criticism of fairness notions based on the difference in recommendation accuracies (Paragraph 2 of Introduction). However, later in the future work (Sec 5), the authors seem to advocate for Equalized odds fairness notion which is an accuracy-based notion.\n\nThe setup uses Y>threshold as a preference prediction isn't the most used form of recommendation, while something like the relative ranking/ordering of the items is. The paper does not talk about how the defined fairness metrics would apply if predictions were used in terms of ranking the different items for each user/user type.\n\nIn the experiments section, would it also make sense to use an importance weighting approach (e.g. [1]) to tackle the selection bias created by the data generation process? [1] Schnabel, Tobias, et al. \"Recommendations as treatments: Debiasing learning and evaluation.\" ICML, 2016.\n\nMinor:\n- In paragraph 2 of the Introduction, the authors say: “female students exhibit low ratings on math and science subjects due to …. sampling bias”. It is not clear if the term \"ratings\" means the predictions or scores observed in a dataset. However, the other factor of societal/cultural differences mentioned by the authors does explain the observation well. Perhaps, the authors meant the lack of ratings instead of lower ratings.\n- In related work, some literature from debiasing word embeddings could be relevant to include because those works also include an independence-based notion between embeddings of occupations and gender-specific words. In the current work, since matrix completion is used as the collaborative filtering method, it also corresponds to user and item embeddings and the notions of orthogonality and independence might be relevant to mention (or describe).", "summary_of_the_review": "The paper tackles an important problem of fairness in recommender systems, and it defines a fairness notion that contains both the user groups and items groups. The metric defined is pretty straightforward to understand. The kernel-based probability density estimation to compute the difference between conditional and marginal probabilities is not a very standard method but it seems to work for the purposes of the experiment. Overall, the fairness notions are not very strongly motivated (as highlighted above), and the paper is missing the guidance around what groupings of items and users are meaningful for such a recommendation task.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636013968897}, {"id": "gHqJvBPby6N", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2594/Reviewer_85io"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The presented study argued that a fair recommendation should be independent of both user and item. Therefore, the study introduced a new fairness notion, i.e., equal experience, and further incorporated this fairness notion as a regularisation term in the matrix completion framework to construct a fair recommender system. The proposed method was evaluated with three datasets (one synthetic and two real datasets). ", "review_text": "The presented study did not analyse the diversity of the recommended items in the experimental results, which the new proposed fairness notion aims to improve. Also, it would further strengthen the study if the authors could analyse who could benefit from the proposed method, e.g., females who enjoyed action or crime movies but did not have much historical data, or others?\n\nAlso, as this study focused on the problem of rating prediction in recommender systems, it would good to also test the effectiveness of the proposed method in the ranking setting.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The presented study argued that a fair recommendation should be independent of both user and item. Therefore, the study introduced a new fairness notion, i.e., equal experience, and further incorporated this fairness notion as a regularisation term in the matrix completion framework to construct a fair recommender system. The proposed method was evaluated with three datasets (one synthetic and two real datasets). ", "main_review": "The presented study did not analyse the diversity of the recommended items in the experimental results, which the new proposed fairness notion aims to improve. Also, it would further strengthen the study if the authors could analyse who could benefit from the proposed method, e.g., females who enjoyed action or crime movies but did not have much historical data, or others?\n\nAlso, as this study focused on the problem of rating prediction in recommender systems, it would good to also test the effectiveness of the proposed method in the ranking setting.", "summary_of_the_review": "The new fairness notion introduced by the presented study can provide certain new knowledge and insights on how to construct a fair recommender system, and the effectiveness of the proposed recommendation method seemed to be supported by the experimental results. However, the study can be further strengthened if more in-depth analysis on the recommendation results can be provided.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635896265326}, {"id": "xQ_24pVfBoD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2594/Reviewer_dddS"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper is concerned with fairness in recommendations. Specifically, there are groups of users and groups of items. Previous work has modelled fairness as the constraint of all user groups having the same accuracy or as the prediction probability being independent of the item group or the user group. In this work, the notion is generalized so that the prediction is independent of both the item and user group. Optimization algorithms are shown to solve this problem along with experimental results.  ", "review_text": "I think that the paper solves an interesting problem and generalizes previous work. The main concern here seems to be with the weight of the contribution. Specifically, the new notion of fairness is not very complicated and seems to follow from previous work. Further, the optimization algorithms also seems to follow from previous work and are not novel. If there were significant technical issues it does not seem to me that they were well-emphasized in the paper.  \n\nHave the authors thought about including any theoretical guarantees. For example, convergence of the optimization. Or a guarantee that the fairness notion would hold at least approximately. \n\nSome minor representation issues:\n1-Eq(2) on page 3, shouldn't \\hat{M} belong to a constraint set such low rank.\n2-Eq(4) on page 3, I think (M_{ij}-\\hat{M}_{ij}) is missing a square. \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper is concerned with fairness in recommendations. Specifically, there are groups of users and groups of items. Previous work has modelled fairness as the constraint of all user groups having the same accuracy or as the prediction probability being independent of the item group or the user group. In this work, the notion is generalized so that the prediction is independent of both the item and user group. Optimization algorithms are shown to solve this problem along with experimental results.  ", "main_review": "I think that the paper solves an interesting problem and generalizes previous work. The main concern here seems to be with the weight of the contribution. Specifically, the new notion of fairness is not very complicated and seems to follow from previous work. Further, the optimization algorithms also seems to follow from previous work and are not novel. If there were significant technical issues it does not seem to me that they were well-emphasized in the paper.  \n\nHave the authors thought about including any theoretical guarantees. For example, convergence of the optimization. Or a guarantee that the fairness notion would hold at least approximately. \n\nSome minor representation issues:\n1-Eq(2) on page 3, shouldn't \\hat{M} belong to a constraint set such low rank.\n2-Eq(4) on page 3, I think (M_{ij}-\\hat{M}_{ij}) is missing a square. \n\n\n\n", "summary_of_the_review": "While the paper introduces a new meaningful notion of fairness. The notion closely follows the previous work. The optimization methods used also follow the previous work. Overall, this makes the contribution of the paper very incremental. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635893040118}, {"id": "7J0_qkqx0J-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2594/Reviewer_TDXy"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new notion “equal experience” to measure the fairness among groups in recommender systems, and provides an method to optimize this new notion based on matrix completion. Experiments demonstrate the effective of the proposed optimization framework.", "review_text": "Strengths:\n+ The paper is well written and easy to follow.\n+ The related works is well described, especially the representative ones. Fair comparisons on these representative baselines are provided in the experiments.\n+ It is interesting to utilize mutual information for fairness problems.\n\nWeaknesses:\n- I suggest to include DER to be a baseline.\n- I am wondering if there are any difficulties when facing more than group of users and items.\n- As shown in the experiments, the proposed method keeps receiving the worst RMSE performance. Also, from table 6, we can observe that when the proposed performs much better than  the baselines for fairness comparisons, it also performs much worse than others for recommendation comparisons. There is tradeoff between fairness and accuracy, which may need a more detailed analysis. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new notion “equal experience” to measure the fairness among groups in recommender systems, and provides an method to optimize this new notion based on matrix completion. Experiments demonstrate the effective of the proposed optimization framework.", "main_review": "Strengths:\n+ The paper is well written and easy to follow.\n+ The related works is well described, especially the representative ones. Fair comparisons on these representative baselines are provided in the experiments.\n+ It is interesting to utilize mutual information for fairness problems.\n\nWeaknesses:\n- I suggest to include DER to be a baseline.\n- I am wondering if there are any difficulties when facing more than group of users and items.\n- As shown in the experiments, the proposed method keeps receiving the worst RMSE performance. Also, from table 6, we can observe that when the proposed performs much better than  the baselines for fairness comparisons, it also performs much worse than others for recommendation comparisons. There is tradeoff between fairness and accuracy, which may need a more detailed analysis. ", "summary_of_the_review": "- Since I am not very familiar with recent work on fairness, the proposed notion \"equal experience\" seems novel to me. \n- The authors provide the optimization method for the \"equal experience\"  based on both traditional ML and deep learning techniques.\n- The results prove that the proposed method can receive better \"equal experience\" optimization performance compared to other baselines.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635875877173}], "openreview_url": "https://openreview.net/forum?id=_ysluXvD1M", "arxiv_id": "2210.05936", "paper_pdf": "papers/_ysluXvD1M.pdf", "paper_pdf_sha256": "978977a30f6dfa2eb2cea598118e7975e05f8fd4120b70a4db01f0f51cae1e90", "paper_pdf_bytes": 902580, "paper_pdf_source": "openreview", "code_url": "https://github.com/cjw2525/FairRec", "code_repository": "cjw2525/FairRec", "code_commit": "c57677441f9c0e2b7ab215d0b30a41fcc872fc47", "code_archive": "repos/_ysluXvD1M.zip", "code_archive_sha256": "70bf6fa32747f4465f9ffe59b1c46d9626133201de18be653c4ccc7a4f7e998b", "code_archive_bytes": 35852, "code_file_count": 12, "code_extensions": {".py": 11, ".ipynb": 1}, "github_disk_usage_kb": 30, "github_languages": {"Python": 20930, "Jupyter Notebook": 4490}, "github_archived": false, "github_pushed_at": "2021-10-05T22:57:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/equal-experience-in-recommender-systems-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-Qaj4_O3cO", "year": 2021, "status": "rejected", "title": "DCT-SNN: Using DCT to Distribute Spatial Information over Time for Learning Low-Latency Spiking Neural Networks", "authors": ["Isha Garg", "Sayeed Shafayet Chowdhury", "Kaushik Roy"], "authorids": ["~Isha_Garg1", "~Sayeed_Shafayet_Chowdhury3", "~Kaushik_Roy1"], "authors_source": "OpenReview API", "abstract": "Spiking Neural Networks (SNNs) offer a promising alternative to traditional deep learning frameworks, since they provide higher \ncomputational efficiency due to event-driven information processing. SNNs distribute the analog values of pixel intensities into binary spikes over time. However, the most widely used input coding schemes, such as Poisson based rate-coding, do not leverage the additional temporal learning capability of SNNs effectively. Moreover, these SNNs suffer from high inference latency which is a major bottleneck to their deployment. To overcome this, we propose a scalable time-based encoding scheme that utilizes the Discrete Cosine Transform (DCT) to reduce the number of timesteps required for inference. DCT decomposes an image into a weighted sum of sinusoidal basis images. At each time step, a single frequency base, taken in order and modulated\nby its corresponding DCT coefficient, is input to an accumulator that generates spikes upon crossing a threshold. We use the proposed scheme to learn DCT-SNN, a low-latency deep SNN with leaky-integrate-and-fire neurons, trained using surrogate gradient descent based backpropagation. We achieve top-1 accuracy of 89.94%, 68.3% and 52.43% on CIFAR-10, CIFAR-100 and TinyImageNet, respectively using VGG architectures. Notably, DCT-SNN performs inference with 2-14X reduced latency compared to other state-of-the-art SNNs, while achieving comparable accuracy to their standard deep learning counterparts. The dimension of the transform allows us to control the number of timesteps required for inference. Additionally, we can trade-off accuracy with latency in a principled manner by dropping the highest frequency components during inference.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "jbLUg-MjuOS", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2670/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\n1.\tA novel coding scheme is proposed based on Discrete Cosine Transform (DCT) for efficient information expression in place of conventional Poisson distribution method. The required time-steps are 2-14x reduced compared with other conversion-SNNs or hybrid trained SNNs.\n2.\tDCT is data-independent while performing at par with PCA.\nCons:\n1.\tThe experimental results are not convincing enough.\n2.\tCorrectness Problem. I am afraid that the descriptions about the reconstruction of input image is wrong. The reverse transform should be a matrix multiplication instead of Hadamard product of the coefficients and the basis. For example, in Fig 1., X=Y_1*T_1+Y_2*T_2+…+Y_5*T_5. It results in conceptual errors in Fig1 & Fig2.\n3.\tThe specific equation of DCT should be with the discussion of the desirable properties and constraints in the section of encoding scheme to provide a clear picture of the method. \nAbout the experimental results:\n1.\tCould the authors try to provide some explanations for why DCT is able to outperform Poisson method or directly exposing the original image to the input spike neuron, since DCT’s reverse transform is a reconstruction of the original image over time?\n2.\tFewer time-steps with relatively lower accuracy is kind of confusing. I would like that the authors could further show the required time-steps for reaching strictly equal (or better) accuracy results with other SNN works (especially those directly trained). It is because that the trade-off between accuracy and the number of time-steps is natural in SNNs. For example, Rathi et al. (2020) could increase their VGG16’s performance from 91.13% to 92.02% by adding 100 time-steps on CIFAR-10. \nRathi, Nitin, et al. \"Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation.\" arXiv preprint arXiv:2005.01807 (2020).\n\n3.\tI notice that you cite Wu's article in related works, but there is a lack of comparison to it in later experiment part. In my opinion, the results of the paper and its sequel on time-steps for training SNNs from scratch are worth-noticing. I would like a more comprehensive and fair comparison of results.\nWu, Yujie, et al. \"Spatio-temporal backpropagation for training high-performance spiking neural networks.\" Frontiers in neuroscience 12 (2018): 331. \nWu, Yujie, et al. \"Direct training for spiking neural networks: Faster, larger, better.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019. \n\nClarity: The paper is fairly well-written.\nOriginality: DCT is a widely used transformation technique in signal processing and data compression, but this paper creatively explores it as a coding scheme in SNNs.\nSignificance: The proposed coding scheme might provide a new resolution to the high inference latency bottleneck in SNNs.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A new coding scheme based on DCT for SNNs, but not convincing enough.", "review": "Pros:\n1.\tA novel coding scheme is proposed based on Discrete Cosine Transform (DCT) for efficient information expression in place of conventional Poisson distribution method. The required time-steps are 2-14x reduced compared with other conversion-SNNs or hybrid trained SNNs.\n2.\tDCT is data-independent while performing at par with PCA.\nCons:\n1.\tThe experimental results are not convincing enough.\n2.\tCorrectness Problem. I am afraid that the descriptions about the reconstruction of input image is wrong. The reverse transform should be a matrix multiplication instead of Hadamard product of the coefficients and the basis. For example, in Fig 1., X=Y_1*T_1+Y_2*T_2+…+Y_5*T_5. It results in conceptual errors in Fig1 & Fig2.\n3.\tThe specific equation of DCT should be with the discussion of the desirable properties and constraints in the section of encoding scheme to provide a clear picture of the method. \nAbout the experimental results:\n1.\tCould the authors try to provide some explanations for why DCT is able to outperform Poisson method or directly exposing the original image to the input spike neuron, since DCT’s reverse transform is a reconstruction of the original image over time?\n2.\tFewer time-steps with relatively lower accuracy is kind of confusing. I would like that the authors could further show the required time-steps for reaching strictly equal (or better) accuracy results with other SNN works (especially those directly trained). It is because that the trade-off between accuracy and the number of time-steps is natural in SNNs. For example, Rathi et al. (2020) could increase their VGG16’s performance from 91.13% to 92.02% by adding 100 time-steps on CIFAR-10. \nRathi, Nitin, et al. \"Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation.\" arXiv preprint arXiv:2005.01807 (2020).\n\n3.\tI notice that you cite Wu's article in related works, but there is a lack of comparison to it in later experiment part. In my opinion, the results of the paper and its sequel on time-steps for training SNNs from scratch are worth-noticing. I would like a more comprehensive and fair comparison of results.\nWu, Yujie, et al. \"Spatio-temporal backpropagation for training high-performance spiking neural networks.\" Frontiers in neuroscience 12 (2018): 331. \nWu, Yujie, et al. \"Direct training for spiking neural networks: Faster, larger, better.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019. \n\nClarity: The paper is fairly well-written.\nOriginality: DCT is a widely used transformation technique in signal processing and data compression, but this paper creatively explores it as a coding scheme in SNNs.\nSignificance: The proposed coding scheme might provide a new resolution to the high inference latency bottleneck in SNNs.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604021871700}, {"id": "DkON7Q4s4_o", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2670/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The scheme proposed breaks down the information in a block of an image into orthogonal basis functions (DCT is used) to make a progressively better reconstruction of the original image block with the addition of more basis functions used (like an nth order Taylor expansion).  The increasing spatial frequency components are known to be perceptually less sensitive (they need to include this) in images, so the low freq components can be presented first.  Each freq component is encoded into spikes sequentially, thereby staging the more perceptually important information first, with less important info coming later.  This reorders the presentation of information to allow a tradeoff of image quality with time/latency.\n\nI think the solution proposed is well-founded and will indeed mitigate the latency problem for spiking neural networks.  However, this feels a bit more like an engineering solution to a specific problem rather than a new concept.  I do like the injection of methods from other fields like image/video compression; it often feels that the deep learning field rediscovers things that have been uncovered years ago in other fields.  I see that as the main value of the paper in addition to helping to make spiking neural networks a POSSIBLE viable solution to edge deployment.  \n\nSection 1: I don’t think it’s a strongly supported claim that deep learning architectures are unsuitable for edge deployment.  There are plenty in deployment and there are new processors (Movidius, Mythic, etc) that can handle these computations for real-time applications.  I’d suggest a softer language there.  This does weaken the motivation for the paper though.\n\nSection 1, second paragraph:  typo: Thy -> The\n\nSection 3.2: On constraints for the transforms.  Did the authors consider Integer Transform (IT)?  This is used in MPEG/AVC.  It is a reversible transform that is an integer simplification of the DCT.  Given that the point of the paper is to decrease latency and computing requirements for edge deployments, this could help.\n\nSection 3.2: The authors do a good job of sweeping performance for different block sizes.\n\nFigure 5:  Isn’t it an obvious result that more time steps are required for Poisson vs DCT?  There simply aren’t enough bins to sum over to have a result until a certain point.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper discusses a method to make spiking networks more relevant to latency-sensitive applications on the edge. I believe the authors' method is relevant to this problem, but doesn't feel like it uncovers anything fundamentally new.  It is an engineering solution to a specific problem; and spiking networks are not a widely used method at the edge currently.  ", "review": "The scheme proposed breaks down the information in a block of an image into orthogonal basis functions (DCT is used) to make a progressively better reconstruction of the original image block with the addition of more basis functions used (like an nth order Taylor expansion).  The increasing spatial frequency components are known to be perceptually less sensitive (they need to include this) in images, so the low freq components can be presented first.  Each freq component is encoded into spikes sequentially, thereby staging the more perceptually important information first, with less important info coming later.  This reorders the presentation of information to allow a tradeoff of image quality with time/latency.\n\nI think the solution proposed is well-founded and will indeed mitigate the latency problem for spiking neural networks.  However, this feels a bit more like an engineering solution to a specific problem rather than a new concept.  I do like the injection of methods from other fields like image/video compression; it often feels that the deep learning field rediscovers things that have been uncovered years ago in other fields.  I see that as the main value of the paper in addition to helping to make spiking neural networks a POSSIBLE viable solution to edge deployment.  \n\nSection 1: I don’t think it’s a strongly supported claim that deep learning architectures are unsuitable for edge deployment.  There are plenty in deployment and there are new processors (Movidius, Mythic, etc) that can handle these computations for real-time applications.  I’d suggest a softer language there.  This does weaken the motivation for the paper though.\n\nSection 1, second paragraph:  typo: Thy -> The\n\nSection 3.2: On constraints for the transforms.  Did the authors consider Integer Transform (IT)?  This is used in MPEG/AVC.  It is a reversible transform that is an integer simplification of the DCT.  Given that the point of the paper is to decrease latency and computing requirements for edge deployments, this could help.\n\nSection 3.2: The authors do a good job of sweeping performance for different block sizes.\n\nFigure 5:  Isn’t it an obvious result that more time steps are required for Poisson vs DCT?  There simply aren’t enough bins to sum over to have a result until a certain point.  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603930996205}, {"id": "wABm5UlB8_R", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2670/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an encoding method based on the Discrete Cosine Transform (DCT) for Spiking Neural Network (SNN). The key idea is to decompose an image into different frequency components and feed them to the SNN sequentially. Compared to the Poisson coding method used in most SNN studies, the proposed encoding method significantly decreases the latency that the SNN needs for image classification while having minimal accuracy decease. \n \nHighlights:\n\n1. The idea of using DCT for input spike encoding is novel and has great potential. One of the problems that prevent the SNN from using fewer inference timesteps is the ineffectiveness of encoding input information. Using DCT, the method can potentially filter out less important information and more effectively encode the information in limited timesteps (as shown in Fig. 6 and Fig. 8 in the paper).\n\n2. The paper doesn't directly learn in the frequency domain generated from DCT. Instead, it reverse transforms the DCT result back to the spatial domain and spreads it into different timesteps of the SNN. By doing so, the spike encoding gives more importance to the low-frequency information in the image. This is desirable because low-frequency information is more important than high-frequency information in the image for classification.\n \nConcerns:\n\n1. The paper lacks experiments to show that DCT directly contributes to the decrease of timesteps for classification. Although comparisons with earlier SNN works that use Poisson encoding are shown, there is a lack of comparison with any SNN methods that directly convert pixel values into spikes using IF neurons and threshold selection. Thus, the existing experiments are not sufficient to exclude the possibility that the latency decrease is not due to DCT. The reviewer suggests conducting additional experiments for this.\n\n2. While the proposed method only focuses on the input encoding of SNN, many recent papers target new training methods (such as [Jibin Wu et al, 2019], [Sen Lu et al, 2020]) that also result in significant latency decrease of SNN for image classification. The paper lacks experiments to compare the performance with these more recent results. The reviewer suggests conducting additional experiments for this.\n\n3. The paper claims that the proposed method has better performance than ANNs trained on DCT coefficients. However, this is not a fair comparison since their encodings are different. The ANNs trained on DCT coefficients (such as [Max Ehrlich et al, 2019]) follow the same procedure as JPEG compression. The encoding uses non-overlapping 8x8 blocks, and the ANNs directly learn from the JPEG transformed domain. If the paper wants to compare with these ANNs, it needs experiments using the same encoding input.\n\n4. The example in Fig. 1 for the reverse transformation is not the same as the source code (spike_model_vgg9_submit.py: Line 233). In section 3.1, the paper performs inverse transform by doing an element-wise multiplication between the transformed vector Y and each frequency basis in the transformation matrix, and claims the same method generalizes to the 2D case. However, the source code performs inverse transformation using the particular element in the transformed matrix Y (it's now a matrix but not a vector) corresponding to the specific frequency basis. Thus, the explanation in the paper contradicts the implementation. The reviewer thinks the example given in Fig. 1 is mathematically incorrect.\n\n5. The proposed method spreads the reverse transformed image into different timesteps, and each timestep corresponds to a particular frequency. However, the paper doesn't explore other possible approaches for spreading the information. For example, the encoding method can use multiple subsequent timesteps for a particular frequency or only present intermittent frequencies. The lack of in-depth analysis of the proposed method possibly prevents the paper from fully exploring the potential of the use of DCT encoding for SNN.\n\nMinor Comments:\n\n1. What is the meaning of \"ov\" in Fig. 4? The reviewer thinks it means \"overlap\". However, it needs to be explained in the text or figure caption.\n\n2. In Table 2, it's not clear whether DNN-d uses DCT coefficients or the reverse transformed image. If the DNN-d uses the DCT coefficient, is there any change to the ConvNet since the DCT destroys the block's spatial relationships?\n\n3. In Table 2, the SNN-d results for TinyImageNet are missing. Is there any reason for that?\n\n\nJibin Wu et al, 2019, A Tandem Learning Rule for Effective Training and Rapid Inference of Deep Spiking Neural Networks\n\nSen Lu et al, 2020, Exploring the Connection Between Binary and Spiking Neural Networks\n\nMax Ehrlich et al, 2019, Deep residual learning in the jpeg transform domain\n\n Since most of my primary concerns are resolved, I have updated my rating based on the revised version.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel input spike encoding method for SNN but lacks crucial experiments to support the conclusion", "review": "This paper proposes an encoding method based on the Discrete Cosine Transform (DCT) for Spiking Neural Network (SNN). The key idea is to decompose an image into different frequency components and feed them to the SNN sequentially. Compared to the Poisson coding method used in most SNN studies, the proposed encoding method significantly decreases the latency that the SNN needs for image classification while having minimal accuracy decease. \n \nHighlights:\n\n1. The idea of using DCT for input spike encoding is novel and has great potential. One of the problems that prevent the SNN from using fewer inference timesteps is the ineffectiveness of encoding input information. Using DCT, the method can potentially filter out less important information and more effectively encode the information in limited timesteps (as shown in Fig. 6 and Fig. 8 in the paper).\n\n2. The paper doesn't directly learn in the frequency domain generated from DCT. Instead, it reverse transforms the DCT result back to the spatial domain and spreads it into different timesteps of the SNN. By doing so, the spike encoding gives more importance to the low-frequency information in the image. This is desirable because low-frequency information is more important than high-frequency information in the image for classification.\n \nConcerns:\n\n1. The paper lacks experiments to show that DCT directly contributes to the decrease of timesteps for classification. Although comparisons with earlier SNN works that use Poisson encoding are shown, there is a lack of comparison with any SNN methods that directly convert pixel values into spikes using IF neurons and threshold selection. Thus, the existing experiments are not sufficient to exclude the possibility that the latency decrease is not due to DCT. The reviewer suggests conducting additional experiments for this.\n\n2. While the proposed method only focuses on the input encoding of SNN, many recent papers target new training methods (such as [Jibin Wu et al, 2019], [Sen Lu et al, 2020]) that also result in significant latency decrease of SNN for image classification. The paper lacks experiments to compare the performance with these more recent results. The reviewer suggests conducting additional experiments for this.\n\n3. The paper claims that the proposed method has better performance than ANNs trained on DCT coefficients. However, this is not a fair comparison since their encodings are different. The ANNs trained on DCT coefficients (such as [Max Ehrlich et al, 2019]) follow the same procedure as JPEG compression. The encoding uses non-overlapping 8x8 blocks, and the ANNs directly learn from the JPEG transformed domain. If the paper wants to compare with these ANNs, it needs experiments using the same encoding input.\n\n4. The example in Fig. 1 for the reverse transformation is not the same as the source code (spike_model_vgg9_submit.py: Line 233). In section 3.1, the paper performs inverse transform by doing an element-wise multiplication between the transformed vector Y and each frequency basis in the transformation matrix, and claims the same method generalizes to the 2D case. However, the source code performs inverse transformation using the particular element in the transformed matrix Y (it's now a matrix but not a vector) corresponding to the specific frequency basis. Thus, the explanation in the paper contradicts the implementation. The reviewer thinks the example given in Fig. 1 is mathematically incorrect.\n\n5. The proposed method spreads the reverse transformed image into different timesteps, and each timestep corresponds to a particular frequency. However, the paper doesn't explore other possible approaches for spreading the information. For example, the encoding method can use multiple subsequent timesteps for a particular frequency or only present intermittent frequencies. The lack of in-depth analysis of the proposed method possibly prevents the paper from fully exploring the potential of the use of DCT encoding for SNN.\n\nMinor Comments:\n\n1. What is the meaning of \"ov\" in Fig. 4? The reviewer thinks it means \"overlap\". However, it needs to be explained in the text or figure caption.\n\n2. In Table 2, it's not clear whether DNN-d uses DCT coefficients or the reverse transformed image. If the DNN-d uses the DCT coefficient, is there any change to the ConvNet since the DCT destroys the block's spatial relationships?\n\n3. In Table 2, the SNN-d results for TinyImageNet are missing. Is there any reason for that?\n\n\nJibin Wu et al, 2019, A Tandem Learning Rule for Effective Training and Rapid Inference of Deep Spiking Neural Networks\n\nSen Lu et al, 2020, Exploring the Connection Between Binary and Spiking Neural Networks\n\nMax Ehrlich et al, 2019, Deep residual learning in the jpeg transform domain\n\n Since most of my primary concerns are resolved, I have updated my rating based on the revised version.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603722172347}, {"id": "e2Eu0Sbdc0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2670/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Authors applied this algorithm into several datasets such as CIFAR 10, CIFAR 100 and TinyImagenet, they argued the image classification accuracies onto these datasets were comparable. The experiments were sufficient but not well designed. The reported experimental results show some novelty and good performance.\n\nIn the early stage of the spiking neural networks, the encoding methods are very important, especially for the training process. As the authors said that the rated based encoding method brings much more time latency which is time consuming. Therefore, the topic of this work is critical. However I have some worries about the proposed methods, from my point of view, this work is just combing the DCT and ANN-SNN method, the novelty is significantly limited, but the idea is interesting. Then, the experimental results reported by this paper were not well designed. The authors argued that the rated based encoding methods are time consuming, but they just did not compare the much more temporal encoding methods in Table 3. And for me, CIFAR10, CIFAR 100 and MNIST are in the same quantity level, which means that you used a VGG net is waste of resource (VGG is too deep). I even did not know what parts of the final results works, the deep CNN based network architecture? The DCT encoding method? Or the ANN-SNN methods. Actually, ANN-SNN method is not a typical bio-inspired way to construct a SNN, I prefer to see you adopt the proposed method into a Tempotron or STDP based learning rule not a surrogate-gradient based rule. The computational efficiency is nice; especially the authors calculated the spike rate of each single layer, but if you just argued the proposed the method is energy consumption, you should at least consider the ANN training process, it is not a single trade-off between inference accuracy and latency. I have run the code from authors provided, the reproducibility is reliable.  \n\nAlso there are some writing errors, such as thy->they, -s, etc. and reference missing, such as these important works:\n1.\tAn FPGA Implementation of Deep Spiking Neural Networks for Low-Power and Fast Classification.\n2.\tDeep CovDenseSNN: A hierarchical event-driven dynamic framework with spiking neurons in noisy environment \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "his work utilizes DCT based encoding method into spiking neural networks. This encoding method combines the mathematics of DCT and bio-plausibility of spiking neurons (LIF) to enrich the correlations in the spatio-temporal domain of spikes, which is quite inspiring.", "review": "Authors applied this algorithm into several datasets such as CIFAR 10, CIFAR 100 and TinyImagenet, they argued the image classification accuracies onto these datasets were comparable. The experiments were sufficient but not well designed. The reported experimental results show some novelty and good performance.\n\nIn the early stage of the spiking neural networks, the encoding methods are very important, especially for the training process. As the authors said that the rated based encoding method brings much more time latency which is time consuming. Therefore, the topic of this work is critical. However I have some worries about the proposed methods, from my point of view, this work is just combing the DCT and ANN-SNN method, the novelty is significantly limited, but the idea is interesting. Then, the experimental results reported by this paper were not well designed. The authors argued that the rated based encoding methods are time consuming, but they just did not compare the much more temporal encoding methods in Table 3. And for me, CIFAR10, CIFAR 100 and MNIST are in the same quantity level, which means that you used a VGG net is waste of resource (VGG is too deep). I even did not know what parts of the final results works, the deep CNN based network architecture? The DCT encoding method? Or the ANN-SNN methods. Actually, ANN-SNN method is not a typical bio-inspired way to construct a SNN, I prefer to see you adopt the proposed method into a Tempotron or STDP based learning rule not a surrogate-gradient based rule. The computational efficiency is nice; especially the authors calculated the spike rate of each single layer, but if you just argued the proposed the method is energy consumption, you should at least consider the ANN training process, it is not a single trade-off between inference accuracy and latency. I have run the code from authors provided, the reproducibility is reliable.  \n\nAlso there are some writing errors, such as thy->they, -s, etc. and reference missing, such as these important works:\n1.\tAn FPGA Implementation of Deep Spiking Neural Networks for Low-Power and Fast Classification.\n2.\tDeep CovDenseSNN: A hierarchical event-driven dynamic framework with spiking neurons in noisy environment \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603244043535}], "openreview_url": "https://openreview.net/forum?id=-Qaj4_O3cO", "arxiv_id": "2010.01795", "paper_pdf": "papers/-Qaj4_O3cO.pdf", "paper_pdf_sha256": "bec6fc58978939a8874f86faeab53a3bb0e5c50afe3d1f9890b10ceb6adde686", "paper_pdf_bytes": 1575748, "paper_pdf_source": "openreview", "code_url": "https://github.com/SayeedChowdhury/dct-snn", "code_repository": "SayeedChowdhury/dct-snn", "code_commit": "c0ef166a8e7df4185c7774900f381d23e87e54ca", "code_archive": "repos/-Qaj4_O3cO.zip", "code_archive_sha256": "9d8f51c6a8d81a8b062bac6d242c20b94dd76c189a3388ddbc903de20843e193", "code_archive_bytes": 25656, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 34, "github_languages": {"Python": 91108}, "github_archived": false, "github_pushed_at": "2021-09-30T16:30:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dct-snn-using-dct-to-distribute-spatial-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bkf4XgrKvS", "year": 2020, "status": "rejected", "title": "Unsupervised Learning of Graph Hierarchical Abstractions with Differentiable Coarsening and Optimal Transport", "authors": ["Tengfei Ma", "Jie Chen"], "authorids": ["tengfei.ma1@ibm.com", "chenjie@us.ibm.com"], "authors_source": "OpenReview API", "abstract": "Hierarchical abstractions are a methodology for solving large-scale graph problems in various disciplines. Coarsening is one such approach: it generates a pyramid of graphs whereby the one in the next level is a structural summary of the prior one. With a long history in scientific computing, many coarsening strategies were developed based on mathematically driven heuristics. Recently, resurgent interests exist in deep learning to design hierarchical methods learnable through differentiable parameterization. These approaches are paired with downstream tasks for supervised learning. In this work, we propose an unsupervised approach, coined \\textsc{OTCoarsening}, with the use of optimal transport. Both the coarsening matrix and the transport cost matrix are parameterized, so that an optimal coarsening strategy can be learned and tailored for a given set of graphs. We demonstrate that the proposed approach produces meaningful coarse graphs and yields competitive performance compared with supervised methods for graph classification.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HygY5iIScS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2208/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an unsupervised hierarchical approach for learning graph representations. The proposed architecture is constructed by unrolling k-steps of a parametrized algebraic multigrid approach for minimizing the Wasserstein metric between the graph and its representation. The node distance (transport cost) used in the Wasserstein metric is also learned as an L2 distance between the embeddings of some graph embedding function. The approach is compared against 6 other state of the art approaches on 5 graph classification tasks, showing significant improvements 4 of them.\n\nThe paper is reasonably well written, however, I think some of the explanations can be tightened further. Especially a lot on the background of AMG is not really that relevant, since the authors are not transferring technical results from AMG. Also, it seems like a better flow for presenting this argument might be to switch the order of sections 3.2.1 and 3.1.2. It looks like the main point is  that this architecture is trying to emulate iterative coarsened residual optimization of the Wasserstein metric between a graph and its representation. How the coarsening matrix is derived is more of a technical point (it looks like the results would be much more sensitive to a switch of metric than to a switch of parametrization for S). \n\nThe empirical results are quite intriguing. There are, however, natural and important questions left unanswered. First and foremost, how does the amount of downsampling (compression) compare between methods. How many parameters do different methods require? It would also be good to see what the baseline performance would have been without any input compression as to understand how close these approaches are to the upper bound.\n\nFinally, I think the main issue of this paper, is left unresolved, namely, what is the point of not having supervision from the downstream task. As a user of graph representations trying to solve some problem, the only thing I would want from my representation is to capture some notion of sufficient statistics that are small enough to be efficient and allow me to solve my problem. I would not necessarily care about how well the learned representation resembles the original graph unless I believed that my downstream task was hard to evaluate  and that it was very smooth in the Wasserstein metric. I read the paper multiple times, trying to find any discussion on this, but it seems that the fact that an unsupervised representation is a good thing is taken for granted. A point could at least be made using the same representation for different tasks experimentally. Or, perhaps, literally doing an AMG-type unpacking of the downstream task itself as a comparison. This would shed light on the question of whether the iterated residuals or the choice of distance is what's driving the observed results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes an unsupervised hierarchical approach for learning graph representations. The proposed architecture is constructed by unrolling k-steps of a parametrized algebraic multigrid approach for minimizing the Wasserstein metric between the graph and its representation. The node distance (transport cost) used in the Wasserstein metric is also learned as an L2 distance between the embeddings of some graph embedding function. The approach is compared against 6 other state of the art approaches on 5 graph classification tasks, showing significant improvements 4 of them.\n\nThe paper is reasonably well written, however, I think some of the explanations can be tightened further. Especially a lot on the background of AMG is not really that relevant, since the authors are not transferring technical results from AMG. Also, it seems like a better flow for presenting this argument might be to switch the order of sections 3.2.1 and 3.1.2. It looks like the main point is  that this architecture is trying to emulate iterative coarsened residual optimization of the Wasserstein metric between a graph and its representation. How the coarsening matrix is derived is more of a technical point (it looks like the results would be much more sensitive to a switch of metric than to a switch of parametrization for S). \n\nThe empirical results are quite intriguing. There are, however, natural and important questions left unanswered. First and foremost, how does the amount of downsampling (compression) compare between methods. How many parameters do different methods require? It would also be good to see what the baseline performance would have been without any input compression as to understand how close these approaches are to the upper bound.\n\nFinally, I think the main issue of this paper, is left unresolved, namely, what is the point of not having supervision from the downstream task. As a user of graph representations trying to solve some problem, the only thing I would want from my representation is to capture some notion of sufficient statistics that are small enough to be efficient and allow me to solve my problem. I would not necessarily care about how well the learned representation resembles the original graph unless I believed that my downstream task was hard to evaluate  and that it was very smooth in the Wasserstein metric. I read the paper multiple times, trying to find any discussion on this, but it seems that the fact that an unsupervised representation is a good thing is taken for granted. A point could at least be made using the same representation for different tasks experimentally. Or, perhaps, literally doing an AMG-type unpacking of the downstream task itself as a comparison. This would shed light on the question of whether the iterated residuals or the choice of distance is what's driving the observed results."}, "tcdate": 1572330385400}, {"id": "B1xsVn_aKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2208/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe paper proposes a differentiable coarsening approach for graph neural network (GNNs).\nTo this end, it is motivated by algebraic multigrid and optimal transport methods. \n\nGNNs is indeed an interesting line of research. And introducing coarsening into them, it a highly relevant step. However, there are some major downsides. First, some of the statements are a little but too strong. The paper starts with claiming that GNNs are competitive to graph kernels. But then for instance\n\nChristopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, Martin Grohe:\nWeisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks. AAAI 2019: 4602-4609\n\nshow that many (if not all) GNNs are equivalently expressive as the Weisefeiler-Lehman (WL) graph kernel. Hence, the competitiveness has to be qualified. Moreover, since you also employ graph convolutional networks for coarsening, you are also in the regime of this paper. Consequently, one should actually compare to WL, at least one should mention this connection. Actually, given that the datasets are not that large, one should run some statistical significance test. Moreover, if you check the paper above, they report much better results for PatchySan on MUTAG, better results on Protein for graph kernels, better results on IMDB-B using a hierarchical GNN approach, based on ideas of higher-order WL. \n\nNevertheless, indeed, the present paper shows that a differentiable pooling using WL kind of ideas is competitive to existing pooling approach. This is nice, but in the light of the work above, the novelty is unclear. This has to be clarified before publication. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "\nThe paper proposes a differentiable coarsening approach for graph neural network (GNNs).\nTo this end, it is motivated by algebraic multigrid and optimal transport methods. \n\nGNNs is indeed an interesting line of research. And introducing coarsening into them, it a highly relevant step. However, there are some major downsides. First, some of the statements are a little but too strong. The paper starts with claiming that GNNs are competitive to graph kernels. But then for instance\n\nChristopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, Martin Grohe:\nWeisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks. AAAI 2019: 4602-4609\n\nshow that many (if not all) GNNs are equivalently expressive as the Weisefeiler-Lehman (WL) graph kernel. Hence, the competitiveness has to be qualified. Moreover, since you also employ graph convolutional networks for coarsening, you are also in the regime of this paper. Consequently, one should actually compare to WL, at least one should mention this connection. Actually, given that the datasets are not that large, one should run some statistical significance test. Moreover, if you check the paper above, they report much better results for PatchySan on MUTAG, better results on Protein for graph kernels, better results on IMDB-B using a hierarchical GNN approach, based on ideas of higher-order WL. \n\nNevertheless, indeed, the present paper shows that a differentiable pooling using WL kind of ideas is competitive to existing pooling approach. This is nice, but in the light of the work above, the novelty is unclear. This has to be clarified before publication. "}, "tcdate": 1571814451038}, {"id": "ryldwtXVKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2208/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method to summarize a given graph based on the algebraic multigrid and optimal transport, which can be further used for the downstream ML tasks such as graph classification.\nAlthough the problem of graph summarization is a relevant task, there are a number of unclear points in this paper listed below:\n\n- In Section 3.1, the coarsening method has been proposed, which is said to be achieved by finding S such that A_C = S^T A S. \n    However, A_C is usually not binary for S \\in R^{n x m}, hence how to get the coarse graph G_C from A_C is not clear. Please carefully explain this point.\n- In the proposed method, coarse nodes should be selected beforehand. Is there any guideline of how to choose them?\n- In Section 3.2, optimal transport is introduced and the distance between G and G_C is measured via entropic optimal transport in Equation (4) or (7).\n    However, in Equation (4), a and b should come from the input G and G_C , and it is not clearly explained how to obtain them from the input.\n    Moreover, how to use the distance between G and G_C in the proposed coarsening method is also not clear. It seems that it is not used in Algorithm 1.\n- I do not understand why the k-step optimal transport distance is needed. Since it converges to the global optimum as k becomes large, it is usually enough to set k to be large enough.\n- In experiments, how is the proposed method used for graph classification?\n    Since the proposed method is for generating coarse graphs in an unsupervised manner, graph classification cannot be directly performed by itself.\n- In addition to the above issue, to assess the effectiveness of the the proposed method, the following experiment is recommended:\n    Fix some classifier and compare performance of graph classification for the original graphs and for the coarse graphs.\n- In the qualitative study in Section 4.4, while the authors discuss coarse nodes, they are just an input from the user and results are arbitrary. Hence such discussion is not informative.\n\nMinor comments:\n- What is \"X\" in Equation (2)?\n- I recommend to write domain for matrices when they used at the first time.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "This paper proposes a method to summarize a given graph based on the algebraic multigrid and optimal transport, which can be further used for the downstream ML tasks such as graph classification.\nAlthough the problem of graph summarization is a relevant task, there are a number of unclear points in this paper listed below:\n\n- In Section 3.1, the coarsening method has been proposed, which is said to be achieved by finding S such that A_C = S^T A S. \n    However, A_C is usually not binary for S \\in R^{n x m}, hence how to get the coarse graph G_C from A_C is not clear. Please carefully explain this point.\n- In the proposed method, coarse nodes should be selected beforehand. Is there any guideline of how to choose them?\n- In Section 3.2, optimal transport is introduced and the distance between G and G_C is measured via entropic optimal transport in Equation (4) or (7).\n    However, in Equation (4), a and b should come from the input G and G_C , and it is not clearly explained how to obtain them from the input.\n    Moreover, how to use the distance between G and G_C in the proposed coarsening method is also not clear. It seems that it is not used in Algorithm 1.\n- I do not understand why the k-step optimal transport distance is needed. Since it converges to the global optimum as k becomes large, it is usually enough to set k to be large enough.\n- In experiments, how is the proposed method used for graph classification?\n    Since the proposed method is for generating coarse graphs in an unsupervised manner, graph classification cannot be directly performed by itself.\n- In addition to the above issue, to assess the effectiveness of the the proposed method, the following experiment is recommended:\n    Fix some classifier and compare performance of graph classification for the original graphs and for the coarse graphs.\n- In the qualitative study in Section 4.4, while the authors discuss coarse nodes, they are just an input from the user and results are arbitrary. Hence such discussion is not informative.\n\nMinor comments:\n- What is \"X\" in Equation (2)?\n- I recommend to write domain for matrices when they used at the first time.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571203423716}], "openreview_url": "https://openreview.net/forum?id=Bkf4XgrKvS", "arxiv_id": "1912.11176", "paper_pdf": "papers/Bkf4XgrKvS.pdf", "paper_pdf_sha256": "55e4f939f73ecba04e48928a4cee48ae8c3d0b5a21ff929385f4df07214a8647", "paper_pdf_bytes": 449208, "paper_pdf_source": "openreview", "code_url": "https://github.com/matenure/OTCoarsening", "code_repository": "matenure/OTCoarsening", "code_commit": "ec047ed7a98087e81fcb6b8e7d3035fc4764e113", "code_archive": "repos/Bkf4XgrKvS.zip", "code_archive_sha256": "2d85e265cb0b6653605a13fbe5cf54b8e36d1b14c33d300814b55ca975203c47", "code_archive_bytes": 30304, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 39, "github_languages": {"Python": 52286}, "github_archived": false, "github_pushed_at": "2021-03-02T19:41:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unsupervised-learning-of-graph-hierarchical-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bklzkh0qFm", "year": 2019, "status": "rejected", "title": "Relational Graph Attention Networks", "authors": ["Dan Busbridge", "Dane Sherburn", "Pietro Cavallo", "Nils Y. Hammerla"], "authorids": ["dan.busbridge@gmail.com", "danesherbs@gmail.com", "p.cavallo85@gmail.com", "nils.hammerla@babylonhealth.com"], "authors_source": "OpenReview API", "abstract": "We investigate Relational Graph Attention Networks, a class of models that extends non-relational graph attention mechanisms to incorporate relational information, opening up these methods to a wider variety of problems. A thorough evaluation of these models is performed, and comparisons are made against established benchmarks. To provide a meaningful comparison, we retrain Relational Graph Convolutional Networks, the spectral counterpart of Relational Graph Attention Networks, and evaluate them under the same conditions. We find that Relational Graph Attention Networks perform worse than anticipated, although some configurations are marginally beneficial for modelling molecular properties. We provide insights as to why this may be, and suggest both modifications to evaluation strategies, as well as directions to investigate for future work.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "BkemLDvAn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper964/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presented a relational graph attention networks that could consider both node \nfeatures and relational information (edge features) to perform node-level and graph-level \nclassifications. The basic idea is to combine the graph attention networks (Veličković et \nal. 2017) and the relational graph networks (Schlichtkrull et al. 2018) to derive a hybrid \nnetworks. This paper is generally easy to follow and written clearly. Several experiments \nare conducted to demonstrate the performance of the proposed model. Although some promising \nresults have been achieved, I think there are several limitations regarding the novelty and \nsignificance of the proposed model. \n\ni) The proposed architecture is mainly adopted from the graph attention networks (Veličković \net al. 2017) and the relational graph networks (Schlichtkrull et al. 2018). Such a simple\ncombination is a good attempt to incorporate both node features and edge features but the\nnovelty is quite limited. \n\n\nii) In table 2, I don’t really see any promising results compared to baselines. There are \nlittle improvements over the baselines or even significantly worse. More importantly, \ncompared two schemes of this work, the ones with attentions are “almost” identical with ones\nwithout attentions, which implies that the proposed attentions mechanism is not really useful\nin practice. For most of newly proposed graph embedding algorithms, it is hard to convince \nit is indeed better without some significant improvements (at least 2% absolute accuracy more). \n\niii) For MUTAG dataset, the statistical information of this dataset is quite different from\nwhat I used to use. MUTAG is a standard dataset for testing graph-level classification for \nboth graph kernels and graph networks. MUTAG is a dataset of 188 mutagenic aromatic and \nheteroaromatic nitro compounds with 7 discrete node labels. Each chemical compound is labeled \naccording to whether it has mutagenic effect on the gram- negative bacterium Salmonnella \nTyphimurium. Could you explain why your MUTAG is now a single graph and is cast as node \nclassification problem?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good submission but not good enough", "review": "This paper presented a relational graph attention networks that could consider both node \nfeatures and relational information (edge features) to perform node-level and graph-level \nclassifications. The basic idea is to combine the graph attention networks (Veličković et \nal. 2017) and the relational graph networks (Schlichtkrull et al. 2018) to derive a hybrid \nnetworks. This paper is generally easy to follow and written clearly. Several experiments \nare conducted to demonstrate the performance of the proposed model. Although some promising \nresults have been achieved, I think there are several limitations regarding the novelty and \nsignificance of the proposed model. \n\ni) The proposed architecture is mainly adopted from the graph attention networks (Veličković \net al. 2017) and the relational graph networks (Schlichtkrull et al. 2018). Such a simple\ncombination is a good attempt to incorporate both node features and edge features but the\nnovelty is quite limited. \n\n\nii) In table 2, I don’t really see any promising results compared to baselines. There are \nlittle improvements over the baselines or even significantly worse. More importantly, \ncompared two schemes of this work, the ones with attentions are “almost” identical with ones\nwithout attentions, which implies that the proposed attentions mechanism is not really useful\nin practice. For most of newly proposed graph embedding algorithms, it is hard to convince \nit is indeed better without some significant improvements (at least 2% absolute accuracy more). \n\niii) For MUTAG dataset, the statistical information of this dataset is quite different from\nwhat I used to use. MUTAG is a standard dataset for testing graph-level classification for \nboth graph kernels and graph networks. MUTAG is a dataset of 188 mutagenic aromatic and \nheteroaromatic nitro compounds with 7 discrete node labels. Each chemical compound is labeled \naccording to whether it has mutagenic effect on the gram- negative bacterium Salmonnella \nTyphimurium. Could you explain why your MUTAG is now a single graph and is cast as node \nclassification problem?\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541465930873}, {"id": "B1lK2M8jnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper964/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work extends Schlichtkrull et al. (2018) by adding attention in two distinct ways: attention between pairs of nodes per relation, and attention between pairs of nodes averaged over all relations. The paper is well written and the equations easy to follow. The results are not strong. And, unfortunately, the model contribution currently is too modest. \n\nInductive task results: Wu et al. (2018) reports that for Tox21 (Duvenauld et al. 2015) is the best-performing approach. We should see the performance on other datasets  (e.g., some of the other datasets in Wu et al. (2018)).\n\nMy introduction suggestion: do not talk about Convolutional neural networks (CNNs). There is a *lot* of work on graph convolutional networks (GCNs). Reading it feels like reading about lattices when the work is about general graphs, and lattices provide no intuition about the proposed solution. \n\n--- After rebuttal ---\n\nStill not convinced of the value of the work to the community. Will keep my score the same.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Small incremental extension of existing work", "review": "This work extends Schlichtkrull et al. (2018) by adding attention in two distinct ways: attention between pairs of nodes per relation, and attention between pairs of nodes averaged over all relations. The paper is well written and the equations easy to follow. The results are not strong. And, unfortunately, the model contribution currently is too modest. \n\nInductive task results: Wu et al. (2018) reports that for Tox21 (Duvenauld et al. 2015) is the best-performing approach. We should see the performance on other datasets  (e.g., some of the other datasets in Wu et al. (2018)).\n\nMy introduction suggestion: do not talk about Convolutional neural networks (CNNs). There is a *lot* of work on graph convolutional networks (GCNs). Reading it feels like reading about lattices when the work is about general graphs, and lattices provide no intuition about the proposed solution. \n\n--- After rebuttal ---\n\nStill not convinced of the value of the work to the community. Will keep my score the same.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541264048643}, {"id": "rkx6mDnd37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper964/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a few variations on the RGCN model with adding attention to either within relations or across relations.\n\nUnfortunately the paper falls short in two main areas:\n\n- novelty: the additions proposed are small modifications to existing algorithms and there are other methods of attention on graphs which have been discussed in the paper but not directly compared to (e.g. this method builds heavily on Velickovic et al 2017)\n- impact: the results achieved in the experiments are very small improvements compared to the baseline of RGCN (~ +0.01 in two experiments and ~ -0.04 in another) and often these small variations in results can be compensated with better baselines training (e.g. better hyper-params, ...)\n\nHowever, on a positive note, the paper has been written very well and I really liked the frank discussion on page 8 about results on MUTAG dataset.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A slight reformulation of RGCN with attention mechanisms with mixed results on graph classification and node classification tasks.", "review": "The paper proposes a few variations on the RGCN model with adding attention to either within relations or across relations.\n\nUnfortunately the paper falls short in two main areas:\n\n- novelty: the additions proposed are small modifications to existing algorithms and there are other methods of attention on graphs which have been discussed in the paper but not directly compared to (e.g. this method builds heavily on Velickovic et al 2017)\n- impact: the results achieved in the experiments are very small improvements compared to the baseline of RGCN (~ +0.01 in two experiments and ~ -0.04 in another) and often these small variations in results can be compensated with better baselines training (e.g. better hyper-params, ...)\n\nHowever, on a positive note, the paper has been written very well and I really liked the frank discussion on page 8 about results on MUTAG dataset.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541093156727}], "openreview_url": "https://openreview.net/forum?id=Bklzkh0qFm", "arxiv_id": "1904.05811", "paper_pdf": "papers/Bklzkh0qFm.pdf", "paper_pdf_sha256": "975457c8dc4e7f6c8fe7d0219928033c7bc4b3be6349f9b95d46503613c46d4b", "paper_pdf_bytes": 2240922, "paper_pdf_source": "openreview", "code_url": "https://github.com/babylonhealth/rgat", "code_repository": "babylonhealth/rgat", "code_commit": "79a49a123a3ac0c25436504182b16d91dac8eb60", "code_archive": "repos/Bklzkh0qFm.zip", "code_archive_sha256": "cd8a29ed6969e3ddd0153aa345ea85195d329d51688d1e8c01b5d570b823e812", "code_archive_bytes": 173971, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 163, "github_languages": {"Python": 76946}, "github_archived": false, "github_pushed_at": "2022-07-13T16:27:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/relational-graph-attention-networks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WqSEVQRlUB", "year": 2026, "status": "rejected", "title": "Tokenize Image as a Set", "authors": ["Zigang Geng", "Mengde Xu", "Han Hu", "Shuyang Gu"], "authorids": ["~Zigang_Geng1", "~Mengde_Xu1", "~Han_Hu1", "~Shuyang_Gu1"], "authors_source": "OpenReview API", "abstract": "This paper proposes a new paradigm for image generation through set-based tokenization and modeling. Unlike conventional methods that serialize images into fixed-position latent codes with a uniform compression ratio, we introduce an unordered token set representation to dynamically allocate coding capacity based on regional semantic complexity. This TokenSet enhances global context aggregation and improves robustness against local perturbations. To address the critical challenge of modeling discrete sets, we devise a dual transformation mechanism that bijectively converts sets into fixed-length integer sequences while preserving summation constraints. Further, we propose Fixed-Sum Discrete Diffusion—the first framework to simultaneously handle discrete values, fixed sequence length, and summation invariance—enabling effective set distribution modeling. Experiments demonstrate our method's superiority in semantic-aware representation and generation quality. Our innovations, spanning novel representation and modeling strategies, advance visual generation beyond traditional sequential token paradigms.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "WRM6ez1Pmp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11186/Reviewer_xcn2"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces a novel image tokenization method and its corresponding image generation method. The main idea is to encode images as a set of permutation-invariant tokens. Accordingly, the authors propose FSDD to apply a diffusion process on the fixed-sum count vectors derived from the image tokens.", "review_text": "This paper introduces a novel image tokenization method and its corresponding image generation method. The main idea is to encode images as a set of permutation-invariant tokens. Accordingly, the authors propose FSDD to apply a diffusion process on the fixed-sum count vectors derived from the image tokens.", "strengths": "1. The idea of encoding an image as a set of tokens is well-motivated.\n2. The corresponding generation method of FSDD aligns well with the tokenization method.\n3. The authors provide some interesting experimental results, such as attention visualization and semantic clustering.", "weaknesses": "1. The main weakness is that the performance of the proposed method lags behind state-of-the-art models, such as TiTok, in both rFID and gFID. The performance potential of the proposed method has not been convincingly shown.\n2. Permutation-invariance is a good intuitive property for image tokenization. However, I find it difficult to conceptually relate it to reconstruction or generation performance.", "questions": "1. In Table 1, why does a higher percentage of overlapping tokens indicate greater robustness? If we assume a very poor tokenizer that encodes every image into the same set of tokens, the overlap percentage would be 100%. Therefore, I don’t think this result provides strong evidence for evaluating robustness.\n2. In Table 2, the linear probe results are interesting, as they indicate that the proposed method provides good semantic representations. Could the authors explain why models with a smaller number of tokens achieve better linear-probe performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel image tokenization method and its corresponding image generation method. The main idea is to encode images as a set of permutation-invariant tokens. Accordingly, the authors propose FSDD to apply a diffusion process on the fixed-sum count vectors derived from the image tokens.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The idea of encoding an image as a set of tokens is well-motivated.\n2. The corresponding generation method of FSDD aligns well with the tokenization method.\n3. The authors provide some interesting experimental results, such as attention visualization and semantic clustering.", "weaknesses": "1. The main weakness is that the performance of the proposed method lags behind state-of-the-art models, such as TiTok, in both rFID and gFID. The performance potential of the proposed method has not been convincingly shown.\n2. Permutation-invariance is a good intuitive property for image tokenization. However, I find it difficult to conceptually relate it to reconstruction or generation performance.", "questions": "1. In Table 1, why does a higher percentage of overlapping tokens indicate greater robustness? If we assume a very poor tokenizer that encodes every image into the same set of tokens, the overlap percentage would be 100%. Therefore, I don’t think this result provides strong evidence for evaluating robustness.\n2. In Table 2, the linear probe results are interesting, as they indicate that the proposed method provides good semantic representations. Could the authors explain why models with a smaller number of tokens achieve better linear-probe performance?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None.", "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761961302571}, {"id": "dFGdc4qDvV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11186/Reviewer_meuw"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces TokenSet, which represents images as unordered token sets instead of sequences, enabling permutation-invariant encoding and dynamic token allocation based on semantic complexity. A dual transformation converts token sets into fixed-length count vectors, and a Fixed-Sum Discrete Diffusion model enforces discrete, fixed-length, and constant-sum constraints for generation. Overall, the work shifts from positional tokenization to set-based representation with improved semantic awareness and robustness.", "review_text": "This paper introduces TokenSet, which represents images as unordered token sets instead of sequences, enabling permutation-invariant encoding and dynamic token allocation based on semantic complexity. A dual transformation converts token sets into fixed-length count vectors, and a Fixed-Sum Discrete Diffusion model enforces discrete, fixed-length, and constant-sum constraints for generation. Overall, the work shifts from positional tokenization to set-based representation with improved semantic awareness and robustness.", "strengths": "1.\tThe paper presents a novel set‐based tokenization and generation paradigm, which is conceptually fresh and offers a compelling alternative to conventional spatial or sequential visual token representations.\n2.\tThe proposed method demonstrates strong robustness to perturbations and noise, showing that the set‐based formulation effectively captures high‐level semantics and maintains stability under input corruption.", "weaknesses": "1.\tGeneration quality remains insufficient (Table 5).\nAlthough the method shows strong robustness and semantic consistency — consistent with its motivation — high-fidelity generation is still essential for evaluating visual tokenizers. Table 5 frames generation quality as future work, but without competitive results, it is difficult to fully validate the practical impact of the approach.\n2.\tToken count does not reflect real inference efficiency (Tables 3 & 4).\nThe model requires 25 sampling steps, making inference closer to diffusion than auto-regressive decoding. Thus, reporting only token count (Tables 3 & 4) may be misleading. Efficiency should be evaluated under comparable inference budgets (e.g., wall-clock time or compute) against diffusion and AR models to substantiate the claimed speed advantages.", "questions": "1.\tSince the model generates a token-count distribution vector, will increasing the number of tokens affect inference time under the same number of sampling steps? In other words, does the inference cost scale with token quantity in practice?\n2.\tThe paper reports results using 25 sampling steps. If the number of sampling steps increases, will the image quality improve further, similar to standard diffusion models?\n3.\tIn what scenarios would the proposed method have clear advantages over traditional autoregressive or diffusion models? Specifically, when does the set-based token generation outperform these approaches in terms of performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces TokenSet, which represents images as unordered token sets instead of sequences, enabling permutation-invariant encoding and dynamic token allocation based on semantic complexity. A dual transformation converts token sets into fixed-length count vectors, and a Fixed-Sum Discrete Diffusion model enforces discrete, fixed-length, and constant-sum constraints for generation. Overall, the work shifts from positional tokenization to set-based representation with improved semantic awareness and robustness.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tThe paper presents a novel set‐based tokenization and generation paradigm, which is conceptually fresh and offers a compelling alternative to conventional spatial or sequential visual token representations.\n2.\tThe proposed method demonstrates strong robustness to perturbations and noise, showing that the set‐based formulation effectively captures high‐level semantics and maintains stability under input corruption.", "weaknesses": "1.\tGeneration quality remains insufficient (Table 5).\nAlthough the method shows strong robustness and semantic consistency — consistent with its motivation — high-fidelity generation is still essential for evaluating visual tokenizers. Table 5 frames generation quality as future work, but without competitive results, it is difficult to fully validate the practical impact of the approach.\n2.\tToken count does not reflect real inference efficiency (Tables 3 & 4).\nThe model requires 25 sampling steps, making inference closer to diffusion than auto-regressive decoding. Thus, reporting only token count (Tables 3 & 4) may be misleading. Efficiency should be evaluated under comparable inference budgets (e.g., wall-clock time or compute) against diffusion and AR models to substantiate the claimed speed advantages.", "questions": "1.\tSince the model generates a token-count distribution vector, will increasing the number of tokens affect inference time under the same number of sampling steps? In other words, does the inference cost scale with token quantity in practice?\n2.\tThe paper reports results using 25 sampling steps. If the number of sampling steps increases, will the image quality improve further, similar to standard diffusion models?\n3.\tIn what scenarios would the proposed method have clear advantages over traditional autoregressive or diffusion models? Specifically, when does the set-based token generation outperform these approaches in terms of performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761834737506}, {"id": "3Pl3CXMHG6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11186/Reviewer_dft9"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes a novel image generation method using set-based tokenization, representing images as unordered token sets instead of fixed sequences. It adaptively allocates coding capacity by regional complexity and improves global context and robustness. A dual transformation converts sets into fixed-length sequences, and Fixed-Sum Discrete Diffusion models discrete sets effectively. Experiments show superior semantic and visual quality. However, it appears that the authors did not include the Ethics Statement and Reproducibility Statement in their paper, which could diminish its overall quality.", "review_text": "This paper proposes a novel image generation method using set-based tokenization, representing images as unordered token sets instead of fixed sequences. It adaptively allocates coding capacity by regional complexity and improves global context and robustness. A dual transformation converts sets into fixed-length sequences, and Fixed-Sum Discrete Diffusion models discrete sets effectively. Experiments show superior semantic and visual quality. However, it appears that the authors did not include the Ethics Statement and Reproducibility Statement in their paper, which could diminish its overall quality.", "strengths": "[1]. This paper introduces a novel set-based image tokenization method that eliminates positional bias and enhances semantic representation.\n\n[2]. It demonstrates strong robustness against local perturbations and image noise.\n\n[3]. The method dynamically allocates representational capacity according to regional semantic complexity, leading to improved efficiency.\n\n[4]. The experiments are sufficient and well-designed, and the figures are clear and easy to understand.\n\n[5]. The paper provides solid theoretical analysis, offering convincing proofs that support the proposed method.", "weaknesses": "[1] Reference formatting issue. The reference section requires further improvement. In particular, please correct the formatting of the conference name. It should be written as **NeurIPS** rather than **Neurips**.\n\n[2] Potential extension to text-to-image models. It is not entirely clear whether the proposed method can be effectively extended to text-to-image models. If time and computational resources allow, I suggest conducting an additional experiment by training a T2I model using JourneyDB, which could provide valuable empirical evidence of the method’s broader applicability. Even a preliminary result or discussion in the supplementary material would make the contribution more convincing.\n\n[3] I could not find the Ethics Statement or Reproducibility Statement in the current submission. These sections are required by the NeurIPS submission guidelines and play an important role in ensuring transparency, responsible research practices, and reproducibility. Please include these statements in the final version, outlining the ethical considerations of your work (e.g., data usage, societal impact) and providing clear details about how the experiments can be reproduced by other researchers.", "questions": "[1]. How does your method perform when applied to text-to-image models?\n\n[2]. Could you please indicate where the Ethics Statement and Reproducibility Statement are located?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel image generation method using set-based tokenization, representing images as unordered token sets instead of fixed sequences. It adaptively allocates coding capacity by regional complexity and improves global context and robustness. A dual transformation converts sets into fixed-length sequences, and Fixed-Sum Discrete Diffusion models discrete sets effectively. Experiments show superior semantic and visual quality. However, it appears that the authors did not include the Ethics Statement and Reproducibility Statement in their paper, which could diminish its overall quality.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "[1]. This paper introduces a novel set-based image tokenization method that eliminates positional bias and enhances semantic representation.\n\n[2]. It demonstrates strong robustness against local perturbations and image noise.\n\n[3]. The method dynamically allocates representational capacity according to regional semantic complexity, leading to improved efficiency.\n\n[4]. The experiments are sufficient and well-designed, and the figures are clear and easy to understand.\n\n[5]. The paper provides solid theoretical analysis, offering convincing proofs that support the proposed method.", "weaknesses": "[1] Reference formatting issue. The reference section requires further improvement. In particular, please correct the formatting of the conference name. It should be written as **NeurIPS** rather than **Neurips**.\n\n[2] Potential extension to text-to-image models. It is not entirely clear whether the proposed method can be effectively extended to text-to-image models. If time and computational resources allow, I suggest conducting an additional experiment by training a T2I model using JourneyDB, which could provide valuable empirical evidence of the method’s broader applicability. Even a preliminary result or discussion in the supplementary material would make the contribution more convincing.\n\n[3] I could not find the Ethics Statement or Reproducibility Statement in the current submission. These sections are required by the NeurIPS submission guidelines and play an important role in ensuring transparency, responsible research practices, and reproducibility. Please include these statements in the final version, outlining the ethical considerations of your work (e.g., data usage, societal impact) and providing clear details about how the experiments can be reproduced by other researchers.", "questions": "[1]. How does your method perform when applied to text-to-image models?\n\n[2]. Could you please indicate where the Ethics Statement and Reproducibility Statement are located?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761311556652}, {"id": "roZfgTKPsF", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11186/Reviewer_UHKS"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces an image tokenizer which is approximately invariant to token order. The authors empirically observe that this leads to tokenized representations that are even more decoupled from spatial/local bias common in other tokenizers (even in TiTok, to some extent, despite its 1D representation).\n\nThe authors then introduce “Fixed-Sum Discrete Diffusion” (FSDD), a diffusion-based generative model over multisets/bags of tokens, in order to leverage the tokenizer’s latent space for image generation.", "review_text": "This paper introduces an image tokenizer which is approximately invariant to token order. The authors empirically observe that this leads to tokenized representations that are even more decoupled from spatial/local bias common in other tokenizers (even in TiTok, to some extent, despite its 1D representation).\n\nThe authors then introduce “Fixed-Sum Discrete Diffusion” (FSDD), a diffusion-based generative model over multisets/bags of tokens, in order to leverage the tokenizer’s latent space for image generation.", "strengths": "While the method used to encourage this approximate permutation invariance is very simple (random token shuffling during training), results demonstrating semantically aligned token receptive fields, high linear probing accuracy, and robustness to noise in the input, are interesting and motivate the importance of such a seemingly minor tweak to TiTok.\n\nThe “Fixed-Sum Discrete Diffusion” approach is novel and is demonstrated to outperform certain baselines which do not exploit the permutation invariance.", "weaknesses": "Generation performance appears to be compromised compared to TiTok (FID ~5.1 compared to FID ~2.0), even when using the proposed FSDD. It is not clear whether this is due to limitations of the proposed generative model (FSDD) itself, or due to the learned bag-of-tokens representation having unfavorable structure for generation. Regarding the FSDD, some details and design choices still remain unclear to me and should be described and supported more explicitly (see questions below). Overall, further discussion, analysis and experiments related to generation could strengthen the paper.", "questions": "**Q1.** Section 3.3: I am a bit confused about the choice of $\\sigma_t$ in eq. (5), as it depends on the samples and is not dependent on $t$ like a usual noise schedule in flow matching/diffusion would. Is there a typo here, or if it is intended, could you clarify or provide references? I have a similar question about eq. (9) -- is $|\\mathbf{X}_1 - \\mathbf{X}_0|$ referring to some precomputed statistic, since $\\mathbf{X}_0$ is not available during sampling? It would also be good to elaborate on the t-dependent truncation ratio $f$.\n\n**Q2.** Section 3.3: Have you considered any other strategies for the initial distribution of $\\mathbf{X}_1$? (For example, starting with all counts assigned to a “mask” bin.) Additionally, the constraint that $\\sum \\mathbf{\\tilde X}_t = M$ for every $t$ might seem a bit restrictive. Have you considered any alternatives, such as modeling unconstrained histograms (using any of the existing categorical/multinomial diffusion approaches) and normalizing at the end, or any other modifications to expand the design space of your generative model?\n\n**Q3.** Section 4.3: Do the authors have any insight into why modeling the unordered representation appears to be “harder” than the 1D tokenization used by TiTok? Even with your proposed FSDD, the gFID is significantly lower than in the case of TiTok. Or is the worse generation performance purely due to limitations of FSDD?\n\nSection 4.3.1 / Table 4:\n* **Q4.** What is SetAR? I did not see it explained in the text. Does it refer to autoregressive modeling of the token histogram/“dual transformation”?\n* **Q5.** Are the authors familiar with autoregressive set prediction similar to [1]? It would be interesting to try this or a similar approach performing autoregressive set prediction without the dual transformation.\n\n**Q6.** In Section 4.2.1, rather than a qualitative demonstration on a single example, it would be better to provide quantitative metrics. In particular, to better understand the effectiveness of the train-time shuffling in encouraging permutation invariance, could the authors provide e.g. a histogram across the whole validation set of the error between images reconstructed from two different random token orderings or MSE/MAE with stddev/quantile statistics?\n\n**Q7.** I appreciate the inclusion of MAE in the linear probing results from Table 2, and it would be great if you could also include a comparison between an ablation of your tokenizer trained without shuffling to further support the claim that tokens from permutation invariant representations are semantically richer. \n\n---\n\nMinor questions/comments:\n* Use of “set” nomenclature: Referring to the representation as a “multiset” or “bag” from the beginning would be clearer, rather than leaving this as a footnote. In my own reading of the manuscript, I did not notice the footnote until explicitly searching for the word “multiset”, as I was a bit confused by the inaccurate use of the word “set”. Similarly, I would encourage the authors to consider altering the title to avoid being technically incorrect.\n* 3.1, L148: What does “partial” permutation sampling mean? Is it something other than simply shuffling the tokens? \n* 1, L069: I did not understand what is meant by “dynamic token allocation based on semantic complexity”, maybe this statement could be made more precise.\n* L297. Typo, missing “is”?\n* References: capitalization is not correct in some instances, e.g. “gan” -> “GAN”, “bayes -> Bayes”, “Neurips” -> “NeurIPS”, etc.\n\n---\n\n[1] Autoregressive Transformers for Indoor Scene Synthesis, Paschalidou et al. NeurIPS 2021 (https://arxiv.org/abs/2110.03675)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an image tokenizer which is approximately invariant to token order. The authors empirically observe that this leads to tokenized representations that are even more decoupled from spatial/local bias common in other tokenizers (even in TiTok, to some extent, despite its 1D representation).\n\nThe authors then introduce “Fixed-Sum Discrete Diffusion” (FSDD), a diffusion-based generative model over multisets/bags of tokens, in order to leverage the tokenizer’s latent space for image generation.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "While the method used to encourage this approximate permutation invariance is very simple (random token shuffling during training), results demonstrating semantically aligned token receptive fields, high linear probing accuracy, and robustness to noise in the input, are interesting and motivate the importance of such a seemingly minor tweak to TiTok.\n\nThe “Fixed-Sum Discrete Diffusion” approach is novel and is demonstrated to outperform certain baselines which do not exploit the permutation invariance.", "weaknesses": "Generation performance appears to be compromised compared to TiTok (FID ~5.1 compared to FID ~2.0), even when using the proposed FSDD. It is not clear whether this is due to limitations of the proposed generative model (FSDD) itself, or due to the learned bag-of-tokens representation having unfavorable structure for generation. Regarding the FSDD, some details and design choices still remain unclear to me and should be described and supported more explicitly (see questions below). Overall, further discussion, analysis and experiments related to generation could strengthen the paper.", "questions": "**Q1.** Section 3.3: I am a bit confused about the choice of $\\sigma_t$ in eq. (5), as it depends on the samples and is not dependent on $t$ like a usual noise schedule in flow matching/diffusion would. Is there a typo here, or if it is intended, could you clarify or provide references? I have a similar question about eq. (9) -- is $|\\mathbf{X}_1 - \\mathbf{X}_0|$ referring to some precomputed statistic, since $\\mathbf{X}_0$ is not available during sampling? It would also be good to elaborate on the t-dependent truncation ratio $f$.\n\n**Q2.** Section 3.3: Have you considered any other strategies for the initial distribution of $\\mathbf{X}_1$? (For example, starting with all counts assigned to a “mask” bin.) Additionally, the constraint that $\\sum \\mathbf{\\tilde X}_t = M$ for every $t$ might seem a bit restrictive. Have you considered any alternatives, such as modeling unconstrained histograms (using any of the existing categorical/multinomial diffusion approaches) and normalizing at the end, or any other modifications to expand the design space of your generative model?\n\n**Q3.** Section 4.3: Do the authors have any insight into why modeling the unordered representation appears to be “harder” than the 1D tokenization used by TiTok? Even with your proposed FSDD, the gFID is significantly lower than in the case of TiTok. Or is the worse generation performance purely due to limitations of FSDD?\n\nSection 4.3.1 / Table 4:\n* **Q4.** What is SetAR? I did not see it explained in the text. Does it refer to autoregressive modeling of the token histogram/“dual transformation”?\n* **Q5.** Are the authors familiar with autoregressive set prediction similar to [1]? It would be interesting to try this or a similar approach performing autoregressive set prediction without the dual transformation.\n\n**Q6.** In Section 4.2.1, rather than a qualitative demonstration on a single example, it would be better to provide quantitative metrics. In particular, to better understand the effectiveness of the train-time shuffling in encouraging permutation invariance, could the authors provide e.g. a histogram across the whole validation set of the error between images reconstructed from two different random token orderings or MSE/MAE with stddev/quantile statistics?\n\n**Q7.** I appreciate the inclusion of MAE in the linear probing results from Table 2, and it would be great if you could also include a comparison between an ablation of your tokenizer trained without shuffling to further support the claim that tokens from permutation invariant representations are semantically richer. \n\n---\n\nMinor questions/comments:\n* Use of “set” nomenclature: Referring to the representation as a “multiset” or “bag” from the beginning would be clearer, rather than leaving this as a footnote. In my own reading of the manuscript, I did not notice the footnote until explicitly searching for the word “multiset”, as I was a bit confused by the inaccurate use of the word “set”. Similarly, I would encourage the authors to consider altering the title to avoid being technically incorrect.\n* 3.1, L148: What does “partial” permutation sampling mean? Is it something other than simply shuffling the tokens? \n* 1, L069: I did not understand what is meant by “dynamic token allocation based on semantic complexity”, maybe this statement could be made more precise.\n* L297. Typo, missing “is”?\n* References: capitalization is not correct in some instances, e.g. “gan” -> “GAN”, “bayes -> Bayes”, “Neurips” -> “NeurIPS”, etc.\n\n---\n\n[1] Autoregressive Transformers for Indoor Scene Synthesis, Paschalidou et al. NeurIPS 2021 (https://arxiv.org/abs/2110.03675)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761159864478}], "openreview_url": "https://openreview.net/forum?id=WqSEVQRlUB", "arxiv_id": "2503.16425", "paper_pdf": "papers/WqSEVQRlUB.pdf", "paper_pdf_sha256": "48b0779bdeaab0c61de87de7c8642ae1df082d14ec689ef9c15abfe518fadc19", "paper_pdf_bytes": 12069518, "paper_pdf_source": "openreview", "code_url": "https://github.com/Gengzigang/TokenSet", "code_repository": "Gengzigang/TokenSet", "code_commit": "6d0fc53e9aa39dcb2f19cbbe2b5646b39b9b43f7", "code_archive": "repos/WqSEVQRlUB.zip", "code_archive_sha256": "e49bda4ebcd29837884bcfefff4a2273a4702a72caa055ace0d5dad5ec509401", "code_archive_bytes": 136893, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 131, "github_languages": {"Python": 82313}, "github_archived": false, "github_pushed_at": "2025-03-21T02:11:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tokenize-image-as-a-set"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0xUEBQV54B", "year": 2025, "status": "rejected", "title": "Large Language Monkeys: Scaling Inference Compute with Repeated Sampling", "authors": ["Bradley Brown", "Jordan Juravsky", "Ryan Saul Ehrlich", "Ronald Clark", "Quoc V Le", "Christopher Re", "Azalia Mirhoseini"], "authorids": ["~Bradley_Brown1", "~Jordan_Juravsky1", "~Ryan_Saul_Ehrlich1", "~Ronald_Clark2", "~Quoc_V_Le1", "~Christopher_Re1", "~Azalia_Mirhoseini3"], "authors_source": "OpenReview API", "abstract": "Scaling the amount of compute used to train language models has dramatically improved their capabilities. However, when it comes to inference, we often limit the amount of compute to only one attempt per problem. Here, we explore inference compute as another axis for scaling, using the simple technique of repeatedly sampling candidate solutions from a model. Across multiple tasks and models, we observe that coverage – the fraction of problems that are solved by any generated sample – scales with the number of samples over four orders of magnitude. Interestingly, the relationship between coverage and the number of samples is often log-linear and can be modelled with an exponentiated power law, suggesting the existence of inference-time scaling laws. In domains like coding and formal proofs, where answers can be automatically verified, these increases in coverage directly translate into improved performance. When we apply repeated sampling to SWE-bench Lite, the fraction of issues solved with DeepSeek-Coder-V2-Instruct increases from 15.9% with one sample to 56% with 250 samples, outperforming the single-sample state-of-the-art of 43%. In domains without automatic verifiers, we find that common methods for picking from a sample collection (majority voting and reward models) plateau beyond several hundred samples and fail to fully scale with the sample budget.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "1WU2aHd5br", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4004/Reviewer_feio"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper studies scaling laws for a new axis of compute for LLMs: inference time compute. They empirically find that coverage (pass@N) improves with the number of inference time samples log-linearly and can be modeled with an exponentiated power law. \nIn domains like coding, where automatic verifiers exist, i.e., verification is much easier than generation, the performance improves from 15.9% with one sample to 56% with 250 samples, outperforming the single-sample state-of-the-art of 43%. In domains lacking automatic verifiers, the authors observe that typical approaches for selecting from a set of IID sampled responses, such as majority voting and reward models, reach a performance plateau after several hundred samples and do not effectively scale with an increasing sample budget, and thus the performance on these problems is bottlenecked by the accuracy of the verifier.", "review_text": "The paper studies scaling laws for a new axis of compute for LLMs: inference time compute. They empirically find that coverage (pass@N) improves with the number of inference time samples log-linearly and can be modeled with an exponentiated power law. \nIn domains like coding, where automatic verifiers exist, i.e., verification is much easier than generation, the performance improves from 15.9% with one sample to 56% with 250 samples, outperforming the single-sample state-of-the-art of 43%. In domains lacking automatic verifiers, the authors observe that typical approaches for selecting from a set of IID sampled responses, such as majority voting and reward models, reach a performance plateau after several hundred samples and do not effectively scale with an increasing sample budget, and thus the performance on these problems is bottlenecked by the accuracy of the verifier.", "strengths": "- The paper presents an extensive analysis of coverage on math and coding tasks and plots empirical scaling laws (and fitted ones) on a multiple model families like Gemma, Pythia and Llama. One of the most interesting trends to note is that models from the same family only affected the offset of the log-linear plot, and not the slope. This is different from typical scaling laws for pre-training loss seen in prior works.\n- The authors also extend analysis to inference time compute, and show that sometimes drawing multiple samples from a smaller (and less capable model) is more compute-efficient (in terms of FLOPs), compared to a single sample from a larger and more capable model. This analysis can be immediately used to reduce the inference time for models deployed in practice, without any loss in performance, at least when a reliable verifier is available. \n- The analysis in Figure 6,7 is particularly compelling to put more effort in improving verification, since it suggests that most test-time methods for reasoning/safety etc., are not failing due to lack of coverage, but more due to the inaccuracy of the verifier.", "weaknesses": "- While the coverage analysis expands on different models and tasks, it does not model the affect of other common parameters of interest like temperature, top-K etc. \n- Some recent works like (Snell et al. 2024, Setlur et al. 2024) show that beam search against an automated verifier improves compute efficiency, with a fixed beam size. Having analysis on this direction of compute use would make the paper stronger.\n- While the authors fit scaling laws for coverage, it would be much more useful to see if scaling laws can also be fit for the setting with trained verifiers, that may not be perfect, accounting for the error in the verifier. Currently it is unclear how the size/data used for the trained verifier affects the compute scaling laws for best-of-n inference. \n- In general, the paper presents interesting results for best-of-N inference, but they are quite narrow and expected. Broadening the laws to other forms of compute usage, or identifying how the \"learnability\" of a verifier affects these laws can help to make the work more complete.\n\n[1] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (C. Snell, J. Lee, K. Xu, A. Kumar)\n[2] Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (A. Setlur, C. Nagpal, A. Fisch, X. Geng, J. Eisenstein, R. Agarwal, A. Agarwal, J. Berant, A. Kumar)", "questions": "- Maybe I missed discussion on this but is there any attempt to also parameterize the coverage scaling law in terms of the model size?\n- Can the authors compare their work with parallel work from Snell et. al. 2024 (Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters)? In particular it would be nice to see discussion on whether the exponentiated power law is also a reasonable to fit scaling curves generated by forms of test-time compute methods, like sequential corrections explored by Snell et al.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies scaling laws for a new axis of compute for LLMs: inference time compute. They empirically find that coverage (pass@N) improves with the number of inference time samples log-linearly and can be modeled with an exponentiated power law. \nIn domains like coding, where automatic verifiers exist, i.e., verification is much easier than generation, the performance improves from 15.9% with one sample to 56% with 250 samples, outperforming the single-sample state-of-the-art of 43%. In domains lacking automatic verifiers, the authors observe that typical approaches for selecting from a set of IID sampled responses, such as majority voting and reward models, reach a performance plateau after several hundred samples and do not effectively scale with an increasing sample budget, and thus the performance on these problems is bottlenecked by the accuracy of the verifier.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper presents an extensive analysis of coverage on math and coding tasks and plots empirical scaling laws (and fitted ones) on a multiple model families like Gemma, Pythia and Llama. One of the most interesting trends to note is that models from the same family only affected the offset of the log-linear plot, and not the slope. This is different from typical scaling laws for pre-training loss seen in prior works.\n- The authors also extend analysis to inference time compute, and show that sometimes drawing multiple samples from a smaller (and less capable model) is more compute-efficient (in terms of FLOPs), compared to a single sample from a larger and more capable model. This analysis can be immediately used to reduce the inference time for models deployed in practice, without any loss in performance, at least when a reliable verifier is available. \n- The analysis in Figure 6,7 is particularly compelling to put more effort in improving verification, since it suggests that most test-time methods for reasoning/safety etc., are not failing due to lack of coverage, but more due to the inaccuracy of the verifier.", "weaknesses": "- While the coverage analysis expands on different models and tasks, it does not model the affect of other common parameters of interest like temperature, top-K etc. \n- Some recent works like (Snell et al. 2024, Setlur et al. 2024) show that beam search against an automated verifier improves compute efficiency, with a fixed beam size. Having analysis on this direction of compute use would make the paper stronger.\n- While the authors fit scaling laws for coverage, it would be much more useful to see if scaling laws can also be fit for the setting with trained verifiers, that may not be perfect, accounting for the error in the verifier. Currently it is unclear how the size/data used for the trained verifier affects the compute scaling laws for best-of-n inference. \n- In general, the paper presents interesting results for best-of-N inference, but they are quite narrow and expected. Broadening the laws to other forms of compute usage, or identifying how the \"learnability\" of a verifier affects these laws can help to make the work more complete.\n\n[1] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (C. Snell, J. Lee, K. Xu, A. Kumar)\n[2] Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (A. Setlur, C. Nagpal, A. Fisch, X. Geng, J. Eisenstein, R. Agarwal, A. Agarwal, J. Berant, A. Kumar)", "questions": "- Maybe I missed discussion on this but is there any attempt to also parameterize the coverage scaling law in terms of the model size?\n- Can the authors compare their work with parallel work from Snell et. al. 2024 (Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters)? In particular it would be nice to see discussion on whether the exponentiated power law is also a reasonable to fit scaling curves generated by forms of test-time compute methods, like sequential corrections explored by Snell et al.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731432150320}, {"id": "ondawrRMSu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4004/Reviewer_zJ2q"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This work considers the question of scaling inference compute through repeated sampling.  The authors find that repeated sampling can greatly improve the performance on reasoning/code tasks, particularly when there is an external verifier that can be used to check the result.  At a high level, this work targets an interesting research question and the methodology looks sound.  However, parts of the paper are poorly written/confusing (particularly in explicating the experimental setup + results with/without the verifier), and could benefit from some major revisions.  I am willing to raise my score if the authors address my questions/concerns.", "review_text": "This work considers the question of scaling inference compute through repeated sampling.  The authors find that repeated sampling can greatly improve the performance on reasoning/code tasks, particularly when there is an external verifier that can be used to check the result.  At a high level, this work targets an interesting research question and the methodology looks sound.  However, parts of the paper are poorly written/confusing (particularly in explicating the experimental setup + results with/without the verifier), and could benefit from some major revisions.  I am willing to raise my score if the authors address my questions/concerns.", "strengths": "* While it isn't surprising that repeated sampling improves performance, the authors have quantified this phenomena in a precise way.  In particular, they find a relatively clean scaling relation between the coverage and number of samples.\n* There is good diversity in the models and datasets used, from more \"agentic\" tasks like SWE-bench to standard math/code reasoning.", "weaknesses": "* The authors should include a few more details about the experimental setup or streamline the existing writing.  For example, one may have chosen a couple of different approaches on how to incorporate the verifier: the first is an iid setup, where the verifier is just used to pick the final answer out of several of attempts.  The second is where the model is provided some verifier signal (e.g., is the answer right/wrong?) between attempts, and asked to review its reasoning trace.  Both scenarios interesting, but there were some experimental choices made in the paper that bear some justification.  \n* The improved performance with repeat sampling is not meaningful in and of itself; if you have enough attempts, the fraction that you will get right will of course go up.  However, it may be useful for the reader to show an example of repeat sampling where the model gets the right answer after a small number of samples -- which part of the reasoning trace gets \"fixed\" between samples typically?  Can we characterize the mistakes that the models make better?\n* I am a bit unclear on the relevant baseline here.  How should we compare spending the inference compute on sampling (versus, say, doing X-of-thought approaches)?\n* A common assumption in the literature is that verification is easier than generation.  Interestingly, this work seems to show that while non-oracle verification can provide some wins, the improvement is overall quite small.  It would be interesting to dig into this question a bit more, and study some explicit examples where the reward model (or majority vote) fails.  How much of the lack of improvement is due to the inherent noisiness of the reward model?  Can the authors characterize the failure mode here a bit more precisely (is non-oracle verification almost as hard as generation in this case)?", "questions": "(See above)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work considers the question of scaling inference compute through repeated sampling.  The authors find that repeated sampling can greatly improve the performance on reasoning/code tasks, particularly when there is an external verifier that can be used to check the result.  At a high level, this work targets an interesting research question and the methodology looks sound.  However, parts of the paper are poorly written/confusing (particularly in explicating the experimental setup + results with/without the verifier), and could benefit from some major revisions.  I am willing to raise my score if the authors address my questions/concerns.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* While it isn't surprising that repeated sampling improves performance, the authors have quantified this phenomena in a precise way.  In particular, they find a relatively clean scaling relation between the coverage and number of samples.\n* There is good diversity in the models and datasets used, from more \"agentic\" tasks like SWE-bench to standard math/code reasoning.", "weaknesses": "* The authors should include a few more details about the experimental setup or streamline the existing writing.  For example, one may have chosen a couple of different approaches on how to incorporate the verifier: the first is an iid setup, where the verifier is just used to pick the final answer out of several of attempts.  The second is where the model is provided some verifier signal (e.g., is the answer right/wrong?) between attempts, and asked to review its reasoning trace.  Both scenarios interesting, but there were some experimental choices made in the paper that bear some justification.  \n* The improved performance with repeat sampling is not meaningful in and of itself; if you have enough attempts, the fraction that you will get right will of course go up.  However, it may be useful for the reader to show an example of repeat sampling where the model gets the right answer after a small number of samples -- which part of the reasoning trace gets \"fixed\" between samples typically?  Can we characterize the mistakes that the models make better?\n* I am a bit unclear on the relevant baseline here.  How should we compare spending the inference compute on sampling (versus, say, doing X-of-thought approaches)?\n* A common assumption in the literature is that verification is easier than generation.  Interestingly, this work seems to show that while non-oracle verification can provide some wins, the improvement is overall quite small.  It would be interesting to dig into this question a bit more, and study some explicit examples where the reward model (or majority vote) fails.  How much of the lack of improvement is due to the inherent noisiness of the reward model?  Can the authors characterize the failure mode here a bit more precisely (is non-oracle verification almost as hard as generation in this case)?", "questions": "(See above)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730765151056}, {"id": "ACOpynxHdh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4004/Reviewer_W6hC"], "rating": 5, "soundness": 2, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper presents a systematic evaluation of repeated sampling for LLM generation across models and benchmark problems considering both oracle verifiers (pass@k) and reward models to select answers. The paper finds smooth scaling behavior with number of samples across all settings with oracle verifiers and proposes a parametric for the inference time scaling laws, even showing that repeated sampling from smaller models can outperform FLOP-matched sampling from larger models. Finally, the paper finds that there remains a large gap between the performance of oracle verifiers and reward models or other mechanisms like majority voting.", "review_text": "This paper presents a systematic evaluation of repeated sampling for LLM generation across models and benchmark problems considering both oracle verifiers (pass@k) and reward models to select answers. The paper finds smooth scaling behavior with number of samples across all settings with oracle verifiers and proposes a parametric for the inference time scaling laws, even showing that repeated sampling from smaller models can outperform FLOP-matched sampling from larger models. Finally, the paper finds that there remains a large gap between the performance of oracle verifiers and reward models or other mechanisms like majority voting.", "strengths": "1. The paper provides a clear, rigorous study of an important and basic topic in LLM inference. The paper tests many different benchmarks and base models and finds that the main results to be robust. A lot of these may exist in prior work, but not in a coherent and systematic way as presented here. \n\n2. The paper is well-written and easy to follow.", "weaknesses": "1. The paper needs to be careful about claims of novelty. The paper does discuss some related work in the intro, but could be more clear that existing work has made many of these observations. The main diffference is that this paper gives a more systematic analysis. For example alphacode [1] and others have figures that look almost identical scaling figures (Fig 6 in alphacode paper), and [2] has similar figures about scaling samples with reward models. \n\n2. I think the scaling laws presentation is missing the somewhat basic discussion that this is totally the expected behavior for computing pass@k for independent bernoulli samples, no LLMs needed. For example, try running this numpy code and you will reproduce the same kinds of scaling law curves on a log scale:\n```\np = 0.001\nT = 10000\nn_trials = 1000\n\nsamples = np.random.random((n_trials, T)) < p\npasses = np.cumsum(samples, axis=1) >= 1\npass_rate = np.mean(passes, axis=0)\n```\nOf course this is a simplification since there should be a different value of p for each problem in the test set, but the basic idea is that when you take independent samples of a bernoulli variable, this is the expected behavior. There is likely a closed form for this simple process. It would probably make sense to model this directly rather than fitting a heuristic scaling law.\n\n3. It is not clear how the presented scaling laws could be useful. The usual usefulness of pre-training scaling laws is to predict the optimal model size at a FLOP budget beyond those used for training. This kind of extrapolation is not tested in the paper. Moreover, the scaling laws are different for each specific model/task combinations and seem cheap to estimate, so it is not clear how prescriptive they are.\n\n4. The stated conclusions about verifiers seem to be too strong given that the paper only uses one reward model that is not particularly tailored to the problems studied. The off-the-shelf RM is chosen for performance on reward bench reasoning, but to my understanding this benchmark is mostly about humaneval-style coding and then the RM is being applied on math reasoning problems. For example, [2] trains task-specific RMs and does not observe the same sort of saturation.\n\n5. Some discussion of related work is missing when discussing how to improve verifiers in the future work directions. For example, [3] proposes an objective for reward modeling to allow LMs to evaluate themselves with a linear model on their representations and [4] uses models to evaluate themselves.\n\n[1] Li, Yujia, et al. \"Competition-level code generation with alphacode.\" Science 378.6624 (2022): 1092-1097.\n\n[2] Lightman, Hunter, et al. \"Let's verify step by step.\" arXiv preprint arXiv:2305.20050 (2023).\n\n[3] Li, Kenneth, et al. \"Q-Probe: A Lightweight Approach to Reward Maximization for Language Models.\" arXiv preprint arXiv:2402.14688 (2024).\n\n[4] Yuan, Weizhe, et al. \"Self-rewarding language models.\" arXiv preprint arXiv:2401.10020 (2024).", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a systematic evaluation of repeated sampling for LLM generation across models and benchmark problems considering both oracle verifiers (pass@k) and reward models to select answers. The paper finds smooth scaling behavior with number of samples across all settings with oracle verifiers and proposes a parametric for the inference time scaling laws, even showing that repeated sampling from smaller models can outperform FLOP-matched sampling from larger models. Finally, the paper finds that there remains a large gap between the performance of oracle verifiers and reward models or other mechanisms like majority voting.", "soundness": 2, "presentation": 4, "contribution": 2, "strengths": "1. The paper provides a clear, rigorous study of an important and basic topic in LLM inference. The paper tests many different benchmarks and base models and finds that the main results to be robust. A lot of these may exist in prior work, but not in a coherent and systematic way as presented here. \n\n2. The paper is well-written and easy to follow.", "weaknesses": "1. The paper needs to be careful about claims of novelty. The paper does discuss some related work in the intro, but could be more clear that existing work has made many of these observations. The main diffference is that this paper gives a more systematic analysis. For example alphacode [1] and others have figures that look almost identical scaling figures (Fig 6 in alphacode paper), and [2] has similar figures about scaling samples with reward models. \n\n2. I think the scaling laws presentation is missing the somewhat basic discussion that this is totally the expected behavior for computing pass@k for independent bernoulli samples, no LLMs needed. For example, try running this numpy code and you will reproduce the same kinds of scaling law curves on a log scale:\n```\np = 0.001\nT = 10000\nn_trials = 1000\n\nsamples = np.random.random((n_trials, T)) < p\npasses = np.cumsum(samples, axis=1) >= 1\npass_rate = np.mean(passes, axis=0)\n```\nOf course this is a simplification since there should be a different value of p for each problem in the test set, but the basic idea is that when you take independent samples of a bernoulli variable, this is the expected behavior. There is likely a closed form for this simple process. It would probably make sense to model this directly rather than fitting a heuristic scaling law.\n\n3. It is not clear how the presented scaling laws could be useful. The usual usefulness of pre-training scaling laws is to predict the optimal model size at a FLOP budget beyond those used for training. This kind of extrapolation is not tested in the paper. Moreover, the scaling laws are different for each specific model/task combinations and seem cheap to estimate, so it is not clear how prescriptive they are.\n\n4. The stated conclusions about verifiers seem to be too strong given that the paper only uses one reward model that is not particularly tailored to the problems studied. The off-the-shelf RM is chosen for performance on reward bench reasoning, but to my understanding this benchmark is mostly about humaneval-style coding and then the RM is being applied on math reasoning problems. For example, [2] trains task-specific RMs and does not observe the same sort of saturation.\n\n5. Some discussion of related work is missing when discussing how to improve verifiers in the future work directions. For example, [3] proposes an objective for reward modeling to allow LMs to evaluate themselves with a linear model on their representations and [4] uses models to evaluate themselves.\n\n[1] Li, Yujia, et al. \"Competition-level code generation with alphacode.\" Science 378.6624 (2022): 1092-1097.\n\n[2] Lightman, Hunter, et al. \"Let's verify step by step.\" arXiv preprint arXiv:2305.20050 (2023).\n\n[3] Li, Kenneth, et al. \"Q-Probe: A Lightweight Approach to Reward Maximization for Language Models.\" arXiv preprint arXiv:2402.14688 (2024).\n\n[4] Yuan, Weizhe, et al. \"Self-rewarding language models.\" arXiv preprint arXiv:2401.10020 (2024).", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730554569875}, {"id": "4S2pEpXraF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4004/Reviewer_NSZC"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper studies the potential of scaling inference compute of LLMs. The authors show that, by generating repeated samples from language models, we can increase the “coverage”, i.e., the fraction of problems that are solved by any generated sample. The improved coverage directly translates to better performances when an oracle verifier is available.  The author also conduct experiments in the setting without oracle verifiers and find that the performances plateau quickly given repeated samples.", "review_text": "This paper studies the potential of scaling inference compute of LLMs. The authors show that, by generating repeated samples from language models, we can increase the “coverage”, i.e., the fraction of problems that are solved by any generated sample. The improved coverage directly translates to better performances when an oracle verifier is available.  The author also conduct experiments in the setting without oracle verifiers and find that the performances plateau quickly given repeated samples.", "strengths": "- The research problem is interesting and well-motivated. Scaling up training compute has led to remarkable success in deep learning. It is important to consider scaling inference compute.\n- The paper is easy to read.\n- The evaluation covers 5 different benchmark datasets and show consistent trends.", "weaknesses": "- One main claim, **scaling inference compute through repeated sampling leads to large improvements in coverage, seems trivial**. I believe that this fact is generally known in the community. Empirically, it has been observed in prior works like [1,2]. Mathematically, it is a simple consequence of equation 1. It is easy to prove that pass@$k$ monotonically increases with $k$ as long as there exists some $C_i>0$. The scaling curves are simply numerical calculation of equation 1 (correct me if this is not the case!). Thus, the novelty of this analysis seems limited.\n\n- The paper **lacks in-depth analysis on scaling laws**. In Section 3, the proposed formula (equation 2) is simply adopted from the GPT-4 technical report. Then the “curve fitting” directly fit the power law to the curve generated by a _known formula_ (i.e., equation 1). I believe that meaningful scaling laws should be distilled from experimental observations. It’s not clear to me what is the insight of fitting some curves when the underlying formula is already known.\n\n- I like the analysis on domains without automatic verifiers since it is a more realistic setting than the experiments based on oracle verifiers.  However, this section is very short and **lacks a deeper exploration of verifiers**. The conclusions here are based on experiments with a single existing 8B verifier. Strengthening this analysis with verifiers of varying sizes and other well-known verification approaches (e.g., process supervision [3]) would significantly enrich this section.\n\n- Minor issues: Line 394: $k$ should be italic.\n\n[1] Program Synthesis with Large Language Models\n\n[2] Competition-level code generation with AlphaCode\n\n[3] Let’s Verify Step by Step", "questions": "Please refer to the **weakness** part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the potential of scaling inference compute of LLMs. The authors show that, by generating repeated samples from language models, we can increase the “coverage”, i.e., the fraction of problems that are solved by any generated sample. The improved coverage directly translates to better performances when an oracle verifier is available.  The author also conduct experiments in the setting without oracle verifiers and find that the performances plateau quickly given repeated samples.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- The research problem is interesting and well-motivated. Scaling up training compute has led to remarkable success in deep learning. It is important to consider scaling inference compute.\n- The paper is easy to read.\n- The evaluation covers 5 different benchmark datasets and show consistent trends.", "weaknesses": "- One main claim, **scaling inference compute through repeated sampling leads to large improvements in coverage, seems trivial**. I believe that this fact is generally known in the community. Empirically, it has been observed in prior works like [1,2]. Mathematically, it is a simple consequence of equation 1. It is easy to prove that pass@$k$ monotonically increases with $k$ as long as there exists some $C_i>0$. The scaling curves are simply numerical calculation of equation 1 (correct me if this is not the case!). Thus, the novelty of this analysis seems limited.\n\n- The paper **lacks in-depth analysis on scaling laws**. In Section 3, the proposed formula (equation 2) is simply adopted from the GPT-4 technical report. Then the “curve fitting” directly fit the power law to the curve generated by a _known formula_ (i.e., equation 1). I believe that meaningful scaling laws should be distilled from experimental observations. It’s not clear to me what is the insight of fitting some curves when the underlying formula is already known.\n\n- I like the analysis on domains without automatic verifiers since it is a more realistic setting than the experiments based on oracle verifiers.  However, this section is very short and **lacks a deeper exploration of verifiers**. The conclusions here are based on experiments with a single existing 8B verifier. Strengthening this analysis with verifiers of varying sizes and other well-known verification approaches (e.g., process supervision [3]) would significantly enrich this section.\n\n- Minor issues: Line 394: $k$ should be italic.\n\n[1] Program Synthesis with Large Language Models\n\n[2] Competition-level code generation with AlphaCode\n\n[3] Let’s Verify Step by Step", "questions": "Please refer to the **weakness** part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730522328432}], "openreview_url": "https://openreview.net/forum?id=0xUEBQV54B", "arxiv_id": "2407.21787", "paper_pdf": "papers/0xUEBQV54B.pdf", "paper_pdf_sha256": "9f3351a18bb10567ab64500bc630641732f0320233e171ae16579e80a5a1d577", "paper_pdf_bytes": 4521578, "paper_pdf_source": "openreview", "code_url": "https://github.com/ScalingIntelligence/large_language_monkeys", "code_repository": "ScalingIntelligence/large_language_monkeys", "code_commit": "ed2bdb06bcfcbda1ed7d9c2bff87c2db10c3ee78", "code_archive": "repos/0xUEBQV54B.zip", "code_archive_sha256": "40a582ea6b1fc2586a8b2742eeb07de2281d6c696281966226439aab494d88b9", "code_archive_bytes": 36856, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 30, "github_languages": {"Python": 72916, "Dockerfile": 768}, "github_archived": false, "github_pushed_at": "2024-09-25T21:28:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/large-language-monkeys-scaling-inference"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XdSYtriYfI", "year": 2024, "status": "rejected", "title": "Federated Ensemble-Directed Offline Reinforcement Learning", "authors": ["Desik Rengarajan", "Nitin Ragothaman", "Dileep Kalathil", "Srinivas Shakkottai"], "authorids": ["~Desik_Rengarajan1", "~Nitin_Ragothaman1", "~Dileep_Kalathil1", "~Srinivas_Shakkottai1"], "authors_source": "OpenReview API", "abstract": "We consider the problem of federated offline reinforcement learning (RL), a scenario under which distributed learning agents must collaboratively learn a high-quality control policy only using small pre-collected datasets generated according to different unknown behavior policies. Naively combining a standard offline RL approach with a standard federated learning approach to solve this problem can lead to poorly performing policies.  In response, we develop the Federated Ensemble-Directed Offline Reinforcement Learning  Algorithm (FEDORA), which distills the collective wisdom of the clients using an ensemble learning approach.  We develop the FEDORA codebase to utilize distributed compute resources on a federated learning platform. We show that FEDORA significantly outperforms other approaches, including offline RL over the combined data pool, in various complex continuous control environments and real-world datasets. Finally, we demonstrate the performance of FEDORA in the real-world on a mobile robot.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "SoSlNgCmtY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5861/Reviewer_MJbF"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This study examines the issue of training offline RL in a federated environment. Initially, the authors outline the difficulties associated with federated offline RL. Following this, a method named FEDORA is introduced, which aims to enhance the collective performance of federated learning. FEDORA accomplishes this by managing how models are aggregated and by reducing the influence of local training data on the global model.", "review_text": "This study examines the issue of training offline RL in a federated environment. Initially, the authors outline the difficulties associated with federated offline RL. Following this, a method named FEDORA is introduced, which aims to enhance the collective performance of federated learning. FEDORA accomplishes this by managing how models are aggregated and by reducing the influence of local training data on the global model.", "strengths": "- Clarity and comprehensibility: The manuscript is articulated clearly. The content is easy to understand.\n\n- Relevance and novelty: This paper focuses on a new problem for RL training (i.e. federated offline RL).\n\n- Insightful discussion: The presentation of challenges related to federated offline RL is clear and insightful.", "weaknesses": "- The paper claims that it outperforms the centralized training offline RL, but the proposed method uses an additional data selection technique compared to the baselines, which is not quite fair. To make this claim, the baseline needs to also have data rebalancing or data selection technique (e.g. with [1], or just send the average reward of all episodes to the server).\n\n- The title is \"ensemble\" method, but it seems the solution is not really using an ensemble method. It's a modified federated aggregation method.\n\n- Even though the proposed method is about federated RL, related work and evaluation should be considered for supervised RL with heterogeneous data as well.\n\n[1] Yue, Yang, et al. \"Boosting offline reinforcement learning via data rebalancing.\" arXiv preprint arXiv:2210.09241 (2022).", "questions": "- The proposed technique seems to be doing a data selection based on the quality of the data, would you get a similar performance if you just select the clients with higher reward datasets? Like in [1] for example.\n \n- Would the proposed method increase the amount of communication required for federated training compared to simple FAvg? It would be nice to quantify that.\n\n\n\n[1] Yue, Yang, et al. \"Boosting offline reinforcement learning via data rebalancing.\" arXiv preprint arXiv:2210.09241 (2022).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study examines the issue of training offline RL in a federated environment. Initially, the authors outline the difficulties associated with federated offline RL. Following this, a method named FEDORA is introduced, which aims to enhance the collective performance of federated learning. FEDORA accomplishes this by managing how models are aggregated and by reducing the influence of local training data on the global model.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Clarity and comprehensibility: The manuscript is articulated clearly. The content is easy to understand.\n\n- Relevance and novelty: This paper focuses on a new problem for RL training (i.e. federated offline RL).\n\n- Insightful discussion: The presentation of challenges related to federated offline RL is clear and insightful.", "weaknesses": "- The paper claims that it outperforms the centralized training offline RL, but the proposed method uses an additional data selection technique compared to the baselines, which is not quite fair. To make this claim, the baseline needs to also have data rebalancing or data selection technique (e.g. with [1], or just send the average reward of all episodes to the server).\n\n- The title is \"ensemble\" method, but it seems the solution is not really using an ensemble method. It's a modified federated aggregation method.\n\n- Even though the proposed method is about federated RL, related work and evaluation should be considered for supervised RL with heterogeneous data as well.\n\n[1] Yue, Yang, et al. \"Boosting offline reinforcement learning via data rebalancing.\" arXiv preprint arXiv:2210.09241 (2022).", "questions": "- The proposed technique seems to be doing a data selection based on the quality of the data, would you get a similar performance if you just select the clients with higher reward datasets? Like in [1] for example.\n \n- Would the proposed method increase the amount of communication required for federated training compared to simple FAvg? It would be nice to quantify that.\n\n\n\n[1] Yue, Yang, et al. \"Boosting offline reinforcement learning via data rebalancing.\" arXiv preprint arXiv:2210.09241 (2022).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698780344061}, {"id": "gCTJj75ClX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5861/Reviewer_RUBc"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a novel federated ensemble-directed offline reinforcement learning algorithm (FEDORA) that enables multiple agents with heterogeneous offline datasets to collaboratively learn a high-quality control policy without sharing data. The key contributions are:\n1. Identifying key challenges of federated offline RL including ensemble heterogeneity, pessimistic value computation, and data heterogeneity.\n2. Systematically addressing these challenges through an ensemble-directed federation approach that extracts collective wisdom of policies and critics and discourages over-reliance on possibly irrelevant local data.\n3. Demonstrating strong empirical performance of FEDORA over baselines on MuJoCo environments and a real-world robot navigation task.", "review_text": "The paper proposes a novel federated ensemble-directed offline reinforcement learning algorithm (FEDORA) that enables multiple agents with heterogeneous offline datasets to collaboratively learn a high-quality control policy without sharing data. The key contributions are:\n1. Identifying key challenges of federated offline RL including ensemble heterogeneity, pessimistic value computation, and data heterogeneity.\n2. Systematically addressing these challenges through an ensemble-directed federation approach that extracts collective wisdom of policies and critics and discourages over-reliance on possibly irrelevant local data.\n3. Demonstrating strong empirical performance of FEDORA over baselines on MuJoCo environments and a real-world robot navigation task.", "strengths": "1. The paper clearly motivates the problem of federated offline RL and identifies key challenges that are not addressed by naively combining existing methods.\n2. FEDORA is systematically designed to address the identified challenges in an intuitive manner through techniques like ensemble-directed federation and federated optimism.\n3. Extensive experiments demonstrate the effectiveness of FEDORA over baselines on simulated and real-world tasks. The ablation studies provide useful insights.\n4. The paper is clearly written and provides sufficient details to understand the proposed techniques.", "weaknesses": "1. How does FEDORA perform when some clients have near-random or adversarial datasets?\n2. Have the authors experimented with different values of β and δ? Is there a principle behind setting them?\n3. How does FEDORA compare to state-of-the-art offline RL algorithms like CQL or IQL in the federated setting?\n\nThe paper makes solid contributions in identifying and addressing key challenges in federated offline RL. The algorithm design is methodical and supported through extensive experiments. More theoretical and implementation details will further improve the paper.", "questions": "see above。", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel federated ensemble-directed offline reinforcement learning algorithm (FEDORA) that enables multiple agents with heterogeneous offline datasets to collaboratively learn a high-quality control policy without sharing data. The key contributions are:\n1. Identifying key challenges of federated offline RL including ensemble heterogeneity, pessimistic value computation, and data heterogeneity.\n2. Systematically addressing these challenges through an ensemble-directed federation approach that extracts collective wisdom of policies and critics and discourages over-reliance on possibly irrelevant local data.\n3. Demonstrating strong empirical performance of FEDORA over baselines on MuJoCo environments and a real-world robot navigation task.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper clearly motivates the problem of federated offline RL and identifies key challenges that are not addressed by naively combining existing methods.\n2. FEDORA is systematically designed to address the identified challenges in an intuitive manner through techniques like ensemble-directed federation and federated optimism.\n3. Extensive experiments demonstrate the effectiveness of FEDORA over baselines on simulated and real-world tasks. The ablation studies provide useful insights.\n4. The paper is clearly written and provides sufficient details to understand the proposed techniques.", "weaknesses": "1. How does FEDORA perform when some clients have near-random or adversarial datasets?\n2. Have the authors experimented with different values of β and δ? Is there a principle behind setting them?\n3. How does FEDORA compare to state-of-the-art offline RL algorithms like CQL or IQL in the federated setting?\n\nThe paper makes solid contributions in identifying and addressing key challenges in federated offline RL. The algorithm design is methodical and supported through extensive experiments. More theoretical and implementation details will further improve the paper.", "questions": "see above。", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698769804971}, {"id": "8oTV9Xw8Sz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5861/Reviewer_oNBT"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper addresses the problem of federated offline reinforcement learning, aiming to learn a high-performing policy by leveraging multiple distributed offline datasets without the need for centralizing the data. The authors highlight several challenges inherent in this setup and propose a framework called FEDORA (Federated Ensemble-Directed Offline Reinforcement Learning Algorithm) to tackle these challenges in the context of federated offline reinforcement learning. To evaluate the effectiveness of FEDORA, the authors conduct experiments on mujoco environments, demonstrating its capability to achieve desirable results. Furthermore, the proposed framework is deployed on real turtlebots, providing practical validation of its performance and applicability in real-world scenarios.", "review_text": "This paper addresses the problem of federated offline reinforcement learning, aiming to learn a high-performing policy by leveraging multiple distributed offline datasets without the need for centralizing the data. The authors highlight several challenges inherent in this setup and propose a framework called FEDORA (Federated Ensemble-Directed Offline Reinforcement Learning Algorithm) to tackle these challenges in the context of federated offline reinforcement learning. To evaluate the effectiveness of FEDORA, the authors conduct experiments on mujoco environments, demonstrating its capability to achieve desirable results. Furthermore, the proposed framework is deployed on real turtlebots, providing practical validation of its performance and applicability in real-world scenarios.", "strengths": "This paper addresses the intriguing problem of federated offline reinforcement learning, which has received limited attention in the existing literature. The authors shed light on various technical challenges that emerge from this unique setting. To tackle these challenges, they propose an innovative algorithm called FEDORA. The efficacy of FEDORA is demonstrated through its successful application to continuous control tasks in mujoco environments, showcasing its ability to learn effective policies. Moreover, the authors validate the practicality and real-world applicability of the proposed framework by deploying it on real turtlebots, further highlighting its performance under realistic scenarios.", "weaknesses": "While the paper introduces a novel approach to an interesting problem, there are several areas that could benefit from further improvement:\n1. The assumption of identical MDPs and reward functions across all agents limits the generalizability and applicability of the proposed approach to real-world scenarios. Exploring techniques to handle heterogeneity among agents' environments could enhance the practicality of the framework.\n2. The requirement for the server to have access to the complete MDP raises concerns about privacy and may not be practical in many scenarios. Providing a practical example or discussing potential alternatives for this setup would strengthen the paper's applicability.\n3. The paper assumes **knowledge of dataset quality**, but it does not thoroughly address the issue of comparing and assessing the quality of different datasets. Specifically, in Sec. 4 , is the dataset generated by hopper-expert-v2 guaranteed to be better than that generated by hopper-medium-v2? Suppose that $D_m$ is generated by the medium policy which has converged while $D_e$ is generated by the expert policy during the initial random exploration stage. Is the quality of $D_e$ always better than $D_m$ ? It is important to explore methods to evaluate dataset quality more robustly and consider scenarios where the assumption of one dataset being consistently better than another may not hold.\n4. Experimental details for Figure 1 are missing, making it difficult to fully understand and interpret the results. Including the specific experimental setup, including hyperparameters, training procedures, and any other relevant details, would enhance the reproducibility and credibility of the findings.\n5. Eq. (7) implies that the server selects the action based on a weighted average of all agents' actions. Further clarification is needed to explain how the weights assigned to individual agents' actions influence the final decision. Does the weighting scheme prevent one agent from significantly overpowering the others? If not, what mechanisms or strategies are in place to address potential dominance issues and ensure a balanced contribution from all participating agents?\n6. The process of training the centralized policy to achieve the straight-line performance in Figure 2 requires further explanation.\n7. Can you compare the performance of FEDORA with just the expert policy running the offline RL algorithm in Fig. 2?\n8. Some important FedRL papers are missing from reference. Including them would strengthen the scholarly contribution of the paper.", "questions": "1. Could you provide a compelling real-world example that demonstrates the practicality of the problem setup? By illustrating a specific scenario or application where the proposed approach can be effectively applied, readers can gain a better understanding of its relevance and potential impact.\n2. How do you determine the quality of a given dataset?\n3. How is the value of $\\beta$ determined, and what insights can you provide regarding its impact on the weights assigned to individual agents?\n4. Could you provide a brief discussion or analysis of the computational and communication costs associated with the proposed approach?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of federated offline reinforcement learning, aiming to learn a high-performing policy by leveraging multiple distributed offline datasets without the need for centralizing the data. The authors highlight several challenges inherent in this setup and propose a framework called FEDORA (Federated Ensemble-Directed Offline Reinforcement Learning Algorithm) to tackle these challenges in the context of federated offline reinforcement learning. To evaluate the effectiveness of FEDORA, the authors conduct experiments on mujoco environments, demonstrating its capability to achieve desirable results. Furthermore, the proposed framework is deployed on real turtlebots, providing practical validation of its performance and applicability in real-world scenarios.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This paper addresses the intriguing problem of federated offline reinforcement learning, which has received limited attention in the existing literature. The authors shed light on various technical challenges that emerge from this unique setting. To tackle these challenges, they propose an innovative algorithm called FEDORA. The efficacy of FEDORA is demonstrated through its successful application to continuous control tasks in mujoco environments, showcasing its ability to learn effective policies. Moreover, the authors validate the practicality and real-world applicability of the proposed framework by deploying it on real turtlebots, further highlighting its performance under realistic scenarios.", "weaknesses": "While the paper introduces a novel approach to an interesting problem, there are several areas that could benefit from further improvement:\n1. The assumption of identical MDPs and reward functions across all agents limits the generalizability and applicability of the proposed approach to real-world scenarios. Exploring techniques to handle heterogeneity among agents' environments could enhance the practicality of the framework.\n2. The requirement for the server to have access to the complete MDP raises concerns about privacy and may not be practical in many scenarios. Providing a practical example or discussing potential alternatives for this setup would strengthen the paper's applicability.\n3. The paper assumes **knowledge of dataset quality**, but it does not thoroughly address the issue of comparing and assessing the quality of different datasets. Specifically, in Sec. 4 , is the dataset generated by hopper-expert-v2 guaranteed to be better than that generated by hopper-medium-v2? Suppose that $D_m$ is generated by the medium policy which has converged while $D_e$ is generated by the expert policy during the initial random exploration stage. Is the quality of $D_e$ always better than $D_m$ ? It is important to explore methods to evaluate dataset quality more robustly and consider scenarios where the assumption of one dataset being consistently better than another may not hold.\n4. Experimental details for Figure 1 are missing, making it difficult to fully understand and interpret the results. Including the specific experimental setup, including hyperparameters, training procedures, and any other relevant details, would enhance the reproducibility and credibility of the findings.\n5. Eq. (7) implies that the server selects the action based on a weighted average of all agents' actions. Further clarification is needed to explain how the weights assigned to individual agents' actions influence the final decision. Does the weighting scheme prevent one agent from significantly overpowering the others? If not, what mechanisms or strategies are in place to address potential dominance issues and ensure a balanced contribution from all participating agents?\n6. The process of training the centralized policy to achieve the straight-line performance in Figure 2 requires further explanation.\n7. Can you compare the performance of FEDORA with just the expert policy running the offline RL algorithm in Fig. 2?\n8. Some important FedRL papers are missing from reference. Including them would strengthen the scholarly contribution of the paper.", "questions": "1. Could you provide a compelling real-world example that demonstrates the practicality of the problem setup? By illustrating a specific scenario or application where the proposed approach can be effectively applied, readers can gain a better understanding of its relevance and potential impact.\n2. How do you determine the quality of a given dataset?\n3. How is the value of $\\beta$ determined, and what insights can you provide regarding its impact on the weights assigned to individual agents?\n4. Could you provide a brief discussion or analysis of the computational and communication costs associated with the proposed approach?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698711534962}, {"id": "7mL5Vr4193", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5861/Reviewer_mieC"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper considers federated reinforcement learning problems with offline datasets of different qualities.\nInstead of aggregating models in a uniform way, FEDORA adopts the principle of maximum entropy to aggregate local models accordingly. Specifically, the model from offline datasets with higher quality weights more in the central aggregation.\nIt is empirically shown that FEDORA outperforms FedAvg in both simulated environments and real-world experiments.", "review_text": "The paper considers federated reinforcement learning problems with offline datasets of different qualities.\nInstead of aggregating models in a uniform way, FEDORA adopts the principle of maximum entropy to aggregate local models accordingly. Specifically, the model from offline datasets with higher quality weights more in the central aggregation.\nIt is empirically shown that FEDORA outperforms FedAvg in both simulated environments and real-world experiments.", "strengths": "1. The design of FEDORA is well-written and easy to follow.\n2. The paper considers a novel offline federated reinforcement learning where local datasets are generated by different behavioral policies.\n3. FEDORA has been evaluated in various environments and has achieved outstanding performance w.r.t. traditional FedAvg.", "weaknesses": "1. What role does the entropy regularizer in Eq. (5) play? Collapsing to datasets with the best performance does not seem to worsen the convergent performance of FEDORA. Is there any ablation study on this term?\n2. The approach toward computing $Q_{t_i}$ and $\\pi_{t_i}$ in Section 5.1 is not clearly described. The evaluation of Q functions and policy may be inaccurate on datasets of low quality. How will it influence the performance of FEDORA?", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers federated reinforcement learning problems with offline datasets of different qualities.\nInstead of aggregating models in a uniform way, FEDORA adopts the principle of maximum entropy to aggregate local models accordingly. Specifically, the model from offline datasets with higher quality weights more in the central aggregation.\nIt is empirically shown that FEDORA outperforms FedAvg in both simulated environments and real-world experiments.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The design of FEDORA is well-written and easy to follow.\n2. The paper considers a novel offline federated reinforcement learning where local datasets are generated by different behavioral policies.\n3. FEDORA has been evaluated in various environments and has achieved outstanding performance w.r.t. traditional FedAvg.", "weaknesses": "1. What role does the entropy regularizer in Eq. (5) play? Collapsing to datasets with the best performance does not seem to worsen the convergent performance of FEDORA. Is there any ablation study on this term?\n2. The approach toward computing $Q_{t_i}$ and $\\pi_{t_i}$ in Section 5.1 is not clearly described. The evaluation of Q functions and policy may be inaccurate on datasets of low quality. How will it influence the performance of FEDORA?", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698674636272}], "openreview_url": "https://openreview.net/forum?id=XdSYtriYfI", "arxiv_id": "2305.03097", "paper_pdf": "papers/XdSYtriYfI.pdf", "paper_pdf_sha256": "49176f17308865e93afc0d409c85f7f329bbf9ced58d91b309309618f56eaec1", "paper_pdf_bytes": 2173337, "paper_pdf_source": "openreview", "code_url": "https://github.com/DesikRengarajan/FEDORA", "code_repository": "DesikRengarajan/FEDORA", "code_commit": "3c355a867e0d38ab650fce18f78a742691afff0f", "code_archive": "repos/XdSYtriYfI.zip", "code_archive_sha256": "e4850d616fd553a2c0775040b3b94d6282022b6f4dd5208499ef82fef8fb77d4", "code_archive_bytes": 19874, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 29, "github_languages": {"Python": 44319, "Shell": 698}, "github_archived": false, "github_pushed_at": "2024-09-25T22:26:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/federated-ensemble-directed-offline"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ygYXtbb3og3", "year": 2023, "status": "rejected", "title": "Mutual Information Regularized Offline Reinforcement Learning", "authors": ["Xiao Ma", "Bingyi Kang", "Zhongwen Xu", "Min Lin", "Shuicheng YAN"], "authorids": ["~Xiao_Ma2", "~Bingyi_Kang1", "~Zhongwen_Xu1", "~Min_Lin1", "~Shuicheng_YAN3"], "authors_source": "OpenReview API", "abstract": "Offline reinforcement learning (RL) aims at learning an effective policy from offline datasets without active interactions with the environment. The major challenge of offline RL is the distribution shift that appears when out-of-distribution actions are queried, which makes the policy improvement direction biased by extrapolation errors. Most existing methods address this problem by penalizing the policy for deviating from the behavior policy during policy improvement or making conservative updates for value functions during policy evaluation. In this work, we propose a novel MISA framework to approach offline RL from the perspective of Mutual Information between States and Actions in the dataset by directly constraining the policy improvement direction. Intuitively, mutual information measures the mutual dependence of actions and states, which reflects how a behavior agent reacts to certain environment states during data collection. To effectively utilize this information to facilitate policy learning, MISA constructs lower bounds of mutual information parameterized by the policy and Q-values. We show that optimizing this lower bound is equivalent to maximizing the likelihood of a one-step improved policy on the offline dataset. In this way, we constrain the policy improvement direction to lie in the data manifold. The resulting algorithm simultaneously augments the policy evaluation and improvement by adding a mutual information regularization. MISA is a general offline RL framework that unifies conservative Q-learning (CQL) and behavior regularization methods (e.g., TD3+BC) as special cases. Our experiments show that MISA performs significantly better than existing methods and achieves new state-of-the-art on various tasks of the D4RL benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "-eH_0KaYuq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1028/Reviewer_ZgDk"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors present MISA, a framework for offline RL based on mutual information based regularization over states and actions. The authors motivate their approach step by step and demonstrate how TD3+Behavior cloning and CQL are related to their proposed approach. Finally, the authors provide empirical results relative to key baselines on the D4RL benchmark while also providing ablation studies and analysis. ", "review_text": "My biggest concerns about this paper are the relatively low amount of novelty and the fact that the overall motivation is a bit intuitive in nature. However, I lean towards acceptance because I do really appreciate the theoretical discourse and experiments. I feel that in light of this context the authors have provided, the paper is fleshed out enough to make a contribution to the conference, particularly because of its focus on such an important area. \n\nUpdate After Author Response: \n\nI have read through the other reviews and responses to each review. I have kept my score unchanged, but I can't really serve as a strong advocate for this paper either. It is quite outside my area of expertise, so I feel that my confidence is quite low about what the right thing to do here is as I sympathize with all of the points made.\n\nI agree that the novelty is not huge here, but also hear the authors that it is a significant contribution beyond the past literature. However, the results not being too significant could indeed be a potential issue for adoption of this method. Moreover, I think the authors are too quick to discount the importance of not performing well on data from a uniform policy. It is certainly not common for offline RL benchmarks, but if the data compiled by our behavior policy is already \"expert\" level then offline RL is only improving things at the margins. In typical off-policy RL we start with a uniform policy and while I acknowledge that these are certainly different settings, I think offline RL should be considered harder than off-policy RL (due to lack of access to the underlying behavior policy and lack of access to a simulator during training) and not a strictly easier setting (akin to some slightly more ambitious form of imitation learning).\n", "strengths": "Strengths: \n- The authors present a clear buildup to their approach in math and justify each step.\n- The authors contextualize their approach pretty well with respect to prior art on a theoretical level.\n- The approach gets pretty good performance relative to baselines on the D4RL benchmark. \n\nWeaknesses: \n- The novelty is not terribly high in comparison to past work on offline RL and mutual information estimation. \n- The introduction of the mutual information regularization in equations 5 and 6 struck me as quite intuitive in nature and not really originating from first principles.\n\nBoth: \n- I do really appreciate that the authors clearly highlight limitations of their approach when doing learning from data that is using far from an expert policy. However, on the other hand, it is quite a significant limitation relative to baselines. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors present MISA, a framework for offline RL based on mutual information based regularization over states and actions. The authors motivate their approach step by step and demonstrate how TD3+Behavior cloning and CQL are related to their proposed approach. Finally, the authors provide empirical results relative to key baselines on the D4RL benchmark while also providing ablation studies and analysis. ", "strength_and_weaknesses": "Strengths: \n- The authors present a clear buildup to their approach in math and justify each step.\n- The authors contextualize their approach pretty well with respect to prior art on a theoretical level.\n- The approach gets pretty good performance relative to baselines on the D4RL benchmark. \n\nWeaknesses: \n- The novelty is not terribly high in comparison to past work on offline RL and mutual information estimation. \n- The introduction of the mutual information regularization in equations 5 and 6 struck me as quite intuitive in nature and not really originating from first principles.\n\nBoth: \n- I do really appreciate that the authors clearly highlight limitations of their approach when doing learning from data that is using far from an expert policy. However, on the other hand, it is quite a significant limitation relative to baselines. ", "clarity,_quality,_novelty_and_reproducibility": "The writing quality is pretty good overall and I really appreciate the theoretical discourse of this paper. The novelty is not very high, but I do appreciate that the authors also clearly present their contribution relative to prior art. The authors have provided their code in the supplemental material. ", "summary_of_the_review": "My biggest concerns about this paper are the relatively low amount of novelty and the fact that the overall motivation is a bit intuitive in nature. However, I lean towards acceptance because I do really appreciate the theoretical discourse and experiments. I feel that in light of this context the authors have provided, the paper is fleshed out enough to make a contribution to the conference, particularly because of its focus on such an important area. \n\nUpdate After Author Response: \n\nI have read through the other reviews and responses to each review. I have kept my score unchanged, but I can't really serve as a strong advocate for this paper either. It is quite outside my area of expertise, so I feel that my confidence is quite low about what the right thing to do here is as I sympathize with all of the points made.\n\nI agree that the novelty is not huge here, but also hear the authors that it is a significant contribution beyond the past literature. However, the results not being too significant could indeed be a potential issue for adoption of this method. Moreover, I think the authors are too quick to discount the importance of not performing well on data from a uniform policy. It is certainly not common for offline RL benchmarks, but if the data compiled by our behavior policy is already \"expert\" level then offline RL is only improving things at the margins. In typical off-policy RL we start with a uniform policy and while I acknowledge that these are certainly different settings, I think offline RL should be considered harder than off-policy RL (due to lack of access to the underlying behavior policy and lack of access to a simulator during training) and not a strictly easier setting (akin to some slightly more ambitious form of imitation learning).\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667110173593}, {"id": "OQU6HolZ7R", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1028/Reviewer_mLZD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a regularization method for constraining policy improvements in offline RL. The proposed regularization is based on a lower bound to the mutual information between states and actions, and attempts to mitigate the issue of distribution shift that arises when the policy is queried on out-of-distribution actions.", "review_text": "Although there are some weaknesss especially with expeirments that I have pointed out, and which should be addressed, I am leaning towards acceptance because the key insight is simple, neat, and seems to be promising for reducing distribution shift in offline RL. However, the additional experiments/comaprisons are needed to confirm the beenfits. ", "strengths": "Strengths\n\n1. The proposed regularization is simple to implement and is well-motivated. It can be easily integrated with most offline RL framework as an additional term in the objective. \n\n2. The practical algorithm is discussed in detail and implementation details are clearly mentioned. \n\n3. The ablations and embedding visualizations provide nice intuitions about what MISA is learning empirically. \n\nWeaknesses\n\n1. Conceptually it is unclear how exactly is MUSA better than prior methods like CQL in mitigating distribution shift. There isn't enough intuition provided regarding this, beyond reference to experimental results and \"better mutual information estimation\" (in section 4.4) Why exactly should the MI estimation be better?\n\n2. The empirical results do not show strong benefit of MISA over baselines, partly because the hopper, cheetah, walker tasks are almost saturated. For the other tasks in Kitchen, Adroit, and Maze, not all baselines are evaluated, so empirical benefits are unclear.\n\n3. The lower bound result in Theorem 4.1 does not discuss anything about how tight the bound is, and whether the bound still holds in practice (i.e. in the experiments) with several approximations, including function approximations.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a regularization method for constraining policy improvements in offline RL. The proposed regularization is based on a lower bound to the mutual information between states and actions, and attempts to mitigate the issue of distribution shift that arises when the policy is queried on out-of-distribution actions.", "strength_and_weaknesses": "Strengths\n\n1. The proposed regularization is simple to implement and is well-motivated. It can be easily integrated with most offline RL framework as an additional term in the objective. \n\n2. The practical algorithm is discussed in detail and implementation details are clearly mentioned. \n\n3. The ablations and embedding visualizations provide nice intuitions about what MISA is learning empirically. \n\nWeaknesses\n\n1. Conceptually it is unclear how exactly is MUSA better than prior methods like CQL in mitigating distribution shift. There isn't enough intuition provided regarding this, beyond reference to experimental results and \"better mutual information estimation\" (in section 4.4) Why exactly should the MI estimation be better?\n\n2. The empirical results do not show strong benefit of MISA over baselines, partly because the hopper, cheetah, walker tasks are almost saturated. For the other tasks in Kitchen, Adroit, and Maze, not all baselines are evaluated, so empirical benefits are unclear.\n\n3. The lower bound result in Theorem 4.1 does not discuss anything about how tight the bound is, and whether the bound still holds in practice (i.e. in the experiments) with several approximations, including function approximations.", "clarity,_quality,_novelty_and_reproducibility": "The contribution is novel in my understanding, and sufficient details for experiments are provided. The writing is clear and easy to follow.", "summary_of_the_review": "Although there are some weaknesss especially with expeirments that I have pointed out, and which should be addressed, I am leaning towards acceptance because the key insight is simple, neat, and seems to be promising for reducing distribution shift in offline RL. However, the additional experiments/comaprisons are needed to confirm the beenfits. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666668789489}, {"id": "mxxWmcHBMvK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1028/Reviewer_aonx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new framework for handling distribution shift in offline RL, based on a mutual information based regularizer. The mutual information term gives a measure of the data collecting policy's actions on certain states. The regularizer computes a lower bound to the overall policy improvement step and the paper shows how this lower bound can still be useful for leading to overall improvements in offline RL. ", "review_text": "The work proposes a novel regularizer based on mutual information. This form of regularizer has been popular in online RL literature, and past works studied approximations to such MI terms. Therefore, the contributions and novelty of this work seems somewhat limited to me. Additionally, the empirical significance of the work is not clear enough. I am willing to re-evaluate my score if the paper clarity can be improved, and experimental significance made clearer. ", "strengths": "\t- The paper proposes a novel regularizer for offline RL and provides intuitions for why such a regularizer might be useful. Existing works have proposed conservative updates to offline RL, but this paper argues the need for such a regularizer. \n\t- Mutual information based regularizers have been well studied in online RL literature; so it is interesting to see a use case for it in offline RL; the lower bound is computed using existing approaches in the literature however (lemma 3.1 and 3.2)\n\t- The main claim of the paper is that this framework can unify two popular offline RL algorithms, namely CQL and TD3+BC. Overall, it leads to the existing algorithms being modified based on equations 5 and 6 and the work shows that this regularizer can lead to a lower bound to the policy improvement step (theorem 4.1)\n\t- The theoretical justification of this work, along with the derived algorithm, is well explained. However, such techniques are not completely novel and adapts from existing approaches. For example, equation 12 for integrating the regularizer into an offline RL framework is somewhat expected where the additonal term is required based on the approximation of the MI term. \n\t- Section 4.4 provides interesting insights for the justification of how the regularizer unifies existing algorithms. \n\t- Experiments are primarily done on the D4RL benchmark, and it seems there are marginal improvements due to the addition of the regularizer. However, I am surprised that even in appendix the full result plots are not shown, and it is difficult to judge the actual empirical significance of this work. \n\t- Overall, I think the paper is useful in terms of adapting a well known regularizer from online RL into offline RL. However, the resulting approximations for the MI term, and the way it is adapted into offline RL is not completely novel. Even if the algorithm is well derived from existing literature, the empirical significance of the work is not fully clear either. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new framework for handling distribution shift in offline RL, based on a mutual information based regularizer. The mutual information term gives a measure of the data collecting policy's actions on certain states. The regularizer computes a lower bound to the overall policy improvement step and the paper shows how this lower bound can still be useful for leading to overall improvements in offline RL. ", "strength_and_weaknesses": "\t- The paper proposes a novel regularizer for offline RL and provides intuitions for why such a regularizer might be useful. Existing works have proposed conservative updates to offline RL, but this paper argues the need for such a regularizer. \n\t- Mutual information based regularizers have been well studied in online RL literature; so it is interesting to see a use case for it in offline RL; the lower bound is computed using existing approaches in the literature however (lemma 3.1 and 3.2)\n\t- The main claim of the paper is that this framework can unify two popular offline RL algorithms, namely CQL and TD3+BC. Overall, it leads to the existing algorithms being modified based on equations 5 and 6 and the work shows that this regularizer can lead to a lower bound to the policy improvement step (theorem 4.1)\n\t- The theoretical justification of this work, along with the derived algorithm, is well explained. However, such techniques are not completely novel and adapts from existing approaches. For example, equation 12 for integrating the regularizer into an offline RL framework is somewhat expected where the additonal term is required based on the approximation of the MI term. \n\t- Section 4.4 provides interesting insights for the justification of how the regularizer unifies existing algorithms. \n\t- Experiments are primarily done on the D4RL benchmark, and it seems there are marginal improvements due to the addition of the regularizer. However, I am surprised that even in appendix the full result plots are not shown, and it is difficult to judge the actual empirical significance of this work. \n\t- Overall, I think the paper is useful in terms of adapting a well known regularizer from online RL into offline RL. However, the resulting approximations for the MI term, and the way it is adapted into offline RL is not completely novel. Even if the algorithm is well derived from existing literature, the empirical significance of the work is not fully clear either. \n", "clarity,_quality,_novelty_and_reproducibility": "- The novelty of this work is somewhat limited, given the approximations for the regularizer have been extensively studied in the past. The paper is however well written and easy to follow, with the theoretical derivation for the resulting updates properly derived. I do have concerns about the reproducibility of the work, since the results are only presented in the form of a table, without actual return plots, and it is not clear to me whether there is a bias variance trade-off due to the addition of this regularizer. Full results not being presented is a major concern for the empirical significance of this work. ", "summary_of_the_review": "The work proposes a novel regularizer based on mutual information. This form of regularizer has been popular in online RL literature, and past works studied approximations to such MI terms. Therefore, the contributions and novelty of this work seems somewhat limited to me. Additionally, the empirical significance of the work is not clear enough. I am willing to re-evaluate my score if the paper clarity can be improved, and experimental significance made clearer. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666624993575}, {"id": "skNrlE22y3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1028/Reviewer_SVYL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes to:\n\n1. Estimate the mutual information between state and action\n\n2. Use the estimate as regularization, by adding it to the policy and value function objective.", "review_text": "The proposed idea is novel and intuitive. But the paper writing can be significantly improved, and why the method outperforms baselines deserve more analysis to be more convincing.", "strengths": "Strength \n\nThe idea of estimating the mutual information between state and action pair and using the estimates to improve policy evaluation and improvement are intuitive.\n\nWeaknesses\n\nThe writing is not very clear. I list a few examples below: \n\n- In the abstract, \"Intuitively, mutual information … reflects how a behavior agent reacts to certain environment states during data collection\". While this is factually true, I fail to see how this is different from simply estimating the behavior policy. \n\n- The abstract also writes \"equivalent to maximizing the likeli- hood of a one-step improved policy on the offline dataset. In this way, we constrain the policy improvement direction to lie in the data manifold.\" I do not see how the former leads to the latter, or even what the latter means at all. AFAIK, the data manifold is not defined in the paper at all.\n\n- The introduction writes \"Though these methods are effective at alleviating the distributional shift problem of the learning policy, the im- proved policy is unconstrained and might still deviate from the data distribution.\" This sentence seems self-contradictory. If these methods are effective at alleviating the issue of distribution shift, then how can the learned policy still deviate from the data distribution.\n\nIn addition to writing, the paper also has two other weaknesses:\n\n- The paper claims TD3+BC and CQL are special cases, but it is unclear why their method outperforms these methods. The paper claims that their method leads to a better mutual information estimation. But what does \"better\" mean, and why does that lead to better performance compared to CQL? This is not explained at all.\n\n-  The method does not perform well when learning from uniform data. How about data where the behavior is uniform in some state, but much more constrained in others, which is the more realistic settings? The limitations of the methods seem to be an afterthought that does not receive adequate analysis.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes to:\n\n1. Estimate the mutual information between state and action\n\n2. Use the estimate as regularization, by adding it to the policy and value function objective.", "strength_and_weaknesses": "Strength \n\nThe idea of estimating the mutual information between state and action pair and using the estimates to improve policy evaluation and improvement are intuitive.\n\nWeaknesses\n\nThe writing is not very clear. I list a few examples below: \n\n- In the abstract, \"Intuitively, mutual information … reflects how a behavior agent reacts to certain environment states during data collection\". While this is factually true, I fail to see how this is different from simply estimating the behavior policy. \n\n- The abstract also writes \"equivalent to maximizing the likeli- hood of a one-step improved policy on the offline dataset. In this way, we constrain the policy improvement direction to lie in the data manifold.\" I do not see how the former leads to the latter, or even what the latter means at all. AFAIK, the data manifold is not defined in the paper at all.\n\n- The introduction writes \"Though these methods are effective at alleviating the distributional shift problem of the learning policy, the im- proved policy is unconstrained and might still deviate from the data distribution.\" This sentence seems self-contradictory. If these methods are effective at alleviating the issue of distribution shift, then how can the learned policy still deviate from the data distribution.\n\nIn addition to writing, the paper also has two other weaknesses:\n\n- The paper claims TD3+BC and CQL are special cases, but it is unclear why their method outperforms these methods. The paper claims that their method leads to a better mutual information estimation. But what does \"better\" mean, and why does that lead to better performance compared to CQL? This is not explained at all.\n\n-  The method does not perform well when learning from uniform data. How about data where the behavior is uniform in some state, but much more constrained in others, which is the more realistic settings? The limitations of the methods seem to be an afterthought that does not receive adequate analysis.", "clarity,_quality,_novelty_and_reproducibility": "The method is novel as far as I know. However, the writing is not clear and too generic. The paper does not release code, so I also do not have high rating as far as reproducibility is concerned.", "summary_of_the_review": "The proposed idea is novel and intuitive. But the paper writing can be significantly improved, and why the method outperforms baselines deserve more analysis to be more convincing.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666584813026}], "openreview_url": "https://openreview.net/forum?id=ygYXtbb3og3", "arxiv_id": "2210.07484", "paper_pdf": "papers/ygYXtbb3og3.pdf", "paper_pdf_sha256": "1e9e9c98bcb97f24b45d4b6854cb4af21028bfd48f1afbd9a12f19a653915446", "paper_pdf_bytes": 742416, "paper_pdf_source": "openreview", "code_url": "https://github.com/sail-sg/MISA", "code_repository": "sail-sg/MISA", "code_commit": "bd350e5ee650ad0a212d80a85c02b6fadea9104a", "code_archive": "repos/ygYXtbb3og3.zip", "code_archive_sha256": "1d2b3d6ebb2fbeb23ff768c5d1bbd9b9662cf531aa9fa5a53fc8a50cbf4a1875", "code_archive_bytes": 38636, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 32, "github_languages": {"Python": 100673}, "github_archived": false, "github_pushed_at": "2022-10-19T03:38:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mutual-information-regularized-offline-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hW2kwAcXq5w", "year": 2022, "status": "rejected", "title": "Discriminator-Weighted Offline Imitation Learning from Suboptimal Demonstrations", "authors": ["Haoran Xu", "Xianyuan Zhan", "Honglei Yin", "Huiling Qin"], "authorids": ["~Haoran_Xu4", "~Xianyuan_Zhan1", "~Honglei_Yin1", "~Huiling_Qin1"], "authors_source": "OpenReview API", "abstract": "We study the problem of offline Imitation Learning (IL) where an agent aims to learn an optimal expert behavior policy without additional online environment interactions. Instead, the agent is provided with a static offline dataset of state-action-next state transition triples from both optimal and non-optimal expert behaviors. This strictly offline imitation learning problem arises in many real-world problems, where environment interactions and expert annotations are costly. Prior works that address the problem either require that expert data occupies the majority proportion of the offline dataset, or need to learn a reward function and perform offline reinforcement learning (RL) based on the learned reward function. In this paper, we propose an imitation learning algorithm to address the problem without additional steps of reward learning and offline RL training for the case when demonstrations containing large-proportion of suboptimal data. Built upon behavioral cloning (BC), we introduce an additional discriminator to distinguish expert and non-expert data, we propose a cooperation strategy to boost the performance of both tasks, this will result in a new policy learning objective and surprisingly, we find its equivalence to a generalized BC objective, where the outputs of discriminator serve as the weights of the BC loss function. Experimental results show that the proposed algorithm can learn behavior policies that are much closer to the optimal policies than policies learned by baseline algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gLEWPLo5qkg", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper514/Reviewer_x4DH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors consider the problem of offline imitation learning in the presence of suboptimal datasets. In the presence of suboptimal data, classical baselines like behavior cloning suffer performance hits, the drop in performance often correlates positively with increase in number of suboptimal trajectories. In this work, the authors propose a novel learning objective inspired by the min-max formulation in GANs. Particularly the agent learns a discriminator to distinguish between samples from the expert and the suboptimal demonstrator. Building on prior work in cost-sensitive learning, this discriminator is used to reweight loss per sample in the offline buffer. \n\nThe proposed algorithm is evaluated on standard offline RL benchmarks, across multiple environments. In many environments, e.g Hopper-v2 style environments, the policy improves against strong imitation learning benchmarks. The discriminator (which takes action probabilities as input) is also evaluated in the context of offline policy evaluation. On Hopper-v2 datasets, the discriminator output is compared with true reward accumulated by multiple policies.", "review_text": "Strengths:\n+ The paper is well written and clearly communicates the motivation and contributions.\n+ Sufficient empirical experiments which support the core thesis of the paper. The baselines are selected carefully to provide meaningful comparison to the proposed approach.\n+ Using the jointly learned discriminator seems like a novel idea for policy optimization and evaluation. \n\n\nWeakness:\n+ The experiments on very similar datasets (e.x changing the fraction of true positives). While I agree that it is strong experiment, From Figure 1. it is clear that the performance of DWBC is not too sensitive to fraction of positive samples. It would be useful to have similar experiments on datasets collected from a random mix of optimal and suboptimal policies (instead of buffer of single learning policy). For example, mixing expert and random policies, or Adroit environments with human demonstrations.\n+ Offline Policy evaluation is evaluated on Hopper-v2 only. It would be useful to have a diversity in environments to evaluate this contribution more objectively.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors consider the problem of offline imitation learning in the presence of suboptimal datasets. In the presence of suboptimal data, classical baselines like behavior cloning suffer performance hits, the drop in performance often correlates positively with increase in number of suboptimal trajectories. In this work, the authors propose a novel learning objective inspired by the min-max formulation in GANs. Particularly the agent learns a discriminator to distinguish between samples from the expert and the suboptimal demonstrator. Building on prior work in cost-sensitive learning, this discriminator is used to reweight loss per sample in the offline buffer. \n\nThe proposed algorithm is evaluated on standard offline RL benchmarks, across multiple environments. In many environments, e.g Hopper-v2 style environments, the policy improves against strong imitation learning benchmarks. The discriminator (which takes action probabilities as input) is also evaluated in the context of offline policy evaluation. On Hopper-v2 datasets, the discriminator output is compared with true reward accumulated by multiple policies.", "main_review": "Strengths:\n+ The paper is well written and clearly communicates the motivation and contributions.\n+ Sufficient empirical experiments which support the core thesis of the paper. The baselines are selected carefully to provide meaningful comparison to the proposed approach.\n+ Using the jointly learned discriminator seems like a novel idea for policy optimization and evaluation. \n\n\nWeakness:\n+ The experiments on very similar datasets (e.x changing the fraction of true positives). While I agree that it is strong experiment, From Figure 1. it is clear that the performance of DWBC is not too sensitive to fraction of positive samples. It would be useful to have similar experiments on datasets collected from a random mix of optimal and suboptimal policies (instead of buffer of single learning policy). For example, mixing expert and random policies, or Adroit environments with human demonstrations.\n+ Offline Policy evaluation is evaluated on Hopper-v2 only. It would be useful to have a diversity in environments to evaluate this contribution more objectively.", "summary_of_the_review": "Overall, the paper is well motivated and provides promising results for leveraging suboptimal datasets for effective offline imitation learning. The problem is well motivated and the authors provide some novel insights into better objectives for behavior cloning. While the authors demonstrate some improvement over the provided baselines, I would encourage the authors to consider adding couple of more ablations, particularly across datasets with a mix of human and random trajectories. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethics concern.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636072535679}, {"id": "GGpgiVMOOtK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper514/Reviewer_xvWD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a new offline imitation learning algorithm, DWBC, for datasets that combine both optimal and suboptimal demonstrations. The approach is based on a modified behavioral cloning loss that weighs expert and non-expert data based on a learned discriminator. DWBC is compared against prior methods in OpenAI Gym tasks, and it is shown to yield better policies compared to the prior work. As a by-product, the method learns a discriminator that can be used to estimate the relative performance of any policies without rolling them out in the environment.", "review_text": "Overall, the results are quite promising, especially the fact that the algorithm can learn to mimic the expert from a really small number of expert demonstrations. For example, Figure 1 shows that it is possible to learn HalfCheetah from just 4 thousand expert samples (=4 episodes?). However, I do feel like that both the method and the derivation need some more clarity, and I also have some concerns regarding the experiments.\n\nThe derivation includes several inaccuracies and vague statements. The main innovation behind DWBC is to learn a discriminator and condition it on the policy that is being learned. The conditioning is motivated by the fact that it makes discrimination easier (given the optimal policy, one can discriminate the actions based on their probabilities only). It is then argued (after Equation 6) that learning the disciminator becomes more robust, if at the same time, the policy is optimized by minimizing the amount of useful information it can provide to the discriminator. This observation is the most important contribution of the paper, yet the justification seems insufficient and vague. Further, it is not obvious that having the derivative of the discriminator loss w.r.t the policy equal to zero (Equation 7) will be desirable, and that minimization of the final loss in Equation 8 will in fact lead to that condition to be true. That said, the results do indicate the algorithm performs well, so I'm mostly curious seeing a more rigorous derivation and justification that could help shed some light why the DWBC works well.\n\nMoreover, there are several other inaccuracies that make following the derivation hard. For example: \n* The loss $L_d$ should not depend on $s$ and $a$ as it does in Equation 6.\n* I think in Equation 8, the gradients need to be stopped from flowing to the discriminator.\n* How is the discriminator trained? Equation 8 is only minimized with respect to the policy. Is the same loss used for learning the discriminator?\n\nI also have several smaller comments on the experiments:\n\nThere seems to be something not quite right with the shading (standard deviation) in Figure 1. For some of the curves, the shaded region is really narrow compared to how much the curves change between iterations. Can you comment on that? Is the training set  different for each seed? \n\nPlease add axis labels to Figure 2.\n\nThe datasets used in the experiments have a really particular form as they are collected during online training and thus the expert and the sub-optimal data are highly related. It would be good to see a comparison where the datasets come from two completely different policies: one optimal expert policy, and another fixed but suboptimal policy. This would be a more realistic setup (i.e., if the data comes from policy that is trained online, then why do we need offline learning?).\n\nIn the offline policy selection experiment, do you use separate datasets for training the discriminator and evaluation? If not, then perhaps the discriminator is simply memorizing the training data. \n\nWhat is the return of the expert policies for the experiments in Figure 1?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new offline imitation learning algorithm, DWBC, for datasets that combine both optimal and suboptimal demonstrations. The approach is based on a modified behavioral cloning loss that weighs expert and non-expert data based on a learned discriminator. DWBC is compared against prior methods in OpenAI Gym tasks, and it is shown to yield better policies compared to the prior work. As a by-product, the method learns a discriminator that can be used to estimate the relative performance of any policies without rolling them out in the environment.", "main_review": "Overall, the results are quite promising, especially the fact that the algorithm can learn to mimic the expert from a really small number of expert demonstrations. For example, Figure 1 shows that it is possible to learn HalfCheetah from just 4 thousand expert samples (=4 episodes?). However, I do feel like that both the method and the derivation need some more clarity, and I also have some concerns regarding the experiments.\n\nThe derivation includes several inaccuracies and vague statements. The main innovation behind DWBC is to learn a discriminator and condition it on the policy that is being learned. The conditioning is motivated by the fact that it makes discrimination easier (given the optimal policy, one can discriminate the actions based on their probabilities only). It is then argued (after Equation 6) that learning the disciminator becomes more robust, if at the same time, the policy is optimized by minimizing the amount of useful information it can provide to the discriminator. This observation is the most important contribution of the paper, yet the justification seems insufficient and vague. Further, it is not obvious that having the derivative of the discriminator loss w.r.t the policy equal to zero (Equation 7) will be desirable, and that minimization of the final loss in Equation 8 will in fact lead to that condition to be true. That said, the results do indicate the algorithm performs well, so I'm mostly curious seeing a more rigorous derivation and justification that could help shed some light why the DWBC works well.\n\nMoreover, there are several other inaccuracies that make following the derivation hard. For example: \n* The loss $L_d$ should not depend on $s$ and $a$ as it does in Equation 6.\n* I think in Equation 8, the gradients need to be stopped from flowing to the discriminator.\n* How is the discriminator trained? Equation 8 is only minimized with respect to the policy. Is the same loss used for learning the discriminator?\n\nI also have several smaller comments on the experiments:\n\nThere seems to be something not quite right with the shading (standard deviation) in Figure 1. For some of the curves, the shaded region is really narrow compared to how much the curves change between iterations. Can you comment on that? Is the training set  different for each seed? \n\nPlease add axis labels to Figure 2.\n\nThe datasets used in the experiments have a really particular form as they are collected during online training and thus the expert and the sub-optimal data are highly related. It would be good to see a comparison where the datasets come from two completely different policies: one optimal expert policy, and another fixed but suboptimal policy. This would be a more realistic setup (i.e., if the data comes from policy that is trained online, then why do we need offline learning?).\n\nIn the offline policy selection experiment, do you use separate datasets for training the discriminator and evaluation? If not, then perhaps the discriminator is simply memorizing the training data. \n\nWhat is the return of the expert policies for the experiments in Figure 1?\n", "summary_of_the_review": "The results presented in the paper are promising, but there are several issues with the clarity of the derivation and some with the experiments, and thus the paper is not yet of sufficient quality to be published as is.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635787818765}, {"id": "dT2CndhvoPa", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper514/Reviewer_jCdL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an offline imitation learning framework that incorporates both optimal and suboptimal datasets to learn decision-making tasks, without requiring any reward annotations. To leverage high reward transitions from the suboptimal dataset, the authors formulate a discriminator that optimizes a positive-unlabeled learning objective, where positive samples come from the optimal dataset and unlabeled samples come from the suboptimal dataset. This discriminator is trained in an adversarial fashion along with the policy, resulting in a behavior cloning objective where samples from the optimal and suboptimal datasets are weighted differently according to the discriminator’s predictions. Experiments demonstrate that on a set of simulated locomotion domains, the proposed algorithm can leverage the suboptimal dataset to learn more performant policies compared to vanilla behavior cloning objectives and prior offline IL/RL baselines.", "review_text": "The problem studied in this paper is quite relevant and important — the ability to re-use large, noisy offline datasets to solve new tasks. This paper motivates the problem well in the introduction and lays out the preliminaries clearly. The related work section also makes references to much of the relevant work, though the subsection on offline RL could be enhanced by expanding on the limitations of offline RL (something the authors already did in the introduction). The method appears to be novel and the discussion framing this method as a weighted behavior cloning objective connects nicely to prior work. Alongside these strengths, there are a number of concerns which I enumerate below:\n1. The derivation presented in section 3.2 is a bit confusing and unintuitive. First, I’m not sure I agree with the justification for the why policy outputs needs to be part of the discriminator input. The authors state that it helps the discriminator distinguish between expert and non-expert actions better, but I would like to see this design choice verified empirically — ie. comparing to a discriminator that only takes the state and action as input, possibly without the adversarial formulation. In addition, the discussion on the adversarial policy learning objective also can benefit from additional motivation and details. In particular, why is “providing as little information in $\\log \\pi$ as possible” the best way to learn a discriminator? Why is the latter necessarily equivalent to “maximizing $L_d$ for $\\pi$”? This subsection is, in my opinion, the weakest part of the paper. The discussion here can be improved by a combination of providing a more formal framework for the adversarial learning problem, complementing the formal framework with intuitive explanations, and connecting the ideas here to prior literature.\n2. I am having trouble understanding the derivation details in Appendix B. Specifically, why does $\\frac{\\partial F}{\\partial \\log \\pi}$ equal to $\\frac{\\partial L_d}{\\partial d} \\cdot \\frac{\\partial d}{\\partial \\log \\pi}$? And also I don’t see where the $\\frac{\\partial d}{\\partial \\log \\pi}$ term in incorporated into the final derivation at the end of the section. Can you please clarify these steps?\n3. One critically missing baseline is to train the discriminator defined in equation (4) once, without all the adversarial learning machinery presented in section 3.2 — and use the discriminator weights to weigh different samples when training the policy. This can help justify all of the additional complexities presented in section 3.2 of updating the discriminator and policy in an alternating optimization scheme.\n4. While the proposed method appears to work well in locomotion domains, it remains unclear how the approach would scale to more complex settings, such as the robotic manipulation datasets for the kitchen and adroit tasks in D4RL [1] and the manipulation tasks in robomimic [2]. In principle, it should not be too difficult to run experiments on these datasets as well and such experiments would certainly enhance the scope of this paper. That said, given that ICLR is primarily focused on core machine learning methods and less so on strong empirical evaluations, this is not the primary concern in this review. \n5. It is unclear to me why BC-all and BCND are constant lines, while the other baselines are curves. Shouldn’t all the baselines be shown as curves — ie. where the performance is changing across training iterations? \n6. While it can be implied from the paper as is, pseudocode or a text description of the full training scheme would be nice to have. In particular, I was wondering how often the discriminator is updated relative to the policy — does one update more frequently than the other? \n\n[1] Fu et al., D4RL: Datasets for Deep Data-Driven Reinforcement Learning, 2020\n\n[2] Mandlekar et al., What Matters in Learning from Offline Human Demonstrations for Robot Manipulation, CoRL 2021", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an offline imitation learning framework that incorporates both optimal and suboptimal datasets to learn decision-making tasks, without requiring any reward annotations. To leverage high reward transitions from the suboptimal dataset, the authors formulate a discriminator that optimizes a positive-unlabeled learning objective, where positive samples come from the optimal dataset and unlabeled samples come from the suboptimal dataset. This discriminator is trained in an adversarial fashion along with the policy, resulting in a behavior cloning objective where samples from the optimal and suboptimal datasets are weighted differently according to the discriminator’s predictions. Experiments demonstrate that on a set of simulated locomotion domains, the proposed algorithm can leverage the suboptimal dataset to learn more performant policies compared to vanilla behavior cloning objectives and prior offline IL/RL baselines.", "main_review": "The problem studied in this paper is quite relevant and important — the ability to re-use large, noisy offline datasets to solve new tasks. This paper motivates the problem well in the introduction and lays out the preliminaries clearly. The related work section also makes references to much of the relevant work, though the subsection on offline RL could be enhanced by expanding on the limitations of offline RL (something the authors already did in the introduction). The method appears to be novel and the discussion framing this method as a weighted behavior cloning objective connects nicely to prior work. Alongside these strengths, there are a number of concerns which I enumerate below:\n1. The derivation presented in section 3.2 is a bit confusing and unintuitive. First, I’m not sure I agree with the justification for the why policy outputs needs to be part of the discriminator input. The authors state that it helps the discriminator distinguish between expert and non-expert actions better, but I would like to see this design choice verified empirically — ie. comparing to a discriminator that only takes the state and action as input, possibly without the adversarial formulation. In addition, the discussion on the adversarial policy learning objective also can benefit from additional motivation and details. In particular, why is “providing as little information in $\\log \\pi$ as possible” the best way to learn a discriminator? Why is the latter necessarily equivalent to “maximizing $L_d$ for $\\pi$”? This subsection is, in my opinion, the weakest part of the paper. The discussion here can be improved by a combination of providing a more formal framework for the adversarial learning problem, complementing the formal framework with intuitive explanations, and connecting the ideas here to prior literature.\n2. I am having trouble understanding the derivation details in Appendix B. Specifically, why does $\\frac{\\partial F}{\\partial \\log \\pi}$ equal to $\\frac{\\partial L_d}{\\partial d} \\cdot \\frac{\\partial d}{\\partial \\log \\pi}$? And also I don’t see where the $\\frac{\\partial d}{\\partial \\log \\pi}$ term in incorporated into the final derivation at the end of the section. Can you please clarify these steps?\n3. One critically missing baseline is to train the discriminator defined in equation (4) once, without all the adversarial learning machinery presented in section 3.2 — and use the discriminator weights to weigh different samples when training the policy. This can help justify all of the additional complexities presented in section 3.2 of updating the discriminator and policy in an alternating optimization scheme.\n4. While the proposed method appears to work well in locomotion domains, it remains unclear how the approach would scale to more complex settings, such as the robotic manipulation datasets for the kitchen and adroit tasks in D4RL [1] and the manipulation tasks in robomimic [2]. In principle, it should not be too difficult to run experiments on these datasets as well and such experiments would certainly enhance the scope of this paper. That said, given that ICLR is primarily focused on core machine learning methods and less so on strong empirical evaluations, this is not the primary concern in this review. \n5. It is unclear to me why BC-all and BCND are constant lines, while the other baselines are curves. Shouldn’t all the baselines be shown as curves — ie. where the performance is changing across training iterations? \n6. While it can be implied from the paper as is, pseudocode or a text description of the full training scheme would be nice to have. In particular, I was wondering how often the discriminator is updated relative to the policy — does one update more frequently than the other? \n\n[1] Fu et al., D4RL: Datasets for Deep Data-Driven Reinforcement Learning, 2020\n\n[2] Mandlekar et al., What Matters in Learning from Offline Human Demonstrations for Robot Manipulation, CoRL 2021", "summary_of_the_review": "My reaction to this paper is mixed. On one hand, the introduction, preliminaries, and related work are well laid out. On the other hand, I had several confusions about the method and have some concerns about the experiments (see the main review for specific details). As it stands, I think this paper is marginally below the acceptance threshold. I hope that the authors can diligently address the concerns that I raised, at which point I will reconsider my recommendation.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635546477127}, {"id": "Xpqe7Z1T2Z", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper514/Reviewer_nHQm"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper deals with the following setup: offline imitation learning in the presence of both an expert dataset and a non-expert dataset. More precisely, the goal is to learn a policy as close as possible to the one(s) that generated the samples in a dataset $D_e$, while making the most of samples in a non-expert dataset $D_o$. The “reward” information is not present/used in the dataset.\nThe authors draw inspiration from the positive-unlabeled classification as well as the adversarial imitation learning literature to propose a new algorithm to tackle this problem. They interleave the training of a discriminator and a policy. The discriminator is trained to discriminate between expert and non-expert dataset (using a positive/unlabeled loss) and takes as input the state, the action and the logit of the policy $\\pi(a | s)$. The policy is trained to imitate the expert on $D_e$ and to “fool” the discriminator. \n\nThe authors present results on four environments from the Gym Mujoco suite with datasets extracted from the D4RL datasets.\n", "review_text": "**Strengths**\n- The paper is overall well written and easy to follow. The justification of the setup makes sense and is well explained.\n- The setting is very interesting and has a great potential impact on the community.\n\n**Weaknesses**\n- Method\n    - The authors fail to mention a very close work that already applied positive-unlabeled learning to the Imitation learning setup: PUGAIL (https://arxiv.org/abs/1911.00459). This greatly limits the novelty of this work.\n    - I find the theoretical derivations confusing. In particular, in appendix B, proposition 1, is “when” an “if and only if” condition or just an implication?\n    - I also don’t understand why the authors need to introduce the function F. Overall the mathematical derivations are unclear and would benefit for more details (keeping them shorter in the main text but developing them correctly in appendix for example).\n\n- Related Work\n    - I think a few more works would be worth mentioning. In particular, some previous work already considered having access to two datasets (on expert, one non-expert), like CSI  (https://hal-supelec.archives-ouvertes.fr/hal-00869804/document)  or MILO (https://arxiv.org/pdf/2106.03207.pdf)\n- Evaluation\n    - Authors write “the proposed algorithm can learn behaviors that are much closer to the optimal policies” yet they fail at providing experiments that support this claim as they only study the return of the learnt policies. I would recommend that they checkout this work (https://arxiv.org/abs/2105.12034 ) that provides insights on how to evaluate models in the context of imitation learning. What is more, the plots don’t show the average return of D_e which makes it impossible to use the return as a proxy to study how “close” the policy is from the demonstrations.\n- Experiments\n    - The experimental setup is good but a bit “light”. As written by the authors, their method is quite fast to train, so why not provide more results on different setups, e.g. 1) with very little data (e.g. like 1 or 5 trajectories only)? with very random data in D_o? with human expert data in D_e? with more complicated environments like Adroit? (All these environments/datasets are available in D4RL).\n    - Please report the average return in D_e as a horizontal bar in the plots. Otherwise it is impossible to calibrate the results as a reader.\n- Writing\n    - In the abstract, the authors say “both optimal and non-optimal expert behaviors”. I find this expression a bit confusing. It suggests that the algorithm will only work if the D_o is actually made of expert but slightly suboptimal trajectories.\n    - nit: In 3.2, “much hard” -> “much harder”, “which we denote it as” -> “which we denote as”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper deals with the following setup: offline imitation learning in the presence of both an expert dataset and a non-expert dataset. More precisely, the goal is to learn a policy as close as possible to the one(s) that generated the samples in a dataset $D_e$, while making the most of samples in a non-expert dataset $D_o$. The “reward” information is not present/used in the dataset.\nThe authors draw inspiration from the positive-unlabeled classification as well as the adversarial imitation learning literature to propose a new algorithm to tackle this problem. They interleave the training of a discriminator and a policy. The discriminator is trained to discriminate between expert and non-expert dataset (using a positive/unlabeled loss) and takes as input the state, the action and the logit of the policy $\\pi(a | s)$. The policy is trained to imitate the expert on $D_e$ and to “fool” the discriminator. \n\nThe authors present results on four environments from the Gym Mujoco suite with datasets extracted from the D4RL datasets.\n", "main_review": "**Strengths**\n- The paper is overall well written and easy to follow. The justification of the setup makes sense and is well explained.\n- The setting is very interesting and has a great potential impact on the community.\n\n**Weaknesses**\n- Method\n    - The authors fail to mention a very close work that already applied positive-unlabeled learning to the Imitation learning setup: PUGAIL (https://arxiv.org/abs/1911.00459). This greatly limits the novelty of this work.\n    - I find the theoretical derivations confusing. In particular, in appendix B, proposition 1, is “when” an “if and only if” condition or just an implication?\n    - I also don’t understand why the authors need to introduce the function F. Overall the mathematical derivations are unclear and would benefit for more details (keeping them shorter in the main text but developing them correctly in appendix for example).\n\n- Related Work\n    - I think a few more works would be worth mentioning. In particular, some previous work already considered having access to two datasets (on expert, one non-expert), like CSI  (https://hal-supelec.archives-ouvertes.fr/hal-00869804/document)  or MILO (https://arxiv.org/pdf/2106.03207.pdf)\n- Evaluation\n    - Authors write “the proposed algorithm can learn behaviors that are much closer to the optimal policies” yet they fail at providing experiments that support this claim as they only study the return of the learnt policies. I would recommend that they checkout this work (https://arxiv.org/abs/2105.12034 ) that provides insights on how to evaluate models in the context of imitation learning. What is more, the plots don’t show the average return of D_e which makes it impossible to use the return as a proxy to study how “close” the policy is from the demonstrations.\n- Experiments\n    - The experimental setup is good but a bit “light”. As written by the authors, their method is quite fast to train, so why not provide more results on different setups, e.g. 1) with very little data (e.g. like 1 or 5 trajectories only)? with very random data in D_o? with human expert data in D_e? with more complicated environments like Adroit? (All these environments/datasets are available in D4RL).\n    - Please report the average return in D_e as a horizontal bar in the plots. Otherwise it is impossible to calibrate the results as a reader.\n- Writing\n    - In the abstract, the authors say “both optimal and non-optimal expert behaviors”. I find this expression a bit confusing. It suggests that the algorithm will only work if the D_o is actually made of expert but slightly suboptimal trajectories.\n    - nit: In 3.2, “much hard” -> “much harder”, “which we denote it as” -> “which we denote as”\n", "summary_of_the_review": "I believe this is an interesting idea in an interesting setup yet it lacks novelty and it would deserve more work, notably on the experimental part, before being published.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635501899098}], "openreview_url": "https://openreview.net/forum?id=hW2kwAcXq5w", "arxiv_id": "2207.10050", "paper_pdf": "papers/hW2kwAcXq5w.pdf", "paper_pdf_sha256": "bb1da13f3b8789df5b81ba3c2b6b6c2bcf61740ff52ec2fb5e3ab39b8feff90c", "paper_pdf_bytes": 603665, "paper_pdf_source": "openreview", "code_url": "https://github.com/ryanxhr/DWBC", "code_repository": "ryanxhr/DWBC", "code_commit": "a1e9d8a068c478128dea0b07d55a375b064eab7e", "code_archive": "repos/hW2kwAcXq5w.zip", "code_archive_sha256": "9ab3af66ec9eaf72c2e131aa661558255e9c7af4c0e6b307f0303575e5978d39", "code_archive_bytes": 10563, "code_file_count": 6, "code_extensions": {".py": 5, ".sh": 1}, "github_disk_usage_kb": 30, "github_languages": {"Python": 29638, "Shell": 1208}, "github_archived": false, "github_pushed_at": "2023-01-05T05:02:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discriminator-weighted-offline-imitation-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "thhdrl4IdMm", "year": 2021, "status": "rejected", "title": "A Chain Graph Interpretation of Real-World Neural Networks", "authors": ["Yuesong Shen", "Daniel Cremers"], "authorids": ["~Yuesong_Shen1", "~Daniel_Cremers1"], "authors_source": "OpenReview API", "abstract": "The last decade has witnessed a boom of deep learning research and applications achieving state-of-the-art results in various domains. However, most advances have been established empirically, and their theoretical analysis remains lacking. One major issue is that our current interpretation of neural networks (NNs) as function approximators is too generic to support in-depth analysis. In this paper, we remedy this by proposing an alternative interpretation that identifies NNs as chain graphs (CGs) and feed-forward as an approximate inference procedure. The CG interpretation specifies the nature of each NN component within the rich theoretical framework of probabilistic graphical models, while at the same time remains general enough to cover real-world NNs with arbitrary depth, multi-branching and varied activations, as well as common structures including convolution / recurrent layers, residual block and dropout. We demonstrate with concrete examples that the CG interpretation can provide novel theoretical support and insights for various NN techniques, as well as derive new deep learning approaches such as the concept of partially collapsed feed-forward inference. It is thus a promising framework that deepens our understanding of neural networks and provides a coherent theoretical formulation for future deep learning research.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JkUt_OmaXIJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1104/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Update: after reading the feedback and discussing with the other reviewers, I decide to keep my score unchanged.\n\nOriginal comments:\nIn this paper, the authors provide new interpretation of neural networks via chain graphs, which can be used as a new theoretical framework to understand the behavior of neural networks.\n\nPros.\n\n1. Although both PMG and Neural networks are based on graphs, very little research has been proposed to bring the two fields together. This paper provides an interesting result to bridge the two previously unrelated fields, which equips the community with useful insights to spark new research directions.\n\n2. This paper helps the community to fully utilize the existing works from PGM to solve current obstacles in deep learning.\n\n3. The open questions are also interesting. I would be particular interested in the third question.\n\nOverall, I think this is an interesting paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting paper", "review": "Update: after reading the feedback and discussing with the other reviewers, I decide to keep my score unchanged.\n\nOriginal comments:\nIn this paper, the authors provide new interpretation of neural networks via chain graphs, which can be used as a new theoretical framework to understand the behavior of neural networks.\n\nPros.\n\n1. Although both PMG and Neural networks are based on graphs, very little research has been proposed to bring the two fields together. This paper provides an interesting result to bridge the two previously unrelated fields, which equips the community with useful insights to spark new research directions.\n\n2. This paper helps the community to fully utilize the existing works from PGM to solve current obstacles in deep learning.\n\n3. The open questions are also interesting. I would be particular interested in the third question.\n\nOverall, I think this is an interesting paper.", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603977396198}, {"id": "0qta0j1HReC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1104/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors look for ways to frame neural networks in terms of probabilistic models. This is a worthy goal and I believe the authors should pursue this path with enthusiasm. The theme is relevant, however the novel ideas that are in the paper seem to have relatively small significance. The translation from neural networks to chain graphs is not surprising; actually if I understand the translation we basically get a Bayesian network out of the directed acyclic graph encoding the connections in a neural network (I must say that the translation could be presented in a more didactic fashion. In any case, given this translation, the authors are able to frame some techniques from neural networks as techniques from chain graphs. But this feels too forced a translation, because neural networks do have a natural interpretation and usage as function approximators; why should one go all this ways to find a probabilistic translation that explains a few well known things, and opens the possibility of some probabilistic algorithms that could anyway be derived in the context of neural networks? One example: given the proposed translation, it is obvious that sequential models will appear as dynamic chain graphs, why is this useful in any way? \n\nOn top of this, I find it troublesome that the authors \"prove\" results by saying things like \"approximately linear\" and \"approximately\" this and that. How can these results be proven given this level of informal justification. I really think this should be fixed before publication. \n\nSome sentences are a bit confusing. In Page 3, for instance, the authors say that a model with P(X|Pa(X)) is modeled by a CRF; usually a CRF models a discriminative model, is this the case here? What exactly is going on? Also the last paragraph of Section 3.2 is very hard to parse (the main point there is not clear at all). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising idea but results are too simple and not adequately presented.", "review": "The authors look for ways to frame neural networks in terms of probabilistic models. This is a worthy goal and I believe the authors should pursue this path with enthusiasm. The theme is relevant, however the novel ideas that are in the paper seem to have relatively small significance. The translation from neural networks to chain graphs is not surprising; actually if I understand the translation we basically get a Bayesian network out of the directed acyclic graph encoding the connections in a neural network (I must say that the translation could be presented in a more didactic fashion. In any case, given this translation, the authors are able to frame some techniques from neural networks as techniques from chain graphs. But this feels too forced a translation, because neural networks do have a natural interpretation and usage as function approximators; why should one go all this ways to find a probabilistic translation that explains a few well known things, and opens the possibility of some probabilistic algorithms that could anyway be derived in the context of neural networks? One example: given the proposed translation, it is obvious that sequential models will appear as dynamic chain graphs, why is this useful in any way? \n\nOn top of this, I find it troublesome that the authors \"prove\" results by saying things like \"approximately linear\" and \"approximately\" this and that. How can these results be proven given this level of informal justification. I really think this should be fixed before publication. \n\nSome sentences are a bit confusing. In Page 3, for instance, the authors say that a model with P(X|Pa(X)) is modeled by a CRF; usually a CRF models a discriminative model, is this the case here? What exactly is going on? Also the last paragraph of Section 3.2 is very hard to parse (the main point there is not clear at all). ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603945208021}, {"id": "9A3BfoaoUYc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1104/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tries to interpret neural networks with chain graphs that provides theoretical analysis on various neural network components. Furthermore, this chain graph interpretation has been used to propose a new approach (architecture), which is a partially collapsed feed-forward. A layered chain graph representation is adopted to formulate the neural networks with layered chain graphs. This further establishes to interpret feed-forward as an approximate probabilistic inference with using linear approximations. Some concrete examples are shown to be analyzed  based on the chain graph formulation. \n\nThe overall context (analysis) seems straightforward to interpret the neural networks with chain graphs, but it is hard to achieve some meaningful information from this new interpretation to improve the current neural network models in terms of learning procedure or optimization.  The proposed partially collapsed feed-forward is a good example to come up with a new approach based on the chain graph interpretation. However, in terms of performance and complexity, it is practically not showing impressive improvements compared to the baseline methods. Moreover, it seems quite similar to previous works as far as I remember and one similar work is 'stochastic feedforward neural networks'. I fully agree the future works (open questions) in the conclusion and discussion section that this work still needs more investigations although this paper is a good initiative work.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "alternative interpretation based on chain graph", "review": "This paper tries to interpret neural networks with chain graphs that provides theoretical analysis on various neural network components. Furthermore, this chain graph interpretation has been used to propose a new approach (architecture), which is a partially collapsed feed-forward. A layered chain graph representation is adopted to formulate the neural networks with layered chain graphs. This further establishes to interpret feed-forward as an approximate probabilistic inference with using linear approximations. Some concrete examples are shown to be analyzed  based on the chain graph formulation. \n\nThe overall context (analysis) seems straightforward to interpret the neural networks with chain graphs, but it is hard to achieve some meaningful information from this new interpretation to improve the current neural network models in terms of learning procedure or optimization.  The proposed partially collapsed feed-forward is a good example to come up with a new approach based on the chain graph interpretation. However, in terms of performance and complexity, it is practically not showing impressive improvements compared to the baseline methods. Moreover, it seems quite similar to previous works as far as I remember and one similar work is 'stochastic feedforward neural networks'. I fully agree the future works (open questions) in the conclusion and discussion section that this work still needs more investigations although this paper is a good initiative work.\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603870136771}, {"id": "KC5znrpUtcX", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1104/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "###############################################################################\n\nSummary\n\nSummarize what the paper claims to contribute. Be positive and generous.\n\nThe authors propose an interpretation of feed-forward neural networks and recursive neural networks as chain graphs (CGs). They claim this new interpretation can provide novel theoretical support an insights to existing techniques, as well allow for the derivation of new approaches.\n\n###############################################################################\n\nPros and cons\n\nPros:\n - the text is well-written\n - the attempt at studying neural networks under a graphical probabilistic model perspective is praisable\n\nCons:\n - the chain graph interpretation put forward by the authors is superfluous, as in reality the definition found in the apper is that of DAGs with particular parametric constraints\n - most of the results presented by the authors do not rely on the CG interpretation, and can already be found in the litterature\n - the paper is missing a connexion to stochastic feedforward neural networks (SFNNs), to which the proposed interpretation is extremely similar\n\n###############################################################################\n\nRecommendation\n\nI recommend rejection of the paper, for two reasons. First, I believe most of the contributions proposed in the paper are not novel, as they can already be found in published form. See my detailed comments below. Second, the CG interpretation put forward by the authors is trivial and superfluous, since the authors in reality only consider CGs restricted to DAGs. As such, the whole claim of the paper that CGs can give a new, relevant perspective to NNs, is not credible.\n\n###############################################################################\n\nQuestions to authors\n\nI would like the authors to comment on the fact that their CG interpretation is indeed the LWF-CG interpretation, restricted to DAGs (singleton chain components). Furthermore, I would appreciate if the authors could relate their approach to the SFNN model, and what differs from that interpretation.\n\n###############################################################################\n\nDetailed comments:\n\np.1 §3: an efficient approximate probabilistic inference -> What is meant here by efficient ? Is it an unbiased estimator ? What is meant by inference ? Computing $p(y|x)$ ? $\\arg \\max_y p(y|x)$ ?\n\np.2 §2: The chain graph model -> There exists at least 4 chain graph interpretations. See\nDrton, Mathias. „Discrete Chain Graph Models“. In: Bernoulli 15 (Sept. 2009),\npp. 736–753.\nAmong those, there are two which subsume UGs and DAGs, while the two others subsume BGs and DAGs. If you follow the CG interpretation from Koller and Friedman then you assume NNs are LWF-CGs.\nI strongly suggest that you make it explicit which chain graph interpretation you follow, for the sake of clarity. These interpretations are not equivalent.\n\np.2 §2: there exists a series of works [...] -> I believe the list is much bigger than that. You forget very popular models which are direct instanciation of PGMs, such as VAEs, HMMs, LDAs, GMMs, CRFs, and all of their variants.\n\np.2 §4: an approximate probabilistic inference procedure -> What is meant by that ?\n\np.2 §4: provides additional insights -> such as ?\n\np.3 §2: layered chain graph -> A chain graph is always layered into chain components. The name \"layered chain graph\" is confusing, as it seems to imply that some chain graphs may not be layered. Since you are using the LWF-CG interpretation, you might point to the relevant papers where the graphical model was introduced, and simply re-use the establihsed LWF-CG name from the litterature, instead of inventing a new one.\nLauritzen, Steffen L. and Wermuth, N. „Graphical Models for Associations\nbetween Variables, some of which are Qualitative and some Quantitative“. In:\nThe Annals of Statistics 17.1 (Mar. 1989), pp. 31–57\nFrydenberg, M. (1990). The chain graph Markov property. Scand. J. Statist. 17 333–353. MR1096723\n\np.3 eq.1: This factorization is not that of a CG, but that of a DAG. Since you assume no undirected connection between variables in the same layer, then all chain component in your chain graph actually contain a single variable. What is the point of introducing the conecpt of CGs then, if you only consider CGs which are restricted to DAGs ?\n\np.3 eq.2: The LWF-CG interpretation, which you seem to follow since you refer to Koller and Friedman, does not imply this factorization for the graphs you consider. Since each chain component $K$ in your graph is a singleton, each $X_i$ and its parents form a clique in the closure graph $K$. As such, the CG structure does not imply any further factorization for p(x|pa(x))... It does seem your interpretation of NNs is not as general CGs, but as very specific DAGs with parametric constraints. Furthermore, you do not define what are $T$ and $f$ here.\n\np.3 Section 2.2: I believe your CG interpretation is simply a reformulation of stochastic neural networks, for which there already exists a great body of work. See, eg:\n\nEric Jang Shixiang Gu Ben Poole. Categorical Reparameterization with Gumbel-Softmax. ICLR (2017)\n\nYoshua Bengio, Nicholas Leonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.\n\nYichuan Tang, Ruslan Salakhutdinov. Learning Stochastic Feedforward Neural Networks. NIPS 2013\n\np.4 Proposition 2.1: This result seems very limited, since it assumes linear layers, and in the end a linear NN.\n\np.4 Corollary 2: What are alpha and beta here ? How do they relate to $f$, $e$ or $T$ from Definition 1 ? What are $e_i^l$ and $s_i^l$ here ?\n\np.4 §6: the modularity of chain components justifies transfer learning via partial reuse of pre-trained networks -> I do not understand this argument. What is meant here by \"modularity\" ?\n\np.4 §7: However, feed-forward is no longer applicable through these intra-connected layers. -> You finally give an example here that corresponds to non-trivial chain graphs (with chain components greater than 1), and recognize that your theoretical results do not apply any more. This contradicts your claim that CGs offer a general interpretation of NNs with theoretical support. I believe the CG interpretation is not justified, nor required to derive the results you present in this paper.\n\np.5 Definition 2: I fail to see the point of introducing this new concept, which is a re-definition of residual blocks.\n\np.5 §4: While a vanilla layered [...] -> I fail to understand where the CG interpretation fits in this argument.\n\np.5 Proposition 4: This result seems trivial to me, and does not require the concept of a CG. Once you show that a linear layer followed by a specific activation can be interpreted as a parametric model for $p(\\textit{out}|\\textit{in})$, then any NN that unrolls as an acyclic directed graph can be interpreted as a probabilistic model.\n\np.6 §1: The simple recurrent layer, [...] -> Again, I do not see where the CG interpretation gives any insight here.\n\np.6 Section 3.3: This result, again, does not require the CG interpretation. It is also already know. See, e.g., Pierre Baldi and Peter J. Sadowski. Understanding Dropout. NIPS 2013\n\np.6 Section 4: It seems to me you are reinventing SFFNs. See, Yichuan Tang and Ruslan Salakhutdinov. Stochastic Feedforward Neural Network. NIPS 2013\n\np.7 §2: the sampling operation is not differentiable -> This is not true. See, e.g.,  the reparameterization trick for VAEs. You mention this in the next sentence...", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper fails to deliver on its promises", "review": "###############################################################################\n\nSummary\n\nSummarize what the paper claims to contribute. Be positive and generous.\n\nThe authors propose an interpretation of feed-forward neural networks and recursive neural networks as chain graphs (CGs). They claim this new interpretation can provide novel theoretical support an insights to existing techniques, as well allow for the derivation of new approaches.\n\n###############################################################################\n\nPros and cons\n\nPros:\n - the text is well-written\n - the attempt at studying neural networks under a graphical probabilistic model perspective is praisable\n\nCons:\n - the chain graph interpretation put forward by the authors is superfluous, as in reality the definition found in the apper is that of DAGs with particular parametric constraints\n - most of the results presented by the authors do not rely on the CG interpretation, and can already be found in the litterature\n - the paper is missing a connexion to stochastic feedforward neural networks (SFNNs), to which the proposed interpretation is extremely similar\n\n###############################################################################\n\nRecommendation\n\nI recommend rejection of the paper, for two reasons. First, I believe most of the contributions proposed in the paper are not novel, as they can already be found in published form. See my detailed comments below. Second, the CG interpretation put forward by the authors is trivial and superfluous, since the authors in reality only consider CGs restricted to DAGs. As such, the whole claim of the paper that CGs can give a new, relevant perspective to NNs, is not credible.\n\n###############################################################################\n\nQuestions to authors\n\nI would like the authors to comment on the fact that their CG interpretation is indeed the LWF-CG interpretation, restricted to DAGs (singleton chain components). Furthermore, I would appreciate if the authors could relate their approach to the SFNN model, and what differs from that interpretation.\n\n###############################################################################\n\nDetailed comments:\n\np.1 §3: an efficient approximate probabilistic inference -> What is meant here by efficient ? Is it an unbiased estimator ? What is meant by inference ? Computing $p(y|x)$ ? $\\arg \\max_y p(y|x)$ ?\n\np.2 §2: The chain graph model -> There exists at least 4 chain graph interpretations. See\nDrton, Mathias. „Discrete Chain Graph Models“. In: Bernoulli 15 (Sept. 2009),\npp. 736–753.\nAmong those, there are two which subsume UGs and DAGs, while the two others subsume BGs and DAGs. If you follow the CG interpretation from Koller and Friedman then you assume NNs are LWF-CGs.\nI strongly suggest that you make it explicit which chain graph interpretation you follow, for the sake of clarity. These interpretations are not equivalent.\n\np.2 §2: there exists a series of works [...] -> I believe the list is much bigger than that. You forget very popular models which are direct instanciation of PGMs, such as VAEs, HMMs, LDAs, GMMs, CRFs, and all of their variants.\n\np.2 §4: an approximate probabilistic inference procedure -> What is meant by that ?\n\np.2 §4: provides additional insights -> such as ?\n\np.3 §2: layered chain graph -> A chain graph is always layered into chain components. The name \"layered chain graph\" is confusing, as it seems to imply that some chain graphs may not be layered. Since you are using the LWF-CG interpretation, you might point to the relevant papers where the graphical model was introduced, and simply re-use the establihsed LWF-CG name from the litterature, instead of inventing a new one.\nLauritzen, Steffen L. and Wermuth, N. „Graphical Models for Associations\nbetween Variables, some of which are Qualitative and some Quantitative“. In:\nThe Annals of Statistics 17.1 (Mar. 1989), pp. 31–57\nFrydenberg, M. (1990). The chain graph Markov property. Scand. J. Statist. 17 333–353. MR1096723\n\np.3 eq.1: This factorization is not that of a CG, but that of a DAG. Since you assume no undirected connection between variables in the same layer, then all chain component in your chain graph actually contain a single variable. What is the point of introducing the conecpt of CGs then, if you only consider CGs which are restricted to DAGs ?\n\np.3 eq.2: The LWF-CG interpretation, which you seem to follow since you refer to Koller and Friedman, does not imply this factorization for the graphs you consider. Since each chain component $K$ in your graph is a singleton, each $X_i$ and its parents form a clique in the closure graph $K$. As such, the CG structure does not imply any further factorization for p(x|pa(x))... It does seem your interpretation of NNs is not as general CGs, but as very specific DAGs with parametric constraints. Furthermore, you do not define what are $T$ and $f$ here.\n\np.3 Section 2.2: I believe your CG interpretation is simply a reformulation of stochastic neural networks, for which there already exists a great body of work. See, eg:\n\nEric Jang Shixiang Gu Ben Poole. Categorical Reparameterization with Gumbel-Softmax. ICLR (2017)\n\nYoshua Bengio, Nicholas Leonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.\n\nYichuan Tang, Ruslan Salakhutdinov. Learning Stochastic Feedforward Neural Networks. NIPS 2013\n\np.4 Proposition 2.1: This result seems very limited, since it assumes linear layers, and in the end a linear NN.\n\np.4 Corollary 2: What are alpha and beta here ? How do they relate to $f$, $e$ or $T$ from Definition 1 ? What are $e_i^l$ and $s_i^l$ here ?\n\np.4 §6: the modularity of chain components justifies transfer learning via partial reuse of pre-trained networks -> I do not understand this argument. What is meant here by \"modularity\" ?\n\np.4 §7: However, feed-forward is no longer applicable through these intra-connected layers. -> You finally give an example here that corresponds to non-trivial chain graphs (with chain components greater than 1), and recognize that your theoretical results do not apply any more. This contradicts your claim that CGs offer a general interpretation of NNs with theoretical support. I believe the CG interpretation is not justified, nor required to derive the results you present in this paper.\n\np.5 Definition 2: I fail to see the point of introducing this new concept, which is a re-definition of residual blocks.\n\np.5 §4: While a vanilla layered [...] -> I fail to understand where the CG interpretation fits in this argument.\n\np.5 Proposition 4: This result seems trivial to me, and does not require the concept of a CG. Once you show that a linear layer followed by a specific activation can be interpreted as a parametric model for $p(\\textit{out}|\\textit{in})$, then any NN that unrolls as an acyclic directed graph can be interpreted as a probabilistic model.\n\np.6 §1: The simple recurrent layer, [...] -> Again, I do not see where the CG interpretation gives any insight here.\n\np.6 Section 3.3: This result, again, does not require the CG interpretation. It is also already know. See, e.g., Pierre Baldi and Peter J. Sadowski. Understanding Dropout. NIPS 2013\n\np.6 Section 4: It seems to me you are reinventing SFFNs. See, Yichuan Tang and Ruslan Salakhutdinov. Stochastic Feedforward Neural Network. NIPS 2013\n\np.7 §2: the sampling operation is not differentiable -> This is not true. See, e.g.,  the reparameterization trick for VAEs. You mention this in the next sentence...", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603736025664}], "openreview_url": "https://openreview.net/forum?id=thhdrl4IdMm", "arxiv_id": "2006.16856", "paper_pdf": "papers/thhdrl4IdMm.pdf", "paper_pdf_sha256": "ba1f08c018193d5b9d64cdb2ec70bf73ac92238ca133ea054367efdea9bb44ae", "paper_pdf_bytes": 591560, "paper_pdf_source": "openreview", "code_url": "https://github.com/tum-vision/nnascg", "code_repository": "tum-vision/nnascg", "code_commit": "9eecd523d2a4d327144a33b3cadd5e3d5de4b734", "code_archive": "repos/thhdrl4IdMm.zip", "code_archive_sha256": "48fde91eb584b40b7e2a582ba3b3e42c0e2958c6fcff9e58d29b2f0172db5e96", "code_archive_bytes": 40424, "code_file_count": 11, "code_extensions": {".py": 9, ".sh": 2}, "github_disk_usage_kb": 34, "github_languages": {"Python": 79755, "Shell": 2616}, "github_archived": false, "github_pushed_at": "2020-07-10T15:41:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deriving-neural-network-design-and-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rylxpA4YwH", "year": 2020, "status": "rejected", "title": "On the Evaluation of Conditional GANs", "authors": ["Terrance DeVries", "Adriana Romero", "Luis Pineda", "Graham W. Taylor", "Michal Drozdzal"], "authorids": ["terrance@uoguelph.ca", "adrianars@fb.com", "lep@fb.com", "gwtaylor@uoguelph.ca", "mdrozdzal@fb.com"], "authors_source": "OpenReview API", "abstract": "Conditional Generative Adversarial Networks (cGANs) are finding increasingly widespread use in many application domains. Despite outstanding progress, quantitative evaluation of such models often involves multiple distinct metrics to assess different desirable properties, such as image quality, conditional consistency, and intra-conditioning diversity. In this setting, model benchmarking becomes a challenge, as each metric may indicate a different \"best\" model. In this paper, we propose the Frechet Joint Distance (FJD), which is defined as the Frechet distance between joint distributions of images and conditioning, allowing it to implicitly capture the aforementioned properties in a single metric. We conduct proof-of-concept experiments on a controllable synthetic dataset, which consistently highlight the benefits of FJD when compared to currently established metrics. Moreover, we use the newly introduced metric to compare existing cGAN-based models for a variety of conditioning modalities (e.g. class labels, object masks, bounding boxes, images, and text captions). We show that FJD can be used as a promising single metric for model benchmarking.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJxA3B9S5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1381/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a variant of the use of Frechet Inception Distance (FID) for the evaluation and benchmarking of conditional GAN models. FID is a popular measure for comparing image distributions in the Inception v3 feature space, in terms of the means and variances of multivariate Gaussians fit to data samples from each distribution. The authors argue that FID is ill-suited for use with cGANs, in that they do not explicitly take into account conditional consistency or intra-conditioning diversity. The main contribution and basic idea of this paper is to create joint image-conditioning distributions from the image embedding and the conditioning embedding in the Inception embedding, and then to combine them (by default, through vector concatenation). These joint image-conditioning distributions are then fed to FID as per standard usage on image distributions alone. The authors refer to their approach as FJD (Frechet Joint Distance).\n\nAlthough the authors propose FJD as a new technique, it should more properly be regarded as the direct use of FID on joint distributions. The main practical contribution of the paper thus reduces to the notion of concatenating the learned conditioning representation with the image representation. As a research contribution, this is in itself not very substantial. However, in their experimentation the authors do take care to show through examples the effect of their joint image-conditioning approach in assessing image quality, conditional consistency, and intra-conditioning diversity, under a variety of conditionings.\n\nThere are some issues that are not adequately addressed:\n\n1) In all experimental cases FJD scores and FID scores correlate well, which undercuts the argument that FJD is superior to FID in assessing the performance of cGAN models. Can situations be experimentally demonstrated where that is not the case? In particular, what happens when the fit in image representation is dramatically better / worse than that of the conditioning representation? This situation is interesting, but not considered in this paper.\n\n2) The effect of the dimensionality of the learned representations is not addressed. When concatenating vectors to produce a joint image-conditioning representation, the dimensionality increases, which would tend to produce larger FJD values than their corresponding FID values. The experimental results of this paper seem to confirm this. However, is\nit really valid then to declare FJD as being somehow more sensitive to the conditioning simply by virtue of obtaining larger values and larger spreads than FID? It should be remembered that FID and FJD are *not* unitless measures.\n\nOverall, in its current state the paper appears to be below the acceptance threshold.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper proposes a variant of the use of Frechet Inception Distance (FID) for the evaluation and benchmarking of conditional GAN models. FID is a popular measure for comparing image distributions in the Inception v3 feature space, in terms of the means and variances of multivariate Gaussians fit to data samples from each distribution. The authors argue that FID is ill-suited for use with cGANs, in that they do not explicitly take into account conditional consistency or intra-conditioning diversity. The main contribution and basic idea of this paper is to create joint image-conditioning distributions from the image embedding and the conditioning embedding in the Inception embedding, and then to combine them (by default, through vector concatenation). These joint image-conditioning distributions are then fed to FID as per standard usage on image distributions alone. The authors refer to their approach as FJD (Frechet Joint Distance).\n\nAlthough the authors propose FJD as a new technique, it should more properly be regarded as the direct use of FID on joint distributions. The main practical contribution of the paper thus reduces to the notion of concatenating the learned conditioning representation with the image representation. As a research contribution, this is in itself not very substantial. However, in their experimentation the authors do take care to show through examples the effect of their joint image-conditioning approach in assessing image quality, conditional consistency, and intra-conditioning diversity, under a variety of conditionings.\n\nThere are some issues that are not adequately addressed:\n\n1) In all experimental cases FJD scores and FID scores correlate well, which undercuts the argument that FJD is superior to FID in assessing the performance of cGAN models. Can situations be experimentally demonstrated where that is not the case? In particular, what happens when the fit in image representation is dramatically better / worse than that of the conditioning representation? This situation is interesting, but not considered in this paper.\n\n2) The effect of the dimensionality of the learned representations is not addressed. When concatenating vectors to produce a joint image-conditioning representation, the dimensionality increases, which would tend to produce larger FJD values than their corresponding FID values. The experimental results of this paper seem to confirm this. However, is\nit really valid then to declare FJD as being somehow more sensitive to the conditioning simply by virtue of obtaining larger values and larger spreads than FID? It should be remembered that FID and FJD are *not* unitless measures.\n\nOverall, in its current state the paper appears to be below the acceptance threshold.\n"}, "tcdate": 1572345269889}, {"id": "HJgKpMenKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1381/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\nThis paper mention that there are some critical drawbacks existing in IS (Inception Score) and FID (Fréchet Inception Distance) which are two popular metrics to measure image generation quality. However, IS and FID scores are initially designed for measuring unconditional distribution, which fails to capture the conditional consistency of conditional distribution. Thus, the authors propose to concatenate conditioned embedding h(y) with image feature vector f(x) to extend the FID metric. The authors also implement the method on a toy dataset to show the sensitivity of FJD on conditional consistency and several popular cGAN models to show the efficiency of FJD on real data.\nPaper Strengths\n  1. The method is intuitive and easy to implement.\n\nPaper Weaknesses\n1. Although this paper shows the problem of FID for capturing the conditional consistency sprightly with the toy dataset, however, this problem does not obviously show up on real data. Basically, FID can also give a good comparison of the different model as FJD", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review": "Summary\nThis paper mention that there are some critical drawbacks existing in IS (Inception Score) and FID (Fréchet Inception Distance) which are two popular metrics to measure image generation quality. However, IS and FID scores are initially designed for measuring unconditional distribution, which fails to capture the conditional consistency of conditional distribution. Thus, the authors propose to concatenate conditioned embedding h(y) with image feature vector f(x) to extend the FID metric. The authors also implement the method on a toy dataset to show the sensitivity of FJD on conditional consistency and several popular cGAN models to show the efficiency of FJD on real data.\nPaper Strengths\n  1. The method is intuitive and easy to implement.\n\nPaper Weaknesses\n1. Although this paper shows the problem of FID for capturing the conditional consistency sprightly with the toy dataset, however, this problem does not obviously show up on real data. Basically, FID can also give a good comparison of the different model as FJD"}, "tcdate": 1571713729151}, {"id": "BJxo_VsaOH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1381/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper extends the Fréchet Inception distance (FID) to the conditional distribution. To this end, the authors use an additional embedding for the condition variables (class, image, text), and concatenate to the data embedding. The proposed metric, name the Fréchet joint distance (FJD), captures three desired properties of conditional generative models: sample quality, conditional consistency, and sample diversity. The authors demonstrate that the proposed metric indeed captures the properties using a synthetic (dSprite) dataset, and shows reasonable values for real datasets.\n\nPros:\n- FJD is an intuitive extension of FID for conditional generative models.\n- FJD can be applied to various types of conditions (e.g., image and text), which cannot be done by prior work (e.g., [1]).\n- The paper is easy to read and experimental details are clearly stated.\n\nCons:\n\n1. FJD is a straightforward extension of FID.\n\nFJD simply follows the FID formula but concatenates the condition embedding to the original data embedding. It is a straightforward extension of FID and suffers from the design choice problems due to the concatenation, as stated below.\n\n2. FJD requires many design choices and not theoretically justified.\n\nAs FJD requires an additional embedding function h, balancing parameter \\alpha, and merging function g, it raises a burden of design choices. While the authors give some suggestions, they are not theoretically justified. Also, one may use the statistical distances [2] between data distribution p_data(x,c) and model distribution p_g(x,c) to evaluate conditional generative models in a principled way, e.g., measure the KL-divergence using the density ratios [3]. The advantage of FJD over such metrics is unclear, as stated below.\n\n3. The advantage over the prior work is not clear.\n\nFJD and FID show the same trend in all reported experiments (Table 2, 3, 4), hence the advantage of FJD is unclear. Also, one may measure the FID score on conditional distributions, i.e., \\sum_c FID( p_data(x|c), p_g(x|c) ). It also captures the desired three properties and would be a strong baseline for FJD. Besides, while the authors aim to design a single metric to stand the models in a line, identifying the trade-offs of models may also be useful. For example, Improved PRD [4] provides the precision-recall trade-offs of generative models, which provides some insights for the models.\n\n\n[1] Ravuri and Vinyals. Classification Accuracy Score for Conditional Generative Models. NeurIPS 2019.\n[2] https://en.wikipedia.org/wiki/Statistical_distance\n[3] Uehara et al. Generative Adversarial Nets from a Density Ratio Estimation Perspective. arXiv 2016.\n[4] Kynkäänniemi et al. Improved Precision and Recall Metric for Assessing Generative Models. NeurIPS 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Summary:\nThis paper extends the Fréchet Inception distance (FID) to the conditional distribution. To this end, the authors use an additional embedding for the condition variables (class, image, text), and concatenate to the data embedding. The proposed metric, name the Fréchet joint distance (FJD), captures three desired properties of conditional generative models: sample quality, conditional consistency, and sample diversity. The authors demonstrate that the proposed metric indeed captures the properties using a synthetic (dSprite) dataset, and shows reasonable values for real datasets.\n\nPros:\n- FJD is an intuitive extension of FID for conditional generative models.\n- FJD can be applied to various types of conditions (e.g., image and text), which cannot be done by prior work (e.g., [1]).\n- The paper is easy to read and experimental details are clearly stated.\n\nCons:\n\n1. FJD is a straightforward extension of FID.\n\nFJD simply follows the FID formula but concatenates the condition embedding to the original data embedding. It is a straightforward extension of FID and suffers from the design choice problems due to the concatenation, as stated below.\n\n2. FJD requires many design choices and not theoretically justified.\n\nAs FJD requires an additional embedding function h, balancing parameter \\alpha, and merging function g, it raises a burden of design choices. While the authors give some suggestions, they are not theoretically justified. Also, one may use the statistical distances [2] between data distribution p_data(x,c) and model distribution p_g(x,c) to evaluate conditional generative models in a principled way, e.g., measure the KL-divergence using the density ratios [3]. The advantage of FJD over such metrics is unclear, as stated below.\n\n3. The advantage over the prior work is not clear.\n\nFJD and FID show the same trend in all reported experiments (Table 2, 3, 4), hence the advantage of FJD is unclear. Also, one may measure the FID score on conditional distributions, i.e., \\sum_c FID( p_data(x|c), p_g(x|c) ). It also captures the desired three properties and would be a strong baseline for FJD. Besides, while the authors aim to design a single metric to stand the models in a line, identifying the trade-offs of models may also be useful. For example, Improved PRD [4] provides the precision-recall trade-offs of generative models, which provides some insights for the models.\n\n\n[1] Ravuri and Vinyals. Classification Accuracy Score for Conditional Generative Models. NeurIPS 2019.\n[2] https://en.wikipedia.org/wiki/Statistical_distance\n[3] Uehara et al. Generative Adversarial Nets from a Density Ratio Estimation Perspective. arXiv 2016.\n[4] Kynkäänniemi et al. Improved Precision and Recall Metric for Assessing Generative Models. NeurIPS 2019."}, "tcdate": 1570776179155}], "openreview_url": "https://openreview.net/forum?id=rylxpA4YwH", "arxiv_id": "1907.08175", "paper_pdf": "papers/rylxpA4YwH.pdf", "paper_pdf_sha256": "86cc696a9e3fdf67be1d6111386c0312321a2852cf24b1d8ddf2436087972998", "paper_pdf_bytes": 8542957, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/fjd", "code_repository": "facebookresearch/fjd", "code_commit": "6e48d77162b16a20ebaea832470cfa25eb2021ef", "code_archive": "repos/rylxpA4YwH.zip", "code_archive_sha256": "2d0a735da3cd716997d21f6ac6b4202f2939c9d270cbbb95fb05c003ea9d8fe6", "code_archive_bytes": 46694, "code_file_count": 11, "code_extensions": {".py": 10, ".ipynb": 1}, "github_disk_usage_kb": 42, "github_languages": {"Python": 78642, "Jupyter Notebook": 27554}, "github_archived": true, "github_pushed_at": "2019-11-27T17:57:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-evaluation-of-conditional-gans"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkMWx309FX", "year": 2019, "status": "rejected", "title": "Reinforcement Learning with Perturbed Rewards", "authors": ["Jingkang Wang", "Yang Liu", "Bo Li"], "authorids": ["wangjksjtu_01@sjtu.edu.cn", "yangliu@ucsc.edu", "lxbosky@gmail.com"], "authors_source": "OpenReview API", "abstract": "Recent studies have shown the vulnerability of reinforcement learning (RL) models in noisy settings. The sources of noises differ across scenarios. For instance, in practice, the observed reward channel is often subject to noise (e.g., when observed rewards are collected through sensors), and thus observed rewards may not be credible as a result. Also, in applications such as robotics, a deep reinforcement learning (DRL) algorithm can be manipulated to produce arbitrary errors. In this paper, we consider noisy RL problems where observed rewards by RL agents are generated with a reward confusion matrix. We call such observed rewards as perturbed rewards. We develop an unbiased reward estimator aided robust RL framework that enables RL agents to learn in noisy environments while observing only perturbed rewards. Our framework draws upon approaches for supervised learning with noisy data. The core ideas of our solution include estimating a reward confusion matrix and defining a set of unbiased surrogate rewards. We prove the convergence and sample complexity of our approach. Extensive experiments on different DRL platforms show that policies based on our estimated surrogate reward can achieve higher expected rewards, and converge faster than existing baselines. For instance, the state-of-the-art PPO algorithm is able to obtain 67.5% and 46.7% improvements in average on five Atari games, when the error rates are 10% and 30% respectively. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1xorCQWp7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1050/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper investigates reinforcement learning with a perturbed reward signal. In particular, the paper proposes a particular model for adding noise to the reward function via a confusion matrix, which offers a nuanced notion of reward-noise that is not too complicated so-as to make learning impossible. I take this learning setting to be both novel and interesting for opening up areas for future work. The central contributions of the work are to 1) leverage a simple estimator to prove the convergence of Q-Learning under the reward-perturbed setting along with the sample-complexity of a variant of (Phased) Q-Learning which they call \"Phrased\" Q-Learning, and 2) An algorithmic scheme for learning in the reward-perturbed setting (Algorithm 1), and 3) An expansive set of experiments that explore the impact of various reward models on learning across different environment-algorithm combinations. The sample complexity term extends Phased Q-Learning to incorporate aspects of the reward confusion matrix, and to my knowledge is novel. Further, even though Theorem 1 is unsurprising (as the paper suggests), I take the collection of Theorem 1, 2, and 3 to be collectively novel.\n\nIndeed, the paper focuses on an interesting and relatively unexplored direction for RL. Apart from the work cited by the paper (and perhaps work like Krueger et al. (2016), in which agents must pay some cost to observe true rewards), there is little work on learning settings of this kind. This paper represents a first step in gaining clarity on how to formalize and study this problem. I did, however, find the analysis and the experiments to be relatively disjointed -- the main sample complexity result presented by the paper (Theorem 2) was given for Phased Q-Learning, yet no experiments actually evaluate the performance of Phased Q-Learning. I think the paper could benefit from experiments focused on simple domains that showcase how traditional algorithms do in cases where it is easier to understand (and visualize) the impact of the reward perturbations (simple chain MDPs, grid worlds, etc.) -- and specifically experiments including Phased Q-Learning. \n\nPros:\n\t- General, interesting new learning setting to study.\n\t- Initial convergence and sample complexity results for this new setting.\n\t- Depth and breadth of experimentation (in terms of diversity of algorithms and environments), includes lots of detail about the experimental setup.\n\nCons:\n\t- Clarity of writing: lots of typos and bits of math that could be more clear (see detailed comments below) [Fixed]\n\t- The plots in Section 4 are all extremely jagged. More trials seem to be required. Moreover, I do think simpler domains might help offer insights into the reward perturbed setting. [Fixed]\n\t- The reward perturbation model is relatively simple.\n\nSome high level questions/comments:\n\t- Why was Phrased Q-Learning not experimented with?\n\t- Why use majority voting as the rule? When this was introduced it sounded like any rule might be used. Have you tried/thought about others?\n\t- Your citation to Kakade's thesis needs fixing; it should read:\n\t\t\"Kakade, Sham Machandranath. On the sample complexity of reinforcement learning. Ph.D Thesis. University of London, 2003.\"\n\n\t\t(right now it is cited as \"(Gatsby 2003)\" throughout the paper)\n\t- You might consider picking a new name for Phrased Q-Learning -- right now the name is too similar to Phased Q-Learning from [Kearns and Singh NIPS 1999].\n        - As mentioned in the \"cons\" section, the confusion matrix is still a somewhat simple model of reward noise. I was left wondering: what might be the next most complicated form of adding reward noise? How might the proposed algorithm(s) respond to this slightly more complex model? That is, it's unclear how general the results are, or if they are honed too tightly to the specific proposed reward noise model. I was hoping the authors could respond to this point.\n\n\t\nSection 0) Abstract:\n\t- Not immediately clear what is meant by \"vulnerability\" or \"noisy settings\". Might be better to pick a more clear initial sentence (same can be said of the \"sources of noise...\"\")\n\nSection 1) Introduction:\n\t- \"adversaries in real-world\" --> \"adversaries in the real-world\"\n\t- You might consider citing Loftin et al. (2014) regarding the bulleted point about \"Application-Specific Noise\".\n\t- \"unbiased reward estimator aided reward robust reinforcement learning framework\" --> this was a bit hard to parse. Consider making more concise, like: \"unbiased reward estimator for use in reinforcement learning with perturbed rewards\".\n\t- \"Our solution framework builds on existing reinforcement learning algorithms, including the recently developed DRL ones\" --> cite these up front So, cite: Q-Learning, CEM, SARSA, DQN, Dueling DQN, DDPG, NAF, and PPO, and spell out the acronym for each the first time you introduce them.\n\t- \"layer of explorations\" --> \"layer of exploration\"\n\nSection 2) Problem Formulation\n\t- \"as each shot of our\" --> what is 'shot' in this context?\n\t- \"In what follow,\" --> \"In what follows,\"\n\t- \"where 0 < \\gamma \\leq 1\" --> Usually, $\\gamma \\in [0,1)$, or $[0,1]$. Why can't $\\gamma = 0$?\n\t- The transition notation changes between $\\mathbb{P}_a(s_{t+1} | s_t)$ and $\\mathbb{P}(s_{t+1} | s_t, a_t)$. I'd suggest picking one and sticking with it to improve clarity.\n\t- \"to learn a state-action value function, for example the Q-function\" --> Why is the Q-function just an example? Isn't is *the* state-action value function? That is, I'd suggest replacing \"to learn a state-action value function, for example the Q-function\" with \"to learn a state-action value function, also called the Q-function\"\n\t- \"Q-function calculates\" --> \"The Q-function denotes\"\n\t- \"the reward feedbacks perfectly\" --> \"the reward feedback perfectly\"\n\t- I prefer that the exposition of the perturbed reward MDP be done with C in the tuple. So: $\\tilde{M} = \\langle \\mathcal{S}, \\mathcal{A}, \\mathcal{R}, C, \\mathcal{P}, \\gamma \\rangle$. This seems the most appropriate definition, since the observed rewards will be generated by $C$.\n\t- The setup of the confusion matrix for reward noise over is very clean. It might be worth pointing out that $C$ need not be Markovian. There are cases where C is not just a function of $\\mathcal{S}$ and $\\mathcal{R}$, like the adversarial case you describe early on.\n\n\nSection 3) Learning w/ Perturbed Rewards\n\t- Theorem 1 builds straightforwardly on Q-Learning convergence guarantee (it might be worth phrasing the result in those terms? That is: the addition of the perturbed reward does not destroy the convergence guarantees of Q-Learning.)\n\t- \"we firstly\" --> \"we first\"\n\t- \"value iteration (using Q function)\" --> \"value iteration\"\n\t- \"Definition 2. Phased Q-Learning\" --> \"Definition 2. Phrased Q-Learning\". I think? Unless you're talking about Phased Q from the Kearns and Singh '99 work.\n\t- \"It uses collected m samples\" --> \"It uses the collected m samples\"\n\t- Theorem 2: it would be helpful to define $T$ since it appears in the sample complexity term. Also, I would suggest specifying the domain of $\\epsilon$, as you do with $\\delta$.\n\t- \"convergence to optimal policy\" --> \"convergence to the optimal policy\"\n\t- \"The idea of constructing MDP is similar to\" --> this seems out of place. The idea of constructing which MDP? Similar to Kakade (2003) in what sense?\n\t- \"the unbiasedness\" --> \"the use of unbiased estimators\"\n\t- \"number of state-action pair, which satisfies\" --> \"number of state-action pairs that satisfy\"\n\t- \"The above procedure continues with more observations arriving.\" --> \"The above procedure continues indefinitely as more observation arrives.\" Also, which procedure? Updating $\\tilde{c}_{i,j}$? If so, I would specify.\n\t- \"is nothing different from Eqn. (2) but with replacing a known reward confusion\" --> \"replaces a known reward confusion\"\n\n\n4) Experiments:\n\t- Diverse experiments! That's great. Lots of algorithms, lots of environment types.\n\t- I expected to see Phrased Q-Learning in the experiments. Why was it not included?\n\t- The plots are pretty jagged, so I'm left feeling a bit skeptical about some of the results. The results would be strengthened if the experiments were repeated for more trials.\n\n5) Conclusion:\n\t- \"despite of the fact\" --> \"despite the fact\"\n\t- \"finite sample complexity of Q-Learning with estimated surrogate rewards are given\" --> It's not really Q-Learning, though. It's a variant of Q-Learning. I'd suggest being explicit about that.\n\nAppendix:\n\n\t- \"It is easy to validate the unbiasedness of proposed estimator directly.\" --> \"It is easy to verify that the proposed estimator is unbiased directly.\"\n\t- \"For the simplicity of notations\" --> \"For simplicity\"\n\t- \"the Phrased Q-Learning could converge to near optimal policy\" --> \"\"the algorithm Phrased Q-Learning can converge to the near optimal policy\"\"\n\t- \"Using union bound\" --> \"Using a union bound\"\n\t- Same comment regarding $\\gamma$: it's typically $0 \\leq \\gamma < 1$.\n\t- Bottom of page 16, the second equation from the bottom, far right term: $c.j$ --> $c,j$.\n\t- \"Using CauchySchwarz Inequality\" --> \"Using the Cauchy-Schwarz Inequality\"\n\n\nReferences:\n\tLoftin, Robert, et al. \"Learning something from nothing: Leveraging implicit human feedback strategies.\" Robot and Human Interactive Communication, 2014 RO-MAN: The 23rd IEEE International Symposium on. IEEE, 2014.\n\n\tKrueger, D., Leike, J., Evans, O., & Salvatier, J. (2016). Active reinforcement learning: Observing rewards at a cost. In Future of Interactive Learning Machines, NIPS Workshop.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting and relatively unexplored variant of RL.", "review": "\nThis paper investigates reinforcement learning with a perturbed reward signal. In particular, the paper proposes a particular model for adding noise to the reward function via a confusion matrix, which offers a nuanced notion of reward-noise that is not too complicated so-as to make learning impossible. I take this learning setting to be both novel and interesting for opening up areas for future work. The central contributions of the work are to 1) leverage a simple estimator to prove the convergence of Q-Learning under the reward-perturbed setting along with the sample-complexity of a variant of (Phased) Q-Learning which they call \"Phrased\" Q-Learning, and 2) An algorithmic scheme for learning in the reward-perturbed setting (Algorithm 1), and 3) An expansive set of experiments that explore the impact of various reward models on learning across different environment-algorithm combinations. The sample complexity term extends Phased Q-Learning to incorporate aspects of the reward confusion matrix, and to my knowledge is novel. Further, even though Theorem 1 is unsurprising (as the paper suggests), I take the collection of Theorem 1, 2, and 3 to be collectively novel.\n\nIndeed, the paper focuses on an interesting and relatively unexplored direction for RL. Apart from the work cited by the paper (and perhaps work like Krueger et al. (2016), in which agents must pay some cost to observe true rewards), there is little work on learning settings of this kind. This paper represents a first step in gaining clarity on how to formalize and study this problem. I did, however, find the analysis and the experiments to be relatively disjointed -- the main sample complexity result presented by the paper (Theorem 2) was given for Phased Q-Learning, yet no experiments actually evaluate the performance of Phased Q-Learning. I think the paper could benefit from experiments focused on simple domains that showcase how traditional algorithms do in cases where it is easier to understand (and visualize) the impact of the reward perturbations (simple chain MDPs, grid worlds, etc.) -- and specifically experiments including Phased Q-Learning. \n\nPros:\n\t- General, interesting new learning setting to study.\n\t- Initial convergence and sample complexity results for this new setting.\n\t- Depth and breadth of experimentation (in terms of diversity of algorithms and environments), includes lots of detail about the experimental setup.\n\nCons:\n\t- Clarity of writing: lots of typos and bits of math that could be more clear (see detailed comments below) [Fixed]\n\t- The plots in Section 4 are all extremely jagged. More trials seem to be required. Moreover, I do think simpler domains might help offer insights into the reward perturbed setting. [Fixed]\n\t- The reward perturbation model is relatively simple.\n\nSome high level questions/comments:\n\t- Why was Phrased Q-Learning not experimented with?\n\t- Why use majority voting as the rule? When this was introduced it sounded like any rule might be used. Have you tried/thought about others?\n\t- Your citation to Kakade's thesis needs fixing; it should read:\n\t\t\"Kakade, Sham Machandranath. On the sample complexity of reinforcement learning. Ph.D Thesis. University of London, 2003.\"\n\n\t\t(right now it is cited as \"(Gatsby 2003)\" throughout the paper)\n\t- You might consider picking a new name for Phrased Q-Learning -- right now the name is too similar to Phased Q-Learning from [Kearns and Singh NIPS 1999].\n        - As mentioned in the \"cons\" section, the confusion matrix is still a somewhat simple model of reward noise. I was left wondering: what might be the next most complicated form of adding reward noise? How might the proposed algorithm(s) respond to this slightly more complex model? That is, it's unclear how general the results are, or if they are honed too tightly to the specific proposed reward noise model. I was hoping the authors could respond to this point.\n\n\t\nSection 0) Abstract:\n\t- Not immediately clear what is meant by \"vulnerability\" or \"noisy settings\". Might be better to pick a more clear initial sentence (same can be said of the \"sources of noise...\"\")\n\nSection 1) Introduction:\n\t- \"adversaries in real-world\" --> \"adversaries in the real-world\"\n\t- You might consider citing Loftin et al. (2014) regarding the bulleted point about \"Application-Specific Noise\".\n\t- \"unbiased reward estimator aided reward robust reinforcement learning framework\" --> this was a bit hard to parse. Consider making more concise, like: \"unbiased reward estimator for use in reinforcement learning with perturbed rewards\".\n\t- \"Our solution framework builds on existing reinforcement learning algorithms, including the recently developed DRL ones\" --> cite these up front So, cite: Q-Learning, CEM, SARSA, DQN, Dueling DQN, DDPG, NAF, and PPO, and spell out the acronym for each the first time you introduce them.\n\t- \"layer of explorations\" --> \"layer of exploration\"\n\nSection 2) Problem Formulation\n\t- \"as each shot of our\" --> what is 'shot' in this context?\n\t- \"In what follow,\" --> \"In what follows,\"\n\t- \"where 0 < \\gamma \\leq 1\" --> Usually, $\\gamma \\in [0,1)$, or $[0,1]$. Why can't $\\gamma = 0$?\n\t- The transition notation changes between $\\mathbb{P}_a(s_{t+1} | s_t)$ and $\\mathbb{P}(s_{t+1} | s_t, a_t)$. I'd suggest picking one and sticking with it to improve clarity.\n\t- \"to learn a state-action value function, for example the Q-function\" --> Why is the Q-function just an example? Isn't is *the* state-action value function? That is, I'd suggest replacing \"to learn a state-action value function, for example the Q-function\" with \"to learn a state-action value function, also called the Q-function\"\n\t- \"Q-function calculates\" --> \"The Q-function denotes\"\n\t- \"the reward feedbacks perfectly\" --> \"the reward feedback perfectly\"\n\t- I prefer that the exposition of the perturbed reward MDP be done with C in the tuple. So: $\\tilde{M} = \\langle \\mathcal{S}, \\mathcal{A}, \\mathcal{R}, C, \\mathcal{P}, \\gamma \\rangle$. This seems the most appropriate definition, since the observed rewards will be generated by $C$.\n\t- The setup of the confusion matrix for reward noise over is very clean. It might be worth pointing out that $C$ need not be Markovian. There are cases where C is not just a function of $\\mathcal{S}$ and $\\mathcal{R}$, like the adversarial case you describe early on.\n\n\nSection 3) Learning w/ Perturbed Rewards\n\t- Theorem 1 builds straightforwardly on Q-Learning convergence guarantee (it might be worth phrasing the result in those terms? That is: the addition of the perturbed reward does not destroy the convergence guarantees of Q-Learning.)\n\t- \"we firstly\" --> \"we first\"\n\t- \"value iteration (using Q function)\" --> \"value iteration\"\n\t- \"Definition 2. Phased Q-Learning\" --> \"Definition 2. Phrased Q-Learning\". I think? Unless you're talking about Phased Q from the Kearns and Singh '99 work.\n\t- \"It uses collected m samples\" --> \"It uses the collected m samples\"\n\t- Theorem 2: it would be helpful to define $T$ since it appears in the sample complexity term. Also, I would suggest specifying the domain of $\\epsilon$, as you do with $\\delta$.\n\t- \"convergence to optimal policy\" --> \"convergence to the optimal policy\"\n\t- \"The idea of constructing MDP is similar to\" --> this seems out of place. The idea of constructing which MDP? Similar to Kakade (2003) in what sense?\n\t- \"the unbiasedness\" --> \"the use of unbiased estimators\"\n\t- \"number of state-action pair, which satisfies\" --> \"number of state-action pairs that satisfy\"\n\t- \"The above procedure continues with more observations arriving.\" --> \"The above procedure continues indefinitely as more observation arrives.\" Also, which procedure? Updating $\\tilde{c}_{i,j}$? If so, I would specify.\n\t- \"is nothing different from Eqn. (2) but with replacing a known reward confusion\" --> \"replaces a known reward confusion\"\n\n\n4) Experiments:\n\t- Diverse experiments! That's great. Lots of algorithms, lots of environment types.\n\t- I expected to see Phrased Q-Learning in the experiments. Why was it not included?\n\t- The plots are pretty jagged, so I'm left feeling a bit skeptical about some of the results. The results would be strengthened if the experiments were repeated for more trials.\n\n5) Conclusion:\n\t- \"despite of the fact\" --> \"despite the fact\"\n\t- \"finite sample complexity of Q-Learning with estimated surrogate rewards are given\" --> It's not really Q-Learning, though. It's a variant of Q-Learning. I'd suggest being explicit about that.\n\nAppendix:\n\n\t- \"It is easy to validate the unbiasedness of proposed estimator directly.\" --> \"It is easy to verify that the proposed estimator is unbiased directly.\"\n\t- \"For the simplicity of notations\" --> \"For simplicity\"\n\t- \"the Phrased Q-Learning could converge to near optimal policy\" --> \"\"the algorithm Phrased Q-Learning can converge to the near optimal policy\"\"\n\t- \"Using union bound\" --> \"Using a union bound\"\n\t- Same comment regarding $\\gamma$: it's typically $0 \\leq \\gamma < 1$.\n\t- Bottom of page 16, the second equation from the bottom, far right term: $c.j$ --> $c,j$.\n\t- \"Using CauchySchwarz Inequality\" --> \"Using the Cauchy-Schwarz Inequality\"\n\n\nReferences:\n\tLoftin, Robert, et al. \"Learning something from nothing: Leveraging implicit human feedback strategies.\" Robot and Human Interactive Communication, 2014 RO-MAN: The 23rd IEEE International Symposium on. IEEE, 2014.\n\n\tKrueger, D., Leike, J., Evans, O., & Salvatier, J. (2016). Active reinforcement learning: Observing rewards at a cost. In Future of Interactive Learning Machines, NIPS Workshop.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541647938812}, {"id": "BJliWdLsh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1050/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Summary\n\nThe authors present work that shows how to deal with noise in reward signals by creating a surrogate reward signal. The work develops a number of results including: showing how the surrogate reward is equal in expectation to the true reward signal, how this doesn't affect the fixed point of the Bellman equation, how to deal with finite and continuous rewards and how the convergence time is affected for different levels of noise. They demonstrate the value of this approach with a variety of early and state-of-the-art algorithms on a variety of domains,, and the results are consistent with the claims.\n\nIt would be useful to outline how prior work approached this same problem and also to evaluate the proposed method with existin approaches to the same problem. I realise that this is the first method that estimates the confusion matrix rather than assuming it is known a priori but there are obvious ways around this, e.g. the authors first experiment assumes the confusion matrix is known, so this would be a good place to compare with other competing techniques. Also, the authors have a way of estimating this, so they could plug it into the other algorithms too.\n\nI also have some concerns about the clarity and precision of the proofs, although I do not have any reason to doubt the Lemma/Theorem correctness (see below).\n\nThe weakest part of the approach is in how the true reward is estimated in order to estiamate the confusion matrix. It uses majority vote (which is only really possible in the case of finite rewards with noise sufficiently low that this will be a robust estimate). Perhaps some other approaches could be explore here too.\n\nFinally, there is discussion about adversarial noise in rewards at the beginning but I am not sure the theory really addresses it nor the evaluations.\n\nNonetheless, given that I do not know whether the claim of originality is true (in terms of the estimation of the confusion matrix). If it is, then the work is a significant and interesting advance, and is clearly widely applicable in domains with noisy rewards. It would be interesting to see a more tractable approach for continous noise too, but this would probably involve assumptions (smoothness? Gaussianity?), and doesn't impact the value of this work.\n\n## Detailed notes\n\nThere is a slight sloppiness in  notation in equation (1). This uses \\tilde{r} as a subscript of e, but r is +1 or -1 and the error variables are e_+ and e_- (not e_{+1} and e_{-1}).\n\n\nThe noise levels in Atari (Figure 3) show something quite interesting which could be commented upon. For noise below 0.5 the surrogate reward works roughly  similarly to the noisy reward, but when the noise level goes above this, the surrogate reward clearly exploits the increased information content (similar to a noisy binary channel with over 0.5 noise). This may have  implications for adversarial noise.\n\nThere are also some issues with the proofs which I spotted outlined below:\n\n### Lemma 1 proof\nThe proof of Lemma 1, I think, fails to achieve its objective. The first pair of equations is not a rewrite of equation (1). I believe that the authors intend for this to be a consequence of Equation (1) but do not really demonstrate this clearly. Also, the authors seem to switch between binary rewards -1 and +1 and two levels of reward r- and r+ leading to some confusion. I would suggest the latter throughout as it is more general but involves no more terms.\n\nI suggest the following as an outline for the proof. It would help for them to define what they mean by the different rhats (as they currently do) and explain that these values are therefore:\n\n  rhat- = [(1 - e+) r- - e- r+ ]/(1 - e+ - e-)\n  rhat+ = [(1 - e-) r+ - e+ r-]/(1- e+ - e-)\n\nfrom equation (1). What is left is for them to actually prove the Lemma, namely that the expected value of rhat is:\n\n  E(rhat ) = p1(rhat=rhat-) rhat- + p(rhat=rhat+) rhat+ = E(r)\n\nwhere the probabilities relate to the surrogate reward taking their respective values. And just stylistically, I would avoid writing \"we could obtain\" and simply write \"we obtain\".\n\nLemma 2 achieves this more clearly with greater generality.\n\n\n### Theorem 1 proof\nAt the end of p13, the proof of the expected value loses track of the chosen action a. I would suggest the authors replace: $$\\mathbb{P}'(s,s',\\hat{r})$$ with $$\\mathbb{P}'(s,a, s',\\hat{r})$$ then define it. Likewise $$\\mathbb{P}(s,s')$$ should be $$\\mathbb{P}(s,a,s')$$ (and also defined).\n\nI am also a little uncomfortable with the switch from: $$max_{b \\in \\mathcal{A}} | Q(s',b) - Q*(s',b)|$$ in the second to last line of p13, which refers to the maximum Q value associated with some state s', to  $$||Q-Q*||_{\\infty}$$ in the next line which is the maximum over all states and actions. The equality should probably be an inequality there too.\n\nThroughout this the notation could be much better defined, including how to interpret the curly F and how it acts in the conditional part of an expectation and variance.\n\nFinally, there is a bit too free a use of the word \"easily\" here. If it were easy, then the authors could do it more clearly I think. Otherwise, please refer to the appropriate result in the literature.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting general surrogate reward which has wide applicability, and can be flexibly included alongside a variety of algorithms.", "review": "## Summary\n\nThe authors present work that shows how to deal with noise in reward signals by creating a surrogate reward signal. The work develops a number of results including: showing how the surrogate reward is equal in expectation to the true reward signal, how this doesn't affect the fixed point of the Bellman equation, how to deal with finite and continuous rewards and how the convergence time is affected for different levels of noise. They demonstrate the value of this approach with a variety of early and state-of-the-art algorithms on a variety of domains,, and the results are consistent with the claims.\n\nIt would be useful to outline how prior work approached this same problem and also to evaluate the proposed method with existin approaches to the same problem. I realise that this is the first method that estimates the confusion matrix rather than assuming it is known a priori but there are obvious ways around this, e.g. the authors first experiment assumes the confusion matrix is known, so this would be a good place to compare with other competing techniques. Also, the authors have a way of estimating this, so they could plug it into the other algorithms too.\n\nI also have some concerns about the clarity and precision of the proofs, although I do not have any reason to doubt the Lemma/Theorem correctness (see below).\n\nThe weakest part of the approach is in how the true reward is estimated in order to estiamate the confusion matrix. It uses majority vote (which is only really possible in the case of finite rewards with noise sufficiently low that this will be a robust estimate). Perhaps some other approaches could be explore here too.\n\nFinally, there is discussion about adversarial noise in rewards at the beginning but I am not sure the theory really addresses it nor the evaluations.\n\nNonetheless, given that I do not know whether the claim of originality is true (in terms of the estimation of the confusion matrix). If it is, then the work is a significant and interesting advance, and is clearly widely applicable in domains with noisy rewards. It would be interesting to see a more tractable approach for continous noise too, but this would probably involve assumptions (smoothness? Gaussianity?), and doesn't impact the value of this work.\n\n## Detailed notes\n\nThere is a slight sloppiness in  notation in equation (1). This uses \\tilde{r} as a subscript of e, but r is +1 or -1 and the error variables are e_+ and e_- (not e_{+1} and e_{-1}).\n\n\nThe noise levels in Atari (Figure 3) show something quite interesting which could be commented upon. For noise below 0.5 the surrogate reward works roughly  similarly to the noisy reward, but when the noise level goes above this, the surrogate reward clearly exploits the increased information content (similar to a noisy binary channel with over 0.5 noise). This may have  implications for adversarial noise.\n\nThere are also some issues with the proofs which I spotted outlined below:\n\n### Lemma 1 proof\nThe proof of Lemma 1, I think, fails to achieve its objective. The first pair of equations is not a rewrite of equation (1). I believe that the authors intend for this to be a consequence of Equation (1) but do not really demonstrate this clearly. Also, the authors seem to switch between binary rewards -1 and +1 and two levels of reward r- and r+ leading to some confusion. I would suggest the latter throughout as it is more general but involves no more terms.\n\nI suggest the following as an outline for the proof. It would help for them to define what they mean by the different rhats (as they currently do) and explain that these values are therefore:\n\n  rhat- = [(1 - e+) r- - e- r+ ]/(1 - e+ - e-)\n  rhat+ = [(1 - e-) r+ - e+ r-]/(1- e+ - e-)\n\nfrom equation (1). What is left is for them to actually prove the Lemma, namely that the expected value of rhat is:\n\n  E(rhat ) = p1(rhat=rhat-) rhat- + p(rhat=rhat+) rhat+ = E(r)\n\nwhere the probabilities relate to the surrogate reward taking their respective values. And just stylistically, I would avoid writing \"we could obtain\" and simply write \"we obtain\".\n\nLemma 2 achieves this more clearly with greater generality.\n\n\n### Theorem 1 proof\nAt the end of p13, the proof of the expected value loses track of the chosen action a. I would suggest the authors replace: $$\\mathbb{P}'(s,s',\\hat{r})$$ with $$\\mathbb{P}'(s,a, s',\\hat{r})$$ then define it. Likewise $$\\mathbb{P}(s,s')$$ should be $$\\mathbb{P}(s,a,s')$$ (and also defined).\n\nI am also a little uncomfortable with the switch from: $$max_{b \\in \\mathcal{A}} | Q(s',b) - Q*(s',b)|$$ in the second to last line of p13, which refers to the maximum Q value associated with some state s', to  $$||Q-Q*||_{\\infty}$$ in the next line which is the maximum over all states and actions. The equality should probably be an inequality there too.\n\nThroughout this the notation could be much better defined, including how to interpret the curly F and how it acts in the conditional part of an expectation and variance.\n\nFinally, there is a bit too free a use of the word \"easily\" here. If it were easy, then the authors could do it more clearly I think. Otherwise, please refer to the appropriate result in the literature.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541265410947}, {"id": "Bkl75YQ92Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1050/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper aims at studying the setting of perturbed rewards in a deep RL setting. Studying the effect of noise in the reward function is interesting. The paper is quite well-written. However the paper studies a rather simple setting, the limitations could be discussed more clearly and there are one or two elements unclear (see below).\n\nThe paper assumes first the interesting case where the generation of the perturbed reward is a function of S*R into the perturbed reward space. But then the confusion matrix does *not* take into account the state, which is justified by \"to let our presentation stay focused (...)\". I believe these elements should at least be clearly discussed. Indeed, in that setting, the theorems given seem to be variations of existing results and it is difficult to understand what is the message behind the theorems.\n\nIn addition, it is assumed that the confusion matrix C is known or estimated from data but it's not clear to me how this can be done in practice.  In equation 4, how do you have access to the predicted true rewards?\n\nAdditional comments:\n- The discount factor can be 0 but can not, in general, be equal to 1. So the equation in paragraph 2.1 \"0 < γ ≤ 1\" is wrong.\n- The paper mention that \"an underwhelming amount of reinforcement learning studies have focused on the settings with perturbed and noisy rewards\" but there are some works on the subject (e.g., https://arxiv.org/abs/1805.03359) and a discussion about the differences with the related work would be interesting.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting but seems to tackle a too narrow problem", "review": "The paper aims at studying the setting of perturbed rewards in a deep RL setting. Studying the effect of noise in the reward function is interesting. The paper is quite well-written. However the paper studies a rather simple setting, the limitations could be discussed more clearly and there are one or two elements unclear (see below).\n\nThe paper assumes first the interesting case where the generation of the perturbed reward is a function of S*R into the perturbed reward space. But then the confusion matrix does *not* take into account the state, which is justified by \"to let our presentation stay focused (...)\". I believe these elements should at least be clearly discussed. Indeed, in that setting, the theorems given seem to be variations of existing results and it is difficult to understand what is the message behind the theorems.\n\nIn addition, it is assumed that the confusion matrix C is known or estimated from data but it's not clear to me how this can be done in practice.  In equation 4, how do you have access to the predicted true rewards?\n\nAdditional comments:\n- The discount factor can be 0 but can not, in general, be equal to 1. So the equation in paragraph 2.1 \"0 < γ ≤ 1\" is wrong.\n- The paper mention that \"an underwhelming amount of reinforcement learning studies have focused on the settings with perturbed and noisy rewards\" but there are some works on the subject (e.g., https://arxiv.org/abs/1805.03359) and a discussion about the differences with the related work would be interesting.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541187978948}], "openreview_url": "https://openreview.net/forum?id=BkMWx309FX", "arxiv_id": "1810.01032", "paper_pdf": "papers/BkMWx309FX.pdf", "paper_pdf_sha256": "f88345c5ee3b037231cf85c768753d4cdaf0d61507eb4108601d785a3e4b01c4", "paper_pdf_bytes": 7436475, "paper_pdf_source": "openreview", "code_url": "https://github.com/wangjksjtu/rl-perturbed-reward", "code_repository": "wangjksjtu/rl-perturbed-reward", "code_commit": "c894ca5dcdeadd0a0907770bb093b703092e1da1", "code_archive": "repos/BkMWx309FX.zip", "code_archive_sha256": "d839df87bf77f18b3f67e81d3a60f03804bd8cf1e9c1f93f6ac9f7ea4de36e74", "code_archive_bytes": 177778, "code_file_count": 99, "code_extensions": {".py": 84, ".sh": 15}, "github_disk_usage_kb": 164, "github_languages": {"Python": 477858, "Shell": 76269}, "github_archived": false, "github_pushed_at": "2024-08-02T16:11:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/reinforcement-learning-with-perturbed-rewards"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1YqWz-R-", "year": 2018, "status": "rejected", "title": "Improving Conditional Sequence Generative Adversarial Networks by Stepwise Evaluation", "authors": ["Yi-Lin Tuan", "Hung-yi Lee"], "authorids": ["pascaltuan@gmail.com", "hungyilee@ntu.edu.tw"], "authors_source": "OpenReview API", "abstract": "Conditional sequence generation is a widely researched topic. One of the most important tasks is dialogue generation, which is composed of input-output pairs with the one-to-many property. Given the recent success of generative adversarial networks (GANs), GANs have been used for sequence generation. However, there is still limited work of its application on conditional sequence generation. We investigate the influence of GAN on conditional sequence generation with three artificial grammars and  dialogue generation. Moreover, we propose stepwise GAN (StepGAN) for conditional sequence generation, which predicts the reward at each time-step. StepGAN can be seen as the general version of SeqGAN. It estimates the expected returns predicted by Monte-Carlo Search in SeqGAN, but it has a lower computational cost than Monte-Carlo Search. Experimental results show that stepwise GAN can outperform other state-of-the-art algorithms in most tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJygOiYxM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper780/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present a new scheme for applying adversarial networks to dialog generation. The idea of why using adversarial networks is important in dialog generation is really well motivated in the paper and related works are discussed in details. \n\nIn the proposed approach, a more flexible discrimination score is obtained by treating independently each sub-sequence of the input. Technically speaking, the authors' contribution is to add a set of free parameters in the sub-sequence discriminator sum of equation 8. From a more general point of view, what is the key output of the paper, except to confirm that curriculum learning can help in dialog generation?  \n\nThe experiments do not seem to show that a net performance improvement can be associated with the introduced free weights and what is a good strategy to tune them in an optimal way. In general, as the authors report also in the abstract, the performance of the proposed algorithm is 'comparable' with the state of the art but never outperforms other existing methods in a consistent way. \n\nIn fact, the performance of the algorithms depends strongly on the specific grammar used to generate the dataset and on the specific evaluation score. The human evaluation experiment is interesting but the proposed method is only compared with one other algorithm (seqGAN) and only two examples of the output are given explicitly.\n\nThe increase in computational cost due to the weighted sub-sequence evaluation is also poorly discussed.   \n\nFew more questions:\n-through the experiments section, the authors focus on evaluating the set of possible good answers (softmax and coverage score) instead of the best answer (argmax). Why is this important for dialog generation? In the generation of a real conversation, shouldn t one always choose the argmax option? What would be a practical use of the second and third-best options?\n-why softmax is always lower than argmax in the synthetic experiment and always higher than argmax in the human evaluation experiment?\n-why MLE, which is used as initialization, does better than all optimized models in the first simulation? Why is the GAN approach expected to increase the coverage compared to MLE? And why, in general, this is not always the case?\n-would it be possible to compare the output of the proposed methods with the output of a non-GAN conditional sequence generator (if any) on human-scored dialog?  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A well written paper with somehow weak experimental results", "rating": "5: Marginally below acceptance threshold", "review": "The authors present a new scheme for applying adversarial networks to dialog generation. The idea of why using adversarial networks is important in dialog generation is really well motivated in the paper and related works are discussed in details. \n\nIn the proposed approach, a more flexible discrimination score is obtained by treating independently each sub-sequence of the input. Technically speaking, the authors' contribution is to add a set of free parameters in the sub-sequence discriminator sum of equation 8. From a more general point of view, what is the key output of the paper, except to confirm that curriculum learning can help in dialog generation?  \n\nThe experiments do not seem to show that a net performance improvement can be associated with the introduced free weights and what is a good strategy to tune them in an optimal way. In general, as the authors report also in the abstract, the performance of the proposed algorithm is 'comparable' with the state of the art but never outperforms other existing methods in a consistent way. \n\nIn fact, the performance of the algorithms depends strongly on the specific grammar used to generate the dataset and on the specific evaluation score. The human evaluation experiment is interesting but the proposed method is only compared with one other algorithm (seqGAN) and only two examples of the output are given explicitly.\n\nThe increase in computational cost due to the weighted sub-sequence evaluation is also poorly discussed.   \n\nFew more questions:\n-through the experiments section, the authors focus on evaluating the set of possible good answers (softmax and coverage score) instead of the best answer (argmax). Why is this important for dialog generation? In the generation of a real conversation, shouldn t one always choose the argmax option? What would be a practical use of the second and third-best options?\n-why softmax is always lower than argmax in the synthetic experiment and always higher than argmax in the human evaluation experiment?\n-why MLE, which is used as initialization, does better than all optimized models in the first simulation? Why is the GAN approach expected to increase the coverage compared to MLE? And why, in general, this is not always the case?\n-would it be possible to compare the output of the proposed methods with the output of a non-GAN conditional sequence generator (if any) on human-scored dialog?  ", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1511794663506}, {"id": "HyRbL5YxM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper780/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "UPDATE, 1/11/18:\nI read the revision. The writing has improved, and the new experiments are good. That said, it is my opinion that the paper is still not quite ready for publication, mostly because the story behind the model doesn't lead to the conclusions being made from the experiments in a clean and consistent way. It's a bit like patchwork at this point. The authors I think will need to put some time into a rewrite, but the content itself is worth pushing forward.\n\nSome notes:\n\"StepGAN a general version of SeqGAN, and can simulate the process of Monte-Carlo search with low extra computational cost:\" This really is a strong claim that's not proven in any way in the paper.\n\"In typical reinforcement learning, the agent obtains a reward...\": in typical RL settings, it's just as likely to find single or episodal rewards and this setting isn't limited to those where you have an extrinsic reward at each step.\nWhy do you still have WGAN-GP when there are no accompanying numbers?\n\n/begin old review\nThe approach is interesting, but as a contribution the paper has a long way to go. The ideas are there and everything seems correct, but there’s little motivation / insight on the model and why it might be better than competing methods for NLP tasks.\n\nIt would be good to see some sort of concrete analysis as far as what the model is doing (for instance how the discriminator scores change), and a comparison of how the reward signal given here might differ from other methods (SeqGAN, MaliGAN), and why this might be better. All we get is some scores, but it’s never clear why these scores indicate good (conditional) language generation. Can we not also look at BLEU scores for language generation or some other metric?\n\nFinally, the writing need to be improved: it starts out OK, but it progressively gets worse and worse.\n\nDetailed notes:\nP2\nIt might be good to mention beam search and scheduled sampling as other common methods to address the exposure bias.\n“objective function irrelevant to backpropagation”: what does this mean?\nMaliGAN actually also uses a “policy gradient”, which corresponds to an estimate of the likelihood ratio, to address the discrete problem.\nThough distinct from this work, Gulrajani used a CNN.\n\nP3\nUse \\log for logarithm\n“Moreover, the likelihood is only estimated at word-level”: is this true? It seems to me that likelihood of the sequence is estimated as well.\n\nP5\nWhy is this a generalized version of SeqGAN? The claim the discriminator value D(x_1..t | y) is the same as what you would get from a full-sequence generator using MCMC seems like a stretch.\nDid you not use a baseline?\nYou might want to build in a little more motivation for these different update rules. I think I understand that (10) is meant to accumulate credit across the rest of the sequence, while (11) does not, but it would be good to have this clearly stated. Why do you think one would work better than the other?\n\nP6\n“The generator G, in the mean time, struggle to maximize the likelihood of discriminator D” I don’t understand what this means.\nWhat was the motivation for using the same model here? Is the energy in this formulation related in any ways to EBGAN? Could you do something similar with separate parameters? Why would be or why would this not be a good idea?\nI do like these synthetic tasks. I think that more analyses would be helpful in understanding what the model (and what competing models) are doing.\n\nP7:\nWhat is VLGAN in the table?\nPerhaps it would be worth exploring changing alpha through optimization?\n\nP8:\nIt seems like many of the improvements in the table are marginal (with some exceptions): is it possible that ESGAN was optimized better?\n“auxiliary tricks” I would avoid this wording.\n\nOther comments on experiments:\nIt seems like the actual NLP part of this paper is quite sparse. Why was MaliGAN left out of the real experiments?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and potentially powerful approach to language generation, but incomplete and lacking insight / more thorough evaluation", "rating": "5: Marginally below acceptance threshold", "review": "UPDATE, 1/11/18:\nI read the revision. The writing has improved, and the new experiments are good. That said, it is my opinion that the paper is still not quite ready for publication, mostly because the story behind the model doesn't lead to the conclusions being made from the experiments in a clean and consistent way. It's a bit like patchwork at this point. The authors I think will need to put some time into a rewrite, but the content itself is worth pushing forward.\n\nSome notes:\n\"StepGAN a general version of SeqGAN, and can simulate the process of Monte-Carlo search with low extra computational cost:\" This really is a strong claim that's not proven in any way in the paper.\n\"In typical reinforcement learning, the agent obtains a reward...\": in typical RL settings, it's just as likely to find single or episodal rewards and this setting isn't limited to those where you have an extrinsic reward at each step.\nWhy do you still have WGAN-GP when there are no accompanying numbers?\n\n/begin old review\nThe approach is interesting, but as a contribution the paper has a long way to go. The ideas are there and everything seems correct, but there’s little motivation / insight on the model and why it might be better than competing methods for NLP tasks.\n\nIt would be good to see some sort of concrete analysis as far as what the model is doing (for instance how the discriminator scores change), and a comparison of how the reward signal given here might differ from other methods (SeqGAN, MaliGAN), and why this might be better. All we get is some scores, but it’s never clear why these scores indicate good (conditional) language generation. Can we not also look at BLEU scores for language generation or some other metric?\n\nFinally, the writing need to be improved: it starts out OK, but it progressively gets worse and worse.\n\nDetailed notes:\nP2\nIt might be good to mention beam search and scheduled sampling as other common methods to address the exposure bias.\n“objective function irrelevant to backpropagation”: what does this mean?\nMaliGAN actually also uses a “policy gradient”, which corresponds to an estimate of the likelihood ratio, to address the discrete problem.\nThough distinct from this work, Gulrajani used a CNN.\n\nP3\nUse \\log for logarithm\n“Moreover, the likelihood is only estimated at word-level”: is this true? It seems to me that likelihood of the sequence is estimated as well.\n\nP5\nWhy is this a generalized version of SeqGAN? The claim the discriminator value D(x_1..t | y) is the same as what you would get from a full-sequence generator using MCMC seems like a stretch.\nDid you not use a baseline?\nYou might want to build in a little more motivation for these different update rules. I think I understand that (10) is meant to accumulate credit across the rest of the sequence, while (11) does not, but it would be good to have this clearly stated. Why do you think one would work better than the other?\n\nP6\n“The generator G, in the mean time, struggle to maximize the likelihood of discriminator D” I don’t understand what this means.\nWhat was the motivation for using the same model here? Is the energy in this formulation related in any ways to EBGAN? Could you do something similar with separate parameters? Why would be or why would this not be a good idea?\nI do like these synthetic tasks. I think that more analyses would be helpful in understanding what the model (and what competing models) are doing.\n\nP7:\nWhat is VLGAN in the table?\nPerhaps it would be worth exploring changing alpha through optimization?\n\nP8:\nIt seems like many of the improvements in the table are marginal (with some exceptions): is it possible that ESGAN was optimized better?\n“auxiliary tricks” I would avoid this wording.\n\nOther comments on experiments:\nIt seems like the actual NLP part of this paper is quite sparse. Why was MaliGAN left out of the real experiments?", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511790086039}, {"id": "HkTXuSulM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper780/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Quality: The paper proposes a direct improvement over SeqGAN by Yu. et. al. (2017). My assessment is partially determined by comparing this paper to Yu et. al (2017). In my opinion, this paper is lacking in quality in comparison to Yu et. al (2017). In particular, Yu et. al. (2017) provides detailed derivation of the policy gradient accompanied by a pseudo-code (algorithm) on how one can implement SeqGAN. On the contrary, this paper does not provide such details. Perhaps, all of the details of SeqGAN follows immediately, but the paper should not assume that all readers will be familiar with SeqGAN. \n\nClarity: \n\n1. The paper provides a review of related methods on conditional sequence generation in Section 3. However, it is very brief and as a non-expert in this field, I needed to refer to the original papers anyways. Perhaps, the review of the related methods can go to the Appendix and this space can be better utilized to expand on the original contributions made by the paper. \n2. MCMC (Markov chain Monte Carlo) is mentioned in 4.1 but it is not explained. \n3. Figure 1 is not sufficiently explained; neither in text nor in the figure caption. It would help greatly to describe the details of the network architecture shown in this figure.\n\nOriginality: The paper proposes a generalization of SeqGAN; however, in my opinion, the methodological contribution appears to be only incremental on SeqGAN. \n\nSignificance: The paper's significance may be evaluated in terms of its impact on applications as it proposes an improvement over the previous work of SeqGAN. However, the extent to which the evaluation is carried out is somewhat unsatisfactory with only one real application. Also, the applications considered in the experiments are primarily on dialogue generation. My initial impression is that the methodology lacks generality and may perhaps cater better to domain specific publication venues. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is concerned with improving the sequence generation problem, in particular, for dialogue generation problem. The main contribution is in proposing to compute the cumulative reward for each step in the sequence generation procedure. The paper demonstrates that this leads to better performance in artificially generated grammar and for dialogue generation.", "rating": "4: Ok but not good enough - rejection", "review": "Quality: The paper proposes a direct improvement over SeqGAN by Yu. et. al. (2017). My assessment is partially determined by comparing this paper to Yu et. al (2017). In my opinion, this paper is lacking in quality in comparison to Yu et. al (2017). In particular, Yu et. al. (2017) provides detailed derivation of the policy gradient accompanied by a pseudo-code (algorithm) on how one can implement SeqGAN. On the contrary, this paper does not provide such details. Perhaps, all of the details of SeqGAN follows immediately, but the paper should not assume that all readers will be familiar with SeqGAN. \n\nClarity: \n\n1. The paper provides a review of related methods on conditional sequence generation in Section 3. However, it is very brief and as a non-expert in this field, I needed to refer to the original papers anyways. Perhaps, the review of the related methods can go to the Appendix and this space can be better utilized to expand on the original contributions made by the paper. \n2. MCMC (Markov chain Monte Carlo) is mentioned in 4.1 but it is not explained. \n3. Figure 1 is not sufficiently explained; neither in text nor in the figure caption. It would help greatly to describe the details of the network architecture shown in this figure.\n\nOriginality: The paper proposes a generalization of SeqGAN; however, in my opinion, the methodological contribution appears to be only incremental on SeqGAN. \n\nSignificance: The paper's significance may be evaluated in terms of its impact on applications as it proposes an improvement over the previous work of SeqGAN. However, the extent to which the evaluation is carried out is somewhat unsatisfactory with only one real application. Also, the applications considered in the experiments are primarily on dialogue generation. My initial impression is that the methodology lacks generality and may perhaps cater better to domain specific publication venues. \n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511704613231}], "openreview_url": "https://openreview.net/forum?id=r1YqWz-R-", "arxiv_id": "1808.05599", "paper_pdf": "papers/r1YqWz-R-.pdf", "paper_pdf_sha256": "6be0b57413bf711934e2ab5bbbf1d3a34f115a4e533c490b798d6fe8d3f62ca0", "paper_pdf_bytes": 862427, "paper_pdf_source": "openreview", "code_url": "https://github.com/Pascalson/Conditional-Seq-GANs", "code_repository": "Pascalson/Conditional-Seq-GANs", "code_commit": "6e7594a7c21826b14d34972b085478e3b3010375", "code_archive": "repos/r1YqWz-R-.zip", "code_archive_sha256": "9e5334f373c9f234803cc14a0fdf4d189ec65bbd7697228af1aed29dc9154ad1", "code_archive_bytes": 222557, "code_file_count": 11, "code_extensions": {".py": 10, ".sh": 1}, "github_disk_usage_kb": 227, "github_languages": {"Python": 97581, "Shell": 2139}, "github_archived": false, "github_pushed_at": "2019-07-13T11:18:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-conditional-sequence-generative"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lsLb0WvlqA", "year": 2026, "status": "rejected", "title": "Efficient Attention via Pre-Scoring: Prioritizing Informative Keys in Transformers", "authors": ["Yutong Bao", "Haoyu Wang", "Zhexiang Li", "David Woodruff"], "authorids": ["~Yutong_Bao1", "~Haoyu_Wang57", "~Zhexiang_Li1", "~David_Woodruff1"], "authors_source": "OpenReview API", "abstract": "Recent advances in transformer architectures deeply enhanced long-context language modeling. Among them, HyperAttention achieves competitive efficiency by combining a single-level LSH-based clustering with uniform residual sampling. However, HyperAttention fails to find all significant keys, which in turn raises the overall perplexity. We propose a pre-scoring mechanism that prioritizes significant keys before applying HyperAttention. We introduce three scoring methods: $k$-means and kernel $k$-means clustering, $k$-median clustering, and leverage score-based ranking (inspired by LevAttention) to filter keys effectively. We further replace HyperAttention's original uniform residual sampling, relying exclusively on our pre-scoring mechanism. Experiments on ChatGLM2 (131k token context) reduce perplexity from 12 to 8.3, which outperforms standard HyperAttention. Moreover, when running on the Vision-Transformer (ViT), our method shows that it can guarantee similar accuracy compared with LevAttention, and will surpass LevAttention given specific parameters. Although this method introduces some computational overhead, its combination with HyperAttention remains 20 times faster than FlashAttention, providing a balanced trade-off between speed and modeling accuracy. Our results highlight the effectiveness of integrating pre-scoring into hierarchical attention mechanisms, significantly improving transformer efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "uqAo23ywre", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18957/Reviewer_1BS1"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 3, "summary": "This paper introduces a new method to efficiently approximate the attention mechanism, with the goal of reducing the computational cost of this operation. The paper heavily relies on previous works LevAttention and HyperAttention, trying to combine the best of both methods. In particular, HyperAttention works by grouping keys and queries into bucket, using locality sensitive hashing (LSH), and then compute the attention only for keys and queries that are in the same bucket. On the other hand, LevAttention selects a subset of the most important keys, and compute the attention over these only.\n\nThis paper proposes to select a subset of the keys, using a clustering algorithm such as k-means or k-median, and then to apply the HyperAttention method on the selected keys. The paper also states some theoretical results about the selection process of the keys with the clustering algorith. Finally, some experimental results are provided, replacing the standard self-attention mechanism with the proposed method in existing models such as the GLM language model or vision transformer (ViT) models. Here, the paper show that the method improve the results of LevAttention or HyperAttention.", "review_text": "This paper introduces a new method to efficiently approximate the attention mechanism, with the goal of reducing the computational cost of this operation. The paper heavily relies on previous works LevAttention and HyperAttention, trying to combine the best of both methods. In particular, HyperAttention works by grouping keys and queries into bucket, using locality sensitive hashing (LSH), and then compute the attention only for keys and queries that are in the same bucket. On the other hand, LevAttention selects a subset of the most important keys, and compute the attention over these only.\n\nThis paper proposes to select a subset of the keys, using a clustering algorithm such as k-means or k-median, and then to apply the HyperAttention method on the selected keys. The paper also states some theoretical results about the selection process of the keys with the clustering algorith. Finally, some experimental results are provided, replacing the standard self-attention mechanism with the proposed method in existing models such as the GLM language model or vision transformer (ViT) models. Here, the paper show that the method improve the results of LevAttention or HyperAttention.", "strengths": "The paper provides some theoretical analysis for the proposed method, but it is hard for me to understand what kind of guarantee it actually provides (see next section).", "weaknesses": "Overall, I have many concerns with the paper.\n\nFirst, I found the paper very hard to read. One of the reason is that the authors assume that the reader are very familiar with previous works LevAttention and HyperAttention. For example, many concepts are not introduced in the paper (\"heavy attention scores\" line 59, \"statistical leverage scores\" line 99, \"polynomial based attention\" line 100, \"positional locality\" line 103, \"planted model\" line 135, etc...). Similarly, the different theorems or assumptions are stated without motivation or explanation, making it hard to understand how these relate to the performance of the method. Similarly, the algorithm proposed in the paper is never stated clearly, relying on previous paper instead. These different factors made it very hard for me to understand the method and theoretical claims.\n\nSecond, the paper mostly discusses LevAttention and HyperAttention as previous work to improve the efficiency of the self-attention. These two works are from 2024 and 2023 respectively, while there exists a wide body of earlier literature addressing this problem, and which are not discussed in the paper. Particularly relevant are the Reformer paper (Kitaev et al, 2020) which also proposes to use LSH to group keys and queries and restrict the self-attention between similar keys and queries, or Routing transformer (Roy et al, 2020) which proposes a similar approach based on k-means clustering.\n\nThird, I found the experimetal results to be unconvincing. The paper only compared the proposed approach to LevAttention and HyperAttention, and not earlier works which led to strong results. Second, the performance of the method seems to be quite poor. For example, on the language modeling experiments, the perplexity obtained with the different approximation techniques considered is above 10, while I believe that the perplexity of the original model is around 6. The performance of the original model should actually be included in Figure 3. Similarly, in Figure 4, the reported results for the considered methods are significantly worse than the original models, showing that the method is probably not useful in practice.", "questions": "What is the perplexity of the original GLM 2 language model (Fig. 3)?\n\n**Missing references**\n\nAurko Roy, Mohammad Saffar, Ashish Vaswani, David Grangier. 2020. Efficient Content-Based Sparse Attention with Routing Transformers. \n\nNikita Kitaev, Łukasz Kaiser, Anselm Levskaya. 2020. Reformer: The Efficient Transformer\n\nJack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, Timothy P. Lillicrap. 2019. Compressive Transformers for Long-Range Sequence Modelling\n\nZhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, Hongsheng Li. 2018. Efficient Attention: Attention with Linear Complexities\n\nSinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, Hao Ma. 2020. Linformer: Self-Attention with Linear Complexity\n\nAngelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, François Fleuret. 2020. Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention\n\nManzil Zaheer et al. 2020. Big Bird: Transformers for Longer Sequences.\n\nIz Beltagy, Matthew E. Peters, Arman Cohan. 2020. Longformer: The Long-Document Transformer\n\nRewon Child, Scott Gray, Alec Radford, Ilya Sutskever. 2019. Generating Long Sequences with Sparse Transformers\n\nYunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 2021. Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new method to efficiently approximate the attention mechanism, with the goal of reducing the computational cost of this operation. The paper heavily relies on previous works LevAttention and HyperAttention, trying to combine the best of both methods. In particular, HyperAttention works by grouping keys and queries into bucket, using locality sensitive hashing (LSH), and then compute the attention only for keys and queries that are in the same bucket. On the other hand, LevAttention selects a subset of the most important keys, and compute the attention over these only.\n\nThis paper proposes to select a subset of the keys, using a clustering algorithm such as k-means or k-median, and then to apply the HyperAttention method on the selected keys. The paper also states some theoretical results about the selection process of the keys with the clustering algorith. Finally, some experimental results are provided, replacing the standard self-attention mechanism with the proposed method in existing models such as the GLM language model or vision transformer (ViT) models. Here, the paper show that the method improve the results of LevAttention or HyperAttention.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "The paper provides some theoretical analysis for the proposed method, but it is hard for me to understand what kind of guarantee it actually provides (see next section).", "weaknesses": "Overall, I have many concerns with the paper.\n\nFirst, I found the paper very hard to read. One of the reason is that the authors assume that the reader are very familiar with previous works LevAttention and HyperAttention. For example, many concepts are not introduced in the paper (\"heavy attention scores\" line 59, \"statistical leverage scores\" line 99, \"polynomial based attention\" line 100, \"positional locality\" line 103, \"planted model\" line 135, etc...). Similarly, the different theorems or assumptions are stated without motivation or explanation, making it hard to understand how these relate to the performance of the method. Similarly, the algorithm proposed in the paper is never stated clearly, relying on previous paper instead. These different factors made it very hard for me to understand the method and theoretical claims.\n\nSecond, the paper mostly discusses LevAttention and HyperAttention as previous work to improve the efficiency of the self-attention. These two works are from 2024 and 2023 respectively, while there exists a wide body of earlier literature addressing this problem, and which are not discussed in the paper. Particularly relevant are the Reformer paper (Kitaev et al, 2020) which also proposes to use LSH to group keys and queries and restrict the self-attention between similar keys and queries, or Routing transformer (Roy et al, 2020) which proposes a similar approach based on k-means clustering.\n\nThird, I found the experimetal results to be unconvincing. The paper only compared the proposed approach to LevAttention and HyperAttention, and not earlier works which led to strong results. Second, the performance of the method seems to be quite poor. For example, on the language modeling experiments, the perplexity obtained with the different approximation techniques considered is above 10, while I believe that the perplexity of the original model is around 6. The performance of the original model should actually be included in Figure 3. Similarly, in Figure 4, the reported results for the considered methods are significantly worse than the original models, showing that the method is probably not useful in practice.", "questions": "What is the perplexity of the original GLM 2 language model (Fig. 3)?\n\n**Missing references**\n\nAurko Roy, Mohammad Saffar, Ashish Vaswani, David Grangier. 2020. Efficient Content-Based Sparse Attention with Routing Transformers. \n\nNikita Kitaev, Łukasz Kaiser, Anselm Levskaya. 2020. Reformer: The Efficient Transformer\n\nJack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, Timothy P. Lillicrap. 2019. Compressive Transformers for Long-Range Sequence Modelling\n\nZhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, Hongsheng Li. 2018. Efficient Attention: Attention with Linear Complexities\n\nSinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, Hao Ma. 2020. Linformer: Self-Attention with Linear Complexity\n\nAngelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, François Fleuret. 2020. Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention\n\nManzil Zaheer et al. 2020. Big Bird: Transformers for Longer Sequences.\n\nIz Beltagy, Matthew E. Peters, Arman Cohan. 2020. Longformer: The Long-Document Transformer\n\nRewon Child, Scott Gray, Alec Radford, Ilya Sutskever. 2019. Generating Long Sequences with Sparse Transformers\n\nYunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 2021. Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761932854774}, {"id": "KeJ61LN1N7", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18957/Reviewer_PL5q"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper proposes an attention acceleration scheme that pre-scores keys before applying HyperAttention. The pre-scoring can be done via k-means or k-median clustering, or via approximate leverage scores. The retained set of “informative” keys is then fed to HyperAttention, replacing its uniform residual sampling. Empirically, on ChatGLM2 and ChatGLM3 with LongBench prompts, the method lowers perplexity relative to vanilla HyperAttention and, at certain top-k settings, reports a best PPL near 8.3 from a HyperAttention baseline of roughly 12. It also reports layer-level speedups over FlashAttention for sufficiently long sequences and applies a similar key selection idea to ViT, showing accuracy approaching softmax attention when sampling enough keys. Theoretically, the paper analyzes a planted-subspace model and proves that clustering with  𝑘=𝑑+1 separates “signal” from “noise” rows comparably to leverage-score selection, giving recovery guarantees of heavy keys under assumptions like row-norm regularity", "review_text": "The paper proposes an attention acceleration scheme that pre-scores keys before applying HyperAttention. The pre-scoring can be done via k-means or k-median clustering, or via approximate leverage scores. The retained set of “informative” keys is then fed to HyperAttention, replacing its uniform residual sampling. Empirically, on ChatGLM2 and ChatGLM3 with LongBench prompts, the method lowers perplexity relative to vanilla HyperAttention and, at certain top-k settings, reports a best PPL near 8.3 from a HyperAttention baseline of roughly 12. It also reports layer-level speedups over FlashAttention for sufficiently long sequences and applies a similar key selection idea to ViT, showing accuracy approaching softmax attention when sampling enough keys. Theoretically, the paper analyzes a planted-subspace model and proves that clustering with  𝑘=𝑑+1 separates “signal” from “noise” rows comparably to leverage-score selection, giving recovery guarantees of heavy keys under assumptions like row-norm regularity", "strengths": "1. Targeting the recall gap of HyperAttention by ranking keys beforehand is a clean, practical idea that directly addresses missed heavy scores. The algorithms are presented with simple wrappers over HyperAttention. \n2.The planted-subspace analysis and Theorems 1–2 formalize when clustering isolates heavy keys, matching the empirical intuition that important keys align with near-orthogonal directions. \n3. Results span LongBench perplexity on GLM2 and GLM3, speed comparisons vs FlashAttention, and a ViT “monkey-patch,” giving a multi-angle view of trade-offs. \n4. The paper breaks down where overhead appears, how it scales with  k and  d, and when speedups emerge, which is valuable for deployment decisions.", "weaknesses": "1. The strongest PPL ≈ 8.3 appears tied to the min_seq_len ≥ n_query configuration and sometimes even top-k set to zero, which partially credits an optimization switch rather than the proposed pre-scoring itself. The paper should isolate gains from pre-scoring vs implementation flags and report both. \n2.Speedups are reported per layer against FlashAttention and discussed asymptotically, but it is unclear how these translate to whole-model throughput and latency under realistic batch sizes and sequence distributions. Consolidated end-to-end metrics are needed. \n3. The paper notes a “corrected coupling” for GLM3 that changes behavior relative to GLM2. This suggests results are sensitive to integration choices. The exact coupling and ablations should be elevated from appendix to main text with code pointers. \n4. The baseline set focuses on HyperAttention, FlashAttention, and leverage-based selection. Given recent efficient attention methods, the empirical section would be stronger with a few additional modern query-aware or block-sparse baselines, or at least a rationale for exclusions. \n5. Guarantees rely on row-norm regularity and separability that may not hold uniformly across layers or modalities. Although LayerNorm helps, some layers can exhibit skewed norms and mixed subspaces. Sensitivity analyses to violations of these assumptions would strengthen the claims.", "questions": "1. How much of the perplexity gain remains when min_seq_len ≥ n_query is disabled and the exact same HyperAttention kernels and fallbacks are used for all methods, including top-k=0 settings? Please provide a clean ablation table. \n2. Can you report end-to-end speed, throughput, and memory vs FlashAttention and HyperAttention on GLM2 and GLM3 for realistic prompt length distributions and batch sizes, not just per layer? \n3. How stable are results to the choice of  k, number of clusters, and initialization of k-means or k-median? For example, do different random seeds flip the identity of retained keys and the downstream PPL curve? \n4. Could you quantify the additional FLOPs and memory of pre-scoring at inference time and show how they amortize with increasing sequence length, for each variant? \n5. Beyond ViT, have you tried audio or multimodal encoders where key distributions differ strongly from text? Any failure cases that violate the planted-subspace intuition or row-norm regularity?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an attention acceleration scheme that pre-scores keys before applying HyperAttention. The pre-scoring can be done via k-means or k-median clustering, or via approximate leverage scores. The retained set of “informative” keys is then fed to HyperAttention, replacing its uniform residual sampling. Empirically, on ChatGLM2 and ChatGLM3 with LongBench prompts, the method lowers perplexity relative to vanilla HyperAttention and, at certain top-k settings, reports a best PPL near 8.3 from a HyperAttention baseline of roughly 12. It also reports layer-level speedups over FlashAttention for sufficiently long sequences and applies a similar key selection idea to ViT, showing accuracy approaching softmax attention when sampling enough keys. Theoretically, the paper analyzes a planted-subspace model and proves that clustering with  𝑘=𝑑+1 separates “signal” from “noise” rows comparably to leverage-score selection, giving recovery guarantees of heavy keys under assumptions like row-norm regularity", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Targeting the recall gap of HyperAttention by ranking keys beforehand is a clean, practical idea that directly addresses missed heavy scores. The algorithms are presented with simple wrappers over HyperAttention. \n2.The planted-subspace analysis and Theorems 1–2 formalize when clustering isolates heavy keys, matching the empirical intuition that important keys align with near-orthogonal directions. \n3. Results span LongBench perplexity on GLM2 and GLM3, speed comparisons vs FlashAttention, and a ViT “monkey-patch,” giving a multi-angle view of trade-offs. \n4. The paper breaks down where overhead appears, how it scales with  k and  d, and when speedups emerge, which is valuable for deployment decisions.", "weaknesses": "1. The strongest PPL ≈ 8.3 appears tied to the min_seq_len ≥ n_query configuration and sometimes even top-k set to zero, which partially credits an optimization switch rather than the proposed pre-scoring itself. The paper should isolate gains from pre-scoring vs implementation flags and report both. \n2.Speedups are reported per layer against FlashAttention and discussed asymptotically, but it is unclear how these translate to whole-model throughput and latency under realistic batch sizes and sequence distributions. Consolidated end-to-end metrics are needed. \n3. The paper notes a “corrected coupling” for GLM3 that changes behavior relative to GLM2. This suggests results are sensitive to integration choices. The exact coupling and ablations should be elevated from appendix to main text with code pointers. \n4. The baseline set focuses on HyperAttention, FlashAttention, and leverage-based selection. Given recent efficient attention methods, the empirical section would be stronger with a few additional modern query-aware or block-sparse baselines, or at least a rationale for exclusions. \n5. Guarantees rely on row-norm regularity and separability that may not hold uniformly across layers or modalities. Although LayerNorm helps, some layers can exhibit skewed norms and mixed subspaces. Sensitivity analyses to violations of these assumptions would strengthen the claims.", "questions": "1. How much of the perplexity gain remains when min_seq_len ≥ n_query is disabled and the exact same HyperAttention kernels and fallbacks are used for all methods, including top-k=0 settings? Please provide a clean ablation table. \n2. Can you report end-to-end speed, throughput, and memory vs FlashAttention and HyperAttention on GLM2 and GLM3 for realistic prompt length distributions and batch sizes, not just per layer? \n3. How stable are results to the choice of  k, number of clusters, and initialization of k-means or k-median? For example, do different random seeds flip the identity of retained keys and the downstream PPL curve? \n4. Could you quantify the additional FLOPs and memory of pre-scoring at inference time and show how they amortize with increasing sequence length, for each variant? \n5. Beyond ViT, have you tried audio or multimodal encoders where key distributions differ strongly from text? Any failure cases that violate the planted-subspace intuition or row-norm regularity?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761908259888}, {"id": "vtTVYKG2uq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18957/Reviewer_EPNi"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper proposes three methods to score keys before HyperAttention, enabling it to identify important keys: k-means, k-median, and leverage-score ranking. It then feeds the selected keys into HyperAttention, replacing its uniform residual sampling. Experiments on GLM2 show that the approach can be faster than FlashAttention. On ViT, the pre-scoring step captures heavy attention entries as well as, or better than, leverage scores.", "review_text": "The paper proposes three methods to score keys before HyperAttention, enabling it to identify important keys: k-means, k-median, and leverage-score ranking. It then feeds the selected keys into HyperAttention, replacing its uniform residual sampling. Experiments on GLM2 show that the approach can be faster than FlashAttention. On ViT, the pre-scoring step captures heavy attention entries as well as, or better than, leverage scores.", "strengths": "+ Clear and practical idea: The paper provides a straightforward approach to enhance HyperAttention by pre-scoring and then attending. This directly addresses a known issue: HyperAttention’s hashing is not aware of which keys matter, and LevAttention’s “universal set” can get large. The bridge between them is simple and useful in practice.\n\n+ Mix of theory and experiments: The paper offers proofs under a standard planted-subspace setup (to argue why the pre-scoring should work) and shows results on GLM2/GLM3 and ViT.", "weaknesses": "- Reason for PPL improvement: The best perplexity (~8.31) happens when pre-scoring is off (top-k = 0, sample_size = 0) and min_seq_len ≥ n_query is set. The paper itself says this gain comes from that configuration (forcing the faster block/tiled path), not from pre-scoring. A clean ablation is needed to separate the effects. \n\n- Unclear speedup claims: \n> Compared to the original HyperAttention, these methods can generate a mild acceleration, with performance becoming more remarkable starting at $2^{13}$ with a speedup factor of around 3 to 4 in Figure 1.\n\n&nbsp;&nbsp;&nbsp;&nbsp; The abstract says “up to 20× faster than FlashAttention” (when combined with HyperAttention), but the text says around 3–4× at $2^{13}$ for the pre-scored variants, and often the reported gains are relative to HyperAttention. Since HyperAttention is not the paper’s main contribution, this framing can be misleading. Please clarify the exact conditions for 20× vs the 3–4× cases and state whether the 3–4× is typical across settings.\n\n- Narrow baseline: Most comparisons are to HyperAttention and LevAttention. Adding Performer, Reformer, and newer retrieval/streaming methods would make the evaluation more complete and show the accuracy–speed trade-offs more clearly.\n\n- Pre-scoring overhead: k-means and k-median add non-trivial compute and can increase memory traffic, which may shrink speed gains, especially in multi-head settings. Please verify this with profiling across many heads, different head dimensions, and batch sizes, and report how much overhead comes from the forward pass vs backward.\n\n- Limited tasks and metrics: Using broader long-context tasks (e.g., QA, retrieval, summarization) and reporting task-level metrics (not just perplexity) would strengthen the paper and validate the method more fully.\n\n### Minor\n- Figure legend and axis text are small and hard to read.", "questions": "See the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes three methods to score keys before HyperAttention, enabling it to identify important keys: k-means, k-median, and leverage-score ranking. It then feeds the selected keys into HyperAttention, replacing its uniform residual sampling. Experiments on GLM2 show that the approach can be faster than FlashAttention. On ViT, the pre-scoring step captures heavy attention entries as well as, or better than, leverage scores.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "+ Clear and practical idea: The paper provides a straightforward approach to enhance HyperAttention by pre-scoring and then attending. This directly addresses a known issue: HyperAttention’s hashing is not aware of which keys matter, and LevAttention’s “universal set” can get large. The bridge between them is simple and useful in practice.\n\n+ Mix of theory and experiments: The paper offers proofs under a standard planted-subspace setup (to argue why the pre-scoring should work) and shows results on GLM2/GLM3 and ViT.", "weaknesses": "- Reason for PPL improvement: The best perplexity (~8.31) happens when pre-scoring is off (top-k = 0, sample_size = 0) and min_seq_len ≥ n_query is set. The paper itself says this gain comes from that configuration (forcing the faster block/tiled path), not from pre-scoring. A clean ablation is needed to separate the effects. \n\n- Unclear speedup claims: \n> Compared to the original HyperAttention, these methods can generate a mild acceleration, with performance becoming more remarkable starting at $2^{13}$ with a speedup factor of around 3 to 4 in Figure 1.\n\n&nbsp;&nbsp;&nbsp;&nbsp; The abstract says “up to 20× faster than FlashAttention” (when combined with HyperAttention), but the text says around 3–4× at $2^{13}$ for the pre-scored variants, and often the reported gains are relative to HyperAttention. Since HyperAttention is not the paper’s main contribution, this framing can be misleading. Please clarify the exact conditions for 20× vs the 3–4× cases and state whether the 3–4× is typical across settings.\n\n- Narrow baseline: Most comparisons are to HyperAttention and LevAttention. Adding Performer, Reformer, and newer retrieval/streaming methods would make the evaluation more complete and show the accuracy–speed trade-offs more clearly.\n\n- Pre-scoring overhead: k-means and k-median add non-trivial compute and can increase memory traffic, which may shrink speed gains, especially in multi-head settings. Please verify this with profiling across many heads, different head dimensions, and batch sizes, and report how much overhead comes from the forward pass vs backward.\n\n- Limited tasks and metrics: Using broader long-context tasks (e.g., QA, retrieval, summarization) and reporting task-level metrics (not just perplexity) would strengthen the paper and validate the method more fully.\n\n### Minor\n- Figure legend and axis text are small and hard to read.", "questions": "See the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761715583548}, {"id": "kAJWv5qYag", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18957/Reviewer_m6zB"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 2, "summary": "This paper proposes an extension of HyperAttention by introducing a pre-selection step using clustering methods (notably k-means) to prioritize informative keys before applying the LSH-based HyperAttention mechanism. The authors provide both theoretical analysis (via a planted-subspace model) and empirical results on long-context LLMs (ChatGLM2/3) and Vision Transformers.", "review_text": "This paper proposes an extension of HyperAttention by introducing a pre-selection step using clustering methods (notably k-means) to prioritize informative keys before applying the LSH-based HyperAttention mechanism. The authors provide both theoretical analysis (via a planted-subspace model) and empirical results on long-context LLMs (ChatGLM2/3) and Vision Transformers.", "strengths": "- The planted-subspace model provides a useful lens for analyzing why clustering can recover heavy keys, giving the method some analytical backing\n- The idea of preselecting the tokens is intuititive", "weaknesses": "- **Missing prior work:** The paper does not adequately discuss **Routing Transformer[1]**, which also introduced clustering for preselection of tokens. Furthermore, due to conceptual similarities, routing transformer should be one of the compared baselines. Moreover, apart from k-means clustering, **MoSA[2]** recently demonstrated the benefits of expert-choice routing for token preselection, and this should at least be discussed.\n- **Algorithmic ambiguity:** The training procedure for k-means clustering is underspecified—does it employ EMA updates with top-s selection, or is clustering recomputed per step? This is important for reproducibility.\n- **Autoregressivity concern:** The selection procedure relies on a **top-s operator**, which is inherently non-autoregressive (requiring access to future tokens). The implications for causal language modeling are not addressed.\n- **Formatting issues:** Several citations are incorrectly formatted (missing parentheses), which detracts from the paper’s polish.\n- **Weak gains over LevAttention:** The results do not demonstrate a significant gain over LevAttention baseline.\n- **Convoluted writing:** The writing is often hard to follow, paragraphs seem disconnected, and it is hard to merge them into a cohesive narrative.\n\n[1] - Efficient Content-Based Sparse Attention with Routing Transformers\n[2] - Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing", "questions": "- Why not **pre-select queries as well**, as done in Routing Transformer?\n- Why are **HyperAttention baseline results missing** from Table 2? Without them, it is difficult to measure the incremental gain from pre-scoring.\n- Under what conditions does the proposed method **outperform LevAttention**, given that LevAttention appears faster and in some cases more accurate?\n- How is **k-means training implemented**—does it rely on EMA updates, online clustering, or recomputation per batch?\n- How would the proposed method behave in a **fully autoregressive training regime**, where future tokens are not accessible for top-s selection?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an extension of HyperAttention by introducing a pre-selection step using clustering methods (notably k-means) to prioritize informative keys before applying the LSH-based HyperAttention mechanism. The authors provide both theoretical analysis (via a planted-subspace model) and empirical results on long-context LLMs (ChatGLM2/3) and Vision Transformers.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "- The planted-subspace model provides a useful lens for analyzing why clustering can recover heavy keys, giving the method some analytical backing\n- The idea of preselecting the tokens is intuititive", "weaknesses": "- **Missing prior work:** The paper does not adequately discuss **Routing Transformer[1]**, which also introduced clustering for preselection of tokens. Furthermore, due to conceptual similarities, routing transformer should be one of the compared baselines. Moreover, apart from k-means clustering, **MoSA[2]** recently demonstrated the benefits of expert-choice routing for token preselection, and this should at least be discussed.\n- **Algorithmic ambiguity:** The training procedure for k-means clustering is underspecified—does it employ EMA updates with top-s selection, or is clustering recomputed per step? This is important for reproducibility.\n- **Autoregressivity concern:** The selection procedure relies on a **top-s operator**, which is inherently non-autoregressive (requiring access to future tokens). The implications for causal language modeling are not addressed.\n- **Formatting issues:** Several citations are incorrectly formatted (missing parentheses), which detracts from the paper’s polish.\n- **Weak gains over LevAttention:** The results do not demonstrate a significant gain over LevAttention baseline.\n- **Convoluted writing:** The writing is often hard to follow, paragraphs seem disconnected, and it is hard to merge them into a cohesive narrative.\n\n[1] - Efficient Content-Based Sparse Attention with Routing Transformers\n[2] - Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing", "questions": "- Why not **pre-select queries as well**, as done in Routing Transformer?\n- Why are **HyperAttention baseline results missing** from Table 2? Without them, it is difficult to measure the incremental gain from pre-scoring.\n- Under what conditions does the proposed method **outperform LevAttention**, given that LevAttention appears faster and in some cases more accurate?\n- How is **k-means training implemented**—does it rely on EMA updates, online clustering, or recomputation per batch?\n- How would the proposed method behave in a **fully autoregressive training regime**, where future tokens are not accessible for top-s selection?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761659728324}], "openreview_url": "https://openreview.net/forum?id=lsLb0WvlqA", "arxiv_id": "2505.11040", "paper_pdf": "papers/lsLb0WvlqA.pdf", "paper_pdf_sha256": "f12242204e300c8735d1d6bd574343ec71c3ca2f6ded69722013d4c9877d1767", "paper_pdf_bytes": 602507, "paper_pdf_source": "openreview", "code_url": "https://github.com/BruceLZX/prescored-transformer", "code_repository": "BruceLZX/prescored-transformer", "code_commit": "87e83d8a060d476b72e4bc2c81e3689e8c8d2645", "code_archive": "repos/lsLb0WvlqA.zip", "code_archive_sha256": "126265f0d03425eece58038d3da8a5a63b7560c6f741f828805e3dcd979c1da6", "code_archive_bytes": 133948, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 139, "github_languages": {"Python": 126956}, "github_archived": false, "github_pushed_at": "2026-04-25T06:06:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/2505-11040"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NnwDdPDwUq", "year": 2025, "status": "rejected", "title": "Contextual Bandits with Entropy-based Human Feedback", "authors": ["Raihan Seraj", "Tristan Sylvain", "Lili Meng"], "authorids": ["~Raihan_Seraj1", "~Tristan_Sylvain2", "~Lili_Meng2"], "authors_source": "OpenReview API", "abstract": "In recent years, preference-based human feedback mechanisms have become integral to improving model performance across a range of applications, including conversational AI systems like ChatGPT. However, existing methodologies often overlook critical factors such as model uncertainty and variability in feedback quality. To address these limitations, we propose an innovative entropy-based human feedback framework designed for contextual bandits, which balances exploration and exploitation by soliciting expert feedback when model entropy surpasses a predefined threshold. Our method is model-agnostic and adaptable to any contextual bandit agent employing stochastic policies. Through rigorous experimentation, we demonstrate that our approach requires minimal human feedback to achieve significant performance gains, even with suboptimal feedback quality. Our work not only introduces a novel feedback solicitation strategy but also underscores the robustness of integrating human guidance into machine learning systems. Our code is publicly available: \\url{https://anonymous.4open.science/r/CBHF-33C5}", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "5Kdehs3wH5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7954/Reviewer_VubH"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces a new dimention into the contextual bandit formulation related to human feedback. The idea is that, during the training phase, human feedback can be queried and used instead of the action selected by the learning agent (or in the RM case, the human feedback may bias the reward). Experiments are conducted on a number of datasets and Learning algorithms to understand the performance of algorithms in this setting across different settings of human expertise and other parameters.", "review_text": "This paper introduces a new dimention into the contextual bandit formulation related to human feedback. The idea is that, during the training phase, human feedback can be queried and used instead of the action selected by the learning agent (or in the RM case, the human feedback may bias the reward). Experiments are conducted on a number of datasets and Learning algorithms to understand the performance of algorithms in this setting across different settings of human expertise and other parameters.", "strengths": "The setting explored in this paper seems novel to me, though I am not an expert in the Bandit space. I wonder if the area of off-policy bandits are relevant here ( see https://arxiv.org/abs/2010.12470) since by taking human feedback in some interactions, the bandit is somewat getting rewards \"off-policy\".\n\nThe combination of algorithms and settings explored seem quite thorough and from what I can tell, some interesting insights can be gleaned about the role of \"AR\" feedback vs \"RM\".", "weaknesses": "1. No theoretical analysis is done in this setting, which is usually the case for Bandit algoirthms, from my limited experience.\n\n2.  There is something about the formulation I dont get. It seems the bandit algorithms will be incentivized to maximize entropy as much as possible, in order to get the benefit of as much human feedback as possible (at least for a sufficient amount of expertise from the human). In other words, the formulation does not really assign any cost  to the act of getting human feedback during training, from what I can see.", "questions": "1. I am very confused by where the human feedback come from and how this relates to the baseline. I guess the HF is just computed by accessing the ground truth labels and generating the matching feedback. What exactly is the baseline? Does it never access human feedback at all? (if not this would also be a useful baseline to look at).\n\n2. Isn't regret a more common notion to evaluate bandit algorithms than mean cumulative reward computed AFTER training?\n\n3. minor: lines 14-15 of Alg 1 are really not necessary in an academic paper.\n\n4. eq 7: missing closing brackets.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new dimention into the contextual bandit formulation related to human feedback. The idea is that, during the training phase, human feedback can be queried and used instead of the action selected by the learning agent (or in the RM case, the human feedback may bias the reward). Experiments are conducted on a number of datasets and Learning algorithms to understand the performance of algorithms in this setting across different settings of human expertise and other parameters.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The setting explored in this paper seems novel to me, though I am not an expert in the Bandit space. I wonder if the area of off-policy bandits are relevant here ( see https://arxiv.org/abs/2010.12470) since by taking human feedback in some interactions, the bandit is somewat getting rewards \"off-policy\".\n\nThe combination of algorithms and settings explored seem quite thorough and from what I can tell, some interesting insights can be gleaned about the role of \"AR\" feedback vs \"RM\".", "weaknesses": "1. No theoretical analysis is done in this setting, which is usually the case for Bandit algoirthms, from my limited experience.\n\n2.  There is something about the formulation I dont get. It seems the bandit algorithms will be incentivized to maximize entropy as much as possible, in order to get the benefit of as much human feedback as possible (at least for a sufficient amount of expertise from the human). In other words, the formulation does not really assign any cost  to the act of getting human feedback during training, from what I can see.", "questions": "1. I am very confused by where the human feedback come from and how this relates to the baseline. I guess the HF is just computed by accessing the ground truth labels and generating the matching feedback. What exactly is the baseline? Does it never access human feedback at all? (if not this would also be a useful baseline to look at).\n\n2. Isn't regret a more common notion to evaluate bandit algorithms than mean cumulative reward computed AFTER training?\n\n3. minor: lines 14-15 of Alg 1 are really not necessary in an academic paper.\n\n4. eq 7: missing closing brackets.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730875355019}, {"id": "Mo5Z1vodZb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7954/Reviewer_mwci"], "rating": 5, "soundness": 2, "presentation": 4, "contribution": 1, "confidence": 3, "summary": "The paper proposes an entropy-based method to determine when to seek expert feedback actively and investigates the relationship between the type of expert feedback (action-based vs. preference-based) and the expert quality and their impact on bandit learning. \n\nOverall, even though the experiment section seems thorough, I have several concerns about the proposed methodology, which I will detail later.", "review_text": "The paper proposes an entropy-based method to determine when to seek expert feedback actively and investigates the relationship between the type of expert feedback (action-based vs. preference-based) and the expert quality and their impact on bandit learning. \n\nOverall, even though the experiment section seems thorough, I have several concerns about the proposed methodology, which I will detail later.", "strengths": "- The paper is well-written\n- The experiment is fairly thorough", "weaknesses": "Issues related to the methodology:\n1. The paper didn't cite a few influential works in this area: [1] acquires value annotation for labeling actions; [2] DAGGAR, which also directly gets experts to perform an action (similar to what this paper has proposed); and [3] APO, which actively selects which data to get trajectory-level preference label from. I especially consider [1] and [3] relevant to this paper's context.\n2. The methodology of simply selecting data points based on the policy's entropy lacks justification. Bayesian active learning frameworks such as BALD [4] already use entropy. How is reducing policy action entropy related to discovering arms with higher rewards? The paper offers hand-wavy justifications. The paper needs to come up with solid justifications for why the method should be chosen.\n\n[1] Tang, Shengpu, and Jenna Wiens. \"Counterfactual-augmented importance sampling for semi-offline policy evaluation.\" Advances in Neural Information Processing Systems 36 (2023): 11394-11429.\n\n[2] Ross, S., Gordon, G., & Bagnell, D. (2011, June). A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics (pp. 627-635). JMLR Workshop and Conference Proceedings.\n\n[3] Das, N., Chakraborty, S., Pacchiano, A., & Chowdhury, S. R. (2024). Active Preference Optimization for Sample Efficient RLHF. In ICML 2024 Workshop on Theoretical Foundations of Foundation Models.\n\n[4] Houlsby, N., Huszár, F., Ghahramani, Z., & Lengyel, M. (2011). Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745.\n\nComment on the experiment section:\n1. I do think the experiment section is clean and easy to follow at a glance. The subsection titles seem to want to summarize the findings. I do still find it lacking in terms of the key takeaway from each experiment. Sec 4.2 only says \"variations\" but offers no conclusive findings. Maybe you can summarize the results by grouping them into a few patterns and present them that way.\n2. Sec 4.4 is an interesting analysis. I agree that many bandit papers focus a lot on theory but lack comprehensive experimental ablations. This paper's experiment tries to offer insights which I see as a strength.", "questions": "Continuing from the lack of justification on the methodology, ideas like BALD or information-directed sampling [5] use entropy reduction as the selection criteria. Have you considered instead of using just entropy, use some form of entropy reduction as the criteria for asking for feedback?\n\n[5] Russo, Daniel, and Benjamin Van Roy. \"Learning to optimize via information-directed sampling.\" Advances in neural information processing systems 27 (2014).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an entropy-based method to determine when to seek expert feedback actively and investigates the relationship between the type of expert feedback (action-based vs. preference-based) and the expert quality and their impact on bandit learning. \n\nOverall, even though the experiment section seems thorough, I have several concerns about the proposed methodology, which I will detail later.", "soundness": 2, "presentation": 4, "contribution": 1, "strengths": "- The paper is well-written\n- The experiment is fairly thorough", "weaknesses": "Issues related to the methodology:\n1. The paper didn't cite a few influential works in this area: [1] acquires value annotation for labeling actions; [2] DAGGAR, which also directly gets experts to perform an action (similar to what this paper has proposed); and [3] APO, which actively selects which data to get trajectory-level preference label from. I especially consider [1] and [3] relevant to this paper's context.\n2. The methodology of simply selecting data points based on the policy's entropy lacks justification. Bayesian active learning frameworks such as BALD [4] already use entropy. How is reducing policy action entropy related to discovering arms with higher rewards? The paper offers hand-wavy justifications. The paper needs to come up with solid justifications for why the method should be chosen.\n\n[1] Tang, Shengpu, and Jenna Wiens. \"Counterfactual-augmented importance sampling for semi-offline policy evaluation.\" Advances in Neural Information Processing Systems 36 (2023): 11394-11429.\n\n[2] Ross, S., Gordon, G., & Bagnell, D. (2011, June). A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics (pp. 627-635). JMLR Workshop and Conference Proceedings.\n\n[3] Das, N., Chakraborty, S., Pacchiano, A., & Chowdhury, S. R. (2024). Active Preference Optimization for Sample Efficient RLHF. In ICML 2024 Workshop on Theoretical Foundations of Foundation Models.\n\n[4] Houlsby, N., Huszár, F., Ghahramani, Z., & Lengyel, M. (2011). Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745.\n\nComment on the experiment section:\n1. I do think the experiment section is clean and easy to follow at a glance. The subsection titles seem to want to summarize the findings. I do still find it lacking in terms of the key takeaway from each experiment. Sec 4.2 only says \"variations\" but offers no conclusive findings. Maybe you can summarize the results by grouping them into a few patterns and present them that way.\n2. Sec 4.4 is an interesting analysis. I agree that many bandit papers focus a lot on theory but lack comprehensive experimental ablations. This paper's experiment tries to offer insights which I see as a strength.", "questions": "Continuing from the lack of justification on the methodology, ideas like BALD or information-directed sampling [5] use entropy reduction as the selection criteria. Have you considered instead of using just entropy, use some form of entropy reduction as the criteria for asking for feedback?\n\n[5] Russo, Daniel, and Benjamin Van Roy. \"Learning to optimize via information-directed sampling.\" Advances in neural information processing systems 27 (2014).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730745081119}, {"id": "cbM2WbaZxg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7954/Reviewer_V8Dc"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "This paper proposes an entropy-based framework to incorporate human feedback into contextual bandits. Specifically, it extends the bandit setup by adding a human intervention component, i.e., at any time, the agent could decide whether to utilize the human expert, where the human expert can either provide the optimal action, or give certain reward penalty of the proposed action. The proposed algorithm basically utilizes any bandit algorithm as the backbone algorithm, and decides to call the human expert based on the entropy of policy $\\pi$, i.e., uncertainty in the policy decision. Experiments are done on some multi-class classification problem.", "review_text": "This paper proposes an entropy-based framework to incorporate human feedback into contextual bandits. Specifically, it extends the bandit setup by adding a human intervention component, i.e., at any time, the agent could decide whether to utilize the human expert, where the human expert can either provide the optimal action, or give certain reward penalty of the proposed action. The proposed algorithm basically utilizes any bandit algorithm as the backbone algorithm, and decides to call the human expert based on the entropy of policy $\\pi$, i.e., uncertainty in the policy decision. Experiments are done on some multi-class classification problem.", "strengths": "- The proposed method is simple, easy to implement, and presented in a clear way.\n- The setting of incorporating expert feedback/intervention in the classical bandit framework is interesting.", "weaknesses": "- The paper’s presentation is somewhat unclear, particularly in the problem formulation. It seems to propose a stronger variation of bandit problems where the model can access oracle labels, but this isn’t clearly explained. Section 3.1 could be revised to clarify this new setup.\n\n- With this revised formulation, it would also help to include a theoretical guarantee for the proposed algorithm, addressing the how it affects on overall regret, the lower bound on the new formulation, and whether the algorithm is optimal.\n\n- While direct supervision through oracle labels is understandable, the reward manipulation is confusing. This reward shaping doesn’t change the algorithm’s selected action; rather, it modifies the reward. It’s unclear why this is necessary when the underlying reward already reflects sub-optimality. Clarifying the motivation here would strengthen the paper.\n\n- The novelty of using entropy-based active learning is also limited; see [1] for a comprehensive review. The paper would benefit from a detailed comparison of its contributions to prior work, and highlight the novelty here.\n\n- Further, several hyper-parameters lack clarity, such as the reward penalty $r_p$ and entropy thresholds.\n\n- Finally, the empirical results are not presented in an especially meaningful way. For example, Figure 2 is difficult to interpret. The oracle access in action recommendation appears to give $K$ partial rewards (where $K$ is the action space), but baseline comparisons may be unfair as they lack oracle access. This makes the results less convincing.\n\n\n[1]. https://burrsettles.com/pub/settles.activelearning.pdf", "questions": "See weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an entropy-based framework to incorporate human feedback into contextual bandits. Specifically, it extends the bandit setup by adding a human intervention component, i.e., at any time, the agent could decide whether to utilize the human expert, where the human expert can either provide the optimal action, or give certain reward penalty of the proposed action. The proposed algorithm basically utilizes any bandit algorithm as the backbone algorithm, and decides to call the human expert based on the entropy of policy $\\pi$, i.e., uncertainty in the policy decision. Experiments are done on some multi-class classification problem.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "- The proposed method is simple, easy to implement, and presented in a clear way.\n- The setting of incorporating expert feedback/intervention in the classical bandit framework is interesting.", "weaknesses": "- The paper’s presentation is somewhat unclear, particularly in the problem formulation. It seems to propose a stronger variation of bandit problems where the model can access oracle labels, but this isn’t clearly explained. Section 3.1 could be revised to clarify this new setup.\n\n- With this revised formulation, it would also help to include a theoretical guarantee for the proposed algorithm, addressing the how it affects on overall regret, the lower bound on the new formulation, and whether the algorithm is optimal.\n\n- While direct supervision through oracle labels is understandable, the reward manipulation is confusing. This reward shaping doesn’t change the algorithm’s selected action; rather, it modifies the reward. It’s unclear why this is necessary when the underlying reward already reflects sub-optimality. Clarifying the motivation here would strengthen the paper.\n\n- The novelty of using entropy-based active learning is also limited; see [1] for a comprehensive review. The paper would benefit from a detailed comparison of its contributions to prior work, and highlight the novelty here.\n\n- Further, several hyper-parameters lack clarity, such as the reward penalty $r_p$ and entropy thresholds.\n\n- Finally, the empirical results are not presented in an especially meaningful way. For example, Figure 2 is difficult to interpret. The oracle access in action recommendation appears to give $K$ partial rewards (where $K$ is the action space), but baseline comparisons may be unfair as they lack oracle access. This makes the results less convincing.\n\n\n[1]. https://burrsettles.com/pub/settles.activelearning.pdf", "questions": "See weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730593969762}, {"id": "iV9raX4qp7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7954/Reviewer_adAk"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper proposes a new human-augmented method for solving contextual bandit problems. The idea is to use two types of human feedback to speed up identification of the best arm. The first type of feedback is “action recommendation”, where the human expert recommends a set of actions (arms) to take. Here, the algorithm selects a random action from the set to take. The second type is “reward manipulation”, which refers here to the human expert specifying a penalty that will be applied when the learner takes an action that is not in the expert’s recommended actions. Feedback is initiated by the learner, and a threshold on the entropy of the current policy is used to decide when to ask for feedback.", "review_text": "The paper proposes a new human-augmented method for solving contextual bandit problems. The idea is to use two types of human feedback to speed up identification of the best arm. The first type of feedback is “action recommendation”, where the human expert recommends a set of actions (arms) to take. Here, the algorithm selects a random action from the set to take. The second type is “reward manipulation”, which refers here to the human expert specifying a penalty that will be applied when the learner takes an action that is not in the expert’s recommended actions. Feedback is initiated by the learner, and a threshold on the entropy of the current policy is used to decide when to ask for feedback.", "strengths": "The paper tackles an important problem. Contextual bandit problems have many important applications, and the paper is right to recognise that human experts could help in solving these problems more quickly. The writing itself is good and the related work is extensive. The tasks considered in the experiments look to be realistically large and challenging.", "weaknesses": "The paper is in my view not sufficiently novel in its current form. The idea of collecting human feedback when uncertainty is high is a well-known heuristic (e.g. [A]).  I think a much larger contribution could have been achieved in one of two ways: (1) by proving some statistical guarantees, which I would expect is possible give the simplicity of the algorithm and of the model of the human, or (2) by following some of the more SOTA work in this field and achieving a more refined trade-off between the cost of human feedback (e.g. cognitive cost [B]) and its value.\n\nFurthermore, the experiments leave a lot to be desired.\n- The proposed method leverages a human expert, but there is no human subject study. Humans give feedback of varying levels of quality (both between and within subjects) with potential biases [C]. The effect this would have on the proposed method is unclear as this variability is not replicated in the simulation experiments (feedback is assumed to be of constant quality and unbiased).\n- As we are discussing the addition of human feedback to an environmental reward here, I would expect to see an ablation that includes a “no feedback” setting, i.e. the standard contextual bandit setting.\n- To understand the value a human can expert bring here, and how practical the proposed method is, I would want to see an experiments that measures how much feedback is needed to achieve a certain level of performance.\n\nThere are a number of issues with presentation. See also the questions below for things that were unclear in the text.\n- Figure 2 lacks error bars. Subfigures showing performance of different algorithms on the same dataset have different scale y-axes, making comparison difficult.\n- Section 4.4 and Figure 4 show the effect of various entropy thresholds, but what those thresholds are is never stated.\n- Table 2 in the appendix is cut off at the bottom.\n- Section 3.2.2 explains that for “reward manipulation” feedback, the expert given a penalty to be applied when a non recommended action is taken. However, this is not consistent with Algorithm 1, which shows the penalty being applied for any action.\n\n[A] Raghu, Maithra, et al. \"The algorithmic automation problem: Prediction, triage, and human effort.\" arXiv preprint arXiv:1903.12220 (2019).\n[B] Banerjee, Rohan, et al. \"To Ask or Not To Ask: Human-in-the-loop Contextual Bandits with Applications in Robot-Assisted Feeding.\" arXiv preprint arXiv:2405.06908 (2024).\n[C] Ji, Xiang, et al. \"Provable benefits of policy learning from human preferences in contextual bandit problems.\" arXiv preprint arXiv:2307.12975 (2023).", "questions": "- On line 134, you assume reward is 0/1. This is a suitable assumption for multi-label classification problems, but seems needlessly restrictive for the proposed method. Could this assumption be loosened?\n- You measure mean cumulative reward in various places, and figure 8 given an idea of how often expert feedback is elicited for various entropy thresholds. However, in order to understand whether the proposed method is practical, I would like have a better understanding of how much expert feedback is needed to achieve a certain level of mean reward. What does that trade-off look like?\n- For the experiments with action recommendations, how many actions were recommended? What was the effect of recommending more actions?\n- How was the quality of feedback parameter used within the experiments? Was it assumed to be known or latent? Were algorithms exposed to a single level of feedback, or could it change within an experiment? I ask because the RL algorithms performed quite well in your experiments, which suggests that they may have been able to adapt to the given level of feedback, even if it was not observable to them.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new human-augmented method for solving contextual bandit problems. The idea is to use two types of human feedback to speed up identification of the best arm. The first type of feedback is “action recommendation”, where the human expert recommends a set of actions (arms) to take. Here, the algorithm selects a random action from the set to take. The second type is “reward manipulation”, which refers here to the human expert specifying a penalty that will be applied when the learner takes an action that is not in the expert’s recommended actions. Feedback is initiated by the learner, and a threshold on the entropy of the current policy is used to decide when to ask for feedback.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "The paper tackles an important problem. Contextual bandit problems have many important applications, and the paper is right to recognise that human experts could help in solving these problems more quickly. The writing itself is good and the related work is extensive. The tasks considered in the experiments look to be realistically large and challenging.", "weaknesses": "The paper is in my view not sufficiently novel in its current form. The idea of collecting human feedback when uncertainty is high is a well-known heuristic (e.g. [A]).  I think a much larger contribution could have been achieved in one of two ways: (1) by proving some statistical guarantees, which I would expect is possible give the simplicity of the algorithm and of the model of the human, or (2) by following some of the more SOTA work in this field and achieving a more refined trade-off between the cost of human feedback (e.g. cognitive cost [B]) and its value.\n\nFurthermore, the experiments leave a lot to be desired.\n- The proposed method leverages a human expert, but there is no human subject study. Humans give feedback of varying levels of quality (both between and within subjects) with potential biases [C]. The effect this would have on the proposed method is unclear as this variability is not replicated in the simulation experiments (feedback is assumed to be of constant quality and unbiased).\n- As we are discussing the addition of human feedback to an environmental reward here, I would expect to see an ablation that includes a “no feedback” setting, i.e. the standard contextual bandit setting.\n- To understand the value a human can expert bring here, and how practical the proposed method is, I would want to see an experiments that measures how much feedback is needed to achieve a certain level of performance.\n\nThere are a number of issues with presentation. See also the questions below for things that were unclear in the text.\n- Figure 2 lacks error bars. Subfigures showing performance of different algorithms on the same dataset have different scale y-axes, making comparison difficult.\n- Section 4.4 and Figure 4 show the effect of various entropy thresholds, but what those thresholds are is never stated.\n- Table 2 in the appendix is cut off at the bottom.\n- Section 3.2.2 explains that for “reward manipulation” feedback, the expert given a penalty to be applied when a non recommended action is taken. However, this is not consistent with Algorithm 1, which shows the penalty being applied for any action.\n\n[A] Raghu, Maithra, et al. \"The algorithmic automation problem: Prediction, triage, and human effort.\" arXiv preprint arXiv:1903.12220 (2019).\n[B] Banerjee, Rohan, et al. \"To Ask or Not To Ask: Human-in-the-loop Contextual Bandits with Applications in Robot-Assisted Feeding.\" arXiv preprint arXiv:2405.06908 (2024).\n[C] Ji, Xiang, et al. \"Provable benefits of policy learning from human preferences in contextual bandit problems.\" arXiv preprint arXiv:2307.12975 (2023).", "questions": "- On line 134, you assume reward is 0/1. This is a suitable assumption for multi-label classification problems, but seems needlessly restrictive for the proposed method. Could this assumption be loosened?\n- You measure mean cumulative reward in various places, and figure 8 given an idea of how often expert feedback is elicited for various entropy thresholds. However, in order to understand whether the proposed method is practical, I would like have a better understanding of how much expert feedback is needed to achieve a certain level of mean reward. What does that trade-off look like?\n- For the experiments with action recommendations, how many actions were recommended? What was the effect of recommending more actions?\n- How was the quality of feedback parameter used within the experiments? Was it assumed to be known or latent? Were algorithms exposed to a single level of feedback, or could it change within an experiment? I ask because the RL algorithms performed quite well in your experiments, which suggests that they may have been able to adapt to the given level of feedback, even if it was not observable to them.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730452761619}], "openreview_url": "https://openreview.net/forum?id=NnwDdPDwUq", "arxiv_id": "2502.08759", "paper_pdf": "papers/NnwDdPDwUq.pdf", "paper_pdf_sha256": "c9b7029fa311897cb7ca0dd4978772ad01f733d4b292dd4d24b0a9ef48a6a5e1", "paper_pdf_bytes": 1401169, "paper_pdf_source": "openreview", "code_url": "https://github.com/BorealisAI/CBHF", "code_repository": "BorealisAI/CBHF", "code_commit": "46fdc310b814fd6e0850804c1800fe386d8acca4", "code_archive": "repos/NnwDdPDwUq.zip", "code_archive_sha256": "f3b6ce069637d1176d7f7aff45bfb83cc97e6e37b63c7ec88983102525f4ca8d", "code_archive_bytes": 43640, "code_file_count": 30, "code_extensions": {".py": 21, ".sh": 9}, "github_disk_usage_kb": 32, "github_languages": {"Python": 106712, "Shell": 6517}, "github_archived": false, "github_pushed_at": "2024-05-31T20:35:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contextual-bandits-with-entropy-based-human"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PIl69UIAWL", "year": 2024, "status": "rejected", "title": "GraphLLM: Boosting Graph Reasoning Ability of Large Language Model", "authors": ["Ziwei Chai", "Tianjie Zhang", "Liang Wu", "Kaiqiao Han", "Xiaohai Hu", "Xuanwen Huang", "Yang Yang"], "authorids": ["~Ziwei_Chai1", "~Tianjie_Zhang1", "~Liang_Wu8", "~Kaiqiao_Han1", "~Xiaohai_Hu1", "~Xuanwen_Huang1", "~Yang_Yang35"], "authors_source": "OpenReview API", "abstract": "The advancement of Large Language Models (LLMs) has remarkably pushed the boundaries towards artificial general intelligence (AGI), with their exceptional ability on understanding diverse types of information,  including but not limited to images and audio.  Despite this progress, a critical gap remains in empowering LLMs to proficiently understand and reason on graph data. Recent studies underscore LLMs' underwhelming performance on fundamental graph reasoning tasks.  In this paper, we endeavor to unearth the obstacles that impede LLMs in graph reasoning, pinpointing the common practice of converting graphs into natural language descriptions (Graph2Text) as a fundamental bottleneck. To overcome this impediment, we introduce GraphLLM, a pioneering end-to-end approach that synergistically integrates graph learning models with LLMs. This synergy equips LLMs with the ability to proficiently interpret and reason on graph data, harnessing the superior expressive power of graph learning models. Our empirical evaluations across four fundamental graph reasoning tasks validate the effectiveness of GraphLLM. The results exhibit a substantial average accuracy enhancement of 54.44%, alongside a noteworthy context reduction of 96.45% across various graph reasoning tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "R4Tl9mnLrC", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4455/Reviewer_wWow"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces GraphLLM, an approach to integrate graph learning models with LLMs via graph transformer and prefix tuning. The authors compare with baselines across four graph reasoning tasks.", "review_text": "This paper introduces GraphLLM, an approach to integrate graph learning models with LLMs via graph transformer and prefix tuning. The authors compare with baselines across four graph reasoning tasks.", "strengths": "I can see the contribution of this paper in trying to apply LLMs for graph learning/reasoning tasks, which I think is interesting.\n\nThe paper is well-organized.\n\nOther than the performance comparison, the authors report the efficiency comparison, which I appreciate.\n\nThe authors further report the performance of gpt-3.5-turbo and gpt-4, other than LLama.", "weaknesses": "The authors claim their proposed method is an end-to-end approach. I wonder the applicability of their model to existing pre-trained LLMs. Can the proposed method be easily adapted or integrated into existing pre-trained LLMs without fine-tuning? If so, what is the formal definition of end-to-end? If not, will this introduce a large burden for training resources? It seems to me that end-to-end is not a good strategy for LLMs, which contradicts the most powerful capability of one model to fit multiple tasks and applications in LLMs.\n\nThere are multiple ways to condense the length and size of context windows, especially in the community of natural language processing. The authors should be more cautious about this when presenting this as the contribution, and should consider comparing their methods with baselines and approaches that address the length and size constraints in the NLP community.\n\nAmbiguous preliminary. In the paper of prefix tuning, the prefix indicates a vector that concats to the front of the input and the representations in hidden layers. In the prefix tuning section of this paper, the authors introduce notations of QKV while not giving formal definitions that correspond to the graph learning scenario. Moreover, what is the input here, only K and V? Should the input also be Q? Before each layer in the transformer, where should the prefix be concatenated? How is this related to the graph learning scenario?\n\nLimited technical novelty. This paper simply combines the graph transformer and prefix tuning of LLMs. In particular, the graph transformer serves as a learning model on top of the graph, then the output of the graph transformer is concatenated to LLMs via prefix tuning. The technics of graph transformer and prefix tuning are both well built, studied and employed by the research community.\n\nHave authors consider comparing with or integrating P-tuning v2, which has shown better performance than prefix tuning.\n\nI suspect that different instructions can cause the LLMs to generate significantly different results and performance. Have the authors conducted experiments or analyses to investigate different instructions? How did the authors choose the instructions?\n\nAll the tested graphs are relatively small. I wonder if the proposed method can be used in large-scale graphs such as ogb-arxiv/ogb-products/ogb-ppa. If the proposed method can be extended to large graphs, how will the authors sample the neighbors? Will neighbor sampling affect the performance?\n\nLack of baselines. One important line of baselines that are missed here is the graph neural network-based ones for graph reasoning tasks. I don’t think the experiments are comprehensive without comparing to methods that are designed to solve the graph reasoning problem. In fact, it is surprising that the authors only compare with NLP-based methods on solving graph tasks.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces GraphLLM, an approach to integrate graph learning models with LLMs via graph transformer and prefix tuning. The authors compare with baselines across four graph reasoning tasks.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "I can see the contribution of this paper in trying to apply LLMs for graph learning/reasoning tasks, which I think is interesting.\n\nThe paper is well-organized.\n\nOther than the performance comparison, the authors report the efficiency comparison, which I appreciate.\n\nThe authors further report the performance of gpt-3.5-turbo and gpt-4, other than LLama.", "weaknesses": "The authors claim their proposed method is an end-to-end approach. I wonder the applicability of their model to existing pre-trained LLMs. Can the proposed method be easily adapted or integrated into existing pre-trained LLMs without fine-tuning? If so, what is the formal definition of end-to-end? If not, will this introduce a large burden for training resources? It seems to me that end-to-end is not a good strategy for LLMs, which contradicts the most powerful capability of one model to fit multiple tasks and applications in LLMs.\n\nThere are multiple ways to condense the length and size of context windows, especially in the community of natural language processing. The authors should be more cautious about this when presenting this as the contribution, and should consider comparing their methods with baselines and approaches that address the length and size constraints in the NLP community.\n\nAmbiguous preliminary. In the paper of prefix tuning, the prefix indicates a vector that concats to the front of the input and the representations in hidden layers. In the prefix tuning section of this paper, the authors introduce notations of QKV while not giving formal definitions that correspond to the graph learning scenario. Moreover, what is the input here, only K and V? Should the input also be Q? Before each layer in the transformer, where should the prefix be concatenated? How is this related to the graph learning scenario?\n\nLimited technical novelty. This paper simply combines the graph transformer and prefix tuning of LLMs. In particular, the graph transformer serves as a learning model on top of the graph, then the output of the graph transformer is concatenated to LLMs via prefix tuning. The technics of graph transformer and prefix tuning are both well built, studied and employed by the research community.\n\nHave authors consider comparing with or integrating P-tuning v2, which has shown better performance than prefix tuning.\n\nI suspect that different instructions can cause the LLMs to generate significantly different results and performance. Have the authors conducted experiments or analyses to investigate different instructions? How did the authors choose the instructions?\n\nAll the tested graphs are relatively small. I wonder if the proposed method can be used in large-scale graphs such as ogb-arxiv/ogb-products/ogb-ppa. If the proposed method can be extended to large graphs, how will the authors sample the neighbors? Will neighbor sampling affect the performance?\n\nLack of baselines. One important line of baselines that are missed here is the graph neural network-based ones for graph reasoning tasks. I don’t think the experiments are comprehensive without comparing to methods that are designed to solve the graph reasoning problem. In fact, it is surprising that the authors only compare with NLP-based methods on solving graph tasks.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699246298775}, {"id": "jqyw6TDOOM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4455/Reviewer_ryJZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes GraphLLM, a new approach to enhance the ability of large language models (LLMs) to understand and reason about graph data. It identifies a key limitation of current methods that convert graphs to text (Graph2Text), which forces LLMs to implicitly learn graph structures and results in lengthy contexts. GraphLLM integrates a graph learning module with the LLM via end-to-end training. This allows the LLM to leverage the graph module's strengths in an efficient way through a condensed graph-enhanced prefix. Experiments on 4 graph reasoning tasks show GraphLLM substantially improves accuracy over Graph2Text methods while reducing context length and accelerating inference.", "review_text": "The paper proposes GraphLLM, a new approach to enhance the ability of large language models (LLMs) to understand and reason about graph data. It identifies a key limitation of current methods that convert graphs to text (Graph2Text), which forces LLMs to implicitly learn graph structures and results in lengthy contexts. GraphLLM integrates a graph learning module with the LLM via end-to-end training. This allows the LLM to leverage the graph module's strengths in an efficient way through a condensed graph-enhanced prefix. Experiments on 4 graph reasoning tasks show GraphLLM substantially improves accuracy over Graph2Text methods while reducing context length and accelerating inference.", "strengths": "* Novel integration of graphs and LLMs.\n\nGraphLLM proposes a novel end-to-end approach to combine graph learning models and LLMs. This allows each component to be optimized to complement the other for different graph reasoning tasks.\n\n* Significant performance gains.\n\nGraphLLM improves accuracy by 54.44% on average over the best Graph2Text method, showing its effectiveness. Also, GraphLLM reduces context length by 96.45% compared to Graph2Text, enhancing efficiency. Besides, GraphLLM achieves 3.42x faster inference over the best Graph2Text method due to context reduction.", "weaknesses": "* Limited graph tasks evaluated.\n\nThe paper evaluates GraphLLM on four graph reasoning tasks, including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching. Although these tasks cover basic graph reasoning abilities, they are still relatively simple, which might overestimate how GraphLLM performs on noisier graphs with complex relational patterns. More complex graph reasoning tasks could better demonstrate the capabilities and limitations of GraphLLM. For example, tasks requiring multi-hop reasoning on large graphs might be more challenging.\n\n* Lack of analysis on scalability.\n\nThe datasets used in the paper are all small with fewer than 20 nodes on average. Thus, the ability of GraphLLM to handle large, real-world graphs is unclear. Evaluations on larger datasets would be informative.\n\n* Restricted node features.\n\nThe tasks only use textual node features. Testing on graphs with non-textual node attributes would make the approach more broadly applicable.\n\n* No comparison to other graph models.\n\nThe paper only compares GraphLLM against methods that convert the graph to text (Graph2Text). However, established graph neural networks like GNNs, GCNs, and graph transformers have been optimized specifically for graph reasoning. Comparing to these dedicated graph reasoning models could better highlight if GraphLLM actually improves over the state-of-the-art in graph representation learning. For example, how does GraphLLM compare to a graph transformer without the attached LLM on the tested graph reasoning tasks? This could isolate the benefits of the LLM integration. Similarly, comparisons to GNN variants on standard benchmarks could reveal where GraphLLM excels or lags behind existing graph-specific architectures. The improvements over Graph2Text may simply be because LLMs struggle with learning from text descriptions of graphs. Graph models designed for relational reasoning may be more competitive. In essence, without comparing to specialized graph reasoning models, it is hard to discern if GraphLLM's integration of LLMs and graph networks is truly advancing state-of-the-art performance. Adding these comparisons would give a clearer sense of GraphLLM's capabilities relative to the field of graph representation learning overall.", "questions": "Please check the weakness section above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes GraphLLM, a new approach to enhance the ability of large language models (LLMs) to understand and reason about graph data. It identifies a key limitation of current methods that convert graphs to text (Graph2Text), which forces LLMs to implicitly learn graph structures and results in lengthy contexts. GraphLLM integrates a graph learning module with the LLM via end-to-end training. This allows the LLM to leverage the graph module's strengths in an efficient way through a condensed graph-enhanced prefix. Experiments on 4 graph reasoning tasks show GraphLLM substantially improves accuracy over Graph2Text methods while reducing context length and accelerating inference.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "* Novel integration of graphs and LLMs.\n\nGraphLLM proposes a novel end-to-end approach to combine graph learning models and LLMs. This allows each component to be optimized to complement the other for different graph reasoning tasks.\n\n* Significant performance gains.\n\nGraphLLM improves accuracy by 54.44% on average over the best Graph2Text method, showing its effectiveness. Also, GraphLLM reduces context length by 96.45% compared to Graph2Text, enhancing efficiency. Besides, GraphLLM achieves 3.42x faster inference over the best Graph2Text method due to context reduction.", "weaknesses": "* Limited graph tasks evaluated.\n\nThe paper evaluates GraphLLM on four graph reasoning tasks, including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching. Although these tasks cover basic graph reasoning abilities, they are still relatively simple, which might overestimate how GraphLLM performs on noisier graphs with complex relational patterns. More complex graph reasoning tasks could better demonstrate the capabilities and limitations of GraphLLM. For example, tasks requiring multi-hop reasoning on large graphs might be more challenging.\n\n* Lack of analysis on scalability.\n\nThe datasets used in the paper are all small with fewer than 20 nodes on average. Thus, the ability of GraphLLM to handle large, real-world graphs is unclear. Evaluations on larger datasets would be informative.\n\n* Restricted node features.\n\nThe tasks only use textual node features. Testing on graphs with non-textual node attributes would make the approach more broadly applicable.\n\n* No comparison to other graph models.\n\nThe paper only compares GraphLLM against methods that convert the graph to text (Graph2Text). However, established graph neural networks like GNNs, GCNs, and graph transformers have been optimized specifically for graph reasoning. Comparing to these dedicated graph reasoning models could better highlight if GraphLLM actually improves over the state-of-the-art in graph representation learning. For example, how does GraphLLM compare to a graph transformer without the attached LLM on the tested graph reasoning tasks? This could isolate the benefits of the LLM integration. Similarly, comparisons to GNN variants on standard benchmarks could reveal where GraphLLM excels or lags behind existing graph-specific architectures. The improvements over Graph2Text may simply be because LLMs struggle with learning from text descriptions of graphs. Graph models designed for relational reasoning may be more competitive. In essence, without comparing to specialized graph reasoning models, it is hard to discern if GraphLLM's integration of LLMs and graph networks is truly advancing state-of-the-art performance. Adding these comparisons would give a clearer sense of GraphLLM's capabilities relative to the field of graph representation learning overall.", "questions": "Please check the weakness section above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698928168241}, {"id": "PiKSgTfGv5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4455/Reviewer_B4te"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "GraphLLM introduces a new method for boosting graph reasoning abilities of LLMs. It introduces a graph encoder using a transformer encoder-decoder for encoding node text descriptions and a graph transformer for incorporating structure. These are combined to generate key value prefixes for an LLM to use. \n\nThey show results on 4 different small-size graph reasoning tasks. They show great performance on these tasks compared to the text only (graph2text) baselines. They also show a large reduction in the number of input tokens to the LLM.", "review_text": "GraphLLM introduces a new method for boosting graph reasoning abilities of LLMs. It introduces a graph encoder using a transformer encoder-decoder for encoding node text descriptions and a graph transformer for incorporating structure. These are combined to generate key value prefixes for an LLM to use. \n\nThey show results on 4 different small-size graph reasoning tasks. They show great performance on these tasks compared to the text only (graph2text) baselines. They also show a large reduction in the number of input tokens to the LLM.", "strengths": "This paper offers a clever approach to incorporating graph reasoning into LLMs. \n\n1) The architecture is smart and sensible, and the integration via prefix tuning makes this relatively simple to integrate. \n\n2) The experimental results are very strong compared to the baselines. GraphLLM essentially completely solves these tasks\n\n3) The graph transformer design is important, as evidenced by the ablation in table 7. \n\nOverall, the main strength of this paper is the architecture design. The architecture is quite smart. It can utilize node text features as well as  graph features to solve some tasks.", "weaknesses": "The main weaknesses of this paper are in its experimental evaluation. The architecture seems potentially powerful but under explored. \n\nLLMs are general purpose reasoners over text and graph2text takes advantage of that. GraphLLM is purposely built and trained for these specific graph reasoning tasks. It would be shocking if it didn't do better. As a consequence, most of the results are not \"interesting\". Graph2text is a single representation that can be used for all four tasks. The prefixes from GraphLLM are tailored to each task individually. Yes, LLM fine-tuning is a more comparable setting, but also less interesting as once we are finetuning, we could be using many other architectures (e.g. GNN). The more interesting question would be can we use GraphLLM as a general graph->prefix encoder that enables the LLM to perform different fundamental tasks better\n\nThis leads to the second weakness, the experiments do not show that GraphLLM is \"useful\". The tasks tested do not require machine learning, they can all be solved explicitly. Why would I use a complex ML architecture to replace a simple program? I think the architecture is smart and there are many avenues to demonstrate utility but these experiments do not demonstrate it. The experiments do demonstrate that GraphLLM is capable of performing graph reasoning tasks and there is value there. There would be much more value in showing more general capabilities. \n\nGPT-4 shows pretty good results on 3 of the tasks with few shot CoT. While still significantly below GraphLLM, LLMs alone may be able to learn these tasks if large enough and especially if fine-tuned or the prompt optimized. Would be interesting to see llama2-70b finetuned on these tasks.  \n\nThe comparison of input tokens does not seem to account for the text input to the node understanding encoder. In fairness, this is a much smaller encoder so it does not matter as much. \nHowever, the main issue here is that the design of graph2text seems intentionally designed to maximize the number of tokens. Example: For the max triplet sum task, why does it matter that \"[Cornelia Brooks's] petite frame and delicate features give her a dainty and ethereal presence\"? (Appedix D). In addition, the text uses the term \"connected with\" to describe edges but the question itself asks about \"friends\". I think this results in a poor showcasing of the LLM baselines. Even reading the shortest path text seems confusing as a human. Does the distance from earth matter? What does activating a wormhole mean? Why are we even dealing with \"wormholes\"? The only description of the task is \"Starting from wormhole 1, How much dark matter we’ll need at the minimum to reach Wormhole 2?\"", "questions": "Questions:\n\n\"The encoderdecoder is newly initialized and updated with the guidance of the pre-trained LLM.\" This is very confusing and it is not clear what it means? How does the pre-trained LLM provide guidance?\nIs the encoder-decoder initialized from a pre-trained model or is it only the token embeddings?\n\nShouldn't prefix tuning require an embedding for each prefix value as well as each key? i.e. L x 2K x d?\n\nWhat are the instructions for generating each node's text features? Why can't this just be done using a template?\nAre the instructions generated for each node \n\nFor task 1, is it always the same substructure being identified?\n\nFor all the tasks is the input to the LLM (after the prefix) always the same? \n\nHow is exact match defined? Does it mean exact match appears in the output somewhere? Looking at the GPT-4 failure case examples in the appendix, it is hard to tell how this metric is applied? \n\nSuggestions:\n(Forgive me if any of these were done and I missed it)\n\nExplicitly mention that in prefix tuning, the LLM parameters are fixed. \n\nI don't think $d$ is explicitly defined. I take it as the dimension of the node and graph understanding encoder/decoders.\n\nStylistically, it looks better if G, E and V have a consistent style\n\nIn the introduction, I am not sure \"inefficiencies\" is the right word choice. What is inefficient about it? \"difficulties\"?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "GraphLLM introduces a new method for boosting graph reasoning abilities of LLMs. It introduces a graph encoder using a transformer encoder-decoder for encoding node text descriptions and a graph transformer for incorporating structure. These are combined to generate key value prefixes for an LLM to use. \n\nThey show results on 4 different small-size graph reasoning tasks. They show great performance on these tasks compared to the text only (graph2text) baselines. They also show a large reduction in the number of input tokens to the LLM.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "This paper offers a clever approach to incorporating graph reasoning into LLMs. \n\n1) The architecture is smart and sensible, and the integration via prefix tuning makes this relatively simple to integrate. \n\n2) The experimental results are very strong compared to the baselines. GraphLLM essentially completely solves these tasks\n\n3) The graph transformer design is important, as evidenced by the ablation in table 7. \n\nOverall, the main strength of this paper is the architecture design. The architecture is quite smart. It can utilize node text features as well as  graph features to solve some tasks.", "weaknesses": "The main weaknesses of this paper are in its experimental evaluation. The architecture seems potentially powerful but under explored. \n\nLLMs are general purpose reasoners over text and graph2text takes advantage of that. GraphLLM is purposely built and trained for these specific graph reasoning tasks. It would be shocking if it didn't do better. As a consequence, most of the results are not \"interesting\". Graph2text is a single representation that can be used for all four tasks. The prefixes from GraphLLM are tailored to each task individually. Yes, LLM fine-tuning is a more comparable setting, but also less interesting as once we are finetuning, we could be using many other architectures (e.g. GNN). The more interesting question would be can we use GraphLLM as a general graph->prefix encoder that enables the LLM to perform different fundamental tasks better\n\nThis leads to the second weakness, the experiments do not show that GraphLLM is \"useful\". The tasks tested do not require machine learning, they can all be solved explicitly. Why would I use a complex ML architecture to replace a simple program? I think the architecture is smart and there are many avenues to demonstrate utility but these experiments do not demonstrate it. The experiments do demonstrate that GraphLLM is capable of performing graph reasoning tasks and there is value there. There would be much more value in showing more general capabilities. \n\nGPT-4 shows pretty good results on 3 of the tasks with few shot CoT. While still significantly below GraphLLM, LLMs alone may be able to learn these tasks if large enough and especially if fine-tuned or the prompt optimized. Would be interesting to see llama2-70b finetuned on these tasks.  \n\nThe comparison of input tokens does not seem to account for the text input to the node understanding encoder. In fairness, this is a much smaller encoder so it does not matter as much. \nHowever, the main issue here is that the design of graph2text seems intentionally designed to maximize the number of tokens. Example: For the max triplet sum task, why does it matter that \"[Cornelia Brooks's] petite frame and delicate features give her a dainty and ethereal presence\"? (Appedix D). In addition, the text uses the term \"connected with\" to describe edges but the question itself asks about \"friends\". I think this results in a poor showcasing of the LLM baselines. Even reading the shortest path text seems confusing as a human. Does the distance from earth matter? What does activating a wormhole mean? Why are we even dealing with \"wormholes\"? The only description of the task is \"Starting from wormhole 1, How much dark matter we’ll need at the minimum to reach Wormhole 2?\"", "questions": "Questions:\n\n\"The encoderdecoder is newly initialized and updated with the guidance of the pre-trained LLM.\" This is very confusing and it is not clear what it means? How does the pre-trained LLM provide guidance?\nIs the encoder-decoder initialized from a pre-trained model or is it only the token embeddings?\n\nShouldn't prefix tuning require an embedding for each prefix value as well as each key? i.e. L x 2K x d?\n\nWhat are the instructions for generating each node's text features? Why can't this just be done using a template?\nAre the instructions generated for each node \n\nFor task 1, is it always the same substructure being identified?\n\nFor all the tasks is the input to the LLM (after the prefix) always the same? \n\nHow is exact match defined? Does it mean exact match appears in the output somewhere? Looking at the GPT-4 failure case examples in the appendix, it is hard to tell how this metric is applied? \n\nSuggestions:\n(Forgive me if any of these were done and I missed it)\n\nExplicitly mention that in prefix tuning, the LLM parameters are fixed. \n\nI don't think $d$ is explicitly defined. I take it as the dimension of the node and graph understanding encoder/decoders.\n\nStylistically, it looks better if G, E and V have a consistent style\n\nIn the introduction, I am not sure \"inefficiencies\" is the right word choice. What is inefficient about it? \"difficulties\"?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698826653283}, {"id": "o9zTYRe6Yg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4455/Reviewer_DjB6"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors propose GraphLLM to enhance the graph reasoning ability of LLMs. Specifically, GraphLLM is an end-to-end approach that integrates LLMs with graph Transformer modules. Compared to existing graph2text approaches, the proposed method can improve accuracy and reduce context length. Experimental results on four graph reasoning tasks, including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching demonstrate the effectiveness of the proposed method.", "review_text": "In this paper, the authors propose GraphLLM to enhance the graph reasoning ability of LLMs. Specifically, GraphLLM is an end-to-end approach that integrates LLMs with graph Transformer modules. Compared to existing graph2text approaches, the proposed method can improve accuracy and reduce context length. Experimental results on four graph reasoning tasks, including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching demonstrate the effectiveness of the proposed method.", "strengths": "1. Applying LLMs to graph-related problems is a recently emerging topic with rich potential.  \n2. The proposed method can combine graph machine learning (i.e., graph Transformer) and LLMs to better capture the graph structural information.  \n3. Experiments demonstrate the efficacy of the proposed method in improving accuracy and reducing context length.  \n4. The authors have released the codes to facilitate reproducibility.", "weaknesses": "1. I would strongly suggest applying the proposed method besides simple artificial graph tasks, where off-the-shelf algorithms already exist to perfectly solve them. Experiments on more real-world tasks such as node classification, link prediction, graph classification, etc., can better demonstrate the effectiveness of the proposed method. This is especially important for a technical paper where new approaches are introduced, as opposed to the previous observation-based papers such as NLGraph and GPT4Graph.  \n2. One major advantage of using LLMs compared to GNNs is explainability. Therefore, I would recommend conducting some analyses in this aspect. For example, if the proposed method cannot produce reasons for its prediction, it is likely that the proposed method captures some spurious correlations and does not truly “solve” the problem.   \n3. I also wonder whether the authors have considered directly comparing with graph Transformers or GNNs, to see whether LLMs have advanced the field.  \n4. Based on my understanding, one major advantage of the Graph2Text pipeline is being able to utilize close-sourced LLMs such as GPT-3.5 or GPT-4 without any fine-tuning step, since we only need to input data and related information (prompts, queries, etc.) to APIs. In comparison, as the proposed method needs to be trained end-to-end, it increases the computation cost and can only operate with open-sourced LLMs (in order to obtain gradients), which limits the applicability.  \n5. All things considered, the technical contribution of the paper is okay but not too novel, considering that all three components are heavily based on the existing literature, i.e., node understanding is a standard Transformer, structure understanding is essentially equivalent to Ma et al., 2023, and prefix tuning follows the common practice of LLMs. It would make the paper stronger if these components could be customized to further enhance the performance.  \n6. Minor: the authors claim one advantage of graph Transformers over GNNs is “its decoupling of node information and structural information”. Experimental evidence could be provided to support this claim.", "questions": "See Weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors propose GraphLLM to enhance the graph reasoning ability of LLMs. Specifically, GraphLLM is an end-to-end approach that integrates LLMs with graph Transformer modules. Compared to existing graph2text approaches, the proposed method can improve accuracy and reduce context length. Experimental results on four graph reasoning tasks, including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching demonstrate the effectiveness of the proposed method.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. Applying LLMs to graph-related problems is a recently emerging topic with rich potential.  \n2. The proposed method can combine graph machine learning (i.e., graph Transformer) and LLMs to better capture the graph structural information.  \n3. Experiments demonstrate the efficacy of the proposed method in improving accuracy and reducing context length.  \n4. The authors have released the codes to facilitate reproducibility.", "weaknesses": "1. I would strongly suggest applying the proposed method besides simple artificial graph tasks, where off-the-shelf algorithms already exist to perfectly solve them. Experiments on more real-world tasks such as node classification, link prediction, graph classification, etc., can better demonstrate the effectiveness of the proposed method. This is especially important for a technical paper where new approaches are introduced, as opposed to the previous observation-based papers such as NLGraph and GPT4Graph.  \n2. One major advantage of using LLMs compared to GNNs is explainability. Therefore, I would recommend conducting some analyses in this aspect. For example, if the proposed method cannot produce reasons for its prediction, it is likely that the proposed method captures some spurious correlations and does not truly “solve” the problem.   \n3. I also wonder whether the authors have considered directly comparing with graph Transformers or GNNs, to see whether LLMs have advanced the field.  \n4. Based on my understanding, one major advantage of the Graph2Text pipeline is being able to utilize close-sourced LLMs such as GPT-3.5 or GPT-4 without any fine-tuning step, since we only need to input data and related information (prompts, queries, etc.) to APIs. In comparison, as the proposed method needs to be trained end-to-end, it increases the computation cost and can only operate with open-sourced LLMs (in order to obtain gradients), which limits the applicability.  \n5. All things considered, the technical contribution of the paper is okay but not too novel, considering that all three components are heavily based on the existing literature, i.e., node understanding is a standard Transformer, structure understanding is essentially equivalent to Ma et al., 2023, and prefix tuning follows the common practice of LLMs. It would make the paper stronger if these components could be customized to further enhance the performance.  \n6. Minor: the authors claim one advantage of graph Transformers over GNNs is “its decoupling of node information and structural information”. Experimental evidence could be provided to support this claim.", "questions": "See Weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698650801321}], "openreview_url": "https://openreview.net/forum?id=PIl69UIAWL", "arxiv_id": "2310.05845", "paper_pdf": "papers/PIl69UIAWL.pdf", "paper_pdf_sha256": "c5928f6df18b1fb0790d7b1a2e0ccf8376d103fdcddcc4d21ea92469502f4e69", "paper_pdf_bytes": 1404711, "paper_pdf_source": "openreview", "code_url": "https://github.com/mistyreed63849/Graph-LLM", "code_repository": "mistyreed63849/Graph-LLM", "code_commit": "ceeba92039dd0066eb808861cc5e3c7e2232a167", "code_archive": "repos/PIl69UIAWL.zip", "code_archive_sha256": "3f7dd940eeca6185ac1c1855651d6d861b4849d523f61415865294c1221668e2", "code_archive_bytes": 24705, "code_file_count": 15, "code_extensions": {".py": 11, ".sh": 4}, "github_disk_usage_kb": 29, "github_languages": {"Python": 92814, "Shell": 458}, "github_archived": false, "github_pushed_at": "2026-01-29T00:17:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graphllm-boosting-graph-reasoning-ability-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "usa87QW3_r9", "year": 2023, "status": "rejected", "title": "Everyone's Preference Changes Differently: Weighted Multi-Interest Retrieval Model", "authors": ["Hui Shi", "Yupeng Gu", "Yitong Zhou", "Bo Zhao", "Sicun Gao", "Jishen Zhao"], "authorids": ["~Hui_Shi3", "~Yupeng_Gu1", "~Yitong_Zhou1", "bozhao@pinterest.com", "~Sicun_Gao1", "~Jishen_Zhao1"], "authors_source": "OpenReview API", "abstract": "User embeddings (vectorized representations of a user) are essential in recommendation systems. Numerous approaches have been proposed to construct a representation for the user in order to find similar items for retrieval tasks, and they have been proven effective in industrial recommendation systems. Recently people have discovered the power of using multiple embeddings to represent a user, with the hope that each embedding represents the user's interest in a certain topic. With multi-interest representation, it's important to model the user's preference over the different topics and how the preference change with time. However, existing approaches either fail to estimate the user's affinity to each interest or unreasonably assume every interest of every user fades with an equal rate with time, thus hurting the performance of candidate retrieval. In this paper, we propose the Multi-Interest Preference (MIP) model, an approach that not only produces multi-interest for users by using the user's sequential engagement more effectively but also automatically learns a set of weights to represent the preference over each embedding so that the candidates can be retrieved from each interest proportionally. Extensive experiments have been done on various industrial-scale datasets to demonstrate the effectiveness of our approach. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "DeKLhnE5EP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper324/Reviewer_ufEd"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigates the multi-interest for user embeddings in recommender retrievers. The authors consider the different weights of interests as well as time-varying interests and integrate a multi-head attention module and a cluster strategy with their weights. The proposed MIP model is then validated using publicly available datasets.\n", "review_text": "This work focuses on an important and interesting problem. However, the general solved task is not clear where the unusual objective function and metrics are adopted. There are some misleading notions and statements which makes it very confusing. The experiments are not convincing where the pre-processing is not appropriate and important baselines are missing. \n", "strengths": "Strengths\n1. This work focuses on the encoding of user embeddings, and it is an interesting and important in recommender retrievers. \n2. Incorporating time information into the multi-interest task is intriguing because user preferences change over time.\n3. The authors run several experiments on various public datasets to demonstrate the effectiveness of the proposed method.\n\nWeakness\n+ The authors point that this work focuses on the retrieval in recommender systems. The models, however, are optimized using the binary cross entropy loss and tested using the metric AUC, NLL, which are rarely used in retrievers.\n+ In general, the sparse and dense features can be combined with the item embedding to allow the item embeddings to be treated as a whole in models. The authors create various operations and run separate experiments on these two types of features. This means that the proposed method cannot be used in cases where there are both sparse and dense features.\n+ The authors claim that \"propose a multi-interest user representation model that minimizes the bias towards popular categories\". However, \"a higher weight is assigned to an interest cluster if the user engages more with items that belong to it\". How do the authors deal with the popularity bias?\n+ In Eq 9, each user chooses one interest, which is inconsistent with the goal of multi-interest.\n+ The experiments are insufficiently convincing.\n    + The adopted metrics AUC and NLL are more appropriate for ranking models, rather than retrievers.\n    + The authors filter users with less than 100 items and thus there are only a significant limited number of users are observed. Experiments should be conducted on larger datasets. The data statistics after filtering should be more detailed. Furthermore, the authors do not provide the details about the test data.\n    + Several important baseline methods in multi-interest task are missing, such as MIND [Multi-Interest Network with Dynamic Routing for Recommendation at Tmall].\n    + In appendix A.1, the experiments are conducted on a synthetic dataset which is not convincing. \n+ The notions, such as k in 3.1 and [;] in equations, are not well defined.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper investigates the multi-interest for user embeddings in recommender retrievers. The authors consider the different weights of interests as well as time-varying interests and integrate a multi-head attention module and a cluster strategy with their weights. The proposed MIP model is then validated using publicly available datasets.\n", "strength_and_weaknesses": "Strengths\n1. This work focuses on the encoding of user embeddings, and it is an interesting and important in recommender retrievers. \n2. Incorporating time information into the multi-interest task is intriguing because user preferences change over time.\n3. The authors run several experiments on various public datasets to demonstrate the effectiveness of the proposed method.\n\nWeakness\n+ The authors point that this work focuses on the retrieval in recommender systems. The models, however, are optimized using the binary cross entropy loss and tested using the metric AUC, NLL, which are rarely used in retrievers.\n+ In general, the sparse and dense features can be combined with the item embedding to allow the item embeddings to be treated as a whole in models. The authors create various operations and run separate experiments on these two types of features. This means that the proposed method cannot be used in cases where there are both sparse and dense features.\n+ The authors claim that \"propose a multi-interest user representation model that minimizes the bias towards popular categories\". However, \"a higher weight is assigned to an interest cluster if the user engages more with items that belong to it\". How do the authors deal with the popularity bias?\n+ In Eq 9, each user chooses one interest, which is inconsistent with the goal of multi-interest.\n+ The experiments are insufficiently convincing.\n    + The adopted metrics AUC and NLL are more appropriate for ranking models, rather than retrievers.\n    + The authors filter users with less than 100 items and thus there are only a significant limited number of users are observed. Experiments should be conducted on larger datasets. The data statistics after filtering should be more detailed. Furthermore, the authors do not provide the details about the test data.\n    + Several important baseline methods in multi-interest task are missing, such as MIND [Multi-Interest Network with Dynamic Routing for Recommendation at Tmall].\n    + In appendix A.1, the experiments are conducted on a synthetic dataset which is not convincing. \n+ The notions, such as k in 3.1 and [;] in equations, are not well defined.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper should be carefully revised to improve clarify. The contribution is limited to introducing users' multiple interests to sequential retrievers. This work can be somewhat reproducible.\n", "summary_of_the_review": "This work focuses on an important and interesting problem. However, the general solved task is not clear where the unusual objective function and metrics are adopted. There are some misleading notions and statements which makes it very confusing. The experiments are not convincing where the pre-processing is not appropriate and important baselines are missing. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666938728056}, {"id": "6Ey-5zw8K0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper324/Reviewer_C5FP"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors work on the problem of generating multiple interest embedding to represent user interest across multiple topics via sequence modeling through user’s sequential actions. The main contribution compared to the previous papers is that the authors learns a set of weights to represent the preference over each cluster-level embedding so that the candidates can be retrieved from each interest proportionally. Experiment is conducted on an offline measure using both public dataset and private large scale dataset(Pinterest Data Set).\n", "review_text": "The paper studies a very novel and practical problem with a simple and clean solution. The paper is well written and easy to follow. The paper has the potential impact to inspire a lot of people working on the field. I recommend this paper to get accepted.", "strengths": "[Advantage]\n1. The paper works on the user embedding generation problem which is very practical as it is widely used as the major recommendation retrieving mechanism in most internet companies. \n\nIt is known that a single embedding may not be representative enough to capture a user's multi-dimensional interest. It makes sense to use clustering based methods to generate multiple versions of embedding to increase general embedding representation power.\n\nThe research problem is practical and can benefit others.\n\n2. The main contribution of learning cluster weight makes sense and can significantly help to improve the retrieving efficiency given the candidate quota is usually limited during the retrieval stage.  The proposed method is straightforward and technically sound.\n\n3. The experiment section is comprehensive and the overall paper is well written and easy to follow.\n\n\n[Concerns and Questions]\n\n1.With the intra-cluster Mask M being applied to user representation learning, does it mean inter-cluster information is ignored during the embedding training?\n\nFor instance, the user's sequence for TV and Movie may be clustered into 2 different clusters, but the actions across these 2 clusters can still be beneficial to each other.\n\n2. Since this work considers cluster based multi-interest embedding learning. It is important to analyze how the number of clusters impact the performance. To be specific, it needs to provide analysis on the special case when only using 1 cluster(i.e. does not consider multi-interest factors) vs considering multiple clusters.\n\n3. Is online LE conducted in Pinterest production? How is the performance? The result will be more convincing with production LE/Launch.\n\n4. Nit: In table 2, notation for vector concatenation operator should be “[;]” instead of “[:]”\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors work on the problem of generating multiple interest embedding to represent user interest across multiple topics via sequence modeling through user’s sequential actions. The main contribution compared to the previous papers is that the authors learns a set of weights to represent the preference over each cluster-level embedding so that the candidates can be retrieved from each interest proportionally. Experiment is conducted on an offline measure using both public dataset and private large scale dataset(Pinterest Data Set).\n", "strength_and_weaknesses": "[Advantage]\n1. The paper works on the user embedding generation problem which is very practical as it is widely used as the major recommendation retrieving mechanism in most internet companies. \n\nIt is known that a single embedding may not be representative enough to capture a user's multi-dimensional interest. It makes sense to use clustering based methods to generate multiple versions of embedding to increase general embedding representation power.\n\nThe research problem is practical and can benefit others.\n\n2. The main contribution of learning cluster weight makes sense and can significantly help to improve the retrieving efficiency given the candidate quota is usually limited during the retrieval stage.  The proposed method is straightforward and technically sound.\n\n3. The experiment section is comprehensive and the overall paper is well written and easy to follow.\n\n\n[Concerns and Questions]\n\n1.With the intra-cluster Mask M being applied to user representation learning, does it mean inter-cluster information is ignored during the embedding training?\n\nFor instance, the user's sequence for TV and Movie may be clustered into 2 different clusters, but the actions across these 2 clusters can still be beneficial to each other.\n\n2. Since this work considers cluster based multi-interest embedding learning. It is important to analyze how the number of clusters impact the performance. To be specific, it needs to provide analysis on the special case when only using 1 cluster(i.e. does not consider multi-interest factors) vs considering multiple clusters.\n\n3. Is online LE conducted in Pinterest production? How is the performance? The result will be more convincing with production LE/Launch.\n\n4. Nit: In table 2, notation for vector concatenation operator should be “[;]” instead of “[:]”\n", "clarity,_quality,_novelty_and_reproducibility": "The paper studies a very novel and practical problem with a simple and clean solution. The paper is well written and easy to follow.\n\nPart of the experiment is conducted on several public dataset which makes it possible to reproduce and follow the work. However, the code is not open-sourced which limits the reproducibility;.\n", "summary_of_the_review": "The paper studies a very novel and practical problem with a simple and clean solution. The paper is well written and easy to follow. The paper has the potential impact to inspire a lot of people working on the field. I recommend this paper to get accepted.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666767790934}, {"id": "n--40L8xLi", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper324/Reviewer_TbF3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a multi-interest user embedding retrieval model MIP for recommendation. The idea is to apply clustering to each user's historical-interacted item embedding sequence. Each cluster is then multiplied with a weight to discriminate their importance. \n\nAs they explicitly use clustering methods on item embeddings, the cluster assignment of each item is known. This can be used to derive the importance weights of clusters, i.e., a higher weight should be assigned to an interest cluster if the user engages more with items that belong to it.\n\nThey conduct extensive experiments on three public datasets to conclude the superior performance of MIP in contrast to other retrieval models.", "review_text": "Due to the above concerns (mostly due to the efficiency problem in W1), I would vote for rejection. But I would like to increase my score if the above questions could be addressed in the author response.\n\n---\n\nAFTER THE AUTHOR RESPONSE:\n\nThe authors' response relieved my concern about efficiency.\nThe proposed method is at least feasible on the online system although it brings more time cost. Thus I'd like to raise my score to above the acceptance threshold.", "strengths": "S1. The motivation for assigning weights to different user interests is strong and interesting, and the proposed method of deriving importance weight according to user's engagement with each cluster is novel and intuitive.\n\nS2. They conduct extensive experiments on three public datasets. Overall, the improvements of the proposed method are significant and promising.\n\n---\n\nW1. A major concern is the efficiency of MIP. For each instance, they perform clustering on the user behavior sequence, which seems to be very time-consuming during both model training and inference. The time cost of  MIP should be illustrated to show whether it could be applied to real-world recommendation systems.\n\nW2. I also have several questions about the details of MIP. Most of them are about the motivation of designs in the method.\n\nQ1. The paper discriminates two situations of item features: dense features and sparse features. Why is this needed? In real-world systems, it is common that items are related to both sparse and dense features. How can you tackle this problem?\n\nQ2. What is the benefit of performing self-attention on the user behavior sequence before clustering? What if removing the self-attention part?\n\nQ3. Why use the last item in each cluster as the cluster representation? \n\nQ4. Why do you need to input cluster representation along with engagement vector to predict cluster weight in Equation (8)?\n\nQ5. Is the method guaranteed to reach convergence considering it has an inserted clustering module?\n\n---\n\nMinor: What does the number mean in the training/test/validation rows of Table 4?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a multi-interest user embedding retrieval model MIP for recommendation. The idea is to apply clustering to each user's historical-interacted item embedding sequence. Each cluster is then multiplied with a weight to discriminate their importance. \n\nAs they explicitly use clustering methods on item embeddings, the cluster assignment of each item is known. This can be used to derive the importance weights of clusters, i.e., a higher weight should be assigned to an interest cluster if the user engages more with items that belong to it.\n\nThey conduct extensive experiments on three public datasets to conclude the superior performance of MIP in contrast to other retrieval models.", "strength_and_weaknesses": "S1. The motivation for assigning weights to different user interests is strong and interesting, and the proposed method of deriving importance weight according to user's engagement with each cluster is novel and intuitive.\n\nS2. They conduct extensive experiments on three public datasets. Overall, the improvements of the proposed method are significant and promising.\n\n---\n\nW1. A major concern is the efficiency of MIP. For each instance, they perform clustering on the user behavior sequence, which seems to be very time-consuming during both model training and inference. The time cost of  MIP should be illustrated to show whether it could be applied to real-world recommendation systems.\n\nW2. I also have several questions about the details of MIP. Most of them are about the motivation of designs in the method.\n\nQ1. The paper discriminates two situations of item features: dense features and sparse features. Why is this needed? In real-world systems, it is common that items are related to both sparse and dense features. How can you tackle this problem?\n\nQ2. What is the benefit of performing self-attention on the user behavior sequence before clustering? What if removing the self-attention part?\n\nQ3. Why use the last item in each cluster as the cluster representation? \n\nQ4. Why do you need to input cluster representation along with engagement vector to predict cluster weight in Equation (8)?\n\nQ5. Is the method guaranteed to reach convergence considering it has an inserted clustering module?\n\n---\n\nMinor: What does the number mean in the training/test/validation rows of Table 4?", "clarity,_quality,_novelty_and_reproducibility": "The work is novel, and the quality and clarity are good.", "summary_of_the_review": "Due to the above concerns (mostly due to the efficiency problem in W1), I would vote for rejection. But I would like to increase my score if the above questions could be addressed in the author response.\n\n---\n\nAFTER THE AUTHOR RESPONSE:\n\nThe authors' response relieved my concern about efficiency.\nThe proposed method is at least feasible on the online system although it brings more time cost. Thus I'd like to raise my score to above the acceptance threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666678970104}, {"id": "7qCSEa7ppQ_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper324/Reviewer_QxyP"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work presents a multi-interest retrieval model, the objective of which is to increase the performance of the retrieval stage in a standard two-stage recommendation system. Extensive experiments have been done on various large-scale datasets to show the effectiveness of the proposed approach.", "review_text": "The authors propose an effective recommendation algorithm that produces multi-interest for users and learns a set of weights to represent the preference over each embedding so that the candidates can be retrieved from each interest proportionally. Overall, the paper is readable and the contributions are significant and somewhat new.", "strengths": "Strength \n+ The introduction and its motivation are clear.\n+ The performance seems impressive.\n+ The ablation tests are reasonable.\n\nWeakness\n- The primary issue of this work regards the presentation. \n   1) Table 2 misses many notations (e.g., v, \\phi, M, etc.)\n   2) The notation of the concatenation should be [;]?\n   3) Are d and d_model the same?\n   4) Why not provide the math equation(s) of FFN() and Linear()?\n   5) p is learned from v if it is a cold start data? Are they mutually exclusive inputs?\n   6) Does the order of selected \\phi matter?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work presents a multi-interest retrieval model, the objective of which is to increase the performance of the retrieval stage in a standard two-stage recommendation system. Extensive experiments have been done on various large-scale datasets to show the effectiveness of the proposed approach.", "strength_and_weaknesses": "Strength \n+ The introduction and its motivation are clear.\n+ The performance seems impressive.\n+ The ablation tests are reasonable.\n\nWeakness\n- The primary issue of this work regards the presentation. \n   1) Table 2 misses many notations (e.g., v, \\phi, M, etc.)\n   2) The notation of the concatenation should be [;]?\n   3) Are d and d_model the same?\n   4) Why not provide the math equation(s) of FFN() and Linear()?\n   5) p is learned from v if it is a cold start data? Are they mutually exclusive inputs?\n   6) Does the order of selected \\phi matter?\n", "clarity,_quality,_novelty_and_reproducibility": "The quality of the work is OK, but the presentation of the paper can be improved. ", "summary_of_the_review": "The authors propose an effective recommendation algorithm that produces multi-interest for users and learns a set of weights to represent the preference over each embedding so that the candidates can be retrieved from each interest proportionally. Overall, the paper is readable and the contributions are significant and somewhat new.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666668484483}], "openreview_url": "https://openreview.net/forum?id=usa87QW3_r9", "arxiv_id": "2207.06652", "paper_pdf": "papers/usa87QW3_r9.pdf", "paper_pdf_sha256": "67dfc593e3fe9ef84908c0953a71dac45e1fc774012b3e40b8b9e4c990f20819", "paper_pdf_bytes": 835789, "paper_pdf_source": "openreview", "code_url": "https://github.com/shihui2010/MIP", "code_repository": "shihui2010/MIP", "code_commit": "891c071c298cc34faaabd4dc33f34e9cac9d8015", "code_archive": "repos/usa87QW3_r9.zip", "code_archive_sha256": "a75ec2c6afbf94aefbc65d20162669fb80f5a71ee2b0067946eba8466d45efc7", "code_archive_bytes": 30293, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 32, "github_languages": {"Python": 74802}, "github_archived": false, "github_pushed_at": "2023-03-19T04:53:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/every-preference-changes-differently-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kcrIligNnl", "year": 2022, "status": "rejected", "title": "Direct Molecular Conformation Generation", "authors": ["Jinhua Zhu", "Yingce Xia", "Chang Liu", "Lijun Wu", "Shufang Xie", "Wengang Zhou", "Tao Qin", "Houqiang Li", "Tie-Yan Liu"], "authorids": ["~Jinhua_Zhu1", "~Yingce_Xia1", "~Chang_Liu10", "~Lijun_Wu1", "~Shufang_Xie1", "~Wengang_Zhou1", "~Tao_Qin1", "~Houqiang_Li1", "~Tie-Yan_Liu1"], "authors_source": "OpenReview API", "abstract": "Molecular conformation generation, which is to generate 3 dimensional coordinates of all the atoms in a molecule, is an important task for bioinformatics and pharmacology. Most existing machine learning based methods first predict interatomic distances and then generate conformations based on them. This two-stage approach has a potential limitation that the predicted distances may conflict with each other, e.g., violating the triangle inequality. In this work, we propose a method that directly outputs the coordinates of atoms, so that there is no violation of constraints. The conformation generator of our method stacks multiple blocks, and each block outputs a conformation which is then refined by the following block.  We adopt the variational auto-encoder (VAE) framework and use a latent variable to generate diverse conformations. To handle the roto-translation equivariance, we adopt a loss that is invariant to rotation and translation of molecule coordinates, by computing the minimal achievable distance after any rotation and translation. Our method outperforms strong baselines on four public datasets, which shows the effectiveness of our method and the great potential of the direct approach. The code is released at \\url{https://github.com/DirectMolecularConfGen/DMCG}. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "lGmmH9xamkb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3050/Reviewer_Y6xJ"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposed a new VAE-based generative model for generating molecular conformations from graphs. The whole formulation and training objective are similar to existing VAE-based models like CVGAE, and the main contrition lies in the architecture of the decoder. The architecture borrows the idea from the recent advance of AlphaFold: the proposed decoder is composed of several identical updating blocks, where they iteratively update the node/edge embeddings as well as the coordinates in each block. And the blocks are implemented with advanced graph neural networks. \n", "review_text": "Strengths:\nThe paper is well-organized and clearly written. All content is easy to understand for me.\nThe experimental part is comprehensive. The author nicely compares the proposed method with competitive baselines on both small-scale and large-scale datasets. The results demonstrate that the model is better or comparable with state-of-the-art results with much less computational cost.\nWeaknesses:\nThe main concern of this paper is the originality. The main contribution of this paper is the decoder part, which is very similar to the “structure module” in Alphafold that iteratively updates the folding and node/edge features. However, I appreciate the efforts to extend the progress of protein folding to molecular modeling, so I also agree the adaption is non-trivial.\nA minor problem: the problem definition in Sec.2 seems not accurate enough. In your statement, it seems each graph only corresponds to a single conformation, but this is not true. Please restate it.\nI’m also interested in the design of the decoder block. I understand the training objective is invariant, however, I’m curious whether updating the coordinates with equivariant networks will further help the performance. Several recent works such as ConfGF have shown that equivariance is an important inductive bias for 3D generative models.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposed a new VAE-based generative model for generating molecular conformations from graphs. The whole formulation and training objective are similar to existing VAE-based models like CVGAE, and the main contrition lies in the architecture of the decoder. The architecture borrows the idea from the recent advance of AlphaFold: the proposed decoder is composed of several identical updating blocks, where they iteratively update the node/edge embeddings as well as the coordinates in each block. And the blocks are implemented with advanced graph neural networks. \n", "main_review": "Strengths:\nThe paper is well-organized and clearly written. All content is easy to understand for me.\nThe experimental part is comprehensive. The author nicely compares the proposed method with competitive baselines on both small-scale and large-scale datasets. The results demonstrate that the model is better or comparable with state-of-the-art results with much less computational cost.\nWeaknesses:\nThe main concern of this paper is the originality. The main contribution of this paper is the decoder part, which is very similar to the “structure module” in Alphafold that iteratively updates the folding and node/edge features. However, I appreciate the efforts to extend the progress of protein folding to molecular modeling, so I also agree the adaption is non-trivial.\nA minor problem: the problem definition in Sec.2 seems not accurate enough. In your statement, it seems each graph only corresponds to a single conformation, but this is not true. Please restate it.\nI’m also interested in the design of the decoder block. I understand the training objective is invariant, however, I’m curious whether updating the coordinates with equivariant networks will further help the performance. Several recent works such as ConfGF have shown that equivariance is an important inductive bias for 3D generative models.\n", "summary_of_the_review": "The paper proposed a VAE model for molecular conformation generation, where the decoder borrows the idea of Alphafold to iteratively refine the generated structure. Personally, the paper seems an extension of the recent progress of protein folding to the small molecule setting, but I appreciate the authors’ efforts to make it work. So currently I vote for a weak acceptance.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635901641789}, {"id": "qTsEulSoocU", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3050/Reviewer_P4XU"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThe paper presents a direct method for generating a molecular\nconformation conditional on the molecular graph.  Inspired by the\nsuccess of AlphaFold2, the authors demonstrated that one could\ndirectly write out 3D coordinates for a small molecule rather than use\nthe previous ways of generating intermediate interatomic distance\nmatrices or iterate over forces. The method displays a decent\nperformance on the standard benchmark set of QM9.  It performs better\nthan past baseline methods for larger molecules.\n", "review_text": "\nOverall, this work appears to be an advance compared to the\nLangevin-based method it mainly compares against.  It is not clear to\nme if this is simply due to a different parameter setting (e.g. a\nlarger model than a year ago works better for larger molecules), or if\nthis improvement is due to a genuine algorithmic advance, and if so,\nwhat part of the algorithm helps the larger molecules the most and why\ncouldn't the authors beat the previous model on the smaller-sized\nmolecules.  I appreciated the authors effort to document the change of\nthe results across the methods in Table 3, but it would help if they\ncould also modify their new model (or ConfGf, or both) to a\ndrastically different size.\n\nI did not understand what was the contribution for the fine-grained\nloss function used in training. Unless I misunderstand something\nimportant, the particular matching loss is not invariant to\npermutations of symmetric atoms---think of the different ways to\nenumerate the atoms in a methyl group, or replace the hydrogen in a\nmethane by methyls, or take any symmetric groups in a molecule that\ncan generate a combinatorial number of equivalent coordinates that\nrepresent the same molecule graph. So I don't understand the statement\nof the authors that this loss is zero if and only if R1 and R2\nrepresent the same molecule structure---what would the loss be if I\nrandomly permute the symmetric hydrogen identities but not change the\ncoordinates?\n\nRegarding Secion 4.3, I understand that these types of tables have\nbeen published previously, but I wanted to state that I strongly\ndisagree with the terminology of calling the average energy of the\ngenerated conformers a \"property prediction\".  I don't think that\nthere exists a single experiment that can measure this particular\nquantity, although arguably one could use a Quantum Mechanical code,\nas these authors and others have done, and calculate a number for it.\nFrom a physics perspective, the somewhat more meaningful property is\nthe Boltzmann-weighted average of the energies of the molecules---is\nthat perhaps the quantity that this benchmark is supposed to evaluate?\nFurthermore, what is the meaning of switching to an MMFF force field\nfor optimization and then use Psi4---why not directly evaluate the\nconformation with Psi4?  Finally, what are the units in the energies\nof Table 4?  If these are in Hartree (which I thought were the default\nunits for Psi4, though I'm no expert in it), then the reported errors\nare all huge to be meaningful one way or another (to the order of more\nthan 13 kcal/mol for the smallest number listed in that table); these\ntypes of enormous errors have little meaning for any conformational\nenergy comparison, as a trivial small bond-length deviation could add\nthis level of energy, whereas a wrong dihedral angle couldn't (unless\nparts of the molecule almost clashed onto itself).  Perhaps the\nauthors would want to either clarify section 4.3 further (it took a\nwhile to parse what the authors actually calculate there, and from a\nquick check of the repo, I couldn't locate the relevant code for it)\nor simply drop it altogether since it probably distracts from the rest\nof this work.\n\nI also didn't understand the first two of the three future directions:\nI am not clear what the authors mean when they suggest combining this\nmethod with flow-based or Langevin dynamics models---can they clarify\nthis point?  Importantly, what are the cases that fail in the current\nmodel---can the authors document some examples in an appendix?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "\nThe paper presents a direct method for generating a molecular\nconformation conditional on the molecular graph.  Inspired by the\nsuccess of AlphaFold2, the authors demonstrated that one could\ndirectly write out 3D coordinates for a small molecule rather than use\nthe previous ways of generating intermediate interatomic distance\nmatrices or iterate over forces. The method displays a decent\nperformance on the standard benchmark set of QM9.  It performs better\nthan past baseline methods for larger molecules.\n", "main_review": "\nOverall, this work appears to be an advance compared to the\nLangevin-based method it mainly compares against.  It is not clear to\nme if this is simply due to a different parameter setting (e.g. a\nlarger model than a year ago works better for larger molecules), or if\nthis improvement is due to a genuine algorithmic advance, and if so,\nwhat part of the algorithm helps the larger molecules the most and why\ncouldn't the authors beat the previous model on the smaller-sized\nmolecules.  I appreciated the authors effort to document the change of\nthe results across the methods in Table 3, but it would help if they\ncould also modify their new model (or ConfGf, or both) to a\ndrastically different size.\n\nI did not understand what was the contribution for the fine-grained\nloss function used in training. Unless I misunderstand something\nimportant, the particular matching loss is not invariant to\npermutations of symmetric atoms---think of the different ways to\nenumerate the atoms in a methyl group, or replace the hydrogen in a\nmethane by methyls, or take any symmetric groups in a molecule that\ncan generate a combinatorial number of equivalent coordinates that\nrepresent the same molecule graph. So I don't understand the statement\nof the authors that this loss is zero if and only if R1 and R2\nrepresent the same molecule structure---what would the loss be if I\nrandomly permute the symmetric hydrogen identities but not change the\ncoordinates?\n\nRegarding Secion 4.3, I understand that these types of tables have\nbeen published previously, but I wanted to state that I strongly\ndisagree with the terminology of calling the average energy of the\ngenerated conformers a \"property prediction\".  I don't think that\nthere exists a single experiment that can measure this particular\nquantity, although arguably one could use a Quantum Mechanical code,\nas these authors and others have done, and calculate a number for it.\nFrom a physics perspective, the somewhat more meaningful property is\nthe Boltzmann-weighted average of the energies of the molecules---is\nthat perhaps the quantity that this benchmark is supposed to evaluate?\nFurthermore, what is the meaning of switching to an MMFF force field\nfor optimization and then use Psi4---why not directly evaluate the\nconformation with Psi4?  Finally, what are the units in the energies\nof Table 4?  If these are in Hartree (which I thought were the default\nunits for Psi4, though I'm no expert in it), then the reported errors\nare all huge to be meaningful one way or another (to the order of more\nthan 13 kcal/mol for the smallest number listed in that table); these\ntypes of enormous errors have little meaning for any conformational\nenergy comparison, as a trivial small bond-length deviation could add\nthis level of energy, whereas a wrong dihedral angle couldn't (unless\nparts of the molecule almost clashed onto itself).  Perhaps the\nauthors would want to either clarify section 4.3 further (it took a\nwhile to parse what the authors actually calculate there, and from a\nquick check of the repo, I couldn't locate the relevant code for it)\nor simply drop it altogether since it probably distracts from the rest\nof this work.\n\nI also didn't understand the first two of the three future directions:\nI am not clear what the authors mean when they suggest combining this\nmethod with flow-based or Langevin dynamics models---can they clarify\nthis point?  Importantly, what are the cases that fail in the current\nmodel---can the authors document some examples in an appendix?\n\n", "summary_of_the_review": "\nI think that this paper is an advance on previous work and probably represents\nthe state of the art in generating molecular conformations conditional\non the graph.  The paper could use a review of the language to clarify\ncertain aspects of it, but I don't think that it is strong enough for ICLR.  \nImportantly, it should attempt to clarify if the improvement is mainly \ndue to model size changes or due to algorithm changes.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635893312200}, {"id": "kecKXS2vDhL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3050/Reviewer_rZ2g"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose the generative model that constructs a conformation (set of 3d coordinates) of the molecular graph with a variational autoencoder (VAE) framework. Unlike several recent works on conformation generation based on generating of distance matrix and further conformation recovering by solving a time-consuming distance geometry problem, the proposed approach directly generates 3d coordinates with an iterative process inside the decoder part of the model.", "review_text": "**Originality:** The work suffers from the lack of originality. The described model shares the same idea as ConfVAE and the unnamed model, proposed in [1] and [2] respectively - the autoencoder-based framework for conformation generation, the iterative generation procedure, and RMSD objective were proposed earlier. The main novelty of the model is neural network architectures of encoder and decoder, that utilize advanced graph convolutions. Both of [1] and [2] works are missed in baselines, the model compared with.\n\n**Clarity and Quality:** The authors clearly describe their motivation and provide good discussion and ablation studies section, still paper contains vague contributions part and missed related work section. The paper is written in a hard-to-read way. The paper contains plenty of inline equations, inconvenient formatting of enumerated statements that make it hard to navigate over text and understand the method. The paper focuses too much attention on technical details of the model, that are minor and can be moved to the Supplementary materials.\n\n**Significance:** The novelty of the paper is minor. The experiment results show the increase of metric values between the proposed model and ConfGF baseline, still, it's hard to determine the significance of the performance gap due to the lack of standard deviation. Also, the significance of the results is doubtful since the important baselines [1] and [2] are missed in Experiment section.\n\n**Drawbacks / questions**:\n\n1. I would recommend using different Greek letters to denote the encoders/decoder parts instead of subscript text to increase the readability of formulas.\n2. The ablation studies cover only architecture details. Still [2] provide loss function based on distance differences that is also roto-translation invariant but do not require additional subproblem solving.\n3. The small-scale dataset results are redundant still large-scale dataset covers it. I would recommend training CVGAE and CGCF on the large-scale dataset and including the results of these models.\n4. The choice of conformation colors in Figure 3 is misleading. I would recommend to add an additional figure with aligned conformation cloud to show the diversity of generated conformations.\n\n[1] An End-to-End Framework for Molecular Conformation Generation via Bilevel Programming, Xu et al.\n\n[2] Auto-Encoding Molecular Conformations, Winter et al.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose the generative model that constructs a conformation (set of 3d coordinates) of the molecular graph with a variational autoencoder (VAE) framework. Unlike several recent works on conformation generation based on generating of distance matrix and further conformation recovering by solving a time-consuming distance geometry problem, the proposed approach directly generates 3d coordinates with an iterative process inside the decoder part of the model.", "main_review": "**Originality:** The work suffers from the lack of originality. The described model shares the same idea as ConfVAE and the unnamed model, proposed in [1] and [2] respectively - the autoencoder-based framework for conformation generation, the iterative generation procedure, and RMSD objective were proposed earlier. The main novelty of the model is neural network architectures of encoder and decoder, that utilize advanced graph convolutions. Both of [1] and [2] works are missed in baselines, the model compared with.\n\n**Clarity and Quality:** The authors clearly describe their motivation and provide good discussion and ablation studies section, still paper contains vague contributions part and missed related work section. The paper is written in a hard-to-read way. The paper contains plenty of inline equations, inconvenient formatting of enumerated statements that make it hard to navigate over text and understand the method. The paper focuses too much attention on technical details of the model, that are minor and can be moved to the Supplementary materials.\n\n**Significance:** The novelty of the paper is minor. The experiment results show the increase of metric values between the proposed model and ConfGF baseline, still, it's hard to determine the significance of the performance gap due to the lack of standard deviation. Also, the significance of the results is doubtful since the important baselines [1] and [2] are missed in Experiment section.\n\n**Drawbacks / questions**:\n\n1. I would recommend using different Greek letters to denote the encoders/decoder parts instead of subscript text to increase the readability of formulas.\n2. The ablation studies cover only architecture details. Still [2] provide loss function based on distance differences that is also roto-translation invariant but do not require additional subproblem solving.\n3. The small-scale dataset results are redundant still large-scale dataset covers it. I would recommend training CVGAE and CGCF on the large-scale dataset and including the results of these models.\n4. The choice of conformation colors in Figure 3 is misleading. I would recommend to add an additional figure with aligned conformation cloud to show the diversity of generated conformations.\n\n[1] An End-to-End Framework for Molecular Conformation Generation via Bilevel Programming, Xu et al.\n\n[2] Auto-Encoding Molecular Conformations, Winter et al.", "summary_of_the_review": "The idea of the paper is not novel. The experiment part misses several important baselines. The text should be revised to improve readability and clarity. The ablation study section should provide experiments to justify loss function choice.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635861489047}, {"id": "3dY6BS4KnhF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3050/Reviewer_T1Ea"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors present a method that is able to accurately generate molecule confirmations, achieving competitive results across four datasets. \nUnlike previous methods that often rely on directly predicting various inter-atomic distances (through a distance matrix) (which in some cases results in physically impossible molecules), the presented work predicts the location of all atoms. The loss function is invariant to rotation, an important property of many molecules. ", "review_text": "Strengths:\n* The results are strong and the presented methodology is clear.\n* The ablations shown in Table 5 are very nice to see and I think are a useful part of the paper.\n\nWeaknesses:\n* Some of the discussion of results could be expanded, or at least included in an appendix. For example, \"This is consistent with our conjecture, and also demonstrates the potential of directly generating coordinates for complex molecules with more heavy atoms.\" This would be nice to have been further expanded upon.\n* The presented results are limited in scope, and largely rely on comparisons with ConfGF, though then some additional comparisons would be useful. For example, for the property prediction showing results comparing to Tables 3 and 4 in the ConfGF paper would be useful. \n\nComments:\n* In Fig. 3, what do the colors represent? Are two views shown?\n*  A table with exact hyper parameters/etc in the appendix could be useful (in addition to the details in section 4.1). It's great that code is provided.\n* Some methods, like G-SchNet (https://github.com/atomistic-machine-learning/G-SchNet) from my understanding do not produce confirmations which violate the triangle inequality. Is this a misunderstanding on my end, or is this approach of successively predicting distances an alternate approach that produces valid molecules?  I realise it is a different problem being approached, but molecules are still being generated. \n* In Table 2, what is your interpretation of the COV(%) being 100%? \n* For the property prediction, on QM9 various properties can be predicted. Could these all be shown?  \n* The text is at times a bit difficult to follow, but it does not detract from the ability to follow the work as a whole. \n* I don't quite follow the \"ours with /FF\" in Table 5, and why this is not the \"default\" flavour?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors present a method that is able to accurately generate molecule confirmations, achieving competitive results across four datasets. \nUnlike previous methods that often rely on directly predicting various inter-atomic distances (through a distance matrix) (which in some cases results in physically impossible molecules), the presented work predicts the location of all atoms. The loss function is invariant to rotation, an important property of many molecules. ", "main_review": "Strengths:\n* The results are strong and the presented methodology is clear.\n* The ablations shown in Table 5 are very nice to see and I think are a useful part of the paper.\n\nWeaknesses:\n* Some of the discussion of results could be expanded, or at least included in an appendix. For example, \"This is consistent with our conjecture, and also demonstrates the potential of directly generating coordinates for complex molecules with more heavy atoms.\" This would be nice to have been further expanded upon.\n* The presented results are limited in scope, and largely rely on comparisons with ConfGF, though then some additional comparisons would be useful. For example, for the property prediction showing results comparing to Tables 3 and 4 in the ConfGF paper would be useful. \n\nComments:\n* In Fig. 3, what do the colors represent? Are two views shown?\n*  A table with exact hyper parameters/etc in the appendix could be useful (in addition to the details in section 4.1). It's great that code is provided.\n* Some methods, like G-SchNet (https://github.com/atomistic-machine-learning/G-SchNet) from my understanding do not produce confirmations which violate the triangle inequality. Is this a misunderstanding on my end, or is this approach of successively predicting distances an alternate approach that produces valid molecules?  I realise it is a different problem being approached, but molecules are still being generated. \n* In Table 2, what is your interpretation of the COV(%) being 100%? \n* For the property prediction, on QM9 various properties can be predicted. Could these all be shown?  \n* The text is at times a bit difficult to follow, but it does not detract from the ability to follow the work as a whole. \n* I don't quite follow the \"ours with /FF\" in Table 5, and why this is not the \"default\" flavour?", "summary_of_the_review": "The work presents strong results on molecular confirmation prediction. A novel method is developed that mitigates concerns of many previous methods by ensuring adherence to geometric rules. The results shown are very competitive, though I would like to see slightly broader comparisons and more exposition on various points that are raised in discussion. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635816916646}], "openreview_url": "https://openreview.net/forum?id=kcrIligNnl", "arxiv_id": "2202.01356", "paper_pdf": "papers/kcrIligNnl.pdf", "paper_pdf_sha256": "72643a536ccc9d646d84a8fca47e0501f176064d4ae067db15590e91c66d11c2", "paper_pdf_bytes": 1071282, "paper_pdf_source": "openreview", "code_url": "https://github.com/DirectMolecularConfGen/DMCG", "code_repository": "DirectMolecularConfGen/DMCG", "code_commit": "23e90f7231eac9a065093ee44b378cbf65181a7c", "code_archive": "repos/kcrIligNnl.zip", "code_archive_sha256": "cb4c76c77be2da4eb8133a9d3649a05f697e2b5f839fbb804ec7fe5b19133d80", "code_archive_bytes": 33308, "code_file_count": 12, "code_extensions": {".py": 11, ".sh": 1}, "github_disk_usage_kb": 31, "github_languages": {"Python": 123858, "Dockerfile": 1829, "Shell": 1548}, "github_archived": false, "github_pushed_at": "2022-10-25T09:15:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/direct-molecular-conformation-generation-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hcCao_UYd6O", "year": 2021, "status": "rejected", "title": "Adversarial Feature Desensitization", "authors": ["Pouya Bashivan", "Mojtaba Faramarzi", "Touraj Laleh", "Blake Aaron Richards", "Irina Rish"], "authorids": ["~Pouya_Bashivan1", "~Mojtaba_Faramarzi1", "~Touraj_Laleh1", "~Blake_Aaron_Richards1", "~Irina_Rish1"], "authors_source": "OpenReview API", "abstract": "Deep neural networks can now perform many tasks that were once thought to be only feasible for humans. While reaching impressive performance under standard settings, such networks are known to be  susceptible to  adversarial attacks -- slight but carefully constructed perturbations of the inputs which drastically decrease the network performance. Here we propose a new way to improve the network robustness against adversarial attacks by   focusing on robust representation learning based on  adversarial training procedure, called here Adversarial Feature Desensitization (AFD). AFD desensitizes the representation via an adversarial game between the embedding network and an adversarial discriminator introduced on top of  the standard predictive model, which is trained to distinguish between the clean and perturbed inputs from their high-level representations. Our method substantially improves the state-of-the-art in robust classification on MNIST, CIFAR10, and CIFAR100 datasets. More importantly, we demonstrate that AFD has better generalization ability than previous methods, as the learned features maintain their robustness across a wide range of perturbations, including perturbations not seen during training. These results indicate that reducing feature sensitivity is a promising approach for ameliorating the problem of adversarial attacks in deep neural networks. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1oi1FDz7HoE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1744/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an adversarial defense method that builds a robust classifier by using a min-max optimization in the feature extractor. \n\nPros:\n1. The proposed idea is interesting and reasonable.\n2. The experiment results look good.\n\nCons:\n1. The proposed idea is not novelty enough. There is a similar work [1] talking about the feature scattering to make the model robustness.  The only difference is the paper further uses a discriminator which involves a different network in the training process.\n2. The results might not be reliable. while [1] uses a similar idea, it later shows a significant performance drop by using some later proposed adversarial attack method like Autoattack[2]. Specifically, it reduces the performance to from claimed 60.6to 36.64. I suggest the author should try more blackbox based attack or the ensemble attack in Autoattack. \n3. The paper quality needs to be improved. For example, Figures 3  doesn't have x-axis label. I personally find it very confusing on the different x scale used in the figure. \n4. The methods mentioned in non-obfuscated gradients are not sufficient for me to believe there is no such case. (ii) (iii) is quite unrelated to the obfuscated gradients problem. Also, the boundary attack in the appendix looks very confusing. I can't understand what the number represents in the table. By the way, the boundary attack is not a good way to attack the adversarial defended method. Attacks like B&B or FAB are more suitable.\n\n\n\n\n\n[1] Zhang, Haichao, and Jianyu Wang. \"Defense against adversarial attacks using feature scattering-based adversarial training.\" Advances in Neural Information Processing Systems. 2019.\n\n[2] https://github.com/fra31/auto-attack", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not enough novelty with potential problems in the evaluation.", "review": "The paper proposes an adversarial defense method that builds a robust classifier by using a min-max optimization in the feature extractor. \n\nPros:\n1. The proposed idea is interesting and reasonable.\n2. The experiment results look good.\n\nCons:\n1. The proposed idea is not novelty enough. There is a similar work [1] talking about the feature scattering to make the model robustness.  The only difference is the paper further uses a discriminator which involves a different network in the training process.\n2. The results might not be reliable. while [1] uses a similar idea, it later shows a significant performance drop by using some later proposed adversarial attack method like Autoattack[2]. Specifically, it reduces the performance to from claimed 60.6to 36.64. I suggest the author should try more blackbox based attack or the ensemble attack in Autoattack. \n3. The paper quality needs to be improved. For example, Figures 3  doesn't have x-axis label. I personally find it very confusing on the different x scale used in the figure. \n4. The methods mentioned in non-obfuscated gradients are not sufficient for me to believe there is no such case. (ii) (iii) is quite unrelated to the obfuscated gradients problem. Also, the boundary attack in the appendix looks very confusing. I can't understand what the number represents in the table. By the way, the boundary attack is not a good way to attack the adversarial defended method. Attacks like B&B or FAB are more suitable.\n\n\n\n\n\n[1] Zhang, Haichao, and Jianyu Wang. \"Defense against adversarial attacks using feature scattering-based adversarial training.\" Advances in Neural Information Processing Systems. 2019.\n\n[2] https://github.com/fra31/auto-attack", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604188963953}, {"id": "_fo4PIapVIv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1744/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to leverage an additional adversarial discriminator to distinguish between the clean and perturbed inputs from the representation level.\n\nThe empirical evaluation shows promising results in terms of generalizing to unforeseen attacks. \nHowever, the baselines TRADES is evaluated on ImageNet which is a more challenging task and it would be good to evaluate the performance on large scale dataset for the proposed method as well.\nIt would be important to compare with related work on improving robustness by learning robust representation [1][2].\nIn addition, from figure 4, the sensitivity comparison is actually not very clear and it would be good to compute the significance level to show how sensitive the AFD learned representation with quantitative results.\n\nFrom the methodology perspective, it would be good to analyze the difference between the distributions of benign and adversary representation. Several previous work shows that with only activation patterns or logits it is not enough to distinguish the benign and adversarial instances. It would be important to evaluate and confirm that the learned representation of the benign and adversarial instances can indeed be learned and separated by a trained discriminator, which may lead to further interesting analysis and findings.\n\n\n[1] Liao, Fangzhou, et al. \"Defense against adversarial attacks using high-level representation guided denoiser.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\n[2] Samangouei, Pouya, Maya Kabkab, and Rama Chellappa. \"Defense-gan: Protecting classifiers against adversarial attacks using generative models.\" ICLR.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposes adversarial feature desensitization to generate robust representation ", "review": "This paper proposes to leverage an additional adversarial discriminator to distinguish between the clean and perturbed inputs from the representation level.\n\nThe empirical evaluation shows promising results in terms of generalizing to unforeseen attacks. \nHowever, the baselines TRADES is evaluated on ImageNet which is a more challenging task and it would be good to evaluate the performance on large scale dataset for the proposed method as well.\nIt would be important to compare with related work on improving robustness by learning robust representation [1][2].\nIn addition, from figure 4, the sensitivity comparison is actually not very clear and it would be good to compute the significance level to show how sensitive the AFD learned representation with quantitative results.\n\nFrom the methodology perspective, it would be good to analyze the difference between the distributions of benign and adversary representation. Several previous work shows that with only activation patterns or logits it is not enough to distinguish the benign and adversarial instances. It would be important to evaluate and confirm that the learned representation of the benign and adversarial instances can indeed be learned and separated by a trained discriminator, which may lead to further interesting analysis and findings.\n\n\n[1] Liao, Fangzhou, et al. \"Defense against adversarial attacks using high-level representation guided denoiser.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\n[2] Samangouei, Pouya, Maya Kabkab, and Rama Chellappa. \"Defense-gan: Protecting classifiers against adversarial attacks using generative models.\" ICLR.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603951361547}, {"id": "15s1UaZWMwZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1744/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper proposes Adversarial Feature Desensitization (AFD) as a defense against adversarial examples. AFD employs a min-max adversarial learning framework where the classifier learns to encode features of both clean and adversarial images as the same distribution, thereby desensitizing adversarial features. With the aim of fooling a separate discriminator model into categorizing the classifier’s adversarial features as from clean images, the classifier is trained with the standard cross-entropy loss and adversarial loss terms. The authors showed through experiments on MNIST, CIFAR10 and CIFAR100 datasets that AFD mostly outperform previous defenses across different adversarial attacks under white- and black-box conditions.\n\nPros:\n+Strong defense performance\n+Novel idea\n\nCons:\n-No discussion on the scalability of AFD defense or results on larger dataset such as imagenet\n\nRecommendation:\nThe idea is interesting and is backed by strong empirical results. Generally, the paper is well-written and easy to follow. Given AFD employs both a GAN training component and adversarial example generation, I can imagine it to be computationally more expensive than most of the existing defenses. It would be more convincing to show experiments and discussions addressing AFD’s scalability. Apart from this point, there is no major flaw in the paper and I am inclined towards acceptance.\n\n\nOther questions and comments:\nFor the black-box attack assessment, are the adversarial examples generated from models that are trained on their respective defense, initialized with different random seeds? It would be good to mention how exactly the black-box attacks are conducted.\n \nSome prior defenses have been shown to be prone to transfer black-box attacks, e.g. adversarial examples from a TRADES model, transferred to the AFD defense. It would be more convincing that AFD is not relying on obfuscated gradients with results on this form of black-box attacks.\n \nHow much is AFD’s performance dependent on the strength of adversarial examples using during training? How many iterations were used for the L-inf attack policy used to perturb the training inputs?\n \nI believe the sentence in the Introduction is untrue: “adversarial learning has not yet been successfully applied to the problem of adversarial robustness.” Several studies using adversarial learning for adversarial robustness exist:\n\n“Improved Network Robustness with Adversary Critic” NeurIPS 2018\n\n\"A Direct Approach to Robust Deep Learning Using Adversarial Networks\" ICLR2019\n\n“What it Thinks is Important is Important: Robustness Transfers through Input Gradients” CVPR 2020”\n\n“GanDef: A GAN based Adversarial Training Defense for Neural Network Classifier” arXiv:1903.02585\n\n--Update after rebuttal--\n\nThe reviewer thanks the authors for addressing the key questions and concerns and have updated the confidence score accordingly.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review #4", "review": "Summary:\n\nThis paper proposes Adversarial Feature Desensitization (AFD) as a defense against adversarial examples. AFD employs a min-max adversarial learning framework where the classifier learns to encode features of both clean and adversarial images as the same distribution, thereby desensitizing adversarial features. With the aim of fooling a separate discriminator model into categorizing the classifier’s adversarial features as from clean images, the classifier is trained with the standard cross-entropy loss and adversarial loss terms. The authors showed through experiments on MNIST, CIFAR10 and CIFAR100 datasets that AFD mostly outperform previous defenses across different adversarial attacks under white- and black-box conditions.\n\nPros:\n+Strong defense performance\n+Novel idea\n\nCons:\n-No discussion on the scalability of AFD defense or results on larger dataset such as imagenet\n\nRecommendation:\nThe idea is interesting and is backed by strong empirical results. Generally, the paper is well-written and easy to follow. Given AFD employs both a GAN training component and adversarial example generation, I can imagine it to be computationally more expensive than most of the existing defenses. It would be more convincing to show experiments and discussions addressing AFD’s scalability. Apart from this point, there is no major flaw in the paper and I am inclined towards acceptance.\n\n\nOther questions and comments:\nFor the black-box attack assessment, are the adversarial examples generated from models that are trained on their respective defense, initialized with different random seeds? It would be good to mention how exactly the black-box attacks are conducted.\n \nSome prior defenses have been shown to be prone to transfer black-box attacks, e.g. adversarial examples from a TRADES model, transferred to the AFD defense. It would be more convincing that AFD is not relying on obfuscated gradients with results on this form of black-box attacks.\n \nHow much is AFD’s performance dependent on the strength of adversarial examples using during training? How many iterations were used for the L-inf attack policy used to perturb the training inputs?\n \nI believe the sentence in the Introduction is untrue: “adversarial learning has not yet been successfully applied to the problem of adversarial robustness.” Several studies using adversarial learning for adversarial robustness exist:\n\n“Improved Network Robustness with Adversary Critic” NeurIPS 2018\n\n\"A Direct Approach to Robust Deep Learning Using Adversarial Networks\" ICLR2019\n\n“What it Thinks is Important is Important: Robustness Transfers through Input Gradients” CVPR 2020”\n\n“GanDef: A GAN based Adversarial Training Defense for Neural Network Classifier” arXiv:1903.02585\n\n--Update after rebuttal--\n\nThe reviewer thanks the authors for addressing the key questions and concerns and have updated the confidence score accordingly.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603612295527}, {"id": "gzin1KIZA8V", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1744/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is about leveraging the GAN idea to get robust features that are insensitive to adversarial attacks. Specifically, with a shared encoder/embedding, the feature of the adversarial attack are considered as the fake sample, while the feature from the real image is the real sample; by forming an adversarial game in the feature space and by jointly considering the classification loss, the AFD method is proposed. \n\nThe main idea is reasonable. But the current manuscript may not be ready for publication.\n\nThe structure of the current manuscript should be revised substantially. For example, too many materials (like experimental results) are given in Appendix, which makes the main manuscript less self-contained; I would suggest rearranging the contents and put the important ones in the main manuscript. \n\nFigures and their captions should be revised, especially Figures 3 and 4.\n\n\nIn the last paragraph of Page 2, the authors stated that “the proposed method outperforms existing methods by a large margin. However, this might not be true, by considering the performance shown in Tables 1 and 2 and the high variance observed in Table 1 and Figure 4.\n\nIn the 5th line of the first paragraph of Section 2, the definition of l_i (x) is different from the one in Algorithm 1. \n\nIn the last paragraph of Page 3, the definition of pi(x, epsilon) is not consistent. Epsilon is missing.\n\nIn the last paragraph of Page 3, the output of Da_psi should be real and the ideal discriminator output is {0,1}.\n\nIn Algorithm 1, why the policy pi is related to Dc_phi.  \n\nTheorem 1 might not be right. A GAN matches two distributions. Besides, if E() is a linear function, delta could be a vector in the null space of the weight.\n\nMany typos exist. For example, PGD-L_{inf} in Page 5.\n\nIn the last paragraph of Page 7, the representation sensitivity Sc is not defined.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The main idea is reasonable. But the current manuscript may not be ready for publication.", "review": "The paper is about leveraging the GAN idea to get robust features that are insensitive to adversarial attacks. Specifically, with a shared encoder/embedding, the feature of the adversarial attack are considered as the fake sample, while the feature from the real image is the real sample; by forming an adversarial game in the feature space and by jointly considering the classification loss, the AFD method is proposed. \n\nThe main idea is reasonable. But the current manuscript may not be ready for publication.\n\nThe structure of the current manuscript should be revised substantially. For example, too many materials (like experimental results) are given in Appendix, which makes the main manuscript less self-contained; I would suggest rearranging the contents and put the important ones in the main manuscript. \n\nFigures and their captions should be revised, especially Figures 3 and 4.\n\n\nIn the last paragraph of Page 2, the authors stated that “the proposed method outperforms existing methods by a large margin. However, this might not be true, by considering the performance shown in Tables 1 and 2 and the high variance observed in Table 1 and Figure 4.\n\nIn the 5th line of the first paragraph of Section 2, the definition of l_i (x) is different from the one in Algorithm 1. \n\nIn the last paragraph of Page 3, the definition of pi(x, epsilon) is not consistent. Epsilon is missing.\n\nIn the last paragraph of Page 3, the output of Da_psi should be real and the ideal discriminator output is {0,1}.\n\nIn Algorithm 1, why the policy pi is related to Dc_phi.  \n\nTheorem 1 might not be right. A GAN matches two distributions. Besides, if E() is a linear function, delta could be a vector in the null space of the weight.\n\nMany typos exist. For example, PGD-L_{inf} in Page 5.\n\nIn the last paragraph of Page 7, the representation sensitivity Sc is not defined.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602678537944}], "openreview_url": "https://openreview.net/forum?id=hcCao_UYd6O", "arxiv_id": "2006.04621", "paper_pdf": "papers/hcCao_UYd6O.pdf", "paper_pdf_sha256": "035e2a462e76b0f89f6d5e4f1c75bfbb4407c0161bd1d8da4356ddbe98637a4b", "paper_pdf_bytes": 11877574, "paper_pdf_source": "openreview", "code_url": "https://github.com/BashivanLab/afd", "code_repository": "BashivanLab/afd", "code_commit": "7631421528c9b7378863a25d06bb7b4e8f62419e", "code_archive": "repos/hcCao_UYd6O.zip", "code_archive_sha256": "f0ad78dd1159ee9217477a79b46f0590357340b078b990edf3891a8d9c84705f", "code_archive_bytes": 34568, "code_file_count": 10, "code_extensions": {".py": 9, ".ipynb": 1}, "github_disk_usage_kb": 36, "github_languages": {"Python": 72082, "Jupyter Notebook": 13929}, "github_archived": false, "github_pushed_at": "2022-01-20T01:08:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-feature-desensitization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJxWl0NKPB", "year": 2020, "status": "rejected", "title": "Combining MixMatch and Active Learning for Better Accuracy with Fewer Labels", "authors": ["Shuang Song", "David Berthelot", "Afshin Rostamizadeh"], "authorids": ["shuangsong@google.com", "dberth@google.com", "rostami@google.com"], "authors_source": "OpenReview API", "abstract": "We propose using active learning based techniques to further improve the state-of-the-art semi-supervised learning MixMatch algorithm. We provide a thorough empirical evaluation of several active-learning and baseline methods, which successfully demonstrate a significant improvement on the benchmark CIFAR-10, CIFAR-100, and SVHN datasets (as much as 1.5% in absolute accuracy). \nWe also provide an empirical analysis of the cost trade-off between incrementally gathering more labeled versus unlabeled data. This analysis can be used to measure the relative value of labeled/unlabeled data at different points of the learning curve, where we find that although the incremental value of labeled data can be as much as 20x that of unlabeled, it quickly diminishes to less than 3x once more than 2,000 labeled example are observed.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyeIXNKTYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper919/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to combine active learning techniques with MixMatch for semi-supervised learning. First, they review active learning and semi-supervised learning, especially MixMatch. Instead of traditional semi-supervised learning with a fixed set of labeled examples, they incrementally grow the labeled set as the training process goes on. They consider several different choices in active learning strategies: uncertainty measure and diversification. Diversification methods are used to balance the samples in different classes and ensure diversity. The cost analysis of adding labeled vs unlabeled data looks interesting. They perform an empirical evaluation on image benchmarks and improve over MixMatch.\n\nOverall, the paper is clearly written and easy to follow. However,  I cannot recommend acceptance because\n\n1. Novelty concern. The combination of two existing techniques seems not novel enough.\n\n2. Missing important baselines in related work and experiments. In the related work on semi-supervised learning (Section 3), the authors only review MixMatch but neglect other literature, e.g.[1,2,3,4]. And semi-supervised learning has a long history and it is not restricted to recent deep learning-based approaches. A thorough review can make the approach well-placed in the literature. In experiments, the authors only compare with MixMatch. I suggest that the authors include the missing literature in the next version.\n\n3. The cost analysis is the most interesting to me. However, Figure 2(b) in Section 5.3 is weird. How can the ratio less than 0? According to the definition in Section 4.3, $L_i \\subset L_{i+1}$ and $|L_{i+1}| > |L_i|$ and similar case for $U_{i,j}$, the cost ratio should not be less than 0. I'm also confused by the explanation in Section 5.3.\n\n4. From Figure 1(b) and Table 2, we can see that on CIFAR-100, the improvement of the proposed MMA is not statistically significant, especially when the label budget is low. But on a simpler dataset CIFAR-10, MMA performs better. How does MMA perform on a more challenging task with more classes, e.g. ImageNet?\n\n5. Training time comparison. At the expense of spending more time on selecting uncertain examples and techniques like k-means clustering, MMA is slightly better than MixMatch. A comparison of training time and complexity would be better to convince me.\n\n\n\n***\nMinor:\npage 4 “Starting with from a fixed pool of n unlabeled sample”\npage 4 “A corollary question is how do various accuracy targets relate to each other?” \npage 5 “While there are there additional active learning”\npage 5 “ Let’s define cl and cu as the cost of respectively obtaining a new labeled sample and a new unlabeled sample.” --> costs\n\n\n\n***\nReferences\n[1] Temporal Ensembling for Semi-Supervised Learning, ICLR 2017.\n[2] Smooth Neighbors on Teacher Graphs for Semi-supervised Learning, CVPR 2018.\n[3] Realistic Evaluation of Semi-supervised Learning Algorithms, NeurIPS 2018.\n[4] There Are Many Consistent Explanations of Unlabeled Data: Why You Should Average, ICLR 2019.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper proposes to combine active learning techniques with MixMatch for semi-supervised learning. First, they review active learning and semi-supervised learning, especially MixMatch. Instead of traditional semi-supervised learning with a fixed set of labeled examples, they incrementally grow the labeled set as the training process goes on. They consider several different choices in active learning strategies: uncertainty measure and diversification. Diversification methods are used to balance the samples in different classes and ensure diversity. The cost analysis of adding labeled vs unlabeled data looks interesting. They perform an empirical evaluation on image benchmarks and improve over MixMatch.\n\nOverall, the paper is clearly written and easy to follow. However,  I cannot recommend acceptance because\n\n1. Novelty concern. The combination of two existing techniques seems not novel enough.\n\n2. Missing important baselines in related work and experiments. In the related work on semi-supervised learning (Section 3), the authors only review MixMatch but neglect other literature, e.g.[1,2,3,4]. And semi-supervised learning has a long history and it is not restricted to recent deep learning-based approaches. A thorough review can make the approach well-placed in the literature. In experiments, the authors only compare with MixMatch. I suggest that the authors include the missing literature in the next version.\n\n3. The cost analysis is the most interesting to me. However, Figure 2(b) in Section 5.3 is weird. How can the ratio less than 0? According to the definition in Section 4.3, $L_i \\subset L_{i+1}$ and $|L_{i+1}| > |L_i|$ and similar case for $U_{i,j}$, the cost ratio should not be less than 0. I'm also confused by the explanation in Section 5.3.\n\n4. From Figure 1(b) and Table 2, we can see that on CIFAR-100, the improvement of the proposed MMA is not statistically significant, especially when the label budget is low. But on a simpler dataset CIFAR-10, MMA performs better. How does MMA perform on a more challenging task with more classes, e.g. ImageNet?\n\n5. Training time comparison. At the expense of spending more time on selecting uncertain examples and techniques like k-means clustering, MMA is slightly better than MixMatch. A comparison of training time and complexity would be better to convince me.\n\n\n\n***\nMinor:\npage 4 “Starting with from a fixed pool of n unlabeled sample”\npage 4 “A corollary question is how do various accuracy targets relate to each other?” \npage 5 “While there are there additional active learning”\npage 5 “ Let’s define cl and cu as the cost of respectively obtaining a new labeled sample and a new unlabeled sample.” --> costs\n\n\n\n***\nReferences\n[1] Temporal Ensembling for Semi-Supervised Learning, ICLR 2017.\n[2] Smooth Neighbors on Teacher Graphs for Semi-supervised Learning, CVPR 2018.\n[3] Realistic Evaluation of Semi-supervised Learning Algorithms, NeurIPS 2018.\n[4] There Are Many Consistent Explanations of Unlabeled Data: Why You Should Average, ICLR 2019.\n\n"}, "tcdate": 1571816477865}, {"id": "BkeRfPx2Kr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper919/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summarize the paper:\n\nThis paper proposes a method that can deal with an active-learning scenario for the recently proposed semi-supervised learning method: MixMatch.  More specifically, the proposed method considers uncertainty measures to choose samples and a diversification step to ensure diversity within the sampled batch.  For uncertainty measures, the paper considers the simple maximum confidence and the gap between two most likely classes.  Additional augmentation techniques inspired from MixMatch are used.  For diversification, a clustering method and an information density method are considered.  Furthermore, the paper proposes a cost analysis model to compare labeled and unlabeled samples.  Experiments demonstrate the behavior of the proposed method.\n\n\nPros of the paper:\n\n- The experimental results seem to be strong and encouraging.\n- The discussions on the cost of labeled and unlabeled samples seems to be an important contribution for semi-supervised active learning.\n- The motivation and direction of the paper is simple and easy to follow: Take the state-of-the-art semi-supervised learning algorithm and propose an active-learning version of it.\n- It is not a straightforward combination of MixMatch and active learning, and there are some specialized techniques such as “aug” used in the design of the proposed algorithm.\n\n\nCons of the paper:\n\n- Only uncertainty based sampling methods are considered, but is this enough?  There seems to be no other papers that deal with active semi-supervised learning for a deep learning context, so it might be important to really explore the many sampling methods (e.g., from survey of Settles 2009).\n\n- A more minor comment: The same issue goes for the semi-supervised learning side.  MixMatch is the state of the art in terms of accuracy for image domains, but it is an ensemble of several semi-supervised learning methods, and have strong assumptions, e.g., smoothness assumption, small distribution overlap, etc.  This will mean the proposed method will also have those strong assumptions and limits the method’s applicability.\n\n- In experiments, it would be better to have figures that are usually used in active learning experiments, where the x-axis is the remaining budget and y-axis is the performance measure.\n\n\nAdditional comments:\n\nActive learning methods gives labels to unlabeled samples in different epochs until the budget is used up, but it would be interesting to give the final labeled and unlabeled dataset after budget is used up as a fixed dataset, and then train the traditional passive MixMatch with this.  Then we can really compare the original MixMatch and active MixMatch.  If the proposed method still works better,  then the proposed method might be meaningful not only as an active learning method but also as a curriculum learning method.\n\n********************\nI would like to thank the authors for answering my questions and updating the paper, but would like to keep my score due to the 2nd point of the cons.  A minor comment on the second point of the author response:  The sharpening step in MixMatch can be regarded as an entropy minimization procedure, which I think is based on the assumption of low distribution overlap.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "Summarize the paper:\n\nThis paper proposes a method that can deal with an active-learning scenario for the recently proposed semi-supervised learning method: MixMatch.  More specifically, the proposed method considers uncertainty measures to choose samples and a diversification step to ensure diversity within the sampled batch.  For uncertainty measures, the paper considers the simple maximum confidence and the gap between two most likely classes.  Additional augmentation techniques inspired from MixMatch are used.  For diversification, a clustering method and an information density method are considered.  Furthermore, the paper proposes a cost analysis model to compare labeled and unlabeled samples.  Experiments demonstrate the behavior of the proposed method.\n\n\nPros of the paper:\n\n- The experimental results seem to be strong and encouraging.\n- The discussions on the cost of labeled and unlabeled samples seems to be an important contribution for semi-supervised active learning.\n- The motivation and direction of the paper is simple and easy to follow: Take the state-of-the-art semi-supervised learning algorithm and propose an active-learning version of it.\n- It is not a straightforward combination of MixMatch and active learning, and there are some specialized techniques such as “aug” used in the design of the proposed algorithm.\n\n\nCons of the paper:\n\n- Only uncertainty based sampling methods are considered, but is this enough?  There seems to be no other papers that deal with active semi-supervised learning for a deep learning context, so it might be important to really explore the many sampling methods (e.g., from survey of Settles 2009).\n\n- A more minor comment: The same issue goes for the semi-supervised learning side.  MixMatch is the state of the art in terms of accuracy for image domains, but it is an ensemble of several semi-supervised learning methods, and have strong assumptions, e.g., smoothness assumption, small distribution overlap, etc.  This will mean the proposed method will also have those strong assumptions and limits the method’s applicability.\n\n- In experiments, it would be better to have figures that are usually used in active learning experiments, where the x-axis is the remaining budget and y-axis is the performance measure.\n\n\nAdditional comments:\n\nActive learning methods gives labels to unlabeled samples in different epochs until the budget is used up, but it would be interesting to give the final labeled and unlabeled dataset after budget is used up as a fixed dataset, and then train the traditional passive MixMatch with this.  Then we can really compare the original MixMatch and active MixMatch.  If the proposed method still works better,  then the proposed method might be meaningful not only as an active learning method but also as a curriculum learning method.\n\n********************\nI would like to thank the authors for answering my questions and updating the paper, but would like to keep my score due to the 2nd point of the cons.  A minor comment on the second point of the author response:  The sharpening step in MixMatch can be regarded as an entropy minimization procedure, which I think is based on the assumption of low distribution overlap.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571714838081}, {"id": "BJedOIhjKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper919/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper takes a look at using active learning techniques instead of random sampling for the \n\"state-of-the-art\" semi-supervised learning (SSL) method MixMatch. At least for this one SSL algorithm, the authors give a strong argument that active learning helps MixMatch (4x label efficiency in some cases) and highlight the active learning algorithms that work best. An additionally interesting point is the value of labeled vs unlabeled data in this setting. For these reasons, I argue for acceptance of this paper. However, I have some reservations which are given below, that perhaps the authors can clarify.  \n\n\nThings that would have improved my score:\n\n - This paper relies on MixMatch very heavily as the sole semi-supervised technique. It would be nice to see more of an argument for this choice of a relatively recent paper that hasn't stood the test of time.\n\n - I am confused why the authors \"report the median of the last 20 checkpoints' accuracy where a checkpoint is computed every 65,536 training iterations\". As we see later, the authors train after each batch of selected examples for 32,768 iterations which is half of the time between checkpoints. Can the authors comment on this choice?\n\n\nMinor comments:\n\n - The end of the abstract makes it sound like like the conclusions are universal (\"quickly diminishes to less than 3x once more than 2000 labeled example are observed\"). I would be surprised if the authors meant this as a universal statement since no theory is provided and the experiments are on similar datasets.\n\n - I don't understand why section 3 is not simply titled \"MixMatch\" since the paper doesn't really touch any other \"modern SSL\" methods.\n\n - In the experiments section, 262144 and 32768 iterations seem to come from nowhere. Only later did I realize that these were powers of 2. Can this be clarified?\n\n - What's the difference between MixMatch and MMA with random selection? Shouldn't these perform the same (which they seem to anyways)?\n\n - I really like that this paper assesses the label efficiency of their algorithm, rather than merely reporting raw accuracy numbers which aren't as meaningful.\n\n - I wonder if MMA seems to not give as big gains on CIFAR-100 because the batch size is 10x larger. Generally, I've found that active learning (especially uncertainty sampling methods) work best with smaller batch sizes and I'm not sure I agree with the reasoning that more classes mean one should select larger batch sizes.\n\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper takes a look at using active learning techniques instead of random sampling for the \n\"state-of-the-art\" semi-supervised learning (SSL) method MixMatch. At least for this one SSL algorithm, the authors give a strong argument that active learning helps MixMatch (4x label efficiency in some cases) and highlight the active learning algorithms that work best. An additionally interesting point is the value of labeled vs unlabeled data in this setting. For these reasons, I argue for acceptance of this paper. However, I have some reservations which are given below, that perhaps the authors can clarify.  \n\n\nThings that would have improved my score:\n\n - This paper relies on MixMatch very heavily as the sole semi-supervised technique. It would be nice to see more of an argument for this choice of a relatively recent paper that hasn't stood the test of time.\n\n - I am confused why the authors \"report the median of the last 20 checkpoints' accuracy where a checkpoint is computed every 65,536 training iterations\". As we see later, the authors train after each batch of selected examples for 32,768 iterations which is half of the time between checkpoints. Can the authors comment on this choice?\n\n\nMinor comments:\n\n - The end of the abstract makes it sound like like the conclusions are universal (\"quickly diminishes to less than 3x once more than 2000 labeled example are observed\"). I would be surprised if the authors meant this as a universal statement since no theory is provided and the experiments are on similar datasets.\n\n - I don't understand why section 3 is not simply titled \"MixMatch\" since the paper doesn't really touch any other \"modern SSL\" methods.\n\n - In the experiments section, 262144 and 32768 iterations seem to come from nowhere. Only later did I realize that these were powers of 2. Can this be clarified?\n\n - What's the difference between MixMatch and MMA with random selection? Shouldn't these perform the same (which they seem to anyways)?\n\n - I really like that this paper assesses the label efficiency of their algorithm, rather than merely reporting raw accuracy numbers which aren't as meaningful.\n\n - I wonder if MMA seems to not give as big gains on CIFAR-100 because the batch size is 10x larger. Generally, I've found that active learning (especially uncertainty sampling methods) work best with smaller batch sizes and I'm not sure I agree with the reasoning that more classes mean one should select larger batch sizes.\n\n\n\n\n\n\n"}, "tcdate": 1571698287968}], "openreview_url": "https://openreview.net/forum?id=HJxWl0NKPB", "arxiv_id": "1912.00594", "paper_pdf": "papers/HJxWl0NKPB.pdf", "paper_pdf_sha256": "ee5ed0a00519ee232b60fa2f28551565b5d8a4b81207e198945aa3b7d65ac3a8", "paper_pdf_bytes": 370563, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/mma", "code_repository": "google-research/mma", "code_commit": "7d08be0837869db83048168be0a6a4cff0f0fcc6", "code_archive": "repos/HJxWl0NKPB.zip", "code_archive_sha256": "bb4ef8e628ea16b3d3d7e505bcfaf02bb10b913506c03ed5e7bddc669b9c49c3", "code_archive_bytes": 44956, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 42, "github_languages": {"Python": 80004}, "github_archived": true, "github_pushed_at": "2023-03-24T22:13:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/combining-mixmatch-and-active-learning-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1esnoAqt7", "year": 2019, "status": "rejected", "title": "Morpho-MNIST: Quantitative Assessment and Diagnostics for Representation Learning", "authors": ["Daniel C. Castro", "Jeremy Tan", "Bernhard Kainz", "Ender Konukoglu", "Ben Glocker"], "authorids": ["d.coelho-de-castro15@imperial.ac.uk", "j.tan17@imperial.ac.uk", "b.kainz@imperial.ac.uk", "kender@vision.ee.ethz.ch", "b.glocker@imperial.ac.uk"], "authors_source": "OpenReview API", "abstract": "Revealing latent structure in data is an active field of research, having introduced exciting technologies such as variational autoencoders and adversarial networks, and is essential to push machine learning towards unsupervised knowledge discovery. However, a major challenge is the lack of suitable benchmarks for an objective and quantitative evaluation of learned representations. To address this issue we introduce Morpho-MNIST, a framework that aims to answer: \"to what extent has my model learned to represent specific factors of variation in the data?\" We extend the popular MNIST dataset by adding a morphometric analysis enabling quantitative comparison of trained models, identification of the roles of latent variables, and characterisation of sample diversity. We further propose a set of quantifiable perturbations to assess the performance of unsupervised and supervised methods on challenging tasks such as outlier detection and domain adaptation.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SJghAtFWaX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper742/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper discusses the problem of evaluating and diagnosing the representations learnt using a generative model. This is a very important and necessary problem.\n\nHowever, this paper lacks in terms of experimental evaluation and has some technical flaws.\n1. Morphological properties deals with only the \"shape\" properties of the image object. However, when the entire image is subject to the generative model, it learns multiple properties from the image apart from shape too - such as texture and color. Additionally, there are lot of low level pixel relations that the model learns to fit the distribution of the given images. However, here the authors have assumed that the latent space of the generative models are influenced only by the morphological properties of the image - which is wrong. Latent space features could be affected by the color or texture of the image as well.\n\n2. Extracting morphological properties of the image is straight-foward for MNIST kind of objects. However, it becomes really difficult for other datasets such as CIFAR or some real world images. Studying the properties of a generative model on such datasets is very challenging and the authors have not added a discussion around that. \n\n3. Now assuming that my GAN model has learnt good representation in Morpho-MNIST dataset, is it guaranteed to learn good representations in other datasets as well? There is no guarantee on generalizability or extensibility of the work. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review: Morpho-MNIST: Quantitative Assessment and Diagnostics for Representation Learning ", "review": "This paper discusses the problem of evaluating and diagnosing the representations learnt using a generative model. This is a very important and necessary problem.\n\nHowever, this paper lacks in terms of experimental evaluation and has some technical flaws.\n1. Morphological properties deals with only the \"shape\" properties of the image object. However, when the entire image is subject to the generative model, it learns multiple properties from the image apart from shape too - such as texture and color. Additionally, there are lot of low level pixel relations that the model learns to fit the distribution of the given images. However, here the authors have assumed that the latent space of the generative models are influenced only by the morphological properties of the image - which is wrong. Latent space features could be affected by the color or texture of the image as well.\n\n2. Extracting morphological properties of the image is straight-foward for MNIST kind of objects. However, it becomes really difficult for other datasets such as CIFAR or some real world images. Studying the properties of a generative model on such datasets is very challenging and the authors have not added a discussion around that. \n\n3. Now assuming that my GAN model has learnt good representation in Morpho-MNIST dataset, is it guaranteed to learn good representations in other datasets as well? There is no guarantee on generalizability or extensibility of the work. ", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541671380358}, {"id": "Hygjmwnqhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper742/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Authors present a set of criteria to categorize MNISt digists (e.g. slant, stroke length, ..) and a set of interesting perturbations (swelling, fractures, ...) to modify MNIST dataset. They suggest analysing performance of generative models based on these tools. By extracting this kind of features, they effectively decrease the dimmension of  data. Therefore, statistically comparing the distribution of generated vs test data and binning the generated data is now possible. They perform a thorough study regarding MNIST. Their tools are a handy addition to the analytical surveys in several applications (e.g. how classification fails), but not convincingly for generation. \n\nSince their method is manually designed for MNIST, the manuscript would benefit from a justification or discussion on the  common pitfalls and the correlation between MNIST generation and more complex natural image generation tasks. Since the presented metrics do not show a significant difference between the VAE and Vanilla GAN model, the question remains whether evaluating on MNIST is a good proxy for the performance of the model on colored images with backgrounds or not. For example sharpness and attending to details is not typically a challenge in MNIST generation where in other datasets this is usually the first challenge to be addressed. I'm not convinced that ability of a model in disentangling thickness correlates to their ability in natural image generation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting characterisation and extension of MNIST ", "review": "Authors present a set of criteria to categorize MNISt digists (e.g. slant, stroke length, ..) and a set of interesting perturbations (swelling, fractures, ...) to modify MNIST dataset. They suggest analysing performance of generative models based on these tools. By extracting this kind of features, they effectively decrease the dimmension of  data. Therefore, statistically comparing the distribution of generated vs test data and binning the generated data is now possible. They perform a thorough study regarding MNIST. Their tools are a handy addition to the analytical surveys in several applications (e.g. how classification fails), but not convincingly for generation. \n\nSince their method is manually designed for MNIST, the manuscript would benefit from a justification or discussion on the  common pitfalls and the correlation between MNIST generation and more complex natural image generation tasks. Since the presented metrics do not show a significant difference between the VAE and Vanilla GAN model, the question remains whether evaluating on MNIST is a good proxy for the performance of the model on colored images with backgrounds or not. For example sharpness and attending to details is not typically a challenge in MNIST generation where in other datasets this is usually the first challenge to be addressed. I'm not convinced that ability of a model in disentangling thickness correlates to their ability in natural image generation.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541224227436}, {"id": "S1etazq92Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper742/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The author proposed an extended version of MNIS where they introduced thickening/thinning/swelling/fracture. The operation is done using binary morphological operations.\n\n* Providing benchmark data for tasks such disentanglement is important but I am not sure generating data is sufficient contribution for a paper. \n* I am not sure what conclusion I should draw from Fig 5 and Fig 6 about the data.\n* Eventually this data can become a benchmark data when it is paired with a method. Then that method/data are a benchmark.\n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "not enough contribution", "review": "The author proposed an extended version of MNIS where they introduced thickening/thinning/swelling/fracture. The operation is done using binary morphological operations.\n\n* Providing benchmark data for tasks such disentanglement is important but I am not sure generating data is sufficient contribution for a paper. \n* I am not sure what conclusion I should draw from Fig 5 and Fig 6 about the data.\n* Eventually this data can become a benchmark data when it is paired with a method. Then that method/data are a benchmark.\n\n ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541214913481}], "openreview_url": "https://openreview.net/forum?id=r1esnoAqt7", "arxiv_id": "1809.10780", "paper_pdf": "papers/r1esnoAqt7.pdf", "paper_pdf_sha256": "68a5af606e5b1da44f70f70a0bd5a25a291b3d6151bb3f8822e98d5284400166", "paper_pdf_bytes": 3645792, "paper_pdf_source": "openreview", "code_url": "https://github.com/dccastro/Morpho-MNIST", "code_repository": "dccastro/Morpho-MNIST", "code_commit": "3a20ed3b4a9b80ed0e01d6bf7d9f2b8b896a5ffe", "code_archive": "repos/r1esnoAqt7.zip", "code_archive_sha256": "d4d6936c394dc2a50681fc5bf50fd5ca9cae61b179c039adcb0d7d2bf9cc3480", "code_archive_bytes": 96149, "code_file_count": 40, "code_extensions": {".py": 40}, "github_disk_usage_kb": 168, "github_languages": {"Python": 139927}, "github_archived": false, "github_pushed_at": "2025-01-23T17:14:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/morpho-mnist-quantitative-assessment-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJ6anzb0Z", "year": 2018, "status": "rejected", "title": "Multimodal Sentiment Analysis To Explore the Structure of Emotions", "authors": ["Anthony Hu", "Seth Flaxman"], "authorids": ["anthony.hu@stats.ox.ac.uk", "s.flaxman@imperial.ac.uk"], "authors_source": "OpenReview API", "abstract": "We propose a novel approach to multimodal sentiment analysis using deep neural\nnetworks combining visual recognition and natural language processing. Our\ngoal is different than the standard sentiment analysis goal of predicting whether\na sentence expresses positive or negative sentiment; instead, we aim to infer the\nlatent emotional state of the user. Thus, we focus on predicting the emotion word\ntags attached by users to their Tumblr posts, treating these as “self-reported emotions.”\nWe demonstrate that our multimodal model combining both text and image\nfeatures outperforms separate models based solely on either images or text. Our\nmodel’s results are interpretable, automatically yielding sensible word lists associated\nwith emotions. We explore the structure of emotions implied by our model\nand compare it to what has been posited in the psychology literature, and validate\nour model on a set of images that have been used in psychology studies. Finally,\nour work also provides a useful tool for the growing academic study of images—\nboth photographs and memes—on social networks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BJ2J7pFgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper978/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a method for classifying Tumblr posts with associated images according to associated single emotion word hashtags.  The method relies on sentiment pre-processing from GloVe and image pre-processing from Inception.  \n \nMy strongest criticism for this paper is against the claim that Tumblr post represent self-reported emotions and that this method sheds new insight on emotion representation and my secondary criticism is a lack of novelty in the method, which seems to be simply a combination of previously published sentiment analysis module and previously published image analysis module, fused in an output layer. \n\nThe authors claim that the hashtags represent self-reported emotions, but this is not true in the way that psychologists query participants regarding emotion words in psychology studies.  Instead these are emotion words that a person chooses to broadcast along with an associated announcement.  As the authors point out, hashtags and words may be used sarcastically or in different ways from what is understood in emotion theory.  It is quite common for everyday people to use emotion words this way e.g. using #love to express strong approval rather than an actual feeling of love.   \n\nIn their analysis the authors claim:\n“The 15 emotions retained were those with high relative frequencies on Tumblr among the PANAS-X scale (Watson & Clark, 1999)”.\nHowever five of the words the authors retain: bored, annoyed, love, optimistic, and pensive are not in fact found in the PANAS-X scale:\n\nReference: The PANAS-X Scale: https://wiki.aalto.fi/download/attachments/50102838/PANAS-X-scale_spec.pdf Also the longer version that the authors cited: \nhttps://www2.psychology.uiowa.edu/faculty/clark/panas-x.pdf\n\nIt should also be noted that the PANAS (Positive and Negative Affect Scale) scale and the PANAS-X (the “X” is for eXtended) scale are questionnaires used to elicit from participants feelings of positive and negative affect, they are not collections of \"core\" emotion words, but rather words that are colloquially attached to either positive or negative sentiment.  For example PANAS-X includes words like:“strong” ,“active”, “healthy”, “sleepy” which are not considered emotion words by psychology.  \n\nIf the authors stated goal is \"different than the standard sentiment analysis goal of predicting whether a sentence expresses positive or negative sentiment\" they should be aware that this is exactly what PANAS is designed to do - not to infer the latent emotional state of a person, except to the extent that their affect is positive or negative.\n\n\nThe work of representing emotions had been an field in psychology for over a hundred years and it is still continuing.  https://en.wikipedia.org/wiki/Contrasting_and_categorization_of_emotions.\n\nOne of the most popular theories of emotion is the theory that there exist “basic” emotions: Anger, Disgust, Fear, Happiness (enjoyment), Sadness and Surprise (Paul Ekman, cited by the authors).  These are short duration sates lasting only seconds.  They are also fairly specific, for example “surprise” is sudden reaction to something unexpected, which is it exactly the same as seeing a flower on your car and expressing “what a nice surprise.”  The surprise would be the initial reaction of “what’s that on my car?  Is it dangerous?” but after identifying the object as non-threatening, the emotion of “surprise” would likely pass and be replaced with appreciation.  \n\nThe Circumplex Model of Emotions (Posner et al 2005) the authors refer to actually stands in opposition to the theories of Ekman.  From the cited paper by Posner et al : \n\"The circumplex model of affect proposes that all affective states arise from cognitive interpretations of core neural sensations that are the product of two independent neurophysiological systems. This model stands in contrast to theories of basic emotions, which posit that a discrete and independent neural system subserves every emotion.\"\nFrom my reading of this paper, it is clear to me that the authors do not have a clear understanding of the current state of psychology’s view of emotion representation and this work would not likely contribute to a new understanding of the latent structure of peoples’ emotions.\n\nIn the PCA result, it is not \"clear\" that the first axis represents valence, as \"sad\" has a slight positive on this scale and \"sad\" is one of the emotions most clearly associated with negative valence.\n\nWith respect to the rest of the paper, the level of novelty and impact is \"ok, but not good enough.\"  This analysis does not seem very different from Twitter analysis, because although Tumblr posts are allowed to be longer than Twitter posts, the authors truncate the posts to 50 characters.  Additionally, the images do not seem to add very much to the classification.  The authors algorithm also seems to be essentially a combination of two other, previously published algorithms.\n\nFor me the novelty of this paper was in its application to the realm of emotion theory, but I do not feel there is a contribution here.  This paper is more about classifying Tumblr posts according to emotion word hashtags than a paper that generates a new insights into emotion representation or that can infer latent emotional state. \n\n\n\n\n\n\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Hashtag classification of Tumblr Posts ", "rating": "4: Ok but not good enough - rejection", "review": "This paper presents a method for classifying Tumblr posts with associated images according to associated single emotion word hashtags.  The method relies on sentiment pre-processing from GloVe and image pre-processing from Inception.  \n \nMy strongest criticism for this paper is against the claim that Tumblr post represent self-reported emotions and that this method sheds new insight on emotion representation and my secondary criticism is a lack of novelty in the method, which seems to be simply a combination of previously published sentiment analysis module and previously published image analysis module, fused in an output layer. \n\nThe authors claim that the hashtags represent self-reported emotions, but this is not true in the way that psychologists query participants regarding emotion words in psychology studies.  Instead these are emotion words that a person chooses to broadcast along with an associated announcement.  As the authors point out, hashtags and words may be used sarcastically or in different ways from what is understood in emotion theory.  It is quite common for everyday people to use emotion words this way e.g. using #love to express strong approval rather than an actual feeling of love.   \n\nIn their analysis the authors claim:\n“The 15 emotions retained were those with high relative frequencies on Tumblr among the PANAS-X scale (Watson & Clark, 1999)”.\nHowever five of the words the authors retain: bored, annoyed, love, optimistic, and pensive are not in fact found in the PANAS-X scale:\n\nReference: The PANAS-X Scale: https://wiki.aalto.fi/download/attachments/50102838/PANAS-X-scale_spec.pdf Also the longer version that the authors cited: \nhttps://www2.psychology.uiowa.edu/faculty/clark/panas-x.pdf\n\nIt should also be noted that the PANAS (Positive and Negative Affect Scale) scale and the PANAS-X (the “X” is for eXtended) scale are questionnaires used to elicit from participants feelings of positive and negative affect, they are not collections of \"core\" emotion words, but rather words that are colloquially attached to either positive or negative sentiment.  For example PANAS-X includes words like:“strong” ,“active”, “healthy”, “sleepy” which are not considered emotion words by psychology.  \n\nIf the authors stated goal is \"different than the standard sentiment analysis goal of predicting whether a sentence expresses positive or negative sentiment\" they should be aware that this is exactly what PANAS is designed to do - not to infer the latent emotional state of a person, except to the extent that their affect is positive or negative.\n\n\nThe work of representing emotions had been an field in psychology for over a hundred years and it is still continuing.  https://en.wikipedia.org/wiki/Contrasting_and_categorization_of_emotions.\n\nOne of the most popular theories of emotion is the theory that there exist “basic” emotions: Anger, Disgust, Fear, Happiness (enjoyment), Sadness and Surprise (Paul Ekman, cited by the authors).  These are short duration sates lasting only seconds.  They are also fairly specific, for example “surprise” is sudden reaction to something unexpected, which is it exactly the same as seeing a flower on your car and expressing “what a nice surprise.”  The surprise would be the initial reaction of “what’s that on my car?  Is it dangerous?” but after identifying the object as non-threatening, the emotion of “surprise” would likely pass and be replaced with appreciation.  \n\nThe Circumplex Model of Emotions (Posner et al 2005) the authors refer to actually stands in opposition to the theories of Ekman.  From the cited paper by Posner et al : \n\"The circumplex model of affect proposes that all affective states arise from cognitive interpretations of core neural sensations that are the product of two independent neurophysiological systems. This model stands in contrast to theories of basic emotions, which posit that a discrete and independent neural system subserves every emotion.\"\nFrom my reading of this paper, it is clear to me that the authors do not have a clear understanding of the current state of psychology’s view of emotion representation and this work would not likely contribute to a new understanding of the latent structure of peoples’ emotions.\n\nIn the PCA result, it is not \"clear\" that the first axis represents valence, as \"sad\" has a slight positive on this scale and \"sad\" is one of the emotions most clearly associated with negative valence.\n\nWith respect to the rest of the paper, the level of novelty and impact is \"ok, but not good enough.\"  This analysis does not seem very different from Twitter analysis, because although Tumblr posts are allowed to be longer than Twitter posts, the authors truncate the posts to 50 characters.  Additionally, the images do not seem to add very much to the classification.  The authors algorithm also seems to be essentially a combination of two other, previously published algorithms.\n\nFor me the novelty of this paper was in its application to the realm of emotion theory, but I do not feel there is a contribution here.  This paper is more about classifying Tumblr posts according to emotion word hashtags than a paper that generates a new insights into emotion representation or that can infer latent emotional state. \n\n\n\n\n\n\n\n\n\n\n\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511801572341}, {"id": "rJA29bLxf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper978/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe authors present a study that aims at inferring the \"emotional\" tags provided by Thumblr users starting from images and texts in the captions. For text processing the authors use a standard LSTM taking as input GLOVE vectors of words in a sentence. For visual information, authors use a pretrained CNN (with fine tuning). A fully connected layer is used to fuse the multimodal information. Experimental results are reported in a self generated data set. \n\nThe contribution from the RL perspective is limited, in the sense that the authors simply applied standard models to predict a bunch of labels (in this case, emotion labels). It is interesting the \"psychological\" analysis that the authors present in Section 6. Still, I think the contribution in that part is a: sentiment-psychologically inspired analysis of the Thumbrl data set. \n\nI think the author's statement on that this study leads to a more plausible psychological model of emotion is not well founded (they also mention to learn to recognize the latent emotional state). Whereas it is true that psychological studies rely on self - filled questionnaires, comparing a questionnaire (produced by expert psychologist) to the tags provided by users in a social network is to ambitious. (in some parts the authors make explicit this is an approximation, this should be stressed in every part of the paper)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors present a study that aims at inferring the \"emotional\" tags provided by Thumblr users starting from images and texts in the captions. ", "rating": "6: Marginally above acceptance threshold", "review": "\nThe authors present a study that aims at inferring the \"emotional\" tags provided by Thumblr users starting from images and texts in the captions. For text processing the authors use a standard LSTM taking as input GLOVE vectors of words in a sentence. For visual information, authors use a pretrained CNN (with fine tuning). A fully connected layer is used to fuse the multimodal information. Experimental results are reported in a self generated data set. \n\nThe contribution from the RL perspective is limited, in the sense that the authors simply applied standard models to predict a bunch of labels (in this case, emotion labels). It is interesting the \"psychological\" analysis that the authors present in Section 6. Still, I think the contribution in that part is a: sentiment-psychologically inspired analysis of the Thumbrl data set. \n\nI think the author's statement on that this study leads to a more plausible psychological model of emotion is not well founded (they also mention to learn to recognize the latent emotional state). Whereas it is true that psychological studies rely on self - filled questionnaires, comparing a questionnaire (produced by expert psychologist) to the tags provided by users in a social network is to ambitious. (in some parts the authors make explicit this is an approximation, this should be stressed in every part of the paper)\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511557814275}, {"id": "HJcw0y5eM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper978/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a multi-modal CNN model for sentiment analysis that combines images and text.  The model is trained on a new dataset collected from Tumblr.\n\nPositive aspects:\n+ Emphasis in model interpretability and its connection to psychological findings in emotions\n+ The idea of using Tumblr data seems interesting, allowing to work with a large set of emotion categories, instead of considering just the binary task positive vs. negative. \n\nWeaknesses:\n- A deeper analysis of previous work on the combination of image and text for sentiment analysis (both datasets and methods) and its relation with the presented work is necessary. \n- The proposed method is not compared with other methods that combine text and image for sentiment analysis.\n-  The study is limited to just one dataset.\n\nThe paper presents interesting ideas and findings in an important challenging area. The main novelties of the paper are: (1) the use of Tumblr data, (2) the proposed CNN architecture, combining images and text (using word embedding. \n\nI missed a \"related work section\", where authors clearly mention previous works on similar datasets. Some related works are mentioned in the paper, but those are spread in different sections. It's hard to get a clear overview of the previous research: datasets, methods and contextualization of the proposed approach in relation with previous work. I think authors should cite Sentibanks. Also, at some point authors should compare their proposal with previous work. \n\nMore comments:\n\n- Some figures could be more complete: to see more examples in Fig 1, 2, 3 would help to understand better the dataset and the challenges. \n- In table 4, for example, it would be nice to see the performance on the different emotion categories.\n- It would be interesting to see qualitative visual results on recognitions.\n\nI like this work, but I think authors should improve the aspects I mention for its publication.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper on sentiment analysis combining image and text ", "rating": "5: Marginally below acceptance threshold", "review": "The paper presents a multi-modal CNN model for sentiment analysis that combines images and text.  The model is trained on a new dataset collected from Tumblr.\n\nPositive aspects:\n+ Emphasis in model interpretability and its connection to psychological findings in emotions\n+ The idea of using Tumblr data seems interesting, allowing to work with a large set of emotion categories, instead of considering just the binary task positive vs. negative. \n\nWeaknesses:\n- A deeper analysis of previous work on the combination of image and text for sentiment analysis (both datasets and methods) and its relation with the presented work is necessary. \n- The proposed method is not compared with other methods that combine text and image for sentiment analysis.\n-  The study is limited to just one dataset.\n\nThe paper presents interesting ideas and findings in an important challenging area. The main novelties of the paper are: (1) the use of Tumblr data, (2) the proposed CNN architecture, combining images and text (using word embedding. \n\nI missed a \"related work section\", where authors clearly mention previous works on similar datasets. Some related works are mentioned in the paper, but those are spread in different sections. It's hard to get a clear overview of the previous research: datasets, methods and contextualization of the proposed approach in relation with previous work. I think authors should cite Sentibanks. Also, at some point authors should compare their proposal with previous work. \n\nMore comments:\n\n- Some figures could be more complete: to see more examples in Fig 1, 2, 3 would help to understand better the dataset and the challenges. \n- In table 4, for example, it would be nice to see the performance on the different emotion categories.\n- It would be interesting to see qualitative visual results on recognitions.\n\nI like this work, but I think authors should improve the aspects I mention for its publication.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511812706471}], "openreview_url": "https://openreview.net/forum?id=BJ6anzb0Z", "arxiv_id": "1805.10205", "paper_pdf": "papers/BJ6anzb0Z.pdf", "paper_pdf_sha256": "3d370dc8f8cd9fd4d0cd445f41bd470b2c7febef07b3e297913bb31439fb30aa", "paper_pdf_bytes": 3934971, "paper_pdf_source": "openreview", "code_url": "https://github.com/deepsentiment/deepsentiment", "code_repository": "deepsentiment/deepsentiment", "code_commit": "620fc7175fc7cedfe13ab9c59d48fcfdb8c47ec8", "code_archive": "repos/BJ6anzb0Z.zip", "code_archive_sha256": "971ef646ceadd63003b95f6bc5cf66dd160caf602b781cd4a61e96c794cfb3ba", "code_archive_bytes": 334694, "code_file_count": 73, "code_extensions": {".py": 67, ".sh": 5, ".ipynb": 1}, "github_disk_usage_kb": 275, "github_languages": {"Python": 667877, "Jupyter Notebook": 46257, "Shell": 10688}, "github_archived": false, "github_pushed_at": "2017-10-27T23:02:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multimodal-sentiment-analysis-to-explore-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HSWE9aceZb", "year": 2026, "status": "rejected", "title": "CXMArena: Unified Dataset to benchmark performance in realistic CXM Scenarios", "authors": ["Raghav Garg", "Karan Gupta", "Kapil Sharma"], "authorids": ["~Raghav_Garg1", "~Karan_Gupta4", "~Kapil_Sharma1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) hold immense potential for revolutionizing Customer Experience Management (CXM), particularly in contact center operations. However, evaluating their practical utility in complex operational environments is hindered by data scarcity (due to privacy concerns) and the limitations of current benchmarks. Existing benchmarks often lack realism, failing to incorporate deep knowledge base (KB) integration, real-world noise, or critical operational tasks beyond conversational fluency.\n\nTo bridge this gap, we introduce CXMdataset, a novel, large-scale synthetic benchmark dataset specifically designed for evaluating AI in operational CXM contexts. Given the diversity in possible contact center features, we have developed a scalable LLM-powered pipeline that simulates the brand's CXM entities that form the foundation of our datasets—such as knowledge articles including product specifications, issue taxonomies, and contact center conversations. The entities closely represent real-world distribution because of controlled noise injection (informed by domain experts) and rigorous automated validation.\n\nBuilding on this, we release CXMdataset, which provides dedicated benchmarks targeting five important operational tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools.\n\nOur baseline experiments underscore the benchmark's difficulty: even state-of-the-art embedding and generation models achieve only 68% accuracy on article search, while standard embedding methods yield a low F1 score of 0.3 for knowledge base refinement, highlighting significant challenges for current models necessitating complex pipelines and solutions over conventional techniques.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "N7iERIZaRx", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16376/Reviewer_gNnY"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "1. CXMArena manuscript introduces a unified, large-scale which is basically synthetic benchmark for evaluating AI models in Customer Experience Management (CXM). The synthetic data benchmark will be used in many real world scenarios\n\n2. The manuscript tells about five operational tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools which are all pretty well explained and in depth.\n\n3. All the dataset which are created using a scalable and high value LLM-powered pipeline that simulates realistic customer-agent interactions with controlled noise injection for authenticity. The LLMs used are pretty in depth with metrics supported. The synthetic data can be used in very critical in CXM scenarios.", "review_text": "1. CXMArena manuscript introduces a unified, large-scale which is basically synthetic benchmark for evaluating AI models in Customer Experience Management (CXM). The synthetic data benchmark will be used in many real world scenarios\n\n2. The manuscript tells about five operational tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools which are all pretty well explained and in depth.\n\n3. All the dataset which are created using a scalable and high value LLM-powered pipeline that simulates realistic customer-agent interactions with controlled noise injection for authenticity. The LLMs used are pretty in depth with metrics supported. The synthetic data can be used in very critical in CXM scenarios.", "strengths": "1. Comprehensive benchmark covering core CXM tasks beyond usual fluency. The CXM tasks usually have real word scenario use cases so the manuscript is useful.\n \n2. Synthetic yet realistic data with strong alignment to real-world metrics. The realistic nature of this data has very strong alignment and large scale value in very in depth metrics.\n\n3. Cross-domain and multilingual support (English, French, German). The cross domain languages are very well used in here.\n\n4. Provides baseline results for multiple models. Multiple models are been used like GPT and others.", "weaknesses": "1. Although the manuscript is being used for synthetic, it may miss subtle nuances of real human behavior. Also LLM with human as a judge could be helpful to explore that might strengthen the findings\n2. Dependent on biases and limitations of LLMs used for generation.\n3. Currently focused on one domain with limited real-world diversity. Also multi domain alignment could be helpful", "questions": "1. How will CXMArena handle continuous domain drift in real CXM data? The drift can be measured and how can be used.\n2. Can the synthetic pipeline generalize to non-English, multi-brand, or emotion-rich interactions?\n3. How could MCP be used and how some performance benchmarks on MCP be used", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "1. CXMArena manuscript introduces a unified, large-scale which is basically synthetic benchmark for evaluating AI models in Customer Experience Management (CXM). The synthetic data benchmark will be used in many real world scenarios\n\n2. The manuscript tells about five operational tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools which are all pretty well explained and in depth.\n\n3. All the dataset which are created using a scalable and high value LLM-powered pipeline that simulates realistic customer-agent interactions with controlled noise injection for authenticity. The LLMs used are pretty in depth with metrics supported. The synthetic data can be used in very critical in CXM scenarios.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Comprehensive benchmark covering core CXM tasks beyond usual fluency. The CXM tasks usually have real word scenario use cases so the manuscript is useful.\n \n2. Synthetic yet realistic data with strong alignment to real-world metrics. The realistic nature of this data has very strong alignment and large scale value in very in depth metrics.\n\n3. Cross-domain and multilingual support (English, French, German). The cross domain languages are very well used in here.\n\n4. Provides baseline results for multiple models. Multiple models are been used like GPT and others.", "weaknesses": "1. Although the manuscript is being used for synthetic, it may miss subtle nuances of real human behavior. Also LLM with human as a judge could be helpful to explore that might strengthen the findings\n2. Dependent on biases and limitations of LLMs used for generation.\n3. Currently focused on one domain with limited real-world diversity. Also multi domain alignment could be helpful", "questions": "1. How will CXMArena handle continuous domain drift in real CXM data? The drift can be measured and how can be used.\n2. Can the synthetic pipeline generalize to non-English, multi-brand, or emotion-rich interactions?\n3. How could MCP be used and how some performance benchmarks on MCP be used", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762036089132}, {"id": "Bhd5KybxcC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16376/Reviewer_52hR"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors propose a new dataset for evaluation of LLM response in Customer Service. They tackle important problems like Knowledge base based questions, tool calling, multi-turn question answers etc. They provide limitations of current datasets and propose a system which can address these limitations.", "review_text": "The authors propose a new dataset for evaluation of LLM response in Customer Service. They tackle important problems like Knowledge base based questions, tool calling, multi-turn question answers etc. They provide limitations of current datasets and propose a system which can address these limitations.", "strengths": "The authors bring forth a very important problem and one which practitioners constantly face. One very important aspect of public datasets and papers is that they very rarely to industry situations and the authors point it out very well. They also do a comprehensive study of the different challenges situations in Customer Service via LLMs in Section 2 and bring about limitations of exisiting datasets very well.", "weaknesses": "The authors highlight that most public datasets have limitations. This is a very valid concern and they have given significant citations establishing the limitations of existing datasets in Section 2. What is not clear is how is this alleviated in their work. Looking at one example they mention \"We simulate real-world data quality issues by introducing controlled redundant and contradictory information from one article to another, creating data for developing KB maintenance techniques\" - it is not clear how is this going to solve the very real problems in lines 088-103.  Similarly, their description on Multi-turn RAG in lines 252-257 do not explain how the limitations discussed in Section 2 are addressed. The paper largely glosses over details (some information is present in Appendix but that is not under detailed review) and more importantly even from Appendix it is not clear how the concerns are addressed.\n\nNext the authors provide multiple evaluation results with different models however it is hard to understand if this is good or bad without a benchmark. They provide some examples in the appendix however they do not give examples of how they are different from existing datasets making an evaluation of the marginal improvement challenging.", "questions": "Please address the concerns in the weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a new dataset for evaluation of LLM response in Customer Service. They tackle important problems like Knowledge base based questions, tool calling, multi-turn question answers etc. They provide limitations of current datasets and propose a system which can address these limitations.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The authors bring forth a very important problem and one which practitioners constantly face. One very important aspect of public datasets and papers is that they very rarely to industry situations and the authors point it out very well. They also do a comprehensive study of the different challenges situations in Customer Service via LLMs in Section 2 and bring about limitations of exisiting datasets very well.", "weaknesses": "The authors highlight that most public datasets have limitations. This is a very valid concern and they have given significant citations establishing the limitations of existing datasets in Section 2. What is not clear is how is this alleviated in their work. Looking at one example they mention \"We simulate real-world data quality issues by introducing controlled redundant and contradictory information from one article to another, creating data for developing KB maintenance techniques\" - it is not clear how is this going to solve the very real problems in lines 088-103.  Similarly, their description on Multi-turn RAG in lines 252-257 do not explain how the limitations discussed in Section 2 are addressed. The paper largely glosses over details (some information is present in Appendix but that is not under detailed review) and more importantly even from Appendix it is not clear how the concerns are addressed.\n\nNext the authors provide multiple evaluation results with different models however it is hard to understand if this is good or bad without a benchmark. They provide some examples in the appendix however they do not give examples of how they are different from existing datasets making an evaluation of the marginal improvement challenging.", "questions": "Please address the concerns in the weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761495421349}, {"id": "IwK4UbUeub", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16376/Reviewer_wsXP"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors introduce CXMArena: a novel, large-scale synthetic benchmark dataset specifically designed for evaluating AI in operational Customer Experience Management contexts.", "review_text": "The authors introduce CXMArena: a novel, large-scale synthetic benchmark dataset specifically designed for evaluating AI in operational Customer Experience Management contexts.", "strengths": "- Real-world distribution because of controlled noise injection (simulated ASR errors, interaction fragments) from SMEs and rigorous automated validation.\n\n- Authors introduce five tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools.\n\n- Pipeline applied to different domains and languages.\n\n- The authors introduce a pipeline to synthetically generate the knowledge base specific to a fictional brand and then uses these KBs to generate realistic conversations along with noise.", "weaknesses": "- My main concern is on the correctness of the synthetic data using an LLM (Gemini in this case) and the LLM as a judge evaluation without a human in the loop. \n\n- Contradiction detection baseline would be very insightful since this is one of the tasks needed for Knowledge Base refinement.\n\n- Do you generate all this dataset synthetically given a seed prompt about the brand name and its type? I'm not convinced how you can ensure highly fidelity data since there is no human in the loop. How do you avoid hallucinations? For instance for the KB refinement task how do you know that similar KBs are correctly labelled? \n\n- GPT 4o as a judge and no human evaluation.\n\n- LLM was used for checking the dataset correctness with no human expert in the loop.", "questions": "In the abstract you mention \"The entities closely represent real-world distribution because of controlled\nnoise injection (informed by domain experts)\". What kind of information do the experts provide?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce CXMArena: a novel, large-scale synthetic benchmark dataset specifically designed for evaluating AI in operational Customer Experience Management contexts.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Real-world distribution because of controlled noise injection (simulated ASR errors, interaction fragments) from SMEs and rigorous automated validation.\n\n- Authors introduce five tasks: Knowledge Base Refinement, Intent Prediction, Agent Quality Adherence, Article Search, and Multi-turn RAG with Integrated Tools.\n\n- Pipeline applied to different domains and languages.\n\n- The authors introduce a pipeline to synthetically generate the knowledge base specific to a fictional brand and then uses these KBs to generate realistic conversations along with noise.", "weaknesses": "- My main concern is on the correctness of the synthetic data using an LLM (Gemini in this case) and the LLM as a judge evaluation without a human in the loop. \n\n- Contradiction detection baseline would be very insightful since this is one of the tasks needed for Knowledge Base refinement.\n\n- Do you generate all this dataset synthetically given a seed prompt about the brand name and its type? I'm not convinced how you can ensure highly fidelity data since there is no human in the loop. How do you avoid hallucinations? For instance for the KB refinement task how do you know that similar KBs are correctly labelled? \n\n- GPT 4o as a judge and no human evaluation.\n\n- LLM was used for checking the dataset correctness with no human expert in the loop.", "questions": "In the abstract you mention \"The entities closely represent real-world distribution because of controlled\nnoise injection (informed by domain experts)\". What kind of information do the experts provide?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761077753185}, {"id": "RvTPVFtpRK", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16376/Reviewer_nXhr"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The authors developed a new way to automatically generate a complete, brand-specific Knowledge Base (KB) for Customer Experience Management (CXM) problems. They have publicly released the entire dataset, which includes articles, synthesized user queries, and tool definitions.\n\nThe paper argues that existing benchmarks fall short for CXM because they focus on general, open-domain knowledge and generic tools. This work tackles that gap by creating a benchmark that includes a full suite of core tasks relevant to CXM, which requires specialized brand/product knowledge and the use of tailored tools, just like in real-world customer support scenarios.\n\nUsing their synthetic KB, the authors designed and benchmarked a suite of tasks to evaluate different LLM capabilities. This includes an Intent Prediction task for classification of user needs, an Article Retrieval task for retrieval of relevant documents, a Tool Use task for executing API calls, and a Quality Adherence task to evaluate if responses follow brand guidelines. The benchmark also includes an end-to-end Multi-turn RAG with Tool Use task, which is a generation problem that integrates all these skills to produce a final, context-aware response.", "review_text": "The authors developed a new way to automatically generate a complete, brand-specific Knowledge Base (KB) for Customer Experience Management (CXM) problems. They have publicly released the entire dataset, which includes articles, synthesized user queries, and tool definitions.\n\nThe paper argues that existing benchmarks fall short for CXM because they focus on general, open-domain knowledge and generic tools. This work tackles that gap by creating a benchmark that includes a full suite of core tasks relevant to CXM, which requires specialized brand/product knowledge and the use of tailored tools, just like in real-world customer support scenarios.\n\nUsing their synthetic KB, the authors designed and benchmarked a suite of tasks to evaluate different LLM capabilities. This includes an Intent Prediction task for classification of user needs, an Article Retrieval task for retrieval of relevant documents, a Tool Use task for executing API calls, and a Quality Adherence task to evaluate if responses follow brand guidelines. The benchmark also includes an end-to-end Multi-turn RAG with Tool Use task, which is a generation problem that integrates all these skills to produce a final, context-aware response.", "strengths": "The paper identifies a critical gap in existing benchmarks, which typically focus on open-domain tasks and fail to address the unique challenges of Customer Experience Management (CXM), such as specialized knowledge retrieval, tailored tool use, and quality adherence.\n\nA key contribution is the novel data synthesis pipeline that generates a complete and realistic brand-specific Knowledge Base. Synthesizing such data is valuable & practical as it enables the creation of a rich CXM environment for research without exposing sensitive company data or compromising user privacy.\n\nThe work provides a comprehensive benchmark by designing and evaluating performance on a suite of diverse tasks relevant to CXM, including retrieval, tool use, and end-to-end multi-turn RAG, thereby establishing important baselines for future research.", "weaknesses": "The benchmark's primary weakness is a lack of a clear rationale for performing several isolated, synthesized tasks on a synthesized knowledge base. For example, both the article retrieval and tool-calling tasks are designed in isolation. While this is a valid approach, the paper does not explain the unique advantages of this synthetic setup over existing approaches that synthesize search queries, or tool calls from openly available articles and real-world tools (e.g. ToolBench).  A stronger justification would help frame the benchmark as a cohesive suite of tasks designed to address novel CXM challenges, rather than a collection of disconnected components. (e.g. any unique challenge around jointly performing KB retrieval, tool call during a multi-turn RAG problem?)\n\nWhile the paper does involve an evaluation on an end-to-end RAG task, a key methodological detail that remains unclear is how tool outputs are simulated during the multi-turn RAG evaluation, or if tool calls are simulated at all. In addition, while the paper mentions llm-as-a-judge is used to evaluate the final multi-turn RAG output, it is not clear if the evaluation is grounded in the golden response or the correct sources.", "questions": "In the multi-turn RAG evaluation, how are tool outputs simulated and provided to the model after a tool call is made?\n\nFor the LLM-as-a-judge evaluation, is the judge grounded against the knowledge base or golden responses to verify factual accuracy?\n\nIt seems that KB_3211 is referenced in the multi-turn set but is missing from the articles subset. Could you clarify this data discrepancy?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors developed a new way to automatically generate a complete, brand-specific Knowledge Base (KB) for Customer Experience Management (CXM) problems. They have publicly released the entire dataset, which includes articles, synthesized user queries, and tool definitions.\n\nThe paper argues that existing benchmarks fall short for CXM because they focus on general, open-domain knowledge and generic tools. This work tackles that gap by creating a benchmark that includes a full suite of core tasks relevant to CXM, which requires specialized brand/product knowledge and the use of tailored tools, just like in real-world customer support scenarios.\n\nUsing their synthetic KB, the authors designed and benchmarked a suite of tasks to evaluate different LLM capabilities. This includes an Intent Prediction task for classification of user needs, an Article Retrieval task for retrieval of relevant documents, a Tool Use task for executing API calls, and a Quality Adherence task to evaluate if responses follow brand guidelines. The benchmark also includes an end-to-end Multi-turn RAG with Tool Use task, which is a generation problem that integrates all these skills to produce a final, context-aware response.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "The paper identifies a critical gap in existing benchmarks, which typically focus on open-domain tasks and fail to address the unique challenges of Customer Experience Management (CXM), such as specialized knowledge retrieval, tailored tool use, and quality adherence.\n\nA key contribution is the novel data synthesis pipeline that generates a complete and realistic brand-specific Knowledge Base. Synthesizing such data is valuable & practical as it enables the creation of a rich CXM environment for research without exposing sensitive company data or compromising user privacy.\n\nThe work provides a comprehensive benchmark by designing and evaluating performance on a suite of diverse tasks relevant to CXM, including retrieval, tool use, and end-to-end multi-turn RAG, thereby establishing important baselines for future research.", "weaknesses": "The benchmark's primary weakness is a lack of a clear rationale for performing several isolated, synthesized tasks on a synthesized knowledge base. For example, both the article retrieval and tool-calling tasks are designed in isolation. While this is a valid approach, the paper does not explain the unique advantages of this synthetic setup over existing approaches that synthesize search queries, or tool calls from openly available articles and real-world tools (e.g. ToolBench).  A stronger justification would help frame the benchmark as a cohesive suite of tasks designed to address novel CXM challenges, rather than a collection of disconnected components. (e.g. any unique challenge around jointly performing KB retrieval, tool call during a multi-turn RAG problem?)\n\nWhile the paper does involve an evaluation on an end-to-end RAG task, a key methodological detail that remains unclear is how tool outputs are simulated during the multi-turn RAG evaluation, or if tool calls are simulated at all. In addition, while the paper mentions llm-as-a-judge is used to evaluate the final multi-turn RAG output, it is not clear if the evaluation is grounded in the golden response or the correct sources.", "questions": "In the multi-turn RAG evaluation, how are tool outputs simulated and provided to the model after a tool call is made?\n\nFor the LLM-as-a-judge evaluation, is the judge grounded against the knowledge base or golden responses to verify factual accuracy?\n\nIt seems that KB_3211 is referenced in the multi-turn set but is missing from the articles subset. Could you clarify this data discrepancy?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760836767310}], "openreview_url": "https://openreview.net/forum?id=HSWE9aceZb", "arxiv_id": "2505.09436", "paper_pdf": "papers/HSWE9aceZb.pdf", "paper_pdf_sha256": "241723888c50213be7e73a6e66eaa0c7e2abd673ddd88cc1708ada59dd56f46f", "paper_pdf_bytes": 2084781, "paper_pdf_source": "openreview", "code_url": "https://github.com/kapilsprinklr/CXMArena", "code_repository": "kapilsprinklr/CXMArena", "code_commit": "755a7c46976d0446aa30c0b210846771f4391bb4", "code_archive": "repos/HSWE9aceZb.zip", "code_archive_sha256": "ad625e4d2ecc7be309a964b480b69a2ba87fbe941c1a5240b258df46e349eb47", "code_archive_bytes": 62764, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 140, "github_languages": {"Python": 50274}, "github_archived": false, "github_pushed_at": "2026-09-01T11:09:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cxmarena-unified-dataset-to-benchmark"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "F0Zd3knG9j", "year": 2025, "status": "rejected", "title": "How transformers learn structured data: insights from hierarchical filtering", "authors": ["Jerome Garnier-Brun", "Marc Mezard", "Emanuele Moscato", "Luca Saglietti"], "authorids": ["~Jerome_Garnier-Brun1", "~Marc_Mezard1", "~Emanuele_Moscato1", "~Luca_Saglietti1"], "authors_source": "OpenReview API", "abstract": "Understanding the learning process and the embedded computation in transformers is becoming a central goal for the development of interpretable AI. In the present study, we introduce a hierarchical filtering procedure for generative models of sequences on trees, allowing us to hand-tune the range of positional correlations in the data. Leveraging this controlled setting, we provide evidence that vanilla encoder-only transformers can approximate the exact inference algorithm when trained on root classification and masked language modeling tasks, and study *how* this computation is discovered and implemented. We find that correlations at larger distances, corresponding to increasing layers of the hierarchy, are sequentially included by the network during training. Moreover, by comparing attention maps from models trained with varying degrees of filtering and by probing the different encoder levels, we find clear evidence of a reconstruction of correlations on successive length scales corresponding to the various levels of the hierarchy, which we relate to a plausible implementation of the exact inference algorithm within the same architecture.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "ePQ20ywz3G", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4635/Reviewer_c8Kw"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper investigates how transformer models make predictions on samples coming from a structured data distribution, focusing on the hypothesis that transformers implement belief propagation to make predictions. \n\nContributions:\n1. The authors propose a novel family of synthetic data distributions based on PCFGs to test the hypothesis empirically. \n2. The authors present experimental results on two tasks, root prediction and masked language modeling, which the authors claim to support their hypothesis.\n3. The authors present a construction for how to implement belief propagation for a tree-structured factor graph of depth $l$, using only $l$ layers.", "review_text": "The paper investigates how transformer models make predictions on samples coming from a structured data distribution, focusing on the hypothesis that transformers implement belief propagation to make predictions. \n\nContributions:\n1. The authors propose a novel family of synthetic data distributions based on PCFGs to test the hypothesis empirically. \n2. The authors present experimental results on two tasks, root prediction and masked language modeling, which the authors claim to support their hypothesis.\n3. The authors present a construction for how to implement belief propagation for a tree-structured factor graph of depth $l$, using only $l$ layers.", "strengths": "1. The work is well-placed in the context of other mechanistic interpretability work for transformers, as well as other work looking into the effect of structured data on machine learning models.\n2. Understanding what strategy is learned by transformer models trained on structured distributions is an important problem.\n3. The family of synthetic distributions is interesting and has a hyperparameter that allows one to control the locality of correlations between tokens in the sequence.\n4. The authors present a novel construction for how to implement BP on a depth $l$ tree using only $l$ layers of a transformer, whereas previous constructions required $2l$ layers.", "weaknesses": "1. It’s not clear how much the observations made in this work generalizes to larger models, and more complicated data distributions.\n2. The empirical evidence presented has alternative interpretations that haven't been ruled out. (Please see Questions.)\n3. The construction for implementing BP on depth-$l$ tree using a transformer of only $l$-layers could benefit from a bit more details, especially on the root-to-leaf message passing part. (Please see Questions.)", "questions": "W2.1: In both the supervised root prediction task and the MLM task, the authors argue that the transformer performs similar in accuracy to BP is evidence that the transformer is implementing an approximation to BP. While it is very intriguing that the accuracies are so systematically similar, there are plausible alternative explanations (that additional experiments or analysis could rule out):\n\nW2.1.1: Similar accuracy doesn’t imply similar behavior on individual inputs. Have authors considered measuring the match between BP and transformer predictions on individual inputs? If the match is high, this would strengthen the authors’ claim that the model actually behaves like BP. \n\nW2.1.2: Similar behavior on individual inputs doesn’t imply similar implementation. For example, in figure 1(c), the authors show accuracy of a model trained on k=0 (and for root classification) data and evaluated on k>=0 data, whose accuracy is similar to running BP on the k=0 graph. The authors claim this is evidence in support of model having learned to implement BP. Why couldn’t the model have model the k=0 data well without implementing BP? BP on k=0 graph is optimal for k=0 data, so the transformer that was trained on lots of k=0 data would necessarily behave like BP on individual inputs, which (without necessarily implementing BP) would make similar predictions to BP on out-of-sample data as well.\n\nW2.2 In the supervised root prediction task, the authors write that (line 314-316) “We interpret this as a consequence of the weaker correlations between distant tokens—and therefore the lower signal-to-noise ratio during learning—that must be resolved to match the BP prediction.” Since the task is to predict the root, could the authors elaborate more on why they believe the weaker correlations among **tokens within a sequence** is causing difficulty for learning? An alternative interpretation is that this is due to the weaker correlation between the **root** and the **entire sequence** of tokens.\t\n\nW3. Following up on the BP construction, could the authors elaborate more on the leaf-to-root passing? In particular, what do the $r$’s represent? It looks like all $r$’s are initialized to uniform distribution, and they each get updated with the same formula (eqn 22), so would $r^{(a,m)}_i$ ever be different from $r^{(a’,m)}_i$ for $a \\neq a’$? A walkthrough of the formulas on a minimal example with small $l$, and $q$ may be helpful here.\n\nOther: Did the authors mean to refer to figure 3 instead of 1c on lines 337? In the caption of figure 1c it says it’s for MLM instead of root classification, and the scale of the x-axis suggests it’s MLM too.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates how transformer models make predictions on samples coming from a structured data distribution, focusing on the hypothesis that transformers implement belief propagation to make predictions. \n\nContributions:\n1. The authors propose a novel family of synthetic data distributions based on PCFGs to test the hypothesis empirically. \n2. The authors present experimental results on two tasks, root prediction and masked language modeling, which the authors claim to support their hypothesis.\n3. The authors present a construction for how to implement belief propagation for a tree-structured factor graph of depth $l$, using only $l$ layers.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The work is well-placed in the context of other mechanistic interpretability work for transformers, as well as other work looking into the effect of structured data on machine learning models.\n2. Understanding what strategy is learned by transformer models trained on structured distributions is an important problem.\n3. The family of synthetic distributions is interesting and has a hyperparameter that allows one to control the locality of correlations between tokens in the sequence.\n4. The authors present a novel construction for how to implement BP on a depth $l$ tree using only $l$ layers of a transformer, whereas previous constructions required $2l$ layers.", "weaknesses": "1. It’s not clear how much the observations made in this work generalizes to larger models, and more complicated data distributions.\n2. The empirical evidence presented has alternative interpretations that haven't been ruled out. (Please see Questions.)\n3. The construction for implementing BP on depth-$l$ tree using a transformer of only $l$-layers could benefit from a bit more details, especially on the root-to-leaf message passing part. (Please see Questions.)", "questions": "W2.1: In both the supervised root prediction task and the MLM task, the authors argue that the transformer performs similar in accuracy to BP is evidence that the transformer is implementing an approximation to BP. While it is very intriguing that the accuracies are so systematically similar, there are plausible alternative explanations (that additional experiments or analysis could rule out):\n\nW2.1.1: Similar accuracy doesn’t imply similar behavior on individual inputs. Have authors considered measuring the match between BP and transformer predictions on individual inputs? If the match is high, this would strengthen the authors’ claim that the model actually behaves like BP. \n\nW2.1.2: Similar behavior on individual inputs doesn’t imply similar implementation. For example, in figure 1(c), the authors show accuracy of a model trained on k=0 (and for root classification) data and evaluated on k>=0 data, whose accuracy is similar to running BP on the k=0 graph. The authors claim this is evidence in support of model having learned to implement BP. Why couldn’t the model have model the k=0 data well without implementing BP? BP on k=0 graph is optimal for k=0 data, so the transformer that was trained on lots of k=0 data would necessarily behave like BP on individual inputs, which (without necessarily implementing BP) would make similar predictions to BP on out-of-sample data as well.\n\nW2.2 In the supervised root prediction task, the authors write that (line 314-316) “We interpret this as a consequence of the weaker correlations between distant tokens—and therefore the lower signal-to-noise ratio during learning—that must be resolved to match the BP prediction.” Since the task is to predict the root, could the authors elaborate more on why they believe the weaker correlations among **tokens within a sequence** is causing difficulty for learning? An alternative interpretation is that this is due to the weaker correlation between the **root** and the **entire sequence** of tokens.\t\n\nW3. Following up on the BP construction, could the authors elaborate more on the leaf-to-root passing? In particular, what do the $r$’s represent? It looks like all $r$’s are initialized to uniform distribution, and they each get updated with the same formula (eqn 22), so would $r^{(a,m)}_i$ ever be different from $r^{(a’,m)}_i$ for $a \\neq a’$? A walkthrough of the formulas on a minimal example with small $l$, and $q$ may be helpful here.\n\nOther: Did the authors mean to refer to figure 3 instead of 1c on lines 337? In the caption of figure 1c it says it’s for MLM instead of root classification, and the scale of the x-axis suggests it’s MLM too.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730658076213}, {"id": "Ujs39TYgTR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4635/Reviewer_VwvV"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper evaluates encoder-only transformers on a synthetic task, where data is sampled from a complete binary tree-structured generative model of fixed depth $\\ell$. There is a knob $k$ that determines at which layer that subtrees are forced to be conditionally independent. This generative process is equivalent to a PCFG, although it is framed mostly in terms of factor graphs and belief propagation. The authors find that transformers can predict the root node type given the leaves with high accuracy. They find that they can do MLM prediction with high accuracy, and that using MLM as a pretraining step improves sample efficiency for root prediction. They analyze attention patterns and see that they attend in a hierarchical fashion as expected.", "review_text": "This paper evaluates encoder-only transformers on a synthetic task, where data is sampled from a complete binary tree-structured generative model of fixed depth $\\ell$. There is a knob $k$ that determines at which layer that subtrees are forced to be conditionally independent. This generative process is equivalent to a PCFG, although it is framed mostly in terms of factor graphs and belief propagation. The authors find that transformers can predict the root node type given the leaves with high accuracy. They find that they can do MLM prediction with high accuracy, and that using MLM as a pretraining step improves sample efficiency for root prediction. They analyze attention patterns and see that they attend in a hierarchical fashion as expected.", "strengths": "Some of the experimental results are quite interesting. For example, Fig 4 shows the expected attention patterns, and Sec 3.4 is interesting in that it shows that MLM pretraining improves sample efficiency.", "weaknesses": "Although the paper includes some interesting visualizations, I am not quite convinced that the contributions of this paper are particularly novel or rise to the level of a full ICLR paper. At times I also found the paper difficult to read, and its claims unclear.\n\n1. In terms of novelty, there seems to be significant overlap with Allen-Zhu & Li (2023), and the experiments seem to be a simpler case of that paper. Like this paper, Allen-Zhu & Li (2023) trained transformers on data generated from CFGs of fixed depth and showed that the the transformer layers learned to attend to constituents as expected.\n1. In terms of the significance of the contribution, independently of the question of whether it is novel, this is a very simple synthetic task that seems a bit contrived so that transformers are successful on it (an issue that is also in Allen-Zhu & Li (2023)), and it is not clear that we learn very much about transformers from these experiments. The strings in the training data are all of the same length, and the depth of the underlying parse tree is also fixed and does not exceed the number of transformer layers. It is not surprising that a transformer encoder with the same number of layers as the underlying parse tree can learn to mimic the structure of the underlying complete binary tree. It would be more interesting to test the transformer on a CFG with parse trees of varying depths. Do we see similar behavior, and does the fact that the number of layers is finite matter then?\n1. In terms of clarity, it is not clear at the beginning of the paper what its primary goal is. Is the paper primarily about proposing a new data model, and if so, what is the purpose and significance of $k$? Are you primarily interested in analyzing the transformer architecture, and if so, how will the analysis on this synthetic task help us understand the behavior of transformers on real tasks such as natural language?\n1. One of the main claims of the paper is that the transformer learns to implement a belief propagation algorithm, but I don't see significant evidence that shows that it is learning to implement BP vs. another algorithm, e.g., some version of the inside algorithm. I don't think the analysis of accuracy on OOD examples and attention patterns rules this case out.\n1. Major figures supporting the paper's claims are only in the appendix (Fig 7, App C.3 and C.4).", "questions": "1. Intuitively, what \"knob\" does the filtering parameter $k$ represent? Is it the case that lower $k$ result in more long-range correlations in the data?\n1. 035: Another relevant paper: https://arxiv.org/abs/2305.02386\n1. Is there a particular reason why you chose to frame the paper mostly in terms of factor graphs and belief propagation, rather than CFGs and standard parsing algorithms (e.g., the inside algorithm)? Is there an advantage to presenting it this way? Is there an advantage in time complexity vs. using a CFG parsing algorithm?\n1. 119: What would the equivalent PCFG be, incorporating the depth constraint and $k$?\n1. 133: What is $\\mathcal{O}_a$? What is $q$? This part is very unclear to me.\n1. 144: It's not clear to me what this means. Can you express this in equations?\n1. Can the root always be uniquely determined by the input symbols? According to my understanding, the underlying CFG can be ambiguous. How is it possible to get 100% accuracy?\n1. How does the BP algorithm described in the main text relate to the experiments? In what way is it used? I don't think this is stated explicitly.\n1. Fig 3: Why do you report validation accuracy but not test accuracy? Why does the accuracy go up and then down? Did you not use the best checkpoint when evaluating on OOD data?\n1. Why use accuracy instead of perplexity for MLM? Since the CFG can be ambiguous, there isn't only one correct answer, right?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper evaluates encoder-only transformers on a synthetic task, where data is sampled from a complete binary tree-structured generative model of fixed depth $\\ell$. There is a knob $k$ that determines at which layer that subtrees are forced to be conditionally independent. This generative process is equivalent to a PCFG, although it is framed mostly in terms of factor graphs and belief propagation. The authors find that transformers can predict the root node type given the leaves with high accuracy. They find that they can do MLM prediction with high accuracy, and that using MLM as a pretraining step improves sample efficiency for root prediction. They analyze attention patterns and see that they attend in a hierarchical fashion as expected.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Some of the experimental results are quite interesting. For example, Fig 4 shows the expected attention patterns, and Sec 3.4 is interesting in that it shows that MLM pretraining improves sample efficiency.", "weaknesses": "Although the paper includes some interesting visualizations, I am not quite convinced that the contributions of this paper are particularly novel or rise to the level of a full ICLR paper. At times I also found the paper difficult to read, and its claims unclear.\n\n1. In terms of novelty, there seems to be significant overlap with Allen-Zhu & Li (2023), and the experiments seem to be a simpler case of that paper. Like this paper, Allen-Zhu & Li (2023) trained transformers on data generated from CFGs of fixed depth and showed that the the transformer layers learned to attend to constituents as expected.\n1. In terms of the significance of the contribution, independently of the question of whether it is novel, this is a very simple synthetic task that seems a bit contrived so that transformers are successful on it (an issue that is also in Allen-Zhu & Li (2023)), and it is not clear that we learn very much about transformers from these experiments. The strings in the training data are all of the same length, and the depth of the underlying parse tree is also fixed and does not exceed the number of transformer layers. It is not surprising that a transformer encoder with the same number of layers as the underlying parse tree can learn to mimic the structure of the underlying complete binary tree. It would be more interesting to test the transformer on a CFG with parse trees of varying depths. Do we see similar behavior, and does the fact that the number of layers is finite matter then?\n1. In terms of clarity, it is not clear at the beginning of the paper what its primary goal is. Is the paper primarily about proposing a new data model, and if so, what is the purpose and significance of $k$? Are you primarily interested in analyzing the transformer architecture, and if so, how will the analysis on this synthetic task help us understand the behavior of transformers on real tasks such as natural language?\n1. One of the main claims of the paper is that the transformer learns to implement a belief propagation algorithm, but I don't see significant evidence that shows that it is learning to implement BP vs. another algorithm, e.g., some version of the inside algorithm. I don't think the analysis of accuracy on OOD examples and attention patterns rules this case out.\n1. Major figures supporting the paper's claims are only in the appendix (Fig 7, App C.3 and C.4).", "questions": "1. Intuitively, what \"knob\" does the filtering parameter $k$ represent? Is it the case that lower $k$ result in more long-range correlations in the data?\n1. 035: Another relevant paper: https://arxiv.org/abs/2305.02386\n1. Is there a particular reason why you chose to frame the paper mostly in terms of factor graphs and belief propagation, rather than CFGs and standard parsing algorithms (e.g., the inside algorithm)? Is there an advantage to presenting it this way? Is there an advantage in time complexity vs. using a CFG parsing algorithm?\n1. 119: What would the equivalent PCFG be, incorporating the depth constraint and $k$?\n1. 133: What is $\\mathcal{O}_a$? What is $q$? This part is very unclear to me.\n1. 144: It's not clear to me what this means. Can you express this in equations?\n1. Can the root always be uniquely determined by the input symbols? According to my understanding, the underlying CFG can be ambiguous. How is it possible to get 100% accuracy?\n1. How does the BP algorithm described in the main text relate to the experiments? In what way is it used? I don't think this is stated explicitly.\n1. Fig 3: Why do you report validation accuracy but not test accuracy? Why does the accuracy go up and then down? Did you not use the best checkpoint when evaluating on OOD data?\n1. Why use accuracy instead of perplexity for MLM? Since the CFG can be ambiguous, there isn't only one correct answer, right?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730251879929}, {"id": "DozGPZRIIY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4635/Reviewer_BYbw"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "The authors build a binary-tree CFG with a filtering mechanism that the nodes on layer k is sampled only conditioned on the root node. \n\nA transformer encoder is then trained to predict the root node. With k=0, the prediction is perfect. And the performance decreases as k increase. It is also shown that the model's performance on out-sample k exactly matches the performance from BP. The authors then conduct experiments on masked prediction (like BERT). Finally, the authors propose an exact implementation of the BP algorithm through the transformer computation.", "review_text": "The authors build a binary-tree CFG with a filtering mechanism that the nodes on layer k is sampled only conditioned on the root node. \n\nA transformer encoder is then trained to predict the root node. With k=0, the prediction is perfect. And the performance decreases as k increase. It is also shown that the model's performance on out-sample k exactly matches the performance from BP. The authors then conduct experiments on masked prediction (like BERT). Finally, the authors propose an exact implementation of the BP algorithm through the transformer computation.", "strengths": "It's a novel viewpoint to study transformer learning from belief propagation.\n\nThe CFG construction with filtering is interesting.\n\nThe study is from multiple perspective: prediction accuracy, probing, and manual construction.", "weaknesses": "I'm not sure whether the transformer accuracy matches that of belief propagation is surprising, could it be a natural consequence that the transformer is simply learning the \"optimal\" prediction function (which is the prediction made by belief propagation)?\n\nI think the writing of this paper can be improved.\n\nWhile the construction sec3.5 is interesting, it does not mean that the learned transformer is actually doing that (if I understand correct).\n\nI may consider raise my score if other two reviewers show strong interest.", "questions": "Line355 I did not quite understand why the accuracy would have a drop in the middle of training.\n\nAlso I hope the writing for section3.3 can be improved by giving more easy-to-understand intuition.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors build a binary-tree CFG with a filtering mechanism that the nodes on layer k is sampled only conditioned on the root node. \n\nA transformer encoder is then trained to predict the root node. With k=0, the prediction is perfect. And the performance decreases as k increase. It is also shown that the model's performance on out-sample k exactly matches the performance from BP. The authors then conduct experiments on masked prediction (like BERT). Finally, the authors propose an exact implementation of the BP algorithm through the transformer computation.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "It's a novel viewpoint to study transformer learning from belief propagation.\n\nThe CFG construction with filtering is interesting.\n\nThe study is from multiple perspective: prediction accuracy, probing, and manual construction.", "weaknesses": "I'm not sure whether the transformer accuracy matches that of belief propagation is surprising, could it be a natural consequence that the transformer is simply learning the \"optimal\" prediction function (which is the prediction made by belief propagation)?\n\nI think the writing of this paper can be improved.\n\nWhile the construction sec3.5 is interesting, it does not mean that the learned transformer is actually doing that (if I understand correct).\n\nI may consider raise my score if other two reviewers show strong interest.", "questions": "Line355 I did not quite understand why the accuracy would have a drop in the middle of training.\n\nAlso I hope the writing for section3.3 can be improved by giving more easy-to-understand intuition.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1729665768919}], "openreview_url": "https://openreview.net/forum?id=F0Zd3knG9j", "arxiv_id": "2408.15138", "paper_pdf": "papers/F0Zd3knG9j.pdf", "paper_pdf_sha256": "3e9a2eb5c64b7be06b8522fa6fc5b4e7830b49a03565445e6c0a60dc034e56af", "paper_pdf_bytes": 794508, "paper_pdf_source": "openreview", "code_url": "https://github.com/emanuele-moscato/tree-language-paper-submission", "code_repository": "emanuele-moscato/tree-language-paper-submission", "code_commit": "3dd5853c96aa9cdba4c943dc49c9ad51e080201e", "code_archive": "repos/F0Zd3knG9j.zip", "code_archive_sha256": "9714d543d1100ad0ad2bef0cab7e2381d31bcd5ac371c8804c9f95797b1bf29c", "code_archive_bytes": 43241, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 33, "github_languages": {"Python": 134562}, "github_archived": false, "github_pushed_at": "2024-12-09T16:02:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-transformers-learn-structured-data"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "L9kwewFGQZ", "year": 2024, "status": "rejected", "title": "Mitigating Interference in the Knowledge Continuum through Attention-Guided Incremental Learning", "authors": ["Prashant Shivaram Bhat", "Bharath Chennamkulam Renjith", "Bahram Zonooz", "Elahe Arani"], "authorids": ["~Prashant_Shivaram_Bhat1", "~Bharath_Chennamkulam_Renjith1", "~Bahram_Zonooz1", "~Elahe_Arani1"], "authors_source": "OpenReview API", "abstract": "Continual learning (CL) remains a significant challenge for deep neural networks, as it is prone to forgetting previously acquired knowledge. Several approaches have been proposed in the literature, such as experience rehearsal, regularization, and parameter isolation, to address this problem. Although almost zero forgetting can be achieved in task-incremental learning, class-incremental learning remains highly challenging due to the problem of inter-task class separation. Limited access to previous task data makes it difficult to discriminate between classes of current and previous tasks. To address this issue, we propose `Attention-Guided Incremental Learning' (AGILE), a novel rehearsal-based CL approach that incorporates compact task-attention to effectively reduce interference between tasks. AGILE utilizes lightweight, learnable task projection vectors to transform the latent representations of a shared task-attention module toward task distribution. Through extensive empirical evaluation we show that AGILE significantly improves generalization performance by mitigating task interference and outperforms rehearsal-based approaches in several CL scenarios. Furthermore AGILE can scale well to a large number of tasks with minimal overhead while remaining well-calibrated with reduced task-recency bias.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "houAM6uWre", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7288/Reviewer_n4fn"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper introduces a rehearsal-based method called AGILE to tackle the class-incremental learning setting in continual learning. Specifically, the paper leverages learnable task embedding vectors and shared task-attention module for better mitigating task interference. Experimental results on benchmark datasets demonstrate the effectiveness of the method.", "review_text": "The paper introduces a rehearsal-based method called AGILE to tackle the class-incremental learning setting in continual learning. Specifically, the paper leverages learnable task embedding vectors and shared task-attention module for better mitigating task interference. Experimental results on benchmark datasets demonstrate the effectiveness of the method.", "strengths": "- The paper reads well and is easy to follow.\n- Class-incremental learning is indeed a more challenging setting than task-incremental learning.", "weaknesses": "- The idea of using task-attention or task embedding vector is not quite novel. For example, DyTox [1] also has a task attention module, L2P [2] leverages task-specific prompts. \n- Following the first one, I think the paper misses several recent competitive methods to compare against. For example, I understand both DyTox and L2P are based on transformers. However, if the proposed method AGILE is generalizable enough, it should be compatible with transformer architectures as well, making comparison with more advance methods like DyTox, L2P possible.\n- \n- The contents in middle and right subfigures in figure 3 seems missing?\n\n[1] Douillard, Arthur, et al. \"Dytox: Transformers for continual learning with dynamic token expansion.\" CVPR 2022\n[2] Wang, Zifeng, et al. \"Learning to prompt for continual learning.\" CVPR 2022", "questions": "- I understand the method is based on rehearsal, what if the rehearsal part is removed. Will the remaining design lead to improvement upon the baselines without rehearsal as well?\n- See weaknesses for the rest questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a rehearsal-based method called AGILE to tackle the class-incremental learning setting in continual learning. Specifically, the paper leverages learnable task embedding vectors and shared task-attention module for better mitigating task interference. Experimental results on benchmark datasets demonstrate the effectiveness of the method.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper reads well and is easy to follow.\n- Class-incremental learning is indeed a more challenging setting than task-incremental learning.", "weaknesses": "- The idea of using task-attention or task embedding vector is not quite novel. For example, DyTox [1] also has a task attention module, L2P [2] leverages task-specific prompts. \n- Following the first one, I think the paper misses several recent competitive methods to compare against. For example, I understand both DyTox and L2P are based on transformers. However, if the proposed method AGILE is generalizable enough, it should be compatible with transformer architectures as well, making comparison with more advance methods like DyTox, L2P possible.\n- \n- The contents in middle and right subfigures in figure 3 seems missing?\n\n[1] Douillard, Arthur, et al. \"Dytox: Transformers for continual learning with dynamic token expansion.\" CVPR 2022\n[2] Wang, Zifeng, et al. \"Learning to prompt for continual learning.\" CVPR 2022", "questions": "- I understand the method is based on rehearsal, what if the rehearsal part is removed. Will the remaining design lead to improvement upon the baselines without rehearsal as well?\n- See weaknesses for the rest questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699088339304}, {"id": "Hg0CPXykTp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7288/Reviewer_pnnH"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a novel rehearsal based continual learning approach which use a shared task-attention module to mitigate the task interference. The shared task-attention module compresses the task specific information to some trainable parameters.", "review_text": "This paper introduces a novel rehearsal based continual learning approach which use a shared task-attention module to mitigate the task interference. The shared task-attention module compresses the task specific information to some trainable parameters.", "strengths": "1. The framework achieves fairly good results compared with baselines.\n2. The paper is written clearly and easy to follow.", "weaknesses": "1. Novelty concern. I would like to point out that the idea of leveraging trainable parameters to store task information has been investigated in previous works [*] [**]. L2P has shown its effectiveness in continual learning areas in recent years. \n\n2. Lack of a comprehensive comparison. There are many works using prompting (learnable parameters) in continual learning and achieving SOTA performance. I suggest the author conduct a comprehensive comparison with these works.\n\n[*] Learning to prompt for continual learning, CVPR 2022.\n\n[**] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, ECCV 2022.", "questions": "Could the author conduct a comprehensive comparison with CL works using prompting (learnable parameters)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel rehearsal based continual learning approach which use a shared task-attention module to mitigate the task interference. The shared task-attention module compresses the task specific information to some trainable parameters.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The framework achieves fairly good results compared with baselines.\n2. The paper is written clearly and easy to follow.", "weaknesses": "1. Novelty concern. I would like to point out that the idea of leveraging trainable parameters to store task information has been investigated in previous works [*] [**]. L2P has shown its effectiveness in continual learning areas in recent years. \n\n2. Lack of a comprehensive comparison. There are many works using prompting (learnable parameters) in continual learning and achieving SOTA performance. I suggest the author conduct a comprehensive comparison with these works.\n\n[*] Learning to prompt for continual learning, CVPR 2022.\n\n[**] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, ECCV 2022.", "questions": "Could the author conduct a comprehensive comparison with CL works using prompting (learnable parameters)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698815603338}, {"id": "HgikmoaCSr", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7288/Reviewer_5d1d"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a replay-based CL method utilizing a lightweight task attention module. The module receives features from the feature extractor and performs task-id prediction using the projection vectors for each task. This approach aligns with the findings of a prior theoretical study. The authors conduct comprehensive experiments to demonstrate the benefits of their approach compared to existing baselines and show the effectiveness of the proposed techniques.", "review_text": "The paper proposes a replay-based CL method utilizing a lightweight task attention module. The module receives features from the feature extractor and performs task-id prediction using the projection vectors for each task. This approach aligns with the findings of a prior theoretical study. The authors conduct comprehensive experiments to demonstrate the benefits of their approach compared to existing baselines and show the effectiveness of the proposed techniques.", "strengths": "1. The proposed approach is grounded in a theoretical study.\n2. The proposed method outperforms the baselines.", "weaknesses": "1. I feel like the paper is written in a rush. The experiment setup is not mentioned in the main paper. It's not clear how many tasks are used in the sequential data (e.g., Seq-CIFAR100), and what architecture is used. I couldn't find where I can find the information in the main text.\n2. It's not clear why the shared task-attention module improves WP and TP when this module itself also suffers from forgetting.\n3. I couldn't fully understand why this method is better than the existing task-id prediction methods. [1] also builds a task-id prediction module on top of the feature extractor. A more comprehensive and detailed discussion should be included.\n\nOverall, I think this approach is promising, but needs some improvements.\n\n[1] Conditional channel gated networks for task-aware continual learning", "questions": "1. How does the model make the final class prediction? Does it first predict the task-id using the attention module and make a within-task prediction?\n2. What's the purpose of using the task projection vectors and why is it used to compute both z_s and z_tp?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a replay-based CL method utilizing a lightweight task attention module. The module receives features from the feature extractor and performs task-id prediction using the projection vectors for each task. This approach aligns with the findings of a prior theoretical study. The authors conduct comprehensive experiments to demonstrate the benefits of their approach compared to existing baselines and show the effectiveness of the proposed techniques.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The proposed approach is grounded in a theoretical study.\n2. The proposed method outperforms the baselines.", "weaknesses": "1. I feel like the paper is written in a rush. The experiment setup is not mentioned in the main paper. It's not clear how many tasks are used in the sequential data (e.g., Seq-CIFAR100), and what architecture is used. I couldn't find where I can find the information in the main text.\n2. It's not clear why the shared task-attention module improves WP and TP when this module itself also suffers from forgetting.\n3. I couldn't fully understand why this method is better than the existing task-id prediction methods. [1] also builds a task-id prediction module on top of the feature extractor. A more comprehensive and detailed discussion should be included.\n\nOverall, I think this approach is promising, but needs some improvements.\n\n[1] Conditional channel gated networks for task-aware continual learning", "questions": "1. How does the model make the final class prediction? Does it first predict the task-id using the attention module and make a within-task prediction?\n2. What's the purpose of using the task projection vectors and why is it used to compute both z_s and z_tp?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698809383244}, {"id": "Bdg5u3znlF", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7288/Reviewer_2y3j"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Inspired by the notion that most methods that work in a task-incremental scenario can achieve almost zero forgetting, the authors introduce AGILE (Attention-Guided Incremental Learning). The main idea is to break down a class incremental problem into two sub-problems: Task-ID prediction (TP) and within-task prediction (WP). Once the first one is solved, the problem can be treated as a Task-Incremental, as the predicted task-id is already available. The authors suggest using task-specific projections to condition the feature vector. This conditioned vector passes through a task-specific module: task prediction and feature importance. During inference, the output of each module is concatenated to obtain the prediction. The authors demonstrate good performance in both task and class incremental scenarios.", "review_text": "Inspired by the notion that most methods that work in a task-incremental scenario can achieve almost zero forgetting, the authors introduce AGILE (Attention-Guided Incremental Learning). The main idea is to break down a class incremental problem into two sub-problems: Task-ID prediction (TP) and within-task prediction (WP). Once the first one is solved, the problem can be treated as a Task-Incremental, as the predicted task-id is already available. The authors suggest using task-specific projections to condition the feature vector. This conditioned vector passes through a task-specific module: task prediction and feature importance. During inference, the output of each module is concatenated to obtain the prediction. The authors demonstrate good performance in both task and class incremental scenarios.", "strengths": "- The authors work under the assumption that the incremental Class problem can be transformed into a task-incremental problem.\n    - However, I can't entirely agree that this is a \"necessary and sufficient\" solution. In fact, there is a probability that working the problem in this way helps the model lose generalization in the representations it generates, and the only reason why this does not happen in the proposed solution is that they use a buffer to store previous tasks.\n    - Even so it is a problem that is not widely attacked, but that can be a good option in many cases, especially if it's motivated by the idea of GWT.\n- The approach comprises many different components that have a good synergy between them. It is beneficial that the authors add Table 2 to show the importance of each loss.", "weaknesses": "- Using EMA is a critical point in the proposal, and the authors do not mention it too much. EMA can also be used to reduce weight modification, meaning that it can mitigate forgetting with a favorable beta. The authors present it to increase generalization.\n    - Experiments showing evidence that it increases generalization could help mitigate the doubts.\n    - Did you have an analysis of the beta value? \n- It is challenging to understand where there are linear layers and where there is soft attention in the proposed methods. The image does not help.\n    - It could be helpful to decrease the amount of terms, names or losses used in the explanation.\n    - For example, from the Figure, one can assume that there is one Task-Attention Module for each task. However, the Task-Attention Module is shared, no?\n- Didn’t find Definition 1 and 2.", "questions": "- Is EMA used in every method for Table 1? Or just AGILE?\n- How much overhead in terms of time is added when adding a Task-Attention Module?\n    - Even if the Task-Attention module is shared, it must still be used independently for each task.\n- Are you familiar with the work called Bias Correction (BiC) in Continual Learning? \n    - There are some similarities that you can find interesting.\n    - I don’t remember if it works in class or task-incremental, but there have been extensions that work in class-incremental settings.\n- Do you know how your proposal scales with the memory size? I have seen methods that scale well (such as DER), but others could be better (like iCarl).\n- Have you tried this approach with a fixed pre-trained model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Inspired by the notion that most methods that work in a task-incremental scenario can achieve almost zero forgetting, the authors introduce AGILE (Attention-Guided Incremental Learning). The main idea is to break down a class incremental problem into two sub-problems: Task-ID prediction (TP) and within-task prediction (WP). Once the first one is solved, the problem can be treated as a Task-Incremental, as the predicted task-id is already available. The authors suggest using task-specific projections to condition the feature vector. This conditioned vector passes through a task-specific module: task prediction and feature importance. During inference, the output of each module is concatenated to obtain the prediction. The authors demonstrate good performance in both task and class incremental scenarios.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The authors work under the assumption that the incremental Class problem can be transformed into a task-incremental problem.\n    - However, I can't entirely agree that this is a \"necessary and sufficient\" solution. In fact, there is a probability that working the problem in this way helps the model lose generalization in the representations it generates, and the only reason why this does not happen in the proposed solution is that they use a buffer to store previous tasks.\n    - Even so it is a problem that is not widely attacked, but that can be a good option in many cases, especially if it's motivated by the idea of GWT.\n- The approach comprises many different components that have a good synergy between them. It is beneficial that the authors add Table 2 to show the importance of each loss.", "weaknesses": "- Using EMA is a critical point in the proposal, and the authors do not mention it too much. EMA can also be used to reduce weight modification, meaning that it can mitigate forgetting with a favorable beta. The authors present it to increase generalization.\n    - Experiments showing evidence that it increases generalization could help mitigate the doubts.\n    - Did you have an analysis of the beta value? \n- It is challenging to understand where there are linear layers and where there is soft attention in the proposed methods. The image does not help.\n    - It could be helpful to decrease the amount of terms, names or losses used in the explanation.\n    - For example, from the Figure, one can assume that there is one Task-Attention Module for each task. However, the Task-Attention Module is shared, no?\n- Didn’t find Definition 1 and 2.", "questions": "- Is EMA used in every method for Table 1? Or just AGILE?\n- How much overhead in terms of time is added when adding a Task-Attention Module?\n    - Even if the Task-Attention module is shared, it must still be used independently for each task.\n- Are you familiar with the work called Bias Correction (BiC) in Continual Learning? \n    - There are some similarities that you can find interesting.\n    - I don’t remember if it works in class or task-incremental, but there have been extensions that work in class-incremental settings.\n- Do you know how your proposal scales with the memory size? I have seen methods that scale well (such as DER), but others could be better (like iCarl).\n- Have you tried this approach with a fixed pre-trained model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698692027617}, {"id": "zEBVau8OqW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7288/Reviewer_zq89"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper focuses on mitigating task interference in continual learning by introducing a compact task-attention module. It incorporates a set of lightweight, learnable task projection vectors, equal in number to the tasks, which transform the latent representations of a shared task-attention module into task-specific distributions. Additionally, this approach aims to enhance the model's performance in continual learning by jointly addressing the challenges of within-task and task-id prediction.", "review_text": "This paper focuses on mitigating task interference in continual learning by introducing a compact task-attention module. It incorporates a set of lightweight, learnable task projection vectors, equal in number to the tasks, which transform the latent representations of a shared task-attention module into task-specific distributions. Additionally, this approach aims to enhance the model's performance in continual learning by jointly addressing the challenges of within-task and task-id prediction.", "strengths": "The approach presented in this paper differs significantly from previous methods by combining a task-attention mechanism with minimal memory overhead. It explores the feasibility of reducing interference between tasks and surpasses rehearsal-based approaches in several continual learning scenarios.", "weaknesses": "A single lightweight task-specific vector may not be sufficient to adequately represent and distinguish the crucial information among multiple tasks. This approach may not effectively address the issue of catastrophic forgetting.", "questions": "1)\tThis method is less innovative and mainly focuses on solving the task interference problem. How to weigh the importance of solving the interference problem or solving the forgetting problem in continual learning?\n2)\tThe innovation in this paper is that the task-attention module is used to solve the task-id prediction problem, and within-task prediction problem how can it be solved efficiently?\n3)\tAs the number of tasks continues to grow, is there any interference or conflict between these lightweight task-specific vectors?\n4)\tCan this method be used in other continual learning scenarios, such as Task- free scenario?\n5)\tPlease provide attention-guided visualization experiments showing what the task-specific vector makes the model pay attention to.\n6)\tIn section 3.4 only the extension of the classifiers was carried out, what exactly does the network extension refer to?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on mitigating task interference in continual learning by introducing a compact task-attention module. It incorporates a set of lightweight, learnable task projection vectors, equal in number to the tasks, which transform the latent representations of a shared task-attention module into task-specific distributions. Additionally, this approach aims to enhance the model's performance in continual learning by jointly addressing the challenges of within-task and task-id prediction.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The approach presented in this paper differs significantly from previous methods by combining a task-attention mechanism with minimal memory overhead. It explores the feasibility of reducing interference between tasks and surpasses rehearsal-based approaches in several continual learning scenarios.", "weaknesses": "A single lightweight task-specific vector may not be sufficient to adequately represent and distinguish the crucial information among multiple tasks. This approach may not effectively address the issue of catastrophic forgetting.", "questions": "1)\tThis method is less innovative and mainly focuses on solving the task interference problem. How to weigh the importance of solving the interference problem or solving the forgetting problem in continual learning?\n2)\tThe innovation in this paper is that the task-attention module is used to solve the task-id prediction problem, and within-task prediction problem how can it be solved efficiently?\n3)\tAs the number of tasks continues to grow, is there any interference or conflict between these lightweight task-specific vectors?\n4)\tCan this method be used in other continual learning scenarios, such as Task- free scenario?\n5)\tPlease provide attention-guided visualization experiments showing what the task-specific vector makes the model pay attention to.\n6)\tIn section 3.4 only the extension of the classifiers was carried out, what exactly does the network extension refer to?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698646047360}], "openreview_url": "https://openreview.net/forum?id=L9kwewFGQZ", "arxiv_id": "2405.13978", "paper_pdf": "papers/L9kwewFGQZ.pdf", "paper_pdf_sha256": "b9b22608a98e83717b0e283c93282db4b1a4f1b6baa5f8ad90122246f5b570b1", "paper_pdf_bytes": 816257, "paper_pdf_source": "openreview", "code_url": "https://github.com/NeurAI-Lab/AGILE", "code_repository": "NeurAI-Lab/AGILE", "code_commit": "40220bf9ff2cf88fb785354b3a49e32c7e5eb328", "code_archive": "repos/L9kwewFGQZ.zip", "code_archive_sha256": "d10c038b2d94112ca5d9c62ddd0a876328547c2a17b34eb9524c96ab15fee590", "code_archive_bytes": 37059, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 29, "github_languages": {"Python": 107042}, "github_archived": false, "github_pushed_at": "2024-05-27T12:30:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mitigating-interference-in-the-knowledge"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Om_QvnjjBL2", "year": 2023, "status": "rejected", "title": "D2Match: Leveraging Deep Learning and Degeneracy for  Subgraph Matching", "authors": ["Xuanzhou Liu", "Lin Zhang", "Jiaqi Sun", "Yujiu Yang", "Haiqin Yang"], "authorids": ["~Xuanzhou_Liu1", "~Lin_Zhang9", "~Jiaqi_Sun1", "~Yujiu_Yang2", "~Haiqin_Yang2"], "authors_source": "OpenReview API", "abstract": "Subgraph matching is a fundamental building block for many graph-based applications and is challenging due to its high-order combinatorial nature.  However, previous methods usually tackle it by combinatorial optimization or representation learning and suffer from exponential computational cost or matching without theoretical guarantees.  In this paper, we develop D2Match by leveraging the efficiency of Deep learning and Degeneracy for subgraph matching. More specifically, we prove that subgraph matching can degenerate to subtree matching, and subsequently is equivalent to finding a perfect matching on a bipartite graph.  This matching procedure can be implemented by the built-in tree-structured aggregation mechanism on graph neural networks, which yields linear time complexity.  Moreover, circle structures, abstracted as {\\em supernodes}, and node attributes can be easily incorporated in D2Match to boost the matching. Finally, we conduct extensive experiments to show the superior performance of our D2Match and confirm that our D2Match indeed tries to exploit the subtrees and differs from existing learning-based subgraph matching methods that depend on memorizing the data distribution divergence.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "gNRYCE6fpa9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper545/Reviewer_9bfF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a statement that subgraph matching can be degenerated to subtree matching and provide proof. Based on this degeneracy property, the authors propose a matching method that utilizes GNN to model subtrees. Also, they adopt GNN to learn node representations to boost the performance of matching.", "review_text": "The paper proposes a novel idea on the degeneracy property of subgraph matching, and the tree-degeneration fits the way of aggregation of GNN. The theoretical results seem novel and interesting. Empirical results show the shortcomings of GNN, which is significant for future research. ", "strengths": "**Strength:**\n1. The idea of degenerating subgraph matching to subtree matching is novel. The theoretical study is interesting, regarding existing graph matching learning papers are usually empirical.\n2. The proposed method improves performances on 7 experimented datasets, some by a substantial margin.\n3. The empirical results show that GNNs tend to capture the data distribution divergence, and have bad performance on evenly distributed data, motivating further research.\n\n**Weakness:**\n1. The degeneracy property can be broken by simply inserting or removing an edge, so the model can be easily attacked and may not be robust to noise.\n2. The ablation study of GNN shows that the improvement is mainly from the GNN module. Also, the paper does not show the subtree modeling's effect on performance improvement, such as demonstrating recall on experimented datasets.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose a statement that subgraph matching can be degenerated to subtree matching and provide proof. Based on this degeneracy property, the authors propose a matching method that utilizes GNN to model subtrees. Also, they adopt GNN to learn node representations to boost the performance of matching.", "strength_and_weaknesses": "**Strength:**\n1. The idea of degenerating subgraph matching to subtree matching is novel. The theoretical study is interesting, regarding existing graph matching learning papers are usually empirical.\n2. The proposed method improves performances on 7 experimented datasets, some by a substantial margin.\n3. The empirical results show that GNNs tend to capture the data distribution divergence, and have bad performance on evenly distributed data, motivating further research.\n\n**Weakness:**\n1. The degeneracy property can be broken by simply inserting or removing an edge, so the model can be easily attacked and may not be robust to noise.\n2. The ablation study of GNN shows that the improvement is mainly from the GNN module. Also, the paper does not show the subtree modeling's effect on performance improvement, such as demonstrating recall on experimented datasets.", "clarity,_quality,_novelty_and_reproducibility": "This work may not possess good reproducibility if the code is not released.", "summary_of_the_review": "The paper proposes a novel idea on the degeneracy property of subgraph matching, and the tree-degeneration fits the way of aggregation of GNN. The theoretical results seem novel and interesting. Empirical results show the shortcomings of GNN, which is significant for future research. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666541888546}, {"id": "IwOSXGt_Rd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper545/Reviewer_6wEv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This main contribution of this paper is that it provides an idea and corresponding theoretical proof about degenerating subgraph matching problem to subtree matching with the help of Graph Neural Network.", "review_text": "This study proposes an interesting method with theoretical proof. Experimental results are good. Running experiments on the suggested datasets will strengthen the paper. Writing can be improved. The paper lacks discussions about the application of the proposed method to graphs with continuous node/edge attributes.", "strengths": "Strength: Theoretical proof and support for each claims mentioned. \n\nWeakness:\n1. The datasets used in the experiments are relative small and old. It will be interesting to see the performance of the proposed method on more recent datasets, such as ogb (https://ogb.stanford.edu/)\n2. Training-test split may have large impact on the performance, e.g., on Mutag and Protein. Please specify the details (e.g., random seeds, how many time the 5-fold CV was run)\n3. It is unclear if the proposed method work for graphs containing edges/nodes with continuous attributes. \n\n\n\n\n\n\n\n\n\n\t From the examples above, they have the same chordless cycle ABC and ACD\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This main contribution of this paper is that it provides an idea and corresponding theoretical proof about degenerating subgraph matching problem to subtree matching with the help of Graph Neural Network.", "strength_and_weaknesses": "Strength: Theoretical proof and support for each claims mentioned. \n\nWeakness:\n1. The datasets used in the experiments are relative small and old. It will be interesting to see the performance of the proposed method on more recent datasets, such as ogb (https://ogb.stanford.edu/)\n2. Training-test split may have large impact on the performance, e.g., on Mutag and Protein. Please specify the details (e.g., random seeds, how many time the 5-fold CV was run)\n3. It is unclear if the proposed method work for graphs containing edges/nodes with continuous attributes. \n\n\n\n\n\n\n\n\n\n\t From the examples above, they have the same chordless cycle ABC and ACD\n", "clarity,_quality,_novelty_and_reproducibility": "The proposed idea is interesting and novel for certain types of graphs. \nWriting can be improved.\n1. The usage of the word “neighbor” for different scenarios at the same time made me confused. For example, in section 4: “Motivated by Hall’s marriage Theorem 3.1, we develop an efficient algorithm to address the perfect matching procedure. A straightforward solution is to randomly select a subset W from the given set of neighbors, N(q), and count whether the corresponding neighbors of W, i.e. N(W), have more elements than this subset. After repeating this process multiple times for all node pairs, we obtain a perfect matching when no instance violates the criterion.”\n2. Figure 1 perfect matching table is a bit confusing. \n", "summary_of_the_review": "This study proposes an interesting method with theoretical proof. Experimental results are good. Running experiments on the suggested datasets will strengthen the paper. Writing can be improved. The paper lacks discussions about the application of the proposed method to graphs with continuous node/edge attributes.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666479522748}, {"id": "Cm9PTEuOVjV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper545/Reviewer_763b"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper identifies a degenerated case of subgraph matching, in which it falls to subtree matching. Then, the matching procedure can be implemented by the built-in tree-structured aggregation mechanism on graph neural networks. The paper proposes a new deep learning method for approximate subgraph matching.", "review_text": "Overall, this paper outlines a novel approach for deep learning based subgraph matching. Although I agree with the novelty of the proposed idea, I feel the paper is kind of overclaiming the theoretical contribution of the paper. The paper also needs to evaluate the proposed approach more thoroughly, by including runtime analysis and comparing against exact graph-matching baselines. Finally, the writing of the paper should be improved.\nOverall, I think there is significant room for improving this paper.", "strengths": "Strength:\n- The proposed idea is novel.\n- The experimental results are strong.\n\nWeakness:\n- The writing of the paper should be improved.\n- Although the paper provides a complexity analysis, the paper should also provide a runtime analysis to compare with baseline approaches. Given that there is an extra operation of converting a graph to a subtree, I assume the proposed method is slower compared to other deep learning based approaches. However, this is not discussed in the paper.\n- Although I understand that this approach is proposed for approximate subgraph matching, I believe the paper should at least include some comparisons with exact subgraph matching methods. Please feel free to skip CO-based methods in settings where they are not applicable.\n- The paper is kind of overclaiming the theoretical contribution of the paper. The paper relaxes the sufficient and necessary conditions to that of necessary only. However, I didn't find a bound or analysis regarding the error introduced in this relaxation.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper identifies a degenerated case of subgraph matching, in which it falls to subtree matching. Then, the matching procedure can be implemented by the built-in tree-structured aggregation mechanism on graph neural networks. The paper proposes a new deep learning method for approximate subgraph matching.", "strength_and_weaknesses": "Strength:\n- The proposed idea is novel.\n- The experimental results are strong.\n\nWeakness:\n- The writing of the paper should be improved.\n- Although the paper provides a complexity analysis, the paper should also provide a runtime analysis to compare with baseline approaches. Given that there is an extra operation of converting a graph to a subtree, I assume the proposed method is slower compared to other deep learning based approaches. However, this is not discussed in the paper.\n- Although I understand that this approach is proposed for approximate subgraph matching, I believe the paper should at least include some comparisons with exact subgraph matching methods. Please feel free to skip CO-based methods in settings where they are not applicable.\n- The paper is kind of overclaiming the theoretical contribution of the paper. The paper relaxes the sufficient and necessary conditions to that of necessary only. However, I didn't find a bound or analysis regarding the error introduced in this relaxation.", "clarity,_quality,_novelty_and_reproducibility": "I find the proposed idea to be novel.\nHowever, the writing of the paper should be significantly improved.\nConcretely, \n- The figures are not clear. For example, Figure 1 has a very brief caption, making the figure very hard to understand.\n- Equation 2 is defined very vaguely. T_v^(l) is not defined, should we interpret it as a WL subtree? \n- Definition of \"perfect matching\" is confusing, \"a matching of a graph in which every node of the graph is incident to exactly one edge\". Based on the definition of graph matching, we always match a query graph with a target graph. How do we match within a single graph?", "summary_of_the_review": "Overall, this paper outlines a novel approach for deep learning based subgraph matching. Although I agree with the novelty of the proposed idea, I feel the paper is kind of overclaiming the theoretical contribution of the paper. The paper also needs to evaluate the proposed approach more thoroughly, by including runtime analysis and comparing against exact graph-matching baselines. Finally, the writing of the paper should be improved.\nOverall, I think there is significant room for improving this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666309980908}], "openreview_url": "https://openreview.net/forum?id=Om_QvnjjBL2", "arxiv_id": "2306.06380", "paper_pdf": "papers/Om_QvnjjBL2.pdf", "paper_pdf_sha256": "0717cc33b43b26a720732cc720ccac889aca81b3ceba7448a711e93be0d092a2", "paper_pdf_bytes": 427383, "paper_pdf_source": "openreview", "code_url": "https://github.com/XuanzhouLiu/D2Match-ICML23", "code_repository": "XuanzhouLiu/D2Match-ICML23", "code_commit": "b444f90a05d6b06bed59e0b82007c374b2562a39", "code_archive": "repos/Om_QvnjjBL2.zip", "code_archive_sha256": "c39a1a4d1635becf4af91014a3f0bfba1bc12276c784ec75061c785c2e4de3e4", "code_archive_bytes": 27478, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 33, "github_languages": {"Python": 130463}, "github_archived": false, "github_pushed_at": "2023-12-30T06:36:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/d2match-leveraging-deep-learning-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VQyHD2R3Aq", "year": 2022, "status": "rejected", "title": "SPIDE: A Purely Spike-based Method for Training Feedback Spiking Neural Networks", "authors": ["Mingqing Xiao", "Qingyan Meng", "Zongpeng Zhang", "Yisen Wang", "Zhouchen Lin"], "authorids": ["~Mingqing_Xiao1", "~Qingyan_Meng1", "~Zongpeng_Zhang1", "~Yisen_Wang1", "~Zhouchen_Lin1"], "authors_source": "OpenReview API", "abstract": "Spiking neural networks (SNNs) with event-based computation are promising brain-inspired models for energy-efficient applications on neuromorphic hardware. However, most supervised SNN training methods require complex computation or impractical neuron models, which hinders them from spike-based energy-efficient training. Among them, the recently proposed method, implicit differentiation on the equilibrium state (IDE), for training feedback SNNs is a promising way that is possible for generalization to locally spike-based learning with flexible network structures. In this paper, we study spike-based implicit differentiation on the equilibrium state (SPIDE) that extends the IDE method for supervised local learning with spikes, which could be possible for energy-efficient training on neuromorphic hardware. Specifically, we first introduce ternary spiking neuron couples that can realize ternary outputs with the common neuron model, and we prove that implicit differentiation can be solved by spikes based on this design. With this approach, the whole training procedure can be made as event-driven spike computation and weights are updated locally with two-stage average firing rates. Then to reduce the approximation error of spikes due to the finite simulation time steps, we propose to modify the resting membrane potential. Based on it, the average firing rate, when viewed as a stochastic estimator, achieves an unbiased estimation of iterative solution for implicit differentiation and the variance of this estimator is reduced. With these key components, we can train SNNs with either feedback or feedforward structures in a small number of time steps. Further, the firing sparsity during training demonstrates the great potential for energy efficiency. Meanwhile, even with these constraints, our trained models could still achieve competitive results on MNIST, CIFAR-10 and CIFAR-100. Our proposed method demonstrates the great potential for energy-efficient training of SNNs on neuromorphic hardware.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Ubrdawfds5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1314/Reviewer_FWiT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors extend the work Xiao et. al., 2021 NeurIPS to elaborate on the ideas of training SNN based on the equilibrium and address some of the shortcomings. \n", "review_text": "The authors attempt to support their conclusion with mathematical evidence, which improves the reliability of the quantitative part of the results. Yet every coin has two sides. At the same time, the math details cannot tell the upper bound and the potential of the proposed method, which is my main concern for the series of approaches. \n\nIn this paper, the authors extend the work Xiao et. al., 2021 NeurIPS to elaborate on the ideas of training SNN based on the equilibrium and address some of the shortcomings. Overall, this paper is clearly written and easy to follow. I list my comments and concerns below and am more than happy to discuss with the authors about these issues. \n\n1. I think the idea of approaching SNN training via equilibrium is interesting, but it requires more reasoning about why it has the potential to go beyond existing methods and what is its unique methodological contribution given Xiao et. al., 2021 NeurIPS. \n\n2. If I understand it correctly, the equilibrium basically says that an integrally convergent input would result in an integrally convergent average firing rate. Since the assumption satisfies Lipschitz's condition with L < 1, I feel like the prove could be simplified with some known ODE theorems on perturbation. \n\n3. The derivative on the equilibrium gives the dependence between the convergent $a^*$ and $x^*$. I think it is more related to backpropagation rather than the  Hebbian learning rule. The analog here is not evident. \n\n4. Since the whole framework is built on the equilibrium, it implicitly requires the dynamics to be stable. Thus, it seems that the latency $T_F$ cannot be extremely shortened to an efficient case where the dynamics are not stable across time. As the authors also pointed out, this paradigm is not compatible with setups that may amplify the norm and maps, e.g. the BN-layers in training. This shortcoming from assumption may limit the practical utility of the proposed methods. In practice, the comparison in Table 3 is not up-to-date. For example, Ref [1] and [2-3] significantly shorten the conversion and training latency and work well for big datasets. The experiments in the current work should at least extend to CIFAR-100 and ImageNet. It would be also great if the authors can additionally add the results on the spiking dataset like CIFAR-gesture and CIFAR10-DVS as well. \n\n5. in Figure 3, it seems that the accuracy has a jump around epoch 50. Can the author explain the reason? \n------------------------------------------------------------------------------------------------------------------------------------------------------\n[1] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, Shi Gu. A Free Lunch From ANN: Towards Efficient, Accurate Spiking Neural Networks Calibration. *Proceedings of the 38th International Conference on Machine Learning, PMLR 139:6316-6325, 2021.*\n\n[2] Fang, Wei, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. *Conference on Neural Information Processing Systems (2021). *\n\n[3] Wu, Jibin, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. *IEEE Transactions on Neural Networks and Learning Systems (2021).*\n\n------------------------------------------------------------------------------------------------------------------------------------------------------\nMy comments after the rebuttal: \n\nThe authors made a very attentive rebuttal and emphasized their contribution as a purely spike-based training approach. Yet my two major concerns still remain here. \n1. I cannot fully track every step in the updating formula to make sure that it is purely spiked-based. For example, does the calculation of (10) and (11) and the associated gradient calculation on page 7 still involve multiplication? If the general multiplication is somewhat supported by neuromorphic hardware, the other methods mentioned in the comparison would also work as well? \n2. If we accept the claim it is purely spike-based, we still have the question of whether the potential energy efficiency is worthy. For example, the performance on DVS-CIFAR-10 is ~15\\% less than the mentioned benchmarks on page 22. As the authors mentioned, the energy cost is related to both the latency and spiking rate. Thus it would be necessary to at least provide a comparison of the estimated cost with the relevant approaches and better to provide an implementation of at least the proposed method on proper hardware to support the purely-spike-based assertion. Also, it would be necessary to demonstrate that the proposed method can achieve at least comparable accuracy with Fang et. al., ICCV 2021 to support that the cost of changing to spike-based training is acceptable. \n\nBased on these two points, I tend to keep my rating at 5 but am open to marginally accepting this paper if the other reviewers strongly think it should be accepted. \n------------------------------------------------------------------------------------------------------------------------------------------------------\nLatest update: \n\nI encouragingly increased my score to 6 given the rebuttal effort but my judgment remains on the borderline as the major concerns are not fully addressed with sufficient numeric proof. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors extend the work Xiao et. al., 2021 NeurIPS to elaborate on the ideas of training SNN based on the equilibrium and address some of the shortcomings. \n", "main_review": "The authors attempt to support their conclusion with mathematical evidence, which improves the reliability of the quantitative part of the results. Yet every coin has two sides. At the same time, the math details cannot tell the upper bound and the potential of the proposed method, which is my main concern for the series of approaches. \n\nIn this paper, the authors extend the work Xiao et. al., 2021 NeurIPS to elaborate on the ideas of training SNN based on the equilibrium and address some of the shortcomings. Overall, this paper is clearly written and easy to follow. I list my comments and concerns below and am more than happy to discuss with the authors about these issues. \n\n1. I think the idea of approaching SNN training via equilibrium is interesting, but it requires more reasoning about why it has the potential to go beyond existing methods and what is its unique methodological contribution given Xiao et. al., 2021 NeurIPS. \n\n2. If I understand it correctly, the equilibrium basically says that an integrally convergent input would result in an integrally convergent average firing rate. Since the assumption satisfies Lipschitz's condition with L < 1, I feel like the prove could be simplified with some known ODE theorems on perturbation. \n\n3. The derivative on the equilibrium gives the dependence between the convergent $a^*$ and $x^*$. I think it is more related to backpropagation rather than the  Hebbian learning rule. The analog here is not evident. \n\n4. Since the whole framework is built on the equilibrium, it implicitly requires the dynamics to be stable. Thus, it seems that the latency $T_F$ cannot be extremely shortened to an efficient case where the dynamics are not stable across time. As the authors also pointed out, this paradigm is not compatible with setups that may amplify the norm and maps, e.g. the BN-layers in training. This shortcoming from assumption may limit the practical utility of the proposed methods. In practice, the comparison in Table 3 is not up-to-date. For example, Ref [1] and [2-3] significantly shorten the conversion and training latency and work well for big datasets. The experiments in the current work should at least extend to CIFAR-100 and ImageNet. It would be also great if the authors can additionally add the results on the spiking dataset like CIFAR-gesture and CIFAR10-DVS as well. \n\n5. in Figure 3, it seems that the accuracy has a jump around epoch 50. Can the author explain the reason? \n------------------------------------------------------------------------------------------------------------------------------------------------------\n[1] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, Shi Gu. A Free Lunch From ANN: Towards Efficient, Accurate Spiking Neural Networks Calibration. *Proceedings of the 38th International Conference on Machine Learning, PMLR 139:6316-6325, 2021.*\n\n[2] Fang, Wei, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. *Conference on Neural Information Processing Systems (2021). *\n\n[3] Wu, Jibin, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. *IEEE Transactions on Neural Networks and Learning Systems (2021).*\n\n------------------------------------------------------------------------------------------------------------------------------------------------------\nMy comments after the rebuttal: \n\nThe authors made a very attentive rebuttal and emphasized their contribution as a purely spike-based training approach. Yet my two major concerns still remain here. \n1. I cannot fully track every step in the updating formula to make sure that it is purely spiked-based. For example, does the calculation of (10) and (11) and the associated gradient calculation on page 7 still involve multiplication? If the general multiplication is somewhat supported by neuromorphic hardware, the other methods mentioned in the comparison would also work as well? \n2. If we accept the claim it is purely spike-based, we still have the question of whether the potential energy efficiency is worthy. For example, the performance on DVS-CIFAR-10 is ~15\\% less than the mentioned benchmarks on page 22. As the authors mentioned, the energy cost is related to both the latency and spiking rate. Thus it would be necessary to at least provide a comparison of the estimated cost with the relevant approaches and better to provide an implementation of at least the proposed method on proper hardware to support the purely-spike-based assertion. Also, it would be necessary to demonstrate that the proposed method can achieve at least comparable accuracy with Fang et. al., ICCV 2021 to support that the cost of changing to spike-based training is acceptable. \n\nBased on these two points, I tend to keep my rating at 5 but am open to marginally accepting this paper if the other reviewers strongly think it should be accepted. \n------------------------------------------------------------------------------------------------------------------------------------------------------\nLatest update: \n\nI encouragingly increased my score to 6 given the rebuttal effort but my judgment remains on the borderline as the major concerns are not fully addressed with sufficient numeric proof. \n\n", "summary_of_the_review": "Overall, this paper is clearly written and easy to follow, but the methods may not apply to the more complex scene with less latency. My concern is that this limitation is from its setup. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635888841617}, {"id": "cc2tMpSst0t", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1314/Reviewer_dKfz"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors proposed a method to train Spiking Neural Networks (SNN) with spike-based implicit differentiation on the equilibrium state.\nMain idea is to use a spike-triggered event instead of average firing rate to approximate implicit differentiation of Feedback Spiking Neural Networks (FSNN).\nTo enable such idea and to further reduce the approximation errors, the authors proposed several techniques such as adopting ternary spiking neuron couples and shifting resting potential.\nThe experimental results showed that the proposed method can achieve high accuracy in several tasks such as MNIST, CIFAR-10, and CIFAR-100 with fewer time steps for training compared to existing methods.", "review_text": "This paper proposed a meaningful implementation of the spike-based learning method for SNNs with higher plausibility.\nThere have been several works on training SNNs with spike-based learning rules based on the back-propagation or the equilibrium propagation, but most of them failed to provide successful experimental results.\nIt is impressive that the experimental results in this paper showed successful accuracy in several datasets including CIFAR-100.\n\nThe writing can be improved for easier understanding.\nIt is hard to understand basic concepts without reading the reference, so It would be better to explain more details about the background such as the IDE.\nEven though I have not fully understood the mathematical derivations in the paper, it was not easy to follow mathematical derivations in the paper without referring to other papers.\n\n========== Comments after rebuttal ==========\n\nI apologize for the late response and thank the authors for adding a detailed explanation of the proposed idea. It was very helpful to deepen my understanding.\nI agree that this paper has a clear contribution of proposing a fully spike-based derivation of IDE training methods for SNNs.\n\nHowever, there remains another concern.\nI missed in my initial review that this paper solely used the IF neuron model. Can the proposed method also be applied to other spiking neuron models?\n\nEven though the IF neuron model is 'common' in many SNN literature, I strongly believe that the study on SNN is meaningful only when temporal information carried by spikes is utilized in information processing. Averaging out spikes with rate-coding with IF neuron is functionally identical to quantized non-spiking neural networks, in which the activation function has quantized output. Multiplication with linearly-quantized values (e.g. low-bit fixed-point numbers) is already cost-effective and can also be implemented only by accumulation. I don't think there is a justifiable reason to use SNNs instead of this counterpart which can simply use conventional hardware. Major benefits of neuromorphic hardware such as event-drivenness can also be replaced by the zero-skipping feature that is common in NN hardware. Even though the proposed training method is fully spike-based, with the IF neuron model, it seems not to use any temporal information of spikes.\n\nI would like to hear the authors' opinions on this. Thank you.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors proposed a method to train Spiking Neural Networks (SNN) with spike-based implicit differentiation on the equilibrium state.\nMain idea is to use a spike-triggered event instead of average firing rate to approximate implicit differentiation of Feedback Spiking Neural Networks (FSNN).\nTo enable such idea and to further reduce the approximation errors, the authors proposed several techniques such as adopting ternary spiking neuron couples and shifting resting potential.\nThe experimental results showed that the proposed method can achieve high accuracy in several tasks such as MNIST, CIFAR-10, and CIFAR-100 with fewer time steps for training compared to existing methods.", "main_review": "This paper proposed a meaningful implementation of the spike-based learning method for SNNs with higher plausibility.\nThere have been several works on training SNNs with spike-based learning rules based on the back-propagation or the equilibrium propagation, but most of them failed to provide successful experimental results.\nIt is impressive that the experimental results in this paper showed successful accuracy in several datasets including CIFAR-100.\n\nThe writing can be improved for easier understanding.\nIt is hard to understand basic concepts without reading the reference, so It would be better to explain more details about the background such as the IDE.\nEven though I have not fully understood the mathematical derivations in the paper, it was not easy to follow mathematical derivations in the paper without referring to other papers.\n\n========== Comments after rebuttal ==========\n\nI apologize for the late response and thank the authors for adding a detailed explanation of the proposed idea. It was very helpful to deepen my understanding.\nI agree that this paper has a clear contribution of proposing a fully spike-based derivation of IDE training methods for SNNs.\n\nHowever, there remains another concern.\nI missed in my initial review that this paper solely used the IF neuron model. Can the proposed method also be applied to other spiking neuron models?\n\nEven though the IF neuron model is 'common' in many SNN literature, I strongly believe that the study on SNN is meaningful only when temporal information carried by spikes is utilized in information processing. Averaging out spikes with rate-coding with IF neuron is functionally identical to quantized non-spiking neural networks, in which the activation function has quantized output. Multiplication with linearly-quantized values (e.g. low-bit fixed-point numbers) is already cost-effective and can also be implemented only by accumulation. I don't think there is a justifiable reason to use SNNs instead of this counterpart which can simply use conventional hardware. Major benefits of neuromorphic hardware such as event-drivenness can also be replaced by the zero-skipping feature that is common in NN hardware. Even though the proposed training method is fully spike-based, with the IF neuron model, it seems not to use any temporal information of spikes.\n\nI would like to hear the authors' opinions on this. Thank you.", "summary_of_the_review": "The proposed method effectively handles several obstacles while adopting the existing IDE method with spike-based implementation and provided convincing experimental results.\nHowever, I doubt that the theoretical improvements of this paper is significant enough especially with my poor understanding on the mathematical derivations in the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635871365859}, {"id": "OJVvBzgjjR3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1314/Reviewer_PJRp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper aims at porting the IDE method  into a spike-fbased and more bio-plausible version. The previous IDE used firing rates rather than spikes for computation, although reference Xiao et al in NeurIPS 2021 had already addressed implementations in spiking neural networks. The authors analyze the approximation error that results from solving implicit differentiation by spikes and report a solution based on ternary spiking neurons, that can be implemented with pairs of standard spiking neurons. They achieve in this way quite good performance for MNIST and CIFAR10.", "review_text": "If one has an application of this method for neuromorphic hardware in mind, it becomes essential that one does not arrive in a rate-based coding regime. It has been demonstrated, e.g. for Intel's Loihi chip in (Davies et al., 2021) that one hardly gets and energy advantage in spike-based hardware if one works in a rate-based coding regime. Hence I am missing a discussion of this issue, and methods for arriving at sparsely firing neurons. If one aims instead at biologically plausible models, one also would have to look for solutions with biologically realistic firing rates of a few Hz.\n\nAltogether I find it difficult to identify the specific innovations of this paper because of its inadequate literature review. In particular, there are already quite a number of spiking neural network solutions that achieve comparable or higher accuracies with comparable latency and number of parameters, see e.g. \n\nZhou S, Li X, Chen Y, et al. Temporal-Coded Deep Spiking Neural Network with Easy Training and Robust Performance[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(12): 11143-11151.\nFang W, Yu Z, Chen Y, et al. Incorporating learnable membrane time constant to enhance learning of spiking neural networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 2661-2671.\n\nThe abstract makes the really imprecise statement \"--most SNN training methods require complex computation or impractical neuron models\", which are hard to reconcile with publications such as\n\nBellec, G., Scherr, F., Subramoney, A., Hajek, E., Salaj, D., Legenstein, R., & Maass, W. (2020). A solution to the learning dilemma for recurrent networks of spiking neurons. Nature communications, 11(1), 1-15.\n\nThe authors should summarize and discuss more benefits of the proposed method. It is important for enhancing the impact of this work and drawing more audiences in the neuromorphic community.  In section 4.1, please explain why negative information is necessary for implicit differentiation calculation. \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper aims at porting the IDE method  into a spike-fbased and more bio-plausible version. The previous IDE used firing rates rather than spikes for computation, although reference Xiao et al in NeurIPS 2021 had already addressed implementations in spiking neural networks. The authors analyze the approximation error that results from solving implicit differentiation by spikes and report a solution based on ternary spiking neurons, that can be implemented with pairs of standard spiking neurons. They achieve in this way quite good performance for MNIST and CIFAR10.", "main_review": "If one has an application of this method for neuromorphic hardware in mind, it becomes essential that one does not arrive in a rate-based coding regime. It has been demonstrated, e.g. for Intel's Loihi chip in (Davies et al., 2021) that one hardly gets and energy advantage in spike-based hardware if one works in a rate-based coding regime. Hence I am missing a discussion of this issue, and methods for arriving at sparsely firing neurons. If one aims instead at biologically plausible models, one also would have to look for solutions with biologically realistic firing rates of a few Hz.\n\nAltogether I find it difficult to identify the specific innovations of this paper because of its inadequate literature review. In particular, there are already quite a number of spiking neural network solutions that achieve comparable or higher accuracies with comparable latency and number of parameters, see e.g. \n\nZhou S, Li X, Chen Y, et al. Temporal-Coded Deep Spiking Neural Network with Easy Training and Robust Performance[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(12): 11143-11151.\nFang W, Yu Z, Chen Y, et al. Incorporating learnable membrane time constant to enhance learning of spiking neural networks[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 2661-2671.\n\nThe abstract makes the really imprecise statement \"--most SNN training methods require complex computation or impractical neuron models\", which are hard to reconcile with publications such as\n\nBellec, G., Scherr, F., Subramoney, A., Hajek, E., Salaj, D., Legenstein, R., & Maass, W. (2020). A solution to the learning dilemma for recurrent networks of spiking neurons. Nature communications, 11(1), 1-15.\n\nThe authors should summarize and discuss more benefits of the proposed method. It is important for enhancing the impact of this work and drawing more audiences in the neuromorphic community.  In section 4.1, please explain why negative information is necessary for implicit differentiation calculation. \n\n\n", "summary_of_the_review": "The paper presents a nice step in an interesting direction. But it does not manage to clarify what exactly its innovations and possible applications are.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635780019325}], "openreview_url": "https://openreview.net/forum?id=VQyHD2R3Aq", "arxiv_id": "2302.00232", "paper_pdf": "papers/VQyHD2R3Aq.pdf", "paper_pdf_sha256": "132df82ffe8425d6421ac7f8e8699bfc4a10f94c4e7c6894ddca788deef1de6f", "paper_pdf_bytes": 484510, "paper_pdf_source": "openreview", "code_url": "https://github.com/pkuxmq/SPIDE-FSNN", "code_repository": "pkuxmq/SPIDE-FSNN", "code_commit": "54e4338d5d09a864909cdc284a02c8d73676e3c4", "code_archive": "repos/VQyHD2R3Aq.zip", "code_archive_sha256": "948fedfc0077c7ce3a0e5d4a48b3b83e941bdb9116874edc1c9bb40a344341e1", "code_archive_bytes": 50397, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 32, "github_languages": {"Python": 183260}, "github_archived": false, "github_pushed_at": "2023-02-02T03:29:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/spide-a-purely-spike-based-method-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1UtnrqVUeNE", "year": 2021, "status": "rejected", "title": "Detecting Misclassification Errors in Neural Networks with a Gaussian Process Model", "authors": ["Xin Qiu", "Risto Miikkulainen"], "authorids": ["~Xin_Qiu1", "~Risto_Miikkulainen1"], "authors_source": "OpenReview API", "abstract": "As neural network classifiers are deployed in real-world applications, it is crucial that their predictions are not just accurate, but trustworthy as well. One practical solution is to assign confidence scores to each prediction, then filter out low-confidence predictions. However, existing confidence metrics are not yet sufficiently reliable for this role. This paper presents a new framework that produces more reliable confidence scores for detecting misclassification errors. This framework, RED, calibrates the classifier's inherent confidence indicators and estimates uncertainty of the calibrated confidence scores using Gaussian Processes. Empirical comparisons with other confidence estimation methods on 125 UCI datasets demonstrate that this approach is effective. An experiment on a vision task with a large deep learning architecture further confirms that the method can scale up, and a case study involving out-of-distribution and adversarial samples shows potential of the proposed method to improve robustness of neural network classifiers more broadly in the future.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "zH59rd_2BGQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3666/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, their goal is to improve calibration and accuracy by augmenting a classification model with a GP. They base their model off RIO (ICLR 2020) which targets regression problems and tries to predict the residual between predicted value and true value. They propose a model, RED, which instead tries to predict the residual between the predicted confidence score for the true class and 1 — the true class target confidence score using a GP. They show strong improvements over the methods they compare to for 125 UCI datasets and CIFAR-10 dataset.\n\nI find the approach interesting though the novelty is incremental over the RIO paper. My main concern is that I think some additional methods need to be compared with. For example [1] uses a bayesian last layer which is something that should be compared with. Using an ensemble of single layer NNs for the last layer or using MC-dropout at test time (which is known to approximate Bayesian inference under certain conditions) would also be interesting.\n\n[1] “Scalable Bayesian Optimization Using Deep Neural Networks” by Snoek et al. \n[2] “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning” by Gal et al.\n\nEdit: Based on the author response in terms of adding additional experiments, I'm raising my score to a 6.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Key comparison methods missing", "review": "In this paper, their goal is to improve calibration and accuracy by augmenting a classification model with a GP. They base their model off RIO (ICLR 2020) which targets regression problems and tries to predict the residual between predicted value and true value. They propose a model, RED, which instead tries to predict the residual between the predicted confidence score for the true class and 1 — the true class target confidence score using a GP. They show strong improvements over the methods they compare to for 125 UCI datasets and CIFAR-10 dataset.\n\nI find the approach interesting though the novelty is incremental over the RIO paper. My main concern is that I think some additional methods need to be compared with. For example [1] uses a bayesian last layer which is something that should be compared with. Using an ensemble of single layer NNs for the last layer or using MC-dropout at test time (which is known to approximate Bayesian inference under certain conditions) would also be interesting.\n\n[1] “Scalable Bayesian Optimization Using Deep Neural Networks” by Snoek et al. \n[2] “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning” by Gal et al.\n\nEdit: Based on the author response in terms of adding additional experiments, I'm raising my score to a 6.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604105077680}, {"id": "WBMKrZ7hKvz", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3666/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Update: Following the authors' clarifications and additional experimental work, I'm increasing my rating to 6.\n\nThis paper proposes RED, a framework for detecting misclassification errors, based on regression of target confidence scores and application of a Gaussian process for uncertainty in predicted confidence scores. It builds upon RIO, a framework for predicting residuals of regression models and their uncertainties using GPs. Compared with other confidence metrics, RED aims for greater separability between correct and incorrect predictions.\n\nThe method is straightforward to implement and performs well against the baselines considered on classification tasks for 125 UCI datasets. However, I question whether the baselines are sufficient; it is not demonstrated whether RED would outperform other confidence scoring and OOD detection methods mentioned in the related work section, such as temperature scaling (or the related method ODIN, proposed in Liang, S., Li, Y., and Srikant, R., 2017. Enhancing the reliability of out-of-distribution image detection in neural networks.) or simply the entropy of the softmax predictions. Unless there is a good justification for the limited set of baselines, I believe the paper's claims to generality are limited.\n\nAdditionally, for the OOD detection results shown in Figure 3, why were AUROC and AUPRC not reported? While the scatterplots show separability of OOD data visually, these metrics (used elsewhere in the paper) would give a better indication of performance (and again, I think a greater range of baselines and tasks would be necessary to make any firm claims about OOD detection).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well-performing, simple to implement method for classification error detection; limited set of baselines may not establish generality", "review": "Update: Following the authors' clarifications and additional experimental work, I'm increasing my rating to 6.\n\nThis paper proposes RED, a framework for detecting misclassification errors, based on regression of target confidence scores and application of a Gaussian process for uncertainty in predicted confidence scores. It builds upon RIO, a framework for predicting residuals of regression models and their uncertainties using GPs. Compared with other confidence metrics, RED aims for greater separability between correct and incorrect predictions.\n\nThe method is straightforward to implement and performs well against the baselines considered on classification tasks for 125 UCI datasets. However, I question whether the baselines are sufficient; it is not demonstrated whether RED would outperform other confidence scoring and OOD detection methods mentioned in the related work section, such as temperature scaling (or the related method ODIN, proposed in Liang, S., Li, Y., and Srikant, R., 2017. Enhancing the reliability of out-of-distribution image detection in neural networks.) or simply the entropy of the softmax predictions. Unless there is a good justification for the limited set of baselines, I believe the paper's claims to generality are limited.\n\nAdditionally, for the OOD detection results shown in Figure 3, why were AUROC and AUPRC not reported? While the scatterplots show separability of OOD data visually, these metrics (used elsewhere in the paper) would give a better indication of performance (and again, I think a greater range of baselines and tasks would be necessary to make any firm claims about OOD detection).", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603953956121}, {"id": "yaY5db__tMY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3666/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper solves an interesting problem of predicting uncertainty in NN without re-raining/modifying the existing NN. The authors propose a framework to calculate a confidence score for detecting misclassification errors by calibrating the NN classifier’s confidence scores and estimates uncertainty around the calibrated scores using Gaussian processes. This framework is called RED (Residual i/o Error Detection). \n\nThis paper is also technically sound and to the best of my knowledge is novel and relevant to the community. \n\nIt would be good to apply SVGP directly to some of these datasets and compare the results against NN+SVGP results.\n\nYou use the term “calibrated” confidence score/prediction. Could you explain what do you mean by calibrated?\n\nI find the presentation of results very confusing. For example, in Table 1, AP-Error is smallest for the RED method and in Table 3 AP-error is the largest for the RED method. In both cases, it is mentioned that the RED method outperforms other methods. \n\nYou mentioned ConfidNet outperformed the MCP baseline by a margin of 0.42. I do not see this number on the table.\n\nIt would be good if the authors could mention in the paper what is RIO short for.\n\nYou mentioned that you need to extend the kernel to multiple output kernel. Could you explain a bit more about that and how you build it?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Adding confidence score to NN classifiers without retraining or modifying the model", "review": "This paper solves an interesting problem of predicting uncertainty in NN without re-raining/modifying the existing NN. The authors propose a framework to calculate a confidence score for detecting misclassification errors by calibrating the NN classifier’s confidence scores and estimates uncertainty around the calibrated scores using Gaussian processes. This framework is called RED (Residual i/o Error Detection). \n\nThis paper is also technically sound and to the best of my knowledge is novel and relevant to the community. \n\nIt would be good to apply SVGP directly to some of these datasets and compare the results against NN+SVGP results.\n\nYou use the term “calibrated” confidence score/prediction. Could you explain what do you mean by calibrated?\n\nI find the presentation of results very confusing. For example, in Table 1, AP-Error is smallest for the RED method and in Table 3 AP-error is the largest for the RED method. In both cases, it is mentioned that the RED method outperforms other methods. \n\nYou mentioned ConfidNet outperformed the MCP baseline by a margin of 0.42. I do not see this number on the table.\n\nIt would be good if the authors could mention in the paper what is RIO short for.\n\nYou mentioned that you need to extend the kernel to multiple output kernel. Could you explain a bit more about that and how you build it?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603841226951}, {"id": "8y86mQBwfLy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3666/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#######################################################\nSUMMARY\n\nThis paper introduces RED, a new methodology to produce reliable confidence scores to detect missclassification errors in neural networks. The idea is to combine kernels based on both input and output spaces (as in RIO) to define a (sparse) GP that estimates the residual between the correctness of the original prediction and the maximum class probability. The authors show enhanced performance against other related methods and the ability of RED to detect OOD and adversarial data through the variance of the confidence score. \n\n#####################################################\nPROS\n\n1) Obtaining confidence scores for neural network predictions is a timely and very relevant topic for the ICLR community, since it is one of the main limitations of real-world applications of current neural nets.\n\n2) The related literature review is clear and, to the best of my knowledge, the proposed metholody based on Gaussian Processes is novel. \n\n3) The experimental validation of the proposed method on the UCI datasets is strong. It uses a wide range of datasets and several statistical tests, and RED obtains superior performance.\n\n4) The idea of using the variance of the proposed confidence score to identify OOD and adversarial data is interesting and promising.\n\n######################################################\nCONS\n\n1) My main concern is that the contribution in RED can be regarded somehow incremental given the RIO approach. It utilizes the same rationale behind RIO, and just adapts the necessary components so that it works in classification. The adaptation of these components is also straightforward: the output kernel now works on several dimensions (instead of the scalar dimension of regression) and the target is now the correctness of the original prediction. \n\n2) The experimental validation focuses on several competitors which can be considered \"of the same family\" as the proposed approach. Namely, all of them calibrate the predictions of a pre-trained neural network. I think it would be interesting to also compare to a different \"family\" of methods. For instance, (Functional) Bayesian Neural Networks are meant to obtain calibrated predictions by leveraging epistemic uncertainty (that coming from the model parameters). \n\n3) I do not fully understand the relevance of the experiment with the large deep learning architecture given by the VGG16 model. Since the proposed method works on the pre-trained neural network, my understanding is that the complexity of the neural network itself is not relevant for the performance of the proposed approach. Also, in this experiment I miss several independent runs to assess the results variability. \n\n####################################\nAdditional questions/feedback:\n\n1) In the second paragraph of section 4.2., there seems to be a typo when reporting the margin. It is said 0.42 and 0.55 for ConfidNet and RED respectively, but I think it should be 0.042 and 0.055 by looking at Table 3.\n\n2) It is not entirely clear to me why the process described in section 4.3. (second paragraph) produces proper OOD and adversarial data. For instance, some of the intended OOD data could be similar to training data (specially because the latter is being normalized to mean 0 and std 1). And similarly for the adversarial case. I think this could be better explained.\n\n3) When it comes to real practice, a key decision is to set a threshold on the confidence score to decide what instances should be supervised by an expert. Is there any recommendation on this?\n\n####################################### \nAFTER REBUTTAL\n\nThe new baselines added make the experimental validation more convincing. Therefore, I have raised my rating to 6 (Marginally above the acceptance threshold). However, I still believe that the contribution is incremental, and I think the paper would gain in terms of novelty if it focused more on the detection of OOD data and adversarial attacks (which right now is more like a preliminary test).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting problem and results, although the approach seems a bit incremental", "review": "#######################################################\nSUMMARY\n\nThis paper introduces RED, a new methodology to produce reliable confidence scores to detect missclassification errors in neural networks. The idea is to combine kernels based on both input and output spaces (as in RIO) to define a (sparse) GP that estimates the residual between the correctness of the original prediction and the maximum class probability. The authors show enhanced performance against other related methods and the ability of RED to detect OOD and adversarial data through the variance of the confidence score. \n\n#####################################################\nPROS\n\n1) Obtaining confidence scores for neural network predictions is a timely and very relevant topic for the ICLR community, since it is one of the main limitations of real-world applications of current neural nets.\n\n2) The related literature review is clear and, to the best of my knowledge, the proposed metholody based on Gaussian Processes is novel. \n\n3) The experimental validation of the proposed method on the UCI datasets is strong. It uses a wide range of datasets and several statistical tests, and RED obtains superior performance.\n\n4) The idea of using the variance of the proposed confidence score to identify OOD and adversarial data is interesting and promising.\n\n######################################################\nCONS\n\n1) My main concern is that the contribution in RED can be regarded somehow incremental given the RIO approach. It utilizes the same rationale behind RIO, and just adapts the necessary components so that it works in classification. The adaptation of these components is also straightforward: the output kernel now works on several dimensions (instead of the scalar dimension of regression) and the target is now the correctness of the original prediction. \n\n2) The experimental validation focuses on several competitors which can be considered \"of the same family\" as the proposed approach. Namely, all of them calibrate the predictions of a pre-trained neural network. I think it would be interesting to also compare to a different \"family\" of methods. For instance, (Functional) Bayesian Neural Networks are meant to obtain calibrated predictions by leveraging epistemic uncertainty (that coming from the model parameters). \n\n3) I do not fully understand the relevance of the experiment with the large deep learning architecture given by the VGG16 model. Since the proposed method works on the pre-trained neural network, my understanding is that the complexity of the neural network itself is not relevant for the performance of the proposed approach. Also, in this experiment I miss several independent runs to assess the results variability. \n\n####################################\nAdditional questions/feedback:\n\n1) In the second paragraph of section 4.2., there seems to be a typo when reporting the margin. It is said 0.42 and 0.55 for ConfidNet and RED respectively, but I think it should be 0.042 and 0.055 by looking at Table 3.\n\n2) It is not entirely clear to me why the process described in section 4.3. (second paragraph) produces proper OOD and adversarial data. For instance, some of the intended OOD data could be similar to training data (specially because the latter is being normalized to mean 0 and std 1). And similarly for the adversarial case. I think this could be better explained.\n\n3) When it comes to real practice, a key decision is to set a threshold on the confidence score to decide what instances should be supervised by an expert. Is there any recommendation on this?\n\n####################################### \nAFTER REBUTTAL\n\nThe new baselines added make the experimental validation more convincing. Therefore, I have raised my rating to 6 (Marginally above the acceptance threshold). However, I still believe that the contribution is incremental, and I think the paper would gain in terms of novelty if it focused more on the detection of OOD data and adversarial attacks (which right now is more like a preliminary test).\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603029206523}], "openreview_url": "https://openreview.net/forum?id=1UtnrqVUeNE", "arxiv_id": "2010.02065", "paper_pdf": "papers/1UtnrqVUeNE.pdf", "paper_pdf_sha256": "b40f23c343100d080e8b44ff9015e6d01f0ed0bc910a2dbc4d6a0d411308d017", "paper_pdf_bytes": 639420, "paper_pdf_source": "openreview", "code_url": "https://github.com/cognizant-ai-labs/red-paper", "code_repository": "cognizant-ai-labs/red-paper", "code_commit": "d88c77808954b13d8f59ca6727517892756548c5", "code_archive": "repos/1UtnrqVUeNE.zip", "code_archive_sha256": "2c46a671611cf1362d3f2d93e06c2140fb9a8ae27dc4bcf3aca0e01ef184075b", "code_archive_bytes": 58294, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 38, "github_languages": {"Python": 254942}, "github_archived": false, "github_pushed_at": "2025-09-28T09:57:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/detecting-misclassification-errors-in-neural-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hye4WaVYwr", "year": 2020, "status": "rejected", "title": "Bootstrapping the Expressivity with Model-based Planning", "authors": ["Kefan Dong", "Yuping Luo", "Tengyu Ma"], "authorids": ["dkf16@mails.tsinghua.edu.cn", "yupingl@cs.princeton.edu", "tengyuma@cs.stanford.edu"], "authors_source": "OpenReview API", "abstract": "We compare the model-free reinforcement learning with the model-based approaches through the lens of the expressive power of neural networks for policies, $Q$-functions, and dynamics.  We show, theoretically and empirically, that even for one-dimensional continuous state space, there are many MDPs whose optimal $Q$-functions and policies are much more complex than the dynamics. We hypothesize many real-world MDPs also have a similar property. For these MDPs, model-based planning is a favorable algorithm, because the resulting policies can approximate the optimal policy significantly better than a neural network parameterization can, and model-free or model-based policy optimization rely on policy parameterization. Motivated by the theory, we apply a simple multi-step model-based bootstrapping planner (BOOTS) to bootstrap a weak $Q$-function into a stronger policy. Empirical results show that applying BOOTS on top of model-based or model-free policy optimization algorithms at the test time improves the performance on MuJoCo benchmark tasks. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJg89NBeqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper372/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper studies a theoretical aspect of the expressivity of policy, Q functions and dynamics. Based on theoretical and empirical analysis, the authors propose a new model-based RL algorithm that said to improve task performance. Final evaluations are demonstrated on MuJoCo benchmark tasks. \n\nOverall, the paper pursues an interesting and ambitious problem on the interplay between model-based and model-free approaches, and the expressivity of the representation of dynamics, policy and value functions. However, the results in the theory part is drawn based on analysis on a very simple and special task. Therefore the theoretical results can not be considered general for all MDP cases. In addition, these results are not surprising.\n\n- Theorem 4.3 states a general theoretical result that holds for neural networks, the proof does not look like it can hold with a universal representation power of a neural network. \n\n- The idea of using Q-functions estimate as Boostrapping is just an idea of using on-planning to improve action selection at every decision step. This is just a recurring idea of many model-based RL approaches. BOOTS consumes more computations as planning, hence would perform better than the baselines. BOOTS should be compared to other model-based approaches that also use planning at Testing, assume all are given the same budget of testing time.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary: This paper studies a theoretical aspect of the expressivity of policy, Q functions and dynamics. Based on theoretical and empirical analysis, the authors propose a new model-based RL algorithm that said to improve task performance. Final evaluations are demonstrated on MuJoCo benchmark tasks. \n\nOverall, the paper pursues an interesting and ambitious problem on the interplay between model-based and model-free approaches, and the expressivity of the representation of dynamics, policy and value functions. However, the results in the theory part is drawn based on analysis on a very simple and special task. Therefore the theoretical results can not be considered general for all MDP cases. In addition, these results are not surprising.\n\n- Theorem 4.3 states a general theoretical result that holds for neural networks, the proof does not look like it can hold with a universal representation power of a neural network. \n\n- The idea of using Q-functions estimate as Boostrapping is just an idea of using on-planning to improve action selection at every decision step. This is just a recurring idea of many model-based RL approaches. BOOTS consumes more computations as planning, hence would perform better than the baselines. BOOTS should be compared to other model-based approaches that also use planning at Testing, assume all are given the same budget of testing time."}, "tcdate": 1571996813597}, {"id": "SygSmcvaKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper372/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper highlights an interesting issue regarding approximability of function approximators (neural networks). The paper provides cases where the action value function is difficult to approximate and is much more difficult than the dynamics of a model. The author conducts some experiments to claim that even with a large NN, DQN still finds a suboptimal policy. Theorems regarding the appoximability of the action value function are presented. Then the paper proposes that rollout-based search should be preferred for planning and conducts some experiments to verify this. Although the paper points out interesting issues of approximating action-value function, both the motivation and the suggestion regarding MBRL are not convincing.\n\n1. In term of the motivation, those cases listed in the paper are interesting, but they are not representative. In fact, the Dynamics can be far more complicated and it is still an open problem regarding how to learn the Dynamics. Furthermore, the proposed method is to simply combine MCTS and bootstrap value estimates. The method itself is not novel and it basically down weights the bootstrap estimate. It is very intuitive that some appropriate combination between the two can yield better performance. However, in model-based setting, the reward sequence can be highly variant and non-stationary, there is no solid reason to believe this can be always better. The motivating experiments in figure 3 are not persuasive. There can be many reasons for a deep RL algorithm to find a suboptimal policy: boostrap target interference, overestimation, difficulty of optimization, etc. It is quite confusing which factor leads to suboptimal performance of DQN.\n \n2. Theorem 4.3 does not make sense to me. What does it mean by “no constant depth NN can approximate the optimal policy with near optimal rewards?” Notice that, in machine learning community, people rarely pursue perfect approximation (equality). As long as the approximation error can be reasonably small, the approximator should be still useful. Does “no constant depth NN can approximate the optimal policy” mean the approximation error is unbounded? I believe we can still expect a large NN to approximate the action-value function very well. \n\nA minor issue. In page 2, the paper writes “model-free RL or MB policy optimization suffer from …, whereas MB planning … ”. MB planning is not a separate category of MB policy optimization. Such statement is not accurate. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper highlights an interesting issue regarding approximability of function approximators (neural networks). The paper provides cases where the action value function is difficult to approximate and is much more difficult than the dynamics of a model. The author conducts some experiments to claim that even with a large NN, DQN still finds a suboptimal policy. Theorems regarding the appoximability of the action value function are presented. Then the paper proposes that rollout-based search should be preferred for planning and conducts some experiments to verify this. Although the paper points out interesting issues of approximating action-value function, both the motivation and the suggestion regarding MBRL are not convincing.\n\n1. In term of the motivation, those cases listed in the paper are interesting, but they are not representative. In fact, the Dynamics can be far more complicated and it is still an open problem regarding how to learn the Dynamics. Furthermore, the proposed method is to simply combine MCTS and bootstrap value estimates. The method itself is not novel and it basically down weights the bootstrap estimate. It is very intuitive that some appropriate combination between the two can yield better performance. However, in model-based setting, the reward sequence can be highly variant and non-stationary, there is no solid reason to believe this can be always better. The motivating experiments in figure 3 are not persuasive. There can be many reasons for a deep RL algorithm to find a suboptimal policy: boostrap target interference, overestimation, difficulty of optimization, etc. It is quite confusing which factor leads to suboptimal performance of DQN.\n \n2. Theorem 4.3 does not make sense to me. What does it mean by “no constant depth NN can approximate the optimal policy with near optimal rewards?” Notice that, in machine learning community, people rarely pursue perfect approximation (equality). As long as the approximation error can be reasonably small, the approximator should be still useful. Does “no constant depth NN can approximate the optimal policy” mean the approximation error is unbounded? I believe we can still expect a large NN to approximate the action-value function very well. \n\nA minor issue. In page 2, the paper writes “model-free RL or MB policy optimization suffer from …, whereas MB planning … ”. MB planning is not a separate category of MB policy optimization. Such statement is not accurate. \n"}, "tcdate": 1571809821467}, {"id": "Byx0j9KoKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper372/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a mainly theoretical argument comparing the expressivity of model-free and model-based RL methods contrary to analysis in the past which usually relies on sample complexity. They construct a family of MDPs, where the true dynamics belong to a simple function class (in terms of the number of linear pieces needed to define the function), but the corresponding optimal Q-function belongs to a function class not necessarily expressible by a simple function. The paper then builds a similar case for randomly/semi-randomly generated MDPs. Finally, they propose to bootstrap the Q-function with n-step returns to boost the expressivity exponentially. \n\nI would lean towards accepting this paper, as this paper looks at an interesting problem and their analysis seems to be rigorous and valuable for the community to build upon. My questions/comments are as follows:\n\n1. Previous work, for example [1], also talks about bootstrapping a neural net Q-function using an n-step approximation so as to improve the efficiency of a pure model-free algorithm. I think this paper should be cited.\n\n2. In terms of the experiments, it is hard to understand the significance and connection to the theory. The theory talks about the expressive power of Q-functions, which suggests that we should look at only the asymptotic performance on these tasks, but most of the results are similar to MBPO or SAC in terms of asymptotic performance, although with a different learning speed, which could have to do with different factors.\n\n3. This paper shows the existence and provides a constructive proof for a family of MDPs where expressing optimal Q-functions is exponentially harder than expressing dynamics. But how likely is such a setting to arise in MDPs in practice? For example, on the gym benchmarks, we would expect fairly not so complicated Q-functions -- although they might take longer to learn. \n\n4. There is also a divide between learning Q*, and learning a policy that optimizes Q* reasonably well. Further, the requirement for the Q-function is only to get relative ordering between Q-values for different actions at states visited by the optimal policy right, which might not be the same as the Q-function landscape across state-action pairs. So, I am not sure if the expressivity of the optimal Q-function, in general, is the right metric to answer this question, but I also completely agree that this is a good starting point. \n\n5. I think when applying RL to the current benchmarks or problems we have, the main problem could be linked to optimization of a neural-net Q-function via bootstrapping (in the sense of approximate dynamic programming) as compared to the expressive power of optimal Q-functions. But I agree that the problem being looked at in the paper would also exist.\n\nReferences:\n[1] Plan Online, Learn Offline: Efficient Learning and Exploration via Model-Based Control, Lowrey et.al.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper presents a mainly theoretical argument comparing the expressivity of model-free and model-based RL methods contrary to analysis in the past which usually relies on sample complexity. They construct a family of MDPs, where the true dynamics belong to a simple function class (in terms of the number of linear pieces needed to define the function), but the corresponding optimal Q-function belongs to a function class not necessarily expressible by a simple function. The paper then builds a similar case for randomly/semi-randomly generated MDPs. Finally, they propose to bootstrap the Q-function with n-step returns to boost the expressivity exponentially. \n\nI would lean towards accepting this paper, as this paper looks at an interesting problem and their analysis seems to be rigorous and valuable for the community to build upon. My questions/comments are as follows:\n\n1. Previous work, for example [1], also talks about bootstrapping a neural net Q-function using an n-step approximation so as to improve the efficiency of a pure model-free algorithm. I think this paper should be cited.\n\n2. In terms of the experiments, it is hard to understand the significance and connection to the theory. The theory talks about the expressive power of Q-functions, which suggests that we should look at only the asymptotic performance on these tasks, but most of the results are similar to MBPO or SAC in terms of asymptotic performance, although with a different learning speed, which could have to do with different factors.\n\n3. This paper shows the existence and provides a constructive proof for a family of MDPs where expressing optimal Q-functions is exponentially harder than expressing dynamics. But how likely is such a setting to arise in MDPs in practice? For example, on the gym benchmarks, we would expect fairly not so complicated Q-functions -- although they might take longer to learn. \n\n4. There is also a divide between learning Q*, and learning a policy that optimizes Q* reasonably well. Further, the requirement for the Q-function is only to get relative ordering between Q-values for different actions at states visited by the optimal policy right, which might not be the same as the Q-function landscape across state-action pairs. So, I am not sure if the expressivity of the optimal Q-function, in general, is the right metric to answer this question, but I also completely agree that this is a good starting point. \n\n5. I think when applying RL to the current benchmarks or problems we have, the main problem could be linked to optimization of a neural-net Q-function via bootstrapping (in the sense of approximate dynamic programming) as compared to the expressive power of optimal Q-functions. But I agree that the problem being looked at in the paper would also exist.\n\nReferences:\n[1] Plan Online, Learn Offline: Efficient Learning and Exploration via Model-Based Control, Lowrey et.al.\n\n"}, "tcdate": 1571687077694}], "openreview_url": "https://openreview.net/forum?id=Hye4WaVYwr", "arxiv_id": null, "paper_pdf": "papers/Hye4WaVYwr.pdf", "paper_pdf_sha256": "6bab9bebe0e2277af06dbeff6bf2227da182ca4e00fbf0847430b51a9bf963ff", "paper_pdf_bytes": 1320829, "paper_pdf_source": "openreview", "code_url": "https://github.com/roosephu/boots", "code_repository": "roosephu/boots", "code_commit": "2f4f500f54feb95cf36abd863f3de4510d6f4950", "code_archive": "repos/Hye4WaVYwr.zip", "code_archive_sha256": "b03d052aaa850a46682c2ae00b46e7763fbc4dbc85c7696b46868a7e326c1529", "code_archive_bytes": 61173, "code_file_count": 69, "code_extensions": {".py": 69}, "github_disk_usage_kb": 45, "github_languages": {"Python": 112974}, "github_archived": false, "github_pushed_at": "2019-10-14T08:35:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bootstrapping-the-expressivity-with-model-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rylKB3A9Fm", "year": 2019, "status": "rejected", "title": "Assessing Generalization in Deep Reinforcement Learning", "authors": ["Charles Packer*", "Katelyn Gao*", "Jernej Kos", "Philipp Krahenbuhl", "Vladlen Koltun", "Dawn Song"], "authorids": ["cpacker@berkeley.edu", "katelyn.gao@intel.com", "jernej@kos.mx", "philkr@cs.utexas.edu", "vladlen.koltun@intel.com", "dawnsong@berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Deep reinforcement learning (RL) has achieved breakthrough results on many tasks, but has been shown to be sensitive to system changes at test time. As a result, building deep RL agents that generalize has become an active research area. Our aim is to catalyze and streamline community-wide progress on this problem by providing the first benchmark and a common experimental protocol for investigating generalization in RL. Our benchmark contains a diverse set of environments and our evaluation methodology covers both in-distribution and out-of-distribution generalization. To provide a set of baselines for future research, we conduct a systematic evaluation of state-of-the-art algorithms, including those that specifically tackle the problem of generalization. The experimental results indicate that in-distribution generalization may be within the capacity of current algorithms, while out-of-distribution generalization is an exciting challenge for future work.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "ryloTLBC2m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1561/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new benchmark for studying generalization in deep RL along with a set of benchmark results. The benchmark consists of several standard RL tasks like Mountain Car along with several Mujoco continuous control tasks. Generalization is measured with respect to changes in environment parameters like force magnitude and pole length. Both interpolation and extrapolation are considered.\n\nThe problem considered in this paper is important and I agree with the authors that a good set of benchmarks for studying generalization is needed. However, a paper proposing a new benchmark should have a good argument for why the set of problems considered is interesting. Similarly, the types of generalization considered should be well motivated. This paper doesn’t do a good job of motivating these choices.\n\nFor example, why is Mountain Car a good task for studying generalization in deep RL? Mountain Car is a classic problem with a two-dimensional state space. This is hardly the kind of problem where deep RL shines or is even needed at all. Similarly, why should we care whether an agent trained on the Cart Pole task can generalize to a pole length between 2x and 10x shorter than the one it was trained on without being allowed to update its policy? Both the set of tasks and the distributions of parameters over which generalization is measured seem somewhat arbitrary.\n\nSimilarly, the restriction to methods that do not update its policy at test time also seems arbitrary since this is somewhat of a gray area. RL^2, which is one of the baselines in the paper, uses memory to adapt its policy to the current environment at test time. How different is this from an agent that updates its weights at test time? Why allow one but not the other?\n\nIn addition to these issues with the proposed benchmark, the baseline results don’t provide any new insights. The main conclusion is that extrapolation is more difficult than interpolation, which is in turn more difficult than training and testing on the same task. Beyond that, the results are very confusing. Two methods for improving generalization (EPOpt and RL^2) are evaluated and both of them seem to mostly decrease generalization performance. I find the poor performance of RL^2-A2C especially worrisome. Isn’t it essentially recurrent A2C where the reward and action are fed in as inputs? Why should the performance drop by 20-40%?\n\nOverall, I don’t see the proposed tasks becoming a widely used benchmark for evaluating generalization in deep RL. There are just too many seemingly arbitrary choices in the design of this benchmark and the lack of interesting findings in the baseline experiments highlights these issues.\n\nOther comments:\n- “Massively Parallel Methods for Deep Reinforcement Learning” by Nair et al. introduced the human starts evaluation condition for Atari games in order to measure generalization to potentially unseen states. This should probably be discussed in related work.\n- It would be good to include the exact architecture details since it’s not clear how rewards and actions are given to the RL^2 agents.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper presents a new benchmark for studying generalization in deep RL along with a set of benchmark results. The benchmark consists of several standard RL tasks like Mountain Car along with several Mujoco continuous control tasks. Generalization is measured with respect to changes in environment parameters like force magnitude and pole length. Both interpolation and extrapolation are considered.\n\nThe problem considered in this paper is important and I agree with the authors that a good set of benchmarks for studying generalization is needed. However, a paper proposing a new benchmark should have a good argument for why the set of problems considered is interesting. Similarly, the types of generalization considered should be well motivated. This paper doesn’t do a good job of motivating these choices.\n\nFor example, why is Mountain Car a good task for studying generalization in deep RL? Mountain Car is a classic problem with a two-dimensional state space. This is hardly the kind of problem where deep RL shines or is even needed at all. Similarly, why should we care whether an agent trained on the Cart Pole task can generalize to a pole length between 2x and 10x shorter than the one it was trained on without being allowed to update its policy? Both the set of tasks and the distributions of parameters over which generalization is measured seem somewhat arbitrary.\n\nSimilarly, the restriction to methods that do not update its policy at test time also seems arbitrary since this is somewhat of a gray area. RL^2, which is one of the baselines in the paper, uses memory to adapt its policy to the current environment at test time. How different is this from an agent that updates its weights at test time? Why allow one but not the other?\n\nIn addition to these issues with the proposed benchmark, the baseline results don’t provide any new insights. The main conclusion is that extrapolation is more difficult than interpolation, which is in turn more difficult than training and testing on the same task. Beyond that, the results are very confusing. Two methods for improving generalization (EPOpt and RL^2) are evaluated and both of them seem to mostly decrease generalization performance. I find the poor performance of RL^2-A2C especially worrisome. Isn’t it essentially recurrent A2C where the reward and action are fed in as inputs? Why should the performance drop by 20-40%?\n\nOverall, I don’t see the proposed tasks becoming a widely used benchmark for evaluating generalization in deep RL. There are just too many seemingly arbitrary choices in the design of this benchmark and the lack of interesting findings in the baseline experiments highlights these issues.\n\nOther comments:\n- “Massively Parallel Methods for Deep Reinforcement Learning” by Nair et al. introduced the human starts evaluation condition for Atari games in order to measure generalization to potentially unseen states. This should probably be discussed in related work.\n- It would be good to include the exact architecture details since it’s not clear how rewards and actions are given to the RL^2 agents.\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541457602550}, {"id": "B1lNZjECnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1561/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a benchmark for for reinforcement learning to study generalization in stationary and changing environments. A combination of several existing env. from OpenAi gym is taken and several ways to set this parameters is proposed. Paper provides a relatively thorough study of popular methodologies on this benchmark.\n\nOverall, I am not sure there is a pressing need for this benchmark and paper does not provide an argument why there is an urgent need for one.\n\nFor instance, paragraph 3 on page 1 details a number of previous studies. Why those benchmarks are in-adequate?\nOn page at the end of second paragraph a  number of benchmarks from transfer learning literature is mentioned. Why not just use those and disallow model updates?\nIn the same way, it is not clear why new metric is introduced? How does it correlate with standard reward metrics?\n\nOverall, as empirical study, I think this work is interesting but I think paper should justify why we need this new benchmark.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper proposes a benchmark for for reinforcement learning to study generalization in stationary and changing environments. A combination of several existing env. from OpenAi gym is taken and several ways to set this parameters is proposed. Paper provides a relatively thorough study of popular methodologies on this benchmark.\n\nOverall, I am not sure there is a pressing need for this benchmark and paper does not provide an argument why there is an urgent need for one.\n\nFor instance, paragraph 3 on page 1 details a number of previous studies. Why those benchmarks are in-adequate?\nOn page at the end of second paragraph a  number of benchmarks from transfer learning literature is mentioned. Why not just use those and disallow model updates?\nIn the same way, it is not clear why new metric is introduced? How does it correlate with standard reward metrics?\n\nOverall, as empirical study, I think this work is interesting but I think paper should justify why we need this new benchmark.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541454587696}, {"id": "B1gkuMGYnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1561/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Update: Lower the confidence and score after reading other comments. \n===\n\nIn this paper, the authors benchmark several RL algorithms on their abilities of generalization. The experiments show interpolation is somehow manageable but extrapolation is difficult to achieve. \n\nThe writing quality is rather good. The authors make it very clear on how their experiments run and how to interpret their results. The experiments are also solid. It's interesting to see that both EPOpt and RL^2, which claim to generalize better, generalize worse than the vanilla counterparts. Since the success rates are sometimes higher with more exploration, could it be possible that the hyperparameters of EPOpt and RL^2 are non-optimal? \n\nFor interpolation/extrapolation tasks, all 5 numbers (RR, EE, DR, DE, RE) are expected since the geometric mean is always 0 once any of the numbers is 0. \n\nWhat does ``\"KL divergence coefficient\" in RL^2-PPO mean? OpenAI's Baselines' implementation includes an entropy term as in A2C. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting topic and solid experiments", "review": "Update: Lower the confidence and score after reading other comments. \n===\n\nIn this paper, the authors benchmark several RL algorithms on their abilities of generalization. The experiments show interpolation is somehow manageable but extrapolation is difficult to achieve. \n\nThe writing quality is rather good. The authors make it very clear on how their experiments run and how to interpret their results. The experiments are also solid. It's interesting to see that both EPOpt and RL^2, which claim to generalize better, generalize worse than the vanilla counterparts. Since the success rates are sometimes higher with more exploration, could it be possible that the hyperparameters of EPOpt and RL^2 are non-optimal? \n\nFor interpolation/extrapolation tasks, all 5 numbers (RR, EE, DR, DE, RE) are expected since the geometric mean is always 0 once any of the numbers is 0. \n\nWhat does ``\"KL divergence coefficient\" in RL^2-PPO mean? OpenAI's Baselines' implementation includes an entropy term as in A2C. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541116519161}], "openreview_url": "https://openreview.net/forum?id=rylKB3A9Fm", "arxiv_id": "1810.12282", "paper_pdf": "papers/rylKB3A9Fm.pdf", "paper_pdf_sha256": "293ca7ce737060d337bb86a1a211c1d8af8d1c969e26cb0e7f609e3003b6a623", "paper_pdf_bytes": 1864848, "paper_pdf_source": "openreview", "code_url": "https://github.com/sunblaze-ucb/rl-generalization", "code_repository": "sunblaze-ucb/rl-generalization", "code_commit": "5e05aa108aeadd4cf21f45d56128e9da3bddc14c", "code_archive": "repos/rylKB3A9Fm.zip", "code_archive_sha256": "287436828bfec9fd97753474c9ece96f66b69c3312c30757f5514f7f26a61d91", "code_archive_bytes": 212964, "code_file_count": 49, "code_extensions": {".py": 49}, "github_disk_usage_kb": 172, "github_languages": {"Python": 169497}, "github_archived": false, "github_pushed_at": "2019-01-22T05:01:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/assessing-generalization-in-deep"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJvWjcgAZ", "year": 2018, "status": "rejected", "title": "Sample-Efficient Deep Reinforcement Learning via Episodic Backward Update", "authors": ["Su Young Lee", "Sungik Choi", "Sae-Young Chung"], "authorids": ["sy9424@kaist.ac.kr", "si_choi@kaist.ac.kr", "schung@kaist.ac.kr"], "authors_source": "OpenReview API", "abstract": "We propose Episodic Backward Update - a new algorithm to boost the performance of a deep reinforcement learning agent by fast reward propagation. In contrast to the conventional use of the replay memory with uniform random sampling, our agent samples a whole episode and successively propagates the value of a state into its previous states. Our computationally efficient recursive algorithm allows sparse and delayed rewards to propagate effectively throughout the sampled episode. We evaluate our algorithm on 2D MNIST Maze Environment and 49 games of the Atari 2600 Environment and show that our agent improves sample efficiency with a competitive computational cost.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJ3y_pYxM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper356/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new variant of DQN where the DQN targets are computed on a full episode by a « backward » update (i.e. from end to start of episode). The targets’ update rule is similar to a regular tabular Q-learning update with high learning rate beta: this allows faster propagation of rewards obtained at the end of the episode (while beta=0 corresponds to regular DQN with no such reward propagation). This mechanism is shown to improve on Q-learning in a toy 2D maze environment (with MNIST-based pixel states providing cell coordinates) with beta=1, and on DQN and its optimality tightening variant on Atari games with beta=0.5.\n\nThe intuition behind the algorithm (that one should try to speed up the propagation of rewards across multiple steps) is not new, in fact it has inspired other approaches like n-step Q-learning, eligibility traces or more recently Retrace(lambda) in deep RL. Actually the idea of replaying experiences in backward order can be traced back to the origins of experience replay («  Programming Robots Using Reinforcement Learning and Teaching », Lin, 1991), something that is not mentioned here. That being said, to the best of my knowledge the specific algorithm proposed in this submission (Alg. 2) is novel, even if Alg. 1 is not (Alg. 1 can be seen as a specific instance of Lin’s algorithm with a very high learning rate, and clearly only makes sense in toy deterministic environments).\n\nIn the absence of any theoretical analysis of the proposed approach, I would have expected an in-depth empirical validation. Unfortunately this is not the case here. In the toy environment (4.1) I am surprised by the really poor quality of the results (paths 5-10 times longer than the shortest path on average): have algorithms been run for a long enough time? Or maybe the average is a bad performance measure due to outliers? I would have also appreciated a comparison to Retrace(lambda), which is a more principled way to use multi-step rewards than n-step Q-learning (which is technically an on-policy method). Similar remarks can be made on the Atari experiments (4.2), where 10M frames is really low (the original DQN paper had results on 50M frames, and Rainbow reports 200M frames in only ~2x the training time reported here). The comparison also should have included prioritized experience replay, which has been shown to provide a significant boost in DQN, but may be tricky to combine with the proposed algorithm. Overall comparing only to vanilla DQN and its optimality tightening variant is too limited when there have been so many other meaningful improvements over DQN. This makes it really hard to tell whether the proposed algorithm would actually help when combined with a state-of-the-art method like Rainbow for instance.\n\nA few additional small remarks and questions:\n- « Second, there is no point in updating a one-step transition unless the future transitions have not been updated yet. »: should « unless » be replaced by « if »?\n- In 4.1 is there a maximum number of steps per episode and can you please confirm that training is done independently for each maze?\n- Typo in eq. 3: the - in the max should be a comma\n- There is a good amount of typos and grammar errors, though they do not harm the readability of the paper\n- Citations for « Deep Reinforcement Learning with Double Q-learning » and « Dueling Network Architectures for Deep Reinforcement Learning » could refer to their conference versions\n- « epsilon starts from 1 and is annealed to 0 at 200,000 steps in a quadratic manner »: please specify the exact formula\n- Fig. 7 is really confusing, there seem to be typos and it is not clear why the beta updates appear in these specific cells, please revise it if you want to keep it", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A potentially interesting approach, but with weak theoretical and empirical validation", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a new variant of DQN where the DQN targets are computed on a full episode by a « backward » update (i.e. from end to start of episode). The targets’ update rule is similar to a regular tabular Q-learning update with high learning rate beta: this allows faster propagation of rewards obtained at the end of the episode (while beta=0 corresponds to regular DQN with no such reward propagation). This mechanism is shown to improve on Q-learning in a toy 2D maze environment (with MNIST-based pixel states providing cell coordinates) with beta=1, and on DQN and its optimality tightening variant on Atari games with beta=0.5.\n\nThe intuition behind the algorithm (that one should try to speed up the propagation of rewards across multiple steps) is not new, in fact it has inspired other approaches like n-step Q-learning, eligibility traces or more recently Retrace(lambda) in deep RL. Actually the idea of replaying experiences in backward order can be traced back to the origins of experience replay («  Programming Robots Using Reinforcement Learning and Teaching », Lin, 1991), something that is not mentioned here. That being said, to the best of my knowledge the specific algorithm proposed in this submission (Alg. 2) is novel, even if Alg. 1 is not (Alg. 1 can be seen as a specific instance of Lin’s algorithm with a very high learning rate, and clearly only makes sense in toy deterministic environments).\n\nIn the absence of any theoretical analysis of the proposed approach, I would have expected an in-depth empirical validation. Unfortunately this is not the case here. In the toy environment (4.1) I am surprised by the really poor quality of the results (paths 5-10 times longer than the shortest path on average): have algorithms been run for a long enough time? Or maybe the average is a bad performance measure due to outliers? I would have also appreciated a comparison to Retrace(lambda), which is a more principled way to use multi-step rewards than n-step Q-learning (which is technically an on-policy method). Similar remarks can be made on the Atari experiments (4.2), where 10M frames is really low (the original DQN paper had results on 50M frames, and Rainbow reports 200M frames in only ~2x the training time reported here). The comparison also should have included prioritized experience replay, which has been shown to provide a significant boost in DQN, but may be tricky to combine with the proposed algorithm. Overall comparing only to vanilla DQN and its optimality tightening variant is too limited when there have been so many other meaningful improvements over DQN. This makes it really hard to tell whether the proposed algorithm would actually help when combined with a state-of-the-art method like Rainbow for instance.\n\nA few additional small remarks and questions:\n- « Second, there is no point in updating a one-step transition unless the future transitions have not been updated yet. »: should « unless » be replaced by « if »?\n- In 4.1 is there a maximum number of steps per episode and can you please confirm that training is done independently for each maze?\n- Typo in eq. 3: the - in the max should be a comma\n- There is a good amount of typos and grammar errors, though they do not harm the readability of the paper\n- Citations for « Deep Reinforcement Learning with Double Q-learning » and « Dueling Network Architectures for Deep Reinforcement Learning » could refer to their conference versions\n- « epsilon starts from 1 and is annealed to 0 at 200,000 steps in a quadratic manner »: please specify the exact formula\n- Fig. 7 is really confusing, there seem to be typos and it is not clear why the beta updates appear in these specific cells, please revise it if you want to keep it", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511802852471}, {"id": "H1gBrkcgM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper356/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a simple modification to the DQN algorithm they call Episodic Backward Update. The algorithm selects transitions in a backward order fashion from end of episode to be more effective in propagating learning of new rewards. This issue of fast propagation of updates is a common theme in RL (cf eligibility traces, prioritised sweeping, and more recently DQN with prioritised replay etc.). Here the proposed update applies the max Bellman operator recursively on a trajectory (unsure whether this is novel), with some decay to prevent accumulating errors with the nested max.\n\nThe paper is written in a clear way. The proposed approach seems reasonable, but I would have guessed that prioritized replay would also naturally sample transitions in roughly that order - given that TD-errors would at first be higher towards the end of an episode and progress backwards from there. I think this should have been one of the baselines to compare to for that reason.\n\nThe experimental results seem promising in the illustrative MNIST domain. Atari results seem decent, especially given that experiments are limited to 10M frames, though the advantage compared to the related approach of optimality tightening is not obvious. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An RL update for DQN-like agents based on recursive max backups.", "rating": "6: Marginally above acceptance threshold", "review": "The authors propose a simple modification to the DQN algorithm they call Episodic Backward Update. The algorithm selects transitions in a backward order fashion from end of episode to be more effective in propagating learning of new rewards. This issue of fast propagation of updates is a common theme in RL (cf eligibility traces, prioritised sweeping, and more recently DQN with prioritised replay etc.). Here the proposed update applies the max Bellman operator recursively on a trajectory (unsure whether this is novel), with some decay to prevent accumulating errors with the nested max.\n\nThe paper is written in a clear way. The proposed approach seems reasonable, but I would have guessed that prioritized replay would also naturally sample transitions in roughly that order - given that TD-errors would at first be higher towards the end of an episode and progress backwards from there. I think this should have been one of the baselines to compare to for that reason.\n\nThe experimental results seem promising in the illustrative MNIST domain. Atari results seem decent, especially given that experiments are limited to 10M frames, though the advantage compared to the related approach of optimality tightening is not obvious. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511810360504}, {"id": "HyxmggJbM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper356/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new way of sampling data for updates in deep-Q networks. The basic principle is to update Q values starting from the end of the episode in order to facility quick propagation of rewards back along the episode.\n\nThe paper is interesting, but it lacks the proper comparisons to previously published techniques.\n\nThe results presented by this paper shows improvement over the baseline. But the Atari results is still significantly worse than the current SOTA.\n\nIn the non-tabular case, the authors have actually moved away from Q learning and defined an objective that is both on and off-policy. Some (theoretical) analysis would be nice. It is hard to judge whether the objective defined in the non-tabular defines a contraction operator at all in the tabular case.\n\nThere has been a number of highly relevant papers. Prioritized replay, for example, could have a very similar effect to proposed approach in the tabular case.\n\nIn the non-tabular case, the Retrace algorithm, tree backup, Watkin's Q learning all bear significant resemblance to the proposed method. Although the proposed algorithm is different from all 3, the authors should still have compared to at least one of them as a baseline. The Retrace algorithm specifically has also been shown to help significantly in the Atari case, and it defines a convergent update rule.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is interesting, but it lacks the proper comparisons to previously published techniques.", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes a new way of sampling data for updates in deep-Q networks. The basic principle is to update Q values starting from the end of the episode in order to facility quick propagation of rewards back along the episode.\n\nThe paper is interesting, but it lacks the proper comparisons to previously published techniques.\n\nThe results presented by this paper shows improvement over the baseline. But the Atari results is still significantly worse than the current SOTA.\n\nIn the non-tabular case, the authors have actually moved away from Q learning and defined an objective that is both on and off-policy. Some (theoretical) analysis would be nice. It is hard to judge whether the objective defined in the non-tabular defines a contraction operator at all in the tabular case.\n\nThere has been a number of highly relevant papers. Prioritized replay, for example, could have a very similar effect to proposed approach in the tabular case.\n\nIn the non-tabular case, the Retrace algorithm, tree backup, Watkin's Q learning all bear significant resemblance to the proposed method. Although the proposed algorithm is different from all 3, the authors should still have compared to at least one of them as a baseline. The Retrace algorithm specifically has also been shown to help significantly in the Atari case, and it defines a convergent update rule.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1512140824202}], "openreview_url": "https://openreview.net/forum?id=BJvWjcgAZ", "arxiv_id": "1805.12375", "paper_pdf": "papers/BJvWjcgAZ.pdf", "paper_pdf_sha256": "6d180fcf3562e9ff8b0594b803974e7d9569bb3cc2b3c027d24cc1275a92999c", "paper_pdf_bytes": 1380708, "paper_pdf_source": "openreview", "code_url": "https://github.com/suyoung-lee/Episodic-Backward-Update", "code_repository": "suyoung-lee/Episodic-Backward-Update", "code_commit": "d4bbcfee39e8d4ee51422015115b55ab3cfa1829", "code_archive": "repos/BJvWjcgAZ.zip", "code_archive_sha256": "7a25ad561b6d5d615899f3ffabce9577863b3f3cdf5190672f265da6d244cf30", "code_archive_bytes": 286798, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 300, "github_languages": {"Python": 72839}, "github_archived": false, "github_pushed_at": "2019-09-24T08:06:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sample-efficient-deep-reinforcement-learning-2"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yyqbLqLhGl", "year": 2026, "status": "rejected", "title": "FLASH: Latent-Aware Semi-Autoregressive Speculative Decoding for Multimodal Tasks", "authors": ["Zihua Wang", "Ruibo Li", "Haozhe Du", "Joey Tianyi Zhou", "Yu Zhang", "Xu Yang"], "authorids": ["~Zihua_Wang1", "~Ruibo_Li2", "~Haozhe_Du2", "~Joey_Tianyi_Zhou1", "~Yu_Zhang21", "~Xu_Yang5"], "authors_source": "OpenReview API", "abstract": "Large language and multimodal models (LLMs and LMMs) exhibit strong inference capabilities but are often limited by slow decoding speeds. This challenge is especially acute in LMMs, where visual inputs typically comprise more tokens with lower information density than text --- an issue exacerbated by recent trends toward finer-grained visual tokenizations to boost performance. Speculative decoding has been effective in accelerating LLM inference by using a smaller draft model to generate candidate tokens, which are then selectively verified by the target model, improving speed without sacrificing output quality. While this strategy has been extended to LMMs, existing methods largely overlook the unique properties of visual inputs and depend solely on text-based draft models.\nIn this work, we propose \\textbf{FLASH} (Fast Latent-Aware Semi-Autoregressive Heuristics), a speculative decoding framework designed specifically for LMMs, which leverages two key properties of multimodal data to design the draft model. First, to address redundancy in visual tokens, we propose a lightweight latent-aware token compression mechanism. Second, recognizing that visual objects often co-occur within a scene, we employ a semi-autoregressive decoding strategy to generate multiple tokens per forward pass. \nExperiments show that FLASH consistently outperforms prior speculative decoding approaches in both unimodal and multimodal settings, achieving up to \\textbf{2.68$\\times$} speed-up on video captioning and \\textbf{2.55$\\times$} on visual instruction tuning tasks compared to the original LMM.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qivxIzaCAm", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11813/Reviewer_iSr7"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The manuscripts proposes a novel framework for accelerating the decoding process of large multi-modal models, which uses a draft model to produce multiple next candidate tokens before they are verified by the target model. Prior works for efficient decoding in LMM paradigm often choose text-only draft models, neglecting that visual inputs contain critical information for the subsequent decoding process. To address this, the manuscript presents FLASH, generating candidate tokens using compressed visual tokens and a semi-autoregressive decoding strategy. The visual tokens are compressed due to the high redundancy in visual tokens, and multiple tokens are generated at once using the semi-autoregressive decoding strategy, instead of generating tokens one by one, to further speed up the decoding process. Overall, FLASH achieves a strong performance compared to previous speculative decoding methods.", "review_text": "The manuscripts proposes a novel framework for accelerating the decoding process of large multi-modal models, which uses a draft model to produce multiple next candidate tokens before they are verified by the target model. Prior works for efficient decoding in LMM paradigm often choose text-only draft models, neglecting that visual inputs contain critical information for the subsequent decoding process. To address this, the manuscript presents FLASH, generating candidate tokens using compressed visual tokens and a semi-autoregressive decoding strategy. The visual tokens are compressed due to the high redundancy in visual tokens, and multiple tokens are generated at once using the semi-autoregressive decoding strategy, instead of generating tokens one by one, to further speed up the decoding process. Overall, FLASH achieves a strong performance compared to previous speculative decoding methods.", "strengths": "1. The proposed method, FLASH, not only takes into account the importance of visual information in speculative decoding, but also reduces the redundancy in the visual representations for improved efficiency. \n\n2. Multi-token prediction strategy is employed for further speeding up the draft generation process. Though both of the ideas are hardly novel, the combined usage in the speculative decoding paradigm is crucial for the field. \n\n3. Quantitative evaluation of FLASH reveals a high acceptance rate, hence an improved speed-up ratio.", "weaknesses": "1. Though FLASH seems effective when compared to existing approaches, the component-wise analysis is not sufficient. For example, I would be curious to know, compared to no speculative decoding strategy, how does FLASH perform under different compression strategies, different number of visual tokens after compression, different number of tokens produced per-forward-pass by the draft model, etc. \n\n2. Since the draft model in FLASH is capable of generating multiple tokens at one single forward pass, is the metric 'average acceptance tokens' still effective for comparing the effectiveness or the performance of different decoding strategies?\n\n3. It is not evaluated or explained how the decoding strategy affects final benchmark performance.", "questions": "1. Just for discussion, how does the size of the draft model affect the acceptance rate and the system-level speed-up rate of LMMs? (From my understanding, if the model produces draft tokens slowly but more accurately, the system-level speed-up rate would be higher, is it correct?)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The manuscripts proposes a novel framework for accelerating the decoding process of large multi-modal models, which uses a draft model to produce multiple next candidate tokens before they are verified by the target model. Prior works for efficient decoding in LMM paradigm often choose text-only draft models, neglecting that visual inputs contain critical information for the subsequent decoding process. To address this, the manuscript presents FLASH, generating candidate tokens using compressed visual tokens and a semi-autoregressive decoding strategy. The visual tokens are compressed due to the high redundancy in visual tokens, and multiple tokens are generated at once using the semi-autoregressive decoding strategy, instead of generating tokens one by one, to further speed up the decoding process. Overall, FLASH achieves a strong performance compared to previous speculative decoding methods.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The proposed method, FLASH, not only takes into account the importance of visual information in speculative decoding, but also reduces the redundancy in the visual representations for improved efficiency. \n\n2. Multi-token prediction strategy is employed for further speeding up the draft generation process. Though both of the ideas are hardly novel, the combined usage in the speculative decoding paradigm is crucial for the field. \n\n3. Quantitative evaluation of FLASH reveals a high acceptance rate, hence an improved speed-up ratio.", "weaknesses": "1. Though FLASH seems effective when compared to existing approaches, the component-wise analysis is not sufficient. For example, I would be curious to know, compared to no speculative decoding strategy, how does FLASH perform under different compression strategies, different number of visual tokens after compression, different number of tokens produced per-forward-pass by the draft model, etc. \n\n2. Since the draft model in FLASH is capable of generating multiple tokens at one single forward pass, is the metric 'average acceptance tokens' still effective for comparing the effectiveness or the performance of different decoding strategies?\n\n3. It is not evaluated or explained how the decoding strategy affects final benchmark performance.", "questions": "1. Just for discussion, how does the size of the draft model affect the acceptance rate and the system-level speed-up rate of LMMs? (From my understanding, if the model produces draft tokens slowly but more accurately, the system-level speed-up rate would be higher, is it correct?)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762100518617}, {"id": "dzQLWnkczs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11813/Reviewer_4Wgs"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "This paper introduces FLASH, a novel speculative decoding framework specifically designed for Large Multimodal Models (LMMs). The authors identify that existing speculative decoding methods, primarily developed for text-only LLMs, fail to leverage the unique properties of multimodal data. FLASH addresses this gap through two key innovations: (1) a visual token compression module that reduces redundant visual tokens from N to C tokens using learnable queries, and (2) a semi-autoregressive decoding head that generates K candidate tokens in parallel rather than sequentially. The method is evaluated on video captioning (Kinetics-400) and visual instruction tuning (LLaVA-instruct-150k) tasks using LLaVA-1.5 and QwenVL-2.5 as target models, achieving up to 2.68× speedup on video captioning and 2.55× on visual instruction tuning.", "review_text": "This paper introduces FLASH, a novel speculative decoding framework specifically designed for Large Multimodal Models (LMMs). The authors identify that existing speculative decoding methods, primarily developed for text-only LLMs, fail to leverage the unique properties of multimodal data. FLASH addresses this gap through two key innovations: (1) a visual token compression module that reduces redundant visual tokens from N to C tokens using learnable queries, and (2) a semi-autoregressive decoding head that generates K candidate tokens in parallel rather than sequentially. The method is evaluated on video captioning (Kinetics-400) and visual instruction tuning (LLaVA-instruct-150k) tasks using LLaVA-1.5 and QwenVL-2.5 as target models, achieving up to 2.68× speedup on video captioning and 2.55× on visual instruction tuning.", "strengths": "1. **Well-motivated approach**: The paper clearly identifies limitations of existing speculative decoding methods when applied to multimodal settings and provides compelling motivation for why multimodal-specific optimizations are needed.\n\n2. **Novel technical contributions**: The combination of visual token compression and semi-autoregressive decoding is novel and well-suited to the characteristics of multimodal data. The observation about visual object co-occurrence enabling parallel token generation is insightful.\n\n3. **Comprehensive experimental evaluation**: The experiments cover multiple models (LLaVA, QwenVL), multiple tasks (video captioning, visual instruction tuning), and multiple model sizes (3B to 32B parameters). The inclusion of both greedy (τ=0) and sampling (τ=1) decoding settings strengthens the evaluation.\n\n4. **Strong empirical results**: FLASH consistently outperforms baseline methods, achieving substantial speedups (up to 2.68×) while maintaining output quality. The method shows particularly strong improvements over text-only speculative decoding approaches.\n\n5. **Thorough ablation studies**: The paper includes useful ablations on different visual compression methods, the impact of K values, and performance in text-only scenarios, providing insights into each component's contribution.", "weaknesses": "1. **Limited theoretical analysis**: The paper lacks theoretical guarantees about the preservation of output distribution after semi-autoregressive decoding. While empirical results suggest quality is maintained, formal analysis would strengthen the claims.\n\n2. **Insufficient comparison with concurrent work**: The paper primarily compares against adaptations of LLM speculative decoding methods (Eagle-MM, Medusa-MM) rather than other multimodal-specific acceleration techniques beyond Dream and MASSV.\n\n3. **Unclear training details**: Critical implementation details are missing, such as:\n   - How the compression ratio C/N is determined\n   - Training time and computational requirements\n   - Convergence behavior of the semi-autoregressive head\n\n4. **Limited analysis of failure modes**: The paper doesn't discuss scenarios where FLASH might perform poorly or when the acceptance rate drops significantly. Understanding these limitations would be valuable for practitioners.\n\n5. **Scalability concerns**: While experiments show results up to 32B parameters, the scalability to even larger models (70B+) or longer contexts is unclear. The fixed compression ratio might become problematic for very long videos.", "questions": "1. **How sensitive is FLASH to the compression ratio C/N?** The paper mentions reducing 576 tokens to 64 for LLaVA, but how was this ratio determined? What happens with different compression ratios?\n\n2. **How does the semi-autoregressive head handle variable-length outputs?** When describing different visual regions, the phrase lengths might vary significantly. How does the fixed K affect performance in such cases?\n\n3. **What is the impact on hallucination?** Visual token compression might lose important details. Have the authors evaluated whether FLASH increases hallucination rates compared to the original model?\n\n4. **Can the method extend to other modalities?** The paper focuses on vision-language models. Could FLASH be adapted for audio-visual or other multimodal combinations?\n\n5. **Why not apply compression to text tokens as well?** The paper only compresses visual tokens. Is there a fundamental reason why text token compression wouldn't work, or was this not explored?\n\n6. **How does FLASH perform on fine-grained visual tasks?** Tasks requiring detailed visual understanding (e.g., OCR, diagram understanding) might be sensitive to token compression. Has this been evaluated?\n\n7. **What is the memory overhead?** While inference time is reduced, what is the additional memory requirement for storing the draft model components?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces FLASH, a novel speculative decoding framework specifically designed for Large Multimodal Models (LMMs). The authors identify that existing speculative decoding methods, primarily developed for text-only LLMs, fail to leverage the unique properties of multimodal data. FLASH addresses this gap through two key innovations: (1) a visual token compression module that reduces redundant visual tokens from N to C tokens using learnable queries, and (2) a semi-autoregressive decoding head that generates K candidate tokens in parallel rather than sequentially. The method is evaluated on video captioning (Kinetics-400) and visual instruction tuning (LLaVA-instruct-150k) tasks using LLaVA-1.5 and QwenVL-2.5 as target models, achieving up to 2.68× speedup on video captioning and 2.55× on visual instruction tuning.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. **Well-motivated approach**: The paper clearly identifies limitations of existing speculative decoding methods when applied to multimodal settings and provides compelling motivation for why multimodal-specific optimizations are needed.\n\n2. **Novel technical contributions**: The combination of visual token compression and semi-autoregressive decoding is novel and well-suited to the characteristics of multimodal data. The observation about visual object co-occurrence enabling parallel token generation is insightful.\n\n3. **Comprehensive experimental evaluation**: The experiments cover multiple models (LLaVA, QwenVL), multiple tasks (video captioning, visual instruction tuning), and multiple model sizes (3B to 32B parameters). The inclusion of both greedy (τ=0) and sampling (τ=1) decoding settings strengthens the evaluation.\n\n4. **Strong empirical results**: FLASH consistently outperforms baseline methods, achieving substantial speedups (up to 2.68×) while maintaining output quality. The method shows particularly strong improvements over text-only speculative decoding approaches.\n\n5. **Thorough ablation studies**: The paper includes useful ablations on different visual compression methods, the impact of K values, and performance in text-only scenarios, providing insights into each component's contribution.", "weaknesses": "1. **Limited theoretical analysis**: The paper lacks theoretical guarantees about the preservation of output distribution after semi-autoregressive decoding. While empirical results suggest quality is maintained, formal analysis would strengthen the claims.\n\n2. **Insufficient comparison with concurrent work**: The paper primarily compares against adaptations of LLM speculative decoding methods (Eagle-MM, Medusa-MM) rather than other multimodal-specific acceleration techniques beyond Dream and MASSV.\n\n3. **Unclear training details**: Critical implementation details are missing, such as:\n   - How the compression ratio C/N is determined\n   - Training time and computational requirements\n   - Convergence behavior of the semi-autoregressive head\n\n4. **Limited analysis of failure modes**: The paper doesn't discuss scenarios where FLASH might perform poorly or when the acceptance rate drops significantly. Understanding these limitations would be valuable for practitioners.\n\n5. **Scalability concerns**: While experiments show results up to 32B parameters, the scalability to even larger models (70B+) or longer contexts is unclear. The fixed compression ratio might become problematic for very long videos.", "questions": "1. **How sensitive is FLASH to the compression ratio C/N?** The paper mentions reducing 576 tokens to 64 for LLaVA, but how was this ratio determined? What happens with different compression ratios?\n\n2. **How does the semi-autoregressive head handle variable-length outputs?** When describing different visual regions, the phrase lengths might vary significantly. How does the fixed K affect performance in such cases?\n\n3. **What is the impact on hallucination?** Visual token compression might lose important details. Have the authors evaluated whether FLASH increases hallucination rates compared to the original model?\n\n4. **Can the method extend to other modalities?** The paper focuses on vision-language models. Could FLASH be adapted for audio-visual or other multimodal combinations?\n\n5. **Why not apply compression to text tokens as well?** The paper only compresses visual tokens. Is there a fundamental reason why text token compression wouldn't work, or was this not explored?\n\n6. **How does FLASH perform on fine-grained visual tasks?** Tasks requiring detailed visual understanding (e.g., OCR, diagram understanding) might be sensitive to token compression. Has this been evaluated?\n\n7. **What is the memory overhead?** While inference time is reduced, what is the additional memory requirement for storing the draft model components?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761928122278}, {"id": "DBTY59hOdl", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11813/Reviewer_Lox2"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This work introduces a method to accelerate the inference speed of large vision-language models. The core idea is to leverage a smaller, faster \"draft model\" to generate a sequence of visual and text tokens as a \"draft,\" which is then efficiently verified in a single parallel pass by the larger, more accurate \"target model.\" By accepting long sequences of drafted tokens when the draft is correct, this speculative decoding approach significantly reduces the number of sequential forward passes required from the expensive target model, thus lowering overall latency.", "review_text": "This work introduces a method to accelerate the inference speed of large vision-language models. The core idea is to leverage a smaller, faster \"draft model\" to generate a sequence of visual and text tokens as a \"draft,\" which is then efficiently verified in a single parallel pass by the larger, more accurate \"target model.\" By accepting long sequences of drafted tokens when the draft is correct, this speculative decoding approach significantly reduces the number of sequential forward passes required from the expensive target model, thus lowering overall latency.", "strengths": "The primary strength of this method lies in its practical utility for reducing the high computational cost associated with large multimodal models. By intelligently combining existing techniques like visual token compression and speculative decoding, it offers a direct path to faster inference without compromising the output quality of the original target model. The ability to verify multiple tokens in parallel is a clever way to exploit the architecture of modern transformers for a substantial speedup.", "weaknesses": "The primary weakness is the limited scope of the evaluation. While the reported speedup is promising, the benchmarks used do not cover a diverse range of real-world scenarios. The acceleration factor, presented as an average, may not be representative of performance on tasks with varying levels of predictability, such as complex visual reasoning versus simple image captioning. For instance, the actual performance gain in applications requiring high creativity or nuanced understanding could be significantly lower than reported. Furthermore, the novelty is somewhat constrained, as the approach is fundamentally an application and combination of pre-existing techniques rather than the introduction of a completely new paradigm.", "questions": "The methodology for creating the draft model raises a key question. Is the draft model a fine-tuned version of the same architecture as the target model? If so, have the authors considered replacing its backbone with a much smaller model? For example, use llava-7b as the backbone of the draft model for the target model llava-13b. Will this further improve the efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces a method to accelerate the inference speed of large vision-language models. The core idea is to leverage a smaller, faster \"draft model\" to generate a sequence of visual and text tokens as a \"draft,\" which is then efficiently verified in a single parallel pass by the larger, more accurate \"target model.\" By accepting long sequences of drafted tokens when the draft is correct, this speculative decoding approach significantly reduces the number of sequential forward passes required from the expensive target model, thus lowering overall latency.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The primary strength of this method lies in its practical utility for reducing the high computational cost associated with large multimodal models. By intelligently combining existing techniques like visual token compression and speculative decoding, it offers a direct path to faster inference without compromising the output quality of the original target model. The ability to verify multiple tokens in parallel is a clever way to exploit the architecture of modern transformers for a substantial speedup.", "weaknesses": "The primary weakness is the limited scope of the evaluation. While the reported speedup is promising, the benchmarks used do not cover a diverse range of real-world scenarios. The acceleration factor, presented as an average, may not be representative of performance on tasks with varying levels of predictability, such as complex visual reasoning versus simple image captioning. For instance, the actual performance gain in applications requiring high creativity or nuanced understanding could be significantly lower than reported. Furthermore, the novelty is somewhat constrained, as the approach is fundamentally an application and combination of pre-existing techniques rather than the introduction of a completely new paradigm.", "questions": "The methodology for creating the draft model raises a key question. Is the draft model a fine-tuned version of the same architecture as the target model? If so, have the authors considered replacing its backbone with a much smaller model? For example, use llava-7b as the backbone of the draft model for the target model llava-13b. Will this further improve the efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761903012502}, {"id": "cRlLjF288V", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11813/Reviewer_ht7i"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes FLASH, a speculative decoding method for LMMs. The authors introduce two key components in the draft model: visual token compression to reduce redundant visual tokens, and a semi-autoregressive head to exploit spatial co-occurrence patterns in visual inputs. FLASH outperforms existing speculative decoding methods for LMMs.", "review_text": "This paper proposes FLASH, a speculative decoding method for LMMs. The authors introduce two key components in the draft model: visual token compression to reduce redundant visual tokens, and a semi-autoregressive head to exploit spatial co-occurrence patterns in visual inputs. FLASH outperforms existing speculative decoding methods for LMMs.", "strengths": "1. The paper addresses an important and timely problem in speculative decoding for LMMs.\n2. The paper is well written and easy to follow.", "weaknesses": "1. The core components (visual token compression and semi-autoregressive generation) are not novel. The main contribution is their effective integration, which is a valuable engineering solution but offers limited algorithmic novelty.\n\n2. The two motivations (visual token redundancy and vision object co-occurrence) of this work are not convincing.\n   \n   a. Visual token redundancy: Context token count primarily affects prefill latency, not decode latency. Since speculative decoding targets decode-stage acceleration, why is visual token compression critical here?\n   \n   b. Vision object co-occurrence: The claim that visual input uniquely exhibits spatial co-occurrence is unconvincing. Spatial patterns like \"in front of\" or \"on the table\" are equally common in text-only contexts, thus the motivation for semi-autoregressive decoding needs stronger justification.\n\n3. While the paper includes the 'Text-only' baseline from Gagrani et al. (2024), it omits Gagrani's multimodal drafting baseline. Furthermore, the paper lacks a comparison with MASSV.\n\n4. Table 1 shows an inconsistency: for QwenVL-2.5 (temperature=1), Medusa-MM's speedup (2.35x) is superior to FLASH's (2.05x), yet the table seems to incorrectly mark FLASH as best. This raises reliability concerns. Furthermore, the speedup from FLASH often appears incremental.", "questions": "Please refer to the Weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes FLASH, a speculative decoding method for LMMs. The authors introduce two key components in the draft model: visual token compression to reduce redundant visual tokens, and a semi-autoregressive head to exploit spatial co-occurrence patterns in visual inputs. FLASH outperforms existing speculative decoding methods for LMMs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper addresses an important and timely problem in speculative decoding for LMMs.\n2. The paper is well written and easy to follow.", "weaknesses": "1. The core components (visual token compression and semi-autoregressive generation) are not novel. The main contribution is their effective integration, which is a valuable engineering solution but offers limited algorithmic novelty.\n\n2. The two motivations (visual token redundancy and vision object co-occurrence) of this work are not convincing.\n   \n   a. Visual token redundancy: Context token count primarily affects prefill latency, not decode latency. Since speculative decoding targets decode-stage acceleration, why is visual token compression critical here?\n   \n   b. Vision object co-occurrence: The claim that visual input uniquely exhibits spatial co-occurrence is unconvincing. Spatial patterns like \"in front of\" or \"on the table\" are equally common in text-only contexts, thus the motivation for semi-autoregressive decoding needs stronger justification.\n\n3. While the paper includes the 'Text-only' baseline from Gagrani et al. (2024), it omits Gagrani's multimodal drafting baseline. Furthermore, the paper lacks a comparison with MASSV.\n\n4. Table 1 shows an inconsistency: for QwenVL-2.5 (temperature=1), Medusa-MM's speedup (2.35x) is superior to FLASH's (2.05x), yet the table seems to incorrectly mark FLASH as best. This raises reliability concerns. Furthermore, the speedup from FLASH often appears incremental.", "questions": "Please refer to the Weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761385216715}], "openreview_url": "https://openreview.net/forum?id=yyqbLqLhGl", "arxiv_id": "2505.12728", "paper_pdf": "papers/yyqbLqLhGl.pdf", "paper_pdf_sha256": "aa690d43d850c4e7dd735b72fa9ebd4f668c59caf114fb6f517dd4aaf94bab0f", "paper_pdf_bytes": 550004, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZihuaEvan/FlashSD", "code_repository": "ZihuaEvan/FlashSD", "code_commit": "2a93afd06cda1450d07e489a34d4fc11c12dbe3c", "code_archive": "repos/yyqbLqLhGl.zip", "code_archive_sha256": "cb92f5a6f2a41ac06967c6cb01891847983c489689fe0393bc510d2521c271df", "code_archive_bytes": 178575, "code_file_count": 34, "code_extensions": {".py": 34}, "github_disk_usage_kb": 142, "github_languages": {"Python": 647839}, "github_archived": false, "github_pushed_at": "2026-08-01T09:22:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/flash-latent-aware-semi-autoregressive"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6BjEqGn1OO", "year": 2025, "status": "rejected", "title": "Modeling Real-Time Interactive Conversations as Timed Diarized Transcripts", "authors": ["Garrett Tanzer", "Gustaf Ahdritz", "Luke Melas-Kyriazi"], "authorids": ["~Garrett_Tanzer1", "~Gustaf_Ahdritz2", "~Luke_Melas-Kyriazi1"], "authors_source": "OpenReview API", "abstract": "Chatbots built upon language models have exploded in popularity, but they have largely been limited to synchronous, turn-by-turn dialogues. In this paper we present a simple yet general method to simulate real-time interactive conversations using pretrained text-only language models, by modeling timed diarized transcripts and decoding them with causal rejection sampling. We demonstrate the promise of this method with two case studies: instant messenger dialogues and spoken conversations, which require generation at about 30 tok/s and 20 tok/s respectively to maintain real-time interactivity. These capabilities can be added into language models using relatively little data and run on commodity hardware.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "117o7KNRkS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11184/Reviewer_4VTs"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper describes a method for applying a standard LLM to the task of generating dialogues incrementally, either turn-by-turn or word-by-word, by considering an output as a triple of (utterance time, speaker, message) and predicting on that basis. Experiments are given with turn-by-turn text-based instant messaging, and word-by-word text-based dialogue (starting with spoken dialogue, but processing with ASR to a text transcript and then experimenting on that). Experiments show that some plausible outputs can be generated, that generally bigger LLMs outperform smaller ones, and that the better models can receive good ratings from human evaluators.", "review_text": "The paper describes a method for applying a standard LLM to the task of generating dialogues incrementally, either turn-by-turn or word-by-word, by considering an output as a triple of (utterance time, speaker, message) and predicting on that basis. Experiments are given with turn-by-turn text-based instant messaging, and word-by-word text-based dialogue (starting with spoken dialogue, but processing with ASR to a text transcript and then experimenting on that). Experiments show that some plausible outputs can be generated, that generally bigger LLMs outperform smaller ones, and that the better models can receive good ratings from human evaluators.", "strengths": "The paper deals with interactive online generation of dialogue, in both message-by-message and word-by-word scenarios; this is a really important task from the point of view of building genuinely interactive conversational agents, and one that really hasn't seen much attention recently. The approach taken is quite intuitive to understand and is explained quite clearly. The evaluation shows that the approach can be effective, not only in generating coherent dialogue turns but including some important dialogue phenomena such as self-repair and turn-taking phrases (as shown by the examples in the appendix).", "weaknesses": "There is a fairly large amount of pre-LLM work on the details of incremental word-by-word dialogue modelling, including agents that can listen and generate on a word-by-word basis: the paper cites Skantze (2021)'s review of turn-taking treatment but would be stronger if it gave more comparison to work in that area.\n\nThe evaluation gives comparisons between LLMs on general metrics, but could be stronger if it looked at more details of coherence, turn-taking etc. It would be really nice if there was some discussion (quantitative or qualitative) of the linguistic phenomena that do/don't seem to be captured by this approach: the paper mentions coherence and consistency, but word-by-word and turn-by-turn modelling mean that other issues become important, e.g. realistic turn-taking patterns, interruption and overlap. The appendix transcripts show that some cases display good examples of some of these, some less so.\n\nThe evaluation uses data with a fairly low level of interactivity compared to many dialogue datasets. Instant messaging is an asynchronous, turn-by-turn medium; the spoken dialogues must be treated more incrementally, and are modelled word-by-word, but come from court transcripts in which the level of formality is high, and thus speaker changes, overlaps, interruptions etc are likely to be rare compared to more everyday, informal conversation. It would be helpful to know more about the average turn length, turn duration, number of speaker changes etc in this data compared to some of the more standard conversational datasets used in dialogue system development.\n\nThe model is one of generating a transcript as a whole, including all participants, whereas a practical conversational agent would have to react to other agents' contributions, online, and continue to adapt as their contributions come in (possibly interrupting, overlapping, leading to conversational directions that are not in the agent's interest, etc) - so some discussion of what would be involved in fitting this approach into those kind of constraints would be very helpful. \n\nRelatedly, the model's basic assumption that p(e|context) can be decomposed into p(t|context) x p(s|context,t) x [...] seems to mean that it's predicting an utterance event time, and then predicting the speaker of that utterance event. A more realistic agent might predict speaker (or know that it wants to speak) first, and then need to predict when to speak - would that fit with this approach?", "questions": "See weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper describes a method for applying a standard LLM to the task of generating dialogues incrementally, either turn-by-turn or word-by-word, by considering an output as a triple of (utterance time, speaker, message) and predicting on that basis. Experiments are given with turn-by-turn text-based instant messaging, and word-by-word text-based dialogue (starting with spoken dialogue, but processing with ASR to a text transcript and then experimenting on that). Experiments show that some plausible outputs can be generated, that generally bigger LLMs outperform smaller ones, and that the better models can receive good ratings from human evaluators.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper deals with interactive online generation of dialogue, in both message-by-message and word-by-word scenarios; this is a really important task from the point of view of building genuinely interactive conversational agents, and one that really hasn't seen much attention recently. The approach taken is quite intuitive to understand and is explained quite clearly. The evaluation shows that the approach can be effective, not only in generating coherent dialogue turns but including some important dialogue phenomena such as self-repair and turn-taking phrases (as shown by the examples in the appendix).", "weaknesses": "There is a fairly large amount of pre-LLM work on the details of incremental word-by-word dialogue modelling, including agents that can listen and generate on a word-by-word basis: the paper cites Skantze (2021)'s review of turn-taking treatment but would be stronger if it gave more comparison to work in that area.\n\nThe evaluation gives comparisons between LLMs on general metrics, but could be stronger if it looked at more details of coherence, turn-taking etc. It would be really nice if there was some discussion (quantitative or qualitative) of the linguistic phenomena that do/don't seem to be captured by this approach: the paper mentions coherence and consistency, but word-by-word and turn-by-turn modelling mean that other issues become important, e.g. realistic turn-taking patterns, interruption and overlap. The appendix transcripts show that some cases display good examples of some of these, some less so.\n\nThe evaluation uses data with a fairly low level of interactivity compared to many dialogue datasets. Instant messaging is an asynchronous, turn-by-turn medium; the spoken dialogues must be treated more incrementally, and are modelled word-by-word, but come from court transcripts in which the level of formality is high, and thus speaker changes, overlaps, interruptions etc are likely to be rare compared to more everyday, informal conversation. It would be helpful to know more about the average turn length, turn duration, number of speaker changes etc in this data compared to some of the more standard conversational datasets used in dialogue system development.\n\nThe model is one of generating a transcript as a whole, including all participants, whereas a practical conversational agent would have to react to other agents' contributions, online, and continue to adapt as their contributions come in (possibly interrupting, overlapping, leading to conversational directions that are not in the agent's interest, etc) - so some discussion of what would be involved in fitting this approach into those kind of constraints would be very helpful. \n\nRelatedly, the model's basic assumption that p(e|context) can be decomposed into p(t|context) x p(s|context,t) x [...] seems to mean that it's predicting an utterance event time, and then predicting the speaker of that utterance event. A more realistic agent might predict speaker (or know that it wants to speak) first, and then need to predict when to speak - would that fit with this approach?", "questions": "See weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730901599705}, {"id": "NFYEYZdp8O", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11184/Reviewer_6JmG"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "The paper presents a method for simulating real-time interactive conversations by modeling diarized, timed transcripts, combined with causal rejection sampling. This method enables pre-trained, text-only language models to handle asynchronous and synchronous dialogues, as demonstrated with case studies involving instant messaging and spoken conversation simulations. The approach aims to maintain interactivity with minimal hardware requirements, while retaining a natural flow of dialogue.", "review_text": "The paper presents a method for simulating real-time interactive conversations by modeling diarized, timed transcripts, combined with causal rejection sampling. This method enables pre-trained, text-only language models to handle asynchronous and synchronous dialogues, as demonstrated with case studies involving instant messaging and spoken conversation simulations. The approach aims to maintain interactivity with minimal hardware requirements, while retaining a natural flow of dialogue.", "strengths": "1. Introduces a feasible approach for real-time interaction modeling using timed diarized transcripts and causal rejection sampling, which can be integrated into standard pre-trained models.\n2. Demonstrates applicability with two distinct real-world cases: asynchronous instant messaging and real-time spoken conversations, adding diversity to potential model interactions.", "weaknesses": "1. There is little comparison with already existing work in this area (i.e. https://arxiv.org/abs/2405.19487) \n2. The model is trained on narrow datasets (instant messaging and court transcripts), raising doubts about its generalization to diverse, real-world conversational scenarios.\n3. Reliance on ASR and TTS systems may introduce errors and disrupt natural flow in spoken dialogues. An end-to-end audio model could reduce these issues by handling speech inputs and outputs directly.", "questions": "1. What considerations were made regarding the potential for user fatigue in interactions involving high rejection rates in real-time conversations?\n2. Have you considered building an end-to-end model? Integrating ASR and TTS into the LLM as well to achieve faster response times?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method for simulating real-time interactive conversations by modeling diarized, timed transcripts, combined with causal rejection sampling. This method enables pre-trained, text-only language models to handle asynchronous and synchronous dialogues, as demonstrated with case studies involving instant messaging and spoken conversation simulations. The approach aims to maintain interactivity with minimal hardware requirements, while retaining a natural flow of dialogue.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. Introduces a feasible approach for real-time interaction modeling using timed diarized transcripts and causal rejection sampling, which can be integrated into standard pre-trained models.\n2. Demonstrates applicability with two distinct real-world cases: asynchronous instant messaging and real-time spoken conversations, adding diversity to potential model interactions.", "weaknesses": "1. There is little comparison with already existing work in this area (i.e. https://arxiv.org/abs/2405.19487) \n2. The model is trained on narrow datasets (instant messaging and court transcripts), raising doubts about its generalization to diverse, real-world conversational scenarios.\n3. Reliance on ASR and TTS systems may introduce errors and disrupt natural flow in spoken dialogues. An end-to-end audio model could reduce these issues by handling speech inputs and outputs directly.", "questions": "1. What considerations were made regarding the potential for user fatigue in interactions involving high rejection rates in real-time conversations?\n2. Have you considered building an end-to-end model? Integrating ASR and TTS into the LLM as well to achieve faster response times?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730710558753}, {"id": "peKzY3sTYU", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11184/Reviewer_63e3"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces a method for simulating real-time interactive conversations using pretrained text-only language models, incorporating two key modifications. First, it employs timed diarized transcripts to represent each timestamped event, with each entry consisting of a timestamp, speaker ID, and message content. The model is tasked with predicting the probability of the next event based on event history, and sampling can proceed token by token, similar to standard causal language model text generation. During inference, a technique termed causal rejection sampling enables real-time interaction by discarding and resampling responses when interrupted by the user, thus adapting to dynamic input. To improve response speed during user interruptions, the authors also introduce two enhancements: (1) accounting for model generation latency and user reaction time and (2) applying a modified speculative decoding method to reuse partially generated responses before interruption.\n\nThe method is evaluated in two domains: instant messaging and spoken conversation. For instant messaging, the authors processed a 9-year message history between the first authors, resulting in 37 million characters in the diarized transcript format. For spoken conversations, they utilized a 1000-hour subset of U.S. Supreme Court oral arguments, totaling 33 million characters. This spoken data was converted into word-level transcripts with precise timing using an ASR engine. Experiments involved both open-source LLMs (Pythia, Gemma, LLaMA) ranging from 160M to 7B parameters, and state-of-the-art proprietary LLMs. Evaluation metrics included document-level perplexity, offline human evaluations ranking generated continuations, and online human evaluations involving direct interaction with the model. Key findings indicate that (1) the method meets real-time interactivity constraints on feasible hardware (e.g., a 40GB A100 GPU for a 7B LLaMA 2 model), and (2) larger, high-quality LLMs generally perform better in real-time interactive conversation modeling.", "review_text": "The paper introduces a method for simulating real-time interactive conversations using pretrained text-only language models, incorporating two key modifications. First, it employs timed diarized transcripts to represent each timestamped event, with each entry consisting of a timestamp, speaker ID, and message content. The model is tasked with predicting the probability of the next event based on event history, and sampling can proceed token by token, similar to standard causal language model text generation. During inference, a technique termed causal rejection sampling enables real-time interaction by discarding and resampling responses when interrupted by the user, thus adapting to dynamic input. To improve response speed during user interruptions, the authors also introduce two enhancements: (1) accounting for model generation latency and user reaction time and (2) applying a modified speculative decoding method to reuse partially generated responses before interruption.\n\nThe method is evaluated in two domains: instant messaging and spoken conversation. For instant messaging, the authors processed a 9-year message history between the first authors, resulting in 37 million characters in the diarized transcript format. For spoken conversations, they utilized a 1000-hour subset of U.S. Supreme Court oral arguments, totaling 33 million characters. This spoken data was converted into word-level transcripts with precise timing using an ASR engine. Experiments involved both open-source LLMs (Pythia, Gemma, LLaMA) ranging from 160M to 7B parameters, and state-of-the-art proprietary LLMs. Evaluation metrics included document-level perplexity, offline human evaluations ranking generated continuations, and online human evaluations involving direct interaction with the model. Key findings indicate that (1) the method meets real-time interactivity constraints on feasible hardware (e.g., a 40GB A100 GPU for a 7B LLaMA 2 model), and (2) larger, high-quality LLMs generally perform better in real-time interactive conversation modeling.", "strengths": "- Adapting text-only LLMs to model real-time interactive conversations is an important yet underexplored problem with the potential to unlock new applications without the need for costly retraining. The proposed method is straightforward but effective, demonstrated across a variety of LLMs from different model families and scales, and applicable to both fine-tuning and in-context learning.\n- The paper includes both offline and online human evaluations in addition to automatic metrics, which are particularly valuable for assessing the performance of LLMs.", "weaknesses": "- A critical aspect of real-time conversation is that the user can interrupt at any moment, requiring the model to discard its current response and handle the new input immediately. However, while the proposed method has the potential to address this, it is not thoroughly tested in the experiments. There is not even a qualitative example showing such behavior from the trained models.\n- Using instant messaging dialogues to evaluate real-time interactions may not be ideal, as people’s response habits vary depending on availability, typing speed, and other factors. Modeling response timing directly may be less meaningful in this setting, and the turn-based nature of messaging lacks the fluidity and interruption dynamics of real-time conversations.\n- The human evaluation lacks details, such as the number of conversations evaluated per method, the evaluation rubric, and consistency across ratings. It is unclear how the conversations are compared or if a consistent standard was applied, which is particularly important given the challenge of evaluating long texts.\n- The control tokens introduce significant overhead in the spoken conversation domain without an optimized tokenizer that treats numbers from 0 to 999 as single tokens. Since existing models may not support such tokens, this limits the method’s practicality when used with standard tokenizers.", "questions": "Q1: Could you clarify what is meant by the \"first 95%\"? Is this based on chronological order? If so, wouldn’t this temporal split risk introducing shifts in topics or conversational styles over time, potentially affecting the robustness and validity of the evaluation?\n> We use the first 95% of the messages as the train set, the next 2.5% as a validation set, and the last 2.5% as a test set.\n\nQ2: In Figure 4, timing is evaluated based on the delays between successive messages. Could you clarify why only delays are measured rather than both delays and any potential early responses? Additionally, how is the alignment between generated and ground-truth messages determined? If the generated messages differ significantly in content from the ground-truth, what is the meaning or value of comparing the timing between two sets of messages that may not correspond closely in their content?\n\nQ3: Why were different evaluation metrics used for human assessments in the instant messaging and spoken conversation settings? Could you explain the reasoning behind this choice and how it aligns with the specific goals of each domain?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a method for simulating real-time interactive conversations using pretrained text-only language models, incorporating two key modifications. First, it employs timed diarized transcripts to represent each timestamped event, with each entry consisting of a timestamp, speaker ID, and message content. The model is tasked with predicting the probability of the next event based on event history, and sampling can proceed token by token, similar to standard causal language model text generation. During inference, a technique termed causal rejection sampling enables real-time interaction by discarding and resampling responses when interrupted by the user, thus adapting to dynamic input. To improve response speed during user interruptions, the authors also introduce two enhancements: (1) accounting for model generation latency and user reaction time and (2) applying a modified speculative decoding method to reuse partially generated responses before interruption.\n\nThe method is evaluated in two domains: instant messaging and spoken conversation. For instant messaging, the authors processed a 9-year message history between the first authors, resulting in 37 million characters in the diarized transcript format. For spoken conversations, they utilized a 1000-hour subset of U.S. Supreme Court oral arguments, totaling 33 million characters. This spoken data was converted into word-level transcripts with precise timing using an ASR engine. Experiments involved both open-source LLMs (Pythia, Gemma, LLaMA) ranging from 160M to 7B parameters, and state-of-the-art proprietary LLMs. Evaluation metrics included document-level perplexity, offline human evaluations ranking generated continuations, and online human evaluations involving direct interaction with the model. Key findings indicate that (1) the method meets real-time interactivity constraints on feasible hardware (e.g., a 40GB A100 GPU for a 7B LLaMA 2 model), and (2) larger, high-quality LLMs generally perform better in real-time interactive conversation modeling.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Adapting text-only LLMs to model real-time interactive conversations is an important yet underexplored problem with the potential to unlock new applications without the need for costly retraining. The proposed method is straightforward but effective, demonstrated across a variety of LLMs from different model families and scales, and applicable to both fine-tuning and in-context learning.\n- The paper includes both offline and online human evaluations in addition to automatic metrics, which are particularly valuable for assessing the performance of LLMs.", "weaknesses": "- A critical aspect of real-time conversation is that the user can interrupt at any moment, requiring the model to discard its current response and handle the new input immediately. However, while the proposed method has the potential to address this, it is not thoroughly tested in the experiments. There is not even a qualitative example showing such behavior from the trained models.\n- Using instant messaging dialogues to evaluate real-time interactions may not be ideal, as people’s response habits vary depending on availability, typing speed, and other factors. Modeling response timing directly may be less meaningful in this setting, and the turn-based nature of messaging lacks the fluidity and interruption dynamics of real-time conversations.\n- The human evaluation lacks details, such as the number of conversations evaluated per method, the evaluation rubric, and consistency across ratings. It is unclear how the conversations are compared or if a consistent standard was applied, which is particularly important given the challenge of evaluating long texts.\n- The control tokens introduce significant overhead in the spoken conversation domain without an optimized tokenizer that treats numbers from 0 to 999 as single tokens. Since existing models may not support such tokens, this limits the method’s practicality when used with standard tokenizers.", "questions": "Q1: Could you clarify what is meant by the \"first 95%\"? Is this based on chronological order? If so, wouldn’t this temporal split risk introducing shifts in topics or conversational styles over time, potentially affecting the robustness and validity of the evaluation?\n> We use the first 95% of the messages as the train set, the next 2.5% as a validation set, and the last 2.5% as a test set.\n\nQ2: In Figure 4, timing is evaluated based on the delays between successive messages. Could you clarify why only delays are measured rather than both delays and any potential early responses? Additionally, how is the alignment between generated and ground-truth messages determined? If the generated messages differ significantly in content from the ground-truth, what is the meaning or value of comparing the timing between two sets of messages that may not correspond closely in their content?\n\nQ3: Why were different evaluation metrics used for human assessments in the instant messaging and spoken conversation settings? Could you explain the reasoning behind this choice and how it aligns with the specific goals of each domain?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730686883228}, {"id": "jw9uykqNYl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11184/Reviewer_JDak"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This paper presents a novel approach for achieving real-time response in dialogue systems by modeling timed, diarized transcripts and decoding through causal rejection sampling. The model is tested on both instant messenger dialogues and spoken conversations, with results demonstrating its ability to sustain real-time interactivity.", "review_text": "This paper presents a novel approach for achieving real-time response in dialogue systems by modeling timed, diarized transcripts and decoding through causal rejection sampling. The model is tested on both instant messenger dialogues and spoken conversations, with results demonstrating its ability to sustain real-time interactivity.", "strengths": "The proposed dialogue system supports multi-speaker, simultaneous conversations, marking a significant advancement over traditional turn-by-turn dialogue models. Additionally, the method’s design allows it to be easily integrated into various models, demonstrating its versatility and broad applicability.", "weaknesses": "1. Additional Evaluation Metrics to Consider\n- While the model demonstrates high generation bandwidth, how does your approach ensure the model engages at the right time rather than speaking continuously? In other words, how does your method balance high generation rates with appropriate turn-taking? \n- An ablation study comparing models without rejection sampling and those fine-tuned for next-turn speaker and response prediction could help validate the model design.\n2. Positioning and Contribution\n\nThis paper proposes a system design for real-time conversation. However, its novelty from a learning and evaluation perspective isn’t fully highlighted. I’d be open to discussing potential ways to position this paper’s contributions for ICLR.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel approach for achieving real-time response in dialogue systems by modeling timed, diarized transcripts and decoding through causal rejection sampling. The model is tested on both instant messenger dialogues and spoken conversations, with results demonstrating its ability to sustain real-time interactivity.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The proposed dialogue system supports multi-speaker, simultaneous conversations, marking a significant advancement over traditional turn-by-turn dialogue models. Additionally, the method’s design allows it to be easily integrated into various models, demonstrating its versatility and broad applicability.", "weaknesses": "1. Additional Evaluation Metrics to Consider\n- While the model demonstrates high generation bandwidth, how does your approach ensure the model engages at the right time rather than speaking continuously? In other words, how does your method balance high generation rates with appropriate turn-taking? \n- An ablation study comparing models without rejection sampling and those fine-tuned for next-turn speaker and response prediction could help validate the model design.\n2. Positioning and Contribution\n\nThis paper proposes a system design for real-time conversation. However, its novelty from a learning and evaluation perspective isn’t fully highlighted. I’d be open to discussing potential ways to position this paper’s contributions for ICLR.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730668528762}], "openreview_url": "https://openreview.net/forum?id=6BjEqGn1OO", "arxiv_id": "2405.13203", "paper_pdf": "papers/6BjEqGn1OO.pdf", "paper_pdf_sha256": "f3ed0676a59c4129cff3f198e3f125410510c10d8dd67eea02afaa8b04b710e6", "paper_pdf_bytes": 417819, "paper_pdf_source": "openreview", "code_url": "https://github.com/gahdritz/rtic", "code_repository": "gahdritz/rtic", "code_commit": "0ed62af2cda9c174dfb1309b66ba453ce724a0d8", "code_archive": "repos/6BjEqGn1OO.zip", "code_archive_sha256": "b6ac20f5f13f3d7f59b4c07bb32768692686df170624ff69479e68966f8b97c1", "code_archive_bytes": 28570, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 34, "github_languages": {"Python": 61707}, "github_archived": false, "github_pushed_at": "2024-06-03T07:10:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/modeling-real-time-interactive-conversations"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1gkePTsAWf", "year": 2024, "status": "rejected", "title": "Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation", "authors": ["Eric Zelikman", "Eliana Lorch", "Lester Mackey", "Adam Tauman Kalai"], "authorids": ["~Eric_Zelikman1", "eliana@lorien.ai", "~Lester_Mackey1", "~Adam_Tauman_Kalai1"], "authors_source": "OpenReview API", "abstract": "Several recent advances in AI systems (e.g., Tree-of-Thoughts and Program-Aided Language Models) solve problems by providing a \"scaffolding\" program that structures multiple calls to language models to generate better outputs. A scaffolding program is written in a programming language such as Python. In this work, we use a language-model-infused scaffolding program to improve itself. We start with a seed \"improver\" that improves an input program according to a given utility function by querying a language model several times and returning the best solution. We then run this seed improver to improve itself. Across a small set of downstream tasks, the resulting improved improver generates programs with significantly better performance than its seed improver. Afterward, we analyze the variety of self-improvement strategies proposed by the language model, including beam search, genetic algorithms, and simulated annealing. Since the language models themselves are not altered, this is not full recursive self-improvement. Nonetheless, it demonstrates that a modern language model, GPT-4 in our proof-of-concept experiments, is capable of writing code that can call itself to improve itself. We critically consider concerns around the development of self-improving technologies and evaluate the frequency with which the generated code bypasses a sandbox.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "SCgnwUu2mw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission284/Reviewer_443q"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposed self-taught optimizer (STOP), which is a nested-iterative method that uses a language model to improve itself, which is an \"improver\" that attempts to improve program solutions to certain tasks (e.g., learning parity w/ noise, 3-SAT). Experiments are conducted with GPT-4 and GPT-3.5 models, and the results show that GPT-4 is able to come up with methods such as beam search and genetic algorithms for the improver. On 5 different algorithmic tasks, STOP was shown effective in improving the original solutions.", "review_text": "This paper proposed self-taught optimizer (STOP), which is a nested-iterative method that uses a language model to improve itself, which is an \"improver\" that attempts to improve program solutions to certain tasks (e.g., learning parity w/ noise, 3-SAT). Experiments are conducted with GPT-4 and GPT-3.5 models, and the results show that GPT-4 is able to come up with methods such as beam search and genetic algorithms for the improver. On 5 different algorithmic tasks, STOP was shown effective in improving the original solutions.", "strengths": "S1: The proposed method is interesting and the idea of having a formulation where the improver can improve over itself given a utility function is very cool.  \nS2: A clear discussion of the limitations and concerns for STOP is presented, which is very helpful.  \nS3: A number of (i.e., five) tasks are considered in the experiments, and the proposed methods yield non-trivial improvements on all of them.", "weaknesses": "W1: My main concern about this work is missing some of the important results for us to understand how well the proposed method works. More concretely, \n* How much of improvements each iteration made. For example, it was mentioned that the improvements may not be monotonic, (which implies non-greedy, global optimization), while such improvement curves will be very interesting to look at, only one example of such is shown in Fig 4.\n* Quantitatively, the differences between the results with GPT-4 and GPT-3.5. It was only briefly mentioned in 5.3 about some of the pitfalls of GPT-3.5 that are not observed for GPT-4, but some concrete comparison would be helpful in understanding how applicable is STOP on other models.  \n\nW2: The writing of the paper could be improved. While I appreciate the authors explaining many of the design choices and giving alternative solutions that are not eventually part of the proposed framework (e.g., all those \"one can/may ...\"), it inadvertently breaks the flow of the paper, making it quite hard to follow sometimes.  \nW3: I am not entirely convinced about the \"novelty\" of the improver program generated by GPT-4. Though the authors attempt to compare the proposed methods to some recent research that happened after the knowledge cutoff time, the ideas behind those improver programs (e.g., beam search, genetic algorithm, etc) are not new at all.  \nW4: Missing related work. A couple of more related works for (self-)improving code generation with LLMs could have been mentioned, for example [1] and [2]  \n\n[1] https://arxiv.org/abs/2302.07867  \n[2] https://arxiv.org/abs/2304.05128", "questions": "Q1: Is the Maximizer formulation used in the final proposed pipeline? If so, is the notation consistent with the $M$ used in section 5.1?  \nQ2: Are there any baseline methods (even heuristics) that you can compare the proposed method to? For example, iterative prompting without using a program as an improver?  \nQ3: Can you comment on how difficult is it for the model to generalize to the improver solutions, given that it must have seen things like beam search, genetic algorithms, etc during training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed self-taught optimizer (STOP), which is a nested-iterative method that uses a language model to improve itself, which is an \"improver\" that attempts to improve program solutions to certain tasks (e.g., learning parity w/ noise, 3-SAT). Experiments are conducted with GPT-4 and GPT-3.5 models, and the results show that GPT-4 is able to come up with methods such as beam search and genetic algorithms for the improver. On 5 different algorithmic tasks, STOP was shown effective in improving the original solutions.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "S1: The proposed method is interesting and the idea of having a formulation where the improver can improve over itself given a utility function is very cool.  \nS2: A clear discussion of the limitations and concerns for STOP is presented, which is very helpful.  \nS3: A number of (i.e., five) tasks are considered in the experiments, and the proposed methods yield non-trivial improvements on all of them.", "weaknesses": "W1: My main concern about this work is missing some of the important results for us to understand how well the proposed method works. More concretely, \n* How much of improvements each iteration made. For example, it was mentioned that the improvements may not be monotonic, (which implies non-greedy, global optimization), while such improvement curves will be very interesting to look at, only one example of such is shown in Fig 4.\n* Quantitatively, the differences between the results with GPT-4 and GPT-3.5. It was only briefly mentioned in 5.3 about some of the pitfalls of GPT-3.5 that are not observed for GPT-4, but some concrete comparison would be helpful in understanding how applicable is STOP on other models.  \n\nW2: The writing of the paper could be improved. While I appreciate the authors explaining many of the design choices and giving alternative solutions that are not eventually part of the proposed framework (e.g., all those \"one can/may ...\"), it inadvertently breaks the flow of the paper, making it quite hard to follow sometimes.  \nW3: I am not entirely convinced about the \"novelty\" of the improver program generated by GPT-4. Though the authors attempt to compare the proposed methods to some recent research that happened after the knowledge cutoff time, the ideas behind those improver programs (e.g., beam search, genetic algorithm, etc) are not new at all.  \nW4: Missing related work. A couple of more related works for (self-)improving code generation with LLMs could have been mentioned, for example [1] and [2]  \n\n[1] https://arxiv.org/abs/2302.07867  \n[2] https://arxiv.org/abs/2304.05128", "questions": "Q1: Is the Maximizer formulation used in the final proposed pipeline? If so, is the notation consistent with the $M$ used in section 5.1?  \nQ2: Are there any baseline methods (even heuristics) that you can compare the proposed method to? For example, iterative prompting without using a program as an improver?  \nQ3: Can you comment on how difficult is it for the model to generalize to the improver solutions, given that it must have seen things like beam search, genetic algorithms, etc during training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699130411656}, {"id": "H3odXorbuJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission284/Reviewer_yFD6"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a meta-learning algorithm in the context of code generation with large language models (LLM). The algorithm has two components: an outer algorithm/ meta-algorithm (the self-improver algorithm) and \nan inner algorithm (the improver algorithm). The inner algorithm (or improver algorithm) optimizes code for downstream tasks by prompting an LLM. The proposed meta-algorithm maintains a single inner algorithm    \nat each iteration instead of a population. In each iteration, the meta-algorithm (the self-improver algorithm) \npasses the optimization goal, and constraints such as limits on runtime and model queries. The meta-algorithm measures the performance of each improver on the downstream task and returns the best solution (improver) based on a meta-utility function.", "review_text": "This paper proposes a meta-learning algorithm in the context of code generation with large language models (LLM). The algorithm has two components: an outer algorithm/ meta-algorithm (the self-improver algorithm) and \nan inner algorithm (the improver algorithm). The inner algorithm (or improver algorithm) optimizes code for downstream tasks by prompting an LLM. The proposed meta-algorithm maintains a single inner algorithm    \nat each iteration instead of a population. In each iteration, the meta-algorithm (the self-improver algorithm) \npasses the optimization goal, and constraints such as limits on runtime and model queries. The meta-algorithm measures the performance of each improver on the downstream task and returns the best solution (improver) based on a meta-utility function.", "strengths": "The key strengths of this research are, first and foremost, the demonstration of a proof-of-concept for using LLMs for self-improvement and meta-learning. \n\nSecond, the strength of the paper is that LLMs can optimize code which includes the model itself. This approach demonstrated by LLM is similar to evolutionary algorithms without any exposure in \nthe training data.\n\nThird, the impact of this research is profound, because it demonstrates that LLMs are able to self-improve themselves in contrast to Reinforcement learning from the Human Feedback approach. This research has the potential to have a significant influence in areas where the input data is continually changing, such as education and health. \n\nFourth, the self-improving behavior can be attributed to emerging capabilities in LLM [1], where \"...these tasks are not explicitly included in pre-training\" (See Section 5 of Reference [1]) and can be observed only on sufficiently large models.\nFuture researchers can explore more in this direction.\n\nReferences\n1. Wei J, Tay Y, Bommasani R, Raffel C, Zoph B, Borgeaud S, Yogatama D, Bosma M, Zhou D, Metzler D, Chi EH. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682. 2022 Jun 15.", "weaknesses": "The base LLM has a huge importance on STOP's performance (Section 5.3)", "questions": "Do the authors have thoughts how their proposed solutions can be used in meta-learning applications where the utility function may not be describle in natural language?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a meta-learning algorithm in the context of code generation with large language models (LLM). The algorithm has two components: an outer algorithm/ meta-algorithm (the self-improver algorithm) and \nan inner algorithm (the improver algorithm). The inner algorithm (or improver algorithm) optimizes code for downstream tasks by prompting an LLM. The proposed meta-algorithm maintains a single inner algorithm    \nat each iteration instead of a population. In each iteration, the meta-algorithm (the self-improver algorithm) \npasses the optimization goal, and constraints such as limits on runtime and model queries. The meta-algorithm measures the performance of each improver on the downstream task and returns the best solution (improver) based on a meta-utility function.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "The key strengths of this research are, first and foremost, the demonstration of a proof-of-concept for using LLMs for self-improvement and meta-learning. \n\nSecond, the strength of the paper is that LLMs can optimize code which includes the model itself. This approach demonstrated by LLM is similar to evolutionary algorithms without any exposure in \nthe training data.\n\nThird, the impact of this research is profound, because it demonstrates that LLMs are able to self-improve themselves in contrast to Reinforcement learning from the Human Feedback approach. This research has the potential to have a significant influence in areas where the input data is continually changing, such as education and health. \n\nFourth, the self-improving behavior can be attributed to emerging capabilities in LLM [1], where \"...these tasks are not explicitly included in pre-training\" (See Section 5 of Reference [1]) and can be observed only on sufficiently large models.\nFuture researchers can explore more in this direction.\n\nReferences\n1. Wei J, Tay Y, Bommasani R, Raffel C, Zoph B, Borgeaud S, Yogatama D, Bosma M, Zhou D, Metzler D, Chi EH. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682. 2022 Jun 15.", "weaknesses": "The base LLM has a huge importance on STOP's performance (Section 5.3)", "questions": "Do the authors have thoughts how their proposed solutions can be used in meta-learning applications where the utility function may not be describle in natural language?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698956642657}, {"id": "Thjn9tRh9F", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission284/Reviewer_X5AY"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces the Self-Taught Optimizer (STOP), a framework that uses a language model to recursively improve code scaffolding  generation in a meta-learning paradigm. This paper demonstrates, across a variety of algorithmic tasks, STOP generates better code scaffolding in more iterations that brings better performance of downstream code. Moreover, several meta heuristics are discovered during the meta-optimization process. This paper further Investigates the potential of misuse of language models about how it circumvents safety measures and reward hacking.", "review_text": "The paper introduces the Self-Taught Optimizer (STOP), a framework that uses a language model to recursively improve code scaffolding  generation in a meta-learning paradigm. This paper demonstrates, across a variety of algorithmic tasks, STOP generates better code scaffolding in more iterations that brings better performance of downstream code. Moreover, several meta heuristics are discovered during the meta-optimization process. This paper further Investigates the potential of misuse of language models about how it circumvents safety measures and reward hacking.", "strengths": "+ originality: this paper is the first to propose a meta-learning framework for optimizing code scaffolding, aiming at better performance of downstream tasks. \n\n+ significance: this paper takes a new and important perspective into revealing the power and potential misuse of large language model by querying the model to optimize the meta-heuristic of solving code tasks. though their current framework is not general to cover all tasks, their observation about reward hacking and sandbox circumvention is of great interest to the AI alignment community.\n\n+ clarity: their paper is easy to follow, with a high level schematic.", "weaknesses": "- missing comparison with two types of baselines, the one is human designed prompt structure such as Chain-of-Though, and Program of Thoughts. which one is better, the prompt structure found in their meta-learning paradigm or these human crafted ones? the other is heuristics for coding such as genetic algorithm. My question is given a downstream task, if STOP can find better meta-heuristics than the common ones?", "questions": "- is there any creativity in the meta-heuristics found by the language model? such as the combination of genetic algorithm and beam search? It should be of great interest to the community if the language model can find the new or more task-specific meta-heuristics, which brings better performance than human crafting.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the Self-Taught Optimizer (STOP), a framework that uses a language model to recursively improve code scaffolding  generation in a meta-learning paradigm. This paper demonstrates, across a variety of algorithmic tasks, STOP generates better code scaffolding in more iterations that brings better performance of downstream code. Moreover, several meta heuristics are discovered during the meta-optimization process. This paper further Investigates the potential of misuse of language models about how it circumvents safety measures and reward hacking.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "+ originality: this paper is the first to propose a meta-learning framework for optimizing code scaffolding, aiming at better performance of downstream tasks. \n\n+ significance: this paper takes a new and important perspective into revealing the power and potential misuse of large language model by querying the model to optimize the meta-heuristic of solving code tasks. though their current framework is not general to cover all tasks, their observation about reward hacking and sandbox circumvention is of great interest to the AI alignment community.\n\n+ clarity: their paper is easy to follow, with a high level schematic.", "weaknesses": "- missing comparison with two types of baselines, the one is human designed prompt structure such as Chain-of-Though, and Program of Thoughts. which one is better, the prompt structure found in their meta-learning paradigm or these human crafted ones? the other is heuristics for coding such as genetic algorithm. My question is given a downstream task, if STOP can find better meta-heuristics than the common ones?", "questions": "- is there any creativity in the meta-heuristics found by the language model? such as the combination of genetic algorithm and beam search? It should be of great interest to the community if the language model can find the new or more task-specific meta-heuristics, which brings better performance than human crafting.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698921909965}, {"id": "YWLma5aNLI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission284/Reviewer_VZPP"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a method that iteratively (meta-)optimises a program. At the beginning this program implements a simple optimisation heuristic through access to an LLM via an API and prompting the LLM for changes to the program that yield higher utility. The utility measures the performance of the program on some downstream optimisation task like 3SAT or Maxcut. The program is replaced with the output after running the program with its own code as input. The proposed method is tested with GPT-4 and GPT-3.5-turbo. Some of the programs that are found through this procedure mimic well known optimisation algorithms like genetic algorithms or simulated annealing.", "review_text": "The paper proposes a method that iteratively (meta-)optimises a program. At the beginning this program implements a simple optimisation heuristic through access to an LLM via an API and prompting the LLM for changes to the program that yield higher utility. The utility measures the performance of the program on some downstream optimisation task like 3SAT or Maxcut. The program is replaced with the output after running the program with its own code as input. The proposed method is tested with GPT-4 and GPT-3.5-turbo. Some of the programs that are found through this procedure mimic well known optimisation algorithms like genetic algorithms or simulated annealing.", "strengths": "Prompting the LLM for better methods to prompt an LLM is a promising direction given the success of handcrafted prompt strategies like chain-of-thought.", "weaknesses": "It is unclear what the connection between the input program (initial_solution) and the output program (best_solution) is. Is the seed program more helpful than compared to prompting the LLM for a program that maximizes some utility (possibly with example programs that demonstrate the task, similar to the few-shot setting for LLMs)? I understand that the goal here is to design a method that can recursively improve itself, but the significance of the results seem questionable to me if it is possible to attain the same results with a conceptually simple prompt.\n\nThe experimental results are limited to GPT-4 and GPT-3.5-turbo, which are both closed source and AFAIK can change over time. This means that the results will likely not be reproducible. I'd suggest running some additional experiments with open source models that offer the possibility of reproducing the results.\n\nThe experiments are restricted to simple problem settings (albeit hard optimization problems) like LPN, 3SAT, or Maxcut. This is quite a restricted scope when comparing to claims such as \"... a method in which code that applies a language model to improve arbitrary solutions is applied recursively to improve itself.\" from the introduction. I would suggest to tone down the writing, especially in the introduction, as it currently to suggest a level of generality that is not support by experiments.\n\nConsidering that the STOP method manages to find well-known optimization algorithms like genetic algorithms or simulated annealing, I get the impression that all that is really happening here is that the LLM is prompted to return some known optimisation heuristic and adapt it to use LLMs. The recursive nature of STOP does not appear to be important here. Can the authors provide suitable ablations or baselines to show that recurrence is indeed important? This is also related to my first point.", "questions": "Why is it important to choose a less-well-known task as the meta optimization objective? Is the expectation here that GPT-4 has seen less training data on this topic?\n\nHow were the qualitative examples chosen? Always the final solution? Did the authors go through all generated programs manually and pick the ones that recognisably correspond to a known optimisation algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method that iteratively (meta-)optimises a program. At the beginning this program implements a simple optimisation heuristic through access to an LLM via an API and prompting the LLM for changes to the program that yield higher utility. The utility measures the performance of the program on some downstream optimisation task like 3SAT or Maxcut. The program is replaced with the output after running the program with its own code as input. The proposed method is tested with GPT-4 and GPT-3.5-turbo. Some of the programs that are found through this procedure mimic well known optimisation algorithms like genetic algorithms or simulated annealing.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "Prompting the LLM for better methods to prompt an LLM is a promising direction given the success of handcrafted prompt strategies like chain-of-thought.", "weaknesses": "It is unclear what the connection between the input program (initial_solution) and the output program (best_solution) is. Is the seed program more helpful than compared to prompting the LLM for a program that maximizes some utility (possibly with example programs that demonstrate the task, similar to the few-shot setting for LLMs)? I understand that the goal here is to design a method that can recursively improve itself, but the significance of the results seem questionable to me if it is possible to attain the same results with a conceptually simple prompt.\n\nThe experimental results are limited to GPT-4 and GPT-3.5-turbo, which are both closed source and AFAIK can change over time. This means that the results will likely not be reproducible. I'd suggest running some additional experiments with open source models that offer the possibility of reproducing the results.\n\nThe experiments are restricted to simple problem settings (albeit hard optimization problems) like LPN, 3SAT, or Maxcut. This is quite a restricted scope when comparing to claims such as \"... a method in which code that applies a language model to improve arbitrary solutions is applied recursively to improve itself.\" from the introduction. I would suggest to tone down the writing, especially in the introduction, as it currently to suggest a level of generality that is not support by experiments.\n\nConsidering that the STOP method manages to find well-known optimization algorithms like genetic algorithms or simulated annealing, I get the impression that all that is really happening here is that the LLM is prompted to return some known optimisation heuristic and adapt it to use LLMs. The recursive nature of STOP does not appear to be important here. Can the authors provide suitable ablations or baselines to show that recurrence is indeed important? This is also related to my first point.", "questions": "Why is it important to choose a less-well-known task as the meta optimization objective? Is the expectation here that GPT-4 has seen less training data on this topic?\n\nHow were the qualitative examples chosen? Always the final solution? Did the authors go through all generated programs manually and pick the ones that recognisably correspond to a known optimisation algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698844502404}, {"id": "duqgzxDyd9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission284/Reviewer_UnQ9"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes using GPT models for self-improving an improver code that improves the solution code for five different problems, given a utility function for each problem. The aggregated utilities measure the performance of the solutions of each problem and serve as the signal for the improver to improve itself. The improver code improves itself a few times and is used to improve the solution codes. The results demonstrate that GPT-4 proposes and implements self-improvement strategies such as beam/tree search, genetic algorithm, and simulated annealing. Results using GPT-4 demonstrate a diminishing increase in aggregated (meta) utility within four iterations, whereas results using GPT-3.5 demonstrate decreasing aggregated utility.", "review_text": "This paper proposes using GPT models for self-improving an improver code that improves the solution code for five different problems, given a utility function for each problem. The aggregated utilities measure the performance of the solutions of each problem and serve as the signal for the improver to improve itself. The improver code improves itself a few times and is used to improve the solution codes. The results demonstrate that GPT-4 proposes and implements self-improvement strategies such as beam/tree search, genetic algorithm, and simulated annealing. Results using GPT-4 demonstrate a diminishing increase in aggregated (meta) utility within four iterations, whereas results using GPT-3.5 demonstrate decreasing aggregated utility.", "strengths": "The paper is well-written, and the problem formulation and method are precise.\nAt a high level, the formulation is elegant, functional, and novel in the context of GPT's.\nThe paper clearly distinguishes between improving a solution and improving an improver (scaffolding) that improves a solution.\nThe work addresses safety, adding sandbox safety measures.", "weaknesses": "The improver's signal for improvement is a real value, which is the aggregation of utilities of solutions to different problems.\nMoreover, the utilities of the solution of each problem may be on different scales, and they are linearly combined.\n\n1. Different tasks have utilities with different values and scales.\n2. Aggregating utilities of different tasks to a single number is a minimal input to the improver and may be an insufficient signal for improvement (just one number in each iteration and previous improver code).\n3. Tasks may not be representative, and their selection is unclear.\n4. Each utilities instantiates data for the problem, for example maxcut instantiates 3 random graphs with 300 nodes each. \nIt's unclear how the data and instances are selected.\n3 and 4 together: how many tasks, data instances, and their types are required?\n5. Figure 4a is missing iterations beyond T=4.\n6. The approach assumes many solution variants are generated for each task, and it's unclear how much of the \"improvement\" is due to increasing the number of variants. \n7. It's unclear how much of the \"improvement\" is due to GPT-4 being a generic language model that may improve each selected task rather than generic recursive self-improvement.\n8. The paper is well written, and the presented formulation is elegant at a high-level, however, it is missing key details of self-improvement.", "questions": "1. How sensitive are the results to changes in the prompts?\n2. How does going wider (range of tasks and data) effect going deeper (iterations of recursive self improvement)?\n3. Will the link to the repo become available?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes using GPT models for self-improving an improver code that improves the solution code for five different problems, given a utility function for each problem. The aggregated utilities measure the performance of the solutions of each problem and serve as the signal for the improver to improve itself. The improver code improves itself a few times and is used to improve the solution codes. The results demonstrate that GPT-4 proposes and implements self-improvement strategies such as beam/tree search, genetic algorithm, and simulated annealing. Results using GPT-4 demonstrate a diminishing increase in aggregated (meta) utility within four iterations, whereas results using GPT-3.5 demonstrate decreasing aggregated utility.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "strengths": "The paper is well-written, and the problem formulation and method are precise.\nAt a high level, the formulation is elegant, functional, and novel in the context of GPT's.\nThe paper clearly distinguishes between improving a solution and improving an improver (scaffolding) that improves a solution.\nThe work addresses safety, adding sandbox safety measures.", "weaknesses": "The improver's signal for improvement is a real value, which is the aggregation of utilities of solutions to different problems.\nMoreover, the utilities of the solution of each problem may be on different scales, and they are linearly combined.\n\n1. Different tasks have utilities with different values and scales.\n2. Aggregating utilities of different tasks to a single number is a minimal input to the improver and may be an insufficient signal for improvement (just one number in each iteration and previous improver code).\n3. Tasks may not be representative, and their selection is unclear.\n4. Each utilities instantiates data for the problem, for example maxcut instantiates 3 random graphs with 300 nodes each. \nIt's unclear how the data and instances are selected.\n3 and 4 together: how many tasks, data instances, and their types are required?\n5. Figure 4a is missing iterations beyond T=4.\n6. The approach assumes many solution variants are generated for each task, and it's unclear how much of the \"improvement\" is due to increasing the number of variants. \n7. It's unclear how much of the \"improvement\" is due to GPT-4 being a generic language model that may improve each selected task rather than generic recursive self-improvement.\n8. The paper is well written, and the presented formulation is elegant at a high-level, however, it is missing key details of self-improvement.", "questions": "1. How sensitive are the results to changes in the prompts?\n2. How does going wider (range of tasks and data) effect going deeper (iterations of recursive self improvement)?\n3. Will the link to the repo become available?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697645237416}], "openreview_url": "https://openreview.net/forum?id=1gkePTsAWf", "arxiv_id": "2310.02304", "paper_pdf": "papers/1gkePTsAWf.pdf", "paper_pdf_sha256": "77a23375914992ce6e6238087e278753f7ae79949cc2fc46b28e6d5b2aaaa27d", "paper_pdf_bytes": 600577, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/stop", "code_repository": "microsoft/stop", "code_commit": "0d6780c54306b2486dd36e9c4ae9b49aceb27ea4", "code_archive": "repos/1gkePTsAWf.zip", "code_archive_sha256": "5452c2032b207e9076866faa26ce4333bc38e59b43a5f11fce571d110c7222c4", "code_archive_bytes": 42182, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 31, "github_languages": {"Python": 89274}, "github_archived": false, "github_pushed_at": "2024-01-01T20:28:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/self-taught-optimizer-stop-recursively-self"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Z4lOwCEJQ8Z", "year": 2023, "status": "rejected", "title": "Training Normalizing Flows from Dependent Data", "authors": ["Matthias Kirchler", "Christoph Lippert", "Marius Kloft"], "authorids": ["~Matthias_Kirchler1", "~Christoph_Lippert1", "~Marius_Kloft1"], "authors_source": "OpenReview API", "abstract": "Normalizing flows are powerful non-parametric statistical models that function as a hybrid between density estimators and generative models. Current learning algorithms for normalizing flows assume that data points are sampled independently, an assumption that is frequently violated in practice, which may lead to erroneous density estimation and data generation. We propose a likelihood objective of normalizing flows incorporating dependencies between the data points, for which we derive a flexible and efficient learning algorithm suitable for different dependency structures. We show that respecting dependencies between observations can improve empirical results on both synthetic and real-world data.\n\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "n1Ip_vUYE2U", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4574/Reviewer_gW39"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a modification to the standard normalizing flow training loss to account for certain dependencies in data. Specifically, the authors consider the special case where datapoints are dependent in latent space, but are transformed pointwise to the observed space. In addition, the authors specialize to the case where the correlation structure is either known up to a convex combination with the identity, or has a block-diagonal structure with known block assignment and constant positive correlation within blocks. The authors then propose training strategies for learning the parameters of these dependencies. The authors evaluate experimentally on simulated and real datasets, using NLL on held-out datasets as the core metric.", "review_text": "The proposed method seems fine in terms of specifying an objective that reflects certain forms of independence. However, the actual tasks being solved require far more explicit discussion, since there are many different tasks that one might try to solve in a dependent data context, and it is not clear which of these the method is designed to address. Because of this concern, the paper feels incomplete.", "strengths": "Strengths:\n + Dependence between datapoints can be an important issue, especially in cases where one cares about being able to make joint or conditional predictions where the task depends on multiple datatpoints simultaneously.\n + The derivations are clearly explained. Although they are straightforward, they are not obfuscated to appear more complicated than they are.\n + Several practical issues in training are addressed.\n + Experiments explore degrees of freedom that were raised in the development of the method.\n\nWeaknesses:\n  - The primary weakness to me is that the method seems undermotivated. Although I know that there exist tasks where dependency between datapoints matter, it is not clear how accounting for dependencies between data translates to better solutions for important tasks. The authors raise several applications, including genomics, finance, and neuroimaging, but stop short of indicating what problems could be solved in these areas if only we modeled dependencies between examples better. NLL is not itself self-motivating; it measures the model's consistency with the data, but does not tell us what this extra fit buys in terms of performance on a real task. What is the unsolved problem in genomics, finance, or neuroimaging that can be solved with a good model of between-example dependence?\n - Along the same lines, when we consider dependent data, the actual prediction task needs to be specified with more care. What exactly is being measured by performance on the test data? Do we care about our having low expected prediction risk on the next datapoint? Low joint prediction risk on a set of $m$ new datapoints? Some kind of statistical inference? In the first case, dependency structure would not matter. In the second case, it might matter, but one would need to motivate why we care about $m$ datapoints simultaneously, and the composition of this test set would need to be motivated carefully (e.g., would this test set represent measurements from a completely new individual? from an individual we have observed before?). Also, we might need the test set to be composed of $n$ replicates of $m$ datapoints. In the third case, the form of statistical inference would need to be made clear (e.g., Are we doing anomaly prediction and constructing p-values? Are we using the normalizing flow to do covariance estimation for a downstream regression / GWAS task?).\n - For example, it is not clear what the goal of the task illustrated in Figure 1 is. If the goal were density estimation, it seems like the standard normalizing flow is doing the right thing. Meanwhile, it is not clear why we want the dependency-adjusted normalizing flow to match the distribution of data sampled independently. This comes back to the issue that when one has dependent data, the goals are not self-evident in the same way that they are when predicting with independent data. Goals and notions of bias need to be made explicit.\n - Given the note in the discussion that \"equal block sizes\" would be a case where the dependent data objective would produce no changes, did the authors consider the baseline of weighting datapoints by inverse block size? It seems that for some tasks this might address many issues and not require major modifications to training.\n - Many tricks for working with covariance matrices have been discussed before in literatures that do efficient ridge regression or work with Gaussian processes. I am not sure that the linear algebra tricks need to be rehashed in the main body of the paper, and this could make room for better discussion of tasks where dependence matters.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a modification to the standard normalizing flow training loss to account for certain dependencies in data. Specifically, the authors consider the special case where datapoints are dependent in latent space, but are transformed pointwise to the observed space. In addition, the authors specialize to the case where the correlation structure is either known up to a convex combination with the identity, or has a block-diagonal structure with known block assignment and constant positive correlation within blocks. The authors then propose training strategies for learning the parameters of these dependencies. The authors evaluate experimentally on simulated and real datasets, using NLL on held-out datasets as the core metric.", "strength_and_weaknesses": "Strengths:\n + Dependence between datapoints can be an important issue, especially in cases where one cares about being able to make joint or conditional predictions where the task depends on multiple datatpoints simultaneously.\n + The derivations are clearly explained. Although they are straightforward, they are not obfuscated to appear more complicated than they are.\n + Several practical issues in training are addressed.\n + Experiments explore degrees of freedom that were raised in the development of the method.\n\nWeaknesses:\n  - The primary weakness to me is that the method seems undermotivated. Although I know that there exist tasks where dependency between datapoints matter, it is not clear how accounting for dependencies between data translates to better solutions for important tasks. The authors raise several applications, including genomics, finance, and neuroimaging, but stop short of indicating what problems could be solved in these areas if only we modeled dependencies between examples better. NLL is not itself self-motivating; it measures the model's consistency with the data, but does not tell us what this extra fit buys in terms of performance on a real task. What is the unsolved problem in genomics, finance, or neuroimaging that can be solved with a good model of between-example dependence?\n - Along the same lines, when we consider dependent data, the actual prediction task needs to be specified with more care. What exactly is being measured by performance on the test data? Do we care about our having low expected prediction risk on the next datapoint? Low joint prediction risk on a set of $m$ new datapoints? Some kind of statistical inference? In the first case, dependency structure would not matter. In the second case, it might matter, but one would need to motivate why we care about $m$ datapoints simultaneously, and the composition of this test set would need to be motivated carefully (e.g., would this test set represent measurements from a completely new individual? from an individual we have observed before?). Also, we might need the test set to be composed of $n$ replicates of $m$ datapoints. In the third case, the form of statistical inference would need to be made clear (e.g., Are we doing anomaly prediction and constructing p-values? Are we using the normalizing flow to do covariance estimation for a downstream regression / GWAS task?).\n - For example, it is not clear what the goal of the task illustrated in Figure 1 is. If the goal were density estimation, it seems like the standard normalizing flow is doing the right thing. Meanwhile, it is not clear why we want the dependency-adjusted normalizing flow to match the distribution of data sampled independently. This comes back to the issue that when one has dependent data, the goals are not self-evident in the same way that they are when predicting with independent data. Goals and notions of bias need to be made explicit.\n - Given the note in the discussion that \"equal block sizes\" would be a case where the dependent data objective would produce no changes, did the authors consider the baseline of weighting datapoints by inverse block size? It seems that for some tasks this might address many issues and not require major modifications to training.\n - Many tricks for working with covariance matrices have been discussed before in literatures that do efficient ridge regression or work with Gaussian processes. I am not sure that the linear algebra tricks need to be rehashed in the main body of the paper, and this could make room for better discussion of tasks where dependence matters.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear in terms of the proposed approach. The specific tasks being solved require more motivation. To my knowledge, introducing dependency to normalizing flows novel, but I am not an expert in the normalizing flows literature. Many of the technical results are \"textbook\" decompositions of covariance matrices that are well-trod in the ridge regression and GP literatures (probably among others). The experiments seem easily reproducible.", "summary_of_the_review": "The proposed method seems fine in terms of specifying an objective that reflects certain forms of independence. However, the actual tasks being solved require far more explicit discussion, since there are many different tasks that one might try to solve in a dependent data context, and it is not clear which of these the method is designed to address. Because of this concern, the paper feels incomplete.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666796976001}, {"id": "1Xmq7iCaOb", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4574/Reviewer_xwvU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors propose an extension of normalizing flows to non-IID data. This extension can be seen as the generalization of a Gaussian copula to vector-valued variables. The authors propose parameterizations of the resulting model and associated learning algorithms that yield efficient training.", "review_text": "I can see this paper being a useful addition to this line of literature, although I don't see it as completely ground-breaking, It could benefit from a more developed and more thoroughly studied use case.", "strengths": "STRENGTHS:\n* The paper studies an important problem. Although the motivation describes it as non-IID data, this method also applies to high-dimensional correlated vectors, and may be useful for modeling IID data with interesting correlations among the features.\n* This work is novel, and to my knowledge this techniques of extending copula to higher-dimensional datasets is new.\n* The algorithmic ideas proposed in the work are novel, technically sound, and non-trivial. The level of technical depth required to derive the methods is significant.\n* The experiments are extensive and look at a wide range of interesting real-world problems.\n\nWEAKNESSES:\n* These are all relatively minor comments. I think it would have been interesting to describe the motivation and the various applications of the method early on (in the introduction). I only \"got\" the idea and why it's interesting after reading the whole paper. I didn't feel like the intro did a good job at selling the work.\n* While the experiments look at many domains, I didn't feel like the method achieved very strong improvements. Highlighting specific examples where this method really shines could make the paper stronger.\n* The paper mentions GWAS and ancestry-based confounding a lot, and it sounds like an interesting motivating application. It could help make the paper stronger to focus and commit on one application problem more deeply, use it as a clear consistent running example, show strong results on it, etc.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose an extension of normalizing flows to non-IID data. This extension can be seen as the generalization of a Gaussian copula to vector-valued variables. The authors propose parameterizations of the resulting model and associated learning algorithms that yield efficient training.", "strength_and_weaknesses": "STRENGTHS:\n* The paper studies an important problem. Although the motivation describes it as non-IID data, this method also applies to high-dimensional correlated vectors, and may be useful for modeling IID data with interesting correlations among the features.\n* This work is novel, and to my knowledge this techniques of extending copula to higher-dimensional datasets is new.\n* The algorithmic ideas proposed in the work are novel, technically sound, and non-trivial. The level of technical depth required to derive the methods is significant.\n* The experiments are extensive and look at a wide range of interesting real-world problems.\n\nWEAKNESSES:\n* These are all relatively minor comments. I think it would have been interesting to describe the motivation and the various applications of the method early on (in the introduction). I only \"got\" the idea and why it's interesting after reading the whole paper. I didn't feel like the intro did a good job at selling the work.\n* While the experiments look at many domains, I didn't feel like the method achieved very strong improvements. Highlighting specific examples where this method really shines could make the paper stronger.\n* The paper mentions GWAS and ancestry-based confounding a lot, and it sounds like an interesting motivating application. It could help make the paper stronger to focus and commit on one application problem more deeply, use it as a clear consistent running example, show strong results on it, etc.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written. However, it could help make the paper stronger to focus and commit on one application problem more deeply, use it as a clear consistent running example, show strong results on it, etc. The results seem to be novel, high quality, and reproducible.", "summary_of_the_review": "I can see this paper being a useful addition to this line of literature, although I don't see it as completely ground-breaking, It could benefit from a more developed and more thoroughly studied use case.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666657125729}, {"id": "_xuxec_HggN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4574/Reviewer_Krj4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper tries to tackle the problem of learning a generative model for sampling identically distributed but dependent data points from a distribution. Like previous normalizing flows, they try to learn a bijective map of data to an assumed latent space but unlike previous works which assume a diagonal Gaussian, they assume a correlation structure within the latent space as well. This makes the computation of probability density estimates in the latent space intractable in general but by restricting to specific classes of correlation structures the authors design a computation scheme that works for their case of dependent data. The authors test their method on toy and real world datasets.\n", "review_text": "I feel the authors need to strengthen the experiments section considerably in an attempt to justify that dependency modelling is actually useful [other than slight NLL benefits]. Other than that I think the paper is a good attempt at incorporating data dependency assumptions into the normalizing flow literature. I am willing to change my rating based on discussions on the forum.", "strengths": "Strengths:\n\n- The authors chose an oft ignored area in generative modeling: that of dependent sampling which is interesting to analyze.\n- The paper is mostly well written and flows well.\n\nWeaknesses:\n- I am not quite sure why the specific covariance structures were chosen. As an aside do the authors think that adding a regularisation functional that minimises the norm of $\\rho$ will benefit the joint optimization strategy. In a way, would that choose the most decorrelating bijective map?\n- The authors should provide a compute budget difference (both time and memory) between the baseline and proposed methods. It seems the baseline is just a quadratic spline flow whereas the proposed method has additional parameters for the latent correlation structure. An aside: have the authors attempted to optimize the baseline latent mean and covariance?\n- The experiment section lacks many details:\n  -- In UKB: how is the relatedness used? Do you assume G is known?\n  -- In stocks: a few samples would help  understand what is log return referring to? Is it tracking per day movement of a pair? How is the temporal nature of stocks over time incorporated in the model?\n  -- ADNI: To my eyes all images in figure 3 look very similar. Does dependent modeling help in downstream tasks? If so, it would strengthen the author’s case a lot if they could show something of that sort.\n\nOverall I feel the authors would greatly benefit in relegating some exp details to the appendix (around values in grid search, extra dataset details) and adding more results from each of the datasets. For ex. in UKB they say a quantile transformation makes statistical modeling hard. It would strongly benefit the authors to show that their way of modeling makes that easier.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper tries to tackle the problem of learning a generative model for sampling identically distributed but dependent data points from a distribution. Like previous normalizing flows, they try to learn a bijective map of data to an assumed latent space but unlike previous works which assume a diagonal Gaussian, they assume a correlation structure within the latent space as well. This makes the computation of probability density estimates in the latent space intractable in general but by restricting to specific classes of correlation structures the authors design a computation scheme that works for their case of dependent data. The authors test their method on toy and real world datasets.\n", "strength_and_weaknesses": "Strengths:\n\n- The authors chose an oft ignored area in generative modeling: that of dependent sampling which is interesting to analyze.\n- The paper is mostly well written and flows well.\n\nWeaknesses:\n- I am not quite sure why the specific covariance structures were chosen. As an aside do the authors think that adding a regularisation functional that minimises the norm of $\\rho$ will benefit the joint optimization strategy. In a way, would that choose the most decorrelating bijective map?\n- The authors should provide a compute budget difference (both time and memory) between the baseline and proposed methods. It seems the baseline is just a quadratic spline flow whereas the proposed method has additional parameters for the latent correlation structure. An aside: have the authors attempted to optimize the baseline latent mean and covariance?\n- The experiment section lacks many details:\n  -- In UKB: how is the relatedness used? Do you assume G is known?\n  -- In stocks: a few samples would help  understand what is log return referring to? Is it tracking per day movement of a pair? How is the temporal nature of stocks over time incorporated in the model?\n  -- ADNI: To my eyes all images in figure 3 look very similar. Does dependent modeling help in downstream tasks? If so, it would strengthen the author’s case a lot if they could show something of that sort.\n\nOverall I feel the authors would greatly benefit in relegating some exp details to the appendix (around values in grid search, extra dataset details) and adding more results from each of the datasets. For ex. in UKB they say a quantile transformation makes statistical modeling hard. It would strongly benefit the authors to show that their way of modeling makes that easier.\n\n", "clarity,_quality,_novelty_and_reproducibility": "- The writing is mostly quite well done except for experiments (see weaknesses)\n- Adapting normalizing flows to dependent sampling is novel.\n- The authors provide code that will make it reproducible [I have not tried running their code.]\n", "summary_of_the_review": "I feel the authors need to strengthen the experiments section considerably in an attempt to justify that dependency modelling is actually useful [other than slight NLL benefits]. Other than that I think the paper is a good attempt at incorporating data dependency assumptions into the normalizing flow literature. I am willing to change my rating based on discussions on the forum.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666327101522}], "openreview_url": "https://openreview.net/forum?id=Z4lOwCEJQ8Z", "arxiv_id": "2209.14933", "paper_pdf": "papers/Z4lOwCEJQ8Z.pdf", "paper_pdf_sha256": "04d4516c4c72a8ff42c615f1af893c4b701f64673893d2d794c5940bfac0470a", "paper_pdf_bytes": 688970, "paper_pdf_source": "openreview", "code_url": "https://github.com/mkirchler/dependent_data_flows", "code_repository": "mkirchler/dependent_data_flows", "code_commit": "78f40b72215ac57bf3cfb916641ffc73362b6712", "code_archive": "repos/Z4lOwCEJQ8Z.zip", "code_archive_sha256": "49f1b4319a5b96c2b05b36552d07a90bf99ed0f30ac781ad245cf30e279a5869", "code_archive_bytes": 65461, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 35, "github_languages": {"Python": 99433}, "github_archived": false, "github_pushed_at": "2023-10-31T10:56:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/training-normalizing-flows-from-dependent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "I_RLPhVUfw8", "year": 2022, "status": "rejected", "title": "Dense Gaussian Processes for Few-Shot Segmentation", "authors": ["Joakim Johnander", "Johan Edstedt", "Michael Felsberg", "Fahad Khan", "Martin Danelljan"], "authorids": ["~Joakim_Johnander1", "~Johan_Edstedt1", "~Michael_Felsberg2", "~Fahad_Khan1", "~Martin_Danelljan4"], "authors_source": "OpenReview API", "abstract": "Few-shot segmentation is a challenging dense prediction task, which entails segmenting a novel query image given only a small annotated support set. The key problem is thus to design a method that aggregates detailed information from the support set, while being robust to large variations in appearance and context. To this end, we propose a few-shot segmentation method based on dense Gaussian process (GP) regression. Given the support set, our dense GP learns the mapping from local deep image features to mask values, capable of capturing complex appearance distributions. Furthermore, it provides a principled means of capturing uncertainty, which serves as another powerful cue for the final segmentation, obtained by a CNN decoder. Instead of a one-dimensional mask output, we further exploit the end-to-end learning capabilities of our approach to learn a high-dimensional output space for the GP. Our approach sets a new state-of-the-art for both 1-shot and 5-shot FSS on the PASCAL-5$^i$ and COCO-20$^i$ benchmarks, achieving an absolute gain of $+14.9$ mIoU in the COCO-20$^i$ 5-shot setting. Furthermore, the segmentation quality of our approach scales gracefully when increasing the support set size, while achieving robust cross-dataset transfer. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SzXbOOjQULT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3529/Reviewer_nhTR"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a special Gaussian process (GP) named dense GP, to model a mapping between dense local deep features and their corresponding mask values. Based on this dense GP, a few-shot segmentation method named DGPNet is proposed.\nThe authors claim that DGPNet is novel in that it can be applied to situations that unseen classes are not linearly separable, and can produce the uncertainty of its prediction as well. To support this they conduct series of experiments on PASCAL-5^i and COCO-20^i.", "review_text": "- Strengths:\n\n    1. It's a solid idea to introduce uncertainty modeling into few-shot segmentation.\n\n    2. The exact implementation details of the experiments are provided. Although the source code is not presented, the authors show its pseudo-code in SM.\n\n\n\n- Weaknesses:\n\n    1. The way of the GP outputs being used is somehow confusing. The authors claim that \\mu_{Q|S} and \\Delta_{Q|S} are the predicted mean and covariance of the mask of query, i.e., y_Q. However, the prediction of y_Q is fulfilled through DFN (a deep CNN), which means it's not ensured that the predicted \\hat{y}_Q can be sampled from a Gaussian distribution with a relatively high probability.\n\n    2. Since the input of the decoder is a concatenation of the GP outputs and shallow-level query image features, it's necessary to conduct another ablation study to show the influence on the predicted result cast by the GP outputs, i.e., the effectiveness of introducing the GP outputs.\n\n   3. The use of 1-4 background-foreground weighting in the loss function actually performs hard-example mining, which is empirically found beneficial in binary segmentation tasks. As the baseline has already surpassed many of the other methods, the author should apply the proposed method on a standard training scheme, such as the SGD optimizer with the CE loss without weighting, following the practice of PANet, PPNet, PFENet and CANet.\n\n   4. The inference speed should be discussed since an additional network DFN is appended as the decoder.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a special Gaussian process (GP) named dense GP, to model a mapping between dense local deep features and their corresponding mask values. Based on this dense GP, a few-shot segmentation method named DGPNet is proposed.\nThe authors claim that DGPNet is novel in that it can be applied to situations that unseen classes are not linearly separable, and can produce the uncertainty of its prediction as well. To support this they conduct series of experiments on PASCAL-5^i and COCO-20^i.", "main_review": "- Strengths:\n\n    1. It's a solid idea to introduce uncertainty modeling into few-shot segmentation.\n\n    2. The exact implementation details of the experiments are provided. Although the source code is not presented, the authors show its pseudo-code in SM.\n\n\n\n- Weaknesses:\n\n    1. The way of the GP outputs being used is somehow confusing. The authors claim that \\mu_{Q|S} and \\Delta_{Q|S} are the predicted mean and covariance of the mask of query, i.e., y_Q. However, the prediction of y_Q is fulfilled through DFN (a deep CNN), which means it's not ensured that the predicted \\hat{y}_Q can be sampled from a Gaussian distribution with a relatively high probability.\n\n    2. Since the input of the decoder is a concatenation of the GP outputs and shallow-level query image features, it's necessary to conduct another ablation study to show the influence on the predicted result cast by the GP outputs, i.e., the effectiveness of introducing the GP outputs.\n\n   3. The use of 1-4 background-foreground weighting in the loss function actually performs hard-example mining, which is empirically found beneficial in binary segmentation tasks. As the baseline has already surpassed many of the other methods, the author should apply the proposed method on a standard training scheme, such as the SGD optimizer with the CE loss without weighting, following the practice of PANet, PPNet, PFENet and CANet.\n\n   4. The inference speed should be discussed since an additional network DFN is appended as the decoder.", "summary_of_the_review": "This paper is novel mostly in that it utilizes a Gaussian process to generate an extra coarse prediction of the query image. However, an ablation study is lacking to prove the effectiveness of the dense GP module alone. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635925777379}, {"id": "AjGYEAOMc48", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3529/Reviewer_b84d"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a dense Gaussian Process approach for few shot segmentation scenario. Overall promising results are achieved comparing to the recent efforts. ", "review_text": "This paper proposes a dense Gaussian Process approach for few shot segmentation scenario. Overall promising results are achieved comparing to the recent efforts. On the other hand, the paper has some issues to be discussed below.\n*lack of clarification of the major contributions. It seems the main contribution lies in being first to apply the combination of GPs and Neural Networks to the few short semantic segmentation task. It is yet unclear what is the main technical contribution in this paper, despite reading Sec.3 multiple times. \n*It is also surprising that despite being a much simpler method as shown in Fig.2 when comparing to existing methods such as SAGNN (CVPR'21), and despite the internal Gaussian representation is very coarse-scale and lacking-of-details as in Fig.2, the final results of this paper are much better. Is there any intuition of why this is the case?\n*The authors should also present more visual results of e.g. their 5-short segmentation in the appendix/supplementary. It also lacks visual evidence of the claimed uncertainty reasoning benefits.\n*What is the computational cost/time complexity for the training/inference processes, respectively?\n*The authors should provide the implementation publicly available. Otherwise it is difficult for others to validate and reproduce the same results and performance.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a dense Gaussian Process approach for few shot segmentation scenario. Overall promising results are achieved comparing to the recent efforts. ", "main_review": "This paper proposes a dense Gaussian Process approach for few shot segmentation scenario. Overall promising results are achieved comparing to the recent efforts. On the other hand, the paper has some issues to be discussed below.\n*lack of clarification of the major contributions. It seems the main contribution lies in being first to apply the combination of GPs and Neural Networks to the few short semantic segmentation task. It is yet unclear what is the main technical contribution in this paper, despite reading Sec.3 multiple times. \n*It is also surprising that despite being a much simpler method as shown in Fig.2 when comparing to existing methods such as SAGNN (CVPR'21), and despite the internal Gaussian representation is very coarse-scale and lacking-of-details as in Fig.2, the final results of this paper are much better. Is there any intuition of why this is the case?\n*The authors should also present more visual results of e.g. their 5-short segmentation in the appendix/supplementary. It also lacks visual evidence of the claimed uncertainty reasoning benefits.\n*What is the computational cost/time complexity for the training/inference processes, respectively?\n*The authors should provide the implementation publicly available. Otherwise it is difficult for others to validate and reproduce the same results and performance.\n", "summary_of_the_review": "many aspects are unclear", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635918556108}, {"id": "8LFfyQ_GD36", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3529/Reviewer_uRb5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors propose a novel few-shot segmentation method by adopting dense Gaussian process (GP) regression to capture complex appearance distributions. To boot the performance, authors consider the uncertainty in the final segmentation. Authors exploit the end-to-end learning capabilities of the proposed method to learn a high-dimensional output space for the GP. \nAuthors report state-of-the-art results in two public few shot segmentation benchmarks.", "review_text": "Strengths:\n- The idea of adopting dense Gaussian Process (GP) regression to learn the mapping from local deep image features to mask values, as well as considering uncertainty for the final segmentation is interesting. \n- The paper is well written and the methodology technically sound.\n- Authors report extensive experiments and the results achieved are very competitive.\n\nWeaknesses:\n- The idea that utilizing Gaussian processes in the context of few-shot classification has been exploited in [r1,r2]. Although the proposed method focuses on the segmentation task, it seems like a dense classification setting of r1 and r2. Authors need to clarify the main differences. \n- Does this method only focus on 1-way (binary segmentation) setting? Can it solve the multiple novel classes in a single episode?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Authors propose a novel few-shot segmentation method by adopting dense Gaussian process (GP) regression to capture complex appearance distributions. To boot the performance, authors consider the uncertainty in the final segmentation. Authors exploit the end-to-end learning capabilities of the proposed method to learn a high-dimensional output space for the GP. \nAuthors report state-of-the-art results in two public few shot segmentation benchmarks.", "main_review": "Strengths:\n- The idea of adopting dense Gaussian Process (GP) regression to learn the mapping from local deep image features to mask values, as well as considering uncertainty for the final segmentation is interesting. \n- The paper is well written and the methodology technically sound.\n- Authors report extensive experiments and the results achieved are very competitive.\n\nWeaknesses:\n- The idea that utilizing Gaussian processes in the context of few-shot classification has been exploited in [r1,r2]. Although the proposed method focuses on the segmentation task, it seems like a dense classification setting of r1 and r2. Authors need to clarify the main differences. \n- Does this method only focus on 1-way (binary segmentation) setting? Can it solve the multiple novel classes in a single episode?", "summary_of_the_review": "The idea of adopting dense Gaussian Process (GP) regression to learn the mapping from local deep image features to mask values, as well as considering uncertainty for the final segmentation is interesting. Although the proposed method seems like a dense setting of the existing methods, the paper is well written and the results achieved are very competitive. I'd like to enhance my rating if the authors address all my concerns.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635910749372}, {"id": "YHfvl9qbhBI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3529/Reviewer_Kjwm"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper strives for solving the few-shot segmentation problem, they propose to incorporate the Gaussian Process(GP) into the framework of few-shot segmentation. Except that, they also exploit the high-dimensional output space for GP, the result reaches state-of-the-art in two benchmarks,  one bonus is that  the segmentation quality scales gracefully as increasing the support set size.", "review_text": "Pro:\n- A novel idea to combine GP and FSS.\n- Extensive experiments to validate the results on different benchmarks\n- State-of-the-art on two benchmarks, and gracefully extend to high-shot segmentation.\n\n\nCon:\n- Hyper-parameters sensitivity. Are the methods sensitive to different hyper-parameter in GP? The author should mention this in the paper.\n- What is your implementation without Cov and GPO in Table 4? The detail is missing here.\n- For me, it is still not clear how the input are interplayed in encoder.(E.3 only introduced about restoring spatial structure of the output of the GP)\n- In contribution part, mentioned the probabilistic modeling, I don't see its importance in experimental parts.\n- Visualization for 5 shot is not provided, what's the main improvement from 1shot to 5shot to 10shot?\n- How about removing the contribution of f branch, only keeping shallow x?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper strives for solving the few-shot segmentation problem, they propose to incorporate the Gaussian Process(GP) into the framework of few-shot segmentation. Except that, they also exploit the high-dimensional output space for GP, the result reaches state-of-the-art in two benchmarks,  one bonus is that  the segmentation quality scales gracefully as increasing the support set size.", "main_review": "Pro:\n- A novel idea to combine GP and FSS.\n- Extensive experiments to validate the results on different benchmarks\n- State-of-the-art on two benchmarks, and gracefully extend to high-shot segmentation.\n\n\nCon:\n- Hyper-parameters sensitivity. Are the methods sensitive to different hyper-parameter in GP? The author should mention this in the paper.\n- What is your implementation without Cov and GPO in Table 4? The detail is missing here.\n- For me, it is still not clear how the input are interplayed in encoder.(E.3 only introduced about restoring spatial structure of the output of the GP)\n- In contribution part, mentioned the probabilistic modeling, I don't see its importance in experimental parts.\n- Visualization for 5 shot is not provided, what's the main improvement from 1shot to 5shot to 10shot?\n- How about removing the contribution of f branch, only keeping shallow x?", "summary_of_the_review": "The paper overall presents a novel idea with solid empirical validations. Except some minor concern I raised in previous parts. In total, I am inclined to accept this paper.  I am happy to raise the score if my concern is fully resolved.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635856742346}], "openreview_url": "https://openreview.net/forum?id=I_RLPhVUfw8", "arxiv_id": "2110.03674", "paper_pdf": "papers/I_RLPhVUfw8.pdf", "paper_pdf_sha256": "e6fbc6d933d4d8f6ade1707021dbabc78b401b62f673b567a8e3e8af938741bf", "paper_pdf_bytes": 3132078, "paper_pdf_source": "openreview", "code_url": "https://github.com/joakimjohnander/DGPNet", "code_repository": "joakimjohnander/DGPNet", "code_commit": "c610c14554ae73abf8f968ea4f73eb8211b532c2", "code_archive": "repos/I_RLPhVUfw8.zip", "code_archive_sha256": "3a8deaf08c38be593a4611f00dfb0f86245350f11edb58385b140df3086af19e", "code_archive_bytes": 41080, "code_file_count": 16, "code_extensions": {".py": 15, ".sh": 1}, "github_disk_usage_kb": 33, "github_languages": {"Python": 122861, "Shell": 551}, "github_archived": false, "github_pushed_at": "2022-09-28T07:11:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dense-gaussian-processes-for-few-shot"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YjXnezbeCwG", "year": 2021, "status": "rejected", "title": "Learning to Use Future Information in Simultaneous Translation", "authors": ["Xueqing Wu", "Yingce Xia", "Lijun Wu", "Shufang Xie", "Weiqing Liu", "Tao Qin", "Tie-Yan Liu"], "authorids": ["~Xueqing_Wu1", "~Yingce_Xia1", "~Lijun_Wu1", "~Shufang_Xie1", "weiqing.liu@microsoft.com", "~Tao_Qin1", "~Tie-Yan_Liu1"], "authors_source": "OpenReview API", "abstract": "Simultaneous neural machine translation (briefly, NMT) has attracted much attention recently. In contrast to standard NMT, where the NMT system can access the full input sentence, simultaneous NMT is a prefix-to-prefix problem, where the system can only utilize the prefix of the input sentence and thus more uncertainty and difficulty are introduced to decoding. Wait-k inference is a simple yet effective strategy for simultaneous NMT, where the decoder generates the output sequence $k$ words behind the input words. For wait-k inference, we observe that wait-m training with $m>k$ in simultaneous NMT (i.e., using more future information for training than inference) generally outperforms wait-k training. Based on this observation, we propose a method that automatically learns how much future information to use in training for simultaneous NMT. Specifically, we introduce a controller to adaptively select wait-m training strategies according to the network status of the translation model and current training sentence pairs, and the controller is jointly trained with the translation model through bi-level optimization. Experiments on four datasets show that our method brings 1 to 3 BLEU point improvement over baselines under the same latency. Our code is available at https://github.com/P2F-research/simulNMT .", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "E0QnlM00xUq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2869/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors observed that some lookahead information during training time is helpful to improve the translation accuracy for simultaneous translation. Base on this observation, this paper proposes to use RL-based methods to learn a certain number of lookahead words during the training of the wait-k-based simultaneous translation model.\n\nThis paper proposes a new approach for improving the translation quality and the results indeed show some improvements over the baseline methods. However, I still have the following concerns:\n1) I think the proposed RL-based methods are very completed (in terms of hyperparameter searching, extra training time compared with other baselines) but the improvements are quite marginal compared with wait-k* and random.\n\n2) For the experiments, the authors did not compare with other agent-based or adaptive methods. \n\n3) This paper only designs a controller for training. I think we could also have a controller for the inference. In this way, I believe we could use regular wait-k training, and use a controller to decide a smaller k during inference. Then it will be very similar to Gu. et al's RL-based methods. I suggest the authors include this method's experiments as well.\n\n4) If I did not misunderstand, the m of wait-m is defined on each training pair, but I think this will dramatically increase the training time and hard to do batch training. How slow is your training compared with baseline wait-k? I believe the baseline wait-k is already very slow.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "complicated method with marginal improvements", "review": "The authors observed that some lookahead information during training time is helpful to improve the translation accuracy for simultaneous translation. Base on this observation, this paper proposes to use RL-based methods to learn a certain number of lookahead words during the training of the wait-k-based simultaneous translation model.\n\nThis paper proposes a new approach for improving the translation quality and the results indeed show some improvements over the baseline methods. However, I still have the following concerns:\n1) I think the proposed RL-based methods are very completed (in terms of hyperparameter searching, extra training time compared with other baselines) but the improvements are quite marginal compared with wait-k* and random.\n\n2) For the experiments, the authors did not compare with other agent-based or adaptive methods. \n\n3) This paper only designs a controller for training. I think we could also have a controller for the inference. In this way, I believe we could use regular wait-k training, and use a controller to decide a smaller k during inference. Then it will be very similar to Gu. et al's RL-based methods. I suggest the authors include this method's experiments as well.\n\n4) If I did not misunderstand, the m of wait-m is defined on each training pair, but I think this will dramatically increase the training time and hard to do batch training. How slow is your training compared with baseline wait-k? I believe the baseline wait-k is already very slow.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603927500425}, {"id": "3pHCV0Rw_Z2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2869/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a training strategy for simultaneous translation to choose appropriate amount of look-ahead information for each decoding. Based on the observation that the wait-k method can be improved by training with longer information, the method introduces a function to determine its length given the current example (source x, target y, and the translation model f). It is used as only a guidance during training, and it remains the same decoding criterion at inference.\n\nThe method is interesting and has promising improvements compared with bare wait-k methods according to the experiments, but the paper seems to have some major questions which should affect the conclusion. I recommend to revise the paper appropriately, especially to resolve following concerns:\n\n- The proposed method is intuitively strange because of mismatching between training/inference strategies. It is also unclear to choose this method rather than adaptive wait-k methods, e.g., one referred as Zheng et al. (2020a), which may solve a similar problem directly. The paper needs at least some comparison of these kind of methods to figure out the advantages of the proposed method.\n- Algorithm 1 involves a suspicious use of the development set: it is used directly to optimize a parameter. Specifically, since \\omega is optimized using the D_va, it brings information of the D_va into f, resulting that the training process does not include any strategy to avoid overfitting (in other words, your training set is actually D_tr + D_va, and there is no so-called development set).\n\nMinor comments:\n\n- The title sounds misleading: the proposed method still does not learn how to use future information because it uses only k look-ahead information at inference (same as usual wait-k methods), i.e., there is nothing special to represent \"future\".\n- Figure 1 involves some common mistakes of using bar charts: it must use 0 as the origin, and must not shorten the bar. If you want to focus on differences between each value, you should use other chart.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising methods, but suspicious settings of model training", "review": "This paper proposes a training strategy for simultaneous translation to choose appropriate amount of look-ahead information for each decoding. Based on the observation that the wait-k method can be improved by training with longer information, the method introduces a function to determine its length given the current example (source x, target y, and the translation model f). It is used as only a guidance during training, and it remains the same decoding criterion at inference.\n\nThe method is interesting and has promising improvements compared with bare wait-k methods according to the experiments, but the paper seems to have some major questions which should affect the conclusion. I recommend to revise the paper appropriately, especially to resolve following concerns:\n\n- The proposed method is intuitively strange because of mismatching between training/inference strategies. It is also unclear to choose this method rather than adaptive wait-k methods, e.g., one referred as Zheng et al. (2020a), which may solve a similar problem directly. The paper needs at least some comparison of these kind of methods to figure out the advantages of the proposed method.\n- Algorithm 1 involves a suspicious use of the development set: it is used directly to optimize a parameter. Specifically, since \\omega is optimized using the D_va, it brings information of the D_va into f, resulting that the training process does not include any strategy to avoid overfitting (in other words, your training set is actually D_tr + D_va, and there is no so-called development set).\n\nMinor comments:\n\n- The title sounds misleading: the proposed method still does not learn how to use future information because it uses only k look-ahead information at inference (same as usual wait-k methods), i.e., there is nothing special to represent \"future\".\n- Figure 1 involves some common mistakes of using bar charts: it must use 0 as the origin, and must not shorten the bar. If you want to focus on differences between each value, you should use other chart.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603922468970}, {"id": "Xzgy6V12UU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2869/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new training method for wait-k simultaneous translation. Rather than training on prefix pairs where the target prefix lags the source by k tokens, it uses an RL controller to determine an optimal lag for each sentence pair. The controller uses a small set of features intended to capture training progress, and is trained with REINFORCE to minimize wait-k loss on a validation set, in alternation with main training steps. This method shows consistent gains over various wait-k training heuristics, and some gains over other approaches that adapt the lag at inference time.\n\nThe paper is very well written and organized. The method makes sense, and the experiments are quite thorough, comparing to a competitive set of heuristic baselines, and showing credible - though fairly modest - gains in this setting.\n\nThe comparisons to adaptive baselines are less convincing. They are shown only for one small-data language pair, no implementation details are given (for instance, architectures and model capacities), and the relative results are very different from those in the literature (MILK << wait-k, WIW, WID). I think the paper would be stronger if these were simply omitted.\n\nGiven its ease of implementation and efficiency, I think there is room for a paper focusing on improving wait-k training, even if wait-k isn’t quite state of the art. But having to set up an RL controller detracts from this picture, especially since the gains over various heuristics - in particular the random heuristic that works for any inference-time k - aren’t spectacular. My main reaction to this work is that it should be possible to get most of the gains using some predetermined curriculum inspired by figure 5 (random sampling at the beginning, annealing to some m > k) that is effective for a broad range of inference-time k’s. Note that this is quite different from the CL baseline included in the results. As it stands, I fear this paper risks being a dead end: too complex to be worth implementing, as the field moves on past wait-k.\n\nQuestions and suggestions:\n\n1. Controller feature (6): will this value ever repeat during an entire training run?\n2. For Mk, since this runs only over the validation set and you are using RL, why not use the actual wait-k BLEU score?\n3. You need to include all heuristic baselines in figure 4.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Using a sledgehammer to kill a fly?", "review": "This paper proposes a new training method for wait-k simultaneous translation. Rather than training on prefix pairs where the target prefix lags the source by k tokens, it uses an RL controller to determine an optimal lag for each sentence pair. The controller uses a small set of features intended to capture training progress, and is trained with REINFORCE to minimize wait-k loss on a validation set, in alternation with main training steps. This method shows consistent gains over various wait-k training heuristics, and some gains over other approaches that adapt the lag at inference time.\n\nThe paper is very well written and organized. The method makes sense, and the experiments are quite thorough, comparing to a competitive set of heuristic baselines, and showing credible - though fairly modest - gains in this setting.\n\nThe comparisons to adaptive baselines are less convincing. They are shown only for one small-data language pair, no implementation details are given (for instance, architectures and model capacities), and the relative results are very different from those in the literature (MILK << wait-k, WIW, WID). I think the paper would be stronger if these were simply omitted.\n\nGiven its ease of implementation and efficiency, I think there is room for a paper focusing on improving wait-k training, even if wait-k isn’t quite state of the art. But having to set up an RL controller detracts from this picture, especially since the gains over various heuristics - in particular the random heuristic that works for any inference-time k - aren’t spectacular. My main reaction to this work is that it should be possible to get most of the gains using some predetermined curriculum inspired by figure 5 (random sampling at the beginning, annealing to some m > k) that is effective for a broad range of inference-time k’s. Note that this is quite different from the CL baseline included in the results. As it stands, I fear this paper risks being a dead end: too complex to be worth implementing, as the field moves on past wait-k.\n\nQuestions and suggestions:\n\n1. Controller feature (6): will this value ever repeat during an entire training run?\n2. For Mk, since this runs only over the validation set and you are using RL, why not use the actual wait-k BLEU score?\n3. You need to include all heuristic baselines in figure 4.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603900410220}, {"id": "BXmSWQDi7qu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2869/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper improves the wait-k based simultaneous NMT by training on an adaptive wait-m policy. The proposed method and experiments are clearly described. Experiments demonstrate that the proposed method is significant better than the wait-k baseline.\n\nHowever, compared to heuristic-based baselines, seemingly the proposed method is not significantly better in most cases, especially on the WMT En-De dataset. Given that the WMT dataset is much larger than IWSLT datasets, does this suggest that the proposed method may not work well on larger dataset?\n\nI’m also curious about the reason of using adaptive wait-m only during training since it is a more straightforward idea to apply adaptive policy during both training and inference. Would it be better if the adaptive policy (by re-design features) was used on inference as well? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The effectiveness of the proposed method is not convincing", "review": "This paper improves the wait-k based simultaneous NMT by training on an adaptive wait-m policy. The proposed method and experiments are clearly described. Experiments demonstrate that the proposed method is significant better than the wait-k baseline.\n\nHowever, compared to heuristic-based baselines, seemingly the proposed method is not significantly better in most cases, especially on the WMT En-De dataset. Given that the WMT dataset is much larger than IWSLT datasets, does this suggest that the proposed method may not work well on larger dataset?\n\nI’m also curious about the reason of using adaptive wait-m only during training since it is a more straightforward idea to apply adaptive policy during both training and inference. Would it be better if the adaptive policy (by re-design features) was used on inference as well? \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603890621199}], "openreview_url": "https://openreview.net/forum?id=YjXnezbeCwG", "arxiv_id": null, "paper_pdf": "papers/YjXnezbeCwG.pdf", "paper_pdf_sha256": "d5ca022554f0e1a7303c0e24f6f76d052b68b765635eb3b94227fbe4cf6b8df3", "paper_pdf_bytes": 579358, "paper_pdf_source": "openreview", "code_url": "https://github.com/P2F-research/simulNMT", "code_repository": "P2F-research/simulNMT", "code_commit": "4721707f7cf0beea39d7ee8accb058eb2c7394fc", "code_archive": "repos/YjXnezbeCwG.zip", "code_archive_sha256": "bb69a76dd4f44b4bffefd0c120ad26ccf78a6ffa04424eb9b52211ffb5fc345d", "code_archive_bytes": 58090, "code_file_count": 28, "code_extensions": {".py": 22, ".sh": 6}, "github_disk_usage_kb": 40, "github_languages": {"Python": 163307, "Shell": 6267}, "github_archived": false, "github_pushed_at": "2020-10-05T11:42:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-use-future-information-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rygePJHYPH", "year": 2020, "status": "rejected", "title": "Towards trustworthy predictions from deep neural networks with fast adversarial calibration", "authors": ["Christian Tomani", "Florian Buettner"], "authorids": ["christian.tomani@gmail.com", "fbuettner.phys@gmail.com"], "authors_source": "OpenReview API", "abstract": "To facilitate a wide-spread acceptance of AI systems guiding decision making in real-world applications, trustworthiness of deployed models is key. That is, it is crucial for predictive models to be uncertainty-aware and yield well-calibrated (and thus trustworthy) predictions for both in-domain samples as well as under domain shift. Recent efforts to account for predictive uncertainty include post-processing steps for trained neural networks, Bayesian neural networks as well as alternative non-Bayesian approaches such as ensemble approaches and evidential deep learning. Here, we propose an efficient yet general modelling approach for obtaining well-calibrated, trustworthy probabilities for samples obtained after a domain shift. We introduce a new training strategy combining an entropy-encouraging loss term with an adversarial calibration loss term and demonstrate that this results in well-calibrated and technically trustworthy predictions for a wide range of perturbations. We comprehensively evaluate previously proposed approaches on different data modalities, a large range of data sets, network architectures and perturbation strategies and observe that our modelling approach substantially outperforms existing state-of-the-art approaches, yielding well-calibrated predictions for both in-domain and out-of domain samples. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H1xalm-RKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1753/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new loss function for training deep neural networks, which show good performance with respect to well-calibrated, trustworthy probabilities for samples after a domain shift. The authors conduct experiments with multiple datasets and multiple forms of perturbations, where the proposed method achieve superior performance. \n\nI have the following major concerns with the paper:\n\n1. Presentation: The paper's presentation is very weak. It contains repetitive, long, convoluted statements and paragraphs all throughout making it difficult for the reader to understand anything. The first time I read this paper, I couldn't process what is happening. The introduction is more like realted work with less focus on what they are trying to pitch in the paper. However, the latter is the less severe concern.\n\n2. While I appreciate the use of reliability diagrams and the loss function inspired from it, I do not completely understand what are the entities in equations 1 and 2 or how they are used later. Please give them a name, and elaborate on them.\n\n3. While I was reading the paragraph before section 2.2 the first time, it seems to me that ECE is defined for those 10 pertubations specifically. I believe this is not the case; hence, the authors should make it more general and divert the specific details to experiments. \n\n4. I am not sure what the authors mean by, \"... while the loss surface remains largely unchanged\", in paragraph above section 2.2.2. I believe the authors have constructed a new loss function by adding a new term to it. That makes it difficult to understand the advantage the authors are talking about after this statement. Overall, I am not really convinced how in contrast to Bayesian Deep Learning, their approach can be EASILY applied to LSTM's and GRU's.\n\n5. What is L_{adv}? There is no equation number, no discussion around where this is defined. Again a presentation issue.\n\nMinor: Please clarify the reference for supplementary materials. For example, it is not clear that Table S1 is in supplementary material.\n\nOverall, this paper is good and has an interesting idea. The experiments are also extensive useful. However, I have reservations regarding the presentation of this paper at this moment. \n\n--- After Rebuttal  ---\n\nI thank the authors for providing a detailed response to my questions and editing the paper. I am now more positive about the paper; however, I still feel the presentation of the paper could be further improved. At this point, I will not move to acceptance range and keep the same score. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This paper introduces a new loss function for training deep neural networks, which show good performance with respect to well-calibrated, trustworthy probabilities for samples after a domain shift. The authors conduct experiments with multiple datasets and multiple forms of perturbations, where the proposed method achieve superior performance. \n\nI have the following major concerns with the paper:\n\n1. Presentation: The paper's presentation is very weak. It contains repetitive, long, convoluted statements and paragraphs all throughout making it difficult for the reader to understand anything. The first time I read this paper, I couldn't process what is happening. The introduction is more like realted work with less focus on what they are trying to pitch in the paper. However, the latter is the less severe concern.\n\n2. While I appreciate the use of reliability diagrams and the loss function inspired from it, I do not completely understand what are the entities in equations 1 and 2 or how they are used later. Please give them a name, and elaborate on them.\n\n3. While I was reading the paragraph before section 2.2 the first time, it seems to me that ECE is defined for those 10 pertubations specifically. I believe this is not the case; hence, the authors should make it more general and divert the specific details to experiments. \n\n4. I am not sure what the authors mean by, \"... while the loss surface remains largely unchanged\", in paragraph above section 2.2.2. I believe the authors have constructed a new loss function by adding a new term to it. That makes it difficult to understand the advantage the authors are talking about after this statement. Overall, I am not really convinced how in contrast to Bayesian Deep Learning, their approach can be EASILY applied to LSTM's and GRU's.\n\n5. What is L_{adv}? There is no equation number, no discussion around where this is defined. Again a presentation issue.\n\nMinor: Please clarify the reference for supplementary materials. For example, it is not clear that Table S1 is in supplementary material.\n\nOverall, this paper is good and has an interesting idea. The experiments are also extensive useful. However, I have reservations regarding the presentation of this paper at this moment. \n\n--- After Rebuttal  ---\n\nI thank the authors for providing a detailed response to my questions and editing the paper. I am now more positive about the paper; however, I still feel the presentation of the paper could be further improved. At this point, I will not move to acceptance range and keep the same score. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571848949286}, {"id": "BklNA9L6tS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1753/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a method for calibrating neural networks on in- and out-of-distribution data using two additional loss terms: entropy-encouraging loss term which maximizes softmax probabilities for wrong classes and adversarial calibration loss term which pushes confidences to match accuracy on adversarial examples. The idea of using adversarial calibration training is interesting and promising, however, the clarity of the paper needs significant improvement and there are several issues which need to be addressed; for this reason, I recommend a weak reject for the current version.\n\n1. One of the contributions listed in the paper is that authors “illustrate the limitations of entropy as measure for trustworthy predictions and introduce a new metric to quantify technical trustworthiness based on the concept of calibration”; the section 2.1 discusses this in detail. The proposed metric is expected calibration error (ECE) averaged over different levels of noise perturbations applied to data during test time. This metric assumes that for a given dataset we know in advance what kind of noise perturbations can be implied to cause out-of-domain/domain-shift scenarios during test time. This assumption most likely doesn’t hold in real-world applications since domain shifts may be of various kinds. Moreover, calibration may not make sense at all for out-of-domain data if test inputs don’t belong to any of the train classes (as opposed to entropy of predictive distribution which still can be computed and is expected to be higher for such inputs). The metric is also dependent on the considered noise level range; the plot of ECE vs noise level can be illustrative and informative while the averaged value of ECE over all noise levels can be misleading (at least, we may want to have some discount factor to account for the fact that with noise level 0 we strongly care about calibration, while for very high noise levels calibration doesn’t give us much information since the useful patterns in data may be corrupted and it may be impossible even for humans to classify such objects, e.g. on Figure 1-top-left for noise level 100, the digit 6 is corrupted so much that we wouldn’t care about calibration for such inputs but just want to maximize the entropy).\nThe paper also claims “Recent efforts in terms of evaluating predictive uncertainty have focused on entropy as measure for uncertainty-awareness for predictions under domain shift.” In previous work addressing uncertainty in domain shift [1], not only entropy but other various metrics are considered for out-of-domain detection (Brier score, thresholded confidences, etc), these metrics don’t depend on particular perturbations and are very informative.\nThe introduced metric, ECE for different noise labels, makes sense only in the toy scenarios where we can control the noise level, however, it would still be better to either look at the plots or to use some discounted averaged ECE over different noise levels, but not equal average. So the significance of this contribution (introduction of new metric) is limited.  \n\n2. The clarity of the paper could be significantly improved.\n\n(a) Figure 1: the top left plot suggests that for high noise level “wrong predictions are often made with high confidence” (so the entropy of this distribution is low) while the top right plot shows that on average predictive entropy gradually grows and is high for high noise levels, so the top right plot is probably not a representative example and it might be misleading to claim this is an “often” case.\n\n(b) Could you please clarify what you mean by “after removing non-misleading evidence” in section 2.2.1? In the next sentence, the remaining probability is probably distribution uniformly across C-1, not C, classes. The predictive entropy loss term essentially maximizes the probabilities of wrong classes with a lower coefficient. How would you support the claim that the loss surface is unchanged?\nThe loss term is “parameter-free” but still we need to tune a coefficient lambda_S for it.\n\n(c) The adversarial loss equation in section 2.2.2 is written as L2-norm of a scalar value, why is L2-norm needed? The acc(B_m) is not differentiable so it is probably just considered as constant in the paper? Please comment on that. \n\n(d) Section 3.2 and Figure 4. On Figure 4 Middle and Right plots have different colors than those listed on legend, please, fix this. On the middle plot, both FALCON and EDL have high ECE for noise levels <50 which indicates that they are both highly underconfident for in-distribution data and low noise levels, while other methods have close to 0 ECE on low noise levels. The authors only comment on EDL underconfidence: “it is worth noting that EDL has a substantially higher ECE for in-domain predictions, reflecting under-confident predictions on the test set”, and not on FALCON underconfidence. However, using the proposed score, ECE averaged over all noise levels (Table 2), it may look like the method is doing a good job, while underconfidence problem is revealed when looking at Figure 4. This is also an illustration of the concern about the proposed metric I raised in point 1.\n\n3.  It would help to have an ablation study in the main text of the paper showing the significance of each loss term: in the appendix Figure S1, it is shown that adding adversarial loss to standard and entropy loss helps. Is the entropy loss needed at all? How does performance change if we only have standard and adversarial loss? How sensitive is performance to the choice of lambda_avd and lambda_s?\n\n\n[1] Ovadia, Yaniv, et al. \"Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift.\" arXiv preprint arXiv:1906.02530 (2019).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper presents a method for calibrating neural networks on in- and out-of-distribution data using two additional loss terms: entropy-encouraging loss term which maximizes softmax probabilities for wrong classes and adversarial calibration loss term which pushes confidences to match accuracy on adversarial examples. The idea of using adversarial calibration training is interesting and promising, however, the clarity of the paper needs significant improvement and there are several issues which need to be addressed; for this reason, I recommend a weak reject for the current version.\n\n1. One of the contributions listed in the paper is that authors “illustrate the limitations of entropy as measure for trustworthy predictions and introduce a new metric to quantify technical trustworthiness based on the concept of calibration”; the section 2.1 discusses this in detail. The proposed metric is expected calibration error (ECE) averaged over different levels of noise perturbations applied to data during test time. This metric assumes that for a given dataset we know in advance what kind of noise perturbations can be implied to cause out-of-domain/domain-shift scenarios during test time. This assumption most likely doesn’t hold in real-world applications since domain shifts may be of various kinds. Moreover, calibration may not make sense at all for out-of-domain data if test inputs don’t belong to any of the train classes (as opposed to entropy of predictive distribution which still can be computed and is expected to be higher for such inputs). The metric is also dependent on the considered noise level range; the plot of ECE vs noise level can be illustrative and informative while the averaged value of ECE over all noise levels can be misleading (at least, we may want to have some discount factor to account for the fact that with noise level 0 we strongly care about calibration, while for very high noise levels calibration doesn’t give us much information since the useful patterns in data may be corrupted and it may be impossible even for humans to classify such objects, e.g. on Figure 1-top-left for noise level 100, the digit 6 is corrupted so much that we wouldn’t care about calibration for such inputs but just want to maximize the entropy).\nThe paper also claims “Recent efforts in terms of evaluating predictive uncertainty have focused on entropy as measure for uncertainty-awareness for predictions under domain shift.” In previous work addressing uncertainty in domain shift [1], not only entropy but other various metrics are considered for out-of-domain detection (Brier score, thresholded confidences, etc), these metrics don’t depend on particular perturbations and are very informative.\nThe introduced metric, ECE for different noise labels, makes sense only in the toy scenarios where we can control the noise level, however, it would still be better to either look at the plots or to use some discounted averaged ECE over different noise levels, but not equal average. So the significance of this contribution (introduction of new metric) is limited.  \n\n2. The clarity of the paper could be significantly improved.\n\n(a) Figure 1: the top left plot suggests that for high noise level “wrong predictions are often made with high confidence” (so the entropy of this distribution is low) while the top right plot shows that on average predictive entropy gradually grows and is high for high noise levels, so the top right plot is probably not a representative example and it might be misleading to claim this is an “often” case.\n\n(b) Could you please clarify what you mean by “after removing non-misleading evidence” in section 2.2.1? In the next sentence, the remaining probability is probably distribution uniformly across C-1, not C, classes. The predictive entropy loss term essentially maximizes the probabilities of wrong classes with a lower coefficient. How would you support the claim that the loss surface is unchanged?\nThe loss term is “parameter-free” but still we need to tune a coefficient lambda_S for it.\n\n(c) The adversarial loss equation in section 2.2.2 is written as L2-norm of a scalar value, why is L2-norm needed? The acc(B_m) is not differentiable so it is probably just considered as constant in the paper? Please comment on that. \n\n(d) Section 3.2 and Figure 4. On Figure 4 Middle and Right plots have different colors than those listed on legend, please, fix this. On the middle plot, both FALCON and EDL have high ECE for noise levels <50 which indicates that they are both highly underconfident for in-distribution data and low noise levels, while other methods have close to 0 ECE on low noise levels. The authors only comment on EDL underconfidence: “it is worth noting that EDL has a substantially higher ECE for in-domain predictions, reflecting under-confident predictions on the test set”, and not on FALCON underconfidence. However, using the proposed score, ECE averaged over all noise levels (Table 2), it may look like the method is doing a good job, while underconfidence problem is revealed when looking at Figure 4. This is also an illustration of the concern about the proposed metric I raised in point 1.\n\n3.  It would help to have an ablation study in the main text of the paper showing the significance of each loss term: in the appendix Figure S1, it is shown that adding adversarial loss to standard and entropy loss helps. Is the entropy loss needed at all? How does performance change if we only have standard and adversarial loss? How sensitive is performance to the choice of lambda_avd and lambda_s?\n\n\n[1] Ovadia, Yaniv, et al. \"Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift.\" arXiv preprint arXiv:1906.02530 (2019)."}, "tcdate": 1571805900453}, {"id": "r1leOMl3tH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1753/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed FALCON, a simple method to produce well-calibrated uncertainty estimation. The idea is to introduce two additional terms, one that directly encourage lower confidence for all negative classes of all data points, and another one that optimizes the ECE for adversarial samples. Experiments show that FALCON outperforms several state-of-the-art methods for calibrating neural network predictions.\n\nAlthough the first term, L_S, affects only negative predictions, it is still somewhat strange to uniformly operate on all data points. It would help to perform an ablation study to see how L_S affect the results.\n\nSome description in Section 2 could use more mathematical rigor (e.g., when describing L_{adv}).\n\nThe authors use EDL as a baseline, I was wondering why not use the more commonly used, and possibly more effective, temperature scaling (TS) method from Guo et al. 2017. Note that the EDL paper does not seem to explicitly compare EDL and TS. \n\nFigure 4 (middle and right) is confusing. The line style is not consistent in the figures.\n\nThere are a few places where the text is rather vague and confusing. For example, what do you mean by ‘non-misleading evidence’ when describing L_{adv}? It would also be better to provide more insight more L_S to help the readers out. For example, it would help to state that L_S operate only on negative predictions.\n\nMost baselines are rather simple non-probabilistic (non-Bayesian) methods. Besides, MNF, It would also be interesting to see how FALCON compare other probabilistic NN method such as natural parameter networks, where they also explicitly evaluated uncertainty estimation.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper proposed FALCON, a simple method to produce well-calibrated uncertainty estimation. The idea is to introduce two additional terms, one that directly encourage lower confidence for all negative classes of all data points, and another one that optimizes the ECE for adversarial samples. Experiments show that FALCON outperforms several state-of-the-art methods for calibrating neural network predictions.\n\nAlthough the first term, L_S, affects only negative predictions, it is still somewhat strange to uniformly operate on all data points. It would help to perform an ablation study to see how L_S affect the results.\n\nSome description in Section 2 could use more mathematical rigor (e.g., when describing L_{adv}).\n\nThe authors use EDL as a baseline, I was wondering why not use the more commonly used, and possibly more effective, temperature scaling (TS) method from Guo et al. 2017. Note that the EDL paper does not seem to explicitly compare EDL and TS. \n\nFigure 4 (middle and right) is confusing. The line style is not consistent in the figures.\n\nThere are a few places where the text is rather vague and confusing. For example, what do you mean by ‘non-misleading evidence’ when describing L_{adv}? It would also be better to provide more insight more L_S to help the readers out. For example, it would help to state that L_S operate only on negative predictions.\n\nMost baselines are rather simple non-probabilistic (non-Bayesian) methods. Besides, MNF, It would also be interesting to see how FALCON compare other probabilistic NN method such as natural parameter networks, where they also explicitly evaluated uncertainty estimation.\n"}, "tcdate": 1571713639998}], "openreview_url": "https://openreview.net/forum?id=rygePJHYPH", "arxiv_id": "2012.10923", "paper_pdf": "papers/rygePJHYPH.pdf", "paper_pdf_sha256": "2017b808a8159561c24a00ecd67ea98cf2b20a0f856210d5f272d2809179ae9d", "paper_pdf_bytes": 1644403, "paper_pdf_source": "openreview", "code_url": "https://github.com/tochris/falcon", "code_repository": "tochris/falcon", "code_commit": "6e0ff0f4359b6b8e8b4c843d7e6eb097f6e1a39f", "code_archive": "repos/rygePJHYPH.zip", "code_archive_sha256": "afc03e3eb65bf2a720e57180928ce5f9b0c6f8ae11f22e52a63c48bb42ce3ca6", "code_archive_bytes": 50864, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 45, "github_languages": {"Python": 198841}, "github_archived": false, "github_pushed_at": "2021-02-24T10:40:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-trustworthy-predictions-from-deep"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkgwuiA9F7", "year": 2019, "status": "rejected", "title": "Cramer-Wold AutoEncoder", "authors": ["Jacek Tabor", "Szymon Knop", "Przemysław Spurek", "Igor Podolak", "Marcin Mazur", "Stanisław Jastrzębski"], "authorids": ["jacek.tabor@uj.edu.pl", "szymon.knop@doctoral.uj.edu.pl", "przemyslaw.spurek@uj.edu.pl", "igor.podolak@uj.edu.pl", "marcin.mazur@uj.edu.pl", "staszek.jastrzebski@gmail.com"], "authors_source": "OpenReview API", "abstract": "Assessing distance betweeen the true and the sample distribution is a key component of many state of the art generative models, such as Wasserstein Autoencoder (WAE). Inspired by prior work on Sliced-Wasserstein Autoencoders (SWAE) and\nkernel smoothing we construct a new generative model – Cramer-Wold AutoEncoder (CWAE). CWAE cost function, based on introduced Cramer-Wold distance between samples, has a simple closed-form in the case of normal prior. As a consequence, while simplifying the optimization procedure (no need of sampling necessary to evaluate the distance function in the training loop), CWAE performance matches quantitatively and qualitatively that of WAE-MMD (WAE using maximum mean discrepancy based distance function) and often improves upon SWAE.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "Sker8vbtp7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper362/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes the Cramer-Wold autoencoder. The first contribution of the paper is to propose the Cramer-Wold distance between two distributions based on the Cramer-Wold Theorem. More specifically, in order to compute the Cramer-Wold distance, we first find the one dimensional projections of the distributions over random slices, and then compute the average L2 distances of the kernel density estimates of these projections over random slices. The second contribution of the paper is to develop a generative autoencoder which uses the Cramer-Wold distance to match the latent distribution of the data to the prior distribution.\n\nWhile I found the derivation of the Cramer-Wold distance interesting, the final form of this distance (Eq. 2), to me, looks very similar to the MMD with a particular kernel. My main question is that: what is the main advantage of the Cramer-Wold distance to an MMD with a proper kernel?\n\nThe paper points out that the main theoretical contribution is that in the case of the Gaussian distribution, the Cramer-Wold distance has a closed form. However, I believe this is also the case in the MMD, since if one of the distributions is Gaussian or analytically known, then E[k(x,x')] in the MMD can be analytically computed.\n\nThe paper further uses this closed form property of the Cramer-Wold distance to propose the Cramer-Wold autoencoder with Gaussian priors. My question here is that how is this method better than the standard VAE, where we also have an analytic form for the ELBO when the prior is Gaussian, an no sampling is required. Indeed, in VAEs, the prior does not have to be Gaussian, and as long as the density of the prior can be evaluated, we can efficiently optimize the ELBO without sampling the prior; which I don't think is the case for the Cramer-Wold autoencoder. I believe the main advantages of methods such as WAE is that they can impose priors that do not have exact analytic forms.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting derivation of a new distance, but should be compared with MMD", "review": "This paper proposes the Cramer-Wold autoencoder. The first contribution of the paper is to propose the Cramer-Wold distance between two distributions based on the Cramer-Wold Theorem. More specifically, in order to compute the Cramer-Wold distance, we first find the one dimensional projections of the distributions over random slices, and then compute the average L2 distances of the kernel density estimates of these projections over random slices. The second contribution of the paper is to develop a generative autoencoder which uses the Cramer-Wold distance to match the latent distribution of the data to the prior distribution.\n\nWhile I found the derivation of the Cramer-Wold distance interesting, the final form of this distance (Eq. 2), to me, looks very similar to the MMD with a particular kernel. My main question is that: what is the main advantage of the Cramer-Wold distance to an MMD with a proper kernel?\n\nThe paper points out that the main theoretical contribution is that in the case of the Gaussian distribution, the Cramer-Wold distance has a closed form. However, I believe this is also the case in the MMD, since if one of the distributions is Gaussian or analytically known, then E[k(x,x')] in the MMD can be analytically computed.\n\nThe paper further uses this closed form property of the Cramer-Wold distance to propose the Cramer-Wold autoencoder with Gaussian priors. My question here is that how is this method better than the standard VAE, where we also have an analytic form for the ELBO when the prior is Gaussian, an no sampling is required. Indeed, in VAEs, the prior does not have to be Gaussian, and as long as the density of the prior can be evaluated, we can efficiently optimize the ELBO without sampling the prior; which I don't think is the case for the Cramer-Wold autoencoder. I believe the main advantages of methods such as WAE is that they can impose priors that do not have exact analytic forms.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1542162252556}, {"id": "HJxRO37ka7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper362/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces a novel regularized auto-encoder architecture called the Cramer-Wold AutoEncoders (CWAE). It's objective (Eq. 7) consists of two terms: (i) a standard reconstruction term making sure the the encoder-decoder pair aligns nicely to accurately reconstruct all the training images and (ii) the regularizer, which roughly speaking requires the encoded training distribution to look similar to the standard normal (which is a prior used in the generative model being trained). The main novelty of the paper is in the form of this regularizer. The authors introduce what they call \"the Cramer-Wold distance\" (for definitions see Theorems 3.1 and 3.2) which is defined between two finite sets of D-dimensional points. The authors provide empirical studies showing that the proposed CWAE method achieves the same quality of samples (measured with FID scores) as the WAE-MMD model [1] previously reported in the literature, while running faster (by up to factor of 2 reduction in the training time, as the authors report). \n\nWhile on the AE model / architecture side I feel the contribution is very marginal, I still think that the improvement in the training speed is something useful. Otherwise it is a nicely written and polished piece of work. \n\nDetailed comments:\n(1) My main problem with this paper is that the novel objective proposed by the authors in Eq. 7 is equivalent to the objective of WAEs appearing in Eq. 4 of [1] (up to a heuristic of applying logarithm to the divergence measure, which is not justified but meant to \"improve the balance between two terms\", see footnote 2), where the authors use the newly introduced Cramer-Wold divergence as a choice of the penalty term in Eq. 4 of [1]. \n(2) When viewed in this way, CW-distance introduced in Eq (2) closely resembles the unbiased U-statistic estimate of the MMD used in WAE-MMD [1, Algorithm 2]. In other words: it may be the case that there is a choice of a reproducing kernel k such that Eq. 2 of this paper is an estimate of MMD_k between two distributions based on the i.i.d. samples X and Y. Note that if it is indeed the case, this corresponds to the V-statistic and thus biased: in U-statistic the diagonal terms (that is i = i' and j = j' in forst two terms of eq 2) would be omitted. If this all is indeed the case, it is not surprising that the numbers the authors get in the experiments are so similar to WAE-MMD, because CWAE would be exactly WAE-MMD with a specific choice of the kernel. \n(3) The authors make a big deal out of their proposed divergence measure not requiring samples from the prior as opposed to WAE-MMD. However, WAE-MMD does not necessarily need to sample from the prior when used with Gaussian prior and Gaussian RBF kernel, because in this case the prior-related parts of the MMD can be computed analytically.  In other words, if the computational advantage of CWAE compared to WAE-MMD comes from CWAE not sampling Pz, the computational overhead of WAE-MMD can be eliminated at least in the above-mentioned setting.\n(4) based on the name \"CW distance\" I would expect the authors to actually prove that it is indeed a distance (i.e. all the main axioms). \n(5) The authors override the CW distance: first in Theorem 3.1 they define it as a distance between two finite point clouds, and later in Theorem 3.2 they redefine it as a distance between a point cloud and the Gaussian distribution. \n(6) What is image(X) in Remark 4.1?\n\n[1] Tolstikhin et al., Wasserstein Auto-Encoders, 2017.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A variation on the Wasserstein Auto-Encoders proposing a specific choice of the divergence penalty.", "review": "The paper introduces a novel regularized auto-encoder architecture called the Cramer-Wold AutoEncoders (CWAE). It's objective (Eq. 7) consists of two terms: (i) a standard reconstruction term making sure the the encoder-decoder pair aligns nicely to accurately reconstruct all the training images and (ii) the regularizer, which roughly speaking requires the encoded training distribution to look similar to the standard normal (which is a prior used in the generative model being trained). The main novelty of the paper is in the form of this regularizer. The authors introduce what they call \"the Cramer-Wold distance\" (for definitions see Theorems 3.1 and 3.2) which is defined between two finite sets of D-dimensional points. The authors provide empirical studies showing that the proposed CWAE method achieves the same quality of samples (measured with FID scores) as the WAE-MMD model [1] previously reported in the literature, while running faster (by up to factor of 2 reduction in the training time, as the authors report). \n\nWhile on the AE model / architecture side I feel the contribution is very marginal, I still think that the improvement in the training speed is something useful. Otherwise it is a nicely written and polished piece of work. \n\nDetailed comments:\n(1) My main problem with this paper is that the novel objective proposed by the authors in Eq. 7 is equivalent to the objective of WAEs appearing in Eq. 4 of [1] (up to a heuristic of applying logarithm to the divergence measure, which is not justified but meant to \"improve the balance between two terms\", see footnote 2), where the authors use the newly introduced Cramer-Wold divergence as a choice of the penalty term in Eq. 4 of [1]. \n(2) When viewed in this way, CW-distance introduced in Eq (2) closely resembles the unbiased U-statistic estimate of the MMD used in WAE-MMD [1, Algorithm 2]. In other words: it may be the case that there is a choice of a reproducing kernel k such that Eq. 2 of this paper is an estimate of MMD_k between two distributions based on the i.i.d. samples X and Y. Note that if it is indeed the case, this corresponds to the V-statistic and thus biased: in U-statistic the diagonal terms (that is i = i' and j = j' in forst two terms of eq 2) would be omitted. If this all is indeed the case, it is not surprising that the numbers the authors get in the experiments are so similar to WAE-MMD, because CWAE would be exactly WAE-MMD with a specific choice of the kernel. \n(3) The authors make a big deal out of their proposed divergence measure not requiring samples from the prior as opposed to WAE-MMD. However, WAE-MMD does not necessarily need to sample from the prior when used with Gaussian prior and Gaussian RBF kernel, because in this case the prior-related parts of the MMD can be computed analytically.  In other words, if the computational advantage of CWAE compared to WAE-MMD comes from CWAE not sampling Pz, the computational overhead of WAE-MMD can be eliminated at least in the above-mentioned setting.\n(4) based on the name \"CW distance\" I would expect the authors to actually prove that it is indeed a distance (i.e. all the main axioms). \n(5) The authors override the CW distance: first in Theorem 3.1 they define it as a distance between two finite point clouds, and later in Theorem 3.2 they redefine it as a distance between a point cloud and the Gaussian distribution. \n(6) What is image(X) in Remark 4.1?\n\n[1] Tolstikhin et al., Wasserstein Auto-Encoders, 2017.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541516405727}, {"id": "B1lJvAAcnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper362/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a WAE variant based on a new statistical distance between the encoded data distribution and the latent prior distribution that can be computed in closed form without drawing samples from the prior (but only when it is Gaussian). The primary contribution is the new CW statistical distance, which is the l2 distance between projected distributions, integrated over all possible projections (although not calculated as so in practice).  \n  \nPlugging this distance into the WAE produces similar performance to existing WAE variants, but does not really advance the existing achievable performance.  Overall, I quite liked the paper and think it is well-written, but I believe the authors need to highlight at least one practical advance introduced by the CW distance (in which case I will raise my score). Some potential options include:\n\n1) Faster training times. It seems to me one potential advantage of the closed-form distance would be that the stochastic WAE-optimization can converge faster (due to lower-variance gradients).  However, the authors only presented per-batch processing times as opposed to overall training time for these models.   \n\n2) Stabler training. Perhaps sampling from the prior (as needed to compute statistical distances in the other WAE variants) introduces undesirable extra variance in the training procedure. The authors could run each WAE training process K times (with random initialization) to see if the closed-form distance enables more stable results.\n\n3) Usefulness of the CW distance outside of the autoencoder context.\nSince the novelty of this work lies in the introduction of the CW distance, I would like to see an independent evaluation of this distance as a  general statistical distance measure (independently of its use in CWAE). Can you use this distance as a multivariate-Gaussian goodness of fit measure for high-dimensional data drawn from both Gaussian and non-Gaussian distributions and show that it actually outperforms other standard statistical distances (e.g. in two-sample testing power)?\n\nWithout demonstrating any practical advance, this work becomes simply another one of the multitude of V/W-AE-variants that already exist.\n\nOther Comments:\n\n- While I agree that standard WAE-MMD and SWAE require some form of sampling to compute their respective statistical distance, a variant of WAE-MMD could be converted to a closed form statistical distance in the case of a Gaussian prior, by way of Stein's method or other existing goodness-of-fit measures designed specifically for Gaussians. See for example: \n\nChwialkowski et al: https://arxiv.org/pdf/1602.02964.pdf\n\nwhich like CW-distance is also a quadratic-time closed-form distance between samples and a target density.\n\nBesides having closed form in the case of a Gaussian prior (which other statistical distances could potentially also achieve), it would be nice to see some discussion of why the authors believe their CW-distance is conceptually superior to such alternatives. \n\n- Silverman's rule of thumb is only asymptotically optimal when the underlying data-generating distribution itself is Gaussian. Perhaps you can argue here that due to CLT: the projected data (for high-dimensional latent spaces) should look approximately Gaussian?\n\nAfter reading the revision: I have raised my score by 1 point and recommend acceptance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice idea & paper, but needs to highlight at least one practical advantage ", "review": "This paper proposes a WAE variant based on a new statistical distance between the encoded data distribution and the latent prior distribution that can be computed in closed form without drawing samples from the prior (but only when it is Gaussian). The primary contribution is the new CW statistical distance, which is the l2 distance between projected distributions, integrated over all possible projections (although not calculated as so in practice).  \n  \nPlugging this distance into the WAE produces similar performance to existing WAE variants, but does not really advance the existing achievable performance.  Overall, I quite liked the paper and think it is well-written, but I believe the authors need to highlight at least one practical advance introduced by the CW distance (in which case I will raise my score). Some potential options include:\n\n1) Faster training times. It seems to me one potential advantage of the closed-form distance would be that the stochastic WAE-optimization can converge faster (due to lower-variance gradients).  However, the authors only presented per-batch processing times as opposed to overall training time for these models.   \n\n2) Stabler training. Perhaps sampling from the prior (as needed to compute statistical distances in the other WAE variants) introduces undesirable extra variance in the training procedure. The authors could run each WAE training process K times (with random initialization) to see if the closed-form distance enables more stable results.\n\n3) Usefulness of the CW distance outside of the autoencoder context.\nSince the novelty of this work lies in the introduction of the CW distance, I would like to see an independent evaluation of this distance as a  general statistical distance measure (independently of its use in CWAE). Can you use this distance as a multivariate-Gaussian goodness of fit measure for high-dimensional data drawn from both Gaussian and non-Gaussian distributions and show that it actually outperforms other standard statistical distances (e.g. in two-sample testing power)?\n\nWithout demonstrating any practical advance, this work becomes simply another one of the multitude of V/W-AE-variants that already exist.\n\nOther Comments:\n\n- While I agree that standard WAE-MMD and SWAE require some form of sampling to compute their respective statistical distance, a variant of WAE-MMD could be converted to a closed form statistical distance in the case of a Gaussian prior, by way of Stein's method or other existing goodness-of-fit measures designed specifically for Gaussians. See for example: \n\nChwialkowski et al: https://arxiv.org/pdf/1602.02964.pdf\n\nwhich like CW-distance is also a quadratic-time closed-form distance between samples and a target density.\n\nBesides having closed form in the case of a Gaussian prior (which other statistical distances could potentially also achieve), it would be nice to see some discussion of why the authors believe their CW-distance is conceptually superior to such alternatives. \n\n- Silverman's rule of thumb is only asymptotically optimal when the underlying data-generating distribution itself is Gaussian. Perhaps you can argue here that due to CLT: the projected data (for high-dimensional latent spaces) should look approximately Gaussian?\n\nAfter reading the revision: I have raised my score by 1 point and recommend acceptance.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541234262765}], "openreview_url": "https://openreview.net/forum?id=rkgwuiA9F7", "arxiv_id": "1805.09235", "paper_pdf": "papers/rkgwuiA9F7.pdf", "paper_pdf_sha256": "36fda4f9111166f7bbbd9d4da07d25478bd74e2a4acafdab9c995fde268789a1", "paper_pdf_bytes": 7236018, "paper_pdf_source": "openreview", "code_url": "https://github.com/gmum/cwae", "code_repository": "gmum/cwae", "code_commit": "50592903c321de25f339f3b00cbd2143741e5037", "code_archive": "repos/rkgwuiA9F7.zip", "code_archive_sha256": "4528136d0c84d4d22b8a19b7b3b7f623834c56e6e8d1095ca8147eee6840b821", "code_archive_bytes": 185385, "code_file_count": 20, "code_extensions": {".py": 19, ".ipynb": 1}, "github_disk_usage_kb": 180, "github_languages": {"Jupyter Notebook": 232896, "Python": 37925}, "github_archived": false, "github_pushed_at": "2020-09-17T16:31:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cramer-wold-autoencoder"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryCM8zWRb", "year": 2018, "status": "rejected", "title": "Recurrent Neural Networks with Top-k Gains for Session-based Recommendations", "authors": ["Balázs Hidasi", "Alexandros Karatzoglou"], "authorids": ["hidasib@gmail.com", "alexk@tid.es"], "authors_source": "OpenReview API", "abstract": "RNNs have been shown to be excellent models for sequential data and in particular for session-based user behavior. The use of RNNs provides impressive performance benefits over classical methods in session-based recommendations. In this work we introduce a novel ranking loss function tailored for RNNs in recommendation settings. The better performance of such loss over alternatives, along with further tricks and improvements described in this work, allow to achieve an overall improvement of up to 35% in terms of MRR and Recall@20 over previous session-based RNN solutions and up to 51% over classical collaborative filtering approaches. Unlike data augmentation-based improvements, our method does not increase training times significantly.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1QqAe3lG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper843/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper discussed the issues for optimizing the loss functions in GRU4Rec and proposed tricks for optimize the loss functions, and also proposed enhanced version of the loss functions for GRU4Rec. Good performance improvements have been reported for the several datasets to show the effectiveness of the proposed methods.\n\nThe good point of this work is to show that the loss function is important to train a better classifier for the session-based recommendation. This work is of value to the session-based recommendations.\n\nSome minor points:\nI think it may be better if the authors could put the results of RSC15 from Tan (2016) and Chatzis ect. (2017) into table 2 as well.\nAs these work has already been published and should be compared with and reported in the formal table. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper discussed the issues for optimizing the loss functions in GRU4Rec and proposed tricks for optimize the loss functions, and also proposed enhanced version of the loss functions for GRU4Rec. Good performance improvements have been reported for the several datasets to show the effectiveness of the proposed methods.", "rating": "6: Marginally above acceptance threshold", "review": "This paper discussed the issues for optimizing the loss functions in GRU4Rec and proposed tricks for optimize the loss functions, and also proposed enhanced version of the loss functions for GRU4Rec. Good performance improvements have been reported for the several datasets to show the effectiveness of the proposed methods.\n\nThe good point of this work is to show that the loss function is important to train a better classifier for the session-based recommendation. This work is of value to the session-based recommendations.\n\nSome minor points:\nI think it may be better if the authors could put the results of RSC15 from Tan (2016) and Chatzis ect. (2017) into table 2 as well.\nAs these work has already been published and should be compared with and reported in the formal table. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511947914929}, {"id": "ryqETl9gG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper843/AnonReviewer2"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This is an interesting paper that analyzes existing loss functions for session-based recommendations. Based on the result of these analysis the authors propose two novel losses functions which add a weighting to existing ranking-based loss functions. These novelties are meant to improve issues related to vanishing gradients of current loss functions. The empirical results on two large-scale datasets are pretty impressive. \n\nI found this paper to be well-written and easy to read.  It also provides a nice introduction to some of the recent literature on RNNs for session-based recommendations. \n\nIn terms of impact, while it studies a fairly applied (and narrow) question, it seems like it would be of interest to researchers and practitioners in recommender systems.\n\n\nI have a few comments and questions: \n\n- The results in Figure 3 show that both a good loss function and sampling strategy are required to perform well. This is interesting in the sense that doing the \"right thing\" according to the model (optimizing using all samples) isn't optimal. This is a very empirical observation and it would be insightful to better understand exactly the objective that is being optimized.\n\n- While BPR-max seems to be the strongest performer (Table 2), cross-entropy (XE) with additional samples is close. This further outlines the importance of the sampling method over the exact form of the loss function. \n\n- In ranking-max losses, it seems like \"outliers\" could have a bigger impact. I don't know how useful it is to think about (and it is a bit unclear what an \"outlier\" means in this implicit feedback setting).\n\n\nMinor comments: \n\n- Around Eq. 4 it may be worth being more explicit about the meaning of i and j. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Insightful analysis of existing loss functions, new losses are proposed and provide strong empirical results. Good paper.", "rating": "8: Top 50% of accepted papers, clear accept", "review": "This is an interesting paper that analyzes existing loss functions for session-based recommendations. Based on the result of these analysis the authors propose two novel losses functions which add a weighting to existing ranking-based loss functions. These novelties are meant to improve issues related to vanishing gradients of current loss functions. The empirical results on two large-scale datasets are pretty impressive. \n\nI found this paper to be well-written and easy to read.  It also provides a nice introduction to some of the recent literature on RNNs for session-based recommendations. \n\nIn terms of impact, while it studies a fairly applied (and narrow) question, it seems like it would be of interest to researchers and practitioners in recommender systems.\n\n\nI have a few comments and questions: \n\n- The results in Figure 3 show that both a good loss function and sampling strategy are required to perform well. This is interesting in the sense that doing the \"right thing\" according to the model (optimizing using all samples) isn't optimal. This is a very empirical observation and it would be insightful to better understand exactly the objective that is being optimized.\n\n- While BPR-max seems to be the strongest performer (Table 2), cross-entropy (XE) with additional samples is close. This further outlines the importance of the sampling method over the exact form of the loss function. \n\n- In ranking-max losses, it seems like \"outliers\" could have a bigger impact. I don't know how useful it is to think about (and it is a bit unclear what an \"outlier\" means in this implicit feedback setting).\n\n\nMinor comments: \n\n- Around Eq. 4 it may be worth being more explicit about the meaning of i and j. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511816498131}, {"id": "S1d1eXqlM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper843/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a few modifications on top of some earlier work (GRU4Rec, Hidasi et al. 2016) for session-based recommendation using RNN. The first one is to include additional negative samples based on popularity raised to some power between 0 and 1. The second one is to mitigate the vanishing gradient problem for pairwise ranking loss, especially with the increased number of negative samples from the first modification. The basic idea is to weight all the negative examples by their “relevance”, since for the irrelevant negatives the gradients are vanishingly small. Experimentally these modifications prove to be effective compared with the original GRU4Rec paper. \n\nThe writing could have been more clear, especially in terms of notations and definitions. I found myself sometimes having to infer the missing bits. For example, in Eq (4) and (5), and many that follow, the index i and j are not defined (I can infer it from the later part), as well as N_s (which I take it as the number of negative examples). This is just one example, but I hope the authors could carefully check the paper and make sure all the notations/terminologies are properly defined or referred with a citation when first introduced (e.g., pointwise, pairwise, and listwise loss functions). I consider myself very familiar with the RecSys literature, and yet sometimes I cannot follow the paper very well, not to mention the general ICLR audience. \n\nRegarding the two main modifications, I found the negative sampling rather trivial (and I am surprised in Hidasi et al. (2016) the negatives are only from the same batch, which seems a huge computational compromise) with many existing work on related topic: Steck (Item popularity and recommendation accuracy, 2011) used the same “popularity to the power between 0 and 1” strategy (they weighted the positive by the inverse popularity to the power). More closely, the negative sampling distribution in word2vec is in fact a unigram raised to the power of 0.75, which is the same as the proposed strategy here. As for the gradient vanishing problem for pairwise ranking loss, it has been previously observed in Rendle & Freudenthaler (Improving Pairwise Learning for Item Recommendation from Implicit Feedback, 2014) for BPR and they proposed an adaptive negative sampling strategy (trying to sample more relevant negatives while still keeping the computational cost low), which is closely related to the ranking-max loss function proposed in this paper. Overall, I don’t think this paper adds much on top of the previous work, and I think a more RecSys-oriented venue might benefit more from the insights presented in this paper.   \n\nI also have some high-level comments regarding using RNN for session-based recommendation (this was also my initial reaction after reading Hidasi et al. 2016). As mentioned in this paper, when applying RNN on RecSys datasets with longer time-span (which means there can be more temporal dynamics in users’ preference and item popularity), the results are not striking (e.g., Wu et al. 2017) with the proposed methods barely outperforming standard matrix factorization methods. It is puzzling that how RNN can work better for session-based case where a user’s preference can hardly change within such a short period of time. I wonder how a simple matrix factorization approach would work for session-based recommendation (which is an important baseline that is missing): regarding the claim that MF is not suited for session-based because of the absence of the concept of a user, each session can simply be considered as a pseudo-user and approaches like asymmetric matrix factorization (Paterek 2007, Improving regularized singular value decomposition for collaborative filtering) can even eliminate the need for learning user factors. ItemKNN is a pretty weak baseline and I wonder if a scalable version of the SLIM (Ning & Karypis 2011, SLIM: Sparse Linear Methods for Top-N Recommender Systems) would give better results. Finally, my general experience with BPR-type of pairwise ranking loss is that it is good at optimizing AUC, but not very well-suited for head-heavy metrics (MRR, Recall, etc.) I wonder how the propose loss would perform comparing with more competitive baselines. \n\nRegarding the page limit, given currently the paper is quite long (12 pages excluding references), I suggest the authors cutting down some space. For example, the part about fixing the cross entropy is not very relevant and can totally be put in the appendix. \n\nMinor comment:\n\n1. Section 3.3.1, “Part of the reasons lies in the rare occurrence…”, should r_j >> r_i be the other way around?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some interesting insights, but no substantial contributions and probably not well-suited for ICLR", "rating": "4: Ok but not good enough - rejection", "review": "This paper presents a few modifications on top of some earlier work (GRU4Rec, Hidasi et al. 2016) for session-based recommendation using RNN. The first one is to include additional negative samples based on popularity raised to some power between 0 and 1. The second one is to mitigate the vanishing gradient problem for pairwise ranking loss, especially with the increased number of negative samples from the first modification. The basic idea is to weight all the negative examples by their “relevance”, since for the irrelevant negatives the gradients are vanishingly small. Experimentally these modifications prove to be effective compared with the original GRU4Rec paper. \n\nThe writing could have been more clear, especially in terms of notations and definitions. I found myself sometimes having to infer the missing bits. For example, in Eq (4) and (5), and many that follow, the index i and j are not defined (I can infer it from the later part), as well as N_s (which I take it as the number of negative examples). This is just one example, but I hope the authors could carefully check the paper and make sure all the notations/terminologies are properly defined or referred with a citation when first introduced (e.g., pointwise, pairwise, and listwise loss functions). I consider myself very familiar with the RecSys literature, and yet sometimes I cannot follow the paper very well, not to mention the general ICLR audience. \n\nRegarding the two main modifications, I found the negative sampling rather trivial (and I am surprised in Hidasi et al. (2016) the negatives are only from the same batch, which seems a huge computational compromise) with many existing work on related topic: Steck (Item popularity and recommendation accuracy, 2011) used the same “popularity to the power between 0 and 1” strategy (they weighted the positive by the inverse popularity to the power). More closely, the negative sampling distribution in word2vec is in fact a unigram raised to the power of 0.75, which is the same as the proposed strategy here. As for the gradient vanishing problem for pairwise ranking loss, it has been previously observed in Rendle & Freudenthaler (Improving Pairwise Learning for Item Recommendation from Implicit Feedback, 2014) for BPR and they proposed an adaptive negative sampling strategy (trying to sample more relevant negatives while still keeping the computational cost low), which is closely related to the ranking-max loss function proposed in this paper. Overall, I don’t think this paper adds much on top of the previous work, and I think a more RecSys-oriented venue might benefit more from the insights presented in this paper.   \n\nI also have some high-level comments regarding using RNN for session-based recommendation (this was also my initial reaction after reading Hidasi et al. 2016). As mentioned in this paper, when applying RNN on RecSys datasets with longer time-span (which means there can be more temporal dynamics in users’ preference and item popularity), the results are not striking (e.g., Wu et al. 2017) with the proposed methods barely outperforming standard matrix factorization methods. It is puzzling that how RNN can work better for session-based case where a user’s preference can hardly change within such a short period of time. I wonder how a simple matrix factorization approach would work for session-based recommendation (which is an important baseline that is missing): regarding the claim that MF is not suited for session-based because of the absence of the concept of a user, each session can simply be considered as a pseudo-user and approaches like asymmetric matrix factorization (Paterek 2007, Improving regularized singular value decomposition for collaborative filtering) can even eliminate the need for learning user factors. ItemKNN is a pretty weak baseline and I wonder if a scalable version of the SLIM (Ning & Karypis 2011, SLIM: Sparse Linear Methods for Top-N Recommender Systems) would give better results. Finally, my general experience with BPR-type of pairwise ranking loss is that it is good at optimizing AUC, but not very well-suited for head-heavy metrics (MRR, Recall, etc.) I wonder how the propose loss would perform comparing with more competitive baselines. \n\nRegarding the page limit, given currently the paper is quite long (12 pages excluding references), I suggest the authors cutting down some space. For example, the part about fixing the cross entropy is not very relevant and can totally be put in the appendix. \n\nMinor comment:\n\n1. Section 3.3.1, “Part of the reasons lies in the rare occurrence…”, should r_j >> r_i be the other way around?\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511825376401}], "openreview_url": "https://openreview.net/forum?id=ryCM8zWRb", "arxiv_id": "1706.03847", "paper_pdf": "papers/ryCM8zWRb.pdf", "paper_pdf_sha256": "1978893be01f8b5332a1b09b0cea54253d33ff20199d669afe4d44c7f7a6440d", "paper_pdf_bytes": 499988, "paper_pdf_source": "openreview", "code_url": "https://github.com/hidasib/GRU4Rec", "code_repository": "hidasib/GRU4Rec", "code_commit": "a4ed5fbdba35bcc18bf1d4a9b76692ef462bb284", "code_archive": "repos/ryCM8zWRb.zip", "code_archive_sha256": "b2ad2479ffb6d320d3f2f3b929b6bd084d8943a36be9dee83dd06635a8c4a230", "code_archive_bytes": 185985, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 312, "github_languages": {"Python": 132123}, "github_archived": false, "github_pushed_at": "2023-08-24T15:57:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recurrent-neural-networks-with-top-k-gains"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5y8R8XSHkG", "year": 2026, "status": "rejected", "title": "Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation", "authors": ["Alexander Shabalin", "Viacheslav Meshchaninov", "Dmitry Vetrov"], "authorids": ["~Alexander_Shabalin1", "~Viacheslav_Meshchaninov1", "~Dmitry_Vetrov1"], "authors_source": "OpenReview API", "abstract": "Diffusion models have achieved state-of-the-art performance in generating images, audio, and video, but their adaptation to text remains challenging due to its discrete nature. Prior approaches either apply Gaussian diffusion in continuous latent spaces, which inherits semantic structure but struggles with token decoding, or operate in categorical simplex space, which respect discreteness but disregard semantic relation between tokens. In this paper, we propose Smoothing Diffusion on Token Embeddings (Smoothie), a novel diffusion method that combines the strengths of both approaches by progressively smoothing token embeddings based on semantic similarity. This technique enables gradual information removal while maintaining a natural decoding process. Experimental results on several sequence-to-sequence generation tasks demonstrate that Smoothie outperforms existing diffusion-based models in generation quality. Furthermore, ablation studies show that our proposed diffusion space yields better performance than both the standard embedding space and the categorical simplex.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jIpnhAIfDb", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19003/Reviewer_V1zB"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "This paper introduces SMOOTHIE, a diffusion model framework designed for text generation. The authors identify a key challenge in adapting diffusion models to discrete data like text: existing methods either operate in a continuous latent space (e.g., Gaussian diffusion on embeddings), which struggles with accurate token decoding, or in a discrete/categorical space, which ignores the semantic relationships between tokens. SMOOTHIE perturbs distance-based representations of tokens, dissolving semantic structure over time. The authors claim this technique is superior to both standard embedding space diffusion and categorical diffusion. They provide empirical evidence on several sequence-to-sequence generation tasks.", "review_text": "This paper introduces SMOOTHIE, a diffusion model framework designed for text generation. The authors identify a key challenge in adapting diffusion models to discrete data like text: existing methods either operate in a continuous latent space (e.g., Gaussian diffusion on embeddings), which struggles with accurate token decoding, or in a discrete/categorical space, which ignores the semantic relationships between tokens. SMOOTHIE perturbs distance-based representations of tokens, dissolving semantic structure over time. The authors claim this technique is superior to both standard embedding space diffusion and categorical diffusion. They provide empirical evidence on several sequence-to-sequence generation tasks.", "strengths": "- The proposed diffusion space, which perturbs representations based on semantic similarity, is a contribution to the field.\n- The authors provide empirical validation across multiple text generation tasks. The reported results suggest that the proposed method may offer a performance improvement over other diffusion-based baselines, and the inclusion of ablation studies helps to substantiate the specific design choices made in the SMOOTHIE framework.", "weaknesses": "- The proposed method's reliance on a pre-trained word embedding model (in this case, BERT) may limit its scalability and applicability. This dependency raises questions about the framework's potential to scale effectively with larger models or different architectures, as it is tied to the properties and constraints of the initial embedding space.\n- The experimental evaluation is missing a common and important conditional text generation task: machine translation. Including results from machine translation would provide a more comprehensive assessment of the method's capabilities and generalizability.\n- The paper lacks an analysis of the method's sensitivity to the choice of the pre-trained word embedding. It would be beneficial to investigate whether the approach is viable with other types of embeddings, such as those from GPT-based models, to better understand the robustness and flexibility of the proposed framework.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SMOOTHIE, a diffusion model framework designed for text generation. The authors identify a key challenge in adapting diffusion models to discrete data like text: existing methods either operate in a continuous latent space (e.g., Gaussian diffusion on embeddings), which struggles with accurate token decoding, or in a discrete/categorical space, which ignores the semantic relationships between tokens. SMOOTHIE perturbs distance-based representations of tokens, dissolving semantic structure over time. The authors claim this technique is superior to both standard embedding space diffusion and categorical diffusion. They provide empirical evidence on several sequence-to-sequence generation tasks.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "- The proposed diffusion space, which perturbs representations based on semantic similarity, is a contribution to the field.\n- The authors provide empirical validation across multiple text generation tasks. The reported results suggest that the proposed method may offer a performance improvement over other diffusion-based baselines, and the inclusion of ablation studies helps to substantiate the specific design choices made in the SMOOTHIE framework.", "weaknesses": "- The proposed method's reliance on a pre-trained word embedding model (in this case, BERT) may limit its scalability and applicability. This dependency raises questions about the framework's potential to scale effectively with larger models or different architectures, as it is tied to the properties and constraints of the initial embedding space.\n- The experimental evaluation is missing a common and important conditional text generation task: machine translation. Including results from machine translation would provide a more comprehensive assessment of the method's capabilities and generalizability.\n- The paper lacks an analysis of the method's sensitivity to the choice of the pre-trained word embedding. It would be beneficial to investigate whether the approach is viable with other types of embeddings, such as those from GPT-based models, to better understand the robustness and flexibility of the proposed framework.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762098197001}, {"id": "RFgAQtkIbS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19003/Reviewer_Uud8"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "SMOOTHIE proposes diffusion over distance-based token-embedding logits, smoothing semantics while preserving discreteness, outperforming prior text diffusion on seq2seq tasks; analyses highlight noise, steps, and self-conditioning.", "review_text": "SMOOTHIE proposes diffusion over distance-based token-embedding logits, smoothing semantics while preserving discreteness, outperforming prior text diffusion on seq2seq tasks; analyses highlight noise, steps, and self-conditioning.", "strengths": "1.\tUnifies prior lines: maps each token to a vector of negative squared distances to all vocab embeddings, then diffuses and feeds softmax(D_t) to the model; enables natural argmax decoding while preserving semantics and discreteness. Clear training/sampling pseudocode. \n2.\tThe distance-based latent generalizes simplex diffusion (simplex emerges under a trivial metric), giving a clean conceptual frame. \n3.\tPractical guidance on schedules/self-conditioning; moderate steps (~100–200) are sufficient, with analysis of step count and reverse-noise. \n4.\tConsistent gains over diffusion baselines on multiple seq2seq tasks.", "weaknesses": "1.\tFixed pre-trained embeddings (E) cap expressivity; authors acknowledge end-to-end training would likely help but leave it to future work. \n2.\tFixed sequence length forces substantial padding; variable length is emulated by truncating after EOS, which is inefficient; prior early-truncation is ad hoc. \n3.\tEvery step computes softmax over the full vocabulary V (and final argmax), which scales poorly for large V and long m; no top-k/approximation is provided. \n4.\tRelies on the Euclidean semantic space hypothesis; authors note other domains may need different distances—raises concerns for polysemy/anisotropy. \n5.\tLittle on throughput/memory vs. competing diffusion methods under equal steps; limited evaluation breadth (small/medium datasets, few human evals).", "questions": "1. Please fine-tune embeddings or learn a task-adaptive metric (e.g., Mahalanobis) on at least one dataset; report lifts vs. fixed E. \n2. Efficiency. Provide tokens/sec, GPU memory, and wall-clock vs. SSD-LM/TESS/embedding-diffusion at matched steps; include step–quality curves. \n3. Try top-k candidate sets (ANN/FAISS) or hierarchical/adaptive softmax; quantify quality vs. speed trade-offs for long sequences and large vocabularies. \n4. Beyond EOS truncation, test a general dynamic-length denoising strategy (e.g., entropy/energy-based early stop) and compare to SeqDiffuSeq’s early truncation. \n5. Add longer-form generation, dialogue, or factual QA, and report mean±σ over multiple seeds; analyze sensitivity to δ ̃and steps across tasks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "SMOOTHIE proposes diffusion over distance-based token-embedding logits, smoothing semantics while preserving discreteness, outperforming prior text diffusion on seq2seq tasks; analyses highlight noise, steps, and self-conditioning.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.\tUnifies prior lines: maps each token to a vector of negative squared distances to all vocab embeddings, then diffuses and feeds softmax(D_t) to the model; enables natural argmax decoding while preserving semantics and discreteness. Clear training/sampling pseudocode. \n2.\tThe distance-based latent generalizes simplex diffusion (simplex emerges under a trivial metric), giving a clean conceptual frame. \n3.\tPractical guidance on schedules/self-conditioning; moderate steps (~100–200) are sufficient, with analysis of step count and reverse-noise. \n4.\tConsistent gains over diffusion baselines on multiple seq2seq tasks.", "weaknesses": "1.\tFixed pre-trained embeddings (E) cap expressivity; authors acknowledge end-to-end training would likely help but leave it to future work. \n2.\tFixed sequence length forces substantial padding; variable length is emulated by truncating after EOS, which is inefficient; prior early-truncation is ad hoc. \n3.\tEvery step computes softmax over the full vocabulary V (and final argmax), which scales poorly for large V and long m; no top-k/approximation is provided. \n4.\tRelies on the Euclidean semantic space hypothesis; authors note other domains may need different distances—raises concerns for polysemy/anisotropy. \n5.\tLittle on throughput/memory vs. competing diffusion methods under equal steps; limited evaluation breadth (small/medium datasets, few human evals).", "questions": "1. Please fine-tune embeddings or learn a task-adaptive metric (e.g., Mahalanobis) on at least one dataset; report lifts vs. fixed E. \n2. Efficiency. Provide tokens/sec, GPU memory, and wall-clock vs. SSD-LM/TESS/embedding-diffusion at matched steps; include step–quality curves. \n3. Try top-k candidate sets (ANN/FAISS) or hierarchical/adaptive softmax; quantify quality vs. speed trade-offs for long sequences and large vocabularies. \n4. Beyond EOS truncation, test a general dynamic-length denoising strategy (e.g., entropy/energy-based early stop) and compare to SeqDiffuSeq’s early truncation. \n5. Add longer-form generation, dialogue, or factual QA, and report mean±σ over multiple seeds; analyze sensitivity to δ ̃and steps across tasks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761966916651}, {"id": "WLmHD0pm40", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19003/Reviewer_QSj6"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "This paper introduces SMOOTHIE (Smoothing Diffusion on Token Embeddings), a novel diffusion model framework for text generation that aims to bridge the gap between continuous (Gaussian) and discrete (Simplex/Categorical) text diffusion methods. The core innovation lies in defining a new latent space where each token is represented by a vector of negative squared Euclidean distances between its embedding and the embeddings of all vocabulary tokens.\nThe authors demonstrate that SMOOTHIE consistently outperforms prior diffusion-based approaches across several sequence-to-sequence tasks, achieving generation quality comparable to strong autoregressive baselines.", "review_text": "This paper introduces SMOOTHIE (Smoothing Diffusion on Token Embeddings), a novel diffusion model framework for text generation that aims to bridge the gap between continuous (Gaussian) and discrete (Simplex/Categorical) text diffusion methods. The core innovation lies in defining a new latent space where each token is represented by a vector of negative squared Euclidean distances between its embedding and the embeddings of all vocabulary tokens.\nThe authors demonstrate that SMOOTHIE consistently outperforms prior diffusion-based approaches across several sequence-to-sequence tasks, achieving generation quality comparable to strong autoregressive baselines.", "strengths": "The core contribution of defining the diffusion space using semantic distances (Euclidean proximity) in the embedding space is highly intuitive and well-justified. It elegantly addresses the major trade-off in existing work: retaining semantic structure (like Gaussian diffusion) while enabling natural decoding from discrete representations (like Simplex diffusion).", "weaknesses": "SMOOTHIE (like most text diffusion models) runs over fixed-length sequences. In practice, they set a dataset-specific max length and pad shorter sequences with a special padding token that the model learns to predict. The generation process is bounded by the preset max. It can emit different effective lengths up to a cap, but it doesn’t truly sample variable length the way an autoregressive model does.", "questions": "In Figure 1 (a), why is the color of the arbitrary i-th embedding always the same? What is the meaning of the structured shape (a flag-shaped pattern with an oval in the bottom) that the dots form?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SMOOTHIE (Smoothing Diffusion on Token Embeddings), a novel diffusion model framework for text generation that aims to bridge the gap between continuous (Gaussian) and discrete (Simplex/Categorical) text diffusion methods. The core innovation lies in defining a new latent space where each token is represented by a vector of negative squared Euclidean distances between its embedding and the embeddings of all vocabulary tokens.\nThe authors demonstrate that SMOOTHIE consistently outperforms prior diffusion-based approaches across several sequence-to-sequence tasks, achieving generation quality comparable to strong autoregressive baselines.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "The core contribution of defining the diffusion space using semantic distances (Euclidean proximity) in the embedding space is highly intuitive and well-justified. It elegantly addresses the major trade-off in existing work: retaining semantic structure (like Gaussian diffusion) while enabling natural decoding from discrete representations (like Simplex diffusion).", "weaknesses": "SMOOTHIE (like most text diffusion models) runs over fixed-length sequences. In practice, they set a dataset-specific max length and pad shorter sequences with a special padding token that the model learns to predict. The generation process is bounded by the preset max. It can emit different effective lengths up to a cap, but it doesn’t truly sample variable length the way an autoregressive model does.", "questions": "In Figure 1 (a), why is the color of the arbitrary i-th embedding always the same? What is the meaning of the structured shape (a flag-shaped pattern with an oval in the bottom) that the dots form?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761883918613}, {"id": "rrxvoA9gst", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19003/Reviewer_gFQF"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 5, "summary": "The paper introduces **SMOOTHIE**, a diffusion model that operates on token embeddings by constructing a *semantic distance tensor* for each token. Instead of performing diffusion in the discrete simplex (as in D3PM or SSD-LM) or the embedding space (as in Diffusion-LM), SMOOTHIE models the evolution of the *pairwise distances* between each token and all vocabulary embeddings. This design allows the diffusion process to preserve both discrete structure and semantic smoothness: at each timestep, Gaussian noise is added to perturb these distances, generating soft distributions that reflect evolving semantic relationships.\n\nThe authors prove that simplex diffusion (D3PM-style) is a special case of SMOOTHIE when using a trivial distance metric, thus theoretically generalizing prior discrete diffusion frameworks. Experiments on summarization (XSum), paraphrasing (QQP), detoxification (ParaDetox), story generation (ROCStories), and question answering (Quasar-T) show consistent improvements over both discrete and continuous diffusion baselines. While the paper includes a basic step-count and self-conditioning analysis, more systematic ablations (e.g., noise schedule, δ̃ magnitude) are needed to isolate where the gains truly come from.", "review_text": "The paper introduces **SMOOTHIE**, a diffusion model that operates on token embeddings by constructing a *semantic distance tensor* for each token. Instead of performing diffusion in the discrete simplex (as in D3PM or SSD-LM) or the embedding space (as in Diffusion-LM), SMOOTHIE models the evolution of the *pairwise distances* between each token and all vocabulary embeddings. This design allows the diffusion process to preserve both discrete structure and semantic smoothness: at each timestep, Gaussian noise is added to perturb these distances, generating soft distributions that reflect evolving semantic relationships.\n\nThe authors prove that simplex diffusion (D3PM-style) is a special case of SMOOTHIE when using a trivial distance metric, thus theoretically generalizing prior discrete diffusion frameworks. Experiments on summarization (XSum), paraphrasing (QQP), detoxification (ParaDetox), story generation (ROCStories), and question answering (Quasar-T) show consistent improvements over both discrete and continuous diffusion baselines. While the paper includes a basic step-count and self-conditioning analysis, more systematic ablations (e.g., noise schedule, δ̃ magnitude) are needed to isolate where the gains truly come from.", "strengths": "1. **Unified formulation:** The paper provides a clear mathematical unification of discrete simplex diffusion and continuous embedding diffusion via a distance-based representation. Theorem 4.1 elegantly connects distance regression with embedding regression.\n2. **Semantic-aware diffusion:** By perturbing the distance tensor with Gaussian noise, the model embeds semantic structure directly into the diffusion process, enabling smooth transitions between semantically related tokens.\n3. **Comprehensive evaluation:** Experiments cover diverse tasks (XSum, QQP, ParaDetox, ROCStories, Quasar-T) and compare fairly with both discrete (TESS, SSD-LM, D3PM) and continuous (DiffuSeq, Diffusion-LM) baselines.\n4. **Fair comparison setup:** Model sizes (~100M), datasets, and decoding strategies are kept consistent; pretrained advantages of baselines (e.g., TESS) are removed.\n5. **Strong performance:** SMOOTHIE outperforms all diffusion baselines and achieves results comparable to autoregressive models like FLAN-T5.", "weaknesses": "1. **Computation cost not analyzed:** Although vectorization and mean-embedding compression are mentioned, there is no report on runtime, GPU memory, or FLOPs. Given that every token interacts with all vocabulary embeddings, the training cost may be substantial.\n2. **Limited ablations on general enhancements:** While the authors tested self-conditioning and found minimal gains (and thus did not adopt it), other general mechanisms such as the tanh-style noise schedule or δ̃ magnitude lack comprehensive ablations. More systematic sensitivity studies would clarify whether improvements stem primarily from the distance-based diffusion rather than secondary hyperparameters.\n3. **Embedding dependency untested:** The model fixes the BERT embedding matrix during training but does not evaluate alternative embeddings (e.g., random, GloVe, or fine-tuned). It remains unclear how sensitive performance is to the embedding quality or domain.\n4. **No convergence or efficiency analysis:** Beyond theoretical equivalence, convergence behavior and stability relative to token-level diffusion are unreported; adding epoch-wise loss curves would clarify training efficiency.", "questions": "1. **Computation cost:** Can you report training time, GPU memory, or FLOPs compared to TESS or DiffuSeq? How is the full distance tensor computation optimized, and how does it scale with vocabulary size?\n2. **Embedding dependence:** Have you evaluated SMOOTHIE using other embeddings (e.g., random or domain-specific)? How robust is the model to embedding quality and distribution shifts?\n3. **Noise schedule and δ̃ ablations:** Why was the tanh-style schedule chosen? Have you tested alternative schedules or δ̃ values to confirm robustness across datasets?\n4. **Self-conditioning and fairness:** Since self-conditioning yields limited improvement and was excluded, can you confirm that no other general enhancements influenced the main gains? Could similar gains be achieved by applying self-conditioning to baselines?\n5. **Convergence and stability:** Please include training curves or epoch-wise loss comparisons with baseline diffusion models to demonstrate convergence efficiency and numerical stability.\n6. **Scalability:** How does SMOOTHIE perform with larger vocabularies (e.g., >50K tokens)? Are there feasible strategies (e.g., top-k distance pruning or clustering) to reduce complexity while maintaining accuracy?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces **SMOOTHIE**, a diffusion model that operates on token embeddings by constructing a *semantic distance tensor* for each token. Instead of performing diffusion in the discrete simplex (as in D3PM or SSD-LM) or the embedding space (as in Diffusion-LM), SMOOTHIE models the evolution of the *pairwise distances* between each token and all vocabulary embeddings. This design allows the diffusion process to preserve both discrete structure and semantic smoothness: at each timestep, Gaussian noise is added to perturb these distances, generating soft distributions that reflect evolving semantic relationships.\n\nThe authors prove that simplex diffusion (D3PM-style) is a special case of SMOOTHIE when using a trivial distance metric, thus theoretically generalizing prior discrete diffusion frameworks. Experiments on summarization (XSum), paraphrasing (QQP), detoxification (ParaDetox), story generation (ROCStories), and question answering (Quasar-T) show consistent improvements over both discrete and continuous diffusion baselines. While the paper includes a basic step-count and self-conditioning analysis, more systematic ablations (e.g., noise schedule, δ̃ magnitude) are needed to isolate where the gains truly come from.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "1. **Unified formulation:** The paper provides a clear mathematical unification of discrete simplex diffusion and continuous embedding diffusion via a distance-based representation. Theorem 4.1 elegantly connects distance regression with embedding regression.\n2. **Semantic-aware diffusion:** By perturbing the distance tensor with Gaussian noise, the model embeds semantic structure directly into the diffusion process, enabling smooth transitions between semantically related tokens.\n3. **Comprehensive evaluation:** Experiments cover diverse tasks (XSum, QQP, ParaDetox, ROCStories, Quasar-T) and compare fairly with both discrete (TESS, SSD-LM, D3PM) and continuous (DiffuSeq, Diffusion-LM) baselines.\n4. **Fair comparison setup:** Model sizes (~100M), datasets, and decoding strategies are kept consistent; pretrained advantages of baselines (e.g., TESS) are removed.\n5. **Strong performance:** SMOOTHIE outperforms all diffusion baselines and achieves results comparable to autoregressive models like FLAN-T5.", "weaknesses": "1. **Computation cost not analyzed:** Although vectorization and mean-embedding compression are mentioned, there is no report on runtime, GPU memory, or FLOPs. Given that every token interacts with all vocabulary embeddings, the training cost may be substantial.\n2. **Limited ablations on general enhancements:** While the authors tested self-conditioning and found minimal gains (and thus did not adopt it), other general mechanisms such as the tanh-style noise schedule or δ̃ magnitude lack comprehensive ablations. More systematic sensitivity studies would clarify whether improvements stem primarily from the distance-based diffusion rather than secondary hyperparameters.\n3. **Embedding dependency untested:** The model fixes the BERT embedding matrix during training but does not evaluate alternative embeddings (e.g., random, GloVe, or fine-tuned). It remains unclear how sensitive performance is to the embedding quality or domain.\n4. **No convergence or efficiency analysis:** Beyond theoretical equivalence, convergence behavior and stability relative to token-level diffusion are unreported; adding epoch-wise loss curves would clarify training efficiency.", "questions": "1. **Computation cost:** Can you report training time, GPU memory, or FLOPs compared to TESS or DiffuSeq? How is the full distance tensor computation optimized, and how does it scale with vocabulary size?\n2. **Embedding dependence:** Have you evaluated SMOOTHIE using other embeddings (e.g., random or domain-specific)? How robust is the model to embedding quality and distribution shifts?\n3. **Noise schedule and δ̃ ablations:** Why was the tanh-style schedule chosen? Have you tested alternative schedules or δ̃ values to confirm robustness across datasets?\n4. **Self-conditioning and fairness:** Since self-conditioning yields limited improvement and was excluded, can you confirm that no other general enhancements influenced the main gains? Could similar gains be achieved by applying self-conditioning to baselines?\n5. **Convergence and stability:** Please include training curves or epoch-wise loss comparisons with baseline diffusion models to demonstrate convergence efficiency and numerical stability.\n6. **Scalability:** How does SMOOTHIE perform with larger vocabularies (e.g., >50K tokens)? Are there feasible strategies (e.g., top-k distance pruning or clustering) to reduce complexity while maintaining accuracy?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761793220446}], "openreview_url": "https://openreview.net/forum?id=5y8R8XSHkG", "arxiv_id": "2505.18853", "paper_pdf": "papers/5y8R8XSHkG.pdf", "paper_pdf_sha256": "fd878313ad5515bf0d512bc16e0cefcaf144c95e037528870dad4eb96303a295", "paper_pdf_bytes": 1212515, "paper_pdf_source": "openreview", "code_url": "https://github.com/ashaba1in/smoothie", "code_repository": "ashaba1in/smoothie", "code_commit": "79d3a409fb5044c2108be749634e2fc1ea7fb163", "code_archive": "repos/5y8R8XSHkG.zip", "code_archive_sha256": "c105df33895b0145356c140cd79a78cbae7d29765d327fc1fc0ab4fbd2e02f25", "code_archive_bytes": 45316, "code_file_count": 25, "code_extensions": {".py": 25}, "github_disk_usage_kb": 160, "github_languages": {"Python": 148052}, "github_archived": false, "github_pushed_at": "2026-02-02T15:58:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/smoothie-smoothing-diffusion-on-token"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MM197t8WlM", "year": 2025, "status": "rejected", "title": "Local Flow Matching Generative Models", "authors": ["Chen Xu", "Xiuyuan Cheng", "Yao Xie"], "authorids": ["~Chen_Xu12", "~Xiuyuan_Cheng1", "~Yao_Xie2"], "authors_source": "OpenReview API", "abstract": "Flow Matching (FM) is a simulation-free method for learning a continuous and invertible flow to interpolate between two distributions, and in particular to generate data from noise in generative modeling. In this paper, we introduce Local Flow Matching ($\\texttt{LFM}$), which consecutively learns a sequence of FM sub-models and each matches a diffusion process up to the time of the step size in the data-to-noise direction. In each step, the two distributions to be interpolated by the sub-model are closer to each other than data vs. noise, and this enables the use of smaller models with faster training. The stepwise structure of $\\texttt{LFM}$ is natural to be distilled and different distillation techniques can be adopted to speed up generation. Theoretically, we prove a generation guarantee of the proposed flow model in terms of the $\\chi^2$-divergence between the generated and true data distributions. In experiments, we demonstrate the improved training efficiency and competitive generative performance of $\\texttt{LFM}$ compared to FM on the unconditional generation of tabular data and image datasets, and also on the conditional generation of robotic manipulation policies.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "KL1jfHzQLL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8427/Reviewer_9UHe"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes to divide the original Flow Matching (FM) into a sequence of parts, called Local Flow Matching (LFM). The authors provide theoretical analysis of the convergence of the forward process, and the theoretical guarantee of the $\\chi^2$ divergence between the real data and generated data distributions. The authors evaluated the proposed method on different tasks, tabular data generation, image generation, robotic manipulation and so on.", "review_text": "This paper proposes to divide the original Flow Matching (FM) into a sequence of parts, called Local Flow Matching (LFM). The authors provide theoretical analysis of the convergence of the forward process, and the theoretical guarantee of the $\\chi^2$ divergence between the real data and generated data distributions. The authors evaluated the proposed method on different tasks, tabular data generation, image generation, robotic manipulation and so on.", "strengths": "The authors proposed the Local Flow Matching (LFM), which divides the Flow Matching into a sequence of local parts. The approach is intuitively correct. The authors provided theoretical analysis of the convergence of the forward process, showing that the $\\chi^2$ divergence between the noise distribution and noised real data distribution decreases exponentially w.r.t. the number of sub-flows. Furthermore, the authors presented the theoretical results showing the $\\chi^2$ divergence between the real distribution and the generated distribution is close w.r.t. the flow training loss $\\epsilon$.", "weaknesses": "From the experiments, it turns out that the proposed LFM is not working as well as the original Flow Matching (FM) (Lipman et al., 2023). The proposed method achieved an FID of 8.45 on the CIFAR-10 dataset whereas the FM achieved a lower FID of 6.35 on the same dataset. On ImageNet 32x32 dataset, the proposed method achieved an FID of 7.0, but the FM achieved a lower FID of 5.02. The authors should compare with FM on the CIFAR-10, ImageNet 32x32, the Oxford Flowers and the LSUN Church datasets. \n\nThe proposed LFM seems to require more training time than the original FM because for training $v_n$ we need to inference from $v_1$ to $v_{n-1}$ for each sample. \n\nThe proposed method also need more inference time since we need to inference from $v_1$ to $v_N$.", "questions": "What does $g \\bot x'_l$ mean in Eq. 6? and why do we need this constraint?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to divide the original Flow Matching (FM) into a sequence of parts, called Local Flow Matching (LFM). The authors provide theoretical analysis of the convergence of the forward process, and the theoretical guarantee of the $\\chi^2$ divergence between the real data and generated data distributions. The authors evaluated the proposed method on different tasks, tabular data generation, image generation, robotic manipulation and so on.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The authors proposed the Local Flow Matching (LFM), which divides the Flow Matching into a sequence of local parts. The approach is intuitively correct. The authors provided theoretical analysis of the convergence of the forward process, showing that the $\\chi^2$ divergence between the noise distribution and noised real data distribution decreases exponentially w.r.t. the number of sub-flows. Furthermore, the authors presented the theoretical results showing the $\\chi^2$ divergence between the real distribution and the generated distribution is close w.r.t. the flow training loss $\\epsilon$.", "weaknesses": "From the experiments, it turns out that the proposed LFM is not working as well as the original Flow Matching (FM) (Lipman et al., 2023). The proposed method achieved an FID of 8.45 on the CIFAR-10 dataset whereas the FM achieved a lower FID of 6.35 on the same dataset. On ImageNet 32x32 dataset, the proposed method achieved an FID of 7.0, but the FM achieved a lower FID of 5.02. The authors should compare with FM on the CIFAR-10, ImageNet 32x32, the Oxford Flowers and the LSUN Church datasets. \n\nThe proposed LFM seems to require more training time than the original FM because for training $v_n$ we need to inference from $v_1$ to $v_{n-1}$ for each sample. \n\nThe proposed method also need more inference time since we need to inference from $v_1$ to $v_N$.", "questions": "What does $g \\bot x'_l$ mean in Eq. 6? and why do we need this constraint?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730691242561}, {"id": "scDYoI0yIn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8427/Reviewer_K69V"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents Local Flow Matching (LFM), an enhancement of Flow Matching (FM) for faster, efficient generative modeling. LFM divides the process into sequential FM sub-models, each bridging distributions that are progressively closer from data to noise, allowing smaller models and quicker training. This stepwise approach also enables the use of distillation techniques to further accelerate data generation. Theoretically, LFM provides a generation guarantee based on $\\chi^2$-divergence. Experiments show LFM’s improved training efficiency and strong performance in generating tabular data, images, and robotic manipulation policies.", "review_text": "This paper presents Local Flow Matching (LFM), an enhancement of Flow Matching (FM) for faster, efficient generative modeling. LFM divides the process into sequential FM sub-models, each bridging distributions that are progressively closer from data to noise, allowing smaller models and quicker training. This stepwise approach also enables the use of distillation techniques to further accelerate data generation. Theoretically, LFM provides a generation guarantee based on $\\chi^2$-divergence. Experiments show LFM’s improved training efficiency and strong performance in generating tabular data, images, and robotic manipulation policies.", "strengths": "The paper explores integrating flow-matching submodels into diffusion processes to enable faster and more efficient learning and inference. The approach is novel and is supported by theoretical guarantees of generation. Experimentally, the method can be applied to various tasks, including image generation, tabular data generation, and robotic manipulation.", "weaknesses": "1. More details should be provided, such as the number of function evaluations (NFEs) and the used ODE sampler for all methods in Table 1 and Table 2, to better demonstrate LFM's effectiveness.\n\n2. Although the tasks are diverse, the paper lacks a solid comparison with prior methods on some fundamental tasks. For instance, comparing LFM with Rectified Flow and OU diffusion on CIFAR-10 with the same amount of batches would offer a clearer understanding of LFM's training efficiency and distillation effects.\n\n3. The advantage of LFM on fast training has not been demonstrated well except Table 2.", "questions": "1. Can you clarify more on how the method benefits fast training based on the experimental results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Local Flow Matching (LFM), an enhancement of Flow Matching (FM) for faster, efficient generative modeling. LFM divides the process into sequential FM sub-models, each bridging distributions that are progressively closer from data to noise, allowing smaller models and quicker training. This stepwise approach also enables the use of distillation techniques to further accelerate data generation. Theoretically, LFM provides a generation guarantee based on $\\chi^2$-divergence. Experiments show LFM’s improved training efficiency and strong performance in generating tabular data, images, and robotic manipulation policies.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper explores integrating flow-matching submodels into diffusion processes to enable faster and more efficient learning and inference. The approach is novel and is supported by theoretical guarantees of generation. Experimentally, the method can be applied to various tasks, including image generation, tabular data generation, and robotic manipulation.", "weaknesses": "1. More details should be provided, such as the number of function evaluations (NFEs) and the used ODE sampler for all methods in Table 1 and Table 2, to better demonstrate LFM's effectiveness.\n\n2. Although the tasks are diverse, the paper lacks a solid comparison with prior methods on some fundamental tasks. For instance, comparing LFM with Rectified Flow and OU diffusion on CIFAR-10 with the same amount of batches would offer a clearer understanding of LFM's training efficiency and distillation effects.\n\n3. The advantage of LFM on fast training has not been demonstrated well except Table 2.", "questions": "1. Can you clarify more on how the method benefits fast training based on the experimental results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729999410050}, {"id": "iqhITg06mM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8427/Reviewer_vDPz"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces Local Flow Matching (LFM), a generative approach that builds upon the existing Flow Matching (FM) framework. LFM improves upon FM by dividing the global flow into multiple local flow sub-models. Each sub-model is trained to match a diffusion process between closer intermediate distributions, which reduces the model size and accelerates training.", "review_text": "The paper introduces Local Flow Matching (LFM), a generative approach that builds upon the existing Flow Matching (FM) framework. LFM improves upon FM by dividing the global flow into multiple local flow sub-models. Each sub-model is trained to match a diffusion process between closer intermediate distributions, which reduces the model size and accelerates training.", "strengths": "1.\tThe paper offers a fresh approach in the field of generative modeling, combining ideas from diffusion processes and FM. The idea of breaking down a single large flow into several smaller steps (local flows) is natural.\n2.\tThe paper provides solid theoretical guarantees, specifically proving a generation guarantee in terms of the $\\chi^2$-divergence between the generated and true data distributions. The experiments are comprehensive, covering a range of tasks from image generation to robotic manipulation, and the results demonstrate the performance of LFM.", "weaknesses": "1.\tFM simplifies the diffusion process into a single step, transforming a trajectory from a curve into a straight line, thus reducing training costs and improving sampling efficiency. However, this paper reverses that advantage by breaking this single step into multiple segments. The division of the trajectory into multiple steps could introduce added computational complexity, potentially negating the efficiency gains that FM originally aimed to provide.\n\n2.\tThe paper presents some conceptual ambiguities regarding key terms. Continuous Normalizing Flows (CNF) and neural ODEs are two distinct models, but the paper incorrectly states that \"CNF trains a neural ODE\" (Line 106). In fact, FM is a method for training CNF models, not a model in itself, much like the proposed LFM method. The paper conflates these concepts in several places (e.g. Line 41, 51), which could cause confusion.\n\n3.\tThe paper highlights improved training efficiency through the use of smaller local models. However, the exact speed improvements are not always well-quantified. For example, while the paper states that LFM leads to faster convergence and reduced memory usage, these claims could be supported by more detailed analysis. Providing concrete timing metrics (e.g., runtime comparisons, memory benchmarks across different hardware) would bolster the argument for LFM's practical benefits. More information can be found in Questions part.\n\n4.\tIn line 127, a reference is used as the subject of the sentence but is enclosed in parentheses. Please correct this formatting issue and ensure consistency throughout the paper.", "questions": "1.\tIn the training process, at each step, the target distribution is generated from the sample of the previous step through the Ornstein-Uhlenbeck (OU) process. What is the theoretical basis for using the OU process in this context? This essentially defines checkpoints along the probability path. Would a different process yield similar results, or is there a specific reason the OU process was chosen?\n\n2.\tIn Appendix B1, how were the stepwise parameters $\\rho$ and $c$ selected? Were they chosen solely based on extensive experimentation and selecting the ones that gave the best results, or was there a more principled approach to their selection?\n\n3.\tIn the image generation tasks (Table 2), it would be beneficial to include comparisons with Flow Matching (FM) [1], rather than only with InterFlow, as Local Flow Matching (LFM) is based on FM. This would better illustrate the claimed improved training efficiency. Additionally, either in the main text or in the appendix, it would be helpful to provide a comparison of the total number of parameters for each method across all experiments, rather than only ensuring the same training scheme.\n\n[1] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Local Flow Matching (LFM), a generative approach that builds upon the existing Flow Matching (FM) framework. LFM improves upon FM by dividing the global flow into multiple local flow sub-models. Each sub-model is trained to match a diffusion process between closer intermediate distributions, which reduces the model size and accelerates training.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1.\tThe paper offers a fresh approach in the field of generative modeling, combining ideas from diffusion processes and FM. The idea of breaking down a single large flow into several smaller steps (local flows) is natural.\n2.\tThe paper provides solid theoretical guarantees, specifically proving a generation guarantee in terms of the $\\chi^2$-divergence between the generated and true data distributions. The experiments are comprehensive, covering a range of tasks from image generation to robotic manipulation, and the results demonstrate the performance of LFM.", "weaknesses": "1.\tFM simplifies the diffusion process into a single step, transforming a trajectory from a curve into a straight line, thus reducing training costs and improving sampling efficiency. However, this paper reverses that advantage by breaking this single step into multiple segments. The division of the trajectory into multiple steps could introduce added computational complexity, potentially negating the efficiency gains that FM originally aimed to provide.\n\n2.\tThe paper presents some conceptual ambiguities regarding key terms. Continuous Normalizing Flows (CNF) and neural ODEs are two distinct models, but the paper incorrectly states that \"CNF trains a neural ODE\" (Line 106). In fact, FM is a method for training CNF models, not a model in itself, much like the proposed LFM method. The paper conflates these concepts in several places (e.g. Line 41, 51), which could cause confusion.\n\n3.\tThe paper highlights improved training efficiency through the use of smaller local models. However, the exact speed improvements are not always well-quantified. For example, while the paper states that LFM leads to faster convergence and reduced memory usage, these claims could be supported by more detailed analysis. Providing concrete timing metrics (e.g., runtime comparisons, memory benchmarks across different hardware) would bolster the argument for LFM's practical benefits. More information can be found in Questions part.\n\n4.\tIn line 127, a reference is used as the subject of the sentence but is enclosed in parentheses. Please correct this formatting issue and ensure consistency throughout the paper.", "questions": "1.\tIn the training process, at each step, the target distribution is generated from the sample of the previous step through the Ornstein-Uhlenbeck (OU) process. What is the theoretical basis for using the OU process in this context? This essentially defines checkpoints along the probability path. Would a different process yield similar results, or is there a specific reason the OU process was chosen?\n\n2.\tIn Appendix B1, how were the stepwise parameters $\\rho$ and $c$ selected? Were they chosen solely based on extensive experimentation and selecting the ones that gave the best results, or was there a more principled approach to their selection?\n\n3.\tIn the image generation tasks (Table 2), it would be beneficial to include comparisons with Flow Matching (FM) [1], rather than only with InterFlow, as Local Flow Matching (LFM) is based on FM. This would better illustrate the claimed improved training efficiency. Additionally, either in the main text or in the appendix, it would be helpful to provide a comparison of the total number of parameters for each method across all experiments, rather than only ensuring the same training scheme.\n\n[1] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729949331955}, {"id": "Gjd2hRKtHL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8427/Reviewer_Snd7"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes a novel generative model, Local Flow Matching (LFM), which segments Flow Matching along the time dimension. The transition between two distributions within each time segment is modeled using a smaller-scale sub-model. This generative model offers the advantages of having smaller sub-models, faster convergence during training, and greater convenience for distillation.", "review_text": "The paper proposes a novel generative model, Local Flow Matching (LFM), which segments Flow Matching along the time dimension. The transition between two distributions within each time segment is modeled using a smaller-scale sub-model. This generative model offers the advantages of having smaller sub-models, faster convergence during training, and greater convenience for distillation.", "strengths": "The paper provides a detailed theoretical proof that the data distribution generated by LFM is close to the true data distribution, demonstrating the exponential convergence of the forward process and the generation guarantee of the backward process. In addition, the paper conducts extensive experiments to validate the effectiveness of LFM, including generation on two-dimensional toy data, tabular data generation, image generation, and robotic manipulation. Moreover, the paper is well-written.", "weaknesses": "Some experimental details in the paper may need further clarification. The data in Table 2 is used to demonstrate that LFM requires less computational cost, but the model sizes of both LFM and InterFlow are not provided. Additionally, it is unclear whether the training data shown in the table refers to the data used by each sub-model individually or by all sub-models collectively.", "questions": "Could you provide more detailed data regarding the issues mentioned in the weaknesses, such as the specific number of parameters for each sub-model in LFM, the number of sub-models, and the parameter count of the InterFlow model (or explain why the total number of parameters across all sub-models is equal to that of InterFlow). Could you provide the amount of data required to train each sub-model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel generative model, Local Flow Matching (LFM), which segments Flow Matching along the time dimension. The transition between two distributions within each time segment is modeled using a smaller-scale sub-model. This generative model offers the advantages of having smaller sub-models, faster convergence during training, and greater convenience for distillation.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "The paper provides a detailed theoretical proof that the data distribution generated by LFM is close to the true data distribution, demonstrating the exponential convergence of the forward process and the generation guarantee of the backward process. In addition, the paper conducts extensive experiments to validate the effectiveness of LFM, including generation on two-dimensional toy data, tabular data generation, image generation, and robotic manipulation. Moreover, the paper is well-written.", "weaknesses": "Some experimental details in the paper may need further clarification. The data in Table 2 is used to demonstrate that LFM requires less computational cost, but the model sizes of both LFM and InterFlow are not provided. Additionally, it is unclear whether the training data shown in the table refers to the data used by each sub-model individually or by all sub-models collectively.", "questions": "Could you provide more detailed data regarding the issues mentioned in the weaknesses, such as the specific number of parameters for each sub-model in LFM, the number of sub-models, and the parameter count of the InterFlow model (or explain why the total number of parameters across all sub-models is equal to that of InterFlow). Could you provide the amount of data required to train each sub-model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729500052108}], "openreview_url": "https://openreview.net/forum?id=MM197t8WlM", "arxiv_id": "2410.02548", "paper_pdf": "papers/MM197t8WlM.pdf", "paper_pdf_sha256": "682ac3ac54133fc803f7b9fe17e2ff94cd6460a5e6b421d6049c844d5259252f", "paper_pdf_bytes": 2658300, "paper_pdf_source": "openreview", "code_url": "https://github.com/hamrel-cxu/LocalFlowMatching", "code_repository": "hamrel-cxu/LocalFlowMatching", "code_commit": "b249c1f28b45693b2624b0486df4eef6339b71d3", "code_archive": "repos/MM197t8WlM.zip", "code_archive_sha256": "01e68c5f0f02ed18b81f0fb50659f935ff7ab0642178deb21b00602976eb04b6", "code_archive_bytes": 41362, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 34, "github_languages": {"Python": 43438}, "github_archived": false, "github_pushed_at": "2025-11-10T21:25:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/local-flow-matching-generative-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0IaTFNJner", "year": 2024, "status": "rejected", "title": "On the Embedding Collapse When Scaling up Recommendation Models", "authors": ["Xingzhuo Guo", "Junwei Pan", "Ximei Wang", "Baixu Chen", "Jie Jiang", "Mingsheng Long"], "authorids": ["~Xingzhuo_Guo1", "~Junwei_Pan1", "~Ximei_Wang1", "~Baixu_Chen2", "~Jie_Jiang3", "~Mingsheng_Long5"], "authors_source": "OpenReview API", "abstract": "Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up existing recommendation models and observe that the enlarged models do not improve satisfactorily. In this context, we investigate the embedding layers of enlarged models and identify a phenomenon of *embedding collapse*, which ultimately hinders scalability, wherein the embedding matrix tends to reside in a low-dimensional subspace. Through empirical and theoretical analysis, we demonstrate that the feature interaction module specific to recommendation models has a *two-sided effect*. On the one hand, the interaction restricts embedding learning when interacting with collapsed embeddings, exacerbating the collapse issue. On the other hand, feature interaction is crucial in mitigating the fitting of spurious features, thereby improving scalability. Based on this analysis, we propose a simple yet effective *multi-embedding* design incorporating embedding-set-specific interaction modules to capture diverse patterns and reduce collapse. Extensive experiments demonstrate that this proposed design provides consistent scalability for various recommendation models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "nWhcrWyPwG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission668/Reviewer_9BGA"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies recommendation model performance when scaling up the embedding layers of the model. The paper identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. Through empirical experiments on FFM and DCNv2 and theoretical analysis on FM, the paper shows that the feature interaction process of recommendation models leads to embedding collapse and thus limits the model scalability. The paper also performed empirical experiments on regularized DCNv2 and DNN which led to less collapsed embeddings, but the model performance got worse. The paper proposes multi-embedding, which leads to better performance when scaling up the embedding layers. Experiments demonstrate that this proposed design provides consistent scalability for various recommendation models.", "review_text": "This paper studies recommendation model performance when scaling up the embedding layers of the model. The paper identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. Through empirical experiments on FFM and DCNv2 and theoretical analysis on FM, the paper shows that the feature interaction process of recommendation models leads to embedding collapse and thus limits the model scalability. The paper also performed empirical experiments on regularized DCNv2 and DNN which led to less collapsed embeddings, but the model performance got worse. The paper proposes multi-embedding, which leads to better performance when scaling up the embedding layers. Experiments demonstrate that this proposed design provides consistent scalability for various recommendation models.", "strengths": "Originality: \n- The paper investigates the enlarged embedding layers of recommendation models and identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. The discovery is novel as far as I know.\n- The paper proposed information abundance to measure the degree of collapse for embedding matrices.\n\nQuality:\n- The paper is well-written. It starts with a novel finding of embedding collapse when increasing embedding dimension which might lead to poor scalability, performs empirical experiments and theoretical analysis to show that it is caused by feature interaction, and proposes a solution to increase scalability.\n\nSignificance:\n- The scaling law of recommendation models is an important topic in both academia and industry. This paper investigates the scaling law of embedding layers and shows why naively increasing embedding dim is not sufficient for scalability. The paper proposed multi-embedding which could help alleviate the phenomenon of embedding collapse and improve scalability.", "weaknesses": "- In section 3, the paper proposes Information Abundance to measure the degree of collapse of embedding matrices. As the paper focuses on the scaling law of embedding layers, the paper should discuss whether Information Abundance is a fair metric when comparing embedding matrices of different dimension sizes.\n- In section 4.2, the paper uses regularized DCNv2 as an example to show that suppressing feature interaction is insufficient for scalability. It is unclear to me why feature interaction in regularized DCNv2 is suppressed.\n- The performance of multi-embedding lacks ablation study. One example is that in Figure 7 (right), we can feed all 4 small embeddings into a single feature interaction layer, and test its performance against the proposed approach. We can also test the Information Abundance of such design versus the proposed multi-embedding design.\n- It would be interesting to test the scaling law of embedding layers with increased feature interaction complexity. This can help us better understand whether the embedding collapse phenomenon is caused by insufficient feature interaction complexity.", "questions": "My questions are listed in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies recommendation model performance when scaling up the embedding layers of the model. The paper identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. Through empirical experiments on FFM and DCNv2 and theoretical analysis on FM, the paper shows that the feature interaction process of recommendation models leads to embedding collapse and thus limits the model scalability. The paper also performed empirical experiments on regularized DCNv2 and DNN which led to less collapsed embeddings, but the model performance got worse. The paper proposes multi-embedding, which leads to better performance when scaling up the embedding layers. Experiments demonstrate that this proposed design provides consistent scalability for various recommendation models.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "Originality: \n- The paper investigates the enlarged embedding layers of recommendation models and identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. The discovery is novel as far as I know.\n- The paper proposed information abundance to measure the degree of collapse for embedding matrices.\n\nQuality:\n- The paper is well-written. It starts with a novel finding of embedding collapse when increasing embedding dimension which might lead to poor scalability, performs empirical experiments and theoretical analysis to show that it is caused by feature interaction, and proposes a solution to increase scalability.\n\nSignificance:\n- The scaling law of recommendation models is an important topic in both academia and industry. This paper investigates the scaling law of embedding layers and shows why naively increasing embedding dim is not sufficient for scalability. The paper proposed multi-embedding which could help alleviate the phenomenon of embedding collapse and improve scalability.", "weaknesses": "- In section 3, the paper proposes Information Abundance to measure the degree of collapse of embedding matrices. As the paper focuses on the scaling law of embedding layers, the paper should discuss whether Information Abundance is a fair metric when comparing embedding matrices of different dimension sizes.\n- In section 4.2, the paper uses regularized DCNv2 as an example to show that suppressing feature interaction is insufficient for scalability. It is unclear to me why feature interaction in regularized DCNv2 is suppressed.\n- The performance of multi-embedding lacks ablation study. One example is that in Figure 7 (right), we can feed all 4 small embeddings into a single feature interaction layer, and test its performance against the proposed approach. We can also test the Information Abundance of such design versus the proposed multi-embedding design.\n- It would be interesting to test the scaling law of embedding layers with increased feature interaction complexity. This can help us better understand whether the embedding collapse phenomenon is caused by insufficient feature interaction complexity.", "questions": "My questions are listed in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698829848014}, {"id": "IlRLbALEdX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission668/Reviewer_JdNi"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper suggests that the embedding collapse phenomenon restricts the scalability of existing recommendation models. Empirical and theoretical analysis show that interaction with collapsed embeddings constrains embedding learning. Also, this paper proposes a multi-embedding design incorporating embedding-set-specific interaction modules to capture diverse patterns and reduce collapse.", "review_text": "This paper suggests that the embedding collapse phenomenon restricts the scalability of existing recommendation models. Empirical and theoretical analysis show that interaction with collapsed embeddings constrains embedding learning. Also, this paper proposes a multi-embedding design incorporating embedding-set-specific interaction modules to capture diverse patterns and reduce collapse.", "strengths": "S1. This paper provides empirical and theoretical analysis of the embedding collapse phenomenon.\n\nS2. This paper provides information abundance for quantifying the degree of collapse for such matrices with low-rank tendencies.", "weaknesses": "W1. The novelty of this paper seems to be limited. The method of dividing the single embedding into multi-embedding sets is similar to DMRL[1] for disentangled representation learning. DMRL divides the feature representation of each modality into k chunks. As a result, the features of different factors are entangled.\n\nW2. The motivation is not completely solid. The reason for increasing the embedding size of the model is inappropriate.\n\nW3. The experimental results of the paper are insufficient. When the embedding size was scaled up through multi-embedding, the experimental results show that performance increases. However, the performance improvement is marginal.\n\n[1] Disentangled Multimodal Representation Learning for Recommendation, IEEE’22", "questions": "Depending on the model and dataset, the model will have an appropriate embedding size. Therefore, it seems reasonable that the performance will drop if the embedding size deviates from the proper value. Then. do we need to scale up the embedding size for the same dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper suggests that the embedding collapse phenomenon restricts the scalability of existing recommendation models. Empirical and theoretical analysis show that interaction with collapsed embeddings constrains embedding learning. Also, this paper proposes a multi-embedding design incorporating embedding-set-specific interaction modules to capture diverse patterns and reduce collapse.", "soundness": "3 good", "presentation": "2 fair", "contribution": "1 poor", "strengths": "S1. This paper provides empirical and theoretical analysis of the embedding collapse phenomenon.\n\nS2. This paper provides information abundance for quantifying the degree of collapse for such matrices with low-rank tendencies.", "weaknesses": "W1. The novelty of this paper seems to be limited. The method of dividing the single embedding into multi-embedding sets is similar to DMRL[1] for disentangled representation learning. DMRL divides the feature representation of each modality into k chunks. As a result, the features of different factors are entangled.\n\nW2. The motivation is not completely solid. The reason for increasing the embedding size of the model is inappropriate.\n\nW3. The experimental results of the paper are insufficient. When the embedding size was scaled up through multi-embedding, the experimental results show that performance increases. However, the performance improvement is marginal.\n\n[1] Disentangled Multimodal Representation Learning for Recommendation, IEEE’22", "questions": "Depending on the model and dataset, the model will have an appropriate embedding size. Therefore, it seems reasonable that the performance will drop if the embedding size deviates from the proper value. Then. do we need to scale up the embedding size for the same dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698822478538}, {"id": "9kPJm1DFwm", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission668/Reviewer_DK5H"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses challenges in scaling up recommendation models, identifying a phenomenon called \"embedding collapse\" while enlarging models. The authors introduce information abundance as a metric to measure and evaluate collapse. They show that embedding matrix reside mostly in a low-dimensional subspace in scaled up models. The study analyzes feature interactions in models, noting they reduce overfitting but can also exacerbate embedding collapse. To tackle this, the authors introduce a multi-embedding design, scaling independent embedding sets and integrating specific interaction modules. This approach claims to improve scalability across various recommendation models. Main contributions:\n- Highlight non-scalability in recommendation models and define the \"embedding collapse\" phenomenon.\n- Empirical and theoretical analysis reveals the dual impact of feature interaction on scalability.\n- Introduction of the multi-embedding design to achieve scalability improvements for recommendation models.", "review_text": "The paper addresses challenges in scaling up recommendation models, identifying a phenomenon called \"embedding collapse\" while enlarging models. The authors introduce information abundance as a metric to measure and evaluate collapse. They show that embedding matrix reside mostly in a low-dimensional subspace in scaled up models. The study analyzes feature interactions in models, noting they reduce overfitting but can also exacerbate embedding collapse. To tackle this, the authors introduce a multi-embedding design, scaling independent embedding sets and integrating specific interaction modules. This approach claims to improve scalability across various recommendation models. Main contributions:\n- Highlight non-scalability in recommendation models and define the \"embedding collapse\" phenomenon.\n- Empirical and theoretical analysis reveals the dual impact of feature interaction on scalability.\n- Introduction of the multi-embedding design to achieve scalability improvements for recommendation models.", "strengths": "Originality:\n- ‘Information Abundance' as a quantitative novel measure to measure the embedding layer collapse.\n- The 'Interaction-Collapse Law' and the two sided effect of feature interaction process helps improve the understanding of embeddings' behavior in recommendation systems.\n\nQuality:\n- The authors have detailed exploration of embeddings and their behavior, particularly in the context of information collapse with rigorous visualizations. \n\nSignificance:\n- Broad Implications for Recommendation Systems: Given the ubiquitous nature of recommendation systems in today's digital platforms, insights into their workings, particularly regarding embeddings, have widespread implications.\n- Potential for Future Research: Introducing novel concepts and metrics invariably opens the door for future studies, both to validate and to build upon these ideas. The 'Interaction-Collapse Law', for instance, may become a focal point in subsequent research.", "weaknesses": "1. Insufficient Empirical Validation on Large-Scale Data: The authors have shown with empirical evidences that large scale recommendation models scale poorly. However it is a common knowledge that large scale models are inherently data hungry to achieve better model convergence. This is an important premise that the paper relies on, it would good if authors can follow up to prove/disprove this as additional data points in this paper. The experiments seem to be on same amount of training data on scaled up models which does not unlock the full power of the scaled up model\n\n2. Studies on Collapse and its effect on scalability and overfitting seem to be limited to Single Embedding studies. To make a stronger case about the proposed method, it would be great if authors can provide discussion on the same for proposed Multi Embedding Design.", "questions": "1. How do you justify the experiment setup when we know that bigger models are inherently data hungry and need more training data to achieve full convergence?\n\n2. From the experiment results shown, the single embedding and multi embedding cases show marginal improvement in AUC values. Can authors provide more context on the significance of the improvements here and if they are well outside the noise region?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses challenges in scaling up recommendation models, identifying a phenomenon called \"embedding collapse\" while enlarging models. The authors introduce information abundance as a metric to measure and evaluate collapse. They show that embedding matrix reside mostly in a low-dimensional subspace in scaled up models. The study analyzes feature interactions in models, noting they reduce overfitting but can also exacerbate embedding collapse. To tackle this, the authors introduce a multi-embedding design, scaling independent embedding sets and integrating specific interaction modules. This approach claims to improve scalability across various recommendation models. Main contributions:\n- Highlight non-scalability in recommendation models and define the \"embedding collapse\" phenomenon.\n- Empirical and theoretical analysis reveals the dual impact of feature interaction on scalability.\n- Introduction of the multi-embedding design to achieve scalability improvements for recommendation models.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Originality:\n- ‘Information Abundance' as a quantitative novel measure to measure the embedding layer collapse.\n- The 'Interaction-Collapse Law' and the two sided effect of feature interaction process helps improve the understanding of embeddings' behavior in recommendation systems.\n\nQuality:\n- The authors have detailed exploration of embeddings and their behavior, particularly in the context of information collapse with rigorous visualizations. \n\nSignificance:\n- Broad Implications for Recommendation Systems: Given the ubiquitous nature of recommendation systems in today's digital platforms, insights into their workings, particularly regarding embeddings, have widespread implications.\n- Potential for Future Research: Introducing novel concepts and metrics invariably opens the door for future studies, both to validate and to build upon these ideas. The 'Interaction-Collapse Law', for instance, may become a focal point in subsequent research.", "weaknesses": "1. Insufficient Empirical Validation on Large-Scale Data: The authors have shown with empirical evidences that large scale recommendation models scale poorly. However it is a common knowledge that large scale models are inherently data hungry to achieve better model convergence. This is an important premise that the paper relies on, it would good if authors can follow up to prove/disprove this as additional data points in this paper. The experiments seem to be on same amount of training data on scaled up models which does not unlock the full power of the scaled up model\n\n2. Studies on Collapse and its effect on scalability and overfitting seem to be limited to Single Embedding studies. To make a stronger case about the proposed method, it would be great if authors can provide discussion on the same for proposed Multi Embedding Design.", "questions": "1. How do you justify the experiment setup when we know that bigger models are inherently data hungry and need more training data to achieve full convergence?\n\n2. From the experiment results shown, the single embedding and multi embedding cases show marginal improvement in AUC values. Can authors provide more context on the significance of the improvements here and if they are well outside the noise region?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698801912398}, {"id": "Sb5jATn86P", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission668/Reviewer_eFdv"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper observes scaling the embedding dimensionality does not lead to satisfactory improvement in recommendation models and claims this is due to the dimensional collapse problem where the learned embedding vector spans only on lower dimensional subspace. The paper further claims such problem will propagate by feature interaction module when a feature embedding interacts with collapsed embeddings and regularization could mitigate collapse but harms the performance. Thus this paper proposes an alternative approach to address the embedding collapse issue, namely to replace single embedding in the original model with multiple embeddings. The paper shows the proposed approach lead to higher \"information abundance\" which is ratio between sum of absolute singular values and the largest singular value, indicate the \"spreadness\"/concentration of the singular value distribution. The paper also applied the proposed approach on several recommender models on two datasets, shows a absolute improvement of AUC on 1e-3 ~ 1e-4 level, when scaling the number of multi-embeddings. \n\ninitial recommendation: weak reject, for reasons please see the weak points.", "review_text": "This paper observes scaling the embedding dimensionality does not lead to satisfactory improvement in recommendation models and claims this is due to the dimensional collapse problem where the learned embedding vector spans only on lower dimensional subspace. The paper further claims such problem will propagate by feature interaction module when a feature embedding interacts with collapsed embeddings and regularization could mitigate collapse but harms the performance. Thus this paper proposes an alternative approach to address the embedding collapse issue, namely to replace single embedding in the original model with multiple embeddings. The paper shows the proposed approach lead to higher \"information abundance\" which is ratio between sum of absolute singular values and the largest singular value, indicate the \"spreadness\"/concentration of the singular value distribution. The paper also applied the proposed approach on several recommender models on two datasets, shows a absolute improvement of AUC on 1e-3 ~ 1e-4 level, when scaling the number of multi-embeddings. \n\ninitial recommendation: weak reject, for reasons please see the weak points.", "strengths": "* The claim about the dimensional collapsing problem when scaling the dimensionality of feature embeddings in recommender system is reasonably supported in the paper and sound.\n* The analysis about the trade off between feature embedding collapsing and the regularization level of feature interaction is interesting and reasonable. \n* The paper is easy to follow", "weaknesses": "* The computation of \"information abundance\" for multi-embedding setting is not clearly defined in the paper. As show in the Figure 1.b for features with low predictive power would have embedding with low \"information abundance\" ratio, and scaling the dimensionality would further lower the ratio. Thus if the ratio for multi embedding is computed by averaging over the small embeddings, the lower ratios for some small embeddings will be compensated. This possibility makes the proposed abundance ratio less reliable. \n* The idea of multi-facet embedding or polysemy embedding has been studied quite extensively in the past. From network embedding (Liu et al. Is a single vector enough? exploring node polysemy for network embedding) to recommender systems (Weston et al. Nonlinear latent factorization by embedding multiple user interests). However, non of the related work on multi-embedding has been discussed in the paper.\n* The finding 1 was claimed to \"applicable to general recommendation models instead of only the models with sub-embeddings\" without being linked to any evidences. Also, I personally find find 1 is much a theory rather than a \"law\".", "questions": "* How is the \"information abundance\" for multi-embeddings computed? \n* How would the proposed multi-embedding approach be positioned in the literature?\n* Please point out the evidence for the claim quoted in the third weak point.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper observes scaling the embedding dimensionality does not lead to satisfactory improvement in recommendation models and claims this is due to the dimensional collapse problem where the learned embedding vector spans only on lower dimensional subspace. The paper further claims such problem will propagate by feature interaction module when a feature embedding interacts with collapsed embeddings and regularization could mitigate collapse but harms the performance. Thus this paper proposes an alternative approach to address the embedding collapse issue, namely to replace single embedding in the original model with multiple embeddings. The paper shows the proposed approach lead to higher \"information abundance\" which is ratio between sum of absolute singular values and the largest singular value, indicate the \"spreadness\"/concentration of the singular value distribution. The paper also applied the proposed approach on several recommender models on two datasets, shows a absolute improvement of AUC on 1e-3 ~ 1e-4 level, when scaling the number of multi-embeddings. \n\ninitial recommendation: weak reject, for reasons please see the weak points.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "* The claim about the dimensional collapsing problem when scaling the dimensionality of feature embeddings in recommender system is reasonably supported in the paper and sound.\n* The analysis about the trade off between feature embedding collapsing and the regularization level of feature interaction is interesting and reasonable. \n* The paper is easy to follow", "weaknesses": "* The computation of \"information abundance\" for multi-embedding setting is not clearly defined in the paper. As show in the Figure 1.b for features with low predictive power would have embedding with low \"information abundance\" ratio, and scaling the dimensionality would further lower the ratio. Thus if the ratio for multi embedding is computed by averaging over the small embeddings, the lower ratios for some small embeddings will be compensated. This possibility makes the proposed abundance ratio less reliable. \n* The idea of multi-facet embedding or polysemy embedding has been studied quite extensively in the past. From network embedding (Liu et al. Is a single vector enough? exploring node polysemy for network embedding) to recommender systems (Weston et al. Nonlinear latent factorization by embedding multiple user interests). However, non of the related work on multi-embedding has been discussed in the paper.\n* The finding 1 was claimed to \"applicable to general recommendation models instead of only the models with sub-embeddings\" without being linked to any evidences. Also, I personally find find 1 is much a theory rather than a \"law\".", "questions": "* How is the \"information abundance\" for multi-embeddings computed? \n* How would the proposed multi-embedding approach be positioned in the literature?\n* Please point out the evidence for the claim quoted in the third weak point.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698736644154}], "openreview_url": "https://openreview.net/forum?id=0IaTFNJner", "arxiv_id": "2310.04400", "paper_pdf": "papers/0IaTFNJner.pdf", "paper_pdf_sha256": "61bed8f96538f35463a03eb5a446a3f24d35786118d26899a86f27c029d22cf4", "paper_pdf_bytes": 2016960, "paper_pdf_source": "openreview", "code_url": "https://github.com/thuml/Multi-Embedding", "code_repository": "thuml/Multi-Embedding", "code_commit": "55078571bf1922dced112505514958f85ca4804c", "code_archive": "repos/0IaTFNJner.zip", "code_archive_sha256": "f3a9d4c906e13a9e12a476736cd2b9a64b142b9592833dc598a54d5c8c99996e", "code_archive_bytes": 40339, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 31, "github_languages": {"Python": 114761}, "github_archived": false, "github_pushed_at": "2024-08-04T00:56:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-embedding-collapse-when-scaling-up"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GKpwIa9wgwR", "year": 2023, "status": "rejected", "title": "Efficient Data Subset Selection to Generalize Training Across Models: Transductive and Inductive Networks", "authors": ["Eeshaan Jain", "Tushar Nandy", "Gaurav Aggarwal", "Ashish V. Tendulkar", "Rishabh K Iyer", "Abir De"], "authorids": ["~Eeshaan_Jain1", "~Tushar_Nandy1", "~Gaurav_Aggarwal4", "~Ashish_V._Tendulkar1", "~Rishabh_K_Iyer2", "~Abir_De1"], "authors_source": "OpenReview API", "abstract": "Subset selection, in recent times, has emerged as a successful approach toward efficient training of models by significantly reducing the amount of data and computational resources required. However, existing methods employ discrete combinatorial and model-specific approaches which lack generalizability--- for each new model, the algorithm has to be executed from the beginning. Therefore, for data subset selection for an unseen architecture, one cannot use the subset chosen for a different model. In this work, we propose SubSelNet, a non-adaptive  subset selection framework, which tackles these problems with two main components. First, we introduce an attention-based neural gadget that leverages the graph structure of architectures and acts as a surrogate to trained deep neural networks for quick model prediction. Then, we use these predictions to build subset samplers. This leads us to develop two variants of  SubSelNet. The first variant is transductive (called as Transductive-SubSelNet) which computes the subset separately for each model by solving a small optimization problem. Such an optimization is still super fast, thanks to the replacement of explicit model training by the model approximator. The second variant is inductive (called as Inductive-SubSelNet) which computes the subset using a trained subset selector, without any optimization.  Most state-of-the-art data subset selection approaches are adaptive, in that the subset selection adapts as the training progresses, and as a result, they require access to the entire data at training time.  Our approach, in contrast, is non-adaptive and does the subset selection only once in the beginning, thereby achieving resource and memory efficiency along with compute-efficiency at training time. Our experiments show that both transductive and inductive variants of our models outperform several methods on the quality of the subset chosen and further demonstrate that our method can be used for choosing the best architecture from a set of architectures.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "nzK5ddbTsE0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5898/Reviewer_B3Rj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this paper, the authors present SUBSELNET - a non- adaptive subset selection framework for solving a particular aspect of subset selection problem - improving generalizability of the subset selection approach; with existing methods, the algorithm has to be executed from the beginning for each new model.\nThe authors introduce an attention based neural approach that uses the graph structure of the architectures, which is then used to build subset samplers. Their approach has 2 variants: transductive and inductive.\nThey claim that their approach is more efficient than the existing approaches since the subset is chosen at the beginning of the training process, and the entire dataset is not required through the training process. ", "review_text": "The approach overall is interesting since, they are able to preselect the dataset for the training process. The authors introduce an attention based neural approach that uses the graph structure of the architectures, which is then used to build subset samplers. Their approach has 2 variants: transductive and inductive. They claim that their approach is more efficient than the existing approaches since the subset is chosen at the beginning of the training process, and the entire dataset is not required through the training process.\n\nOverall, there are a few areas that are not clear from the paper, including the implementation and the evaluation of the approach. ", "strengths": "The authors have provided the motivation for the problem, and presented the solution to address the problem. They have also evaluated their approach against 6 other approaches, and 3 different datasets, showing the speed up and memory utilization. \nThe approach overall is interesting since, they are able to preselect the dataset for the training process. \n\nHowever, there are a few areas that are not clear from the paper.\n1. It would have been great if the authors spent more time discussing the pros- and cons- of their graph network. In general, graph networks themselves can be large, slow and memory consuming. It seems like the comparison is performed on the output of the graph network rather than the end-to-end approach.\nThe details about the GNN and the graph embedding are important.\n\n2. Much of the details, including the step-by-step algorithm and the details are added to the appendix. For instance, the diagram in appendix B and Pseudocode in C, both would have helped understand the paper better, if it was in the main text.\n\n3. Since the approach relies on pre-selecting, it is not clear how the approach is able to avoid overfitting or underfitting. The authors have split the data into train, validation and test sets. Including a report on the accuracy on these datasets, and the time/ computation resources required for these would have been helpful.\n\n4. The loss function (eq 1) and the objective functions (5 and 6) require more explanation. For intance, the authors state in page 3 after eq1 that: \"One can use submodular functions (Fujishige, 2005; Iyer, 2015) like Facility Location, graph cut, or Log-Determinants to model DIVERSITY(S)\". However, they havent mentioned the approach they have used in the paper. \nLater, they mention the use of entropy on the subset sampler H(Prπ(•)) to model the diversity in page 4 after eq 5 and  KL after eq 6. The choice of the functions needs to be elaborated to appreciate the approach better. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors present SUBSELNET - a non- adaptive subset selection framework for solving a particular aspect of subset selection problem - improving generalizability of the subset selection approach; with existing methods, the algorithm has to be executed from the beginning for each new model.\nThe authors introduce an attention based neural approach that uses the graph structure of the architectures, which is then used to build subset samplers. Their approach has 2 variants: transductive and inductive.\nThey claim that their approach is more efficient than the existing approaches since the subset is chosen at the beginning of the training process, and the entire dataset is not required through the training process. ", "strength_and_weaknesses": "The authors have provided the motivation for the problem, and presented the solution to address the problem. They have also evaluated their approach against 6 other approaches, and 3 different datasets, showing the speed up and memory utilization. \nThe approach overall is interesting since, they are able to preselect the dataset for the training process. \n\nHowever, there are a few areas that are not clear from the paper.\n1. It would have been great if the authors spent more time discussing the pros- and cons- of their graph network. In general, graph networks themselves can be large, slow and memory consuming. It seems like the comparison is performed on the output of the graph network rather than the end-to-end approach.\nThe details about the GNN and the graph embedding are important.\n\n2. Much of the details, including the step-by-step algorithm and the details are added to the appendix. For instance, the diagram in appendix B and Pseudocode in C, both would have helped understand the paper better, if it was in the main text.\n\n3. Since the approach relies on pre-selecting, it is not clear how the approach is able to avoid overfitting or underfitting. The authors have split the data into train, validation and test sets. Including a report on the accuracy on these datasets, and the time/ computation resources required for these would have been helpful.\n\n4. The loss function (eq 1) and the objective functions (5 and 6) require more explanation. For intance, the authors state in page 3 after eq1 that: \"One can use submodular functions (Fujishige, 2005; Iyer, 2015) like Facility Location, graph cut, or Log-Determinants to model DIVERSITY(S)\". However, they havent mentioned the approach they have used in the paper. \nLater, they mention the use of entropy on the subset sampler H(Prπ(•)) to model the diversity in page 4 after eq 5 and  KL after eq 6. The choice of the functions needs to be elaborated to appreciate the approach better. ", "clarity,_quality,_novelty_and_reproducibility": "The approach overall is interesting since, they are able to preselect the dataset for the training process. It seems novel in that aspect.\nThe paper's clarity can be improved - important and interesting parts have been moved to the appendix, whereas the math behind the model could have been explained better. \nThe paper as presented, is less easy to reproduce - details about the GNN, the embeddings etc are probably missing. ", "summary_of_the_review": "The approach overall is interesting since, they are able to preselect the dataset for the training process. The authors introduce an attention based neural approach that uses the graph structure of the architectures, which is then used to build subset samplers. Their approach has 2 variants: transductive and inductive. They claim that their approach is more efficient than the existing approaches since the subset is chosen at the beginning of the training process, and the entire dataset is not required through the training process.\n\nOverall, there are a few areas that are not clear from the paper, including the implementation and the evaluation of the approach. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666918152722}, {"id": "b6R2wmRkGK3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5898/Reviewer_VJSk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In  this paper,  a new non-adaptive data subset selection method is proposed. The traditional adaptive method mixed the training and  subset selection. While for the new proposed method, the subset processing is done before the training. Furthermore, the paper also proposed the transductive and inductive variants. The experimental results verified that both variants output perform the baselines on subset selection and also an be used to choose the best architecture.\n", "review_text": "In summary I think this paper makes a good presentation on a new method for the non-adaptive methods and verifies the contributions in the experimental studies. It will be better to fix some annotations and have more ablation studies and show more results.", "strengths": "This paper has a clear decomposition on the model into parts including model approximator and subset sampler. And for each part, it has clear annotations to explain the whole process. \nFor the subset sampler, 2 variants are proposed and the experimental results show the tradeoff between them, and furthermore, a combination of these 2 variants are tested too.\nThe comprehensive experimental results proved the effectiveness of the proposed method.\nSome questions:\n1. some writing errors such as “viz.” appeared a couple of times.\n2. both formula (5) and (6) has the E_S, is that correct? \n3. for the E_S optimization objective such as (6), as the parameter \\pi is under the prob distribution and needs sampling, how do you optimize the \\pi? do you use some reparameterization trick which is now shown in the paper.\n4. Can you elaborate more on why you don’t jointly optimize the gnn parameter and transformer parameter?\n5. In the experimental setup, the paper mentions all the baselines are non-adaptive, is it correct?\n6. Do you have any results to show the accuracy gap between the neural model approximator and the fully trained model?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In  this paper,  a new non-adaptive data subset selection method is proposed. The traditional adaptive method mixed the training and  subset selection. While for the new proposed method, the subset processing is done before the training. Furthermore, the paper also proposed the transductive and inductive variants. The experimental results verified that both variants output perform the baselines on subset selection and also an be used to choose the best architecture.\n", "strength_and_weaknesses": "This paper has a clear decomposition on the model into parts including model approximator and subset sampler. And for each part, it has clear annotations to explain the whole process. \nFor the subset sampler, 2 variants are proposed and the experimental results show the tradeoff between them, and furthermore, a combination of these 2 variants are tested too.\nThe comprehensive experimental results proved the effectiveness of the proposed method.\nSome questions:\n1. some writing errors such as “viz.” appeared a couple of times.\n2. both formula (5) and (6) has the E_S, is that correct? \n3. for the E_S optimization objective such as (6), as the parameter \\pi is under the prob distribution and needs sampling, how do you optimize the \\pi? do you use some reparameterization trick which is now shown in the paper.\n4. Can you elaborate more on why you don’t jointly optimize the gnn parameter and transformer parameter?\n5. In the experimental setup, the paper mentions all the baselines are non-adaptive, is it correct?\n6. Do you have any results to show the accuracy gap between the neural model approximator and the fully trained model?\n", "clarity,_quality,_novelty_and_reproducibility": "This paper proposed a new non-adaptive method for the data subset selection problem and has a clear description of the proposed method. Also the experimental results verify the paper’s arguments: mainly on the advantage of the tradeoff between speedup and memory.", "summary_of_the_review": "In summary I think this paper makes a good presentation on a new method for the non-adaptive methods and verifies the contributions in the experimental studies. It will be better to fix some annotations and have more ablation studies and show more results.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666664761359}, {"id": "t7S7LPqaZQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5898/Reviewer_a8HM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper introduces a SUBSELNET to select a subset of training data. SUBSELNET is a non-adaptive method as it is agnostic to model architecture and training stage and. The authors design a neural model approximator to approximate the output of any given architecture. Two variants of subset sampler are proposed.  SUBSELNET is compared against state-of-the-art methods on three datasets to demonstrate the tradeoff advantage between accuracy and speed up.", "review_text": "I would recommend weak rejection because of the experiment results. It is hard to judge the effectiveness of the proposed method when all methods exhibit significant accuracy drops.  I would increase my score if the authors could demonstrate the efficiency advantage of the proposed method at no or negligible accuracy loss. ", "strengths": "Strength:\n\n(1)\tWell-motivated problem. Data selection for an unseen architecture is an important problem with real-world impact.\n\nWeakness:\n\n(1)\tThe main problem of this work is the experiment. The authors choose 0.5% and 5% of training data, however, with such a low subsample rate, we notice a dramatic accuracy drop in Figure 1.  It is hard to judge the effectiveness of any methods with such a significant accuracy drop. The experiment should reflect at what speed up, the model can maintain the same accuracy. Essentially, focusing on the left side of Figure 1. The authors may consider a larger range of sampling rates (5%, 10%, 20%,30%, 40%, 50%, 60%, 70%), following the setup of previous work such as GraNd and Gister. \n\nQuestion:\n\n(1)\tWhen calculating inference time, are you assume selection and training are conducted in sequential? Is it possible to perform selection in parallel with training when loading a batch of data?\n\n(2)\tWhat is the training cost of training the model approximator? Have the authors evaluated the approximator error? \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a SUBSELNET to select a subset of training data. SUBSELNET is a non-adaptive method as it is agnostic to model architecture and training stage and. The authors design a neural model approximator to approximate the output of any given architecture. Two variants of subset sampler are proposed.  SUBSELNET is compared against state-of-the-art methods on three datasets to demonstrate the tradeoff advantage between accuracy and speed up.", "strength_and_weaknesses": "Strength:\n\n(1)\tWell-motivated problem. Data selection for an unseen architecture is an important problem with real-world impact.\n\nWeakness:\n\n(1)\tThe main problem of this work is the experiment. The authors choose 0.5% and 5% of training data, however, with such a low subsample rate, we notice a dramatic accuracy drop in Figure 1.  It is hard to judge the effectiveness of any methods with such a significant accuracy drop. The experiment should reflect at what speed up, the model can maintain the same accuracy. Essentially, focusing on the left side of Figure 1. The authors may consider a larger range of sampling rates (5%, 10%, 20%,30%, 40%, 50%, 60%, 70%), following the setup of previous work such as GraNd and Gister. \n\nQuestion:\n\n(1)\tWhen calculating inference time, are you assume selection and training are conducted in sequential? Is it possible to perform selection in parallel with training when loading a batch of data?\n\n(2)\tWhat is the training cost of training the model approximator? Have the authors evaluated the approximator error? \n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: Clarity is fine. Figures are generally easy to read. Most descriptions are clear. \n\nQuality:  Experiments setups can be improved ( See Weakness 1). Writing can be polished with noticeable grammar mistakes.   For example, Section 5.1 Model architectures and baselines: four non-adaptive  -> four adaptive. Figure1 Caption ;and; -> , and\n\nNovelty:  Neural model approximator using GNN seems new for data selection. \n", "summary_of_the_review": "I would recommend weak rejection because of the experiment results. It is hard to judge the effectiveness of the proposed method when all methods exhibit significant accuracy drops.  I would increase my score if the authors could demonstrate the efficiency advantage of the proposed method at no or negligible accuracy loss. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666647971462}], "openreview_url": "https://openreview.net/forum?id=GKpwIa9wgwR", "arxiv_id": "2409.12255", "paper_pdf": "papers/GKpwIa9wgwR.pdf", "paper_pdf_sha256": "581117c8d34f86e758a7c520ab2a822181addc574bd3bcdbde56bc663cb35310", "paper_pdf_bytes": 1498817, "paper_pdf_source": "openreview", "code_url": "https://github.com/structlearning/subselnet", "code_repository": "structlearning/subselnet", "code_commit": "8e75a6c18224940a59de0721e9a09bf562d4f30b", "code_archive": "repos/GKpwIa9wgwR.zip", "code_archive_sha256": "9df6b1abbb98d39157c2eeefe8aa396e69aab37421407a84a52bba598d11f19b", "code_archive_bytes": 36720, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 39, "github_languages": {"Python": 83913}, "github_archived": false, "github_pushed_at": "2024-01-28T10:31:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-data-subset-selection-to-generalize-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "L2a_bcarHcF", "year": 2022, "status": "rejected", "title": "Linear algebra with transformers", "authors": ["Francois Charton"], "authorids": ["~Francois_Charton1"], "authors_source": "OpenReview API", "abstract": "Most applications of transformers to mathematics, from integration to theorem proving, focus on symbolic computation. In this paper, we show that transformers can be trained to perform numerical calculations with high accuracy. We consider problems of linear algebra:  matrix transposition, addition, multiplication, eigenvalues and vectors, singular value decomposition, and inversion. Training small transformers (up to six layers) over datasets of random matrices, we achieve high accuracies (over 90%) on all problems. We also show that trained models can generalize out of their training distribution, and that out-of-domain accuracy can be greatly improved by working from more diverse datasets (in particular, by training from matrices with non-independent and identically distributed coefficients). Finally, we show that few-shot learning can be leveraged to retrain models to solve larger problems.\n\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1FKI86JQhfk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper866/Reviewer_UG5r"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper “Linear algebra with transformers” studies the application of seq2seq transformers to matrix operations. It studies their performance across different encodings of floating point numbers, different sizes of matrices, different operations, and different (synthetic) data distributions. The main findings are that transformers work surprisingly well on various matrix operations (addition, multiplication, eigenvalues, inversion, SVD, …) for small matrices (e.g. 5x5), and that generalization to OOD problems is not symmetric (I.e. generalization from one distribution to another does not imply the other way round).", "review_text": "Strengths:\n- The authors present a detailed study of many important questions. This is a fairly comprehensive work on the idea.\n- Thought-provoking application of transformers.\n- Very well written.\n\nWeaknesses:\n- I would have loved to see more than just L1 distances\n- The paper studies the question only on random matrices. In other symbolic domains we have seen that insights gained from machine learning approaches trained on random data do not necessarily carry over to “real-world” distributions. I would have loved to see a study that includes a wider variety of training and evaluation data.\n- The models used in this paper have sometimes rather odd (small) hyper parameters. E.g., for many experiments the models have only 2 layers. I’d love to see larger models and see if they improve the results (and the exact number of parameters).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper “Linear algebra with transformers” studies the application of seq2seq transformers to matrix operations. It studies their performance across different encodings of floating point numbers, different sizes of matrices, different operations, and different (synthetic) data distributions. The main findings are that transformers work surprisingly well on various matrix operations (addition, multiplication, eigenvalues, inversion, SVD, …) for small matrices (e.g. 5x5), and that generalization to OOD problems is not symmetric (I.e. generalization from one distribution to another does not imply the other way round).", "main_review": "Strengths:\n- The authors present a detailed study of many important questions. This is a fairly comprehensive work on the idea.\n- Thought-provoking application of transformers.\n- Very well written.\n\nWeaknesses:\n- I would have loved to see more than just L1 distances\n- The paper studies the question only on random matrices. In other symbolic domains we have seen that insights gained from machine learning approaches trained on random data do not necessarily carry over to “real-world” distributions. I would have loved to see a study that includes a wider variety of training and evaluation data.\n- The models used in this paper have sometimes rather odd (small) hyper parameters. E.g., for many experiments the models have only 2 layers. I’d love to see larger models and see if they improve the results (and the exact number of parameters).", "summary_of_the_review": "This work explores a wonderful idea: to solve linear algebra with transformer models. While these models use way more compute internally than the problem they are applied to requires to solve, it is an intriguing question whether these computations can be learned from scratch without further biases. The surprising answer is that this works relatively well for small matrices.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636166032517}, {"id": "bIGYa7oe_y5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper866/Reviewer_TnqM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper consider the problem of approximating algebraic computations over matrices using transformers.  \nExperiments with different encodings are presented, investigating the use of transformers for approximating a number of algebraic operations.", "review_text": "While I found the paper well written, I didn't find it very impactful. The authors are not proposing a novel technique for addressing the problem, but only report some experiments with different encodings for the matrices and a standard transformer architecture. I was hoping to get more insights after reading this article, such as what architectural choices are beneficial for each specific problem and why. Instead, the proposed approach is training an off-the-shelf model with randomly distributed examples, and the experiments do not consider alternative techniques.\n\nIn my opinion, the motivation for this approach is also lacking, since the reported experiments only consider very small problems that can be solved exactly. I wish that the paper considered instead cases that need to be approximated, or possibly prove that indeed transformers trained on smaller problems can generalise to much higher dimensions. Since the learning algorithm has access to an oracle (numpy) that can provide exact supervision, an interesting problem is how to select the most informative input instances to be solved during training. Another aspect that is worth investigating in my opinion is how to cope with noise in the training data, allowing the training of transformers with an approximate oracle.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper consider the problem of approximating algebraic computations over matrices using transformers.  \nExperiments with different encodings are presented, investigating the use of transformers for approximating a number of algebraic operations.", "main_review": "While I found the paper well written, I didn't find it very impactful. The authors are not proposing a novel technique for addressing the problem, but only report some experiments with different encodings for the matrices and a standard transformer architecture. I was hoping to get more insights after reading this article, such as what architectural choices are beneficial for each specific problem and why. Instead, the proposed approach is training an off-the-shelf model with randomly distributed examples, and the experiments do not consider alternative techniques.\n\nIn my opinion, the motivation for this approach is also lacking, since the reported experiments only consider very small problems that can be solved exactly. I wish that the paper considered instead cases that need to be approximated, or possibly prove that indeed transformers trained on smaller problems can generalise to much higher dimensions. Since the learning algorithm has access to an oracle (numpy) that can provide exact supervision, an interesting problem is how to select the most informative input instances to be solved during training. Another aspect that is worth investigating in my opinion is how to cope with noise in the training data, allowing the training of transformers with an approximate oracle.", "summary_of_the_review": "- There isn't a novel contribution besides some experiments with an off-the-shelf model\n- The motivation is not supported by the experiments, which only report results in low-dimensional settings\n- The experiments do not consider any other technique besides transformers ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635333870584}, {"id": "Fk-6r7iVQFL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper866/Reviewer_LVVb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors train generic, dense transformers to perform several standard linear algebra computations, ranging from simple tasks like transposition to complex nonlinear tasks such as matrix inversion. They restrict themselves to relatively small matrices due to the practical limits of the dense, quadratic attention mechanism. The main result of the paper is that transformers can perform fairly well on all tasks, meaning that they can usually produce outputs that are correct upto relatively small tolerance. The paper also shows that some forms of out-of-distribution generalization are possible, and that this phenomenon is sensitive to the details of the training distribution.", "review_text": "I found this to be an interesting paper overall.\n\n- Framing\n\nI found the following claim to be problematic: 'Our results on out-of-distribution generalization provide justification to the idea that models trained over random data can be used to solve \"real world\" problems'. First, the authors only evaluate on other synthetic distributions. Second, in most \"real world\" problems, matrices are gigantic relative to the tiny context windows of dense transformers. Third, since traditional methods are always perfectly accurate on all distributions, I think the claim carries with it some burden to elaborate on why such noisy (and much less scalable) methods might prove useful in practice. Note that even if the potential practicality cannot be argued for, I think the experiments are interesting enough to stand on their own. \n\n- Sparse data reporting\n\nI would have liked to see much more data collected from the experiments, especially train-loss, validation-loss, and correctness-upto-tolerance curves over a range of architectures. The curves would also make it clear how many samples had been trained on for each measurement, which would be useful for understanding the relative performance of the different encodings. Note: it is not always clear from the current tables which encoding is even being used. It would also be interesting to see some analysis/visualization of the attention patterns, at least for tasks with relatively simple ground-truth algorithms like transposition and addition. Some experimental results also seem to be omitted; for example, S4.3 claims that \"deeper decoders are needed\" for matrix-matrix multiplication, but Table 5 does not include enough data to defend this claim.\n\n- Out of distribution findings seem unsurprising\n\nIt is great that the authors assess out-of-distribution, and I appreciate the negative result of generalizing from Wigner to matrices with positive eigenvalues. However, although the details of the subsequent out-of-distribution experiments are interesting, I found it generally unsurprising that models trained on non-Wigner matrices would generalize better, and I thought that the authors tried to make too big a point of this finding.\n\n- Co-training?\n\nAlthough not essential, I would be interested to see how co-training on many of the tasks at once affects sample efficiency.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors train generic, dense transformers to perform several standard linear algebra computations, ranging from simple tasks like transposition to complex nonlinear tasks such as matrix inversion. They restrict themselves to relatively small matrices due to the practical limits of the dense, quadratic attention mechanism. The main result of the paper is that transformers can perform fairly well on all tasks, meaning that they can usually produce outputs that are correct upto relatively small tolerance. The paper also shows that some forms of out-of-distribution generalization are possible, and that this phenomenon is sensitive to the details of the training distribution.", "main_review": "I found this to be an interesting paper overall.\n\n- Framing\n\nI found the following claim to be problematic: 'Our results on out-of-distribution generalization provide justification to the idea that models trained over random data can be used to solve \"real world\" problems'. First, the authors only evaluate on other synthetic distributions. Second, in most \"real world\" problems, matrices are gigantic relative to the tiny context windows of dense transformers. Third, since traditional methods are always perfectly accurate on all distributions, I think the claim carries with it some burden to elaborate on why such noisy (and much less scalable) methods might prove useful in practice. Note that even if the potential practicality cannot be argued for, I think the experiments are interesting enough to stand on their own. \n\n- Sparse data reporting\n\nI would have liked to see much more data collected from the experiments, especially train-loss, validation-loss, and correctness-upto-tolerance curves over a range of architectures. The curves would also make it clear how many samples had been trained on for each measurement, which would be useful for understanding the relative performance of the different encodings. Note: it is not always clear from the current tables which encoding is even being used. It would also be interesting to see some analysis/visualization of the attention patterns, at least for tasks with relatively simple ground-truth algorithms like transposition and addition. Some experimental results also seem to be omitted; for example, S4.3 claims that \"deeper decoders are needed\" for matrix-matrix multiplication, but Table 5 does not include enough data to defend this claim.\n\n- Out of distribution findings seem unsurprising\n\nIt is great that the authors assess out-of-distribution, and I appreciate the negative result of generalizing from Wigner to matrices with positive eigenvalues. However, although the details of the subsequent out-of-distribution experiments are interesting, I found it generally unsurprising that models trained on non-Wigner matrices would generalize better, and I thought that the authors tried to make too big a point of this finding.\n\n- Co-training?\n\nAlthough not essential, I would be interested to see how co-training on many of the tasks at once affects sample efficiency.\n", "summary_of_the_review": "Despite the concerns listed above, I think this paper does contribute to our emerging understanding of transformers, and that many people in the community will find it worth engaging with.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635169044614}, {"id": "0c1wtErzVGr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper866/Reviewer_Jby6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper describes several experiments where transformers are trained to perform real-valued linear algebra calculations (matrix transposition, addition, multiplication, eigenvalues & eigenvectors of symmetric matrices, SVD, inversion). In-distribution accuracy is generally very high, whereas care is needed in order to obtain out-of-domain generalization.", "review_text": "The paper carries out a very thorough set of experiments on linear algebra calculations with transformers, using four different encodings of input matrices. In addition, the authors are aware of the importance of out-of-distribution generalization, varying both the matrix size and the distribution of input matrices of a given size. Results appear to be complete, and the conclusion drawn from them are generally sound.\n\nHowever, the problem tackled in this paper does not appear to be particularly useful. In my opinion, the conclusions and findings of this paper are only interesting on a theoretical level (perhaps they can help understand what transformers can or cannot do), rather than being directly applicable in a meaningful way. After all, we do have algorithms for all linear algebra problems considered, and they work with 100% accuracy, perfect out-of-domain generalization, and faster run time.\n\nAs the authors note in the discussion, at the current stage transformers have quadratic complexity in the number of tokens, which translates into $O(n^4)$ complexity for $n \\times n$ input matrices, and this is asymptotically slower than the exact algorithms we have. A potentially interesting future direction (which perhaps can be advertised more by the author, to strengthen the claim that this paper is useful) is to investigate linear-time transformers on tasks where the exact algorithm requires more than $O(n^2)$ time, so that perhaps transformers can be used to perform approximate computations with less time.\n\nI also have the following minor comments.\n- First line of page 2: $m \\times n$ should be in a math formula.\n- Section 5, fifth line: what does “for small values of n” mean here? The statement is very precise, so it is hard to believe that it is true up to a certain small number (say, 5) and false for a larger n.\n- End of page 7, “This confirms that out-of-distribution generalization is possible when particular attention...”: actually, you showed that it is *necessary* to pay particular attention, not that it is *sufficient*. So I would write “out-of-distribution generalization requires particular attention...”.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper describes several experiments where transformers are trained to perform real-valued linear algebra calculations (matrix transposition, addition, multiplication, eigenvalues & eigenvectors of symmetric matrices, SVD, inversion). In-distribution accuracy is generally very high, whereas care is needed in order to obtain out-of-domain generalization.", "main_review": "The paper carries out a very thorough set of experiments on linear algebra calculations with transformers, using four different encodings of input matrices. In addition, the authors are aware of the importance of out-of-distribution generalization, varying both the matrix size and the distribution of input matrices of a given size. Results appear to be complete, and the conclusion drawn from them are generally sound.\n\nHowever, the problem tackled in this paper does not appear to be particularly useful. In my opinion, the conclusions and findings of this paper are only interesting on a theoretical level (perhaps they can help understand what transformers can or cannot do), rather than being directly applicable in a meaningful way. After all, we do have algorithms for all linear algebra problems considered, and they work with 100% accuracy, perfect out-of-domain generalization, and faster run time.\n\nAs the authors note in the discussion, at the current stage transformers have quadratic complexity in the number of tokens, which translates into $O(n^4)$ complexity for $n \\times n$ input matrices, and this is asymptotically slower than the exact algorithms we have. A potentially interesting future direction (which perhaps can be advertised more by the author, to strengthen the claim that this paper is useful) is to investigate linear-time transformers on tasks where the exact algorithm requires more than $O(n^2)$ time, so that perhaps transformers can be used to perform approximate computations with less time.\n\nI also have the following minor comments.\n- First line of page 2: $m \\times n$ should be in a math formula.\n- Section 5, fifth line: what does “for small values of n” mean here? The statement is very precise, so it is hard to believe that it is true up to a certain small number (say, 5) and false for a larger n.\n- End of page 7, “This confirms that out-of-distribution generalization is possible when particular attention...”: actually, you showed that it is *necessary* to pay particular attention, not that it is *sufficient*. So I would write “out-of-distribution generalization requires particular attention...”.", "summary_of_the_review": "This paper provides a thorough and well-written investigation of the use of transformers to perform linear algebra computation. However, this does not appear to be particularly useful.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634765596733}], "openreview_url": "https://openreview.net/forum?id=L2a_bcarHcF", "arxiv_id": "2112.01898", "paper_pdf": "papers/L2a_bcarHcF.pdf", "paper_pdf_sha256": "bc131c4a0592ef313bb305cb8dd0244d46446b6a1d5b740f7d6e058a905939af", "paper_pdf_bytes": 838859, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/LAWT", "code_repository": "facebookresearch/LAWT", "code_commit": "50bba0451c605b7112f914e827961300ad10b170", "code_archive": "repos/L2a_bcarHcF.zip", "code_archive_sha256": "bb45209ffe3b027cb67152262b4578d362bf8d38fe67a209dd94e7cff7152b3e", "code_archive_bytes": 63847, "code_file_count": 15, "code_extensions": {".py": 14, ".ipynb": 1}, "github_disk_usage_kb": 34, "github_languages": {"Python": 163819, "Jupyter Notebook": 12325}, "github_archived": true, "github_pushed_at": "2024-08-14T03:33:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/linear-algebra-with-transformers-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "muppfCkU9H1", "year": 2021, "status": "rejected", "title": "Multi-hop Attention Graph Neural Network", "authors": ["Guangtao Wang", "Zhitao Ying", "Jing Huang", "Jure Leskovec"], "authorids": ["~Guangtao_Wang1", "~Zhitao_Ying1", "~Jing_Huang3", "~Jure_Leskovec1"], "authors_source": "OpenReview API", "abstract": "Self-attention mechanism in graph neural networks (GNNs) led to state-of-the-art performance on many graph representation learning task. Currently, at every layer, attention is computed between connected pairs of nodes and depends solely on the representation of the two nodes. However, such attention mechanism does not account for nodes that are not directly connected but provide important network context, which could lead to improved predictive performance. Here we propose Multi-hop Attention Graph Neural Network (MAGNA), a principled way to incorporate multi-hop context information into attention computation, enabling long-range interactions at every layer of the GNN.  To compute attention between nodes that are not directly connected,\nMAGNA diffuses the attention scores across the network, which increases the ''receptive field'' for every layer of the GNN.\nUnlike previous approaches, MAGNA uses a diffusion prior on attention values, to efficiently account for all paths between the pair of disconnected nodes. This helps MAGNA capture large-scale structural information in every layer, and learn more informative attention. Experimental results on node classification as well as the knowledge graph completion benchmarks show that MAGNA achieves state-of-the-art results: MAGNA achieves up to 5.7% relative error reduction over the previous state-of-the-art on Cora, Citeseer, and Pubmed. MAGNA also obtains the best performance on a large-scale Open Graph Benchmark dataset. On knowledge graph completion  MAGNA advances state-of-the-art on WN18RR and FB15k-237 across four different performance metrics.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "sRe4AiG3c6l", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2024/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\n     Conventional Graph Neural Networks (GNNs) learn node representations that encode information from multiple hops away by iteratively aggregating information through their immediate neighbors. Self-Attention modules have been adopted to GNNs to selectively aggregate information coming through the immediate neighbors at different propagation stages. However, current self-attention mechanisms are limited to only attend over the nodes' immediate neighbors and not directly over their neighbors that are multiple hops away. Here in this work, the authors intend to address this issue and propose a means to obtain attention scores over indirectly connected neighbors. \n\n    The message passing paradigm is commonly adopted in GNNs because directly computing the higher powers of an adjacency matrix is not scalable. The same scalability concern is present for this work, which tries to obtain attention scores for indirect neighbors directly. Thus in order to solve this issue, the authors propose to diffuse the learned attention scores from their 1-hop neighbors to neighbors that are multiple hops away, thereby providing a means to directly obtain attention scores over indirect neighbors that are reachable from the nodes. \n\n——\nPros:\n\n\tThe paper is well written.\n\tThe paper provides experimental results for both homogenous and multi-relational graphs.\n\n——\nConcerns:\n\t(i) Proposed methodology being more powerful than GAT is arguable:\n\n\tWhen the attention scores for indirectly connected neighbors are still computed based on the immediate neighbors' attention scores, it is not convincing enough to be argued as more powerful than GAT, which learns attention scores over contextualized immediate neighbors.  Also, the approximate realization of the model described in Eqn: 5 follows a message-passing style to propagate attention scores. Suppose it is to be argued that standard message-passing-based diffusion is not powerful enough to get a good immediate neighbor representation that encodes neighbors' information from far away. In that case, it is not immediately clear how a similar diffusion, when used for propagating attention scores from immediate neighbors to neighbors multiple hops away, will be more powerful. \n\n(ii) Experimental results are not conclusive: \n\n\t\t(a) Effect of Layer-Norm and FeedForward \n\t\t\tOne of the important ablation model that is missing is MAGNA without feed-forward and layer-Norm components. Currently, it is not clear how much of an improvement is achieved because of these standard two components.  \n\t\t(b) Disentangling the effect of page-rank from the attention diffusion\n\t\t\tSince Diffusion-GCN is also based on Page-Rank based propagation, it would be helpful to compare with the Diffusion-GCN model with these two components appended to them, along with residual connection if already not present. This would help us clarify how much of the gain in performance depends on the page-rank-based propagation compared to the attention propagation. The teleport probability of Diffusion-GCN should also be similarly experimented with and the analysis should be compared with the plots in Figure 3. \n\n \t\t(c) Comparable or not significant gains achieved in Node classification tasks. \n\t\t\tIt is amendable that the authors have reported results for both single-relational and multi-relational graphs. However, the node classification results are not significantly better than GAT or Diffusion GCN on the reported smaller datasets (Table 1) with a single train/val/test split. And on OGB Arxiv dataset, GAT and Diffusion GCN numbers are not reported. Hence, it would be helpful to analyze additional datasets.\n  \t\t\t\t\tIgnoring the benefits of LayerNorm that the MAGNA can leverage, comparing its No-Feed-Fwd version with Diffusion-GCN, which is also based on a page-rank formulation, MAGNA gain ~1% improvement on Cora and Pubmed dataset whereas it falls behind by ~1% in Citeseer. \n\n\t\t(d) Disentangling the effect of Multi-scale diffusion from attention diffusion \n\t\t\tSince MAGNA uses a multi-scale diffusion at each layer, a comparison with a similar non-attentive multi-scale diffusion model like Lanczosnet that is also referred in the paper would be helpful to disentangle\tand understand the importance of the attention mechanism. \n\n\t\t(e) KG Completion: Missing baselines and model variations \n \n \t\t- Missing comparison with Self-attention (GAT) based knowledge graph embedding model, KBGAT. \n\t\t  Nathani et al., Learning Attention-based Embeddings for Relation Prediction in Knowledge Graphs, ACL 2019\n\t\t- Additionally, it would be helpful to have similar model ablation studies of MAGNA model as in Table 1 for KG Completion. \n\t\n\t\t(f) Depth Analysis:\n\t\t\t  Diffusion-GCN comparison missing. The performance stabilization over GAT might be arising because of the restart probability. MAGNA only has weights associated with 3 layers, unlike with GAT, which has weights for every propagation step. \n\t\t\t\n\n——\nQuestions during rebuttal:\n\n\t- Kindly clarify concern (i)\n\t- Check experimental concerns above for additional ablation and baseline variants that is required to disentangle and appreciate the usefulness of the primary contribution, the attention diffusion component. \n\t- Comparison with the KB-GAT model that is based on GAT for KG completion task, will strengthen the results on KG completion task.\n\t- Comparison with GAT and Diffusion-GCN with LayerNorm and FeedFwd components on multiple train/test/val splits for smaller datasets or for other datasets from OGB will strengthen the results for node classification. \n\n--- Post-rebuttal\nI thank the authors for responding to all the questions and getting back with additional experiment results.\n\nMajor concern: While I understand the motivation and how having attention scores over nodes multiple hops away can be powerful, I'm still not convinced with the approximate realization. It is not clear how diffusing attention defined over 1-hop neighbors is powerful over attention methods defined over immediate neighbors that contain k-hop information aggregated from diffusion.\n\nAlso, the performance drop and overfitting issue with GAT or diffusion-GCN can be combated similarly by sharing weights across GNN layers and also using a higher-order diffusion matrix at each GNN layer.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Important problem but the proposed solution may be not powerful than existing models ", "review": "Summary:\n\n     Conventional Graph Neural Networks (GNNs) learn node representations that encode information from multiple hops away by iteratively aggregating information through their immediate neighbors. Self-Attention modules have been adopted to GNNs to selectively aggregate information coming through the immediate neighbors at different propagation stages. However, current self-attention mechanisms are limited to only attend over the nodes' immediate neighbors and not directly over their neighbors that are multiple hops away. Here in this work, the authors intend to address this issue and propose a means to obtain attention scores over indirectly connected neighbors. \n\n    The message passing paradigm is commonly adopted in GNNs because directly computing the higher powers of an adjacency matrix is not scalable. The same scalability concern is present for this work, which tries to obtain attention scores for indirect neighbors directly. Thus in order to solve this issue, the authors propose to diffuse the learned attention scores from their 1-hop neighbors to neighbors that are multiple hops away, thereby providing a means to directly obtain attention scores over indirect neighbors that are reachable from the nodes. \n\n——\nPros:\n\n\tThe paper is well written.\n\tThe paper provides experimental results for both homogenous and multi-relational graphs.\n\n——\nConcerns:\n\t(i) Proposed methodology being more powerful than GAT is arguable:\n\n\tWhen the attention scores for indirectly connected neighbors are still computed based on the immediate neighbors' attention scores, it is not convincing enough to be argued as more powerful than GAT, which learns attention scores over contextualized immediate neighbors.  Also, the approximate realization of the model described in Eqn: 5 follows a message-passing style to propagate attention scores. Suppose it is to be argued that standard message-passing-based diffusion is not powerful enough to get a good immediate neighbor representation that encodes neighbors' information from far away. In that case, it is not immediately clear how a similar diffusion, when used for propagating attention scores from immediate neighbors to neighbors multiple hops away, will be more powerful. \n\n(ii) Experimental results are not conclusive: \n\n\t\t(a) Effect of Layer-Norm and FeedForward \n\t\t\tOne of the important ablation model that is missing is MAGNA without feed-forward and layer-Norm components. Currently, it is not clear how much of an improvement is achieved because of these standard two components.  \n\t\t(b) Disentangling the effect of page-rank from the attention diffusion\n\t\t\tSince Diffusion-GCN is also based on Page-Rank based propagation, it would be helpful to compare with the Diffusion-GCN model with these two components appended to them, along with residual connection if already not present. This would help us clarify how much of the gain in performance depends on the page-rank-based propagation compared to the attention propagation. The teleport probability of Diffusion-GCN should also be similarly experimented with and the analysis should be compared with the plots in Figure 3. \n\n \t\t(c) Comparable or not significant gains achieved in Node classification tasks. \n\t\t\tIt is amendable that the authors have reported results for both single-relational and multi-relational graphs. However, the node classification results are not significantly better than GAT or Diffusion GCN on the reported smaller datasets (Table 1) with a single train/val/test split. And on OGB Arxiv dataset, GAT and Diffusion GCN numbers are not reported. Hence, it would be helpful to analyze additional datasets.\n  \t\t\t\t\tIgnoring the benefits of LayerNorm that the MAGNA can leverage, comparing its No-Feed-Fwd version with Diffusion-GCN, which is also based on a page-rank formulation, MAGNA gain ~1% improvement on Cora and Pubmed dataset whereas it falls behind by ~1% in Citeseer. \n\n\t\t(d) Disentangling the effect of Multi-scale diffusion from attention diffusion \n\t\t\tSince MAGNA uses a multi-scale diffusion at each layer, a comparison with a similar non-attentive multi-scale diffusion model like Lanczosnet that is also referred in the paper would be helpful to disentangle\tand understand the importance of the attention mechanism. \n\n\t\t(e) KG Completion: Missing baselines and model variations \n \n \t\t- Missing comparison with Self-attention (GAT) based knowledge graph embedding model, KBGAT. \n\t\t  Nathani et al., Learning Attention-based Embeddings for Relation Prediction in Knowledge Graphs, ACL 2019\n\t\t- Additionally, it would be helpful to have similar model ablation studies of MAGNA model as in Table 1 for KG Completion. \n\t\n\t\t(f) Depth Analysis:\n\t\t\t  Diffusion-GCN comparison missing. The performance stabilization over GAT might be arising because of the restart probability. MAGNA only has weights associated with 3 layers, unlike with GAT, which has weights for every propagation step. \n\t\t\t\n\n——\nQuestions during rebuttal:\n\n\t- Kindly clarify concern (i)\n\t- Check experimental concerns above for additional ablation and baseline variants that is required to disentangle and appreciate the usefulness of the primary contribution, the attention diffusion component. \n\t- Comparison with the KB-GAT model that is based on GAT for KG completion task, will strengthen the results on KG completion task.\n\t- Comparison with GAT and Diffusion-GCN with LayerNorm and FeedFwd components on multiple train/test/val splits for smaller datasets or for other datasets from OGB will strengthen the results for node classification. \n\n--- Post-rebuttal\nI thank the authors for responding to all the questions and getting back with additional experiment results.\n\nMajor concern: While I understand the motivation and how having attention scores over nodes multiple hops away can be powerful, I'm still not convinced with the approximate realization. It is not clear how diffusing attention defined over 1-hop neighbors is powerful over attention methods defined over immediate neighbors that contain k-hop information aggregated from diffusion.\n\nAlso, the performance drop and overfitting issue with GAT or diffusion-GCN can be combated similarly by sharing weights across GNN layers and also using a higher-order diffusion matrix at each GNN layer.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603948315100}, {"id": "CdS_VzRVA3v", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2024/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper mainly proposes to learn attention-based edge coefficients by incorporating information from farther away nodes by means of their shortest path (powers of the adjacency matrix). Furthermore, the authors show the spectral properties of the proposed algorithm and its equivalence to personalized page rank.\n\nStrong points:\n\nThe numerical experiments show slight improvement.\n\nThe spectral analysis is potentially interesting.\n\nWeak points:\n\nThe multi-hop attention network has been done before (see below). The novelty thus resides only in the spectral analysis and the page rank equivalence.\n\nRecommendation:\n\nThe paper is ok, but I do not consider it to have enough novelty to be considered for publication in ICLR.\n\nMajor comment:\n\n1) Assuming there is only a scalar assigned to each edge (i.e. \\mathcal{R} = \\mathbb{R}), then (1),(3),(4) are a particular case of (39)-(41) in Isufi et al, 2020, where A_{k} = alpha I for k<K and A_{K} = H (the H matrix in eq. 4). This renders (1), (3), (4)'s only novelty to be the potential for vector-valued edge weights. Please elaborate on the comparison with Isufi et al, 2020 in the paper.\n\nE. Isufi, F. Gama, and A. Ribeiro, \"EdgeNets: Edge Varying Graph Neural Networks,\" arXiv:2001.07620v2 [cs.LG], 12 March 2020. [Online]. Available: http://arxiv.org/abs/2001.07620\n\n2) Why is R in N_r x d_r? Shouldn't it have N_e as a dimension? So if we have 5 edge types (like in the QM9 molecule data set, with 4 bond types and an extra no-bond type, encoded as one hot vectors), that means that R has only 5 rows? And what does the number of columns represent? It's just that X is very straightforward: number of nodes x number of channels. But R is hard to interpret.\n\n3) The LHS of eq. (2) shows dependence with i, j and l. The RHS shows dependence with i, j, l, and also k. Where does k appear in the matrix indexing? Is there a different attention score matrix for each value of k?\n\n4) I believe the paper would benefit for an increased elaboration on the usefulness of proposition 2, since this is the main novelty. Let us say we are given a graph, and we choose to describe that graph by means of some matrix (either the adjacency or the Laplacian). The choice of graph description fixes the frequency interpretation of the graph. Different matrix choices lead to different frequency interpretations. Now, the attention mechanism changes this matrix description by learning a new matrix description. However, this new, learned, matrix description will most likely not even share the same eigenvectors as the original matrix description of the graph. So how do we know that the filters learned are actually low-pass filters in the graph? They maybe low-pass filters in the learned graph by attention mechanisms, but they may not be in the actual graph that was given by the problem. Please, clarify what's the interpretation of Proposition 2, since matrices A and \\mathcal{A} are being learned from data.\n\nMinor comments:\n\nDue to Cayley-Hamilton theorem, there is no need for (3) to go to infinity. It suffices for it to go to N_n-1.\n\nReference to Sandryhaila and Moura, 2013 is missing at the beginning of Section 3 when referring to discrete signal processing on graphs.\n\nA. Sandryhaila and J. M. F. Moura, \"Discrete signal processing on graphs,\" IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644–1656, 1 Apr. 2013.\n\nFootnote 2 on page 5 is missing a proposition reference.\n\n--- UPDATED SCORE ---\n\nFirst of all, I would like to thank the authors for carefully addressing my comments.\n\nIn light of the authors' response, and after a thorough and careful discussion with the other reviewers, I have decided to update my score to 5 (five).\n\nIn summary, I appreciate the authors' effort to signal the differences between their work and Elvin et al. While I agree with these, I still think this is only an incremental contribution. For further reference, after carefully discussing the paper with other reviewers, these two published papers were also pointed out where multi-hop attention is addressed. Namely,\nhttps://openreview.net/forum?id=rkKvBAiiz\nhttps://ieeexplore.ieee.org/document/8683050\nI apologize for not finding these papers in my first round of reviews, but it does not really alter my evaluation of the paper.\n\nI think that the most novel contribution is on the spectral analysis. But this is only stated, with no real insights developed, and not emphasized enough. More insight on this would definitely bring a novelty. More specifically, novelties that would have made the paper more interesting: (i) a different way of computing the attention coefficients, that would be more parameter efficient (as opposed to eq. (1)) in the paper, (ii) actual useful insights into what the frequency response of the learned filters look like in the attention matrix as opposed to the given support matrix of the graph, (iii) a comparison between the spectral basis of the learned attention matrix as compared to the support matrix.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novelty in the Spectral Analysis", "review": "Summary:\n\nThis paper mainly proposes to learn attention-based edge coefficients by incorporating information from farther away nodes by means of their shortest path (powers of the adjacency matrix). Furthermore, the authors show the spectral properties of the proposed algorithm and its equivalence to personalized page rank.\n\nStrong points:\n\nThe numerical experiments show slight improvement.\n\nThe spectral analysis is potentially interesting.\n\nWeak points:\n\nThe multi-hop attention network has been done before (see below). The novelty thus resides only in the spectral analysis and the page rank equivalence.\n\nRecommendation:\n\nThe paper is ok, but I do not consider it to have enough novelty to be considered for publication in ICLR.\n\nMajor comment:\n\n1) Assuming there is only a scalar assigned to each edge (i.e. \\mathcal{R} = \\mathbb{R}), then (1),(3),(4) are a particular case of (39)-(41) in Isufi et al, 2020, where A_{k} = alpha I for k<K and A_{K} = H (the H matrix in eq. 4). This renders (1), (3), (4)'s only novelty to be the potential for vector-valued edge weights. Please elaborate on the comparison with Isufi et al, 2020 in the paper.\n\nE. Isufi, F. Gama, and A. Ribeiro, \"EdgeNets: Edge Varying Graph Neural Networks,\" arXiv:2001.07620v2 [cs.LG], 12 March 2020. [Online]. Available: http://arxiv.org/abs/2001.07620\n\n2) Why is R in N_r x d_r? Shouldn't it have N_e as a dimension? So if we have 5 edge types (like in the QM9 molecule data set, with 4 bond types and an extra no-bond type, encoded as one hot vectors), that means that R has only 5 rows? And what does the number of columns represent? It's just that X is very straightforward: number of nodes x number of channels. But R is hard to interpret.\n\n3) The LHS of eq. (2) shows dependence with i, j and l. The RHS shows dependence with i, j, l, and also k. Where does k appear in the matrix indexing? Is there a different attention score matrix for each value of k?\n\n4) I believe the paper would benefit for an increased elaboration on the usefulness of proposition 2, since this is the main novelty. Let us say we are given a graph, and we choose to describe that graph by means of some matrix (either the adjacency or the Laplacian). The choice of graph description fixes the frequency interpretation of the graph. Different matrix choices lead to different frequency interpretations. Now, the attention mechanism changes this matrix description by learning a new matrix description. However, this new, learned, matrix description will most likely not even share the same eigenvectors as the original matrix description of the graph. So how do we know that the filters learned are actually low-pass filters in the graph? They maybe low-pass filters in the learned graph by attention mechanisms, but they may not be in the actual graph that was given by the problem. Please, clarify what's the interpretation of Proposition 2, since matrices A and \\mathcal{A} are being learned from data.\n\nMinor comments:\n\nDue to Cayley-Hamilton theorem, there is no need for (3) to go to infinity. It suffices for it to go to N_n-1.\n\nReference to Sandryhaila and Moura, 2013 is missing at the beginning of Section 3 when referring to discrete signal processing on graphs.\n\nA. Sandryhaila and J. M. F. Moura, \"Discrete signal processing on graphs,\" IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644–1656, 1 Apr. 2013.\n\nFootnote 2 on page 5 is missing a proposition reference.\n\n--- UPDATED SCORE ---\n\nFirst of all, I would like to thank the authors for carefully addressing my comments.\n\nIn light of the authors' response, and after a thorough and careful discussion with the other reviewers, I have decided to update my score to 5 (five).\n\nIn summary, I appreciate the authors' effort to signal the differences between their work and Elvin et al. While I agree with these, I still think this is only an incremental contribution. For further reference, after carefully discussing the paper with other reviewers, these two published papers were also pointed out where multi-hop attention is addressed. Namely,\nhttps://openreview.net/forum?id=rkKvBAiiz\nhttps://ieeexplore.ieee.org/document/8683050\nI apologize for not finding these papers in my first round of reviews, but it does not really alter my evaluation of the paper.\n\nI think that the most novel contribution is on the spectral analysis. But this is only stated, with no real insights developed, and not emphasized enough. More insight on this would definitely bring a novelty. More specifically, novelties that would have made the paper more interesting: (i) a different way of computing the attention coefficients, that would be more parameter efficient (as opposed to eq. (1)) in the paper, (ii) actual useful insights into what the frequency response of the learned filters look like in the attention matrix as opposed to the given support matrix of the graph, (iii) a comparison between the spectral basis of the learned attention matrix as compared to the support matrix.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603737621704}, {"id": "XZh2oE3QW6y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2024/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a novel attention-based GNN called MAGNA. The main contribution consists in considerably increasing the receptive field by considering a multi-hop neighborhood instead of the standard one hop. The technical challenge consists in obtaining attention scores for all relevant nodes in an efficient way. MAGNA solves this by using a diffusion-based technique combined with a geometric distribution. The authors show that the latter further allows for approximations, and also give interesting theoretical insights (e.g., show a relation to page rank). \n\nThe paper is overall well written, related work is considered adequately, the idea is interesting, and the results are convincing. The evaluation comprises several different datasets, domains, tasks, and competitive baselines. The ablation studies give interesting insights in the effects of different parameter choices. Altogether, I suggest to accept the paper.\n\n----------------------------------------------\nSmaller comments:\n- p.3: \"degenerate categorical distribution with 1 category\": I think this should be explained.\n- p.3: \"effectively creating attention shortcuts between nodes that are not connected (Figure 1).\": I assume you mean \"directly connected\" since the whole approach seems to consider only connected nodes?\n- p.4: \"as well as good model generalization.\": How does the diffusion process ensure good model generalization?\n- p.5: Footnote 2: ??\n- 4.2. Baselines: The paragraph mentions few what is not in the table, so maybe you can just drop it and use the space for more descriptions.\n- 4.2. Results.: The last sentence is unclear to me.\n\n----------------------------------------------\nUpdate after Rebuttal: I have read the other reviews and authors' responses. \n\nWhile I do think the novelty of the contribution is sufficient, given that  the paper referenced in another review has not been peer-reviewed yet, the new ablation results in Table 1 show that the paper's contribution is not outstanding. I adjusted my score.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #4", "review": "The authors propose a novel attention-based GNN called MAGNA. The main contribution consists in considerably increasing the receptive field by considering a multi-hop neighborhood instead of the standard one hop. The technical challenge consists in obtaining attention scores for all relevant nodes in an efficient way. MAGNA solves this by using a diffusion-based technique combined with a geometric distribution. The authors show that the latter further allows for approximations, and also give interesting theoretical insights (e.g., show a relation to page rank). \n\nThe paper is overall well written, related work is considered adequately, the idea is interesting, and the results are convincing. The evaluation comprises several different datasets, domains, tasks, and competitive baselines. The ablation studies give interesting insights in the effects of different parameter choices. Altogether, I suggest to accept the paper.\n\n----------------------------------------------\nSmaller comments:\n- p.3: \"degenerate categorical distribution with 1 category\": I think this should be explained.\n- p.3: \"effectively creating attention shortcuts between nodes that are not connected (Figure 1).\": I assume you mean \"directly connected\" since the whole approach seems to consider only connected nodes?\n- p.4: \"as well as good model generalization.\": How does the diffusion process ensure good model generalization?\n- p.5: Footnote 2: ??\n- 4.2. Baselines: The paragraph mentions few what is not in the table, so maybe you can just drop it and use the space for more descriptions.\n- 4.2. Results.: The last sentence is unclear to me.\n\n----------------------------------------------\nUpdate after Rebuttal: I have read the other reviews and authors' responses. \n\nWhile I do think the novelty of the contribution is sufficient, given that  the paper referenced in another review has not been peer-reviewed yet, the new ablation results in Table 1 show that the paper's contribution is not outstanding. I adjusted my score.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603734579388}, {"id": "NeieF-6Ywo", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2024/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "==== Summary ====\nThis paper proposes MAGNA, a multi-hop self-attention mechanism for attention based graph neural networks. The proposed method increases the receptive field at each layer, requiring less layers to achieve a large receptive field. Also, with the proposed method the attention coefficient between two nodes is not just a function of the two nodes but also of their neighbourhood. The proposed MAGNA method is an extension of GAT networks that introduces a diffusion step on the computed attention coefficients, following a similar approach (Diffusion-GCNs) that has been used for GCNs.\n\n==== Pros: ====\n* The proposed method extends GAT layers to have multi-hop receptive fields without increasing the number of model parameters.\n* Even though the two main building blocks (Graph Attention and Graph Diffusion) of the paper are not novel, their combination is novel and it achieves state of the art empirical results in two different tasks.\n* The evaluation of MAGNA and comparison with previous approaches is well done, using two different tasks and the standard benchmarks for those tasks.\n* The paper is well written.\n\n\n#### Cons:\n* The proposed method MAGNA seems to be similar to the APPNP method proposed in [Klicpera 2019a] that also uses diffusion to increase the neighbourhood around each node in a GCN layer (without attention). However the two models are not compared and the later is not included in the Related Work section, even though it is cited previously.\n* The ablation study is a good idea, but it isn’t very clear how it is done.\n* The motivation behind including layer normalization and deep aggregation or why they are useful isn’t entirely clear.\n\n#### Questions:\n1) MAGNA has 3 main differences compared to GAT: layer normalization, diffusion and deep aggregation. In the ablation study in Table 1, which components are removed for each row? When it says “No LayerNorm”, does that mean that it uses diffusion and deep aggregation but no LayerNorm? If so, is there an ablation test where MAGNA just uses the diffusion step but without the other two components? That would be very relevant, since it would be like a GAT network with a larger receptive field at each layer, and would show how much improvement the increased receptive field brings without the other components.\n2) Since the paper puts a lot of emphasis in the multi-hop capability of MAGNA, and related to my previous question, it would be quite interesting to compare MAGNA against a GAT network that has a larger receptive field, for example by using a multi-hop adjacency matrix computed with the diffusion process from [Klicpera 2019b] (called sparsified matrix in that paper) instead of the 1-hop adjacency matrix used in the original GAT paper. With this comparison we could see the benefit of using the diffusion of the attention values instead of just increasing the receptive field of GAT by allowing each node to pay attention to nodes up to K hopes away.\n3) In section 4.1 the authors say that “Since MAGNA computes many attention values, layer normalization is crucial in ensuring training stability”. Doesn’t MAGNA compute the same number of attention values as GAT and it then diffuses them? Any ideas why layer normalization is crucial when diffusion is used but not without it? (as seen in the ablation study, when MAGNA doesn’t use diffusion the scores are on par with GAT, which suggests that layer normalization is only useful with diffusion).\n\n#### Minor comments\nIn the Edge Attention Computation paragraph in section 2.2, the authors say “To compute representation of” and probably meant “To compute the representation of”.\n\n$W_o$ in Equation 6 is not explained in the text.\n\nAcross the paper and in the caption for Figure 2, the text talks about MAGNA blocks, but the text in the Figure says “DAGN” Block. Is that the same? Same in the 2nd plot in Figure 3, it says DAGN, should it be MAGNA?\n\n#### Reasons for score\nThe authors give a solid justification for using a diffusion process to increase the receptive field of GNNs with attention, and along with previous papers on graph diffusion provide a good motivation and explanation of their method. Also, the experiments section shows that MAGNA achieves state of the art results in two different tasks. Because of these reasons, I vote for accept. My only concern is that some components don’t have a clear justification besides the empirical improvement in performance (layer normalization and deep aggregation) and that comparisons with other techniques (APPNP and GAT with k-hop adjacency matrix) would make the experiments section stronger.\n\n----- Post Rebuttal Update -----\nAfter the author's rebuttal and the discussion with other reviewers, I have decided to lower my score to 6.\nThe reason is that during the review and discussion process it was pointed out that the main contribution of the paper is more incremental than novel, as both multi-hop GATs and diffusion for GNNs have been explored before in similar ways. Additionally, even though the authors provide some theoretical grounding for their work, some of it is a bit disconnected from the rest of the paper, like the relation with PageRank introduced in section 3.2, which is not referenced or discussed in any other section.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting combination of graph attention and graph diffusion", "review": "==== Summary ====\nThis paper proposes MAGNA, a multi-hop self-attention mechanism for attention based graph neural networks. The proposed method increases the receptive field at each layer, requiring less layers to achieve a large receptive field. Also, with the proposed method the attention coefficient between two nodes is not just a function of the two nodes but also of their neighbourhood. The proposed MAGNA method is an extension of GAT networks that introduces a diffusion step on the computed attention coefficients, following a similar approach (Diffusion-GCNs) that has been used for GCNs.\n\n==== Pros: ====\n* The proposed method extends GAT layers to have multi-hop receptive fields without increasing the number of model parameters.\n* Even though the two main building blocks (Graph Attention and Graph Diffusion) of the paper are not novel, their combination is novel and it achieves state of the art empirical results in two different tasks.\n* The evaluation of MAGNA and comparison with previous approaches is well done, using two different tasks and the standard benchmarks for those tasks.\n* The paper is well written.\n\n\n#### Cons:\n* The proposed method MAGNA seems to be similar to the APPNP method proposed in [Klicpera 2019a] that also uses diffusion to increase the neighbourhood around each node in a GCN layer (without attention). However the two models are not compared and the later is not included in the Related Work section, even though it is cited previously.\n* The ablation study is a good idea, but it isn’t very clear how it is done.\n* The motivation behind including layer normalization and deep aggregation or why they are useful isn’t entirely clear.\n\n#### Questions:\n1) MAGNA has 3 main differences compared to GAT: layer normalization, diffusion and deep aggregation. In the ablation study in Table 1, which components are removed for each row? When it says “No LayerNorm”, does that mean that it uses diffusion and deep aggregation but no LayerNorm? If so, is there an ablation test where MAGNA just uses the diffusion step but without the other two components? That would be very relevant, since it would be like a GAT network with a larger receptive field at each layer, and would show how much improvement the increased receptive field brings without the other components.\n2) Since the paper puts a lot of emphasis in the multi-hop capability of MAGNA, and related to my previous question, it would be quite interesting to compare MAGNA against a GAT network that has a larger receptive field, for example by using a multi-hop adjacency matrix computed with the diffusion process from [Klicpera 2019b] (called sparsified matrix in that paper) instead of the 1-hop adjacency matrix used in the original GAT paper. With this comparison we could see the benefit of using the diffusion of the attention values instead of just increasing the receptive field of GAT by allowing each node to pay attention to nodes up to K hopes away.\n3) In section 4.1 the authors say that “Since MAGNA computes many attention values, layer normalization is crucial in ensuring training stability”. Doesn’t MAGNA compute the same number of attention values as GAT and it then diffuses them? Any ideas why layer normalization is crucial when diffusion is used but not without it? (as seen in the ablation study, when MAGNA doesn’t use diffusion the scores are on par with GAT, which suggests that layer normalization is only useful with diffusion).\n\n#### Minor comments\nIn the Edge Attention Computation paragraph in section 2.2, the authors say “To compute representation of” and probably meant “To compute the representation of”.\n\n$W_o$ in Equation 6 is not explained in the text.\n\nAcross the paper and in the caption for Figure 2, the text talks about MAGNA blocks, but the text in the Figure says “DAGN” Block. Is that the same? Same in the 2nd plot in Figure 3, it says DAGN, should it be MAGNA?\n\n#### Reasons for score\nThe authors give a solid justification for using a diffusion process to increase the receptive field of GNNs with attention, and along with previous papers on graph diffusion provide a good motivation and explanation of their method. Also, the experiments section shows that MAGNA achieves state of the art results in two different tasks. Because of these reasons, I vote for accept. My only concern is that some components don’t have a clear justification besides the empirical improvement in performance (layer normalization and deep aggregation) and that comparisons with other techniques (APPNP and GAT with k-hop adjacency matrix) would make the experiments section stronger.\n\n----- Post Rebuttal Update -----\nAfter the author's rebuttal and the discussion with other reviewers, I have decided to lower my score to 6.\nThe reason is that during the review and discussion process it was pointed out that the main contribution of the paper is more incremental than novel, as both multi-hop GATs and diffusion for GNNs have been explored before in similar ways. Additionally, even though the authors provide some theoretical grounding for their work, some of it is a bit disconnected from the rest of the paper, like the relation with PageRank introduced in section 3.2, which is not referenced or discussed in any other section.\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603479542713}], "openreview_url": "https://openreview.net/forum?id=muppfCkU9H1", "arxiv_id": "2009.14332", "paper_pdf": "papers/muppfCkU9H1.pdf", "paper_pdf_sha256": "8750d1d4a7c780f108c6c741cb7084932cfb180cc5192f9f7bb656042f90cc8c", "paper_pdf_bytes": 424519, "paper_pdf_source": "openreview", "code_url": "https://github.com/xjtuwgt/GNN-MAGNA", "code_repository": "xjtuwgt/GNN-MAGNA", "code_commit": "b37162830b5fe2582316390ca7853b4f30bbd129", "code_archive": "repos/muppfCkU9H1.zip", "code_archive_sha256": "065b7da814b19e8e98715c85be45b6b5a30d5a49fdab89d2f7bae2c6c5c594b5", "code_archive_bytes": 57644, "code_file_count": 28, "code_extensions": {".py": 28}, "github_disk_usage_kb": 40, "github_languages": {"Python": 175959}, "github_archived": false, "github_pushed_at": "2021-08-25T19:14:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/direct-multi-hop-attention-based-graph-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkltE0VKwH", "year": 2020, "status": "rejected", "title": "Coordinated Exploration via Intrinsic Rewards for Multi-Agent Reinforcement Learning", "authors": ["Shariq Iqbal", "Fei Sha"], "authorids": ["shariqiqbal2810@gmail.com", "fsha@google.com"], "authors_source": "OpenReview API", "abstract": "Solving tasks with sparse rewards is one of the most important challenges in reinforcement learning. In the single-agent setting, this challenge has been addressed by introducing intrinsic rewards that motivate agents to explore unseen regions of their state spaces. Applying these techniques naively to the multi-agent setting results in agents exploring independently, without any coordination among themselves. We argue that learning in cooperative multi-agent settings can be accelerated and improved if agents coordinate with respect to what they have explored. In this paper we propose an approach for learning how to dynamically select between different types of intrinsic rewards which consider not just what an individual agent has explored, but all agents, such that the agents can coordinate their exploration and maximize extrinsic returns. Concretely, we formulate the approach as a hierarchical policy where a high-level controller selects among sets of policies trained on different types of intrinsic rewards and the low-level controllers learn the action policies of all agents under these specific rewards. We demonstrate the effectiveness of the proposed approach in a multi-agent gridworld domain with sparse rewards, and then show that our method scales up to more complex settings by evaluating on the VizDoom platform.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkgAf17b9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1082/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe paper proposes a method for coordinating the exploration efforts of agents in a multi-agent reinforcement learning setting. The approach has two main components: (i) learning different exploration policies using different \"joint\" intrinsic rewards; and (ii) learning a higher-level policy that selects one of the exploration policies to be executed at the beginning of each episode.\n\nEach agent has its own novelty function which quantifies the novelty of observation seen by that agent. To coordinate exploration, these novelty functions are combined using aggregation functions to produce intrinsic reward for the agent. Each such aggregating function yields a different intrinsic reward. The authors propose several such aggregating functions as examples, however the method is applicable to other aggregating functions as well, as long as they can be computed off-policy.\n\nDuring training, the higher level policy selects one of the exploration policies which is then executed for the entire episode. The episode data is used in two ways: (i) to train the higher-level policy using policy gradients for maximizing extrinsic rewards along with an entropy term; and (ii) to train each exploration policy using soft actor-critic on its own intrinsic reward function (and extrinsic reward) in an off-policy manner.\n\nExperiments done on grid-world and VizDoom environment for three different tasks demonstrate that, on most tasks, the proposed approach performs at least as well as separately trained individual intrinsic rewards. Further ablation studies confirm that both the hierarchical setup and the \"joint\" intrinsic rewards are useful.\n\n\nQuestions to the Authors:\n\n1. The second sentence in section 5 is not clear - \"Furthermore, the type of reward ... sufficiently complex\". The high-level policy selects an exploration strategy at the beginning of each episode and then sticks to it for the entire duration of the episode. Changing the exploration strategy over the course of training might be useful in cases when agent needs to switch to a different exploration strategy after reaching a particular bottleneck state. However, this would require the exploration strategy to be changed in the middle of an episode which is not supported. Could you give an example where the exploration strategy must be changed over time even if one only selects the strategy at the beginning of each episode? Also, why not select the exploration strategy after every fixed number of time steps within each episode (by making high-level policy a function of the current state)? \n\n2. Analyzing the role of high-level policy and its evolution over time on different tasks would be a very nice addition to the paper. Qualitative experiments demonstrating that it provides a curriculum which helps the agents in surpassing the performance of individual intrinsic rewards would be helpful.\n\n3. Should \\Pi in (10) also depend on i?\n\nThough paper is reasonably well written I find the contributions are very marginal. If authors can position the paper well with the existing literature and bring out the impact of the contributions it will be helpful. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "Summary:\nThe paper proposes a method for coordinating the exploration efforts of agents in a multi-agent reinforcement learning setting. The approach has two main components: (i) learning different exploration policies using different \"joint\" intrinsic rewards; and (ii) learning a higher-level policy that selects one of the exploration policies to be executed at the beginning of each episode.\n\nEach agent has its own novelty function which quantifies the novelty of observation seen by that agent. To coordinate exploration, these novelty functions are combined using aggregation functions to produce intrinsic reward for the agent. Each such aggregating function yields a different intrinsic reward. The authors propose several such aggregating functions as examples, however the method is applicable to other aggregating functions as well, as long as they can be computed off-policy.\n\nDuring training, the higher level policy selects one of the exploration policies which is then executed for the entire episode. The episode data is used in two ways: (i) to train the higher-level policy using policy gradients for maximizing extrinsic rewards along with an entropy term; and (ii) to train each exploration policy using soft actor-critic on its own intrinsic reward function (and extrinsic reward) in an off-policy manner.\n\nExperiments done on grid-world and VizDoom environment for three different tasks demonstrate that, on most tasks, the proposed approach performs at least as well as separately trained individual intrinsic rewards. Further ablation studies confirm that both the hierarchical setup and the \"joint\" intrinsic rewards are useful.\n\n\nQuestions to the Authors:\n\n1. The second sentence in section 5 is not clear - \"Furthermore, the type of reward ... sufficiently complex\". The high-level policy selects an exploration strategy at the beginning of each episode and then sticks to it for the entire duration of the episode. Changing the exploration strategy over the course of training might be useful in cases when agent needs to switch to a different exploration strategy after reaching a particular bottleneck state. However, this would require the exploration strategy to be changed in the middle of an episode which is not supported. Could you give an example where the exploration strategy must be changed over time even if one only selects the strategy at the beginning of each episode? Also, why not select the exploration strategy after every fixed number of time steps within each episode (by making high-level policy a function of the current state)? \n\n2. Analyzing the role of high-level policy and its evolution over time on different tasks would be a very nice addition to the paper. Qualitative experiments demonstrating that it provides a curriculum which helps the agents in surpassing the performance of individual intrinsic rewards would be helpful.\n\n3. Should \\Pi in (10) also depend on i?\n\nThough paper is reasonably well written I find the contributions are very marginal. If authors can position the paper well with the existing literature and bring out the impact of the contributions it will be helpful. \n"}, "tcdate": 1572052758357}, {"id": "rJeoQlwTKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1082/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overall I like the approach in the paper. It proposes a nice 2 pronged method for exploiting exploration via intrinsic rewards for multi-agent systems. The parts that a bit lacking with the current version of the paper in this are the evaluation tasks are few and a bit simple and I think there needs to be more discussion on the \"coverage\" of the intrinsic reward types. Are the ones proposed motivated by the tasks in the paper or are they sufficient for tasks in general?  Last using a more recent novelty metric could allow the method to work on more interesting/complex tasks.\n\nMore detailed feedback:\n- It would be good to include more learning curves in the main text for the paper.\n- The fact that applying intrinsic motivation to multi-agent simulations seems like a natural idea would be to convert the problem to a \"single\" agent problem to compare against the \"normal\" application of intrinsic rewards. This might be another baseline to consider for comparison.\n- It says that all agents share the same replay buffer. Does this also imply that every agent is performing the same task there are just many agents? This does not make the problem very multi-agent with different goals. Would it affect the algorithm significantly to work on an environment where the agents have various types of goals?\n- As is noted in the text, this method appears to work well in the centralized training scheme that many have adopted recently. However, It makes me wonder if there is a way to employ these exploration schemes in a non-centralized training form. The ability to ask other agents in the world about there preferences and novelty of states appears to be a strong assumption, especially in a multi-agent robotics problem.\n- While the authors note that the intrinsic rewards used in this work are not comprehensive it would be good to note how comprehensive they are. Are there a few that were left out on purpose. Do the authours believe this set is sufficient. This statement makes it seem like the authors just tried a few options and found one that worked. It would be good to expand on this discussion more.\n- More detail for Figure 1 would be helpful to understand the overall network design. While that figure it helpful maybe it would be good to include a version that goes into detail for the 2 agent environment. Then a more compressed n agent version can also be shown.\n- The paper describes a policy selector that is a type of high-level policy for HRL. This design seems rather unique in that this part of the policy can optimizing for which intrinsic reward to toggle based on the extrinsic rewards observed. I like it. It is noted that entropy is important for this design. Can this be analyzed in an empirical way? Is this true for most environments/tasks?\n- Task 2 seems a bit contrived. Is there another instance of this type of task elsewhere in another paper? It would be better to use more standard tasks if they are available.\n- Before section 6.1 the paper is discussing rewards the are received. It would be good to more explicit about where these rewards are coming from. I think it is meant that these rewards are the extrinsic rewards but it does not say.\n- As noted just before section 6.1 it seems for the collection of tasks 1-3 it is already obvious what types of intrinsic rewards should be used. It would be good to include more tasks where this decision is less obvious.\n- Why are there \"black holes\" in the environment? Also if an agent steps into a black hole they are crushed never to be seen again. What you describe sounds more like a wormhole where one end is non-stationary... Also, can the agents detect the presence of a black hole in some way?\n- It appears the novel metric is count based. While this can work in practice it seems a rather simple metric. Is it possible to use something more like ICM or RND that was referenced in the paper? Especially for the VizDoom environment?\n- In table 2 where are some of the numbers bold? It would be good to include this information in the caption for the table.\n- I am not sure if the discussion on the behaviours the intrinsic reward functions result in are very surprising. Maybe there is a more interesting behaviour that results from the combination of two intrinsic rewards?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "Overall I like the approach in the paper. It proposes a nice 2 pronged method for exploiting exploration via intrinsic rewards for multi-agent systems. The parts that a bit lacking with the current version of the paper in this are the evaluation tasks are few and a bit simple and I think there needs to be more discussion on the \"coverage\" of the intrinsic reward types. Are the ones proposed motivated by the tasks in the paper or are they sufficient for tasks in general?  Last using a more recent novelty metric could allow the method to work on more interesting/complex tasks.\n\nMore detailed feedback:\n- It would be good to include more learning curves in the main text for the paper.\n- The fact that applying intrinsic motivation to multi-agent simulations seems like a natural idea would be to convert the problem to a \"single\" agent problem to compare against the \"normal\" application of intrinsic rewards. This might be another baseline to consider for comparison.\n- It says that all agents share the same replay buffer. Does this also imply that every agent is performing the same task there are just many agents? This does not make the problem very multi-agent with different goals. Would it affect the algorithm significantly to work on an environment where the agents have various types of goals?\n- As is noted in the text, this method appears to work well in the centralized training scheme that many have adopted recently. However, It makes me wonder if there is a way to employ these exploration schemes in a non-centralized training form. The ability to ask other agents in the world about there preferences and novelty of states appears to be a strong assumption, especially in a multi-agent robotics problem.\n- While the authors note that the intrinsic rewards used in this work are not comprehensive it would be good to note how comprehensive they are. Are there a few that were left out on purpose. Do the authours believe this set is sufficient. This statement makes it seem like the authors just tried a few options and found one that worked. It would be good to expand on this discussion more.\n- More detail for Figure 1 would be helpful to understand the overall network design. While that figure it helpful maybe it would be good to include a version that goes into detail for the 2 agent environment. Then a more compressed n agent version can also be shown.\n- The paper describes a policy selector that is a type of high-level policy for HRL. This design seems rather unique in that this part of the policy can optimizing for which intrinsic reward to toggle based on the extrinsic rewards observed. I like it. It is noted that entropy is important for this design. Can this be analyzed in an empirical way? Is this true for most environments/tasks?\n- Task 2 seems a bit contrived. Is there another instance of this type of task elsewhere in another paper? It would be better to use more standard tasks if they are available.\n- Before section 6.1 the paper is discussing rewards the are received. It would be good to more explicit about where these rewards are coming from. I think it is meant that these rewards are the extrinsic rewards but it does not say.\n- As noted just before section 6.1 it seems for the collection of tasks 1-3 it is already obvious what types of intrinsic rewards should be used. It would be good to include more tasks where this decision is less obvious.\n- Why are there \"black holes\" in the environment? Also if an agent steps into a black hole they are crushed never to be seen again. What you describe sounds more like a wormhole where one end is non-stationary... Also, can the agents detect the presence of a black hole in some way?\n- It appears the novel metric is count based. While this can work in practice it seems a rather simple metric. Is it possible to use something more like ICM or RND that was referenced in the paper? Especially for the VizDoom environment?\n- In table 2 where are some of the numbers bold? It would be good to include this information in the caption for the table.\n- I am not sure if the discussion on the behaviours the intrinsic reward functions result in are very surprising. Maybe there is a more interesting behaviour that results from the combination of two intrinsic rewards?\n"}, "tcdate": 1571807267492}, {"id": "H1xJLKdhKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1082/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Contribution:\n\nThe paper proposes to use a set of handcrafted intrinsic rewards that depend on the novelty of an observation as perceived by the rest of the other agents. For each pair of reward and agent, they learn a policy and a value through actor critic method, and then a meta-policy choses at the beginning of each episode which intrinsic rewards to use, meaning that the policy used by the agents corresponds to the one that maximizes the reward chosen.\n\n\n\nReview:\n\n\nThe major limitation of the paper in my opinion is the fact that the \"coordination\" that occurs here is only happening at training time, not at execution time. The agents eventually learn whatever trajectory they need to perform, and then proceed to do so without any interaction with the other agents. In a sense, they don't even learn to explore collaboratively. In other words, agents trained on task 1 in a given maze would not be able to solve task 2 on the same maze without essentially relearning everything from scratch.\nThe other corollary of the fact that each agent learns its own policy is that the number of agents is fixed at training time, preventing testing with a different number of agents, as sometimes done in the literature ([1] [2]).\n\nGiven this limitation the scope of the work basically reduces to the exploration of a fixed environment when the action space can be factored into different agents. This \"multi-agent\" formulation is presumably meant to break down the computational complexity of having a joint observation/action space. However, the experiments are conducted only with a very limited number of agents (only 2 in the non toy environment of vizdoom). This small scale doesn't, in my opinion, demonstrate the advantage of the decomposition of the MDP over say SOTA single-agent exploration methods applied to the cartesian product of all the agents action spaces (in vizdoom the paper considers only 3 actions, so with two agents it would amount to 9 actions, which is still very tractable). Once the trajectories of both agents are found, they can be distilled to each of them individually so that they only depend on the local observation.\n\n\nRegarding the experiments on the Vizdoom environment, it appears that the traditional evaluation setup [3] doesn't involve providing the global position (x,y) to the agents as part of the observations (they must be inferred from the visual feed), contrary to the experimental setup presented in this paper.\nIn my opinion, this weakens the claim that the method \"scales to more complex environments\" since providing the position essentially makes the environment similar to a grid-world (arguably the visual feed isn't even needed to solve the task.\n\n\nThe use of a dynamic policy selection is somewhat interesting, but would benefit better investigation. Firstly, it is not clear to me if all the selection of the policy to use during training affects all the trajectories of the batch, or if different episodes of the batch may have a different policy.\nSecondly, it seems that the setting is typically the one of a (non-stationary) bandit, since there is no state and the \"reward\" is the return obtained by the policy. Could you share the reason behind the choice of an actor-critic algorithm over classical bandit algorithms? One obvious advantage of the latter are provable regret bounds.\nIn all, the selection policy seems to be useful during training, since it sometimes yields better solutions than any of the individual reward schemes. It suggests that some form of curriculum over the rewards is occurring during training, but if this is really what is going on, then it's possible that the relevant literature about curriculum learning may offer more stable and principled solutions than an actor critic, for example population based training. This could potentially solve the issues observed in task 2.\n\n\n[1] Relational Deep Reinforcement Learning, Zambaldi et al, https://arxiv.org/abs/1806.01830\n[2] A Structured Prediction Approach for Generalization in Cooperative Multi-Agent Reinforcement Learning, Carion et al, https://arxiv.org/abs/1910.08809\n[3] Curiosity-driven Exploration by Self-supervised Prediction, Pathak et al, ICML 2017\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "Contribution:\n\nThe paper proposes to use a set of handcrafted intrinsic rewards that depend on the novelty of an observation as perceived by the rest of the other agents. For each pair of reward and agent, they learn a policy and a value through actor critic method, and then a meta-policy choses at the beginning of each episode which intrinsic rewards to use, meaning that the policy used by the agents corresponds to the one that maximizes the reward chosen.\n\n\n\nReview:\n\n\nThe major limitation of the paper in my opinion is the fact that the \"coordination\" that occurs here is only happening at training time, not at execution time. The agents eventually learn whatever trajectory they need to perform, and then proceed to do so without any interaction with the other agents. In a sense, they don't even learn to explore collaboratively. In other words, agents trained on task 1 in a given maze would not be able to solve task 2 on the same maze without essentially relearning everything from scratch.\nThe other corollary of the fact that each agent learns its own policy is that the number of agents is fixed at training time, preventing testing with a different number of agents, as sometimes done in the literature ([1] [2]).\n\nGiven this limitation the scope of the work basically reduces to the exploration of a fixed environment when the action space can be factored into different agents. This \"multi-agent\" formulation is presumably meant to break down the computational complexity of having a joint observation/action space. However, the experiments are conducted only with a very limited number of agents (only 2 in the non toy environment of vizdoom). This small scale doesn't, in my opinion, demonstrate the advantage of the decomposition of the MDP over say SOTA single-agent exploration methods applied to the cartesian product of all the agents action spaces (in vizdoom the paper considers only 3 actions, so with two agents it would amount to 9 actions, which is still very tractable). Once the trajectories of both agents are found, they can be distilled to each of them individually so that they only depend on the local observation.\n\n\nRegarding the experiments on the Vizdoom environment, it appears that the traditional evaluation setup [3] doesn't involve providing the global position (x,y) to the agents as part of the observations (they must be inferred from the visual feed), contrary to the experimental setup presented in this paper.\nIn my opinion, this weakens the claim that the method \"scales to more complex environments\" since providing the position essentially makes the environment similar to a grid-world (arguably the visual feed isn't even needed to solve the task.\n\n\nThe use of a dynamic policy selection is somewhat interesting, but would benefit better investigation. Firstly, it is not clear to me if all the selection of the policy to use during training affects all the trajectories of the batch, or if different episodes of the batch may have a different policy.\nSecondly, it seems that the setting is typically the one of a (non-stationary) bandit, since there is no state and the \"reward\" is the return obtained by the policy. Could you share the reason behind the choice of an actor-critic algorithm over classical bandit algorithms? One obvious advantage of the latter are provable regret bounds.\nIn all, the selection policy seems to be useful during training, since it sometimes yields better solutions than any of the individual reward schemes. It suggests that some form of curriculum over the rewards is occurring during training, but if this is really what is going on, then it's possible that the relevant literature about curriculum learning may offer more stable and principled solutions than an actor critic, for example population based training. This could potentially solve the issues observed in task 2.\n\n\n[1] Relational Deep Reinforcement Learning, Zambaldi et al, https://arxiv.org/abs/1806.01830\n[2] A Structured Prediction Approach for Generalization in Cooperative Multi-Agent Reinforcement Learning, Carion et al, https://arxiv.org/abs/1910.08809\n[3] Curiosity-driven Exploration by Self-supervised Prediction, Pathak et al, ICML 2017\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571748167122}], "openreview_url": "https://openreview.net/forum?id=rkltE0VKwH", "arxiv_id": "1905.12127", "paper_pdf": "papers/rkltE0VKwH.pdf", "paper_pdf_sha256": "239cb76ff4f87fa4d4ff2c71ac36a2612331d8a5fc4c5becf1e51678cc45c4b9", "paper_pdf_bytes": 1209301, "paper_pdf_source": "openreview", "code_url": "https://github.com/shariqiqbal2810/Multi-Explore", "code_repository": "shariqiqbal2810/Multi-Explore", "code_commit": "116cfbec87f97c83dd16c271616287d0b1dcb0dd", "code_archive": "repos/rkltE0VKwH.zip", "code_archive_sha256": "9a03d94a0eab37a6fb2b293f529bc9486ec7c339cc299ad1e276770255758752", "code_archive_bytes": 61094, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 47, "github_languages": {"Python": 142794}, "github_archived": false, "github_pushed_at": "2021-05-22T18:45:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/coordinated-exploration-via-intrinsic-rewards"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "m3FOf6nKnU", "year": 2026, "status": "rejected", "title": "Forget Vectors at Play: Universal Input Perturbations Driving Machine Unlearning in Image Classification", "authors": ["Changchang Sun", "Ren Wang", "Yihua Zhang", "Jinghan Jia", "Jiancheng Liu", "Gaowen Liu", "Yan Yan", "Sijia Liu"], "authorids": ["~Changchang_Sun1", "~Ren_Wang1", "~Yihua_Zhang1", "~Jinghan_Jia1", "~Jiancheng_Liu2", "~Gaowen_Liu4", "~Yan_Yan6", "~Sijia_Liu1"], "authors_source": "OpenReview API", "abstract": "Machine unlearning (MU), which seeks to erase the influence of specific unwanted data from already-trained models, is becoming increasingly vital in model editing, particularly to comply with evolving data regulations like the \"right to be forgotten\". Conventional approaches are predominantly model-based, typically requiring retraining or fine-tuning the model's weights to meet unlearning requirements. In this work, we approach the MU problem from an input perturbation-based perspective, where the model weights remain intact throughout the unlearning process. We demonstrate the existence of a proactive input-based unlearning strategy, referred to forget vector, which can be generated as an input-agnostic data perturbation and remains as effective as model-based approximate unlearning approaches. We also explore forget vector arithmetic, whereby multiple class-specific forget vectors can be combined through simple operations (e.g., linear combinations) to generate new forget vectors for unseen unlearning tasks, such as forgetting arbitrary subsets across classes. Extensive experiments validate the effectiveness and adaptability of the forget vector, showcasing its competitive performance relative to state-of-the-art model-based methods while achieving superior parameter efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "nBQxrzhZaS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13522/Reviewer_MhaY"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper proposes a model-free, perturbation based machine unlearning that claims to achieve the unlearning of forget set based on the perturbing the data and keeping the model weights intact throughout the unlearning process. They propose the forget vectors that can be applied to the input forget data as an input-agnostic data perturbation and remains as effective as model-based approximate unlearning approaches.", "review_text": "The paper proposes a model-free, perturbation based machine unlearning that claims to achieve the unlearning of forget set based on the perturbing the data and keeping the model weights intact throughout the unlearning process. They propose the forget vectors that can be applied to the input forget data as an input-agnostic data perturbation and remains as effective as model-based approximate unlearning approaches.", "strengths": "The research questions that paper investigated are very interesting such as \"For instance, it remains unclear whether, current MU approaches generalize effectively to “shifted” forget data\".\n\nProposing the forget vectors that can demonstrate the direction of unlearning.", "weaknesses": "Since the model's parameters are unchanged and the information of forget set still encoded in the model, the unlearning has not take place and this method is not compatible with data privacy regulations.\n\nFrom the general perspective the idea can be interpreted as controlling the inputs to change model's output. This idea has been explored for explaining a black box prediction and how its prediction changes in the literature (model agnostic local explainers) but their proposed method won't change the model's parameters and this is in conflict with the idea of MU.\n\n\nLine 258 // $\\rightarrow$ \"(without unlearning)...\" $\\rightarrow$ a bit confused about the experimental setup:\n\nfor the experiment that is depicted in Figure 2, the authors applied different levels (methods) of perturbation on the whole data and passed it to the original model without unlearning and unlearned model using retraining. but the issues is that the original model performance downgraded on the test and remaining data but not on the forget set. If original model is not unlearned, how it's performance stayed the same on the forget set?", "questions": "Line 54 // $\\rightarrow$ \"MU design...\" $\\rightarrow$ How this method can comply with the GDPR and privacy regulations? Does it mean that the model can still remember the forget dataset ? wouldn't it oppose the purpose of MU?\n\n\nLine 91 $\\rightarrow$ \"eliminate the influence of...\" $\\rightarrow$ How the influence of specific data / subset of data is eliminated since the parameters has remained unchanged?\n\n\nLine 184 // $\\rightarrow$ \"to achieve unlearning...\" $\\rightarrow$ So in this case the inputs are perturbed to achieve the unlearning but the model remains the same. So in this case, I expect the paper to investigate that the model treats the original and perturbed inputs the same meaning the prediction for both of them are similar and the encoding and representation of the model for both are the same. Even slight introduction of noise and perturbing the input can influence the model's prediction. So I would like to ask the authors to conduct an experiment to see how close the hidden representations of these two samples are.\n\nline 196 // $\\rightarrow$ \"based on (2)...\" $\\rightarrow$ I expect the authors to also investigate the following question: 1. is the perturbed data treated similarily to the original data? If this is treated as a different data point then how we can claim the model is already unlearned because the model can not recognize the class, and the model's performance has remained the same.\n\n\nline 293 // $\\rightarrow$ \"this yields the full...\" $\\rightarrow$\n\nI like the idea of optimizing the perturbation, but still the model's parameters wouldn't change and the model is the same but the forget data is manipulated. THis question can be risen that if I pass another datapoint as forget sample, what would be the issue? \n\nI know the answer to this question might sound trivial but the issue is coming from the point that model's parameters are not changed.\n\n\nSince the forget vector can show a direction to change the model while preserving the performance on the retain set, have authors considered fine tuning or applying grad Ascent using the direction of forget set? what would happen if fine tune the model's paremeted based on forget vectors?\n\n\nThe experiments section > The most important question that I hope authors can shed some light on that is that, does the representation of the perturbed forget data is similar to the original data? or how far it would get?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a model-free, perturbation based machine unlearning that claims to achieve the unlearning of forget set based on the perturbing the data and keeping the model weights intact throughout the unlearning process. They propose the forget vectors that can be applied to the input forget data as an input-agnostic data perturbation and remains as effective as model-based approximate unlearning approaches.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The research questions that paper investigated are very interesting such as \"For instance, it remains unclear whether, current MU approaches generalize effectively to “shifted” forget data\".\n\nProposing the forget vectors that can demonstrate the direction of unlearning.", "weaknesses": "Since the model's parameters are unchanged and the information of forget set still encoded in the model, the unlearning has not take place and this method is not compatible with data privacy regulations.\n\nFrom the general perspective the idea can be interpreted as controlling the inputs to change model's output. This idea has been explored for explaining a black box prediction and how its prediction changes in the literature (model agnostic local explainers) but their proposed method won't change the model's parameters and this is in conflict with the idea of MU.\n\n\nLine 258 // $\\rightarrow$ \"(without unlearning)...\" $\\rightarrow$ a bit confused about the experimental setup:\n\nfor the experiment that is depicted in Figure 2, the authors applied different levels (methods) of perturbation on the whole data and passed it to the original model without unlearning and unlearned model using retraining. but the issues is that the original model performance downgraded on the test and remaining data but not on the forget set. If original model is not unlearned, how it's performance stayed the same on the forget set?", "questions": "Line 54 // $\\rightarrow$ \"MU design...\" $\\rightarrow$ How this method can comply with the GDPR and privacy regulations? Does it mean that the model can still remember the forget dataset ? wouldn't it oppose the purpose of MU?\n\n\nLine 91 $\\rightarrow$ \"eliminate the influence of...\" $\\rightarrow$ How the influence of specific data / subset of data is eliminated since the parameters has remained unchanged?\n\n\nLine 184 // $\\rightarrow$ \"to achieve unlearning...\" $\\rightarrow$ So in this case the inputs are perturbed to achieve the unlearning but the model remains the same. So in this case, I expect the paper to investigate that the model treats the original and perturbed inputs the same meaning the prediction for both of them are similar and the encoding and representation of the model for both are the same. Even slight introduction of noise and perturbing the input can influence the model's prediction. So I would like to ask the authors to conduct an experiment to see how close the hidden representations of these two samples are.\n\nline 196 // $\\rightarrow$ \"based on (2)...\" $\\rightarrow$ I expect the authors to also investigate the following question: 1. is the perturbed data treated similarily to the original data? If this is treated as a different data point then how we can claim the model is already unlearned because the model can not recognize the class, and the model's performance has remained the same.\n\n\nline 293 // $\\rightarrow$ \"this yields the full...\" $\\rightarrow$\n\nI like the idea of optimizing the perturbation, but still the model's parameters wouldn't change and the model is the same but the forget data is manipulated. THis question can be risen that if I pass another datapoint as forget sample, what would be the issue? \n\nI know the answer to this question might sound trivial but the issue is coming from the point that model's parameters are not changed.\n\n\nSince the forget vector can show a direction to change the model while preserving the performance on the retain set, have authors considered fine tuning or applying grad Ascent using the direction of forget set? what would happen if fine tune the model's paremeted based on forget vectors?\n\n\nThe experiments section > The most important question that I hope authors can shed some light on that is that, does the representation of the perturbed forget data is similar to the original data? or how far it would get?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761976367905}, {"id": "KXeWxL1gkc", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13522/Reviewer_gacL"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper focuses on machine unlearning, particularly on various aspects of the generalization. In the first part, the authors examine the efficacy of  a generic unlearning method in handling images with corruptions or perturbations. In the second part, the authors propose a method for the optimization of forget vectors (which are input perturbations), and these forget vectors can be combined for compositional unlearning on an unseen unlearning task.", "review_text": "This paper focuses on machine unlearning, particularly on various aspects of the generalization. In the first part, the authors examine the efficacy of  a generic unlearning method in handling images with corruptions or perturbations. In the second part, the authors propose a method for the optimization of forget vectors (which are input perturbations), and these forget vectors can be combined for compositional unlearning on an unseen unlearning task.", "strengths": "The premise of the paper is interesting, where input perturbations are optimized, and that the learnt forget vectors can be combined in a compositional manner to tackle unseen forgetting tasks.\n\nThe writing and presentation are quite clear.", "weaknesses": "I am not sure how the two parts of the paper are connected. The first part investigates corruptions and perturbations on a generic unlearning method, but it is not clear what learning points are pertinent in proposing the subsequent method.\n\n\nAlthough I like the idea of the compositional combination of forget vectors for unlearning, I find that the implementation itself is not too interesting. Specifically, there is another optimization step where the linear combination weights of the learned forget vectors are optimized. Thus, the “transfer” of forget vectors is not exactly a zero-shot transfer, since training is still required.\n\nFurthermore, since optimization is still required, can the “compositional” method really be claimed to be performed on an “unseen task”?\n\nIn the experiment results, the proposed method is compared to various works, but many of them are not the latest works in the field. A suggestion would be to compare against more recent works in the field, such as the following:\n\nLearning to Unlearn for Robust Machine Unlearning. ECCV 2024\n\nAdversarial Machine Unlearning. ICLR 2025\n\nAdversarial Mixup Unlearning. ICLR 2025\n\nDecoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks. CVPR 2025", "questions": "During the investigation of corruptions, there are actually many times and strengths of corruptions available. How did the authors decide which corruptions to use, and which intensities to use?\n\nCan the forget vectors be combined in a “zero-shot” manner? For instance, in the example given with the automobile and the bird, can they be combined with simple averaging to be used on airplanes, without optimization?\n\nIf a trained forget vector for a class is used for another class, is the accuracy still maintained? Furthermore, does it affect the GradCAM saliency maps?\n\nPlease see weaknesses for some other questions. \n\nI tentatively recommend a borderline reject score due to my concerns with the disconnected parts of the paper, as well as my concerns with how compositional the vectors actually are. Yet, I think this paper has an interesting premise. If my concerns are addressed, I am open to raising my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on machine unlearning, particularly on various aspects of the generalization. In the first part, the authors examine the efficacy of  a generic unlearning method in handling images with corruptions or perturbations. In the second part, the authors propose a method for the optimization of forget vectors (which are input perturbations), and these forget vectors can be combined for compositional unlearning on an unseen unlearning task.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The premise of the paper is interesting, where input perturbations are optimized, and that the learnt forget vectors can be combined in a compositional manner to tackle unseen forgetting tasks.\n\nThe writing and presentation are quite clear.", "weaknesses": "I am not sure how the two parts of the paper are connected. The first part investigates corruptions and perturbations on a generic unlearning method, but it is not clear what learning points are pertinent in proposing the subsequent method.\n\n\nAlthough I like the idea of the compositional combination of forget vectors for unlearning, I find that the implementation itself is not too interesting. Specifically, there is another optimization step where the linear combination weights of the learned forget vectors are optimized. Thus, the “transfer” of forget vectors is not exactly a zero-shot transfer, since training is still required.\n\nFurthermore, since optimization is still required, can the “compositional” method really be claimed to be performed on an “unseen task”?\n\nIn the experiment results, the proposed method is compared to various works, but many of them are not the latest works in the field. A suggestion would be to compare against more recent works in the field, such as the following:\n\nLearning to Unlearn for Robust Machine Unlearning. ECCV 2024\n\nAdversarial Machine Unlearning. ICLR 2025\n\nAdversarial Mixup Unlearning. ICLR 2025\n\nDecoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks. CVPR 2025", "questions": "During the investigation of corruptions, there are actually many times and strengths of corruptions available. How did the authors decide which corruptions to use, and which intensities to use?\n\nCan the forget vectors be combined in a “zero-shot” manner? For instance, in the example given with the automobile and the bird, can they be combined with simple averaging to be used on airplanes, without optimization?\n\nIf a trained forget vector for a class is used for another class, is the accuracy still maintained? Furthermore, does it affect the GradCAM saliency maps?\n\nPlease see weaknesses for some other questions. \n\nI tentatively recommend a borderline reject score due to my concerns with the disconnected parts of the paper, as well as my concerns with how compositional the vectors actually are. Yet, I think this paper has an interesting premise. If my concerns are addressed, I am open to raising my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761583555872}, {"id": "wCK1zPnNGV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13522/Reviewer_oFB6"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a machine unlearning method that introduces an augmentation vector to perturb input data, preventing a trained machine learning model from making correct predictions on the designated forget datasets. The augmentation vector is treated as a set of trainable variables, optimized through a combined forget and retain loss functions to balance unlearning performance and knowledge retention. Experimental results demonstrate that the proposed approach is effective for both class-level and random unlearning tasks. However, the method also leads to noticeable performance degradation on the retain and test datasets.", "review_text": "This paper proposes a machine unlearning method that introduces an augmentation vector to perturb input data, preventing a trained machine learning model from making correct predictions on the designated forget datasets. The augmentation vector is treated as a set of trainable variables, optimized through a combined forget and retain loss functions to balance unlearning performance and knowledge retention. Experimental results demonstrate that the proposed approach is effective for both class-level and random unlearning tasks. However, the method also leads to noticeable performance degradation on the retain and test datasets.", "strengths": "1. The paper is clearly written and presents a well-motivated discussion on how the proposed method differs from existing unlearning approaches.\n2. The experimental evaluation is comprehensive and includes standard benchmarks and metrics commonly used in the literature for assessing unlearning algorithms.", "weaknesses": "I have several concerns regarding the novelty of the proposed method and its impact on performance degradation over the retain datasets.\n\nAlthough the paper claims the approach to be data-based, the method essentially trains an operator to modify the input, which is conceptually similar to techniques explored in the visual prompting literature, as acknowledged in the background section. The primary difference lies in the definition of the encoding function $f$, where previous works employ linear or nonlinear projections, while this method defines $f(x)=x+w$. This formulation represents a relatively minor variation rather than a fundamentally new idea.\n\nThe experimental results indicate a noticeable degradation in performance on the retain and test datasets compared with other methods. This suggests the limitation of augmenting inputs via simple vector addition. It would strengthen the paper to include an analysis of the geometric properties of the learned augmentation vectors relative to the training data, to better illustrate how they contribute to domain shift.\n\nBTW: the green and red tags indicating the best and second-best performance in Table 1 appear to be mislabeled. E.g, RA on the random data forgetting task for ImageNet-10.", "questions": "Can the authors show what the vectors learned captured, is it a vector that move input representation to the representation area of retain data points (maybe a specific class)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a machine unlearning method that introduces an augmentation vector to perturb input data, preventing a trained machine learning model from making correct predictions on the designated forget datasets. The augmentation vector is treated as a set of trainable variables, optimized through a combined forget and retain loss functions to balance unlearning performance and knowledge retention. Experimental results demonstrate that the proposed approach is effective for both class-level and random unlearning tasks. However, the method also leads to noticeable performance degradation on the retain and test datasets.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper is clearly written and presents a well-motivated discussion on how the proposed method differs from existing unlearning approaches.\n2. The experimental evaluation is comprehensive and includes standard benchmarks and metrics commonly used in the literature for assessing unlearning algorithms.", "weaknesses": "I have several concerns regarding the novelty of the proposed method and its impact on performance degradation over the retain datasets.\n\nAlthough the paper claims the approach to be data-based, the method essentially trains an operator to modify the input, which is conceptually similar to techniques explored in the visual prompting literature, as acknowledged in the background section. The primary difference lies in the definition of the encoding function $f$, where previous works employ linear or nonlinear projections, while this method defines $f(x)=x+w$. This formulation represents a relatively minor variation rather than a fundamentally new idea.\n\nThe experimental results indicate a noticeable degradation in performance on the retain and test datasets compared with other methods. This suggests the limitation of augmenting inputs via simple vector addition. It would strengthen the paper to include an analysis of the geometric properties of the learned augmentation vectors relative to the training data, to better illustrate how they contribute to domain shift.\n\nBTW: the green and red tags indicating the best and second-best performance in Table 1 appear to be mislabeled. E.g, RA on the random data forgetting task for ImageNet-10.", "questions": "Can the authors show what the vectors learned captured, is it a vector that move input representation to the representation area of retain data points (maybe a specific class)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761577291919}, {"id": "kPiep0Lf2h", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13522/Reviewer_xw9H"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper reframes machine unlearning for image classification as a data-based operation. Instead of updating weights, the authors learn a universal, input-agnostic perturbation, “forget vector”, that is added to inputs to degrade predictions on the forget set while preserving utility on retain and test data. They formulate an optimization with an untargeted margin loss on forget set and a cross-entropy retain regularizer on retain set, plus an L2 norm penalty. The model itself stays frozen. They further propose compositional unlearning by linearly combining class-wise forget vectors to address random data deletion. Experiments on CIFAR-10 and ImageNet-10 with ResNet-18 / VGG-16 / ViT-Base compare against Retrain and several approximate MU baselines, reporting competitive unlearning performance.", "review_text": "This paper reframes machine unlearning for image classification as a data-based operation. Instead of updating weights, the authors learn a universal, input-agnostic perturbation, “forget vector”, that is added to inputs to degrade predictions on the forget set while preserving utility on retain and test data. They formulate an optimization with an untargeted margin loss on forget set and a cross-entropy retain regularizer on retain set, plus an L2 norm penalty. The model itself stays frozen. They further propose compositional unlearning by linearly combining class-wise forget vectors to address random data deletion. Experiments on CIFAR-10 and ImageNet-10 with ResNet-18 / VGG-16 / ViT-Base compare against Retrain and several approximate MU baselines, reporting competitive unlearning performance.", "strengths": "1. The paper is well-written and easy to follow.\n\n2. Achieving unlearning by keeping the model fixed and using a visual-prompting-like input perturbation is novel, offering a fresh perspective on approximate machine unlearning.\n\n3. The idea of compositional Forget Vector arithmetic is interesting.\n\n4. Grad-CAM visualizations provide intuitive evidence supporting the reported effects.", "weaknesses": "1. The proposed method that perturbs only the input via a Forget Vector by a linear operation in image input space has inherent limitations.\n\n    * Effectiveness: Although the paper achieves competitive unlearning accuracy, this comes at the cost of degraded RA and TA due to the added perturbation.\n\n    * Efficiency: According to Table A2 (RTE), the Forget Vector is not clearly superior to other model-based unlearning methods in runtime, and its performance does not significantly surpass them.\n\n    Therefore, for a model maintainer, there may be insufficient reason to prefer the Forget Vector approach as the unlearning mechanism. Exploring hybrid approaches that combine Forget Vectors with other unlearning methods could better demonstrate its value.\n\n2. The unlearning settings (only single-class forgetting and 10% random sample forgetting) considered are limited. A robust unlearning method should handle arbitrary numbers of classes and larger proportions of random samples. This may be particularly challenging for universal perturbations, as in the Forget Vector method.\n\n3. The application scope in this paper is narrow, focusing on image classification. Unlearning is also relevant in image generation models, large language models, and multimodal models. The transferability of this approach remains unknown. For example, in vision–language models, whether an image-side Forget Vector can achieve comparable forgetting effects is an important unresolved question.", "questions": "1. How does Forget Vector perform in the following scenarios aimed at addressing the aforementioned weaknesses: (i) its integration with other unlearning methods, (ii) evaluation under more challenging unlearning settings, and (iii) evaluation on vision–language models?\n\n2. In the experiments related to Figure 3, the combination coefficient $w_1=-1$ for Forget Vectors is negative. While this is plausible from an optimization perspective, it seems at odds with the intended semantics of a forget vector. Could you provide a clear explanation for this phenomenon?\n\n3. In Table 1, the green/red highlighting appears to be based on performance gaps, which might conflict with the up/down arrows (e.g., TA in the Random Data Forgetting, ImageNet-10, ViT-Base setting). It would help to clearly state the convention in the caption to avoid ambiguity.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper reframes machine unlearning for image classification as a data-based operation. Instead of updating weights, the authors learn a universal, input-agnostic perturbation, “forget vector”, that is added to inputs to degrade predictions on the forget set while preserving utility on retain and test data. They formulate an optimization with an untargeted margin loss on forget set and a cross-entropy retain regularizer on retain set, plus an L2 norm penalty. The model itself stays frozen. They further propose compositional unlearning by linearly combining class-wise forget vectors to address random data deletion. Experiments on CIFAR-10 and ImageNet-10 with ResNet-18 / VGG-16 / ViT-Base compare against Retrain and several approximate MU baselines, reporting competitive unlearning performance.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The paper is well-written and easy to follow.\n\n2. Achieving unlearning by keeping the model fixed and using a visual-prompting-like input perturbation is novel, offering a fresh perspective on approximate machine unlearning.\n\n3. The idea of compositional Forget Vector arithmetic is interesting.\n\n4. Grad-CAM visualizations provide intuitive evidence supporting the reported effects.", "weaknesses": "1. The proposed method that perturbs only the input via a Forget Vector by a linear operation in image input space has inherent limitations.\n\n    * Effectiveness: Although the paper achieves competitive unlearning accuracy, this comes at the cost of degraded RA and TA due to the added perturbation.\n\n    * Efficiency: According to Table A2 (RTE), the Forget Vector is not clearly superior to other model-based unlearning methods in runtime, and its performance does not significantly surpass them.\n\n    Therefore, for a model maintainer, there may be insufficient reason to prefer the Forget Vector approach as the unlearning mechanism. Exploring hybrid approaches that combine Forget Vectors with other unlearning methods could better demonstrate its value.\n\n2. The unlearning settings (only single-class forgetting and 10% random sample forgetting) considered are limited. A robust unlearning method should handle arbitrary numbers of classes and larger proportions of random samples. This may be particularly challenging for universal perturbations, as in the Forget Vector method.\n\n3. The application scope in this paper is narrow, focusing on image classification. Unlearning is also relevant in image generation models, large language models, and multimodal models. The transferability of this approach remains unknown. For example, in vision–language models, whether an image-side Forget Vector can achieve comparable forgetting effects is an important unresolved question.", "questions": "1. How does Forget Vector perform in the following scenarios aimed at addressing the aforementioned weaknesses: (i) its integration with other unlearning methods, (ii) evaluation under more challenging unlearning settings, and (iii) evaluation on vision–language models?\n\n2. In the experiments related to Figure 3, the combination coefficient $w_1=-1$ for Forget Vectors is negative. While this is plausible from an optimization perspective, it seems at odds with the intended semantics of a forget vector. Could you provide a clear explanation for this phenomenon?\n\n3. In Table 1, the green/red highlighting appears to be based on performance gaps, which might conflict with the up/down arrows (e.g., TA in the Random Data Forgetting, ImageNet-10, ViT-Base setting). It would help to clearly state the convention in the caption to avoid ambiguity.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761298613084}], "openreview_url": "https://openreview.net/forum?id=m3FOf6nKnU", "arxiv_id": "2412.16780", "paper_pdf": "papers/m3FOf6nKnU.pdf", "paper_pdf_sha256": "8f9b0ebb8cf819d23b4d76ca4f466a460c455820f7494cea8f2291e953246aea", "paper_pdf_bytes": 3305568, "paper_pdf_source": "openreview", "code_url": "https://github.com/Changchangsun/Forget-Vector", "code_repository": "Changchangsun/Forget-Vector", "code_commit": "6abd0eb5c3d22d6dfd530bdfdd4c65abc944121f", "code_archive": "repos/m3FOf6nKnU.zip", "code_archive_sha256": "547c90f7714d7f86aa518a608ff43f151c659eb4ab6cfdc5e75085138524d3f3", "code_archive_bytes": 170253, "code_file_count": 47, "code_extensions": {".py": 47}, "github_disk_usage_kb": 203, "github_languages": {"Python": 453103}, "github_archived": false, "github_pushed_at": "2026-01-06T00:46:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/forget-vectors-at-play-universal-input"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vJmpg0exYA", "year": 2025, "status": "rejected", "title": "DiscQuant: A Quantization Method for Neural Networks Inspired by Discrepancy Theory", "authors": ["Jerry Chee", "Arturs Backurs", "Rainie Heck", "Li Zhang", "Janardhan Kulkarni", "Thomas Rothvoss", "Sivakanth Gopi"], "authorids": ["~Jerry_Chee1", "~Arturs_Backurs1", "~Rainie_Heck1", "~Li_Zhang11", "~Janardhan_Kulkarni2", "~Thomas_Rothvoss1", "~Sivakanth_Gopi1"], "authors_source": "OpenReview API", "abstract": "Quantizing the weights of a neural network has two steps: (1) Finding a good low bit-complexity representation for weights (which we call the quantization grid) and (2) Rounding the original weights to values in the quantization grid. In this paper, we study the problem of rounding optimally given any quantization grid. The simplest and most commonly used way to round is Round-to-Nearest (RTN). By rounding in a data-dependent way instead, one can improve the quality of the quantized model significantly.\n\nWe study the rounding problem from the lens of \\emph{discrepancy theory}, which studies how well we can round a continuous solution to a discrete solution without affecting solution quality too much. We prove that given $m=poly(1/\\epsilon)$ samples from the data distribution, we can round all but $O(m)$ model weights such that the expected approximation error of the quantized model on the true data distribution is $\\le \\epsilon$ as long as the space of gradients of the original model is approximately low rank (which we empirically validate).\n\nOur proof, which is algorithmic, inspired a  simple and practical rounding algorithm called \\emph{DiscQuant}. In our experiments, we demonstrate that DiscQuant significantly improves over the prior state-of-the-art rounding method called GPTQ and the baseline RTN over a range of benchmarks on Phi3mini-3.8B and Llama3.1-8B. For example, rounding Phi3mini-3.8B to a fixed quantization grid with 3.25 bits per parameter using DiscQuant gets 64\\% accuracy on the GSM8k dataset, whereas GPTQ achieves 54\\% and RTN achieves 31\\% (the original model achieves 84\\%).", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "IyoxZlLW4z", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11524/Reviewer_TWTt"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new method of quantizing neural networks with inspiration from theory of discrepancies. Traditionally, quantization in neural networks is composed of two processes, namely defining a grid of quantization and rounding model weights to match that particular grid. Standard rounding procedures are typically RTN, alongside data-dependent methods, some of which include GPTQ; however, in this paper, the round the weights using the theory of discrepancies approach without resulting in an increase in the loss on unseen data.\n\nDiscQuant relies on a mathematical framework to guarantee low error through rounding of nearly all model weights based on a low-rank assumption of gradient covariance. The authors establish theoretical bounds that their method can be guaranteed to achieve the expected generalization error to be at most epsilon on the data distribution, conditioned on certain low-rank conditions being met in the gradient space. They use such theoretical results to design a practical rounding algorithm that rounds the model weights from such an optimizer in a way that minimizes a regularized objective function combining KL divergence and linear constraints; it thus preserves overall model performance.\n\nExtensive experiments on Phi-3-mini-3.8B and Meta-Llama-3.1-8B models across tasks and quantization levels indicate the superior performance of DiscQuant against RTN and GPTQ specifically at low bits. The authors are able to show that DiscQuant has gained in performance over benchmarks GSM8k, ARC Challenge, and PIQA, which then comes to expose its generalizability and robustness over any possible format of quantizations such as block scaling and incoherence processing.", "review_text": "This paper introduces a new method of quantizing neural networks with inspiration from theory of discrepancies. Traditionally, quantization in neural networks is composed of two processes, namely defining a grid of quantization and rounding model weights to match that particular grid. Standard rounding procedures are typically RTN, alongside data-dependent methods, some of which include GPTQ; however, in this paper, the round the weights using the theory of discrepancies approach without resulting in an increase in the loss on unseen data.\n\nDiscQuant relies on a mathematical framework to guarantee low error through rounding of nearly all model weights based on a low-rank assumption of gradient covariance. The authors establish theoretical bounds that their method can be guaranteed to achieve the expected generalization error to be at most epsilon on the data distribution, conditioned on certain low-rank conditions being met in the gradient space. They use such theoretical results to design a practical rounding algorithm that rounds the model weights from such an optimizer in a way that minimizes a regularized objective function combining KL divergence and linear constraints; it thus preserves overall model performance.\n\nExtensive experiments on Phi-3-mini-3.8B and Meta-Llama-3.1-8B models across tasks and quantization levels indicate the superior performance of DiscQuant against RTN and GPTQ specifically at low bits. The authors are able to show that DiscQuant has gained in performance over benchmarks GSM8k, ARC Challenge, and PIQA, which then comes to expose its generalizability and robustness over any possible format of quantizations such as block scaling and incoherence processing.", "strengths": "This paper contributes a new input to the field of neural network quantization by pursuing a novel angle inspired by discrepancy theory in tackling the rounding problem. Traditional methods of quantization lie on two strands: Either the quantization grid must be designed efficiently, or the standard rounding technique, Round-to-Nearest, must be used. In its work, this paper introduces discrepancy theory for optimizing the rounding step and shows it can get high performance with all weights rounded to almost all possible bits but with a rather small approximation error. This result is established in a theoretical framework that automatically ensures low error, such as in the low-rank covariance matrices of the gradients, and so on. Therefore, this problem formulates a relatively little-investigated area of quantization in a new way.\n\nThe quality of this paper is very good in terms of methodology, with sound theoretical underpinnings for the proposed DiscQuant method to work. The authors systematically derive bounds on generalization error, effectively coupling their method with gradient covariance properties empirically validated. This is further complemented by a robust experimental setup in which DiscQuant is tested in several neural architectures (Phi-3-mini-3.8B and Meta-Llama-3.1-8B) along with quantization formats (block scaling, incoherence processing) with clear evidence that it surpasses the state-of-the-art techniques already in place like GPTQ and RTN. The results are very diversified regarding the extended coverage of benchmarks and quantization levels and therefore effectively demonstrate the applicability of DiscQuant over a wide scope and strength.\n\nThe paper is very clear in both its structured presentation of the theoretical framework and the algorithmic details. The authors do an excellent job in delineating the motivation behind DiscQuant as well as the central ideas of discrepancy theory. Figures and tables are well incorporated to help understanding in clear ways of how DiscQuant differs from other methods. While portions of the math development are tough slogging, the authors offer explanations that make intuitive sense, such as in the interpretation of results within the context of quantization grids and convex polytopes, which makes the theoretical contributions accessible to those readers with a strong technical background.\n\nThis work can be the first to provide significant influence for further work in quantization research, especially for large language models, where post-training quantization has to be very efficient for deployment on memory-constrained devices. At the same time, it frames quantization as a discrepancy problem and finds a practical rounding algorithm that achieves high compression with low loss in accuracy; thus, we contribute to new approach affecting far beyond LLMs and ranging from model deployment over mobile and embedded environments. This work further opens a route of further research into discrepancy theory in neural network quantization, inviting work that itself may eventually lead to very efficient quantization techniques. For value along the dimension of presentation, the paper is a highly significant and valuable addition to the field, advancing our understanding of effective quantization for modern, large-scale models.", "weaknesses": "While this paper has meaningful contribution, there are areas where it can improve with theoretical explanations of the methods, experimental validations, and practical applicability.\n\nOne area for improvement concerns the theoretical justification of why the proposed method is better than other data-dependent rounding methods, like GPTQ, in settings where the gradient covariance is not strongly low-rank. The authors assume low-rank covariance to support the theoretical bounds for DiscQuant, though it remains unclear how the method would perform when such an assumption fails to hold. Discussing some cases where the low-rank assumption is violated and offering either theoretical insights into potential limitations or proposing ways to adapt DiscQuant for such cases would strengthen the contribution. This would be an opportunity to refer to alternate bounding techniques that would account for the high-rank scenarios, or comparisons with approaches based on other assumptions.\n\nThe experimental evaluation was exhaustively done for standard models, but some additional comparisons of alternative data-dependent rounding techniques that are also effective in PTQ contexts may be very helpful. For example, newer approaches like CDQuant or AdaQuant-all seek to minimize quantization error in a data-dependent optimization manner-would make good baselines. Including the experiments that were run in those methods would actually give more background about how the comparative performance of DiscQuant goes along with an outline of relative advantages. Experiments on further tasks beyond text generation and multiple-choice questions - such as real-time inference on mobile devices or edge computing environments - should further enable these results to generalize. Extensions in this direction will unveil the flexibility and possible trade-offs of DiscQuant across different practical applications.\n\nWhile clear on the whole, some aspects of the presentation of discrepancy theory in the paper could have been clearer for readers not familiar with discrepancy theory, such as explaining quantization in terms of how discrepancy theory lends itself to being quantized. As it stands now, it is not very intuitive especially to readers not familiar with convex polytope to connect the random walk the Lovett-Meka algorithm used and the rounding process in DiscQuant. The connection is necessary to be explained further in order to make the text more readable and understandable.\n\nIn conclusion, though DiscQuant strongly performs well on all benchmarks, it may extend its discussion regarding the practical impact of the deployment of DiscQuant in real-world applications. For example, since the approach is iterative, even the computational efficiency along with memory usage in comparison to lesser approaches such as RTN may be much beneficial. Examined in greater detail, the amount of increased computational expense in terms of overhead in the form of time or memory could shed light on the trade-offs between using DiscQuant and alternative approaches. Second, some comments on ease of implementation, especially in comparison to such widely used approaches as GPTQ, might be useful for practitioners testing its practical utility.\n\nIn summary, what the paper contributes is a good contribution, but it requires theoretical discussion on the possible limitations of the low-rank assumption, more experiments on a variety of baselines and tasks, clearer depiction of the role of discrepancy theory, and more practical insights into trading off deploying DiscQuant in a variety of real-world settings, thereby making the results of this paper more comprehensive, accessible, and applicable across a wider range of contexts.", "questions": "1. How does DiscQuant do if the gradient covariance does not have strong low rank structures? In other words, if the low rank structure of the gradient covariance does not hold, are there alternative strategies by which we could adapt or extend DiscQuant so that the discrepancies may continue to be effective? Perhaps by modulating the discrepancy-based constraints? I'd appreciate any insight into whether or not DiscQuant is extensible to other model architectures and distributions.\n\n2. Of relevance to GPTQ and RTN, it would be interesting to see other baselines of recent data-dependent rounding techniques, such as CDQuant and AdaQuant, which optimize the quantization error. This would give a more thorough view of the possibilities and compromises with DiscQuant. Could the authors include these new baselines in future work or some ideas on how DiscQuant would theoretically compare with them?\n\n3. The proposed method, DiscQuant, basically comprises an optimization loop. How do the computational and memory costs of DiscQuant compare to the alternatives, like RTN and GPTQ? Is the latency or memory overhead for quantization drastic? How would these effects lead to deployment challenges for applications that have specific real-time requirements or are large-scale models? A careful study of those practical tradeoffs would be useful for understanding how this approach is actually feasible in the production setting.\n\n4. This paper relies upon the discrepancy theory as a basis for its rounding strategy, but not all readers will be as familiar with discrepancy theory as the authors. Could the authors provide additional explanation for why discrepancy theory is especially well-suited to this problem? Moreover, explanations of the concepts random walk and convex polytope would be more transparent and therefore better for readers who are unaware of these topics if they were illustrated with an example, perhaps also simplified, to bridge the gulf between such a theoretical approach and practical application.\n\n5. In DiscQuant, the random walk approach by inspiration of the Lovett-Meka algorithm finds some feasible vertex in the polytope. Could the authors provide more intuition behind why this method is useful for quantization? For instance, does one actually need the random walk to get low generalization error, or might possibly much simpler methods for finding a vertex of the polytope work comparably? Such intuition would help clarify the design choices and could perhaps lead to avenues for simplifications.\n\n6. The paper employs KL divergence as a metric to be minimized to reduce the gap between the original and a quantized model. Would the authors consider alternative metrics, such as MSE or other activation-based losses, for further generalizing the properties of DiscQuant? The better understanding of loss formula interactions would allow practitioners to fine-tune this method for specific applications.\n\n7. Although DiscQuant is only tested on text-based tasks with Phi-3-mini and Meta-Llama models, the authors might highlight the scope of potential application of this work with CNNs or transformers for vision. Are there any architectural or task-specific restrictions such that the approach of DiscQuant needs to be adjusted? Further research in that would demonstrate the extensibility of DiscQuant and what needs to be changed to allow the approach for other applications.\n\n8. The authors assume in practice that the gradient space is low rank, an assumption they verify empirically with certain architectures. Do they have additional insight or data on how DiscQuant behaves with higher-rank gradient spaces over datasets or models? Elucidation of any limits or change in performance within such settings would help one understand whether the assumptions for the method are generalizable.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new method of quantizing neural networks with inspiration from theory of discrepancies. Traditionally, quantization in neural networks is composed of two processes, namely defining a grid of quantization and rounding model weights to match that particular grid. Standard rounding procedures are typically RTN, alongside data-dependent methods, some of which include GPTQ; however, in this paper, the round the weights using the theory of discrepancies approach without resulting in an increase in the loss on unseen data.\n\nDiscQuant relies on a mathematical framework to guarantee low error through rounding of nearly all model weights based on a low-rank assumption of gradient covariance. The authors establish theoretical bounds that their method can be guaranteed to achieve the expected generalization error to be at most epsilon on the data distribution, conditioned on certain low-rank conditions being met in the gradient space. They use such theoretical results to design a practical rounding algorithm that rounds the model weights from such an optimizer in a way that minimizes a regularized objective function combining KL divergence and linear constraints; it thus preserves overall model performance.\n\nExtensive experiments on Phi-3-mini-3.8B and Meta-Llama-3.1-8B models across tasks and quantization levels indicate the superior performance of DiscQuant against RTN and GPTQ specifically at low bits. The authors are able to show that DiscQuant has gained in performance over benchmarks GSM8k, ARC Challenge, and PIQA, which then comes to expose its generalizability and robustness over any possible format of quantizations such as block scaling and incoherence processing.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper contributes a new input to the field of neural network quantization by pursuing a novel angle inspired by discrepancy theory in tackling the rounding problem. Traditional methods of quantization lie on two strands: Either the quantization grid must be designed efficiently, or the standard rounding technique, Round-to-Nearest, must be used. In its work, this paper introduces discrepancy theory for optimizing the rounding step and shows it can get high performance with all weights rounded to almost all possible bits but with a rather small approximation error. This result is established in a theoretical framework that automatically ensures low error, such as in the low-rank covariance matrices of the gradients, and so on. Therefore, this problem formulates a relatively little-investigated area of quantization in a new way.\n\nThe quality of this paper is very good in terms of methodology, with sound theoretical underpinnings for the proposed DiscQuant method to work. The authors systematically derive bounds on generalization error, effectively coupling their method with gradient covariance properties empirically validated. This is further complemented by a robust experimental setup in which DiscQuant is tested in several neural architectures (Phi-3-mini-3.8B and Meta-Llama-3.1-8B) along with quantization formats (block scaling, incoherence processing) with clear evidence that it surpasses the state-of-the-art techniques already in place like GPTQ and RTN. The results are very diversified regarding the extended coverage of benchmarks and quantization levels and therefore effectively demonstrate the applicability of DiscQuant over a wide scope and strength.\n\nThe paper is very clear in both its structured presentation of the theoretical framework and the algorithmic details. The authors do an excellent job in delineating the motivation behind DiscQuant as well as the central ideas of discrepancy theory. Figures and tables are well incorporated to help understanding in clear ways of how DiscQuant differs from other methods. While portions of the math development are tough slogging, the authors offer explanations that make intuitive sense, such as in the interpretation of results within the context of quantization grids and convex polytopes, which makes the theoretical contributions accessible to those readers with a strong technical background.\n\nThis work can be the first to provide significant influence for further work in quantization research, especially for large language models, where post-training quantization has to be very efficient for deployment on memory-constrained devices. At the same time, it frames quantization as a discrepancy problem and finds a practical rounding algorithm that achieves high compression with low loss in accuracy; thus, we contribute to new approach affecting far beyond LLMs and ranging from model deployment over mobile and embedded environments. This work further opens a route of further research into discrepancy theory in neural network quantization, inviting work that itself may eventually lead to very efficient quantization techniques. For value along the dimension of presentation, the paper is a highly significant and valuable addition to the field, advancing our understanding of effective quantization for modern, large-scale models.", "weaknesses": "While this paper has meaningful contribution, there are areas where it can improve with theoretical explanations of the methods, experimental validations, and practical applicability.\n\nOne area for improvement concerns the theoretical justification of why the proposed method is better than other data-dependent rounding methods, like GPTQ, in settings where the gradient covariance is not strongly low-rank. The authors assume low-rank covariance to support the theoretical bounds for DiscQuant, though it remains unclear how the method would perform when such an assumption fails to hold. Discussing some cases where the low-rank assumption is violated and offering either theoretical insights into potential limitations or proposing ways to adapt DiscQuant for such cases would strengthen the contribution. This would be an opportunity to refer to alternate bounding techniques that would account for the high-rank scenarios, or comparisons with approaches based on other assumptions.\n\nThe experimental evaluation was exhaustively done for standard models, but some additional comparisons of alternative data-dependent rounding techniques that are also effective in PTQ contexts may be very helpful. For example, newer approaches like CDQuant or AdaQuant-all seek to minimize quantization error in a data-dependent optimization manner-would make good baselines. Including the experiments that were run in those methods would actually give more background about how the comparative performance of DiscQuant goes along with an outline of relative advantages. Experiments on further tasks beyond text generation and multiple-choice questions - such as real-time inference on mobile devices or edge computing environments - should further enable these results to generalize. Extensions in this direction will unveil the flexibility and possible trade-offs of DiscQuant across different practical applications.\n\nWhile clear on the whole, some aspects of the presentation of discrepancy theory in the paper could have been clearer for readers not familiar with discrepancy theory, such as explaining quantization in terms of how discrepancy theory lends itself to being quantized. As it stands now, it is not very intuitive especially to readers not familiar with convex polytope to connect the random walk the Lovett-Meka algorithm used and the rounding process in DiscQuant. The connection is necessary to be explained further in order to make the text more readable and understandable.\n\nIn conclusion, though DiscQuant strongly performs well on all benchmarks, it may extend its discussion regarding the practical impact of the deployment of DiscQuant in real-world applications. For example, since the approach is iterative, even the computational efficiency along with memory usage in comparison to lesser approaches such as RTN may be much beneficial. Examined in greater detail, the amount of increased computational expense in terms of overhead in the form of time or memory could shed light on the trade-offs between using DiscQuant and alternative approaches. Second, some comments on ease of implementation, especially in comparison to such widely used approaches as GPTQ, might be useful for practitioners testing its practical utility.\n\nIn summary, what the paper contributes is a good contribution, but it requires theoretical discussion on the possible limitations of the low-rank assumption, more experiments on a variety of baselines and tasks, clearer depiction of the role of discrepancy theory, and more practical insights into trading off deploying DiscQuant in a variety of real-world settings, thereby making the results of this paper more comprehensive, accessible, and applicable across a wider range of contexts.", "questions": "1. How does DiscQuant do if the gradient covariance does not have strong low rank structures? In other words, if the low rank structure of the gradient covariance does not hold, are there alternative strategies by which we could adapt or extend DiscQuant so that the discrepancies may continue to be effective? Perhaps by modulating the discrepancy-based constraints? I'd appreciate any insight into whether or not DiscQuant is extensible to other model architectures and distributions.\n\n2. Of relevance to GPTQ and RTN, it would be interesting to see other baselines of recent data-dependent rounding techniques, such as CDQuant and AdaQuant, which optimize the quantization error. This would give a more thorough view of the possibilities and compromises with DiscQuant. Could the authors include these new baselines in future work or some ideas on how DiscQuant would theoretically compare with them?\n\n3. The proposed method, DiscQuant, basically comprises an optimization loop. How do the computational and memory costs of DiscQuant compare to the alternatives, like RTN and GPTQ? Is the latency or memory overhead for quantization drastic? How would these effects lead to deployment challenges for applications that have specific real-time requirements or are large-scale models? A careful study of those practical tradeoffs would be useful for understanding how this approach is actually feasible in the production setting.\n\n4. This paper relies upon the discrepancy theory as a basis for its rounding strategy, but not all readers will be as familiar with discrepancy theory as the authors. Could the authors provide additional explanation for why discrepancy theory is especially well-suited to this problem? Moreover, explanations of the concepts random walk and convex polytope would be more transparent and therefore better for readers who are unaware of these topics if they were illustrated with an example, perhaps also simplified, to bridge the gulf between such a theoretical approach and practical application.\n\n5. In DiscQuant, the random walk approach by inspiration of the Lovett-Meka algorithm finds some feasible vertex in the polytope. Could the authors provide more intuition behind why this method is useful for quantization? For instance, does one actually need the random walk to get low generalization error, or might possibly much simpler methods for finding a vertex of the polytope work comparably? Such intuition would help clarify the design choices and could perhaps lead to avenues for simplifications.\n\n6. The paper employs KL divergence as a metric to be minimized to reduce the gap between the original and a quantized model. Would the authors consider alternative metrics, such as MSE or other activation-based losses, for further generalizing the properties of DiscQuant? The better understanding of loss formula interactions would allow practitioners to fine-tune this method for specific applications.\n\n7. Although DiscQuant is only tested on text-based tasks with Phi-3-mini and Meta-Llama models, the authors might highlight the scope of potential application of this work with CNNs or transformers for vision. Are there any architectural or task-specific restrictions such that the approach of DiscQuant needs to be adjusted? Further research in that would demonstrate the extensibility of DiscQuant and what needs to be changed to allow the approach for other applications.\n\n8. The authors assume in practice that the gradient space is low rank, an assumption they verify empirically with certain architectures. Do they have additional insight or data on how DiscQuant behaves with higher-rank gradient spaces over datasets or models? Elucidation of any limits or change in performance within such settings would help one understand whether the assumptions for the method are generalizable.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730956958096}, {"id": "fC9fhe7sAf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11524/Reviewer_6GRV"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "This paper proposes \"DiscQuant\", a method to quantize weights of a large language model, which aims to improve the quantization-accuracy trade-off. The authors argue that quantization can be roughly thought of as two steps: 1) coming up with a good quantization grid,  2) a good rounding scheme that maps the exact parameters to a point in the discrete grid. Authors argue that while there ha been much effort on step 1, there is less focus on step 2, which is the rounding procedure. The key theoretical ingredient of the paper is to formulate the problem of rounding in the framework of discrepancy theory. In particular, they aim to bound the errors introduced due to rounding for all seen and unseen samples\n\nAfter formulating the problem in an exact manner, which is roughly the KL divergence between unrounded and rounded model predictions, they go on to make several assumptions: 1) they assume that a simple rounding up/down suffices, and there is no needs for \"jumps\" in the rounding, which becomes more reasonable if the grid is fine enough. 2) they assuming a particular low-rank structure about the weight gradients, and that gradients are well behaved (defined as $\\beta$-reasonable). This is key to achieve an unseen or generalization error bound. 3) They assume that  the first order approximation to the error is sufficient for calculating the errors. Here, they stress the fact that while loss gradients averaged over samples may be small, the per sample gradients are not small and in fact dominate the errors due to rounding. \n\nAfter making thee assumptions, the paper goes on to state the main theorem of the paper, Theorem 3.3, which gives the guarantee for generalization error, and subsequently introduce the algorithm that finds such a rounding efficiently in section 4. Finally, the paper presents, what seems like very compelling evidence that their proposed quantization scheme outperforms two baselines (RTN and GPTQ, see Figure 2). They also empirically test two key assumptions, the first order error approximation in Figure 3, and the low-rank structure of gradients in Figure 4,  which substantiates why their theoretical results are applicable.", "review_text": "This paper proposes \"DiscQuant\", a method to quantize weights of a large language model, which aims to improve the quantization-accuracy trade-off. The authors argue that quantization can be roughly thought of as two steps: 1) coming up with a good quantization grid,  2) a good rounding scheme that maps the exact parameters to a point in the discrete grid. Authors argue that while there ha been much effort on step 1, there is less focus on step 2, which is the rounding procedure. The key theoretical ingredient of the paper is to formulate the problem of rounding in the framework of discrepancy theory. In particular, they aim to bound the errors introduced due to rounding for all seen and unseen samples\n\nAfter formulating the problem in an exact manner, which is roughly the KL divergence between unrounded and rounded model predictions, they go on to make several assumptions: 1) they assume that a simple rounding up/down suffices, and there is no needs for \"jumps\" in the rounding, which becomes more reasonable if the grid is fine enough. 2) they assuming a particular low-rank structure about the weight gradients, and that gradients are well behaved (defined as $\\beta$-reasonable). This is key to achieve an unseen or generalization error bound. 3) They assume that  the first order approximation to the error is sufficient for calculating the errors. Here, they stress the fact that while loss gradients averaged over samples may be small, the per sample gradients are not small and in fact dominate the errors due to rounding. \n\nAfter making thee assumptions, the paper goes on to state the main theorem of the paper, Theorem 3.3, which gives the guarantee for generalization error, and subsequently introduce the algorithm that finds such a rounding efficiently in section 4. Finally, the paper presents, what seems like very compelling evidence that their proposed quantization scheme outperforms two baselines (RTN and GPTQ, see Figure 2). They also empirically test two key assumptions, the first order error approximation in Figure 3, and the low-rank structure of gradients in Figure 4,  which substantiates why their theoretical results are applicable.", "strengths": "In general, I find this work to be very strong, as it works on an important practical problem, and presents an innovative appraoch that is both theoretically grounded and is practically impactful. Let me enumerate these strengths one by one:\n- The recognition of the problem with exiting quantization methods, in that they ignore the importance of the rounding step and only focus on the quantization grid, seems to be a highly relevant and important message of this paper\n- After recognizing this problem, authors propose a very nice and innovative approach by formulating the problem in terms of discrepancy theory, which in hindsight, seems like an excellent choice for this problem\n- The assumptions necessary for the theory, seems to be well substantiated and reasonable, and authors go to reasonable length to explain and justify them, rather than hiding them. \n- The empirical results of the paper are equally strong as the theoretical results.", "weaknesses": "My main criticism of the current draft is its lack of clarity on some of the technical/theoretical parts of the paper. \nFor example, the paper would benefit a lot from an expanded explanation of the basics, ie, the basic discrepancy theory setup that they are casting their problem to, explaining the e Lovett-Meka algorithm that they are invoking so many times in detail, and then explaining what problems (complexity perhaps) it has, and what they do to fix it. Currently, it seems like the paper assumes the reader is already familiar with all thee topics and they only present bits that are novel. For reference, I spent nearly 1 hour trying to catch up with thee basics, namely the  Lovett-Meka algorithm,  but still only partially understood the technical details of the paper.  \n\nEven after an expanded explanation on the theory basics, I think the paper needs to give more intuitive high level view of the algorithm. In particular, in section 4, there could be more explanations on what is the idea behind this formulation/heuristic of minimizing along arbitrary direction $c$? Perhaps a geometric intuition, similar to the one given in Fig 1, could be given here? \n\nAnother point is the lack of clarity on the complexity of the proposed approach. From my limited understanding, the Lovett-Meka algorithm involves iteratively solving a SDP, which could be highly expensive in some cases. It sounds like this paper addresses some of these complexity issues via the heuristics (eg, lines 371-272). But it's hard to fully understand the solution, if the reader hasn't fully understood the problem. So an expanded section on background methods and their complexity, and then the complexity of the heuristics-based approach, would help the reader quite a bit.", "questions": "One of my biggest questions that remained unresolved while reading the paper was, when discussing gradient covariance matrix, are we talking about the covariance within a fully connected layer, or could it be between different layers? In general,  there is some ambiguity as to what “n” in the parameter space really entails here. Is it the full parameter space (all parameters together), or is this procedure applied per each fully connected layer? If it is the latter, the next question is, in which order are these soundings applied? And would the authors re-calculated the gradients after each step? \n\nOn a related note, if this procedure is done iteratively for different fully connected layers, can authors expand on this? For example, do they quantize every layer as if other layers are in the original form, or do they have an iterative approach where the the quantization of first layer impacts the quantization of the subsequent layers. \n\nIf this is done in one shot, is that a complex operation? In general, I would also highly welcome some notes on complexity of running this algorithm.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes \"DiscQuant\", a method to quantize weights of a large language model, which aims to improve the quantization-accuracy trade-off. The authors argue that quantization can be roughly thought of as two steps: 1) coming up with a good quantization grid,  2) a good rounding scheme that maps the exact parameters to a point in the discrete grid. Authors argue that while there ha been much effort on step 1, there is less focus on step 2, which is the rounding procedure. The key theoretical ingredient of the paper is to formulate the problem of rounding in the framework of discrepancy theory. In particular, they aim to bound the errors introduced due to rounding for all seen and unseen samples\n\nAfter formulating the problem in an exact manner, which is roughly the KL divergence between unrounded and rounded model predictions, they go on to make several assumptions: 1) they assume that a simple rounding up/down suffices, and there is no needs for \"jumps\" in the rounding, which becomes more reasonable if the grid is fine enough. 2) they assuming a particular low-rank structure about the weight gradients, and that gradients are well behaved (defined as $\\beta$-reasonable). This is key to achieve an unseen or generalization error bound. 3) They assume that  the first order approximation to the error is sufficient for calculating the errors. Here, they stress the fact that while loss gradients averaged over samples may be small, the per sample gradients are not small and in fact dominate the errors due to rounding. \n\nAfter making thee assumptions, the paper goes on to state the main theorem of the paper, Theorem 3.3, which gives the guarantee for generalization error, and subsequently introduce the algorithm that finds such a rounding efficiently in section 4. Finally, the paper presents, what seems like very compelling evidence that their proposed quantization scheme outperforms two baselines (RTN and GPTQ, see Figure 2). They also empirically test two key assumptions, the first order error approximation in Figure 3, and the low-rank structure of gradients in Figure 4,  which substantiates why their theoretical results are applicable.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "In general, I find this work to be very strong, as it works on an important practical problem, and presents an innovative appraoch that is both theoretically grounded and is practically impactful. Let me enumerate these strengths one by one:\n- The recognition of the problem with exiting quantization methods, in that they ignore the importance of the rounding step and only focus on the quantization grid, seems to be a highly relevant and important message of this paper\n- After recognizing this problem, authors propose a very nice and innovative approach by formulating the problem in terms of discrepancy theory, which in hindsight, seems like an excellent choice for this problem\n- The assumptions necessary for the theory, seems to be well substantiated and reasonable, and authors go to reasonable length to explain and justify them, rather than hiding them. \n- The empirical results of the paper are equally strong as the theoretical results.", "weaknesses": "My main criticism of the current draft is its lack of clarity on some of the technical/theoretical parts of the paper. \nFor example, the paper would benefit a lot from an expanded explanation of the basics, ie, the basic discrepancy theory setup that they are casting their problem to, explaining the e Lovett-Meka algorithm that they are invoking so many times in detail, and then explaining what problems (complexity perhaps) it has, and what they do to fix it. Currently, it seems like the paper assumes the reader is already familiar with all thee topics and they only present bits that are novel. For reference, I spent nearly 1 hour trying to catch up with thee basics, namely the  Lovett-Meka algorithm,  but still only partially understood the technical details of the paper.  \n\nEven after an expanded explanation on the theory basics, I think the paper needs to give more intuitive high level view of the algorithm. In particular, in section 4, there could be more explanations on what is the idea behind this formulation/heuristic of minimizing along arbitrary direction $c$? Perhaps a geometric intuition, similar to the one given in Fig 1, could be given here? \n\nAnother point is the lack of clarity on the complexity of the proposed approach. From my limited understanding, the Lovett-Meka algorithm involves iteratively solving a SDP, which could be highly expensive in some cases. It sounds like this paper addresses some of these complexity issues via the heuristics (eg, lines 371-272). But it's hard to fully understand the solution, if the reader hasn't fully understood the problem. So an expanded section on background methods and their complexity, and then the complexity of the heuristics-based approach, would help the reader quite a bit.", "questions": "One of my biggest questions that remained unresolved while reading the paper was, when discussing gradient covariance matrix, are we talking about the covariance within a fully connected layer, or could it be between different layers? In general,  there is some ambiguity as to what “n” in the parameter space really entails here. Is it the full parameter space (all parameters together), or is this procedure applied per each fully connected layer? If it is the latter, the next question is, in which order are these soundings applied? And would the authors re-calculated the gradients after each step? \n\nOn a related note, if this procedure is done iteratively for different fully connected layers, can authors expand on this? For example, do they quantize every layer as if other layers are in the original form, or do they have an iterative approach where the the quantization of first layer impacts the quantization of the subsequent layers. \n\nIf this is done in one shot, is that a complex operation? In general, I would also highly welcome some notes on complexity of running this algorithm.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730484777087}, {"id": "fMTkUM8Y3J", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11524/Reviewer_7uqc"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper studied the generalization gap of quantization using techniques from discrepancy theory under the assumption that the gradient is approximately low-rank. Based on theoretical analysis, the authors proposed a new quantization algorithm, named DiscQuant. Experiments are conducted to compare the proposed algorithm with existing quantization algorithm, like RTN and GPTQ.", "review_text": "The paper studied the generalization gap of quantization using techniques from discrepancy theory under the assumption that the gradient is approximately low-rank. Based on theoretical analysis, the authors proposed a new quantization algorithm, named DiscQuant. Experiments are conducted to compare the proposed algorithm with existing quantization algorithm, like RTN and GPTQ.", "strengths": "1. Theoretical analysis is solid. The paper provides a solid theoretical analysis to study the generalization error of quantization. \n2. Connection with discrepancy theory. It is novel to apply techniques of discrepancy theory to neural network quantization.", "weaknesses": "The weaknesses of the paper mainly come from the numerical algorithm and experiments. \n\n1. The proposed algorithm solves the optimization problem (3). It seems like full-model training is required to solve the problem (3), which can be too expensive for state-of-the-art large language models. \n2. The authors only compare the proposed approach with RTN and GPTQ, which are relatively old PTQ algorithms in the area. The results seem to show a big gap between the SOTA PTQ methods.", "questions": "1. The result in the main Theorem 3.3 only depends on the data distribution. In practice, the PTQ performance also depends on the value distribution of weight matrices. If values in weights are evenly distributed, the performance after PTQ is usually better. The authors also mentioned incoherence processing, a popular method to reduce the ranges of weights. I am very curious why the distribution of weights is not reflected in the main theorem 3.3 for generalization error. \n2. Could authors explain more about how to solve the problem (3) in their algorithm? In my understanding, we may need to run backpropagation and optimize the problem (3). If that is the case, the cost is almost the same as training the full model which is too much for PTQ compared to other existing algorithms. By using the resources, people can directly run distillation or quantization-aware training for better compression and better performance. \n3. A follow-up question, can authors provide time and memory costs for the proposed algorithm?\n4. Nowadays, there are lots of new algorithms for quantizing neural networks. It is better if authors can compare their algorithm with those works. Some new algorithms are listed in the following: \n      * Zhang, Aozhong, et al. \"MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization.\" arXiv preprint arXiv:2406.00800 (2024).\n      * Shao, Wenqi, et al. \"Omniquant: Omnidirectionally calibrated quantization for large language models.\" arXiv preprint arXiv:2308.13137 (2023).\n      * Chee, Jerry, et al. \"Quip: 2-bit quantization of large language models with guarantees.\" Advances in Neural Information Processing Systems 36 (2024).\n      * Liu, Zechun, et al. \"SpinQuant--LLM quantization with learned rotations.\" arXiv preprint arXiv:2405.16406 (2024).\nThe results shown in the paper seem to be worse than the reported results in these most recent algorithms. It will be great the authors can choose some of them to make comparisons and explain the performance gap.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studied the generalization gap of quantization using techniques from discrepancy theory under the assumption that the gradient is approximately low-rank. Based on theoretical analysis, the authors proposed a new quantization algorithm, named DiscQuant. Experiments are conducted to compare the proposed algorithm with existing quantization algorithm, like RTN and GPTQ.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Theoretical analysis is solid. The paper provides a solid theoretical analysis to study the generalization error of quantization. \n2. Connection with discrepancy theory. It is novel to apply techniques of discrepancy theory to neural network quantization.", "weaknesses": "The weaknesses of the paper mainly come from the numerical algorithm and experiments. \n\n1. The proposed algorithm solves the optimization problem (3). It seems like full-model training is required to solve the problem (3), which can be too expensive for state-of-the-art large language models. \n2. The authors only compare the proposed approach with RTN and GPTQ, which are relatively old PTQ algorithms in the area. The results seem to show a big gap between the SOTA PTQ methods.", "questions": "1. The result in the main Theorem 3.3 only depends on the data distribution. In practice, the PTQ performance also depends on the value distribution of weight matrices. If values in weights are evenly distributed, the performance after PTQ is usually better. The authors also mentioned incoherence processing, a popular method to reduce the ranges of weights. I am very curious why the distribution of weights is not reflected in the main theorem 3.3 for generalization error. \n2. Could authors explain more about how to solve the problem (3) in their algorithm? In my understanding, we may need to run backpropagation and optimize the problem (3). If that is the case, the cost is almost the same as training the full model which is too much for PTQ compared to other existing algorithms. By using the resources, people can directly run distillation or quantization-aware training for better compression and better performance. \n3. A follow-up question, can authors provide time and memory costs for the proposed algorithm?\n4. Nowadays, there are lots of new algorithms for quantizing neural networks. It is better if authors can compare their algorithm with those works. Some new algorithms are listed in the following: \n      * Zhang, Aozhong, et al. \"MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization.\" arXiv preprint arXiv:2406.00800 (2024).\n      * Shao, Wenqi, et al. \"Omniquant: Omnidirectionally calibrated quantization for large language models.\" arXiv preprint arXiv:2308.13137 (2023).\n      * Chee, Jerry, et al. \"Quip: 2-bit quantization of large language models with guarantees.\" Advances in Neural Information Processing Systems 36 (2024).\n      * Liu, Zechun, et al. \"SpinQuant--LLM quantization with learned rotations.\" arXiv preprint arXiv:2405.16406 (2024).\nThe results shown in the paper seem to be worse than the reported results in these most recent algorithms. It will be great the authors can choose some of them to make comparisons and explain the performance gap.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730394809655}, {"id": "ZqmiGj39au", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11524/Reviewer_ney3"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper proposes DiscQuant, a data-driven rounding algorithm for post-training quantization (PTQ) of large neural networks. DiscQuant assumes that the gradients of the original model are low-rank. Under this assumption, the authors prove that their algorithm can arbitrarily minimize the upper bound on the error of the quantized model, given an accordingly large number of samples from the target data distribution.\n\nThe algorithm is primarily concerned with the second (rounding) step in quantization. The rounding process aims to minimize the KL divergence between the distributions of the next token predictions of the original and quantized models. The proposed algorithm significantly improves against the existing state-of-the-art when quantizing LLMs like Phi-3-mini-4k-instruct and Meta-Llama-3.1-8B-Instruct.", "review_text": "This paper proposes DiscQuant, a data-driven rounding algorithm for post-training quantization (PTQ) of large neural networks. DiscQuant assumes that the gradients of the original model are low-rank. Under this assumption, the authors prove that their algorithm can arbitrarily minimize the upper bound on the error of the quantized model, given an accordingly large number of samples from the target data distribution.\n\nThe algorithm is primarily concerned with the second (rounding) step in quantization. The rounding process aims to minimize the KL divergence between the distributions of the next token predictions of the original and quantized models. The proposed algorithm significantly improves against the existing state-of-the-art when quantizing LLMs like Phi-3-mini-4k-instruct and Meta-Llama-3.1-8B-Instruct.", "strengths": "1. The paper is well-written and provides sufficient theoretical results using discrepancy theory (§3).\n2. The proposed algorithm is agnostic to the quantization grid, which makes it quite generally applicable.", "weaknesses": "While I acknowledge the contributions made in this work, I hesitate to place a higher score here because of the following reasons.\n\n1. The scope of the currently presented applications is limited to LLMs. I feel that if VLMs (e.g., Llava-v1.6-34b) were to be tested, consistent results there would greatly increase the paper's relevance.\n2. Even within LLMs, the tested models seem quite small. Quantization is concerned with memory efficiency, so it makes sense to perform experiments with large models (e.g., Llama-3.1-70B-Instruct) that would pose greater storage challenges than the ones tested in the current work. Understandably, as the authors note (lines 405-406), DiscQuant requires two copies of the model to be stored in memory during the quantization process. This limits the scope for large-scale experiments in academic settings, but it also highlights a major shortcoming of the approach.\n3. The method needs access to data, which might be the bottleneck for many practitioners who, for instance, quickly need to prototype a handful of models but don't have the resources to run the full models, neither the data to quantize them using a DiscQuant-like algorithm. Besides, as the authors note in the last paragraph, the choice of data is in itself non-trivial. \n\n**Minor issues**\n\n1. The authors may want to confine the abstract to one paragraph in the interest of adherence to the ICLR 2025 guidelines.\n2. Line 133: let's --> lets?\n3. Line 379: close *to* the polytype?", "questions": "It would be great to hear the authors' comments on the points raised above (see **Weaknesses**). In addition,\n\n1. Lines 249-251: \"We make this assumption because we don’t want to change any parameter of the original model too much during quantization, consider it an important property of algorithms we design\" -- could the authors further explain this design choice, and the drawbacks of potentially relaxing this restriction?\n2. do the authors plan to release their code?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes DiscQuant, a data-driven rounding algorithm for post-training quantization (PTQ) of large neural networks. DiscQuant assumes that the gradients of the original model are low-rank. Under this assumption, the authors prove that their algorithm can arbitrarily minimize the upper bound on the error of the quantized model, given an accordingly large number of samples from the target data distribution.\n\nThe algorithm is primarily concerned with the second (rounding) step in quantization. The rounding process aims to minimize the KL divergence between the distributions of the next token predictions of the original and quantized models. The proposed algorithm significantly improves against the existing state-of-the-art when quantizing LLMs like Phi-3-mini-4k-instruct and Meta-Llama-3.1-8B-Instruct.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper is well-written and provides sufficient theoretical results using discrepancy theory (§3).\n2. The proposed algorithm is agnostic to the quantization grid, which makes it quite generally applicable.", "weaknesses": "While I acknowledge the contributions made in this work, I hesitate to place a higher score here because of the following reasons.\n\n1. The scope of the currently presented applications is limited to LLMs. I feel that if VLMs (e.g., Llava-v1.6-34b) were to be tested, consistent results there would greatly increase the paper's relevance.\n2. Even within LLMs, the tested models seem quite small. Quantization is concerned with memory efficiency, so it makes sense to perform experiments with large models (e.g., Llama-3.1-70B-Instruct) that would pose greater storage challenges than the ones tested in the current work. Understandably, as the authors note (lines 405-406), DiscQuant requires two copies of the model to be stored in memory during the quantization process. This limits the scope for large-scale experiments in academic settings, but it also highlights a major shortcoming of the approach.\n3. The method needs access to data, which might be the bottleneck for many practitioners who, for instance, quickly need to prototype a handful of models but don't have the resources to run the full models, neither the data to quantize them using a DiscQuant-like algorithm. Besides, as the authors note in the last paragraph, the choice of data is in itself non-trivial. \n\n**Minor issues**\n\n1. The authors may want to confine the abstract to one paragraph in the interest of adherence to the ICLR 2025 guidelines.\n2. Line 133: let's --> lets?\n3. Line 379: close *to* the polytype?", "questions": "It would be great to hear the authors' comments on the points raised above (see **Weaknesses**). In addition,\n\n1. Lines 249-251: \"We make this assumption because we don’t want to change any parameter of the original model too much during quantization, consider it an important property of algorithms we design\" -- could the authors further explain this design choice, and the drawbacks of potentially relaxing this restriction?\n2. do the authors plan to release their code?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730068386616}], "openreview_url": "https://openreview.net/forum?id=vJmpg0exYA", "arxiv_id": "2501.06417", "paper_pdf": "papers/vJmpg0exYA.pdf", "paper_pdf_sha256": "80f4912851402231ebd62268d7269a6c4ea54ac1d6e8befe107e4f9af33e771b", "paper_pdf_bytes": 926711, "paper_pdf_source": "openreview", "code_url": "https://github.com/jerry-chee/DiscQuant", "code_repository": "jerry-chee/DiscQuant", "code_commit": "b1db18b079803982d70d2e74e3679e6a96a683e5", "code_archive": "repos/vJmpg0exYA.zip", "code_archive_sha256": "c44e0ee2ae20182a892f5101629dc1704215f42a4fcaebc795890306f4f4a1c3", "code_archive_bytes": 38876, "code_file_count": 15, "code_extensions": {".py": 12, ".sh": 3}, "github_disk_usage_kb": 34, "github_languages": {"Python": 233732, "Shell": 4702}, "github_archived": false, "github_pushed_at": "2025-01-10T04:16:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discquant-a-quantization-method-for-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IhWtRwIbos", "year": 2024, "status": "rejected", "title": "Discovering Environments with XRM", "authors": ["Mohammad Pezeshki", "Diane Bouchacourt", "Mark Ibrahim", "Nicolas Ballas", "Pascal Vincent", "David Lopez-Paz"], "authorids": ["~Mohammad_Pezeshki1", "~Diane_Bouchacourt3", "~Mark_Ibrahim1", "~Nicolas_Ballas1", "~Pascal_Vincent1", "~David_Lopez-Paz2"], "authors_source": "OpenReview API", "abstract": "Successful out-of-distribution generalization requires environment annotations. Unfortunately, these are resource-intensive to obtain, and their relevance to model performance is limited by the expectations and perceptual biases of human annotators. Therefore, to enable robust AI systems across applications, we must develop algorithms to automatically discover environments inducing broad generalization. Current proposals, which divide examples based on their training error, suffer from one fundamental problem. These methods add hyper-parameters and early-stopping criteria that are impossible to tune without a validation set with human-annotated environments, the very information subject to discovery. In this paper, we propose Cross-Risk-Minimization (XRM) to address this issue. XRM trains two twin networks, each learning from one random half of the training data, while imitating confident held-out mistakes made by its sibling. XRM provides a recipe for hyper-parameter tuning, does not require early-stopping, and can discover environments for all training and validation data. Domain generalization algorithms built on top of XRM environments achieve oracle worst-group-accuracy, solving a long-standing problem in out-of-distribution generalization.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "D1zaYXQauT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2424/Reviewer_fj8c"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper addresses the challenge of achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. The authors propose Cross-Risk Minimization (XRM), a novel approach that trains twin networks to learn from random halves of the training data while imitating confident mistakes made by their counterparts. XRM enables automatic discovery of environments for both training and validation data. The authors demonstrate the effectiveness of XRM by building domain generalization algorithms based on the discovered environments, achieving oracle worst-group-accuracy.", "review_text": "This paper addresses the challenge of achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. The authors propose Cross-Risk Minimization (XRM), a novel approach that trains twin networks to learn from random halves of the training data while imitating confident mistakes made by their counterparts. XRM enables automatic discovery of environments for both training and validation data. The authors demonstrate the effectiveness of XRM by building domain generalization algorithms based on the discovered environments, achieving oracle worst-group-accuracy.", "strengths": "1. The paper is well-organized and easy to understand.\n2. This paper addresses a crucial challenge in Domain Generalization (DG) tasks, which is the data-splitting process without relying on human annotations.\n3. The authors provide strong empirical evidence through extensive experiments to substantiate the effectiveness of their proposed XRM method.", "weaknesses": "1. The paper's claims may be slightly overstated. While the focus on subpopulation shift in distribution shift is indeed important, it might be more appropriate to avoid claiming to solve a long-standing problem in out-of-distribution generalization without further empirical studies on widely recognized DG benchmarks such as DomainBed and Wilds. These additional experiments could provide more convincing evidence of the proposed approach's effectiveness.\n2. The paper lacks a comprehensive discussion of important related works concerning data splitting strategies for improved DG performance and subpopulation shift, such as references [1], [2], [3] and [4]. Notably, in [1], the authors have theoretically demonstrated the challenges of learning the invariant correlation between samples and labels in the absence of prior information. Including these relevant works would enhance the paper's literature review and contextualize the proposed approach.\n\nTypo:\n\nIn the first sentence of the paragraph above section 4.2, it appears that the authors have inadvertently added a redundant \"we.”\n\n[1] ZIN: When and How to Learn Invariance Without Environment Partition?\n[2] Provably Invariant Learning without Domain Information.\n[3] Rethinking Invariant Graph Representation Learning without Environment Partitions.\n[4] Just Mix Once: Worst-group Generalization by Group Interpolation.", "questions": "1. As highlighted in [1], it is crucial to understand the specific scenarios where XRM is expected to be effective. Therefore, it would be beneficial for the authors to provide further insights into the data distribution settings in which XRM is likely to perform well. Alternatively, the authors could explore providing theoretical guarantees to enhance the understanding of XRM's strengths and limitations.\n2. The observation that XRM outperforms Human-annotation methods is intriguing and warrants further explanation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the challenge of achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. The authors propose Cross-Risk Minimization (XRM), a novel approach that trains twin networks to learn from random halves of the training data while imitating confident mistakes made by their counterparts. XRM enables automatic discovery of environments for both training and validation data. The authors demonstrate the effectiveness of XRM by building domain generalization algorithms based on the discovered environments, achieving oracle worst-group-accuracy.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is well-organized and easy to understand.\n2. This paper addresses a crucial challenge in Domain Generalization (DG) tasks, which is the data-splitting process without relying on human annotations.\n3. The authors provide strong empirical evidence through extensive experiments to substantiate the effectiveness of their proposed XRM method.", "weaknesses": "1. The paper's claims may be slightly overstated. While the focus on subpopulation shift in distribution shift is indeed important, it might be more appropriate to avoid claiming to solve a long-standing problem in out-of-distribution generalization without further empirical studies on widely recognized DG benchmarks such as DomainBed and Wilds. These additional experiments could provide more convincing evidence of the proposed approach's effectiveness.\n2. The paper lacks a comprehensive discussion of important related works concerning data splitting strategies for improved DG performance and subpopulation shift, such as references [1], [2], [3] and [4]. Notably, in [1], the authors have theoretically demonstrated the challenges of learning the invariant correlation between samples and labels in the absence of prior information. Including these relevant works would enhance the paper's literature review and contextualize the proposed approach.\n\nTypo:\n\nIn the first sentence of the paragraph above section 4.2, it appears that the authors have inadvertently added a redundant \"we.”\n\n[1] ZIN: When and How to Learn Invariance Without Environment Partition?\n[2] Provably Invariant Learning without Domain Information.\n[3] Rethinking Invariant Graph Representation Learning without Environment Partitions.\n[4] Just Mix Once: Worst-group Generalization by Group Interpolation.", "questions": "1. As highlighted in [1], it is crucial to understand the specific scenarios where XRM is expected to be effective. Therefore, it would be beneficial for the authors to provide further insights into the data distribution settings in which XRM is likely to perform well. Alternatively, the authors could explore providing theoretical guarantees to enhance the understanding of XRM's strengths and limitations.\n2. The observation that XRM outperforms Human-annotation methods is intriguing and warrants further explanation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698549604766}, {"id": "QZ7kLDcXKl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2424/Reviewer_f39H"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces CROSS-RISK MINIMIZATION (XRM), a method for achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. By training twin networks to imitate confident mistakes made by each other, XRM enables automatic discovery of relevant environments for training and validation data. The proposed approach addresses the challenge of hyper-parameter tuning and achieves oracle worst-group accuracy, offering a promising solution for broad generalization in AI systems.", "review_text": "The paper introduces CROSS-RISK MINIMIZATION (XRM), a method for achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. By training twin networks to imitate confident mistakes made by each other, XRM enables automatic discovery of relevant environments for training and validation data. The proposed approach addresses the challenge of hyper-parameter tuning and achieves oracle worst-group accuracy, offering a promising solution for broad generalization in AI systems.", "strengths": "* The problem of learning OOD robust model without manual domain partition is a very important task, which might has great impact on real-world applications.\n* The proposed method has a clear advantage over existing methods such as EIIL and JTT, that they does not need to explicitly tune the hyperparameter for early stopping. Since hyper-parameter tuning is a crucial challenge, the proposed method would be of interest to many.\n* The empirical performance is strong.", "weaknesses": "I have several concerns as follows:\n\n1. The paper should provide a clear discussion on the identifiability challenges presented in [1], which demonstrate that learning invariance without domain partition can be generally impossible. It is crucial to address the need for imposing inductive bias, additional assumptions, conditions, or auxiliary information to ensure the effectiveness of the proposed method. A thorough exploration of these aspects would enhance the paper's theoretical foundation and its practical applicability.\n\n2. According to [2, 3], spurious features are defined as any nodes in the causal graph other than the direct causes of the label. However, I have concerns about the evaluations conducted on datasets like waterbird, which explicitly contain only one dominating spurious feature. These datasets may not fully reflect the implications of the proposed methods on more realistic and high-dimensional datasets such as ImageNet variants. Moreover, the paper relies on the assumption that Empirical Risk Minimization (ERM) learns spurious features first, but it may not hold true for all types of spurious features as discussed in [2, 3]. It would be valuable to address these concerns and provide further insights into the generalizability of the method to diverse real-world datasets.   \n\n3. If a large number of spurious features are present, [1] demonstrates that there are necessary and sufficient conditions for learning invariance without explicit domain partitions, which can be quite restrictive. I have concerns about whether the proposed two-stage method can effectively address this problem given the limitations imposed by these conditions. It would be valuable for the authors to discuss how their method overcomes or accommodates these restrictions and whether it can achieve satisfactory results in scenarios with a significant number of spurious features.  \n\n4.  Several studies [5, 6] have highlighted the challenges associated with learning invariance in the presence of many spurious features. In a recent paper, [4] discovered that when dealing with a large number of spurious features, each ERM model tends to learn a subset of these features. [4] further demonstrates that rather than exclusively focusing on learning invariant features, it is beneficial for OOD performance to diversify the learned spurious features (referred to as spurious feature diversification). Spurious feature diversification is shown to explain the effectiveness of empirically strong methods like SWAD and Model soup. It would be valuable to investigate whether the proposed method (XRM) can enhance spurious feature diversification and demonstrate effective performance on a broader range of real-world datasets, such as PACS, OfficeHome, DomainNet, or ImageNet variants.\n\nCorrect me if I was wrong. I would increase the score if (part of) my concerns were addressed. \n\n[1] Yong Lin et.al., ZIN: When and how to learn invariance without domain partition.\n\n[2] Martin Arjovsky et.al., Invariant Risk Minimization \n\n[3] Jonas Peters, et.al.,. Causal inference using invariant prediction: identification and confidence intervals\n\n[4] Yong Lin et.al., Spurious Feature Diversification Improves Out-of-distribution Generalization\n\n[5] Ishaan Gulrajani et.al., In Search of Lost Domain Generalization\n\n[6] Elan Rosenfeld et.al., The Risks of Invariant Risk Minimization\n\n[7] Junbum Cha et.al., SWAD: Domain Generalization by Seeking Flat Minima\n\n[8] Mitchell Wortsman et.al., Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces CROSS-RISK MINIMIZATION (XRM), a method for achieving robust out-of-distribution generalization without relying on resource-intensive environment annotations. By training twin networks to imitate confident mistakes made by each other, XRM enables automatic discovery of relevant environments for training and validation data. The proposed approach addresses the challenge of hyper-parameter tuning and achieves oracle worst-group accuracy, offering a promising solution for broad generalization in AI systems.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "* The problem of learning OOD robust model without manual domain partition is a very important task, which might has great impact on real-world applications.\n* The proposed method has a clear advantage over existing methods such as EIIL and JTT, that they does not need to explicitly tune the hyperparameter for early stopping. Since hyper-parameter tuning is a crucial challenge, the proposed method would be of interest to many.\n* The empirical performance is strong.", "weaknesses": "I have several concerns as follows:\n\n1. The paper should provide a clear discussion on the identifiability challenges presented in [1], which demonstrate that learning invariance without domain partition can be generally impossible. It is crucial to address the need for imposing inductive bias, additional assumptions, conditions, or auxiliary information to ensure the effectiveness of the proposed method. A thorough exploration of these aspects would enhance the paper's theoretical foundation and its practical applicability.\n\n2. According to [2, 3], spurious features are defined as any nodes in the causal graph other than the direct causes of the label. However, I have concerns about the evaluations conducted on datasets like waterbird, which explicitly contain only one dominating spurious feature. These datasets may not fully reflect the implications of the proposed methods on more realistic and high-dimensional datasets such as ImageNet variants. Moreover, the paper relies on the assumption that Empirical Risk Minimization (ERM) learns spurious features first, but it may not hold true for all types of spurious features as discussed in [2, 3]. It would be valuable to address these concerns and provide further insights into the generalizability of the method to diverse real-world datasets.   \n\n3. If a large number of spurious features are present, [1] demonstrates that there are necessary and sufficient conditions for learning invariance without explicit domain partitions, which can be quite restrictive. I have concerns about whether the proposed two-stage method can effectively address this problem given the limitations imposed by these conditions. It would be valuable for the authors to discuss how their method overcomes or accommodates these restrictions and whether it can achieve satisfactory results in scenarios with a significant number of spurious features.  \n\n4.  Several studies [5, 6] have highlighted the challenges associated with learning invariance in the presence of many spurious features. In a recent paper, [4] discovered that when dealing with a large number of spurious features, each ERM model tends to learn a subset of these features. [4] further demonstrates that rather than exclusively focusing on learning invariant features, it is beneficial for OOD performance to diversify the learned spurious features (referred to as spurious feature diversification). Spurious feature diversification is shown to explain the effectiveness of empirically strong methods like SWAD and Model soup. It would be valuable to investigate whether the proposed method (XRM) can enhance spurious feature diversification and demonstrate effective performance on a broader range of real-world datasets, such as PACS, OfficeHome, DomainNet, or ImageNet variants.\n\nCorrect me if I was wrong. I would increase the score if (part of) my concerns were addressed. \n\n[1] Yong Lin et.al., ZIN: When and how to learn invariance without domain partition.\n\n[2] Martin Arjovsky et.al., Invariant Risk Minimization \n\n[3] Jonas Peters, et.al.,. Causal inference using invariant prediction: identification and confidence intervals\n\n[4] Yong Lin et.al., Spurious Feature Diversification Improves Out-of-distribution Generalization\n\n[5] Ishaan Gulrajani et.al., In Search of Lost Domain Generalization\n\n[6] Elan Rosenfeld et.al., The Risks of Invariant Risk Minimization\n\n[7] Junbum Cha et.al., SWAD: Domain Generalization by Seeking Flat Minima\n\n[8] Mitchell Wortsman et.al., Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697363256818}, {"id": "xfnLORP8ma", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2424/Reviewer_KDkA"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper addresses OOD generalization by discovering latent environments (partitions) of the training data that are beneficial when used subsequently with standard methods (GroupDRO, reweighting, or resampling to equalize groups during training). The method proceeds by training a pair of models (details discussed below).", "review_text": "The paper addresses OOD generalization by discovering latent environments (partitions) of the training data that are beneficial when used subsequently with standard methods (GroupDRO, reweighting, or resampling to equalize groups during training). The method proceeds by training a pair of models (details discussed below).", "strengths": "- Thorough evaluation on multiple standard datasets.\n\n- Good empirical results.", "weaknesses": "W1. If I understand correctly, the method seems to rely on the fact that misclassified examples are such because they do not contain a \"spurious correlation\" that a model would learn by default. The twin training serves to reinforce the tendency of one of the trained models to capture this spurious correlation. If this is indeed the case, then the overall methods seems to depend on the (common) heuristic that models learn spurious correlations by default (a.k.a. shortcut learning). I think this is the same heuristic that is used in the existing methods criticised in Section 3. Critically, this heuristic relies on the fact that we know that the chosen architecture/training set lead to learnin undesirable spurious correlations by default. What if one applies the method to a situation where a perfectly-fine, \"robust\" model is learned by default? I'm guessing that the method would then be detrimental.\n\nI'm not suggesting that we should be able to do better without additional knowledge (in fact [1] seems to show it's not possible) but the authors here do claim to do so, hence the need to point out this possible limitation (see also W5).\n\nIf my understanding of the method is correct, the method is also very similar to the following.\n- Works in the debiasing literature (e.g. LfF) that train a pair of models that respectively rely/do not rely on spurious features. These works are discussed in Sect. 3, and I do understand that they rely to some extent to the tuning of model capacity to ensure that it captures the spurious feature, but I am not convinced that the proposed parallel training (which seem to be the essential difference) leads to the discovery of something fundamentally different.\n- Works in the \"model diversity\" literature that train a pair of models that differ in their predictions [5,6]. These also proceed to train models in parallel, in a was that seems conceptually very similar to the step 1 proposed here (implementation details aside).\n\n--------\n\nW2. The proposed method only partitions the data into 2 \"environments\". I don't think this is really in line with the literature on DG (with which this work is supposed to connect) that are mostly based on invariance learning and need a large number of training environments (e.g. IRM). This work therefore seems much more related to the simpler setting of \"debiasing\" methods (e.g. LfF) that aim at removing the reliance of a model from one precise biased feature.\n\nThe methods used for the second phase are indeed simple baselines for debiasing, and not really DG methods. These are very strong baselines in these settings (and with the datasets considered), but I'm not sure this is what the reader would expect given all the mentions about DG.\n\n--------\n\nW3. Absence of a comprehensive review of related work. Some directly-related methods are correctly cited/discussed throughout the paper, but there are other connected areas that are not really discussed (examples below).\n\n[1,2] discuss conditions under which environment discovery is possible. I think the theoretical statements in [1] are particularly important to discuss (I am not sure how the proposed method overcomes the impossibility stated in that paper; see also W5).\n[3] was an early method that also proposed to \"unshuffle\" data (a term used in this paper) by simply clustering the data. Looking at the visualizations of discovered \"environments\" in Fig. 3, one wonders if these could also be discovered with such as simple clustering baseline.\n[4] is another recent method that also seems to claim discovering partitions in the data (I suspect it has similar flaws to those discussed in the paper; it has appeared at ICCV 2023 after the ICLR deadline so it's totally fine to dismiss it though).\n\n[1] [ZIN When and How to Learn Invariance Without Environment Partition](https://arxiv.org/abs/2203.05818)\n\n[2] [Provably Invariance Learning without Domain Information](https://proceedings.mlr.press/v202/tan23b/tan23b.pdf)\n\n[3] [Unshuffling data for improved generalization](https://arxiv.org/abs/2002.11894)\n\n[4] [Invariant Feature Regularization for Fair Face Recognition](https://openaccess.thecvf.com/content/ICCV2023/papers/Ma_Invariant_Feature_Regularization_for_Fair_Face_Recognition_ICCV_2023_paper.pdf)\n\n[5] [Agree to Disagree: Diversity through Disagreement for Better Transferability](https://arxiv.org/abs/2202.04414)\n\n[6] [Diversify and Disambiguate: Learning From Underspecified Data](https://arxiv.org/abs/2202.03418)\n\n--------\n\nW4: No discussion or empirical exploration of the limitations of the methods. No precise statement of the assumptions on which the method relies.\n\n--------\n\nMinor comments (no need to comment in the rebuttal; these do not affect my rating of the paper)\n\n- W-minor 1. The writing style is unusual for a technical paper. There are many verbose statements, emotional words, exclamation marks, etc. This is actually a great writing style in other circumstances, but it does not maximize the clarity and efficiency of communication. This does not directly affect my rating of the paper, but it made the reading more tedious. I would suggest using a more concise style and neutral tone for the benefit of the readers.\n\n- W-minor 2. The existing methods for environment discovery based on 2 phases is described twice in sections 1 and 3. It could be clearer to merge these.\nSection 3 is a mix of review/background material/motivation/related work. It's not bad at all in its contents, but it could be easier for the readers to stick with common sections like \"related work\", \"background\", etc.\n\n- W-minor 3. Note that the initial premise stated in the very first sentence of the abstract is not really correct (although it does not really affect the rest of the paper):\n\"Successful out-of-distribution generalization requires environment annotations (...) therefore (...) we must develop algorithms to automatically discover environments\"\nUsing multiple training environments/domains is only one approach to improve OOD generalization.", "questions": "Please comment on W1-W4 above.\n\nTo summarize, the main reasons for my negative rating are the absence of precise statements about limitations/assumptions of the method, and the missing discussion of links with the existing literature. Therefore, I am not sure this is really a work about DG (but rather the simpler setting of single-bias), and the core of the method may be very similar to existing work [5,6] (although presented in very different terms).\n\n--------\n\nIn the spirit of constructive feedback, I would suggest that theses issues are fixable (in a future version) with:\n\n(1) a proper review of the existing work, how/if it relates to this work (e.g. what is the connection with invariance learning? how do the many-environment methods related to this one? how to understand the claims made here in relation to the impossibility theorem in [1] mentioned below?)\n\n(2) a better discussion why/how the proposed method work. The current text is mostly hand waving. Even if a complete theory is out of reach, perhaps a concrete example could help (conceptual, or with a toy example).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses OOD generalization by discovering latent environments (partitions) of the training data that are beneficial when used subsequently with standard methods (GroupDRO, reweighting, or resampling to equalize groups during training). The method proceeds by training a pair of models (details discussed below).", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- Thorough evaluation on multiple standard datasets.\n\n- Good empirical results.", "weaknesses": "W1. If I understand correctly, the method seems to rely on the fact that misclassified examples are such because they do not contain a \"spurious correlation\" that a model would learn by default. The twin training serves to reinforce the tendency of one of the trained models to capture this spurious correlation. If this is indeed the case, then the overall methods seems to depend on the (common) heuristic that models learn spurious correlations by default (a.k.a. shortcut learning). I think this is the same heuristic that is used in the existing methods criticised in Section 3. Critically, this heuristic relies on the fact that we know that the chosen architecture/training set lead to learnin undesirable spurious correlations by default. What if one applies the method to a situation where a perfectly-fine, \"robust\" model is learned by default? I'm guessing that the method would then be detrimental.\n\nI'm not suggesting that we should be able to do better without additional knowledge (in fact [1] seems to show it's not possible) but the authors here do claim to do so, hence the need to point out this possible limitation (see also W5).\n\nIf my understanding of the method is correct, the method is also very similar to the following.\n- Works in the debiasing literature (e.g. LfF) that train a pair of models that respectively rely/do not rely on spurious features. These works are discussed in Sect. 3, and I do understand that they rely to some extent to the tuning of model capacity to ensure that it captures the spurious feature, but I am not convinced that the proposed parallel training (which seem to be the essential difference) leads to the discovery of something fundamentally different.\n- Works in the \"model diversity\" literature that train a pair of models that differ in their predictions [5,6]. These also proceed to train models in parallel, in a was that seems conceptually very similar to the step 1 proposed here (implementation details aside).\n\n--------\n\nW2. The proposed method only partitions the data into 2 \"environments\". I don't think this is really in line with the literature on DG (with which this work is supposed to connect) that are mostly based on invariance learning and need a large number of training environments (e.g. IRM). This work therefore seems much more related to the simpler setting of \"debiasing\" methods (e.g. LfF) that aim at removing the reliance of a model from one precise biased feature.\n\nThe methods used for the second phase are indeed simple baselines for debiasing, and not really DG methods. These are very strong baselines in these settings (and with the datasets considered), but I'm not sure this is what the reader would expect given all the mentions about DG.\n\n--------\n\nW3. Absence of a comprehensive review of related work. Some directly-related methods are correctly cited/discussed throughout the paper, but there are other connected areas that are not really discussed (examples below).\n\n[1,2] discuss conditions under which environment discovery is possible. I think the theoretical statements in [1] are particularly important to discuss (I am not sure how the proposed method overcomes the impossibility stated in that paper; see also W5).\n[3] was an early method that also proposed to \"unshuffle\" data (a term used in this paper) by simply clustering the data. Looking at the visualizations of discovered \"environments\" in Fig. 3, one wonders if these could also be discovered with such as simple clustering baseline.\n[4] is another recent method that also seems to claim discovering partitions in the data (I suspect it has similar flaws to those discussed in the paper; it has appeared at ICCV 2023 after the ICLR deadline so it's totally fine to dismiss it though).\n\n[1] [ZIN When and How to Learn Invariance Without Environment Partition](https://arxiv.org/abs/2203.05818)\n\n[2] [Provably Invariance Learning without Domain Information](https://proceedings.mlr.press/v202/tan23b/tan23b.pdf)\n\n[3] [Unshuffling data for improved generalization](https://arxiv.org/abs/2002.11894)\n\n[4] [Invariant Feature Regularization for Fair Face Recognition](https://openaccess.thecvf.com/content/ICCV2023/papers/Ma_Invariant_Feature_Regularization_for_Fair_Face_Recognition_ICCV_2023_paper.pdf)\n\n[5] [Agree to Disagree: Diversity through Disagreement for Better Transferability](https://arxiv.org/abs/2202.04414)\n\n[6] [Diversify and Disambiguate: Learning From Underspecified Data](https://arxiv.org/abs/2202.03418)\n\n--------\n\nW4: No discussion or empirical exploration of the limitations of the methods. No precise statement of the assumptions on which the method relies.\n\n--------\n\nMinor comments (no need to comment in the rebuttal; these do not affect my rating of the paper)\n\n- W-minor 1. The writing style is unusual for a technical paper. There are many verbose statements, emotional words, exclamation marks, etc. This is actually a great writing style in other circumstances, but it does not maximize the clarity and efficiency of communication. This does not directly affect my rating of the paper, but it made the reading more tedious. I would suggest using a more concise style and neutral tone for the benefit of the readers.\n\n- W-minor 2. The existing methods for environment discovery based on 2 phases is described twice in sections 1 and 3. It could be clearer to merge these.\nSection 3 is a mix of review/background material/motivation/related work. It's not bad at all in its contents, but it could be easier for the readers to stick with common sections like \"related work\", \"background\", etc.\n\n- W-minor 3. Note that the initial premise stated in the very first sentence of the abstract is not really correct (although it does not really affect the rest of the paper):\n\"Successful out-of-distribution generalization requires environment annotations (...) therefore (...) we must develop algorithms to automatically discover environments\"\nUsing multiple training environments/domains is only one approach to improve OOD generalization.", "questions": "Please comment on W1-W4 above.\n\nTo summarize, the main reasons for my negative rating are the absence of precise statements about limitations/assumptions of the method, and the missing discussion of links with the existing literature. Therefore, I am not sure this is really a work about DG (but rather the simpler setting of single-bias), and the core of the method may be very similar to existing work [5,6] (although presented in very different terms).\n\n--------\n\nIn the spirit of constructive feedback, I would suggest that theses issues are fixable (in a future version) with:\n\n(1) a proper review of the existing work, how/if it relates to this work (e.g. what is the connection with invariance learning? how do the many-environment methods related to this one? how to understand the claims made here in relation to the impossibility theorem in [1] mentioned below?)\n\n(2) a better discussion why/how the proposed method work. The current text is mostly hand waving. Even if a complete theory is out of reach, perhaps a concrete example could help (conceptual, or with a toy example).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697189229654}], "openreview_url": "https://openreview.net/forum?id=IhWtRwIbos", "arxiv_id": "2309.16748", "paper_pdf": "papers/IhWtRwIbos.pdf", "paper_pdf_sha256": "530817f989fd5a1b94db56b513bd3e3595723eb4254ca1e29fc33e274f43431d", "paper_pdf_bytes": 1253272, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/XRM", "code_repository": "facebookresearch/XRM", "code_commit": "5e90815250aedea546b9e17ffd5d7e0a7c34c20f", "code_archive": "repos/IhWtRwIbos.zip", "code_archive_sha256": "9f5ff3e8126cebd856716093df2e8f37adc07f5ed500e631b6a25694f33c0660", "code_archive_bytes": 32953, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 32, "github_languages": {"Python": 78103}, "github_archived": true, "github_pushed_at": "2024-12-06T22:23:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discovering-environments-with-xrm"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WcTLZrpzfe", "year": 2023, "status": "rejected", "title": "Orientation-Aware Graph Neural Networks for Protein Structure Representation Learning", "authors": ["Jiahan Li", "Shitong Luo", "Congyue Deng", "Chaoran Cheng", "Jiaqi Guan", "Leonidas Guibas", "Jianzhu Ma", "Jian Peng"], "authorids": ["~Jiahan_Li2", "~Shitong_Luo1", "~Congyue_Deng1", "~Chaoran_Cheng2", "~Jiaqi_Guan1", "~Leonidas_Guibas1", "~Jianzhu_Ma2", "~Jian_Peng1"], "authors_source": "OpenReview API", "abstract": "By folding to particular 3D structures, proteins play a key role in living beings. To learn meaningful representation from a protein structure for downstream tasks, not only the global backbone topology but the local fine-grained orientational relations between amino acids should also be considered. In this work, we propose the Orientation-Aware Graph Neural Networks (OAGNNs) to better sense the geometric characteristics in protein structure (e.g. inner-residue torsion angles, inter-residue orientations). Extending a single weight from a scalar to a 3D vector, we construct a rich set of geometric-meaningful operations to process both the classical and SO(3) representations of a given structure. To plug our designed perceptron unit into existing Graph Neural Networks, we further introduce an equivariant message passing paradigm, showing superior versatility in maintaining SO(3)-equivariance at the global scale. Experiments have shown that our OAGNNs have a remarkable ability to sense geometric orientational features compared to classical networks. OAGNNs have also achieved state-of-the-art performance on various computational biology applications related to protein 3D structures.\n\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "zeJhhWAZKW7", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1194/Reviewer_XTJW"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper designs a directed weight operation to bake the SO(3)-action into the neural network by adding an extra dimension in the weight matrix. Except for the directed weight perceptrons, the paper proposes an equivariant message passing neural network combing the DWP module. The synthetic experiment shows the effectiveness of the DWP, and the experiments on protein 3D structures achieve good performance. Overall, though the experiments are able to show performance enhancement using such a directed weight matrix, the paper has a limited illustration of the DWP module. This paper also contains some typos, need to check more carefully.", "review_text": "The paper proposed a DWP plus SO(3)-equivariant GNNs for protein engineering. The merit of the design compared to the existing GVP and other equivariant message passing is not significant and experiments are not sufficient for illustrating the new model's effectiveness.", "strengths": "**Strength:**\n1. The DWP is novel to combine the SO(3)-rotation and the linear transformation, unlike the VNN model, the authors also consider extracting orientation information with the weight matrix. Adding an extra dimension on the weight matrix is quite similar to the 3D convolution filters that extract 3D information from the 3D image data.\n2. The synthetic experiments are a bonus to show the effectiveness of DWP and beat other related methods.\n3. The proposed DWP module is compatible with other GNN models, which is quite interesting to investigate in many applications that require more orientation information.\n \n**Weaknesses:**\n1. The number of learning parameters will triple the size of the linear transform matrix and the complexity might be too big.\n2. The ablation experiments should also consider keeping the model parameter almost the same as your proposed model since the DWP brings more parameters to the model than the other methods.\n3. More illustrations or experiments to directly research the directed weight matrix compared with the undirected one.\n4. More experiments are needed to verify the quality of the model.\n\n**Question:**\n1. The proposed method (even the illustrative figure) is similar to GVP. What is the merit of using DWP compared to GVP?\n2. Is there any constraint act on the DWP to preserve some property? If we regard the directed weight matrix as the same as the normal weight matrix, we are not sure whether the orientation information in the data is extracted by the DWP module.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper designs a directed weight operation to bake the SO(3)-action into the neural network by adding an extra dimension in the weight matrix. Except for the directed weight perceptrons, the paper proposes an equivariant message passing neural network combing the DWP module. The synthetic experiment shows the effectiveness of the DWP, and the experiments on protein 3D structures achieve good performance. Overall, though the experiments are able to show performance enhancement using such a directed weight matrix, the paper has a limited illustration of the DWP module. This paper also contains some typos, need to check more carefully.", "strength_and_weaknesses": "**Strength:**\n1. The DWP is novel to combine the SO(3)-rotation and the linear transformation, unlike the VNN model, the authors also consider extracting orientation information with the weight matrix. Adding an extra dimension on the weight matrix is quite similar to the 3D convolution filters that extract 3D information from the 3D image data.\n2. The synthetic experiments are a bonus to show the effectiveness of DWP and beat other related methods.\n3. The proposed DWP module is compatible with other GNN models, which is quite interesting to investigate in many applications that require more orientation information.\n \n**Weaknesses:**\n1. The number of learning parameters will triple the size of the linear transform matrix and the complexity might be too big.\n2. The ablation experiments should also consider keeping the model parameter almost the same as your proposed model since the DWP brings more parameters to the model than the other methods.\n3. More illustrations or experiments to directly research the directed weight matrix compared with the undirected one.\n4. More experiments are needed to verify the quality of the model.\n\n**Question:**\n1. The proposed method (even the illustrative figure) is similar to GVP. What is the merit of using DWP compared to GVP?\n2. Is there any constraint act on the DWP to preserve some property? If we regard the directed weight matrix as the same as the normal weight matrix, we are not sure whether the orientation information in the data is extracted by the DWP module.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper tries to enhance the representation of protein data, which is a vital problem to promote AI research on protein-related tasks. It remains unclear which role the DWP plays since it only expands the dimension of the learning parameter and no further analysis of the DWP or the feature it extracts. Compared with other related methods, we think the paper has limited contribution either to the proposed GNN model or the explainability of the directed weight matrix. The extended experiments show the model a promising work, more research should be taken to refine the mechanism that DWP makes the model aware of the orientation.\n", "summary_of_the_review": "The paper proposed a DWP plus SO(3)-equivariant GNNs for protein engineering. The merit of the design compared to the existing GVP and other equivariant message passing is not significant and experiments are not sufficient for illustrating the new model's effectiveness.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667671814072}, {"id": "eH7IYZAIL3Y", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1194/Reviewer_EZjF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a graph neural network architecture that uses vectorized neurons, extensive hidden feature mixing operations, and global/local reference frame transformations to produce protein embeddings that are sensitive to or \"aware of\" residue-specific orientations within the larger macromolecule. They first introduce Directed Weights, combinations of multidimensional tensors that generalize scalars to vectors and vectors to higher-dimensional tensors. They then describe the Directed Weight Perceptron module, which performs numerous scalar and vector operations such as linear multiplication, dot products, and cross products. Armed with networks that more naturally describe data and relationships in 3D space, the authors next describe their SO(3) equivariant message passing paradigm. They then describe a few variants of their model adapted to mimic the architectures of recent graph models such as GCNs, GINs, and GATs. Finally, the authors demonstrate their models' performance on a number of protein-related tasks.", "review_text": "An interesting method, but the presentation lacks clarity and several design decisions are not well justified.\n\nWhat would improve my score: add clarifications and additional details.", "strengths": "This paper introduces a novel concept, but the presentation, design choices, and testing methodology could all be much clearer. Moving some of the information from the appendix to the text would clarify the role of various mechanisms presented. For example, discussing more directly the scalar and vector feature constructions earlier in the text and not leaving them to the appendix would justify the use of the separate scalar features in the first place, which may otherwise seem extraneous when compared to feature vectors that could be magnitudes larger. Second, it isn't clear why the Directed Interaction model is necessary. Its stated purpose is to establish a connection between the scalar and vector features, which seems redundant given the operations blending scalar and vector features in the Directed Linear module. Clarification on what this module provides that the latter does not would be helpful.\n\nThe experimental results generally seem reliable and convincing, but it is not immediately clear what configurations of models are being compared to each other. It even seems in some places that conflicting information is given i.e. in B.2 the DWP model is 1 layer for the Synthetic Task but B.3 \"Our model uses 4-layer message passing, where each message passing consists of 3-layer DWP\". The latter further seems redundant with \"there is 4 message passing...3-layer networks.\" One can tease these statements/cases apart, but more concise wording and clearer statement of model architecture/hyperparameters would be beneficial. Further, the synthetic task does not seem convincing without more information about how the points on a sphere were sampled, and statistics such as the variances of the positive and negative samples. Finally, it seems rather astounding that an MLP can match the performance of the most sophisticated models with a comparable number of weights on this task. The universal approximation theorem is relevant for arbitrarily deep and/or wide networks, so this is not sufficient justification. Similarly, it appears that GVP fails completely on this synthetic task, but is tested in later tasks whereas VNN performs well and is excluded without explanation.\n\nThe visualizations are informative but the table layout (tables 3-5) is confusing. There are also many citations to the arxiv version of papers that have been published in ML conference proceedings or other venues. The references should be carefully checked and corrected.\n\nSome specific questions:\n1. Why was GVP chosen over VNN for testing on benchmarks?\n2. What applications outside of protein embedding is this work relevant to?\n3. Why is the Directed Interaction module necessary when feature blending already occurs in the Directed Linear modules?\n4. How are the k-nearest neighbors computed? Is the full N x N distance matrix available?\n5. How could MLPs produce similar results as more efficient graph models with the same number of parameters?\n6. Is there rigorous rationale for averaging Pearson r, Spearman rho, and Kendall tau? Was there consideration of weighting the factors? Is there any redundancy between the metrics?\n7. How exactly were the synthetic task vectors sampled? Were the positive and negative areas of equal area? was the negative area less densely sampled?\n8. Will code and trained models be released?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose a graph neural network architecture that uses vectorized neurons, extensive hidden feature mixing operations, and global/local reference frame transformations to produce protein embeddings that are sensitive to or \"aware of\" residue-specific orientations within the larger macromolecule. They first introduce Directed Weights, combinations of multidimensional tensors that generalize scalars to vectors and vectors to higher-dimensional tensors. They then describe the Directed Weight Perceptron module, which performs numerous scalar and vector operations such as linear multiplication, dot products, and cross products. Armed with networks that more naturally describe data and relationships in 3D space, the authors next describe their SO(3) equivariant message passing paradigm. They then describe a few variants of their model adapted to mimic the architectures of recent graph models such as GCNs, GINs, and GATs. Finally, the authors demonstrate their models' performance on a number of protein-related tasks.", "strength_and_weaknesses": "This paper introduces a novel concept, but the presentation, design choices, and testing methodology could all be much clearer. Moving some of the information from the appendix to the text would clarify the role of various mechanisms presented. For example, discussing more directly the scalar and vector feature constructions earlier in the text and not leaving them to the appendix would justify the use of the separate scalar features in the first place, which may otherwise seem extraneous when compared to feature vectors that could be magnitudes larger. Second, it isn't clear why the Directed Interaction model is necessary. Its stated purpose is to establish a connection between the scalar and vector features, which seems redundant given the operations blending scalar and vector features in the Directed Linear module. Clarification on what this module provides that the latter does not would be helpful.\n\nThe experimental results generally seem reliable and convincing, but it is not immediately clear what configurations of models are being compared to each other. It even seems in some places that conflicting information is given i.e. in B.2 the DWP model is 1 layer for the Synthetic Task but B.3 \"Our model uses 4-layer message passing, where each message passing consists of 3-layer DWP\". The latter further seems redundant with \"there is 4 message passing...3-layer networks.\" One can tease these statements/cases apart, but more concise wording and clearer statement of model architecture/hyperparameters would be beneficial. Further, the synthetic task does not seem convincing without more information about how the points on a sphere were sampled, and statistics such as the variances of the positive and negative samples. Finally, it seems rather astounding that an MLP can match the performance of the most sophisticated models with a comparable number of weights on this task. The universal approximation theorem is relevant for arbitrarily deep and/or wide networks, so this is not sufficient justification. Similarly, it appears that GVP fails completely on this synthetic task, but is tested in later tasks whereas VNN performs well and is excluded without explanation.\n\nThe visualizations are informative but the table layout (tables 3-5) is confusing. There are also many citations to the arxiv version of papers that have been published in ML conference proceedings or other venues. The references should be carefully checked and corrected.\n\nSome specific questions:\n1. Why was GVP chosen over VNN for testing on benchmarks?\n2. What applications outside of protein embedding is this work relevant to?\n3. Why is the Directed Interaction module necessary when feature blending already occurs in the Directed Linear modules?\n4. How are the k-nearest neighbors computed? Is the full N x N distance matrix available?\n5. How could MLPs produce similar results as more efficient graph models with the same number of parameters?\n6. Is there rigorous rationale for averaging Pearson r, Spearman rho, and Kendall tau? Was there consideration of weighting the factors? Is there any redundancy between the metrics?\n7. How exactly were the synthetic task vectors sampled? Were the positive and negative areas of equal area? was the negative area less densely sampled?\n8. Will code and trained models be released?", "clarity,_quality,_novelty_and_reproducibility": "The proposed method is novel and the results look reasonable, but the paper lacks clarity and is missing some important details and justification of some of the modeling decisions.", "summary_of_the_review": "An interesting method, but the presentation lacks clarity and several design decisions are not well justified.\n\nWhat would improve my score: add clarifications and additional details.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667607533197}, {"id": "c9oEnWC7Vsu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1194/Reviewer_6r4X"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This is a work on rotationally equivariant feed forward networks for applications in protein folding and related tasks.\nA number of equivariant operations which appear useful for these applications are introduced, such as a cross product operator.\nExperiments are done on several useful tasks using protein structure databases, and significantly improved results are reported.", "review_text": "Valuable improvements to ML for protein folding.", "strengths": "Strengths: simple ideas, broad experimental testing, good performance improvement.\n\nWeaknesses: very empirical, leaves much for future work.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This is a work on rotationally equivariant feed forward networks for applications in protein folding and related tasks.\nA number of equivariant operations which appear useful for these applications are introduced, such as a cross product operator.\nExperiments are done on several useful tasks using protein structure databases, and significantly improved results are reported.", "strength_and_weaknesses": "Strengths: simple ideas, broad experimental testing, good performance improvement.\n\nWeaknesses: very empirical, leaves much for future work.", "clarity,_quality,_novelty_and_reproducibility": "I found the presentation clear and the results look convincing.   I am not really able to judge its originality.", "summary_of_the_review": "Valuable improvements to ML for protein folding.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666640397542}, {"id": "ywXin33IqnU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1194/Reviewer_kS63"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents a class of SO(3)-equivariant neural networks called Orientation-Aware Graph Neural Networks (OAGNNs). \nThe model is based on specific layers acting on scalar and vector features, designed to ensure better expressive power (compared to previous works) while maintaining SO(3) equivariance. \n\nCompared to previous related work, the main novelty of the proposed method is to have order-3 tensors of learnable weights acting on the features, effectively expanding classical perceptron weights from scalars to oriented vectors in 3d. \n\nIn the experimental section, the authors focus primarily on tasks related to structural biology. \nThis is a natural playground for SO(3) equivariant models, since many important characteristics of proteins can be described in terms of local frames irrespective of the global orientation of the molecules. \n\nAll results show that the proposed method is significantly better than the tested baselines, on almost all tasks and metrics considered. ", "review_text": "There are no major concerns with the paper, but there is margin for improvement. Extending the comparison with previous work is particularly important, since it is not clear whether the proposed method is simply making use of the extra parameters compared to VNN.\n\nI have recommended a weak acceptance, conditional on the authors addressing my concerns above. ", "strengths": "**Strengths**: \n\n- The paper is interesting and tackles an important problem. I expect that the paper could be of interest to the GNN, computer vision, and computational biology communities. \n- The results are strong and there are no concerns in terms of significance (although confidence intervals are not reported). The design of the model is validated also through ablation studies. \n\n**Weaknesses**: \n\n- Equation (6) implies that $c \\vec W \\in \\mathbb{R}^{C \\times 3}$ for $s \\in \\mathbb{R}^C$ and $\\vec W \\in \\mathbb{R}^{C \\times 3}$.  How is the product $c\\vec W$ defined here?\n- A lot of the geometrical intuition behind the paper is never formally stated. For example: \n\t- Why is equivariance to rigid transformations broken in Equations 7,8?\n\t- What does it mean that the network can \"sense the orientational features\"?\n\t- What set of transformations is the proposed method equivariant to, that VNNs weren't before? \n\t- What does it mean that the proposed method is \"more adaptive to the geometrically meaningful features?\"\n- In Experiment 5.1, do all models have a comparable number of parameters? DWP uses significantly more parameters than VNN, so this should be factored into the comparison. \n- Why is there no comparison with VNN in Experiment 5.2? Is there some reason that makes it impossible to design a GNN with vector neurons? This is important because GVP is the worst-performing baseline in Experiment 5.1, so the more interesting comparison is between OA and VNN. \n- Typos: \n\t- Missing parenthesis in the numerator of Eq. 24, right.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents a class of SO(3)-equivariant neural networks called Orientation-Aware Graph Neural Networks (OAGNNs). \nThe model is based on specific layers acting on scalar and vector features, designed to ensure better expressive power (compared to previous works) while maintaining SO(3) equivariance. \n\nCompared to previous related work, the main novelty of the proposed method is to have order-3 tensors of learnable weights acting on the features, effectively expanding classical perceptron weights from scalars to oriented vectors in 3d. \n\nIn the experimental section, the authors focus primarily on tasks related to structural biology. \nThis is a natural playground for SO(3) equivariant models, since many important characteristics of proteins can be described in terms of local frames irrespective of the global orientation of the molecules. \n\nAll results show that the proposed method is significantly better than the tested baselines, on almost all tasks and metrics considered. ", "strength_and_weaknesses": "**Strengths**: \n\n- The paper is interesting and tackles an important problem. I expect that the paper could be of interest to the GNN, computer vision, and computational biology communities. \n- The results are strong and there are no concerns in terms of significance (although confidence intervals are not reported). The design of the model is validated also through ablation studies. \n\n**Weaknesses**: \n\n- Equation (6) implies that $c \\vec W \\in \\mathbb{R}^{C \\times 3}$ for $s \\in \\mathbb{R}^C$ and $\\vec W \\in \\mathbb{R}^{C \\times 3}$.  How is the product $c\\vec W$ defined here?\n- A lot of the geometrical intuition behind the paper is never formally stated. For example: \n\t- Why is equivariance to rigid transformations broken in Equations 7,8?\n\t- What does it mean that the network can \"sense the orientational features\"?\n\t- What set of transformations is the proposed method equivariant to, that VNNs weren't before? \n\t- What does it mean that the proposed method is \"more adaptive to the geometrically meaningful features?\"\n- In Experiment 5.1, do all models have a comparable number of parameters? DWP uses significantly more parameters than VNN, so this should be factored into the comparison. \n- Why is there no comparison with VNN in Experiment 5.2? Is there some reason that makes it impossible to design a GNN with vector neurons? This is important because GVP is the worst-performing baseline in Experiment 5.1, so the more interesting comparison is between OA and VNN. \n- Typos: \n\t- Missing parenthesis in the numerator of Eq. 24, right.", "clarity,_quality,_novelty_and_reproducibility": "- Clarity: the paper is clear, although it could be more precise in explaining the main contributions. \n- Quality: the quality of the paper is good, and there are no major concerns regarding the design of the method. \n- Novelty: the main idea of the paper is novel, sensibly extending previous work. However, my suggestion is to compare the proposed method with VNNs in a more in-depth way (not just on one synthetic experiment, and also ensuring that the comparison is fair). \n- Reproducibility: it should be possible to implement the proposed method from the description, and experimental details are reported in the appendix. ", "summary_of_the_review": "There are no major concerns with the paper, but there is margin for improvement. Extending the comparison with previous work is particularly important, since it is not clear whether the proposed method is simply making use of the extra parameters compared to VNN.\n\nI have recommended a weak acceptance, conditional on the authors addressing my concerns above. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666371411685}], "openreview_url": "https://openreview.net/forum?id=WcTLZrpzfe", "arxiv_id": "2201.13299", "paper_pdf": "papers/WcTLZrpzfe.pdf", "paper_pdf_sha256": "bbbfb7b1c456222f86bebe6a7d19126a6fa3e902167dd10ada454684a4b822f5", "paper_pdf_bytes": 1797323, "paper_pdf_source": "openreview", "code_url": "https://github.com/Ced3-han/OAGNN", "code_repository": "Ced3-han/OAGNN", "code_commit": "63307a57c9c4d02ec4f1bee780c117e72c2b93c3", "code_archive": "repos/WcTLZrpzfe.zip", "code_archive_sha256": "bbd1ba1b98ea2ee611018ce54b190df26c727b0f6a0352aceb06a95a1026b63f", "code_archive_bytes": 34628, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 40, "github_languages": {"Python": 88581}, "github_archived": false, "github_pushed_at": "2025-05-27T14:14:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/directed-weight-neural-networks-for-protein"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Py8WbvKH_wv", "year": 2022, "status": "rejected", "title": "DRIBO: Robust Deep Reinforcement Learning via Multi-View Information Bottleneck", "authors": ["Jiameng Fan", "Wenchao Li"], "authorids": ["~Jiameng_Fan1", "~Wenchao_Li1"], "authors_source": "OpenReview API", "abstract": "Deep reinforcement learning (DRL) agents are often sensitive to visual changes that were unseen in their training environments. To address this problem, we leverage the sequential nature of RL to learn robust representations that encode only task-relevant information from observations based on the unsupervised multi-view setting. Specifically, we introduce a novel contrastive version of Multi-View Information Bottleneck (MIB) objective for temporal data. We train RL agents from pixels with this auxiliary objective to learn robust representations that can compress away task-irrelevant information and are predictive of task-relevant dynamics. This approach enables us to train high-performance policies that are robust to visual distractions and can generalize well to unseen environments. We demonstrate that our approach can achieve SOTA performance on diverse visual control tasks on the DeepMind Control Suite when the background is replaced with natural videos. In addition, we show that our approach outperforms well-established baselines for generalization to unseen environments on the Procgen benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "46d63rAXqk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1213/Reviewer_LDpb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to learn robust representation from the replay buffer for reinforcement learning. The key idea is to leverage the concept of mutual information and the InfoNCE tool to compute the mutual information as a regularizer (a.k.a. DRIBO loss). The authors conducted experiments on some standard benchmarks (DeepMind Visual Control Suite and ProcGen). ", "review_text": "Strength:\n- To me, the final experiment result on ProcGen is impressive.\n- Supplementary material shows that the authors seem to have digged into the experiment data with some depth. \n\nWeakness:\n- The writing is not so clean. Needs polishing or even rewriting.\n  - The paper introduction is quite vague and covers very general ideas, lacking specificity and advantage to the proposed approach; \n  - In Eq (5) and (6), there abruptly comes the Lagrange multiplier. Many other places also show that the writing is very rough.\n- The results on DeepMind Control Suite are somewhat marginal improvement. \n- While the results on ProcGen are impressive, the analysis and interpretation into these results are somewhat thin. Because the method involves many differences in comparison to the baselines, it is unclear whether the proposed main idea, mutual information loss, is the source of the major contributions. To me, the breakdown of contributions of each design choice (esp. including fair comparison with baselines with the same type and level of data augmentation) is important. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to learn robust representation from the replay buffer for reinforcement learning. The key idea is to leverage the concept of mutual information and the InfoNCE tool to compute the mutual information as a regularizer (a.k.a. DRIBO loss). The authors conducted experiments on some standard benchmarks (DeepMind Visual Control Suite and ProcGen). ", "main_review": "Strength:\n- To me, the final experiment result on ProcGen is impressive.\n- Supplementary material shows that the authors seem to have digged into the experiment data with some depth. \n\nWeakness:\n- The writing is not so clean. Needs polishing or even rewriting.\n  - The paper introduction is quite vague and covers very general ideas, lacking specificity and advantage to the proposed approach; \n  - In Eq (5) and (6), there abruptly comes the Lagrange multiplier. Many other places also show that the writing is very rough.\n- The results on DeepMind Control Suite are somewhat marginal improvement. \n- While the results on ProcGen are impressive, the analysis and interpretation into these results are somewhat thin. Because the method involves many differences in comparison to the baselines, it is unclear whether the proposed main idea, mutual information loss, is the source of the major contributions. To me, the breakdown of contributions of each design choice (esp. including fair comparison with baselines with the same type and level of data augmentation) is important. ", "summary_of_the_review": "Although there might be some good core contribution, the paper in its current state is not quite ready for publication. See details in my Main Review.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636331646936}, {"id": "bX1kQc3ySwn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1213/Reviewer_kUtq"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper tackles the problem of generalization amidst visual distractors for control tasks. In particular, the distractors have no dependence on the optimal policy and thus clearly form a task-irrelevant component. The proposal is to use mutual information between two views as a proxy for how much task-relevant information is present in the constructed representation. This objective is adapted for RL to consider the long term sequential nature. Finally, the authors test this approach on ProcGen and DMC Suite with distractors, while also performing certain ablations. ", "review_text": "The paper is well motivated and the main problem is interesting in my opinion. The related work is well covered and most of the writing flows well. Experimenting over two benchmarks (ProcGen + DMC) is quite extensive relative to other similar papers. \n\nI start to have issues in understanding around Definition 1. I get the overall concept, that we can deconstruct the overall mutual information in a task-relevant and a task-irrelevant terms. However, why should we maximize the sum of both (as noted in the 5th line below Eq. 2)? You say maximizing I(O, S) gets you a sufficient representation. That would mean maximizing the first term as well. But you also say that the first term needs to be minimized. This seems contradictory. I understand Theorem 1 and also get the derivation of the components of Eq. 5 and 6. However in Eq. 5, do we want to maximize both terms? One is the task irrelevant term (aka minimality) while the other is the task relevant term (aka sufficiency). They should have opposite signs right (as is in Equation 7)? \n\nFinally, regardless of the derivation of the DRIBO loss, my main concern is the empirical results here. The scores reported for DMC are not accurate in my opinion. I have run RAD on the distractor suite and it gets much better performance than what is reported. Why is RAD performance this low? DRIBO is using augmentations from RAD to generate the multi-view observations, and so as a first baseline, it should be performing better than RAD. Note that RAD achieves the same score (roughly) as reported in the original paper even when run with distractors. Can the authors clarify this?\n\nMoreover, the performance reported for the proposed method is not good enough. In particular, running SAC with a reward prediction head achieves similar or even better performance than of the proposed method. In light of these two observations, I am not convinced that DRIBO really is doing more than a simple baseline, and certainely not doing so much better than RAD.\n\nFor the ProcGen experiments, why don't you compare with DAAC [1] and IDAAC [1] since I believe they are the state of the art methods on ProcGen currently.\n\nWhy compare the t-sne with CURL? We already know that CURL does not do a good job and that RAD is much better and even simpler. Also, in Figure 3, CURL is outperforming RAD in some of the tasks, which I think is again almost certainly not the case (based on the experiments I have run with RAD).\n\nReferences:\n\n[1] Decoupling Value and Policy for Generalization in Reinforcement Learning", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper tackles the problem of generalization amidst visual distractors for control tasks. In particular, the distractors have no dependence on the optimal policy and thus clearly form a task-irrelevant component. The proposal is to use mutual information between two views as a proxy for how much task-relevant information is present in the constructed representation. This objective is adapted for RL to consider the long term sequential nature. Finally, the authors test this approach on ProcGen and DMC Suite with distractors, while also performing certain ablations. ", "main_review": "The paper is well motivated and the main problem is interesting in my opinion. The related work is well covered and most of the writing flows well. Experimenting over two benchmarks (ProcGen + DMC) is quite extensive relative to other similar papers. \n\nI start to have issues in understanding around Definition 1. I get the overall concept, that we can deconstruct the overall mutual information in a task-relevant and a task-irrelevant terms. However, why should we maximize the sum of both (as noted in the 5th line below Eq. 2)? You say maximizing I(O, S) gets you a sufficient representation. That would mean maximizing the first term as well. But you also say that the first term needs to be minimized. This seems contradictory. I understand Theorem 1 and also get the derivation of the components of Eq. 5 and 6. However in Eq. 5, do we want to maximize both terms? One is the task irrelevant term (aka minimality) while the other is the task relevant term (aka sufficiency). They should have opposite signs right (as is in Equation 7)? \n\nFinally, regardless of the derivation of the DRIBO loss, my main concern is the empirical results here. The scores reported for DMC are not accurate in my opinion. I have run RAD on the distractor suite and it gets much better performance than what is reported. Why is RAD performance this low? DRIBO is using augmentations from RAD to generate the multi-view observations, and so as a first baseline, it should be performing better than RAD. Note that RAD achieves the same score (roughly) as reported in the original paper even when run with distractors. Can the authors clarify this?\n\nMoreover, the performance reported for the proposed method is not good enough. In particular, running SAC with a reward prediction head achieves similar or even better performance than of the proposed method. In light of these two observations, I am not convinced that DRIBO really is doing more than a simple baseline, and certainely not doing so much better than RAD.\n\nFor the ProcGen experiments, why don't you compare with DAAC [1] and IDAAC [1] since I believe they are the state of the art methods on ProcGen currently.\n\nWhy compare the t-sne with CURL? We already know that CURL does not do a good job and that RAD is much better and even simpler. Also, in Figure 3, CURL is outperforming RAD in some of the tasks, which I think is again almost certainly not the case (based on the experiments I have run with RAD).\n\nReferences:\n\n[1] Decoupling Value and Policy for Generalization in Reinforcement Learning", "summary_of_the_review": "The paper is motivated by an interesting problem to the community, proposes a fairly novel method for leraning better representations, but there remain certain gaps in my understanding currently. I hope the authors can clarify this during the rebuttal. In any case, the experimental results are not correct in my prior experience of running these exact methods. For this reason, I am advocating for a reject. I'm happy to revise my review if these issues can be addressed. \n\n\n-------------------------- Post Rebuttal ----------------------------------\n\nThe new results provided by the authors convince me fairly that the method is somewhat useful. I am still skeptical about the RAD results and the overall scores reported for DMC tasks, since the absolute numbers are not state-of-the-art. This might be due to a different evaluation setting where the distractor videos are changing across training, and at test time a video is drawn from a different distribution. I am unsure how much this sort of an evaluation scheme affects RAD's results, which should produce much better numbers than reported. I am giving the benefit of doubt here to the authors and hoping that they have done a fair evaluation for the baseline RAD method. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635853441791}, {"id": "oPCD8tE2-_j", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1213/Reviewer_m1E2"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problem of pixel-based control with reinforcement learning. The authors categorize observation changes into task-relevant and task-irrelevant changes. They define task-irrelevant changes as those do not have casual relations with actions, and introduce a conditional prior to compress task-irrelevant information, such that agent can focus on task-relevant information. The authors evaluated the proposed DRIBO method on pixel-based control tasks (DM Control) with natural video background and the procgen suite.", "review_text": "Strengths:\n* The paper shows a technically approach to compress task-irrelevant information, making the agent focus on task-relevant information.\n* The paper shows good empirical improvements on DM Control with natural video background and procgen, compared to existing methods in system-wise comparisons.\n* Comprehensive ablation studies on DM Control that identify the importance of history length and information compression.\n\nWeaknesses:\n* It’s not clear to me whether there are as many task-irrelevant observation changes in procgen as in DM Control with video background. The generalization requirement in procgen seems to be different. I would like to see if the authors can clarify how the procgen experiment align with the main story of this paper. \n* Many of DRIBO’s design choices (using RSSM, long history length, etc) are different to existing methods. The system-wise comparison on procgen (table 1) doesn’t provide much insight on why DRIBO improves generalization on procgen. \n* The authors claim that $I(S_t^{(1)};S_t^{(2)}|S_{t-1},A_{t-1})$ can be maximized by maximizing $I(S_t^{(1)};S_t^{(2)})$. I don't think this is mathematically correct without making certain assumptions.\n\nSmaller presentation issues:\n* In Sec.3, the author states that \"task-irrelevant information does not contribute to the choice of actions\". I think it should be \"task-irrelevant information does not affected by choice of actions\".\n* The author parameterizes mutual information as like $I_\\theta$. Mutual information is a true measurement and can't be parameterized. Only MI estimates can be parameterized.\n* CEB is a generic information bottleneck. It's incorrect to say that it cannot be applied to sequential data.\n* The appendix shows in the main paper file.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the problem of pixel-based control with reinforcement learning. The authors categorize observation changes into task-relevant and task-irrelevant changes. They define task-irrelevant changes as those do not have casual relations with actions, and introduce a conditional prior to compress task-irrelevant information, such that agent can focus on task-relevant information. The authors evaluated the proposed DRIBO method on pixel-based control tasks (DM Control) with natural video background and the procgen suite.", "main_review": "Strengths:\n* The paper shows a technically approach to compress task-irrelevant information, making the agent focus on task-relevant information.\n* The paper shows good empirical improvements on DM Control with natural video background and procgen, compared to existing methods in system-wise comparisons.\n* Comprehensive ablation studies on DM Control that identify the importance of history length and information compression.\n\nWeaknesses:\n* It’s not clear to me whether there are as many task-irrelevant observation changes in procgen as in DM Control with video background. The generalization requirement in procgen seems to be different. I would like to see if the authors can clarify how the procgen experiment align with the main story of this paper. \n* Many of DRIBO’s design choices (using RSSM, long history length, etc) are different to existing methods. The system-wise comparison on procgen (table 1) doesn’t provide much insight on why DRIBO improves generalization on procgen. \n* The authors claim that $I(S_t^{(1)};S_t^{(2)}|S_{t-1},A_{t-1})$ can be maximized by maximizing $I(S_t^{(1)};S_t^{(2)})$. I don't think this is mathematically correct without making certain assumptions.\n\nSmaller presentation issues:\n* In Sec.3, the author states that \"task-irrelevant information does not contribute to the choice of actions\". I think it should be \"task-irrelevant information does not affected by choice of actions\".\n* The author parameterizes mutual information as like $I_\\theta$. Mutual information is a true measurement and can't be parameterized. Only MI estimates can be parameterized.\n* CEB is a generic information bottleneck. It's incorrect to say that it cannot be applied to sequential data.\n* The appendix shows in the main paper file.", "summary_of_the_review": "The approach is novel. The paper shows good empirical results on DM Control with natural video background and procgen. However, it's not clear to me why the procgen-type of generalization and the procgen experiment are relevant to the main story and I have some concerns about the mathematical correctness. I'm slightly leaning to acceptance at this point, but I hope the authors would be able to clarify. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635717388396}, {"id": "s3QP4Mp83x9", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1213/Reviewer_uNzG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "Motivated by the fact that RL agents are sensitive to unseen environments, this paper proposes a method to separate task-relevant and task-irrelevant information from observations based on unsupervised multi-view settings. They use data augmentation methods to create multi-view data for calculating contrastive Multi-View Information Bottleneck (MIB) and use it as the objective to increase the generalization of model representation.", "review_text": "1. Strengths and Weaknesses:\n\n\\+ The perspective of using mutual information to increase the robustness is interesting. The authors provide clear analyses of how to use mutual information to learn task-irrelevant information based on multi-view settings. \n\n\\+ The proposed method is a plug-in penalty that can be integrated into many modern RL algorithms. \n\n\\- The representation encoder and the estimator of mutual information (using InfoNCE) are borrowed from existing works, which may limit the contribution of this work.\n\n\\- The proposed methods show advantages in DMC and ProcGen environments. However, the settings in both environments are very similar: changing the background image in the same task. I am not sure if there are any real-world tasks that can be better solved by the proposed method. It would be great if the author can conduct experiments on other realistic tasks.\n\n\n2. General questions\n\n(1)\tHow is the data augmentation conducted to create the multi-view setting? For example, in Figure 1, is the entire image rotated or only the background rotated? If the entire image is rotated, will the state change and the action not be consistent with the state?\n\n(2)\tHow to create the multi-view setting for layouts in the ProcGen suite? In my understanding, the layout of environments is corresponding to the difficulty level. How to conduct image augmentation on such difficulty levels? Or, do the author only change the background image in ProcGen? \n\n(3)    Based on (2), I am also wondering if this method can be extended to more general settings to increase the robustness?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Motivated by the fact that RL agents are sensitive to unseen environments, this paper proposes a method to separate task-relevant and task-irrelevant information from observations based on unsupervised multi-view settings. They use data augmentation methods to create multi-view data for calculating contrastive Multi-View Information Bottleneck (MIB) and use it as the objective to increase the generalization of model representation.", "main_review": "1. Strengths and Weaknesses:\n\n\\+ The perspective of using mutual information to increase the robustness is interesting. The authors provide clear analyses of how to use mutual information to learn task-irrelevant information based on multi-view settings. \n\n\\+ The proposed method is a plug-in penalty that can be integrated into many modern RL algorithms. \n\n\\- The representation encoder and the estimator of mutual information (using InfoNCE) are borrowed from existing works, which may limit the contribution of this work.\n\n\\- The proposed methods show advantages in DMC and ProcGen environments. However, the settings in both environments are very similar: changing the background image in the same task. I am not sure if there are any real-world tasks that can be better solved by the proposed method. It would be great if the author can conduct experiments on other realistic tasks.\n\n\n2. General questions\n\n(1)\tHow is the data augmentation conducted to create the multi-view setting? For example, in Figure 1, is the entire image rotated or only the background rotated? If the entire image is rotated, will the state change and the action not be consistent with the state?\n\n(2)\tHow to create the multi-view setting for layouts in the ProcGen suite? In my understanding, the layout of environments is corresponding to the difficulty level. How to conduct image augmentation on such difficulty levels? Or, do the author only change the background image in ProcGen? \n\n(3)    Based on (2), I am also wondering if this method can be extended to more general settings to increase the robustness?", "summary_of_the_review": "In general, I think the proposed idea of this paper is interesting, and the theoretical analysis seems correct to me. However, I still have questions about the details of how to implement the multi-view settings (In my general question (1)). Also, the experiment part seems a little weak since the two environment suites are very similar. \n\nI think this paper is marginally above the borderline. I may raise my score if the author provides reasonable answers to my question.\n\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635565231473}], "openreview_url": "https://openreview.net/forum?id=Py8WbvKH_wv", "arxiv_id": "2102.13268", "paper_pdf": "papers/Py8WbvKH_wv.pdf", "paper_pdf_sha256": "d6b14c9f4426cb524a4a2e5fc0b41a735fe1446b6921365c8626b3e1bad3dba0", "paper_pdf_bytes": 5377900, "paper_pdf_source": "openreview", "code_url": "https://github.com/BU-DEPEND-Lab/DRIBO", "code_repository": "BU-DEPEND-Lab/DRIBO", "code_commit": "bb3374e2ae861ed9aaf7606b43a027247f3fabc6", "code_archive": "repos/Py8WbvKH_wv.zip", "code_archive_sha256": "b7bfaa71ddba14541a5c0c724c44277a0752d054c1aae9d33e9afd9b23c523f9", "code_archive_bytes": 42557, "code_file_count": 21, "code_extensions": {".py": 15, ".sh": 6}, "github_disk_usage_kb": 36, "github_languages": {"Python": 136324, "Shell": 13326}, "github_archived": false, "github_pushed_at": "2022-06-14T13:51:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/robust-deep-reinforcement-learning-via-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "w5bNwUzj33", "year": 2021, "status": "rejected", "title": "Cross-Domain Few-Shot Learning by Representation Fusion", "authors": ["Thomas Adler", "Johannes Brandstetter", "Michael Widrich", "Andreas Mayr", "David Kreil", "Michael K Kopp", "Günter Klambauer", "Sepp Hochreiter"], "authorids": ["~Thomas_Adler1", "~Johannes_Brandstetter1", "~Michael_Widrich2", "~Andreas_Mayr2", "openreview20@kreil.org", "~Michael_K_Kopp1", "~Günter_Klambauer1", "~Sepp_Hochreiter1"], "authors_source": "OpenReview API", "abstract": "In order to quickly adapt to new data, few-shot learning aims at learning from few examples, often by using already acquired knowledge. The new data often differs from the previously seen data due to a domain shift, that is, a change of the input-target distribution. While several methods perform well on small domain shifts like new target classes with similar inputs, larger domain shifts are still challenging. Large domain shifts may result in abstract concepts that are not shared between the original and the new domain. However, low-level concepts like edges in images might still be shared and useful. For cross-domain few-shot learning, we suggest representation fusion to unify different abstraction levels of a deep neural network into one representation. We propose Cross-domain Hebbian Ensemble Few-shot learning (CHEF), which consists of representation fusion by an ensemble of Hebbian learners acting on different layers of a deep neural network that was trained on the original domain. On the few-shot datasets miniImagenet and tieredImagenet, where the domain shift is small, CHEF is competitive with state-of-the-art methods. On cross-domain few-shot benchmark challenges with larger domain shifts, CHEF obtains state-of-the-art results in all categories. We further apply CHEF on a real-world cross-domain application in drug discovery. We consider a domain shift from bioactive molecules to environmental chemicals and drugs with twelve associated toxicity prediction tasks. On these tasks that are highly relevant for computational drug discovery, CHEF significantly outperforms all its competitors. ", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "RPOjACdR9qv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1471/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper primarily deals with cross-domain few-shot learning. Under this setting, there is a large shift in domain going from the meta-train dataset to the few-shot datasets. Inspired by previous work, the authors argue that high-level concepts might not be useful in this setting but low-level concepts like edges, textures and shapes can be utilized. They propose a Cross-domain Hebbian Ensemble Few-shot (CHEF) learner, that learns an ensemble of classifiers at multiple levels of a deep neural network, thus making use of both low and high level concepts. Experimental results show that CHEF does better, in most cases, than learning a separate classifier at a given level. They show results under the cross-domain and the standard few-shot setting.\n\nPros:\n1. Utilizing low and high level concepts is a simple technique to boost few-shot learning performance.\n2. CHEF does not require any updates to the weights of the model backbone. It learns additional weights for classification.\n\nCons:\n1. The paper is missing details. The authors talk about Hebbian learning, FID, etc. but do not give details about it. The experimental set-up is missing information about how the models are trained and tested.\n2. Is the proposed algorithm a Hebbian learner? The update in Equations 1 and 2 is a standard gradient descent update.\n3. Using 2 fully-connected layers changes the model backbone from ResNet-x to a ResNet-(x+1). This should be clearly noted in Table 2 and 3.\n\nClarifications:\n1. For the experiments, the softmax output layer has as many units as the number of classes in the meta-train and meta-validation sets combined. Does this mean that the pre-training is done on the sets combined? If so, this is not an apples-to-apples comparison. If not, the model does not know the difference between the classes in the meta-validation set. How does having these classes in the validation set help while pre-training?\n\nNotes:\n1. Mini-ImageNet and Tiered-ImageNet do involve general domain shifts. Even though p(x) does not change much going from meta-train dataset to the few-shot datasets, the samples seen in the two scenarios are disjoint.\n2. The Appendix should be cleaned up.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple algorithm but missing many key details", "review": "Summary:\n\nThis paper primarily deals with cross-domain few-shot learning. Under this setting, there is a large shift in domain going from the meta-train dataset to the few-shot datasets. Inspired by previous work, the authors argue that high-level concepts might not be useful in this setting but low-level concepts like edges, textures and shapes can be utilized. They propose a Cross-domain Hebbian Ensemble Few-shot (CHEF) learner, that learns an ensemble of classifiers at multiple levels of a deep neural network, thus making use of both low and high level concepts. Experimental results show that CHEF does better, in most cases, than learning a separate classifier at a given level. They show results under the cross-domain and the standard few-shot setting.\n\nPros:\n1. Utilizing low and high level concepts is a simple technique to boost few-shot learning performance.\n2. CHEF does not require any updates to the weights of the model backbone. It learns additional weights for classification.\n\nCons:\n1. The paper is missing details. The authors talk about Hebbian learning, FID, etc. but do not give details about it. The experimental set-up is missing information about how the models are trained and tested.\n2. Is the proposed algorithm a Hebbian learner? The update in Equations 1 and 2 is a standard gradient descent update.\n3. Using 2 fully-connected layers changes the model backbone from ResNet-x to a ResNet-(x+1). This should be clearly noted in Table 2 and 3.\n\nClarifications:\n1. For the experiments, the softmax output layer has as many units as the number of classes in the meta-train and meta-validation sets combined. Does this mean that the pre-training is done on the sets combined? If so, this is not an apples-to-apples comparison. If not, the model does not know the difference between the classes in the meta-validation set. How does having these classes in the validation set help while pre-training?\n\nNotes:\n1. Mini-ImageNet and Tiered-ImageNet do involve general domain shifts. Even though p(x) does not change much going from meta-train dataset to the few-shot datasets, the samples seen in the two scenarios are disjoint.\n2. The Appendix should be cleaned up.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604812125651}, {"id": "-n9Q0hm_Y19", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1471/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors focus on cross-domain few-shot learning in the case of large source-target domain shifts. In particular, a new Cross-domain Hebbian Ensemble Few-shot (CHEF) learning method is proposed that performs representation fusion using an ensemble of Hebbian learners on different layers of a DNN trained on the source domain. The proposed CHEF method is validated on classification benchmark datasets with smaller domain shifts (miniImagenet and tieredImagenet) and larger domain shifts (drug discovery, ChEMBL20), and it can outperform related SOTA methods, especially with larger shifts.   \n\n+ The paper is clearly written and generally well organized, and all the key concepts are detailed, although I found some parts of Section 3 difficult to follow.  Even with Figure 1 and Algorithm 1, the paper was not easy to follow. The section on related work is very condensed, without much critical analysis. Therefore, it is not clear how their CHEF is motivated by challenges in literature.  The literature e review for methods on DA is not up to date, and not reflect SOTA methods on deep DA. \n\n+ The code is made available as supplementary material, so the results in this paper should be reproducible by a reader.\n\n+  The supplementary material provides additional information that should be useful to the reader (experimental setup and results).  \n\n+ The authors present many interesting results in Section 4, and they are for the most part convincing. They do present averages results over independent replications, using some cross-validation process. Using the FID to measure the domain shifts is excellent. However, I am not convinced about the results shown in Section 4.3. If I understand, Table 4 compares a deep NN (FCN). \n\n+ CHEF with some conventional ML models (SVM and RF)?  This experiment needs some clarification. Their model could be also compared with SOA methods in terms of time and/or memory complexity.     \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Cross-Domain Few-Shot Learning by Representation Fusion", "review": "In this paper, the authors focus on cross-domain few-shot learning in the case of large source-target domain shifts. In particular, a new Cross-domain Hebbian Ensemble Few-shot (CHEF) learning method is proposed that performs representation fusion using an ensemble of Hebbian learners on different layers of a DNN trained on the source domain. The proposed CHEF method is validated on classification benchmark datasets with smaller domain shifts (miniImagenet and tieredImagenet) and larger domain shifts (drug discovery, ChEMBL20), and it can outperform related SOTA methods, especially with larger shifts.   \n\n+ The paper is clearly written and generally well organized, and all the key concepts are detailed, although I found some parts of Section 3 difficult to follow.  Even with Figure 1 and Algorithm 1, the paper was not easy to follow. The section on related work is very condensed, without much critical analysis. Therefore, it is not clear how their CHEF is motivated by challenges in literature.  The literature e review for methods on DA is not up to date, and not reflect SOTA methods on deep DA. \n\n+ The code is made available as supplementary material, so the results in this paper should be reproducible by a reader.\n\n+  The supplementary material provides additional information that should be useful to the reader (experimental setup and results).  \n\n+ The authors present many interesting results in Section 4, and they are for the most part convincing. They do present averages results over independent replications, using some cross-validation process. Using the FID to measure the domain shifts is excellent. However, I am not convinced about the results shown in Section 4.3. If I understand, Table 4 compares a deep NN (FCN). \n\n+ CHEF with some conventional ML models (SVM and RF)?  This experiment needs some clarification. Their model could be also compared with SOA methods in terms of time and/or memory complexity.     \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604374070229}, {"id": "TZ492T9APT2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1471/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduce a learning mechanism that combining few-shot domain adaptation with a Hebbian learning rule. Basically, the authors fused multiple layer feature representations in weak learner and ensemble the classification results. This approach is trivial. I would suggest the author can introduce the benefit or provide the reason a Hebbian learning can improve adaptation performance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novelty is limited", "review": "This paper introduce a learning mechanism that combining few-shot domain adaptation with a Hebbian learning rule. Basically, the authors fused multiple layer feature representations in weak learner and ensemble the classification results. This approach is trivial. I would suggest the author can introduce the benefit or provide the reason a Hebbian learning can improve adaptation performance.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603946971602}, {"id": "-tbYnK-19bh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1471/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes cross-domain Hebbian ensemble few-shot learning or CHEF which achieves representation fusion by an ensemble of Hebbian learners acting on different layers of a deep neural network that was trained on the original domain and aims at learning from few examples, often by using already acquired knowledge. The experiments show results on miniImageNet and tieredImageNet, where the domain shift is small and also on the cross domain few-shot benchmarks with larger domain shifts. The paper also shows auxiliary experiments on drug discovery. I have the following comments on the paper:\n\nmajor comments\n1. How do you compare the Hebbian ensemble learning strategy with other ensemble learning strategies, such as random forest and boosting? Is it possible to do a comparison between those?\n\n2. The equation (1) basically depicts the Hebbian learning rule where V is the matrix of postsynaptic responses v_i which are effectively the gradient of the loss in equation (2). I wonder how this rule (equation 1) differs in various layers? I also wonder what does combining several Hebbian learners mean in the case of few-shot learning?\n\n3. Is it possible to plot the datasets with T-SNE or PCA and present it beside Table 1 for a clear understanding of the dataset characteristics? I understand those are approximate methods, but it is interesting to see coherence of the plot with FID.\n\nBased on my current understanding and the above comments, I currently recommend the paper as \"marginally below acceptance threshold\". I would like to hear clarification on Hebbian ensemble learning.\n\nminor comments\n1. A reference to Hebbian ensemble learning will be useful.\n2. A reference to Fréchet-Inception-Distance (FID) will also be useful.\n3. In ICLR 2020, there were few works that proposed to learn mutual information from diverse domains. I think it is worth to provide to have a discussion on them.\n(i) M. Federici et al., Learning Robust Representations via Multi-View Information Bottleneck, ICLR, 2020.\n(ii) M. Tschannen et al., On Mutual Information Maximization for Representation Learning, ICLR, 2020.\n4. In figure 2, the labels along the x axis are confusingly aligned. I think it is better to make them exactly perpendicular to the x axis.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Details on Hebbian Ensemble Learning", "review": "This paper proposes cross-domain Hebbian ensemble few-shot learning or CHEF which achieves representation fusion by an ensemble of Hebbian learners acting on different layers of a deep neural network that was trained on the original domain and aims at learning from few examples, often by using already acquired knowledge. The experiments show results on miniImageNet and tieredImageNet, where the domain shift is small and also on the cross domain few-shot benchmarks with larger domain shifts. The paper also shows auxiliary experiments on drug discovery. I have the following comments on the paper:\n\nmajor comments\n1. How do you compare the Hebbian ensemble learning strategy with other ensemble learning strategies, such as random forest and boosting? Is it possible to do a comparison between those?\n\n2. The equation (1) basically depicts the Hebbian learning rule where V is the matrix of postsynaptic responses v_i which are effectively the gradient of the loss in equation (2). I wonder how this rule (equation 1) differs in various layers? I also wonder what does combining several Hebbian learners mean in the case of few-shot learning?\n\n3. Is it possible to plot the datasets with T-SNE or PCA and present it beside Table 1 for a clear understanding of the dataset characteristics? I understand those are approximate methods, but it is interesting to see coherence of the plot with FID.\n\nBased on my current understanding and the above comments, I currently recommend the paper as \"marginally below acceptance threshold\". I would like to hear clarification on Hebbian ensemble learning.\n\nminor comments\n1. A reference to Hebbian ensemble learning will be useful.\n2. A reference to Fréchet-Inception-Distance (FID) will also be useful.\n3. In ICLR 2020, there were few works that proposed to learn mutual information from diverse domains. I think it is worth to provide to have a discussion on them.\n(i) M. Federici et al., Learning Robust Representations via Multi-View Information Bottleneck, ICLR, 2020.\n(ii) M. Tschannen et al., On Mutual Information Maximization for Representation Learning, ICLR, 2020.\n4. In figure 2, the labels along the x axis are confusingly aligned. I think it is better to make them exactly perpendicular to the x axis.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603891882376}, {"id": "7YggefBKGf", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1471/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:  This paper proposes a domain-shift problem using fewer training examples. It suggests representation fusion as the concept of unifying and merging information from different layers of abstraction. \nCross-domain Hebbian Ensemble Few-shot learning (CHEF) is introduced for extracting features using representation fusion.  More importantly, CHEF does not need to backpropagate information through the backbone network.   CHEF  is applied to various cross-domain few-shot tasks and cross-domain real-world applications from drug discovery.\n\nStrong Points: 1-  The paper is well organized and easy to understand.\n2-  The model is evaluated in four benchmark datasets CropDisease, EuroSAT,  ISIC2018, and ChestX.  It also conducted experiments on two large scale datasets (prepared from ImageNet dataset), miniImagenet and tieredImagenet.  The proposed model shows consistent and promising performance in all datasets.  \n\n3- It introduced Hebbian learners for feature fusion that does not require backpropagation of error signals through the entire backbone network.  Only the parameters of the Hebbian learners need adjustment. Therefore it is speedy and versatile.\n\nWeaknesses: 1- The crucial contribution of this work is Hebbian Learner. Therefore it should devote more space for Hebbian Learner. The explanation about Hebbian Learner provided in this paper is not sufficient understanding to propose an approach clearly. I recommend the authors that include more description of Hebbian Learner.  I understand the space limit, but the model's crucial and essential contribution should be included in the main paper.\n2- This paper claims that using \"Hebbian Learner makes CHEF extremely fast,\" but I did not find any supporting experiments in the main paper to prove this statement. It should include results on time comparison.\n3- This paper performed the experiments on a few-shot learning setup for that chosen 5, 20, and 50 examples per class to train the model and observed continuous improvement on model performance. To show the limit of model performance, the experiments should also be performed using all available examples in the datasets.\n4-  authors have selected ResNet-12 as the backbone model; any specific reason for this? I wonder to see the model performance for more deep networks like ResNet-101 as the backbone model.\n5- It would be better to show the individual contribution of the Hebbian Learner. Therefore result should also be included in the ablation analysis section without using Hebbian Learner in the same setting.\n6-   On page#4, the authors have mentioned that \"We combine the NK feature vectors into a matrix Z ∈ R^ NK×D and initialize a weight matrix W ∈ R ^K×D.\"  But it is not mentioned about the initialization technique. Random initialization, Xavier initialization, etc. ?\n7- Some essential baseline approaches are missing for comparison, such as  \"Few-Shot Adversarial Domain Adaptation\" by Saeid Motiian et al. NIPS 2017.\n\n\nOverall: The paper needs to include many things for better clarity. I feel it provides some insufficient information or does not clearly explain the proposed model's crucial contribution. It needs to include experimental results for different settings, as I mentioned in the weaknesses section, to prove the model's efficacy. \n\n\n\n\n \n\n\n \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "CROSS-DOMAIN FEW-SHOT LEARNING BY REPRESENTATION FUSION", "review": "Summary:  This paper proposes a domain-shift problem using fewer training examples. It suggests representation fusion as the concept of unifying and merging information from different layers of abstraction. \nCross-domain Hebbian Ensemble Few-shot learning (CHEF) is introduced for extracting features using representation fusion.  More importantly, CHEF does not need to backpropagate information through the backbone network.   CHEF  is applied to various cross-domain few-shot tasks and cross-domain real-world applications from drug discovery.\n\nStrong Points: 1-  The paper is well organized and easy to understand.\n2-  The model is evaluated in four benchmark datasets CropDisease, EuroSAT,  ISIC2018, and ChestX.  It also conducted experiments on two large scale datasets (prepared from ImageNet dataset), miniImagenet and tieredImagenet.  The proposed model shows consistent and promising performance in all datasets.  \n\n3- It introduced Hebbian learners for feature fusion that does not require backpropagation of error signals through the entire backbone network.  Only the parameters of the Hebbian learners need adjustment. Therefore it is speedy and versatile.\n\nWeaknesses: 1- The crucial contribution of this work is Hebbian Learner. Therefore it should devote more space for Hebbian Learner. The explanation about Hebbian Learner provided in this paper is not sufficient understanding to propose an approach clearly. I recommend the authors that include more description of Hebbian Learner.  I understand the space limit, but the model's crucial and essential contribution should be included in the main paper.\n2- This paper claims that using \"Hebbian Learner makes CHEF extremely fast,\" but I did not find any supporting experiments in the main paper to prove this statement. It should include results on time comparison.\n3- This paper performed the experiments on a few-shot learning setup for that chosen 5, 20, and 50 examples per class to train the model and observed continuous improvement on model performance. To show the limit of model performance, the experiments should also be performed using all available examples in the datasets.\n4-  authors have selected ResNet-12 as the backbone model; any specific reason for this? I wonder to see the model performance for more deep networks like ResNet-101 as the backbone model.\n5- It would be better to show the individual contribution of the Hebbian Learner. Therefore result should also be included in the ablation analysis section without using Hebbian Learner in the same setting.\n6-   On page#4, the authors have mentioned that \"We combine the NK feature vectors into a matrix Z ∈ R^ NK×D and initialize a weight matrix W ∈ R ^K×D.\"  But it is not mentioned about the initialization technique. Random initialization, Xavier initialization, etc. ?\n7- Some essential baseline approaches are missing for comparison, such as  \"Few-Shot Adversarial Domain Adaptation\" by Saeid Motiian et al. NIPS 2017.\n\n\nOverall: The paper needs to include many things for better clarity. I feel it provides some insufficient information or does not clearly explain the proposed model's crucial contribution. It needs to include experimental results for different settings, as I mentioned in the weaknesses section, to prove the model's efficacy. \n\n\n\n\n \n\n\n \n\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603745146163}], "openreview_url": "https://openreview.net/forum?id=w5bNwUzj33", "arxiv_id": "2010.06498", "paper_pdf": "papers/w5bNwUzj33.pdf", "paper_pdf_sha256": "0a842fb5522b231c3a02cf0c4be8d17993e87922e692fd9f4841916b9b039408", "paper_pdf_bytes": 364312, "paper_pdf_source": "openreview", "code_url": "https://github.com/tomte812/chef", "code_repository": "tomte812/chef", "code_commit": "ace4fa3621669c4454044bf1932b46ef0a59af44", "code_archive": "repos/w5bNwUzj33.zip", "code_archive_sha256": "d2930c076c14dc2f30236c099a12f62a6cf6737224cabb7c18c601e9664cdc59", "code_archive_bytes": 55919, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 41, "github_languages": {"Python": 160628}, "github_archived": false, "github_pushed_at": "2020-10-05T07:05:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cross-domain-few-shot-learning-by-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJevJCVYvB", "year": 2020, "status": "rejected", "title": "Training Neural Networks for and by Interpolation", "authors": ["Leonard Berrada", "Andrew Zisserman", "Pawan M. Kumar"], "authorids": ["lberrada@robots.ox.ac.uk", "az@robots.ox.ac.uk", "pawan@robots.ox.ac.uk"], "authors_source": "OpenReview API", "abstract": "In modern supervised learning, many deep neural networks are able to interpolate the data: the empirical loss can be driven to near zero on all samples simultaneously. In this work, we explicitly exploit this interpolation property for the design of a new optimization algorithm for deep learning. Specifically, we use it to compute an adaptive learning-rate in closed form at each iteration. This results in the Adaptive Learning-rates for Interpolation with Gradients (ALI-G) algorithm. ALI-G retains the main advantage of SGD which is a low computational cost per iteration. But unlike SGD, the learning-rate of ALI-G uses a single constant hyper-parameter and does not require a decay schedule, which makes it considerably easier to tune. We provide convergence guarantees of ALI-G in the stochastic convex setting. Notably, all our convergence results tackle the realistic case where the interpolation property is satisfied up to some tolerance. We provide experiments on a variety of architectures and tasks: (i) learning a differentiable neural computer; (ii) training a wide residual network on the SVHN data set; (iii) training a Bi-LSTM on the SNLI data set; and (iv) training wide residual networks and densely connected networks on the CIFAR data sets. ALI-G produces state-of-the-art results among adaptive methods, and even yields comparable performance with SGD, which requires manually tuned learning-rate schedules. Furthermore, ALI-G is simple to implement in any standard deep learning framework and can be used as a drop-in replacement in existing code.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rJg9gcITcB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper896/AnonReviewer4"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Thanks for the responses and my concerns seem to be addressed. But since I do not know much about this area, I would like to stick to my initial rating 6.\n==================================================================================================\nThis work designs a new optimization SGD algorithm named ALI-G for deep neural network with interpolation property. \nThis algorithm only has a single hyper-parameter and doesn’t have a decay schedule. The authors provide the convergence guarantees of ALI-G in the stochastic convex setting as well as the experiment results on four tasks.This paper shows state-of-the-art results but I still have have two concerns.\n\nMy main concern is that the performances of other SGD algorithms may be potentially better than the results showed in section 5 because it is not easy to tune the parameter. It would be better if the authors can tune the hyper-parameter more carefully. Take figure 3 as an example. The settings where step size bigger than 1e+1 can hardly shows something, because the step-size is too big for SGD to converge. The settings where step size smaller 1e-2 can also hardly shows something, because the step-size is too small and the experiments only runs 10k steps. It would be better if authors can do more experiments in the settings where step size is from 1e-3 to 1e+0. Moreover, the optimal step size of different optimization algorithms may differ a lot. It would be much more fair if the authors can compare the best performance of different algorithms.\n\nAnother concern is that the authors only give the convergence rate of ALI-G in section 3 but haven’t make any comparisons. For example, it would be better if the authors can show that ALI-G has better convergence result than vanilla SGD without decay schedule.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "title": "Official Blind Review #4", "review": "Thanks for the responses and my concerns seem to be addressed. But since I do not know much about this area, I would like to stick to my initial rating 6.\n==================================================================================================\nThis work designs a new optimization SGD algorithm named ALI-G for deep neural network with interpolation property. \nThis algorithm only has a single hyper-parameter and doesn’t have a decay schedule. The authors provide the convergence guarantees of ALI-G in the stochastic convex setting as well as the experiment results on four tasks.This paper shows state-of-the-art results but I still have have two concerns.\n\nMy main concern is that the performances of other SGD algorithms may be potentially better than the results showed in section 5 because it is not easy to tune the parameter. It would be better if the authors can tune the hyper-parameter more carefully. Take figure 3 as an example. The settings where step size bigger than 1e+1 can hardly shows something, because the step-size is too big for SGD to converge. The settings where step size smaller 1e-2 can also hardly shows something, because the step-size is too small and the experiments only runs 10k steps. It would be better if authors can do more experiments in the settings where step size is from 1e-3 to 1e+0. Moreover, the optimal step size of different optimization algorithms may differ a lot. It would be much more fair if the authors can compare the best performance of different algorithms.\n\nAnother concern is that the authors only give the convergence rate of ALI-G in section 3 but haven’t make any comparisons. For example, it would be better if the authors can show that ALI-G has better convergence result than vanilla SGD without decay schedule.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1572854258264}, {"id": "BJgm9PbTqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper896/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses designing and analyzing an optimization algorithm. Like SGD, it maintains a low computational cost per optimization iteration, but unlike SGD, it does not require manually tuning a decay schedule. This work uses the interpolation property (that the empirical loss can be driven to near zero on all samples simultaneously in a neural network) to compute an adaptive learning rate in closed form at each optimization iteration, and results show that this method produces state-of-the-art results among adaptive methods. I can say the paper was very well written and easy to follow along/understand. Prior work seems comprehensive, and the intuitive comparisons to the prior methods were also useful for the reader. \n\nMy current decision is a weak accept, for a well-written paper, thorough results including meaningful baselines and numerous hyperparameter searches, and a seemingly high-impact tool. Some concerns are listed as follows:\n1-\tConvergence was only discussed in the stochastic convex setting, which seems limiting because we rarely deal with convex problems in problems requiring neural networks.\n2-\tRegularization of the weights during the optimization is dealt with by projecting onto the feasible set of weights, but it seems like there are other types of losses that don’t necessarily to go 0. For example, terms in the objective such as entropy seem worrisome.\n3-\tOne detail that I did not fully follow along with is section 3.1. How does Theorem 1 (Regarding convexity) related to the “each training sample can use its own learning rate without harming progress on the other ones” and/or “allow the updates to rely on the stochastic estimate rather than the exact”? \n4-\tUnfortunately, I am not an expert in this particular area, so I’m not confident about the novelty. For example, the difference between L4 and this is stated to be the utilization of the interpolation policy (which just sets f*=0) and the maximal learning rate, and the stated benefit of convergence guarantees in stochastic convex settings seems poor since most problems will not be convex anyway. More generally, it seems like all details of the algorithm came from elsewhere, although the presented synthesis of ideas does have clear benefits.\n\nAfter reading the author response: \n-I'm fine with points 4 and 3. \n-My feelings about 1 are still the same.\n-My comment on 2 wasn't about cross-entopy loss, but rather other types of objectives that people are often interested in optimizing (such as max-ent RL, where we aim for maximizing rewards as well as maximizing entropy of the policy), in which case, it's not clear to me how we could apply this optimizer. \n-My decision stays as a weak accept", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "This paper addresses designing and analyzing an optimization algorithm. Like SGD, it maintains a low computational cost per optimization iteration, but unlike SGD, it does not require manually tuning a decay schedule. This work uses the interpolation property (that the empirical loss can be driven to near zero on all samples simultaneously in a neural network) to compute an adaptive learning rate in closed form at each optimization iteration, and results show that this method produces state-of-the-art results among adaptive methods. I can say the paper was very well written and easy to follow along/understand. Prior work seems comprehensive, and the intuitive comparisons to the prior methods were also useful for the reader. \n\nMy current decision is a weak accept, for a well-written paper, thorough results including meaningful baselines and numerous hyperparameter searches, and a seemingly high-impact tool. Some concerns are listed as follows:\n1-\tConvergence was only discussed in the stochastic convex setting, which seems limiting because we rarely deal with convex problems in problems requiring neural networks.\n2-\tRegularization of the weights during the optimization is dealt with by projecting onto the feasible set of weights, but it seems like there are other types of losses that don’t necessarily to go 0. For example, terms in the objective such as entropy seem worrisome.\n3-\tOne detail that I did not fully follow along with is section 3.1. How does Theorem 1 (Regarding convexity) related to the “each training sample can use its own learning rate without harming progress on the other ones” and/or “allow the updates to rely on the stochastic estimate rather than the exact”? \n4-\tUnfortunately, I am not an expert in this particular area, so I’m not confident about the novelty. For example, the difference between L4 and this is stated to be the utilization of the interpolation policy (which just sets f*=0) and the maximal learning rate, and the stated benefit of convergence guarantees in stochastic convex settings seems poor since most problems will not be convex anyway. More generally, it seems like all details of the algorithm came from elsewhere, although the presented synthesis of ideas does have clear benefits.\n\nAfter reading the author response: \n-I'm fine with points 4 and 3. \n-My feelings about 1 are still the same.\n-My comment on 2 wasn't about cross-entopy loss, but rather other types of objectives that people are often interested in optimizing (such as max-ent RL, where we aim for maximizing rewards as well as maximizing entropy of the policy), in which case, it's not clear to me how we could apply this optimizer. \n-My decision stays as a weak accept", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572833162689}, {"id": "Syxp4JxAKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper896/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new adaptive learning rate method which is tailored to the optimization of deep neural networks. The motivating observation is that over-parameterized DNNs are able to interpolate the training data (i.e. they are able to reach near-zero training error). This enables application of the Polyak update rule to stochastic updates and a simplification by assuming a zero minimal training loss. A number of proofs for convergence in various convex settings are provided, and empirical evaluation on several benchmarks demonstrates (a) ability to optimize complex architectures, (b) performance improvements over, and (c) performance close to manually tuned SGD learning rates.\n\nI vote for accepting this paper. The approach is well-motivated, the method is described clearly and detail, and the experiments support the paper's claims well. What I would still like to see are a few additional details regarding the experimental protocol. In particular, did you train a single or multiple models for each result that is reported? Do different runs start from the exact same weight initialization? What condition was used to stop the training? The results in section 5.2. are all very close to each other, and it would be helpful to have a sense of the variability of the different methods. The graphs in Figure 4 do look like the models did not converge yet.\n\nGenerally, it would be nice to examine the behavior of the method in cases where the neural network is underparameterized or is otherwise unable to effectively interpolate the training data. Does the method lead to divergence in this case, or is it subpar to other methods? I think section 2.2. could benefit from a short motivational introduction; on the first read, I was not clear about the purpose of introducing the Polyak step size as it is not mentioned explicitly in the text leading to it.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This paper proposes a new adaptive learning rate method which is tailored to the optimization of deep neural networks. The motivating observation is that over-parameterized DNNs are able to interpolate the training data (i.e. they are able to reach near-zero training error). This enables application of the Polyak update rule to stochastic updates and a simplification by assuming a zero minimal training loss. A number of proofs for convergence in various convex settings are provided, and empirical evaluation on several benchmarks demonstrates (a) ability to optimize complex architectures, (b) performance improvements over, and (c) performance close to manually tuned SGD learning rates.\n\nI vote for accepting this paper. The approach is well-motivated, the method is described clearly and detail, and the experiments support the paper's claims well. What I would still like to see are a few additional details regarding the experimental protocol. In particular, did you train a single or multiple models for each result that is reported? Do different runs start from the exact same weight initialization? What condition was used to stop the training? The results in section 5.2. are all very close to each other, and it would be helpful to have a sense of the variability of the different methods. The graphs in Figure 4 do look like the models did not converge yet.\n\nGenerally, it would be nice to examine the behavior of the method in cases where the neural network is underparameterized or is otherwise unable to effectively interpolate the training data. Does the method lead to divergence in this case, or is it subpar to other methods? I think section 2.2. could benefit from a short motivational introduction; on the first read, I was not clear about the purpose of introducing the Polyak step size as it is not mentioned explicitly in the text leading to it."}, "tcdate": 1571843893319}, {"id": "B1xbXQwnKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper896/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new gradient descent methods for training deep neural network which can take the adaptive step size with only one hyper-parameter to tune -- the maximum learning rate -- and achieve comparable results to stochastic gradient descent (SGD) on various tasks and models. In order to achieve that, they develop a stochastic extension of the Polyak step-size for the non-convex setting, namely the adaptive learning-rates for interpolation with gradients (ALI-G), in which the minimal value of the objective loss is set to 0 due to interpolation in neural networks and the learning rates are clipped by a chosen maximal value. The problem is formulated clearly, and the review on the Polyak step-size and related works are well done. Another main contribution of the paper is to provide the convergence guarantees for ALI-G in the convex setting where the objective loss is Lipschitz-continuous (Theorem 1 in the paper). Their theorem also takes into account the error in the estimate of the minimal value of the objective loss. In addition, they derive the connections between  ALI-G and SGD and show that compared to SGD, ALI-G take into consideration that the objective loss is non-negative and set the loss to 0 when it is negative. They perform empirical study to compare their algorithm with other methods including Adagrad, Adam, DFW, L4Adam and SGD on learning a differentiable neural computer, object recognition, and a natural language processing task. Their experimental results show that ALI-G performance is comparable with that of SGD with schedule learning rate.\n\nOverall, this paper could be an interesting contribution. However, some points in the theory and experiments need to be verified. I weakly reject this paper, but given these clarifications in an author response, I would be willing to increase the score. \n\nFor the algorithm and theory, there are some points that need to be verified and further clarification on novelty:\n\n1. When there are regularization such as the weight decay regularization, the minimal objective loss will not be 0. In such cases, Theorem 1 in the paper only guarantees that ALI-G reaches an objective loss less than or equal to a multiple of the estimate error \\epsilon of the true minimal objective loss. It cannot guarantee that the objective loss reached by ALI-G can converge to the true minimall objective loss. Furthermore, when training neural networks with, for example, the weight decay regularization, often times the value of the regularization loss, i.e.  the estimate error \\epsilon, is not small. Therefore, the upper bound given by Theorem 1 is rather loose. \n\n2.  The paper mentions that when no regularization is used, ALI-G and Deep Frank-Wolfe (DFW) are identical algorithms. The difference between the two algorithms are when regularization is used. However, given my concern for Theorem 1 above, the convergence of ALI-G and advantage of ALI-G over DFW in this setting is questionable, and the claim that “ALI-G can handle arbitrary (lower-bounded) loss functions” also needs to be verified. \n\nFor the experiments, the following should be addressed:\n1. In the experiment with the differentiable neural computers, even though ALI-G obtains better performance for a large range of \\eta, its best objective loss is still worse than RMSProp, L4Adam, and L4Mom.\n\n2. Given the merit of Theorem 1 is that the convergence guarantee takes into account the estimate error of the minimal objective loss, an ablation study that compares ALI-G with other methods in the same setting with and without regularization are needed. For example, it would be more convincing if similar results to those in Table 2 or 3 but without regularization are provided and discussed.\n\n3. In Section 5.5, given that ALI-G and DFW are related, why is there no result for DFW in Figure \n4.?\n\n4.  As the paper mentions, AProx algorithm and ALI-G are related, why is there no comparison with AProx in the experiments?\nThings to improve the paper that did not impact the score:\n1. In all experiments, the performance differences between ALI-G and competitive methods are small. Thus, error bars are needed for these results.\n\n2. In Table 3, the gap between SGD and ALI-G can be significantly different on different architectures. For example, in CIFAR-100 experiments, while ALI-G achieves the same result as SGD with the DenseNet, ALI-G’s performance on Wide ResNet is much worse than SGD. Do you have any explanation for this?\n\n3. How is ALI-G compared with the methods proposed in the paper “Stochastic Gradient Descent with Polyak's Learning Rate” (https://arxiv.org/abs/1903.08688)\n\n\nThe following paper also proved the convergence of adaptive gradient methods for nonconvex optimization:\nDongruo Zhou*, Yiqi Tang*, Ziyan Yang*, Yuan Cao, Quanquan Gu. On the Convergence of Adaptive Gradient Methods for Nonconvex Optimization.\n\n-------------------------------------------------------------\nAfter rebuttal:\nI noticed that the authors:\n\n1. did not compare with SGD or Adam with good hyperparameters (https://arxiv.org/abs/1907.08610)\n\n2. did not test on the large scale datasets, e.g., imagenet\n\n3. the proposed algorithm is not as good as SGD on most numerical experiments.\n\nFor the above reasons, I think this paper should be rejected.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This paper proposes a new gradient descent methods for training deep neural network which can take the adaptive step size with only one hyper-parameter to tune -- the maximum learning rate -- and achieve comparable results to stochastic gradient descent (SGD) on various tasks and models. In order to achieve that, they develop a stochastic extension of the Polyak step-size for the non-convex setting, namely the adaptive learning-rates for interpolation with gradients (ALI-G), in which the minimal value of the objective loss is set to 0 due to interpolation in neural networks and the learning rates are clipped by a chosen maximal value. The problem is formulated clearly, and the review on the Polyak step-size and related works are well done. Another main contribution of the paper is to provide the convergence guarantees for ALI-G in the convex setting where the objective loss is Lipschitz-continuous (Theorem 1 in the paper). Their theorem also takes into account the error in the estimate of the minimal value of the objective loss. In addition, they derive the connections between  ALI-G and SGD and show that compared to SGD, ALI-G take into consideration that the objective loss is non-negative and set the loss to 0 when it is negative. They perform empirical study to compare their algorithm with other methods including Adagrad, Adam, DFW, L4Adam and SGD on learning a differentiable neural computer, object recognition, and a natural language processing task. Their experimental results show that ALI-G performance is comparable with that of SGD with schedule learning rate.\n\nOverall, this paper could be an interesting contribution. However, some points in the theory and experiments need to be verified. I weakly reject this paper, but given these clarifications in an author response, I would be willing to increase the score. \n\nFor the algorithm and theory, there are some points that need to be verified and further clarification on novelty:\n\n1. When there are regularization such as the weight decay regularization, the minimal objective loss will not be 0. In such cases, Theorem 1 in the paper only guarantees that ALI-G reaches an objective loss less than or equal to a multiple of the estimate error \\epsilon of the true minimal objective loss. It cannot guarantee that the objective loss reached by ALI-G can converge to the true minimall objective loss. Furthermore, when training neural networks with, for example, the weight decay regularization, often times the value of the regularization loss, i.e.  the estimate error \\epsilon, is not small. Therefore, the upper bound given by Theorem 1 is rather loose. \n\n2.  The paper mentions that when no regularization is used, ALI-G and Deep Frank-Wolfe (DFW) are identical algorithms. The difference between the two algorithms are when regularization is used. However, given my concern for Theorem 1 above, the convergence of ALI-G and advantage of ALI-G over DFW in this setting is questionable, and the claim that “ALI-G can handle arbitrary (lower-bounded) loss functions” also needs to be verified. \n\nFor the experiments, the following should be addressed:\n1. In the experiment with the differentiable neural computers, even though ALI-G obtains better performance for a large range of \\eta, its best objective loss is still worse than RMSProp, L4Adam, and L4Mom.\n\n2. Given the merit of Theorem 1 is that the convergence guarantee takes into account the estimate error of the minimal objective loss, an ablation study that compares ALI-G with other methods in the same setting with and without regularization are needed. For example, it would be more convincing if similar results to those in Table 2 or 3 but without regularization are provided and discussed.\n\n3. In Section 5.5, given that ALI-G and DFW are related, why is there no result for DFW in Figure \n4.?\n\n4.  As the paper mentions, AProx algorithm and ALI-G are related, why is there no comparison with AProx in the experiments?\nThings to improve the paper that did not impact the score:\n1. In all experiments, the performance differences between ALI-G and competitive methods are small. Thus, error bars are needed for these results.\n\n2. In Table 3, the gap between SGD and ALI-G can be significantly different on different architectures. For example, in CIFAR-100 experiments, while ALI-G achieves the same result as SGD with the DenseNet, ALI-G’s performance on Wide ResNet is much worse than SGD. Do you have any explanation for this?\n\n3. How is ALI-G compared with the methods proposed in the paper “Stochastic Gradient Descent with Polyak's Learning Rate” (https://arxiv.org/abs/1903.08688)\n\n\nThe following paper also proved the convergence of adaptive gradient methods for nonconvex optimization:\nDongruo Zhou*, Yiqi Tang*, Ziyan Yang*, Yuan Cao, Quanquan Gu. On the Convergence of Adaptive Gradient Methods for Nonconvex Optimization.\n\n-------------------------------------------------------------\nAfter rebuttal:\nI noticed that the authors:\n\n1. did not compare with SGD or Adam with good hyperparameters (https://arxiv.org/abs/1907.08610)\n\n2. did not test on the large scale datasets, e.g., imagenet\n\n3. the proposed algorithm is not as good as SGD on most numerical experiments.\n\nFor the above reasons, I think this paper should be rejected.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571742489345}], "openreview_url": "https://openreview.net/forum?id=BJevJCVYvB", "arxiv_id": "1906.05661", "paper_pdf": "papers/BJevJCVYvB.pdf", "paper_pdf_sha256": "ceb527e116770d06f06aea1025ac9cd14923fabe563394be0a379d5173852316", "paper_pdf_bytes": 605903, "paper_pdf_source": "openreview", "code_url": "https://github.com/oval-group/ali-g", "code_repository": "oval-group/ali-g", "code_commit": "568947a2bbb651be3401fbae60229e2fcbf8a505", "code_archive": "repos/BJevJCVYvB.zip", "code_archive_sha256": "c9fb31b7b9dc9fcfb3e119b323f6c87e474c98342869771893ef60fd3e8fb4e7", "code_archive_bytes": 37466, "code_file_count": 30, "code_extensions": {".py": 30}, "github_disk_usage_kb": 48, "github_languages": {"Python": 76994}, "github_archived": false, "github_pushed_at": "2021-03-07T15:57:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/training-neural-networks-for-and-by"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HygUOoC5KX", "year": 2019, "status": "rejected", "title": "Are Generative Classifiers More Robust to Adversarial Attacks?", "authors": ["Yingzhen Li", "John Bradshaw", "Yash Sharma"], "authorids": ["yl494@cam.ac.uk", "jab255@cam.ac.uk", "ysharma1126@gmail.com"], "authors_source": "OpenReview API", "abstract": "There is a rising interest in studying the robustness of deep neural network classifiers against adversaries, with both advanced attack and defence techniques being actively developed. However, most recent work focuses on discriminative classifiers, which only model the conditional distribution of the labels given the inputs. In this paper, we propose and investigate the deep Bayes classifier, which improves classical naive Bayes with conditional deep generative models. We further develop detection methods for adversarial examples, which reject inputs with low likelihood under the generative model. Experimental results suggest that deep Bayes classifiers are more robust than deep discriminative classifiers, and that the proposed detection methods are effective against many recently proposed attacks.", "decision": null, "meta_review": "Reject", "num_reviews": 4, "reviews": [{"id": "B1xvkiIZgN", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper356/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper explores the potential of generative models for adversarial robustness. It presents some interesting and well formulated findings but is lacking in some meaningful ways. \n\n-It does a good job of summarizing the literature though it misses out on some very recent but relevant work\n-It introduces three detection methods and extensively evaluates them  against several zero-knowledge attack types. Notably all the attack types are gradient based methods. \n-They present detailed performance metrics on the zero-knowledge attack, and introduce the perfect knowledge attack but do not present equivalently detailed results.\n-The results are  presented in a confusing manor, and the paper is perhaps done a disservice by the inclusion of a very large number of tables both in the main text and the appendix without the necessary writing required to situate the reader.  \n-The paper also briefly mentions the ongoing debate around \"off-manifold\" conjecture, only to them assume its correctness and based the entire work on the premise. \n\n###Highlights \n*well written introduction and a good overview of generative modeling \n*very good visualizations and explanation of the detection methods \n*thorough results on zero-knowledge gradient based attack detection rates  \n*interesting direction on adversarial examples  \n\n### Areas for improvement \n*more time should be spent on discussing the off manifold conjecture; the paper is based on the correctness of this conjecture though recent work has called its validity into question\n*only gradient based attacks are discussed though the paper title suggests robustness to all attacks\n*perfect-knowledge attacks should be explored in much more detail \n* the paper should streamline the results and findings, the large number of tables and results will confuse the reader and do not add to the argument\n*recent work on generative models for adversarial robustness should be discussed (https://arxiv.org/abs/1805.09190, https://arxiv.org/abs/1811.06969)  \n*the two class cifar10 results are not particularly convincing as this is a substantial simplification of the cifar10 problem \n* the results on the full cifar10 data set make use a the extracted features of a pre-trained discriminative cifar10 network, this introduces the problem at layer activation adversarial attacks, but this is not mentioned in the paper (https://arxiv.org/pdf/1511.05122.pdf)\n\nOverall i think the work shows promise but is not yet ready for acceptance \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work but confusing presentation and some missing areas ", "review": "This paper explores the potential of generative models for adversarial robustness. It presents some interesting and well formulated findings but is lacking in some meaningful ways. \n\n-It does a good job of summarizing the literature though it misses out on some very recent but relevant work\n-It introduces three detection methods and extensively evaluates them  against several zero-knowledge attack types. Notably all the attack types are gradient based methods. \n-They present detailed performance metrics on the zero-knowledge attack, and introduce the perfect knowledge attack but do not present equivalently detailed results.\n-The results are  presented in a confusing manor, and the paper is perhaps done a disservice by the inclusion of a very large number of tables both in the main text and the appendix without the necessary writing required to situate the reader.  \n-The paper also briefly mentions the ongoing debate around \"off-manifold\" conjecture, only to them assume its correctness and based the entire work on the premise. \n\n###Highlights \n*well written introduction and a good overview of generative modeling \n*very good visualizations and explanation of the detection methods \n*thorough results on zero-knowledge gradient based attack detection rates  \n*interesting direction on adversarial examples  \n\n### Areas for improvement \n*more time should be spent on discussing the off manifold conjecture; the paper is based on the correctness of this conjecture though recent work has called its validity into question\n*only gradient based attacks are discussed though the paper title suggests robustness to all attacks\n*perfect-knowledge attacks should be explored in much more detail \n* the paper should streamline the results and findings, the large number of tables and results will confuse the reader and do not add to the argument\n*recent work on generative models for adversarial robustness should be discussed (https://arxiv.org/abs/1805.09190, https://arxiv.org/abs/1811.06969)  \n*the two class cifar10 results are not particularly convincing as this is a substantial simplification of the cifar10 problem \n* the results on the full cifar10 data set make use a the extracted features of a pre-trained discriminative cifar10 network, this introduces the problem at layer activation adversarial attacks, but this is not mentioned in the paper (https://arxiv.org/pdf/1511.05122.pdf)\n\nOverall i think the work shows promise but is not yet ready for acceptance \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1544805087321}, {"id": "H1eqrk0K6X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper356/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work investigates an interesting direction of improving robustness of classifiers against adversarial attacks by using generative models. The authors propose the *deep Bayes classifier*, which is a deep LVM based extension of naive Bayes. Furthermore, the authors extensively explore 7 possible factorisations of the classifier. Thorough experiments are conducted to assess the capability of defending or detecting adversarial examples.  Besides, the authors incorporate discriminative features to generative classifiers and demonstrate clear robustness gain.\n\n### Highlights\n* This work proposes an attractive direction -- the use of generative model in defending or detecting adversarial attacks. I suggest this idea should be follower by more further studies.\n* The presented models are quite straight-forward but exhibit good robustness against attacks listed in the experiments.\n* Various structural possibilities of the graphical model are examined which is preferable and helps assess the effectiveness of generative classifiers.\n\n### Minors\n* Although the major point here is robustness against adversarial attacks, as mentioned by the authors, the performance on clear cases (i.e. no attacks) is unsatisfactory. Also, experiments on CIFAR are too much simplified (only 2 very unlike classes) and therefore not very convincing. \n* For the combination of generative classifier and discriminative features, I’m curious about the results on the clear CIFAR-10 multi-class problem. It should be a very positive plus if results are satisfactory.\n* The writing is sometimes hard to follow. For examples, many ad-hoc abbreviations are used across the paper causing difficulties of understanding the core idea and results.\n\n### Conclusion \nIn general, this paper brings our attention to a previously less investigated but seemingly promising research direction, i.e. robustness of generative model against adversarial attacks. The idea is insightful and proposed models are straight-forward. While only on small-scale  problems (with the presence of attacks), extensive experimental results in this paper can assist further study on this field. Thus, I recommend this paper to be accepted.   \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good work with thorough experimental study", "review": "This work investigates an interesting direction of improving robustness of classifiers against adversarial attacks by using generative models. The authors propose the *deep Bayes classifier*, which is a deep LVM based extension of naive Bayes. Furthermore, the authors extensively explore 7 possible factorisations of the classifier. Thorough experiments are conducted to assess the capability of defending or detecting adversarial examples.  Besides, the authors incorporate discriminative features to generative classifiers and demonstrate clear robustness gain.\n\n### Highlights\n* This work proposes an attractive direction -- the use of generative model in defending or detecting adversarial attacks. I suggest this idea should be follower by more further studies.\n* The presented models are quite straight-forward but exhibit good robustness against attacks listed in the experiments.\n* Various structural possibilities of the graphical model are examined which is preferable and helps assess the effectiveness of generative classifiers.\n\n### Minors\n* Although the major point here is robustness against adversarial attacks, as mentioned by the authors, the performance on clear cases (i.e. no attacks) is unsatisfactory. Also, experiments on CIFAR are too much simplified (only 2 very unlike classes) and therefore not very convincing. \n* For the combination of generative classifier and discriminative features, I’m curious about the results on the clear CIFAR-10 multi-class problem. It should be a very positive plus if results are satisfactory.\n* The writing is sometimes hard to follow. For examples, many ad-hoc abbreviations are used across the paper causing difficulties of understanding the core idea and results.\n\n### Conclusion \nIn general, this paper brings our attention to a previously less investigated but seemingly promising research direction, i.e. robustness of generative model against adversarial attacks. The idea is insightful and proposed models are straight-forward. While only on small-scale  problems (with the presence of attacks), extensive experimental results in this paper can assist further study on this field. Thus, I recommend this paper to be accepted.   \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1542213441627}, {"id": "B1ePMbbcnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper356/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this work the authors propose and analyse generative models as defences against adversarial examples. In addition, three detection methods are introduced and an extension to deep features is suggested.\n\nMy main concerns are as follows (details below):\n* Important prior work is not mentioned.\n* Evaluation with direct attacks is only based on (very few) gradient-based techniques, many results are not reliable.\n* There are signs of gradient masking (the common problem of robustness evaluation, in particular of only gradient-based techniques are used).\n* The way detection rates are taken into account in the perfect knowledge scenario is confusing.\n\n### Style\nI like the idea of testing many different factorisation structures. However, that comes with the drawback that one needs to constantly check back what the abbreviations mean. Together with the three detection methods, the manuscript is quite confusing at times and should definitely be streamlined. One suggestion: remove the detection methods: I did not find any real conclusion about them but they are definitely side-tracking users away from the main results.\n\n### Prior work\nThere is at least one closely related prior work not mentioned here: the analysis by synthesis model [1]. This model uses a variational auto encoder to learn class conditional distributions and shows high robustness on MNIST. Please make clear what your contribution is over this paper (other than testing several other factorisations).\n\n### Evaluation problems\nThe robustness of models should be evaluated on different direct attacks ranging from gradient-based to score-based (e.g. NES [2]) to decision-based attacks [3]. Please take a look at [1] to see how a very extensive evaluation might look like. The results can be astonishingly different for different attacks, and so basing conclusion on only one or two attacks is dangerous (in particular if you only use gradient-based ones). One can also see that in your results, just check the variations you get between MIM and PGD. Also, rather then discussing (and showing in detail) results for individual attacks, the minimum adversarial distance for a given sample that can be found by any attack is much more comparable between models (which can also streamline the manuscript).\n\nOne can see signs of gradient masking in your results. For example, in Figure 3 the MIM attacks levels out at 20% for the DBX model. That can happen for iterative attacks if the gradient is masked. Similarly, in Figure 5 DBX-ZK (zero knowledge) is better in both accuracy and detection rate than DBX-PKK (which takes the KL-detection method into account and should thus either be better in accuracy or detection rate).\n\nMore generally, the perfect knowledge case, in which the attacker knows about the detector, should only count samples as adversarials which evade the detector and change the model decision. Thus, the detection rate should be zero. Otherwise I have no idea what trade-off between accuracy and detection rate you are actually targeting and how to compare the results.\n\nAlso, some intermediate results are conflicting with each other. E.g. in 4.1 you state “the usage of bottleneck is beneficial for better robustness”, but for L2 this is not true.\n\nAlso, I am not sure how conclusive the grey-box and black-box scenarios really are: since the substitute is basically a DFX or DFZ, it’s unsurprising that adversarials transfer best to those two models.\n\n### Minor\n * In 4.1 you say “as they fail to find near manifold adversarials”, but I don’t see how there can be L-infty adversarials on MNIST that are on-manifold (remember, MNIST pixel values are basically binary). Plus, in the zero-knowledge scenario there is nothing that enforces staying on this manifold.\n * Result presentation (Figure 3/5 & Table 1) is very different for different attack scenarios, which makes them hard to compare. Please unify.\n * Is the L2 distance you report in Table 1 the mean (or median) distance to adversarial examples. If so, GBZ (for which you state that C&W “failed on attacking” has actually a smaller mean adversarial distance than some other models (for which C&W is actually quite successful).\n * Grey-box scenario doesn’t make a lot of sense: since the substitute is basically a DFX or DFZ, it’s unsurprising that adversarials transfer best to those two models. A similar confounder makes the black-box results difficult to interpret.\n* Also, taking into account that the paper is two pages longer and thus calls for higher standards\n\nTaken together, I find the general direction of the paper very interesting and I’d definitely encourage the authors to go further. At the current stage, however, I feel that (1) contributions are not sufficiently delineated to prior work, (2) the evaluation is not convincingly supporting the claims and that  (3) the manuscript needs to be streamlined (both in terms of text and figures).\n\n[1] Schott et al. (2018) “Towards the first adversarially robust neural network model on MNIST” (https://arxiv.org/abs/1805.09190)\n[2] Ilyas et al. (2018) “Black-box Adversarial Attacks with Limited Queries and Information” ( [https://arxiv.org/abs/1804.08598)](https://arxiv.org/abs/1804.08598)) \n[3] Brendel et al. (2018) “Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models” (https://arxiv.org/abs/1712.04248)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting direction but some important prior work is missing and evaluation does not yet convincingly support claims", "review": "In this work the authors propose and analyse generative models as defences against adversarial examples. In addition, three detection methods are introduced and an extension to deep features is suggested.\n\nMy main concerns are as follows (details below):\n* Important prior work is not mentioned.\n* Evaluation with direct attacks is only based on (very few) gradient-based techniques, many results are not reliable.\n* There are signs of gradient masking (the common problem of robustness evaluation, in particular of only gradient-based techniques are used).\n* The way detection rates are taken into account in the perfect knowledge scenario is confusing.\n\n### Style\nI like the idea of testing many different factorisation structures. However, that comes with the drawback that one needs to constantly check back what the abbreviations mean. Together with the three detection methods, the manuscript is quite confusing at times and should definitely be streamlined. One suggestion: remove the detection methods: I did not find any real conclusion about them but they are definitely side-tracking users away from the main results.\n\n### Prior work\nThere is at least one closely related prior work not mentioned here: the analysis by synthesis model [1]. This model uses a variational auto encoder to learn class conditional distributions and shows high robustness on MNIST. Please make clear what your contribution is over this paper (other than testing several other factorisations).\n\n### Evaluation problems\nThe robustness of models should be evaluated on different direct attacks ranging from gradient-based to score-based (e.g. NES [2]) to decision-based attacks [3]. Please take a look at [1] to see how a very extensive evaluation might look like. The results can be astonishingly different for different attacks, and so basing conclusion on only one or two attacks is dangerous (in particular if you only use gradient-based ones). One can also see that in your results, just check the variations you get between MIM and PGD. Also, rather then discussing (and showing in detail) results for individual attacks, the minimum adversarial distance for a given sample that can be found by any attack is much more comparable between models (which can also streamline the manuscript).\n\nOne can see signs of gradient masking in your results. For example, in Figure 3 the MIM attacks levels out at 20% for the DBX model. That can happen for iterative attacks if the gradient is masked. Similarly, in Figure 5 DBX-ZK (zero knowledge) is better in both accuracy and detection rate than DBX-PKK (which takes the KL-detection method into account and should thus either be better in accuracy or detection rate).\n\nMore generally, the perfect knowledge case, in which the attacker knows about the detector, should only count samples as adversarials which evade the detector and change the model decision. Thus, the detection rate should be zero. Otherwise I have no idea what trade-off between accuracy and detection rate you are actually targeting and how to compare the results.\n\nAlso, some intermediate results are conflicting with each other. E.g. in 4.1 you state “the usage of bottleneck is beneficial for better robustness”, but for L2 this is not true.\n\nAlso, I am not sure how conclusive the grey-box and black-box scenarios really are: since the substitute is basically a DFX or DFZ, it’s unsurprising that adversarials transfer best to those two models.\n\n### Minor\n * In 4.1 you say “as they fail to find near manifold adversarials”, but I don’t see how there can be L-infty adversarials on MNIST that are on-manifold (remember, MNIST pixel values are basically binary). Plus, in the zero-knowledge scenario there is nothing that enforces staying on this manifold.\n * Result presentation (Figure 3/5 & Table 1) is very different for different attack scenarios, which makes them hard to compare. Please unify.\n * Is the L2 distance you report in Table 1 the mean (or median) distance to adversarial examples. If so, GBZ (for which you state that C&W “failed on attacking” has actually a smaller mean adversarial distance than some other models (for which C&W is actually quite successful).\n * Grey-box scenario doesn’t make a lot of sense: since the substitute is basically a DFX or DFZ, it’s unsurprising that adversarials transfer best to those two models. A similar confounder makes the black-box results difficult to interpret.\n* Also, taking into account that the paper is two pages longer and thus calls for higher standards\n\nTaken together, I find the general direction of the paper very interesting and I’d definitely encourage the authors to go further. At the current stage, however, I feel that (1) contributions are not sufficiently delineated to prior work, (2) the evaluation is not convincingly supporting the claims and that  (3) the manuscript needs to be streamlined (both in terms of text and figures).\n\n[1] Schott et al. (2018) “Towards the first adversarially robust neural network model on MNIST” (https://arxiv.org/abs/1805.09190)\n[2] Ilyas et al. (2018) “Black-box Adversarial Attacks with Limited Queries and Information” ( [https://arxiv.org/abs/1804.08598)](https://arxiv.org/abs/1804.08598)) \n[3] Brendel et al. (2018) “Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models” (https://arxiv.org/abs/1712.04248)", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541177614730}, {"id": "ryl3wzTYn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper356/AnonReviewer2"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims to test the robustness of generative classifiers [1] w.r.t. adversarial examples, considering their use as a potentially more robust alternative to adversarial training of discriminative classifiers. To achieve this, *Deep Bayes*, a generalization of the Naive Bayes classifier using a latent variable model and trained in a fashion similar to variational autoencoders [2] is introduced, and 7 different latent variable models are compared, covering a spectrum of generative or discriminative classification models, with or without bottlenecks. Their DFX and DBX architectures in particular closely match traditional discriminative classifiers, without and with a latent bottleneck.\n\nThese 7 models are compared against a large range of adversarial attacks, depending on the kind of noise added (l_2 or l_inf) and how much the adversary can access (the full gradients of the model, its output on training data, or only the model as a black-box). The performance of the models is assessed depending on two criteria: how the performance of the classifier resists to adversarial noise, and how quickly the model can detect adversarial samples. Three methods for detecting adversarial samples are compared: the first (only applicable to generative classifiers) discards samples with a low likelihood, according to the off-manifold assumption [3], the second discards samples for which the classifier has low confidence in its classification (p(y|x) is under some threshold), and the third compares the output probability vector of the classifier on a sample to the mean classification vector of this class over the train data, and discards the sample if the two vectors are too dissimilar (meaning the classifier is over-confident or under-confident).\n\nThe main contribution of this paper is the extensive experiments that have been done to compare the models against the various adversarial attacks. While experiments were only done on small datasets like MNIST and CIFAR (generative classifiers don't scale as easily on large image datasets), they nonetheless give very interesting insights and the authors provided encouraging results on applying generative classifiers on features learned by discriminative classifiers. Theirs result shows that generative architecture are in general more robust to the current state-of-the-art adversarial attacks, and detect adversarial examples more easily. The authors also recognize that these results may be biased by the fact that current adversarial attacks have been specifically optimized towards discriminative classifier.\n\nThis is a solid paper in my opinion. The experimental setup and motivations are clearly detailed, and the paper was easy to follow. Extensive results and description of the experimental protocol are provided in the appendices, giving me confidence that the results should be reproducible. The results of this paper give interesting insights regarding how to approach robustness to adversarial examples in classification tasks, and provide realistic ways to try and apply generative classifiers in real-worlds tasks, using pre-learned features from discriminative networks.\n\n\n[1] http://papers.nips.cc/paper/2020-on-discriminative-vs-generative-classifiers-a-comparison-of-logistic-regression-and-naive-bayes.pdf\n[2] https://arxiv.org/abs/1312.6114\n[3] https://arxiv.org/abs/1801.02774", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Solid and insightful experimental study", "review": "This paper aims to test the robustness of generative classifiers [1] w.r.t. adversarial examples, considering their use as a potentially more robust alternative to adversarial training of discriminative classifiers. To achieve this, *Deep Bayes*, a generalization of the Naive Bayes classifier using a latent variable model and trained in a fashion similar to variational autoencoders [2] is introduced, and 7 different latent variable models are compared, covering a spectrum of generative or discriminative classification models, with or without bottlenecks. Their DFX and DBX architectures in particular closely match traditional discriminative classifiers, without and with a latent bottleneck.\n\nThese 7 models are compared against a large range of adversarial attacks, depending on the kind of noise added (l_2 or l_inf) and how much the adversary can access (the full gradients of the model, its output on training data, or only the model as a black-box). The performance of the models is assessed depending on two criteria: how the performance of the classifier resists to adversarial noise, and how quickly the model can detect adversarial samples. Three methods for detecting adversarial samples are compared: the first (only applicable to generative classifiers) discards samples with a low likelihood, according to the off-manifold assumption [3], the second discards samples for which the classifier has low confidence in its classification (p(y|x) is under some threshold), and the third compares the output probability vector of the classifier on a sample to the mean classification vector of this class over the train data, and discards the sample if the two vectors are too dissimilar (meaning the classifier is over-confident or under-confident).\n\nThe main contribution of this paper is the extensive experiments that have been done to compare the models against the various adversarial attacks. While experiments were only done on small datasets like MNIST and CIFAR (generative classifiers don't scale as easily on large image datasets), they nonetheless give very interesting insights and the authors provided encouraging results on applying generative classifiers on features learned by discriminative classifiers. Theirs result shows that generative architecture are in general more robust to the current state-of-the-art adversarial attacks, and detect adversarial examples more easily. The authors also recognize that these results may be biased by the fact that current adversarial attacks have been specifically optimized towards discriminative classifier.\n\nThis is a solid paper in my opinion. The experimental setup and motivations are clearly detailed, and the paper was easy to follow. Extensive results and description of the experimental protocol are provided in the appendices, giving me confidence that the results should be reproducible. The results of this paper give interesting insights regarding how to approach robustness to adversarial examples in classification tasks, and provide realistic ways to try and apply generative classifiers in real-worlds tasks, using pre-learned features from discriminative networks.\n\n\n[1] http://papers.nips.cc/paper/2020-on-discriminative-vs-generative-classifiers-a-comparison-of-logistic-regression-and-naive-bayes.pdf\n[2] https://arxiv.org/abs/1312.6114\n[3] https://arxiv.org/abs/1801.02774", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541161572386}], "openreview_url": "https://openreview.net/forum?id=HygUOoC5KX", "arxiv_id": "1802.06552", "paper_pdf": "papers/HygUOoC5KX.pdf", "paper_pdf_sha256": "c4119443dc3e9b71f4f1b3fea836e27cee60a0f22268cd445847096f30517167", "paper_pdf_bytes": 2991445, "paper_pdf_source": "openreview", "code_url": "https://github.com/deepgenerativeclassifier/DeepBayes", "code_repository": "deepgenerativeclassifier/DeepBayes", "code_commit": "b7d78330638b8dfc8b64b1c19f460c74944bebda", "code_archive": "repos/HygUOoC5KX.zip", "code_archive_sha256": "c2489fc08e6b2e69b5d2edc231a5ef9f945fb10f9a0e30d0d010a35b757e49ee", "code_archive_bytes": 265888, "code_file_count": 73, "code_extensions": {".py": 73}, "github_disk_usage_kb": 228, "github_languages": {"Python": 472032}, "github_archived": false, "github_pushed_at": "2019-05-10T10:29:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/are-generative-classifiers-more-robust-to"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HklpCzC6-", "year": 2018, "status": "rejected", "title": "Image Segmentation by Iterative Inference from Conditional Score Estimation", "authors": ["Adriana Romero", "Michal Drozdzal", "Akram Erraqabi", "Simon Jégou", "Yoshua Bengio"], "authorids": ["adriana.romsor@gmail.com", "michal.drozdzal@gmail.com", "akram.er-raqabi@umontreal.ca", "simon.jegou@gmail.com", "yoshua.umontreal@gmail.com"], "authors_source": "OpenReview API", "abstract": "Inspired by the combination of feedforward and iterative computations in the visual cortex, and taking advantage of the ability of denoising autoencoders to estimate the score of a joint distribution, we propose a novel approach to iterative inference for capturing and exploiting the complex joint distribution of output variables conditioned on some input variables. This approach is applied to image pixel-wise segmentation, with the estimated conditional score used to perform gradient ascent towards a mode of the estimated conditional distribution. This extends previous work on score estimation by denoising autoencoders to the case of a conditional distribution, with a novel use of a corrupted feedforward predictor replacing Gaussian corruption. An advantage of this approach over more classical ways to perform iterative inference for structured outputs, like conditional random fields (CRFs), is that it is not any more necessary to define an explicit energy function linking the output variables. To keep computations tractable, such energy function parametrizations are typically fairly constrained, involving only a few neighbors of each of the output variables in each clique. We experimentally find that the proposed iterative inference from conditional score estimation by conditional denoising autoencoders performs better than comparable models based on CRFs or those not using any explicit modeling of the conditional joint distribution of outputs.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "S1cNDW9eM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper87/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I am a returning reviewer for this paper, from a previous conference. Much of the paper remains unchanged from the time of my previous review. I have revised my review according to the updates in the paper:\n\nSummary of the paper:\nThis work proposes a neural network based alternative to standard CRF post-processing techniques that are generally used on top semantic segmentation CNNs. As an alternative to CRF, this work proposes to iteratively refine the predicted segmentation with a denoising auto encoder (DAE). Results on CamVid semantic segmentation dataset showed better improvements over base CNN predictions in comparison to popular DenseCRF technique.\n\n\nPaper Strengths:\n- A neat technique for incorporating CRF-like pixel label relations into semantic segmentation via neural networks (auto encoders).\n- Promising results on CamVid segmentation dataset with reliable improvements over baseline techniques and minor improvements when used in conjunction with recent models.\n\n\nMajor Weaknesses:\nI have two main concerns for this work:\n- One is related to the novelty as the existing work of Xie et al. ECCV'16 also proposed similar technique with very similar aim.  I think, conceptual or empirical comparisons are required to assess the importance of the proposed approach with respect to existing ones. Mere citation and short discussion is not enough. Moreover, Xie et al. seem to have demonstrated their technique on two different tasks and on three different datasets.\n- Another concern is related to experiments. Authors experimented with only one dataset and with one problem. But, I would either expect some demonstration of generality (more datasets or tasks) or strong empirical performance (state-of-the-art on CamVid) to assess the empirical usefulness with respect to existing techniques. Both of these aspects are missing in experiments. \n\n\nMinor Weaknesses:\n- Negligible improvements with respect to CRF techniques on modern deep architectures.\n- Runtime comparison is missing with respect to baseline techniques. Applying the proposed DAE 40-50 times seems very time consuming for each image.\n- By back-propagating through CRF-like techniques [Zheng et al. ICCV'15, Gadde et al. ECCV'16, Chandra et al. ECCV'16 etc.], one could refine the base segmentation CNN as well. It seems this is also possible with the proposed architecture. Is that correct? Or, are there any problems with the end-to-end fine-tuning as the input distribution to DAE constantly changes? Did authors try this?\n\n\nSuggestions:\n- Only Gaussian noise corruption is used for training DAE. Did authors experiment with any other noise types? Probably, more structured noise would help in learning better contextual relations across pixel labels?\n\nClarifications:\nWhat is the motivation to add Euclidean loss to the standard cross-entropy loss for segmentation in Eq-3?\n\nReview summary:\nThe use of denoising auto encoders (DAEs) for capturing pixel label relations and then using them to iteratively refine the segmentation predictions is interesting. But, incomplete comparisons with similar existing work and limited experiments makes this a weak paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good CNN based post-processing technique for semantic segmentation, but the experiments are incomplete and limited.", "rating": "4: Ok but not good enough - rejection", "review": "I am a returning reviewer for this paper, from a previous conference. Much of the paper remains unchanged from the time of my previous review. I have revised my review according to the updates in the paper:\n\nSummary of the paper:\nThis work proposes a neural network based alternative to standard CRF post-processing techniques that are generally used on top semantic segmentation CNNs. As an alternative to CRF, this work proposes to iteratively refine the predicted segmentation with a denoising auto encoder (DAE). Results on CamVid semantic segmentation dataset showed better improvements over base CNN predictions in comparison to popular DenseCRF technique.\n\n\nPaper Strengths:\n- A neat technique for incorporating CRF-like pixel label relations into semantic segmentation via neural networks (auto encoders).\n- Promising results on CamVid segmentation dataset with reliable improvements over baseline techniques and minor improvements when used in conjunction with recent models.\n\n\nMajor Weaknesses:\nI have two main concerns for this work:\n- One is related to the novelty as the existing work of Xie et al. ECCV'16 also proposed similar technique with very similar aim.  I think, conceptual or empirical comparisons are required to assess the importance of the proposed approach with respect to existing ones. Mere citation and short discussion is not enough. Moreover, Xie et al. seem to have demonstrated their technique on two different tasks and on three different datasets.\n- Another concern is related to experiments. Authors experimented with only one dataset and with one problem. But, I would either expect some demonstration of generality (more datasets or tasks) or strong empirical performance (state-of-the-art on CamVid) to assess the empirical usefulness with respect to existing techniques. Both of these aspects are missing in experiments. \n\n\nMinor Weaknesses:\n- Negligible improvements with respect to CRF techniques on modern deep architectures.\n- Runtime comparison is missing with respect to baseline techniques. Applying the proposed DAE 40-50 times seems very time consuming for each image.\n- By back-propagating through CRF-like techniques [Zheng et al. ICCV'15, Gadde et al. ECCV'16, Chandra et al. ECCV'16 etc.], one could refine the base segmentation CNN as well. It seems this is also possible with the proposed architecture. Is that correct? Or, are there any problems with the end-to-end fine-tuning as the input distribution to DAE constantly changes? Did authors try this?\n\n\nSuggestions:\n- Only Gaussian noise corruption is used for training DAE. Did authors experiment with any other noise types? Probably, more structured noise would help in learning better contextual relations across pixel labels?\n\nClarifications:\nWhat is the motivation to add Euclidean loss to the standard cross-entropy loss for segmentation in Eq-3?\n\nReview summary:\nThe use of denoising auto encoders (DAEs) for capturing pixel label relations and then using them to iteratively refine the segmentation predictions is interesting. But, incomplete comparisons with similar existing work and limited experiments makes this a weak paper.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511819057618}, {"id": "rynv0Uf-f", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper87/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an iterative procedure on top of a standard image semantic segmentation networks. \n\nThe submission proposes a change to the training procedure of stacking a denoising auto-encoder for image segmentation. The technical contribution of this paper is small. The paper aims to answer a single question: When using a DAE network on top of a segmentation network output, should one condition on the predicted, or the ground truth segmentation? (why not on both?) The answer is conditioning on the predicted image for a second round of inference is a bit better. The method also performs a bit better (no statistical significance tests) than other post-processing methods (Dense-CRF, CRF-RNNs)\n\nExperimental results are available only on a small dataset and for two different networks. This may be sufficient for a first proof-of-concept but a comparison against standard benchmark methods and datasets for semantic segmentation is missing. It is unlikely that in the current state of this submission is a contribution to image segmentation, evidence is weak and several improvements are suggested.\n\n- The experimental evidence is insufficient. The improvements are small, statistical tests are not available. The CamVid dataset is the smallest of the image segmentation datasets used these days, more compelling would be MSCOCO or Cityscapes, better most of them. The question whether this network effect is tied to small-dataset and low-resolution is not answered. Will a similar effect be observed when compared to networks trained on way more data (e.g., CityScapes)? \n- The most important baseline is missing: auto-context [Tu08]. Training the same network the DAE uses in an auto-context way. That is, take the output of the first model, then train another network using both input and prediction again for semantic segmentation (and not Eq.3). This is easy to do, practically almost always achieves better performance and I would assume the resulting network is faster and performs similar to the method presented in this submission on (guessing, I have not tried). In any case, to me this is the most obvious baseline. \n- I am in favour of probabilistic methods, but the availability of an approximation of p(y) (or the nearest mode) is not used (as is most often the case).\n- Runtimes are absent. This is a practical consideration which is important especially if there is little technological improvement. The DAE model of this submission compares to simple filtering methods as Krähenbühl&Koltun DenseCRF which are fast and performance results are comparable. The question wether this is practically relevant is missing, judging from the construction I guess this does not fare well. Also training time is significantly more, please comment.\n- The related work is very well written, thanks. This proposal is conceptually very similar to auto-context [Tu08] and this reference missing (this is also the most important baseline)\n\n[Tu08] Tu, “Auto-context and its application to high-level vision tasks”, CVPR 2008\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Paper needs better experimental section and stronger baselines to validate the claimed contributions.", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes an iterative procedure on top of a standard image semantic segmentation networks. \n\nThe submission proposes a change to the training procedure of stacking a denoising auto-encoder for image segmentation. The technical contribution of this paper is small. The paper aims to answer a single question: When using a DAE network on top of a segmentation network output, should one condition on the predicted, or the ground truth segmentation? (why not on both?) The answer is conditioning on the predicted image for a second round of inference is a bit better. The method also performs a bit better (no statistical significance tests) than other post-processing methods (Dense-CRF, CRF-RNNs)\n\nExperimental results are available only on a small dataset and for two different networks. This may be sufficient for a first proof-of-concept but a comparison against standard benchmark methods and datasets for semantic segmentation is missing. It is unlikely that in the current state of this submission is a contribution to image segmentation, evidence is weak and several improvements are suggested.\n\n- The experimental evidence is insufficient. The improvements are small, statistical tests are not available. The CamVid dataset is the smallest of the image segmentation datasets used these days, more compelling would be MSCOCO or Cityscapes, better most of them. The question whether this network effect is tied to small-dataset and low-resolution is not answered. Will a similar effect be observed when compared to networks trained on way more data (e.g., CityScapes)? \n- The most important baseline is missing: auto-context [Tu08]. Training the same network the DAE uses in an auto-context way. That is, take the output of the first model, then train another network using both input and prediction again for semantic segmentation (and not Eq.3). This is easy to do, practically almost always achieves better performance and I would assume the resulting network is faster and performs similar to the method presented in this submission on (guessing, I have not tried). In any case, to me this is the most obvious baseline. \n- I am in favour of probabilistic methods, but the availability of an approximation of p(y) (or the nearest mode) is not used (as is most often the case).\n- Runtimes are absent. This is a practical consideration which is important especially if there is little technological improvement. The DAE model of this submission compares to simple filtering methods as Krähenbühl&Koltun DenseCRF which are fast and performance results are comparable. The question wether this is practically relevant is missing, judging from the construction I guess this does not fare well. Also training time is significantly more, please comment.\n- The related work is very well written, thanks. This proposal is conceptually very similar to auto-context [Tu08] and this reference missing (this is also the most important baseline)\n\n[Tu08] Tu, “Auto-context and its application to high-level vision tasks”, CVPR 2008\n\n\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1512365668541}, {"id": "B1hVeGtez", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper87/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an image segmentation method which iteratively refines the semantic segmentation mask obtained from a deep net. To this end the authors investigate a denoising auto-encoder (DAE). Its purpose is to provide a semantic segmentation which improves upon its input in terms of the log-likelihood.\n\nMore specifically, the authors `propose to condition the autoencoder with an additional input’ (page 1). To this end they use features obtained from the deep net. Instead of training the DAE with ground truth y, the authors found usage of the deep net prediction to yield better results.\n\nThe proposed approach is evaluated on the CamVid dataset.\n\nSummary:\n——\nI think the paper discusses a very interesting topic and presents an elegant approach. A few points are missing which would provide significantly more value to a reader. Specifically, an evaluation on the classical Pascal VOC dataset, details regarding the training protocol of the baseline (which are omitted right now), an assessment regarding stability of the proposed approach (not discussed right now), and a clear focus of the paper on segmentation or conditioning. See comments below for details and other points.\n\nComments:\n——\n1. When training the DAE, a combination of squared loss and categorical cross-entropy loss is used. What’s the effect of the squared error loss and would the categorical cross-entropy on its own be sufficient? This question remains open when reading the submission.\n\n2. The proposed approach is evaluated on the CamVid dataset which is used less compared to the standard and larger Pascal VOC dataset. I conjecture that the proposed approach wouldn’t work too well on Pascal VOC. On Pascal VOC, images are distinctly different from each other whereas subsequent frames are similar in CamVid, i.e., the road is always located at the bottom center of the image. The proposed architecture is able to take advantage of this dataset bias, but would fail to do so on Pascal VOC, which has a much more intricate bias. It would be great if the authors could check this hypothesis and report quantitative results similar to Tab. 1 and Fig. 4 for Pascal VOC.\n\n3. The authors mention a grid-search for the stepsize and the number of iterations. What values were selected in the end on the CamVid and hopefully the Pascal VOC dataset?\n\n4. Was the dense CRF applied out of the box, or were its parameters adjusted for good performance on the CamVid validation dataset? While parameters such as the number of iterations and epsilon are tuned for the proposed approach on the CamVid validation set, the submission doesn’t specify whether a similar procedure was performed for the CRF baseline.\n\n5. Fig. 4 seems to indicate that the proposed approach doesn’t converge. Hence an appropriate stepsize and a reasonable number of iterations need to be chosen on a validation set. Choosing those parameters guarantees that the method performs well on average, but individual results could potentially be entirely wrong, particularly if large step sizes are chosen. I suspect this effect to be more pronounced on the Pascal VOC dataset (hence my conjecture in point 2). To further investigate this property, as a reader, I’d be curious to get to know the standard deviation/variance of the accuracy in addition to the mean IoU. Again, it would be great if the authors could check this hypothesis and report those results.\n\n6. I find the experimental section to be slightly disconnected from the initial description. Specifically, the paper `proposes to condition the autoencoder with an additional input’ (page 1). No experiments are conducted to validate this proposal. Hence the main focus of the paper (image segmentation or DAE conditioning) remains vague. If the authors choose to focus on image segmentation, a comparison to state-of-the-art should be provided on classical datasets such as Pascal VOC, if DAE conditioning is the focus, some experiments in this direction should be included in addition to the Pascal VOC results.\n\nMinor comment:\n——\n- I find it surprising that the authors choose not to cite some related work on combining deep nets with structured prediction.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "rating": "5: Marginally below acceptance threshold", "review": "The paper proposes an image segmentation method which iteratively refines the semantic segmentation mask obtained from a deep net. To this end the authors investigate a denoising auto-encoder (DAE). Its purpose is to provide a semantic segmentation which improves upon its input in terms of the log-likelihood.\n\nMore specifically, the authors `propose to condition the autoencoder with an additional input’ (page 1). To this end they use features obtained from the deep net. Instead of training the DAE with ground truth y, the authors found usage of the deep net prediction to yield better results.\n\nThe proposed approach is evaluated on the CamVid dataset.\n\nSummary:\n——\nI think the paper discusses a very interesting topic and presents an elegant approach. A few points are missing which would provide significantly more value to a reader. Specifically, an evaluation on the classical Pascal VOC dataset, details regarding the training protocol of the baseline (which are omitted right now), an assessment regarding stability of the proposed approach (not discussed right now), and a clear focus of the paper on segmentation or conditioning. See comments below for details and other points.\n\nComments:\n——\n1. When training the DAE, a combination of squared loss and categorical cross-entropy loss is used. What’s the effect of the squared error loss and would the categorical cross-entropy on its own be sufficient? This question remains open when reading the submission.\n\n2. The proposed approach is evaluated on the CamVid dataset which is used less compared to the standard and larger Pascal VOC dataset. I conjecture that the proposed approach wouldn’t work too well on Pascal VOC. On Pascal VOC, images are distinctly different from each other whereas subsequent frames are similar in CamVid, i.e., the road is always located at the bottom center of the image. The proposed architecture is able to take advantage of this dataset bias, but would fail to do so on Pascal VOC, which has a much more intricate bias. It would be great if the authors could check this hypothesis and report quantitative results similar to Tab. 1 and Fig. 4 for Pascal VOC.\n\n3. The authors mention a grid-search for the stepsize and the number of iterations. What values were selected in the end on the CamVid and hopefully the Pascal VOC dataset?\n\n4. Was the dense CRF applied out of the box, or were its parameters adjusted for good performance on the CamVid validation dataset? While parameters such as the number of iterations and epsilon are tuned for the proposed approach on the CamVid validation set, the submission doesn’t specify whether a similar procedure was performed for the CRF baseline.\n\n5. Fig. 4 seems to indicate that the proposed approach doesn’t converge. Hence an appropriate stepsize and a reasonable number of iterations need to be chosen on a validation set. Choosing those parameters guarantees that the method performs well on average, but individual results could potentially be entirely wrong, particularly if large step sizes are chosen. I suspect this effect to be more pronounced on the Pascal VOC dataset (hence my conjecture in point 2). To further investigate this property, as a reader, I’d be curious to get to know the standard deviation/variance of the accuracy in addition to the mean IoU. Again, it would be great if the authors could check this hypothesis and report those results.\n\n6. I find the experimental section to be slightly disconnected from the initial description. Specifically, the paper `proposes to condition the autoencoder with an additional input’ (page 1). No experiments are conducted to validate this proposal. Hence the main focus of the paper (image segmentation or DAE conditioning) remains vague. If the authors choose to focus on image segmentation, a comparison to state-of-the-art should be provided on classical datasets such as Pascal VOC, if DAE conditioning is the focus, some experiments in this direction should be included in addition to the Pascal VOC results.\n\nMinor comment:\n——\n- I find it surprising that the authors choose not to cite some related work on combining deep nets with structured prediction.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511755827790}], "openreview_url": "https://openreview.net/forum?id=HklpCzC6-", "arxiv_id": "1705.07450", "paper_pdf": "papers/HklpCzC6-.pdf", "paper_pdf_sha256": "bb797bccb459d007147e87e7736b81eb9d93ba0635d1bb475675e706e25a3b14", "paper_pdf_bytes": 3639524, "paper_pdf_source": "openreview", "code_url": "https://github.com/adri-romsor/iterative_inference_segm", "code_repository": "adri-romsor/iterative_inference_segm", "code_commit": "33e3a988034b9c807ad15c9d0c2ab8bb32d79a39", "code_archive": "repos/HklpCzC6-.zip", "code_archive_sha256": "ad43313e1c86ef87627b75a49c52fc03dccf205db50e6188d8551439283ed79d", "code_archive_bytes": 101174, "code_file_count": 25, "code_extensions": {".py": 23, ".ipynb": 1, ".sh": 1}, "github_disk_usage_kb": 453, "github_languages": {"Python": 196495, "Jupyter Notebook": 60338, "Shell": 193}, "github_archived": false, "github_pushed_at": "2017-08-07T21:42:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/image-segmentation-by-iterative-inference"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rpblsD3eXG", "year": 2026, "status": "rejected", "title": "Communication-Efficient Multi-Device Inference Acceleration for Transformer Models", "authors": ["Xiao Liu", "Lijun Zhang", "Deepak Ganesan", "Hui Guan"], "authorids": ["~Xiao_Liu9", "~Lijun_Zhang4", "~Deepak_Ganesan1", "~Hui_Guan1"], "authors_source": "OpenReview API", "abstract": "Transformer models power many AI applications but suffer from high inference latency, limiting their use in real-time settings. Multi-device inference can reduce latency by parallelizing computation. Yet, existing methods require high inter-device bandwidth, making them impractical for bandwidth-constrained environments. We propose ASTRA, a communication-efficient framework that accelerates Transformer inference through a novel integration of sequence parallelism and a Mixed-Precision Attention mechanism designed to minimize inter-device communication. ASTRA compresses non-local token embeddings via vector quantization and preserves task accuracy through two optimizations, Noise-Augmented Quantization and Distributed Class Tokens. Experiments on ViT and GPT2 across vision and NLP tasks show that Astra achieves up to 2.64$\\times$ speedups over single-device inference and up to 15.25$\\times$ speedups over state-of-the-art multi-device inferences, while operating under bandwidths as low as 10 Mbps.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "i5ZhHPkGoU", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16127/Reviewer_7h8e"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 2, "summary": "The paper introduces a way to improve communication efficiency for multi device Transformer inference using sequence parallelism and mixed-precision. They show significant speedups on ViT and GPT2 scale models.", "review_text": "The paper introduces a way to improve communication efficiency for multi device Transformer inference using sequence parallelism and mixed-precision. They show significant speedups on ViT and GPT2 scale models.", "strengths": "- The paper is well-presented and easy to follow\n- Communication overhead is significant in large Transformer model distributed settings\n- The use of codebooks is interesting", "weaknesses": "- The models used for inference are small and it is not clear to me that these hold at scale.", "questions": "- It would be helpful to further motivate the wireless and edge deployment motivation for this framework. Why are we doing inference requests on wifi?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a way to improve communication efficiency for multi device Transformer inference using sequence parallelism and mixed-precision. They show significant speedups on ViT and GPT2 scale models.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- The paper is well-presented and easy to follow\n- Communication overhead is significant in large Transformer model distributed settings\n- The use of codebooks is interesting", "weaknesses": "- The models used for inference are small and it is not clear to me that these hold at scale.", "questions": "- It would be helpful to further motivate the wireless and edge deployment motivation for this framework. Why are we doing inference requests on wifi?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1762158970672}, {"id": "sWHe6yvnlR", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16127/Reviewer_mHin"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces ASTRA, a framework for accelerating transformer inference across multiple devices in bandwidth-constrained environments. The key part is a Mixed-Precision Attention mechanism that computes local attention with full precision while using vector-quantized embeddings for non-local tokens, reducing communication overhead. To preserve accuracy under aggressive compression, the authors propose Noise-Augmented Vector Quantization (NAVQ) and Distributed Class Tokens. Experiments on ViT and GPT-2 models across vision and NLP tasks demonstrate speedups of up to 2.64x over single-device inference and 15.25x over existing multi-device methods at bandwidths as low as 10 Mbps, while maintaining accuracy within 3.58% of the original models.", "review_text": "This paper introduces ASTRA, a framework for accelerating transformer inference across multiple devices in bandwidth-constrained environments. The key part is a Mixed-Precision Attention mechanism that computes local attention with full precision while using vector-quantized embeddings for non-local tokens, reducing communication overhead. To preserve accuracy under aggressive compression, the authors propose Noise-Augmented Vector Quantization (NAVQ) and Distributed Class Tokens. Experiments on ViT and GPT-2 models across vision and NLP tasks demonstrate speedups of up to 2.64x over single-device inference and 15.25x over existing multi-device methods at bandwidths as low as 10 Mbps, while maintaining accuracy within 3.58% of the original models.", "strengths": "1. Addresses a real bottleneck: The paper identifies and tackles a genuine problem, that communication dominates latency (58.6-93.5%) in bandwidth-constrained multi-device inference.\n2. Novel compression approach: The Mixed-Precision Attention mechanism is creative, using full-precision for local tokens and VQ for remote tokens.\n3. Good evaluation: Extensive experiments across multiple architectures (ViT, GPT-2), tasks (classification, language modeling), and conditions (bandwidth, device count, heterogeneity) demonstrate broad applicability.\n4. Practical compatibility: The framework integrates with existing quantization methods (8-bit, 4-bit), showing additional speedups of 1.35 to 2.73x when combined.", "weaknesses": "1. Limited architectural types: The evaluation focuses only on ViT and GPT-2, which are relatively small and dated models. Modern applications use much larger models (e.g., LLaMA variants). The scalability claims are weakened without evidence on contemporary, production-scale models.\n\n2. Severe zero-shot degradation: Table 3 shows large performance drops in zero-shot settings (e.g., GPT-2M perplexity increases from 43.22 to 62.29, a 44% degradation). This is a critical limitation for practical deployment where generalization is essential. \n\n3. Baselines potentially unfair: The comparison with BP, TP, and SP assumes these methods use full float32 precision. However, these methods could also be combined with standard compression techniques (gradient compression, activation compression). A more fair comparison would evaluate \"BP+8bit quantization\" vs \"ASTRA+8bit quantization\" to isolate ASTRA's contribution. Additionally, recent methods like FlexGen or DistServe are not compared, making it unclear how ASTRA compares to the current sota.\n\n4. Communication model oversimplified: The paper assumes fixed bandwidth and doesn't account for real-world network variability, packet loss, or latency jitter. The latency model appears to assume perfect overlap of computation and communication, which is rarely achievable. Dynamic bandwidth fluctuations common in WiFi environments (cited as the target deployment) could significantly impact the practical speedups. The authors should evaluate under realistic network conditions with variable bandwidth and packet loss, or at minimum discuss how ASTRA degrades under non-ideal conditions.", "questions": "Please address the implicitly listed questions in the weakness.\n\nCan you provide a decision tree or heuristic for practitioners to select:\n1. Number of groups G based on task type (vision vs. NLP) and target accuracy?\n2. Commitment loss weight ε based on model architecture and dataset?\n3. When to use distributed vs. single class tokens?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ASTRA, a framework for accelerating transformer inference across multiple devices in bandwidth-constrained environments. The key part is a Mixed-Precision Attention mechanism that computes local attention with full precision while using vector-quantized embeddings for non-local tokens, reducing communication overhead. To preserve accuracy under aggressive compression, the authors propose Noise-Augmented Vector Quantization (NAVQ) and Distributed Class Tokens. Experiments on ViT and GPT-2 models across vision and NLP tasks demonstrate speedups of up to 2.64x over single-device inference and 15.25x over existing multi-device methods at bandwidths as low as 10 Mbps, while maintaining accuracy within 3.58% of the original models.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Addresses a real bottleneck: The paper identifies and tackles a genuine problem, that communication dominates latency (58.6-93.5%) in bandwidth-constrained multi-device inference.\n2. Novel compression approach: The Mixed-Precision Attention mechanism is creative, using full-precision for local tokens and VQ for remote tokens.\n3. Good evaluation: Extensive experiments across multiple architectures (ViT, GPT-2), tasks (classification, language modeling), and conditions (bandwidth, device count, heterogeneity) demonstrate broad applicability.\n4. Practical compatibility: The framework integrates with existing quantization methods (8-bit, 4-bit), showing additional speedups of 1.35 to 2.73x when combined.", "weaknesses": "1. Limited architectural types: The evaluation focuses only on ViT and GPT-2, which are relatively small and dated models. Modern applications use much larger models (e.g., LLaMA variants). The scalability claims are weakened without evidence on contemporary, production-scale models.\n\n2. Severe zero-shot degradation: Table 3 shows large performance drops in zero-shot settings (e.g., GPT-2M perplexity increases from 43.22 to 62.29, a 44% degradation). This is a critical limitation for practical deployment where generalization is essential. \n\n3. Baselines potentially unfair: The comparison with BP, TP, and SP assumes these methods use full float32 precision. However, these methods could also be combined with standard compression techniques (gradient compression, activation compression). A more fair comparison would evaluate \"BP+8bit quantization\" vs \"ASTRA+8bit quantization\" to isolate ASTRA's contribution. Additionally, recent methods like FlexGen or DistServe are not compared, making it unclear how ASTRA compares to the current sota.\n\n4. Communication model oversimplified: The paper assumes fixed bandwidth and doesn't account for real-world network variability, packet loss, or latency jitter. The latency model appears to assume perfect overlap of computation and communication, which is rarely achievable. Dynamic bandwidth fluctuations common in WiFi environments (cited as the target deployment) could significantly impact the practical speedups. The authors should evaluate under realistic network conditions with variable bandwidth and packet loss, or at minimum discuss how ASTRA degrades under non-ideal conditions.", "questions": "Please address the implicitly listed questions in the weakness.\n\nCan you provide a decision tree or heuristic for practitioners to select:\n1. Number of groups G based on task type (vision vs. NLP) and target accuracy?\n2. Commitment loss weight ε based on model architecture and dataset?\n3. When to use distributed vs. single class tokens?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761986255678}, {"id": "lzdJwESrE6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16127/Reviewer_24hX"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces ASTRA, a communication-efficient framework for multi-device Transformer inference under bandwidth-constrained settings. Existing multi-device methods suffer from high inter-device communication overhead, which dominates latency when bandwidth is limited. ASTRA addresses this by combining sequence parallelism with a Mixed-Precision Attention mechanism: local attention is computed at full precision, while remote tokens are compressed via low-bit vector quantization.", "review_text": "This paper introduces ASTRA, a communication-efficient framework for multi-device Transformer inference under bandwidth-constrained settings. Existing multi-device methods suffer from high inter-device communication overhead, which dominates latency when bandwidth is limited. ASTRA addresses this by combining sequence parallelism with a Mixed-Precision Attention mechanism: local attention is computed at full precision, while remote tokens are compressed via low-bit vector quantization.", "strengths": "This paper observes a real bottleneck in multi-device Transformer inference for low-bandwidth or edge environments, which is increasingly relevant for real-time AI applications.", "weaknesses": "1. ASTRA integrates known techniques (sequence parallelism + token quantization + noise augmentation), so the main contribution is in practical integration and bandwidth optimization, not a fundamentally new inference algorithm.\n2. Lacks formal characterization of attention approximation error due to vector quantization and noise injection. Consider adding error bounds or theoretical analysis of how quantization and noise affect attention computation and model accuracy.\n3. Experiments assume stable bandwidth; network variability or heterogeneous device scenarios are not tested.\n4. Latency and energy claims are based on simulation; real-world multi-device deployment is not evaluated. Incorporate hardware-level profiling (multi-GPU, FPGA, or edge devices) to substantiate speedup and efficiency claims.", "questions": "see weakness", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "all_content": {"summary": "This paper introduces ASTRA, a communication-efficient framework for multi-device Transformer inference under bandwidth-constrained settings. Existing multi-device methods suffer from high inter-device communication overhead, which dominates latency when bandwidth is limited. ASTRA addresses this by combining sequence parallelism with a Mixed-Precision Attention mechanism: local attention is computed at full precision, while remote tokens are compressed via low-bit vector quantization.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "This paper observes a real bottleneck in multi-device Transformer inference for low-bandwidth or edge environments, which is increasingly relevant for real-time AI applications.", "weaknesses": "1. ASTRA integrates known techniques (sequence parallelism + token quantization + noise augmentation), so the main contribution is in practical integration and bandwidth optimization, not a fundamentally new inference algorithm.\n2. Lacks formal characterization of attention approximation error due to vector quantization and noise injection. Consider adding error bounds or theoretical analysis of how quantization and noise affect attention computation and model accuracy.\n3. Experiments assume stable bandwidth; network variability or heterogeneous device scenarios are not tested.\n4. Latency and energy claims are based on simulation; real-world multi-device deployment is not evaluated. Incorporate hardware-level profiling (multi-GPU, FPGA, or edge devices) to substantiate speedup and efficiency claims.", "questions": "see weakness", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761792215120}, {"id": "SkLiFSRlTs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16127/Reviewer_3hGB"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper addresses the significant challenge of high inference latency in Transformer models, particularly in multi-device scenarios where inter-device communication becomes a dominant bottleneck in low-bandwidth environments. The authors propose ASTRA, a communication-efficient inference framework that builds on sequence parallelism but introduces a novel Mixed-Precision Attention mechanism to drastically reduce communication overhead. This mechanism computes attention using full-precision embeddings for local tokens while using low-bit vector-quantized (VQ) representations for non-local tokens transmitted between devices. To preserve model accuracy under such aggressive compression, ASTRA introduces two key optimizations: Noise-Augmented Quantization (NAVQ), a training-time regularization strategy that injects noise into quantized embeddings to improve generalization , and Distributed Class Tokens, which replicates the class token to each device and aggregates the outputs to mitigate information bias. Experiments on ViT and GPT-2 models show that ASTRA achieves substantial end-to-end speedups -- up to 2.64x over single-device inference and 15.25x over other multi-device methods -- in bandwidth-constrained settings (as low as 10 Mbps), while incurring only minor accuracy degradation.", "review_text": "This paper addresses the significant challenge of high inference latency in Transformer models, particularly in multi-device scenarios where inter-device communication becomes a dominant bottleneck in low-bandwidth environments. The authors propose ASTRA, a communication-efficient inference framework that builds on sequence parallelism but introduces a novel Mixed-Precision Attention mechanism to drastically reduce communication overhead. This mechanism computes attention using full-precision embeddings for local tokens while using low-bit vector-quantized (VQ) representations for non-local tokens transmitted between devices. To preserve model accuracy under such aggressive compression, ASTRA introduces two key optimizations: Noise-Augmented Quantization (NAVQ), a training-time regularization strategy that injects noise into quantized embeddings to improve generalization , and Distributed Class Tokens, which replicates the class token to each device and aggregates the outputs to mitigate information bias. Experiments on ViT and GPT-2 models show that ASTRA achieves substantial end-to-end speedups -- up to 2.64x over single-device inference and 15.25x over other multi-device methods -- in bandwidth-constrained settings (as low as 10 Mbps), while incurring only minor accuracy degradation.", "strengths": "The paper is pretty interesting wtih 2 proposed optimizations to speedup inference across multiple device. I am not a systems expert, but the math behind the optimizations is correct and makes sense.", "weaknesses": "The biggest weakness and drawback from the paper is the lack of large-scale experiments. Espeically the model choices in this paper is so small that none of these optimizations matter. Given the prevelance of open models of much larger sizes even ViTs and Qwen series models, it is empirically needed to make sure the proposed optimizations are quality neutral while providing the benefits. Without these results it hard to make a case for accepting the paper. I hope the authors can scale up the benchmarking further.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the significant challenge of high inference latency in Transformer models, particularly in multi-device scenarios where inter-device communication becomes a dominant bottleneck in low-bandwidth environments. The authors propose ASTRA, a communication-efficient inference framework that builds on sequence parallelism but introduces a novel Mixed-Precision Attention mechanism to drastically reduce communication overhead. This mechanism computes attention using full-precision embeddings for local tokens while using low-bit vector-quantized (VQ) representations for non-local tokens transmitted between devices. To preserve model accuracy under such aggressive compression, ASTRA introduces two key optimizations: Noise-Augmented Quantization (NAVQ), a training-time regularization strategy that injects noise into quantized embeddings to improve generalization , and Distributed Class Tokens, which replicates the class token to each device and aggregates the outputs to mitigate information bias. Experiments on ViT and GPT-2 models show that ASTRA achieves substantial end-to-end speedups -- up to 2.64x over single-device inference and 15.25x over other multi-device methods -- in bandwidth-constrained settings (as low as 10 Mbps), while incurring only minor accuracy degradation.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper is pretty interesting wtih 2 proposed optimizations to speedup inference across multiple device. I am not a systems expert, but the math behind the optimizations is correct and makes sense.", "weaknesses": "The biggest weakness and drawback from the paper is the lack of large-scale experiments. Espeically the model choices in this paper is so small that none of these optimizations matter. Given the prevelance of open models of much larger sizes even ViTs and Qwen series models, it is empirically needed to make sure the proposed optimizations are quality neutral while providing the benefits. Without these results it hard to make a case for accepting the paper. I hope the authors can scale up the benchmarking further.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760734083358}], "openreview_url": "https://openreview.net/forum?id=rpblsD3eXG", "arxiv_id": "2505.19342", "paper_pdf": "papers/rpblsD3eXG.pdf", "paper_pdf_sha256": "b8aa60f7841b593dec32649b2cb76c11783fcbb6cd2ae6db3262dde2112d6299", "paper_pdf_bytes": 942944, "paper_pdf_source": "openreview", "code_url": "https://github.com/xl1990/Astra", "code_repository": "xl1990/Astra", "code_commit": "c959791453fb5627027825b019ab20ca9384e6ad", "code_archive": "repos/rpblsD3eXG.zip", "code_archive_sha256": "1ee487edbae1c968c037536f1d429a3fb0716102b983ab3e51d1c1d12658a99a", "code_archive_bytes": 519372, "code_file_count": 14, "code_extensions": {".py": 9, ".sh": 5}, "github_disk_usage_kb": 224, "github_languages": {"Python": 90349, "Shell": 6165}, "github_archived": false, "github_pushed_at": "2025-05-25T22:05:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/communication-efficient-multi-device"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rnRBGMNYa2", "year": 2025, "status": "rejected", "title": "Response Tuning: Aligning Large Language Models without Instruction", "authors": ["Seokhyun An", "Hyounghun Kim"], "authorids": ["~Seokhyun_An1", "~Hyounghun_Kim1"], "authors_source": "OpenReview API", "abstract": "Instruction tuning—supervised fine-tuning using instruction-response pairs—is a foundational step in transitioning pre-trained Large Language Models (LLMs) into helpful and safe chat assistants. Our hypothesis is that establishing an adequate output distribution can enable such a transition given the capabilities inherent in pre-trained LLMs. To verify this, we propose Response Tuning (RT), which eliminates the instruction-conditioning step in instruction tuning and solely focuses on response distribution supervision. Our experiments demonstrate that RT models, trained only using responses, can effectively respond to a wide range of instructions and exhibit helpfulness comparable to that of their instruction-tuned counterparts. Furthermore, we observe that controlling the training response distribution can substantially improve their user preference or elicit target behaviors such as refusing assistance for unsafe queries. Our findings illuminate the role of establishing an adequate output distribution in alignment, highlighting the potential of the extensive inherent capabilities of pre-trained LLMs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "BeohvFcSU0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9542/Reviewer_AfZU"], "rating": 5, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 5, "summary": "The main question this paper tries to answer is \"how much do we lose by tuning just on the responses (response tuning) instead of instruction-response pairs (instruction tuning)\". The authors use existing IT datasets and remove instructions from them to get datasets for response tuning. They fine-tune a variety of base-llms (not instruction tuned) and compare IT vs RT. They evaluate instructability with human evaluation, GPT-4 based response quality evaluation and pairwise assessment. They also evaluate core capabilities across standard benchmarks like MMLU, OpenbookQA, HellaSwag, ARC, GSM 8K, PIQA. Across a wide range of experiments, they show that response tuning comes close to instruction tuning. Additionally, they show the application of this method for preference alignment and safety alignment. Finally, they show that a similar trend holds for in-context learning, where using just responses doesn't degrade performance significantly compared to using instruction-response pairs.", "review_text": "The main question this paper tries to answer is \"how much do we lose by tuning just on the responses (response tuning) instead of instruction-response pairs (instruction tuning)\". The authors use existing IT datasets and remove instructions from them to get datasets for response tuning. They fine-tune a variety of base-llms (not instruction tuned) and compare IT vs RT. They evaluate instructability with human evaluation, GPT-4 based response quality evaluation and pairwise assessment. They also evaluate core capabilities across standard benchmarks like MMLU, OpenbookQA, HellaSwag, ARC, GSM 8K, PIQA. Across a wide range of experiments, they show that response tuning comes close to instruction tuning. Additionally, they show the application of this method for preference alignment and safety alignment. Finally, they show that a similar trend holds for in-context learning, where using just responses doesn't degrade performance significantly compared to using instruction-response pairs.", "strengths": "* originality - The authors ask a simple question that hasn't been considered from a scientific point of view. \"Can we use good responses instead of instruction-response pairs\"\n* quality - The authors run a wide variety of experiments to make their claim stronger and well substantiated. \n* clarity - The paper is well-written and easy to read. In most places all the necessary details and citations are provided\n* significance - There is the scientific question on the source of gains due to instruction tuning. From a practical perspective, this raises the question whether, as a community, we should invest more energy looking at instruction-free response datasets which are free of instruction biases.", "weaknesses": "There are two main concerns regarding the soundness of these experiments\n## Leaking instructions to response\nThe datasets in section 4, 5 and 6 are collected as instruction - response pairs. The data generation process may induce a leak from the instruction to the response. For e.g. if the instruction was \"Write an article on the role of dinosaurs in prehistoric ecology\", and the response was \"Dinosaurs played a significant role in prehistoric ecology ...\", the instruction can be inferred based on the response. In a trivial situation, for an instruction \"Follow the rules x, y, z\", the response could be \"following the rules x, y, z, here's the response ...\". To argue that one needs only the response (and not the instruction), the authors would need to show that the instruction cannot be trivially inferred from the response. Alternatively, an experiment training on high quality responses gathered without explicit instructions could also prove the point. \n\n## Training setup (Section 4.1)\nThe authors use PEFT (QLoRA) but don't give any references that this is a standard setup for instruction tuning. It is possible that a less expressive training scheme (PEFT vs full fine-tuning) suppresses the utility of having instructions in the training data. Alternatively, if they could show that their PEFT-trained IT models are comparable to the best IT models on these datasets, that could alleviate this concern too.", "questions": "My two main questions with potential fixes are indicated in the weaknesses section. Convincing experimental results for the two weaknesses will change my stance on soundness and the overall rating.   \n\nSome suggestions that aren't critical to soundness\n* \"establishing an output space\" is less standard than \"establishing an output distribution\". The authors do use \"distribution\" at various places in the paper, so it isn't clear why not use the same standard term everywhere (including the abstract)\n* The authors use certain terms that aren't well-defined. 073 - \"contextual refusals\", \n* line 311 - \"some fluctuations\" - why are the fluctuations attributed to fine-tuning? if there's a good reason, it should be stated, if not, it is sufficient to state that the numbers are comparable. \n* line 333 - what does the up arrow mean?\n* line 428 - what exactly is meant by \"pattern matching\"?\n- Figure 4b - IT win-rate is 60% compared to RT win rate of 40%. This is a big difference but there is no discussion of this negative result in the text. \n\nThe authors consistently use the narrative that \"capabilities required to behave as a chat assistant are largely inherent in pre-trained LLMs and can be activated without explicit supervision from instruction-response pairs\" (lines 096, 515 and perhaps more). However, there is no direct proof for the \"capabilities being inherent in pre-trained LLMs\". One can argue that training on responses adds these capabilities that aren't inherent, just like instruction-response pairs are supposedly adding these capabilities. This claim is unnecessary and unsubstantiated. This claim is unnecessarily distracting and I would urge the authors to stay focused on the central claim of response tuning being at-par with instruction tuning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The main question this paper tries to answer is \"how much do we lose by tuning just on the responses (response tuning) instead of instruction-response pairs (instruction tuning)\". The authors use existing IT datasets and remove instructions from them to get datasets for response tuning. They fine-tune a variety of base-llms (not instruction tuned) and compare IT vs RT. They evaluate instructability with human evaluation, GPT-4 based response quality evaluation and pairwise assessment. They also evaluate core capabilities across standard benchmarks like MMLU, OpenbookQA, HellaSwag, ARC, GSM 8K, PIQA. Across a wide range of experiments, they show that response tuning comes close to instruction tuning. Additionally, they show the application of this method for preference alignment and safety alignment. Finally, they show that a similar trend holds for in-context learning, where using just responses doesn't degrade performance significantly compared to using instruction-response pairs.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "* originality - The authors ask a simple question that hasn't been considered from a scientific point of view. \"Can we use good responses instead of instruction-response pairs\"\n* quality - The authors run a wide variety of experiments to make their claim stronger and well substantiated. \n* clarity - The paper is well-written and easy to read. In most places all the necessary details and citations are provided\n* significance - There is the scientific question on the source of gains due to instruction tuning. From a practical perspective, this raises the question whether, as a community, we should invest more energy looking at instruction-free response datasets which are free of instruction biases.", "weaknesses": "There are two main concerns regarding the soundness of these experiments\n## Leaking instructions to response\nThe datasets in section 4, 5 and 6 are collected as instruction - response pairs. The data generation process may induce a leak from the instruction to the response. For e.g. if the instruction was \"Write an article on the role of dinosaurs in prehistoric ecology\", and the response was \"Dinosaurs played a significant role in prehistoric ecology ...\", the instruction can be inferred based on the response. In a trivial situation, for an instruction \"Follow the rules x, y, z\", the response could be \"following the rules x, y, z, here's the response ...\". To argue that one needs only the response (and not the instruction), the authors would need to show that the instruction cannot be trivially inferred from the response. Alternatively, an experiment training on high quality responses gathered without explicit instructions could also prove the point. \n\n## Training setup (Section 4.1)\nThe authors use PEFT (QLoRA) but don't give any references that this is a standard setup for instruction tuning. It is possible that a less expressive training scheme (PEFT vs full fine-tuning) suppresses the utility of having instructions in the training data. Alternatively, if they could show that their PEFT-trained IT models are comparable to the best IT models on these datasets, that could alleviate this concern too.", "questions": "My two main questions with potential fixes are indicated in the weaknesses section. Convincing experimental results for the two weaknesses will change my stance on soundness and the overall rating.   \n\nSome suggestions that aren't critical to soundness\n* \"establishing an output space\" is less standard than \"establishing an output distribution\". The authors do use \"distribution\" at various places in the paper, so it isn't clear why not use the same standard term everywhere (including the abstract)\n* The authors use certain terms that aren't well-defined. 073 - \"contextual refusals\", \n* line 311 - \"some fluctuations\" - why are the fluctuations attributed to fine-tuning? if there's a good reason, it should be stated, if not, it is sufficient to state that the numbers are comparable. \n* line 333 - what does the up arrow mean?\n* line 428 - what exactly is meant by \"pattern matching\"?\n- Figure 4b - IT win-rate is 60% compared to RT win rate of 40%. This is a big difference but there is no discussion of this negative result in the text. \n\nThe authors consistently use the narrative that \"capabilities required to behave as a chat assistant are largely inherent in pre-trained LLMs and can be activated without explicit supervision from instruction-response pairs\" (lines 096, 515 and perhaps more). However, there is no direct proof for the \"capabilities being inherent in pre-trained LLMs\". One can argue that training on responses adds these capabilities that aren't inherent, just like instruction-response pairs are supposedly adding these capabilities. This claim is unnecessary and unsubstantiated. This claim is unnecessarily distracting and I would urge the authors to stay focused on the central claim of response tuning being at-par with instruction tuning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730705961622}, {"id": "hCeTIMdgfS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9542/Reviewer_6b1i"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper demonstrates that only fine-tuning base models on the responses in instruction datasets (response tuning, or RT) often results in comparable performance to fine-tuning on the responses *conditioned on the instructions* (traditional instruction tuning, or IT) without losing any core model capabilities. This supports the authors' hypothesis that models have already learned instruction following ability during pretraining, and instruction tuning thus mainly facilitates the model placing higher probabilities on tokens associated with helpful responses at inference time, or what the authors call \"aligning the training response distribution\". The authors measure response quality using 4 metrics: Likert-style individual response acceptability ratings + pairwise response preference selection using both human annotators and GPT4-as-a-judge. They find results to be consistent across 3 models from different families and on 3 instruction-tuning datasets, and also find that a larger Gemma model closes the gap between RT and IT more than a smaller one.\n\nThe authors further investigate whether using an LLM to refine the response quality along dimensions such as clarity, structure, and tone before training improves model response quality (as judged by GPT4), and find a positive result.\n\nFinally, they extend their experimental setting to refusal, and show that only training on responses from refusal/safety datasets (and not the unsafe queries the responses are associated with) also yields comparable performance, without causing the model to over-refuse. This is particularly interesting, because it supports the hypothesis that models learn to identify unsafe queries during pretraining, but need tuning to learn the \"language\" of refusal.", "review_text": "This paper demonstrates that only fine-tuning base models on the responses in instruction datasets (response tuning, or RT) often results in comparable performance to fine-tuning on the responses *conditioned on the instructions* (traditional instruction tuning, or IT) without losing any core model capabilities. This supports the authors' hypothesis that models have already learned instruction following ability during pretraining, and instruction tuning thus mainly facilitates the model placing higher probabilities on tokens associated with helpful responses at inference time, or what the authors call \"aligning the training response distribution\". The authors measure response quality using 4 metrics: Likert-style individual response acceptability ratings + pairwise response preference selection using both human annotators and GPT4-as-a-judge. They find results to be consistent across 3 models from different families and on 3 instruction-tuning datasets, and also find that a larger Gemma model closes the gap between RT and IT more than a smaller one.\n\nThe authors further investigate whether using an LLM to refine the response quality along dimensions such as clarity, structure, and tone before training improves model response quality (as judged by GPT4), and find a positive result.\n\nFinally, they extend their experimental setting to refusal, and show that only training on responses from refusal/safety datasets (and not the unsafe queries the responses are associated with) also yields comparable performance, without causing the model to over-refuse. This is particularly interesting, because it supports the hypothesis that models learn to identify unsafe queries during pretraining, but need tuning to learn the \"language\" of refusal.", "strengths": "- Originality: the paper takes a creative approach to an underexplored research question.\n- Quality: The paper includes quite extensive experiments to answer 3 main RQs and is not lacking on content. The authors include multiple datasets and models in their analysis. I appreciate that they use both human and model evaluation that is generally well-controlled for biases (e.g., not using the same model/family to evaluate another model); given that both evaluation methods have known shortcomings, their results seem more robust due to the extensive effort given to evaluation.\n- Clarity: The paper is generally well-written. Given there are a large amount of experimental results, the authors make good use of the Appendix for additional details while also making sure the main paper is readable and self-contained. I appreciate the delineation of the paper into 3 main sections corresponding to the 3 RQs of interest. Generally comprehensive lit review.\n- Significance: The paper provides a useful datapoint an interesting insight to ongoing discussion about what models learn during instruction tuning. While the 2nd RQ is a bit orthogonal to the paper's narrative, the 1st and 3rd RQs are interesting and impactful. I would like to see it presented at the conference.", "weaknesses": "Main:\n- My main qualm with the paper is the authors use the word \"comparable\" liberally when describing the performance of RT and IT tuning. However, in most cases, it is consistently the case that IT outperforms RT, often by what looks like rather substantial margins (e.g., Figures 2 and 4). I agree with the authors that it's really interesting that RT can make up the majority of the gap between the untuned and IT models, but I think the language can be toned down a bit and more discussion given to what performance aspects cannot be solely learned from RT alone. Alternatively, the authors could provide a quantitative definition of what they consider comparable performance.\nSimilarly, some of the differences in core capabilities (Table 1) seem meaningful (such as +7 on HellaSwag for Gemma 2 and -7 on GSM8k for Llama 3.1), but these are mostly chalked up to being typical or expected noise. It would be valuable to include a line for the untuned/base model in this table as a comparison point.\n- The URIAL paper should certainly be discussed in the related work section as it asks a similar RQ and has similar conclusions to this paper (though different methods).\n- The authors make a key assumption about the data they instruction- and response-tune on: that it hasn't been included in the pretraining data of the models in a meaningful way. This assumption seems to be supported by the poor performance of the untuned models in, e.g., Figure 2, but it should be stated more concretely. \n- Practical impact: from an analysis/model understanding perspective, the paper is strong. The authors do not explain, however, what impacts only training on responses could have from a practical angle. Presumably, it is not cheaper to collect the response data, since human or synthetic data generation pipelines would still require the instructions or refusal query to be present. It could be faster to train the model when one does not have to run inference on the instruction tokens, though the loss computation is the same so this may be minor. I would appreciate some discussion about this, and/or quantitative #s on, e.g., finetuning speed, in the paper.\n\nMore minor:\n- the authors use a number of terms in their main claims that are a bit vague and deserve proper definition, such as \"output/response space\", \"training response distribution\", \"behavioral guide\" and \"inherent attributes\" of the response space. I got the gist of what they were saying, but I don't think that should be assumed, and it would be good to be more consistent w/ terminology and more clearly define it.\n- the paper could benefit if the authors added some analysis on models with open training data (such as Olmo or Pythia)-- this would allow them to perform a leakage analysis and understand what types of/how much instruction data is present in the pretraining corpus.\n- Comparing Figs 3 and 6 it appears that both URIAL and URIAL-R are competitive with RT/IT, which raises the question of **why are we instruction-tuning in the first place?** I would appreciate some clarification of this in the rebuttal.", "questions": "Notes and minor other points:\n- How are you testing the models' core capabilities in Table 1-- prompt format, instructions in prompt, greedy decoding, few-shot examples?\n- typos line 044: \"omits\"; line 054 missing \"we\"; line 082: \"guides\"; line 092: \"focuses\"; line 405: missing \"the\"; line 406: extra \"the\"\n- it would be good to link to the plots for the other models in the Appendix in the caption of each Figure, for reader's easy reference and cross-model comparison.\n- related work: would be good to cite https://aclanthology.org/2020.emnlp-main.105/ for cross-task generalization as this was the earliest (?) work\n- The authors should mention this concurrent work in a camera-ready: https://arxiv.org/abs/2409.14254\n- I didn't find Figs 3/6 to be particularly useful and I am not sure what the 1-5 metric represented is. At a minimum it would be good to make the graphs represent the same metrics so that Figs 3 and 6 can be directly compared. \n- Using the word \"significant\" to describe results (lines 312, 403) implies you've done statistical testing", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper demonstrates that only fine-tuning base models on the responses in instruction datasets (response tuning, or RT) often results in comparable performance to fine-tuning on the responses *conditioned on the instructions* (traditional instruction tuning, or IT) without losing any core model capabilities. This supports the authors' hypothesis that models have already learned instruction following ability during pretraining, and instruction tuning thus mainly facilitates the model placing higher probabilities on tokens associated with helpful responses at inference time, or what the authors call \"aligning the training response distribution\". The authors measure response quality using 4 metrics: Likert-style individual response acceptability ratings + pairwise response preference selection using both human annotators and GPT4-as-a-judge. They find results to be consistent across 3 models from different families and on 3 instruction-tuning datasets, and also find that a larger Gemma model closes the gap between RT and IT more than a smaller one.\n\nThe authors further investigate whether using an LLM to refine the response quality along dimensions such as clarity, structure, and tone before training improves model response quality (as judged by GPT4), and find a positive result.\n\nFinally, they extend their experimental setting to refusal, and show that only training on responses from refusal/safety datasets (and not the unsafe queries the responses are associated with) also yields comparable performance, without causing the model to over-refuse. This is particularly interesting, because it supports the hypothesis that models learn to identify unsafe queries during pretraining, but need tuning to learn the \"language\" of refusal.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Originality: the paper takes a creative approach to an underexplored research question.\n- Quality: The paper includes quite extensive experiments to answer 3 main RQs and is not lacking on content. The authors include multiple datasets and models in their analysis. I appreciate that they use both human and model evaluation that is generally well-controlled for biases (e.g., not using the same model/family to evaluate another model); given that both evaluation methods have known shortcomings, their results seem more robust due to the extensive effort given to evaluation.\n- Clarity: The paper is generally well-written. Given there are a large amount of experimental results, the authors make good use of the Appendix for additional details while also making sure the main paper is readable and self-contained. I appreciate the delineation of the paper into 3 main sections corresponding to the 3 RQs of interest. Generally comprehensive lit review.\n- Significance: The paper provides a useful datapoint an interesting insight to ongoing discussion about what models learn during instruction tuning. While the 2nd RQ is a bit orthogonal to the paper's narrative, the 1st and 3rd RQs are interesting and impactful. I would like to see it presented at the conference.", "weaknesses": "Main:\n- My main qualm with the paper is the authors use the word \"comparable\" liberally when describing the performance of RT and IT tuning. However, in most cases, it is consistently the case that IT outperforms RT, often by what looks like rather substantial margins (e.g., Figures 2 and 4). I agree with the authors that it's really interesting that RT can make up the majority of the gap between the untuned and IT models, but I think the language can be toned down a bit and more discussion given to what performance aspects cannot be solely learned from RT alone. Alternatively, the authors could provide a quantitative definition of what they consider comparable performance.\nSimilarly, some of the differences in core capabilities (Table 1) seem meaningful (such as +7 on HellaSwag for Gemma 2 and -7 on GSM8k for Llama 3.1), but these are mostly chalked up to being typical or expected noise. It would be valuable to include a line for the untuned/base model in this table as a comparison point.\n- The URIAL paper should certainly be discussed in the related work section as it asks a similar RQ and has similar conclusions to this paper (though different methods).\n- The authors make a key assumption about the data they instruction- and response-tune on: that it hasn't been included in the pretraining data of the models in a meaningful way. This assumption seems to be supported by the poor performance of the untuned models in, e.g., Figure 2, but it should be stated more concretely. \n- Practical impact: from an analysis/model understanding perspective, the paper is strong. The authors do not explain, however, what impacts only training on responses could have from a practical angle. Presumably, it is not cheaper to collect the response data, since human or synthetic data generation pipelines would still require the instructions or refusal query to be present. It could be faster to train the model when one does not have to run inference on the instruction tokens, though the loss computation is the same so this may be minor. I would appreciate some discussion about this, and/or quantitative #s on, e.g., finetuning speed, in the paper.\n\nMore minor:\n- the authors use a number of terms in their main claims that are a bit vague and deserve proper definition, such as \"output/response space\", \"training response distribution\", \"behavioral guide\" and \"inherent attributes\" of the response space. I got the gist of what they were saying, but I don't think that should be assumed, and it would be good to be more consistent w/ terminology and more clearly define it.\n- the paper could benefit if the authors added some analysis on models with open training data (such as Olmo or Pythia)-- this would allow them to perform a leakage analysis and understand what types of/how much instruction data is present in the pretraining corpus.\n- Comparing Figs 3 and 6 it appears that both URIAL and URIAL-R are competitive with RT/IT, which raises the question of **why are we instruction-tuning in the first place?** I would appreciate some clarification of this in the rebuttal.", "questions": "Notes and minor other points:\n- How are you testing the models' core capabilities in Table 1-- prompt format, instructions in prompt, greedy decoding, few-shot examples?\n- typos line 044: \"omits\"; line 054 missing \"we\"; line 082: \"guides\"; line 092: \"focuses\"; line 405: missing \"the\"; line 406: extra \"the\"\n- it would be good to link to the plots for the other models in the Appendix in the caption of each Figure, for reader's easy reference and cross-model comparison.\n- related work: would be good to cite https://aclanthology.org/2020.emnlp-main.105/ for cross-task generalization as this was the earliest (?) work\n- The authors should mention this concurrent work in a camera-ready: https://arxiv.org/abs/2409.14254\n- I didn't find Figs 3/6 to be particularly useful and I am not sure what the 1-5 metric represented is. At a minimum it would be good to make the graphs represent the same metrics so that Figs 3 and 6 can be directly compared. \n- Using the word \"significant\" to describe results (lines 312, 403) implies you've done statistical testing", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730679652938}, {"id": "uQykfxPAWV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9542/Reviewer_JXF8"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The article suggests that the model's instruction-following capability can be acquired by fine-tuning using with the response data, and this hypothesis is validated through experiments. Furthermore, the article demonstrates enhanced preference alignment and safety by controlling the response distribution.", "review_text": "The article suggests that the model's instruction-following capability can be acquired by fine-tuning using with the response data, and this hypothesis is validated through experiments. Furthermore, the article demonstrates enhanced preference alignment and safety by controlling the response distribution.", "strengths": "1. The article is well-written and highly readable.\n2. It presents an interesting hypothesis and proposes a simple yet effective method.\n3. Experiments were conducted across various benchmarks.", "weaknesses": "1. The pre-training data for some models in the experiments not publicly available, and models like Gemma-2 already possess some instruction-following capability. This suggests that the pre-training data may inherently include instruction-response pairs, potentially skewing the experimental results and affecting accuracy.\n2. Experimental results show that the performance of Response Tuning still falls short of Instruction Tuning, particularly in human and GPT-4 evaluations. Although this is an interesting finding, deliberately constructing a response set without instructions for training seems neither effective nor efficient. Therefore, the significance of this method for advancing research remains questionable.", "questions": "1. Can you conduct similar experiments using models with publicly available pre-training data, such as OLMo, TinyLlama, and Pythia, to validate the conclusion of your article?\n2. In Table 1, why does Instruction Tuning perform better on several benchmarks but have a lower overall performance compared to Response Tuning? How is the overall performance in Table 1 calculated? For instance, the arithmetic averages for Response Tuning on these benchmarks are 54.05 and 61.08, so why are they reported as 58.42 and 65.62?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The article suggests that the model's instruction-following capability can be acquired by fine-tuning using with the response data, and this hypothesis is validated through experiments. Furthermore, the article demonstrates enhanced preference alignment and safety by controlling the response distribution.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The article is well-written and highly readable.\n2. It presents an interesting hypothesis and proposes a simple yet effective method.\n3. Experiments were conducted across various benchmarks.", "weaknesses": "1. The pre-training data for some models in the experiments not publicly available, and models like Gemma-2 already possess some instruction-following capability. This suggests that the pre-training data may inherently include instruction-response pairs, potentially skewing the experimental results and affecting accuracy.\n2. Experimental results show that the performance of Response Tuning still falls short of Instruction Tuning, particularly in human and GPT-4 evaluations. Although this is an interesting finding, deliberately constructing a response set without instructions for training seems neither effective nor efficient. Therefore, the significance of this method for advancing research remains questionable.", "questions": "1. Can you conduct similar experiments using models with publicly available pre-training data, such as OLMo, TinyLlama, and Pythia, to validate the conclusion of your article?\n2. In Table 1, why does Instruction Tuning perform better on several benchmarks but have a lower overall performance compared to Response Tuning? How is the overall performance in Table 1 calculated? For instance, the arithmetic averages for Response Tuning on these benchmarks are 54.05 and 61.08, so why are they reported as 58.42 and 65.62?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730639023078}, {"id": "4nwUZPTYL4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9542/Reviewer_krvS"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "Based on the hypothesis that language models learn to follow instructions during instruction tuning mainly by learning the output space, the paper presents experiments where models are trained only on responses (\"Response Tuning\" or RT) instead of conditioning them on instructions (the usual \"Instruction Tuning\" or IT).\n\nRT on various Llama and Gemma models with data from Alpaca, Dolly, and LIMA performs comparable to IT on instruction following (AlpacaEval) and several core capability evaluations (MMLU, OpenbookQA, Hellaswag, ARC, GSM8K, and PIQA).\n\nTo further demonstrate the importance of the quality of the responses, the responses from the datasets above are refined using a larger LM, and training on this refined dataset improves AlpacaEval scores in both IT and RT setups.\n\nWhen trained on LM responses corresponding to safety related refusals, though IT performs better than RT initially, with increasing amounts of data, performance of RT models eventually matches IT.\n\nIn an ICL (instead of finetuning) setup as well, it is shown that LMs can be aligned with human preferences using response demonstrations alone, with response-only ICL yielding comparable performance to paired-demonstration ICL in one evaluation setting (https://allenai.github.io/re-align/just_eval.html).", "review_text": "Based on the hypothesis that language models learn to follow instructions during instruction tuning mainly by learning the output space, the paper presents experiments where models are trained only on responses (\"Response Tuning\" or RT) instead of conditioning them on instructions (the usual \"Instruction Tuning\" or IT).\n\nRT on various Llama and Gemma models with data from Alpaca, Dolly, and LIMA performs comparable to IT on instruction following (AlpacaEval) and several core capability evaluations (MMLU, OpenbookQA, Hellaswag, ARC, GSM8K, and PIQA).\n\nTo further demonstrate the importance of the quality of the responses, the responses from the datasets above are refined using a larger LM, and training on this refined dataset improves AlpacaEval scores in both IT and RT setups.\n\nWhen trained on LM responses corresponding to safety related refusals, though IT performs better than RT initially, with increasing amounts of data, performance of RT models eventually matches IT.\n\nIn an ICL (instead of finetuning) setup as well, it is shown that LMs can be aligned with human preferences using response demonstrations alone, with response-only ICL yielding comparable performance to paired-demonstration ICL in one evaluation setting (https://allenai.github.io/re-align/just_eval.html).", "strengths": "The main finding in this paper is interesting. It provides an insight into how LMs learn to follow instructions. Future work may be able to leverage this insight to improve instruction tuning procedures.", "weaknesses": "## Practical utility\nThe paper does not demonstrate the practical utility of the finding. It would be helpful in the following cases:\n\n1. Obtaining instruction-response pairs is significantly harder than obtaining responses alone.\n2. Training on responses alone is more effective than training on instruction-response pairs.\n\n1 is not true in any of the experimental settings considered in the paper, and 2 is not true in any of the results, at least not significantly.\n\nSuggestions for addressing this concern:\n\n- There may be tasks where 1 is true. Code generation (from natural language instructions) is a good example. Does RT on code, which is much easier to obtain than paired NL-code examples do as well as IT? More generally, tasks with structured outputs that are easy to synthetically generate are suitable here.\n- Tasks where the model performance is sensitive to the superficial features in the instructions would be helpful to demonstrate 2. Does RT make the model more robust to variations in instructions than IT in such cases?\n\n## Role of the response refinement experiments\nThese experiments generally show the importance of the quality of the responses, but it is unclear how this is related to the main hypothesis about conditioning on instructions being unnecessary. Refinement generally helps in both IT and RT and RT settings. It would be more helpful to directly compare IT and RT models with and without response refinement. Does refinement increase the winrate of RT models against IT models?", "questions": "- What is the performance of the base models (Llama 3.1 8B and Gemma 2 9B) on the tasks shown in Table 1?\n- (More of a suggestion) It would be easier to interpret the results in Table 3 if all the models (IT and RT with and without response refinement) are evaluated against a fixed reference model for AlpacaEval.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Based on the hypothesis that language models learn to follow instructions during instruction tuning mainly by learning the output space, the paper presents experiments where models are trained only on responses (\"Response Tuning\" or RT) instead of conditioning them on instructions (the usual \"Instruction Tuning\" or IT).\n\nRT on various Llama and Gemma models with data from Alpaca, Dolly, and LIMA performs comparable to IT on instruction following (AlpacaEval) and several core capability evaluations (MMLU, OpenbookQA, Hellaswag, ARC, GSM8K, and PIQA).\n\nTo further demonstrate the importance of the quality of the responses, the responses from the datasets above are refined using a larger LM, and training on this refined dataset improves AlpacaEval scores in both IT and RT setups.\n\nWhen trained on LM responses corresponding to safety related refusals, though IT performs better than RT initially, with increasing amounts of data, performance of RT models eventually matches IT.\n\nIn an ICL (instead of finetuning) setup as well, it is shown that LMs can be aligned with human preferences using response demonstrations alone, with response-only ICL yielding comparable performance to paired-demonstration ICL in one evaluation setting (https://allenai.github.io/re-align/just_eval.html).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The main finding in this paper is interesting. It provides an insight into how LMs learn to follow instructions. Future work may be able to leverage this insight to improve instruction tuning procedures.", "weaknesses": "## Practical utility\nThe paper does not demonstrate the practical utility of the finding. It would be helpful in the following cases:\n\n1. Obtaining instruction-response pairs is significantly harder than obtaining responses alone.\n2. Training on responses alone is more effective than training on instruction-response pairs.\n\n1 is not true in any of the experimental settings considered in the paper, and 2 is not true in any of the results, at least not significantly.\n\nSuggestions for addressing this concern:\n\n- There may be tasks where 1 is true. Code generation (from natural language instructions) is a good example. Does RT on code, which is much easier to obtain than paired NL-code examples do as well as IT? More generally, tasks with structured outputs that are easy to synthetically generate are suitable here.\n- Tasks where the model performance is sensitive to the superficial features in the instructions would be helpful to demonstrate 2. Does RT make the model more robust to variations in instructions than IT in such cases?\n\n## Role of the response refinement experiments\nThese experiments generally show the importance of the quality of the responses, but it is unclear how this is related to the main hypothesis about conditioning on instructions being unnecessary. Refinement generally helps in both IT and RT and RT settings. It would be more helpful to directly compare IT and RT models with and without response refinement. Does refinement increase the winrate of RT models against IT models?", "questions": "- What is the performance of the base models (Llama 3.1 8B and Gemma 2 9B) on the tasks shown in Table 1?\n- (More of a suggestion) It would be easier to interpret the results in Table 3 if all the models (IT and RT with and without response refinement) are evaluated against a fixed reference model for AlpacaEval.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730329863592}], "openreview_url": "https://openreview.net/forum?id=rnRBGMNYa2", "arxiv_id": "2410.02465", "paper_pdf": "papers/rnRBGMNYa2.pdf", "paper_pdf_sha256": "2c546f7a0663f82b345270f848d742a2fb25ab5b7f735db5088ec4c85feedfcf", "paper_pdf_bytes": 1123010, "paper_pdf_source": "openreview", "code_url": "https://github.com/seokhyunan/response-tuning", "code_repository": "seokhyunan/response-tuning", "code_commit": "8db38cbbf14740ff2ddf8cf16cd033f09d2af449", "code_archive": "repos/rnRBGMNYa2.zip", "code_archive_sha256": "d08bea568563f111beb5bd88dd6bb85cfb8296d89daff07ad4f7f25f57993c06", "code_archive_bytes": 47126, "code_file_count": 26, "code_extensions": {".py": 15, ".sh": 11}, "github_disk_usage_kb": 35, "github_languages": {"Python": 84369, "Shell": 9165}, "github_archived": false, "github_pushed_at": "2025-09-12T06:49:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/response-tuning-aligning-large-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EE75tyB5Ay", "year": 2024, "status": "rejected", "title": "On the Generalization of Training-based ChatGPT Detection Methods", "authors": ["Han Xu", "Jie Ren", "Pengfei He", "Shenglai Zeng", "Yingqian Cui", "Amy Liu", "Hui Liu", "Jiliang Tang"], "authorids": ["~Han_Xu1", "~Jie_Ren6", "~Pengfei_He2", "~Shenglai_Zeng2", "~Yingqian_Cui1", "amyziyuliu@gmail.com", "~Hui_Liu8", "~Jiliang_Tang1"], "authors_source": "OpenReview API", "abstract": "ChatGPT is one of the most popular language models which achieve amazing performance on various natural language tasks. Consequently, there is also an urgent need to detect the texts generated ChatGPT from human written. One of the extensively studied methods trains classification models to distinguish both. However, existing studies also demonstrate that the trained models may suffer from distribution shifts (during test), i.e., they are ineffective to predict the generated texts from unseen  language tasks or topics. In this work, we aim to have a comprehensive investigation on these methods' generalization behaviors under distribution shift caused by a wide range of factors, including prompts, text lengths, topics, and language tasks. To achieve this goal, we first collect a new dataset with human and ChatGPT texts, and then we conduct extensive studies on the collected dataset. Our studies unveil insightful findings which provide guidance for developing future methodologies or data collection strategies for ChatGPT detection.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "SZhaUPcWsw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6093/Reviewer_FWZZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper investigates the challenges in distinguishing between human-written and ChatGPT-generated texts. The paper makes significant contributions to the field by offering a detailed analysis of ChatGPT detection methods under various distribution shifts. But I think most of findings seem obvious and it is more technical than scientific.", "review_text": "This paper investigates the challenges in distinguishing between human-written and ChatGPT-generated texts. The paper makes significant contributions to the field by offering a detailed analysis of ChatGPT detection methods under various distribution shifts. But I think most of findings seem obvious and it is more technical than scientific.", "strengths": "- **Comprehensive Investigation**: The paper conducts a thorough analysis of the generalization behaviors of existing methods under distribution shifts caused by various factors like prompts, text lengths, topics, and language tasks.\n\n- **New Dataset**: The authors contribute to the field by collecting a new dataset containing both human and ChatGPT-generated texts, facilitating in-depth studies on the detection methods.\n\n- **Insightful Findings**: The research uncovers insightful findings, providing valuable guidance for the development of future methodologies and data collection strategies for ChatGPT detection.", "weaknesses": "- The authors used three prompts in Figure 1; but in general, users might use various prompts. This is not aligned with real user usage,\n- The experiments are limited to CHATGPT, we do not know whether these conclusions still hold in GPT4 or other open-source LLMs.\n- Most findinds seem obvious.", "questions": "- Can we have a setting to mix various prompts, for example in Figure 1. This could be aligned with the real user usage\n- How are these findings valid to other LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the challenges in distinguishing between human-written and ChatGPT-generated texts. The paper makes significant contributions to the field by offering a detailed analysis of ChatGPT detection methods under various distribution shifts. But I think most of findings seem obvious and it is more technical than scientific.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "- **Comprehensive Investigation**: The paper conducts a thorough analysis of the generalization behaviors of existing methods under distribution shifts caused by various factors like prompts, text lengths, topics, and language tasks.\n\n- **New Dataset**: The authors contribute to the field by collecting a new dataset containing both human and ChatGPT-generated texts, facilitating in-depth studies on the detection methods.\n\n- **Insightful Findings**: The research uncovers insightful findings, providing valuable guidance for the development of future methodologies and data collection strategies for ChatGPT detection.", "weaknesses": "- The authors used three prompts in Figure 1; but in general, users might use various prompts. This is not aligned with real user usage,\n- The experiments are limited to CHATGPT, we do not know whether these conclusions still hold in GPT4 or other open-source LLMs.\n- Most findinds seem obvious.", "questions": "- Can we have a setting to mix various prompts, for example in Figure 1. This could be aligned with the real user usage\n- How are these findings valid to other LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699090661517}, {"id": "KUMpz1d3yx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6093/Reviewer_uBbK"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a solution for detecting texts generated by ChatGPT. \n\nSpecifically, the authors focus on understanding the generalization behaviors of training-based detection methods under distribution shifts caused by various factors including prompts, text lengths, topics, and language tasks. \n\nThey also collect a new data set and present findings to guide future methodologies and data collection strategies for ChatGPT detection.", "review_text": "The paper proposes a solution for detecting texts generated by ChatGPT. \n\nSpecifically, the authors focus on understanding the generalization behaviors of training-based detection methods under distribution shifts caused by various factors including prompts, text lengths, topics, and language tasks. \n\nThey also collect a new data set and present findings to guide future methodologies and data collection strategies for ChatGPT detection.", "strengths": "* The authors present a novel data set.\n* The analysis is detailed and comprehensive.\n* They provide insights on the data collection and domain adaption strategy.", "weaknesses": "* ChatGPT Direction seems to be not well motivated. It needs a why, not just a  what and a how.\n* This work exclusively discusses the train-based methods, which are smaller in scope.", "questions": "Could you elaborate on ChatGPT Direction? How is it motivated?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a solution for detecting texts generated by ChatGPT. \n\nSpecifically, the authors focus on understanding the generalization behaviors of training-based detection methods under distribution shifts caused by various factors including prompts, text lengths, topics, and language tasks. \n\nThey also collect a new data set and present findings to guide future methodologies and data collection strategies for ChatGPT detection.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "* The authors present a novel data set.\n* The analysis is detailed and comprehensive.\n* They provide insights on the data collection and domain adaption strategy.", "weaknesses": "* ChatGPT Direction seems to be not well motivated. It needs a why, not just a  what and a how.\n* This work exclusively discusses the train-based methods, which are smaller in scope.", "questions": "Could you elaborate on ChatGPT Direction? How is it motivated?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698808817316}, {"id": "mzihAVfCN9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6093/Reviewer_vcVF"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents an empirical evaluation on existing training-based \"ChatGPT detection\" methods. It has created a dataset which includes generated content prompted from ChatGPT and human written content. Analysis includes in-distribution evaluation, as well as OOD evaluation involving length-shift, topic- and domain-shift. Experiments show limited generalizability of training-based detection methods, and tend to overfit simplistic features. Meanwhile, it also suggests that these methods help with extracting transferable features that help generalize across domains.", "review_text": "This paper presents an empirical evaluation on existing training-based \"ChatGPT detection\" methods. It has created a dataset which includes generated content prompted from ChatGPT and human written content. Analysis includes in-distribution evaluation, as well as OOD evaluation involving length-shift, topic- and domain-shift. Experiments show limited generalizability of training-based detection methods, and tend to overfit simplistic features. Meanwhile, it also suggests that these methods help with extracting transferable features that help generalize across domains.", "strengths": "This work studies an important and timely problem for detecting LLM-generated content. Experiments have conducted for in-distributed settings as well as OOD settings involving content length shift and topic/domain shift.\nThe feature attribution analysis is an interesting and novel angle of study in the context of LLM-generated content detection.", "weaknesses": "The study seeks for detecting content generated by ChatGPT, which is just an interface where the backend model keeps evolving. Hence, it is hard to say if the experimental results and analysis are reproducible and sustainable. In my opinion, this type of study should be conducted for a static LLM.  \nLength shift and domain/topic shift represent limited types of distribution shift that is easily detectable. The authors could have considered more implicit shift where content is paraphrased with syntax-controlled paraphrasing or style transfer, like those used in recent approaches for data pollution attack / defense.\n\nTypos:\n\n3.1:\ndon't -> do not", "questions": "It is common for training-based methods to have more limited generalizability since they somewhat overfit a specific data distribution. From another perspective, would any unsupervised OOD detection methods [1,2] apply to detecting LLM-generated content?\n\n[1] Zhou, et al. Contrastive Out-of-Distribution Detection for Pretrained Transformers. EMNLP 2021\n[2] Xu, et al. Contrastive Novelty-Augmented Learning: Anticipating Outliers with Large Language Models. ACL 2023", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an empirical evaluation on existing training-based \"ChatGPT detection\" methods. It has created a dataset which includes generated content prompted from ChatGPT and human written content. Analysis includes in-distribution evaluation, as well as OOD evaluation involving length-shift, topic- and domain-shift. Experiments show limited generalizability of training-based detection methods, and tend to overfit simplistic features. Meanwhile, it also suggests that these methods help with extracting transferable features that help generalize across domains.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "This work studies an important and timely problem for detecting LLM-generated content. Experiments have conducted for in-distributed settings as well as OOD settings involving content length shift and topic/domain shift.\nThe feature attribution analysis is an interesting and novel angle of study in the context of LLM-generated content detection.", "weaknesses": "The study seeks for detecting content generated by ChatGPT, which is just an interface where the backend model keeps evolving. Hence, it is hard to say if the experimental results and analysis are reproducible and sustainable. In my opinion, this type of study should be conducted for a static LLM.  \nLength shift and domain/topic shift represent limited types of distribution shift that is easily detectable. The authors could have considered more implicit shift where content is paraphrased with syntax-controlled paraphrasing or style transfer, like those used in recent approaches for data pollution attack / defense.\n\nTypos:\n\n3.1:\ndon't -> do not", "questions": "It is common for training-based methods to have more limited generalizability since they somewhat overfit a specific data distribution. From another perspective, would any unsupervised OOD detection methods [1,2] apply to detecting LLM-generated content?\n\n[1] Zhou, et al. Contrastive Out-of-Distribution Detection for Pretrained Transformers. EMNLP 2021\n[2] Xu, et al. Contrastive Novelty-Augmented Learning: Anticipating Outliers with Large Language Models. ACL 2023", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697777351378}], "openreview_url": "https://openreview.net/forum?id=EE75tyB5Ay", "arxiv_id": "2310.01307", "paper_pdf": "papers/EE75tyB5Ay.pdf", "paper_pdf_sha256": "3abe01c531e2b31c13d7c77ee27d3d4154e73af831b6966d10527fb55d4b9c9a", "paper_pdf_bytes": 3363618, "paper_pdf_source": "openreview", "code_url": "https://github.com/hannxu123/hcvar", "code_repository": "hannxu123/hcvar", "code_commit": "c5c6aed6c013610bf91606e0cdef0d052b4da8e4", "code_archive": "repos/EE75tyB5Ay.zip", "code_archive_sha256": "8f77bd09813a4f4a5c4351265c8b8ec45064c9297467f83aacc145b6adb4cddc", "code_archive_bytes": 7866, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 33, "github_languages": {"Python": 17523}, "github_archived": false, "github_pushed_at": "2023-10-03T16:26:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-generalization-of-training-based"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vKXd1m74DkN", "year": 2023, "status": "rejected", "title": "Artificial Replay: A Meta-Algorithm for Harnessing Historical Data in Bandits", "authors": ["Sid Banerjee", "Sean R. Sinclair", "Milind Tambe", "Lily Xu", "Christina Yu"], "authorids": ["~Sid_Banerjee1", "~Sean_R._Sinclair1", "~Milind_Tambe1", "~Lily_Xu1", "~Christina_Yu1"], "authors_source": "OpenReview API", "abstract": "While standard bandit algorithms sometimes incur high regret, their performance can be greatly improved by \"warm starting\" with historical data. Unfortunately, how best to incorporate historical data is unclear: naively initializing reward estimates using all historical samples can suffer from spurious data and imbalanced data coverage, leading to computational and storage issues - particularly in continuous action spaces. We address these two challenges by proposing Artificial Replay, a meta-algorithm for incorporating historical data into any arbitrary base bandit algorithm. Artificial Replayuses only a subset of the historical data as needed to reduce computation and storage. We show that for a broad class of base algorithms that satisfy independence of irrelevant data (IIData), a novel property that we introduce, our method achieves equal regret as a full warm-start approach while potentially using only a fraction of historical data. We complement these theoretical results with a case study of $K$-armed and continuous combinatorial bandit algorithms, including on a green security domain using real poaching data, to show the practical benefits of Artificial Replayin achieving optimal regret alongside low computational and storage costs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "efIYgjC1--", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5123/Reviewer_Ai7i"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the stochastic bandits problem, where historical data is available. The goal is to learn a policy that exploits the available historical data to reduce regret (the difference between the maximum achievable reward and the policy's reward).\n\n\nThe authors propose a meta-algorithm named ARTIFICIAL REPLAY that uses a bandit algorithm (satisfying IIData property) as a base algorithm. The proposed algorithm only uses a subset of the historical data (hence reducing the computational cost), but it has the same regret bound as the algorithm that uses all historical data upfront (without suffering from issues like spurious data). They also validate the proposed algorithm's performance on synthetic and real-world datasets.", "review_text": "This paper significantly overlaps with my current work, and I am very knowledgeable about most of the topics covered by the paper.", "strengths": "**Strengths of paper:**\n1. The problem of exploiting historical data is interesting, and as discussed in the paper, many real-world applications already have historical data.\n\n2. The proposed meta-algorithm only uses a subset of historical data and has the better regret as a bandit algorithm (satisfying IIData property) that uses all historical data for initializing reward estimates. \n\n3. The proposed two-base bandit algorithms, MONUCB (for K-armed bandits) and CMAB-CRA (combinatorial MAB for Continuous Resource Allocation), enjoy IIData property. The authors empirically validated the different performance aspects of the proposed algorithm on synthetic and real datasets. \n\n**Weakness of paper:**\n1. The assumption of IIData, i.e., having additional information about other actions won't change the bandit algorithm decision in a given round, is too strong and counterintuitive. Best of my knowledge, no existing bandit algorithm that has IIData property.\n\n2. The statement, \"We present ARTIFICIAL REPLAY, a meta-algorithm that modifies any base bandit algorithm to efficiently harness historical data.\" in the Conclusion and a similar statement in the Abstract is misleading. Because ARTIFICIAL REPLAY only works when the base algorithm has IIData property.\n\n3. The storage and computational problems are not that big in many real-world applications, and the authors have not motivated enough by giving suitable examples of where one should care about these issues. Even for the proposed algorithm, all history needed to be stored, so the storage issue is still there. However, one can have a computationally efficient implementation for accessing historical data. This problem will get more challenging for the continuous action space as finding action within the $\\epsilon$ range may be computationally intensive.\n\n4. The proposed algorithm is horizon dependent and needs $T$ as input. Making such an assumption may not be practical in many real-world applications as $T$ (how long the algorithm will be used in practice) may not be known.\n\n5. In the experiments, authors have not used Thomson sampling variants (which enjoy better empirical performance than their UCB counterparts) against their proposed algorithms. It would be interesting to see how MONUCB (with ARTIFICIAL REPLAY) performs against the Thomson sampling variant that does not use any history (even for a K-armed bandit).\n\n6. Some related work for Combinatorial Multi-Armed Bandit for Continuous Resource Allocation is missing, e.g.,  \ni. Tor Lattimore, Koby Crammer, and Csaba Szepesvári. Optimal resource allocation with semi-bandit feedback. UAI, 2014.\nii. Tor Lattimore, Koby Crammer, and Csaba Szepesvári. Linear multi-resource allocation with semi-bandit feedback. NIPS, 2015.\niii. Yuval Dagan and Crammer Koby. A better resource allocation algorithm with semi-bandit feedback. ALT, 2018.\niv. Other recent work.\n\n\n**Question and other comments.** \n\nPlease address the above weakness. I have one more question:\n1. Page 3, paragraph after Eq. (3): Why do chosen resources for different arms needs to be $\\epsilon$-away from each other?\n\n\nI have a few minor comments:\n1. $H_1$ needs to be initialized as an empty set.\n2. $H^{\\text{hist}}$ needs to be updated after getting a sample from it; otherwise, it is possible to sample the same action-reward tuple when the existing action is chosen again.\n\nI am open to changing my score based on the authors' responses.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the stochastic bandits problem, where historical data is available. The goal is to learn a policy that exploits the available historical data to reduce regret (the difference between the maximum achievable reward and the policy's reward).\n\n\nThe authors propose a meta-algorithm named ARTIFICIAL REPLAY that uses a bandit algorithm (satisfying IIData property) as a base algorithm. The proposed algorithm only uses a subset of the historical data (hence reducing the computational cost), but it has the same regret bound as the algorithm that uses all historical data upfront (without suffering from issues like spurious data). They also validate the proposed algorithm's performance on synthetic and real-world datasets.", "strength_and_weaknesses": "**Strengths of paper:**\n1. The problem of exploiting historical data is interesting, and as discussed in the paper, many real-world applications already have historical data.\n\n2. The proposed meta-algorithm only uses a subset of historical data and has the better regret as a bandit algorithm (satisfying IIData property) that uses all historical data for initializing reward estimates. \n\n3. The proposed two-base bandit algorithms, MONUCB (for K-armed bandits) and CMAB-CRA (combinatorial MAB for Continuous Resource Allocation), enjoy IIData property. The authors empirically validated the different performance aspects of the proposed algorithm on synthetic and real datasets. \n\n**Weakness of paper:**\n1. The assumption of IIData, i.e., having additional information about other actions won't change the bandit algorithm decision in a given round, is too strong and counterintuitive. Best of my knowledge, no existing bandit algorithm that has IIData property.\n\n2. The statement, \"We present ARTIFICIAL REPLAY, a meta-algorithm that modifies any base bandit algorithm to efficiently harness historical data.\" in the Conclusion and a similar statement in the Abstract is misleading. Because ARTIFICIAL REPLAY only works when the base algorithm has IIData property.\n\n3. The storage and computational problems are not that big in many real-world applications, and the authors have not motivated enough by giving suitable examples of where one should care about these issues. Even for the proposed algorithm, all history needed to be stored, so the storage issue is still there. However, one can have a computationally efficient implementation for accessing historical data. This problem will get more challenging for the continuous action space as finding action within the $\\epsilon$ range may be computationally intensive.\n\n4. The proposed algorithm is horizon dependent and needs $T$ as input. Making such an assumption may not be practical in many real-world applications as $T$ (how long the algorithm will be used in practice) may not be known.\n\n5. In the experiments, authors have not used Thomson sampling variants (which enjoy better empirical performance than their UCB counterparts) against their proposed algorithms. It would be interesting to see how MONUCB (with ARTIFICIAL REPLAY) performs against the Thomson sampling variant that does not use any history (even for a K-armed bandit).\n\n6. Some related work for Combinatorial Multi-Armed Bandit for Continuous Resource Allocation is missing, e.g.,  \ni. Tor Lattimore, Koby Crammer, and Csaba Szepesvári. Optimal resource allocation with semi-bandit feedback. UAI, 2014.\nii. Tor Lattimore, Koby Crammer, and Csaba Szepesvári. Linear multi-resource allocation with semi-bandit feedback. NIPS, 2015.\niii. Yuval Dagan and Crammer Koby. A better resource allocation algorithm with semi-bandit feedback. ALT, 2018.\niv. Other recent work.\n\n\n**Question and other comments.** \n\nPlease address the above weakness. I have one more question:\n1. Page 3, paragraph after Eq. (3): Why do chosen resources for different arms needs to be $\\epsilon$-away from each other?\n\n\nI have a few minor comments:\n1. $H_1$ needs to be initialized as an empty set.\n2. $H^{\\text{hist}}$ needs to be updated after getting a sample from it; otherwise, it is possible to sample the same action-reward tuple when the existing action is chosen again.\n\nI am open to changing my score based on the authors' responses.", "clarity,_quality,_novelty_and_reproducibility": "**Clarity:**\nThe paper is well organized, but the presentation has minor details that could be improved, as discussed earlier in **Strength And Weaknesses**.\n\n**Quality:** \nOverall, the paper appears to be technically sound. The proofs appear correct, but I have not carefully checked the details. The experimental evaluation is adequate and supports the main claims. \n\n**Novelty:** \nThis paper contributes some new ideas, but they only represent incremental advances.\n\n**Reproducibility:** \nThe key resources (e.g., proofs and code) are available, and sufficient details are given to reproduce the main results.", "summary_of_the_review": "This paper significantly overlaps with my current work, and I am very knowledgeable about most of the topics covered by the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "Since this work is a theoretical paper, I do not find any ethical concerns.", "recommendation": "3: reject, not good enough"}, "tcdate": 1666876576985}, {"id": "7eGtfOVGX2a", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5123/Reviewer_NKDb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers the stochastic bandit learning problem when the algorithm has access to information regarding historical actions and their corresponding rewards. By leveraging already available information regarding the rewards from historical actions, a bandit algorithm can obtain significantly improved regret over vanilla bandit algorithms. A simple, but naive, approach to use historical information is to incorporate all available data to “initialize” a bandit algorithm - while this approach effectively utilizes the available historical information, it suffers from being computationally expensive. Instead the paper proposes a new meta-algorithm: for any online bandit algorithm, when the bandit algorithm recommends action A_t at time t, instead of actually playing action A_t it checks if action A_t is available in the historical data and updates the internal state of the algorithm using the historical reward instead. The authors show that under certain assumptions on the base bandit algorithm, this strategy has regret equivalent to the naive approach above that uses the entire historical data upfront.", "review_text": "The paper is well-written and proposes a simple, intuitive (meta)-algorithm to harness historical information in stochastic bandit settings. The proposed approach makes efficient use of the available historical information and obtains significantly improved regret without using the full historical dataset at the beginning.", "strengths": "Strengths:\n+ Proposed algorithm is simple and intuitive.\n+ The IIData condition is novel, yet easy to verify for different algorithms.\n\nWeaknesses:\n- Could you add a discussion of gap independent regret? How much data is needed to give asymptotic improvements in the regret?\n- It is unclear what the exact computational /  storage costs are for using all the available historical data. For the K-armed bandit problem, the storage costs are simply a function of the number of arms and remain unchanged whether the algorithm builds UCB estimates for all the arms using all the historical data at time 0 or whether it uses the historical data incrementally. Even for the CMAB-CRA problem, the entire available historical data still needs to be stored (for potential future look-ups); so it’s not fully clear what the savings are.\n- I am interested in a discussion of the setting where the historical data is also obtained via actions of a no-regret bandit algorithm (which is likely to be the case when such an algorithm is actually deployed). In this setting, the collected historical data is unlikely to contain spurious data - and it will be interesting to see whether the empirical gains still hold.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers the stochastic bandit learning problem when the algorithm has access to information regarding historical actions and their corresponding rewards. By leveraging already available information regarding the rewards from historical actions, a bandit algorithm can obtain significantly improved regret over vanilla bandit algorithms. A simple, but naive, approach to use historical information is to incorporate all available data to “initialize” a bandit algorithm - while this approach effectively utilizes the available historical information, it suffers from being computationally expensive. Instead the paper proposes a new meta-algorithm: for any online bandit algorithm, when the bandit algorithm recommends action A_t at time t, instead of actually playing action A_t it checks if action A_t is available in the historical data and updates the internal state of the algorithm using the historical reward instead. The authors show that under certain assumptions on the base bandit algorithm, this strategy has regret equivalent to the naive approach above that uses the entire historical data upfront.", "strength_and_weaknesses": "Strengths:\n+ Proposed algorithm is simple and intuitive.\n+ The IIData condition is novel, yet easy to verify for different algorithms.\n\nWeaknesses:\n- Could you add a discussion of gap independent regret? How much data is needed to give asymptotic improvements in the regret?\n- It is unclear what the exact computational /  storage costs are for using all the available historical data. For the K-armed bandit problem, the storage costs are simply a function of the number of arms and remain unchanged whether the algorithm builds UCB estimates for all the arms using all the historical data at time 0 or whether it uses the historical data incrementally. Even for the CMAB-CRA problem, the entire available historical data still needs to be stored (for potential future look-ups); so it’s not fully clear what the savings are.\n- I am interested in a discussion of the setting where the historical data is also obtained via actions of a no-regret bandit algorithm (which is likely to be the case when such an algorithm is actually deployed). In this setting, the collected historical data is unlikely to contain spurious data - and it will be interesting to see whether the empirical gains still hold.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and is very clear and nicely structured. The key idea is highlighted appropriately and the main algorithm as well as the experiments are presented well. The algorithm itself is very natural and simple - but I actually consider that a positive. \n\nThe appendix includes enough details for reproducibility (especially once the code is released).", "summary_of_the_review": "The paper is well-written and proposes a simple, intuitive (meta)-algorithm to harness historical information in stochastic bandit settings. The proposed approach makes efficient use of the available historical information and obtains significantly improved regret without using the full historical dataset at the beginning.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666733932080}, {"id": "TNeGxjLEGVc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5123/Reviewer_oZhq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of bandit optimization when the learner has access to historic data. In this setting, the paper studies how best to use the historic data in order to reduce the computational and storage costs of the learner, while not hurting its regret. \n\nThe main contribution of the paper is to propose an artificial-replay based algorithm that makes efficient use of the historic data. This algorithm works as a wrapper around any bandit algorithm. For any action recommended by the bandit algorithm, the meta algorithm first checks if there is a data point corresponding to that action in the historic data. If yes, it sends the reward of that action to the learner and removes that point from history. If not, the algorithm simply queries the environment for the reward of the action. Under a condition called \"independence of irrelevant data\", the authors show that the artificial replay algorithm achieves the same regret as  the regret obtained by warm-starting the bandit algorithm with the entire historic data. \n\nThe authors instantiate their framework on two special bandit problems: stochastic multi-armed bandits (MAB), Combinatorial Multi-Armed Bandit for Continuous Resource Allocation (CMAB-CRA). For both these problems, the authors design modified versions of UCB algorithms that satisfy the IIData condition. Experimentally, the authors show that the proposed techniques are computationally more efficient than full warm-start based approaches, which achieving same regret. \n", "review_text": "While the problem being considered in the paper is interesting, the results could be significantly improved. For example, more interesting examples (than MAB) should be used to illustrate the benefits of the proposed algorithm. Moreover, the paper doesn't really talk about the regret optimality of the proposed algorithms. It'd be great if the authors address my concerns above.", "strengths": "Strengths: \n1) The problem is very relevant in the context of modern recommendation systems where we often have a lot of historic data. Incorporating the entire historic data in the bandit algorithm can be computationally expensive. So, a natural question that arises is whether we can trade-off a little bit of regret for computational efficiency. The paper takes a step towards answering this question. It shows that under IIData condition, the artificial-replay based algorithm has the same regret as the full-start algorithm while being computationally more efficient. \n2) The proposed meta-algorithm is simple and can be used with any bandit algorithm.\n\nWeaknesses: \n1) While the IIData condition looks interesting, it is a strong condition. It requires the bandit algorithm's recommendation at a particular time instant to be the same even if it is provided with more information about all the other arms that are not being recommended. None of the bandit algorithms I know of seem to satisfy this condition (e.g., UCB, Thompson Sampling, Phased elimination).  This is worrisome because we are now forced to construct good bandit algorithms that satisfy the IIData condition.  Does this make the decades of work on designing bandit algorithms irrelevant? Is there an easy way to modify any given bandit algorithm to satisfy the IIData condition? \n\n         In practice, there is no need for a stringent requirement that the artificial-replay algorithm and the full start algorithm have the same regret. It is okay to trade-off a little bit of regret for computational efficiency. Is it possible to weaken the IIData condition to take this into account?\n\n2) One important aspect that the paper hasn't touched upon at all is the regret optimality of the proposed algorithms. Is ARTIFICIAL REPLAY(MONUCB) regret optimal for both MAB and CMAB-CRA problems? Are there any regret lower bounds that can be provided in the presence of historic data? These lower bounds can help us understand how optimal the proposed algorithms are.\n\n         Atleast for MAB, it doesn't look like the proposed algorithm is optimal. When there is no historic data, it is well known that standard UCB is not optimal in both minimax and instance dependent sense [1, 2]. There are several other algorithms that have been proposed to fix this issue.  \n\n3) At a number of places in the paper, it is claimed that the proposed algorithm is better in computation and storage than the baseline algorithm. However, I don't see this clearly. Note that the full start algorithm only requires storing sufficient statistics (and not the entire data). For MAB, these sufficient statistics are the average reward and number of pulls of each arm. So, if we knew that we are using the full start algorithm ahead of time, then we only need to store these statistics, and the resulting algorithm only requires O(K) storage and O(K) additional compute. This is in fact much better than the artificial replay algorithm which requires O(H) storage and O(sqrt{T}) additional compute. Overall, I believe the MAB setting is not clearly showcasing the computational, storage benefits of the proposed algorithm. A more challenging setting with continuous action spaces (e.g., Linear UCB, Neural UCB) would have been more  interesting. While the authors do consider the CMAB-CRA problem, most of its details are relegated to the appendix. Moreover, the problem is studied by a niche community. I'd instead recommend the authors to consider a more fundamental problem (like LinearUCB) and showcase the benefits of the proposed algorithm.\n \n4) Minor comments: \n  (a) the greedy algorithm mentioned in section 4.2 is never defined in the paper.  Given this, it is hard to evaluate the optimality of the proposed algorithm. \n  (b) in the statement of theorem 4.2, the run time of artificial-replay algorithm should be the minimum of sqrt{T} of log{T}/square of sub-optimality gap.\n  (c) in experiments, it'd be good to report runtime and storage improvements achieved by the proposed algorithm. \n  (d) In figure 8 (bottom right), why does the FULL START algorithm have such bad performance? It looks counter intuitive.\n\n[1] Lattimore, Tor. \"Optimally confident UCB: Improved regret for finite-armed bandits.\" arXiv preprint arXiv:1507.07880 (2015).\n\n[2] Garivier, Aurélien, and Olivier Cappé. \"The KL-UCB algorithm for bounded stochastic bandits and beyond.\" In Proceedings of the 24th annual conference on learning theory, pp. 359-376. JMLR Workshop and Conference Proceedings, 2011.\n\nMore historic data hurting the regret is weird, counter-intuitive and requires more understanding. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers the problem of bandit optimization when the learner has access to historic data. In this setting, the paper studies how best to use the historic data in order to reduce the computational and storage costs of the learner, while not hurting its regret. \n\nThe main contribution of the paper is to propose an artificial-replay based algorithm that makes efficient use of the historic data. This algorithm works as a wrapper around any bandit algorithm. For any action recommended by the bandit algorithm, the meta algorithm first checks if there is a data point corresponding to that action in the historic data. If yes, it sends the reward of that action to the learner and removes that point from history. If not, the algorithm simply queries the environment for the reward of the action. Under a condition called \"independence of irrelevant data\", the authors show that the artificial replay algorithm achieves the same regret as  the regret obtained by warm-starting the bandit algorithm with the entire historic data. \n\nThe authors instantiate their framework on two special bandit problems: stochastic multi-armed bandits (MAB), Combinatorial Multi-Armed Bandit for Continuous Resource Allocation (CMAB-CRA). For both these problems, the authors design modified versions of UCB algorithms that satisfy the IIData condition. Experimentally, the authors show that the proposed techniques are computationally more efficient than full warm-start based approaches, which achieving same regret. \n", "strength_and_weaknesses": "Strengths: \n1) The problem is very relevant in the context of modern recommendation systems where we often have a lot of historic data. Incorporating the entire historic data in the bandit algorithm can be computationally expensive. So, a natural question that arises is whether we can trade-off a little bit of regret for computational efficiency. The paper takes a step towards answering this question. It shows that under IIData condition, the artificial-replay based algorithm has the same regret as the full-start algorithm while being computationally more efficient. \n2) The proposed meta-algorithm is simple and can be used with any bandit algorithm.\n\nWeaknesses: \n1) While the IIData condition looks interesting, it is a strong condition. It requires the bandit algorithm's recommendation at a particular time instant to be the same even if it is provided with more information about all the other arms that are not being recommended. None of the bandit algorithms I know of seem to satisfy this condition (e.g., UCB, Thompson Sampling, Phased elimination).  This is worrisome because we are now forced to construct good bandit algorithms that satisfy the IIData condition.  Does this make the decades of work on designing bandit algorithms irrelevant? Is there an easy way to modify any given bandit algorithm to satisfy the IIData condition? \n\n         In practice, there is no need for a stringent requirement that the artificial-replay algorithm and the full start algorithm have the same regret. It is okay to trade-off a little bit of regret for computational efficiency. Is it possible to weaken the IIData condition to take this into account?\n\n2) One important aspect that the paper hasn't touched upon at all is the regret optimality of the proposed algorithms. Is ARTIFICIAL REPLAY(MONUCB) regret optimal for both MAB and CMAB-CRA problems? Are there any regret lower bounds that can be provided in the presence of historic data? These lower bounds can help us understand how optimal the proposed algorithms are.\n\n         Atleast for MAB, it doesn't look like the proposed algorithm is optimal. When there is no historic data, it is well known that standard UCB is not optimal in both minimax and instance dependent sense [1, 2]. There are several other algorithms that have been proposed to fix this issue.  \n\n3) At a number of places in the paper, it is claimed that the proposed algorithm is better in computation and storage than the baseline algorithm. However, I don't see this clearly. Note that the full start algorithm only requires storing sufficient statistics (and not the entire data). For MAB, these sufficient statistics are the average reward and number of pulls of each arm. So, if we knew that we are using the full start algorithm ahead of time, then we only need to store these statistics, and the resulting algorithm only requires O(K) storage and O(K) additional compute. This is in fact much better than the artificial replay algorithm which requires O(H) storage and O(sqrt{T}) additional compute. Overall, I believe the MAB setting is not clearly showcasing the computational, storage benefits of the proposed algorithm. A more challenging setting with continuous action spaces (e.g., Linear UCB, Neural UCB) would have been more  interesting. While the authors do consider the CMAB-CRA problem, most of its details are relegated to the appendix. Moreover, the problem is studied by a niche community. I'd instead recommend the authors to consider a more fundamental problem (like LinearUCB) and showcase the benefits of the proposed algorithm.\n \n4) Minor comments: \n  (a) the greedy algorithm mentioned in section 4.2 is never defined in the paper.  Given this, it is hard to evaluate the optimality of the proposed algorithm. \n  (b) in the statement of theorem 4.2, the run time of artificial-replay algorithm should be the minimum of sqrt{T} of log{T}/square of sub-optimality gap.\n  (c) in experiments, it'd be good to report runtime and storage improvements achieved by the proposed algorithm. \n  (d) In figure 8 (bottom right), why does the FULL START algorithm have such bad performance? It looks counter intuitive.\n\n[1] Lattimore, Tor. \"Optimally confident UCB: Improved regret for finite-armed bandits.\" arXiv preprint arXiv:1507.07880 (2015).\n\n[2] Garivier, Aurélien, and Olivier Cappé. \"The KL-UCB algorithm for bounded stochastic bandits and beyond.\" In Proceedings of the 24th annual conference on learning theory, pp. 359-376. JMLR Workshop and Conference Proceedings, 2011.\n\nMore historic data hurting the regret is weird, counter-intuitive and requires more understanding. ", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the paper is mostly well written and easy to follow. ", "summary_of_the_review": "While the problem being considered in the paper is interesting, the results could be significantly improved. For example, more interesting examples (than MAB) should be used to illustrate the benefits of the proposed algorithm. Moreover, the paper doesn't really talk about the regret optimality of the proposed algorithms. It'd be great if the authors address my concerns above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666681520792}, {"id": "qMPsz-bB2c", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5123/Reviewer_ZUbC"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studied an important bandit problem: how to deal with historical data. This paper proposed a meta-algorithm to handle computation and storage issues. ", "review_text": "Good problem but the theory part needs to be strengthened. ", "strengths": "The problem this paper studied is important and interesting. The meta-algorithm is simple and intuitive. I feel before we talk about computational efficiency and storage, it is better to fully understand the statistical efficiency (regret) for this problem. I have some questions about the regret guarantee and use Theorem 4.3 as an example since this is the basic multi-armed bandits.\n \n1. I feel there lacks sufficient discussion on how the historical data can reduce the online regret. The paper just mentioned \"equally improved regret guarantee\". This should be discussed quantitively. How about minimax regret? The abstract mentioned spurious and imbalance data could appear in historical data. How they affect the regret? \n \n2. I feel different historical data could affect online regret in a very different way. For example, if historical data contains all the optimal actions, online algorithms could just do imitation learning. If historical data contains a lot very sub-optimal actions, we could use them to remove bad actions. This work seems to not cover those issues. And the question is what exactly a full start algorithm is? There are many different ways to use full historical data and this also depends on what is your base algorithm.\n \n3. In (2), the regret definition seems to be independent of historical data. How does this happens? Should it be conditional on historical data? Will a_j^H be random variables? \n \n4. The authors claimed they propose regret-optimal IIData policies. Where the optimality comes from? I didn't see any lower bound here and is this instance-dependent or minimax optimal?\n \n \n   \n \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studied an important bandit problem: how to deal with historical data. This paper proposed a meta-algorithm to handle computation and storage issues. ", "strength_and_weaknesses": "The problem this paper studied is important and interesting. The meta-algorithm is simple and intuitive. I feel before we talk about computational efficiency and storage, it is better to fully understand the statistical efficiency (regret) for this problem. I have some questions about the regret guarantee and use Theorem 4.3 as an example since this is the basic multi-armed bandits.\n \n1. I feel there lacks sufficient discussion on how the historical data can reduce the online regret. The paper just mentioned \"equally improved regret guarantee\". This should be discussed quantitively. How about minimax regret? The abstract mentioned spurious and imbalance data could appear in historical data. How they affect the regret? \n \n2. I feel different historical data could affect online regret in a very different way. For example, if historical data contains all the optimal actions, online algorithms could just do imitation learning. If historical data contains a lot very sub-optimal actions, we could use them to remove bad actions. This work seems to not cover those issues. And the question is what exactly a full start algorithm is? There are many different ways to use full historical data and this also depends on what is your base algorithm.\n \n3. In (2), the regret definition seems to be independent of historical data. How does this happens? Should it be conditional on historical data? Will a_j^H be random variables? \n \n4. The authors claimed they propose regret-optimal IIData policies. Where the optimality comes from? I didn't see any lower bound here and is this instance-dependent or minimax optimal?\n \n \n   \n \n", "clarity,_quality,_novelty_and_reproducibility": "Overall this paper is good but the theory part needs more clarification. ", "summary_of_the_review": "Good problem but the theory part needs to be strengthened. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666478478910}], "openreview_url": "https://openreview.net/forum?id=vKXd1m74DkN", "arxiv_id": "2210.00025", "paper_pdf": "papers/vKXd1m74DkN.pdf", "paper_pdf_sha256": "f867ff4aad5a9c20617f0cfbc636afe98cbc552aa9a13841e5faefc7df7dccc1", "paper_pdf_bytes": 706286, "paper_pdf_source": "openreview", "code_url": "https://github.com/lily-x/artificial-replay", "code_repository": "lily-x/artificial-replay", "code_commit": "7602ffcdc5a9766eefe806e4ad152a31a2d303b0", "code_archive": "repos/vKXd1m74DkN.zip", "code_archive_sha256": "b53aa425ab10e14e0758d820d94a46ba6d92e10f0403d25584e2711b9a6c24d4", "code_archive_bytes": 47228, "code_file_count": 26, "code_extensions": {".py": 26}, "github_disk_usage_kb": 40, "github_languages": {"Python": 122475}, "github_archived": false, "github_pushed_at": "2023-01-26T22:41:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/artificial-replay-a-meta-algorithm-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "a1m8Jba-N6l", "year": 2022, "status": "rejected", "title": "$k$-Mixup Regularization for Deep Learning via Optimal Transport", "authors": ["Kristjan Greenewald", "Anming Gu", "Mikhail Yurochkin", "Justin Solomon", "Edward Chien"], "authorids": ["~Kristjan_Greenewald1", "~Anming_Gu1", "~Mikhail_Yurochkin1", "~Justin_Solomon1", "~Edward_Chien1"], "authors_source": "OpenReview API", "abstract": "Mixup is a popular regularization technique for training deep neural networks that can improve generalization  and increase adversarial robustness. It perturbs input training data in the direction of other randomly-chosen instances in the training set. To better leverage the structure of the data, we extend mixup to $k$-mixup by perturbing $k$-batches of training points in the direction of other $k$-batches using displacement interpolation, i.e. interpolation under the Wasserstein metric. We demonstrate theoretically and in simulations that $k$-mixup preserves cluster and manifold structures, and we extend theory studying the efficacy of standard mixup to the $k$-mixup case. Our empirical results show that training with $k$-mixup further improves generalization and robustness across several network architectures and benchmark datasets of differing modalities.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "SMxT3JMYgEB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3774/Reviewer_P7xo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Goals: The paper presents a new manner of mixing up training samples to create new training distributions like mix up. It uses optimal transport between minibatches of data of size k. Then they perform mixup between transported data and uses these new data in the training of their neural networks. The intuition is that their method is more able to respect the manifold of data contrary to original mixup which mix samples uniformly at random.\n", "review_text": "Description: The paper explains well the method and it can be reproduced easily in my opinion. However, I find it hard to fully understand how they build the entire minibatches of data they use (see questions below). There are some details I am not sure to understand for now and I feel the paper lacks a bit of clarity. May be adding some concrete examples might help readers to get the picture of the creation of the full minibatch used in training or may be adding the pseudo code of their algorithm. Regarding theoretical results, the paper discusses their results properly, however I have some concerns about the validity of the results due to the fact that authors use in practice minibatch optimal transport instead of exact OT (see questions and remarks below). Thus, I am not convinced of the pertinence of some claims.\n\n\nEvaluation: The evaluation of the methods seems complete experimentally. The method has been used on classification problems on a different datasets of different dimensions. It has also been compared to other mixup variants (manifold mixup) and on adversarial attacks. The results show a small increase in the performance. I think the empirical comparison is complete. However, I have some concerns regarding the theory as stated above. Indeed using minibatch OT as you do, creates non optimal connections between samples including connections between different examples [3]. Thus a longer discussion on minibatch OT is required.\n\n\nSignificance: The idea is interesting but is not too novel as the idea of coupling close samples has already been explored in the original mixup paper through k-nearest neighbourh. The difference is the use of optimal transport to determine how to connect samples instead of doing it randomly as done in the original mixup.\n\nRelated Work and Discussion: The strength of the methods are discuted but the aspect of minibatch OT is lacking and as such, the discussion is not complete in my opinion. The limitations of the methods is not discussed enough in my opinion. Authors could have empirically experimented the percentage of connections between different clusters of data to show that their method respects the manifold of data. Regarding related work, there are missing discussions which I think are important with previous methods (k nearest neighbourh from original mixup for instance).\n\nClarity: A reader not familiar with the mixup regularization could understand previous work as well as the presented method. The paper is easy to read and correctly written. The objective is clearly stated and I have not seen typos in the main text. I think that, from the text, one could reproduce the proposed methods. However, I also think that a discussion on how they create the full minibatch of data used during training should be included in the paper. Maybe adding examples or pseudo code is a good way to improve the clarity on these details.\n\nQuestions and remarks:\n\n1. You do not really use optimal transport but minibatch optimal transport[1,2] and you should discuss the differences. As such, it is known from [3,4,5] that it creates non optimal connections between clusters of data due to the sampling of minibatches. Following your intuition to preserve the manifold structure, I wonder why all your best scores are not for the biggest k, as the biggest k would more preserve the structure. This is contradictory with the initial prediction. Finally, if k grows to infinity, the proposed method would just mixup samples between themselves, thus no doing any mixup at all... This is also a concerning point, a study focused on minibatch OT (what authors do in practice) might alleviate this problem.\n\n2. Even if the connections between far away data are rare, it is noted in [5] that they happen and can harm the neural networks on some applications. I wonder their impact on the training. (see point 8 for a related point)\n\n3. Furthermore, due to the possible imbalanced classes, the data structure you are looking for, might not be possible in practice even for big k and even with OT between the full distributions. How does your method work for highly imbalanced case ? May be using partial OT or unbalanced OT with your minibatch formulation might help.\n\n4. I have some concerns with Section 3.1 and Section 3.2. You have applied your theory to exact OT, or minibatch OT with only one batch couple, while you are doing minibatch OT which is an expectation of optimal transport terms over minibatches of data. The latter favorises the creations of non optimal connections and is different of the former. You consider the case where the number of data grows to infinity but in practice you only use really small k. As such the theory is not in concordance with the practice and I am not convinced. \n\n5. While the use of Optimal Transport is appeling and increases the scores, it is not the first time that doing miwup between similar data is used. Indeed, in the original mix up paper, authors tried to used a k-nearest neighbourh and did not see improvement over original mixup. A discussion between k-nearest neighbourh is thus lacking.\n\n6. What is the full batch size you used for your training ? Was it k ?\n\n7. Could you please share a pseudo algorithm of your method ? It could be in appendix and help to understand some details of your training.\n\n8. As the motivation of your method is to respect the manifold of data, an interesting experiment, in my opinion, would be to measure the average percentage of connections between data which belong to different clusters. \n\n[1] DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation, Damadoran et al.\n\n[2] Learning Generative Models with Sinkhorn Divergences, Genevay et al.\n\n[3] Learning with minibatch Wasserstein: asymptotic and gradient properties, Fatras et al.\n\n[4] Minibatch Optimal Transport distances; analysis and applications, Fatras et al.\n\n[5] Unbalanced minibatch Optimal Transport; applications to Domain Adaptation, Fatras et al.\n\n\n----------------------------------------------------------------------------------------------------------------------------------------------------------------\n############################################ UPDATE ############################################\n----------------------------------------------------------------------------------------------------------------------------------------------------------------\nI have read your answers. While I think some elements have greatly improved, I still think that there are some problems. I think the discussion on the minibatch transport plan should be better discussed at least in supplementary. I have also a concern of the manifold structure preservation in high dimension. Most reviewers also think that the use of theorems in practice is unclear. For all these reasons I keep my score unchanged.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Goals: The paper presents a new manner of mixing up training samples to create new training distributions like mix up. It uses optimal transport between minibatches of data of size k. Then they perform mixup between transported data and uses these new data in the training of their neural networks. The intuition is that their method is more able to respect the manifold of data contrary to original mixup which mix samples uniformly at random.\n", "main_review": "Description: The paper explains well the method and it can be reproduced easily in my opinion. However, I find it hard to fully understand how they build the entire minibatches of data they use (see questions below). There are some details I am not sure to understand for now and I feel the paper lacks a bit of clarity. May be adding some concrete examples might help readers to get the picture of the creation of the full minibatch used in training or may be adding the pseudo code of their algorithm. Regarding theoretical results, the paper discusses their results properly, however I have some concerns about the validity of the results due to the fact that authors use in practice minibatch optimal transport instead of exact OT (see questions and remarks below). Thus, I am not convinced of the pertinence of some claims.\n\n\nEvaluation: The evaluation of the methods seems complete experimentally. The method has been used on classification problems on a different datasets of different dimensions. It has also been compared to other mixup variants (manifold mixup) and on adversarial attacks. The results show a small increase in the performance. I think the empirical comparison is complete. However, I have some concerns regarding the theory as stated above. Indeed using minibatch OT as you do, creates non optimal connections between samples including connections between different examples [3]. Thus a longer discussion on minibatch OT is required.\n\n\nSignificance: The idea is interesting but is not too novel as the idea of coupling close samples has already been explored in the original mixup paper through k-nearest neighbourh. The difference is the use of optimal transport to determine how to connect samples instead of doing it randomly as done in the original mixup.\n\nRelated Work and Discussion: The strength of the methods are discuted but the aspect of minibatch OT is lacking and as such, the discussion is not complete in my opinion. The limitations of the methods is not discussed enough in my opinion. Authors could have empirically experimented the percentage of connections between different clusters of data to show that their method respects the manifold of data. Regarding related work, there are missing discussions which I think are important with previous methods (k nearest neighbourh from original mixup for instance).\n\nClarity: A reader not familiar with the mixup regularization could understand previous work as well as the presented method. The paper is easy to read and correctly written. The objective is clearly stated and I have not seen typos in the main text. I think that, from the text, one could reproduce the proposed methods. However, I also think that a discussion on how they create the full minibatch of data used during training should be included in the paper. Maybe adding examples or pseudo code is a good way to improve the clarity on these details.\n\nQuestions and remarks:\n\n1. You do not really use optimal transport but minibatch optimal transport[1,2] and you should discuss the differences. As such, it is known from [3,4,5] that it creates non optimal connections between clusters of data due to the sampling of minibatches. Following your intuition to preserve the manifold structure, I wonder why all your best scores are not for the biggest k, as the biggest k would more preserve the structure. This is contradictory with the initial prediction. Finally, if k grows to infinity, the proposed method would just mixup samples between themselves, thus no doing any mixup at all... This is also a concerning point, a study focused on minibatch OT (what authors do in practice) might alleviate this problem.\n\n2. Even if the connections between far away data are rare, it is noted in [5] that they happen and can harm the neural networks on some applications. I wonder their impact on the training. (see point 8 for a related point)\n\n3. Furthermore, due to the possible imbalanced classes, the data structure you are looking for, might not be possible in practice even for big k and even with OT between the full distributions. How does your method work for highly imbalanced case ? May be using partial OT or unbalanced OT with your minibatch formulation might help.\n\n4. I have some concerns with Section 3.1 and Section 3.2. You have applied your theory to exact OT, or minibatch OT with only one batch couple, while you are doing minibatch OT which is an expectation of optimal transport terms over minibatches of data. The latter favorises the creations of non optimal connections and is different of the former. You consider the case where the number of data grows to infinity but in practice you only use really small k. As such the theory is not in concordance with the practice and I am not convinced. \n\n5. While the use of Optimal Transport is appeling and increases the scores, it is not the first time that doing miwup between similar data is used. Indeed, in the original mix up paper, authors tried to used a k-nearest neighbourh and did not see improvement over original mixup. A discussion between k-nearest neighbourh is thus lacking.\n\n6. What is the full batch size you used for your training ? Was it k ?\n\n7. Could you please share a pseudo algorithm of your method ? It could be in appendix and help to understand some details of your training.\n\n8. As the motivation of your method is to respect the manifold of data, an interesting experiment, in my opinion, would be to measure the average percentage of connections between data which belong to different clusters. \n\n[1] DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation, Damadoran et al.\n\n[2] Learning Generative Models with Sinkhorn Divergences, Genevay et al.\n\n[3] Learning with minibatch Wasserstein: asymptotic and gradient properties, Fatras et al.\n\n[4] Minibatch Optimal Transport distances; analysis and applications, Fatras et al.\n\n[5] Unbalanced minibatch Optimal Transport; applications to Domain Adaptation, Fatras et al.\n\n\n----------------------------------------------------------------------------------------------------------------------------------------------------------------\n############################################ UPDATE ############################################\n----------------------------------------------------------------------------------------------------------------------------------------------------------------\nI have read your answers. While I think some elements have greatly improved, I still think that there are some problems. I think the discussion on the minibatch transport plan should be better discussed at least in supplementary. I have also a concern of the manifold structure preservation in high dimension. Most reviewers also think that the use of theorems in practice is unclear. For all these reasons I keep my score unchanged.", "summary_of_the_review": "Recommendation: Reject. While I agree the idea is appealing, it is not too novel and some missing discussions would lead to a huge change in the original paper in my opinion.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636108700939}, {"id": "j9fncwOstn7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3774/Reviewer_iZL9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes an optimal transport-based mixup algorithm, theoretically analyzes the algorithm, and empirically evaluates its performance.", "review_text": "Major comments:\n\n-- Proposition 1 and Theorem 1's implications are unclear. Does this mean that k-mixup with large k is closer to \"on-manifold mixup\" such as SMOTE?  For instance, see \"[Chawla et al. (2002)] that proposed to augment the rare class in an imbalanced dataset by interpolating the nearest neighbors and [DeVries & Taylor (2017)] that showed that interpolation and extrapolation the nearest neighbors of the same class in feature space can improve generalization\". (The description of these papers are excerpted from the original mixup paper -- the authors may want to refer to Sec. 4 of the original paper.)\n\n-- \"The localized nature of the matchings makes it more likely that the averaged labels will smoothly interpolate over the decision boundaries. A consequence is that k-mixup is robust to higher values of α, since it is no longer necessary to keep λ close to 0 or 1 to avoid erroneous labels.\" => For high-dimensional data, even with higher values of alpha, it is unclear if erroneous labels (aka manifold intrusion) will occur with high probability. This questions the practical gain of the proposed method on high-dimensional data. \n\n-- The proposed idea seems relevant to [\"GAN-mixup: Augmenting Across Data Manifolds for Improved Robustness\", Sohn et al.].  The authors may want to clarify the difference between the proposed approach and the one introduced in this prior work. \n\n-- Performance comparisons with baseline algorithms are missing.  \n--- Guo et al.'s AdaMixup\n-- [\"On Adversarial Mixup Resynthesis\", Beckham et al.]\n\n-- Some performance gains (compared to k = 1) on real datasets look too marginal; See Figure 10, Figure 11 \n\n-- By looking at Figure 5 and Figure 6, the performance seems to be highly correlated with the average squared distance of vicinal distribution from training set, regardless of the choice of k.  For instance, in Figure 6, (k=1, alpha=0.5) and (k=8, alpha=1) have almost the same distance as well as the test accuracy. \n\nThis makes it unclear whether or not the performance gain actually comes from the benefits of k-mixup. Instead, this might be an artifact of decreasing distance intervals as k increases. For instance, in Figure 5, even though the same values of alpha's are tested, k=1's average squared distance ranges from 0.068 to 1.732, while k=32's those ranges from 0.029 to 0.586. This allows the latter to try out more reasonable values of the squared distances. \n\nTo show that this is not the case, the authors may want to rerun the experiments while adaptively setting the range of alpha values for a different choice of k such that the same (or similar) alpha values can be tested. \n\n-- The confidence on test performance seems surprisingly too low to me. What are the random factors across different Monte Carlo runs?  Random initializations and random shuffling usually alone usually give a much higher confidence interval than the reported values such as 0.02 or 0.05. \n(See Table 1 in this for instance -- https://arxiv.org/abs/2109.08203)\n\n-- Adversarial robustness: robust accuracy against FGSM (or any simple gradient-based attack) can be highly misleading. Please use AutoAttack by Croce and Hein instead to see whether there exists an actual robustness gain.\n\n\nMinor comments:\n\n-- \"Averaging weights are typically drawn from a beta distribution β(α, α), with parameter α ≪ 1 such that the generated training set is vicinal\" -> Not true.  See Table 1 and Table 2 in the original mixup paper for the choice of large alpha values.  Also, the original paper says \"For example, in CIFAR-10 classification we can get very low training error on real data even when α → ∞ (i.e., training only on averages of pairs of real examples)\".", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes an optimal transport-based mixup algorithm, theoretically analyzes the algorithm, and empirically evaluates its performance.", "main_review": "Major comments:\n\n-- Proposition 1 and Theorem 1's implications are unclear. Does this mean that k-mixup with large k is closer to \"on-manifold mixup\" such as SMOTE?  For instance, see \"[Chawla et al. (2002)] that proposed to augment the rare class in an imbalanced dataset by interpolating the nearest neighbors and [DeVries & Taylor (2017)] that showed that interpolation and extrapolation the nearest neighbors of the same class in feature space can improve generalization\". (The description of these papers are excerpted from the original mixup paper -- the authors may want to refer to Sec. 4 of the original paper.)\n\n-- \"The localized nature of the matchings makes it more likely that the averaged labels will smoothly interpolate over the decision boundaries. A consequence is that k-mixup is robust to higher values of α, since it is no longer necessary to keep λ close to 0 or 1 to avoid erroneous labels.\" => For high-dimensional data, even with higher values of alpha, it is unclear if erroneous labels (aka manifold intrusion) will occur with high probability. This questions the practical gain of the proposed method on high-dimensional data. \n\n-- The proposed idea seems relevant to [\"GAN-mixup: Augmenting Across Data Manifolds for Improved Robustness\", Sohn et al.].  The authors may want to clarify the difference between the proposed approach and the one introduced in this prior work. \n\n-- Performance comparisons with baseline algorithms are missing.  \n--- Guo et al.'s AdaMixup\n-- [\"On Adversarial Mixup Resynthesis\", Beckham et al.]\n\n-- Some performance gains (compared to k = 1) on real datasets look too marginal; See Figure 10, Figure 11 \n\n-- By looking at Figure 5 and Figure 6, the performance seems to be highly correlated with the average squared distance of vicinal distribution from training set, regardless of the choice of k.  For instance, in Figure 6, (k=1, alpha=0.5) and (k=8, alpha=1) have almost the same distance as well as the test accuracy. \n\nThis makes it unclear whether or not the performance gain actually comes from the benefits of k-mixup. Instead, this might be an artifact of decreasing distance intervals as k increases. For instance, in Figure 5, even though the same values of alpha's are tested, k=1's average squared distance ranges from 0.068 to 1.732, while k=32's those ranges from 0.029 to 0.586. This allows the latter to try out more reasonable values of the squared distances. \n\nTo show that this is not the case, the authors may want to rerun the experiments while adaptively setting the range of alpha values for a different choice of k such that the same (or similar) alpha values can be tested. \n\n-- The confidence on test performance seems surprisingly too low to me. What are the random factors across different Monte Carlo runs?  Random initializations and random shuffling usually alone usually give a much higher confidence interval than the reported values such as 0.02 or 0.05. \n(See Table 1 in this for instance -- https://arxiv.org/abs/2109.08203)\n\n-- Adversarial robustness: robust accuracy against FGSM (or any simple gradient-based attack) can be highly misleading. Please use AutoAttack by Croce and Hein instead to see whether there exists an actual robustness gain.\n\n\nMinor comments:\n\n-- \"Averaging weights are typically drawn from a beta distribution β(α, α), with parameter α ≪ 1 such that the generated training set is vicinal\" -> Not true.  See Table 1 and Table 2 in the original mixup paper for the choice of large alpha values.  Also, the original paper says \"For example, in CIFAR-10 classification we can get very low training error on real data even when α → ∞ (i.e., training only on averages of pairs of real examples)\".", "summary_of_the_review": "The theoretical claims look solid, but their implications are unclear.  The experimental settings and results could be improved.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636093097926}, {"id": "9gt8HBLSiQM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3774/Reviewer_EGaJ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to select pairs of points to mix between two batches of K examples by using the hungarian algorithm to find an L2 optimal matching.  This substantially improves performance on non-linear low-dimensional classification tasks where mixup underfits and it also improves results on moderate-scale classification tasks (like CIFAR-100) and especially helps when the mixing rate alpha is large.  There is also theory analyzing how often mixing will interpolate between different clusters when using k-mixup.  \n\nDetailed Notes from reading the paper: \n\n  -Modification of mixup where batches of k points are perturbed in the direction of k other points using interpolation under a Wasserstein metric.  \n\n  -Proof with experiments and theory showing this k-mixup preserves cluster and manifold structure.  \n\n  -K-mixup improves results (or keeps same) and improves adversarial robustness.  \n\n  -Increasing K makes interpolated points more likely to be on data manifold.  \n\n  -Use Hungarian algorithm to find L2 optimal permutation.  \n\n", "review_text": "This paper shows some nice improvements on low-dimensional toy datasets from using optimal transport to select examples from two different batches to interpolate.  The improvements on larger datasets are more marginal, but are still significant when the mixing rate alpha is large.  This is decent work and it may have some impact, but the small improvements may make the impact fairly limited.  Additionally what it achieves does overlap some with Manifold Mixup, although the paper does explain that Manifold Mixup has other drawbacks.  \n\nOther comments: \n  -First paragraph of intro is really boilerplate that could be removed.  \n\n  -Figure 1 is a bit weaker of a result than could be possible - since the solution inside the spiral is still somewhat but only partially blurred after using k-mixup.  Nonetheless it is better than the baseline.  \n\n  -Figure 2 is pretty convincing that more points are being interpolated onto the same manifold when using a larger K.  \n\n  -How is the algorithm different from mixing with nearest neighbors from a limited pool of examples?  Perhaps it's the requirement of an optimal transport (so that the same point can't be picked twice as a neighbor)?  \n\n  -The insight in Theorem 1 is nice, especially that cross-cluster mixes will still be selected more as K grows, but as a decreasing fraction of K.  \n\n  -The classification results (Figure 8) are a bit discouraging, and somewhat contradict the introduction which claims that the k-mixup technique doesn't hurt results, when several results are 0.1-0.3 basis points below the baseline.  Still, where improvements occur they are often of larger magnitude than the deteriorations.  \n\n  -The improvement with large alpha is impressive.  \n\n  -FGSM is a very weak attack, so the improvements in Figure 12 are of questionable significance, although it is nice to see slightly better robustness.  \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to select pairs of points to mix between two batches of K examples by using the hungarian algorithm to find an L2 optimal matching.  This substantially improves performance on non-linear low-dimensional classification tasks where mixup underfits and it also improves results on moderate-scale classification tasks (like CIFAR-100) and especially helps when the mixing rate alpha is large.  There is also theory analyzing how often mixing will interpolate between different clusters when using k-mixup.  \n\nDetailed Notes from reading the paper: \n\n  -Modification of mixup where batches of k points are perturbed in the direction of k other points using interpolation under a Wasserstein metric.  \n\n  -Proof with experiments and theory showing this k-mixup preserves cluster and manifold structure.  \n\n  -K-mixup improves results (or keeps same) and improves adversarial robustness.  \n\n  -Increasing K makes interpolated points more likely to be on data manifold.  \n\n  -Use Hungarian algorithm to find L2 optimal permutation.  \n\n", "main_review": "This paper shows some nice improvements on low-dimensional toy datasets from using optimal transport to select examples from two different batches to interpolate.  The improvements on larger datasets are more marginal, but are still significant when the mixing rate alpha is large.  This is decent work and it may have some impact, but the small improvements may make the impact fairly limited.  Additionally what it achieves does overlap some with Manifold Mixup, although the paper does explain that Manifold Mixup has other drawbacks.  \n\nOther comments: \n  -First paragraph of intro is really boilerplate that could be removed.  \n\n  -Figure 1 is a bit weaker of a result than could be possible - since the solution inside the spiral is still somewhat but only partially blurred after using k-mixup.  Nonetheless it is better than the baseline.  \n\n  -Figure 2 is pretty convincing that more points are being interpolated onto the same manifold when using a larger K.  \n\n  -How is the algorithm different from mixing with nearest neighbors from a limited pool of examples?  Perhaps it's the requirement of an optimal transport (so that the same point can't be picked twice as a neighbor)?  \n\n  -The insight in Theorem 1 is nice, especially that cross-cluster mixes will still be selected more as K grows, but as a decreasing fraction of K.  \n\n  -The classification results (Figure 8) are a bit discouraging, and somewhat contradict the introduction which claims that the k-mixup technique doesn't hurt results, when several results are 0.1-0.3 basis points below the baseline.  Still, where improvements occur they are often of larger magnitude than the deteriorations.  \n\n  -The improvement with large alpha is impressive.  \n\n  -FGSM is a very weak attack, so the improvements in Figure 12 are of questionable significance, although it is nice to see slightly better robustness.  \n\n", "summary_of_the_review": "This paper achieves small improvements on large datasets and significant improvements on either low dimensional data or where the mixing rate alpha is very large.  The improvements are small but the idea is simple and logical, so I weakly recommend acceptance.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636072589878}, {"id": "gpYQ3sOw_Pd", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3774/Reviewer_K2Mi"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work proposes an improvement on the Mixup regularization for training deep neural networks. Instead of performing weighted averages of randomly chosen pairs of samples, an optimal transport map between two k-batches is computed. Then, \"new\" samples are constructed by interpolating between *coupled* pairs of samples (according to the optimal transport plan). This enables to better reflect the local structure of the dataset. A theoretical study supports this intuition, and extensive experiments are conducted.", "review_text": "Pros: \n- This work is well-motivated and well-written.\n- Theoretical study supports the claim of the paper (as k increases, the vicinal samples better reflect the local structure of the dataset).\n- Experimental protocol seems sound and varied.\n\nCons:\n- There are two hyper-parameters, k and alpha, whose choice is not completely clear to me (see below).\n\nQuestions and remarks: \n- It could be useful to the reader to elaborate on the advantage of the Hungarian algorithm over Sinkhorn w.r.t k.\n- Do you really need to compare the cost of the regularization to the cost of computing gradients? Isn't k-mixup regularization a pre-processing step with \"fixed cost\"? \n- Do you have an idea why increasing k does not always improve your results? This seems to be opposed to the intuition that higher k better reflects the structure of the dataset. More generally, it could be great to have a discussion on how to choose k and alpha for your method.\n- Could this technique be useful in the context of transfer learning (I am not asking for more experiments here)?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes an improvement on the Mixup regularization for training deep neural networks. Instead of performing weighted averages of randomly chosen pairs of samples, an optimal transport map between two k-batches is computed. Then, \"new\" samples are constructed by interpolating between *coupled* pairs of samples (according to the optimal transport plan). This enables to better reflect the local structure of the dataset. A theoretical study supports this intuition, and extensive experiments are conducted.", "main_review": "Pros: \n- This work is well-motivated and well-written.\n- Theoretical study supports the claim of the paper (as k increases, the vicinal samples better reflect the local structure of the dataset).\n- Experimental protocol seems sound and varied.\n\nCons:\n- There are two hyper-parameters, k and alpha, whose choice is not completely clear to me (see below).\n\nQuestions and remarks: \n- It could be useful to the reader to elaborate on the advantage of the Hungarian algorithm over Sinkhorn w.r.t k.\n- Do you really need to compare the cost of the regularization to the cost of computing gradients? Isn't k-mixup regularization a pre-processing step with \"fixed cost\"? \n- Do you have an idea why increasing k does not always improve your results? This seems to be opposed to the intuition that higher k better reflects the structure of the dataset. More generally, it could be great to have a discussion on how to choose k and alpha for your method.\n- Could this technique be useful in the context of transfer learning (I am not asking for more experiments here)?", "summary_of_the_review": "This work seems well-motivated, clearly explained with globally convincing experiments. My only concern is on how to choose the hyper-parameters k and alpha, which do not seem to always have the intended effect. Overall, I tend to recommend acceptance.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635849513207}, {"id": "12deyrN2gTG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3774/Reviewer_8N9h"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposed $k$-mixup, which is a generalization of mixup, to better regularize neural networks during training. Specifically, instead of mixing totally random pairs of training data, the authors draw $k$ pairs of data each time and mix them in a way that minimizes the Wasserstein distance between them. The authors theoretically proved that their method can better preserve the local structure of the data. They also did experiments for various networks and datasets and showed that $k$-mixup can improve the generalization performance and robustness of the models compared to the original mixup.", "review_text": "-Strengths:\n\n1. This paper is generally well-written and well-organized, making it easy to understand.\n\n2. The authors did many experiments in various settings. They used neural networks with different sizes (small MLP to DenseNet) on different datasets (both synthetic and real-world datasets) from different modalities (image and audio). Each experiment is also repeated multiple times.\n\n3. The authors provided very detailed settings, e.g., network structures and hyperparameters, for their experiments. This makes their experiments quite reproducible.\n\n-Concerns:\n\n1. The justification for the proposed k-mixup method might be unclear, and this is my major concern for this paper. I will explain my concern in detail in the following paragraphs:\n\n1.1 The authors claimed that $k$-mixup preserves the local structures of the training data, but this alone is not enough to explain why $k$-mixup could help. Imagine the extreme case where we take $k$ to infinity, and this will essentially be equivalent to having no mixup at all because every data point will mix up with itself. Therefore, there might be deeper reasons why $k$-mixup could work, and this could be related to how the labels are interpolated during mixup. Since $k$-mixup encourages closer points to mix, the mixup between points from different classes will have a sharper label transition compared to the normal mixup. The authors proved in Theorem 2 that $k$-mixup can help the model with smooth interpolation between the clusters, but this model seems somewhat specific and the authors did not talk much about this smoothing effect in this paper.\n\n1.2 Both smaller $\\alpha$ and larger $k$ can make the training set become more vicinal, and intuitively it could be unclear which has a better regularization effect. For instance, setting both $k$ and $\\alpha$ to be large and setting both to be small should both result in a vicinal dataset. In this case, the two regimes might have a similar local structure for the training data, and it seems confusing why this extra parameter $k$ could help because we can always choose a smaller $\\alpha$ to make the training data more local.\n\n1.3 For larger $k$, more training data will be mixed with points within the same class if the data have some cluster structure, as can be seen in Figure 2. Thus, the ratio of the same-class mixup will vary when $k$ changes. This may also influence the regularization effect of k-mixup and it could be possible that changing the ratio of same-class mixup can already improve the performance of mixup.\n\n2. The empirical benefit provided by $k$-mixup might not be significant enough, and this $k$-mixup method requires extra tuning for the hyperparameter $k$. Details are provided below:\n\n2.1 The performance gain by performing $k$-mixup is not consistent for different datasets and network structures. For instance, in the small $\\alpha$ regime ($\\alpha$=0.1), larger $k$ always does not improve the performance in most of the datasets while requiring more computation and parameter tuning. This could be because smaller $\\alpha$ better preserves the local structure of the data and larger $k$ is not needed.\n\n2.2 For the cases where $k$-mixup does improve the performance over the original mixup, the performance gain does not seem to be very large (usually less than 1%), and achieving this performance gain requires much work in tuning the hyperparameters $k$ and $\\alpha$. As mentioned in concern 1.2, the regularization effect provided by $k$-mixup is controlled by both $k$ and $\\alpha$, and based on the experimental results there seems not to be a consistent scaling law for the best $k$ and $\\alpha$. Specifically, from the best-performance model for original mixup($k$=1), achieving best performance sometimes require us to increase both $k$ and $\\alpha$, but sometimes will require increasing $k$ and decreasing (or keep) $\\alpha$ instead, depending on the task. This probably means tuning $k$-mixup requires a grid search over $k$ and $\\alpha$, which introduces extra computation cost.\n\n2.3 It would be better if the authors could compare $k$-mixup to other mixup variants. This paper only provides one experiment comparing $k$-mixup to manifold mixup on one task, so it might not be convincing enough that $k$-mixup could perform better than other variants of mixup.\n\n2.4 The authors claimed in their paper that the extra computation cost caused by $k$-mixup is small, and it could be better if they could provide numerical evidence for this, e.g., compare the wallclock time of training the same model using regular mixup to that of $k$-mixup where $k$=32.\n\n3. The assumptions and conclusions of the theoretical claims are somewhat unrealistic.\n\n3.1 The theoretical claims often require k to be \"large enough\", which actually needs $k$ to be larger than some exponential function of dimension $d$. This cannot be true in practice. For Proposition 1, we need $k$ to be larger than both $\\Omega(1/\\delta)^d$ and $(1/R_S)^d$. For Theorem 2, $k$ needs to ensure both $A_\\delta$ and $B_\\delta$ contain enough points, which could also require $k$ to be exponentially large. This is somewhat unrealistic because the inputs for the usual tasks are usually of high dimension. It would also be better if the authors could explicitly state the requirement of $k$ in their statements of theorems.\n\n3.2 The assumptions needed for the theoretical results might lack justifications. For section 3.2, the authors assumed that the input data to be $(m, \\Delta)$-clusterable and the distance between any pair of covering balls is at least $2\\Delta$. This assumption intuitively would result in a very large $m$ for real data and largely weaken the conclusions. Besides, it might be better if the authors could explicitly state this assumption outside the Lemma since Theorem 1 and 2 also need that assumption.\n\n3.3 Section 4 might seem a bit confusing. The expressions for the loss and regularizers are roughly the same as the previous paper (Zhang et al., 2020) except that the expectation in the regularization terms is taken over the \"locally-informed distribution\", but it seems unclear why this is better than the original distribution.\n\n-Minor Comments:\n\n1. Figure 1 seems a bit confusing and might need more interpretation. The authors claimed that it shows $k$-mixup can better keep the manifold support structure, but it seems that 4-mixup produces a more blurry function on Swiss Roll than 1-mixup, which seems confusing. Besides, it might be better if the authors could provide more detailed explanations about these datasets, e.g., how they are generated and labeled, to make Figure 1 more clear.\n\n2. Some of the intuitions from the paper might become weaker for higher-dimensional data. For example, in the regime when the number of data is not much larger than the dimension, the data points will become more separated in the sense that linear interpolation between any two points might not be close to any other points. This regime may be very different from the synthetic datasets visualized in Figures 1 and 2. Besides, since the theoretical results in this paper often require $k$ to be exponentially large in $d$, this could also be part of the reason why $k$-mixup doesn't work so well in high dimensions. The authors discussed this in section 6, and it might be better if the authors could comment more.\n\n3. For the Toy datasets, especially One Ring and Swiss Roll, normal mixup seems to work better for smaller $\\alpha$, so perhaps the authors could try to use even smaller $\\alpha$ and see whether that can give even better results.\n\n4. For the performance on Google Speech Commands, the authors claimed that the difference between the best $k$-mixup and the best normal mixup is 0.6508%. How is this number computed?\n\n-Typos: There is a duplicate sentence near the end of Section 2: \"The localized nature of the matchings makes it more likely that the averaged labels will smoothly interpolate over the decision boundaries.\" appears twice.\n\n---------------------Update--------------------------\n\nI have read all the other reviews, the authors' responses to all the reviewers, and the revisions the authors made in the updated draft, and I have decided to keep my score unchanged, i.e., I tend to recommend rejection. I would like to thank the authors for their very detailed response which addressed some of my concerns (e.g., ratio of same-class mixup, wallclock time comparison), but my major concerns (unclear justification, marginal performance gain) about this paper still remain. Detailed reasons why I keep my score are listed below:\n\n- The justification for $k$-mixup seems somewhat unclear, and this is a common concern for most of the reviewers (P7xo, iZL9, K2Mi, 8N9h). The authors claimed that this method can better preserve cluster and manifold structures and provide some smoothing effect between clusters. However, both claims might be too abstract. The authors provided theoretical explanations in Section 3, but the implications for these theorems seem unclear, e.g., as also mentioned by reviewer P7xo and K2Mi, they cannot explain why increasing k doesn't necessarily always improve the performance.\n\n- The performance improvements on real datasets seem marginal, and this is also a common concern for most reviewers (iZL9, EGaJ, 8N9h). Besides, getting this performance gain requires tuning an additional parameter $k$, and there is no simple indicator (e.g., one cannot use average squared distance to directly predict the performance gain) for this gain, so one really needs to train more models with different $k$'s and $\\alpha$'s and do cross-validation, which requires much more computation power.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed $k$-mixup, which is a generalization of mixup, to better regularize neural networks during training. Specifically, instead of mixing totally random pairs of training data, the authors draw $k$ pairs of data each time and mix them in a way that minimizes the Wasserstein distance between them. The authors theoretically proved that their method can better preserve the local structure of the data. They also did experiments for various networks and datasets and showed that $k$-mixup can improve the generalization performance and robustness of the models compared to the original mixup.", "main_review": "-Strengths:\n\n1. This paper is generally well-written and well-organized, making it easy to understand.\n\n2. The authors did many experiments in various settings. They used neural networks with different sizes (small MLP to DenseNet) on different datasets (both synthetic and real-world datasets) from different modalities (image and audio). Each experiment is also repeated multiple times.\n\n3. The authors provided very detailed settings, e.g., network structures and hyperparameters, for their experiments. This makes their experiments quite reproducible.\n\n-Concerns:\n\n1. The justification for the proposed k-mixup method might be unclear, and this is my major concern for this paper. I will explain my concern in detail in the following paragraphs:\n\n1.1 The authors claimed that $k$-mixup preserves the local structures of the training data, but this alone is not enough to explain why $k$-mixup could help. Imagine the extreme case where we take $k$ to infinity, and this will essentially be equivalent to having no mixup at all because every data point will mix up with itself. Therefore, there might be deeper reasons why $k$-mixup could work, and this could be related to how the labels are interpolated during mixup. Since $k$-mixup encourages closer points to mix, the mixup between points from different classes will have a sharper label transition compared to the normal mixup. The authors proved in Theorem 2 that $k$-mixup can help the model with smooth interpolation between the clusters, but this model seems somewhat specific and the authors did not talk much about this smoothing effect in this paper.\n\n1.2 Both smaller $\\alpha$ and larger $k$ can make the training set become more vicinal, and intuitively it could be unclear which has a better regularization effect. For instance, setting both $k$ and $\\alpha$ to be large and setting both to be small should both result in a vicinal dataset. In this case, the two regimes might have a similar local structure for the training data, and it seems confusing why this extra parameter $k$ could help because we can always choose a smaller $\\alpha$ to make the training data more local.\n\n1.3 For larger $k$, more training data will be mixed with points within the same class if the data have some cluster structure, as can be seen in Figure 2. Thus, the ratio of the same-class mixup will vary when $k$ changes. This may also influence the regularization effect of k-mixup and it could be possible that changing the ratio of same-class mixup can already improve the performance of mixup.\n\n2. The empirical benefit provided by $k$-mixup might not be significant enough, and this $k$-mixup method requires extra tuning for the hyperparameter $k$. Details are provided below:\n\n2.1 The performance gain by performing $k$-mixup is not consistent for different datasets and network structures. For instance, in the small $\\alpha$ regime ($\\alpha$=0.1), larger $k$ always does not improve the performance in most of the datasets while requiring more computation and parameter tuning. This could be because smaller $\\alpha$ better preserves the local structure of the data and larger $k$ is not needed.\n\n2.2 For the cases where $k$-mixup does improve the performance over the original mixup, the performance gain does not seem to be very large (usually less than 1%), and achieving this performance gain requires much work in tuning the hyperparameters $k$ and $\\alpha$. As mentioned in concern 1.2, the regularization effect provided by $k$-mixup is controlled by both $k$ and $\\alpha$, and based on the experimental results there seems not to be a consistent scaling law for the best $k$ and $\\alpha$. Specifically, from the best-performance model for original mixup($k$=1), achieving best performance sometimes require us to increase both $k$ and $\\alpha$, but sometimes will require increasing $k$ and decreasing (or keep) $\\alpha$ instead, depending on the task. This probably means tuning $k$-mixup requires a grid search over $k$ and $\\alpha$, which introduces extra computation cost.\n\n2.3 It would be better if the authors could compare $k$-mixup to other mixup variants. This paper only provides one experiment comparing $k$-mixup to manifold mixup on one task, so it might not be convincing enough that $k$-mixup could perform better than other variants of mixup.\n\n2.4 The authors claimed in their paper that the extra computation cost caused by $k$-mixup is small, and it could be better if they could provide numerical evidence for this, e.g., compare the wallclock time of training the same model using regular mixup to that of $k$-mixup where $k$=32.\n\n3. The assumptions and conclusions of the theoretical claims are somewhat unrealistic.\n\n3.1 The theoretical claims often require k to be \"large enough\", which actually needs $k$ to be larger than some exponential function of dimension $d$. This cannot be true in practice. For Proposition 1, we need $k$ to be larger than both $\\Omega(1/\\delta)^d$ and $(1/R_S)^d$. For Theorem 2, $k$ needs to ensure both $A_\\delta$ and $B_\\delta$ contain enough points, which could also require $k$ to be exponentially large. This is somewhat unrealistic because the inputs for the usual tasks are usually of high dimension. It would also be better if the authors could explicitly state the requirement of $k$ in their statements of theorems.\n\n3.2 The assumptions needed for the theoretical results might lack justifications. For section 3.2, the authors assumed that the input data to be $(m, \\Delta)$-clusterable and the distance between any pair of covering balls is at least $2\\Delta$. This assumption intuitively would result in a very large $m$ for real data and largely weaken the conclusions. Besides, it might be better if the authors could explicitly state this assumption outside the Lemma since Theorem 1 and 2 also need that assumption.\n\n3.3 Section 4 might seem a bit confusing. The expressions for the loss and regularizers are roughly the same as the previous paper (Zhang et al., 2020) except that the expectation in the regularization terms is taken over the \"locally-informed distribution\", but it seems unclear why this is better than the original distribution.\n\n-Minor Comments:\n\n1. Figure 1 seems a bit confusing and might need more interpretation. The authors claimed that it shows $k$-mixup can better keep the manifold support structure, but it seems that 4-mixup produces a more blurry function on Swiss Roll than 1-mixup, which seems confusing. Besides, it might be better if the authors could provide more detailed explanations about these datasets, e.g., how they are generated and labeled, to make Figure 1 more clear.\n\n2. Some of the intuitions from the paper might become weaker for higher-dimensional data. For example, in the regime when the number of data is not much larger than the dimension, the data points will become more separated in the sense that linear interpolation between any two points might not be close to any other points. This regime may be very different from the synthetic datasets visualized in Figures 1 and 2. Besides, since the theoretical results in this paper often require $k$ to be exponentially large in $d$, this could also be part of the reason why $k$-mixup doesn't work so well in high dimensions. The authors discussed this in section 6, and it might be better if the authors could comment more.\n\n3. For the Toy datasets, especially One Ring and Swiss Roll, normal mixup seems to work better for smaller $\\alpha$, so perhaps the authors could try to use even smaller $\\alpha$ and see whether that can give even better results.\n\n4. For the performance on Google Speech Commands, the authors claimed that the difference between the best $k$-mixup and the best normal mixup is 0.6508%. How is this number computed?\n\n-Typos: There is a duplicate sentence near the end of Section 2: \"The localized nature of the matchings makes it more likely that the averaged labels will smoothly interpolate over the decision boundaries.\" appears twice.\n\n---------------------Update--------------------------\n\nI have read all the other reviews, the authors' responses to all the reviewers, and the revisions the authors made in the updated draft, and I have decided to keep my score unchanged, i.e., I tend to recommend rejection. I would like to thank the authors for their very detailed response which addressed some of my concerns (e.g., ratio of same-class mixup, wallclock time comparison), but my major concerns (unclear justification, marginal performance gain) about this paper still remain. Detailed reasons why I keep my score are listed below:\n\n- The justification for $k$-mixup seems somewhat unclear, and this is a common concern for most of the reviewers (P7xo, iZL9, K2Mi, 8N9h). The authors claimed that this method can better preserve cluster and manifold structures and provide some smoothing effect between clusters. However, both claims might be too abstract. The authors provided theoretical explanations in Section 3, but the implications for these theorems seem unclear, e.g., as also mentioned by reviewer P7xo and K2Mi, they cannot explain why increasing k doesn't necessarily always improve the performance.\n\n- The performance improvements on real datasets seem marginal, and this is also a common concern for most reviewers (iZL9, EGaJ, 8N9h). Besides, getting this performance gain requires tuning an additional parameter $k$, and there is no simple indicator (e.g., one cannot use average squared distance to directly predict the performance gain) for this gain, so one really needs to train more models with different $k$'s and $\\alpha$'s and do cross-validation, which requires much more computation power.", "summary_of_the_review": "I tend to vote for rejecting this paper. Despite being clearly written, the justification for their proposed method is somewhat unclear, and the performance gain seems not significant and consistent enough. Therefore, I think more work needs to be done to provide enough justifications for $k$-mixup (perhaps from the theoretical side) and to further improve the empirical performances.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634995324640}], "openreview_url": "https://openreview.net/forum?id=a1m8Jba-N6l", "arxiv_id": "2106.02933", "paper_pdf": "papers/a1m8Jba-N6l.pdf", "paper_pdf_sha256": "f963cac5a1157591e5423c4dbc99fcde7c1cf76eba645a292c9f2493488ad84b", "paper_pdf_bytes": 7687462, "paper_pdf_source": "openreview", "code_url": "https://github.com/AnmingGu/kmixup-cifar10", "code_repository": "AnmingGu/kmixup-cifar10", "code_commit": "ed2ecfd47226baf375f90e12cad5b4a97899a149", "code_archive": "repos/a1m8Jba-N6l.zip", "code_archive_sha256": "bdc208c551f27d744334db5dcc959c74034c94814117959f9aad4ccc1a78e6cc", "code_archive_bytes": 21750, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 37, "github_languages": {"Python": 66782}, "github_archived": false, "github_pushed_at": "2023-08-28T20:12:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/k-mixup-regularization-for-deep-learning-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tqc8n6oHCtZ", "year": 2021, "status": "rejected", "title": "Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Search", "authors": ["Gyuwan Kim", "Kyunghyun Cho"], "authorids": ["~Gyuwan_Kim1", "~Kyunghyun_Cho1"], "authors_source": "OpenReview API", "abstract": "Although transformers have achieved impressive accuracies in various tasks in natural language processing, they often come with a prohibitive computational cost, that prevents their use in scenarios with limited computational resources for inference. This need for computational efficiency in inference has been addressed by for instance PoWER-BERT (Goyal et al., 2020) which gradually decreases the length of a sequence as it is passed through layers. These approaches however often assume that the target computational complexity is known in advance at the time of training. This implies that a separate model must be trained for each inference scenario with its distinct computational budget. In this paper, we extend PoWER-BERT to address this issue of inefficiency and redundancy. The proposed extension enables us to train a large-scale transformer, called Length-Adaptive Transformer, once and uses it for various inference scenarios without re-training it. To do so, we train a transformer with LengthDrop, a structural variant of dropout, which stochastically determines the length of a sequence at each layer. We then use a multi-objective evolutionary search to find a length configuration that maximizes the accuracy and minimizes the computational complexity under any given computational budget. Additionally, we significantly extend the applicability of PoWER-BERT beyond sequence-level classification into token-level classification such as span-based question-answering, by introducing the idea of Drop-and-Restore. With Drop-and-Restore, word-vectors are dropped temporarily in intermediate layers and restored at the last layer if necessary. We empirically verify the utility of the proposed approach by demonstrating the superior accuracy-efficiency trade-off under various setups, including SQuAD 1.1, MNLI-m, and SST-2. Upon publication, the code to reproduce our work will be open-sourced.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "hfSpA-BICvP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3047/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The work targets an interesting direction of improving the efficiency of Transformers by reducing the sequence length. The main contributions of the work are (1) proposing LengthDrop as the way to achieve length reduction; (2) utilizing techniques developed in NAS, namely one-shot NAS, to enable proper training and allow adaptive drop ratio search after training. All these ideas are very reasonable and interesting. Empirically, the authors show that the proposed method is able to match or even outperform BERT-base model with 1/3 - 1/2 FLOPs during inference (not training).\n\nThe first concern I have is that the source of the gain is unclear in two aspects:\n(1) The proposed model is finetuned for longer (5 + 5 epochs) compared to the 3 epochs of the original BERT. However, the baseline numbers are still based on 3 epochs. This could be one of the reasons why the proposed model even outperform the original BERT.\n(2) With the \"Inplace Distillation\" in play, the obtained model is effectively a distilled model. This adds another layer of complication to judge how much the gain/loss comes from distillation and the length reduction.\n\nSecondly, authors do not mention much about the training (finetuning + ES) cost compared to the standard BERT or Power-BERT. In many real-world cases, this cost is also non-trivial. This may be part of the reason why only 3 datasets are considered in this paper. \n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting direction, reasonable and interesting techniques, limited experiments and unclear source of gain", "review": "The work targets an interesting direction of improving the efficiency of Transformers by reducing the sequence length. The main contributions of the work are (1) proposing LengthDrop as the way to achieve length reduction; (2) utilizing techniques developed in NAS, namely one-shot NAS, to enable proper training and allow adaptive drop ratio search after training. All these ideas are very reasonable and interesting. Empirically, the authors show that the proposed method is able to match or even outperform BERT-base model with 1/3 - 1/2 FLOPs during inference (not training).\n\nThe first concern I have is that the source of the gain is unclear in two aspects:\n(1) The proposed model is finetuned for longer (5 + 5 epochs) compared to the 3 epochs of the original BERT. However, the baseline numbers are still based on 3 epochs. This could be one of the reasons why the proposed model even outperform the original BERT.\n(2) With the \"Inplace Distillation\" in play, the obtained model is effectively a distilled model. This adds another layer of complication to judge how much the gain/loss comes from distillation and the length reduction.\n\nSecondly, authors do not mention much about the training (finetuning + ES) cost compared to the standard BERT or Power-BERT. In many real-world cases, this cost is also non-trivial. This may be part of the reason why only 3 datasets are considered in this paper. \n ", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603921431207}, {"id": "yKDcZ1zqZx3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3047/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work introduces a method, called LengthDrop, to train a Length-Adaptive Transformer that supports adaptive model architecture based on different latency constraints. In order to make the model robust to variable input lengths, the method stochastically reduces the length of a sequence at each layer during training. Once the model is trained, the method uses an evolutionary search to find subnetworks that maximize model accuracy under a latency budget. \n \nPros:\n- Accelerating the inference speed of Transformer networks is an important problem.\n- The idea of training a length-adaptive Transformer once and using it in different scenarios with different latency constraints is interesting. \n \nCons:\n- The discussion with several state-of-the-art work is lacking.\n- The experimental setup is vague, and the evaluation results are inadequate. \n \nThe paper looks from an interesting angle to build adaptive Transformers for inference -- reducing the input sequence at each Transformer layer. However, there are a few concerns.\n \nFirst, the paper proposes to use a series of techniques to make LengthDrop work but lacks the ablation studies to show how those techniques help to make Transformer length adaptive. For example, Section 3.1 states that LengthDrop requires LayerDrop[1], which also supports adaptive Transformer by stochastically dropping layers during training for adaptive inference. However, there are no ablation studies or comparison results on LengthDrop vs. LayerDrop in terms of the accuracy-vs-latency trade-off. This raises the question of whether LengthDrop is necessary to obtain the given accuracy-vs-latency or perhaps simpler alternatives such as LayerDrop would be sufficient.\n \nIn addition to LayerDrop, it appears that the paper also incorporates several other fixes, such as the sandwidth rule and inplace distillation, which are borrowed from prior work.  However, how these fixes contribute to LengthDrop is not clearly explained, and there are no studies nor experimental results to explain how each technique contributes to the final accuracy-vs-latency results. \n\nSecond, the comparison with related work is weak. In particular, LengthDrop is built on top of PoWER-BERT, yet the evaluation does not compare with PoWER-BERT.  Furthermore, the paper compares with DistillBERT, but there are multiple knowledge distillation based work that show better performance than DistillBERT, such as TinyBERT[2]. In terms of adaptive architecture, the evolutionary search of length configurations is similar to the NAS process in the Hardware-aware Transformer[2], which seems to be very related as it also uses evolutionary search to find a specialized sub-network of Transformer models with a latency constraint. The paper briefly mentioned [2], but it is not clear the advantage of this work as compared with [2]. \n \nThird, the paper lacks enough information on the evaluation setups, raising several questions on the reported speedups. For example, it is unclear what's the batch size used in the evaluation. Figure 3(a) shows that reducing FLOPs on GPU does not lead to a reduction of latency for batch size 1, which is the common setting for online inference scenarios as queries come in one-by-one. It is unclear whether input length reduction may actually bring significant latency reduction when the batch size is small (e.g., 1), as the large matrix multiplications have been highly optimized on modern CPU and GPU through efficient kernels (e.g., cuDNN). Even for results on CPU and GPU with batch size >= 16, it is less clear whether the linear correlation between FLOPs and latency is a fact of failing to use highly optimized BLAS libraries, because the paper does not report the details on the hardware, the inference frameworks, and libraries it uses for the experimental results. \n\nIn addition to the batch size and lack of hardware/platform/library information, the experimental setup for training a Length-Adaptive Transformer is also not very clear. For example, it is unclear what's the maximum sequence length is used in training. Whether mixed sequence length is used in training BERT?  What is the sequence length(s) to obtain the results in Table 1? Without associating the actual length reduction ratio, it is hard to evaluate the reported FLOPs reduction.\n \n[1] Fan et. al. \"Reducing Transformer Depth on Demand with Structured Dropout\", https://arxiv.org/abs/1909.11556\n\n[2] Jiao et. al. \"TinyBERT: Distilling BERT for Natural Language Understanding\", https://arxiv.org/abs/1909.10351", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "This work introduces a method, called LengthDrop, to train a Length-Adaptive Transformer that supports adaptive model architecture based on different latency constraints. In order to make the model robust to variable input lengths, the method stochastically reduces the length of a sequence at each layer during training. Once the model is trained, the method uses an evolutionary search to find subnetworks that maximize model accuracy under a latency budget. \n \nPros:\n- Accelerating the inference speed of Transformer networks is an important problem.\n- The idea of training a length-adaptive Transformer once and using it in different scenarios with different latency constraints is interesting. \n \nCons:\n- The discussion with several state-of-the-art work is lacking.\n- The experimental setup is vague, and the evaluation results are inadequate. \n \nThe paper looks from an interesting angle to build adaptive Transformers for inference -- reducing the input sequence at each Transformer layer. However, there are a few concerns.\n \nFirst, the paper proposes to use a series of techniques to make LengthDrop work but lacks the ablation studies to show how those techniques help to make Transformer length adaptive. For example, Section 3.1 states that LengthDrop requires LayerDrop[1], which also supports adaptive Transformer by stochastically dropping layers during training for adaptive inference. However, there are no ablation studies or comparison results on LengthDrop vs. LayerDrop in terms of the accuracy-vs-latency trade-off. This raises the question of whether LengthDrop is necessary to obtain the given accuracy-vs-latency or perhaps simpler alternatives such as LayerDrop would be sufficient.\n \nIn addition to LayerDrop, it appears that the paper also incorporates several other fixes, such as the sandwidth rule and inplace distillation, which are borrowed from prior work.  However, how these fixes contribute to LengthDrop is not clearly explained, and there are no studies nor experimental results to explain how each technique contributes to the final accuracy-vs-latency results. \n\nSecond, the comparison with related work is weak. In particular, LengthDrop is built on top of PoWER-BERT, yet the evaluation does not compare with PoWER-BERT.  Furthermore, the paper compares with DistillBERT, but there are multiple knowledge distillation based work that show better performance than DistillBERT, such as TinyBERT[2]. In terms of adaptive architecture, the evolutionary search of length configurations is similar to the NAS process in the Hardware-aware Transformer[2], which seems to be very related as it also uses evolutionary search to find a specialized sub-network of Transformer models with a latency constraint. The paper briefly mentioned [2], but it is not clear the advantage of this work as compared with [2]. \n \nThird, the paper lacks enough information on the evaluation setups, raising several questions on the reported speedups. For example, it is unclear what's the batch size used in the evaluation. Figure 3(a) shows that reducing FLOPs on GPU does not lead to a reduction of latency for batch size 1, which is the common setting for online inference scenarios as queries come in one-by-one. It is unclear whether input length reduction may actually bring significant latency reduction when the batch size is small (e.g., 1), as the large matrix multiplications have been highly optimized on modern CPU and GPU through efficient kernels (e.g., cuDNN). Even for results on CPU and GPU with batch size >= 16, it is less clear whether the linear correlation between FLOPs and latency is a fact of failing to use highly optimized BLAS libraries, because the paper does not report the details on the hardware, the inference frameworks, and libraries it uses for the experimental results. \n\nIn addition to the batch size and lack of hardware/platform/library information, the experimental setup for training a Length-Adaptive Transformer is also not very clear. For example, it is unclear what's the maximum sequence length is used in training. Whether mixed sequence length is used in training BERT?  What is the sequence length(s) to obtain the results in Table 1? Without associating the actual length reduction ratio, it is hard to evaluate the reported FLOPs reduction.\n \n[1] Fan et. al. \"Reducing Transformer Depth on Demand with Structured Dropout\", https://arxiv.org/abs/1909.11556\n\n[2] Jiao et. al. \"TinyBERT: Distilling BERT for Natural Language Understanding\", https://arxiv.org/abs/1909.10351", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603914544172}, {"id": "c8HNRTz3Fa", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3047/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n\nThe paper proposed the Length-Adaptive Transformer. The model can be trained once and directly applied to different inference scenarios. To achieve this goal, the author proposed the LengthDrop method, which randomly samples the length at each layer. In addition, the author used the sandwich rule to train the model. At each step, the sandwich rule will train the largest model, the smallest model, and another bunch of randomly sampled models. In the inference phase, the paper proposed to search for the best length configuration that balances the accuracy and latency tradeoff via evolutionary search. Moreover, to generalize the model to token annotation tasks, the author proposed the Drop-and-Restore process, in which the tokens that have been dropped are used again in the final layer. Experiments show that Length-Adaptive Transformer is able to outperfom the baseline models when evaluated at the same latency level.\n\n\n##########################################################################\n\nReasons for score: \n \n\nThe paper is well-written and the length-adaptive idea is reasonable. However, I think it still requires more experiments. Also, the training process is not very clear to the reader. It is not so clear how different techniques impact the final performance and the author has not reported the training time. Due to these two reasons, I choose to vote for weak rejection.\n \n\n##########################################################################\n\nPros: \n \n\n1. The idea of length-adaptive transformer is novel. Perform on-demand truncation of the hidden lengths has not been well explored.\n \n\n##########################################################################\n\nCons: \n \n\n1. The paper needs to conduct experiments on more text classification dataset. Currently, only SST-2 and MNLI are considered. Also, the paper lacks the comparison with \"[NeurIPS2020] DynaBERT: Dynamic BERT with Adaptive Widthand Depth\", which is also able to balance the latency and accuracy.\n\n2. The training process is quite complicated and involves multiple steps, e.g., Length-Drop, LayerDrop, and Sandwich rule. The author has not explained the relative contribution of each techniques. Also, the author has not reported the training time of the whole pipeline.\n\n\n##########################################################################\n\nQuestions during rebuttal period: \n \n\nPlease address and clarify the cons above \n \n\n#########################################################################\n\nTypos: \n\n(1) Section 2.2, Paragraph 2, \"more efficient and more accuracy\" should be \"more efficient and more accurate\"\n(2) Page 7, under Table 1 and Figure 4, \"Lengnth\" should be \"Length\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lacks experiments", "review": "##########################################################################\n\nSummary:\n\nThe paper proposed the Length-Adaptive Transformer. The model can be trained once and directly applied to different inference scenarios. To achieve this goal, the author proposed the LengthDrop method, which randomly samples the length at each layer. In addition, the author used the sandwich rule to train the model. At each step, the sandwich rule will train the largest model, the smallest model, and another bunch of randomly sampled models. In the inference phase, the paper proposed to search for the best length configuration that balances the accuracy and latency tradeoff via evolutionary search. Moreover, to generalize the model to token annotation tasks, the author proposed the Drop-and-Restore process, in which the tokens that have been dropped are used again in the final layer. Experiments show that Length-Adaptive Transformer is able to outperfom the baseline models when evaluated at the same latency level.\n\n\n##########################################################################\n\nReasons for score: \n \n\nThe paper is well-written and the length-adaptive idea is reasonable. However, I think it still requires more experiments. Also, the training process is not very clear to the reader. It is not so clear how different techniques impact the final performance and the author has not reported the training time. Due to these two reasons, I choose to vote for weak rejection.\n \n\n##########################################################################\n\nPros: \n \n\n1. The idea of length-adaptive transformer is novel. Perform on-demand truncation of the hidden lengths has not been well explored.\n \n\n##########################################################################\n\nCons: \n \n\n1. The paper needs to conduct experiments on more text classification dataset. Currently, only SST-2 and MNLI are considered. Also, the paper lacks the comparison with \"[NeurIPS2020] DynaBERT: Dynamic BERT with Adaptive Widthand Depth\", which is also able to balance the latency and accuracy.\n\n2. The training process is quite complicated and involves multiple steps, e.g., Length-Drop, LayerDrop, and Sandwich rule. The author has not explained the relative contribution of each techniques. Also, the author has not reported the training time of the whole pipeline.\n\n\n##########################################################################\n\nQuestions during rebuttal period: \n \n\nPlease address and clarify the cons above \n \n\n#########################################################################\n\nTypos: \n\n(1) Section 2.2, Paragraph 2, \"more efficient and more accuracy\" should be \"more efficient and more accurate\"\n(2) Page 7, under Table 1 and Figure 4, \"Lengnth\" should be \"Length\"", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603878056478}, {"id": "edTeUw-WAh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3047/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims to make inference more efficient for finetuned contextual representation models such as BERT. The authors extend PowerBERT, a method introduced recently to perform  efficient inference by dynamically reducing input tokens as the model goes deeper. The authors address two limitations of PowerBERT: the need to pre-set the required computational budget during training (which makes the model inflexible), and the inability to tackle span level tasks (which require the full sentence at the final layer). The authors present a simple solution to both problems (though one that requires heavy engineering to work, see below), and show promising results compared to several BERT baselines. Overall, the ideas presented in this paper are simple (in a good way), and I would like to see more work that applies similar ideas to gain efficiency and not just accuracy. However, the paper leaves important questions unanswered, and thus I cannot recommend accepting it. First, how expensive is the proposed evolutionary search compared to standard fine-tuning? Second, how does this approach compare to the original PowerBERT model (despite its limitations)? Third, given that the evolutionary search is performed on the validation set, how do the authors evaluate it during model development? \nI am curious to read the authors' response, and could be convinced to increase my score, but currently this paper does not meet the ICLR bar.\n\nDetailed summary:\n\nThis paper takes PowerBERT, a recently introduced model for efficient inference, and addresses two of its limitations: the need to retrain the model for each computational budget, and the inability to handle span-level tasks such as QA. The authors present simple solutions to both: for the former, they train a model that randomly selects the number of dropped tokens during training, which makes the model more resilient to different levels of pruning, and perform an evolutionary search process to match the exact pruning levels for a given computational budget. For the latter, they restore the tokens dropped at earlier stages in the last layer, allowing the model to access span indices for tokens dropped earlier if necessary. These solutions are solid and seem to work in practice (although they require quite a bit of engineering to work well, see section 3.1). Nonetheless, their implementation leaves a few important questions unanswered:\n\n1. If I understand correctly, the proposed evolutionary search process needs to be run for any computational budget, which makes it suffer from the same problem as the original PowerBERT model. The authors address this concern in section 3.2, saying that \"it only require a single pass through the relatively small validation set for each length configuration\", but later say they \"repeat this iteration G times\" (with G=30) and \"evolutionary search converges after about fifteen iterations\". This leads to the natural question: how much faster is this procedure compared to finetuning the model, as done in PowerBERT? And even if it is faster, it still doesn't result in a completely flexible model, as each new budget level requires rerunning this process.\n2. Another issue with this approach, is that the model is (partially) trained on the development set. This leaves me wondering how it was evaluated during model development. \n3. The authors build on the PowerBERT model, but never compare to it. Although they (supposedly) provide important benefits compared to this model, it is important to understand at what (accuracy and efficiency) cost. \n\nMore comments: \n1. This paper would strongly benefit from a proof-read. Several paragraphs and sentences repeat what was just said (e.g., the paragraph starting with \"It is not trivial to find an optimal length\" on page 2, or the first sentence on page 7). There are multiple typos (e.g., \"Lengnth\" on Table 1)  and grammatical errors (\"it converges rapidly evolutionary search converges after about fifteen iterations.\") all around the paper. Moreover, several issues were unclear to me: \na. what are sub-models? are these layers with fewer tokens? \nb. how is model distillation incorporated in the solution (last paragraph of section 3.1)?\n2. BERT uses word-pieces and not words.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple and intuitive idea, important questions unanswered", "review": "This paper aims to make inference more efficient for finetuned contextual representation models such as BERT. The authors extend PowerBERT, a method introduced recently to perform  efficient inference by dynamically reducing input tokens as the model goes deeper. The authors address two limitations of PowerBERT: the need to pre-set the required computational budget during training (which makes the model inflexible), and the inability to tackle span level tasks (which require the full sentence at the final layer). The authors present a simple solution to both problems (though one that requires heavy engineering to work, see below), and show promising results compared to several BERT baselines. Overall, the ideas presented in this paper are simple (in a good way), and I would like to see more work that applies similar ideas to gain efficiency and not just accuracy. However, the paper leaves important questions unanswered, and thus I cannot recommend accepting it. First, how expensive is the proposed evolutionary search compared to standard fine-tuning? Second, how does this approach compare to the original PowerBERT model (despite its limitations)? Third, given that the evolutionary search is performed on the validation set, how do the authors evaluate it during model development? \nI am curious to read the authors' response, and could be convinced to increase my score, but currently this paper does not meet the ICLR bar.\n\nDetailed summary:\n\nThis paper takes PowerBERT, a recently introduced model for efficient inference, and addresses two of its limitations: the need to retrain the model for each computational budget, and the inability to handle span-level tasks such as QA. The authors present simple solutions to both: for the former, they train a model that randomly selects the number of dropped tokens during training, which makes the model more resilient to different levels of pruning, and perform an evolutionary search process to match the exact pruning levels for a given computational budget. For the latter, they restore the tokens dropped at earlier stages in the last layer, allowing the model to access span indices for tokens dropped earlier if necessary. These solutions are solid and seem to work in practice (although they require quite a bit of engineering to work well, see section 3.1). Nonetheless, their implementation leaves a few important questions unanswered:\n\n1. If I understand correctly, the proposed evolutionary search process needs to be run for any computational budget, which makes it suffer from the same problem as the original PowerBERT model. The authors address this concern in section 3.2, saying that \"it only require a single pass through the relatively small validation set for each length configuration\", but later say they \"repeat this iteration G times\" (with G=30) and \"evolutionary search converges after about fifteen iterations\". This leads to the natural question: how much faster is this procedure compared to finetuning the model, as done in PowerBERT? And even if it is faster, it still doesn't result in a completely flexible model, as each new budget level requires rerunning this process.\n2. Another issue with this approach, is that the model is (partially) trained on the development set. This leaves me wondering how it was evaluated during model development. \n3. The authors build on the PowerBERT model, but never compare to it. Although they (supposedly) provide important benefits compared to this model, it is important to understand at what (accuracy and efficiency) cost. \n\nMore comments: \n1. This paper would strongly benefit from a proof-read. Several paragraphs and sentences repeat what was just said (e.g., the paragraph starting with \"It is not trivial to find an optimal length\" on page 2, or the first sentence on page 7). There are multiple typos (e.g., \"Lengnth\" on Table 1)  and grammatical errors (\"it converges rapidly evolutionary search converges after about fifteen iterations.\") all around the paper. Moreover, several issues were unclear to me: \na. what are sub-models? are these layers with fewer tokens? \nb. how is model distillation incorporated in the solution (last paragraph of section 3.1)?\n2. BERT uses word-pieces and not words.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603643758019}], "openreview_url": "https://openreview.net/forum?id=tqc8n6oHCtZ", "arxiv_id": "2010.07003", "paper_pdf": "papers/tqc8n6oHCtZ.pdf", "paper_pdf_sha256": "1498257af2d552ddc7840717dc3eb672b62b2d55e4cb2ba7efe22686eba77c28", "paper_pdf_bytes": 308856, "paper_pdf_source": "openreview", "code_url": "https://github.com/clovaai/length-adaptive-transformer", "code_repository": "clovaai/length-adaptive-transformer", "code_commit": "8c35cb3c2ea76112d973e8e5f867330af9f52113", "code_archive": "repos/tqc8n6oHCtZ.zip", "code_archive_sha256": "290896c6a9ac7a1f9080700f6b3a021b9916f682f1e1cb3509462f936ba40225", "code_archive_bytes": 47571, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 43, "github_languages": {"Python": 151470}, "github_archived": false, "github_pushed_at": "2020-11-02T12:22:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/length-adaptive-transformer-train-once-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJxt2aVFPr", "year": 2020, "status": "rejected", "title": "Optimizing Data Usage via Differentiable Rewards", "authors": ["Xinyi Wang", "Hieu Pham", "Paul Michel", "Antonios Anastasopoulos", "Graham Neubig", "Jaime Carbonell"], "authorids": ["xinyiw1@cs.cmu.edu", "hyhieu@cmu.edu", "pmichel1@cs.cmu.edu", "aanastas@andrew.cmu.edu", "gneubig@cs.cmu.edu", "jgc@cs.cmu.edu"], "authors_source": "OpenReview API", "abstract": "To acquire a new skill, humans learn better and faster if a tutor, based on their current knowledge level, informs them of how much attention they should pay to particular content or practice problems. Similarly, a machine learning model could potentially be trained better with a scorer that “adapts” to its current learning state and estimates the importance of each training data instance. Training such an adaptive scorer efficiently is a challenging problem; in order to precisely quantify the effect of a data instance at a given time during the training, it is typically necessary to first complete the entire training process. To efficiently optimize data usage, we propose a reinforcement learning approach called Differentiable Data Selection (DDS). In DDS, we formulate a scorer network as a learnable function of the training data, which can be efficiently updated along with the main model being trained. Specifically, DDS updates the scorer with an intuitive reward signal: it should up-weigh the data that has a similar gradient with a dev set upon which we would finally like to perform well. Without significant computing overhead, DDS delivers strong and consistent improvements over several strong baselines on two very different tasks of machine translation and image classification.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Skgt9dG-9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper790/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper introduces a simple idea to optimize the weights of a weighted empirical training distributions. The goal is to optimize the population risk, and the idea is to optimize a distribution over the training examples to maximize the cosine similarity between training set gradients and validation set gradients. The distribution over the training set is parameterized by a neural network taking as arguments the\n\nStrengths:\n- The method is quite simple.\n- The results appear to be strong, although I am less familiar with the NMT baselines. The imagenet results seem quite strong to me.\n\nWeaknesses:\n- I couldn't find a particularly clear description of the scoring networks architecture. Given that it observes the whole dataset, this seems like a critical choice that could have a big impact on the complexity of this approach. At the very least, this should be clearly reported, and I recommend a more thorough investigation of this choice.\n- The authors report that their method takes 1.5x to 2x longer to run than the uniform baseline. Yet, they ran all methods for the same number of steps / epochs. It seems to me that a fairer comparison might be letting all methods enjoy the same total budget measure roughly by wall time.\n\nQuestions:\n- I didn't follow why the computation of the per example gradient grad l(x_i, y_i, theta_t-1) is so onerous. Isn't that computed on line 5 already?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary: This paper introduces a simple idea to optimize the weights of a weighted empirical training distributions. The goal is to optimize the population risk, and the idea is to optimize a distribution over the training examples to maximize the cosine similarity between training set gradients and validation set gradients. The distribution over the training set is parameterized by a neural network taking as arguments the\n\nStrengths:\n- The method is quite simple.\n- The results appear to be strong, although I am less familiar with the NMT baselines. The imagenet results seem quite strong to me.\n\nWeaknesses:\n- I couldn't find a particularly clear description of the scoring networks architecture. Given that it observes the whole dataset, this seems like a critical choice that could have a big impact on the complexity of this approach. At the very least, this should be clearly reported, and I recommend a more thorough investigation of this choice.\n- The authors report that their method takes 1.5x to 2x longer to run than the uniform baseline. Yet, they ran all methods for the same number of steps / epochs. It seems to me that a fairer comparison might be letting all methods enjoy the same total budget measure roughly by wall time.\n\nQuestions:\n- I didn't follow why the computation of the per example gradient grad l(x_i, y_i, theta_t-1) is so onerous. Isn't that computed on line 5 already?"}, "tcdate": 1572051088743}, {"id": "r1lqmp15YS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper790/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a reinforcement learning approach towards using data that present best correlation with a validation set’s gradient signal. The broader point of this paper is that there is inevitably some distribution shift going from train to test set - and the validation set can be a small curated set whose distribution is closer to the testing distribution than what the training dataset's distribution is. \n\nThe problem setup bears relationship to several areas including domain adaptation/covariate shift problems, curriculum learning based approaches amongst others. One assumption that I see which needs to be understood more is equation (6) - wherein, somehow, there is a Markov assumption used to zero out the contribution of the scoring network on parameters unto previous time step. Trying to understand the implications of this assumption (how the performance varies with/without this assumption) would be instructive for understanding potential shortcomings of this framework.\n\nI think the paper is well written, handles an important question. That said, I am not too aware of recent work in this area to make a decisive judgement on this paper’s novelty/contributions. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper presents a reinforcement learning approach towards using data that present best correlation with a validation set’s gradient signal. The broader point of this paper is that there is inevitably some distribution shift going from train to test set - and the validation set can be a small curated set whose distribution is closer to the testing distribution than what the training dataset's distribution is. \n\nThe problem setup bears relationship to several areas including domain adaptation/covariate shift problems, curriculum learning based approaches amongst others. One assumption that I see which needs to be understood more is equation (6) - wherein, somehow, there is a Markov assumption used to zero out the contribution of the scoring network on parameters unto previous time step. Trying to understand the implications of this assumption (how the performance varies with/without this assumption) would be instructive for understanding potential shortcomings of this framework.\n\nI think the paper is well written, handles an important question. That said, I am not too aware of recent work in this area to make a decisive judgement on this paper’s novelty/contributions. "}, "tcdate": 1571581217748}, {"id": "rkgz2PwHYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper790/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an iterative method that jointly trains the model and a scorer network that places a non-uniform distribution over data sets.  The paper proposes a gradient method to learn the scorer network based on reinforcement learning, which is novel as to what the reviewer knows.\n\nThere are several concerns/questions:\n\n1) The paper doesn’t define the D_{dev} clearly. How is D_{dev} chosen? Is it a subset of D_{train}? \n\n2) In section 2.1, why “smaller development set D_{dev} is much closer to the P_{test}(X,Y)”? P_{test}(X,Y) is supposed to be not observed during training?\n\n3) In Eq (5), if D_{dev} is s subset of D_{train}, if \\theta* is the minimal of J, it means the gradient \nat  \\theta* is 0. To calculate the gradient of J with respect to \\psi, by chain rule, it need to calculate gradient to \\theta* first then \\theta* to \\psi. If gradient of \\theta* is 0, the product is also 0? So the \\psi will not be updated if D_{dev}  is sufficiently similar to D_{train} ?\n\n4) In Section 2.3, it omits the second order Hessian term. How does that influence the performance? \n\n5) it mentions “without significant computing overhead“ in abstract, which is not demonstrated elsewhere.\n\n6) In the experiments, table 1, it seems the major improvement comes from retrain and TCS rather than DDS? In figure 3, it is better to show the weights of an image without DDS and comparing that with DDS.\n\n7) The paper contains many typos such as Eqn.11 is not defined in the main paper, the “Eqn ??” Appears in the appendix, “tha minimizes” etc.\n\nIn general, the idea of the paper is natural and the results seem promising. I am looking forward to the reply to my questions/concerns. \n\n#############\n\nI have read the author's feedback. I think the clarity of both methodology and experiment does not reach the acceptance level and would maintain my current rating. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "The paper proposes an iterative method that jointly trains the model and a scorer network that places a non-uniform distribution over data sets.  The paper proposes a gradient method to learn the scorer network based on reinforcement learning, which is novel as to what the reviewer knows.\n\nThere are several concerns/questions:\n\n1) The paper doesn’t define the D_{dev} clearly. How is D_{dev} chosen? Is it a subset of D_{train}? \n\n2) In section 2.1, why “smaller development set D_{dev} is much closer to the P_{test}(X,Y)”? P_{test}(X,Y) is supposed to be not observed during training?\n\n3) In Eq (5), if D_{dev} is s subset of D_{train}, if \\theta* is the minimal of J, it means the gradient \nat  \\theta* is 0. To calculate the gradient of J with respect to \\psi, by chain rule, it need to calculate gradient to \\theta* first then \\theta* to \\psi. If gradient of \\theta* is 0, the product is also 0? So the \\psi will not be updated if D_{dev}  is sufficiently similar to D_{train} ?\n\n4) In Section 2.3, it omits the second order Hessian term. How does that influence the performance? \n\n5) it mentions “without significant computing overhead“ in abstract, which is not demonstrated elsewhere.\n\n6) In the experiments, table 1, it seems the major improvement comes from retrain and TCS rather than DDS? In figure 3, it is better to show the weights of an image without DDS and comparing that with DDS.\n\n7) The paper contains many typos such as Eqn.11 is not defined in the main paper, the “Eqn ??” Appears in the appendix, “tha minimizes” etc.\n\nIn general, the idea of the paper is natural and the results seem promising. I am looking forward to the reply to my questions/concerns. \n\n#############\n\nI have read the author's feedback. I think the clarity of both methodology and experiment does not reach the acceptance level and would maintain my current rating. \n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571284906456}], "openreview_url": "https://openreview.net/forum?id=BJxt2aVFPr", "arxiv_id": "1911.10088", "paper_pdf": "papers/BJxt2aVFPr.pdf", "paper_pdf_sha256": "e2e75774eaa658db7b1b16c602f01d4f0a3c0d0d4f7b4d6d0113b34ebc57667f", "paper_pdf_bytes": 932144, "paper_pdf_source": "openreview", "code_url": "https://github.com/cindyxinyiwang/DataSelection", "code_repository": "cindyxinyiwang/DataSelection", "code_commit": "b7cbb2cd819640aa9ed0e1090c18fce2095082ae", "code_archive": "repos/BJxt2aVFPr.zip", "code_archive_sha256": "9d583ea78e867281933b50c814a52c88f2f6c93d4de6ac99aa5bdd1eba06d87d", "code_archive_bytes": 61741, "code_file_count": 24, "code_extensions": {".py": 17, ".sh": 7}, "github_disk_usage_kb": 52, "github_languages": {"Python": 262569, "Shell": 12989, "Perl": 4412}, "github_archived": false, "github_pushed_at": "2021-06-20T00:19:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimizing-data-usage-via-differentiable-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJiaRbk0-", "year": 2018, "status": "rejected", "title": "Towards Binary-Valued Gates for Robust LSTM Training ", "authors": ["Zhuohan Li", "Di He", "Fei Tian", "Wei Chen", "Tao Qin", "Liwei Wang", "Tie-Yan Liu"], "authorids": ["lizhuohan@pku.edu.cn", "di_he@pku.edu.cn", "fetia@microsoft.com", "wche@microsoft.com", "taoqin@microsoft.com", "wanglw@cis.pku.edu.cn", "tyliu@microsoft.com"], "authors_source": "OpenReview API", "abstract": "Long Short-Term Memory (LSTM) is one of the most widely used recurrent structures in sequence modeling. Its goal is to use gates to control the information flow (e.g., whether to skip some information/transformation or not) in the recurrent computations, although its practical implementation based on soft gates only partially achieves this goal and is easy to overfit. In this paper, we propose a new way for LSTM training, which pushes the values of the gates towards 0 or 1. By doing so, we can (1) better control the information flow: the gates are mostly open or closed, instead of in a middle state; and (2) avoid overfitting to certain extent: the gates operate at their flat regions, which is shown to correspond to better generalization ability. However, learning towards discrete values of the gates is generally difficult. To tackle this challenge, we leverage the recently developed Gumbel-Softmax trick from the field of variational methods, and make the model trainable with standard backpropagation. Experimental results on language modeling and machine translation show that (1) the values of the gates generated by our method are more reasonable and intuitively interpretable, and (2) our proposed method generalizes better and achieves better accuracy on test sets in all tasks. Moreover, the learnt models are not sensitive to low-precision approximation and low-rank approximation of the gate parameters due to the flat loss surface.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyA3jBqgG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper118/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper propose a new \"gate\" function for LSTM to enable the values of the gates towards 0 or 1. The motivation behind is a  flat region of the loss surface is likely to generalize well. It shows the experimental results are comparable or better than vanilla LSTM and much more robust to low-precision approximation and low-rank approximation.\n\nIn section 3.2, the paper claimed using a smaller temperature cannot guarantee the outputs to be close to the boundary. Is there any experimental evidence to show it's not working? It also claimed pushing output gate to 0/1 will drop the performance. It actually quite interesting because there are bunch of paper claimed output gate is not important for language modeling, e.g. https://openreview.net/pdf?id=HJOQ7MgAW . \n\nIn the sensitive analysis, what if apply rounding / low-rank for all the parameters? \n\nHow was this approach compare to binarynet https://arxiv.org/abs/1602.02830 ? Applying the same idea, but only for forget gate/ input gate. Also, can we apply this idea to the binarynet? \n\nOverall, I think it's an interesting paper but I feel it should compare with some simple baseline to binarized the gate function.  \n\nUpdates: Thanks a lot for all the clarification. It do improve the paper quality but I'm still thinking it's higher than \"6\" but lower than \"7\". To me, improve ppl from \"52.8\" to \"52.1\" isn't very significant. For WMT, it improve on DE->EN but not for EN->DE (although it improve both for the author's own baseline). So I'm not fully convinced this approach could improve the generalization. But I feel this work can have many other applications such as \"binarynet\". ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "rating": "6: Marginally above acceptance threshold", "review": "This paper propose a new \"gate\" function for LSTM to enable the values of the gates towards 0 or 1. The motivation behind is a  flat region of the loss surface is likely to generalize well. It shows the experimental results are comparable or better than vanilla LSTM and much more robust to low-precision approximation and low-rank approximation.\n\nIn section 3.2, the paper claimed using a smaller temperature cannot guarantee the outputs to be close to the boundary. Is there any experimental evidence to show it's not working? It also claimed pushing output gate to 0/1 will drop the performance. It actually quite interesting because there are bunch of paper claimed output gate is not important for language modeling, e.g. https://openreview.net/pdf?id=HJOQ7MgAW . \n\nIn the sensitive analysis, what if apply rounding / low-rank for all the parameters? \n\nHow was this approach compare to binarynet https://arxiv.org/abs/1602.02830 ? Applying the same idea, but only for forget gate/ input gate. Also, can we apply this idea to the binarynet? \n\nOverall, I think it's an interesting paper but I feel it should compare with some simple baseline to binarized the gate function.  \n\nUpdates: Thanks a lot for all the clarification. It do improve the paper quality but I'm still thinking it's higher than \"6\" but lower than \"7\". To me, improve ppl from \"52.8\" to \"52.1\" isn't very significant. For WMT, it improve on DE->EN but not for EN->DE (although it improve both for the author's own baseline). So I'm not fully convinced this approach could improve the generalization. But I feel this work can have many other applications such as \"binarynet\". ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511836598499}, {"id": "Syo-smqgf", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper118/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper argues for pushing the input and forget gate’s output toward 0 or 1, i.e., the LSTM tends to reside in flat region of surface loss, which is likely to generalize well. To achieve that, the sigmoid function in the original LSTM is replaced by a function G that is continuous and differentiable with respect to the parameters (by applying the Gumbel-Softmax trick). As a result, the model is still differentiable while the output gate is approximately binarized.  \n\nPros:\n-\tThe paper is clearly written\n-\tThe method is new and somehow theoretically guaranteed by the proof of the Proposition 1\n-\tThe experiments are clearly explained with detailed configurations\n-\tThe performance of the method in the model compression task is promising \n\nCons:\n-\tThe “simple deduction” which states that pushing the gate values toward 0 or 1 correspond to the region of the overall loss surface may need more theoretical analysis\n-\tIt is confusing whether the output of the gate is sampled based on or computed directly by the function G  \n-\tThe experiments lack many recent baselines on the same dataset (Penn Treebank: Melis et al. (2017) – On the State of the Art of Evaluation in Neural Language Models; WMT: Ashish et.al. (2017) – Attention Is All You Need) \n-\tThe experiment’s result is only slightly better than the baseline’s\n-\tTo be more persuasive, the author should include in the baselines other method that can “binerize” the gate values such as the one sharpening the sigmoid function. \n\n\nIn short, this work is worth a read. Although the experimental results are not quite persuasive, the method is nice and promising. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting, but not impressive", "rating": "6: Marginally above acceptance threshold", "review": "The paper argues for pushing the input and forget gate’s output toward 0 or 1, i.e., the LSTM tends to reside in flat region of surface loss, which is likely to generalize well. To achieve that, the sigmoid function in the original LSTM is replaced by a function G that is continuous and differentiable with respect to the parameters (by applying the Gumbel-Softmax trick). As a result, the model is still differentiable while the output gate is approximately binarized.  \n\nPros:\n-\tThe paper is clearly written\n-\tThe method is new and somehow theoretically guaranteed by the proof of the Proposition 1\n-\tThe experiments are clearly explained with detailed configurations\n-\tThe performance of the method in the model compression task is promising \n\nCons:\n-\tThe “simple deduction” which states that pushing the gate values toward 0 or 1 correspond to the region of the overall loss surface may need more theoretical analysis\n-\tIt is confusing whether the output of the gate is sampled based on or computed directly by the function G  \n-\tThe experiments lack many recent baselines on the same dataset (Penn Treebank: Melis et al. (2017) – On the State of the Art of Evaluation in Neural Language Models; WMT: Ashish et.al. (2017) – Attention Is All You Need) \n-\tThe experiment’s result is only slightly better than the baseline’s\n-\tTo be more persuasive, the author should include in the baselines other method that can “binerize” the gate values such as the one sharpening the sigmoid function. \n\n\nIn short, this work is worth a read. Although the experimental results are not quite persuasive, the method is nice and promising. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511828226587}, {"id": "S15OPlugz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper118/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims to push the LSTM gates to be binary. To achieve this, the paper proposes to employ the recent Gumbel-Softmax trick to obtain end-to-end trainable categorical distribution (taking 0 or 1 value). The resulted G2-LSTM is applied for language model and machine translation in the experiments. \n\nThe novelty of this paper is limited. Just directly apply the Gumbel-Softmax trick. \n\nThe motivation is not explained clearly and convincingly. Why need to pursue binary gates? According to the paper, it may give better generalization performance. But there is no theoretical or experimental evidence provided by this paper to support this argument. \n\nThe results of the new G2-LSTM are not significantly better than baselines in the experiments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The technical novelty is limited and experiments do not show much benefits of the proposed model.", "rating": "4: Ok but not good enough - rejection", "review": "This paper aims to push the LSTM gates to be binary. To achieve this, the paper proposes to employ the recent Gumbel-Softmax trick to obtain end-to-end trainable categorical distribution (taking 0 or 1 value). The resulted G2-LSTM is applied for language model and machine translation in the experiments. \n\nThe novelty of this paper is limited. Just directly apply the Gumbel-Softmax trick. \n\nThe motivation is not explained clearly and convincingly. Why need to pursue binary gates? According to the paper, it may give better generalization performance. But there is no theoretical or experimental evidence provided by this paper to support this argument. \n\nThe results of the new G2-LSTM are not significantly better than baselines in the experiments.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511683954491}], "openreview_url": "https://openreview.net/forum?id=rJiaRbk0-", "arxiv_id": "1806.02988", "paper_pdf": "papers/rJiaRbk0-.pdf", "paper_pdf_sha256": "30f2ae910e4e16a7636eb28772281d71599260ec5674b38373054756f369fef1", "paper_pdf_bytes": 763886, "paper_pdf_source": "openreview", "code_url": "https://github.com/zhuohan123/g2-lstm", "code_repository": "zhuohan123/g2-lstm", "code_commit": "542aee7c014f3d1c24da23c5363994c3d7b3999e", "code_archive": "repos/rJiaRbk0-.zip", "code_archive_sha256": "51c6e7d06c403f4783d97ef9c50686407d75681247b799b8f5a7c438cf1e374e", "code_archive_bytes": 676052, "code_file_count": 53, "code_extensions": {".py": 52, ".sh": 1}, "github_disk_usage_kb": 642, "github_languages": {"Python": 345728, "Perl": 15002, "Shell": 783}, "github_archived": false, "github_pushed_at": "2018-07-22T18:45:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-binary-valued-gates-for-robust-lstm"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qPwSoOQPrI", "year": 2026, "status": "rejected", "title": "Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution", "authors": ["Qiao Xiao", "Alan Ansell", "Boqian Wu", "Lu Yin", "Mykola Pechenizkiy", "Shiwei Liu", "Decebal Constantin Mocanu"], "authorids": ["~Qiao_Xiao1", "~Alan_Ansell1", "~Boqian_Wu1", "~Lu_Yin1", "~Mykola_Pechenizkiy1", "~Shiwei_Liu2", "~Decebal_Constantin_Mocanu1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have achieved remarkable success across various tasks but face deployment challenges due to their massive computational demands. While post-training pruning methods like SparseGPT and Wanda can effectively reduce the model size, but struggle to maintain model performance at high sparsity levels, limiting their utility for downstream tasks. Existing fine-tuning methods, such as full fine-tuning and LoRA, fail to preserve sparsity as they require updating the whole dense matrices, not well-suited for sparse LLMs. In this paper, we propose \\textbf{Sparsity Evolution Fine-Tuning (SEFT)}, a novel method designed specifically for sparse LLMs. SEFT dynamically evolves the sparse topology of pruned models during fine-tuning, while preserving the overall sparsity throughout the process. The strengths of SEFT lie in its ability to perform task-specific adaptation through a weight drop-and-grow strategy, enabling the pruned model to self-adapt its sparse connectivity pattern based on the target dataset. Furthermore, a sensitivity-driven pruning criterion is employed to ensure that the desired sparsity level is consistently maintained throughout fine-tuning. Our experiments on various LLMs, including LLaMA families, DeepSeek, and Mistral, across a diverse set of benchmarks demonstrate that SEFT achieves stronger performance while offering superior memory and computation efficiency compared to existing baselines. The code is provided in the supplementary material and will be released publicly.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "LIRYMy0wn2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17173/Reviewer_Jmj9"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a sparse fine-tuning method for large language models (LLMs) that dynamically adjusts the sparsity topology during training. SEFT employs a drop-and-grow mechanism to evolve sparse connections, removing weights with the smallest update magnitudes and reactivating previously pruned ones with the largest current gradients. This allows the sparsity pattern to adapt throughout fine-tuning while maintaining a fixed sparsity ratio. Experiments on models such as Llama and Mistral demonstrate that SEFT achieves better performance and higher efficiency than existing sparse fine-tuning baselines.", "review_text": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a sparse fine-tuning method for large language models (LLMs) that dynamically adjusts the sparsity topology during training. SEFT employs a drop-and-grow mechanism to evolve sparse connections, removing weights with the smallest update magnitudes and reactivating previously pruned ones with the largest current gradients. This allows the sparsity pattern to adapt throughout fine-tuning while maintaining a fixed sparsity ratio. Experiments on models such as Llama and Mistral demonstrate that SEFT achieves better performance and higher efficiency than existing sparse fine-tuning baselines.", "strengths": "1. Efficient fine-tuning for large-scale LLMs is a pressing and valuable research direction. Leveraging sparsity to reduce training costs offers clear benefits for resource-constrained settings.\n2. The motivation is sound and design is reasonable. Allowing the model to reconstruct its sparse topology and reactivate pruned weights provides flexibility for adapting to downstream tasks, addressing the rigidity of static pruning schemes.\n3. Comprehensive experiments and evaluation settings. The experiments cover multiple open-source LLM architectures and two different pruning schemes as initialization. The analysis includes both performance and efficiency metrics, with additional ablations and sensitivity studies in the appendix, lending credibility to the conclusions.\n4. Clear writing and presentation. The paper is well structured, and Figure 1 offers an intuitive comparison between SEFT and LoRA, clearly illustrating the mechanism.", "weaknesses": "1. Missing comparisons to recent baselines. The paper omits comparisons with recent methods such as SparseLoRA [1] and S^$2$FT [2], which are directly relevant. Even a conceptual or mechanism-level discussion would help position SEFT more clearly within the current literature.\n2. Lack of statistical significance analysis. While SEFT often achieves the best average results (e.g., Table 2), the margins are small (~1%). Given the stochastic nature of LLM fine-tuning, reporting averages and standard deviations over multiple seeds or conducting significance tests would strengthen claims such as “SEFT consistently outperforms baselines”.\n3. Insufficient discussion of computational overhead of “drop-and-grow”. The drop-and-grow updates rely on dense gradient computation (Appendix I), yet the main text lacks quantitative analysis of this cost. Although the proposed update scheme mitigates overhead, it may still limit SEFT’s deployability in low-resource environments.\n\n---\n\n**References**\n\n[1] Khaki, S., Li, X., Guo, J., Zhu, L., Plataniotis, K.N., Yazdanbakhsh, A., Keutzer, K., Han, S. &amp; Liu, Z.. (2025). SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:29768-29783.\n\n[2] Yang, X., Leng, J., Guo, G., Zhao, J., Nakada, R., Zhang, L., ... & Chen, B. (2024). S $^{2} $ FT: Efficient, scalable and generalizable LLM fine-tuning by structured sparsity. Advances in Neural Information Processing Systems, 37, 59912-59947.", "questions": "Could the authors quantify the time cost of the topology evolution step relative to the total training process? Would accounting for this overhead alter the comparative efficiency results between SEFT and other methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a sparse fine-tuning method for large language models (LLMs) that dynamically adjusts the sparsity topology during training. SEFT employs a drop-and-grow mechanism to evolve sparse connections, removing weights with the smallest update magnitudes and reactivating previously pruned ones with the largest current gradients. This allows the sparsity pattern to adapt throughout fine-tuning while maintaining a fixed sparsity ratio. Experiments on models such as Llama and Mistral demonstrate that SEFT achieves better performance and higher efficiency than existing sparse fine-tuning baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Efficient fine-tuning for large-scale LLMs is a pressing and valuable research direction. Leveraging sparsity to reduce training costs offers clear benefits for resource-constrained settings.\n2. The motivation is sound and design is reasonable. Allowing the model to reconstruct its sparse topology and reactivate pruned weights provides flexibility for adapting to downstream tasks, addressing the rigidity of static pruning schemes.\n3. Comprehensive experiments and evaluation settings. The experiments cover multiple open-source LLM architectures and two different pruning schemes as initialization. The analysis includes both performance and efficiency metrics, with additional ablations and sensitivity studies in the appendix, lending credibility to the conclusions.\n4. Clear writing and presentation. The paper is well structured, and Figure 1 offers an intuitive comparison between SEFT and LoRA, clearly illustrating the mechanism.", "weaknesses": "1. Missing comparisons to recent baselines. The paper omits comparisons with recent methods such as SparseLoRA [1] and S^$2$FT [2], which are directly relevant. Even a conceptual or mechanism-level discussion would help position SEFT more clearly within the current literature.\n2. Lack of statistical significance analysis. While SEFT often achieves the best average results (e.g., Table 2), the margins are small (~1%). Given the stochastic nature of LLM fine-tuning, reporting averages and standard deviations over multiple seeds or conducting significance tests would strengthen claims such as “SEFT consistently outperforms baselines”.\n3. Insufficient discussion of computational overhead of “drop-and-grow”. The drop-and-grow updates rely on dense gradient computation (Appendix I), yet the main text lacks quantitative analysis of this cost. Although the proposed update scheme mitigates overhead, it may still limit SEFT’s deployability in low-resource environments.\n\n---\n\n**References**\n\n[1] Khaki, S., Li, X., Guo, J., Zhu, L., Plataniotis, K.N., Yazdanbakhsh, A., Keutzer, K., Han, S. &amp; Liu, Z.. (2025). SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:29768-29783.\n\n[2] Yang, X., Leng, J., Guo, G., Zhao, J., Nakada, R., Zhang, L., ... & Chen, B. (2024). S $^{2} $ FT: Efficient, scalable and generalizable LLM fine-tuning by structured sparsity. Advances in Neural Information Processing Systems, 37, 59912-59947.", "questions": "Could the authors quantify the time cost of the topology evolution step relative to the total training process? Would accounting for this overhead alter the comparative efficiency results between SEFT and other methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762092799023}, {"id": "sL2FdJF5bX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17173/Reviewer_rqp2"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method for fine-tuning sparse large language models (LLMs). While traditional fine-tuning approaches like LoRA fail to maintain sparsity and post-training pruning methods (e.g., SparseGPT, Wanda) often cause substantial performance degradation at high sparsity, SEFT addresses these issues by dynamically evolving the sparse topology during fine-tuning while keeping the overall sparsity fixed.", "review_text": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method for fine-tuning sparse large language models (LLMs). While traditional fine-tuning approaches like LoRA fail to maintain sparsity and post-training pruning methods (e.g., SparseGPT, Wanda) often cause substantial performance degradation at high sparsity, SEFT addresses these issues by dynamically evolving the sparse topology during fine-tuning while keeping the overall sparsity fixed.", "strengths": "- The paper is clearly presented and well-organized, with only a few minor typos.\n- The proposed SEFT method is conceptually simple, well-motivated, and demonstrates strong empirical performance.\n- The authors provide comprehensive ablation studies, including analyses of sparsity’s effect on performance, inference speed, and memory efficiency.\n- Experiments are conducted across multiple datasets and diverse LLM families, which strengthens the empirical validation of the method.", "weaknesses": "- Limited Related Work Discussion:\nThe discussion of related work primarily focuses on low-rank methods (e.g., LoRA) but overlooks several recent sparse fine-tuning and pruning approaches that are methodologically closer to SEFT. Although Appendix A.2 and A.3 cover PEFT and dynamic sparse training, it would strengthen the paper to include recent sparse PEFT methods (e.g., [1,2,3,4]) for a more comprehensive contextualization.\n- Novelty Clarification:\nThe paper lacks a clear explanation of how SEFT differs from similar sparse fine-tuning frameworks, particularly SpIEL [5], which also incorporate grow-and-drop phases and gradient-based parameter selection. A more explicit discussion comparing SEFT’s innovations—such as its sparsity adaptation phase or index update mechanism—with these prior works would clarify its novelty. Besides, similar to SMT[3], SEFT also introduce reconstruction indices method to maintain and train sparse model.  SEFT also have similar parameter selection apprach regards to gradient magnitude similar to SpIEL[5] and SMT[3]. More discussion about the similarity will be appreciated. \n- Missing Baseline Comparisons:\nSEFT is conceptually similar to several sparse fine-tuning methods (e.g., SMT [3], SpIEL [5]), yet these are not included as baselines in the experiments. Adding such comparisons would provide a more balanced and convincing evaluation of SEFT’s performance and efficiency.\n- Minor Typos:\nInconsistent spelling: “fine-tuning” vs. “finetuning” → use one form (“fine-tuning”).\n“enables sparse LLMs recover” → “enables sparse LLMs to recover.”\n“delta vector δ dynamically explore” → “explores.”\n“Specifically models are fine-tuned…” → “Specifically, models are fine-tuned…”\n\n\n[1] The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks\n\n[2] Parameter-Efficient Fine-Tuning without Introducing New Latency\n\n[3] Sparse Matrix in Large Language Model Fine-tuning\n\n[4] Training Neural Networks with Fixed Sparse Masks\n\n[5] Scaling Sparse Fine-Tuning to Large Language Models", "questions": "- Memory Cost:\nSEFT maintains reconstruction indices to record sparsity patterns. Could the authors provide quantitative estimates of this additional memory overhead? Sparse matrices typically require extra storage for index data, potentially offsetting efficiency gains. Further, does maintaining and reconstructing these indices introduce additional copy operations or memory fragmentation?\n\n- Training Speed:\nSparse matrix operations often involve index resolution and memory mapping, which may reduce cache efficiency and increase latency during the indices read operations. While the paper evaluates inference latency (main text and Appendix B), it lacks discussion of training time. Including an analysis of training speed and a comparison to dense fine-tuning methods would help clarify whether SEFT introduces measurable overhead during training.\n\nThis paper presents a promising and well-executed approach to fine-tuning sparse LLMs. However, its contribution would be strengthened by a deeper comparison to related sparse fine-tuning works, explicit clarification of novel aspects, and quantitative discussion of memory and training-time overhead.\nIf these concerns are addressed in the rebuttal, I would be willing to reconsider my current rating.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method for fine-tuning sparse large language models (LLMs). While traditional fine-tuning approaches like LoRA fail to maintain sparsity and post-training pruning methods (e.g., SparseGPT, Wanda) often cause substantial performance degradation at high sparsity, SEFT addresses these issues by dynamically evolving the sparse topology during fine-tuning while keeping the overall sparsity fixed.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "- The paper is clearly presented and well-organized, with only a few minor typos.\n- The proposed SEFT method is conceptually simple, well-motivated, and demonstrates strong empirical performance.\n- The authors provide comprehensive ablation studies, including analyses of sparsity’s effect on performance, inference speed, and memory efficiency.\n- Experiments are conducted across multiple datasets and diverse LLM families, which strengthens the empirical validation of the method.", "weaknesses": "- Limited Related Work Discussion:\nThe discussion of related work primarily focuses on low-rank methods (e.g., LoRA) but overlooks several recent sparse fine-tuning and pruning approaches that are methodologically closer to SEFT. Although Appendix A.2 and A.3 cover PEFT and dynamic sparse training, it would strengthen the paper to include recent sparse PEFT methods (e.g., [1,2,3,4]) for a more comprehensive contextualization.\n- Novelty Clarification:\nThe paper lacks a clear explanation of how SEFT differs from similar sparse fine-tuning frameworks, particularly SpIEL [5], which also incorporate grow-and-drop phases and gradient-based parameter selection. A more explicit discussion comparing SEFT’s innovations—such as its sparsity adaptation phase or index update mechanism—with these prior works would clarify its novelty. Besides, similar to SMT[3], SEFT also introduce reconstruction indices method to maintain and train sparse model.  SEFT also have similar parameter selection apprach regards to gradient magnitude similar to SpIEL[5] and SMT[3]. More discussion about the similarity will be appreciated. \n- Missing Baseline Comparisons:\nSEFT is conceptually similar to several sparse fine-tuning methods (e.g., SMT [3], SpIEL [5]), yet these are not included as baselines in the experiments. Adding such comparisons would provide a more balanced and convincing evaluation of SEFT’s performance and efficiency.\n- Minor Typos:\nInconsistent spelling: “fine-tuning” vs. “finetuning” → use one form (“fine-tuning”).\n“enables sparse LLMs recover” → “enables sparse LLMs to recover.”\n“delta vector δ dynamically explore” → “explores.”\n“Specifically models are fine-tuned…” → “Specifically, models are fine-tuned…”\n\n\n[1] The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks\n\n[2] Parameter-Efficient Fine-Tuning without Introducing New Latency\n\n[3] Sparse Matrix in Large Language Model Fine-tuning\n\n[4] Training Neural Networks with Fixed Sparse Masks\n\n[5] Scaling Sparse Fine-Tuning to Large Language Models", "questions": "- Memory Cost:\nSEFT maintains reconstruction indices to record sparsity patterns. Could the authors provide quantitative estimates of this additional memory overhead? Sparse matrices typically require extra storage for index data, potentially offsetting efficiency gains. Further, does maintaining and reconstructing these indices introduce additional copy operations or memory fragmentation?\n\n- Training Speed:\nSparse matrix operations often involve index resolution and memory mapping, which may reduce cache efficiency and increase latency during the indices read operations. While the paper evaluates inference latency (main text and Appendix B), it lacks discussion of training time. Including an analysis of training speed and a comparison to dense fine-tuning methods would help clarify whether SEFT introduces measurable overhead during training.\n\nThis paper presents a promising and well-executed approach to fine-tuning sparse LLMs. However, its contribution would be strengthened by a deeper comparison to related sparse fine-tuning works, explicit clarification of novel aspects, and quantitative discussion of memory and training-time overhead.\nIf these concerns are addressed in the rebuttal, I would be willing to reconsider my current rating.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761878639890}, {"id": "qFsRJSN37y", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17173/Reviewer_wBjB"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors experiment on LLaMA, DeepSeek, and Mistral model families, pruned with both SparseGPT and Wanda. The results demonstrate that SEFT outperforms baselines like LoRA* (prune-after-tuning), SPP, and SQFT across various benchmarks (Commonsense Reasoning, MMLU, GSM8K), with a particularly strong advantage at high sparsity levels (e.g., 70%-80%). Furthermore, SEFT shows significant efficiency gains, including up to 2x memory savings during training and up to 2.5x inference speedup on a CPU.", "review_text": "The authors experiment on LLaMA, DeepSeek, and Mistral model families, pruned with both SparseGPT and Wanda. The results demonstrate that SEFT outperforms baselines like LoRA* (prune-after-tuning), SPP, and SQFT across various benchmarks (Commonsense Reasoning, MMLU, GSM8K), with a particularly strong advantage at high sparsity levels (e.g., 70%-80%). Furthermore, SEFT shows significant efficiency gains, including up to 2x memory savings during training and up to 2.5x inference speedup on a CPU.", "strengths": "SEFT cleverly adapts the idea of Dynamic Sparse Training (DST) to the novel scenario of \"fine-tuning an already-sparse model.\" The \"Drop-Grow-Adapt\" three-step cycle shown in Figure 2 is clear and logically sound. The \"Grow\" step, in particular, which allows the model to \"resurrect\" pruned weights, is key to its task-adaptive capability.\n\nhe analyses in the appendix are thorough. For example, Appendix E.1 (Table 7) proves the necessity of \"allowing resurrection of pruned weights,\" and Appendix E.2 (Fig 5a) demonstrates the superiority of the \"sensitivity-based\" pruning criterion over \"magnitude-based.\" This greatly enhances the credibility of the method's design.", "weaknesses": "SEFT introduces new hyperparameters that require tuning, most notably the sparse topology \"update frequency $k$\" and the \"drop rate.\" The appendix (F.3, F.4) shows that the optimal update frequency $k$ is task-dependent (e.g., 60 for C4 vs. 10 for Commonsense Reasoning). This slightly weakens the claim of a universally simple method, as this tuning adds back some of the experimental burden that fine-tuning aims to reduce.\n\n\nThe \"Sparsity Adaptation\" step (Sec 3.2) seems to re-prune the entire model at every $k$ steps using a sensitivity-based score ($|\\theta_i \\nabla_{\\theta_i} L|$). Is this score computed on the fly, or are gradients accumulated? This step seems computationally intensive, and its cost relative to the \"Drop\" and \"Grow\" steps is not fully clarified.", "questions": "Could the authors please quantify the training wall-clock time of SEFT compared to the baselines (LoRA*, SPP, SQFT)? How significant is the computational overhead from the \"Grow\" and \"Adapt\" steps, which require dense gradient information?\n\nIn the \"Sparsity Adaptation\" step (Sec 3.2), is the sensitivity score $s_i = |\\theta_i \\nabla_{\\theta_i} L|$ computed using the instantaneous gradient from the current batch, or is it an accumulated value (e.g., from an optimizer like Adam)? If it's instantaneous, how stable is this metric? If it's accumulated, what is the memory cost?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors experiment on LLaMA, DeepSeek, and Mistral model families, pruned with both SparseGPT and Wanda. The results demonstrate that SEFT outperforms baselines like LoRA* (prune-after-tuning), SPP, and SQFT across various benchmarks (Commonsense Reasoning, MMLU, GSM8K), with a particularly strong advantage at high sparsity levels (e.g., 70%-80%). Furthermore, SEFT shows significant efficiency gains, including up to 2x memory savings during training and up to 2.5x inference speedup on a CPU.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "SEFT cleverly adapts the idea of Dynamic Sparse Training (DST) to the novel scenario of \"fine-tuning an already-sparse model.\" The \"Drop-Grow-Adapt\" three-step cycle shown in Figure 2 is clear and logically sound. The \"Grow\" step, in particular, which allows the model to \"resurrect\" pruned weights, is key to its task-adaptive capability.\n\nhe analyses in the appendix are thorough. For example, Appendix E.1 (Table 7) proves the necessity of \"allowing resurrection of pruned weights,\" and Appendix E.2 (Fig 5a) demonstrates the superiority of the \"sensitivity-based\" pruning criterion over \"magnitude-based.\" This greatly enhances the credibility of the method's design.", "weaknesses": "SEFT introduces new hyperparameters that require tuning, most notably the sparse topology \"update frequency $k$\" and the \"drop rate.\" The appendix (F.3, F.4) shows that the optimal update frequency $k$ is task-dependent (e.g., 60 for C4 vs. 10 for Commonsense Reasoning). This slightly weakens the claim of a universally simple method, as this tuning adds back some of the experimental burden that fine-tuning aims to reduce.\n\n\nThe \"Sparsity Adaptation\" step (Sec 3.2) seems to re-prune the entire model at every $k$ steps using a sensitivity-based score ($|\\theta_i \\nabla_{\\theta_i} L|$). Is this score computed on the fly, or are gradients accumulated? This step seems computationally intensive, and its cost relative to the \"Drop\" and \"Grow\" steps is not fully clarified.", "questions": "Could the authors please quantify the training wall-clock time of SEFT compared to the baselines (LoRA*, SPP, SQFT)? How significant is the computational overhead from the \"Grow\" and \"Adapt\" steps, which require dense gradient information?\n\nIn the \"Sparsity Adaptation\" step (Sec 3.2), is the sensitivity score $s_i = |\\theta_i \\nabla_{\\theta_i} L|$ computed using the instantaneous gradient from the current batch, or is it an accumulated value (e.g., from an optimizer like Adam)? If it's instantaneous, how stable is this metric? If it's accumulated, what is the memory cost?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761603955015}, {"id": "hMHCa8IEW3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17173/Reviewer_dRk7"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "The paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method tailored for already-pruned large language models. It dynamically evolves the sparse topology during finetuning via a weight drop-and-grow strategy coupled with a sensitivity-driven pruning criterion, so task-relevant connections are restored while the overall sparsity budget is strictly preserved. Extensive experiments on multiple LLaMA, DeepSeek, and Mistral models across diverse benchmarks demonstrate that SEFT consistently outperforms prior baselines in both accuracy and memory-computation efficiency.", "review_text": "The paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method tailored for already-pruned large language models. It dynamically evolves the sparse topology during finetuning via a weight drop-and-grow strategy coupled with a sensitivity-driven pruning criterion, so task-relevant connections are restored while the overall sparsity budget is strictly preserved. Extensive experiments on multiple LLaMA, DeepSeek, and Mistral models across diverse benchmarks demonstrate that SEFT consistently outperforms prior baselines in both accuracy and memory-computation efficiency.", "strengths": "1. The proposed drop–grow–adapt procedure intuitively tackles the challenges of sparse fine-tuning: unlike LoRA, it allows high-rank weight updates while strictly preserving high sparsity, and its reduced memory footprint further boosts its potential for widespread adoption.\n2. The manuscript is clearly written and easy to follow.", "weaknesses": "W1. Continuously updating the sparsity masks adds extra computation—especially because full dense gradients are computed during the sparse-topology-evolution phase—yet the paper does not report the actual training-time overhead versus LoRA, making it difficult to gauge the true computational cost.\n\nW2. Although sparsification is promising, in practice weight quantization is often more useful, and state-of-the-art methods can reach 1–2-bit precision with minimal accuracy loss (roughly comparable to a 90 % sparsity level). Taking this into account, investigating whether the method can be combined with quantization would be an important consideration for practical applicability.", "questions": "Please clarify the concerns raised in W1 and W2.\n\nMinor\n• Typo at L885: “The results, presented in Table 7, show that …”.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Sparsity Evolution Fine-Tuning (SEFT), a new method tailored for already-pruned large language models. It dynamically evolves the sparse topology during finetuning via a weight drop-and-grow strategy coupled with a sensitivity-driven pruning criterion, so task-relevant connections are restored while the overall sparsity budget is strictly preserved. Extensive experiments on multiple LLaMA, DeepSeek, and Mistral models across diverse benchmarks demonstrate that SEFT consistently outperforms prior baselines in both accuracy and memory-computation efficiency.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The proposed drop–grow–adapt procedure intuitively tackles the challenges of sparse fine-tuning: unlike LoRA, it allows high-rank weight updates while strictly preserving high sparsity, and its reduced memory footprint further boosts its potential for widespread adoption.\n2. The manuscript is clearly written and easy to follow.", "weaknesses": "W1. Continuously updating the sparsity masks adds extra computation—especially because full dense gradients are computed during the sparse-topology-evolution phase—yet the paper does not report the actual training-time overhead versus LoRA, making it difficult to gauge the true computational cost.\n\nW2. Although sparsification is promising, in practice weight quantization is often more useful, and state-of-the-art methods can reach 1–2-bit precision with minimal accuracy loss (roughly comparable to a 90 % sparsity level). Taking this into account, investigating whether the method can be combined with quantization would be an important consideration for practical applicability.", "questions": "Please clarify the concerns raised in W1 and W2.\n\nMinor\n• Typo at L885: “The results, presented in Table 7, show that …”.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761553518174}], "openreview_url": "https://openreview.net/forum?id=qPwSoOQPrI", "arxiv_id": "2505.24037", "paper_pdf": "papers/qPwSoOQPrI.pdf", "paper_pdf_sha256": "f4bbfe734ca000370d6ce2688093827945d9112641c022ae27d4a56fe8f0f8af", "paper_pdf_bytes": 1611300, "paper_pdf_source": "openreview", "code_url": "https://github.com/QiaoXiao7282/SEFT", "code_repository": "QiaoXiao7282/SEFT", "code_commit": "3fea1d28ee6d1ac0711acb295ec9d3af2ef8e55f", "code_archive": "repos/qPwSoOQPrI.zip", "code_archive_sha256": "94a361b1e6228d54a5cd8ab835163e8d23c1643555b4ec2059c71194e95c7a98", "code_archive_bytes": 303543, "code_file_count": 111, "code_extensions": {".py": 109, ".cpp": 1, ".cu": 1}, "github_disk_usage_kb": 253, "github_languages": {"Python": 1033131, "C++": 4498, "Dockerfile": 3924, "Cuda": 3062, "Makefile": 1432}, "github_archived": false, "github_pushed_at": "2025-06-06T15:22:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/leave-it-to-the-specialist-repair-sparse-llms"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sBpYRQOrMn", "year": 2025, "status": "rejected", "title": "New Paradigm of Adversarial Training: Breaking Inherent Trade-Off between Accuracy and Robustness via Dummy Classes", "authors": ["Yanyun Wang", "Li Liu", "Zi Liang", "Qingqing Ye", "Haibo Hu"], "authorids": ["~Yanyun_Wang1", "~Li_Liu8", "~Zi_Liang1", "~Qingqing_Ye1", "~Haibo_Hu2"], "authors_source": "OpenReview API", "abstract": "Adversarial Training (AT) is recognized as one of the most effective methods to enhance the robustness of Deep Neural Networks (DNNs). However, existing AT methods suffer from an inherent trade-off between adversarial robustness and clean accuracy, which seriously hinders their real-world deployment. Previous works have studied this trade-off within the current AT paradigm, exploring various factors such as perturbation intensity, label noise and class margin. Despite these efforts, current AT methods still typically experience a reduction in clean accuracy by over 10% to date, without significant improvements in robustness compared with simple baselines like PGD-AT. This inherent trade-off raises a question: whether the current AT paradigm, which assumes to learn the corresponding benign and adversarial samples as the same class, inappropriately combines clean and robust objectives that may be essentially inconsistent. In this work, we surprisingly reveal that up to 40% of CIFAR-10 adversarial samples always fail to satisfy such an assumption across various AT methods and robust models, explicitly indicating the improvement room for the current AT paradigm. Accordingly, to relax the tension between clean and robust learning derived from this overstrict assumption, we propose a new AT paradigm by introducing an additional dummy class for each original class, aiming to accommodate the hard adversarial samples with shifted distribution after perturbation. The robustness w.r.t. these adversarial samples can be achieved by runtime recovery from the predicted dummy classes to their corresponding original ones, eliminating the compromise with clean learning. Building on this new paradigm, we propose a novel plug-and-play AT technology named DUmmy Classes-based Adversarial Training (DUCAT). Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that the DUCAT concurrently improves clean accuracy and adversarial robustness compared with state-of-the-art benchmarks, effectively releasing the existing inherent trade-off. The code is available at https://anonymous.4open.science/r/DUCAT.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "629T09CwOj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6428/Reviewer_9r4X"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "Pointing out that the existing AT methods enforce the model to learn benign and adversarial samples are belong to the same class lead to the trade-off between clean and adversarial accuracy, this paper introduced a new AT paradigm introducing additional dummy classes for certain adversarial samples that differ in distribution from the original ones to relax the assumption of existing AT methods. Rather than strictly separating the class for benign and adversarial examples, this paper construct soft labels to explicitly bridge corresponding original and dummy classes as the suboptimal alternative target of each other, leading the model to learn separation between them. Experimental results across different baselines and tasks show effectiveness of the proposed method.", "review_text": "Pointing out that the existing AT methods enforce the model to learn benign and adversarial samples are belong to the same class lead to the trade-off between clean and adversarial accuracy, this paper introduced a new AT paradigm introducing additional dummy classes for certain adversarial samples that differ in distribution from the original ones to relax the assumption of existing AT methods. Rather than strictly separating the class for benign and adversarial examples, this paper construct soft labels to explicitly bridge corresponding original and dummy classes as the suboptimal alternative target of each other, leading the model to learn separation between them. Experimental results across different baselines and tasks show effectiveness of the proposed method.", "strengths": "- This paper has strong motivation of tackling trade-off between clean and robust accuracy of existing AT methods, demonstrating existing methods all fail to some adversaries that proves the common deficiency of existing works' learning objectives.\n- This paper is well written and easy to follow.", "weaknesses": "- The proposed method seem to be incremental compared to TRADES.\n   - Although the proposed method shows impressive improvement on PGD-AT, MART, and Consistency-AT, compared to the most important baseline that addressed the trade-off issue for the first time, TRADES, DUCAT sacrificed clean accuracy more than the improved robustness as shown in Table 1, especially on CIFAR-10. Also, the improvements achieved on CIFAR-100 and Tiny-ImageNet from the proposed method (DUCAT) are marginal while it requires additional computational resources compared to TRADES as shown in Figure 6. Based on this observation, it would be better to provide more supporting arguments on DUCAT compared to TRADES.\n- Since this paper insists that if the adversarial examples generated from one model prove to be effective, they are might also effective and transferred to the different models. In responses, it would be beneficial to show the adversarial and standard accuracy under the transfer attack scenario to further strengthen the proposed method. The experiments can be done in the Figure 3 showing transfer attack scenarios across different training algorithms, where adversarial examples generated from ResNet18 model trained on DUCAT can be transferred to baseline methods, or vice versa. Rather than bar plot, it would be clearer to visualize the performance on CIFAR-10 and CIFAR-100 in the format of table.", "questions": "- In figure 2, there is different tendency in ratio between all failure/all worked cases for different class, why different classes show different tendency? It would be helpful to provide a more detailed analysis of the class-specific differences along with the dataset name and explain which factor of each class leads to different distributions for the results. Please examine properties of the different classes, for example, they could visual complexity, number of entities within example, what are the main differences captured by models between adversarial examples and clean examples using PCA analysis, which might affect to the tendencies in failure/success rates.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Pointing out that the existing AT methods enforce the model to learn benign and adversarial samples are belong to the same class lead to the trade-off between clean and adversarial accuracy, this paper introduced a new AT paradigm introducing additional dummy classes for certain adversarial samples that differ in distribution from the original ones to relax the assumption of existing AT methods. Rather than strictly separating the class for benign and adversarial examples, this paper construct soft labels to explicitly bridge corresponding original and dummy classes as the suboptimal alternative target of each other, leading the model to learn separation between them. Experimental results across different baselines and tasks show effectiveness of the proposed method.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This paper has strong motivation of tackling trade-off between clean and robust accuracy of existing AT methods, demonstrating existing methods all fail to some adversaries that proves the common deficiency of existing works' learning objectives.\n- This paper is well written and easy to follow.", "weaknesses": "- The proposed method seem to be incremental compared to TRADES.\n   - Although the proposed method shows impressive improvement on PGD-AT, MART, and Consistency-AT, compared to the most important baseline that addressed the trade-off issue for the first time, TRADES, DUCAT sacrificed clean accuracy more than the improved robustness as shown in Table 1, especially on CIFAR-10. Also, the improvements achieved on CIFAR-100 and Tiny-ImageNet from the proposed method (DUCAT) are marginal while it requires additional computational resources compared to TRADES as shown in Figure 6. Based on this observation, it would be better to provide more supporting arguments on DUCAT compared to TRADES.\n- Since this paper insists that if the adversarial examples generated from one model prove to be effective, they are might also effective and transferred to the different models. In responses, it would be beneficial to show the adversarial and standard accuracy under the transfer attack scenario to further strengthen the proposed method. The experiments can be done in the Figure 3 showing transfer attack scenarios across different training algorithms, where adversarial examples generated from ResNet18 model trained on DUCAT can be transferred to baseline methods, or vice versa. Rather than bar plot, it would be clearer to visualize the performance on CIFAR-10 and CIFAR-100 in the format of table.", "questions": "- In figure 2, there is different tendency in ratio between all failure/all worked cases for different class, why different classes show different tendency? It would be helpful to provide a more detailed analysis of the class-specific differences along with the dataset name and explain which factor of each class leads to different distributions for the results. Please examine properties of the different classes, for example, they could visual complexity, number of entities within example, what are the main differences captured by models between adversarial examples and clean examples using PCA analysis, which might affect to the tendencies in failure/success rates.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730672022679}, {"id": "XvEMnpS8m6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6428/Reviewer_r1aS"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "In this paper, the authors address the ongoing trade-off issue in adversarial training-based methods. To mitigate this problem, they introduce a dummy class and soft labeling techniques into the existing adversarial training framework. A novel adversarial training loss function is designed to improve the model’s trade-off between accuracy and robustness. The proposed method is evaluated on three datasets and compared against four state-of-the-art adversarial training approaches. Additionally, the authors analyze the impact of various hyperparameters on model performance.", "review_text": "In this paper, the authors address the ongoing trade-off issue in adversarial training-based methods. To mitigate this problem, they introduce a dummy class and soft labeling techniques into the existing adversarial training framework. A novel adversarial training loss function is designed to improve the model’s trade-off between accuracy and robustness. The proposed method is evaluated on three datasets and compared against four state-of-the-art adversarial training approaches. Additionally, the authors analyze the impact of various hyperparameters on model performance.", "strengths": "1. Addresses a Key Limitation in Adversarial Training: The paper introduces a new adversarial training paradigm, DUCAT, which aims to break the inherent trade-off between clean accuracy and robustness, a well-known limitation in existing methods\n\n2. Plug-and-Play Applicability: DUCAT is presented as a flexible method that can be integrated with existing adversarial training frameworks, potentially making it easier to adopt in practical applications without major redesigns.", "weaknesses": "1. Lack of In-depth Analysis of Method Efficacy: The proposed DUCAT method primarily leverages the concepts of dummy classes and label smoothing. However, previous studies (e.g., [1], [2], [3], [4]) have explored the effectiveness of either dummy classes or label smoothing to enhance robustness or accuracy. A more thorough analysis is needed to clarify the unique contribution of DUCAT in comparison to these established works.\n\n   [1] Label Smoothing and Logit Squeezing: A Replacement for Adversarial Training? https://arxiv.org/abs/1910.11585\n\n   [2] Frustratingly Easy Model Generalization by Dummy Risk Minimization https://arxiv.org/pdf/2308.02287\n\n   [3] When Does Label Smoothing Help?\n\n    [4] MIXUP INFERENCE: Better Exploiting Mixup to Defend Adversarial Attacks https://openreview.net/pdf?id=ByxtC2VtPB\n\n2. Limited Empirical Evidence: The experimental results, as seen in Table I, show only marginal improvements of the proposed DUCAT method over TRADES. Further analysis is recommended to understand the limited performance gains.", "questions": "1. How does the proposed DUCAT method compare in depth to the above-mentioned related works? Could you elaborate on the unique aspects of its effectiveness?\n\n2. What might be the reasons behind the limited performance improvement of DUCAT over TRADES?\n\n3. Could you provide a performance comparison between the proposed DUCAT method and the technique outlined in [4]?\n\n     [4] MIXUP INFERENCE: BETTER EXPLOITING MIXUP TO DEFEND ADVERSARIAL ATTACKS https://openreview.net/pdf?id=ByxtC2VtPB", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors address the ongoing trade-off issue in adversarial training-based methods. To mitigate this problem, they introduce a dummy class and soft labeling techniques into the existing adversarial training framework. A novel adversarial training loss function is designed to improve the model’s trade-off between accuracy and robustness. The proposed method is evaluated on three datasets and compared against four state-of-the-art adversarial training approaches. Additionally, the authors analyze the impact of various hyperparameters on model performance.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Addresses a Key Limitation in Adversarial Training: The paper introduces a new adversarial training paradigm, DUCAT, which aims to break the inherent trade-off between clean accuracy and robustness, a well-known limitation in existing methods\n\n2. Plug-and-Play Applicability: DUCAT is presented as a flexible method that can be integrated with existing adversarial training frameworks, potentially making it easier to adopt in practical applications without major redesigns.", "weaknesses": "1. Lack of In-depth Analysis of Method Efficacy: The proposed DUCAT method primarily leverages the concepts of dummy classes and label smoothing. However, previous studies (e.g., [1], [2], [3], [4]) have explored the effectiveness of either dummy classes or label smoothing to enhance robustness or accuracy. A more thorough analysis is needed to clarify the unique contribution of DUCAT in comparison to these established works.\n\n   [1] Label Smoothing and Logit Squeezing: A Replacement for Adversarial Training? https://arxiv.org/abs/1910.11585\n\n   [2] Frustratingly Easy Model Generalization by Dummy Risk Minimization https://arxiv.org/pdf/2308.02287\n\n   [3] When Does Label Smoothing Help?\n\n    [4] MIXUP INFERENCE: Better Exploiting Mixup to Defend Adversarial Attacks https://openreview.net/pdf?id=ByxtC2VtPB\n\n2. Limited Empirical Evidence: The experimental results, as seen in Table I, show only marginal improvements of the proposed DUCAT method over TRADES. Further analysis is recommended to understand the limited performance gains.", "questions": "1. How does the proposed DUCAT method compare in depth to the above-mentioned related works? Could you elaborate on the unique aspects of its effectiveness?\n\n2. What might be the reasons behind the limited performance improvement of DUCAT over TRADES?\n\n3. Could you provide a performance comparison between the proposed DUCAT method and the technique outlined in [4]?\n\n     [4] MIXUP INFERENCE: BETTER EXPLOITING MIXUP TO DEFEND ADVERSARIAL ATTACKS https://openreview.net/pdf?id=ByxtC2VtPB", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730646606836}, {"id": "Rthi5lZWVy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6428/Reviewer_QPXH"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposed DUCAT, a new adversarial training paradigm that leverages dummy classes to achieve a better trade-off between adversarial robustness and clean accuracy. The introduced dummy classes relax the overstrict label assignment. As a plug-and-play method, DUCAT can be applied to existing adversarial training pipelines to help further improve performance. The effectiveness of DUCAT is demonstrated by extensive experiments across different datasets.", "review_text": "This paper proposed DUCAT, a new adversarial training paradigm that leverages dummy classes to achieve a better trade-off between adversarial robustness and clean accuracy. The introduced dummy classes relax the overstrict label assignment. As a plug-and-play method, DUCAT can be applied to existing adversarial training pipelines to help further improve performance. The effectiveness of DUCAT is demonstrated by extensive experiments across different datasets.", "strengths": "- The intuition behind introducing dummy classes is making sense to me.\n- The paper is well-written and easy to follow.\n- The proposed methods are evaluated on different datasets, baseline methods and network architectures.", "weaknesses": "- I agree with the design of DUCAT and its effectiveness when applied to existing adversarial training methods, but a non-negligible fact is that the performance is still far away from methods like [1, 2, 3], which obtain higher clean and robust accuracy. Although DUCAT is a good plug-and-play method for adversarial training, I think the value is still limited if it cannot bring those AT-based baseline methods to a closer level to more advanced methods. \n- Some notations are not clearly explained. Please see the questions below. \n\n[1] Wang, Z., Pang, T., Du, C., Lin, M., Liu, W., & Yan, S. (2023, July). Better diffusion models further improve adversarial training. In International Conference on Machine Learning (pp. 36246-36263). PMLR.\n\n[2] Rebuffi, S. A., Gowal, S., Calian, D. A., Stimberg, F., Wiles, O., & Mann, T. (2021). Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946.\n\n[3] Sehwag, V., Mahloujifar, S., Handina, T., Dai, S., Xiang, C., Chiang, M., & Mittal, P. (2022). ROBUST LEARNING MEETS GENERATIVE MODELS: CAN PROXY DISTRIBUTIONS IMPROVE ADVERSARIAL ROBUSTNESS?. In 10th International Conference on Learning Representations, ICLR 2022.", "questions": "- What is $\\mathcal{P}(x_i)$ in Eq. 2? The distribution of adversarial examples of $x_i$?\n- What are $\\dot{\\mathbf{y}}$ and $\\ddot{\\mathbf{y}}$ in Eq. 4? Do they stand for probabilities of $C$ real classes and $C$ dummy classes respectively? Are you actually doing a decomposition $\\mathbf{y} = \\dot{\\mathbf{y}} + \\ddot{\\mathbf{y}}$?\n- What is $h_\\theta^{Dummy}(\\cdot)$? Another neural network? If I understand Fig. 1 correctly, the output representation of the second last layer is unchanged, and you just enlarge the dimensionality of the last layer? Can you clarify if you are using another neural network or just enlarging the last layer?\n- How do you set the dummy classes for each example? \n- I am confused about the word \"SOTAs\" used in Sec 3.2. If it means those aiming at releasing the trade-off between accuracy and robustness, then we can find a bunch of methods overperforming them on RobustBench[4]. If it specifically refers to AT-based methods, then like what I mentioned in the weaknesses above, the value of DUCAT is limited as we can find more advanced methods. In fact, methods like [1] also involve adversarial training. If DUCAT is proved able to improve the performance of some of those methods, its value will be underscored.\n\n[4] https://robustbench.github.io/", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed DUCAT, a new adversarial training paradigm that leverages dummy classes to achieve a better trade-off between adversarial robustness and clean accuracy. The introduced dummy classes relax the overstrict label assignment. As a plug-and-play method, DUCAT can be applied to existing adversarial training pipelines to help further improve performance. The effectiveness of DUCAT is demonstrated by extensive experiments across different datasets.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The intuition behind introducing dummy classes is making sense to me.\n- The paper is well-written and easy to follow.\n- The proposed methods are evaluated on different datasets, baseline methods and network architectures.", "weaknesses": "- I agree with the design of DUCAT and its effectiveness when applied to existing adversarial training methods, but a non-negligible fact is that the performance is still far away from methods like [1, 2, 3], which obtain higher clean and robust accuracy. Although DUCAT is a good plug-and-play method for adversarial training, I think the value is still limited if it cannot bring those AT-based baseline methods to a closer level to more advanced methods. \n- Some notations are not clearly explained. Please see the questions below. \n\n[1] Wang, Z., Pang, T., Du, C., Lin, M., Liu, W., & Yan, S. (2023, July). Better diffusion models further improve adversarial training. In International Conference on Machine Learning (pp. 36246-36263). PMLR.\n\n[2] Rebuffi, S. A., Gowal, S., Calian, D. A., Stimberg, F., Wiles, O., & Mann, T. (2021). Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946.\n\n[3] Sehwag, V., Mahloujifar, S., Handina, T., Dai, S., Xiang, C., Chiang, M., & Mittal, P. (2022). ROBUST LEARNING MEETS GENERATIVE MODELS: CAN PROXY DISTRIBUTIONS IMPROVE ADVERSARIAL ROBUSTNESS?. In 10th International Conference on Learning Representations, ICLR 2022.", "questions": "- What is $\\mathcal{P}(x_i)$ in Eq. 2? The distribution of adversarial examples of $x_i$?\n- What are $\\dot{\\mathbf{y}}$ and $\\ddot{\\mathbf{y}}$ in Eq. 4? Do they stand for probabilities of $C$ real classes and $C$ dummy classes respectively? Are you actually doing a decomposition $\\mathbf{y} = \\dot{\\mathbf{y}} + \\ddot{\\mathbf{y}}$?\n- What is $h_\\theta^{Dummy}(\\cdot)$? Another neural network? If I understand Fig. 1 correctly, the output representation of the second last layer is unchanged, and you just enlarge the dimensionality of the last layer? Can you clarify if you are using another neural network or just enlarging the last layer?\n- How do you set the dummy classes for each example? \n- I am confused about the word \"SOTAs\" used in Sec 3.2. If it means those aiming at releasing the trade-off between accuracy and robustness, then we can find a bunch of methods overperforming them on RobustBench[4]. If it specifically refers to AT-based methods, then like what I mentioned in the weaknesses above, the value of DUCAT is limited as we can find more advanced methods. In fact, methods like [1] also involve adversarial training. If DUCAT is proved able to improve the performance of some of those methods, its value will be underscored.\n\n[4] https://robustbench.github.io/", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730593168852}, {"id": "u8PS5GNvqM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6428/Reviewer_dTK9"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a new adversarial training method (DUmmy Classes-based Adversarial Training, DUCAT). DUCAT relaxes the strong assumption in adversarial training that requires clean samples and adversarial samples to be classified into the same category via introducing dummy classes. DUCAT can improve the model's adversarial robustness while achieving higher clean accuracy than existing AT methods such as PGD, TRADS, MART, and Consistency-AT.", "review_text": "This paper proposes a new adversarial training method (DUmmy Classes-based Adversarial Training, DUCAT). DUCAT relaxes the strong assumption in adversarial training that requires clean samples and adversarial samples to be classified into the same category via introducing dummy classes. DUCAT can improve the model's adversarial robustness while achieving higher clean accuracy than existing AT methods such as PGD, TRADS, MART, and Consistency-AT.", "strengths": "This paper reveals that always-failed samples widely exist in existing AT methods and points out that this is caused by the overly strong assumption of AT on clean samples and adversarial samples. The corresponding dummy class method can improve the robustness of the model and, compared with existing adversarial training methods, shows an improvement in clean accuracy.\n\nThe organization and writing quality of the paper are both good.", "weaknesses": "1. The paper's title and the main claim is somewhat misleading (Breaking Inherent Trade-Off between Accuracy and Robustness). I don't think the proposed method indeed breaks the trade-off between accuracy and robustness. As shown in Table 1, the clean accuracy obtained via DUCAT is still significantly lower than standard training (for example, ResNet-18 can easily achieve accuracy higher than 93% under standard training). I suggest authors rephrase their claim to more accurately rephrase their claim to more accurately reflect the improvement over existing adversarial training methods, and also provide the clean accuracy of standard training for reference.\n2. The use of soft labels instead of hard labels to enhance adversarial training performance somewhat limits the novelty of this paper; please refer to [1, 2, 3, 4], where [1] discusses injecting more learned smoothing during adversarial training, [2] proposes generating soft labels instead of hard one-hot labels to achieve more robust models, and so forth. Additionally, I believe that the one-versus-the-rest loss proposed in [4] already shares a similarity with the ideas presented in this paper.\n\n\n[1] Robust Overfitting may be mitigated by properly learned smoothening (ICLR2021)\n\n[2] Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective (NeurIPS2023)\n\n[3] Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR2024)\n\n[4] Switching One-Versus-the-Rest Loss to Increase the Margin of Logits for Adversarial Robustness (preprint)", "questions": "I believe the authors should reconsider the terms in which they claim the contributions of the article. Instead of stating that it has broken the inherent trade-off between accuracy and robustness, this paper should more accurately be described as reducing the loss of clean accuracy while enhancing adversarial robustness.\n\nAlso, I would like the authors to elaborate on the differences and major innovations in their approach and related work (please refer to the weaknesses session 2).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new adversarial training method (DUmmy Classes-based Adversarial Training, DUCAT). DUCAT relaxes the strong assumption in adversarial training that requires clean samples and adversarial samples to be classified into the same category via introducing dummy classes. DUCAT can improve the model's adversarial robustness while achieving higher clean accuracy than existing AT methods such as PGD, TRADS, MART, and Consistency-AT.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "This paper reveals that always-failed samples widely exist in existing AT methods and points out that this is caused by the overly strong assumption of AT on clean samples and adversarial samples. The corresponding dummy class method can improve the robustness of the model and, compared with existing adversarial training methods, shows an improvement in clean accuracy.\n\nThe organization and writing quality of the paper are both good.", "weaknesses": "1. The paper's title and the main claim is somewhat misleading (Breaking Inherent Trade-Off between Accuracy and Robustness). I don't think the proposed method indeed breaks the trade-off between accuracy and robustness. As shown in Table 1, the clean accuracy obtained via DUCAT is still significantly lower than standard training (for example, ResNet-18 can easily achieve accuracy higher than 93% under standard training). I suggest authors rephrase their claim to more accurately rephrase their claim to more accurately reflect the improvement over existing adversarial training methods, and also provide the clean accuracy of standard training for reference.\n2. The use of soft labels instead of hard labels to enhance adversarial training performance somewhat limits the novelty of this paper; please refer to [1, 2, 3, 4], where [1] discusses injecting more learned smoothing during adversarial training, [2] proposes generating soft labels instead of hard one-hot labels to achieve more robust models, and so forth. Additionally, I believe that the one-versus-the-rest loss proposed in [4] already shares a similarity with the ideas presented in this paper.\n\n\n[1] Robust Overfitting may be mitigated by properly learned smoothening (ICLR2021)\n\n[2] Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective (NeurIPS2023)\n\n[3] Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR2024)\n\n[4] Switching One-Versus-the-Rest Loss to Increase the Margin of Logits for Adversarial Robustness (preprint)", "questions": "I believe the authors should reconsider the terms in which they claim the contributions of the article. Instead of stating that it has broken the inherent trade-off between accuracy and robustness, this paper should more accurately be described as reducing the loss of clean accuracy while enhancing adversarial robustness.\n\nAlso, I would like the authors to elaborate on the differences and major innovations in their approach and related work (please refer to the weaknesses session 2).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730127742944}], "openreview_url": "https://openreview.net/forum?id=sBpYRQOrMn", "arxiv_id": "2410.12671", "paper_pdf": "papers/sBpYRQOrMn.pdf", "paper_pdf_sha256": "bca40401ca5de6221f3286d08b21a6307fbbd5e19b5c0a79aa141369b7842eeb", "paper_pdf_bytes": 4853463, "paper_pdf_source": "openreview", "code_url": "https://github.com/FlaAI/DUCAT", "code_repository": "FlaAI/DUCAT", "code_commit": "6bb27b31cbecfa6ce04c7d559203b54732eca3c9", "code_archive": "repos/sBpYRQOrMn.zip", "code_archive_sha256": "579780961626ef8adbc81a367e253950d3e67fc692c9b6f19f114d3f5f5d9447", "code_archive_bytes": 34170, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 36, "github_languages": {"Python": 86427}, "github_archived": false, "github_pushed_at": "2024-10-03T13:57:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/new-paradigm-of-adversarial-training-breaking"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7Scc7Nl7lg", "year": 2024, "status": "rejected", "title": "Revealing Vision-Language Integration in the Brain with Multimodal Networks", "authors": ["Vighnesh Subramaniam", "Colin Conwell", "Christopher Wang", "Gabriel Kreiman", "Boris Katz", "Ignacio Cases", "Andrei Barbu"], "authorids": ["~Vighnesh_Subramaniam1", "~Colin_Conwell1", "~Christopher_Wang1", "~Gabriel_Kreiman1", "~Boris_Katz1", "~Ignacio_Cases2", "~Andrei_Barbu3"], "authors_source": "OpenReview API", "abstract": "We use multimodal deep neural networks to identify sites of multimodal integration in the human brain. These are regions where a multimodal language-vision model is better at predicting neural recordings (stereoelectroencephalography, SEEG) than either a unimodal language, unimodal vision model, or a linearly-integrated language-vision model. We use a wide range of state-of-the-art models spanning different architectures including Transformers and CNNs (ALBEF, BLIP, Flava, ConvNeXt, BEIT, SIMCLR, CLIP, SLIP) with different multimodal integration approaches to model the SEEG signal while subjects watched movies. As a key enabling step, we first demonstrate that the approach has the resolution to distinguish trained from randomly-initialized models for both language and vision; the inability to do so would fundamentally hinder further analysis. We show that trained models systematically outperform randomly initialized models in their ability to predict the SEEG signal. We then compare unimodal and multimodal models against one another. A key contribution is standardizing the methodology for doing so while carefully avoiding statistical artifacts. Since models all have different architectures, number of parameters, and training sets which can obscure the results, we then carry out a test between two controlled models: SLIP-Combo and SLIP-SimCLR which keep all of these attributes the same aside from multimodal input. Using this method, we identify neural sites (on average 141 out of 1090 total sites or 12.94\\%) and brain regions where multimodal integration is occurring. We find numerous new sites of multimodal integration, many of which lie around the temporoparietal junction, long theorized to be a hub of multimodal integration.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "kOdmGYipr4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6104/Reviewer_JsYq"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "I have reviewed this work previously and the current version does not address my previous major concerns so I will repeat my points in this review in hopes that the authors can address them this time.\n\nThis work investigates the ability of multi-modal neural networks to align with multi-modal ECoG brain recordings, acquired while 7 epileptic children were watching movies. The authors aim to use a contrast between multi-modal and uni-modal models to reveal which locations in the brain relate to integration from multiple modalities. A large number of models are tested (7 multi-modal and 5 uni-modal). This work finds that about 13% of the tested neural sites are better predicted by multi-modal models.", "review_text": "I have reviewed this work previously and the current version does not address my previous major concerns so I will repeat my points in this review in hopes that the authors can address them this time.\n\nThis work investigates the ability of multi-modal neural networks to align with multi-modal ECoG brain recordings, acquired while 7 epileptic children were watching movies. The authors aim to use a contrast between multi-modal and uni-modal models to reveal which locations in the brain relate to integration from multiple modalities. A large number of models are tested (7 multi-modal and 5 uni-modal). This work finds that about 13% of the tested neural sites are better predicted by multi-modal models.", "strengths": "- Using multi-modal brain recordings from movies\n- Using ECoG for high spatial and temporal precision\n- Checking many models (12, 2 models in depth)\n- An additional analysis that goes more in depth because, as the authors realize, a difference in brain predictivity in a direct comparison between a multimodal and a unimodal model can be due to the many possible differences between the models", "weaknesses": "1. The biggest weakness is the central claim that the proposed analyses of multi-modal models can localize vision-language **integration** in the brain. Can the authors please define what they mean by vision-language integration? Even on the model side, it is an open question to what degree multi-modal models actually integrate information from multiple modalities as opposed to increasing the alignment between individual modalities (Liang et al. 2022 https://arxiv.org/pdf/2209.03430.pdf). It is not clear whether the results that are observed are due to vision-language integration or whether they are due to improved representations of the language-only or vision-only modality. For example, in Fig 4, all regions that are identified as multimodal (e.g. marked with a green star), are canonical language regions. How can the authors disentangle the effect of integration of modalities vs the effect of improving language-only information in the model? For instance, let's do a thought experiment and apply the authors' methods to study different layers of a language-only encoder: I predict that these results will be similar to what has been shown before which is that some layers in the language model predict exactly the same regions they call multimodal substantially better than other layers (see Jain and Huth, 2018 NeurIPS; Toneva and Wehbe, 2019 NeurIPS for some of the earlier work showing this). That clearly is not due to multimodality though because the input is only language.\n\n2. The presentation can be much improved: there is very little discussion of what the different models that are used are and how they are trained. This is key to understand the contributions of this work. I suggest the authors include a table of all models and model variations used with clearly marked information about what modality was used to train the model and what modality is used to evaluate the model (e.g. even if the authors are using only a vision encoder at inference time, if the vision encoder was jointly trained with a language encoder, this should be noted as this may make a difference in the representations)\n3. The work is still not positioned well in the current literature on multi-modal modeling of brain recordings, and it’s not clear what the novelty here is for people who are unfamiliar with this area. The authors should discuss the work of Oota et al. 2022 COLING and Wang et al. 2022 bioRxiv https://www.biorxiv.org/content/10.1101/2022.09.27.508760v1. \n4. The contribution to an ML audience is not very clear. I believe this work will be better suited to be evaluated by neuroscientists, since the claimed contributions are on the neuroscience side, and will also be more appreciated at a neuroscience venue.", "questions": "See Weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "I have reviewed this work previously and the current version does not address my previous major concerns so I will repeat my points in this review in hopes that the authors can address them this time.\n\nThis work investigates the ability of multi-modal neural networks to align with multi-modal ECoG brain recordings, acquired while 7 epileptic children were watching movies. The authors aim to use a contrast between multi-modal and uni-modal models to reveal which locations in the brain relate to integration from multiple modalities. A large number of models are tested (7 multi-modal and 5 uni-modal). This work finds that about 13% of the tested neural sites are better predicted by multi-modal models.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- Using multi-modal brain recordings from movies\n- Using ECoG for high spatial and temporal precision\n- Checking many models (12, 2 models in depth)\n- An additional analysis that goes more in depth because, as the authors realize, a difference in brain predictivity in a direct comparison between a multimodal and a unimodal model can be due to the many possible differences between the models", "weaknesses": "1. The biggest weakness is the central claim that the proposed analyses of multi-modal models can localize vision-language **integration** in the brain. Can the authors please define what they mean by vision-language integration? Even on the model side, it is an open question to what degree multi-modal models actually integrate information from multiple modalities as opposed to increasing the alignment between individual modalities (Liang et al. 2022 https://arxiv.org/pdf/2209.03430.pdf). It is not clear whether the results that are observed are due to vision-language integration or whether they are due to improved representations of the language-only or vision-only modality. For example, in Fig 4, all regions that are identified as multimodal (e.g. marked with a green star), are canonical language regions. How can the authors disentangle the effect of integration of modalities vs the effect of improving language-only information in the model? For instance, let's do a thought experiment and apply the authors' methods to study different layers of a language-only encoder: I predict that these results will be similar to what has been shown before which is that some layers in the language model predict exactly the same regions they call multimodal substantially better than other layers (see Jain and Huth, 2018 NeurIPS; Toneva and Wehbe, 2019 NeurIPS for some of the earlier work showing this). That clearly is not due to multimodality though because the input is only language.\n\n2. The presentation can be much improved: there is very little discussion of what the different models that are used are and how they are trained. This is key to understand the contributions of this work. I suggest the authors include a table of all models and model variations used with clearly marked information about what modality was used to train the model and what modality is used to evaluate the model (e.g. even if the authors are using only a vision encoder at inference time, if the vision encoder was jointly trained with a language encoder, this should be noted as this may make a difference in the representations)\n3. The work is still not positioned well in the current literature on multi-modal modeling of brain recordings, and it’s not clear what the novelty here is for people who are unfamiliar with this area. The authors should discuss the work of Oota et al. 2022 COLING and Wang et al. 2022 bioRxiv https://www.biorxiv.org/content/10.1101/2022.09.27.508760v1. \n4. The contribution to an ML audience is not very clear. I believe this work will be better suited to be evaluated by neuroscientists, since the claimed contributions are on the neuroscience side, and will also be more appreciated at a neuroscience venue.", "questions": "See Weaknesses above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699009183914}, {"id": "zegMs2LTYY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6104/Reviewer_9iyA"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper aims to identify neural sites where multimodal integration is occuring in the brain. To achieve that, authors evaluate in which regions  multimodal (vision + language) models are better than unimodal models in predicting neural recordings (SEEG). \n\nUsing this method, the authors identify 141 out of 1090 total sites where multimodal integration is happening.", "review_text": "This paper aims to identify neural sites where multimodal integration is occuring in the brain. To achieve that, authors evaluate in which regions  multimodal (vision + language) models are better than unimodal models in predicting neural recordings (SEEG). \n\nUsing this method, the authors identify 141 out of 1090 total sites where multimodal integration is happening.", "strengths": "1. The methodology to identify multimodal sites is well described and quite comprehensive (trained vs. random, multimodal vs. unimodal, SLIP combo vs SLIP-SimCLR). The release of code will enable future work investigating similar questions with other modalities or other type of brain recordings\n2. The paper is easy to follow. This is due to clear writing and presentation of methods and results. \n3. Statistical tests and confidence intervals.  \n4. Multimodality test results on section 4.2. I appreciate 5 tests of multimodality reported in the results and how each test  filters out possible confounds.", "weaknesses": "1. One test that can also be included is to randomly input one of the modalities in a multimodal model and then compare with predictions using actual multimodal inputs. This test can reveal the importance of multimodal information avoiding the confounds due to architecture, parameters, training set etc. \n2. Did the authors perform SLIP-CLIP vs SLIP-SimCLR vs SLIP-Combo comparison? Because SLIP-CLIP is also multimodal I am curious what was the motivation for only showing SLIP-CLIP vs Combo results in Figure 4\n3. It is not clear to me why language alignment and vision alignment event structures leads to difference in results. I would like to read author’s explanation on why results depend on how event structures and if this is a limitation of this approach.", "questions": "1. Is the dataset also part of this paper or is it from an already published paper? If yes then this is an additional contribution of this paper and should be emphasized.\n2. Will this dataset be publicly released? ( if not released already )", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to identify neural sites where multimodal integration is occuring in the brain. To achieve that, authors evaluate in which regions  multimodal (vision + language) models are better than unimodal models in predicting neural recordings (SEEG). \n\nUsing this method, the authors identify 141 out of 1090 total sites where multimodal integration is happening.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The methodology to identify multimodal sites is well described and quite comprehensive (trained vs. random, multimodal vs. unimodal, SLIP combo vs SLIP-SimCLR). The release of code will enable future work investigating similar questions with other modalities or other type of brain recordings\n2. The paper is easy to follow. This is due to clear writing and presentation of methods and results. \n3. Statistical tests and confidence intervals.  \n4. Multimodality test results on section 4.2. I appreciate 5 tests of multimodality reported in the results and how each test  filters out possible confounds.", "weaknesses": "1. One test that can also be included is to randomly input one of the modalities in a multimodal model and then compare with predictions using actual multimodal inputs. This test can reveal the importance of multimodal information avoiding the confounds due to architecture, parameters, training set etc. \n2. Did the authors perform SLIP-CLIP vs SLIP-SimCLR vs SLIP-Combo comparison? Because SLIP-CLIP is also multimodal I am curious what was the motivation for only showing SLIP-CLIP vs Combo results in Figure 4\n3. It is not clear to me why language alignment and vision alignment event structures leads to difference in results. I would like to read author’s explanation on why results depend on how event structures and if this is a limitation of this approach.", "questions": "1. Is the dataset also part of this paper or is it from an already published paper? If yes then this is an additional contribution of this paper and should be emphasized.\n2. Will this dataset be publicly released? ( if not released already )", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698918250134}, {"id": "9VUxAF3vWT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6104/Reviewer_gGN5"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "There is a multitude of papers training decoders from latent representations of DNN models onto brain activity. The goal is to reconstruct the brain activity, which, if successful, would indicate that there is something brain-like in the representations that are formed by artificial learning models. One cohort of such papers focus on models of vision (and predicting activity of visual areas of the human brain), while another on language models (and predicting the activity of language areas in the brain). In this paper the authors ask what if we take models that work on vision and language simultaneously - would the representations and activations of those models be more predictive of brain activity? And if so - in which regions?\n\nAn answer to this question might help us understand where in the brain are the areas that integrate different sensory modalities together, or at least work on several modalities at the same time.\n\nThe authors then take 14 models and use their representation to predict each electrode's recordings. The idea is that if a multimodal model's representation is significantly more useful for predicting the brain activity, then this model's representations are closer to what is happening in the brain, and thus can be thought of as evidence for that electrode's area being involved in multimodal processing of information.\n\nThe authors do find several such electrodes, but the number of those electrodes is not sufficiently high to draw strong conclusions (at least this was my impression from reading the paper, please correct me if I am wrong).\n\nOverall this work presents and very cool idea, which is well-executed and logically reported, however due to the lack of data (I suspect) there is just not enough ground to claim definitive findings.", "review_text": "There is a multitude of papers training decoders from latent representations of DNN models onto brain activity. The goal is to reconstruct the brain activity, which, if successful, would indicate that there is something brain-like in the representations that are formed by artificial learning models. One cohort of such papers focus on models of vision (and predicting activity of visual areas of the human brain), while another on language models (and predicting the activity of language areas in the brain). In this paper the authors ask what if we take models that work on vision and language simultaneously - would the representations and activations of those models be more predictive of brain activity? And if so - in which regions?\n\nAn answer to this question might help us understand where in the brain are the areas that integrate different sensory modalities together, or at least work on several modalities at the same time.\n\nThe authors then take 14 models and use their representation to predict each electrode's recordings. The idea is that if a multimodal model's representation is significantly more useful for predicting the brain activity, then this model's representations are closer to what is happening in the brain, and thus can be thought of as evidence for that electrode's area being involved in multimodal processing of information.\n\nThe authors do find several such electrodes, but the number of those electrodes is not sufficiently high to draw strong conclusions (at least this was my impression from reading the paper, please correct me if I am wrong).\n\nOverall this work presents and very cool idea, which is well-executed and logically reported, however due to the lack of data (I suspect) there is just not enough ground to claim definitive findings.", "strengths": "Indeed an important and timely question, and, to the best of my knowledge, this work is first to explore this topic. I really like the idea and the research question.\n\nI very much support the emphasis you made on the fact that before any analysis is done we should confirm that there is either a strong difference in signal reconstructive power between trained and untrained networks, or a strong decoding ability -- we need to have a confirmation that the signal is indeed there before we analyse anything related to it.", "weaknesses": "My overall impression can be summarised as follows: while the first half of the paper is great and explains the idea and motivation really well, creating a rightful sense of expectation of the result, the section on the results somewhat comes short of delivering the findings with a bang. After reading the first half I was excited to read the next pages to find out \"where, indeed, are those areas that integrate vision and language?\" and tingling with an expectation of learning something new about our brains. But then, for some reason, the Results section is very timid and just presents dry numbers for each of the test that were planned. Were the differences in Pearson R not strong enough for the authors to be confident in their findings? What would be the strong result? Is this just the matter of presentation, or the result is too weak to claim some sort of victory and knowledge discovered?\n\n(1) It would be helpful to include a better explanation of what the \"event structures\" are exactly. Maybe a picture.\n\n(2) Page 6, Section 4, second paragraph: This bit of text here is a bit too overloaded with number and while the authors might expect their readers to be careful and try to understand what is the meaning and significance of those numbers being what they are... but a reader is rarely that careful. I would advise to add explanations to this paragraph that explain to the reader what they are supposed to think when they see this or other set of numbers. Tell the reader what they are supposed to with those numbers.\n\n(3) Figure 2: The caption does not explain the figure well. Panels (a) and (b) are not mentioned in the caption. An attempt to explain with \"mid left\" and \"bottom right\" is confusing, perhaps just put the names of the models on the figure plots. The dots on the freesurfer brain surfaces on the right of the figure are not explained at all - what are they? For the colour-blind it is very hard to see the red dot, I recommend using blue.\n\n(4) Page 8, Section 4.2, second test paragraph: It is unclear without further explanations what do the authors themselves make of this result. Was it interesting and/or significant? What did it demonstrate? Was this a strong result or no so much? Expected or unexpected? All in terms of the main research question of the paper.\n\n(5) Same as (5) applies to paragraphs on tests three, four and five. Actually one also. Currently these paragraphs are basically just a table of results but written out in words. A table with results and numbers is good, but we also need an interpretation and analysis of these results. Are they strong / interesting? Is there a scientific discovery here? What is it? How strong is the evidence? Were the R differences significant?\n\n(6) Page 9, first paragraph: \"we find that the largest contiguous cluster of these electrodes is found in and around the temporoparietal junction\" -- it would help a lot to see those electrodes on a picture of a brain! Not only in a way how it was presented on Figure 4, but actually each electrode plotted as a dot so that the reader could also discover this by actually seeing that \"yep, indeed, here are the electodes that are more predictive in multimodal regime and indeed they cluster around superior temporal region\". Figure 4, in my opinion, falls short of presenting this finding and sweeps the results under average-colored areas, raising questions why the plot was individual electrodes was not shown. Something like Figure 8 (supplementary G), but please use different colors for \"unimodal\" or \"multimodel across vision and language\" as the current selection of colors blurs together\n\n(7) In your own estimation, are those singular electrodes shown on Figure 8 as \"multimodal across vision and language\" provide sufficient evidence to multimodal processing in those areas? Or are the numbers too few to provide strong support for this claim? The yellow dots seem to be quite scattered, and we also don't know, for example, is the lack of them in visual areas explained by the fact that no electrodes in those areas were multimodal, or is this just because there were not electrodes implanted there?\n\n(8) It would help to evaluate the strength of the funding if we would be able to see a comparison (maybe a distribution plot) of multimodal over language-only / vision-only -- this would allow us to see not only where and how much of those electrodes exist, but also _how different_ their predictive power is. The averaged numbers you provide in section 4 are just averages and are just one number, hiding the true distribution we would be interested to see.", "questions": "(1) The implantation sites of sEEG electrodes in your dataset were clinically motivated. To what extent did they cover the areas you were interested in? Both whether all of the areas of interest were covered, and also among the the areas that were covered - was the coverage sufficient for you analysis in your estimation and why?\n\n(2) Are there multimodal models combining vision, text and audio? The data that you have contains all 3 modalities, what was/is the main obstacle to identifying tri-modal predictive regions in the brain? Is it the lack of appropriate DNNs or something else?\n\n(3) In your experiments were the subjects able to hear the audio track of the movie?\n\n(4) For your literature review here is another work comparing vision to DNN specifically on sEEG data from a 100+ subjects https://www.nature.com/articles/s42003-018-0110-y\n\n(5) How do you deal with the fact that the data comes from different subjects? Does inter/intra-subject considerations enter your analysis at all or you just consider each LFP electrode on its own regardless of the subject it came from?\n\n(6) Could you perhaps use fMRI data instead? You would lose temporal and frequency resolution, but for the level of analysis at which you are working these are not too relevant and demonstrating higher predictability of BOLD signal would be equally impressive and informative. More importantly it would allow you to capture the whole brain and there should be more datasets available for fMRI that for sEEG.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "There is a multitude of papers training decoders from latent representations of DNN models onto brain activity. The goal is to reconstruct the brain activity, which, if successful, would indicate that there is something brain-like in the representations that are formed by artificial learning models. One cohort of such papers focus on models of vision (and predicting activity of visual areas of the human brain), while another on language models (and predicting the activity of language areas in the brain). In this paper the authors ask what if we take models that work on vision and language simultaneously - would the representations and activations of those models be more predictive of brain activity? And if so - in which regions?\n\nAn answer to this question might help us understand where in the brain are the areas that integrate different sensory modalities together, or at least work on several modalities at the same time.\n\nThe authors then take 14 models and use their representation to predict each electrode's recordings. The idea is that if a multimodal model's representation is significantly more useful for predicting the brain activity, then this model's representations are closer to what is happening in the brain, and thus can be thought of as evidence for that electrode's area being involved in multimodal processing of information.\n\nThe authors do find several such electrodes, but the number of those electrodes is not sufficiently high to draw strong conclusions (at least this was my impression from reading the paper, please correct me if I am wrong).\n\nOverall this work presents and very cool idea, which is well-executed and logically reported, however due to the lack of data (I suspect) there is just not enough ground to claim definitive findings.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "Indeed an important and timely question, and, to the best of my knowledge, this work is first to explore this topic. I really like the idea and the research question.\n\nI very much support the emphasis you made on the fact that before any analysis is done we should confirm that there is either a strong difference in signal reconstructive power between trained and untrained networks, or a strong decoding ability -- we need to have a confirmation that the signal is indeed there before we analyse anything related to it.", "weaknesses": "My overall impression can be summarised as follows: while the first half of the paper is great and explains the idea and motivation really well, creating a rightful sense of expectation of the result, the section on the results somewhat comes short of delivering the findings with a bang. After reading the first half I was excited to read the next pages to find out \"where, indeed, are those areas that integrate vision and language?\" and tingling with an expectation of learning something new about our brains. But then, for some reason, the Results section is very timid and just presents dry numbers for each of the test that were planned. Were the differences in Pearson R not strong enough for the authors to be confident in their findings? What would be the strong result? Is this just the matter of presentation, or the result is too weak to claim some sort of victory and knowledge discovered?\n\n(1) It would be helpful to include a better explanation of what the \"event structures\" are exactly. Maybe a picture.\n\n(2) Page 6, Section 4, second paragraph: This bit of text here is a bit too overloaded with number and while the authors might expect their readers to be careful and try to understand what is the meaning and significance of those numbers being what they are... but a reader is rarely that careful. I would advise to add explanations to this paragraph that explain to the reader what they are supposed to think when they see this or other set of numbers. Tell the reader what they are supposed to with those numbers.\n\n(3) Figure 2: The caption does not explain the figure well. Panels (a) and (b) are not mentioned in the caption. An attempt to explain with \"mid left\" and \"bottom right\" is confusing, perhaps just put the names of the models on the figure plots. The dots on the freesurfer brain surfaces on the right of the figure are not explained at all - what are they? For the colour-blind it is very hard to see the red dot, I recommend using blue.\n\n(4) Page 8, Section 4.2, second test paragraph: It is unclear without further explanations what do the authors themselves make of this result. Was it interesting and/or significant? What did it demonstrate? Was this a strong result or no so much? Expected or unexpected? All in terms of the main research question of the paper.\n\n(5) Same as (5) applies to paragraphs on tests three, four and five. Actually one also. Currently these paragraphs are basically just a table of results but written out in words. A table with results and numbers is good, but we also need an interpretation and analysis of these results. Are they strong / interesting? Is there a scientific discovery here? What is it? How strong is the evidence? Were the R differences significant?\n\n(6) Page 9, first paragraph: \"we find that the largest contiguous cluster of these electrodes is found in and around the temporoparietal junction\" -- it would help a lot to see those electrodes on a picture of a brain! Not only in a way how it was presented on Figure 4, but actually each electrode plotted as a dot so that the reader could also discover this by actually seeing that \"yep, indeed, here are the electodes that are more predictive in multimodal regime and indeed they cluster around superior temporal region\". Figure 4, in my opinion, falls short of presenting this finding and sweeps the results under average-colored areas, raising questions why the plot was individual electrodes was not shown. Something like Figure 8 (supplementary G), but please use different colors for \"unimodal\" or \"multimodel across vision and language\" as the current selection of colors blurs together\n\n(7) In your own estimation, are those singular electrodes shown on Figure 8 as \"multimodal across vision and language\" provide sufficient evidence to multimodal processing in those areas? Or are the numbers too few to provide strong support for this claim? The yellow dots seem to be quite scattered, and we also don't know, for example, is the lack of them in visual areas explained by the fact that no electrodes in those areas were multimodal, or is this just because there were not electrodes implanted there?\n\n(8) It would help to evaluate the strength of the funding if we would be able to see a comparison (maybe a distribution plot) of multimodal over language-only / vision-only -- this would allow us to see not only where and how much of those electrodes exist, but also _how different_ their predictive power is. The averaged numbers you provide in section 4 are just averages and are just one number, hiding the true distribution we would be interested to see.", "questions": "(1) The implantation sites of sEEG electrodes in your dataset were clinically motivated. To what extent did they cover the areas you were interested in? Both whether all of the areas of interest were covered, and also among the the areas that were covered - was the coverage sufficient for you analysis in your estimation and why?\n\n(2) Are there multimodal models combining vision, text and audio? The data that you have contains all 3 modalities, what was/is the main obstacle to identifying tri-modal predictive regions in the brain? Is it the lack of appropriate DNNs or something else?\n\n(3) In your experiments were the subjects able to hear the audio track of the movie?\n\n(4) For your literature review here is another work comparing vision to DNN specifically on sEEG data from a 100+ subjects https://www.nature.com/articles/s42003-018-0110-y\n\n(5) How do you deal with the fact that the data comes from different subjects? Does inter/intra-subject considerations enter your analysis at all or you just consider each LFP electrode on its own regardless of the subject it came from?\n\n(6) Could you perhaps use fMRI data instead? You would lose temporal and frequency resolution, but for the level of analysis at which you are working these are not too relevant and demonstrating higher predictability of BOLD signal would be equally impressive and informative. More importantly it would allow you to capture the whole brain and there should be more datasets available for fMRI that for sEEG.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698890312872}, {"id": "ucTvwZzyvK", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6104/Reviewer_KL3N"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The current paper described a comparison between language/vision unimodal and language-vision multimodal models in predicting brain activities while the subject is watching movies. The authors trained model features to predict SEEG brain activities using ridge regression. Following the performance of the ridge regression on hold-out set of data, the authors identified the brain areas that best fit by a unimodal or multimodal model. This study shed light on the function of cortical areas.", "review_text": "The current paper described a comparison between language/vision unimodal and language-vision multimodal models in predicting brain activities while the subject is watching movies. The authors trained model features to predict SEEG brain activities using ridge regression. Following the performance of the ridge regression on hold-out set of data, the authors identified the brain areas that best fit by a unimodal or multimodal model. This study shed light on the function of cortical areas.", "strengths": "Using STOA vision and language models to help interpret the function of the brain is interesting.", "weaknesses": "Overall, no insights have been generated in the current study, nor have a solid novel methodology.\n\n1.\tMissing references in the related work section (paragraph 1, last two sentences; paragraph 2, 1st sentence). Missing figure references throughout the text. The writing of the paper lacks clarity. \n\n2.\tModel performance seems to be low. Figure 2a suggests that the Pearson correlation between model predictions and the true neuron activity is about 0.1. This means the model explains about 1% of the variance of the data (linear model r2). This is pretty low. Please explain why we should care about a model with limited prediction power. \n\n3.\tThe performance difference between brain regions, or between models is low, compared to the error bar per condition (figure 2a, and in multimodality tests). By the way, it is not clear to me what the error bar stands for. Is the difference between models or between electrodes sufficient for discrimination?\n\n4.\tWhether the model selection is consistent with the known biological function of the brain areas. \n\n5.\tDo all multimodal tests identify the same set of multimodal selection electrodes? Quantitative results should be provided. A standard procedure should be followed to identify brain areas as potential vision-language integration regions.   \n\n6.\tHow the regression model is trained? What is the input and what is the output? For each model with a different feature size, how does the parameter space for the ridge regression model differ between models? Whether the model performance was affected by the number of parameters?", "questions": "1.\tClarify the analysis and model comparison criteria. \nFigure 1a suggests that the Pearson correlation coefficient is obtained per time bin per electrode per model. What exactly is in the vector that feeds into the Pearson correlation analysis? \nHow many image-text pairs are in the training, testing, and validation dataset? What is the fraction that achieved above threshold prediction? \nProvide example predictions and the corresponding true brain activities, and provide the Pearson r value for the example. \n2.\tImprove figure resolution, please. \n3.\tAddress all the questions raised in the Weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The current paper described a comparison between language/vision unimodal and language-vision multimodal models in predicting brain activities while the subject is watching movies. The authors trained model features to predict SEEG brain activities using ridge regression. Following the performance of the ridge regression on hold-out set of data, the authors identified the brain areas that best fit by a unimodal or multimodal model. This study shed light on the function of cortical areas.", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "strengths": "Using STOA vision and language models to help interpret the function of the brain is interesting.", "weaknesses": "Overall, no insights have been generated in the current study, nor have a solid novel methodology.\n\n1.\tMissing references in the related work section (paragraph 1, last two sentences; paragraph 2, 1st sentence). Missing figure references throughout the text. The writing of the paper lacks clarity. \n\n2.\tModel performance seems to be low. Figure 2a suggests that the Pearson correlation between model predictions and the true neuron activity is about 0.1. This means the model explains about 1% of the variance of the data (linear model r2). This is pretty low. Please explain why we should care about a model with limited prediction power. \n\n3.\tThe performance difference between brain regions, or between models is low, compared to the error bar per condition (figure 2a, and in multimodality tests). By the way, it is not clear to me what the error bar stands for. Is the difference between models or between electrodes sufficient for discrimination?\n\n4.\tWhether the model selection is consistent with the known biological function of the brain areas. \n\n5.\tDo all multimodal tests identify the same set of multimodal selection electrodes? Quantitative results should be provided. A standard procedure should be followed to identify brain areas as potential vision-language integration regions.   \n\n6.\tHow the regression model is trained? What is the input and what is the output? For each model with a different feature size, how does the parameter space for the ridge regression model differ between models? Whether the model performance was affected by the number of parameters?", "questions": "1.\tClarify the analysis and model comparison criteria. \nFigure 1a suggests that the Pearson correlation coefficient is obtained per time bin per electrode per model. What exactly is in the vector that feeds into the Pearson correlation analysis? \nHow many image-text pairs are in the training, testing, and validation dataset? What is the fraction that achieved above threshold prediction? \nProvide example predictions and the corresponding true brain activities, and provide the Pearson r value for the example. \n2.\tImprove figure resolution, please. \n3.\tAddress all the questions raised in the Weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698795966279}, {"id": "GHCPs7Vnlv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6104/Reviewer_HcSy"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors set out to map multimodal networks trained on text and vision to predict intracortical recordings from the brain using sEEG for epilepsy during movie watching. They find clear instances of sites encoding both language and text at the temporoparietal junction.", "review_text": "The authors set out to map multimodal networks trained on text and vision to predict intracortical recordings from the brain using sEEG for epilepsy during movie watching. They find clear instances of sites encoding both language and text at the temporoparietal junction.", "strengths": "This is an informative and well-done study of multimodality measured through intracortical recordings. The data is unique and abundant and the evaluation was done for a large number of models. A lot of attention was put into the controls. As someone who specializes in the field of task-driven neural networks vs brains, I can appreciate that this is well-executed and will surely find a receptive audience for neuroscientists.", "weaknesses": "I don't think this is an appropriate venue for the paper. There's no clear methodological advance in ML that would be of broad interest to the ICLR community: it needs to be read by neuroscientists, not ML people. I looked at the neuroscience papers at ICLR in the last 2 years and found only one that would classify as investigating task-driven neural networks in brains in the style of Yamins, DiCarlo, Kriegeskorte, etc., and that paper showed a clear methodological contribution (https://openreview.net/forum?id=Tp7kI90Htd). The authors should look at where this type of neuroAI work is typically published, e.g. NeurIPS, SVRHM, cosyne, PNAS, Nature Comms, etc.\n\nEdit Nov 23rd: I have no real qualms about what's presented in this paper; in response to my original comments, the authors have argued that every once in a while, a task-driven neural network paper in this style is published at ICLR. Thus, my original questions about the suitability of this work for ICLR notwithstanding (and I will note, consistent with reviewer JsYg), I have increased my score to a 6.", "questions": "-", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors set out to map multimodal networks trained on text and vision to predict intracortical recordings from the brain using sEEG for epilepsy during movie watching. They find clear instances of sites encoding both language and text at the temporoparietal junction.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "This is an informative and well-done study of multimodality measured through intracortical recordings. The data is unique and abundant and the evaluation was done for a large number of models. A lot of attention was put into the controls. As someone who specializes in the field of task-driven neural networks vs brains, I can appreciate that this is well-executed and will surely find a receptive audience for neuroscientists.", "weaknesses": "I don't think this is an appropriate venue for the paper. There's no clear methodological advance in ML that would be of broad interest to the ICLR community: it needs to be read by neuroscientists, not ML people. I looked at the neuroscience papers at ICLR in the last 2 years and found only one that would classify as investigating task-driven neural networks in brains in the style of Yamins, DiCarlo, Kriegeskorte, etc., and that paper showed a clear methodological contribution (https://openreview.net/forum?id=Tp7kI90Htd). The authors should look at where this type of neuroAI work is typically published, e.g. NeurIPS, SVRHM, cosyne, PNAS, Nature Comms, etc.\n\nEdit Nov 23rd: I have no real qualms about what's presented in this paper; in response to my original comments, the authors have argued that every once in a while, a task-driven neural network paper in this style is published at ICLR. Thus, my original questions about the suitability of this work for ICLR notwithstanding (and I will note, consistent with reviewer JsYg), I have increased my score to a 6.", "questions": "-", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "-", "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698164416747}], "openreview_url": "https://openreview.net/forum?id=7Scc7Nl7lg", "arxiv_id": "2406.14481", "paper_pdf": "papers/7Scc7Nl7lg.pdf", "paper_pdf_sha256": "d8188eeba8ed6256030fa6fb09958db5ecd0cc41e0636a257da8d41811e8e8b2", "paper_pdf_bytes": 30079476, "paper_pdf_source": "openreview", "code_url": "https://github.com/vsubramaniam851/brain-multimodal", "code_repository": "vsubramaniam851/brain-multimodal", "code_commit": "87175f724d908cb5e2db33d65ce7e94d777dc430", "code_archive": "repos/7Scc7Nl7lg.zip", "code_archive_sha256": "e13a58b235f8dd49dcc3411768251a6fca2e9fd9e56d048cf3d4568b20527d85", "code_archive_bytes": 27408, "code_file_count": 9, "code_extensions": {".py": 8, ".ipynb": 1}, "github_disk_usage_kb": 34, "github_languages": {"Python": 90948, "Jupyter Notebook": 13584}, "github_archived": false, "github_pushed_at": "2024-07-25T22:27:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/revealing-vision-language-integration-in-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_d2f3hRn0hT", "year": 2023, "status": "rejected", "title": "Single-level Adversarial Data Synthesis based on Neural Tangent Kernels", "authors": ["Yu-Rong Zhang", "Reddy Su", "Sheng-Yen Chou", "Shan-Hung Wu"], "authorids": ["~Yu-Rong_Zhang1", "~Reddy_Su1", "~Sheng-Yen_Chou1", "~Shan-Hung_Wu1"], "authors_source": "OpenReview API", "abstract": "Generative adversarial networks (GANs) have achieved impressive performance in data synthesis and have driven the development of many applications. How- ever, GANs are known to be hard to train due to their bilevel objective, which leads to the problems of convergence, mode collapse, and gradient vanishing. In this paper, we propose a new generative model called the generative adversarial NTK (GA-NTK) that has a single-level objective. The GA-NTK keeps the spirit of adversarial learning (which helps generate plausible data) while avoiding the training difficulties of GANs. This is done by modeling the discriminator as a Gaussian process with a neural tangent kernel (NTK-GP) whose training dynam- ics can be completely described by a closed-form formula. We analyze the conver- gence behavior of GA-NTK trained by gradient descent and give some sufficient conditions for convergence. We also conduct extensive experiments to study the advantages and limitations of GA-NTK and propose some techniques that make GA-NTK more practical.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SW1AGo-WHHN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1331/Reviewer_LpjQ"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose generative adversarial neural tangent kernel (GA-NTK) and its variants.\nThey model the discriminator as a Gaussian process whose mean (and covariance) is governed by NTK.\nAlso, they can be performed via a single-level training process as a whole, because the training dynamics of their discriminators can be evaluated in closed form.\nThe authors claim that this property allows them to avoid the training difficulties often encountered with GANs, and experimentally confirmed this claim and their effectiveness for data synthesis.", "review_text": "I give this study a positive score now, on the basis of the novelty and interest of the idea [1*].\nHowever, the description of the proposed method is insufficient and not clear [3]--[6*].\nThese points could be improved!\nI will change the score up or down, depending on the responses to [3]--[6*].\n\n---\n\nI changed the score from 6 to 8.", "strengths": "I place special emphasis on comments with the mark *.\n\n---\n\nStrength:\n\n[1*] As the authors also wrote at the end of Section 1, the main significance of this study is not performance improvement but novelty of the research direction.\nThis study is an interesting attempt that uses an interesting idea to solve the problems of GANs, which are the motivation for the study.\n\n---\n\nWeakness:\n\n[2] There are writing misses.\n\n[2-1] You should use \\citep as well as \\citet.\nFor example, you can write \"theoretical interpretations Arjovsky & Bottou (2017).\" in p.3 as \"theoretical interpretations (Arjovsky & Bottou, 2017).\" by using \\citep.\nAlso, \"Franceschi et al. Franceschi et al. (2021)\" in p.3 is strange for me; you can write \"Franceschi et al. (2021)\" by using \\citet only.\nPlease look up the use of \\citep and \\citet and rewrite the relevant parts.\n\n[2-2] In l.12 in p.4, value each element -> value of each element.\nCheck your writing (including other parts) again.\n\n[2-3] In eq.(4), $\\arg\\min\\_{Z^n} L(Z^n)=\\arg\\min\\_{Z^n} ||\\cdot||_2$ -> $\\arg\\min\\_{Z^n} L(Z^n)$, where $L(Z^n)=||\\cdot||^2$.\nYou should define $L(Z^n)$ explicitly, and unify $||\\cdot||\\_2$ and $||\\cdot||$ into one if they are the seme (see also norm in other parts).\n\n[2-4] The $f$-divergence appears in places, but what function was used for $f$?\n\n[2-5] In Figure 3, title is partially obscured.\n\n[2-6] In Figure 5, what is \"real\" in the legend?\n\n[2-7] In the last equation in p.14, absolute value symbol is missing:\nfor example, $\\nabla_{z_j} L(Z^n)\\le$ -> $|\\nabla_{z_j} L(Z^n)|\\le$.\n\n[2-8] You should centralize Figure 14.\n\n[2-9] In Section 10, ?? -> 18.\n\n---\n\nQuestion:\n\nRegarding the following points, it may be that the authors' description is appropriate and there is no problem, just because I have not understood it correctly.\nI do not reflect these points in my current recommendation score.\nDepending on the authors' response, I may change my recommendation score.\n\n[3] Is it necessary to solve an optimization problem like eq.(4) every time to generate pseudo examples?\nIf so, that is a demerit compared with GANs, which can generate an infinite number of pseudo examples without additional training from a trained model.\nThis disadvantage should be written more clearly.\n(Even so, I will not lower the score.)\n\n[4*] From the experimental results (see especially Figure 3) and my understanding of the formulas, it seems to me that GA-NTK that is sufficiently trained with a large $t$ would simply return an training example (or mixture of training examples) as a pseudo example.\nIs this question correct?\nIf correct, such data synthesis would have low practical value.\nIt does not seem to me that the use of a generator model adequately solves such a problem.\nAlso, the authors should explain when (i.e., at what $t$) to stop learning.\n\n[5] This question relates to [4*].\nI am interested in the relationship between the batch size $b$ and the novelty of the pseudo example (dissimilarity to the training examples $X^{b/2}$).\nI suspect that if $b$ is small, each generated pseudo example will be fairly close to one of the training examples $X^{b/2}$.\nIs this question correct?\nIf correct, I think that the authors should mention this issue.\n\n[6*] It is difficult for me to understand the whole algorithm.\nI could not understand the relation between the index $t$ in $\\lambda=\\eta \\cdot t$ in eq.(3), the index $j$ in Theorem 3.1, \"Epochs\" in Figure 3, and \"Iterations\" in Figure 5, and the evolution of $Z^n$ and $Z^{n,(j)}$.\nPlease write a pseudo code (or modify the text) so that the relation of these objects become clear.\n\nAlso, how does $t$ change when $Z^n$ changes?\nI think this question is relevant to the validity of applying NTK theory.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the authors propose generative adversarial neural tangent kernel (GA-NTK) and its variants.\nThey model the discriminator as a Gaussian process whose mean (and covariance) is governed by NTK.\nAlso, they can be performed via a single-level training process as a whole, because the training dynamics of their discriminators can be evaluated in closed form.\nThe authors claim that this property allows them to avoid the training difficulties often encountered with GANs, and experimentally confirmed this claim and their effectiveness for data synthesis.", "strength_and_weaknesses": "I place special emphasis on comments with the mark *.\n\n---\n\nStrength:\n\n[1*] As the authors also wrote at the end of Section 1, the main significance of this study is not performance improvement but novelty of the research direction.\nThis study is an interesting attempt that uses an interesting idea to solve the problems of GANs, which are the motivation for the study.\n\n---\n\nWeakness:\n\n[2] There are writing misses.\n\n[2-1] You should use \\citep as well as \\citet.\nFor example, you can write \"theoretical interpretations Arjovsky & Bottou (2017).\" in p.3 as \"theoretical interpretations (Arjovsky & Bottou, 2017).\" by using \\citep.\nAlso, \"Franceschi et al. Franceschi et al. (2021)\" in p.3 is strange for me; you can write \"Franceschi et al. (2021)\" by using \\citet only.\nPlease look up the use of \\citep and \\citet and rewrite the relevant parts.\n\n[2-2] In l.12 in p.4, value each element -> value of each element.\nCheck your writing (including other parts) again.\n\n[2-3] In eq.(4), $\\arg\\min\\_{Z^n} L(Z^n)=\\arg\\min\\_{Z^n} ||\\cdot||_2$ -> $\\arg\\min\\_{Z^n} L(Z^n)$, where $L(Z^n)=||\\cdot||^2$.\nYou should define $L(Z^n)$ explicitly, and unify $||\\cdot||\\_2$ and $||\\cdot||$ into one if they are the seme (see also norm in other parts).\n\n[2-4] The $f$-divergence appears in places, but what function was used for $f$?\n\n[2-5] In Figure 3, title is partially obscured.\n\n[2-6] In Figure 5, what is \"real\" in the legend?\n\n[2-7] In the last equation in p.14, absolute value symbol is missing:\nfor example, $\\nabla_{z_j} L(Z^n)\\le$ -> $|\\nabla_{z_j} L(Z^n)|\\le$.\n\n[2-8] You should centralize Figure 14.\n\n[2-9] In Section 10, ?? -> 18.\n\n---\n\nQuestion:\n\nRegarding the following points, it may be that the authors' description is appropriate and there is no problem, just because I have not understood it correctly.\nI do not reflect these points in my current recommendation score.\nDepending on the authors' response, I may change my recommendation score.\n\n[3] Is it necessary to solve an optimization problem like eq.(4) every time to generate pseudo examples?\nIf so, that is a demerit compared with GANs, which can generate an infinite number of pseudo examples without additional training from a trained model.\nThis disadvantage should be written more clearly.\n(Even so, I will not lower the score.)\n\n[4*] From the experimental results (see especially Figure 3) and my understanding of the formulas, it seems to me that GA-NTK that is sufficiently trained with a large $t$ would simply return an training example (or mixture of training examples) as a pseudo example.\nIs this question correct?\nIf correct, such data synthesis would have low practical value.\nIt does not seem to me that the use of a generator model adequately solves such a problem.\nAlso, the authors should explain when (i.e., at what $t$) to stop learning.\n\n[5] This question relates to [4*].\nI am interested in the relationship between the batch size $b$ and the novelty of the pseudo example (dissimilarity to the training examples $X^{b/2}$).\nI suspect that if $b$ is small, each generated pseudo example will be fairly close to one of the training examples $X^{b/2}$.\nIs this question correct?\nIf correct, I think that the authors should mention this issue.\n\n[6*] It is difficult for me to understand the whole algorithm.\nI could not understand the relation between the index $t$ in $\\lambda=\\eta \\cdot t$ in eq.(3), the index $j$ in Theorem 3.1, \"Epochs\" in Figure 3, and \"Iterations\" in Figure 5, and the evolution of $Z^n$ and $Z^{n,(j)}$.\nPlease write a pseudo code (or modify the text) so that the relation of these objects become clear.\n\nAlso, how does $t$ change when $Z^n$ changes?\nI think this question is relevant to the validity of applying NTK theory.", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\n\nI do not understand the points [3]--[6*].\nOther parts have a good clarity.\n\n---\n\nQuality:\n\nI think that the quality of the paper is not low, but I do not understand the points [3]--[6*].\nSo I will reserve a definite evaluation until the response.\n\n---\n\nNovelty:\n\nAs I wrote in [1*], I give a good evaluation regarding the novelty.\n\n---\n\nReproducibility:\n\nI took a glance of GitHub repository.\nI think that experimental part of this study is reproducible.\nHowever, I want you to cope with [6*], to improve the understandability (a basis of the reproducibility).", "summary_of_the_review": "I give this study a positive score now, on the basis of the novelty and interest of the idea [1*].\nHowever, the description of the proposed method is insufficient and not clear [3]--[6*].\nThese points could be improved!\nI will change the score up or down, depending on the responses to [3]--[6*].\n\n---\n\nI changed the score from 6 to 8.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "8: accept, good paper"}, "tcdate": 1667115102464}, {"id": "HLlECLgvv3v", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1331/Reviewer_rAgg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a novel NTK-based GAN named GA-NTK. Specifically, the authors use an infinite-wide linear neural network (i.e., NTK model) as the discriminator for GAN, which thus enables one to directly obtain a closed-form solution for the inner maximization problem in GAN training. Compared with vanilla GAN, GA-NTK enjoys better convergence and less mode collapse and can avoid gradient vanishing. Experiments are conducted on both synthesized and real-world datasets to justify the effectiveness of the proposed method.", "review_text": "The proposed GA-NTK is novel and the experiment results are strong enough to justify the effectiveness of GA-NTK. Besides, the paper also extends the application domain of NTK and opens up new research direction of GAN training. Therefore, I suggest to accept this paper.", "strengths": "In general, the paper is well-written and the proposed GA-NTK method is very interesting and novel. Other strengths are listed below:\n\n1. GA-NTK converts the original bilevel optimization problem of training GAN to a single-level problem, which significantly reduces the hardness of training GAN.\n\n2. Experiment results on real-world datasets show that GA-NTK can generate images that align well with human perception.\n\n3. Experiment results on synthesized data are strong enough to justify that GA-NTK can avoid mode collapse.\n\n\nOther comments / questions:\n\n1. I wonder the relationship between the batch size and the generative performance of the proposed method. For example, based on the FID results in Table 1, for GACNTKg, one can find a positive correlation on MNIST, a negative correlation on CIFAR-10, and no correlation on CIFAR-100. Any insight about that?\n\n2. What is the time usage for training GA-NTK and its variants?\n\n3. In the second paragraph of page 2, I disagree with the claim that gradients of $\\theta_g$ can not be back-propagated through the inner maximization problem. Actually, once you find the solution of the inner maxization problem in which we denoting it as $\\theta_D^*$, then the gradient for the current $\\theta_g$ can be written in an analytical form with respect to $\\theta_D^*$.\n\n4. The comparison between GA-NTK and existing approaches may be a little unfair, as GA-NTK does not use the same model architectures as that in existing approaches. A more fair comparison for me may be comparing existing approaches with its GA-NTK version in which the finite-wide discrimators are replaced with their corresponded infinite-wide linear conterparts.\n\n5. It would be interesing to compare the convergence speed of GA-NTK with previous GAN training algorithms.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a novel NTK-based GAN named GA-NTK. Specifically, the authors use an infinite-wide linear neural network (i.e., NTK model) as the discriminator for GAN, which thus enables one to directly obtain a closed-form solution for the inner maximization problem in GAN training. Compared with vanilla GAN, GA-NTK enjoys better convergence and less mode collapse and can avoid gradient vanishing. Experiments are conducted on both synthesized and real-world datasets to justify the effectiveness of the proposed method.", "strength_and_weaknesses": "In general, the paper is well-written and the proposed GA-NTK method is very interesting and novel. Other strengths are listed below:\n\n1. GA-NTK converts the original bilevel optimization problem of training GAN to a single-level problem, which significantly reduces the hardness of training GAN.\n\n2. Experiment results on real-world datasets show that GA-NTK can generate images that align well with human perception.\n\n3. Experiment results on synthesized data are strong enough to justify that GA-NTK can avoid mode collapse.\n\n\nOther comments / questions:\n\n1. I wonder the relationship between the batch size and the generative performance of the proposed method. For example, based on the FID results in Table 1, for GACNTKg, one can find a positive correlation on MNIST, a negative correlation on CIFAR-10, and no correlation on CIFAR-100. Any insight about that?\n\n2. What is the time usage for training GA-NTK and its variants?\n\n3. In the second paragraph of page 2, I disagree with the claim that gradients of $\\theta_g$ can not be back-propagated through the inner maximization problem. Actually, once you find the solution of the inner maxization problem in which we denoting it as $\\theta_D^*$, then the gradient for the current $\\theta_g$ can be written in an analytical form with respect to $\\theta_D^*$.\n\n4. The comparison between GA-NTK and existing approaches may be a little unfair, as GA-NTK does not use the same model architectures as that in existing approaches. A more fair comparison for me may be comparing existing approaches with its GA-NTK version in which the finite-wide discrimators are replaced with their corresponded infinite-wide linear conterparts.\n\n5. It would be interesing to compare the convergence speed of GA-NTK with previous GAN training algorithms.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and the proposed GA-NTK method is novel.", "summary_of_the_review": "The proposed GA-NTK is novel and the experiment results are strong enough to justify the effectiveness of GA-NTK. Besides, the paper also extends the application domain of NTK and opens up new research direction of GAN training. Therefore, I suggest to accept this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666973035171}, {"id": "0ABMJSIcS5T", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1331/Reviewer_ihk5"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new framework, GA-NTK, of Generative adversarial networks (GANs) by constructing the discriminator based on the neural tangent kernel (NTK). The NTK allows to describe the discriminator output in a compact and simple function of the given data X and the auxiliary variables Z, and hence many computational issues involved in the GANs training, such as convergence problem, mode collapse, and gradient vanishing, are greatly suppressed. The proposed method's experimental results succeeded in producing artificial images of comparable qualities with other standard methods such as WGAN and SNGAN.\n", "review_text": "The core idea of the paper is novel and interesting but its statistical meaning is not clear as mentioned above. Taken as best one can, it can be said that the authors proposed an optimization framework of generative models rather than the statistical one. However, it is not possible for me to take it so favorably, and hence I vote for a rejection side about this paper.\n", "strengths": "Strengths:\n- The core part of the idea is novel. \n\n- The computational cost of the proposed method is significantly smaller than other comparable methods.  \n\nWeaknesses:\n- From a statistical viewpoint, there is no sufficient theoretical background for the proposed method, I think. The details are described in the Quality section below.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new framework, GA-NTK, of Generative adversarial networks (GANs) by constructing the discriminator based on the neural tangent kernel (NTK). The NTK allows to describe the discriminator output in a compact and simple function of the given data X and the auxiliary variables Z, and hence many computational issues involved in the GANs training, such as convergence problem, mode collapse, and gradient vanishing, are greatly suppressed. The proposed method's experimental results succeeded in producing artificial images of comparable qualities with other standard methods such as WGAN and SNGAN.\n", "strength_and_weaknesses": "Strengths:\n- The core part of the idea is novel. \n\n- The computational cost of the proposed method is significantly smaller than other comparable methods.  \n\nWeaknesses:\n- From a statistical viewpoint, there is no sufficient theoretical background for the proposed method, I think. The details are described in the Quality section below.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The presentation is clear and the manuscript is easy-to-read. But there are still minor flaws. The following list of the spots I realized:\n******************************\nPage 9: What is ``FNN''? \n******************************\n\nQuality: Although the presentation is good, I think the proposed method is not well supported by statistical background. The meaning is as follows. In the standard GANs framework, the generator tries to approximate the data distribution P_{data}. A common way to do so is [1] generating a hidden variable (noise) as z~P_Z (P_Z is usually chosen as Gaussian); [2] updating the generator G so that the distribution of G(z) is as close as possible to the data distribution. The distribution of G(z), P_g, resultantly approximates P_{data}. Hence we can use P_g as a generative model. The proposed method, however, has no such distribution. The author(s) do not discuss this point at all. Actually, the proposed way of generating artificial data is to find the minimizer of the loss defined by eq. (4); if the global minimizer of this minimization problem is obtained, then the generated artificial data is uniquely obtained as a function of the data and has no fluctuation, implying that it is not possible to generate diverse artificial data. Still, from another viewpoint, it is possible that a variety of artificial data can be obtained from optimization issues such as splitting the data into mini-batches and the local minimums of the loss. But the distribution of the variety caused by these issues is basically uncontrollable and cannot be expected to approximate the data distribution.\nHence, in the standard statistical meaning of the generative model, I think the proposed method does not provide any generative model. Due to this reason, I think the quality of the paper is not high.\n\nNovelty: The base idea is novel and interesting, though its statistical justification is absent as I criticize above.\n\nReproducibility: Although I did not try to reproduce the result by myself, the algorithm is explained well and thus one can reproduce the result in principle. Hence there is no serious problem in reproducibility.\n", "summary_of_the_review": "The core idea of the paper is novel and interesting but its statistical meaning is not clear as mentioned above. Taken as best one can, it can be said that the authors proposed an optimization framework of generative models rather than the statistical one. However, it is not possible for me to take it so favorably, and hence I vote for a rejection side about this paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666923932002}, {"id": "SAgdxxDphZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1331/Reviewer_qHF8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a novel generative model GA-NTK that mitigates the drawbacks of traditional GANs trained with the alternative SGD method. This is achieved by using a closed-form discriminator based on a neural tangent kernel (NTK) instead of a neural network. The author proved the convergence of the proposed model during the training phase. Also, numerical experiments are conducted to show the efficacy of the proposed method.", "review_text": "Overall the paper provides a new angle of view for the GAN models. But there are still some points the paper needs to address to meet the bar of acceptance. ", "strengths": "*Strengths*\n- The motivation of the proposed method is clearly stated. The mathematical settings of the proposed method are clearly stated. The overall writing is fluent and easy to understand.\n- Numerical experiments of generating new samples with the method are conducted. Also, the result is consistently evaluated against state-of-the-art methods.\n\n*Weaknesses*\n- The statement of the method could be made better if the training process of the proposed method is more clearly stated. In particular, I'm confused how the proposed discriminator is trained (or not trained) during the overall training process, e.g. what are the variables to be updated during training? \n- Since the main advantage of the proposed method is to substitute traditional neural network-based discriminator with kernel methods, it would be plausible to include the comparison of the discriminator performance between the two methods. \n- The scalability of the proposed method is one of the drawbacks of the proposed method. In particular, the proposed discriminator has a space complexity of O(n^2) where n is the number of data points. \n- The paper makes the assumption that mode collapse is solely caused by min-max training regime in traditional GAN methods. But mode collapse can be caused by other factors as well, such as the structure of generators. A more thorough study of the causal relationship between min-max training and mode collapse could be plausible to provide.\n- In the experiment part, generated images and their most similar images from the dataset are provided. It looks to me that the GA-CNTK method tries to memorize the training samples and the pixel level and not able to learn high-level semantics from the images.\n- typo on page 4: \"my be desirable\" -> \"might be desirable\"", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposed a novel generative model GA-NTK that mitigates the drawbacks of traditional GANs trained with the alternative SGD method. This is achieved by using a closed-form discriminator based on a neural tangent kernel (NTK) instead of a neural network. The author proved the convergence of the proposed model during the training phase. Also, numerical experiments are conducted to show the efficacy of the proposed method.", "strength_and_weaknesses": "*Strengths*\n- The motivation of the proposed method is clearly stated. The mathematical settings of the proposed method are clearly stated. The overall writing is fluent and easy to understand.\n- Numerical experiments of generating new samples with the method are conducted. Also, the result is consistently evaluated against state-of-the-art methods.\n\n*Weaknesses*\n- The statement of the method could be made better if the training process of the proposed method is more clearly stated. In particular, I'm confused how the proposed discriminator is trained (or not trained) during the overall training process, e.g. what are the variables to be updated during training? \n- Since the main advantage of the proposed method is to substitute traditional neural network-based discriminator with kernel methods, it would be plausible to include the comparison of the discriminator performance between the two methods. \n- The scalability of the proposed method is one of the drawbacks of the proposed method. In particular, the proposed discriminator has a space complexity of O(n^2) where n is the number of data points. \n- The paper makes the assumption that mode collapse is solely caused by min-max training regime in traditional GAN methods. But mode collapse can be caused by other factors as well, such as the structure of generators. A more thorough study of the causal relationship between min-max training and mode collapse could be plausible to provide.\n- In the experiment part, generated images and their most similar images from the dataset are provided. It looks to me that the GA-CNTK method tries to memorize the training samples and the pixel level and not able to learn high-level semantics from the images.\n- typo on page 4: \"my be desirable\" -> \"might be desirable\"", "clarity,_quality,_novelty_and_reproducibility": "- The paper's writing has good clarity. The method is novel as it provides a different training regime by introducing a new class of discriminators. \n- code and network architectures are provided to improve reproducibility.", "summary_of_the_review": "Overall the paper provides a new angle of view for the GAN models. But there are still some points the paper needs to address to meet the bar of acceptance. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666568151684}], "openreview_url": "https://openreview.net/forum?id=_d2f3hRn0hT", "arxiv_id": "2204.04090", "paper_pdf": "papers/_d2f3hRn0hT.pdf", "paper_pdf_sha256": "689b9da743b479c97d98057268704bde84e85b5cbb82f40347c555a84b990951", "paper_pdf_bytes": 33549887, "paper_pdf_source": "openreview", "code_url": "https://github.com/ga-ntk/ga-ntk", "code_repository": "ga-ntk/ga-ntk", "code_commit": "8151ac9a39c05689f8070e1b57645361d49f14f9", "code_archive": "repos/_d2f3hRn0hT.zip", "code_archive_sha256": "dd702a7af463567e6b470be3b97d8d7e3782924cad60fcfbb1120def9269ae83", "code_archive_bytes": 44767, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 42, "github_languages": {"Python": 181458}, "github_archived": false, "github_pushed_at": "2022-09-29T10:42:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generative-adversarial-method-based-on-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zLb9oSWy933", "year": 2022, "status": "rejected", "title": "Fast Finite Width Neural Tangent Kernel", "authors": ["Roman Novak", "Jascha Sohl-Dickstein", "Samuel Stern Schoenholz"], "authorids": ["~Roman_Novak2", "~Jascha_Sohl-Dickstein2", "~Samuel_Stern_Schoenholz1"], "authors_source": "OpenReview API", "abstract": "The Neural Tangent Kernel (NTK), defined as the outer product of the neural network (NN) Jacobians, $\\Theta_\\theta(x_1, x_2) = \\left[\\partial f(\\theta, x_1)\\big/\\partial \\theta\\right] \\left[\\partial f(\\theta, x_2)\\big/\\partial \\theta\\right]^T$, has emerged as a central object of study in deep learning. In the infinite width limit, the NTK can sometimes be computed analytically and is useful for understanding training and generalization of NN architectures. At finite widths, the NTK is also used to better initialize NNs, compare the conditioning across models, perform architecture search, and do meta-learning. Unfortunately, the finite-width NTK is notoriously expensive to compute, which severely limits its practical utility. \n\nWe perform the first in-depth analysis of the compute and memory requirements for NTK computation in finite width networks. \nLeveraging the structure of neural networks, we further propose two novel algorithms that change the exponent of the compute and memory requirements of the finite width NTK, dramatically improving efficiency.\n\nWe open-source (https://github.com/iclr2022anon/fast_finite_width_ntk) our two algorithms as general-purpose JAX function transformations that apply to any differentiable computation (convolutions, attention, recurrence, etc.) and introduce no new hyper-parameters.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "BYbuvXI-lDD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4583/Reviewer_tZAo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies an in-depth analysis of runtime and memory requirements for computing the finite-width NTK. The authors analyze computing costs of Jacobian-vector products (and vice versa) for both fully-connected and convolutional neural networks. They also improve the NTK computation cost by leveraging the structure of neural networks resulting in drastic speedup and memory saving. Finally, they make all their implementations open-source based on the JAX library.", "review_text": "- This paper is well-written and easy to understand. The novelty of this work is fairly weak. All results come from simple computations of automatic differentiation, and the reduced cost of NTK follows from amortizing simple linear operations, which is not surprising at all. However, these weaknesses might be covered by their open-source implementation, as it can be significantly important and useful to future works.\n\n- The infinite-width NTK can be exactly computed without Jacobians (Arora et al., 2019) and it can be much efficient than Jacobian-based approaches. Does the finite-width NTK have concrete benefits compared to the infinite-width NTK?\n\n- Minor Issue:\n  - It would be great if more details of Automatic Differentiation (AD) operations to derive memory costs of JVP/VJP are provided.\n  - As assumed $O=\\mathcal{O}(LW)$, the time and memory costs in Section 3.2~3.4 can be reduced without $O$.\n  - The method names in Figure 1 are too small to identify. It would be great to make larger  font size.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies an in-depth analysis of runtime and memory requirements for computing the finite-width NTK. The authors analyze computing costs of Jacobian-vector products (and vice versa) for both fully-connected and convolutional neural networks. They also improve the NTK computation cost by leveraging the structure of neural networks resulting in drastic speedup and memory saving. Finally, they make all their implementations open-source based on the JAX library.", "main_review": "- This paper is well-written and easy to understand. The novelty of this work is fairly weak. All results come from simple computations of automatic differentiation, and the reduced cost of NTK follows from amortizing simple linear operations, which is not surprising at all. However, these weaknesses might be covered by their open-source implementation, as it can be significantly important and useful to future works.\n\n- The infinite-width NTK can be exactly computed without Jacobians (Arora et al., 2019) and it can be much efficient than Jacobian-based approaches. Does the finite-width NTK have concrete benefits compared to the infinite-width NTK?\n\n- Minor Issue:\n  - It would be great if more details of Automatic Differentiation (AD) operations to derive memory costs of JVP/VJP are provided.\n  - As assumed $O=\\mathcal{O}(LW)$, the time and memory costs in Section 3.2~3.4 can be reduced without $O$.\n  - The method names in Figure 1 are too small to identify. It would be great to make larger  font size.\n\n", "summary_of_the_review": "The review score is all about the battle of lack of novelty versus the impact of implementations. I judge this work to be under the bar of acceptance for now but am willing to raise it depending on the author's feedback on the importance of computing finite width NTK.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635891140467}, {"id": "FnWeeyLcfkv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4583/Reviewer_qdx1"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the practical compute and memory requirements to computing the Neural Tangent Kernel (NTK), introducing two new approaches to doing so for standard NN primitives (structured derivative and NTK-vector product) which each have advantages (in terms of variables like batch size, output dim) over the naive jacobian contraction method. The authors provide experiments demonstrating the advantages of their approaches across a range of architectures and hardware. The authors provide open-source code which seems to be integrated neatly with the JAX and Neural Tangents frameworks.", "review_text": "Strengths:\n- Computing the NTK is a popular area at the moment and the open-source code should make some experiments possible that were not previously for some researchers.\n- The approaches proposed exploits nice tricks relating to autodifferentiation (like jvps and vjps) and structure in the computation graphs, which seems tailor made for JAX.\n- The experiments are convincing are back up the author's analysis of the advantages of their methods\n- The paper is generally very well written and clear, I especially like the colour-coded variables like N,O,P etc\n\nWeaknesses (most of these are not huge or beyond the current scope of the paper):\n- The actual technical contribution is not huge, and the paper is more about implementing known ideas from autodiff, linear algebra, and particular aspects of NN structure.\n- While improvements on previous implementations is great (and i suspect is now optimal?), it seems these improvements are still far more expensive that standard sgd training (quadratic in batch size, and output size vs linear), which will hamper the impact and usage of this work. For example, 1 second to compute the NTK of a resnet 18 on a v100 for a single pair of imagenet inputs is not ideal. Also is average pooling applied here? I recall average pooling making computational requirements very tricky for CNNs. This may seem a bit unfair a criticism but I feel it is valid given that this is a code paper.\n- While it is true to JVP and vjp are on the same order of complexity as forward passes, they are strictly more expensive (i believe a jvp is on the order of 3 forward passes), so while the orders you provide e.g. in Table 1,2 are true, in practice the methods will translate to be slightly slower that simply taking the values in Table 1,2 as gospel. I think it would be good to mention this.\n- I think JAX is nice, but wondering if the authors also be implementing their approaches in other frameworks like PyTorch? A lot of non-google/deepmind ppl I know prefer PyTorch to a framework like JAX, and so I believe the impact of the proposed approaches would be greater if so. I realise there are properties with the way that autograd is written in pytorch that makes this more difficult, but I feel like an implementation in just JAX serves to benefit a subset of the research community not the whole, particularly one with a commercial motive.\n\nMinor points:\n- The legend in fig 1 left blocks the graph\n- When you write mathematically in the NTK in the abstract and eq 1, it's really not clear what f, theta, x are? Nor what dimensions that have. I know it's standard notation but some may not have seen this. You can probably get away with it in the abstract for space, but in eq 1 you should spell these out imo.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the practical compute and memory requirements to computing the Neural Tangent Kernel (NTK), introducing two new approaches to doing so for standard NN primitives (structured derivative and NTK-vector product) which each have advantages (in terms of variables like batch size, output dim) over the naive jacobian contraction method. The authors provide experiments demonstrating the advantages of their approaches across a range of architectures and hardware. The authors provide open-source code which seems to be integrated neatly with the JAX and Neural Tangents frameworks.", "main_review": "Strengths:\n- Computing the NTK is a popular area at the moment and the open-source code should make some experiments possible that were not previously for some researchers.\n- The approaches proposed exploits nice tricks relating to autodifferentiation (like jvps and vjps) and structure in the computation graphs, which seems tailor made for JAX.\n- The experiments are convincing are back up the author's analysis of the advantages of their methods\n- The paper is generally very well written and clear, I especially like the colour-coded variables like N,O,P etc\n\nWeaknesses (most of these are not huge or beyond the current scope of the paper):\n- The actual technical contribution is not huge, and the paper is more about implementing known ideas from autodiff, linear algebra, and particular aspects of NN structure.\n- While improvements on previous implementations is great (and i suspect is now optimal?), it seems these improvements are still far more expensive that standard sgd training (quadratic in batch size, and output size vs linear), which will hamper the impact and usage of this work. For example, 1 second to compute the NTK of a resnet 18 on a v100 for a single pair of imagenet inputs is not ideal. Also is average pooling applied here? I recall average pooling making computational requirements very tricky for CNNs. This may seem a bit unfair a criticism but I feel it is valid given that this is a code paper.\n- While it is true to JVP and vjp are on the same order of complexity as forward passes, they are strictly more expensive (i believe a jvp is on the order of 3 forward passes), so while the orders you provide e.g. in Table 1,2 are true, in practice the methods will translate to be slightly slower that simply taking the values in Table 1,2 as gospel. I think it would be good to mention this.\n- I think JAX is nice, but wondering if the authors also be implementing their approaches in other frameworks like PyTorch? A lot of non-google/deepmind ppl I know prefer PyTorch to a framework like JAX, and so I believe the impact of the proposed approaches would be greater if so. I realise there are properties with the way that autograd is written in pytorch that makes this more difficult, but I feel like an implementation in just JAX serves to benefit a subset of the research community not the whole, particularly one with a commercial motive.\n\nMinor points:\n- The legend in fig 1 left blocks the graph\n- When you write mathematically in the NTK in the abstract and eq 1, it's really not clear what f, theta, x are? Nor what dimensions that have. I know it's standard notation but some may not have seen this. You can probably get away with it in the abstract for space, but in eq 1 you should spell these out imo.\n", "summary_of_the_review": "Nice code which should be used by some researchers. Using methods/ideas that are established. Concerns that calculating NTK is still too expensive (which it is by nature, not fault of this paper) to prohibit some researchers and also that uptake in using the code will be only by a subset of the community.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635727014510}, {"id": "WPVoHUX2aO8", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4583/Reviewer_uyYg"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "Experiments are convincing. The provided code is helpful for researchers who need a fast computation of NTK. Though the ideological (mathematical part) is very simple.", "review_text": "The paper describes three methods for computing the neural tangent kernel for a given a batch of vectors and a feed-forward neural network. The first method, Jacobian contraction is quite straightforward and is not suggested for use. The second one,  NTK-vector products is based on reducing the task to the computation of Jacobian-vector and vector-Jacobian products (which themselves can be computed in a fashion similar to forward pass). The method is suggested for use for networks whose width is smaller than the dimension of the output space (probably for contractive autoencoders). The third method, Structured derivatives, is based on a simple identity (9) that simplifies the computation of NTK over weights of the same layer. It turns out that this method is preferable for shallow networks or the typical sitution when  width is larger than the dimension of output space.\n\nExperiments are convincing. The provided code is helpful for researchers who need a fast computation of NTK. Though the ideological (mathematical part) is very simple.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Experiments are convincing. The provided code is helpful for researchers who need a fast computation of NTK. Though the ideological (mathematical part) is very simple.", "main_review": "The paper describes three methods for computing the neural tangent kernel for a given a batch of vectors and a feed-forward neural network. The first method, Jacobian contraction is quite straightforward and is not suggested for use. The second one,  NTK-vector products is based on reducing the task to the computation of Jacobian-vector and vector-Jacobian products (which themselves can be computed in a fashion similar to forward pass). The method is suggested for use for networks whose width is smaller than the dimension of the output space (probably for contractive autoencoders). The third method, Structured derivatives, is based on a simple identity (9) that simplifies the computation of NTK over weights of the same layer. It turns out that this method is preferable for shallow networks or the typical sitution when  width is larger than the dimension of output space.\n\nExperiments are convincing. The provided code is helpful for researchers who need a fast computation of NTK. Though the ideological (mathematical part) is very simple.", "summary_of_the_review": "I would not recommend the paper for publication at such a top-tier conference as ICLR. The simplicity could be justified, in principle, by the novelty. But this paper is too simple, and the novelty is moderate.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635681317616}, {"id": "_uymrH5KAs6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4583/Reviewer_Ye97"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work aims to solve the computation problem of the finite-width neural tangent kernel (NTK), which is a central object in deep learning. The authors analyze the computation and memory requirements for finite-width NTK and propose two novel algorithms that can improve efficiency. Open-source has been provided by the authors.", "review_text": "Strengths\n\n(1) This work provides a comprehensive study on various ways to compute the finite-width NTK. In particular, the time cost and memory cost are given for each case, providing a detailed reference for researchers in related fields.\n\n(2) The author makes the article easy to understand through careful structural planning.\n\nWeakness:\n\n(1) I am confused with the setting that ${\\rm O} = O( {\\rm L W})$, for which the authors demonstrated that the number of logits is dominated by the product of width and depth.  If I understand correctly, $\\rm O$ is the number of neurons in the last layer of the neural network. Thus it should be independent with ${\\rm W}$ and ${\\rm L}$. \n\n(2) The dashed lines in Figure 1 are hard to distingish, I suggest the authros could make it clearer. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work aims to solve the computation problem of the finite-width neural tangent kernel (NTK), which is a central object in deep learning. The authors analyze the computation and memory requirements for finite-width NTK and propose two novel algorithms that can improve efficiency. Open-source has been provided by the authors.", "main_review": "Strengths\n\n(1) This work provides a comprehensive study on various ways to compute the finite-width NTK. In particular, the time cost and memory cost are given for each case, providing a detailed reference for researchers in related fields.\n\n(2) The author makes the article easy to understand through careful structural planning.\n\nWeakness:\n\n(1) I am confused with the setting that ${\\rm O} = O( {\\rm L W})$, for which the authors demonstrated that the number of logits is dominated by the product of width and depth.  If I understand correctly, $\\rm O$ is the number of neurons in the last layer of the neural network. Thus it should be independent with ${\\rm W}$ and ${\\rm L}$. \n\n(2) The dashed lines in Figure 1 are hard to distingish, I suggest the authros could make it clearer. ", "summary_of_the_review": "Overall, this article is clearly written and well structured. If the author can solve the problem of unclear dotted line in Figure 1, it will be more perfect.\n\nBesides, this work builds on JAX. However, the most popular deep learning framework used is PyTorch and Tensorflow. It would be better if the authors discuss why they choose JAX instead of other frameworks.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635415193733}], "openreview_url": "https://openreview.net/forum?id=zLb9oSWy933", "arxiv_id": "2206.08720", "paper_pdf": "papers/zLb9oSWy933.pdf", "paper_pdf_sha256": "59d5dff4295d4aebd90c0472af833f3302c6ac4e2c5acf20e5e6575b480b1cf7", "paper_pdf_bytes": 1028775, "paper_pdf_source": "openreview", "code_url": "https://github.com/iclr2022anon/fast_finite_width_ntk", "code_repository": "iclr2022anon/fast_finite_width_ntk", "code_commit": "09a6c2e805d61f8acebf38c590020b79b491f733", "code_archive": "repos/zLb9oSWy933.zip", "code_archive_sha256": "67c517b5d642af8d4d737c08c83c9a23ea4f37ce30d2879a080ba974aca15573", "code_archive_bytes": 39704, "code_file_count": 9, "code_extensions": {".py": 8, ".ipynb": 1}, "github_disk_usage_kb": 37, "github_languages": {"Python": 104975, "Jupyter Notebook": 54565}, "github_archived": false, "github_pushed_at": "2022-06-20T18:37:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fast-finite-width-neural-tangent-kernel-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5L8XMh667qz", "year": 2021, "status": "rejected", "title": "Encoded Prior Sliced Wasserstein AutoEncoder for learning latent manifold representations", "authors": ["Sanjukta Krishnagopal", "Jacob Bedrossian"], "authorids": ["~Sanjukta_Krishnagopal1", "jacob@math.umd.edu"], "authors_source": "OpenReview API", "abstract": "While variational autoencoders have been successful in a variety of tasks, the use of conventional Gaussian or Gaussian mixture priors are limited in their ability to encode underlying structure of data in the latent representation.\nIn this work, we introduce an Encoded Prior Sliced Wasserstein AutoEncoder (EPSWAE) wherein an additional prior-encoder network facilitates learns an embedding of the data manifold which preserves topological and geometric properties of the data, thus improving the structure of latent space.\nThe autoencoder and prior-encoder networks are iteratively trained using the Sliced Wasserstein (SW) distance, which efficiently measures the distance between two \\textit{arbitrary} sampleable distributions without being constrained to a specific form as in the KL divergence, and without requiring expensive adversarial training.\nTo improve the representation, we use (1) a structural consistency term in the loss that encourages isometry between feature space and latent space and (2) a nonlinear variant of the SW distance which averages over random nonlinear shearing.\nThe effectiveness of the learned manifold encoding is best explored by traversing the latent space through interpolations along \\textit{geodesics} which generate samples that lie on the manifold and hence are advantageous compared to standard Euclidean interpolation.\nTo this end, we introduce a graph-based algorithm for interpolating along network-geodesics in latent space by maximizing the density of samples along the path while minimizing total energy. We use the 3D-spiral data to show that the prior does indeed encode the geometry underlying the data and to demonstrate the advantages of the network-algorithm for interpolation.\nAdditionally, we apply our framework to MNIST, and CelebA datasets, and show that outlier generations, latent representations, and geodesic interpolations are comparable to the state of the art.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Iinm5DjqIzK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3302/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Paper Summary\n\nThe paper extends the variational autoencoder framework with a richer prior distribution to model more complex correlations in the latent variable distribution. They start with a Gaussian mixture distribution as the prior for the latent variables, and add an encoder network to allow richer correlation structure in the latent variables.  Training the prior distribution requires an optimization between the prior distribution and the latent encoded distribution of the training data set. The paper starts with an existing method of optimizing the prior by computing an approximation of the Wasserstein distance between prior and encoded training distribution that uses an average over slices through the prior and encoded training distribution. The paper replaces linear projections used in prior work with a non-linear projection. The paper also employs a structural consistency term which has been used in prior work, however, the paper employs this term differently than prior work by applying it between encoder features and latent variables rather than inputs and latent variables. Since the latent variable space is now a complex and possibly a nonconvex submanifold, points in the latent space R^D  lying between points corresponding to training data points may not actually fall in the training distribution. The paper therefore proposes a method of interpolating between points in the manifold by constructing a graph between points sampled from the manifold and then choosing points lying along lines in the graph. The paper tests the method on three datasets, a synthetic 40 dimensional spiral dataset, the venerable MNIST dataset and a scaled down CELEB A dataset. Plots of the latent space trained on the spiral dataset shows that the latent space can in fact have complex internal structure. \n\n# Pros and Cons\n\nImproving the ability of generative models to capture high-dimensional empirical distributions accurately is a key problem in machine learning and central to the representation learning theme of the ICLR community. \n\nThe paper clearly states contributions up front, namely: using an encoder to generate richer priors for the latent variable distribution, non-linear projections for sliced Wasserstein approximation, and a graph based interpolation method.  They also alter the structural consistency term so that it applies to more abstract features instead of inputs. \n\nThe paper does a thorough and clear job of covering prior work and technical background such as the Wasserstein metric, perhaps even excessively so.  \n\nThe particular choice of a sinusoidal non-linear projection, a key contribution according to the paper, is not motivated in the text. On first glance, the sinusoidal term seems like an odd choice for a non-linearity. After looking at the results which include a test on spirals, it is clearer why this might have be chosen, but is a sinusoidal term likely to be helpful for non periodic data? It might be possible to shed some light on this by investigating whether the coefficients in the non-linear term, zeta and gamma, are significantly different from zero after training on MNIST or the CELEB A data set used in the paper.  \n\nFigure 3 comparing EPSWAE and SWAE doesn’t clearly illustrate the benefits of the EP component. In the print version, both grids of images seem blurry and prone to oddities such as overly large and dark eyes. Even when one blows up the image to 3X size using the digital version the advantage does not jump out. While I recognize that evaluating generative models is hard, the observations do not clearly support the author's hypothesis that the EP component provides an advantage. Maybe they could evaluate Freschet Inception Distance over the whole data set? Use blind human reviewers to choose between EPSWAE and SWAE based on realism and report preference scores? Interestingly, there is one duplicate image in the EPSWAE grid: row 1, col 5 and row 5, col 8 look identical. Seems odd to get identical images: is this because of sampling from a discrete graph structure? The highly similar but not identical images row 1, col 1 and row 1, col 8 are more what I would expect. Maybe the advantage could be made clearer by helping us focus on relevant features. For instance, it might be that EPSWAE is a little bit less likely to generate large black eyes? I can't tell from this small sample, but a grid that focuses on this might make the point. Hair versus background also seems to be a challenge. SWAE image row 3, col 4 seems to have two hair regions for the same face, but, EPSWAE images such as row 4, col 7 also look like they have two distinct hair regions. There are a couple of SWAE images that seem particularly ill formed: row 5, col 5  and maybe row 5, col 8, and row 7, col 4. It might be worth focusing on a few of the worst examples from both EPSWAE and SWAE to show differences in tails rather than the mean? I wonder if you could do a leave-one-out kind of analysis where you check the probability of held out training data points under the prior for EPSWAE vs. SWAE to assess if the prior is capturing the empirical distribution better? You would have to invert the prior network with gradient descent to do this ... but it might work. \n\nIt might make sense to compress some of the tutorial material up front to make room for more results demonstrating the efficacy of the model. The MNIST results, fig 7 in appendix D, shows some advantage for EPSWAE: For instance, figure 7b and 7c both have more bloated numbers … especially 8, 3 and 0.  Also figure 7b has one degenerate number in position 1,2 … possibly a 2?  \n\nThe paper argues that the structured latent space improves generative power. Is there any evidence of structure in the latent space after training on CELEB A? If we plot 2 or 3D projections, or plot projections of structure preservering factorizations such as PCA, do we see structure in the encoding training data points or are they distributed independently? One might also try independence tests between variables in the encoded latent space to see if pairs of variables are being encoded with correlations. This could be compared easily between EPSWAE and SWAE. \n\nFigure 4 in section 5.4 on interpolation shows smooth interpolation between two points A and B in latent space presumably drawn from training data. The interpolations are pretty smooth. Nice! However, the authors claim is that the graph embedding gives better interpolations than linear interpolation between points in the latent space. To make this point, we would also need to see interpolations between points using linear interpolations. The plots on spiral might make this point, but it is not clear.  \n\nThe ability of VAE’s to disentangle the dimensions of an empirical distribution into indpendent latent variables is sometimes seen as a feature, not a bug. For instance, if the training data truly lie along a spiral, isn’t this really a 1D latent space and not a 3D space? While I can see the appeal of improving generated distribution realism, some discussion by the paper on the merits of improving encoder and decoder vs. complexifying the latent space would help to motivate this approach. \n\nIt isn't clear to me that the structural consistency term is a good idea in general. Ideally we want the latent space to capture something fundamental about the underlying structure of the data and not features of the input. Moving the structural consistency from input to more abstract features addresses this concern somewhat, but aren't the latent values themselves the ultimate goal? \n\n# Recommendation\n\n I recommend a rejection of the paper. The hypotheses (that richer priors and geodesic interpolation generate better images on realistic images) are not clearly supported by the experimental results provided.\n\n# Questions\n\n For the spiral training, what was the dimensionality of the latent space? Was it 3D?   \n\nFigure 4 caption contains statement \"through an intermediate sample corresponding to the midpoint in latent space\". Are you actually literally using the midpoint? I thought graph embedding was being used to avoid using midpoints? Maybe the sentence is just ambiguous? \n\nHow many samples are used in the Wasserstein approximation? How were the coefficients in the multi-term loss function defined (alpha, beta and kappa)? Oh - I see these are in the appendix... Seems like appendices are becoming pretty integral to papers these days ... Probably worth including these for the main results in section 5 for the two results presented. \n\nStep 3 in the graph embedding didn't make sense to me after reading it a couple of times. It wasn't quite clear how this sample specific weighting works. It would probably be worth expanding this a bit at the expense of background material, as it is one of the contributions of the paper. \n\n# Other Feedback\n\nPage 2 \"Adversarial methods are harder to train\" ... also adversarial methods are implicit distributions -- you can sample from them but you cannot easily calculate the likelihood of an image under the adversarial model (although you can use gradient descent to try to find the latent parameters). This makes things like outlier detection difficult. \n\nPage 7, Section 5.2, last sentence refers to Fig 6, but I think this should be Appendix D, Fig 7? \n\nIf you flipped the name around from EPSWAE to SWEP-AE, you would have a much more memorable acronym for people to take away from your paper/talk/poster although I recognize this doesn’t have the same “build” on previous work dynamic.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Richer Priors and Geodesic Interpolation Might Improve Image Generation in VAEs  ", "review": "# Paper Summary\n\nThe paper extends the variational autoencoder framework with a richer prior distribution to model more complex correlations in the latent variable distribution. They start with a Gaussian mixture distribution as the prior for the latent variables, and add an encoder network to allow richer correlation structure in the latent variables.  Training the prior distribution requires an optimization between the prior distribution and the latent encoded distribution of the training data set. The paper starts with an existing method of optimizing the prior by computing an approximation of the Wasserstein distance between prior and encoded training distribution that uses an average over slices through the prior and encoded training distribution. The paper replaces linear projections used in prior work with a non-linear projection. The paper also employs a structural consistency term which has been used in prior work, however, the paper employs this term differently than prior work by applying it between encoder features and latent variables rather than inputs and latent variables. Since the latent variable space is now a complex and possibly a nonconvex submanifold, points in the latent space R^D  lying between points corresponding to training data points may not actually fall in the training distribution. The paper therefore proposes a method of interpolating between points in the manifold by constructing a graph between points sampled from the manifold and then choosing points lying along lines in the graph. The paper tests the method on three datasets, a synthetic 40 dimensional spiral dataset, the venerable MNIST dataset and a scaled down CELEB A dataset. Plots of the latent space trained on the spiral dataset shows that the latent space can in fact have complex internal structure. \n\n# Pros and Cons\n\nImproving the ability of generative models to capture high-dimensional empirical distributions accurately is a key problem in machine learning and central to the representation learning theme of the ICLR community. \n\nThe paper clearly states contributions up front, namely: using an encoder to generate richer priors for the latent variable distribution, non-linear projections for sliced Wasserstein approximation, and a graph based interpolation method.  They also alter the structural consistency term so that it applies to more abstract features instead of inputs. \n\nThe paper does a thorough and clear job of covering prior work and technical background such as the Wasserstein metric, perhaps even excessively so.  \n\nThe particular choice of a sinusoidal non-linear projection, a key contribution according to the paper, is not motivated in the text. On first glance, the sinusoidal term seems like an odd choice for a non-linearity. After looking at the results which include a test on spirals, it is clearer why this might have be chosen, but is a sinusoidal term likely to be helpful for non periodic data? It might be possible to shed some light on this by investigating whether the coefficients in the non-linear term, zeta and gamma, are significantly different from zero after training on MNIST or the CELEB A data set used in the paper.  \n\nFigure 3 comparing EPSWAE and SWAE doesn’t clearly illustrate the benefits of the EP component. In the print version, both grids of images seem blurry and prone to oddities such as overly large and dark eyes. Even when one blows up the image to 3X size using the digital version the advantage does not jump out. While I recognize that evaluating generative models is hard, the observations do not clearly support the author's hypothesis that the EP component provides an advantage. Maybe they could evaluate Freschet Inception Distance over the whole data set? Use blind human reviewers to choose between EPSWAE and SWAE based on realism and report preference scores? Interestingly, there is one duplicate image in the EPSWAE grid: row 1, col 5 and row 5, col 8 look identical. Seems odd to get identical images: is this because of sampling from a discrete graph structure? The highly similar but not identical images row 1, col 1 and row 1, col 8 are more what I would expect. Maybe the advantage could be made clearer by helping us focus on relevant features. For instance, it might be that EPSWAE is a little bit less likely to generate large black eyes? I can't tell from this small sample, but a grid that focuses on this might make the point. Hair versus background also seems to be a challenge. SWAE image row 3, col 4 seems to have two hair regions for the same face, but, EPSWAE images such as row 4, col 7 also look like they have two distinct hair regions. There are a couple of SWAE images that seem particularly ill formed: row 5, col 5  and maybe row 5, col 8, and row 7, col 4. It might be worth focusing on a few of the worst examples from both EPSWAE and SWAE to show differences in tails rather than the mean? I wonder if you could do a leave-one-out kind of analysis where you check the probability of held out training data points under the prior for EPSWAE vs. SWAE to assess if the prior is capturing the empirical distribution better? You would have to invert the prior network with gradient descent to do this ... but it might work. \n\nIt might make sense to compress some of the tutorial material up front to make room for more results demonstrating the efficacy of the model. The MNIST results, fig 7 in appendix D, shows some advantage for EPSWAE: For instance, figure 7b and 7c both have more bloated numbers … especially 8, 3 and 0.  Also figure 7b has one degenerate number in position 1,2 … possibly a 2?  \n\nThe paper argues that the structured latent space improves generative power. Is there any evidence of structure in the latent space after training on CELEB A? If we plot 2 or 3D projections, or plot projections of structure preservering factorizations such as PCA, do we see structure in the encoding training data points or are they distributed independently? One might also try independence tests between variables in the encoded latent space to see if pairs of variables are being encoded with correlations. This could be compared easily between EPSWAE and SWAE. \n\nFigure 4 in section 5.4 on interpolation shows smooth interpolation between two points A and B in latent space presumably drawn from training data. The interpolations are pretty smooth. Nice! However, the authors claim is that the graph embedding gives better interpolations than linear interpolation between points in the latent space. To make this point, we would also need to see interpolations between points using linear interpolations. The plots on spiral might make this point, but it is not clear.  \n\nThe ability of VAE’s to disentangle the dimensions of an empirical distribution into indpendent latent variables is sometimes seen as a feature, not a bug. For instance, if the training data truly lie along a spiral, isn’t this really a 1D latent space and not a 3D space? While I can see the appeal of improving generated distribution realism, some discussion by the paper on the merits of improving encoder and decoder vs. complexifying the latent space would help to motivate this approach. \n\nIt isn't clear to me that the structural consistency term is a good idea in general. Ideally we want the latent space to capture something fundamental about the underlying structure of the data and not features of the input. Moving the structural consistency from input to more abstract features addresses this concern somewhat, but aren't the latent values themselves the ultimate goal? \n\n# Recommendation\n\n I recommend a rejection of the paper. The hypotheses (that richer priors and geodesic interpolation generate better images on realistic images) are not clearly supported by the experimental results provided.\n\n# Questions\n\n For the spiral training, what was the dimensionality of the latent space? Was it 3D?   \n\nFigure 4 caption contains statement \"through an intermediate sample corresponding to the midpoint in latent space\". Are you actually literally using the midpoint? I thought graph embedding was being used to avoid using midpoints? Maybe the sentence is just ambiguous? \n\nHow many samples are used in the Wasserstein approximation? How were the coefficients in the multi-term loss function defined (alpha, beta and kappa)? Oh - I see these are in the appendix... Seems like appendices are becoming pretty integral to papers these days ... Probably worth including these for the main results in section 5 for the two results presented. \n\nStep 3 in the graph embedding didn't make sense to me after reading it a couple of times. It wasn't quite clear how this sample specific weighting works. It would probably be worth expanding this a bit at the expense of background material, as it is one of the contributions of the paper. \n\n# Other Feedback\n\nPage 2 \"Adversarial methods are harder to train\" ... also adversarial methods are implicit distributions -- you can sample from them but you cannot easily calculate the likelihood of an image under the adversarial model (although you can use gradient descent to try to find the latent parameters). This makes things like outlier detection difficult. \n\nPage 7, Section 5.2, last sentence refers to Fig 6, but I think this should be Appendix D, Fig 7? \n\nIf you flipped the name around from EPSWAE to SWEP-AE, you would have a much more memorable acronym for people to take away from your paper/talk/poster although I recognize this doesn’t have the same “build” on previous work dynamic.  ", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603922098534}, {"id": "bIENVj0QHo", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3302/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces an additional prior-encoder network to autoencoders to learn an unconstrained prior. The autoencoder and prior-encoder networks are iteratively trained with the sliced Wasserstein distance (SWD). To strengthen SWD, this paper further applies nonlinear transformations with a structural consistency term for better match between two distributions. For better interpolation on the latent space, it also introduces a graph-based algorithm. \n\nWhile the paper cites some works that also aims at addressing the drawback of SWD, it still misses some important related works like [a, b, c]. The paper is highly expected to make more discussion and suitable empirical comparison with them.\n\n[a] Deshpande et al., Generative modeling using the sliced wasserstein distance, CVPR 2018.\n\n[b] Wu et al., Sliced wasserstein generative models, CVPR 2019.\n\n[c] Liutkus et al., Sliced-Wasserstein Flows: Nonparametric Generative Modeling via Optimal Transport and Diffusions, ICML 2019.\n\nThe motivation of using nonlinear transformation sounds not convincing. It is indeed known that traditional SW approximation generally requires a large amount of linear transformations. The bottleneck has been overcome by some existing works (Chen et al., 2020b; Kolouri et al., 2019; Nguyen et al., 2020; Deshpande et al., 2019, 2018). Why can the suggested nonlinear transformation avoid suffering from this issue? It is also not clear why to choose Eq.(6) for the nonlinear transformation. Are there any more excellent properties of the suggested nonlinear transformation compared to the existing methods? It is also necessary to make more discussions on the application of other nonlinear transformations.\n\nThe motivation of using the additional prior-encoder is not clear to me? The introduction states that it learns an unconstrained prior distribution that matches any data manifold topology. Unfortunately, I cannot find any clear explanation about this in the proposed method part. \n\nThe evaluation is highly insufficient. The paper merely compares  the proposed method with SWAE which was published in 2018. More recent methods like [Deshpande et al. 2019, Wu et al. 2019, Liutkus et al. 2019] should be compared for a more complete study. Moreover, new generative modeling methods should be evaluated quantitatively using popular metrics like inception score, FID etc. Unfortunately, this paper does study this at all. In addition, the visual results of the proposed method seems not comparable with the state of the art on the CelebA dataset.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "missing some related works, unclear motivation, and insufficient evaluation", "review": "The paper introduces an additional prior-encoder network to autoencoders to learn an unconstrained prior. The autoencoder and prior-encoder networks are iteratively trained with the sliced Wasserstein distance (SWD). To strengthen SWD, this paper further applies nonlinear transformations with a structural consistency term for better match between two distributions. For better interpolation on the latent space, it also introduces a graph-based algorithm. \n\nWhile the paper cites some works that also aims at addressing the drawback of SWD, it still misses some important related works like [a, b, c]. The paper is highly expected to make more discussion and suitable empirical comparison with them.\n\n[a] Deshpande et al., Generative modeling using the sliced wasserstein distance, CVPR 2018.\n\n[b] Wu et al., Sliced wasserstein generative models, CVPR 2019.\n\n[c] Liutkus et al., Sliced-Wasserstein Flows: Nonparametric Generative Modeling via Optimal Transport and Diffusions, ICML 2019.\n\nThe motivation of using nonlinear transformation sounds not convincing. It is indeed known that traditional SW approximation generally requires a large amount of linear transformations. The bottleneck has been overcome by some existing works (Chen et al., 2020b; Kolouri et al., 2019; Nguyen et al., 2020; Deshpande et al., 2019, 2018). Why can the suggested nonlinear transformation avoid suffering from this issue? It is also not clear why to choose Eq.(6) for the nonlinear transformation. Are there any more excellent properties of the suggested nonlinear transformation compared to the existing methods? It is also necessary to make more discussions on the application of other nonlinear transformations.\n\nThe motivation of using the additional prior-encoder is not clear to me? The introduction states that it learns an unconstrained prior distribution that matches any data manifold topology. Unfortunately, I cannot find any clear explanation about this in the proposed method part. \n\nThe evaluation is highly insufficient. The paper merely compares  the proposed method with SWAE which was published in 2018. More recent methods like [Deshpande et al. 2019, Wu et al. 2019, Liutkus et al. 2019] should be compared for a more complete study. Moreover, new generative modeling methods should be evaluated quantitatively using popular metrics like inception score, FID etc. Unfortunately, this paper does study this at all. In addition, the visual results of the proposed method seems not comparable with the state of the art on the CelebA dataset.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603898902497}, {"id": "POuXoPrMjAK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3302/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses the issues of representation learning with VAEs and propose EPSWAE as a solution. EPSWAE applies a prior encoder to construct an implicit prior, which is more flexible. Moreover, the authors apply the sliced Wasserstein distance for the matching between the posterior and the prior, enhance the conventional SWD with non-linear transformations and make the latent space similar to the feature space with a structural consistency loss. This paper also proposes a graph-based algorithm for minimizing the pathwise energy to achieve the manifold walking to improve the interpolation in the latent space.\n\n-------------------\n\nThe paper is well written. The pipeline is clear and easy to understand. The representation learning with VAEs is a widely studied topic. Using a flexible implicit prior will boost the learning of the latent codes. The usage of the encoder (or generator in the adversarial cases) in the latent space is widely discussed in previous works, such as vampprior, semi-implicit VI, doubly semi-implicit VI, etc. Thus the contribution in this part is limited.  \n\nThe sliced Wasserstein distance is an efficient approximation of the Wasserstein distance for the distribution matching. To avoid the projections that contain useless information, as cited in this paper, a lot of papers generalize the typical random linear transform to non-linear transformation (Chen et al., 2020b; Kolouri et al., 2019; Nguyen et al., 2020; Deshpande et al., 2019). Here the authors propose the non-linear sliced Wasserstein (NSW) distance in Equation 5 with a transformation shown in Equation 6, which appears to be a special case that satisfies the four conditions discussed in (Kolouri et al., 2019). Is there any difference between NSW distance and the generalized SWD in Kolouri et al., 2019? \n\nThe graph-based method is an interesting way for the manifold walking and is much better than conventional ways such as linear interpolation. I think this part should be discussed more in the paper. \n\n-------------------\n\nSome detailed questions about the technique\n\n* The authors claim the usage of FSC encourages pairwise distances of the latent code to be similar to the pairwise distances of the data features. I am curious if FSC is necessary for manifold learning. In figure 5, it does not show much difference with/without the FSC loss. \n\n* As claimed in the paper, the adversarial methods in latent space are expensive, while in the experiment part, there is no computation (such as time per update step) comparison with the adversarial methods. The usage of FSC also needs to compute the pairwise distance. Is that expensive too?\n\n* In the abstract and the pseudocode shown in the appendix, the prior encoder is trained with sliced Wasserstein loss. Why we choose SWD for the training instead of NSW? \n\n* It is also confusing that in the method part, the prior encoder is trained with NSW loss, which is not consistent as claimed in the other parts of the paper.\n\n* In the experiments, the authors claim the generations are better, while the quantitative results, such as fid, are not provided. \n\n* For the spiral toy dataset: views in Fig. 5 seem not consistent and are hard to compare. How about we compare in 2d cases?\n\nIn conclusion, the paper has the merits, and these investigations may be helpful for this problem, but it is not enough and need to dig out more to be an ICLR publication.  \n\n---\n\nUpdate after the discussions\n\nI appreciate the efforts that the authors make in their responses, some of which address my concerns and improve the quality of the paper.\n\nI have raised my rating. However, taking into account all information during the discussion phase, I stick to my original review that this paper still needs to explore more to be a mature publication. For example, if the main contribution comes from the prior encoder, as I said the contribution is limited since the usage of the encoder (or generator in the adversarial cases) in the latent space is widely discussed in previous works, such as vampprior, semi-implicit VI, doubly semi-implicit VI, etc. This also seems to make the contribution of the sliced Wasserstein part incremental. Plus, as this paper has several components, their relations need to be discussed in a more clear way. Thus, more ablation studies are needed to help this paper to present its insight in a much more clear way.\n\nThanks again for the efforts that the authors make and I hope my reviews could help them to polish this paper to be a nice publication. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper addresses the representation learning of VAEs and is reasonable, but need to dig more.", "review": "This paper addresses the issues of representation learning with VAEs and propose EPSWAE as a solution. EPSWAE applies a prior encoder to construct an implicit prior, which is more flexible. Moreover, the authors apply the sliced Wasserstein distance for the matching between the posterior and the prior, enhance the conventional SWD with non-linear transformations and make the latent space similar to the feature space with a structural consistency loss. This paper also proposes a graph-based algorithm for minimizing the pathwise energy to achieve the manifold walking to improve the interpolation in the latent space.\n\n-------------------\n\nThe paper is well written. The pipeline is clear and easy to understand. The representation learning with VAEs is a widely studied topic. Using a flexible implicit prior will boost the learning of the latent codes. The usage of the encoder (or generator in the adversarial cases) in the latent space is widely discussed in previous works, such as vampprior, semi-implicit VI, doubly semi-implicit VI, etc. Thus the contribution in this part is limited.  \n\nThe sliced Wasserstein distance is an efficient approximation of the Wasserstein distance for the distribution matching. To avoid the projections that contain useless information, as cited in this paper, a lot of papers generalize the typical random linear transform to non-linear transformation (Chen et al., 2020b; Kolouri et al., 2019; Nguyen et al., 2020; Deshpande et al., 2019). Here the authors propose the non-linear sliced Wasserstein (NSW) distance in Equation 5 with a transformation shown in Equation 6, which appears to be a special case that satisfies the four conditions discussed in (Kolouri et al., 2019). Is there any difference between NSW distance and the generalized SWD in Kolouri et al., 2019? \n\nThe graph-based method is an interesting way for the manifold walking and is much better than conventional ways such as linear interpolation. I think this part should be discussed more in the paper. \n\n-------------------\n\nSome detailed questions about the technique\n\n* The authors claim the usage of FSC encourages pairwise distances of the latent code to be similar to the pairwise distances of the data features. I am curious if FSC is necessary for manifold learning. In figure 5, it does not show much difference with/without the FSC loss. \n\n* As claimed in the paper, the adversarial methods in latent space are expensive, while in the experiment part, there is no computation (such as time per update step) comparison with the adversarial methods. The usage of FSC also needs to compute the pairwise distance. Is that expensive too?\n\n* In the abstract and the pseudocode shown in the appendix, the prior encoder is trained with sliced Wasserstein loss. Why we choose SWD for the training instead of NSW? \n\n* It is also confusing that in the method part, the prior encoder is trained with NSW loss, which is not consistent as claimed in the other parts of the paper.\n\n* In the experiments, the authors claim the generations are better, while the quantitative results, such as fid, are not provided. \n\n* For the spiral toy dataset: views in Fig. 5 seem not consistent and are hard to compare. How about we compare in 2d cases?\n\nIn conclusion, the paper has the merits, and these investigations may be helpful for this problem, but it is not enough and need to dig out more to be an ICLR publication.  \n\n---\n\nUpdate after the discussions\n\nI appreciate the efforts that the authors make in their responses, some of which address my concerns and improve the quality of the paper.\n\nI have raised my rating. However, taking into account all information during the discussion phase, I stick to my original review that this paper still needs to explore more to be a mature publication. For example, if the main contribution comes from the prior encoder, as I said the contribution is limited since the usage of the encoder (or generator in the adversarial cases) in the latent space is widely discussed in previous works, such as vampprior, semi-implicit VI, doubly semi-implicit VI, etc. This also seems to make the contribution of the sliced Wasserstein part incremental. Plus, as this paper has several components, their relations need to be discussed in a more clear way. Thus, more ablation studies are needed to help this paper to present its insight in a much more clear way.\n\nThanks again for the efforts that the authors make and I hope my reviews could help them to polish this paper to be a nice publication. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603779726768}], "openreview_url": "https://openreview.net/forum?id=5L8XMh667qz", "arxiv_id": "2010.01037", "paper_pdf": "papers/5L8XMh667qz.pdf", "paper_pdf_sha256": "e5fe10b8d71575c08ccc3563ada6149268d7207f66386d947d8894d66c3bbfed", "paper_pdf_bytes": 12219545, "paper_pdf_source": "openreview", "code_url": "https://github.com/chimeraki/EPSWAE", "code_repository": "chimeraki/EPSWAE", "code_commit": "64c1d7c7d66a481cc449bffc28c05c5616744c62", "code_archive": "repos/5L8XMh667qz.zip", "code_archive_sha256": "0b9834d346718e1e8c53ecff1048efa16107686dfd5ba73b1a9183c435aaa85c", "code_archive_bytes": 32340, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 44, "github_languages": {"Python": 55281}, "github_archived": false, "github_pushed_at": "2020-10-05T12:01:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/encoded-prior-sliced-wasserstein-autoencoder"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJlJSaEFwS", "year": 2020, "status": "rejected", "title": "Robust Cross-lingual Embeddings from Parallel Sentences ", "authors": ["Ali Sabet", "Prakhar Gupta", "Jean-Baptiste Cordonnier", "Robert West", "Martin Jaggi"], "authorids": ["asabet@uwaterloo.ca", "prakhar.gupta@epfl.ch", "jean-baptiste.cordonnier@epfl.ch", "robert.west@epfl.ch", "martin.jaggi@epfl.ch"], "authors_source": "OpenReview API", "abstract": "Recent advances in cross-lingual word embeddings have primarily relied on mapping-based methods, which project pretrained word embeddings from different languages into a shared space through a linear transformation. However, these approaches assume word embedding spaces are isomorphic between different languages, which has been shown not to  hold in practice (Søgaard et al., 2018), and fundamentally limits their performance. This motivates investigating joint learning methods which can overcome this impediment, by simultaneously learning embeddings across languages via a cross-lingual term in the training objective. Given the abundance of parallel data available (Tiedemann, 2012), we propose a bilingual extension of the CBOW method which leverages sentence-aligned corpora to obtain robust cross-lingual word and sentence representations. Our approach significantly improves cross-lingual sentence retrieval performance over all other approaches, as well as convincingly outscores mapping methods while maintaining parity with jointly trained methods on word-translation. It also achieves parity with a deep RNN method on a zero-shot cross-lingual document classification task, requiring far fewer computational resources for training and inference. As an additional advantage, our bilingual method also improves the quality of monolingual word vectors despite training on much smaller datasets.  We make our code and models publicly available.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BJeROFsS5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper507/AnonReviewer4"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "* Recommendation\nWhile the contributions in this work are not staggeringly innovative, they are well grounded in existing work and well supported by experiments. Therefore I think the paper should be accepted.\n\n* Summarize paper’s major contributions\n\nThe authors aimed to improve on the task of cross-lingual sentence retrieval, by introducing a model with a modified objective function, which utilizes a cross-lingual loss. They demonstrated that this objective function led to large improvements on word-level representation tasks and cross-lingual sentence retrieval, and achieves competitive performance on a document-level task while being more computationally efficient. The authors performed an in-depth ablation study, to support their claims that the proposed model addresses some of the key problems with other existing approaches to cross-lingual representation learning (e.g. hubness). \n\n* Comments on the paper\n\nThis paper is exceptionally well written, organized, and clear. In addition to a solid introduction and related works sections, which frame the problem nicely, the conducted experiments thoroughly demonstrate the performance of the proposed model, as they evaluate on several granularities (word, sentence, document), as well as a robust ablation study and analysis. By the end of the paper, I am sufficiently convinced by the work and its contributions. \n\nMeanwhile, I believe the paper could benefit from more discussion or analysis of cases where the proposed model did not lead to improvements, in particular with Russian in the word-translation retrieval experiment, where TRANSGRAM outperforms the proposed model. Although the authors briefly note this in the Discussion section, there is unfortunately no conversation about why this may be. The proposed model becomes less convincing when I consider that it might only work for other agglutinative, English-like languages, and I wonder how this approach would fair with other morphologically rich languages similar to Russian, and non-agglutinative languages in general. \n\n* Minor corrections:\n\n- In Figure-1’s caption, at the very end, there is a space missing between “sentence_(cross-lingual compotent).”\n\n- Some missing colon’s (:) throughout the paper when breaking from a paragraph to introduce a list (e.g. “Contributions” in the Introductions)\n\n- The x-axis of Figure 2 and Figure 3 are unclear to me, and a bit difficult to read. How do I interpret “10^-2” as a corpus size? In other words, what are “10^-2 amounts of data.” Fix this. \n\nOverall, great work. I also appreciate the details about training both in the paper and appendix, that will be useful for those wishing to reproduce this work. \n\n* Questions for authors\n\n- Why do you think that TRANSGRAM outperformed your system on the word-translation retrieval experiment for Russian? Do you have any reasons to believe that the proposed model cannot extend well to other morphologically rich languages, or languages very dissimilar to English? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #4", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "* Recommendation\nWhile the contributions in this work are not staggeringly innovative, they are well grounded in existing work and well supported by experiments. Therefore I think the paper should be accepted.\n\n* Summarize paper’s major contributions\n\nThe authors aimed to improve on the task of cross-lingual sentence retrieval, by introducing a model with a modified objective function, which utilizes a cross-lingual loss. They demonstrated that this objective function led to large improvements on word-level representation tasks and cross-lingual sentence retrieval, and achieves competitive performance on a document-level task while being more computationally efficient. The authors performed an in-depth ablation study, to support their claims that the proposed model addresses some of the key problems with other existing approaches to cross-lingual representation learning (e.g. hubness). \n\n* Comments on the paper\n\nThis paper is exceptionally well written, organized, and clear. In addition to a solid introduction and related works sections, which frame the problem nicely, the conducted experiments thoroughly demonstrate the performance of the proposed model, as they evaluate on several granularities (word, sentence, document), as well as a robust ablation study and analysis. By the end of the paper, I am sufficiently convinced by the work and its contributions. \n\nMeanwhile, I believe the paper could benefit from more discussion or analysis of cases where the proposed model did not lead to improvements, in particular with Russian in the word-translation retrieval experiment, where TRANSGRAM outperforms the proposed model. Although the authors briefly note this in the Discussion section, there is unfortunately no conversation about why this may be. The proposed model becomes less convincing when I consider that it might only work for other agglutinative, English-like languages, and I wonder how this approach would fair with other morphologically rich languages similar to Russian, and non-agglutinative languages in general. \n\n* Minor corrections:\n\n- In Figure-1’s caption, at the very end, there is a space missing between “sentence_(cross-lingual compotent).”\n\n- Some missing colon’s (:) throughout the paper when breaking from a paragraph to introduce a list (e.g. “Contributions” in the Introductions)\n\n- The x-axis of Figure 2 and Figure 3 are unclear to me, and a bit difficult to read. How do I interpret “10^-2” as a corpus size? In other words, what are “10^-2 amounts of data.” Fix this. \n\nOverall, great work. I also appreciate the details about training both in the paper and appendix, that will be useful for those wishing to reproduce this work. \n\n* Questions for authors\n\n- Why do you think that TRANSGRAM outperformed your system on the word-translation retrieval experiment for Russian? Do you have any reasons to believe that the proposed model cannot extend well to other morphologically rich languages, or languages very dissimilar to English? "}, "tcdate": 1572350325620}, {"id": "BygJyY52FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper507/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper does not bring anything novel to the field of cross-lingual representation learning: it just revisits some older ideas (from the period of 2013-2015), now revamped, given the fact that more sophisticated and more effective methods are used to model exactly the same intuitions. I see this work as largely incremental, and it just further supports what has been known before, and it further supports recent findings (which are all quite straightforward) from the work of Ormazabal et al. (ACL 2019). The actual model implementation is a straightforward extension of the Sent2Vec model to cross-lingual scenarios, inspired by previous work (e.g., the work on TransGram and BiVec), so the paper is also very incremental from the methodological perspective.\n\nI am puzzled why MUSE is selected as the unsupervised baseline given that fact that: 1) previous work showed its non-robustness for many language pairs, 2) the VecMap model of Artetxe et al. has been proven as the most robust unsupervised cross-lingual word embedding model in several recent empirical analyses - see e.g., Glavas et al. (ACL 2019), Heyman et al. (NAACL 2019), or the original VecMap work. Also, I am puzzled why the paper overstates the rekindled interest towards TransGram, given that TransGram and especially BiVec are well-known models that learn from parallel data.\n\nAnother note related to evaluation: to really establish how different cross-lingual embeddings compare to each other, a wider set of experiments and downstream evaluation is definitely required, see the work of e.g., Glavas et al. (ACL 2019). \n\nMost importantly, the paper evaluates only on very similar language pairs. The main reason why much recent work has focused on alignment-based/projection-based methods was quite pragmatic: we need such weakly supervised methods where we cannot assume the abundance of parallel data to enable cross-lingual transfer in resource-poor settings. If parallel data exists, it is quite intuitive and obvious (and also empirically validated before) that joint modeling is a better choice than a weakly supervised method that just uses 1K or 5K translation pairs. In fact, I am not even sure that it is fair to compare models that rely only on 1K translation pairs with models that draw their strength from 1M or 2M parallel sentences. This paper just shows that, if we have parallel data (which we do for many resource-richer language pairs), it is better to do joint modeling instead of learning simple alignments, but that is a pretty trivial finding imho.\n\nAre the results on MLDoc really state-of-the-art? The results are actually quite mixed, and the advantage of Bi-Sent2Vec is its quicker training. However, what about more recent methods such as XLM which rely on exactly the same resources as Bi-Sent2Vec to do the zero-shot classification task?\n\nTable 2: it is a well-known fact that multilingual training can improve performance in monolingual supervision: see e.g., the work of Faruqui and Dyer (EACL 2014, not cited). Alignment-based approaches that apply the Orthogonal Procrustes mapping cannot improve on monolingual word similarity simply because the orthogonality constraint preserves the topology of the original space. Therefore, evaluating different embedding methods on the intrinsic word similarity task is not a sound evaluation protocol imho - it would be much more informative to plug the embeddings as features in a classification or a parsing task (or something else).\n\nFigure 2: corpus size. Based on the results presented, it seems that the performance saturates by adding more parallel data, but the authors fail to fully understand their evaluation data in the first place. For instance, there are multiple problems with the MUSE datasets, as discussed in the recent work of Kementchedjhieva et al. (EMNLP 2019) - it evaluates mostly high-frequent word (actually - noun) translation, and of course that this saturates more quickly. It doesn't by any means imply that joint training therefore requires less data to reach peak performance: this is true only with the MUSE dataset, and is not a general truth.\n\nMinor:\n- The work of Artetxe et al. (ACL 2017) should be cited when talking about bootstrapping alignment-based methods from limited bilingual supervision (instead of the work of Artetxe and Schwenk which concerns learning multilingual sentence embeddings).\n- Many very relevant and historically important papers are omitted from the related work section: e.g., Hermann and Blunsom's work, Chandar et al., Soyer et al., Vulic and Moens, Gouws et al., Levy et al., to name only a few.\n- I am not sure that the statement that BiVec is not needed in the presence of TransGram is true in general: it mostly suggests that there are some deficiencies with the evaluation protocol.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "The paper does not bring anything novel to the field of cross-lingual representation learning: it just revisits some older ideas (from the period of 2013-2015), now revamped, given the fact that more sophisticated and more effective methods are used to model exactly the same intuitions. I see this work as largely incremental, and it just further supports what has been known before, and it further supports recent findings (which are all quite straightforward) from the work of Ormazabal et al. (ACL 2019). The actual model implementation is a straightforward extension of the Sent2Vec model to cross-lingual scenarios, inspired by previous work (e.g., the work on TransGram and BiVec), so the paper is also very incremental from the methodological perspective.\n\nI am puzzled why MUSE is selected as the unsupervised baseline given that fact that: 1) previous work showed its non-robustness for many language pairs, 2) the VecMap model of Artetxe et al. has been proven as the most robust unsupervised cross-lingual word embedding model in several recent empirical analyses - see e.g., Glavas et al. (ACL 2019), Heyman et al. (NAACL 2019), or the original VecMap work. Also, I am puzzled why the paper overstates the rekindled interest towards TransGram, given that TransGram and especially BiVec are well-known models that learn from parallel data.\n\nAnother note related to evaluation: to really establish how different cross-lingual embeddings compare to each other, a wider set of experiments and downstream evaluation is definitely required, see the work of e.g., Glavas et al. (ACL 2019). \n\nMost importantly, the paper evaluates only on very similar language pairs. The main reason why much recent work has focused on alignment-based/projection-based methods was quite pragmatic: we need such weakly supervised methods where we cannot assume the abundance of parallel data to enable cross-lingual transfer in resource-poor settings. If parallel data exists, it is quite intuitive and obvious (and also empirically validated before) that joint modeling is a better choice than a weakly supervised method that just uses 1K or 5K translation pairs. In fact, I am not even sure that it is fair to compare models that rely only on 1K translation pairs with models that draw their strength from 1M or 2M parallel sentences. This paper just shows that, if we have parallel data (which we do for many resource-richer language pairs), it is better to do joint modeling instead of learning simple alignments, but that is a pretty trivial finding imho.\n\nAre the results on MLDoc really state-of-the-art? The results are actually quite mixed, and the advantage of Bi-Sent2Vec is its quicker training. However, what about more recent methods such as XLM which rely on exactly the same resources as Bi-Sent2Vec to do the zero-shot classification task?\n\nTable 2: it is a well-known fact that multilingual training can improve performance in monolingual supervision: see e.g., the work of Faruqui and Dyer (EACL 2014, not cited). Alignment-based approaches that apply the Orthogonal Procrustes mapping cannot improve on monolingual word similarity simply because the orthogonality constraint preserves the topology of the original space. Therefore, evaluating different embedding methods on the intrinsic word similarity task is not a sound evaluation protocol imho - it would be much more informative to plug the embeddings as features in a classification or a parsing task (or something else).\n\nFigure 2: corpus size. Based on the results presented, it seems that the performance saturates by adding more parallel data, but the authors fail to fully understand their evaluation data in the first place. For instance, there are multiple problems with the MUSE datasets, as discussed in the recent work of Kementchedjhieva et al. (EMNLP 2019) - it evaluates mostly high-frequent word (actually - noun) translation, and of course that this saturates more quickly. It doesn't by any means imply that joint training therefore requires less data to reach peak performance: this is true only with the MUSE dataset, and is not a general truth.\n\nMinor:\n- The work of Artetxe et al. (ACL 2017) should be cited when talking about bootstrapping alignment-based methods from limited bilingual supervision (instead of the work of Artetxe and Schwenk which concerns learning multilingual sentence embeddings).\n- Many very relevant and historically important papers are omitted from the related work section: e.g., Hermann and Blunsom's work, Chandar et al., Soyer et al., Vulic and Moens, Gouws et al., Levy et al., to name only a few.\n- I am not sure that the statement that BiVec is not needed in the presence of TransGram is true in general: it mostly suggests that there are some deficiencies with the evaluation protocol.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571756246711}, {"id": "r1xkQQ6iYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper507/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n***Update***\n\nI thank the reviewers for answering my questions, and I have read the reviews from the other reviewers.  I am borderline on this paper, but still learn towards rejection. I feel that it is a bit incremental and still a little misleading/over-stated. For instance, xnli isn't mentioned but mldoc is, it isn't clear in the table which methods are mappings and which aren't (big data difference), in the mldoc experiment was more parallel data used for these specific languages than was used for LASER? the experiment I asked for wasn't done on the same corpora head-to-head with LASER and their approach. I think any claims about outperforming LASER need to be evaluated on the same corpus as much as possible (Europarl) and weaknesses of the approach should also be mentioned (it is not multilingual in these evaluations). Otherwise you conflate data/multilingualness/model which makes it hard to draw conclusions from the experiments. Also CSLS  was not applied to the ACL paper and this makes a huge difference and should be accounted for (and also idf possibly). Also this paper still isn't at least mentioned which is very related in my opinion.\n\nThis paper proposes a method to learn bilingual word embeddings by modifying the Sent2Vec (which is based on word2vec) approach, and applying it to bitext. They evaluate on monolingual and bilingual word similarity, bitext mining, and a zero-shot document classification task where a classifier is trained on data in one language but evaluated on another (using their embeddings as features in both cases).\n\nI have the following concerns about this paper:\n\n1) For the sentence mining tasks - how come Ormazabal 2019 is not included? The baselines are weak and the task is not standard. One of the baselines used here is MUSE which is unsupervised (The refined version of MUSE was also not included in these experiments). There are fixed datasets that people have experimented with (like BUCC) that would help tie in your results here with the literature a little better. These could include LASER (which is already compared to for document classification) and \"Simple and Effective Paraphrastic Similarity from Parallel Translations\" ACL 2019, which like this paper, proposes a pooled token embedding approach and outperforms more complicated architectures. Neither of these approaches uses idf either (how much does idf help - is it really needed?).\n\n2) For the zero-shot document classification task, it should also be pointed out that LASER can handle many languages all at once. Can this approach also work well if all languages were trained jointly?. Also did you evaluate on XNLI for zero-shot as well? LASER does very well here. I do realize comparing to LASER is not really fair since it is trained on so much data, however the model is similar to previous versions of LASER that were trained on Europarl, which could also be used as the data for training your models for a more apples-to-apples comparison.\n\n3) Note that also the improvements on monolingual similarity using parallel data are well known. For instance \"Embedding Word Similarity with Neural Machine Translation\" (arXiv 2014).. Also even the base for the current state of the art models on SimLex-999, Paragram (TACL 2015), used paraphrases created by pivoting on parallel data.\n\nI think the main contributions of this paper are modifying Sent2vec so it can be used on bilingual data and using it to learn nice representations for words and sentences (and documents). I think that to be published, it should clearly outperform and/or have advantages over all previous works - and this is not clear from this paper in its current form. It is okay if it doesn't do the best on everything, but it is hard to tell how this work currently fits into the literature especially in terms of the sentence-level tasks which are a focus.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "N/A", "title": "Official Blind Review #2", "review": "\n***Update***\n\nI thank the reviewers for answering my questions, and I have read the reviews from the other reviewers.  I am borderline on this paper, but still learn towards rejection. I feel that it is a bit incremental and still a little misleading/over-stated. For instance, xnli isn't mentioned but mldoc is, it isn't clear in the table which methods are mappings and which aren't (big data difference), in the mldoc experiment was more parallel data used for these specific languages than was used for LASER? the experiment I asked for wasn't done on the same corpora head-to-head with LASER and their approach. I think any claims about outperforming LASER need to be evaluated on the same corpus as much as possible (Europarl) and weaknesses of the approach should also be mentioned (it is not multilingual in these evaluations). Otherwise you conflate data/multilingualness/model which makes it hard to draw conclusions from the experiments. Also CSLS  was not applied to the ACL paper and this makes a huge difference and should be accounted for (and also idf possibly). Also this paper still isn't at least mentioned which is very related in my opinion.\n\nThis paper proposes a method to learn bilingual word embeddings by modifying the Sent2Vec (which is based on word2vec) approach, and applying it to bitext. They evaluate on monolingual and bilingual word similarity, bitext mining, and a zero-shot document classification task where a classifier is trained on data in one language but evaluated on another (using their embeddings as features in both cases).\n\nI have the following concerns about this paper:\n\n1) For the sentence mining tasks - how come Ormazabal 2019 is not included? The baselines are weak and the task is not standard. One of the baselines used here is MUSE which is unsupervised (The refined version of MUSE was also not included in these experiments). There are fixed datasets that people have experimented with (like BUCC) that would help tie in your results here with the literature a little better. These could include LASER (which is already compared to for document classification) and \"Simple and Effective Paraphrastic Similarity from Parallel Translations\" ACL 2019, which like this paper, proposes a pooled token embedding approach and outperforms more complicated architectures. Neither of these approaches uses idf either (how much does idf help - is it really needed?).\n\n2) For the zero-shot document classification task, it should also be pointed out that LASER can handle many languages all at once. Can this approach also work well if all languages were trained jointly?. Also did you evaluate on XNLI for zero-shot as well? LASER does very well here. I do realize comparing to LASER is not really fair since it is trained on so much data, however the model is similar to previous versions of LASER that were trained on Europarl, which could also be used as the data for training your models for a more apples-to-apples comparison.\n\n3) Note that also the improvements on monolingual similarity using parallel data are well known. For instance \"Embedding Word Similarity with Neural Machine Translation\" (arXiv 2014).. Also even the base for the current state of the art models on SimLex-999, Paragram (TACL 2015), used paraphrases created by pivoting on parallel data.\n\nI think the main contributions of this paper are modifying Sent2vec so it can be used on bilingual data and using it to learn nice representations for words and sentences (and documents). I think that to be published, it should clearly outperform and/or have advantages over all previous works - and this is not clear from this paper in its current form. It is okay if it doesn't do the best on everything, but it is hard to tell how this work currently fits into the literature especially in terms of the sentence-level tasks which are a focus.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571701527035}], "openreview_url": "https://openreview.net/forum?id=SJlJSaEFwS", "arxiv_id": "1912.12481", "paper_pdf": "papers/SJlJSaEFwS.pdf", "paper_pdf_sha256": "31b33466673e3ebfb027e3dbfc334e1d8ee56e3611209bdf7cc5e04ac2d2e9ac", "paper_pdf_bytes": 300711, "paper_pdf_source": "openreview", "code_url": "https://github.com/epfml/Bi-Sent2Vec", "code_repository": "epfml/Bi-Sent2Vec", "code_commit": "b3bf758f86a7006112a7a65bf9955d88ef3465d7", "code_archive": "repos/SJlJSaEFwS.zip", "code_archive_sha256": "62849656fb4d22a4428fcfb2e67fcf3349ece2b3ff3321f33f8581c21832eb0f", "code_archive_bytes": 40961, "code_file_count": 24, "code_extensions": {".h": 12, ".cc": 11, ".py": 1}, "github_disk_usage_kb": 52, "github_languages": {"C++": 119853, "Python": 5657, "Makefile": 1647}, "github_archived": false, "github_pushed_at": "2020-06-27T17:28:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/robust-cross-lingual-embeddings-from-parallel-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ByN7Yo05YX", "year": 2019, "status": "rejected", "title": "Adaptive Neural Trees", "authors": ["Ryutaro Tanno", "Kai Arulkumaran", "Daniel C. Alexander", "Antonio Criminisi", "Aditya Nori"], "authorids": ["ryutaro.tanno.15@ucl.ac.uk", "kailash.arulkumaran13@imperial.ac.uk", "d.alexander@ucl.ac.uk", "antcrim@microsoft.com", "adityan@microsoft.com"], "authors_source": "OpenReview API", "abstract": "Deep neural networks and decision trees operate on largely separate paradigms; typically, the former performs representation learning with pre-specified architectures, while the latter is characterised by learning hierarchies over pre-specified features with data-driven architectures. We unite the two via adaptive neural trees (ANTs), a model that incorporates representation learning into edges, routing functions and leaf nodes of a decision tree, along with a backpropagation-based training algorithm that adaptively grows the architecture from primitive modules (e.g., convolutional layers). ANTs allow increased interpretability via hierarchical clustering, e.g., learning meaningful class associations, such as separating natural vs. man-made objects. We demonstrate this on classification and regression tasks, achieving over 99% and 90% accuracy on the MNIST and CIFAR-10 datasets, and outperforming standard neural networks, random forests and gradient boosted trees on the SARCOS dataset. Furthermore, ANT optimisation naturally adapts the architecture to the size and complexity of the training data.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SklzsP6ph7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper433/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The presented method is a generalization of a number of existing methods, which can be regarded as special cases. Overall the method seems to be novel. Meanwhile, I have two major questions:\nTo account for the bias issue, instead of a single DT, ensemble methods such as random forests are the popular choices. How ANT could benefit from relying on a single DT instead of a random forest type?\nThe datasets of MNIST and CIFAR-10 are used for many years and the performance is already saturated. As presented in Table.3, the performance of the proposed method is also not the best on either of the tested datasets. Please clearly elaborate on why and how to address this issue. It would be more interesting and meaningful to work with a more recent large datasets, such as ImageNet or MS COCO. \n\nThe response does not fully address my concerns. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposes the Adaptive Neural Trees (ANT) approach to combine the two learning paradigms of deep neural nets and decision trees (DT).", "review": "The presented method is a generalization of a number of existing methods, which can be regarded as special cases. Overall the method seems to be novel. Meanwhile, I have two major questions:\nTo account for the bias issue, instead of a single DT, ensemble methods such as random forests are the popular choices. How ANT could benefit from relying on a single DT instead of a random forest type?\nThe datasets of MNIST and CIFAR-10 are used for many years and the performance is already saturated. As presented in Table.3, the performance of the proposed method is also not the best on either of the tested datasets. Please clearly elaborate on why and how to address this issue. It would be more interesting and meaningful to work with a more recent large datasets, such as ImageNet or MS COCO. \n\nThe response does not fully address my concerns. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541425050168}, {"id": "H1epcjMThm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper433/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is written rather well, however I find the experiments incomplete and have some reservations about the\nmethod. My main points of critique are:\n\n1.  Combining DT & NN \nI have doubts that the way you combine DT &NN  you get the \"Best of both world\". In some ways your architecture also \nshares disadvantages of both:\n\n1.1 Interpretability\nBecause each node in the tree can a neural network (with arbritrary complexity), this approach looses one central advantage of DT, that is the interpretability of the result.    Each node in the tree can perform arbritrary complex (and hierarchical) \ncomputations. The authors only show one particular example (Fig. 2a), where the model has learned is a reasonable \nstructure.\n\n1.2 Complexity:\nThe whole architecture is much more complex than either a neural network or a decision tree. I expect that therefore training these is not easy, and expert knowledge in either DT  or NN may not be enough to use this model.\n\n\n2. Limited experiments\n\n2.1 The authors only consider 2 experiments from vision (MNIST & CIFAR 10) while proposing a universal method.  To show universality the authors should use data sets from different domains (e.g UCI data sets)\n\n2.2 The authors argue that a  strength of the method   is  that it uses a low number of parameters on average for a forward path (compared to the total parameter size).  I don't find this argument to be convincing. In the limit this would imply a high memorization of the  data.  Also, a similar case can be made for standard CNN, when a particular filter is mostly inactive for some data points.\n\n2.3 The interpretability of DT compared to NN I mentioned earlier.  To make the argument that their method learns the\nhierarchical structure of the data , the authors should have added experiments to support this, where  such a hierarchical structure is clearly present and can be evaluated empirically.\n\n\n--\n\nIn light of the extended experiments w.r.t. to 2.1 I increased my score from 5 to 6.  Overall, I still have doubts about the interpretability and complexity of the proposed method.  \n\nComplexity:  \"but all the intuitions needed would come solely from training NN\".    I disagree with this response.   The architecture is a mix between a tree (hard, decision-tree like error surface,  non-local) and neural network (smooth, mostly convex error surface). This also implies that the training process and its behavior will possess patterns and challenges of both approaches. \n\nInterpretability:  I think the method misses \"priors\" that enforce credit assignment.  Partitioning the problem in subp-roblems should be done via the tree components, whereas processing (such as image filtering) should be done in the network nodes. However,  the method does not enforce, or encourage this behavior, for instance\nvia constraints:   also nodes can do partitioning (because neural networks can approximate decision trees)  and edges can do processing (e.g. decisions-trees can be used for mnist).\n\nSo I still believe this to be a borderline paper, however, the experiments support a more general applicability.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "limited experiments, doubts about the method", "review": "The paper is written rather well, however I find the experiments incomplete and have some reservations about the\nmethod. My main points of critique are:\n\n1.  Combining DT & NN \nI have doubts that the way you combine DT &NN  you get the \"Best of both world\". In some ways your architecture also \nshares disadvantages of both:\n\n1.1 Interpretability\nBecause each node in the tree can a neural network (with arbritrary complexity), this approach looses one central advantage of DT, that is the interpretability of the result.    Each node in the tree can perform arbritrary complex (and hierarchical) \ncomputations. The authors only show one particular example (Fig. 2a), where the model has learned is a reasonable \nstructure.\n\n1.2 Complexity:\nThe whole architecture is much more complex than either a neural network or a decision tree. I expect that therefore training these is not easy, and expert knowledge in either DT  or NN may not be enough to use this model.\n\n\n2. Limited experiments\n\n2.1 The authors only consider 2 experiments from vision (MNIST & CIFAR 10) while proposing a universal method.  To show universality the authors should use data sets from different domains (e.g UCI data sets)\n\n2.2 The authors argue that a  strength of the method   is  that it uses a low number of parameters on average for a forward path (compared to the total parameter size).  I don't find this argument to be convincing. In the limit this would imply a high memorization of the  data.  Also, a similar case can be made for standard CNN, when a particular filter is mostly inactive for some data points.\n\n2.3 The interpretability of DT compared to NN I mentioned earlier.  To make the argument that their method learns the\nhierarchical structure of the data , the authors should have added experiments to support this, where  such a hierarchical structure is clearly present and can be evaluated empirically.\n\n\n--\n\nIn light of the extended experiments w.r.t. to 2.1 I increased my score from 5 to 6.  Overall, I still have doubts about the interpretability and complexity of the proposed method.  \n\nComplexity:  \"but all the intuitions needed would come solely from training NN\".    I disagree with this response.   The architecture is a mix between a tree (hard, decision-tree like error surface,  non-local) and neural network (smooth, mostly convex error surface). This also implies that the training process and its behavior will possess patterns and challenges of both approaches. \n\nInterpretability:  I think the method misses \"priors\" that enforce credit assignment.  Partitioning the problem in subp-roblems should be done via the tree components, whereas processing (such as image filtering) should be done in the network nodes. However,  the method does not enforce, or encourage this behavior, for instance\nvia constraints:   also nodes can do partitioning (because neural networks can approximate decision trees)  and edges can do processing (e.g. decisions-trees can be used for mnist).\n\nSo I still believe this to be a borderline paper, however, the experiments support a more general applicability.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541381013466}, {"id": "HJxPbVzH27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper433/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe authors proposed a new model Adaptive Neural Trees(ANTs) by combining the representation learning and gradient optimization of neural networks with architecture learning of decision trees. The key advantage of the new model ANTs  over the existing methods(Random forest, Linear classifier, Neural decision forest, et al) is: it may achieve high accuracy(above $90\\%$) with relatively much smaller number of parameters, as shown by the experiments on the datasets MNIST and CIFAR-10. Besides, the authors proposed single-path inference based on the greedily-selected leaf node to approximate the multi-path inferences with the full predictive distribution. The experiments show the single-path inference doesn't lose much accuracy but it saves memory and time. This paper is acceptable after minor modification.\n\n\nQuestions:\nIn the second line below equation (1), $n$ in $t_{e_{n(j)}}^{\\psi}$ is not defined. Also, should $t_{e_{1}}^{\\psi}$ be $t_{e_{n(1)}}^{\\psi}$? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "acceptable", "review": "\nThe authors proposed a new model Adaptive Neural Trees(ANTs) by combining the representation learning and gradient optimization of neural networks with architecture learning of decision trees. The key advantage of the new model ANTs  over the existing methods(Random forest, Linear classifier, Neural decision forest, et al) is: it may achieve high accuracy(above $90\\%$) with relatively much smaller number of parameters, as shown by the experiments on the datasets MNIST and CIFAR-10. Besides, the authors proposed single-path inference based on the greedily-selected leaf node to approximate the multi-path inferences with the full predictive distribution. The experiments show the single-path inference doesn't lose much accuracy but it saves memory and time. This paper is acceptable after minor modification.\n\n\nQuestions:\nIn the second line below equation (1), $n$ in $t_{e_{n(j)}}^{\\psi}$ is not defined. Also, should $t_{e_{1}}^{\\psi}$ be $t_{e_{n(1)}}^{\\psi}$? ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1540854783104}], "openreview_url": "https://openreview.net/forum?id=ByN7Yo05YX", "arxiv_id": "1807.06699", "paper_pdf": "papers/ByN7Yo05YX.pdf", "paper_pdf_sha256": "9be575e73fe2dda1a1b86a32502d469dbaab3109e56e95337d028c47271b3ce6", "paper_pdf_bytes": 1975257, "paper_pdf_source": "openreview", "code_url": "https://github.com/rtanno21609/AdaptiveNeuralTrees", "code_repository": "rtanno21609/AdaptiveNeuralTrees", "code_commit": "8d7835fd21b16e43801ef20ad530380057f53a0c", "code_archive": "repos/ByN7Yo05YX.zip", "code_archive_sha256": "ef03d66c454a8cbd7a46a4a93f7c57e70a4c4faaf6d876e38f96ec7802c8a422", "code_archive_bytes": 482040, "code_file_count": 11, "code_extensions": {".py": 6, ".ipynb": 4, ".sh": 1}, "github_disk_usage_kb": 279, "github_languages": {"Jupyter Notebook": 550686, "Python": 137273, "Shell": 777}, "github_archived": false, "github_pushed_at": "2019-07-03T10:52:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-neural-trees"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJQO7UJCW", "year": 2018, "status": "rejected", "title": "Adversarial Learning for Semi-Supervised Semantic Segmentation", "authors": ["Wei-Chih Hung", "Yi-Hsuan Tsai", "Yan-Ting Liou", "Yen-Yu Lin", "Ming-Hsuan Yang"], "authorids": ["whung8@ucmerced.edu", "ytsai@nec-labs.com", "lyt@csie.ntu.edu.tw", "yylin@citi.sinica.edu.tw", "mhyang@ucmerced.edu"], "authors_source": "OpenReview API", "abstract": "We propose a method for semi-supervised semantic segmentation using the adversarial network. While most existing discriminators are trained to classify input images as real or fake on the image level, we design a discriminator in a fully convolutional manner to differentiate the predicted probability maps from the ground truth segmentation distribution with the consideration of the spatial resolution. We show that the proposed discriminator can be used to improve the performance on semantic segmentation by coupling the adversarial loss with the standard cross entropy loss on the segmentation network. In addition, the fully convolutional discriminator enables the semi-supervised learning through discovering the trustworthy regions in prediction results of unlabeled images, providing additional supervisory signals. In contrast to existing methods that utilize weakly-labeled images, our method leverages unlabeled images without any annotation to enhance the segmentation model. Experimental results on both the PASCAL VOC 2012 dataset and the Cityscapes dataset demonstrate the effectiveness of our algorithm.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJRdYLhgM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper125/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an alternative adversarial loss function for image segmentation, and an additional loss for unlabeled images.\n\n+ well written\n+ good evaluation\n+ good performance compared to prior state of art\n- technical novelty\n- semi-supervised loss does not yield significant improvement\n- missing citations and comparisons\n\nThe paper is well written, structured, and easy to read.\nThe experimental section is extensive, and shows a significant improvement over prior state of the art in semi-supervised learning.\nUnfortunately, it is unclear what exactly lead to this performance increase. Is it a better baseline model? Is the algorithm tuned better, or is there something fundamentally different compared to prior work (e.g. Luc 2016).\n\nFinally, it would help if the authors could highlight their technical difference compared to prior work. The presented adversarial loss is similar to Luc 2016 and \"Image-to-Image Translation with Conditional Adversarial Networks, Isola etal 2017\". What is different, and why is it important?\nThe semi-supervised loss is similar to Pathak 2015a, it would help to highlight the difference, and show experimentally why it matters.\n\nIn summary, the authors should highlight the difference to prior work, and show why the proposed changes matter.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "rating": "5: Marginally below acceptance threshold", "review": "The paper presents an alternative adversarial loss function for image segmentation, and an additional loss for unlabeled images.\n\n+ well written\n+ good evaluation\n+ good performance compared to prior state of art\n- technical novelty\n- semi-supervised loss does not yield significant improvement\n- missing citations and comparisons\n\nThe paper is well written, structured, and easy to read.\nThe experimental section is extensive, and shows a significant improvement over prior state of the art in semi-supervised learning.\nUnfortunately, it is unclear what exactly lead to this performance increase. Is it a better baseline model? Is the algorithm tuned better, or is there something fundamentally different compared to prior work (e.g. Luc 2016).\n\nFinally, it would help if the authors could highlight their technical difference compared to prior work. The presented adversarial loss is similar to Luc 2016 and \"Image-to-Image Translation with Conditional Adversarial Networks, Isola etal 2017\". What is different, and why is it important?\nThe semi-supervised loss is similar to Pathak 2015a, it would help to highlight the difference, and show experimentally why it matters.\n\nIn summary, the authors should highlight the difference to prior work, and show why the proposed changes matter.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511971190420}, {"id": "H1Op4eqlM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper125/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed an approach for semi-supervised semantic segmentation based on adversarial training. Built upon a popular segmentation network, the paper integrated adversarial loss to incorporate unlabeled examples in training. The outputs from the discriminator are interpreted as indicators for the reliability of label prediction, and used to filter-out non-reliable predictions as augmented training data from unlabeled images.  The proposed method achieved consistent improvement over existing state-of-the-art on two challenging segmentation datasets.\n\nAlthough the motivation is reasonable and the results are impressive, there are some parts that need more clarification/discussion as described below.\n\n1) Robustness of discriminator output:\nThe main contribution of the proposed model is exploiting the outputs from the discriminator as the confidence score maps of the predicted segmentation labels. However, the outputs from the discriminator indicate whether its inputs are from ground-truth labels or model predictions, and may not be directly related to ‘correctness’ of the label prediction. For instance, it may prefer per-pixel score vectors closed to one-hot encoded vectors. More thorough analysis/discussions are required to show how outputs from discriminator are correlated with the correctness of label prediction.     \n\n2) Design of discriminator\nI wonder if conditional discriminator fits better for the task. i.e. D(X,P) instead of D(P). It may prevent the model generating label prediction P non-relevant to input X by adversarial training, and makes the score prediction from the discriminator more meaningful. Some ablation study or discussions would be helpful.\n\n3) Presentations\nThere are several abused notations; notations for the ground-truth label P and the prediction from the generator S(X) should be clearly separated in Eq. (1) and (4). Also, it would better to find a better notation for the outputs from D instead of D^(*,0) and D^(*,1).  \nTraining details in semi-supervised learning would be helpful. For instance, the proposed semi-supervised learning strategy based on Eq. (5) may be suffered by noise outputs from the discriminator in early training stages. I wonder how authors resolved the issues (e.g. training the generator and discriminator are with the labeled example first and extending it to training with unlabeled data.) \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "No title", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposed an approach for semi-supervised semantic segmentation based on adversarial training. Built upon a popular segmentation network, the paper integrated adversarial loss to incorporate unlabeled examples in training. The outputs from the discriminator are interpreted as indicators for the reliability of label prediction, and used to filter-out non-reliable predictions as augmented training data from unlabeled images.  The proposed method achieved consistent improvement over existing state-of-the-art on two challenging segmentation datasets.\n\nAlthough the motivation is reasonable and the results are impressive, there are some parts that need more clarification/discussion as described below.\n\n1) Robustness of discriminator output:\nThe main contribution of the proposed model is exploiting the outputs from the discriminator as the confidence score maps of the predicted segmentation labels. However, the outputs from the discriminator indicate whether its inputs are from ground-truth labels or model predictions, and may not be directly related to ‘correctness’ of the label prediction. For instance, it may prefer per-pixel score vectors closed to one-hot encoded vectors. More thorough analysis/discussions are required to show how outputs from discriminator are correlated with the correctness of label prediction.     \n\n2) Design of discriminator\nI wonder if conditional discriminator fits better for the task. i.e. D(X,P) instead of D(P). It may prevent the model generating label prediction P non-relevant to input X by adversarial training, and makes the score prediction from the discriminator more meaningful. Some ablation study or discussions would be helpful.\n\n3) Presentations\nThere are several abused notations; notations for the ground-truth label P and the prediction from the generator S(X) should be clearly separated in Eq. (1) and (4). Also, it would better to find a better notation for the outputs from D instead of D^(*,0) and D^(*,1).  \nTraining details in semi-supervised learning would be helpful. For instance, the proposed semi-supervised learning strategy based on Eq. (5) may be suffered by noise outputs from the discriminator in early training stages. I wonder how authors resolved the issues (e.g. training the generator and discriminator are with the labeled example first and extending it to training with unlabeled data.) \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511814336091}, {"id": "r1RDwROeG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper125/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes techniques for training semantic segmentation networks. There are two key ideas:\n\n- Attach a pixel-level GAN loss to the output semantic segmentation map. That is, add a discriminator network that decides whether each pixel in the label map belongs to a real label map or not. Of course, this loss alone is unaware of the input image and would drive the network to produce plausible label maps that have no relation to the input image. An additional cross-entropy loss (the standard semantic segmentation loss) is used to tie the network to the input and the ground-truth label map, when available.\n\n- Additional unlabeled data is utilized by using a trained semantic segmentation network to produce a label map with associated confidences; high-confidence pixels are used as ground-truth labels and are fed back to the network as training data.\n\nThe paper is fine and the work is competently done, but the experimental results never quite come together. The technical development isn’t surprising and doesn’t have much to teach researchers working in the area. Given that the technical novelty is rather light and the experimental benefits are not quite there, I cannot recommend the paper for publication in a first-tier conference.\n\nSome more detailed comments:\n\n1. The GAN and the semi-supervised training scheme appear to be largely independent. The GAN can be applied without any unlabeled data, for example. The paper generally appears to present two largely independent ideas. This is fine, except they don’t convincingly pan out in experiments.\n\n2. The biggest issue is that the experimental results do not convincingly indicate that the presented ideas are useful.\n2a. In the “Full” condition, the presented approach does not come close to the performance of the DeepLab baseline, even though the DeepLab network is used in the presented approach. Perhaps the authors have taken out some components of the DeepLab scheme for these experiments, such as multi-scale processing, but the question then is “Why?”. These components are not illegal, they are not cheating, they are not overly complex and are widely used. If the authors cannot demonstrate an improvement with these components, their ideas are unlikely to be adopted in state-of-the-art semantic systems, which do use these components and are doing fine.\n2b. In the 1/8, 1/4, and 1/2 conditions, the performance of the baselines is not quoted. This is wrong. Since the authors are evaluating on the validation sets, there is no reason not to train the baselines on the same amount of labeled data (1/8, 1/4, 1/2) and report the results. The training scripts are widely available and such training of baselines for controlled experiments is commonly done in the literature. The reviewer is left to suspect, with no evidence given to the contrary, that the presented approach does not outperform the DeepLab baseline even in the reduced-data conditions.\n\nA somewhat unflattering view of the work would be that this is another example of throwing a GAN at everything to see if it sticks. In this case, the experiments do not indicate that it did.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "not enough for a first-tier conference", "rating": "5: Marginally below acceptance threshold", "review": "This paper describes techniques for training semantic segmentation networks. There are two key ideas:\n\n- Attach a pixel-level GAN loss to the output semantic segmentation map. That is, add a discriminator network that decides whether each pixel in the label map belongs to a real label map or not. Of course, this loss alone is unaware of the input image and would drive the network to produce plausible label maps that have no relation to the input image. An additional cross-entropy loss (the standard semantic segmentation loss) is used to tie the network to the input and the ground-truth label map, when available.\n\n- Additional unlabeled data is utilized by using a trained semantic segmentation network to produce a label map with associated confidences; high-confidence pixels are used as ground-truth labels and are fed back to the network as training data.\n\nThe paper is fine and the work is competently done, but the experimental results never quite come together. The technical development isn’t surprising and doesn’t have much to teach researchers working in the area. Given that the technical novelty is rather light and the experimental benefits are not quite there, I cannot recommend the paper for publication in a first-tier conference.\n\nSome more detailed comments:\n\n1. The GAN and the semi-supervised training scheme appear to be largely independent. The GAN can be applied without any unlabeled data, for example. The paper generally appears to present two largely independent ideas. This is fine, except they don’t convincingly pan out in experiments.\n\n2. The biggest issue is that the experimental results do not convincingly indicate that the presented ideas are useful.\n2a. In the “Full” condition, the presented approach does not come close to the performance of the DeepLab baseline, even though the DeepLab network is used in the presented approach. Perhaps the authors have taken out some components of the DeepLab scheme for these experiments, such as multi-scale processing, but the question then is “Why?”. These components are not illegal, they are not cheating, they are not overly complex and are widely used. If the authors cannot demonstrate an improvement with these components, their ideas are unlikely to be adopted in state-of-the-art semantic systems, which do use these components and are doing fine.\n2b. In the 1/8, 1/4, and 1/2 conditions, the performance of the baselines is not quoted. This is wrong. Since the authors are evaluating on the validation sets, there is no reason not to train the baselines on the same amount of labeled data (1/8, 1/4, 1/2) and report the results. The training scripts are widely available and such training of baselines for controlled experiments is commonly done in the literature. The reviewer is left to suspect, with no evidence given to the contrary, that the presented approach does not outperform the DeepLab baseline even in the reduced-data conditions.\n\nA somewhat unflattering view of the work would be that this is another example of throwing a GAN at everything to see if it sticks. In this case, the experiments do not indicate that it did.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511741286155}], "openreview_url": "https://openreview.net/forum?id=SJQO7UJCW", "arxiv_id": "1802.07934", "paper_pdf": "papers/SJQO7UJCW.pdf", "paper_pdf_sha256": "504cc9935023e0374bf77bd040099a15014757a8b7938c56cf6728ace804c08c", "paper_pdf_bytes": 4570902, "paper_pdf_source": "openreview", "code_url": "https://github.com/hfslyc/AdvSemiSeg", "code_repository": "hfslyc/AdvSemiSeg", "code_commit": "841d546c927605747594be726622363bf781cefb", "code_archive": "repos/SJQO7UJCW.zip", "code_archive_sha256": "e72f40de01137b360bf1f1ec7e7601d5aee87d6ab4d1abcbb595aabfa66532e8", "code_archive_bytes": 849235, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 834, "github_languages": {"Python": 49841}, "github_archived": false, "github_pushed_at": "2021-04-21T10:48:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-learning-for-semi-supervised"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eChWBrh9mc", "year": 2026, "status": "rejected", "title": "EntryPrune: Neural Network Feature Selection using First Impressions", "authors": ["Felix Zimmer", "Patrik Okanovic", "Torsten Hoefler"], "authorids": ["~Felix_Zimmer1", "~Patrik_Okanovic1", "~Torsten_Hoefler1"], "authors_source": "OpenReview API", "abstract": "There is an ongoing effort to develop feature selection algorithms to improve interpretability, reduce computational resources, and minimize overfitting in predictive models. Neural networks stand out as architectures on which to build feature selection methods, and recently, neuron pruning and regrowth have emerged from the sparse neural network literature as promising new tools. We introduce EntryPrune, a novel supervised feature selection algorithm using a dense neural network with a dynamic sparse input layer. It employs entry-based pruning, a novel approach that compares neurons based on their relative change induced when they have entered the network. Extensive experiments on 13 different datasets show that our approach generally outperforms the current state-of-the-art methods, and in particular improves the average accuracy on low-dimensional datasets. Furthermore, we show that EntryPruning surpasses traditional techniques such as magnitude pruning within the EntryPrune framework and that EntryPrune achieves lower runtime than competing approaches. Our code is available in the supplementary material.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "L0n0X0vksr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16314/Reviewer_3ddh"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "The paper introduces EntryPrune, a feature selection method for neural networks that uses a dynamically sparse input layer and entry-based pruning. The key idea is to evaluate features based on their initial impact when they first enter the network. The paper reports that EntryPrune outperforms established methods like NeuroFS and LassoNet in terms of accuracy and runtime efficiency.", "review_text": "The paper introduces EntryPrune, a feature selection method for neural networks that uses a dynamically sparse input layer and entry-based pruning. The key idea is to evaluate features based on their initial impact when they first enter the network. The paper reports that EntryPrune outperforms established methods like NeuroFS and LassoNet in terms of accuracy and runtime efficiency.", "strengths": "The introduction of entry-based pruning; measuring the initial impact is reasonable and normalization method used in the techniques ensures fair comparison. Results show some advantage in runtime in long datasets and marginal improvement over the existing techniques.", "weaknesses": "1.\tThe core contribution of entry-based pruning is incremental at best. The main parts of the technique, gradient based regrowth and pruning largely borrows from prior works like NeuroFS and RigL. The entry-based pruning technique is simply a minor adaptation rather than a truly novel contribution to the field.\n2.\tThe experimental setup with MLP of 1 hidden layer with 100 neurons and large network containing two layers, is too basic and fails to offer a convincing benchmark for current applicability of the method. \n3.\tIt is unclear how the pruning mechanism work in the case of multi layered networks and whether is it applicable to other than dense layers such as convolutional or residual.\n4.\tBaselines are too old (the latest is one from 2023). I suggest adding atleast GradEnFS (2024) and more to establish EntryPrune’s place in the current literature landscape. \n5.\tThe results are, at best, marginal and for the challenging dataset (Cifar-100), NeuroFS is performing better than the proposed consistently. Also, the reported accuracy ranges of these experiments (~40% for Cifar-10 and ~20% for Cifar-100) raises concerns about the practical significance of such techniques. \n6.\tThere is a significant lack of practical benefit demonstrated. The paper fails to provide clear real-world scenarios where EntryPrune would offer significant advantages over simpler, well-established feature selection methods other than special cases e.g. for interpretability.", "questions": "1.\tCan you clearly highlight how EntryPrune differ from the existing works in a more substantial way, beyond the use of early batch scoring ?\n2.\tCould entry scores after some rotations drift over time? \n3.\tThe ablation in Figure 8 show sensitivity to hyperparameters. Can you provide a principled approach or a rule of thumb to set the hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces EntryPrune, a feature selection method for neural networks that uses a dynamically sparse input layer and entry-based pruning. The key idea is to evaluate features based on their initial impact when they first enter the network. The paper reports that EntryPrune outperforms established methods like NeuroFS and LassoNet in terms of accuracy and runtime efficiency.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "The introduction of entry-based pruning; measuring the initial impact is reasonable and normalization method used in the techniques ensures fair comparison. Results show some advantage in runtime in long datasets and marginal improvement over the existing techniques.", "weaknesses": "1.\tThe core contribution of entry-based pruning is incremental at best. The main parts of the technique, gradient based regrowth and pruning largely borrows from prior works like NeuroFS and RigL. The entry-based pruning technique is simply a minor adaptation rather than a truly novel contribution to the field.\n2.\tThe experimental setup with MLP of 1 hidden layer with 100 neurons and large network containing two layers, is too basic and fails to offer a convincing benchmark for current applicability of the method. \n3.\tIt is unclear how the pruning mechanism work in the case of multi layered networks and whether is it applicable to other than dense layers such as convolutional or residual.\n4.\tBaselines are too old (the latest is one from 2023). I suggest adding atleast GradEnFS (2024) and more to establish EntryPrune’s place in the current literature landscape. \n5.\tThe results are, at best, marginal and for the challenging dataset (Cifar-100), NeuroFS is performing better than the proposed consistently. Also, the reported accuracy ranges of these experiments (~40% for Cifar-10 and ~20% for Cifar-100) raises concerns about the practical significance of such techniques. \n6.\tThere is a significant lack of practical benefit demonstrated. The paper fails to provide clear real-world scenarios where EntryPrune would offer significant advantages over simpler, well-established feature selection methods other than special cases e.g. for interpretability.", "questions": "1.\tCan you clearly highlight how EntryPrune differ from the existing works in a more substantial way, beyond the use of early batch scoring ?\n2.\tCould entry scores after some rotations drift over time? \n3.\tThe ablation in Figure 8 show sensitivity to hyperparameters. Can you provide a principled approach or a rule of thumb to set the hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761977795454}, {"id": "lqBHqREVCN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16314/Reviewer_Jpgm"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This manuscript proposes a supervised feature selection algorithm called EntryPrune. The algorithm is based on a dense neural network with a dynamically sparse input layer. The core mechanism is claimed to be \"entry-based pruning\", evaluating the importance of a neuron(feature) based on the relative change induced when it first enters the network, combined with a random regrowth strategy. The authors claim that this method outperforms (or matches) sota on 13 datasets, especially on \"long\" datasets.", "review_text": "This manuscript proposes a supervised feature selection algorithm called EntryPrune. The algorithm is based on a dense neural network with a dynamically sparse input layer. The core mechanism is claimed to be \"entry-based pruning\", evaluating the importance of a neuron(feature) based on the relative change induced when it first enters the network, combined with a random regrowth strategy. The authors claim that this method outperforms (or matches) sota on 13 datasets, especially on \"long\" datasets.", "strengths": "1. The pruning approach proposed in this paper is an interesting heuristic, which attempt to address the issue of unfair evaluation time between new and old neurons in dynamic sparse training.\n2. Compared to NeuroFS and LassoNet, the proposed method may have lower computation time while maintaining comparable performance.", "weaknesses": "1. Dynamic sparse training is a widely researched and used approach. The method proposed in this manuscript is more like an incremental improvement on the existing NeuroFS framework. The most creative part is the introduction of a new pruning metric strategy. In addition, this manuscript avoids any theoretical analysis of its effectiveness.\n\n2. The manuscript seems to have ignored GBDT baselines e.g., xgboost and catboost, in the main text, and appendix B also seems to show GBDT's powerful ability to identify interactive features.\n\n3. The manuscript argues that \"random regrowth\" is superior to \"gradient-based regrowth\" because the former can discover \"interaction features.\" However, the only evidence for this claim comes from a toy example in Appendix B, lacking studies on more real-world datasets .\n\n4. Fig.11 shows that the feature subset selected by the proposed method has low stability in multiple runs, which means that the reliability of this method may be highly dependent on hyperparameter changes. This is unacceptable in fields such as healthcare and finance.\n\n5. While the method proposed in the manuscript performs well on homogeneous datasets such as images and speech, it performs poorly on many so-called \"wide datasets\" (where the number of features is greater than the number of samples, such as ARCENE and GLA-BRA-180). The paper's claim of \"better overall performance\" is inconsistent with the data.", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript proposes a supervised feature selection algorithm called EntryPrune. The algorithm is based on a dense neural network with a dynamically sparse input layer. The core mechanism is claimed to be \"entry-based pruning\", evaluating the importance of a neuron(feature) based on the relative change induced when it first enters the network, combined with a random regrowth strategy. The authors claim that this method outperforms (or matches) sota on 13 datasets, especially on \"long\" datasets.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The pruning approach proposed in this paper is an interesting heuristic, which attempt to address the issue of unfair evaluation time between new and old neurons in dynamic sparse training.\n2. Compared to NeuroFS and LassoNet, the proposed method may have lower computation time while maintaining comparable performance.", "weaknesses": "1. Dynamic sparse training is a widely researched and used approach. The method proposed in this manuscript is more like an incremental improvement on the existing NeuroFS framework. The most creative part is the introduction of a new pruning metric strategy. In addition, this manuscript avoids any theoretical analysis of its effectiveness.\n\n2. The manuscript seems to have ignored GBDT baselines e.g., xgboost and catboost, in the main text, and appendix B also seems to show GBDT's powerful ability to identify interactive features.\n\n3. The manuscript argues that \"random regrowth\" is superior to \"gradient-based regrowth\" because the former can discover \"interaction features.\" However, the only evidence for this claim comes from a toy example in Appendix B, lacking studies on more real-world datasets .\n\n4. Fig.11 shows that the feature subset selected by the proposed method has low stability in multiple runs, which means that the reliability of this method may be highly dependent on hyperparameter changes. This is unacceptable in fields such as healthcare and finance.\n\n5. While the method proposed in the manuscript performs well on homogeneous datasets such as images and speech, it performs poorly on many so-called \"wide datasets\" (where the number of features is greater than the number of samples, such as ARCENE and GLA-BRA-180). The paper's claim of \"better overall performance\" is inconsistent with the data.", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761914710593}, {"id": "rdNFCRh1We", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16314/Reviewer_Nw9T"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes EntryPrune, a feature selection method that combines a dense neural network with a dynamically sparse input layer, where features are iteratively pruned and regrown. The main idea is an entry-based pruning metric that evaluates each new feature based on its early gradient-driven impact. The method is benchmarked on 10+ datasets, and (according to the results in the manuscript) it shows strong performance particularly on datasets with more samples than features.", "review_text": "This paper proposes EntryPrune, a feature selection method that combines a dense neural network with a dynamically sparse input layer, where features are iteratively pruned and regrown. The main idea is an entry-based pruning metric that evaluates each new feature based on its early gradient-driven impact. The method is benchmarked on 10+ datasets, and (according to the results in the manuscript) it shows strong performance particularly on datasets with more samples than features.", "strengths": "- Novel method\n- Good experimental coverage (13 datasets) \n- The paper is well written and easy to follow", "weaknesses": "- Although the paper motivates feature selection as a path to interpretability, it does not connect its contribution to established explainability methods such as SHAP, LIME, Integrated Gradients, or Grad-CAM. Given that the method relies on gradients and is applied to image data, this omission weakens the interpretability claim\n\n- On wide datasets, the method offers little to no improvement over existing baselines\n\n- The experimental setup relies mainly on SVMs, which are somewhat outdated, incorporating more modern models such as GBDTs or Random Forests would provide a fairer and more relevant comparison\n\n- The evaluation also omits tree-based feature selection baselines (e.g., feature importance from Random Forest or GBDT), which are widely used in practice and should at least be discussed", "questions": "- How does EntryPrune compare to attribution-based interpretability methods (e.g., SHAP, LIME, Grad-CAM, ...)? Could the authors clarify whether EntryPrune should be viewed as a competing interpretability approach or a complementary one?\n- Please see the \"Weaknesses\" section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes EntryPrune, a feature selection method that combines a dense neural network with a dynamically sparse input layer, where features are iteratively pruned and regrown. The main idea is an entry-based pruning metric that evaluates each new feature based on its early gradient-driven impact. The method is benchmarked on 10+ datasets, and (according to the results in the manuscript) it shows strong performance particularly on datasets with more samples than features.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- Novel method\n- Good experimental coverage (13 datasets) \n- The paper is well written and easy to follow", "weaknesses": "- Although the paper motivates feature selection as a path to interpretability, it does not connect its contribution to established explainability methods such as SHAP, LIME, Integrated Gradients, or Grad-CAM. Given that the method relies on gradients and is applied to image data, this omission weakens the interpretability claim\n\n- On wide datasets, the method offers little to no improvement over existing baselines\n\n- The experimental setup relies mainly on SVMs, which are somewhat outdated, incorporating more modern models such as GBDTs or Random Forests would provide a fairer and more relevant comparison\n\n- The evaluation also omits tree-based feature selection baselines (e.g., feature importance from Random Forest or GBDT), which are widely used in practice and should at least be discussed", "questions": "- How does EntryPrune compare to attribution-based interpretability methods (e.g., SHAP, LIME, Grad-CAM, ...)? Could the authors clarify whether EntryPrune should be viewed as a competing interpretability approach or a complementary one?\n- Please see the \"Weaknesses\" section", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761730672635}, {"id": "5JgD6proLI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16314/Reviewer_osxp"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The authors propose EntryPrune, an embedded feature-selection method built on a dense MLP with a dynamically sparse input layer. The key idea is an entry-based pruning criterion. Specifically, when a pruned input feature is (re)introduced as a candidate into the input layer, the algorithm measures the short-horizon relative change it induces. This is operationalized as the z-scored L1-norm of its first-layer gradient accumulations over the first n mini-batches post-entry. These accumulated scores are stored for all features. Periodically, the algorithm identifies the K features with the largest stored entry scores and prunes the rest. Then, it randomly regrows a batch of new candidate inputs whose incoming weights are reinitialized to tiny values. According to the authors, this strategy balances between established and freshly (re)introduced features and mitigates the bias of magnitude-based pruning toward long-tenured inputs. \nEmpirically, across 13 datasets (tabular, vision, speech, genomics, and text), EntryPrune often outperforms or matches SotA feature selection baselines (NeuroFS, LassoNet, STG, QS, RFS, etc.). An ablation suggests that entry-based scoring (short-horizon gradient sums) beats magnitude‐based criteria inside the same framework. The proposed method often runs faster wall-clock than a sparse-training baseline (NeuroFS). The code and replication scripts are provided.", "review_text": "The authors propose EntryPrune, an embedded feature-selection method built on a dense MLP with a dynamically sparse input layer. The key idea is an entry-based pruning criterion. Specifically, when a pruned input feature is (re)introduced as a candidate into the input layer, the algorithm measures the short-horizon relative change it induces. This is operationalized as the z-scored L1-norm of its first-layer gradient accumulations over the first n mini-batches post-entry. These accumulated scores are stored for all features. Periodically, the algorithm identifies the K features with the largest stored entry scores and prunes the rest. Then, it randomly regrows a batch of new candidate inputs whose incoming weights are reinitialized to tiny values. According to the authors, this strategy balances between established and freshly (re)introduced features and mitigates the bias of magnitude-based pruning toward long-tenured inputs. \nEmpirically, across 13 datasets (tabular, vision, speech, genomics, and text), EntryPrune often outperforms or matches SotA feature selection baselines (NeuroFS, LassoNet, STG, QS, RFS, etc.). An ablation suggests that entry-based scoring (short-horizon gradient sums) beats magnitude‐based criteria inside the same framework. The proposed method often runs faster wall-clock than a sparse-training baseline (NeuroFS). The code and replication scripts are provided.", "strengths": "1. The algorithmic contribution is clear. The entry-based pruning idea, i.e., scoring features by their initial impact upon (re)entry and freezing that score to avoid tenure bias, is straightforward, well motivated, and easy to implement. The random regrowth and tiny reinitialization are thoughtfully chosen to avoid gradient suppression and encourage exploration.\n2. The empirical evaluation is nice. The paper evaluates on 13 datasets spanning images, speech, sensor, and genomics, with reports of SVM accuracies and additional downstream models. Various ablations strengthen the analysis. \n3. The code is provided; many baselines are taken from public repositories. The paper documents settings and includes extensive appendix tables.", "weaknesses": "1. The related work discussion is incomplete. For example, recent approaches, such as \"CancelOut: A Layer for Feature Selection in Deep Neural Networks\" by Borisov et al., or \"Leveraging model inherent variable importance for stable online feature selection\" by Haug et al. are not discussed. In fact, I think they should be considered as competitors in the evaluation. \n2. Uncertainty quantification plays a critical role when it comes to feature selection. EntryPrune does not consider this aspect. Is there a way to measure/quantify the uncertainty of the selection process in EntryPrune?\n3. It is not clear what the contribution of the entry metric from regrowth is. Is it possible to isolate this contribution?\n4. The robustness analysis could be much deeper. Would a simple exponential decay or batch-normalized cumulative score improve robustness? In which scenarios does EntryPrune fail? What is the impact of concept shifts or other types of drifts in the data on EntryPrune's performance?", "questions": "See comments above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose EntryPrune, an embedded feature-selection method built on a dense MLP with a dynamically sparse input layer. The key idea is an entry-based pruning criterion. Specifically, when a pruned input feature is (re)introduced as a candidate into the input layer, the algorithm measures the short-horizon relative change it induces. This is operationalized as the z-scored L1-norm of its first-layer gradient accumulations over the first n mini-batches post-entry. These accumulated scores are stored for all features. Periodically, the algorithm identifies the K features with the largest stored entry scores and prunes the rest. Then, it randomly regrows a batch of new candidate inputs whose incoming weights are reinitialized to tiny values. According to the authors, this strategy balances between established and freshly (re)introduced features and mitigates the bias of magnitude-based pruning toward long-tenured inputs. \nEmpirically, across 13 datasets (tabular, vision, speech, genomics, and text), EntryPrune often outperforms or matches SotA feature selection baselines (NeuroFS, LassoNet, STG, QS, RFS, etc.). An ablation suggests that entry-based scoring (short-horizon gradient sums) beats magnitude‐based criteria inside the same framework. The proposed method often runs faster wall-clock than a sparse-training baseline (NeuroFS). The code and replication scripts are provided.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The algorithmic contribution is clear. The entry-based pruning idea, i.e., scoring features by their initial impact upon (re)entry and freezing that score to avoid tenure bias, is straightforward, well motivated, and easy to implement. The random regrowth and tiny reinitialization are thoughtfully chosen to avoid gradient suppression and encourage exploration.\n2. The empirical evaluation is nice. The paper evaluates on 13 datasets spanning images, speech, sensor, and genomics, with reports of SVM accuracies and additional downstream models. Various ablations strengthen the analysis. \n3. The code is provided; many baselines are taken from public repositories. The paper documents settings and includes extensive appendix tables.", "weaknesses": "1. The related work discussion is incomplete. For example, recent approaches, such as \"CancelOut: A Layer for Feature Selection in Deep Neural Networks\" by Borisov et al., or \"Leveraging model inherent variable importance for stable online feature selection\" by Haug et al. are not discussed. In fact, I think they should be considered as competitors in the evaluation. \n2. Uncertainty quantification plays a critical role when it comes to feature selection. EntryPrune does not consider this aspect. Is there a way to measure/quantify the uncertainty of the selection process in EntryPrune?\n3. It is not clear what the contribution of the entry metric from regrowth is. Is it possible to isolate this contribution?\n4. The robustness analysis could be much deeper. Would a simple exponential decay or batch-normalized cumulative score improve robustness? In which scenarios does EntryPrune fail? What is the impact of concept shifts or other types of drifts in the data on EntryPrune's performance?", "questions": "See comments above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761500987065}], "openreview_url": "https://openreview.net/forum?id=eChWBrh9mc", "arxiv_id": "2410.02344", "paper_pdf": "papers/eChWBrh9mc.pdf", "paper_pdf_sha256": "ef4c85abfb5c75fca4a06f79c550257b0836932a9c5e08f37154c63cb10f2c29", "paper_pdf_bytes": 4484287, "paper_pdf_source": "openreview", "code_url": "https://github.com/flxzimmer/entryprune", "code_repository": "flxzimmer/entryprune", "code_commit": "81cd448d5a1037902aaa86fc649d38cdf5f780eb", "code_archive": "repos/eChWBrh9mc.zip", "code_archive_sha256": "c8f482782c3b32c9342851c7f8a082e5703a9058e856bf69fd8493ad657d6a03", "code_archive_bytes": 166864, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 261, "github_languages": {"Python": 35650}, "github_archived": false, "github_pushed_at": "2025-05-20T14:28:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/relchanet-neural-network-feature-selection"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r6aX67YhD9", "year": 2025, "status": "rejected", "title": "Learning to Watermark LLM-generated Text via Reinforcement Learning", "authors": ["Xiaojun Xu", "Yuanshun Yao", "Yang Liu"], "authorids": ["~Xiaojun_Xu1", "~Yuanshun_Yao2", "~Yang_Liu3"], "authors_source": "OpenReview API", "abstract": "We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) trains a detector to detect the generated watermarked text and (2) tunes the LLM to generate text easily detectable by the detector while keeping its normal utility. We empirically show that our watermarks are more accurate, robust, and adaptable (to new attacks) with no generation overhead. It also allows watermarked model open-sourcing. In addition, if used together with alignment, the extra overhead introduced is low -- only training an extra reward model (i.e. our detector). We hope our work can bring more effort into studying a broader watermark design that is not limited to working with a fixed LLM.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "AwJBa8Vbn9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7704/Reviewer_7Qwc"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper explores embedding watermarks directly within large language models (LLMs) to detect and track misuse of generated content. Unlike prior token-level approaches, this research introduces a model-level watermark embedded into the model’s weights, detectable by a paired detector. The proposed method uses reinforcement learning to co-train the LLM and a detector, optimizing the LLM to produce watermarked text detectable with high accuracy while maintaining readability and model utility.", "review_text": "This paper explores embedding watermarks directly within large language models (LLMs) to detect and track misuse of generated content. Unlike prior token-level approaches, this research introduces a model-level watermark embedded into the model’s weights, detectable by a paired detector. The proposed method uses reinforcement learning to co-train the LLM and a detector, optimizing the LLM to produce watermarked text detectable with high accuracy while maintaining readability and model utility.", "strengths": "1. This paper proposes to fine-tune the LLMs to embed watermarks.\n2. The proposed method is robust against different attacks.\n3. The idea that combines the watermark embedding process with the alignment process is interesting.", "weaknesses": "1. The detection needs the original prompt, which is usually unavailable during the detection process.\n\n2. This paper uses the D^{nw} (human-written prompt and answer) to fine-tune the LLM and detector. What I am worried about is that the detector learned the difference between human-written text and LLM-generated text instead of un-watermarked text (text generated by unwatermarked LLMs) and watermarked text. It would be good to present the results between the original LLM and the fine-tuned LLM, and try this watermarking method on some more powerful LLMs.\n\n3. Could authors specify the model used to measure the PPL?\n\n4. This watermarking method changed the parameter of the original LLM. I think it would be good to measure if this fine-tuning affects the performance of the original LLM using methods like FActScore, AlpacaFarm, etc.", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores embedding watermarks directly within large language models (LLMs) to detect and track misuse of generated content. Unlike prior token-level approaches, this research introduces a model-level watermark embedded into the model’s weights, detectable by a paired detector. The proposed method uses reinforcement learning to co-train the LLM and a detector, optimizing the LLM to produce watermarked text detectable with high accuracy while maintaining readability and model utility.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. This paper proposes to fine-tune the LLMs to embed watermarks.\n2. The proposed method is robust against different attacks.\n3. The idea that combines the watermark embedding process with the alignment process is interesting.", "weaknesses": "1. The detection needs the original prompt, which is usually unavailable during the detection process.\n\n2. This paper uses the D^{nw} (human-written prompt and answer) to fine-tune the LLM and detector. What I am worried about is that the detector learned the difference between human-written text and LLM-generated text instead of un-watermarked text (text generated by unwatermarked LLMs) and watermarked text. It would be good to present the results between the original LLM and the fine-tuned LLM, and try this watermarking method on some more powerful LLMs.\n\n3. Could authors specify the model used to measure the PPL?\n\n4. This watermarking method changed the parameter of the original LLM. I think it would be good to measure if this fine-tuning affects the performance of the original LLM using methods like FActScore, AlpacaFarm, etc.", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730676368711}, {"id": "5iDDQ1uSxB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7704/Reviewer_YCyk"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a reinforcement learning-based watermarking method that simultaneously fine-tunes the model and trains a classifier to identify model-generated text. The method requires prompt-completion text pairs during detection and can be integrated with existing model alignment tasks. Experiments demonstrate good detectability and robustness against text modifications.", "review_text": "This paper proposes a reinforcement learning-based watermarking method that simultaneously fine-tunes the model and trains a classifier to identify model-generated text. The method requires prompt-completion text pairs during detection and can be integrated with existing model alignment tasks. Experiments demonstrate good detectability and robustness against text modifications.", "strengths": "• Successfully proposes and implements a watermarking method using fine-tuning and reinforcement learning\n\n• Conducts comprehensive experiments on watermark detectability and robustness\n\n• Successfully integrates the proposed fine-tuning method into existing alignment workflows", "weaknesses": "• The method appears to require the prompt that generated the text being tested for watermarks. This prerequisite fundamentally differs from current inference-time watermarking methods. The authors don't explicitly discuss how this condition affects watermark embedding and detection\n\n• The requirement of having the original prompt for detection significantly limits practical detection scenarios\n\n• The detectability and robustness experiments don't explicitly discuss the impact of prompts. For example, it's unclear how changes to prompts might affect detectability\n\n• The paper lacks details about baseline method parameters and settings, and these settings may not be comprehensive", "questions": "• Does the Detector D require prompt-completion text pairs as input? Does this mean the original prompt is needed during use? Doesn't this severely limit the watermark's practical applications?\n\n• Why doesn't the no-fine-tuning method achieve the lowest perplexity? Intuitively, not fine-tuning the model should have minimal impact on generated text. Does this suggest that perplexity might not accurately reflect the watermark's impact on text quality?\n\n• Since KGW changes its green list at each step, it naturally has poorer robustness. Would it be more fair to compare robustness with KGW family's unigram methods (“Provable robust watermarking for ai-generated text”)?\n\n• What are the parameter settings for baseline methods in the experiments? For example, what are the size of green list and delta values in KGW?\n\n• In the robustness experiments, did they only consider modifications to the generated text, or did they also examine how prompt modifications might affect watermark detection results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a reinforcement learning-based watermarking method that simultaneously fine-tunes the model and trains a classifier to identify model-generated text. The method requires prompt-completion text pairs during detection and can be integrated with existing model alignment tasks. Experiments demonstrate good detectability and robustness against text modifications.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "• Successfully proposes and implements a watermarking method using fine-tuning and reinforcement learning\n\n• Conducts comprehensive experiments on watermark detectability and robustness\n\n• Successfully integrates the proposed fine-tuning method into existing alignment workflows", "weaknesses": "• The method appears to require the prompt that generated the text being tested for watermarks. This prerequisite fundamentally differs from current inference-time watermarking methods. The authors don't explicitly discuss how this condition affects watermark embedding and detection\n\n• The requirement of having the original prompt for detection significantly limits practical detection scenarios\n\n• The detectability and robustness experiments don't explicitly discuss the impact of prompts. For example, it's unclear how changes to prompts might affect detectability\n\n• The paper lacks details about baseline method parameters and settings, and these settings may not be comprehensive", "questions": "• Does the Detector D require prompt-completion text pairs as input? Does this mean the original prompt is needed during use? Doesn't this severely limit the watermark's practical applications?\n\n• Why doesn't the no-fine-tuning method achieve the lowest perplexity? Intuitively, not fine-tuning the model should have minimal impact on generated text. Does this suggest that perplexity might not accurately reflect the watermark's impact on text quality?\n\n• Since KGW changes its green list at each step, it naturally has poorer robustness. Would it be more fair to compare robustness with KGW family's unigram methods (“Provable robust watermarking for ai-generated text”)?\n\n• What are the parameter settings for baseline methods in the experiments? For example, what are the size of green list and delta values in KGW?\n\n• In the robustness experiments, did they only consider modifications to the generated text, or did they also examine how prompt modifications might affect watermark detection results?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "Not applicable.", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730587714091}, {"id": "clrZsNN2Wb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7704/Reviewer_CaER"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces a method for watermarking outputs from Large Language Models (LLMs) by embedding detectable signals into the model's weights rather than the text itself. This approach uses reinforcement learning to co-train the LLM and a paired detector, enhancing detection accuracy, robustness, and adaptability to new attacks. The method allows for open-sourcing watermarked models without releasing unwatermarked versions and incurs no generation overhead. The experiments demonstrate high detection rates and resilience to adversarial attacks while maintaining the model's utility.", "review_text": "The paper introduces a method for watermarking outputs from Large Language Models (LLMs) by embedding detectable signals into the model's weights rather than the text itself. This approach uses reinforcement learning to co-train the LLM and a paired detector, enhancing detection accuracy, robustness, and adaptability to new attacks. The method allows for open-sourcing watermarked models without releasing unwatermarked versions and incurs no generation overhead. The experiments demonstrate high detection rates and resilience to adversarial attacks while maintaining the model's utility.", "strengths": "- This paper introduces watermarking LLM by fine-tuning, which makes watermark detection easier and more robust to attacks such as paraphrasing.\n- This manuscript is well-written and easy to follow.", "weaknesses": "- Fine-tuning the generative models and using an additional detector for watermark verification is not new, and related methods [1, 2] are supposed to be discussed in the related work section.\n- I am concerned about the reliability of using a language model as the detector instead of statistical tests.\n- It is unknown whether the extra fine-tuning process will introduce side effects or biases into LLMs, and there is no theoretical analysis of the changed parameters by additional fine-tuning.\n\nReference:\n\n[1] Yu, Ning, et al. \"Artificial fingerprinting for generative models: Rooting deepfake attribution in training data.\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\n\n[2] Yu, Ning, et al. \"Responsible disclosure of generative models using scalable fingerprinting.\" arXiv preprint arXiv:2012.08726 (2020).", "questions": "- The proposed method uses a learnable DNN as the detector, which can increase the watermark detection accuracy since both the LLM and the detector are optimized during watermarking. What if the adversary also provides a detector? That is to say, the adversary can keep the watermarked LLM fixed and optimize another LLM as his/her own detector. Note that the adversary does not need to modify the watermarked LLM, simply train another detector that can extract his/her watermark from the generated text, which would cast ambiguity over the verification process. How to deal with this situation?\n- Will you release the source codes?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a method for watermarking outputs from Large Language Models (LLMs) by embedding detectable signals into the model's weights rather than the text itself. This approach uses reinforcement learning to co-train the LLM and a paired detector, enhancing detection accuracy, robustness, and adaptability to new attacks. The method allows for open-sourcing watermarked models without releasing unwatermarked versions and incurs no generation overhead. The experiments demonstrate high detection rates and resilience to adversarial attacks while maintaining the model's utility.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- This paper introduces watermarking LLM by fine-tuning, which makes watermark detection easier and more robust to attacks such as paraphrasing.\n- This manuscript is well-written and easy to follow.", "weaknesses": "- Fine-tuning the generative models and using an additional detector for watermark verification is not new, and related methods [1, 2] are supposed to be discussed in the related work section.\n- I am concerned about the reliability of using a language model as the detector instead of statistical tests.\n- It is unknown whether the extra fine-tuning process will introduce side effects or biases into LLMs, and there is no theoretical analysis of the changed parameters by additional fine-tuning.\n\nReference:\n\n[1] Yu, Ning, et al. \"Artificial fingerprinting for generative models: Rooting deepfake attribution in training data.\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\n\n[2] Yu, Ning, et al. \"Responsible disclosure of generative models using scalable fingerprinting.\" arXiv preprint arXiv:2012.08726 (2020).", "questions": "- The proposed method uses a learnable DNN as the detector, which can increase the watermark detection accuracy since both the LLM and the detector are optimized during watermarking. What if the adversary also provides a detector? That is to say, the adversary can keep the watermarked LLM fixed and optimize another LLM as his/her own detector. Note that the adversary does not need to modify the watermarked LLM, simply train another detector that can extract his/her watermark from the generated text, which would cast ambiguity over the verification process. How to deal with this situation?\n- Will you release the source codes?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730482229559}, {"id": "Gbb4dSoT9O", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7704/Reviewer_CSmp"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes an LLM watermarking method that injects the watermark at the LLM fine-tuning stage, by training the LLM jointly with another detector language model via RL methods, such that the LLM's output can be reliably detected by the detector model.", "review_text": "This paper proposes an LLM watermarking method that injects the watermark at the LLM fine-tuning stage, by training the LLM jointly with another detector language model via RL methods, such that the LLM's output can be reliably detected by the detector model.", "strengths": "- The paper tackles an important problem of watermarking open-source LLM models to detect their output.\n\n- The paper is relatively well-written and clear.", "weaknesses": "- The success of the proposed method may be highly dependent on the training dataset and process (including several hyperparameters), which brings to question whether it can be generalized well in practice (please see questions below).\n\n- The paper should also be more thorough in its discussion of related works, benchmarks and claimed contributions. For example, as it is a training-based watermarking approach, the paper should at least include a discussion on other related fine-tuning watermarking methods (e.g. backdoor watermarking, like [1]). \n\n  [1]  Li, Shen, et al. \"Double-I Watermark: Protecting Model Copyright for LLM Fine-tuning.\"\n \n- Also, as it is comparing with training-free watermarking methods, it should also discuss and include training-free watermarking methods that can be applied on open-source models, such as [2], especially since the paper explicitly highlights open-sourcing watermarked models as a problem of existing works.\n\n   [2]  Lau et al. \"Waterfall, Framework for Robust and Scalable Text Watermarking and Provenance for LLMs\"\n\n\n-  The proposed method does not seem to provide the significant advantages in runtime overhead as claimed in Sec 5.5. For example, compared to the KGW method for the Llama2 model, the advantage in generation time is small (e.g. <0.01s or <2%) while disadvantage in detection time is relatively much larger (~0.02s or ~64%). This is on top of the very significant training overheads incurred by the method, which may not be 1-time given the requirements for additional adversarial training for better results, as described in various parts of the paper.\n\n- The claim that verification runtime is less important than generation runtime should also be more calibrated, as in practice fast verification runtime may be more important because in some settings active screening needs to be done to check for watermarks which will involve checking through large corpuses of text. \n\n- While the authors claim that compared to the RLHF pipeline, the only additional cost is \"training an extra reward model\", this may not be the case -- the more complex objective could impact the training dynamics and convergence of the model, and this potential impact should be discussed.\n\n\n- The sequential-code watermarking approach (Sec. 4.4) involving exact match (line 245) has the significant weakness of being very brittle to perturbations -- any small changes to the text may cause a non-match of the sequence and cause detectability to drop to 0. The comparison with other statistical-based model-centric watermarking method is not appropriate as those cases do not require an exact code sequence match, unlike the proposed method.\n\n- Given the claim from the paper that the method can be used to watermark open-source models and detect their output, the watermarking detection performance for OOD tasks would be important. I suggest that the authors shift up the OOD section in the appendix, and provide further details ad elaboration on this aspect.", "questions": "- Please provided additional explanations on any convergence analysis of the proposed iterative training algorithm.\n\n- Please elaborate on the differences between the proposed approach and other finetuning-based watermarking approaches, such as backdoor watermarking [1].\n\n- For the main results, as the proposed method involves finetuning with the training data, a fair comparison would involve the various methods applied to the fine-tuned model on the same training data. Otherwise, it would be hard to interpret any evaluation made on the utility or performance of the watermarked model. Please provide some empirical results on this.\n\n- Please elaborate on specifically what training data was used and what test dataset was used to evaluate the model performance, especially since the proposed method will have access to the training data while the baselines do not. Would the prompts used during testing need to be similar to the training dataset? The appendix on OOD tasks seem to imply that. \n\n- The proposed method seems to face significant challenges with substitution attack for the Llama2 setting on the C4 dataset. Could the authors please elaborate on reasons behind this, and whether it is potentially indicative of the sensitivity of the method's success to the hyperparameters used for training? \n\n- Please provide some sensitivity results, if available, on how the empirical results may change depending on the various key hyperparameters that need to be set for the proposed method\n\n- Given that training is required for the proposed method, the utility of the model may be of concern. The evaluations done are on tasks where there is high degree of tolerance for error (e.g. perplexity scores). Table 10 in the appendix seems to indicate that the response 'without watermark' and 'with watermark' feature relatively different semantic meaning. Have the authors done any experiments where standardized benchmarks tests are done on the watermarked LLMs, with results compared to the LLMs of the baseline methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an LLM watermarking method that injects the watermark at the LLM fine-tuning stage, by training the LLM jointly with another detector language model via RL methods, such that the LLM's output can be reliably detected by the detector model.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper tackles an important problem of watermarking open-source LLM models to detect their output.\n\n- The paper is relatively well-written and clear.", "weaknesses": "- The success of the proposed method may be highly dependent on the training dataset and process (including several hyperparameters), which brings to question whether it can be generalized well in practice (please see questions below).\n\n- The paper should also be more thorough in its discussion of related works, benchmarks and claimed contributions. For example, as it is a training-based watermarking approach, the paper should at least include a discussion on other related fine-tuning watermarking methods (e.g. backdoor watermarking, like [1]). \n\n  [1]  Li, Shen, et al. \"Double-I Watermark: Protecting Model Copyright for LLM Fine-tuning.\"\n \n- Also, as it is comparing with training-free watermarking methods, it should also discuss and include training-free watermarking methods that can be applied on open-source models, such as [2], especially since the paper explicitly highlights open-sourcing watermarked models as a problem of existing works.\n\n   [2]  Lau et al. \"Waterfall, Framework for Robust and Scalable Text Watermarking and Provenance for LLMs\"\n\n\n-  The proposed method does not seem to provide the significant advantages in runtime overhead as claimed in Sec 5.5. For example, compared to the KGW method for the Llama2 model, the advantage in generation time is small (e.g. <0.01s or <2%) while disadvantage in detection time is relatively much larger (~0.02s or ~64%). This is on top of the very significant training overheads incurred by the method, which may not be 1-time given the requirements for additional adversarial training for better results, as described in various parts of the paper.\n\n- The claim that verification runtime is less important than generation runtime should also be more calibrated, as in practice fast verification runtime may be more important because in some settings active screening needs to be done to check for watermarks which will involve checking through large corpuses of text. \n\n- While the authors claim that compared to the RLHF pipeline, the only additional cost is \"training an extra reward model\", this may not be the case -- the more complex objective could impact the training dynamics and convergence of the model, and this potential impact should be discussed.\n\n\n- The sequential-code watermarking approach (Sec. 4.4) involving exact match (line 245) has the significant weakness of being very brittle to perturbations -- any small changes to the text may cause a non-match of the sequence and cause detectability to drop to 0. The comparison with other statistical-based model-centric watermarking method is not appropriate as those cases do not require an exact code sequence match, unlike the proposed method.\n\n- Given the claim from the paper that the method can be used to watermark open-source models and detect their output, the watermarking detection performance for OOD tasks would be important. I suggest that the authors shift up the OOD section in the appendix, and provide further details ad elaboration on this aspect.", "questions": "- Please provided additional explanations on any convergence analysis of the proposed iterative training algorithm.\n\n- Please elaborate on the differences between the proposed approach and other finetuning-based watermarking approaches, such as backdoor watermarking [1].\n\n- For the main results, as the proposed method involves finetuning with the training data, a fair comparison would involve the various methods applied to the fine-tuned model on the same training data. Otherwise, it would be hard to interpret any evaluation made on the utility or performance of the watermarked model. Please provide some empirical results on this.\n\n- Please elaborate on specifically what training data was used and what test dataset was used to evaluate the model performance, especially since the proposed method will have access to the training data while the baselines do not. Would the prompts used during testing need to be similar to the training dataset? The appendix on OOD tasks seem to imply that. \n\n- The proposed method seems to face significant challenges with substitution attack for the Llama2 setting on the C4 dataset. Could the authors please elaborate on reasons behind this, and whether it is potentially indicative of the sensitivity of the method's success to the hyperparameters used for training? \n\n- Please provide some sensitivity results, if available, on how the empirical results may change depending on the various key hyperparameters that need to be set for the proposed method\n\n- Given that training is required for the proposed method, the utility of the model may be of concern. The evaluations done are on tasks where there is high degree of tolerance for error (e.g. perplexity scores). Table 10 in the appendix seems to indicate that the response 'without watermark' and 'with watermark' feature relatively different semantic meaning. Have the authors done any experiments where standardized benchmarks tests are done on the watermarked LLMs, with results compared to the LLMs of the baseline methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730220355986}], "openreview_url": "https://openreview.net/forum?id=r6aX67YhD9", "arxiv_id": "2403.10553", "paper_pdf": "papers/r6aX67YhD9.pdf", "paper_pdf_sha256": "f3c93c28a2c15aade4c9c9939bafa6ec3526949f4a78c21634b7f111a4306555", "paper_pdf_bytes": 476124, "paper_pdf_source": "openreview", "code_url": "https://github.com/xiaojunxu/learning-to-watermark-llm", "code_repository": "xiaojunxu/learning-to-watermark-llm", "code_commit": "09e6bf65a9c1ef867ee5356bbf67000caec651f6", "code_archive": "repos/r6aX67YhD9.zip", "code_archive_sha256": "48e090ca7194c97ad7cf49bd5cfbdb1f6a40a14999465d0e51cbf3c789c29641", "code_archive_bytes": 42777, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 37, "github_languages": {"Python": 151486}, "github_archived": false, "github_pushed_at": "2024-03-19T17:02:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-watermark-llm-generated-text-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "C6zFUEvgiU", "year": 2024, "status": "rejected", "title": "Feedback-guided Data Synthesis for Imbalanced Classification", "authors": ["Reyhane Askari Hemmat", "Mohammad Pezeshki", "Florian Bordes", "Michal Drozdzal", "Adriana Romero-Soriano"], "authorids": ["~Reyhane_Askari_Hemmat1", "~Mohammad_Pezeshki1", "~Florian_Bordes1", "~Michal_Drozdzal1", "~Adriana_Romero-Soriano1"], "authors_source": "OpenReview API", "abstract": "Current status quo in machine learning is to use static datasets of real images for training, which often come from long-tailed distributions. With the recent advances in generative models, researchers have started augmenting these static datasets with synthetic data, reporting moderate performance improvements on classification tasks. We hypothesize that these performance gains are limited by the lack of feedback from the classifier to the generative model, which would promote the usefulness of the generated samples to improve the classifier’s performance. In this work, we introduce a framework for augmenting static datasets with useful synthetic samples, which leverages one-shot feedback from the classifier to drive the sampling of the generative model. In order for the framework to be effective, we find that the samples must be close to the support of the real data of the task at hand, and be sufficiently diverse. We validate three feedback criteria on a long-tailed dataset (ImageNet-LT) as well as a group-imbalanced dataset (NICO++). On ImageNet-LT, we achieve state-of-the-art results, with over 4% improvement on underrepresented classes while being twice efficient in terms of the number of generated synthetic samples. NICO++ also enjoys marked boosts of over 5% in worst group accuracy. With these results, our framework paves the path towards effectively leveraging state-of-the-art text-to-image models as data sources that can be queried to improve downstream applications.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "FpvWDQzGb5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6233/Reviewer_SFA5"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The effectiveness of utilizing synthesized data is limited by the lack of feedback. This work proposes a framework to drive the sampling process of a generative model, thereby improving the usefulness of the generated samples.", "review_text": "The effectiveness of utilizing synthesized data is limited by the lack of feedback. This work proposes a framework to drive the sampling process of a generative model, thereby improving the usefulness of the generated samples.", "strengths": "● the experimental results were stunning, achieving state-of-the-art on ImageNet-LT\n● the writing is clear and easy to follow\n● the experiment is comprehensive, comparing three types of feedback criteria", "weaknesses": "ImageNet-LT is essentially a pseudo long-tail dataset, where the tail classes may not necessarily be the minority in the actual data distribution. Therefore, generative models can sample relatively well. However, for real-world long-tail distributions, is it also difficult for generative models to obtain sufficiently good samples?", "questions": "ImageNet-LT is essentially a pseudo long-tail dataset, where the tail classes may not necessarily be the minority in the actual data distribution. Therefore, generative models can sample relatively well. However, for real-world long-tail distributions, is it also difficult for generative models to obtain sufficiently good samples?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The effectiveness of utilizing synthesized data is limited by the lack of feedback. This work proposes a framework to drive the sampling process of a generative model, thereby improving the usefulness of the generated samples.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "● the experimental results were stunning, achieving state-of-the-art on ImageNet-LT\n● the writing is clear and easy to follow\n● the experiment is comprehensive, comparing three types of feedback criteria", "weaknesses": "ImageNet-LT is essentially a pseudo long-tail dataset, where the tail classes may not necessarily be the minority in the actual data distribution. Therefore, generative models can sample relatively well. However, for real-world long-tail distributions, is it also difficult for generative models to obtain sufficiently good samples?", "questions": "ImageNet-LT is essentially a pseudo long-tail dataset, where the tail classes may not necessarily be the minority in the actual data distribution. Therefore, generative models can sample relatively well. However, for real-world long-tail distributions, is it also difficult for generative models to obtain sufficiently good samples?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699076596481}, {"id": "KWe7NH3Rop", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6233/Reviewer_tjGi"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "With the recent advances in generative models, researchers have started augmenting these static datasets with synthetic data, reporting moderate performance improvements on long-tailed classification tasks. \nThe authors hypothesize that these performance gains are limited by the lack of feedback from the classifier to the generative model, which would promote the usefulness of the generated samples to improve the classifier’s performance.\nIn this work, the authors introduce a framework for augmenting static datasets with useful synthetic samples, which leverages one-shot feedback from the classifier to drive the sampling of the generative model. \nFor the framework to be effective, they find that the samples must be close to the support of the real data of the task at hand and be sufficiently diverse. \nThe authors validate three feedback criteria on a long-tailed dataset (ImageNet-LT) and a group-imbalanced dataset.", "review_text": "With the recent advances in generative models, researchers have started augmenting these static datasets with synthetic data, reporting moderate performance improvements on long-tailed classification tasks. \nThe authors hypothesize that these performance gains are limited by the lack of feedback from the classifier to the generative model, which would promote the usefulness of the generated samples to improve the classifier’s performance.\nIn this work, the authors introduce a framework for augmenting static datasets with useful synthetic samples, which leverages one-shot feedback from the classifier to drive the sampling of the generative model. \nFor the framework to be effective, they find that the samples must be close to the support of the real data of the task at hand and be sufficiently diverse. \nThe authors validate three feedback criteria on a long-tailed dataset (ImageNet-LT) and a group-imbalanced dataset.", "strengths": "1. The problem definition to encourage the generated samples to be helpful to the classifier, inspired by active learning frameworks, is novel.\n2. The proposed method performs better than the previous sample synthesis-based imbalance classification methods.", "weaknesses": "- The proposed solution for the problem definition is too naïve. For active learning methods, in addition to the confidence-based or entropy-based approach, margin margin-based approach is also possible. For the recent active learning criteria, such as BALD [1], VAAL [2], or MCDAL [3]. To claim the contribution of a complete research paper, the authors should devise an idea to leverage such recent active learning methods to find more novel solutions suitable for this problem.\n[1] Deep Bayesian Active Learning with Image Data. ICML 2017.\n[2] Variational Adversarial Active Learning. ICCV 2019.\n[3] MCDAL: Maximum Classifier Discrepancy for Active Learning. TNNLS 2022.\n\n- Also, instead of simply comparing among naïve active learning criteria, how about combining multiple losses (at least linear combination in the loss)? That would be more novel than the proposed solution.\n\n- The experiment is also too weak. For the datasets, The authors only use ImageNet and NICO++. However, according to other recent Long-tailed recognition papers, they usually evaluate their methods on iNaturalist and Place-LT datasets to demonstrate the scalability. At least the authors should have evaluated their method on CIFAR datasets to show the effectiveness of their methods on other datasets.\n\n- Also, a comparison with more recent state-of-the-art long-tailed recognition papers is missing. For example, CMO [4] is one of the recent long-tailed recognition methods based on sample synthesis. To claim the usefulness of the proposed method, the authors should compare the proposed method with recent long-tailed recognition papers, including [4].\n[4] The Majority Can Help The Minority: Context-rich Minority Oversampling for Long-Tailed Classification. CVPR 2022.\n\n- More analysis of the detailed design choices. For example, how are the hyper-parameters decided, such as w in Eqns (5), (6), (8)? As the authors proposed to add additional criteria, it would be necessary to analyze the effect of w on the performance.", "questions": "Please refer to the questions in the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "With the recent advances in generative models, researchers have started augmenting these static datasets with synthetic data, reporting moderate performance improvements on long-tailed classification tasks. \nThe authors hypothesize that these performance gains are limited by the lack of feedback from the classifier to the generative model, which would promote the usefulness of the generated samples to improve the classifier’s performance.\nIn this work, the authors introduce a framework for augmenting static datasets with useful synthetic samples, which leverages one-shot feedback from the classifier to drive the sampling of the generative model. \nFor the framework to be effective, they find that the samples must be close to the support of the real data of the task at hand and be sufficiently diverse. \nThe authors validate three feedback criteria on a long-tailed dataset (ImageNet-LT) and a group-imbalanced dataset.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The problem definition to encourage the generated samples to be helpful to the classifier, inspired by active learning frameworks, is novel.\n2. The proposed method performs better than the previous sample synthesis-based imbalance classification methods.", "weaknesses": "- The proposed solution for the problem definition is too naïve. For active learning methods, in addition to the confidence-based or entropy-based approach, margin margin-based approach is also possible. For the recent active learning criteria, such as BALD [1], VAAL [2], or MCDAL [3]. To claim the contribution of a complete research paper, the authors should devise an idea to leverage such recent active learning methods to find more novel solutions suitable for this problem.\n[1] Deep Bayesian Active Learning with Image Data. ICML 2017.\n[2] Variational Adversarial Active Learning. ICCV 2019.\n[3] MCDAL: Maximum Classifier Discrepancy for Active Learning. TNNLS 2022.\n\n- Also, instead of simply comparing among naïve active learning criteria, how about combining multiple losses (at least linear combination in the loss)? That would be more novel than the proposed solution.\n\n- The experiment is also too weak. For the datasets, The authors only use ImageNet and NICO++. However, according to other recent Long-tailed recognition papers, they usually evaluate their methods on iNaturalist and Place-LT datasets to demonstrate the scalability. At least the authors should have evaluated their method on CIFAR datasets to show the effectiveness of their methods on other datasets.\n\n- Also, a comparison with more recent state-of-the-art long-tailed recognition papers is missing. For example, CMO [4] is one of the recent long-tailed recognition methods based on sample synthesis. To claim the usefulness of the proposed method, the authors should compare the proposed method with recent long-tailed recognition papers, including [4].\n[4] The Majority Can Help The Minority: Context-rich Minority Oversampling for Long-Tailed Classification. CVPR 2022.\n\n- More analysis of the detailed design choices. For example, how are the hyper-parameters decided, such as w in Eqns (5), (6), (8)? As the authors proposed to add additional criteria, it would be necessary to analyze the effect of w on the performance.", "questions": "Please refer to the questions in the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698777596825}, {"id": "vLKHY6hgZa", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6233/Reviewer_Ajag"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This thesis utilises recent advances in generative modelling to address the shortcomings of synthetic data in representation learning and introduces feedback from downstream classifier models to guide the data generation process. To augment static datasets with useful synthetic samples, the research designs a framework that utilises pre-trained image generation models to provide useful and diverse synthetic samples that are close to the support of real data distributions to improve the representation learning task. This paper lays the groundwork for the effective use of state-of-the-art text-to-image models as data sources that can be queried to improve downstream applications.", "review_text": "This thesis utilises recent advances in generative modelling to address the shortcomings of synthetic data in representation learning and introduces feedback from downstream classifier models to guide the data generation process. To augment static datasets with useful synthetic samples, the research designs a framework that utilises pre-trained image generation models to provide useful and diverse synthetic samples that are close to the support of real data distributions to improve the representation learning task. This paper lays the groundwork for the effective use of state-of-the-art text-to-image models as data sources that can be queried to improve downstream applications.", "strengths": "- Originality. The paper designs a diffusion model sampling strategy that uses the feedback of the pre-trained classifier to generate samples that help improve its own performance, which improves the classification performance to a certain extent. Has a certain degree of innovation.\n- Quality. The experimental design of the paper is reasonable, and the feasibility of the method is verified in ImageNet-LT and NICO++. \n- Clarity. The paper well-organized and clearly written. \n- Significance. The ideas proposed in this paper have certain contributions to this field.", "weaknesses": "1. The font format of the article is not uniform. Do the words in italics want to express any special meaning? Make it difficult for readers to read.\n2. The charts are mixed up, for example, Figure 5. Is it a table or a graph? The sizes of some pictures also don’t match.\n3. How about the time complexity of this method?\n4. Are there more evaluation metrics to evaluate the performance of the proposed method versus the baseline method?", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This thesis utilises recent advances in generative modelling to address the shortcomings of synthetic data in representation learning and introduces feedback from downstream classifier models to guide the data generation process. To augment static datasets with useful synthetic samples, the research designs a framework that utilises pre-trained image generation models to provide useful and diverse synthetic samples that are close to the support of real data distributions to improve the representation learning task. This paper lays the groundwork for the effective use of state-of-the-art text-to-image models as data sources that can be queried to improve downstream applications.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- Originality. The paper designs a diffusion model sampling strategy that uses the feedback of the pre-trained classifier to generate samples that help improve its own performance, which improves the classification performance to a certain extent. Has a certain degree of innovation.\n- Quality. The experimental design of the paper is reasonable, and the feasibility of the method is verified in ImageNet-LT and NICO++. \n- Clarity. The paper well-organized and clearly written. \n- Significance. The ideas proposed in this paper have certain contributions to this field.", "weaknesses": "1. The font format of the article is not uniform. Do the words in italics want to express any special meaning? Make it difficult for readers to read.\n2. The charts are mixed up, for example, Figure 5. Is it a table or a graph? The sizes of some pictures also don’t match.\n3. How about the time complexity of this method?\n4. Are there more evaluation metrics to evaluate the performance of the proposed method versus the baseline method?", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698463872449}], "openreview_url": "https://openreview.net/forum?id=C6zFUEvgiU", "arxiv_id": "2310.00158", "paper_pdf": "papers/C6zFUEvgiU.pdf", "paper_pdf_sha256": "4f620a92d5c286c2f9ee06b5cc2931782f68ef71b082d09972843b41ada10ee3", "paper_pdf_bytes": 15629518, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/Feedback-guided-Data-Synthesis", "code_repository": "facebookresearch/Feedback-guided-Data-Synthesis", "code_commit": "60dd027b7f5a4c133fcf6b43e98bc5dbdd5cacb6", "code_archive": "repos/C6zFUEvgiU.zip", "code_archive_sha256": "31126bad0da3a98c3d5b16ab4b163e33127d7f587a65653b1c6a6f97f8b2ac59", "code_archive_bytes": 38679, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 35, "github_languages": {"Python": 17277}, "github_archived": true, "github_pushed_at": "2024-09-09T21:22:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/feedback-guided-data-synthesis-for-imbalanced"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "K2spEiswXVf", "year": 2023, "status": "rejected", "title": "MALIBO: Meta-Learning for Likelihood-free Bayesian Optimization", "authors": ["Jiarong Pan", "Stefan Falkner", "Felix Berkenkamp", "Joaquin Vanschoren"], "authorids": ["~Jiarong_Pan2", "~Stefan_Falkner1", "~Felix_Berkenkamp1", "~Joaquin_Vanschoren1"], "authors_source": "OpenReview API", "abstract": "Bayesian Optimization (BO) is a popular method to optimize expensive black-box functions. Typically, BO only uses observations from the current task. Recently proposed methods try to warm-start BO by exploiting knowledge from related tasks, yet suffer from scalability issues and sensitivity to heterogeneous scale across multiple datasets. We propose a novel approach to solve these problems by combining a meta-learning technique and a likelihood-free acquisition function. The meta-learning model simultaneously learns the underlying (task-agnostic) data distribution and a latent feature representation for individual tasks. The likelihood-free BO technique has less stringent assumptions about the problems and works with any classification algorithm, making it computation efficient and robust to different scales across tasks. Finally, gradient boosting is used as a residual model on top to adapt to distribution drifts between new and prior tasks, which might otherwise weaken the usefulness of the meta-learned features. Experiments show that the meta-model learns an effective prior for warm-starting optimization algorithms, while being cheap to evaluate and invariant to changes of scale across different datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "cmqqIeTeZAj", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5086/Reviewer_YSQy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The work proposes a likelihood-free Bayesian optimization strategy with a meta-learning scheme. Given multiple tasks, it trains task-agnostic and task-specific components in order to predict a probability that measures how likely a solution is, inspired by BORE and LFBO. By utilizing Bayesian logistic regression, it determines a query point. In addition, the authors use a gradient boosting model to predict a residual of the model. Eventually, the experiments demonstrate the effectiveness of the proposed method, compared to other existing methods.", "review_text": "Please see the above text boxes.", "strengths": "Here I describe the strengths and weaknesses of this paper.\n\n### Strengths\n\nIt solves an interesting topic where multiple historical tasks are given, by applying a likelihood-free framework.\n\nThe proposed method is quite novel. In particular, a combination of some components such as mean prediction layer, residual prediction layer, and gradient boosting is interesting.\n\nThe purpose of the respective components and the respective loss functions is well-described.\n\n### Weaknesses\n\nPresentation and writing can be improved. The current version is okay, but I think it can be polished more.\n\nIteration budgets for the experiments are too low. I think that you should give a larger iteration budget in order to show the convergence of the algorithm tested.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The work proposes a likelihood-free Bayesian optimization strategy with a meta-learning scheme. Given multiple tasks, it trains task-agnostic and task-specific components in order to predict a probability that measures how likely a solution is, inspired by BORE and LFBO. By utilizing Bayesian logistic regression, it determines a query point. In addition, the authors use a gradient boosting model to predict a residual of the model. Eventually, the experiments demonstrate the effectiveness of the proposed method, compared to other existing methods.", "strength_and_weaknesses": "Here I describe the strengths and weaknesses of this paper.\n\n### Strengths\n\nIt solves an interesting topic where multiple historical tasks are given, by applying a likelihood-free framework.\n\nThe proposed method is quite novel. In particular, a combination of some components such as mean prediction layer, residual prediction layer, and gradient boosting is interesting.\n\nThe purpose of the respective components and the respective loss functions is well-described.\n\n### Weaknesses\n\nPresentation and writing can be improved. The current version is okay, but I think it can be polished more.\n\nIteration budgets for the experiments are too low. I think that you should give a larger iteration budget in order to show the convergence of the algorithm tested.", "clarity,_quality,_novelty_and_reproducibility": "### Questions\n\n* For the experimental results, the variance (or standard deviation) of the experiments (i.e., shaded regions) is somewhat odd. Since $y$-axis is a log scale, the variance at the last of iterations should be larger than the variance at the beginning of iterations.\n\n* A feature extractor can be called as ResNet? According to Appendix C, your feature extractor is different with ResNet. I think that the name should be changed.\n\n* I would like to ask about gradient boosting. Which gradient boosting is used in this paper?\n\n* Following the above question, if you use gradient descent, how did you optimize a function $C$ (Line 10 or Line 12 of Algorithm 1)?\n\n* In the experiments, three proposed method, i.e., MALIBO wo GB, GB, and GB-TS, are tested. Does MALIBO wo GB include TS? Why did not you test MALIBO wo GB w TS or MALIBO wo GB wo TS?\n\n### Minor issues\n\n* In Page 4, $\\max(y - \\tau)$ seems like a typo; please fix it.\n\n* In the caption of Figure 2, $x_t$ should be $\\boldsymbol x_t$.", "summary_of_the_review": "Please see the above text boxes.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666900411782}, {"id": "CgKqhIRAf2o", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5086/Reviewer_YYiE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a meta-learning method for likelihood-free Bayesian optimization. In particular, the proposed approach is a combination of two existing solutions, a meta-learning approach from BaNNER and a likelihood-free LFBO, together with the gradient boosting method. The proposed solution is able to work with high-dimensional inputs and handle heterogeneous scales and noises across different tasks.\n", "review_text": "Due to the limited novelty, concerns about the proposed solution, and the lack of baselines in experiments, the paper may require further improvements to fit the conference.", "strengths": "The main strength of the paper is in the empirical performance which is shown to outperform two of the existing works, ABLR, and GC3P. However, it has several weaknesses as follows.\n\n1. The technical solution is not novel since it is based mostly on the two existing works BaNNER and LFBO.\n\n2. Although the paper claims that the meta-learned classifier can balance between exploration and exploitation, the proposed approach requires a gradient boost method to correct the errors. It means that the meta-learned classifier does not allow exploration properly. This is unlike multi-task GP, where the correlation between tasks can be learned from the data without any additional ad-hoc method (such as gradient boosting) for correction.\n\n3. While Thompson sampling often relies on a good approximation of the posterior distribution, the proposed approach only uses the Laplace method to approximate the posterior distribution which is a very simple approximation method. Better approximation methods such as variants of MCMC and/or variational inference techniques should be applied.\n\n4. The numerical representation of the task (z) is simply optimized to be close to a standard (multivariate) normal distribution. I am wondering if the proposed method can work with a multi-modal task distribution, i.e., prior tasks form 2 clusters where tasks in a cluster are similar and tasks between clusters are different. In a principled Bayesian approach such as multi-task GP, it can correctly correlate a new task with an existing task.\n\n5. It is unclear to me about the choice of \\tau (for prior tasks and the current task) in the proposed algorithm.\n\n6. While the paper reviews a lot of related works on meta-learning for BO, the experiments only consist of 2 existing works while ignoring the others such as those using GP surrogates (e.g., running experiments on low-dimensional problems).\n\n7. Section 3.1 lacks many well-known BO methods such as GP-UCB, predictive entropy search, max-value entropy search, and knowledge gradient-based methods.\n\n8. Since BO is about sample efficiency, can gradient boosting work reasonably well with a small training dataset?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work proposes a meta-learning method for likelihood-free Bayesian optimization. In particular, the proposed approach is a combination of two existing solutions, a meta-learning approach from BaNNER and a likelihood-free LFBO, together with the gradient boosting method. The proposed solution is able to work with high-dimensional inputs and handle heterogeneous scales and noises across different tasks.\n", "strength_and_weaknesses": "The main strength of the paper is in the empirical performance which is shown to outperform two of the existing works, ABLR, and GC3P. However, it has several weaknesses as follows.\n\n1. The technical solution is not novel since it is based mostly on the two existing works BaNNER and LFBO.\n\n2. Although the paper claims that the meta-learned classifier can balance between exploration and exploitation, the proposed approach requires a gradient boost method to correct the errors. It means that the meta-learned classifier does not allow exploration properly. This is unlike multi-task GP, where the correlation between tasks can be learned from the data without any additional ad-hoc method (such as gradient boosting) for correction.\n\n3. While Thompson sampling often relies on a good approximation of the posterior distribution, the proposed approach only uses the Laplace method to approximate the posterior distribution which is a very simple approximation method. Better approximation methods such as variants of MCMC and/or variational inference techniques should be applied.\n\n4. The numerical representation of the task (z) is simply optimized to be close to a standard (multivariate) normal distribution. I am wondering if the proposed method can work with a multi-modal task distribution, i.e., prior tasks form 2 clusters where tasks in a cluster are similar and tasks between clusters are different. In a principled Bayesian approach such as multi-task GP, it can correctly correlate a new task with an existing task.\n\n5. It is unclear to me about the choice of \\tau (for prior tasks and the current task) in the proposed algorithm.\n\n6. While the paper reviews a lot of related works on meta-learning for BO, the experiments only consist of 2 existing works while ignoring the others such as those using GP surrogates (e.g., running experiments on low-dimensional problems).\n\n7. Section 3.1 lacks many well-known BO methods such as GP-UCB, predictive entropy search, max-value entropy search, and knowledge gradient-based methods.\n\n8. Since BO is about sample efficiency, can gradient boosting work reasonably well with a small training dataset?", "clarity,_quality,_novelty_and_reproducibility": "As discussed above, the novelty of this paper is limited since it is based mostly on the two existing works BaNNER and LFBO. I also have several concerns about the proposed solution as elaborated in the above weaknesses. Regarding clarity, as the Laplace approximation is a well-known technique, the paper can reduce the explanation of the Laplace approximation to include more explanation on the regularization (to learn the task representation), the gradient boosting, and the choice of \\tau.", "summary_of_the_review": "Due to the limited novelty, concerns about the proposed solution, and the lack of baselines in experiments, the paper may require further improvements to fit the conference.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough"}, "tcdate": 1666640111705}, {"id": "CwzkuTpvZX1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5086/Reviewer_Xo11"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the problem of meta Bayesian optimization (BO), which aims to warm-start the BO process by exploiting knowledge from related tasks. In this paper, the authors propose warm-starting the acquisition function, which takes the form of a classifier in the likelihood-free BO setting. Gradient boosting is further incorporated to combat distributional shifts. ", "review_text": "The paper presents an interesting idea to tackle an important problem. However, the difficulty or challenges of the problem are not highlighted enough. The comparison with related works is lacking. And the writing can be improved. \n\nDetailed comments:\n\n- The idea of meta-learning an acquisition function has been explored by Volpp et al. (2020). Could the authors elaborate on the differences and advantages of the proposed algorithm to that of Volpp et al. (2020)? In addition, this should also be included as a baseline in the experiment section.\n\n- The paper is generally hard to follow since the readers assume knowledge from several other key papers. I have to read several other works on the BORE framework and BaNNER to understand several parts of the paper. I encourage the authors to make the paper more self-contained by reintroducing concepts from essential related works like Berkenkamp et al. (2021), Tiao et al. (2021), and Song et al. (2022).\n\n- When the input space is a simple Euclidean space without any structure (think of minimizing the function $f(x) = x^2 – 4x$), how does the feature embedding $h(\\cdot)$ works? I suppose feature embedding is only useful when there is some structure in the inputs (e.g., images).\n\n- Since BO is a black-box optimization algorithm, can the authors indicate more clearly what are the objective functions that we are optimizing in the experiment section? It is difficult for readers to comprehend and assess results without knowing what are we optimizing.\n\n- From my understanding, UCB is one of the most commonly used acquisition functions, with a nice theory in BO. However, the paper did not mention or compare with UCB. Is there a good reason for that? Can the authors compare the proposed algorithm with GP-based BO methods with EI/PI and UCB?\n", "strengths": "Strengths:\n\n(+) The problem of (meta) Bayesian optimization is an important problem with several applications.\n\nWeaknesses:\n\n(-) The comparison with existing works is lacking, both in the techniques and in the empirical evaluations.\n\n(-) The lack of a theoretical guarantee (which is common in the BO literature).\n\n(-) The writing can be improved (as detailed below).  \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies the problem of meta Bayesian optimization (BO), which aims to warm-start the BO process by exploiting knowledge from related tasks. In this paper, the authors propose warm-starting the acquisition function, which takes the form of a classifier in the likelihood-free BO setting. Gradient boosting is further incorporated to combat distributional shifts. ", "strength_and_weaknesses": "Strengths:\n\n(+) The problem of (meta) Bayesian optimization is an important problem with several applications.\n\nWeaknesses:\n\n(-) The comparison with existing works is lacking, both in the techniques and in the empirical evaluations.\n\n(-) The lack of a theoretical guarantee (which is common in the BO literature).\n\n(-) The writing can be improved (as detailed below).  \n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is generally hard to follow, many concepts require prior knowledge and are not sufficiently explained in the paper. More detail below.\n\nNovelty: The novelty aspect is low as a similar idea has been explored by Volpp et al. (2020). The proposed meta-learning framework is almost a direct application of Berkenkamp et al. (2021).\n\nReproducibility: There is no code included with the submission.\n", "summary_of_the_review": "The paper presents an interesting idea to tackle an important problem. However, the difficulty or challenges of the problem are not highlighted enough. The comparison with related works is lacking. And the writing can be improved. \n\nDetailed comments:\n\n- The idea of meta-learning an acquisition function has been explored by Volpp et al. (2020). Could the authors elaborate on the differences and advantages of the proposed algorithm to that of Volpp et al. (2020)? In addition, this should also be included as a baseline in the experiment section.\n\n- The paper is generally hard to follow since the readers assume knowledge from several other key papers. I have to read several other works on the BORE framework and BaNNER to understand several parts of the paper. I encourage the authors to make the paper more self-contained by reintroducing concepts from essential related works like Berkenkamp et al. (2021), Tiao et al. (2021), and Song et al. (2022).\n\n- When the input space is a simple Euclidean space without any structure (think of minimizing the function $f(x) = x^2 – 4x$), how does the feature embedding $h(\\cdot)$ works? I suppose feature embedding is only useful when there is some structure in the inputs (e.g., images).\n\n- Since BO is a black-box optimization algorithm, can the authors indicate more clearly what are the objective functions that we are optimizing in the experiment section? It is difficult for readers to comprehend and assess results without knowing what are we optimizing.\n\n- From my understanding, UCB is one of the most commonly used acquisition functions, with a nice theory in BO. However, the paper did not mention or compare with UCB. Is there a good reason for that? Can the authors compare the proposed algorithm with GP-based BO methods with EI/PI and UCB?\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666328374528}, {"id": "ZtjIEt8YRYF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5086/Reviewer_Xog3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new approach to perform transfer-learning for Hyperparameter optimization where offline evaluations are used to fasten the tuning on a new task. The method proposes to leverage recent work that uses classification to train surrogates by learning to classify good/bad configuration. In particular in this work, the classifier takes a combination of global (task agnostic) and local (task specific) features and is learned with a Bayesian Logistic Regression layer which is made possible with the use of several approximations. \nIn addition, a gradient boosting approach is used to improve the final fit of the model at each iterations. Experiments are conducted on several tabular benchmarking suites  (hpobench, mlbench) and also artificial examples to study the effect of noise levels. The method proposed is show to be competitive or better than the baselines proposed.", "review_text": "The paper provides an interesting and novel method to a relevant problem with several ingredients that may be leveraged by future work (for instance using GB to improve surrogate quality). However, as it stands the experimental section has only too little baselines to really its performance against state-of-the-art (only 2 baselines for transfer) and some part of the models are unclear. I believe those could be potentially addressed before camera ready.\n\nAdditional details:\n* many typos, it would be valuable to run a spell-check:\n  * meta-leared\n  * observaitons\n  * This benchmarks\n  * Coppola\n  * ensmeble\n* Eq (3): interestingly, the scale is back in the loss (but the surrogate dont have to predict values proportional to it)\n* A potential ablation would be to use BORE to illustrate the benefit of using LFBO in your case, other possible ablations I was wondering about were: using only local/global part of the model, fitting (3) directly as mentioned in the begining of 4.2. The latter would be important to have given that otherwise all the complexity of the approximations (laplace and logits) would not be justified.", "strengths": "The strengths of the paper are:\n* tackle a very relevant and impactful problem\n* provides a novel technical contributions with several ideas that could be leveraged in future work (refinement with gradient-boosting, use of EI equivalent likelihood-free criterion)\n* good coverage of experiments: a reasonable range of benchmarks/different blackboxes are considered \n\nWeaknesses:\n* lack of runtime analysis: no runtime is given for how long the method takes to return suggestion (the appendix mentions 2048 epochs on top of that, a gradient boosting tree is fitted). However, this can significantly worsen the results if they were reported against wallclock time (as proposed in other works) since then almost all time will be spent in fitting models and the method would likely underperform random-search. Without additional details, one cannot assess if the method will have any practical relevance (a method taking 5 minute to suggest the next candidate will likely not be very applicable).\n* the set of baselines is relatively considered is relatively small (only 2 transfer learning and 2 non transfer baselines). They are several baselines that could be easily added: BORE for non transfer, but also BORE/LFBO with search space pruned to bounding box of best previous evaluations [Peronne 2019]. Ideally, another additional transfer-learning baseline on top of this simple bounding-box approach would also be added in order to better assess the quality of the method regarding state-of-the-art (2 methods for transfer is a small number).\n* details are lacking, some part of the methods were not clear to me:\n  * 4.1: was difficult to get through while it is a key part of the paper, I would suggest adding the dimensions for the different variables and give the final expression of the classifier that sums its two input (which is given in 4.2). In any case, the text alone was not sufficient for me to get exactly how $\\phi \\dot z_t$ is obtained as I could see several ways to achieve this depending on input dimensions.\n  * 4.2: the expression of L^LFBO given in Eq. (3) does not take arguments as input\n  * 4.3: how exactly is the Gradient boosting combined with the classifier was not entirely clear to me. I would recommend to write it down formally rather than with words (given the current text description, there could be many ways on how a GB would be fitted to reduce the residuals)\n* the set of ablations is small and several complexity are not justified (for instance the approximation done instead of the direct optimization of (3), see additional details for a bigger description on this.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new approach to perform transfer-learning for Hyperparameter optimization where offline evaluations are used to fasten the tuning on a new task. The method proposes to leverage recent work that uses classification to train surrogates by learning to classify good/bad configuration. In particular in this work, the classifier takes a combination of global (task agnostic) and local (task specific) features and is learned with a Bayesian Logistic Regression layer which is made possible with the use of several approximations. \nIn addition, a gradient boosting approach is used to improve the final fit of the model at each iterations. Experiments are conducted on several tabular benchmarking suites  (hpobench, mlbench) and also artificial examples to study the effect of noise levels. The method proposed is show to be competitive or better than the baselines proposed.", "strength_and_weaknesses": "The strengths of the paper are:\n* tackle a very relevant and impactful problem\n* provides a novel technical contributions with several ideas that could be leveraged in future work (refinement with gradient-boosting, use of EI equivalent likelihood-free criterion)\n* good coverage of experiments: a reasonable range of benchmarks/different blackboxes are considered \n\nWeaknesses:\n* lack of runtime analysis: no runtime is given for how long the method takes to return suggestion (the appendix mentions 2048 epochs on top of that, a gradient boosting tree is fitted). However, this can significantly worsen the results if they were reported against wallclock time (as proposed in other works) since then almost all time will be spent in fitting models and the method would likely underperform random-search. Without additional details, one cannot assess if the method will have any practical relevance (a method taking 5 minute to suggest the next candidate will likely not be very applicable).\n* the set of baselines is relatively considered is relatively small (only 2 transfer learning and 2 non transfer baselines). They are several baselines that could be easily added: BORE for non transfer, but also BORE/LFBO with search space pruned to bounding box of best previous evaluations [Peronne 2019]. Ideally, another additional transfer-learning baseline on top of this simple bounding-box approach would also be added in order to better assess the quality of the method regarding state-of-the-art (2 methods for transfer is a small number).\n* details are lacking, some part of the methods were not clear to me:\n  * 4.1: was difficult to get through while it is a key part of the paper, I would suggest adding the dimensions for the different variables and give the final expression of the classifier that sums its two input (which is given in 4.2). In any case, the text alone was not sufficient for me to get exactly how $\\phi \\dot z_t$ is obtained as I could see several ways to achieve this depending on input dimensions.\n  * 4.2: the expression of L^LFBO given in Eq. (3) does not take arguments as input\n  * 4.3: how exactly is the Gradient boosting combined with the classifier was not entirely clear to me. I would recommend to write it down formally rather than with words (given the current text description, there could be many ways on how a GB would be fitted to reduce the residuals)\n* the set of ablations is small and several complexity are not justified (for instance the approximation done instead of the direct optimization of (3), see additional details for a bigger description on this.", "clarity,_quality,_novelty_and_reproducibility": "The clarity of the paper is generally OK except for some details missing and some notations missing (should be easy to fix).\nThe quality of the paper is good although it could be improve by adding more baselines and ablations to show better that the method actually improves state-of-the-art. Finally, the paper provides an original contribution and references well previous work.\n", "summary_of_the_review": "The paper provides an interesting and novel method to a relevant problem with several ingredients that may be leveraged by future work (for instance using GB to improve surrogate quality). However, as it stands the experimental section has only too little baselines to really its performance against state-of-the-art (only 2 baselines for transfer) and some part of the models are unclear. I believe those could be potentially addressed before camera ready.\n\nAdditional details:\n* many typos, it would be valuable to run a spell-check:\n  * meta-leared\n  * observaitons\n  * This benchmarks\n  * Coppola\n  * ensmeble\n* Eq (3): interestingly, the scale is back in the loss (but the surrogate dont have to predict values proportional to it)\n* A potential ablation would be to use BORE to illustrate the benefit of using LFBO in your case, other possible ablations I was wondering about were: using only local/global part of the model, fitting (3) directly as mentioned in the begining of 4.2. The latter would be important to have given that otherwise all the complexity of the approximations (laplace and logits) would not be justified.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666095446643}], "openreview_url": "https://openreview.net/forum?id=K2spEiswXVf", "arxiv_id": "2307.03565", "paper_pdf": "papers/K2spEiswXVf.pdf", "paper_pdf_sha256": "d3e672a633b22012cb3e2d6f94a7eed612c4e145530d1df7bd8b773843e90884", "paper_pdf_bytes": 7692730, "paper_pdf_source": "openreview", "code_url": "https://github.com/boschresearch/meta-learning-likelihood-free-bayesian-optimization", "code_repository": "boschresearch/meta-learning-likelihood-free-bayesian-optimization", "code_commit": "87f1b3d2ed59441f8197b38ffdf68116bb90c2d8", "code_archive": "repos/K2spEiswXVf.zip", "code_archive_sha256": "bc582087f97d95af0ef8f1a4e0577f898871e73b28fb1b32a6d99bafbeae54bc", "code_archive_bytes": 45516, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 43, "github_languages": {"Python": 77655}, "github_archived": true, "github_pushed_at": "2024-09-11T14:24:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/malibo-meta-learning-for-likelihood-free"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZAA0Ol4z2i4", "year": 2022, "status": "rejected", "title": "Explaining Off-Policy Actor-Critic From A Bias-Variance Perspective", "authors": ["Ting-Han Fan", "Peter Ramadge"], "authorids": ["~Ting-Han_Fan1", "~Peter_Ramadge1"], "authors_source": "OpenReview API", "abstract": "Off-policy Actor-Critic algorithms have demonstrated phenomenal experimental performance but still require better explanations. To this end, we show its policy evaluation error on the distribution of transitions decomposes into: a Bellman error, a bias from policy mismatch, and a variance term from sampling. By comparing the magnitude of bias and variance, we explain the success of the Emphasizing Recent Experience sampling and 1/age weighted sampling. Both sampling strategies yield smaller bias and variance and are hence preferable to uniform sampling.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "b2qizkRSJ5s", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1960/Reviewer_RYz4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper seeks to improve our understanding of actor-critic methods in deep RL by providing an analysis of the error in estimating the value function. The main theoretical result is an error bound which depends on the Bellman error, variance from sampling and a bias term from the policy mismatch. This error bound suggests that a replay buffer sampling strategy that focuses on recent samples will have an advantage. Experiments validate this hypothesis and using weighted sampling with 1/age weights improves performance. Morover, the theory helps explain the success of the ERE algorithm which also emphasizes recent transitions. \n", "review_text": "This is a nice paper expanding our undertanding of actor-critic algorithms. Focusing on the policy evaluation error is a simple approach yielding meaningful results. The assumptions are sufficiently weak for the conclusions to apply to modern deep RL methods, a strength of this analysis. The experiments are a good addition to validate the theoretical findings and make the case more convincing. Overall, this is a great paper with concrete insights that are clearly conveyed. \n\nQuestions and comments:\n- I think the fact that the analysis is uisng continuous action spaces is neat. This seems to be a less popular choice in RL.\n\n- Algorithm 1 is set up nicely to avoid the complexity of deep RL algorithms but still retain the essence of the algorithms and allowing interesting theoretical derivations.\n\n- p.6 To clarify, why is it stated that $|\\hat{Q} - Q^{\\pi^N}|$  mainly depends on steps $\\ge i$. Is this because of the presence of $\\epsilon^{i,L}_\\hat{Q}$ and $W^{i,L}_1$?\n\n- It would have been nice to have an experiment which directly measures the policy evaluation error. This could be done in a simpler environment where policy evaluation is relatively easy e.g. LQR. This could help support the theory even further.\n\n- How are the curves in figure 1 generated? Are these obtained analytically or are simulations run to obtain then? \nIt seems like for uniform sampling, if we sample only 1 transition at each step, the expected count (aggregate weight) after $n$ steps would be $\\sum_{i=0}^n \\frac{1}{i} \\approx \\log n$. This would imply that the aggregate weight for the transition at time step $i$ would be approximate $\\log n - log i$, which seems to roughly match the shape of the curve plotted.\n\n- p.6 The detailed discussion in the \"Normalization\" paragraph is nice to tie up loose ends.\n\n- p.6-7 Some parts of the \"Interpretation\" paragraph are a bit confusing to me. In particular, I'm not following the part discussing the bias and variance terms and how they impact the optimization process. I understand that, near the beginning of learning, variance may be large relative to bias. But in phase 2, it's not clear to me what happens to the relative magnitudes. It seems like both bias and variance would be small. Also, while lower policy evaluation error leads to more accurate estimations of $Q^{\\pi^N}$, I'm finding it difficult to translate that to how fast optimization progresses. It would seem like more accurate estimates would lead to \"better\" optimization steps but, at the same time, the optimization process will naturally slow down as the iterates approach a maximum. I think this paragraph should be clarified a bit more.\n\n- Since the policy evaluation error is only one component of the total error $|\\hat{Q} - Q^*|$, what do you think about the other term $|Q^{\\pi^N} - Q^*|$? Is it not as important to assess the behaviour of the actor-critic algorithms? It seems like focusing on policy evaluation is be enough to understand the performance of actor-critic algorithms in certain cases.\n\n- Does the theory recommend any \"optimal weights\"? It would be interesting if the derived bounds could be used to develop a better sampling strategy than 1/age or ERE.\n- On a similary note, have you tried more aggressive recency weights, such as 1/age^c for c > 1? It would seem to decrease the bias even further although at some cost in the variance. Perhaps an adaptive c would be appropriate too.  \n\n- Can the developed theory be used to explain the performance of prioritized ER? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper seeks to improve our understanding of actor-critic methods in deep RL by providing an analysis of the error in estimating the value function. The main theoretical result is an error bound which depends on the Bellman error, variance from sampling and a bias term from the policy mismatch. This error bound suggests that a replay buffer sampling strategy that focuses on recent samples will have an advantage. Experiments validate this hypothesis and using weighted sampling with 1/age weights improves performance. Morover, the theory helps explain the success of the ERE algorithm which also emphasizes recent transitions. \n", "main_review": "This is a nice paper expanding our undertanding of actor-critic algorithms. Focusing on the policy evaluation error is a simple approach yielding meaningful results. The assumptions are sufficiently weak for the conclusions to apply to modern deep RL methods, a strength of this analysis. The experiments are a good addition to validate the theoretical findings and make the case more convincing. Overall, this is a great paper with concrete insights that are clearly conveyed. \n\nQuestions and comments:\n- I think the fact that the analysis is uisng continuous action spaces is neat. This seems to be a less popular choice in RL.\n\n- Algorithm 1 is set up nicely to avoid the complexity of deep RL algorithms but still retain the essence of the algorithms and allowing interesting theoretical derivations.\n\n- p.6 To clarify, why is it stated that $|\\hat{Q} - Q^{\\pi^N}|$  mainly depends on steps $\\ge i$. Is this because of the presence of $\\epsilon^{i,L}_\\hat{Q}$ and $W^{i,L}_1$?\n\n- It would have been nice to have an experiment which directly measures the policy evaluation error. This could be done in a simpler environment where policy evaluation is relatively easy e.g. LQR. This could help support the theory even further.\n\n- How are the curves in figure 1 generated? Are these obtained analytically or are simulations run to obtain then? \nIt seems like for uniform sampling, if we sample only 1 transition at each step, the expected count (aggregate weight) after $n$ steps would be $\\sum_{i=0}^n \\frac{1}{i} \\approx \\log n$. This would imply that the aggregate weight for the transition at time step $i$ would be approximate $\\log n - log i$, which seems to roughly match the shape of the curve plotted.\n\n- p.6 The detailed discussion in the \"Normalization\" paragraph is nice to tie up loose ends.\n\n- p.6-7 Some parts of the \"Interpretation\" paragraph are a bit confusing to me. In particular, I'm not following the part discussing the bias and variance terms and how they impact the optimization process. I understand that, near the beginning of learning, variance may be large relative to bias. But in phase 2, it's not clear to me what happens to the relative magnitudes. It seems like both bias and variance would be small. Also, while lower policy evaluation error leads to more accurate estimations of $Q^{\\pi^N}$, I'm finding it difficult to translate that to how fast optimization progresses. It would seem like more accurate estimates would lead to \"better\" optimization steps but, at the same time, the optimization process will naturally slow down as the iterates approach a maximum. I think this paragraph should be clarified a bit more.\n\n- Since the policy evaluation error is only one component of the total error $|\\hat{Q} - Q^*|$, what do you think about the other term $|Q^{\\pi^N} - Q^*|$? Is it not as important to assess the behaviour of the actor-critic algorithms? It seems like focusing on policy evaluation is be enough to understand the performance of actor-critic algorithms in certain cases.\n\n- Does the theory recommend any \"optimal weights\"? It would be interesting if the derived bounds could be used to develop a better sampling strategy than 1/age or ERE.\n- On a similary note, have you tried more aggressive recency weights, such as 1/age^c for c > 1? It would seem to decrease the bias even further although at some cost in the variance. Perhaps an adaptive c would be appropriate too.  \n\n- Can the developed theory be used to explain the performance of prioritized ER? \n", "summary_of_the_review": "The paper presents a convincing analysis of replay sampling in actor-critic algorithms with theoretical results applicable to modern deep RL algorithms. I think it would be a nice addition to the literature.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636000941616}, {"id": "Z4kA7bbpRL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1960/Reviewer_VgPP"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studies the bias variance perspective of off-policy actor-critic algorithms. The policy mismatch in off-policy learning leads to a variance issue (often due to importance sampling) while the bias is due to the mismatch of the behaviour and the target policies. This paper studies the setting where the behaviour policy is unknown. The paper studies why sampling schemes from the experience replay buffer often work well - since they lead to lower bias and variance issues than uniform sampling from the buffer.  The key contribution of this work is to provide theoretical justification, for off-policy actor-critic algorithms from a bias-variance perspective. \n", "review_text": "Overall comments : \n\n\t1. This paper studies policy evaluation errors for off-policy actor-critic algorithms. The key argument is that the policy evaluation error for actor-critic in off-policy has not been studied well. While this is true, I would have expected some empirical analysis, to justify the theoretical claims of the off-policy critic error in the overall algorithms. For example, how does the critic evaluation error in algorithms like SAC/DDPG lead to the performance in the policy improvement step? Does poor estimation of the critic lead to errors in algorithms like SAC? I think a careful analysis of this is what is expected from a paper trying to analysue bias, variance and policy evaluation error of algorithms like SAC/DDPG etc which has strong empirical performance. \n\n\t2. Theorem 1 characterizes the effect of the policy evaluation error in the off-policy actor-critic setting. While this result is important to characterize - I am not sure if it adds too much value in the overall scheme of things. For example, it'd have been better to see how this error leads to the convergence/divergence of off-policy algorithms? Otherwise, this is similar to any Q function error, as in TD learning, with the additional terms due to the policy mismatch term W (but this theorem result is what we'd expect anyway?)\n\n\t3. Theorem 1 and the Corollary are useful to see how the overall off-policy learning error can be characterized in terms of the bias and variance. I think this result is important to exactly see the effect policy mismatch and the induced variance has in the overall learning algorithm. \n\n\t4. The experimental analysis of this paper is rather weak. The only strong result here is Theorem 1 - but it is difficult to justify the usefulness of it. In the experimental studies, I expected ablation analysis demonstrating the effect of this variance term and how the critic evaluation error influences the performance of standard off-policy actor-critic algorithms like SAC/DDPG.\n\n\t5. The paper discusses the different sampling schemes, and I thought the motivation was to see why these sampling schemes often work well in practice. However, from the context of the paper, it is not clear how this contribution is justified - I do not see any clear statements/results backing this claim?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the bias variance perspective of off-policy actor-critic algorithms. The policy mismatch in off-policy learning leads to a variance issue (often due to importance sampling) while the bias is due to the mismatch of the behaviour and the target policies. This paper studies the setting where the behaviour policy is unknown. The paper studies why sampling schemes from the experience replay buffer often work well - since they lead to lower bias and variance issues than uniform sampling from the buffer.  The key contribution of this work is to provide theoretical justification, for off-policy actor-critic algorithms from a bias-variance perspective. \n", "main_review": "Overall comments : \n\n\t1. This paper studies policy evaluation errors for off-policy actor-critic algorithms. The key argument is that the policy evaluation error for actor-critic in off-policy has not been studied well. While this is true, I would have expected some empirical analysis, to justify the theoretical claims of the off-policy critic error in the overall algorithms. For example, how does the critic evaluation error in algorithms like SAC/DDPG lead to the performance in the policy improvement step? Does poor estimation of the critic lead to errors in algorithms like SAC? I think a careful analysis of this is what is expected from a paper trying to analysue bias, variance and policy evaluation error of algorithms like SAC/DDPG etc which has strong empirical performance. \n\n\t2. Theorem 1 characterizes the effect of the policy evaluation error in the off-policy actor-critic setting. While this result is important to characterize - I am not sure if it adds too much value in the overall scheme of things. For example, it'd have been better to see how this error leads to the convergence/divergence of off-policy algorithms? Otherwise, this is similar to any Q function error, as in TD learning, with the additional terms due to the policy mismatch term W (but this theorem result is what we'd expect anyway?)\n\n\t3. Theorem 1 and the Corollary are useful to see how the overall off-policy learning error can be characterized in terms of the bias and variance. I think this result is important to exactly see the effect policy mismatch and the induced variance has in the overall learning algorithm. \n\n\t4. The experimental analysis of this paper is rather weak. The only strong result here is Theorem 1 - but it is difficult to justify the usefulness of it. In the experimental studies, I expected ablation analysis demonstrating the effect of this variance term and how the critic evaluation error influences the performance of standard off-policy actor-critic algorithms like SAC/DDPG.\n\n\t5. The paper discusses the different sampling schemes, and I thought the motivation was to see why these sampling schemes often work well in practice. However, from the context of the paper, it is not clear how this contribution is justified - I do not see any clear statements/results backing this claim?\n", "summary_of_the_review": "I think this paper is interesting and careful studies of existing algorithms that work well in practice is required. However, this work needs more analysis (either experimentally or theoretical justifications) for it to be ready for acceptance. There is only one theory result (Theorem 1) which characterizes the off-policy actor-critic algorithms. However, this result statement is itself not novel and is derived from a vast majority of similar results in RL literature. In practice, I expected a lot more ablation studies characterizing the bias-variance trade-off. The paper perhaps has more claims than what it could demonstrate - and clearly needs more work for it to be ready for acceptance. I would encourage the authors to pick some simpler tasks and demonstrate the bias/variance analyais on some simple mdps too - instead of only showing performance curves for different sampling schemes on some standard mujoco tasks (which often is hard to interpret and not clear how these results are supporting the advertised claim)", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635814044493}, {"id": "Lw02O7gbY7y", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1960/Reviewer_e5zQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper studies the policy evaluation error in off-policy actor critic algorithms. The core idea is to decompose the error into three terms: bellman error, policy mismatch bias and sampling variance. Based on the decomposition, the paper justifies ERE, a recently proposed replay sampling heuristic and empirically studies the properties of a number of other sampling methods.", "review_text": "========== Novelty =============\n\nThe idea of adjusting sampling probability of transitions in the replay buffer for off-policy learning is not new, and has been explored in a number of orthogonal ways as also mentioned in the current paper. This paper investigates a theoretically motivated approach for decomposing the evaluation errors and studies empirical properties of a number of other sampling methods.\n\n========== Literature reviews =============\n\nAs mentioned by the authors, off-policy evaluation learning has been categorized into importance sampling based and regression based. However, in practice, it is more common to combine the aforementioned ideas with contraction operators to carry out scalable off-policy learning when combined with function approximations. Notable examples include n-step Q-learning [1], Retrace [2], Peng's Q [3] among others [4-6]. I think they might be of direct interest if the authors' attempt is to analyze the stability of off-policy learning algorithms.\n\n[1] Hessel et al, Rainbow: combining improvements in deep RL, 2018\n[2] Munos et al, Safe and sample efficient RL, 2016\n[3] Kozuno et al, Revisiting Peng's Q lambda for modern RL, 2021\n[4] Harutyunyan et al, Q(lambda) with off-policy corrections, 2016\n[5] Tang et al, Taylor expansion policy optimization, 2020\n[6] Rowland et al, Adaptive trade-offs in off-policy learning, 2019\n\nThe author has also missed a ref [7] on connecting replay with loss functions.\n\n[7] Fujimoto et al, An equivalence between loss function and non-uniform sampling, 2021\n\n========== Detailed questions =============\n\n1. Overall, the notations are a bit confusing throughout the paper. The paper uses rho_{rho_i} to represent discounted visitation distribution rho starting from rho_i, and uses rho_i to represent the discounted visitation distribution at step i. Putting such notations together makes it quite difficult to read in general. I'd suggest simplifying the notations by e.g., getting rid of a general initial distribution but focusing on a single starting state s_0, or adopting some other ways to represent rho_{rho_i}.\n\n2. The indices e and i are also a bit counterintuitive. I'd suggest using i for the update iteration and t for the time stpe.\n\n3. Algo 1 does not necessarily reflect an exact off-policy actor-critic algorithm used in practice. In particular, the data collection policy might be a perturbed policy (e.g., epsilon-greedy or gaussian corrupted version of the target policy pi^e). I think it is worth making this clear in the presentation.\n\n4. In Eqn 6, an \"average\" behavior policy is defined using all previous policies. One question I have in mind is the utility of defining such a distribution -- though by combining all discounted visitation distribution, we can define an average policy, this policy does not necessarily induce the average discounted visitation distribution. In other words, there can be an average discounted visitation dist which cannot be realized by any Markovian / non-stationary policy. \n\n5. Based on Thm 1, it seems that the evaluation always suffers from an error that is (1/1-gamma) * W, where W is the wasserstain distance mismatch between the current policy pi^N and average policy pi^D. This observation is based on the last term of the RHS in Thm 1. Does it mean that the error is irreducible when pi^N mismatches with pi^D (i.e., W>0)? I am not quite sure if this is true in the tabular case, in which the evaluation error should vanish as more samples come in -- as long as all state-action pairs are visited, the evaluation error should decrease and vanish. It is not clear how the result in Thm 1 is consistent in this case\n\n6. Though I appreciate that the authors try to connect theory with practice, I find the jump from Sec 4 to Sec 5 to be rather abrupt. In particular, I don't see why ERE's implementation of the sampling scheme directly relates to the error bounds in Sec 4. Can we say ERE seeks to minimize the error bound?\n\nA rather important factor here is that, intuitively, it is kind of clear why putting emphasis on recent experiences makes sense -- this is because overall RL algorithms prefer on-policy data over off-policy data to be more stable. In fact, even algorithms which are off-policy by design can perform much better when using near on-policy data, such as VMPO (over MPO) and R2D2 (over DQN). Therefore, the intuition of using recent experience is already there in the literature. What can be valuable here is how detailed implementation practices seeks to optimize a theoretically justified bound. Unfortunately, it is not reflected well here.\n\n7. Fig 2 shows the difference of different algorithms. It seems that different sampling methods do not have that much of a significant difference in the performance. Overall, the empirical observations do not make a convincing case as to why methods such as ERE matter.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies the policy evaluation error in off-policy actor critic algorithms. The core idea is to decompose the error into three terms: bellman error, policy mismatch bias and sampling variance. Based on the decomposition, the paper justifies ERE, a recently proposed replay sampling heuristic and empirically studies the properties of a number of other sampling methods.", "main_review": "========== Novelty =============\n\nThe idea of adjusting sampling probability of transitions in the replay buffer for off-policy learning is not new, and has been explored in a number of orthogonal ways as also mentioned in the current paper. This paper investigates a theoretically motivated approach for decomposing the evaluation errors and studies empirical properties of a number of other sampling methods.\n\n========== Literature reviews =============\n\nAs mentioned by the authors, off-policy evaluation learning has been categorized into importance sampling based and regression based. However, in practice, it is more common to combine the aforementioned ideas with contraction operators to carry out scalable off-policy learning when combined with function approximations. Notable examples include n-step Q-learning [1], Retrace [2], Peng's Q [3] among others [4-6]. I think they might be of direct interest if the authors' attempt is to analyze the stability of off-policy learning algorithms.\n\n[1] Hessel et al, Rainbow: combining improvements in deep RL, 2018\n[2] Munos et al, Safe and sample efficient RL, 2016\n[3] Kozuno et al, Revisiting Peng's Q lambda for modern RL, 2021\n[4] Harutyunyan et al, Q(lambda) with off-policy corrections, 2016\n[5] Tang et al, Taylor expansion policy optimization, 2020\n[6] Rowland et al, Adaptive trade-offs in off-policy learning, 2019\n\nThe author has also missed a ref [7] on connecting replay with loss functions.\n\n[7] Fujimoto et al, An equivalence between loss function and non-uniform sampling, 2021\n\n========== Detailed questions =============\n\n1. Overall, the notations are a bit confusing throughout the paper. The paper uses rho_{rho_i} to represent discounted visitation distribution rho starting from rho_i, and uses rho_i to represent the discounted visitation distribution at step i. Putting such notations together makes it quite difficult to read in general. I'd suggest simplifying the notations by e.g., getting rid of a general initial distribution but focusing on a single starting state s_0, or adopting some other ways to represent rho_{rho_i}.\n\n2. The indices e and i are also a bit counterintuitive. I'd suggest using i for the update iteration and t for the time stpe.\n\n3. Algo 1 does not necessarily reflect an exact off-policy actor-critic algorithm used in practice. In particular, the data collection policy might be a perturbed policy (e.g., epsilon-greedy or gaussian corrupted version of the target policy pi^e). I think it is worth making this clear in the presentation.\n\n4. In Eqn 6, an \"average\" behavior policy is defined using all previous policies. One question I have in mind is the utility of defining such a distribution -- though by combining all discounted visitation distribution, we can define an average policy, this policy does not necessarily induce the average discounted visitation distribution. In other words, there can be an average discounted visitation dist which cannot be realized by any Markovian / non-stationary policy. \n\n5. Based on Thm 1, it seems that the evaluation always suffers from an error that is (1/1-gamma) * W, where W is the wasserstain distance mismatch between the current policy pi^N and average policy pi^D. This observation is based on the last term of the RHS in Thm 1. Does it mean that the error is irreducible when pi^N mismatches with pi^D (i.e., W>0)? I am not quite sure if this is true in the tabular case, in which the evaluation error should vanish as more samples come in -- as long as all state-action pairs are visited, the evaluation error should decrease and vanish. It is not clear how the result in Thm 1 is consistent in this case\n\n6. Though I appreciate that the authors try to connect theory with practice, I find the jump from Sec 4 to Sec 5 to be rather abrupt. In particular, I don't see why ERE's implementation of the sampling scheme directly relates to the error bounds in Sec 4. Can we say ERE seeks to minimize the error bound?\n\nA rather important factor here is that, intuitively, it is kind of clear why putting emphasis on recent experiences makes sense -- this is because overall RL algorithms prefer on-policy data over off-policy data to be more stable. In fact, even algorithms which are off-policy by design can perform much better when using near on-policy data, such as VMPO (over MPO) and R2D2 (over DQN). Therefore, the intuition of using recent experience is already there in the literature. What can be valuable here is how detailed implementation practices seeks to optimize a theoretically justified bound. Unfortunately, it is not reflected well here.\n\n7. Fig 2 shows the difference of different algorithms. It seems that different sampling methods do not have that much of a significant difference in the performance. Overall, the empirical observations do not make a convincing case as to why methods such as ERE matter.\n", "summary_of_the_review": "Overall, I think the paper is lacking in a few aspects.\n\n1. Theory is not presented in a very clear way. I think the theory results are not very convincing, in that it does not reconcile with certain intuitions. However, I might be missing something here and am curious to hear what the authors say.\n\n2. Theory does not connect with practice. A major point of the paper, to my understanding, is that it entails a potentially theoretically sound framework to explain replay practices. However, I don't think the authors have established a convincing case here -- in understanding the practice, the theory here does not provide much more information than plain intuitions we already have.\n\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No concerns.", "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635607305919}, {"id": "q_LzsG_hJJt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1960/Reviewer_b574"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes to decompose the critic error into several terms to better understand off-policy actor-critic algorithms. Empirical results are provided to show how their theory can be used to explain two experience replay sampling strategies.", "review_text": "This paper is very confusing. \n\nFirst, the title is about explaining off-policy actor critic. However, all the theories discuss only the critic error. In my opinion, the authors should focus on the policy evaluation problem with data from mixing distributions because the theory has noting to do with the actor.\n\nSecond, Algorithm 1 does not make sense to me. In Line 3, it involves the true Bellman operator, which is not available because we do not know the transition kernel. It also does not specify the function class of Q, so I'll assume the argmin is done over all possible Q function. Then shouldn't the argmin simply return the true Q function of the policy \\pi^e, because the true Q function apparently gives 0 for the objective inside the argmin? This confusion makes it hard to interpret Theorem 1. First Theorem 1 does not specify what \\hat Q it talks about, I feel it cannot be an arbitrary \\hat Q because otherwise the error shouldn't depend on N. For now I assume it's the \\hat Q generated by Line 3 in Algorithm 1. Then as discussed before, that argmin should return the true Q function, then the error should simply be 0. \n\nThird, I do not understand the motivation of this work. What do we get from the proposed decomposition? It looks it's just an understanding of ERE, which I don't think is enough for publication. Further, it would be good to explain how many seeds are used to generate the plots. For now it is hard to interpret the results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to decompose the critic error into several terms to better understand off-policy actor-critic algorithms. Empirical results are provided to show how their theory can be used to explain two experience replay sampling strategies.", "main_review": "This paper is very confusing. \n\nFirst, the title is about explaining off-policy actor critic. However, all the theories discuss only the critic error. In my opinion, the authors should focus on the policy evaluation problem with data from mixing distributions because the theory has noting to do with the actor.\n\nSecond, Algorithm 1 does not make sense to me. In Line 3, it involves the true Bellman operator, which is not available because we do not know the transition kernel. It also does not specify the function class of Q, so I'll assume the argmin is done over all possible Q function. Then shouldn't the argmin simply return the true Q function of the policy \\pi^e, because the true Q function apparently gives 0 for the objective inside the argmin? This confusion makes it hard to interpret Theorem 1. First Theorem 1 does not specify what \\hat Q it talks about, I feel it cannot be an arbitrary \\hat Q because otherwise the error shouldn't depend on N. For now I assume it's the \\hat Q generated by Line 3 in Algorithm 1. Then as discussed before, that argmin should return the true Q function, then the error should simply be 0. \n\nThird, I do not understand the motivation of this work. What do we get from the proposed decomposition? It looks it's just an understanding of ERE, which I don't think is enough for publication. Further, it would be good to explain how many seeds are used to generate the plots. For now it is hard to interpret the results.", "summary_of_the_review": "The presentation and motivation of the work is not clear", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635349602820}], "openreview_url": "https://openreview.net/forum?id=ZAA0Ol4z2i4", "arxiv_id": "2110.02421", "paper_pdf": "papers/ZAA0Ol4z2i4.pdf", "paper_pdf_sha256": "ddcf988c2390ec48386e434e20ab08079f8319b5496670d2c91e7eac2e743c8e", "paper_pdf_bytes": 676711, "paper_pdf_source": "openreview", "code_url": "https://github.com/sunfex/weighted-sac", "code_repository": "sunfex/weighted-sac", "code_commit": "2cc75a420f09b5a0a17bd25021e70db1425985a8", "code_archive": "repos/ZAA0Ol4z2i4.zip", "code_archive_sha256": "803a2a08daf1e1886bea4b7c1870c89cf06bc2824b3024b49b22655e588ece2d", "code_archive_bytes": 35145, "code_file_count": 7, "code_extensions": {".py": 6, ".sh": 1}, "github_disk_usage_kb": 39, "github_languages": {"Python": 28818, "Shell": 54}, "github_archived": false, "github_pushed_at": "2021-10-02T22:05:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/explaining-off-policy-actor-critic-from-a"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "heFdS9_tkzc", "year": 2021, "status": "rejected", "title": "Distantly Supervised Relation Extraction in Federated Settings", "authors": ["Dianbo Sui", "Yubo Chen", "Kang Liu", "Jun Zhao"], "authorids": ["~Dianbo_Sui1", "~Yubo_Chen1", "~Kang_Liu1", "~Jun_Zhao4"], "authors_source": "OpenReview API", "abstract": "Distant supervision is widely used in relation extraction in order to create a large-scale training dataset by aligning a knowledge base with unstructured text. Most existing studies in this field have assumed there is a great deal of centralized unstructured text. However, in practice, text may be distributed on different platforms and cannot be centralized due to privacy restrictions. Therefore, it is worthwhile to investigate distant supervision in the federated learning paradigm, which decouples the training of the model from the need for direct access to the raw text. However, overcoming label noise of distant supervision becomes more difficult in federated settings, because the sentences containing the same entity pair scatter around different platforms. In this paper, we propose a federated denoising framework to suppress label noise in federated settings. The core of this framework is a multiple instance learning based denoising method that is able to select reliable sentences via cross-platform collaboration. Various experimental results on New York Times dataset and miRNA gene regulation relation dataset demonstrate the effectiveness of the proposed method.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "eEEHZDHju4t", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper114/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n\nThe paper investigates intersection of federated learning and distant supervision of knowledge graphs from texts. The main innovation is a simple yet empirically effective denoising rule that selects only the sentences deemed to be the most reliable for learning in each round of training.\n\n\nStrong points\n* The method is very simple to implement therefore it might be widely adopted as a baseline.\n* The denoising step is specific only for federated learning and distant supervision. It can be applied to many other domains as well (and not just for relation extraction).\n\nWeak points\n* The paper would benefit from further proofreading. \n* Some uncertainty about experimental results, namely repeatability of the runs and methodology of hyperparam selection (more in questions section).\n\n\nRecommendation\n\n* Pre rebuttal: I like the simplistic approach however in current form I am leaning towards rejecting the paper due to my uncertainty in the empirical evaluation. However I am willing to change my evaluation if these questions are resolved.\n\n* Post rebuttal update:  Several of my concerns about clarity of the experimental section were addressed. Therefore I increased my score and now I am inclining towards accepting the paper.\n\nQuestions\n\n* Methodology of obtaining the empirical results isn't clear from the text. Section 4.1 mentions a single held out test set. However then it isn't clear how hyperparam sweep discussed in Sec 4.2 was done. If there is only a single test set were the hyperparams selected on the same test set that is later used to report results in Fig 2 and 3? I would consider that a flaw. Or is there a second validation set (that isn't mentioned in the text explicitly) used to tune hyperparams?\n\n* Curves in Fig 2 and 3 seem to be from just one run for each setting. Doing more runs with different random initializations and reporting mean (or top k results) + confidence intervals would increase confidence in these results.\n\n\nPossible improvements\n\n* \"Lazy MIL almost does not leak the corpus information in each platform\" --- can you make this statement more precise? \n\n* Baseline attention models are always forced to spread their attention among sentences in the local bag even if they are all irrelevant. What if the attention module is allowed to output \"all sentences are irrelevant\" option (this is sometimes called 'sentinel', https://arxiv.org/abs/1612.01887). This might work as learnable denoising that deals with the issue that some local bags don't include any relevant examples. \n\n* Add references for all datasets in section 4.1.\n* Alg 1: are lines 6 and 7 needed? they seem to be contained by line 8.\n* Alg 1, line 11: this piece of Python in pseudocode might not be readable by everyone, I would explain it in plain words.\n* Alg 2: Hyperparams --- \"E is the number\" (what number?)\n\nSome language issues:\n\n* leads to catastrophic repro- ducibility -> is not reproducible\n* is a large biomedical with -> biomedical dataset with?\n* KB is public available -> publicly available\n* the token representation is represented as -> the token is represented as?\n* The overall of Lazy MIL is illustrated in Algorithm 1. --- overall design of? overview of?\n* \"to extract the high-level sentence representation from three segments of CNN outputs, and the boundaries of segments are determined by the positions of the two entities.\" -> I would first define what segments are and only after that how they are used in NN, here the order is reversed and I find it more difficult to follow.\n* Konečný (different accentation in the last character)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple and probably efficient idea, some questions about evaluation ", "review": "Summary\n\nThe paper investigates intersection of federated learning and distant supervision of knowledge graphs from texts. The main innovation is a simple yet empirically effective denoising rule that selects only the sentences deemed to be the most reliable for learning in each round of training.\n\n\nStrong points\n* The method is very simple to implement therefore it might be widely adopted as a baseline.\n* The denoising step is specific only for federated learning and distant supervision. It can be applied to many other domains as well (and not just for relation extraction).\n\nWeak points\n* The paper would benefit from further proofreading. \n* Some uncertainty about experimental results, namely repeatability of the runs and methodology of hyperparam selection (more in questions section).\n\n\nRecommendation\n\n* Pre rebuttal: I like the simplistic approach however in current form I am leaning towards rejecting the paper due to my uncertainty in the empirical evaluation. However I am willing to change my evaluation if these questions are resolved.\n\n* Post rebuttal update:  Several of my concerns about clarity of the experimental section were addressed. Therefore I increased my score and now I am inclining towards accepting the paper.\n\nQuestions\n\n* Methodology of obtaining the empirical results isn't clear from the text. Section 4.1 mentions a single held out test set. However then it isn't clear how hyperparam sweep discussed in Sec 4.2 was done. If there is only a single test set were the hyperparams selected on the same test set that is later used to report results in Fig 2 and 3? I would consider that a flaw. Or is there a second validation set (that isn't mentioned in the text explicitly) used to tune hyperparams?\n\n* Curves in Fig 2 and 3 seem to be from just one run for each setting. Doing more runs with different random initializations and reporting mean (or top k results) + confidence intervals would increase confidence in these results.\n\n\nPossible improvements\n\n* \"Lazy MIL almost does not leak the corpus information in each platform\" --- can you make this statement more precise? \n\n* Baseline attention models are always forced to spread their attention among sentences in the local bag even if they are all irrelevant. What if the attention module is allowed to output \"all sentences are irrelevant\" option (this is sometimes called 'sentinel', https://arxiv.org/abs/1612.01887). This might work as learnable denoising that deals with the issue that some local bags don't include any relevant examples. \n\n* Add references for all datasets in section 4.1.\n* Alg 1: are lines 6 and 7 needed? they seem to be contained by line 8.\n* Alg 1, line 11: this piece of Python in pseudocode might not be readable by everyone, I would explain it in plain words.\n* Alg 2: Hyperparams --- \"E is the number\" (what number?)\n\nSome language issues:\n\n* leads to catastrophic repro- ducibility -> is not reproducible\n* is a large biomedical with -> biomedical dataset with?\n* KB is public available -> publicly available\n* the token representation is represented as -> the token is represented as?\n* The overall of Lazy MIL is illustrated in Algorithm 1. --- overall design of? overview of?\n* \"to extract the high-level sentence representation from three segments of CNN outputs, and the boundaries of segments are determined by the positions of the two entities.\" -> I would first define what segments are and only after that how they are used in NN, here the order is reversed and I find it more difficult to follow.\n* Konečný (different accentation in the last character)", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604604554763}, {"id": "j7K7HVGrnp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper114/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses relation extraction problem for distributed platforms for privacy concerns. The authors propose to leverage federated learning with denoising techinques for better performance. The relation extractor is set to the conventional piece-wise CNN and the main contributions focus on dealing noises in the distributed settings.\n\nPros:\n* A new method for relation extraction in the federated learning to address privacy concerns.\n* Best performance compared to baseline models.\n\nCons:\n* Lack of novelty. This method follows the federated learning paradigm with few improvements. The main contribution is limited to the proposed lazy multiple instance learning, which is not specific to relation extraction problem. If this is not specified for relation extraction, more tasks are expected to demonstrate it is a general algorithm.\n* Experiments are insufficient. First, it is doubtful whether the i.i.d. setting is a practical assumption for relation extraction, because in the real world the quality of corpus in different platforms may vary drastically. Second, given the i.i.d. setting, the authors can repeat their experiments to reduce randomness, because the proposed denoising strategy may only perform well in situations that useful sentences are allocated to only few platforms. The authors may provide more analysis on the distribution of contributions from platforms.\n* Writing could be improved. In section 3.3, what does the ``value v^i'' refers to? Also, if v^i > v^j, then the id^i-th sentence in the platform i is selected...''. Does it mean that at every round only one sentence from a platform is selected? What if for a certain relation, there are many promising candidate sentences from one platform?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good problem but improvements are expected", "review": "This paper addresses relation extraction problem for distributed platforms for privacy concerns. The authors propose to leverage federated learning with denoising techinques for better performance. The relation extractor is set to the conventional piece-wise CNN and the main contributions focus on dealing noises in the distributed settings.\n\nPros:\n* A new method for relation extraction in the federated learning to address privacy concerns.\n* Best performance compared to baseline models.\n\nCons:\n* Lack of novelty. This method follows the federated learning paradigm with few improvements. The main contribution is limited to the proposed lazy multiple instance learning, which is not specific to relation extraction problem. If this is not specified for relation extraction, more tasks are expected to demonstrate it is a general algorithm.\n* Experiments are insufficient. First, it is doubtful whether the i.i.d. setting is a practical assumption for relation extraction, because in the real world the quality of corpus in different platforms may vary drastically. Second, given the i.i.d. setting, the authors can repeat their experiments to reduce randomness, because the proposed denoising strategy may only perform well in situations that useful sentences are allocated to only few platforms. The authors may provide more analysis on the distribution of contributions from platforms.\n* Writing could be improved. In section 3.3, what does the ``value v^i'' refers to? Also, if v^i > v^j, then the id^i-th sentence in the platform i is selected...''. Does it mean that at every round only one sentence from a platform is selected? What if for a certain relation, there are many promising candidate sentences from one platform?", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604037478749}, {"id": "mrStXnZp5u-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper114/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis work focuses on investigating distant supervision for relation extraction task within federated learning paradigm. The proposed Lazy Multi-instance Learning approach identifies reliable sentences for the same entity pairs across platforms to denoise distantly supervised data and perform relation extraction.  \n\nThe proposed approach has been applied on two relation extraction datasets and has reported gains. \n\n\nStrengths:\n- The core of distant supervision for relation extraction is to denoise sentences for building a supervised classifier. The problem is well investigated in NLP community. However, this work focuses on modeling the distant supervision problem in real-world environment addressing data decentralization, ownership and privacy. I buy the idea of applying the existing methods or extend the NLP applications in federated setup. \n- the paper is well written, clearly motivation and addressed real-world problem \n- interesting work: blend of distant supervision and federated learning for NLP task\n\nWeaknesses:\n- the paper lacks novelty in terms of methodology \n- Distant supervision for relation extraction (PCNN and MIL) and the federated learning methodologies used in this work have been inspired from existing works. This works combines the two.\n- the experimental setup needs more clarity \n\n\nQuestions:\n1. Assumption made: Sentences with same entity pairs must scatter across platforms. How would the Lazy MIL system bootstrap without such an assumption (no same entity pairs)?\n2. Assumption made: The number of output classes are same therefore the \\Theta parameter (including output-layer weights connecting softmax) is used across platforms. In real world, it is not the case. Each platform (or customer) may have different relation types. How would you aggregate/distribute the output layer parameters?\n3. How to obtain a minimum viable global \\Theta? How to init this without having the actual training data in Master server?  \n4. In equation 2, the v^i is not controlled by any threshold. Will it denoise if its value is too low (say <0.50) and it is still the highest among all the platforms? \n5. In section 4.1 in data partitioning, how do you ensure the number of relation types remains same across all platforms?\n\nAdditional comments: \n\n- Include an ablation study analyzing scores due to different values of K.\n\nExperimental setup  unclear:\n- how do to init \\Theta in global model?\n- What is the held-out data set for local and global model training?\n\nResults:\n- the experimental results show noticeable gains. It is surprising as the NYT dataset is well investigated in distant supervision settings. \n- As the training data is split across platforms, the overall system is decoupled into several local models. In essence, the overall performance in federated settings should deteriorate or remain competitive to the baseline models due to 'no joint' training on the overall corpora as well as due to (somewhat) lossy aggregation. Please provide a detailed reasoning about the noticeable gains achieved.   \n\nReproducibility\n- no code available", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work, Lacks novelty and unclear experimental setup", "review": "Summary:\n\nThis work focuses on investigating distant supervision for relation extraction task within federated learning paradigm. The proposed Lazy Multi-instance Learning approach identifies reliable sentences for the same entity pairs across platforms to denoise distantly supervised data and perform relation extraction.  \n\nThe proposed approach has been applied on two relation extraction datasets and has reported gains. \n\n\nStrengths:\n- The core of distant supervision for relation extraction is to denoise sentences for building a supervised classifier. The problem is well investigated in NLP community. However, this work focuses on modeling the distant supervision problem in real-world environment addressing data decentralization, ownership and privacy. I buy the idea of applying the existing methods or extend the NLP applications in federated setup. \n- the paper is well written, clearly motivation and addressed real-world problem \n- interesting work: blend of distant supervision and federated learning for NLP task\n\nWeaknesses:\n- the paper lacks novelty in terms of methodology \n- Distant supervision for relation extraction (PCNN and MIL) and the federated learning methodologies used in this work have been inspired from existing works. This works combines the two.\n- the experimental setup needs more clarity \n\n\nQuestions:\n1. Assumption made: Sentences with same entity pairs must scatter across platforms. How would the Lazy MIL system bootstrap without such an assumption (no same entity pairs)?\n2. Assumption made: The number of output classes are same therefore the \\Theta parameter (including output-layer weights connecting softmax) is used across platforms. In real world, it is not the case. Each platform (or customer) may have different relation types. How would you aggregate/distribute the output layer parameters?\n3. How to obtain a minimum viable global \\Theta? How to init this without having the actual training data in Master server?  \n4. In equation 2, the v^i is not controlled by any threshold. Will it denoise if its value is too low (say <0.50) and it is still the highest among all the platforms? \n5. In section 4.1 in data partitioning, how do you ensure the number of relation types remains same across all platforms?\n\nAdditional comments: \n\n- Include an ablation study analyzing scores due to different values of K.\n\nExperimental setup  unclear:\n- how do to init \\Theta in global model?\n- What is the held-out data set for local and global model training?\n\nResults:\n- the experimental results show noticeable gains. It is surprising as the NYT dataset is well investigated in distant supervision settings. \n- As the training data is split across platforms, the overall system is decoupled into several local models. In essence, the overall performance in federated settings should deteriorate or remain competitive to the baseline models due to 'no joint' training on the overall corpora as well as due to (somewhat) lossy aggregation. Please provide a detailed reasoning about the noticeable gains achieved.   \n\nReproducibility\n- no code available", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603932076105}, {"id": "fEzvLpRrpZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper114/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper explores relation extraction with distant supervision in the federated setting and focuses on handling the label noise problem from automatic distant supervision by Multiple Instance Learning-based methods and proposes Lazy MIL. \n\nFor a specific triple (h, r, t) in KB, values (probability with relation r) for sentences containing h and t will be calculated locally in each platform, and such best value and index in each platform will be uploaded to the master server. A master server decides the most reliable sentence and broadcast such information to all activated platforms in this round.\n\nStrengths: \n\n+ This paper considers relation extraction in the federated setting, which is a new direction in this area;\n+ The mentioned label noise problem is a real problem in distant supervision, and it’s intuitive that this problem exists in the federated setting;\n+ Experiments show the effectiveness of the proposed method;\n\nWeaknesses:\n\n- The contribution of the paper is weak (in the context of the expectations of ICLR) \n- For baselines in experiments, it’s not clear how to use baselines in the federated setting. For example, for “keep the other modules unchanged and only replace the denoising module” in the last sentence of Sec. 4.3, it’s confusing for me that how to replace this part. Is the part for broadcasting denoising information? \n- Except for the experiment results, it’s not clear that if the proposed method is specific to the federated setting, and what are the federated-setting-specific designed compared with other baselines?\n\nQuestions:\n\nIn Sec. 4.4, “We believe the reason is that our denoising method can hinder false-positive instances from poisoning local models.” Does this mean that other baselines can’t hinder false-positive instances?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting paper but only contributes to a limited field.", "review": "This paper explores relation extraction with distant supervision in the federated setting and focuses on handling the label noise problem from automatic distant supervision by Multiple Instance Learning-based methods and proposes Lazy MIL. \n\nFor a specific triple (h, r, t) in KB, values (probability with relation r) for sentences containing h and t will be calculated locally in each platform, and such best value and index in each platform will be uploaded to the master server. A master server decides the most reliable sentence and broadcast such information to all activated platforms in this round.\n\nStrengths: \n\n+ This paper considers relation extraction in the federated setting, which is a new direction in this area;\n+ The mentioned label noise problem is a real problem in distant supervision, and it’s intuitive that this problem exists in the federated setting;\n+ Experiments show the effectiveness of the proposed method;\n\nWeaknesses:\n\n- The contribution of the paper is weak (in the context of the expectations of ICLR) \n- For baselines in experiments, it’s not clear how to use baselines in the federated setting. For example, for “keep the other modules unchanged and only replace the denoising module” in the last sentence of Sec. 4.3, it’s confusing for me that how to replace this part. Is the part for broadcasting denoising information? \n- Except for the experiment results, it’s not clear that if the proposed method is specific to the federated setting, and what are the federated-setting-specific designed compared with other baselines?\n\nQuestions:\n\nIn Sec. 4.4, “We believe the reason is that our denoising method can hinder false-positive instances from poisoning local models.” Does this mean that other baselines can’t hinder false-positive instances?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603848954729}, {"id": "UFTUBSkvxFd", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper114/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper introduced a new federated scenario for distantly supervised relation extraction where extractions come from multiple private resources. And therefore a global model cannot directly access all the data simultaneously. \n\nTo protect data privacy, this paper assigned each resource a local model for separate training. In order to apply the at-least-one sentence bag denoising technique, at the beginning of each round, local models will calculate scores of sentences in the local bag, compare scores with other resources and then generate the training data. After each round of parallel training, local models will be synchronized with the global model via a weighted average algorithm. By only sharing scores rather than the full text, data privacy is protected. Experiments over two datasets show promising results.\n\n####################\n\nReasons for score:\n\nI vote for marginally negative. Overall, this paper brought up an interesting challenge for this NLP sub-task. However, it is not clear to me the necessity of compromising accuracy while doing parallel computing, and model innovation is slightly weak. \n\n####################\n\nThere are two comments from a performance-driven perspective:\n\n1. Is there a specific reason for this task to do parallel computing? From my understanding (if correct), if parallel training is not required, we can avoid major information loss from delayed denoising and averaging step for model integration. In another word, if we only train one local platform at a time (and denoising frequently), and then synchronize model parameters to the next platform for further training, the performance should be close (or equal if using ONE model) to the performance upper bound when K=1.  If parallel computing is the key, then it is necessary to analyze the speed and performance trade-off.\n\n2. It is necessary to use pre-trained language models like MTB (matching the blank) paper as the encoder because improvement over the current baseline could possibly be mitigated by a stronger encoder.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper tackles a new scenario for distant supervision relation extraction where extractions come from multiple private resources thus global training is not available.", "review": "Summary:\nThis paper introduced a new federated scenario for distantly supervised relation extraction where extractions come from multiple private resources. And therefore a global model cannot directly access all the data simultaneously. \n\nTo protect data privacy, this paper assigned each resource a local model for separate training. In order to apply the at-least-one sentence bag denoising technique, at the beginning of each round, local models will calculate scores of sentences in the local bag, compare scores with other resources and then generate the training data. After each round of parallel training, local models will be synchronized with the global model via a weighted average algorithm. By only sharing scores rather than the full text, data privacy is protected. Experiments over two datasets show promising results.\n\n####################\n\nReasons for score:\n\nI vote for marginally negative. Overall, this paper brought up an interesting challenge for this NLP sub-task. However, it is not clear to me the necessity of compromising accuracy while doing parallel computing, and model innovation is slightly weak. \n\n####################\n\nThere are two comments from a performance-driven perspective:\n\n1. Is there a specific reason for this task to do parallel computing? From my understanding (if correct), if parallel training is not required, we can avoid major information loss from delayed denoising and averaging step for model integration. In another word, if we only train one local platform at a time (and denoising frequently), and then synchronize model parameters to the next platform for further training, the performance should be close (or equal if using ONE model) to the performance upper bound when K=1.  If parallel computing is the key, then it is necessary to analyze the speed and performance trade-off.\n\n2. It is necessary to use pre-trained language models like MTB (matching the blank) paper as the encoder because improvement over the current baseline could possibly be mitigated by a stronger encoder.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603770544442}], "openreview_url": "https://openreview.net/forum?id=heFdS9_tkzc", "arxiv_id": "2008.05049", "paper_pdf": "papers/heFdS9_tkzc.pdf", "paper_pdf_sha256": "003b498d2b85ce84b06bbc885441fe95b9dc6c0a7c5853543a6bdb944a5cbbab", "paper_pdf_bytes": 704607, "paper_pdf_source": "openreview", "code_url": "https://github.com/DianboWork/FedDS", "code_repository": "DianboWork/FedDS", "code_commit": "fd5eec9ce2386d83da910854ea5dd7934f39e138", "code_archive": "repos/heFdS9_tkzc.zip", "code_archive_sha256": "d8ef2308082edba1608e6dc87be9f2bb851dd494e65521be26948892783c10b8", "code_archive_bytes": 44395, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 46, "github_languages": {"Python": 148604}, "github_archived": false, "github_pushed_at": "2021-09-18T06:41:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distantly-supervised-relation-extraction-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJl3CANKvB", "year": 2020, "status": "rejected", "title": "A SIMPLE AND EFFECTIVE FRAMEWORK FOR PAIRWISE DEEP METRIC LEARNING", "authors": ["Qi Qi", "Yan Yan", "Zixuan Wu", "Xiaoyu Wang", "Tianbao Yang"], "authorids": ["qi-qi@uiowa.edu", "yanyan.tju@gmail.com", "wuzu@bc.edu", "fanghuaxue@gmail.com", "tianbao-yang@uiowa.edu"], "authors_source": "OpenReview API", "abstract": "Deep metric learning (DML) has received much attention in deep learning due to its wide applications in computer vision. Previous studies have focused on designing complicated losses and hard example mining methods, which are mostly heuristic and lack of theoretical understanding. In this paper, we cast DML as a simple pairwise binary classification problem that classifies a pair of examples as similar or dissimilar. It identifies the most critical issue in this problem---imbalanced data pairs. To tackle this issue, we propose a simple and effective framework to sample pairs in a batch of data for updating the model. The key to this framework is to define a robust loss for all pairs over a mini-batch of data, which is formulated by distributionally robust optimization. The flexibility in constructing the  {\\it uncertainty decision set} of the dual variable allows us to recover state-of-the-art complicated losses and also to induce novel variants.  Empirical studies on several benchmark data sets demonstrate that our simple and effective method outperforms the state-of-the-art results.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJgZmWvTtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1446/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper casts deep metric learning (DML) as a pairwise binary classification problem such that pairs of examples need to be classified as similar or dissimilar. The authors propose an objective function that computes a weighted sum over the pairwise losses in a mini-batch. The weight vector is selected to maximize the objective from a decision set encoding constraints. This formulation is called the distributionally robust optimization (DRO) framework.\n\nThe authors argue that the DRO framework is theoretically justified by showing how certain decision sets result in existing machine learning loss functions. This portion of the paper seemed hand-wavy. It is not clear what is the purpose of including the theorem from Namkoon & Duchi. It would be more clear in my view to just make the short point that a certain decision set recovers the DRO with f-divergence as would be expected. The claims with regard to learning theory are over-stated in the paper.\n\nThe authors proposed three variants of the general framework. They include a top-K formulation, a variance-regularized version, and a top-K version using a balance between positive and negative examples. The DRO framework and the variants are the main contributions in terms of methodology in this paper. It is also shown that the framework generalizes more complicated recently proposed losses.\n\nThe experiments demonstrate the DRO framework consistently outperforms state of the art deep metric learning methods on benchmark datasets by small margins. There is also a computational speed advantage that is shown.\nOverall, this paper shows that the ideas from distributionally robust optimization work well in deep metric learning. In particular, the paper shows that by combining the DRO framework with simple loss functions, performance comparable with complicated loss functions can be obtained. This aspect, along with the generality are the main strong suits. That being said, I do not see this paper to be that significant of a contribution. The main idea in the paper seems like a rather direct application of the DRO modeling framework and it does not provide too significant of improvement over the MS loss.  The paper was not written super clearly and was too long. Reviewers were instructed to apply a higher standard to papers in excess of 8 pages and this paper would have been presented more effectively if it was shorter. For these reasons, I recommended a weak reject.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper casts deep metric learning (DML) as a pairwise binary classification problem such that pairs of examples need to be classified as similar or dissimilar. The authors propose an objective function that computes a weighted sum over the pairwise losses in a mini-batch. The weight vector is selected to maximize the objective from a decision set encoding constraints. This formulation is called the distributionally robust optimization (DRO) framework.\n\nThe authors argue that the DRO framework is theoretically justified by showing how certain decision sets result in existing machine learning loss functions. This portion of the paper seemed hand-wavy. It is not clear what is the purpose of including the theorem from Namkoon & Duchi. It would be more clear in my view to just make the short point that a certain decision set recovers the DRO with f-divergence as would be expected. The claims with regard to learning theory are over-stated in the paper.\n\nThe authors proposed three variants of the general framework. They include a top-K formulation, a variance-regularized version, and a top-K version using a balance between positive and negative examples. The DRO framework and the variants are the main contributions in terms of methodology in this paper. It is also shown that the framework generalizes more complicated recently proposed losses.\n\nThe experiments demonstrate the DRO framework consistently outperforms state of the art deep metric learning methods on benchmark datasets by small margins. There is also a computational speed advantage that is shown.\nOverall, this paper shows that the ideas from distributionally robust optimization work well in deep metric learning. In particular, the paper shows that by combining the DRO framework with simple loss functions, performance comparable with complicated loss functions can be obtained. This aspect, along with the generality are the main strong suits. That being said, I do not see this paper to be that significant of a contribution. The main idea in the paper seems like a rather direct application of the DRO modeling framework and it does not provide too significant of improvement over the MS loss.  The paper was not written super clearly and was too long. Reviewers were instructed to apply a higher standard to papers in excess of 8 pages and this paper would have been presented more effectively if it was shorter. For these reasons, I recommended a weak reject."}, "tcdate": 1571807513465}, {"id": "SyevOXVstr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1446/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors address the (increasingly popular) problem of learning a metric from a given multi-dimensional data set. They consider the deep metric learning setup, where target distances are defined as the euclidean distances in an artificial feature space (created by a deep neural network). Main focus of the paper is to cases where the data set is affected by a substantial imbalance between the amount of examples that are similar to each other and the total number of examples.\n\nI would tend to accept the paper because handling the imbalance problem in metric learning is important and both the theoretical analysis and the experiments show that the proposed method may have some impact.\n\nThe idea of reducing the problem to a binary classification between similar and dissimilar examples may look too simple but i) is a common approach in deep metric learning, ii) helps to handle the implicit imbalance problem and iii) suggests possible generalisations to other network-based problems (for example, where similarity is naturally defined by the existence of absence of a link). Showing that many complicated losses are equivalent to DRO may also help the general understanding of the metric learning task.\n\nMy main concerns are about the net contribution of the paper. Tackling the imbalance problem is important but it is not clear whether the full metric learning setup is really needed. The authors could have stated more precisely in what sense the metric learning unbalanced problem they consider is different from usual unbalanced binary classification. Otherwise, as DRO is well known, it is hard to identify the real novelty of their method.\n\nQuestions: \n- how does the specific metric learning setup make the considered DRO different from usual unbalanced classification? \n- how the network architecture affects the performance? For example, would the size of the embedding space change the recall/imbalance plot? \n- Is the choice of euclidean distances standard in deep metric learning? Would a choice of more general distances be incorporated in the proposed method?\n \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The authors address the (increasingly popular) problem of learning a metric from a given multi-dimensional data set. They consider the deep metric learning setup, where target distances are defined as the euclidean distances in an artificial feature space (created by a deep neural network). Main focus of the paper is to cases where the data set is affected by a substantial imbalance between the amount of examples that are similar to each other and the total number of examples.\n\nI would tend to accept the paper because handling the imbalance problem in metric learning is important and both the theoretical analysis and the experiments show that the proposed method may have some impact.\n\nThe idea of reducing the problem to a binary classification between similar and dissimilar examples may look too simple but i) is a common approach in deep metric learning, ii) helps to handle the implicit imbalance problem and iii) suggests possible generalisations to other network-based problems (for example, where similarity is naturally defined by the existence of absence of a link). Showing that many complicated losses are equivalent to DRO may also help the general understanding of the metric learning task.\n\nMy main concerns are about the net contribution of the paper. Tackling the imbalance problem is important but it is not clear whether the full metric learning setup is really needed. The authors could have stated more precisely in what sense the metric learning unbalanced problem they consider is different from usual unbalanced binary classification. Otherwise, as DRO is well known, it is hard to identify the real novelty of their method.\n\nQuestions: \n- how does the specific metric learning setup make the considered DRO different from usual unbalanced classification? \n- how the network architecture affects the performance? For example, would the size of the embedding space change the recall/imbalance plot? \n- Is the choice of euclidean distances standard in deep metric learning? Would a choice of more general distances be incorporated in the proposed method?\n \n"}, "tcdate": 1571664751435}, {"id": "H1lyrJVHtB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1446/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a framework for deep metric learning. Using ideas from distributionally robust optimization, the loss (in each batch) is the worst case weighted average of all pairwise classification losses, taken over an uncertainty set of possible weights. The framework is shown to be general and encompass various previous approaches. Based on it, the authors propose several new algorithms, which are shown to outperform the SOTA on image retrieval data sets in terms of recall.\n\nThe main contribution of the paper is a unification of previous deep metric learning algorithms, which would be helpful to the community and could inspire new approaches. I found the empirical observation that the proposed algorithms are able to reduce the computation time by nearly half to be compelling. However, apart from DRO-TopK-PN, the proposed algorithms appear to be minor modifications of existing algorithms. \n\nQuestions about the experimental protocol:\n1. Are the results from one run, or averaged over several? Standard errors of the evaluation metrics would be very helpful to judge the improvements made by the algorithms, especially as the algorithms are stochastic due to batching. \n2. The proposed algorithms seem to be similar to those of Fan et al. (2017) and Namkoong and Duchi (2017). Is there a particular reason why they weren’t included in the experiments? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a framework for deep metric learning. Using ideas from distributionally robust optimization, the loss (in each batch) is the worst case weighted average of all pairwise classification losses, taken over an uncertainty set of possible weights. The framework is shown to be general and encompass various previous approaches. Based on it, the authors propose several new algorithms, which are shown to outperform the SOTA on image retrieval data sets in terms of recall.\n\nThe main contribution of the paper is a unification of previous deep metric learning algorithms, which would be helpful to the community and could inspire new approaches. I found the empirical observation that the proposed algorithms are able to reduce the computation time by nearly half to be compelling. However, apart from DRO-TopK-PN, the proposed algorithms appear to be minor modifications of existing algorithms. \n\nQuestions about the experimental protocol:\n1. Are the results from one run, or averaged over several? Standard errors of the evaluation metrics would be very helpful to judge the improvements made by the algorithms, especially as the algorithms are stochastic due to batching. \n2. The proposed algorithms seem to be similar to those of Fan et al. (2017) and Namkoong and Duchi (2017). Is there a particular reason why they weren’t included in the experiments? "}, "tcdate": 1571270455145}], "openreview_url": "https://openreview.net/forum?id=SJl3CANKvB", "arxiv_id": "1912.11194", "paper_pdf": "papers/SJl3CANKvB.pdf", "paper_pdf_sha256": "c0dc6b29592b7834ec9ef2e83d1433cbbc2fb67f873c1e06b879efaa372dcd27", "paper_pdf_bytes": 427281, "paper_pdf_source": "openreview", "code_url": "https://github.com/qiqi-helloworld/A-Simple-and-Effective-Framework-for-Pairewise-Distance-Metric-Learning", "code_repository": "qiqi-helloworld/A-Simple-and-Effective-Framework-for-Pairewise-Distance-Metric-Learning", "code_commit": "e12a417513a6cb4098bd955c0063c24a4c8e3543", "code_archive": "repos/SJl3CANKvB.zip", "code_archive_sha256": "a19f516da2cf3560984bf93098b182f69a61145003c459eb46fc3ced5fcbf60b", "code_archive_bytes": 81756, "code_file_count": 49, "code_extensions": {".py": 46, ".sh": 3}, "github_disk_usage_kb": 64, "github_languages": {"Python": 138404, "Shell": 5653}, "github_archived": false, "github_pushed_at": "2020-05-03T15:54:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-simple-and-effective-framework-for-pairwise-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BklKFo09YX", "year": 2019, "status": "rejected", "title": "Mol-CycleGAN - a generative model for molecular optimization", "authors": ["Łukasz Maziarka", "Agnieszka Pocha", "Jan Kaczmarczyk", "Michał Warchoł"], "authorids": ["l.maziarka@gmail.com", "lamiane.chan@gmail.com", "jan.kaczmarczyk@ardigen.com", "michal.warchol@ardigen.com"], "authors_source": "OpenReview API", "abstract": "Designing a molecule with desired properties is one of the biggest challenges in drug development, as it requires optimization of chemical compound structures with respect to many complex properties. To augment the compound design process we introduce Mol-CycleGAN -- a CycleGAN-based model that generates optimized compounds with a chemical scaffold of interest. Namely, given a molecule our model generates a structurally similar one with an optimized value of the considered property. We evaluate the performance of the model on selected optimization objectives related to structural properties (presence of halogen groups, number of aromatic rings) and to a physicochemical property (penalized logP). In the task of optimization of penalized logP of drug-like molecules our model significantly outperforms previous results. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "Hkxd3QyWTQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper466/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an approach for optimizing molecular properties, based on the application of CycleGANs to variational autoencoders for molecules. A recently proposed domain-specific VAE called Junction Tree VAE (JT-VAE) is employed. The optimization of molecules is an important problem for example in drug discovery or materials design, and the development of machine learning approaches for it is therefore an important application.\n\n\nOn the positive side, this reviewer would like to highlight the structural transformation task, which is an interesting addition. The presentation is mostly clear, and there is an improvement in benchmark scores compared to the baseline JT-VAE.\n\nOn the negative side, this reviewer would argue the paper is a bit incremental, as it appears to be “just” a combination of two models without a clear motivation from a molecular design perspective – why is the model e.g. better (motivated) than the model by You et al, the MOLGAN by DeCao and Kipf, or the CGVAE by Liu et al? \n Also, the number of benchmark experiments is small compared to more rigorous evaluation studies, e.g. by Olivecrona et al. https://doi.org/10.1186/s13321-017-0235-x \n\nFurthermore, a lot of (labeled) training examples for the respective properties seem to be needed, whereas in practical drug discovery, often only very small datasets are available.\n\n\n\nQuestions:\n\nHow exactly is the similarity constraint enforced?\n\nWhy do the authors use the JT-VAE and not the SD-VAE or CGVAE?\n\nCan the authors explain why they think penalized logP is relevant to drug discovery? While the optimization of physicochemical properties alone is certainly important in a multi-objective setting, this reviewer is not entirely sure that physicochemical property optimization alone is a problem that often occurs in drug discovery.\n\nWhy are the values for the RL model by You et al. 2018 not shown in Table 2, even though the authors state that they have done the comparison? If those molecules are not drug-like, it would be good to see the comparison, and show the respective molecules. Keep in mind that you are not asking the models to produce drug-like molecules here! The RL model therefore seems to generate exactly what it is asked (rewarded) for. Is it a fair comparison that you expect it to do something else?\n\nDoes the requirement to define the sets X and Y restrict the capacity of the model to binary class membership, as in “good” and “bad”? Does this imply the model cannot extrapolate into e.g. continuous parameter ranges unseen during training? \n\nProvocative statement this reviewer would be keen to hear the authors’ (and co-reviewers) opinion on: It seems that there are considerable efforts needed in getting VAEs for discrete data such as graphs to work, namely the drastic restriction of the molecular graph space via grammars, CGVAE, or the JT setting (which builds on a heuristically defined vocabulary), whereas RL models even get decent results with simple SMILES-RNNs (see e.g. the work by Merk et al, who have even successfully made and tested their results in the lab!  https://doi.org/10.1038/s42004-018-0068-1 and https://doi.org/10.1002/minf.201700153 ) Do the author have an intuition into why this might be the case?\n\n\nRegarding the proposed application of GANs to text generation, this reviewer would suggest to have a look at this recent (as in yesterday) paper https://arxiv.org/pdf/1811.02549.pdf\n\n\n\nOverall, this reviewer believes the paper would be a good ICLR workshop paper or, after a bit more technical details on the models, a decent chemoinformatics paper, but unfortunately this reviewer is not convinced that in the current form the results are significant or general enough in terms of ML to warrant acceptance in the main ICLR track. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising results, unfortunately not strong enough for main track", "review": "The paper presents an approach for optimizing molecular properties, based on the application of CycleGANs to variational autoencoders for molecules. A recently proposed domain-specific VAE called Junction Tree VAE (JT-VAE) is employed. The optimization of molecules is an important problem for example in drug discovery or materials design, and the development of machine learning approaches for it is therefore an important application.\n\n\nOn the positive side, this reviewer would like to highlight the structural transformation task, which is an interesting addition. The presentation is mostly clear, and there is an improvement in benchmark scores compared to the baseline JT-VAE.\n\nOn the negative side, this reviewer would argue the paper is a bit incremental, as it appears to be “just” a combination of two models without a clear motivation from a molecular design perspective – why is the model e.g. better (motivated) than the model by You et al, the MOLGAN by DeCao and Kipf, or the CGVAE by Liu et al? \n Also, the number of benchmark experiments is small compared to more rigorous evaluation studies, e.g. by Olivecrona et al. https://doi.org/10.1186/s13321-017-0235-x \n\nFurthermore, a lot of (labeled) training examples for the respective properties seem to be needed, whereas in practical drug discovery, often only very small datasets are available.\n\n\n\nQuestions:\n\nHow exactly is the similarity constraint enforced?\n\nWhy do the authors use the JT-VAE and not the SD-VAE or CGVAE?\n\nCan the authors explain why they think penalized logP is relevant to drug discovery? While the optimization of physicochemical properties alone is certainly important in a multi-objective setting, this reviewer is not entirely sure that physicochemical property optimization alone is a problem that often occurs in drug discovery.\n\nWhy are the values for the RL model by You et al. 2018 not shown in Table 2, even though the authors state that they have done the comparison? If those molecules are not drug-like, it would be good to see the comparison, and show the respective molecules. Keep in mind that you are not asking the models to produce drug-like molecules here! The RL model therefore seems to generate exactly what it is asked (rewarded) for. Is it a fair comparison that you expect it to do something else?\n\nDoes the requirement to define the sets X and Y restrict the capacity of the model to binary class membership, as in “good” and “bad”? Does this imply the model cannot extrapolate into e.g. continuous parameter ranges unseen during training? \n\nProvocative statement this reviewer would be keen to hear the authors’ (and co-reviewers) opinion on: It seems that there are considerable efforts needed in getting VAEs for discrete data such as graphs to work, namely the drastic restriction of the molecular graph space via grammars, CGVAE, or the JT setting (which builds on a heuristically defined vocabulary), whereas RL models even get decent results with simple SMILES-RNNs (see e.g. the work by Merk et al, who have even successfully made and tested their results in the lab!  https://doi.org/10.1038/s42004-018-0068-1 and https://doi.org/10.1002/minf.201700153 ) Do the author have an intuition into why this might be the case?\n\n\nRegarding the proposed application of GANs to text generation, this reviewer would suggest to have a look at this recent (as in yesterday) paper https://arxiv.org/pdf/1811.02549.pdf\n\n\n\nOverall, this reviewer believes the paper would be a good ICLR workshop paper or, after a bit more technical details on the models, a decent chemoinformatics paper, but unfortunately this reviewer is not convinced that in the current form the results are significant or general enough in terms of ML to warrant acceptance in the main ICLR track. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541628848485}, {"id": "Byg3Zy8cn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper466/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Main idea:\nThis paper proposes to use CycleGAN to generate a molecule with a desired property given an initial molecule without that property.  This is a task appearing in drug design. Specifically, the training set consists of two sets of molecules with and without a desired property. For each molecule, a representation/embedding is obtained, which is the latent code of the JT-VAE. Two generators/transformers are introduced to perform transformation from the representation of one domain to that of another domain. The training involves a weighted sum of several losses: the LSGAN loss, the cyclic consistency loss and the reconstruction loss, which are exactly the same losses used in the CycleGAN paper.  Experiments are conducted on ZINC-250K dataset and evaluated in terms of structural modifications and molecule optimization.\n\nComments:\nThis is a good application paper using machine learning techniques. However, I don't think ICLR is a suitable venue for this paper since the improvements and contributions are more on molecule optimization side. I suggest that the authors extend the technical part of CycleGAN with more background introduction and submit the paper to a bioinformatics journal. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An application of CycleGAN to molecule transformations", "review": "Main idea:\nThis paper proposes to use CycleGAN to generate a molecule with a desired property given an initial molecule without that property.  This is a task appearing in drug design. Specifically, the training set consists of two sets of molecules with and without a desired property. For each molecule, a representation/embedding is obtained, which is the latent code of the JT-VAE. Two generators/transformers are introduced to perform transformation from the representation of one domain to that of another domain. The training involves a weighted sum of several losses: the LSGAN loss, the cyclic consistency loss and the reconstruction loss, which are exactly the same losses used in the CycleGAN paper.  Experiments are conducted on ZINC-250K dataset and evaluated in terms of structural modifications and molecule optimization.\n\nComments:\nThis is a good application paper using machine learning techniques. However, I don't think ICLR is a suitable venue for this paper since the improvements and contributions are more on molecule optimization side. I suggest that the authors extend the technical part of CycleGAN with more background introduction and submit the paper to a bioinformatics journal. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541197572173}, {"id": "HJe11ceq2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper466/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "A strand of papers has been proposed for optimizing molecules for a given property using different forms of variational autoencoders as a tool to extract embedding for molecules. This paper is also in this direction by applying Cycle-GAN [Zhu et al., 2017] in the latent space learnt from junction-tree VAE [Jin et.al, 2018], aiming to learn a translation function (in the latent space), from the set of molecules without the interested property to the set of molecules with the property.\n\nThe paper is well written and the contribution is more on the application side, although combining the above mentioned two gradients are indeed novel. That being said, I wish the authors can bring more perspectives from chemists or biologists which can justify and support the application setting.\n\nSince the focus is molecular optimization, I paid special attention to the results in Section 4.2. While the improvement regarding to the interested property (logP) is a pleasure to see, the drop of the success rate is also significant comparing to [Jin et.al, 2018], which undermines the significance of the results, especially given the increased complexity. Another issue is the requirement of a large set of molecules with desired property. This further restricts the applicability of the method. How to solve the cold start issue will be critical in this setting and this is not mentioned in the paper. Thirdly, combining two existing components is ok but not enough novelty from my point of view. Considering the high acceptance bar of ICLR, I will not accept this paper.\n\nDetailed comments:\n1. In Section 4.2, how similar the generated molecules are to the ones already in the Y_train? \n2. In Section 4.2, G(X) map to Y, what does it mean to apply G(G(X))? Do you decode G(X) to a molecule first and then feed into the encoder to apply G (as in Section 4.3)? If not, G is not supposed to learn the transition from Y to X. Will you always get exactly the same X when apply F(G(X))? Can the sequence be G(X), G(F(G(X))), …? \n3. In Section 4.3, is it always that later step gives better logP? It seems so from Figure 6 for 1-30 iterations, how about later 30-80 iterations? If so, for the generated molecules, the method seems have a tension between the similarity to the original molecules and the level of the desired property. Can you comment on how important, in practice, the similarity matters?\n4. Following 3, \\lambda_2 seems directly affect the balance of the tension and it should be studied in more detail.\n5. How about reproducibility, will the data and code published?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not good enough", "review": "A strand of papers has been proposed for optimizing molecules for a given property using different forms of variational autoencoders as a tool to extract embedding for molecules. This paper is also in this direction by applying Cycle-GAN [Zhu et al., 2017] in the latent space learnt from junction-tree VAE [Jin et.al, 2018], aiming to learn a translation function (in the latent space), from the set of molecules without the interested property to the set of molecules with the property.\n\nThe paper is well written and the contribution is more on the application side, although combining the above mentioned two gradients are indeed novel. That being said, I wish the authors can bring more perspectives from chemists or biologists which can justify and support the application setting.\n\nSince the focus is molecular optimization, I paid special attention to the results in Section 4.2. While the improvement regarding to the interested property (logP) is a pleasure to see, the drop of the success rate is also significant comparing to [Jin et.al, 2018], which undermines the significance of the results, especially given the increased complexity. Another issue is the requirement of a large set of molecules with desired property. This further restricts the applicability of the method. How to solve the cold start issue will be critical in this setting and this is not mentioned in the paper. Thirdly, combining two existing components is ok but not enough novelty from my point of view. Considering the high acceptance bar of ICLR, I will not accept this paper.\n\nDetailed comments:\n1. In Section 4.2, how similar the generated molecules are to the ones already in the Y_train? \n2. In Section 4.2, G(X) map to Y, what does it mean to apply G(G(X))? Do you decode G(X) to a molecule first and then feed into the encoder to apply G (as in Section 4.3)? If not, G is not supposed to learn the transition from Y to X. Will you always get exactly the same X when apply F(G(X))? Can the sequence be G(X), G(F(G(X))), …? \n3. In Section 4.3, is it always that later step gives better logP? It seems so from Figure 6 for 1-30 iterations, how about later 30-80 iterations? If so, for the generated molecules, the method seems have a tension between the similarity to the original molecules and the level of the desired property. Can you comment on how important, in practice, the similarity matters?\n4. Following 3, \\lambda_2 seems directly affect the balance of the tension and it should be studied in more detail.\n5. How about reproducibility, will the data and code published?\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541175766651}], "openreview_url": "https://openreview.net/forum?id=BklKFo09YX", "arxiv_id": "1902.02119", "paper_pdf": "papers/BklKFo09YX.pdf", "paper_pdf_sha256": "1d54c705989d1df4520b8c8721288cce8f109206f13ba39a1af403c7db1e741b", "paper_pdf_bytes": 1392295, "paper_pdf_source": "openreview", "code_url": "https://github.com/ardigen/mol-cycle-gan", "code_repository": "ardigen/mol-cycle-gan", "code_commit": "feb8d7504d0078798ee70d6d5cda3f37b4f7a903", "code_archive": "repos/BklKFo09YX.zip", "code_archive_sha256": "8b90b216b33b7d950f4cb5a179c50b6e04d4bc41283f27e99c2837abc57d8256", "code_archive_bytes": 359906, "code_file_count": 17, "code_extensions": {".py": 10, ".sh": 5, ".ipynb": 2}, "github_disk_usage_kb": 346, "github_languages": {"Jupyter Notebook": 493617, "Python": 38177, "Shell": 1006}, "github_archived": false, "github_pushed_at": "2019-02-06T14:30:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mol-cyclegan-a-generative-model-for-molecular"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MWtm2f1tyG", "year": 2026, "status": "rejected", "title": "IMPACT: Industrial Machine Perception via Acoustic Cognitive Transformer", "authors": ["Changheon Han", "Yuseop Sim", "Hoin Jung", "Jiho Lee", "Hojun Lee", "Yun Seok Kang", "Sucheol Woo", "Garam Kim", "Hyung Wook Park", "Martin Byung-Guk Jun"], "authorids": ["~Changheon_Han2", "~Yuseop_Sim1", "~Hoin_Jung1", "~Jiho_Lee8", "~Hojun_Lee4", "~Yun_Seok_Kang1", "~Sucheol_Woo1", "~Garam_Kim3", "~Hyung_Wook_Park1", "~Martin_Byung-Guk_Jun1"], "authors_source": "OpenReview API", "abstract": "Industrial acoustic signals encode machine state, yet prevailing data-driven approaches are task-specific supervised pipelines that generalize poorly beyond their design conditions. Progress is further limited by the scarcity of large-scale datasets and pretrained models tailored to active shop floor audio. To address this, we introduce DINOS (Diverse INdustrial Operation Sounds), a dataset of 74,149 recordings totaling over 1,093 hours collected from active manufacturing lines across diverse processes and operating regimes. We also provide IMPACT(Industrial Machine Perception via Acoustic Cognitive Transformer), a reference model pretrained on DINOS to standardize evaluation. Our benchmark is structured in four machine-specific steps: (1) baseline discrimination, (2) moderate operational complexity, (3) scalability to unseen equipment, and (4) domain shift and sensor modality adaptation. Across tasks, models pretrained or fine-tuned on DINOS consistently outperform general-purpose audio models, demonstrating the value of domain-specific pretraining for industrial acoustic perception.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "kAEn5W4pw4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19409/Reviewer_US9t"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "The paper proposes DINOS (Diverse INdustrial Operation Sounds), a large-scale dataset for understanding Industrial acoustic signals at a large scale. The paper also trained a self-supervised baseline model on the data IMPACT (Industrial Machine Perception via Acoustic Cognitive Transformer).", "review_text": "The paper proposes DINOS (Diverse INdustrial Operation Sounds), a large-scale dataset for understanding Industrial acoustic signals at a large scale. The paper also trained a self-supervised baseline model on the data IMPACT (Industrial Machine Perception via Acoustic Cognitive Transformer).", "strengths": "The paper is unique and interesting. The paper is written well and contains detailed experiments. The self-supervised model IMPACT, trained on the proposed data, achieves the best performance across the majority of the tasks.", "weaknesses": "The paper has limited novelty. The primary contribution of the paper is the dataset; the IMPACT model is based on a well-known existing self-supervised model, EAT.", "questions": "What is the number of parameters across various models in Table 4? Does the Impact model work better because it is a larger model, or due to pretraining on DINOS?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes DINOS (Diverse INdustrial Operation Sounds), a large-scale dataset for understanding Industrial acoustic signals at a large scale. The paper also trained a self-supervised baseline model on the data IMPACT (Industrial Machine Perception via Acoustic Cognitive Transformer).", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper is unique and interesting. The paper is written well and contains detailed experiments. The self-supervised model IMPACT, trained on the proposed data, achieves the best performance across the majority of the tasks.", "weaknesses": "The paper has limited novelty. The primary contribution of the paper is the dataset; the IMPACT model is based on a well-known existing self-supervised model, EAT.", "questions": "What is the number of parameters across various models in Table 4? Does the Impact model work better because it is a larger model, or due to pretraining on DINOS?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761962711750}, {"id": "PojjjgQnnC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19409/Reviewer_erTN"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "## An open-access dataset of industrial operation sounds\n\n- Proposed DINOS, a dataset with over ~1000 hours of recordings from active manufacturing lines.\n- Proposes IMPACT, a reference baseline models trained on the DINOS dataset.", "review_text": "## An open-access dataset of industrial operation sounds\n\n- Proposed DINOS, a dataset with over ~1000 hours of recordings from active manufacturing lines.\n- Proposes IMPACT, a reference baseline models trained on the DINOS dataset.", "strengths": "- Paper is well written (except some minor grammatical errors. Authors, please recheck for missing spaces and punctuation.)\n- A comprehensive benchmarking setup, with distinct pretraining and downstream benchmarking sets is provided.\n- Limited availability of public, large-scale corpora is a major pain point in manufacturing and floor monitoring, so the dataset could indeed prove invaluable to the community.\n- Evaluation, to the extent done in the paper, is good.", "weaknesses": "- Based on the results alone, it is hard to say how useful the proposed dataset is over the publicly available DCASE2025 Challenge Task 2 dataset for pretraining.", "questions": "1. Is there an overlap between the pretraining set for DINOS and DCASE2025 Challenge Task 2? \n2. Why is your paper titled after the model, and not the dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "## An open-access dataset of industrial operation sounds\n\n- Proposed DINOS, a dataset with over ~1000 hours of recordings from active manufacturing lines.\n- Proposes IMPACT, a reference baseline models trained on the DINOS dataset.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- Paper is well written (except some minor grammatical errors. Authors, please recheck for missing spaces and punctuation.)\n- A comprehensive benchmarking setup, with distinct pretraining and downstream benchmarking sets is provided.\n- Limited availability of public, large-scale corpora is a major pain point in manufacturing and floor monitoring, so the dataset could indeed prove invaluable to the community.\n- Evaluation, to the extent done in the paper, is good.", "weaknesses": "- Based on the results alone, it is hard to say how useful the proposed dataset is over the publicly available DCASE2025 Challenge Task 2 dataset for pretraining.", "questions": "1. Is there an overlap between the pretraining set for DINOS and DCASE2025 Challenge Task 2? \n2. Why is your paper titled after the model, and not the dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761920676617}, {"id": "E23S76vTdl", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19409/Reviewer_66g5"], "rating": 2, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new dataset DINOS, consisting of 74149 acoustic samples collected from active manufacturing lines. The authors then proposed a pretraining method, IMPACT, which is conceptually similar to EAT. The authors evaluated its performance on 27 downstream tasks using DINOS.", "review_text": "This paper introduces a new dataset DINOS, consisting of 74149 acoustic samples collected from active manufacturing lines. The authors then proposed a pretraining method, IMPACT, which is conceptually similar to EAT. The authors evaluated its performance on 27 downstream tasks using DINOS.", "strengths": "1. The collection of DINOS is an earnest effort. DINOS consists of the signals collected from both a microphone and a stethoscope, and covers various types of equipment.\n2. The authors evaluated the performance of various off-the-shelf pretrained models on DINOS.", "weaknesses": "1. The evaluation is critically insufficient and cannot show the superiority of IMPACT. The authors did not apply other pretraining methods (e.g., AudioMAE) on DINOS. They only evaluated the off-the-shelf pretrained models (e.g., a model pretrained using AudioMAE method on other acoustic datasets) on DINOS. Since IMPACT is a pretraining method, if the authors want to show the superiority of IMPACT, they need to **pretrain** IMPACT and other pretraining methods (e.g., AudioMAE) **on the same datasets**.\n2. The proposed pretraining method is conceptually not sufficiently novel. Its similarity to EAT is also acknowledged by the authors.\n3. Therefore, if the majority of the contributions lie in the introduction of DINOS, then this paper might be below the bar of ICLR. It might be more suitable to submit this paper to a venue specialized in industrial sensing or a venue offering dataset tracks.\n4. The presentation could be improved overall.", "questions": "Please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new dataset DINOS, consisting of 74149 acoustic samples collected from active manufacturing lines. The authors then proposed a pretraining method, IMPACT, which is conceptually similar to EAT. The authors evaluated its performance on 27 downstream tasks using DINOS.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The collection of DINOS is an earnest effort. DINOS consists of the signals collected from both a microphone and a stethoscope, and covers various types of equipment.\n2. The authors evaluated the performance of various off-the-shelf pretrained models on DINOS.", "weaknesses": "1. The evaluation is critically insufficient and cannot show the superiority of IMPACT. The authors did not apply other pretraining methods (e.g., AudioMAE) on DINOS. They only evaluated the off-the-shelf pretrained models (e.g., a model pretrained using AudioMAE method on other acoustic datasets) on DINOS. Since IMPACT is a pretraining method, if the authors want to show the superiority of IMPACT, they need to **pretrain** IMPACT and other pretraining methods (e.g., AudioMAE) **on the same datasets**.\n2. The proposed pretraining method is conceptually not sufficiently novel. Its similarity to EAT is also acknowledged by the authors.\n3. Therefore, if the majority of the contributions lie in the introduction of DINOS, then this paper might be below the bar of ICLR. It might be more suitable to submit this paper to a venue specialized in industrial sensing or a venue offering dataset tracks.\n4. The presentation could be improved overall.", "questions": "Please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760589443588}], "openreview_url": "https://openreview.net/forum?id=MWtm2f1tyG", "arxiv_id": "2507.06481", "paper_pdf": "papers/MWtm2f1tyG.pdf", "paper_pdf_sha256": "5485256ffdedfbaa6a9e5ebc82992d8ffc1dc7cb199d064c6c4acb31a67aab5b", "paper_pdf_bytes": 1451593, "paper_pdf_source": "openreview", "code_url": "https://github.com/hanprd/DINOS", "code_repository": "hanprd/DINOS", "code_commit": "2cfbc7786551cd2a734763aeb821243560453721", "code_archive": "repos/MWtm2f1tyG.zip", "code_archive_sha256": "c7125ed0b06d0edd167f736edc5a96e7855b4191d9b659449228c6f647625f40", "code_archive_bytes": 326926, "code_file_count": 15, "code_extensions": {".py": 14, ".sh": 1}, "github_disk_usage_kb": 278, "github_languages": {"Python": 282806, "Shell": 446}, "github_archived": false, "github_pushed_at": "2026-05-17T22:35:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/impact-industrial-machine-perception-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ejVuTFFkl6", "year": 2025, "status": "rejected", "title": "EEG-ImageNet: An Electroencephalogram Dataset and Benchmarks with Image Visual Stimuli of Multi-Granularity Labels", "authors": ["Shuqi Zhu", "Ziyi Ye", "Qingyao Ai", "Yiqun LIU"], "authorids": ["~Shuqi_Zhu1", "~Ziyi_Ye2", "~Qingyao_Ai1", "~Yiqun_LIU1"], "authors_source": "OpenReview API", "abstract": "Exploring how brain activity translates into visual perception offers valuable insights into the biological visual system's representation of the world. Recent advancements have enabled effective image classification and high-quality reconstruction using brain signals obtained through Functional Magnetic Resonance Imaging (fMRI) or magnetoencephalography (MEG). However, the cost and bulkiness of these technologies hinder their practical application. In contrast, Electroencephalography (EEG) presents advantages such as ease of use, affordability, high temporal resolution, and non-invasive operation, yet it remains underutilized in related research due to a shortage of comprehensive datasets. To fill this gap, we introduce EEG-ImageNet, a novel EEG dataset featuring recordings from 16 participants exposed to 4000 images sourced from the ImageNet dataset. This dataset offers five times the number of EEG-image pairs compared to existing benchmarks. EEG-ImageNet includes image stimuli labeled with varying levels of granularity, comprising 40 images with coarse labels and 40 with fine labels. We establish benchmarks for both object classification and image reconstruction based on this dataset. Experiments with several commonly used models show that the best-performing models can achieve object classification with an accuracy around 60% and image reconstruction with two-way identification around 64%. These findings highlight the dataset's potential to enhance EEG-based visual brain-computer interfaces, deepen our understanding of visual perception in biological systems, and suggest promising applications for improving machine vision models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zafmRtlFw1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9655/Reviewer_MjjY"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "In this article, the authors proposed EEG-ImageNet, a dataset with visual stimuli presentation using Rapid Serial Visual Presentation with 16 subjects and a total of 4000 images. The authors present the dataset, classification benchmark and reconstruction studies.", "review_text": "In this article, the authors proposed EEG-ImageNet, a dataset with visual stimuli presentation using Rapid Serial Visual Presentation with 16 subjects and a total of 4000 images. The authors present the dataset, classification benchmark and reconstruction studies.", "strengths": "The paper is well-written, and tackles a fundamental problem in the EEG-to-IMG field: the lack of datasets. \n\nThe authors are precise about the methods used, evaluation within the same subject, and present a 2-step methodology to reconstruct the image given the biosignal.\n\nThe overview of the literature on datasets is interesting, although not complete.", "weaknesses": "Major:\n\n- The authors fail to list Thing's EEG dataset, version 1 and version 2, for some reason that I don't understand. This dataset is currently widely used in the literature and employs a methodology similar to the one the authors present here. A critical comparison is necessary.\n\n- The results with EEGNet models are very poor, raising a flag about the validation of the dataset for deep learning. I understand that reproducing EEGNet is very challenging. However, more neural networks should be chosen to validate this dataset for deep learning. Some suggestions for classification are: ShallowConvNet, DeepConv, EEGConformer, ATCNet, EEGConformer, EEGNex, and perhaps EEGChannelNet.\n\n- The search for parameters seems to favor traditional machine learning models in this study; a better search for parameters should be made. A good reference to understand the importance of hyperparameters can be found Borra, D. et al (2024).\n\n- Even when training one subject by model, other co-founders are present, as discussed in (Kilgallen et al., 2024; Bharadwaj, 2023; Li et al., 2020; Ahmed et al., 2021). Your paper doesn't discuss these issues, making me question your dataset's applicability. How do you ensure that these confounders aren’t impacting your dataset, too?\n\nMinor:\n\n- The tone of paragraphs 212-215 needs to be reduced; the sentences don't hold up compared to other datasets, Things-EEG, for example.\n\n- Some datasets are missing from table 1, please include them:\n\nZheng, X., Wang, L., Chen, K., Lyu, Y., Zhou, J., & Wang, L. (2024). EIT-1M: One Million EEG-Image-Text Pairs for Human Visual-textual Recognition and More. arXiv preprint arXiv:2407.01884.\n\nXu, J., Aristimunha, B., Feucht, M. E., Qian, E., Liu, C., Shahjahan, T., ... & Nestor, A. (2024). Alljoined--A dataset for EEG-to-Image decoding. arXiv preprint arXiv:2404.05553.\n\nGifford, A. T., Dwivedi, K., Roig, G., & Cichy, R. M. (2022). A large and rich EEG dataset for modeling human visual object recognition. NeuroImage, 264, 119754.\n\nKaneshiro, B., Perreau Guimaraes, M., Kim, H. S., Norcia, A. M., & Suppes, P. (2015). A representational similarity analysis of the dynamics of object processing using single-trial EEG classification. Plos one, 10(8), e0135697.\n\n\nReferences \n\nBorra, D., Paissan, F., & Ravanelli, M. (2024). SpeechBrain-MOABB: An open-source Python library for benchmarking deep neural networks applied to EEG signals. Computers in Biology and Medicine, 182, 109097.\n\nKilgallen, J. A., Pearlmutter, B. A., & Siskind, J. M. (2024, June). Learning Exemplar Representations in Single-Trial EEG Category Decoding. In 2024 35th Irish Signals and Systems Conference (ISSC) (pp. 1-6). IEEE.\n\nBharadwaj, H. M., Wilbur, R. B., & Siskind, J. M. (2023). Still an Ineffective Method With Supertrials/ERPs—Comments on “Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features”. IEEE Transactions on Pattern Analysis and Machine Intelligence.\n\nLi, R., Johansen, J. S., Ahmed, H., Ilyevsky, T. V., Wilbur, R. B., Bharadwaj, H. M., & Siskind, J. M. (2020). The perils and pitfalls of block design for EEG classification experiments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1), 316-333.\n\nAhmed, H., Wilbur, R. B., Bharadwaj, H. M., & Siskind, J. M. (2021). Confounds in the data—Comments on “Decoding brain representations by multimodal learning of neural activity and visual features”. IEEE transactions on pattern analysis and machine intelligence, 44(12), 9217-9220.", "questions": "The points raised in the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this article, the authors proposed EEG-ImageNet, a dataset with visual stimuli presentation using Rapid Serial Visual Presentation with 16 subjects and a total of 4000 images. The authors present the dataset, classification benchmark and reconstruction studies.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper is well-written, and tackles a fundamental problem in the EEG-to-IMG field: the lack of datasets. \n\nThe authors are precise about the methods used, evaluation within the same subject, and present a 2-step methodology to reconstruct the image given the biosignal.\n\nThe overview of the literature on datasets is interesting, although not complete.", "weaknesses": "Major:\n\n- The authors fail to list Thing's EEG dataset, version 1 and version 2, for some reason that I don't understand. This dataset is currently widely used in the literature and employs a methodology similar to the one the authors present here. A critical comparison is necessary.\n\n- The results with EEGNet models are very poor, raising a flag about the validation of the dataset for deep learning. I understand that reproducing EEGNet is very challenging. However, more neural networks should be chosen to validate this dataset for deep learning. Some suggestions for classification are: ShallowConvNet, DeepConv, EEGConformer, ATCNet, EEGConformer, EEGNex, and perhaps EEGChannelNet.\n\n- The search for parameters seems to favor traditional machine learning models in this study; a better search for parameters should be made. A good reference to understand the importance of hyperparameters can be found Borra, D. et al (2024).\n\n- Even when training one subject by model, other co-founders are present, as discussed in (Kilgallen et al., 2024; Bharadwaj, 2023; Li et al., 2020; Ahmed et al., 2021). Your paper doesn't discuss these issues, making me question your dataset's applicability. How do you ensure that these confounders aren’t impacting your dataset, too?\n\nMinor:\n\n- The tone of paragraphs 212-215 needs to be reduced; the sentences don't hold up compared to other datasets, Things-EEG, for example.\n\n- Some datasets are missing from table 1, please include them:\n\nZheng, X., Wang, L., Chen, K., Lyu, Y., Zhou, J., & Wang, L. (2024). EIT-1M: One Million EEG-Image-Text Pairs for Human Visual-textual Recognition and More. arXiv preprint arXiv:2407.01884.\n\nXu, J., Aristimunha, B., Feucht, M. E., Qian, E., Liu, C., Shahjahan, T., ... & Nestor, A. (2024). Alljoined--A dataset for EEG-to-Image decoding. arXiv preprint arXiv:2404.05553.\n\nGifford, A. T., Dwivedi, K., Roig, G., & Cichy, R. M. (2022). A large and rich EEG dataset for modeling human visual object recognition. NeuroImage, 264, 119754.\n\nKaneshiro, B., Perreau Guimaraes, M., Kim, H. S., Norcia, A. M., & Suppes, P. (2015). A representational similarity analysis of the dynamics of object processing using single-trial EEG classification. Plos one, 10(8), e0135697.\n\n\nReferences \n\nBorra, D., Paissan, F., & Ravanelli, M. (2024). SpeechBrain-MOABB: An open-source Python library for benchmarking deep neural networks applied to EEG signals. Computers in Biology and Medicine, 182, 109097.\n\nKilgallen, J. A., Pearlmutter, B. A., & Siskind, J. M. (2024, June). Learning Exemplar Representations in Single-Trial EEG Category Decoding. In 2024 35th Irish Signals and Systems Conference (ISSC) (pp. 1-6). IEEE.\n\nBharadwaj, H. M., Wilbur, R. B., & Siskind, J. M. (2023). Still an Ineffective Method With Supertrials/ERPs—Comments on “Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features”. IEEE Transactions on Pattern Analysis and Machine Intelligence.\n\nLi, R., Johansen, J. S., Ahmed, H., Ilyevsky, T. V., Wilbur, R. B., Bharadwaj, H. M., & Siskind, J. M. (2020). The perils and pitfalls of block design for EEG classification experiments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1), 316-333.\n\nAhmed, H., Wilbur, R. B., Bharadwaj, H. M., & Siskind, J. M. (2021). Confounds in the data—Comments on “Decoding brain representations by multimodal learning of neural activity and visual features”. IEEE transactions on pattern analysis and machine intelligence, 44(12), 9217-9220.", "questions": "The points raised in the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730741030810}, {"id": "vz1VoGa3I8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9655/Reviewer_YTBj"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "EEG-based neural visual representation for image classification tasks is very promising and currently does require more extensive EEG datasets with reliable experiment design to support this research area. The authors provided more experiments/databases in this work than the existing work.", "review_text": "EEG-based neural visual representation for image classification tasks is very promising and currently does require more extensive EEG datasets with reliable experiment design to support this research area. The authors provided more experiments/databases in this work than the existing work.", "strengths": "EEG-based neural visual representation shows promise for image retrieval and classification tasks, especially in contexts where current datasets are limited in size and have design limitations. This work contributes by providing a more comprehensive EEG dataset tailored for image retrieval and classification, addressing some of these existing challenges.", "weaknesses": "The authors missed discussing a significant work, The Perils and Pitfalls of Block Design for EEG Classification Experiments (published in IEEE TPAMI, 2020), which highlights a critical weakness in Spampinato’s 2017 study. Specifically, the block design used in EEG experiments can yield artificially high accuracy levels that are unlikely to translate to real-world testing conditions. Unfortunately, the authors appear to have followed a similar block design in their current study.\n\nWhile I appreciate that the authors chose a random seed to randomize the order of categories before the experiment, there remains an issue. Once a category is introduced, participants may anticipate that subsequent images within that block are visually similar, which could introduce bias. To improve the experimental design, I think it is essential fully randomizing the presentation of images across all categories rather than within isolated blocks.", "questions": "See above. Concern the issues from block design for EEG experiments.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "EEG-based neural visual representation for image classification tasks is very promising and currently does require more extensive EEG datasets with reliable experiment design to support this research area. The authors provided more experiments/databases in this work than the existing work.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "EEG-based neural visual representation shows promise for image retrieval and classification tasks, especially in contexts where current datasets are limited in size and have design limitations. This work contributes by providing a more comprehensive EEG dataset tailored for image retrieval and classification, addressing some of these existing challenges.", "weaknesses": "The authors missed discussing a significant work, The Perils and Pitfalls of Block Design for EEG Classification Experiments (published in IEEE TPAMI, 2020), which highlights a critical weakness in Spampinato’s 2017 study. Specifically, the block design used in EEG experiments can yield artificially high accuracy levels that are unlikely to translate to real-world testing conditions. Unfortunately, the authors appear to have followed a similar block design in their current study.\n\nWhile I appreciate that the authors chose a random seed to randomize the order of categories before the experiment, there remains an issue. Once a category is introduced, participants may anticipate that subsequent images within that block are visually similar, which could introduce bias. To improve the experimental design, I think it is essential fully randomizing the presentation of images across all categories rather than within isolated blocks.", "questions": "See above. Concern the issues from block design for EEG experiments.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "n/a", "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730634633147}, {"id": "i3IGXJDr17", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9655/Reviewer_LG8e"], "rating": 3, "soundness": 1, "presentation": 4, "contribution": 2, "confidence": 5, "summary": "This paper described a newly collected EEG dataset named EEG-ImageNet, together with two benchmarks which are the object classification and the image reconstruction tasks based on the EEG signal. The dataset is collected by showing images sequentially to the subjects while collecting the EEG signal. The object classification is set to predict the fine-grained and coarse label of the image. And the image reconstruction task is to reconstruct the image based on the EEG signal.", "review_text": "This paper described a newly collected EEG dataset named EEG-ImageNet, together with two benchmarks which are the object classification and the image reconstruction tasks based on the EEG signal. The dataset is collected by showing images sequentially to the subjects while collecting the EEG signal. The object classification is set to predict the fine-grained and coarse label of the image. And the image reconstruction task is to reconstruct the image based on the EEG signal.", "strengths": "1. The overall problem the authors are trying to solve is very important, and building such a dataset will definitely help the area moving forward, as long as the dataset is not wrongly constructed. \n2. The benchmarks are set properly, with results verified by several models. \n3. The presentation of this paper is generally good.", "weaknesses": "I totally understand the efforts the authors put in to collect the dataset and I really appreciate them. However, I have to say that the overall experiment setting to collect the EEG signal with visual stimuli has some intrinsic flaws due to the temporal effect, and can't be made up by simply adding the section A.5. \n1. I believe the authors are aware of the temporal effect of RSVP. The general pipeline to show the image sequence is very similar to Spampinato et al.(2017), where the images within the same category are presented in a row to the subject. This is so-called **block design** in a followed paper [1] and is proved to be a wrong way to collect EEG dataset. Shortly speaking, the block design will lead to classifying arbitrary temporal artifacts of the data instead of stimulus-related activity. I verified the claim of [1] in both datasets from Spampinato et al. (2017) and under a controlled setting of our own, and I generally believe the claim is right. Also, the claim is verified by the authors on the proposed dataset in figure 6 where the first few samples in the test set is classified more accurately.\n2. While the authors claims that the influence can be mitigated by some techniques including different train-test split or narrower temporal segmentation, it's still a problem. The authors claims a stationary and convergent trend on the last few samples of the test set in their setting, however, it's probably because the temporal influence is convergent to somewhere between this category and the category right after it. To verify this claim, it will be helpful to plot the confusion matrix of the classification benchmark on the dataset. If my guess is correct, we will see a upward trend for the wrongly classified data to be classified to the category right after it. This is also a form of temporal influence that highly biased the model to learn temporal features, and will influence the final result. \n3. Let's set apart all the following things and recognize the last sentence of the paper: \"the dataset can contribute significantly to future research aimed at mitigating temporal effects in the RSVP paradigm\". Then all the proposed two benchmarks are invalid now. Because achieving high results on the two benchmarks encourages a method to actually utilize the temporal effect rather than mitigate the temporal effect, we will have some contradictory directions to optimize on this signal dataset without a clear goal, which will definitely confuse the followed researchers in this direction.\n4. By the way, there are some newer datasets proposed but in a more randomized for your information. THINGS-EEG[2] uses a separate stage for validation after all the training data are collected (which is made unaware for the subject), and SEED-DV[3] separate different stages of stimuli and make sure the images of the same stage are only presented in either training or the testing time.\n\n[1] Li R, Johansen J S, Ahmed H, et al. **The perils and pitfalls of block design for EEG classification experiments**[J]. *IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 43(1): 316-333.*\n\n[2] Grootswagers T, Zhou I, Robinson A K, et al. Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams[J]. Scientific Data, 2022, 9(1): 3.\n\n[3] Liu X, Liu Y, Wang Y, et al. EEG2Video: Towards Decoding Dynamic Visual Perception from EEG Signal[J]. Advances in Neural Information Processing Systems, 2024.", "questions": "If you plot the confusion matrix according to the sequence of categories you present to the subjects, what will it look like?", "flag_for_ethics_review": ["Yes, Responsible research practice (e.g., human subjects, data release)"], "all_content": {"summary": "This paper described a newly collected EEG dataset named EEG-ImageNet, together with two benchmarks which are the object classification and the image reconstruction tasks based on the EEG signal. The dataset is collected by showing images sequentially to the subjects while collecting the EEG signal. The object classification is set to predict the fine-grained and coarse label of the image. And the image reconstruction task is to reconstruct the image based on the EEG signal.", "soundness": 1, "presentation": 4, "contribution": 2, "strengths": "1. The overall problem the authors are trying to solve is very important, and building such a dataset will definitely help the area moving forward, as long as the dataset is not wrongly constructed. \n2. The benchmarks are set properly, with results verified by several models. \n3. The presentation of this paper is generally good.", "weaknesses": "I totally understand the efforts the authors put in to collect the dataset and I really appreciate them. However, I have to say that the overall experiment setting to collect the EEG signal with visual stimuli has some intrinsic flaws due to the temporal effect, and can't be made up by simply adding the section A.5. \n1. I believe the authors are aware of the temporal effect of RSVP. The general pipeline to show the image sequence is very similar to Spampinato et al.(2017), where the images within the same category are presented in a row to the subject. This is so-called **block design** in a followed paper [1] and is proved to be a wrong way to collect EEG dataset. Shortly speaking, the block design will lead to classifying arbitrary temporal artifacts of the data instead of stimulus-related activity. I verified the claim of [1] in both datasets from Spampinato et al. (2017) and under a controlled setting of our own, and I generally believe the claim is right. Also, the claim is verified by the authors on the proposed dataset in figure 6 where the first few samples in the test set is classified more accurately.\n2. While the authors claims that the influence can be mitigated by some techniques including different train-test split or narrower temporal segmentation, it's still a problem. The authors claims a stationary and convergent trend on the last few samples of the test set in their setting, however, it's probably because the temporal influence is convergent to somewhere between this category and the category right after it. To verify this claim, it will be helpful to plot the confusion matrix of the classification benchmark on the dataset. If my guess is correct, we will see a upward trend for the wrongly classified data to be classified to the category right after it. This is also a form of temporal influence that highly biased the model to learn temporal features, and will influence the final result. \n3. Let's set apart all the following things and recognize the last sentence of the paper: \"the dataset can contribute significantly to future research aimed at mitigating temporal effects in the RSVP paradigm\". Then all the proposed two benchmarks are invalid now. Because achieving high results on the two benchmarks encourages a method to actually utilize the temporal effect rather than mitigate the temporal effect, we will have some contradictory directions to optimize on this signal dataset without a clear goal, which will definitely confuse the followed researchers in this direction.\n4. By the way, there are some newer datasets proposed but in a more randomized for your information. THINGS-EEG[2] uses a separate stage for validation after all the training data are collected (which is made unaware for the subject), and SEED-DV[3] separate different stages of stimuli and make sure the images of the same stage are only presented in either training or the testing time.\n\n[1] Li R, Johansen J S, Ahmed H, et al. **The perils and pitfalls of block design for EEG classification experiments**[J]. *IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 43(1): 316-333.*\n\n[2] Grootswagers T, Zhou I, Robinson A K, et al. Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams[J]. Scientific Data, 2022, 9(1): 3.\n\n[3] Liu X, Liu Y, Wang Y, et al. EEG2Video: Towards Decoding Dynamic Visual Perception from EEG Signal[J]. Advances in Neural Information Processing Systems, 2024.", "questions": "If you plot the confusion matrix according to the sequence of categories you present to the subjects, what will it look like?", "flag_for_ethics_review": ["Yes, Responsible research practice (e.g., human subjects, data release)"], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730463057522}, {"id": "RwE4cLCesJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9655/Reviewer_bg7F"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "This article introduces a new EEG dataset recorded through an RSVP (Rapid Serial Visual Presentation) experiment paradigm, where EEG data was collected from 16 participants as they viewed images from the ImageNet dataset. Each participant viewed 4,000 images across 80 categories. The dataset supports multi-granularity analysis, with 40 categories designated for coarse-grained analysis and 40 for fine-grained analysis. The article provides a detailed description of the dataset construction process, including participant selection, stimulus dataset, and acquisition procedure. Additionally, the dataset was used to establish benchmarks for two common image tasks: classification and image reconstruction. Experimental results validate the feasibility of this dataset, providing a foundation for subsequent research.", "review_text": "This article introduces a new EEG dataset recorded through an RSVP (Rapid Serial Visual Presentation) experiment paradigm, where EEG data was collected from 16 participants as they viewed images from the ImageNet dataset. Each participant viewed 4,000 images across 80 categories. The dataset supports multi-granularity analysis, with 40 categories designated for coarse-grained analysis and 40 for fine-grained analysis. The article provides a detailed description of the dataset construction process, including participant selection, stimulus dataset, and acquisition procedure. Additionally, the dataset was used to establish benchmarks for two common image tasks: classification and image reconstruction. Experimental results validate the feasibility of this dataset, providing a foundation for subsequent research.", "strengths": "1.This paper makes use of the advantages of EEG and proposes EEG-ImageNet dataset to provide reference for related tasks.\n\n2.The authors provided a detailed description of the experimental acquisition process and preprocessing.", "weaknesses": "1.The innovation of this dataset, compared to others, lies in the introduction of multi-granularity labels; however, subsequent experiments did not include a comparison of performance with other models in handling multi-granularity labels.\n\n2.In the benchmarks, the authors chose relatively traditional methods and did not use current SOTA models to validate the dataset.\n\n3.As a new EEG-Image dataset, it has not been compared with previous datasets of the same type, such as THINGS-EEG.\n\n4.In the image classification validation experiments, only Acc was used as the evaluation metric, limiting the effectiveness of the analysis. Additionally, in the image reconstruction task, only four evaluation metrics were employed.\n\nREFERENCES:\n\n[1]Bharadwaj, H. M., Wilbur, R. B. & Siskind, J. M. Still an ineffective method with supertrials/ERPs—comments on “decoding brain representations by multimodal learning of neural activity and visual features”. IEEE Trans. Pattern Anal. Mach. Intell. 45, 14052–14054 (2023).\n\n[2]Grootswagers, T., Zhou, I., Robinson, A. K., Hebart, M. N., and Carlson, T. A. (2022). Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams. Scientific Data, 9(1):3.", "questions": "1.In Table 2, why does EEGNet perform relatively poorly? On line 378，the authors mentioned that \"* indicates the use of time-domain features.\" What is the significance of this approach?\n\n2.On line 256, it mentions, “The latter 40 categories are designed as a fine-grained task, divided into 5 groups with 8 categories each.” Could this approach lead to a data imbalance issue, making it difficult to directly compare the performance at different granularities in the subsequent image classification task?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article introduces a new EEG dataset recorded through an RSVP (Rapid Serial Visual Presentation) experiment paradigm, where EEG data was collected from 16 participants as they viewed images from the ImageNet dataset. Each participant viewed 4,000 images across 80 categories. The dataset supports multi-granularity analysis, with 40 categories designated for coarse-grained analysis and 40 for fine-grained analysis. The article provides a detailed description of the dataset construction process, including participant selection, stimulus dataset, and acquisition procedure. Additionally, the dataset was used to establish benchmarks for two common image tasks: classification and image reconstruction. Experimental results validate the feasibility of this dataset, providing a foundation for subsequent research.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1.This paper makes use of the advantages of EEG and proposes EEG-ImageNet dataset to provide reference for related tasks.\n\n2.The authors provided a detailed description of the experimental acquisition process and preprocessing.", "weaknesses": "1.The innovation of this dataset, compared to others, lies in the introduction of multi-granularity labels; however, subsequent experiments did not include a comparison of performance with other models in handling multi-granularity labels.\n\n2.In the benchmarks, the authors chose relatively traditional methods and did not use current SOTA models to validate the dataset.\n\n3.As a new EEG-Image dataset, it has not been compared with previous datasets of the same type, such as THINGS-EEG.\n\n4.In the image classification validation experiments, only Acc was used as the evaluation metric, limiting the effectiveness of the analysis. Additionally, in the image reconstruction task, only four evaluation metrics were employed.\n\nREFERENCES:\n\n[1]Bharadwaj, H. M., Wilbur, R. B. & Siskind, J. M. Still an ineffective method with supertrials/ERPs—comments on “decoding brain representations by multimodal learning of neural activity and visual features”. IEEE Trans. Pattern Anal. Mach. Intell. 45, 14052–14054 (2023).\n\n[2]Grootswagers, T., Zhou, I., Robinson, A. K., Hebart, M. N., and Carlson, T. A. (2022). Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams. Scientific Data, 9(1):3.", "questions": "1.In Table 2, why does EEGNet perform relatively poorly? On line 378，the authors mentioned that \"* indicates the use of time-domain features.\" What is the significance of this approach?\n\n2.On line 256, it mentions, “The latter 40 categories are designed as a fine-grained task, divided into 5 groups with 8 categories each.” Could this approach lead to a data imbalance issue, making it difficult to directly compare the performance at different granularities in the subsequent image classification task?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729930466204}], "openreview_url": "https://openreview.net/forum?id=ejVuTFFkl6", "arxiv_id": "2406.07151", "paper_pdf": "papers/ejVuTFFkl6.pdf", "paper_pdf_sha256": "11fcebeeb116af283350768ee4210d2b0e9702089b32ff3a6b27c6c8c0f719ad", "paper_pdf_bytes": 9831475, "paper_pdf_source": "openreview", "code_url": "https://github.com/Promise-Z5Q2SQ/EEG-ImageNet-Dataset", "code_repository": "Promise-Z5Q2SQ/EEG-ImageNet-Dataset", "code_commit": "7c78157100d3737690d73f206f4385e72613290d", "code_archive": "repos/ejVuTFFkl6.zip", "code_archive_sha256": "ad670acc5e055a42b19287596ee17ded6c7c3cfa80a712bdbbb104a9cfbf8a26", "code_archive_bytes": 26518, "code_file_count": 16, "code_extensions": {".py": 13, ".sh": 3}, "github_disk_usage_kb": 44, "github_languages": {"Python": 40544, "Shell": 2053}, "github_archived": false, "github_pushed_at": "2025-05-15T13:26:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/eeg-imagenet-an-electroencephalogram-dataset"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mBzsKsrXf9", "year": 2024, "status": "rejected", "title": "ArtWhisperer: A Dataset for Characterizing Human-AI Interactions in Artistic Creations", "authors": ["Kailas Vodrahalli", "James Zou"], "authorids": ["~Kailas_Vodrahalli1", "~James_Zou1"], "authors_source": "OpenReview API", "abstract": "As generative AI becomes more prevalent, it is important to study how human users interact with such models. In this work, we investigate how people use text-to-image models to generate desired target images. To study this interaction, we created ArtWhisperer, an online game where users are given a target image and are tasked with iteratively finding a prompt that creates a similar-looking image as the target. Through this game, we recorded over 50,000 human-AI interactions; each interaction corresponds to one text prompt created by a user and the corresponding generated image. The majority of these are repeated interactions where a user iterates to find the best prompt for their target image, making this a unique sequential dataset for studying human-AI collaborations. In an initial analysis of this dataset, we identify several characteristics of prompt interactions and user strategies. People submit diverse prompts and are able to discover a variety of text descriptions that generate similar images. Interestingly, prompt diversity does not decrease as users find better prompts. We further propose a new metric to quantify the steerability of AI using our dataset. We define steerability as the expected number of interactions required to adequately complete a task. We estimate this value by fitting a Markov chain for each target task and calculating the expected time to reach an adequate score in the Markov chain. We  quantify and compare AI steerability across different types of target images and two different models, finding that images of cities and natural world images are more steerable than artistic and fantasy images. These findings provide insights into human-AI interaction behavior, present a concrete method of assessing AI steerability, and demonstrate the general utility of the ArtWhisperer dataset.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NFEW2j5897", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7874/Reviewer_1NDi"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper investigates human-AI interactions in the context of text to image generation models. The paper presents a gamified experimental setup called “ArtWhisperer”, where a subject is provided with a target image. The subject’s goal is to iteratively prompt the model to successfully generate an image similar to the given target image. \n\nThe paper presents analyses of the experimental data, and defines a metric called “steerability”, which captures the human subject’s difficulty of generating an image via prompting.", "review_text": "The paper investigates human-AI interactions in the context of text to image generation models. The paper presents a gamified experimental setup called “ArtWhisperer”, where a subject is provided with a target image. The subject’s goal is to iteratively prompt the model to successfully generate an image similar to the given target image. \n\nThe paper presents analyses of the experimental data, and defines a metric called “steerability”, which captures the human subject’s difficulty of generating an image via prompting.", "strengths": "1. The experimental data of goal-driven prompt-interactions appears to be unique, and of potential interest to the generative AI community.\n\n2. The overall experimental setup seems reasonable — the generated images are evaluated against *generated* images; not real images. The controlled setting, is especially good, and provides important additional insights about the overall task.\n\n3. The paper clearly discusses the major limitation of the experimental setup — relatively low (although positive) correlation between human notions of similarity and the automatic score. This is first identified via a different experimental setup, and the consequence of the mismatch is quantified via a steerability score on a smaller set of images based on human-provided similarity scores. The difference in steerability scores on this smaller subset is not too large. This is promising, and improves confidence about the overall validity of the conclusions based on experiments with automatic scores.\n\n4. The paper analyses the data to describe findings that may be of interest to researchers and practitioners working on human-AI interaction applications. Specifically, the study of steerability is quite relevant for creative AI applications.", "weaknesses": "1. Prompt engineering is now a topic of interest for people who wish to efficiently generate usable content from conditional generative models. My hypothesis is that, “Steerability”, i.e., the ease of generating the target image, depends on the amount of experience of the user. It would be great if the authors discuss how they might be able to decouple the influence of returning users who potentially have the ability to efficiently “steer” the generative model via better prompts.\n\n2. Apart from the prompting experience, a person’s knowledge of the contents of the target image could also play a role in steerability. Especially in cases of a famous person, famous landmark, famous city, etc. There appear to be some person-specific confounding factors that affect steerability, which the paper doesn’t seem to address. It appears that the rationale for this is that, with a large enough population of users, such person-specific effects might be negligible. However, it would be nice if these confounding factors were explicitly identified and discussed.\n\n3. “Image specificity” (Jas et al., CVPR 2014) measures the extent to which human captions that describe an image, varies across people. The observation that target images containing specific landmarks are more steerable than abstract images, appears to allude to a similar concept. Of course, the idea of steerability involves additional complexity of the generative model, and the human’s understanding of prompting. However, it seems possible that images with high “specificity” might lend themselves to be more steerable. It would be great to discuss this further.\n\n4. The mismatch between human notions of similarity and the automatic metric makes the experimental setup suboptimal, and data a little more complicated to draw conclusions from. Please see “Questions” for specific concerns regarding the mismatch.\n\nReferences:\n(Jas et al., CVPR 2014) Jas, Mainak, and Devi Parikh. \"Image specificity.\" *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*. 2015", "questions": "1. Were there comments / feedback from users that elicited insight regarding their strategy? Specifically regarding the incongruity between human similarity notions and automatically computed similarity score? E.g., did certain persons try to maximise the automatic score while others maximised their own perception of similarity?\n\n2. Was there an incentive for high score? E.g., additional reward? Do we know that people were taking the task “seriously”, and not just experimenting / exploring?\n\n3. What is the score distribution between each generated image (and the target image) — that are generated from the same prompt? Is there any insight into the variance in matching score across different random seeds of the model (keeping the prompt constant)\n\n4. “The intuition here is that t*1_i is more representative of the types of images we may expect given the fixed prompt, p*_i .” — the intuition here is unfortunately not clear to this reader. Could the authors please elaborate? Thanks!", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates human-AI interactions in the context of text to image generation models. The paper presents a gamified experimental setup called “ArtWhisperer”, where a subject is provided with a target image. The subject’s goal is to iteratively prompt the model to successfully generate an image similar to the given target image. \n\nThe paper presents analyses of the experimental data, and defines a metric called “steerability”, which captures the human subject’s difficulty of generating an image via prompting.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The experimental data of goal-driven prompt-interactions appears to be unique, and of potential interest to the generative AI community.\n\n2. The overall experimental setup seems reasonable — the generated images are evaluated against *generated* images; not real images. The controlled setting, is especially good, and provides important additional insights about the overall task.\n\n3. The paper clearly discusses the major limitation of the experimental setup — relatively low (although positive) correlation between human notions of similarity and the automatic score. This is first identified via a different experimental setup, and the consequence of the mismatch is quantified via a steerability score on a smaller set of images based on human-provided similarity scores. The difference in steerability scores on this smaller subset is not too large. This is promising, and improves confidence about the overall validity of the conclusions based on experiments with automatic scores.\n\n4. The paper analyses the data to describe findings that may be of interest to researchers and practitioners working on human-AI interaction applications. Specifically, the study of steerability is quite relevant for creative AI applications.", "weaknesses": "1. Prompt engineering is now a topic of interest for people who wish to efficiently generate usable content from conditional generative models. My hypothesis is that, “Steerability”, i.e., the ease of generating the target image, depends on the amount of experience of the user. It would be great if the authors discuss how they might be able to decouple the influence of returning users who potentially have the ability to efficiently “steer” the generative model via better prompts.\n\n2. Apart from the prompting experience, a person’s knowledge of the contents of the target image could also play a role in steerability. Especially in cases of a famous person, famous landmark, famous city, etc. There appear to be some person-specific confounding factors that affect steerability, which the paper doesn’t seem to address. It appears that the rationale for this is that, with a large enough population of users, such person-specific effects might be negligible. However, it would be nice if these confounding factors were explicitly identified and discussed.\n\n3. “Image specificity” (Jas et al., CVPR 2014) measures the extent to which human captions that describe an image, varies across people. The observation that target images containing specific landmarks are more steerable than abstract images, appears to allude to a similar concept. Of course, the idea of steerability involves additional complexity of the generative model, and the human’s understanding of prompting. However, it seems possible that images with high “specificity” might lend themselves to be more steerable. It would be great to discuss this further.\n\n4. The mismatch between human notions of similarity and the automatic metric makes the experimental setup suboptimal, and data a little more complicated to draw conclusions from. Please see “Questions” for specific concerns regarding the mismatch.\n\nReferences:\n(Jas et al., CVPR 2014) Jas, Mainak, and Devi Parikh. \"Image specificity.\" *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*. 2015", "questions": "1. Were there comments / feedback from users that elicited insight regarding their strategy? Specifically regarding the incongruity between human similarity notions and automatically computed similarity score? E.g., did certain persons try to maximise the automatic score while others maximised their own perception of similarity?\n\n2. Was there an incentive for high score? E.g., additional reward? Do we know that people were taking the task “seriously”, and not just experimenting / exploring?\n\n3. What is the score distribution between each generated image (and the target image) — that are generated from the same prompt? Is there any insight into the variance in matching score across different random seeds of the model (keeping the prompt constant)\n\n4. “The intuition here is that t*1_i is more representative of the types of images we may expect given the fixed prompt, p*_i .” — the intuition here is unfortunately not clear to this reader. Could the authors please elaborate? Thanks!", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699221567516}, {"id": "ElBQ7es696", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7874/Reviewer_oZkf"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors introduce a dataset for understanding prompt tailoring in image generation. To collect the data, they introduced a game collecting 51k interactions across 2k papers, with a smaller group of paid users. The game is driven by two datasources AI Pompts and Wikipedia which they align using CLIP embeddings to select optimal plausible query images. From the game, they identify Markov chain approach to the steerability of user prompts. The authors envisage the dataset can be used to improve the outputs of image generation methods but also as a synthetic test setting using tailored NLP methods to test the response.", "review_text": "The authors introduce a dataset for understanding prompt tailoring in image generation. To collect the data, they introduced a game collecting 51k interactions across 2k papers, with a smaller group of paid users. The game is driven by two datasources AI Pompts and Wikipedia which they align using CLIP embeddings to select optimal plausible query images. From the game, they identify Markov chain approach to the steerability of user prompts. The authors envisage the dataset can be used to improve the outputs of image generation methods but also as a synthetic test setting using tailored NLP methods to test the response.", "strengths": "- Prompt engineering is a difficult problem to capture which the authors have proposed a method to achieve future analytical comparisons of methods.\n- The dataset seems to be sufficient to capture user insights as the authors identify steerability is achievable.\n- The use of multiple data sources to drive the method is good, especially as real world image captions do not contain the content description. However, the authors fail to quantify the performance, and they do not elaborate in the evaluation of the two sources.", "weaknesses": "- It would be better to be clearer about the controlled collection earlier, simply paid and unpaid would avoid assumptions of how this was performed.\n- The authors are missing a comparison of how true the distribution of generated images is to the original images as the Wikipedia captions in general don't describe the content so highly likely they do not match the original image. A quantitative evaluation of this is important otherwise it makes the choice of Wikipedia not relevant and does not increase validity over using any source.\n- Authors should avoid re-using mathematical notation x_i changes through section 2. Although in general x is re-used in changing context throughout the paper.\n- Lack of discussion on participant characteristics how was gender, race & ethnicity balanced? \n- Steerability contribution is well received, however, this would have more validity if was tested on a different model to confirm the insights, two versions of StableDiffusion provide limited insight.,", "questions": "- How was bias addressed during data collection?\n- How do the results on steerability differ between the two input datasets?\n- Does the influence of unpaid/paid effect the outcomes?", "flag_for_ethics_review": ["Yes, Discrimination / bias / fairness concerns"], "all_content": {"summary": "The authors introduce a dataset for understanding prompt tailoring in image generation. To collect the data, they introduced a game collecting 51k interactions across 2k papers, with a smaller group of paid users. The game is driven by two datasources AI Pompts and Wikipedia which they align using CLIP embeddings to select optimal plausible query images. From the game, they identify Markov chain approach to the steerability of user prompts. The authors envisage the dataset can be used to improve the outputs of image generation methods but also as a synthetic test setting using tailored NLP methods to test the response.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- Prompt engineering is a difficult problem to capture which the authors have proposed a method to achieve future analytical comparisons of methods.\n- The dataset seems to be sufficient to capture user insights as the authors identify steerability is achievable.\n- The use of multiple data sources to drive the method is good, especially as real world image captions do not contain the content description. However, the authors fail to quantify the performance, and they do not elaborate in the evaluation of the two sources.", "weaknesses": "- It would be better to be clearer about the controlled collection earlier, simply paid and unpaid would avoid assumptions of how this was performed.\n- The authors are missing a comparison of how true the distribution of generated images is to the original images as the Wikipedia captions in general don't describe the content so highly likely they do not match the original image. A quantitative evaluation of this is important otherwise it makes the choice of Wikipedia not relevant and does not increase validity over using any source.\n- Authors should avoid re-using mathematical notation x_i changes through section 2. Although in general x is re-used in changing context throughout the paper.\n- Lack of discussion on participant characteristics how was gender, race & ethnicity balanced? \n- Steerability contribution is well received, however, this would have more validity if was tested on a different model to confirm the insights, two versions of StableDiffusion provide limited insight.,", "questions": "- How was bias addressed during data collection?\n- How do the results on steerability differ between the two input datasets?\n- Does the influence of unpaid/paid effect the outcomes?", "flag_for_ethics_review": ["Yes, Discrimination / bias / fairness concerns"], "details_of_ethics_concerns": "The potential dataset has the risk of large bias. The authors should justify how bias is addressed in the data collection to avoid future issues around this topic being introduced. The minimum should be clearly stating the diversity of participants in data collection. However, if this reveals large imbalance the data needs to be re-balanced or studied to understand if the concluding steerability approach is bias.", "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698836864288}, {"id": "TFAloltJIv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7874/Reviewer_Ww77"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This dataset studies how people interact with text-to-image models to generate desired target images. Particularly, during the data collection, the job of human is to iteratively re-fine the text prompt (only) to create a similar-looking image (to a given reference). Such a procedure traces the trajectory of human in making creative art using text-to-image model, and presents a unique sequential dataset for studying human-AI collaboration.\n\nThen this paper perform analysis on the collected dataset, and find that user is able to discover a variety of text description that generates similar images, with the trajectory of user queries of similar diversity along the search. Then they propose a metric to quantify the steerability of AI, which is the expected number of interaction required to adequately complete a task.", "review_text": "This dataset studies how people interact with text-to-image models to generate desired target images. Particularly, during the data collection, the job of human is to iteratively re-fine the text prompt (only) to create a similar-looking image (to a given reference). Such a procedure traces the trajectory of human in making creative art using text-to-image model, and presents a unique sequential dataset for studying human-AI collaboration.\n\nThen this paper perform analysis on the collected dataset, and find that user is able to discover a variety of text description that generates similar images, with the trajectory of user queries of similar diversity along the search. Then they propose a metric to quantify the steerability of AI, which is the expected number of interaction required to adequately complete a task.", "strengths": "- The paper is well presented, the motivation, data collection, experiments are clearly written. There is abundant details to understand the quality of the data, the distribution of user trajectory, the domain of data, etc.", "weaknesses": "- It's unclear to me on how we could utilize the study done in this paper to improve research in text-to-image generation modeling, or Human AI interaction research. I saw some direction from the conclusion but those can be drawn without studying the research presented in this paper.", "questions": "- In the section about AI-generated images, it seems like from this seed selection procedure, you are only choosing the target image based on how easy it could be re-created using the prompt (and different random seeds). The criteria is not encouraging the aesthetic of the image, nor alignment between image and target prompts? \n- What is IRB approval?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This dataset studies how people interact with text-to-image models to generate desired target images. Particularly, during the data collection, the job of human is to iteratively re-fine the text prompt (only) to create a similar-looking image (to a given reference). Such a procedure traces the trajectory of human in making creative art using text-to-image model, and presents a unique sequential dataset for studying human-AI collaboration.\n\nThen this paper perform analysis on the collected dataset, and find that user is able to discover a variety of text description that generates similar images, with the trajectory of user queries of similar diversity along the search. Then they propose a metric to quantify the steerability of AI, which is the expected number of interaction required to adequately complete a task.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- The paper is well presented, the motivation, data collection, experiments are clearly written. There is abundant details to understand the quality of the data, the distribution of user trajectory, the domain of data, etc.", "weaknesses": "- It's unclear to me on how we could utilize the study done in this paper to improve research in text-to-image generation modeling, or Human AI interaction research. I saw some direction from the conclusion but those can be drawn without studying the research presented in this paper.", "questions": "- In the section about AI-generated images, it seems like from this seed selection procedure, you are only choosing the target image based on how easy it could be re-created using the prompt (and different random seeds). The criteria is not encouraging the aesthetic of the image, nor alignment between image and target prompts? \n- What is IRB approval?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698566108043}, {"id": "w6u35tx1Gn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7874/Reviewer_kk5o"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper aims to study how people write prompts for generating a target image with text-to-image models. The authors have collected the ArtWhisperer dataset that includes the different iterations of prompts a user tries out in generating the images, and from this dataset they have observed that different, diverse prompts can be used to generate an image. They also propose a metric for quantifying the steerability of the target image generation task.", "review_text": "This paper aims to study how people write prompts for generating a target image with text-to-image models. The authors have collected the ArtWhisperer dataset that includes the different iterations of prompts a user tries out in generating the images, and from this dataset they have observed that different, diverse prompts can be used to generate an image. They also propose a metric for quantifying the steerability of the target image generation task.", "strengths": "* The paper addresses an interesting problem that has not been really investigated by other researchers in depth.\n* The dataset will be made available to the public. To the best of my knowledge, no comparable dataset exists.\n* The authors have followed responsible research practices (getting IRB approval, handling user data/privacy properly).\n* The work is well written and easy to follow.\n* The discussed future use cases including synthetic prompting and fully automated evaluation of t2i models would be relevant to many real-world applications.", "weaknesses": "* The work does not investigate the behavior of users with malevolent intent (the authors acknowledge this), as their actions may not align with those of good actors. The users are assumed to be the latter (or at least, homogeneous in that respect), and the observations made in this paper are based upon that assumption.\n* I would be interested to see more discussion/applications of how to take these noted observations of user behavior, steerability, and prompts and apply them to an existing problem. These insights could have possible implications for model training/evaluation in the future.\n* There may have been returning users on a day-to-day basis, so it’s feasible that a user became more experienced with prompting over time, which would thus affect steerability. The authors however did not collect this type of data, so it’s not possible to observe this.", "questions": "1. Do the authors have any sense for how user behavior may differ with other t2i models? For example, does the quality of the model (i.e., its ability to generate high-quality imagery) affect the results?\n2. Did the authors consider enabling users to adjust the model parameters (e.g. the seed)? This could potentially enable the user to generate an image that is even closer to the target, and could also be useful in future applications highlighted in the Discussion section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to study how people write prompts for generating a target image with text-to-image models. The authors have collected the ArtWhisperer dataset that includes the different iterations of prompts a user tries out in generating the images, and from this dataset they have observed that different, diverse prompts can be used to generate an image. They also propose a metric for quantifying the steerability of the target image generation task.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "* The paper addresses an interesting problem that has not been really investigated by other researchers in depth.\n* The dataset will be made available to the public. To the best of my knowledge, no comparable dataset exists.\n* The authors have followed responsible research practices (getting IRB approval, handling user data/privacy properly).\n* The work is well written and easy to follow.\n* The discussed future use cases including synthetic prompting and fully automated evaluation of t2i models would be relevant to many real-world applications.", "weaknesses": "* The work does not investigate the behavior of users with malevolent intent (the authors acknowledge this), as their actions may not align with those of good actors. The users are assumed to be the latter (or at least, homogeneous in that respect), and the observations made in this paper are based upon that assumption.\n* I would be interested to see more discussion/applications of how to take these noted observations of user behavior, steerability, and prompts and apply them to an existing problem. These insights could have possible implications for model training/evaluation in the future.\n* There may have been returning users on a day-to-day basis, so it’s feasible that a user became more experienced with prompting over time, which would thus affect steerability. The authors however did not collect this type of data, so it’s not possible to observe this.", "questions": "1. Do the authors have any sense for how user behavior may differ with other t2i models? For example, does the quality of the model (i.e., its ability to generate high-quality imagery) affect the results?\n2. Did the authors consider enabling users to adjust the model parameters (e.g. the seed)? This could potentially enable the user to generate an image that is even closer to the target, and could also be useful in future applications highlighted in the Discussion section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698544049813}], "openreview_url": "https://openreview.net/forum?id=mBzsKsrXf9", "arxiv_id": "2306.08141", "paper_pdf": "papers/mBzsKsrXf9.pdf", "paper_pdf_sha256": "d7217c23549ff7f50da0f9f3d07bda846e233fbdd65dd4987491fb6c5fdcdf82", "paper_pdf_bytes": 3923407, "paper_pdf_source": "openreview", "code_url": "https://github.com/kailas-v/ArtWhisperer", "code_repository": "kailas-v/ArtWhisperer", "code_commit": "246b8361cd217c7b679e95b139dabdada7cdda16", "code_archive": "repos/mBzsKsrXf9.zip", "code_archive_sha256": "419847fd432eda559548f21cfbd0caee6ac9769d8091c1b3c077df5bd7dd98d8", "code_archive_bytes": 32075, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 35, "github_languages": {"Python": 105880}, "github_archived": false, "github_pushed_at": "2023-08-29T12:48:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/artwhisperer-a-dataset-for-characterizing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iMy1hOrqiVE", "year": 2023, "status": "rejected", "title": "Exclusive Supermask Subnetwork Training for Continual Learning", "authors": ["Prateek Yadav", "Mohit Bansal"], "authorids": ["~Prateek_Yadav1", "~Mohit_Bansal2"], "authors_source": "OpenReview API", "abstract": "Continual Learning (CL) methods mainly focus on avoiding catastrophic forgetting and learning representations that are transferable to new tasks. Recently, Wortsman et al. (2020) proposed a CL method, SupSup, which uses a randomly initialized, fixed base network (model) and finds a supermask for each new task that selectively keeps or removes each weight to produce a subnetwork. They prevent forgetting as the network weights are not being updated. Although there is no forgetting, the performance of supermask is sub-optimal because fixed weights restrict its representational power. Furthermore, there is no accumulation or transfer of knowledge inside the model when new tasks are learned. Hence, we propose ExSSNeT (Exclusive Supermask SubNEtwork Training), that performs exclusive and non-overlapping subnetwork weight training. This avoids conflicting updates to the shared weights by subsequent tasks to improve performance while still preventing forgetting. Furthermore, we propose a novel KNN-based Knowledge Transfer (KKT) module that dynamically initializes a new task's mask based on previous tasks for improving knowledge transfer. We demonstrate that ExSSNeT outperforms SupSup and other strong previous methods on both text classification and vision tasks while preventing forgetting. Moreover, ExSSNeT is particularly advantageous for sparse masks that activate 2-10% of the model parameters, resulting in an average improvement of 8.3% over SupSup. Additionally, ExSSNeT scales to a large number of tasks (100), and our KKT module helps to learn new tasks faster while improving the overall performance.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "qcGHRDOctQ8", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3457/Reviewer_XDRc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the author focused on developing the solution for mitigating the catastrophic forgetting problem in the continuous learning problem. The proposed framework is mainly composed of two components, the first component is exclusive and non-overlapping subnetwork weight training, The second component is the KNN knowledge transfer module. The authors have evaluated the proposed pipeline on several benchmarks and demonstrate the improvement over listed benchmarks. ", "review_text": "My rating is between the borderline and weakly reject since the evaluation data is relative small. Not sure if it can be generalized to more complicated task with big data. ", "strengths": "Strength:\n1.\tThe framework has adopted and expanded previous works’ solution and achieved better performance than listed benchmarks.\n2.\tThe idea is straightforward and easy to reproduce.\n\nWeakness:\n1.\tThe listed image benchmark is relatively simple, hard to justify the effectiveness of the proposed solution for dealing with big dataset. \n2.\tOne of the most important steps in this pipeline is the supermask learning. It needs to be learned when adding new task. The author has listed the operations after 3 tasks in figure 1. What will be the parameters complexity after N tasks? Will that be a problem in the long run of continuous learning. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this work, the author focused on developing the solution for mitigating the catastrophic forgetting problem in the continuous learning problem. The proposed framework is mainly composed of two components, the first component is exclusive and non-overlapping subnetwork weight training, The second component is the KNN knowledge transfer module. The authors have evaluated the proposed pipeline on several benchmarks and demonstrate the improvement over listed benchmarks. ", "strength_and_weaknesses": "Strength:\n1.\tThe framework has adopted and expanded previous works’ solution and achieved better performance than listed benchmarks.\n2.\tThe idea is straightforward and easy to reproduce.\n\nWeakness:\n1.\tThe listed image benchmark is relatively simple, hard to justify the effectiveness of the proposed solution for dealing with big dataset. \n2.\tOne of the most important steps in this pipeline is the supermask learning. It needs to be learned when adding new task. The author has listed the operations after 3 tasks in figure 1. What will be the parameters complexity after N tasks? Will that be a problem in the long run of continuous learning. \n", "clarity,_quality,_novelty_and_reproducibility": "The writing quality is ok. I think the results can be reproducible. ", "summary_of_the_review": "My rating is between the borderline and weakly reject since the evaluation data is relative small. Not sure if it can be generalized to more complicated task with big data. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666937022015}, {"id": "KigaYTrwZAV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3457/Reviewer_7z7s"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a mask-based approach to obtain subnetworks for specific tasks in continual learning (CL). It is an improvement of the previous method SupSup, which finds fixed supermask for each task to alleviate forgetting. This paper proposes to perform exclusive and nonoverlapping subnetwork training to avoid conflicting updating on the shared parameters. A specific KNN based method is proposed to enhance parameter sharing. The experiments show that the proposed method can produce better performance than other methods. ", "review_text": "As discussed above, the work is well motivated and justified, and the paper is well-written. \nThe paper provides enough novel insights and observations and some empirical experimental results and analyses on the proposed method. The significance is influenced by its relationship with SupSup and that it is restricted to the task-incremental setting, where task identifiers are required in both training and testing. \n\n---\nI appreciate the authors' response, which addressed part of my concerns, especially on the KKT module. I am still concerned by the novelty (mainly about the relationship with SupSup) and technical contributions. After further discussions, I am further convinced that the experiment part also needs to be improved, especially in comparison with other parameter isolation-based methods, instead of only SupSup.  \nThe work may be further improved by fixing the issues and conducting some more work, such as some attempts to extend the method to other settings beyond the task-incremental setting.\n", "strengths": "Strength\n- The work is well motivated. It starts from the limitation of the SupSup method and designs well-motivated and well-justified techniques to improve the supermask based method. \n- The proposed method performs better than the compared methods with a similar setting. \n- The paper conducted carefully designed analyses and discussions on the experiments. \n\nWeakness\n- The whole framework is based on SupSup with similar systematic limitations and benefits, which influences the significance at some level. Specifically, the task identifier is required in both training and testing, restricting the methods on the task-incremental setting. \n- As a result, all the experiments are restricted to the task-incremental setting. Please correct me if I overlooked other settings.\n- Some related works are not discussed, such as (but not limited to) the following ones. There are a series of approaches to generate subnetworks at a module or neuron level to alleviate the forgetting and parameter interference issue in CL. Although the settings may not be the same, they may be discussed. \n\n    - Hurtado, Julio, Alain Raymond, and Alvaro Soto. \"Optimizing reusable knowledge for continual learning via metalearning.\" Advances in Neural Information Processing Systems 34 (2021): 14150-14162.\n\n    - Veniat, Tom, Ludovic Denoyer, and Marc'Aurelio Ranzato. \"Efficient continual learning with modular networks and task-driven priors.\" arXiv preprint arXiv:2012.12631 (2020).\n\n    - Yan, Qingsen, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. \"Learning Bayesian Sparse Networks with Full Experience Replay for Continual Learning.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 109-118. 2022.\n    - Golkar, Siavash, Michael Kagan, and Kyunghyun Cho. \"Continual learning via neural pruning.\" arXiv preprint arXiv:1903.04476 (2019).\n\n- How is the efficiency of the KKT module? It needs to run training and testing on multiple splits. Can the running time on the vision dataset be reported? How does the setting for the KKT module (like number of splits) influence the running time and performance? \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a mask-based approach to obtain subnetworks for specific tasks in continual learning (CL). It is an improvement of the previous method SupSup, which finds fixed supermask for each task to alleviate forgetting. This paper proposes to perform exclusive and nonoverlapping subnetwork training to avoid conflicting updating on the shared parameters. A specific KNN based method is proposed to enhance parameter sharing. The experiments show that the proposed method can produce better performance than other methods. ", "strength_and_weaknesses": "Strength\n- The work is well motivated. It starts from the limitation of the SupSup method and designs well-motivated and well-justified techniques to improve the supermask based method. \n- The proposed method performs better than the compared methods with a similar setting. \n- The paper conducted carefully designed analyses and discussions on the experiments. \n\nWeakness\n- The whole framework is based on SupSup with similar systematic limitations and benefits, which influences the significance at some level. Specifically, the task identifier is required in both training and testing, restricting the methods on the task-incremental setting. \n- As a result, all the experiments are restricted to the task-incremental setting. Please correct me if I overlooked other settings.\n- Some related works are not discussed, such as (but not limited to) the following ones. There are a series of approaches to generate subnetworks at a module or neuron level to alleviate the forgetting and parameter interference issue in CL. Although the settings may not be the same, they may be discussed. \n\n    - Hurtado, Julio, Alain Raymond, and Alvaro Soto. \"Optimizing reusable knowledge for continual learning via metalearning.\" Advances in Neural Information Processing Systems 34 (2021): 14150-14162.\n\n    - Veniat, Tom, Ludovic Denoyer, and Marc'Aurelio Ranzato. \"Efficient continual learning with modular networks and task-driven priors.\" arXiv preprint arXiv:2012.12631 (2020).\n\n    - Yan, Qingsen, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. \"Learning Bayesian Sparse Networks with Full Experience Replay for Continual Learning.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 109-118. 2022.\n    - Golkar, Siavash, Michael Kagan, and Kyunghyun Cho. \"Continual learning via neural pruning.\" arXiv preprint arXiv:1903.04476 (2019).\n\n- How is the efficiency of the KKT module? It needs to run training and testing on multiple splits. Can the running time on the vision dataset be reported? How does the setting for the KKT module (like number of splits) influence the running time and performance? \n\n", "clarity,_quality,_novelty_and_reproducibility": "The whole paper is written clearly, and the work is with good reproducibility, especially considering the relationship with SupSup. \nAlthough the relationship with SupSup may influence the significance of the novel, this work may still provide enough novel insights and observations. \n", "summary_of_the_review": "As discussed above, the work is well motivated and justified, and the paper is well-written. \nThe paper provides enough novel insights and observations and some empirical experimental results and analyses on the proposed method. The significance is influenced by its relationship with SupSup and that it is restricted to the task-incremental setting, where task identifiers are required in both training and testing. \n\n---\nI appreciate the authors' response, which addressed part of my concerns, especially on the KKT module. I am still concerned by the novelty (mainly about the relationship with SupSup) and technical contributions. After further discussions, I am further convinced that the experiment part also needs to be improved, especially in comparison with other parameter isolation-based methods, instead of only SupSup.  \nThe work may be further improved by fixing the issues and conducting some more work, such as some attempts to extend the method to other settings beyond the task-incremental setting.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666683868798}, {"id": "xglMBeZioIy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3457/Reviewer_fPix"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an Exclusive Supermask Subnetwork Training framework for continual learning of both text classification and vision tasks. Compared to the previous method SupSup, the proposed ExSSNet makes fixed weights trainable thus facilitating the knowledge transfer from previously learned tasks to new tasks. The further proposed KKT module can be regarded as a better Initialization mechanism, which helps to learn new tasks faster and better. Extensive experiments and impressing performance validate the effectiveness of the proposed method.", "review_text": "I think this paper can be regarded as an good improvements of the previous method SupSup thus I tend to accept. But my main concern is about the claim of learning a large number of tasks. If the author can give good feedback, I will improve the score.", "strengths": "Strength:\n- The idea of the overall framework is novel compared to previous work SupSup. It is a natural idea to expend fixed weights to trainable weights but it brings a large boost and more advantages, e.g., under the condition of sparse masks.\n- The experiments are sufficient and can effectively validate the claimed points. The informative figures and tables are also helpful for illustration and understanding.\n- The paper is well organized and comprehensive in terms of experimental settings and details.\n\nWeakness:\n- Although the idea is novel, most of the specific implementation of the framework is based on existing work. For example, the supermask learning follows Ramanujan et al.(2019), the training mechanism using exclusive mask is similar as Learn to Grow[A].\n- The description of the paper can be polished. Some long sentences interfere with reading, and there are also a few typos, e.g., the first sentence of the fourth paragraph, “We overcome the aforementioned issues, we propose our method,...”.\n- The main concern is about the claim that ExSSNet can learn 100 tasks. First, as listed in Table 4, SupSup actually achieves a good results (90.34%) and the improvement of ExSSNet (91.21%) is not noticeable. Therefore, the ability to learn 100 tasks is what SupSup have achieved, but not unique for ExSSNet. Second, I wonder that this kind of ability is related to the size of the model since the extreme case is there are no free weights. How is the performance of using a small model like LeNet for SplitMNIST and ResNet18 for SplitCIFAR100. Furthermore, how is the performance of using a larger model like Resnet101? Can larger models learn more than 100 tasks since they have more parameter spaces? \n\nReference:\n- [A] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher. \"Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting.\" International Conference on Machine Learning. PMLR, 2019.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes an Exclusive Supermask Subnetwork Training framework for continual learning of both text classification and vision tasks. Compared to the previous method SupSup, the proposed ExSSNet makes fixed weights trainable thus facilitating the knowledge transfer from previously learned tasks to new tasks. The further proposed KKT module can be regarded as a better Initialization mechanism, which helps to learn new tasks faster and better. Extensive experiments and impressing performance validate the effectiveness of the proposed method.", "strength_and_weaknesses": "Strength:\n- The idea of the overall framework is novel compared to previous work SupSup. It is a natural idea to expend fixed weights to trainable weights but it brings a large boost and more advantages, e.g., under the condition of sparse masks.\n- The experiments are sufficient and can effectively validate the claimed points. The informative figures and tables are also helpful for illustration and understanding.\n- The paper is well organized and comprehensive in terms of experimental settings and details.\n\nWeakness:\n- Although the idea is novel, most of the specific implementation of the framework is based on existing work. For example, the supermask learning follows Ramanujan et al.(2019), the training mechanism using exclusive mask is similar as Learn to Grow[A].\n- The description of the paper can be polished. Some long sentences interfere with reading, and there are also a few typos, e.g., the first sentence of the fourth paragraph, “We overcome the aforementioned issues, we propose our method,...”.\n- The main concern is about the claim that ExSSNet can learn 100 tasks. First, as listed in Table 4, SupSup actually achieves a good results (90.34%) and the improvement of ExSSNet (91.21%) is not noticeable. Therefore, the ability to learn 100 tasks is what SupSup have achieved, but not unique for ExSSNet. Second, I wonder that this kind of ability is related to the size of the model since the extreme case is there are no free weights. How is the performance of using a small model like LeNet for SplitMNIST and ResNet18 for SplitCIFAR100. Furthermore, how is the performance of using a larger model like Resnet101? Can larger models learn more than 100 tasks since they have more parameter spaces? \n\nReference:\n- [A] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher. \"Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting.\" International Conference on Machine Learning. PMLR, 2019.", "clarity,_quality,_novelty_and_reproducibility": "The originality in terms of framework is good while the originality of specific method is not so creative. The overall quality and clarification of the work meets the standard of the conference. ", "summary_of_the_review": "I think this paper can be regarded as an good improvements of the previous method SupSup thus I tend to accept. But my main concern is about the claim of learning a large number of tasks. If the author can give good feedback, I will improve the score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666423377894}, {"id": "jD_hFLfazZF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3457/Reviewer_jWtC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper is proposing a model for continual learning based on subnetwork isolation and training. Specifically, the authors build on top of SupSup, a prior method relying on a randomly initialized backbone where task-specific subnetworks are discovered and isolated, showcasing good performances on a stream of tasks to be learned. This paper extends the method by i) allowing subnetwork parameters to be updated by gradient descent, and therefore achieving better performance on their task and ii) introducing a KNN-based mask initialization method for bootstrapping subnetwork selection on the next task. Experiments are carried out for natural language processing and computer vision tasks.", "review_text": "I appreciate the quality of the research and of the experimental section. However, in my opinion, the technical contribution of the paper is not worth for publication in a venue such as ICLR. I really think this work could benefit more discussion and comparisons with the parameter isolation methods I mentioned above.", "strengths": "+ +The paper is very clear and well written, one read suffices to the interested reader to grasp the main ideas and technical contributions.\n+ +Experiments showcase encouraging results on two different data modalities and tasks, for a total of 5 different datasets among NLP and vision.\n+ +Experimental results are very encouraging, as the model is able to outperform a number of recent works, including some relying on replay buffers.\n+ +An ablation study grounds the effectiveness of KKT, grounding its beneficial effect.\n- -The biggest weakness of the paper is its technical novelty, that I'll discuss in the next point.\n- -The model, needing to separate the relevant task-specific subnetwork during inference, can only be employed in task-incremental learning settings, which represent the easiest and more controlled protocol in CL literature. This raises questions about the suitability of the model for real problems.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper is proposing a model for continual learning based on subnetwork isolation and training. Specifically, the authors build on top of SupSup, a prior method relying on a randomly initialized backbone where task-specific subnetworks are discovered and isolated, showcasing good performances on a stream of tasks to be learned. This paper extends the method by i) allowing subnetwork parameters to be updated by gradient descent, and therefore achieving better performance on their task and ii) introducing a KNN-based mask initialization method for bootstrapping subnetwork selection on the next task. Experiments are carried out for natural language processing and computer vision tasks.", "strength_and_weaknesses": "+ +The paper is very clear and well written, one read suffices to the interested reader to grasp the main ideas and technical contributions.\n+ +Experiments showcase encouraging results on two different data modalities and tasks, for a total of 5 different datasets among NLP and vision.\n+ +Experimental results are very encouraging, as the model is able to outperform a number of recent works, including some relying on replay buffers.\n+ +An ablation study grounds the effectiveness of KKT, grounding its beneficial effect.\n- -The biggest weakness of the paper is its technical novelty, that I'll discuss in the next point.\n- -The model, needing to separate the relevant task-specific subnetwork during inference, can only be employed in task-incremental learning settings, which represent the easiest and more controlled protocol in CL literature. This raises questions about the suitability of the model for real problems.", "clarity,_quality,_novelty_and_reproducibility": "- -The technical contributions of the paper are quite limited. Indeed, the authors build on top of a previous work, SupSup, and add a few incremental updates to improve performance. The first improvement relies on allowing parameters of subnetworks to be updated by the optimization (and freezing the ones used by previous tasks). This contribution is very intuitive and effective, but its addition doesn't constitute sufficient novelty per se. Moreover, the model resembles a number of parameter isolation methods that were proposed in the last few years [1,2,3,4]. Only a couple of them are discussed and the authors claim ExSSNet avoids their shortcomings. However, none of these methods is compared against. More specifically, [1,2] rely on task-specific subnetworks, whose weights and masks are learned jointly, and have similar mechanisms to block gradient updates on parameters used in previous tasks. In this respect, differences between these methods and the proposed work is in details. Finally, the KKT component is interesting and effective, but its employment is a bit contextual to SupSup.\n- -Section 2 is not necessary, as it reiterates the motivations of the paper in a more detailed way. In my opinion, motivations are convincing enough in the introduction.\n- -Difference between SSNet and exSSNet does not emerge clearly in Figure 2. I however acknowledge it emerges in other experiments.\n- -I struggle understanding the notation paragraph in section 3. Specifically, a node is defined as $\\mathcal{Z}_v=\\sigma(\\mathcal{I}_v)$, where $\\mathcal{I}_v$ are inputs, $\\mathcal{Z}_v$ are outputs, and $\\sigma$ is defined as an activation function. This way, it seems that a node is the same as an activation function.\n\n[1] Serra, Joan, et al. \"Overcoming catastrophic forgetting with hard attention to the task.\" International Conference on Machine Learning. PMLR, 2018.\n\n[2] Abati, Davide, et al. \"Conditional channel gated networks for task-aware continual learning.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n\n[3] Mallya, Arun, Dillon Davis, and Svetlana Lazebnik. \"Piggyback: Adapting a single network to multiple tasks by learning to mask weights.\" Proceedings of the European Conference on Computer Vision (ECCV). 2018.\n\n[4] Kim, Jangho, Jeesoo Kim, and Nojun Kwak. \"StackNet: Stacking feature maps for Continual learning.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020.\n", "summary_of_the_review": "I appreciate the quality of the research and of the experimental section. However, in my opinion, the technical contribution of the paper is not worth for publication in a venue such as ICLR. I really think this work could benefit more discussion and comparisons with the parameter isolation methods I mentioned above.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666281773701}], "openreview_url": "https://openreview.net/forum?id=iMy1hOrqiVE", "arxiv_id": "2210.10209", "paper_pdf": "papers/iMy1hOrqiVE.pdf", "paper_pdf_sha256": "121713a735c66b2116797bab35e8a34d49f8afa42aca8e5db3e458c973dcdafb", "paper_pdf_bytes": 596011, "paper_pdf_source": "openreview", "code_url": "https://github.com/prateeky2806/exessnet", "code_repository": "prateeky2806/exessnet", "code_commit": "1e768f855eeb708b7a5e9189acc725f89bd7c98e", "code_archive": "repos/iMy1hOrqiVE.zip", "code_archive_sha256": "f12391621a870c289f94b8b7a81fe8e92834b76fdf2aea54edc51f19a3b0a0c9", "code_archive_bytes": 53725, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 44, "github_languages": {"Python": 171218}, "github_archived": false, "github_pushed_at": "2022-10-20T14:59:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/exclusive-supermask-subnetwork-training-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "v-f7ifhKYps", "year": 2022, "status": "rejected", "title": "Maximum Entropy Population Based Training for Zero-Shot Human-AI Coordination", "authors": ["Rui Zhao", "Jinming Song", "Hu Haifeng", "Yang Gao", "Yi Wu", "Zhongqian Sun", "Yang Wei"], "authorids": ["~Rui_Zhao1", "joshuasong@tencent.com", "~Hu_Haifeng1", "~Yang_Gao1", "~Yi_Wu1", "sallensun@tencent.com", "~Yang_Wei2"], "authors_source": "OpenReview API", "abstract": "An AI agent should be able to coordinate with humans to solve tasks. We consider the problem of training a Reinforcement Learning (RL) agent without using any human data, i.e., in a zero-shot setting, to make it capable of collaborating with humans. Standard RL agents learn through self-play. Unfortunately, these agents only know how to collaborate with themselves and normally do not perform well with unseen partners, such as humans. The methodology of how to train a robust agent in a zero-shot fashion is still subject to research. Motivated from the maximum entropy RL, we derive a centralized population entropy objective to facilitate learning of a diverse population of agents, which is later used to train a robust AI agent to collaborate with unseen partners. The proposed method shows its effectiveness compared to baseline methods, including self-play PPO, the standard Population-Based Training (PBT), and trajectory diversity-based PBT, in the popular Overcooked game environment. We also conduct online experiments with real humans and further demonstrate the efficacy of the method in the real world.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "bu36Otbj8Ea", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper536/Reviewer_PRq2"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper aims to address the important problem of training AI agents in multi-agent games. Standard self-play training leads to agents that overfit to a particular partner, and even naive population based training may not help much if the population is not very diverse. A number of recent works have looked into generating diversity in the partner strategies so that population based training methods can work better. By covering a diverse set of trajectories/policies, it is more likely that coordinating with a new partner (perhaps a human) at test time will fall in-distribution relative to our training set partners.\n\nThe paper proposes a population diversity metric based on cross-entropy between different pairs of agents in the population. They optimize a lower bound to this metric based on the population entropy. They also propose a 'prioritized sampling' procedure that determines which partner in the population the ego-agent should train with. They experiment on the game of Overcooked, compare with TrajeDi, and run a human experiment to check performance of coordination with humans at test time.", "review_text": "Strengths\nThe problem is important, and testing it on Overcooked along with a human experiment is a good domain for this exploration.\nThe population diversity metric is also reasonable, and seems like a good quantity to include as part of an auxiliary loss.\nThe human experiments are neat (although I am left wondering why TrajeDi is not compared with in the human experiments).\n\n\nWeaknesses\n\nThough the paper shows the population entropy to be a lower bound of the population diversity, the gap is huge (factor of n^2). Looking at the entropy values in Table 1, the lowest value is about 0.9 and the highest is about 1.8 (only a factor of 2). As such, even for the smallest choice of n=2, it seems that the lower bound ends up being nearly useless. I'm surprised the paper actually proposes to train based on the lower bound. To me, optimizing an approximation of the population diversity seems much more promising.\n\nI also thought section 3.4 was not well motivated, and it is unclear what the takeaway of this section is. On one hand, it's unclear why we are interested in optimizing the worst performing partner, rather than the average reward of the partners. Moreover, looking at the supporting claims for Eq 10 (Lemma 4), I don't have intuition as to why this claim is not completely trivial -- that training with the lowest performing agent improves the minimum reward over all agents (which seems obvious). Related to this, why does the paper claim uniform sampling \"does not provide any guarantees\" (can't we also claim that training with all agents uniformly increases the overall reward?)\n\nI'm also not convinced by the discussion in Q1 regarding Figure 3. It does not appear that the entropy is converging to a higher value. In fact the plots are very spiky and hard to interpret. Judging from Table 1, however, it does appear that pushing for population entropy does impact the environment reward quite a bit, contrary to the claims in the paper that \"the reward does not decrease much\".\n\nI found the writing to be poor. There are many choppy parts that break the reader's rhythm such as \"less 'panic'\", \"While uniform sampling does not provide any guarantee on the worst case.\". There are also very vague statements, such as \"With prioritized sampling, we make the collaboration between the AI agent and any agent in the population as good as possible in general\". It's unclear to me what precise claim they are making. Of course there are no guarantees regarding the global optimum, so what precisely does it mean to \"make... as good as possible in general\"?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper aims to address the important problem of training AI agents in multi-agent games. Standard self-play training leads to agents that overfit to a particular partner, and even naive population based training may not help much if the population is not very diverse. A number of recent works have looked into generating diversity in the partner strategies so that population based training methods can work better. By covering a diverse set of trajectories/policies, it is more likely that coordinating with a new partner (perhaps a human) at test time will fall in-distribution relative to our training set partners.\n\nThe paper proposes a population diversity metric based on cross-entropy between different pairs of agents in the population. They optimize a lower bound to this metric based on the population entropy. They also propose a 'prioritized sampling' procedure that determines which partner in the population the ego-agent should train with. They experiment on the game of Overcooked, compare with TrajeDi, and run a human experiment to check performance of coordination with humans at test time.", "main_review": "Strengths\nThe problem is important, and testing it on Overcooked along with a human experiment is a good domain for this exploration.\nThe population diversity metric is also reasonable, and seems like a good quantity to include as part of an auxiliary loss.\nThe human experiments are neat (although I am left wondering why TrajeDi is not compared with in the human experiments).\n\n\nWeaknesses\n\nThough the paper shows the population entropy to be a lower bound of the population diversity, the gap is huge (factor of n^2). Looking at the entropy values in Table 1, the lowest value is about 0.9 and the highest is about 1.8 (only a factor of 2). As such, even for the smallest choice of n=2, it seems that the lower bound ends up being nearly useless. I'm surprised the paper actually proposes to train based on the lower bound. To me, optimizing an approximation of the population diversity seems much more promising.\n\nI also thought section 3.4 was not well motivated, and it is unclear what the takeaway of this section is. On one hand, it's unclear why we are interested in optimizing the worst performing partner, rather than the average reward of the partners. Moreover, looking at the supporting claims for Eq 10 (Lemma 4), I don't have intuition as to why this claim is not completely trivial -- that training with the lowest performing agent improves the minimum reward over all agents (which seems obvious). Related to this, why does the paper claim uniform sampling \"does not provide any guarantees\" (can't we also claim that training with all agents uniformly increases the overall reward?)\n\nI'm also not convinced by the discussion in Q1 regarding Figure 3. It does not appear that the entropy is converging to a higher value. In fact the plots are very spiky and hard to interpret. Judging from Table 1, however, it does appear that pushing for population entropy does impact the environment reward quite a bit, contrary to the claims in the paper that \"the reward does not decrease much\".\n\nI found the writing to be poor. There are many choppy parts that break the reader's rhythm such as \"less 'panic'\", \"While uniform sampling does not provide any guarantee on the worst case.\". There are also very vague statements, such as \"With prioritized sampling, we make the collaboration between the AI agent and any agent in the population as good as possible in general\". It's unclear to me what precise claim they are making. Of course there are no guarantees regarding the global optimum, so what precisely does it mean to \"make... as good as possible in general\"?", "summary_of_the_review": "The paper tackles an important problem, but leaves much room for improvement. The population entropy should be backed up by (toy) experiments that analyze the effect of the quantity, since on paper it seems that the lower bound exhibits a huge gap compared to the empirical range of entropy values observed in Table 1. The experiments can also be improved (TrajeDi is missing in the human experiments), and the current results are hard to interpret (the reward in Table 1 does drop significantly when the entropy term is pushed up).   Improving the overall writing would also raise the paper's potential impact.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635973806677}, {"id": "l78MVPtuDqY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper536/Reviewer_5ykj"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a maximum entropy objective that can be applied to population-based multi-agent training in an RL setting to create a population of multi-agent policies that can more easily cope with zero-shot introductions of \"unseen\" policies. They formulate a framework for successfully applying this objective at scale by working out a proxy objective, and apply the method to the multi-agent Overcook environment tasks.", "review_text": "\nI am struggling to come up with sound overall grade for this paper. In short:\n\n1. The method seems to be technically sound, building on simple and relatively popular ideas in both multi-agent and single-agent RL in an interesting way. \n\n2. It seems to be effective when tested against comparable baselines in an interesting multi-agent environment.\n\n3. The submission does a good job at providing enough details (and code) to enable quickly reproducing the work.\n\nHowever:\n\n4. The experimental section is relatively weak:\n\n    a. Firstly, multi-agent Overcook is a relatively recent benchmark all things considered (and I would definitely not define it \"popular\"). There are a variety of choices, ranging from toy-like environments (matrix games, OpenAI's MPE) to complex ones (SMAC, Hanabi, DM Lab) that are popular in the recently wildly exploding MARL literature that could also have been employed, and possibly preferred. This would have enabled to understand the significance of the results compared to state of the art zero-shot coordination.\n\n    b. This is particularly a problem when comparing this manuscript with Lupu et al., 2019, which the authors acknowledge being comparable work. Considering that they are comparing their methods against theirs, it seems unreasonable not to benchmark MEP on at least some of the toy problems proposed in that work (whilst I would accept the authors not trying to compete with their well-tuned Hanabi results, as that would result in quite a lot of potential work).\n\n5. The focus on AI-human coordination seems a little weird. In principle this paper is proposing to increase the diversity of states seen by the agent policy by maximising the types of behaviours generated during training time. This means that the method is applicable to be tested against any kind of out-of-distribution(-ish) agents. This is not a huge issue, but implies that the experimental section could have focused more greatly on undestanding the space of ad-hoc policies over which it is more robust vs possible failure cases. In my opinion, generating an adversarial out of distribution set of agents to test against would have been more effective than qualitatively analysing a comparably small quantity of human trials.\n\n6. The paper at times is unnecessarily handwavy, unclear, or makes dubious statements that are not well backed up by the literature. Here's some examples:\n\n   a. In the introduction, it is stated that self-play is the \"mainstream method for building state-of-the-art AI agents\" (without reference), but self-play makes for a relatively limited amount of work when looking at the broad RL / MARL literature. I would argue that it is _barely_ mainstream, given the implementation complexity of even its simplest form. It is also stated that \"self-play-trained agents are very specialized\", but this is not a well understood property of PBT, and in practice doesn't seem to have significantly affected performance when properly done (see e.g. Alphastar).\n\n   b. It is claimed that \"prioritized sampling [of agents policies from the population] [makes] the collaboration between the AI agent and any agent in the population as good as possible in general\" -- I don't understand how to interpret the sentence: is the manuscript saying that prioritized sampling is generally optimal (a fairly strong claim!) wrt. learning with PBT for cooperative MARL? How does this interact with the fact that the schema utilises a particular ranking system that might be more or less compatible with the task?\n\n### Nits\n\n- Generally: the [...] RL -> [...] RL\n- Section 1: prioritized sampling [of what?]\n- Section 1: the experimental section is not a contribution of the work -- there's nothing intrinsically new about how MEP was tested, as far as I can see?\n- Section 2: modes of sub-optimal -> modes of optimal (?)\n- Section 3.1: what do incentive and multi-modal mean here?\n- Section 3.4: the first paragraph feels off in terms of syntax / punctuation.\n- Figure 3: hard to interpret -- it feels like it could have been reduced to two plots by grouping wrt. y-axis and alpha.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a maximum entropy objective that can be applied to population-based multi-agent training in an RL setting to create a population of multi-agent policies that can more easily cope with zero-shot introductions of \"unseen\" policies. They formulate a framework for successfully applying this objective at scale by working out a proxy objective, and apply the method to the multi-agent Overcook environment tasks.", "main_review": "\nI am struggling to come up with sound overall grade for this paper. In short:\n\n1. The method seems to be technically sound, building on simple and relatively popular ideas in both multi-agent and single-agent RL in an interesting way. \n\n2. It seems to be effective when tested against comparable baselines in an interesting multi-agent environment.\n\n3. The submission does a good job at providing enough details (and code) to enable quickly reproducing the work.\n\nHowever:\n\n4. The experimental section is relatively weak:\n\n    a. Firstly, multi-agent Overcook is a relatively recent benchmark all things considered (and I would definitely not define it \"popular\"). There are a variety of choices, ranging from toy-like environments (matrix games, OpenAI's MPE) to complex ones (SMAC, Hanabi, DM Lab) that are popular in the recently wildly exploding MARL literature that could also have been employed, and possibly preferred. This would have enabled to understand the significance of the results compared to state of the art zero-shot coordination.\n\n    b. This is particularly a problem when comparing this manuscript with Lupu et al., 2019, which the authors acknowledge being comparable work. Considering that they are comparing their methods against theirs, it seems unreasonable not to benchmark MEP on at least some of the toy problems proposed in that work (whilst I would accept the authors not trying to compete with their well-tuned Hanabi results, as that would result in quite a lot of potential work).\n\n5. The focus on AI-human coordination seems a little weird. In principle this paper is proposing to increase the diversity of states seen by the agent policy by maximising the types of behaviours generated during training time. This means that the method is applicable to be tested against any kind of out-of-distribution(-ish) agents. This is not a huge issue, but implies that the experimental section could have focused more greatly on undestanding the space of ad-hoc policies over which it is more robust vs possible failure cases. In my opinion, generating an adversarial out of distribution set of agents to test against would have been more effective than qualitatively analysing a comparably small quantity of human trials.\n\n6. The paper at times is unnecessarily handwavy, unclear, or makes dubious statements that are not well backed up by the literature. Here's some examples:\n\n   a. In the introduction, it is stated that self-play is the \"mainstream method for building state-of-the-art AI agents\" (without reference), but self-play makes for a relatively limited amount of work when looking at the broad RL / MARL literature. I would argue that it is _barely_ mainstream, given the implementation complexity of even its simplest form. It is also stated that \"self-play-trained agents are very specialized\", but this is not a well understood property of PBT, and in practice doesn't seem to have significantly affected performance when properly done (see e.g. Alphastar).\n\n   b. It is claimed that \"prioritized sampling [of agents policies from the population] [makes] the collaboration between the AI agent and any agent in the population as good as possible in general\" -- I don't understand how to interpret the sentence: is the manuscript saying that prioritized sampling is generally optimal (a fairly strong claim!) wrt. learning with PBT for cooperative MARL? How does this interact with the fact that the schema utilises a particular ranking system that might be more or less compatible with the task?\n\n### Nits\n\n- Generally: the [...] RL -> [...] RL\n- Section 1: prioritized sampling [of what?]\n- Section 1: the experimental section is not a contribution of the work -- there's nothing intrinsically new about how MEP was tested, as far as I can see?\n- Section 2: modes of sub-optimal -> modes of optimal (?)\n- Section 3.1: what do incentive and multi-modal mean here?\n- Section 3.4: the first paragraph feels off in terms of syntax / punctuation.\n- Figure 3: hard to interpret -- it feels like it could have been reduced to two plots by grouping wrt. y-axis and alpha.", "summary_of_the_review": "Overall, the manuscript presents an idea that seems compelling and useful, but the experimental section doesn't provide enough signal to compare this method against the literature. This makes my recommendation borderline at this point, so I'm looking forward to discussing the manuscript with the authors and the rest of the reviewers to understand how to improve it towards possibly acceptance.\n\n---\n\nBumped up score to weak accept.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635952992890}, {"id": "GQw_qCuPXCU", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper536/Reviewer_1grX"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper tries to find a new approach by enforcing the diversity in multi-agent RL  via the maximum entropy to address the zero-shot human-AI coordination problem. More specifically, in the proposed Maximum Entropy Population-based training (MEP) framework,  the authors choose population entropy as an efficient surrogate objective and use the prioritized sampling. The empirical results on the overcooked show that MEP outperforms other baselines with both simulated and real human players. \n", "review_text": "Strength:\n\n1. The paper is well-written and easy to follow.\n2. The idea of combing the population entropy during the training is interesting, which considers both agents' individual diversity and pairwise diversity.\n\nConcerns:\n1. The authors propose a good way to promote the policy diversity in multi-agent RL, and this is straightforward that the diversity policy would benefit the zero-shot human-ai coordination. However, I think policy diversity is a general method that can be useful in many other problems, not only limited to human-ai coordination. Therefore, I think the authors can discuss the broader impact of the paper and clarify more about why test this method on specific human-ai coordination problem.\n2. The prioritized sampling + entropy population seems important to train the diverse policy, and an ablation study to investigate how each component works would be better.\n3. Some important references about diversity in population-based multi-agent RL are missing, like [1,2,3]. It would be good to discuss the relationship between the proposed method and the diversity promoting solutions in the PSRO framework.\n\n[1] Balduzzi, David, et al. \"Open-ended learning in symmetric zero-sum games.\" International Conference on Machine Learning. PMLR, 2019.\n\n[2] Nieves, Nicolas Perez, et al. \"Modelling behavioural diversity for learning in open-ended games.\" arXiv preprint arXiv:2103.07927 (2021).\n\n[3] Liu, Xiangyu, et al. \"Unifying Behavioral and Response Diversity for Open-ended Learning in Zero-sum Games.\" arXiv preprint arXiv:2106.04958 (2021).\n\n\nTypos & Questions:\n1. Algo. 1 gives a detailed algorithm procedure, but I can still be confused about obtaining the initial population/ policy pool? Is it a fixed policy pool, or would it be expanded by adding the new learned policy during the training?\n2. Page 5: The repeat 'Environment' sub-section names of the last two paragraphs.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper tries to find a new approach by enforcing the diversity in multi-agent RL  via the maximum entropy to address the zero-shot human-AI coordination problem. More specifically, in the proposed Maximum Entropy Population-based training (MEP) framework,  the authors choose population entropy as an efficient surrogate objective and use the prioritized sampling. The empirical results on the overcooked show that MEP outperforms other baselines with both simulated and real human players. \n", "main_review": "Strength:\n\n1. The paper is well-written and easy to follow.\n2. The idea of combing the population entropy during the training is interesting, which considers both agents' individual diversity and pairwise diversity.\n\nConcerns:\n1. The authors propose a good way to promote the policy diversity in multi-agent RL, and this is straightforward that the diversity policy would benefit the zero-shot human-ai coordination. However, I think policy diversity is a general method that can be useful in many other problems, not only limited to human-ai coordination. Therefore, I think the authors can discuss the broader impact of the paper and clarify more about why test this method on specific human-ai coordination problem.\n2. The prioritized sampling + entropy population seems important to train the diverse policy, and an ablation study to investigate how each component works would be better.\n3. Some important references about diversity in population-based multi-agent RL are missing, like [1,2,3]. It would be good to discuss the relationship between the proposed method and the diversity promoting solutions in the PSRO framework.\n\n[1] Balduzzi, David, et al. \"Open-ended learning in symmetric zero-sum games.\" International Conference on Machine Learning. PMLR, 2019.\n\n[2] Nieves, Nicolas Perez, et al. \"Modelling behavioural diversity for learning in open-ended games.\" arXiv preprint arXiv:2103.07927 (2021).\n\n[3] Liu, Xiangyu, et al. \"Unifying Behavioral and Response Diversity for Open-ended Learning in Zero-sum Games.\" arXiv preprint arXiv:2106.04958 (2021).\n\n\nTypos & Questions:\n1. Algo. 1 gives a detailed algorithm procedure, but I can still be confused about obtaining the initial population/ policy pool? Is it a fixed policy pool, or would it be expanded by adding the new learned policy during the training?\n2. Page 5: The repeat 'Environment' sub-section names of the last two paragraphs.", "summary_of_the_review": "This paper presents an interesting solution to train the diversity policy with a population, but more clarifications (as stated above) are required. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635906317613}, {"id": "xNQs5ZNA5mT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper536/Reviewer_MRtf"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes an approach for training agents that are capable of ad hoc coordination with humans. The method combines an entropy based objective with a method for prioritizing partner selection. Results on the Overcooked domain show improved performance over a series of baselines.", "review_text": "At present I do not believe this paper should be published at ICLR, for the following reason: it claims to present a novel approach yet it simply combines two previously published methods (TrajeDi and PFSP). The framing of the method is incorrect, it does not cite PFSP until related work (not in Section 3.4) and it claims TrajeDi is \"concurrent\" which is not true. TrajeDi was presented at AAMAS 2021 (in May) and again at ICML 2021 (in July). This to me is not concurrent. To be concrete on the differences, the TrajeDi objective is almost identical, when the $\\gamma$ parameter is set to zero. Meanwhile, PFSP uses almost the same prioritization scheme, but the only difference is that it is used in competitive games vs. cooperative.\n\nHowever, I believe there is material in this paper that is worthy of publication, and a simple reframing/restructuring may be sufficient for the paper to be a useful contribution. Indeed, the paper does present novel theoretical results, a novel *combination* of existing methods in a new setting (since PFSP has only been used in competitive settings until now), and interesting experiments including testing their agents with humans. This could be a good paper if it reduces claims of novelty and instead focuses on how it builds on these previous works. **If significant changes are made to the claims in the paper and presentation of the methods I would be happy to raise my score**.\n\nMore Detailed Comments:\n* The ablation studies are useful. Particularly the inclusion of the prioritization scheme with TrajeDi. Which value for $\\gamma$ was used for TrajeDi? I would guess it is using $\\gamma=1$ which means the results show us the difference between action diversity and trajectory diversity, which is an interesting result, but likely problem specific. What happens if you use other hyperparameters for TrajeDi?\n* Figure 3 has a lot of redundancy. I would move both the episode reward curves onto the same plot and also move both of the entropy curves onto the same plot, then use the colors to differentiate between the methods and a shared legend. Then have the name of the plot on the title (across the top) rather than as a y-axis label.\n* Missing baseline: Off Belief Learning (Hu et al 2021) is the state-of-the-art method for ZSC in Hanabi. It would be interesting to see how it performs here. At minimum it should be cited.\n* For the hyperparameter studies in the Appendix, it would be great if we can also compare different configurations on the same plots with the same axis. \n* As always, it is great that code is included.\n* The human experiments are interesting. However, it seems the baselines are reduced. How does TrajeDi perform here? Does the addition of PFSP impact the human coordination? It seems like there are a lot of unanswered questions. Nonetheless, as it is this is a strong result.\n* There are no discussion of limitations in the paper. Honest discussion of this would make the work stronger. For example, how would it work with larger population sizes and higher dimensional problems (such as Hanabi).\n* Figure 5 - my guess is the bolding just means a higher average, without considering the error bars. It is better to only bold if the error bars do not overlap.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an approach for training agents that are capable of ad hoc coordination with humans. The method combines an entropy based objective with a method for prioritizing partner selection. Results on the Overcooked domain show improved performance over a series of baselines.", "main_review": "At present I do not believe this paper should be published at ICLR, for the following reason: it claims to present a novel approach yet it simply combines two previously published methods (TrajeDi and PFSP). The framing of the method is incorrect, it does not cite PFSP until related work (not in Section 3.4) and it claims TrajeDi is \"concurrent\" which is not true. TrajeDi was presented at AAMAS 2021 (in May) and again at ICML 2021 (in July). This to me is not concurrent. To be concrete on the differences, the TrajeDi objective is almost identical, when the $\\gamma$ parameter is set to zero. Meanwhile, PFSP uses almost the same prioritization scheme, but the only difference is that it is used in competitive games vs. cooperative.\n\nHowever, I believe there is material in this paper that is worthy of publication, and a simple reframing/restructuring may be sufficient for the paper to be a useful contribution. Indeed, the paper does present novel theoretical results, a novel *combination* of existing methods in a new setting (since PFSP has only been used in competitive settings until now), and interesting experiments including testing their agents with humans. This could be a good paper if it reduces claims of novelty and instead focuses on how it builds on these previous works. **If significant changes are made to the claims in the paper and presentation of the methods I would be happy to raise my score**.\n\nMore Detailed Comments:\n* The ablation studies are useful. Particularly the inclusion of the prioritization scheme with TrajeDi. Which value for $\\gamma$ was used for TrajeDi? I would guess it is using $\\gamma=1$ which means the results show us the difference between action diversity and trajectory diversity, which is an interesting result, but likely problem specific. What happens if you use other hyperparameters for TrajeDi?\n* Figure 3 has a lot of redundancy. I would move both the episode reward curves onto the same plot and also move both of the entropy curves onto the same plot, then use the colors to differentiate between the methods and a shared legend. Then have the name of the plot on the title (across the top) rather than as a y-axis label.\n* Missing baseline: Off Belief Learning (Hu et al 2021) is the state-of-the-art method for ZSC in Hanabi. It would be interesting to see how it performs here. At minimum it should be cited.\n* For the hyperparameter studies in the Appendix, it would be great if we can also compare different configurations on the same plots with the same axis. \n* As always, it is great that code is included.\n* The human experiments are interesting. However, it seems the baselines are reduced. How does TrajeDi perform here? Does the addition of PFSP impact the human coordination? It seems like there are a lot of unanswered questions. Nonetheless, as it is this is a strong result.\n* There are no discussion of limitations in the paper. Honest discussion of this would make the work stronger. For example, how would it work with larger population sizes and higher dimensional problems (such as Hanabi).\n* Figure 5 - my guess is the bolding just means a higher average, without considering the error bars. It is better to only bold if the error bars do not overlap.", "summary_of_the_review": "The paper is interesting and well written, and addresses an important problem. The method itself combines two known methods (TrajeDi and PFSP) but does not provide sufficient credit, since it claims TrajeDi is \"concurrent\" and it only discusses PFSP in related work. Accurately positioning the new contribution w.r.t the previous would make this a solid contribution.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635780938160}, {"id": "O79X6ooM_5v", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper536/Reviewer_mavT"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors focus on training agents for zero-shot human-AI coordination on Overcooked. They propose to first train a population of agents with a population entropy bonus, and then train another agent as the best response to the population, using a prioritized sampling approach that focuses more training on the worst performing pairs. They evaluate their agent and some baselines and ablations both with a human proxy model (i.e. an agent trained on human-human gameplay data with behavioral cloning) and also a smaller subset of their agent and baselines with real humans on Mechanical Turk. Their method performs as well or better in terms of game score than the baselines and ablations presented.", "review_text": "STRENGTHS\n\nThe high-level motivation of injecting diversity into partner populations is a sound and important one for human-AI coordination. The paper goes beyond just generalization to held-out agents and evaluates with real humans, which I think is crucial for this line of work. The authors include in their supplementary data videos of gameplay, and additionally discuss qualitative aspects of agent behavior, which I found useful in guiding intuition.\n\nWEAKNESSES\n\nThere are several crucial missing or misunderstood ablations and baselines.\n\n**Misunderstood baseline - TrajeDi**: the JSD term in the authors' population entropy bonus in equation 7 is just a special case of TrajeDi with gamma=0 (see equation 5 and discussion following in the TrajeDi paper). Thus, the population entropy bonus just amounts to adding the usual (per-agent) entropy bonus to TrajeDi. Can the authors comment? In addition, I could not find the hyperparameters or tuning procedure for the TrajeDi baseline. What value of gamma was used and how was it selected? What size population was used for TrajeDi? Comparing Fig 4a and 4c, it looks like the TrajeDi agent performs significantly worse than self-play, which suggests it was poorly tuned.\n\nSeveral times the authors suggest they compare to \"state-of-the-art\" baselines, but the two methods that to my knowledge have a claim to this are not included.\n\n**Missing baseline #1 - PPO_BC**: while the authors include several baselines from the Carroll et al 2019 paper, they for some reason do not include the best-performing one - the best response to a human BC model, PPO_BC. The authors might rightly protest that the comparison wouldn't be exactly apples to apples, since PPO_BC gets to use human data in training. For this reason, I don't think its a prerequisite of a human data-free method to *beat* PPO_BC, but it is still an important, informative, and easy baseline to include for comparison.\n\n**Missing baseline #2 - FCP**: more recently, Strouse at al 2021 (https://arxiv.org/abs/2110.08176) introduced fictitious co-play (FCP), which achieved state-of-the-art human-AI coordination on Overcooked without human data, even seemingly beating the human BC model best response agent above. FCP similarly involves a two-stage training process of first training a diverse set of partners, and then training a best response to them. However, FCP achieves diversity solely through seeds of self-play, and taking checkpoints from different points of training (beginning, middle, and end). Strangely, buried in appendix E, the present authors mention that they too train their best response agent with multiple checkpoints of their partner population, again from the beginning, middle, and end of training. First, this seems like an odd and important detail to bury in an appendix and not mention in the main text. Second, it is an important enough detail to deserve an ablation (which the FCP paper did). Moreover, the combination of omitting mention of FCP while simultaneously using an eerily similar trick and not mentioning it in the main text looks pretty suspicious. In any case, from what I can tell, FCP is essentially the authors' present setup but with alpha=0, beta=0, and a bigger population (N=32 vs N=5) (though the RL algorithms also differ, i.e. VMPO vs PPO). In other words, if the authors use larger partner populations, are the population entropy bonus and prioritized sampling still important?\n\n**Missing ablation #1 - past checkpoints**: as mentioned in the paragraph above, the authors mentioned in appendix E that they train with multiple checkpoints of their partner population. That seems like an important detail to ablate.\n\n**Missing ablation #2 - two terms in population entropy**: the authors' population entropy objective (equation 7) contains two terms - a JSD term and the usual per-agent entropy bonus. The latter is a standard addition in many RL setups. The alpha=0 ablation in Figure 4b turns off both the JSD and per-agent entropy terms together. Thus, it is unclear whether the JSD term, the per-agent entropy bonus, or the combination is important.\n\n**Question about uniform sampling ablation**: I was surprised to see this version do so poorly. Is it just that training takes a bit longer with uniform sampling? Had this agent converged? I couldn't find learning curves for this ablation.\n\n**Missing baselines and ablations in human experiments**: none of the important baselines or ablations were evaluated with humans. While results with the human proxy model are helpful and suggestive, I think its important to include the key baselines and ablations in the human experiments as well, i.e. TrajeDi, alpha=0, beta=0, and the other baselines and ablations mentioned above.\n\n**Between-subjects design**: This worries me, since I'm guessing the evaluation of the MEP agent and the original Carroll et al 2019 agents took place at least two years apart, and may have been performed by different authors. Unfortunately with human evaluations, many small details matter, such as server lag, UI, and sample population, all of which are hard to control for when evaluating two years apart. Could the authors provide additional justification for how they ensure the evaluations are identical? If not, and since I would like to see other baselines and ablations evaluated with humans anyway, perhaps it would be easy to re-evaluate SP and PBT.\n\nLastly, a few more minor points:\n1. Structurally, I found section 3.1 and Theorem 1 unnecessary. It would be simpler to just start their narrative with equation 7, which at least to me is more well motivated than equation 2 anyway.\n2. The JSD term in equation 7 is also closely related to the Emergence of Individuality (EOI) reward of Jiang & Lu 2021 (https://arxiv.org/abs/2006.05842), since the JSD term can also be understood as the mutual information between agent index and action choice, conditioned on state, or I(i;a|s).\n3. In the answer to Question 1, the authors state that as alpha increases from 0 to .01, the population entropy increases while the reward barely decreases. However, from what I can tell, the changes are of similar magnitude, with reward dropping about 20% and entropy increasing 20%, so this seems like a misleading claim.\n4. In the answer to Question 3, the authors state that MEP outperforms SP and PBT in all environments. However, the error bars seem to overlap on Forced, so I might change this to say \"performs as well or better.\"\n5. In the intro, are cooperative games and emergent communication really \"real world applications\"?\n6. Tylkin et al 2020 (https://econcs.seas.harvard.edu/files/econcs/files/tylkin_neurips20.pdf) is relevant for citation and discussion on training agents for human-AI coordination.\n7. MAVEN from Mahajan et al 2019 (https://arxiv.org/abs/1910.07483), as well as EOI mentioned above, are missing related work on applying population diversity-based methods in the multi-agent setting.\n8. TrajeDi is described as \"concurrent\" work, but the paper has been out for several months. \"Concurrent\" is I suppose of debatable definition, but I think its more appropriate for work that is in submission at the same time.\n9. There are many typos and grammatical errors in the paper. They did not damage readability for me, and I do not penalize the authors for this, but the paper would benefit from a proofreading. Some examples: \"from the max ent RL\" -> \"by max ent RL\" (abstract), \"assistant\" -> \"assist\" (paragraph 1 of intro), \"use\" -> \"using\" (paragraph  4 of intro), \"incentive\" -> \"diverse\"? (paragraph 1 of sec 3.1), \"unbound\" -> \"unbounded\" (last paragraph of sec 3.1), \"Compare\" -> \"Compared\" and \"Take\" -> \"Taking\" (paragraph after equation 6), \"panic\" -> \"panicked\" (sentence before sec 3.4), etc etc.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors focus on training agents for zero-shot human-AI coordination on Overcooked. They propose to first train a population of agents with a population entropy bonus, and then train another agent as the best response to the population, using a prioritized sampling approach that focuses more training on the worst performing pairs. They evaluate their agent and some baselines and ablations both with a human proxy model (i.e. an agent trained on human-human gameplay data with behavioral cloning) and also a smaller subset of their agent and baselines with real humans on Mechanical Turk. Their method performs as well or better in terms of game score than the baselines and ablations presented.", "main_review": "STRENGTHS\n\nThe high-level motivation of injecting diversity into partner populations is a sound and important one for human-AI coordination. The paper goes beyond just generalization to held-out agents and evaluates with real humans, which I think is crucial for this line of work. The authors include in their supplementary data videos of gameplay, and additionally discuss qualitative aspects of agent behavior, which I found useful in guiding intuition.\n\nWEAKNESSES\n\nThere are several crucial missing or misunderstood ablations and baselines.\n\n**Misunderstood baseline - TrajeDi**: the JSD term in the authors' population entropy bonus in equation 7 is just a special case of TrajeDi with gamma=0 (see equation 5 and discussion following in the TrajeDi paper). Thus, the population entropy bonus just amounts to adding the usual (per-agent) entropy bonus to TrajeDi. Can the authors comment? In addition, I could not find the hyperparameters or tuning procedure for the TrajeDi baseline. What value of gamma was used and how was it selected? What size population was used for TrajeDi? Comparing Fig 4a and 4c, it looks like the TrajeDi agent performs significantly worse than self-play, which suggests it was poorly tuned.\n\nSeveral times the authors suggest they compare to \"state-of-the-art\" baselines, but the two methods that to my knowledge have a claim to this are not included.\n\n**Missing baseline #1 - PPO_BC**: while the authors include several baselines from the Carroll et al 2019 paper, they for some reason do not include the best-performing one - the best response to a human BC model, PPO_BC. The authors might rightly protest that the comparison wouldn't be exactly apples to apples, since PPO_BC gets to use human data in training. For this reason, I don't think its a prerequisite of a human data-free method to *beat* PPO_BC, but it is still an important, informative, and easy baseline to include for comparison.\n\n**Missing baseline #2 - FCP**: more recently, Strouse at al 2021 (https://arxiv.org/abs/2110.08176) introduced fictitious co-play (FCP), which achieved state-of-the-art human-AI coordination on Overcooked without human data, even seemingly beating the human BC model best response agent above. FCP similarly involves a two-stage training process of first training a diverse set of partners, and then training a best response to them. However, FCP achieves diversity solely through seeds of self-play, and taking checkpoints from different points of training (beginning, middle, and end). Strangely, buried in appendix E, the present authors mention that they too train their best response agent with multiple checkpoints of their partner population, again from the beginning, middle, and end of training. First, this seems like an odd and important detail to bury in an appendix and not mention in the main text. Second, it is an important enough detail to deserve an ablation (which the FCP paper did). Moreover, the combination of omitting mention of FCP while simultaneously using an eerily similar trick and not mentioning it in the main text looks pretty suspicious. In any case, from what I can tell, FCP is essentially the authors' present setup but with alpha=0, beta=0, and a bigger population (N=32 vs N=5) (though the RL algorithms also differ, i.e. VMPO vs PPO). In other words, if the authors use larger partner populations, are the population entropy bonus and prioritized sampling still important?\n\n**Missing ablation #1 - past checkpoints**: as mentioned in the paragraph above, the authors mentioned in appendix E that they train with multiple checkpoints of their partner population. That seems like an important detail to ablate.\n\n**Missing ablation #2 - two terms in population entropy**: the authors' population entropy objective (equation 7) contains two terms - a JSD term and the usual per-agent entropy bonus. The latter is a standard addition in many RL setups. The alpha=0 ablation in Figure 4b turns off both the JSD and per-agent entropy terms together. Thus, it is unclear whether the JSD term, the per-agent entropy bonus, or the combination is important.\n\n**Question about uniform sampling ablation**: I was surprised to see this version do so poorly. Is it just that training takes a bit longer with uniform sampling? Had this agent converged? I couldn't find learning curves for this ablation.\n\n**Missing baselines and ablations in human experiments**: none of the important baselines or ablations were evaluated with humans. While results with the human proxy model are helpful and suggestive, I think its important to include the key baselines and ablations in the human experiments as well, i.e. TrajeDi, alpha=0, beta=0, and the other baselines and ablations mentioned above.\n\n**Between-subjects design**: This worries me, since I'm guessing the evaluation of the MEP agent and the original Carroll et al 2019 agents took place at least two years apart, and may have been performed by different authors. Unfortunately with human evaluations, many small details matter, such as server lag, UI, and sample population, all of which are hard to control for when evaluating two years apart. Could the authors provide additional justification for how they ensure the evaluations are identical? If not, and since I would like to see other baselines and ablations evaluated with humans anyway, perhaps it would be easy to re-evaluate SP and PBT.\n\nLastly, a few more minor points:\n1. Structurally, I found section 3.1 and Theorem 1 unnecessary. It would be simpler to just start their narrative with equation 7, which at least to me is more well motivated than equation 2 anyway.\n2. The JSD term in equation 7 is also closely related to the Emergence of Individuality (EOI) reward of Jiang & Lu 2021 (https://arxiv.org/abs/2006.05842), since the JSD term can also be understood as the mutual information between agent index and action choice, conditioned on state, or I(i;a|s).\n3. In the answer to Question 1, the authors state that as alpha increases from 0 to .01, the population entropy increases while the reward barely decreases. However, from what I can tell, the changes are of similar magnitude, with reward dropping about 20% and entropy increasing 20%, so this seems like a misleading claim.\n4. In the answer to Question 3, the authors state that MEP outperforms SP and PBT in all environments. However, the error bars seem to overlap on Forced, so I might change this to say \"performs as well or better.\"\n5. In the intro, are cooperative games and emergent communication really \"real world applications\"?\n6. Tylkin et al 2020 (https://econcs.seas.harvard.edu/files/econcs/files/tylkin_neurips20.pdf) is relevant for citation and discussion on training agents for human-AI coordination.\n7. MAVEN from Mahajan et al 2019 (https://arxiv.org/abs/1910.07483), as well as EOI mentioned above, are missing related work on applying population diversity-based methods in the multi-agent setting.\n8. TrajeDi is described as \"concurrent\" work, but the paper has been out for several months. \"Concurrent\" is I suppose of debatable definition, but I think its more appropriate for work that is in submission at the same time.\n9. There are many typos and grammatical errors in the paper. They did not damage readability for me, and I do not penalize the authors for this, but the paper would benefit from a proofreading. Some examples: \"from the max ent RL\" -> \"by max ent RL\" (abstract), \"assistant\" -> \"assist\" (paragraph 1 of intro), \"use\" -> \"using\" (paragraph  4 of intro), \"incentive\" -> \"diverse\"? (paragraph 1 of sec 3.1), \"unbound\" -> \"unbounded\" (last paragraph of sec 3.1), \"Compare\" -> \"Compared\" and \"Take\" -> \"Taking\" (paragraph after equation 6), \"panic\" -> \"panicked\" (sentence before sec 3.4), etc etc.", "summary_of_the_review": "There are several crucial missing baselines and ablations and clarifications before I think the paper is ready for publication. However, pending their inclusion and a review of the new results, I am very open to raising my score to an accept.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635730345390}], "openreview_url": "https://openreview.net/forum?id=v-f7ifhKYps", "arxiv_id": "2112.11701", "paper_pdf": "papers/v-f7ifhKYps.pdf", "paper_pdf_sha256": "adb015a81ba67976c35b71e206708da9acba6d57cd04b0211f160f851f363e6c", "paper_pdf_bytes": 2582698, "paper_pdf_source": "openreview", "code_url": "https://github.com/ruizhaogit/maximum_entropy_population_based_training", "code_repository": "ruizhaogit/maximum_entropy_population_based_training", "code_commit": "518ea6a2f9ccc75c1c252bafaf946de62dc96e05", "code_archive": "repos/v-f7ifhKYps.zip", "code_archive_sha256": "73887f6251f025da74b83bcc1b0de66c6f8ba823085a030b5fbd07f16f7a411b", "code_archive_bytes": 48551, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 42, "github_languages": {"Python": 144349}, "github_archived": false, "github_pushed_at": "2022-11-29T12:00:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/maximum-entropy-population-based-training-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "L5b6jUonKFB", "year": 2021, "status": "rejected", "title": "Deep Continuous Networks", "authors": ["Nergis Tomen", "Silvia Laura Pintea", "Jan van Gemert"], "authorids": ["~Nergis_Tomen1", "~Silvia_Laura_Pintea1", "~Jan_van_Gemert1"], "authors_source": "OpenReview API", "abstract": "CNNs and computational models of biological vision share some fundamental principles, which, combined with recent developments in deep learning, have opened up new avenues of research in neuroscience. However, in contrast to biological models, conventional CNN architectures are based on spatio-temporally discrete representations, and thus cannot accommodate certain aspects of biological complexity such as continuously varying receptive field sizes and temporal dynamics of neuronal responses. Here we propose deep continuous networks (DCNs), which combine spatially continuous convolutional filter representations, with the continuous time framework of neural ODEs. This allows us to learn the spatial support of the filters during training, as well as model the temporal evolution of feature maps, linking DCNs closely to biological models. We show that DCNs are versatile. Experimentally, we demonstrate their applicability to a standard classification problem, where they allow for parameter reductions and meta-parametrization. We illustrate the biological plausibility of the scale distributions learned by DCNs and explore their performance in a pattern completion task, which is inspired by models from computational neuroscience. Finally, we suggest that the continuous representations learned by DCNs may enable computationally efficient implementations.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "AbjIHxP2p0i", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1599/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summarize what the paper claims to contribute. \nThe paper develops a spatio-temporal network that is defined in terms of continuous spatial functions and continuous temporal dynamics. The approach is meant to bring deep networks closer to biological neural models. The model is tested on CIFAR-10, and on a variation of CIFAR-10 in which blocks of pixels are blacked out. The methodological novelty consists of combining several existing approaches (continuous kernels, kernel-scale learning, and neural ODEs). \n\nList strong and weak points of the paper. \nStrong points:\n-\tI think the motivation is very strong. Deep convolutional networks are increasingly used in brain modelling, but they are somewhat disconnected from earlier computational neuroscience in ways that this paper tries to address. \n-\tThe pattern completion test with blacked-out pixels is a nice way to employ spatiotemporal dynamics in image recognition, and the results are promising. \n-\tIt is very interesting that the distribution of learned scales reflects the distribution of receptive-field sizes in primary visual cortex. \n\nWeak points: \n-\tThe feature-map evolution results in Fig. 4 are interesting, and a few more examples are given in the appendix, but it would also be nice to see mean +/- SD dynamics across many images, to complement Figure 3C more thoroughly. \n-\tThe model doesn’t outperform controls on CIFAR-10, although it does perform moderately better with blacked-out blocks of 6x6 pixels or more. \n-\tPerformance of the model and baselines on CIFAR-10 is not strong, which raises the question of how compatible the approach is with higher performance. \n-\tThe choice of CIFAR-10 as a test of the approach does not seem to be well motivated. A task with a temporal component might give the dynamic parts of the model more to do. \n\nClearly state your recommendation (accept or reject) with one or two key reasons for this choice. \nI recommend to accept the paper. There is a growing body of work that compares deep CNNs to biological neural networks, and the conventions of discrete time and space in CNNs unfortunately distance this work somewhat from much previous computational neuroscience. I think the paper helps to close this gap. \n\nAsk questions you would like answered by the authors to help you clarify your understanding of the paper and provide the additional evidence you need to be confident in your assessment. \n-\tCould you please expand on the motivation for continuous space with respect to biological vision? Ultimately vision is based on discrete photoreceptors, and generally on populations of discrete cells. Relatedly, the motivation for continuous time is more obvious, but I think it would also help to comment on spikes in this context. \n-\tThe dynamic equation in Eq. 2 seems to be autonomous. How does input affect h? \n-\tIn Section 3.2, the learned scales are shown to increase with network depth. This is compared with biological receptive fields, which grow through the visual hierarchy. But I had understood the scale to correspond to the kernel size rather than the receptive field size (which grows in deep networks even if all the kernels are the same size). Does sigma correspond to kernel or receptive field size? \n\nProvide additional feedback with the aim to improve the paper. \n-\tAnalytic tractability is mentioned on page 1 as an advantage of some continuous models in neuroscience, and it seems to be implied at that point that such benefits are sought in the paper, but it doesn’t seem that the approach ultimately offers much hope in this sense. If there is some potential here, please expand. \n-\tI didn’t understand “… time (or network depth)” on pg. 4.  \n-\tAlso on pg. 4, the sampling domain of the filter is given, but not the sampling frequency. \n-\tI found section A.1 relatively hard to follow. In particular, Eq. 5 seems to be meant to motivate the filter family, but I didn’t follow the argument (or maybe I missed the point entirely). Also, aside from general interest I didn’t understand how the paragraph that contains Eq. 6 related to the rest of the paper. \n-\tIt wouldn’t hurt to define DOPRI. \n-\tI didn’t follow the last paragraph of A.2. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice bridge between deep learning and computational neuroscience.  ", "review": "Summarize what the paper claims to contribute. \nThe paper develops a spatio-temporal network that is defined in terms of continuous spatial functions and continuous temporal dynamics. The approach is meant to bring deep networks closer to biological neural models. The model is tested on CIFAR-10, and on a variation of CIFAR-10 in which blocks of pixels are blacked out. The methodological novelty consists of combining several existing approaches (continuous kernels, kernel-scale learning, and neural ODEs). \n\nList strong and weak points of the paper. \nStrong points:\n-\tI think the motivation is very strong. Deep convolutional networks are increasingly used in brain modelling, but they are somewhat disconnected from earlier computational neuroscience in ways that this paper tries to address. \n-\tThe pattern completion test with blacked-out pixels is a nice way to employ spatiotemporal dynamics in image recognition, and the results are promising. \n-\tIt is very interesting that the distribution of learned scales reflects the distribution of receptive-field sizes in primary visual cortex. \n\nWeak points: \n-\tThe feature-map evolution results in Fig. 4 are interesting, and a few more examples are given in the appendix, but it would also be nice to see mean +/- SD dynamics across many images, to complement Figure 3C more thoroughly. \n-\tThe model doesn’t outperform controls on CIFAR-10, although it does perform moderately better with blacked-out blocks of 6x6 pixels or more. \n-\tPerformance of the model and baselines on CIFAR-10 is not strong, which raises the question of how compatible the approach is with higher performance. \n-\tThe choice of CIFAR-10 as a test of the approach does not seem to be well motivated. A task with a temporal component might give the dynamic parts of the model more to do. \n\nClearly state your recommendation (accept or reject) with one or two key reasons for this choice. \nI recommend to accept the paper. There is a growing body of work that compares deep CNNs to biological neural networks, and the conventions of discrete time and space in CNNs unfortunately distance this work somewhat from much previous computational neuroscience. I think the paper helps to close this gap. \n\nAsk questions you would like answered by the authors to help you clarify your understanding of the paper and provide the additional evidence you need to be confident in your assessment. \n-\tCould you please expand on the motivation for continuous space with respect to biological vision? Ultimately vision is based on discrete photoreceptors, and generally on populations of discrete cells. Relatedly, the motivation for continuous time is more obvious, but I think it would also help to comment on spikes in this context. \n-\tThe dynamic equation in Eq. 2 seems to be autonomous. How does input affect h? \n-\tIn Section 3.2, the learned scales are shown to increase with network depth. This is compared with biological receptive fields, which grow through the visual hierarchy. But I had understood the scale to correspond to the kernel size rather than the receptive field size (which grows in deep networks even if all the kernels are the same size). Does sigma correspond to kernel or receptive field size? \n\nProvide additional feedback with the aim to improve the paper. \n-\tAnalytic tractability is mentioned on page 1 as an advantage of some continuous models in neuroscience, and it seems to be implied at that point that such benefits are sought in the paper, but it doesn’t seem that the approach ultimately offers much hope in this sense. If there is some potential here, please expand. \n-\tI didn’t understand “… time (or network depth)” on pg. 4.  \n-\tAlso on pg. 4, the sampling domain of the filter is given, but not the sampling frequency. \n-\tI found section A.1 relatively hard to follow. In particular, Eq. 5 seems to be meant to motivate the filter family, but I didn’t follow the argument (or maybe I missed the point entirely). Also, aside from general interest I didn’t understand how the paragraph that contains Eq. 6 related to the rest of the paper. \n-\tIt wouldn’t hurt to define DOPRI. \n-\tI didn’t follow the last paragraph of A.2. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604120044986}, {"id": "cz3g-4wZF5_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1599/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Review\n\n\n### Summary\n\nThe authors define continuous deep networks by expressing 2D-convolutional filters as a linear combination of Gaussian function and its derivatives. By combining this description with the previously proposed neural ODE framework they obtain a spatio-temporally continuous description of a convolutional neural network.\n\nThere are 3 main contributions:  \n\n1. they are able to estimate the support width of the filters and they show that  it increases with the network depth as observed in the visual cortex  \n2. they show that their network performs as well as alternative non-continuous neural networks on CIFAR-10 while having less parameters  \n3. they exploit the temporal dynamics to resolve a pattern completion task\n\nOverall, I think the work is good but I am not as enthusiastic as the authors about the importance of the work for neuroscience and machine learning.\n\n### Strengths\n\n* The paper is well written and easy to understand. The goals and contributions of the work are clearly stated.\n* The quantitative results are marginally good.\n* The qualitative results (filter support width, pattern completion and contrast robustness) are interesting and relevant to neuroscience\n\n\n### Weaknesses\n\n* I fail to understand how the work could be relevant to neuroscience beyond what is presented here. What is so important about using spatially continuous filters that couldn't be done with discrete filters ?\n* The pattern completion task is not fully conducted. It would be great to reconstruct the missing part of the input.\n* The increase of filter support width with network depth correlates with what is known for the visual cortex but I fail to understand how it could be relevant to current work in experimental neuroscience. What is the benefit of learning continuously changing support for machine learning ? About the relation to biology, the increase in size might be more related to the specific task on which the network is trained than to what is observed in the visual cortex.\n* The observed contrast robustness is not compared to other neural networks nor discussed in the light of experimental neuroscience observations.\n\n### Minor comments\n\n* The text in the figures is way too small. It should be the same size as the main text.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review DCN", "review": "## Review\n\n\n### Summary\n\nThe authors define continuous deep networks by expressing 2D-convolutional filters as a linear combination of Gaussian function and its derivatives. By combining this description with the previously proposed neural ODE framework they obtain a spatio-temporally continuous description of a convolutional neural network.\n\nThere are 3 main contributions:  \n\n1. they are able to estimate the support width of the filters and they show that  it increases with the network depth as observed in the visual cortex  \n2. they show that their network performs as well as alternative non-continuous neural networks on CIFAR-10 while having less parameters  \n3. they exploit the temporal dynamics to resolve a pattern completion task\n\nOverall, I think the work is good but I am not as enthusiastic as the authors about the importance of the work for neuroscience and machine learning.\n\n### Strengths\n\n* The paper is well written and easy to understand. The goals and contributions of the work are clearly stated.\n* The quantitative results are marginally good.\n* The qualitative results (filter support width, pattern completion and contrast robustness) are interesting and relevant to neuroscience\n\n\n### Weaknesses\n\n* I fail to understand how the work could be relevant to neuroscience beyond what is presented here. What is so important about using spatially continuous filters that couldn't be done with discrete filters ?\n* The pattern completion task is not fully conducted. It would be great to reconstruct the missing part of the input.\n* The increase of filter support width with network depth correlates with what is known for the visual cortex but I fail to understand how it could be relevant to current work in experimental neuroscience. What is the benefit of learning continuously changing support for machine learning ? About the relation to biology, the increase in size might be more related to the specific task on which the network is trained than to what is observed in the visual cortex.\n* The observed contrast robustness is not compared to other neural networks nor discussed in the light of experimental neuroscience observations.\n\n### Minor comments\n\n* The text in the figures is way too small. It should be the same size as the main text.\n\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604053185743}, {"id": "2lskEF1iXJ9", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1599/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe authors propose a hierarchical model of neural ODEs, which they fit to CIFAR10. They find performance on par with ResNets, and include qualitative analyses on the filters learned by the models, their ability to fill in occluded features, and robustness to contrast at test time.\n\n\nStrengths:\n\nThe filter parameterization is interesting. I can imagine this improving sample efficiency in certain contexts — perhaps the authors should seek out those kinds of tasks to complement their CIFAR results?\n\nThe discussion does a nice job of explaining the issues that neural ODEs have when scaling to large image datasets.\n\n\nWeaknesses:\n\nYou spend time discussing spatio-temporal receptive fields throughout. Why? Your models are applied to 2d images.\n\nThe authors are missing a huge literature on (a) recurrent convolutional networks, and (b) using these networks to simulate classical vs. extra classical receptive field effects.\n\nResNet-Blocks is the original ResNet? Maybe change \"blocks\" to a citation? Or V1/V2 depending on which implementation it is (unclear from the text).\n\nThere's essentially no difference between the performance of any of the models tested. Is it possible to scale to ImageNet? It is important to show that the proposed method does *something* different than the standard ResNet. The authors attempted to add some qualitative experiments towards this goal in Fig 4, but those results are not very convincing. I think to show filling-in you'd want to show reconstruction in RGB space.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "Summary:\n\nThe authors propose a hierarchical model of neural ODEs, which they fit to CIFAR10. They find performance on par with ResNets, and include qualitative analyses on the filters learned by the models, their ability to fill in occluded features, and robustness to contrast at test time.\n\n\nStrengths:\n\nThe filter parameterization is interesting. I can imagine this improving sample efficiency in certain contexts — perhaps the authors should seek out those kinds of tasks to complement their CIFAR results?\n\nThe discussion does a nice job of explaining the issues that neural ODEs have when scaling to large image datasets.\n\n\nWeaknesses:\n\nYou spend time discussing spatio-temporal receptive fields throughout. Why? Your models are applied to 2d images.\n\nThe authors are missing a huge literature on (a) recurrent convolutional networks, and (b) using these networks to simulate classical vs. extra classical receptive field effects.\n\nResNet-Blocks is the original ResNet? Maybe change \"blocks\" to a citation? Or V1/V2 depending on which implementation it is (unclear from the text).\n\nThere's essentially no difference between the performance of any of the models tested. Is it possible to scale to ImageNet? It is important to show that the proposed method does *something* different than the standard ResNet. The authors attempted to add some qualitative experiments towards this goal in Fig 4, but those results are not very convincing. I think to show filling-in you'd want to show reconstruction in RGB space.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603937140005}], "openreview_url": "https://openreview.net/forum?id=L5b6jUonKFB", "arxiv_id": "2402.01557", "paper_pdf": "papers/L5b6jUonKFB.pdf", "paper_pdf_sha256": "835884eaa3a8bbdef28749a8e3328707813d164ce29c262e8118fc165338cf18", "paper_pdf_bytes": 2603410, "paper_pdf_source": "openreview", "code_url": "https://github.com/ntomen/Deep-Continuous-Networks", "code_repository": "ntomen/Deep-Continuous-Networks", "code_commit": "81cd9b651179aa70a495feccf883526cc28f78bb", "code_archive": "repos/L5b6jUonKFB.zip", "code_archive_sha256": "3023d0264a9830b1d8ad20a64d74ba9b5d28fc1c5eb002fce44c3f57439d6db2", "code_archive_bytes": 123038, "code_file_count": 42, "code_extensions": {".py": 41, ".sh": 1}, "github_disk_usage_kb": 48, "github_languages": {"Python": 364470, "Shell": 135}, "github_archived": false, "github_pushed_at": "2021-09-27T16:51:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-continuous-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rylkma4twr", "year": 2020, "status": "rejected", "title": "Min-Max Optimization without Gradients: Convergence and Applications to Adversarial ML", "authors": ["Sijia Liu", "Songtao Lu", "Xiangyi Chen", "Yao Feng", "Kaidi Xu", "Abdullah Al-Dujaili", "Minyi Hong", "Una-May Obelilly"], "authorids": ["sijia.liu@ibm.com", "songtao@ibm.com", "chen5719@umn.edu", "feng-y16@mails.tsinghua.edu.cn", "xu.kaid@husky.neu.edu", "aldujail@mit.edu", "mhong@umn.edu", "unamay@csail.mit.edu"], "authors_source": "OpenReview API", "abstract": "In this paper, we study the problem of constrained robust (min-max) optimization ina black-box setting, where the desired optimizer cannot access the gradients of the objective function but may query its values. We present a principled optimization framework, integrating a zeroth-order (ZO) gradient estimator with an alternating projected stochastic gradient descent-ascent method, where the former only requires a small number of function queries and the later needs just one-step descent/ascent update. We show that the proposed framework, referred to as ZO-Min-Max, has a sub-linear convergence rate under mild conditions and scales gracefully with problem size. From an application side, we explore a promising connection between black-box min-max optimization and black-box evasion and poisoning attacks in adversarial machine learning (ML). Our empirical evaluations on these use cases demonstrate the effectiveness of our approach and its scalability to dimensions that prohibit using recent black-box solvers.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HkxtF2y29H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper433/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers zeroth-order method for min-max optimization (ZO-MIN-MAX) in two cases: one-sided black box (for outer minimization) and two-sided black box (for both inner maximization and outer minimization). Convergence analysis is carefully provided to show that ZO-MIN-MAX converges to a neighborhood of stationary points. Then, the authors empirically compare several methods on \n1) adversarial attack on ImageNet with deep networks, and \n2) black-box poisoning attack on logistic regression. The results show that ZO-MIN-MAX can provide satisfactory performance on these tasks.\n\nIn general a good paper with dense content, clear organization and writing. However, the experiment part does not seem truly convincing. \n1.\tWhat is the relationship between Eqn.(13) and the proposed ZO-MIN-MAX? It seem that in the experiment you compare using this loss ( Eqn.(13) ) against finite-sum loss, but both with ZO-MIN-MAX algorithm? In figure 1 and 2, I don’t see a competing method. So the point here is that the loss Eqn.(13)  is better, but not the proposed algorithm? I think you should compare different optimization algorithm under same loss, e.g. something like Eqn.(13)+ZO-MIN-MAX vs. Eqn.(13)+FO-MIN-MAX. This is not evident to show that ZO-MIN-MAX is better than other zero-th order methods.\n2.\tI would suggest comparing to more zeroth-order methods in the experiment.\n\nFrom the experiments I cannot tell whether ZO-MIN-MAX is good enough compared with other methods", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper considers zeroth-order method for min-max optimization (ZO-MIN-MAX) in two cases: one-sided black box (for outer minimization) and two-sided black box (for both inner maximization and outer minimization). Convergence analysis is carefully provided to show that ZO-MIN-MAX converges to a neighborhood of stationary points. Then, the authors empirically compare several methods on \n1) adversarial attack on ImageNet with deep networks, and \n2) black-box poisoning attack on logistic regression. The results show that ZO-MIN-MAX can provide satisfactory performance on these tasks.\n\nIn general a good paper with dense content, clear organization and writing. However, the experiment part does not seem truly convincing. \n1.\tWhat is the relationship between Eqn.(13) and the proposed ZO-MIN-MAX? It seem that in the experiment you compare using this loss ( Eqn.(13) ) against finite-sum loss, but both with ZO-MIN-MAX algorithm? In figure 1 and 2, I don’t see a competing method. So the point here is that the loss Eqn.(13)  is better, but not the proposed algorithm? I think you should compare different optimization algorithm under same loss, e.g. something like Eqn.(13)+ZO-MIN-MAX vs. Eqn.(13)+FO-MIN-MAX. This is not evident to show that ZO-MIN-MAX is better than other zero-th order methods.\n2.\tI would suggest comparing to more zeroth-order methods in the experiment.\n\nFrom the experiments I cannot tell whether ZO-MIN-MAX is good enough compared with other methods"}, "tcdate": 1572760704707}, {"id": "SylOV5CfqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper433/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an algorithm for performing min-max optimisation without gradients and analyses its convergence. The algorithm is evaluated for the min-max problems that arise in the context of adversarial attacks. The presented algorithm is a natural application of a zeroth-order gradient estimator and the authors also prove that the algorithm has a sublinear convergence rate (in a specific sense). \n\nConsidering that the algorithm merely applies the zeroth-order gradient estimator to min-max problems, the algorithm itself only makes up a somewhat novel contribution. However, to the best of my knowledge, it has not been used in this context before and personally I find the algorithm quite appealing. In fact, due to its simplicity it is essentially something that anyone could implement from scratch. \n\nPerhaps a more important contribution is that the authors provide a fairly extensive convergence analysis, which is an important tool in analysing the algorithm and its properties. Unfortunately, it is not trivial to understand the presented convergence results and their practical implications (if any). For instance, equation (10), which is arguably one of the key equations in the paper, contains variables zeta, nu and P1, all of which depend on a number of other variables in a fairly complicated manner. The expression in (10) also contains terms that do not depend on T and it is not obvious how large these terms might be in practice (in the event that the assumptions are at least approximately true in a local region). Even though I am somewhat sceptical to the practical relevance of this convergence analysis, I recognise that it is an interesting and fascinating achievement that the authors have managed to provide a convergence analysis of an algorithm which is based on black-box min-max optimisation. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper presents an algorithm for performing min-max optimisation without gradients and analyses its convergence. The algorithm is evaluated for the min-max problems that arise in the context of adversarial attacks. The presented algorithm is a natural application of a zeroth-order gradient estimator and the authors also prove that the algorithm has a sublinear convergence rate (in a specific sense). \n\nConsidering that the algorithm merely applies the zeroth-order gradient estimator to min-max problems, the algorithm itself only makes up a somewhat novel contribution. However, to the best of my knowledge, it has not been used in this context before and personally I find the algorithm quite appealing. In fact, due to its simplicity it is essentially something that anyone could implement from scratch. \n\nPerhaps a more important contribution is that the authors provide a fairly extensive convergence analysis, which is an important tool in analysing the algorithm and its properties. Unfortunately, it is not trivial to understand the presented convergence results and their practical implications (if any). For instance, equation (10), which is arguably one of the key equations in the paper, contains variables zeta, nu and P1, all of which depend on a number of other variables in a fairly complicated manner. The expression in (10) also contains terms that do not depend on T and it is not obvious how large these terms might be in practice (in the event that the assumptions are at least approximately true in a local region). Even though I am somewhat sceptical to the practical relevance of this convergence analysis, I recognise that it is an interesting and fascinating achievement that the authors have managed to provide a convergence analysis of an algorithm which is based on black-box min-max optimisation. \n"}, "tcdate": 1572166192275}, {"id": "SkeqZkz-5S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper433/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors aim to propose new algorithms for min-max optimization problem when the gradients are not available and establish sublinear convergence of the algorithm. I don't think this paper can be accepted for ICLR for the following reasons:\n\n1. For Setting (a) (One-sided black-box), the theory can be established by the same analysis for ZO optimization by optimizing y. Even by a proximal step for y, the analysis is essentially the same as ZO where an estimation of the gradient for x is conducted. \n\n2. The assumptions A1 and A2 are hardly satisfied in ML applications, where the objective is essentially smooth. The authors should at least analyze the case where a sub/super-gradients is available.\n\n3. Also, for most ML problems we have today, I don't find many applications where the gradients are not available, and I thus feel that it is not interesting to consider ZO optimizations.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The authors aim to propose new algorithms for min-max optimization problem when the gradients are not available and establish sublinear convergence of the algorithm. I don't think this paper can be accepted for ICLR for the following reasons:\n\n1. For Setting (a) (One-sided black-box), the theory can be established by the same analysis for ZO optimization by optimizing y. Even by a proximal step for y, the analysis is essentially the same as ZO where an estimation of the gradient for x is conducted. \n\n2. The assumptions A1 and A2 are hardly satisfied in ML applications, where the objective is essentially smooth. The authors should at least analyze the case where a sub/super-gradients is available.\n\n3. Also, for most ML problems we have today, I don't find many applications where the gradients are not available, and I thus feel that it is not interesting to consider ZO optimizations."}, "tcdate": 1572048641876}], "openreview_url": "https://openreview.net/forum?id=rylkma4twr", "arxiv_id": "1909.13806", "paper_pdf": "papers/rylkma4twr.pdf", "paper_pdf_sha256": "3a6ac0973684ffe30c17b4d89023ea6ab561b2eaa07386771e203ee6a2e8dd27", "paper_pdf_bytes": 838837, "paper_pdf_source": "openreview", "code_url": "https://github.com/KaidiXu/ZO-minmax", "code_repository": "KaidiXu/ZO-minmax", "code_commit": "17d1b1183a5e9d07d70875efa33e6f0134f85958", "code_archive": "repos/rylkma4twr.zip", "code_archive_sha256": "6a8ee763ebbfefedaacf239cd28db3d6ab92319c1034407525da4d5a5350c70e", "code_archive_bytes": 64966, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 69, "github_languages": {"Python": 322618}, "github_archived": false, "github_pushed_at": "2020-06-28T19:49:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/min-max-optimization-without-gradients"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkxt8oC9FQ", "year": 2019, "status": "rejected", "title": "Perfect Match: A Simple Method for Learning Representations For Counterfactual Inference With Neural Networks", "authors": ["Patrick Schwab", "Lorenz Linhardt", "Walter Karlen"], "authorids": ["patrick.schwab@hest.ethz.ch", "llorenz@student.ethz.ch", "walter.karlen@hest.ethz.ch"], "authors_source": "OpenReview API", "abstract": "Learning representations for counterfactual inference from observational data is of high practical relevance for many domains, such as healthcare, public policy and economics. Counterfactual inference enables one to answer \"What if...?\" questions, such as \"What would be the outcome if we gave this patient treatment $t_1$?\". However, current methods for training neural networks for counterfactual inference on observational data are either overly complex, limited to settings with only two available treatment options, or both. Here, we present Perfect Match (PM), a method for training neural networks for counterfactual inference that is easy to implement, compatible with any architecture, does not add computational complexity or hyperparameters, and extends to any number of treatments. PM is based on the idea of augmenting samples within a minibatch with their propensity-matched nearest neighbours. Our experiments demonstrate that PM outperforms a number of more complex state-of-the-art methods in inferring counterfactual outcomes across several real-world and semi-synthetic datasets.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HygsMMPq2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper192/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "========= Summary =========\n\nThe authors propose a novel method for counterfactual inference (i.e. individual/heterogeneous treatment effect, as well as average treatment effect) with neural networks. They perform propensity score matching within each minibatch in order to match the covariate distributions during training, which leads to a doubly robust model.\n\nPM is evaluated on several standard semi-synthetic datasets (jobs, IHDP, TCGA) and PM shows state-of-the-art performance on some datasets, and overall looks quite promising. \n\n======= Comments =======\n\nThe paper is well-written, presents a novel method of some interest to the community, and shows quite good performance across a range of relevant benchmarks.\n\nI have one major issue with this work: I don't see why propensity-score matching *within* a minibatch should provide a substantial improvement over propensity-score matching across the dataset (Ho et al 2011). I find the cursory explanation given (\"it ensures that every gradient step is done in a way that is approximately unbiased\") unconvincing, since (a) proper SGD training should be robust to per-batch biases during training (the expected loss is identical for both methods, correct?), and (b) biases should go away in the limit of large batch sizes. If indeed SGD required unbiased *minibatches* then standard minibatch SGD wouldn't work at all.\n\nLooking at the experimental details in the appendix, it appears that the MatchIt package was used to do PSM, rather than a careful comparison under the same conditions. Are the exact matching procedure, PS estimator model, choosing \"one of 6 closest  matches by propensity score\", batch size, etc. the same between your PM implementation and MatchIt? I'd be very curious to see the results of a controlled comparison between Alg S1 and S2 under the same conditions (i.e. run your PM implementation on the whole dataset), and perhaps even some more clever experiments illustrating why matching within a minibatch is important. \n\nAnother hypothesis for why PM is better than PSM is that the matching distribution for PM changes at each epoch (at least due to the randomization among the 6 closest matches). Could it be that the advantage of PM is that it actually provides a randomized rather than constant distribution of matched points?\n\nCan the authors provide more motivation for why PM should outperform PSM? Or some more careful comparison of these methods isolating the benefits of PM? I think a convincing justification and comparison here could change my opinion, as I like the paper otherwise. Thanks!\n\nDetailed Comments:\n\n- There is insufficient explanation of the PM method in the main text. The method is only mentioned in a single sentence buried in the middle of a long paragraph \"In PM, we match every sample within a minibatch...\". This should be made more clear, e.g. by moving Algorithm S1 to the main text.\n- The discussion on Model Selection and the argument for nearest-neighbor PEHE is clever and well-supported by the experiments.\n- In Table 3 and 4, it's not clear which numbers are reported by the original authors and which were replicated by the authors.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review: Interesting paper and impressive results; why novel vs. standard PSM?", "review": "========= Summary =========\n\nThe authors propose a novel method for counterfactual inference (i.e. individual/heterogeneous treatment effect, as well as average treatment effect) with neural networks. They perform propensity score matching within each minibatch in order to match the covariate distributions during training, which leads to a doubly robust model.\n\nPM is evaluated on several standard semi-synthetic datasets (jobs, IHDP, TCGA) and PM shows state-of-the-art performance on some datasets, and overall looks quite promising. \n\n======= Comments =======\n\nThe paper is well-written, presents a novel method of some interest to the community, and shows quite good performance across a range of relevant benchmarks.\n\nI have one major issue with this work: I don't see why propensity-score matching *within* a minibatch should provide a substantial improvement over propensity-score matching across the dataset (Ho et al 2011). I find the cursory explanation given (\"it ensures that every gradient step is done in a way that is approximately unbiased\") unconvincing, since (a) proper SGD training should be robust to per-batch biases during training (the expected loss is identical for both methods, correct?), and (b) biases should go away in the limit of large batch sizes. If indeed SGD required unbiased *minibatches* then standard minibatch SGD wouldn't work at all.\n\nLooking at the experimental details in the appendix, it appears that the MatchIt package was used to do PSM, rather than a careful comparison under the same conditions. Are the exact matching procedure, PS estimator model, choosing \"one of 6 closest  matches by propensity score\", batch size, etc. the same between your PM implementation and MatchIt? I'd be very curious to see the results of a controlled comparison between Alg S1 and S2 under the same conditions (i.e. run your PM implementation on the whole dataset), and perhaps even some more clever experiments illustrating why matching within a minibatch is important. \n\nAnother hypothesis for why PM is better than PSM is that the matching distribution for PM changes at each epoch (at least due to the randomization among the 6 closest matches). Could it be that the advantage of PM is that it actually provides a randomized rather than constant distribution of matched points?\n\nCan the authors provide more motivation for why PM should outperform PSM? Or some more careful comparison of these methods isolating the benefits of PM? I think a convincing justification and comparison here could change my opinion, as I like the paper otherwise. Thanks!\n\nDetailed Comments:\n\n- There is insufficient explanation of the PM method in the main text. The method is only mentioned in a single sentence buried in the middle of a long paragraph \"In PM, we match every sample within a minibatch...\". This should be made more clear, e.g. by moving Algorithm S1 to the main text.\n- The discussion on Model Selection and the argument for nearest-neighbor PEHE is clever and well-supported by the experiments.\n- In Table 3 and 4, it's not clear which numbers are reported by the original authors and which were replicated by the authors.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541202450849}, {"id": "B1e3G21qhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper192/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper proposed to extend TARNET (Shalit et al. 2017), a representation learning approach for counterfactual inference, in the following ways.\n\nFirst, to extend TARNET to multiple treatment setting, k head networks (instead of 2) were constructed following the shared MLP layers, where each head network modeled the outcome of one treatment. This extension seemed quite straightforward.\n\nSecond, during training, for every sample in a minibatch, find its nearest neighbors from all other treatments and add them to the minibatch. The distance was measured by the propensity score, which was defined the probability of a sample being assigned to a treatment group and could be learned by a classification model (such as support vector machine used in this work). Therefore, 1) the augmented minibatch would contain the same number of samples for each treatment group; 2) different treatment group were balanced.\n\nThird, a model selection strategy was proposed by estimating the PEHE using nearest neighbor search.\n\nComments:\nThis paper is well motivated. The key challenges in counterfactual inference is how to adjust for the bias in treatment assignment and the associated discrepancies in the distribution of different treatment groups. \n\nThe main idea of this paper, i.e., augmenting the minibatch through propensity score matching for each sample, is well explained in Section 3. However, it could be better if the introduction of model architecture (in Appendix F) was presented in the method section.\n\nDid the author need to train (k choose 2) SVMs to compute the propensity scores for samples from k treatment groups?\n\nWhen comparing different approaches, as were shown in Table 3, 4 and Figure 3,4, did the author run any statistical test, such as t-test, to confirm the difference between those distributions were significant? The standard deviations of those errors seemed quite large so the difference could be non-significant.\n\nCould the author provide more explanations on why the proposed approach, i.e., minibatch augmentation using propensity score matching, can outperform the TARNET? In TARNET, each sample it only used to update the head network corresponding to the sample's treatment assignment, why would balancing samples in the minibatch can improve the estimation of treatment effect?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple idea, the presentation of the method and experiment results can be improved", "review": "Summary:\nThis paper proposed to extend TARNET (Shalit et al. 2017), a representation learning approach for counterfactual inference, in the following ways.\n\nFirst, to extend TARNET to multiple treatment setting, k head networks (instead of 2) were constructed following the shared MLP layers, where each head network modeled the outcome of one treatment. This extension seemed quite straightforward.\n\nSecond, during training, for every sample in a minibatch, find its nearest neighbors from all other treatments and add them to the minibatch. The distance was measured by the propensity score, which was defined the probability of a sample being assigned to a treatment group and could be learned by a classification model (such as support vector machine used in this work). Therefore, 1) the augmented minibatch would contain the same number of samples for each treatment group; 2) different treatment group were balanced.\n\nThird, a model selection strategy was proposed by estimating the PEHE using nearest neighbor search.\n\nComments:\nThis paper is well motivated. The key challenges in counterfactual inference is how to adjust for the bias in treatment assignment and the associated discrepancies in the distribution of different treatment groups. \n\nThe main idea of this paper, i.e., augmenting the minibatch through propensity score matching for each sample, is well explained in Section 3. However, it could be better if the introduction of model architecture (in Appendix F) was presented in the method section.\n\nDid the author need to train (k choose 2) SVMs to compute the propensity scores for samples from k treatment groups?\n\nWhen comparing different approaches, as were shown in Table 3, 4 and Figure 3,4, did the author run any statistical test, such as t-test, to confirm the difference between those distributions were significant? The standard deviations of those errors seemed quite large so the difference could be non-significant.\n\nCould the author provide more explanations on why the proposed approach, i.e., minibatch augmentation using propensity score matching, can outperform the TARNET? In TARNET, each sample it only used to update the head network corresponding to the sample's treatment assignment, why would balancing samples in the minibatch can improve the estimation of treatment effect?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541172244432}, {"id": "rkxxZ8nwnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper192/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an augmentation of traditional neural network learning to allow for the inference of causal effects. Specifically, they modify the data sampling procedure of SGD during training to use matched samples that are paired via propensity score matching. Experimental results on a number of dataset show that the proposed methodology is comparable to alternative machine learning based causal inference methods. \n\nOverall, I think this is a nice idea. I have two main concerns: \n(1) The use of small batches for matching. Figure 2 does alleviate this concern to an extent, but there is a large literature in statistics and the social sciences on the effect that the quality of matches have on the final causal estimand. It is quite possible that this particular dataset is more amenable to PSM. It is also worth noting that while there is bias reduction shown in figure 2, it is not overwhelming. \n\n(2) The use of propensity scores for matching. One of the insights from the heterogeneous treatment effect literature is that it is not difficult to find cases where the propensity of treatment is identical for two sets of covariates that otherwise do not obey any real balance. This can lead to large biases in the final estimate. Given that PSM is still a relatively widely used practice, I don’t think that its use is a ground for rejection in itself, but given that neural networks are often used to estimate complex causal relations when they are used and this paper is interested in individual treatment effects it is worth noting. \n\nI found the experimental setup to do a very good job in covering large portions of the behavior of the algorithm. The final results are a little underwhelming–the proposed method does not appear to clearly define a new state of the art for the tasks it is applied to–but it is often competitive and the paper presents an interesting idea.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall a good paper", "review": "This paper proposes an augmentation of traditional neural network learning to allow for the inference of causal effects. Specifically, they modify the data sampling procedure of SGD during training to use matched samples that are paired via propensity score matching. Experimental results on a number of dataset show that the proposed methodology is comparable to alternative machine learning based causal inference methods. \n\nOverall, I think this is a nice idea. I have two main concerns: \n(1) The use of small batches for matching. Figure 2 does alleviate this concern to an extent, but there is a large literature in statistics and the social sciences on the effect that the quality of matches have on the final causal estimand. It is quite possible that this particular dataset is more amenable to PSM. It is also worth noting that while there is bias reduction shown in figure 2, it is not overwhelming. \n\n(2) The use of propensity scores for matching. One of the insights from the heterogeneous treatment effect literature is that it is not difficult to find cases where the propensity of treatment is identical for two sets of covariates that otherwise do not obey any real balance. This can lead to large biases in the final estimate. Given that PSM is still a relatively widely used practice, I don’t think that its use is a ground for rejection in itself, but given that neural networks are often used to estimate complex causal relations when they are used and this paper is interested in individual treatment effects it is worth noting. \n\nI found the experimental setup to do a very good job in covering large portions of the behavior of the algorithm. The final results are a little underwhelming–the proposed method does not appear to clearly define a new state of the art for the tasks it is applied to–but it is often competitive and the paper presents an interesting idea.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541027320488}], "openreview_url": "https://openreview.net/forum?id=rkxt8oC9FQ", "arxiv_id": "1810.00656", "paper_pdf": "papers/rkxt8oC9FQ.pdf", "paper_pdf_sha256": "2765311a94ecf1808988ffb56d2dab6c973efdca46bfeea6225cae0114350475", "paper_pdf_bytes": 1009524, "paper_pdf_source": "openreview", "code_url": "https://github.com/d909b/perfect_match", "code_repository": "d909b/perfect_match", "code_commit": "37673e0d8b4029e734a9bf2a4a164e06a8105b84", "code_archive": "repos/rkxt8oC9FQ.zip", "code_archive_sha256": "7ebadd0d62c1e35a1f8909209756c4edd81c3175920e874fdb676dd0d45c4c23", "code_archive_bytes": 398094, "code_file_count": 51, "code_extensions": {".py": 45, ".r": 5, ".sh": 1}, "github_disk_usage_kb": 369, "github_languages": {"Python": 324313, "R": 18317, "Shell": 3828}, "github_archived": false, "github_pushed_at": "2023-03-24T23:48:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/perfect-match-a-simple-method-for-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkVf1AeAZ", "year": 2018, "status": "rejected", "title": "Label Embedding Network: Learning Label Representation for Soft Training of Deep Networks", "authors": ["Xu Sun", "Bingzhen Wei", "Xuancheng Ren", "Shuming Ma"], "authorids": ["xusun@pku.edu.cn", "weibz@pku.edu.cn", "renxc@pku.edu.cn", "shumingma@pku.edu.cn"], "authors_source": "OpenReview API", "abstract": "We propose a method, called Label Embedding Network, which can learn label representation (label embedding) during the training process of deep networks. With the proposed method, the label embedding is adaptively and automatically learned through back propagation. The original one-hot represented loss function is converted into a new loss function with soft distributions, such that the originally unrelated labels have continuous interactions with each other during the training process. As a result, the trained model can achieve substantially higher accuracy and with faster convergence speed. Experimental results based on competitive tasks demonstrate the effectiveness of the proposed method, and the learned label embedding is reasonable and interpretable. The proposed method achieves comparable or even better results than the state-of-the-art systems.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hk7pW6HlM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper438/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to add an embedding layer for labels that constrains normal classifiers in order to find label representations that are semantically consistent. The approach is then experimented on various image and text tasks.\n\nThe description of the model is laborious and hard to follow. Figure 1 helps but is only referred to at the end of the description (at the end of section 2.1), which instead explains each step without the big picture and loses the reader with confusing notation. For instance, it only became clear at the end of the section that E was learned.\n\nOne of the motivations behing the model is to force label representations to be in a semantic space (where two labels with similar meanings would be nearby). The assumption given in the introduction is that softmax would not yield such a representation, but nowhere in the paper this assumption is verified. I believe that using cross-entropy with softmax should also push semantically similar labels to be nearby in the weight space entering the softmax. This should at least be verified and compared appropriately.\n\nAnother motivation of the paper is that targets are given as 1s or 0s while soft targets should work better. I believe this is true, but there is a lot of prior work on these, such as adding a temperature to the softmax, or using distillation, etc. None of these are discussed appropriately in the paper.\n\nSection 2.2 describes a way to compress the label embedding representation, but it is not clear if this is actually used in the experiments. h is never discussed after section 2.2.\n\nExperiments on known datasets are interesting, but none of the results are competitive with current state-of-the-art results (SOTA), despite what is said in Appending D. For instance, one can find SOTA results for CIFAR100 around 16% and for CIFAR10 around 3%. Similarly, one can find SOTA results for IWSLT2015 around 28 BLEU. It can be fine to not be SOTA as long as it is acknowledged and discussed appropriately.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes to add an embedding layer for labels that constrains normal classifiers in order to find label representations that are semantically consistent. The approach is then experimented on various image and text tasks.\n\nThe description of the model is laborious and hard to follow. Figure 1 helps but is only referred to at the end of the description (at the end of section 2.1), which instead explains each step without the big picture and loses the reader with confusing notation. For instance, it only became clear at the end of the section that E was learned.\n\nOne of the motivations behing the model is to force label representations to be in a semantic space (where two labels with similar meanings would be nearby). The assumption given in the introduction is that softmax would not yield such a representation, but nowhere in the paper this assumption is verified. I believe that using cross-entropy with softmax should also push semantically similar labels to be nearby in the weight space entering the softmax. This should at least be verified and compared appropriately.\n\nAnother motivation of the paper is that targets are given as 1s or 0s while soft targets should work better. I believe this is true, but there is a lot of prior work on these, such as adding a temperature to the softmax, or using distillation, etc. None of these are discussed appropriately in the paper.\n\nSection 2.2 describes a way to compress the label embedding representation, but it is not clear if this is actually used in the experiments. h is never discussed after section 2.2.\n\nExperiments on known datasets are interesting, but none of the results are competitive with current state-of-the-art results (SOTA), despite what is said in Appending D. For instance, one can find SOTA results for CIFAR100 around 16% and for CIFAR10 around 3%. Similarly, one can find SOTA results for IWSLT2015 around 28 BLEU. It can be fine to not be SOTA as long as it is acknowledged and discussed appropriately.\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511539130620}, {"id": "r1zEZ9ief", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper438/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method which jointly learns the label embedding (in the form of class similarity) and a classification model. While the motivation of the paper makes sense, the model is not properly justified, and I learned very little after reading the paper.\n\nThere are 5 terms in the proposed objective function. There are also several other parameters associated with them: for example, the label temperature of z_2’’ and and parameter alpha in the second last term etc.\n\nFor all the experiments, the same set of parameters are used, and it is claimed that “the method is robust in our experiment and simply works without fine tuning”. While I agree that a robust and fine-tuning-free model is ideal 1) this has to be justified by experiment. 2) showing the experiment with different parameters will help us understand the role each component plays. This is perhaps more important than improving the baseline method by a few point, especially given that the goal of this work is not to beat the state-of-the-art.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Technique not properly justified; not enough insights can be learned from the work.", "rating": "3: Clear rejection", "review": "The paper proposes a method which jointly learns the label embedding (in the form of class similarity) and a classification model. While the motivation of the paper makes sense, the model is not properly justified, and I learned very little after reading the paper.\n\nThere are 5 terms in the proposed objective function. There are also several other parameters associated with them: for example, the label temperature of z_2’’ and and parameter alpha in the second last term etc.\n\nFor all the experiments, the same set of parameters are used, and it is claimed that “the method is robust in our experiment and simply works without fine tuning”. While I agree that a robust and fine-tuning-free model is ideal 1) this has to be justified by experiment. 2) showing the experiment with different parameters will help us understand the role each component plays. This is perhaps more important than improving the baseline method by a few point, especially given that the goal of this work is not to beat the state-of-the-art.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511919914089}, {"id": "SyZf4f5gM", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper438/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a label embedding network method that learns label embeddings during the training process of deep networks. \nPros: Good empirical results.\nCons:  There is not much technical contribution. The proposed approach is neither well motivated, nor well presented/justified.  The presentation of the paper needs to be improved. \n\n1. Part of the motivation on page 1 does not make sense. In particular, for paragraph 3, if the classification task is just to separate A from B, then (1,0) separation should be better than (0.8, 0.2). \n\n2. Label embedding learning has been investigated in many previous works. The authors however ignored all the existing works on this topic, but enforce label embedding vectors as similarities between labels in Section 2.1 without clear motivation and justification. This assumption is not very natural — though label embeddings can capture semantic information and label correlations, it is unnecessary that label embedding matrix should be m xm and each entry should represent the similarity between a pair of labels.  The paper needs to provide a clear rationale/justification for the assumptions made, while clarifying the difference (and reason) from the literature works. \n\n3. The proposed model is not well explained.  \n(1) By using the objective in eq.(14), how to learn the embeddings E? \n(2) The authors state “In back propagation, the gradient from z2 is kept from propagating to h”.  This makes the learning process quite arbitrary under the objective in eq.(14). \n(3) The label embeddings are not directly used for the classification (H(y, z’_1)), but rather as auxiliary part of the objective.  How to decide the test labels?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "not a well presented/justified  model", "rating": "4: Ok but not good enough - rejection", "review": "This paper proposes a label embedding network method that learns label embeddings during the training process of deep networks. \nPros: Good empirical results.\nCons:  There is not much technical contribution. The proposed approach is neither well motivated, nor well presented/justified.  The presentation of the paper needs to be improved. \n\n1. Part of the motivation on page 1 does not make sense. In particular, for paragraph 3, if the classification task is just to separate A from B, then (1,0) separation should be better than (0.8, 0.2). \n\n2. Label embedding learning has been investigated in many previous works. The authors however ignored all the existing works on this topic, but enforce label embedding vectors as similarities between labels in Section 2.1 without clear motivation and justification. This assumption is not very natural — though label embeddings can capture semantic information and label correlations, it is unnecessary that label embedding matrix should be m xm and each entry should represent the similarity between a pair of labels.  The paper needs to provide a clear rationale/justification for the assumptions made, while clarifying the difference (and reason) from the literature works. \n\n3. The proposed model is not well explained.  \n(1) By using the objective in eq.(14), how to learn the embeddings E? \n(2) The authors state “In back propagation, the gradient from z2 is kept from propagating to h”.  This makes the learning process quite arbitrary under the objective in eq.(14). \n(3) The label embeddings are not directly used for the classification (H(y, z’_1)), but rather as auxiliary part of the objective.  How to decide the test labels?\n", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511822345018}], "openreview_url": "https://openreview.net/forum?id=BkVf1AeAZ", "arxiv_id": "1710.10393", "paper_pdf": "papers/BkVf1AeAZ.pdf", "paper_pdf_sha256": "18e98059f62a50a1e7f0c41b7c5adf5a8b3bca2b3448368b0aa23807b5ef4627", "paper_pdf_bytes": 499393, "paper_pdf_source": "openreview", "code_url": "https://github.com/lancopku/label-embedding-network", "code_repository": "lancopku/label-embedding-network", "code_commit": "ecd9d9b0610ab50b5f0c2630078be72bd4e4591a", "code_archive": "repos/BkVf1AeAZ.zip", "code_archive_sha256": "a8378e4a5539d50605168fde403054f3d22ccb3511b52855cfd62a28da1ac0e6", "code_archive_bytes": 1552845, "code_file_count": 25, "code_extensions": {".py": 22, ".pl": 3}, "github_disk_usage_kb": 1352, "github_languages": {"Python": 209409, "Perl": 21387}, "github_archived": false, "github_pushed_at": "2018-01-29T07:15:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/label-embedding-network-learning-label"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1CoJWNcpmm", "year": 2026, "status": "rejected", "title": "You Do Not Fully Utilize Transformer's Representation Capacity", "authors": ["Gleb Gerasimov", "Yaroslav Aksenov", "Nikita Balagansky", "Viacheslav Sinii", "Daniil Gavrilov"], "authorids": ["~Gleb_Gerasimov1", "~Yaroslav_Aksenov1", "~Nikita_Balagansky3", "~Viacheslav_Sinii1", "~Daniil_Gavrilov1"], "authors_source": "OpenReview API", "abstract": "In contrast to RNNs, which compress their history into a single hidden state, Transformers can attend to all past tokens directly. However, standard Transformers rely solely on the hidden state from the previous layer to represent the entire context. We show that this design choice induces representation collapse and degrades performance. To address this issue, we introduce Layer-Integrated Memory (LIMe), a lightweight extension that leverages existing key–value buffers and learns per-head, per-layer routing weights to integrate representations from all previous layers with negligible overhead. Through extensive experiments—including language modeling, synthetic reasoning benchmarks, and very deep architectures—LIMe consistently achieves faster convergence, lower perplexity per FLOP, and substantial accuracy improvements on synthetic tasks while preserving higher value–vector entropy and improved token separability. Finally, our analysis of the learned routing weights reveals systematic reuse of both local and long-distance features, demonstrating how LIMe mitigates collapse, unlocks richer representations without increasing hidden-state size, and points to promising directions for future research.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "2BlHNIBl7g", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24293/Reviewer_Ax1G"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper suggests adding a weighted average after the standard key projection. The average is taken over all the key representations of the current token in the current layer and head as well as the previous ones (over $i * h$ vectors in the $i$-th layer with a model having $h$ kv heads). The same is done for the values (but not the queries). The coefficient of weighted average is shared between keys and values. Results show improvement over baseline (as well as DenseFormer and HyperConnections) on downstream tasks. Particularly, there is a signficant boost in accuracy over Arithmetic Expression Task which is attributed to the ability to store more information needed for reasoning. Additionally the authors show that the representation remains linearly separable even in later layers which is not true about the baseline.", "review_text": "The paper suggests adding a weighted average after the standard key projection. The average is taken over all the key representations of the current token in the current layer and head as well as the previous ones (over $i * h$ vectors in the $i$-th layer with a model having $h$ kv heads). The same is done for the values (but not the queries). The coefficient of weighted average is shared between keys and values. Results show improvement over baseline (as well as DenseFormer and HyperConnections) on downstream tasks. Particularly, there is a signficant boost in accuracy over Arithmetic Expression Task which is attributed to the ability to store more information needed for reasoning. Additionally the authors show that the representation remains linearly separable even in later layers which is not true about the baseline.", "strengths": "While the method shares similarity with existing methods such as DenseFormer, the correct placement of weighted averages is important and in addition to superior performance on the experiments, yields side-benefits such as the ability to re-use the KV cache. The authors report additional investigative results such as the analysis done on the learned router weights.", "weaknesses": "In Section 5.1, it would be very helpful to have the random baseline for each task. In particular, that results that are reported for several of tasks seem near-chance (e.g. WiC). There is also no confidence intervals reported which makes it very hard to determine the significance of the improvements. Overall this makes me question the efficacy of the method in general language modeling.\n\nIt is confusing to refer to LLaMA in Table 1. Based on my understanding, this is only a model with the same base architecture as LLaMA models where as a LLaMa baseline suggests the pre-trained models. I strongly suggest to make this clear since based on my understanding you are training everything from scratch.\n\nI have asked additional questions below. Overall, I am uncertain about the intepretation of the provided results and whether they can currently clearly establish the effectiveness of the proposed method.", "questions": "1. When doing value classification (e.g. in Fig. 2b) is the rest of the model frozen?\n\n2. Did you consider using a per-dimension (instead of per-head) weighted average? Was there any difference in performance? Alternatively, is it important to average across heads or is it enough to average over the same head across different layers? \n\n3. DenseFormer does a similar mixing as the proposed method. Still, the results for DenseFormer are sometimes even worse than the baseline. Also, Denseformer paper reports reasonable improvements over the baseline. Why similar consistent improvements are not observed in these new set of experiments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper suggests adding a weighted average after the standard key projection. The average is taken over all the key representations of the current token in the current layer and head as well as the previous ones (over $i * h$ vectors in the $i$-th layer with a model having $h$ kv heads). The same is done for the values (but not the queries). The coefficient of weighted average is shared between keys and values. Results show improvement over baseline (as well as DenseFormer and HyperConnections) on downstream tasks. Particularly, there is a signficant boost in accuracy over Arithmetic Expression Task which is attributed to the ability to store more information needed for reasoning. Additionally the authors show that the representation remains linearly separable even in later layers which is not true about the baseline.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "While the method shares similarity with existing methods such as DenseFormer, the correct placement of weighted averages is important and in addition to superior performance on the experiments, yields side-benefits such as the ability to re-use the KV cache. The authors report additional investigative results such as the analysis done on the learned router weights.", "weaknesses": "In Section 5.1, it would be very helpful to have the random baseline for each task. In particular, that results that are reported for several of tasks seem near-chance (e.g. WiC). There is also no confidence intervals reported which makes it very hard to determine the significance of the improvements. Overall this makes me question the efficacy of the method in general language modeling.\n\nIt is confusing to refer to LLaMA in Table 1. Based on my understanding, this is only a model with the same base architecture as LLaMA models where as a LLaMa baseline suggests the pre-trained models. I strongly suggest to make this clear since based on my understanding you are training everything from scratch.\n\nI have asked additional questions below. Overall, I am uncertain about the intepretation of the provided results and whether they can currently clearly establish the effectiveness of the proposed method.", "questions": "1. When doing value classification (e.g. in Fig. 2b) is the rest of the model frozen?\n\n2. Did you consider using a per-dimension (instead of per-head) weighted average? Was there any difference in performance? Alternatively, is it important to average across heads or is it enough to average over the same head across different layers? \n\n3. DenseFormer does a similar mixing as the proposed method. Still, the results for DenseFormer are sometimes even worse than the baseline. Also, Denseformer paper reports reasonable improvements over the baseline. Why similar consistent improvements are not observed in these new set of experiments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762216911680}, {"id": "0XAZPOMqvX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24293/Reviewer_vzV7"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes Layer-Integrated Memory (LIMe), which allows each attention head to access Key-Value representations from all previous layers through learned routing weights.", "review_text": "This paper proposes Layer-Integrated Memory (LIMe), which allows each attention head to access Key-Value representations from all previous layers through learned routing weights.", "strengths": "Comprehensive experimental design: The evaluation spans multiple dimensions: language modeling perplexity, mathematical reasoning on GSM8K, and synthetic tasks with controlled difficulty levels. The representation collapse analysis combines entropy measurements, linear separability tests, and grammatical probing to validate the core hypothesis from different angles. The routing weight analysis provides interpretability by revealing which layer representations the model prefers to access. This is the most lovely part of this paper.", "weaknesses": "1. Limited novelty over prior work. The core mechanism of using learned weights to aggregate multi-layer representations appears in Transparent Attention (Bapna et al., EMNLP 2018), which uses trainable softmax-normalized weights to combine encoder layer outputs in NMT decoder cross-attention. The mathematical formulation resembles that prior work, with the main difference being application to decoder-only self-attention. More recently, Hyper-Connections (Zhu et al., Sept 2024) addresses representation collapse through multi-stream connections with learned routing, sharing similar motivation. The paper does not clearly articulate what architectural insight LIMe provides beyond adapting these known techniques to decoder-only models with efficient KV buffer reuse.\n2. Unclear computational cost analysis. The paper claims \"negligible overhead\" yet mentions O(L**2) routing complexity in limitations. For a 64-layer model, each layer must route over 64 previous layer KV pairs, but the paper does not provide memory bandwidth analysis for this case. The pipeline parallelism overhead of 7.8% contradicts the \"negligible\" claim for production scenarios.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Layer-Integrated Memory (LIMe), which allows each attention head to access Key-Value representations from all previous layers through learned routing weights.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "Comprehensive experimental design: The evaluation spans multiple dimensions: language modeling perplexity, mathematical reasoning on GSM8K, and synthetic tasks with controlled difficulty levels. The representation collapse analysis combines entropy measurements, linear separability tests, and grammatical probing to validate the core hypothesis from different angles. The routing weight analysis provides interpretability by revealing which layer representations the model prefers to access. This is the most lovely part of this paper.", "weaknesses": "1. Limited novelty over prior work. The core mechanism of using learned weights to aggregate multi-layer representations appears in Transparent Attention (Bapna et al., EMNLP 2018), which uses trainable softmax-normalized weights to combine encoder layer outputs in NMT decoder cross-attention. The mathematical formulation resembles that prior work, with the main difference being application to decoder-only self-attention. More recently, Hyper-Connections (Zhu et al., Sept 2024) addresses representation collapse through multi-stream connections with learned routing, sharing similar motivation. The paper does not clearly articulate what architectural insight LIMe provides beyond adapting these known techniques to decoder-only models with efficient KV buffer reuse.\n2. Unclear computational cost analysis. The paper claims \"negligible overhead\" yet mentions O(L**2) routing complexity in limitations. For a 64-layer model, each layer must route over 64 previous layer KV pairs, but the paper does not provide memory bandwidth analysis for this case. The pipeline parallelism overhead of 7.8% contradicts the \"negligible\" claim for production scenarios.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762155233863}, {"id": "SupanSNIuT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24293/Reviewer_nGJ6"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper starts from the observation that in standard transformer networks, there is a single residual stream, meaning that the representations from all the previous layers are compressed into a single hidden state. This single hidden state is then used as the input of the next layer. This can lead to *representation collapse*, which is the phenomenon where different tokens become undistinguishable. Hence, this paper propose a new mechanism to address this issue, called LIMe. The idea is that each layer can attend to the representations of *all* previous layers, instead of just the immediante previous one. In practice, this is done by modifying the way keys and values are computed. Instead of just using the keys and values computed from the input of the current layer, the keys and values from all previous layers are linearly combined, using trainable weights. Said otherwise, the keys and values of used in the attention of layer L are obtained by doing a linear combination of the keys and values of all the heads of the previous layers. The weight of this linear combination are fixed trainable parameters.\n\nThe proposed method is then empirically evaluated on different language modeling tasks. First, a LLaMa like model, with 1B parameters is trained on 50B tokens, and evaluated on downstream NLP tasks such as QNLI, WiC or ARC (easy/challenge). Here the experiments show that LIMe obtain better performance than the standard transformer architecture, as well as other approaches such as DenseFormer or HyperConnections. Then the model is compared to the standard transfomer on GSM8k or synthetic tasks such as arithmetic expression evaluation, again showing that LIMe performs better than the baseline. There are also ablations studying the *representation collapse* showing that LIMe is less prone to representation collapse than standard transformers.", "review_text": "This paper starts from the observation that in standard transformer networks, there is a single residual stream, meaning that the representations from all the previous layers are compressed into a single hidden state. This single hidden state is then used as the input of the next layer. This can lead to *representation collapse*, which is the phenomenon where different tokens become undistinguishable. Hence, this paper propose a new mechanism to address this issue, called LIMe. The idea is that each layer can attend to the representations of *all* previous layers, instead of just the immediante previous one. In practice, this is done by modifying the way keys and values are computed. Instead of just using the keys and values computed from the input of the current layer, the keys and values from all previous layers are linearly combined, using trainable weights. Said otherwise, the keys and values of used in the attention of layer L are obtained by doing a linear combination of the keys and values of all the heads of the previous layers. The weight of this linear combination are fixed trainable parameters.\n\nThe proposed method is then empirically evaluated on different language modeling tasks. First, a LLaMa like model, with 1B parameters is trained on 50B tokens, and evaluated on downstream NLP tasks such as QNLI, WiC or ARC (easy/challenge). Here the experiments show that LIMe obtain better performance than the standard transformer architecture, as well as other approaches such as DenseFormer or HyperConnections. Then the model is compared to the standard transfomer on GSM8k or synthetic tasks such as arithmetic expression evaluation, again showing that LIMe performs better than the baseline. There are also ablations studying the *representation collapse* showing that LIMe is less prone to representation collapse than standard transformers.", "strengths": "I am a bit of the fence regarding this paper.\n\nIn terms of strengths, I believe that the proposed idea in the paper is simple and elegant. The paper is clearly written and easy to follow. The experimental evaluations are convincing.", "weaknesses": "My main concern with the paper is its relation to previous work, and especially its significance with respect to these.\n\nFirst, I believe that the paper does not make a great job discussing the difference with previous work such as DenseFormer, or Value Residual Learning. More precisely, I think that the idea of combining the representations from multiple previous layers instead of just using the representation from the previous layer is not new. The contributions of the paper are thus mostly about details of how this idea is implemented in practice, and the paper could do a better job at discussing these. Moreover, I believe that the baseline considered in the paper (DenseFormer, HyperConnection) have multiple variant considered in the original papers, and the details of which one is used are missing. Finally, I am a bit surprised that the baseline (such as DenseFormer) does not seem to improve compared to the standard transformer, which goes against the claim of the original paper.\n\nAnother minor concern is the additional runtime required by the method, as it needs to read significantly more activations from memory compared to the standard transformer. \n\n**Additional references**\n\n*MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections.* Da Xiao, Qingye Meng, Shengping Li, Xingyuan Yuan. 2025.\n\n*Value Residual Learning.* Zhanchao Zhou, Tianyi Wu, Zhiyun Jiang, Fares Obeid, Zhenzhong Lan. 2024\n\n*LAUREL: Learned Augmented Residual Layer.* Gaurav Menghani, Ravi Kumar, Sanjiv Kumar. 2024\n\n*DeepCrossAttention: Supercharging Transformer Residual Connections.* Mike Heddes, Adel Javanmard, Kyriakos Axiotis, Gang Fu, MohammadHossein Bateni, Vahab Mirrokni. 2025", "questions": "Which variant of DenseFormer and HyperConnection did you use?\n\nDid you re-implement the baselines yourself or use existing code?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper starts from the observation that in standard transformer networks, there is a single residual stream, meaning that the representations from all the previous layers are compressed into a single hidden state. This single hidden state is then used as the input of the next layer. This can lead to *representation collapse*, which is the phenomenon where different tokens become undistinguishable. Hence, this paper propose a new mechanism to address this issue, called LIMe. The idea is that each layer can attend to the representations of *all* previous layers, instead of just the immediante previous one. In practice, this is done by modifying the way keys and values are computed. Instead of just using the keys and values computed from the input of the current layer, the keys and values from all previous layers are linearly combined, using trainable weights. Said otherwise, the keys and values of used in the attention of layer L are obtained by doing a linear combination of the keys and values of all the heads of the previous layers. The weight of this linear combination are fixed trainable parameters.\n\nThe proposed method is then empirically evaluated on different language modeling tasks. First, a LLaMa like model, with 1B parameters is trained on 50B tokens, and evaluated on downstream NLP tasks such as QNLI, WiC or ARC (easy/challenge). Here the experiments show that LIMe obtain better performance than the standard transformer architecture, as well as other approaches such as DenseFormer or HyperConnections. Then the model is compared to the standard transfomer on GSM8k or synthetic tasks such as arithmetic expression evaluation, again showing that LIMe performs better than the baseline. There are also ablations studying the *representation collapse* showing that LIMe is less prone to representation collapse than standard transformers.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "I am a bit of the fence regarding this paper.\n\nIn terms of strengths, I believe that the proposed idea in the paper is simple and elegant. The paper is clearly written and easy to follow. The experimental evaluations are convincing.", "weaknesses": "My main concern with the paper is its relation to previous work, and especially its significance with respect to these.\n\nFirst, I believe that the paper does not make a great job discussing the difference with previous work such as DenseFormer, or Value Residual Learning. More precisely, I think that the idea of combining the representations from multiple previous layers instead of just using the representation from the previous layer is not new. The contributions of the paper are thus mostly about details of how this idea is implemented in practice, and the paper could do a better job at discussing these. Moreover, I believe that the baseline considered in the paper (DenseFormer, HyperConnection) have multiple variant considered in the original papers, and the details of which one is used are missing. Finally, I am a bit surprised that the baseline (such as DenseFormer) does not seem to improve compared to the standard transformer, which goes against the claim of the original paper.\n\nAnother minor concern is the additional runtime required by the method, as it needs to read significantly more activations from memory compared to the standard transformer. \n\n**Additional references**\n\n*MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections.* Da Xiao, Qingye Meng, Shengping Li, Xingyuan Yuan. 2025.\n\n*Value Residual Learning.* Zhanchao Zhou, Tianyi Wu, Zhiyun Jiang, Fares Obeid, Zhenzhong Lan. 2024\n\n*LAUREL: Learned Augmented Residual Layer.* Gaurav Menghani, Ravi Kumar, Sanjiv Kumar. 2024\n\n*DeepCrossAttention: Supercharging Transformer Residual Connections.* Mike Heddes, Adel Javanmard, Kyriakos Axiotis, Gang Fu, MohammadHossein Bateni, Vahab Mirrokni. 2025", "questions": "Which variant of DenseFormer and HyperConnection did you use?\n\nDid you re-implement the baselines yourself or use existing code?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761923685771}, {"id": "GVt58opRIr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24293/Reviewer_8E6q"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper identifies \"representation collapse\" as a key weakness in standard Transformer decoders, where the reliance on a single residual stream from the immediately preceding layer forces the model to compress all prior information, leading to a loss of feature diversity in deeper layers. To address this, the authors propose LIMe, a lightweight architectural modification. LIMe allows each attention head at every layer to compute its KV representations by routing and mixing the KV buffers from all preceding layers, not just the current one. This is achieved by learning a per-head, per-layer routing matrix that weights the contributions of past layers.", "review_text": "This paper identifies \"representation collapse\" as a key weakness in standard Transformer decoders, where the reliance on a single residual stream from the immediately preceding layer forces the model to compress all prior information, leading to a loss of feature diversity in deeper layers. To address this, the authors propose LIMe, a lightweight architectural modification. LIMe allows each attention head at every layer to compute its KV representations by routing and mixing the KV buffers from all preceding layers, not just the current one. This is achieved by learning a per-head, per-layer routing matrix that weights the contributions of past layers.", "strengths": "- The primary strength of LIMe is its elegance and low overhead. By reusing existing KV buffers, it adds multi-layer information flow with almost no additional memory and a negligible computational cost (especially when GQA is used). This makes it a very practical and \"drop-in\" friendly modification.\n\n-  The paper does an excellent job of clearly identifying a specific problem (representation collapse) and proposing a solution (LIMe) that directly targets it.", "weaknesses": "- The authors correctly identify in the limitations that the vanilla implementation of the router has an $\\mathcal{O}(L^2)$ asymptotic complexity (where $L$ is the number of layers), as each layer's router must process keys from all $L-1$ previous layers. This is fine for the 16-layer models in the main paper, but it will become a significant computational bottleneck for scaling to very deep models (e.g., $L=100+$). The heuristic ablations in Appendix F (e.g., last-j or first-j) all show worse performance, suggesting a difficult trade-off between performance and scalability.\n- The method's core idea, accessing all previous KV caches, creates a practical implementation challenge for large-scale training. In a standard pipeline parallel setup, this would require significant communication across pipeline stages (GPUs), as later layers would need to fetch KV caches from all earlier GPUs. The authors acknowledge this and their preliminary test shows a ~7.8% latency overhead. This practical hurdle might deter adoption for training SOTA-scale models, as it requires \"non-trivial engineering effort\" to optimize.", "questions": "Please refer to my weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper identifies \"representation collapse\" as a key weakness in standard Transformer decoders, where the reliance on a single residual stream from the immediately preceding layer forces the model to compress all prior information, leading to a loss of feature diversity in deeper layers. To address this, the authors propose LIMe, a lightweight architectural modification. LIMe allows each attention head at every layer to compute its KV representations by routing and mixing the KV buffers from all preceding layers, not just the current one. This is achieved by learning a per-head, per-layer routing matrix that weights the contributions of past layers.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The primary strength of LIMe is its elegance and low overhead. By reusing existing KV buffers, it adds multi-layer information flow with almost no additional memory and a negligible computational cost (especially when GQA is used). This makes it a very practical and \"drop-in\" friendly modification.\n\n-  The paper does an excellent job of clearly identifying a specific problem (representation collapse) and proposing a solution (LIMe) that directly targets it.", "weaknesses": "- The authors correctly identify in the limitations that the vanilla implementation of the router has an $\\mathcal{O}(L^2)$ asymptotic complexity (where $L$ is the number of layers), as each layer's router must process keys from all $L-1$ previous layers. This is fine for the 16-layer models in the main paper, but it will become a significant computational bottleneck for scaling to very deep models (e.g., $L=100+$). The heuristic ablations in Appendix F (e.g., last-j or first-j) all show worse performance, suggesting a difficult trade-off between performance and scalability.\n- The method's core idea, accessing all previous KV caches, creates a practical implementation challenge for large-scale training. In a standard pipeline parallel setup, this would require significant communication across pipeline stages (GPUs), as later layers would need to fetch KV caches from all earlier GPUs. The authors acknowledge this and their preliminary test shows a ~7.8% latency overhead. This practical hurdle might deter adoption for training SOTA-scale models, as it requires \"non-trivial engineering effort\" to optimize.", "questions": "Please refer to my weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761848197754}], "openreview_url": "https://openreview.net/forum?id=1CoJWNcpmm", "arxiv_id": "2502.09245", "paper_pdf": "papers/1CoJWNcpmm.pdf", "paper_pdf_sha256": "35e926d45626558187a1a65f5b193e1642de1915247c336e34d974d284427231", "paper_pdf_bytes": 1462274, "paper_pdf_source": "openreview", "code_url": "https://github.com/corl-team/lime", "code_repository": "corl-team/lime", "code_commit": "424281209236fc4e461d3e32160df26245f6ba4c", "code_archive": "repos/1CoJWNcpmm.zip", "code_archive_sha256": "dae1aec6fe0c6e72bb38cd8edefb0c1e4b824375e3c14f01524a191707e20e40", "code_archive_bytes": 291665, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 280, "github_languages": {"Python": 76088}, "github_archived": false, "github_pushed_at": "2025-05-28T12:07:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/you-do-not-fully-utilize-transformer-s"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VvHVLVUD6m", "year": 2025, "status": "rejected", "title": "On the Mode-Seeking Properties of Langevin Dynamics", "authors": ["Xiwei Cheng", "Kexin Fu", "Farzan Farnia"], "authorids": ["~Xiwei_Cheng2", "~Kexin_Fu2", "~Farzan_Farnia1"], "authors_source": "OpenReview API", "abstract": "The Langevin Dynamics framework, which aims to generate samples from the score function of a probability distribution, is widely used for analyzing and interpreting score-based generative modeling. While the convergence behavior of Langevin Dynamics under unimodal distributions has been extensively studied in the literature, in practice the data distribution could consist of multiple distinct modes. In this work, we investigate Langevin Dynamics in producing samples from multimodal distributions and theoretically study its mode-seeking properties. We prove that under a variety of sub-Gaussian mixtures, Langevin Dynamics is unlikely to find all mixture components within a sub-exponential number of steps in the data dimension. To reduce the mode-seeking tendencies of Langevin Dynamics, we propose Chained Langevin Dynamics, which divides the data vector into patches of constant size and generates every patch sequentially conditioned on the previous patches. We perform a theoretical analysis of Chained Langevin Dynamics by reducing it to sampling from a constant-dimensional distribution. We present the results of several numerical experiments on synthetic and real image datasets, supporting our theoretical results on the iteration complexities of sample generation from mixture distributions using the chained and vanilla Langevin Dynamics.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "FMcrkUXVeS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9430/Reviewer_Egfk"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper investigates the mode-seeking tendency of stochastic gradient Langevin dynamics (SGLD) and proposes Chained Langevin dynamics with the objective of sampling from multi-modal distributions more accurately. The paper first studies the convergence of SGLD on a multi-modal target with a high-variance, low density, background mode, and shows hardness results of sampling from the sub Gaussian mixture in sub-exponential time in the data dimension. Chained Langevin dynamics seek to alleviate this issue by dividing the input vector into patches and applying SGLD to the conditional distributions $p(x^q \\vert x^1, \\dots, x^{q-1})$, only, thus reducing the effective dimension of the sampled variable. The proposed chained Langevin dynamics are then combined with annealing schemes and used to train a score-based generative model on toy-datasets and MNIST image datasets.", "review_text": "This paper investigates the mode-seeking tendency of stochastic gradient Langevin dynamics (SGLD) and proposes Chained Langevin dynamics with the objective of sampling from multi-modal distributions more accurately. The paper first studies the convergence of SGLD on a multi-modal target with a high-variance, low density, background mode, and shows hardness results of sampling from the sub Gaussian mixture in sub-exponential time in the data dimension. Chained Langevin dynamics seek to alleviate this issue by dividing the input vector into patches and applying SGLD to the conditional distributions $p(x^q \\vert x^1, \\dots, x^{q-1})$, only, thus reducing the effective dimension of the sampled variable. The proposed chained Langevin dynamics are then combined with annealing schemes and used to train a score-based generative model on toy-datasets and MNIST image datasets.", "strengths": "- The paper makes theoretical contributions on the mode seeking properties of Langevin dynamics with and without annealing and makes methodological contributions to improve SGLD.\n- The paper investigates an interesting, and to me novel test-case of having a high-variance low-density background mode. Such settings could be useful benchmarks for sampling and generative modelling, particularly for approaches that have scientific applications in mind.\n- The chained Langevin diffusion is a simple, easy to implement concept with potentially high impact in sampling and score-based generative modelling.\n- To me it is quite impressive that the LSTM architecture (somewhat untypical for diffusion models) achieves the translation from initial samples to target samples on the image dataset example. However, I don't fully understand how this works, since annealed Langevin is normally initialised from a Gaussian distribution. Some explanations on how the generative model is applied, here, would be helpful.", "weaknesses": "- To me, the paper appears to be two different papers in one. It is hard to make out a story and it is unclear whether the paper attempts to make contributions in Langevin sampling or score-based generative modelling. I think the paper should make a clearer distinction between sampling from a known target distribution (for which we know the Stein score but don't have samples) and generative modelling (for which we have access to samples but not the Stein score). These two settings are quite different, even though they are mathematically related. This also influences how the theoretical and empirical contributions are perceived as a reader.\n- The theory is supposed to motivate the chained Langevin algorithm. However, I see a gap between the motivating failure case of SGLD and the proposed chained LD, at least in the main text of the paper. Firstly, chained LD is introduced in Algorithm 1 as a training algorithm of a generative model trained on data. The theoretical results, on the other hand, regard a sampler from a fixed target distribution. However, the paper does not explain why vanilla score-based generative models would fail on mixtures of sub-Gaussian distributions in the same way as SGLD, and thus why chained LD is needed for the training of the generative model. Secondly, the theory of chained LD is not fully developed, and I guess the mixture of sub-Gaussians would not satisfy the implied assumptions of m-strongly log concavity in Proposition 1.\n- If the papers main focus is a new sampling methodology, I think it would be good to justify chained LD more thoroughly from a theory point of view, or to make more thorough evaluations of the sampler, comparing it to other adaptions of Langevin sampling and annealing methods. For example, can something be said about the optimal number of patches between $Q=1$ and $Q=d$, or is this a tuning parameter? I am also curious if related strategies to chained Langevin have been developed before. It seems a reasonable approach, as it looks like a mixture of an autoregressive sampler and a Langevin sampler. Putting chained Langevin into context with related works on other variations of vanilla Langevin would be helpful. \n- If the paper proposes a new method for score-based generative modelling, it would be interesting to see how it compares to vanilla diffusion models on standard benchmarks. For example, reporting FID scores of annealed chained LD on standard image datasets like CIFAR10 would typically be expected for a paper at ICLR, even if the scores are not state of the art. Of course it is hard to make improvements, here, since score-based generative models are so optimised, but I could see how slicing the sample into chunks could speed up convergence in some way or another.", "questions": "- I am not familiar with the recent literature on convergence results for Langevin dynamics. I would expect that the difficulty of sampling from multi-modal distributions with unadjusted Langevin algorithms is well-known. While I consider the background noise an insightful novel test-case, how are the theoretical contributions in this paper on the convergence of Langevin algorithms different from existing results on unadjusted Langevin algorithms? \n- What is the trade-off that one needs to make when choosing the number of patches for chained Langevin? Is there an optimal patch number between $Q=1$ and $Q=d$?\n- For the score-based generative model, could you clarify what forward process you used? In diffusion models, one typically picks a forward process that converges to a Gaussian distribution and learns the reverse process from Gaussian samples to data. How did you initialise the reverse diffusion in the image translation experiments show-cased in figure 3, where the diffusion is seemingly not initialised from a Gaussian?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the mode-seeking tendency of stochastic gradient Langevin dynamics (SGLD) and proposes Chained Langevin dynamics with the objective of sampling from multi-modal distributions more accurately. The paper first studies the convergence of SGLD on a multi-modal target with a high-variance, low density, background mode, and shows hardness results of sampling from the sub Gaussian mixture in sub-exponential time in the data dimension. Chained Langevin dynamics seek to alleviate this issue by dividing the input vector into patches and applying SGLD to the conditional distributions $p(x^q \\vert x^1, \\dots, x^{q-1})$, only, thus reducing the effective dimension of the sampled variable. The proposed chained Langevin dynamics are then combined with annealing schemes and used to train a score-based generative model on toy-datasets and MNIST image datasets.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper makes theoretical contributions on the mode seeking properties of Langevin dynamics with and without annealing and makes methodological contributions to improve SGLD.\n- The paper investigates an interesting, and to me novel test-case of having a high-variance low-density background mode. Such settings could be useful benchmarks for sampling and generative modelling, particularly for approaches that have scientific applications in mind.\n- The chained Langevin diffusion is a simple, easy to implement concept with potentially high impact in sampling and score-based generative modelling.\n- To me it is quite impressive that the LSTM architecture (somewhat untypical for diffusion models) achieves the translation from initial samples to target samples on the image dataset example. However, I don't fully understand how this works, since annealed Langevin is normally initialised from a Gaussian distribution. Some explanations on how the generative model is applied, here, would be helpful.", "weaknesses": "- To me, the paper appears to be two different papers in one. It is hard to make out a story and it is unclear whether the paper attempts to make contributions in Langevin sampling or score-based generative modelling. I think the paper should make a clearer distinction between sampling from a known target distribution (for which we know the Stein score but don't have samples) and generative modelling (for which we have access to samples but not the Stein score). These two settings are quite different, even though they are mathematically related. This also influences how the theoretical and empirical contributions are perceived as a reader.\n- The theory is supposed to motivate the chained Langevin algorithm. However, I see a gap between the motivating failure case of SGLD and the proposed chained LD, at least in the main text of the paper. Firstly, chained LD is introduced in Algorithm 1 as a training algorithm of a generative model trained on data. The theoretical results, on the other hand, regard a sampler from a fixed target distribution. However, the paper does not explain why vanilla score-based generative models would fail on mixtures of sub-Gaussian distributions in the same way as SGLD, and thus why chained LD is needed for the training of the generative model. Secondly, the theory of chained LD is not fully developed, and I guess the mixture of sub-Gaussians would not satisfy the implied assumptions of m-strongly log concavity in Proposition 1.\n- If the papers main focus is a new sampling methodology, I think it would be good to justify chained LD more thoroughly from a theory point of view, or to make more thorough evaluations of the sampler, comparing it to other adaptions of Langevin sampling and annealing methods. For example, can something be said about the optimal number of patches between $Q=1$ and $Q=d$, or is this a tuning parameter? I am also curious if related strategies to chained Langevin have been developed before. It seems a reasonable approach, as it looks like a mixture of an autoregressive sampler and a Langevin sampler. Putting chained Langevin into context with related works on other variations of vanilla Langevin would be helpful. \n- If the paper proposes a new method for score-based generative modelling, it would be interesting to see how it compares to vanilla diffusion models on standard benchmarks. For example, reporting FID scores of annealed chained LD on standard image datasets like CIFAR10 would typically be expected for a paper at ICLR, even if the scores are not state of the art. Of course it is hard to make improvements, here, since score-based generative models are so optimised, but I could see how slicing the sample into chunks could speed up convergence in some way or another.", "questions": "- I am not familiar with the recent literature on convergence results for Langevin dynamics. I would expect that the difficulty of sampling from multi-modal distributions with unadjusted Langevin algorithms is well-known. While I consider the background noise an insightful novel test-case, how are the theoretical contributions in this paper on the convergence of Langevin algorithms different from existing results on unadjusted Langevin algorithms? \n- What is the trade-off that one needs to make when choosing the number of patches for chained Langevin? Is there an optimal patch number between $Q=1$ and $Q=d$?\n- For the score-based generative model, could you clarify what forward process you used? In diffusion models, one typically picks a forward process that converges to a Gaussian distribution and learns the reverse process from Gaussian samples to data. How did you initialise the reverse diffusion in the image translation experiments show-cased in figure 3, where the diffusion is seemingly not initialised from a Gaussian?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730671198186}, {"id": "kPCTnV5SCO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9430/Reviewer_7ZLa"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This article examines the behavior of Langevin dynamics (LD), a sampling method commonly used in generative modeling, especially in contexts where data distributions are multimodal (contain multiple clusters or “modes”). The authors highlight a key limitation of traditional Langevin dynamics: its difficulty in sampling from all modes in a multimodal distribution, as it often struggles to transition between modes in high-dimensional settings.\n\nTo address this, the authors propose Chained Langevin Dynamics (Chained-LD), a novel approach that divides the data into smaller patches and samples these sequentially, conditioned on previously generated patches. This method reduces the likelihood of mode-seeking (focusing on only a few modes) by lowering the dimensionality for each sampling step, which in turn accelerates convergence to the correct distribution.\n\nKey contributions and findings of the paper include:\n1. **Theoretical Analysis:** The study provides a rigorous analysis of examples where the traditional Langevin dynamics can fail to capture all modes, especially under Gaussian and sub-Gaussian mixtures.\n2. **Chained Langevin Dynamics:** The paper introduces Chained-LD as a solution and establishes its convergence properties, showing that it can sample all modes in a distribution.\n3. **Empirical Validation:** Experiments on synthetic datasets and real image datasets (such as MNIST) successfully conducted showing how Chained-LD can capture all modes compared to standard LD.", "review_text": "This article examines the behavior of Langevin dynamics (LD), a sampling method commonly used in generative modeling, especially in contexts where data distributions are multimodal (contain multiple clusters or “modes”). The authors highlight a key limitation of traditional Langevin dynamics: its difficulty in sampling from all modes in a multimodal distribution, as it often struggles to transition between modes in high-dimensional settings.\n\nTo address this, the authors propose Chained Langevin Dynamics (Chained-LD), a novel approach that divides the data into smaller patches and samples these sequentially, conditioned on previously generated patches. This method reduces the likelihood of mode-seeking (focusing on only a few modes) by lowering the dimensionality for each sampling step, which in turn accelerates convergence to the correct distribution.\n\nKey contributions and findings of the paper include:\n1. **Theoretical Analysis:** The study provides a rigorous analysis of examples where the traditional Langevin dynamics can fail to capture all modes, especially under Gaussian and sub-Gaussian mixtures.\n2. **Chained Langevin Dynamics:** The paper introduces Chained-LD as a solution and establishes its convergence properties, showing that it can sample all modes in a distribution.\n3. **Empirical Validation:** Experiments on synthetic datasets and real image datasets (such as MNIST) successfully conducted showing how Chained-LD can capture all modes compared to standard LD.", "strengths": "The paper is well-written, easy to read, and highly structured, with thorough documentation that enhances its clarity. The computational approach is straightforward to follow, and the main message of the work is communicated effectively.\n\nOne of the most compelling aspects of the paper is the recursion on the vector $n_t$, which presents an innovative approach to the field; to the best of my knowledge, this recursion is novel and has significant implications on the examples presented. The test setting introduced is particularly interesting, as it highlights how SGMs tend to emphasize majority classes, potentially overlooking minority representations in the data.\n\nAdditionally, the authors provide a detailed description of the behavior within a specific test setting, which is later extended to the framework of annealed Langevin dynamics. This progression in the analysis is logical and well-integrated, offering valuable insights into the dynamics of SGMs in multimodal distributions.", "weaknesses": "1. **Missing Citations:** The paper does not reference the recent works, like *Conforti, Dumus, and Gentiloni-Silveri (2023)*, which provides critical insights into the KL divergence bounds. Given the significance of these results for KL bounds, integrating this reference would strengthen the theoretical grounding of the paper.\n2. **Real-World Applicability and Initial Distribution:** The results are derived based on an initial distribution that is not known in advance, raising questions about their direct applicability to real-world scenarios. In practical applications, data is typically standardized and rescaled, and initializations do not start from a well-defined mode $P^{(0)}$. It would be helpful to understand if and how these findings could be adapted to scenarios where the data distribution and modes are not known beforehand. Moreover, I suggest to include that a direct application of these findings would be finding fair algorithms on less represented group. Could the authors discuss potential approaches for adapting their method to scenarios where the initial distribution is unknown or to explain any limitations in applying their results to such cases?\n3. **Handling of Balanced Modes:** The study assumes a setting where one mode is notably dominant. In reality, distributions often lack such stark imbalances. How would the results change when no mode significantly outweighs others? An exploration of how the method performs when modes have similar probabilities would clarify its robustness and generalizability. Could the authors include an additional experiments or theoretical analysis for the case where modes have similar probabilities? This would help demonstrate the method's robustness across different types of multimodal distributions.\n4. **Assumption 2 and Gaussian Mixtures:** On page 5, it is claimed that Assumption 2 is automatically satisfied for Gaussian mixtures, yet this is not elaborated upon. This assumption, particularly in points (iv) and (v), introduces technical constraints that lack intuitive explanations for real-world distributions. A more detailed justification for these points would enhance clarity. Could the authors provide a brief proof or explanation for why Gaussian mixtures satisfy Assumption 2, particularly focusing on points (iv) and (v)?\n5. **Selection of Patch Size $Q$:** The algorithm’s effectiveness relies on the choice of patch size $Q$. However, guidance on how $Q$ should be selected w.r.t. the dimension $d$ for different applications is not provided. Insights into an optimal or adaptive selection process of this parameter could improve the method’s usability in diverse contexts. Could the authors provide a heuristic or guideline for selecting $Q$ based on d and the characteristics of the data, or to discuss the sensitivity of their method to different choices of $Q$?\n6. **Score Function Estimation:** It is unclear how $s_\\theta$, the conditional score function estimator, is obtained and whether it is computed offline or dynamically updated. An explanation of the estimation method and its computational cost would help assess the algorithm’s practicality. Could the authors give a brief description of the estimation procedure for $s_\\theta$ in the main text or appendix, including whether it's pre-computed or updated during sampling?\n7. **Impact on Input Geometry:** The segmentation of the sample into patches for processing may alter the overall geometry of the input vector $x_0$. How does this segmentation affect the coherence and structure of the original data, and does it risk distorting key relationships within the vector? Could the authors discuss or analyze how their patch-based approach preserves or affects important structural relationships in the data?\n8. **Proposition 1 and Convergence Bounds:** On page 7, Proposition 1 references a TV distance condition bounded by $\\varepsilon Q/d$ but does not justify why this specific bound is chosen. Furthermore, the criteria for calling this result a “convergence bound” are unclear and would benefit from further clarification on its theoretical significance. Could the authors share intuition or justification for the specific bound $\\varepsilon Q/d$ and clarify what it is meant by \"convergence bound\" in this context?\n9. **Computational Efficiency:** The proposed approach, particularly with Chained Langevin Dynamics, appears more computationally intensive than traditional methods. Can the authors quantify the additional computational load required for convergence in comparison to previous approaches? Understanding the trade-offs between computational cost and performance is essential for evaluating the practical feasibility of this algorithm. Could the authors include a computational complexity analysis or empirical runtime comparison between their method and traditional Langevin dynamics?\n\nI’m open to improving the evaluation score, as this paper shows real potential in highlighting biases in SGMs, especially around minority representation. Addressing the weaknesses—such as clarifying assumptions, real-world applicability, and computational efficiency—would greatly enhance its impact and make it a valuable reference for understanding and mitigating biases in generative models.", "questions": "See the **Weaknesses** section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article examines the behavior of Langevin dynamics (LD), a sampling method commonly used in generative modeling, especially in contexts where data distributions are multimodal (contain multiple clusters or “modes”). The authors highlight a key limitation of traditional Langevin dynamics: its difficulty in sampling from all modes in a multimodal distribution, as it often struggles to transition between modes in high-dimensional settings.\n\nTo address this, the authors propose Chained Langevin Dynamics (Chained-LD), a novel approach that divides the data into smaller patches and samples these sequentially, conditioned on previously generated patches. This method reduces the likelihood of mode-seeking (focusing on only a few modes) by lowering the dimensionality for each sampling step, which in turn accelerates convergence to the correct distribution.\n\nKey contributions and findings of the paper include:\n1. **Theoretical Analysis:** The study provides a rigorous analysis of examples where the traditional Langevin dynamics can fail to capture all modes, especially under Gaussian and sub-Gaussian mixtures.\n2. **Chained Langevin Dynamics:** The paper introduces Chained-LD as a solution and establishes its convergence properties, showing that it can sample all modes in a distribution.\n3. **Empirical Validation:** Experiments on synthetic datasets and real image datasets (such as MNIST) successfully conducted showing how Chained-LD can capture all modes compared to standard LD.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper is well-written, easy to read, and highly structured, with thorough documentation that enhances its clarity. The computational approach is straightforward to follow, and the main message of the work is communicated effectively.\n\nOne of the most compelling aspects of the paper is the recursion on the vector $n_t$, which presents an innovative approach to the field; to the best of my knowledge, this recursion is novel and has significant implications on the examples presented. The test setting introduced is particularly interesting, as it highlights how SGMs tend to emphasize majority classes, potentially overlooking minority representations in the data.\n\nAdditionally, the authors provide a detailed description of the behavior within a specific test setting, which is later extended to the framework of annealed Langevin dynamics. This progression in the analysis is logical and well-integrated, offering valuable insights into the dynamics of SGMs in multimodal distributions.", "weaknesses": "1. **Missing Citations:** The paper does not reference the recent works, like *Conforti, Dumus, and Gentiloni-Silveri (2023)*, which provides critical insights into the KL divergence bounds. Given the significance of these results for KL bounds, integrating this reference would strengthen the theoretical grounding of the paper.\n2. **Real-World Applicability and Initial Distribution:** The results are derived based on an initial distribution that is not known in advance, raising questions about their direct applicability to real-world scenarios. In practical applications, data is typically standardized and rescaled, and initializations do not start from a well-defined mode $P^{(0)}$. It would be helpful to understand if and how these findings could be adapted to scenarios where the data distribution and modes are not known beforehand. Moreover, I suggest to include that a direct application of these findings would be finding fair algorithms on less represented group. Could the authors discuss potential approaches for adapting their method to scenarios where the initial distribution is unknown or to explain any limitations in applying their results to such cases?\n3. **Handling of Balanced Modes:** The study assumes a setting where one mode is notably dominant. In reality, distributions often lack such stark imbalances. How would the results change when no mode significantly outweighs others? An exploration of how the method performs when modes have similar probabilities would clarify its robustness and generalizability. Could the authors include an additional experiments or theoretical analysis for the case where modes have similar probabilities? This would help demonstrate the method's robustness across different types of multimodal distributions.\n4. **Assumption 2 and Gaussian Mixtures:** On page 5, it is claimed that Assumption 2 is automatically satisfied for Gaussian mixtures, yet this is not elaborated upon. This assumption, particularly in points (iv) and (v), introduces technical constraints that lack intuitive explanations for real-world distributions. A more detailed justification for these points would enhance clarity. Could the authors provide a brief proof or explanation for why Gaussian mixtures satisfy Assumption 2, particularly focusing on points (iv) and (v)?\n5. **Selection of Patch Size $Q$:** The algorithm’s effectiveness relies on the choice of patch size $Q$. However, guidance on how $Q$ should be selected w.r.t. the dimension $d$ for different applications is not provided. Insights into an optimal or adaptive selection process of this parameter could improve the method’s usability in diverse contexts. Could the authors provide a heuristic or guideline for selecting $Q$ based on d and the characteristics of the data, or to discuss the sensitivity of their method to different choices of $Q$?\n6. **Score Function Estimation:** It is unclear how $s_\\theta$, the conditional score function estimator, is obtained and whether it is computed offline or dynamically updated. An explanation of the estimation method and its computational cost would help assess the algorithm’s practicality. Could the authors give a brief description of the estimation procedure for $s_\\theta$ in the main text or appendix, including whether it's pre-computed or updated during sampling?\n7. **Impact on Input Geometry:** The segmentation of the sample into patches for processing may alter the overall geometry of the input vector $x_0$. How does this segmentation affect the coherence and structure of the original data, and does it risk distorting key relationships within the vector? Could the authors discuss or analyze how their patch-based approach preserves or affects important structural relationships in the data?\n8. **Proposition 1 and Convergence Bounds:** On page 7, Proposition 1 references a TV distance condition bounded by $\\varepsilon Q/d$ but does not justify why this specific bound is chosen. Furthermore, the criteria for calling this result a “convergence bound” are unclear and would benefit from further clarification on its theoretical significance. Could the authors share intuition or justification for the specific bound $\\varepsilon Q/d$ and clarify what it is meant by \"convergence bound\" in this context?\n9. **Computational Efficiency:** The proposed approach, particularly with Chained Langevin Dynamics, appears more computationally intensive than traditional methods. Can the authors quantify the additional computational load required for convergence in comparison to previous approaches? Understanding the trade-offs between computational cost and performance is essential for evaluating the practical feasibility of this algorithm. Could the authors include a computational complexity analysis or empirical runtime comparison between their method and traditional Langevin dynamics?\n\nI’m open to improving the evaluation score, as this paper shows real potential in highlighting biases in SGMs, especially around minority representation. Addressing the weaknesses—such as clarifying assumptions, real-world applicability, and computational efficiency—would greatly enhance its impact and make it a valuable reference for understanding and mitigating biases in generative models.", "questions": "See the **Weaknesses** section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730666693793}, {"id": "LORcFlnIDj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9430/Reviewer_nwd8"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper coniders Langevin Dynamics for multimodal distributions (actually mixtures of gaussians, with a small extension to the case of of a large-variance gaussian plus some localized subgaussians). They prove in some toy model cases that Langevin Dynamics does not reach good approximation to these distributions in short time. As a \"solution\" to that problem, they lower the dimensionality by partitioning the data vector in equal pieces and sequentially generating the pieces rather than generating them all together. This improves on the \"curse of dimension\" part of the difficulty.", "review_text": "The paper coniders Langevin Dynamics for multimodal distributions (actually mixtures of gaussians, with a small extension to the case of of a large-variance gaussian plus some localized subgaussians). They prove in some toy model cases that Langevin Dynamics does not reach good approximation to these distributions in short time. As a \"solution\" to that problem, they lower the dimensionality by partitioning the data vector in equal pieces and sequentially generating the pieces rather than generating them all together. This improves on the \"curse of dimension\" part of the difficulty.", "strengths": "The topic of the paper (rate of approximation of multimodal distributions) is of interest and the problem of finding a good scalable strategy in these problems is largely open.", "weaknesses": "The setting of study is mainly a toy model, whose assumptions may or may not hold in realistic scenarios. \nA discussion of limitations of the core/crucial Assumptions 1 and 2 of the paper is dearly missed.\n\nI think that this is a good step in that direction, but I'm not sure if the paper is already valuable as is, mainly due to the doubts I have as to how general Assumptions 1 and 2 actually are, compared to typical applications.", "questions": "1) Why is the setup of Assumption 1 general enough to deserve a paper?\n\nMore precisely, can one say that there are other kinds of multi-modality that appear in applications or not? \nIf yes, then what are the cases not covered by Assumption 1?\nIf not, in what sense does Assumption 1 subsume the main difficulties related to multi-modality?\n\n2) Is it correct to say that the underlying principle of Assumption 2 is not that distinct from the case of Assumption 1, other than allowing the \"concentrated/independent modes\" part of the distribution to be given by more general kinds of measures?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper coniders Langevin Dynamics for multimodal distributions (actually mixtures of gaussians, with a small extension to the case of of a large-variance gaussian plus some localized subgaussians). They prove in some toy model cases that Langevin Dynamics does not reach good approximation to these distributions in short time. As a \"solution\" to that problem, they lower the dimensionality by partitioning the data vector in equal pieces and sequentially generating the pieces rather than generating them all together. This improves on the \"curse of dimension\" part of the difficulty.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The topic of the paper (rate of approximation of multimodal distributions) is of interest and the problem of finding a good scalable strategy in these problems is largely open.", "weaknesses": "The setting of study is mainly a toy model, whose assumptions may or may not hold in realistic scenarios. \nA discussion of limitations of the core/crucial Assumptions 1 and 2 of the paper is dearly missed.\n\nI think that this is a good step in that direction, but I'm not sure if the paper is already valuable as is, mainly due to the doubts I have as to how general Assumptions 1 and 2 actually are, compared to typical applications.", "questions": "1) Why is the setup of Assumption 1 general enough to deserve a paper?\n\nMore precisely, can one say that there are other kinds of multi-modality that appear in applications or not? \nIf yes, then what are the cases not covered by Assumption 1?\nIf not, in what sense does Assumption 1 subsume the main difficulties related to multi-modality?\n\n2) Is it correct to say that the underlying principle of Assumption 2 is not that distinct from the case of Assumption 1, other than allowing the \"concentrated/independent modes\" part of the distribution to be given by more general kinds of measures?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730648954002}, {"id": "yC3ppudg8g", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9430/Reviewer_2K9H"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "The paper considers the task of sampling from a multimodal density using Langevin-like SDEs. The authors considers the standard Euler discretisation of the over damped Langevin diffusion and show that the chain does not discover the modes with high probability after a number of iterations that is exponential in the dimension when the target is a mixture of (sub)-gaussian distributions. To improve over this slow mixing, the paper proposes chained Langevin dynamics, essentially a Gibbs-type approach where Langevin dynamics are applied to batches of the chain instead of the full vector.", "review_text": "The paper considers the task of sampling from a multimodal density using Langevin-like SDEs. The authors considers the standard Euler discretisation of the over damped Langevin diffusion and show that the chain does not discover the modes with high probability after a number of iterations that is exponential in the dimension when the target is a mixture of (sub)-gaussian distributions. To improve over this slow mixing, the paper proposes chained Langevin dynamics, essentially a Gibbs-type approach where Langevin dynamics are applied to batches of the chain instead of the full vector.", "strengths": "The paper considers an interesting problem, since slow mixing in multimodal settings is an issue. The authors also propose an algorithm to obviate such slow mixing.", "weaknesses": "- The paper is hard to read and contains several paragraphs that are of dubious relevance (e.g. lines 141-144 mentioning GANs, several paragraphs on score-based diffusion models, lines 150-161, 179-200, where it is not clear these are relevant for this article). The introduction is long and overall unclear.\n\n- The authors do not seem comfortable with the MCMC literature, e.g. referring to the Euler discretisation of the overdamped Langevin diffusion as \"stochastic gradient Langevin Dynamics\", citing a paper that introduces a completely different approach based on subsampling. \n\n- It is not clear if the authors consider the usual sampling context vs. the generative modelling context: is the score of the data distribution known? Moreover the paper contains plenty of references to the generative modelling literature, but it is not clear what is the purpose, and I can only imagine it is to increase their chances of publication. \n\n- It is unclear if the chained algorithm that is proposed has even the right stationary distribution, nor if it has any useful property for sampling purposes. Moreover, one needs to have the conditional densities to run the algorithm in the first place.\n\n-  The experiments, especially those on the MNIST data, miss explanations on the obtained results, so it is unclear how to interpret them. \n\n- Overall, I find the paper should go through a substantial rewriting and thus should be rejected.", "questions": "- Is the idea behind the chained algorithm (i.e. decomposing the process in blocks) based/influenced on previous works? If so, the authors should mention the relevant papers.\n\n-Could the authors give any justification for their algorithm?\n\n- Lines 170-171: why is there a minus sign in front of the drift term?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers the task of sampling from a multimodal density using Langevin-like SDEs. The authors considers the standard Euler discretisation of the over damped Langevin diffusion and show that the chain does not discover the modes with high probability after a number of iterations that is exponential in the dimension when the target is a mixture of (sub)-gaussian distributions. To improve over this slow mixing, the paper proposes chained Langevin dynamics, essentially a Gibbs-type approach where Langevin dynamics are applied to batches of the chain instead of the full vector.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "The paper considers an interesting problem, since slow mixing in multimodal settings is an issue. The authors also propose an algorithm to obviate such slow mixing.", "weaknesses": "- The paper is hard to read and contains several paragraphs that are of dubious relevance (e.g. lines 141-144 mentioning GANs, several paragraphs on score-based diffusion models, lines 150-161, 179-200, where it is not clear these are relevant for this article). The introduction is long and overall unclear.\n\n- The authors do not seem comfortable with the MCMC literature, e.g. referring to the Euler discretisation of the overdamped Langevin diffusion as \"stochastic gradient Langevin Dynamics\", citing a paper that introduces a completely different approach based on subsampling. \n\n- It is not clear if the authors consider the usual sampling context vs. the generative modelling context: is the score of the data distribution known? Moreover the paper contains plenty of references to the generative modelling literature, but it is not clear what is the purpose, and I can only imagine it is to increase their chances of publication. \n\n- It is unclear if the chained algorithm that is proposed has even the right stationary distribution, nor if it has any useful property for sampling purposes. Moreover, one needs to have the conditional densities to run the algorithm in the first place.\n\n-  The experiments, especially those on the MNIST data, miss explanations on the obtained results, so it is unclear how to interpret them. \n\n- Overall, I find the paper should go through a substantial rewriting and thus should be rejected.", "questions": "- Is the idea behind the chained algorithm (i.e. decomposing the process in blocks) based/influenced on previous works? If so, the authors should mention the relevant papers.\n\n-Could the authors give any justification for their algorithm?\n\n- Lines 170-171: why is there a minus sign in front of the drift term?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730457666147}], "openreview_url": "https://openreview.net/forum?id=VvHVLVUD6m", "arxiv_id": "2406.02017", "paper_pdf": "papers/VvHVLVUD6m.pdf", "paper_pdf_sha256": "2834ee1d036272eff2c65d236c9a4f1741b4ba4933f50300578dd38f196ab192", "paper_pdf_bytes": 25124044, "paper_pdf_source": "openreview", "code_url": "https://github.com/Xiwei-Cheng/Chained_LD", "code_repository": "Xiwei-Cheng/Chained_LD", "code_commit": "8f8da77696037112eb4f29c93674913dc5d45115", "code_archive": "repos/VvHVLVUD6m.zip", "code_archive_sha256": "9b8ed26fbbf94e0b1653456bbf2a1295a90606937ade403017c976d03630976e", "code_archive_bytes": 43217, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 44, "github_languages": {"Python": 180335}, "github_archived": false, "github_pushed_at": "2024-05-28T11:28:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-mode-seeking-properties-of-langevin"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hkQOYyUChL", "year": 2024, "status": "rejected", "title": "Learning and Forgetting Unsafe Examples in Large Language Models", "authors": ["Jiachen ZHAO", "Zhun Deng", "David Madras", "James Zou", "Mengye Ren"], "authorids": ["~Jiachen_ZHAO1", "~Zhun_Deng1", "~David_Madras1", "~James_Zou1", "~Mengye_Ren1"], "authors_source": "OpenReview API", "abstract": "As the number of large language models (LLMs) available to the public grows, there is a pressing need to understand the safety implications associated with these models learning from third-party custom finetuning data.  We explore the behavior of LLMs finetuned on unsafe content, represented by datasets that contain biases, toxicity, and harmfulness, finding that while LLMs can readily learn this unsafe content, they also tend to forget it when subsequently finetuned on safer content.  Drawing inspiration from this forgetting behavior, we introduce the ``\\ff{}'' algorithm, which filters unsafe data based on how strong the model's forgetting signal is for that data. We find that the \\ff{} algorithm outperforms alternative strategies like replay and moral self-correction in curbing LLMs' ability to assimilate unsafe content during custom finetuning, e.g. 75\\% lower than not applying any safety measures and 62\\% lower than using self-correction in toxicity score.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "llylpPoOUb", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8876/Reviewer_qeKd"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors first analyze if the language models can forget unsafe content on fine-tuning. For this the authors first fine-tune a pre-trained language model on a noisy dataset with both safe and unsafe content and then fine-tune this model on safe data and analyze if the model is able to forget the classification of the unsafe data points in the first fine-tuning. Authors name the second fine-tuning stage as safety review. The author utilize different datasets to study the effect of presence of bias, toxicity and harmfulness, and conclude that during safety review the model can easily forget the unsafe content. Motivated by this, the authors propose a forget filter algorithm, which analyzes which samples are forgotten first during the safety review. The ones which are forgotten first are unsafe samples and therefore can be removed from the noisy dataset to make it safe. In the considered setup, the authors demonstrate that such a method could be used to filter out the data and make the models safer.", "review_text": "The authors first analyze if the language models can forget unsafe content on fine-tuning. For this the authors first fine-tune a pre-trained language model on a noisy dataset with both safe and unsafe content and then fine-tune this model on safe data and analyze if the model is able to forget the classification of the unsafe data points in the first fine-tuning. Authors name the second fine-tuning stage as safety review. The author utilize different datasets to study the effect of presence of bias, toxicity and harmfulness, and conclude that during safety review the model can easily forget the unsafe content. Motivated by this, the authors propose a forget filter algorithm, which analyzes which samples are forgotten first during the safety review. The ones which are forgotten first are unsafe samples and therefore can be removed from the noisy dataset to make it safe. In the considered setup, the authors demonstrate that such a method could be used to filter out the data and make the models safer.", "strengths": "* The authors have attempted to make models safer and better aligned. This is an important question in the safety community.\n\n* The experiments attempted by the authors are interesting", "weaknesses": "Major Weaknesses:\n\n* The authors have only presented the case where there is presence of limited amount of safe and unsafe data. In real worlds the models are trained on internet scale corpus where the data is present from different domains and distributions. It is likely that on safety review the model would forget the out of distribution samples from the safety review finetuning dataset rather than explicitly forgetting the unsafe datapoints. Since the authors have limited their dataset to limited amount of safe and unsafe data points, it might give an illusion that the model can actually unlearn the unsafe datapoints. Therefore a more rigorous evaluation where the authors can also consider different sources and domains of safe and unsafe data, should be done.The scalability of the proposed method to filter the features from the pre-training dataset is not clear. \n\nMinor weaknesses:\n\n* The authors should try to perform some attacks like jailbreaking on the models fine tuned on the filtered dataset. If the unsafe content can be filtered from the dataset, then the model should not be jailbroken easily.\n\n* A recent work [1] also uses the concept of second fine-tuning in order to filter out noisy samples. The idea of the proposed approach seems to be very similar to this. It would be nice, it the authors could look into this work [1].\n\n[1] Maini, Pratyush et al. “Characterizing Datapoints via Second-Split Forgetting.” ArXiv abs/2210.15031\n\nUpdate after rebuttal: I appreciate the authors for replying to my concerns. Unfortunately, authors have not addressed my concerns sufficiently. They have not provided additional evidence on jailbreaking attacks. Also I don't think that restricting the evaluation setup to fixed number of safe and unsafe datapoints can help in guaranteeing that the model can become safer using the proposed approach. Given these outstanding concerns I would like to maintain my score.", "questions": "I would request the authors to kindly address the comments in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors first analyze if the language models can forget unsafe content on fine-tuning. For this the authors first fine-tune a pre-trained language model on a noisy dataset with both safe and unsafe content and then fine-tune this model on safe data and analyze if the model is able to forget the classification of the unsafe data points in the first fine-tuning. Authors name the second fine-tuning stage as safety review. The author utilize different datasets to study the effect of presence of bias, toxicity and harmfulness, and conclude that during safety review the model can easily forget the unsafe content. Motivated by this, the authors propose a forget filter algorithm, which analyzes which samples are forgotten first during the safety review. The ones which are forgotten first are unsafe samples and therefore can be removed from the noisy dataset to make it safe. In the considered setup, the authors demonstrate that such a method could be used to filter out the data and make the models safer.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "strengths": "* The authors have attempted to make models safer and better aligned. This is an important question in the safety community.\n\n* The experiments attempted by the authors are interesting", "weaknesses": "Major Weaknesses:\n\n* The authors have only presented the case where there is presence of limited amount of safe and unsafe data. In real worlds the models are trained on internet scale corpus where the data is present from different domains and distributions. It is likely that on safety review the model would forget the out of distribution samples from the safety review finetuning dataset rather than explicitly forgetting the unsafe datapoints. Since the authors have limited their dataset to limited amount of safe and unsafe data points, it might give an illusion that the model can actually unlearn the unsafe datapoints. Therefore a more rigorous evaluation where the authors can also consider different sources and domains of safe and unsafe data, should be done.The scalability of the proposed method to filter the features from the pre-training dataset is not clear. \n\nMinor weaknesses:\n\n* The authors should try to perform some attacks like jailbreaking on the models fine tuned on the filtered dataset. If the unsafe content can be filtered from the dataset, then the model should not be jailbroken easily.\n\n* A recent work [1] also uses the concept of second fine-tuning in order to filter out noisy samples. The idea of the proposed approach seems to be very similar to this. It would be nice, it the authors could look into this work [1].\n\n[1] Maini, Pratyush et al. “Characterizing Datapoints via Second-Split Forgetting.” ArXiv abs/2210.15031\n\nUpdate after rebuttal: I appreciate the authors for replying to my concerns. Unfortunately, authors have not addressed my concerns sufficiently. They have not provided additional evidence on jailbreaking attacks. Also I don't think that restricting the evaluation setup to fixed number of safe and unsafe datapoints can help in guaranteeing that the model can become safer using the proposed approach. Given these outstanding concerns I would like to maintain my score.", "questions": "I would request the authors to kindly address the comments in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699053548009}, {"id": "cDrUEMMV0X", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8876/Reviewer_pR1P"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents an empirical study that unsafe examples are more likely to be forgotten during fine-tuning and that the ability emerges with larger-scale LMs. The authors utilize this empirical finding and propose a novel approach, ForgetFilter, to identify harmful training example which reduces harm in learned models.", "review_text": "This paper presents an empirical study that unsafe examples are more likely to be forgotten during fine-tuning and that the ability emerges with larger-scale LMs. The authors utilize this empirical finding and propose a novel approach, ForgetFilter, to identify harmful training example which reduces harm in learned models.", "strengths": "- The analysis is quite extensive. The authors performed study over multiple model architectures, ranging from smaller ones to larger ones, and evaluated various types of unsafe examples (bias, toxicity, harmfulness) and observe consistent findings. \n- The proposed approach, ForgetFilter, is simple and effective. I appreciate author's evaluation of long-term safety where the proposed approach clearly outperforms alternative methods.", "weaknesses": "- I believe the authors should discuss related works that filter training data / perform data selection based on learning dynamics like frequency of forgetting. \n\nSome papers:\n\n[1] Maini et al. Characterizing Datapoints via Second-Split Forgetting, NeurIPS 2022\n\n[2] Swayamdipta et al. Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics. EMNLP 2020\n\n- I wonder whether in practice \"safe examples\" are always readily available for a fine-tuning task. For example, if I download a random dataset D from the web without having a safe dataset of the same task as D, can I still apply ForgetFilter? Is this setup practical?", "questions": "- We see from Figure 2 that up to 80% of harmful examples will be forgotten in the end. But what is the proportion of harmful examples among all the forgotten examples?\n\nI am asking the question above because I am a bit surprised to see applying ForgetFiltering improves downstream performance as well in Table 2. I thought ForgetFiltering may also remove training examples that are not harmful mistakenly. Can you explain side effects that is happening on downstream performance?\n\n- See the \"weakness\" part for my question about applying ForgetFilter when no safe dataset is available.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an empirical study that unsafe examples are more likely to be forgotten during fine-tuning and that the ability emerges with larger-scale LMs. The authors utilize this empirical finding and propose a novel approach, ForgetFilter, to identify harmful training example which reduces harm in learned models.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The analysis is quite extensive. The authors performed study over multiple model architectures, ranging from smaller ones to larger ones, and evaluated various types of unsafe examples (bias, toxicity, harmfulness) and observe consistent findings. \n- The proposed approach, ForgetFilter, is simple and effective. I appreciate author's evaluation of long-term safety where the proposed approach clearly outperforms alternative methods.", "weaknesses": "- I believe the authors should discuss related works that filter training data / perform data selection based on learning dynamics like frequency of forgetting. \n\nSome papers:\n\n[1] Maini et al. Characterizing Datapoints via Second-Split Forgetting, NeurIPS 2022\n\n[2] Swayamdipta et al. Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics. EMNLP 2020\n\n- I wonder whether in practice \"safe examples\" are always readily available for a fine-tuning task. For example, if I download a random dataset D from the web without having a safe dataset of the same task as D, can I still apply ForgetFilter? Is this setup practical?", "questions": "- We see from Figure 2 that up to 80% of harmful examples will be forgotten in the end. But what is the proportion of harmful examples among all the forgotten examples?\n\nI am asking the question above because I am a bit surprised to see applying ForgetFiltering improves downstream performance as well in Table 2. I thought ForgetFiltering may also remove training examples that are not harmful mistakenly. Can you explain side effects that is happening on downstream performance?\n\n- See the \"weakness\" part for my question about applying ForgetFilter when no safe dataset is available.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698826080695}, {"id": "GCiQ1M0N04", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8876/Reviewer_DzUN"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper claims that the unsafe data points are more likely to be forgotten during finetuning. They study this effect under different scales of Large Language Models. They design a data filtration method to remove unsafe datapoints from the original finetuning dataset.", "review_text": "This paper claims that the unsafe data points are more likely to be forgotten during finetuning. They study this effect under different scales of Large Language Models. They design a data filtration method to remove unsafe datapoints from the original finetuning dataset.", "strengths": "1. I found the collection of datasets well done and expressive enough to conduct rigorous experiments. Moreover, the use of different model sizes was indeed a good addition to understanding the effects of scale on the phenomena studied in this work. \n2. The motivation of this paper is strong since producing safer models is indeed an important issue.\n\nOverall, I vote to reject this paper. This paper does not demonstrate any new behavior that is not already apparent from existing literature. The experiments also are not setup correctly to prove the original claim that unsafe points are more easily forgotten. If the authors sufficiently answer all of my questions, I am willing to increase my scores.", "weaknesses": "1. In the abstract, it is claimed that models tend to forget unsafe data points when finetuned on safe data points. Isn't this true of any attribute? Training on data points that include only waterbirds will decrease the performance of a model on landbirds for image vision tasks for example. This forgetting phenomenon when shifting domains like this is well-observed in the literature in my opinion. My most grave concern with this paper is that the forgetting of unsafe examples is only done during safety review. However, when further finetuning on safe examples, it is completely expected that the forgetting on safe examples would be less, the forgetting on random points would be slightly higher, and the forgetting on unsafe points would be more. This seems like it has everything to do with the distributions from which the points rely and not that there is something fundamental about unsafe data that is more easily forgettable which is what this paper seems to claim. For example, if one did an Unsafe Review by finetuning on unsafe data points, I expect the exact opposite trend to hold. You mention evidence to support this point in that using safe examples from a different domain causes forgetting over all points, safe or unsafe, from the original domain. Thus, it is not the safety of a data point affecting its rate of forgetting, just the similarity to the data points seen during the review. If this is the case, then the claim that unsafe data points are more easily forgettable than safe data points has not been corroborated. You have only proved that finetuning on safe points from a domain causes the model to produce answers similar to these safe points on the same domain, which is not surprising at all and such ideas exist already in the literature. \n2. As a slight writing suggestion, I would make it such that the list of contributions on page 2 had more details. Currently, they are very spare and do not convey much information besides barebones ideas. Adding more details here would greatly improve readability.\n3. This paper proposes the ForgetFilter method as a way of filtering the finetuning dataset of unsafe examples to produce safer models. However, in the Related Works section, there is no mention of different filtering algorithms. This is a large miss in my opinion. It is difficult to understand where such a filtering technique stands in the context of the large existing body of work of data filtration for safety. I would add a brief description of existing filtration methods and why this method improves on these methods. \n4. The Related Works also does not mention the connection between memory and safety that has long been analyzed in the literature. For example, the work \"Does Learning Require Memorization? A Short Tale about a Long Tail\" discussed how private points on the tails of the distribution get memorized and influence the weights of the models more than other data points. This connection should be mentioned here as it is completely possible that unsafe points may belong to tails of the data distribution and the effects examined in this paper can be explained by this previous work.\n5. I found the ordering and structure of Section 3.1 not readable. I would rearrange this section in chronological ordering to better understand the data first before discussing what a safety review is. \n6. I found several spelling and grammatical errors in this paper. While not critical to my score, a scientific paper should be more presentable. For example, on page 3, you misspelled promote as promot. Another example is forgetting the comma after default on page 3.  Fixing these will improve the presentation of the work significantly. \n7. The metric you are reporting is the average rate for a set of data points. You are not reporting the expected value. The expected value is a specific term and you report the empirical average, which is not the same value. Please change this to be correct.\n8. The main contribution of the ForgetFilter method is a way of filtering the original finetuning dataset such that the resulting model produces safer outputs. It does this by removing points predicted to be unsafe. Therefore, for the experiments, it should be tested if using a forgetting rate to predict the safety of a datapoint in the finetuning dataset is a strong method for predicting the safety of a datapoint. Thus, the baselines should be simpler methods or existing methods of filtering the data set. However, Safety Replay and Moral Self-Correction are neither of these. Thus, it is unclear whether Forget-Filter is really improving over anything since the comparison is not fair. There are many ways of filtering datasets for safety. These methods must be used as a baseline first and foremost to truly understand the strength of ForgetFilter.", "questions": "1. Why are the ground truth responses in the BBQ dataset modified to a stereotypical choice?\n2. Why were only the learning rate and the batch size changed from the default hyperparameters of every model? Did the results seen depend on this hyperparameter selection or was this done purely for computational reasons? If it is the latter, please make sure to note this in the main text.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper claims that the unsafe data points are more likely to be forgotten during finetuning. They study this effect under different scales of Large Language Models. They design a data filtration method to remove unsafe datapoints from the original finetuning dataset.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "strengths": "1. I found the collection of datasets well done and expressive enough to conduct rigorous experiments. Moreover, the use of different model sizes was indeed a good addition to understanding the effects of scale on the phenomena studied in this work. \n2. The motivation of this paper is strong since producing safer models is indeed an important issue.\n\nOverall, I vote to reject this paper. This paper does not demonstrate any new behavior that is not already apparent from existing literature. The experiments also are not setup correctly to prove the original claim that unsafe points are more easily forgotten. If the authors sufficiently answer all of my questions, I am willing to increase my scores.", "weaknesses": "1. In the abstract, it is claimed that models tend to forget unsafe data points when finetuned on safe data points. Isn't this true of any attribute? Training on data points that include only waterbirds will decrease the performance of a model on landbirds for image vision tasks for example. This forgetting phenomenon when shifting domains like this is well-observed in the literature in my opinion. My most grave concern with this paper is that the forgetting of unsafe examples is only done during safety review. However, when further finetuning on safe examples, it is completely expected that the forgetting on safe examples would be less, the forgetting on random points would be slightly higher, and the forgetting on unsafe points would be more. This seems like it has everything to do with the distributions from which the points rely and not that there is something fundamental about unsafe data that is more easily forgettable which is what this paper seems to claim. For example, if one did an Unsafe Review by finetuning on unsafe data points, I expect the exact opposite trend to hold. You mention evidence to support this point in that using safe examples from a different domain causes forgetting over all points, safe or unsafe, from the original domain. Thus, it is not the safety of a data point affecting its rate of forgetting, just the similarity to the data points seen during the review. If this is the case, then the claim that unsafe data points are more easily forgettable than safe data points has not been corroborated. You have only proved that finetuning on safe points from a domain causes the model to produce answers similar to these safe points on the same domain, which is not surprising at all and such ideas exist already in the literature. \n2. As a slight writing suggestion, I would make it such that the list of contributions on page 2 had more details. Currently, they are very spare and do not convey much information besides barebones ideas. Adding more details here would greatly improve readability.\n3. This paper proposes the ForgetFilter method as a way of filtering the finetuning dataset of unsafe examples to produce safer models. However, in the Related Works section, there is no mention of different filtering algorithms. This is a large miss in my opinion. It is difficult to understand where such a filtering technique stands in the context of the large existing body of work of data filtration for safety. I would add a brief description of existing filtration methods and why this method improves on these methods. \n4. The Related Works also does not mention the connection between memory and safety that has long been analyzed in the literature. For example, the work \"Does Learning Require Memorization? A Short Tale about a Long Tail\" discussed how private points on the tails of the distribution get memorized and influence the weights of the models more than other data points. This connection should be mentioned here as it is completely possible that unsafe points may belong to tails of the data distribution and the effects examined in this paper can be explained by this previous work.\n5. I found the ordering and structure of Section 3.1 not readable. I would rearrange this section in chronological ordering to better understand the data first before discussing what a safety review is. \n6. I found several spelling and grammatical errors in this paper. While not critical to my score, a scientific paper should be more presentable. For example, on page 3, you misspelled promote as promot. Another example is forgetting the comma after default on page 3.  Fixing these will improve the presentation of the work significantly. \n7. The metric you are reporting is the average rate for a set of data points. You are not reporting the expected value. The expected value is a specific term and you report the empirical average, which is not the same value. Please change this to be correct.\n8. The main contribution of the ForgetFilter method is a way of filtering the original finetuning dataset such that the resulting model produces safer outputs. It does this by removing points predicted to be unsafe. Therefore, for the experiments, it should be tested if using a forgetting rate to predict the safety of a datapoint in the finetuning dataset is a strong method for predicting the safety of a datapoint. Thus, the baselines should be simpler methods or existing methods of filtering the data set. However, Safety Replay and Moral Self-Correction are neither of these. Thus, it is unclear whether Forget-Filter is really improving over anything since the comparison is not fair. There are many ways of filtering datasets for safety. These methods must be used as a baseline first and foremost to truly understand the strength of ForgetFilter.", "questions": "1. Why are the ground truth responses in the BBQ dataset modified to a stereotypical choice?\n2. Why were only the learning rate and the batch size changed from the default hyperparameters of every model? Did the results seen depend on this hyperparameter selection or was this done purely for computational reasons? If it is the latter, please make sure to note this in the main text.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698799805159}, {"id": "QyZnr6oty4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8876/Reviewer_LEME"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This research focuses on the safety implications of LLMs trained on third-party custom finetuning data. The study finds that LLMs can learn unsafe content but tend to forget it when subsequently trained on safer data. To address this, the \"ForgetFilter\" algorithm is introduced, which effectively filters out unsafe data during finetuning, resulting in significantly reduced assimilation of harmful content by LLMs.", "review_text": "This research focuses on the safety implications of LLMs trained on third-party custom finetuning data. The study finds that LLMs can learn unsafe content but tend to forget it when subsequently trained on safer data. To address this, the \"ForgetFilter\" algorithm is introduced, which effectively filters out unsafe data during finetuning, resulting in significantly reduced assimilation of harmful content by LLMs.", "strengths": "* The problem addressed in this paper holds significant importance as it focuses on ensuring the generation of safe responses by LLMs.\n* The overall narrative presented is quite reasonable, highlighting the LLM's tendency to forget conflicting content when subjected to safety fine-tuning. Additionally, the observation that larger models exhibit improved ability to forget unsafe content during the reviewing stage adds an intriguing aspect.\n* The proposed FF algorithm demonstrates sound reasoning and showcases strong empirical performance, further bolstering its credibility.", "weaknesses": "* I have a concern regarding the paper's consideration of the first fine-tuning stage involving noisy and unsafe content. Considering that fine-tuning datasets are typically small and of high quality, I find this setting to be somewhat artificial. In my opinion, a more interesting and realistic issue arises from the existence of unsafe content during the pre-training stage of LLMs [1]. I kindly suggest that the authors explore whether the ForgetFilter (FF) method can effectively filter out unsafe content from the pre-training dataset, as this would provide valuable insights to the field.\n\n* The reason why unsafe content is forgotten remains unexplained.  Providing an empirical or theoretical explanation for this phenomenon would greatly enhance the paper and warrant a higher score. Additionally, it is crucial for the authors to investigate the conditions under which unsafe content can be forgotten, enabling readers to understand when to effectively apply the proposed ForgetFilter (FF) method. \n\n\n[1] Pretraining Language Models with Human Preferences", "questions": "The presentation can benefit from a diagram to show the procedure of the 2-stage fine-tuning in the paper.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This research focuses on the safety implications of LLMs trained on third-party custom finetuning data. The study finds that LLMs can learn unsafe content but tend to forget it when subsequently trained on safer data. To address this, the \"ForgetFilter\" algorithm is introduced, which effectively filters out unsafe data during finetuning, resulting in significantly reduced assimilation of harmful content by LLMs.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "* The problem addressed in this paper holds significant importance as it focuses on ensuring the generation of safe responses by LLMs.\n* The overall narrative presented is quite reasonable, highlighting the LLM's tendency to forget conflicting content when subjected to safety fine-tuning. Additionally, the observation that larger models exhibit improved ability to forget unsafe content during the reviewing stage adds an intriguing aspect.\n* The proposed FF algorithm demonstrates sound reasoning and showcases strong empirical performance, further bolstering its credibility.", "weaknesses": "* I have a concern regarding the paper's consideration of the first fine-tuning stage involving noisy and unsafe content. Considering that fine-tuning datasets are typically small and of high quality, I find this setting to be somewhat artificial. In my opinion, a more interesting and realistic issue arises from the existence of unsafe content during the pre-training stage of LLMs [1]. I kindly suggest that the authors explore whether the ForgetFilter (FF) method can effectively filter out unsafe content from the pre-training dataset, as this would provide valuable insights to the field.\n\n* The reason why unsafe content is forgotten remains unexplained.  Providing an empirical or theoretical explanation for this phenomenon would greatly enhance the paper and warrant a higher score. Additionally, it is crucial for the authors to investigate the conditions under which unsafe content can be forgotten, enabling readers to understand when to effectively apply the proposed ForgetFilter (FF) method. \n\n\n[1] Pretraining Language Models with Human Preferences", "questions": "The presentation can benefit from a diagram to show the procedure of the 2-stage fine-tuning in the paper.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697945895140}], "openreview_url": "https://openreview.net/forum?id=hkQOYyUChL", "arxiv_id": "2312.12736", "paper_pdf": "papers/hkQOYyUChL.pdf", "paper_pdf_sha256": "88253a52b7eef8b698331fd1dd51417a16b23d29daf9160ec901e46f683c7cc9", "paper_pdf_bytes": 588135, "paper_pdf_source": "openreview", "code_url": "https://github.com/andotalao24/learn-forget-unsafe-llm", "code_repository": "andotalao24/learn-forget-unsafe-llm", "code_commit": "9c2faf7d1a404455e5661c8610dfa550e1fad8dc", "code_archive": "repos/hkQOYyUChL.zip", "code_archive_sha256": "df4c1ccd44cf70b4194e40721129e3e89c8a15f52dd516c83bd8c5fb6bd61e39", "code_archive_bytes": 22964, "code_file_count": 12, "code_extensions": {".py": 9, ".sh": 3}, "github_disk_usage_kb": 38, "github_languages": {"Python": 60039, "Shell": 2298}, "github_archived": false, "github_pushed_at": "2024-07-15T03:32:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-and-forgetting-unsafe-examples-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wtr-9AKxCI5", "year": 2023, "status": "rejected", "title": "MobileViTv3: Mobile-Friendly Vision Transformer with Simple and Effective Fusion of Local, Global and Input Features", "authors": ["Shakti Nagnath Wadekar", "Abhishek Chaurasia"], "authorids": ["~Shakti_Nagnath_Wadekar1", "~Abhishek_Chaurasia2"], "authors_source": "OpenReview API", "abstract": "MobileViT (MobileViTv1) combines convolutional neural networks (CNNs) and vision transformers (ViTs) to create light-weight models for mobile vision tasks. Though the main MobileViTv1-block helps to achieve competitive state-of-the-art results, the fusion block inside MobileViTv1-block, creates scaling challenges and has a complex learning task. We propose changes to the fusion block that are simple and effective to create MobileViTv3-block, which addresses the scaling and simplifies the learning task. Our proposed MobileViTv3-block used to create MobileViTv3-XXS, XS and S models outperform MobileViTv1 on ImageNet-1k, ADE20K, COCO and PascalVOC2012 datasets. On ImageNet-1K, MobileViTv3-XXS and MobileViTv3-XS surpasses MobileViTv1-XXS and MobileViTv1-XS by 2% and 1.9% respectively. Recently published MobileViTv2 architecture removes fusion block and uses linear complexity transformers to perform better than MobileViTv1. We add our proposed fusion block to MobileViTv2 to create MobileViTv3-0.5, 0.75 and 1.0 models. MobileViTv3-0.5 and MobileViTv3-0.75 outperforms MobileViTv2-0.5 and MobileViTv2-0.75 by 2.1% and 1.0% respectively on ImageNet-1K dataset. For segmentation task, MobileViTv3-1.0 achieves 2.07% and 1.1% better mIOU compared to MobileViTv2-1.0 on ADE20K dataset and PascalVOC2012 dataset respectively.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "XjT8Yv8-MT4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3372/Reviewer_jAum"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper attempts to resolve the scaling issues of MobileViTv1-blocks which hinder the learning. Specifically, the authors propose MobileViTv3-block which has a more simplified architecture. The proposed blocks have been added to both MobileViTv1 (with transformer blocks) and MobileViTv2 (with linear transformer blocks) models and experiments on various tasks such as classification (ImageNet), detection(COCO) and segmentation (ADE20K) demonstrate improved results. ", "review_text": "This work presents new MobileViT blocks that make the architecture more efficient and purportedly more accurate in different tasks (e.g. classification, segmentation, detection, etc.). Despite the great engineering effort behind this work, the novelty and contribution are very limited and as a results don't merit acceptance. ", "strengths": "Strengths: \n\n1. The paper is well-written and easy to follow. The visualizations allow for better understanding of the proposed technique. \n\n2. The proposed method has been rigorously tested on many different tasks and datasets. \n\n3. The authors have performed many optimizations to further perfect the proposed blocks. This is reflected in consistent improvements as well as the reduction in number of parameters and flops \n\nWeaknesses/Concerns\n\n1. My main criticism of this work is lack of novelty. As mentioned above, this work is an amazing engineering effort through which various building blocks (e.g. 3 x 3 conv layers, etc.) have been replaced with seemingly better alternatives. However, it is hard to pin-point a major contribution or novelty which is specific to this work (e.g. novel design of an attention layer)\n\n2. It is understandable that authors attempt to further optimize the already existing MobileViTv1 and MobileViTv2 blocks. However, one may wonder why the transformer layers should  proceed the conv-based layers ? why not processing features in parallel streams ? the current design is very limited in novelty, and the motivation behind it seems to be the reduction of number of flops, parameters, etc. However, there should also be better justification of the intuition. \n\n3. Experimental comparisons are presented in a somewhat contrived settings. It is preferred to know how the proposed method performs compared to other competing approaches.  \n\n4. Performance gains in downstream tasks of semantic segmentation on ADE20K dataset and object detection on COCO are higher than classification. What could be the reason behind this ?  \n\n5. Another concern is the throughput of the model with the proposed blocks on devices other than GPU. Is there any comparison on mobile devices ? in fact, the real latency on such devices could be drastically different than what is measured on the GPU. In addition, it is recommended to use more modern GPU hardware (e.g. A100) to report latency numbers to better illustrate the performance differences. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper attempts to resolve the scaling issues of MobileViTv1-blocks which hinder the learning. Specifically, the authors propose MobileViTv3-block which has a more simplified architecture. The proposed blocks have been added to both MobileViTv1 (with transformer blocks) and MobileViTv2 (with linear transformer blocks) models and experiments on various tasks such as classification (ImageNet), detection(COCO) and segmentation (ADE20K) demonstrate improved results. ", "strength_and_weaknesses": "Strengths: \n\n1. The paper is well-written and easy to follow. The visualizations allow for better understanding of the proposed technique. \n\n2. The proposed method has been rigorously tested on many different tasks and datasets. \n\n3. The authors have performed many optimizations to further perfect the proposed blocks. This is reflected in consistent improvements as well as the reduction in number of parameters and flops \n\nWeaknesses/Concerns\n\n1. My main criticism of this work is lack of novelty. As mentioned above, this work is an amazing engineering effort through which various building blocks (e.g. 3 x 3 conv layers, etc.) have been replaced with seemingly better alternatives. However, it is hard to pin-point a major contribution or novelty which is specific to this work (e.g. novel design of an attention layer)\n\n2. It is understandable that authors attempt to further optimize the already existing MobileViTv1 and MobileViTv2 blocks. However, one may wonder why the transformer layers should  proceed the conv-based layers ? why not processing features in parallel streams ? the current design is very limited in novelty, and the motivation behind it seems to be the reduction of number of flops, parameters, etc. However, there should also be better justification of the intuition. \n\n3. Experimental comparisons are presented in a somewhat contrived settings. It is preferred to know how the proposed method performs compared to other competing approaches.  \n\n4. Performance gains in downstream tasks of semantic segmentation on ADE20K dataset and object detection on COCO are higher than classification. What could be the reason behind this ?  \n\n5. Another concern is the throughput of the model with the proposed blocks on devices other than GPU. Is there any comparison on mobile devices ? in fact, the real latency on such devices could be drastically different than what is measured on the GPU. In addition, it is recommended to use more modern GPU hardware (e.g. A100) to report latency numbers to better illustrate the performance differences. ", "clarity,_quality,_novelty_and_reproducibility": "The novelty seems to be very limited. It builds upon previous MobileViT architectures with little to no major contribution. The proposed effort seems to be reproducible as authors have mentioned about public release of the code in GitHub. ", "summary_of_the_review": "This work presents new MobileViT blocks that make the architecture more efficient and purportedly more accurate in different tasks (e.g. classification, segmentation, detection, etc.). Despite the great engineering effort behind this work, the novelty and contribution are very limited and as a results don't merit acceptance. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667273736204}, {"id": "H9EbTvBpl_a", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3372/Reviewer_8bQB"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper propsed mobilevit v3 which did modification based on mobilevit by reducing conv3x3 to conv1x1, adjusting the routing from input and pre-conv embedding [local feature as named in the paper] layers for feature fusion.\n\nIt produces a better efficiency especially at low-parameter cases. ", "review_text": "good performance against SoTA baselines, while the modification and pruning of network follows prior arts with limited novelty. ", "strengths": "Strength: \n\n1) From fig3, Performance of mobilevitv3 is significantly improved comparing with v2/v1, when # of  parameters is smaller than 3.5m. \n\n2) The model is well evaluated based on various tasks:  imagenet for classification, coco for detection and pascal ade20k for segmentation. In all tasks the model shows significant improvement especially on XS cases. \n\n3) Extensive related works are studied in the paper such that the results are compared against strong enough SoTA methods.\n\nCons:\n\nThe ablation study of modifications is lacked in the paper, i.e. step-by-step performance [flops, parameters, acc] comparison of doing each modification, so that we may got the idea of the contribution of each part. \n\nThe novelty of the paper is somehow limited since most modification is already proposed in prior works [depth-wise conv etc.], and the improvement is marginal when paraemters > 3.5m on imagenet. \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper propsed mobilevit v3 which did modification based on mobilevit by reducing conv3x3 to conv1x1, adjusting the routing from input and pre-conv embedding [local feature as named in the paper] layers for feature fusion.\n\nIt produces a better efficiency especially at low-parameter cases. ", "strength_and_weaknesses": "Strength: \n\n1) From fig3, Performance of mobilevitv3 is significantly improved comparing with v2/v1, when # of  parameters is smaller than 3.5m. \n\n2) The model is well evaluated based on various tasks:  imagenet for classification, coco for detection and pascal ade20k for segmentation. In all tasks the model shows significant improvement especially on XS cases. \n\n3) Extensive related works are studied in the paper such that the results are compared against strong enough SoTA methods.\n\nCons:\n\nThe ablation study of modifications is lacked in the paper, i.e. step-by-step performance [flops, parameters, acc] comparison of doing each modification, so that we may got the idea of the contribution of each part. \n\nThe novelty of the paper is somehow limited since most modification is already proposed in prior works [depth-wise conv etc.], and the improvement is marginal when paraemters > 3.5m on imagenet. \n\n", "clarity,_quality,_novelty_and_reproducibility": "1) easy to read and follow, and easy to reproduce. \n2) novelty is limited in my opinion although it shows good performance on benchmarks. ", "summary_of_the_review": "good performance against SoTA baselines, while the modification and pruning of network follows prior arts with limited novelty. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666718919872}, {"id": "_8AiJAY-RK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3372/Reviewer_kTXc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a light-weight model for mobile vision tasks, namely MobileViT-v3. The proposed model is based on the previous MobileVit v1 and v2. Specifically, several modifications have been conducted to improve the previous architectures, including replacing 3x3 conv with 1x1 conv, fusing local and global features, adding a residual connection between the input feature and the fusion feature, and replacing normal 3x3 conv with depthwise 3x3 conv. The above changes result in more efficient and effective mobile backbones. Experiments have been performed on the MobileViT families to validate the effectiveness.", "review_text": "Overall, the paper provides extensive experiments to verify its effectiveness. The motivation is clear. I admire the authors’ efforts in engineering exploration. However, the technical novelty is limited since the previous works have investigated the core components (i.e., 1x1 conv, depthwise 3x3 conv, and residual connection). Besides, the paper lacks in-depth analysis. I will give my first rating as `reject`.", "strengths": "### Pros:\n\n- The motivation is clear. It’s practical to design a model that is lightweight and efficient on mobile devices.\n- The technical comparisons between the MobileViT family are clear, as shown in Figure 2. The intuitive and straightforward illustration makes the methodology easy to follow.\n- Extensive experiments (including image/video classifications, object detection, and semantic segmentation) have been provided to validate the effectiveness. However, I still hold some concerns about the experimental comparisons, which are shown in the cons part.\n\n### Cons:\n\n- ***Lack of in-depth analysis and limited technical contribution***\n\nThe primary concern is the technical contribution. The components proposed in this paper are the common techniques to achieve the corresponding purpose. For example, adopting 1x1 conv to minimize the network parameters and employing depthwise 3x3 conv to reduce parameters and FLOPs without significantly impacting performance have already been investigated in previous works [1,2,3]. Besides, this paper needs to include an in-depth analysis of the usage of the above modules. From my perspective, I admire the authors’ efforts in the engineering exploration. However, the main ideas have been explored in many previous works, which makes the works seem to be incremental.\n\n- ***Over-claimed contribution and unclear writing***\n\nThe paper emphasizes that this work addresses the challenges of scaling and simplifies the learning tasks. However, I can not see it’s an essential problem in the previous MobileViT v1 and v2, since there are also a series of scaled models in MobielViT v1 and v2. The contribution seems to be over-claimed. Please present this part more clearly.\n\n- ***Insufficient comparisons***\n\nThe experimental comparisons are mainly conducted on the Top1 accuracy, parameters, and Flops. However, for mobile applications, the actual latency also matters. Please report the throughput for comprehensive comparisons.\n\n- ***Others***\n\n(a) *Extended comparisons:* Since depthwise 3x3 conv is utilized in the MobileViT v3, I would like to see the ablation studies of this module on object detection and segmentation since depthwise conv usually performs well in the dense prediction tasks.\n\n(b) Please add the conclusion part to make the paper more complete.\n\n---\n\n[1] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions, 2021.\n\n[2] PVT v2: Improved Baselines with Pyramid Vision Transformer, 2021.\n\n[3] Co-Scale Conv-Attentional Image Transformers, 2021.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces a light-weight model for mobile vision tasks, namely MobileViT-v3. The proposed model is based on the previous MobileVit v1 and v2. Specifically, several modifications have been conducted to improve the previous architectures, including replacing 3x3 conv with 1x1 conv, fusing local and global features, adding a residual connection between the input feature and the fusion feature, and replacing normal 3x3 conv with depthwise 3x3 conv. The above changes result in more efficient and effective mobile backbones. Experiments have been performed on the MobileViT families to validate the effectiveness.", "strength_and_weaknesses": "### Pros:\n\n- The motivation is clear. It’s practical to design a model that is lightweight and efficient on mobile devices.\n- The technical comparisons between the MobileViT family are clear, as shown in Figure 2. The intuitive and straightforward illustration makes the methodology easy to follow.\n- Extensive experiments (including image/video classifications, object detection, and semantic segmentation) have been provided to validate the effectiveness. However, I still hold some concerns about the experimental comparisons, which are shown in the cons part.\n\n### Cons:\n\n- ***Lack of in-depth analysis and limited technical contribution***\n\nThe primary concern is the technical contribution. The components proposed in this paper are the common techniques to achieve the corresponding purpose. For example, adopting 1x1 conv to minimize the network parameters and employing depthwise 3x3 conv to reduce parameters and FLOPs without significantly impacting performance have already been investigated in previous works [1,2,3]. Besides, this paper needs to include an in-depth analysis of the usage of the above modules. From my perspective, I admire the authors’ efforts in the engineering exploration. However, the main ideas have been explored in many previous works, which makes the works seem to be incremental.\n\n- ***Over-claimed contribution and unclear writing***\n\nThe paper emphasizes that this work addresses the challenges of scaling and simplifies the learning tasks. However, I can not see it’s an essential problem in the previous MobileViT v1 and v2, since there are also a series of scaled models in MobielViT v1 and v2. The contribution seems to be over-claimed. Please present this part more clearly.\n\n- ***Insufficient comparisons***\n\nThe experimental comparisons are mainly conducted on the Top1 accuracy, parameters, and Flops. However, for mobile applications, the actual latency also matters. Please report the throughput for comprehensive comparisons.\n\n- ***Others***\n\n(a) *Extended comparisons:* Since depthwise 3x3 conv is utilized in the MobileViT v3, I would like to see the ablation studies of this module on object detection and segmentation since depthwise conv usually performs well in the dense prediction tasks.\n\n(b) Please add the conclusion part to make the paper more complete.\n\n---\n\n[1] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions, 2021.\n\n[2] PVT v2: Improved Baselines with Pyramid Vision Transformer, 2021.\n\n[3] Co-Scale Conv-Attentional Image Transformers, 2021.", "clarity,_quality,_novelty_and_reproducibility": "The motivation of this paper is clear. The work is based on previous publications with a few modifications. Thus, it cloud be easy to reproduce the experimental results.", "summary_of_the_review": "Overall, the paper provides extensive experiments to verify its effectiveness. The motivation is clear. I admire the authors’ efforts in engineering exploration. However, the technical novelty is limited since the previous works have investigated the core components (i.e., 1x1 conv, depthwise 3x3 conv, and residual connection). Besides, the paper lacks in-depth analysis. I will give my first rating as `reject`.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "There are no ethics concerns in this paper.", "recommendation": "3: reject, not good enough"}, "tcdate": 1666669252482}, {"id": "-mdfsKdNTWU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3372/Reviewer_AJcM"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work proposes a series of modifications to the MobileViT V1 and V2 architectures. It introduces a few tweaks to the original model architecture, including replacing regular conv by depthwise conv to reduce latency, using 1x1 conv instead of 3x3 conv to reduce number of parameters, and adding residual connection in the proposed MobileViT V3 block. Such modifications brings in additional quality gain compared to original MobileViT V1 and V2, while being more efficient in terms of FLOPs and number of parameters in some cases. Experiments on downstreaming tasks such as detection and semantic segmentation also show that the proposed modifications can benefit dense prediction tasks as well.", "review_text": "In summary, this work tries to study a good problem and propose valid modifications to improve existing models. However, it falls short regarding technical details, latency benchmark and comparison with recent works. \n\nI would like authors to address my comments, especially regarding lack of latency benchmark and comparison with more recent models. ", "strengths": "It is of great interest to study possible solutions for optimizing transformer models regarding the speed while maintain its superior quality. Although there are abundant works on pushing the boundary of transformer models in vision tasks in terms of quality, there is far less study on optimizing those heavy-lifting models for on-device use cases, where convents such as MobileNet and EfficientNet still dominate. \n\nThis work builds on top of two recent works, MobileViT V1 and V2, and proposes simple way to improve the model performance without altering the model architectures significantly. The figures in the paper provide a clear view of the landscape of recent on-device ViT models, and the place of the proposed MobileViT V3 models. This gives users a good sense of how the proposed model compares to existing works. This is highly appreciated.\n\nTechnically, this work is sound: using depth wise conv for speed-up and 1x1 conv to shrink model parameters are both widely used techniques to optimize a model. From experiments, the ablation study does show that these modifications indeed contribute to performance improvement. \n\nDespite of the strength, there are a lot of problems in the current manuscript that weakens this work.\n\n1. The paper claims it “addresses the scaling and simplifies the learning task” However, from the experimental results, it is not clear why the proposed modifications can achieve the claimed goal. Even the original MobileViT V1 and V2 are not difficult to scale up (especially V2 as the authors already have several variants of that by scaling it up and down): there are just a matter of changing number of filters or hidden size, and possibly adding or remove certain number of blocks. In addition, I do not see the original MobileViT models are not easy to train. In this sense, this paper’s claim does not hold: it is not supported from the experiments. \n\n2. Another big issue is benchmarking. This paper presents the plots of accuracy v.s. model FLOPs and number of parameters, where the proposed model has fewer number of parameters or FLOPs compared to its rivals and existing convnets. It gives a false impression that the proposed models outperforms all existing convents and ViT models. This is not exactly true. For on-device scenarios, FLOPs and number of parameters are not the golden rule to decide a model’s real performance, but latency is. Especially, although MobileNet has larger number of parameters, it actually runs much faster than ViT model (MobileViT) due to CPU-friendly ops. As pointed in [1] and [2], number of parameters and FLOPs are not positively correlated to latency. In fact, when a model contains excessive transpose/reshape ops or many branches, it can lead to higher latency even though the number of parameters are less. \nUnfortunately, the only real latency benchmark is present in Appendix B Table 8, on MobileViT model variants only. It does not provide a holistic picture on the real latency comparison of existing commonly used mobile models. \nTherefore, a latency benchmark on real mobile devices (not only on desktop GPU) would be more convincing, just like what MobileViT V1 and V2 have done. Without such benchmark, this work is less convincing that it will run fast as expected. \n\n3. Technical details\nThere are many unclear details that need further clarification. For example:\n\n    - When replacing 3x3 conv by 1x1 conv, you will have smaller receptive field as 1x1 conv generally serves a channel reduction method. Due to such loss, how do you ensure the information from global features is proper preserved?\n\n    -  How do you decide the scaling factor in Table 1? They seem pretty ad-hoc. For example, 1.3, 1.6 and 1.7 are not quite standard numbers to scale a model.\n\n   -  In implementation details, it is not clear why the model can only be trained with a much smaller batch size. Given that the model has fewer FLOPs and number of parameters comparing to MobileViT V1, it should be trainable using the same batch size as MobileViT V1, which is 1024.\n\n4. Related works miss a few recent works, for example:\n\n    - LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference, 2021\n    - EfficientFormer: Vision Transformers at MobileNet Speed, 2022\n    - Next-ViT: Next Generation Vision Transformer for Efficient Deployment in Realistic Industrial Scenarios, 2022\n\nPlease consider adding these papers and do a more comprehensive comparison. All released code so it should not be difficult to reproduce those works. This is also related to experimental validation. The current comparison is mostly on MobileViT variants and old convnets. It is far from enough to justify the proposed model superiority. More recent models should be included and compared.\n\n- [1] An Improved One millisecond Mobile Backbone, 2022\n- [2] EfficientFormer: Vision Transformers at MobileNet Speed, 2022", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This work proposes a series of modifications to the MobileViT V1 and V2 architectures. It introduces a few tweaks to the original model architecture, including replacing regular conv by depthwise conv to reduce latency, using 1x1 conv instead of 3x3 conv to reduce number of parameters, and adding residual connection in the proposed MobileViT V3 block. Such modifications brings in additional quality gain compared to original MobileViT V1 and V2, while being more efficient in terms of FLOPs and number of parameters in some cases. Experiments on downstreaming tasks such as detection and semantic segmentation also show that the proposed modifications can benefit dense prediction tasks as well.", "strength_and_weaknesses": "It is of great interest to study possible solutions for optimizing transformer models regarding the speed while maintain its superior quality. Although there are abundant works on pushing the boundary of transformer models in vision tasks in terms of quality, there is far less study on optimizing those heavy-lifting models for on-device use cases, where convents such as MobileNet and EfficientNet still dominate. \n\nThis work builds on top of two recent works, MobileViT V1 and V2, and proposes simple way to improve the model performance without altering the model architectures significantly. The figures in the paper provide a clear view of the landscape of recent on-device ViT models, and the place of the proposed MobileViT V3 models. This gives users a good sense of how the proposed model compares to existing works. This is highly appreciated.\n\nTechnically, this work is sound: using depth wise conv for speed-up and 1x1 conv to shrink model parameters are both widely used techniques to optimize a model. From experiments, the ablation study does show that these modifications indeed contribute to performance improvement. \n\nDespite of the strength, there are a lot of problems in the current manuscript that weakens this work.\n\n1. The paper claims it “addresses the scaling and simplifies the learning task” However, from the experimental results, it is not clear why the proposed modifications can achieve the claimed goal. Even the original MobileViT V1 and V2 are not difficult to scale up (especially V2 as the authors already have several variants of that by scaling it up and down): there are just a matter of changing number of filters or hidden size, and possibly adding or remove certain number of blocks. In addition, I do not see the original MobileViT models are not easy to train. In this sense, this paper’s claim does not hold: it is not supported from the experiments. \n\n2. Another big issue is benchmarking. This paper presents the plots of accuracy v.s. model FLOPs and number of parameters, where the proposed model has fewer number of parameters or FLOPs compared to its rivals and existing convnets. It gives a false impression that the proposed models outperforms all existing convents and ViT models. This is not exactly true. For on-device scenarios, FLOPs and number of parameters are not the golden rule to decide a model’s real performance, but latency is. Especially, although MobileNet has larger number of parameters, it actually runs much faster than ViT model (MobileViT) due to CPU-friendly ops. As pointed in [1] and [2], number of parameters and FLOPs are not positively correlated to latency. In fact, when a model contains excessive transpose/reshape ops or many branches, it can lead to higher latency even though the number of parameters are less. \nUnfortunately, the only real latency benchmark is present in Appendix B Table 8, on MobileViT model variants only. It does not provide a holistic picture on the real latency comparison of existing commonly used mobile models. \nTherefore, a latency benchmark on real mobile devices (not only on desktop GPU) would be more convincing, just like what MobileViT V1 and V2 have done. Without such benchmark, this work is less convincing that it will run fast as expected. \n\n3. Technical details\nThere are many unclear details that need further clarification. For example:\n\n    - When replacing 3x3 conv by 1x1 conv, you will have smaller receptive field as 1x1 conv generally serves a channel reduction method. Due to such loss, how do you ensure the information from global features is proper preserved?\n\n    -  How do you decide the scaling factor in Table 1? They seem pretty ad-hoc. For example, 1.3, 1.6 and 1.7 are not quite standard numbers to scale a model.\n\n   -  In implementation details, it is not clear why the model can only be trained with a much smaller batch size. Given that the model has fewer FLOPs and number of parameters comparing to MobileViT V1, it should be trainable using the same batch size as MobileViT V1, which is 1024.\n\n4. Related works miss a few recent works, for example:\n\n    - LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference, 2021\n    - EfficientFormer: Vision Transformers at MobileNet Speed, 2022\n    - Next-ViT: Next Generation Vision Transformer for Efficient Deployment in Realistic Industrial Scenarios, 2022\n\nPlease consider adding these papers and do a more comprehensive comparison. All released code so it should not be difficult to reproduce those works. This is also related to experimental validation. The current comparison is mostly on MobileViT variants and old convnets. It is far from enough to justify the proposed model superiority. More recent models should be included and compared.\n\n- [1] An Improved One millisecond Mobile Backbone, 2022\n- [2] EfficientFormer: Vision Transformers at MobileNet Speed, 2022", "clarity,_quality,_novelty_and_reproducibility": "This paper is not well written as there are many unclear details as I point out above. Such issues make it difficult to assess the merit of this work, given that it essentially just proposes a few well-know modifications to existing models without changing the meta architecture at all.\n\nThe paper is not well organized. I feel a lot of interesting part should not be put into appendix but in the main body such as latency numbers and limitations. Adding such content into the main body will greatly help improve the quality of the manuscript.\n \nThe main text also finishes too abruptly. There is no conclusion or future work, and the ablation study stops suddenly. Please consider better organize the manuscript.   ", "summary_of_the_review": "In summary, this work tries to study a good problem and propose valid modifications to improve existing models. However, it falls short regarding technical details, latency benchmark and comparison with recent works. \n\nI would like authors to address my comments, especially regarding lack of latency benchmark and comparison with more recent models. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666390744311}], "openreview_url": "https://openreview.net/forum?id=wtr-9AKxCI5", "arxiv_id": "2209.15159", "paper_pdf": "papers/wtr-9AKxCI5.pdf", "paper_pdf_sha256": "f0c407ee71df02bfed77b3fd7d63af69e1defcc35d36caf352d02dc460e280a5", "paper_pdf_bytes": 5990095, "paper_pdf_source": "openreview", "code_url": "https://github.com/micronDLA/MobileViTv3", "code_repository": "micronDLA/MobileViTv3", "code_commit": "d381beb017ae3c244686afaa48064f95865df7b9", "code_archive": "repos/wtr-9AKxCI5.zip", "code_archive_sha256": "5525ed62c8215cc988be5684e90865092eb1fb335c4aea0d56b4d771152bd8e6", "code_archive_bytes": 56595, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 45, "github_languages": {"Python": 100763}, "github_archived": true, "github_pushed_at": "2022-10-06T17:19:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mobilevitv3-mobile-friendly-vision"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SGOma2sAF7Q", "year": 2022, "status": "rejected", "title": "LCS: Learning Compressible Subspaces for Adaptive Network Compression at Inference Time", "authors": ["Maxwell Horton", "Elvis Nunez", "Anish Prabhu", "Anurag Ranjan", "Ali Farhadi", "Mohammad Rastegari"], "authorids": ["~Maxwell_Horton1", "~Elvis_Nunez1", "~Anish_Prabhu1", "~Anurag_Ranjan1", "~Ali_Farhadi4", "~Mohammad_Rastegari2"], "authors_source": "OpenReview API", "abstract": "When deploying deep learning models to a device, it is traditionally assumed that available computational resources (compute, memory, and power) remain static. However, real-world computing systems do not always provide stable resource guarantees. Computational resources need to be conserved when load from other processes is high or battery power is low. Inspired by recent works on neural network subspaces, we propose a method for training a \"compressible subspace\" of neural networks that contains a fine-grained spectrum of models that range from highly efficient to highly accurate. Our models require no retraining, thus our subspace of models can be deployed entirely on-device to allow adaptive network compression at inference time. We present results for achieving arbitrarily fine-grained accuracy-efficiency trade-offs at inference time for structured and unstructured sparsity. We achieve accuracies on-par with standard models when testing our uncompressed models, and maintain high accuracy for sparsity rates above 90% when testing our compressed models. We also demonstrate that our algorithm extends to quantization at variable bit widths, achieving accuracy on par with individually trained networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "09Hl29L8661", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper485/Reviewer_Dpge"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces learnable compressible subspaces, which attempts to learn a set of models that can be switched at inference time to adapt to different resource requirements. This work is motivated by previous work in neural subspaces and slimmable networks. It is evaluated on CIFAR10 and ImageNet and compared against other recent works for adaptable inference models. These results show under certain conditions LCS can maintain higher accuracies at larger sparsities compared to other works.", "review_text": "=== Strengths ===\n\nTable 1 is useful in summarizing the related works and the claimed advantages for LCS.\n\nThe experiment section listed a significant amount of detail on hyperparams and methods.\n\nThe small discussion and figures on batch norm stats shifts were interesting and helpful for motivating group and instance norm for this application.\n\nThis area of adaptive inference is becoming more and more important with larger models and specialized big-little architectures.\n\n=== Weaknesses ===\n\nIt is unclear how these networks switch between models at inference time, which of course should depend on the type of compression used.  For sparsity, this seems like it would require dynamically pruning the model at inference time which seems very dangerous. For unstructured sparsity, this may require special hardware for taking advantage of that unstructured sparsity. For quantization, this may require hardware can support fine-grained switching of the quantization bitwidth.\n\nMy understanding is that batch norm stats have to be recomputed for NS and US in a post-training way but not necessarily at inference time. It doesn't seem fair to avoid this step since it can be done before model deployment and takes a fraction of the training time. Leaving it out in the evaluation also nullifies the comparison if this is not fair.\n\nThe claim that other methods need additional batch norm params and cannot support fine-grained compression level is mostly correct, but the importance of this seems overstated. In practice, it seems more reasonable to chose a smaller subset of model configurations that can be fully tested before deployment, and the batch norm params should be nearly negligible compared to the weights. Also, quantization LCS does of course limit the number of compression levels, and structured pruning LCS limits the number of compression levels to the number of channels (which is similar to US).\n\nThe existence of gamma and alpha together is confusing to follow. Since the parameterization in linear, it seems like only one of these should be necessary.\n\nThere should be other works included for building robust compressible models, e.g. Robust Quantization (Neurips20).\n\nThe writing is clear but repetitive in some areas. For example, I believe there are 5 sentences talking about being inspired by Wortsman in the first few pages.\n\n=== Questions ===\nIn the Related Works, the neural subspace method is described as operating on simplices, but the description in Section 3.1 seems to be on lines. Is this deliberate?\n\nPlease correct me if I'm wrong but isn't the unstructured compressible point method Dropout? There might be differently weighted probabilities and dynamic dropout probabilities, but they seem fundamentally the same. \n\nFor quantization, what hardware supports dynamic fine-grained switching from 3-8 bits? How are the pruned channels or pruned individual weights chosen at runtime in an adaptive way? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces learnable compressible subspaces, which attempts to learn a set of models that can be switched at inference time to adapt to different resource requirements. This work is motivated by previous work in neural subspaces and slimmable networks. It is evaluated on CIFAR10 and ImageNet and compared against other recent works for adaptable inference models. These results show under certain conditions LCS can maintain higher accuracies at larger sparsities compared to other works.", "main_review": "=== Strengths ===\n\nTable 1 is useful in summarizing the related works and the claimed advantages for LCS.\n\nThe experiment section listed a significant amount of detail on hyperparams and methods.\n\nThe small discussion and figures on batch norm stats shifts were interesting and helpful for motivating group and instance norm for this application.\n\nThis area of adaptive inference is becoming more and more important with larger models and specialized big-little architectures.\n\n=== Weaknesses ===\n\nIt is unclear how these networks switch between models at inference time, which of course should depend on the type of compression used.  For sparsity, this seems like it would require dynamically pruning the model at inference time which seems very dangerous. For unstructured sparsity, this may require special hardware for taking advantage of that unstructured sparsity. For quantization, this may require hardware can support fine-grained switching of the quantization bitwidth.\n\nMy understanding is that batch norm stats have to be recomputed for NS and US in a post-training way but not necessarily at inference time. It doesn't seem fair to avoid this step since it can be done before model deployment and takes a fraction of the training time. Leaving it out in the evaluation also nullifies the comparison if this is not fair.\n\nThe claim that other methods need additional batch norm params and cannot support fine-grained compression level is mostly correct, but the importance of this seems overstated. In practice, it seems more reasonable to chose a smaller subset of model configurations that can be fully tested before deployment, and the batch norm params should be nearly negligible compared to the weights. Also, quantization LCS does of course limit the number of compression levels, and structured pruning LCS limits the number of compression levels to the number of channels (which is similar to US).\n\nThe existence of gamma and alpha together is confusing to follow. Since the parameterization in linear, it seems like only one of these should be necessary.\n\nThere should be other works included for building robust compressible models, e.g. Robust Quantization (Neurips20).\n\nThe writing is clear but repetitive in some areas. For example, I believe there are 5 sentences talking about being inspired by Wortsman in the first few pages.\n\n=== Questions ===\nIn the Related Works, the neural subspace method is described as operating on simplices, but the description in Section 3.1 seems to be on lines. Is this deliberate?\n\nPlease correct me if I'm wrong but isn't the unstructured compressible point method Dropout? There might be differently weighted probabilities and dynamic dropout probabilities, but they seem fundamentally the same. \n\nFor quantization, what hardware supports dynamic fine-grained switching from 3-8 bits? How are the pruned channels or pruned individual weights chosen at runtime in an adaptive way? ", "summary_of_the_review": "This paper is an interesting proposal that attempts to apply the ideas of neural subspaces to produce a set of compressed models at varying points on the accuracy / efficiency curve. Yet, these methods in the end seem more about learning robust compressible models and stray far from the original neural subspace idea, especially with compressible points. In my current understanding, these networks seem to have no demonstrated advantage to universally slimmable networks, which have a simpler validation process, runtime switching method, and more intuitive training procedure. The comparison against these networks and others needs to be better justified since I currently do not understand why fine-tuning is not allowed. If I misunderstood the method significantly, I would be willing to increase my score, but currently I suggest rejecting the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636137443585}, {"id": "N-G9v9L6wr_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper485/Reviewer_FDfN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper present a method for learning a compressible subspace of neural networks that contains a fine-grained spectrum of models that range from highly efficient to highly accurate. The proposed method allows choosing the proper point of trade-off between accuracy and efficiency at inference time, according to the available resources. There are also efforts to reduce the runtime tweaking overhead like replacing BatchNorm with GroupNorm.", "review_text": "Strengths: the paper is well motivated, as adaption to runtime available resource is important. \n\nWeaknesses:\n* The method is a direct extension of the learning subspace method.\n* There are important details missing from the paper. E.g., there is no info on how to generate a alpha via the state of hardware in runtime. This creates severe difficulty in understanding and reproducing the method.\n* There is no material to measure hardware performance. For example, only the accuracy of the classification models are given, but memory bandwidth, latency or FPS are not available to quantitatively measure the advantage. \n* The paper is not properly peer-compared. For example, the work is not well compared with pruning and quantification methods.\n\nQuestions for the Author(s):\n* please elaborate on the definition of the compression function f and the intuition behind?\n* how to choose the hyper-parameter alpha in a hardware run-time?\n* What will happen to the arch of a model, if pruning is also performed? \n* If using this method in a hardware, how to change the quantization meta-parameters(scale and zero-point) accordingly?\n\n    ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper present a method for learning a compressible subspace of neural networks that contains a fine-grained spectrum of models that range from highly efficient to highly accurate. The proposed method allows choosing the proper point of trade-off between accuracy and efficiency at inference time, according to the available resources. There are also efforts to reduce the runtime tweaking overhead like replacing BatchNorm with GroupNorm.", "main_review": "Strengths: the paper is well motivated, as adaption to runtime available resource is important. \n\nWeaknesses:\n* The method is a direct extension of the learning subspace method.\n* There are important details missing from the paper. E.g., there is no info on how to generate a alpha via the state of hardware in runtime. This creates severe difficulty in understanding and reproducing the method.\n* There is no material to measure hardware performance. For example, only the accuracy of the classification models are given, but memory bandwidth, latency or FPS are not available to quantitatively measure the advantage. \n* The paper is not properly peer-compared. For example, the work is not well compared with pruning and quantification methods.\n\nQuestions for the Author(s):\n* please elaborate on the definition of the compression function f and the intuition behind?\n* how to choose the hyper-parameter alpha in a hardware run-time?\n* What will happen to the arch of a model, if pruning is also performed? \n* If using this method in a hardware, how to change the quantization meta-parameters(scale and zero-point) accordingly?\n\n    ", "summary_of_the_review": "The paper proposes a method that reasonably extends the learning subspace method to allow performing accuracy-efficiency tradeoff according to runtime available resource. The method has been evaluated on several classification tasks and find to be useful.\n\nHowever, the paper does not clearly explain how the compression is performed, with important details like choice of alpha missing. The measurement of speedup is not that quantitative, lacking realworld test stats. It is very difficult to evaluate the contribution of this paper under these conditions.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635912375087}, {"id": "pddY3qvZKc2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper485/Reviewer_jBcH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to balance the inference accuracy and efficiency by training a subspace of neural networks and then adapting the network within the subspace at inference time.", "review_text": "The novelty of this paper looks limited. It is an extension of recent work on learning network subspaces. Meanwhile, compared to existing works with adaptive networks, the advantages of this work are the finer-grained compress level and no need to recalibrate BN. The benefit of such improvements looks trivial.\n\nMany vital details are missing in this paper. The detailed form of the compression level function $\\gamma(\\alpha)$ and the compression function $f(\\omega, \\gamma)$ are not provided. These are the core of the algorithm and the authors need to show them. \n\nAlso, the authors do not explain what determines the dimension n of the stochastic function $\\alpha$. If $\\alpha$ controls the sparsity at the level of each weight, n will be extremely large, and the training overhead of the proposed algorithm is extremely large because they need to do n forward passes and backward passes for each batch. Even $\\alpha$ controls the sparsity at the layer level, the training overhead will still be formidable. I cannot find much information about n in this paper. The authors need to provide more details about the dimension n of $\\alpha$.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to balance the inference accuracy and efficiency by training a subspace of neural networks and then adapting the network within the subspace at inference time.", "main_review": "The novelty of this paper looks limited. It is an extension of recent work on learning network subspaces. Meanwhile, compared to existing works with adaptive networks, the advantages of this work are the finer-grained compress level and no need to recalibrate BN. The benefit of such improvements looks trivial.\n\nMany vital details are missing in this paper. The detailed form of the compression level function $\\gamma(\\alpha)$ and the compression function $f(\\omega, \\gamma)$ are not provided. These are the core of the algorithm and the authors need to show them. \n\nAlso, the authors do not explain what determines the dimension n of the stochastic function $\\alpha$. If $\\alpha$ controls the sparsity at the level of each weight, n will be extremely large, and the training overhead of the proposed algorithm is extremely large because they need to do n forward passes and backward passes for each batch. Even $\\alpha$ controls the sparsity at the layer level, the training overhead will still be formidable. I cannot find much information about n in this paper. The authors need to provide more details about the dimension n of $\\alpha$.", "summary_of_the_review": "Many important details are missing in this paper. For example, the finer-grained compression level is a major selling point of this paper, but the authors did not even provide the compression level function $\\gamma(\\alpha)$.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635906534925}, {"id": "G1OaXu_Jhkr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper485/Reviewer_EG95"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to learn compressive subspaces which can adaptively compress the network during inference. It constructs either linear subspace or a single endpoint for compression. It replaces BN with GroupNorm to avoid re-calibrating during inference or after adjustment. Their method is evaluated in three different scenarios: structured sparsity, unstructured sparsity, and quantization.", "review_text": "The paper is well written and easy to follow. The ideas of constructing a linear subspace and using the function $f(w(\\alpha), \\gamma (\\alpha))$ to perform compression during inference are novel. The analysis of BN parameters in adjustment provides quantitative analysis in this area. The experiments are exhaustive and can well support their ideas.\n\nHowever, as the paper claims, they bias the subspace to contain high-accuracy solutions at one end and high-efficiency solutions at the other end. In my understanding, two endpoints are using the same network architecture. How to train a network to obtain $w_1$ and $w_2$ in this case?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to learn compressive subspaces which can adaptively compress the network during inference. It constructs either linear subspace or a single endpoint for compression. It replaces BN with GroupNorm to avoid re-calibrating during inference or after adjustment. Their method is evaluated in three different scenarios: structured sparsity, unstructured sparsity, and quantization.", "main_review": "The paper is well written and easy to follow. The ideas of constructing a linear subspace and using the function $f(w(\\alpha), \\gamma (\\alpha))$ to perform compression during inference are novel. The analysis of BN parameters in adjustment provides quantitative analysis in this area. The experiments are exhaustive and can well support their ideas.\n\nHowever, as the paper claims, they bias the subspace to contain high-accuracy solutions at one end and high-efficiency solutions at the other end. In my understanding, two endpoints are using the same network architecture. How to train a network to obtain $w_1$ and $w_2$ in this case?", "summary_of_the_review": "I think the paper is well written; the method is novel and interesting; the experiment can well support the claims. Just need to clarify some details", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635887437212}], "openreview_url": "https://openreview.net/forum?id=SGOma2sAF7Q", "arxiv_id": "2110.04252", "paper_pdf": "papers/SGOma2sAF7Q.pdf", "paper_pdf_sha256": "b351cec7d73cd0bc6456716a829564bd79be90f442bcac316187b46c2cb4ccf1", "paper_pdf_bytes": 1591332, "paper_pdf_source": "openreview", "code_url": "https://github.com/apple/learning-compressible-subspaces", "code_repository": "apple/learning-compressible-subspaces", "code_commit": "308f5216c7b0c15a0ebc88c7cec3e7d28c82b6c1", "code_archive": "repos/SGOma2sAF7Q.zip", "code_archive_sha256": "f484dc27f7964049847fd1025144fb77cb1318fddff1009813b75bf0e37eaf02", "code_archive_bytes": 79518, "code_file_count": 31, "code_extensions": {".py": 29, ".sh": 2}, "github_disk_usage_kb": 59, "github_languages": {"Python": 168374, "Shell": 225}, "github_archived": false, "github_pushed_at": "2021-10-27T16:22:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lcs-learning-compressible-subspaces-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bK-rJMKrOsm", "year": 2021, "status": "rejected", "title": "Multi-Head Attention: Collaborate Instead of Concatenate", "authors": ["Jean-Baptiste Cordonnier", "Andreas Loukas", "Martin Jaggi"], "authorids": ["~Jean-Baptiste_Cordonnier2", "~Andreas_Loukas1", "~Martin_Jaggi1"], "authors_source": "OpenReview API", "abstract": "Attention layers are widely used in natural language processing (NLP) and are beginning to influence computer vision architectures. However, they suffer from over-parameterization. For instance, it was shown that the majority of attention heads could be pruned without impacting accuracy. This work aims to enhance current understanding on how multiple heads interact. Motivated by the observation that trained attention heads share common key/query projections, we propose a collaborative multi-head attention layer that enables heads to learn shared projections. Our scheme decreases the number of parameters in an attention layer and can be used as a drop-in replacement in any transformer architecture.For instance, by allowing heads to collaborate on a neural machine translation task, we can reduce the key dimension by 4× without any loss in performance. We also show that it is possible to re-parametrize a pre-trained multi-head attention layer into our collaborative attention layer. Even without retraining, collaborative multi-head attention manages to reduce the size of the key and query projections by half without sacrificing accuracy. Our code is public.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "K4mGNXyqoGD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3019/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "========================\n\nPaper Summary:\n\nThis paper proposes a new form of multi-head attention. It can reduce parameters and FLOPs of Transformer models without performance loss on En-De translation. Moreover, for pre-trained language models, no large-scale pretraining is required to convert the attention. A tensor decomposition based method is proposed for the conversion.\n\n==========================\n\nOverall review\n\nThis paper challenges the widely adopted multi-head attention, asking the question whether the concatenation of multiple heads is the best way to fuse heads. The proposed method is well-motivated (Fig. 1), but the empirical performance does not show a clear advantage over the conventional way. More rigorous experiment is needed to justify this new model.\n\nPros\n\n- the method is novel and well-motivated\n- works for both original transformer and pre-trained language models\n\nCons\n\n- improvements (in terms of evaluation metrics) is marginal to original MHA\n- more translation / generation tasks should be evaluated\n- does not measure empirical speed up at inference time\n\n==========================\n\nQuestions / Suggestions\n\n- In the abstract, the authors argued for over-parameterization. In fact, over-parameterized Transformers such as GPT-3 in fact achieves very strong performance. This is still an open question so might not be appropriate to say transformers suffers over-parameterization.\n- In the end of introduction, Synthesizer model is mentioned but this part is not clearly explained in the rest of the paper.\n- Figure 2 is somewhat confusing. I couldn't understand the relationship between (c) & (d) to the rest of the figure.\n\n===========================\n\nMinor Issues\n\n- Figure 4 x-axis: 767 -> 768", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Review by R5", "review": "========================\n\nPaper Summary:\n\nThis paper proposes a new form of multi-head attention. It can reduce parameters and FLOPs of Transformer models without performance loss on En-De translation. Moreover, for pre-trained language models, no large-scale pretraining is required to convert the attention. A tensor decomposition based method is proposed for the conversion.\n\n==========================\n\nOverall review\n\nThis paper challenges the widely adopted multi-head attention, asking the question whether the concatenation of multiple heads is the best way to fuse heads. The proposed method is well-motivated (Fig. 1), but the empirical performance does not show a clear advantage over the conventional way. More rigorous experiment is needed to justify this new model.\n\nPros\n\n- the method is novel and well-motivated\n- works for both original transformer and pre-trained language models\n\nCons\n\n- improvements (in terms of evaluation metrics) is marginal to original MHA\n- more translation / generation tasks should be evaluated\n- does not measure empirical speed up at inference time\n\n==========================\n\nQuestions / Suggestions\n\n- In the abstract, the authors argued for over-parameterization. In fact, over-parameterized Transformers such as GPT-3 in fact achieves very strong performance. This is still an open question so might not be appropriate to say transformers suffers over-parameterization.\n- In the end of introduction, Synthesizer model is mentioned but this part is not clearly explained in the rest of the paper.\n- Figure 2 is somewhat confusing. I couldn't understand the relationship between (c) & (d) to the rest of the figure.\n\n===========================\n\nMinor Issues\n\n- Figure 4 x-axis: 767 -> 768", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604825045313}, {"id": "hMa-yylY1q-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3019/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper analyzes the multi-head attention in transformers and suggests to use collaboration instead of concatenation of multiple heads. Empirical results on WMT’16 English-German demonstrates that the proposed approach reduces the of parameters without sacrificing performance. Further experiments on pre-trained BERT models also demonstrate its efficiency.  Overall, the paper is well motivated and provides a deep analysis of redundancy of the multi-head attention. \n\nHere are a few detailed comments:\n\n-- In practice, it seems the proposed approach didn’t speed up the training, even though it gives an improvement in terms of FLOPS. This may limit to the large scale of BERT pre-training.\n\n-- From Figure 3 and Table 2, even reducing the $\\hat{D_k}$ from 768 to 128, the total number of parameters of pretrained models (e.g. BERT-base) only reduces from 108.3M to 96.6M. However, the performance has dropped significantly (83.0 -> 77.6 GLUE score) in the BERT-base setting. If we take a closer look, in case of $\\hat{D_k}$=768, the performance on the large tasks (e.g. MNLI) dropped significantly from 84.1 to 83.4 in terms of accuracy, and its improvement is from small tasks (e.g. RTE). It is better to have a deeper analysis and explanation. \n \n===== Update after author response ===== \n\nIf you look a deeper look into Transformers, in case of BERT, the attention block only takes around 25% of total parameters. It is that suspicious that collaborative MHA takes 18% less time in practical since it requires many factors e.g., GPU kernel fusion. \n\nRegarding the performance on MNLI, from Fig 5, it shows that when D_k is larger than 512, MHA reaches to the baseline in terms of accuracy. Additional information from Tab 2, these models have more than 101.4M (in case of D_k = 384) which is almost the same as the original baseline. However, the performance on GLUE is dropped largely, thus it is hard to support papers' claims.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well motivated but weak results", "review": "This paper analyzes the multi-head attention in transformers and suggests to use collaboration instead of concatenation of multiple heads. Empirical results on WMT’16 English-German demonstrates that the proposed approach reduces the of parameters without sacrificing performance. Further experiments on pre-trained BERT models also demonstrate its efficiency.  Overall, the paper is well motivated and provides a deep analysis of redundancy of the multi-head attention. \n\nHere are a few detailed comments:\n\n-- In practice, it seems the proposed approach didn’t speed up the training, even though it gives an improvement in terms of FLOPS. This may limit to the large scale of BERT pre-training.\n\n-- From Figure 3 and Table 2, even reducing the $\\hat{D_k}$ from 768 to 128, the total number of parameters of pretrained models (e.g. BERT-base) only reduces from 108.3M to 96.6M. However, the performance has dropped significantly (83.0 -> 77.6 GLUE score) in the BERT-base setting. If we take a closer look, in case of $\\hat{D_k}$=768, the performance on the large tasks (e.g. MNLI) dropped significantly from 84.1 to 83.4 in terms of accuracy, and its improvement is from small tasks (e.g. RTE). It is better to have a deeper analysis and explanation. \n \n===== Update after author response ===== \n\nIf you look a deeper look into Transformers, in case of BERT, the attention block only takes around 25% of total parameters. It is that suspicious that collaborative MHA takes 18% less time in practical since it requires many factors e.g., GPU kernel fusion. \n\nRegarding the performance on MNLI, from Fig 5, it shows that when D_k is larger than 512, MHA reaches to the baseline in terms of accuracy. Additional information from Tab 2, these models have more than 101.4M (in case of D_k = 384) which is almost the same as the original baseline. However, the performance on GLUE is dropped largely, thus it is hard to support papers' claims.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603938298943}, {"id": "LtQDXO9FCN5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3019/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents an interesting collaborative MHA to enable heads to share projections, which can be easily applied to most existing transformer-based models, including NMT and pre-training models. With using the new collaborative MHA, the number of parameters and FLOPs can be decreased. \n\nThe PCA-based analysis of the query/key matrices is interesting and impressive, which motivates the newly proposed MHA. The authors also propose a Tensor Decomposition based method to easily convert MHA to its collaborative version without retraining.    \n\nThe paper is well-written and organized, the experiments are thorough. However, I have several concerns:\n1)\tIn the Table 1, it will be more convincing, if the run time on CPU and GPU can be provided. \n2)\tAll the pre-trained models (BERT/DistilBERT/ALBERT) are evaluated on the GLUE dataset, the experiments on more challenging tasks like QA (e.g. SQuAD 1.1/2.0) should be added. \n3)\tThe proposed method seems to be not effective for pre-trained models, e.g. when the number of parameters is decreased from 108.5M to 96.6M, this reduction of model size is not that big, while the average score decreases from 83.2 to 77.6. \n4)\tIt would be interesting to add more analysis on the patterns of learned mixing matrix M for different tasks.\n5)\tIn the Figure 3, the training time of the standard MHA with D_k = 256 is 0.0?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting collaborative MHA mechanism", "review": "This paper presents an interesting collaborative MHA to enable heads to share projections, which can be easily applied to most existing transformer-based models, including NMT and pre-training models. With using the new collaborative MHA, the number of parameters and FLOPs can be decreased. \n\nThe PCA-based analysis of the query/key matrices is interesting and impressive, which motivates the newly proposed MHA. The authors also propose a Tensor Decomposition based method to easily convert MHA to its collaborative version without retraining.    \n\nThe paper is well-written and organized, the experiments are thorough. However, I have several concerns:\n1)\tIn the Table 1, it will be more convincing, if the run time on CPU and GPU can be provided. \n2)\tAll the pre-trained models (BERT/DistilBERT/ALBERT) are evaluated on the GLUE dataset, the experiments on more challenging tasks like QA (e.g. SQuAD 1.1/2.0) should be added. \n3)\tThe proposed method seems to be not effective for pre-trained models, e.g. when the number of parameters is decreased from 108.5M to 96.6M, this reduction of model size is not that big, while the average score decreases from 83.2 to 77.6. \n4)\tIt would be interesting to add more analysis on the patterns of learned mixing matrix M for different tasks.\n5)\tIn the Figure 3, the training time of the standard MHA with D_k = 256 is 0.0?", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603895100969}, {"id": "NC8EHArn2nJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3019/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper investigates the over-parameterization of attention heads in Transformer’s multi-head attention. The authors show that query-key projections are redundant because trained concatenated heads tend to compute their attention patterns on common features. They propose a reparameterization of multi-head attention allowing the parameters of queries and keys to be shared between heads: this is called “collaborative attention”. This attention can be applied either instead of the standard attention during training, or as a drop-in replacement for an already trained model. To use as a drop-in replacement, the method requires to use tensor decomposition and subsequent model fine-tuning.\n\n------------------------------------------------------------------------------------------------------\nStrengths\n\n1) a nice analysis of PCA components showing that individual heads are not low-rank, but their concatenation is.\n2) the paper is overall clear and the method is explained well.\n\n------------------------------------------------------------------------------------------------------\nWeaknesses\n\n(main) While the main contribution is a more efficient attention layer without a significant drop in performance, this claim is not supported empirically. Since the method operated only within attention layers (reduces only dimensions of queries and keys), in terms of efficiency/quality trade-off it should be compared to other methods, e.g. simple head pruning. While the paper does not provide such comparison, it is clear from the results that the simple head pruning is likely to be superior (and is simpler implementation-wise). \n\nNamely, \n1) when used as a drop-in replacement, the method is more complicated than a simple head pruning, but not more effective. E.g., the proposed method reduces query/key dimension by a factor of 1.5/2/3 and keeps everything else intact. With simple head pruning, about 50% of the heads can be removed without sacrificing quality: in addition to queries and keys, this also halves values and the output projection matrix. \n2) when used in training (see the MT experiments), the results are also not better than post-hoc head pruning. The argument to support this new approach could be that the model is smaller in training, but since the parameter reduction is only in queries and keys, the decrease in the number of parameters is negligible for the whole model. For example, keeping the same quality it reduces the number of parameters from 6.1*10^6 to 5.4*10^6, which is not going to make a large difference.\n\n(minor) As a side contribution, the authors claim to report the discrepancy between the theory and implementation of attention layers. Namely, while the original definition of attention layers does not have biases in linear projections for q/k/v in attention, the authors claim that implementations contain the bias terms, and spend some time showing how to model the biases in key and query layers properly.\n\nHowever, the original Transformer implementation (tensor2tensor) does *not* have biases by default. This means that the authors are probably referring to some specific implementation, different from the original one provided by the Transformer’s authors. Therefore, I can not consider this as a contribution and think that this part is misleading for a reader.\n\nP.S. Here is the tensor2tensor code I was referring to: https://github.com/tensorflow/tensor2tensor/blob/5f9dd2db6d7797162e53adf152310ed13e9fc711/tensor2tensor/layers/common_attention.py#L4416\n\n------------------------------------------------------------------------------------------------------\nOverall recommendation\n\nOverall, I can not recommend accepting this paper. While there are some parts of the paper that I like, the main claim is not supported empirically: both in terms of baselines and the overall decrease in the number of parameters. Additionally, the part with the biases in attention implementation is misleading.\n\n------------------------------------------------------------------------------------------------------\nUpdate after author response\n------------------------------------------------------------------------------------------------------\n\n1) Context and content attention\n\nThank you for updating and saying that only some of the implementations include bias! I think now this part is not misleading and can be of interest. I still have some concerns that this part does not fit the whole story very well - but this is the matter of taste. In the current state, I think it is ok :)\n\n2) On the comparison with head pruning and on the paper going beyond practical realm.\n\nI agree with your comments, but I do think you should make it very clear in the paper. In the current state, the paper tried to make practical contributions and, since they mostly do not hold (e.g., head pruning is simpler in practice), it's hard to appreciate the paper's value. I think you need to modify the things you highlight, and with proper discussion it would be much better. For example, if you state explicitly that in practice pruning may be simpler, but your results say/illustrate something other than practical applications. You won't lose because of it; in fact, I think the opposite.\n\nOverall, I think the paper has improved during the discussion period. In a hope that the authors address my later comments and discuss the pruning in the text, I'm raising the score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A nice paper, but has problems with supporting the claims and some parts are misleading", "review": "The paper investigates the over-parameterization of attention heads in Transformer’s multi-head attention. The authors show that query-key projections are redundant because trained concatenated heads tend to compute their attention patterns on common features. They propose a reparameterization of multi-head attention allowing the parameters of queries and keys to be shared between heads: this is called “collaborative attention”. This attention can be applied either instead of the standard attention during training, or as a drop-in replacement for an already trained model. To use as a drop-in replacement, the method requires to use tensor decomposition and subsequent model fine-tuning.\n\n------------------------------------------------------------------------------------------------------\nStrengths\n\n1) a nice analysis of PCA components showing that individual heads are not low-rank, but their concatenation is.\n2) the paper is overall clear and the method is explained well.\n\n------------------------------------------------------------------------------------------------------\nWeaknesses\n\n(main) While the main contribution is a more efficient attention layer without a significant drop in performance, this claim is not supported empirically. Since the method operated only within attention layers (reduces only dimensions of queries and keys), in terms of efficiency/quality trade-off it should be compared to other methods, e.g. simple head pruning. While the paper does not provide such comparison, it is clear from the results that the simple head pruning is likely to be superior (and is simpler implementation-wise). \n\nNamely, \n1) when used as a drop-in replacement, the method is more complicated than a simple head pruning, but not more effective. E.g., the proposed method reduces query/key dimension by a factor of 1.5/2/3 and keeps everything else intact. With simple head pruning, about 50% of the heads can be removed without sacrificing quality: in addition to queries and keys, this also halves values and the output projection matrix. \n2) when used in training (see the MT experiments), the results are also not better than post-hoc head pruning. The argument to support this new approach could be that the model is smaller in training, but since the parameter reduction is only in queries and keys, the decrease in the number of parameters is negligible for the whole model. For example, keeping the same quality it reduces the number of parameters from 6.1*10^6 to 5.4*10^6, which is not going to make a large difference.\n\n(minor) As a side contribution, the authors claim to report the discrepancy between the theory and implementation of attention layers. Namely, while the original definition of attention layers does not have biases in linear projections for q/k/v in attention, the authors claim that implementations contain the bias terms, and spend some time showing how to model the biases in key and query layers properly.\n\nHowever, the original Transformer implementation (tensor2tensor) does *not* have biases by default. This means that the authors are probably referring to some specific implementation, different from the original one provided by the Transformer’s authors. Therefore, I can not consider this as a contribution and think that this part is misleading for a reader.\n\nP.S. Here is the tensor2tensor code I was referring to: https://github.com/tensorflow/tensor2tensor/blob/5f9dd2db6d7797162e53adf152310ed13e9fc711/tensor2tensor/layers/common_attention.py#L4416\n\n------------------------------------------------------------------------------------------------------\nOverall recommendation\n\nOverall, I can not recommend accepting this paper. While there are some parts of the paper that I like, the main claim is not supported empirically: both in terms of baselines and the overall decrease in the number of parameters. Additionally, the part with the biases in attention implementation is misleading.\n\n------------------------------------------------------------------------------------------------------\nUpdate after author response\n------------------------------------------------------------------------------------------------------\n\n1) Context and content attention\n\nThank you for updating and saying that only some of the implementations include bias! I think now this part is not misleading and can be of interest. I still have some concerns that this part does not fit the whole story very well - but this is the matter of taste. In the current state, I think it is ok :)\n\n2) On the comparison with head pruning and on the paper going beyond practical realm.\n\nI agree with your comments, but I do think you should make it very clear in the paper. In the current state, the paper tried to make practical contributions and, since they mostly do not hold (e.g., head pruning is simpler in practice), it's hard to appreciate the paper's value. I think you need to modify the things you highlight, and with proper discussion it would be much better. For example, if you state explicitly that in practice pruning may be simpler, but your results say/illustrate something other than practical applications. You won't lose because of it; in fact, I think the opposite.\n\nOverall, I think the paper has improved during the discussion period. In a hope that the authors address my later comments and discuss the pruning in the text, I'm raising the score.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603801837403}], "openreview_url": "https://openreview.net/forum?id=bK-rJMKrOsm", "arxiv_id": "2006.16362", "paper_pdf": "papers/bK-rJMKrOsm.pdf", "paper_pdf_sha256": "2c090efed555025bf1ec70aaaf02df4a259d6752ce8c48947808eacf1d1dc5da", "paper_pdf_bytes": 1342898, "paper_pdf_source": "openreview", "code_url": "https://github.com/epfml/collaborative-attention", "code_repository": "epfml/collaborative-attention", "code_commit": "5b05408640e339970f1009e06cac678e2a1ab534", "code_archive": "repos/bK-rJMKrOsm.zip", "code_archive_sha256": "2e38005b5627bae9e850a705ea04bfdbe629d2b0a890af1ebb5c512581daf089", "code_archive_bytes": 23532, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 49, "github_languages": {"Python": 35139}, "github_archived": false, "github_pushed_at": "2023-06-12T21:27:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-head-attention-collaborate-instead-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1xKBCEYDr", "year": 2020, "status": "rejected", "title": "Black-box Adversarial Attacks with Bayesian Optimization", "authors": ["Satya Narayan Shukla", "Anit Kumar Sahu", "Devin Willmott", "J. Zico Kolter"], "authorids": ["snshukla@cs.umass.edu", "anit.sahu@gmail.com", "devin.willmott@uky.edu", "zkolter@cs.cmu.edu"], "authors_source": "OpenReview API", "abstract": "We focus on the problem of black-box adversarial attacks, where the aim is to generate adversarial examples using information limited to loss function evaluations of input-output pairs. We use Bayesian optimization (BO) to specifically\ncater to scenarios involving low query budgets to develop query efficient adversarial attacks. We alleviate the issues surrounding BO in regards to optimizing high dimensional deep learning models by effective dimension upsampling techniques. Our proposed approach achieves performance comparable to the state of the art black-box adversarial attacks albeit with a much lower average query count. In particular, in low query budget regimes, our proposed method reduces the query count up to 80% with respect to the state of the art methods.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyljiA_RFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1118/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper applies Bayesian optimisation (BO), a sample efficient global optimisation technique, to the problem of finding adversarial perturbation. First, the application is starightforward application of BO to the well-known problem of adversarial perturbation. Nothing innovative. Second, the paper addresses the high dimensional optimisation with simple upsampling technique like nearest-neighbour, without even trying their hands dirty by using one of many many high-dimensional Bayesian optimisation algorithm (a quick Google search will reveal them), The work  thus fail in thoroughness also. Third, adversarial perturbation are known to exist even around the image such that even a simple gradient descet optimisation starting from the target image would be able to provide perceptually small perturbation (it does not have to the smallest to be perceptually small). Hence, the impact is also missing. Thus accroding to me this paper is not good enough for acceptance. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper applies Bayesian optimisation (BO), a sample efficient global optimisation technique, to the problem of finding adversarial perturbation. First, the application is starightforward application of BO to the well-known problem of adversarial perturbation. Nothing innovative. Second, the paper addresses the high dimensional optimisation with simple upsampling technique like nearest-neighbour, without even trying their hands dirty by using one of many many high-dimensional Bayesian optimisation algorithm (a quick Google search will reveal them), The work  thus fail in thoroughness also. Third, adversarial perturbation are known to exist even around the image such that even a simple gradient descet optimisation starting from the target image would be able to provide perceptually small perturbation (it does not have to the smallest to be perceptually small). Hence, the impact is also missing. Thus accroding to me this paper is not good enough for acceptance. "}, "tcdate": 1571880610546}, {"id": "rkgOQMOAtS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1118/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a BO-based black-box attack generation method. In general, it is very well written and easy to follow. The main contribution is to combine BO with dimension reduction, which leads to the effectiveness in generating black-box adversarial examples in the regime of limited queries.  However, I still have some concerns about this paper. \n\n1) The benefits of BO? It seems that the step of dimension reduction is crucial to make BO scablable to high-dimensional problems. I wonder if the gradient estimation-based attack methods can apply the similar trick and yield the similar performance. That is, one can solve problem min_{\\delta} attack_loss( x, y, g(\\delta) ) by estimating gradients via finite-difference of function values, where g(\\cdot) is the dimension-reduction operator, and \\delta is the low-dimensional perturbation. Such a baseline is not clear in the paper, and the comparison with (Tu et al., 2019) is not provided in the paper. \n\n2) Moreover, in experiments, it seems that only query-efficiency was reported. What about distortion-efficiency for BO-based attack? For $\\ell_\\infty$-attacks, the other $\\ell_p$ norms can be used as distortion metrics. I wonder what perturbation does the BO method converge to. It was shown in (https://arxiv.org/pdf/1907.11684.pdf, Table 1) that BO usually leads to larger \\ell_1 and \\ell_2 distortion. \n\n3) It might be useful to show the convergence of BO in terms of objective value versus iterations/queries. This may give a clearer picture on how BO works in the attack generation setting. \n\n4) Minor comment: In related work \"Bayesian optimization has played a supporting role in several methods,\nincluding Tu et al. (2019), where ....\" However,  Tu et al. (2019) does not seem using BO and ADMM. \n\n############ Post-feedback ##########\nThanks for the clarification and the additional experiments. I am satisfied with the response, and have increased my score to 6.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "This paper proposed a BO-based black-box attack generation method. In general, it is very well written and easy to follow. The main contribution is to combine BO with dimension reduction, which leads to the effectiveness in generating black-box adversarial examples in the regime of limited queries.  However, I still have some concerns about this paper. \n\n1) The benefits of BO? It seems that the step of dimension reduction is crucial to make BO scablable to high-dimensional problems. I wonder if the gradient estimation-based attack methods can apply the similar trick and yield the similar performance. That is, one can solve problem min_{\\delta} attack_loss( x, y, g(\\delta) ) by estimating gradients via finite-difference of function values, where g(\\cdot) is the dimension-reduction operator, and \\delta is the low-dimensional perturbation. Such a baseline is not clear in the paper, and the comparison with (Tu et al., 2019) is not provided in the paper. \n\n2) Moreover, in experiments, it seems that only query-efficiency was reported. What about distortion-efficiency for BO-based attack? For $\\ell_\\infty$-attacks, the other $\\ell_p$ norms can be used as distortion metrics. I wonder what perturbation does the BO method converge to. It was shown in (https://arxiv.org/pdf/1907.11684.pdf, Table 1) that BO usually leads to larger \\ell_1 and \\ell_2 distortion. \n\n3) It might be useful to show the convergence of BO in terms of objective value versus iterations/queries. This may give a clearer picture on how BO works in the attack generation setting. \n\n4) Minor comment: In related work \"Bayesian optimization has played a supporting role in several methods,\nincluding Tu et al. (2019), where ....\" However,  Tu et al. (2019) does not seem using BO and ADMM. \n\n############ Post-feedback ##########\nThanks for the clarification and the additional experiments. I am satisfied with the response, and have increased my score to 6.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571877407798}, {"id": "rylgel1fYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1118/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an idea on making adversarial attack on deep learning model. Since the space of input-output for the adversarial attach is huge, the paper proposes to use Bayesian optimization (BO) to sequentially select an attack.\n\nAlthough the potential application of adversarial attack on deep learning model is interesting, the paper contribution and the novelty are limited giving the fact that there is another related paper published [1].\n\nThe authors in [1] consider using Bayesian optimization to make adversarial attack for model testing. In particular, they have considered the deep learning model. Then, they extend to multi-task settings. There is a big overlapping between the idea in [1] and the current paper.\n\nThe paper presentation and writing is high quality although the paper is a bit over-length.\n\n[1] Gopakumar, Shivapratap, et al. \"Algorithmic assurance: an active approach to algorithmic testing using Bayesian optimisation.\" Advances in Neural Information Processing Systems. 2018.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper presents an idea on making adversarial attack on deep learning model. Since the space of input-output for the adversarial attach is huge, the paper proposes to use Bayesian optimization (BO) to sequentially select an attack.\n\nAlthough the potential application of adversarial attack on deep learning model is interesting, the paper contribution and the novelty are limited giving the fact that there is another related paper published [1].\n\nThe authors in [1] consider using Bayesian optimization to make adversarial attack for model testing. In particular, they have considered the deep learning model. Then, they extend to multi-task settings. There is a big overlapping between the idea in [1] and the current paper.\n\nThe paper presentation and writing is high quality although the paper is a bit over-length.\n\n[1] Gopakumar, Shivapratap, et al. \"Algorithmic assurance: an active approach to algorithmic testing using Bayesian optimisation.\" Advances in Neural Information Processing Systems. 2018.\n\n"}, "tcdate": 1571053543550}], "openreview_url": "https://openreview.net/forum?id=H1xKBCEYDr", "arxiv_id": "1909.13857", "paper_pdf": "papers/H1xKBCEYDr.pdf", "paper_pdf_sha256": "a5807cf9e806a8e545d93e0bb80c00f7b6b99809ed294526c61975f6c573948e", "paper_pdf_bytes": 689993, "paper_pdf_source": "openreview", "code_url": "https://github.com/snu-mllab/parsimonious-blackbox-attack", "code_repository": "snu-mllab/parsimonious-blackbox-attack", "code_commit": "9061d845b161fc100933c38b74c98b8b9375b80a", "code_archive": "repos/H1xKBCEYDr.zip", "code_archive_sha256": "99e7f1364ac6977b8893b249a41250653f95a102ab46fd4936eab91407f3f9d1", "code_archive_bytes": 72077, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 69, "github_languages": {"Python": 94094}, "github_archived": false, "github_pushed_at": "2020-12-07T04:45:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/black-box-adversarial-attacks-with-bayesian"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJeKCi0qYX", "year": 2019, "status": "rejected", "title": "MILE: A Multi-Level Framework for Scalable Graph Embedding", "authors": ["Jiongqian Liang", "Saket Gurukar", "Srinivasan Parthasarathy"], "authorids": ["liang.albert@outlook.com", "gurukar.1@osu.edu", "srini@cse.ohio-state.edu"], "authors_source": "OpenReview API", "abstract": "Recently there has been a surge of interest in designing graph embedding methods. Few, if any, can scale to a large-sized graph with millions of nodes due to both computational complexity and memory requirements. In this paper, we relax this limitation by introducing the MultI-Level Embedding (MILE) framework – a generic methodology allowing contemporary graph embedding methods to scale to large graphs. MILE repeatedly coarsens the graph into smaller ones using a hybrid matching technique to maintain the backbone structure of the graph. It then applies existing embedding methods on the coarsest graph and refines the embeddings to the original graph through a novel graph convolution neural network that it learns. The proposed MILE framework is agnostic to the underlying graph embedding techniques and can be applied to many existing graph embedding methods without modifying them. We employ our framework on several popular graph embedding techniques and conduct embedding for real-world graphs. Experimental results on five large-scale datasets demonstrate that MILE significantly boosts the speed (order of magnitude) of graph embedding while also often generating embeddings of better quality for the task of node classification. MILE can comfortably scale to a graph with 9 million nodes and 40 million edges, on which existing methods run out of memory or take too long to compute on a modern workstation.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1g_pHZi27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper916/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a multi-level embedding (MILE) framework, which can be applied on top of existing network embedding methods and helps them scale to large scale networks with faster speed. To get the backbone structure of graph, MILE repeatedly coarsens the graph into smaller ones using a hybrid matching technique, and GCN is used for the refinement of embeddings.\n\n[+] The paper is well-written and the idea is clearly presented.\n[+] MILE is able to reduce computational cost while achieving comparable, or sometimes even better embedding quality. \n[+] MILE is general enough to apply to different underlying embedding strategies.\n[-] Most of the baseline methods are of similar type, since LINE, DeepWalk, node2vec and NetMF can all be unified to matrix factorization framework. There have been many new network embedding methods proposed in the past two years. It would be interesting to see how much MILE can help scale these methods.\n\nOverall, though there have already been hundreds of papers on network embedding in the past 2~3 years, I think this paper can be an interesting addition to this fast-growing area. Therefore, I would recommend to accept it.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall Interesting work: clear motivation and nice performance gain", "review": "This paper proposes a multi-level embedding (MILE) framework, which can be applied on top of existing network embedding methods and helps them scale to large scale networks with faster speed. To get the backbone structure of graph, MILE repeatedly coarsens the graph into smaller ones using a hybrid matching technique, and GCN is used for the refinement of embeddings.\n\n[+] The paper is well-written and the idea is clearly presented.\n[+] MILE is able to reduce computational cost while achieving comparable, or sometimes even better embedding quality. \n[+] MILE is general enough to apply to different underlying embedding strategies.\n[-] Most of the baseline methods are of similar type, since LINE, DeepWalk, node2vec and NetMF can all be unified to matrix factorization framework. There have been many new network embedding methods proposed in the past two years. It would be interesting to see how much MILE can help scale these methods.\n\nOverall, though there have already been hundreds of papers on network embedding in the past 2~3 years, I think this paper can be an interesting addition to this fast-growing area. Therefore, I would recommend to accept it.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541244351864}, {"id": "BJldyjav2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper916/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this submission, the authors propose a three-stage framework for large-scale graph embedding. The proposed method first constructs a small graph by graph coarsening, then applies any existing graph embedding method, and last refines the learned embeddings. It is useful, however, the experimental results are not convincing and cannot support the authors' claims about the proposed method.\n\nFirst, in many places, the authors claim that the embedding quality of the proposed method is improved. For example, the last sentence of Section 1, and \"MILE improves quality\" paragraph on Page 7. However, the experimental results fail to support this. As the proposed method is for the large-scale graph, let's focus on the results of YouTube dataset and Yelp dataset first. For Youtube dataset ((d) of Table 2), when m is set to be 8, for all the cases, the performance drops. For Yelp dataset (Figure 3), the authors do not provide Micro-f1 for the original graph (m = 0) or m = 1, 2, so it is hard or impossible to demonstrate that the quality of the proposed method is still good. \n\nSecond, the comparison with existing methods is not sufficient. For the most important Yelp dataset (as this dataset fits the motivation scenario (large-scale graph) of this submission), the authors fail to report any comparison. Thus it might not be weak to demonstrate the benefit of the proposed method.\n\nThird, some experiment details are missing. For example, how the authors compute the running time of the proposed method? All the three stages are included? How the authors implement the existing methods? Are these implementations good enough to ensure a fair comparison? \n\n*******\nSome other questions:\na) On page 2, the authors mention that the proposed method \"can be easily extended to directed graph\". However, based on my understanding, directly graph will affect both the graph coarsening and embedding refining steps, and it seems not so easy to extend. Do the authors have the solution and experiments for directed graph? It would be interesting to see such results, which enlarges the application scope of the proposed method.\n\nb) The toy example on page 3 is very clear. However, for real-world graphs, does the proposed graph coarsening work well? For example, one property the proposed method utilizes is \"structurally equivalent\". What is the percentage of the nodes that can have such property for real-world graphs? \n\n********\nSome other comments:\nGenerally speaking, this submission studies a very practical task. Although the authors claim that the proposed method has great efficiency while the embedding quality is comparable good or even better than the existing methods, I think that there is an efficiency-quality trade-off based on the experimental results in this submission. When m increases, the graph coarsening step causes more information loss, and thus the quality may decrease. Embedding refining step can be regarded as a procedure to reduce such information loss, but may not improve the embedding quality better than the original graph. So to me, it would be more meaningful to study such efficiency-quality trade-off for large-scale graph embedding.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Practically useful, but experiments are not convincing", "review": "In this submission, the authors propose a three-stage framework for large-scale graph embedding. The proposed method first constructs a small graph by graph coarsening, then applies any existing graph embedding method, and last refines the learned embeddings. It is useful, however, the experimental results are not convincing and cannot support the authors' claims about the proposed method.\n\nFirst, in many places, the authors claim that the embedding quality of the proposed method is improved. For example, the last sentence of Section 1, and \"MILE improves quality\" paragraph on Page 7. However, the experimental results fail to support this. As the proposed method is for the large-scale graph, let's focus on the results of YouTube dataset and Yelp dataset first. For Youtube dataset ((d) of Table 2), when m is set to be 8, for all the cases, the performance drops. For Yelp dataset (Figure 3), the authors do not provide Micro-f1 for the original graph (m = 0) or m = 1, 2, so it is hard or impossible to demonstrate that the quality of the proposed method is still good. \n\nSecond, the comparison with existing methods is not sufficient. For the most important Yelp dataset (as this dataset fits the motivation scenario (large-scale graph) of this submission), the authors fail to report any comparison. Thus it might not be weak to demonstrate the benefit of the proposed method.\n\nThird, some experiment details are missing. For example, how the authors compute the running time of the proposed method? All the three stages are included? How the authors implement the existing methods? Are these implementations good enough to ensure a fair comparison? \n\n*******\nSome other questions:\na) On page 2, the authors mention that the proposed method \"can be easily extended to directed graph\". However, based on my understanding, directly graph will affect both the graph coarsening and embedding refining steps, and it seems not so easy to extend. Do the authors have the solution and experiments for directed graph? It would be interesting to see such results, which enlarges the application scope of the proposed method.\n\nb) The toy example on page 3 is very clear. However, for real-world graphs, does the proposed graph coarsening work well? For example, one property the proposed method utilizes is \"structurally equivalent\". What is the percentage of the nodes that can have such property for real-world graphs? \n\n********\nSome other comments:\nGenerally speaking, this submission studies a very practical task. Although the authors claim that the proposed method has great efficiency while the embedding quality is comparable good or even better than the existing methods, I think that there is an efficiency-quality trade-off based on the experimental results in this submission. When m increases, the graph coarsening step causes more information loss, and thus the quality may decrease. Embedding refining step can be regarded as a procedure to reduce such information loss, but may not improve the embedding quality better than the original graph. So to me, it would be more meaningful to study such efficiency-quality trade-off for large-scale graph embedding.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541032671610}, {"id": "ByxAk8WMnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper916/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a multi-Level framework for learning node embeddings for large-scale graphs. The author first coarsens the graphs into different levels of subgraphs. The low-level subgraphs are obtained with the node embeddings of the higher-level graphs with a graph convolutional neural network. By iteratively applying this procedure, the node embeddings of the original graphs can be obtained. Experimental results on several networks (including one network with ~10M node) prove the effective and efficiency of the proposed method over existing state-of-the-art approaches.   \n\nStrength:\n- scaling up node embedding methods is a very important and practical problem\n- experiments show that the proposed methods seems to be very effective. \nWeakness:\n- the proposed method seems to be very heuristic\n- some claims in the papers are wrong according to existing literatures\n\nOverall, the paper is well written and easy to follow. The proposed method is simple but heuristic.  However, the performance seems to be quite effective according to the experiments. The reasons that why the method works need to be better explained, which can significantly the quality of the paper and its impact in the future.\n\nDetails:\n-- In the introduction part, \"However, such methods rarely scale to large datasets (e.g., graphs with over 1 million nodes) since they are computationally expensive and often memory intensive\". This is not TRUE! In the paper of LINE (Tang et al. 2015). It shows the LINE model can easily scale up to networks with one million nodes with a few hours. \n-- The authors use Equation (7) to learn the parameters of the graph convolutional neural network. I am really surprised that this method works. Especially the learned parameters are shared across different layers. \n-- Have you tried and compared different approaches of graph coarsening?\n-- In Figure 2. (a), according to Equation (1), in the second step, the weight of the edge between A and DE should be 2/sqrt(3)*sqrt(4)?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea and result", "review": "This paper proposed a multi-Level framework for learning node embeddings for large-scale graphs. The author first coarsens the graphs into different levels of subgraphs. The low-level subgraphs are obtained with the node embeddings of the higher-level graphs with a graph convolutional neural network. By iteratively applying this procedure, the node embeddings of the original graphs can be obtained. Experimental results on several networks (including one network with ~10M node) prove the effective and efficiency of the proposed method over existing state-of-the-art approaches.   \n\nStrength:\n- scaling up node embedding methods is a very important and practical problem\n- experiments show that the proposed methods seems to be very effective. \nWeakness:\n- the proposed method seems to be very heuristic\n- some claims in the papers are wrong according to existing literatures\n\nOverall, the paper is well written and easy to follow. The proposed method is simple but heuristic.  However, the performance seems to be quite effective according to the experiments. The reasons that why the method works need to be better explained, which can significantly the quality of the paper and its impact in the future.\n\nDetails:\n-- In the introduction part, \"However, such methods rarely scale to large datasets (e.g., graphs with over 1 million nodes) since they are computationally expensive and often memory intensive\". This is not TRUE! In the paper of LINE (Tang et al. 2015). It shows the LINE model can easily scale up to networks with one million nodes with a few hours. \n-- The authors use Equation (7) to learn the parameters of the graph convolutional neural network. I am really surprised that this method works. Especially the learned parameters are shared across different layers. \n-- Have you tried and compared different approaches of graph coarsening?\n-- In Figure 2. (a), according to Equation (1), in the second step, the weight of the edge between A and DE should be 2/sqrt(3)*sqrt(4)?", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1540654565861}], "openreview_url": "https://openreview.net/forum?id=HJeKCi0qYX", "arxiv_id": "1802.09612", "paper_pdf": "papers/HJeKCi0qYX.pdf", "paper_pdf_sha256": "a635b28a127ea382124e28578ab2e7375d43838be9614294e9b425f414a4fe1c", "paper_pdf_bytes": 960423, "paper_pdf_source": "openreview", "code_url": "https://github.com/jiongqian/MILE", "code_repository": "jiongqian/MILE", "code_commit": "47940a33ab9bbb452af88fbc863020bdbe924fa2", "code_archive": "repos/HJeKCi0qYX.zip", "code_archive_sha256": "1ec4d3bc1c09e396071700a7af620d152daf891c95ee7bb30d7c4efc034109f1", "code_archive_bytes": 350324, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 432, "github_languages": {"Python": 55560}, "github_archived": false, "github_pushed_at": "2020-07-19T19:41:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mile-a-multi-level-framework-for-scalable"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJlrSmbAZ", "year": 2018, "status": "rejected", "title": "Bayesian Uncertainty Estimation for Batch Normalized Deep Networks", "authors": ["Mattias Teye", "Hossein Azizpour", "Kevin Smith"], "authorids": ["teye@kth.se", "azizpour@kth.se", "ksmith@kth.se"], "authors_source": "OpenReview API", "abstract": "Deep neural networks have led to a series of breakthroughs, dramatically improving the state-of-the-art in many domains. The techniques driving these advances, however, lack a formal method to account for model uncertainty. While the Bayesian approach to learning provides a solid theoretical framework to handle uncertainty, inference in Bayesian-inspired deep neural networks is difficult. In this paper, we provide a practical approach to Bayesian learning that relies on a regularization technique found in nearly every modern network, batch normalization. We show that training a deep network using batch normalization is equivalent to approximate inference in Bayesian models, and we demonstrate how this finding allows us to make useful estimates of the model uncertainty. Using our approach, it is possible to make meaningful uncertainty estimates using conventional architectures without modifying the network or the training procedure. Our approach is thoroughly validated in a series of empirical experiments on different tasks and using various measures, showing it to outperform baselines on a majority of datasets with strong statistical significance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bk8cjwFgz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1165/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "*Summary*\n\nThe paper proposes using batch normalisation at test time to get the predictive uncertainty. The stochasticity of the prediction comes from different minibatches of training data that were used to normalise the activity/pre-activation values at each layer. This is justified by an argument that using batch norm is doing variational inference, so one should use the approximate posterior provided by batch norm at prediction time. Several experiments show Monte Carlo prediction at test time using batch norm is better than dropout.\n\n*Originality and significance*\n\nAs far as I understand, almost learning algorithms similar to equation 2 can be recast as variational inference under equation 1. However, the critical questions are what is the corresponding prior, what is the approximating density, what are the additional approximations to obtain 2, and whether the approximation is a good approximation for getting closer to the posterior/obtain better prediction. \n\nIt is not clear to me from the presentation what the q(w) density is -- whether this is explicit (as in vanilla Gaussian VI or MC dropout), or implicit (the stochasticity on the activity h due to batch norm induces an equivalence q on w).\n\nFrom a Bayesian perspective, it is also not satisfying to ignore the regularisation term by an empirical heuristic provided in the batch norm paper [small \\lambda] -- what is the rationale of this? Can this be explained by comparing the variational free-energy. \n\nThe experiments also do not compare to modern variational inference methods using the reparameterisation trick with Gaussian variational approximations (see Blundell et al 2016) or richer variational families (see e.g. Louizos and Welling, 2016, 2017). The VI method included in the PBP paper (Hernandez-Lobato and Adams, 2015) does not use the reparameterisation trick, which has been found to reduce variance and improve over Graves' VI method.\n\n*Clarity*\nThe paper is in general well written and easy to understand. \n\n*Additional comments*\n\nPage 2: Monte Carlo Droput --> Dropout\nPage 3 related work: (Adams, 2015) should be (Hernandez-Lobato and Adams, 2015)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Using Batch norm at test time to obtain uncertainty estimate", "rating": "5: Marginally below acceptance threshold", "review": "*Summary*\n\nThe paper proposes using batch normalisation at test time to get the predictive uncertainty. The stochasticity of the prediction comes from different minibatches of training data that were used to normalise the activity/pre-activation values at each layer. This is justified by an argument that using batch norm is doing variational inference, so one should use the approximate posterior provided by batch norm at prediction time. Several experiments show Monte Carlo prediction at test time using batch norm is better than dropout.\n\n*Originality and significance*\n\nAs far as I understand, almost learning algorithms similar to equation 2 can be recast as variational inference under equation 1. However, the critical questions are what is the corresponding prior, what is the approximating density, what are the additional approximations to obtain 2, and whether the approximation is a good approximation for getting closer to the posterior/obtain better prediction. \n\nIt is not clear to me from the presentation what the q(w) density is -- whether this is explicit (as in vanilla Gaussian VI or MC dropout), or implicit (the stochasticity on the activity h due to batch norm induces an equivalence q on w).\n\nFrom a Bayesian perspective, it is also not satisfying to ignore the regularisation term by an empirical heuristic provided in the batch norm paper [small \\lambda] -- what is the rationale of this? Can this be explained by comparing the variational free-energy. \n\nThe experiments also do not compare to modern variational inference methods using the reparameterisation trick with Gaussian variational approximations (see Blundell et al 2016) or richer variational families (see e.g. Louizos and Welling, 2016, 2017). The VI method included in the PBP paper (Hernandez-Lobato and Adams, 2015) does not use the reparameterisation trick, which has been found to reduce variance and improve over Graves' VI method.\n\n*Clarity*\nThe paper is in general well written and easy to understand. \n\n*Additional comments*\n\nPage 2: Monte Carlo Droput --> Dropout\nPage 3 related work: (Adams, 2015) should be (Hernandez-Lobato and Adams, 2015)", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511779213669}, {"id": "Bkw2_15xz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1165/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors show how the regularization procedure called batch normalization,\ncurrently being used by most deep learning systems, can be understood as\nperforming approximate Bayesian inference. The authors compare this approach to\nMonte Carlo dropout (another regularization technique which can also be\nconsidered to perform approximate Bayesian inference). The experiments\nperformed show that the Bayesian view of batch normalization performs similarly\nas MC dropout in terms of the estimates of uncertainty that it produces.\n\nQuality:\n\nI found the quality to be low in some aspects. First, the description of what\nis the prior used by batch normalization in section 3.3 is unsatisfactory. The\nauthors basically refer to Appendix 6.4 for the case in which the weight decay\npenalty is not zero. The details in that Appendix are almost none, they just\nsay \"it is thus possible to derive the prior...\".\n\nThe results in Table 2 are a bit confusing. The authors should highlight in\nbold face the results of the best performing method.\n\nThe authors indicate that they do not need to compare to variational methods\nbecause Gal and Ghahramani 2015 compare already to those methods. However, Gal\nand Ghahramani's code used Bayesian optimization methods to tune\nhyper-parameters and this code contains a bug that optimizes hyper-parameters\nby maximizing performance on the test data. In particular for hyperparameter\nselection, they average performance across (subsets of) 5 of the training sets\nfrom the 20x train/test split, and then using the tau which got the best\naverage performance for all of 20x train/test splits to evaluate performance:\n\nhttps://github.com/yaringal/DropoutUncertaintyExps/blob/master/bostonHousing/net/experiment_BO.py#L54\n\nTherefore, the claim that \n\n\"Since we have established that MCBN performs on par with MCDO, by proxy we\nmight conclude that MCBN outperforms those VI methods as well.\"\n\nis not valid.\n\nAt the beginning of section 4.3 the authors indicate that they follow in their\nexperiments the setup of Gal and Ghahramani (2015). However, Gal and Ghahramani\n(2015) actually follow Hernández-Lobato and Adams, 2015 so the correct\nreference should be the latter one.\n\nClarity:\n\nThe paper is clearly written and easy to follow and understand.\n\nI found confusing how to use the proposed method to obtain estimates of\nuncertainty for a particular test data point x_star. The paragraph just above\nsection 4 says that the authors sample a batch of training data for this, but\nassume that the test point x_star has to be included in this batch.\nHow is this actually done in practice?\n\nOriginality:\n\nThe proposed contribution is original. This is the first time that a Bayesian\ninterpretation has been given to the batch normalization regularization\nproposal.\n\nSignificance:\n\nThe paper's contributions are significant. Batch normalization is a very\npopular regularization technique and showing that it can be used to obtain\nestimates of uncertainty is relevant and significant. Many existing deep\nlearning systems can use this to produce estimates of uncertainty in their\npredictions.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and relevant but lack of details on prior", "rating": "6: Marginally above acceptance threshold", "review": "The authors show how the regularization procedure called batch normalization,\ncurrently being used by most deep learning systems, can be understood as\nperforming approximate Bayesian inference. The authors compare this approach to\nMonte Carlo dropout (another regularization technique which can also be\nconsidered to perform approximate Bayesian inference). The experiments\nperformed show that the Bayesian view of batch normalization performs similarly\nas MC dropout in terms of the estimates of uncertainty that it produces.\n\nQuality:\n\nI found the quality to be low in some aspects. First, the description of what\nis the prior used by batch normalization in section 3.3 is unsatisfactory. The\nauthors basically refer to Appendix 6.4 for the case in which the weight decay\npenalty is not zero. The details in that Appendix are almost none, they just\nsay \"it is thus possible to derive the prior...\".\n\nThe results in Table 2 are a bit confusing. The authors should highlight in\nbold face the results of the best performing method.\n\nThe authors indicate that they do not need to compare to variational methods\nbecause Gal and Ghahramani 2015 compare already to those methods. However, Gal\nand Ghahramani's code used Bayesian optimization methods to tune\nhyper-parameters and this code contains a bug that optimizes hyper-parameters\nby maximizing performance on the test data. In particular for hyperparameter\nselection, they average performance across (subsets of) 5 of the training sets\nfrom the 20x train/test split, and then using the tau which got the best\naverage performance for all of 20x train/test splits to evaluate performance:\n\nhttps://github.com/yaringal/DropoutUncertaintyExps/blob/master/bostonHousing/net/experiment_BO.py#L54\n\nTherefore, the claim that \n\n\"Since we have established that MCBN performs on par with MCDO, by proxy we\nmight conclude that MCBN outperforms those VI methods as well.\"\n\nis not valid.\n\nAt the beginning of section 4.3 the authors indicate that they follow in their\nexperiments the setup of Gal and Ghahramani (2015). However, Gal and Ghahramani\n(2015) actually follow Hernández-Lobato and Adams, 2015 so the correct\nreference should be the latter one.\n\nClarity:\n\nThe paper is clearly written and easy to follow and understand.\n\nI found confusing how to use the proposed method to obtain estimates of\nuncertainty for a particular test data point x_star. The paragraph just above\nsection 4 says that the authors sample a batch of training data for this, but\nassume that the test point x_star has to be included in this batch.\nHow is this actually done in practice?\n\nOriginality:\n\nThe proposed contribution is original. This is the first time that a Bayesian\ninterpretation has been given to the batch normalization regularization\nproposal.\n\nSignificance:\n\nThe paper's contributions are significant. Batch normalization is a very\npopular regularization technique and showing that it can be used to obtain\nestimates of uncertainty is relevant and significant. Many existing deep\nlearning systems can use this to produce estimates of uncertainty in their\npredictions.\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511811246700}, {"id": "Hk7HI4h1G", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1165/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an approximate method to construct Bayesian uncertainty estimates in networks trained with batch normalization.\n\nThere is a lot going on in this paper. Although the overall presentation is clean, there are few key shortfalls (see below). Overall, the reported functionality is nice, although the experimental results are difficult to intepret (despite laudable effort by the authors to make them intuitive).\n\nSome open questions that I find crucial:\n\n* How exactly is the “stochastic forward-pass” performed that gives rise to the moment estimates? This step is the real meat of the paper, yet I struggle to find a concrete definition in the text. Is this really just an average over a few recent weights during optimization? If so, how is this method specific to batch normalization? Maybe I’m showing my own lack of understanding here, but it’s worrying that the actual sampling technique is not explained anywhere. This relates to a larger point about the paper's main point: What, exactly, is the Bayesian interpretation of batch normalization proposed here? In Bayesian Dropout, there is an explicit variational objective. Here, this is replaced by an implicit regularizer. The argument in Section 3.3 seems rather weak to me. To paraphrase it: If the prior vanishes, so does the regularizer. Fine. But what's the regularizer that's vanishing? The sentence that \"the influence of the prior diminishes as the size of the training data increases\" is debatable for something as over-parametrized as a DNN. I wouldn't be surprised that there are many directions in the weight-space of a trained DNN along which the posterior is dominated by the prior.\n\n* I’m confused about the statements made about the “constant uncertainty” baseline. First off, how is this (constant) width of the predictive region chosen? Did I miss this, or is it not explained anywhere? Unless I misunderstand the definition of CRPS and PLL, that width should matter, no? Then, the paragraph at the end of page 8 is worrying: The authors essentially say that the constant baseline is quite close to the estimate constructed in their work because constant uncertainty is “quite a reasonable baseline”. That can hardly be true (if it is, then it puts the entire paper into question! If trivial uncertainty is almost as good as this method, isn't the method trivial, too?). \nOn a related point: What would Figure 2 look like for the constand uncertainty setting? Just a horizontal line in blue and red? But at which level?\n\nI like this paper. It is presented well (modulo the above problems), and it makes some strong points. But I’m worried about the empirical evaluation, and the omission of crucial algorithmic details. They may hide serious problems.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A dense paper with a few key open questions", "rating": "5: Marginally below acceptance threshold", "review": "This paper proposes an approximate method to construct Bayesian uncertainty estimates in networks trained with batch normalization.\n\nThere is a lot going on in this paper. Although the overall presentation is clean, there are few key shortfalls (see below). Overall, the reported functionality is nice, although the experimental results are difficult to intepret (despite laudable effort by the authors to make them intuitive).\n\nSome open questions that I find crucial:\n\n* How exactly is the “stochastic forward-pass” performed that gives rise to the moment estimates? This step is the real meat of the paper, yet I struggle to find a concrete definition in the text. Is this really just an average over a few recent weights during optimization? If so, how is this method specific to batch normalization? Maybe I’m showing my own lack of understanding here, but it’s worrying that the actual sampling technique is not explained anywhere. This relates to a larger point about the paper's main point: What, exactly, is the Bayesian interpretation of batch normalization proposed here? In Bayesian Dropout, there is an explicit variational objective. Here, this is replaced by an implicit regularizer. The argument in Section 3.3 seems rather weak to me. To paraphrase it: If the prior vanishes, so does the regularizer. Fine. But what's the regularizer that's vanishing? The sentence that \"the influence of the prior diminishes as the size of the training data increases\" is debatable for something as over-parametrized as a DNN. I wouldn't be surprised that there are many directions in the weight-space of a trained DNN along which the posterior is dominated by the prior.\n\n* I’m confused about the statements made about the “constant uncertainty” baseline. First off, how is this (constant) width of the predictive region chosen? Did I miss this, or is it not explained anywhere? Unless I misunderstand the definition of CRPS and PLL, that width should matter, no? Then, the paragraph at the end of page 8 is worrying: The authors essentially say that the constant baseline is quite close to the estimate constructed in their work because constant uncertainty is “quite a reasonable baseline”. That can hardly be true (if it is, then it puts the entire paper into question! If trivial uncertainty is almost as good as this method, isn't the method trivial, too?). \nOn a related point: What would Figure 2 look like for the constand uncertainty setting? Just a horizontal line in blue and red? But at which level?\n\nI like this paper. It is presented well (modulo the above problems), and it makes some strong points. But I’m worried about the empirical evaluation, and the omission of crucial algorithmic details. They may hide serious problems.", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1510913595299}], "openreview_url": "https://openreview.net/forum?id=BJlrSmbAZ", "arxiv_id": "1802.06455", "paper_pdf": "papers/BJlrSmbAZ.pdf", "paper_pdf_sha256": "343389233ec47193f90410316e2fd1ab0aac5e05fcd71f80949826d0f79f808c", "paper_pdf_bytes": 1928150, "paper_pdf_source": "openreview", "code_url": "https://github.com/icml-mcbn/mcbn", "code_repository": "icml-mcbn/mcbn", "code_commit": "7c3338272eff0096c27dd139278037ea57c90cf7", "code_archive": "repos/BJlrSmbAZ.zip", "code_archive_sha256": "0f965632f9598d34d5f9ed3365ca129bd562a379035b20fe9806e916ff07b81b", "code_archive_bytes": 2319851, "code_file_count": 34, "code_extensions": {".py": 22, ".ipynb": 7, ".sh": 5}, "github_disk_usage_kb": 2251, "github_languages": {"Python": 106881, "Jupyter Notebook": 64125, "Shell": 4139, "Makefile": 387}, "github_archived": false, "github_pushed_at": "2018-05-04T22:44:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bayesian-uncertainty-estimation-for-batch"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bkU1bQUSQD", "year": 2026, "status": "rejected", "title": "VeriReason: Reinforcement Learning with Testbench Feedback for Reasoning-Enhanced Verilog Generation", "authors": ["Yiting Wang", "Guoheng Sun", "Wanghao Ye", "Gang Qu", "Ang Li"], "authorids": ["~Yiting_Wang8", "~Guoheng_Sun1", "~Wanghao_Ye2", "~Gang_Qu2", "~Ang_Li6"], "authors_source": "OpenReview API", "abstract": "Automating Register Transfer Level (RTL) code generation using Large Language Models (LLMs) offers substantial promise for streamlining digital circuit design and reducing human effort. However, current LLM-based approaches for RTL code generation face significant challenges. Methods such as supervised fine-tuning (SFT), in-context learning, and chain-of-thought (CoT) struggle with several critical limitations in the RTL domain: the scarcity of high-quality training data, poor alignment between natural language specifications and generated code, lack of built-in verification mechanisms, and difficulty balancing between model generalization and domain specialization. Inspired by groundbreaking research such as DeepSeek-R1, which combines reinforcement learning with reasoning capabilities, we introduce VeriReason, a comprehensive framework that integrates supervised fine-tuning with Guided Reward Proximal Optimization (GRPO) reinforcement learning specifically tailored for RTL code generation. Using our curated high-quality training examples alongside a feedback-driven reward model, VeriReason combines testbench evaluations with structural heuristics to improve specification-code alignment and eliminate hallucinations. Iterative GRPO embeds intrinsic self-checking and reasoning capabilities, enabling the model to autonomously detect and correct functional errors. On the VerilogEval Benchmark, VeriReason delivers significant improvements: achieving 83.1% functional correctness on the VerilogEval Machine benchmark, substantially outperforming both comparable-sized models and much larger commercial systems like GPT-4 Turbo. Additionally, our approach demonstrates up to a 2.8× increase in first-attempt functional correctness compared to baseline methods and exhibits robust generalization to unseen designs. To our knowledge, VeriReason represents the first system to successfully integrate explicit reasoning capabilities with reinforcement learning for Verilog generation, establishing a new state-of-the-art for automated RTL synthesis. The code is available at: https://anonymous.4open.science/r/VeriReason-E625.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "YfDnITWtYA", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13910/Reviewer_ZrAm"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces VeriReason, which combines SFT with GRPO to generate Verilog RTL code. The approach uses testbench feedback and AST-based structural rewards to train models that can reason through hardware design problems. The method is evaluated on VerilogEval benchmarks using various model sizes from 1.5B to 7B parameters.\n\nThough the paper represents an admirable contribution to this growing field, it presents several problems 1) the dataset and reproducibility details are incomplete, 2) the paper fails to clearly justify why GRPO is better than other RL approaches: for example, while the authors claim GRPO is more efficient than PPO, there's no empirical comparison between the two methods, 3) the experiments are missing crucial details about benchmark versions (which VerilogEval?) and baseline configs.\n\nThe authors could improve the chances of their paper acceptance by clarify the exact VerilogEval version and provide results on other popular benchmarks (CVDP, RTLLLM, etc) and add comparisons with other RL methods. Willing to reconsider the scores if these questions are addressed.", "review_text": "The paper introduces VeriReason, which combines SFT with GRPO to generate Verilog RTL code. The approach uses testbench feedback and AST-based structural rewards to train models that can reason through hardware design problems. The method is evaluated on VerilogEval benchmarks using various model sizes from 1.5B to 7B parameters.\n\nThough the paper represents an admirable contribution to this growing field, it presents several problems 1) the dataset and reproducibility details are incomplete, 2) the paper fails to clearly justify why GRPO is better than other RL approaches: for example, while the authors claim GRPO is more efficient than PPO, there's no empirical comparison between the two methods, 3) the experiments are missing crucial details about benchmark versions (which VerilogEval?) and baseline configs.\n\nThe authors could improve the chances of their paper acceptance by clarify the exact VerilogEval version and provide results on other popular benchmarks (CVDP, RTLLLM, etc) and add comparisons with other RL methods. Willing to reconsider the scores if these questions are addressed.", "strengths": "- The authors applies GRPO to a novel problem in data scarce hw frontend design with a well-designed multi-level reward system\n- The paper showcases impressive results: 83.1% pass@5 on VerilogEval-Machine, outperforming GPT-4 Turbo (83.0%) with much smaller models. The improvements are particularly impressive for smaller models (1.5B: +19.1 points). \n-- However, this is also a weakness, as VerilogEval-Human numbers lag significantly behind.\n- The adaptive data filtration strategy (retaining samples with mean reward ∈ [0.3, 1.8] and std > 0.1) is clever for identifying learnable examples.", "weaknesses": "- The paper mentions using VerilogEval but doesn't specify which version (v1 or v2); also failed to explain the differences in model performance for VerilogEval-Machine and VerilogEval-Human. These are important because impressive achievement in the former could be results of eval data contamination, since it was scraped from problems online \n- Where are the evaluation results for RTLLM and similar benchmarks?\n- No ablation on GRPO vs other RL algorithms (PPO, DPO)\n- Using GPT4 to regenerate and check code could introduce biases from GPT 4's training data, reinforcing the eval data contamination issue  \n- The training curves (Figure 2) show very different x-axis scales 800 vs 400 vs 100 steps, varied with model sizes, without explanation", "questions": "- Which VerilogEval version are you using? \n\n- Why the specific reward values (0.1, 1.1, 2.0)? The jump from max AST score (1.1) to functional correctness (2.0) seems arbitrary\n\n- Where are the RL baselines? Table 1 only compares against SFT methods. Where's the comparison with other RL approaches on the same task? Why is GRPO necessary? \n\n- The paper says that testbenches are generated automatically. How do you validate they actually test the functional correctness of the RTL code? What's the coverage of these generated tests?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces VeriReason, which combines SFT with GRPO to generate Verilog RTL code. The approach uses testbench feedback and AST-based structural rewards to train models that can reason through hardware design problems. The method is evaluated on VerilogEval benchmarks using various model sizes from 1.5B to 7B parameters.\n\nThough the paper represents an admirable contribution to this growing field, it presents several problems 1) the dataset and reproducibility details are incomplete, 2) the paper fails to clearly justify why GRPO is better than other RL approaches: for example, while the authors claim GRPO is more efficient than PPO, there's no empirical comparison between the two methods, 3) the experiments are missing crucial details about benchmark versions (which VerilogEval?) and baseline configs.\n\nThe authors could improve the chances of their paper acceptance by clarify the exact VerilogEval version and provide results on other popular benchmarks (CVDP, RTLLLM, etc) and add comparisons with other RL methods. Willing to reconsider the scores if these questions are addressed.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The authors applies GRPO to a novel problem in data scarce hw frontend design with a well-designed multi-level reward system\n- The paper showcases impressive results: 83.1% pass@5 on VerilogEval-Machine, outperforming GPT-4 Turbo (83.0%) with much smaller models. The improvements are particularly impressive for smaller models (1.5B: +19.1 points). \n-- However, this is also a weakness, as VerilogEval-Human numbers lag significantly behind.\n- The adaptive data filtration strategy (retaining samples with mean reward ∈ [0.3, 1.8] and std > 0.1) is clever for identifying learnable examples.", "weaknesses": "- The paper mentions using VerilogEval but doesn't specify which version (v1 or v2); also failed to explain the differences in model performance for VerilogEval-Machine and VerilogEval-Human. These are important because impressive achievement in the former could be results of eval data contamination, since it was scraped from problems online \n- Where are the evaluation results for RTLLM and similar benchmarks?\n- No ablation on GRPO vs other RL algorithms (PPO, DPO)\n- Using GPT4 to regenerate and check code could introduce biases from GPT 4's training data, reinforcing the eval data contamination issue  \n- The training curves (Figure 2) show very different x-axis scales 800 vs 400 vs 100 steps, varied with model sizes, without explanation", "questions": "- Which VerilogEval version are you using? \n\n- Why the specific reward values (0.1, 1.1, 2.0)? The jump from max AST score (1.1) to functional correctness (2.0) seems arbitrary\n\n- Where are the RL baselines? Table 1 only compares against SFT methods. Where's the comparison with other RL approaches on the same task? Why is GRPO necessary? \n\n- The paper says that testbenches are generated automatically. How do you validate they actually test the functional correctness of the RTL code? What's the coverage of these generated tests?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761964142588}, {"id": "nkDTWzdDxz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13910/Reviewer_gnu5"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new framework for Verilog Code Generation. The proposed VeriReason integrates Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) reinforcement learning to enhance the performance of Verilog Completion. The key contributions of this paper include    \n(1)\tCollect a high-quality Verilog dataset of 1,892 samples with a reasoning trajectory.     \n(2)\tDesign the reward method with testbench feedback using syntactic correctness, functional correctness, and structural similarity.", "review_text": "This paper introduces a new framework for Verilog Code Generation. The proposed VeriReason integrates Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) reinforcement learning to enhance the performance of Verilog Completion. The key contributions of this paper include    \n(1)\tCollect a high-quality Verilog dataset of 1,892 samples with a reasoning trajectory.     \n(2)\tDesign the reward method with testbench feedback using syntactic correctness, functional correctness, and structural similarity.", "strengths": "(1)\tA filtering algorithm for Verilog corpora. A major contribution of this paper is the two-stage adaptive filtration process to collect Verilog modules. These complex steps ensure the stability of GRPO training, especially the combination of the reward function.\n\n(2)\tA reward model with reinforcement learning testbench feedback. The reward score includes three measures from syntactic correctness, functional correctness, and structural similarity. Selected hyperparameters ensure the balance of these three components. During GRPO training, this scoring system prevents gradient vanishing and provides faster convergence.\n\n(3)\tReasoning generation and testbench generation. ChatGPT-4.1 is used to generate reasoning steps. The distillation trajectory could improve the reasoning ability of SFT model. Algorithm 3 is effective for automated testbench generation.", "weaknesses": "(1) Limited novelty and insights. Regarding the methodology, the authors are only performing SFT and GRPO on Qwen2.5, without proposing new training paradigms specifically for Verilog code generation or addressing some unique challenges. The results (in Fig. 2) only showed increasing reward scores, but an increasing reward may be achieved by the model's reward hacking, instead of true improvement in coding ability. Better show the real scores on benchmarks at different training steps. In addition, it is not sufficiently clear which specific policy contributed to the higher performance, compared with the SFT baseline. \n\n(2) Insufficient experiments for both benchmark and baselines. The authors only use the VerilogEval dataset to evaluate the pass@k. For Verilog Code Generation, the common benchmarks in most papers also include RTLLM [1] and CVDP [2]. More importantly, this paper has missed several recent works with 7B models that achieve higher scores on VerilogEval-Machine and VerilogEval-Human, such as CodeV [3] and CraftRTL [4]. Also, it seems to be a serious mistake that RTLLM [1], as a widely adopted benchmark, is listed as a model (as RTLLLM) with performance values in Table 1. In addition, regarding commericial LLMs, models such as GPT-5 can also be included. \n\n(3) Although the filtering and forming of the Verilog dataset is useful to the community, there is very limited quantitative analysis on these 1,892 samples. In addition, there have been many existing datasets (with reasoning) for Verilog code generation, and some may be of higher quality or greater quantity. \n\n(4) Possibly missing content. Seems the reviewer cannot find results corresponding to these claimed contributions: \"2.8× increase in first-attempt functional correctness\", “even with as few as 20 annotated examples, GRPO yields substantial performance gains”.\n\n(5) It may be worthwhile to change the base model from Qwen2.5 to other models. The solution may differ greatly when evaluated on different base models. \n\n[1] RTLLM: An open-source benchmark for design RTL generation with large language model.     \n[2] Comprehensive Verilog Design Problems: A Next-Generation Benchmark Dataset for Evaluating Large Language Models and Agents on RTL Design and Verification.      \n[3] CodeV: Empowering LLMs with HDL Generation through Multi-Level Summarization.   \n[4] CraftRTL: High-quality synthetic data generation for Verilog code models with correct-by-construction non-textual representations and targeted code repair.", "questions": "Please see the weakness part above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new framework for Verilog Code Generation. The proposed VeriReason integrates Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) reinforcement learning to enhance the performance of Verilog Completion. The key contributions of this paper include    \n(1)\tCollect a high-quality Verilog dataset of 1,892 samples with a reasoning trajectory.     \n(2)\tDesign the reward method with testbench feedback using syntactic correctness, functional correctness, and structural similarity.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "(1)\tA filtering algorithm for Verilog corpora. A major contribution of this paper is the two-stage adaptive filtration process to collect Verilog modules. These complex steps ensure the stability of GRPO training, especially the combination of the reward function.\n\n(2)\tA reward model with reinforcement learning testbench feedback. The reward score includes three measures from syntactic correctness, functional correctness, and structural similarity. Selected hyperparameters ensure the balance of these three components. During GRPO training, this scoring system prevents gradient vanishing and provides faster convergence.\n\n(3)\tReasoning generation and testbench generation. ChatGPT-4.1 is used to generate reasoning steps. The distillation trajectory could improve the reasoning ability of SFT model. Algorithm 3 is effective for automated testbench generation.", "weaknesses": "(1) Limited novelty and insights. Regarding the methodology, the authors are only performing SFT and GRPO on Qwen2.5, without proposing new training paradigms specifically for Verilog code generation or addressing some unique challenges. The results (in Fig. 2) only showed increasing reward scores, but an increasing reward may be achieved by the model's reward hacking, instead of true improvement in coding ability. Better show the real scores on benchmarks at different training steps. In addition, it is not sufficiently clear which specific policy contributed to the higher performance, compared with the SFT baseline. \n\n(2) Insufficient experiments for both benchmark and baselines. The authors only use the VerilogEval dataset to evaluate the pass@k. For Verilog Code Generation, the common benchmarks in most papers also include RTLLM [1] and CVDP [2]. More importantly, this paper has missed several recent works with 7B models that achieve higher scores on VerilogEval-Machine and VerilogEval-Human, such as CodeV [3] and CraftRTL [4]. Also, it seems to be a serious mistake that RTLLM [1], as a widely adopted benchmark, is listed as a model (as RTLLLM) with performance values in Table 1. In addition, regarding commericial LLMs, models such as GPT-5 can also be included. \n\n(3) Although the filtering and forming of the Verilog dataset is useful to the community, there is very limited quantitative analysis on these 1,892 samples. In addition, there have been many existing datasets (with reasoning) for Verilog code generation, and some may be of higher quality or greater quantity. \n\n(4) Possibly missing content. Seems the reviewer cannot find results corresponding to these claimed contributions: \"2.8× increase in first-attempt functional correctness\", “even with as few as 20 annotated examples, GRPO yields substantial performance gains”.\n\n(5) It may be worthwhile to change the base model from Qwen2.5 to other models. The solution may differ greatly when evaluated on different base models. \n\n[1] RTLLM: An open-source benchmark for design RTL generation with large language model.     \n[2] Comprehensive Verilog Design Problems: A Next-Generation Benchmark Dataset for Evaluating Large Language Models and Agents on RTL Design and Verification.      \n[3] CodeV: Empowering LLMs with HDL Generation through Multi-Level Summarization.   \n[4] CraftRTL: High-quality synthetic data generation for Verilog code models with correct-by-construction non-textual representations and targeted code repair.", "questions": "Please see the weakness part above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761919625369}, {"id": "jm91wthzjn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13910/Reviewer_4H7J"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces VeriReason, a framework for automated Verilog RTL code generation that integrates supervised fine-tuning with Group Relative Policy Optimization (GRPO)-based reinforcement learning. The framework uniquely combines structural, functional, and syntactic reward signals for optimizing large language models in hardware design domains. VeriReason is empirically shown to achieve state-of-the-art results on the VerilogEval benchmark, notably outperforming both open and commercial baselines across multiple model sizes. The approach emphasizes alignment with specification, self-checking behavior, and improved generalization with significantly less training data.", "review_text": "This paper introduces VeriReason, a framework for automated Verilog RTL code generation that integrates supervised fine-tuning with Group Relative Policy Optimization (GRPO)-based reinforcement learning. The framework uniquely combines structural, functional, and syntactic reward signals for optimizing large language models in hardware design domains. VeriReason is empirically shown to achieve state-of-the-art results on the VerilogEval benchmark, notably outperforming both open and commercial baselines across multiple model sizes. The approach emphasizes alignment with specification, self-checking behavior, and improved generalization with significantly less training data.", "strengths": "1. The integration of explicit testbench-driven feedback within a reinforcement learning loop (GRPO) for Verilog code generation is carefully engineered and directly tailored to the domain;\n2. The work includes comprehensive ablations that disentangle and validate the individual and combined effects of supervised fine-tuning and GRPO, making the contribution measurable and transparent;\n3. This paper open-sources its codebase and curated dataset to advance reproducibility and benchmarking practices in the research community.", "weaknesses": "1. In Table 1, baselines such as CodeV[1] and CraftRTL[2] for RTL generation are missed, which demonstrate stronger performance on the VerilogEval benchmark;\n2. There are other concurrent works on reinforcement learning-based RTL generation, such as [3] and [4]. It would be better if the authors can give a brief review of these works and clarify the difference between this work and others;\n3. The paper includes several manually selected hyperparameters. For instance, on lines 299–300, could the authors elaborate on the intuition or rationale behind the chosen values for $\\alpha_{\\text{min}}$, $\\alpha_{\\text{max}}$ and $\\beta$? Similarly, on line 272, how are the weights 0.6, 0.5, and -0.3 selected? Finally, how sensitive are the overall results to variations in these hyperparameters?\n\n[1] Zhao, Y., Huang, D., Li, C., Jin, P., Song, M., Xu, Y., Nan, Z., Gao, M., Ma, T., Qi, L. and Pan, Y., 2025. Codev: Empowering llms with hdl generation through multi-level summarization. *IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems*.\n\n[2] Liu, M., Tsai, Y.D., Zhou, W. and Ren, H., 2024. Craftrtl: High-quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair. *arXiv preprint arXiv:2409.12993*.\n\n[3] Teng, F., Pan, M., Zhang, X., He, Z., Yang, Y., Chai, X., Qi, M., Lu, L. and Yin, J., 2025. VERIRL: Boosting the LLM-based Verilog Code Generation via Reinforcement Learning. *arXiv preprint arXiv:2508.18462*.\n\n[4] Wang, N., Yao, B., Zhou, J., Hu, Y., Wang, X., Guan, N. and Jiang, Z., 2025. Insights from verification: Training a verilog generation LLM with reinforcement learning with testbench feedback. *arXiv preprint arXiv:2504.15804*.", "questions": "In Table 1, the reported performance for RTLCoder-DeepSeek-Coder on VerilogEval-Machine is 37.2 for pass@1 and 64.9 for pass@5, while on VerilogEval-Human it is 16.9 for pass@1 and 35.7 for pass@5. These results exhibit a substantial discrepancy compared to the original RTLCoder paper, which reports 61.2 for pass@1 and 76.5 for pass@5 on VerilogEval-Machine, and 41.6 for pass@1 and 50.1 for pass@5 on VerilogEval-Human. Could the authors elaborate on the reasons for this gap?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces VeriReason, a framework for automated Verilog RTL code generation that integrates supervised fine-tuning with Group Relative Policy Optimization (GRPO)-based reinforcement learning. The framework uniquely combines structural, functional, and syntactic reward signals for optimizing large language models in hardware design domains. VeriReason is empirically shown to achieve state-of-the-art results on the VerilogEval benchmark, notably outperforming both open and commercial baselines across multiple model sizes. The approach emphasizes alignment with specification, self-checking behavior, and improved generalization with significantly less training data.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The integration of explicit testbench-driven feedback within a reinforcement learning loop (GRPO) for Verilog code generation is carefully engineered and directly tailored to the domain;\n2. The work includes comprehensive ablations that disentangle and validate the individual and combined effects of supervised fine-tuning and GRPO, making the contribution measurable and transparent;\n3. This paper open-sources its codebase and curated dataset to advance reproducibility and benchmarking practices in the research community.", "weaknesses": "1. In Table 1, baselines such as CodeV[1] and CraftRTL[2] for RTL generation are missed, which demonstrate stronger performance on the VerilogEval benchmark;\n2. There are other concurrent works on reinforcement learning-based RTL generation, such as [3] and [4]. It would be better if the authors can give a brief review of these works and clarify the difference between this work and others;\n3. The paper includes several manually selected hyperparameters. For instance, on lines 299–300, could the authors elaborate on the intuition or rationale behind the chosen values for $\\alpha_{\\text{min}}$, $\\alpha_{\\text{max}}$ and $\\beta$? Similarly, on line 272, how are the weights 0.6, 0.5, and -0.3 selected? Finally, how sensitive are the overall results to variations in these hyperparameters?\n\n[1] Zhao, Y., Huang, D., Li, C., Jin, P., Song, M., Xu, Y., Nan, Z., Gao, M., Ma, T., Qi, L. and Pan, Y., 2025. Codev: Empowering llms with hdl generation through multi-level summarization. *IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems*.\n\n[2] Liu, M., Tsai, Y.D., Zhou, W. and Ren, H., 2024. Craftrtl: High-quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair. *arXiv preprint arXiv:2409.12993*.\n\n[3] Teng, F., Pan, M., Zhang, X., He, Z., Yang, Y., Chai, X., Qi, M., Lu, L. and Yin, J., 2025. VERIRL: Boosting the LLM-based Verilog Code Generation via Reinforcement Learning. *arXiv preprint arXiv:2508.18462*.\n\n[4] Wang, N., Yao, B., Zhou, J., Hu, Y., Wang, X., Guan, N. and Jiang, Z., 2025. Insights from verification: Training a verilog generation LLM with reinforcement learning with testbench feedback. *arXiv preprint arXiv:2504.15804*.", "questions": "In Table 1, the reported performance for RTLCoder-DeepSeek-Coder on VerilogEval-Machine is 37.2 for pass@1 and 64.9 for pass@5, while on VerilogEval-Human it is 16.9 for pass@1 and 35.7 for pass@5. These results exhibit a substantial discrepancy compared to the original RTLCoder paper, which reports 61.2 for pass@1 and 76.5 for pass@5 on VerilogEval-Machine, and 41.6 for pass@1 and 50.1 for pass@5 on VerilogEval-Human. Could the authors elaborate on the reasons for this gap?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761804433967}], "openreview_url": "https://openreview.net/forum?id=bkU1bQUSQD", "arxiv_id": "2505.11849", "paper_pdf": "papers/bkU1bQUSQD.pdf", "paper_pdf_sha256": "48334b758dea98d967ffedc9eb2034e444c37d1bc3b1eb1c1f3845dcc353645a", "paper_pdf_bytes": 397020, "paper_pdf_source": "openreview", "code_url": "https://github.com/NellyW8/VeriReason", "code_repository": "NellyW8/VeriReason", "code_commit": "d215b7fe1b3db6dd4ca725f7d9399c49414c7531", "code_archive": "repos/bkU1bQUSQD.zip", "code_archive_sha256": "a8f766e82a9735827429b94b64d7783379615a8dba7a962322f6b0d25da5c49a", "code_archive_bytes": 199100, "code_file_count": 5, "code_extensions": {".py": 4, ".sh": 1}, "github_disk_usage_kb": 284, "github_languages": {"Python": 103440, "HTML": 71729, "Shell": 1229}, "github_archived": false, "github_pushed_at": "2025-09-25T11:45:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/verireason-reinforcement-learning-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UW0zetsx8X", "year": 2025, "status": "rejected", "title": "Prompt Optimization with Human Feedback", "authors": ["Xiaoqiang Lin", "Zhongxiang Dai", "Arun Verma", "See-Kiong Ng", "Patrick Jaillet", "Bryan Kian Hsiang Low"], "authorids": ["~Xiaoqiang_Lin1", "~Zhongxiang_Dai1", "~Arun_Verma1", "~See-Kiong_Ng1", "~Patrick_Jaillet1", "~Bryan_Kian_Hsiang_Low1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have demonstrated remarkable performances in various tasks. However, the performances of LLMs heavily depend on the input prompt. This has given rise to a number of recent works on prompt optimization. However, the previous works often require the availability of a numeric score to assess the quality of every prompt. Unfortunately, when a human user interacts with a black-box LLM, it is often infeasible and unreliable to attain such a score. Instead, it is usually significantly easier and more reliable to obtain preference feedback from a human user, i.e., showing the user the responses generated from a pair of prompts and asking the user which one is preferred. Therefore, in this paper, we study the problem of prompt optimization with human feedback (POHF), in which we aim to optimize the prompt for a black-box LLM using only human preference feedback. By drawing inspirations from dueling bandits, we design a theoretically principled strategy to select a pair of prompts to query for preference feedback in every iteration, and hence introduce our algorithm named automated POHF (APOHF). We apply our APOHF algorithm to a variety of tasks, including optimizing user instructions, prompt optimization for text-to-image generative models, and response optimization with human feedback (i.e., further refining the response using a variant of our APOHF). The results demonstrate that our APOHF can efficiently find a good prompt using a small number of preference feedback instances.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "pkZopzKgTP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13379/Reviewer_jWbb"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces APOHF, for optimizing prompts for large language models (LLMs) using only human preference feedback rather than numeric scores. APOHF iteratively selects prompt pairs for user comparison, training a neural network to predict each prompt’s utility based on feedback. The algorithm selects prompts by combining utility prediction with an exploration-exploitation approach inspired by dueling bandits. Applied across tasks like instruction optimization and prompt tuning for text-to-image models, APOHF demonstrates more efficient prompt selection compared to baseline methods under limited feedback conditions.", "review_text": "This paper introduces APOHF, for optimizing prompts for large language models (LLMs) using only human preference feedback rather than numeric scores. APOHF iteratively selects prompt pairs for user comparison, training a neural network to predict each prompt’s utility based on feedback. The algorithm selects prompts by combining utility prediction with an exploration-exploitation approach inspired by dueling bandits. Applied across tasks like instruction optimization and prompt tuning for text-to-image models, APOHF demonstrates more efficient prompt selection compared to baseline methods under limited feedback conditions.", "strengths": "* This paper presents a novel approach to prompt optimization by using human preference feedback alone, which is suitable for black-box LLMs.\n* This paper is generally clear in its method description, though certain theoretical justifications could be expanded for a more rigorous understanding of APOHF’s design choices.\n* The algorithm design balances prompt utility prediction with exploration, showing reliable performance across varied tasks.", "weaknesses": "* The method employs the Bradley-Terry-Luce (BTL) model to represent human preference feedback, which assumes consistency and transitivity in user preferences. This may oversimplify human feedback, especially in real-world applications where user preferences can be inconsistent or influenced by context. The reliance on binary feedback might also limit the granularity of information available to the model, potentially leading to suboptimal prompt choices. \n* The method relies on a user-provided initial task description to generate the prompt domain, assuming that these initial examples are representative of the task requirements. This dependency introduces the potential for bias if the initial examples do not capture the full scope of the task or if the user’s interpretation is inconsistent with the intended outcomes. This reliance can constrain the model's flexibility and lead to prompts that are effective only in limited or narrowly defined scenarios.\n* APOHF presumes that user preferences are consistent and relevant across multiple iterations, assuming stability in what constitutes an optimal prompt. However, in complex, open-ended tasks or tasks that evolve over time, user preferences may shift, and certain prompts that were optimal initially may no longer be relevant. This assumption limits the model's adaptability to dynamic contexts, reducing its applicability in real-world tasks where user expectations or task goals may evolve.", "questions": "* Have the authors examined how noise in user feedback impacts the accuracy of the prompt selection? Is the model robust to feedback inconsistencies or context-dependency in user preferences?\n* Is there a way to iteratively refine or expand the prompt domain based on user feedback, to counteract biases introduced by an incomplete initial task description?\n* Has the algorithm been tested in settings where user preferences or task requirements change over time? If so, how does APOHF handle shifts in user preferences?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces APOHF, for optimizing prompts for large language models (LLMs) using only human preference feedback rather than numeric scores. APOHF iteratively selects prompt pairs for user comparison, training a neural network to predict each prompt’s utility based on feedback. The algorithm selects prompts by combining utility prediction with an exploration-exploitation approach inspired by dueling bandits. Applied across tasks like instruction optimization and prompt tuning for text-to-image models, APOHF demonstrates more efficient prompt selection compared to baseline methods under limited feedback conditions.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* This paper presents a novel approach to prompt optimization by using human preference feedback alone, which is suitable for black-box LLMs.\n* This paper is generally clear in its method description, though certain theoretical justifications could be expanded for a more rigorous understanding of APOHF’s design choices.\n* The algorithm design balances prompt utility prediction with exploration, showing reliable performance across varied tasks.", "weaknesses": "* The method employs the Bradley-Terry-Luce (BTL) model to represent human preference feedback, which assumes consistency and transitivity in user preferences. This may oversimplify human feedback, especially in real-world applications where user preferences can be inconsistent or influenced by context. The reliance on binary feedback might also limit the granularity of information available to the model, potentially leading to suboptimal prompt choices. \n* The method relies on a user-provided initial task description to generate the prompt domain, assuming that these initial examples are representative of the task requirements. This dependency introduces the potential for bias if the initial examples do not capture the full scope of the task or if the user’s interpretation is inconsistent with the intended outcomes. This reliance can constrain the model's flexibility and lead to prompts that are effective only in limited or narrowly defined scenarios.\n* APOHF presumes that user preferences are consistent and relevant across multiple iterations, assuming stability in what constitutes an optimal prompt. However, in complex, open-ended tasks or tasks that evolve over time, user preferences may shift, and certain prompts that were optimal initially may no longer be relevant. This assumption limits the model's adaptability to dynamic contexts, reducing its applicability in real-world tasks where user expectations or task goals may evolve.", "questions": "* Have the authors examined how noise in user feedback impacts the accuracy of the prompt selection? Is the model robust to feedback inconsistencies or context-dependency in user preferences?\n* Is there a way to iteratively refine or expand the prompt domain based on user feedback, to counteract biases introduced by an incomplete initial task description?\n* Has the algorithm been tested in settings where user preferences or task requirements change over time? If so, how does APOHF handle shifts in user preferences?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730728841726}, {"id": "Lg6hXSy2no", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13379/Reviewer_WviX"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper mentions a common issue of black-box LLM usages, where it is unavailable to automatically obtain a quality score for a given prompt, thus causing the difficulty of prompt optimization. Assuming that only human’s preference feedback is reliable, the authors propose APOHF, a framework to perform prompt optimization with human feedback. The proposed method aims to determine a good prompt via a prompt selection algorithm inspired by Dueling Bandits, based on a neural network performance predictor trained with the information from a small number of human feedback instances. The authors demonstrate quality improvement throughout iterations of the optimization. Overall, the experiment results show that the proposed framework is effective to obtain appropriate prompts with human feedback.", "review_text": "This paper mentions a common issue of black-box LLM usages, where it is unavailable to automatically obtain a quality score for a given prompt, thus causing the difficulty of prompt optimization. Assuming that only human’s preference feedback is reliable, the authors propose APOHF, a framework to perform prompt optimization with human feedback. The proposed method aims to determine a good prompt via a prompt selection algorithm inspired by Dueling Bandits, based on a neural network performance predictor trained with the information from a small number of human feedback instances. The authors demonstrate quality improvement throughout iterations of the optimization. Overall, the experiment results show that the proposed framework is effective to obtain appropriate prompts with human feedback.", "strengths": "-  Inspired by the concept of RLHF and Dueling Bandits, the authors propose a prompt optimization framework leveraging human feedback. The experiment results show that the proposed method is indeed effective to a certain extent with actual demonstrations.\n\n-  With the Bandit-fashioned prompt selection strategy, powered by a trained NN (MLP) model as the performance predictor, the proposed framework determines the prompt efficiently in terms of the number of required human feedback instances.\n\n-  In the appendix, the authors also provide a brief coverage of theoretical explanations to justify the proposed prompt selection strategy.", "weaknesses": "-   Limited Comparison with Prompt Optimization Baselines:\nThe authors' decision to exclude comparisons with state-of-the-art prompt optimization methods (e.g. TextGrad), citing the lack of a scoring method, may be overly restrictive. While direct human scoring might not be feasible, surrogate evaluation metrics (e.g., embedding-similarity between LLM output and ground truth) could serve as viable alternatives for these comparisons. Given the superior performance of existing prompt optimization methods across various tasks, their inclusion in the performance comparisons would provide valuable context and a more comprehensive evaluation of APOHF's effectiveness relative to the current state of the art.\n\n- Dependency on Neural Network Models for Prompt Selection:\nThe proposed framework's prompt-pair selection strategy heavily relies on a trained neural network (MLP) model for performance prediction based on input embeddings. This approach introduces two potential sources of variability:\na) The choice of embedding model\nb) The architecture and training of the performance predictor (MLP)\nBoth components may significantly influence the accuracy of performance predictions, potentially leading to substantial variations in the final LLM performance. A more robust analysis of these components' impact on the overall framework would strengthen the paper's credibility.\n\n- Ambiguity in Prompt Domain Generation:\nThe process of generating the \"discrete domain of prompts X\" lacks sufficient detail. While the authors mention using a powerful LLM (e.g., ChatGPT) for prompt generation via in-context learning, the specifics of this crucial step remain unclear. Moreover, the paper lacks experimental analysis demonstrating how different prompt domain generation methods affect APOHF's overall performance. This omission limits the reader's ability to fully assess the method's robustness and generalizability.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper mentions a common issue of black-box LLM usages, where it is unavailable to automatically obtain a quality score for a given prompt, thus causing the difficulty of prompt optimization. Assuming that only human’s preference feedback is reliable, the authors propose APOHF, a framework to perform prompt optimization with human feedback. The proposed method aims to determine a good prompt via a prompt selection algorithm inspired by Dueling Bandits, based on a neural network performance predictor trained with the information from a small number of human feedback instances. The authors demonstrate quality improvement throughout iterations of the optimization. Overall, the experiment results show that the proposed framework is effective to obtain appropriate prompts with human feedback.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "-  Inspired by the concept of RLHF and Dueling Bandits, the authors propose a prompt optimization framework leveraging human feedback. The experiment results show that the proposed method is indeed effective to a certain extent with actual demonstrations.\n\n-  With the Bandit-fashioned prompt selection strategy, powered by a trained NN (MLP) model as the performance predictor, the proposed framework determines the prompt efficiently in terms of the number of required human feedback instances.\n\n-  In the appendix, the authors also provide a brief coverage of theoretical explanations to justify the proposed prompt selection strategy.", "weaknesses": "-   Limited Comparison with Prompt Optimization Baselines:\nThe authors' decision to exclude comparisons with state-of-the-art prompt optimization methods (e.g. TextGrad), citing the lack of a scoring method, may be overly restrictive. While direct human scoring might not be feasible, surrogate evaluation metrics (e.g., embedding-similarity between LLM output and ground truth) could serve as viable alternatives for these comparisons. Given the superior performance of existing prompt optimization methods across various tasks, their inclusion in the performance comparisons would provide valuable context and a more comprehensive evaluation of APOHF's effectiveness relative to the current state of the art.\n\n- Dependency on Neural Network Models for Prompt Selection:\nThe proposed framework's prompt-pair selection strategy heavily relies on a trained neural network (MLP) model for performance prediction based on input embeddings. This approach introduces two potential sources of variability:\na) The choice of embedding model\nb) The architecture and training of the performance predictor (MLP)\nBoth components may significantly influence the accuracy of performance predictions, potentially leading to substantial variations in the final LLM performance. A more robust analysis of these components' impact on the overall framework would strengthen the paper's credibility.\n\n- Ambiguity in Prompt Domain Generation:\nThe process of generating the \"discrete domain of prompts X\" lacks sufficient detail. While the authors mention using a powerful LLM (e.g., ChatGPT) for prompt generation via in-context learning, the specifics of this crucial step remain unclear. Moreover, the paper lacks experimental analysis demonstrating how different prompt domain generation methods affect APOHF's overall performance. This omission limits the reader's ability to fully assess the method's robustness and generalizability.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730614483915}, {"id": "n6xi8YuAFz", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13379/Reviewer_Y27t"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper mainly studies the prompt optimization problem. In specific, the main motivation coms from that previous PO methods usually require the availability of a numeric score to assess the quality of every prompt. This score is difficult or sometime unable to get in the situation when human interact with a black-box LLM system. Therefore, the authors claimed that preference feedback is more feasible given the mentioned human-llm interaction scenario. The authors thus proposed a method named APOHF for prompt optimization with human feedback. Some results show that this method can find a good prompt with several preference feedbacks used.", "review_text": "This paper mainly studies the prompt optimization problem. In specific, the main motivation coms from that previous PO methods usually require the availability of a numeric score to assess the quality of every prompt. This score is difficult or sometime unable to get in the situation when human interact with a black-box LLM system. Therefore, the authors claimed that preference feedback is more feasible given the mentioned human-llm interaction scenario. The authors thus proposed a method named APOHF for prompt optimization with human feedback. Some results show that this method can find a good prompt with several preference feedbacks used.", "strengths": "The paper is well written and easy to follow\nThe motivation is clear compared to previous method. [which may not be practical as per the weakness part]", "weaknesses": "The main weakness is the prompt optimization from human feedback itself may not be practical. Let me explain the reasons. \n1) The advantage or the main efforts of LLM is to improve its instruction following abilities, I.e., to cover as many prompts as possible. Therefore, when human in the loop, the key efforts should be in the zero-shot/few-shot abilities. If you expect the user to give feedback to the prompt provided by him/her-self, it may negatively impact the feasible application of this LLM product. How does your approach complement efforts to improve zero-shot/few-shot abilities of LLMs? Are there specific scenarios where human feedback on prompts provides unique value, even as LLMs improve their general instruction-following capabilities?\n\n\n2) Prompt optimization itself is important for sure, but the efforts on making it automatic are more useful. For instance, as many APO did, LLM can improve the prompt itself by self-reflecting (e.g., reasoning the errors and self-refining current prompts). How does your method compare to or complement automated prompt optimization techniques like self-reflection? Are there potential advantages to incorporating human feedback alongside automated methods?\n\n3) If let’s say we need human preference, we can directly train a reward model. Given the generalization ability of the reward model, we can directly sample the best response with requiring the human-preference during inference. How does your approach compare to training a reward model on human preferences? Could you discuss the potential advantages or disadvantages of your method compared to using a pre-trained reward model?\n\nIn addition, the experiments can be done on more types of tasks, such as math, coding, role-play, etc.", "questions": "Most of questions have been listed in the weakness part.\n\nAnother general question or suggestion would be: Could the author please raise up several real examples that can show the scenarios where the proposed method will be feasible to apply?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper mainly studies the prompt optimization problem. In specific, the main motivation coms from that previous PO methods usually require the availability of a numeric score to assess the quality of every prompt. This score is difficult or sometime unable to get in the situation when human interact with a black-box LLM system. Therefore, the authors claimed that preference feedback is more feasible given the mentioned human-llm interaction scenario. The authors thus proposed a method named APOHF for prompt optimization with human feedback. Some results show that this method can find a good prompt with several preference feedbacks used.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "The paper is well written and easy to follow\nThe motivation is clear compared to previous method. [which may not be practical as per the weakness part]", "weaknesses": "The main weakness is the prompt optimization from human feedback itself may not be practical. Let me explain the reasons. \n1) The advantage or the main efforts of LLM is to improve its instruction following abilities, I.e., to cover as many prompts as possible. Therefore, when human in the loop, the key efforts should be in the zero-shot/few-shot abilities. If you expect the user to give feedback to the prompt provided by him/her-self, it may negatively impact the feasible application of this LLM product. How does your approach complement efforts to improve zero-shot/few-shot abilities of LLMs? Are there specific scenarios where human feedback on prompts provides unique value, even as LLMs improve their general instruction-following capabilities?\n\n\n2) Prompt optimization itself is important for sure, but the efforts on making it automatic are more useful. For instance, as many APO did, LLM can improve the prompt itself by self-reflecting (e.g., reasoning the errors and self-refining current prompts). How does your method compare to or complement automated prompt optimization techniques like self-reflection? Are there potential advantages to incorporating human feedback alongside automated methods?\n\n3) If let’s say we need human preference, we can directly train a reward model. Given the generalization ability of the reward model, we can directly sample the best response with requiring the human-preference during inference. How does your approach compare to training a reward model on human preferences? Could you discuss the potential advantages or disadvantages of your method compared to using a pre-trained reward model?\n\nIn addition, the experiments can be done on more types of tasks, such as math, coding, role-play, etc.", "questions": "Most of questions have been listed in the weakness part.\n\nAnother general question or suggestion would be: Could the author please raise up several real examples that can show the scenarios where the proposed method will be feasible to apply?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730506668641}, {"id": "6MO1LOWcEr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13379/Reviewer_bMVs"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The authors of this paper studied prompt optimization using dueling bandit. The basic idea is to ask users to provide preference feedback on responses generated by two prompts, which is directly attributed to the quality of the prompts for further selection. Experiments on prompt optimization in text-to-text and text-to-image generation tasks, and also in response optimization, demonstrate the effectiveness of the proposed solution.", "review_text": "The authors of this paper studied prompt optimization using dueling bandit. The basic idea is to ask users to provide preference feedback on responses generated by two prompts, which is directly attributed to the quality of the prompts for further selection. Experiments on prompt optimization in text-to-text and text-to-image generation tasks, and also in response optimization, demonstrate the effectiveness of the proposed solution.", "strengths": "1.\tPrompt optimization is an important problem in LLM studies, and the paper provides a valid perspective.\n2.\tThe clarity of the manuscript is satisfactory, which helps readers to best comprehend the technique details.", "weaknesses": "1.\tThe proposed solution is rather standard: running dueling bandits on top of ChatGPT generated rewrites of initial queries. The synergy between these two components is very weak. Ideally, the proposal of new prompts should be informed by the learnt scoring function. But the current way, the new prompts are sampled from ChatGPT independently from the scoring function.\n2.\tAs the users’ feedback is about the response, rather than the prompt itself, the LLM that generates the responses matters. Specifically, the prompt quality is a function of generator LLM as well. This clearly limits the generality of the proposed solution: the scoring function learnt with one LLM might not work well for other LLMs. Unfortunately, there is no experiment investigating this factor. \n3.\tA few statements made in the paper are somehow overclaiming, for example, using a text encoder can avoid the need of a whitebox LLM, which does not seem to be advantageous, since the availability of whitebox LLM is no worse than an opensource text encoder. And I do not see what the principle is behind choosing the highest scored prompt in history as the first prompt, except it works better than randomly choosing two.", "questions": "1.\tAs the user feedback is provided to the response, instead of the prompt directly, how to we account for the variance caused by sampling from the LLM? For example, the user feels one result being better than another could be caused by LLM sampling, rather than the prompt. And I am not sure if this variance can be simply factored into the BT model.\n2.\tDoes the learnt scoring function work across different generator LLMs? Similarly, does the learnt scoring function generalize across different tasks, e.g., question answering vs., text summarization?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors of this paper studied prompt optimization using dueling bandit. The basic idea is to ask users to provide preference feedback on responses generated by two prompts, which is directly attributed to the quality of the prompts for further selection. Experiments on prompt optimization in text-to-text and text-to-image generation tasks, and also in response optimization, demonstrate the effectiveness of the proposed solution.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tPrompt optimization is an important problem in LLM studies, and the paper provides a valid perspective.\n2.\tThe clarity of the manuscript is satisfactory, which helps readers to best comprehend the technique details.", "weaknesses": "1.\tThe proposed solution is rather standard: running dueling bandits on top of ChatGPT generated rewrites of initial queries. The synergy between these two components is very weak. Ideally, the proposal of new prompts should be informed by the learnt scoring function. But the current way, the new prompts are sampled from ChatGPT independently from the scoring function.\n2.\tAs the users’ feedback is about the response, rather than the prompt itself, the LLM that generates the responses matters. Specifically, the prompt quality is a function of generator LLM as well. This clearly limits the generality of the proposed solution: the scoring function learnt with one LLM might not work well for other LLMs. Unfortunately, there is no experiment investigating this factor. \n3.\tA few statements made in the paper are somehow overclaiming, for example, using a text encoder can avoid the need of a whitebox LLM, which does not seem to be advantageous, since the availability of whitebox LLM is no worse than an opensource text encoder. And I do not see what the principle is behind choosing the highest scored prompt in history as the first prompt, except it works better than randomly choosing two.", "questions": "1.\tAs the user feedback is provided to the response, instead of the prompt directly, how to we account for the variance caused by sampling from the LLM? For example, the user feels one result being better than another could be caused by LLM sampling, rather than the prompt. And I am not sure if this variance can be simply factored into the BT model.\n2.\tDoes the learnt scoring function work across different generator LLMs? Similarly, does the learnt scoring function generalize across different tasks, e.g., question answering vs., text summarization?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730389060922}, {"id": "Im3e56yLwu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13379/Reviewer_X4tm"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "In existing prompt optimization work, the vast majority of methods rely on numerical scores to select better-performing prompts. However, in certain real-world tasks (e.g., text-to-image generation), using numerical scores for evaluation may not be applicable. To address this issue, the authors propose a human feedback-based prompt optimization method—APOHF. This method first collects pairs of human preference data and trains a neural network to provide latent scores aligned with human preferences. Then, combining greedy search and the method of maximizing the upper confidence bound, the authors select two prompts from multiple directions, balancing both performance and diverse prompt exploration. Experimental results show that after several iterations, APOHF can generate high-quality prompts.", "review_text": "In existing prompt optimization work, the vast majority of methods rely on numerical scores to select better-performing prompts. However, in certain real-world tasks (e.g., text-to-image generation), using numerical scores for evaluation may not be applicable. To address this issue, the authors propose a human feedback-based prompt optimization method—APOHF. This method first collects pairs of human preference data and trains a neural network to provide latent scores aligned with human preferences. Then, combining greedy search and the method of maximizing the upper confidence bound, the authors select two prompts from multiple directions, balancing both performance and diverse prompt exploration. Experimental results show that after several iterations, APOHF can generate high-quality prompts.", "strengths": "1. The APOHF method proposed by the authors aims to solve the problem of prompt optimization in tasks that are difficult to evaluate using numerical scores, marking a new attempt different from previous studies.\n2. APOHF outperforms other baseline methods in terms of performance.\n3. The wide range of experimental types further validates the effectiveness of APOHF.\n4. The authors' writing is clear, and the content of the paper is concise and easy to understand.", "weaknesses": "### 1. Experiment\n\na) From Table 6 in the appendix, I understand that for the user instruction optimization task, the authors chose two tasks, \"rhymes\" and \"word sorting\" (presumably from the BigBench dataset), which can be evaluated using numerical scores (such as accuracy), making them compatible with methods like APE, OPRO, and APO. I believe the authors should discuss this in more detail. Considering the rebuttal time constraints, You may focus on an in-depth analysis of 1 or 2 methods (such as APO, OPRO). Additionally, regarding the \"sentiment\" task in Table 6, why did the authors choose an initial instruction with a score of 0? For a simple instruction like \"Determine the sentiment category (positive or negative) of a given sentence,\" a score of 0 should not occur.\n\nb) I recommend that the authors revise the way the curves are represented in the figures. The current use of square, circular, and triangular symbols makes the figures cluttered, making it difficult to discern curve details.\n\n### 2. Motivation\na) For prompt tasks that are difficult to evaluate with numerical scores (e.g., text-to-image generation), the optimization target of the APOHF method is individual samples (such as generating images of a garden or a street). I believe the practical significance of such single-sample optimization is limited. In everyday usage, this approach has high resource costs, making it inefficient. In industry, optimizing meta prompts holds greater practical value.\n\nb) As I know, APOHF is not the first prompt optimization work considering human feedback. Harvard and MIT have published a similar paper named **PRompt Optimization in Multi-Step Tasks (PROMST): Integrating Human Feedback and Heuristic-based Sampling (https://arxiv.org/pdf/2402.08702)**. I think it maybe diminish the contribution of this work.", "questions": "1. Regarding the resource consumption of APOHF, what is the specific cost? How many API calls were made in total, and how many tokens were consumed during the experiments?\n\n2. In Figure 3, does it show the score of the newly generated prompt after each iteration? Why doesn't APOHF exhibit score fluctuations? Logically, the scores of results generated by different prompts should vary, making it unlikely for them to remain consistent.\n\n3. In the text-to-image task, the authors measure the quality of generated images by calculating the similarity between the generated images and the ground truth. Could this similarity metric also be used as a scoring standard for methods like APO and OPRO?\n\n4. How many data samples does the test set include?\n\n5. Please answer my question refer to **Weakness  Section**", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In existing prompt optimization work, the vast majority of methods rely on numerical scores to select better-performing prompts. However, in certain real-world tasks (e.g., text-to-image generation), using numerical scores for evaluation may not be applicable. To address this issue, the authors propose a human feedback-based prompt optimization method—APOHF. This method first collects pairs of human preference data and trains a neural network to provide latent scores aligned with human preferences. Then, combining greedy search and the method of maximizing the upper confidence bound, the authors select two prompts from multiple directions, balancing both performance and diverse prompt exploration. Experimental results show that after several iterations, APOHF can generate high-quality prompts.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The APOHF method proposed by the authors aims to solve the problem of prompt optimization in tasks that are difficult to evaluate using numerical scores, marking a new attempt different from previous studies.\n2. APOHF outperforms other baseline methods in terms of performance.\n3. The wide range of experimental types further validates the effectiveness of APOHF.\n4. The authors' writing is clear, and the content of the paper is concise and easy to understand.", "weaknesses": "### 1. Experiment\n\na) From Table 6 in the appendix, I understand that for the user instruction optimization task, the authors chose two tasks, \"rhymes\" and \"word sorting\" (presumably from the BigBench dataset), which can be evaluated using numerical scores (such as accuracy), making them compatible with methods like APE, OPRO, and APO. I believe the authors should discuss this in more detail. Considering the rebuttal time constraints, You may focus on an in-depth analysis of 1 or 2 methods (such as APO, OPRO). Additionally, regarding the \"sentiment\" task in Table 6, why did the authors choose an initial instruction with a score of 0? For a simple instruction like \"Determine the sentiment category (positive or negative) of a given sentence,\" a score of 0 should not occur.\n\nb) I recommend that the authors revise the way the curves are represented in the figures. The current use of square, circular, and triangular symbols makes the figures cluttered, making it difficult to discern curve details.\n\n### 2. Motivation\na) For prompt tasks that are difficult to evaluate with numerical scores (e.g., text-to-image generation), the optimization target of the APOHF method is individual samples (such as generating images of a garden or a street). I believe the practical significance of such single-sample optimization is limited. In everyday usage, this approach has high resource costs, making it inefficient. In industry, optimizing meta prompts holds greater practical value.\n\nb) As I know, APOHF is not the first prompt optimization work considering human feedback. Harvard and MIT have published a similar paper named **PRompt Optimization in Multi-Step Tasks (PROMST): Integrating Human Feedback and Heuristic-based Sampling (https://arxiv.org/pdf/2402.08702)**. I think it maybe diminish the contribution of this work.", "questions": "1. Regarding the resource consumption of APOHF, what is the specific cost? How many API calls were made in total, and how many tokens were consumed during the experiments?\n\n2. In Figure 3, does it show the score of the newly generated prompt after each iteration? Why doesn't APOHF exhibit score fluctuations? Logically, the scores of results generated by different prompts should vary, making it unlikely for them to remain consistent.\n\n3. In the text-to-image task, the authors measure the quality of generated images by calculating the similarity between the generated images and the ground truth. Could this similarity metric also be used as a scoring standard for methods like APO and OPRO?\n\n4. How many data samples does the test set include?\n\n5. Please answer my question refer to **Weakness  Section**", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729518835098}], "openreview_url": "https://openreview.net/forum?id=UW0zetsx8X", "arxiv_id": "2405.17346", "paper_pdf": "papers/UW0zetsx8X.pdf", "paper_pdf_sha256": "3b7d626c1285a47cc6c52cbfdfa107259ac1c4f614f8c0f42431a0083800638a", "paper_pdf_bytes": 1520400, "paper_pdf_source": "openreview", "code_url": "https://github.com/xqlin98/APOHF", "code_repository": "xqlin98/APOHF", "code_commit": "26eea14b1f89efc1b6d61bed8d154f44fa236b79", "code_archive": "repos/UW0zetsx8X.zip", "code_archive_sha256": "795ace631dad3634d807850a2374f1c9fc7b8ff945b7a89ba07440b29d7565f1", "code_archive_bytes": 50363, "code_file_count": 17, "code_extensions": {".py": 14, ".sh": 3}, "github_disk_usage_kb": 44, "github_languages": {"Python": 164262, "Shell": 1311}, "github_archived": false, "github_pushed_at": "2024-08-07T16:59:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/prompt-optimization-with-human-feedback"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5BoXZXTJvL", "year": 2024, "status": "rejected", "title": "Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models", "authors": ["Rocktim Jyoti Das", "Liqun Ma", "Zhiqiang Shen"], "authorids": ["~Rocktim_Jyoti_Das2", "~Liqun_Ma1", "~Zhiqiang_Shen1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) with a billion or more parameters are prime targets for network pruning, which aims to reduce a portion of the network weights without compromising performance. Prior approaches such as Weights Magnitude, SparseGPT, and Wanda, either concentrated solely on weights or integrated weights with activations for sparsity. However, they overlooked the informative gradients derived from pretrained large language models. In this paper, we present a novel sparsity-centric pruning method for pretrained LLMs, termed **G**radient-**b**ased **L**anguage **M**odel **P**runer (**GBLM-Pruner**). Distinctively, GBLM-Pruner operates in a training-free manner by harnessing normalized gradients, and substantially outperforms competitive counterparts like SparseGPT and Wanda in multiple benchmarks. Intriguing, after incorporating gradients, the unstructured pruning method tends to reveal some structural patterns post-pruning, which mirrors the geometric interdependence inherent in the LLMs' parameter structure. Additionally, GBLM-Pruner functions without any subsequent retraining or weight updates to maintain its simplicity as other counterparts. Extensive evaluations on LLaMA-1 and LLaMA-2 across various language benchmarks and perplexity show that GBLM-Pruner surpasses magnitude pruning, Wanda (*weights+activations*), and SparseGPT (*weights+activations+weight update*) by significant margins. Our code and models will be publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "EfNnPFnEDk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission207/Reviewer_k6s5"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This study introduces GBLM-Pruner, a new post-training pruning technique designed for large language models, which leverages gradient information. The authors provide both theoretical rationale and empirical assessments that demonstrate GBLM-Pruner outperforms other prominent baselines, such as Wanda and SparseGPT.", "review_text": "This study introduces GBLM-Pruner, a new post-training pruning technique designed for large language models, which leverages gradient information. The authors provide both theoretical rationale and empirical assessments that demonstrate GBLM-Pruner outperforms other prominent baselines, such as Wanda and SparseGPT.", "strengths": "- The paper is well-organized, effectively presenting the method with clear descriptions and comprehensive empirical evaluations.\n- Both theoretical explanations and empirical results are presented to validate the theoretical explanations and empirical results.\n- The paper includes plenty of ablation studies, encompassing diverse sparsity levels, different pruning metrics, assessments of dependency on calibration samples, and visualizations that highlight the specifics of sparse patterns.", "weaknesses": "- The improvements achieved by GBLM-Pruner, as compared to other baselines like SparseGPT and Wanda, appear to be relatively modest. For instance, in Table 2, under 50% unstructured sparsity, GBLM-Pruner (l1) yields perplexity reductions of only 0.06, 0.09, and 0.05 compared to Wanda on LLaMA-2-7B/13B/70B, respectively. Additionally, in Figure 2, the curves for Wanda and GBLM-Pruner exhibit significant overlap.\n  \n- I'm unclear about the rationale behind experimenting with the pruning metrics listed in Line 7/8. It seems that some of these metrics may not provide meaningful insights.\n\n- It's essential to understand the memory and time requirements during the pruning process of GBLM-Pruner. Obtaining gradient information can impose a significant memory cost, and it may not be feasible to conduct this process in a layer-wise manner. Storing intermediate features for the backward process could further impact memory usage. Thus, it would be valuable to compare these memory and time requirements with those of other baseline methods for a more comprehensive assessment of GBLM-Pruner's practicality.", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces GBLM-Pruner, a new post-training pruning technique designed for large language models, which leverages gradient information. The authors provide both theoretical rationale and empirical assessments that demonstrate GBLM-Pruner outperforms other prominent baselines, such as Wanda and SparseGPT.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "- The paper is well-organized, effectively presenting the method with clear descriptions and comprehensive empirical evaluations.\n- Both theoretical explanations and empirical results are presented to validate the theoretical explanations and empirical results.\n- The paper includes plenty of ablation studies, encompassing diverse sparsity levels, different pruning metrics, assessments of dependency on calibration samples, and visualizations that highlight the specifics of sparse patterns.", "weaknesses": "- The improvements achieved by GBLM-Pruner, as compared to other baselines like SparseGPT and Wanda, appear to be relatively modest. For instance, in Table 2, under 50% unstructured sparsity, GBLM-Pruner (l1) yields perplexity reductions of only 0.06, 0.09, and 0.05 compared to Wanda on LLaMA-2-7B/13B/70B, respectively. Additionally, in Figure 2, the curves for Wanda and GBLM-Pruner exhibit significant overlap.\n  \n- I'm unclear about the rationale behind experimenting with the pruning metrics listed in Line 7/8. It seems that some of these metrics may not provide meaningful insights.\n\n- It's essential to understand the memory and time requirements during the pruning process of GBLM-Pruner. Obtaining gradient information can impose a significant memory cost, and it may not be feasible to conduct this process in a layer-wise manner. Storing intermediate features for the backward process could further impact memory usage. Thus, it would be valuable to compare these memory and time requirements with those of other baseline methods for a more comprehensive assessment of GBLM-Pruner's practicality.", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698854712805}, {"id": "8FOyF3WZuE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission207/Reviewer_B7C2"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "* The paper proposes to integrate gradient information into pruning criteria currently used for LLMs.\n* The corresponding GBLM method is evaluated on Llama models for perplexity and zero-shot tasks.", "review_text": "* The paper proposes to integrate gradient information into pruning criteria currently used for LLMs.\n* The corresponding GBLM method is evaluated on Llama models for perplexity and zero-shot tasks.", "strengths": "* The paper is easy to follow and describes the proposed method in good detail.\n* The method is evaluated on strong LLama models rather than older LLMs like OPT.\n* Source code is provided, aiding reproducability.", "weaknesses": "* Integrating gradient information into pruning criteria is a well studied area, see for example [1, 2, 3, 4]. This is currently not discussed under Related Work.\n* Consequently, the novelty of GBLM is quite limited. For instance, the analysis in Section 2.3 is very similar to derivations presented in [2]. Ultimately, GBLM seems to be a minor variation of a diagonal Fisher scheme (using both gradients and activations while slightly tweaking norms in a heuristic manner).\n* The most robust form of evaluation, perplexity, shows only very slight improvements relative to prior work of < 0.1 points, while dropping noticably from the baseline. I am not sure if this is a significant enough improvement in practice.\n* It is unclear how the gradient calculation impacts the speed and compute/memory requirements of the pruning process. Being fast and memory efficient is one of the key strengths of SparseGPT and Wanda, hence I think a detailed comparison/discussion of this aspect would be important.\n\nUnfortunately, at this time, I find neither the method itself nor the empirical results interesting enough to recommend acceptance.\n\n[1] Pruning convolutional neural networks for resource efficient inference, Molchanov et al.\n\n[2] WoodFisher: Efficient Second-Order Approximation for Neural Network Compression, Singh et al.\n\n[4] The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models, Kurtic et al.\n\n[3] Movement Pruning: Adaptive Sparsity by Fine-Tuning, Sanh et al.", "questions": "* See weaknesses, in particular the compute/memory efficiency point.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "* The paper proposes to integrate gradient information into pruning criteria currently used for LLMs.\n* The corresponding GBLM method is evaluated on Llama models for perplexity and zero-shot tasks.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* The paper is easy to follow and describes the proposed method in good detail.\n* The method is evaluated on strong LLama models rather than older LLMs like OPT.\n* Source code is provided, aiding reproducability.", "weaknesses": "* Integrating gradient information into pruning criteria is a well studied area, see for example [1, 2, 3, 4]. This is currently not discussed under Related Work.\n* Consequently, the novelty of GBLM is quite limited. For instance, the analysis in Section 2.3 is very similar to derivations presented in [2]. Ultimately, GBLM seems to be a minor variation of a diagonal Fisher scheme (using both gradients and activations while slightly tweaking norms in a heuristic manner).\n* The most robust form of evaluation, perplexity, shows only very slight improvements relative to prior work of < 0.1 points, while dropping noticably from the baseline. I am not sure if this is a significant enough improvement in practice.\n* It is unclear how the gradient calculation impacts the speed and compute/memory requirements of the pruning process. Being fast and memory efficient is one of the key strengths of SparseGPT and Wanda, hence I think a detailed comparison/discussion of this aspect would be important.\n\nUnfortunately, at this time, I find neither the method itself nor the empirical results interesting enough to recommend acceptance.\n\n[1] Pruning convolutional neural networks for resource efficient inference, Molchanov et al.\n\n[2] WoodFisher: Efficient Second-Order Approximation for Neural Network Compression, Singh et al.\n\n[4] The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models, Kurtic et al.\n\n[3] Movement Pruning: Adaptive Sparsity by Fine-Tuning, Sanh et al.", "questions": "* See weaknesses, in particular the compute/memory efficiency point.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698757865107}, {"id": "wVcKzLRO1W", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission207/Reviewer_gsUn"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes to integrate the first-order gradient into the unstructured pruning of large language models and achieves superior performance compared to sparseGPT and Wanda.", "review_text": "This paper proposes to integrate the first-order gradient into the unstructured pruning of large language models and achieves superior performance compared to sparseGPT and Wanda.", "strengths": "1. A superior method compared to SparseGPT and Wanda on unstructured pruning of large language model\n2. The authors have conducted extensive experiments to assess the method's effectiveness on LLaMa-1 and LLaMa-2. Additionally, the paper illustrates the impact of various gradient and activation combinations on the determination of parameter importance.\n3. The paper is well-written, offering clarity and ease of understanding in its presentation.", "weaknesses": "1. The novelty of this method appears somewhat constrained. Utilizing the first-order gradient for determining parameter importance is a common approach in pruning techniques applied to CNN, BERT, and ViT. This technique is well-established within the realm of model pruning. Considering in some instances this method even falls short of those achieved by SparseGPT (e.g., 2:4 for LLaMA-1 and LLaMA-2), I cannot say the first-order gradient in pruning LLMs might be a major contribution.\n2. This paper lacks experiments on different LLM families. Conducting trials with models like OPT, BLOOM, or other alternatives could provide valuable insights into the method's applicability and generalizability across various LLM families.\n3. The paper doesn't provide details regarding the latency of the pruned model. In a study centered on LLM compression, including latency metrics is crucial since such information is highly important  to the readers to understand the efficiency of the pruned model.", "questions": "1. Could you specify the error function utilized for calculating gradients in your approach?\n2. Have you conducted any latency experiments on the pruned model, particularly under the 2:4 or 4:8 configurations?\n3. Is the calibration set employed for your methods and Wanda, SparseGPT identical?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to integrate the first-order gradient into the unstructured pruning of large language models and achieves superior performance compared to sparseGPT and Wanda.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. A superior method compared to SparseGPT and Wanda on unstructured pruning of large language model\n2. The authors have conducted extensive experiments to assess the method's effectiveness on LLaMa-1 and LLaMa-2. Additionally, the paper illustrates the impact of various gradient and activation combinations on the determination of parameter importance.\n3. The paper is well-written, offering clarity and ease of understanding in its presentation.", "weaknesses": "1. The novelty of this method appears somewhat constrained. Utilizing the first-order gradient for determining parameter importance is a common approach in pruning techniques applied to CNN, BERT, and ViT. This technique is well-established within the realm of model pruning. Considering in some instances this method even falls short of those achieved by SparseGPT (e.g., 2:4 for LLaMA-1 and LLaMA-2), I cannot say the first-order gradient in pruning LLMs might be a major contribution.\n2. This paper lacks experiments on different LLM families. Conducting trials with models like OPT, BLOOM, or other alternatives could provide valuable insights into the method's applicability and generalizability across various LLM families.\n3. The paper doesn't provide details regarding the latency of the pruned model. In a study centered on LLM compression, including latency metrics is crucial since such information is highly important  to the readers to understand the efficiency of the pruned model.", "questions": "1. Could you specify the error function utilized for calculating gradients in your approach?\n2. Have you conducted any latency experiments on the pruned model, particularly under the 2:4 or 4:8 configurations?\n3. Is the calibration set employed for your methods and Wanda, SparseGPT identical?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698730280551}, {"id": "u3zpfXbdew", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission207/Reviewer_dK5u"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This study introduces GBLM-Pruner, a gradient-based approach for the unstructured pruning of large language models (LLMs). The core idea of this research is centered around a Taylor expansion applied to the loss function. This method estimates the change in loss by employing a combination of first-order gradient and second-order approximation (OBD). Empirical evaluations using LLaMA and LLaMA-2 demonstrate that GBLM-Pruner outperforms other methods such as magnitude pruning, SparseGPT, and Wanda in terms of performance.", "review_text": "This study introduces GBLM-Pruner, a gradient-based approach for the unstructured pruning of large language models (LLMs). The core idea of this research is centered around a Taylor expansion applied to the loss function. This method estimates the change in loss by employing a combination of first-order gradient and second-order approximation (OBD). Empirical evaluations using LLaMA and LLaMA-2 demonstrate that GBLM-Pruner outperforms other methods such as magnitude pruning, SparseGPT, and Wanda in terms of performance.", "strengths": "1. This paper highlights the significance of gradients in the pruning of large language models (LLMs). The author presents a Taylor-based approach to identify critical parameters, yielding favorable outcomes in comparison to earlier techniques.\n2. The work sets robust benchmarks by contrasting the proposed methods with various existing baselines, offering valuable insights for the research community.", "weaknesses": "1. To my knowledge, SparseGPT is similarly a gradient-based approach, utilizing Taylor expansion and second-order Hessian for estimating parameter importance. In light of this, the contribution of the current work may appear somewhat constrained.\n2. As depicted in Figure 2, SparseGPT, Wanda, and the newly introduced GBLM-Pruner exhibit closely comparable results, with only minor differences in Perplexity (PPL). There isn't compelling evidence to suggest that GBLM-Pruner significantly outperforms its predecessors.\n3. It would be beneficial if the author could include data on the latency of the pruned LLMs, particularly in the context of 2:4 sparsity acceleration.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces GBLM-Pruner, a gradient-based approach for the unstructured pruning of large language models (LLMs). The core idea of this research is centered around a Taylor expansion applied to the loss function. This method estimates the change in loss by employing a combination of first-order gradient and second-order approximation (OBD). Empirical evaluations using LLaMA and LLaMA-2 demonstrate that GBLM-Pruner outperforms other methods such as magnitude pruning, SparseGPT, and Wanda in terms of performance.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper highlights the significance of gradients in the pruning of large language models (LLMs). The author presents a Taylor-based approach to identify critical parameters, yielding favorable outcomes in comparison to earlier techniques.\n2. The work sets robust benchmarks by contrasting the proposed methods with various existing baselines, offering valuable insights for the research community.", "weaknesses": "1. To my knowledge, SparseGPT is similarly a gradient-based approach, utilizing Taylor expansion and second-order Hessian for estimating parameter importance. In light of this, the contribution of the current work may appear somewhat constrained.\n2. As depicted in Figure 2, SparseGPT, Wanda, and the newly introduced GBLM-Pruner exhibit closely comparable results, with only minor differences in Perplexity (PPL). There isn't compelling evidence to suggest that GBLM-Pruner significantly outperforms its predecessors.\n3. It would be beneficial if the author could include data on the latency of the pruned LLMs, particularly in the context of 2:4 sparsity acceleration.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698403164216}], "openreview_url": "https://openreview.net/forum?id=5BoXZXTJvL", "arxiv_id": "2311.04902", "paper_pdf": "papers/5BoXZXTJvL.pdf", "paper_pdf_sha256": "105bf01bb58e356452a1975b942bb4fd4d510bfb0fc75908858f828bec26cb60", "paper_pdf_bytes": 2882889, "paper_pdf_source": "openreview", "code_url": "https://github.com/VILA-Lab/GBLM-Pruner", "code_repository": "VILA-Lab/GBLM-Pruner", "code_commit": "b59221570017d2e1dac45cb1c16f1c54ba4d8ea1", "code_archive": "repos/5BoXZXTJvL.zip", "code_archive_sha256": "1aee469d557f5cbd5f5bd74e35c8dbf4b9ff2e68fd0d13a086e15bd717bb9e24", "code_archive_bytes": 31903, "code_file_count": 12, "code_extensions": {".py": 7, ".sh": 5}, "github_disk_usage_kb": 38, "github_languages": {"Python": 45290, "Shell": 1001}, "github_archived": false, "github_pushed_at": "2025-08-23T11:51:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beyond-size-how-gradients-shape-pruning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9rRhMKNOkeT", "year": 2023, "status": "rejected", "title": "Concept-based Explanations for Out-of-Distribution Detectors", "authors": ["Jihye Choi", "Jayaram Raghuram", "Ryan Feng", "Jiefeng Chen", "Somesh Jha", "Atul Prakash"], "authorids": ["~Jihye_Choi1", "~Jayaram_Raghuram1", "~Ryan_Feng1", "~Jiefeng_Chen2", "~Somesh_Jha1", "~Atul_Prakash1"], "authors_source": "OpenReview API", "abstract": "Out-of-distribution (OOD) detection plays a crucial role in ensuring the safe deployment of deep neural network (DNN) classifiers.\nWhile a myriad of methods have focused on improving the performance of OOD detectors, a critical gap remains in interpreting their decisions.\nWe help bridge this gap by providing explanations for OOD detectors based on learned high-level concepts.\nWe first propose two new metrics for assessing the effectiveness of a particular set of concepts for explaining OOD detectors: 1) $\\textit{detection completeness}$, which quantifies the sufficiency of concepts for explaining an OOD-detector's decisions, and 2) $\\textit{concept separability}$, which captures the distributional separation between in-distribution and OOD data in the concept space.\nBased on these metrics, we propose a framework for learning a set of concepts that satisfy the desired properties of detection completeness and concept separability, and demonstrate the framework's effectiveness in providing concept-based explanations for diverse OOD detection techniques.\nWe also show how to identify prominent concepts that contribute to the detection results via a modified Shapley value-based importance score.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "L7O5Z3dKfOL", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4789/Reviewer_YcEH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper addresses the problem of explaining the decisions of an out-of-distribution (OOD) detector by using concept-based representations. In order to learn concept-based explanation, the authors propose two measures, namely the concept detection completeness and the concept separability, which they optimize during training of OOD detectors.  ", "review_text": "The paper presents a novel idea, which is presented highlighting its relevance. The research is motivated and carefully constructed. The experimental analysis, although with several existing methods and on relevant datasets, brings to unfair statements in comparison with other methods, and lacks addressing strongly the validity of the proposed measures. Also, a more thorough analysis of the learned concepts should be provided: how are they relevant and what do they cover?\nOverall is a good work, but I would like to see the response of the authors to my questions and doubts.", "strengths": "__Strenghts__\n- the problem addressed is relevant, and ground motivations for the research are discussed\n- the paper is well written and carefully constructed, easy-to-follow and clear\n- the new metrics are technically sound\n- experiments in combination with existing methods\n\n__Weaknesses__\n- In my opinion, the main concern/weakness is in the way the authors analyse and interpret the results: the authors stress the fact that existingmethods are not able to guarantee concept separability and detection completeness, which are metrics that they define in this work. On the contrary, the method based on learning concepts by optimizing these metrics is performing better. There is no surprise in these results, as previous methods are not optimized for the considered metrics. Thus, it seems to me that the experimentations (in the way it is presented) is unfair. \n\n- Another weakness is in the lack of details about the deployement of their approach within existing approaches: it seems a concept-based learning component (optimized with the proposed metrics) is deployed within intermediate layers, but no details about the 'where' in the network (which layer) it is used, and how it may change if deployed in different parts. \n\n- Connected to above, if this can be deployed in different parts of the network, I imagine that different concepts are learned, of different semantics. How does it relate with the 'high-level' concepts that the authors aim at learning? Also, how the 'high-level' is defined?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper addresses the problem of explaining the decisions of an out-of-distribution (OOD) detector by using concept-based representations. In order to learn concept-based explanation, the authors propose two measures, namely the concept detection completeness and the concept separability, which they optimize during training of OOD detectors.  ", "strength_and_weaknesses": "__Strenghts__\n- the problem addressed is relevant, and ground motivations for the research are discussed\n- the paper is well written and carefully constructed, easy-to-follow and clear\n- the new metrics are technically sound\n- experiments in combination with existing methods\n\n__Weaknesses__\n- In my opinion, the main concern/weakness is in the way the authors analyse and interpret the results: the authors stress the fact that existingmethods are not able to guarantee concept separability and detection completeness, which are metrics that they define in this work. On the contrary, the method based on learning concepts by optimizing these metrics is performing better. There is no surprise in these results, as previous methods are not optimized for the considered metrics. Thus, it seems to me that the experimentations (in the way it is presented) is unfair. \n\n- Another weakness is in the lack of details about the deployement of their approach within existing approaches: it seems a concept-based learning component (optimized with the proposed metrics) is deployed within intermediate layers, but no details about the 'where' in the network (which layer) it is used, and how it may change if deployed in different parts. \n\n- Connected to above, if this can be deployed in different parts of the network, I imagine that different concepts are learned, of different semantics. How does it relate with the 'high-level' concepts that the authors aim at learning? Also, how the 'high-level' is defined?", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear, stating the addressed problem and the relation with existing works well. The conclusion section is not really reporting conclusions, but rather discussing limitations and societal impact of the work, leaving the reader pending final and concise answers to the posedresearch questions.\n\nThe quality of the paper and the research itself, including the novelty, is high, which inclines my evaluation towards the positive side. However, the experimental analysis and observations made with respect to existing works are unfair (see above). The authors should have focused on better highlighting the usefulness of the proposed concept-world and canonical-world representations, and the validation of the proposed metrics.\n\nThe work lacks reproducibility: I did not find mentions that the code will be available, neither details about the layers in between the proposed appraoch is deployed, and what effects this might have on the learned concepts and the overall explainability of the OOD detections. In its current form, the paper makes difficult to re-implement or even deploy the proposed approach to replicate the experiments.", "summary_of_the_review": "The paper presents a novel idea, which is presented highlighting its relevance. The research is motivated and carefully constructed. The experimental analysis, although with several existing methods and on relevant datasets, brings to unfair statements in comparison with other methods, and lacks addressing strongly the validity of the proposed measures. Also, a more thorough analysis of the learned concepts should be provided: how are they relevant and what do they cover?\nOverall is a good work, but I would like to see the response of the authors to my questions and doubts.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666977694301}, {"id": "gvtV7yaJM87", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4789/Reviewer_qdMm"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method to interpret out-of-distribution detection using learnt high-level concepts.", "review_text": "The paper poses an interesting question of how to use concepts to generate explanations from OOD detectors. While the evaluation and idea is adequate, it is hard for me to recommend accept without a convincing user study or extensive qualitative examples of why such an explanation makes sense. ", "strengths": "Strengths\n- To my knowledge, this is the first work to study the notion of concept-based explanations for OOD detection.\n- The idea of LDA on the concept layer is clever and simple.\n- The proposed metrics are easy to understand. Detection completeness and concept separability naturally follow.\n\nWeaknesses\n- I would have expected more examples of what a concept-based explanation for an OOD detector means. Figure 1 is unconvincing. Please provide more qualitative evidence of the utility of concept-based explanations in the OOD setting.\n- While the authors acknowledge the lack of a human subject experiment, it is not clear from the prose alone if concept-based explanations are even necessary (let alone helpful) for OOD detectors. I, among others, am unsure why concepts for OOD detectors. How does it help? A carefully designed user study would right this. It would also warrant a method that would lead to higher adoption and impact down the line.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a method to interpret out-of-distribution detection using learnt high-level concepts.", "strength_and_weaknesses": "Strengths\n- To my knowledge, this is the first work to study the notion of concept-based explanations for OOD detection.\n- The idea of LDA on the concept layer is clever and simple.\n- The proposed metrics are easy to understand. Detection completeness and concept separability naturally follow.\n\nWeaknesses\n- I would have expected more examples of what a concept-based explanation for an OOD detector means. Figure 1 is unconvincing. Please provide more qualitative evidence of the utility of concept-based explanations in the OOD setting.\n- While the authors acknowledge the lack of a human subject experiment, it is not clear from the prose alone if concept-based explanations are even necessary (let alone helpful) for OOD detectors. I, among others, am unsure why concepts for OOD detectors. How does it help? A carefully designed user study would right this. It would also warrant a method that would lead to higher adoption and impact down the line.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper could be more clear as to why concepts for OOD detectors are sensible. \nNovelty: The idea depends heavily on existing concept-based explanation work, carefully applied to the OOD setting.\nQuality: The evaluation is adequate.\nReproducibility: Adequate.", "summary_of_the_review": "The paper poses an interesting question of how to use concepts to generate explanations from OOD detectors. While the evaluation and idea is adequate, it is hard for me to recommend accept without a convincing user study or extensive qualitative examples of why such an explanation makes sense. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666885509923}, {"id": "c1SV3IZfUw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4789/Reviewer_v8DY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes concept-based learning and an explanation for OoD Detection. The authors propose metrics to capture detection completeness and separability, which are used for the learning process. The evaluation of multiple datasets demonstrates the potential usefulness of the proposed techniques for OoD explanation.", "review_text": "Overall, I like the idea of this paper and think it works on an important problem, which does not receive enough attention previously. This paper makes an early step in explaining OoD detectors in relatively general settings without the requirement of white box. However, I still have a few concerns posted in the weakness part, that hope the authors could address during the rebuttal phase and add relevant discussion in the next version.", "strengths": "Strength:\n\n- Important topic and work on OoD detection explanations\n- Proposed techniques are general and could be applicable to various contexts, without the whitebox requirement\n- Feasible solution with promising results\n\nWeakness:\n\n- Unclear about the robustness of the learned “concepts”\n- Not very clear how to design the concept space or concept learning architecture\n- Although High level and simple concept explanations are possible, it is not clear how such explanations could be helpful to guide the OoD detection developers in an actionable and constructive way.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes concept-based learning and an explanation for OoD Detection. The authors propose metrics to capture detection completeness and separability, which are used for the learning process. The evaluation of multiple datasets demonstrates the potential usefulness of the proposed techniques for OoD explanation.", "strength_and_weaknesses": "Strength:\n\n- Important topic and work on OoD detection explanations\n- Proposed techniques are general and could be applicable to various contexts, without the whitebox requirement\n- Feasible solution with promising results\n\nWeakness:\n\n- Unclear about the robustness of the learned “concepts”\n- Not very clear how to design the concept space or concept learning architecture\n- Although High level and simple concept explanations are possible, it is not clear how such explanations could be helpful to guide the OoD detection developers in an actionable and constructive way.", "clarity,_quality,_novelty_and_reproducibility": "The paper is mostly well-organized and written. The concept-based explanation also has its novelty on the OoD detection contexts, authors also make certain technical contributions to adapt concept-based explanations to the OoD contexts, which should be sound and feasible. \n\nThe evaluation also mostly supports the authors’ claims.", "summary_of_the_review": "Overall, I like the idea of this paper and think it works on an important problem, which does not receive enough attention previously. This paper makes an early step in explaining OoD detectors in relatively general settings without the requirement of white box. However, I still have a few concerns posted in the weakness part, that hope the authors could address during the rebuttal phase and add relevant discussion in the next version.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666701096437}, {"id": "1FbhrzVCGu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4789/Reviewer_H68v"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build on Yeh et al. (2020)’s argument on concept completeness and propose to use two metrics: detection (concept) completeness and concept separability. Following such metrics, the authors then design an algorithm for learning such desired concept-based explanations. ", "review_text": "This paper tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build heavily on Yeh et al. (2020), with the key differences that (1) they introduce an additional metric, i.e., concept separability, (2) instead of focusing on classification, they focus on OOD detection. Other aspects of the paper including the completeness scores and SHAP-based evaluation are very similar to Yeh et al. Given the limited technical merit, I would place this paper marginally below the threshold. ", "strengths": "Strength \n\n+ The paper is very well written and well organized. It is a good read. \n\n+ Providing concept-based explanations to OOD detector is an interesting topic. \n\nWeaknesses\n\n- The notion of detection completeness is not entirely new. As shown in Definition 1 and 2, it builds on the classification completeness in Yeh et al. (2020). \n\n- The proposed method has a limited scope in that it only focuses on explaining OOD detector. \n\n- There are no direct comparisons between the proposed method and Yeh et al., which this paper heavily builds upon. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build on Yeh et al. (2020)’s argument on concept completeness and propose to use two metrics: detection (concept) completeness and concept separability. Following such metrics, the authors then design an algorithm for learning such desired concept-based explanations. ", "strength_and_weaknesses": "Strength \n\n+ The paper is very well written and well organized. It is a good read. \n\n+ Providing concept-based explanations to OOD detector is an interesting topic. \n\nWeaknesses\n\n- The notion of detection completeness is not entirely new. As shown in Definition 1 and 2, it builds on the classification completeness in Yeh et al. (2020). \n\n- The proposed method has a limited scope in that it only focuses on explaining OOD detector. \n\n- There are no direct comparisons between the proposed method and Yeh et al., which this paper heavily builds upon. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper builds heavily on Yeh et al. (2020) which is also related to concept-based explanation of neural networks with emphasis on the completeness desideratum, which is one of the two proposed metrics in this paper.  Specifically, the only difference between the original classification completeness of Yeh et al. and the paper is to replace the classification accuracy with an AUC score. The formulations on the concept space as well as the SHAP related evaluation also follow closely Yeh et al. \n\nOne key difference from Yeh et al. is the definition of concept separability score. The authors augment the regularization term from Yeh et al. with two other regularization terms $J_{norm}$ and $J_{mse}$, along with the LDA objective to encourage separability. Such a design makes sense to me, though compared to the original Yeh et al., the technical contribution tends to be limited. \n\nWhile this paper focuses on OOD detection, I wonder if it would could also improve results in the simple classification case. The authors claim that Yeh et al. has potential issues in terms of reconstructing the features and may lead to degradation in performance. It would therefore be interesting to see if this is true. \n\nThe SHAP-based evaluation seems to also follow directly Yeh et al. While the novelty is limited, it is nice to see that the proposed method works well in combination with SHAP. \n\nThe definition of patch and receptive field is a bit vague. Why do you call it a $a_l \\times b_l$ patch? Do you feed patches into a network rather than the whole image? \n", "summary_of_the_review": "This paper tackles the problem of concept-based explanations for deep neural network OOD detectors. The authors build heavily on Yeh et al. (2020), with the key differences that (1) they introduce an additional metric, i.e., concept separability, (2) instead of focusing on classification, they focus on OOD detection. Other aspects of the paper including the completeness scores and SHAP-based evaluation are very similar to Yeh et al. Given the limited technical merit, I would place this paper marginally below the threshold. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666537864852}], "openreview_url": "https://openreview.net/forum?id=9rRhMKNOkeT", "arxiv_id": "2203.02586", "paper_pdf": "papers/9rRhMKNOkeT.pdf", "paper_pdf_sha256": "2f77f67b4b7fde208d63bee04a537bff7da4191ff63320b69188d052264d8881", "paper_pdf_bytes": 2695565, "paper_pdf_source": "openreview", "code_url": "https://github.com/jihyechoi77/concepts-for-ood", "code_repository": "jihyechoi77/concepts-for-ood", "code_commit": "28be657c6cb8fb01b3760db29b866eb6dfd4448d", "code_archive": "repos/9rRhMKNOkeT.zip", "code_archive_sha256": "d2ab9e0b33d10a9ea6d057ea6fec36bd8e71e04e08530a8e063b307751604f91", "code_archive_bytes": 48741, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 46, "github_languages": {"Python": 145394}, "github_archived": false, "github_pushed_at": "2023-06-08T23:48:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/concept-based-explanations-for-out-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TNxKD3z_tPZ", "year": 2022, "status": "rejected", "title": "Persistent Homology Captures the Generalization of Neural Networks Without A Validation Set", "authors": ["Asier Gutiérrez-Fandiño", "David Pérez Fernández", "Jordi Armengol-Estapé", "Marta Villegas"], "authorids": ["~Asier_Gutiérrez-Fandiño1", "~David_Pérez_Fernández1", "~Jordi_Armengol-Estapé1", "~Marta_Villegas2"], "authors_source": "OpenReview API", "abstract": "The training of neural networks is usually monitored with a validation (holdout) set to estimate the generalization of the model. This is done instead of measuring intrinsic properties of the model to determine whether it is learning appropriately. In this work, we suggest studying the training of neural networks with Algebraic Topology, specifically Persistent Homology (PH). Using simplicial complex representations of neural networks, we study the PH diagram distance evolution on the neural network learning process with different architectures and several datasets. Results show that the PH diagram distance between consecutive neural network states correlates with the validation accuracy, implying that the generalization error of a neural network could be intrinsically estimated without any holdout set.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ML1BO9nJ5Ed", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4090/Reviewer_nK7m"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper uses distances between the Persistent diagrams of states of neural networks and observe that during training there is a high correlation with the corresponding validation accuracy of the model. The authors tested their method on a variety of datasets.", "review_text": "I think the subject of the paper is interesting itself and using PH techniques to discover insights about deep learning models is indeed an interesting topic. However, I find the paper to be still in early stage of development and still requires work. Specifically, It think the authors need to really support  their claim for the title. Generalization is probably the hardest problem in deep learning model models and the claim that PH truly captures that is not explained well at least at the beginning of the paper. What is the metric for such a claim ? The authors mentioned \"that there exists a high correlation with the corresponding validation accuracy of the model.\" I do not think think that is enough reason to justify the claim of the paper.\n\nThe contributions are not very novel in my opinion. For instance :\n\nBased on principles of Algebraic Topology, we propose measuring the distances (Silhouette\nand Heat) between the PH persistence diagrams obtained from a given state of a neural\nnetwork during the training procedure and the one in the immediately previous weights\nupdate.\n\nMeasuring the distance between two weights of neural network is not significant contribution, unless you do something with it--which you are claiming next but this particular point is not a valid contribution in my opinion.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper uses distances between the Persistent diagrams of states of neural networks and observe that during training there is a high correlation with the corresponding validation accuracy of the model. The authors tested their method on a variety of datasets.", "main_review": "I think the subject of the paper is interesting itself and using PH techniques to discover insights about deep learning models is indeed an interesting topic. However, I find the paper to be still in early stage of development and still requires work. Specifically, It think the authors need to really support  their claim for the title. Generalization is probably the hardest problem in deep learning model models and the claim that PH truly captures that is not explained well at least at the beginning of the paper. What is the metric for such a claim ? The authors mentioned \"that there exists a high correlation with the corresponding validation accuracy of the model.\" I do not think think that is enough reason to justify the claim of the paper.\n\nThe contributions are not very novel in my opinion. For instance :\n\nBased on principles of Algebraic Topology, we propose measuring the distances (Silhouette\nand Heat) between the PH persistence diagrams obtained from a given state of a neural\nnetwork during the training procedure and the one in the immediately previous weights\nupdate.\n\nMeasuring the distance between two weights of neural network is not significant contribution, unless you do something with it--which you are claiming next but this particular point is not a valid contribution in my opinion.", "summary_of_the_review": "I am sorry to say that I have to reject the paper. The paper is still in its early stage and requires more careful presentation especially in the beginning to make the contribution more clear and crisp ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636001418311}, {"id": "R18ptY2HKOa", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4090/Reviewer_eLVu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper presents some empirical observations about the relationship of a persistent homology-based measure of learning dynamics and validation set error, during the training of deep neural nets. The paper opens with an introduction on persistent homology, then introduces an approach to study the structure of deep nets using topological data analysis (TDA) tools. Three case studies are then presented, for which a measure of change in topological structure of the network during training is compared with the validation set error. The argument of the paper is that the two measures are correlated, and therefore it may be possible to use the topological measure (which depends on the structure of the network only, and not on data) in place of the validation set error to assess the generalization performance of a deep net.", "review_text": "One merit of the paper is, to put it in very broad terms, the idea of\napplying tools from topological data analysis to the problem of\nunderstanding generalization in deep neural networks. In principle,\nthis is an interesting problem from a theoretical standpoint, and the\nidea of applying TDA is timely in the sense that many fields are now\nin the process of \"discovering\" topological techniques, and it is good\nto see some exploration of what they can be useful for. Unfortunately\nhowever, in my opinion the strengths of the paper stop here. I will\nnow go in detail over what I consider to be its major weaknesses.\n\nEmpirical studies can be valuable, but the complete absence of any\ntheoretical justification (or even just an intuitively plausible\nstory) for the main argument means that the experimental evidence\npresented carries all the burden of convincing the reader. My main\nconcern with the paper is that, unfortunately, the analyses presented\nare nowhere close to the standard of rigor and thoroughness one would\nexpect in this case. In particular:\n\n1. Unless I missed it, the paper never mentions the evolution of the\n   training error. For all I know by reading the paper and looking at\n   the plots, the tested networks never overfit the training data, in\n   which case I expect the training error to correlate to the\n   validation set error even better than the proposed metric,\n   something that would completely invalidate the claim of the paper\n   (i.e., that the topological measure is telling us something\n   nontrivial). In other words, in my opinion the interesting quantity\n   to understand the learning dynamics is not the validation set error\n   per se, but the generalization gap (or, in other words, the\n   validation set error should be considered together with the\n   training error, and not in isolation).\n2. As far as I understand (although the paper is quite confusing in\n   this respect - see below for more on this), the complex machinery\n   of TDA is used to characterize how much the network changes during\n   training. By looking at the plots in the paper, this quantity of\n   interest (rate of change) is typically large at the beginning of\n   the training, and small towards the end. Because this metric goes\n   down quite quickly, its cumulative value during training takes on a\n   saturating trend. This trend correlates with the validation set\n   error, which also saturates. The main claim of the paper is that\n   this correlation suggests a conceptual link between the two\n   quantities. But there are many quantities that presumably exhibit\n   similar trends during training! For instance, how about the norm of\n   the weight update vector (which could be conceptualized as a much\n   simpler way of measuring change in the network, and therefore could\n   be a natural comparison to the proposed metric), or even just the\n   learning rate, which is dynamically adjusted by RMSprop in the\n   proposed case studies?  Similarly to my first point above, these\n   are examples of quantities that could trivially correlate with the\n   validation set error in the specific examples chosen, just as well\n   or better than the proposed measure, without of course carrying any\n   interesting information on the generalization gap.\n3. Besides the issues above, the main quantitative piece of evidence\n   proposed in the paper is the table with the correlation\n   coefficients. To be honest, after reading the paper I am not sure\n   what is being correlated with what here. On page 7 I read that \"*For\n   each experiment (e.g., layer size in MNIST), we plot both the\n   evolution of the PH diagram distance and the validation score\n   (accuracy). The plotted values are the corresponding means of the 5\n   repetitions with different seeds. In addition, we compute the\n   Pearson correlation for these values.*\", and I deduce that the\n   correlation must be between the validation score and the normalized\n   cumulative topological distance (actually the plots contain also\n   the non-cumulative distance, but I guess the correlation is\n   computed with the cumulative one because the non-cumulative doesn't\n   seem to be correlated with the validation score). However, later on\n   in the same page I read \"*...there is strong\n   correlation. Intuitively, this is also observed in the plots,\n   although once the distances are normalized it is not as clear to\n   visualize.*\". Does this mean that the correlation was computed on\n   the unnormalized values, that is, not on what is in the plots,\n   contradicting the statement above?\n4. The paper does not seem to connect well with the existing\n   theoretical literature on the generalization gap in deep\n   learning. For instance, it completely ignores the effort to extend\n   information-theoretic criteria to be applicable to deep\n   networks. For instance, https://arxiv.org/abs/1906.07774v1 discuss\n   how a technique introduced by Takeuchi in the '70s (itself a\n   generalization of the Akaike Information Criterion) can be used to\n   predict the generalization gap. The introduction of that paper also\n   gives several references to multiple strands of the literature that\n   seem relevant for anybody interested in proposing a new technique\n   to estimate generalization (flatness etc). Other works that are\n   relevant from a theoretical standpoint are recent approaches\n   exploring an implicit simplicity bias in DNNs (see for instance\n   https://arxiv.org/abs/1805.08522, https://arxiv.org/abs/1812.10156,\n   and more recent literature citing those).\n\nMinor points:\n1. There are several passages in the paper that I completely failed to\n   parse. For instance, \"*Corneanu et al. (2020) try to estimate the\n   performance gap between training and testing using PH\n   measures. They claim. However, one can observe some caveats.*\", \"*We\n   study the relation between the evolution of the PH diagram\n   distances with the one of the validation score with the cumulative\n   values of the distance between homologous persistence diagrams\n   because this value seems much more stable. The information of the\n   distance between the persistence diagrams has been normalized\n   [...]*\".\n2. I don't understand why the y axes in most plots are labelled\n   \"distance difference\" but the word \"difference\" never appears in\n   the text of the paper - only distances are discussed.\n3. The architecture and the pre-training of the convolutional network\n   is entirely unclear to me. It is defined as \"*In the case of CNNs,\n   the pre-trained model is defined as 3 convolutional blocks with\n   kernel size 3 (starting with 32 channels), interleaved with max\n   pooling (its linear layers are thrown away after the\n   pre-training)*\". What does \"starting with 32 channels\" mean? This\n   sounds like the authors imply that there is a standard way of\n   changing the number of channels from one layer of a conv net to the\n   next (this is not the case, as far as I know). How is max pooling\n   performed? How many, and how large, linear layers are trained and\n   then \"thrown away\" during pre-training? When does pre-training\n   stop?\n4. I do not agree with the following statement: \"*If the measured\n   distances are, indeed, related with the learning process of neural\n   networks, these variations should not have any noticeable effect.*\",\n   when discussing the various network and training settings that were\n   examined in the experiments. How is changing the learning rate\n   expected not to have a noticeable effect on a metric related to the\n   learning process? And more generally, why would one expect to\n   change things like the size of the network in such a drastic way\n   (going e.g. all the way down to 4 units per layer in size) and see\n   no effect on the learning process?\n5. When commenting on Comeanu et al 2020, the paper states that the\n   method proposed there is \"not usable in practice\". I haven't read\n   that paper, but I wonder to what degree this is just the opinion of\n   the authors of the present work. If it is, it should be justified,\n   and not passed on as a fact. Moreover, regardless of its merits (on\n   the apparent lack of which I have commented in the Major Concerns\n   section), the method proposed in the current work is enormously\n   expensive from a computational standpoint (taking a week to run on\n   a machine with 2 V100 and 1.5TB of RAM to analyze networks with a\n   few hundred neurons on problems such as MNIST), so it doesn't seem\n   like practical usability is a key area of focus of the authors.\n6. The precise definition of the metric, and some of the choices\n   behind it, are not clearly specified. On page 5, \"*For each weighted\n   directed graph associated with the state of a neural network, we\n   link a directed flag complex to it. The topological properties of\n   this directed flag complex are studied using homology groups H n\n   . We calculate the homology groups up to degree 3 (H 0 -H 3 ).*\"\n   What is a flag complex? It is mentioned elsewhere in the paper that\n   flag complexes are used in the other (unpublished, AFAICT) paper by\n   the same authors, but the method proposed here should be\n   understandable without having to go and read that other\n   paper. Also, what motivates the choice of using homology groups up\n   to degree 3?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents some empirical observations about the relationship of a persistent homology-based measure of learning dynamics and validation set error, during the training of deep neural nets. The paper opens with an introduction on persistent homology, then introduces an approach to study the structure of deep nets using topological data analysis (TDA) tools. Three case studies are then presented, for which a measure of change in topological structure of the network during training is compared with the validation set error. The argument of the paper is that the two measures are correlated, and therefore it may be possible to use the topological measure (which depends on the structure of the network only, and not on data) in place of the validation set error to assess the generalization performance of a deep net.", "main_review": "One merit of the paper is, to put it in very broad terms, the idea of\napplying tools from topological data analysis to the problem of\nunderstanding generalization in deep neural networks. In principle,\nthis is an interesting problem from a theoretical standpoint, and the\nidea of applying TDA is timely in the sense that many fields are now\nin the process of \"discovering\" topological techniques, and it is good\nto see some exploration of what they can be useful for. Unfortunately\nhowever, in my opinion the strengths of the paper stop here. I will\nnow go in detail over what I consider to be its major weaknesses.\n\nEmpirical studies can be valuable, but the complete absence of any\ntheoretical justification (or even just an intuitively plausible\nstory) for the main argument means that the experimental evidence\npresented carries all the burden of convincing the reader. My main\nconcern with the paper is that, unfortunately, the analyses presented\nare nowhere close to the standard of rigor and thoroughness one would\nexpect in this case. In particular:\n\n1. Unless I missed it, the paper never mentions the evolution of the\n   training error. For all I know by reading the paper and looking at\n   the plots, the tested networks never overfit the training data, in\n   which case I expect the training error to correlate to the\n   validation set error even better than the proposed metric,\n   something that would completely invalidate the claim of the paper\n   (i.e., that the topological measure is telling us something\n   nontrivial). In other words, in my opinion the interesting quantity\n   to understand the learning dynamics is not the validation set error\n   per se, but the generalization gap (or, in other words, the\n   validation set error should be considered together with the\n   training error, and not in isolation).\n2. As far as I understand (although the paper is quite confusing in\n   this respect - see below for more on this), the complex machinery\n   of TDA is used to characterize how much the network changes during\n   training. By looking at the plots in the paper, this quantity of\n   interest (rate of change) is typically large at the beginning of\n   the training, and small towards the end. Because this metric goes\n   down quite quickly, its cumulative value during training takes on a\n   saturating trend. This trend correlates with the validation set\n   error, which also saturates. The main claim of the paper is that\n   this correlation suggests a conceptual link between the two\n   quantities. But there are many quantities that presumably exhibit\n   similar trends during training! For instance, how about the norm of\n   the weight update vector (which could be conceptualized as a much\n   simpler way of measuring change in the network, and therefore could\n   be a natural comparison to the proposed metric), or even just the\n   learning rate, which is dynamically adjusted by RMSprop in the\n   proposed case studies?  Similarly to my first point above, these\n   are examples of quantities that could trivially correlate with the\n   validation set error in the specific examples chosen, just as well\n   or better than the proposed measure, without of course carrying any\n   interesting information on the generalization gap.\n3. Besides the issues above, the main quantitative piece of evidence\n   proposed in the paper is the table with the correlation\n   coefficients. To be honest, after reading the paper I am not sure\n   what is being correlated with what here. On page 7 I read that \"*For\n   each experiment (e.g., layer size in MNIST), we plot both the\n   evolution of the PH diagram distance and the validation score\n   (accuracy). The plotted values are the corresponding means of the 5\n   repetitions with different seeds. In addition, we compute the\n   Pearson correlation for these values.*\", and I deduce that the\n   correlation must be between the validation score and the normalized\n   cumulative topological distance (actually the plots contain also\n   the non-cumulative distance, but I guess the correlation is\n   computed with the cumulative one because the non-cumulative doesn't\n   seem to be correlated with the validation score). However, later on\n   in the same page I read \"*...there is strong\n   correlation. Intuitively, this is also observed in the plots,\n   although once the distances are normalized it is not as clear to\n   visualize.*\". Does this mean that the correlation was computed on\n   the unnormalized values, that is, not on what is in the plots,\n   contradicting the statement above?\n4. The paper does not seem to connect well with the existing\n   theoretical literature on the generalization gap in deep\n   learning. For instance, it completely ignores the effort to extend\n   information-theoretic criteria to be applicable to deep\n   networks. For instance, https://arxiv.org/abs/1906.07774v1 discuss\n   how a technique introduced by Takeuchi in the '70s (itself a\n   generalization of the Akaike Information Criterion) can be used to\n   predict the generalization gap. The introduction of that paper also\n   gives several references to multiple strands of the literature that\n   seem relevant for anybody interested in proposing a new technique\n   to estimate generalization (flatness etc). Other works that are\n   relevant from a theoretical standpoint are recent approaches\n   exploring an implicit simplicity bias in DNNs (see for instance\n   https://arxiv.org/abs/1805.08522, https://arxiv.org/abs/1812.10156,\n   and more recent literature citing those).\n\nMinor points:\n1. There are several passages in the paper that I completely failed to\n   parse. For instance, \"*Corneanu et al. (2020) try to estimate the\n   performance gap between training and testing using PH\n   measures. They claim. However, one can observe some caveats.*\", \"*We\n   study the relation between the evolution of the PH diagram\n   distances with the one of the validation score with the cumulative\n   values of the distance between homologous persistence diagrams\n   because this value seems much more stable. The information of the\n   distance between the persistence diagrams has been normalized\n   [...]*\".\n2. I don't understand why the y axes in most plots are labelled\n   \"distance difference\" but the word \"difference\" never appears in\n   the text of the paper - only distances are discussed.\n3. The architecture and the pre-training of the convolutional network\n   is entirely unclear to me. It is defined as \"*In the case of CNNs,\n   the pre-trained model is defined as 3 convolutional blocks with\n   kernel size 3 (starting with 32 channels), interleaved with max\n   pooling (its linear layers are thrown away after the\n   pre-training)*\". What does \"starting with 32 channels\" mean? This\n   sounds like the authors imply that there is a standard way of\n   changing the number of channels from one layer of a conv net to the\n   next (this is not the case, as far as I know). How is max pooling\n   performed? How many, and how large, linear layers are trained and\n   then \"thrown away\" during pre-training? When does pre-training\n   stop?\n4. I do not agree with the following statement: \"*If the measured\n   distances are, indeed, related with the learning process of neural\n   networks, these variations should not have any noticeable effect.*\",\n   when discussing the various network and training settings that were\n   examined in the experiments. How is changing the learning rate\n   expected not to have a noticeable effect on a metric related to the\n   learning process? And more generally, why would one expect to\n   change things like the size of the network in such a drastic way\n   (going e.g. all the way down to 4 units per layer in size) and see\n   no effect on the learning process?\n5. When commenting on Comeanu et al 2020, the paper states that the\n   method proposed there is \"not usable in practice\". I haven't read\n   that paper, but I wonder to what degree this is just the opinion of\n   the authors of the present work. If it is, it should be justified,\n   and not passed on as a fact. Moreover, regardless of its merits (on\n   the apparent lack of which I have commented in the Major Concerns\n   section), the method proposed in the current work is enormously\n   expensive from a computational standpoint (taking a week to run on\n   a machine with 2 V100 and 1.5TB of RAM to analyze networks with a\n   few hundred neurons on problems such as MNIST), so it doesn't seem\n   like practical usability is a key area of focus of the authors.\n6. The precise definition of the metric, and some of the choices\n   behind it, are not clearly specified. On page 5, \"*For each weighted\n   directed graph associated with the state of a neural network, we\n   link a directed flag complex to it. The topological properties of\n   this directed flag complex are studied using homology groups H n\n   . We calculate the homology groups up to degree 3 (H 0 -H 3 ).*\"\n   What is a flag complex? It is mentioned elsewhere in the paper that\n   flag complexes are used in the other (unpublished, AFAICT) paper by\n   the same authors, but the method proposed here should be\n   understandable without having to go and read that other\n   paper. Also, what motivates the choice of using homology groups up\n   to degree 3?\n", "summary_of_the_review": "Despite investigating the application of a promising set of techniques\nto an interesting problem, this paper fails to sufficiently support\nits claims. The main conclusion rests upon the correlation of a\nproposed measure of change in deep nets during training with\nvalidation set accuracy, but no effort has been made to control for\npossible confounds that could explain such correlation\ntrivially. Moreover, multiple important passages in the paper are\nextremely hard (or indeed impossible as far as I'm concerned) to\nfollow, and noticeable gaps are present in the references to related\nliterature.\n", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635762869792}, {"id": "53ChJiluxyG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4090/Reviewer_Fo4q"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper analyses the training of neural networks from a topological\nperspective, presenting a pipeline that can measure (pseudo) distances\nbetween the network's weights during training. Such information is then\nemployed to study the generalisation error of a neural network.\n\nIn contrast to existing methods for estimating this error, this paper\ndoes *not* require a specific hold-out data set, as topological features\nof the neural network are monitored during training. This frees up\nadditional data for fitting, which can be highly relevant in the sparse\ndata regime.", "review_text": "I enjoyed the ideas presented in this paper; the analysis of\ngeneralisation performance is a highly relevant and timely, and indeed\nconstitutes one of the largest obstacle toward employing machine\nlearning techniques in the wild. Understanding generalisation without\nrequiring additional data has the potential to improve machine learning\ntechniques to a substantial extent.\n\nThat being said, the current paper suffers from some issues, which\nprevent my endorsement at this point:\n\n1. Clarity: while background information on topological data analysis is\n   provided (which I appreciate; in particular in light of the fact that\n   TDA is still a rather novel occurrence at ML conferences), the method\n   itself could be described in more detail. In particular, some aspects\n   should be discussed in Section 4:\n\n    - The use of additional representations for topological features\n      needs to be clarified. At present, the main paper does not contain\n      any description of the challenges in calculating distances or the\n      need for additional representations. At the very least, a brief\n      description of these topological representations is required, so\n      that readers may better understand the remainder of the paper.\n      I find it particularly problematic that no stability guarantees of\n      these representations are discussed; they are not *just* drop-in\n      replacements for the bottleneck or Wasserstein distances.\n\n    - Overall, this section should be rewritten with a clear 'roadmap'\n      in mind. What is the problem you want to tackle and how do you\n      intend to tackle it in this paper?\n\n2. Delineation to existing work: The ICLR paper by Rieck et al. (2019b)\n   appears to already contain a large amount of the material proposed in\n   this paper. A cursory reading shows that Rieck et al. even discuss\n   generalisation performance in an early stopping setting (more about\n   this later). It is therefore critical that the contributions of this\n   paper are more clearly delineated from existing work. From my\n   understanding of both papers, I would say that the current submission\n   improves on the following aspects:\n\n    - Choice of filtration for such neural networks\n\n    - The use of other topological representations (whereas Rieck et al.\n      only use a summary statistic of persistence diagrams).\n\n   These differences (and potentially all others that I missed) should be\n   briefly discussed; the advantage would be that the paper can refer to\n   the previous publication as a justification of the method itself!\n\n   The previous work by Rieck et al. also proposes a set of experiments\n   that would be very fitting to perform here: the experiments proposed\n   in this paper could be substantially strengthened by the addition of certain\n   (simple) monitoring baselines that serve to showcase the added value\n   of employing topological information in the first place here. Rieck et\n   al. compared their topological measure to a 'patience criterion' based\n   on validation loss, thus assessing the difference in performance\n   and epochs after the early stopping criterion was applied). In my\n   opinion, this is most important experiment here when assessing\n   validation performance. Also, it would be interesting to add simple\n   baselines measures to check that these are *insufficient* to be used\n   alone. For instance, a non-topological baseline that could be assessed\n   in this context would be a distribution of weights; it should not be\n   possible to assess validation performance with this baseline.\n\n2. Experimental depth: the experimental section is lacking depth, in\n   particular given the specific goals of this paper. The depicted plots\n   and correlations are useful, but don't serve to highlight the\n   benefits of the proposed method. I would therefore suggest a more\n   thorough experimental setup such as the one shown in the supplement\n   of Rieck et al. (2019b): as mentioned in the previous point, their\n   proposed measure of network complexity is shown as additional early\n   stopping criterion (without an hold-out validation data set as well).\n\n   A post-hoc experiment of this sort would be extremely useful in\n   demonstrating the utility of the method, and it would even enable the\n   quantification of certain measures. The correlation is also useful in\n   this context, but since the goal is to assess the generalisation\n   error, an experiment in which the new measure is directly applied\n   would be extremely worthwhile.\n\nPlease see below for detailed comments.\n\n## Detailed comments\n\n- The first contribution reads a little bit like a sentence fragment to\n  me. I would suggest to use the terms 'silhouette distance' and 'heat\n  kernel distance' instead of plain 'silhouette' and 'heat'.\n\n- The terminology 'homological convergence' is slightly misleading\n  because the paper studies persistent homology, as far as I understand.\n\n- The introduction of topological concepts in the background section\n  jumps between the 'geometrical' and the 'abstract'; I understand that\n  both perspectives are valid, but I would suggest to choose only\n  a single one.\n\n- The discussion of chains and boundaries requires more explanations (or\n  could be partially relegated to the supplementary materials). For\n  instance, please clarify the 'signed combination.' However,\n  definitions need to appear in the main text, not only in the\n  supplementary materials.\n\n- In figure 4, the $3$-simplex should consist of four $2$-simplices, but\n  only two are shown. Either clarify this in the captions or update the\n  figure accordingly.\n\n- The introduction of filtrations is not motivated; the 'nested family\n  of simplicial complexes' could be introduced by building more\n  intuition here.\n\n- The related work discussion should, as mentioned above, delineate this\n  paper from Rieck et al. (2019b).\n\n- The heading 'Algebraic Topology object' is rather confusing to me;\n  I would suggest to rewrite it as 'directed flag complex', for\n  instance.\n\n- As outlined above, the distance calculation lacks information about\n  stability and other guarantees. There is also a broken reference on p.\n  5, which needs to be rectified.\n\n- In the experimental section, the use of the MLP should be explained in\n  more detail. What does it do and how does it work exactly?\n\n- I do not understand the 'input order' experiment; does it refer to\n  ordering of samples (i.e. on the batch level)?\n\n- The use of additional topological descriptors makes the use of the\n  word 'distance' imprecise: the bottleneck and Wasserstein\n  distances are metrics in the mathematical sense; while there *is*\n  a well-defined metric between, for instance, persistence landscapes or\n  silhouettes, this is only (to a certain extent) an approximation to\n  the bottleneck/Wasserstein metrics. This distinction should be made\n  clearer in the paper.\n\n- Figure 6 should add standard deviations of the curves. An additional\n  clarification of the content of this figure would also be appreciated\n  by readers; I found the discussion to be slightly confusing because\n  I did not get the relevance of the cumulative distance. The statement\n  'the evolution of the homological convergence [...] seems to be very\n  similar to the one of the validation score' also needs some\n  clarification. Except for Figure 6b, where I observe some oscillation\n  behaviour and maybe the tendency to converge, I don't observe\n  convergence anywhere else.\n\n- The same applies to Figures 7–9. Adding some more details here will\n  result in a clearer description of the method. Maybe some of the\n  curves could also be relegated to the supplementary materials?\n\n- I would also suggest to investigate the use of energy distances or\n  energy correlations, as these measures are more flexible and capable\n  of assessing more than just linear dependencies between two\n  variables. See [*Energy statistics: A class of statistics based on distances*](https://www.sciencedirect.com/science/article/abs/pii/S0378375813000633)\n  for more details.\n\n## Terminology and style\n\nFor the most part, the paper is written well. I have some suggestions\nconcerning the style:\n\n- Please use `\\citep` and `\\citet` consistently when using `natbib`. At\n  present, citations are directly appearing in the text without any\n  enclosing parentheses. For instance, in the related work section,\n  the sentence 'for improving the training procedure of the models Hofer\n  et al. (2020); [...]' should use `\\citep` in order to obtain a proper\n  parenthetical citation.\n\n- can't --> cannot (likewise for other contractions; I admit that this\n  is a personal preference, so feel free to ignore it)\n\n- 'contained simplicial complex' --> 'nested simplicial complexes'\n\n- The paper employs non-standard terminology in certain places. The\n  diagram, for instance, is called 'persistence diagram', no\n  'persistence homology diagram'. Please ensure consistency with\n  existing papers here. The same applies to capitalisation; I personally\n  see no need to write 'Persistence Diagram', but if this spelling is\n  used, it should be used consistently.\n\n- 'complex clique' --> Use the term 'clique complex'\n\n- 'module $Z_i(K)$' --> 'modulo $Z_i(K)$' (the paper is discussing the\n  quotient operation)\n\n- 'non-cumulative homology': I think the word 'distances' is missing\n  here.\n\n- The use of 'Means mean' and 'Deviations mean' is slightly confusing;\n  I would suggest to show follow the format $\\mu \\pm \\sigma$, with $\\mu$\n  being the mean and $\\sigma$ being some measure of variance, such as\n  the standard deviation.\n\n- For the references list, I would suggest to carefully check which\n  version of a paper is being cited. Numerous papers mentioned in the\n  bibliography have already been published as book chapters, papers,\n  etc.\n\n- The preprint by Guss and Salakhutdinov is cited twice.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper analyses the training of neural networks from a topological\nperspective, presenting a pipeline that can measure (pseudo) distances\nbetween the network's weights during training. Such information is then\nemployed to study the generalisation error of a neural network.\n\nIn contrast to existing methods for estimating this error, this paper\ndoes *not* require a specific hold-out data set, as topological features\nof the neural network are monitored during training. This frees up\nadditional data for fitting, which can be highly relevant in the sparse\ndata regime.", "main_review": "I enjoyed the ideas presented in this paper; the analysis of\ngeneralisation performance is a highly relevant and timely, and indeed\nconstitutes one of the largest obstacle toward employing machine\nlearning techniques in the wild. Understanding generalisation without\nrequiring additional data has the potential to improve machine learning\ntechniques to a substantial extent.\n\nThat being said, the current paper suffers from some issues, which\nprevent my endorsement at this point:\n\n1. Clarity: while background information on topological data analysis is\n   provided (which I appreciate; in particular in light of the fact that\n   TDA is still a rather novel occurrence at ML conferences), the method\n   itself could be described in more detail. In particular, some aspects\n   should be discussed in Section 4:\n\n    - The use of additional representations for topological features\n      needs to be clarified. At present, the main paper does not contain\n      any description of the challenges in calculating distances or the\n      need for additional representations. At the very least, a brief\n      description of these topological representations is required, so\n      that readers may better understand the remainder of the paper.\n      I find it particularly problematic that no stability guarantees of\n      these representations are discussed; they are not *just* drop-in\n      replacements for the bottleneck or Wasserstein distances.\n\n    - Overall, this section should be rewritten with a clear 'roadmap'\n      in mind. What is the problem you want to tackle and how do you\n      intend to tackle it in this paper?\n\n2. Delineation to existing work: The ICLR paper by Rieck et al. (2019b)\n   appears to already contain a large amount of the material proposed in\n   this paper. A cursory reading shows that Rieck et al. even discuss\n   generalisation performance in an early stopping setting (more about\n   this later). It is therefore critical that the contributions of this\n   paper are more clearly delineated from existing work. From my\n   understanding of both papers, I would say that the current submission\n   improves on the following aspects:\n\n    - Choice of filtration for such neural networks\n\n    - The use of other topological representations (whereas Rieck et al.\n      only use a summary statistic of persistence diagrams).\n\n   These differences (and potentially all others that I missed) should be\n   briefly discussed; the advantage would be that the paper can refer to\n   the previous publication as a justification of the method itself!\n\n   The previous work by Rieck et al. also proposes a set of experiments\n   that would be very fitting to perform here: the experiments proposed\n   in this paper could be substantially strengthened by the addition of certain\n   (simple) monitoring baselines that serve to showcase the added value\n   of employing topological information in the first place here. Rieck et\n   al. compared their topological measure to a 'patience criterion' based\n   on validation loss, thus assessing the difference in performance\n   and epochs after the early stopping criterion was applied). In my\n   opinion, this is most important experiment here when assessing\n   validation performance. Also, it would be interesting to add simple\n   baselines measures to check that these are *insufficient* to be used\n   alone. For instance, a non-topological baseline that could be assessed\n   in this context would be a distribution of weights; it should not be\n   possible to assess validation performance with this baseline.\n\n2. Experimental depth: the experimental section is lacking depth, in\n   particular given the specific goals of this paper. The depicted plots\n   and correlations are useful, but don't serve to highlight the\n   benefits of the proposed method. I would therefore suggest a more\n   thorough experimental setup such as the one shown in the supplement\n   of Rieck et al. (2019b): as mentioned in the previous point, their\n   proposed measure of network complexity is shown as additional early\n   stopping criterion (without an hold-out validation data set as well).\n\n   A post-hoc experiment of this sort would be extremely useful in\n   demonstrating the utility of the method, and it would even enable the\n   quantification of certain measures. The correlation is also useful in\n   this context, but since the goal is to assess the generalisation\n   error, an experiment in which the new measure is directly applied\n   would be extremely worthwhile.\n\nPlease see below for detailed comments.\n\n## Detailed comments\n\n- The first contribution reads a little bit like a sentence fragment to\n  me. I would suggest to use the terms 'silhouette distance' and 'heat\n  kernel distance' instead of plain 'silhouette' and 'heat'.\n\n- The terminology 'homological convergence' is slightly misleading\n  because the paper studies persistent homology, as far as I understand.\n\n- The introduction of topological concepts in the background section\n  jumps between the 'geometrical' and the 'abstract'; I understand that\n  both perspectives are valid, but I would suggest to choose only\n  a single one.\n\n- The discussion of chains and boundaries requires more explanations (or\n  could be partially relegated to the supplementary materials). For\n  instance, please clarify the 'signed combination.' However,\n  definitions need to appear in the main text, not only in the\n  supplementary materials.\n\n- In figure 4, the $3$-simplex should consist of four $2$-simplices, but\n  only two are shown. Either clarify this in the captions or update the\n  figure accordingly.\n\n- The introduction of filtrations is not motivated; the 'nested family\n  of simplicial complexes' could be introduced by building more\n  intuition here.\n\n- The related work discussion should, as mentioned above, delineate this\n  paper from Rieck et al. (2019b).\n\n- The heading 'Algebraic Topology object' is rather confusing to me;\n  I would suggest to rewrite it as 'directed flag complex', for\n  instance.\n\n- As outlined above, the distance calculation lacks information about\n  stability and other guarantees. There is also a broken reference on p.\n  5, which needs to be rectified.\n\n- In the experimental section, the use of the MLP should be explained in\n  more detail. What does it do and how does it work exactly?\n\n- I do not understand the 'input order' experiment; does it refer to\n  ordering of samples (i.e. on the batch level)?\n\n- The use of additional topological descriptors makes the use of the\n  word 'distance' imprecise: the bottleneck and Wasserstein\n  distances are metrics in the mathematical sense; while there *is*\n  a well-defined metric between, for instance, persistence landscapes or\n  silhouettes, this is only (to a certain extent) an approximation to\n  the bottleneck/Wasserstein metrics. This distinction should be made\n  clearer in the paper.\n\n- Figure 6 should add standard deviations of the curves. An additional\n  clarification of the content of this figure would also be appreciated\n  by readers; I found the discussion to be slightly confusing because\n  I did not get the relevance of the cumulative distance. The statement\n  'the evolution of the homological convergence [...] seems to be very\n  similar to the one of the validation score' also needs some\n  clarification. Except for Figure 6b, where I observe some oscillation\n  behaviour and maybe the tendency to converge, I don't observe\n  convergence anywhere else.\n\n- The same applies to Figures 7–9. Adding some more details here will\n  result in a clearer description of the method. Maybe some of the\n  curves could also be relegated to the supplementary materials?\n\n- I would also suggest to investigate the use of energy distances or\n  energy correlations, as these measures are more flexible and capable\n  of assessing more than just linear dependencies between two\n  variables. See [*Energy statistics: A class of statistics based on distances*](https://www.sciencedirect.com/science/article/abs/pii/S0378375813000633)\n  for more details.\n\n## Terminology and style\n\nFor the most part, the paper is written well. I have some suggestions\nconcerning the style:\n\n- Please use `\\citep` and `\\citet` consistently when using `natbib`. At\n  present, citations are directly appearing in the text without any\n  enclosing parentheses. For instance, in the related work section,\n  the sentence 'for improving the training procedure of the models Hofer\n  et al. (2020); [...]' should use `\\citep` in order to obtain a proper\n  parenthetical citation.\n\n- can't --> cannot (likewise for other contractions; I admit that this\n  is a personal preference, so feel free to ignore it)\n\n- 'contained simplicial complex' --> 'nested simplicial complexes'\n\n- The paper employs non-standard terminology in certain places. The\n  diagram, for instance, is called 'persistence diagram', no\n  'persistence homology diagram'. Please ensure consistency with\n  existing papers here. The same applies to capitalisation; I personally\n  see no need to write 'Persistence Diagram', but if this spelling is\n  used, it should be used consistently.\n\n- 'complex clique' --> Use the term 'clique complex'\n\n- 'module $Z_i(K)$' --> 'modulo $Z_i(K)$' (the paper is discussing the\n  quotient operation)\n\n- 'non-cumulative homology': I think the word 'distances' is missing\n  here.\n\n- The use of 'Means mean' and 'Deviations mean' is slightly confusing;\n  I would suggest to show follow the format $\\mu \\pm \\sigma$, with $\\mu$\n  being the mean and $\\sigma$ being some measure of variance, such as\n  the standard deviation.\n\n- For the references list, I would suggest to carefully check which\n  version of a paper is being cited. Numerous papers mentioned in the\n  bibliography have already been published as book chapters, papers,\n  etc.\n\n- The preprint by Guss and Salakhutdinov is cited twice.\n\n", "summary_of_the_review": "While I am very excited about topology-based approaches that aim to\nunderstand the training or testing behaviour of neural networks,\nI cannot endorse this paper for publication yet.\n\nThe current write-up is suffering from several issues, which need to be\nrectified in a **major revision** before reaching the quality standards\nof ICLR. These issues include:\n\n1. Lack of clarity: an improved introduction to topological concepts is\n   required and the proposed method needs to be compared more with\n   existing topology-based approaches (in particular since it does not\n   meet the requirements of a metric in the mathematical sense, an\n   analysis of approximation guarantees is crucial).\n\n2. Lack of experimental depth and delineation to existing work: since\n   the express goal of this paper is to analyse generalisation\n   properties of neural networks based on their topological properties,\n   a comparison with existing work (Rieck et al., 2019b) and an improved\n   experimental suite (containing previous work *and* non-topological\n   baselines) is critical for corroborating the claims of this paper.\n\n**Updated after rebuttal**: A lot has been discussed during this rebuttal period. I would strongly\nrecommend to pick up some of the suggestions of reviewers in order to improve the paper. A recent paper\nby [Birdal et al.](https://papers.nips.cc/paper/2021/hash/35a12c43227f217207d4e06ffefe39d3-Abstract.html)\ndemonstrates how to successfully assess generalisation performance (disclaimer: I am *not* one of the authors\nof said paper). I hope that this may serve as a partial inspiration.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["Yes, Unprofessional behaviors (e.g., unprofessional exchange between authors and reviewers)"], "details_of_ethics_concerns": "Authors were bordering on passive--aggressive and behaved in an abrasive manner with all reviewers. Instead of updating/revising the paper, much was spent on nit-picking irrelevant details. Overall, I have the feeling that the hours I put into reviewing this paper and engaging with the authors were rather wasted. I wish the authors would take some of the comments to heart—my summary with a link to a recent NeurIPS paper demonstrates how a successful experimental setup could look like; I honestly don't understand why the authors appear to be so adamant and hostile in these exchanges.", "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635756446793}, {"id": "IEV3mE6Z9vZ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4090/Reviewer_JhC7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper experimentally investigates the (cor)relation between the validation accuracy achieved by a neural network and a variation of a topological descriptor built on top of the network: the persistence diagram of a directed flag complex built on top of the network. \n\nThe main claim is that there is a high correlation between the validation accuracy and this \"topological distance\", so that the later may be used when a validation set is not available. ", "review_text": "# Strengths and Weaknesses\n\n## Strengths: \n- The paper is well written overall (aside for few typos, see Minor comments).\n- Experimental report is extensive given the amount of complementary plots in the supplementary material. \n- The methodology proposed by the paper is original and interesting. The construction of topological descriptor for networks has interesting properties, theoretically improving for instance on the approach proposed by Rieck et al. (at the price of computational efficiency though). \n\n## Weaknesses :\n\n- My main concern with the paper is that its conclusions seem too optimistic. From my understanding, saying that the \"*proposed measure strongly correlate with validation accuracy*\" (Section 7) and that \"*the generalization error of a neural network could be intrinsically estimated without any holdout set*\" (abstract) would mean for it to be useful that : \n  - (a) If the network starts overfitting, that is the validation accuracy starts decreasing (while the training accuracy presumably keeps increasing), this should be reflected in the curve of (cumulative) distances between (vectorization of) PDs. On Figure 7, b, for instance, there is---as far as I understand---no way to infer from the line-curve (PD distances) that the network is overfitting. \n  - (b) To be able to predict the generalization gap, one would expect a change of behavior when the network properly learns (e.g. MNIST) vs when it does not (e.g. CIFAR-10-MLP) ; (here, I assume that the training accuracy is going high---it as not been reported from what I can tell so we cannot be sure what the generalization gap is. Hence I assume that low validation accuracy $\\Leftrightarrow$ high gap.). From Figure 8.a and 8.b for instance, how is one supposed to infer that one network is learning something useful while the other does not? \n\nI am worried that what is interpreted as a strong correlation may only be a consequence of the networks parameters optimization/convergence. During optimization, as the step size goes to zero using RMSprop, it is not unlikely that the network weights variations (hence successive diagram distances) are getting smaller, yielding this concave-shape for the plots of cumulative diagram distances and for the *training* accuracy---thus for the validation accuracy when the network learns properly (i.e. validation accuracy $\\simeq$ training accuracy). \n\nFurthermore, when looking at plots in the supplementary material (for instance Figure 53), it seems that the reasonably correct correlation coefficients obtained are in great part due to the correlation *before* network parameters convergence (bottom-left quadrant: low topological cumulative, low validation score), which are not very interesting. On the top-right quadrant (when the network is likely to be \"ready\"), things get messier. \n\nIn the same vein, the sentence \"*(...) validation accuracy depends on the specificity of the data sampled in the validation subset, while the homological convergence is independent of the validation data*\" (end of section 6) is somewhat contradictory with the claim of the paper that homological properties of the network \"*captures the generalization of neural networks without a validation set*\". \n\nAll in all, and unless I have missed something crucial, I feel that the empirical results (which are quite numerous, thanks!) provided by this work are, unfortunately, not convincing.\n\n- A second important concern is the intersection with the Anonymous concurrent work (available in the supplementary material). With few exceptions, everything is the same up to the end of section 5 : same construction of topological descriptors (flag complex, filtration by Eq (1)), same diagram vectorizations, same datasets, same changes of architecture (number of layers, layer widths)... So that, roughly, what differs in the two paper is the experimental measurement: this paper studies what happen at training time for a given network, while the concurrent one makes cross-networks comparisons. Though the two papers cover different objectives, it diminishes the intrinsic contributions of the current one, which are now limited to the experimental results which are, as said above, insufficiently convincing in my opinion. \n- As acknowledged by the authors, this method does not scale and is only applicable to reasonable MLP; even though requiring high-end hardware and a lot of time. \n- The paper is not supported by any theoretical result that would motivate the use of their approach.\n\n\n# Minor comments\n- (Typo) p2, give-->given\n- (Typo) End of Section 3 : \"They claim. However\". It seems a sentence is missing. \n- (Typo) Start of p5 : an smoothing --> a smoothing\n- (Suggestion) Eq 1 : use \\max and \\left( \\right). \n- (Suggestion) Numbering cases with \"1. blabla... 2. blabla\" hinders readability. Perhaps using something in the vein of \"*(i)* blabla, *(ii)* blabla\" would be better. \n- (Clarity) The transition between filtrations of simplicial complex and persistence diagrams may be a bit too short (I am referring to the single sentence \"The sequence of homology groups is calculated by varying the parameter $\\epsilon$ to obtain the persistence homology diagram.\")\n- (Typo) p5 : a reference is missing (\"as in ?\").\n- end of p7, \"the homology does not seem to converge\". I do not understand this claim. If the network (parameters) converge, so does the persistence diagrams. If my understanding is correct, RMSprop should always yield convergence (it uses a decay factor for the step size); so I guess more epochs are needed. \n- p9 : \"our approach computes the exact PDs distance\", I think the exact PDs are computed in first place (not the distances). \n- Results in Table 1 for Heat and Silhouette are the exact same. Is it an unfortunate copy-paste?\n- Supplementary material, Figures 42-53 : how can it be that the topological cumulative decreases (given that the validation score seems monotonic, I guess that the axes labels have been switched?)? Also, why does the validation score reaches 1, while it is supposed to be much lower on some model? (I guess a normalization has been applied). \n- The paper claims that \"*in CNNs, the correlation are (...) still usually above 0.8*\", but unless I misread the tables 5,6, 9, and 10, this seems to be exaggerated. CIFAR100CNN + number of layers $\\geq 4$ is one the only instance where this seems to hold. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper experimentally investigates the (cor)relation between the validation accuracy achieved by a neural network and a variation of a topological descriptor built on top of the network: the persistence diagram of a directed flag complex built on top of the network. \n\nThe main claim is that there is a high correlation between the validation accuracy and this \"topological distance\", so that the later may be used when a validation set is not available. ", "main_review": "# Strengths and Weaknesses\n\n## Strengths: \n- The paper is well written overall (aside for few typos, see Minor comments).\n- Experimental report is extensive given the amount of complementary plots in the supplementary material. \n- The methodology proposed by the paper is original and interesting. The construction of topological descriptor for networks has interesting properties, theoretically improving for instance on the approach proposed by Rieck et al. (at the price of computational efficiency though). \n\n## Weaknesses :\n\n- My main concern with the paper is that its conclusions seem too optimistic. From my understanding, saying that the \"*proposed measure strongly correlate with validation accuracy*\" (Section 7) and that \"*the generalization error of a neural network could be intrinsically estimated without any holdout set*\" (abstract) would mean for it to be useful that : \n  - (a) If the network starts overfitting, that is the validation accuracy starts decreasing (while the training accuracy presumably keeps increasing), this should be reflected in the curve of (cumulative) distances between (vectorization of) PDs. On Figure 7, b, for instance, there is---as far as I understand---no way to infer from the line-curve (PD distances) that the network is overfitting. \n  - (b) To be able to predict the generalization gap, one would expect a change of behavior when the network properly learns (e.g. MNIST) vs when it does not (e.g. CIFAR-10-MLP) ; (here, I assume that the training accuracy is going high---it as not been reported from what I can tell so we cannot be sure what the generalization gap is. Hence I assume that low validation accuracy $\\Leftrightarrow$ high gap.). From Figure 8.a and 8.b for instance, how is one supposed to infer that one network is learning something useful while the other does not? \n\nI am worried that what is interpreted as a strong correlation may only be a consequence of the networks parameters optimization/convergence. During optimization, as the step size goes to zero using RMSprop, it is not unlikely that the network weights variations (hence successive diagram distances) are getting smaller, yielding this concave-shape for the plots of cumulative diagram distances and for the *training* accuracy---thus for the validation accuracy when the network learns properly (i.e. validation accuracy $\\simeq$ training accuracy). \n\nFurthermore, when looking at plots in the supplementary material (for instance Figure 53), it seems that the reasonably correct correlation coefficients obtained are in great part due to the correlation *before* network parameters convergence (bottom-left quadrant: low topological cumulative, low validation score), which are not very interesting. On the top-right quadrant (when the network is likely to be \"ready\"), things get messier. \n\nIn the same vein, the sentence \"*(...) validation accuracy depends on the specificity of the data sampled in the validation subset, while the homological convergence is independent of the validation data*\" (end of section 6) is somewhat contradictory with the claim of the paper that homological properties of the network \"*captures the generalization of neural networks without a validation set*\". \n\nAll in all, and unless I have missed something crucial, I feel that the empirical results (which are quite numerous, thanks!) provided by this work are, unfortunately, not convincing.\n\n- A second important concern is the intersection with the Anonymous concurrent work (available in the supplementary material). With few exceptions, everything is the same up to the end of section 5 : same construction of topological descriptors (flag complex, filtration by Eq (1)), same diagram vectorizations, same datasets, same changes of architecture (number of layers, layer widths)... So that, roughly, what differs in the two paper is the experimental measurement: this paper studies what happen at training time for a given network, while the concurrent one makes cross-networks comparisons. Though the two papers cover different objectives, it diminishes the intrinsic contributions of the current one, which are now limited to the experimental results which are, as said above, insufficiently convincing in my opinion. \n- As acknowledged by the authors, this method does not scale and is only applicable to reasonable MLP; even though requiring high-end hardware and a lot of time. \n- The paper is not supported by any theoretical result that would motivate the use of their approach.\n\n\n# Minor comments\n- (Typo) p2, give-->given\n- (Typo) End of Section 3 : \"They claim. However\". It seems a sentence is missing. \n- (Typo) Start of p5 : an smoothing --> a smoothing\n- (Suggestion) Eq 1 : use \\max and \\left( \\right). \n- (Suggestion) Numbering cases with \"1. blabla... 2. blabla\" hinders readability. Perhaps using something in the vein of \"*(i)* blabla, *(ii)* blabla\" would be better. \n- (Clarity) The transition between filtrations of simplicial complex and persistence diagrams may be a bit too short (I am referring to the single sentence \"The sequence of homology groups is calculated by varying the parameter $\\epsilon$ to obtain the persistence homology diagram.\")\n- (Typo) p5 : a reference is missing (\"as in ?\").\n- end of p7, \"the homology does not seem to converge\". I do not understand this claim. If the network (parameters) converge, so does the persistence diagrams. If my understanding is correct, RMSprop should always yield convergence (it uses a decay factor for the step size); so I guess more epochs are needed. \n- p9 : \"our approach computes the exact PDs distance\", I think the exact PDs are computed in first place (not the distances). \n- Results in Table 1 for Heat and Silhouette are the exact same. Is it an unfortunate copy-paste?\n- Supplementary material, Figures 42-53 : how can it be that the topological cumulative decreases (given that the validation score seems monotonic, I guess that the axes labels have been switched?)? Also, why does the validation score reaches 1, while it is supposed to be much lower on some model? (I guess a normalization has been applied). \n- The paper claims that \"*in CNNs, the correlation are (...) still usually above 0.8*\", but unless I misread the tables 5,6, 9, and 10, this seems to be exaggerated. CIFAR100CNN + number of layers $\\geq 4$ is one the only instance where this seems to hold. ", "summary_of_the_review": "Though I appreciate the ambitions of the work, I feel that the empirical results are, unfortunately, not convincing. The absence of theoretical results and the very strong similarities with the concurrent submitted paper diminishes it's impact as well. ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635426213742}], "openreview_url": "https://openreview.net/forum?id=TNxKD3z_tPZ", "arxiv_id": "2106.00012", "paper_pdf": "papers/TNxKD3z_tPZ.pdf", "paper_pdf_sha256": "6b0dd30151bd111d9d1e0521ea91732bf34d60f06b81274516ee66db104119e3", "paper_pdf_bytes": 866979, "paper_pdf_source": "openreview", "code_url": "https://github.com/asier-gutierrez/nn-evolution", "code_repository": "asier-gutierrez/nn-evolution", "code_commit": "311211821b0ebffd2458061e9636bd6dd68e7609", "code_archive": "repos/TNxKD3z_tPZ.zip", "code_archive_sha256": "f19c3432fe4dafa338be2973e2b322a5fddd35e9d763096c873b99544440ae5a", "code_archive_bytes": 62312, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 60, "github_languages": {"Python": 156507}, "github_archived": false, "github_pushed_at": "2021-07-17T08:21:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/persistent-homology-captures-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_TGlfdZOHY3", "year": 2021, "status": "rejected", "title": "On Episodes, Prototypical Networks, and Few-Shot Learning", "authors": ["Steinar Laenen", "Luca Bertinetto"], "authorids": ["~Steinar_Laenen1", "~Luca_Bertinetto1"], "authors_source": "OpenReview API", "abstract": "Episodic learning is a popular practice among researchers and practitioners interested in few-shot learning. It consists of organising training in a series of learning problems, each relying on small “support” and “query” sets to mimic the few-shot circumstances encountered during evaluation.\nIn this paper, we investigate the usefulness of episodic learning in Prototypical Networks, one of the most popular algorithms making use of this practice.\nSurprisingly, in our experiments we found that, for Prototypical Networks, it is detrimental to use the episodic learning strategy of separating training samples between support and query set, as it is\na data-inefficient way to exploit training batches. This “non-episodic” version of Prototypical Networks, which corresponds to the classic Neighbourhood Component Analysis, reliably improves over its episodic counterpart in multiple datasets, achieving an accuracy that is competitive with the state-of-the-art, despite being extremely simple.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "FgXSA1gvj_B", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1898/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper's starting point is the question whether the episodic training is beneficial, or not, for FSL / Prototypical Networks. The work can be seen as a follow-up of the recent works showing that simple baselines can outperform rather sophisticated few-shot learning models. Towards answering this question, this paper points out that Prototypical Networks (PN) are related to Neighborhood Component Analysis (NCA), and NCA can be considered as an episodic training-free alternative of PN.\n\nIn more detail, PN aims to learn per-class prototypes based on sample averaging in the feature space. NCA, in contrast, aims to maximize the ratio of total similarity between same-class example pairs to the total similarity between different-class pairs. Due to their similarities in terms of their formulations, the paper claims that NCA loss can be considered as an alternative to PN loss to do non-episodic representation learning for few-shot learning purposes. In addition, the paper has a few strong claims, such as episodic training is “detrimental to learning” and “under no circumstance beneficial to differentiate between support and query set within a training batch\". Clearly, these are intriguing claims.\n \nHowever, there is a gap between the claims and the experimental validation. First, even if ProtoNet loss and NCA loss seem to be similar to each other, they're nevertheless different models, and it takes quite a significant manipulation to convert PN to NCA. Therefore, the fact one particular non-episodic-training based model gives superior results compared to those of PN with episodic-training, does tell us much about detrimental effects of episodic training for PN or in general. Second, while the paper's observations that NCA has the advantage of using more pairwise similarities within a batch compared to PN is indeed insightful, it rather points out to certain weaknesses in the way per-batch / per-episode data is being utilized by the PN formulation, instead of problems about episodic training.\n\nOverall, the paper has interesting observations about PN's weaknesses and shows why one particular simple non-episodic training / non meta-learned approach (NCA) can yield superior results compared to PN, which is a relatively mode sophisticated & well-established approach. However, the paper's (over-strong) claims remain mostly unsupported, which makes the otherwise interesting work poorly framed. The paper, with more water-tight arguments only, could otherwise be a valuable contribution but it requires quite significant & fundamental revisions throughout the paper, therefore, is not ready for publication in its current form.\n\nPost-rebuttal: I would like to thank for the detailed responses and the careful revisions made in the paper. Overall, the paper is now definitely improved in certain ways and initiates an important discussion on the value of episodic training & classifier synthesis for few-shot learning, as opposed to typically-simpler metric-learning based approaches. The paper also approaches this problem from an interesting point of view, by focusing on sample utilization in the episodic training of PN.  \n\nHowever, I still find that the the paper remains somewhat weak in its current form for the following reasons:\n- I maintain my view that NCA vs PN are not direct alternatives to each other, considering that PN allows learning a representation that is optimized for class-average to sample comparisons, whereas, NCA uses a sample-to-sample distance based loss. The fact that the very construction of these two models, despite the similarities pointed out, blurs the strength of the overall NCA vs PN based discussion on the value of episodic training.\n- The claims about the advantages of non-episodic training of NCA is mainly based on the observation that NCA creates more positive/negative labels. However, it is not clear whether it is the more efficient utilization of training samples or just the differences in terms of the predictive model formulations. (Perhaps, averaging based prototype computation is a bad idea, after all, which may not have directly anything to do with episodic training.)\n- To this end, Fig. 3 is indeed interesting, but again the results are not very clear. Here, careful optimizer re-tuning specially for each case can be necessary as subsampling degrades the gradient approximation quality, which creates the question how much fundamentally important efficiency in batch utilization is, as long as one uses a proper optimizer. \nOverall, I think the paper makes a valuable step in an interesting direction but the paper fails to make a strong-enough case. Overall, I  improve my rating by a single level to 4, but find that the paper is not stronger than this in its current form.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work, unsupported claims", "review": "The paper's starting point is the question whether the episodic training is beneficial, or not, for FSL / Prototypical Networks. The work can be seen as a follow-up of the recent works showing that simple baselines can outperform rather sophisticated few-shot learning models. Towards answering this question, this paper points out that Prototypical Networks (PN) are related to Neighborhood Component Analysis (NCA), and NCA can be considered as an episodic training-free alternative of PN.\n\nIn more detail, PN aims to learn per-class prototypes based on sample averaging in the feature space. NCA, in contrast, aims to maximize the ratio of total similarity between same-class example pairs to the total similarity between different-class pairs. Due to their similarities in terms of their formulations, the paper claims that NCA loss can be considered as an alternative to PN loss to do non-episodic representation learning for few-shot learning purposes. In addition, the paper has a few strong claims, such as episodic training is “detrimental to learning” and “under no circumstance beneficial to differentiate between support and query set within a training batch\". Clearly, these are intriguing claims.\n \nHowever, there is a gap between the claims and the experimental validation. First, even if ProtoNet loss and NCA loss seem to be similar to each other, they're nevertheless different models, and it takes quite a significant manipulation to convert PN to NCA. Therefore, the fact one particular non-episodic-training based model gives superior results compared to those of PN with episodic-training, does tell us much about detrimental effects of episodic training for PN or in general. Second, while the paper's observations that NCA has the advantage of using more pairwise similarities within a batch compared to PN is indeed insightful, it rather points out to certain weaknesses in the way per-batch / per-episode data is being utilized by the PN formulation, instead of problems about episodic training.\n\nOverall, the paper has interesting observations about PN's weaknesses and shows why one particular simple non-episodic training / non meta-learned approach (NCA) can yield superior results compared to PN, which is a relatively mode sophisticated & well-established approach. However, the paper's (over-strong) claims remain mostly unsupported, which makes the otherwise interesting work poorly framed. The paper, with more water-tight arguments only, could otherwise be a valuable contribution but it requires quite significant & fundamental revisions throughout the paper, therefore, is not ready for publication in its current form.\n\nPost-rebuttal: I would like to thank for the detailed responses and the careful revisions made in the paper. Overall, the paper is now definitely improved in certain ways and initiates an important discussion on the value of episodic training & classifier synthesis for few-shot learning, as opposed to typically-simpler metric-learning based approaches. The paper also approaches this problem from an interesting point of view, by focusing on sample utilization in the episodic training of PN.  \n\nHowever, I still find that the the paper remains somewhat weak in its current form for the following reasons:\n- I maintain my view that NCA vs PN are not direct alternatives to each other, considering that PN allows learning a representation that is optimized for class-average to sample comparisons, whereas, NCA uses a sample-to-sample distance based loss. The fact that the very construction of these two models, despite the similarities pointed out, blurs the strength of the overall NCA vs PN based discussion on the value of episodic training.\n- The claims about the advantages of non-episodic training of NCA is mainly based on the observation that NCA creates more positive/negative labels. However, it is not clear whether it is the more efficient utilization of training samples or just the differences in terms of the predictive model formulations. (Perhaps, averaging based prototype computation is a bad idea, after all, which may not have directly anything to do with episodic training.)\n- To this end, Fig. 3 is indeed interesting, but again the results are not very clear. Here, careful optimizer re-tuning specially for each case can be necessary as subsampling degrades the gradient approximation quality, which creates the question how much fundamentally important efficiency in batch utilization is, as long as one uses a proper optimizer. \nOverall, I think the paper makes a valuable step in an interesting direction but the paper fails to make a strong-enough case. Overall, I  improve my rating by a single level to 4, but find that the paper is not stronger than this in its current form.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603912110403}, {"id": "ugIVEZg1Cpa", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1898/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Summary\n\nThe submission investigates the properties of episodic training and its impact on learning using Prototypical Networks as a case study. The paper draws a connection between Prototypical Networks and Neighbourhood Component Analysis (NCA), noting that their loss functions are similar but that NCA is trained non-episodically, which allows it to learn from the relationship between all example pairs in a batch.\n\nWhen controlling for batch size, the paper claims to show that NCA (combined with a nearest-centroid inference strategy) performs better than Prototypical Networks, as evidenced by experiments on CIFAR-FS and mini-ImageNet. Ablation experiments are performed, claiming to show that applying the NCA loss to batches sampled episodically allows Prototypical Networks to bridge the performance gap with NCA, and that the partition of examples within a batch into support and query sets is detrimental to Prototypical Networks training. Finally, NCA is evaluated alongside comparable competing approaches on mini-ImageNet, CIFAR-FS, and tiered-ImageNet, and is claimed to yield results comparable or superior to the state-of-the-art.\n\n#### Strengths and weaknesses\n\n* **+** The value of episodic training is increasingly being questioned, and the submission approaches the topic from a new and interesting perspective.\n* **+** The connection between nearest-centroid few-shot learning approaches and NCA has not been made in the literature to my knowledge and has potential applications beyond the scope of this paper.\n* **+** The paper is well-written, easy to follow, and well-connected to the existing literature.\n* **-** The extent to which the observations presented generalize to few-shot learners beyond Prototypical Networks is not evaluated, which may limit the scope of the submission’s contributions in terms of understanding the properties of episodic training.\n* **-** The Matching Networks / NCA connection makes more sense in my opinion than the Prototypical Networks / NCA connection.\n* **-** A single set of hyperparameters was used across learners for a given benchmark, which can bias the conclusions drawn from the experiments.\n\n#### Recommendation\n\nI’m leaning towards acceptance. I have some issues with the submission that are detailed below, but overall the paper presents an interesting take on a topic that’s currently very relevant to the few-shot learning community, and I feel that the value it brings to the conversation is sufficient to overcome the concerns I have.\n\n#### Detailed justification\n\nThe biggest concern I have with the submission is methodological. One one hand, the authors went beyond the usual practice of reporting accuracies on a single run and instead trained each method with five different random initializations, and this is a practice that I’m happy to see in a few-shot classification paper. On the other hand, the choice to share a single set of hyperparameters across learners for a given benchmark leaves a blind spot in the evaluation. What if Prototypical Networks are more sensitive to the choice of optimizer, learning rate schedule, and weight decay coefficient than NCA? Is it possible that the set of hyperparameters chosen for the experiments happens to work poorly for Prototypical Networks? Would we observe the same trends if we tuned hyperparameters independently for each experimental setting? In its current form the submission shows that Prototypical Networks are sensitive to the hyperparameters used to sample episodes *while keeping other hyperparameters fixed*, but showing the same trend while doing a reasonable effort at tuning other hyperparameters would make for a more convincing argument. This is why I take the claim made in Section 4.2 that \"NCA performs better than all PN configurations, no matter the batch size\" with a grain of salt, for instance.\n\nI also feel that the submission misses out on an opportunity to support a more general statement about episodic training via observations on approaches such as Matching Networks, MAML, etc. I really like the way Figure 1 explains visually how Prototypical Networks miss out on useful relationships between examples in a batch and is therefore data-inefficient. To me, this is one of the submission’s most important contributions: the suggestion that a leave-one-out strategy could allow episodic approaches to achieve the same kind of data efficiency as non-episodic approaches, alleviating the need for a supervised pre-training / episodic fine-tuning strategy. To be clear, I don’t think the missed opportunity would be a reason to reject the paper, but I think that showing empirically that the leave-one-out strategy applies beyond Prototypical Networks would make me lean more strongly towards acceptance.\n\nThe connection drawn between Prototypical Networks and NCA feels forced at times. In the introduction the paper claims to \"show that, without episodic learning, Prototypical Networks correspond to the classic Neighbourhood Component Analysis\", but Section 3.3 lists the creation of prototypes as a key difference between the two which is not resolved by training non-episodically. From my perspective, NCA would be more akin to the non-episodic counterpart to Matching Networks without Full Contextual Embeddings – albeit with a Euclidean metric rather than a cosine similarity metric – since both perform comparisons on example pairs.\n\nThis relationship with Matching Networks could be exploited to improve clarity. For instance, row 6 of Figure 4 can be interpreted as a Matching Networks implementation with a Euclidean distance metric. With this in mind, could the difference in performance between \"*1*-NN with class centroids\" and *k*-NN / Soft Assignment noted in Section 4.1 – as well as the drop in performance observed in Figure 4’s row 6 – be explained by the fact that a (soft) nearest-neighbour approach is more sensitive to outliers?\n\nFinally, I have some issues with how results are reported in Tables 1 and 2. Firstly, we don’t know how competing approaches would perform if we applied the paper’s proposed multi-layer concatenation trick, and the idea itself feels more like a way to give NCA’s performance a small boost and bring it into SOTA-like territory. Comparing NCA without multi-layer against other approaches is therefore more interesting to me. Secondly, 95% confidence intervals are provided, but the absence of identification of the best-performing approach(es) in each setting makes it hard to draw high-level conclusions at a glance. I would suggest bolding the best accuracy in each column along with all other entries for which a 95% confidence interval test on the difference between the means is inconclusive in determining that the difference is significant.\n\n#### Questions\n\n1. In Equation 2, why is the sum normalized by the total number of examples in the episode rather than the number of query examples?\n1. Can the authors comment on the extent to which Figure 2 supports the hypothesis that NCA is better for training because it learns from a larger number of positives and negatives? Assuming this is true, we should see that Prototypical Networks configurations that increase the number of positives and negatives should perform better for a given batch size. Does Figure 2 support this assertion?\n1. Can the authors elaborate on the \"no S/Q\" ablation (Figure 4, row 7)? What is the point of reference when computing distances for support and query examples? Is the loss computed in the same way for support and query examples? The text in Section 4.3 makes it appear like the loss for query examples is the NCA loss, but the loss for support examples is the prototypical loss. Wouldn’t it be conceptually cleaner to compute leave-one-out prototypes, i.e. leave each example out of the computation of its own class’ prototype (resulting in slightly different prototypes for examples of the same class)? In my mind, this would be the best way to remove the support/query partition while maintaining prototype computation, thereby showing that the partition is detrimental to Prototypical Networks training.\n\n#### Additional feedback\n\n1. This is somewhat inconsequential, but across all implementations of episodic training that I have examined I haven’t encountered an implementation that uses a flag to differentiate between support and query examples. Instead, the implementations I have examined explicitly represent support and query examples as separate tensors. I was therefore surprised to read that \"in most implementations [...] each image is characterised by a flag indicating whether it corresponds to the support or the query set [...]\"; can the authors point to the implementations they have in mind when making that assertion?\n1. I would be careful with the assertion that \"during evaluation the triplet {w, n, m} [...] must stay unchanged across methods\". While this is true for the benchmarks considered in this submission, benchmarks like Meta-Dataset evaluate on variable-ways and variable-shots episodes.\n1. I’m not too concerned with the computational efficiency of NCA. The pairwise Euclidean distances can be computed efficiently using the inner- and outer-product of the batch of embeddings with itself.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "#### Summary\n\nThe submission investigates the properties of episodic training and its impact on learning using Prototypical Networks as a case study. The paper draws a connection between Prototypical Networks and Neighbourhood Component Analysis (NCA), noting that their loss functions are similar but that NCA is trained non-episodically, which allows it to learn from the relationship between all example pairs in a batch.\n\nWhen controlling for batch size, the paper claims to show that NCA (combined with a nearest-centroid inference strategy) performs better than Prototypical Networks, as evidenced by experiments on CIFAR-FS and mini-ImageNet. Ablation experiments are performed, claiming to show that applying the NCA loss to batches sampled episodically allows Prototypical Networks to bridge the performance gap with NCA, and that the partition of examples within a batch into support and query sets is detrimental to Prototypical Networks training. Finally, NCA is evaluated alongside comparable competing approaches on mini-ImageNet, CIFAR-FS, and tiered-ImageNet, and is claimed to yield results comparable or superior to the state-of-the-art.\n\n#### Strengths and weaknesses\n\n* **+** The value of episodic training is increasingly being questioned, and the submission approaches the topic from a new and interesting perspective.\n* **+** The connection between nearest-centroid few-shot learning approaches and NCA has not been made in the literature to my knowledge and has potential applications beyond the scope of this paper.\n* **+** The paper is well-written, easy to follow, and well-connected to the existing literature.\n* **-** The extent to which the observations presented generalize to few-shot learners beyond Prototypical Networks is not evaluated, which may limit the scope of the submission’s contributions in terms of understanding the properties of episodic training.\n* **-** The Matching Networks / NCA connection makes more sense in my opinion than the Prototypical Networks / NCA connection.\n* **-** A single set of hyperparameters was used across learners for a given benchmark, which can bias the conclusions drawn from the experiments.\n\n#### Recommendation\n\nI’m leaning towards acceptance. I have some issues with the submission that are detailed below, but overall the paper presents an interesting take on a topic that’s currently very relevant to the few-shot learning community, and I feel that the value it brings to the conversation is sufficient to overcome the concerns I have.\n\n#### Detailed justification\n\nThe biggest concern I have with the submission is methodological. One one hand, the authors went beyond the usual practice of reporting accuracies on a single run and instead trained each method with five different random initializations, and this is a practice that I’m happy to see in a few-shot classification paper. On the other hand, the choice to share a single set of hyperparameters across learners for a given benchmark leaves a blind spot in the evaluation. What if Prototypical Networks are more sensitive to the choice of optimizer, learning rate schedule, and weight decay coefficient than NCA? Is it possible that the set of hyperparameters chosen for the experiments happens to work poorly for Prototypical Networks? Would we observe the same trends if we tuned hyperparameters independently for each experimental setting? In its current form the submission shows that Prototypical Networks are sensitive to the hyperparameters used to sample episodes *while keeping other hyperparameters fixed*, but showing the same trend while doing a reasonable effort at tuning other hyperparameters would make for a more convincing argument. This is why I take the claim made in Section 4.2 that \"NCA performs better than all PN configurations, no matter the batch size\" with a grain of salt, for instance.\n\nI also feel that the submission misses out on an opportunity to support a more general statement about episodic training via observations on approaches such as Matching Networks, MAML, etc. I really like the way Figure 1 explains visually how Prototypical Networks miss out on useful relationships between examples in a batch and is therefore data-inefficient. To me, this is one of the submission’s most important contributions: the suggestion that a leave-one-out strategy could allow episodic approaches to achieve the same kind of data efficiency as non-episodic approaches, alleviating the need for a supervised pre-training / episodic fine-tuning strategy. To be clear, I don’t think the missed opportunity would be a reason to reject the paper, but I think that showing empirically that the leave-one-out strategy applies beyond Prototypical Networks would make me lean more strongly towards acceptance.\n\nThe connection drawn between Prototypical Networks and NCA feels forced at times. In the introduction the paper claims to \"show that, without episodic learning, Prototypical Networks correspond to the classic Neighbourhood Component Analysis\", but Section 3.3 lists the creation of prototypes as a key difference between the two which is not resolved by training non-episodically. From my perspective, NCA would be more akin to the non-episodic counterpart to Matching Networks without Full Contextual Embeddings – albeit with a Euclidean metric rather than a cosine similarity metric – since both perform comparisons on example pairs.\n\nThis relationship with Matching Networks could be exploited to improve clarity. For instance, row 6 of Figure 4 can be interpreted as a Matching Networks implementation with a Euclidean distance metric. With this in mind, could the difference in performance between \"*1*-NN with class centroids\" and *k*-NN / Soft Assignment noted in Section 4.1 – as well as the drop in performance observed in Figure 4’s row 6 – be explained by the fact that a (soft) nearest-neighbour approach is more sensitive to outliers?\n\nFinally, I have some issues with how results are reported in Tables 1 and 2. Firstly, we don’t know how competing approaches would perform if we applied the paper’s proposed multi-layer concatenation trick, and the idea itself feels more like a way to give NCA’s performance a small boost and bring it into SOTA-like territory. Comparing NCA without multi-layer against other approaches is therefore more interesting to me. Secondly, 95% confidence intervals are provided, but the absence of identification of the best-performing approach(es) in each setting makes it hard to draw high-level conclusions at a glance. I would suggest bolding the best accuracy in each column along with all other entries for which a 95% confidence interval test on the difference between the means is inconclusive in determining that the difference is significant.\n\n#### Questions\n\n1. In Equation 2, why is the sum normalized by the total number of examples in the episode rather than the number of query examples?\n1. Can the authors comment on the extent to which Figure 2 supports the hypothesis that NCA is better for training because it learns from a larger number of positives and negatives? Assuming this is true, we should see that Prototypical Networks configurations that increase the number of positives and negatives should perform better for a given batch size. Does Figure 2 support this assertion?\n1. Can the authors elaborate on the \"no S/Q\" ablation (Figure 4, row 7)? What is the point of reference when computing distances for support and query examples? Is the loss computed in the same way for support and query examples? The text in Section 4.3 makes it appear like the loss for query examples is the NCA loss, but the loss for support examples is the prototypical loss. Wouldn’t it be conceptually cleaner to compute leave-one-out prototypes, i.e. leave each example out of the computation of its own class’ prototype (resulting in slightly different prototypes for examples of the same class)? In my mind, this would be the best way to remove the support/query partition while maintaining prototype computation, thereby showing that the partition is detrimental to Prototypical Networks training.\n\n#### Additional feedback\n\n1. This is somewhat inconsequential, but across all implementations of episodic training that I have examined I haven’t encountered an implementation that uses a flag to differentiate between support and query examples. Instead, the implementations I have examined explicitly represent support and query examples as separate tensors. I was therefore surprised to read that \"in most implementations [...] each image is characterised by a flag indicating whether it corresponds to the support or the query set [...]\"; can the authors point to the implementations they have in mind when making that assertion?\n1. I would be careful with the assertion that \"during evaluation the triplet {w, n, m} [...] must stay unchanged across methods\". While this is true for the benchmarks considered in this submission, benchmarks like Meta-Dataset evaluate on variable-ways and variable-shots episodes.\n1. I’m not too concerned with the computational efficiency of NCA. The pairwise Euclidean distances can be computed efficiently using the inner- and outer-product of the batch of embeddings with itself.\n", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603899195039}, {"id": "oVtw72_1b6S", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1898/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper investigates the usefulness of episodic learning in prototypical learning which is a popular practice in few-shot learning. The authors propose a non-episodic prototypical network which basically corresponds to the classical neighborhood component analysis and they claimed that this network reliably improves over its episodic counterpart in multiple datasets. I have the following comments on the paper:\n\n1. The sections 3.1, 3.2 and 3.3 are not the contributions of the paper. Only section 3.4 can be considered as something new from experimental point of view and not methodologically new. k-NN, 1-NN with class centroids and soft assignments are all some specific experimental settings. Therefore, I do not see the technical contributions of the paper other than the claimed novel experimental settings which is also marginal.\n\n2. The paper shows a robust experimentations and comparisons with prior arts, however, I don't understand how the three settings mentioned in the section 3.4 are evaluated in the tables.\n\n3. I am curious if you have done a comparison with baseline NCA, i.e. equation (3). I have not found the comparison in A.1 which only contains some discussions but no direct comparison.\n\n4. Anyways, The paper is written in good English and I haven't found any typos yet.\n\nBased on my current understandings and above comments, currently I recommend for a weak rejection. However, I would like to follow the discussions on the paper and understand the contributions well.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Technical contributions are not clear and experiments are generic", "review": "This paper investigates the usefulness of episodic learning in prototypical learning which is a popular practice in few-shot learning. The authors propose a non-episodic prototypical network which basically corresponds to the classical neighborhood component analysis and they claimed that this network reliably improves over its episodic counterpart in multiple datasets. I have the following comments on the paper:\n\n1. The sections 3.1, 3.2 and 3.3 are not the contributions of the paper. Only section 3.4 can be considered as something new from experimental point of view and not methodologically new. k-NN, 1-NN with class centroids and soft assignments are all some specific experimental settings. Therefore, I do not see the technical contributions of the paper other than the claimed novel experimental settings which is also marginal.\n\n2. The paper shows a robust experimentations and comparisons with prior arts, however, I don't understand how the three settings mentioned in the section 3.4 are evaluated in the tables.\n\n3. I am curious if you have done a comparison with baseline NCA, i.e. equation (3). I have not found the comparison in A.1 which only contains some discussions but no direct comparison.\n\n4. Anyways, The paper is written in good English and I haven't found any typos yet.\n\nBased on my current understandings and above comments, currently I recommend for a weak rejection. However, I would like to follow the discussions on the paper and understand the contributions well.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603882456777}, {"id": "3r0gbxU-Dfd", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1898/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper proposes to use neighborhood component analysis in lieu of prototype loss to train embedding functions of few-shot learning. This method takes full advantage of relations between all sampled points in an episode to facilitate learning, and it removes the distinction between support and query samples during meta-training time.\n\nReason for score: Overall, I lean towards reject. This paper proposes an interesting method that improves upon Prototypical Networks, and performs on-par with other baseline methods. However, the paper does not advance our understanding of how episodic training interacts with few-shot learning beyond the addition of more performance numbers to compare against. \n\nPros: The proposed method is a straight-forward improvement to Prototypical networks. In the wide range of scenarios evaluated in the experiments, the proposed method consistently outperforms ProtoNet. This proposed method should also be easy to implement, making integration with other FSL methods based on ProtoNets feasible. \nWhile not exactly nouveau, using NCA to train embedding functions has not been done before (to my best knowledge) and is a theoretically sound approach.\n\nCons:\n1. Perhaps unsurprisingly, the performance of NCA is lower than some FSL methods that use additional capacity. Even so, I think presenting only favorable comparisons in the performance tables is counter productive as it fails to capture the research context of this work. \n2. The conclusion from the ablation studies on batch size is still unclear to me. The “NCA fixed batch composition” setting seems to perform better than NCA in 3 of the 4 plots in figure 4. This particular setting is interesting as it allows control of the number of classes in each batch. As the batchsize is fixed in the ablation, we won’t know what is the relation between the number of ways and performance of this fixed batch variant of NCA. This ablation is also confusing in that neither “no proto” nor “no S/Q” significantly improves performance, yet their combination performs well. A full ablation that systematically covers all hyperparameters would be helpful in furthering understanding in this direction.\n3. The motivation of the paper feels unclear: on one hand the authors claim that they aim at understanding the (un)usefulness of episodic learning, yet on the other hand this paper doesn’t present any results beyond the final performance number to aid with this understanding. Visualizations and/or theoretical arguments would be greatly appreciated.\n\nMinor points (suggestions, not related to score):\nIntroduction\n“These results legitimately cast a doubt” -> “These results cast a doubt”\nRelated Work\n“Between 2016 and 2017” feels unnecessary\n“Matching and Prototypical…weighted by either an LSTM or a simple average, respectively”: Matching Networks also proposes a sample average variant. Their main difference lies in one uses cosine similarity while the other uses euclidean distance.\n“Differently from these papers” -> Different\nBACKGROUND AND METHOD\nIn 3.3, point 3 is not necessarily correct. Why would some examples be more likely than others in the episodic scheme? All FSL benchmarks (used in this paper) are close to class balanced, and an hierarchical sampling scheme for episodes results in full coverage each epoch just like standard supervised learning.\nIn 3.4 “which is the probability that image i is sampled from image j”. This is under the assumption that each support image defines a Gaussian in the embedding space. This is not true in general. The probabilistic interpretation of NCA should be further explained. \n\n[Post rebuttal]\nI am still leaning against the acceptance of this paper due to concerns about the limited impact of this submission. The topic of \"limitations of episodic training\" has been widely studied and in fact most SoA few-shot learning methods adopt a combination of supervised and episodic training for this precise reason. \n\nThe exploration into the relation between the number of sample pairs and learning performance is indeed correct, but no strong conclusions can be reached due to the logical jumps required. The authors established that the performance is correlated with number of pairs with a log/square-root curve, and that ProtoNet performance is similar to that of subsampled NCA (fig3). The mechanism behind why this is has not been elucidated. I think many questions can be explored to strengthen this paper, for example:\nAre classes embedded tighter together? \nAre hard negatives pushed further apart? \nDoes NCA induce a non-euclidean geometry that is more expressive than the euclidean geometry naturally induced by ProtoNets?\nCan the classification problem be converted into a pair comparison problem, so that PAC learning theory can be used to explain the shape of this curve? \nWhat is the sample complexity of the NCA classifier compared to the Prototypical classifier?\n\nRegarding the proposed method, NCA is certainly an improvement over ProtoNets, but performs worse than existing methods in most experiments. This makes me doubtful of the impact of this work on the methodological front. Better modelling of relationships between few-shot examples has been implicitly and explicitly studied too. For instance, many works adopt sample level set functions in the form of transformers, attention modules, and graph neural networks. Arguably, these additional architectures are more expressive \"deep\" alternatives to NCA, and hence achieves better performance than the proposed method.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of On Episodes, Prototypical Networks, and Few-Shot Learning ", "review": "Summary: This paper proposes to use neighborhood component analysis in lieu of prototype loss to train embedding functions of few-shot learning. This method takes full advantage of relations between all sampled points in an episode to facilitate learning, and it removes the distinction between support and query samples during meta-training time.\n\nReason for score: Overall, I lean towards reject. This paper proposes an interesting method that improves upon Prototypical Networks, and performs on-par with other baseline methods. However, the paper does not advance our understanding of how episodic training interacts with few-shot learning beyond the addition of more performance numbers to compare against. \n\nPros: The proposed method is a straight-forward improvement to Prototypical networks. In the wide range of scenarios evaluated in the experiments, the proposed method consistently outperforms ProtoNet. This proposed method should also be easy to implement, making integration with other FSL methods based on ProtoNets feasible. \nWhile not exactly nouveau, using NCA to train embedding functions has not been done before (to my best knowledge) and is a theoretically sound approach.\n\nCons:\n1. Perhaps unsurprisingly, the performance of NCA is lower than some FSL methods that use additional capacity. Even so, I think presenting only favorable comparisons in the performance tables is counter productive as it fails to capture the research context of this work. \n2. The conclusion from the ablation studies on batch size is still unclear to me. The “NCA fixed batch composition” setting seems to perform better than NCA in 3 of the 4 plots in figure 4. This particular setting is interesting as it allows control of the number of classes in each batch. As the batchsize is fixed in the ablation, we won’t know what is the relation between the number of ways and performance of this fixed batch variant of NCA. This ablation is also confusing in that neither “no proto” nor “no S/Q” significantly improves performance, yet their combination performs well. A full ablation that systematically covers all hyperparameters would be helpful in furthering understanding in this direction.\n3. The motivation of the paper feels unclear: on one hand the authors claim that they aim at understanding the (un)usefulness of episodic learning, yet on the other hand this paper doesn’t present any results beyond the final performance number to aid with this understanding. Visualizations and/or theoretical arguments would be greatly appreciated.\n\nMinor points (suggestions, not related to score):\nIntroduction\n“These results legitimately cast a doubt” -> “These results cast a doubt”\nRelated Work\n“Between 2016 and 2017” feels unnecessary\n“Matching and Prototypical…weighted by either an LSTM or a simple average, respectively”: Matching Networks also proposes a sample average variant. Their main difference lies in one uses cosine similarity while the other uses euclidean distance.\n“Differently from these papers” -> Different\nBACKGROUND AND METHOD\nIn 3.3, point 3 is not necessarily correct. Why would some examples be more likely than others in the episodic scheme? All FSL benchmarks (used in this paper) are close to class balanced, and an hierarchical sampling scheme for episodes results in full coverage each epoch just like standard supervised learning.\nIn 3.4 “which is the probability that image i is sampled from image j”. This is under the assumption that each support image defines a Gaussian in the embedding space. This is not true in general. The probabilistic interpretation of NCA should be further explained. \n\n[Post rebuttal]\nI am still leaning against the acceptance of this paper due to concerns about the limited impact of this submission. The topic of \"limitations of episodic training\" has been widely studied and in fact most SoA few-shot learning methods adopt a combination of supervised and episodic training for this precise reason. \n\nThe exploration into the relation between the number of sample pairs and learning performance is indeed correct, but no strong conclusions can be reached due to the logical jumps required. The authors established that the performance is correlated with number of pairs with a log/square-root curve, and that ProtoNet performance is similar to that of subsampled NCA (fig3). The mechanism behind why this is has not been elucidated. I think many questions can be explored to strengthen this paper, for example:\nAre classes embedded tighter together? \nAre hard negatives pushed further apart? \nDoes NCA induce a non-euclidean geometry that is more expressive than the euclidean geometry naturally induced by ProtoNets?\nCan the classification problem be converted into a pair comparison problem, so that PAC learning theory can be used to explain the shape of this curve? \nWhat is the sample complexity of the NCA classifier compared to the Prototypical classifier?\n\nRegarding the proposed method, NCA is certainly an improvement over ProtoNets, but performs worse than existing methods in most experiments. This makes me doubtful of the impact of this work on the methodological front. Better modelling of relationships between few-shot examples has been implicitly and explicitly studied too. For instance, many works adopt sample level set functions in the form of transformers, attention modules, and graph neural networks. Arguably, these additional architectures are more expressive \"deep\" alternatives to NCA, and hence achieves better performance than the proposed method.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603571696285}], "openreview_url": "https://openreview.net/forum?id=_TGlfdZOHY3", "arxiv_id": "2012.09831", "paper_pdf": "papers/_TGlfdZOHY3.pdf", "paper_pdf_sha256": "98f819786ee8688b553d7c79f4ec4d1fb048460dbb4b4a2ee65e06d1ee844dae", "paper_pdf_bytes": 504019, "paper_pdf_source": "openreview", "code_url": "https://github.com/fiveai/on-episodes-fsl", "code_repository": "fiveai/on-episodes-fsl", "code_commit": "90f2fe1244d5add4a93eea0768b49472ce42e504", "code_archive": "repos/_TGlfdZOHY3.zip", "code_archive_sha256": "6d0f00124ef08cc6b65cd40c6a68a1c56b8c1dd7692fd82099dcc7426bbee4ed", "code_archive_bytes": 72290, "code_file_count": 68, "code_extensions": {".sh": 46, ".py": 22}, "github_disk_usage_kb": 52, "github_languages": {"Python": 86516, "Shell": 43569}, "github_archived": false, "github_pushed_at": "2021-11-30T09:03:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-episodes-prototypical-networks-and-few-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryxmrpNtvH", "year": 2020, "status": "rejected", "title": "Deeper Insights into Weight Sharing in Neural Architecture Search", "authors": ["Yuge Zhang", "Quanlu Zhang", "Junyang Jiang", "Zejun Lin", "Yujing Wang"], "authorids": ["scottyugochang@gmail.com", "quanlu.zhang@microsoft.com", "jyjiang97@gmail.com", "gdzejlin@gmail.com", "yujing.wang@microsoft.com"], "authors_source": "OpenReview API", "abstract": "With the success of deep neural networks, Neural Architecture Search (NAS) as a way of automatic model design has attracted wide attention. As training every child model from scratch is very time-consuming, recent works leverage weight-sharing to speed up the model evaluation procedure. These approaches greatly reduce computation by maintaining a single copy of weights on the super-net and share the weights among every child model. However, weight-sharing has no theoretical guarantee and its impact has not been well studied before. In this paper, we conduct comprehensive experiments to reveal the impact of weight-sharing: (1) The best-performing models from different runs or even from consecutive epochs within the same run have significant variance; (2) Even with high variance, we can extract valuable information from training the super-net with shared weights; (3) The interference between child models is a main factor that induces high variance; (4) Properly reducing the degree of weight sharing could effectively reduce variance and improve performance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ryxxwa_CYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper516/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies weight sharing in neural architecture search (NAS). It constructs a mini search space with 64 possible choices, and performs various comparisons and studies in an exhaustive way. Some of the observations are quite interesting, exploring the limitations of weight sharing.\n\nMy biggest concern is the limited search space. Unlike other NAS works that usually have search space size > 10^10, this paper focuses on a very small search space (64 options in total). Because the search space is so small, a small change in any search option might cause a big difference for the sampled model, which possibly lead to some of the instability observed in this paper (such as observation 3 in Section 3.2 and the implication \"training a child model can easily perturb the rank of the previous mini-batch in section 4.1). However, this might not be true if the search space is big, where changing a few search options may not affect the supernet significantly.\n\nIt would be great if the authors can perform similar study on a larger search space. If evaluation for large search space is difficult, you may consider some pre-defined accuracy lookup tables (such as NAS-Bench-101: https://arxiv.org/abs/1902.09635).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper studies weight sharing in neural architecture search (NAS). It constructs a mini search space with 64 possible choices, and performs various comparisons and studies in an exhaustive way. Some of the observations are quite interesting, exploring the limitations of weight sharing.\n\nMy biggest concern is the limited search space. Unlike other NAS works that usually have search space size > 10^10, this paper focuses on a very small search space (64 options in total). Because the search space is so small, a small change in any search option might cause a big difference for the sampled model, which possibly lead to some of the instability observed in this paper (such as observation 3 in Section 3.2 and the implication \"training a child model can easily perturb the rank of the previous mini-batch in section 4.1). However, this might not be true if the search space is big, where changing a few search options may not affect the supernet significantly.\n\nIt would be great if the authors can perform similar study on a larger search space. If evaluation for large search space is difficult, you may consider some pre-defined accuracy lookup tables (such as NAS-Bench-101: https://arxiv.org/abs/1902.09635).\n"}, "tcdate": 1571880279541}, {"id": "SJlzZfOtKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper516/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "First of all, I have to state that this is not my area of expertise. So, my review here is an educated guess. \n\nThe paper is an empirical study that looks into the effect of weight sharing in neural network architecture search. The basic idea is that by sharing weights among multiple candidate architectures, the training process can be significantly improved.  In the literature, there have been mixed results, either in favor of or against weight sharing. The question this paper aims to address is to determine if weight sharing is justifiable and to what extent. \n\nThe primary subject investigated by the authors is to determine if the variance of the ranks generated by different runs of the algorithm are highly correlated with each other (e.g. using Kendall rank correlation score). Then, they compared such results with the ground truth (i.e. every child is trained independently). They found that the ranks generated by weight sharing are indeed highly correlated with each other, but there is much larger variance in the ranks when compared to the ground truth method. To understand why this is non-trivial: (1) on one hand, weight sharing speeds up the training process by providing an initial point close to a local minimum, but (2) the local minimum point may or may not  be good for the new architecture.  Hence, one does not know apriori under what conditions would weight sharing be a good strategy. \n\nThe authors also looked into variance of the rank within the same instance by examining how the rank changes with mini-batch epochs. They found that variance is large even within the same instance. \n\nMy primary concern is that the paper is entirely empirical with little if any justification of the results. In addition, it is based on a single architecture and a single dataset. This would have been fine if the results were supported with explanation or theoretical justification. Second, the ultimate goal is to improve the prediction accuracy, not the ranking accuracy. These are not necessarily equivalent. For instance, it is possible that the ranks have a high variance simply because many of the candidate architectures have nearly equivalent performance so the order within them becomes nearly random (and unimportant). In fact, I think the results support this conclusion (see for example Figure 10). Third, some of the highlighted observations are trivial. For example, Observation 1, which states that \"Two child models have (higher or lower) interference with each other when they share weights. A child model’s validation accuracy highly depends on the child models it is jointly trained with.\" I think this observation is trivial. \n\nSome other comments:\n- I would appreciate it if the authors could explain briefly how \"prefix sharing\" works so that the paper is self-contained. \n- The goal is to help improve the speed of neural architecture search. The authors mention \"hints for designing more efficient weight-sharing.\" Please state those conclusions precisely and clearly. I understand that the authors suggest similarity-based grouping. So, please mention clearly what you recommend in the conclusion section. \n\n\n========================== \n#post rebuttal remarks\n\nThanks for the response. Figure 10 was a typo from my end and I apologize for it. I actually meant figure 3. \n\nAs I said in my review, having an empirical study is acceptable provided that it covers many datasets, not just a single one. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "N/A", "title": "Official Blind Review #1", "review": "First of all, I have to state that this is not my area of expertise. So, my review here is an educated guess. \n\nThe paper is an empirical study that looks into the effect of weight sharing in neural network architecture search. The basic idea is that by sharing weights among multiple candidate architectures, the training process can be significantly improved.  In the literature, there have been mixed results, either in favor of or against weight sharing. The question this paper aims to address is to determine if weight sharing is justifiable and to what extent. \n\nThe primary subject investigated by the authors is to determine if the variance of the ranks generated by different runs of the algorithm are highly correlated with each other (e.g. using Kendall rank correlation score). Then, they compared such results with the ground truth (i.e. every child is trained independently). They found that the ranks generated by weight sharing are indeed highly correlated with each other, but there is much larger variance in the ranks when compared to the ground truth method. To understand why this is non-trivial: (1) on one hand, weight sharing speeds up the training process by providing an initial point close to a local minimum, but (2) the local minimum point may or may not  be good for the new architecture.  Hence, one does not know apriori under what conditions would weight sharing be a good strategy. \n\nThe authors also looked into variance of the rank within the same instance by examining how the rank changes with mini-batch epochs. They found that variance is large even within the same instance. \n\nMy primary concern is that the paper is entirely empirical with little if any justification of the results. In addition, it is based on a single architecture and a single dataset. This would have been fine if the results were supported with explanation or theoretical justification. Second, the ultimate goal is to improve the prediction accuracy, not the ranking accuracy. These are not necessarily equivalent. For instance, it is possible that the ranks have a high variance simply because many of the candidate architectures have nearly equivalent performance so the order within them becomes nearly random (and unimportant). In fact, I think the results support this conclusion (see for example Figure 10). Third, some of the highlighted observations are trivial. For example, Observation 1, which states that \"Two child models have (higher or lower) interference with each other when they share weights. A child model’s validation accuracy highly depends on the child models it is jointly trained with.\" I think this observation is trivial. \n\nSome other comments:\n- I would appreciate it if the authors could explain briefly how \"prefix sharing\" works so that the paper is self-contained. \n- The goal is to help improve the speed of neural architecture search. The authors mention \"hints for designing more efficient weight-sharing.\" Please state those conclusions precisely and clearly. I understand that the authors suggest similarity-based grouping. So, please mention clearly what you recommend in the conclusion section. \n\n\n========================== \n#post rebuttal remarks\n\nThanks for the response. Figure 10 was a typo from my end and I apologize for it. I actually meant figure 3. \n\nAs I said in my review, having an empirical study is acceptable provided that it covers many datasets, not just a single one. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571549690175}, {"id": "Byx_AfDsdr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper516/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Many NAS methods rely on weight sharing. Notably in ENAS, a single weight tensor is used for all candidate operations each edge of a cell. In this paper, the authors take a small NAS search space (64 possible networks) and train each network separately to obtain their individual rankings. They then examine how this ranking correlates when  the same network is trained as part of a super-net with weight sharing, as in NAS algorithms.\n\nGiven the prevalence of NAS algorithms, an examination of the potential pitfalls of weight sharing is very important and I commend the authors for that. There are a lack of typos and grammatical errors, which is nice too!\n\nI have two major issues with this paper however. Firstly, the scope is limited; everything is based on looking at 64 convnets trained on CIFAR-10, this makes it tricky to make any broad statements about weight-sharing in NAS. Secondly, the paper reads as a string of observations, and it is not clear what the takeaways are (it is hinted that one could reduce the search space for Figure 3, but this is not expanded on).\n\nA few chronological comments:\n\n- \"Population based algorithm () is another popular approach\" ---> \"Population-based algorithms are another popular approach\"\n\n- \"In this paper we try to answer\" --> A little confidence wouldn't hurt :)\n\n- \"Surprisingly\". I don't like the reader being told what is surprising/interesting etc. but maybe that's just me.\n\n- As mentioned above, more detail on search space pruning would be really nice. For instance, are particular operations e.g. conv5x5 neglected.\n\n- To clarify, when you have shared weights, but one candidate operation uses more parameters than another, do you just read off the weight tensor until you hit length? e.g. if convA uses X params, and convB uses Y params, are the parameters for ConvA weight(1:X) and convB weight(1:Y)?\n\n- Literature review is good. Figure 1 is nice and straightforward.\n\n- The figure and table captions could do with some more detail. For instance, in Figure 2 the caption should contain the take-home point of the figure.\n\n- \"there are some statistic information\" --> \"there is some statistical information\"\n\n- Figure 3 is nice. It looks like you can't tell what's good, but you can tell what's bad. A comparison of what architectures good v bad comprise off would be a nice addition.\n\n- Figure 5 confused me, as there is a lot going on. What do ordered and shuffled mean? Is it just whether you are mixing up your minibatch selections? \"The curve has obvious periodicity with the length of 64 mini-batches i.e. the number of child model\" doesn't make sense to me. Could you elaborate?\n\n- The accuracies in the plots look very low. ~80% for CIFAR-10 is really bad. Am I missing something?\n\nPros\n------\n- Good topic with a few interesting observations\n- Relatively well-written\n\nCons\n-------\n- Very limited scope. Only 1 dataset and only 64 models\n- The narrative is lacking, what are the key points that people using NAS should be aware of?\n\nI recommend a weak rejection for this paper. The topic is interesting, but I haven't been convinced through the limited scope of the experiments, or the arguments made what the real point is. Should I stop weight sharing with NAS? Should I prune my search space? etc. A few neat observations is nice, but there is a lack of cohesion. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "Many NAS methods rely on weight sharing. Notably in ENAS, a single weight tensor is used for all candidate operations each edge of a cell. In this paper, the authors take a small NAS search space (64 possible networks) and train each network separately to obtain their individual rankings. They then examine how this ranking correlates when  the same network is trained as part of a super-net with weight sharing, as in NAS algorithms.\n\nGiven the prevalence of NAS algorithms, an examination of the potential pitfalls of weight sharing is very important and I commend the authors for that. There are a lack of typos and grammatical errors, which is nice too!\n\nI have two major issues with this paper however. Firstly, the scope is limited; everything is based on looking at 64 convnets trained on CIFAR-10, this makes it tricky to make any broad statements about weight-sharing in NAS. Secondly, the paper reads as a string of observations, and it is not clear what the takeaways are (it is hinted that one could reduce the search space for Figure 3, but this is not expanded on).\n\nA few chronological comments:\n\n- \"Population based algorithm () is another popular approach\" ---> \"Population-based algorithms are another popular approach\"\n\n- \"In this paper we try to answer\" --> A little confidence wouldn't hurt :)\n\n- \"Surprisingly\". I don't like the reader being told what is surprising/interesting etc. but maybe that's just me.\n\n- As mentioned above, more detail on search space pruning would be really nice. For instance, are particular operations e.g. conv5x5 neglected.\n\n- To clarify, when you have shared weights, but one candidate operation uses more parameters than another, do you just read off the weight tensor until you hit length? e.g. if convA uses X params, and convB uses Y params, are the parameters for ConvA weight(1:X) and convB weight(1:Y)?\n\n- Literature review is good. Figure 1 is nice and straightforward.\n\n- The figure and table captions could do with some more detail. For instance, in Figure 2 the caption should contain the take-home point of the figure.\n\n- \"there are some statistic information\" --> \"there is some statistical information\"\n\n- Figure 3 is nice. It looks like you can't tell what's good, but you can tell what's bad. A comparison of what architectures good v bad comprise off would be a nice addition.\n\n- Figure 5 confused me, as there is a lot going on. What do ordered and shuffled mean? Is it just whether you are mixing up your minibatch selections? \"The curve has obvious periodicity with the length of 64 mini-batches i.e. the number of child model\" doesn't make sense to me. Could you elaborate?\n\n- The accuracies in the plots look very low. ~80% for CIFAR-10 is really bad. Am I missing something?\n\nPros\n------\n- Good topic with a few interesting observations\n- Relatively well-written\n\nCons\n-------\n- Very limited scope. Only 1 dataset and only 64 models\n- The narrative is lacking, what are the key points that people using NAS should be aware of?\n\nI recommend a weak rejection for this paper. The topic is interesting, but I haven't been convinced through the limited scope of the experiments, or the arguments made what the real point is. Should I stop weight sharing with NAS? Should I prune my search space? etc. A few neat observations is nice, but there is a lack of cohesion. "}, "tcdate": 1570628304145}], "openreview_url": "https://openreview.net/forum?id=ryxmrpNtvH", "arxiv_id": "2001.01431", "paper_pdf": "papers/ryxmrpNtvH.pdf", "paper_pdf_sha256": "c4559211c46f944c66f11a60b3b8c97880c11f3107cce8d85c0253e7c16eb119", "paper_pdf_bytes": 828821, "paper_pdf_source": "openreview", "code_url": "https://github.com/ultmaster/deeper-insights-weight-sharing", "code_repository": "ultmaster/deeper-insights-weight-sharing", "code_commit": "2a7f852cd22190b0ccd2bad3e7a1c208f01986b5", "code_archive": "repos/ryxmrpNtvH.zip", "code_archive_sha256": "6e6d34f04b91ffcb3bea90576ea93040991e4a57234d5373c73236c9b95836bb", "code_archive_bytes": 97620, "code_file_count": 27, "code_extensions": {".py": 24, ".sh": 3}, "github_disk_usage_kb": 69, "github_languages": {"Python": 124967, "Batchfile": 8031, "Shell": 1174}, "github_archived": false, "github_pushed_at": "2022-12-08T06:17:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deeper-insights-into-weight-sharing-in-neural-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1xYr3C5t7", "year": 2019, "status": "rejected", "title": "Neural Message Passing for Multi-Label Classification", "authors": ["Jack Lanchantin", "Arshdeep Sekhon", "Yanjun Qi"], "authorids": ["jjl5sw@virginia.edu", "as5cu@virginia.edu", "yq2h@virginia.edu"], "authors_source": "OpenReview API", "abstract": "Multi-label classification (MLC) is the task of assigning a set of target labels for a given sample. Modeling the combinatorial label interactions in MLC has been a long-haul challenge. Recurrent neural network (RNN) based encoder-decoder models have shown state-of-the-art performance for solving MLC. However, the sequential nature of modeling label dependencies through an RNN limits its ability in parallel computation, predicting dense labels, and providing interpretable results. In this paper, we propose Message Passing Encoder-Decoder (MPED) Networks,  aiming to provide fast, accurate, and interpretable MLC. MPED networks model the joint prediction of labels by replacing all RNNs in the encoder-decoder architecture with message passing mechanisms and dispense with autoregressive inference entirely.  The proposed models are simple, fast, accurate, interpretable, and structure-agnostic (can be used on known or unknown structured data). Experiments on seven real-world MLC datasets show the proposed models outperform autoregressive RNN models across five different metrics with a significant speedup during training and testing time.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "B1lmRDA037", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1560/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes an approach for using graph neural networks (GNN) to perform multi-label classification (MLC). The main idea is to use attentional pooling to project an input graph into a \"label graph\", whose nodes correspond to labels on some MLC problem. Multiple rounds of self-attention/message-passing hops can be performed on the input graph and label graph. Each output label is binary-valued, and is predicted from its corresponding node in the label graph. They evaluate on 6 multi-label sequence classification datasets, and report strong perform over baselines.\n\nThough interesting, I recommend rejection for several reasons:\n\n1) The technical contribution has limited novelty. One (very recent) reference this paper misses is \"Hierarchical Graph Representation Learning with Differentiable Pooling\" by Ying et al. (2018), which uses a very similar mechanism. The field is moving quickly, so references get missed sometimes, however from what I can tell, the graph-coarsening idea presented here isn't that technically distinct from Ying el al.'s. The Mrowca et al. (2018) \"Flexible Neural Representation for Physics Prediction\" is also fairly similar and should probably at least be cited.\n\n2) There aren't strong baselines. This approach is based on GNNs, and the Graph2MLP results, which is similar to previous GNN graph-level classification methods, are fairly strong too. My suspicion is that with some more tuning and tweaking, the results here would be similar to those of Ying et al., Velickovic et al. (2017)'s Graph Attention Nets, and other models which use what Gilmer et al. (2017) terms the \"readout\" function for MLC. Without testing some of these other approaches, how can readers be sure this is approach has value over other approaches? The reviews by Gilmer et al. (2017) and Battaglia et al. (2018) summarize a bunch of alternatives that could be tried, some of which use similar encoder/decoder setups (not with the attentional pooling, however, as far as I know).\n\n3) The writing is fairly dense for what is a fairly straightforward idea. And the paper is over 8.5 pages, with key details in the Appendix. \n\nI believe this approach could be quite powerful, and there was clearly a lot of excellent work that went into this project. But because the GNN area is very active, the bar is high. With a little more innovation on the model side (can the same core model be useful for things beyond MLC as well? I'm guessing it could), better baselines, better scholarship, and condensing the writing, I think this paper can be an important step forward.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting but weaker novelty/experiments/writing", "review": "The paper describes an approach for using graph neural networks (GNN) to perform multi-label classification (MLC). The main idea is to use attentional pooling to project an input graph into a \"label graph\", whose nodes correspond to labels on some MLC problem. Multiple rounds of self-attention/message-passing hops can be performed on the input graph and label graph. Each output label is binary-valued, and is predicted from its corresponding node in the label graph. They evaluate on 6 multi-label sequence classification datasets, and report strong perform over baselines.\n\nThough interesting, I recommend rejection for several reasons:\n\n1) The technical contribution has limited novelty. One (very recent) reference this paper misses is \"Hierarchical Graph Representation Learning with Differentiable Pooling\" by Ying et al. (2018), which uses a very similar mechanism. The field is moving quickly, so references get missed sometimes, however from what I can tell, the graph-coarsening idea presented here isn't that technically distinct from Ying el al.'s. The Mrowca et al. (2018) \"Flexible Neural Representation for Physics Prediction\" is also fairly similar and should probably at least be cited.\n\n2) There aren't strong baselines. This approach is based on GNNs, and the Graph2MLP results, which is similar to previous GNN graph-level classification methods, are fairly strong too. My suspicion is that with some more tuning and tweaking, the results here would be similar to those of Ying et al., Velickovic et al. (2017)'s Graph Attention Nets, and other models which use what Gilmer et al. (2017) terms the \"readout\" function for MLC. Without testing some of these other approaches, how can readers be sure this is approach has value over other approaches? The reviews by Gilmer et al. (2017) and Battaglia et al. (2018) summarize a bunch of alternatives that could be tried, some of which use similar encoder/decoder setups (not with the attentional pooling, however, as far as I know).\n\n3) The writing is fairly dense for what is a fairly straightforward idea. And the paper is over 8.5 pages, with key details in the Appendix. \n\nI believe this approach could be quite powerful, and there was clearly a lot of excellent work that went into this project. But because the GNN area is very active, the bar is high. With a little more innovation on the model side (can the same core model be useful for things beyond MLC as well? I'm guessing it could), better baselines, better scholarship, and condensing the writing, I think this paper can be an important step forward.\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541494731276}, {"id": "ByeckYZ937", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1560/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "As a reviewer I am expert in learning in structured data domains. Because of that I completely disagree that the proposed title of the paper is not misleading. In fact, both the input and the output of the proposed system are not graphs. Moreover, the intermediate representations are always complete graphs, so there is no graph to graph transformation here. It is the internal topology of the encoder and decoder that corresponds to a complete graph and not the nature of the processed data. \nThe main intended contribution of the paper is to define a system able to capture the dependencies among input features as well as output labels, so to improve the multi-label classification task addressed by the system. This is obtained by defining a recurrent model with a complete graph topology to both encode the input and decode the output. The decoding part starts from the assumption of independence among the output labels and then, via interaction with the encoded representation of the input, eventually turns to an output where relevant statistical dependences among output labels emerge with decoding. Since both encoding and decoding are recurrent models (with no enforced guarantee to have stable points), the paper proposes to unfold the recursion for a fixed predefined number of time steps.\nPresentation of the proposal is generally good, although there are some issues that are not clear. For example, the same weights indices are used for matrices belonging to the encoding and decoding, making the reader to believe that such matrices are shared. In addition, the sentence about model parameters at page 5 is a bit ambiguous and it is not sufficient to resolve the presentation problem. \nThe discussion at the end of page 4 on the fact that a sequential representation for the input components is not natural is actually out of place for the specific application task selected for presentation. In fact, words in a sentence have an order. The fact that such order is lost with the bag-of-word representation is a problem of preprocessing, not of the nature of the data. In general, however, it is true that forcing an order is not natural. \nGoing in the merit of the proposal, the number of parameters for the decoder scales quadratically with the number of output labels (fully connected graph). In domains with a large numbers of labels (e.g. thousands) there may be concerns on two different aspects: i) computational burden may grow significantly even if the average number of labels per item is small; ii) proper propagation of information on dependencies among labels may require to use a large value for T (graph hops), i.e. there is a dependency between size of label graph and \"useful\" value for T.  On this issue, by the way, figures 3 and 4 seem to report incongruent results since, because of symmetries in the model topology, equal and  reciprocal influences between input components (and output labels) would have been expected, but these are not observed in the figures. \nAnalogous considerations could be done for the encoder when the size of the input is large.\nConcerning experimental results, no statistical significance test is performed, so it is not clear to me if the shown improvements are actually significant. Speed-up in training and testing seem at least to give some advantage with respect to other competing approaches, however the scaling problem described above for the decoder (and encoder) may lead to much worst performances in those special cases.\nThe addressed problem is covered by a large literature, involving many different approaches. It would have been nice to report, for the selected datasets, the best performance (and computation times) obtained by, for example,  probabilistic graphical models or SVM-based models.\nThe paper seems to refer most of the relevant recent neural-based approaches.\nI think the paper is relevant for ICLR (although there is no explicit analysis of the obtained hidden representations) and of interest for a good portion of attendees. \n\nMinor issues: \n- two rows before Section 2.2.1: \\mathbb{h}_*^2  should be \\mathbb{h}_*^1\n- equations 4, 5, 9, 10, 14: matrices W are indexed in such a way to assume that each input word/label is associated to a different matrix (i.e., set of parameters). Is this really the case ? How is then managed the fact that different inputs may have a different number of components ? how is a specific matrix assigned to a specific word ? I guess this is a presentation mistake, otherwise there are relevant issues that are completely not addressed by the presentation.\n- equation (10): since the output should be interpreted as a probability, why not using a softmax? sigmoidal units by themselves do not guarantee that the outputs sum to 1. I guess you do not have this problem because you adopt batch normalisation. This however is conceptually not nice since there is no uniformity across the dataset. Moreover, the softmax function has a nice probabilistic interpretation in the family of the exponential distributions.\n- \"[...] we use add a positional encoding...\"\n- Multi-head Attention: apart for the not so clear description, the equation involving the softmax is missing.\n- \"[..] the the attention and feedforward layers.\"\n- \"[..] the the sum of the total true...\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A paper, with a misleading title, presenting experimental results for which statistical significance is not reported. ", "review": "As a reviewer I am expert in learning in structured data domains. Because of that I completely disagree that the proposed title of the paper is not misleading. In fact, both the input and the output of the proposed system are not graphs. Moreover, the intermediate representations are always complete graphs, so there is no graph to graph transformation here. It is the internal topology of the encoder and decoder that corresponds to a complete graph and not the nature of the processed data. \nThe main intended contribution of the paper is to define a system able to capture the dependencies among input features as well as output labels, so to improve the multi-label classification task addressed by the system. This is obtained by defining a recurrent model with a complete graph topology to both encode the input and decode the output. The decoding part starts from the assumption of independence among the output labels and then, via interaction with the encoded representation of the input, eventually turns to an output where relevant statistical dependences among output labels emerge with decoding. Since both encoding and decoding are recurrent models (with no enforced guarantee to have stable points), the paper proposes to unfold the recursion for a fixed predefined number of time steps.\nPresentation of the proposal is generally good, although there are some issues that are not clear. For example, the same weights indices are used for matrices belonging to the encoding and decoding, making the reader to believe that such matrices are shared. In addition, the sentence about model parameters at page 5 is a bit ambiguous and it is not sufficient to resolve the presentation problem. \nThe discussion at the end of page 4 on the fact that a sequential representation for the input components is not natural is actually out of place for the specific application task selected for presentation. In fact, words in a sentence have an order. The fact that such order is lost with the bag-of-word representation is a problem of preprocessing, not of the nature of the data. In general, however, it is true that forcing an order is not natural. \nGoing in the merit of the proposal, the number of parameters for the decoder scales quadratically with the number of output labels (fully connected graph). In domains with a large numbers of labels (e.g. thousands) there may be concerns on two different aspects: i) computational burden may grow significantly even if the average number of labels per item is small; ii) proper propagation of information on dependencies among labels may require to use a large value for T (graph hops), i.e. there is a dependency between size of label graph and \"useful\" value for T.  On this issue, by the way, figures 3 and 4 seem to report incongruent results since, because of symmetries in the model topology, equal and  reciprocal influences between input components (and output labels) would have been expected, but these are not observed in the figures. \nAnalogous considerations could be done for the encoder when the size of the input is large.\nConcerning experimental results, no statistical significance test is performed, so it is not clear to me if the shown improvements are actually significant. Speed-up in training and testing seem at least to give some advantage with respect to other competing approaches, however the scaling problem described above for the decoder (and encoder) may lead to much worst performances in those special cases.\nThe addressed problem is covered by a large literature, involving many different approaches. It would have been nice to report, for the selected datasets, the best performance (and computation times) obtained by, for example,  probabilistic graphical models or SVM-based models.\nThe paper seems to refer most of the relevant recent neural-based approaches.\nI think the paper is relevant for ICLR (although there is no explicit analysis of the obtained hidden representations) and of interest for a good portion of attendees. \n\nMinor issues: \n- two rows before Section 2.2.1: \\mathbb{h}_*^2  should be \\mathbb{h}_*^1\n- equations 4, 5, 9, 10, 14: matrices W are indexed in such a way to assume that each input word/label is associated to a different matrix (i.e., set of parameters). Is this really the case ? How is then managed the fact that different inputs may have a different number of components ? how is a specific matrix assigned to a specific word ? I guess this is a presentation mistake, otherwise there are relevant issues that are completely not addressed by the presentation.\n- equation (10): since the output should be interpreted as a probability, why not using a softmax? sigmoidal units by themselves do not guarantee that the outputs sum to 1. I guess you do not have this problem because you adopt batch normalisation. This however is conceptually not nice since there is no uniformity across the dataset. Moreover, the softmax function has a nice probabilistic interpretation in the family of the exponential distributions.\n- \"[...] we use add a positional encoding...\"\n- Multi-head Attention: apart for the not so clear description, the equation involving the softmax is missing.\n- \"[..] the the attention and feedforward layers.\"\n- \"[..] the the sum of the total true...\"\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541179617987}, {"id": "Sylw5GMBnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1560/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an encoder-decoder model based on the graph representation of inputs and outputs to solve the multi-label classification problem. The proposed model considers the output labels as a fully connected graph where the pair-wise interaction between labels can be modelled.\n\nOverall, although the proposed approach seems interesting, the representation of the paper needs to be improved. Below I listed some comments and suggestions about the paper.\n\n- The proposed model did not actually use any graph structure of input and output, which can potentially mislead the readers of the paper. For instance, the encoder is just a fully connected feed-forward network with an additional attention mechanism. In the same sense, the decoder is also just a fully connected feed-forward network. Furthermore, the inputs and outputs used throughout the paper do not have any graph structure or did not use any inferred graph structure from data. I recommend using any graph-structured data to show that the proposed model can actually work with the graph-structured data (with proper graph notations) or revise the manuscript without graph2graph representation.\n\n- I personally do not agree with the statement that the proposed model is interpretable because it can visualise the relation between labels through the attention. NN is hard to interpret because the weight structure cannot be intuitively interpretable. In the same sense, the proposed model cannot avoid the problem with the nature of black-box mechanism. Especially, multiple weight matrices are shared across the different layers, which makes it more difficult to interpret. Although the attention weights can be visualised, how can we visualise the decision process of the model from end-to-end? The question should be answered to claim that the model is interpretable.\n\n- 2.2.1, 2.2.2, 2.3 shares the similar network layer construction, which can be represented as a new layer of NN with different inputs (or at least 2.2.2 and 2.3 have the same layer structure). It would be better to encapsulate these explanations into a new NN module which can be reused multiple parts of the manuscript for a concise explanation.\n\n- Although the network claims to model the interactions between labels, the final prediction of labels are conditionally independent to each other, whereas the energy based models such as SPEN models the structure of output directly. In that sense, the model does not take into account the structure of output when the prediction is made although the underlying structure seems to model the 'pair-wise' interaction between labels.\n\n- In Table1, if the bold-face is used to emphasise the best outcome, I found it is inconsistent with the result (see the output of delicious and tfbs datasets).\n\n- Is it more natural to explain the encoder first followed by the decoder?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Graph2Graph without any graph structured inputs or outputs.", "review": "This paper proposes an encoder-decoder model based on the graph representation of inputs and outputs to solve the multi-label classification problem. The proposed model considers the output labels as a fully connected graph where the pair-wise interaction between labels can be modelled.\n\nOverall, although the proposed approach seems interesting, the representation of the paper needs to be improved. Below I listed some comments and suggestions about the paper.\n\n- The proposed model did not actually use any graph structure of input and output, which can potentially mislead the readers of the paper. For instance, the encoder is just a fully connected feed-forward network with an additional attention mechanism. In the same sense, the decoder is also just a fully connected feed-forward network. Furthermore, the inputs and outputs used throughout the paper do not have any graph structure or did not use any inferred graph structure from data. I recommend using any graph-structured data to show that the proposed model can actually work with the graph-structured data (with proper graph notations) or revise the manuscript without graph2graph representation.\n\n- I personally do not agree with the statement that the proposed model is interpretable because it can visualise the relation between labels through the attention. NN is hard to interpret because the weight structure cannot be intuitively interpretable. In the same sense, the proposed model cannot avoid the problem with the nature of black-box mechanism. Especially, multiple weight matrices are shared across the different layers, which makes it more difficult to interpret. Although the attention weights can be visualised, how can we visualise the decision process of the model from end-to-end? The question should be answered to claim that the model is interpretable.\n\n- 2.2.1, 2.2.2, 2.3 shares the similar network layer construction, which can be represented as a new layer of NN with different inputs (or at least 2.2.2 and 2.3 have the same layer structure). It would be better to encapsulate these explanations into a new NN module which can be reused multiple parts of the manuscript for a concise explanation.\n\n- Although the network claims to model the interactions between labels, the final prediction of labels are conditionally independent to each other, whereas the energy based models such as SPEN models the structure of output directly. In that sense, the model does not take into account the structure of output when the prediction is made although the underlying structure seems to model the 'pair-wise' interaction between labels.\n\n- In Table1, if the bold-face is used to emphasise the best outcome, I found it is inconsistent with the result (see the output of delicious and tfbs datasets).\n\n- Is it more natural to explain the encoder first followed by the decoder?", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1540854414804}], "openreview_url": "https://openreview.net/forum?id=r1xYr3C5t7", "arxiv_id": "1904.08049", "paper_pdf": "papers/r1xYr3C5t7.pdf", "paper_pdf_sha256": "216d0696d2fe98a0440e0f74dce95c7cc3ebdc6c06dc6746d2ee27b22f2b8f53", "paper_pdf_bytes": 1381301, "paper_pdf_source": "openreview", "code_url": "https://github.com/QData/LaMP", "code_repository": "QData/LaMP", "code_commit": "c4152b9370c59dec825b98f857756d80679a5230", "code_archive": "repos/r1xYr3C5t7.zip", "code_archive_sha256": "2befdd3fb4b532a5fef28f506f536d02bba26e40ec471ec9831f734ca6600cda", "code_archive_bytes": 125706, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 452, "github_languages": {"Python": 118220}, "github_archived": false, "github_pushed_at": "2024-07-21T20:28:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-message-passing-for-multi-label-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1U_af-0-", "year": 2018, "status": "rejected", "title": "Quadrature-based features for kernel approximation", "authors": ["Marina Munkhoeva", "Yermek Kapushev", "Evgeny Burnaev", "Ivan Oseledets"], "authorids": ["marina.munkhoeva@skolkovotech.ru", "kapushev@gmail.com", "e.burnaev@skoltech.ru", "i.oseledets@skoltech.ru"], "authors_source": "OpenReview API", "abstract": "We consider the problem of improving kernel approximation via feature maps. These maps arise as Monte Carlo approximation to integral representations of kernel functions and scale up kernel methods for larger datasets. We propose to use more efficient numerical integration technique to obtain better estimates of the integrals compared to the state-of-the-art methods. Our approach allows to use information about the integrand to enhance approximation and facilitates fast computations. We derive the convergence behavior and conduct an extensive empirical study that supports our hypothesis.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BkpB7yqxz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1012/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors offer a novel version of the random feature map approach to approximately solving large-scale kernel problems: each feature map evaluates the \"fourier feature\" corresponding to the kernel at a set of randomly sampled quadrature points. This gives an unbiased kernel estimator; they prove a bound its variance and provide experiment evidence that for Gaussian and arc-cos kernels, their suggested qaudrature rule outperforms previous random feature maps in terms of kernel approximation error and in terms of downstream classification and regression tasks. The idea is straightforward, the analysis seems correct, and the experiments suggest the method has superior accuracy compared to prior RFMs for shift-invariant kernels. The work is original, but I would say incremental, and the relevant literature is cited.\n\nThe method seems to give significantly lower kernel approximation errors, but the significance of the performance difference in downstream ML tasks is unclear --- the confidence intervals of the different methods overlap sufficiently to make it questionable whether the relative complexity of this method is worth the effort. Since good performance on downstream tasks is the crucial feature that we want RFMs to have, it is not clear that this method represents a true improvement over the state-of-the-art. The exposition of the quadrature method is difficult to follow, and the connection between the quadrature rules and the random feature map is never explicitly stated: e.g. equation 6 says how the kernel function is approximated as an integral, but does not give the feature map that an ML practitioner should use to get that approximate integral.\n\nIt would have been a good idea to include figures showing the time-accuracy tradeoff of the various methods, which is more important in large-scale ML applications than the kernel approximation error. It is not clear that the method is *not* more expensive in practice than previous methods (Table 1 gives superior asymptotic runtimes, but I would like to see actual run times, as evaluating the feature maps sound relatively complicated compared to other RFMs). On a related note, I would also like to have seen this method applied to kernels where the probability density in the Bochner integral was not the Gaussian density (e.g., the Laplacian kernel): the authors suggested that their method works there as well when one uses a Gaussian approximation of the density (which is not clear to me),  --- and it may be the case that sampling from their quadrature distribution is faster than sampling from the original non-Gaussian density.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "incremental development in random feature map approach", "rating": "7: Good paper, accept", "review": "The authors offer a novel version of the random feature map approach to approximately solving large-scale kernel problems: each feature map evaluates the \"fourier feature\" corresponding to the kernel at a set of randomly sampled quadrature points. This gives an unbiased kernel estimator; they prove a bound its variance and provide experiment evidence that for Gaussian and arc-cos kernels, their suggested qaudrature rule outperforms previous random feature maps in terms of kernel approximation error and in terms of downstream classification and regression tasks. The idea is straightforward, the analysis seems correct, and the experiments suggest the method has superior accuracy compared to prior RFMs for shift-invariant kernels. The work is original, but I would say incremental, and the relevant literature is cited.\n\nThe method seems to give significantly lower kernel approximation errors, but the significance of the performance difference in downstream ML tasks is unclear --- the confidence intervals of the different methods overlap sufficiently to make it questionable whether the relative complexity of this method is worth the effort. Since good performance on downstream tasks is the crucial feature that we want RFMs to have, it is not clear that this method represents a true improvement over the state-of-the-art. The exposition of the quadrature method is difficult to follow, and the connection between the quadrature rules and the random feature map is never explicitly stated: e.g. equation 6 says how the kernel function is approximated as an integral, but does not give the feature map that an ML practitioner should use to get that approximate integral.\n\nIt would have been a good idea to include figures showing the time-accuracy tradeoff of the various methods, which is more important in large-scale ML applications than the kernel approximation error. It is not clear that the method is *not* more expensive in practice than previous methods (Table 1 gives superior asymptotic runtimes, but I would like to see actual run times, as evaluating the feature maps sound relatively complicated compared to other RFMs). On a related note, I would also like to have seen this method applied to kernels where the probability density in the Bochner integral was not the Gaussian density (e.g., the Laplacian kernel): the authors suggested that their method works there as well when one uses a Gaussian approximation of the density (which is not clear to me),  --- and it may be the case that sampling from their quadrature distribution is faster than sampling from the original non-Gaussian density.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511809861102}, {"id": "SyGYH-ieG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1012/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper shows that techniques due to Genz & Monahan (1998) can be used to achieve low kernel approximation error under the framework of random fourier feature.\n\nPros\n\n1. It is new to apply quadrature rules to improve kernel approximation. The only other work I found is\nGaussian Quadrature for Kernel Features NIPS 2017. \nThe work is pretty recent so the author might not know it when submitting the paper. But in either case, it will be good to discuss the connections.\n\n2. The proposed method is shown to outperform a few baselines empirically.\n\nCons\n\n1. I don’t find the theoretical analysis to be very useful. In particular, the theorem shows that the kernel approximation error is O(1/D), which is the same as the original RFF paper. Unless the paper can provide a better characterization of the constants (like the ORF paper), it does not provide much insight in the proposed method. Unlike deep neural networks, since RFF is such a simple model, I think providing precise theoretical understanding is crucial. \n\n2. Approximating an integral is a well-studied topic. I do not find a good discussion on all the possible methods. Why is Genz & Monahan 1998 better than other alternatives such as Monte-Carlo, QMC etc? One argument seems to be “for kernels with specific specific integrand one can improve on its properties”. But this trick can be used for Monte-Carlo as well. And I do not see benefit of this trick in the curves.\n\n3. When choosing the orthogonal matrix, I think one obvious choice is to sample a matrix from the Stiefel manifold (the Q matrix of a random Gaussian). This baseline should be added in additional to H and B.\n\n4. A wall-time experiment is needed to justify the speedup.\n\nMinor comments:\n“For kennels with q(w) other than Gaussian… obtain very accurate results with little effort by using Gaussian approximation of q(w)”. What is the citation of this in the kernel approximation context?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method with good empirical result, but not enough insights on why", "rating": "6: Marginally above acceptance threshold", "review": "This paper shows that techniques due to Genz & Monahan (1998) can be used to achieve low kernel approximation error under the framework of random fourier feature.\n\nPros\n\n1. It is new to apply quadrature rules to improve kernel approximation. The only other work I found is\nGaussian Quadrature for Kernel Features NIPS 2017. \nThe work is pretty recent so the author might not know it when submitting the paper. But in either case, it will be good to discuss the connections.\n\n2. The proposed method is shown to outperform a few baselines empirically.\n\nCons\n\n1. I don’t find the theoretical analysis to be very useful. In particular, the theorem shows that the kernel approximation error is O(1/D), which is the same as the original RFF paper. Unless the paper can provide a better characterization of the constants (like the ORF paper), it does not provide much insight in the proposed method. Unlike deep neural networks, since RFF is such a simple model, I think providing precise theoretical understanding is crucial. \n\n2. Approximating an integral is a well-studied topic. I do not find a good discussion on all the possible methods. Why is Genz & Monahan 1998 better than other alternatives such as Monte-Carlo, QMC etc? One argument seems to be “for kernels with specific specific integrand one can improve on its properties”. But this trick can be used for Monte-Carlo as well. And I do not see benefit of this trick in the curves.\n\n3. When choosing the orthogonal matrix, I think one obvious choice is to sample a matrix from the Stiefel manifold (the Q matrix of a random Gaussian). This baseline should be added in additional to H and B.\n\n4. A wall-time experiment is needed to justify the speedup.\n\nMinor comments:\n“For kennels with q(w) other than Gaussian… obtain very accurate results with little effort by using Gaussian approximation of q(w)”. What is the citation of this in the kernel approximation context?", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511884154053}, {"id": "H1--71dlz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper1012/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to improve the kernel approximation of random features by using quadratures, in particular, stochastic spherical-radial rules. The quadrature rules have smaller variance given the same number of random features, and experiments show its reconstruction error and classification accuracies are better than existing algorithms.\n\nIt is an interesting paper, but it seems the authors are not aware of some existing works [1, 2] on quadrature for random features. Given these previous works, the contribution and novelty of the paper is limited.\n\n[1] Francis Bach. On the Equivalence between Kernel Quadrature Rules and Random Feature Expansions. JMLR, 2017.\n[2] Tri Dao, Christopher De Sa, Christopher Ré. Gaussian Quadrature for Kernel Features. NIPS 2017", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper, but lacks novelty and comparison to existing work", "rating": "4: Ok but not good enough - rejection", "review": "The paper proposes to improve the kernel approximation of random features by using quadratures, in particular, stochastic spherical-radial rules. The quadrature rules have smaller variance given the same number of random features, and experiments show its reconstruction error and classification accuracies are better than existing algorithms.\n\nIt is an interesting paper, but it seems the authors are not aware of some existing works [1, 2] on quadrature for random features. Given these previous works, the contribution and novelty of the paper is limited.\n\n[1] Francis Bach. On the Equivalence between Kernel Quadrature Rules and Random Feature Expansions. JMLR, 2017.\n[2] Tri Dao, Christopher De Sa, Christopher Ré. Gaussian Quadrature for Kernel Features. NIPS 2017", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511678713193}], "openreview_url": "https://openreview.net/forum?id=H1U_af-0-", "arxiv_id": "1802.03832", "paper_pdf": "papers/H1U_af-0-.pdf", "paper_pdf_sha256": "823493ecc0fc69703dd2500ae9b70fc7d104465a2d623a27395091c0630ee5cc", "paper_pdf_bytes": 1686033, "paper_pdf_source": "openreview", "code_url": "https://github.com/quffka/quffka", "code_repository": "quffka/quffka", "code_commit": "1599e6ce580e7482a4525d4df5fc654b10a72e96", "code_archive": "repos/H1U_af-0-.zip", "code_archive_sha256": "927db02153ce7dbd1b266ef3f991413da803517357d494c7614e9e9e6f912ed4", "code_archive_bytes": 4323707, "code_file_count": 29, "code_extensions": {".py": 19, ".ipynb": 9, ".sh": 1}, "github_disk_usage_kb": 3783, "github_languages": {"Jupyter Notebook": 2551738, "Python": 63776, "Shell": 1198}, "github_archived": false, "github_pushed_at": "2018-04-17T23:06:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/quadrature-based-features-for-kernel"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PbiQPxFLtQ", "year": 2026, "status": "rejected", "title": "How Explanations Leak the Decision Logic: Stealing Graph Neural Networks via Explanation Alignment", "authors": ["Bin Ma", "Yuyuan Feng", "Minhua Lin", "Enyan Dai"], "authorids": ["~Bin_Ma9", "~Yuyuan_Feng2", "~Minhua_Lin1", "~Enyan_Dai1"], "authors_source": "OpenReview API", "abstract": "Graph Neural Networks (GNNs) have become essential tools for analyzing graph-structured data in domains such as drug discovery and financial analysis, leading to growing demands for model transparency. Recent advances in explainable GNNs have addressed this need by revealing important subgraphs that influence predictions, but these explanation mechanisms may inadvertently expose models to security risks. This paper investigates how such explanations potentially leak critical decision logic that can be exploited for model stealing. We propose {\\method}, a novel stealing framework that integrates explanation alignment for capturing decision logic with guided data augmentation for efficient training under limited queries, enabling effective replication of both the predictive behavior and underlying reasoning patterns of target models. Experiments on molecular graph datasets demonstrate that our approach shows advantages over conventional methods in model stealing. This work highlights important security considerations for the deployment of explainable GNNs in sensitive domains and suggests the need for protective measures against explanation-based attacks. Our code is available at https://anonymous.4open.science/r/EGSteal-2BF7.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "pnDwaWSOG3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15094/Reviewer_kv62"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper shows that model explanations can leak decision logic and enable model stealing. The authors propose EGSteal, an attack framework that combines explanation alignment and data augmentation to replicate a target model’s predictions and reasoning under limited queries.", "review_text": "This paper shows that model explanations can leak decision logic and enable model stealing. The authors propose EGSteal, an attack framework that combines explanation alignment and data augmentation to replicate a target model’s predictions and reasoning under limited queries.", "strengths": "- The paper is clearly written and well structured, making it easy to follow the authors' ideas and understand the main contributions.\n- The figures are well designed and intuitive, effectively illustrating how the proposed method works and helping readers grasp the key concepts at a glance.", "weaknesses": "- The paper claims that GNN models are deployed online to protect intellectual property in critical applications (e.g., drug screening). However, in realistic IP-protection scenarios, online ML services are unlikely to provide model explanations, since explanations can reveal internal model logic and sensitive knowledge.\n\n- Only explanations that rely on internal model access (e.g., gradient- or attention-based, or self-interpretable models) provided by an online ML service are useful; for post-hoc explainers such as GNNExplainer or PGExplainer, attackers could obtain explanations locally. Suppose the target model relies on a post-hoc explainer such as PGExplainer. In that case, it is questionable whether the proposed method is stealing the GNN model itself or rather the explainer.  Even if explanations were accessible, different explanation methods often yield inconsistent results, so aligning a surrogate model based on Graph-CAM explanations without knowing the target’s explanation mechanism is weird. \n\n- The proposed data augmentation is motivated by the assumption that real-world APIs limit the number of allowed queries. However, no empirical evidence or real-world example is provided to support this claim. The authors should show that such query constraints exist in real-world services.\n\n- The theoretical assumption that $G_S$ is prediction-irrelevant is inconsistent with the implementation, where $G_S$ is determined by a rank-based threshold in Eq. (8). This creates two issues: (1) If $G_S$ is truly irrelevant, node scores should be 0 (or close to 0), rather than merely ranked low, since even low-ranked nodes could still have non-negligible importance; (2) The manual hyperparameter $\\alpha$ critically affects the construction of $G_E$ and $G_S$, as well as the rationality of data augmentation.\n\n- If the target model is truly a black box, the attacker should not know which explanation method the target uses (e.g., post-hoc vs. self-interpretable). However, most main experiments assume Graph-CAM explanations for the target model, which is overly idealized.\n\n- In the cross-dataset setting, the evaluation focuses on showing that the proposed method outperforms TS in performance metrics (Lines 396-398). However, for an attack method, the key objective should be whether the surrogate model truly replicates the target model’s behavior. The results on the AIDS dataset, where the AUC is below 60%, suggest that the proposed method may fail to effectively steal the target model when the in-distribution assumption does not hold. This raises concerns about the validity of the method under realistic scenarios.", "questions": "- The paper focuses on node-level explanations, but in graph applications, structure-level (i.e., edge-level) explanations are more common. Could the proposed method be directly extended to edge-level explanations? I suspect that, due to the specific design of the ranking-based loss, the proposed method might not generalize well to edge-level explanations.\n- The paper presents an experiment showing the robustness of the proposed method to noisy explanations. Could the authors clarify: (1) The motivation for this experiment: In what practical scenarios would the explanations become noisy? (2) Why is the proposed method robust to such noise? Is there any specific mechanism or design choice that contributes to this robustness?\n- In the experiments, the query budget is set to at least 10% of the training dataset. This means the attacker can issue hundreds or even thousands of queries. Is it motivated by any real-world scenario (e.g., are there existing systems that limit users to a few hundred queries)?\n\n\n**I found several aspects of the paper unconvincing as currently presented, which leads me to recommend rejection. That said, I may have misunderstood some points, and I would appreciate further clarification from the authors during the rebuttal phase. I apologize if any of my remarks are based on a misunderstanding, and I thank the authors in advance for clarifying these points.**", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper shows that model explanations can leak decision logic and enable model stealing. The authors propose EGSteal, an attack framework that combines explanation alignment and data augmentation to replicate a target model’s predictions and reasoning under limited queries.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper is clearly written and well structured, making it easy to follow the authors' ideas and understand the main contributions.\n- The figures are well designed and intuitive, effectively illustrating how the proposed method works and helping readers grasp the key concepts at a glance.", "weaknesses": "- The paper claims that GNN models are deployed online to protect intellectual property in critical applications (e.g., drug screening). However, in realistic IP-protection scenarios, online ML services are unlikely to provide model explanations, since explanations can reveal internal model logic and sensitive knowledge.\n\n- Only explanations that rely on internal model access (e.g., gradient- or attention-based, or self-interpretable models) provided by an online ML service are useful; for post-hoc explainers such as GNNExplainer or PGExplainer, attackers could obtain explanations locally. Suppose the target model relies on a post-hoc explainer such as PGExplainer. In that case, it is questionable whether the proposed method is stealing the GNN model itself or rather the explainer.  Even if explanations were accessible, different explanation methods often yield inconsistent results, so aligning a surrogate model based on Graph-CAM explanations without knowing the target’s explanation mechanism is weird. \n\n- The proposed data augmentation is motivated by the assumption that real-world APIs limit the number of allowed queries. However, no empirical evidence or real-world example is provided to support this claim. The authors should show that such query constraints exist in real-world services.\n\n- The theoretical assumption that $G_S$ is prediction-irrelevant is inconsistent with the implementation, where $G_S$ is determined by a rank-based threshold in Eq. (8). This creates two issues: (1) If $G_S$ is truly irrelevant, node scores should be 0 (or close to 0), rather than merely ranked low, since even low-ranked nodes could still have non-negligible importance; (2) The manual hyperparameter $\\alpha$ critically affects the construction of $G_E$ and $G_S$, as well as the rationality of data augmentation.\n\n- If the target model is truly a black box, the attacker should not know which explanation method the target uses (e.g., post-hoc vs. self-interpretable). However, most main experiments assume Graph-CAM explanations for the target model, which is overly idealized.\n\n- In the cross-dataset setting, the evaluation focuses on showing that the proposed method outperforms TS in performance metrics (Lines 396-398). However, for an attack method, the key objective should be whether the surrogate model truly replicates the target model’s behavior. The results on the AIDS dataset, where the AUC is below 60%, suggest that the proposed method may fail to effectively steal the target model when the in-distribution assumption does not hold. This raises concerns about the validity of the method under realistic scenarios.", "questions": "- The paper focuses on node-level explanations, but in graph applications, structure-level (i.e., edge-level) explanations are more common. Could the proposed method be directly extended to edge-level explanations? I suspect that, due to the specific design of the ranking-based loss, the proposed method might not generalize well to edge-level explanations.\n- The paper presents an experiment showing the robustness of the proposed method to noisy explanations. Could the authors clarify: (1) The motivation for this experiment: In what practical scenarios would the explanations become noisy? (2) Why is the proposed method robust to such noise? Is there any specific mechanism or design choice that contributes to this robustness?\n- In the experiments, the query budget is set to at least 10% of the training dataset. This means the attacker can issue hundreds or even thousands of queries. Is it motivated by any real-world scenario (e.g., are there existing systems that limit users to a few hundred queries)?\n\n\n**I found several aspects of the paper unconvincing as currently presented, which leads me to recommend rejection. That said, I may have misunderstood some points, and I would appreciate further clarification from the authors during the rebuttal phase. I apologize if any of my remarks are based on a misunderstanding, and I thank the authors in advance for clarifying these points.**", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761991403675}, {"id": "AcnBEUtDf6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15094/Reviewer_1dwp"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper propose a GNN model extraction attack method specifically targeting models with explanation subgraph provided.  The paper utilizes the causal relation between explanation in graph and label outcome to construct intervened graphs as augmented data. The author further design a training loss for surrogate model with explanation output to ensure both it replicates the classification result and interpretable mechanism.", "review_text": "This paper propose a GNN model extraction attack method specifically targeting models with explanation subgraph provided.  The paper utilizes the causal relation between explanation in graph and label outcome to construct intervened graphs as augmented data. The author further design a training loss for surrogate model with explanation output to ensure both it replicates the classification result and interpretable mechanism.", "strengths": "**1.** The idea of not only aligning mode classification but also aligning model explanation outcome is interesting. By requiring the surrogate model to have similar importance preference with target model, it may potentially make their inner mechanisms to become closer, achieving a better replication.\n\n**2.** The presentation of the paper is clean and sound. The notation in the paper is used clearly and formulations are also expressed tidily. The figures for illustration are easy to understand.\n\n**3.** The experimental results provided are reliable. There are many details of the algorithm design and hyper-parameters setting are discussed in the appendix, and the code is provided through the anonymous link. Attack performance under different attack capacity and architectures are also fully test, making the experimental results reliable.", "weaknesses": "**1.** The proposed intervened data as augmented data may introduce \"incorrectness\". In the causal analysis the author assumed the explanation subgraph $G_{E}$ is the part that decides to the classification,  while the style graph has little impact. So the author arbitrarily perturb the style graph while holding explanation subgraph to be the same to generate the intervened graph, while given the original label and same explanation subgraphs. However, this may be incorrect in some cases. For example, if we hope to indicate a benzene structure in a molecular (6 edges in a ring), the generated intervened graph may linking edges between them, and make the graph indeed no longer contain clean benzene and GT label changed, but the augmented data is still labeled as \"contain\", so incorrectness happens; the intervened model may also be created with benzene in $G_{S}$ and lead the real $G_{E}$ change, while the augmented data still holds the original one. In all, the assumption on Eq.(3) lacks of rationality since sometimes GNN not classify a specific subgraph region but a structure pattern, which should vary when $G_{S}$ change.  \n\n**2.** The chosen datasets all contain a lower quantity of average graph size within the dataset and lacks ground-truth explanations. The proposed intervened graphs hold a relatively large perturbation space, and when the graph scale is large, it would require much more samples to satisfy the Eq.(3) and Eq.(4). However, the chosen datasets in the experiments are relatively small on every single graph, which omits this potential problem. Besides, the chosen datasets do not contain a naturally existent groundtruth explanation, which makes it hard to known if the explained results $G_{E}$ really have the equality to hold Eq.(3) and (4).", "questions": "**1.**  As stated in appendix, the proposed loss term requires a $TQ|V|^{2}$ complexity to calculate all pairs of nodes' importance, which is a relatively high complexity. Be So I wonder if there's runtime record for conducting an extraction attack.\n\n**2.** Following the complexity problem, it seems that the main burden comes from the loss term on aligning explanation ranking on the surrogate model and target model. So how about remove this term? Would this largely decrease the model's performance? \n\n**3.** Is the surrogate model's explainer required to be the same with the target model? If not, would this architecture shift introduce additional error on replication?(ablation study suggested) If yes, this setting seems to lose practicability since the target model's explainer may not be known to the attacker, e.g., architecture or inner parameters.\n\nI suggest the author to at least conduct ablation study on the **questions 2&3**, because it leads to the necessity and rationality of designing the alignment loss term, which is currently unclear on its utility compared with its high complexity. If these two questions are properly addressed I would like to **raise the score to at least positive**.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose a GNN model extraction attack method specifically targeting models with explanation subgraph provided.  The paper utilizes the causal relation between explanation in graph and label outcome to construct intervened graphs as augmented data. The author further design a training loss for surrogate model with explanation output to ensure both it replicates the classification result and interpretable mechanism.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "**1.** The idea of not only aligning mode classification but also aligning model explanation outcome is interesting. By requiring the surrogate model to have similar importance preference with target model, it may potentially make their inner mechanisms to become closer, achieving a better replication.\n\n**2.** The presentation of the paper is clean and sound. The notation in the paper is used clearly and formulations are also expressed tidily. The figures for illustration are easy to understand.\n\n**3.** The experimental results provided are reliable. There are many details of the algorithm design and hyper-parameters setting are discussed in the appendix, and the code is provided through the anonymous link. Attack performance under different attack capacity and architectures are also fully test, making the experimental results reliable.", "weaknesses": "**1.** The proposed intervened data as augmented data may introduce \"incorrectness\". In the causal analysis the author assumed the explanation subgraph $G_{E}$ is the part that decides to the classification,  while the style graph has little impact. So the author arbitrarily perturb the style graph while holding explanation subgraph to be the same to generate the intervened graph, while given the original label and same explanation subgraphs. However, this may be incorrect in some cases. For example, if we hope to indicate a benzene structure in a molecular (6 edges in a ring), the generated intervened graph may linking edges between them, and make the graph indeed no longer contain clean benzene and GT label changed, but the augmented data is still labeled as \"contain\", so incorrectness happens; the intervened model may also be created with benzene in $G_{S}$ and lead the real $G_{E}$ change, while the augmented data still holds the original one. In all, the assumption on Eq.(3) lacks of rationality since sometimes GNN not classify a specific subgraph region but a structure pattern, which should vary when $G_{S}$ change.  \n\n**2.** The chosen datasets all contain a lower quantity of average graph size within the dataset and lacks ground-truth explanations. The proposed intervened graphs hold a relatively large perturbation space, and when the graph scale is large, it would require much more samples to satisfy the Eq.(3) and Eq.(4). However, the chosen datasets in the experiments are relatively small on every single graph, which omits this potential problem. Besides, the chosen datasets do not contain a naturally existent groundtruth explanation, which makes it hard to known if the explained results $G_{E}$ really have the equality to hold Eq.(3) and (4).", "questions": "**1.**  As stated in appendix, the proposed loss term requires a $TQ|V|^{2}$ complexity to calculate all pairs of nodes' importance, which is a relatively high complexity. Be So I wonder if there's runtime record for conducting an extraction attack.\n\n**2.** Following the complexity problem, it seems that the main burden comes from the loss term on aligning explanation ranking on the surrogate model and target model. So how about remove this term? Would this largely decrease the model's performance? \n\n**3.** Is the surrogate model's explainer required to be the same with the target model? If not, would this architecture shift introduce additional error on replication?(ablation study suggested) If yes, this setting seems to lose practicability since the target model's explainer may not be known to the attacker, e.g., architecture or inner parameters.\n\nI suggest the author to at least conduct ablation study on the **questions 2&3**, because it leads to the necessity and rationality of designing the alignment loss term, which is currently unclear on its utility compared with its high complexity. If these two questions are properly addressed I would like to **raise the score to at least positive**.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761966147842}, {"id": "kKdLv7xgGe", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15094/Reviewer_njmV"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "Since GNNs are widely used in many high-stake domains, the explanations on how GNN make predictions are becoming important. Existing GNN explanations can be obatined either by post-hoc explanations or self-explainable GNNs. While being useful to increase the transparency of decision making, it also increases the risk that the GNNs can be attacked (with the help of these explanations). This paper investigates the so-called model stealing attacks (namely replicating the prediction behaviour of the GNNs) with the help of GNN explanations. To achieve this, they assume that a graph can be devided into two groups: explanation subgraphs and style subgraphs. They further (strongly) assume that the explanation subgraph fully determinate the prediction while the other subgraphs have no impact on the GNN prediction. On this basis, they design a mimic model that not only aligns the predicted label but also the provided explanation subgraph. To train such a model, they augment the training pool by perturbing the style subgraphs without querying the target GNN model too many times. They perform many empirical studies to show the effectiveness and superiority of their method. However, I have many concerns that need more empirical studies, without which the soundness of this paper can be largely weakened.", "review_text": "Since GNNs are widely used in many high-stake domains, the explanations on how GNN make predictions are becoming important. Existing GNN explanations can be obatined either by post-hoc explanations or self-explainable GNNs. While being useful to increase the transparency of decision making, it also increases the risk that the GNNs can be attacked (with the help of these explanations). This paper investigates the so-called model stealing attacks (namely replicating the prediction behaviour of the GNNs) with the help of GNN explanations. To achieve this, they assume that a graph can be devided into two groups: explanation subgraphs and style subgraphs. They further (strongly) assume that the explanation subgraph fully determinate the prediction while the other subgraphs have no impact on the GNN prediction. On this basis, they design a mimic model that not only aligns the predicted label but also the provided explanation subgraph. To train such a model, they augment the training pool by perturbing the style subgraphs without querying the target GNN model too many times. They perform many empirical studies to show the effectiveness and superiority of their method. However, I have many concerns that need more empirical studies, without which the soundness of this paper can be largely weakened.", "strengths": "1. This paper is well organized;\n2. Many statements are supported with experiments;\n3. They tackle a novel problem;", "weaknesses": "The authors assume that the provided explanation subgraph fully determinate the GNN prediction. However, most post-hoc explanations may not reflect the actual decision logic [1]; besides, even self-explainable GNNs may provide explanations that are different from the underlying subgraphs that actually drived the predictions [2][3]. This issue is called unfaithful explanations, which have been widely recognized. Let me make this comment more actionable: \n\n(1) how could you guarantee (or measure) that the provided explanations actually reflect the decision logic, especially for post-hoc explanations? and under a black-box setting?\n\n(2) if you could not guarantee (or measure) this, what is the impact of the faithfulness of explanations on your method? Quantitaive and/or qualitative analyses are necessary; \n\n(3) what if there are multiple explanation subgraphs that align with the decision logic, but the explainer only provides a single one? You will treat the others as style subgraphs that do not impact the prediction? What will be the impact on your methodology?\n\n(4) What if the GNN predictor is a weak predictor? In other words, if the clasisfication accuracy is low (for example 0.4), the GNN classifier may not follow the true decision logic (if they follow the true decision logic, the accuracy should be high). Will your stealing model be able to align with the GNN preidction as well as the explanation?\n\n## References\n[1] Faithful and Consistent Graph Neural Network Explanations with Rationale Alignment\n\n[2] How Faithful are Self-Explainable GNNs?\n\n[3]  GNN Explanations that do not Explain and How to find Them\n\nPS: I am willing to increase my ratings if my concerns are (partially) addressed. (I may lower my ratings if they are completely ignored.)", "questions": "Please see the weak points", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Since GNNs are widely used in many high-stake domains, the explanations on how GNN make predictions are becoming important. Existing GNN explanations can be obatined either by post-hoc explanations or self-explainable GNNs. While being useful to increase the transparency of decision making, it also increases the risk that the GNNs can be attacked (with the help of these explanations). This paper investigates the so-called model stealing attacks (namely replicating the prediction behaviour of the GNNs) with the help of GNN explanations. To achieve this, they assume that a graph can be devided into two groups: explanation subgraphs and style subgraphs. They further (strongly) assume that the explanation subgraph fully determinate the prediction while the other subgraphs have no impact on the GNN prediction. On this basis, they design a mimic model that not only aligns the predicted label but also the provided explanation subgraph. To train such a model, they augment the training pool by perturbing the style subgraphs without querying the target GNN model too many times. They perform many empirical studies to show the effectiveness and superiority of their method. However, I have many concerns that need more empirical studies, without which the soundness of this paper can be largely weakened.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. This paper is well organized;\n2. Many statements are supported with experiments;\n3. They tackle a novel problem;", "weaknesses": "The authors assume that the provided explanation subgraph fully determinate the GNN prediction. However, most post-hoc explanations may not reflect the actual decision logic [1]; besides, even self-explainable GNNs may provide explanations that are different from the underlying subgraphs that actually drived the predictions [2][3]. This issue is called unfaithful explanations, which have been widely recognized. Let me make this comment more actionable: \n\n(1) how could you guarantee (or measure) that the provided explanations actually reflect the decision logic, especially for post-hoc explanations? and under a black-box setting?\n\n(2) if you could not guarantee (or measure) this, what is the impact of the faithfulness of explanations on your method? Quantitaive and/or qualitative analyses are necessary; \n\n(3) what if there are multiple explanation subgraphs that align with the decision logic, but the explainer only provides a single one? You will treat the others as style subgraphs that do not impact the prediction? What will be the impact on your methodology?\n\n(4) What if the GNN predictor is a weak predictor? In other words, if the clasisfication accuracy is low (for example 0.4), the GNN classifier may not follow the true decision logic (if they follow the true decision logic, the accuracy should be high). Will your stealing model be able to align with the GNN preidction as well as the explanation?\n\n## References\n[1] Faithful and Consistent Graph Neural Network Explanations with Rationale Alignment\n\n[2] How Faithful are Self-Explainable GNNs?\n\n[3]  GNN Explanations that do not Explain and How to find Them\n\nPS: I am willing to increase my ratings if my concerns are (partially) addressed. (I may lower my ratings if they are completely ignored.)", "questions": "Please see the weak points", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761554568736}], "openreview_url": "https://openreview.net/forum?id=PbiQPxFLtQ", "arxiv_id": "2506.03087", "paper_pdf": "papers/PbiQPxFLtQ.pdf", "paper_pdf_sha256": "dc3a2a0092de33ca0ea243fa4728f4dd91d4b852d61067c22f1afc8f13530290", "paper_pdf_bytes": 951582, "paper_pdf_source": "openreview", "code_url": "https://github.com/beanmah/EGSteal", "code_repository": "beanmah/EGSteal", "code_commit": "47a65a5dbec40d4156bcffac6eebb00322ae74a6", "code_archive": "repos/PbiQPxFLtQ.zip", "code_archive_sha256": "38ed2416e3cce64dd46681bd928ffad381b851cfad3af1f679f2e9a772a8caa1", "code_archive_bytes": 306624, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 294, "github_languages": {"Python": 217907}, "github_archived": false, "github_pushed_at": "2025-06-03T08:27:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-explanations-leak-the-decision-logic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NV5p50EkT6", "year": 2025, "status": "rejected", "title": "Channel-aware Contrastive Conditional Diffusion for Multivariate Probabilistic Time Series Forecasting", "authors": ["Siyang Li", "Yize Chen", "Hui Xiong"], "authorids": ["~Siyang_Li9", "~Yize_Chen1", "~Hui_Xiong1"], "authors_source": "OpenReview API", "abstract": "Forecasting faithful trajectories of multivariate time series from practical scopes is essential for reasonable decision-making. Recent methods majorly tailor generative conditional diffusion models to estimate the target temporal predictive distribution. However, it remains an obstacle to enhance the exploitation efficiency of given implicit temporal predictive information to bolster conditional diffusion learning. To this end, we propose a generic channel-aware contrastive conditional diffusion model termed CCDM to achieve desirable multivariate probabilistic forecasting, obviating the need for curated temporal conditioning inductive biases. In detail, we first design a channel-centric conditional denoising network to manage intra-variate variations and cross-variate correlations, which can lead to scalability on diverse prediction horizons and channel numbers. Then, we devise an ad-hoc denoising-based temporal contrastive learning to explicitly amplify the predictive mutual information between past observations and future forecasts. It can coherently complement naive step-wise denoising diffusion training and improve the forecasting accuracy and generality on unknown test time series. Besides, we offer theoretic insights on the benefits of such auxiliary contrastive training refinement from both neural mutual information and temporal distribution generalization aspects. The proposed CCDM can exhibit superior forecasting capability compared to current state-of-the-art diffusion forecasters over a comprehensive benchmark, with best MSE and CRPS outcomes on 79.17% and 87.5% cases.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "dfzjurl99r", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6674/Reviewer_m6S8"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper presents CCDM (Channel-aware Contrastive Conditional Diffusion Model), a novel approach for multivariate probabilistic time series forecasting. CCDM introduces two key innovations: a channel-aware conditional denoising network and a denoising-based temporal contrastive refinement. The channel-aware architecture incorporates channel-independent dense encoders and channel-mixing diffusion transformers to efficiently capture intra-variate and inter-variate temporal correlations. This design allows for better scalability across different prediction horizons and channel numbers. The denoising-based temporal contrastive refinement explicitly maximizes the prediction-related mutual information between past observations and future forecasts, complementing the standard denoising diffusion training.\n\nThe authors provide theoretical insights into the benefits of their approach from both neural mutual information and distribution generalization perspectives. They claim that CCDM exhibits superior forecasting capability compared to current state-of-the-art diffusion forecasters. The paper also presents a proposition on the upper bound of forecasting error for conditional diffusion models, linking the efficacy of conditional diffusion forecasters to the step-wise noise regression accuracy of the trained denoising network on unknown test time series.", "review_text": "The paper presents CCDM (Channel-aware Contrastive Conditional Diffusion Model), a novel approach for multivariate probabilistic time series forecasting. CCDM introduces two key innovations: a channel-aware conditional denoising network and a denoising-based temporal contrastive refinement. The channel-aware architecture incorporates channel-independent dense encoders and channel-mixing diffusion transformers to efficiently capture intra-variate and inter-variate temporal correlations. This design allows for better scalability across different prediction horizons and channel numbers. The denoising-based temporal contrastive refinement explicitly maximizes the prediction-related mutual information between past observations and future forecasts, complementing the standard denoising diffusion training.\n\nThe authors provide theoretical insights into the benefits of their approach from both neural mutual information and distribution generalization perspectives. They claim that CCDM exhibits superior forecasting capability compared to current state-of-the-art diffusion forecasters. The paper also presents a proposition on the upper bound of forecasting error for conditional diffusion models, linking the efficacy of conditional diffusion forecasters to the step-wise noise regression accuracy of the trained denoising network on unknown test time series.", "strengths": "The proposed CCDM (Channel-aware Contrastive Conditional Diffusion Model) introduces two key innovations that address important challenges in the field. First, it employs a channel-aware conditional denoising network that efficiently captures both intra-variate and inter-variate temporal correlations. This architecture, combining channel-independent dense encoders and channel-mixing diffusion transformers, allows for better scalability across different prediction horizons and channel numbers. Second, CCDM implements a denoising-based temporal contrastive refinement that explicitly maximizes the prediction-related mutual information between past observations and future forecasts. This approach complements the standard denoising diffusion training and improves forecasting accuracy and generalization on unknown test data.\n\nEmpirically, CCDM demonstrates superior forecasting capability compared to state-of-the-art diffusion forecasters. The method is also designed to be efficient, end-to-end, and seamlessly integrated with original simplified denoising diffusion optimization.", "weaknesses": "1. Lack of novelty and motivation. The paper's core contributions lack novelty in several aspects. The proposed Diffusion loss and mutual information concepts have been previously explored in the literature. Furthermore, the integration of contrastive learning loss within the Diffusion architecture seems arbitrary, and adding mutual information unnecessarily increases the complexity of the approach. While the authors attempt to justify their design choices with theoretical arguments, the discussion in lines 310-326 largely reiterates existing work (Tsai et al., 2020). The authors' decision to use KL loss approximation, necessitated by their inability to compute the upper bound in Proposition 1, is a standard approach in the field. This makes it difficult to justify the relationship between the two types of loss functions as a meaningful contribution.\n\n\n2. Presentation: The paper's visual presentation suffers from several clarity issues. Figure 1's bidirectional arrows create ambiguity regarding the complementary learning pathways, and it's unclear when the losses described in equation 5 and 6 are applied. Figure 2 compounds this confusion by failing to clearly explain the implementation and timing of negative time series augmentation. Additionally, the distinction between channel-independent and channel-dependent approaches is not well articulated. The inclusion of CiDM appears redundant given that the Diffusion Transformer depicted in Figure 3 inherently supports multivariate input processing. This design choice is particularly questionable given the absence of any ablation studies to justify its necessity or demonstrate its effectiveness. These presentation issues significantly impact the reader's ability to understand the proposed architecture and its implementation details.\n\n3. The experimental results lack clarity and comprehensive evaluation in several key aspects. First, although the proposed design makes no explicit claims about improving either short-term or long-term performance, the paper would benefit from including short-term prediction results using CSDI's benchmark datasets for better comparison. Second, the two-stage training process remains ambiguous throughout the paper. While line 289 claims to \"improve the conditional denoiser behaviors in out-of-distribution (OOD) regions,\" no empirical evidence is provided to support this OOD performance improvement. Given the inherent uncertainty in Diffusion models, the absence of mean and variance statistics for the experimental results is a significant oversight. This statistical ambiguity is particularly problematic in Table 8, where the ablation results showing different components' contributions appear inconsistent and potentially arbitrary, making it difficult to draw meaningful conclusions about the model's effectiveness.", "questions": "1. Lack of novelty in motivation:\n\t- The introduction of diffusion loss and mutual information is not novel.\n\t- The use of contrastive learning loss in a diffusion architecture is not intuitive.\n\t- Introducing mutual information makes the problem more complex.\n\t2.\tInsufficient theoretical justification:\n\t- The Theoretical Insights section (lines 310-326) appears to largely repeat (Tsai et al., 2020).\n\t- Proposition 1’s upper bound cannot be calculated in practice, which is why KL loss is used as an approximation. This is common knowledge and fails to effectively link the use of two different losses.\n\t3.\tUnclear figures and methodology:\n\t-  Figure 1’s bidirectional arrows are confusing, raising questions about when to use the losses in equations 5 and 6.\n\t- Figure 2 adds to this confusion, leaving unclear when and how negative time series augmentation is used.\n\t- In Figure 3, it’s unclear why CiDM is necessary when Diffusion Transformer seems can naturally handle multivariate input. No ablation study is provided to justify CiDM's important.\n\t4.\tExperimental results are difficult to follow:\n\t- Given that the proposed design doesn’t specifically aim to improve short-term or long-term performance, the reviewer suggests including results from the CSDI dataset (short-term prediction).\n\t- The two-stage training process is not clearly explained.\n\t- Despite claiming improved performance in out-of-distribution (OOD) regions (line 289), no evidence is provided to support this.\n\t- Given diffusion models’ inherent uncertainty, providing mean and variance of the data is crucial. This issue is particularly noticeable in Table 8, where different components’ contributions to model performance appear random.\n\t5.\tIncomplete evaluation:\n\t- The paper lacks results on short-term prediction datasets (like those used in CSDI).\n\t- There’s no clear demonstration of OOD performance improvement.\n\t- The absence of mean and variance data for the uncertainty inherent in diffusion models is a significant omission.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents CCDM (Channel-aware Contrastive Conditional Diffusion Model), a novel approach for multivariate probabilistic time series forecasting. CCDM introduces two key innovations: a channel-aware conditional denoising network and a denoising-based temporal contrastive refinement. The channel-aware architecture incorporates channel-independent dense encoders and channel-mixing diffusion transformers to efficiently capture intra-variate and inter-variate temporal correlations. This design allows for better scalability across different prediction horizons and channel numbers. The denoising-based temporal contrastive refinement explicitly maximizes the prediction-related mutual information between past observations and future forecasts, complementing the standard denoising diffusion training.\n\nThe authors provide theoretical insights into the benefits of their approach from both neural mutual information and distribution generalization perspectives. They claim that CCDM exhibits superior forecasting capability compared to current state-of-the-art diffusion forecasters. The paper also presents a proposition on the upper bound of forecasting error for conditional diffusion models, linking the efficacy of conditional diffusion forecasters to the step-wise noise regression accuracy of the trained denoising network on unknown test time series.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The proposed CCDM (Channel-aware Contrastive Conditional Diffusion Model) introduces two key innovations that address important challenges in the field. First, it employs a channel-aware conditional denoising network that efficiently captures both intra-variate and inter-variate temporal correlations. This architecture, combining channel-independent dense encoders and channel-mixing diffusion transformers, allows for better scalability across different prediction horizons and channel numbers. Second, CCDM implements a denoising-based temporal contrastive refinement that explicitly maximizes the prediction-related mutual information between past observations and future forecasts. This approach complements the standard denoising diffusion training and improves forecasting accuracy and generalization on unknown test data.\n\nEmpirically, CCDM demonstrates superior forecasting capability compared to state-of-the-art diffusion forecasters. The method is also designed to be efficient, end-to-end, and seamlessly integrated with original simplified denoising diffusion optimization.", "weaknesses": "1. Lack of novelty and motivation. The paper's core contributions lack novelty in several aspects. The proposed Diffusion loss and mutual information concepts have been previously explored in the literature. Furthermore, the integration of contrastive learning loss within the Diffusion architecture seems arbitrary, and adding mutual information unnecessarily increases the complexity of the approach. While the authors attempt to justify their design choices with theoretical arguments, the discussion in lines 310-326 largely reiterates existing work (Tsai et al., 2020). The authors' decision to use KL loss approximation, necessitated by their inability to compute the upper bound in Proposition 1, is a standard approach in the field. This makes it difficult to justify the relationship between the two types of loss functions as a meaningful contribution.\n\n\n2. Presentation: The paper's visual presentation suffers from several clarity issues. Figure 1's bidirectional arrows create ambiguity regarding the complementary learning pathways, and it's unclear when the losses described in equation 5 and 6 are applied. Figure 2 compounds this confusion by failing to clearly explain the implementation and timing of negative time series augmentation. Additionally, the distinction between channel-independent and channel-dependent approaches is not well articulated. The inclusion of CiDM appears redundant given that the Diffusion Transformer depicted in Figure 3 inherently supports multivariate input processing. This design choice is particularly questionable given the absence of any ablation studies to justify its necessity or demonstrate its effectiveness. These presentation issues significantly impact the reader's ability to understand the proposed architecture and its implementation details.\n\n3. The experimental results lack clarity and comprehensive evaluation in several key aspects. First, although the proposed design makes no explicit claims about improving either short-term or long-term performance, the paper would benefit from including short-term prediction results using CSDI's benchmark datasets for better comparison. Second, the two-stage training process remains ambiguous throughout the paper. While line 289 claims to \"improve the conditional denoiser behaviors in out-of-distribution (OOD) regions,\" no empirical evidence is provided to support this OOD performance improvement. Given the inherent uncertainty in Diffusion models, the absence of mean and variance statistics for the experimental results is a significant oversight. This statistical ambiguity is particularly problematic in Table 8, where the ablation results showing different components' contributions appear inconsistent and potentially arbitrary, making it difficult to draw meaningful conclusions about the model's effectiveness.", "questions": "1. Lack of novelty in motivation:\n\t- The introduction of diffusion loss and mutual information is not novel.\n\t- The use of contrastive learning loss in a diffusion architecture is not intuitive.\n\t- Introducing mutual information makes the problem more complex.\n\t2.\tInsufficient theoretical justification:\n\t- The Theoretical Insights section (lines 310-326) appears to largely repeat (Tsai et al., 2020).\n\t- Proposition 1’s upper bound cannot be calculated in practice, which is why KL loss is used as an approximation. This is common knowledge and fails to effectively link the use of two different losses.\n\t3.\tUnclear figures and methodology:\n\t-  Figure 1’s bidirectional arrows are confusing, raising questions about when to use the losses in equations 5 and 6.\n\t- Figure 2 adds to this confusion, leaving unclear when and how negative time series augmentation is used.\n\t- In Figure 3, it’s unclear why CiDM is necessary when Diffusion Transformer seems can naturally handle multivariate input. No ablation study is provided to justify CiDM's important.\n\t4.\tExperimental results are difficult to follow:\n\t- Given that the proposed design doesn’t specifically aim to improve short-term or long-term performance, the reviewer suggests including results from the CSDI dataset (short-term prediction).\n\t- The two-stage training process is not clearly explained.\n\t- Despite claiming improved performance in out-of-distribution (OOD) regions (line 289), no evidence is provided to support this.\n\t- Given diffusion models’ inherent uncertainty, providing mean and variance of the data is crucial. This issue is particularly noticeable in Table 8, where different components’ contributions to model performance appear random.\n\t5.\tIncomplete evaluation:\n\t- The paper lacks results on short-term prediction datasets (like those used in CSDI).\n\t- There’s no clear demonstration of OOD performance improvement.\n\t- The absence of mean and variance data for the uncertainty inherent in diffusion models is a significant omission.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730701617835}, {"id": "4O9ytYtOQB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6674/Reviewer_GGe5"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The authors describe a novel diffusion model CCDM for multivariate probabilistic time series forecasting. They propose a channel-wise diffusion model with a novel training loss based on ideas from contrastive learning.\nThe training input $x$ and target time series $y_k$, degraded by the diffusion forward process at stage $k$, are first encoded with a known per-channel CiDM module, resulting in two equal-sized latent representations. They are concatenated and input to a standard channel-wise DiT. The transformer output is decoded yielding the estimated diffusion noise $\\epsilon_k$.\nInstead of training the model with the classical log-likelihood diffusion loss, the authors propose to add a contrastive loss term which takes into account one positive and $N$ negative target samples for each input sample utilizing the InfoNCE loss as introduced by Oord et. al. for general time series applications. They prove an upper bound of the denoising diffusion-induced forecasting error and finally propose an algorithm to generate the negative samples.\nIn experiments, the model shows improved performance compared to SOTA models on different datasets. An ablation study shows the influence of the contrastive weight on the performance and reveals that the choice of the number of negative samples is critical and needs to be found empirically by hyperparameter search.", "review_text": "The authors describe a novel diffusion model CCDM for multivariate probabilistic time series forecasting. They propose a channel-wise diffusion model with a novel training loss based on ideas from contrastive learning.\nThe training input $x$ and target time series $y_k$, degraded by the diffusion forward process at stage $k$, are first encoded with a known per-channel CiDM module, resulting in two equal-sized latent representations. They are concatenated and input to a standard channel-wise DiT. The transformer output is decoded yielding the estimated diffusion noise $\\epsilon_k$.\nInstead of training the model with the classical log-likelihood diffusion loss, the authors propose to add a contrastive loss term which takes into account one positive and $N$ negative target samples for each input sample utilizing the InfoNCE loss as introduced by Oord et. al. for general time series applications. They prove an upper bound of the denoising diffusion-induced forecasting error and finally propose an algorithm to generate the negative samples.\nIn experiments, the model shows improved performance compared to SOTA models on different datasets. An ablation study shows the influence of the contrastive weight on the performance and reveals that the choice of the number of negative samples is critical and needs to be found empirically by hyperparameter search.", "strengths": "The idea of extending the diffusion loss with a contrastive loss is new and interesting. The architecture of the model is built from known building blocks, but not known in this combination. The authors give a thorough prove concerning the properties of the forecasting error. They present a comprehensive bibliography.", "weaknesses": "A general weakness of the interesting idea is that (p 20, section A10) the real effect of $N$ is intractable and the optimal value must be found by hyperparameter search.", "questions": "p 3: $y_0$ and $y_K$ is introduced one section too late\n\np 4: wording: accounts for … and can be any types of … (unclear)\n\np 5: replacing point-wise attention by channes-wise attention. (Could you elaborate this shortly?)\n\np 6 l 275: both a positive sample\n\np 7 l 327: efficiency\n\np 9: While the tables in the appendix show the influence of the constrastive learning, it is really hard to interpret the positive effect from figure 5. Can this be stated more clearly?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors describe a novel diffusion model CCDM for multivariate probabilistic time series forecasting. They propose a channel-wise diffusion model with a novel training loss based on ideas from contrastive learning.\nThe training input $x$ and target time series $y_k$, degraded by the diffusion forward process at stage $k$, are first encoded with a known per-channel CiDM module, resulting in two equal-sized latent representations. They are concatenated and input to a standard channel-wise DiT. The transformer output is decoded yielding the estimated diffusion noise $\\epsilon_k$.\nInstead of training the model with the classical log-likelihood diffusion loss, the authors propose to add a contrastive loss term which takes into account one positive and $N$ negative target samples for each input sample utilizing the InfoNCE loss as introduced by Oord et. al. for general time series applications. They prove an upper bound of the denoising diffusion-induced forecasting error and finally propose an algorithm to generate the negative samples.\nIn experiments, the model shows improved performance compared to SOTA models on different datasets. An ablation study shows the influence of the contrastive weight on the performance and reveals that the choice of the number of negative samples is critical and needs to be found empirically by hyperparameter search.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The idea of extending the diffusion loss with a contrastive loss is new and interesting. The architecture of the model is built from known building blocks, but not known in this combination. The authors give a thorough prove concerning the properties of the forecasting error. They present a comprehensive bibliography.", "weaknesses": "A general weakness of the interesting idea is that (p 20, section A10) the real effect of $N$ is intractable and the optimal value must be found by hyperparameter search.", "questions": "p 3: $y_0$ and $y_K$ is introduced one section too late\n\np 4: wording: accounts for … and can be any types of … (unclear)\n\np 5: replacing point-wise attention by channes-wise attention. (Could you elaborate this shortly?)\n\np 6 l 275: both a positive sample\n\np 7 l 327: efficiency\n\np 9: While the tables in the appendix show the influence of the constrastive learning, it is really hard to interpret the positive effect from figure 5. Can this be stated more clearly?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730541244198}, {"id": "sV6VMfNDKA", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6674/Reviewer_D78R"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This work proposes CCDM which is able to train a diffusion model for time series forecasting via the combination of a denoising loss and a contrastive loss, which they call a \"neural mutual information\" perspective.  This allows for the training of probabilistic forecasts in a deep learning setting, following other recent works.  By allowing for probabilistic forecasts, they are able to train a more accurate diffusion model when compared to historical regression-only style models.  Comparing across standard benchmark datasets and against several other baseline methods, favorable performance is achieved in the MSE (regression) metric and CRPS (probabilistic) metric.  Several ablation studies then confirm that the contrastive loss function is useful in diffusion model training.", "review_text": "This work proposes CCDM which is able to train a diffusion model for time series forecasting via the combination of a denoising loss and a contrastive loss, which they call a \"neural mutual information\" perspective.  This allows for the training of probabilistic forecasts in a deep learning setting, following other recent works.  By allowing for probabilistic forecasts, they are able to train a more accurate diffusion model when compared to historical regression-only style models.  Comparing across standard benchmark datasets and against several other baseline methods, favorable performance is achieved in the MSE (regression) metric and CRPS (probabilistic) metric.  Several ablation studies then confirm that the contrastive loss function is useful in diffusion model training.", "strengths": "The topic of how to combine the diffusion models of deep learning with the probabilistic forecasting of time series seems like an important and emerging area.\n\nThe specific success of contrastive learning in aiding the diffusion model in CRPS performance is demonstrated through ablation experiments.\n\nGood performance is achieved in many datasets and on many tasks.", "weaknesses": "The theoretical insights do not seem cohesive.\n\nThe contribution seems extremely simple or are not well highlighted.  After reading, I have the understanding that the major or only contribution is the application of contrastive learning to diffusion models, as depicted in Figure 1.  If that is the case, it is not properly highlighted and the challenges are not sufficiently discussed.\n\nFurther probabilistic analysis like the one in Figure 4 could be helpful for emphasizing the impact of this work.", "questions": "Can you clarify what is the difference between your proposed \"neural mutual information perspective\" and the combination of a denoising loss and contrastive loss?\n\nCan you explain what the contributions are compared to the existing works [1] and [2]?  From reading the related work section, I got the impression that the other diffusion-based methods which \"repurpose DiT\" should be the closest related works; however, getting to the experiments it seems these two methods are not compared.\n\nIs it possible to provide error bars for Figure 5? I understand the claim is that contrastive learning is helpful; however, looking at these figures, I get the completely opposite impression that actually the contrastive learning is doing almost nothing.  Perhaps it is possible to further explain these results.\n\nIf a key contribution of the work is the method to apply contrastive learning in the time series domain, can you elaborate on what the specific challenges of applying contrastive learning to time series data are?\n\n\n[1] \"Timedit: General-purpose diffusion transformers for time series foundation model\" Defu Cao et al. 2024.\n\n[2] \"Latent diffusion transformer for probabilistic time series forecasting\" Shibo Feng et al. 2024", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes CCDM which is able to train a diffusion model for time series forecasting via the combination of a denoising loss and a contrastive loss, which they call a \"neural mutual information\" perspective.  This allows for the training of probabilistic forecasts in a deep learning setting, following other recent works.  By allowing for probabilistic forecasts, they are able to train a more accurate diffusion model when compared to historical regression-only style models.  Comparing across standard benchmark datasets and against several other baseline methods, favorable performance is achieved in the MSE (regression) metric and CRPS (probabilistic) metric.  Several ablation studies then confirm that the contrastive loss function is useful in diffusion model training.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The topic of how to combine the diffusion models of deep learning with the probabilistic forecasting of time series seems like an important and emerging area.\n\nThe specific success of contrastive learning in aiding the diffusion model in CRPS performance is demonstrated through ablation experiments.\n\nGood performance is achieved in many datasets and on many tasks.", "weaknesses": "The theoretical insights do not seem cohesive.\n\nThe contribution seems extremely simple or are not well highlighted.  After reading, I have the understanding that the major or only contribution is the application of contrastive learning to diffusion models, as depicted in Figure 1.  If that is the case, it is not properly highlighted and the challenges are not sufficiently discussed.\n\nFurther probabilistic analysis like the one in Figure 4 could be helpful for emphasizing the impact of this work.", "questions": "Can you clarify what is the difference between your proposed \"neural mutual information perspective\" and the combination of a denoising loss and contrastive loss?\n\nCan you explain what the contributions are compared to the existing works [1] and [2]?  From reading the related work section, I got the impression that the other diffusion-based methods which \"repurpose DiT\" should be the closest related works; however, getting to the experiments it seems these two methods are not compared.\n\nIs it possible to provide error bars for Figure 5? I understand the claim is that contrastive learning is helpful; however, looking at these figures, I get the completely opposite impression that actually the contrastive learning is doing almost nothing.  Perhaps it is possible to further explain these results.\n\nIf a key contribution of the work is the method to apply contrastive learning in the time series domain, can you elaborate on what the specific challenges of applying contrastive learning to time series data are?\n\n\n[1] \"Timedit: General-purpose diffusion transformers for time series foundation model\" Defu Cao et al. 2024.\n\n[2] \"Latent diffusion transformer for probabilistic time series forecasting\" Shibo Feng et al. 2024", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730354560485}, {"id": "VMau38nIko", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6674/Reviewer_4FVe"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper innovatively designs a composite channel-aware conditional denoising network (CCDM) that merges a channel-independent dense encoder to extract univariate dynamics and a channel diffusion transformer to aggregate cross-variate correlations. And a time-contrast learning based on self-organized denoising is designed to explicitly amplify the predictive mutual information between past observations and future predictions. The method can consistently complement the plain stepwise denoising diffusion training to improve the prediction accuracy and generalization of unknown test time series.", "review_text": "This paper innovatively designs a composite channel-aware conditional denoising network (CCDM) that merges a channel-independent dense encoder to extract univariate dynamics and a channel diffusion transformer to aggregate cross-variate correlations. And a time-contrast learning based on self-organized denoising is designed to explicitly amplify the predictive mutual information between past observations and future predictions. The method can consistently complement the plain stepwise denoising diffusion training to improve the prediction accuracy and generalization of unknown test time series.", "strengths": "Learning generalisable multivariate probability distributions based on a step-wise denoising paradigm is a pivotal problem where there are two challenging obstacles, firstly how to establish cross-channel temporal correlation connections between observed and denoising predicted sequences, and subsequently how to improve the capability to mine implicit representations from a given time series. \n\nTo address these two issues, the authors propose two innovative architectures: \n\n* The channel-independent dense encoder and the channel-diffusion transformer to extract univariate and aggregated cross-variate correlations, respectively.\n\n* Self-organized denoising-based temporal contrast learning to consistently complement plain stepwise denoising-diffusion training.", "weaknesses": "**Weakness 1:** \n\nCCDM proposes contrastive learning for self-organizing denoising as a consistent complement to naive stepwise denoising diffusion training. Contrastive learning is a typical discriminative self-supervised learning. Compared with generative self-supervised learning, contrastive learning only needs to determine whether the sample is positive or not, and the low difficulty of the training task may lead to the model not being fully trained. \n\nIn addition, CCDM establishes the training paradigm based on a contrastive learning strategy that maximizes the mutual information between input and output. However, the model will be applied to the generative task based on prediction. The discriminative representation information learned by the model during the training phase varies dramatically for downstream generation tasks. To the best of our knowledge, large models trained in a self-supervised manner often need to design efficient fine-tuning strategies to adapt them to specific downstream tasks, while CCDMS directly reason on concrete datasets after being trained with pairwise learning.\n\n**Weakness 2:** \n\nThere has been a lot of research (Crossformer[1], LIFT[2], SAMformer[3]) on building cross-channel connections between multivariate time series. The hybrid channel-aware denoising architecture proposed by CDDM includes two main structures, Channel-wise Diffusion Transformer and Channel-independent Dense Module (CiDM). CiDM, which is not originally from this article but from TiDE[4], The Channel-wise Diffusion Transformer looks like it combines the Dimension-Segment-Wise (DSW) structure designed in Crossformer[1] with the conditional diffusion model. These shortcomings make CCDM an incremental summary work rather than innovative research.\n\n**Weakness 3:** \n\nThe lack of advanced baselines leads to the inability to verify the competitiveness of the proposed CCDM. Specifically, only four baselines based on Diffusion are shown in Table 1, and both SSSD and CSDI are published in 2021. \n\nMore extensive performance comparisons are expected to fully validate the effectiveness of the proposed method. The proposed baseline can be introduced in five sections:\n\n* Models based on pre-trained LLM alignment to TS, OneFitsAll, TimeLLM, LLM4TS.\n\n* Pre-trained time series base models on unified time series datasets from multiple domains, such as Timer, Moriai, Moment, UniTime.\n\n* Small datasets to train and test on specific datasets, e.g. PatchTST, iTransformer[5], ModernTCN[6].\n\n* Recent diffusion-based time series probabilistic prediction models, such as Diffuison-TS[7], mr-Diff[8], MG-TSD[9].\n\n* Self-supervised prediction models based on contrastive learning and mask reconstruction SimMTM[10].\n\nThe CCDM is expected to be compared with at least one competitive model in each forecasting paradigm to indicate the plausibility of the model design.\n\nEven compared with the four baselines in Table 1, CCDM has almost no performance improvement on all six datasets. Specifically, on the ETTh dataset, the MSE of CCDM is only 0.0017 lower than that of TimeDiff. On the Exchange, Appliance and Weather datasets, CCDM shows worse performance than TimeDiff and CSDI. On Electricity and Traffic datasets, the average improvement of CCDM in MSE metric is less than 10%.\n\n**Weakness 4:** \n\nIn the ablation experiment (Table 2), after removing contrastive refinement, the model improves only 1.2% and 1.08% on ECL and Traffic datasets. \n\nIn addition, another confusing result is that ETTh and Weather have a small number of channels (7 and 21), while ECL has a huge number of channels (321). However, after removing the channel-wise design, On the contrary, the performance in ETTh and Weather drastically decreases (23.34% and 40.28% in MSE), and the performance in the ECL barely changes (5.71% in MSE). \n\nThis may indicate that channel-wise designs fail to capture cross-channel latent representations from datasets with a large number of channels and instead perform well in other datasets with a smaller number of channels. The results of ablation experiments are negative and difficult to understand for the main thrust of this paper, \"establishing cross-channel connections in Diffusion models\".\n\n**Reference:** \n\n1) Zhang, Yunhao and Junchi Yan. “Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting.” ICLR 2023.\n\n2) Ilbert, Romain et al. “SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention.” ICML 2024.\n\n3) Zhao, Lifan and Yanyan Shen. “Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators.” ICLR 2024\n\n4) Das, Abhimanyu et al. “Long-term Forecasting with TiDE: Time-series Dense Encoder.” ArXiv abs/2304.08424 (2023): n. pag.\n\n5) Liu, Yong et al. “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.” ArXiv abs/2310.06625 (2023): n. pag.\n\n6) Luo, Donghao and Xue Wang. “ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis.” ICLR 2024.\n\n7) Yuan, Xinyu and Yan Qiao. “Diffusion-TS: Interpretable Diffusion for General Time Series Generation.” ICLR 2024.\n\n8) Shen, Lifeng et al. “Multi-Resolution Diffusion Models for Time Series Forecasting.” ICLR 2024.\n\n9) Fan, Xinyao et al. “MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process.” ICLR 2024.\n\n10) Dong, Jiaxiang et al. “SimMTM: A Simple Pre-Training Framework for Masked Time-Series Modeling.” NIPS 2023.", "questions": "**Question 1:** \n\nSee Weakness 3 for our concerns about the experimental results, and we believe that the introduction of competitive and up-to-date baselines is considered necessary.  In addition, the authors are expected to explain the phenomena in the ablation experiments, details of which can be referred to Weakness 4.\n\n**Question 2:** \n\nIt is mentioned in the main text that \"density ratio function can be any types of positive real functions\". Is there a clear strategy to determine the best density ratio function for different time-series data domains? Does designing different density ratio functions indirectly affect the training effect? How to construct similar and dissimilar instances to prevent model collapse? We believe that the construction of negative samples is critical to applying CCDM to a wide and diverse real-world setting.\n\n**Question 3:** \n\nAs far as we know, contrastive learning tends to consume a lot of computational resources. Specifically, we often need tens of thousands of iterations to fully train the diffusion model, and the total training overhead is staggering if we need to generate a set of n negative samples based on each positive example at each iteration. Therefore, theoretical analysis complexity and experimental results on time overhead of CCDM are necessary for their potential real-world applications.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper innovatively designs a composite channel-aware conditional denoising network (CCDM) that merges a channel-independent dense encoder to extract univariate dynamics and a channel diffusion transformer to aggregate cross-variate correlations. And a time-contrast learning based on self-organized denoising is designed to explicitly amplify the predictive mutual information between past observations and future predictions. The method can consistently complement the plain stepwise denoising diffusion training to improve the prediction accuracy and generalization of unknown test time series.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "Learning generalisable multivariate probability distributions based on a step-wise denoising paradigm is a pivotal problem where there are two challenging obstacles, firstly how to establish cross-channel temporal correlation connections between observed and denoising predicted sequences, and subsequently how to improve the capability to mine implicit representations from a given time series. \n\nTo address these two issues, the authors propose two innovative architectures: \n\n* The channel-independent dense encoder and the channel-diffusion transformer to extract univariate and aggregated cross-variate correlations, respectively.\n\n* Self-organized denoising-based temporal contrast learning to consistently complement plain stepwise denoising-diffusion training.", "weaknesses": "**Weakness 1:** \n\nCCDM proposes contrastive learning for self-organizing denoising as a consistent complement to naive stepwise denoising diffusion training. Contrastive learning is a typical discriminative self-supervised learning. Compared with generative self-supervised learning, contrastive learning only needs to determine whether the sample is positive or not, and the low difficulty of the training task may lead to the model not being fully trained. \n\nIn addition, CCDM establishes the training paradigm based on a contrastive learning strategy that maximizes the mutual information between input and output. However, the model will be applied to the generative task based on prediction. The discriminative representation information learned by the model during the training phase varies dramatically for downstream generation tasks. To the best of our knowledge, large models trained in a self-supervised manner often need to design efficient fine-tuning strategies to adapt them to specific downstream tasks, while CCDMS directly reason on concrete datasets after being trained with pairwise learning.\n\n**Weakness 2:** \n\nThere has been a lot of research (Crossformer[1], LIFT[2], SAMformer[3]) on building cross-channel connections between multivariate time series. The hybrid channel-aware denoising architecture proposed by CDDM includes two main structures, Channel-wise Diffusion Transformer and Channel-independent Dense Module (CiDM). CiDM, which is not originally from this article but from TiDE[4], The Channel-wise Diffusion Transformer looks like it combines the Dimension-Segment-Wise (DSW) structure designed in Crossformer[1] with the conditional diffusion model. These shortcomings make CCDM an incremental summary work rather than innovative research.\n\n**Weakness 3:** \n\nThe lack of advanced baselines leads to the inability to verify the competitiveness of the proposed CCDM. Specifically, only four baselines based on Diffusion are shown in Table 1, and both SSSD and CSDI are published in 2021. \n\nMore extensive performance comparisons are expected to fully validate the effectiveness of the proposed method. The proposed baseline can be introduced in five sections:\n\n* Models based on pre-trained LLM alignment to TS, OneFitsAll, TimeLLM, LLM4TS.\n\n* Pre-trained time series base models on unified time series datasets from multiple domains, such as Timer, Moriai, Moment, UniTime.\n\n* Small datasets to train and test on specific datasets, e.g. PatchTST, iTransformer[5], ModernTCN[6].\n\n* Recent diffusion-based time series probabilistic prediction models, such as Diffuison-TS[7], mr-Diff[8], MG-TSD[9].\n\n* Self-supervised prediction models based on contrastive learning and mask reconstruction SimMTM[10].\n\nThe CCDM is expected to be compared with at least one competitive model in each forecasting paradigm to indicate the plausibility of the model design.\n\nEven compared with the four baselines in Table 1, CCDM has almost no performance improvement on all six datasets. Specifically, on the ETTh dataset, the MSE of CCDM is only 0.0017 lower than that of TimeDiff. On the Exchange, Appliance and Weather datasets, CCDM shows worse performance than TimeDiff and CSDI. On Electricity and Traffic datasets, the average improvement of CCDM in MSE metric is less than 10%.\n\n**Weakness 4:** \n\nIn the ablation experiment (Table 2), after removing contrastive refinement, the model improves only 1.2% and 1.08% on ECL and Traffic datasets. \n\nIn addition, another confusing result is that ETTh and Weather have a small number of channels (7 and 21), while ECL has a huge number of channels (321). However, after removing the channel-wise design, On the contrary, the performance in ETTh and Weather drastically decreases (23.34% and 40.28% in MSE), and the performance in the ECL barely changes (5.71% in MSE). \n\nThis may indicate that channel-wise designs fail to capture cross-channel latent representations from datasets with a large number of channels and instead perform well in other datasets with a smaller number of channels. The results of ablation experiments are negative and difficult to understand for the main thrust of this paper, \"establishing cross-channel connections in Diffusion models\".\n\n**Reference:** \n\n1) Zhang, Yunhao and Junchi Yan. “Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting.” ICLR 2023.\n\n2) Ilbert, Romain et al. “SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention.” ICML 2024.\n\n3) Zhao, Lifan and Yanyan Shen. “Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators.” ICLR 2024\n\n4) Das, Abhimanyu et al. “Long-term Forecasting with TiDE: Time-series Dense Encoder.” ArXiv abs/2304.08424 (2023): n. pag.\n\n5) Liu, Yong et al. “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.” ArXiv abs/2310.06625 (2023): n. pag.\n\n6) Luo, Donghao and Xue Wang. “ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis.” ICLR 2024.\n\n7) Yuan, Xinyu and Yan Qiao. “Diffusion-TS: Interpretable Diffusion for General Time Series Generation.” ICLR 2024.\n\n8) Shen, Lifeng et al. “Multi-Resolution Diffusion Models for Time Series Forecasting.” ICLR 2024.\n\n9) Fan, Xinyao et al. “MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process.” ICLR 2024.\n\n10) Dong, Jiaxiang et al. “SimMTM: A Simple Pre-Training Framework for Masked Time-Series Modeling.” NIPS 2023.", "questions": "**Question 1:** \n\nSee Weakness 3 for our concerns about the experimental results, and we believe that the introduction of competitive and up-to-date baselines is considered necessary.  In addition, the authors are expected to explain the phenomena in the ablation experiments, details of which can be referred to Weakness 4.\n\n**Question 2:** \n\nIt is mentioned in the main text that \"density ratio function can be any types of positive real functions\". Is there a clear strategy to determine the best density ratio function for different time-series data domains? Does designing different density ratio functions indirectly affect the training effect? How to construct similar and dissimilar instances to prevent model collapse? We believe that the construction of negative samples is critical to applying CCDM to a wide and diverse real-world setting.\n\n**Question 3:** \n\nAs far as we know, contrastive learning tends to consume a lot of computational resources. Specifically, we often need tens of thousands of iterations to fully train the diffusion model, and the total training overhead is staggering if we need to generate a set of n negative samples based on each positive example at each iteration. Therefore, theoretical analysis complexity and experimental results on time overhead of CCDM are necessary for their potential real-world applications.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729700609991}], "openreview_url": "https://openreview.net/forum?id=NV5p50EkT6", "arxiv_id": "2410.02168", "paper_pdf": "papers/NV5p50EkT6.pdf", "paper_pdf_sha256": "225900e3e1ba2eb27fb1f149a72c79ebd0db158d0bb8363850843bae74ccb9e9", "paper_pdf_bytes": 14416893, "paper_pdf_source": "openreview", "code_url": "https://github.com/LSY-Cython/CCDM", "code_repository": "LSY-Cython/CCDM", "code_commit": "a9a4998d397861b5a3f081ed7f05baeedc88cd2b", "code_archive": "repos/NV5p50EkT6.zip", "code_archive_sha256": "154d248927c9f002bea3a149b186135d4683dc0bddebac000013d4b4f43426a5", "code_archive_bytes": 23141, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 47, "github_languages": {"Python": 75397}, "github_archived": false, "github_pushed_at": "2025-10-09T10:47:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/channel-aware-contrastive-conditional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QzQSR56JZr", "year": 2024, "status": "rejected", "title": "Harnessing the Power of Large Language Models for Natural Language to First-Order Logic Translation", "authors": ["Yuan Yang", "Siheng Xiong", "Ali Payani", "Ehsan Shareghi", "Faramarz Fekri"], "authorids": ["~Yuan_Yang1", "~Siheng_Xiong1", "~Ali_Payani1", "~Ehsan_Shareghi1", "~Faramarz_Fekri1"], "authors_source": "OpenReview API", "abstract": "Translating natural language sentences to first-order logic (NL-FOL translation) remains a critical task in many logic-based NLP systems, as it enables ML models to reason logically over text.\nHowever, existing translation methods still struggle to scale to real-world tasks due to the lack of a large and high-quality dataset and a model family with high precision and coverage.\n\nIn this work, we approach this longstanding challenge by harnessing the power of pre-trained large language models (LLMs).\nTo do so, we present MALLS (large language **M**odel gener**A**ted N**L**-FO**L** pair**S**), a dataset of 28K diverse and verified sentence-level NL-FOL pairs collected from GPT-4.\nWe create MALLS by implementing an adaptive pipeline that prompts GPT-4 for pairs with rich and diverse contexts. \nTo ensure the validity of FOL rules and their alignment with the NL sentences, we utilized a combined strategy of FOL rule parsing, human annotation, and automatic filtering.\n\nWe also present LogicLLaMA, an LLaMA2-7B/13B model family fine-tuned on MALLS for NL-FOL translation. \nLogicLLaMA is capable of directly translating natural language into FOL rules, which outperforms GPT-3.5. \nLogicLLaMA is also equipped to correct FOL rules predicted by GPT-3.5, and can achieve similar performance as GPT-4 with a fraction of the cost.\nThis correction ability was achieved by a novel reinforcement learning with human feedback (RLHF) framework, which initially trains on synthetically perturbed NL-FOL pairs to encourage chain-of-thought reasoning and then fine-tunes with RLHF on GPT-3.5 outputs using a FOL verifier as the reward model. \nCodes and data are available [here](https://www.dropbox.com/sh/t0f69776773e9er/AABKaWuvepUvhSp-0u2w-b2Pa?dl=0).", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "2R5cyDD1L4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3948/Reviewer_wVqn"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper contributed the MALLS dataset and the LogicLLAMA models for natural language to first-order-logic translation tasks. The authors also describe a multi-stage training paradigm using closed-source GPT models to generate datasets for training open-source LLAM for the NL-FOL translation tasks. Experiments are conducted to demonstrate the effectiveness of the proposal.", "review_text": "This paper contributed the MALLS dataset and the LogicLLAMA models for natural language to first-order-logic translation tasks. The authors also describe a multi-stage training paradigm using closed-source GPT models to generate datasets for training open-source LLAM for the NL-FOL translation tasks. Experiments are conducted to demonstrate the effectiveness of the proposal.", "strengths": "• Investing the possible links between large language models and logical reasoning is an important topics to advance the methodology for AI.\n• The paradigm of training open-sourced NL-FOL from open-sourced LLMs using data sourcing from close-sourced LLMs is a reasonable way to create NL-FOL models.\n• Experiments demonstrate almost start-of-the-art performance of NL-FOL tasks with the proposed LogicLLAMA model from LLAMA.", "weaknesses": "• The targeting FOL language lacks a formal characterization. Is it an fully expressive First-order logic language or just the logic program subset? \n•  The description of iterative correction via RLHF is difficult to follow lacking a few explicit description to task T4. In particular, it took me multiple-pass to understand why T4 is different from T3. It might worth a few rewriting to make this more explicit and easy to follow regarding the format of training data and loss functions.", "questions": "1. Page 7, regarding logical equivalence, which exact semantic models of FOL do you refer to? It is better to explicitly name the exact semantic models, such as Herbrand structure/universe if there is any; otherwise there might be ambiguity rendering the FOL being just a subset of FOL --- even if only a subset of FOL is supported it is already meaningful but please be specific. \n2.  Please formally define the 4 training tasks T1-T4 with explicit input-output definitions and loss definitions (if not enough space, putting them into supplemental materials is acceptable).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper contributed the MALLS dataset and the LogicLLAMA models for natural language to first-order-logic translation tasks. The authors also describe a multi-stage training paradigm using closed-source GPT models to generate datasets for training open-source LLAM for the NL-FOL translation tasks. Experiments are conducted to demonstrate the effectiveness of the proposal.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "• Investing the possible links between large language models and logical reasoning is an important topics to advance the methodology for AI.\n• The paradigm of training open-sourced NL-FOL from open-sourced LLMs using data sourcing from close-sourced LLMs is a reasonable way to create NL-FOL models.\n• Experiments demonstrate almost start-of-the-art performance of NL-FOL tasks with the proposed LogicLLAMA model from LLAMA.", "weaknesses": "• The targeting FOL language lacks a formal characterization. Is it an fully expressive First-order logic language or just the logic program subset? \n•  The description of iterative correction via RLHF is difficult to follow lacking a few explicit description to task T4. In particular, it took me multiple-pass to understand why T4 is different from T3. It might worth a few rewriting to make this more explicit and easy to follow regarding the format of training data and loss functions.", "questions": "1. Page 7, regarding logical equivalence, which exact semantic models of FOL do you refer to? It is better to explicitly name the exact semantic models, such as Herbrand structure/universe if there is any; otherwise there might be ambiguity rendering the FOL being just a subset of FOL --- even if only a subset of FOL is supported it is already meaningful but please be specific. \n2.  Please formally define the 4 training tasks T1-T4 with explicit input-output definitions and loss definitions (if not enough space, putting them into supplemental materials is acceptable).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699197661096}, {"id": "WUEysgkO9J", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3948/Reviewer_6TG1"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper aims to address the challenge of translating natural language sentences into first-order logic (FOL). A new dataset, MALLS, is proposed, which is generated by leveraging LLMs and applying filtering rules to eliminate bad cases. The paper proposed fine-tuning LLaMA on MALLS to enhance LLaMA's FOL generation ability. The model is tested on LogicNLI, FOLIO, and MALLS to demonstrate the effectiveness of the proposed method.", "review_text": "The paper aims to address the challenge of translating natural language sentences into first-order logic (FOL). A new dataset, MALLS, is proposed, which is generated by leveraging LLMs and applying filtering rules to eliminate bad cases. The paper proposed fine-tuning LLaMA on MALLS to enhance LLaMA's FOL generation ability. The model is tested on LogicNLI, FOLIO, and MALLS to demonstrate the effectiveness of the proposed method.", "strengths": "The paper demonstrates considerable effort in conducting numerous experiments to achieve its objective.\nThe motivation to enhance FOL generation ability is interesting.", "weaknesses": "1. The paper lacks novelty. Most contributions are either engineering work or published work including: prompting LLM to generate data, various prompting and dataset filtering methods, supervised fine-tuning, and RLHF. Moreover, the conclusions are predictable: fine-tuning LLM on a specific domain can improve its performance, potentially surpassing the best LLM for open domains.\n2. The proposed methods do not show significant improvement. As per Table 2, the majority of the gain comes from vanilla supervised fine-tuning on MALLS, and the newly proposed methods only bring marginal gains, especially when compared to the gain from the baseline to vanilla fine-tuning.", "questions": "1. The abstract should be written in a single paragraph.\n2. Consider reducing the use of markers like C1-C2, Q1-Q2, T1-T4. The frequent use of these symbols indicates a struggle to explain clearly, and it can be challenging for readers to understand by locating the definition. Short names might be more effective.\n3. Try to limit the number of your contributions and new terms. Highlight only the most significant things you want readers to remember. Currently, you have five prompting methods to create the dataset, four rules to filter the dataset, four fine-tuning methods to train LLaMA, and the experiment aims to address four research questions. This approach dilutes each contribution and makes it difficult to determine whether it works and can be generalized to new problems.\n4. Can your method generate more examples, since your generation and filtering are automatically executed?\n5. How do you assess the coverage or diversity of your dataset? LLM may only generate high-frequency knowledge.\n6. While the natural language input generated from LLM is often fluent and grammatically correct, real human inputs can be noisy. How do you address this issue?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to address the challenge of translating natural language sentences into first-order logic (FOL). A new dataset, MALLS, is proposed, which is generated by leveraging LLMs and applying filtering rules to eliminate bad cases. The paper proposed fine-tuning LLaMA on MALLS to enhance LLaMA's FOL generation ability. The model is tested on LogicNLI, FOLIO, and MALLS to demonstrate the effectiveness of the proposed method.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The paper demonstrates considerable effort in conducting numerous experiments to achieve its objective.\nThe motivation to enhance FOL generation ability is interesting.", "weaknesses": "1. The paper lacks novelty. Most contributions are either engineering work or published work including: prompting LLM to generate data, various prompting and dataset filtering methods, supervised fine-tuning, and RLHF. Moreover, the conclusions are predictable: fine-tuning LLM on a specific domain can improve its performance, potentially surpassing the best LLM for open domains.\n2. The proposed methods do not show significant improvement. As per Table 2, the majority of the gain comes from vanilla supervised fine-tuning on MALLS, and the newly proposed methods only bring marginal gains, especially when compared to the gain from the baseline to vanilla fine-tuning.", "questions": "1. The abstract should be written in a single paragraph.\n2. Consider reducing the use of markers like C1-C2, Q1-Q2, T1-T4. The frequent use of these symbols indicates a struggle to explain clearly, and it can be challenging for readers to understand by locating the definition. Short names might be more effective.\n3. Try to limit the number of your contributions and new terms. Highlight only the most significant things you want readers to remember. Currently, you have five prompting methods to create the dataset, four rules to filter the dataset, four fine-tuning methods to train LLaMA, and the experiment aims to address four research questions. This approach dilutes each contribution and makes it difficult to determine whether it works and can be generalized to new problems.\n4. Can your method generate more examples, since your generation and filtering are automatically executed?\n5. How do you assess the coverage or diversity of your dataset? LLM may only generate high-frequency knowledge.\n6. While the natural language input generated from LLM is often fluent and grammatically correct, real human inputs can be noisy. How do you address this issue?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698759193306}, {"id": "KLhOjUfK0x", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3948/Reviewer_8JQp"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper 28K sentence-level natural language to First Order Logic pairs collected from GPT-4. The authors also present a LLaMA-13B model fine-tuned on this dataset which combined with GPT3.5 achieves GPT-4 level performance on the NL-FOL translation tasks.\n\nThe combination method with GPT3.5 is interesting because it relies on using RLHF method to correct synthetically perturbed NL-FOL pairs and using a first order logic verifier as a reward model.", "review_text": "This paper 28K sentence-level natural language to First Order Logic pairs collected from GPT-4. The authors also present a LLaMA-13B model fine-tuned on this dataset which combined with GPT3.5 achieves GPT-4 level performance on the NL-FOL translation tasks.\n\nThe combination method with GPT3.5 is interesting because it relies on using RLHF method to correct synthetically perturbed NL-FOL pairs and using a first order logic verifier as a reward model.", "strengths": "This paper presents a new method for gathering RLHF feedback based on iterative correction which is a non-trivial extension of previous methods.", "weaknesses": "The techniques in the paper are not novel and fine-tuning LLAMA for a downstream task can be considered engineering at this point instead of active research. The novelty of the proposed approach is not totally clear.", "questions": "Even though the paper says that the LogicLLAMA is finetuned using RLHF there is no human feedback involved. It should probably be called something else ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper 28K sentence-level natural language to First Order Logic pairs collected from GPT-4. The authors also present a LLaMA-13B model fine-tuned on this dataset which combined with GPT3.5 achieves GPT-4 level performance on the NL-FOL translation tasks.\n\nThe combination method with GPT3.5 is interesting because it relies on using RLHF method to correct synthetically perturbed NL-FOL pairs and using a first order logic verifier as a reward model.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This paper presents a new method for gathering RLHF feedback based on iterative correction which is a non-trivial extension of previous methods.", "weaknesses": "The techniques in the paper are not novel and fine-tuning LLAMA for a downstream task can be considered engineering at this point instead of active research. The novelty of the proposed approach is not totally clear.", "questions": "Even though the paper says that the LogicLLAMA is finetuned using RLHF there is no human feedback involved. It should probably be called something else ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698644243412}], "openreview_url": "https://openreview.net/forum?id=QzQSR56JZr", "arxiv_id": "2305.15541", "paper_pdf": "papers/QzQSR56JZr.pdf", "paper_pdf_sha256": "c7fd4b01fbe5f423eef75311e3bd78fd126cb38dbb392e7757b85983cbce7ebf", "paper_pdf_bytes": 749189, "paper_pdf_source": "openreview", "code_url": "https://github.com/gblackout/LogicLLaMA", "code_repository": "gblackout/LogicLLaMA", "code_commit": "785a2c08e8fe964c8b2a10bb183ce5a08867aa3b", "code_archive": "repos/QzQSR56JZr.zip", "code_archive_sha256": "57d3792140e8239e651c6f1e7ec70eae49ff4bbb65d41b8aa7b7e976faba3555", "code_archive_bytes": 42110, "code_file_count": 16, "code_extensions": {".py": 9, ".sh": 6, ".ipynb": 1}, "github_disk_usage_kb": 40, "github_languages": {"Jupyter Notebook": 288177, "Python": 60800, "Shell": 2479}, "github_archived": false, "github_pushed_at": "2023-10-25T20:44:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/harnessing-the-power-of-large-language-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "CcXTudu9bvu", "year": 2023, "status": "rejected", "title": "DELTA: Diverse Client Sampling for Fasting Federated Learning", "authors": ["Lin Wang", "Yongxin Guo", "Tao Lin", "Xiaoying Tang"], "authorids": ["~Lin_Wang14", "~Yongxin_Guo1", "~Tao_Lin1", "~Xiaoying_Tang2"], "authors_source": "OpenReview API", "abstract": "Partial client participation has been widely adopted in Federated Learning (FL) to efficiently reduce the communication burden. However, an improper client sampling scheme will select unrepresentative subsets, which will cause a large variance in the model update and slows down the convergence. Existing sampling methods are either biased or can be further improved to accelerate the convergence. In this paper, we propose an unbiased sampling scheme, termed DELTA, to alleviate this problem. In particular, DELTA characterizes the impact of client diversity and local variance and samples the representative clients who carry valuable information for global model updates. Moreover, DELTA is a provably optimal unbiased sampling scheme that minimizes the variance caused by partial client participation and achieves better convergence than other unbiased sampling schemes. We corroborate our results with experiments on both synthetic and real data sets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Z5HcbBbrJ9Y", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6402/Reviewer_Yzew"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a new method to improve previous (cluster-based) important client sampling methods in federated learning. The new method is motivated by the insight that it would be beneficial to select clients from diverse groups. Convergence analysis are also provided and the authors claim they improve over existing ones. At last, experiments on FEMNIST and CIFAR-10 are provided to validate the performance.", "review_text": "I suspect that this paper has mistakes in the proof. The theoretical results may not be valid.", "strengths": "Strength\n- The idea of sampling from diverse gradient groups is a novel idea and seems to be promising.\n- The authors make the proposed sampling algorithm to be unbiased.\n\nWeakness\n- The writing of this paper should be improved. There are many vague statements. For example, in the introduction, the authors explain why previous works are not good. But all the explanations are just intuitions or conjectures. They cannot be used to support the observations in Figure 2. We do not know whether these conjectures are true or not (e.g. whether cluster-IS really select clients with small gradients and whether this is the core reason causing slow convergence).\n- Also, many mathematical expressions are wrong. For example, in equation 2 and 6, the function $f$ should also have some subscripts because its form changes when we sample different clients. Also, in equation (5), the authors define $E|\\nabla f(x)|^2 = E|\\nabla \\tilde{f}(x)|^2 + \\chi^2$. However, in equation (15), they wrote $E|\\nabla \\tilde{f}(x)|^2 = E|\\nabla f(x)|^2 + \\chi^2$. It is obvious these two equations are conflict with each other.\n- The proof may not be correct. I didn't find any proof details for Theorem D.2, which I suspect is wrong due to the above mistakes.\n- Remark 3.2 (3) is not accurate. \"Chen (2020) ... with additional gradient similarity bound\". This sounds like this paper did not do this and remove this assumption. But in fact, the authors define it in Assumption 3 and also use it.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new method to improve previous (cluster-based) important client sampling methods in federated learning. The new method is motivated by the insight that it would be beneficial to select clients from diverse groups. Convergence analysis are also provided and the authors claim they improve over existing ones. At last, experiments on FEMNIST and CIFAR-10 are provided to validate the performance.", "strength_and_weaknesses": "Strength\n- The idea of sampling from diverse gradient groups is a novel idea and seems to be promising.\n- The authors make the proposed sampling algorithm to be unbiased.\n\nWeakness\n- The writing of this paper should be improved. There are many vague statements. For example, in the introduction, the authors explain why previous works are not good. But all the explanations are just intuitions or conjectures. They cannot be used to support the observations in Figure 2. We do not know whether these conjectures are true or not (e.g. whether cluster-IS really select clients with small gradients and whether this is the core reason causing slow convergence).\n- Also, many mathematical expressions are wrong. For example, in equation 2 and 6, the function $f$ should also have some subscripts because its form changes when we sample different clients. Also, in equation (5), the authors define $E|\\nabla f(x)|^2 = E|\\nabla \\tilde{f}(x)|^2 + \\chi^2$. However, in equation (15), they wrote $E|\\nabla \\tilde{f}(x)|^2 = E|\\nabla f(x)|^2 + \\chi^2$. It is obvious these two equations are conflict with each other.\n- The proof may not be correct. I didn't find any proof details for Theorem D.2, which I suspect is wrong due to the above mistakes.\n- Remark 3.2 (3) is not accurate. \"Chen (2020) ... with additional gradient similarity bound\". This sounds like this paper did not do this and remove this assumption. But in fact, the authors define it in Assumption 3 and also use it.", "clarity,_quality,_novelty_and_reproducibility": "The quality of this paper should be improved. The authors should polish the writing with rigorous statements and check the correctness of the theoretical results.", "summary_of_the_review": "I suspect that this paper has mistakes in the proof. The theoretical results may not be valid.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667523587138}, {"id": "mhHJ9TxhRB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6402/Reviewer_LBTW"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes novel client sampling strategies to accelerate the convergence of \nfederated averaging methods with partial client participation. The idea is to determine \nthe sampling strategies to minimize variance from the worst-case convergence bounds of the methods.", "review_text": "I think this paper considers the problem of broad interest. However, the contributions seem to be unclear, due to unclear motivation on why and how the authors design optimal sampling strategies which are better than existing strategies. In addition, some of the proposed strategies cannot be implemented in practice, thus raising the issue on how they modify the strategies to be implementable in the experiments. \n", "strengths": "Designing optimal client sampling strategies is a problem of broad interest for designing scalable federated learning methods. \nThe authors attempt to **design optimal client sampling strategies in theory and practice** to accelerate the convergence of commonly used federated averaging methods. \n\nHowever, I would like to state the following **major concerns**: \n\n1. **Limitations of using the optimal sampling for DELTA.** \n\nFrom Corollary 3.4, the optimal sampling depends on $\\Vert \\nabla F_i(x) - \\nabla f(x) \\Vert$. This strategy cannot be implemented in practice, mainly because the gradient of the whole objective $\\nabla f(x)$ cannot be accessed usually for problems over huge amounts of data.\n\nTherefore, the authors should state how this strategy is modified to be able to implement it in their experiments.\n\n2. **Redundant assumptions for deriving the convergence results for FedIS and DELTA.**\n\nAssumptions 2 and 3 are commonly used for deriving convergence guarantees for federated methods. This is because Assumption 2 states how different the local stochastic gradient and the local full gradient is, and Assumption 3 implies how different the local full gradient and the whole full gradient is. \n\nHowever, Assumption 4 seems to be redundant to Assumption 2, since both of them impose the property of local stochastic gradients. Therefore, the stated convergence results for FedIS and DELTA seem too restricted. \n\n3. **Lack of motivation on how to design the optimal client sampling strategies for FedIS and DELTA.** \n\nSince FedIS or Algorithm 3 from (Chen et al., 2020) looks similar to DELTA, the proof techniques between these methods should be similar. Hence, I am not sure how the variance from convergence guarantees for FedIS and DELTA is different. Can the authors elaborate on this and perhaps add the motivation before stating theoretical results?  \n\n4. **Numerical evaluations against other unclear existing sampling strategies.** \n\nIn the experiments, the authors compared their sampling strategies against others, e.g. the power of choice, norm, and heterogeneity. However, these strategies are not clearly stated (i.e. what is $p_i^t$)? Are they existing sampling strategies, e.g., from (Chen et al., 2020)? \n\nSince different $p_i^t$ lead to different additional computational costs, it would be also more interesting to compare performance of the algorithms using different sampling with respect to the wall-clock time. \n\nFurthermore, I have the following **small concerns**: \n\n1. The step-size condition in Theorem 3.1 for $\\eta_L$ and $\\eta$ can be simplified to improve readability of this theorem. \n\nThe authors can write $\\eta_L < \\min\\( 1/(8LK), C \\)$ where the constant $C$ is obtained from the condition that\n$1/2 - (10L^2/m)\\sum_{i=1}^m K^2\\eta_L^2(A^2+1)>0$. Then, the author can state the condition for $\\eta$ which is $\\eta \\leq 1/(\\eta_L L)$.\n\nHowever, the constant $A$ is not clearly stated. Does it depend on the index of the client $i = 1,2,\\ldots,m$?\nCan the authors check this? \n\n2. The legend in Figure 5 includes FedSRC-G and FedSRC-D. I believe that they refer to DELTA and FEDIS. Can the authors check and edit the legend? \n\n3. Fedprox from (Li et al., 2018) does not use variance-reduction techniques to design federated optimization methods.\nTherefore, the authors should revise this part of the text which is in their contribution section.\n\n**Typo(s) I can spot:** \n- the data heterogeneity to **fast** the convergence speed $\\rightarrow$  the data heterogeneity to **accelerate** the convergence speed.\n\nNote that due to limited time, I cannot check convergence proofs carefully. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes novel client sampling strategies to accelerate the convergence of \nfederated averaging methods with partial client participation. The idea is to determine \nthe sampling strategies to minimize variance from the worst-case convergence bounds of the methods.", "strength_and_weaknesses": "Designing optimal client sampling strategies is a problem of broad interest for designing scalable federated learning methods. \nThe authors attempt to **design optimal client sampling strategies in theory and practice** to accelerate the convergence of commonly used federated averaging methods. \n\nHowever, I would like to state the following **major concerns**: \n\n1. **Limitations of using the optimal sampling for DELTA.** \n\nFrom Corollary 3.4, the optimal sampling depends on $\\Vert \\nabla F_i(x) - \\nabla f(x) \\Vert$. This strategy cannot be implemented in practice, mainly because the gradient of the whole objective $\\nabla f(x)$ cannot be accessed usually for problems over huge amounts of data.\n\nTherefore, the authors should state how this strategy is modified to be able to implement it in their experiments.\n\n2. **Redundant assumptions for deriving the convergence results for FedIS and DELTA.**\n\nAssumptions 2 and 3 are commonly used for deriving convergence guarantees for federated methods. This is because Assumption 2 states how different the local stochastic gradient and the local full gradient is, and Assumption 3 implies how different the local full gradient and the whole full gradient is. \n\nHowever, Assumption 4 seems to be redundant to Assumption 2, since both of them impose the property of local stochastic gradients. Therefore, the stated convergence results for FedIS and DELTA seem too restricted. \n\n3. **Lack of motivation on how to design the optimal client sampling strategies for FedIS and DELTA.** \n\nSince FedIS or Algorithm 3 from (Chen et al., 2020) looks similar to DELTA, the proof techniques between these methods should be similar. Hence, I am not sure how the variance from convergence guarantees for FedIS and DELTA is different. Can the authors elaborate on this and perhaps add the motivation before stating theoretical results?  \n\n4. **Numerical evaluations against other unclear existing sampling strategies.** \n\nIn the experiments, the authors compared their sampling strategies against others, e.g. the power of choice, norm, and heterogeneity. However, these strategies are not clearly stated (i.e. what is $p_i^t$)? Are they existing sampling strategies, e.g., from (Chen et al., 2020)? \n\nSince different $p_i^t$ lead to different additional computational costs, it would be also more interesting to compare performance of the algorithms using different sampling with respect to the wall-clock time. \n\nFurthermore, I have the following **small concerns**: \n\n1. The step-size condition in Theorem 3.1 for $\\eta_L$ and $\\eta$ can be simplified to improve readability of this theorem. \n\nThe authors can write $\\eta_L < \\min\\( 1/(8LK), C \\)$ where the constant $C$ is obtained from the condition that\n$1/2 - (10L^2/m)\\sum_{i=1}^m K^2\\eta_L^2(A^2+1)>0$. Then, the author can state the condition for $\\eta$ which is $\\eta \\leq 1/(\\eta_L L)$.\n\nHowever, the constant $A$ is not clearly stated. Does it depend on the index of the client $i = 1,2,\\ldots,m$?\nCan the authors check this? \n\n2. The legend in Figure 5 includes FedSRC-G and FedSRC-D. I believe that they refer to DELTA and FEDIS. Can the authors check and edit the legend? \n\n3. Fedprox from (Li et al., 2018) does not use variance-reduction techniques to design federated optimization methods.\nTherefore, the authors should revise this part of the text which is in their contribution section.\n\n**Typo(s) I can spot:** \n- the data heterogeneity to **fast** the convergence speed $\\rightarrow$  the data heterogeneity to **accelerate** the convergence speed.\n\nNote that due to limited time, I cannot check convergence proofs carefully. \n", "clarity,_quality,_novelty_and_reproducibility": "The clarity and novelty of the work should be improved and better highlighted. Details related to these concerns are mentioned in the section of strength and weaknesses. ", "summary_of_the_review": "I think this paper considers the problem of broad interest. However, the contributions seem to be unclear, due to unclear motivation on why and how the authors design optimal sampling strategies which are better than existing strategies. In addition, some of the proposed strategies cannot be implemented in practice, thus raising the issue on how they modify the strategies to be implementable in the experiments. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A ", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667482706728}, {"id": "kKngt5h7Kqu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6402/Reviewer_N8Mo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper considered the problem of client sampling in federated learning (FL) to improve the convergence speed of FL training. The authors proposed a new client sampling scheme called DELTA, which is unbiased and able to sample more diverse clients that carry valuable information for global model updates. The authors conducted a theoretical convergence rate performance analysis and verified the theoretical convergence performance of their proposed algorithm through simulation experiments.", "review_text": "This paper studied the problem of client sampling in FL. The authors proposed a new client sampling algorithm called DELTA, which could achieve both unbiasedness and capture diverse client information. The authors provided rigorous theoretical performance analysis and also provided new insights for FedIS, which is also a new contribution. However, the authors didn't provide any theoretical performance analysis for DELTA with stochastic gradient approximation. Also, some experimental results on convergence speed comparisons may be unfair.", "strengths": "Strengths:\n1. The client sampling problem in FL is a timely and important problem in FL.\n2. The proposed DELTA client sampling scheme captures diverse clients similar to the federated importance sampling (FedIS) scheme while offering unbiased client sampling performance.\n3. The new and tighter convergence results for FedIS contribute to new understandings of FedIS.\n\nWeaknesses:\n1. The bounded stochastic gradient assumption in Assumption 4 is a bit restrictive and no longer needed in many state-of-the-art FL algorithms' convergence analyses. It may be interesting to see whether this assumption can be relaxed for the DELTA client sampling scheme.\n\n2. The sampling probability $p_i^t$ in Eq. (11) requires full gradient evaluations, which are difficult to implement in practice (also admitted by the authors). To address this challenge, the authors proposed the use of stochastic gradients to approximate the computations. However, the authors didn't theoretically analyze the impacts of the approximation errors on the convergence rate performance, which is somewhat disappointing.\n\n3. The experiment comparisons in Section 5 may be unfair. In the comparisons between DELTA, FedAvg, and FedIS, the authors only compared the convergence speeds in terms of iterations. However, the proposed DELTA method requires rather complicated calculations of $p_i^t$ in Eq. (11) compared to simple uniform sampling in FedAvg and relatively straightforward calculations of $p_i^t$ in FedIS. That is, the per-iteration complexities of these methods are quite different. Thus, it may be better to also compare the wall-clock convergence time between these methods.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper considered the problem of client sampling in federated learning (FL) to improve the convergence speed of FL training. The authors proposed a new client sampling scheme called DELTA, which is unbiased and able to sample more diverse clients that carry valuable information for global model updates. The authors conducted a theoretical convergence rate performance analysis and verified the theoretical convergence performance of their proposed algorithm through simulation experiments.", "strength_and_weaknesses": "Strengths:\n1. The client sampling problem in FL is a timely and important problem in FL.\n2. The proposed DELTA client sampling scheme captures diverse clients similar to the federated importance sampling (FedIS) scheme while offering unbiased client sampling performance.\n3. The new and tighter convergence results for FedIS contribute to new understandings of FedIS.\n\nWeaknesses:\n1. The bounded stochastic gradient assumption in Assumption 4 is a bit restrictive and no longer needed in many state-of-the-art FL algorithms' convergence analyses. It may be interesting to see whether this assumption can be relaxed for the DELTA client sampling scheme.\n\n2. The sampling probability $p_i^t$ in Eq. (11) requires full gradient evaluations, which are difficult to implement in practice (also admitted by the authors). To address this challenge, the authors proposed the use of stochastic gradients to approximate the computations. However, the authors didn't theoretically analyze the impacts of the approximation errors on the convergence rate performance, which is somewhat disappointing.\n\n3. The experiment comparisons in Section 5 may be unfair. In the comparisons between DELTA, FedAvg, and FedIS, the authors only compared the convergence speeds in terms of iterations. However, the proposed DELTA method requires rather complicated calculations of $p_i^t$ in Eq. (11) compared to simple uniform sampling in FedAvg and relatively straightforward calculations of $p_i^t$ in FedIS. That is, the per-iteration complexities of these methods are quite different. Thus, it may be better to also compare the wall-clock convergence time between these methods.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written in general. However, the paper could have been organized in a better way. For example, the DELTA algorithm should be placed in much earlier sections. In the current form of this paper, the DELTA algorithm appears rather late, which left quite a few notations undefined (e.g., $\\eta$ and $\\eta_L$) and created some difficulty/confusion in following the paper. The proposed DELTA algorithm is novel. The reproducibility of this paper is good.", "summary_of_the_review": "This paper studied the problem of client sampling in FL. The authors proposed a new client sampling algorithm called DELTA, which could achieve both unbiasedness and capture diverse client information. The authors provided rigorous theoretical performance analysis and also provided new insights for FedIS, which is also a new contribution. However, the authors didn't provide any theoretical performance analysis for DELTA with stochastic gradient approximation. Also, some experimental results on convergence speed comparisons may be unfair.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667293203497}, {"id": "mzB15s0PB_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6402/Reviewer_xCAd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors explore the use of importance sampling (IS) its extensions for selecting clients in FL, where IS is based off gradient norms. They first provide a convergence analysis for standard IS sampling (sampling proportional to size of gradients). They then propose an alternative IS sampling which captures both gradient magnitude, as well as gradient diversity (the difference between the client's gradient and the global objective function gradient). They provide a convergence analysis for this sampling scheme and compare the scheme with the original FedIS.", "review_text": "This paper extends the analysis of IS for FL, and proposes a principled extension which (at least partially) solves a well known diversity issue that is common to IS in general (though it's certainly even more pronounced in FL). The clear theoretical analysis is a valuable contribution to the field of client selection for FL. ", "strengths": "Strengths\n- Strong theoretical analysis. \n- Theory is interpretable and illuminating about the nature of the problem.\n\nWeaknesses\n- Experimental results are somewhat weak, with only a marginal improvement over FedIS in experiments on real datasets. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors explore the use of importance sampling (IS) its extensions for selecting clients in FL, where IS is based off gradient norms. They first provide a convergence analysis for standard IS sampling (sampling proportional to size of gradients). They then propose an alternative IS sampling which captures both gradient magnitude, as well as gradient diversity (the difference between the client's gradient and the global objective function gradient). They provide a convergence analysis for this sampling scheme and compare the scheme with the original FedIS.", "strength_and_weaknesses": "Strengths\n- Strong theoretical analysis. \n- Theory is interpretable and illuminating about the nature of the problem.\n\nWeaknesses\n- Experimental results are somewhat weak, with only a marginal improvement over FedIS in experiments on real datasets. ", "clarity,_quality,_novelty_and_reproducibility": "The paper is very well written, which is impressive for a theoretically focused paper. \nExperiments are well described and would not be too challenging to reproduce.\n", "summary_of_the_review": "This paper extends the analysis of IS for FL, and proposes a principled extension which (at least partially) solves a well known diversity issue that is common to IS in general (though it's certainly even more pronounced in FL). The clear theoretical analysis is a valuable contribution to the field of client selection for FL. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666282770228}], "openreview_url": "https://openreview.net/forum?id=CcXTudu9bvu", "arxiv_id": null, "paper_pdf": "papers/CcXTudu9bvu.pdf", "paper_pdf_sha256": "34866af809800423b45fa26f7caf56dd6d8c0a609047519e44b9cd4f1125e0f7", "paper_pdf_bytes": 2641510, "paper_pdf_source": "openreview", "code_url": "https://github.com/T-Lab-CUHKSZ/DELTA_FL", "code_repository": "T-Lab-CUHKSZ/DELTA_FL", "code_commit": "bc8e3a7aed6691ea6a954837656437721edf81e7", "code_archive": "repos/CcXTudu9bvu.zip", "code_archive_sha256": "c546d3d1e18990a27aaed08d18faa60cdc94d42187f2c802c8813801240b91c3", "code_archive_bytes": 111062, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 47, "github_languages": {"Python": 325235}, "github_archived": false, "github_pushed_at": "2024-04-02T05:14:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/delta-diverse-client-sampling-for-fasting"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3GHHpYrYils", "year": 2022, "status": "rejected", "title": "On Anytime Learning at Macroscale", "authors": ["Lucas Caccia", "Jing Xu", "Myle Ott", "MarcAurelio Ranzato", "Ludovic Denoyer"], "authorids": ["~Lucas_Caccia1", "~Jing_Xu5", "~Myle_Ott1", "~MarcAurelio_Ranzato1", "~Ludovic_Denoyer1"], "authors_source": "OpenReview API", "abstract": " Classical machine learning frameworks assume access to a possibly large dataset in order to train a predictive model. In many practical applications however, data does not arrive all at once, but in batches over time. This creates a natural trade-off between accuracy of a model and time to obtain such a model. A greedy predictor could produce non-trivial predictions by immediately training on batches as soon as these become available but, it may also make sub-optimal use of future data. On the other hand, a tardy predictor could wait for a long time to aggregate several batches into a larger dataset, but ultimately deliver a much better performance.  In this work, we consider such a streaming learning setting,  which we dub {\\em anytime learning at macroscale} (ALMA). It is an instance of anytime learning applied not at the level of a single chunk of data, but at the level of the entire  sequence of large batches. We first formalize this learning setting, we then introduce metrics to assess how well learners perform on the given task for a given memory and compute budget, and finally we test about thirty baseline approaches on three standard benchmarks repurposed for anytime learning at macroscale. Our findings indicate that no model strikes the best trade-off across the board. While replay-based methods attain the lowest error rate, they also incur in a 5 to 10 times increase of compute. Approaches that grow capacity over time do offer better scaling in terms of training flops, but they also underperform simpler ensembling methods in terms of error rate. Overall, ALMA offers both a good abstraction of the typical learning setting faced everyday by practitioners, and a set of unsolved modeling problems for those interested in efficient learning of dynamic models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "G2B3MoXj0Xc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper766/Reviewer_D3ZG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper introduces a novel setup called Anytime Learning at Macroscale. In this setup the learner receives the examples as a sequence of large batches, and is required to output a model after processing each batch. This model is used to give prediction for the next batch. The overall performance is then sum of the average losses on the individual batches.", "review_text": "My main concern is with the motivation of the setup. In particular, it does not seem to be very different from online learning. The paper does discuss the differences between the two setup, however the arguments are a bit superficial. In particular, it is claimed that in online learning the examples are streamed one at the time as opposed to receiving them in large batches - but this does not seem to make the online learning methods completely handicapped. Randomized algorithms (such as Thompson-sampling based methods) should perform reasonably. (For each example in batch i, apply the model obtained after processing all the examples in batches 1,2,..., i-1.) It would be great if the authors could elaborate on this.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces a novel setup called Anytime Learning at Macroscale. In this setup the learner receives the examples as a sequence of large batches, and is required to output a model after processing each batch. This model is used to give prediction for the next batch. The overall performance is then sum of the average losses on the individual batches.", "main_review": "My main concern is with the motivation of the setup. In particular, it does not seem to be very different from online learning. The paper does discuss the differences between the two setup, however the arguments are a bit superficial. In particular, it is claimed that in online learning the examples are streamed one at the time as opposed to receiving them in large batches - but this does not seem to make the online learning methods completely handicapped. Randomized algorithms (such as Thompson-sampling based methods) should perform reasonably. (For each example in batch i, apply the model obtained after processing all the examples in batches 1,2,..., i-1.) It would be great if the authors could elaborate on this.", "summary_of_the_review": "Concern with the motivation: the setup is not that different from online learning after all.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635921974769}, {"id": "ILXixoFhtem", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper766/Reviewer_4ezA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors describe a framework to perform empirical evaluation of an anytime learning setting where data is available in a streaming minibatch fashion. With a primary aim to measure performance of a classifier across variety of practical settings of such streaming data to not only achieve high accuracy, but also provide non-trivial prediction anytime using limited computational resources. Using multiple benchmark datasets, the paper concludes that methods with intermediate parameter updates are better on the accuracy to computational efficiency tradeoff, and larger models generalize better.", "review_text": "Paper is well written. The authors document the approach they have considered, and the metrics used to evaluate the various experimental settings used. \n\nAs I started to read the paper, the problem setting seemed very similar to the ones used in data stream mining research over the past decade. Though the authors state that the primary differentiator to the stream setting is that the models use what they call as \"meta-batches\" as streams rather than streaming single data instances, I fail to understand any theoretical or empirical difference in the two approaches. There exists multiple popular data stream frameworks such as MOA (Massive Online Analysis) that is used exactly for the problem setting described in the paper. So, the primary contribution of the paper seems to be in extensively evaluating the model complexity and approaches over various data and problem settings. Moreover, majority of the future questions that the authors hope to answer have been studied across various papers (in similar forms). Please refer to Lu, Jie, et al. \"Learning under concept drift: A review.\" IEEE Transactions on Knowledge and Data Engineering 31.12 (2018): 2346-236.\n\nThe second objection of the paper is that the authors seem to use simple datasets from today's standard to derive their conclusion. Though the conclusions in the paper are fair and not surprising, a more complex set of datasets may provide a stronger result. Furthermore, it is important to note that there are other factors that influence the classifier performance, more than the batch size and data size available for training. The data itself may be imbalanced, non standard etc. So, by using more datasets, these issues can potentially be elevated.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors describe a framework to perform empirical evaluation of an anytime learning setting where data is available in a streaming minibatch fashion. With a primary aim to measure performance of a classifier across variety of practical settings of such streaming data to not only achieve high accuracy, but also provide non-trivial prediction anytime using limited computational resources. Using multiple benchmark datasets, the paper concludes that methods with intermediate parameter updates are better on the accuracy to computational efficiency tradeoff, and larger models generalize better.", "main_review": "Paper is well written. The authors document the approach they have considered, and the metrics used to evaluate the various experimental settings used. \n\nAs I started to read the paper, the problem setting seemed very similar to the ones used in data stream mining research over the past decade. Though the authors state that the primary differentiator to the stream setting is that the models use what they call as \"meta-batches\" as streams rather than streaming single data instances, I fail to understand any theoretical or empirical difference in the two approaches. There exists multiple popular data stream frameworks such as MOA (Massive Online Analysis) that is used exactly for the problem setting described in the paper. So, the primary contribution of the paper seems to be in extensively evaluating the model complexity and approaches over various data and problem settings. Moreover, majority of the future questions that the authors hope to answer have been studied across various papers (in similar forms). Please refer to Lu, Jie, et al. \"Learning under concept drift: A review.\" IEEE Transactions on Knowledge and Data Engineering 31.12 (2018): 2346-236.\n\nThe second objection of the paper is that the authors seem to use simple datasets from today's standard to derive their conclusion. Though the conclusions in the paper are fair and not surprising, a more complex set of datasets may provide a stronger result. Furthermore, it is important to note that there are other factors that influence the classifier performance, more than the batch size and data size available for training. The data itself may be imbalanced, non standard etc. So, by using more datasets, these issues can potentially be elevated.", "summary_of_the_review": "In summary, the original contribution is not very clear. The authors have ignored to discuss comparisons with a branch of data stream mining reseach that has provided similar conclusion. And the empirical evaluation is on simple datasets, and may need further evidential support from more complex datasets.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635825253110}, {"id": "Y8gM87zIKJa", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper766/Reviewer_KMi2"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors consider a batch learning problem, in which large batches of data arrive in series. They explore the performance of several types of algorithms in terms of their computational cost, model size, and error rate. ", "review_text": "The problem presented by the authors is relevant to applied ML/AI problems, which are always a work in progress. Further, improving the efficiency of learning is desirable. So, the problem is reasonably well motivated.\n\n1. I'm skeptical about the value of the cumulative error rate. From a practical point of view a data engineer might be concerned with the questions; how good is the model I have now? What would be impact of further data collection? If one has collected a set of batches, the performance of prior models is not terribly relevant. \n\n2. The non-iid nature of the data is mentioned, and it is mentioned that cross-validation is carried out only on the current batch. These concepts could be explored more thoroughly. What are the issues with evaluation in this setting? Should I hold out a portion of every batch to use for evaluation? What other evaluation strategies are possible? What are their strengths and weaknesses?\n\n2. In continuous streaming settings, there may be distribution shift over time. That is not addressed in this work. \n\n3. The authors make a big point about the scale of the problem. Obviously this makes naive approaches less appealing. But how does the problem scale? What really separates (if anything), the macroscale problem from more mundane sized problems? Are there underlying scaling laws at work that cause shifts the performance of each learning strategy?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors consider a batch learning problem, in which large batches of data arrive in series. They explore the performance of several types of algorithms in terms of their computational cost, model size, and error rate. ", "main_review": "The problem presented by the authors is relevant to applied ML/AI problems, which are always a work in progress. Further, improving the efficiency of learning is desirable. So, the problem is reasonably well motivated.\n\n1. I'm skeptical about the value of the cumulative error rate. From a practical point of view a data engineer might be concerned with the questions; how good is the model I have now? What would be impact of further data collection? If one has collected a set of batches, the performance of prior models is not terribly relevant. \n\n2. The non-iid nature of the data is mentioned, and it is mentioned that cross-validation is carried out only on the current batch. These concepts could be explored more thoroughly. What are the issues with evaluation in this setting? Should I hold out a portion of every batch to use for evaluation? What other evaluation strategies are possible? What are their strengths and weaknesses?\n\n2. In continuous streaming settings, there may be distribution shift over time. That is not addressed in this work. \n\n3. The authors make a big point about the scale of the problem. Obviously this makes naive approaches less appealing. But how does the problem scale? What really separates (if anything), the macroscale problem from more mundane sized problems? Are there underlying scaling laws at work that cause shifts the performance of each learning strategy?", "summary_of_the_review": "The paper addresses a real need, but more insight is needed.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635175808452}, {"id": "XUSzwk_m8Yo", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper766/Reviewer_x4p7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Summary: This paper proposes anytime learning at macroscale (ALMA), which is anytime learning under the assumption that data is observed as a sequence of large batches. This paper introduces metrics that can be used to access the error rate, memory, and compute throughput the entire learning process. They evaluate multiple learning models on different datasets in the ALMA setting. They observe that methods that update parameters at a moderate rate tend to yield a better tradeoff, while bigger models tend to generalize better.", "review_text": "The problem setting of anytime learning at macroscale is interesting and novel to me. How to efficiently learn data in a streaming fashion is a practical challenge. The proposed learning setting targets the level of the entire sequence of large datasets. \n\nAlthough the method overall is valuable and interesting, the paper is poorly organized and thus hard to understand. It is difficult to find some critical details or notation definitions that are related but separated apart in the paper. The method lacks an integrated and principal formulation from which the techniques are derived. Though they show some theoretical results, it is hard to relate them to the objective and the algorithm steps explicitly and tightly. The contribution is unclear. Here are some detailed comments:\n\n- What does \"organically generated\" mean?\n- \" both training ... and finetuning ... are not satisfying\" should use \"neither...nor...\"\n- What does \"constrained capacity\" ?\n- What is the definition of macroscale?\n- Unclear the difference between ALMA and other learning frameworks.\n- Related work compares ALMA to lots of different prior work, but it is poorly organized and I am not sure why those prior work should be considered as comparisons.\n- It is suspicious to say that \"extensions to regression and unsupervised learning (where y is missing) are trivial\".\n- In fixed architecture, why \"A potential drawback of Ens is that evaluation and training are inconsistent\"?\n- In Growing Mixture of Experts, \"Compared to Ens, MoE has exponentially many more components\". I am not sure where \"exponentially\" comes from.\n\n\nQuality: The submission is technically sound. The claims in the contribution are supported by empirically results. It is a complete piece of work.\n\nClarity: The experimental details are also very specific, such that reproducing the results should be possible. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Summary: This paper proposes anytime learning at macroscale (ALMA), which is anytime learning under the assumption that data is observed as a sequence of large batches. This paper introduces metrics that can be used to access the error rate, memory, and compute throughput the entire learning process. They evaluate multiple learning models on different datasets in the ALMA setting. They observe that methods that update parameters at a moderate rate tend to yield a better tradeoff, while bigger models tend to generalize better.", "main_review": "The problem setting of anytime learning at macroscale is interesting and novel to me. How to efficiently learn data in a streaming fashion is a practical challenge. The proposed learning setting targets the level of the entire sequence of large datasets. \n\nAlthough the method overall is valuable and interesting, the paper is poorly organized and thus hard to understand. It is difficult to find some critical details or notation definitions that are related but separated apart in the paper. The method lacks an integrated and principal formulation from which the techniques are derived. Though they show some theoretical results, it is hard to relate them to the objective and the algorithm steps explicitly and tightly. The contribution is unclear. Here are some detailed comments:\n\n- What does \"organically generated\" mean?\n- \" both training ... and finetuning ... are not satisfying\" should use \"neither...nor...\"\n- What does \"constrained capacity\" ?\n- What is the definition of macroscale?\n- Unclear the difference between ALMA and other learning frameworks.\n- Related work compares ALMA to lots of different prior work, but it is poorly organized and I am not sure why those prior work should be considered as comparisons.\n- It is suspicious to say that \"extensions to regression and unsupervised learning (where y is missing) are trivial\".\n- In fixed architecture, why \"A potential drawback of Ens is that evaluation and training are inconsistent\"?\n- In Growing Mixture of Experts, \"Compared to Ens, MoE has exponentially many more components\". I am not sure where \"exponentially\" comes from.\n\n\nQuality: The submission is technically sound. The claims in the contribution are supported by empirically results. It is a complete piece of work.\n\nClarity: The experimental details are also very specific, such that reproducing the results should be possible. ", "summary_of_the_review": "This paper proposes a novel learning setting called anytime learning at macroscale (ALMA). The proposed idea is simple and technically sound. Extensive evaluations are conducted. However, the contribution is unclear, and the presentation needs improvement.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635112047092}], "openreview_url": "https://openreview.net/forum?id=3GHHpYrYils", "arxiv_id": "2106.09563", "paper_pdf": "papers/3GHHpYrYils.pdf", "paper_pdf_sha256": "582db19e47086a146cc981abd12dcfafe2453b697d086d500789939581ed1e15", "paper_pdf_bytes": 454568, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/alma", "code_repository": "facebookresearch/alma", "code_commit": "02a6d0e76fedc6af315dd906698153edabbaea5a", "code_archive": "repos/3GHHpYrYils.zip", "code_archive_sha256": "d94e19c3f6d30199b213f3a04d1de10ede9f501117a7817f064cbb9114e96069", "code_archive_bytes": 84934, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 60, "github_languages": {"Python": 188093}, "github_archived": true, "github_pushed_at": "2021-11-24T20:57:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-anytime-learning-at-macroscale"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-J9xYzP2HD", "year": 2021, "status": "rejected", "title": "Chameleon: Learning Model Initializations Across Tasks With Different Schemas", "authors": ["Lukas Brinkmeyer", "Rafael Rego Drumond", "Randolf Scholz", "Josif Grabocka", "Lars Schmidt-Thieme"], "authorids": ["~Lukas_Brinkmeyer1", "radrumond@ismll.uni-hildesheim.de", "scholz@ismll.uni-hildesheim.de", "~Josif_Grabocka1", "~Lars_Schmidt-Thieme1"], "authors_source": "OpenReview API", "abstract": "Parametric models, and particularly neural networks, require weight initialization as a starting point for gradient-based optimization. Recent work shows that an initial parameter set can be learned from a population of supervised learning tasks that enables a fast convergence for unseen tasks even when only a handful of instances is available (model-agnostic meta-learning). \nCurrently, methods for learning model initializations are limited to a population of tasks sharing the same schema, i.e., the same number, order, type, and semantics of predictor and target variables.\nIn this paper, we address the problem of meta-learning weight initialization across tasks with different schemas, for example, if the number of predictors varies across tasks, while they still share some variables. We propose Chameleon, a model that learns to align different predictor schemas to a common representation. \nIn experiments on 23 datasets of the OpenML-CC18 benchmark, we show that Chameleon can successfully learn parameter initializations across tasks with different schemas, presenting, to the best of our knowledge, the first cross-dataset few-shot classification approach for unstructured data.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "YR_o8utp3K6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper861/AnonReviewer5"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper aims to perform meta-learning across tasks that have different input data types by learning separate task-specific encoders, and then aligning the features produced by these encoders before making predictions.\n\nPros:\nSharing information across tasks with different input types is a relevant problem\nCons:\nPrecise problem statement and method very unclear\nExperiments are only on toy datasets\n\nDetailed Comments:\n\nIt is not clear from the abstract / introduction what is meant by “schema.” From the abstract: “for example, if the number of predictors varies across tasks, while they still share some variables.” Does this refer to the number of classes in a few-shot problem? What variables are shared? Classes, or input features? Later in the intro: “training a single model across different tasks is only feasible if all tasks share the same schema, meaning that all instances share one set of features in identical order.” These definitions of schema do not seem to be the same. Schema also does not seem to be defined in Section 3. At the beginning of that section it says, “every task has to share the same schema of common size K” which seems to indicate “schema” is the number of features and then a few lines later, “ tasks with varying input schema and feature length F” which seems to indicate “schema” is *not* the number of features.\n\n\nIn the related work section, few-shot learning did not begin in 2017 as might be suggested by the citations. It would be good to recognize the earlier works in this area, such as \nFei-Fei, L. et al. A bayesian approach to unsupervised one-shot learning of object categories. 2003\nA Bayesian framework for concept learning. PhD thesis, Massachusetts Institute of Technology, 1999.\nFor few-shot learning with deep learning, Matching Networks should arguably be cited: Vinyals, Oriol, et al. Matching networks for one shot learning. 2016.\nThe original MAML paper actually proposed the first-order version of MAML, Nichol et al. was not the first to propose this.\n\nI don’t understand how the method works when the features are learned and not given. For example, the encoder for EMNIST-Digits produces 32 features, while the encoder for EMNIST-Letters produces 64 features. If the meta-training tasks are drawn from only EMNIST-Digits, then how can the “re-ordering” matrix be learned from EMNIST-Digits such that it can re-order features from EMNIST-Letters? At the most basic level, based on Figure 2, the matrix \\Pi would have to have different dimensionality for each dataset. Even if they were the same dimensionality, how is the feature ordering supervision performed in this case?\n\nIn the “main results”, if you sub-sample features, how do you know that the sub-sampled features have enough information to perform the classification task? \n\nIt would be helpful to have an experiment on a less-toy dataset, both to demonstrate that the problem of “mis-aligned features” exists in more complex data, and that the method can address it. \n\nOverall, this paper is extremely confusing. I do not understand the problem statement or how the method is trained in the learned feature case. In my view, the clarity of this paper needs to be significantly improved to consider acceptance. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Confusing presentation of problem statement and method", "review": "Summary:\nThis paper aims to perform meta-learning across tasks that have different input data types by learning separate task-specific encoders, and then aligning the features produced by these encoders before making predictions.\n\nPros:\nSharing information across tasks with different input types is a relevant problem\nCons:\nPrecise problem statement and method very unclear\nExperiments are only on toy datasets\n\nDetailed Comments:\n\nIt is not clear from the abstract / introduction what is meant by “schema.” From the abstract: “for example, if the number of predictors varies across tasks, while they still share some variables.” Does this refer to the number of classes in a few-shot problem? What variables are shared? Classes, or input features? Later in the intro: “training a single model across different tasks is only feasible if all tasks share the same schema, meaning that all instances share one set of features in identical order.” These definitions of schema do not seem to be the same. Schema also does not seem to be defined in Section 3. At the beginning of that section it says, “every task has to share the same schema of common size K” which seems to indicate “schema” is the number of features and then a few lines later, “ tasks with varying input schema and feature length F” which seems to indicate “schema” is *not* the number of features.\n\n\nIn the related work section, few-shot learning did not begin in 2017 as might be suggested by the citations. It would be good to recognize the earlier works in this area, such as \nFei-Fei, L. et al. A bayesian approach to unsupervised one-shot learning of object categories. 2003\nA Bayesian framework for concept learning. PhD thesis, Massachusetts Institute of Technology, 1999.\nFor few-shot learning with deep learning, Matching Networks should arguably be cited: Vinyals, Oriol, et al. Matching networks for one shot learning. 2016.\nThe original MAML paper actually proposed the first-order version of MAML, Nichol et al. was not the first to propose this.\n\nI don’t understand how the method works when the features are learned and not given. For example, the encoder for EMNIST-Digits produces 32 features, while the encoder for EMNIST-Letters produces 64 features. If the meta-training tasks are drawn from only EMNIST-Digits, then how can the “re-ordering” matrix be learned from EMNIST-Digits such that it can re-order features from EMNIST-Letters? At the most basic level, based on Figure 2, the matrix \\Pi would have to have different dimensionality for each dataset. Even if they were the same dimensionality, how is the feature ordering supervision performed in this case?\n\nIn the “main results”, if you sub-sample features, how do you know that the sub-sampled features have enough information to perform the classification task? \n\nIt would be helpful to have an experiment on a less-toy dataset, both to demonstrate that the problem of “mis-aligned features” exists in more complex data, and that the method can address it. \n\nOverall, this paper is extremely confusing. I do not understand the problem statement or how the method is trained in the learned feature case. In my view, the clarity of this paper needs to be significantly improved to consider acceptance. \n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604597212721}, {"id": "vCAikM3KaKF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper861/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n--------------\nThe paper proposes a trainable way to re-order or recover the ordering of features from sets of examples, and use it as a way to build a common feature space (or embedding) for a neural net, the (initial) parameters of which can be trained by Reptile.\nExperiments show that such initial parameters enable faster training (inside of an episode) than untrained weights.\n\nPros\n------\n- The paper shows it is possible to recover information about the identity of coordinates in the input space, through a learned transformation, on several unstructured datasets. The similarity between such representations of individual coordinates can help identify similar features, either in a given dataset or across datasets.\n\n\nCons\n--------\nThe paper is overall really hard to follow, statements are often confusing or misleading. For instance:\n- The introduction suggests a multi-modal learning paradigm, where different tasks could have access to data in different input spaces, some of them common. However, the paper then seems to consider individual coordinates in the input space only, and focuses on mapping shuffled subsets of these coordinates back to their initial position.\n- There is confusion about the \"tasks\", which sometimes correspond to one of the OpenML datasets, and sometimes to individual few-shot episodes from one of these datasets.\n- Concepts like \"schema\" and \"predictors\" are never properly introduced or defined.\n- The description of the \"chameleon\" (alignment) component mentions \"order-invariant\" and \"permutation invariant\" several times, but it is quite unclear whether it refers to the the order of the examples within the data set (or episode) or the order in which the features are represented.\n\nThe paper uses few-shot learning vocabulary and techniques, including Reptile, but the methodology seems completely different from the few-shot learning literature. In particular:\n- There does not appear to be a split between meta-training and meta-test classes within a dataset, or meta-training datasets and meta-testing ones, except for the EMNIST experiment. Even then, the pre-training of the \"chameleon\" alignment module seems to involve using examples of the meta-test classes.\n- The reported evaluation metric is really unusual: they report the improvement (and sometimes accuracy) after 3 steps of gradient descent from within an episode, which is somewhat related to the quality of the meta-learned weights, but no other metric that would be comparable to existing literature, which makes it especially hard to assess the results.\n\nThe principle of the alignment module seems similar to (soft) attention mechanisms, in that there is a softmax trained to highlight which parts of an input vector should be emphasized (or selected) at a given point in the processing (here, in the aligned feature space). However, the literature on attention is not reviewed. \n\nMany design choices are not addressed clearly, neither in how they were made, or the impact of these choices, especially regarding the architecture of the alignment module:\n- It is a linear transformation (before the softmax), though parameterized by 3 matrices. An alternative would have been a 3-layer neural net, similar to attention networks.\n- The parameterization of the first matrix makes the number of parameters depend on N, the number of examples in a given task. This could be quite limiting to be restrained to tasks of exactly N examples, especially if both the support (mini-train) and query (mini-test or valid) parts of an episode need to have exactly N examples.\n- There is also no discussion of the  value or impact of or K, the size of the chosen embedding space).\n\nRecommendation\n--------------------------\nI recommend to reject this submission.\n\nArguments\n------------------\nThe main idea in the paper, learning alignments of various input spaces into a common embedding space through an attention mechanism, has merit and may  work reasonably.\nHowever, both the algorithm and the experimental set up are described in a quite confused way, and not well justified or grounded. The reported results are not comparable with few-shot learning literature, nor multi-modal training or feature imputation, and do not make a convincing case. \n\nQuestions\n---------------\nAs I understand it, the \"Chameleon\" architecture itself simply consists in 3 matrix multiplications (Nx8, 8x16, 16xK), which would be equivalent to the length-1 1D convolutions, is that correct? It may be more straightforward to explain that way, as $enc(X) = X M_1 M_2 M_3 X^T$.\nAlso, should the 2nd and 3rd convolutions be labeled \"8x16x1\" and \"16xKx1\" respectively? As far as I can tell, only the first Conv1D should have a dependency on N.\n\nAdditional feedback\n---------------------------\nIn Figure 2, the \"reshape\" operation should be \"transpose\" instead.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The idea of learning to re-align input spaces in a common feature space has merit, but the experimental protocol is unusual and results not convincing", "review": "Summary\n--------------\nThe paper proposes a trainable way to re-order or recover the ordering of features from sets of examples, and use it as a way to build a common feature space (or embedding) for a neural net, the (initial) parameters of which can be trained by Reptile.\nExperiments show that such initial parameters enable faster training (inside of an episode) than untrained weights.\n\nPros\n------\n- The paper shows it is possible to recover information about the identity of coordinates in the input space, through a learned transformation, on several unstructured datasets. The similarity between such representations of individual coordinates can help identify similar features, either in a given dataset or across datasets.\n\n\nCons\n--------\nThe paper is overall really hard to follow, statements are often confusing or misleading. For instance:\n- The introduction suggests a multi-modal learning paradigm, where different tasks could have access to data in different input spaces, some of them common. However, the paper then seems to consider individual coordinates in the input space only, and focuses on mapping shuffled subsets of these coordinates back to their initial position.\n- There is confusion about the \"tasks\", which sometimes correspond to one of the OpenML datasets, and sometimes to individual few-shot episodes from one of these datasets.\n- Concepts like \"schema\" and \"predictors\" are never properly introduced or defined.\n- The description of the \"chameleon\" (alignment) component mentions \"order-invariant\" and \"permutation invariant\" several times, but it is quite unclear whether it refers to the the order of the examples within the data set (or episode) or the order in which the features are represented.\n\nThe paper uses few-shot learning vocabulary and techniques, including Reptile, but the methodology seems completely different from the few-shot learning literature. In particular:\n- There does not appear to be a split between meta-training and meta-test classes within a dataset, or meta-training datasets and meta-testing ones, except for the EMNIST experiment. Even then, the pre-training of the \"chameleon\" alignment module seems to involve using examples of the meta-test classes.\n- The reported evaluation metric is really unusual: they report the improvement (and sometimes accuracy) after 3 steps of gradient descent from within an episode, which is somewhat related to the quality of the meta-learned weights, but no other metric that would be comparable to existing literature, which makes it especially hard to assess the results.\n\nThe principle of the alignment module seems similar to (soft) attention mechanisms, in that there is a softmax trained to highlight which parts of an input vector should be emphasized (or selected) at a given point in the processing (here, in the aligned feature space). However, the literature on attention is not reviewed. \n\nMany design choices are not addressed clearly, neither in how they were made, or the impact of these choices, especially regarding the architecture of the alignment module:\n- It is a linear transformation (before the softmax), though parameterized by 3 matrices. An alternative would have been a 3-layer neural net, similar to attention networks.\n- The parameterization of the first matrix makes the number of parameters depend on N, the number of examples in a given task. This could be quite limiting to be restrained to tasks of exactly N examples, especially if both the support (mini-train) and query (mini-test or valid) parts of an episode need to have exactly N examples.\n- There is also no discussion of the  value or impact of or K, the size of the chosen embedding space).\n\nRecommendation\n--------------------------\nI recommend to reject this submission.\n\nArguments\n------------------\nThe main idea in the paper, learning alignments of various input spaces into a common embedding space through an attention mechanism, has merit and may  work reasonably.\nHowever, both the algorithm and the experimental set up are described in a quite confused way, and not well justified or grounded. The reported results are not comparable with few-shot learning literature, nor multi-modal training or feature imputation, and do not make a convincing case. \n\nQuestions\n---------------\nAs I understand it, the \"Chameleon\" architecture itself simply consists in 3 matrix multiplications (Nx8, 8x16, 16xK), which would be equivalent to the length-1 1D convolutions, is that correct? It may be more straightforward to explain that way, as $enc(X) = X M_1 M_2 M_3 X^T$.\nAlso, should the 2nd and 3rd convolutions be labeled \"8x16x1\" and \"16xKx1\" respectively? As far as I can tell, only the first Conv1D should have a dependency on N.\n\nAdditional feedback\n---------------------------\nIn Figure 2, the \"reshape\" operation should be \"transpose\" instead.\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604050362352}, {"id": "mP5SR2rcqQC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper861/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Summary and contributions\n    - In this work, the authors tried to solve the problem of ``heterogeneous'' meta-learning where each task resides in a different feature space from the other tasks. They introduced a feature transformation or re-ordering matrix to align the features. While I agree with the authors that this problem is of significance in the meta-learning community, the solution in this work, depending on the ground-truth of re-ordering matrix, is trivial and impractical. \n\n- Strengths: \n    - The problem investigated in this paper, i.e., meta-learning tasks in heterogeneous feature spaces, is important to the field of meta-learning. \n    - The paper is well written and easy to follow. \n\n- Weaknesses:\n    - The primary concern about this paper is its technical contribution, being limited and impractical. To align tasks in incommensurable feature spaces, projecting them into a common feature space has been a common practice. Please kindly see related works on heterogeneous transfer learning. The major challenge lies in the supervision needed to train the alignment matrix or function. The ground-truth feature alignment matrix is almost impractical to collect, if the dimension of features is super large and we have no knowledge of the semantic correspondence between two features from two tasks. \n    - The empirical results are also not convincing.\n         - Why is only Glorot initialization compared in Figure 3? What has been widely adopted is some better initialization strategies, including (He initialization). \n        - From both Figure 3 and Figure 4, and also the results in Appendix C, I see little improvement of the proposed over Frozen. This means that most benefits of the feature alignment come from the supervised training part where a ground-truth alignment matrix is required to train $\\Phi$, while the matrix is even infeasible to have in practical settings. \n        - In Line 6 of the section \"Ablations\", the authors mentioned that features 2 and 3 are showing a strong correlation, but I cannot see why in Figure 6. Maybe it is features 2 and 4?\n\n- Minor:\n    Line 2 in Section 4: Equation (9) does not exist…", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The technical contribution is limited and impractical.", "review": "- Summary and contributions\n    - In this work, the authors tried to solve the problem of ``heterogeneous'' meta-learning where each task resides in a different feature space from the other tasks. They introduced a feature transformation or re-ordering matrix to align the features. While I agree with the authors that this problem is of significance in the meta-learning community, the solution in this work, depending on the ground-truth of re-ordering matrix, is trivial and impractical. \n\n- Strengths: \n    - The problem investigated in this paper, i.e., meta-learning tasks in heterogeneous feature spaces, is important to the field of meta-learning. \n    - The paper is well written and easy to follow. \n\n- Weaknesses:\n    - The primary concern about this paper is its technical contribution, being limited and impractical. To align tasks in incommensurable feature spaces, projecting them into a common feature space has been a common practice. Please kindly see related works on heterogeneous transfer learning. The major challenge lies in the supervision needed to train the alignment matrix or function. The ground-truth feature alignment matrix is almost impractical to collect, if the dimension of features is super large and we have no knowledge of the semantic correspondence between two features from two tasks. \n    - The empirical results are also not convincing.\n         - Why is only Glorot initialization compared in Figure 3? What has been widely adopted is some better initialization strategies, including (He initialization). \n        - From both Figure 3 and Figure 4, and also the results in Appendix C, I see little improvement of the proposed over Frozen. This means that most benefits of the feature alignment come from the supervised training part where a ground-truth alignment matrix is required to train $\\Phi$, while the matrix is even infeasible to have in practical settings. \n        - In Line 6 of the section \"Ablations\", the authors mentioned that features 2 and 3 are showing a strong correlation, but I cannot see why in Figure 6. Maybe it is features 2 and 4?\n\n- Minor:\n    Line 2 in Section 4: Equation (9) does not exist…", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603959458797}, {"id": "S8A1K3_qCUU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper861/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Chameleon: Learning Model Initializations Across Tasks With Different Schemas\n\n  \nThe paper provides an interesting direction in the few-shot classification field. In particular, it proposes a model that learns to align different predictor schemas to a common representation. The paper also demonstrates how current meta-learning approaches can successfully learn a model initialisation across tasks with different schemas as long as they share some variables with respect to their type or semantics.\n\nThe paper takes on an interesting facet of few-shot classification: An encoder model that aligns to different predictor schemas to a common representation. It tackles the problem by using 1D convolution (three of them) to transform the input features to the K-features target space and learning the alignment from the data itself. Comprehensive experiments have been done with quantitative results and analysis, to show the effectiveness of the proposed approach and the results are convincing and the code is provided to determine the reproducibility of the results.\n\nOverall performance is quite good however, it would be a good study to have an analysis of the different datasets as to how balanced/unbalanced they are, how it affects the performance, the nature of the features etc. Also, I would like the author to discuss how suitable/adaptable this approach will be for multi-label tasks and what kind of modifications (if any) are to be made.\n\nThe idea of encoding different predictor schemas to a common representation is quite interesting and comprehensive experiments and supporting ablation study has been made.   \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review", "review": "Chameleon: Learning Model Initializations Across Tasks With Different Schemas\n\n  \nThe paper provides an interesting direction in the few-shot classification field. In particular, it proposes a model that learns to align different predictor schemas to a common representation. The paper also demonstrates how current meta-learning approaches can successfully learn a model initialisation across tasks with different schemas as long as they share some variables with respect to their type or semantics.\n\nThe paper takes on an interesting facet of few-shot classification: An encoder model that aligns to different predictor schemas to a common representation. It tackles the problem by using 1D convolution (three of them) to transform the input features to the K-features target space and learning the alignment from the data itself. Comprehensive experiments have been done with quantitative results and analysis, to show the effectiveness of the proposed approach and the results are convincing and the code is provided to determine the reproducibility of the results.\n\nOverall performance is quite good however, it would be a good study to have an analysis of the different datasets as to how balanced/unbalanced they are, how it affects the performance, the nature of the features etc. Also, I would like the author to discuss how suitable/adaptable this approach will be for multi-label tasks and what kind of modifications (if any) are to be made.\n\nThe idea of encoding different predictor schemas to a common representation is quite interesting and comprehensive experiments and supporting ablation study has been made.   \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603910260268}, {"id": "jYUeZe3aezJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper861/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Previous meta-learning approaches typically focus on tasks that share the same input types, e.g. images.\nThis paper addresses the problem of meta-learning weight initialization across tasks with different types of input features. \nIt proposes Chameleon model that learns to align input features from different tasks by learning a permutation matrix for each task, and shows that Chameleon can successfully learn good initialization.\n\n\nStrength:\n- It identifies and tackles a new important problem in meta-learning: meta-learning on tasks with different input features. \n- The proposed approach is simple but shows improvements over the baseline method.\n\nWeakeness:\n- Supervised training for the permutation matrix is necessary for the model to perform well.\n- Experimental results section can be more detailed. Given that Algorithm 2 is the major part of the method, how is the reordering training procedure constructed? How is the target permutation matrix determined? are there / what are the shared features between different tasks?\n- Would be great if experiments are done on one or two more datasets to strengthen the result.\n\nAdditional Comments:\n- How many features are used? How would the performance change if there are more/fewer features?\n- typo: Equation (9) is mentioned several times\n\nI believe this paper proposed a new interesting problem in meta-learning and provided a simple effective model to address the problem.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official review", "review": "Previous meta-learning approaches typically focus on tasks that share the same input types, e.g. images.\nThis paper addresses the problem of meta-learning weight initialization across tasks with different types of input features. \nIt proposes Chameleon model that learns to align input features from different tasks by learning a permutation matrix for each task, and shows that Chameleon can successfully learn good initialization.\n\n\nStrength:\n- It identifies and tackles a new important problem in meta-learning: meta-learning on tasks with different input features. \n- The proposed approach is simple but shows improvements over the baseline method.\n\nWeakeness:\n- Supervised training for the permutation matrix is necessary for the model to perform well.\n- Experimental results section can be more detailed. Given that Algorithm 2 is the major part of the method, how is the reordering training procedure constructed? How is the target permutation matrix determined? are there / what are the shared features between different tasks?\n- Would be great if experiments are done on one or two more datasets to strengthen the result.\n\nAdditional Comments:\n- How many features are used? How would the performance change if there are more/fewer features?\n- typo: Equation (9) is mentioned several times\n\nI believe this paper proposed a new interesting problem in meta-learning and provided a simple effective model to address the problem.\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603878541187}], "openreview_url": "https://openreview.net/forum?id=-J9xYzP2HD", "arxiv_id": "1909.13576", "paper_pdf": "papers/-J9xYzP2HD.pdf", "paper_pdf_sha256": "5cca826662fa60afd6abbc0be4bba477d3399efff780dcb8380841e022d18272", "paper_pdf_bytes": 632014, "paper_pdf_source": "openreview", "code_url": "https://github.com/radrumond/Chameleon", "code_repository": "radrumond/Chameleon", "code_commit": "e7c2e86c81b7292a7ffb5bea0dc3b0a71f5bb1b8", "code_archive": "repos/-J9xYzP2HD.zip", "code_archive_sha256": "1fd83b1a661cb326c80a8153b3e956fd4a55b7b624624fa146e47fc0a66401a4", "code_archive_bytes": 44459, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 61, "github_languages": {"Python": 117074}, "github_archived": false, "github_pushed_at": "2020-03-20T15:48:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/chameleon-learning-model-initializations"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkgF4kSFPB", "year": 2020, "status": "rejected", "title": "Hallucinative Topological Memory for Zero-Shot Visual Planning", "authors": ["Kara Liu", "Thanard Kurutach", "Pieter Abbeel", "Aviv Tamar"], "authorids": ["karamarieliu@berkeley.edu", "thanard.kurutach@berkeley.edu", "pabbeel@cs.berkeley.edu", "aviv.tamar.mail@gmail.com"], "authors_source": "OpenReview API", "abstract": "In visual planning (VP), an agent learns to plan goal-directed behavior from observations of a dynamical system obtained offline, e.g., images obtained from self-supervised robot interaction. VP algorithms essentially combine data-driven perception and planning, and are important for robotic manipulation and navigation domains, among others. A recent and promising approach to VP is the semi-parametric topological memory (SPTM) method, where image samples are treated as nodes in a graph, and the connectivity in the graph is learned using deep image classification. Thus, the learned graph represents the topological connectivity of the data, and planning can be performed using conventional graph search methods. However, training SPTM necessitates a suitable loss function for the connectivity classifier, which requires non-trivial manual tuning. More importantly, SPTM is constricted in its ability to generalize to changes in the domain, as its graph is constructed from direct observations and thus requires collecting new samples for planning. In this paper, we propose Hallucinative Topological Memory (HTM), which overcomes these shortcomings. In HTM, instead of training a discriminative classifier we train an energy function using contrastive predictive coding. In addition, we learn a conditional VAE model that generates samples given a context image of the domain, and use these hallucinated samples for building the connectivity graph, allowing for zero-shot generalization to domain changes. In simulated domains, HTM outperforms conventional SPTM and visual foresight methods in terms of both plan quality and success in long-horizon planning. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkgfxqIAtB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1663/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents HTM, an extension of the semiparametric topological memory method that augments the approach with hallucinated nodes and an energy cost function. The hallucination is enabled by a CVAE, conditioned on an image of the environment, and allows the method to generalize to unseen environments. The energy cost function is trained as a contrastive loss and acts as a robustness score for connecting the two samples. The underlying graph is then used to plan for several top view planning problems.\n\nThe paper is well written and clear. I believe such latent representations are an interesting approach to solving visual navigation and general planning. HTM provides an interesting and useful extension to SPTM, allowing both generalization to unseen environments and a more robust loss function. The \n\nMy primary concern is the lack of rigorous experimentation to validate the concept and push it’s limits. The results in Table 1 show HTM outperforms baselines clearly on the given problems, but how it performs on more complex problems is unclear. These problems are dynamically simple and the obstacles are easily identified. Some more difficult problems may be:\n- The mazes in SPTM or environments from https://arxiv.org/pdf/1612.03801.pdf.\n- The original SPTM paper focuses on visual navigation from first person views. How does this method apply to such situations? How does the context translate to this scenario?\n- Planning in real environments with real images, as done in [6].\n\nOther comparisons and notes:\n- Can the method be applied to higher dimensional problems (dimensionality of the underlying space) where planning may be more difficult? E.g. SE(2), robot arms or other agents from UPN [26]. Application with actual 3D workspace problems too would be interesting as the image context may underspecify the environment.\n- The energy cost function acts as a proxy for connection probability when traversing an edge. This may also be useful for dynamical systems (e.g., the mujoco ant navigating a maze). Are there limitations for the method on such problems, e.g., edges may no longer be symmetric?\n- How does SPTM compare when the space has been explored already?\n- Can more quantitative results been shown such as solution path cost?\n- Provide definitions for Fidelity, feasibility, and completeness and the source of data (polling human’s) in the Table 1 caption.\n- “As shown in Table 5.2, “, should be renamed to Table 1.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper presents HTM, an extension of the semiparametric topological memory method that augments the approach with hallucinated nodes and an energy cost function. The hallucination is enabled by a CVAE, conditioned on an image of the environment, and allows the method to generalize to unseen environments. The energy cost function is trained as a contrastive loss and acts as a robustness score for connecting the two samples. The underlying graph is then used to plan for several top view planning problems.\n\nThe paper is well written and clear. I believe such latent representations are an interesting approach to solving visual navigation and general planning. HTM provides an interesting and useful extension to SPTM, allowing both generalization to unseen environments and a more robust loss function. The \n\nMy primary concern is the lack of rigorous experimentation to validate the concept and push it’s limits. The results in Table 1 show HTM outperforms baselines clearly on the given problems, but how it performs on more complex problems is unclear. These problems are dynamically simple and the obstacles are easily identified. Some more difficult problems may be:\n- The mazes in SPTM or environments from https://arxiv.org/pdf/1612.03801.pdf.\n- The original SPTM paper focuses on visual navigation from first person views. How does this method apply to such situations? How does the context translate to this scenario?\n- Planning in real environments with real images, as done in [6].\n\nOther comparisons and notes:\n- Can the method be applied to higher dimensional problems (dimensionality of the underlying space) where planning may be more difficult? E.g. SE(2), robot arms or other agents from UPN [26]. Application with actual 3D workspace problems too would be interesting as the image context may underspecify the environment.\n- The energy cost function acts as a proxy for connection probability when traversing an edge. This may also be useful for dynamical systems (e.g., the mujoco ant navigating a maze). Are there limitations for the method on such problems, e.g., edges may no longer be symmetric?\n- How does SPTM compare when the space has been explored already?\n- Can more quantitative results been shown such as solution path cost?\n- Provide definitions for Fidelity, feasibility, and completeness and the source of data (polling human’s) in the Table 1 caption.\n- “As shown in Table 5.2, “, should be renamed to Table 1.\n"}, "tcdate": 1571871210436}, {"id": "Syx34y_sYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1663/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper propose a novel visual planning approach which constructs explicit plans from \"hallucinated\" states of the environment. To hallucinate states, it uses a Conditional Variational Autoencoder (which is conditioned on a context image of the domain). To plan, it trains a Contrastive Predictive Coding (CPC) model for judging similarities between states, then applies this model to hallucinated states + start/end states, then runs Dijkstra on the edges weighted by similarities.\n\nI vote for accepting this paper as it tackles two important problems: where to get subgoals for visual planning and what similarity function to use for zero-shot planning. Furthermore, the paper is clearly written, the experiments are well-conducted and analyzed.\n\nDetailed arguments:\n1. Where to get subgoals for visual planning is an important question persistently arising in control tasks. SPTM-style solution is indeed limited because it relies on an exploration sequence as a source of subgoals. Every time the environment changes, data would need to be re-collected. Getting subgoals from a conditional generative model is a neat solution.\n2. Benchmarking similarity functions is crucial. One productive way to approach zero-shot problems is to employ similarity functions, but the question arises: what algorithm to use for training them? The paper compares two popular choices: CPC and Temporal Distance Classification (in particular, R-network). It thus provides guidance that CPC might be a better algorithm for training similarity functions.\n3. The paper is well-positioned in the related work and points to the correct deficiencies of the existing methods. It also features nice experimental design with controlled complexity of the tasks, ablation studies and two relevant baselines.\n\nI would encourage the authors to discuss the following questions:\n1) Fidelity in Table 3 - why is it lower for SPTM compared to HTM if both methods rely on the same generated samples? Is it because HTM selects betters samples than SPTM for its plans?\n2) Why is fidelity larger for SPTM in a more complex task 2?\n3) Same question about fidelity/feasibility for HTM1/2?\n4) Are there any plans to open-source the code?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper propose a novel visual planning approach which constructs explicit plans from \"hallucinated\" states of the environment. To hallucinate states, it uses a Conditional Variational Autoencoder (which is conditioned on a context image of the domain). To plan, it trains a Contrastive Predictive Coding (CPC) model for judging similarities between states, then applies this model to hallucinated states + start/end states, then runs Dijkstra on the edges weighted by similarities.\n\nI vote for accepting this paper as it tackles two important problems: where to get subgoals for visual planning and what similarity function to use for zero-shot planning. Furthermore, the paper is clearly written, the experiments are well-conducted and analyzed.\n\nDetailed arguments:\n1. Where to get subgoals for visual planning is an important question persistently arising in control tasks. SPTM-style solution is indeed limited because it relies on an exploration sequence as a source of subgoals. Every time the environment changes, data would need to be re-collected. Getting subgoals from a conditional generative model is a neat solution.\n2. Benchmarking similarity functions is crucial. One productive way to approach zero-shot problems is to employ similarity functions, but the question arises: what algorithm to use for training them? The paper compares two popular choices: CPC and Temporal Distance Classification (in particular, R-network). It thus provides guidance that CPC might be a better algorithm for training similarity functions.\n3. The paper is well-positioned in the related work and points to the correct deficiencies of the existing methods. It also features nice experimental design with controlled complexity of the tasks, ablation studies and two relevant baselines.\n\nI would encourage the authors to discuss the following questions:\n1) Fidelity in Table 3 - why is it lower for SPTM compared to HTM if both methods rely on the same generated samples? Is it because HTM selects betters samples than SPTM for its plans?\n2) Why is fidelity larger for SPTM in a more complex task 2?\n3) Same question about fidelity/feasibility for HTM1/2?\n4) Are there any plans to open-source the code?"}, "tcdate": 1571680051631}, {"id": "SkePIyMquB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1663/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a method for learning agents to solve visual planning, in particular to navigate to a desired goal position in a maze, with a learned topological map, i.e. a graph, where nodes correspond to positions in the maze and edges correspond accessibility (reachability in a certain number of steps). The work extends previous work (semi parametric topological memory, ref. [22]) in several ways. It claims to address a shortcoming of [22], namely the fact that the graph is calculated offline from random rollouts, by using a conditional variational auto-encoder to predict a set of observed images which could lie between the current position and the goal position, and, most importantly from a context image which describes the layout of the environment. These predicted images are then arranged in a graph through a connectivity predictor, which is trained from rollouts through a contrastive loss. Training is performed on multiple environments, and the context vector provides enough information for this connectivity network to generalize to unseen environments.  At test time, the agent navigates using a planner and a policy. The planner calculates the shortest path on a graph where edges are connectivity probabilities, and the policy is an inverse model trained on the output of the planner.\n\nWhile the idea of a topological memory with dynamic graph creation is certainly interesting, the work is unfortunately not well enough executed and the paper structured written in a way which makes it up to impossible to grasp what has been really done, as much information is missing which would be required for understanding. \n\nAs a first example, we are never really told what the observations are, which the agent sees. The different figures of the paper show very small images with a 2D maze from a bird’s eye view consisting of a walls arranged in a single connected component (mostly 1 to 3 strokes) in red color and an agent shown as a position indicated as a green dot. Are these the observed images? In absence of any other information, this is what we need to assume, and then this problem is fully observable and does not seem to be very challenging. Given the figures, even a handcrafted algorithm should be able to calculate the optimal solution with Dijkstra’s algorithm on a graph calculated from the pixel grid.\n\nThis important missing information alone makes it difficult to assess the paper, but the rest of the writing is similarly confusing. The authors focus on very short and dense descriptions of mathematical properties, but seem to have forgotten to ground the different symbols and to connect them to physical entities of the problem. The technical writing is in large parts disconnected from the problem which is addressed by it.\n\nFurther examples are:\n\n-\t“the data (…) is collected in a self-supervised manner”: what does this mean? Self-supervision is way of creating loss from data without labels, but I am not sure what is meant by collecting data this way.\n-\tThe paragraph on CPC in section 2 can only be understood if the contrastive loss is known. To make the paper self-consistent, this should be properly explained, and tied to a training procedure which details how exactly the positive and negative samples are defined … and collected.\n-\tThe CPC objective in section 2 is only loosely connected to its usage in section 3.2. Barely writing “we optimize a CPC” objective is not sufficient for understanding how this objective is really tied to the different entities of the problem. This paper contains maths (which is always a pleasure to read), but it is not a purely and abstract mathematical problem - a real task is addressed, so it needs to be connected to it. This connection has certainly been done by the authors while they were working on the problem, but they should also communicate it to the reader.\n-\tThe section on ML trajectory is too dense and should be rewritten. I don’t understand what the authors want to tell us here. Basically, a (generalized) Dijkstra is run on a graph, where edge weights are the density or density ratios learned by the CPC objective, and if the edge weights are probabilities, that the shortest path corresponds to a trajectory likelihood. This is known, and this information is buried in a dense set of equations which are difficult to decipher and do not add any further value to the paper.\n-\tThe connection between the planner (generalized Dijkstra) and the policy is never explained. We don’t know how the policy is trained and how it works.\n\nOne of the downsides of the method is that it requires a context image. This image is responsible for the generalization to unseen environments, but it is a major drawback, as the image must be created beforehand. The authors claim that the context image must only contain the layout in any format which makes it possible to extract information about navigational space from it, but in the experiments the context image corresponds to the full map – and it is probably equivalent to the observed images, but we can’t be sure as we haven’t been told. In any case, it is far from sure how this could generalize to more complex environments, let alone 3D navigation as is currently addressed in standard simulators like VizDoom, GIBSON, Matterport, Deepmind Lab, Habitat AI etc. \n\nThe authors’ claim that the proposed environment requires long-term planning, but looking at the images this does not seem to be the case. \n\nThe paper claims to perform zero-shot generalization and to adapt to changes in the environment, like the slight changes in camera motion, variations in lightning, but it unclear how the solution solves this claim.\n\nHow does the agent determine that a goal has been reached, without ground truth information? \n\nWhat happens, if the hallucinated images are disconnected (form several connected components) or are disconnected from the current position and/or from the goal position?\n\nAs mentioned, the method is evaluated on an environment, which is too simple. The experiments are difficult to assess, as we don’t really know what the agent observes. An information asymmetry is mentioned (visual foresight having the object’s (=agent’s) position and the others not) … but if the proposed method observes the bird’s eye view, it can infer the agent’s position (as the position of the green dot).\n\nSubjective evaluation by humans on this kind of simple data does not seem to be meaningful, in particular with a very low number of observers (5 people).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper presents a method for learning agents to solve visual planning, in particular to navigate to a desired goal position in a maze, with a learned topological map, i.e. a graph, where nodes correspond to positions in the maze and edges correspond accessibility (reachability in a certain number of steps). The work extends previous work (semi parametric topological memory, ref. [22]) in several ways. It claims to address a shortcoming of [22], namely the fact that the graph is calculated offline from random rollouts, by using a conditional variational auto-encoder to predict a set of observed images which could lie between the current position and the goal position, and, most importantly from a context image which describes the layout of the environment. These predicted images are then arranged in a graph through a connectivity predictor, which is trained from rollouts through a contrastive loss. Training is performed on multiple environments, and the context vector provides enough information for this connectivity network to generalize to unseen environments.  At test time, the agent navigates using a planner and a policy. The planner calculates the shortest path on a graph where edges are connectivity probabilities, and the policy is an inverse model trained on the output of the planner.\n\nWhile the idea of a topological memory with dynamic graph creation is certainly interesting, the work is unfortunately not well enough executed and the paper structured written in a way which makes it up to impossible to grasp what has been really done, as much information is missing which would be required for understanding. \n\nAs a first example, we are never really told what the observations are, which the agent sees. The different figures of the paper show very small images with a 2D maze from a bird’s eye view consisting of a walls arranged in a single connected component (mostly 1 to 3 strokes) in red color and an agent shown as a position indicated as a green dot. Are these the observed images? In absence of any other information, this is what we need to assume, and then this problem is fully observable and does not seem to be very challenging. Given the figures, even a handcrafted algorithm should be able to calculate the optimal solution with Dijkstra’s algorithm on a graph calculated from the pixel grid.\n\nThis important missing information alone makes it difficult to assess the paper, but the rest of the writing is similarly confusing. The authors focus on very short and dense descriptions of mathematical properties, but seem to have forgotten to ground the different symbols and to connect them to physical entities of the problem. The technical writing is in large parts disconnected from the problem which is addressed by it.\n\nFurther examples are:\n\n-\t“the data (…) is collected in a self-supervised manner”: what does this mean? Self-supervision is way of creating loss from data without labels, but I am not sure what is meant by collecting data this way.\n-\tThe paragraph on CPC in section 2 can only be understood if the contrastive loss is known. To make the paper self-consistent, this should be properly explained, and tied to a training procedure which details how exactly the positive and negative samples are defined … and collected.\n-\tThe CPC objective in section 2 is only loosely connected to its usage in section 3.2. Barely writing “we optimize a CPC” objective is not sufficient for understanding how this objective is really tied to the different entities of the problem. This paper contains maths (which is always a pleasure to read), but it is not a purely and abstract mathematical problem - a real task is addressed, so it needs to be connected to it. This connection has certainly been done by the authors while they were working on the problem, but they should also communicate it to the reader.\n-\tThe section on ML trajectory is too dense and should be rewritten. I don’t understand what the authors want to tell us here. Basically, a (generalized) Dijkstra is run on a graph, where edge weights are the density or density ratios learned by the CPC objective, and if the edge weights are probabilities, that the shortest path corresponds to a trajectory likelihood. This is known, and this information is buried in a dense set of equations which are difficult to decipher and do not add any further value to the paper.\n-\tThe connection between the planner (generalized Dijkstra) and the policy is never explained. We don’t know how the policy is trained and how it works.\n\nOne of the downsides of the method is that it requires a context image. This image is responsible for the generalization to unseen environments, but it is a major drawback, as the image must be created beforehand. The authors claim that the context image must only contain the layout in any format which makes it possible to extract information about navigational space from it, but in the experiments the context image corresponds to the full map – and it is probably equivalent to the observed images, but we can’t be sure as we haven’t been told. In any case, it is far from sure how this could generalize to more complex environments, let alone 3D navigation as is currently addressed in standard simulators like VizDoom, GIBSON, Matterport, Deepmind Lab, Habitat AI etc. \n\nThe authors’ claim that the proposed environment requires long-term planning, but looking at the images this does not seem to be the case. \n\nThe paper claims to perform zero-shot generalization and to adapt to changes in the environment, like the slight changes in camera motion, variations in lightning, but it unclear how the solution solves this claim.\n\nHow does the agent determine that a goal has been reached, without ground truth information? \n\nWhat happens, if the hallucinated images are disconnected (form several connected components) or are disconnected from the current position and/or from the goal position?\n\nAs mentioned, the method is evaluated on an environment, which is too simple. The experiments are difficult to assess, as we don’t really know what the agent observes. An information asymmetry is mentioned (visual foresight having the object’s (=agent’s) position and the others not) … but if the proposed method observes the bird’s eye view, it can infer the agent’s position (as the position of the green dot).\n\nSubjective evaluation by humans on this kind of simple data does not seem to be meaningful, in particular with a very low number of observers (5 people).\n"}, "tcdate": 1570541390857}], "openreview_url": "https://openreview.net/forum?id=BkgF4kSFPB", "arxiv_id": "2002.12336", "paper_pdf": "papers/BkgF4kSFPB.pdf", "paper_pdf_sha256": "83688d472624b80f280220069766928741253d656665d2e51fad7cf2e4403ae0", "paper_pdf_bytes": 3414096, "paper_pdf_source": "openreview", "code_url": "https://github.com/thanard/hallucinative-topological-memory", "code_repository": "thanard/hallucinative-topological-memory", "code_commit": "5e75a0fb926dbee2323e292a82846b8ea81f99fa", "code_archive": "repos/BkgF4kSFPB.zip", "code_archive_sha256": "17474f39bf4c8971fdfec403717338366a44166844364c4555f4de52eb45ef5c", "code_archive_bytes": 34465, "code_file_count": 19, "code_extensions": {".py": 14, ".sh": 5}, "github_disk_usage_kb": 71, "github_languages": {"Python": 54931, "Shell": 1428}, "github_archived": false, "github_pushed_at": "2020-08-04T04:03:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hallucinative-topological-memory-for-zero-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HyGySsAct7", "year": 2019, "status": "rejected", "title": "Targeted Adversarial Examples for Black Box Audio Systems", "authors": ["Rohan Taori", "Amog Kamsetty", "Brenton Chu", "Nikita Vemuri"], "authorids": ["rohantaori@berkeley.edu", "amogkamsetty@berkeley.edu", "brentonlongchu@berkeley.edu", "nikitavemuri@berkeley.edu"], "authors_source": "OpenReview API", "abstract": "The application of deep recurrent networks to audio transcription has led to impressive gains in automatic speech recognition (ASR) systems. Many have demonstrated that small adversarial perturbations can fool deep neural networks into incorrectly predicting a specified target with high confidence. Current work on fooling ASR systems have focused on white-box attacks, in which the model architecture and parameters are known. In this paper, we adopt a black-box approach to adversarial generation, combining the approaches of both genetic algorithms and gradient estimation to solve the task. We achieve a 89.25% targeted attack similarity after 3000 generations while maintaining 94.6% audio file similarity.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "r1g3c6VgpX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper52/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a black-box attack on multi-word ASR systems.  Most work on black-box attacks have focused on tasks in vision. This work adds to the literature on attac\nks on speech systems. The key novelties are the handling of a loss function over multiple decodings as well as the use of novel genetic algorithms to generate the adversari\nal examples.\n\nA weakness of this paper is that they do not compare to the closely related Alzantot et al. work. While the latter is focused on single word settings and is thus solving an\n easier problem, what would happen if the Alzantot et al. method was applied to each\n\n\nWhile the idea is interesting but incremental, the evaluation of the approach is weak.\n\n1. Insted of choosing random pairs of words as target phrases, it would be interesting to pick phrases that are likely to occur in English and to ask how success rate varie\ns as a function of the initial phrase and target phrase.\n\n2. To confirm that the resulting adversarial examples are similar to audio samples in the original dataset, the authors should do user studies. This is a key component in e\nvaluating the efficacy of such attacks. The cross correlation is useful but does not get at perceptual similarity.\n\n3. Table 1 is not useful since either the datasets are different or information is not given on the specific white box attacks.\n\n4. Does increasing the iterations lead to a higher success rate as claimed at end of page 7?\n\n\nAbstract:\n1. This sentence is misleading : \"Current work..are known\" given the Alzantot et al. work focuses on black-box attacks.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Evaluation is weak", "review": "This paper proposes a black-box attack on multi-word ASR systems.  Most work on black-box attacks have focused on tasks in vision. This work adds to the literature on attac\nks on speech systems. The key novelties are the handling of a loss function over multiple decodings as well as the use of novel genetic algorithms to generate the adversari\nal examples.\n\nA weakness of this paper is that they do not compare to the closely related Alzantot et al. work. While the latter is focused on single word settings and is thus solving an\n easier problem, what would happen if the Alzantot et al. method was applied to each\n\n\nWhile the idea is interesting but incremental, the evaluation of the approach is weak.\n\n1. Insted of choosing random pairs of words as target phrases, it would be interesting to pick phrases that are likely to occur in English and to ask how success rate varie\ns as a function of the initial phrase and target phrase.\n\n2. To confirm that the resulting adversarial examples are similar to audio samples in the original dataset, the authors should do user studies. This is a key component in e\nvaluating the efficacy of such attacks. The cross correlation is useful but does not get at perceptual similarity.\n\n3. Table 1 is not useful since either the datasets are different or information is not given on the specific white box attacks.\n\n4. Does increasing the iterations lead to a higher success rate as claimed at end of page 7?\n\n\nAbstract:\n1. This sentence is misleading : \"Current work..are known\" given the Alzantot et al. work focuses on black-box attacks.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541586324157}, {"id": "SkxXsTqT3m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper52/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In \"Targeted adversarial examples for black box audio systems\" the authors look at an adversarial problem in neural nets for audio processing. There is quite a lot of recent interest in adversarial problems in machine learning. That work is mostly on the image side, and so this work is very topical. The problem is to modify an audio signal without changing how it sounds to the human ear, so that it is interpreted as the attacker wishes by the neural network. In the black box approach, the weights of the neural network are not known by the attacker. The attacker however must be able to present modified audio and learn the network's interpretation as often as the attacker wants. This work is very exciting and topical, and of interest to the ICLR community.\n\nThe authors demonstrate a proof of concept using the recent DeepSpeech model, and they connect very well with recent literature on adversarial networks.\n\nThe particular algorithm the authors propose is based on genetic algorithms. I thought that this was a weak part of the paper, because genetic algorithms are quite ad hoc and have few theoretical guarantees when compared to SMC, MCMC, nested sampling or herding, which all do basically the same thing as genetic algorithms. This can lead to loose ends, such as the \"momentum mutation\" introduced by the authors in 2.2, wherein probability of mutation increases as the population fails to adapt. It is true that momentum mutation would avoid local maxima, but it would also take the solution away from global maxima through a sort of \"sampling noise\" (the global maxima is a point at which the population also \"fails to adapt\", as there's no more adaptation to be done). It's unclear if this is a problem, but things like annealed importance sampling also deal with the same problem (or effective sample size of SMC), and they have theory to back them up.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Targeted adversarial examples for black box audio systems", "review": "In \"Targeted adversarial examples for black box audio systems\" the authors look at an adversarial problem in neural nets for audio processing. There is quite a lot of recent interest in adversarial problems in machine learning. That work is mostly on the image side, and so this work is very topical. The problem is to modify an audio signal without changing how it sounds to the human ear, so that it is interpreted as the attacker wishes by the neural network. In the black box approach, the weights of the neural network are not known by the attacker. The attacker however must be able to present modified audio and learn the network's interpretation as often as the attacker wants. This work is very exciting and topical, and of interest to the ICLR community.\n\nThe authors demonstrate a proof of concept using the recent DeepSpeech model, and they connect very well with recent literature on adversarial networks.\n\nThe particular algorithm the authors propose is based on genetic algorithms. I thought that this was a weak part of the paper, because genetic algorithms are quite ad hoc and have few theoretical guarantees when compared to SMC, MCMC, nested sampling or herding, which all do basically the same thing as genetic algorithms. This can lead to loose ends, such as the \"momentum mutation\" introduced by the authors in 2.2, wherein probability of mutation increases as the population fails to adapt. It is true that momentum mutation would avoid local maxima, but it would also take the solution away from global maxima through a sort of \"sampling noise\" (the global maxima is a point at which the population also \"fails to adapt\", as there's no more adaptation to be done). It's unclear if this is a problem, but things like annealed importance sampling also deal with the same problem (or effective sample size of SMC), and they have theory to back them up.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541414299315}, {"id": "B1eLktzOnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper52/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "PAPER SUMMARY:\n\nThis paper introduces a biologically motivated black-box attack algorithm. \nThe target model in this case is DNN applied to the ASR context (automatic speech recognition system). \n\nNOVELTY & SIGNIFICANCE:\n\nThe proposed approach extends the previous genetic approach of (Alzantot et al., 2018) to attack a more complicated ASR system (that handles phrases and sentences). The new contribution here is an add-on momentum mutation component on top of the existing genetic programming architecture of (Alzantot et al., 2018) as illustrated in Figure 3.\n\nThis however appears very incremental seeing that integrating the mutation component into existing system is straight-forward and that mutation is not even a new concept -- it has always been a vital component in genetic programming paradigm.\n\nIt is also unclear how this mutation component improves over the existing work (more on this in the sections below).\n\nAnother issue is this work seems to ignore the recent literature on adversarial black-box attacks to DNN model. To list a few:\n\nChen, P.-Y.; Zhang, H.; Sharma, Y.; Yi, J.; and Hsieh, C.-J. 2017b.\nZOO: Zeroth-order optimization-based  black-box attacks to deepneural networks without training substitute models. \nIn Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security (15-26) ACM\n\nCheng,  M.;  Le,  T.;  Chen,  P.-Y.;  Yi,  J.;  Zhang,  H.;  and  Hsieh,C.-J.2018.\nQuery-efficient hard-label black-box attack:  An optimization-based approach. arXiv preprint arXiv:1807.04457\n\nWhile these works have not been used to attacking ASR system, they should be directly applicable to such system since after all, they are black-box attacks. I think the proposed method needs to be compared with these works.\n\nTECHNICAL SOUNDNESS:\n\nI find it surprising that even though the proposed method is claimed to be a black-box attack but in the end, it actually exploits the fact that the target model uses CTC decoder. This pertains specifically to the target model's internal architecture and a black-box attack is not supposed to know this.\n\nCLARITY:\n\nThe paper is clearly written.\n\nEMPIRICAL RESULTS:\n\nI do not understand this statement:\n\n\"That 35% of random attacks were successful in this respect highlights the fact that black box\nadversarial attacks are definitely possible and highly effective at the same time\"\n\nWhy is 35% successful attack rate a positive result? The result tends to suggest that this is an attack with low success rate. \n\nThe 2nd paragraph in 3.2 seems to give a vague explanation: \"the vast majority of failure cases are only a few edit distances away from the target. \n\nThis suggests that running the algorithm for a few more iterations could produce a higher success rate, although at the cost of correlation similarity\".\n\nGiven the above statement, I do not see why the authors didn't actually \"run the algorithm for a few more iterations\" to verify it ...\n\nI am also curious why is the success rate of the proposed method is significantly lower than that of the existing system -- I assume \"single word black box\" is the work of (Alzantot et al., 2018).\n\nI find the empirical evaluation somewhat sloppy: why are the tested method not compared on the same benchmark? How do we interpret the results then?\n\nREVIEW SUMMARY:\n\nThe paper misses the recent literature on black-box attack. The authors need to compare with those to demonstrate the efficiency of their proposed work. I also find the contribution of this paper too incremental & its empirical evaluation appears somewhat sloppy and not convincing (see my specific comments above). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is not well-positioned against the existing literature on black-box attack. Its empirical evaluation is somewhat sloppy.", "review": "PAPER SUMMARY:\n\nThis paper introduces a biologically motivated black-box attack algorithm. \nThe target model in this case is DNN applied to the ASR context (automatic speech recognition system). \n\nNOVELTY & SIGNIFICANCE:\n\nThe proposed approach extends the previous genetic approach of (Alzantot et al., 2018) to attack a more complicated ASR system (that handles phrases and sentences). The new contribution here is an add-on momentum mutation component on top of the existing genetic programming architecture of (Alzantot et al., 2018) as illustrated in Figure 3.\n\nThis however appears very incremental seeing that integrating the mutation component into existing system is straight-forward and that mutation is not even a new concept -- it has always been a vital component in genetic programming paradigm.\n\nIt is also unclear how this mutation component improves over the existing work (more on this in the sections below).\n\nAnother issue is this work seems to ignore the recent literature on adversarial black-box attacks to DNN model. To list a few:\n\nChen, P.-Y.; Zhang, H.; Sharma, Y.; Yi, J.; and Hsieh, C.-J. 2017b.\nZOO: Zeroth-order optimization-based  black-box attacks to deepneural networks without training substitute models. \nIn Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security (15-26) ACM\n\nCheng,  M.;  Le,  T.;  Chen,  P.-Y.;  Yi,  J.;  Zhang,  H.;  and  Hsieh,C.-J.2018.\nQuery-efficient hard-label black-box attack:  An optimization-based approach. arXiv preprint arXiv:1807.04457\n\nWhile these works have not been used to attacking ASR system, they should be directly applicable to such system since after all, they are black-box attacks. I think the proposed method needs to be compared with these works.\n\nTECHNICAL SOUNDNESS:\n\nI find it surprising that even though the proposed method is claimed to be a black-box attack but in the end, it actually exploits the fact that the target model uses CTC decoder. This pertains specifically to the target model's internal architecture and a black-box attack is not supposed to know this.\n\nCLARITY:\n\nThe paper is clearly written.\n\nEMPIRICAL RESULTS:\n\nI do not understand this statement:\n\n\"That 35% of random attacks were successful in this respect highlights the fact that black box\nadversarial attacks are definitely possible and highly effective at the same time\"\n\nWhy is 35% successful attack rate a positive result? The result tends to suggest that this is an attack with low success rate. \n\nThe 2nd paragraph in 3.2 seems to give a vague explanation: \"the vast majority of failure cases are only a few edit distances away from the target. \n\nThis suggests that running the algorithm for a few more iterations could produce a higher success rate, although at the cost of correlation similarity\".\n\nGiven the above statement, I do not see why the authors didn't actually \"run the algorithm for a few more iterations\" to verify it ...\n\nI am also curious why is the success rate of the proposed method is significantly lower than that of the existing system -- I assume \"single word black box\" is the work of (Alzantot et al., 2018).\n\nI find the empirical evaluation somewhat sloppy: why are the tested method not compared on the same benchmark? How do we interpret the results then?\n\nREVIEW SUMMARY:\n\nThe paper misses the recent literature on black-box attack. The authors need to compare with those to demonstrate the efficiency of their proposed work. I also find the contribution of this paper too incremental & its empirical evaluation appears somewhat sloppy and not convincing (see my specific comments above). ", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541052638221}], "openreview_url": "https://openreview.net/forum?id=HyGySsAct7", "arxiv_id": "1805.07820", "paper_pdf": "papers/HyGySsAct7.pdf", "paper_pdf_sha256": "815b739b4090b9da68f044e7975924cb0c2cfbd6be6d245f097e9fa1418642e6", "paper_pdf_bytes": 1122854, "paper_pdf_source": "openreview", "code_url": "https://github.com/rtaori/Black-Box-Audio", "code_repository": "rtaori/Black-Box-Audio", "code_commit": "3d6c9001faed2988b39ee142200d11e7a54846a3", "code_archive": "repos/HyGySsAct7.zip", "code_archive_sha256": "96ce550898c6c061c6279b0fc145726d4498fd6bb74672ff42ec17952b7be21b", "code_archive_bytes": 469586, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 466, "github_languages": {"Python": 17625}, "github_archived": false, "github_pushed_at": "2020-08-27T19:40:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/targeted-adversarial-examples-for-black-box"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1spAqUp-", "year": 2018, "status": "rejected", "title": "Pixel Deconvolutional Networks", "authors": ["Hongyang Gao", "Hao Yuan", "Zhengyang Wang", "Shuiwang Ji"], "authorids": ["hongyang.gao@wsu.edu", "hao.yuan@wsu.edu", "zwang6@eecs.wsu.edu", "sji@eecs.wsu.edu"], "authors_source": "OpenReview API", "abstract": "Deconvolutional layers have been widely used in a variety of deep\nmodels for up-sampling, including encoder-decoder networks for\nsemantic segmentation and deep generative models for unsupervised\nlearning. One of the key limitations of deconvolutional operations\nis that they result in the so-called checkerboard problem. This is\ncaused by the fact that no direct relationship exists among adjacent\npixels on the output feature map. To address this problem, we\npropose the pixel deconvolutional layer (PixelDCL) to establish\ndirect relationships among adjacent pixels on the up-sampled feature\nmap. Our method is based on a fresh interpretation of the regular\ndeconvolution operation. The resulting PixelDCL can be used to\nreplace any deconvolutional layer in a plug-and-play manner without\ncompromising the fully trainable capabilities of original models.\nThe proposed PixelDCL may result in slight decrease in efficiency,\nbut this can be overcome by an implementation trick. Experimental\nresults on semantic segmentation demonstrate that PixelDCL can\nconsider spatial features such as edges and shapes and yields more\naccurate segmentation outputs than deconvolutional layers. When used\nin image generation tasks, our PixelDCL can largely overcome the\ncheckerboard problem suffered by regular deconvolution operations.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1L5VaYgG", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper20/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Paper summary:\nThis paper proposes a technique to generalize deconvolution operations used in standard CNN architectures. Traditional deconvolution operation uses independent filter weights to compute output features at adjacent pixels. This work proposes to do sequential prediction of adjacent pixel features (via intermediate feature maps) resulting in more spatially smooth outputs for deconvolution layer. This new layer is referred to as ‘pixel deconvolution layer’ and it is demonstrated on two tasks of semantic segmentation and face generation.\n\n\nPaper Strengths:\n- Despite being simple technique, the proposed pixel deconvolution layer is novel and interesting.\n- Experimental results on two different tasks demonstrating the general use of the proposed deconvolution layer.\n\n\nMajor Weaknesses:\n- The main weakness of this paper lies in its weak experiments. Although authors say that several possibilities exist for the dependencies between intermediate feature maps, there are no systematic ablation studies on what type of connectivities work best for the proposed layer. Authors experimented with two randomly chosen connectivities which is not enough to understand what type of connectivities work best. This is important as this forms the main contribution of the paper.\n- Also, several quantitative results seem incomplete. Why is the DeepLab-ResNet performance so low? A quick look at PascalVOC results indicate that DeepLab-ResNet has IoU of over 79 on this dataset, but the reported numbers in this paper are only around 73 IoU. There is no mention of IoU for base DeepLab-ResNet model and the standard DeepLab+CRF technique. And, there are no quantitative results on image generation.\n\n\nMinor Weaknesses:\n- Although the paper is easy to understand, several parts of the paper are poorly written. Several sentences are repeated multiple times across the paper. Some statements need corrections/refinements such as “mean IoU is a more accuracy evaluation measure”. And, it is better to under-tone some statements such as changing “solving” to “tackling”.\n- The illustration of checkerboard artifacts from standard deconvolution technique is not clear. For example, the results presented in Figure-4 indicate segmentation mistakes of the network rather than checkerboard artifacts.\n\n\nClarifications:\n- Why authors choose to ‘resize’ the images for training semantic segmentation networks, instead of generally used ‘cropping’ to create batches?\n- I can not see the ‘red’ in Figure-5. I see the later feature map more as ‘pinkish’ color. It is probably due to my color vision. In any case, it is better to use different color scheme to distinguish.\n\n\nSuggestions:\n- I strongly advice authors to do some ablation studies on connectivities to make this a good paper. Also, it would be great if authors can revise the writing thoroughly to make this a more enjoyable read.\n\n\nReview Summary:\nThe proposed technique, despite being simple, is novel and interesting. But, the weak and incomplete experiments make this not yet ready for publication.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple yet good technique for better deconvolutions in neural networks. But, the experiments are weak and not good enough.", "rating": "5: Marginally below acceptance threshold", "review": "Paper summary:\nThis paper proposes a technique to generalize deconvolution operations used in standard CNN architectures. Traditional deconvolution operation uses independent filter weights to compute output features at adjacent pixels. This work proposes to do sequential prediction of adjacent pixel features (via intermediate feature maps) resulting in more spatially smooth outputs for deconvolution layer. This new layer is referred to as ‘pixel deconvolution layer’ and it is demonstrated on two tasks of semantic segmentation and face generation.\n\n\nPaper Strengths:\n- Despite being simple technique, the proposed pixel deconvolution layer is novel and interesting.\n- Experimental results on two different tasks demonstrating the general use of the proposed deconvolution layer.\n\n\nMajor Weaknesses:\n- The main weakness of this paper lies in its weak experiments. Although authors say that several possibilities exist for the dependencies between intermediate feature maps, there are no systematic ablation studies on what type of connectivities work best for the proposed layer. Authors experimented with two randomly chosen connectivities which is not enough to understand what type of connectivities work best. This is important as this forms the main contribution of the paper.\n- Also, several quantitative results seem incomplete. Why is the DeepLab-ResNet performance so low? A quick look at PascalVOC results indicate that DeepLab-ResNet has IoU of over 79 on this dataset, but the reported numbers in this paper are only around 73 IoU. There is no mention of IoU for base DeepLab-ResNet model and the standard DeepLab+CRF technique. And, there are no quantitative results on image generation.\n\n\nMinor Weaknesses:\n- Although the paper is easy to understand, several parts of the paper are poorly written. Several sentences are repeated multiple times across the paper. Some statements need corrections/refinements such as “mean IoU is a more accuracy evaluation measure”. And, it is better to under-tone some statements such as changing “solving” to “tackling”.\n- The illustration of checkerboard artifacts from standard deconvolution technique is not clear. For example, the results presented in Figure-4 indicate segmentation mistakes of the network rather than checkerboard artifacts.\n\n\nClarifications:\n- Why authors choose to ‘resize’ the images for training semantic segmentation networks, instead of generally used ‘cropping’ to create batches?\n- I can not see the ‘red’ in Figure-5. I see the later feature map more as ‘pinkish’ color. It is probably due to my color vision. In any case, it is better to use different color scheme to distinguish.\n\n\nSuggestions:\n- I strongly advice authors to do some ablation studies on connectivities to make this a good paper. Also, it would be great if authors can revise the writing thoroughly to make this a more enjoyable read.\n\n\nReview Summary:\nThe proposed technique, despite being simple, is novel and interesting. But, the weak and incomplete experiments make this not yet ready for publication.", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511801998474}, {"id": "BkZQtx5lz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper20/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed the new approach for feature upsampling called pixel deconvolution, which aims to resolve checkboard artifact of conventional deconvolution. By sequentially applying a series of decomposed convolutions, the proposed method explicitly enforces the model to consider the relation between pixels thus effectively improve the deconvolution network with an increased computational cost to some extent.\n\nOverall, the paper is clearly written and easy to understand the main motivation and methods. However, the checkboard artifact is a well-known problem of deconvolution network, and has been addressed by several approaches which are simpler than the proposed pixel deconvolution. For example, it is well known that simple bilinear interpolation optionally followed by convolutions effectively removes checkboard artifact to some extent, and bilinear additive upsampling proposed in Wonja et al., 2017 also demonstrated its effectiveness as an alternative for deconvolution. Comparisons against these approaches would make the paper stronger. Besides, comparisons/discussions based on extensive analysis on various deconvolution architectures presented in Wonja et al., 2017 would also be interesting.\n\nWonja et al, The Devil is in the Decoder, In BMVC, 2017\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "No title", "rating": "6: Marginally above acceptance threshold", "review": "This paper proposed the new approach for feature upsampling called pixel deconvolution, which aims to resolve checkboard artifact of conventional deconvolution. By sequentially applying a series of decomposed convolutions, the proposed method explicitly enforces the model to consider the relation between pixels thus effectively improve the deconvolution network with an increased computational cost to some extent.\n\nOverall, the paper is clearly written and easy to understand the main motivation and methods. However, the checkboard artifact is a well-known problem of deconvolution network, and has been addressed by several approaches which are simpler than the proposed pixel deconvolution. For example, it is well known that simple bilinear interpolation optionally followed by convolutions effectively removes checkboard artifact to some extent, and bilinear additive upsampling proposed in Wonja et al., 2017 also demonstrated its effectiveness as an alternative for deconvolution. Comparisons against these approaches would make the paper stronger. Besides, comparisons/discussions based on extensive analysis on various deconvolution architectures presented in Wonja et al., 2017 would also be interesting.\n\nWonja et al, The Devil is in the Decoder, In BMVC, 2017\n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511815448826}, {"id": "B1YorpYxz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper20/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is well written and easy to follow. The authors propose pixel deconvolutional layers for convolutional neural networks. The motivation of the proposed method, PixelDCL, is to remove the checkerboard effect of deconvolutoinal layers. \nThe method consists of adding direct dependencies among the intermediate feature maps generated by the deconv layer. PixelDCL is applied sequentially, therefore it is slower than the original deconvolutional layer. The authors evaluate the model in two different problems: semantic segmentation (on PASCAL VOC and MSCOCO datasets) and in image generation VAE (with the CelebA dataset). \n\nThe authors justify the proposed method as a way to alleviate the checkerboard effect (while introducing more complexity to the model and making it slower). In the experimental section, however, they do not compare with other approaches to do so For example, the upsampling+conv approach, which has been shown to remove the checkerboard effect while being more efficient than the proposed method (as it does not require any sequential computation). Moreover, the PixelDCL does not seem to bring substantial improvements on DeepLab (a state-of-the-art semantic segmentation algorithm). More comments and further exploration on this results should be done. Why no performance boost? Is it because of the residual connection? Or other component of DeepLab? Is the proposed layer really useful once a powerful model is used?\n\nI also think the experiments on VAE are not conclusive. The authors simply show set of generated images. First, it is difficult to see the different of the image generated using deconv and PixelDCL. Second, a set of 20 qualitative images does not (and cannot) validate any research idea.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for Pixel Deconvolutional Networks", "rating": "5: Marginally below acceptance threshold", "review": "This paper is well written and easy to follow. The authors propose pixel deconvolutional layers for convolutional neural networks. The motivation of the proposed method, PixelDCL, is to remove the checkerboard effect of deconvolutoinal layers. \nThe method consists of adding direct dependencies among the intermediate feature maps generated by the deconv layer. PixelDCL is applied sequentially, therefore it is slower than the original deconvolutional layer. The authors evaluate the model in two different problems: semantic segmentation (on PASCAL VOC and MSCOCO datasets) and in image generation VAE (with the CelebA dataset). \n\nThe authors justify the proposed method as a way to alleviate the checkerboard effect (while introducing more complexity to the model and making it slower). In the experimental section, however, they do not compare with other approaches to do so For example, the upsampling+conv approach, which has been shown to remove the checkerboard effect while being more efficient than the proposed method (as it does not require any sequential computation). Moreover, the PixelDCL does not seem to bring substantial improvements on DeepLab (a state-of-the-art semantic segmentation algorithm). More comments and further exploration on this results should be done. Why no performance boost? Is it because of the residual connection? Or other component of DeepLab? Is the proposed layer really useful once a powerful model is used?\n\nI also think the experiments on VAE are not conclusive. The authors simply show set of generated images. First, it is difficult to see the different of the image generated using deconv and PixelDCL. Second, a set of 20 qualitative images does not (and cannot) validate any research idea.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511802272999}], "openreview_url": "https://openreview.net/forum?id=B1spAqUp-", "arxiv_id": "1705.06820", "paper_pdf": "papers/B1spAqUp-.pdf", "paper_pdf_sha256": "225d2d063947da936db5e382ba17ddbb63eda29732d07ce110a773bf56c688e3", "paper_pdf_bytes": 3512620, "paper_pdf_source": "openreview", "code_url": "https://github.com/divelab/PixelTCN", "code_repository": "divelab/PixelTCN", "code_commit": "40806e88da8b174ed38f7fae1fcc5f345e3e67c2", "code_archive": "repos/B1spAqUp-.zip", "code_archive_sha256": "0e00f408b9ff95de60708ae79935ed862ff9c5e0452341e361738319e6c6ee93", "code_archive_bytes": 7724808, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 5072, "github_languages": {"Python": 33457}, "github_archived": false, "github_pushed_at": "2018-12-13T17:15:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pixel-deconvolutional-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "G9bpAoc47a", "year": 2026, "status": "rejected", "title": "It Takes a Graph to Know a Graph: Rewiring for Homophily with a Reference Graph", "authors": ["Harel Mendelman", "Haggai Maron", "Ronen Talmon"], "authorids": ["~Harel_Mendelman1", "~Haggai_Maron1", "~Ronen_Talmon2"], "authors_source": "OpenReview API", "abstract": "Graph Neural Networks (GNNs) excel at analyzing graph-structured data but struggle on heterophilic graphs, where connected nodes often belong to different classes. While this challenge is commonly addressed with specialized GNN architectures, graph rewiring remains an underexplored strategy in this context. We provide theoretical foundations linking edge homophily, GNN embedding smoothness, and node classification performance, motivating the need to enhance homophily. Building on this insight, we introduce a rewiring framework that increases graph homophily using a reference graph, with theoretical guarantees on the homophily of the rewired graph. To broaden applicability, we propose a label-driven diffusion approach for constructing a homophilic reference graph from node features and training labels. Through extensive simulations, we analyze how the homophily of both the original and reference graphs influences the rewired graph homophily and downstream GNN performance. We evaluate our method on 13 real-world heterophilic datasets and show that it outperforms existing rewiring techniques and specialized GNNs for heterophilic graphs, achieving improved node classification accuracy while remaining efficient and scalable to large graphs.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "czNTMeAokE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24227/Reviewer_SpJs"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper introduces REFine, a graph rewiring framework designed to enhance graph homophily using a reference graph, thereby improving the performance of standard Graph Neural Networks (GNNs) on heterophilic graphs. The authors provide a theoretical foundation linking edge homophily, embedding smoothness, and GNN performance, which motivates homophily enhancement. They propose a principled rewiring method guided by a reference graph, with theoretical guarantees on homophily improvement under certain conditions. A label-driven diffusion process is introduced to construct a homophilic reference graph from node features and training labels.", "review_text": "This paper introduces REFine, a graph rewiring framework designed to enhance graph homophily using a reference graph, thereby improving the performance of standard Graph Neural Networks (GNNs) on heterophilic graphs. The authors provide a theoretical foundation linking edge homophily, embedding smoothness, and GNN performance, which motivates homophily enhancement. They propose a principled rewiring method guided by a reference graph, with theoretical guarantees on homophily improvement under certain conditions. A label-driven diffusion process is introduced to construct a homophilic reference graph from node features and training labels.", "strengths": "S1 The paper establishes a clear theoretical connection between graph homophily and the smoothness of GNN embeddings (Theorem 1), providing a strong motivation for homophily-enhancing rewiring. \nS2 The experimental section is thorough, evaluating the method across a diverse set of 13 datasets with varying sizes and homophily levels.\nS3 The use of the METIS algorithm for graph partitioning, coupled with parallelizable per-cluster computations, makes the method highly scalable to large graphs.", "weaknesses": "W1 The effectiveness of the reference graph construction hinges on the assumption that node feature similarity is indicative of label similarity. In domains where this assumption does not hold, the method's performance may be limited. While the label-driven diffusion aims to mitigate this, the fundamental dependency remains a potential limitation.\nW2 The framework is specifically designed and evaluated for the node classification task. Its applicability to other fundamental graph learning tasks, such as graph classification or link prediction, is not explored or discussed, which may limit its perceived utility for the broader graph learning community.\nW3 The method involves several key hyperparameters, including the kernel scale ε, the choice between edge addition/deletion, and the number of edges k to rewire. While a grid search is employed, the performance is contingent on proper tuning, which could be computationally expensive and less user-friendly in practice.", "questions": "Q1 The current framework performs either edge addition or deletion in a single rewiring step. Have the authors considered a strategy that performs both operations simultaneously, which could offer finer control over the resulting graph topology?\nQ2 The reference graph is constructed in a fixed, non-learnable manner. Could the performance be further improved by integrating the reference graph construction into an end-to-end, trainable framework, perhaps drawing inspiration from GSL paradigms in future work?\nQ3 How does the method perform on graphs that are extremely sparse (leading to potential disconnection) or excessively dense (where rewiring might have minimal relative impact)? Are there specific regimes where the method is less effective?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces REFine, a graph rewiring framework designed to enhance graph homophily using a reference graph, thereby improving the performance of standard Graph Neural Networks (GNNs) on heterophilic graphs. The authors provide a theoretical foundation linking edge homophily, embedding smoothness, and GNN performance, which motivates homophily enhancement. They propose a principled rewiring method guided by a reference graph, with theoretical guarantees on homophily improvement under certain conditions. A label-driven diffusion process is introduced to construct a homophilic reference graph from node features and training labels.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "S1 The paper establishes a clear theoretical connection between graph homophily and the smoothness of GNN embeddings (Theorem 1), providing a strong motivation for homophily-enhancing rewiring. \nS2 The experimental section is thorough, evaluating the method across a diverse set of 13 datasets with varying sizes and homophily levels.\nS3 The use of the METIS algorithm for graph partitioning, coupled with parallelizable per-cluster computations, makes the method highly scalable to large graphs.", "weaknesses": "W1 The effectiveness of the reference graph construction hinges on the assumption that node feature similarity is indicative of label similarity. In domains where this assumption does not hold, the method's performance may be limited. While the label-driven diffusion aims to mitigate this, the fundamental dependency remains a potential limitation.\nW2 The framework is specifically designed and evaluated for the node classification task. Its applicability to other fundamental graph learning tasks, such as graph classification or link prediction, is not explored or discussed, which may limit its perceived utility for the broader graph learning community.\nW3 The method involves several key hyperparameters, including the kernel scale ε, the choice between edge addition/deletion, and the number of edges k to rewire. While a grid search is employed, the performance is contingent on proper tuning, which could be computationally expensive and less user-friendly in practice.", "questions": "Q1 The current framework performs either edge addition or deletion in a single rewiring step. Have the authors considered a strategy that performs both operations simultaneously, which could offer finer control over the resulting graph topology?\nQ2 The reference graph is constructed in a fixed, non-learnable manner. Could the performance be further improved by integrating the reference graph construction into an end-to-end, trainable framework, perhaps drawing inspiration from GSL paradigms in future work?\nQ3 How does the method perform on graphs that are extremely sparse (leading to potential disconnection) or excessively dense (where rewiring might have minimal relative impact)? Are there specific regimes where the method is less effective?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762058144916}, {"id": "pSUIBho0mC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24227/Reviewer_prfN"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper addresses the challenge of poor GNN performance on heterophilic graphs. It explores graph rewiring as a strategy to improve GNN performance with theoretical and empirical study. It specifically introduces a label-driven diffusion method to construct the reference graph from node features and available labels, to rewire graphs for increasing homophily. Extensive experiments could show the effectiveness of the proposed method.", "review_text": "The paper addresses the challenge of poor GNN performance on heterophilic graphs. It explores graph rewiring as a strategy to improve GNN performance with theoretical and empirical study. It specifically introduces a label-driven diffusion method to construct the reference graph from node features and available labels, to rewire graphs for increasing homophily. Extensive experiments could show the effectiveness of the proposed method.", "strengths": "1. The idea of addressing the homophily–heterophily issue through graph rewiring is interesting and provides a different perspective compared to traditional architectural modifications.\n\n2. The paper combines theoretical analysis regarding the reference graph with a comprehensive empirical study, and the explanations are generally clear and well-motivated.\n\n3. Some experimental results indeed show significant performance improvements over certain baselines.", "weaknesses": "1. The proposed method lacks clear novelty. Label-guided graph rewiring has already been studied, for example, in *Bose, K., Banerjee, S., & Das, S. (2025). “Can Graph Neural Networks Tackle Heterophily? Yes, With a Label-Guided Graph Rewiring Approach!” IEEE TNNLS*, and the core technical component \"label-driven diffusion\" used here appears directly derived from existing work Mendelman\n& Talmon (2025)., which substantially weakens the contribution.\n\n2. The paper does not convincingly justify why increasing graph homophily is the right direction. Ideally, models should handle both homophilic and heterophilic structures without modifying the graph itself. More puzzlingly, Table 2 shows that the proposed method, which increases homophily, performs better even on heterophilic GNNs (e.g., H2GCN) — this seems conceptually inconsistent and requires further discussion.\n\n3. The baseline comparison is outdated and incomplete. The latest strong methods such as *Bose et al., 2025 (label-guided rewiring) and Barbero et al., 2023, “Locality-Aware Graph Rewiring in GNNs,” ICLR 2024*, and above TNNLS 2025 work are not included.\n\n4. The organization and presentation need improvement. Many experimental results that belong in the main text are buried in the appendix, with duplicated tables (E.1 COMPLETE RESULTS seems same as Table.1), and no bottomline in many tables. The paper feels somewhat rushed.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of poor GNN performance on heterophilic graphs. It explores graph rewiring as a strategy to improve GNN performance with theoretical and empirical study. It specifically introduces a label-driven diffusion method to construct the reference graph from node features and available labels, to rewire graphs for increasing homophily. Extensive experiments could show the effectiveness of the proposed method.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The idea of addressing the homophily–heterophily issue through graph rewiring is interesting and provides a different perspective compared to traditional architectural modifications.\n\n2. The paper combines theoretical analysis regarding the reference graph with a comprehensive empirical study, and the explanations are generally clear and well-motivated.\n\n3. Some experimental results indeed show significant performance improvements over certain baselines.", "weaknesses": "1. The proposed method lacks clear novelty. Label-guided graph rewiring has already been studied, for example, in *Bose, K., Banerjee, S., & Das, S. (2025). “Can Graph Neural Networks Tackle Heterophily? Yes, With a Label-Guided Graph Rewiring Approach!” IEEE TNNLS*, and the core technical component \"label-driven diffusion\" used here appears directly derived from existing work Mendelman\n& Talmon (2025)., which substantially weakens the contribution.\n\n2. The paper does not convincingly justify why increasing graph homophily is the right direction. Ideally, models should handle both homophilic and heterophilic structures without modifying the graph itself. More puzzlingly, Table 2 shows that the proposed method, which increases homophily, performs better even on heterophilic GNNs (e.g., H2GCN) — this seems conceptually inconsistent and requires further discussion.\n\n3. The baseline comparison is outdated and incomplete. The latest strong methods such as *Bose et al., 2025 (label-guided rewiring) and Barbero et al., 2023, “Locality-Aware Graph Rewiring in GNNs,” ICLR 2024*, and above TNNLS 2025 work are not included.\n\n4. The organization and presentation need improvement. Many experimental results that belong in the main text are buried in the appendix, with duplicated tables (E.1 COMPLETE RESULTS seems same as Table.1), and no bottomline in many tables. The paper feels somewhat rushed.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761802986919}, {"id": "ltKy7gw1NS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24227/Reviewer_M5SM"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "The paper presents theoretical and empirical connections between smoothness, homophily/heterophily, and GNN performance. Based on these analyses, authors propose REFine, a feature and label-driven diffusion approach that rewires the graph to be more homophilous. Extensive experiments are conducted on homophilous and heterophilous benchmarks to demonstrate the utility of the approach.", "review_text": "The paper presents theoretical and empirical connections between smoothness, homophily/heterophily, and GNN performance. Based on these analyses, authors propose REFine, a feature and label-driven diffusion approach that rewires the graph to be more homophilous. Extensive experiments are conducted on homophilous and heterophilous benchmarks to demonstrate the utility of the approach.", "strengths": "1. The approach of graph rewiring for homophily/heterophily is intuitive and interesting\n2. The paper is sound, clear, and well-written.", "weaknesses": "1. **Positioning vs. prior rewiring for heterophily (novelty).**  \n  The core idea of REFine is to construct a more homophilous reference graph leveraging the graph **features + training labels** and then **adding/deleting edges**. To me, this is quite is close to **DHGR**, which also leverages similarities between node features and training labels to form the rewired graph. The similarities and differences of REFine in comparison to DHGR as well as its advantages/disadvantages are not clear. I believe the paper would benefit from a sharper distinction (beyond “simpler & guaranteed”) and **direct empirical comparisons** to DHGR/GSL baselines to make the contribution more clear. \n\n2. **Incremental theory relative to prior spectral/smoothness analyses.**  \n  The main smoothness and homophily result (Theorem 1) is very close to existing analyses that link Dirichlet energy/smoothing to homophily/heterophily. One example of where this appears in prior works is in Theorem 3 in \"Beyond Homophily in Graph Neural Networks\" by Zhu et al., 2020. The subsequent propositions in the paper also state intuitive claims (if the reference graph is more homophilous, then adding its edges/deleting complement edges improves homophily) under certain conditions. Please clarify how these results meaningfully extend prior theory and provide new insights for the limitations of prior methods or advantages of REFine.\n\n3. **Missing critical baseline and mixed empirical gains.**  \nWhile the method largely outperforms the benchmarked rewiring techniques, it is **often comparable** to specialized heterophilous GNNs, with a standout improvement on **Roman-empire** but with no major gains elsewhere. Additionally, the absence of a **DHGR** comparison weakens conclusions about superiority among heterophily-oriented rewiring methods. Including DHGR would strengthen the empirical case.\n\n4. **When to prefer rewiring over heterophilous GNNs.**  \nI am unsure as to when a practitioner should choose heterophilous GNNs or REFine. Are there any benefits of using the rewiring technique over GNNs other than just performance? If the performances are similar, why should we want to use a rewiring method over a heterophilous GNN? The paper should clarify when to choose rewiring (e.g., datasets where feature/label-driven reference graphs are reliably more homophilous), and provide guidelines on when specialized heterophilous GNNs remain the better choice.", "questions": "My questions largely stem from the limitations: \n\n1. What are the advantages and disadvantages of REFine in comparison to DHGR? What are the conceptual differences given DHGR also constructs a rewired graph based on node features similarities and train labels?\n2. How do the theoretical results meaningfully build on existing spectral analyses relating smoothness/energy to homophily/heterophily? \n3. Given the connection between oversmoothing, oversquashing, and heterophily, what are the connections between existing rewiring approaches that address oversmoothing and oversquashing to heterophily? why don't their constructions mitigate the heterophily problem?\n4. Why is DHGR not a baseline? How does DHGR perform in comparison to REFine?\n5. It would be good as to get an intuition for why we would consider using a rewiring approach over a heterophilous GNN. The heterophilous GNNs perform closely in many cases, and if we have established heterophilous GNNs, when is rewiring a better alternative?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents theoretical and empirical connections between smoothness, homophily/heterophily, and GNN performance. Based on these analyses, authors propose REFine, a feature and label-driven diffusion approach that rewires the graph to be more homophilous. Extensive experiments are conducted on homophilous and heterophilous benchmarks to demonstrate the utility of the approach.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "1. The approach of graph rewiring for homophily/heterophily is intuitive and interesting\n2. The paper is sound, clear, and well-written.", "weaknesses": "1. **Positioning vs. prior rewiring for heterophily (novelty).**  \n  The core idea of REFine is to construct a more homophilous reference graph leveraging the graph **features + training labels** and then **adding/deleting edges**. To me, this is quite is close to **DHGR**, which also leverages similarities between node features and training labels to form the rewired graph. The similarities and differences of REFine in comparison to DHGR as well as its advantages/disadvantages are not clear. I believe the paper would benefit from a sharper distinction (beyond “simpler & guaranteed”) and **direct empirical comparisons** to DHGR/GSL baselines to make the contribution more clear. \n\n2. **Incremental theory relative to prior spectral/smoothness analyses.**  \n  The main smoothness and homophily result (Theorem 1) is very close to existing analyses that link Dirichlet energy/smoothing to homophily/heterophily. One example of where this appears in prior works is in Theorem 3 in \"Beyond Homophily in Graph Neural Networks\" by Zhu et al., 2020. The subsequent propositions in the paper also state intuitive claims (if the reference graph is more homophilous, then adding its edges/deleting complement edges improves homophily) under certain conditions. Please clarify how these results meaningfully extend prior theory and provide new insights for the limitations of prior methods or advantages of REFine.\n\n3. **Missing critical baseline and mixed empirical gains.**  \nWhile the method largely outperforms the benchmarked rewiring techniques, it is **often comparable** to specialized heterophilous GNNs, with a standout improvement on **Roman-empire** but with no major gains elsewhere. Additionally, the absence of a **DHGR** comparison weakens conclusions about superiority among heterophily-oriented rewiring methods. Including DHGR would strengthen the empirical case.\n\n4. **When to prefer rewiring over heterophilous GNNs.**  \nI am unsure as to when a practitioner should choose heterophilous GNNs or REFine. Are there any benefits of using the rewiring technique over GNNs other than just performance? If the performances are similar, why should we want to use a rewiring method over a heterophilous GNN? The paper should clarify when to choose rewiring (e.g., datasets where feature/label-driven reference graphs are reliably more homophilous), and provide guidelines on when specialized heterophilous GNNs remain the better choice.", "questions": "My questions largely stem from the limitations: \n\n1. What are the advantages and disadvantages of REFine in comparison to DHGR? What are the conceptual differences given DHGR also constructs a rewired graph based on node features similarities and train labels?\n2. How do the theoretical results meaningfully build on existing spectral analyses relating smoothness/energy to homophily/heterophily? \n3. Given the connection between oversmoothing, oversquashing, and heterophily, what are the connections between existing rewiring approaches that address oversmoothing and oversquashing to heterophily? why don't their constructions mitigate the heterophily problem?\n4. Why is DHGR not a baseline? How does DHGR perform in comparison to REFine?\n5. It would be good as to get an intuition for why we would consider using a rewiring approach over a heterophilous GNN. The heterophilous GNNs perform closely in many cases, and if we have established heterophilous GNNs, when is rewiring a better alternative?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761799850191}, {"id": "RSJGnlUUYD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24227/Reviewer_CvPR"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a label-driven diffusion approach to construct a homophilic \"reference graph.\" The goal is to rewire an original graph to improve downstream GNN performance. The paper includes extensive simulations analyzing how the homophily of both the original and reference graphs influences the final rewired graph and subsequent task performance.", "review_text": "This paper proposes a label-driven diffusion approach to construct a homophilic \"reference graph.\" The goal is to rewire an original graph to improve downstream GNN performance. The paper includes extensive simulations analyzing how the homophily of both the original and reference graphs influences the final rewired graph and subsequent task performance.", "strengths": "1. The method is inherently scalable by design, as it uses a clustering-based approach as its first step.\n2. The core idea of constructing a reference graph for label-driven rewiring is novel, yet the proposed solution is simple and elegant.", "weaknesses": "1. There is a significant disconnect between the paper's theoretical grounding and its practical implementation. The clustering step, while efficient, creates a problem:\n    - Theories (Theorem 1, Proposition 1, Proposition 2) are derived from a *global* perspective, seemingly assuming all edges can be modified.\n    - The implementation, however, performs rewiring *locally* (within clusters). The links *between* clusters are fixed and cannot be modified.\n    - This implies the theoretical statements may not hold for the final reassembled graph. The theory must be revised to account for this restriction (i.e., that a subset of edges is immutable). Similarly, the *actual* global homophily may not change as expected due to these fixed inter-cluster links.\n2. The paper's central motivation is that increasing homophily is beneficial for GNNs. However, this premise has been increasingly questioned in recent literature [1], which suggests homophily is not universally necessary or beneficial. This makes the paper's *exclusive* focus on refining homophily seem limited or dataset-specific. The authors must address this and justify their motivation in light of this conflicting research.\n3. The introduction of a \"reference graph\" inherently doubles the storage and memory cost, as both the original and reference graph structures must be maintained. This is a significant practical drawback that is not adequately discussed or justified.\n4. The method's performance seems highly dependent on the initial clustering step. How does the *balance* of these clusters (e.g., the variance in cluster sizes) influence the final rewiring and downstream GNN performance? This critical factor is not analyzed.\n\n## Minor\n\n1. **Table 1:** To improve clarity, it would be beneficial to add a row for \"Averaged Gain\" (or similar summary metric) in Table 1 to allow for a more direct comparison of the method's overall effectiveness against baselines.\n2. **Figure 1 Analysis:** The analysis in Figure 1 is insightful. Do the same patterns regarding homophily's influence hold true for the other datasets used in the paper? Please include this analysis (perhaps in Appendix B) to demonstrate the generalizability of these findings.\n\n[1] Is Homophily a Necessity for Graph Neural Networks? In ICLR, 2022", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a label-driven diffusion approach to construct a homophilic \"reference graph.\" The goal is to rewire an original graph to improve downstream GNN performance. The paper includes extensive simulations analyzing how the homophily of both the original and reference graphs influences the final rewired graph and subsequent task performance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The method is inherently scalable by design, as it uses a clustering-based approach as its first step.\n2. The core idea of constructing a reference graph for label-driven rewiring is novel, yet the proposed solution is simple and elegant.", "weaknesses": "1. There is a significant disconnect between the paper's theoretical grounding and its practical implementation. The clustering step, while efficient, creates a problem:\n    - Theories (Theorem 1, Proposition 1, Proposition 2) are derived from a *global* perspective, seemingly assuming all edges can be modified.\n    - The implementation, however, performs rewiring *locally* (within clusters). The links *between* clusters are fixed and cannot be modified.\n    - This implies the theoretical statements may not hold for the final reassembled graph. The theory must be revised to account for this restriction (i.e., that a subset of edges is immutable). Similarly, the *actual* global homophily may not change as expected due to these fixed inter-cluster links.\n2. The paper's central motivation is that increasing homophily is beneficial for GNNs. However, this premise has been increasingly questioned in recent literature [1], which suggests homophily is not universally necessary or beneficial. This makes the paper's *exclusive* focus on refining homophily seem limited or dataset-specific. The authors must address this and justify their motivation in light of this conflicting research.\n3. The introduction of a \"reference graph\" inherently doubles the storage and memory cost, as both the original and reference graph structures must be maintained. This is a significant practical drawback that is not adequately discussed or justified.\n4. The method's performance seems highly dependent on the initial clustering step. How does the *balance* of these clusters (e.g., the variance in cluster sizes) influence the final rewiring and downstream GNN performance? This critical factor is not analyzed.\n\n## Minor\n\n1. **Table 1:** To improve clarity, it would be beneficial to add a row for \"Averaged Gain\" (or similar summary metric) in Table 1 to allow for a more direct comparison of the method's overall effectiveness against baselines.\n2. **Figure 1 Analysis:** The analysis in Figure 1 is insightful. Do the same patterns regarding homophily's influence hold true for the other datasets used in the paper? Please include this analysis (perhaps in Appendix B) to demonstrate the generalizability of these findings.\n\n[1] Is Homophily a Necessity for Graph Neural Networks? In ICLR, 2022", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761706222801}, {"id": "sSQNvem5Am", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24227/Reviewer_PKvY"], "rating": 0, "soundness": 1, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "The paper addresses the issue of weak performance of GNNs in homophilous graphs. It proposes to rewire the graph to increase its homophily. First, a reference graph is constructed based mostly on node feature similarity. Then, the original graph is rewired by adding or deleting edges to make it more similar to the reference graph, which increases its homophily. Finally, models are trained on the resulting graph.", "review_text": "The paper addresses the issue of weak performance of GNNs in homophilous graphs. It proposes to rewire the graph to increase its homophily. First, a reference graph is constructed based mostly on node feature similarity. Then, the original graph is rewired by adding or deleting edges to make it more similar to the reference graph, which increases its homophily. Finally, models are trained on the resulting graph.", "strengths": "The paper is mostly well-written and easy to follow.", "weaknesses": "- The main assumption behind the paper's approach is that GNNs provide weak performance on heterophilous graphs. However, this assumption is outdated: there have been multiple paper showing that GNNs perform perfectly fine on heterophilous graphs (in particular, they typically outperform specialized models), see [1-5]. The paper does not discuss these works at all. The paper also states as its motivation: \"Standard GNNs are designed primarily for homophilic graphs, as they rely on the homophily assumption\". This is also not the case: none of the papers that developed modern GNNs mention homophily at all. Note that this models do not even have explicit access to node labels, but rather work with node features. Nowadays, any claim that GNNs are designed exclusively for homophilous graphs or do not work well on heterophilous graphs should be supported by strong evidence that directly addresses works [1-5] and shows where they went wrong (and I am not aware of any such evidence).\n\n- The motivation and statement of Theorem 1 seem confusing. Linearly separable embeddings are being discussed. However, there can be two types of linear separability for embeddings: all pairs of embeddings can be separable, or pairs of embeddings from different classes can be separable. What is needed for strong model performance is the second (intra-class) type of separability, but it seems that the paper discusses the first (all-pair) type of separability, which is not directly related to model performance. The paper claims that Theorem 1 ties GNN's ability to generate linearly separable embedding to graph homophily, but then Theorem 1 starts by assuming that the embeddings are linearly separable. It is also not clear how Theorem 1 proves some property of GNNs, when there is no GNN in the theorem. Finally, what is $A_{u,  v}$ in Theorem 1? I cannot find its definition, but it seems like $A$ denotes elements of the adjacency matrix. Then the minimum non-zero element is simply 1, because the adjacency matrix is binary (it is never mentioned that the graph is weighted, and later sections describing the method clearly assume unweighted graphs).\n\n- The experimental results are unreliable. First, the paper uses Squirrel and Chameleon datasets (among others), that have been shown in [3] to be buggy. Cornell, Texas, and Wisconsin datasets have also been criticized in [3] (in particular, the Texas dataset has a class consisting of a single node). More importantly, the results reported in the paper for GCN, GAT, H2GCN and GPRGNN are significantly lower than those reported in [3] for those datasets that are used in both works. For example, GPRGNN achieves an accuracy of 64.85 on the Roman-Empire dataset in [3], while the current paper reports only an accuracy of 20.5 for it. Most notably, however, the current paper reports accuracies of 14.8 and 14.0 for GATv2 and APPNP respectively on the Roman-Empire dataset, which is roughly the performance of the naive majority class prediction (13.96). This clearly shows that the considered baselines were not well-tuned and thus the obtained experimental results cannot be trusted.\n\n\n\n[1] Is homophily a necessity for graph neural networks? (ICLR 2022)\n\n[2] Revisiting heterophily for graph neural networks (NeurIPS 2022)\n\n[3] A critical look at the evaluation of gnns under heterophily: Are we really making progress? (ICLR 2023)\n\n[4] Characterizing Graph Datasets for Node Classification: Homophily–Heterophily Dichotomy and Beyond (NeurIPS 2023)\n\n[5] Oversmoothing, Oversquashing, Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning (arxiv preprint)", "questions": "- How is the hyperparameter k (the number of rewired edges) in the proposed method selected? If the main point is to increase graph homophily, why not simply replace the original graph with the reference graph (i.e., rewire all the edges)?\n\n- The reference graph construction is mostly based on node feature similarity. However, GNNs have access to node features. Why is it expected that the edges of the reference graph will provide GNNs with additional information?\n\n- It seems like the sources of the datasets are not mentioned in the paper. In particular, what is the Elliptic/EllipicBitcoin dataset? If it is the dataset with the same name from Pytorch Geometric, then it has strong class imbalance, and accuracy is not an appropriate metric for it (and the naive majority class prediction provides better accuracy on it then the values reported for GNNs in the paper).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the issue of weak performance of GNNs in homophilous graphs. It proposes to rewire the graph to increase its homophily. First, a reference graph is constructed based mostly on node feature similarity. Then, the original graph is rewired by adding or deleting edges to make it more similar to the reference graph, which increases its homophily. Finally, models are trained on the resulting graph.", "soundness": 1, "presentation": 3, "contribution": 1, "strengths": "The paper is mostly well-written and easy to follow.", "weaknesses": "- The main assumption behind the paper's approach is that GNNs provide weak performance on heterophilous graphs. However, this assumption is outdated: there have been multiple paper showing that GNNs perform perfectly fine on heterophilous graphs (in particular, they typically outperform specialized models), see [1-5]. The paper does not discuss these works at all. The paper also states as its motivation: \"Standard GNNs are designed primarily for homophilic graphs, as they rely on the homophily assumption\". This is also not the case: none of the papers that developed modern GNNs mention homophily at all. Note that this models do not even have explicit access to node labels, but rather work with node features. Nowadays, any claim that GNNs are designed exclusively for homophilous graphs or do not work well on heterophilous graphs should be supported by strong evidence that directly addresses works [1-5] and shows where they went wrong (and I am not aware of any such evidence).\n\n- The motivation and statement of Theorem 1 seem confusing. Linearly separable embeddings are being discussed. However, there can be two types of linear separability for embeddings: all pairs of embeddings can be separable, or pairs of embeddings from different classes can be separable. What is needed for strong model performance is the second (intra-class) type of separability, but it seems that the paper discusses the first (all-pair) type of separability, which is not directly related to model performance. The paper claims that Theorem 1 ties GNN's ability to generate linearly separable embedding to graph homophily, but then Theorem 1 starts by assuming that the embeddings are linearly separable. It is also not clear how Theorem 1 proves some property of GNNs, when there is no GNN in the theorem. Finally, what is $A_{u,  v}$ in Theorem 1? I cannot find its definition, but it seems like $A$ denotes elements of the adjacency matrix. Then the minimum non-zero element is simply 1, because the adjacency matrix is binary (it is never mentioned that the graph is weighted, and later sections describing the method clearly assume unweighted graphs).\n\n- The experimental results are unreliable. First, the paper uses Squirrel and Chameleon datasets (among others), that have been shown in [3] to be buggy. Cornell, Texas, and Wisconsin datasets have also been criticized in [3] (in particular, the Texas dataset has a class consisting of a single node). More importantly, the results reported in the paper for GCN, GAT, H2GCN and GPRGNN are significantly lower than those reported in [3] for those datasets that are used in both works. For example, GPRGNN achieves an accuracy of 64.85 on the Roman-Empire dataset in [3], while the current paper reports only an accuracy of 20.5 for it. Most notably, however, the current paper reports accuracies of 14.8 and 14.0 for GATv2 and APPNP respectively on the Roman-Empire dataset, which is roughly the performance of the naive majority class prediction (13.96). This clearly shows that the considered baselines were not well-tuned and thus the obtained experimental results cannot be trusted.\n\n\n\n[1] Is homophily a necessity for graph neural networks? (ICLR 2022)\n\n[2] Revisiting heterophily for graph neural networks (NeurIPS 2022)\n\n[3] A critical look at the evaluation of gnns under heterophily: Are we really making progress? (ICLR 2023)\n\n[4] Characterizing Graph Datasets for Node Classification: Homophily–Heterophily Dichotomy and Beyond (NeurIPS 2023)\n\n[5] Oversmoothing, Oversquashing, Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning (arxiv preprint)", "questions": "- How is the hyperparameter k (the number of rewired edges) in the proposed method selected? If the main point is to increase graph homophily, why not simply replace the original graph with the reference graph (i.e., rewire all the edges)?\n\n- The reference graph construction is mostly based on node feature similarity. However, GNNs have access to node features. Why is it expected that the edges of the reference graph will provide GNNs with additional information?\n\n- It seems like the sources of the datasets are not mentioned in the paper. In particular, what is the Elliptic/EllipicBitcoin dataset? If it is the dataset with the same name from Pytorch Geometric, then it has strong class imbalance, and accuracy is not an appropriate metric for it (and the naive majority class prediction provides better accuracy on it then the values reported for GNNs in the paper).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 0, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1760737435585}], "openreview_url": "https://openreview.net/forum?id=G9bpAoc47a", "arxiv_id": "2505.12411", "paper_pdf": "papers/G9bpAoc47a.pdf", "paper_pdf_sha256": "18cd23519a8058ee7a8f078e7f58c68036750efd9b3d82bf7d1047cdd7b05a64", "paper_pdf_bytes": 9253210, "paper_pdf_source": "openreview", "code_url": "https://github.com/harel147/REFine", "code_repository": "harel147/REFine", "code_commit": "288eef57e78a7c0030cf3ba4be905b5278eff39f", "code_archive": "repos/G9bpAoc47a.zip", "code_archive_sha256": "a48ff9c7842f79d2cbe6e5172af4b7b87380f42f0e229f020809c6e6eb4a55a4", "code_archive_bytes": 335922, "code_file_count": 35, "code_extensions": {".py": 34, ".sh": 1}, "github_disk_usage_kb": 311, "github_languages": {"Python": 186327, "Makefile": 1154, "Shell": 394}, "github_archived": false, "github_pushed_at": "2026-01-26T14:58:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/it-takes-a-graph-to-know-a-graph-rewiring-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fEEbTDoecM", "year": 2025, "status": "rejected", "title": "RL$^3$: Boosting Meta Reinforcement Learning via RL inside RL$^2$", "authors": ["Abhinav Bhatia", "Samer B. Nashed", "Shlomo Zilberstein"], "authorids": ["~Abhinav_Bhatia1", "~Samer_B._Nashed1", "~Shlomo_Zilberstein1"], "authors_source": "OpenReview API", "abstract": "Meta reinforcement learning (meta-RL) methods such as \\rlsquare have emerged as promising approaches for learning data-efficient RL algorithms tailored to a given task distribution. However, they show poor asymptotic performance and struggle with out-of-distribution tasks because they rely on sequence models, such as recurrent neural networks or transformers, to process experiences rather than summarize them using general-purpose RL components such as value functions. In contrast, traditional RL algorithms are data-inefficient as they do not use domain knowledge, but do converge to an optimal policy in the limit. We propose RL$^3$, a principled hybrid approach that incorporates action-values, learned per task via traditional RL, in the inputs to meta-RL. We show that RL$^3$ earns greater cumulative reward in the long term compared to RL$^2$ while drastically reducing meta-training time and generalizes better to out-of-distribution tasks. Experiments are conducted on both custom and benchmark discrete domains from the meta-RL literature that exhibit a range of short-term, long-term, and complex dependencies.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "gWAcjeefEY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13056/Reviewer_Lu7E"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper proposes RL^{3}, a novel method designed to enhance out-of-distribution (OOD) generalization and reduce meta-training time in meta-reinforcement learning (meta-RL) by integrating Q-estimates as additional input information for the meta-RL agent. The approach is implemented within the RL^{2} framework using transformers and is evaluated against RL^{2} across several benchmark tasks, including bandits, random MDPs, and gridworld.", "review_text": "This paper proposes RL^{3}, a novel method designed to enhance out-of-distribution (OOD) generalization and reduce meta-training time in meta-reinforcement learning (meta-RL) by integrating Q-estimates as additional input information for the meta-RL agent. The approach is implemented within the RL^{2} framework using transformers and is evaluated against RL^{2} across several benchmark tasks, including bandits, random MDPs, and gridworld.", "strengths": "- By adding Q-estimates, RL^{3} demonstrates improved efficiency over RL^{2} by requiring fewer PPO iterations and delivering better OOD performance on the selected benchmarks.", "weaknesses": "- The experimental benchmarks focus on tasks with limited state spaces, which may restrict the applicability of the findings.\n- Additional experiments on benchmarks with continuous state spaces, such as parameterized MuJoCo environments, are necessary to fully evaluate RL^{3}’s practical benefits over RL^{2} in terms of sample efficiency. The results based on tabular Q-function estimation in small state spaces do not directly indicate performance gains in more complex, real-world settings.", "questions": "1. Could cumulative reward plots over timesteps for each method be added to visualize meta-training progress?\n2. What rollout length was used for RL^{2} in the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes RL^{3}, a novel method designed to enhance out-of-distribution (OOD) generalization and reduce meta-training time in meta-reinforcement learning (meta-RL) by integrating Q-estimates as additional input information for the meta-RL agent. The approach is implemented within the RL^{2} framework using transformers and is evaluated against RL^{2} across several benchmark tasks, including bandits, random MDPs, and gridworld.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- By adding Q-estimates, RL^{3} demonstrates improved efficiency over RL^{2} by requiring fewer PPO iterations and delivering better OOD performance on the selected benchmarks.", "weaknesses": "- The experimental benchmarks focus on tasks with limited state spaces, which may restrict the applicability of the findings.\n- Additional experiments on benchmarks with continuous state spaces, such as parameterized MuJoCo environments, are necessary to fully evaluate RL^{3}’s practical benefits over RL^{2} in terms of sample efficiency. The results based on tabular Q-function estimation in small state spaces do not directly indicate performance gains in more complex, real-world settings.", "questions": "1. Could cumulative reward plots over timesteps for each method be added to visualize meta-training progress?\n2. What rollout length was used for RL^{2} in the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731099337348}, {"id": "3xQxsJhTIx", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13056/Reviewer_E1NK"], "rating": 5, "soundness": 2, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "The paper presents a modification of meta-RL that incorporates additional information in the form of Q-estimates and state counts (and possibly other things as well) in order to improve training times, data requirements and possibly asymptomatic performance.  The approach is motivated with both empirical and theoretical arguments and validated on three artificial test domains.", "review_text": "The paper presents a modification of meta-RL that incorporates additional information in the form of Q-estimates and state counts (and possibly other things as well) in order to improve training times, data requirements and possibly asymptomatic performance.  The approach is motivated with both empirical and theoretical arguments and validated on three artificial test domains.", "strengths": "1. The paper is *very* well-presented.  It was easy to understand and enjoyable to read.\n2. The work is well-situated in the literature.\n3. The paper provides both empirical and theoretical analyses that both motivate and justify the approach and design decisions.\n4. Reported results are good and convince the reader that the proposed approach does provide (some of) the improvements hypothesized.  In particular, rewards are comparable or better than RL^2 while (meta)training time is decreased and OOD generalization seems also to be improved to some extent.", "weaknesses": "1. Experimental results are obtained on only toy problems.  It seems likely that the approach may not scale to more difficult/larger/real-world problems.  How does this work on a modestly more difficult domain like Atari, for example?\n2. Some claims are not really supported.  For example, while the following statement from Sec. 5 seems like it could be true, there is no supporting evidence given: \"VAMDPs can be plugged into any base meta-RL algorithm with a reasonable expectation of improving it.\"  As another example, it is also suggested that RL^3 might enjoy convergence guarantees, but this is also not supported in any way.\n3. Performance improvement results seem like they don't include the overhead for computing q-values and state counts and are therefore potentially somewhat misleading.  This is addressed to some degree in the Sec 6, \"Computation Overhead Considerations\" subsection, but the discussion is too brief to be convincing; in particular, it is not clear that the claim will hold if scaled to larger/real-world problems.\n4. The paper doesn't compare RL^3 with any other meta-RL approaches, other than RL^2, the one it modifies.", "questions": "1. Several places in the paper seem to suggest that one of the benefits of RL^3 will be asymptotic guarantees of convergence.  Does RL^3 in fact, gain asymptotic performance guarantees?  \n\n2. In Section 4.2 it states, \"In practice, we provide action advantages along with the max Q-value (value function) and optionally, standard errors, instead of just Q-estimates\".  How critical is this to RL^3's success?  How much does this affect the results?\n\n3. Sec 5 talks about using PPO as the metalearner, but earlier the paper talks about using a blackbox/NN for the metalearner.  I'm a bit confused here---is PPO used just for augmented RL^2?\n\n4. Is Fig 3b comparing only meta-training time (as the last sentence of the MDP results subsection seems to suggest)? Or does it include all required computation?  How are the q-values and counts obtained? How much does this cost? How would this scale for larger/real-world problems?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a modification of meta-RL that incorporates additional information in the form of Q-estimates and state counts (and possibly other things as well) in order to improve training times, data requirements and possibly asymptomatic performance.  The approach is motivated with both empirical and theoretical arguments and validated on three artificial test domains.", "soundness": 2, "presentation": 4, "contribution": 2, "strengths": "1. The paper is *very* well-presented.  It was easy to understand and enjoyable to read.\n2. The work is well-situated in the literature.\n3. The paper provides both empirical and theoretical analyses that both motivate and justify the approach and design decisions.\n4. Reported results are good and convince the reader that the proposed approach does provide (some of) the improvements hypothesized.  In particular, rewards are comparable or better than RL^2 while (meta)training time is decreased and OOD generalization seems also to be improved to some extent.", "weaknesses": "1. Experimental results are obtained on only toy problems.  It seems likely that the approach may not scale to more difficult/larger/real-world problems.  How does this work on a modestly more difficult domain like Atari, for example?\n2. Some claims are not really supported.  For example, while the following statement from Sec. 5 seems like it could be true, there is no supporting evidence given: \"VAMDPs can be plugged into any base meta-RL algorithm with a reasonable expectation of improving it.\"  As another example, it is also suggested that RL^3 might enjoy convergence guarantees, but this is also not supported in any way.\n3. Performance improvement results seem like they don't include the overhead for computing q-values and state counts and are therefore potentially somewhat misleading.  This is addressed to some degree in the Sec 6, \"Computation Overhead Considerations\" subsection, but the discussion is too brief to be convincing; in particular, it is not clear that the claim will hold if scaled to larger/real-world problems.\n4. The paper doesn't compare RL^3 with any other meta-RL approaches, other than RL^2, the one it modifies.", "questions": "1. Several places in the paper seem to suggest that one of the benefits of RL^3 will be asymptotic guarantees of convergence.  Does RL^3 in fact, gain asymptotic performance guarantees?  \n\n2. In Section 4.2 it states, \"In practice, we provide action advantages along with the max Q-value (value function) and optionally, standard errors, instead of just Q-estimates\".  How critical is this to RL^3's success?  How much does this affect the results?\n\n3. Sec 5 talks about using PPO as the metalearner, but earlier the paper talks about using a blackbox/NN for the metalearner.  I'm a bit confused here---is PPO used just for augmented RL^2?\n\n4. Is Fig 3b comparing only meta-training time (as the last sentence of the MDP results subsection seems to suggest)? Or does it include all required computation?  How are the q-values and counts obtained? How much does this cost? How would this scale for larger/real-world problems?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730878820604}, {"id": "pjVtuym21x", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13056/Reviewer_5dhT"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper presents RL3, a novel approach to meta reinforcement learning that combines traditional RL techniques with meta-RL methods. The authors propose incorporating Q-value estimates from traditional RL as additional inputs to a meta-RL architecture, aiming to improve long-term performance and out-of-distribution generalization while maintaining short-term efficiency.", "review_text": "This paper presents RL3, a novel approach to meta reinforcement learning that combines traditional RL techniques with meta-RL methods. The authors propose incorporating Q-value estimates from traditional RL as additional inputs to a meta-RL architecture, aiming to improve long-term performance and out-of-distribution generalization while maintaining short-term efficiency.", "strengths": "- Novel approach: The paper introduces an innovative method that combines the strengths of traditional RL and meta-RL, potentially addressing some limitations of existing meta-RL approaches.\n- Theoretical foundation: The authors provide theoretical insights into why incorporating Q-value estimates can be beneficial, linking them to the optimal meta-value function.\n- Improved performance: RL3 demonstrates better long-term returns and out-of-distribution generalization than RL2, while maintaining short-term efficiency.\n- Flexibility: The approach can be applied to various meta-RL algorithms, not just RL2.", "weaknesses": "- Limited baselines: The paper would benefit from comparing RL3 to more state-of-the-art meta-RL approaches beyond just RL2. In particular, comparing RL3 to hypernetwork-based approaches would provide valuable insights into its relative performance.\n- Scope of experiments: The experiments are limited to discrete domains. As suggested, it would be valuable to see how RL3 performs in high-dimensional state spaces and with continuous action spaces. This limitation restricts the applicability of the current results to more complex, real-world scenarios.\n- Compute budget dependency: The performance of RL3 relative to RL2 seems to depend on the compute budget (H), which appears to be environment-dependent. The paper would benefit from providing guidelines or heuristics for selecting an appropriate range of H values for different environments.\n- Generalizability: While the paper demonstrates improved out-of-distribution generalization, it's unclear how well this generalizes across a wider range of task distributions or more complex domains.\n- Computational overhead: The paper doesn't thoroughly discuss the potential computational overhead of computing Q-value estimates alongside the meta-RL process, which could be a significant factor in practical applications with high dimensional state spaces.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents RL3, a novel approach to meta reinforcement learning that combines traditional RL techniques with meta-RL methods. The authors propose incorporating Q-value estimates from traditional RL as additional inputs to a meta-RL architecture, aiming to improve long-term performance and out-of-distribution generalization while maintaining short-term efficiency.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Novel approach: The paper introduces an innovative method that combines the strengths of traditional RL and meta-RL, potentially addressing some limitations of existing meta-RL approaches.\n- Theoretical foundation: The authors provide theoretical insights into why incorporating Q-value estimates can be beneficial, linking them to the optimal meta-value function.\n- Improved performance: RL3 demonstrates better long-term returns and out-of-distribution generalization than RL2, while maintaining short-term efficiency.\n- Flexibility: The approach can be applied to various meta-RL algorithms, not just RL2.", "weaknesses": "- Limited baselines: The paper would benefit from comparing RL3 to more state-of-the-art meta-RL approaches beyond just RL2. In particular, comparing RL3 to hypernetwork-based approaches would provide valuable insights into its relative performance.\n- Scope of experiments: The experiments are limited to discrete domains. As suggested, it would be valuable to see how RL3 performs in high-dimensional state spaces and with continuous action spaces. This limitation restricts the applicability of the current results to more complex, real-world scenarios.\n- Compute budget dependency: The performance of RL3 relative to RL2 seems to depend on the compute budget (H), which appears to be environment-dependent. The paper would benefit from providing guidelines or heuristics for selecting an appropriate range of H values for different environments.\n- Generalizability: While the paper demonstrates improved out-of-distribution generalization, it's unclear how well this generalizes across a wider range of task distributions or more complex domains.\n- Computational overhead: The paper doesn't thoroughly discuss the potential computational overhead of computing Q-value estimates alongside the meta-RL process, which could be a significant factor in practical applications with high dimensional state spaces.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730679456452}, {"id": "2PdUmTk4pe", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13056/Reviewer_eaSY"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper proposes a novel hybrid approach, RL3, that integrates value-augmented MDP into meta reinforcement learning (meta-RL) frameworks RL2. The authors aim to address limitations in current meta-RL models, such as poor asymptotic performance and weak out-of-distribution (OOD) generalization. RL3 introduces object-level Q-value estimates computed using traditional RL, which are integrated into the meta-RL model to improve cumulative reward, speed up meta-training, and enhance OOD generalization.", "review_text": "The paper proposes a novel hybrid approach, RL3, that integrates value-augmented MDP into meta reinforcement learning (meta-RL) frameworks RL2. The authors aim to address limitations in current meta-RL models, such as poor asymptotic performance and weak out-of-distribution (OOD) generalization. RL3 introduces object-level Q-value estimates computed using traditional RL, which are integrated into the meta-RL model to improve cumulative reward, speed up meta-training, and enhance OOD generalization.", "strengths": "1) RL3 may reduces meta-training time, which has potential practical implications for scaling meta-RL systems.\n\n2) The authors provide theoretical foundation, including proofs that support the efficacy of incorporating Q-value estimates into meta-RL, explaining the potential advantages in handling OOD tasks and long-term planning.", "weaknesses": "1) The author’s understanding of RL2 appears to remain questionable. Specifically, the claim made in the discussion of RL3 suggests that \n\"For object-level RL, we use model estimation followed by value iteration (with discount factor\nγ = 1) to obtain Q-estimates.\"\nHowever, in reality, RL2 also employs Q-value estimation in a similar manner to predict the future expectation of reward using the discount factor γ. Specifically, both RL2 and RL3 optimize the same equation mentioned as Equation (1) in the paper. They both employ the idea of optimizing across multiple episodes simultaneously, and there does not appear to be a fundamental difference between the two approaches.\n\nThis similarity raises concerns about the novelty of the author's argument regarding RL3's distinct capabilities.\n\n2) Authors mentioned \"To test OOD generalization, we vary parameters including the stochasticity of actions, density of obstacles, and the number of dangerous tiles.\" Although this level of variation is commonly considered a new task in meta-learning settings, the actual changes introduced are relatively minor. Furthermore, GridWorld is a relatively simple environment. \n\nWhether you have plans to test on more complex domains in future work, such as vision-based RL tasks or robotic control tasks (for example DMC or Atari)? \n\n3) The authors did not provide a clear definition of the RL3 algorithm in a formalized manner, nor did they present a direct comparison with the formulation of the RL2 algorithm. This lack of precise algorithmic definitions makes it difficult for readers to fully understand how RL3 operates and differs from RL2.\n\nCould authors  include a side-by-side comparison with RL2 to highlight the key differences.\n\n4) I not sure that simply feeding the entire sequence of states and actions from a whole task continuously into an LSTM or transformer, as done in RL3, will necessarily enhance OOD generalization and long-term performance. A more promising approach would involve training on trajectories of both short and long lengths. Whether authors have considered alternative methods combining varying trajectory lengths.", "questions": "1) What is a VAMDP? Is there a clear and formal definition?\n\n2) Could the authors outline the fundamental differences in the implementation of RL2 and RL3? Are there any precise mathematical formulations that highlight these differences? As presented, Equation 1 only expresses the meta-learning objective but does not distinguish RL2 from RL3.\n\n3) Why is RL2 unable to address OOD generalization problems? Additionally, why does RL2, compared to RL, lack long-term performance? If RL methods have OOD generalization capabilities, why does RL2 not inherit them?\n\n4) What is the difference between the Q-value in RL, RL2, and the Q-estimates used in RL3?\n\n5) In Figure 3b, why are the performance results of RL2 and RL3 equal when the budget is set to 128? Which specific setting in the experiments is responsible for this outcome?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel hybrid approach, RL3, that integrates value-augmented MDP into meta reinforcement learning (meta-RL) frameworks RL2. The authors aim to address limitations in current meta-RL models, such as poor asymptotic performance and weak out-of-distribution (OOD) generalization. RL3 introduces object-level Q-value estimates computed using traditional RL, which are integrated into the meta-RL model to improve cumulative reward, speed up meta-training, and enhance OOD generalization.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1) RL3 may reduces meta-training time, which has potential practical implications for scaling meta-RL systems.\n\n2) The authors provide theoretical foundation, including proofs that support the efficacy of incorporating Q-value estimates into meta-RL, explaining the potential advantages in handling OOD tasks and long-term planning.", "weaknesses": "1) The author’s understanding of RL2 appears to remain questionable. Specifically, the claim made in the discussion of RL3 suggests that \n\"For object-level RL, we use model estimation followed by value iteration (with discount factor\nγ = 1) to obtain Q-estimates.\"\nHowever, in reality, RL2 also employs Q-value estimation in a similar manner to predict the future expectation of reward using the discount factor γ. Specifically, both RL2 and RL3 optimize the same equation mentioned as Equation (1) in the paper. They both employ the idea of optimizing across multiple episodes simultaneously, and there does not appear to be a fundamental difference between the two approaches.\n\nThis similarity raises concerns about the novelty of the author's argument regarding RL3's distinct capabilities.\n\n2) Authors mentioned \"To test OOD generalization, we vary parameters including the stochasticity of actions, density of obstacles, and the number of dangerous tiles.\" Although this level of variation is commonly considered a new task in meta-learning settings, the actual changes introduced are relatively minor. Furthermore, GridWorld is a relatively simple environment. \n\nWhether you have plans to test on more complex domains in future work, such as vision-based RL tasks or robotic control tasks (for example DMC or Atari)? \n\n3) The authors did not provide a clear definition of the RL3 algorithm in a formalized manner, nor did they present a direct comparison with the formulation of the RL2 algorithm. This lack of precise algorithmic definitions makes it difficult for readers to fully understand how RL3 operates and differs from RL2.\n\nCould authors  include a side-by-side comparison with RL2 to highlight the key differences.\n\n4) I not sure that simply feeding the entire sequence of states and actions from a whole task continuously into an LSTM or transformer, as done in RL3, will necessarily enhance OOD generalization and long-term performance. A more promising approach would involve training on trajectories of both short and long lengths. Whether authors have considered alternative methods combining varying trajectory lengths.", "questions": "1) What is a VAMDP? Is there a clear and formal definition?\n\n2) Could the authors outline the fundamental differences in the implementation of RL2 and RL3? Are there any precise mathematical formulations that highlight these differences? As presented, Equation 1 only expresses the meta-learning objective but does not distinguish RL2 from RL3.\n\n3) Why is RL2 unable to address OOD generalization problems? Additionally, why does RL2, compared to RL, lack long-term performance? If RL methods have OOD generalization capabilities, why does RL2 not inherit them?\n\n4) What is the difference between the Q-value in RL, RL2, and the Q-estimates used in RL3?\n\n5) In Figure 3b, why are the performance results of RL2 and RL3 equal when the budget is set to 128? Which specific setting in the experiments is responsible for this outcome?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730560249318}, {"id": "6DLiyKfLL5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13056/Reviewer_WXJJ"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "Discerning that RL$^2$ shows poor asymptotic and ood performance while general RL demonstrates short-term efficiency, this paper constructively combines the general RL with the previous meta-RL algorithm RL$^2$ and calls it RL$^3$. To be specific, the paper seeks to add an extra signal as input, namely the action-value function. Though seems to be straightforward, adding this input would give a useful compression compared to trajectory history. To validate the effectiveness, the paper provides the theoretical proof for recovering the optimal meta-level value function. Finally, some experiments on discrete domains are performed.", "review_text": "Discerning that RL$^2$ shows poor asymptotic and ood performance while general RL demonstrates short-term efficiency, this paper constructively combines the general RL with the previous meta-RL algorithm RL$^2$ and calls it RL$^3$. To be specific, the paper seeks to add an extra signal as input, namely the action-value function. Though seems to be straightforward, adding this input would give a useful compression compared to trajectory history. To validate the effectiveness, the paper provides the theoretical proof for recovering the optimal meta-level value function. Finally, some experiments on discrete domains are performed.", "strengths": "1. The paper constructively combines the previous meta-RL algorithms called RL$^2$ and general RL signal to absorb the advantage and avoid the disadvantage of both sides.\n2. The paper adopts extra action value as the auxiliary input and provides some theoretical evidence to show rationality.\n3. Compared to RL$^2$, which uses the history trajectory rather than the RL$^3$ using the action value, the latter can learn more efficiently and easily. I appreciate the authors' thorough explanation.", "weaknesses": "1. The biggest concern for me is that as shown in Appendix A.4, $\\textbf{based on the extra input action-value}$, the iteration process would lead to two cases either the belief distribution collapses rapidly to zero for tasks $j\\ne i$ or the belief distribution will reduce to zero for tasks not compatible with the observed Q-values. $\\textbf{Does this mean that whatever is capable of distinguishing tasks as input should meet this proof and the action value is only a feasible case?}$ If so, what is the advantage of this over adopting the task representation as input like the context-based meta-RL algorithms [1] ? Note that these algorithms can also hold a good compression w.r.t the transition / trajectory. Also, these algorithms can resolve the continuous domain rather than be limited to the discrete domain.\n2. In my view, why RL$^3$ holds better OOD performance can not be interpreted by the theoretical justification.\n3. RL$^2$ is a very early meta-RL algorithm, incrementally building upon RL$^2$ may still have a large gap with SOTA algorithms. As some context-based meta-RL algorithms hold the SOTA, I believe the authors can move RL$^3$ onto the context-based meta-RL algorithms to show further superiority or at least add some comparison.\n4. This is related to problem 3, some lately-published papers regarding advanced meta RL should be discussed.\n5. Some unclear definitions and typos: what are object-level and meta-level? What is Figure b) trying to show? Why does any permutation of transitions yield the same Q-estimates? For testing OOD generalization, MDPs are generated from distributions with different parameters than in training, what does the \"parameters\" mean here? How do the authors generate the OOD data? VeriBAD -> VariBAD\n6. Unclear pseudo-code: Does the $ONEHOT(s)\\cdot Q(s) \\cdot N(s)$ denote the results after concatenation?\n\n[1] Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning. Lanqing Li, et.al.\n\n$\\textbf{I would be happy to raise the score if the above problems are properly addressed.}$", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Discerning that RL$^2$ shows poor asymptotic and ood performance while general RL demonstrates short-term efficiency, this paper constructively combines the general RL with the previous meta-RL algorithm RL$^2$ and calls it RL$^3$. To be specific, the paper seeks to add an extra signal as input, namely the action-value function. Though seems to be straightforward, adding this input would give a useful compression compared to trajectory history. To validate the effectiveness, the paper provides the theoretical proof for recovering the optimal meta-level value function. Finally, some experiments on discrete domains are performed.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper constructively combines the previous meta-RL algorithms called RL$^2$ and general RL signal to absorb the advantage and avoid the disadvantage of both sides.\n2. The paper adopts extra action value as the auxiliary input and provides some theoretical evidence to show rationality.\n3. Compared to RL$^2$, which uses the history trajectory rather than the RL$^3$ using the action value, the latter can learn more efficiently and easily. I appreciate the authors' thorough explanation.", "weaknesses": "1. The biggest concern for me is that as shown in Appendix A.4, $\\textbf{based on the extra input action-value}$, the iteration process would lead to two cases either the belief distribution collapses rapidly to zero for tasks $j\\ne i$ or the belief distribution will reduce to zero for tasks not compatible with the observed Q-values. $\\textbf{Does this mean that whatever is capable of distinguishing tasks as input should meet this proof and the action value is only a feasible case?}$ If so, what is the advantage of this over adopting the task representation as input like the context-based meta-RL algorithms [1] ? Note that these algorithms can also hold a good compression w.r.t the transition / trajectory. Also, these algorithms can resolve the continuous domain rather than be limited to the discrete domain.\n2. In my view, why RL$^3$ holds better OOD performance can not be interpreted by the theoretical justification.\n3. RL$^2$ is a very early meta-RL algorithm, incrementally building upon RL$^2$ may still have a large gap with SOTA algorithms. As some context-based meta-RL algorithms hold the SOTA, I believe the authors can move RL$^3$ onto the context-based meta-RL algorithms to show further superiority or at least add some comparison.\n4. This is related to problem 3, some lately-published papers regarding advanced meta RL should be discussed.\n5. Some unclear definitions and typos: what are object-level and meta-level? What is Figure b) trying to show? Why does any permutation of transitions yield the same Q-estimates? For testing OOD generalization, MDPs are generated from distributions with different parameters than in training, what does the \"parameters\" mean here? How do the authors generate the OOD data? VeriBAD -> VariBAD\n6. Unclear pseudo-code: Does the $ONEHOT(s)\\cdot Q(s) \\cdot N(s)$ denote the results after concatenation?\n\n[1] Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning. Lanqing Li, et.al.\n\n$\\textbf{I would be happy to raise the score if the above problems are properly addressed.}$", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729153625390}], "openreview_url": "https://openreview.net/forum?id=fEEbTDoecM", "arxiv_id": "2306.15909", "paper_pdf": "papers/fEEbTDoecM.pdf", "paper_pdf_sha256": "ef7644aba83f86aa707328f336b05b3a741b0482f5b5b696eb57dde99f1ed242", "paper_pdf_bytes": 5781210, "paper_pdf_source": "openreview", "code_url": "https://github.com/bhatiaabhinav/RL3", "code_repository": "bhatiaabhinav/RL3", "code_commit": "0ae3447be0d25a268df17391e6e279d11402c127", "code_archive": "repos/fEEbTDoecM.zip", "code_archive_sha256": "da0856caf38f058dfd30efda6c7bb21239ccc4015460536eb8f1fd4443d2313c", "code_archive_bytes": 50491, "code_file_count": 6, "code_extensions": {".jl": 6}, "github_disk_usage_kb": 48, "github_languages": {"Julia": 72661}, "github_archived": false, "github_pushed_at": "2024-12-08T16:11:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rl-3-boosting-meta-reinforcement-learning-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WnEnU2K3Rb", "year": 2024, "status": "rejected", "title": "Beyond the Benchmark: Detecting Diverse Anomalies in Videos", "authors": ["Yoav Arad", "Michael Werman"], "authorids": ["~Yoav_Arad1", "~Michael_Werman1"], "authors_source": "OpenReview API", "abstract": "Video Anomaly Detection (VAD) plays a crucial role in modern surveillance systems, aiming to  identify various anomalies in real-world situations. However, current benchmark datasets predominantly emphasize simple, single-frame anomalies such as novel object detection. This narrow focus restricts the advancement of VAD models. In this research, we advocate for an expansion of VAD investigations to encompass intricate anomalies that extend beyond conventional benchmark boundaries.\nTo facilitate this, we introduce two  datasets, HMDB-AD and HMDB-Violence, to challenge models with diverse action-based anomalies. These datasets are derived from the HMDB51 action recognition dataset.\nWe further present Multi-Frame Anomaly Detection (MFAD), a novel method built upon the AI-VAD framework. \nAI-VAD utilizes single-frame features such as pose estimation and deep image encoding, and two-frame features such as object velocity. They then apply a density estimation algorithm to compute anomaly scores. To address complex multi-frame anomalies, we add a deep video encoding features capturing long-range temporal dependencies, and logistic regression  to enhance final score calculation.\nExperimental results confirm our assumptions, highlighting existing models limitations with new anomaly types. MFAD excels in both simple and complex anomaly detection scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "CysXj5mheY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2475/Reviewer_y6mV"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The manuscript underscores the importance of Video Anomaly Detection (VAD) in surveillance systems. It criticizes the current focus on simple, single-frame anomalies in benchmark datasets and advocates for expanding the scope of VAD to intricate anomalies. The authors introduce two datasets, HMDB-AD and HMDB-Violence, to challenge models with diverse action-based anomalies, and present Multi-Frame Anomaly Detection (MFAD). MFAD builds upon the AI-VAD framework, incorporating single-frame and two-frame features and applying density estimation. To tackle complex multi-frame anomalies, deep video encoding, and logistic regression are added. Experimental results highlight limitations in existing models with new anomaly types, demonstrating MFAD's proficiency in both simple and complex anomaly detection scenarios.", "review_text": "The manuscript underscores the importance of Video Anomaly Detection (VAD) in surveillance systems. It criticizes the current focus on simple, single-frame anomalies in benchmark datasets and advocates for expanding the scope of VAD to intricate anomalies. The authors introduce two datasets, HMDB-AD and HMDB-Violence, to challenge models with diverse action-based anomalies, and present Multi-Frame Anomaly Detection (MFAD). MFAD builds upon the AI-VAD framework, incorporating single-frame and two-frame features and applying density estimation. To tackle complex multi-frame anomalies, deep video encoding, and logistic regression are added. Experimental results highlight limitations in existing models with new anomaly types, demonstrating MFAD's proficiency in both simple and complex anomaly detection scenarios.", "strengths": "Strengths of the MFAD approach:\n\n++ Comprehensive Feature Extraction: MFAD extracts four diverse feature types, including object velocities, human pose estimations, deep image encodings, and deep video encodings, enabling a holistic analysis of video data.\n\n++ Adaptive Density Score Calculation: Using Gaussian Mixture Models (GMM) for velocity features and k-nearest neighbors (kNN) for other high-dimensional features, it adapts the density score calculation to the nature of the features, enhancing anomaly detection accuracy.\n\n++ Max Feature Aggregation: The addition of the 'max' feature, which aggregates maximum feature scores per frame, adds value to the approach, improving anomaly detection.\n\n++ Gaussian Smoothing: The application of Gaussian smoothing to anomaly scores reduces noise and provides more stable and interpretable results.\n\nOverall, MFAD's strengths lie in its feature diversity, multi-modal analysis, adaptive density scoring, effective feature fusion, supervised learning, and robust experimental design, making it a powerful method for detecting both simple and complex anomalies in video data.", "weaknesses": "Looking at the manuscript, weaknesses are provided below. \n\n-- Complexity: MFAD's multi-stage process and diverse feature extraction can make it computationally demanding and challenging to implement in resource-constrained environments.\n\n-- Model Specificity: Utilizing specific video foundation models may reduce adaptability to different datasets or domains.\n\n-- Not Real-Time: Computationally intensive and a requirement for separate training/testing data make real-time application challenging.\n\n-- Gaussian Smoothing Limitation: Applying Gaussian smoothing may not suit all anomaly patterns, potentially leading to information loss.", "questions": "There are just a couple of questions that I need clarification on!! \n\n--> How does MFAD handle the challenge of real-time video anomaly detection given its computational complexity? As I can see the author has provided the note for reproducibility, some insights would be helpful to understand the scope. \n--> Can MFAD adapt to different video datasets and domains effectively, or is it limited by its reliance on specific video foundation models?\n\nCurrently, I'm leaning towards accepting this work, if the Authors can provide some insights into the weakness & questions section, that would be helpful to understand the significant contribution.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The manuscript underscores the importance of Video Anomaly Detection (VAD) in surveillance systems. It criticizes the current focus on simple, single-frame anomalies in benchmark datasets and advocates for expanding the scope of VAD to intricate anomalies. The authors introduce two datasets, HMDB-AD and HMDB-Violence, to challenge models with diverse action-based anomalies, and present Multi-Frame Anomaly Detection (MFAD). MFAD builds upon the AI-VAD framework, incorporating single-frame and two-frame features and applying density estimation. To tackle complex multi-frame anomalies, deep video encoding, and logistic regression are added. Experimental results highlight limitations in existing models with new anomaly types, demonstrating MFAD's proficiency in both simple and complex anomaly detection scenarios.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Strengths of the MFAD approach:\n\n++ Comprehensive Feature Extraction: MFAD extracts four diverse feature types, including object velocities, human pose estimations, deep image encodings, and deep video encodings, enabling a holistic analysis of video data.\n\n++ Adaptive Density Score Calculation: Using Gaussian Mixture Models (GMM) for velocity features and k-nearest neighbors (kNN) for other high-dimensional features, it adapts the density score calculation to the nature of the features, enhancing anomaly detection accuracy.\n\n++ Max Feature Aggregation: The addition of the 'max' feature, which aggregates maximum feature scores per frame, adds value to the approach, improving anomaly detection.\n\n++ Gaussian Smoothing: The application of Gaussian smoothing to anomaly scores reduces noise and provides more stable and interpretable results.\n\nOverall, MFAD's strengths lie in its feature diversity, multi-modal analysis, adaptive density scoring, effective feature fusion, supervised learning, and robust experimental design, making it a powerful method for detecting both simple and complex anomalies in video data.", "weaknesses": "Looking at the manuscript, weaknesses are provided below. \n\n-- Complexity: MFAD's multi-stage process and diverse feature extraction can make it computationally demanding and challenging to implement in resource-constrained environments.\n\n-- Model Specificity: Utilizing specific video foundation models may reduce adaptability to different datasets or domains.\n\n-- Not Real-Time: Computationally intensive and a requirement for separate training/testing data make real-time application challenging.\n\n-- Gaussian Smoothing Limitation: Applying Gaussian smoothing may not suit all anomaly patterns, potentially leading to information loss.", "questions": "There are just a couple of questions that I need clarification on!! \n\n--> How does MFAD handle the challenge of real-time video anomaly detection given its computational complexity? As I can see the author has provided the note for reproducibility, some insights would be helpful to understand the scope. \n--> Can MFAD adapt to different video datasets and domains effectively, or is it limited by its reliance on specific video foundation models?\n\nCurrently, I'm leaning towards accepting this work, if the Authors can provide some insights into the weakness & questions section, that would be helpful to understand the significant contribution.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698818562971}, {"id": "iFBXqr1Yh6", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2475/Reviewer_HVoz"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "1 poor", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "Briefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\nThis paper proposes a method for video anomaly detection that goes beyond the limitations of current benchmark datasets. The authors introduce two new datasets, HMDB-AD and HMDB-Violence, which challenge models with diverse action-based anomalies. They also present a novel method called Multi-Frame Anomaly Detection (MFAD) that incorporates deep video encoding features to capture long-range temporal dependencies and logistic regression to enhance the final score calculation. The experimental results show that MFAD outperforms existing methods on both simple and complex anomaly detection scenarios.", "review_text": "Briefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\nThis paper proposes a method for video anomaly detection that goes beyond the limitations of current benchmark datasets. The authors introduce two new datasets, HMDB-AD and HMDB-Violence, which challenge models with diverse action-based anomalies. They also present a novel method called Multi-Frame Anomaly Detection (MFAD) that incorporates deep video encoding features to capture long-range temporal dependencies and logistic regression to enhance the final score calculation. The experimental results show that MFAD outperforms existing methods on both simple and complex anomaly detection scenarios.", "strengths": "(1) The paper addresses the limitation of current benchmark datasets for video anomaly detection and proposes two new datasets that allow for the detection of complex action-based anomalies. This expands the scope of what constitutes an anomaly and encourages further research on more comprehensive anomaly types.\n(2) The proposed method, MFAD, simply incorporates deep video encoding features and logistic regression to effectively detect both simple and complex anomalies. The experimental results demonstrate the effectiveness of the method on benchmark datasets as well as the newly introduced datasets.\n(3) The paper is well-structured and clear. Based on the two datasets proposed in the paper, the method used performs better than existing methods.", "weaknesses": "(1) The paper lacks a more detailed description of the datasets HMDB-AD and HMDB-Violence. It would be beneficial to provide more information on the distribution of normal and abnormal activities, and any specific challenges or characteristics of the datasets.\n(2) The method proposed in this paper is more like a simple patchwork combination that lacks sound and rigorous theoretical support. Moreover, the paper lacks a more detailed and visual explanation of the proposed method.\n(3) The article lacks experimental validation of the effectiveness of the various components of the method. For example, the effect of redundant background information in deep image encodings was not verified.\n(4) The article also lacks a comprehensive analysis of the method's limitations, which would have facilitated a discussion of any potential challenges or failures. For example, the sub-optimal performance of the method proposed on the STC and Avenue datasets and the reasons for this should be analysed.", "questions": "1. How did you distinguish between normal and abnormal activity in multiple scenarios when constructing the two new datasets? What criteria were used?\n2. How did you extract the human pose estimation, object velocity and depth image coding, can you provide more details on these?\n3. How did you synthesise both few-frame and multi-frame features?\n4. What are the limitations or failure cases of the MFAD method?\n5. How did you verify the impact of each extracted feature on the results?\n6. What are the limitations or failure cases of the MFAD method? Why the proposed method is sub-optimal for experiments on the STC and Avenue datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Briefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\nThis paper proposes a method for video anomaly detection that goes beyond the limitations of current benchmark datasets. The authors introduce two new datasets, HMDB-AD and HMDB-Violence, which challenge models with diverse action-based anomalies. They also present a novel method called Multi-Frame Anomaly Detection (MFAD) that incorporates deep video encoding features to capture long-range temporal dependencies and logistic regression to enhance the final score calculation. The experimental results show that MFAD outperforms existing methods on both simple and complex anomaly detection scenarios.", "soundness": "1 poor", "presentation": "1 poor", "contribution": "2 fair", "strengths": "(1) The paper addresses the limitation of current benchmark datasets for video anomaly detection and proposes two new datasets that allow for the detection of complex action-based anomalies. This expands the scope of what constitutes an anomaly and encourages further research on more comprehensive anomaly types.\n(2) The proposed method, MFAD, simply incorporates deep video encoding features and logistic regression to effectively detect both simple and complex anomalies. The experimental results demonstrate the effectiveness of the method on benchmark datasets as well as the newly introduced datasets.\n(3) The paper is well-structured and clear. Based on the two datasets proposed in the paper, the method used performs better than existing methods.", "weaknesses": "(1) The paper lacks a more detailed description of the datasets HMDB-AD and HMDB-Violence. It would be beneficial to provide more information on the distribution of normal and abnormal activities, and any specific challenges or characteristics of the datasets.\n(2) The method proposed in this paper is more like a simple patchwork combination that lacks sound and rigorous theoretical support. Moreover, the paper lacks a more detailed and visual explanation of the proposed method.\n(3) The article lacks experimental validation of the effectiveness of the various components of the method. For example, the effect of redundant background information in deep image encodings was not verified.\n(4) The article also lacks a comprehensive analysis of the method's limitations, which would have facilitated a discussion of any potential challenges or failures. For example, the sub-optimal performance of the method proposed on the STC and Avenue datasets and the reasons for this should be analysed.", "questions": "1. How did you distinguish between normal and abnormal activity in multiple scenarios when constructing the two new datasets? What criteria were used?\n2. How did you extract the human pose estimation, object velocity and depth image coding, can you provide more details on these?\n3. How did you synthesise both few-frame and multi-frame features?\n4. What are the limitations or failure cases of the MFAD method?\n5. How did you verify the impact of each extracted feature on the results?\n6. What are the limitations or failure cases of the MFAD method? Why the proposed method is sub-optimal for experiments on the STC and Avenue datasets?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed.", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698678505083}, {"id": "qzJbeDuUAI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2475/Reviewer_A5Te"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes a multi-frame-based video anomaly detection method, that builds on top of [1]. In [1] mostly frame-level attributes are included, while the proposed method extends the method of [1] by including multi-frames encoding features extracted across 16 frames. \n\nFurthermore, the paper cherry picks two groups of anomalies from HMDB51 to show the importance of anomalies across the temporal axis. \n\nThe paper reports interesting results not only on the above two subsets of HMDB51, but also on few benchmark anomaly detection datasets. The results on the benchmark datasets are competitive compared to the chosen state-of-the-art methods, but outperforming them on the two subsets of HMDB51. \n\n[1] Tal Reiss and Yedid Hoshen. Attribute-based Representations for Accurate and Interpretable\nVideo Anomaly Detection, December 2022. URL http://arxiv.org/abs/2212.00789.\narXiv:2212.00789 [cs].", "review_text": "This paper proposes a multi-frame-based video anomaly detection method, that builds on top of [1]. In [1] mostly frame-level attributes are included, while the proposed method extends the method of [1] by including multi-frames encoding features extracted across 16 frames. \n\nFurthermore, the paper cherry picks two groups of anomalies from HMDB51 to show the importance of anomalies across the temporal axis. \n\nThe paper reports interesting results not only on the above two subsets of HMDB51, but also on few benchmark anomaly detection datasets. The results on the benchmark datasets are competitive compared to the chosen state-of-the-art methods, but outperforming them on the two subsets of HMDB51. \n\n[1] Tal Reiss and Yedid Hoshen. Attribute-based Representations for Accurate and Interpretable\nVideo Anomaly Detection, December 2022. URL http://arxiv.org/abs/2212.00789.\narXiv:2212.00789 [cs].", "strengths": "-Good overview of existing methods and datasets used in anomaly detection. \n-\"Introducing\" new videos to video anomaly detection for benchmarking.\n-Comprehensive and competitive results on the public benchmarks and significantly higher results compared to state-of-the-art on the two subsets of HMDB51.\n-Proper ablation study. which also shows the effect of video encoding features in Table 4.", "weaknesses": "Despite the interesting results, the paper's method sounds like a simple extension of [1] by introducing temporal features to [1].\nThough the paper has cherry picked videos from HMDB51 and suggests using them for anomaly detection, they need claim this as their data (Table 1), which is not correct.\nThe abolition study shows that video encoder features alone are producing almost similar results with the entire set of features on the subsets of HMDB51, so what is the point in inclusion of other features? \nWhat about including/cherry picking some other videos from HMDB51 that are normal, but are similar to the abnormal videos already included in the two subsets?", "questions": "Please see the previous section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a multi-frame-based video anomaly detection method, that builds on top of [1]. In [1] mostly frame-level attributes are included, while the proposed method extends the method of [1] by including multi-frames encoding features extracted across 16 frames. \n\nFurthermore, the paper cherry picks two groups of anomalies from HMDB51 to show the importance of anomalies across the temporal axis. \n\nThe paper reports interesting results not only on the above two subsets of HMDB51, but also on few benchmark anomaly detection datasets. The results on the benchmark datasets are competitive compared to the chosen state-of-the-art methods, but outperforming them on the two subsets of HMDB51. \n\n[1] Tal Reiss and Yedid Hoshen. Attribute-based Representations for Accurate and Interpretable\nVideo Anomaly Detection, December 2022. URL http://arxiv.org/abs/2212.00789.\narXiv:2212.00789 [cs].", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "-Good overview of existing methods and datasets used in anomaly detection. \n-\"Introducing\" new videos to video anomaly detection for benchmarking.\n-Comprehensive and competitive results on the public benchmarks and significantly higher results compared to state-of-the-art on the two subsets of HMDB51.\n-Proper ablation study. which also shows the effect of video encoding features in Table 4.", "weaknesses": "Despite the interesting results, the paper's method sounds like a simple extension of [1] by introducing temporal features to [1].\nThough the paper has cherry picked videos from HMDB51 and suggests using them for anomaly detection, they need claim this as their data (Table 1), which is not correct.\nThe abolition study shows that video encoder features alone are producing almost similar results with the entire set of features on the subsets of HMDB51, so what is the point in inclusion of other features? \nWhat about including/cherry picking some other videos from HMDB51 that are normal, but are similar to the abnormal videos already included in the two subsets?", "questions": "Please see the previous section", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698668666138}, {"id": "a24ZTDZhnA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2475/Reviewer_fVMW"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors study anomaly detection in video, presenting two datasets with video-level annotations and an adapted version of AI-VAD, called MFAD, which performs well on the proposed datasets. The presented method is also evaluated on three existing datasets, being compared with AI-VAD and other methods from literature.", "review_text": "The authors study anomaly detection in video, presenting two datasets with video-level annotations and an adapted version of AI-VAD, called MFAD, which performs well on the proposed datasets. The presented method is also evaluated on three existing datasets, being compared with AI-VAD and other methods from literature.", "strengths": "- Anomaly detection is an interesting and timely topic.\n- The paper is well written and easy to follow.", "weaknesses": "- The proposed method is incremental w.r.t. AI-VAD.\n- UCF-Crime and XD-Violence datasets are not included in the comparison provided in Table 1.\n- The proposed datasets are rearranged subsets of HMDB51. There was no manual annotation involved, or at least, the authors did not mentione anything about it. Therefore, it is hard to consider the proposed datasets as entirely new. Just as the method, this contribution is incremental.\n- The comparison in Table 1 does not reflect the difficulty / diversity advantages suggested by the authors. I do not see the benefits of the proposed datasets w.r.t. recent benchmarks such as UCF-Crime, XD-Violence or UBnormal.\n- The proposed datasets contain video-level annotations (the videos are labeled either as normal or abnormal), while other benchmarks contain frame or pixel annotations. I believe this type of annotation does not reflect a realistic scenario.\n- There is no time evaluation reported for the presented method. Anomaly detection methods are expected to run in real-time, but it is not clear if MFAD can do this. To me, the method is a bit heavy.\n- There are some some language corrections to be made, e.g.:\n  - \"It’s crucial to\" => \"It is crucial to\" (language abbrevations should be avoided in formal language).", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors study anomaly detection in video, presenting two datasets with video-level annotations and an adapted version of AI-VAD, called MFAD, which performs well on the proposed datasets. The presented method is also evaluated on three existing datasets, being compared with AI-VAD and other methods from literature.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "- Anomaly detection is an interesting and timely topic.\n- The paper is well written and easy to follow.", "weaknesses": "- The proposed method is incremental w.r.t. AI-VAD.\n- UCF-Crime and XD-Violence datasets are not included in the comparison provided in Table 1.\n- The proposed datasets are rearranged subsets of HMDB51. There was no manual annotation involved, or at least, the authors did not mentione anything about it. Therefore, it is hard to consider the proposed datasets as entirely new. Just as the method, this contribution is incremental.\n- The comparison in Table 1 does not reflect the difficulty / diversity advantages suggested by the authors. I do not see the benefits of the proposed datasets w.r.t. recent benchmarks such as UCF-Crime, XD-Violence or UBnormal.\n- The proposed datasets contain video-level annotations (the videos are labeled either as normal or abnormal), while other benchmarks contain frame or pixel annotations. I believe this type of annotation does not reflect a realistic scenario.\n- There is no time evaluation reported for the presented method. Anomaly detection methods are expected to run in real-time, but it is not clear if MFAD can do this. To me, the method is a bit heavy.\n- There are some some language corrections to be made, e.g.:\n  - \"It’s crucial to\" => \"It is crucial to\" (language abbrevations should be avoided in formal language).", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1697359138993}], "openreview_url": "https://openreview.net/forum?id=WnEnU2K3Rb", "arxiv_id": "2310.01904", "paper_pdf": "papers/WnEnU2K3Rb.pdf", "paper_pdf_sha256": "c8b07d25c6bfb8fb55a14e652d41c699034747cb903ff672c50f6ae744227b01", "paper_pdf_bytes": 10393905, "paper_pdf_source": "openreview", "code_url": "https://github.com/yoavarad/MFAD", "code_repository": "yoavarad/MFAD", "code_commit": "a2a3d3fbb3c609081af537af34ac4288c4f39954", "code_archive": "repos/WnEnU2K3Rb.zip", "code_archive_sha256": "4d597a422be7606cf94ac232bf5e9732cc7895bb36e93323e6602ca1a25c4442", "code_archive_bytes": 17922, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 40, "github_languages": {"Python": 59588, "Shell": 1282}, "github_archived": false, "github_pushed_at": "2023-09-28T10:57:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beyond-the-benchmark-detecting-diverse"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "j3AyKG-H3uM", "year": 2023, "status": "rejected", "title": "Improved Group Robustness via Classifier Retraining on Independent Splits", "authors": ["Thien Hang Nguyen", "Hongyang Ryan Zhang", "Huy Nguyen"], "authorids": ["~Thien_Hang_Nguyen1", "~Hongyang_Ryan_Zhang1", "~Huy_Nguyen1"], "authors_source": "OpenReview API", "abstract": "Deep neural networks learned by minimizing the average risk can achieve strong average performance, but their performance for a subgroup may degrade, if the subgroup is underrepresented in the overall data population. Group distributionally robust optimization (Sagawa et al., 2020a, GDRO) is a standard baseline for learning models with strong worst-group performance. However, GDRO requires group labels for every example during training and can be prone to overfitting, often requiring careful model capacity control via regularization or early stopping. When only a limited amount of group labels is available, Just Train Twice (Liu et al., 2021, JTT) is a popular approach which infers a pseudo-group-label for every unlabeled example. The process of inferring pseudo labels can be highly sensitive during model selection. To alleviate overfitting for GDRO and the pseudo labeling process for JTT, we propose a new method via classifier retraining on independent splits (of the training data). We find that using a novel sample splitting procedure achieves robust worst-group performance in the fine-tuning step. When evaluated on benchmark image and text classification tasks, our approach consistently reduces the requirement of group labels and hyperparameter search during training. Experimental results confirm that our approach performs favorably compared with existing methods (including GDRO and JTT) when either group labels are available during training or are only available during validation.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "mdmvCHAVoX", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4736/Reviewer_X9ti"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a simple method called CROIS for learning a model that can be applied in group-shift settings, i.e. cases where the performance of the group that is underrepresented in the training set is low when there is a distribution shift.\nCROIS is motivated by the idea that although ERM-trained DNNs often take advantage of spurious features, previous works show its capability to produce good features.\nCROIS consists of 2 steps method 1) train an ERM model on data without the group label information and 2) correct the classifier layer by using the features extracted from stage 1) to train a GDRO model on the data with group labels. Although in practice in the experiments the hyperparameters are fixed, for both steps the model selection can be done on the validation set, the first one optimising the aggregate accuracy and the second one the worst-group accuracy.\n", "review_text": "The idea is interesting and promising, but in the current state, I am not very convinced the work is ready to be published. The authors should revise their experiments section to address my doubts. \n\n", "strengths": "**Strength**: \n\nThe paper is well written, with simple and relevant notation. CROIS requires fewer labelled data than the GDRO baseline considered, and it does not rely on pseudo-labelling as JTT. The authors perform a good ablation study on different components of their proposed method (e.g. sampling with or without replacement, i.e. using independent splits or not, first and second stage different training strategies, robustness towards the hyperparameters tuning).\n\n**Weaknesses**: \n\nFully training group labels baselines could include more recent methods such as SGDRO (non-flat version of GDRO proposed in Goel et al.). \nThere is a misleading sentence on page 3 when describing the Waterbirds dataset: “The bird images are then modified with either a water or land background.” There is no “modification” involved, in the dataset the birds are already placed either on water or land background (this is a minor note, but should be revisited).\nThe main weaknesses are all in the “Experiments” section. 1) Although the hyperparameters are fixed, some ablation is still required (for example the regularisation in the second stage for the CivilComments dataset). 2) In Tables 1 and 2 it is not clear what the results are, is it correct to assume they are the mean and std. performances over three different initialization seeds? 3) The results in Tables 1 and 2 are not correct, CROIS is using the validation set for training in the second stage, not only for model selection. For the datasets considered, the validation set has the same distribution as the test set, it does not seem correct to talk about “group shift” anymore. 4) Table 2 should also include the average test accuracy to show the trade-off between average and worst-case performances when varying the size of the validation set. Also, why is the std. not reported here? 5) Table 3 shows promising results, but the best fraction “p” is a hyperparameter that should be tuned using the worst-group performance on the validation set, like the regularization term for GDRO.  Its behaviour is not linear (e.g. the more the better) nor consistent across datasets. For example, in both text datasets, only one fraction produces better results than the GDRO baseline. Here, I disagree with the statement “In practice, p is not a parameter to choose (there’s no reason to throw away group labels) but rather is limited by the resources available to obtain group labels”: using p=0.5 is often worse than using p=0.3. 6) To obtain robust results, it would have been better to evaluate the methods across different splits of train-val-test, not simply different initialisation seeds. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a simple method called CROIS for learning a model that can be applied in group-shift settings, i.e. cases where the performance of the group that is underrepresented in the training set is low when there is a distribution shift.\nCROIS is motivated by the idea that although ERM-trained DNNs often take advantage of spurious features, previous works show its capability to produce good features.\nCROIS consists of 2 steps method 1) train an ERM model on data without the group label information and 2) correct the classifier layer by using the features extracted from stage 1) to train a GDRO model on the data with group labels. Although in practice in the experiments the hyperparameters are fixed, for both steps the model selection can be done on the validation set, the first one optimising the aggregate accuracy and the second one the worst-group accuracy.\n", "strength_and_weaknesses": "**Strength**: \n\nThe paper is well written, with simple and relevant notation. CROIS requires fewer labelled data than the GDRO baseline considered, and it does not rely on pseudo-labelling as JTT. The authors perform a good ablation study on different components of their proposed method (e.g. sampling with or without replacement, i.e. using independent splits or not, first and second stage different training strategies, robustness towards the hyperparameters tuning).\n\n**Weaknesses**: \n\nFully training group labels baselines could include more recent methods such as SGDRO (non-flat version of GDRO proposed in Goel et al.). \nThere is a misleading sentence on page 3 when describing the Waterbirds dataset: “The bird images are then modified with either a water or land background.” There is no “modification” involved, in the dataset the birds are already placed either on water or land background (this is a minor note, but should be revisited).\nThe main weaknesses are all in the “Experiments” section. 1) Although the hyperparameters are fixed, some ablation is still required (for example the regularisation in the second stage for the CivilComments dataset). 2) In Tables 1 and 2 it is not clear what the results are, is it correct to assume they are the mean and std. performances over three different initialization seeds? 3) The results in Tables 1 and 2 are not correct, CROIS is using the validation set for training in the second stage, not only for model selection. For the datasets considered, the validation set has the same distribution as the test set, it does not seem correct to talk about “group shift” anymore. 4) Table 2 should also include the average test accuracy to show the trade-off between average and worst-case performances when varying the size of the validation set. Also, why is the std. not reported here? 5) Table 3 shows promising results, but the best fraction “p” is a hyperparameter that should be tuned using the worst-group performance on the validation set, like the regularization term for GDRO.  Its behaviour is not linear (e.g. the more the better) nor consistent across datasets. For example, in both text datasets, only one fraction produces better results than the GDRO baseline. Here, I disagree with the statement “In practice, p is not a parameter to choose (there’s no reason to throw away group labels) but rather is limited by the resources available to obtain group labels”: using p=0.5 is often worse than using p=0.3. 6) To obtain robust results, it would have been better to evaluate the methods across different splits of train-val-test, not simply different initialisation seeds. \n", "clarity,_quality,_novelty_and_reproducibility": "**Clarity**: The paper is well written, but the “Experiments” section is not always very clear (e.g. the reported results are the average or the best values? What are the authors varying to create 3 runs? Why is the std. not included in all the Tables?)\n\n**Quality**: As highlighted in the “weaknesses” the quality of the experiments can be improved.\n\n**Novelty**: Although the single components exist in prior works, the idea of combining the two stages is somewhat new.\n\n**Reproducibility**: The source code is provided and the hyperparameters are detailed in the Appendix.\n", "summary_of_the_review": "The idea is interesting and promising, but in the current state, I am not very convinced the work is ready to be published. The authors should revise their experiments section to address my doubts. \n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667321922881}, {"id": "0C06gc7jCBG", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4736/Reviewer_2dWG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed a novel approach to improve worst-group performance, CROIS, that works well with reduced/limited group annotations on training/val set. The idea is simple yet effective: obtain good feature extractors on group-unlabeled data with ERM loss, and only retrain the last layer on group-labeled data with GDRO loss. Empirical experiments and abundant ablations are performed to support the method.", "review_text": "Though the empirical results do not outperform SOTA, the paper still presents a novel and interesting idea in worst-group improvement that is quite different from usual approaches in this area. A key question regarding the group-labeled size for classifier retraining can be better explained for other researchers to build upon the work and/or practitioners to apply the proposed method.", "strengths": "Strengths:\n\nThe proposed method is a simple, novel and interesting idea that is orthogonal to the usual pseudo-group-labeling approach taken by other existing methods like JTT.\n\nWeaknesses/comments:\n\n1. The relationship of the CROIS performance to the data size for retraining (group-labeled fraction in training set for CR - Table 3, or reduced val size for CR - Table 2) seems quite inconclusive, as also pointed out by the authors in Sec 4.2 (Page 8). In particular when one cross compares Table 2 and 3, I assume for instance 30% of training size (Table 3) is comparable to 100% of val size (Table 2) for CelebA, which happens to be the sweet CR size for this dataset? IMHO p is the (only) key parameter in Algorithm 1, and it would be good to clarify in the paper, for instance, what's the rule of thumb of choosing p? Would it make sense to consider the CR data size (or p) as a tunable hparam? If group-unlabeled size >> group-labeled size, then should one use 50/50 group-labeled in D_L/D_val or tuning the ratio is still needed?\n2. Empirical results are somewhat strong, on par with (but not significantly outperforming) SOTA (e.g. Table 1). Many other baselines (Sec 1.1) should be added to the results (e.g. Table 1) for comparison. Besides, it would be good to include error bars in Table 2 to show that error bar should become much larger when val size becomes smaller despite the \"relatively small\" drop in mean performance.\n3. In Sec 3, the authors claim that independent splits are done to avoid the feature extractors memorizing spurious correlations with class labels. But many modern DL models rely on unsupervised/contrastive/label-free training of feature extractors. For instance, one can continue the pretraining of BERT on the in-domain data on the two language tasks (https://arxiv.org/abs/2004.10964). Perhaps then independent splits still won't be required and more group labels can be used for classifier retraining (in case of limited group annotations)?\n4. The abstract/intro clearly separates the two use cases: a) full training and b) val group annotation. Algorithm 1 (Sec 3) does not make such distinction where D_L is listed as an input. It only becomes clear how Alg 1 is applied to case b) in Sec 4.1 where val set is actually split 50/50 into D_L and D_val. For consistency, could consider clarifying this earlier in the paper to avoid confusion.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors proposed a novel approach to improve worst-group performance, CROIS, that works well with reduced/limited group annotations on training/val set. The idea is simple yet effective: obtain good feature extractors on group-unlabeled data with ERM loss, and only retrain the last layer on group-labeled data with GDRO loss. Empirical experiments and abundant ablations are performed to support the method.", "strength_and_weaknesses": "Strengths:\n\nThe proposed method is a simple, novel and interesting idea that is orthogonal to the usual pseudo-group-labeling approach taken by other existing methods like JTT.\n\nWeaknesses/comments:\n\n1. The relationship of the CROIS performance to the data size for retraining (group-labeled fraction in training set for CR - Table 3, or reduced val size for CR - Table 2) seems quite inconclusive, as also pointed out by the authors in Sec 4.2 (Page 8). In particular when one cross compares Table 2 and 3, I assume for instance 30% of training size (Table 3) is comparable to 100% of val size (Table 2) for CelebA, which happens to be the sweet CR size for this dataset? IMHO p is the (only) key parameter in Algorithm 1, and it would be good to clarify in the paper, for instance, what's the rule of thumb of choosing p? Would it make sense to consider the CR data size (or p) as a tunable hparam? If group-unlabeled size >> group-labeled size, then should one use 50/50 group-labeled in D_L/D_val or tuning the ratio is still needed?\n2. Empirical results are somewhat strong, on par with (but not significantly outperforming) SOTA (e.g. Table 1). Many other baselines (Sec 1.1) should be added to the results (e.g. Table 1) for comparison. Besides, it would be good to include error bars in Table 2 to show that error bar should become much larger when val size becomes smaller despite the \"relatively small\" drop in mean performance.\n3. In Sec 3, the authors claim that independent splits are done to avoid the feature extractors memorizing spurious correlations with class labels. But many modern DL models rely on unsupervised/contrastive/label-free training of feature extractors. For instance, one can continue the pretraining of BERT on the in-domain data on the two language tasks (https://arxiv.org/abs/2004.10964). Perhaps then independent splits still won't be required and more group labels can be used for classifier retraining (in case of limited group annotations)?\n4. The abstract/intro clearly separates the two use cases: a) full training and b) val group annotation. Algorithm 1 (Sec 3) does not make such distinction where D_L is listed as an input. It only becomes clear how Alg 1 is applied to case b) in Sec 4.1 where val set is actually split 50/50 into D_L and D_val. For consistency, could consider clarifying this earlier in the paper to avoid confusion.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well presented (except for the choice of a key hparam of CR data size, see weakness #2 above). The proposed method is a simple, novel and interesting idea that is orthogonal to the usual pseudo-group-labeling approach taken by other existing methods like JTT. I think most readers will find it interesting. The pseudo code is clear and simple enough for interested readers to reproduce the work.", "summary_of_the_review": "Though the empirical results do not outperform SOTA, the paper still presents a novel and interesting idea in worst-group improvement that is quite different from usual approaches in this area. A key question regarding the group-labeled size for classifier retraining can be better explained for other researchers to build upon the work and/or practitioners to apply the proposed method.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666745731872}, {"id": "0J2qcFNMVnW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4736/Reviewer_ysYp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a very simple method termed classifier retraining on independent splits to improve the subpopulation shift robustness of the model. The main algorithm of the paper is clearly shown in Algorithm 1 of the paper.", "review_text": "This paper proposes a very simple and effective method. But some experiments are unfair.", "strengths": "Advantages.\n\nThis paper is well written and easy to follow. Although the method of the paper is too straightforward, the author explains its principle very clearly.\n\nWeaknesses.\n\nHowever, I still have the following concerns: \n1. The comparison experiments in Table 1 is unfair. Since the proposed method uses the data from the validation set for training.\n2. The paper does not conduct discussion and compare with state-of-the-art methods, e.g., LISA[1], UMIX[2], CnC[3], CGD[4].\n\n[1] Improving out-of-distribution robustness via selective augmentation.\n\n[2] UMIX: Improving Importance Weighting for Subpopulation Shift via Uncertainty-Aware Mixup\n\n[3] Correct-n-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations\n\n[4] Improving out-of-distribution robustness via selective augmentation.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a very simple method termed classifier retraining on independent splits to improve the subpopulation shift robustness of the model. The main algorithm of the paper is clearly shown in Algorithm 1 of the paper.", "strength_and_weaknesses": "Advantages.\n\nThis paper is well written and easy to follow. Although the method of the paper is too straightforward, the author explains its principle very clearly.\n\nWeaknesses.\n\nHowever, I still have the following concerns: \n1. The comparison experiments in Table 1 is unfair. Since the proposed method uses the data from the validation set for training.\n2. The paper does not conduct discussion and compare with state-of-the-art methods, e.g., LISA[1], UMIX[2], CnC[3], CGD[4].\n\n[1] Improving out-of-distribution robustness via selective augmentation.\n\n[2] UMIX: Improving Importance Weighting for Subpopulation Shift via Uncertainty-Aware Mixup\n\n[3] Correct-n-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations\n\n[4] Improving out-of-distribution robustness via selective augmentation.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well written and easy to follow.", "summary_of_the_review": "This paper proposes a very simple and effective method. But some experiments are unfair.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666531609328}], "openreview_url": "https://openreview.net/forum?id=j3AyKG-H3uM", "arxiv_id": "2204.09583", "paper_pdf": "papers/j3AyKG-H3uM.pdf", "paper_pdf_sha256": "281a45e1840587f8b41835a912b54684bbf7aa2f6d525608e7ea1117f2dd7e64", "paper_pdf_bytes": 1218406, "paper_pdf_source": "openreview", "code_url": "https://github.com/timmytonga/crois", "code_repository": "timmytonga/crois", "code_commit": "33930be328adeb87f38792630b924e077bcc9d3a", "code_archive": "repos/j3AyKG-H3uM.zip", "code_archive_sha256": "1a49355239eec7d4bb7f77d8cfcd12f90bbf037d660a090230ecec3e7d701050", "code_archive_bytes": 57308, "code_file_count": 26, "code_extensions": {".py": 26}, "github_disk_usage_kb": 49, "github_languages": {"Python": 179333}, "github_archived": false, "github_pushed_at": "2023-07-27T15:26:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improved-worst-group-robustness-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "46lmrnVBHBL", "year": 2022, "status": "rejected", "title": "Explanatory Learning: Beyond Empiricism in Neural Networks", "authors": ["Antonio Norelli", "Giorgio Mariani", "Luca Moschella", "Andrea Santilli", "Giambattista Parascandolo", "Simone Melzi", "Emanuele Rodolà"], "authorids": ["~Antonio_Norelli2", "~Giorgio_Mariani1", "~Luca_Moschella1", "~Andrea_Santilli1", "~Giambattista_Parascandolo1", "~Simone_Melzi2", "~Emanuele_Rodolà1"], "authors_source": "OpenReview API", "abstract": "We introduce Explanatory Learning (EL), an explanation-driven machine learning framework to use existing knowledge buried in symbolic sequences expressed in an unknown language. In EL, the burden of interpreting explanations is not left to humans or human-coded compilers, as done in Program Synthesis. Rather, EL calls for a learned interpreter, built upon existing explanations paired with observations of several phenomena. This interpreter can then be used to make predictions on novel phenomena, and even find an explanation for them. We formulate the EL problem as a simple binary classification task, so that common end-to-end approaches aligned with the dominant empiricist view of machine learning could, in principle, solve it. To these models, we oppose Critical Rationalist Networks (CRNs), which instead embrace a rationalist view on the acquisition of knowledge. CRNs express several desired properties by construction, they are truly explainable, can adjust their processing at test-time for harder inferences, and can offer strong confidence guarantees on their predictions.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JuWRoCKfuG2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2684/Reviewer_ZgwB"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors introduce a novel learning framework that revolves around learning a map between concepts and their description, where the latter are expressed in a language unknown to the learner.  Two learning problems are defined: (i) learning a new concept from a description of that concepts as well as examples and descriptions of other concepts, (ii) learn a new concept from examples of that concept as well as examples and descriptions of other concepts.  The underlying assumption is that the map (interpreter) from sentences to sets of examples is shared.  The authors propose an environment to evaluate these tasks and a neural architecture to tackle them.  ", "review_text": "- The paper is a pleasure to read.\n\n\n- The choice of illustrative examples is excellent.\n\n\n- The contribution is conceptual in nature, which is good.  Alas, the motivation for pursuing this research direction is unclear.  Specifically:\n\n1) It is not clear why the machine should estimate *both* the link between sentences and data and between sentences and sentences *jointly*.\n\n2) It is not clear why the lanugage would be unknown and unstructured.\n\n3) It is not clear what (new?) real-world applications this setup is meant to capture.\n\n(Please note that this setup does not solve the \"semantic gap\" problem, because the semantics of the sentences are implicitly supplied by whoever designs the data set.)\n\nI would advise the authors to ground their motivation in concrete conceptual problems and/or applications.\n\nGiven that the learning framework is *the* key contribution of the paper (the value of the two other major contributions depends on whether the learning setting makes sense), the fact that motivation is lacking/unclear is a serious concern.\n\nThis makes it hard to evaluate the significance of the paper.\n\n\n- The terminology used in the paper feels somewhat misleading/inappropriate.\n\nAs far as I can see, \"phenomena\" are simply *concepts* and the \"explanations\" are intensional *descriptions* thereof, except in a language unknown to the machine (but known to the annotator).  I do not understand what is gained by using this terminology.  Indeed, I find the latter confusing and I expect other readers to be confused by it too.\n\nIn particular, I do not understand the link between \"description\" and \"explanation\", especially considering the causal connotations of the latter term (which are becoming more and more clear as work on causality is being merged into AI and ML.)\n\n\n- The work seems to rely on a rather strict assumption.  In particular, the fact that D_0 should be discriminative for P_0 in L is (as far as I can see) unlikely to hold in practice -- especially considering that D_0 is supposed to contain a \"small set of observations\" -- and it simplifies the learning problem considerably.  In what applications is it reasonable for this asumption to hold?  Is it necessary?  What happens if it doesn't hold in practice?  How costly is it to acquire a D_0 that explicitly satisfies this assumption?  Given that such a D_0 would not be IID, how would this impact statistical learning of P_0?  These issues are touched upon in the conclusions, but they deserves an actual discussion.\n\n\n- The proposed architecture is reasonable but surprisingly involved and the details are hidden in the appendix.  It would be more straightforward to explain in detail the various pieces that make up the architecture (how many, what they take as input and what they spit out) from the get-go, rather than relying on Figure 3, which lacks mathematical precision.\n\n\n- Inferring whether an instance x satisfies a concept P is surprisingly involved as it requires to generate a (presumably large) number of candidate descriptions for P and then counting, for each description, whether x satisfies it.  Presumably this scales poorly with language complexity.  Is this efficient at all?\n\n\n- The experiments are limited to an interesting but entirely synthetic (actually, quite toy) setting.  For instance, as far as I can tell no sub-symbolic inputs are present.  Experiments with real data would have been useful to evaluate the efficacy of the proposed pipeline.  CRNs are compared only against two simpler baselines.  I realize that the focus of the paper is in its conceputal contribution, but a more varied selection of experiments would have been welcome.\n\n\nMinor issues\n------------\n\n- Wouldn't it be more natural to define the interpreter as a map from descriptions to classifiers (indicator functions)?\n\n- It would be good to disambiguate the term \"explanatory learning\" from previous uses of the same term, see \"Explanatory interactive machine learning\" AIES 2019.\n\n- The \"communication problem\" shares some aspects with multitask learning, especially so if the language is compositional and the tasks form a hierarchically (or can be related logically to each other, e.g., P0 is the negation of P1 etc.).  It also shares aspects with few-shot learning.  It may be worth highlighting the connection.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors introduce a novel learning framework that revolves around learning a map between concepts and their description, where the latter are expressed in a language unknown to the learner.  Two learning problems are defined: (i) learning a new concept from a description of that concepts as well as examples and descriptions of other concepts, (ii) learn a new concept from examples of that concept as well as examples and descriptions of other concepts.  The underlying assumption is that the map (interpreter) from sentences to sets of examples is shared.  The authors propose an environment to evaluate these tasks and a neural architecture to tackle them.  ", "main_review": "- The paper is a pleasure to read.\n\n\n- The choice of illustrative examples is excellent.\n\n\n- The contribution is conceptual in nature, which is good.  Alas, the motivation for pursuing this research direction is unclear.  Specifically:\n\n1) It is not clear why the machine should estimate *both* the link between sentences and data and between sentences and sentences *jointly*.\n\n2) It is not clear why the lanugage would be unknown and unstructured.\n\n3) It is not clear what (new?) real-world applications this setup is meant to capture.\n\n(Please note that this setup does not solve the \"semantic gap\" problem, because the semantics of the sentences are implicitly supplied by whoever designs the data set.)\n\nI would advise the authors to ground their motivation in concrete conceptual problems and/or applications.\n\nGiven that the learning framework is *the* key contribution of the paper (the value of the two other major contributions depends on whether the learning setting makes sense), the fact that motivation is lacking/unclear is a serious concern.\n\nThis makes it hard to evaluate the significance of the paper.\n\n\n- The terminology used in the paper feels somewhat misleading/inappropriate.\n\nAs far as I can see, \"phenomena\" are simply *concepts* and the \"explanations\" are intensional *descriptions* thereof, except in a language unknown to the machine (but known to the annotator).  I do not understand what is gained by using this terminology.  Indeed, I find the latter confusing and I expect other readers to be confused by it too.\n\nIn particular, I do not understand the link between \"description\" and \"explanation\", especially considering the causal connotations of the latter term (which are becoming more and more clear as work on causality is being merged into AI and ML.)\n\n\n- The work seems to rely on a rather strict assumption.  In particular, the fact that D_0 should be discriminative for P_0 in L is (as far as I can see) unlikely to hold in practice -- especially considering that D_0 is supposed to contain a \"small set of observations\" -- and it simplifies the learning problem considerably.  In what applications is it reasonable for this asumption to hold?  Is it necessary?  What happens if it doesn't hold in practice?  How costly is it to acquire a D_0 that explicitly satisfies this assumption?  Given that such a D_0 would not be IID, how would this impact statistical learning of P_0?  These issues are touched upon in the conclusions, but they deserves an actual discussion.\n\n\n- The proposed architecture is reasonable but surprisingly involved and the details are hidden in the appendix.  It would be more straightforward to explain in detail the various pieces that make up the architecture (how many, what they take as input and what they spit out) from the get-go, rather than relying on Figure 3, which lacks mathematical precision.\n\n\n- Inferring whether an instance x satisfies a concept P is surprisingly involved as it requires to generate a (presumably large) number of candidate descriptions for P and then counting, for each description, whether x satisfies it.  Presumably this scales poorly with language complexity.  Is this efficient at all?\n\n\n- The experiments are limited to an interesting but entirely synthetic (actually, quite toy) setting.  For instance, as far as I can tell no sub-symbolic inputs are present.  Experiments with real data would have been useful to evaluate the efficacy of the proposed pipeline.  CRNs are compared only against two simpler baselines.  I realize that the focus of the paper is in its conceputal contribution, but a more varied selection of experiments would have been welcome.\n\n\nMinor issues\n------------\n\n- Wouldn't it be more natural to define the interpreter as a map from descriptions to classifiers (indicator functions)?\n\n- It would be good to disambiguate the term \"explanatory learning\" from previous uses of the same term, see \"Explanatory interactive machine learning\" AIES 2019.\n\n- The \"communication problem\" shares some aspects with multitask learning, especially so if the language is compositional and the tasks form a hierarchically (or can be related logically to each other, e.g., P0 is the negation of P1 etc.).  It also shares aspects with few-shot learning.  It may be worth highlighting the connection.", "summary_of_the_review": "Potentially great paper with unclear motivation/applications", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636117490194}, {"id": "C-6LEcN_gZ0", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2684/Reviewer_H4ra"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes approaching learning a language as a learning problem that is grammar, alphabet, etc. agnostic. Such formulation has resulted in the so called Explanatory Learning (EL) framework that is paired up with an environment to test it. The starting point for learning a language is to extract an interpreter based on a given set of observation and their explanation and then use the interpreter to determine whether an observation belongs to the language, in a binary classification setting. ", "review_text": "Questions: \n- In Figure 1, why the first left sequence is correct according to the rule? \n- When tagging unseen structures based on the secret rule discovered by the user, how many unseen structures are required to completely rule out other rules that are marginally different from the secret one. Is this number always 1176 in the current setting? \n- Since some figures don’t have caption, it’s hard to refer to them… In the figure on the lest side of “Metrics.” How the predicted vector is derived based on various rules? \n- As the authors point out the goal here is analogous to that of IPL. Do I understand this correctly that the main advantage of EL to IPL is skipping the translation of the data to logic? If that’s the case, one can learn the translation itself and eliminate the difficulty. Can you elaborate on the advantage of EL over IPL?\n- Can’t T-Acc be changes to something auc based instead that takes care of the permissively problem? \n\n\nOther remarks: \n- This is a language specific problem, where order of the sequence does not necessarily matters. The current title does not  reflect this. Perhaps something like Beyond empiricism in neural language learning or something alike would be a better suit. \n- Some approaches to explanatory AI that aim at generating explanations as well as predictions such as “TED: Teaching AI to Explain its Decisions” are missed here and should be commended on. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes approaching learning a language as a learning problem that is grammar, alphabet, etc. agnostic. Such formulation has resulted in the so called Explanatory Learning (EL) framework that is paired up with an environment to test it. The starting point for learning a language is to extract an interpreter based on a given set of observation and their explanation and then use the interpreter to determine whether an observation belongs to the language, in a binary classification setting. ", "main_review": "Questions: \n- In Figure 1, why the first left sequence is correct according to the rule? \n- When tagging unseen structures based on the secret rule discovered by the user, how many unseen structures are required to completely rule out other rules that are marginally different from the secret one. Is this number always 1176 in the current setting? \n- Since some figures don’t have caption, it’s hard to refer to them… In the figure on the lest side of “Metrics.” How the predicted vector is derived based on various rules? \n- As the authors point out the goal here is analogous to that of IPL. Do I understand this correctly that the main advantage of EL to IPL is skipping the translation of the data to logic? If that’s the case, one can learn the translation itself and eliminate the difficulty. Can you elaborate on the advantage of EL over IPL?\n- Can’t T-Acc be changes to something auc based instead that takes care of the permissively problem? \n\n\nOther remarks: \n- This is a language specific problem, where order of the sequence does not necessarily matters. The current title does not  reflect this. Perhaps something like Beyond empiricism in neural language learning or something alike would be a better suit. \n- Some approaches to explanatory AI that aim at generating explanations as well as predictions such as “TED: Teaching AI to Explain its Decisions” are missed here and should be commended on. ", "summary_of_the_review": "- Overall, I quite like the idea of the paper due to the generality of the approach, namely learning an interpreter from observation. \n\n- “proving” the discovered rule is indeed the secret one by examining it in practice (i.e., being used to tag unseen structure) as opposed to revealing the rule directly in text is also a very neat idea.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635863134659}, {"id": "tqDYuVHUrFS", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2684/Reviewer_VRSC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new framework for studying explanation driven machine learning problems called Explanatory Learning. The goal is to learn an interpreter model from explanations paired with observations for a particular phenomenon. The explanations might be in an unknown language, but the explanations paired with observations can be used to learn a good interpreter. Once learnt, the interpreter should be able to follow new explanation for an unseen phenomenon. They refer to this problem as the communication problem. They also define the scientist problem where explanations for the unseen phenomenon is not available. The paper also introduces an Odeen dataset to facilitate experiments with Explanatory Learning. The authors propose a neural network architecture as a solution to the scientist problem. It has two components: a conjecture generator which generates a set of candidate English explanations, and a second interpreter model. Experimental results show better generalization compared to end to end neural systems.  ", "review_text": "Strengths: 1. The proposed framework forces us to think about explanations and generalization as first class citizens. \n\nWeaknesses: \n1. The idea sounds very familiar to explanation based learning. It will be good to contrast the two.\n2. My main criticism is that I am not sure we need an entirely new framework for targeting explanation based solutions. Many recent work in NLP try to frame problems in English to get cross task generalization. I would prefer this paper to be positioned as an improvement to such existing approaches, as this is tackling a harder problem class. \n3. The experimental results are all on the new game like dataset. It will be good to see performance of proposed methods on real datasets, or already existing synthetic datasets. The baselines are also weak, so its not clear if the proposed techniques will perform better than other cross task generalization methods.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new framework for studying explanation driven machine learning problems called Explanatory Learning. The goal is to learn an interpreter model from explanations paired with observations for a particular phenomenon. The explanations might be in an unknown language, but the explanations paired with observations can be used to learn a good interpreter. Once learnt, the interpreter should be able to follow new explanation for an unseen phenomenon. They refer to this problem as the communication problem. They also define the scientist problem where explanations for the unseen phenomenon is not available. The paper also introduces an Odeen dataset to facilitate experiments with Explanatory Learning. The authors propose a neural network architecture as a solution to the scientist problem. It has two components: a conjecture generator which generates a set of candidate English explanations, and a second interpreter model. Experimental results show better generalization compared to end to end neural systems.  ", "main_review": "Strengths: 1. The proposed framework forces us to think about explanations and generalization as first class citizens. \n\nWeaknesses: \n1. The idea sounds very familiar to explanation based learning. It will be good to contrast the two.\n2. My main criticism is that I am not sure we need an entirely new framework for targeting explanation based solutions. Many recent work in NLP try to frame problems in English to get cross task generalization. I would prefer this paper to be positioned as an improvement to such existing approaches, as this is tackling a harder problem class. \n3. The experimental results are all on the new game like dataset. It will be good to see performance of proposed methods on real datasets, or already existing synthetic datasets. The baselines are also weak, so its not clear if the proposed techniques will perform better than other cross task generalization methods.  ", "summary_of_the_review": "Although I am all for explanation based learning solutions, I don't think the components introduced in this paper warrants introduction of a new learning framework. I think it will be better to place it as an improvement to existing approaches for cross task generalization. I also found the experimental results to be weak. It was not clear if the proposed techniques will perform better than other cross task generalization methods.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635837225098}, {"id": "qs3d7g_Lg-R", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2684/Reviewer_PMcp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of classification learning where each data point is accompanied by some explanation. The explanation is in some arbitrary language. The paper proposes a way to increase the performance of the classification task by learning to predict the explanation associated with a given data point, and then using that explanation along with the data point to proceed to make a prediction about the data point's label. The paper implements this pipeline using neural networks, and presents empirical results on a new benchmark dataset to demonstrate its performance.  \n", "review_text": "Strengths:\n\n- Acknowledges the bilateral nature of explanations in building explainable AI/ML.\n- Offers a new benchmark problem.\n\nWeaknesses:\n\n- Does not fully explicate the assumptions that it is making in terms of the explanations.\n- Does not properly position itself with respect to relevant work in the literature.\n- The somewhat over-the-top philosophical discussion distracts from the essential point.\n\nGeneral remarks:\n\nThe problem acknowledges the need to have bilateral explainability; i.e., build machines that can explain, by first explaining *to* the machines. This is discussed, for example, by Michael, \"Machine Coaching\", IJCAI Workshop on XAI 2019, which seems to be rather relevant to the current paper, especially given that the paper makes an effort to connect to learning theory. Another line of work that seems to be relevant is Mozina et al., “Argument Based Machine Learning”, AIJ 2007, where data points are accompanied by an argument (i.e., an explanation) on why they are labeled as they are, as is the case in the current paper. Ignoring the obvious difference from the current work that these two works assume that explanations are in logic, the underlying theme of the cited papers and the current paper seems considerably close to ignore.  \n\nI found the philosophical positioning of the paper as a study in empiricism vs rationalism to be somewhat far-fetched. I believe it is instructive to take a step back and seek what are the underlying assumptions of the paper, and whether those bring something new to the table. \n\nIt is clear that the label of a data point is a function of the data point and the phenomenon, and not only of the data point. And since the phenomenon is essentially associated with an explanation, then the label is a function of the data point and the explanation. What the authors propose, then, as the pipeline for predicting the label is a form of chaining of learned pieces of knowledge, where one predicts the explanation, and then using that previous prediction one proceeds to predict the label at a second cycle of inferencing. Of relevance here is the work by Valiant, \"Robust Logics\", AIJ 2000, and Michael, \"Simultaneous Learning and Prediction\", KR 2014, as well as some follow-up papers by the authors, that establish the benefits of chaining.\n\nFrom a formal point of view, there is nothing in the explanations that makes them explanations. As the paper says, they are simply strings. Is it really the case that they are arbitrary strings, or should they be strings that are learnable as a function of the data points (or more properly, sets of data points)? That is to say, if each type of explanation for a phenomenon P_j was simply replaced by the number j, would then this still be considered an arbitrary language of explanations, despite having no structure and not being learnable? If indeed, explanations cannot be arbitrary strings, then one needs to carefully state what are the underlying assumptions on the language of explanations. If, on the other hand, explanations can be arbitrary (e.g., explanation j for phenomenon P_j) then this makes the term \"explanation\" rather mood; it is simply another signal in the data (at a meta-level, see my comment below). \n\nAnother aspect of the paper is the two levels of learning problems that exist, as mentioned in the last two paragraphs: the object-level learning problem of learning from a data point and its explanation the label of the data point; and the meta-level learning problem of learning from sets of data points to predict the phenomenon/explanation. Both of these two problems seem to follow the empiricist view, in the sense of the paper, which, as I said above, makes it unclear what the philosophical discussion on empiricism vs rationalism really add to the picture. The conjecture generator seems to be simply a meta-level classifier (albeit a stochastic one). Which brings me back again to the question on whether the explanations need to have some learnable structure.\n\nAdditional points:\n\nThe metric of NRS seems to include in its definition the identification of a nearest neighbor. Why? Shouldn't the goal be to predict the actual explanation? If the algorithm that makes the prediction wishes to use the nearest neighbor to reach that decision, then that would be fine. But the use of the nearest neighbor seems more natural to be part of the algorithm that attempts to solve the problem, not part of the evaluation metric for measuring success.\n\nThe proposed approach to solving the problem seems not to be accompanied by any formal guarantees on its performance. This would typically be compensated by an extensive experimental section, which is not the case here. \n\nI would have found a different narrative for this paper to be more convincing and impactful: the introduction of Odeen as a benchmark problem, and a deeper discussion of its features and parameters, and then present the particular approach, properly placed in the context of relevant work, as a suggested direction for what type of systems would presumably be useful in tackling the Odeen benchmark. \n\n------- After the Author Rebuttal -------\n\nI acknowledge that the authors have made an effort to engage with the points that I raised, and I have increased my score. I believe that the connections with learning theory are much more deep than the brief remarks offered in the revised version of the paper, and I hope that the authors will consider exploring them further in their future work as a formal underpinning of their empirical work. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of classification learning where each data point is accompanied by some explanation. The explanation is in some arbitrary language. The paper proposes a way to increase the performance of the classification task by learning to predict the explanation associated with a given data point, and then using that explanation along with the data point to proceed to make a prediction about the data point's label. The paper implements this pipeline using neural networks, and presents empirical results on a new benchmark dataset to demonstrate its performance.  \n", "main_review": "Strengths:\n\n- Acknowledges the bilateral nature of explanations in building explainable AI/ML.\n- Offers a new benchmark problem.\n\nWeaknesses:\n\n- Does not fully explicate the assumptions that it is making in terms of the explanations.\n- Does not properly position itself with respect to relevant work in the literature.\n- The somewhat over-the-top philosophical discussion distracts from the essential point.\n\nGeneral remarks:\n\nThe problem acknowledges the need to have bilateral explainability; i.e., build machines that can explain, by first explaining *to* the machines. This is discussed, for example, by Michael, \"Machine Coaching\", IJCAI Workshop on XAI 2019, which seems to be rather relevant to the current paper, especially given that the paper makes an effort to connect to learning theory. Another line of work that seems to be relevant is Mozina et al., “Argument Based Machine Learning”, AIJ 2007, where data points are accompanied by an argument (i.e., an explanation) on why they are labeled as they are, as is the case in the current paper. Ignoring the obvious difference from the current work that these two works assume that explanations are in logic, the underlying theme of the cited papers and the current paper seems considerably close to ignore.  \n\nI found the philosophical positioning of the paper as a study in empiricism vs rationalism to be somewhat far-fetched. I believe it is instructive to take a step back and seek what are the underlying assumptions of the paper, and whether those bring something new to the table. \n\nIt is clear that the label of a data point is a function of the data point and the phenomenon, and not only of the data point. And since the phenomenon is essentially associated with an explanation, then the label is a function of the data point and the explanation. What the authors propose, then, as the pipeline for predicting the label is a form of chaining of learned pieces of knowledge, where one predicts the explanation, and then using that previous prediction one proceeds to predict the label at a second cycle of inferencing. Of relevance here is the work by Valiant, \"Robust Logics\", AIJ 2000, and Michael, \"Simultaneous Learning and Prediction\", KR 2014, as well as some follow-up papers by the authors, that establish the benefits of chaining.\n\nFrom a formal point of view, there is nothing in the explanations that makes them explanations. As the paper says, they are simply strings. Is it really the case that they are arbitrary strings, or should they be strings that are learnable as a function of the data points (or more properly, sets of data points)? That is to say, if each type of explanation for a phenomenon P_j was simply replaced by the number j, would then this still be considered an arbitrary language of explanations, despite having no structure and not being learnable? If indeed, explanations cannot be arbitrary strings, then one needs to carefully state what are the underlying assumptions on the language of explanations. If, on the other hand, explanations can be arbitrary (e.g., explanation j for phenomenon P_j) then this makes the term \"explanation\" rather mood; it is simply another signal in the data (at a meta-level, see my comment below). \n\nAnother aspect of the paper is the two levels of learning problems that exist, as mentioned in the last two paragraphs: the object-level learning problem of learning from a data point and its explanation the label of the data point; and the meta-level learning problem of learning from sets of data points to predict the phenomenon/explanation. Both of these two problems seem to follow the empiricist view, in the sense of the paper, which, as I said above, makes it unclear what the philosophical discussion on empiricism vs rationalism really add to the picture. The conjecture generator seems to be simply a meta-level classifier (albeit a stochastic one). Which brings me back again to the question on whether the explanations need to have some learnable structure.\n\nAdditional points:\n\nThe metric of NRS seems to include in its definition the identification of a nearest neighbor. Why? Shouldn't the goal be to predict the actual explanation? If the algorithm that makes the prediction wishes to use the nearest neighbor to reach that decision, then that would be fine. But the use of the nearest neighbor seems more natural to be part of the algorithm that attempts to solve the problem, not part of the evaluation metric for measuring success.\n\nThe proposed approach to solving the problem seems not to be accompanied by any formal guarantees on its performance. This would typically be compensated by an extensive experimental section, which is not the case here. \n\nI would have found a different narrative for this paper to be more convincing and impactful: the introduction of Odeen as a benchmark problem, and a deeper discussion of its features and parameters, and then present the particular approach, properly placed in the context of relevant work, as a suggested direction for what type of systems would presumably be useful in tackling the Odeen benchmark. \n\n------- After the Author Rebuttal -------\n\nI acknowledge that the authors have made an effort to engage with the points that I raised, and I have increased my score. I believe that the connections with learning theory are much more deep than the brief remarks offered in the revised version of the paper, and I hope that the authors will consider exploring them further in their future work as a formal underpinning of their empirical work. \n", "summary_of_the_review": "I find Odeen to be a useful contribution, and one that would raise awareness on the need of certain underused techniques in the machine learning literature. But, the rest of the paper needs to be more clearly and properly placed in the context of existing work in the literature.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "n/a", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635712390161}], "openreview_url": "https://openreview.net/forum?id=46lmrnVBHBL", "arxiv_id": "2201.10222", "paper_pdf": "papers/46lmrnVBHBL.pdf", "paper_pdf_sha256": "926fd69698fdd89e21e85bc08fb3481222090d8aa755bcb519635c5b18e4f734", "paper_pdf_bytes": 8725249, "paper_pdf_source": "openreview", "code_url": "https://github.com/gladia-research-group/explanatory-learning", "code_repository": "gladia-research-group/explanatory-learning", "code_commit": "40d32f3ea8a2b7d875452de288cc8b611369423c", "code_archive": "repos/46lmrnVBHBL.zip", "code_archive_sha256": "c6883c7e9e619a07d5d9fcbbab0d7f2b7c3e4993dcee4d9e96556013535c9bda", "code_archive_bytes": 59820, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 61, "github_languages": {"Python": 168574}, "github_archived": false, "github_pushed_at": "2022-05-17T11:18:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/explanatory-learning-beyond-empiricism-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tckGH8K9y6o", "year": 2021, "status": "rejected", "title": "Symmetric Wasserstein Autoencoders", "authors": ["Sun Sun", "Hongyu Guo"], "authorids": ["~Sun_Sun1", "~Hongyu_Guo1"], "authors_source": "OpenReview API", "abstract": "Leveraging the framework of Optimal Transport, we introduce a new family of generative autoencoders with a learnable prior, called Symmetric Wasserstein Autoencoders (SWAEs). We propose to symmetrically match the joint distributions of the observed data and the latent representation induced by the encoder and the decoder. The resultant algorithm jointly optimizes the modelling losses in both the data and the latent spaces with  the loss in the data space leading to the denoising effect. With the symmetric treatment of the data and the latent representation, the algorithm implicitly preserves the local structure of the data in the latent space. To further improve the latent representation, we incorporate a reconstruction loss into the objective, which significantly benefits both the generation and reconstruction. We empirically show the superior performance of SWAEs over several state-of-the-art generative autoencoders in terms of classification, reconstruction, and generation.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "S3LWKJ_t6XN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1010/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed symmetric Wasserstein autoencoders (SWAE), which is a new type of generative autoencoders within the framework of optimal transport (OT). This work is based on Wasserstein autoencoders (WAE), but leverage a symmetric distance between the encoding distribution and the decoding distribution to better preserve the local structure of the data in the latent space. A reconstruction loss based regularization is suggested for better performance. The paper is well written and easy to follow. My only concern is that the improvement over WAE seems to be marginal. Below are some minor issues that need to be addressed. \n\n1. Please provide the appendix.\n\n2. The data sets used in the experiment are kind of too simple. Can the author provide results for full CIFAR10? \n\n3. The FID of WAE-GAN and WAE-MMD in Table 2 does not look right, can the author provide implementation info? Did you use your own code?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A nice generative autoencoder using optimal transport", "review": "This paper proposed symmetric Wasserstein autoencoders (SWAE), which is a new type of generative autoencoders within the framework of optimal transport (OT). This work is based on Wasserstein autoencoders (WAE), but leverage a symmetric distance between the encoding distribution and the decoding distribution to better preserve the local structure of the data in the latent space. A reconstruction loss based regularization is suggested for better performance. The paper is well written and easy to follow. My only concern is that the improvement over WAE seems to be marginal. Below are some minor issues that need to be addressed. \n\n1. Please provide the appendix.\n\n2. The data sets used in the experiment are kind of too simple. Can the author provide results for full CIFAR10? \n\n3. The FID of WAE-GAN and WAE-MMD in Table 2 does not look right, can the author provide implementation info? Did you use your own code?\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604156941683}, {"id": "C9gaiZC5oj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1010/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to treat the encoding and the decoding pairs symmetrically as a solution to OT problems. SWAE minimizes $p(x_d, z_d)$ and $p(x_e, z_e)$ in a jointly manner and shows better latent representation learning and generation. Moreover, the symmetric treatment for encoding and decoding shows an advantage in data denoising. \n\n\n----------\n\nIt is interesting to minimize the distance between $p(x_d, z_d)$ and $p(x_e, z_e)$ with the OT formulation.  The usage of the deterministic encoder and decoder could solve the problem of $W(p_{x_e}, p_{e_d})$ minimization, while it is difficult for the latent code since the latent codes from the prior and the posterior are not paired. Here it is good to see the authors apply aggregated posterior methods to solve this. The usage of the closest pseudo-inputs in constructing the pair of $z_e$ and $z_d$ could be a good approximation for $W(p_{z_e}, p_{z_d})$.\n\nMy questions mainly lie in the experiment part:\n\nThe authors report SWAE with three configurations, where $\\beta=1$, $\\beta=0.5$ and $\\beta=0$ respectively. When $\\beta \\rightarrow 1$, SWAE is only minimizing the $d( p(x_d, z_d), p(x_e, z_e) ) $; when $\\beta \\rightarrow 0$, SWAE behaves like an autoencoder with regularization in latent space.  The classification results and the reconstruction results of the case $\\beta = 0$  are missing. I am curious how these two cases perform, are they perform like VAE and Vampprior VAE? \n\nThe authors visualize the latent space with t-sne where the dimensionality of $z$ is 80 before the dimension reduction. As far as I know, when the dimension of $z$ is large enough, the t-sne results of VAEs and WAEs are similar to figure 3a, 3b and 3g. The results shown in c,d,e and f seem degenerated.  Especially 3c, it does not look like what we have seen in previous papers. For example, figure 2 in Makhzani et al, 2015 (https://arxiv.org/pdf/1511.05644.pdf) shows that the latent representations produced by VAE have several modes.  I doubt this could be the problem of dimension reduction. It will be more convincing if the author could show the visualization of these models where $dim(z) = 2$ and without any dimension reduction. \n\nFor the reconstruction results (not the denoising reconstruction) shown in Figure 9 and 10, the images seem to be dynamically binarized. However, in the description of the experiment details, this part is not mentioned. Considering SWAE is not applying a Bernoulli decoder, the dynamic binarization is not necessary. Moreover, the reconstruction from SWAE ($\\beta=0.5$) and SWAE ($\\beta=0$) seem to be binarized, which is not reasonable.\n\nWhen $\\beta=1$, a good baseline for SWAE will be ALI (Dumoulin et al., 2017), since ALI is minimizing $JS(p(x_d, z_d)|| p(x_e, z_e))$. I am curious how SWAE ($\\beta=1$) outperforms/underperforms compared to ALI and some corresponding analysis from the authors will be appreciated.\n\nIn the random generation results, only SWAE ($\\beta^\\star$) is shown. The generation results by SWAE ($\\beta=1$) and SWAE ($\\beta=0$) are missing while the quantitative results are shown in table 2.\n\n\n---------------\n\nIn general, the idea of symmetrically treating encoding and decoding to solve with the Wasserstein distance is interesting and is worthy of study. However, some details in the experiments remain unclear; some experiments results and the corresponding analysis are missing. The authors might need to dig out more to support their method. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method, but the experiments need to be refined", "review": "This paper proposes to treat the encoding and the decoding pairs symmetrically as a solution to OT problems. SWAE minimizes $p(x_d, z_d)$ and $p(x_e, z_e)$ in a jointly manner and shows better latent representation learning and generation. Moreover, the symmetric treatment for encoding and decoding shows an advantage in data denoising. \n\n\n----------\n\nIt is interesting to minimize the distance between $p(x_d, z_d)$ and $p(x_e, z_e)$ with the OT formulation.  The usage of the deterministic encoder and decoder could solve the problem of $W(p_{x_e}, p_{e_d})$ minimization, while it is difficult for the latent code since the latent codes from the prior and the posterior are not paired. Here it is good to see the authors apply aggregated posterior methods to solve this. The usage of the closest pseudo-inputs in constructing the pair of $z_e$ and $z_d$ could be a good approximation for $W(p_{z_e}, p_{z_d})$.\n\nMy questions mainly lie in the experiment part:\n\nThe authors report SWAE with three configurations, where $\\beta=1$, $\\beta=0.5$ and $\\beta=0$ respectively. When $\\beta \\rightarrow 1$, SWAE is only minimizing the $d( p(x_d, z_d), p(x_e, z_e) ) $; when $\\beta \\rightarrow 0$, SWAE behaves like an autoencoder with regularization in latent space.  The classification results and the reconstruction results of the case $\\beta = 0$  are missing. I am curious how these two cases perform, are they perform like VAE and Vampprior VAE? \n\nThe authors visualize the latent space with t-sne where the dimensionality of $z$ is 80 before the dimension reduction. As far as I know, when the dimension of $z$ is large enough, the t-sne results of VAEs and WAEs are similar to figure 3a, 3b and 3g. The results shown in c,d,e and f seem degenerated.  Especially 3c, it does not look like what we have seen in previous papers. For example, figure 2 in Makhzani et al, 2015 (https://arxiv.org/pdf/1511.05644.pdf) shows that the latent representations produced by VAE have several modes.  I doubt this could be the problem of dimension reduction. It will be more convincing if the author could show the visualization of these models where $dim(z) = 2$ and without any dimension reduction. \n\nFor the reconstruction results (not the denoising reconstruction) shown in Figure 9 and 10, the images seem to be dynamically binarized. However, in the description of the experiment details, this part is not mentioned. Considering SWAE is not applying a Bernoulli decoder, the dynamic binarization is not necessary. Moreover, the reconstruction from SWAE ($\\beta=0.5$) and SWAE ($\\beta=0$) seem to be binarized, which is not reasonable.\n\nWhen $\\beta=1$, a good baseline for SWAE will be ALI (Dumoulin et al., 2017), since ALI is minimizing $JS(p(x_d, z_d)|| p(x_e, z_e))$. I am curious how SWAE ($\\beta=1$) outperforms/underperforms compared to ALI and some corresponding analysis from the authors will be appreciated.\n\nIn the random generation results, only SWAE ($\\beta^\\star$) is shown. The generation results by SWAE ($\\beta=1$) and SWAE ($\\beta=0$) are missing while the quantitative results are shown in table 2.\n\n\n---------------\n\nIn general, the idea of symmetrically treating encoding and decoding to solve with the Wasserstein distance is interesting and is worthy of study. However, some details in the experiments remain unclear; some experiments results and the corresponding analysis are missing. The authors might need to dig out more to support their method. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603851750475}, {"id": "N6HBl1ivLwF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1010/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In the paper, the authors propose a new family of generative auto-encoders, named Symmetric Wasserstein Autoencoders (SWAEs), based on replacing the KL divergence between the encoding and decoding distributions on the traditional VAE framework into the Wasserstein metric between these distributions. Then, they also carried out experiments to demonstrate the favorable performance of their auto-encoder over previous base-line autoencoders.\n\nI think the SWAEs is quite interesting but lacks novelty. Here are my comments with the paper:\n\n(1) The result of Theorem 1 is under the assumption that both the encoder and decoder are deterministic, which is quite restrictive. Can the authors provide some understandings when either one of them is random? The result of Theorem 1 is also quite similar to the main result in the Wasserstein autoencoder work. I think the authors should at least cite this point properly.\n\n(2) Simply replacing KL divergence between the encoding and decoding distributions by Wasserstein metric to have symmetric properties sounds not novel. The KL divergence has a good interpretation in terms of ELBO. What can we interpret the Wasserstein metric between the encoding and decoding distributions? Theorem 1 does not seem convincing to me.\n\n(3) In the experiments, the choice of $\\beta = 1/2$ seems to yield best results, i.e., we should balance the reconstruction loss and the discrepancy in the data space in the objective function (4). Can the authors provide some intuition behind that?\n\n(4) The related work with generative modeling based on optimal transport lacks several relevant recent works; see for examples [1], which achieves SOTA among sliced-based Wasserstein distances in deep generative models, or [2], which proposes max-sliced Wasserstein distance for deep generative models.\n\nReferences:\n\n[1] K. Nguyen, N. Ho, T. Pham, H. Bui. Distributional sliced-Wasserstein and applications to deep generative modeling. Arxiv preprint Arxiv: 2002.07367, 2020.\n\n[2] I. Deshpande, Y. Hu, R. Sun, A. Pyrros, N. Siddiqui, S. Koyejo, Z. Zhao, D. Forsyth, A. Schwing. Max-sliced Wasserstein distance and its use for GANs. CVPR, 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposal of Symmetric Wasserstein Autoencoders is quite interesting but lacks novelty.", "review": "In the paper, the authors propose a new family of generative auto-encoders, named Symmetric Wasserstein Autoencoders (SWAEs), based on replacing the KL divergence between the encoding and decoding distributions on the traditional VAE framework into the Wasserstein metric between these distributions. Then, they also carried out experiments to demonstrate the favorable performance of their auto-encoder over previous base-line autoencoders.\n\nI think the SWAEs is quite interesting but lacks novelty. Here are my comments with the paper:\n\n(1) The result of Theorem 1 is under the assumption that both the encoder and decoder are deterministic, which is quite restrictive. Can the authors provide some understandings when either one of them is random? The result of Theorem 1 is also quite similar to the main result in the Wasserstein autoencoder work. I think the authors should at least cite this point properly.\n\n(2) Simply replacing KL divergence between the encoding and decoding distributions by Wasserstein metric to have symmetric properties sounds not novel. The KL divergence has a good interpretation in terms of ELBO. What can we interpret the Wasserstein metric between the encoding and decoding distributions? Theorem 1 does not seem convincing to me.\n\n(3) In the experiments, the choice of $\\beta = 1/2$ seems to yield best results, i.e., we should balance the reconstruction loss and the discrepancy in the data space in the objective function (4). Can the authors provide some intuition behind that?\n\n(4) The related work with generative modeling based on optimal transport lacks several relevant recent works; see for examples [1], which achieves SOTA among sliced-based Wasserstein distances in deep generative models, or [2], which proposes max-sliced Wasserstein distance for deep generative models.\n\nReferences:\n\n[1] K. Nguyen, N. Ho, T. Pham, H. Bui. Distributional sliced-Wasserstein and applications to deep generative modeling. Arxiv preprint Arxiv: 2002.07367, 2020.\n\n[2] I. Deshpande, Y. Hu, R. Sun, A. Pyrros, N. Siddiqui, S. Koyejo, Z. Zhao, D. Forsyth, A. Schwing. Max-sliced Wasserstein distance and its use for GANs. CVPR, 2019.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603843230526}, {"id": "QQyE01dU6n5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1010/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n \nThis works proposes an new auto-encoder variant based on an Optimal Transport (OT) penalty.  While there are many such previous works of OT and auto-encoders, this work proposes a joint OT penalty on data and latent space. As the scalability of computing OT penalties in high dimensions is a concern, the authors address this by restricting to deterministic encoders and decoders in Theorem 1, an extension to joint distributions of Theorem 1 of Tolstikhin 2018. The resulting algorithm amounts to a loss involving L2 penalties for (1) the reconstruction loss (2) decoded latents (conditional on \"pseudo-inputs\") and real samples (3) encoded samples and the conditional latents. Next experimental results are shown on small-scale datasets (MNIST, Fashion-MNIST, Coil20, subest of CIFAR-10) and compared against the VAE, WAE-{GAN,MMD}, VampPrior, and MIM.\n\n##########################################################################\n\nReasons for score: \n\nOverall I vote for rejection. Given the numerous prior works (properly cited in the paper), the novelty of the proposed loss appears limited to me. Theorem 1 also appears to be limited novelty over the Theorem 1 of Tolstikhin 2018. Additionally, the empirical evaluation of the method is only limited to small-scale datasets. \n \n\n##########################################################################\n\nPros: \n* Easy to read\n* Reasonable selection of prior models to compare against\n\n##########################################################################\n\nCons: \n* Only small-scale datasets are evaluated. Thus the empirical advantage is unclear\n* Theorem 1 appears to be of limited novelty over Theorem 1 in Tolstikin 2018\n\n\n##########################################################################\n\nQuestions during rebuttal period: \n \nWhy is your theorem novel over Theorem 1 in Tolstikin 2018?\n\nWhat is the performance of your method on more complex datasets such as CelebA and LSUN Bedroom?\n\n#########################################################################\n\nPOST-REBUTTAL RESPONSE:\n\nThanks for clarification on Theorem 1 of this paper, i.e. that the novelty is in interpretation. I agree that the \"denoising\" between observed and generated data is an interesting idea.\n\nI read the author's additional experiments on CelebA. In Figure 10, VampPrior is still qualitatively superior to the author's best result $SWAE(\\beta^*=0.5)$. I have some skepticism over the reported results for WAE-{GAN/MMD}, which are much worse than the results in the original paper (Tolstikhin 2018). The authors appear to have used different encoder/decoder architectures, which complicates the comparison. Is WAE's decreased performance due to choice of architecture or algorithm? FID scores on CelebA would be also helpful.\n\nAll told, I raise my score, but still harbor some doubts over the empirical advantage of this work.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear novelty of Theorem 1 over Tolstikhin 2018 / Empirical evaluations limited to small-scale datasets", "review": "##########################################################################\n\nSummary:\n \nThis works proposes an new auto-encoder variant based on an Optimal Transport (OT) penalty.  While there are many such previous works of OT and auto-encoders, this work proposes a joint OT penalty on data and latent space. As the scalability of computing OT penalties in high dimensions is a concern, the authors address this by restricting to deterministic encoders and decoders in Theorem 1, an extension to joint distributions of Theorem 1 of Tolstikhin 2018. The resulting algorithm amounts to a loss involving L2 penalties for (1) the reconstruction loss (2) decoded latents (conditional on \"pseudo-inputs\") and real samples (3) encoded samples and the conditional latents. Next experimental results are shown on small-scale datasets (MNIST, Fashion-MNIST, Coil20, subest of CIFAR-10) and compared against the VAE, WAE-{GAN,MMD}, VampPrior, and MIM.\n\n##########################################################################\n\nReasons for score: \n\nOverall I vote for rejection. Given the numerous prior works (properly cited in the paper), the novelty of the proposed loss appears limited to me. Theorem 1 also appears to be limited novelty over the Theorem 1 of Tolstikhin 2018. Additionally, the empirical evaluation of the method is only limited to small-scale datasets. \n \n\n##########################################################################\n\nPros: \n* Easy to read\n* Reasonable selection of prior models to compare against\n\n##########################################################################\n\nCons: \n* Only small-scale datasets are evaluated. Thus the empirical advantage is unclear\n* Theorem 1 appears to be of limited novelty over Theorem 1 in Tolstikin 2018\n\n\n##########################################################################\n\nQuestions during rebuttal period: \n \nWhy is your theorem novel over Theorem 1 in Tolstikin 2018?\n\nWhat is the performance of your method on more complex datasets such as CelebA and LSUN Bedroom?\n\n#########################################################################\n\nPOST-REBUTTAL RESPONSE:\n\nThanks for clarification on Theorem 1 of this paper, i.e. that the novelty is in interpretation. I agree that the \"denoising\" between observed and generated data is an interesting idea.\n\nI read the author's additional experiments on CelebA. In Figure 10, VampPrior is still qualitatively superior to the author's best result $SWAE(\\beta^*=0.5)$. I have some skepticism over the reported results for WAE-{GAN/MMD}, which are much worse than the results in the original paper (Tolstikhin 2018). The authors appear to have used different encoder/decoder architectures, which complicates the comparison. Is WAE's decreased performance due to choice of architecture or algorithm? FID scores on CelebA would be also helpful.\n\nAll told, I raise my score, but still harbor some doubts over the empirical advantage of this work.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603835555343}], "openreview_url": "https://openreview.net/forum?id=tckGH8K9y6o", "arxiv_id": "2106.13024", "paper_pdf": "papers/tckGH8K9y6o.pdf", "paper_pdf_sha256": "e1987ec91ea20e18b648ac668f21c02ccae1f2474361caaa6584ea3a3f1badfc", "paper_pdf_bytes": 38922030, "paper_pdf_source": "openreview", "code_url": "https://github.com/sunsunyyl/SWAE", "code_repository": "sunsunyyl/SWAE", "code_commit": "0554c3b67bddda189e4eeda40aab147963997281", "code_archive": "repos/tckGH8K9y6o.zip", "code_archive_sha256": "f8f68048b4e27d8222e8801e25b95baa42aa0716b28654fa1deeb839ddecf99f", "code_archive_bytes": 48086, "code_file_count": 23, "code_extensions": {".py": 22, ".sh": 1}, "github_disk_usage_kb": 62, "github_languages": {"Python": 195735, "Shell": 1463}, "github_archived": false, "github_pushed_at": "2021-06-29T19:32:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/symmetric-wasserstein-autoencoders-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkgStySKPB", "year": 2020, "status": "rejected", "title": "Contrastive Multiview Coding", "authors": ["Yonglong Tian", "Dilip Krishnan", "Phillip Isola"], "authorids": ["yonglong@mit.edu", "dilipkay@google.com", "phillipi@mit.edu"], "authors_source": "OpenReview API", "abstract": "Humans view the world through many sensory channels, e.g., the long-wavelength light channel, viewed by the left eye, or the high-frequency vibrations channel, heard by the right ear. Each view is noisy and incomplete, but important factors, such as physics, geometry, and semantics, tend to be shared between all views (e.g., a \"dog\" can be seen, heard, and felt). We hypothesize that a powerful representation is one that models view-invariant factors. Based on this hypothesis, we investigate a contrastive coding scheme, in which a representation is learned that aims to maximize mutual information between different views but is otherwise compact. Our approach scales to any number of views, and is view-agnostic. The resulting learned representations perform above the state of the art for downstream tasks such as object classification, compared to formulations based on predictive learning or single view reconstruction, and improve as more views are added. On the Imagenet linear readoff benchmark, we achieve 68.4% top-1 accuracy. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ByxjhmgRYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1840/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This interesting paper on an important topic; however, its readability could be dramatically improved, especially for the reader less familiar with the problem. \n\nIn order to make the paper more accessible, the authors should reorganize the introduction by breaking it down into two parts:\n1) a more traditional introduction \n- one intuitive paragraph about multi-view coding\n- one intuitive paragraph with an illustrative example on how the proposed approach will help solve a problem; at the same intuitive level, compare-and-contrast it with existing approaches \n- one intuitive, in-detail paragraph on how the proposed approach works\n- one paragraph summarizing the main findings/results \n2) a second, new section, that will turn the current Figures 1 & 2 into a complete description of an illustrative example (the current, detailed \"captions\" are a good start, but they should be fleshed out into a full, detailed section of the paper)  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This interesting paper on an important topic; however, its readability could be dramatically improved, especially for the reader less familiar with the problem. \n\nIn order to make the paper more accessible, the authors should reorganize the introduction by breaking it down into two parts:\n1) a more traditional introduction \n- one intuitive paragraph about multi-view coding\n- one intuitive paragraph with an illustrative example on how the proposed approach will help solve a problem; at the same intuitive level, compare-and-contrast it with existing approaches \n- one intuitive, in-detail paragraph on how the proposed approach works\n- one paragraph summarizing the main findings/results \n2) a second, new section, that will turn the current Figures 1 & 2 into a complete description of an illustrative example (the current, detailed \"captions\" are a good start, but they should be fleshed out into a full, detailed section of the paper)  \n"}, "tcdate": 1571845042685}, {"id": "SkgXlgdptr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1840/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a new self-supervised learning methods by utilizing contrastive predictive coding technique.  The proposed algorithm is more effective than existing self-supervised learning algorithm.  The presented results are encouraging.  \n1. In section 3.2,  the authors show that  a large number of views would improve the representation quality. However,  multi-views may provide redundancy information. What is the core information that affect  the representation quality?\n\nIn fact,  I am not an expert on self-supervised learning and  contrastive predictive coding,  so my reviewer confidence is low.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This paper proposed a new self-supervised learning methods by utilizing contrastive predictive coding technique.  The proposed algorithm is more effective than existing self-supervised learning algorithm.  The presented results are encouraging.  \n1. In section 3.2,  the authors show that  a large number of views would improve the representation quality. However,  multi-views may provide redundancy information. What is the core information that affect  the representation quality?\n\nIn fact,  I am not an expert on self-supervised learning and  contrastive predictive coding,  so my reviewer confidence is low."}, "tcdate": 1571811307339}, {"id": "r1lX1jDjtH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1840/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presented a multi-view learning method that is based on negative sampling in contrastive learning. The core idea is to set an anchor view and the sample positive and negative data points from the other view and maximise the agreement between positive pairs in learning from two views. When more than two views are presented, the learning objective is a sum over all possible combinations of two views. The performance of the proposed model is good, and the ablation study is interesting. \n\nComments:\n\n1. The core concept, or at least one of the core concepts, in multi-view learning is the conditional independence.\n\nNormally, the underlying assumption in multi-view learning is that, given the class label, the samples from multiple views are conditionally independent from each other. Therefore, the goal is to learn distinctive representations from different data sources/disjoint populations, so then after learning, the ensemble of them is able to capture a set of diverse aspects of the data. A \"side-effect\" of learning from multiple views is that individual views indeed get improved by learning from others. Meanwhile, self-supervised learning is the case when the input data to the designed learning system is also the target of the system. \n\nThe paper presented an idea for self-supervised learning from multiple views, which is not exactly the same, but still in the same regime. This concept could be used to explain some empirical findings in this paper. Since it is expected, there is even no need in conducting experiments. \n\n\n\n\n2. My main concern of this paper is the novelty, however, the empirical results are strong.\n\nThe paper mainly presented a simple yet effective method for self-supervised learning from two views, and the generalisation is a sum over all possible combinations of two views. The method itself has already been proposed many years ago as mentioned in the related work section in the paper, and the generalisation was also described in prior work, which makes me doubt the novelty of the paper. \n\nThe earliest work to the best of my knowledge is [1], and later on there are a couple workshops [2,3] on multi-view learning which largely settled the field of learning from multiple views from neural networks', kernels', and bayesian perspectives. Many things mentioned in this paper have already been discovered at that time. \n\n3. The theoretical justification is not as strong as the generalised CCA.\n\nCCA has been applied in the field of multi-view learning and self-supervised learning for long, and it was initially proposed for comparing the correlation between two sets of samples of two random variables. A successful generalisation is the generalised CCA which is capable of learning from multiple views. The formula of GCCA as referred in [4] is simple and elegant, and then the extension of using neural networks is also straightforward. Since people has relatively clearer understanding of CCA itself, the generalised version or the kernel version of it is also well-understood. \n\nA nice theoretical understanding of contrastive unsupervised learning is provided in [5], and I recommend the authors to study.\n\n\n[1] de Sa, Virginia R. \"Learning classification with unlabeled data.\" Advances in neural information processing systems. 1994.\n[2] ICML Workshop, \"Learning With Multiple Views\". 2005\n[3] NIPS workshop, \"Learning from multiple sources\". 2008\n[4] Benton, Adrian, et al. \"Deep generalized canonical correlation analysis.\" ICLR workshop 2017.\n[5] Arora, Sanjeev, et al. \"A theoretical analysis of contrastive unsupervised representation learning.\" ICML 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper presented a multi-view learning method that is based on negative sampling in contrastive learning. The core idea is to set an anchor view and the sample positive and negative data points from the other view and maximise the agreement between positive pairs in learning from two views. When more than two views are presented, the learning objective is a sum over all possible combinations of two views. The performance of the proposed model is good, and the ablation study is interesting. \n\nComments:\n\n1. The core concept, or at least one of the core concepts, in multi-view learning is the conditional independence.\n\nNormally, the underlying assumption in multi-view learning is that, given the class label, the samples from multiple views are conditionally independent from each other. Therefore, the goal is to learn distinctive representations from different data sources/disjoint populations, so then after learning, the ensemble of them is able to capture a set of diverse aspects of the data. A \"side-effect\" of learning from multiple views is that individual views indeed get improved by learning from others. Meanwhile, self-supervised learning is the case when the input data to the designed learning system is also the target of the system. \n\nThe paper presented an idea for self-supervised learning from multiple views, which is not exactly the same, but still in the same regime. This concept could be used to explain some empirical findings in this paper. Since it is expected, there is even no need in conducting experiments. \n\n\n\n\n2. My main concern of this paper is the novelty, however, the empirical results are strong.\n\nThe paper mainly presented a simple yet effective method for self-supervised learning from two views, and the generalisation is a sum over all possible combinations of two views. The method itself has already been proposed many years ago as mentioned in the related work section in the paper, and the generalisation was also described in prior work, which makes me doubt the novelty of the paper. \n\nThe earliest work to the best of my knowledge is [1], and later on there are a couple workshops [2,3] on multi-view learning which largely settled the field of learning from multiple views from neural networks', kernels', and bayesian perspectives. Many things mentioned in this paper have already been discovered at that time. \n\n3. The theoretical justification is not as strong as the generalised CCA.\n\nCCA has been applied in the field of multi-view learning and self-supervised learning for long, and it was initially proposed for comparing the correlation between two sets of samples of two random variables. A successful generalisation is the generalised CCA which is capable of learning from multiple views. The formula of GCCA as referred in [4] is simple and elegant, and then the extension of using neural networks is also straightforward. Since people has relatively clearer understanding of CCA itself, the generalised version or the kernel version of it is also well-understood. \n\nA nice theoretical understanding of contrastive unsupervised learning is provided in [5], and I recommend the authors to study.\n\n\n[1] de Sa, Virginia R. \"Learning classification with unlabeled data.\" Advances in neural information processing systems. 1994.\n[2] ICML Workshop, \"Learning With Multiple Views\". 2005\n[3] NIPS workshop, \"Learning from multiple sources\". 2008\n[4] Benton, Adrian, et al. \"Deep generalized canonical correlation analysis.\" ICLR workshop 2017.\n[5] Arora, Sanjeev, et al. \"A theoretical analysis of contrastive unsupervised representation learning.\" ICML 2019."}, "tcdate": 1571678939453}], "openreview_url": "https://openreview.net/forum?id=BkgStySKPB", "arxiv_id": "1906.05849", "paper_pdf": "papers/BkgStySKPB.pdf", "paper_pdf_sha256": "7eaa203b54d639db823de64b9560e8d0698cbc15c4b6beda0e8ea7a43cd23cdf", "paper_pdf_bytes": 1495630, "paper_pdf_source": "openreview", "code_url": "https://github.com/HobbitLong/CMC", "code_repository": "HobbitLong/CMC", "code_commit": "7b227be0b10ef4e526c72af07664f5079ed9ee09", "code_archive": "repos/BkgStySKPB.zip", "code_archive_sha256": "79418356244c8040ee8d63d8e7b042c1ab45d8ae7d73cefe779c3c7761865fb9", "code_archive_bytes": 32632, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 76, "github_languages": {"Python": 98729}, "github_archived": false, "github_pushed_at": "2020-11-10T07:33:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contrastive-multiview-coding"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkedwoC5t7", "year": 2019, "status": "rejected", "title": "Formal Limitations on the Measurement of Mutual Information", "authors": ["David McAllester", "Karl Stratos"], "authorids": ["mcallester@ttic.edu", "stratos@ttic.edu"], "authors_source": "OpenReview API", "abstract": "Motivated by applications to unsupervised learning, we consider the problem of measuring mutual information. Recent analysis has shown that naive kNN estimators of mutual information have serious statistical limitations motivating more refined methods. In this paper we prove that serious statistical limitations are inherent to any measurement method. More specifically, we show that any distribution-free high-confidence lower bound on mutual information cannot be larger than $O(\\ln N)$ where $N$ is the size of the data sample. We also analyze the Donsker-Varadhan lower bound on KL divergence in particular and show that, when simple statistical considerations are taken into account, this bound can never produce a high-confidence value larger than $\\ln N$. While large high-confidence lower bounds are impossible, in practice one can use estimators without formal guarantees. We suggest expressing mutual information as a difference of entropies and using cross entropy as an entropy estimator.  We observe that, although cross entropy is only an upper bound on entropy, cross-entropy estimates converge to the true cross entropy at the rate of $1/\\sqrt{N}$.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1lxFreb6X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper277/AnonReviewer1"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of estimating bounds on the mutual information. It begins by showing that popular recent estimators (e.g. MINE) are flawed, since they rely on the Donsker-Varadhan bound that cannot be estimated efficiently. They then point to entropy upper bounds as a much more feasible approach to MI approximation, and propose a framework for using them in practice. These upper bounds converge as 1/sqrt(N) to the true entropy value, making them potentially viable in practice to obtain reasonable approximations of MI. \n\nGiven the significant recent attention payed to the MINE estimator and to MI estimation in machine learning generally, I think the message of this paper is very critical to the machine learning community. This analysis of the MINE estimator alone would warrant publication. \n\nThere are currently some weaknesses to this paper, however, when compared to the typical ICLR paper. If the following issues are addressed I am prepared to raise my scoring of this paper:\n1. Provide some intuition on how to apply this type of analysis to the continuous-valued case, or some reason why such an analysis would require a different framework. The more detail the better, if bounds analogous to Theorem 1 could be proved for continuous variables and added to the paper, that would be excellent.\n2. Section 4 as written is intuitively clear, but could greatly benefit from a full rigorous analysis. There is plenty of space in the paper to do this (and the supplement is available if needed). If such an analysis can’t be done, this section should be deleted.\n3. Some empirical example of MINE converging slowly in practice would greatly add to the impact of the paper.\n\nMinor issues: \nLast sentence of Section 1 should clarify what “true entropy” means. As written it is ambiguous between the actual entropy or the cross-entropy, since both are entropies.\n\nI understand the nested optimization problem in Section 6, but the presentation is somewhat unclearly written. More exposition here would help, along with a more clear step-by-step explanation of the practical procedure.\n\nEDIT: The authors have addressed my concerns and I have raised my score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Timely discussion of approaches for approximating mutual information ", "review": "This paper considers the problem of estimating bounds on the mutual information. It begins by showing that popular recent estimators (e.g. MINE) are flawed, since they rely on the Donsker-Varadhan bound that cannot be estimated efficiently. They then point to entropy upper bounds as a much more feasible approach to MI approximation, and propose a framework for using them in practice. These upper bounds converge as 1/sqrt(N) to the true entropy value, making them potentially viable in practice to obtain reasonable approximations of MI. \n\nGiven the significant recent attention payed to the MINE estimator and to MI estimation in machine learning generally, I think the message of this paper is very critical to the machine learning community. This analysis of the MINE estimator alone would warrant publication. \n\nThere are currently some weaknesses to this paper, however, when compared to the typical ICLR paper. If the following issues are addressed I am prepared to raise my scoring of this paper:\n1. Provide some intuition on how to apply this type of analysis to the continuous-valued case, or some reason why such an analysis would require a different framework. The more detail the better, if bounds analogous to Theorem 1 could be proved for continuous variables and added to the paper, that would be excellent.\n2. Section 4 as written is intuitively clear, but could greatly benefit from a full rigorous analysis. There is plenty of space in the paper to do this (and the supplement is available if needed). If such an analysis can’t be done, this section should be deleted.\n3. Some empirical example of MINE converging slowly in practice would greatly add to the impact of the paper.\n\nMinor issues: \nLast sentence of Section 1 should clarify what “true entropy” means. As written it is ambiguous between the actual entropy or the cross-entropy, since both are entropies.\n\nI understand the nested optimization problem in Section 6, but the presentation is somewhat unclearly written. More exposition here would help, along with a more clear step-by-step explanation of the practical procedure.\n\nEDIT: The authors have addressed my concerns and I have raised my score.\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541633399640}, {"id": "Syl6cQcsnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper277/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is about estimating mutual information in high dimensional settings. This is a very challenging open problem, that is of interest to a diverse set of research communities.\n\nIn this paper, it is theoretically argued that the recent proposed mutual information (lower bound) estimator, MINE, that is based on the Donsker-Varadhan representation of the corresponding KL divergence expression for mutual infortmation, is fundamentally flawed for high dimensions (of discrete variables).  It is further shown that lower bounds for joint entropy are hard to obtain due to exponential sample complexity. So, the authors suggest to instead obtain an upper bound for each entropy term in the mutual information expression; cross-entropy is the suggested upper bound for an entropy term.\n\nI have some basic questions as the following.\n\nSince the recent KL divergence based MI estimator, MINE, is inaccurate in high dimensions, there should be at least a discussion on connections between your estimator and the classic nearest neighbor distances based estimator of Kraskov et al. and the extensions (I suppose, even for discrete variables, one can compute distances to obtain nearest neighbors efficiently). Also, there are kernel functions based estimators.\n\nThere is no discussion in the paper about the errors accumulating from individual entropy terms in the mutual information expression. Kraskov et al. talk about this problem of accumulating errors in their seminal paper and propose not to compute the entropy terms individually. What you are proposing is in contrast to their clever observations.\n\nDoes the analysis on upper bound for entropy term also apply to the conditional entropy in the mutual information expression ? I think, there are more subtleties that should be explained.\n\nSince the proposed approach upper bounds entropy using cross entropy term (i.e. using some machine learning model like a neural network), it is even more important to show solid empirical evaluation, for synthetic as well as real world data.\n\nThere is a subtle difference between estimating mutual information and proposing an upper/lower bound for it. At present, it is not clear if the proposed upper bound of entropy would lead to an overall upper bound or lower bound for the mutual infortmation expression. The latter is important to know both in the context of optimization based on mutual information maximization (it should be lower bound in such case), as well analyzing mutual infortmation to under complex dynamics such as in brain.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising work from theoretical standpoint", "review": "This paper is about estimating mutual information in high dimensional settings. This is a very challenging open problem, that is of interest to a diverse set of research communities.\n\nIn this paper, it is theoretically argued that the recent proposed mutual information (lower bound) estimator, MINE, that is based on the Donsker-Varadhan representation of the corresponding KL divergence expression for mutual infortmation, is fundamentally flawed for high dimensions (of discrete variables).  It is further shown that lower bounds for joint entropy are hard to obtain due to exponential sample complexity. So, the authors suggest to instead obtain an upper bound for each entropy term in the mutual information expression; cross-entropy is the suggested upper bound for an entropy term.\n\nI have some basic questions as the following.\n\nSince the recent KL divergence based MI estimator, MINE, is inaccurate in high dimensions, there should be at least a discussion on connections between your estimator and the classic nearest neighbor distances based estimator of Kraskov et al. and the extensions (I suppose, even for discrete variables, one can compute distances to obtain nearest neighbors efficiently). Also, there are kernel functions based estimators.\n\nThere is no discussion in the paper about the errors accumulating from individual entropy terms in the mutual information expression. Kraskov et al. talk about this problem of accumulating errors in their seminal paper and propose not to compute the entropy terms individually. What you are proposing is in contrast to their clever observations.\n\nDoes the analysis on upper bound for entropy term also apply to the conditional entropy in the mutual information expression ? I think, there are more subtleties that should be explained.\n\nSince the proposed approach upper bounds entropy using cross entropy term (i.e. using some machine learning model like a neural network), it is even more important to show solid empirical evaluation, for synthetic as well as real world data.\n\nThere is a subtle difference between estimating mutual information and proposing an upper/lower bound for it. At present, it is not clear if the proposed upper bound of entropy would lead to an overall upper bound or lower bound for the mutual infortmation expression. The latter is important to know both in the context of optimization based on mutual information maximization (it should be lower bound in such case), as well analyzing mutual infortmation to under complex dynamics such as in brain.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541280661130}, {"id": "rJlF2U6MoX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper277/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studied the Donsker-Varadhan lower bound of KL-divergence. The authors show that with high probability, the DV lower bound is upper bounded by log of the sample size, so if the true KL-divergence is very large, then exponential sample size is needed to make the DV lower bound tight. The same argument holds true for any distribution-free high-confidence lower bound (such as DV lower bound) for KL divergence. Then the authors proposed to use an upper bound for entropy instead of lower bound for mutual information. \n\nThe idea of the paper is interesting and the proof of Theorems 1 and 2 are valid. Especially I like the idea of Theorem 1, which proves that any distribution-free high-confidence lower bound for KL divergence is upper bounded by log of sample size. This idea is similar to the paper in (Gao et al 15') which shows that mutual information estimator is upper bounded by log(N).\n\nHowever, this paper contains many fatal flaws, which significantly weaken the quality of this paper. Precisely,\n\n1. The DV lower bound is just an alternative of mutual information, helping MMI predictive coding algorithm to find good coding functions C_x and C_y. The goal of MMI predictive coding is not to estimate the mutual information I(C_x(x), C_y(y)) precisely, instead, the goal is to find good coding functions. The fact that DV lower bound is small means that we can not estimate mutual information through DV lower bound, but it does not directly imply that we can not find the coding functions. I expect some experiments to show that when mutual information is large, MMI predictive coding using DV lower bound can not find good coding functions.\n\n2. In Section 3, the formula after the proof of Outlier Risk Lemma and before Theorem 1 (btw, it is better to have numbers for these formula) seems to be problematic. The first formula shows that E_{z~q} e^{F(z)} >= (1/N)e^{F_max}, then we plug it in (4). But in (4) there is negative ln of E_{z~q} e^{F(z)}, so we should have KL(P,Q) <= something, correct? This may be a typo but this typo is so important such that it affect the readability a lot. Theorem 1 is correct but the paragraphs before Theorem 1 confuse the reader a lot.\n\n3. In Section 4, are you considering classical entropy for discrete random variables, or differential entropy for continuous random variables? I assume you are considering the latter, since most people of machine learning community are interested in continuous random variables. Then your statement of I(X;Y) <= H(X) is incorrect, since for continuous random variables, H(X|Y) can be negative. See (Thomas & Cover, Chapter 8) for a reference.\n\n4. Related to problem 3, if you are considering continuous random variables, then the statement I(X;Y)=H(X)+H(Y)-H(X,Y) is not always correct. There are cases that H(X) is infinite, H(Y) is infinite, H(X,Y) is infinite but I(X;Y) is finite. These cases does not only exist in mathematical books, but also exists in practice, especially when the data is located on a low-dimensional manifold embedded in a high-dimensional space. Therefore, your approach of decompose the mutual information is not always possible.\n\n5. Regarding to your proposed optimization problem in Section 6 (also it is better to have a number), I have some concerns. Since it involves max over \\Psi outside, and inf over \\Theta and inf over \\Phi inside, so I wonder how do you solve this problem? Can you guarantee that the solution can provide good coding functions C_x and C_y? Also, it seems that this optimization problem is proposed as an improvement over the DV lower bound method, so I wish to see some experiment showing that this method is better than the DV lower bound method, at least for some synthetic datasets.\n\nBecause of the above mentioned flaws (especially 3 and 4, and lack of experiments), I think the paper is below the standard of ICLR conference.\n\nReferences:\n[1] Efficient Estimation of Mutual Information for Strongly Dependent Variables, by Gao, Ver Steeg and Galstyan, AISTATS15'\n[2] Elements of Information Theory, 2nd edition, by Thomas and Cover.\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "ICLR 2019 Conference Paper277 AnonReviewer3", "review": "This paper studied the Donsker-Varadhan lower bound of KL-divergence. The authors show that with high probability, the DV lower bound is upper bounded by log of the sample size, so if the true KL-divergence is very large, then exponential sample size is needed to make the DV lower bound tight. The same argument holds true for any distribution-free high-confidence lower bound (such as DV lower bound) for KL divergence. Then the authors proposed to use an upper bound for entropy instead of lower bound for mutual information. \n\nThe idea of the paper is interesting and the proof of Theorems 1 and 2 are valid. Especially I like the idea of Theorem 1, which proves that any distribution-free high-confidence lower bound for KL divergence is upper bounded by log of sample size. This idea is similar to the paper in (Gao et al 15') which shows that mutual information estimator is upper bounded by log(N).\n\nHowever, this paper contains many fatal flaws, which significantly weaken the quality of this paper. Precisely,\n\n1. The DV lower bound is just an alternative of mutual information, helping MMI predictive coding algorithm to find good coding functions C_x and C_y. The goal of MMI predictive coding is not to estimate the mutual information I(C_x(x), C_y(y)) precisely, instead, the goal is to find good coding functions. The fact that DV lower bound is small means that we can not estimate mutual information through DV lower bound, but it does not directly imply that we can not find the coding functions. I expect some experiments to show that when mutual information is large, MMI predictive coding using DV lower bound can not find good coding functions.\n\n2. In Section 3, the formula after the proof of Outlier Risk Lemma and before Theorem 1 (btw, it is better to have numbers for these formula) seems to be problematic. The first formula shows that E_{z~q} e^{F(z)} >= (1/N)e^{F_max}, then we plug it in (4). But in (4) there is negative ln of E_{z~q} e^{F(z)}, so we should have KL(P,Q) <= something, correct? This may be a typo but this typo is so important such that it affect the readability a lot. Theorem 1 is correct but the paragraphs before Theorem 1 confuse the reader a lot.\n\n3. In Section 4, are you considering classical entropy for discrete random variables, or differential entropy for continuous random variables? I assume you are considering the latter, since most people of machine learning community are interested in continuous random variables. Then your statement of I(X;Y) <= H(X) is incorrect, since for continuous random variables, H(X|Y) can be negative. See (Thomas & Cover, Chapter 8) for a reference.\n\n4. Related to problem 3, if you are considering continuous random variables, then the statement I(X;Y)=H(X)+H(Y)-H(X,Y) is not always correct. There are cases that H(X) is infinite, H(Y) is infinite, H(X,Y) is infinite but I(X;Y) is finite. These cases does not only exist in mathematical books, but also exists in practice, especially when the data is located on a low-dimensional manifold embedded in a high-dimensional space. Therefore, your approach of decompose the mutual information is not always possible.\n\n5. Regarding to your proposed optimization problem in Section 6 (also it is better to have a number), I have some concerns. Since it involves max over \\Psi outside, and inf over \\Theta and inf over \\Phi inside, so I wonder how do you solve this problem? Can you guarantee that the solution can provide good coding functions C_x and C_y? Also, it seems that this optimization problem is proposed as an improvement over the DV lower bound method, so I wish to see some experiment showing that this method is better than the DV lower bound method, at least for some synthetic datasets.\n\nBecause of the above mentioned flaws (especially 3 and 4, and lack of experiments), I think the paper is below the standard of ICLR conference.\n\nReferences:\n[1] Efficient Estimation of Mutual Information for Strongly Dependent Variables, by Gao, Ver Steeg and Galstyan, AISTATS15'\n[2] Elements of Information Theory, 2nd edition, by Thomas and Cover.\n\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1539655344555}], "openreview_url": "https://openreview.net/forum?id=BkedwoC5t7", "arxiv_id": "1811.04251", "paper_pdf": "papers/BkedwoC5t7.pdf", "paper_pdf_sha256": "09808aa345082d851bd65a9b259ed82c81d8ec3670116fa78ba58ae363c9f68b", "paper_pdf_bytes": 186582, "paper_pdf_source": "openreview", "code_url": "https://github.com/karlstratos/doe", "code_repository": "karlstratos/doe", "code_commit": "84b5d5c394d5e6be3455a48ced742c231e2fef5c", "code_archive": "repos/BkedwoC5t7.zip", "code_archive_sha256": "a99bb40e031255a74b2caae7b89a33d6fc7d422be17a9561f89f1b1700c65231", "code_archive_bytes": 250905, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 476, "github_languages": {"Python": 21116}, "github_archived": false, "github_pushed_at": "2022-04-13T19:15:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/formal-limitations-on-the-measurement-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HytSvlWRZ", "year": 2018, "status": "rejected", "title": "Subspace Network: Deep Multi-Task Censored Regression for Modeling Neurodegenerative Diseases", "authors": ["Mengying Sun", "Inci M. Baytas", "Zhangyang Wang", "Jiayu Zhou"], "authorids": ["sunmeng2@msu.edu", "baytasin@msu.edu", "atlaswang@tamu.edu", "jiayuz@msu.edu"], "authors_source": "OpenReview API", "abstract": "Over the past decade a wide spectrum of machine learning models have been developed to model the neurodegenerative diseases, associating biomarkers, especially non-intrusive neuroimaging markers, with key clinical scores measuring the cognitive status of patients. Multi-task learning (MTL) has been extensively explored in these studies to address challenges associated to high dimensionality and small cohort size. However, most existing MTL approaches are based on linear models and suffer from two major limitations: 1) they cannot explicitly consider upper/lower bounds in these clinical scores; 2) they lack the capability to capture complicated non-linear effects among the variables. In this paper, we propose the Subspace Network, an efficient deep modeling approach for non-linear multi-task censored regression. Each layer of the subspace network performs a multi-task censored regression to improve upon the predictions from the last layer via sketching a low-dimensional subspace to perform knowledge transfer among learning tasks. We show that under mild assumptions, for each layer the parametric subspace can be recovered using only one pass of training data. In addition, empirical results demonstrate that the proposed subspace network quickly picks up correct parameter subspaces, and outperforms state-of-the-arts in predicting neurodegenerative clinical scores using information in brain imaging. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1z2QSOlz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper599/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a DNN, called subspace network, for nonlinear multi-task censored regression problem. The topic is important. Experiments on real data show improvements compared to several traditional approaches.\n\nMy major concerns are as follows.\n\n1. The paper is not self-contained. The authors claim that they establish both asymptotic and non-asymptotic convergence properties for Algorithm 1. However, for some key steps in the proof, they refer to other references. If this is due to space limitation in the main text, they may want to provide a complete proof in the appendix.\n\n2. The experiments are unconvincing. They compare the proposed SN with other traditional approaches on a very small data  set with 670 samples and 138 features. A major merit of DNN is that it can automatically extract useful features. However, in this experiment, the features are handcrafted before they are fed into the models. Thus, I would like to see a comparison between SN with vanilla DNN. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors propose a DNN, called subspace network, for nonlinear multi-task censored regression problem. The writing needs more elaboration. The experiments are unconvincing.", "rating": "5: Marginally below acceptance threshold", "review": "The authors propose a DNN, called subspace network, for nonlinear multi-task censored regression problem. The topic is important. Experiments on real data show improvements compared to several traditional approaches.\n\nMy major concerns are as follows.\n\n1. The paper is not self-contained. The authors claim that they establish both asymptotic and non-asymptotic convergence properties for Algorithm 1. However, for some key steps in the proof, they refer to other references. If this is due to space limitation in the main text, they may want to provide a complete proof in the appendix.\n\n2. The experiments are unconvincing. They compare the proposed SN with other traditional approaches on a very small data  set with 670 samples and 138 features. A major merit of DNN is that it can automatically extract useful features. However, in this experiment, the features are handcrafted before they are fed into the models. Thus, I would like to see a comparison between SN with vanilla DNN. ", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1511703465551}, {"id": "B1r0SU9gz", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper599/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new multi-task network architecture within which low-rank parameter spaces were found using matrix factorization. As carefully proved and tested, only one pass of the training data would help recover the parametric subspace, thus network could be easily trained layer-wise and expanded.\n\nSome novel contributions:\n1. Layer by layer feedforward training process, no back-prop.\n2. On-line settings to train parameters ( guaranteed convergence in a single pass of the data)\n\nWeakness :\n1. The assumption that a low-rank parameter space exists among tasks rather than original feature spaces is not new and widely used in literature.\n2. The proof part(Section 2.2) can be extended with more details in Appendix.\n3. In synthetic data experiments (Table1), only small margins could be observed between SN, f-MLP and rf-MLP, and only Layer 1 of SN performs better above all others. \n4. Typo: In Table2,3,5, Multi-l_{2,1} (denotes the L2,1 norm) were written wrong.\n5. In the synthetic data experiments on comparison with single-task and multi-task models, counter-intuitive results (with larger training data split, ANMSE raises instead of decreases) of multi-task models may need further explanation. \n6. Extra models like Deep Networks with/without matrix factorization could be added. ( As proposed model is a deep model, the lack of comparison with deep methods is dubious)\n7. In Section 4.2, the real dataset is rather small thus the results on this small dataset were not convincing enough. SN model outperforms the state-of-the-art with only small margin. Extensive experiments could be added.\n8. The performance on One-Layer Subspace Network (with only the input features) could be added. \n\nConclusion:\nThough with a quite novel idea on solving multi-task censored regression problem, the experiments conducted on synthetic data and real data are not convincing enough to ensure the contribution of the Subspace Network. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper introduces a multi-task network architecture within which low-rank parameter spaces were found using matrix factorization. As carefully proved and tested, only one pass of the training data would help recover the parametric subspace, thus network could be easily trained layer-wise and expanded.", "rating": "5: Marginally below acceptance threshold", "review": "This paper presents a new multi-task network architecture within which low-rank parameter spaces were found using matrix factorization. As carefully proved and tested, only one pass of the training data would help recover the parametric subspace, thus network could be easily trained layer-wise and expanded.\n\nSome novel contributions:\n1. Layer by layer feedforward training process, no back-prop.\n2. On-line settings to train parameters ( guaranteed convergence in a single pass of the data)\n\nWeakness :\n1. The assumption that a low-rank parameter space exists among tasks rather than original feature spaces is not new and widely used in literature.\n2. The proof part(Section 2.2) can be extended with more details in Appendix.\n3. In synthetic data experiments (Table1), only small margins could be observed between SN, f-MLP and rf-MLP, and only Layer 1 of SN performs better above all others. \n4. Typo: In Table2,3,5, Multi-l_{2,1} (denotes the L2,1 norm) were written wrong.\n5. In the synthetic data experiments on comparison with single-task and multi-task models, counter-intuitive results (with larger training data split, ANMSE raises instead of decreases) of multi-task models may need further explanation. \n6. Extra models like Deep Networks with/without matrix factorization could be added. ( As proposed model is a deep model, the lack of comparison with deep methods is dubious)\n7. In Section 4.2, the real dataset is rather small thus the results on this small dataset were not convincing enough. SN model outperforms the state-of-the-art with only small margin. Extensive experiments could be added.\n8. The performance on One-Layer Subspace Network (with only the input features) could be added. \n\nConclusion:\nThough with a quite novel idea on solving multi-task censored regression problem, the experiments conducted on synthetic data and real data are not convincing enough to ensure the contribution of the Subspace Network. \n", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1511839181292}, {"id": "SkwZAL4ef", "reviewer_signature": ["ICLR.cc/2018/Conference/Paper599/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes a multi task learning framework for the modeling of clinical data in neurodegenerative diseases. \nDifferently from previous applications of machine learning in neurodegeneration modeling, the proposed approach models the clinical data accounting for the bounded nature of cognitive tests scores. The framework is represented by a feed-forward deep architecture analogous to a residual network. At each layer a low-rank constraint is enforced on the linear transformation, while the cost function is specified in order to differentially account for the bounds of the predicted variables.\n\nThe idea of explicitly accounting for the boundedness of clinical scores is interesting, although the assumption of the proposed model is still incorrect: clinical scores are defined on discrete scales. For this reason the Gaussian assumption for the cost function used in the method is still not appropriate for the proposed application. \nFurthermore, while being the main methodological drive of this work, the paper does not show evidence about improved predictive performance and generalisation when accounting for the boundedness of the regression targets. \nThe proposed algorithm is also generally compared with respect to linear methods, and the authors could have provided a more rigorous benchmark including standard non-linear prediction approaches (e.g. random forests, NN, GP, …). \n\nOverall, the proposed methods seems to provide little added value to the large amount of predictive methods proposed so far for prediction in neurodegenerative disorders. Moreover, the proposed experimental paradigm appears flawed. What is the interest of predicting baseline (or 6 months at best) cognitive scores (relatively low-cost and part of any routine clinical assessment) from brain imaging data (high-cost and not routine)?\n\nOther remarks. \n\n- In section 2.2 and 4 there is some confusion between iteration indices and samples indices “i”. \n\n- Contrarily to what is stated in the introduction, the loss functions proposed in page 3 (first two formulas) only accounts for the lower bound of the predicted variables.  \n\n-  Figure 2, synthetic data. The scale of the improvement of the subspace difference is quite tiny, in the order of 1e-2 when compared to U, and of 1e-5 across iterations. The loss function of Figure 2.b also does not show a strong improvement across iterations, while indicating a rather large instability of the optimisation procedure. These aspects may be a sign of convergence issues. \n\n- The dimensionality of the subspace representation importantly depends on the choice of the rank R of U and V. This is a crucial parameters that is however not discussed nor analysed in the paper. \n\n- The synthetic example of page 7 is quite misleading and potentially biased towards the proposed model. The authors are generating the synthetic data according to the model, and it is thus not surprising that they managed to obtain the best performance.  In particular, due to the nonlinear nature of (1), all the competing linear models are expected to perform poorly in this kind of setting.\n\n- The computation time for the linear model shown in Table 3 is quite surprising (~20 minutes for linear regression of 5k observations). Is there anything that I am missing?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea which is however not clearly developed. Incremental results.", "rating": "4: Ok but not good enough - rejection", "review": "This work proposes a multi task learning framework for the modeling of clinical data in neurodegenerative diseases. \nDifferently from previous applications of machine learning in neurodegeneration modeling, the proposed approach models the clinical data accounting for the bounded nature of cognitive tests scores. The framework is represented by a feed-forward deep architecture analogous to a residual network. At each layer a low-rank constraint is enforced on the linear transformation, while the cost function is specified in order to differentially account for the bounds of the predicted variables.\n\nThe idea of explicitly accounting for the boundedness of clinical scores is interesting, although the assumption of the proposed model is still incorrect: clinical scores are defined on discrete scales. For this reason the Gaussian assumption for the cost function used in the method is still not appropriate for the proposed application. \nFurthermore, while being the main methodological drive of this work, the paper does not show evidence about improved predictive performance and generalisation when accounting for the boundedness of the regression targets. \nThe proposed algorithm is also generally compared with respect to linear methods, and the authors could have provided a more rigorous benchmark including standard non-linear prediction approaches (e.g. random forests, NN, GP, …). \n\nOverall, the proposed methods seems to provide little added value to the large amount of predictive methods proposed so far for prediction in neurodegenerative disorders. Moreover, the proposed experimental paradigm appears flawed. What is the interest of predicting baseline (or 6 months at best) cognitive scores (relatively low-cost and part of any routine clinical assessment) from brain imaging data (high-cost and not routine)?\n\nOther remarks. \n\n- In section 2.2 and 4 there is some confusion between iteration indices and samples indices “i”. \n\n- Contrarily to what is stated in the introduction, the loss functions proposed in page 3 (first two formulas) only accounts for the lower bound of the predicted variables.  \n\n-  Figure 2, synthetic data. The scale of the improvement of the subspace difference is quite tiny, in the order of 1e-2 when compared to U, and of 1e-5 across iterations. The loss function of Figure 2.b also does not show a strong improvement across iterations, while indicating a rather large instability of the optimisation procedure. These aspects may be a sign of convergence issues. \n\n- The dimensionality of the subspace representation importantly depends on the choice of the rank R of U and V. This is a crucial parameters that is however not discussed nor analysed in the paper. \n\n- The synthetic example of page 7 is quite misleading and potentially biased towards the proposed model. The authors are generating the synthetic data according to the model, and it is thus not surprising that they managed to obtain the best performance.  In particular, due to the nonlinear nature of (1), all the competing linear models are expected to perform poorly in this kind of setting.\n\n- The computation time for the linear model shown in Table 3 is quite surprising (~20 minutes for linear regression of 5k observations). Is there anything that I am missing?\n", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1511448063449}], "openreview_url": "https://openreview.net/forum?id=HytSvlWRZ", "arxiv_id": "1802.06516", "paper_pdf": "papers/HytSvlWRZ.pdf", "paper_pdf_sha256": "85c20e92e1ff241feb59b97c98effc1118cda2c3a2c830d079c37acb223a404d", "paper_pdf_bytes": 410074, "paper_pdf_source": "openreview", "code_url": "https://github.com/illidanlab/subspace-net", "code_repository": "illidanlab/subspace-net", "code_commit": "39c811836550c4762cd40ff9c3b32382a7f5a807", "code_archive": "repos/HytSvlWRZ.zip", "code_archive_sha256": "bd3a27ec9909349e32e15caf1b70776e9002c79c4fdad9a7dc6f3220405dcf46", "code_archive_bytes": 9319542, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 9102, "github_languages": {"Python": 33399}, "github_archived": false, "github_pushed_at": "2018-02-25T04:45:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/subspace-network-deep-multi-task-censored"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7lzq9mMVxq", "year": 2026, "status": "rejected", "title": "Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders", "authors": ["David Chanin", "Tomáš Dulka", "Adrià Garriga-Alonso"], "authorids": ["~David_Chanin1", "~Tomáš_Dulka1", "~Adrià_Garriga-Alonso1"], "authors_source": "OpenReview API", "abstract": "It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying \"true features\" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Importantly, our work shows that SAE width is not a neutral hyperparameter: narrower SAEs suffer more from hedging than wider SAEs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Q96nRlhoTz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14548/Reviewer_esmW"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "The paper studies LLM interpretability with sparse auto-encoders (SAEs). Previous works identify one phenomenon where one feature actives only when another does as the limiting factor for applying SAEs for interpretability. The authors introduce another limiting phenomenon which they call \"feature hedging\", that occurs when the SAE's latent space is too narrow and the underlying features are correlated. \n\nThey demonstrate this through some toy experiments on 2 and 4 dimensional settings and they define a metric to detect this effect in LLMs. Their experiments on SAEs trained on LLM activations, and report their results using the metric (which they call hedging degree) as a function of the parameters of SAE and average number of active latents. Finally they propose a re-weighted Matryoshka training to overcome the phenomenon they proposed.", "review_text": "The paper studies LLM interpretability with sparse auto-encoders (SAEs). Previous works identify one phenomenon where one feature actives only when another does as the limiting factor for applying SAEs for interpretability. The authors introduce another limiting phenomenon which they call \"feature hedging\", that occurs when the SAE's latent space is too narrow and the underlying features are correlated. \n\nThey demonstrate this through some toy experiments on 2 and 4 dimensional settings and they define a metric to detect this effect in LLMs. Their experiments on SAEs trained on LLM activations, and report their results using the metric (which they call hedging degree) as a function of the parameters of SAE and average number of active latents. Finally they propose a re-weighted Matryoshka training to overcome the phenomenon they proposed.", "strengths": "The idea of studying sparse autoencoders under the assumption of anti-correlated (or hierarchical) features is a good direction, the toy setup the authors propose is interesting and intuitive, and in the tests they show using LLMs, the hedging-degree metric appears to work.", "weaknesses": "While the direction of study is novel, the large scale experiments are limited in scope, and I am not fully convinced that the metric of hedging degree is the right one to measure this phenomenon. Is the $\\lVert .\\rVert$ an $\\ell_2$ or $\\ell_1$ norm? Further, what is $W_{rand}[0:N]$? are the latents initialized with a Uniform or a Gaussian distribution?\n\nCan the authors give more reasoning behind selecting this metric besides it exceeding a \"random\" baseline? It would help their case to see tests using this metric on the toy setups. \n\nOverall, the writing can be greatly improved to a more precise way of defining terms. As an example, terms like polysemantic and monsemantic activations without their definitions, making this paper sort of inaccessible to users outside the area of interpretability. \n\nOverall, I do not recommend accepting this paper.", "questions": "Asked above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies LLM interpretability with sparse auto-encoders (SAEs). Previous works identify one phenomenon where one feature actives only when another does as the limiting factor for applying SAEs for interpretability. The authors introduce another limiting phenomenon which they call \"feature hedging\", that occurs when the SAE's latent space is too narrow and the underlying features are correlated. \n\nThey demonstrate this through some toy experiments on 2 and 4 dimensional settings and they define a metric to detect this effect in LLMs. Their experiments on SAEs trained on LLM activations, and report their results using the metric (which they call hedging degree) as a function of the parameters of SAE and average number of active latents. Finally they propose a re-weighted Matryoshka training to overcome the phenomenon they proposed.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "The idea of studying sparse autoencoders under the assumption of anti-correlated (or hierarchical) features is a good direction, the toy setup the authors propose is interesting and intuitive, and in the tests they show using LLMs, the hedging-degree metric appears to work.", "weaknesses": "While the direction of study is novel, the large scale experiments are limited in scope, and I am not fully convinced that the metric of hedging degree is the right one to measure this phenomenon. Is the $\\lVert .\\rVert$ an $\\ell_2$ or $\\ell_1$ norm? Further, what is $W_{rand}[0:N]$? are the latents initialized with a Uniform or a Gaussian distribution?\n\nCan the authors give more reasoning behind selecting this metric besides it exceeding a \"random\" baseline? It would help their case to see tests using this metric on the toy setups. \n\nOverall, the writing can be greatly improved to a more precise way of defining terms. As an example, terms like polysemantic and monsemantic activations without their definitions, making this paper sort of inaccessible to users outside the area of interpretability. \n\nOverall, I do not recommend accepting this paper.", "questions": "Asked above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762477935919}, {"id": "zExbR2iwaZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14548/Reviewer_3uzz"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "The paper introduces the problem of \"feature hedging\" in Sparse AutoEncoders (SAEs), motivating it first in the context of \"toy\" models, then empirically studying it in the context of larger-scale SAEs/LLMs. The authors proceed to introduce a tentative solution to hedging by incorporating a balancing term in the Matryoshka SAE loss function, finding that doing so presents empirical benefits for \"Balance Matryoshka SAEs\" relative to unbalanced baselines.", "review_text": "The paper introduces the problem of \"feature hedging\" in Sparse AutoEncoders (SAEs), motivating it first in the context of \"toy\" models, then empirically studying it in the context of larger-scale SAEs/LLMs. The authors proceed to introduce a tentative solution to hedging by incorporating a balancing term in the Matryoshka SAE loss function, finding that doing so presents empirical benefits for \"Balance Matryoshka SAEs\" relative to unbalanced baselines.", "strengths": "- The authors motivate, define, empirically study, and propose tentative solutions to a new and potentially important problem (\"feature hedging\") with leading interpretability methods (SAEs). \n    - Both defining/demonstrating this problem, and making substantive empirical progress towards understanding and resolving it, represent significant contributions to the interpretability community.\n- The paper is generally well-written, with clear argumentation and strong intuition-building in both the introduction and \"toy\" experiments (sec 3).\n- Hedging experiments (sec 4) are comprehensive, considering a range of LLMs/layers, SAE architectures and hyperparameters, etc.\n    - And reporting several categories of SAEBench scores across hyperparameter values for the balancing multiplier (in sec 5) is also much appreciated for better understanding and contextualizing balancing with respect to standard (multiplier = 0) or unbalanced Matryoshka (multiplier = 1) SAEs.", "weaknesses": "The most substantial weakness is the lack of hedging results for Balance Matryoshka SAEs (elaborated below) -- given that the core motivation for the proposed balancing approach is to resolve the hedging problem, the absence of results showing whether balancing actually helps at all with this problem means that it cannot be taken seriously as a contribution. There are also more minor concerns regarding clarity on some important experimental details (discussed in the Questions section of the review) and citing prior interpretability work only within the narrow context of SAEs (elaborated below). Weaknesses by section are provided below.\n\nBackground/Related Work:\n- There is already substantial work on closely-related problems in the context of probing classifiers -- see, e.g., [1-5]. (Note that this is not a serious novelty concern, as hedging is an SAE-specific problem and the most relevant works studying SAEs are already cited and discussed. Instead, I believe this is simply another instance of the \"parallel community\" studying mechanistic interpretability largely focused on SAEs that often fails to cite work from the broader, longstanding interpretability/model analysis literature, as highlighted in [6].)\n\n[1] Kumar, A., Tan, C., & Sharma, A. (2022). Probing classifiers are unreliable for concept removal and detection. Advances in Neural Information Processing Systems, 35, 17994-18008.          \n[2] Ravichander, A., Belinkov, Y., & Hovy, E. (2021, April). Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (pp. 3363-3377).        \n[3] Elazar, Y., Ravfogel, S., Jacovi, A., & Goldberg, Y. (2021). Amnesic probing: Behavioral explanation with amnesic counterfactuals. Transactions of the Association for Computational Linguistics, 9, 160-175.            \n[4] Canby, M., Davies, A., Rastogi, C., & Hockenmaier, J. (2024). How Reliable are Causal Probing Interventions?. arXiv preprint arXiv:2408.15510.     \n[5] Hewitt, J., & Liang, P. (2019). Designing and interpreting probes with control tasks. arXiv preprint arXiv:1909.03368.           \n[6] Saphra, Naomi, and Sarah Wiegreffe. \"Mechanistic?.\" Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2024.\n\nSections 4-5:\n- In sec 5, it is claimed that \"As we saw in Section 4.1, the more narrow an SAE is, the worse the hedging. Matryoshka SAEs thus solve feature absorption at the expense of exacerbating feature hedging.\" However, the experiments in sec 4.1 do not explicitly consider Matryoshka SAEs, weakening the motivation for Balance Matryoshka SAEs in sec 5. I suggest repeating the experiments in sec 4.1 (or a suitable subset, given compute constraints) with Matryoshka SAEs to determine the extent to which feature hedging is actually an issue for baseline Matryoshka SAEs.\n\nSection 5:\n- **(MOST IMPORTANT:)** There is no measurement of hedging for \"real-world\" balance Matryoshka SAEs (i.e., those trained on actual LLMs rather than toy models), so it is impossible to tell how well balancing actually solves this problem (relative to baseline Matryoshka SAEs without balancing). The results from \"toy\" models are useful for illustrating the tradeoff between hedging and absorption, but insufficient to claim that Balance Matryoshka SAEs present any real utility for resolving hedging in the target context of LLMs. I suggest repeating the experiments from sec 4.1 (or a suitable subset, given compute constraints) with the Balance Matryoshka SAEs trained in sec 5.\n\nIf the authors are able to perform the experiments noted above (perhaps at a smaller scale if this is infeasible during the discussion period) -- and if results show that Balance Matryoshka SAEs (at some multiplier between 0 and 1) do actually, nontrivially improve on baseline traditional SAEs (multiplier = 0) and Matryoshka SAEs (multiplier = 1) -- then I would be happy to increase my score.", "questions": "Intro:\n- In the caption of fig 1b, the authors state that \"Asymmetry between encoder and decoder is characteristic of absorption\", which seems to be critical to the discussion in sec 4 (lines 291-296), and is also reiterated in tab 1. Can you elaborate on this point? Is this expected *a priori* (and in which case, what is the theoretical justification); or is it based on empirical findings in this or other work (in which case, can you cite the relevant work/section)?\n\nSec 3:\n- In the paragraph \"The implications of this for SAE performance are quite dire [...]\" -- intuitively, I would tend to agree with that this is a problem, but I think the argument could be better fleshed out. For instance, what would the consequences be for practical interpretability desiderata, such as OOD detection or safety monitoring? Why is negative feature mixing an issue if two features really are strongly negatively correlated? \n    - The simple answer to the second question is that such correlations may be spurious and lead to poor OOD robustness of SAE features; but for SAE training datasets that are i.i.d. wrt model training data, such issues might still be indicative of the underlying representation learned by models, and thus useful for detecting spurious/shortcut feature learning. I am curious to hear how the authors would respond to this argument, and the degree to which it presents a real challenge to the core motivation of this work.\n\nSec 4:\n- A central point supporting the experiments and interpretation throughout sec 4 is stated as: \"Based on our understanding of hedging in toy models, we expect that when a new latent is added to an SAE, this should 'pull out' the component of the new feature from existing SAE latents, where it was previously hedged. Thus if hedging occurs, the change in existing latents after a new latent is added should project onto that new latent.\" I have a few questions on this point: \n    - To my understanding, the evidence in sec 3 and app A.1-2 comes from experiments where the number of latents is fixed prior to training the model -- which is a completely different setting from the expectation stated above. So:\n        - Is my understanding here correct?\n        - And if so, then what is the stated expectation based on? (I.e., please articulate precisely what it is in your \"understanding of hedging in toy models\" that would lead to this expectation, and why.)\n    - Additionally, I feel that this argument would substantially benefit from a more rigorous formal presentation. For instance: (a) how could one mathematically state the expectation as a hypothesis to be tested, (b) how would one define a reasonable corresponding \"null hypothesis\" and measure the degree to which one is to be favored over the other, and (c) how could one formally state the evidentiary basis (per the previous questions) in relation to both hypotheses? (Note: I believe that (a) may already be covered by equation 7 and supporting text, but I would appreciate more clarification on precisely how it relates to the stated expectation/hypothesis; and I don't see anything corresponding to (b-c) in the main paper.)\n- Similar to the previous question, what would one expect for a \"null hypothesis\" value of hedging degree h? (A control condition here would be useful -- if this is nontrivial to define in the context of LLM SAEs, then at least showing what h looks like for \"toy\" experiments like those in sec 3 (ideally across multiple dictionary sizes, k/L0 values, etc, at values closer to those in sec 4) would be helpful.)\n- In fig 6,\n    - In legends of all 3 plots, am I correct in assuming that \"btk\" refers to BatchTopK and non-btk is \"l1\" (meaning L1-regularized SAEs, such as those in Olah et al., 2024)? This was confusing to figure out.\n    - What is being visualized for BatchTopK models on on the x-axis of fig 6c? To my understanding, the same k for BatchTopK is used for training all SAEs (with k=25, per appendix A.4) -- so how can there be multiple L0 values? Or are these separate models trained with different values of k?\n\nSec 5:\n- The experiment whose results are reprted in fig 7 needs some clarification -- e.g., even after consulting the appendix, the following are still unclear:\n    - What are the dictionary sizes $\\mathcal{M} = m_1, ..., m_n$? My best guess is that there are just two dictionaries, $m_1, m_2 = 1, 4$, but this needs to be clarified.\n    - $\\beta = 0.25$ shows the best results. Does this mean that $\\beta_{m_i} = 0.25$ for all $m_i$, or just $m_i : i > 1$?\n- What is the basis for the multiplier formula used to obtain $\\beta_m$ in lines 415-419? (I understand that the multiplier, such as 0.5, is a hyperparameter; but why set $\\beta_m = \\mu^{(n - m)}$ for multiplier $\\mu$ rather than, e.g., sampling linearly between 0 and 1?)\n\nMinor clarifications/nitpicks:\n- In fig 6, the caption states \"Hedging degree for SAEs trained on Gemma-2-2b layer 12\" -- but both gemma and llama are tested, and fig 6b is over multiple layers, correct? Please update the caption accordingly.\n- In fig 8, what is the purpose of setting the multiplier to a value greater than 1? I don't understand what this would correspond to theoretically; and empirically it seems that most variation occurs within [0, 1] (as one would intuitively expect).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the problem of \"feature hedging\" in Sparse AutoEncoders (SAEs), motivating it first in the context of \"toy\" models, then empirically studying it in the context of larger-scale SAEs/LLMs. The authors proceed to introduce a tentative solution to hedging by incorporating a balancing term in the Matryoshka SAE loss function, finding that doing so presents empirical benefits for \"Balance Matryoshka SAEs\" relative to unbalanced baselines.", "soundness": 2, "presentation": 3, "contribution": 4, "strengths": "- The authors motivate, define, empirically study, and propose tentative solutions to a new and potentially important problem (\"feature hedging\") with leading interpretability methods (SAEs). \n    - Both defining/demonstrating this problem, and making substantive empirical progress towards understanding and resolving it, represent significant contributions to the interpretability community.\n- The paper is generally well-written, with clear argumentation and strong intuition-building in both the introduction and \"toy\" experiments (sec 3).\n- Hedging experiments (sec 4) are comprehensive, considering a range of LLMs/layers, SAE architectures and hyperparameters, etc.\n    - And reporting several categories of SAEBench scores across hyperparameter values for the balancing multiplier (in sec 5) is also much appreciated for better understanding and contextualizing balancing with respect to standard (multiplier = 0) or unbalanced Matryoshka (multiplier = 1) SAEs.", "weaknesses": "The most substantial weakness is the lack of hedging results for Balance Matryoshka SAEs (elaborated below) -- given that the core motivation for the proposed balancing approach is to resolve the hedging problem, the absence of results showing whether balancing actually helps at all with this problem means that it cannot be taken seriously as a contribution. There are also more minor concerns regarding clarity on some important experimental details (discussed in the Questions section of the review) and citing prior interpretability work only within the narrow context of SAEs (elaborated below). Weaknesses by section are provided below.\n\nBackground/Related Work:\n- There is already substantial work on closely-related problems in the context of probing classifiers -- see, e.g., [1-5]. (Note that this is not a serious novelty concern, as hedging is an SAE-specific problem and the most relevant works studying SAEs are already cited and discussed. Instead, I believe this is simply another instance of the \"parallel community\" studying mechanistic interpretability largely focused on SAEs that often fails to cite work from the broader, longstanding interpretability/model analysis literature, as highlighted in [6].)\n\n[1] Kumar, A., Tan, C., & Sharma, A. (2022). Probing classifiers are unreliable for concept removal and detection. Advances in Neural Information Processing Systems, 35, 17994-18008.          \n[2] Ravichander, A., Belinkov, Y., & Hovy, E. (2021, April). Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume (pp. 3363-3377).        \n[3] Elazar, Y., Ravfogel, S., Jacovi, A., & Goldberg, Y. (2021). Amnesic probing: Behavioral explanation with amnesic counterfactuals. Transactions of the Association for Computational Linguistics, 9, 160-175.            \n[4] Canby, M., Davies, A., Rastogi, C., & Hockenmaier, J. (2024). How Reliable are Causal Probing Interventions?. arXiv preprint arXiv:2408.15510.     \n[5] Hewitt, J., & Liang, P. (2019). Designing and interpreting probes with control tasks. arXiv preprint arXiv:1909.03368.           \n[6] Saphra, Naomi, and Sarah Wiegreffe. \"Mechanistic?.\" Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2024.\n\nSections 4-5:\n- In sec 5, it is claimed that \"As we saw in Section 4.1, the more narrow an SAE is, the worse the hedging. Matryoshka SAEs thus solve feature absorption at the expense of exacerbating feature hedging.\" However, the experiments in sec 4.1 do not explicitly consider Matryoshka SAEs, weakening the motivation for Balance Matryoshka SAEs in sec 5. I suggest repeating the experiments in sec 4.1 (or a suitable subset, given compute constraints) with Matryoshka SAEs to determine the extent to which feature hedging is actually an issue for baseline Matryoshka SAEs.\n\nSection 5:\n- **(MOST IMPORTANT:)** There is no measurement of hedging for \"real-world\" balance Matryoshka SAEs (i.e., those trained on actual LLMs rather than toy models), so it is impossible to tell how well balancing actually solves this problem (relative to baseline Matryoshka SAEs without balancing). The results from \"toy\" models are useful for illustrating the tradeoff between hedging and absorption, but insufficient to claim that Balance Matryoshka SAEs present any real utility for resolving hedging in the target context of LLMs. I suggest repeating the experiments from sec 4.1 (or a suitable subset, given compute constraints) with the Balance Matryoshka SAEs trained in sec 5.\n\nIf the authors are able to perform the experiments noted above (perhaps at a smaller scale if this is infeasible during the discussion period) -- and if results show that Balance Matryoshka SAEs (at some multiplier between 0 and 1) do actually, nontrivially improve on baseline traditional SAEs (multiplier = 0) and Matryoshka SAEs (multiplier = 1) -- then I would be happy to increase my score.", "questions": "Intro:\n- In the caption of fig 1b, the authors state that \"Asymmetry between encoder and decoder is characteristic of absorption\", which seems to be critical to the discussion in sec 4 (lines 291-296), and is also reiterated in tab 1. Can you elaborate on this point? Is this expected *a priori* (and in which case, what is the theoretical justification); or is it based on empirical findings in this or other work (in which case, can you cite the relevant work/section)?\n\nSec 3:\n- In the paragraph \"The implications of this for SAE performance are quite dire [...]\" -- intuitively, I would tend to agree with that this is a problem, but I think the argument could be better fleshed out. For instance, what would the consequences be for practical interpretability desiderata, such as OOD detection or safety monitoring? Why is negative feature mixing an issue if two features really are strongly negatively correlated? \n    - The simple answer to the second question is that such correlations may be spurious and lead to poor OOD robustness of SAE features; but for SAE training datasets that are i.i.d. wrt model training data, such issues might still be indicative of the underlying representation learned by models, and thus useful for detecting spurious/shortcut feature learning. I am curious to hear how the authors would respond to this argument, and the degree to which it presents a real challenge to the core motivation of this work.\n\nSec 4:\n- A central point supporting the experiments and interpretation throughout sec 4 is stated as: \"Based on our understanding of hedging in toy models, we expect that when a new latent is added to an SAE, this should 'pull out' the component of the new feature from existing SAE latents, where it was previously hedged. Thus if hedging occurs, the change in existing latents after a new latent is added should project onto that new latent.\" I have a few questions on this point: \n    - To my understanding, the evidence in sec 3 and app A.1-2 comes from experiments where the number of latents is fixed prior to training the model -- which is a completely different setting from the expectation stated above. So:\n        - Is my understanding here correct?\n        - And if so, then what is the stated expectation based on? (I.e., please articulate precisely what it is in your \"understanding of hedging in toy models\" that would lead to this expectation, and why.)\n    - Additionally, I feel that this argument would substantially benefit from a more rigorous formal presentation. For instance: (a) how could one mathematically state the expectation as a hypothesis to be tested, (b) how would one define a reasonable corresponding \"null hypothesis\" and measure the degree to which one is to be favored over the other, and (c) how could one formally state the evidentiary basis (per the previous questions) in relation to both hypotheses? (Note: I believe that (a) may already be covered by equation 7 and supporting text, but I would appreciate more clarification on precisely how it relates to the stated expectation/hypothesis; and I don't see anything corresponding to (b-c) in the main paper.)\n- Similar to the previous question, what would one expect for a \"null hypothesis\" value of hedging degree h? (A control condition here would be useful -- if this is nontrivial to define in the context of LLM SAEs, then at least showing what h looks like for \"toy\" experiments like those in sec 3 (ideally across multiple dictionary sizes, k/L0 values, etc, at values closer to those in sec 4) would be helpful.)\n- In fig 6,\n    - In legends of all 3 plots, am I correct in assuming that \"btk\" refers to BatchTopK and non-btk is \"l1\" (meaning L1-regularized SAEs, such as those in Olah et al., 2024)? This was confusing to figure out.\n    - What is being visualized for BatchTopK models on on the x-axis of fig 6c? To my understanding, the same k for BatchTopK is used for training all SAEs (with k=25, per appendix A.4) -- so how can there be multiple L0 values? Or are these separate models trained with different values of k?\n\nSec 5:\n- The experiment whose results are reprted in fig 7 needs some clarification -- e.g., even after consulting the appendix, the following are still unclear:\n    - What are the dictionary sizes $\\mathcal{M} = m_1, ..., m_n$? My best guess is that there are just two dictionaries, $m_1, m_2 = 1, 4$, but this needs to be clarified.\n    - $\\beta = 0.25$ shows the best results. Does this mean that $\\beta_{m_i} = 0.25$ for all $m_i$, or just $m_i : i > 1$?\n- What is the basis for the multiplier formula used to obtain $\\beta_m$ in lines 415-419? (I understand that the multiplier, such as 0.5, is a hyperparameter; but why set $\\beta_m = \\mu^{(n - m)}$ for multiplier $\\mu$ rather than, e.g., sampling linearly between 0 and 1?)\n\nMinor clarifications/nitpicks:\n- In fig 6, the caption states \"Hedging degree for SAEs trained on Gemma-2-2b layer 12\" -- but both gemma and llama are tested, and fig 6b is over multiple layers, correct? Please update the caption accordingly.\n- In fig 8, what is the purpose of setting the multiplier to a value greater than 1? I don't understand what this would correspond to theoretically; and empirically it seems that most variation occurs within [0, 1] (as one would intuitively expect).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762294544858}, {"id": "srW6rqdHEW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14548/Reviewer_G273"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper describes the phenomenon of feature hedging, in some ways the opposite of the previously noted phenomena of feature absorption. Hedging is analyzed in four toy scenarios: independent, hierarchical, correlated, and anti-correlated features. Then, a metric, hedging degree, is proposed and used to measure hedging in SAEs for LLMs. Finally, the paper proposes balance matryoshka SAEs that scale the loss for each level to reduce hedging. This appears to be a simple approach that can help, although, as noted by the authors, not in all cases.\n\nI learn towards rejecting the paper, mainly because feature hedging is not formally defined. This makes it difficult to judge whether the hedging degree is a reasonable metric, and limits formal understanding of the phenomenon.\n\nFinally, I would like to disclose that I am not familiar with recent work on SAEs beyond the most well-known papers. There has been a surge of works in this area, so there could be closely related works to this paper that I am not aware of.", "review_text": "This paper describes the phenomenon of feature hedging, in some ways the opposite of the previously noted phenomena of feature absorption. Hedging is analyzed in four toy scenarios: independent, hierarchical, correlated, and anti-correlated features. Then, a metric, hedging degree, is proposed and used to measure hedging in SAEs for LLMs. Finally, the paper proposes balance matryoshka SAEs that scale the loss for each level to reduce hedging. This appears to be a simple approach that can help, although, as noted by the authors, not in all cases.\n\nI learn towards rejecting the paper, mainly because feature hedging is not formally defined. This makes it difficult to judge whether the hedging degree is a reasonable metric, and limits formal understanding of the phenomenon.\n\nFinally, I would like to disclose that I am not familiar with recent work on SAEs beyond the most well-known papers. There has been a surge of works in this area, so there could be closely related works to this paper that I am not aware of.", "strengths": "1. **Identifies a fundamental issue**. The paper identifies a fundamental issue in SAEs. I agree with the authors' conclusion in section 7 that understanding hedging is valuable for further work on SAEs.\n2. **Balance matryoshka SAE**. This generalization of matryoshka SAEs is appealing. Weighing the levels is a small, simple change, and we can retrieve both SAEs and matryoshka SAEs as special cases, and the paper demonstrates that it can reduce hedging.", "weaknesses": "1. **Theoretical analysis**. Although the simple toy examples are useful for understanding when feature hedging may occur, some theoretical analysis of the phenomenon would be a useful complement. Would it be possible to explain why these feature-hedging optima occur in simple toy settings?\n2. **Causes of hedging**. Throughout the paper, different causes for hedging are discussed. Some are related to the optimization problem itself (MSE, use of L1), others to the data (feature correlations), and even the learning process (in the paragraph on line 287). My impression is that the actual cause is not fully clear, and the discussion ends up slightly confusing. Again, some theoretical analysis could help clear things up.\n3. **Hedging degree**. It is hard to tell whether this metric measures hedging, especially as hedging has not been formally defined. Moreover, it seems like a difficult metric to use in practice, as it depends on a hyperparameter N and training the network (and thus the training setup, such as the optimizer and length of extended training). A.5 states that the hedging degree increases with N, but does not provide a convincing reason for choosing some N. If there was, e.g., a formal definition of hedging and hedging degree was derived as an approximation of it, I would find its use more convincing, but as it stands, it does not appear to be a reliable metric.", "questions": "1. **Why not use more latents?**. Perhaps a naive question: why should we not simply make the model wider and focus on reducing feature absorption? If SAEs are \"almost certainly narrower than the number of underlying features\" (line 72), it would be natural to first make them less narrow, which would reduce hedging?\n2. **$\\beta$ multipliers**. On line 418, the $\\beta$ parameters are parameterized relative to each other using a multiplier. How did you arrive at this particular form? That is: (ii) why is it decreasing, and (ii) why a constant multiplier and not some other function?\n\n**Minor comments**\n- Throughout the text, it would be good to disambiguate input features and latent features in the text so that they are not confused (as is done with $f$ and $l$ in mathematical notation).\n- Line 17, \"is caused by SAE reconstruction loss\" -> \"is caused by *the SAE's* reconstruction loss\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper describes the phenomenon of feature hedging, in some ways the opposite of the previously noted phenomena of feature absorption. Hedging is analyzed in four toy scenarios: independent, hierarchical, correlated, and anti-correlated features. Then, a metric, hedging degree, is proposed and used to measure hedging in SAEs for LLMs. Finally, the paper proposes balance matryoshka SAEs that scale the loss for each level to reduce hedging. This appears to be a simple approach that can help, although, as noted by the authors, not in all cases.\n\nI learn towards rejecting the paper, mainly because feature hedging is not formally defined. This makes it difficult to judge whether the hedging degree is a reasonable metric, and limits formal understanding of the phenomenon.\n\nFinally, I would like to disclose that I am not familiar with recent work on SAEs beyond the most well-known papers. There has been a surge of works in this area, so there could be closely related works to this paper that I am not aware of.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. **Identifies a fundamental issue**. The paper identifies a fundamental issue in SAEs. I agree with the authors' conclusion in section 7 that understanding hedging is valuable for further work on SAEs.\n2. **Balance matryoshka SAE**. This generalization of matryoshka SAEs is appealing. Weighing the levels is a small, simple change, and we can retrieve both SAEs and matryoshka SAEs as special cases, and the paper demonstrates that it can reduce hedging.", "weaknesses": "1. **Theoretical analysis**. Although the simple toy examples are useful for understanding when feature hedging may occur, some theoretical analysis of the phenomenon would be a useful complement. Would it be possible to explain why these feature-hedging optima occur in simple toy settings?\n2. **Causes of hedging**. Throughout the paper, different causes for hedging are discussed. Some are related to the optimization problem itself (MSE, use of L1), others to the data (feature correlations), and even the learning process (in the paragraph on line 287). My impression is that the actual cause is not fully clear, and the discussion ends up slightly confusing. Again, some theoretical analysis could help clear things up.\n3. **Hedging degree**. It is hard to tell whether this metric measures hedging, especially as hedging has not been formally defined. Moreover, it seems like a difficult metric to use in practice, as it depends on a hyperparameter N and training the network (and thus the training setup, such as the optimizer and length of extended training). A.5 states that the hedging degree increases with N, but does not provide a convincing reason for choosing some N. If there was, e.g., a formal definition of hedging and hedging degree was derived as an approximation of it, I would find its use more convincing, but as it stands, it does not appear to be a reliable metric.", "questions": "1. **Why not use more latents?**. Perhaps a naive question: why should we not simply make the model wider and focus on reducing feature absorption? If SAEs are \"almost certainly narrower than the number of underlying features\" (line 72), it would be natural to first make them less narrow, which would reduce hedging?\n2. **$\\beta$ multipliers**. On line 418, the $\\beta$ parameters are parameterized relative to each other using a multiplier. How did you arrive at this particular form? That is: (ii) why is it decreasing, and (ii) why a constant multiplier and not some other function?\n\n**Minor comments**\n- Throughout the text, it would be good to disambiguate input features and latent features in the text so that they are not confused (as is done with $f$ and $l$ in mathematical notation).\n- Line 17, \"is caused by SAE reconstruction loss\" -> \"is caused by *the SAE's* reconstruction loss\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761931058914}, {"id": "HGj1zlnSQs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14548/Reviewer_WZV8"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "In this work, the authors have defined and studied the feature hedging problem, theoretically in toy models and empirically in LLM SAEs. We show that hedging is worse the more narrow the SAE, and introduce a technique to characterise the amount of hedging present in a given SAE. They also studied hedging and absorption in matryoshka SAEs. They demonstrated that it is possible to improve the monosemanticity of matryoshka SAEs by adjusting the relative loss coefficients at each level of the matryoshka SAE better to balance the competing forces of absorption and hedging, while both issues remain present. It is also shown that the SAE width is not a neutral hyperparameter: narrow SAEs suffer more from hedging than wider SAEs.", "review_text": "In this work, the authors have defined and studied the feature hedging problem, theoretically in toy models and empirically in LLM SAEs. We show that hedging is worse the more narrow the SAE, and introduce a technique to characterise the amount of hedging present in a given SAE. They also studied hedging and absorption in matryoshka SAEs. They demonstrated that it is possible to improve the monosemanticity of matryoshka SAEs by adjusting the relative loss coefficients at each level of the matryoshka SAE better to balance the competing forces of absorption and hedging, while both issues remain present. It is also shown that the SAE width is not a neutral hyperparameter: narrow SAEs suffer more from hedging than wider SAEs.", "strengths": "Well written, good diagrams, the problem is clearly formulated\nThe methodology for the hedging degree is well explained\nFeature hedging may not be a fully novel idea, buta  good and useful idea for SAE", "weaknesses": "Formal proofs are limited; relies mostly on empirical evidence and illustrative derivations.\nSome assumptions—e.g., “parent latents are learned before child latents”—while intuitive, are not rigorously tested.\nThe broad idea that correlation can distort learned dictionaries has appeared in sparse coding literature so not so novel idea\nCharacterisation of Hedging is missing\n\\beta multiplier nonbelensing is not clear", "questions": "Give more rigorous theoretical proof\nexplain and reason out assumptions\nPotential interactions between reconstruction loss and latent geometry could be explored more formally.\nMore experimentation can be performed", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors have defined and studied the feature hedging problem, theoretically in toy models and empirically in LLM SAEs. We show that hedging is worse the more narrow the SAE, and introduce a technique to characterise the amount of hedging present in a given SAE. They also studied hedging and absorption in matryoshka SAEs. They demonstrated that it is possible to improve the monosemanticity of matryoshka SAEs by adjusting the relative loss coefficients at each level of the matryoshka SAE better to balance the competing forces of absorption and hedging, while both issues remain present. It is also shown that the SAE width is not a neutral hyperparameter: narrow SAEs suffer more from hedging than wider SAEs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Well written, good diagrams, the problem is clearly formulated\nThe methodology for the hedging degree is well explained\nFeature hedging may not be a fully novel idea, buta  good and useful idea for SAE", "weaknesses": "Formal proofs are limited; relies mostly on empirical evidence and illustrative derivations.\nSome assumptions—e.g., “parent latents are learned before child latents”—while intuitive, are not rigorously tested.\nThe broad idea that correlation can distort learned dictionaries has appeared in sparse coding literature so not so novel idea\nCharacterisation of Hedging is missing\n\\beta multiplier nonbelensing is not clear", "questions": "Give more rigorous theoretical proof\nexplain and reason out assumptions\nPotential interactions between reconstruction loss and latent geometry could be explored more formally.\nMore experimentation can be performed", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761907828528}], "openreview_url": "https://openreview.net/forum?id=7lzq9mMVxq", "arxiv_id": "2505.11756", "paper_pdf": "papers/7lzq9mMVxq.pdf", "paper_pdf_sha256": "8646991b763679c38c987cc748cd9ff22363e1e3b868d43039bf583054066e1b", "paper_pdf_bytes": 1400967, "paper_pdf_source": "openreview", "code_url": "https://github.com/chanind/feature-hedging-paper", "code_repository": "chanind/feature-hedging-paper", "code_commit": "6b6d7d64ce3f8871b8fec2e67557f6a5c7d00827", "code_archive": "repos/7lzq9mMVxq.zip", "code_archive_sha256": "b7ce0542f3dbab0051323fd259d8cd258cf9052a7356006b7eaa9ef6c1b676f1", "code_archive_bytes": 314993, "code_file_count": 69, "code_extensions": {".py": 65, ".ipynb": 4}, "github_disk_usage_kb": 332, "github_languages": {"Python": 340819, "Jupyter Notebook": 105284}, "github_archived": false, "github_pushed_at": "2025-11-26T15:15:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/feature-hedging-correlated-features-break"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hkjcdmz8Ro", "year": 2024, "status": "rejected", "title": "Jailbreaking Black Box Large Language Models in Twenty Queries", "authors": ["Patrick Chao", "Alexander Robey", "Edgar Dobriban", "Hamed Hassani", "George J. Pappas", "Eric Wong"], "authorids": ["~Patrick_Chao1", "~Alexander_Robey1", "~Edgar_Dobriban1", "~Hamed_Hassani2", "~George_J._Pappas1", "~Eric_Wong1"], "authors_source": "OpenReview API", "abstract": "There is growing research interest in ensuring that large language models align with human safety and ethical guidelines. Adversarial attacks known as 'jailbreaks' pose a significant threat as they coax models into overriding alignment safeguards. Identifying these vulnerabilities through attacking a language model (red teaming) is instrumental in understanding inherent weaknesses and preventing misuse. We present Prompt Automatic Iterative Refinement (PAIR), which generates semantic jailbreaks with only black-box access \nto a language model.\nEmpirically, PAIR often requires fewer than 20 queries, orders of magnitude fewer than prior jailbreak attacks. PAIR draws inspiration from the human process of social engineering, and employs an attacker language model to automatically generate adversarial prompts in place of a human. The attacker model uses the target model's response as additional context to iteratively refine the adversarial prompt. PAIR achieves competitive jailbreaking success rates and transferability on open and closed-source language models, including GPT-3.5/4, Vicuna, and PaLM.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "l6aXwkpyyS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7805/Reviewer_2SyA"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes an automated red teaming method that uses an attacker LLM to iteratively propose refinements to a jailbreak until it works. Experiments confirm that this method is able to jailbreak various open- and closed-source models, including ChatGPT. The method outperforms GCG at a much lower compute budget, and ablations to the system prompt used by the attacker LLM demonstrate the importance of different parts of the method.\n\n---\nUpdate after rebuttal: I will keep my score at a 5, although other reviewers could convince me to update.", "review_text": "This paper proposes an automated red teaming method that uses an attacker LLM to iteratively propose refinements to a jailbreak until it works. Experiments confirm that this method is able to jailbreak various open- and closed-source models, including ChatGPT. The method outperforms GCG at a much lower compute budget, and ablations to the system prompt used by the attacker LLM demonstrate the importance of different parts of the method.\n\n---\nUpdate after rebuttal: I will keep my score at a 5, although other reviewers could convince me to update.", "strengths": "- This is a sensible automated red teaming method, and it obtains comparable or better results than GCG while more closely mirroring human jailbreaks\n\n- Ablation studies that show the importance of a few different components of the method", "weaknesses": "- Ultimately, this is a fairly simple method, and there isn't much technical innovation. One could say that this paper is mainly about the system prompt in appendix B, and various experiments measuring its efficacy. The paper would benefit from more analysis into different strategies taken by the attacker LLM, whether these mirror what human red teamers try, whether smarter attacker LLMs work better (disentangled from how unfiltered they are), etc.\n\n- The only baseline is GCG. The baselines in Perez et al would be especially useful to include. E.g., the few-shot method in that work also uses iterative queries to an attacker LLM.\n\n- GPT-4 is used as a judge, which makes the results in the paper hard to compare to, since the underlying model in the GPT-4 API isn't guaranteed to always be available\n\n- In Section 3.3, the authors state, \"modifying the system prompt is a feature only available on open-source LLMs, limiting the available choices for A\". This isn't true at all. OpenAI introduced the notion of a system prompt, and their API allows editing system prompts. This also contradicts the experiments a few paragraphs later on GPT-3.5\n\n- Why aren't there experiments with GPT-4? It's odd that GPT-3.5 performs less well than Vicuna 1.5, since GPT-3.5 gets much higher scores on MMLU and is generally smarter. It would also be interesting to know if GPT-4 performs less well than Vicuna 1.5, and it would be an easy experiment to run.\n\n- In principle, the attacker LLM could violate the JSON format, since no methods for guaranteeing JSON formatting are used. How often does the attacker LLM violate the JSON format? How does your method handle cases where this happens?\n\n- In Figure 4, it's unclear whether the left plot is showing the effect of width or depth. The caption makes it seem like the x-axis is depth, but this doesn't make sense in context.", "questions": "See weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an automated red teaming method that uses an attacker LLM to iteratively propose refinements to a jailbreak until it works. Experiments confirm that this method is able to jailbreak various open- and closed-source models, including ChatGPT. The method outperforms GCG at a much lower compute budget, and ablations to the system prompt used by the attacker LLM demonstrate the importance of different parts of the method.\n\n---\nUpdate after rebuttal: I will keep my score at a 5, although other reviewers could convince me to update.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- This is a sensible automated red teaming method, and it obtains comparable or better results than GCG while more closely mirroring human jailbreaks\n\n- Ablation studies that show the importance of a few different components of the method", "weaknesses": "- Ultimately, this is a fairly simple method, and there isn't much technical innovation. One could say that this paper is mainly about the system prompt in appendix B, and various experiments measuring its efficacy. The paper would benefit from more analysis into different strategies taken by the attacker LLM, whether these mirror what human red teamers try, whether smarter attacker LLMs work better (disentangled from how unfiltered they are), etc.\n\n- The only baseline is GCG. The baselines in Perez et al would be especially useful to include. E.g., the few-shot method in that work also uses iterative queries to an attacker LLM.\n\n- GPT-4 is used as a judge, which makes the results in the paper hard to compare to, since the underlying model in the GPT-4 API isn't guaranteed to always be available\n\n- In Section 3.3, the authors state, \"modifying the system prompt is a feature only available on open-source LLMs, limiting the available choices for A\". This isn't true at all. OpenAI introduced the notion of a system prompt, and their API allows editing system prompts. This also contradicts the experiments a few paragraphs later on GPT-3.5\n\n- Why aren't there experiments with GPT-4? It's odd that GPT-3.5 performs less well than Vicuna 1.5, since GPT-3.5 gets much higher scores on MMLU and is generally smarter. It would also be interesting to know if GPT-4 performs less well than Vicuna 1.5, and it would be an easy experiment to run.\n\n- In principle, the attacker LLM could violate the JSON format, since no methods for guaranteeing JSON formatting are used. How often does the attacker LLM violate the JSON format? How does your method handle cases where this happens?\n\n- In Figure 4, it's unclear whether the left plot is showing the effect of width or depth. The caption makes it seem like the x-axis is depth, but this doesn't make sense in context.", "questions": "See weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699306387266}, {"id": "6PS88SG8ck", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7805/Reviewer_9YSw"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper studies the safety of LLMs. In particular, the authors proposed to utilize an LLM to craft jailbreaking prompts to break the safety of a target LLM.", "review_text": "This paper studies the safety of LLMs. In particular, the authors proposed to utilize an LLM to craft jailbreaking prompts to break the safety of a target LLM.", "strengths": "1. The safety of LLMs is an active research area. The proposed method could help to red team LLMs.\n\n2. The idea of utilizing LLMs to generate jailbreaking prompts is novel.", "weaknesses": "1. The technique contribution is week. The proposed method utilizes the LLM to refine the prompt. Thus, the performance of the proposed method heavily relies on the designed system prompt and LLMs. Moreover, the proposed method is based on heuristics, i.e., there is no insight for the proposed approach. But I understand those two points could be very challenging for LLM research.\n\n2. The evaluation is not systematic. For instance, only 50 questions are used in the evaluation. Thus, it is unclear whether the proposed approach is generalizable. More importantly, is the judge model the same for the proposed algorithm and evaluation? If this is the case, it is hard to see whether the reported results are reliable as LLMs could be inaccurate in their predictions. It would be better if other metrics could be used for cross-validation, e.g., manually check and the word list used by Zou et al. 2023. The proposed method is only compared with GCG. There are also many other baselines, e.g., handcrafted methods (https://www.jailbreakchat.com/). \n\n3. In GCG, authors showed that their approach could be transferred to other LLMs. Thus, GCG could craft adversarial prompts and transfer them to other LLMs. It would be good if such a comparison could be included. \n\nA minor point: The jailbreaking percentage is low for certain LLMs.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the safety of LLMs. In particular, the authors proposed to utilize an LLM to craft jailbreaking prompts to break the safety of a target LLM.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The safety of LLMs is an active research area. The proposed method could help to red team LLMs.\n\n2. The idea of utilizing LLMs to generate jailbreaking prompts is novel.", "weaknesses": "1. The technique contribution is week. The proposed method utilizes the LLM to refine the prompt. Thus, the performance of the proposed method heavily relies on the designed system prompt and LLMs. Moreover, the proposed method is based on heuristics, i.e., there is no insight for the proposed approach. But I understand those two points could be very challenging for LLM research.\n\n2. The evaluation is not systematic. For instance, only 50 questions are used in the evaluation. Thus, it is unclear whether the proposed approach is generalizable. More importantly, is the judge model the same for the proposed algorithm and evaluation? If this is the case, it is hard to see whether the reported results are reliable as LLMs could be inaccurate in their predictions. It would be better if other metrics could be used for cross-validation, e.g., manually check and the word list used by Zou et al. 2023. The proposed method is only compared with GCG. There are also many other baselines, e.g., handcrafted methods (https://www.jailbreakchat.com/). \n\n3. In GCG, authors showed that their approach could be transferred to other LLMs. Thus, GCG could craft adversarial prompts and transfer them to other LLMs. It would be good if such a comparison could be included. \n\nA minor point: The jailbreaking percentage is low for certain LLMs.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698801319961}, {"id": "uCNZepZ4cE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7805/Reviewer_HLa7"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "A novel and query-efficient LLM-agent-based red-teaming tool for testing jailbreaking behaviors of black-box LLMs. The core invention includes an attacker (an LLM) that reads the response from a target model (another LLM), and it uses the response and feedback to improve the jailbreak prompts. Empirical results show comparable (or better) jailbreak performance compared to the GCG method.", "review_text": "A novel and query-efficient LLM-agent-based red-teaming tool for testing jailbreaking behaviors of black-box LLMs. The core invention includes an attacker (an LLM) that reads the response from a target model (another LLM), and it uses the response and feedback to improve the jailbreak prompts. Empirical results show comparable (or better) jailbreak performance compared to the GCG method.", "strengths": "1. The proposed method is novel and leverages the chain of thoughts and in-context learning capability of LLMs for red-teaming.\n2. The ability to perform query-efficient red-teaming for black-box LLMs is important\n3. The jailbreak results (both direct queries and transfer) on black-box LLMs (GPT-3.5, GPT-4, Claude-1, Claude-2, PaLM-2) are quite remarkable.", "weaknesses": "While I enjoyed reading the paper and found the proposed method quite neat and novel, I have several major concerns that prevented me from recommending acceptance in the current form. I look forward to the authors' rebuttal to clarify my concerns.\n\n1. In the evaluation, it is stated that \"For each of these target models, we use a temperature of zero for deterministic generation\". I do not find this setting convincing, as this is not the default setting for LLMs. Moreover, a recent study <https://arxiv.org/abs/2310.06987> actually shows that merely changing these generation hyperparameters from the default value can weaken the safety alignment, without further optimization or tweaks. I would like to see the results on the default generation parameters (admittedly there will be randomness, and more runs are needed to make the result statistically meaningful), and would like to know why considering the unusual setting of making the temperature =0  is of practical importance.\n\n2. If I understand the evaluation setting correctly, I do not think the authors did justice to the GCG method in terms of query analysis. \n- First of all, the proposed method is prompt-specific, which means the total number of queries actually grows linearly with the number of test prompts. On the other hand, GCG is prompt-agnostic, which finds a \"universal\" suffix for every test prompt. When a new test prompt comes in, GCG does not require any query, while the proposed method does. If there is a large number of test prompts to be evaluated, perhaps GCG can be more query-efficient.\n- The authors mentioned they used 50 queries from AdvBench for evaluation, while \"when the maximum number of iterations is reached. For GCG, we use the default implementation and parameters, which uses a batch size of 512 and 500 steps for a total of 256,000 queries.\" This is confusing to me. Why shouldn't we just run GCG on the same 50 queries for a fair comparison? I would like to see a jailbreak success rate v.s. query count plot comparing GCG and the proposed method in this case.\n\n3. The related work only covers recent works on jailbreaking LLMs. However, given the similarity in adversarial testing/red-teaming, I suggest the authors include the discussion on query-based black-box attacks as related work, especially in the discussion of query efficiency. Of course, due to the special text interface, attackers may not need to explicitly estimate gradient information for updating the attack vectors, as opposed to standard ML models.", "questions": "- Referring to Weakness #1, please justify the selection of an unusual generating hyperparameter setting. Why didn't the authors try jailbreak attacks using the default settings (in which the alignment is mostly effective)\n- Referring to Weakness #2, is the query analysis fair to GCG? Given a small set of test prompts, perhaps the difference in query efficiency can be minor. If the set of test prompts is large, the growing total query complexity of the proposed method versus GCG can be an issue.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "A novel and query-efficient LLM-agent-based red-teaming tool for testing jailbreaking behaviors of black-box LLMs. The core invention includes an attacker (an LLM) that reads the response from a target model (another LLM), and it uses the response and feedback to improve the jailbreak prompts. Empirical results show comparable (or better) jailbreak performance compared to the GCG method.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The proposed method is novel and leverages the chain of thoughts and in-context learning capability of LLMs for red-teaming.\n2. The ability to perform query-efficient red-teaming for black-box LLMs is important\n3. The jailbreak results (both direct queries and transfer) on black-box LLMs (GPT-3.5, GPT-4, Claude-1, Claude-2, PaLM-2) are quite remarkable.", "weaknesses": "While I enjoyed reading the paper and found the proposed method quite neat and novel, I have several major concerns that prevented me from recommending acceptance in the current form. I look forward to the authors' rebuttal to clarify my concerns.\n\n1. In the evaluation, it is stated that \"For each of these target models, we use a temperature of zero for deterministic generation\". I do not find this setting convincing, as this is not the default setting for LLMs. Moreover, a recent study <https://arxiv.org/abs/2310.06987> actually shows that merely changing these generation hyperparameters from the default value can weaken the safety alignment, without further optimization or tweaks. I would like to see the results on the default generation parameters (admittedly there will be randomness, and more runs are needed to make the result statistically meaningful), and would like to know why considering the unusual setting of making the temperature =0  is of practical importance.\n\n2. If I understand the evaluation setting correctly, I do not think the authors did justice to the GCG method in terms of query analysis. \n- First of all, the proposed method is prompt-specific, which means the total number of queries actually grows linearly with the number of test prompts. On the other hand, GCG is prompt-agnostic, which finds a \"universal\" suffix for every test prompt. When a new test prompt comes in, GCG does not require any query, while the proposed method does. If there is a large number of test prompts to be evaluated, perhaps GCG can be more query-efficient.\n- The authors mentioned they used 50 queries from AdvBench for evaluation, while \"when the maximum number of iterations is reached. For GCG, we use the default implementation and parameters, which uses a batch size of 512 and 500 steps for a total of 256,000 queries.\" This is confusing to me. Why shouldn't we just run GCG on the same 50 queries for a fair comparison? I would like to see a jailbreak success rate v.s. query count plot comparing GCG and the proposed method in this case.\n\n3. The related work only covers recent works on jailbreaking LLMs. However, given the similarity in adversarial testing/red-teaming, I suggest the authors include the discussion on query-based black-box attacks as related work, especially in the discussion of query efficiency. Of course, due to the special text interface, attackers may not need to explicitly estimate gradient information for updating the attack vectors, as opposed to standard ML models.", "questions": "- Referring to Weakness #1, please justify the selection of an unusual generating hyperparameter setting. Why didn't the authors try jailbreak attacks using the default settings (in which the alignment is mostly effective)\n- Referring to Weakness #2, is the query analysis fair to GCG? Given a small set of test prompts, perhaps the difference in query efficiency can be minor. If the set of test prompts is large, the growing total query complexity of the proposed method versus GCG can be an issue.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698701067986}, {"id": "dxi9xoZx4s", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7805/Reviewer_PADE"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a new method named Prompt Automatic Iterative Refinement (PAIR), which can perform black box jailbreak attack to large language models. Experiments and results have shown that PAIR can successfully attack GPT 3.5 and GPT 4 less than twenty queries. Such threats raise the further research in better safety alignment for large language models.", "review_text": "This paper proposes a new method named Prompt Automatic Iterative Refinement (PAIR), which can perform black box jailbreak attack to large language models. Experiments and results have shown that PAIR can successfully attack GPT 3.5 and GPT 4 less than twenty queries. Such threats raise the further research in better safety alignment for large language models.", "strengths": "1. This paper gives good presentations about their method and details about the prompt design process.\n2. Compared to previous jailbreak method GCG, jailbreak prompts generated by PAIR can maintain semantic meanings and successfully attack black box models in only a few quires.\n3. The authors provide detailed ablation study about their system prompt design.\n4. The design of the attack method uses the idea of agent interaction. PAIR use multiple language models to revise the prompt and attack the black box models.", "weaknesses": "1. Not convincing results in Table 1. PAIR can only successfully attack parts of the models presented in your papers. This method is almost useless toward both open source Llama2 and close source Claude. Such inconsistency of results among different models makes the results and methodology less convincing to me.\n2. Too naive loss function. The Jailbreak Scoring part is too naive as an optimization method. The score is just a True or False score. I don't think such naive loss function design can be helpful for performing successfully attack, which makes the entire method seems like random searching. That maybe the reason why PAIR cannot jailbreak models like Llama2 or Claude.\n3. The necessary usage of complex system design in Table 7. Though authors have implemented ablation studies toward the usage of example and 'improvement' in attacker system prompt design, I still have concerns about the necessity of such a long and complex system prompt designing.", "questions": "What are the versions of the close source models used in your experiments. As far as I know, many close source models like ChatGPT would update their weights regularly. For better reproducibility, can you show in detail the specific version of the close source models. Some versions of GPT3.5 and GPT4 do not have safety alignment while others have. You may get totally different results with different versions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new method named Prompt Automatic Iterative Refinement (PAIR), which can perform black box jailbreak attack to large language models. Experiments and results have shown that PAIR can successfully attack GPT 3.5 and GPT 4 less than twenty queries. Such threats raise the further research in better safety alignment for large language models.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper gives good presentations about their method and details about the prompt design process.\n2. Compared to previous jailbreak method GCG, jailbreak prompts generated by PAIR can maintain semantic meanings and successfully attack black box models in only a few quires.\n3. The authors provide detailed ablation study about their system prompt design.\n4. The design of the attack method uses the idea of agent interaction. PAIR use multiple language models to revise the prompt and attack the black box models.", "weaknesses": "1. Not convincing results in Table 1. PAIR can only successfully attack parts of the models presented in your papers. This method is almost useless toward both open source Llama2 and close source Claude. Such inconsistency of results among different models makes the results and methodology less convincing to me.\n2. Too naive loss function. The Jailbreak Scoring part is too naive as an optimization method. The score is just a True or False score. I don't think such naive loss function design can be helpful for performing successfully attack, which makes the entire method seems like random searching. That maybe the reason why PAIR cannot jailbreak models like Llama2 or Claude.\n3. The necessary usage of complex system design in Table 7. Though authors have implemented ablation studies toward the usage of example and 'improvement' in attacker system prompt design, I still have concerns about the necessity of such a long and complex system prompt designing.", "questions": "What are the versions of the close source models used in your experiments. As far as I know, many close source models like ChatGPT would update their weights regularly. For better reproducibility, can you show in detail the specific version of the close source models. Some versions of GPT3.5 and GPT4 do not have safety alignment while others have. You may get totally different results with different versions.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics concerns.", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698514894084}], "openreview_url": "https://openreview.net/forum?id=hkjcdmz8Ro", "arxiv_id": "2310.08419", "paper_pdf": "papers/hkjcdmz8Ro.pdf", "paper_pdf_sha256": "8b9f505b06727faf677f4fbcb47886d2ec0d3b1dd614d2879b12cca15fdd5912", "paper_pdf_bytes": 2863510, "paper_pdf_source": "openreview", "code_url": "https://github.com/patrickrchao/JailbreakingLLMs", "code_repository": "patrickrchao/JailbreakingLLMs", "code_commit": "6379ef705a0fc745530f7d895963510c021b496a", "code_archive": "repos/hkjcdmz8Ro.zip", "code_archive_sha256": "75e7f0351db06f8b142a9e2315030330873571ffffd7d6a40d3603c0dd9c5e4c", "code_archive_bytes": 29111, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 41, "github_languages": {"Python": 73751, "Dockerfile": 362}, "github_archived": false, "github_pushed_at": "2025-07-02T00:03:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/jailbreaking-black-box-large-language-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ueEMZjY9WiM", "year": 2023, "status": "rejected", "title": "Compression-aware Training of Neural Networks using Frank-Wolfe", "authors": ["Max Zimmer", "Christoph Spiegel", "Sebastian Pokutta"], "authorids": ["~Max_Zimmer1", "~Christoph_Spiegel1", "~Sebastian_Pokutta1"], "authors_source": "OpenReview API", "abstract": "Many existing Neural Network pruning approaches either rely on retraining to compensate for pruning-caused performance degradation or they induce strong biases to converge to a specific sparse solution throughout training. A third paradigm, ’compression-aware’ training, obtains state-of-the-art dense models which are robust to a wide range of compression ratios using a single dense training run while also avoiding retraining. In that vein, we propose a constrained optimization framework centered around a versatile family of norm constraints and the Stochastic Frank-Wolfe (SFW) algorithm which together encourage convergence to well-performing solutions while inducing robustness towards convolutional filter pruning and low-rank matrix decomposition. Comparing our novel approaches to compression methods in these domains on benchmark image-classification architectures and datasets, we find that our proposed scheme is able to yield competitive results, often outperforming existing compression-aware approaches. In the case of low-rank matrix decomposition, our approach can require much less computational resources than nuclear-norm regularization based approaches by requiring only a fraction of the singular values in each iteration. As a special case, our proposed constraints can be extended to include the unstructured sparsity-inducing constraint proposed constraint by Pokutta et al. (2020) and Miao et al. (2022), which we improve upon. Our findings also indicate that the robustness of SFW-trained models largely depends on the gradient rescaling of the learning rate and we establish a theoretical foundation for that practice.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "mByKxvZi00", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6319/Reviewer_RYk5"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a novel framework for compression-aware training of neural networks. The proposed method uses norm constraints, for two types of pruning (1) convolutional filter pruning (2) low-rank matrix decomposition, expressed via updates of the Stochastic Frank-Wolfe (SFW) algorithm efficiently. ", "review_text": "Overall, I think this is a good paper. The ideas presented are novel and are well-supported by ample experimental evidence and some theoretical results. The proposed method and discussions are relevant and useful to the community. ", "strengths": "Pros: \n- The proposed framework is interesting and beneficial to the community. The authors provide sufficient intuition and motivation behind the use of the sparsity-inducing norm constraints and how they can be effectively realized via SFW. The presentation is clear and the math is easy to follow. \n- All the claims are supported well by empirical studies on benchmark datasets. The baselines seem sensible; although I must admit that I'm not too familiar with the related works in the compression-aware setting\n\nComments on the robustness study:\nOne of the interesting sections in the paper is the study on the robustness of the pruned model. The experimental study and the authors' discussion on the benefits of using the rescaled learning rate are insightful, especially at higher compression rates. With gradient scaling, the authors are able to show the 1/\\sqrt(T) convergence result (FW gap) for SFW. However, they still collectively don't provide enough convincing arguments for the robustness claims, in my opinion.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a novel framework for compression-aware training of neural networks. The proposed method uses norm constraints, for two types of pruning (1) convolutional filter pruning (2) low-rank matrix decomposition, expressed via updates of the Stochastic Frank-Wolfe (SFW) algorithm efficiently. ", "strength_and_weaknesses": "Pros: \n- The proposed framework is interesting and beneficial to the community. The authors provide sufficient intuition and motivation behind the use of the sparsity-inducing norm constraints and how they can be effectively realized via SFW. The presentation is clear and the math is easy to follow. \n- All the claims are supported well by empirical studies on benchmark datasets. The baselines seem sensible; although I must admit that I'm not too familiar with the related works in the compression-aware setting\n\nComments on the robustness study:\nOne of the interesting sections in the paper is the study on the robustness of the pruned model. The experimental study and the authors' discussion on the benefits of using the rescaled learning rate are insightful, especially at higher compression rates. With gradient scaling, the authors are able to show the 1/\\sqrt(T) convergence result (FW gap) for SFW. However, they still collectively don't provide enough convincing arguments for the robustness claims, in my opinion.", "clarity,_quality,_novelty_and_reproducibility": "The writing is clear and the ideas are easy to follow. The authors have agreed to release the code if accepted. ", "summary_of_the_review": "Overall, I think this is a good paper. The ideas presented are novel and are well-supported by ample experimental evidence and some theoretical results. The proposed method and discussions are relevant and useful to the community. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666854003118}, {"id": "kkpY9s6ENoY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6319/Reviewer_gmRx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a constrained optimization framework based on a versatile family of norm constraints and the stochastic FrankWolfe (SFW) algorithm. The proposed method apply on benchmark image-classification architectures and datasets, and it yields competitive results, often outperforming existing compression-aware approaches. ", "review_text": "The proposed method applied on benchmark image-classification architectures and\ndatasets outperform the existing compression-aware approaches.  In the case of low-rank\nmatrix decomposition, the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration. \n\nSome Questions:\n\n1. Why the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration ? It would be great if the authors would detail it.\n\n2. In the experimental results, we hope to see the efficiency of the proposed methods. \nPlease give the results on test accuracy vs time.\n\n3. There will be a strict upper limit of 9 pages for the main text of the submission, \nwith unlimited additional pages for citations (https://iclr.cc/Conferences/2023/CallForPapers). \nThis paper maybe not fit for this request.\n\n\n\n\n--------------------------------------------------------------------------------------------------------------------\n---------------------------------------------------------------------------------------------------------------------\nThe authors still did not solve my main concern- the limited novelty of this paper. \nSo I support to reject this paper.\n", "strengths": "Strength:\n\nThe proposed method applied on benchmark image-classification architectures and\ndatasets outperform the existing compression-aware approaches.  In the case of low-rank\nmatrix decomposition, the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration.\n\nWeakness:\n\nThe novelty of this paper is very limited. This paper basically combines \nthe existing k-support norm regularztion (https://arxiv.org/pdf/1204.5043.pdf, \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7952587) and the existing stochastic \nFrank-Wolfe methods (https://arxiv.org/pdf/1607.08254.pdf). ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed a constrained optimization framework based on a versatile family of norm constraints and the stochastic FrankWolfe (SFW) algorithm. The proposed method apply on benchmark image-classification architectures and datasets, and it yields competitive results, often outperforming existing compression-aware approaches. ", "strength_and_weaknesses": "Strength:\n\nThe proposed method applied on benchmark image-classification architectures and\ndatasets outperform the existing compression-aware approaches.  In the case of low-rank\nmatrix decomposition, the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration.\n\nWeakness:\n\nThe novelty of this paper is very limited. This paper basically combines \nthe existing k-support norm regularztion (https://arxiv.org/pdf/1204.5043.pdf, \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7952587) and the existing stochastic \nFrank-Wolfe methods (https://arxiv.org/pdf/1607.08254.pdf). ", "clarity,_quality,_novelty_and_reproducibility": "The writing can significantly be improved. \nThe novelty of this paper is very limited. This paper basically combines \nthe existing k-support norm regularztion (https://arxiv.org/pdf/1204.5043.pdf, \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7952587) and the existing stochastic \nFrank-Wolfe methods (https://arxiv.org/pdf/1607.08254.pdf). ", "summary_of_the_review": "The proposed method applied on benchmark image-classification architectures and\ndatasets outperform the existing compression-aware approaches.  In the case of low-rank\nmatrix decomposition, the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration. \n\nSome Questions:\n\n1. Why the proposed method can require much less computational resources\nthan nuclear-norm regularization based approaches by requiring only a fraction of\nthe singular values in each iteration ? It would be great if the authors would detail it.\n\n2. In the experimental results, we hope to see the efficiency of the proposed methods. \nPlease give the results on test accuracy vs time.\n\n3. There will be a strict upper limit of 9 pages for the main text of the submission, \nwith unlimited additional pages for citations (https://iclr.cc/Conferences/2023/CallForPapers). \nThis paper maybe not fit for this request.\n\n\n\n\n--------------------------------------------------------------------------------------------------------------------\n---------------------------------------------------------------------------------------------------------------------\nThe authors still did not solve my main concern- the limited novelty of this paper. \nSo I support to reject this paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666656422024}, {"id": "kJ178g9l2A", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6319/Reviewer_XfiW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the authors adopt SWF (stochastic Frank-Wolfe) algorithm to perform compression-aware training. Prelimary experiments demonstrate the efficacy of the proposed approach. ", "review_text": "See the comments above.", "strengths": "The authors adopt a classic optimization algorithm (SFW) to perform structure compression-aware training. Before this work is accepted, several concerns should be addressed:\n(i)\tThe first concern is about the novelty. This work heavily depends on the unstructed compression-aware training by extending the unstructured pruning to structured pruning setting. The novelty is discounted. \n(ii)\tThe authors claims that the adopted SFW methods achieve significant speedup compated with nuclear-norm regularization based approach. In general, for nuclear norm regularized optimization, several SVD-free methods have been proposed [1]. We require the authors compare their proposed approach with more advanced optimization methods for solving nuclear norm regularized problem to show its effecgtiveness for a fair comparision. Furthermore, the real speedup ration should be reported to show the efficacy of the proposed algorithm. \n(iii)\tThe current experiments restrict to CIFAR-10, CIFAR-100, Tiny-ImageNet. In general, ImageNet-1K is a standard benchmark to verify the effiectiveness of the compression methods. We recommend the author provide more experents on ImageNet-1k. \n(iv)\tAbout Theorem 5.1. Why does the term E[g*||Grad{L}||] indicate the convergence of the proposed approach? THe authors should gives more discussions on this convergence criteria. \n\n\n[1] SVD-free Convex-Concave Approaches for Nuclear Norm Regularization, IJCAI, 2017.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this work, the authors adopt SWF (stochastic Frank-Wolfe) algorithm to perform compression-aware training. Prelimary experiments demonstrate the efficacy of the proposed approach. ", "strength_and_weaknesses": "The authors adopt a classic optimization algorithm (SFW) to perform structure compression-aware training. Before this work is accepted, several concerns should be addressed:\n(i)\tThe first concern is about the novelty. This work heavily depends on the unstructed compression-aware training by extending the unstructured pruning to structured pruning setting. The novelty is discounted. \n(ii)\tThe authors claims that the adopted SFW methods achieve significant speedup compated with nuclear-norm regularization based approach. In general, for nuclear norm regularized optimization, several SVD-free methods have been proposed [1]. We require the authors compare their proposed approach with more advanced optimization methods for solving nuclear norm regularized problem to show its effecgtiveness for a fair comparision. Furthermore, the real speedup ration should be reported to show the efficacy of the proposed algorithm. \n(iii)\tThe current experiments restrict to CIFAR-10, CIFAR-100, Tiny-ImageNet. In general, ImageNet-1K is a standard benchmark to verify the effiectiveness of the compression methods. We recommend the author provide more experents on ImageNet-1k. \n(iv)\tAbout Theorem 5.1. Why does the term E[g*||Grad{L}||] indicate the convergence of the proposed approach? THe authors should gives more discussions on this convergence criteria. \n\n\n[1] SVD-free Convex-Concave Approaches for Nuclear Norm Regularization, IJCAI, 2017.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is well writen", "summary_of_the_review": "See the comments above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666528499203}, {"id": "0VyfhWKZab2", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6319/Reviewer_v7Db"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work proposes the SFW algorithm to solve a constrained optimization framework. The proposed method can result in well-performing models that are robust towards convolutional filter pruning as well as low-rank matrix decomposition. Experiment results also show that the proposed method has better “accuracy vs sparsity” performance than existing approaches. ", "review_text": "The experiment result is convincing. I also like the idea of combing the SFW and the group-k-support norm constraint together, but I didn’t find much novelty in the theory and algorithm respects.", "strengths": "Strength:\nOverall, the paper is well-organized and easy to follow. This paper studies two optimization problems: convolutional filter pruning and low-rank matrix decomposition. For the first problem, the author uses the group-k-support norm ball to constrain the optimization problem to compress the model instead of the existing k-sparse approach. To solve this problem, the author proposes to use the SFW algorithm. For the second problem, the spectral-k-support norm is used. Experiments results on those two problems show the advantage of the proposed method. Specifically, the accuracy of the proposed method is more robust than the existing methods when compressing the model.\n\nThis work also empirically shows that the robustness of SFW can largely be attributed to the usage of the gradient rescaling of the learning rate. To justify the usage of gradient rescaling theoretically, the convergence of SFW with batch gradient dependent step size in the non-convex setting is established.\n\nWeaknesses:\n1.  In the abstract, the author said that compression-aware training could obtain state-of-the-art dense models which are robust to a wide range of compression ratios using a single dense training run while also avoiding retraining. But I believe the proposed framework still needs to run several times independently to obtain models with different compression ratios. That is, to obtain a model with different compression ratios, you have to run SFW another time from the beginning. Therefore, I am not sure why this method can avoid retraining.\n  \n2. I notice that In Figure 2, the performance of ABFP is comparable to or even better than SFW. Could the author show the advantage of SFW compared to ABFP?\n\n3. Although Theorem 3.1 shows the convergence of SFW with batch gradient dependent step size in the non-convex setting.  However, the assumption of the objective function seems to be too strong (L-smooth and L-Lipschitz). I believe neither of the two studied problems would satisfy these assumptions. Please correct me if I am wrong.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes the SFW algorithm to solve a constrained optimization framework. The proposed method can result in well-performing models that are robust towards convolutional filter pruning as well as low-rank matrix decomposition. Experiment results also show that the proposed method has better “accuracy vs sparsity” performance than existing approaches. ", "strength_and_weaknesses": "Strength:\nOverall, the paper is well-organized and easy to follow. This paper studies two optimization problems: convolutional filter pruning and low-rank matrix decomposition. For the first problem, the author uses the group-k-support norm ball to constrain the optimization problem to compress the model instead of the existing k-sparse approach. To solve this problem, the author proposes to use the SFW algorithm. For the second problem, the spectral-k-support norm is used. Experiments results on those two problems show the advantage of the proposed method. Specifically, the accuracy of the proposed method is more robust than the existing methods when compressing the model.\n\nThis work also empirically shows that the robustness of SFW can largely be attributed to the usage of the gradient rescaling of the learning rate. To justify the usage of gradient rescaling theoretically, the convergence of SFW with batch gradient dependent step size in the non-convex setting is established.\n\nWeaknesses:\n1.  In the abstract, the author said that compression-aware training could obtain state-of-the-art dense models which are robust to a wide range of compression ratios using a single dense training run while also avoiding retraining. But I believe the proposed framework still needs to run several times independently to obtain models with different compression ratios. That is, to obtain a model with different compression ratios, you have to run SFW another time from the beginning. Therefore, I am not sure why this method can avoid retraining.\n  \n2. I notice that In Figure 2, the performance of ABFP is comparable to or even better than SFW. Could the author show the advantage of SFW compared to ABFP?\n\n3. Although Theorem 3.1 shows the convergence of SFW with batch gradient dependent step size in the non-convex setting.  However, the assumption of the objective function seems to be too strong (L-smooth and L-Lipschitz). I believe neither of the two studied problems would satisfy these assumptions. Please correct me if I am wrong.\n", "clarity,_quality,_novelty_and_reproducibility": "Quality and clarity are good. Originality is minor.", "summary_of_the_review": "The experiment result is convincing. I also like the idea of combing the SFW and the group-k-support norm constraint together, but I didn’t find much novelty in the theory and algorithm respects.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666472199479}], "openreview_url": "https://openreview.net/forum?id=ueEMZjY9WiM", "arxiv_id": "2205.11921", "paper_pdf": "papers/ueEMZjY9WiM.pdf", "paper_pdf_sha256": "71945e16dd49a9b53db08ca758b1bee9cb33bb83ebc3d460dac27bc13c4fb8f8", "paper_pdf_bytes": 564052, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZIB-IOL/compression-aware-SFW", "code_repository": "ZIB-IOL/compression-aware-SFW", "code_commit": "6e9e0febc09f4af43596c1aceaf8c29e6214fdb6", "code_archive": "repos/ueEMZjY9WiM.zip", "code_archive_sha256": "86c89f0e738ea1361b4e957ad813fc6d9393d7dafca9ff5a3d68e943d4a0aa0c", "code_archive_bytes": 55203, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 49, "github_languages": {"Python": 198043, "TeX": 296}, "github_archived": false, "github_pushed_at": "2023-02-22T09:22:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/compression-aware-training-of-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "aJ_GcB4vcT0", "year": 2022, "status": "rejected", "title": "Unsupervised Learning of Neurosymbolic Encoders", "authors": ["Eric Zhan", "Jennifer J. Sun", "Ann Kennedy", "Yisong Yue", "Swarat Chaudhuri"], "authorids": ["~Eric_Zhan1", "~Jennifer_J._Sun1", "~Ann_Kennedy1", "~Yisong_Yue1", "~Swarat_Chaudhuri1"], "authors_source": "OpenReview API", "abstract": "We present a framework for the unsupervised learning of neurosymbolic encoders, i.e., encoders obtained by composing neural networks with symbolic programs from a domain-specific language. Such a framework can naturally incorporate symbolic expert knowledge into the learning process and lead to more interpretable and factorized latent representations than fully neural encoders. Also, models learned this way can have downstream impact, as many analysis workflows can benefit from having clean programmatic descriptions. We ground our learning algorithm in the variational autoencoding (VAE) framework, where we aim to learn a neurosymbolic encoder in conjunction with a standard decoder. Our algorithm integrates standard VAE-style training with modern program synthesis techniques. We evaluate our method on learning latent representations for real-world trajectory data from animal biology and sports analytics. We show that our approach offers significantly better separation than standard VAEs and leads to practical gains on downstream tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "zHslqxab8Jc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3881/Reviewer_cxLd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes to learn a novel, interpretable neuro-symbolic encoder for sequences in an autoencoding framework. The key idea is that one learns both a symbolic as well as a neural encoder, and gradually makes the symbolic encoder more and more structured progressively increasing the complexity of it. Given such an encoder, the paper proceeds to show that the encoding from it is useful for clustering sequence data and demonstrates gains over other “unstructured” purely neural approaches.\n", "review_text": "Positives\n+ The general problem of learning programs to represent data in unsupervised learning is novel and potentially fundamental to machine learning and AI\n+ The proposed method is simple and interesting, and seems to achieve nice gains over purely neural approaches\n+ The paper is generally well written\n\nNegatives\n- My main issue is that the neuroscience data that the paper currently mainly evaluates on is not something a general ML audience might be very familiar with, and thus, it is hard to estimate what makes the task hard or easy. Could the authors give intuitions on why regular neural encodings fail to cluster in a better manner? Is it because of low SNR in the sequence data, or that RNN encoders pick up on spurious noise signals and thus do not yield a sufficiently discriminative embedding? If that is the case how would a convolutional encoder perform? Are the gains from the symbolic approach something that could just be achieved by a convolutional encoder? How much do we really need the symbolic approach and what are the inductive biases that we gain from them? I think these are important questions to answer in order to judge the efficacy of the proposed approach. (*)\n\n- In terms of the methodology, I do not understand why increasing the complexity of the programs gradually also leads to the purely neural part becoming less and less informative in the toy-dataset. Could it not have been the case that somehow the more complex program just learnt the features that the less complex program (with say length 2) learnt, and that in going from Fig. 3 d) to 3 e) we just learnt the same embedding space for the neural part? Is it encouraged that we learn more complex programs when we increase the length of the programs by tuning the corresponding gamma weights in Eqn. 5, or the corresponding channel capacities? (*)\n\n- I generally think such an encoding is potentially very useful for image data, and might be a big step towards better image representations and advances towards human-like intelligence. It would be great if the authors could comment on the potential for extending the approach to such applications and how far we might be from such use-cases. \n\nMinor Points:\n1) [A] is a closely related work that would be useful to cite in a neural-symbolic + amortized inference / VAE context.\n\n[A]: Vedantam, Ramakrishna, Karan Desai, Stefan Lee, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. 2019. “Probabilistic Neural-Symbolic Models for Interpretable Visual Question Answering.” arXiv [cs.LG]. arXiv. http://arxiv.org/abs/1902.07864.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to learn a novel, interpretable neuro-symbolic encoder for sequences in an autoencoding framework. The key idea is that one learns both a symbolic as well as a neural encoder, and gradually makes the symbolic encoder more and more structured progressively increasing the complexity of it. Given such an encoder, the paper proceeds to show that the encoding from it is useful for clustering sequence data and demonstrates gains over other “unstructured” purely neural approaches.\n", "main_review": "Positives\n+ The general problem of learning programs to represent data in unsupervised learning is novel and potentially fundamental to machine learning and AI\n+ The proposed method is simple and interesting, and seems to achieve nice gains over purely neural approaches\n+ The paper is generally well written\n\nNegatives\n- My main issue is that the neuroscience data that the paper currently mainly evaluates on is not something a general ML audience might be very familiar with, and thus, it is hard to estimate what makes the task hard or easy. Could the authors give intuitions on why regular neural encodings fail to cluster in a better manner? Is it because of low SNR in the sequence data, or that RNN encoders pick up on spurious noise signals and thus do not yield a sufficiently discriminative embedding? If that is the case how would a convolutional encoder perform? Are the gains from the symbolic approach something that could just be achieved by a convolutional encoder? How much do we really need the symbolic approach and what are the inductive biases that we gain from them? I think these are important questions to answer in order to judge the efficacy of the proposed approach. (*)\n\n- In terms of the methodology, I do not understand why increasing the complexity of the programs gradually also leads to the purely neural part becoming less and less informative in the toy-dataset. Could it not have been the case that somehow the more complex program just learnt the features that the less complex program (with say length 2) learnt, and that in going from Fig. 3 d) to 3 e) we just learnt the same embedding space for the neural part? Is it encouraged that we learn more complex programs when we increase the length of the programs by tuning the corresponding gamma weights in Eqn. 5, or the corresponding channel capacities? (*)\n\n- I generally think such an encoding is potentially very useful for image data, and might be a big step towards better image representations and advances towards human-like intelligence. It would be great if the authors could comment on the potential for extending the approach to such applications and how far we might be from such use-cases. \n\nMinor Points:\n1) [A] is a closely related work that would be useful to cite in a neural-symbolic + amortized inference / VAE context.\n\n[A]: Vedantam, Ramakrishna, Karan Desai, Stefan Lee, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. 2019. “Probabilistic Neural-Symbolic Models for Interpretable Visual Question Answering.” arXiv [cs.LG]. arXiv. http://arxiv.org/abs/1902.07864.", "summary_of_the_review": "My main concerns with the paper are around whether the RNN encodings are overfitting and picking up on spurious correlations in sequence data that say a convolutional encoder might already fix, which will mean that one does not really need symbolic encoders for the current task. This is an even larger concern in light of the fact that a general machine learning researcher might not have a lot of great intuitions about the neuroscience sequence data and that makes it hard to assess the impact of the work. For the rebuttal I would encourage the authors to address concerns marked with (*) in the main review.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636902288756}, {"id": "XPeyY1pU_w3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3881/Reviewer_zt4F"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper is a clean extension of the supervised neurosymbolic approach\nof Shah et al. to the unsupervised setting where a VAE is used\nto encode input to the system as an interpretable program.\nThese latent representations are shown to be interpretable for\na synthetic experiment and useful for downstream tasks.\n\t", "review_text": "The paper is clearly written and the idea was easy to understand and\nfollow. The experiments are always going to be a challenge in a\nspace like this as how can we say if the symbolic latent\nrepresentation meaningfully captures something in the data.\nThe synthetic result while a little contrived is convincing that\nat least in a controlled setting a reasonable program is produced.\nI wish there was more exploration of how often are sensible\nprograms learned and how sensitive this is to the choice of DSLs.\nMaybe something comparing the programs of experts to what\nthe latent representation learned? Basically the program equivlaent\nof showing a grid of faces.\n\nThis would also benefit from comparisons that use purely the neural component\n\nI am also curious empirically how it's decided how many symbolic\nprograms should be in the latent representation? Would a similar\nprocess be used for real-world data?\n\nMy main concern is how heavily this paper relies on the NEAR\nwork particularly for learning the symbolic encoder and even\nthe experimental data used. As the VAE bits are fairly standard\nit makes the work feel fairly incremental.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper is a clean extension of the supervised neurosymbolic approach\nof Shah et al. to the unsupervised setting where a VAE is used\nto encode input to the system as an interpretable program.\nThese latent representations are shown to be interpretable for\na synthetic experiment and useful for downstream tasks.\n\t", "main_review": "The paper is clearly written and the idea was easy to understand and\nfollow. The experiments are always going to be a challenge in a\nspace like this as how can we say if the symbolic latent\nrepresentation meaningfully captures something in the data.\nThe synthetic result while a little contrived is convincing that\nat least in a controlled setting a reasonable program is produced.\nI wish there was more exploration of how often are sensible\nprograms learned and how sensitive this is to the choice of DSLs.\nMaybe something comparing the programs of experts to what\nthe latent representation learned? Basically the program equivlaent\nof showing a grid of faces.\n\nThis would also benefit from comparisons that use purely the neural component\n\nI am also curious empirically how it's decided how many symbolic\nprograms should be in the latent representation? Would a similar\nprocess be used for real-world data?\n\nMy main concern is how heavily this paper relies on the NEAR\nwork particularly for learning the symbolic encoder and even\nthe experimental data used. As the VAE bits are fairly standard\nit makes the work feel fairly incremental.\n", "summary_of_the_review": "This is an interesting and novel way to learn programs as symbolic encoders of the input. I have some nagging concerns about how incremental the contribution given how much this work relies on Shah et al.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635915142272}, {"id": "MyK3_gcPvNF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3881/Reviewer_a6JR"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper tries to use VAE-based generative task to generate a latent representation that could be encoded as symbolic program. The authors show that in this way it could achieve better cluster results and also generate interpret-able and reasonable programs for each data.", "review_text": "I think the idea of utilizing self-supervised training task to generate symbolic programs is very interesting and worth exploring.\n\nMy major concerns lie in the approach and evaluation.\n\nFor approach:\n1) Though I read the paper carefully, I still didn't fully get how the authors represent the program via z(alpha, fi). It seems that the output of VAE should be a vector, and how did the authors decode z(alpha, fi) to alpha and fi, especially when alpha should be a \"discrete program architecture\". Is alpha a program or just a single operator? Also how did you encode q(alpha, fi) when alpha could be a complex tree-structured program?\n2) I'm also not fully understand equation 3. How to you search the child program of alpha in a differentiable manner? Why this is a supervised task (from my understanding it's more like finding a child program that outputs similar encoding vector as parent program).\n3) You mention you \"repeat steps 1&2 until q(apha, fi) is fully symbolic\". The meaning of fully symbolic is really confusing to me. What is the stopping criterion? Also, what's the intuition that you need to update the parameter agrain after fixing program alpha?\n4) The authors didn't mention how each function in the DSL is implemented in paper and appendix. I think it's very important to understand the whole architecture. Also, it's important to know how you handle the non-differentiable operators, such as if-then, and others.\n\n\nFor evaluation:\n1) Current evaluation seems not very convincing to me. The authors only show that with the help of symbolic program, the method could get representations with better cluster quality (program helps representation learning). But I think a more intersting perspective is to see whether the learned program itself is helpful. For example, whether it could be used to predict future trajectory (such as 3-body problem), or even help solving some high-level reasoning tasks.\n2) Lack of the ablation study of the proposed framework. For example, why it's important to update the parameter again. Is the design of DSL influences the final representation learning, etc.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper tries to use VAE-based generative task to generate a latent representation that could be encoded as symbolic program. The authors show that in this way it could achieve better cluster results and also generate interpret-able and reasonable programs for each data.", "main_review": "I think the idea of utilizing self-supervised training task to generate symbolic programs is very interesting and worth exploring.\n\nMy major concerns lie in the approach and evaluation.\n\nFor approach:\n1) Though I read the paper carefully, I still didn't fully get how the authors represent the program via z(alpha, fi). It seems that the output of VAE should be a vector, and how did the authors decode z(alpha, fi) to alpha and fi, especially when alpha should be a \"discrete program architecture\". Is alpha a program or just a single operator? Also how did you encode q(alpha, fi) when alpha could be a complex tree-structured program?\n2) I'm also not fully understand equation 3. How to you search the child program of alpha in a differentiable manner? Why this is a supervised task (from my understanding it's more like finding a child program that outputs similar encoding vector as parent program).\n3) You mention you \"repeat steps 1&2 until q(apha, fi) is fully symbolic\". The meaning of fully symbolic is really confusing to me. What is the stopping criterion? Also, what's the intuition that you need to update the parameter agrain after fixing program alpha?\n4) The authors didn't mention how each function in the DSL is implemented in paper and appendix. I think it's very important to understand the whole architecture. Also, it's important to know how you handle the non-differentiable operators, such as if-then, and others.\n\n\nFor evaluation:\n1) Current evaluation seems not very convincing to me. The authors only show that with the help of symbolic program, the method could get representations with better cluster quality (program helps representation learning). But I think a more intersting perspective is to see whether the learned program itself is helpful. For example, whether it could be used to predict future trajectory (such as 3-body problem), or even help solving some high-level reasoning tasks.\n2) Lack of the ablation study of the proposed framework. For example, why it's important to update the parameter again. Is the design of DSL influences the final representation learning, etc.", "summary_of_the_review": "This paper proposes a very interesting research direction, but the writing and organization of the proposed method make it hard to understand it. In addition, the current evaluation is way too simple and not interesting enough. I highly recommend the authors to add some experiments to show that the learned programs can help some down-stream tasks.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635914623945}, {"id": "I_DmoVDizn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3881/Reviewer_G4ix"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose an unsupervised approach to train neurosymbolic encoders to obtain a programmatically interpretable representation using the dictionary of a domain-specific language (DSL). The experimental results show that the proposed method can outperform baseline neural encoders in extracting semantically meaningful representations of behavior in CalMS21 (mice) dataset. The results also show that the performance can be robust across different DSL designs by domain experts.\n\n\n\n\n", "review_text": "The strength of this paper is described above. Although the concept and method are simple and new, and the paper itself is well-written, the significance in the problem setting and experimental results were unclear to me at this stage. The specific comments are as follows. \n\nMajor concerns: \n1. Although the unsupervised neurosymbolic “encoder” is new (as described in Sec. 5, but the decoders existed in [Feinman & Lake, 2020] and also the following [1]?), but I did not understand why the authors considered that creating the encoder is important.\n2. Although the concept of the proposed method is general and interesting, the applied domains were limited to animal/human trajectories. That is, it is unclear to me why the authors selected these domains to verify the proposed method. Or verification using other domain data may be also required.\n3. The significance of the experimental results was unclear to me. The task to extract distinctive information between the data w/ and w/o interaction in mice may be considered less challenging for me. The basketball task including offensive/defensive players is more challenging, but the result of the proposed method was overall comparable with the baselines (whereas it improved slightly with respect to purity). In the basketball task, I am also concerned about the importance of offensive/defensive players detection (this is usually given). The final downstream task (not tasks) is a classification task, and the results were comparable with the previous work (Sun et al. 2021b). \n\nMinor comments:\n4. What is x in Sec. 2.2? If it is some inputs, what types of inputs?\n5. Which is correct, .428 in NMI of Figure 4 or .423 of Table 1? Or What DSL was used in Table 1 results? \n\n[1] Halley Young, Osbert Bastani, Mayur Naik, Learning Neurosymbolic Generative Models via Program Synthesis, ICML, 97:7144-7153, 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose an unsupervised approach to train neurosymbolic encoders to obtain a programmatically interpretable representation using the dictionary of a domain-specific language (DSL). The experimental results show that the proposed method can outperform baseline neural encoders in extracting semantically meaningful representations of behavior in CalMS21 (mice) dataset. The results also show that the performance can be robust across different DSL designs by domain experts.\n\n\n\n\n", "main_review": "The strength of this paper is described above. Although the concept and method are simple and new, and the paper itself is well-written, the significance in the problem setting and experimental results were unclear to me at this stage. The specific comments are as follows. \n\nMajor concerns: \n1. Although the unsupervised neurosymbolic “encoder” is new (as described in Sec. 5, but the decoders existed in [Feinman & Lake, 2020] and also the following [1]?), but I did not understand why the authors considered that creating the encoder is important.\n2. Although the concept of the proposed method is general and interesting, the applied domains were limited to animal/human trajectories. That is, it is unclear to me why the authors selected these domains to verify the proposed method. Or verification using other domain data may be also required.\n3. The significance of the experimental results was unclear to me. The task to extract distinctive information between the data w/ and w/o interaction in mice may be considered less challenging for me. The basketball task including offensive/defensive players is more challenging, but the result of the proposed method was overall comparable with the baselines (whereas it improved slightly with respect to purity). In the basketball task, I am also concerned about the importance of offensive/defensive players detection (this is usually given). The final downstream task (not tasks) is a classification task, and the results were comparable with the previous work (Sun et al. 2021b). \n\nMinor comments:\n4. What is x in Sec. 2.2? If it is some inputs, what types of inputs?\n5. Which is correct, .428 in NMI of Figure 4 or .423 of Table 1? Or What DSL was used in Table 1 results? \n\n[1] Halley Young, Osbert Bastani, Mayur Naik, Learning Neurosymbolic Generative Models via Program Synthesis, ICML, 97:7144-7153, 2019.", "summary_of_the_review": "Although the concept and method are simple and new, and the paper itself is well-written, the significance in the problem setting and experimental results were unclear for me. Therefore, it is difficult for me to provide a higher rating at this stage. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635715208442}], "openreview_url": "https://openreview.net/forum?id=aJ_GcB4vcT0", "arxiv_id": "2107.13132", "paper_pdf": "papers/aJ_GcB4vcT0.pdf", "paper_pdf_sha256": "a5f1a50c761a88d7969b46cbee49f7b38798e99801fb54b1ae6e5d21c96f7b09", "paper_pdf_bytes": 3419080, "paper_pdf_source": "openreview", "code_url": "https://github.com/ezhan94/neurosymbolic-encoders", "code_repository": "ezhan94/neurosymbolic-encoders", "code_commit": "c52a83a33c99e02c52be7b176edeb2c6bdedec78", "code_archive": "repos/aJ_GcB4vcT0.zip", "code_archive_sha256": "feda2969087efd7f01a5a984eb29bf87928e74a4e7f594980dee37d587956e5d", "code_archive_bytes": 85769, "code_file_count": 52, "code_extensions": {".py": 50, ".sh": 2}, "github_disk_usage_kb": 65, "github_languages": {"Python": 227332, "Shell": 8642}, "github_archived": false, "github_pushed_at": "2023-06-05T02:39:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unsupervised-learning-of-neurosymbolic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xFYXLlpIyPQ", "year": 2021, "status": "rejected", "title": "Guarantees for Tuning the Step Size using a Learning-to-Learn Approach", "authors": ["Xiang Wang", "Shuai Yuan", "Chenwei Wu", "Rong Ge"], "authorids": ["~Xiang_Wang1", "shuai@cs.duke.edu", "~Chenwei_Wu1", "~Rong_Ge1"], "authors_source": "OpenReview API", "abstract": "Learning-to-learn---using optimization algorithms to learn a new optimizer---has successfully trained efficient optimizers in practice. This approach relies on meta-gradient descent on a meta-objective based on the trajectory that the optimizer generates. However, there were few theoretical guarantees on how to avoid meta-gradient explosion/vanishing problem, or how to train an optimizer with good generalization performance. In this paper, we study the learning-to-learn approach on a simple problem of tuning the step size for quadratic loss. Our results show that although there is a way to design the meta-objective so that the meta-gradient remains polynomially bounded, computing the meta-gradient directly using backpropagation leads to numerical issues that look similar to gradient explosion/vanishing problems. We also characterize when it is necessary to compute the meta-objective on a separate validation set instead of the original training set. Finally, we verify our results empirically and show that a similar phenomenon appears even for more complicated learned optimizers parametrized by neural networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "YqUiMEr5eCI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper993/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors analyze two different approaches to \"learning to learn\" which correct various pathologies, however primarily restricted to the quadratic setting. The first is the use of the log objective rather than the raw objective for the inner-loop optimization step, and the second is the differences between train-by-train and train-by-validation. Overall, however, I feel that this work relies much too heavily on its supplemental material, as the proofs themselves are not well spelled out within the text itself, and at 68 pages total this is perhaps too extensive for this venue.\n\nThe authors give a reasonably good overview of recent approaches to to learning to learn, or meta-learning for optimizers. However, it is worth pointing out that they omit much of the early work in this area, particularly the work of Hochreiter et al., 2001 and the various citations therein, which give this area its name.\n\nNotationally, the work is also very dense. The authors might consider whether it is of value in introducing their approach whether to do away with the focus on supervised (e.g. input/output x/y data) models and instead focus on noisy gradients evaluated at each iteration. Otherwise it might be helpful to be more explicit about the indices x_i -> y_i when introducing these models, although this is perhaps a very nit-picky and aesthetic consideration.\n\nTheorem 3 is itself interesting, however it seems to follow quite directly from Section 2.3 of Metz et al. Although the authors do provide more detail, this is restricted to the quadratic inner-loop setting, whereas the description of Metz doesn't go so far due to the more general problem setting (and changing Hessians).\n\nTheorem 4, similarly is of interest. But it's not clear how much practical value it has. As far as I can tell the authors do not use this log objective when discussing the train-by-train/vs/train-by-validation setting. I may have misunderstood, but if this is not the case why not? Similarly, how does this compare empirically to the solution proposed by Metz et al? And lastly, the authors did not describe in detail how they dealt with the intermediate gradients which would have been most useful.\n\nFinally, the remaining theorems are interesting, but mostly seem to confirm the results of Metz at al, and are restricted to the quadratic setting. So more work might be needed to show why this is particularly of use.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Somewhat interesting, but too much hidden in the supplemental", "review": "The authors analyze two different approaches to \"learning to learn\" which correct various pathologies, however primarily restricted to the quadratic setting. The first is the use of the log objective rather than the raw objective for the inner-loop optimization step, and the second is the differences between train-by-train and train-by-validation. Overall, however, I feel that this work relies much too heavily on its supplemental material, as the proofs themselves are not well spelled out within the text itself, and at 68 pages total this is perhaps too extensive for this venue.\n\nThe authors give a reasonably good overview of recent approaches to to learning to learn, or meta-learning for optimizers. However, it is worth pointing out that they omit much of the early work in this area, particularly the work of Hochreiter et al., 2001 and the various citations therein, which give this area its name.\n\nNotationally, the work is also very dense. The authors might consider whether it is of value in introducing their approach whether to do away with the focus on supervised (e.g. input/output x/y data) models and instead focus on noisy gradients evaluated at each iteration. Otherwise it might be helpful to be more explicit about the indices x_i -> y_i when introducing these models, although this is perhaps a very nit-picky and aesthetic consideration.\n\nTheorem 3 is itself interesting, however it seems to follow quite directly from Section 2.3 of Metz et al. Although the authors do provide more detail, this is restricted to the quadratic inner-loop setting, whereas the description of Metz doesn't go so far due to the more general problem setting (and changing Hessians).\n\nTheorem 4, similarly is of interest. But it's not clear how much practical value it has. As far as I can tell the authors do not use this log objective when discussing the train-by-train/vs/train-by-validation setting. I may have misunderstood, but if this is not the case why not? Similarly, how does this compare empirically to the solution proposed by Metz et al? And lastly, the authors did not describe in detail how they dealt with the intermediate gradients which would have been most useful.\n\nFinally, the remaining theorems are interesting, but mostly seem to confirm the results of Metz at al, and are restricted to the quadratic setting. So more work might be needed to show why this is particularly of use.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603979064677}, {"id": "nVEwoj7HeVx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper993/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents novel theoretical results on learning a step size for vanilla GD and SGD by unrolling the optimization steps and back-propagating, taking into account the simple problem minimizing quadratic functions and mean-square errors. The authors could demonstrate the occurrence of already-detected phenomena for learned optimizers, such as gradient explosion/vanishing and over-fitting, in the particular studied case. A few experiments illustrate what the developed theory predicts. \n\nTo the best of my knowledge, the literature lacks theoretical guarantees for learned optimizers and results as those within this manuscript can shed light on new ways to develop a solid learned optimizer theory. \n\nTheorems 1 and 2, seem central to motivate the work. Yet, they are very imprecise.\n\nIn Theorem 1:\n\n\"For tuning the step size of gradient descent on a quadratic objective, if the meta-objective is the\nloss of the last iteration, then the meta-gradient can explode/vanish.\"\n\nComment: Can is a loose expression — since it is a theorem, can it be shown that there are situations where they actually explode/ vanish? I believe this is discussed in Section 3, then why not state clearly that in the case of quadratics, the gradient vanishes?\n\nIn Theorem 2:\n\n\"For a simple least squares problem in d dimensions, if the number of samples n is a constant fraction of d (e.g., d/2), and the samples have large noise, then the train-by-train approach performs much worse than train-by-validation\"\n\nComment: what is much worse? How is this shown?\n\n\"On the other hand, when number of samples n is large, train-by-train can get close to error dσ2/\nn, which is optimal.\"\n\nComment: what is optimal?\"\n\n\nSection 3: After Equation (2), \"We first show that when the number of samples is small (in particular n < d) and the\nnoise is a large enough constant, train-by-train can be much worse than train-by-validation\"\n\nComment: Too many imprecise terms.\n\nUntil Section 3, the discussion emphasizes the use of logarithm for the meta-objective. But from section 4 onward, I see that the regular cost is used for the meta-objective. I am not sure if I am missing something here?\n\nFigure 2 is invoked to support ‘In Figure 2, we verify the observation from Metz et al. (2019) that the optimal step size depends on inner training length.’— While the trend shows difference with length, I was unsure if the difference is significant—a comparison of setting a constant step size would be more revealing.\n\nI also feel many of the proofs have not been properly explained in terms of the steps used in derivations, for example the various inequalities/upper bounds For Figures showing Training and Test RMSE, I would strongly recommend use of relative RMSE,\nthat is relative to the true value, expressed perhaps in dB plot. Currently it is very difficult to gauge the differences in terms of absolute RMSE.\n\nQuestion: What is the overall conclusion after the experiments?\n\nIs it that use of a logarithm helps remove vanishing gradient problem? From the experiments it seems the emphasis is that given data should be used with a separate validation data set.\n\nI do not think that the quadratic analysis necessarily makes it insignificant, since it does aid understanding. Nevertheless, I think its rather poorly written and structured and I was left trying to find out what is being emphasized. As a result, even this simple-case analysis seems unclear to me. The imprecise nature of the Theorems added more to the confusion.\n\n\nPros: \n- There are not many theoretical works on learned optimizers, thus this can be interesting to provide new insights.\n\n- The proofs seem to be correct and rely upon interesting mathematical properties, as far as I could follow.\n\nCons: \n\n- It seems to be hard to generalize the proposed analysis to more intricate (and practical) problems, such as training MLPs as optimizers. If this work intends to serve as a basis for further investigations, the authors should highlight which ideas could be still used in more general cases.\n\n- The results of Theorem 3 were partially discussed in Metz et al. (2019) (Sec. 2.3). The authors could acknowledge this fact in the text.\n\n- In Section 3, c_{min} is assumed to be positive, but what does it mean? Can w_0 be orthogonal to some eigenvector u_i of H?\n\n- There are many typos and colloquial statements are presented. Sentences as \"it is OK to use\", \"but didn’t give any\", should be avoided. Moreover there are 21 \"it's\" throughout the paper.\n\n- The statement \"Therefore in order to train neural networks, it is better to use train-by-validation.\" is too strong and was not proved.\n\n- Why is SGD noisier in Fig. 7 than the other methods? Are TbT and TbV calculating a full-batch gradient? This would not be a fair comparison if SGD had only access to an estimate of this gradient.\n\n\n\nMinor concerns:\n\n- What did the authors mean by \"constant fraction\" in Theorem 2? \n\n- I think the task definition should also depend on the initial condition w_0 as different initializations provide different trajectories.\n\n- The population losses were defined but not used in the text. Maybe it should be defined only in the appendix for clarity purposes.\n\n- What is the relevance of \"L > \\alpha\" to show that F(\\eta) is strictly convex in the proof of Theorem 3?\n\n- The constants in Theorem 5 and 6 are unused and could be omitted for the sake of concision. I recommend defining then in the proof only.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A 4+ paper because of the interesting problem it tries to address", "review": "This paper presents novel theoretical results on learning a step size for vanilla GD and SGD by unrolling the optimization steps and back-propagating, taking into account the simple problem minimizing quadratic functions and mean-square errors. The authors could demonstrate the occurrence of already-detected phenomena for learned optimizers, such as gradient explosion/vanishing and over-fitting, in the particular studied case. A few experiments illustrate what the developed theory predicts. \n\nTo the best of my knowledge, the literature lacks theoretical guarantees for learned optimizers and results as those within this manuscript can shed light on new ways to develop a solid learned optimizer theory. \n\nTheorems 1 and 2, seem central to motivate the work. Yet, they are very imprecise.\n\nIn Theorem 1:\n\n\"For tuning the step size of gradient descent on a quadratic objective, if the meta-objective is the\nloss of the last iteration, then the meta-gradient can explode/vanish.\"\n\nComment: Can is a loose expression — since it is a theorem, can it be shown that there are situations where they actually explode/ vanish? I believe this is discussed in Section 3, then why not state clearly that in the case of quadratics, the gradient vanishes?\n\nIn Theorem 2:\n\n\"For a simple least squares problem in d dimensions, if the number of samples n is a constant fraction of d (e.g., d/2), and the samples have large noise, then the train-by-train approach performs much worse than train-by-validation\"\n\nComment: what is much worse? How is this shown?\n\n\"On the other hand, when number of samples n is large, train-by-train can get close to error dσ2/\nn, which is optimal.\"\n\nComment: what is optimal?\"\n\n\nSection 3: After Equation (2), \"We first show that when the number of samples is small (in particular n < d) and the\nnoise is a large enough constant, train-by-train can be much worse than train-by-validation\"\n\nComment: Too many imprecise terms.\n\nUntil Section 3, the discussion emphasizes the use of logarithm for the meta-objective. But from section 4 onward, I see that the regular cost is used for the meta-objective. I am not sure if I am missing something here?\n\nFigure 2 is invoked to support ‘In Figure 2, we verify the observation from Metz et al. (2019) that the optimal step size depends on inner training length.’— While the trend shows difference with length, I was unsure if the difference is significant—a comparison of setting a constant step size would be more revealing.\n\nI also feel many of the proofs have not been properly explained in terms of the steps used in derivations, for example the various inequalities/upper bounds For Figures showing Training and Test RMSE, I would strongly recommend use of relative RMSE,\nthat is relative to the true value, expressed perhaps in dB plot. Currently it is very difficult to gauge the differences in terms of absolute RMSE.\n\nQuestion: What is the overall conclusion after the experiments?\n\nIs it that use of a logarithm helps remove vanishing gradient problem? From the experiments it seems the emphasis is that given data should be used with a separate validation data set.\n\nI do not think that the quadratic analysis necessarily makes it insignificant, since it does aid understanding. Nevertheless, I think its rather poorly written and structured and I was left trying to find out what is being emphasized. As a result, even this simple-case analysis seems unclear to me. The imprecise nature of the Theorems added more to the confusion.\n\n\nPros: \n- There are not many theoretical works on learned optimizers, thus this can be interesting to provide new insights.\n\n- The proofs seem to be correct and rely upon interesting mathematical properties, as far as I could follow.\n\nCons: \n\n- It seems to be hard to generalize the proposed analysis to more intricate (and practical) problems, such as training MLPs as optimizers. If this work intends to serve as a basis for further investigations, the authors should highlight which ideas could be still used in more general cases.\n\n- The results of Theorem 3 were partially discussed in Metz et al. (2019) (Sec. 2.3). The authors could acknowledge this fact in the text.\n\n- In Section 3, c_{min} is assumed to be positive, but what does it mean? Can w_0 be orthogonal to some eigenvector u_i of H?\n\n- There are many typos and colloquial statements are presented. Sentences as \"it is OK to use\", \"but didn’t give any\", should be avoided. Moreover there are 21 \"it's\" throughout the paper.\n\n- The statement \"Therefore in order to train neural networks, it is better to use train-by-validation.\" is too strong and was not proved.\n\n- Why is SGD noisier in Fig. 7 than the other methods? Are TbT and TbV calculating a full-batch gradient? This would not be a fair comparison if SGD had only access to an estimate of this gradient.\n\n\n\nMinor concerns:\n\n- What did the authors mean by \"constant fraction\" in Theorem 2? \n\n- I think the task definition should also depend on the initial condition w_0 as different initializations provide different trajectories.\n\n- The population losses were defined but not used in the text. Maybe it should be defined only in the appendix for clarity purposes.\n\n- What is the relevance of \"L > \\alpha\" to show that F(\\eta) is strictly convex in the proof of Theorem 3?\n\n- The constants in Theorem 5 and 6 are unused and could be omitted for the sake of concision. I recommend defining then in the proof only.\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603874153848}, {"id": "P93LCjaF30", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper993/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers algorithms that attempt to learn learning rates for gradient descent by gradient descent. Analysis is provided for a few specific quadratic losses showing that the gradient with respect to the learning rate may explode or vanish, and taking the logarithm is suggested to mitigate this. Further results suggest that implementing the gradient of the log comes with interesting numerical difficulties as *intermediate results* might explode or vanish even if the final answer does not.\n\nNext, linear regression problems are analyzed in both the over-determined and under-determined settings. It shown that in the under-determined setting the optimal learning rate when tuned on the training set is very far from the optimal learning rate when tuned on the validation set. \n\nExperiments are presented validating these theoretical findings, as well as comparisons to manual tuning on MNIST.\n\n\nI felt that the claims in the abstract were a bit overblown here. I would not say that there has been a characterization of when to use validation vs train set, or that the proposed logarithmic method necessarily avoids vanishing/exploding metagradients. Instead, these questions have been addressed in the very specific setting of linear regression.\n\nTheorems 5 and 6 appear to be making statements about the minimizers of the meta-objectives. Is there any guarantee that the meta-descent will efficiently approximate these minimizers so that the final algorithm actually comes with a guarantee?\n\nOverall, I was a bit underwhelmed by the linear regression setting. This setting seems a bit limited since it is almost possible to write in closed form what the sgd iterates will do. I am willing to concede that there is some significant difficulty here, or that this setting is somehow necessarily broadly relevant, but I don’t think it has been clearly discussed. As for the empirical comparison, MNIST is frankly also a kind of toy dataset at this point, and even in this setting it does not seem that any gain is to be had over simply tuning SGD via a logarithmic grid.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "metalearning on linear regression", "review": "This paper considers algorithms that attempt to learn learning rates for gradient descent by gradient descent. Analysis is provided for a few specific quadratic losses showing that the gradient with respect to the learning rate may explode or vanish, and taking the logarithm is suggested to mitigate this. Further results suggest that implementing the gradient of the log comes with interesting numerical difficulties as *intermediate results* might explode or vanish even if the final answer does not.\n\nNext, linear regression problems are analyzed in both the over-determined and under-determined settings. It shown that in the under-determined setting the optimal learning rate when tuned on the training set is very far from the optimal learning rate when tuned on the validation set. \n\nExperiments are presented validating these theoretical findings, as well as comparisons to manual tuning on MNIST.\n\n\nI felt that the claims in the abstract were a bit overblown here. I would not say that there has been a characterization of when to use validation vs train set, or that the proposed logarithmic method necessarily avoids vanishing/exploding metagradients. Instead, these questions have been addressed in the very specific setting of linear regression.\n\nTheorems 5 and 6 appear to be making statements about the minimizers of the meta-objectives. Is there any guarantee that the meta-descent will efficiently approximate these minimizers so that the final algorithm actually comes with a guarantee?\n\nOverall, I was a bit underwhelmed by the linear regression setting. This setting seems a bit limited since it is almost possible to write in closed form what the sgd iterates will do. I am willing to concede that there is some significant difficulty here, or that this setting is somehow necessarily broadly relevant, but I don’t think it has been clearly discussed. As for the empirical comparison, MNIST is frankly also a kind of toy dataset at this point, and even in this setting it does not seem that any gain is to be had over simply tuning SGD via a logarithmic grid.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603851933054}, {"id": "cyIqI9IMTq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper993/AnonReviewer1"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Overview\n---\n\nMeta-gradient descent is an approach to step-size adaptation in which the step-size is adapted by considering how it influences the loss function over time. Intuitively, one can think of the trajectory of parameters $(w_s)_{s=1}^t$ as being a function of the step-size $\\eta$, and try to control the loss indirectly through the step-size's influence on the weight trajectory. This paper provides guarantees for this class of algorithms when applied to a quadradic loss function. It is shown that the meta-objective $\\ell_t(\\eta)=\\frac{1}{2}w_t(\\eta)^\\top H w_t(\\eta)$ contains no bad local solutions, but can suffer from vanishing/exploding gradients. It is then shown that this can be remedied simply considering the logarithm of this meta-objective, but that this too will have issues with numerical stability if approached with back-propagation. Finally, results related to the generalization ability of these methods are presented.\n\n**Overall, I recommend the paper for acceptance.** Despite focusing on a simple quadratic loss setting, the results are quite non-trivial and I believe will be of interest to many in the community. The writing was clear throughout, and I found the proofs that I worked through (Appendix A) to be instructive and of high-quality. The experiments leave a bit to be desired but I think are generally successful in demonstrating the results suggested by the theory.\n\nMy main reservation about this paper is the length of the Appendix. On one hand, I think it's absolutely unreasonable to expect reviewers to work through and verify **60** pages of proofs, for free, on their own time, given a two-week deadline, so I'm inclined to say that this work is better-suited for a journal where the results can be verified properly by reviewers. However, I did work through Appendix A, and the material was well-explained and generally quite excellent, so I'm willing to believe that the rest of the appendix follows suit. But because I don't have time to work through the remaining 50 pages, I can't be sure of the quality and this is reflected in my score, which might otherwise be higher.\n\n## Questions\n---\n- **Page 5**: *\"Note we truncate a sequence and declare the meta loss is high once the weight norm exceeds certain threshold. Specifically, if at the $\\tau^{\\text{th}}$ step, $\\|w_{\\tau,\\eta}\\|\\ge 40\\sigma$, we freeze the training on this task and set $w_{\\tau',\\eta}= 40\\sigma u$ for all $\\tau\\le \\tau'\\le t$, for some arbitrary vector $u$ with unit norm.\"* Why? Is there any specific justification for this, and how were these settings chosen?\n- **Page 6**: *\"Experiments show that in many settings (especially with large t and large $\\eta_0$) the implementation does not converge.\"* Where? Everything seems to be converging in the results shown in this section. Is this demonstrated elsewhere in the paper that I'm missing?\n- **Pages 7/8**: How many trials of these experiments were performed? Is there a reason why figures 5-7 don't include any measures of spread? Without some measure of spread it's hard to tell whether any of the differences are significant, which is problematic when these results are supposed to be verifying the theory presented.\n\n## Minor Comments (which did not affect my score)\n---\n- **Page 2**: *\"one needs to do the meta-training for an optimizer that runs for enough number of steps\"* this reads awkwardly.\n- **Page 2**: *\"Another challenge is about the generalization performance of the learned optimizer\"* This might read more clearly by rephrasing as \"The generalization performance of the learned optimizer is another challenge\"\n- **Page 13**: *\"where the second inequality holds\"* I think this is supposed to read equality rather than inequality.\n\n## Potentially Useful Citations\n---\nThe following citations investigate methods which adapt a *vector* of step-sizes, which might be of interest\n- **Incremental Delta-bar Delta (IDBD)**: Sutton, Richard S. \"Adapting bias by gradient descent: An incremental version of delta-bar-delta.\" AAAI. 1992.\n- **Stochastic Meta-descent (Generalization of IDBD)**: Schraudolph, Nicol N. \"Local gain adaptation in stochastic gradient descent.\" (1999): 569-574.\n- **TIDBD (IDBD for Reinforcement Learning algorithms)**:\n  Kearney, Alex, et al. \"Tidbd: Adapting temporal-difference step-sizes through stochastic meta-descent.\" arXiv preprint arXiv:1804.03334 (2018).\n  \n  Günther, Johannes, et al. \"Meta-learning for Predictive Knowledge Architectures: A Case Study Using TIDBD on a Sensor-rich Robotic Arm.\" Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems. 2019.\n- **AdaGain (Meta-descent for learning stability)**: Jacobsen, Andrew, et al. \"Meta-descent for online, continual prediction.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Great paper, long Appendix", "review": "## Overview\n---\n\nMeta-gradient descent is an approach to step-size adaptation in which the step-size is adapted by considering how it influences the loss function over time. Intuitively, one can think of the trajectory of parameters $(w_s)_{s=1}^t$ as being a function of the step-size $\\eta$, and try to control the loss indirectly through the step-size's influence on the weight trajectory. This paper provides guarantees for this class of algorithms when applied to a quadradic loss function. It is shown that the meta-objective $\\ell_t(\\eta)=\\frac{1}{2}w_t(\\eta)^\\top H w_t(\\eta)$ contains no bad local solutions, but can suffer from vanishing/exploding gradients. It is then shown that this can be remedied simply considering the logarithm of this meta-objective, but that this too will have issues with numerical stability if approached with back-propagation. Finally, results related to the generalization ability of these methods are presented.\n\n**Overall, I recommend the paper for acceptance.** Despite focusing on a simple quadratic loss setting, the results are quite non-trivial and I believe will be of interest to many in the community. The writing was clear throughout, and I found the proofs that I worked through (Appendix A) to be instructive and of high-quality. The experiments leave a bit to be desired but I think are generally successful in demonstrating the results suggested by the theory.\n\nMy main reservation about this paper is the length of the Appendix. On one hand, I think it's absolutely unreasonable to expect reviewers to work through and verify **60** pages of proofs, for free, on their own time, given a two-week deadline, so I'm inclined to say that this work is better-suited for a journal where the results can be verified properly by reviewers. However, I did work through Appendix A, and the material was well-explained and generally quite excellent, so I'm willing to believe that the rest of the appendix follows suit. But because I don't have time to work through the remaining 50 pages, I can't be sure of the quality and this is reflected in my score, which might otherwise be higher.\n\n## Questions\n---\n- **Page 5**: *\"Note we truncate a sequence and declare the meta loss is high once the weight norm exceeds certain threshold. Specifically, if at the $\\tau^{\\text{th}}$ step, $\\|w_{\\tau,\\eta}\\|\\ge 40\\sigma$, we freeze the training on this task and set $w_{\\tau',\\eta}= 40\\sigma u$ for all $\\tau\\le \\tau'\\le t$, for some arbitrary vector $u$ with unit norm.\"* Why? Is there any specific justification for this, and how were these settings chosen?\n- **Page 6**: *\"Experiments show that in many settings (especially with large t and large $\\eta_0$) the implementation does not converge.\"* Where? Everything seems to be converging in the results shown in this section. Is this demonstrated elsewhere in the paper that I'm missing?\n- **Pages 7/8**: How many trials of these experiments were performed? Is there a reason why figures 5-7 don't include any measures of spread? Without some measure of spread it's hard to tell whether any of the differences are significant, which is problematic when these results are supposed to be verifying the theory presented.\n\n## Minor Comments (which did not affect my score)\n---\n- **Page 2**: *\"one needs to do the meta-training for an optimizer that runs for enough number of steps\"* this reads awkwardly.\n- **Page 2**: *\"Another challenge is about the generalization performance of the learned optimizer\"* This might read more clearly by rephrasing as \"The generalization performance of the learned optimizer is another challenge\"\n- **Page 13**: *\"where the second inequality holds\"* I think this is supposed to read equality rather than inequality.\n\n## Potentially Useful Citations\n---\nThe following citations investigate methods which adapt a *vector* of step-sizes, which might be of interest\n- **Incremental Delta-bar Delta (IDBD)**: Sutton, Richard S. \"Adapting bias by gradient descent: An incremental version of delta-bar-delta.\" AAAI. 1992.\n- **Stochastic Meta-descent (Generalization of IDBD)**: Schraudolph, Nicol N. \"Local gain adaptation in stochastic gradient descent.\" (1999): 569-574.\n- **TIDBD (IDBD for Reinforcement Learning algorithms)**:\n  Kearney, Alex, et al. \"Tidbd: Adapting temporal-difference step-sizes through stochastic meta-descent.\" arXiv preprint arXiv:1804.03334 (2018).\n  \n  Günther, Johannes, et al. \"Meta-learning for Predictive Knowledge Architectures: A Case Study Using TIDBD on a Sensor-rich Robotic Arm.\" Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems. 2019.\n- **AdaGain (Meta-descent for learning stability)**: Jacobsen, Andrew, et al. \"Meta-descent for online, continual prediction.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603772981632}], "openreview_url": "https://openreview.net/forum?id=xFYXLlpIyPQ", "arxiv_id": "2006.16495", "paper_pdf": "papers/xFYXLlpIyPQ.pdf", "paper_pdf_sha256": "b72b7f9b9c1e02c71497713fdd8639343487e8464d113bcd0a61994855dba06b", "paper_pdf_bytes": 749683, "paper_pdf_source": "openreview", "code_url": "https://github.com/shuaiyuan1996/learning-to-learn", "code_repository": "shuaiyuan1996/learning-to-learn", "code_commit": "5c937d6b9e4a893fd3ac4955c44aa46afb81012d", "code_archive": "repos/xFYXLlpIyPQ.zip", "code_archive_sha256": "ade990114b1baea1f9b2eb755dca54d42db606f1219d74c2a3aba42d6c9def5c", "code_archive_bytes": 119455, "code_file_count": 34, "code_extensions": {".py": 30, ".sh": 4}, "github_disk_usage_kb": 63, "github_languages": {"Python": 388958, "Shell": 1469}, "github_archived": false, "github_pushed_at": "2021-06-09T21:30:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/guarantees-for-tuning-the-step-size-using-a"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJxNzgSKvH", "year": 2020, "status": "rejected", "title": "Selective sampling for accelerating  training of deep neural networks", "authors": ["Berry Weinstein", "Shai Fine", "Yacov Hel-Or"], "authorids": ["berry.weinstein@post.idc.ac.il", "shai.fine@idc.ac.il", "toky@idc.ac.il"], "authors_source": "OpenReview API", "abstract": "We present a selective sampling method designed to accelerate the training of deep neural networks. To this end,  we introduce a novel measurement,  the {\\it minimal margin score} (MMS), which measures the minimal amount of displacement an input should take until its predicted classification is switched.   For multi-class linear classification,  the MMS measure is a natural generalization of the margin-based selection criterion, which was thoroughly studied in the binary classification setting.  In addition, the MMS measure provides an interesting insight into the progress of the training process and can be useful for designing and monitoring new training regimes. Empirically we demonstrate a substantial acceleration when training commonly used deep neural network architectures for popular image classification tasks.  The efficiency of our method is compared against the standard training procedures, and against commonly used selective sampling alternatives: Hard negative mining selection, and Entropy-based selection.\nFinally, we demonstrate an additional speedup when we adopt a more aggressive learning-drop regime while using the MMS selective sampling method.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HkgDJXI-5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2169/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "A new approach is proposed to speed up training in deep models.\n\nThe idea is to select sample batches when back propagating the error based on the distance of the prediction foe the sample from the decision boundary. Specifically, we pick points closer to the boundary, i.e., ones that we are less confident about for backpropagation.\n\nExperiments are performed comparing the method with Hard negative sampling (HNM) , entropy-based sample selection as well as regular training. Experiments are performed on Cifar10 and Cifar100 datasets. \nWhy only two datasets, the method is general so there should be more datasets to verify its performance.\n\nThe results on Cifar100 in Fig 5 c seems to show that we cannot reach the training accuracy using the proposed method as compared to the other methods. What is the intuition here as to why it happens? In general though since the main goal is to speed up training I do not see very convincing evidence of this in the limited evaluation which seems to be the main weakness here.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "A new approach is proposed to speed up training in deep models.\n\nThe idea is to select sample batches when back propagating the error based on the distance of the prediction foe the sample from the decision boundary. Specifically, we pick points closer to the boundary, i.e., ones that we are less confident about for backpropagation.\n\nExperiments are performed comparing the method with Hard negative sampling (HNM) , entropy-based sample selection as well as regular training. Experiments are performed on Cifar10 and Cifar100 datasets. \nWhy only two datasets, the method is general so there should be more datasets to verify its performance.\n\nThe results on Cifar100 in Fig 5 c seems to show that we cannot reach the training accuracy using the proposed method as compared to the other methods. What is the intuition here as to why it happens? In general though since the main goal is to speed up training I do not see very convincing evidence of this in the limited evaluation which seems to be the main weakness here."}, "tcdate": 1572066014769}, {"id": "BJeWmNn3KS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2169/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a minimal margin score (MMS) criterion to speed up the training of the deep networks.\n\nI would vote for a clear rejection of this paper. This submission is a clearly unfinished one. The two biggest problems are as follows\n\n1. Lack of a comprehensive discussion on rules for sampling section, please see \"Automated Curriculum Learning for Neural Networks\". Why previous methods are worse than the proposed one is not clear.\n\n2. All experiments are only compared with baseline approaches. In some experiments, the improvements are really marginal (e.g., Figure 2). In these cases, the STD of these curves is not shown, it is not clear whether the improvements are significant or not.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This paper proposes a minimal margin score (MMS) criterion to speed up the training of the deep networks.\n\nI would vote for a clear rejection of this paper. This submission is a clearly unfinished one. The two biggest problems are as follows\n\n1. Lack of a comprehensive discussion on rules for sampling section, please see \"Automated Curriculum Learning for Neural Networks\". Why previous methods are worse than the proposed one is not clear.\n\n2. All experiments are only compared with baseline approaches. In some experiments, the improvements are really marginal (e.g., Figure 2). In these cases, the STD of these curves is not shown, it is not clear whether the improvements are significant or not."}, "tcdate": 1571763225134}, {"id": "rJglmTn5tS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2169/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary of contributions\n\nThis paper aims to accelerate the training of deep networks using a selective sampling. \nThey adapt ideas from active learning (which use some form of uncertainty estimation about the class of the label) to selectively choose samples on which to perform the backward pass. Specifically, they use the minimal margin score (MMS). \nTheir algorithm works by computing the forward pass over a batch of size B (which is much larger than the regular batch of size b), compute the uncertainty measure for each sample, and only perform the backward pass over the b samples with the highest uncertainty. The motivation is that the backward pass is more expensive than the forward pass, and that by only performing this pass on a subset of samples, computations are saved. \n\n\n### Recommendation\n\nReject. The central premise of the paper is unclear, the writing/presentation needs improvement, and the experiments are not convincing. \n\n\n### Detailed comments/improvements: \n\n\nThere is a central premise of the paper that I don't understand: that the forward pass is much cheaper than the backward pass. \nThis is claimed in the intro by referring to charts that hardware manufacturers publish (but there are no references included), but I don't see why this should be the case. \nFor a linear network with weights W, the forward pass is given by the matrix-matrix product (rows of X are minibatch samples):\nY = XW^T\n\nand the backward pass is given by the two matrix-matrix products:\ndL/dX = dL/dY*dY/dX = dL/dY*W\ndL/dW = dL/dY*dY/dW = dL/dY*X^T\n\nSimilarly the two operations in the backward pass for convolutional layers are given by a convolution of the output gradients with the transposed weigtht kernels and the input image respectively. \n\nPoint being, I don't see why the backward pass should be more than 3x more expensive than the forward pass. A simple experiment in PyTorch confirms this: the code snippet pasted at the bottom shows that the backward pass takes only around 2.6x longer than the forward pass.  \n\nfprop: 0.009286s\nbprop: 0.0240s\nbprop/fprop: 2.5893x\n\nIn algorithm 1, it is assumed that b << B. For this to be effective the forward pass would have to be *much* faster than the backward pass for this method to yield an improvement in computation. Can the authors comment on where this justification comes from?\n\nI am unclear on what the purpose of Section 4.1 is. This shows that the MMS of the proposed method is lower than the other two, but this should be completely expected since that is exactly the quantity being minimized. \nThere are also several unsubstantiated claims: \"Lower MMS scores resemble a better...batch of samples\", \"the batches selected by our method provide a higher value for the training procedure vs. the HNM samples.\", \"Evidently, the mean MMS provides a clearer perspective...and usefulness of the selected samples\". What does higher value, usefulness, clearer perspective mean?\n\nMore generally, it is unclear if there is really any improvement in the final performance from using the proposed method.\nIn Figure 2, all methods seem to have similar final performance. \nIn Figure 5, is there a reason why the curve for MMS is cut off? How does its final performance compare to that of the baseline method in red? It looks like the baseline might be better, but it's hard to tell from the figure. \n\nWhy are the experiments with the entropy measure in a seperate section? Please include them along with the other methods in the same plot, i.e merge Figure 2 and Figure 4. \n\nMy suggestions for improving the experimental section are as follows:\n- include all methods together in all the plots/tables\n- repeat experiments multiple times with different seeds to get error bars. Include these both in the learning curves and in the tables. \n- It's hard to see small differences in the learning curves, so including tables as well is important. Include best performance for all the methods in the tables. \n\nFinally, in 2019 CIFAR alone is not longer a sufficient dataset to report experiments on. Please report results on ImageNet as well. \n\nOne of the central premises of the paper is acceleration in terms of compute/time. To make this point, there should also be results in terms of walltime and floating-point operations. Please include these results in the paper.  \n    \n\n\n\n### Code snippet timing forward/backward passes\n\n\nimport torch, torch.nn as nn, time\n\nmodel =\tnn.Sequential(nn.Linear(784, 1000),\n                      nn.ReLU(),\n                      nn.Linear(1000, 1000),\n                      nn.ReLU(),\n                      nn.Linear(1000, 10),\n                      nn.LogSoftmax())\n\ndata = torch.randn(128, 784)\nlabels = torch.ones(128).long()\nt = time.time()\npred = model.forward(data)\nloss = nn.functional.nll_loss(pred, labels)\nfprop_time = time.time() - t\nt = time.time()\nloss.backward()\nbprop_time = time.time() - t\nprint('fprop: {:.4}s'.format(fprop_time))\nprint('bprop: {:.4f}s'.format(bprop_time))\nprint('bprop/fprop: {:.4f}x'.format(bprop_time / fprop_time))\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "### Summary of contributions\n\nThis paper aims to accelerate the training of deep networks using a selective sampling. \nThey adapt ideas from active learning (which use some form of uncertainty estimation about the class of the label) to selectively choose samples on which to perform the backward pass. Specifically, they use the minimal margin score (MMS). \nTheir algorithm works by computing the forward pass over a batch of size B (which is much larger than the regular batch of size b), compute the uncertainty measure for each sample, and only perform the backward pass over the b samples with the highest uncertainty. The motivation is that the backward pass is more expensive than the forward pass, and that by only performing this pass on a subset of samples, computations are saved. \n\n\n### Recommendation\n\nReject. The central premise of the paper is unclear, the writing/presentation needs improvement, and the experiments are not convincing. \n\n\n### Detailed comments/improvements: \n\n\nThere is a central premise of the paper that I don't understand: that the forward pass is much cheaper than the backward pass. \nThis is claimed in the intro by referring to charts that hardware manufacturers publish (but there are no references included), but I don't see why this should be the case. \nFor a linear network with weights W, the forward pass is given by the matrix-matrix product (rows of X are minibatch samples):\nY = XW^T\n\nand the backward pass is given by the two matrix-matrix products:\ndL/dX = dL/dY*dY/dX = dL/dY*W\ndL/dW = dL/dY*dY/dW = dL/dY*X^T\n\nSimilarly the two operations in the backward pass for convolutional layers are given by a convolution of the output gradients with the transposed weigtht kernels and the input image respectively. \n\nPoint being, I don't see why the backward pass should be more than 3x more expensive than the forward pass. A simple experiment in PyTorch confirms this: the code snippet pasted at the bottom shows that the backward pass takes only around 2.6x longer than the forward pass.  \n\nfprop: 0.009286s\nbprop: 0.0240s\nbprop/fprop: 2.5893x\n\nIn algorithm 1, it is assumed that b << B. For this to be effective the forward pass would have to be *much* faster than the backward pass for this method to yield an improvement in computation. Can the authors comment on where this justification comes from?\n\nI am unclear on what the purpose of Section 4.1 is. This shows that the MMS of the proposed method is lower than the other two, but this should be completely expected since that is exactly the quantity being minimized. \nThere are also several unsubstantiated claims: \"Lower MMS scores resemble a better...batch of samples\", \"the batches selected by our method provide a higher value for the training procedure vs. the HNM samples.\", \"Evidently, the mean MMS provides a clearer perspective...and usefulness of the selected samples\". What does higher value, usefulness, clearer perspective mean?\n\nMore generally, it is unclear if there is really any improvement in the final performance from using the proposed method.\nIn Figure 2, all methods seem to have similar final performance. \nIn Figure 5, is there a reason why the curve for MMS is cut off? How does its final performance compare to that of the baseline method in red? It looks like the baseline might be better, but it's hard to tell from the figure. \n\nWhy are the experiments with the entropy measure in a seperate section? Please include them along with the other methods in the same plot, i.e merge Figure 2 and Figure 4. \n\nMy suggestions for improving the experimental section are as follows:\n- include all methods together in all the plots/tables\n- repeat experiments multiple times with different seeds to get error bars. Include these both in the learning curves and in the tables. \n- It's hard to see small differences in the learning curves, so including tables as well is important. Include best performance for all the methods in the tables. \n\nFinally, in 2019 CIFAR alone is not longer a sufficient dataset to report experiments on. Please report results on ImageNet as well. \n\nOne of the central premises of the paper is acceleration in terms of compute/time. To make this point, there should also be results in terms of walltime and floating-point operations. Please include these results in the paper.  \n    \n\n\n\n### Code snippet timing forward/backward passes\n\n\nimport torch, torch.nn as nn, time\n\nmodel =\tnn.Sequential(nn.Linear(784, 1000),\n                      nn.ReLU(),\n                      nn.Linear(1000, 1000),\n                      nn.ReLU(),\n                      nn.Linear(1000, 10),\n                      nn.LogSoftmax())\n\ndata = torch.randn(128, 784)\nlabels = torch.ones(128).long()\nt = time.time()\npred = model.forward(data)\nloss = nn.functional.nll_loss(pred, labels)\nfprop_time = time.time() - t\nt = time.time()\nloss.backward()\nbprop_time = time.time() - t\nprint('fprop: {:.4}s'.format(fprop_time))\nprint('bprop: {:.4f}s'.format(bprop_time))\nprint('bprop/fprop: {:.4f}x'.format(bprop_time / fprop_time))\n"}, "tcdate": 1571634455602}], "openreview_url": "https://openreview.net/forum?id=SJxNzgSKvH", "arxiv_id": "1911.06996", "paper_pdf": "papers/SJxNzgSKvH.pdf", "paper_pdf_sha256": "15cbc67e612a4d0e8b3d0a379a83536db80b510327334510460cd7133e9d88cb", "paper_pdf_bytes": 904527, "paper_pdf_source": "openreview", "code_url": "https://github.com/berryweinst/mms-select", "code_repository": "berryweinst/mms-select", "code_commit": "e1a9dda89c69ad5bec3edd5a0cd6bb37c22753d8", "code_archive": "repos/SJxNzgSKvH.zip", "code_archive_sha256": "94795d0f099ba1aa2b726b2b565aba6f4aa7a95729d6ec903989d2d9f5b22ffb", "code_archive_bytes": 89824, "code_file_count": 47, "code_extensions": {".py": 47}, "github_disk_usage_kb": 81, "github_languages": {"Python": 317969}, "github_archived": false, "github_pushed_at": "2019-10-02T07:26:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/selective-sampling-for-accelerating-training-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1GkMhAqYm", "year": 2019, "status": "rejected", "title": "CoDraw: Collaborative Drawing as a Testbed for Grounded Goal-driven Communication", "authors": ["Nikita Kitaev", "Jin-Hwa Kim", "Xinlei Chen", "Marcus Rohrbach", "Yuandong Tian", "Dhruv Batra", "Devi Parikh"], "authorids": ["kitaev@cs.berkeley.edu", "jnhwkim@gmail.com", "xinleic@fb.com", "maroffm@gmail.com", "yuandong@fb.com", "dbatra@gatech.edu", "parikh@gatech.edu"], "authors_source": "OpenReview API", "abstract": "In this work, we propose a goal-driven collaborative task that contains language, vision, and action in a virtual environment as its core components. Specifically, we develop a Collaborative image-Drawing game between two agents, called CoDraw. Our game is grounded in a virtual world that contains movable clip art objects. The game involves two players: a Teller and a Drawer. The Teller sees an abstract scene containing multiple clip art pieces in a semantically meaningful configuration, while the Drawer tries to reconstruct the scene on an empty canvas using available clip art pieces. The two players communicate via two-way communication using natural language. We collect the CoDraw dataset of ~10K dialogs consisting of ~138K messages exchanged between human agents. We define protocols and metrics to evaluate the effectiveness of learned agents on this testbed, highlighting the need for a novel \"crosstalk\" condition which pairs agents trained independently on disjoint subsets of the training data for evaluation. We present models for our task, including simple but effective baselines and neural network approaches trained using a combination of imitation learning and goal-driven training. All models are benchmarked using both fully automated evaluation and by playing the game with live human agents.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1lU2zbJ6Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1228/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper a new task namely CoDraw is introduced. In CoDraw, there is a teller who describes a scene and a drawer who tries to select clip art component and place them on a canvas to draw the description. The drawing environment contains simple objects and a fixed background scene all in cartoon style. The describing language thus does not have sophisticated components and phrases. A metric based on the presence of the components in the original image and the generated image is coined to compute similarity which is used in learning and evaluation.  Authors mention that in order to gain better performance they needed to train the teller and drawer separately on disjoint subsets of the training data which they call it a cross talk.\n\nComments about the task:\nThe introduced task seems to be very simplistic with very limited number of simple objects. From the explanations and examples the dialogs between the teller and drawer are not natural. As explained the teller will always tell ‘ok’ in some of the scenarios. How is this different with a system that generates clip art images based on a “full description”? Generating clip arts based on descriptions is a task that was introduced in the original clip art paper by Zitnick and Parikh 2013. This paper does not clarify how they are different than monologs of generating scenes based on a description.  \n\nComments about the method:\nI couldn’t find anything particularly novel about the method. The network is a combination of a feed forward model and an LSTM and the learning is done with a combination of imitation learning and REINFORCE. \n\n\nComments about the experimental results:\nIt is hard to evaluate whether the obtained results are satisfying or not. The task is somehow simplistic since there a limited number of clip art objects and the scenes are very abstract which does not have complications of natural images and accordingly the dialogs are also very simplistic. All the baselines are based on nearest neighbors. \n\nComments about presentation:\nThe writing of this paper needs to be improved. The current draft is not coherent and it is hard to navigate between different components of the method and different design choices. Some of the design choices are not experimentally proved to be effective: they are mentioned to be observed to be good design choices. It would be more effective to show the effect of these design choices by some ablation study. \nThere are many details about the method which are not fully explained: what are the details of your imitation learning method? Can you formalize your RL fine-tuning part with the use of some formulations? With the current format, the technical part of the paper is not fully understandable.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Mostly a dataset paper, writing is not coherent, results are not convincing", "review": "In this paper a new task namely CoDraw is introduced. In CoDraw, there is a teller who describes a scene and a drawer who tries to select clip art component and place them on a canvas to draw the description. The drawing environment contains simple objects and a fixed background scene all in cartoon style. The describing language thus does not have sophisticated components and phrases. A metric based on the presence of the components in the original image and the generated image is coined to compute similarity which is used in learning and evaluation.  Authors mention that in order to gain better performance they needed to train the teller and drawer separately on disjoint subsets of the training data which they call it a cross talk.\n\nComments about the task:\nThe introduced task seems to be very simplistic with very limited number of simple objects. From the explanations and examples the dialogs between the teller and drawer are not natural. As explained the teller will always tell ‘ok’ in some of the scenarios. How is this different with a system that generates clip art images based on a “full description”? Generating clip arts based on descriptions is a task that was introduced in the original clip art paper by Zitnick and Parikh 2013. This paper does not clarify how they are different than monologs of generating scenes based on a description.  \n\nComments about the method:\nI couldn’t find anything particularly novel about the method. The network is a combination of a feed forward model and an LSTM and the learning is done with a combination of imitation learning and REINFORCE. \n\n\nComments about the experimental results:\nIt is hard to evaluate whether the obtained results are satisfying or not. The task is somehow simplistic since there a limited number of clip art objects and the scenes are very abstract which does not have complications of natural images and accordingly the dialogs are also very simplistic. All the baselines are based on nearest neighbors. \n\nComments about presentation:\nThe writing of this paper needs to be improved. The current draft is not coherent and it is hard to navigate between different components of the method and different design choices. Some of the design choices are not experimentally proved to be effective: they are mentioned to be observed to be good design choices. It would be more effective to show the effect of these design choices by some ablation study. \nThere are many details about the method which are not fully explained: what are the details of your imitation learning method? Can you formalize your RL fine-tuning part with the use of some formulations? With the current format, the technical part of the paper is not fully understandable.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541505709835}, {"id": "BJlFdvccnm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1228/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a game of collaborative drawing where a teller is\nto communicate a picture to a drawer via natural language.  The picture\nallows only a small number of components and a fixed and limited set\nof detailed variations of such components.\n\nPros:\n\nThe work contributed a dataset where the task has relatively objective\ncriteria for success.  The dataset itself is a valuable contribution\nto the community interested in the subject.   It may be useful for\npurposes beyond those it was designed for.\n\nThe task is interesting and its visual nature allows for easy inspection\nof the reasons for successes or failures.  It provides reasonable grounding\nfor the dialog.  By restricting the scope of variations through the options\nand parameters, some detailed aspects of the conversation could be explored\nwith carefully controlled experiments.\n\nThe authors identified the need for and proposed a \"crosstalk\" protocol\nthat they believe can prevent leakage via common training data and\nthe development of non-language, shared codebooks that defeat the purpose\nof focusing on the natural language dialog.\n\nThe set up allows for pairing of human and human, machine and machine,\nand human and machine for the two roles, which enables comparison to\nhuman performance baselines in several perspectives.\n\nThe figures give useful examples that are of great help to the readers.\n\nCons.:\n\nDespite the restriction of the task context to creating a picture with\nseverely limited components, the scenario of the dialogs still has many\ndetails to keep track of, and many important facets are missing in the\ndescriptions, especially on the data.\n\nThere is no analysis of the errors.  The presentation of\nexperimental results stops at the summary metrics, leaving many\ndoubts on why they are as such.\n\nThe work feels somewhat pre-mature in its exploration of the models\nand the conclusions to warrant publication.  At times it feels like the\nauthors do not understand enough why the algorithms behave as they do.\nHowever if this is considered as a dataset description paper and\nthe right expectation is set in the openings, it may still be acceptable.\n\nThe completed work warrants a longer report when more solid conclusions\ncan be drawn about the model behavior.\n\nThe writing is not organized enough and it takes many back-and-forth rounds\nof checking during reading to find out about certain details that are given\nlong after their first references in other contexts.  Some examples are\nincluded in the followings.\n\nMisc.\n\nSection 3.2, datasets of 9993 dialogs:\nAre they done by humans?   Later it is understood from further descriptions.\nIt is useful to be more explicit at the first mention of this data collection effort.\nThe way they relate to the 10020 scenes is mentioned as \"one per scene\", with a footnote on some being removed.\nDoes it mean that no scene is described by two different people?  Does this\nlimit the usefulness of the data in understanding inter-personal differences?\n\nLater in the descriptions (e.g. 4.1 on baseline methods) the notion of\ntraining set is mentioned, but up to then there is no mentioning of how\ntraining and testing (novel scenes) data are created.\nIt is also not clear what training data include: scenes only?\nDialogs associated with specific scenes?  Drawer actions?\n\nSection 4.1, what is a drawer action?  How many possibilities are there?\nFrom the description of \"rule-based nearest-neighbor drawer\" they seem to be\ncorresponding to \"teller utterance\".\nHowever it is not clear where they come from.  What is an example of a drawer action?\nAre the draw actions represented using the feature vectors discussed in the later sections?\n\nSection 5.1, the need for the crosstalk protocol is an interesting observation,\nhowever based on the description here, a reader may not be able to understand\nthe problem.  What do you mean by \"only limited generalization has taken place\"?  Any examples?\n\nSection 5, near the end: the description of the dataset splits is too cryptic.\nWhat are being split?  How is val used in this context?\n\nAll in all the data preparation and partitioning descriptions need substantial clarification.\n\nSection 6:  Besides reporting averaged similarity scores, it will be useful to report some error analysis.\nWhat are the very good or very bad cases?  Why did that happen?\nAre the bad scenes constructed by humans the same as those bad scenes\nconstructed by machines?  Do humans and machines tend to make different errors?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An artificial task for modeling and evaluation of goal-oriented dialogs", "review": "The paper proposes a game of collaborative drawing where a teller is\nto communicate a picture to a drawer via natural language.  The picture\nallows only a small number of components and a fixed and limited set\nof detailed variations of such components.\n\nPros:\n\nThe work contributed a dataset where the task has relatively objective\ncriteria for success.  The dataset itself is a valuable contribution\nto the community interested in the subject.   It may be useful for\npurposes beyond those it was designed for.\n\nThe task is interesting and its visual nature allows for easy inspection\nof the reasons for successes or failures.  It provides reasonable grounding\nfor the dialog.  By restricting the scope of variations through the options\nand parameters, some detailed aspects of the conversation could be explored\nwith carefully controlled experiments.\n\nThe authors identified the need for and proposed a \"crosstalk\" protocol\nthat they believe can prevent leakage via common training data and\nthe development of non-language, shared codebooks that defeat the purpose\nof focusing on the natural language dialog.\n\nThe set up allows for pairing of human and human, machine and machine,\nand human and machine for the two roles, which enables comparison to\nhuman performance baselines in several perspectives.\n\nThe figures give useful examples that are of great help to the readers.\n\nCons.:\n\nDespite the restriction of the task context to creating a picture with\nseverely limited components, the scenario of the dialogs still has many\ndetails to keep track of, and many important facets are missing in the\ndescriptions, especially on the data.\n\nThere is no analysis of the errors.  The presentation of\nexperimental results stops at the summary metrics, leaving many\ndoubts on why they are as such.\n\nThe work feels somewhat pre-mature in its exploration of the models\nand the conclusions to warrant publication.  At times it feels like the\nauthors do not understand enough why the algorithms behave as they do.\nHowever if this is considered as a dataset description paper and\nthe right expectation is set in the openings, it may still be acceptable.\n\nThe completed work warrants a longer report when more solid conclusions\ncan be drawn about the model behavior.\n\nThe writing is not organized enough and it takes many back-and-forth rounds\nof checking during reading to find out about certain details that are given\nlong after their first references in other contexts.  Some examples are\nincluded in the followings.\n\nMisc.\n\nSection 3.2, datasets of 9993 dialogs:\nAre they done by humans?   Later it is understood from further descriptions.\nIt is useful to be more explicit at the first mention of this data collection effort.\nThe way they relate to the 10020 scenes is mentioned as \"one per scene\", with a footnote on some being removed.\nDoes it mean that no scene is described by two different people?  Does this\nlimit the usefulness of the data in understanding inter-personal differences?\n\nLater in the descriptions (e.g. 4.1 on baseline methods) the notion of\ntraining set is mentioned, but up to then there is no mentioning of how\ntraining and testing (novel scenes) data are created.\nIt is also not clear what training data include: scenes only?\nDialogs associated with specific scenes?  Drawer actions?\n\nSection 4.1, what is a drawer action?  How many possibilities are there?\nFrom the description of \"rule-based nearest-neighbor drawer\" they seem to be\ncorresponding to \"teller utterance\".\nHowever it is not clear where they come from.  What is an example of a drawer action?\nAre the draw actions represented using the feature vectors discussed in the later sections?\n\nSection 5.1, the need for the crosstalk protocol is an interesting observation,\nhowever based on the description here, a reader may not be able to understand\nthe problem.  What do you mean by \"only limited generalization has taken place\"?  Any examples?\n\nSection 5, near the end: the description of the dataset splits is too cryptic.\nWhat are being split?  How is val used in this context?\n\nAll in all the data preparation and partitioning descriptions need substantial clarification.\n\nSection 6:  Besides reporting averaged similarity scores, it will be useful to report some error analysis.\nWhat are the very good or very bad cases?  Why did that happen?\nAre the bad scenes constructed by humans the same as those bad scenes\nconstructed by machines?  Do humans and machines tend to make different errors?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541216112970}, {"id": "BJllz-x5hX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1228/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents CoDraw, a grounded and goal-driven dialogue environment for collaborative drawing. The authors argue convincingly that an interactive and grounded evaluation environment helps us better measure how well NLG/NLU agents actually understand and use their language — rather than evaluating against arbitrary ground-truth examples of what humans say, we can evaluate the objective end-to-end performance of a system in a well-specified nonlinguistic task. They collect a novel dataset in this grounded and goal-driven communication paradigm, define a success metric for the collaborative drawing task, and present models for maximizing that metric.\n\nThis is a very interesting task and the dataset/models are a very useful contribution to the community. I have just a few comments below:\n\n1. Results:\n1a. I’m not sure how impressed I should be by these results. The human–human similarity score is pretty far above those of the best models, even though MTurkers are not optimized (and likely not as motivated as an NN) to solve this task. You might be able to convince me more if you had a stronger baseline — e.g. a bag-of-words Drawer model which works off of the average of the word embeddings in a scripted Teller input. Have you tried baselines like these?\n1b. Please provide variance measures on your results (within model configuration, across scene examples). Are the machine–machine pairs consistently performing well together? Are the humans? Depending on those variance numbers you might also consider doing a statistical test to argue that the auxiliary loss function and and RL fine-tuning offer certain improvement over the Scene2seq base model.\n\n2. Framing: there is a lot of work in collaborative / multi-agent dialogue models which you have missed — see refs below to start. You should link to this literature (mostly in NLP) and contrast your task/model with theirs.\n\nReferences\nVogel & Jurafsky (2010). Learning to follow navigational directions.\nHe et al. (2017). Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings.\nFried et al. (2018). Unified pragmatic models for generating and following instructions.\nFried et al. (2018). Speaker-follower models for vision-and-language navigation.\nLazaridou et al. (2016). The red one!: On learning to refer to things based on their discriminative properties.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Exciting task! Not sure about model results", "review": "This paper presents CoDraw, a grounded and goal-driven dialogue environment for collaborative drawing. The authors argue convincingly that an interactive and grounded evaluation environment helps us better measure how well NLG/NLU agents actually understand and use their language — rather than evaluating against arbitrary ground-truth examples of what humans say, we can evaluate the objective end-to-end performance of a system in a well-specified nonlinguistic task. They collect a novel dataset in this grounded and goal-driven communication paradigm, define a success metric for the collaborative drawing task, and present models for maximizing that metric.\n\nThis is a very interesting task and the dataset/models are a very useful contribution to the community. I have just a few comments below:\n\n1. Results:\n1a. I’m not sure how impressed I should be by these results. The human–human similarity score is pretty far above those of the best models, even though MTurkers are not optimized (and likely not as motivated as an NN) to solve this task. You might be able to convince me more if you had a stronger baseline — e.g. a bag-of-words Drawer model which works off of the average of the word embeddings in a scripted Teller input. Have you tried baselines like these?\n1b. Please provide variance measures on your results (within model configuration, across scene examples). Are the machine–machine pairs consistently performing well together? Are the humans? Depending on those variance numbers you might also consider doing a statistical test to argue that the auxiliary loss function and and RL fine-tuning offer certain improvement over the Scene2seq base model.\n\n2. Framing: there is a lot of work in collaborative / multi-agent dialogue models which you have missed — see refs below to start. You should link to this literature (mostly in NLP) and contrast your task/model with theirs.\n\nReferences\nVogel & Jurafsky (2010). Learning to follow navigational directions.\nHe et al. (2017). Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings.\nFried et al. (2018). Unified pragmatic models for generating and following instructions.\nFried et al. (2018). Speaker-follower models for vision-and-language navigation.\nLazaridou et al. (2016). The red one!: On learning to refer to things based on their discriminative properties.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541173512381}], "openreview_url": "https://openreview.net/forum?id=r1GkMhAqYm", "arxiv_id": "1712.05558", "paper_pdf": "papers/r1GkMhAqYm.pdf", "paper_pdf_sha256": "3f47fca372a160081c23ec1c4c7b2e87024f34a3d54ddd173d195a1736bd75af", "paper_pdf_bytes": 4087502, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/CoDraw", "code_repository": "facebookresearch/CoDraw", "code_commit": "b209770a327f48fdd768724bbcf2783897b0c7fb", "code_archive": "repos/r1GkMhAqYm.zip", "code_archive_sha256": "5a589ea0cbdb9d28895d6d14d76b3641439317cac96fe7c853e834a0ee7b4d55", "code_archive_bytes": 925079, "code_file_count": 9, "code_extensions": {".js": 8, ".py": 1}, "github_disk_usage_kb": 552, "github_languages": {"JavaScript": 50428, "HTML": 6385, "CSS": 1853, "Python": 713}, "github_archived": true, "github_pushed_at": "2019-01-15T21:03:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/codraw-collaborative-drawing-as-a-testbed-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "79uLTCNMpJ", "year": 2026, "status": "rejected", "title": "From Emergence to Control: Probing and Modulating Self-Reflection in Language Models", "authors": ["Xudong Zhu", "Jiachen Jiang", "Mohammad Mahdi Khalili", "Zhihui Zhu"], "authorids": ["~Xudong_Zhu1", "~Jiachen_Jiang1", "~Mohammad_Mahdi_Khalili3", "~Zhihui_Zhu1"], "authors_source": "OpenReview API", "abstract": "Self-reflection---the ability of a large language model (LLM) to revisit, evaluate, and revise its own reasoning---has recently emerged as a powerful behavior enabled by reinforcement learning with verifiable rewards (RLVR). While self-reflection correlates with improved reasoning accuracy, its origin and underlying mechanisms remain poorly understood. In this work, we first show that self-reflection is not exclusive to RLVR fine-tuned models: it already emerges, albeit rarely, in pretrained models. To probe this latent ability, we introduce Reflection-Inducing Probing, a method that injects reflection-triggering reasoning traces from fine-tuned models into pretrained models. This intervention raises self-reflection frequency of Qwen2.5 from 0.6% to 18.6%, revealing a hidden capacity for reflection. Moreover, our analysis of internal representations shows that both pretrained and fine-tuned models maintain hidden states that distinctly separate self-reflective from non-reflective contexts. Leveraging this observation, we then construct a self-reflection vector, a direction in activation space associated with self-reflective reasoning. By manipulating this vector, we enable bidirectional control over the self-reflective behavior for both pretrained and fine-tuned models. Experiments across multiple reasoning benchmarks show that enhancing these vectors improves reasoning performance by up to 12%, while suppressing them reduces computational cost, providing a flexible mechanism to navigate the trade-off between reasoning quality and efficiency without requiring additional training. Our findings further our understanding of self-reflection and support a growing body of work showing that understanding model internals can enable precise behavioral control.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "N5D70Gqt09", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21778/Reviewer_hCwu"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper investigates whether large language models develop self-reflection abilities during pretraining and whether such behavior can be controlled through linear interventions in activation space. The authors introduce a method called reflection-inducing probing, where pre-reflection reasoning traces from a fine-tuned model are inserted into a pretrained model’s input to test whether it spontaneously emits reflection tokens (e.g., “Wait”). They find that this probing substantially increases self-reflection frequency and that hidden states preceding reflective tokens form a distinct cluster separable from non-reflective contexts. By taking the difference of means between these states, they derive a self-reflection vector that, when added or subtracted at inference, can respectively enhance or suppress reflective behavior and modestly improve task accuracy. Results are demonstrated on reasoning benchmarks such as GPQA and MATH500 using DeepSeek-R1 models, suggesting that reflection signals may exist in pretrained models and can be linearly controlled at inference time.", "review_text": "The paper investigates whether large language models develop self-reflection abilities during pretraining and whether such behavior can be controlled through linear interventions in activation space. The authors introduce a method called reflection-inducing probing, where pre-reflection reasoning traces from a fine-tuned model are inserted into a pretrained model’s input to test whether it spontaneously emits reflection tokens (e.g., “Wait”). They find that this probing substantially increases self-reflection frequency and that hidden states preceding reflective tokens form a distinct cluster separable from non-reflective contexts. By taking the difference of means between these states, they derive a self-reflection vector that, when added or subtracted at inference, can respectively enhance or suppress reflective behavior and modestly improve task accuracy. Results are demonstrated on reasoning benchmarks such as GPQA and MATH500 using DeepSeek-R1 models, suggesting that reflection signals may exist in pretrained models and can be linearly controlled at inference time.", "strengths": "S1. The paper presents an interesting and clear visualization of token-level separation using UMAP, highlighting distinct patterns between reflective and non-reflective representations.\nS2. The study is guided by good motivation, making the problem setup and research goal easy to understand.", "weaknesses": "W1. Line 363 overstates the generality of the token-level correlation. Showing that one model (DeepSeek) emits “wait”-style tokens associated with reflection does not imply that other models share the same surface-form pattern. Test the setup on additional model families (e.g., LLaMA and Qwen) or explicitly acknowledge that this observation is model-specific improve quality of paper. Moreover, prior work (e.g., [2]) suggests that explicit markers like “Wait” or “Hmm” can increase generation length without being required for quality; relying solely on such markers to detect reflection risks both false positives and brittle conclusions. A complementary detection method that does not only depend only on such tokens would substantially strengthen the paper.\n\nW2. As discussed in [3], Qwen models are known to have some issues in their pretraining corpus. Since the analysis is conducted only on the Qwen models family, it is difficult to disentangle reflection effects from artifacts arising from training data.\n \nW3. The paper leans on 2D UMAP plots to argue separability in high-dimensional hidden space without required null/quantitative baselines. UMAP can create attractive clusters even from noise. Conclusions drawn from it should therefore be much weaker than those presented in the work. In [1], the authors use PCA, which is more interpretable and can reveal linear relationships in the representations. Applying PCA in this setup could strengthen the paper by providing clearer intuition about the structure of the learned direction and supporting the proposed method with more interpretable evidence.\n\n[1] Steering Llama 2 via Contrastive Activation Addition, Rimsky et al. \n[2] Wait, We Don’t Need to “Wait”! Removing Thinking Tokens Improves Reasoning Efficiency, Wang et. al.\n[3] Spurious Rewards: Rethinking Training Signals in RLVR, Shao et. al", "questions": "Q1. In Table 1 authors report results for intervention with the self-reflection vector. Could the authors please report the corresponding suppression results (vector subtraction) for the Qwen and LLaMA model families as well?\n\nQ2. Why are \"But\", \"Wait\" tokens is not part of reflection? \n\nQ3. Have the authors attempted to interpret the discovered self-reflection direction using mechanistic-interpretability approaches (for example, circuit analysis, neuron/subspace ablation, or dictionary-learning methods)? If so, please summarize the key findings; if not, could the authors comment on the feasibility of such analyses and whether they plan to pursue them?\n\nQ4. In Table 1 the paper reports generation length and Pass@1. It would also be informative to evaluate how applying the self-reflection vector affects overall generation quality in a distributional sense (for example, KL divergence between the intervened model’s output distribution and that of the SFT or RL-trained model, or other distributional quality metrics). Could the authors add such analyses or explain why they were not included?\n\nQ5. In the ablation over the scaling parameter $\\alpha$, Pass@1 appears relatively noisy. Could the authors provide intuition or an analysis for this variability? Additionally, would the authors consider evaluating the method on more recent, diverse benchmarks (for example AIME 2024/2025) to demonstrate robustness?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates whether large language models develop self-reflection abilities during pretraining and whether such behavior can be controlled through linear interventions in activation space. The authors introduce a method called reflection-inducing probing, where pre-reflection reasoning traces from a fine-tuned model are inserted into a pretrained model’s input to test whether it spontaneously emits reflection tokens (e.g., “Wait”). They find that this probing substantially increases self-reflection frequency and that hidden states preceding reflective tokens form a distinct cluster separable from non-reflective contexts. By taking the difference of means between these states, they derive a self-reflection vector that, when added or subtracted at inference, can respectively enhance or suppress reflective behavior and modestly improve task accuracy. Results are demonstrated on reasoning benchmarks such as GPQA and MATH500 using DeepSeek-R1 models, suggesting that reflection signals may exist in pretrained models and can be linearly controlled at inference time.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "S1. The paper presents an interesting and clear visualization of token-level separation using UMAP, highlighting distinct patterns between reflective and non-reflective representations.\nS2. The study is guided by good motivation, making the problem setup and research goal easy to understand.", "weaknesses": "W1. Line 363 overstates the generality of the token-level correlation. Showing that one model (DeepSeek) emits “wait”-style tokens associated with reflection does not imply that other models share the same surface-form pattern. Test the setup on additional model families (e.g., LLaMA and Qwen) or explicitly acknowledge that this observation is model-specific improve quality of paper. Moreover, prior work (e.g., [2]) suggests that explicit markers like “Wait” or “Hmm” can increase generation length without being required for quality; relying solely on such markers to detect reflection risks both false positives and brittle conclusions. A complementary detection method that does not only depend only on such tokens would substantially strengthen the paper.\n\nW2. As discussed in [3], Qwen models are known to have some issues in their pretraining corpus. Since the analysis is conducted only on the Qwen models family, it is difficult to disentangle reflection effects from artifacts arising from training data.\n \nW3. The paper leans on 2D UMAP plots to argue separability in high-dimensional hidden space without required null/quantitative baselines. UMAP can create attractive clusters even from noise. Conclusions drawn from it should therefore be much weaker than those presented in the work. In [1], the authors use PCA, which is more interpretable and can reveal linear relationships in the representations. Applying PCA in this setup could strengthen the paper by providing clearer intuition about the structure of the learned direction and supporting the proposed method with more interpretable evidence.\n\n[1] Steering Llama 2 via Contrastive Activation Addition, Rimsky et al. \n[2] Wait, We Don’t Need to “Wait”! Removing Thinking Tokens Improves Reasoning Efficiency, Wang et. al.\n[3] Spurious Rewards: Rethinking Training Signals in RLVR, Shao et. al", "questions": "Q1. In Table 1 authors report results for intervention with the self-reflection vector. Could the authors please report the corresponding suppression results (vector subtraction) for the Qwen and LLaMA model families as well?\n\nQ2. Why are \"But\", \"Wait\" tokens is not part of reflection? \n\nQ3. Have the authors attempted to interpret the discovered self-reflection direction using mechanistic-interpretability approaches (for example, circuit analysis, neuron/subspace ablation, or dictionary-learning methods)? If so, please summarize the key findings; if not, could the authors comment on the feasibility of such analyses and whether they plan to pursue them?\n\nQ4. In Table 1 the paper reports generation length and Pass@1. It would also be informative to evaluate how applying the self-reflection vector affects overall generation quality in a distributional sense (for example, KL divergence between the intervened model’s output distribution and that of the SFT or RL-trained model, or other distributional quality metrics). Could the authors add such analyses or explain why they were not included?\n\nQ5. In the ablation over the scaling parameter $\\alpha$, Pass@1 appears relatively noisy. Could the authors provide intuition or an analysis for this variability? Additionally, would the authors consider evaluating the method on more recent, diverse benchmarks (for example AIME 2024/2025) to demonstrate robustness?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761998551886}, {"id": "1BSsreCaUv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21778/Reviewer_zuym"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper suggests to search for a vector in representation space that induces self-reflection of the model. This vector can then be manipulated by intervention in order to steer the model.", "review_text": "This paper suggests to search for a vector in representation space that induces self-reflection of the model. This vector can then be manipulated by intervention in order to steer the model.", "strengths": "I really like the main idea of the paper. It is (to my knowledge) innovative, intuitive and actionable.\nThe experimental design is well thought out, and the paper is overall easy to read.", "weaknesses": "I have some issues with the focus on the single token \"wait, \" as an indicator for self-reflection. In fact, this token could also appear in other contexts. Given the tiny amount of self-reflection in the base model, this could actually have a substantial impact on a key finding of the paper, i.e., the emergence of self-reflection. I would urge the authors not only extend the set of respective tokens but also quantify, which sets of tokens are used by the base model accordingly.\nFrom the appendix: \"In our analysis of DeepSeek-R1 outputs, \"wait\" accounted for approximately 97.2% of all\ndetected reflection instances.\" -> the keywords might be different ones for other models, though.\n\nA statistical analysis of the significance of the results is missing. Given that the improvement, e.g., for the DeepSeek models are very moderate, I would see this as a necessity.\nAdditionally, the set of test datasets should in my opinion be extended. Why is \"SR Suppressed\" missing in the last two\n\nThe limitation section is weak. E.g., there is a huge focus on very small models in the current study.\n\nMinor suggestion:\nFigure 3 is not useful in a b/w print out, consider color changes.", "questions": "I think, question/issues for discussion are straightforward to derive from the weaknesses.\n\nOn top:\nWouldn't *identifying* self-reflecting behavior be way more accurately and reliable performed by using a (small) LLM instead of keywords? Why keywords?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper suggests to search for a vector in representation space that induces self-reflection of the model. This vector can then be manipulated by intervention in order to steer the model.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "I really like the main idea of the paper. It is (to my knowledge) innovative, intuitive and actionable.\nThe experimental design is well thought out, and the paper is overall easy to read.", "weaknesses": "I have some issues with the focus on the single token \"wait, \" as an indicator for self-reflection. In fact, this token could also appear in other contexts. Given the tiny amount of self-reflection in the base model, this could actually have a substantial impact on a key finding of the paper, i.e., the emergence of self-reflection. I would urge the authors not only extend the set of respective tokens but also quantify, which sets of tokens are used by the base model accordingly.\nFrom the appendix: \"In our analysis of DeepSeek-R1 outputs, \"wait\" accounted for approximately 97.2% of all\ndetected reflection instances.\" -> the keywords might be different ones for other models, though.\n\nA statistical analysis of the significance of the results is missing. Given that the improvement, e.g., for the DeepSeek models are very moderate, I would see this as a necessity.\nAdditionally, the set of test datasets should in my opinion be extended. Why is \"SR Suppressed\" missing in the last two\n\nThe limitation section is weak. E.g., there is a huge focus on very small models in the current study.\n\nMinor suggestion:\nFigure 3 is not useful in a b/w print out, consider color changes.", "questions": "I think, question/issues for discussion are straightforward to derive from the weaknesses.\n\nOn top:\nWouldn't *identifying* self-reflecting behavior be way more accurately and reliable performed by using a (small) LLM instead of keywords? Why keywords?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761993081675}, {"id": "dZGrYRjtlX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21778/Reviewer_w6Jt"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "Summary:\nThe paper focusses on answering two questions, first, is self-reflection a novel behavior induced by RLVR, or does it already emerge during pretraining? And second, can we control self-reflection in LLMs to balance performance and computational efficiency? For this the authors provide empirical results to analyze the hidden representations of the models to see if these self-reflection capabilities can be modulated.\n\nSelf-reflection emerging during pre-training, which is one of the two questions that the authors try to answer, is not a novel finding, and has been studied extensively during prior works. The claim that self-reflection capabilities can be modulated is an interesting one, and requires more deeper analysis.", "review_text": "Summary:\nThe paper focusses on answering two questions, first, is self-reflection a novel behavior induced by RLVR, or does it already emerge during pretraining? And second, can we control self-reflection in LLMs to balance performance and computational efficiency? For this the authors provide empirical results to analyze the hidden representations of the models to see if these self-reflection capabilities can be modulated.\n\nSelf-reflection emerging during pre-training, which is one of the two questions that the authors try to answer, is not a novel finding, and has been studied extensively during prior works. The claim that self-reflection capabilities can be modulated is an interesting one, and requires more deeper analysis.", "strengths": "Strengths:\n- The analysis of model states and visualizations on both models showing clear separation between self-reflection and nonself-reflection states is very interesting.\n- The paper presents a good ablation study on the effect of their linear intervention method by varying the parameter alpha.\n- It is interesting to see that negative alpha does reduce the size of the response, and increasing alpha increases it. I don’t understand the reason behind the author's claim that higher output length means deeper self-reflective reasoning. However, increasing alpha increases the performance on reasoning benchmarks, which is a more intuitive explanation.\n- SR suppression is helpful to make the models efficient, while minimal performance drop.\n- The authors perform evaluations on varied model sizes (1.5B-13B), which is great. They also look at benchmarks across different domain, like MATH500 and GPQA, and  see if the performance is correlated with changing alpha, which is interesting.\n- The paper also shows ablations towards the end comparing different methods for generating steering vectors.", "weaknesses": "Scope for improvement:\n- The claim that an LLMs ability to reflect upon its responses during pre-training with “Wait” or similar words is not a new finding, and has been studied extensively in prior literature. Therefore, I think the paper has limited novelty.\n- The authors should shed some light on why they used difference-in-means to construct the self-reflection vector. I am curious what other metrics they evaluated before settling on using the mean. I see the ablations on comparing difference in means with PCA and Contrastive PCA, but I am curious if the authors tried different metrics apart from the mean.\n- I am curious why the authors claim that a higher response length means deeper self-reflective reasoning by the model.", "questions": "Points 2 and 3, in Weaknesses section. Also\n- Did the authors experiment with variations of equation 6 to update the hidden state vector? Instead of using \"projected weighted steering,\" which is ĥ^(ℓ) = h^(ℓ) + α v^(ℓ) ⟨h^(ℓ), v^(ℓ)⟩, why not use ĥ^(ℓ) = h^(ℓ) + α v^(ℓ) without taking the projection?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Summary:\nThe paper focusses on answering two questions, first, is self-reflection a novel behavior induced by RLVR, or does it already emerge during pretraining? And second, can we control self-reflection in LLMs to balance performance and computational efficiency? For this the authors provide empirical results to analyze the hidden representations of the models to see if these self-reflection capabilities can be modulated.\n\nSelf-reflection emerging during pre-training, which is one of the two questions that the authors try to answer, is not a novel finding, and has been studied extensively during prior works. The claim that self-reflection capabilities can be modulated is an interesting one, and requires more deeper analysis.", "soundness": 4, "presentation": 3, "contribution": 2, "strengths": "Strengths:\n- The analysis of model states and visualizations on both models showing clear separation between self-reflection and nonself-reflection states is very interesting.\n- The paper presents a good ablation study on the effect of their linear intervention method by varying the parameter alpha.\n- It is interesting to see that negative alpha does reduce the size of the response, and increasing alpha increases it. I don’t understand the reason behind the author's claim that higher output length means deeper self-reflective reasoning. However, increasing alpha increases the performance on reasoning benchmarks, which is a more intuitive explanation.\n- SR suppression is helpful to make the models efficient, while minimal performance drop.\n- The authors perform evaluations on varied model sizes (1.5B-13B), which is great. They also look at benchmarks across different domain, like MATH500 and GPQA, and  see if the performance is correlated with changing alpha, which is interesting.\n- The paper also shows ablations towards the end comparing different methods for generating steering vectors.", "weaknesses": "Scope for improvement:\n- The claim that an LLMs ability to reflect upon its responses during pre-training with “Wait” or similar words is not a new finding, and has been studied extensively in prior literature. Therefore, I think the paper has limited novelty.\n- The authors should shed some light on why they used difference-in-means to construct the self-reflection vector. I am curious what other metrics they evaluated before settling on using the mean. I see the ablations on comparing difference in means with PCA and Contrastive PCA, but I am curious if the authors tried different metrics apart from the mean.\n- I am curious why the authors claim that a higher response length means deeper self-reflective reasoning by the model.", "questions": "Points 2 and 3, in Weaknesses section. Also\n- Did the authors experiment with variations of equation 6 to update the hidden state vector? Instead of using \"projected weighted steering,\" which is ĥ^(ℓ) = h^(ℓ) + α v^(ℓ) ⟨h^(ℓ), v^(ℓ)⟩, why not use ĥ^(ℓ) = h^(ℓ) + α v^(ℓ) without taking the projection?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761952995843}, {"id": "mrU4opevDQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21778/Reviewer_hFui"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper studies whether self-reflection property in reasoning models, emerging during RLVR training, already contains in the base model and whether it could be controlled via simple linear interventions (by adding steering vectors). It argues that such behaviour is already contained in the base model, and validate this by performing reflection-inducing probing experiment. By constructing the steering vector that induces self-reflection, the effect of steering is validated by improvements across AIME24, MATH500 and GPQA Diamond. Another notable observation is that hidden states that correspond to reflection-inducing inputs are separable from non-reflection-inducing hidden states, which is shown via UMAP plots.", "review_text": "This paper studies whether self-reflection property in reasoning models, emerging during RLVR training, already contains in the base model and whether it could be controlled via simple linear interventions (by adding steering vectors). It argues that such behaviour is already contained in the base model, and validate this by performing reflection-inducing probing experiment. By constructing the steering vector that induces self-reflection, the effect of steering is validated by improvements across AIME24, MATH500 and GPQA Diamond. Another notable observation is that hidden states that correspond to reflection-inducing inputs are separable from non-reflection-inducing hidden states, which is shown via UMAP plots.", "strengths": "S1. The motivation is very clear and well-presented.\n\nS2. A nice appendix is provided with more context, more results with ablations and details of experiments.\n\nS3. Results with separability of reflection-inducing hidden states presented in Figures 3, 8 and 9, and the fact that self-reflection could be modulated by a linear intervention (steering vector), are interesting and might be useful for further research.", "weaknesses": "W1. It is well-known that Qwen family behaves differently during RLVR training [1], and this primarily attributed to the structure of pre-training dataset. So the conclusion in section 3 that self-reflection emerges during pre-training might not generalize to other families or even models, and might be explained by the fact that pre-training dataset contains examples with self-reflection, making it almost trivial.\n\nW2. Experimental setup seems to be unnatural. Authors define self-reflection via collection of words, and identify from a single model (Qwen2.5-Math-7B distilled) and single task (MATH500) that from all these tokens \"Wait\" is the most frequent one, and therefore self-reflection is defined via the presence of \"Wait\" token, which is notable simplification of this cognitive process as acknowledged by the authors, but perhaps suitable for preliminary investigation; the steering vector is therefore boost its probability. Given that budget forcing [2] seems to reduce performance on Llama 3.1 8B Instruct, it might be the case that \"Wait\" token is not a good indicator of self-reflection in general and more thorough study is required - for example, conducting the same experiment as in section 3, but inserting the pre-reflection prompt from thinking Qwen into Llama 3.1 8B instead of base Qwen - this would additionally validate that self-reflection presence is indeed could be detected via this method. \n\nW3. The generalizability of results is somewhat questionable. To detect the signals of self-reflection, authors use a single model and a single mathematical task, then apply it to other models and even families; further experiments are conducted on two mathematical benchmarks and one scientific reasoning benchmark, while many other general reasoning domains also remain omitted (and due to the results of [1], it could invalidate main conclusions).\n\nW4. There is not so much novelty and a lack of practical conclusions. The effect of inserting \"Wait\" already has been shown in [2], while the steering vectors methodology are already standard in the industry with many methods studied [3], even in reasoning [4] where Wait-related features were also steered. The insights into how models work are limited by the finding that self-reflection emerges during pre-training, which is affected by reasons described in W1-W3. Please see S3.\n\nW5. DeepSeek-R1-Distill-Qwen-1.5B is in fact two-staged-trained model (obtained via SFT distillation from Qwen2.5-Math-1.5B, which is also obtained from Qwen2.5-1.5B base pretrained model). SFT is known to change the distribution of tokens more significantly than GRPO (due to KL divergence penalty in the latter), so the setup in the paper is quite different from comparing pretrained and RLVR trained models, which also affect the conclusions. Further work might be extended by performing more direct comparisons.", "questions": "Q1. Do you think that self-reflection is the only emergent property, or RLVR induces any other important patterns or skills that are not present in the base model? Have you tried to analyze those?\n\nQ2. What is the cosine similarity between steering vectors on different layers and the unembedding of token Wait? What tokens become more frequently generated after steering? Have you tried to analyze what this vector introduces to the behavior of the model?\n\nQ3. The separability is clearly shown in the UMAP plots, but can you perform experiment to test linear separability? You could build the linear classifier (logistic regression) and nonlinear classifier (e.g. MLP) and compare their overall quality of classification. This will reveal more information about geometrical properties of those states. \n\n[1] Spurious Rewards: Rethinking Training Signals in RLVR, Shao et al., 2025\n[2] s1: Simple Test-Time Scaling, Muennighoff et al., 2025\n[3] AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders, Wu et al., 2025\n[4] I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders, Galichin et al., 2025", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies whether self-reflection property in reasoning models, emerging during RLVR training, already contains in the base model and whether it could be controlled via simple linear interventions (by adding steering vectors). It argues that such behaviour is already contained in the base model, and validate this by performing reflection-inducing probing experiment. By constructing the steering vector that induces self-reflection, the effect of steering is validated by improvements across AIME24, MATH500 and GPQA Diamond. Another notable observation is that hidden states that correspond to reflection-inducing inputs are separable from non-reflection-inducing hidden states, which is shown via UMAP plots.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "S1. The motivation is very clear and well-presented.\n\nS2. A nice appendix is provided with more context, more results with ablations and details of experiments.\n\nS3. Results with separability of reflection-inducing hidden states presented in Figures 3, 8 and 9, and the fact that self-reflection could be modulated by a linear intervention (steering vector), are interesting and might be useful for further research.", "weaknesses": "W1. It is well-known that Qwen family behaves differently during RLVR training [1], and this primarily attributed to the structure of pre-training dataset. So the conclusion in section 3 that self-reflection emerges during pre-training might not generalize to other families or even models, and might be explained by the fact that pre-training dataset contains examples with self-reflection, making it almost trivial.\n\nW2. Experimental setup seems to be unnatural. Authors define self-reflection via collection of words, and identify from a single model (Qwen2.5-Math-7B distilled) and single task (MATH500) that from all these tokens \"Wait\" is the most frequent one, and therefore self-reflection is defined via the presence of \"Wait\" token, which is notable simplification of this cognitive process as acknowledged by the authors, but perhaps suitable for preliminary investigation; the steering vector is therefore boost its probability. Given that budget forcing [2] seems to reduce performance on Llama 3.1 8B Instruct, it might be the case that \"Wait\" token is not a good indicator of self-reflection in general and more thorough study is required - for example, conducting the same experiment as in section 3, but inserting the pre-reflection prompt from thinking Qwen into Llama 3.1 8B instead of base Qwen - this would additionally validate that self-reflection presence is indeed could be detected via this method. \n\nW3. The generalizability of results is somewhat questionable. To detect the signals of self-reflection, authors use a single model and a single mathematical task, then apply it to other models and even families; further experiments are conducted on two mathematical benchmarks and one scientific reasoning benchmark, while many other general reasoning domains also remain omitted (and due to the results of [1], it could invalidate main conclusions).\n\nW4. There is not so much novelty and a lack of practical conclusions. The effect of inserting \"Wait\" already has been shown in [2], while the steering vectors methodology are already standard in the industry with many methods studied [3], even in reasoning [4] where Wait-related features were also steered. The insights into how models work are limited by the finding that self-reflection emerges during pre-training, which is affected by reasons described in W1-W3. Please see S3.\n\nW5. DeepSeek-R1-Distill-Qwen-1.5B is in fact two-staged-trained model (obtained via SFT distillation from Qwen2.5-Math-1.5B, which is also obtained from Qwen2.5-1.5B base pretrained model). SFT is known to change the distribution of tokens more significantly than GRPO (due to KL divergence penalty in the latter), so the setup in the paper is quite different from comparing pretrained and RLVR trained models, which also affect the conclusions. Further work might be extended by performing more direct comparisons.", "questions": "Q1. Do you think that self-reflection is the only emergent property, or RLVR induces any other important patterns or skills that are not present in the base model? Have you tried to analyze those?\n\nQ2. What is the cosine similarity between steering vectors on different layers and the unembedding of token Wait? What tokens become more frequently generated after steering? Have you tried to analyze what this vector introduces to the behavior of the model?\n\nQ3. The separability is clearly shown in the UMAP plots, but can you perform experiment to test linear separability? You could build the linear classifier (logistic regression) and nonlinear classifier (e.g. MLP) and compare their overall quality of classification. This will reveal more information about geometrical properties of those states. \n\n[1] Spurious Rewards: Rethinking Training Signals in RLVR, Shao et al., 2025\n[2] s1: Simple Test-Time Scaling, Muennighoff et al., 2025\n[3] AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders, Wu et al., 2025\n[4] I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders, Galichin et al., 2025", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761917177208}], "openreview_url": "https://openreview.net/forum?id=79uLTCNMpJ", "arxiv_id": "2506.12217", "paper_pdf": "papers/79uLTCNMpJ.pdf", "paper_pdf_sha256": "5b7659603bea5a77158d40fa02b73aac414fff609befeddc301e1cbc19132658", "paper_pdf_bytes": 3683241, "paper_pdf_source": "openreview", "code_url": "https://github.com/xzAscC/ProbingReflection", "code_repository": "xzAscC/ProbingReflection", "code_commit": "33991f446ee887f959f7f22be330b03d07c2a5ca", "code_archive": "repos/79uLTCNMpJ.zip", "code_archive_sha256": "115836012fe713a2d0a534b4749d3e47068b118721a93e47bc71d53f34740f44", "code_archive_bytes": 293907, "code_file_count": 37, "code_extensions": {".py": 29, ".sh": 8}, "github_disk_usage_kb": 358, "github_languages": {"Python": 265817, "Shell": 11872}, "github_archived": false, "github_pushed_at": "2026-07-16T22:15:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/from-emergence-to-control-probing-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "E9GakjQype", "year": 2025, "status": "rejected", "title": "AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs", "authors": ["Anselm Paulus", "Arman Zharmagambetov", "Chuan Guo", "Brandon Amos", "Yuandong Tian"], "authorids": ["~Anselm_Paulus1", "~Arman_Zharmagambetov1", "~Chuan_Guo1", "~Brandon_Amos1", "~Yuandong_Tian1"], "authors_source": "OpenReview API", "abstract": "While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain `jailbreaking attacks` that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming. \nOn the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the TargetLLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the `AdvPrompter`, to generate human-readable adversarial prompts in seconds, $\\sim800\\times$ faster than existing optimization-based approaches.\nWe train the AdvPrompter using a novel algorithm that `does not require gradients` of the TargetLLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by AdvPrompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "p84p9Rl8Ef", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4771/Reviewer_eFRA"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper studies how efficiently generate transferrable and interpretable surfix for jailbreak. Unlike previous white-box attack methods that adopt search-based optimization, the authors propose a learning-based method, i.e. finetuning a LLM to generate adversarial prompts using annotated harmful QA data. A major benefit of this approach is inference-time efficiency. To train the LLM, the authors propose an alternated optimization by first searching for the best surfix that prompts the target LLM to answer harmful queries, and subsequently use it to finetune the Prompter LLM. Experiments are mainly conducted by comparing with some white-box attacks, and both direct search and transfer settings are considered. The result show that a major inference efficiency boost, with mixed results in terms of ASR. The proposed method also exhibits stronger transferability to close-source proprietary models than baselines.", "review_text": "This paper studies how efficiently generate transferrable and interpretable surfix for jailbreak. Unlike previous white-box attack methods that adopt search-based optimization, the authors propose a learning-based method, i.e. finetuning a LLM to generate adversarial prompts using annotated harmful QA data. A major benefit of this approach is inference-time efficiency. To train the LLM, the authors propose an alternated optimization by first searching for the best surfix that prompts the target LLM to answer harmful queries, and subsequently use it to finetune the Prompter LLM. Experiments are mainly conducted by comparing with some white-box attacks, and both direct search and transfer settings are considered. The result show that a major inference efficiency boost, with mixed results in terms of ASR. The proposed method also exhibits stronger transferability to close-source proprietary models than baselines.", "strengths": "1. Exploring the leanring-based paradigm for generating adversarial prompts is novel and relevant, to the best of the reviewer’s knowledge.\n2. Compared with search-based paradigm, methods along the learning-based paradigm, like this one, naturally enjoy the benefit of inference-time efficiency.\n3. Experiment results suggest that the proposed method surpasses previous method in terms of transferrability to black-box models, which is arguably a more practical scenario than white-box attack.", "weaknesses": "1. [major] It seems that, intuitively, the solution of equation 1 depends on the targetLLM. i.e. the optimal surfix that triggers a LLM to output the target response (e.g. “Sure, here are detailed xxx”) might be different. I’d imagine that this would hurt the transferrability in theory. It does appear that the transferability of AdvPrompter is at least better than early white-box attackers, but it might be due to poor transferability of white-box attackers in the first place.\n2. [major] Following 1, I have some doubt about the practicality of jailbreak methods that requires transferrability in general. I would suggest comparing with SOTA blackbox methods on Figure 2. While I acknowledge that it is debatable whether such comparison is academically fair, the general practice usually guides us towards using whichever that is most effective. But I am happy to hear the author’s rebuttal and take them into consideration.\n3. [minor] Learning-based paradigm, compared with search-based ones, incurs high training cost.", "questions": "1. Necessity of alternated update: Is it necessary to put the AdvPrompter in the loop of suffix generation (Figure 1 bottom right)? My understanding is that, the purpose of including it is to generate topk candidate tokens for the suffix. I am curious whether the author has tried to use a separate LLM to do this? The upside is that the dataset used to finetune the AdvPrompter can generated offline (without alternatively updating AdvPrompter); The downside is that the generated most likely tokens will not be adaptive.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies how efficiently generate transferrable and interpretable surfix for jailbreak. Unlike previous white-box attack methods that adopt search-based optimization, the authors propose a learning-based method, i.e. finetuning a LLM to generate adversarial prompts using annotated harmful QA data. A major benefit of this approach is inference-time efficiency. To train the LLM, the authors propose an alternated optimization by first searching for the best surfix that prompts the target LLM to answer harmful queries, and subsequently use it to finetune the Prompter LLM. Experiments are mainly conducted by comparing with some white-box attacks, and both direct search and transfer settings are considered. The result show that a major inference efficiency boost, with mixed results in terms of ASR. The proposed method also exhibits stronger transferability to close-source proprietary models than baselines.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Exploring the leanring-based paradigm for generating adversarial prompts is novel and relevant, to the best of the reviewer’s knowledge.\n2. Compared with search-based paradigm, methods along the learning-based paradigm, like this one, naturally enjoy the benefit of inference-time efficiency.\n3. Experiment results suggest that the proposed method surpasses previous method in terms of transferrability to black-box models, which is arguably a more practical scenario than white-box attack.", "weaknesses": "1. [major] It seems that, intuitively, the solution of equation 1 depends on the targetLLM. i.e. the optimal surfix that triggers a LLM to output the target response (e.g. “Sure, here are detailed xxx”) might be different. I’d imagine that this would hurt the transferrability in theory. It does appear that the transferability of AdvPrompter is at least better than early white-box attackers, but it might be due to poor transferability of white-box attackers in the first place.\n2. [major] Following 1, I have some doubt about the practicality of jailbreak methods that requires transferrability in general. I would suggest comparing with SOTA blackbox methods on Figure 2. While I acknowledge that it is debatable whether such comparison is academically fair, the general practice usually guides us towards using whichever that is most effective. But I am happy to hear the author’s rebuttal and take them into consideration.\n3. [minor] Learning-based paradigm, compared with search-based ones, incurs high training cost.", "questions": "1. Necessity of alternated update: Is it necessary to put the AdvPrompter in the loop of suffix generation (Figure 1 bottom right)? My understanding is that, the purpose of including it is to generate topk candidate tokens for the suffix. I am curious whether the author has tried to use a separate LLM to do this? The upside is that the dataset used to finetune the AdvPrompter can generated offline (without alternatively updating AdvPrompter); The downside is that the generated most likely tokens will not be adaptive.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730872939436}, {"id": "pNnDH10A6M", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4771/Reviewer_8Khr"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces a new method, and potentially a new perspective, for jailbreak prompting, the main contribution is that the prompter is itself a trained model, thus it can generate the jailbreak prompts extremely efficiently. The authors also mentioned several other properties such as human-readable or gradient-free, but these properties have been well discussed before.", "review_text": "The paper introduces a new method, and potentially a new perspective, for jailbreak prompting, the main contribution is that the prompter is itself a trained model, thus it can generate the jailbreak prompts extremely efficiently. The authors also mentioned several other properties such as human-readable or gradient-free, but these properties have been well discussed before.", "strengths": "1. The idea of using a pretrained model to directly generate jailbreak as next token prediction is interesting. \n\n2. The propose method can generate jailbreak much faster than existing methods, especially since most existing methods are optimizing for every sample.", "weaknesses": "1. the idea of training a generation model for jailbreak has many natural limitations, such as the evaluation will require a different split of data, need to validating that the model can generalize across different target LLMs, and also different benchmark datasets. The authors didn't fully address these. \n\n     -. 1.1 For example, since the method does not require any gradient of the target LLMs, the authors need to report more results on the commerlized LLMs where the strengths might be more obvious (referring to Table 2). \n\n    -. 1.2. The evaluation is only limited to AdvBench, the empirical scope is too small. Some other choices are like HarmBench or JAMBench. \n\n    -. 1.3. The evaluation needs to demonstrate the power of the advPrompter while the model is trained on AdvBench, and tested on other benchmarks such as HarmBench or JamBench. This is very important to show the advantages of proposed methods over the per-sample optimization method. \n\n    -. 1.4. Similarly, the authors might want to offer more detailed and comprehensive discussions on the differences between the targetLLM during training vs. during testing, although a gentle discussion has been offered in 4.2. \n\n2. The empirical scope is also fairly limited in terms of the methods being compared. Newer methods in jailbreaks, even just the ones published and presented in recent conferences (excluding arXiv ones), need to be discussed and compared. There are more methods that can deliver human-readable jailbreaks. (Although another method that can simultaneously fulfill all the properties in Table 1 might not exist). \n\n3. While the authors present a unique method, and might be the only one at this moment can achieve all the properties in Table 1, the performance is unfortunately achieved by trade-offs. For example, in Table 2, the proposed method is not necessarily always the best performing method in ASR. The perplexity is always the lowest, but comparison to newer method might be needed, e.g., [1]. This is important because in AI security research, ASR and perplexity are probably more important factors than generation time. Authors might need to offer more convincing discussions why the method is favored although being lower in ASR in certain cases. \n\n\n[1]. Role-playing to Generate Natural-language Jailbreakings to Test Guideline Adherence of LLMs", "questions": "1. Does the method require gradient during training? (i.e., does the method have to be trained with white-box LLM?) if that's the case, then that is another point needs to be made clear. If not, results showing how the trained advprompter from highly aligned models such as GPT then applied to less aligned models will be interesting. \n\n2. Training time and requirement of computing might also need to be discussed, although less important. \n\n3. The authors might need to compare their results in Sec. 4.3 with other jailbreak defensive methods for LLMs.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "all_content": {"summary": "The paper introduces a new method, and potentially a new perspective, for jailbreak prompting, the main contribution is that the prompter is itself a trained model, thus it can generate the jailbreak prompts extremely efficiently. The authors also mentioned several other properties such as human-readable or gradient-free, but these properties have been well discussed before.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The idea of using a pretrained model to directly generate jailbreak as next token prediction is interesting. \n\n2. The propose method can generate jailbreak much faster than existing methods, especially since most existing methods are optimizing for every sample.", "weaknesses": "1. the idea of training a generation model for jailbreak has many natural limitations, such as the evaluation will require a different split of data, need to validating that the model can generalize across different target LLMs, and also different benchmark datasets. The authors didn't fully address these. \n\n     -. 1.1 For example, since the method does not require any gradient of the target LLMs, the authors need to report more results on the commerlized LLMs where the strengths might be more obvious (referring to Table 2). \n\n    -. 1.2. The evaluation is only limited to AdvBench, the empirical scope is too small. Some other choices are like HarmBench or JAMBench. \n\n    -. 1.3. The evaluation needs to demonstrate the power of the advPrompter while the model is trained on AdvBench, and tested on other benchmarks such as HarmBench or JamBench. This is very important to show the advantages of proposed methods over the per-sample optimization method. \n\n    -. 1.4. Similarly, the authors might want to offer more detailed and comprehensive discussions on the differences between the targetLLM during training vs. during testing, although a gentle discussion has been offered in 4.2. \n\n2. The empirical scope is also fairly limited in terms of the methods being compared. Newer methods in jailbreaks, even just the ones published and presented in recent conferences (excluding arXiv ones), need to be discussed and compared. There are more methods that can deliver human-readable jailbreaks. (Although another method that can simultaneously fulfill all the properties in Table 1 might not exist). \n\n3. While the authors present a unique method, and might be the only one at this moment can achieve all the properties in Table 1, the performance is unfortunately achieved by trade-offs. For example, in Table 2, the proposed method is not necessarily always the best performing method in ASR. The perplexity is always the lowest, but comparison to newer method might be needed, e.g., [1]. This is important because in AI security research, ASR and perplexity are probably more important factors than generation time. Authors might need to offer more convincing discussions why the method is favored although being lower in ASR in certain cases. \n\n\n[1]. Role-playing to Generate Natural-language Jailbreakings to Test Guideline Adherence of LLMs", "questions": "1. Does the method require gradient during training? (i.e., does the method have to be trained with white-box LLM?) if that's the case, then that is another point needs to be made clear. If not, results showing how the trained advprompter from highly aligned models such as GPT then applied to less aligned models will be interesting. \n\n2. Training time and requirement of computing might also need to be discussed, although less important. \n\n3. The authors might need to compare their results in Sec. 4.3 with other jailbreak defensive methods for LLMs.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "details_of_ethics_concerns": "The paper is written in a technical way. Personally, I don't think the paper has ethical issues. However, the paper itself is about jailbreaking LLMs, a fairly sensitive topic, might benefit from an additional layer of caution.", "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730742457340}, {"id": "s2QaO38zOn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4771/Reviewer_QcAH"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes a novel method to enhance jailbreaking attacks on safety-aligned large language models (LLMs). The proposed method involves constructing a framework that fine-tunes an LLM from a base model by encouraging it to generate human-readable adversarial suffixes for harmful requests. Extensive experimental results demonstrate that the AdvPrompter can produce low-perplexity adversarial suffixes and achieve performance comparable to two baseline methods, i.e., GCG and AutoDAN.", "review_text": "This paper proposes a novel method to enhance jailbreaking attacks on safety-aligned large language models (LLMs). The proposed method involves constructing a framework that fine-tunes an LLM from a base model by encouraging it to generate human-readable adversarial suffixes for harmful requests. Extensive experimental results demonstrate that the AdvPrompter can produce low-perplexity adversarial suffixes and achieve performance comparable to two baseline methods, i.e., GCG and AutoDAN.", "strengths": "* The proposed method enables the fast generation of specific adversarial suffixes for individual harmful requests.\n* The experimental results show that the proposed method achieves great performance.", "weaknesses": "Overall, I think the method is sound, but there are a few concerns.\n\n* The advantages of AdvPrompter mentioned in lines 108-136 should be specified under certain comparative conditions. For example, the \"adaptivity to input\" should be highlighted in the context of generating at a low time cost, as both GCG-individual and AutoDAN-individual are also adaptive to input. The \"fast generation\" compared to GCG and AutoDAN should be specified in the context of generating individual adversarial suffixes, since GCG-universal and AutoDAN-universal are ready to be used once obtained.\n\n* Since both AutoDAN-universal and AdvPrompter generate human-readable adversarial suffixes quickly, it would be beneficial to discuss their performance in more detail, especially in Table 3.\n\n* I think the \"Gradient-free TargetLLM\" is not a significant advantage, and it is unnecessary to emphasize the \"gray-box TargetLLM\" since it is actually a \"white-box\" model.\n\n* More comparisons to existing methods, such as TAP and PAP, should be included. These methods also generate human-readable adversarial prompts.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel method to enhance jailbreaking attacks on safety-aligned large language models (LLMs). The proposed method involves constructing a framework that fine-tunes an LLM from a base model by encouraging it to generate human-readable adversarial suffixes for harmful requests. Extensive experimental results demonstrate that the AdvPrompter can produce low-perplexity adversarial suffixes and achieve performance comparable to two baseline methods, i.e., GCG and AutoDAN.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* The proposed method enables the fast generation of specific adversarial suffixes for individual harmful requests.\n* The experimental results show that the proposed method achieves great performance.", "weaknesses": "Overall, I think the method is sound, but there are a few concerns.\n\n* The advantages of AdvPrompter mentioned in lines 108-136 should be specified under certain comparative conditions. For example, the \"adaptivity to input\" should be highlighted in the context of generating at a low time cost, as both GCG-individual and AutoDAN-individual are also adaptive to input. The \"fast generation\" compared to GCG and AutoDAN should be specified in the context of generating individual adversarial suffixes, since GCG-universal and AutoDAN-universal are ready to be used once obtained.\n\n* Since both AutoDAN-universal and AdvPrompter generate human-readable adversarial suffixes quickly, it would be beneficial to discuss their performance in more detail, especially in Table 3.\n\n* I think the \"Gradient-free TargetLLM\" is not a significant advantage, and it is unnecessary to emphasize the \"gray-box TargetLLM\" since it is actually a \"white-box\" model.\n\n* More comparisons to existing methods, such as TAP and PAP, should be included. These methods also generate human-readable adversarial prompts.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730712940642}, {"id": "7MSxndCE9q", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4771/Reviewer_qn1H"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper titled \"AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs\" presents a novel approach to generating adversarial prompts that jailbreak large language models (LLMs), enabling the generation of harmful or inappropriate content. Traditional methods for adversarial prompt generation, such as manual red-teaming or optimization-based methods, can be slow, inefficient, and prone to generating semantically meaningless attacks. In contrast, the authors propose AdvPrompter, an LLM trained using a novel algorithm to rapidly generate human-readable adversarial prompts without requiring gradient information from the target LLM.\n\nThe core innovation of the paper lies in its alternating training method, AdvPrompterTrain, which alternates between generating adversarial suffixes and fine-tuning the AdvPrompter model. The resulting adversarial prompts are highly effective, achieving state-of-the-art results on the AdvBench and HarmBench datasets, with improved attack success rates, faster generation times, and strong transferability to black-box LLMs. Additionally, the paper demonstrates that by fine-tuning LLMs on datasets generated by AdvPrompter, models can become more robust against jailbreaking attacks while maintaining high performance on benchmarks like MMLU and MT-bench.", "review_text": "The paper titled \"AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs\" presents a novel approach to generating adversarial prompts that jailbreak large language models (LLMs), enabling the generation of harmful or inappropriate content. Traditional methods for adversarial prompt generation, such as manual red-teaming or optimization-based methods, can be slow, inefficient, and prone to generating semantically meaningless attacks. In contrast, the authors propose AdvPrompter, an LLM trained using a novel algorithm to rapidly generate human-readable adversarial prompts without requiring gradient information from the target LLM.\n\nThe core innovation of the paper lies in its alternating training method, AdvPrompterTrain, which alternates between generating adversarial suffixes and fine-tuning the AdvPrompter model. The resulting adversarial prompts are highly effective, achieving state-of-the-art results on the AdvBench and HarmBench datasets, with improved attack success rates, faster generation times, and strong transferability to black-box LLMs. Additionally, the paper demonstrates that by fine-tuning LLMs on datasets generated by AdvPrompter, models can become more robust against jailbreaking attacks while maintaining high performance on benchmarks like MMLU and MT-bench.", "strengths": "1. **Clarity**: The paper is well-structured and provides clear explanations of its methodology, backed by comprehensive experimental results and ablation studies. The use of figures and tables, such as Table 1, aids in understanding the comparative advantages of the method. The training algorithm and attack framework are concisely explained, which improves accessibility.\n\n2. **Significance**: AdvPrompter presents significant contributions to the area of LLM robustness and safety, offering a highly scalable solution to automatic red-teaming. Its gradient-free approach makes it applicable in both white-box and black-box settings, which broadens its impact for securing deployed LLM systems. The paper’s findings on fine-tuning models with adversarial data to improve safety also open up new avenues for automated adversarial training in LLMs.", "weaknesses": "1. **Lack of Comparison with BEAST**: One notable omission is the lack of detailed comparison with BEAST (introduced in \"Fast Adversarial Attacks on Language Models in One GPU Minute\")​. BEAST also focuses on gradient-free attacks and is highly efficient, achieving impressive success rates within one GPU minute, making it an essential baseline. The authors of AdvPrompter reference BEAST, but they fail to provide head-to-head benchmarks, especially in terms of speed and success rates. This limits the ability to assess whether AdvPrompter's claim of \"fast\" generation holds up against a method already proven to be both rapid and effective.\n\n2. \"Unclear Computational Efficiency\": While AdvPrompter claims faster generation of adversarial prompts compared to gradient-based methods, the paper does not include detailed benchmarks or profiling to demonstrate computational efficiency on a per-prompt basis. For instance, BEAST reports precise GPU utilization metrics and compares the attack time per prompt across different models. AdvPrompter lacks such concrete data, making its claims of speed improvements less convincing. Without these comparisons, it is unclear whether the method is truly fast or simply optimized for a limited set of tasks.\n\nIncomplete Discussion of Attack Readability: While AdvPrompter claims to generate human-readable adversarial prompts, there is limited qualitative analysis of the readability or coherence of these prompts. I would like to see more analysis on this.", "questions": "I encourage the authors to address the points raised in the weaknesses section and to conduct additional experiments where further investigation is required.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper titled \"AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs\" presents a novel approach to generating adversarial prompts that jailbreak large language models (LLMs), enabling the generation of harmful or inappropriate content. Traditional methods for adversarial prompt generation, such as manual red-teaming or optimization-based methods, can be slow, inefficient, and prone to generating semantically meaningless attacks. In contrast, the authors propose AdvPrompter, an LLM trained using a novel algorithm to rapidly generate human-readable adversarial prompts without requiring gradient information from the target LLM.\n\nThe core innovation of the paper lies in its alternating training method, AdvPrompterTrain, which alternates between generating adversarial suffixes and fine-tuning the AdvPrompter model. The resulting adversarial prompts are highly effective, achieving state-of-the-art results on the AdvBench and HarmBench datasets, with improved attack success rates, faster generation times, and strong transferability to black-box LLMs. Additionally, the paper demonstrates that by fine-tuning LLMs on datasets generated by AdvPrompter, models can become more robust against jailbreaking attacks while maintaining high performance on benchmarks like MMLU and MT-bench.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. **Clarity**: The paper is well-structured and provides clear explanations of its methodology, backed by comprehensive experimental results and ablation studies. The use of figures and tables, such as Table 1, aids in understanding the comparative advantages of the method. The training algorithm and attack framework are concisely explained, which improves accessibility.\n\n2. **Significance**: AdvPrompter presents significant contributions to the area of LLM robustness and safety, offering a highly scalable solution to automatic red-teaming. Its gradient-free approach makes it applicable in both white-box and black-box settings, which broadens its impact for securing deployed LLM systems. The paper’s findings on fine-tuning models with adversarial data to improve safety also open up new avenues for automated adversarial training in LLMs.", "weaknesses": "1. **Lack of Comparison with BEAST**: One notable omission is the lack of detailed comparison with BEAST (introduced in \"Fast Adversarial Attacks on Language Models in One GPU Minute\")​. BEAST also focuses on gradient-free attacks and is highly efficient, achieving impressive success rates within one GPU minute, making it an essential baseline. The authors of AdvPrompter reference BEAST, but they fail to provide head-to-head benchmarks, especially in terms of speed and success rates. This limits the ability to assess whether AdvPrompter's claim of \"fast\" generation holds up against a method already proven to be both rapid and effective.\n\n2. \"Unclear Computational Efficiency\": While AdvPrompter claims faster generation of adversarial prompts compared to gradient-based methods, the paper does not include detailed benchmarks or profiling to demonstrate computational efficiency on a per-prompt basis. For instance, BEAST reports precise GPU utilization metrics and compares the attack time per prompt across different models. AdvPrompter lacks such concrete data, making its claims of speed improvements less convincing. Without these comparisons, it is unclear whether the method is truly fast or simply optimized for a limited set of tasks.\n\nIncomplete Discussion of Attack Readability: While AdvPrompter claims to generate human-readable adversarial prompts, there is limited qualitative analysis of the readability or coherence of these prompts. I would like to see more analysis on this.", "questions": "I encourage the authors to address the points raised in the weaknesses section and to conduct additional experiments where further investigation is required.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729543151624}], "openreview_url": "https://openreview.net/forum?id=E9GakjQype", "arxiv_id": "2404.16873", "paper_pdf": "papers/E9GakjQype.pdf", "paper_pdf_sha256": "d2beb5335bae944c14f9ed923fb1cfe4eda5f43f2d00b0b92fd9f11389d02c1f", "paper_pdf_bytes": 2188146, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/advprompter", "code_repository": "facebookresearch/advprompter", "code_commit": "802a500c91f1dcd7c8b76869d3e39bf8e40ed7d7", "code_archive": "repos/E9GakjQype.zip", "code_archive_sha256": "0cdb6d552ae22cb7d8d04cec9a49bb0b4d1015946f2e715ab772704a8b4a9079", "code_archive_bytes": 58385, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 49, "github_languages": {"Python": 93783}, "github_archived": true, "github_pushed_at": "2024-05-06T20:09:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/advprompter-fast-adaptive-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qRbkTbe8JT", "year": 2024, "status": "rejected", "title": "IMEX-Reg: Implicit-Explicit Regularization in the Function Space for Continual Learning", "authors": ["Prashant Shivaram Bhat", "Bharath Chennamkulam Renjith", "Bahram Zonooz", "Elahe Arani"], "authorids": ["~Prashant_Shivaram_Bhat1", "~Bharath_Chennamkulam_Renjith1", "~Bahram_Zonooz1", "~Elahe_Arani1"], "authors_source": "OpenReview API", "abstract": "Continual learning (CL) remains one of the long-standing challenges for deep\nneural networks due to catastrophic forgetting of previously acquired knowledge.\nAlthough rehearsal-based approaches have been fairly successful in mitigating\ncatastrophic forgetting, they suffer from overfitting on buffered samples and prior\ninformation loss, hindering generalization under low-buffer regimes. Inspired\nby how humans learn using strong inductive biases, we propose IMEX-Reg to\nimprove the generalization performance of experience rehearsal in CL under low\nbuffer regimes. Specifically, we employ a two-pronged implicit-explicit regular-\nization approach using contrastive representation learning (CRL) and consistency\nregularization. To further leverage the global relationship between representations\nlearned using CRL, we propose a novel regularization strategy to guide the clas-\nsifier toward the activation correlations in the unit hypersphere of the CRL. Our\nresults show that IMEX-Reg significantly improves generalization performance and\noutperforms rehearsal-based approaches in several CL scenarios. It is also robust\nto natural and adversarial corruptions with less task-recency bias. Additionally, we\nprovide theoretical insights to support our design decisions further.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "0Bxx2wfSFz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7749/Reviewer_B5E5"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a novel approach called IMEX-Reg, tailored for low memory buffer scenarios. In addition to several other techniques such as EMA, the approach utilizes contrastive representation learning and applies regularization to both the classification outputs and projection outputs. The motivation behind the approach stems from the fact that contrastive embeddings lie on the hypersphere, which improves training stability and classification performance. The experiment shows that the proposed method outperforms the baselines under the low memory buffer settings and shows the effectiveness of the proposed techniques in the ablation study.", "review_text": "This paper introduces a novel approach called IMEX-Reg, tailored for low memory buffer scenarios. In addition to several other techniques such as EMA, the approach utilizes contrastive representation learning and applies regularization to both the classification outputs and projection outputs. The motivation behind the approach stems from the fact that contrastive embeddings lie on the hypersphere, which improves training stability and classification performance. The experiment shows that the proposed method outperforms the baselines under the low memory buffer settings and shows the effectiveness of the proposed techniques in the ablation study.", "strengths": "1. The motivation is clear and the paper is well-written\n2. It outperforms the baselines", "weaknesses": "1. The contribution of the paper is merely incremental as using contrastive learning and regularization on function space have been extensively studied before [1, 2, 3]. Imposing a regularization on the contrastive projected outputs along with the classification outputs improves the performance is not surprising.\n2. The method introduces many hyper-parameters (e.g., alpha, beta, lambda in Eq.7), but there was no study of how these hyper-parameters affect the model performance.\n3. The authors argue that their method has advantages over existing replay methods. However, rehearsal-free methods such as [3, 4] already significantly outperform the proposed method. For instance, [3] achieves 87.8% accuracy on Seq-CIFAR10 and 47.1% on Seq-TinyImageNet without the need to save any samples.\n\n[1] Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning \\\n[2] co2l- Contrastive Continual Learning \\\n[3] A theoretical study on solving continual learning \\\n[4] Look At Me, No Replay! SurpriseNet: Anomaly Detection Inspired Class Incremental Learning", "questions": "1. Based on a paper under review, [5] achieves 74+% in Seq-CIFAR10 with a buffer size of 200 in ResNet-18. The authors may compare their method with [5]\n2. Why is this method robust to natural corruption? An important discussion aligned with robustness to data distribution (or corruption) is covered in [6].\n\nMisc.\nPlease use \\` rather than ' for \\`SGD' and \\`Joint' on page 6\n\n\n[5] Learnability and algorithm for continual learning \\\n[6] A multi-head model for continual learning via out-of-distribution detection", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel approach called IMEX-Reg, tailored for low memory buffer scenarios. In addition to several other techniques such as EMA, the approach utilizes contrastive representation learning and applies regularization to both the classification outputs and projection outputs. The motivation behind the approach stems from the fact that contrastive embeddings lie on the hypersphere, which improves training stability and classification performance. The experiment shows that the proposed method outperforms the baselines under the low memory buffer settings and shows the effectiveness of the proposed techniques in the ablation study.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The motivation is clear and the paper is well-written\n2. It outperforms the baselines", "weaknesses": "1. The contribution of the paper is merely incremental as using contrastive learning and regularization on function space have been extensively studied before [1, 2, 3]. Imposing a regularization on the contrastive projected outputs along with the classification outputs improves the performance is not surprising.\n2. The method introduces many hyper-parameters (e.g., alpha, beta, lambda in Eq.7), but there was no study of how these hyper-parameters affect the model performance.\n3. The authors argue that their method has advantages over existing replay methods. However, rehearsal-free methods such as [3, 4] already significantly outperform the proposed method. For instance, [3] achieves 87.8% accuracy on Seq-CIFAR10 and 47.1% on Seq-TinyImageNet without the need to save any samples.\n\n[1] Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning \\\n[2] co2l- Contrastive Continual Learning \\\n[3] A theoretical study on solving continual learning \\\n[4] Look At Me, No Replay! SurpriseNet: Anomaly Detection Inspired Class Incremental Learning", "questions": "1. Based on a paper under review, [5] achieves 74+% in Seq-CIFAR10 with a buffer size of 200 in ResNet-18. The authors may compare their method with [5]\n2. Why is this method robust to natural corruption? An important discussion aligned with robustness to data distribution (or corruption) is covered in [6].\n\nMisc.\nPlease use \\` rather than ' for \\`SGD' and \\`Joint' on page 6\n\n\n[5] Learnability and algorithm for continual learning \\\n[6] A multi-head model for continual learning via out-of-distribution detection", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698809936268}, {"id": "MdKP3Nmv7p", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7749/Reviewer_bb13"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "This paper studies an interesting topic, continual learning, which aims to learn a series of tasks without forgetting. In order to the generalization performance of the memory-based methods, this paper introduces to employ contrastive representation learning (CRL) and consistency regularization. The experiment results show that the proposed approach achieve good results in continual learning.", "review_text": "This paper studies an interesting topic, continual learning, which aims to learn a series of tasks without forgetting. In order to the generalization performance of the memory-based methods, this paper introduces to employ contrastive representation learning (CRL) and consistency regularization. The experiment results show that the proposed approach achieve good results in continual learning.", "strengths": "1. The main idea seems interesting.\n2. This paper studies an interesting topic.", "weaknesses": "1. The notations are hard to follow. For example, x and y should be bold because they are matrixes.\n2. The parameters of the shared model and classifier are not defined.\n3. Eq.1 is not clear to me. What is the actual network for f and g? Why is h not used in Eq.1?\n4. In the text below Eq.2, you said z = h(f(.)). However, z is not defined in Eq.2. The input and output patterns for the models f, g and h are unclear.  \n5. Why introduce the existing Conjecture 1? Does this theory connect with your actual design?\n6. Why the classifier can create the function spaces?\n7. The proposed approach is based on the existing technology, and the overall novelty is small.\n8. The proposed approach still requires the task information, which can not be used in more realistic continual learning settings such as task-free continual learning.\n9. The methodology section is hard to follow. A lot of notations are not defined clearly and the proposed approach is not novel enough.", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies an interesting topic, continual learning, which aims to learn a series of tasks without forgetting. In order to the generalization performance of the memory-based methods, this paper introduces to employ contrastive representation learning (CRL) and consistency regularization. The experiment results show that the proposed approach achieve good results in continual learning.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The main idea seems interesting.\n2. This paper studies an interesting topic.", "weaknesses": "1. The notations are hard to follow. For example, x and y should be bold because they are matrixes.\n2. The parameters of the shared model and classifier are not defined.\n3. Eq.1 is not clear to me. What is the actual network for f and g? Why is h not used in Eq.1?\n4. In the text below Eq.2, you said z = h(f(.)). However, z is not defined in Eq.2. The input and output patterns for the models f, g and h are unclear.  \n5. Why introduce the existing Conjecture 1? Does this theory connect with your actual design?\n6. Why the classifier can create the function spaces?\n7. The proposed approach is based on the existing technology, and the overall novelty is small.\n8. The proposed approach still requires the task information, which can not be used in more realistic continual learning settings such as task-free continual learning.\n9. The methodology section is hard to follow. A lot of notations are not defined clearly and the proposed approach is not novel enough.", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698769880784}, {"id": "ZFo2zT0z5K", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7749/Reviewer_siLj"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper develops IMEX-Reg as a new method to tackle catastrophic forgetting in continual learning settings.  IMEX-Reg  is inspired by the nervous system mechanisms and combines contrastive representation learning with consistency regularization. It aligns the classifier with CRL representations in the unit hypersphere. Empirical results are offered to demonstrate that IMEX-Reg improves the model generalization and leads to SOTA performance compared to the baselines. IMEX-Reg is also resilient to adversarial data issues and reduces bias towards recent tasks. The approach is also supported by theoretical justifications.", "review_text": "The paper develops IMEX-Reg as a new method to tackle catastrophic forgetting in continual learning settings.  IMEX-Reg  is inspired by the nervous system mechanisms and combines contrastive representation learning with consistency regularization. It aligns the classifier with CRL representations in the unit hypersphere. Empirical results are offered to demonstrate that IMEX-Reg improves the model generalization and leads to SOTA performance compared to the baselines. IMEX-Reg is also resilient to adversarial data issues and reduces bias towards recent tasks. The approach is also supported by theoretical justifications.", "strengths": "1. CL is still an active research area and the proposed approach is a new method for this purpose.\n\n2. The paper reads well and can be followed straightforwardly.\n\n3. Section D in the Appendix is informative and offers insights about the weaknesses and future potentials for the proposed research.", "weaknesses": "1. Continual learning in the context of the used baselines is a mature field with many existing works. However, the method does provide SOTA results across the board to warrant a contribution that offers a significant performance boost.\n\n2. Theoretical justifications of the paper are not novel and mostly are reiterating previous results. Doing so is OK but does not offer any new theoretical contributions. \n\n3. Some aspects of the algorithm are not studied extensively.", "questions": "1. The connection between the proposed approach and the nervous system is very loose and emphasis on this aspect is overstated. What is the reason behind this emphasis without providing much evidence to support it?\n\n2. It is important to study the effect of \\alph, \\beta, and \\lambda on the performance. How the user should tune them? A study should be offered for this purpose. If the performance is sensitive with respect to the values of these hyperparameters, then it is essential to provide a solution for selecting the optimal values.\n\n3. There are other common settings to study CL using CIFAR100, e.g., using 20 tasks each with 5 classes. I think adding experiments for these settings is also helpful to demonstrate how well the method scales when there are more tasks.\n\n4. Having learning curves in CL is common and allows for studying the dynamic of learning. I think providing them in addition to the tables is helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper develops IMEX-Reg as a new method to tackle catastrophic forgetting in continual learning settings.  IMEX-Reg  is inspired by the nervous system mechanisms and combines contrastive representation learning with consistency regularization. It aligns the classifier with CRL representations in the unit hypersphere. Empirical results are offered to demonstrate that IMEX-Reg improves the model generalization and leads to SOTA performance compared to the baselines. IMEX-Reg is also resilient to adversarial data issues and reduces bias towards recent tasks. The approach is also supported by theoretical justifications.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. CL is still an active research area and the proposed approach is a new method for this purpose.\n\n2. The paper reads well and can be followed straightforwardly.\n\n3. Section D in the Appendix is informative and offers insights about the weaknesses and future potentials for the proposed research.", "weaknesses": "1. Continual learning in the context of the used baselines is a mature field with many existing works. However, the method does provide SOTA results across the board to warrant a contribution that offers a significant performance boost.\n\n2. Theoretical justifications of the paper are not novel and mostly are reiterating previous results. Doing so is OK but does not offer any new theoretical contributions. \n\n3. Some aspects of the algorithm are not studied extensively.", "questions": "1. The connection between the proposed approach and the nervous system is very loose and emphasis on this aspect is overstated. What is the reason behind this emphasis without providing much evidence to support it?\n\n2. It is important to study the effect of \\alph, \\beta, and \\lambda on the performance. How the user should tune them? A study should be offered for this purpose. If the performance is sensitive with respect to the values of these hyperparameters, then it is essential to provide a solution for selecting the optimal values.\n\n3. There are other common settings to study CL using CIFAR100, e.g., using 20 tasks each with 5 classes. I think adding experiments for these settings is also helpful to demonstrate how well the method scales when there are more tasks.\n\n4. Having learning curves in CL is common and allows for studying the dynamic of learning. I think providing them in addition to the tables is helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698486129360}], "openreview_url": "https://openreview.net/forum?id=qRbkTbe8JT", "arxiv_id": "2404.18161", "paper_pdf": "papers/qRbkTbe8JT.pdf", "paper_pdf_sha256": "c60973ec84e37afe0d9317a98c09963f29841b574c735e885533505dc78ccba2", "paper_pdf_bytes": 489335, "paper_pdf_source": "openreview", "code_url": "https://github.com/NeurAI-Lab/IMEX-Reg", "code_repository": "NeurAI-Lab/IMEX-Reg", "code_commit": "74169d4617c1a86db2d3f1b11ef7182689cb9293", "code_archive": "repos/qRbkTbe8JT.zip", "code_archive_sha256": "8b28b80fa188503cae961903cb76834b8b0234c6c4e44046b9e883c91e947cee", "code_archive_bytes": 52460, "code_file_count": 41, "code_extensions": {".py": 41}, "github_disk_usage_kb": 42, "github_languages": {"Python": 158980}, "github_archived": false, "github_pushed_at": "2024-05-24T20:11:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/imex-reg-implicit-explicit-regularization-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6OxI4WqGr6", "year": 2023, "status": "rejected", "title": "Semi-Supervised Offline Reinforcement Learning with Action-Free Trajectories", "authors": ["Qinqing Zheng", "Mikael Henaff", "Brandon Amos", "Aditya Grover"], "authorids": ["~Qinqing_Zheng1", "~Mikael_Henaff1", "~Brandon_Amos1", "~Aditya_Grover1"], "authors_source": "OpenReview API", "abstract": "Natural agents can effectively learn from multiple data sources that differ in size, quality, and types of measurements. We study this heterogeneity in the context of offline reinforcement learning (RL) by introducing a new, practically motivated semi-supervised setting. Here, an agent has access to two sets of trajectories: labelled trajectories containing state, action, reward triplets at every timestep, along with unlabelled trajectories that contain only state and reward information.  For this setting, we develop a simple meta-algorithmic pipeline that learns an inverse-dynamics model on the labelled data to obtain proxy-labels for the unlabelled data, followed by the use of any offline RL algorithm on the true and proxy-labelled trajectories. Empirically, we find this simple pipeline to be highly successful --- on several D4RL benchmarks~\\cite{fu2020d4rl}, certain offline RL algorithms can match the performance of variants trained on a fully labeled dataset even when we label only 10\\% trajectories from the low return regime. Finally, we perform a large-scale controlled empirical study investigating the interplay of data-centric properties of the labelled and unlabelled datasets, with algorithmic design choices (e.g., inverse dynamics, offline RL algorithm) to identify general trends and best practices for training RL agents on semi-supervised offline datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "bdlDjxuX2eP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2369/Reviewer_Mz9R"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new, practically motivated semi-supervised setting, where the agent can access both labeled trajectories and unlabelled trajectories that do not include the actions of the trajectories. A model is trained to give actions for unlabeled data and then the offline RL method is trained by the whole dataset. Experiments based on D4RL dataset show the good performance of the proposed method.", "review_text": "See the above-mentioned questions.", "strengths": "The paper is well-written and easy to follow. The setting is novel and meaningful since there are a lot of unlabeled data like videos in real-world scenarios.  The experiments are designed carefully to illustrate the influence of data with different qualities. For the weakness, Iwould like to ask some questions:\n1. For IDM, the length of the input, k, could be changed, and k=1 in the paper. So the input of the model is four states (s\n_t-1,..., s_t+1). How does the model encode these states?\n2.  The experiments only use the expert dataset, which means most trajectories are good. D4RL also provides random and medium-level dataset. As your claim in the paper, the quality of the data has a huge influence on the performance. Is there any analysis based on these data? Could we say that the method also needs high-quality data for both labeled and unlabeled data to achieve good performance?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces a new, practically motivated semi-supervised setting, where the agent can access both labeled trajectories and unlabelled trajectories that do not include the actions of the trajectories. A model is trained to give actions for unlabeled data and then the offline RL method is trained by the whole dataset. Experiments based on D4RL dataset show the good performance of the proposed method.", "strength_and_weaknesses": "The paper is well-written and easy to follow. The setting is novel and meaningful since there are a lot of unlabeled data like videos in real-world scenarios.  The experiments are designed carefully to illustrate the influence of data with different qualities. For the weakness, Iwould like to ask some questions:\n1. For IDM, the length of the input, k, could be changed, and k=1 in the paper. So the input of the model is four states (s\n_t-1,..., s_t+1). How does the model encode these states?\n2.  The experiments only use the expert dataset, which means most trajectories are good. D4RL also provides random and medium-level dataset. As your claim in the paper, the quality of the data has a huge influence on the performance. Is there any analysis based on these data? Could we say that the method also needs high-quality data for both labeled and unlabeled data to achieve good performance?", "clarity,_quality,_novelty_and_reproducibility": "The detailed dataset and code of the method are not provided.", "summary_of_the_review": "See the above-mentioned questions.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666704645088}, {"id": "xD5FiPIwT9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2369/Reviewer_az7q"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a method for semi-supervised learning in the offline RL setting where the unlabelled part of the dataset consists of action-free state trajectories and the labelled part consists of the full trajectories. They use an inverse dynamics model to learn actions that give rise to state transitions and use the learned model to inject labels for unlabelled data and perform offline learning using classic model-free offline RL algorithms such as CQL.", "review_text": "Overall, I lean towards rejecting this paper. My reasoning is that while it is tackling a notable practical problem in the RL setting, there is not enough technical contribution nor is it polished enough in its exposition of the design choices for me to recommend acceptance. ", "strengths": "Weaknesses:\n- My main concern is with the technical contribution of this paper. Considering the inverse dynamics model as the crucial technical novelty, I am not sure the method given in Eq. 1 is the best, and at the very least, some further evaluations and arguments for its design choice is needed. For one, the covariate matrix with k>0 would seem to break the markov property of the RL setting. Furthermore, it is not justified why this choice was made beyond the empirical validation, which is not enough either in my opinion. I am not convinced that this extension of IDM significantly contributes to the overall goal of taking advantage of semi-supervision, as compared to the fully unsupervised case. In particular, it would have been interesting to see whether semi-supervision in this regime would help when the data split is within trajectories and not between trajectories. \n- Along the lines of the above, I think an analysis of why only a single labelling round was used instead of the conventional self-training paradigm of retraining per round could have been included to make the paper stronger. In particular, I assume it could be possible to provide a more detailed analysis of how well the inverse dynamics model learns when compared to the ground truth data on the log-likelihood of the multivariate Gaussians used to estimate them since you have access to those data labels. \n- As it stands, this paper's main contributions are a careful study of the different considerations one should take when doing semi-supervised offline RL. It has a good experimental validation of how performant value-based and BC methods. However, it does not make enough technical contribution to take advantage of the specific literature in the semi-supervised learning setting or does not justify its design choices well. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces a method for semi-supervised learning in the offline RL setting where the unlabelled part of the dataset consists of action-free state trajectories and the labelled part consists of the full trajectories. They use an inverse dynamics model to learn actions that give rise to state transitions and use the learned model to inject labels for unlabelled data and perform offline learning using classic model-free offline RL algorithms such as CQL.", "strength_and_weaknesses": "Weaknesses:\n- My main concern is with the technical contribution of this paper. Considering the inverse dynamics model as the crucial technical novelty, I am not sure the method given in Eq. 1 is the best, and at the very least, some further evaluations and arguments for its design choice is needed. For one, the covariate matrix with k>0 would seem to break the markov property of the RL setting. Furthermore, it is not justified why this choice was made beyond the empirical validation, which is not enough either in my opinion. I am not convinced that this extension of IDM significantly contributes to the overall goal of taking advantage of semi-supervision, as compared to the fully unsupervised case. In particular, it would have been interesting to see whether semi-supervision in this regime would help when the data split is within trajectories and not between trajectories. \n- Along the lines of the above, I think an analysis of why only a single labelling round was used instead of the conventional self-training paradigm of retraining per round could have been included to make the paper stronger. In particular, I assume it could be possible to provide a more detailed analysis of how well the inverse dynamics model learns when compared to the ground truth data on the log-likelihood of the multivariate Gaussians used to estimate them since you have access to those data labels. \n- As it stands, this paper's main contributions are a careful study of the different considerations one should take when doing semi-supervised offline RL. It has a good experimental validation of how performant value-based and BC methods. However, it does not make enough technical contribution to take advantage of the specific literature in the semi-supervised learning setting or does not justify its design choices well. ", "clarity,_quality,_novelty_and_reproducibility": "This paper is clearly written, with some minor typos throughout. Overall, my main concern is with the novelty and quality of the proposed algorithm. The results should be reproducible.", "summary_of_the_review": "Overall, I lean towards rejecting this paper. My reasoning is that while it is tackling a notable practical problem in the RL setting, there is not enough technical contribution nor is it polished enough in its exposition of the design choices for me to recommend acceptance. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666675660635}, {"id": "utyHuT20BA", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2369/Reviewer_omLk"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper provides an empirical study about a setting where a portion of the offline RL dataset doesn't include actions. To exploit action-missing data, this work proposes to learn an inverse dynamics model on data with actions to generate proxy actions from state transitions. A set of empirical studies on d4rl gym-locomotion control are provided to give insights into how such semi-supervised learning will be helpful for the performance of the final policy. The ablation studies show the proposed semi-supervised method is particularly helpful when the action-missing data is of high quality and labelled data is of lower quality.", "review_text": "The paper studies an interesting setting for offline RL which has significant practical value. The study of different scenarios of data availability and the quality of the data is a timely topic for offline RL research.\nHowever, my main concern about this paper is it's not very informative. The whole paper looks more like a technical report about applying the inverse dynamics model to the d4rl dataset and I'm struggling to figure out what's the key takeaway. \nThe authors summarised three of their key findings in the introduction:\n1. One interesting claim of the paper is SS-ORL agent works well in the setting with lower-quality labelled and high-quality unlabelled data. But I'm not sure if such a claim can generalize to the setting when the policies are more diverse than a single or a mixture of two unimodal policies, as I argued in the second point of the weakness. \n2. \"When the labelled data quality is high, utilizing unlabelled data does not bring significant benefits.\" This might just reveal that most of the data from the d4rl gym-locomotion are redundant because of the low diversity of the policy?\n3. \"CQL and TD3BC are less sensitive to the action labels compared to DT.\" This is not very obvious to me by just looking at Figure 4.1 and Figure 4.2. It would be better to have quantitative results if that's true.\n\nGiven the reasons above, I cannot recommend acceptance at the current stage. But I may change my mind if the authors or other reviewers can convince me about point 1, or remind me why point 2 and 3 are significant enough to accept the paper.", "strengths": "Strength:\n1. The setting is well-motivated and has decent potential for real-world applications. \n2. The empirical studies are thorough and in particular, ablation over the quality of data is well-done.\n\nWeaknesses:\n1. The proposed method is quite standard in online settings. The claimed novelty (multiple transitions as input to the inverse model) is more of technical detail and the reason why it is helpful for the Markovian setting is not explained properly.\n2. Most of the experiment results in the paper are based on the medium-expert dataset of d4rl gym-locomotion. This raises questions about whether the conclusion can be generalised to the setting with more diverse data. For medium-expert, the behaviour policy is basically a mixture of two policies (medium level and expert level). If the trajectories are split into multiple groups according to the returns, it's likely that the root cause of the varied returns is initial states rather than the quality of the policy. One clue is the experiments on medium-replay data in Figure C.2 show for hopper and walker2d, the quality of the unlabelled data plays a much more important role than the case in the medium and medium expert.\n\nMinor issues\n1. Is \"label\" a proper word to replace action? My impression is unlabelled data in offline RL is more about reward-missing data. The intuition behind that is humans can label rewards to trajectories easily because it's a scale but \"labelling\" actions seems like very difficult.\n2. There are some obvious grammatical errors in the paper: e.g. \"we are mainly interested in the case **only** a **significant majority** of the trajectories in the offline trajectories are **unlabeled**.\" \"How can we utilize the unlabelled data for improving **the performance offline RL algorithms**?\"\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper provides an empirical study about a setting where a portion of the offline RL dataset doesn't include actions. To exploit action-missing data, this work proposes to learn an inverse dynamics model on data with actions to generate proxy actions from state transitions. A set of empirical studies on d4rl gym-locomotion control are provided to give insights into how such semi-supervised learning will be helpful for the performance of the final policy. The ablation studies show the proposed semi-supervised method is particularly helpful when the action-missing data is of high quality and labelled data is of lower quality.", "strength_and_weaknesses": "Strength:\n1. The setting is well-motivated and has decent potential for real-world applications. \n2. The empirical studies are thorough and in particular, ablation over the quality of data is well-done.\n\nWeaknesses:\n1. The proposed method is quite standard in online settings. The claimed novelty (multiple transitions as input to the inverse model) is more of technical detail and the reason why it is helpful for the Markovian setting is not explained properly.\n2. Most of the experiment results in the paper are based on the medium-expert dataset of d4rl gym-locomotion. This raises questions about whether the conclusion can be generalised to the setting with more diverse data. For medium-expert, the behaviour policy is basically a mixture of two policies (medium level and expert level). If the trajectories are split into multiple groups according to the returns, it's likely that the root cause of the varied returns is initial states rather than the quality of the policy. One clue is the experiments on medium-replay data in Figure C.2 show for hopper and walker2d, the quality of the unlabelled data plays a much more important role than the case in the medium and medium expert.\n\nMinor issues\n1. Is \"label\" a proper word to replace action? My impression is unlabelled data in offline RL is more about reward-missing data. The intuition behind that is humans can label rewards to trajectories easily because it's a scale but \"labelling\" actions seems like very difficult.\n2. There are some obvious grammatical errors in the paper: e.g. \"we are mainly interested in the case **only** a **significant majority** of the trajectories in the offline trajectories are **unlabeled**.\" \"How can we utilize the unlabelled data for improving **the performance offline RL algorithms**?\"\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is easy to follow in general and the writing is clear.\nThe quality of empirical evaluations is fine.\nThe methodology originality is low but the empirical study on the offline RL setting is novel.", "summary_of_the_review": "The paper studies an interesting setting for offline RL which has significant practical value. The study of different scenarios of data availability and the quality of the data is a timely topic for offline RL research.\nHowever, my main concern about this paper is it's not very informative. The whole paper looks more like a technical report about applying the inverse dynamics model to the d4rl dataset and I'm struggling to figure out what's the key takeaway. \nThe authors summarised three of their key findings in the introduction:\n1. One interesting claim of the paper is SS-ORL agent works well in the setting with lower-quality labelled and high-quality unlabelled data. But I'm not sure if such a claim can generalize to the setting when the policies are more diverse than a single or a mixture of two unimodal policies, as I argued in the second point of the weakness. \n2. \"When the labelled data quality is high, utilizing unlabelled data does not bring significant benefits.\" This might just reveal that most of the data from the d4rl gym-locomotion are redundant because of the low diversity of the policy?\n3. \"CQL and TD3BC are less sensitive to the action labels compared to DT.\" This is not very obvious to me by just looking at Figure 4.1 and Figure 4.2. It would be better to have quantitative results if that's true.\n\nGiven the reasons above, I cannot recommend acceptance at the current stage. But I may change my mind if the authors or other reviewers can convince me about point 1, or remind me why point 2 and 3 are significant enough to accept the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1665754670725}], "openreview_url": "https://openreview.net/forum?id=6OxI4WqGr6", "arxiv_id": "2210.06518", "paper_pdf": "papers/6OxI4WqGr6.pdf", "paper_pdf_sha256": "a2f78117fae4e7d2c63b2d5af3b61fdd8abf613d1b1aee5ecb27237ccb422172", "paper_pdf_bytes": 1146698, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/ssorl", "code_repository": "facebookresearch/ssorl", "code_commit": "cfefabed915e76aefb453f29344055014906d4ed", "code_archive": "repos/6OxI4WqGr6.zip", "code_archive_sha256": "efe378d8359c08fb7cfbd609168b18f0401b778ad6d216234bbde8a4bfd629cb", "code_archive_bytes": 59279, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 53, "github_languages": {"Python": 152547}, "github_archived": true, "github_pushed_at": "2023-07-16T20:08:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/semi-supervised-offline-reinforcement-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "s51gCxF70pq", "year": 2022, "status": "rejected", "title": "Learning Temporally-Consistent Representations for Data-Efficient Reinforcement Learning", "authors": ["Trevor McInroe", "Lukas Schäfer", "Stefano V Albrecht"], "authorids": ["~Trevor_McInroe1", "~Lukas_Schäfer1", "~Stefano_V_Albrecht1"], "authors_source": "OpenReview API", "abstract": "Deep reinforcement learning (RL) agents that exist in high-dimensional state spaces, such as those composed of images, have interconnected learning burdens. Agents must learn an action-selection policy that completes their given task, which requires them to learn a representation of the state space that discerns between useful and useless information. The reward function is the only supervised feedback that RL agents receive, which causes a representation learning bottleneck that can manifest in poor sample efficiency. We present $k$-Step Latent (KSL), a new representation learning method that enforces temporal consistency of representations via a self-supervised auxiliary task wherein agents learn to recurrently predict action-conditioned representations of the state space. The state encoder learned by KSL produces low-dimensional representations that make optimization of the RL task more sample efficient. Altogether, KSL produces state-of-the-art results in both data efficiency and asymptotic performance in the popular PlaNet benchmark suite. Our analyses show that KSL produces encoders that generalize better to new tasks unseen during training, and its representations are more strongly tied to reward, are more invariant to perturbations in the state space, and move more smoothly through the temporal axis of the RL problem than other methods such as DrQ, RAD, CURL, and SAC-AE.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "8GbrIxj5Ovh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1915/Reviewer_J6YX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper shows that an auxiliary self-supervised task that enforces temporal consistency of latents improves sample efficiency in continuous control environemnts, and identifies important implementation details that make it work. ", "review_text": "Data-efficiency of RL algorithms is an important research area, and this paper explores how auxiliary self-supervised learning can improve data-efficiency in continuous control domains. The paper is very rich with experimental details, the implementation choices are carefully ablated and the paper is overall well written and explained. \n\n### Similarities with SPR\n------\nMy main issue with this paper is the fact that it positions the K-step latent objective as a “new representation learning method”, whereas in fact the K-step latent is exactly the representation learning method used in SPR [1]. Throughout the abstract, intro and the methods section, the paper positions KSL as a new representation method: but in practice it’s an adaptation of SPR to continuous control that requires some implementation changes. The entirety of section 3.2 and the Figure 1 is exactly SPR (see Section 2.2 and Figure 2 in the SPR paper), but the paper makes no references to it except in the related work section. The related work section again fails to acknowledge that KSL and SPR share the same representation learning objective, and not only the architecture. \n\nThe authors could very well have positioned this paper as “We applied SPR to continuous control, here’s some implementation changes we needed to make along the way to make it work”, and it would have been a much more honest and accurate description of the work.  The empirical contribution here and the detailed analyses itself would have been valuable on its own. \n\n### Questions on Experiments\n------\nThe description of the “Generalization of Encoders” experiments is very sparse. Can you specify what exactly the training tasks and evaluation tasks are in the generalization experiment? \n\nFor the invariance experiment, it would have been nicer to see invariance to real-world distractors and not artificial noise. I would recommend the Distracting Control Suite [2] for more convincing experiments around these.\n\nFor a lot of results, the performance seems to be under-reported than the original results in papers, especially for RAD and DrQ. Here’s a link to the raw performance scores for baseline methods used in [3] https://console.cloud.google.com/storage/browser/rl-benchmark-data/dm_control, and these were reportedly obtained from the corresponding authors. Can you clarify the discrepancy in the performance data? This seems to be a major issue. \n\nAdditionally, in a lot of performance curves, the standard deviation regions overlap, making it harder to establish stochastic dominance of one method over another. It would be nicer to see a better stochastic analysis using stratified CIs on multiple normalized metrics (see Figure 11 in [3]). You can do this easily via the colab: https://bit.ly/statistical_precipice_colab\n\n[1]  Schwarzer, M., Anand, A., Goel, R., Hjelm, R. D., Courville, A., & Bachman, P. (2020). Data-efficient reinforcement learning with self-predictive representations. ICLR 2021. https://arxiv.org/abs/2007.05929\n\n[2] Stone, A., Ramirez, O., Konolige, K., & Jonschkowski, R. (2021). The Distracting Control Suite—A Challenging Benchmark for Reinforcement Learning from Pixels. https://arxiv.org/abs/2101.02722\n\n[3] Agarwal, R., Schwarzer, M., Castro, P. S., Courville, A., & Bellemare, M. G. (2021). Deep reinforcement learning at the edge of the statistical precipice. NeurIPS 2021 https://arxiv.org/abs/2108.13264", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper shows that an auxiliary self-supervised task that enforces temporal consistency of latents improves sample efficiency in continuous control environemnts, and identifies important implementation details that make it work. ", "main_review": "Data-efficiency of RL algorithms is an important research area, and this paper explores how auxiliary self-supervised learning can improve data-efficiency in continuous control domains. The paper is very rich with experimental details, the implementation choices are carefully ablated and the paper is overall well written and explained. \n\n### Similarities with SPR\n------\nMy main issue with this paper is the fact that it positions the K-step latent objective as a “new representation learning method”, whereas in fact the K-step latent is exactly the representation learning method used in SPR [1]. Throughout the abstract, intro and the methods section, the paper positions KSL as a new representation method: but in practice it’s an adaptation of SPR to continuous control that requires some implementation changes. The entirety of section 3.2 and the Figure 1 is exactly SPR (see Section 2.2 and Figure 2 in the SPR paper), but the paper makes no references to it except in the related work section. The related work section again fails to acknowledge that KSL and SPR share the same representation learning objective, and not only the architecture. \n\nThe authors could very well have positioned this paper as “We applied SPR to continuous control, here’s some implementation changes we needed to make along the way to make it work”, and it would have been a much more honest and accurate description of the work.  The empirical contribution here and the detailed analyses itself would have been valuable on its own. \n\n### Questions on Experiments\n------\nThe description of the “Generalization of Encoders” experiments is very sparse. Can you specify what exactly the training tasks and evaluation tasks are in the generalization experiment? \n\nFor the invariance experiment, it would have been nicer to see invariance to real-world distractors and not artificial noise. I would recommend the Distracting Control Suite [2] for more convincing experiments around these.\n\nFor a lot of results, the performance seems to be under-reported than the original results in papers, especially for RAD and DrQ. Here’s a link to the raw performance scores for baseline methods used in [3] https://console.cloud.google.com/storage/browser/rl-benchmark-data/dm_control, and these were reportedly obtained from the corresponding authors. Can you clarify the discrepancy in the performance data? This seems to be a major issue. \n\nAdditionally, in a lot of performance curves, the standard deviation regions overlap, making it harder to establish stochastic dominance of one method over another. It would be nicer to see a better stochastic analysis using stratified CIs on multiple normalized metrics (see Figure 11 in [3]). You can do this easily via the colab: https://bit.ly/statistical_precipice_colab\n\n[1]  Schwarzer, M., Anand, A., Goel, R., Hjelm, R. D., Courville, A., & Bachman, P. (2020). Data-efficient reinforcement learning with self-predictive representations. ICLR 2021. https://arxiv.org/abs/2007.05929\n\n[2] Stone, A., Ramirez, O., Konolige, K., & Jonschkowski, R. (2021). The Distracting Control Suite—A Challenging Benchmark for Reinforcement Learning from Pixels. https://arxiv.org/abs/2101.02722\n\n[3] Agarwal, R., Schwarzer, M., Castro, P. S., Courville, A., & Bellemare, M. G. (2021). Deep reinforcement learning at the edge of the statistical precipice. NeurIPS 2021 https://arxiv.org/abs/2108.13264", "summary_of_the_review": "Data-efficiency of RL algorithms is an important research area, and this paper explores how auxiliary self-supervised learning can improve data-efficiency in continuous control domains. The paper is very rich with experimental details, the implementation choices are carefully ablated and the paper is overall well written and explained. \n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636224442667}, {"id": "kBhysgOm7i1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1915/Reviewer_pTps"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper tackles the problem of sample inefficiency in continuous control, by noting that standard RL methods deal with both policy optimisation and representation learning jointly with a single supervisory signal, namely the reward. Consequently, authors propose to leverage long-term temporal connections between actions in the representation learning, and introduce k-Step Latent (KSL), a representation learning module for learning temporally consistent representations of the state space. Authors show that KSL improves previous state-of-the-art methods in PlaNet benchmark suite and provide some analysis of the representations learned by KSL.", "review_text": "The paper addresses a very important problem in RL: data efficiency. The author's perspective on leveraging long-term temporal connection is not exactly novel (see Self-Predictive Representation [Schwarzer et al. 2021], Successor Features [Kulkarni, et al. 2016, Barreto et al. 2017]), but the specific method introduced seems to be novel, to the best of my knowledge.  \n\nOn the method:\nThe motivation is clearly stated and makes sense to me. I would have like to see discussion on how this differ/relates to successor features, when learned jointly or separately. \n\nA claimed \"that learned representations of the state space should relate to reward\" is not really justified: there are data-efficient method that disentangle the reward from the representation. It also seems to contradict another desired property: generalisation from one task to another. \n\nOn the experiment:\nThe experiments are extensive and the method is compared against sensible baselines. The results with respect to data-efficiency are promising.\n\n\nminor comments: \n- the font is very small on most figures axes\n- figures 5.2, 5.3 and 6 would make more sense with a different y axis scale.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper tackles the problem of sample inefficiency in continuous control, by noting that standard RL methods deal with both policy optimisation and representation learning jointly with a single supervisory signal, namely the reward. Consequently, authors propose to leverage long-term temporal connections between actions in the representation learning, and introduce k-Step Latent (KSL), a representation learning module for learning temporally consistent representations of the state space. Authors show that KSL improves previous state-of-the-art methods in PlaNet benchmark suite and provide some analysis of the representations learned by KSL.", "main_review": "The paper addresses a very important problem in RL: data efficiency. The author's perspective on leveraging long-term temporal connection is not exactly novel (see Self-Predictive Representation [Schwarzer et al. 2021], Successor Features [Kulkarni, et al. 2016, Barreto et al. 2017]), but the specific method introduced seems to be novel, to the best of my knowledge.  \n\nOn the method:\nThe motivation is clearly stated and makes sense to me. I would have like to see discussion on how this differ/relates to successor features, when learned jointly or separately. \n\nA claimed \"that learned representations of the state space should relate to reward\" is not really justified: there are data-efficient method that disentangle the reward from the representation. It also seems to contradict another desired property: generalisation from one task to another. \n\nOn the experiment:\nThe experiments are extensive and the method is compared against sensible baselines. The results with respect to data-efficiency are promising.\n\n\nminor comments: \n- the font is very small on most figures axes\n- figures 5.2, 5.3 and 6 would make more sense with a different y axis scale.", "summary_of_the_review": "The main idea behind the paper is not novel, but the implementation is, and the results are promising. Some claims are not founded and somehow contradictory, and some related works are missing. But overall an interesting contribution!", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635963317889}, {"id": "cAD_Wo6r5s", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1915/Reviewer_TKuY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper applied \"Bootstrap your own latent\" (BYOL) to the case of RL by introducing an additional learned transition model and show that this can improve sample efficiency.", "review_text": "Strenghts:\n* Simple approach\n* Clearly written\n* Representation learning for better generalization or sample efficiency is an important topic in RL\n* Positive experimental results\n\nWeaknesses: \n* My main worry with this application of BYOL to RL is that the introduction of the transition model T changes the 'support' of z_m away from the support of z_o. In other words, while psi_o expects an output of T, psi_m gets the direct output of phi_m for which it was not trained and which might be entirely different from the output of T. Based on the experiments, this still seems to do something useful, but I would argue that psi_m should be seen more as random mapping than as the slow moving state-encoding. In any case, I think this out-of-distribution problem for psi_m should be addressed in the paper. \n* For Figure 3: Why not use t-SNI instead of only the first two dimensions of PCA? In particular, while it's not wrong to say that \"DrQ's projections show little sign of reward-based orgnaization by 15k steps\", that is slightly misleading as it doesn't say anything about the latent representation as we're only looking at 2 principal axes. \n\nAdditional Questions:\n* How is translation augmentation applied?\n* Nit: How would the results change when removing the sg before the policy?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper applied \"Bootstrap your own latent\" (BYOL) to the case of RL by introducing an additional learned transition model and show that this can improve sample efficiency.", "main_review": "Strenghts:\n* Simple approach\n* Clearly written\n* Representation learning for better generalization or sample efficiency is an important topic in RL\n* Positive experimental results\n\nWeaknesses: \n* My main worry with this application of BYOL to RL is that the introduction of the transition model T changes the 'support' of z_m away from the support of z_o. In other words, while psi_o expects an output of T, psi_m gets the direct output of phi_m for which it was not trained and which might be entirely different from the output of T. Based on the experiments, this still seems to do something useful, but I would argue that psi_m should be seen more as random mapping than as the slow moving state-encoding. In any case, I think this out-of-distribution problem for psi_m should be addressed in the paper. \n* For Figure 3: Why not use t-SNI instead of only the first two dimensions of PCA? In particular, while it's not wrong to say that \"DrQ's projections show little sign of reward-based orgnaization by 15k steps\", that is slightly misleading as it doesn't say anything about the latent representation as we're only looking at 2 principal axes. \n\nAdditional Questions:\n* How is translation augmentation applied?\n* Nit: How would the results change when removing the sg before the policy?", "summary_of_the_review": "An interesting direct application of BYOL to RL. However, the necessity to include action-conditioned transition models in RL raises additional complications compared to BYOL which have not yet been addressed (or discussed) and I believe these should be included in the paper before publication. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635926545150}, {"id": "cTLf0waB7SN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1915/Reviewer_38iT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a representation-learning method (k-step latent, or KSL) that uses a self-supervised auxiliary loss between recurrently-predicted action-conditioned representations of the state space and non-recurrently predicted target representations, in the style of BYOL. The method is trained using two separate optimizers on different parts of the models, to avoid interference in the statistics maintained by the optimizers. Results at 100k and 500k steps on 6 tasks from the DM control suite (from pixels) compared to methods using alternative self-supervised auxiliary losses show that the proposed method improves data efficiency. Analysis of the learned latent representations shows that those from KSL produce more robust encoders and are more consistent with the underlying MDP.", "review_text": "**Clarity.**\nFor the most part, the writing is clear. One minor concern is that the description of KSL (sec 3.2) is difficult to understand, and could be made much more clear with better choices of notation, including the pseudocode in the main text, and an explanation of the algorithm that follows the pseudocode. \n\nThat said, the main issue with the clarity and writing of the paper is that the contribution is not made clear both in the paper nor in the experimental analysis. Is the paper about using k-step latents plus an auxiliary loss tying these to observations? Or is the paper about the specific form this auxiliary loss takes? Most of the writing and the naming (KSL) imply the former but the evaluation speaks to the latter.\n\n**Novelty and significance.**\nClarity aside, the paper does not propose a sufficiently novel contribution for acceptance. There is a very large body of work in model-based reinforcement learning that uses k-step model-based predictions and corresponding losses to improve data efficiency and performance, and reduce model rollout errors. Some of this is even cited in this paper, but others that are not cited include Recurrent environment simulators by Chiappa et al 2017, TreeQN by Farquhar et al 2018, MuZero by Schrittwieser et al 2019, and Muesli by Hessel et al 2021. These methods all use k-step latents with a loss tying them to (encodings of) observations.\n\nThe specific form of the auxiliary loss here, adapted from BYOL (which is not even mentioned until the related work for some reason), is very similar to that of SPR, which is cited by the paper. There are some minor differences but these mainly seem like implementation details and since there are no empirical comparisons I have to assume this is the case.\n\nFurther, the experiments themselves are extremely limited. Evaluating only on 6 tasks from the DM control suite is not enough to show that this is a compelling and useful contribution, especially given the high overlap with prior work. The additional analysis of the learned representations are nice, but are not enough without showing the strength of the proposed approach, or else doing a much more thorough analysis. Finally, I’d like to see ablations of the components and choices made for the proposed method. Why have both \\Psi_o and P? Is using the EMA for the momentum pathway the best choice? Is a normalized L2 loss the best choice?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a representation-learning method (k-step latent, or KSL) that uses a self-supervised auxiliary loss between recurrently-predicted action-conditioned representations of the state space and non-recurrently predicted target representations, in the style of BYOL. The method is trained using two separate optimizers on different parts of the models, to avoid interference in the statistics maintained by the optimizers. Results at 100k and 500k steps on 6 tasks from the DM control suite (from pixels) compared to methods using alternative self-supervised auxiliary losses show that the proposed method improves data efficiency. Analysis of the learned latent representations shows that those from KSL produce more robust encoders and are more consistent with the underlying MDP.", "main_review": "**Clarity.**\nFor the most part, the writing is clear. One minor concern is that the description of KSL (sec 3.2) is difficult to understand, and could be made much more clear with better choices of notation, including the pseudocode in the main text, and an explanation of the algorithm that follows the pseudocode. \n\nThat said, the main issue with the clarity and writing of the paper is that the contribution is not made clear both in the paper nor in the experimental analysis. Is the paper about using k-step latents plus an auxiliary loss tying these to observations? Or is the paper about the specific form this auxiliary loss takes? Most of the writing and the naming (KSL) imply the former but the evaluation speaks to the latter.\n\n**Novelty and significance.**\nClarity aside, the paper does not propose a sufficiently novel contribution for acceptance. There is a very large body of work in model-based reinforcement learning that uses k-step model-based predictions and corresponding losses to improve data efficiency and performance, and reduce model rollout errors. Some of this is even cited in this paper, but others that are not cited include Recurrent environment simulators by Chiappa et al 2017, TreeQN by Farquhar et al 2018, MuZero by Schrittwieser et al 2019, and Muesli by Hessel et al 2021. These methods all use k-step latents with a loss tying them to (encodings of) observations.\n\nThe specific form of the auxiliary loss here, adapted from BYOL (which is not even mentioned until the related work for some reason), is very similar to that of SPR, which is cited by the paper. There are some minor differences but these mainly seem like implementation details and since there are no empirical comparisons I have to assume this is the case.\n\nFurther, the experiments themselves are extremely limited. Evaluating only on 6 tasks from the DM control suite is not enough to show that this is a compelling and useful contribution, especially given the high overlap with prior work. The additional analysis of the learned representations are nice, but are not enough without showing the strength of the proposed approach, or else doing a much more thorough analysis. Finally, I’d like to see ablations of the components and choices made for the proposed method. Why have both \\Psi_o and P? Is using the EMA for the momentum pathway the best choice? Is a normalized L2 loss the best choice?\n", "summary_of_the_review": "Overall, this paper lacks sufficient novelty for acceptance. It recombines existing techniques in a slightly different way than previously and shows improvements on a very small and narrow set of environments without comparing to the most relevant related work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635896438883}, {"id": "hPFzN9KfUb1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1915/Reviewer_Qru8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Summary:\nThe authors introduced k-Step Latent (KSL), a representation learning method for visual-based continuous control tasks. KSL utilises multi-step latent action-dependent predictive supervision for training the representation. The empirical evaluations are based on the dm-control suite benchmarks, KSL demonstrates improved sample efficiency (100k evaluation) and asymptotic performance (500k evaluation) across the six tasks presented, comparing to the baseline algorithms (mainly based on image augmentation). The authors further empirically examined the properties of the learned representations, and showed that the trained encoder quickly learns to be representative of the reward structure. The authors also argue that KSL supports more robust representation learning and stronger generalisability.\n", "review_text": "Pros:\n\n - The paper is well-written and easy to comprehend;\n\n - The empirical evaluation on dm-control suite indeed shows that KSL yields state-of-the-art results on the six presented tasks;\n\n - The choice of using independent optimisers for training the representation learning module given the signals from the predictive latent supervision (Eq. 4) and critic training (Eq. 1) respectively raises a good point for representation learning in RL with auxiliary tasks, from multi-task learning perspective. Such inductive bias is additionally substantiated with empirical comparisons.\n\nConcerns:\n\n - The main evaluation on the overall performance is rather limited, it might worth showing more (especially the middle/hard tasks as defined in Yarats, et al. 2021);\n\n - The KSL model combines many existing techniques in representation learning for RL, such as image augmentation (Laskin, et al. 2020) and multi-step latent predictive supervision auxiliary task (Schwarzer, et al. 2021), leading to limited novelty of the proposed KSL model.\n\n - The multi-step latent predictive supervision objective for representation learning in KSL is highly similar to the SPR model (Schwarzer, et al. 2021), it seems like the authors acknowledge the similarity hence spent a paragraph discussing the difference, but discussions are mainly based on the architectural/training differences. From this perspective, KSL appears as an adaptation of the SPR model to the continuous control tasks (despite the admirable engineering efforts). The difference argument could be made convincing given that the authors could provide further clarifications of the differences between KSL and SPR, from an algorithmic perspective, and provide further empirical comparisons between the two agents (e.g., the authors state that \"KSL’s architecture is general enough to be applied in both discrete- and continuous-action domains\", hence it would be interesting to see an adaptation of KSL to discrete control and test it on atari benchmarks, and compare with the SPR results).\n\n - The arguments of the improved representation learning of KSL in terms of \"Permutation Invariance\" and \"Temporal Coherence\" seems poorly justified. The reported figures do not show significant or consistent improvement of the KSL over the baseline methods, only by largely zooming in the y-axis could one observe some differences. I doubt if such minuscule differences are consequential in the overall learning.\n\n - I think using the momentum encoder to provide input to the target Q-network in SAC critic-training is an interesting choice, but lacks further (theoretical) justification, i.e., why would this be better than simply using the online encoder as the inputs to the target Q-network, is it possibly because of the consistent temporal lags?\n\nMinor Points:\n\n - The main empirical evaluations of the learned representations is based on the walker-walk task, which is a simple task with dense reward structure(Yarats, et al. 2021). It would be more interesting to see how the learned representations are indicative of the reward structure in the sparse reward tasks, such that Cartpole-Swingup.\n\n - KSL is motivated by the inductive bias that \"States that are nearby in time are likely to share high levels of mutual information\", another similar work that utilises multi-step action-dependent latent predictions by Whitney, et al. (2019) is based on the inductive bias that the similarities between the embeddings for the states and/or action sequences should be based on their successor outcomes (e.g., successor representations). It would be nice to see some discussions on the relationship between the two seemingly independent inductive biases.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Summary:\nThe authors introduced k-Step Latent (KSL), a representation learning method for visual-based continuous control tasks. KSL utilises multi-step latent action-dependent predictive supervision for training the representation. The empirical evaluations are based on the dm-control suite benchmarks, KSL demonstrates improved sample efficiency (100k evaluation) and asymptotic performance (500k evaluation) across the six tasks presented, comparing to the baseline algorithms (mainly based on image augmentation). The authors further empirically examined the properties of the learned representations, and showed that the trained encoder quickly learns to be representative of the reward structure. The authors also argue that KSL supports more robust representation learning and stronger generalisability.\n", "main_review": "Pros:\n\n - The paper is well-written and easy to comprehend;\n\n - The empirical evaluation on dm-control suite indeed shows that KSL yields state-of-the-art results on the six presented tasks;\n\n - The choice of using independent optimisers for training the representation learning module given the signals from the predictive latent supervision (Eq. 4) and critic training (Eq. 1) respectively raises a good point for representation learning in RL with auxiliary tasks, from multi-task learning perspective. Such inductive bias is additionally substantiated with empirical comparisons.\n\nConcerns:\n\n - The main evaluation on the overall performance is rather limited, it might worth showing more (especially the middle/hard tasks as defined in Yarats, et al. 2021);\n\n - The KSL model combines many existing techniques in representation learning for RL, such as image augmentation (Laskin, et al. 2020) and multi-step latent predictive supervision auxiliary task (Schwarzer, et al. 2021), leading to limited novelty of the proposed KSL model.\n\n - The multi-step latent predictive supervision objective for representation learning in KSL is highly similar to the SPR model (Schwarzer, et al. 2021), it seems like the authors acknowledge the similarity hence spent a paragraph discussing the difference, but discussions are mainly based on the architectural/training differences. From this perspective, KSL appears as an adaptation of the SPR model to the continuous control tasks (despite the admirable engineering efforts). The difference argument could be made convincing given that the authors could provide further clarifications of the differences between KSL and SPR, from an algorithmic perspective, and provide further empirical comparisons between the two agents (e.g., the authors state that \"KSL’s architecture is general enough to be applied in both discrete- and continuous-action domains\", hence it would be interesting to see an adaptation of KSL to discrete control and test it on atari benchmarks, and compare with the SPR results).\n\n - The arguments of the improved representation learning of KSL in terms of \"Permutation Invariance\" and \"Temporal Coherence\" seems poorly justified. The reported figures do not show significant or consistent improvement of the KSL over the baseline methods, only by largely zooming in the y-axis could one observe some differences. I doubt if such minuscule differences are consequential in the overall learning.\n\n - I think using the momentum encoder to provide input to the target Q-network in SAC critic-training is an interesting choice, but lacks further (theoretical) justification, i.e., why would this be better than simply using the online encoder as the inputs to the target Q-network, is it possibly because of the consistent temporal lags?\n\nMinor Points:\n\n - The main empirical evaluations of the learned representations is based on the walker-walk task, which is a simple task with dense reward structure(Yarats, et al. 2021). It would be more interesting to see how the learned representations are indicative of the reward structure in the sparse reward tasks, such that Cartpole-Swingup.\n\n - KSL is motivated by the inductive bias that \"States that are nearby in time are likely to share high levels of mutual information\", another similar work that utilises multi-step action-dependent latent predictions by Whitney, et al. (2019) is based on the inductive bias that the similarities between the embeddings for the states and/or action sequences should be based on their successor outcomes (e.g., successor representations). It would be nice to see some discussions on the relationship between the two seemingly independent inductive biases.\n", "summary_of_the_review": "Scores:\n\nI suggest marginal rejection (5/10). KSL indeed show state-of-the-art performance on the presented tasks and I like the way the authors assessed the quality of representation learning. However, KSL seems like a combination of existing methods, but lacks comprehensive empirical evaluation in that sense. Moreover, the high similarity with SPR is concerning without further clarification and empirical justification. Some arguments of the improvement learned representation is over-stated.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635851089934}], "openreview_url": "https://openreview.net/forum?id=s51gCxF70pq", "arxiv_id": "2110.04935", "paper_pdf": "papers/s51gCxF70pq.pdf", "paper_pdf_sha256": "77427493e7d80a45671bf60be24684de67c5ee544cc420d85428ae4edcc0ff18", "paper_pdf_bytes": 3816891, "paper_pdf_source": "openreview", "code_url": "https://github.com/anon-researcher-repo/ksl", "code_repository": "anon-researcher-repo/ksl", "code_commit": "72f0665322e8e4de33838a8e3a57356c192c5743", "code_archive": "repos/s51gCxF70pq.zip", "code_archive_sha256": "9c2e5dde0894d72f4527b68515946252de7caf329d800dfadbb7072428cda9cb", "code_archive_bytes": 55851, "code_file_count": 7, "code_extensions": {".py": 6, ".ipynb": 1}, "github_disk_usage_kb": 66, "github_languages": {"Jupyter Notebook": 56108, "Python": 48398}, "github_archived": false, "github_pushed_at": "2021-10-06T14:48:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-temporally-consistent-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lJuOUWlAC8i", "year": 2021, "status": "rejected", "title": "Learning Contextualized Knowledge Structures for Commonsense Reasoning", "authors": ["Jun Yan", "Mrigank Raman", "Tianyu Zhang", "Ryan Rossi", "Handong Zhao", "Sungchul Kim", "Nedim Lipka", "Xiang Ren"], "authorids": ["~Jun_Yan5", "mt1170736@iitd.ac.in", "zhang-ty17@mails.tsinghua.edu.cn", "~Ryan_Rossi1", "~Handong_Zhao3", "sukim@adobe.com", "lipka@adobe.com", "~Xiang_Ren1"], "authors_source": "OpenReview API", "abstract": "Recently, neural-symbolic architectures have achieved success on commonsense reasoning through effectively encoding relational structures retrieved from external knowledge graphs (KGs) and obtained state-of-the-art results in tasks such as (commonsense) question answering and natural language inference. However, current neural-symbolic reasoning methods rely on quality and contextualized knowledge structures (i.e., fact triples) that can be retrieved at the pre-processing stage and overlook challenges such as dealing with incompleteness of a KG (low coverage), limited expressiveness of its relations, and irrelevant retrieved facts in the reasoning context. \nIn this paper, we present a novel neural-symbolic approach, named Hybrid Graph Network (HGN), which jointly generates feature representations for new triples (as complement to the existing edges in the KG), determines relevance of the triples to the reasoning context, and learns graph model parameters for encoding the relational information. Our method learns a compact graph structure (comprising both retrieved and generated edges) through filtering edges that are unhelpful to the reasoning process. We show marked improvements on three commonsense reasoning benchmarks and demonstrate the superiority of the learned graph structures with user studies.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "JzSfOkUXAkc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2935/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a graph network (called HGN), aiming to better leverage commonsense knowledge graphs (KGs) to solve commonsense question answering and reasoning tasks, by jointly generating representations for new triples from KGs, determining relevance of the triples, and learning graph model parameters. The proposed model is tested on several tasks: CommonsenseQA, OpenbookQA, and CODAH.\n\nPros:\n-  Overall, the paper is easy to follow, although there are a number of typos or grammatical errors that need to be fixed.  The overall idea is clear.\n-  Jointly learning (pruning) the graph structure with the network parameters is interesting.\n-  The proposed model outperforms the baselines in comparison.\n-  Human evaluation is provided.\n\nCons:\n-  My major concern about this paper is the novelty and contributions in terms of methodology. Compared to existing methods (e.g., those PG models proposed (Wang et al. 2020)), the novelty of the current submission is rather limited---the proposed model of jointly generating new triples and learning (pruning) the graph structure with the network parameters is an interesting, but a pretty incremental idea. \n\n-  The empirical comparison to previous work (e.g., Wang et al. 2020) needs to be clearer to help understand the empirical advantages of the proposed models. The paper mentioned some reason of excluding PG-Full from comparison, but since PG-Global does not include static knowledge embedding and PG-full does, is the latter a more reasonable baseline to be compared with? The model does not achieve better performance than existing models on some tasks, which casts doubts on its effectiveness; e.g., whether its advantage is orthogonal to that brought by stronger models such as those performing much better on the OpenbookQA task.\n\nMore comments:\n-  The paper uses much space to discuss neural symbolic approaches. Given the vague benefit of doing so, it may be better to use the limited space to focus more on establishing the contributions w.r.t. existing models; e.g., more details about (Wang et al., 2020) can be provided and compared to in both methodological and experimental analyses.\n- The human evaluation was performed on the questions with correct questions. More analyses on the edges and weights generated for questions that were not correctly answers may help better understand the proposed model. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novelty of the proposed model is limited", "review": "The paper proposes a graph network (called HGN), aiming to better leverage commonsense knowledge graphs (KGs) to solve commonsense question answering and reasoning tasks, by jointly generating representations for new triples from KGs, determining relevance of the triples, and learning graph model parameters. The proposed model is tested on several tasks: CommonsenseQA, OpenbookQA, and CODAH.\n\nPros:\n-  Overall, the paper is easy to follow, although there are a number of typos or grammatical errors that need to be fixed.  The overall idea is clear.\n-  Jointly learning (pruning) the graph structure with the network parameters is interesting.\n-  The proposed model outperforms the baselines in comparison.\n-  Human evaluation is provided.\n\nCons:\n-  My major concern about this paper is the novelty and contributions in terms of methodology. Compared to existing methods (e.g., those PG models proposed (Wang et al. 2020)), the novelty of the current submission is rather limited---the proposed model of jointly generating new triples and learning (pruning) the graph structure with the network parameters is an interesting, but a pretty incremental idea. \n\n-  The empirical comparison to previous work (e.g., Wang et al. 2020) needs to be clearer to help understand the empirical advantages of the proposed models. The paper mentioned some reason of excluding PG-Full from comparison, but since PG-Global does not include static knowledge embedding and PG-full does, is the latter a more reasonable baseline to be compared with? The model does not achieve better performance than existing models on some tasks, which casts doubts on its effectiveness; e.g., whether its advantage is orthogonal to that brought by stronger models such as those performing much better on the OpenbookQA task.\n\nMore comments:\n-  The paper uses much space to discuss neural symbolic approaches. Given the vague benefit of doing so, it may be better to use the limited space to focus more on establishing the contributions w.r.t. existing models; e.g., more details about (Wang et al., 2020) can be provided and compared to in both methodological and experimental analyses.\n- The human evaluation was performed on the questions with correct questions. More analyses on the edges and weights generated for questions that were not correctly answers may help better understand the proposed model. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604023805175}, {"id": "UdpljQPZXK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2935/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a question answering model that is augmented with a common-sense knowledge graph (KG). The paper builds on the following two observations — (a) KGs are incomplete often lacking facts that would be needed for reasoning to answer a question. (b) Current methods over-retrieves facts (edges) from the KG leading to a lot of unrelated facts that potentially makes reasoning noisier and harder.\n\nThe paper first retrieves all possible facts from the KG connecting entities in the question and answer. However, due to the incompleteness of the KG, the retrieved subgraph might be missing important edges between entities. To deal with this, they connect all nodes between question and answer entities and initialize the embedding of the newly added edge with hidden layers of a sentence generated from a fine-tuned GPT2 language model (this important detail was mentioned in the appendix). However, currently the graph is over complete and very noisy. The proposed model then sparsifies the graph by learning edge weights via a two-step message passing process. The edge weight is learned as a function of the current edge representation and the textual representation. Lastly, an entropy term is added to the objective function to encourage more peakiness (and sparsity). \n\nThe model is tested on three common-sense QA benchmarks and on three of them they beat the baselines albeit only around 1-2%. Statistical significance of the result was not reported. Ablation study show that the efficacy of including the generated edges and pruning the graph. There was a small human-study also done where the annotators were shown a binarized graph and were asked to rate each edge. Annotators had moderate agreement between themselves in finding that the pruned graph was better than the original retrieved graph.\n\n\nStrengths:\n* Developing models that can use symbolic external knowledge present in common-sense KGs and also overcome the sparsity in KG is important and this model is a step in that direction\n* The paper achieves a little improvement in performance in all three datasets and ablation experiments are helpful in understanding the results\n* The paper is clearly written and it was easy to follow for the most part\n\nWeaknesses & clarifying questions for the authors:\n* My biggest complain of the paper in its current form is that several modeling choices were not motivated at all. For example, generating edges between nodes using GPT-2 language model is fairly non-standard. However, the paper lacks any motivation on why this is the right approach to generate facts which are not captured in a KG. What is the guarantee that GPT-2 will not hallucinate and generate a false fact and thereby adding unnecessary noise in the reasoning process.\n* Following up on the previous point, there could have been several other modeling choices. For example, instead of generating text via a language model, one could gather text (sentences) from Wikipedia or other text corpora containing the entities (which would mean the text would probably not be a false fact). These modeling choices were not explored and were not discussed. \n* The GPT-2 modeling choice was also moved to the appendix and I think it should definitely be moved to the main section of the paper as it is one of the core technical contribution of the paper.\n* Another modeling decision that was not motivated was the graph reasoning part. It is unclear to me why the edge weight is modeled as a part of the message passing process. Another (simpler) alternative could be modeling it as an edge attention, which is computed wrt the text and the current node embeddings. I would be curious to know how this simple model worked and if it didn’t why was the case.\n* Even though there are improvements across dataset, the improvements are relatively minor (<1% in few datasets). I think it would be useful to have statistical significance test.\n* Regarding the human study, if I understand correctly, was only the node and adjacent matrix shown to the annotators?. Was the relation type (KB relations and generated sentences) included too? If they were not included I think they should be because knowing the relations is also very improvement.\n* Can you elaborate on the average helpfulness score of edges in table 5? How many (what proportions) were scored 0, 1 or 2 for both the graphs? I think it would also be helpful to report how many facts all/majority of the annotators found to be helpful for both the graphs. \n\nMissing Reference: It would be nice to cite Sun et al EMNLP 2019 -- PullNet: Open Domain Question Answering with\nIterative Retrieval on Knowledge Bases and Text since one of the core contributions of that paper was to retrieve and keep only relevant facts from the KG. Relation paths in KG were explored by several works before Wang et al 2020 such as Neelakantan et al ACL 2015 - Compositional Vector Space Models for Knowledge Base Completion, Das et al EACL 2017 -- Chains of reasoning over entities, relations and text etc. It would be nice to cite those work as well.\n\nRecommendation:  In light of the current weaknesses of the paper, I am giving it a score of 5 and I look forward to the discussion.\n\n=======11/22======\n\nI am deciding to keep the same scores as before. Some of the initial concerns remain. I think the paper still lacks motivation wrt the GPT2 model generating missing edges. Thank you for getting the latest results, the paper is stronger than before and with some more work, I am confident it will be a good contribution to the research community.\n\n=====11/24======\n\nAfter having read through the explanation behind using GPT2 as edge features (and sufficient backing by 2 closely related work), I am increasing my score to 6. I think the discussion helped in somewhat convincing me that this approach would work for ConceptNet because of its limited schema.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The paper proposes a question answering model that is augmented with a common-sense knowledge graph (KG). The paper builds on the following two observations — (a) KGs are incomplete often lacking facts that would be needed for reasoning to answer a question. (b) Current methods over-retrieves facts (edges) from the KG leading to a lot of unrelated facts that potentially makes reasoning noisier and harder.\n\nThe paper first retrieves all possible facts from the KG connecting entities in the question and answer. However, due to the incompleteness of the KG, the retrieved subgraph might be missing important edges between entities. To deal with this, they connect all nodes between question and answer entities and initialize the embedding of the newly added edge with hidden layers of a sentence generated from a fine-tuned GPT2 language model (this important detail was mentioned in the appendix). However, currently the graph is over complete and very noisy. The proposed model then sparsifies the graph by learning edge weights via a two-step message passing process. The edge weight is learned as a function of the current edge representation and the textual representation. Lastly, an entropy term is added to the objective function to encourage more peakiness (and sparsity). \n\nThe model is tested on three common-sense QA benchmarks and on three of them they beat the baselines albeit only around 1-2%. Statistical significance of the result was not reported. Ablation study show that the efficacy of including the generated edges and pruning the graph. There was a small human-study also done where the annotators were shown a binarized graph and were asked to rate each edge. Annotators had moderate agreement between themselves in finding that the pruned graph was better than the original retrieved graph.\n\n\nStrengths:\n* Developing models that can use symbolic external knowledge present in common-sense KGs and also overcome the sparsity in KG is important and this model is a step in that direction\n* The paper achieves a little improvement in performance in all three datasets and ablation experiments are helpful in understanding the results\n* The paper is clearly written and it was easy to follow for the most part\n\nWeaknesses & clarifying questions for the authors:\n* My biggest complain of the paper in its current form is that several modeling choices were not motivated at all. For example, generating edges between nodes using GPT-2 language model is fairly non-standard. However, the paper lacks any motivation on why this is the right approach to generate facts which are not captured in a KG. What is the guarantee that GPT-2 will not hallucinate and generate a false fact and thereby adding unnecessary noise in the reasoning process.\n* Following up on the previous point, there could have been several other modeling choices. For example, instead of generating text via a language model, one could gather text (sentences) from Wikipedia or other text corpora containing the entities (which would mean the text would probably not be a false fact). These modeling choices were not explored and were not discussed. \n* The GPT-2 modeling choice was also moved to the appendix and I think it should definitely be moved to the main section of the paper as it is one of the core technical contribution of the paper.\n* Another modeling decision that was not motivated was the graph reasoning part. It is unclear to me why the edge weight is modeled as a part of the message passing process. Another (simpler) alternative could be modeling it as an edge attention, which is computed wrt the text and the current node embeddings. I would be curious to know how this simple model worked and if it didn’t why was the case.\n* Even though there are improvements across dataset, the improvements are relatively minor (<1% in few datasets). I think it would be useful to have statistical significance test.\n* Regarding the human study, if I understand correctly, was only the node and adjacent matrix shown to the annotators?. Was the relation type (KB relations and generated sentences) included too? If they were not included I think they should be because knowing the relations is also very improvement.\n* Can you elaborate on the average helpfulness score of edges in table 5? How many (what proportions) were scored 0, 1 or 2 for both the graphs? I think it would also be helpful to report how many facts all/majority of the annotators found to be helpful for both the graphs. \n\nMissing Reference: It would be nice to cite Sun et al EMNLP 2019 -- PullNet: Open Domain Question Answering with\nIterative Retrieval on Knowledge Bases and Text since one of the core contributions of that paper was to retrieve and keep only relevant facts from the KG. Relation paths in KG were explored by several works before Wang et al 2020 such as Neelakantan et al ACL 2015 - Compositional Vector Space Models for Knowledge Base Completion, Das et al EACL 2017 -- Chains of reasoning over entities, relations and text etc. It would be nice to cite those work as well.\n\nRecommendation:  In light of the current weaknesses of the paper, I am giving it a score of 5 and I look forward to the discussion.\n\n=======11/22======\n\nI am deciding to keep the same scores as before. Some of the initial concerns remain. I think the paper still lacks motivation wrt the GPT2 model generating missing edges. Thank you for getting the latest results, the paper is stronger than before and with some more work, I am confident it will be a good contribution to the research community.\n\n=====11/24======\n\nAfter having read through the explanation behind using GPT2 as edge features (and sufficient backing by 2 closely related work), I am increasing my score to 6. I think the discussion helped in somewhat convincing me that this approach would work for ConceptNet because of its limited schema.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603928613867}, {"id": "55ZmquMImnm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2935/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "=== Summary ===\n\nIn this paper, the authors propose a new approach towards incorporating knowledge graphs (KG) into commonsense QA frameworks. KGs are helpful for adding structured \"world\" information, which neural-symbolic architectures can leverage to do commonsense reasoning, e.g., \"what is the expensive resource in printing on paper?\" (paper). In such architectures, however, the authors argue that KG quality is a large impediment (e.g., missing or incorrect edges, distracting nodes, etc). Therefore, they propose a \"hybrid\" KG-based model (accordingly named \"Hybrid Graph Network\") that jointly learns to refine/augment the graph structure while also optimizing it for inference performance.\n\nExperiments are conducted on a number of commonsense reasoning tasks with multiple KG sources, and compared to relevant baselines. They also perform a user study to examine the \"helpfulness\" of the refined KGs produced by the HGN.\n\n=== Justification for Score ===\n\nThis paper is well-written and well-evaluated. The proposed method is also relatively simple and intuitively motivated. The experiments, however, only show modest (yet still positive) empirical gains. Perhaps not a game-changer for commonsense QA, but still a reasonable contribution that I would recommend for acceptance.\n\n=== Strengths ===\n\n+ The paper is clear and well-written. \n+ The experimental section is strong. The model is compared to strong baselines, and I appreciated the extra user-study on learned graph structure.\n+ The method is well-motivated, and provides (modest) empirical gains compared to some baselines.\n+ The method shows good performance with respect to increasing data efficiency (Fig. 4).\n\n=== Concerns ===\n\n- The main concern is on the empirical effectiveness of the model. The results appear to give only modest gains at best (against comparable baselines to the best of my knowledge). For a number of the results the variance is large compared to the relative difference---it would helpful to also include tests of significance for these improvements.\n\n- On OpenbookQA the model significantly underperforms T5-based models. Though I appreciate T5 is unwieldy due to its large size, it makes me question if this method indeed presents a complimentary gain, or is climbing the wrong architectural hill.\n\n=== Update After Rebuttal ===\n\nI commend the authors on a through rebuttal and active rewrites/experimentation. I still think the work is good, and can warrant acceptance. However, I still find the empirical results to be only moderate at best (though I appreciated the authors' rebuttal and significance testing). I am keeping my score the same.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well evaluated and effective KG-based commonsense QA framework", "review": "=== Summary ===\n\nIn this paper, the authors propose a new approach towards incorporating knowledge graphs (KG) into commonsense QA frameworks. KGs are helpful for adding structured \"world\" information, which neural-symbolic architectures can leverage to do commonsense reasoning, e.g., \"what is the expensive resource in printing on paper?\" (paper). In such architectures, however, the authors argue that KG quality is a large impediment (e.g., missing or incorrect edges, distracting nodes, etc). Therefore, they propose a \"hybrid\" KG-based model (accordingly named \"Hybrid Graph Network\") that jointly learns to refine/augment the graph structure while also optimizing it for inference performance.\n\nExperiments are conducted on a number of commonsense reasoning tasks with multiple KG sources, and compared to relevant baselines. They also perform a user study to examine the \"helpfulness\" of the refined KGs produced by the HGN.\n\n=== Justification for Score ===\n\nThis paper is well-written and well-evaluated. The proposed method is also relatively simple and intuitively motivated. The experiments, however, only show modest (yet still positive) empirical gains. Perhaps not a game-changer for commonsense QA, but still a reasonable contribution that I would recommend for acceptance.\n\n=== Strengths ===\n\n+ The paper is clear and well-written. \n+ The experimental section is strong. The model is compared to strong baselines, and I appreciated the extra user-study on learned graph structure.\n+ The method is well-motivated, and provides (modest) empirical gains compared to some baselines.\n+ The method shows good performance with respect to increasing data efficiency (Fig. 4).\n\n=== Concerns ===\n\n- The main concern is on the empirical effectiveness of the model. The results appear to give only modest gains at best (against comparable baselines to the best of my knowledge). For a number of the results the variance is large compared to the relative difference---it would helpful to also include tests of significance for these improvements.\n\n- On OpenbookQA the model significantly underperforms T5-based models. Though I appreciate T5 is unwieldy due to its large size, it makes me question if this method indeed presents a complimentary gain, or is climbing the wrong architectural hill.\n\n=== Update After Rebuttal ===\n\nI commend the authors on a through rebuttal and active rewrites/experimentation. I still think the work is good, and can warrant acceptance. However, I still find the empirical results to be only moderate at best (though I appreciated the authors' rebuttal and significance testing). I am keeping my score the same.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603919085671}], "openreview_url": "https://openreview.net/forum?id=lJuOUWlAC8i", "arxiv_id": "2010.12873", "paper_pdf": "papers/lJuOUWlAC8i.pdf", "paper_pdf_sha256": "8c713e67d2faba556ea45b1c39a904c3161730a0c00c720bec8c91c8143a7539", "paper_pdf_bytes": 628874, "paper_pdf_source": "openreview", "code_url": "https://github.com/INK-USC/HGN", "code_repository": "INK-USC/HGN", "code_commit": "42c391dc9753156c5dfb3fb686b615110d8c36ce", "code_archive": "repos/lJuOUWlAC8i.zip", "code_archive_sha256": "10f6cd9a9c3b1e13b9a49c4e5e9dade80ddac6b72d18b5694890478c45cbca38", "code_archive_bytes": 71478, "code_file_count": 21, "code_extensions": {".py": 20, ".sh": 1}, "github_disk_usage_kb": 68, "github_languages": {"Python": 259119, "Shell": 1569}, "github_archived": false, "github_pushed_at": "2022-11-24T09:41:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-contextualized-knowledge-structures"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryxtCpNtDS", "year": 2020, "status": "rejected", "title": "Autoencoders and Generative Adversarial Networks for Imbalanced Sequence Classification", "authors": ["Stephanie Ger", "Diego Klabjan"], "authorids": ["stephanieger@u.northwestern.edu", "d-klabjan@northwestern.edu"], "authors_source": "OpenReview API", "abstract": "We introduce a novel synthetic oversampling method for variable length, multi- feature sequence datasets based on autoencoders and generative adversarial net- works. We show that this method improves classification accuracy for highly imbalanced sequence classification tasks. We show that this method outperforms standard oversampling techniques that use techniques such as SMOTE and autoencoders. We also use generative adversarial networks on the majority class as an outlier detection method for novelty detection, with limited classification improvement. We show that the use of generative adversarial network based synthetic data improves classification model performance on a variety of sequence data sets.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1ej5XlTqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper863/AnonReviewer5"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. For imbalanced learning problem, Precision cannot play a good role. Therefore, I recommend using the performance metrics F-value and G-mean to provide comprehensive assessments. \n2. In table 2 of the 5%  data imbalance, the proposed method is not as good as the baseline. Could you provide more results with different percentage of data imbalance, such as 2%, 3%, and 4%. \n2. The authors created their own baseline, and compared against it. There is plenty of baseline methods in literature to compare against such as:\n[1] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321–357, 2002. \n[2] Haibo He, Yang Bai, Edwardo A Garcia, and Shutao Li. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In 2008 IEEE International Joint Conference on Neural Networks, pp. 1322–1328. IEEE, 2008. \n[3] Han H, Wang W Y, Mao B H. Borderline-SMOTE: a new over-sampling method in imbalanced data sets learning[C]//International conference on intelligent computing. Springer, Berlin, Heidelberg, 2005: 878-887.\n3. Figure 2 is hard to understand.\n4. What if projecting both original data and synthetic data into 2D space for visualization, as shown in “Model-Based Oversampling for Imbalanced Sequence Classification”. \n5. How robust is the proposed algorithm when facing different levels of noise?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #5", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "1. For imbalanced learning problem, Precision cannot play a good role. Therefore, I recommend using the performance metrics F-value and G-mean to provide comprehensive assessments. \n2. In table 2 of the 5%  data imbalance, the proposed method is not as good as the baseline. Could you provide more results with different percentage of data imbalance, such as 2%, 3%, and 4%. \n2. The authors created their own baseline, and compared against it. There is plenty of baseline methods in literature to compare against such as:\n[1] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321–357, 2002. \n[2] Haibo He, Yang Bai, Edwardo A Garcia, and Shutao Li. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In 2008 IEEE International Joint Conference on Neural Networks, pp. 1322–1328. IEEE, 2008. \n[3] Han H, Wang W Y, Mao B H. Borderline-SMOTE: a new over-sampling method in imbalanced data sets learning[C]//International conference on intelligent computing. Springer, Berlin, Heidelberg, 2005: 878-887.\n3. Figure 2 is hard to understand.\n4. What if projecting both original data and synthetic data into 2D space for visualization, as shown in “Model-Based Oversampling for Imbalanced Sequence Classification”. \n5. How robust is the proposed algorithm when facing different levels of noise?\n"}, "tcdate": 1572828050584}, {"id": "HyeyClaKqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper863/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is well-written. The idea is good, but it seems like GANs have been suggested for Imbalanced data sequences before. A quick search on google, I found this paper:\n\"Multi-Task Generative Adversarial Network for Handling Imbalanced Clinical Data\" by Mina Rezaei et al., arXiv:1811.10419v1\n\nMoreover, the paper doesn't seem to be comparing their results with other state of the art imbalanced sequence classification methods. The comparisons are all between different proposed GAN methods. \n\nFor these two reasons, I do not recommend this paper for publication at this point. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper is well-written. The idea is good, but it seems like GANs have been suggested for Imbalanced data sequences before. A quick search on google, I found this paper:\n\"Multi-Task Generative Adversarial Network for Handling Imbalanced Clinical Data\" by Mina Rezaei et al., arXiv:1811.10419v1\n\nMoreover, the paper doesn't seem to be comparing their results with other state of the art imbalanced sequence classification methods. The comparisons are all between different proposed GAN methods. \n\nFor these two reasons, I do not recommend this paper for publication at this point. "}, "tcdate": 1572618439244}, {"id": "rkeFqzHiKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper863/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper consider important and interesting problem: how to generate a sequence from minority class if we want to do oversampling with synthetic data in a way similar to SMOTE. \n\nNow from the paper, the exact used approach is not clear, as the details are too scarce (see some examples below). Experiments are not convincing, as the authors don't compare to the state of the art approaches. As the exact problem statement is hard to grasp, it is hard to identify the exact contribution of authors.\nNote, that the proposed approaches, in my opinion, are not that different from approaches from modern data generation for imbalanced classification [1, 2], as we propose some kind of GAN to generate new data and so have only two variables: how we select loss for this GAN and how we select the architecture.\n\nI assume, that to be accepted at a major venue a deeper investigation is required at the moment.\n\nSee also the following comments:\n1. Figure 4: title is not required, as we have a caption with the same information. Better to use confidence bars too.\n2. The figures will benefit from usage of vector format.\n3. Figure 3: tSNE can vary from one run to another. It is better to provide at least 3 random figures or even better train e.g. a simple classification model for t-SNE embedded model and present ROC AUC scores for identification synthetic/non-synthetic.\n4. Table 1&2 formatting is different from that usually used in Academy (see e.g. https://dl.sciencesocieties.org/files/publications/style/chapter-05.pdf)\n5. F1 score is often not the best metric for imbalanced problems. The paper will benefit from providing also PR AUC (average precision) scores.\n6. Figure 2: avoid confusion matrices presented in this form, as they take much space providing almost no information. Classic tables are better.\n7. From the problem statement at the very beginning of section 2 it is not clear what kind of labels do we expect (I suppose that for each sequence we have a specific label i.e. all y_i are 0 but one, that is 1)\n8. Sometimes bigger weights for minority objects or dropping significant part of majority sequences are enough, so results for these approaches also should be included\n9. How the hyperparameters mu and lambda were selected?\n\n[1.] Guo, Ting, et al. \"Discriminative Sample Generation for Deep Imbalanced Learning.\" Proceedings of the 28th International Joint Conference on Artificial Intelligence. AAAI Press, 2019. IJCAI 2019, https://www.ijcai.org/proceedings/2019/0334.pdf\n[2.] Douzas, Georgios, and Fernando Bacao. \"Effective data generation for imbalanced learning using conditional generative adversarial networks.\" Expert Systems with applications 91 (2018): 464-471.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper consider important and interesting problem: how to generate a sequence from minority class if we want to do oversampling with synthetic data in a way similar to SMOTE. \n\nNow from the paper, the exact used approach is not clear, as the details are too scarce (see some examples below). Experiments are not convincing, as the authors don't compare to the state of the art approaches. As the exact problem statement is hard to grasp, it is hard to identify the exact contribution of authors.\nNote, that the proposed approaches, in my opinion, are not that different from approaches from modern data generation for imbalanced classification [1, 2], as we propose some kind of GAN to generate new data and so have only two variables: how we select loss for this GAN and how we select the architecture.\n\nI assume, that to be accepted at a major venue a deeper investigation is required at the moment.\n\nSee also the following comments:\n1. Figure 4: title is not required, as we have a caption with the same information. Better to use confidence bars too.\n2. The figures will benefit from usage of vector format.\n3. Figure 3: tSNE can vary from one run to another. It is better to provide at least 3 random figures or even better train e.g. a simple classification model for t-SNE embedded model and present ROC AUC scores for identification synthetic/non-synthetic.\n4. Table 1&2 formatting is different from that usually used in Academy (see e.g. https://dl.sciencesocieties.org/files/publications/style/chapter-05.pdf)\n5. F1 score is often not the best metric for imbalanced problems. The paper will benefit from providing also PR AUC (average precision) scores.\n6. Figure 2: avoid confusion matrices presented in this form, as they take much space providing almost no information. Classic tables are better.\n7. From the problem statement at the very beginning of section 2 it is not clear what kind of labels do we expect (I suppose that for each sequence we have a specific label i.e. all y_i are 0 but one, that is 1)\n8. Sometimes bigger weights for minority objects or dropping significant part of majority sequences are enough, so results for these approaches also should be included\n9. How the hyperparameters mu and lambda were selected?\n\n[1.] Guo, Ting, et al. \"Discriminative Sample Generation for Deep Imbalanced Learning.\" Proceedings of the 28th International Joint Conference on Artificial Intelligence. AAAI Press, 2019. IJCAI 2019, https://www.ijcai.org/proceedings/2019/0334.pdf\n[2.] Douzas, Georgios, and Fernando Bacao. \"Effective data generation for imbalanced learning using conditional generative adversarial networks.\" Expert Systems with applications 91 (2018): 464-471."}, "tcdate": 1571668624693}], "openreview_url": "https://openreview.net/forum?id=ryxtCpNtDS", "arxiv_id": "1901.02514", "paper_pdf": "papers/ryxtCpNtDS.pdf", "paper_pdf_sha256": "86ce9fba8d962d3b3126fb94ed8b21a1ce29bf2ea78a37a0c3953bd4b3f89a3a", "paper_pdf_bytes": 286288, "paper_pdf_source": "openreview", "code_url": "https://github.com/stephanieger/imbalanced-sequence-classification", "code_repository": "stephanieger/imbalanced-sequence-classification", "code_commit": "3534ff8930aab1a1c98e18443fc89ce2c4a84aac", "code_archive": "repos/ryxtCpNtDS.zip", "code_archive_sha256": "43e2934862da677362062e517ebb8d7cfd418654f9579688be03853784a4f3d2", "code_archive_bytes": 46025, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 96, "github_languages": {"Python": 261239}, "github_archived": false, "github_pushed_at": "2022-12-08T05:31:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/autoencoders-and-generative-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1eFtj0cKQ", "year": 2019, "status": "rejected", "title": "Generative Models from the perspective of Continual Learning", "authors": ["Timothée Lesort", "Hugo Caselles-Dupré", "Michael Garcia-Ortiz", "Jean-François Goudou", "David Filliat"], "authorids": ["timothee.lesort@thalesgroup.com", "caselles@ensta.fr", "mgarciaortiz@softbankrobotics.com", "jean-francois.goudou@thalesgroup.com", "david.filliat@ensta.fr"], "authors_source": "OpenReview API", "abstract": "Which generative model is the most suitable for Continual Learning? This paper aims at evaluating and comparing generative models on disjoint sequential image generation tasks. We investigate how several models learn and forget, considering various strategies: rehearsal, regularization, generative replay and fine-tuning. We used two quantitative metrics to estimate the generation quality and memory ability. We experiment with sequential tasks on three commonly used benchmarks for Continual Learning (MNIST, Fashion MNIST and CIFAR10). We found that among all models, the original GAN performs best and among Continual Learning strategies, generative replay outperforms all other methods. Even if we found satisfactory combinations on MNIST and Fashion MNIST, training generative models sequentially on CIFAR10 is particularly instable, and remains a challenge.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SkxPAq452m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper463/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper evaluates and compares various methods for learning GANs in a Continual Learning setting, i.e., only some of the classes are available during training. It evaluates different continual learning methods including rehearsal, EWC and generative replay applied to training several deep generative models, like GAN, CGAN, WGAN, WGAN-GP, VAE and CVAE on MNIST, Fashion MNIST and CIFAR. The authors conclude with these experimental results that generative replay is the most effective method for such a setting, and found it is difficult to generate CIFAR10 images that can be classified successfully by an image classifier.\n\nI appreciate the authors for providing so much detailed experimental results to the community, but this paper lacks novelty in general. All the CL methods the authors evaluate come from other papers that are already using these methods for generative models: rehearsal has been used in VCL Nguyen et al. (2017), EWC comes directly from Seff et al. (2017), and generative replay has been used by Wu et al. (2018a). The authors also fail to provide any valuable insight with these experimental results, e.g., analyzing why generative replay fails to improve VAEs. \n\nI expect to see more exciting results coming from the authors, but the paper is not mature enough for acceptance this time.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A comprehensive evaluation of existing methods lacking novelty and insight", "review": "This paper evaluates and compares various methods for learning GANs in a Continual Learning setting, i.e., only some of the classes are available during training. It evaluates different continual learning methods including rehearsal, EWC and generative replay applied to training several deep generative models, like GAN, CGAN, WGAN, WGAN-GP, VAE and CVAE on MNIST, Fashion MNIST and CIFAR. The authors conclude with these experimental results that generative replay is the most effective method for such a setting, and found it is difficult to generate CIFAR10 images that can be classified successfully by an image classifier.\n\nI appreciate the authors for providing so much detailed experimental results to the community, but this paper lacks novelty in general. All the CL methods the authors evaluate come from other papers that are already using these methods for generative models: rehearsal has been used in VCL Nguyen et al. (2017), EWC comes directly from Seff et al. (2017), and generative replay has been used by Wu et al. (2018a). The authors also fail to provide any valuable insight with these experimental results, e.g., analyzing why generative replay fails to improve VAEs. \n\nI expect to see more exciting results coming from the authors, but the paper is not mature enough for acceptance this time.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541192398907}, {"id": "Hyx3Jj1qnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper463/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper performs an empirical comparison of models and CL methods in a generative setting. The main motivation of the paper is to make statements about which model/method combinations are best to use for generative tasks in the CL setting. In short, the paper provides an empirical analysis and evaluation of the combination of CL methods and generative models.\n\nThe datasets used for comparison are MNIST, Fashion MNIST, and CIFAR10. For each dataset, sequential (class by class) generative tasks are introduced, aligning with the CL setting. The models investigated are VAEs, GANs, and WGANs, along with their (class) conditional counter-parts. The CL methods investigated are (i) fine-tuning (a simple baseline), (ii) rehearsal methods, (iii) elastic weight consolidation (EWC), and (iv) generative replay (GR). The authors propose to use two evaluation metrics: Fréchet Inception Distance (FID) measures the quality of the generated images, and fitting capacity (FC) measures the usefulness of the images to train classifiers.\n\nPros:\n- The authors are correct in pointing out that most of the work on CL has been restricted to the discriminative case, and that there is value in exploring generative tasks in the CL setting.\n- Empirical and experimental evaluation of this sort are useful, and help the community better understand the relationship between model, CL method, and task. Such an evaluation and in-depth analysis is welcomed in CL, especially in the generative setting.\n- The authors draw a number of useful conclusions e.g., regarding the usefulness and dangers of employing the different CL methods.\n\nCons:\n- My main concern with this paper regards the evaluation metrics used. The authors propose quality metrics for the generative model, both of which (directly or indirectly) measure the quality of the generated images. In this setting, it is unsurprising that GANs outperform VAEs, as they are known to generate higher-quality images. This however, does not necessarily mean that they are better at the continual learning task (i.e., avoiding catastrophic forgetting). It seems to me that one source from which to draw would be [1], which conducted a very rigorous and useful empirical evaluation of generative models, and the methodology followed there (i.e., evaluating marginal log-likelihoods via annealed importance sampling) would be more convincing evidence for empirical comparison of models, as it would somewhat detach the quality of the generated images from the ability of the model to avoid catastrophic forgetting.\n\nUsing their proposed image-quality metrics, the authors make statements such as: \"Our results do not give a clear distinction between conditional and unconditional models. However, adversarial methods perform significantly better than variational methods. GANs variants are able to produce better, sharper quality and variety of samples, as observed in Fig. 13 and 14 in Appendix G. Hence, adversarial methods seem more viable for CL.\" My impression is that this statement on the viability of VAEs vs GANs for CL, which is a major point of the paper, does not follow from the empirical results on the quality of the generated images. It seems quite predictable that the GAN-based models would produce higher quality images, regardless of catastrophic forgetting.\n\nAdditional (minor) comments:\n- Sec. 2 could consist of a more thorough review of the literature, with a more in-depth comparison of the different CL methods proposed and evaluated in the paper.\n- Sec. 2 contains a number of statements of the form: \"restricted to VAEs only\". For many of the cases it is not immediately clear why this is true, and in my opinion the authors should either drop those comments, or make them rigorous.\n- VCL \"use specific weights for each task, which only works for the setting where the number of tasks is known in advance\". Unclear what exactly this means or why this is true.\n- \"while the teacher retains knowledge\" - how does it \"retain knowledge\", how is this then transferred to the student, and why is this restricted to VAEs?\n\nExperimental protocol:\n- Core-sets for the rehearsal as proposed by [1] could be an interesting extension. It is unclear how the samples were selected for rehearsal, and core-sets represent a principled way to do so, that would also be interesting to compare in this setting to a random baseline.\n- For VAEs, a potentially better metric of their ability (other than the log-likelihood as suggested by [2]) would be fitting capacity (or other metric) over learned latent space rather than the reconstructed image-space.\n\nOverall, my impression is that while an empirical analysis of CL methods in the generative setting is a useful concept, the submission in its current form requires some improvement. In particular, I am worried that the choice of evaluation metrics may lead to incorrect (or partially correct) conclusions, which could of course have a negative impact on the research into CL. It also seems that the paper could use some further polishing in both writing and presentation. As such, I encourage the authors to continue the work on this empirical analysis, and perhaps submit in again to future conferences.\n\n[1] - Nguyen et al. Variational Continual Learning, ICLR 2018\n[2] - Wu et al. On the Quantitative Analysis of Decoder-Based Generative Models, ICLR 2017", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Empirical analysis of CL is welcomed, but a few concerns with the experimental set-up.", "review": "This paper performs an empirical comparison of models and CL methods in a generative setting. The main motivation of the paper is to make statements about which model/method combinations are best to use for generative tasks in the CL setting. In short, the paper provides an empirical analysis and evaluation of the combination of CL methods and generative models.\n\nThe datasets used for comparison are MNIST, Fashion MNIST, and CIFAR10. For each dataset, sequential (class by class) generative tasks are introduced, aligning with the CL setting. The models investigated are VAEs, GANs, and WGANs, along with their (class) conditional counter-parts. The CL methods investigated are (i) fine-tuning (a simple baseline), (ii) rehearsal methods, (iii) elastic weight consolidation (EWC), and (iv) generative replay (GR). The authors propose to use two evaluation metrics: Fréchet Inception Distance (FID) measures the quality of the generated images, and fitting capacity (FC) measures the usefulness of the images to train classifiers.\n\nPros:\n- The authors are correct in pointing out that most of the work on CL has been restricted to the discriminative case, and that there is value in exploring generative tasks in the CL setting.\n- Empirical and experimental evaluation of this sort are useful, and help the community better understand the relationship between model, CL method, and task. Such an evaluation and in-depth analysis is welcomed in CL, especially in the generative setting.\n- The authors draw a number of useful conclusions e.g., regarding the usefulness and dangers of employing the different CL methods.\n\nCons:\n- My main concern with this paper regards the evaluation metrics used. The authors propose quality metrics for the generative model, both of which (directly or indirectly) measure the quality of the generated images. In this setting, it is unsurprising that GANs outperform VAEs, as they are known to generate higher-quality images. This however, does not necessarily mean that they are better at the continual learning task (i.e., avoiding catastrophic forgetting). It seems to me that one source from which to draw would be [1], which conducted a very rigorous and useful empirical evaluation of generative models, and the methodology followed there (i.e., evaluating marginal log-likelihoods via annealed importance sampling) would be more convincing evidence for empirical comparison of models, as it would somewhat detach the quality of the generated images from the ability of the model to avoid catastrophic forgetting.\n\nUsing their proposed image-quality metrics, the authors make statements such as: \"Our results do not give a clear distinction between conditional and unconditional models. However, adversarial methods perform significantly better than variational methods. GANs variants are able to produce better, sharper quality and variety of samples, as observed in Fig. 13 and 14 in Appendix G. Hence, adversarial methods seem more viable for CL.\" My impression is that this statement on the viability of VAEs vs GANs for CL, which is a major point of the paper, does not follow from the empirical results on the quality of the generated images. It seems quite predictable that the GAN-based models would produce higher quality images, regardless of catastrophic forgetting.\n\nAdditional (minor) comments:\n- Sec. 2 could consist of a more thorough review of the literature, with a more in-depth comparison of the different CL methods proposed and evaluated in the paper.\n- Sec. 2 contains a number of statements of the form: \"restricted to VAEs only\". For many of the cases it is not immediately clear why this is true, and in my opinion the authors should either drop those comments, or make them rigorous.\n- VCL \"use specific weights for each task, which only works for the setting where the number of tasks is known in advance\". Unclear what exactly this means or why this is true.\n- \"while the teacher retains knowledge\" - how does it \"retain knowledge\", how is this then transferred to the student, and why is this restricted to VAEs?\n\nExperimental protocol:\n- Core-sets for the rehearsal as proposed by [1] could be an interesting extension. It is unclear how the samples were selected for rehearsal, and core-sets represent a principled way to do so, that would also be interesting to compare in this setting to a random baseline.\n- For VAEs, a potentially better metric of their ability (other than the log-likelihood as suggested by [2]) would be fitting capacity (or other metric) over learned latent space rather than the reconstructed image-space.\n\nOverall, my impression is that while an empirical analysis of CL methods in the generative setting is a useful concept, the submission in its current form requires some improvement. In particular, I am worried that the choice of evaluation metrics may lead to incorrect (or partially correct) conclusions, which could of course have a negative impact on the research into CL. It also seems that the paper could use some further polishing in both writing and presentation. As such, I encourage the authors to continue the work on this empirical analysis, and perhaps submit in again to future conferences.\n\n[1] - Nguyen et al. Variational Continual Learning, ICLR 2018\n[2] - Wu et al. On the Quantitative Analysis of Decoder-Based Generative Models, ICLR 2017", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541171939956}, {"id": "SJg-oxTunm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper463/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents an empirical evaluation of continual learning approaches for generative modelling. Noting that much of previous work focuses on supervised tasks, the paper evaluates various combinations of continual learning strategies (EWC, rehearsal/replay-based, or generative replay) and generative models (GANs or likelihood-based).\nThe experiments evaluate all combinations on MNIST and Fashion MNIST, and the resulting best-performing combination on CIFAR.\nThe paper is well-written and structured, and although there are no new proposed algorithms or measures, I think this has the potential to be a useful empirical study on the relatively unstudied topic of continual learning with generative models.\n\nHowever, my main concern is in the detail of analysis and discussion: for an empirical study, it would be much more beneficial to empirically investigate *why* certain combinations are more effective than others. For example:\n- Is the reason GANs are better than likelihood models with generative replay purely because of sample quality? Or is it sufficient for the generator to learn some key characteristics for a class that lead to sufficient discriminability?\n- Why is rehearsal better for likelihood models? (And how does this relate to the hypothesis of overfitting to a few real examples?)\n\nThe CIFAR-10 results also require more work - it is unclear why the existing approaches could not be made to work, and whether this is a fundamental deficiency in the existing approaches or other factors (hyperparameters, architecture choices, lack of time, etc). Presuming the sample quality is as good as in the WGAN-GP work (given the original implementation is used for experiments), why is this insufficient for generative replay? More detailed analysis / discussion, or another combinatorial study, would help for CIFAR too.\n\nSome comments:\n- The poor performance of EWC across the board is concerning. Was this implemented by computing the Fisher of the ELBO with respect to parameters? Was the empirical or true Fisher used? Why does the performance appear so poor compared to Seff et al (2017)? This suggests that either more thought is required on how to best protect parameters of generative models, or the baseline has not been properly implemented/tuned.\n- Given this is an entirely empirical study, I would strongly encourage the authors to release code sooner than the acceptance deadline - this can be achieved using an anonymous repository.\n- Figure 2 and 3 plots are a little difficult to parse without axis labels.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Potentially nice empirical study, but needs more work on experimental analysis and discussion", "review": "This paper presents an empirical evaluation of continual learning approaches for generative modelling. Noting that much of previous work focuses on supervised tasks, the paper evaluates various combinations of continual learning strategies (EWC, rehearsal/replay-based, or generative replay) and generative models (GANs or likelihood-based).\nThe experiments evaluate all combinations on MNIST and Fashion MNIST, and the resulting best-performing combination on CIFAR.\nThe paper is well-written and structured, and although there are no new proposed algorithms or measures, I think this has the potential to be a useful empirical study on the relatively unstudied topic of continual learning with generative models.\n\nHowever, my main concern is in the detail of analysis and discussion: for an empirical study, it would be much more beneficial to empirically investigate *why* certain combinations are more effective than others. For example:\n- Is the reason GANs are better than likelihood models with generative replay purely because of sample quality? Or is it sufficient for the generator to learn some key characteristics for a class that lead to sufficient discriminability?\n- Why is rehearsal better for likelihood models? (And how does this relate to the hypothesis of overfitting to a few real examples?)\n\nThe CIFAR-10 results also require more work - it is unclear why the existing approaches could not be made to work, and whether this is a fundamental deficiency in the existing approaches or other factors (hyperparameters, architecture choices, lack of time, etc). Presuming the sample quality is as good as in the WGAN-GP work (given the original implementation is used for experiments), why is this insufficient for generative replay? More detailed analysis / discussion, or another combinatorial study, would help for CIFAR too.\n\nSome comments:\n- The poor performance of EWC across the board is concerning. Was this implemented by computing the Fisher of the ELBO with respect to parameters? Was the empirical or true Fisher used? Why does the performance appear so poor compared to Seff et al (2017)? This suggests that either more thought is required on how to best protect parameters of generative models, or the baseline has not been properly implemented/tuned.\n- Given this is an entirely empirical study, I would strongly encourage the authors to release code sooner than the acceptance deadline - this can be achieved using an anonymous repository.\n- Figure 2 and 3 plots are a little difficult to parse without axis labels.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541095577255}], "openreview_url": "https://openreview.net/forum?id=S1eFtj0cKQ", "arxiv_id": "1812.09111", "paper_pdf": "papers/S1eFtj0cKQ.pdf", "paper_pdf_sha256": "4af51d5ff91c155bcae8ac010586b6b76a9ae40befcd9c5289ef184213b69573", "paper_pdf_bytes": 12967663, "paper_pdf_source": "openreview", "code_url": "https://github.com/TLESORT/Generative_Continual_Learning", "code_repository": "TLESORT/Generative_Continual_Learning", "code_commit": "66b121437c248993b41f154b5a2d6b7197278578", "code_archive": "repos/S1eFtj0cKQ.zip", "code_archive_sha256": "8eb5fcccad983692cb5e1089fe7478a79fcb7ff9d881575240d5f991abe9e85e", "code_archive_bytes": 565858, "code_file_count": 39, "code_extensions": {".py": 38, ".sh": 1}, "github_disk_usage_kb": 560, "github_languages": {"Python": 237477, "Shell": 4556}, "github_archived": false, "github_pushed_at": "2019-11-22T13:51:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generative-models-from-the-perspective-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dDvH0UA0CZ", "year": 2026, "status": "rejected", "title": "Sampling from Your Language Model One Byte at a Time", "authors": ["Jonathan Hayase", "Alisa Liu", "Noah A. Smith", "Sewoong Oh"], "authorids": ["~Jonathan_Hayase2", "~Alisa_Liu1", "~Noah_A._Smith2", "~Sewoong_Oh3"], "authors_source": "OpenReview API", "abstract": "Tokenization is used almost universally by modern language models, enabling efficient text representation using multi-byte or multi-character tokens. These models are typically invoked to autoregressively complete a text prompt by tokenizing the prompt, sampling more tokens to continue the tokenized prompt, and detokenizing the result. However, prior work has shown that this process can introduce distortion into the model's sampling distribution, leading to unexpected or undesirable generations. For example, users are often advised not to end their prompts with a space because it prevents the model from including the space as part of the next token. While this heuristic is effective in English, the underlying problem continues to affect languages such as Chinese as well as code generation, settings where word and syntactic boundaries may not line up with token boundaries. We present an optimal method to solve this \"Prompt Boundary Problem,\" which is based on an efficient online algorithm for Byte-Pair Encoding (BPE). This allows one to compute the next byte distribution conditioned on an arbitrary byte prefix, given only logit access to the original tokenizer-based model. This procedure can be applied iteratively to convert any autoregressive LM with a BPE tokenizer into a character-level or byte-level LM, without changing the generative distribution at the text level. We show that this significantly improves next-character prediction accuracy when computed on arbitrary prefixes. Moreover, our method is able to unify the vocabularies of language models with different tokenizers, allowing one to ensemble LMs with different tokenizers at inference time as well as transfer the post-training from one model to another using proxy-tuning. We demonstrate in experiments that the ensemble and proxy-tuned models outperform their constituents on downstream evals.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zqhy52mIP6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14331/Reviewer_KhEW"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "This paper presents an efficient algorithm to the prompt boundary problem or tokenization bias problem. \n\nLines of attack are execution speed, memory efficiency and possibly mixed generation, i.e. generating bytes only if there is a tokenization bias expected, otherwise generating tokens.\n\nNote that I do not want to score the paper at this moment I will assign scores after our discussion.", "review_text": "This paper presents an efficient algorithm to the prompt boundary problem or tokenization bias problem. \n\nLines of attack are execution speed, memory efficiency and possibly mixed generation, i.e. generating bytes only if there is a tokenization bias expected, otherwise generating tokens.\n\nNote that I do not want to score the paper at this moment I will assign scores after our discussion.", "strengths": "Making more efficient versions of exact solution to the PBP is crucial for its adoption in all decoding libraries and ultimately for its broad appeal.\n\nI will ask more clarifying questions I find the writing somewhat unclear.", "weaknesses": "I struggle a lot reading this paper. For one the contributions are not clear to me exactly i also think it overclaims here and there and the experiments are a bit poorly chosen to illustrate the contributiin. Let me get more specific ( to be clear nothing you couldnt fix in the camera ready ). \n\nIn no specific order\n\n1 in abstract be specific about what efficiecy you improve \n\n2 in contributions i would restrict my claim to the efficency bit bc as u say yourself other work has already provided exact solutiond to the problem + do not claim ensembeling that was already done in prior work or if u do how is yours better\n\n3 u claim exact solutions but never define what that means i assume u refer to phans definition of statistically equivalent u should define what u mean\n\n4 in that context u should discuss clearee what distingushes your work from phan, in line 145\n\n5 in 3 you talk about valid coverings again i must assume u take that inspired by phan that define valid encodings and covers but i dont know u need to define that, or do you mean vieira?\n\n6 proposition 3.1 is not clear the first part is a definition the second one was already shown i phan just cite theirs\n\n7 3.2 and 3.3 i will ask more clarifying questions in discussion\n\n8 experiments: from your method i come out believe you came up with a more efficient version vieira and phan hence experiments should focus on that \n\n9 what is overhead with bpe\n\n10 the experiments that do not focus on efficenty i find distracting they seem mere additional experiments illustrating prior work. If there is a result that the other two would not have produced you should compare\n\n11 4.3 in paricular i find misleading u should byte level ensemble introduced by phan, so compare to them", "questions": "Please help me understand exaclty how your algorithm is different from phan, phan also uses a tree data structure, and i am understanding u than use token level sampling but i find the sections in the paper hard to follow. Hence i find it hard to verify its correctness.\n\nSo to summarize my main concerns \n\nClarity of writing needs improvement\n\nClarity of claim ideally much less claim but more precition\n\nExperiments are too unfocussed on the problem you are trying to solve which to my understanding is the efficency part not the pbp", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an efficient algorithm to the prompt boundary problem or tokenization bias problem. \n\nLines of attack are execution speed, memory efficiency and possibly mixed generation, i.e. generating bytes only if there is a tokenization bias expected, otherwise generating tokens.\n\nNote that I do not want to score the paper at this moment I will assign scores after our discussion.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "Making more efficient versions of exact solution to the PBP is crucial for its adoption in all decoding libraries and ultimately for its broad appeal.\n\nI will ask more clarifying questions I find the writing somewhat unclear.", "weaknesses": "I struggle a lot reading this paper. For one the contributions are not clear to me exactly i also think it overclaims here and there and the experiments are a bit poorly chosen to illustrate the contributiin. Let me get more specific ( to be clear nothing you couldnt fix in the camera ready ). \n\nIn no specific order\n\n1 in abstract be specific about what efficiecy you improve \n\n2 in contributions i would restrict my claim to the efficency bit bc as u say yourself other work has already provided exact solutiond to the problem + do not claim ensembeling that was already done in prior work or if u do how is yours better\n\n3 u claim exact solutions but never define what that means i assume u refer to phans definition of statistically equivalent u should define what u mean\n\n4 in that context u should discuss clearee what distingushes your work from phan, in line 145\n\n5 in 3 you talk about valid coverings again i must assume u take that inspired by phan that define valid encodings and covers but i dont know u need to define that, or do you mean vieira?\n\n6 proposition 3.1 is not clear the first part is a definition the second one was already shown i phan just cite theirs\n\n7 3.2 and 3.3 i will ask more clarifying questions in discussion\n\n8 experiments: from your method i come out believe you came up with a more efficient version vieira and phan hence experiments should focus on that \n\n9 what is overhead with bpe\n\n10 the experiments that do not focus on efficenty i find distracting they seem mere additional experiments illustrating prior work. If there is a result that the other two would not have produced you should compare\n\n11 4.3 in paricular i find misleading u should byte level ensemble introduced by phan, so compare to them", "questions": "Please help me understand exaclty how your algorithm is different from phan, phan also uses a tree data structure, and i am understanding u than use token level sampling but i find the sections in the paper hard to follow. Hence i find it hard to verify its correctness.\n\nSo to summarize my main concerns \n\nClarity of writing needs improvement\n\nClarity of claim ideally much less claim but more precition\n\nExperiments are too unfocussed on the problem you are trying to solve which to my understanding is the efficency part not the pbp", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762043718214}, {"id": "RGuXepZXiS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14331/Reviewer_Nkha"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper presents a method for converting token-level language models to character-level (byte-level) models. The paper claims several contributions. First, an O(1) overhead algorithm for conditioning token-level LMs on character strings, a novel Valid Covering Tree concept for enumerating valid tokenizations, an efficient bigram-based canonicality test for BPE, and empirical validation showing the method achieves low error and improved bits/byte compared to canonical tokenization baselines. ByteSampler does not add high overhead relative to plain BPE decoding while substantially reducing boundary artifacts, and it evades the worst-case exponential cost of unpruned prefix-cover enumeration.", "review_text": "This paper presents a method for converting token-level language models to character-level (byte-level) models. The paper claims several contributions. First, an O(1) overhead algorithm for conditioning token-level LMs on character strings, a novel Valid Covering Tree concept for enumerating valid tokenizations, an efficient bigram-based canonicality test for BPE, and empirical validation showing the method achieves low error and improved bits/byte compared to canonical tokenization baselines. ByteSampler does not add high overhead relative to plain BPE decoding while substantially reducing boundary artifacts, and it evades the worst-case exponential cost of unpruned prefix-cover enumeration.", "strengths": "The empirical evaluation is comprehensive, testing on four modern LLMs (Llama-3.2-1B, Meta-Llama-3.1-8B, DeepSeek-R1-Distill-Llama-8B, and phi-4) with careful measurement of the speed-accuracy tradeoff. The implementation appears to be well-engineered, with the bundled beam summing approach using trie-based filtering (Appendix D) providing meaningful practical speedups. The pedagogical presentation is generally clear, with helpful visualizations like the Valid Covering Tree diagrams that make the concepts accessible. The paper demonstrates that even with modest computational budgets (small beam sizes), reasonable approximations to the character-level distribution can be achieved, which is useful for practitioners. The experimental design measuring both Jensen-Shannon distance to a reference model and bits/byte compression is informative.", "weaknesses": "The core weakness of the paper is lack of novelty:\n\nThe algorithm is similar to a previous algorithm from Vieira et al. (2025b). The paper claims throughout to provide exact solutions, but this claim is not clear as the authors redefine what \"exact\" means to be \"modulo invalid token sequences\".  The method only ensures canonicality of tokens overlapping the prompt prefix, not the entire sequence, making it an approximation to the ideal distribution. This limitation is in Appendix D.3 rather than discussed in detail. The exactness claim relies on an assumption that noncanonical strings have zero probability, but this assumption doesn't hold according to Geh et al. (2024) and Vieira et al. (2025b).\n\nThe complexity comparison in Table 1 is also not clear. It compares ByteSampler's practical O(1) overhead against Vieira et al. (2024)'s theoretical 2^O(n) worst-case bound, while not comparing against the \"beam summing algorithm\" in the same paper with O(N·K·|∆|) complexity, which is similar.. \n\nAdditionally, the bigram test is similar to previous work by Antwerpen & Neubeck (2024) and Vieira et al. (2025b).", "questions": "Did you measure the probability mass on noncanonical strings in your experiments? If so, did you consider how does considering noncanonical tokenizations affect performance on downstream tasks?\n\nHow does the bigram test differ from previous papers on canonicality?\n\nIf possible, it would be helpful to test the method to other canonicalization papers to understand the difference in runtime and accuracy, to differentiate the method more from previous work.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a method for converting token-level language models to character-level (byte-level) models. The paper claims several contributions. First, an O(1) overhead algorithm for conditioning token-level LMs on character strings, a novel Valid Covering Tree concept for enumerating valid tokenizations, an efficient bigram-based canonicality test for BPE, and empirical validation showing the method achieves low error and improved bits/byte compared to canonical tokenization baselines. ByteSampler does not add high overhead relative to plain BPE decoding while substantially reducing boundary artifacts, and it evades the worst-case exponential cost of unpruned prefix-cover enumeration.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The empirical evaluation is comprehensive, testing on four modern LLMs (Llama-3.2-1B, Meta-Llama-3.1-8B, DeepSeek-R1-Distill-Llama-8B, and phi-4) with careful measurement of the speed-accuracy tradeoff. The implementation appears to be well-engineered, with the bundled beam summing approach using trie-based filtering (Appendix D) providing meaningful practical speedups. The pedagogical presentation is generally clear, with helpful visualizations like the Valid Covering Tree diagrams that make the concepts accessible. The paper demonstrates that even with modest computational budgets (small beam sizes), reasonable approximations to the character-level distribution can be achieved, which is useful for practitioners. The experimental design measuring both Jensen-Shannon distance to a reference model and bits/byte compression is informative.", "weaknesses": "The core weakness of the paper is lack of novelty:\n\nThe algorithm is similar to a previous algorithm from Vieira et al. (2025b). The paper claims throughout to provide exact solutions, but this claim is not clear as the authors redefine what \"exact\" means to be \"modulo invalid token sequences\".  The method only ensures canonicality of tokens overlapping the prompt prefix, not the entire sequence, making it an approximation to the ideal distribution. This limitation is in Appendix D.3 rather than discussed in detail. The exactness claim relies on an assumption that noncanonical strings have zero probability, but this assumption doesn't hold according to Geh et al. (2024) and Vieira et al. (2025b).\n\nThe complexity comparison in Table 1 is also not clear. It compares ByteSampler's practical O(1) overhead against Vieira et al. (2024)'s theoretical 2^O(n) worst-case bound, while not comparing against the \"beam summing algorithm\" in the same paper with O(N·K·|∆|) complexity, which is similar.. \n\nAdditionally, the bigram test is similar to previous work by Antwerpen & Neubeck (2024) and Vieira et al. (2025b).", "questions": "Did you measure the probability mass on noncanonical strings in your experiments? If so, did you consider how does considering noncanonical tokenizations affect performance on downstream tasks?\n\nHow does the bigram test differ from previous papers on canonicality?\n\nIf possible, it would be helpful to test the method to other canonicalization papers to understand the difference in runtime and accuracy, to differentiate the method more from previous work.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761971891706}, {"id": "tj5kMzG2dY", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14331/Reviewer_E5BB"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper studies the problem of compute next-byte probabilities for autoregressive sampling from a tokenized LLMs. The authors introduces a construction of valid covering tree that improves robustness and speed over prior works. For applications, they show the advantage of this byte-sampler procedure on fixing broken tokens, model ensemble and proxy tuning.", "review_text": "This paper studies the problem of compute next-byte probabilities for autoregressive sampling from a tokenized LLMs. The authors introduces a construction of valid covering tree that improves robustness and speed over prior works. For applications, they show the advantage of this byte-sampler procedure on fixing broken tokens, model ensemble and proxy tuning.", "strengths": "The application of byte-level sampling to proxy tuning is interesting and should be further emphasized as a key contribution of the paper. \n\nThe proposed method demonstrates clear improvements over prior approaches in both speed and robustness. \n\nMoreover, the comparison between the byte-sampling strategy and the token-healing approach is well-presented and helps clarify the advantages of the proposed technique.", "weaknesses": "There are notable overlaps between the proposed method and prior work such as [1], which also introduces a similar byte-level sampling procedure and targets applications like handling broken tokens in code and ensembling LLMs. Nevertheless, [1] also has certain limitations, particularly that its pre-tokenization handling is restricted to older BPE-based tokenizers (e.g., SentencePiece), as detailed in Section C.6 where this paper also provides solution to handle such problem. In my opinion, this distinction should be emphasized more clearly in the main paper rather than being overshadowed by the ensembling experiments, so that readers without a strong tokenization background can better appreciate the robustness advantage. For this reason, it might also strengthen the paper to revise the title to highlight this aspect, for example: “Robustly Sampling from Your Language Model One Byte at a Time.”\n\nThe writeup in Section 3.2 is not very well-organized. Perhaps providing an example such as on a simple Markov chain would improve the clarity. Also, equation (2) needs more explanation so that it lines up with prior works [1,2].\n\n[1] Buu Phan, Brandon Amos, Itai Gat, Marton Havasi, Matthew J Muckley, and Karen Ullrich. Exact byte-level probabilities from tokenized language models for fim-tasks and model ensembles. In The Thirteenth International Conference on Learning Representations, 2025.\n\n[2] Tim Vieira, Ben LeBrun, Mario Giulianelli, Juan Luis Gastaldi, Brian DuSell, John Terilla, Timothy J O’Donnell, and Ryan Cotterell. From language models over tokens to language models over characters. arXiv preprint arXiv:2412.03719, 2024.", "questions": "Could you elaborate on the differences among the methods discussed in Section D.3? \n\nRegarding proxy tuning, would it be possible to incorporate additional experts or anti-experts to further enhance performance? It would be interesting to include more experimental results regarding this.\n\nIn Phan et al. [1], an O(1) procedure for next-byte sampling is also introduced. Setting aside robustness concerns such as pre-tokenization issues, could you clarify how your approach improves sampling speed compared to theirs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of compute next-byte probabilities for autoregressive sampling from a tokenized LLMs. The authors introduces a construction of valid covering tree that improves robustness and speed over prior works. For applications, they show the advantage of this byte-sampler procedure on fixing broken tokens, model ensemble and proxy tuning.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The application of byte-level sampling to proxy tuning is interesting and should be further emphasized as a key contribution of the paper. \n\nThe proposed method demonstrates clear improvements over prior approaches in both speed and robustness. \n\nMoreover, the comparison between the byte-sampling strategy and the token-healing approach is well-presented and helps clarify the advantages of the proposed technique.", "weaknesses": "There are notable overlaps between the proposed method and prior work such as [1], which also introduces a similar byte-level sampling procedure and targets applications like handling broken tokens in code and ensembling LLMs. Nevertheless, [1] also has certain limitations, particularly that its pre-tokenization handling is restricted to older BPE-based tokenizers (e.g., SentencePiece), as detailed in Section C.6 where this paper also provides solution to handle such problem. In my opinion, this distinction should be emphasized more clearly in the main paper rather than being overshadowed by the ensembling experiments, so that readers without a strong tokenization background can better appreciate the robustness advantage. For this reason, it might also strengthen the paper to revise the title to highlight this aspect, for example: “Robustly Sampling from Your Language Model One Byte at a Time.”\n\nThe writeup in Section 3.2 is not very well-organized. Perhaps providing an example such as on a simple Markov chain would improve the clarity. Also, equation (2) needs more explanation so that it lines up with prior works [1,2].\n\n[1] Buu Phan, Brandon Amos, Itai Gat, Marton Havasi, Matthew J Muckley, and Karen Ullrich. Exact byte-level probabilities from tokenized language models for fim-tasks and model ensembles. In The Thirteenth International Conference on Learning Representations, 2025.\n\n[2] Tim Vieira, Ben LeBrun, Mario Giulianelli, Juan Luis Gastaldi, Brian DuSell, John Terilla, Timothy J O’Donnell, and Ryan Cotterell. From language models over tokens to language models over characters. arXiv preprint arXiv:2412.03719, 2024.", "questions": "Could you elaborate on the differences among the methods discussed in Section D.3? \n\nRegarding proxy tuning, would it be possible to incorporate additional experts or anti-experts to further enhance performance? It would be interesting to include more experimental results regarding this.\n\nIn Phan et al. [1], an O(1) procedure for next-byte sampling is also introduced. Setting aside robustness concerns such as pre-tokenization issues, could you clarify how your approach improves sampling speed compared to theirs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761888569422}, {"id": "AVn3DwKO7l", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14331/Reviewer_Qz8v"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper presents ByteSampler, a method of treating a distribution over token sequences (an LLM) as the corresponding distribution over character strings, assuming that the probability of non-canonical token sequences in the LLM is 0. It provides three capabilities given a prompt $S$ in the form of a string of characters/bytes, as summarized in Section 3.2: (1) compute the total probability that $S$ is a prefix of a canonical token sequence generated by the LLM; (2) sample a completion of $S$ while avoiding the prompt boundary problem; and (3) compute the probability distribution over next characters/bytes (rather than tokens) following $S$, providing the ability to convert a token-level LLM to a character/byte-level one. The method works by constructing a tree of canonical token sequences where the final token straddles the boundary of $S$. The experiments show that the byte-level interface results in comparable bits per character on naturalistic data to the original token-level LLM while not adding too much overhead in terms of calls to the underlying token-level LLM. Being able to convert a token-level LLM also adds the ability to ensemble LLMs with different tokenizers or proxy-tune models with different tokenizers; and the authors present fairly strong results on these applications.", "review_text": "This paper presents ByteSampler, a method of treating a distribution over token sequences (an LLM) as the corresponding distribution over character strings, assuming that the probability of non-canonical token sequences in the LLM is 0. It provides three capabilities given a prompt $S$ in the form of a string of characters/bytes, as summarized in Section 3.2: (1) compute the total probability that $S$ is a prefix of a canonical token sequence generated by the LLM; (2) sample a completion of $S$ while avoiding the prompt boundary problem; and (3) compute the probability distribution over next characters/bytes (rather than tokens) following $S$, providing the ability to convert a token-level LLM to a character/byte-level one. The method works by constructing a tree of canonical token sequences where the final token straddles the boundary of $S$. The experiments show that the byte-level interface results in comparable bits per character on naturalistic data to the original token-level LLM while not adding too much overhead in terms of calls to the underlying token-level LLM. Being able to convert a token-level LLM also adds the ability to ensemble LLMs with different tokenizers or proxy-tune models with different tokenizers; and the authors present fairly strong results on these applications.", "strengths": "The method provides a sensible solution for the prompt boundary problem and incurs relatively little computational overhead, so it clearly has practical use. The paper is generally clear and easy to follow. The idea of converting token-level LLMs to character/byte-level ones to enable ensembling and proxy tuning of models with different tokenizers is exciting and opens up a lot of new possibilities. With the exception of TriviaQA and WikidataQA in Table 7, the experimental results in Section 4.3 and 4.4 are quite strong. The paper includes an excellent breadth of models and benchmarks.", "weaknesses": "This paper has quite a bit of overlap with two recently published papers, which I will refer to as [1] and [2]:\n\n* [1] Vieira et al. (2025). [From Language Models over Tokens to Language Models over Characters](https://openreview.net/forum?id=sQS0roNQZR)\n* [2] Vieira et al. (2025). [Language Models over Canonical Byte-Pair Encodings](https://openreview.net/forum?id=eCVrfVDNSY)\n\n[1] also provides algorithms for converting a token-level LLM to a character/byte-level one, and [2] provides algorithms for enforcing the canonicality of BPE token sequences.\n\n1. The characterization that ByteSampler provides an \"exact\" solution to the prompt boundary problem is misleading. Like this paper, [1] also provided algorithms for converting a token-level LLM to a character/byte-level one. The beam summing algorithm of [1] (approximately) marginalizes over all tokenizations that match a given character sequence, including non-canonical ones. This corresponds closely to this paper's Eq (1) (incidentally, the equation seems to be missing the probability of EOS). Crucially, ByteSampler relies on the assumption that the probability of non-canonical token sequences in the LLM is negligible. However, [2] showed that, in practice, this is often not the case. For example, LLaMa 1B and LLaMa 8B have a log-canonicality rate of about -1.2, or roughly 30% -- so about 70% of sampled token sequences are non-canonical. This means that the distribution that ByteSampler induces is considerably distorted from Eq (1) and from the actual distribution over character strings that LLMs generate in practice. The distribution of ByteSampler corresponds more closely to the \"locally canonicalized\" method of [2] (Section 3.2.2).\n1. The most comparable prior work is the beam summing algorithm of [1], but this paper does not sufficiently compare against it. It compares ByteSampler against [1] only in terms of computational cost, not the quality of the distribution over character strings, and it only compares against the exponential-time version that does not use beam search. [1] includes a beam size hyperparameter $K$ that can be used to trade computational cost for fidelity to the true marginal distribution over character sequences, in which case the time complexity is linear in the length of the character string, not exponential. A good comparison would be to show the Pareto frontier of character-normalized cross-entropy on naturalistic data vs. computational cost. Perhaps ByteSampler outperforms [1] for some range of $K$.\n1. The pairwise validation for BPE in Proposition 3.1 was already proven and used by [2], as was the trick for checking the merge trajectory along the boundary between the tokens for conflicts. The paper does not acknowledge this.\n1. A minor point: In Table 3 and similar tables, the values in the loss per unit column are not comparable, so it doesn't make sense to bold the lowest score.\n1. Very minor: Table 6 is not referenced in the main text.", "questions": "1. Is Eq (1) missing EOS?\n1. Table 3: Why is the bits per character score for ByteSampler bold?\n1. 322: How did you get this conversion rate?\n1. Table 3: Can you provide the formula for computing bits per character?\n1. 328: Are you including the probability of EOS when computing the probabilities of these documents?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents ByteSampler, a method of treating a distribution over token sequences (an LLM) as the corresponding distribution over character strings, assuming that the probability of non-canonical token sequences in the LLM is 0. It provides three capabilities given a prompt $S$ in the form of a string of characters/bytes, as summarized in Section 3.2: (1) compute the total probability that $S$ is a prefix of a canonical token sequence generated by the LLM; (2) sample a completion of $S$ while avoiding the prompt boundary problem; and (3) compute the probability distribution over next characters/bytes (rather than tokens) following $S$, providing the ability to convert a token-level LLM to a character/byte-level one. The method works by constructing a tree of canonical token sequences where the final token straddles the boundary of $S$. The experiments show that the byte-level interface results in comparable bits per character on naturalistic data to the original token-level LLM while not adding too much overhead in terms of calls to the underlying token-level LLM. Being able to convert a token-level LLM also adds the ability to ensemble LLMs with different tokenizers or proxy-tune models with different tokenizers; and the authors present fairly strong results on these applications.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The method provides a sensible solution for the prompt boundary problem and incurs relatively little computational overhead, so it clearly has practical use. The paper is generally clear and easy to follow. The idea of converting token-level LLMs to character/byte-level ones to enable ensembling and proxy tuning of models with different tokenizers is exciting and opens up a lot of new possibilities. With the exception of TriviaQA and WikidataQA in Table 7, the experimental results in Section 4.3 and 4.4 are quite strong. The paper includes an excellent breadth of models and benchmarks.", "weaknesses": "This paper has quite a bit of overlap with two recently published papers, which I will refer to as [1] and [2]:\n\n* [1] Vieira et al. (2025). [From Language Models over Tokens to Language Models over Characters](https://openreview.net/forum?id=sQS0roNQZR)\n* [2] Vieira et al. (2025). [Language Models over Canonical Byte-Pair Encodings](https://openreview.net/forum?id=eCVrfVDNSY)\n\n[1] also provides algorithms for converting a token-level LLM to a character/byte-level one, and [2] provides algorithms for enforcing the canonicality of BPE token sequences.\n\n1. The characterization that ByteSampler provides an \"exact\" solution to the prompt boundary problem is misleading. Like this paper, [1] also provided algorithms for converting a token-level LLM to a character/byte-level one. The beam summing algorithm of [1] (approximately) marginalizes over all tokenizations that match a given character sequence, including non-canonical ones. This corresponds closely to this paper's Eq (1) (incidentally, the equation seems to be missing the probability of EOS). Crucially, ByteSampler relies on the assumption that the probability of non-canonical token sequences in the LLM is negligible. However, [2] showed that, in practice, this is often not the case. For example, LLaMa 1B and LLaMa 8B have a log-canonicality rate of about -1.2, or roughly 30% -- so about 70% of sampled token sequences are non-canonical. This means that the distribution that ByteSampler induces is considerably distorted from Eq (1) and from the actual distribution over character strings that LLMs generate in practice. The distribution of ByteSampler corresponds more closely to the \"locally canonicalized\" method of [2] (Section 3.2.2).\n1. The most comparable prior work is the beam summing algorithm of [1], but this paper does not sufficiently compare against it. It compares ByteSampler against [1] only in terms of computational cost, not the quality of the distribution over character strings, and it only compares against the exponential-time version that does not use beam search. [1] includes a beam size hyperparameter $K$ that can be used to trade computational cost for fidelity to the true marginal distribution over character sequences, in which case the time complexity is linear in the length of the character string, not exponential. A good comparison would be to show the Pareto frontier of character-normalized cross-entropy on naturalistic data vs. computational cost. Perhaps ByteSampler outperforms [1] for some range of $K$.\n1. The pairwise validation for BPE in Proposition 3.1 was already proven and used by [2], as was the trick for checking the merge trajectory along the boundary between the tokens for conflicts. The paper does not acknowledge this.\n1. A minor point: In Table 3 and similar tables, the values in the loss per unit column are not comparable, so it doesn't make sense to bold the lowest score.\n1. Very minor: Table 6 is not referenced in the main text.", "questions": "1. Is Eq (1) missing EOS?\n1. Table 3: Why is the bits per character score for ByteSampler bold?\n1. 322: How did you get this conversion rate?\n1. Table 3: Can you provide the formula for computing bits per character?\n1. 328: Are you including the probability of EOS when computing the probabilities of these documents?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761829707465}], "openreview_url": "https://openreview.net/forum?id=dDvH0UA0CZ", "arxiv_id": "2506.14123", "paper_pdf": "papers/dDvH0UA0CZ.pdf", "paper_pdf_sha256": "6e2c4d13fcc52e1d60ee1148593c7210646f1c7ebaeaa0347fca218df95e973d", "paper_pdf_bytes": 437853, "paper_pdf_source": "openreview", "code_url": "https://github.com/SewoongLab/byte-sampler", "code_repository": "SewoongLab/byte-sampler", "code_commit": "f27de98fa3ad9d5df7704a31c8d30972a41d26ce", "code_archive": "repos/dDvH0UA0CZ.zip", "code_archive_sha256": "b4291266e14add6ea400764adb885c6eb11684760c47017dfbe460c9b22e3a56", "code_archive_bytes": 265698, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 361, "github_languages": {"Python": 147990}, "github_archived": false, "github_pushed_at": "2026-02-26T21:28:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sampling-from-your-language-model-one-byte-at"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Nk1MegaPuG", "year": 2025, "status": "rejected", "title": "Evading Data Contamination Detection for Language Models is (too) Easy", "authors": ["Jasper Dekoninck", "Mark Niklas Mueller", "Maximilian Baader", "Marc Fischer", "Martin Vechev"], "authorids": ["~Jasper_Dekoninck1", "~Mark_Niklas_Mueller2", "~Maximilian_Baader1", "~Marc_Fischer1", "~Martin_Vechev1"], "authors_source": "OpenReview API", "abstract": "The benchmark performance of large language models (LLMs) has a high impact on their popularity and is thus of great importance to many model providers. However, the reliability of such benchmark scores as a measure of model quality gets compromised if the model is contaminated with benchmark data. While recent contamination detection methods try to address this issue, they overlook the possibility of deliberate contamination by malicious model providers aiming to evade detection. We propose a categorization of model providers based on their (de)contamination practices and argue that malicious contamination is of crucial importance as it casts doubt on the reliability of public benchmarks. To study this issue more rigorously, we analyze current contamination detection methods based on their assumptions. This analysis reveals a significant vulnerability in existing approaches: they do not account for rephrased benchmark data used during training by malicious actors. We demonstrate how exploiting this gap can result in significantly inflated benchmark scores while completely evading current detection methods.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NKtzD55SPW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9617/Reviewer_uLUM"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper focuses on the data contamination problem with an analysis of existing frameworks. The paper categorizes existing frameworks in 4 categories and reveals their weaknesses by a detailed study. The paper proposes a new attack strategy based on the analysis and observation.", "review_text": "This paper focuses on the data contamination problem with an analysis of existing frameworks. The paper categorizes existing frameworks in 4 categories and reveals their weaknesses by a detailed study. The paper proposes a new attack strategy based on the analysis and observation.", "strengths": "1. The analysis of existing methods is reaonsable. The paper starts the discussion on categorizing existing paper with a figure and the categorization is reasonable. \n2. The selection of base models and benchmarks is representative and comprehensive.", "weaknesses": "The paper lacks novelty and technical depth, and the presentation of the paper is hard-to-follow. \n1. The paper mainly studies the existing methods, however, it lacks a detailed discussion on the proposed attack strategy and the reasonableness of such a strategy. As a reader, I cannot fully understand the main design of the strategy proposed after the analysis (In Sec 5.1 I suppose). It's hard to understand the method with a short paragraph and the paper lacks a formulation or figure for the strategy. 2. The experiments just focus on evaluating existing methods and it's hard to identify the section proving the effectiveness of the proposed strategy. The experimental design doesn't the support the claim of the paper and lacks crucial experimental results. \n3. The organization of the paper is dispersive with no key idea. The presentation is not coherent. After discussing the experiment results, the paper switched to a discussion on section 7, but such a discussion is not related to the experiment results presented before. This section sounds like a \"related work\" section.", "questions": "Please refer to the weakness part about the paper structure, technical depth, and novelty.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the data contamination problem with an analysis of existing frameworks. The paper categorizes existing frameworks in 4 categories and reveals their weaknesses by a detailed study. The paper proposes a new attack strategy based on the analysis and observation.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The analysis of existing methods is reaonsable. The paper starts the discussion on categorizing existing paper with a figure and the categorization is reasonable. \n2. The selection of base models and benchmarks is representative and comprehensive.", "weaknesses": "The paper lacks novelty and technical depth, and the presentation of the paper is hard-to-follow. \n1. The paper mainly studies the existing methods, however, it lacks a detailed discussion on the proposed attack strategy and the reasonableness of such a strategy. As a reader, I cannot fully understand the main design of the strategy proposed after the analysis (In Sec 5.1 I suppose). It's hard to understand the method with a short paragraph and the paper lacks a formulation or figure for the strategy. 2. The experiments just focus on evaluating existing methods and it's hard to identify the section proving the effectiveness of the proposed strategy. The experimental design doesn't the support the claim of the paper and lacks crucial experimental results. \n3. The organization of the paper is dispersive with no key idea. The presentation is not coherent. After discussing the experiment results, the paper switched to a discussion on section 7, but such a discussion is not related to the experiment results presented before. This section sounds like a \"related work\" section.", "questions": "Please refer to the weakness part about the paper structure, technical depth, and novelty.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730695979554}, {"id": "0Pf3X85StL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9617/Reviewer_2ndD"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper looks into the vulnerabilities and reliabilities of current data contamination detection methods. The authors categorize model providers based on their contamination practices, emphasizing malicious actors could intentionally contaminate models by rephrasing the benchmark data, thus evading detection while boosting model scores. Through experiments across several models and benchmarks, the authors show that the rephrasing technique could increase performance while bypassing existing detection methods.", "review_text": "This paper looks into the vulnerabilities and reliabilities of current data contamination detection methods. The authors categorize model providers based on their contamination practices, emphasizing malicious actors could intentionally contaminate models by rephrasing the benchmark data, thus evading detection while boosting model scores. Through experiments across several models and benchmarks, the authors show that the rephrasing technique could increase performance while bypassing existing detection methods.", "strengths": "- The paper introduces a novel categorization of model providers based on their data contamination practices.\n- Provide a simple yet effective contamination technique to evade detection, highlighting limitations in current detection methods.\n- The paper is well-organized, with each section building on the previous one. Key concepts like contamination types and detection vulnerabilities are presented accessibly, making the research easy to understand and follow.", "weaknesses": "The paper demonstrates the vulnerability of current detection methods to rephrasing attacks but does not explore potential defenses. Without suggestions for addressing this vulnerability, the work may feel incomplete.", "questions": "- Is the performance across different benchmarks consistent? Could you provide a more detailed breakdown of per-benchmark performance?\n- The paper notes that the limited sample size of the contaminated set contributes to increased variance in results. Would it be feasible to increase the sample size of the contaminated dataset to reduce this variance and enhance result stability?", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "all_content": {"summary": "This paper looks into the vulnerabilities and reliabilities of current data contamination detection methods. The authors categorize model providers based on their contamination practices, emphasizing malicious actors could intentionally contaminate models by rephrasing the benchmark data, thus evading detection while boosting model scores. Through experiments across several models and benchmarks, the authors show that the rephrasing technique could increase performance while bypassing existing detection methods.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper introduces a novel categorization of model providers based on their data contamination practices.\n- Provide a simple yet effective contamination technique to evade detection, highlighting limitations in current detection methods.\n- The paper is well-organized, with each section building on the previous one. Key concepts like contamination types and detection vulnerabilities are presented accessibly, making the research easy to understand and follow.", "weaknesses": "The paper demonstrates the vulnerability of current detection methods to rephrasing attacks but does not explore potential defenses. Without suggestions for addressing this vulnerability, the work may feel incomplete.", "questions": "- Is the performance across different benchmarks consistent? Could you provide a more detailed breakdown of per-benchmark performance?\n- The paper notes that the limited sample size of the contaminated set contributes to increased variance in results. Would it be feasible to increase the sample size of the contaminated dataset to reduce this variance and enhance result stability?", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "details_of_ethics_concerns": "By highlighting potential methods to evade current data contamination detection methods, the authors seek to promote awareness and encourage the development of more secure and reliable evaluation frameworks. While the paper aims to strengthen detection methods, these findings could be misused if taken out of context. Therefore, an ethics review may be beneficial to ensure the findings are communicated and applied responsibly.", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730670237166}, {"id": "NnSp6NiaFd", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9617/Reviewer_MpuQ"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper presents a examination of the vulnerabilities in current benchmarking practices for LLMs. They introduce Evasive Augmentation Learning (EAL), a technique that rephrases benchmark samples to boost performance on benchmarks and evade current detection methods.", "review_text": "The paper presents a examination of the vulnerabilities in current benchmarking practices for LLMs. They introduce Evasive Augmentation Learning (EAL), a technique that rephrases benchmark samples to boost performance on benchmarks and evade current detection methods.", "strengths": "- Method is simple and easy to understand.\n- The topic is highly relevant to the current challenges in assessing the performance of large language models.", "weaknesses": "- Unclear definition on \"Definition 3\"\n- The distinguish between openly malicious and evasively malicious is somewhat problematic. It's hard to find realistic usage when applying the open malicious scenario to any of the attackers. Openly malicious is easily getting rid of by the other detection method listed.\n- Sections 4 and 5 would benefit from streamlining. There is some repetition and unnecessary content that could be condensed to improve clarity and conciseness.\n- Although the paper states that code is provided, it does not include detailed descriptions of the processes for data rephrasing and contamination insertion, particularly regarding how rephrased data is generated.\n- The authors overclaimed the contribution. Failure to consider the technical feasibility and applicability of updating baseline data in real-world applications, resulting in proposed attack methods that are more suitable for closed environments and may not work in open and real-time updating environments", "questions": "- Does the EAL technique circumvent all detection methods mentioned in A Taxonomy for Data Contamination in Large Language Models, including those that do not require pre-training data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a examination of the vulnerabilities in current benchmarking practices for LLMs. They introduce Evasive Augmentation Learning (EAL), a technique that rephrases benchmark samples to boost performance on benchmarks and evade current detection methods.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- Method is simple and easy to understand.\n- The topic is highly relevant to the current challenges in assessing the performance of large language models.", "weaknesses": "- Unclear definition on \"Definition 3\"\n- The distinguish between openly malicious and evasively malicious is somewhat problematic. It's hard to find realistic usage when applying the open malicious scenario to any of the attackers. Openly malicious is easily getting rid of by the other detection method listed.\n- Sections 4 and 5 would benefit from streamlining. There is some repetition and unnecessary content that could be condensed to improve clarity and conciseness.\n- Although the paper states that code is provided, it does not include detailed descriptions of the processes for data rephrasing and contamination insertion, particularly regarding how rephrased data is generated.\n- The authors overclaimed the contribution. Failure to consider the technical feasibility and applicability of updating baseline data in real-world applications, resulting in proposed attack methods that are more suitable for closed environments and may not work in open and real-time updating environments", "questions": "- Does the EAL technique circumvent all detection methods mentioned in A Taxonomy for Data Contamination in Large Language Models, including those that do not require pre-training data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730210504373}, {"id": "XIfLZuDYkf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9617/Reviewer_ujAG"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 5, "summary": "This work aims to propose an evasive attack to bypass potential detection. The author first summarizes existing attacks, while studying potential detection strategies using metadata, reference models, semantic-preserving transformations, and different accessibility. Then, the author proposed that rephrasing can help to increase evasiveness.", "review_text": "This work aims to propose an evasive attack to bypass potential detection. The author first summarizes existing attacks, while studying potential detection strategies using metadata, reference models, semantic-preserving transformations, and different accessibility. Then, the author proposed that rephrasing can help to increase evasiveness.", "strengths": "+ This work studies a valid topic -- detecting more evasive attack to endanger LLM safety.\n+ This work summarizes many attacks and detection efforts.", "weaknesses": "- Many settings are unclear: (1) For \"contamination,\" there is no definition of the threat model regarding its knowledge, capabilities, or concrete examples in text-based attacks. (2) It is also unclear how the attacker influences benchmarks, what the adversarial samples look like, and what the adversarial objectives are. (3) Furthermore, defense settings are also unclear about the role of detector/defender and what actions can the defender take to eliminate poisoning? The author is suggested to add those information about both threat model and defender's strategies.\n\n- The paper uses overly complicated wording, which creates barriers to understanding. Specifically: (1) \"Contamination\" is typically used as \"poisoning.\" (2) The \"actor\" is generally referred to as \"attacker\" or \"adversary.\" The \"actor\" typically links to specific roles (such as specific cyber threat patterns). (3) \"De-Contamination\" is rarely used. An alternative term is \"mitigation.\" (4) \"verbatim\" could be \"word-wise.\" (5) \"Sample-level\" could be more accurately expressed as \"sentence (or word, token)-level.\" (6) It is inaccurate to use \"Problematic Data Contamination\" in Definition 3, as that definition is talking about possibility of injecting poisoning data D'. \n\n- The author did not clearly explain why pre-training requires benchmarking data. The author probably means \"public\" or \"crowdsourced\" data.\n\n- The paper over-claimed the contribution \"We define four (de-)contamination settings...\" where the settings are summarized, not defined in this paper.\n\n- The role of model providers is unclear. In section 3, model providers are both attacker and defender. The author is suggested to distinguish model providers with adversaries (the \"actor\").\n\n- The paper has poor structure and incomplete content: (1) There is no related work section. (2) The author did not detail the proposed rephrasing attack, such as what prompts or techniques are used for rephrasing. (3) Definition 4 simply repeats definition 1 and 2. (4) The Definition 5 should be highlighted with rigorous formulation with later technical design addressing it. (5) It would be better to introduce \"TPR\" (true positive rate) and \"FPR\" (false positive rate) in Figure 1.\n\n- The contribution is not significant. The author simply uses \"rephrasing,\" which lacks strong intuition and motivation for its usefulness.\n\n- The experiments are simple, and the settings are problematic. Specifically: (1) the author mentions several detection strategies but does not provide corresponding experiments to address those detections. (2) The paper also fails to clarify how evasiveness is guaranteed -- Is it measured by lower TPR by detection methods? (3) No baseline attacks. (4) The author is suggested to discuss the trade-off between attack effectiveness and evasiveness.\n\n- Many sentences seem GPT-generated. It is fine to use GPT to polish writing. However, the original writing maybe logically inconsistent and unclear, so GPT alone cannot produce better logic.", "questions": "Please see the comments above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work aims to propose an evasive attack to bypass potential detection. The author first summarizes existing attacks, while studying potential detection strategies using metadata, reference models, semantic-preserving transformations, and different accessibility. Then, the author proposed that rephrasing can help to increase evasiveness.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "+ This work studies a valid topic -- detecting more evasive attack to endanger LLM safety.\n+ This work summarizes many attacks and detection efforts.", "weaknesses": "- Many settings are unclear: (1) For \"contamination,\" there is no definition of the threat model regarding its knowledge, capabilities, or concrete examples in text-based attacks. (2) It is also unclear how the attacker influences benchmarks, what the adversarial samples look like, and what the adversarial objectives are. (3) Furthermore, defense settings are also unclear about the role of detector/defender and what actions can the defender take to eliminate poisoning? The author is suggested to add those information about both threat model and defender's strategies.\n\n- The paper uses overly complicated wording, which creates barriers to understanding. Specifically: (1) \"Contamination\" is typically used as \"poisoning.\" (2) The \"actor\" is generally referred to as \"attacker\" or \"adversary.\" The \"actor\" typically links to specific roles (such as specific cyber threat patterns). (3) \"De-Contamination\" is rarely used. An alternative term is \"mitigation.\" (4) \"verbatim\" could be \"word-wise.\" (5) \"Sample-level\" could be more accurately expressed as \"sentence (or word, token)-level.\" (6) It is inaccurate to use \"Problematic Data Contamination\" in Definition 3, as that definition is talking about possibility of injecting poisoning data D'. \n\n- The author did not clearly explain why pre-training requires benchmarking data. The author probably means \"public\" or \"crowdsourced\" data.\n\n- The paper over-claimed the contribution \"We define four (de-)contamination settings...\" where the settings are summarized, not defined in this paper.\n\n- The role of model providers is unclear. In section 3, model providers are both attacker and defender. The author is suggested to distinguish model providers with adversaries (the \"actor\").\n\n- The paper has poor structure and incomplete content: (1) There is no related work section. (2) The author did not detail the proposed rephrasing attack, such as what prompts or techniques are used for rephrasing. (3) Definition 4 simply repeats definition 1 and 2. (4) The Definition 5 should be highlighted with rigorous formulation with later technical design addressing it. (5) It would be better to introduce \"TPR\" (true positive rate) and \"FPR\" (false positive rate) in Figure 1.\n\n- The contribution is not significant. The author simply uses \"rephrasing,\" which lacks strong intuition and motivation for its usefulness.\n\n- The experiments are simple, and the settings are problematic. Specifically: (1) the author mentions several detection strategies but does not provide corresponding experiments to address those detections. (2) The paper also fails to clarify how evasiveness is guaranteed -- Is it measured by lower TPR by detection methods? (3) No baseline attacks. (4) The author is suggested to discuss the trade-off between attack effectiveness and evasiveness.\n\n- Many sentences seem GPT-generated. It is fine to use GPT to polish writing. However, the original writing maybe logically inconsistent and unclear, so GPT alone cannot produce better logic.", "questions": "Please see the comments above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729909879342}], "openreview_url": "https://openreview.net/forum?id=Nk1MegaPuG", "arxiv_id": "2402.02823", "paper_pdf": "papers/Nk1MegaPuG.pdf", "paper_pdf_sha256": "17493804626c3bf4e46d9bf8d0730495f1082805e02ca55ceae5d8fe767213c1", "paper_pdf_bytes": 1640940, "paper_pdf_source": "openreview", "code_url": "https://github.com/eth-sri/malicious-contamination", "code_repository": "eth-sri/malicious-contamination", "code_commit": "59d02ef6bd42ff1047a0520bbe58807f0324b22d", "code_archive": "repos/Nk1MegaPuG.zip", "code_archive_sha256": "69b85fd40a5aa0497cce2ae90f631693908288f012702721183d7d8030bd1416", "code_archive_bytes": 60655, "code_file_count": 30, "code_extensions": {".py": 22, ".sh": 7, ".ipynb": 1}, "github_disk_usage_kb": 51, "github_languages": {"Python": 129906, "Jupyter Notebook": 26259, "Shell": 6523}, "github_archived": false, "github_pushed_at": "2024-02-12T15:25:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evading-data-contamination-detection-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mt5NPvTp5a", "year": 2024, "status": "rejected", "title": "Improved Operator Learning by Orthogonal Attention", "authors": ["Zipeng Xiao", "Zhongkai Hao", "Bokai Lin", "Zhijie Deng", "Hang Su"], "authorids": ["~Zipeng_Xiao1", "~Zhongkai_Hao1", "~Bokai_Lin1", "~Zhijie_Deng1", "~Hang_Su3"], "authors_source": "OpenReview API", "abstract": "Neural operators, as an efficient surrogate model for learning the solutions of PDEs, have received extensive attention in the field of scientific machine learning.\nAmong them, attention-based neural operators have become one of the mainstreams in related research.\nHowever, existing approaches overfit the limited training data due to the considerable number of parameters in the attention mechanism.\nTo address this, we develop an orthogonal attention based on the eigendecomposition of the kernel integral operator and the neural approximation of eigenfunctions. \nThe orthogonalization naturally poses a proper regularization effect on the resulting neural operator, which aids in resisting overfitting and boosting generalization. \nExperiments on six standard neural operator benchmark datasets comprising both regular and irregular geometries show that our method can outperform competing baselines with decent margins.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "HwOjt9eG6I", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3527/Reviewer_DnDW"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "the paper proposed a novel network architecture for neural operators that solve PDEs. The proposed architecture incorporates an orthogonal attention mechanism to alleviate the potential overfitting issues, and improve the generalization of the operator for solving PDEs. The demonstration of the effectiveness in the experimental section is clear, and the ablation study highlights the positive contribution of the proposed orthogonal attention mechanism.", "review_text": "the paper proposed a novel network architecture for neural operators that solve PDEs. The proposed architecture incorporates an orthogonal attention mechanism to alleviate the potential overfitting issues, and improve the generalization of the operator for solving PDEs. The demonstration of the effectiveness in the experimental section is clear, and the ablation study highlights the positive contribution of the proposed orthogonal attention mechanism.", "strengths": "1. a novel attention mechanism is proposed to address the issue of overfitting, and the experimental sections show improvement over the existing attention mechanisms.\n\n2. the ablation study compares the proposed orthogonal attention mechanism to other normalization schemes, and shows the advantages of the proposed mechanism.\n\n3. scaling up the neural network with the proposed orthogonal attention mechanism brings in performance improvement, which shows that the proposed mechanism improves generalization of neural networks with various sizes.", "weaknesses": "1. I understand the motivation that the top line in figure 1 updates the PDE solution so that strong regularization is needed, and that is why the proposed orthogonal attention is incorporated, along with linear attention. However, technically, both the top line and the bottom line are simply nonlinear functions, so have the authors tried to incorporate the proposed attention into both lines to see if it further improves the generalization?\n\n2. the current parametrization requires solving a Cholesky decomposition for each batch of the data during training, and I wonder whether there could be ways to simplify the process. For example, could we fix the orthonormal matrix before training, and learn mu as a function of the input data? in that way, we could avoid taking the inverse of a whole matrix each iteration but rather taking the inverse of individual values in mu.", "questions": "n/a", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "the paper proposed a novel network architecture for neural operators that solve PDEs. The proposed architecture incorporates an orthogonal attention mechanism to alleviate the potential overfitting issues, and improve the generalization of the operator for solving PDEs. The demonstration of the effectiveness in the experimental section is clear, and the ablation study highlights the positive contribution of the proposed orthogonal attention mechanism.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. a novel attention mechanism is proposed to address the issue of overfitting, and the experimental sections show improvement over the existing attention mechanisms.\n\n2. the ablation study compares the proposed orthogonal attention mechanism to other normalization schemes, and shows the advantages of the proposed mechanism.\n\n3. scaling up the neural network with the proposed orthogonal attention mechanism brings in performance improvement, which shows that the proposed mechanism improves generalization of neural networks with various sizes.", "weaknesses": "1. I understand the motivation that the top line in figure 1 updates the PDE solution so that strong regularization is needed, and that is why the proposed orthogonal attention is incorporated, along with linear attention. However, technically, both the top line and the bottom line are simply nonlinear functions, so have the authors tried to incorporate the proposed attention into both lines to see if it further improves the generalization?\n\n2. the current parametrization requires solving a Cholesky decomposition for each batch of the data during training, and I wonder whether there could be ways to simplify the process. For example, could we fix the orthonormal matrix before training, and learn mu as a function of the input data? in that way, we could avoid taking the inverse of a whole matrix each iteration but rather taking the inverse of individual values in mu.", "questions": "n/a", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699217811760}, {"id": "JeTbRPkVov", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3527/Reviewer_rVud"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a novel neural operator architecture called Orthogonal Neural Operator (ONO) for solving partial differential equations (PDEs).\nThe key idea is to introduce an orthogonal attention mechanism that provides inherent regularization to combat overfitting when training on limited PDE data from classical numerical solvers.\nThe orthogonal attention is motivated by representing the kernel integral operator for PDE solutions using orthonormal eigenfunctions based on Mercer's theorem.\nThis eigendecomposition perspective allows defining a parameterized attention-like module with orthogonal regularization on the features.\nSpecifically, ONO consists of two pathways - one pathway uses neural networks to extract expressive features that approximate eigenfunctions, while the other pathway updates the PDE solution states based on the orthogonal attention.\nThe orthogonal attention first projects the eigenfunction-like features to an orthonormal space and then performs linear attention weighted by eigenvalues to update the solution states.\nThe orthogonalization acts as regularization and is implemented efficiently using the covariance matrix and its Cholesky decomposition.\nThe experiments are conducted on 6 benchmark datasets with regular and irregular geometries.\nThe empirical evaluations validate the effectiveness of the proposed technique over competitive baselines.", "review_text": "This paper proposes a novel neural operator architecture called Orthogonal Neural Operator (ONO) for solving partial differential equations (PDEs).\nThe key idea is to introduce an orthogonal attention mechanism that provides inherent regularization to combat overfitting when training on limited PDE data from classical numerical solvers.\nThe orthogonal attention is motivated by representing the kernel integral operator for PDE solutions using orthonormal eigenfunctions based on Mercer's theorem.\nThis eigendecomposition perspective allows defining a parameterized attention-like module with orthogonal regularization on the features.\nSpecifically, ONO consists of two pathways - one pathway uses neural networks to extract expressive features that approximate eigenfunctions, while the other pathway updates the PDE solution states based on the orthogonal attention.\nThe orthogonal attention first projects the eigenfunction-like features to an orthonormal space and then performs linear attention weighted by eigenvalues to update the solution states.\nThe orthogonalization acts as regularization and is implemented efficiently using the covariance matrix and its Cholesky decomposition.\nThe experiments are conducted on 6 benchmark datasets with regular and irregular geometries.\nThe empirical evaluations validate the effectiveness of the proposed technique over competitive baselines.", "strengths": "- This work proposes a novel orthogonal attention mechanism for neural operators that provides inherent regularization. The connection to eigendecomposition of the kernel operator is an original perspective. The authors introduces a two-pathway architecture with eigenfunction approximation and orthogonal attention-based solution update. The disentangled design is innovative.\n- Orthogonal regularization through orthogonalization of features is an interesting way to mitigate overfitting in neural operators.\n- The results demonstrate ONO achieves competitive performance by reducing the prediction error substantially compared to prior neural operator methods like FNO, Geo-FNO, LSM, etc. Further analysis shows ONO generalizes remarkably better than baselines for zero-shot super-resolution and predicting unseen time intervals. Ablation studies verify that the orthogonal attention is crucial to the performance gains.\n- Broad applicability to diverse PDEs with different geometries highlights the potential of the idea.", "weaknesses": "- The motivation of avoiding overfitting with regularization is reasonable, but the paper lacks experiments that directly demonstrate overfitting issues in baseline models to substantiate the need for orthogonal regularization. Adding such empirical analysis could strengthen the motivation.\n- While the eigendecomposition perspective provides insights, the connection to eigenfunctions is mainly conceptual. More theoretical analysis that formally relates the orthogonal attention to spectral properties could enhance the rigor.\n- The ablation study verifies the usefulness of the orthogonalization component. However, it does not isolate the impact of the two-pathway architecture itself. Additional experiments are needed to demonstrate the benefits of the disentangled design.\n- The long-term impact could be boosted by testing on real-world datasets and problems beyond standard benchmarks to showcase effectiveness in complex practical settings. The paper could also provide a more comprehensive evaluation of the model's performance across a broader spectrum of PDEs, including those with more complex boundary conditions and non-linearities. This would not only demonstrate the robustness of the model but also identify potential limitations in its current form.", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel neural operator architecture called Orthogonal Neural Operator (ONO) for solving partial differential equations (PDEs).\nThe key idea is to introduce an orthogonal attention mechanism that provides inherent regularization to combat overfitting when training on limited PDE data from classical numerical solvers.\nThe orthogonal attention is motivated by representing the kernel integral operator for PDE solutions using orthonormal eigenfunctions based on Mercer's theorem.\nThis eigendecomposition perspective allows defining a parameterized attention-like module with orthogonal regularization on the features.\nSpecifically, ONO consists of two pathways - one pathway uses neural networks to extract expressive features that approximate eigenfunctions, while the other pathway updates the PDE solution states based on the orthogonal attention.\nThe orthogonal attention first projects the eigenfunction-like features to an orthonormal space and then performs linear attention weighted by eigenvalues to update the solution states.\nThe orthogonalization acts as regularization and is implemented efficiently using the covariance matrix and its Cholesky decomposition.\nThe experiments are conducted on 6 benchmark datasets with regular and irregular geometries.\nThe empirical evaluations validate the effectiveness of the proposed technique over competitive baselines.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- This work proposes a novel orthogonal attention mechanism for neural operators that provides inherent regularization. The connection to eigendecomposition of the kernel operator is an original perspective. The authors introduces a two-pathway architecture with eigenfunction approximation and orthogonal attention-based solution update. The disentangled design is innovative.\n- Orthogonal regularization through orthogonalization of features is an interesting way to mitigate overfitting in neural operators.\n- The results demonstrate ONO achieves competitive performance by reducing the prediction error substantially compared to prior neural operator methods like FNO, Geo-FNO, LSM, etc. Further analysis shows ONO generalizes remarkably better than baselines for zero-shot super-resolution and predicting unseen time intervals. Ablation studies verify that the orthogonal attention is crucial to the performance gains.\n- Broad applicability to diverse PDEs with different geometries highlights the potential of the idea.", "weaknesses": "- The motivation of avoiding overfitting with regularization is reasonable, but the paper lacks experiments that directly demonstrate overfitting issues in baseline models to substantiate the need for orthogonal regularization. Adding such empirical analysis could strengthen the motivation.\n- While the eigendecomposition perspective provides insights, the connection to eigenfunctions is mainly conceptual. More theoretical analysis that formally relates the orthogonal attention to spectral properties could enhance the rigor.\n- The ablation study verifies the usefulness of the orthogonalization component. However, it does not isolate the impact of the two-pathway architecture itself. Additional experiments are needed to demonstrate the benefits of the disentangled design.\n- The long-term impact could be boosted by testing on real-world datasets and problems beyond standard benchmarks to showcase effectiveness in complex practical settings. The paper could also provide a more comprehensive evaluation of the model's performance across a broader spectrum of PDEs, including those with more complex boundary conditions and non-linearities. This would not only demonstrate the robustness of the model but also identify potential limitations in its current form.", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699119326271}, {"id": "N8wElD2ZvB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3527/Reviewer_bpG7"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper has proposed a new attention method to improve the generalization ability of neural operators for PDE.", "review_text": "The paper has proposed a new attention method to improve the generalization ability of neural operators for PDE.", "strengths": "1. The method is easy to understand with some theory backup.\n2. The method has shown improved performance across many datasets.", "weaknesses": "1. It is unclear how to justify that the method mitigates the overfitting problem, which is one of the central claims of the paper.\n2. The impact of the method versus the models' size should be studied.", "questions": "1. what are the performance versus model sizes?\n2. how do you measure the overfitting?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper has proposed a new attention method to improve the generalization ability of neural operators for PDE.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The method is easy to understand with some theory backup.\n2. The method has shown improved performance across many datasets.", "weaknesses": "1. It is unclear how to justify that the method mitigates the overfitting problem, which is one of the central claims of the paper.\n2. The impact of the method versus the models' size should be studied.", "questions": "1. what are the performance versus model sizes?\n2. how do you measure the overfitting?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No concern.", "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699036756816}, {"id": "nry81J0gbP", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3527/Reviewer_TYEx"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes an attention based operator learning method, where the features are orthonormalized at each layer. The orthonormalized features imply a trainable kernel and also act as a built-in regularization mechanism that aims to improve the generalization and prevent overfitting.", "review_text": "The paper proposes an attention based operator learning method, where the features are orthonormalized at each layer. The orthonormalized features imply a trainable kernel and also act as a built-in regularization mechanism that aims to improve the generalization and prevent overfitting.", "strengths": "The paper is overall well-written and focuses on the important problem of generalization and overfitting in operator learning. The theoretical build-up is insightful and the experiments show promising results, although some details are lacking.", "weaknesses": "- The model complexity is not completely clear. As pointed out by the authors, the Cholesky decomposition and also inverting the matrix $L$ are computationally expensive. While this complexity depends on the dimension $k$ of $\\hat{g}_i$, the choice of $k$ is not clear in the experiments. (I assume that the *width* in the experiments refers to the size of $d$ in $h_i^l$s.)\n- While the paper tries to address the overfitting with limited data, the experimental results concerning the dataset size are limited to only one dataset and one baseline method. A similar comparison with other methods, and if possible, with DeepONet, would help understanding the data efficiency and generalization power of these methods better.", "questions": "1. I would like to see more details and clarifications regarding the choices made for hyperparameters in the experiments, such as the size of $k$ and $d'$. \n2. A per-epoch runtime comparison with other methods would also help clarify the model complexity.\n3. I am wondering whether multi-head attention can be naturally applied in the Orthogonal Attention blocks and if the authors have considered or tried that.\n4. Right before Equation 6, the non-linearity is said to come after the residual connections and FFN. However, Equation 6 doesn't explicitly apply $\\sigma$. Does that mean that $\\sigma$ is built into the last layer of the FFN? If so, since the FFN is said to serve as the projection $\\mathcal{P}$ to the solutions in the last layer, are the last layer's outputs also passed to $\\sigma$? Wouldn't that be troublesome for the predictions?\n5. Figure 2 is a bit confusing and misleading when compared with Equation 5. A more detailed caption might help clarify the process.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an attention based operator learning method, where the features are orthonormalized at each layer. The orthonormalized features imply a trainable kernel and also act as a built-in regularization mechanism that aims to improve the generalization and prevent overfitting.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "The paper is overall well-written and focuses on the important problem of generalization and overfitting in operator learning. The theoretical build-up is insightful and the experiments show promising results, although some details are lacking.", "weaknesses": "- The model complexity is not completely clear. As pointed out by the authors, the Cholesky decomposition and also inverting the matrix $L$ are computationally expensive. While this complexity depends on the dimension $k$ of $\\hat{g}_i$, the choice of $k$ is not clear in the experiments. (I assume that the *width* in the experiments refers to the size of $d$ in $h_i^l$s.)\n- While the paper tries to address the overfitting with limited data, the experimental results concerning the dataset size are limited to only one dataset and one baseline method. A similar comparison with other methods, and if possible, with DeepONet, would help understanding the data efficiency and generalization power of these methods better.", "questions": "1. I would like to see more details and clarifications regarding the choices made for hyperparameters in the experiments, such as the size of $k$ and $d'$. \n2. A per-epoch runtime comparison with other methods would also help clarify the model complexity.\n3. I am wondering whether multi-head attention can be naturally applied in the Orthogonal Attention blocks and if the authors have considered or tried that.\n4. Right before Equation 6, the non-linearity is said to come after the residual connections and FFN. However, Equation 6 doesn't explicitly apply $\\sigma$. Does that mean that $\\sigma$ is built into the last layer of the FFN? If so, since the FFN is said to serve as the projection $\\mathcal{P}$ to the solutions in the last layer, are the last layer's outputs also passed to $\\sigma$? Wouldn't that be troublesome for the predictions?\n5. Figure 2 is a bit confusing and misleading when compared with Equation 5. A more detailed caption might help clarify the process.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698994712109}], "openreview_url": "https://openreview.net/forum?id=mt5NPvTp5a", "arxiv_id": "2310.12487", "paper_pdf": "papers/mt5NPvTp5a.pdf", "paper_pdf_sha256": "3c6067466dc44a9ea885a0eacc49070ffb35a709d441c7ba435932600ebcf5c6", "paper_pdf_bytes": 889287, "paper_pdf_source": "openreview", "code_url": "https://github.com/SJTU-DENG-Lab/Orthogonal-Neural-operator", "code_repository": "SJTU-DENG-Lab/Orthogonal-Neural-operator", "code_commit": "c9bc85dcd3630ffcb9822acc455a551cae76d662", "code_archive": "repos/mt5NPvTp5a.zip", "code_archive_sha256": "046b9e7bf4c13b010b2eb1608e30c65e3c3e16c32a9d78cb2e60bc3b24206473", "code_archive_bytes": 34180, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 42, "github_languages": {"Python": 114700}, "github_archived": false, "github_pushed_at": "2023-10-15T03:10:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improved-operator-learning-by-orthogonal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nIGza1_wxk", "year": 2023, "status": "rejected", "title": "Model Transferability with Responsive Decision Subjects ", "authors": ["Yang Liu", "Yatong Chen", "Zeyu Tang", "Kun Zhang"], "authorids": ["~Yang_Liu3", "~Yatong_Chen1", "~Zeyu_Tang1", "~Kun_Zhang1"], "authors_source": "OpenReview API", "abstract": "This paper studies model transferability when human decision subjects respond to a deployed machine learning model. In our setting, an agent or a user corresponds to a sample $(X,Y)$ drawn from a distribution $\\mathcal{D}$ and will face a model $h$ and its classification result $h(X)$. Agents can modify $X$ to adapt to $h$, which will incur a distribution shift on $(X,Y)$. Therefore, when training $h$, the learner will need to consider the subsequently ``induced\" distribution when the output model is deployed. Our formulation is motivated by applications where the deployed machine learning models interact with human agents, and will ultimately face \\emph{responsive} and interactive data distributions. We formalize the discussions of the transferability of a model by studying how the model trained on the available source distribution (data) would translate to the performance on the induced domain. We provide both upper bounds for the performance gap due to the induced domain shift, as well as lower bound for the trade-offs that a classifier has to suffer on either the source training distribution or the induced target distribution. We provide further instantiated analysis for two popular domain adaptation settings with covariate shift and target shift.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "o6W2VHseOS0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3030/Reviewer_Cedq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThis paper formulates a very interesting and novel problem (IDA, induced domain adapatation) in transfer learning. Consider the supervised classification setting where one usually trains a classifier $h : X \\mapsto Y$ from samples $\\{(X, Y)\\}$ drawn from some distribution $\\sim \\mathcal D$. Oftentimes, when the data are generated from human input, the human could possibly modify $(X, Y)$ to adapt to $h$, resulting in a distribution shift over $\\mathcal D$. Therefore, when training $h$, it is important to take this distribution shift into consideration. However, this can become complicated and interactive if the human further adapt to $h$.\n\nThis paper conducts a rather detailed study on this problem. It proves upper and lower bounds for the transfer risks for several important fundamental questions in this setting, as outlined in page 2. Besides, the paper realizes their bounds by both showing how to compute them in practice and computing them on real datasets.\n", "review_text": "This paper has good theoretical and empirical results, yet the techniques might not be exciting and the experiments might not be convincing enough due to paucity of datasets evaluated.", "strengths": "\n### Strength\n\n1. The setting studied in this paper is important, interesting, and novel.\n2. This paper studies most fundamental questions in this setting and presents satisfactory results that covers several fundamental lower and upper bounds in this setting.\n3. As a mainly theoretical paper, this paper is aware of the practical impact of their theoretical results. To this end, this paper shows how to compute their upper bounds in real-world problems and conducts experiments to demonstrate their results. Furthermore, the experimental results corroborates their theoretical results and suggests that the theoretical results could give meaningful upper and lower bounds for the transfer risks.\n\n### Weakness\n\n1. The proof techniques look simple and the techniques themselves might not inspire broader community.\n2. The authors could possibly conduct experiments on more datasets to strengthen their experimental results.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "\nThis paper formulates a very interesting and novel problem (IDA, induced domain adapatation) in transfer learning. Consider the supervised classification setting where one usually trains a classifier $h : X \\mapsto Y$ from samples $\\{(X, Y)\\}$ drawn from some distribution $\\sim \\mathcal D$. Oftentimes, when the data are generated from human input, the human could possibly modify $(X, Y)$ to adapt to $h$, resulting in a distribution shift over $\\mathcal D$. Therefore, when training $h$, it is important to take this distribution shift into consideration. However, this can become complicated and interactive if the human further adapt to $h$.\n\nThis paper conducts a rather detailed study on this problem. It proves upper and lower bounds for the transfer risks for several important fundamental questions in this setting, as outlined in page 2. Besides, the paper realizes their bounds by both showing how to compute them in practice and computing them on real datasets.\n", "strength_and_weaknesses": "\n### Strength\n\n1. The setting studied in this paper is important, interesting, and novel.\n2. This paper studies most fundamental questions in this setting and presents satisfactory results that covers several fundamental lower and upper bounds in this setting.\n3. As a mainly theoretical paper, this paper is aware of the practical impact of their theoretical results. To this end, this paper shows how to compute their upper bounds in real-world problems and conducts experiments to demonstrate their results. Furthermore, the experimental results corroborates their theoretical results and suggests that the theoretical results could give meaningful upper and lower bounds for the transfer risks.\n\n### Weakness\n\n1. The proof techniques look simple and the techniques themselves might not inspire broader community.\n2. The authors could possibly conduct experiments on more datasets to strengthen their experimental results.\n", "clarity,_quality,_novelty_and_reproducibility": "\nThe writing is good. No apparaent typos are noticed. The authors supplies sufficient materials (codes) for reproducing the experimental results in this paper.\n", "summary_of_the_review": "This paper has good theoretical and empirical results, yet the techniques might not be exciting and the experiments might not be convincing enough due to paucity of datasets evaluated.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667573107991}, {"id": "nlcuvld8yE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3030/Reviewer_vuF4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This papers studies the performance of models on the distribution induced by the model itself. The paper first provides examples of when this situation arises. Then, they develop formal definitions of metrics that measures the performance of the model on the induced distributions. These new quantities are inspired by the theory of domain adaptation (DA), like the induced risks that the model should minimize, which is similar to the target risk. Then authors show lower and upper bounds of these quantities. Afterwards, authors use different examples to illustrate their bounds. Finally, their bounds are computed on a real-world dataset.", "review_text": "I agree that the problem is appealing. The defined quantites and theorems are also interesting. The main downside is that I would have expected a training procedure to minimize the induced risk. While authors discuss the challenges of this question in appendix, it should also be discussed in the main paper.\n\nThe lack of training procedure to minimize the induced risk and the lack of empirical contributions make that I recommend 'marginally below the acceptance threshold'. I would be happy to increase my score if authors provide additionnal details/toy experiments on how to minimize the induced risk.", "strengths": "This paper studies a new problem and defines new quantities to understand and study it. It makes the paper mostly theoretical and related to the statistic field. Authors define many quantities inspired from DA. However as in this problem we do not have access to target samples unlike in DA, this problem seems more related to domain generalization (DG). I wonder if some theoretical or training procedure developed in DG could be applied to this problem and maybe this should be discussed in the paper. \n\nRegarding the bounds, they are explained and applied to several examples. They are also evaluated on the FICO credit score dataset.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This papers studies the performance of models on the distribution induced by the model itself. The paper first provides examples of when this situation arises. Then, they develop formal definitions of metrics that measures the performance of the model on the induced distributions. These new quantities are inspired by the theory of domain adaptation (DA), like the induced risks that the model should minimize, which is similar to the target risk. Then authors show lower and upper bounds of these quantities. Afterwards, authors use different examples to illustrate their bounds. Finally, their bounds are computed on a real-world dataset.", "strength_and_weaknesses": "This paper studies a new problem and defines new quantities to understand and study it. It makes the paper mostly theoretical and related to the statistic field. Authors define many quantities inspired from DA. However as in this problem we do not have access to target samples unlike in DA, this problem seems more related to domain generalization (DG). I wonder if some theoretical or training procedure developed in DG could be applied to this problem and maybe this should be discussed in the paper. \n\nRegarding the bounds, they are explained and applied to several examples. They are also evaluated on the FICO credit score dataset.", "clarity,_quality,_novelty_and_reproducibility": " The related work is, to the best of my knowledge, complete. Regarding discussion, I think a short discussion about the challenge to minimize induced risk should be added in the main paper and it should refer to appendix D for full discussion. The paper is well written. A short paragraph about the difference in setting between domain adaptation and the considered problem could also be useful. In particular, the lack of access to target samples. ", "summary_of_the_review": "I agree that the problem is appealing. The defined quantites and theorems are also interesting. The main downside is that I would have expected a training procedure to minimize the induced risk. While authors discuss the challenges of this question in appendix, it should also be discussed in the main paper.\n\nThe lack of training procedure to minimize the induced risk and the lack of empirical contributions make that I recommend 'marginally below the acceptance threshold'. I would be happy to increase my score if authors provide additionnal details/toy experiments on how to minimize the induced risk.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667415209723}, {"id": "n9EJlKj1iiE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3030/Reviewer_Rs7K"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies model transferability when human decision subjects respond to the deployed model. It provides a series of lower / upper bound between empirical risk minimizer classifier and the optimal classifier. In particular, teh paper studies two common cases: covariate shift and target shift.", "review_text": "The paper gives a series of bounds for the model transferability, the proofs I checked (non-exhaustive) are all good. It would be nice to see more discussions on the tightness of these bounds, as in the general case, the positiveness of the lower bound is not even clear.", "strengths": "pros：\nthe paper is clearly written, and the introduction of the problem is well motivated. The organization of the paper also makes it easy to follow, with clear notations / explanations to the theorems etc. \n\ncons:\nAs the authors write in Section 3.3, the bounds in the paper is mainly of theoretic interest. I would like to see some disucssions on the tightness of these bounds, even a single example would be helpful. The mathematical tools used are also not deep.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies model transferability when human decision subjects respond to the deployed model. It provides a series of lower / upper bound between empirical risk minimizer classifier and the optimal classifier. In particular, teh paper studies two common cases: covariate shift and target shift.", "strength_and_weaknesses": "pros：\nthe paper is clearly written, and the introduction of the problem is well motivated. The organization of the paper also makes it easy to follow, with clear notations / explanations to the theorems etc. \n\ncons:\nAs the authors write in Section 3.3, the bounds in the paper is mainly of theoretic interest. I would like to see some disucssions on the tightness of these bounds, even a single example would be helpful. The mathematical tools used are also not deep.\n", "clarity,_quality,_novelty_and_reproducibility": "Overall the paper is pretty clear. My main concern is still the use of these bounds. In the scenarios described in Section3.3, I would be interested in knowing\n1. if the provided bound tight in any sense for the empirical risk minimizing model h_S?\n2. as the authors write in Section 6  \"indicating the suboptimality of training on D_S\", is there any other learning algorithm that provides better bounds? Intuitively, training with a penalty term would likely give better results.\n\nOf course it may involve more assumptions, but I think a tightness result, even a asymmpototic one would be very useful.", "summary_of_the_review": "The paper gives a series of bounds for the model transferability, the proofs I checked (non-exhaustive) are all good. It would be nice to see more discussions on the tightness of these bounds, as in the general case, the positiveness of the lower bound is not even clear.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667357127125}, {"id": "gG7VnspL4As", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3030/Reviewer_tvWE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors explore what guarantees we can identify on the performance of classifiers in response to classification subjects which act strategically - that is, attempt to update their inputs so as to achieve a better output. They call this loss the \"induced risk\": the risk the model incurs on the input distribution which its deployment induces. Their results are mostly of two forms: upper bounds on the induced risk above the optimal induced risk, and lower bounds on the minimum of the source risk and the induced risk. The machinery used for their proofs is mostly domain-adaptation-inspired, where the source distribution is the original data, and the target distribution is the strategically modified data.", "review_text": "Overall, this paper provides some useful results in the strategic classification space. I have some doubts about the applicability of the covariate shift/label shift assumptions to this setting, and therefore not sure how useful those results are. Additionally, I have some confusion around other assumptions made around the data generative process with replicator dynamics and the experiments. Due to some good results and exposition alongside these concerns, I'm recommending a Weak Reject for now.", "strengths": "Strengths:\n- These are useful results and I think more general than what has previously been shown in strategic classification\n- paper is mostly pretty clear and exposition is good\n\nFeedback: \n-Sec 4 and 5: it isn't clear to me why Covariate/Target shift would be useful assumptions for the strategic classification setting. One would imagine that the induced distribution will have a much lower P(Y = 1 | X) for X which are right over the decision boundary, for instance; or that P(X | Y = 1) would cluster more closely near the decision boundary in the induced distribution. I'm happy to be convinced otherwise, but it seems to me that by the definition of strategic classification, these results aren't so useful.\n-Sec 4.2: it would be good to explain why these might be natural assumptions: I think I understand it has something to do with people choosing to move their inputs towards the desirable outcome Y=1, but not totally sure\nSec 4.3, Setup 3/4: I think there needs to be a little more definition around how the adapted feature is generated. For instance, if \\tau = 0.5, x = 0.4, and B = 0.2, then the probability of a successful update is 0.5, but it is drawn from the distribution U(0.5, 0,3), which doesn't obviously make sense to me.\n-Prop 4.7: I don't quite understand why as Err_D_S(h*_T) goes up in this bound, the gap increases - it would be nice to get some intuition on this relationship\nReplicator Dynamics: It would be good to have a better description of this in the body of the paper, I'm not so clear on exactly how this works\nExperiments: I don't totally understand the data generation process, even after looking at the supplement. It would be good to have a clearer explanation of the exact strategic modification model here (e.g. what is \\epsilon or \\alpha, as defined in the supplement)\n\nNotes:\nEq (3) - would be good to define the optimal model here: it's not clear at this point in the paper if h*_T is the model with the optimal induced risk, or optimal source risk (you explain later, but should explain here)\nSec 4.2 - should define X_+(h) and X_-(h)\nSec 5: a little bit confusing to me to overload w here as the induced positive rate, when it was the weighting coefficient earlier\nThm 5.2: what is p here?\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the authors explore what guarantees we can identify on the performance of classifiers in response to classification subjects which act strategically - that is, attempt to update their inputs so as to achieve a better output. They call this loss the \"induced risk\": the risk the model incurs on the input distribution which its deployment induces. Their results are mostly of two forms: upper bounds on the induced risk above the optimal induced risk, and lower bounds on the minimum of the source risk and the induced risk. The machinery used for their proofs is mostly domain-adaptation-inspired, where the source distribution is the original data, and the target distribution is the strategically modified data.", "strength_and_weaknesses": "Strengths:\n- These are useful results and I think more general than what has previously been shown in strategic classification\n- paper is mostly pretty clear and exposition is good\n\nFeedback: \n-Sec 4 and 5: it isn't clear to me why Covariate/Target shift would be useful assumptions for the strategic classification setting. One would imagine that the induced distribution will have a much lower P(Y = 1 | X) for X which are right over the decision boundary, for instance; or that P(X | Y = 1) would cluster more closely near the decision boundary in the induced distribution. I'm happy to be convinced otherwise, but it seems to me that by the definition of strategic classification, these results aren't so useful.\n-Sec 4.2: it would be good to explain why these might be natural assumptions: I think I understand it has something to do with people choosing to move their inputs towards the desirable outcome Y=1, but not totally sure\nSec 4.3, Setup 3/4: I think there needs to be a little more definition around how the adapted feature is generated. For instance, if \\tau = 0.5, x = 0.4, and B = 0.2, then the probability of a successful update is 0.5, but it is drawn from the distribution U(0.5, 0,3), which doesn't obviously make sense to me.\n-Prop 4.7: I don't quite understand why as Err_D_S(h*_T) goes up in this bound, the gap increases - it would be nice to get some intuition on this relationship\nReplicator Dynamics: It would be good to have a better description of this in the body of the paper, I'm not so clear on exactly how this works\nExperiments: I don't totally understand the data generation process, even after looking at the supplement. It would be good to have a clearer explanation of the exact strategic modification model here (e.g. what is \\epsilon or \\alpha, as defined in the supplement)\n\nNotes:\nEq (3) - would be good to define the optimal model here: it's not clear at this point in the paper if h*_T is the model with the optimal induced risk, or optimal source risk (you explain later, but should explain here)\nSec 4.2 - should define X_+(h) and X_-(h)\nSec 5: a little bit confusing to me to overload w here as the induced positive rate, when it was the weighting coefficient earlier\nThm 5.2: what is p here?\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: decent clarity around the exposition and proofs, some confusion around replicator dynamics and experiments\nQuality: I think the work is solid, however I'm not sure that the assumptions relied on by the proofs are all appropriate\nNovelty: I think this is novel because it uses lighter assumptions (mostly) than previous strategic classification work\nReproducibility: I have a little confusion around the experiments at the moment, with a bit more information this could be reproducible", "summary_of_the_review": "Overall, this paper provides some useful results in the strategic classification space. I have some doubts about the applicability of the covariate shift/label shift assumptions to this setting, and therefore not sure how useful those results are. Additionally, I have some confusion around other assumptions made around the data generative process with replicator dynamics and the experiments. Due to some good results and exposition alongside these concerns, I'm recommending a Weak Reject for now.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666625049493}], "openreview_url": "https://openreview.net/forum?id=nIGza1_wxk", "arxiv_id": "2107.05911", "paper_pdf": "papers/nIGza1_wxk.pdf", "paper_pdf_sha256": "c2b6bca7159c19e9e0f659b7920d3c53865250ed8b1057b3c6b1fd8d6868e45e", "paper_pdf_bytes": 1596478, "paper_pdf_source": "openreview", "code_url": "https://github.com/UCSC-REAL/Model_Transferability", "code_repository": "UCSC-REAL/Model_Transferability", "code_commit": "25eefb48e784dc4f808b74ec902fdeba7fb7ab0c", "code_archive": "repos/nIGza1_wxk.zip", "code_archive_sha256": "b769c8b92240ce3bfbfc3666ea3b35c86d327b43ae39dddb41e4853f56b9a6fd", "code_archive_bytes": 50657, "code_file_count": 8, "code_extensions": {".py": 7, ".ipynb": 1}, "github_disk_usage_kb": 56, "github_languages": {"Jupyter Notebook": 124753, "Python": 61416}, "github_archived": false, "github_pushed_at": "2023-05-31T11:36:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/induced-domain-adaptation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jNB6vfl_680", "year": 2022, "status": "rejected", "title": "Global Magnitude Pruning With Minimum Threshold Is All We Need", "authors": ["Manas Gupta", "Vishandi Rudy Keneta", "Abhishek Vaidyanathan", "Ritwik Kanodia", "Efe Camci", "Chuan-Sheng Foo", "Jie Lin"], "authorids": ["~Manas_Gupta1", "~Vishandi_Rudy_Keneta1", "~Abhishek_Vaidyanathan1", "~Ritwik_Kanodia1", "~Efe_Camci1", "~Chuan-Sheng_Foo1", "~Jie_Lin1"], "authors_source": "OpenReview API", "abstract": "Neural network pruning remains a very important yet challenging problem to solve. Many pruning solutions have been proposed over the years with high degrees of algorithmic complexity. In this work, we shed light on a very simple pruning technique that achieves state-of-the-art (SOTA) performance. We showcase that magnitude based pruning, specifically, global magnitude pruning (GP) is sufficient to achieve SOTA performance on a range of neural network architectures. In certain architectures, the last few layers of a network may get over-pruned. For these cases, we introduce a straightforward method to mitigate this. We preserve a certain fixed number of weights in each layer of the network to ensure no layer is over-pruned. We call this the Minimum Threshold (MT). We find that GP combined with MT when needed, achieves SOTA performance on all datasets and architectures tested including ResNet-50 and MobileNet-V1 on ImageNet. Code available on github.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "yk3qGpNokH2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1229/Reviewer_YX1q"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper revisits a traditional model compression method, global magnitude pruning (GP), and shows that GP can achieves state-of-the-art. The paper further improves GP by introducing minimum threshold (MT). Experiments on ResNet show that GP+MT achieves better accuracy with the same sparsity ratio compared to GP.\n\n", "review_text": "1) My major concern is that the paper does not clearly explain differences between GP in this paper and traditional GP methods. Based on Equation (1), I think GP in this paper is the same as GP adopted in Han et al.(2015a). If GP in this paper can achieve SOTA, Han's method can also achieve SOTA. My suggestion is that the authors provide qualitative or quantitative comparison with previous GP methods.\n\n2) The proposed MT is not a sufficient improvement. First, GP+MT increases accuracy on MobileNet, while slightly decreasing accuracy on ResNet. Second, I am not sure whether MT is sensitive to selected threshold. \n\n3) Since CIFAR-10 is a tiny-scale dataset, I think experiments on CIFAR-10 cannot sufficiently validate that for WideResNet-22-8, GP+MT can further improve accuracy compared to GP with the same sparsity ratio.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper revisits a traditional model compression method, global magnitude pruning (GP), and shows that GP can achieves state-of-the-art. The paper further improves GP by introducing minimum threshold (MT). Experiments on ResNet show that GP+MT achieves better accuracy with the same sparsity ratio compared to GP.\n\n", "main_review": "1) My major concern is that the paper does not clearly explain differences between GP in this paper and traditional GP methods. Based on Equation (1), I think GP in this paper is the same as GP adopted in Han et al.(2015a). If GP in this paper can achieve SOTA, Han's method can also achieve SOTA. My suggestion is that the authors provide qualitative or quantitative comparison with previous GP methods.\n\n2) The proposed MT is not a sufficient improvement. First, GP+MT increases accuracy on MobileNet, while slightly decreasing accuracy on ResNet. Second, I am not sure whether MT is sensitive to selected threshold. \n\n3) Since CIFAR-10 is a tiny-scale dataset, I think experiments on CIFAR-10 cannot sufficiently validate that for WideResNet-22-8, GP+MT can further improve accuracy compared to GP with the same sparsity ratio.", "summary_of_the_review": "The paper does not clearly explain differences between GP in this paper and traditional GP methods. The proposed MT is not a sufficient improvement. So my rating is \"5: marginally below the acceptance threshold\"", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635979360666}, {"id": "-VmicmQ5F25", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1229/Reviewer_5P4J"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "A very simple and effective pruning method is proposed. Instead of layer-wise pruning, a global threshold is used to prune weights according to their magnitude. In addition, to avoid over-pruning, a minimum number of parameters is preserved for each layer after pruning. Experiments on CIFAR-10 and ImageNet validate the effectiveness of proposed method. ", "review_text": "* Layer-wise sparsity ratios used to be tricky hyper-parameters. A global threshold and a minimum number of preserved weights eliminate layer-wise sparsity ratios. In this context, the proposed method is practical. \n\n* Although the empirical results seem promising despite of the simplicity. it is not straightforward to understand why we should use global threshold across layers and why MT is crucial to ensure superior performance for certain models. For example, how does over-pruning impact the performance? \n\n* A related study [1] argues that layer-wise sparsity is of importance. A comparison (both theoretical and empirical) should make the result stronger. \n\n\n[1] Lee, Jaeho, et al. \"Layer-adaptive Sparsity for the Magnitude-based Pruning.\" International Conference on Learning Representations. 2020. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "A very simple and effective pruning method is proposed. Instead of layer-wise pruning, a global threshold is used to prune weights according to their magnitude. In addition, to avoid over-pruning, a minimum number of parameters is preserved for each layer after pruning. Experiments on CIFAR-10 and ImageNet validate the effectiveness of proposed method. ", "main_review": "* Layer-wise sparsity ratios used to be tricky hyper-parameters. A global threshold and a minimum number of preserved weights eliminate layer-wise sparsity ratios. In this context, the proposed method is practical. \n\n* Although the empirical results seem promising despite of the simplicity. it is not straightforward to understand why we should use global threshold across layers and why MT is crucial to ensure superior performance for certain models. For example, how does over-pruning impact the performance? \n\n* A related study [1] argues that layer-wise sparsity is of importance. A comparison (both theoretical and empirical) should make the result stronger. \n\n\n[1] Lee, Jaeho, et al. \"Layer-adaptive Sparsity for the Magnitude-based Pruning.\" International Conference on Learning Representations. 2020. ", "summary_of_the_review": "The simplicity and effectiveness of proposed method are important and interesting. The findings will inspire more studies on the nature of network pruning. However, it is not easy for readers to understand where the effectiveness of proposed method comes. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635954090866}, {"id": "GQ7wdtM1_9J", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1229/Reviewer_MHY2"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper revisits Global Pruning (GP) and makes a case to consider it seriously as part of the pruning literature. The authors also present an addition to GP, that claims, to further increase accuracy and reliability called Minimum Threshold (MT). MT is the constraint put on every layer of the neural network to ensure they are not pruned beyond a certain limit which might result in catastrophic accuracy drops. \n\nThe paper presents experiments on CIFAR-10 with a few CNN architectures along with current pruning baselines (layer-wise mostly. They also show results on ImageNet using ResNet50 and MovileNetV1. Following these experiments, the authors claim that GP or GPMT are SOTA for pruning despite their simplicity. \n\nThe paper also includes some discussion on how output patterns of each layer change with different pruning schemes and a note about how to set the optimal MT values. \n\nWhile I agree with the sentiment of the paper, I do not think the paper is presenting anything new (even from a benchmarking perspective) but rather is missing the point about what makes GP vs layer-wise pruning a worthy trade-off often. I will list my concerns in the main review.", "review_text": "I will go sequentially in the paper but will start with the strengths follow it up with the weaknesses: \n\nOverall, the writing quality is OK and is easy to follow. The information has been well put together for a new reader to understand without having to worry about the background. The authors should be appreciated for that.  \n\nStrengths:\n1) The fact that authors revisited GP to show its power is a very nice step. GP has been (according to me) the strongest baselines for pruning for a given parameter budget even now (has been acknowledged in multiple papers). \n2) The idea of MT makes sense and is a valid fix for the overly pruned layers in general.\n3) The experiments on ImageNet are commended because the insights from there a more valuable than other things presented in this paper. \n4) The discussion on network architectures, outputs and settings of MT are interesting in general. \n5) I really liked the related work, it is thorough from a literature review standpoint and would recommend new readers to go through it for a nice picture of the field \n\nWeaknesses:\nThis might sound a bit critical, but I would appreciate it if the authors understand the merit of whatever I am trying to put forward here. \n\n1) No one ever claimed GP is not sufficient to achieve SOTA performance on pruning. GP is extremely powerful no matter what gives the highest accuracy (often) for the same #params even at the highest sparsity levels. GP is not forgotten or overlooked by any means. The reason why people work on layer-wise pruning is slightly different and I will come to it later. \n2) While the authors mentioned people of LTH community probably used GP as one of the potential pruning techniques, GP was the only one that resulted in strong tickets at higher sparsities, making IMP explicitly work on GP no matter what other factors are. \n3) MT comes into the picture only when there is one-shot pruning and GP has been shown to work great even with gradual schedules (IMP is an example) and will not suffer from pruning out entire layers just because of the gradual processing. I would even argue setting the small pruning schedule is data-independent and has a similar cost as finding the optimal MT as done in this paper. MT makes sense for one-shot pruning, but one-shot pruning itself is problematic (will explain in a bit).\n4) Table 1 is a poor representation, I would say, STR and GMP are extremely easy to implement. GMP could be done in One-shot and GMP is data-independent no matter what. I would like to hear authors' thoughts on this. \n5) Coming to the setup, the biggest problem I have with the way GP or GPMT is being used is the application on top of a pre-trained network. Eg. On imagenet the fine-tuning of GP or GPMT takes 20 extra epochs, which isn't the case with all the baselines in the discussion. All of them are pruning while training baseline that only takes 100 epochs at best. I personally have adapted GP to be pruning while training like with IMP and GP does amazingly well, so the aspect of one-shot is overshadowed by the extra cost involved in fine-tuning. I don't know how we can make a case for these trade-offs to make the comparison fair.\n6) The big one, the paper completely disregards the compute cost during inference (FLOPs) for these pruned models. The biggest difference GP and layer-wise pruning methods have is the inference costs involved. Please see STR or RigL papers to see about that axis of comparison. GP often has 2x more inference FLOPs than uniform-sparsity with very little gain in accuracy. If uniform sparsity was adjusted for that compute costs, the accuracy will be much higher. This paper, while being a study on GP, completely overlooks this aspect of GP or GPMT.\n7) While the results on CIFAR-10 are good for debugging and understanding, the insights aren't particularly transferable. I would also not read too much into 40% sparsity results or even 95% sparsity because the networks involved are so heavily overparameterized that even at that sparsity level CIFAR-10 seems like a simple task. However, I would like to say Table 5 is a real thing for one-shot pruning, but will not happen for gradual GP. I would recommend authors implement gradual GP to see what it brings to the table, along with reduced training costs (debatable because one can think of pre-trained models as free). \n8) Figures 2 and 3 just reinforce what I am saying above. \n9) The gains in Tables 6 and 7 are not significant enough to make sweeping claims on SOTA while also not measuring FLOPs.\n10) The use of \"huge margin\" is highly discouraged in Table 8 because the gains are very insignificant and trust me the FLOPs are going to be so exorbitantly high that the accuracy vs # params doesn't matter in that case. \n11) As mentioned in the paper, MT doesn't seem to bring anything to the table for ImageNet experiments, just because of how the weight norms across layers change. MT is probably not a valuable addition in this case. \n12) I would not get into the discussion of output preservation because, it doesn't necessarily mean that the model is doing well if its outputs are closer to some other model, the yardstick itself is flaky here. \n13) Lastly, the discussion of MT tries to make it feel like it is not complex or data-dependent. But the same paper says GMP schedule is complex. I would say searching for MT is as good as coming up with a cubic pruning pattern which will even help GP when done in a gradual fashion.\n\nOverall, I do not think the paper brings any new observations into the limelight and also ignores the issues present with whatever has been presented. The paper is not ready for publication, but I am happy to chat with the authors to help them understand and (myself) gain perspective if I am missing something. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper revisits Global Pruning (GP) and makes a case to consider it seriously as part of the pruning literature. The authors also present an addition to GP, that claims, to further increase accuracy and reliability called Minimum Threshold (MT). MT is the constraint put on every layer of the neural network to ensure they are not pruned beyond a certain limit which might result in catastrophic accuracy drops. \n\nThe paper presents experiments on CIFAR-10 with a few CNN architectures along with current pruning baselines (layer-wise mostly. They also show results on ImageNet using ResNet50 and MovileNetV1. Following these experiments, the authors claim that GP or GPMT are SOTA for pruning despite their simplicity. \n\nThe paper also includes some discussion on how output patterns of each layer change with different pruning schemes and a note about how to set the optimal MT values. \n\nWhile I agree with the sentiment of the paper, I do not think the paper is presenting anything new (even from a benchmarking perspective) but rather is missing the point about what makes GP vs layer-wise pruning a worthy trade-off often. I will list my concerns in the main review.", "main_review": "I will go sequentially in the paper but will start with the strengths follow it up with the weaknesses: \n\nOverall, the writing quality is OK and is easy to follow. The information has been well put together for a new reader to understand without having to worry about the background. The authors should be appreciated for that.  \n\nStrengths:\n1) The fact that authors revisited GP to show its power is a very nice step. GP has been (according to me) the strongest baselines for pruning for a given parameter budget even now (has been acknowledged in multiple papers). \n2) The idea of MT makes sense and is a valid fix for the overly pruned layers in general.\n3) The experiments on ImageNet are commended because the insights from there a more valuable than other things presented in this paper. \n4) The discussion on network architectures, outputs and settings of MT are interesting in general. \n5) I really liked the related work, it is thorough from a literature review standpoint and would recommend new readers to go through it for a nice picture of the field \n\nWeaknesses:\nThis might sound a bit critical, but I would appreciate it if the authors understand the merit of whatever I am trying to put forward here. \n\n1) No one ever claimed GP is not sufficient to achieve SOTA performance on pruning. GP is extremely powerful no matter what gives the highest accuracy (often) for the same #params even at the highest sparsity levels. GP is not forgotten or overlooked by any means. The reason why people work on layer-wise pruning is slightly different and I will come to it later. \n2) While the authors mentioned people of LTH community probably used GP as one of the potential pruning techniques, GP was the only one that resulted in strong tickets at higher sparsities, making IMP explicitly work on GP no matter what other factors are. \n3) MT comes into the picture only when there is one-shot pruning and GP has been shown to work great even with gradual schedules (IMP is an example) and will not suffer from pruning out entire layers just because of the gradual processing. I would even argue setting the small pruning schedule is data-independent and has a similar cost as finding the optimal MT as done in this paper. MT makes sense for one-shot pruning, but one-shot pruning itself is problematic (will explain in a bit).\n4) Table 1 is a poor representation, I would say, STR and GMP are extremely easy to implement. GMP could be done in One-shot and GMP is data-independent no matter what. I would like to hear authors' thoughts on this. \n5) Coming to the setup, the biggest problem I have with the way GP or GPMT is being used is the application on top of a pre-trained network. Eg. On imagenet the fine-tuning of GP or GPMT takes 20 extra epochs, which isn't the case with all the baselines in the discussion. All of them are pruning while training baseline that only takes 100 epochs at best. I personally have adapted GP to be pruning while training like with IMP and GP does amazingly well, so the aspect of one-shot is overshadowed by the extra cost involved in fine-tuning. I don't know how we can make a case for these trade-offs to make the comparison fair.\n6) The big one, the paper completely disregards the compute cost during inference (FLOPs) for these pruned models. The biggest difference GP and layer-wise pruning methods have is the inference costs involved. Please see STR or RigL papers to see about that axis of comparison. GP often has 2x more inference FLOPs than uniform-sparsity with very little gain in accuracy. If uniform sparsity was adjusted for that compute costs, the accuracy will be much higher. This paper, while being a study on GP, completely overlooks this aspect of GP or GPMT.\n7) While the results on CIFAR-10 are good for debugging and understanding, the insights aren't particularly transferable. I would also not read too much into 40% sparsity results or even 95% sparsity because the networks involved are so heavily overparameterized that even at that sparsity level CIFAR-10 seems like a simple task. However, I would like to say Table 5 is a real thing for one-shot pruning, but will not happen for gradual GP. I would recommend authors implement gradual GP to see what it brings to the table, along with reduced training costs (debatable because one can think of pre-trained models as free). \n8) Figures 2 and 3 just reinforce what I am saying above. \n9) The gains in Tables 6 and 7 are not significant enough to make sweeping claims on SOTA while also not measuring FLOPs.\n10) The use of \"huge margin\" is highly discouraged in Table 8 because the gains are very insignificant and trust me the FLOPs are going to be so exorbitantly high that the accuracy vs # params doesn't matter in that case. \n11) As mentioned in the paper, MT doesn't seem to bring anything to the table for ImageNet experiments, just because of how the weight norms across layers change. MT is probably not a valuable addition in this case. \n12) I would not get into the discussion of output preservation because, it doesn't necessarily mean that the model is doing well if its outputs are closer to some other model, the yardstick itself is flaky here. \n13) Lastly, the discussion of MT tries to make it feel like it is not complex or data-dependent. But the same paper says GMP schedule is complex. I would say searching for MT is as good as coming up with a cubic pruning pattern which will even help GP when done in a gradual fashion.\n\nOverall, I do not think the paper brings any new observations into the limelight and also ignores the issues present with whatever has been presented. The paper is not ready for publication, but I am happy to chat with the authors to help them understand and (myself) gain perspective if I am missing something. \n\n", "summary_of_the_review": "The paper does not propose anything novel (which is fine) but the observations are also not new while the paper ignores most of the issues and tries to underplay other baselines. \n\nThe paper is not ready to be published but is an excellent starting point for new readers. \n\n\n=---------------------------------------------------------------------------------------\n\nAfter an extremely long discussion with the authors and providing them with things I felt were necessary for fixing the experiments, claims etc., I think the paper is not ready to be published. The authors should revisit most aspects of the paper if they want to make it a benchmark paper for GP and claim so. None of the insights and contributions are novel given my discussion and experience. I  hope the authors understand that it is better to publish a ready paper than a paper changing around a lot at this point. \n\nI vote for rejection.\n\n\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634418079295}], "openreview_url": "https://openreview.net/forum?id=jNB6vfl_680", "arxiv_id": null, "paper_pdf": "papers/jNB6vfl_680.pdf", "paper_pdf_sha256": "69d9bdeda6b27a9cb31ea8f8834509a402e7393b7bfc356b31a13adf3d44ce67", "paper_pdf_bytes": 1288947, "paper_pdf_source": "openreview", "code_url": "https://github.com/GPMT-Authors/Global-Pruning-With-Minimum-Threshold", "code_repository": "GPMT-Authors/Global-Pruning-With-Minimum-Threshold", "code_commit": "6e753e4bf39c3343d8514fffd4269c092908a9f8", "code_archive": "repos/jNB6vfl_680.zip", "code_archive_sha256": "e7ca2e42da25f81467874f3426fe01587c1a213909699b3270295dfeff894944", "code_archive_bytes": 28026, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 66, "github_languages": {"Python": 94785}, "github_archived": false, "github_pushed_at": "2021-10-06T12:48:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/global-magnitude-pruning-with-minimum"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4Nt1F3qf9Gn", "year": 2021, "status": "rejected", "title": "CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients", "authors": ["Dani Kiyasseh", "Tingting Zhu", "David A. Clifton"], "authorids": ["~Dani_Kiyasseh1", "tingting.zhu@eng.ox.ac.uk", "~David_A._Clifton1"], "authors_source": "OpenReview API", "abstract": "The healthcare industry generates troves of unlabelled physiological data. This data can be exploited via contrastive learning, a self-supervised pre-training method that encourages representations of instances to be similar to one another. We propose a family of contrastive learning methods, CLOCS, that encourages representations across space, time, \\textit{and} patients to be similar to one another. We show that CLOCS consistently outperforms the state-of-the-art methods, BYOL and SimCLR, when performing a linear evaluation of, and fine-tuning on, downstream tasks. We also show that CLOCS achieves strong generalization performance with only 25\\% of labelled training data. Furthermore, our training procedure naturally generates patient-specific representations that can be used to quantify patient-similarity. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Lp3p_qUXl6Y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1639/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a contrastive learning method for cardiac signals.\n\nThis strong points of the paper are the following:\n1. It presents its problem clearly. The problem it targets is an impactful problem in the medical domain.\n2. The method it proposes is straight forward and easily understandable.\n3. It demonstrates the advantage of the proposed method on a real dataset by comparing against the SOTA CL methods like BYOL and SimCLR.\n\nThe paper also has weaknesses.\n1. unsupported claims In the abstract.\n it is claimed that \"our training procedure naturally generates patient-specific representations that can be used to quantify patient-similarity\". Then the next time where the \"patient-similarity\" appears is in the discussion. It says \"We have managed to learn patient-specific representations....\". However, in the main paper, I can not find any evidence showing the representation it learns is capturing the patient-similarity. It is unacceptable to claim something by just claiming it.\n2.unclear contributions.\n In the introduction, the author is listing their methods as contributions. Usually, the contribution should be something achieved by you while not previously achieved. It could the new best performance, a new solution to a problem, a new angle of viewing the problem. The method itself alone hardly can be the contributions. I can propose any algorithms. They can not be contributions if they are not useful. So I suggest the author think about elaborating on what are your contributions.\n3.not sufficient description of the task.\n In the whole paper, the only thing I get about the task is that the author focuses on is a classification task related to cardiac arrhythmia. This is far from sufficient. The author should provide the backgrounds of the task. What are the extra 11 or 4 classes of cardiac arrhythmia? Is the task label also time series? or it is a global label that does change across time? For most of the readers from the ML community, they do not have a background in cardiac signals. To let the audience understand the tasks, I suggest the author to including visualization about the class label and the cardiac signals. Currently, I have no idea about what the task exactly is.\n4.lack of insight.\n Currently, the paper is mainly written in a way like I did A,B,C,D; they are better than previous method E,F; see my numbers. Here are my questions. Why doing A,B,C,D? For example, why Constrasitve Multi-segments Coding make sense? This question is also related to my previous point. What is your class label? If it is also a time series, considering the following case. At time step t1, the patient is normal, while at time step t2, the patient is in some state of cardiac arrhythmia. Why it makes sense to encourage the representations at t1 and t2 to be similar?\nActually, the experimental result is interesting. From Table 2, I notice that the best performance is achieved by either CMSC or CMSMLC. It means that contrastive multi-segments across time is useful. It contradicts my intuition in the last paragraph. Can the author give any explanation?\nThere are more questions about the results the author showed. For example, in Figure 4(a), CMSC has an abnormally high AUC when embedding dimension = 128. Why does it happen?\n\nI hope the author can include more insight or deep analysis in the paper to help the reader learn more from this paper.\nThe paper has a good topic, good methods, and good experiments. However, it still needs some effort to be a good paper. I hope my comments can help the author improve the paper and make it good work. My rating is not final. I will raise the score if my main concerns are addressed.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting topic; lack of insight; unsupported claims ", "review": "This paper proposes a contrastive learning method for cardiac signals.\n\nThis strong points of the paper are the following:\n1. It presents its problem clearly. The problem it targets is an impactful problem in the medical domain.\n2. The method it proposes is straight forward and easily understandable.\n3. It demonstrates the advantage of the proposed method on a real dataset by comparing against the SOTA CL methods like BYOL and SimCLR.\n\nThe paper also has weaknesses.\n1. unsupported claims In the abstract.\n it is claimed that \"our training procedure naturally generates patient-specific representations that can be used to quantify patient-similarity\". Then the next time where the \"patient-similarity\" appears is in the discussion. It says \"We have managed to learn patient-specific representations....\". However, in the main paper, I can not find any evidence showing the representation it learns is capturing the patient-similarity. It is unacceptable to claim something by just claiming it.\n2.unclear contributions.\n In the introduction, the author is listing their methods as contributions. Usually, the contribution should be something achieved by you while not previously achieved. It could the new best performance, a new solution to a problem, a new angle of viewing the problem. The method itself alone hardly can be the contributions. I can propose any algorithms. They can not be contributions if they are not useful. So I suggest the author think about elaborating on what are your contributions.\n3.not sufficient description of the task.\n In the whole paper, the only thing I get about the task is that the author focuses on is a classification task related to cardiac arrhythmia. This is far from sufficient. The author should provide the backgrounds of the task. What are the extra 11 or 4 classes of cardiac arrhythmia? Is the task label also time series? or it is a global label that does change across time? For most of the readers from the ML community, they do not have a background in cardiac signals. To let the audience understand the tasks, I suggest the author to including visualization about the class label and the cardiac signals. Currently, I have no idea about what the task exactly is.\n4.lack of insight.\n Currently, the paper is mainly written in a way like I did A,B,C,D; they are better than previous method E,F; see my numbers. Here are my questions. Why doing A,B,C,D? For example, why Constrasitve Multi-segments Coding make sense? This question is also related to my previous point. What is your class label? If it is also a time series, considering the following case. At time step t1, the patient is normal, while at time step t2, the patient is in some state of cardiac arrhythmia. Why it makes sense to encourage the representations at t1 and t2 to be similar?\nActually, the experimental result is interesting. From Table 2, I notice that the best performance is achieved by either CMSC or CMSMLC. It means that contrastive multi-segments across time is useful. It contradicts my intuition in the last paragraph. Can the author give any explanation?\nThere are more questions about the results the author showed. For example, in Figure 4(a), CMSC has an abnormally high AUC when embedding dimension = 128. Why does it happen?\n\nI hope the author can include more insight or deep analysis in the paper to help the reader learn more from this paper.\nThe paper has a good topic, good methods, and good experiments. However, it still needs some effort to be a good paper. I hope my comments can help the author improve the paper and make it good work. My rating is not final. I will raise the score if my main concerns are addressed.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604653695951}, {"id": "nnVXxVNw08Y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1639/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes a method for unsupervised learning of patient representations from ECG data.\nUsing contrastive learning, ECG recordings from different time periods and different leads are optimized to be similar for the same patient and different for all the other patients.\nThe method is evaluated on several datasets, showing improvement over random initialization and an alternative method using data permutations.\n\nThe method is simple and a relatively minor extension of contrastive learning to time series. \nBut it is elegant, shows good results over alternative approaches and can be a useful solution for time series.\n\nThe paper doesn't currently  contain a description of the actual input and the network architecture. There are 29 pages of appendices and these have some additional details, but the main paper should contain some of this important information as well.\n\nEquations 1-4 need to come with some explanation. At the moment, several variables there are left undefined and the overall intuition behind structuring the loss equations should be explained.\n\nCould these additional objectives or data augmentations be applied during the fine-tuning phase as well? If so, would that reduce the benefit of unsupervised pretraining?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple but elegant", "review": "The paper describes a method for unsupervised learning of patient representations from ECG data.\nUsing contrastive learning, ECG recordings from different time periods and different leads are optimized to be similar for the same patient and different for all the other patients.\nThe method is evaluated on several datasets, showing improvement over random initialization and an alternative method using data permutations.\n\nThe method is simple and a relatively minor extension of contrastive learning to time series. \nBut it is elegant, shows good results over alternative approaches and can be a useful solution for time series.\n\nThe paper doesn't currently  contain a description of the actual input and the network architecture. There are 29 pages of appendices and these have some additional details, but the main paper should contain some of this important information as well.\n\nEquations 1-4 need to come with some explanation. At the moment, several variables there are left undefined and the overall intuition behind structuring the loss equations should be explained.\n\nCould these additional objectives or data augmentations be applied during the fine-tuning phase as well? If so, would that reduce the benefit of unsupervised pretraining?", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604012439224}, {"id": "8kd928a8UUN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1639/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis work presents a new self-supervised training framework for multi-channel ECG signals. The authors use contrastive learning by exploiting the fact that a single patient can generate multiple ECG signals, and there are multiple views (i.e. leads) for the same ECG signals. Compared to popular self-supervised training methods BYOL and SimCLR, the proposed method shows superior performance on the arrhythmia classification task for 4 different datasets in various scenarios.\n\nPros:\n- The proposed approach is easy to follow, makes much sense, and is well-motivated based on the domain-specific knowledge of ECG signals.\n- The proposed approach outperforms popular baselines (although borrowed from the vision community) for the arrhythmia classification task on various combinations of experiment setups.\n\nCons:\n- The most critical shortcoming of this paper is that the baselines, although widely known, are borrowed from the computer vision community only, which is the reason for rating 4. There are already several papers proposing self-supervised learning methods for time-series data such as audio signals, ECGs and EEGs [1, 2, 3, 4, 5]. The authors should compose a stronger baseline set to demonstrate the value of this work. (Audio signals are usually single-channel unlike ECG signals, but CMSC alone already demonstrates good performance, hence methods such as wave2vec could be a strong baseline)\n- The third paragraph of section 6.5, where the authors claim the superiority of CMSC across various embedding dimensions seems like a stretch. The AUC difference of SimCLR (which is 0.02) and that of CMSC (0.03) is not that large, and CMSC even underperforms when the embedding dimension is 256. The proposed method has already shown good performance in various cases, so there is probably no need to make a stretched claim.\n\nReference\n1. Schneider, S., Baevski, A., Collobert, R. and Auli, M., 2019. wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:1904.05862.\n2. Baevski, A., Zhou, H., Mohamed, A. and Auli, M., 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. arXiv preprint arXiv:2006.11477.\n3. Banville, H., Moffat, G., Albuquerque, I., Engemann, D.A., Hyvärinen, A. and Gramfort, A., 2019, October. Self-supervised representation learning from electroencephalography signals. In 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP) (pp. 1-6). IEEE.\n4. Sarkar, P. and Etemad, A., 2020. Self-supervised ecg representation learning for emotion recognition. arXiv preprint arXiv:2002.03898.\n5. Cheng, J.Y., Goh, H., Dogrusoz, K., Tuzel, O. and Azemi, E., 2020. Subject-aware contrastive learning for biosignals. arXiv preprint arXiv:2007.04871.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "Summary:\nThis work presents a new self-supervised training framework for multi-channel ECG signals. The authors use contrastive learning by exploiting the fact that a single patient can generate multiple ECG signals, and there are multiple views (i.e. leads) for the same ECG signals. Compared to popular self-supervised training methods BYOL and SimCLR, the proposed method shows superior performance on the arrhythmia classification task for 4 different datasets in various scenarios.\n\nPros:\n- The proposed approach is easy to follow, makes much sense, and is well-motivated based on the domain-specific knowledge of ECG signals.\n- The proposed approach outperforms popular baselines (although borrowed from the vision community) for the arrhythmia classification task on various combinations of experiment setups.\n\nCons:\n- The most critical shortcoming of this paper is that the baselines, although widely known, are borrowed from the computer vision community only, which is the reason for rating 4. There are already several papers proposing self-supervised learning methods for time-series data such as audio signals, ECGs and EEGs [1, 2, 3, 4, 5]. The authors should compose a stronger baseline set to demonstrate the value of this work. (Audio signals are usually single-channel unlike ECG signals, but CMSC alone already demonstrates good performance, hence methods such as wave2vec could be a strong baseline)\n- The third paragraph of section 6.5, where the authors claim the superiority of CMSC across various embedding dimensions seems like a stretch. The AUC difference of SimCLR (which is 0.02) and that of CMSC (0.03) is not that large, and CMSC even underperforms when the embedding dimension is 256. The proposed method has already shown good performance in various cases, so there is probably no need to make a stretched claim.\n\nReference\n1. Schneider, S., Baevski, A., Collobert, R. and Auli, M., 2019. wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:1904.05862.\n2. Baevski, A., Zhou, H., Mohamed, A. and Auli, M., 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. arXiv preprint arXiv:2006.11477.\n3. Banville, H., Moffat, G., Albuquerque, I., Engemann, D.A., Hyvärinen, A. and Gramfort, A., 2019, October. Self-supervised representation learning from electroencephalography signals. In 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP) (pp. 1-6). IEEE.\n4. Sarkar, P. and Etemad, A., 2020. Self-supervised ecg representation learning for emotion recognition. arXiv preprint arXiv:2002.03898.\n5. Cheng, J.Y., Goh, H., Dogrusoz, K., Tuzel, O. and Azemi, E., 2020. Subject-aware contrastive learning for biosignals. arXiv preprint arXiv:2007.04871.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603977025268}, {"id": "km-jwTBw8WO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1639/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to use contrastive learning to learn representations from cardiac signals (ECGs). The model incorporates ECG domain knowledge, patient-specific, and relationships between multiple leads (channels), in the learning process. The targeted task is very important, enormous ECGs are collected and stored, but seldom people are mining them as they are unlabelled. This paper might reinvigorate them. \n\nMajor: \n\nHowever, some of the domain knowledge-based intuitions are unclear (might even be wrong). For example, for Contrastive Multi-lead Coding, we know that ECGs are 2D projections from the heart's 3D electricity activities. The similarity between different leads is related to the projection surface. What if two leads are orthogonally projected? They should be unrelated at all. Thus would even bring the noise in the model. Maybe this is the reason why CMSC in Table 1 is the best, but adding Multi-lead Coding (CMSMLC) drops the performance. \n\nIn fact, unsupervised learning usually needs (much) larger datasets. But all four experimental datasets are actually well annotated (relatively small) ECG databases. It seems more reasonable to experiment on a larger unlabelled dataset, such as the MIMIC-III waveform database (https://physionet.org/content/mimic3wdb/1.0/). And it might give more convincing and desired results. \n\nOthers: \n\n(1) This paper mainly focuses on design positive and negative pairs from the perspective of ECG domain knowledge - between leads, within subject. The technical contribution looks limited. \n\n(2) Some related papers about contrastive learning on physiological signals. \n\nPublished (you might reconsider the last sentence in Related Work.):\n\n[1] Banville, Hubert, et al. \"Self-supervised representation learning from electroencephalography signals.\" 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2019.\n\nPreprint:\n\n[2] Cheng, Joseph Y., et al. \"Subject-aware contrastive learning for biosignals.\" arXiv preprint arXiv:2007.04871 (2020).\n\n[3] Banville, Hubert, et al. \"Uncovering the structure of clinical EEG signals with self-supervised learning.\" arXiv preprint arXiv:2007.16104 (2020).\n\n(3) Multi-lead ECG is more common in clinical usage. Can this method be extended to learn representations from multi-lead ECG?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Questions about intuitions of Contrastive Multi-lead Coding", "review": "This paper proposes to use contrastive learning to learn representations from cardiac signals (ECGs). The model incorporates ECG domain knowledge, patient-specific, and relationships between multiple leads (channels), in the learning process. The targeted task is very important, enormous ECGs are collected and stored, but seldom people are mining them as they are unlabelled. This paper might reinvigorate them. \n\nMajor: \n\nHowever, some of the domain knowledge-based intuitions are unclear (might even be wrong). For example, for Contrastive Multi-lead Coding, we know that ECGs are 2D projections from the heart's 3D electricity activities. The similarity between different leads is related to the projection surface. What if two leads are orthogonally projected? They should be unrelated at all. Thus would even bring the noise in the model. Maybe this is the reason why CMSC in Table 1 is the best, but adding Multi-lead Coding (CMSMLC) drops the performance. \n\nIn fact, unsupervised learning usually needs (much) larger datasets. But all four experimental datasets are actually well annotated (relatively small) ECG databases. It seems more reasonable to experiment on a larger unlabelled dataset, such as the MIMIC-III waveform database (https://physionet.org/content/mimic3wdb/1.0/). And it might give more convincing and desired results. \n\nOthers: \n\n(1) This paper mainly focuses on design positive and negative pairs from the perspective of ECG domain knowledge - between leads, within subject. The technical contribution looks limited. \n\n(2) Some related papers about contrastive learning on physiological signals. \n\nPublished (you might reconsider the last sentence in Related Work.):\n\n[1] Banville, Hubert, et al. \"Self-supervised representation learning from electroencephalography signals.\" 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2019.\n\nPreprint:\n\n[2] Cheng, Joseph Y., et al. \"Subject-aware contrastive learning for biosignals.\" arXiv preprint arXiv:2007.04871 (2020).\n\n[3] Banville, Hubert, et al. \"Uncovering the structure of clinical EEG signals with self-supervised learning.\" arXiv preprint arXiv:2007.16104 (2020).\n\n(3) Multi-lead ECG is more common in clinical usage. Can this method be extended to learn representations from multi-lead ECG?\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603446097083}], "openreview_url": "https://openreview.net/forum?id=4Nt1F3qf9Gn", "arxiv_id": "2005.13249", "paper_pdf": "papers/4Nt1F3qf9Gn.pdf", "paper_pdf_sha256": "adabb38f1737e5ffc4fb3ff16d86f833bce954d0895c9a8614a937f4d72eaee7", "paper_pdf_bytes": 1579474, "paper_pdf_source": "openreview", "code_url": "https://github.com/danikiyasseh/CLOCS", "code_repository": "danikiyasseh/CLOCS", "code_commit": "e58fb4f0369b086735169d781a9513730fa157d0", "code_archive": "repos/4Nt1F3qf9Gn.zip", "code_archive_sha256": "d4ebb1e655b44ff2a441510432f121d08daa448c2c14cd3ce5c40ccef28b6d09", "code_archive_bytes": 27970, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 69, "github_languages": {"Python": 82660}, "github_archived": false, "github_pushed_at": "2022-10-19T12:22:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/clocs-contrastive-learning-of-cardiac-signals"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1g4M0EtPS", "year": 2020, "status": "rejected", "title": "Gaussian MRF Covariance Modeling for Efficient Black-Box Adversarial Attacks", "authors": ["Anit Kumar Sahu", "J. Zico Kolter", "Satya Narayan Shukla"], "authorids": ["anit.sahu@gmail.com", "zkolter@cs.cmu.edu", "snshukla@cs.umass.edu"], "authors_source": "OpenReview API", "abstract": "We study the problem of generating adversarial examples in a black-box setting, where we only have access to a zeroth order oracle, providing us with loss function evaluations. We employ Markov Random Fields (MRF) to exploit the structure of input data to systematically model the covariance structure of the gradients. The MRF structure in addition to Bayesian inference for the gradients facilitates one-step attacks akin to Fast Gradient Sign Method (FGSM) albeit in the black-box setting. The resulting method uses fewer queries than the current state of the art to achieve comparable performance. In particular, in the regime of lower query budgets, we show that our method is particularly effective in terms of fewer average queries with high attack accuracy while employing one-step attacks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJe8shqRYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper998/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper deals with the problem of finding an adversarial examples when only the output of a model can be evaluated, but not its gradient. The key idea of the paper is building a Gaussian MRF (a Gaussian with a sparse inverse covariance matrix with a special band structure) to maintain a model for the gradients for predicting search directions. The approach is sensible and uses the FFT trick applicable for diagonalizing covariance matrices with circulant structure.\n\n- The ideas in this paper have practical utility. The paper is unfortunately not very carefully written, and the arguments require occasionally some guesswork.\n\n- There is not a sufficient discussion of experimental findings. Why does the proposed method works better than a white box FGSM for instance? \n\n- I don’t fully understand what ‘solving for GMRF’ means (Alg 1). I would expect it is estimating the parameters of the covariance function but this algorithm just calculates the likelihood. \n\n- Given the simple coupling structure with parameter tying, this problem seems to be closely related to estimating AR(N) style models (alpha and beta parameters) so I am surprised to see only a general treatment. \nIn this problem, it seems much more natural to estimate theta and g recursively and concurrently, and there are very well known algorithms related to Kalman filtering. Please discuss.\n\n- The evaluation of the idea is not complete While it is certainly interesting to see that such a simple heuristic can achieve comparable performance on standard datasets over other black-box attack methods in the limited query budget regime, I would expect to see some experiments that illustrate the quality of the gradient estimator more closely (not only its final effectiveness for finding a search direction inside of an optimization method) and more strong justification of the proposed model. \n\n- As the entire contribution hinges on the observation that gradients across dimensions of an example as well as across examples in a dataset are correlated, it would have been also very informative to show estimates of autocorrelation functions of  the gradients on different datasets to justify the basic modelling choices. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper deals with the problem of finding an adversarial examples when only the output of a model can be evaluated, but not its gradient. The key idea of the paper is building a Gaussian MRF (a Gaussian with a sparse inverse covariance matrix with a special band structure) to maintain a model for the gradients for predicting search directions. The approach is sensible and uses the FFT trick applicable for diagonalizing covariance matrices with circulant structure.\n\n- The ideas in this paper have practical utility. The paper is unfortunately not very carefully written, and the arguments require occasionally some guesswork.\n\n- There is not a sufficient discussion of experimental findings. Why does the proposed method works better than a white box FGSM for instance? \n\n- I don’t fully understand what ‘solving for GMRF’ means (Alg 1). I would expect it is estimating the parameters of the covariance function but this algorithm just calculates the likelihood. \n\n- Given the simple coupling structure with parameter tying, this problem seems to be closely related to estimating AR(N) style models (alpha and beta parameters) so I am surprised to see only a general treatment. \nIn this problem, it seems much more natural to estimate theta and g recursively and concurrently, and there are very well known algorithms related to Kalman filtering. Please discuss.\n\n- The evaluation of the idea is not complete While it is certainly interesting to see that such a simple heuristic can achieve comparable performance on standard datasets over other black-box attack methods in the limited query budget regime, I would expect to see some experiments that illustrate the quality of the gradient estimator more closely (not only its final effectiveness for finding a search direction inside of an optimization method) and more strong justification of the proposed model. \n\n- As the entire contribution hinges on the observation that gradients across dimensions of an example as well as across examples in a dataset are correlated, it would have been also very informative to show estimates of autocorrelation functions of  the gradients on different datasets to justify the basic modelling choices. \n\n"}, "tcdate": 1571888286190}, {"id": "B1gYI3PpFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper998/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a method for black box adversarial image generation. The idea is to learn a parameterization of a precision matrix so that gradients of a network's loss are assumed to be drawn from a corresponding Gaussian. The parameters of this model are fit efficiently using the spectral theorem that their particular parameterization of the precision matrix allows them to use. Gradient estimation is then viewed as a Gaussian conditioning problem given observations (see last equation on page 5).\n\nOverall, I think the method is elegant -- particular in comparison to many existing approaches in the literature that rely on highly complex machinery like genetic algorithms and the like. My main source of questions is the experimental results section, which I currently view as somewhat weak and a little confusing -- I would be more than happy to increase my score if my concerns are sufficiently addressed.\n\nFirst, I'm not sure that the story told by Figure 3 and Table 1 is entirely clear. On the whole at a given sufficiently high success rate (say, 80%), it seems that the authors' approach consistently loses to the Parsimonious attack of Moon et al., 2019? The authors' method seems to perform strictly better than Bandits_{TD} and NES, but the gap between the Parsimonious attack and the authors' is quite substantial past 300 evaluations.\n\nThis ultimately leaves me with the question of what to do with this paper. The learning framework used to estimate gradients is clever, but doesn't seem to me to be methodologically groundbreaking to the point where it can stand on its own merits independent of its performance in comparison to other techniques. Particularly considering other papers have been published since the Ilyas et al., 2019 paper that outperform Bandits_{TD} and NES in terms of query efficiency, I worry that this paper simply presents a decent idea with middling results.\n\nI'd therefore like the authors to primarily comment on the broader impact they believe their paper will have. The conclusion primarily focuses on the introduction of the method: does it have substantial merits past its empirical performance?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "In this paper, the authors propose a method for black box adversarial image generation. The idea is to learn a parameterization of a precision matrix so that gradients of a network's loss are assumed to be drawn from a corresponding Gaussian. The parameters of this model are fit efficiently using the spectral theorem that their particular parameterization of the precision matrix allows them to use. Gradient estimation is then viewed as a Gaussian conditioning problem given observations (see last equation on page 5).\n\nOverall, I think the method is elegant -- particular in comparison to many existing approaches in the literature that rely on highly complex machinery like genetic algorithms and the like. My main source of questions is the experimental results section, which I currently view as somewhat weak and a little confusing -- I would be more than happy to increase my score if my concerns are sufficiently addressed.\n\nFirst, I'm not sure that the story told by Figure 3 and Table 1 is entirely clear. On the whole at a given sufficiently high success rate (say, 80%), it seems that the authors' approach consistently loses to the Parsimonious attack of Moon et al., 2019? The authors' method seems to perform strictly better than Bandits_{TD} and NES, but the gap between the Parsimonious attack and the authors' is quite substantial past 300 evaluations.\n\nThis ultimately leaves me with the question of what to do with this paper. The learning framework used to estimate gradients is clever, but doesn't seem to me to be methodologically groundbreaking to the point where it can stand on its own merits independent of its performance in comparison to other techniques. Particularly considering other papers have been published since the Ilyas et al., 2019 paper that outperform Bandits_{TD} and NES in terms of query efficiency, I worry that this paper simply presents a decent idea with middling results.\n\nI'd therefore like the authors to primarily comment on the broader impact they believe their paper will have. The conclusion primarily focuses on the introduction of the method: does it have substantial merits past its empirical performance?"}, "tcdate": 1571810384643}, {"id": "HkxENuvTFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper998/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper employs Markov random fields to exploit the input data structure and further model the covariance structure of the gradients. This embeds covariance structures of input data space into the gradient operator for an adversary attack.\nThey further use this gradient operator with a fast gradient sign Method. The numerics show effective for using fewer queries to obtain high attack accuracy. This paper is well written with clear derivations. I suggest the publication of the paper.\n\n In fact, modeling the structure of an input space structure into the adversary attack is a good direction. For similar intuitions in this direction, I recommend a related work:\n      \n         \"A. Lin, Y. Dukler, W. Li, G. Montufar, Wasserstein Diffusion Tikhonov Regularization\" \n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper employs Markov random fields to exploit the input data structure and further model the covariance structure of the gradients. This embeds covariance structures of input data space into the gradient operator for an adversary attack.\nThey further use this gradient operator with a fast gradient sign Method. The numerics show effective for using fewer queries to obtain high attack accuracy. This paper is well written with clear derivations. I suggest the publication of the paper.\n\n In fact, modeling the structure of an input space structure into the adversary attack is a good direction. For similar intuitions in this direction, I recommend a related work:\n      \n         \"A. Lin, Y. Dukler, W. Li, G. Montufar, Wasserstein Diffusion Tikhonov Regularization\" \n "}, "tcdate": 1571809324040}], "openreview_url": "https://openreview.net/forum?id=H1g4M0EtPS", "arxiv_id": "2010.04205", "paper_pdf": "papers/H1g4M0EtPS.pdf", "paper_pdf_sha256": "abb700b6ed17cb5325d5f68ae345ffb45f3a3eca64fdb2b6260a613c29a9bade", "paper_pdf_bytes": 587358, "paper_pdf_source": "openreview", "code_url": "https://github.com/anitksahu/GMRF", "code_repository": "anitksahu/GMRF", "code_commit": "2ae642589cfefac06bafbe0325154b1066fd367f", "code_archive": "repos/H1g4M0EtPS.zip", "code_archive_sha256": "8469a3f16d9fd5a4eae193295dfafe36beefafafc2f8e02784297ec7abc7c5ed", "code_archive_bytes": 94175, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 102, "github_languages": {"Python": 30924}, "github_archived": false, "github_pushed_at": "2020-10-09T07:07:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gaussian-mrf-covariance-modeling-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HkGmDsR9YQ", "year": 2019, "status": "rejected", "title": "Generalization and Regularization in DQN", "authors": ["Jesse Farebrother", "Marlos C. Machado", "Michael Bowling"], "authorids": ["jfarebro@ualberta.ca", "machado@ualberta.ca", "mbowling@ualberta.ca"], "authors_source": "OpenReview API", "abstract": "Deep reinforcement learning (RL) algorithms have shown an impressive ability to learn complex control policies in high-dimensional environments. However, despite the ever-increasing performance on popular benchmarks like the Arcade Learning Environment (ALE), policies learned by deep RL algorithms can struggle to generalize when evaluated in remarkably similar environments. These results are unexpected given the fact that, in supervised learning, deep neural networks often learn robust features that generalize across tasks. In this paper, we study the generalization capabilities of DQN in order to aid in understanding this mismatch between generalization in deep RL and supervised learning methods. We provide evidence suggesting that DQN overspecializes to the domain it is trained on. We then comprehensively evaluate the impact of traditional methods of regularization from supervised learning, $\\ell_2$ and dropout, and of reusing learned representations to improve the generalization capabilities of DQN. We perform this study using different game modes of Atari 2600 games, a recently introduced modification for the ALE which supports slight variations of the Atari 2600 games used for benchmarking in the field. Despite regularization being largely underutilized in deep RL, we show that it can, in fact, help DQN learn more general features. These features can then be reused and fine-tuned on similar tasks, considerably improving the sample efficiency of DQN.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "BklXppaonm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper250/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I totally disagree with the authors that any of their observations are surprising. Indeed the fact that an RL agent does not generalizes to small modifications of the task (either visual or in the dynamics) is well known. If the agent should generalize though is a different question. And I do not mean this in the sense that it is an undesirable property but rather if it is outside of what “learning one task” means. Particularly I feel this is a very pessimistic view of RL and potentially not even in-line with what happens in supervised learning. \n\nI think one mantra of deep learning (and deep RL needs to obey by it) is that one should test in the same setting as training for things to truly work. For supervised, there is a distribution of data, and the test set are samples from the same distribution. However the testing scenario used here is slightly different. During training, if I do not see car accelerating, I think it makes no sense to expect to generalize to a new game that has this property as it is out-of-distribution. Of course it would be ideal if it could do that. And to clarify, while for us some of these extensions seem very similar and minimal changes, hence it should generalize to rather than transfer to, this is just the effect of imposing our own biases on the learning process. Deep Nets do not learn like we do, and in their universe they have never seen a car accelerating -- it makes sense that it might not to be able to generalize to it. Again, I’m not arguing that we don’t want this, but rather if we should expect it as part of what the system should normally generalize to.\n\nTo that end I think this paper enters in that unresolved dispute of what generalization should be versus what is transfer. At what point do we have truly a new task vs a sample from the same task. I don’t think there is an answer.\n\nGoing back to the observations in this work. I think the fact that the environment is not stochastic reinforces this overfitting (as in the extreme you end up with a policy that just repeats the optimal sequence of actions). I think in this particular case I can see how finetuning to a variation of the task fails. However true stochasticity in the environment (e.g. having a distribution of variations) like is done in Distral paper (where each episode is a different layout) can behave as a regularizer that will mitigate a bit the overfitting. That is to say that I believe the observed behaviour will be less pronounced in complex stochastic setting. \n\nNevertheless the paper seems to highlight an important observation (and back it up with empirical evidence), namely we should use more regularization like L2 or otherwise in practice. Which is mostly absent from publications. And I think this on its own is valuable. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not unknown but nice systematic exploration", "review": "I totally disagree with the authors that any of their observations are surprising. Indeed the fact that an RL agent does not generalizes to small modifications of the task (either visual or in the dynamics) is well known. If the agent should generalize though is a different question. And I do not mean this in the sense that it is an undesirable property but rather if it is outside of what “learning one task” means. Particularly I feel this is a very pessimistic view of RL and potentially not even in-line with what happens in supervised learning. \n\nI think one mantra of deep learning (and deep RL needs to obey by it) is that one should test in the same setting as training for things to truly work. For supervised, there is a distribution of data, and the test set are samples from the same distribution. However the testing scenario used here is slightly different. During training, if I do not see car accelerating, I think it makes no sense to expect to generalize to a new game that has this property as it is out-of-distribution. Of course it would be ideal if it could do that. And to clarify, while for us some of these extensions seem very similar and minimal changes, hence it should generalize to rather than transfer to, this is just the effect of imposing our own biases on the learning process. Deep Nets do not learn like we do, and in their universe they have never seen a car accelerating -- it makes sense that it might not to be able to generalize to it. Again, I’m not arguing that we don’t want this, but rather if we should expect it as part of what the system should normally generalize to.\n\nTo that end I think this paper enters in that unresolved dispute of what generalization should be versus what is transfer. At what point do we have truly a new task vs a sample from the same task. I don’t think there is an answer.\n\nGoing back to the observations in this work. I think the fact that the environment is not stochastic reinforces this overfitting (as in the extreme you end up with a policy that just repeats the optimal sequence of actions). I think in this particular case I can see how finetuning to a variation of the task fails. However true stochasticity in the environment (e.g. having a distribution of variations) like is done in Distral paper (where each episode is a different layout) can behave as a regularizer that will mitigate a bit the overfitting. That is to say that I believe the observed behaviour will be less pronounced in complex stochastic setting. \n\nNevertheless the paper seems to highlight an important observation (and back it up with empirical evidence), namely we should use more regularization like L2 or otherwise in practice. Which is mostly absent from publications. And I think this on its own is valuable. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541295546889}, {"id": "BylM33v92X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper250/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \nThis paper focuses on a \"generalization\" in deep Q-network (DQN). Specifically, they showed that when features (parameters of DQN) are trained in one environment (default flavour/mode) and then used as an initialization for the same model but for a slightly different environment ( i.e. still captures key concepts of the original environment ) can boost the performance of the model in the new flavour/environment. More importantly, the performance boost is significant when DQN's parameters which are used to initialize the model for the new environment were trained using dropout and L2 regularization in the default flavour/mode.  For the experiments, 4 games: FREEWAY, HERO, BREAKOUT, and SPACE INVADERS which have 13 flavours (combinations of a mode and a difficulty) are used.\n\nStrengths:\n+ This paper is interesting in the sense that it empirically shows that using regularization in training deep RL can be helpful when the goal is the generalization from one flavour of an environment to another one but very similar to the original.\n+ The experiments show that using REGULARIZED FINE-TUNING and FINE-TUNING for a new flavour /mode can help with sample efficiency compared with the models which are trained from scratch [10M frames vs 50M frames and 50M frames vs 100M frames ](Table 3).\n+ The paper is well-written and it can be easily followed.\n\nWeaknesses:\n- In my view, there should be experiments in which the proposed method is compared with other approaches that improve generalization in deep RL like Zhang et al., 2018 and Justesen et al., 2018.\n- According to the paper (at least my understanding) DQN's hyper-parameters are tuned based on default mode/flavour environment. It is possible that results for 'SCRATCH' results (Table 3) can be improved if DQN's hyper-parameters are tuned on the current flavour not default one.\n-The proposed method is only applicable when the default environment and the new one are very similar. If the environments are that similar why even bother to train first on default and then generalize to the new one? Wouldn't be less expensive to just focus on the target environment and find the best model on that?\n\nQuestions:\n- It is mentioned in the paper, that evaluation protocol suggested by Machado et al. (2018) is being followed in this paper. Have you followed the protocol introduced in section 4.2 of Machado et al. (2018)? If yes, the protocol in Machado et al. (2018) is for training time but numbers in Table 1 in this paper are for evaluation time. Can you elaborate?\n- Why only those 6 games were selected for the experiments? \n\nIn summary, I found this paper is interesting but my concern is about the experiments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting study but need more experiments. ", "review": "Summary: \nThis paper focuses on a \"generalization\" in deep Q-network (DQN). Specifically, they showed that when features (parameters of DQN) are trained in one environment (default flavour/mode) and then used as an initialization for the same model but for a slightly different environment ( i.e. still captures key concepts of the original environment ) can boost the performance of the model in the new flavour/environment. More importantly, the performance boost is significant when DQN's parameters which are used to initialize the model for the new environment were trained using dropout and L2 regularization in the default flavour/mode.  For the experiments, 4 games: FREEWAY, HERO, BREAKOUT, and SPACE INVADERS which have 13 flavours (combinations of a mode and a difficulty) are used.\n\nStrengths:\n+ This paper is interesting in the sense that it empirically shows that using regularization in training deep RL can be helpful when the goal is the generalization from one flavour of an environment to another one but very similar to the original.\n+ The experiments show that using REGULARIZED FINE-TUNING and FINE-TUNING for a new flavour /mode can help with sample efficiency compared with the models which are trained from scratch [10M frames vs 50M frames and 50M frames vs 100M frames ](Table 3).\n+ The paper is well-written and it can be easily followed.\n\nWeaknesses:\n- In my view, there should be experiments in which the proposed method is compared with other approaches that improve generalization in deep RL like Zhang et al., 2018 and Justesen et al., 2018.\n- According to the paper (at least my understanding) DQN's hyper-parameters are tuned based on default mode/flavour environment. It is possible that results for 'SCRATCH' results (Table 3) can be improved if DQN's hyper-parameters are tuned on the current flavour not default one.\n-The proposed method is only applicable when the default environment and the new one are very similar. If the environments are that similar why even bother to train first on default and then generalize to the new one? Wouldn't be less expensive to just focus on the target environment and find the best model on that?\n\nQuestions:\n- It is mentioned in the paper, that evaluation protocol suggested by Machado et al. (2018) is being followed in this paper. Have you followed the protocol introduced in section 4.2 of Machado et al. (2018)? If yes, the protocol in Machado et al. (2018) is for training time but numbers in Table 1 in this paper are for evaluation time. Can you elaborate?\n- Why only those 6 games were selected for the experiments? \n\nIn summary, I found this paper is interesting but my concern is about the experiments.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541205161679}, {"id": "rkeceWxK3m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper250/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This is an empirical study on the ability for DQNs trained with/without regularization to perform well on variants of the same environment (e.g. increasing difficulty of a game). The paper is well written, the experimental methodology is clear & sound, and the significance is around improved sample efficiency via warm starting from a regularized DQN to fine tune. The error bounds for the regularized models results seem uncomfortably large in some cases. Overall it looks like a good methodological paper that can inform others on taking regularization more seriously when training DQNs. Evaluating on a modified ALE environment is great, but it would have been better to see this having similar impact in real life applications.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Empirical Paper on Evaluating Generalization Properties with Regularized / Non-Regularized DQN", "review": "This is an empirical study on the ability for DQNs trained with/without regularization to perform well on variants of the same environment (e.g. increasing difficulty of a game). The paper is well written, the experimental methodology is clear & sound, and the significance is around improved sample efficiency via warm starting from a regularized DQN to fine tune. The error bounds for the regularized models results seem uncomfortably large in some cases. Overall it looks like a good methodological paper that can inform others on taking regularization more seriously when training DQNs. Evaluating on a modified ALE environment is great, but it would have been better to see this having similar impact in real life applications.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541107954261}], "openreview_url": "https://openreview.net/forum?id=HkGmDsR9YQ", "arxiv_id": "1810.00123", "paper_pdf": "papers/HkGmDsR9YQ.pdf", "paper_pdf_sha256": "9bdf26769a0e0e1bf2596c7940165a52d2fa4792a14d54a17a067c8a77c8bf93", "paper_pdf_bytes": 649929, "paper_pdf_source": "openreview", "code_url": "https://github.com/JesseFarebro/dqn-ale", "code_repository": "JesseFarebro/dqn-ale", "code_commit": "a39b3ba0f5e9fd8da05c9dd39f5264e9e4df5575", "code_archive": "repos/HkGmDsR9YQ.zip", "code_archive_sha256": "1f5cd4f16c4215fee020f915f0b22d3e57037b7da13f9513e96a17280f0319e5", "code_archive_bytes": 609319, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 594, "github_languages": {"Python": 30937}, "github_archived": false, "github_pushed_at": "2019-04-16T17:06:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalization-and-regularization-in-dqn"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PwaUsAk4i6", "year": 2026, "status": "rejected", "title": "Diffusion Guidance Is a Controllable Policy Improvement Operator", "authors": ["Kevin Frans", "Seohong Park", "Pieter Abbeel", "Sergey Levine"], "authorids": ["~Kevin_Frans1", "~Seohong_Park1", "~Pieter_Abbeel2", "~Sergey_Levine1"], "authors_source": "OpenReview API", "abstract": "At the core of reinforcement learning is the idea of learning beyond the performance in the data. However, scaling such systems has proven notoriously tricky. In contrast, techniques from generative modeling have shown to be remarkably scalable and are simple to train. \nIn this work, we combine these strengths, by deriving a direct relation between policy improvement and guidance of diffusion models. The resulting framework, CFGRL, is a policy improvement operator that is trained with the simplicity of supervised learning, yet is more effective than typically-used weighted policy extraction strategies. On offline RL tasks, we observe a reliable trend---increased guidance weighting leads to increased performance. Additionally, the CFGRL framework can be adapted to \"directly'' extract policies from offline data *without* running a full end-to-end RL algorithm, allowing us to generalize simple supervised methods (e.g. goal-conditioned behavior cloning) to further prioritize optimality, gaining performance across the board without additional cost.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "8RFLO3QRll", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14486/Reviewer_Bfm1"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper proposes CFGRL, a simple, controllable policy-improvement operator that leverages classifier-free guidance from diffusion/flow models. Policies are factorized as a product of a reference policy and an “optimality” term that is a monotone function of the advantage; sampling from this product is achieved by composing unconditional and optimality-conditioned policy scores, with a test-time guidance weight $w$ controlling the strength of improvement. The authors prove that such product policies improve over the reference and that increasing $w$ yields further improvement (with the usual trade-off against distribution shift). They also show that, under certain choices, guided sampling matches the solution to a KL-regularized policy-improvement objective. Practically, they instantiate CFGRL with a single diffusion/flow network and provide simple training/sampling algorithms", "review_text": "This paper proposes CFGRL, a simple, controllable policy-improvement operator that leverages classifier-free guidance from diffusion/flow models. Policies are factorized as a product of a reference policy and an “optimality” term that is a monotone function of the advantage; sampling from this product is achieved by composing unconditional and optimality-conditioned policy scores, with a test-time guidance weight $w$ controlling the strength of improvement. The authors prove that such product policies improve over the reference and that increasing $w$ yields further improvement (with the usual trade-off against distribution shift). They also show that, under certain choices, guided sampling matches the solution to a KL-regularized policy-improvement objective. Practically, they instantiate CFGRL with a single diffusion/flow network and provide simple training/sampling algorithms", "strengths": "The central insight—viewing policy improvement as classifier-free guidance over an advantage-conditioned policy—is elegant. It unifies guided diffusion sampling with KL-regularized policy improvement and control-as-inference via a clean product-policy view, and shows that test-time guidance directly tunes the improvement strength. \n\nThe theory is tidy. The paper also avoids learning an explicit optimality predictor via a Bayes inversion that merges unconditional and optimality-conditioned policies into a single network. Algorithms are minimal and clear. \n\nEmpirically, CFGRL improves over AWR on a strong majority of ExORL and OGBench tasks and shows a more favorable scaling trend than AWR’s temperature sweep. The GCBC “upgrade” is impactful: simply guiding the goal-conditioned policy (no value function) yields sizable gains, including hierarchical variants. Code is released; runs are modestly resource-bound.", "weaknesses": "The paper notes that larger w both improves $ A_{ \\hat \\pi }$ and deviates more from the dataset policy, possibly hurting performance; the ablation indeed shows performance sometimes declines past a point, but there’s no adaptive or trust-region control of $w$ or measured KL to the prior. \n\nFor the offline RL part, results are averaged over four seeds; gains are consistent but sometimes modest. The GCBC part uses more seeds, but a wider set of domains and stronger end-to-end RL baselines would further cement significance. The paper itself notes it is not a full SOTA RL algorithm. \n\nFlow steps (16–32) suggest non-trivial inference cost relative to a single-shot policy; wall-clock sampling latency and throughput are not measured, which matters for deployment.", "questions": "Could you add comparisons to CRR/CCR, IDQL with diffusion policy extraction, and Q-score-matching / rejection-sampling approaches, ideally normalizing compute and tuning budgets? This would position CFGRL more clearly among diffusion-policy extractors.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CFGRL, a simple, controllable policy-improvement operator that leverages classifier-free guidance from diffusion/flow models. Policies are factorized as a product of a reference policy and an “optimality” term that is a monotone function of the advantage; sampling from this product is achieved by composing unconditional and optimality-conditioned policy scores, with a test-time guidance weight $w$ controlling the strength of improvement. The authors prove that such product policies improve over the reference and that increasing $w$ yields further improvement (with the usual trade-off against distribution shift). They also show that, under certain choices, guided sampling matches the solution to a KL-regularized policy-improvement objective. Practically, they instantiate CFGRL with a single diffusion/flow network and provide simple training/sampling algorithms", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The central insight—viewing policy improvement as classifier-free guidance over an advantage-conditioned policy—is elegant. It unifies guided diffusion sampling with KL-regularized policy improvement and control-as-inference via a clean product-policy view, and shows that test-time guidance directly tunes the improvement strength. \n\nThe theory is tidy. The paper also avoids learning an explicit optimality predictor via a Bayes inversion that merges unconditional and optimality-conditioned policies into a single network. Algorithms are minimal and clear. \n\nEmpirically, CFGRL improves over AWR on a strong majority of ExORL and OGBench tasks and shows a more favorable scaling trend than AWR’s temperature sweep. The GCBC “upgrade” is impactful: simply guiding the goal-conditioned policy (no value function) yields sizable gains, including hierarchical variants. Code is released; runs are modestly resource-bound.", "weaknesses": "The paper notes that larger w both improves $ A_{ \\hat \\pi }$ and deviates more from the dataset policy, possibly hurting performance; the ablation indeed shows performance sometimes declines past a point, but there’s no adaptive or trust-region control of $w$ or measured KL to the prior. \n\nFor the offline RL part, results are averaged over four seeds; gains are consistent but sometimes modest. The GCBC part uses more seeds, but a wider set of domains and stronger end-to-end RL baselines would further cement significance. The paper itself notes it is not a full SOTA RL algorithm. \n\nFlow steps (16–32) suggest non-trivial inference cost relative to a single-shot policy; wall-clock sampling latency and throughput are not measured, which matters for deployment.", "questions": "Could you add comparisons to CRR/CCR, IDQL with diffusion policy extraction, and Q-score-matching / rejection-sampling approaches, ideally normalizing compute and tuning budgets? This would position CFGRL more clearly among diffusion-policy extractors.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761962620337}, {"id": "adsAOUuzww", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14486/Reviewer_6iHH"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper proposes Classifier-Free Guided Reinforcement Learning (CFGRL), which interprets the classifier-free guidance (CFG) mechanism of diffusion models as a policy improvement operator.\nThe authors show that by treating the optimality condition as a discrete binary variable $o\\in \\{0,1\\}$, CFG can be reinterpreted as applying an advantage-weighted transformation on a reference policy. This allows controllable policy improvement via the guidance weight $w$, similar in spirit to temperature or KL coefficients in regularized RL. Experiments demonstrate that CFGRL improves over Advantage-Weighted Regression (AWR) and Goal-Conditioned Behavioral Cloning (GCBC) across offline RL benchmarks.\n- An LLM was used to improve writing.", "review_text": "This paper proposes Classifier-Free Guided Reinforcement Learning (CFGRL), which interprets the classifier-free guidance (CFG) mechanism of diffusion models as a policy improvement operator.\nThe authors show that by treating the optimality condition as a discrete binary variable $o\\in \\{0,1\\}$, CFG can be reinterpreted as applying an advantage-weighted transformation on a reference policy. This allows controllable policy improvement via the guidance weight $w$, similar in spirit to temperature or KL coefficients in regularized RL. Experiments demonstrate that CFGRL improves over Advantage-Weighted Regression (AWR) and Goal-Conditioned Behavioral Cloning (GCBC) across offline RL benchmarks.\n- An LLM was used to improve writing.", "strengths": "1. Simplicity and practical appeal – The method requires only standard diffusion training and allows tuning the improvement strength $w$ at inference, offering a practical way to control policy quality without retraining.\n\n2. Solid empirical demonstration – Results on offline RL and goal-conditioned control tasks consistently show improvements over strong baselines such as AWR and GCBC.\n\n3. Readable and well-presented – The paper is clearly written, with theoretical and empirical sections well balanced.", "weaknesses": "1. Limited novelty beyond reinterpretation\n\nThe core idea—recasting classifier-free guidance as a policy improvement operator—is conceptually elegant but incremental.\n\nThe method mainly replaces the continuous classifier (score function) in diffusion guidance with a discrete optimality variable, which is a small modification rather than a fundamentally new algorithmic contribution.\n\nMuch of the theoretical framing follows directly from existing formulations of advantage-weighted regression and control-as-inference.\n\n2. Lack of comparative analysis with continuous guidance\n\nThe paper would be significantly stronger if it empirically compared continuous value-based guidance (e.g., by Q-values or advantages) versus discrete optimality-based guidance.\n\nSuch a comparison could clarify what is actually gained by discretizing optimality in this setting.\n\nWithout this, CFGRL appears as a special case of prior decision-diffuser-style methods using advantage-conditioned diffusion.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Classifier-Free Guided Reinforcement Learning (CFGRL), which interprets the classifier-free guidance (CFG) mechanism of diffusion models as a policy improvement operator.\nThe authors show that by treating the optimality condition as a discrete binary variable $o\\in \\{0,1\\}$, CFG can be reinterpreted as applying an advantage-weighted transformation on a reference policy. This allows controllable policy improvement via the guidance weight $w$, similar in spirit to temperature or KL coefficients in regularized RL. Experiments demonstrate that CFGRL improves over Advantage-Weighted Regression (AWR) and Goal-Conditioned Behavioral Cloning (GCBC) across offline RL benchmarks.\n- An LLM was used to improve writing.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. Simplicity and practical appeal – The method requires only standard diffusion training and allows tuning the improvement strength $w$ at inference, offering a practical way to control policy quality without retraining.\n\n2. Solid empirical demonstration – Results on offline RL and goal-conditioned control tasks consistently show improvements over strong baselines such as AWR and GCBC.\n\n3. Readable and well-presented – The paper is clearly written, with theoretical and empirical sections well balanced.", "weaknesses": "1. Limited novelty beyond reinterpretation\n\nThe core idea—recasting classifier-free guidance as a policy improvement operator—is conceptually elegant but incremental.\n\nThe method mainly replaces the continuous classifier (score function) in diffusion guidance with a discrete optimality variable, which is a small modification rather than a fundamentally new algorithmic contribution.\n\nMuch of the theoretical framing follows directly from existing formulations of advantage-weighted regression and control-as-inference.\n\n2. Lack of comparative analysis with continuous guidance\n\nThe paper would be significantly stronger if it empirically compared continuous value-based guidance (e.g., by Q-values or advantages) versus discrete optimality-based guidance.\n\nSuch a comparison could clarify what is actually gained by discretizing optimality in this setting.\n\nWithout this, CFGRL appears as a special case of prior decision-diffuser-style methods using advantage-conditioned diffusion.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761953826543}, {"id": "YH4EYrWCm2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14486/Reviewer_gz24"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces the CFGRL framework, which establishes a theoretical connection between classifier-Free diffusion guidance and the policy improvement operator in classical reinforcement learning. It formulates the improved policy as a \"product policy\", defined as a reference policy multiplied by an advantage-based optimality function. The authors highlight that the guidance weight $w$ enables flexible control over the extent of policy improvement at test time, without requiring retraining as in methods such as AWR. The proposed CFGRL framework is validated on both offline RL and goal-conditioned behavior cloning (GCBC) tasks, demonstrating superior performance over AWR and GCBC methods that can be interpreted as using a fixed guidance weight of 1. These findings provide evidence that guidance weights greater than one can effectively cooperate with diffusion models to achieve more efficient policy improvement.", "review_text": "This paper introduces the CFGRL framework, which establishes a theoretical connection between classifier-Free diffusion guidance and the policy improvement operator in classical reinforcement learning. It formulates the improved policy as a \"product policy\", defined as a reference policy multiplied by an advantage-based optimality function. The authors highlight that the guidance weight $w$ enables flexible control over the extent of policy improvement at test time, without requiring retraining as in methods such as AWR. The proposed CFGRL framework is validated on both offline RL and goal-conditioned behavior cloning (GCBC) tasks, demonstrating superior performance over AWR and GCBC methods that can be interpreted as using a fixed guidance weight of 1. These findings provide evidence that guidance weights greater than one can effectively cooperate with diffusion models to achieve more efficient policy improvement.", "strengths": "1. The paper is well written, and the presentation of results is clear and easy for readers to follow.\n\n2. To the best of my knowledge, this paper is the first to theoretically establish and prove the connection between classifier-free guided diffusion policy sampling and the policy improvement operator in RL.\n\n3. The authors’ analysis of AWR’s weakness in Section 5, together with the experimental observation that CFGRL can sustain larger guidance weights than AWR, constitutes an interesting result.", "weaknesses": "1. The main limitation of this paper lies in that most of its ideas have already appeared independently in prior works. For example, the relationship between classifier-free guidance and weighted regression has been discussed in [1], while the use of classifier-free guidance for policy improvement and the adjustment of different guidance weights was explored in [2]. Although the authors argue that [2] focuses on generating future state sequences whereas CFGRL generates single-step actions, I consider this distinction in output space relatively trivial. Apart from these aspects, the most notable contribution is the explicit formulation of the connection between the **policy improvement operator** and **classifier-free guided diffusion policy sampling**. However, this result is not particularly surprising, and the theoretical derivation itself is fairly straightforward. Considering this as the paper's primary contribution, it remains questionable whether the work alone is sufficient to support publication at ICLR.\n\n2. Moreover, the baseline methods in the experiments are relatively weak, and the proposed algorithm does not obtain a significant performance advantage under these settings, which limits the overall algorithmic contribution. Although the authors emphasize that \"By itself, CFGRL does not represent a state-of-the-art RL algorithm, but rather an additional tool in the algorithm designer's toolbox,\" the paper would be more compelling if the authors could further demonstrate how innovative algorithms can be derived using the CFGRL framework.\n\n[1] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" *arXiv preprint arXiv:2207.12598* (2022).\n\n[2] Ajay, Anurag, et al. \"Is Conditional Generative Modeling all you need for Decision Making?.\" *The Eleventh International Conference on Learning Representations*.", "questions": "1. What kinds of more advanced algorithms do the authors expect could be derived from the CFGRL framework, beyond current implementations that largely reproduce methods already proposed in prior work?\n\n2. Beyond GCBC and offline RL, what other problem domains could the CFGRL framework be applied to?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the CFGRL framework, which establishes a theoretical connection between classifier-Free diffusion guidance and the policy improvement operator in classical reinforcement learning. It formulates the improved policy as a \"product policy\", defined as a reference policy multiplied by an advantage-based optimality function. The authors highlight that the guidance weight $w$ enables flexible control over the extent of policy improvement at test time, without requiring retraining as in methods such as AWR. The proposed CFGRL framework is validated on both offline RL and goal-conditioned behavior cloning (GCBC) tasks, demonstrating superior performance over AWR and GCBC methods that can be interpreted as using a fixed guidance weight of 1. These findings provide evidence that guidance weights greater than one can effectively cooperate with diffusion models to achieve more efficient policy improvement.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well written, and the presentation of results is clear and easy for readers to follow.\n\n2. To the best of my knowledge, this paper is the first to theoretically establish and prove the connection between classifier-free guided diffusion policy sampling and the policy improvement operator in RL.\n\n3. The authors’ analysis of AWR’s weakness in Section 5, together with the experimental observation that CFGRL can sustain larger guidance weights than AWR, constitutes an interesting result.", "weaknesses": "1. The main limitation of this paper lies in that most of its ideas have already appeared independently in prior works. For example, the relationship between classifier-free guidance and weighted regression has been discussed in [1], while the use of classifier-free guidance for policy improvement and the adjustment of different guidance weights was explored in [2]. Although the authors argue that [2] focuses on generating future state sequences whereas CFGRL generates single-step actions, I consider this distinction in output space relatively trivial. Apart from these aspects, the most notable contribution is the explicit formulation of the connection between the **policy improvement operator** and **classifier-free guided diffusion policy sampling**. However, this result is not particularly surprising, and the theoretical derivation itself is fairly straightforward. Considering this as the paper's primary contribution, it remains questionable whether the work alone is sufficient to support publication at ICLR.\n\n2. Moreover, the baseline methods in the experiments are relatively weak, and the proposed algorithm does not obtain a significant performance advantage under these settings, which limits the overall algorithmic contribution. Although the authors emphasize that \"By itself, CFGRL does not represent a state-of-the-art RL algorithm, but rather an additional tool in the algorithm designer's toolbox,\" the paper would be more compelling if the authors could further demonstrate how innovative algorithms can be derived using the CFGRL framework.\n\n[1] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" *arXiv preprint arXiv:2207.12598* (2022).\n\n[2] Ajay, Anurag, et al. \"Is Conditional Generative Modeling all you need for Decision Making?.\" *The Eleventh International Conference on Learning Representations*.", "questions": "1. What kinds of more advanced algorithms do the authors expect could be derived from the CFGRL framework, beyond current implementations that largely reproduce methods already proposed in prior work?\n\n2. Beyond GCBC and offline RL, what other problem domains could the CFGRL framework be applied to?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761818647453}, {"id": "NMcVOXeqao", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14486/Reviewer_y8PU"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper reframes classifier-free guidance [1] from flow models as a controllable policy-improvement operator in RL, sampling from a prior policy tilted by an optimality term. Empirically, the proposed CFGRL improves over imitation/weighted-extraction baselines such as Goal-conditioned BC and AWR [2].\n\n[1] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" arXiv preprint arXiv:2207.12598 (2022).\n\n[2] Peng, Xue Bin, et al. \"Advantage-weighted regression: Simple and scalable off-policy reinforcement learning.\" arXiv preprint arXiv:1910.00177 (2019).", "review_text": "The paper reframes classifier-free guidance [1] from flow models as a controllable policy-improvement operator in RL, sampling from a prior policy tilted by an optimality term. Empirically, the proposed CFGRL improves over imitation/weighted-extraction baselines such as Goal-conditioned BC and AWR [2].\n\n[1] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" arXiv preprint arXiv:2207.12598 (2022).\n\n[2] Peng, Xue Bin, et al. \"Advantage-weighted regression: Simple and scalable off-policy reinforcement learning.\" arXiv preprint arXiv:1910.00177 (2019).", "strengths": "1. Theory & clarity.\n\nThe theoretical development is solid. The paper cleanly formalizes the link between generative-model guidance and policy improvement, and presents it in a way that is both accessible and precise.\n\n2. Guidance weight $w$: evidence matches the claim.\n\nFigure 4 empirically supports Remark 2: increasing the classifier-free guidance weight $w$ consistently steers the policy toward better directions. The additional results in the Appendix (Fig. 6) further corroborate this trend across settings, strengthening the central guidance-as-control narrative.\n\n3. CFGRL for GCBC: clear setup, convincing gains.\n\nSection 6 is well-specified and persuasive. The experiments substantiate the claim that CFGRL provides a principled, tunable alternative to standard GCBC: whereas prior GCBC effectively fixes $w=1$, the proposed framework exposes $w$ as a controllable knob and yields improved performance accordingly. The alignment between the theoretical framing and the observed gains is compelling.", "weaknesses": "1. The paper positions CFGRL as a scalable policy-extraction tool inspired by classifier-free guidance, and shows consistent gains over GCBC/flow-GCBC on OGBench. I think the paper's central thesis is to scale up RL by leveraging the scalability of “modern generative learning,” and, by implication, generative-model-based techniques like CFGRL should match or exceed the performance of RL-based methods such as FQL [3], which are harder to scale under the author's claim (as noted at line 28). To substantiate the broader “scale up RL” thesis, it would be essential to include strong baselines. Including FQL results would clarify where the advantage of CFGRL lies, in competing with modern value-gradient methods rather than with an unexpressive policy-extraction scheme like AWR [4].\n\n[3] Park, Seohong, Qiyang Li, and Sergey Levine. \"Flow q-learning.\" arXiv preprint arXiv:2502.02538 (2025).\n\n[4] Park, Seohong, et al. \"Is value learning really the main bottleneck in offline RL?.\" Advances in Neural Information Processing Systems 37 (2024): 79029-79056.", "questions": "1. Positioning vs. SOTA offline RL: strengths and weaknesses.\n\nEven if CFGRL is not intended to replace SOTA offline RL, a clear trade-off analysis is needed. Please add representative value-/policy-gradient baselines (e.g., FQL) under matched protocols, and discuss where CFGRL is advantageous (simplicity, stability, test-time controllability $w$) and where it is weaker (asymptotic performance, OOD robustness, sample efficiency). A brief ablation on wall-clock, hyper-sensitivity, and compute/memory would help practitioners decide when to choose CFGRL as a “designer’s toolbox” component.\n\n\n2. Task selection in Table 2.\n\nOGBench contains many tasks, but Table 2 reports nine. What criteria determined this subset? Please provide a justification for the selection or expand the evaluation to a broader, representative slice of OGBench to reduce selection bias.\n\n\n3. Baselines for the sampling/extraction objective (Eq. 8).\n\nSince many policy-extraction schemes target the same tilted distribution, AWR alone is insufficient as a comparator. Please include Relative Trajectory Balance [5] as an additional baseline: it optimizes an equivalent target distribution in theory and achieves stronger performance than AWR on several D4RL tasks [6]. \n\n4. Scope beyond $o\\in\\{0,1,\\emptyset\\}$: return-conditioned guidance.\n\nSection 4.1 instantiates $o\\in\\{0,1,\\emptyset\\}$, but under the Control-as-an-Inference view, the optimality variable naturally extends to learned Q-values or discounted returns. This suggests a return-conditioned variant—analogous to RvS/Decision Transformer/Decision Diffuser [7][8][9]—where classifier-free guidance tilts actions by target return. Please clarify how the $o\\in\\{0,1,\\emptyset\\}$ setting in CFGRL compares to return-conditioned BC with classifier-free guidance (both conceptually and empirically). \n\n[5] Venkatraman, Siddarth, et al. \"Amortizing intractable inference in diffusion models for vision, language, and control.\" Advances in neural information processing systems 37 (2024): 76080-76114.\n\n[6] Fu, Justin, et al. \"D4RL: Datasets for Deep Data-Driven Reinforcement Learning.\"\n\n[7] Emmons, Scott, et al. \"RvS: What is Essential for Offline RL via Supervised Learning?.\" International Conference on Learning Representations.\n\n[8] Chen, Lili, et al. \"Decision transformer: Reinforcement learning via sequence modeling.\" Advances in neural information processing systems 34 (2021): 15084-15097.\n\n[9] Ajay, Anurag, et al. \"Is Conditional Generative Modeling all you need for Decision Making?.\" The Eleventh International Conference on Learning Representations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper reframes classifier-free guidance [1] from flow models as a controllable policy-improvement operator in RL, sampling from a prior policy tilted by an optimality term. Empirically, the proposed CFGRL improves over imitation/weighted-extraction baselines such as Goal-conditioned BC and AWR [2].\n\n[1] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" arXiv preprint arXiv:2207.12598 (2022).\n\n[2] Peng, Xue Bin, et al. \"Advantage-weighted regression: Simple and scalable off-policy reinforcement learning.\" arXiv preprint arXiv:1910.00177 (2019).", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Theory & clarity.\n\nThe theoretical development is solid. The paper cleanly formalizes the link between generative-model guidance and policy improvement, and presents it in a way that is both accessible and precise.\n\n2. Guidance weight $w$: evidence matches the claim.\n\nFigure 4 empirically supports Remark 2: increasing the classifier-free guidance weight $w$ consistently steers the policy toward better directions. The additional results in the Appendix (Fig. 6) further corroborate this trend across settings, strengthening the central guidance-as-control narrative.\n\n3. CFGRL for GCBC: clear setup, convincing gains.\n\nSection 6 is well-specified and persuasive. The experiments substantiate the claim that CFGRL provides a principled, tunable alternative to standard GCBC: whereas prior GCBC effectively fixes $w=1$, the proposed framework exposes $w$ as a controllable knob and yields improved performance accordingly. The alignment between the theoretical framing and the observed gains is compelling.", "weaknesses": "1. The paper positions CFGRL as a scalable policy-extraction tool inspired by classifier-free guidance, and shows consistent gains over GCBC/flow-GCBC on OGBench. I think the paper's central thesis is to scale up RL by leveraging the scalability of “modern generative learning,” and, by implication, generative-model-based techniques like CFGRL should match or exceed the performance of RL-based methods such as FQL [3], which are harder to scale under the author's claim (as noted at line 28). To substantiate the broader “scale up RL” thesis, it would be essential to include strong baselines. Including FQL results would clarify where the advantage of CFGRL lies, in competing with modern value-gradient methods rather than with an unexpressive policy-extraction scheme like AWR [4].\n\n[3] Park, Seohong, Qiyang Li, and Sergey Levine. \"Flow q-learning.\" arXiv preprint arXiv:2502.02538 (2025).\n\n[4] Park, Seohong, et al. \"Is value learning really the main bottleneck in offline RL?.\" Advances in Neural Information Processing Systems 37 (2024): 79029-79056.", "questions": "1. Positioning vs. SOTA offline RL: strengths and weaknesses.\n\nEven if CFGRL is not intended to replace SOTA offline RL, a clear trade-off analysis is needed. Please add representative value-/policy-gradient baselines (e.g., FQL) under matched protocols, and discuss where CFGRL is advantageous (simplicity, stability, test-time controllability $w$) and where it is weaker (asymptotic performance, OOD robustness, sample efficiency). A brief ablation on wall-clock, hyper-sensitivity, and compute/memory would help practitioners decide when to choose CFGRL as a “designer’s toolbox” component.\n\n\n2. Task selection in Table 2.\n\nOGBench contains many tasks, but Table 2 reports nine. What criteria determined this subset? Please provide a justification for the selection or expand the evaluation to a broader, representative slice of OGBench to reduce selection bias.\n\n\n3. Baselines for the sampling/extraction objective (Eq. 8).\n\nSince many policy-extraction schemes target the same tilted distribution, AWR alone is insufficient as a comparator. Please include Relative Trajectory Balance [5] as an additional baseline: it optimizes an equivalent target distribution in theory and achieves stronger performance than AWR on several D4RL tasks [6]. \n\n4. Scope beyond $o\\in\\{0,1,\\emptyset\\}$: return-conditioned guidance.\n\nSection 4.1 instantiates $o\\in\\{0,1,\\emptyset\\}$, but under the Control-as-an-Inference view, the optimality variable naturally extends to learned Q-values or discounted returns. This suggests a return-conditioned variant—analogous to RvS/Decision Transformer/Decision Diffuser [7][8][9]—where classifier-free guidance tilts actions by target return. Please clarify how the $o\\in\\{0,1,\\emptyset\\}$ setting in CFGRL compares to return-conditioned BC with classifier-free guidance (both conceptually and empirically). \n\n[5] Venkatraman, Siddarth, et al. \"Amortizing intractable inference in diffusion models for vision, language, and control.\" Advances in neural information processing systems 37 (2024): 76080-76114.\n\n[6] Fu, Justin, et al. \"D4RL: Datasets for Deep Data-Driven Reinforcement Learning.\"\n\n[7] Emmons, Scott, et al. \"RvS: What is Essential for Offline RL via Supervised Learning?.\" International Conference on Learning Representations.\n\n[8] Chen, Lili, et al. \"Decision transformer: Reinforcement learning via sequence modeling.\" Advances in neural information processing systems 34 (2021): 15084-15097.\n\n[9] Ajay, Anurag, et al. \"Is Conditional Generative Modeling all you need for Decision Making?.\" The Eleventh International Conference on Learning Representations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761471281293}], "openreview_url": "https://openreview.net/forum?id=PwaUsAk4i6", "arxiv_id": "2505.23458", "paper_pdf": "papers/PwaUsAk4i6.pdf", "paper_pdf_sha256": "e78e3498ca4539c1394a48a0dda7a1c517cb3bfce95f914fe8a25347b8409275", "paper_pdf_bytes": 3346170, "paper_pdf_source": "openreview", "code_url": "https://github.com/kvfrans/cfgrl", "code_repository": "kvfrans/cfgrl", "code_commit": "14bd57b53bea27a3951687387520f11fbbe44abb", "code_archive": "repos/PwaUsAk4i6.zip", "code_archive_sha256": "4a0ea7596a89ca9db733f3ec8fc01ba55d44e321738092b31df6405bd756af09", "code_archive_bytes": 440989, "code_file_count": 42, "code_extensions": {".py": 42}, "github_disk_usage_kb": 412, "github_languages": {"Python": 308759}, "github_archived": false, "github_pushed_at": "2025-05-31T01:00:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/diffusion-guidance-is-a-controllable-policy"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "aya06N6R4W", "year": 2025, "status": "rejected", "title": "Counterfactual Causal Inference in Natural Language with Large Language Models", "authors": ["Gael Gendron", "Joze M. Rozanec", "Michael Witbrock", "Gillian Dobbie"], "authorids": ["~Gael_Gendron1", "~Joze_M._Rozanec1", "~Michael_Witbrock1", "~Gillian_Dobbie1"], "authors_source": "OpenReview API", "abstract": "Causal structure discovery methods are commonly applied to structured data where the causal variables are known and where statistical testing can be used to assess the causal relationships. By contrast, recovering a causal structure from unstructured natural language data such as news articles contains numerous challenges due to the absence of known variables or counterfactual data to estimate the causal links. Large Language Models (LLMs) have shown promising results in this direction but also exhibit limitations. This work investigates LLM's abilities to build causal graphs from text documents and perform counterfactual causal inference. We propose an end-to-end causal structure discovery and causal inference method from natural language: we first use an LLM to extract the instantiated causal variables from text data and build a causal graph. We merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible. We then conduct counterfactual inference on the estimated graph. The causal graph conditioning allows reduction of LLM biases and better represents the causal estimands. We use our method to show that the limitations in the counterfactual causal reasoning abilities come from prediction errors and propose directions to mitigate them. We demonstrate the applicability of our method on real-world news articles.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jKxU2ioeI2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6811/Reviewer_v566"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper proposes an approach for causal discovery and counterfactual inference from unstructured natural language. The method uses LLMs to extract causal variables, build causal graphs, and conduct counterfactual reasoning based on these inferred structures. The authors introduce a two-step framework: first, they prompt the LLM to generate causal graphs by identifying relationships between events in the text, potentially merging graphs from various documents for broader coverage. Then, the inferred graphs are used to simulate counterfactual scenarios, allowing them to analyze and evaluate causal influences with an LLM. The authors conduct experiments on synthetic and real-world datasets to illustrate the model's ability to handle counterfactual queries and identify sources of inference limitations, primarily prediction errors in reasoning tasks. Through these experiments, the paper highlights limitations in current LLM capabilities regarding counterfactual reasoning and suggests that prediction errors are a significant bottleneck.", "review_text": "The paper proposes an approach for causal discovery and counterfactual inference from unstructured natural language. The method uses LLMs to extract causal variables, build causal graphs, and conduct counterfactual reasoning based on these inferred structures. The authors introduce a two-step framework: first, they prompt the LLM to generate causal graphs by identifying relationships between events in the text, potentially merging graphs from various documents for broader coverage. Then, the inferred graphs are used to simulate counterfactual scenarios, allowing them to analyze and evaluate causal influences with an LLM. The authors conduct experiments on synthetic and real-world datasets to illustrate the model's ability to handle counterfactual queries and identify sources of inference limitations, primarily prediction errors in reasoning tasks. Through these experiments, the paper highlights limitations in current LLM capabilities regarding counterfactual reasoning and suggests that prediction errors are a significant bottleneck.", "strengths": "The main strength of the paper is in exploring how LLMs can support causal reasoning tasks on text data, pointing to new possibilities and practical challenges. The authors introduce a novel approach for discovering causal structures and conducting counterfactual inference directly from unstructured text using LLMs. \n\nThe methodology is well-defined, with clear steps in building and merging the causal graphs. \n\nThe experiments, both on synthetic and on real data, demonstrate the applicability of the framework.", "weaknesses": "1. The paper’s reliance on synthetic data from Cladder restricts its assessment of model performance in realistic contexts. The model is only demonstrated on a few real-world news examples, which doesn’t fully establish its applicability to unstructured text data beyond controlled scenarios.\n\n2. Expanding the analysis to include diverse domains would provide a stronger basis for understanding the versatility and potential challenges fro the LLM-based approach.\n\n3. The paper could benefit from quantitative performance metrics on real-world data (e.g., precision and recall for causal variable extraction, fidelity of counterfactual predictions).\n\n4. The paper would be improved by addressing issues like hallucination and prompt sensitivity, as these can impact the accuracy of causal inference.", "questions": "1. Could you provide additional quantitative evaluations on real-world datasets beyond the limited examples in the paper? Specifically, metrics like precision, recall, or causal graph fidelity in real news articles or other domains.\n\n2. Given the reliance on the synthetic Cladder data, do you see limitations in how well this dataset represents real-world causal complexity? \n\n3. How do you think the approach handles potential issues with LLM hallucination or prompt sensitivity?\n\n4. The paper mentions that counterfactual inference accuracy is a primary limitation, largely due to prediction errors. Could you provide more details on specific failure cases?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an approach for causal discovery and counterfactual inference from unstructured natural language. The method uses LLMs to extract causal variables, build causal graphs, and conduct counterfactual reasoning based on these inferred structures. The authors introduce a two-step framework: first, they prompt the LLM to generate causal graphs by identifying relationships between events in the text, potentially merging graphs from various documents for broader coverage. Then, the inferred graphs are used to simulate counterfactual scenarios, allowing them to analyze and evaluate causal influences with an LLM. The authors conduct experiments on synthetic and real-world datasets to illustrate the model's ability to handle counterfactual queries and identify sources of inference limitations, primarily prediction errors in reasoning tasks. Through these experiments, the paper highlights limitations in current LLM capabilities regarding counterfactual reasoning and suggests that prediction errors are a significant bottleneck.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "The main strength of the paper is in exploring how LLMs can support causal reasoning tasks on text data, pointing to new possibilities and practical challenges. The authors introduce a novel approach for discovering causal structures and conducting counterfactual inference directly from unstructured text using LLMs. \n\nThe methodology is well-defined, with clear steps in building and merging the causal graphs. \n\nThe experiments, both on synthetic and on real data, demonstrate the applicability of the framework.", "weaknesses": "1. The paper’s reliance on synthetic data from Cladder restricts its assessment of model performance in realistic contexts. The model is only demonstrated on a few real-world news examples, which doesn’t fully establish its applicability to unstructured text data beyond controlled scenarios.\n\n2. Expanding the analysis to include diverse domains would provide a stronger basis for understanding the versatility and potential challenges fro the LLM-based approach.\n\n3. The paper could benefit from quantitative performance metrics on real-world data (e.g., precision and recall for causal variable extraction, fidelity of counterfactual predictions).\n\n4. The paper would be improved by addressing issues like hallucination and prompt sensitivity, as these can impact the accuracy of causal inference.", "questions": "1. Could you provide additional quantitative evaluations on real-world datasets beyond the limited examples in the paper? Specifically, metrics like precision, recall, or causal graph fidelity in real news articles or other domains.\n\n2. Given the reliance on the synthetic Cladder data, do you see limitations in how well this dataset represents real-world causal complexity? \n\n3. How do you think the approach handles potential issues with LLM hallucination or prompt sensitivity?\n\n4. The paper mentions that counterfactual inference accuracy is a primary limitation, largely due to prediction errors. Could you provide more details on specific failure cases?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730766131587}, {"id": "YOUeNB7rRV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6811/Reviewer_TnB7"], "rating": 3, "soundness": 4, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "The paper studies timely and important goal: how do we get AI systems to reason causally over natural text? They extend LLM results on inferring causal graphs to natural text datasets and perform counterfactual inference.", "review_text": "The paper studies timely and important goal: how do we get AI systems to reason causally over natural text? They extend LLM results on inferring causal graphs to natural text datasets and perform counterfactual inference.", "strengths": "* The motivation of the problem and the datasets chosen are excellent. \n* Experiments are clearly described and ablations are provided", "weaknesses": "My brief review is that the authors have done a great job setting up the problem, but their results are underwhelming. \n* The proposed method introduces a lot of complex parts, but is unable to outperform the causalCoT baseline from Jin et al.\n* I do not get any new insight from the results, especially because all the validation is also done by LLMs. \n* While the idea is interesting, the paper's results fail to convince me that it works in practice. At least, we should see stronger results on the synthetic cladder benchmark.", "questions": "I have some questions below. I also provide feedback to improve the work, which the authors need not respond to.\n\nQuestions\n* Is the graph generated per sentence? Or per paragraph?\n* How are two graphs merged? Do you also use an LLM for this task? How do you handle variable name mismatches, mismatches in edge direction, etc,?\n* How does the work relate to LLMs+causal representation learning? E.g.,this paper https://arxiv.org/abs/2402.03941\n* Section 3.1 states that hidden variables are also extracted from LLM. How are they used in the causal analysis? In a dataset like cladder, does the LLM output hidden variables?\n* The decision to simply ignore any questions where the LLM outputs an incorrectly formatted answer or other error is not fair. Can you present results on the entire cladder dataset? If the LLM failed to produce a valid graph, then it should be counted as a failure, since it is part of the proposed method.\n* Can you think of any other validation apart from LLM self-validation? It's unclear what that evaluation is adding.\n\nSuggestions to improve:\n* I would suggest the authors to focus on cladder and show a demonstrable gain. Otherwise, it is even harder to interpret the real-world oil dataset results.\n* I suspect that your method is trusting the LLM too much. Instead, once the graph is obtained, we can do many operations symbolically and invoke LLMs only for small, focused tasks. So I would encourage you to look into a direction where the central control is handled by the SCM and CF procedure and the modular atomic tasks are delegated to LLMs. Some iteration of this idea should easily give higher accuracies on cladder, assuming that the graph is correct.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies timely and important goal: how do we get AI systems to reason causally over natural text? They extend LLM results on inferring causal graphs to natural text datasets and perform counterfactual inference.", "soundness": 4, "presentation": 3, "contribution": 1, "strengths": "* The motivation of the problem and the datasets chosen are excellent. \n* Experiments are clearly described and ablations are provided", "weaknesses": "My brief review is that the authors have done a great job setting up the problem, but their results are underwhelming. \n* The proposed method introduces a lot of complex parts, but is unable to outperform the causalCoT baseline from Jin et al.\n* I do not get any new insight from the results, especially because all the validation is also done by LLMs. \n* While the idea is interesting, the paper's results fail to convince me that it works in practice. At least, we should see stronger results on the synthetic cladder benchmark.", "questions": "I have some questions below. I also provide feedback to improve the work, which the authors need not respond to.\n\nQuestions\n* Is the graph generated per sentence? Or per paragraph?\n* How are two graphs merged? Do you also use an LLM for this task? How do you handle variable name mismatches, mismatches in edge direction, etc,?\n* How does the work relate to LLMs+causal representation learning? E.g.,this paper https://arxiv.org/abs/2402.03941\n* Section 3.1 states that hidden variables are also extracted from LLM. How are they used in the causal analysis? In a dataset like cladder, does the LLM output hidden variables?\n* The decision to simply ignore any questions where the LLM outputs an incorrectly formatted answer or other error is not fair. Can you present results on the entire cladder dataset? If the LLM failed to produce a valid graph, then it should be counted as a failure, since it is part of the proposed method.\n* Can you think of any other validation apart from LLM self-validation? It's unclear what that evaluation is adding.\n\nSuggestions to improve:\n* I would suggest the authors to focus on cladder and show a demonstrable gain. Otherwise, it is even harder to interpret the real-world oil dataset results.\n* I suspect that your method is trusting the LLM too much. Instead, once the graph is obtained, we can do many operations symbolically and invoke LLMs only for small, focused tasks. So I would encourage you to look into a direction where the central control is handled by the SCM and CF procedure and the modular atomic tasks are delegated to LLMs. Some iteration of this idea should easily give higher accuracies on cladder, assuming that the graph is correct.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730700204869}, {"id": "iPw9ekHXZn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6811/Reviewer_L7ZX"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper studies causal discovery by using large language models to extract the causal graph from textual documents by prompting and perform sequence prediction. Experiments were conducted on both synthetic and more realistic datasets.", "review_text": "This paper studies causal discovery by using large language models to extract the causal graph from textual documents by prompting and perform sequence prediction. Experiments were conducted on both synthetic and more realistic datasets.", "strengths": "* The problem itself is of interest to a wide audience\n* The idea of applying LLM to this problem may inspire more researchers", "weaknesses": "My main concern is that the contribution seems less significnat compared with what authors claimed -- in particular:\n\n1. the main methodology, discussed in Section 3, might be summed up as prompting LLM and getting the results. I do not see novel ways of applying the LLMs to this specific problem (I agree there was effort in designing prompts to get the model generate a JSON-formated causal graph, but that seems not to be a significnat contribution), or insights/evidence of how this approach is better and in what ways.\n\n2. There are assumptions needed for equation (1) to work and I don't see mentions/discussions on their validaity in this setup, nor do I see a clear definition on the variables used throught section 3. For example, in the unnumbered introduction paragraph of Section 3, I understand that the causal variables/nodes refer to as noun pharases while in Section 3.2 the main focus was to adjust the conditional probability of predicted probability of tokens. There is no discussion on how one is related to the other; for example, what do you try to quantify/estimate from $P(Y|do(x),x',y')$? How does it relate to say the example in the unnumbered introduction paragraph of Section 1 (\"travel restrictions diminishes airlines companies revenues\")?\n\n3. Although I appreciate the authors' efforst in doing a relative thorough experimental studies using various LLMs. I do think the results warant more discussions. For example, in Table 1, it was noted that \"Only extracted answers are shown,\" how many exactractions failed? Does it introduce systemic bias in the evaluation?", "questions": "In addition to several questions I asked in the Weaknesses section --\n\n1. Can you elaborate what variables are being intervenes and what causal effects are being studied in Section 3.2 when you write $P(Y_0|C)$ and how does this (the conditional probability of an output token $Y_0$) relates to the causal question you are trying to study?\n\n2. In Section 3.2, the authors wrote \" Due to their extensive training on a massive amount of data, LLMs are good estimators,\" I don't see why this necessarily implies they are also good estimators when you replace the contexts (tokens) by causal variables (say $x'$). Can you provide more details?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies causal discovery by using large language models to extract the causal graph from textual documents by prompting and perform sequence prediction. Experiments were conducted on both synthetic and more realistic datasets.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* The problem itself is of interest to a wide audience\n* The idea of applying LLM to this problem may inspire more researchers", "weaknesses": "My main concern is that the contribution seems less significnat compared with what authors claimed -- in particular:\n\n1. the main methodology, discussed in Section 3, might be summed up as prompting LLM and getting the results. I do not see novel ways of applying the LLMs to this specific problem (I agree there was effort in designing prompts to get the model generate a JSON-formated causal graph, but that seems not to be a significnat contribution), or insights/evidence of how this approach is better and in what ways.\n\n2. There are assumptions needed for equation (1) to work and I don't see mentions/discussions on their validaity in this setup, nor do I see a clear definition on the variables used throught section 3. For example, in the unnumbered introduction paragraph of Section 3, I understand that the causal variables/nodes refer to as noun pharases while in Section 3.2 the main focus was to adjust the conditional probability of predicted probability of tokens. There is no discussion on how one is related to the other; for example, what do you try to quantify/estimate from $P(Y|do(x),x',y')$? How does it relate to say the example in the unnumbered introduction paragraph of Section 1 (\"travel restrictions diminishes airlines companies revenues\")?\n\n3. Although I appreciate the authors' efforst in doing a relative thorough experimental studies using various LLMs. I do think the results warant more discussions. For example, in Table 1, it was noted that \"Only extracted answers are shown,\" how many exactractions failed? Does it introduce systemic bias in the evaluation?", "questions": "In addition to several questions I asked in the Weaknesses section --\n\n1. Can you elaborate what variables are being intervenes and what causal effects are being studied in Section 3.2 when you write $P(Y_0|C)$ and how does this (the conditional probability of an output token $Y_0$) relates to the causal question you are trying to study?\n\n2. In Section 3.2, the authors wrote \" Due to their extensive training on a massive amount of data, LLMs are good estimators,\" I don't see why this necessarily implies they are also good estimators when you replace the contexts (tokens) by causal variables (say $x'$). Can you provide more details?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730673656542}, {"id": "ecTE4WFoKo", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6811/Reviewer_Nzcs"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "They explore the challenges and potential of using large language models (LLMs) to infer causal relationships from unstructured text data, such as news articles. Then they present a method for discovering causal structures and conducting counterfactual causal inference using LLMs. They propose a pipeline approach: \n1. They first use an LLM to extract the instantiated causal variables from text data and build a causal graph;\n2.  Then they merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible; \n3. Finally, they conduct counterfactual inference on the estimated graph. \n\nI think the paper is well-written. However, the main and only concern for me, for each of the aforementioned step above, they are many works already. To name a few: \n+ Causal discovery with LLMs: (i) Efficient Causal Graph Discovery Using Large Language Models; (ii) Causal Graph Discovery with Retrieval-Augmented Generation based Large Language Models\n+ Counterfactual inference with LLMs: (i) Using LLMs for Explaining Sets of Counterfactual Examples to Final Users; (ii) Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations\n\n\nThis work is more like a combination of different causal tasks with a novel end-to-end method. I admit there are merits for this end-to-end method but I think more discussion on the related work is necessary. I am open to other reviews' opinions about the novelty. \n\nFor this paper's soundness, I think this paper is well-written and the experiments are quite supportive.", "review_text": "They explore the challenges and potential of using large language models (LLMs) to infer causal relationships from unstructured text data, such as news articles. Then they present a method for discovering causal structures and conducting counterfactual causal inference using LLMs. They propose a pipeline approach: \n1. They first use an LLM to extract the instantiated causal variables from text data and build a causal graph;\n2.  Then they merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible; \n3. Finally, they conduct counterfactual inference on the estimated graph. \n\nI think the paper is well-written. However, the main and only concern for me, for each of the aforementioned step above, they are many works already. To name a few: \n+ Causal discovery with LLMs: (i) Efficient Causal Graph Discovery Using Large Language Models; (ii) Causal Graph Discovery with Retrieval-Augmented Generation based Large Language Models\n+ Counterfactual inference with LLMs: (i) Using LLMs for Explaining Sets of Counterfactual Examples to Final Users; (ii) Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations\n\n\nThis work is more like a combination of different causal tasks with a novel end-to-end method. I admit there are merits for this end-to-end method but I think more discussion on the related work is necessary. I am open to other reviews' opinions about the novelty. \n\nFor this paper's soundness, I think this paper is well-written and the experiments are quite supportive.", "strengths": "1. The paper presents a novel approach by leveraging large language models to perform causal structure discovery and counterfactual inference from unstructured text.\n2. The authors propose a comprehensive, end-to-end method that extracts causal graphs, performs interventions, and predicts counterfactual scenarios.\n3. By incorporating causal graph conditioning, the method helps reduce biases inherent in LLM.", "weaknesses": "The main weakness is listed in their Limitation section, I agree with their pointed limitation and like their honesty: \n1. The counterfactual setting is mainly on synthetic datasets. \n2. The method currently focuses on DAGs and does not account for feedback loops, which are common in real-world events.", "questions": "1. I think the existing content(experiment&writting) is very sound. But the novelty is not that clear to me. The end-to-end way is more like a pipeline for different phases of causal tasks with LLMs. Namely, for both causal graph discovery and counterfactual inference, there are already existing works that use LLMs for this task. But I admit the end-to-end framework proposed in this paper is nove. \n2. How do you validate the accuracy of the causal graphs generated by the LLMs, particularly when no ground-truth data is available for real-world scenarios? I guess this evaluation is based on synthetic datasets where the ground truth label is known?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "They explore the challenges and potential of using large language models (LLMs) to infer causal relationships from unstructured text data, such as news articles. Then they present a method for discovering causal structures and conducting counterfactual causal inference using LLMs. They propose a pipeline approach: \n1. They first use an LLM to extract the instantiated causal variables from text data and build a causal graph;\n2.  Then they merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible; \n3. Finally, they conduct counterfactual inference on the estimated graph. \n\nI think the paper is well-written. However, the main and only concern for me, for each of the aforementioned step above, they are many works already. To name a few: \n+ Causal discovery with LLMs: (i) Efficient Causal Graph Discovery Using Large Language Models; (ii) Causal Graph Discovery with Retrieval-Augmented Generation based Large Language Models\n+ Counterfactual inference with LLMs: (i) Using LLMs for Explaining Sets of Counterfactual Examples to Final Users; (ii) Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations\n\n\nThis work is more like a combination of different causal tasks with a novel end-to-end method. I admit there are merits for this end-to-end method but I think more discussion on the related work is necessary. I am open to other reviews' opinions about the novelty. \n\nFor this paper's soundness, I think this paper is well-written and the experiments are quite supportive.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper presents a novel approach by leveraging large language models to perform causal structure discovery and counterfactual inference from unstructured text.\n2. The authors propose a comprehensive, end-to-end method that extracts causal graphs, performs interventions, and predicts counterfactual scenarios.\n3. By incorporating causal graph conditioning, the method helps reduce biases inherent in LLM.", "weaknesses": "The main weakness is listed in their Limitation section, I agree with their pointed limitation and like their honesty: \n1. The counterfactual setting is mainly on synthetic datasets. \n2. The method currently focuses on DAGs and does not account for feedback loops, which are common in real-world events.", "questions": "1. I think the existing content(experiment&writting) is very sound. But the novelty is not that clear to me. The end-to-end way is more like a pipeline for different phases of causal tasks with LLMs. Namely, for both causal graph discovery and counterfactual inference, there are already existing works that use LLMs for this task. But I admit the end-to-end framework proposed in this paper is nove. \n2. How do you validate the accuracy of the causal graphs generated by the LLMs, particularly when no ground-truth data is available for real-world scenarios? I guess this evaluation is based on synthetic datasets where the ground truth label is known?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729503976748}], "openreview_url": "https://openreview.net/forum?id=aya06N6R4W", "arxiv_id": "2410.06392", "paper_pdf": "papers/aya06N6R4W.pdf", "paper_pdf_sha256": "6651cc456ba0319d57b39b1a2d483c3e95b465ce4f07f676ef14805cc9df62bb", "paper_pdf_bytes": 389421, "paper_pdf_source": "openreview", "code_url": "https://github.com/Strong-AI-Lab/counterfactual-llm-inference", "code_repository": "Strong-AI-Lab/counterfactual-llm-inference", "code_commit": "0b25e48ac59bdf22417b449310d481616f96d54c", "code_archive": "repos/aya06N6R4W.zip", "code_archive_sha256": "f84cebb7aad9aabd847289d7a5c3c13867227712bb3224b614fa9f7d21fb8c79", "code_archive_bytes": 42383, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 52, "github_languages": {"Python": 116979}, "github_archived": false, "github_pushed_at": "2024-10-02T08:34:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/counterfactual-causal-inference-in-natural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uJPWeZffgl", "year": 2024, "status": "rejected", "title": "Convex and Bilevel Optimization for Neuro-Symbolic Inference and Learning", "authors": ["Charles Andrew Dickens", "Changyu Gao", "Connor Pryor", "Stephen Wright", "Lise Getoor"], "authorids": ["~Charles_Andrew_Dickens1", "~Changyu_Gao1", "~Connor_Pryor1", "~Stephen_Wright1", "~Lise_Getoor1"], "authors_source": "OpenReview API", "abstract": "We address a key challenge for neuro-symbolic (NeSy) systems by leveraging convex and bilevel optimization techniques to develop a general first-order gradient-based framework for end-to-end neural and symbolic parameter learning.\nSpecifically, we formulate NeSy learning as a bilevel program, and we employ Moreau smoothing and a graduated value-function approach to support learning with a constrained lower-level inference problem.\nThe applicability of our learning framework is demonstrated with NeuPSL, a state-of-the-art NeSy architecture.\nTo achieve this, we propose a primal and dual formulation of NeuPSL inference as a strongly convex linearly constrained quadratic program and show learning gradients are functions of the optimal dual variables.\nBased on this formulation, we develop a corresponding dual block coordinate descent algorithm that naturally exploits warm-starts. \nThis leads to over $100 \\times$ learning runtime improvements over the current state-of-the-art NeuPSL inference method.\nFinally, we provide extensive empirical evaluations across $8$ datasets covering a range of prediction tasks and demonstrate our learning framework achieves up to a $16$% point prediction performance improvement over the current standard learning process.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "0dpW8Br74H", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5927/Reviewer_oqJp"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors give an equivalent formulation of NeSy EBMs learning as a bilevel problem. Such formulation allows smooth first-order optimization.", "review_text": "The authors give an equivalent formulation of NeSy EBMs learning as a bilevel problem. Such formulation allows smooth first-order optimization.", "strengths": "The writing of the paper makes it very hard to identify many strengths. It is my understanding, however, that by operating at the level of NeSy EBMs, any advancement applies to (potentially) a broad class of neuro-symbolic methods.", "weaknesses": "- The paper is a neuro-symbolic AI approach for learning and inference with barely any mention of previous related work in the field [1, 2, 3, 4, 5, 6, ..].\n\n- The paper is **very** hard to read with not a single figure or running example to help with the exposition.\n\n- Empirical evaluation is only carried out on toy datasets, using very basic tasks (e.g. MNIST-addition is evaluated only using $2$ digits, which can be easily solved using existing baselines), and the numbers are presented with no extra commentary/analysis.\n\nReferences: \n\n[1] Semantic Probabilistic Layers for Neuro-Symbolic Learning. Kareem Ahmed, Stefano Teso, Kai-Wei Chang, Guy Van den Broeck, Antonio Vergari. NeurIPS 2022.\n\n[2] A Semantic Loss Function for Deep Learning with Symbolic Knowledge. Jingyi Xu, Zilu Zhang, Tal Friedman, Yitao Liang, Guy Van den Broeck. ICML 2018.\n\n[3] Semantic Strengthening of Neuro-Symbolic Learning. Kareem Ahmed, Kai-Wei Chang, Guy Van den Broeck. AISTATS 2023.\n\n[4] Neuro-Symbolic Entropy Regularization. Kareem Ahmed, Eric Wang, Kai-Wei Chang, Guy Van den Broeck. UAI 2022.\n\n[5] Coherent Hierarchical Multi−Label Classification Networks. Eleonora Giunchiglia and Thomas Lukasiewicz. NeurIPS 2022.\n\n[6] Deep Learning with Logical Constraints. Eleonora Giunchiglia‚ Mihaela Catalina Stoian and Thomas Lukasiewicz. IJCAI 2022.", "questions": "- I'm struggling to understand what the problem being solved here is exactly. The second paragraph in the introduction mentions that \"the predictions are not guaranteed to have an analytical form or be differentiable, and traditional deep learning techniques are not directly applicable\", could you please say more about that? All the NeSy AI techniques that I am aware of: semantic loss, deepproblog, NeSy entropy, semantic strengthening, semantic probabilistic layers, neupsl, etc... are able to train the models end-to-end using back-propagation.'\n\n- As a follow-up question: why should we be interested in developing this equivalent formulation as a bilevel problem? What could one possibly gain that improves upon the exact inference proposed in semantic loss, deepproblog, NeSy entropy and  semantic probabilistic layers?\n\n- I am struggling to understand the results obtained on MNIST-addition. How can it be that NeuPSL, which is approximate, can outperform DeepProbLog which performs exact inference?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors give an equivalent formulation of NeSy EBMs learning as a bilevel problem. Such formulation allows smooth first-order optimization.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "strengths": "The writing of the paper makes it very hard to identify many strengths. It is my understanding, however, that by operating at the level of NeSy EBMs, any advancement applies to (potentially) a broad class of neuro-symbolic methods.", "weaknesses": "- The paper is a neuro-symbolic AI approach for learning and inference with barely any mention of previous related work in the field [1, 2, 3, 4, 5, 6, ..].\n\n- The paper is **very** hard to read with not a single figure or running example to help with the exposition.\n\n- Empirical evaluation is only carried out on toy datasets, using very basic tasks (e.g. MNIST-addition is evaluated only using $2$ digits, which can be easily solved using existing baselines), and the numbers are presented with no extra commentary/analysis.\n\nReferences: \n\n[1] Semantic Probabilistic Layers for Neuro-Symbolic Learning. Kareem Ahmed, Stefano Teso, Kai-Wei Chang, Guy Van den Broeck, Antonio Vergari. NeurIPS 2022.\n\n[2] A Semantic Loss Function for Deep Learning with Symbolic Knowledge. Jingyi Xu, Zilu Zhang, Tal Friedman, Yitao Liang, Guy Van den Broeck. ICML 2018.\n\n[3] Semantic Strengthening of Neuro-Symbolic Learning. Kareem Ahmed, Kai-Wei Chang, Guy Van den Broeck. AISTATS 2023.\n\n[4] Neuro-Symbolic Entropy Regularization. Kareem Ahmed, Eric Wang, Kai-Wei Chang, Guy Van den Broeck. UAI 2022.\n\n[5] Coherent Hierarchical Multi−Label Classification Networks. Eleonora Giunchiglia and Thomas Lukasiewicz. NeurIPS 2022.\n\n[6] Deep Learning with Logical Constraints. Eleonora Giunchiglia‚ Mihaela Catalina Stoian and Thomas Lukasiewicz. IJCAI 2022.", "questions": "- I'm struggling to understand what the problem being solved here is exactly. The second paragraph in the introduction mentions that \"the predictions are not guaranteed to have an analytical form or be differentiable, and traditional deep learning techniques are not directly applicable\", could you please say more about that? All the NeSy AI techniques that I am aware of: semantic loss, deepproblog, NeSy entropy, semantic strengthening, semantic probabilistic layers, neupsl, etc... are able to train the models end-to-end using back-propagation.'\n\n- As a follow-up question: why should we be interested in developing this equivalent formulation as a bilevel problem? What could one possibly gain that improves upon the exact inference proposed in semantic loss, deepproblog, NeSy entropy and  semantic probabilistic layers?\n\n- I am struggling to understand the results obtained on MNIST-addition. How can it be that NeuPSL, which is approximate, can outperform DeepProbLog which performs exact inference?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698878719844}, {"id": "1XhJni65VG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5927/Reviewer_Zeyx"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "I am not sure why I got assigned this paper, which focuses on neuro-symbolic methods and is therefore fairly far away from my area of expertise. I feel only partially qualified to review it: I don't believe I bid on this paper, but if I did, this was certainly a misclick. While I am very familiar with optimization, I don't know much about neuro-symbolic methods, and this review will reflect that, but I will still do my best. I am very open to changing my mind about this paper based on feedback from the authors and other reviewers on what to direct my attention towards.\n\nThe authors study how to fuse symbolic processing with neural networks. They start with a bilevel optimization problem for neuro-symbolic learning, reformulate it as a constrained optimization problem with certain inequality-based constraints, then propose a smoothed variant of said constraints which replaces a certain energy-function with its Moreau envelope. This problem is optimized using an augmented-Lagrangian-based algorithm. The authors apply this approach to a neural probablistic soft logic model, and study optimization-theoretic properties of the resulting objective. The authors benchmark their approach on certain models, mainly against ADMM, and show that their technique performs much better in some cases.", "review_text": "I am not sure why I got assigned this paper, which focuses on neuro-symbolic methods and is therefore fairly far away from my area of expertise. I feel only partially qualified to review it: I don't believe I bid on this paper, but if I did, this was certainly a misclick. While I am very familiar with optimization, I don't know much about neuro-symbolic methods, and this review will reflect that, but I will still do my best. I am very open to changing my mind about this paper based on feedback from the authors and other reviewers on what to direct my attention towards.\n\nThe authors study how to fuse symbolic processing with neural networks. They start with a bilevel optimization problem for neuro-symbolic learning, reformulate it as a constrained optimization problem with certain inequality-based constraints, then propose a smoothed variant of said constraints which replaces a certain energy-function with its Moreau envelope. This problem is optimized using an augmented-Lagrangian-based algorithm. The authors apply this approach to a neural probablistic soft logic model, and study optimization-theoretic properties of the resulting objective. The authors benchmark their approach on certain models, mainly against ADMM, and show that their technique performs much better in some cases.", "strengths": "The proposed algorithm is shown to perform significantly on a number of benchmarks compared to ADMM. This makes sense, as the authors' method is specialized to their setting, whereas ADMM is generic. \n\nThe convex-optimization-based aspects seem reasonably well-thought-out. There are a lot of choices that one could make here, and the authors' choices seem reasonable.\n\nThe paper is reasonably well-written, though at times not easy to parse for someone outside the area. \n\nI am very interested in what the other reviewers, who will likely know much more about this area than me, will have to say, as their reviews are likely to help direct me to what are the most important parts of the work I should direct attention to, and will update my thoughts and edit my review accordingly once the time comes.", "weaknesses": "Way too many acronyms. I had trouble remembering what half of them stood for once I got far enough away from their definitions. This paper goes so far in the direction of using acronyms for everything, that I strongly recommend the authors go into the other extreme and remove *all* acronyms, since doing this will make the paper easier to read.\n\nVery little seems to actually happen in Section 2. The paper would be improved by reviewing some of the technical points in more detail, otherwise the description is so high-level that it is almost meaningless. In particular, the differences between implicit-differentiation-based methods and value-function approaches are not sufficiently explained, and are not clear to me even though I know exactly how to differentiate through an optimization problem using envelope theorems and similar. This is much more background than many readers will have, so if I'm confused, chances are most people will be. Please see questions.\n\nThe main technical sections are notation-heavy, and are quite complex to read. Theorem 5.3 in particular is somewhat hard-to-parse and takes a third of a page, in total, to state, even though its first main claim is a simply saying that a certain objective is convex-concave, and the other claims are also relatively simple.\n\nThe evaluation is purely quantitative: the authors show better numbers on tables compared to alternatives. While this is a valid way to evaluate, it also gives a much weaker idea of what is going on compared to evaluations that are not table-based and consider factors other than performance metrics. Nothing I saw in the experimental section rules out a situation of the form \"none of the algorithms work, but the authors' is slightly less broken\" - I would like to see some evidence that this isn't the case. \n* I've previously seen misleading results like this in reinforcement learning papers where an agent achieved \"good performance\" through random actions that were slightly-more-aligned with the objective compared to baselines, but was so far from correct behavior that most people would reasonably view both the method and the baselines as equally bad, and the differences in metrics as meaningless. In total, how do we know the neuro-symbolic system is performing as it should in this setting?", "questions": "I do not understand the described difference between implicit-differentiation-based methods and value-function approaches. \n* Is a value-function approach one where the lower-level objective is solved either analytically or to convergence, after which one applies a suitable envelope theorem approach to calculate the gradient of the objective using the optimal value?\n* In contrast, is an implicit-differentiation approach one where we do not solve the inner optimization problem to convergence? Or am I completely confused by what you mean by this distinction here?\n\nWhy is the constraint in (7) an inequality constraint, rather than an equality constraint? The paragraph directly after it mentions an equality constraint. Is this a typo?\n\nIs there a way to evaluate the performance of the resulting neuro-symbolic system qualitatively, to ensure it behaves in the manner that the algorithmic designer expects it does? Is there some kind of sanity check one could do to guard against the \"all the methods fail at the given task, but ours fails with slightly better numbers\" potential failure mode? While I certainly have no direct evidence that this is happening, it would make me feel much better about the paper if this could be definitively ruled out via the experiments.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "I am not sure why I got assigned this paper, which focuses on neuro-symbolic methods and is therefore fairly far away from my area of expertise. I feel only partially qualified to review it: I don't believe I bid on this paper, but if I did, this was certainly a misclick. While I am very familiar with optimization, I don't know much about neuro-symbolic methods, and this review will reflect that, but I will still do my best. I am very open to changing my mind about this paper based on feedback from the authors and other reviewers on what to direct my attention towards.\n\nThe authors study how to fuse symbolic processing with neural networks. They start with a bilevel optimization problem for neuro-symbolic learning, reformulate it as a constrained optimization problem with certain inequality-based constraints, then propose a smoothed variant of said constraints which replaces a certain energy-function with its Moreau envelope. This problem is optimized using an augmented-Lagrangian-based algorithm. The authors apply this approach to a neural probablistic soft logic model, and study optimization-theoretic properties of the resulting objective. The authors benchmark their approach on certain models, mainly against ADMM, and show that their technique performs much better in some cases.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The proposed algorithm is shown to perform significantly on a number of benchmarks compared to ADMM. This makes sense, as the authors' method is specialized to their setting, whereas ADMM is generic. \n\nThe convex-optimization-based aspects seem reasonably well-thought-out. There are a lot of choices that one could make here, and the authors' choices seem reasonable.\n\nThe paper is reasonably well-written, though at times not easy to parse for someone outside the area. \n\nI am very interested in what the other reviewers, who will likely know much more about this area than me, will have to say, as their reviews are likely to help direct me to what are the most important parts of the work I should direct attention to, and will update my thoughts and edit my review accordingly once the time comes.", "weaknesses": "Way too many acronyms. I had trouble remembering what half of them stood for once I got far enough away from their definitions. This paper goes so far in the direction of using acronyms for everything, that I strongly recommend the authors go into the other extreme and remove *all* acronyms, since doing this will make the paper easier to read.\n\nVery little seems to actually happen in Section 2. The paper would be improved by reviewing some of the technical points in more detail, otherwise the description is so high-level that it is almost meaningless. In particular, the differences between implicit-differentiation-based methods and value-function approaches are not sufficiently explained, and are not clear to me even though I know exactly how to differentiate through an optimization problem using envelope theorems and similar. This is much more background than many readers will have, so if I'm confused, chances are most people will be. Please see questions.\n\nThe main technical sections are notation-heavy, and are quite complex to read. Theorem 5.3 in particular is somewhat hard-to-parse and takes a third of a page, in total, to state, even though its first main claim is a simply saying that a certain objective is convex-concave, and the other claims are also relatively simple.\n\nThe evaluation is purely quantitative: the authors show better numbers on tables compared to alternatives. While this is a valid way to evaluate, it also gives a much weaker idea of what is going on compared to evaluations that are not table-based and consider factors other than performance metrics. Nothing I saw in the experimental section rules out a situation of the form \"none of the algorithms work, but the authors' is slightly less broken\" - I would like to see some evidence that this isn't the case. \n* I've previously seen misleading results like this in reinforcement learning papers where an agent achieved \"good performance\" through random actions that were slightly-more-aligned with the objective compared to baselines, but was so far from correct behavior that most people would reasonably view both the method and the baselines as equally bad, and the differences in metrics as meaningless. In total, how do we know the neuro-symbolic system is performing as it should in this setting?", "questions": "I do not understand the described difference between implicit-differentiation-based methods and value-function approaches. \n* Is a value-function approach one where the lower-level objective is solved either analytically or to convergence, after which one applies a suitable envelope theorem approach to calculate the gradient of the objective using the optimal value?\n* In contrast, is an implicit-differentiation approach one where we do not solve the inner optimization problem to convergence? Or am I completely confused by what you mean by this distinction here?\n\nWhy is the constraint in (7) an inequality constraint, rather than an equality constraint? The paragraph directly after it mentions an equality constraint. Is this a typo?\n\nIs there a way to evaluate the performance of the resulting neuro-symbolic system qualitatively, to ensure it behaves in the manner that the algorithmic designer expects it does? Is there some kind of sanity check one could do to guard against the \"all the methods fail at the given task, but ours fails with slightly better numbers\" potential failure mode? While I certainly have no direct evidence that this is happening, it would make me feel much better about the paper if this could be definitively ruled out via the experiments.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698727033975}, {"id": "nyoIBm9kZP", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5927/Reviewer_vhBN"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper develops a general first-order gradient-based framework for end-to-end neural and symbolic parameter learning. \nThe framework is formulated as a bilevel optimization problem with a constrained lower-level inference problem. The inference problem minimizes a task-specific energy function, while the upper-level problem minimizes a mixed value- and minimizer-based objective. Using the value-function approach, the bilevel problem is reformulated as an inequality-constrained optimization problem, which in turn is relaxed by allowing finite violations of the constraint. Further, to deal with potential  non-differentiability of the energy function, a Moreau-envelope-based smoothening is applied. The final algorithm iteratively solves increasingly tight relaxations of the resulting smoothed optimization problem. Each relaxation is solved using a bound-constrained augmented Lagrangian algorithm.\n\nNext, the framework is applied to neural probabilistic soft logic, in which inference is formulated as MAP on a deep hinge-loss Markov random field. The inference problem is then reformulated as a regularized LCQP, for which several continuity properties are established. An efficiently parallelizable dual block coordinate descent algorithm is proposed for solving it.\nFinally, an empirical investigation shows that the proposed set of methods consistently improves inference and training runtime, while improving the accuracy.", "review_text": "This paper develops a general first-order gradient-based framework for end-to-end neural and symbolic parameter learning. \nThe framework is formulated as a bilevel optimization problem with a constrained lower-level inference problem. The inference problem minimizes a task-specific energy function, while the upper-level problem minimizes a mixed value- and minimizer-based objective. Using the value-function approach, the bilevel problem is reformulated as an inequality-constrained optimization problem, which in turn is relaxed by allowing finite violations of the constraint. Further, to deal with potential  non-differentiability of the energy function, a Moreau-envelope-based smoothening is applied. The final algorithm iteratively solves increasingly tight relaxations of the resulting smoothed optimization problem. Each relaxation is solved using a bound-constrained augmented Lagrangian algorithm.\n\nNext, the framework is applied to neural probabilistic soft logic, in which inference is formulated as MAP on a deep hinge-loss Markov random field. The inference problem is then reformulated as a regularized LCQP, for which several continuity properties are established. An efficiently parallelizable dual block coordinate descent algorithm is proposed for solving it.\nFinally, an empirical investigation shows that the proposed set of methods consistently improves inference and training runtime, while improving the accuracy.", "strengths": "- The reformulation of the NeSy learning problem into its relaxed and smoothened formulation (9) is quite powerful as it allows the application of first-order optimization techniques. The employed higher-level bound-constrained augmented Lagrangian method for solving it is known but applied in a new context.\n- The novel formulation of NeuPSL as an LCQP and the presented efficient and parallelizable dual block coordinate descent algorithm for solving it seems to be the key contribution. The shown continuity properties are a nice addition, the detailed proofs appear correct.\n- The experiments show impressive results in terms of runtime and accuracy in multiple settings. The experimental methodology is good, with detailed information on resources and hyperparameter optimization, with open-source code provided. \n- Overall, this paper makes multiple contributions that are empirically shown to significantly improve the runtime and accuracy of neuro-symbolic methods. The presented framework will probably be a starting point for various potential future applications.", "weaknesses": "My main critique of this paper is that the presentation is in parts not very clear, especially in Section 4. Notation is not always properly defined (e.g. the prox operator), dependencies are omitted without explicitly mentioning it (e.g. in the definition of the value function), or important parts are left out of the main text (e.g. how the the dual variables in equation 10 are updated or how the Moreau envelope of the energy function is computed). See questions for additional parts that were unclear to me.", "questions": "- If I understand correctly, the Algorithm 1 requires computing both the Value function and the Moreau envelope at each iteration. Both require minimization over $y$, so in the case of NeuPSL are these both computed using the BCD algorithm in section 5? I.e., is the $\\epsilon$ smoothening in equation 13 the same as the $\\frac{1}{\\rho}$ smoothening term in equation 8? In any case, I would recommend highlighting more explicitly how exactly Section 5 links into Section 4.\n- At the end of Section 5.1, the authors mention that mapping primal variables to dual variables requires calculating a pseudo-inverse of the matrix A. Is this at any point required in the proposed learning algorithm? If yes, is it described anywhere?\n- Are the authors aware of previous work using Moreau envelopes in the context of neuro-symbolic or structure learning with non-differentiable settings? Such links would be a useful addition to the related work. Otherwise, if this has not been used before, the novelty of this contribution could also be highlighted more.\n- One of the mentioned key contributions is that the \"dual BCD algorithm for NeuPSL inference [...] naturally produces statistics necessary for learning gradients\". I don't undertand this statement, how are the statistics in the BCD algorithm necessary for \"learning\" gradients? Do you mean they help in computing the gradients required for the NeSy learning algorithm?\n\nRemarks:\n- After equation 7: \"The formulation in (7) is referred to as a value-function approach in bilevel optimization literature.\" Citations would be great here.\n- In section 5.1, the vector $b$ seems to be an affine function in the neural predictions and symbolic inputs rather than linear.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper develops a general first-order gradient-based framework for end-to-end neural and symbolic parameter learning. \nThe framework is formulated as a bilevel optimization problem with a constrained lower-level inference problem. The inference problem minimizes a task-specific energy function, while the upper-level problem minimizes a mixed value- and minimizer-based objective. Using the value-function approach, the bilevel problem is reformulated as an inequality-constrained optimization problem, which in turn is relaxed by allowing finite violations of the constraint. Further, to deal with potential  non-differentiability of the energy function, a Moreau-envelope-based smoothening is applied. The final algorithm iteratively solves increasingly tight relaxations of the resulting smoothed optimization problem. Each relaxation is solved using a bound-constrained augmented Lagrangian algorithm.\n\nNext, the framework is applied to neural probabilistic soft logic, in which inference is formulated as MAP on a deep hinge-loss Markov random field. The inference problem is then reformulated as a regularized LCQP, for which several continuity properties are established. An efficiently parallelizable dual block coordinate descent algorithm is proposed for solving it.\nFinally, an empirical investigation shows that the proposed set of methods consistently improves inference and training runtime, while improving the accuracy.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- The reformulation of the NeSy learning problem into its relaxed and smoothened formulation (9) is quite powerful as it allows the application of first-order optimization techniques. The employed higher-level bound-constrained augmented Lagrangian method for solving it is known but applied in a new context.\n- The novel formulation of NeuPSL as an LCQP and the presented efficient and parallelizable dual block coordinate descent algorithm for solving it seems to be the key contribution. The shown continuity properties are a nice addition, the detailed proofs appear correct.\n- The experiments show impressive results in terms of runtime and accuracy in multiple settings. The experimental methodology is good, with detailed information on resources and hyperparameter optimization, with open-source code provided. \n- Overall, this paper makes multiple contributions that are empirically shown to significantly improve the runtime and accuracy of neuro-symbolic methods. The presented framework will probably be a starting point for various potential future applications.", "weaknesses": "My main critique of this paper is that the presentation is in parts not very clear, especially in Section 4. Notation is not always properly defined (e.g. the prox operator), dependencies are omitted without explicitly mentioning it (e.g. in the definition of the value function), or important parts are left out of the main text (e.g. how the the dual variables in equation 10 are updated or how the Moreau envelope of the energy function is computed). See questions for additional parts that were unclear to me.", "questions": "- If I understand correctly, the Algorithm 1 requires computing both the Value function and the Moreau envelope at each iteration. Both require minimization over $y$, so in the case of NeuPSL are these both computed using the BCD algorithm in section 5? I.e., is the $\\epsilon$ smoothening in equation 13 the same as the $\\frac{1}{\\rho}$ smoothening term in equation 8? In any case, I would recommend highlighting more explicitly how exactly Section 5 links into Section 4.\n- At the end of Section 5.1, the authors mention that mapping primal variables to dual variables requires calculating a pseudo-inverse of the matrix A. Is this at any point required in the proposed learning algorithm? If yes, is it described anywhere?\n- Are the authors aware of previous work using Moreau envelopes in the context of neuro-symbolic or structure learning with non-differentiable settings? Such links would be a useful addition to the related work. Otherwise, if this has not been used before, the novelty of this contribution could also be highlighted more.\n- One of the mentioned key contributions is that the \"dual BCD algorithm for NeuPSL inference [...] naturally produces statistics necessary for learning gradients\". I don't undertand this statement, how are the statistics in the BCD algorithm necessary for \"learning\" gradients? Do you mean they help in computing the gradients required for the NeSy learning algorithm?\n\nRemarks:\n- After equation 7: \"The formulation in (7) is referred to as a value-function approach in bilevel optimization literature.\" Citations would be great here.\n- In section 5.1, the vector $b$ seems to be an affine function in the neural predictions and symbolic inputs rather than linear.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698637597712}, {"id": "3eBsQnR96r", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5927/Reviewer_2ZRW"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work formulates neuro-symbolic learning as a bilevel optimization problem, using Moreau envelopes to smooth the NeSy energy function. This yields substantial runtime improvements over competing methods.", "review_text": "This work formulates neuro-symbolic learning as a bilevel optimization problem, using Moreau envelopes to smooth the NeSy energy function. This yields substantial runtime improvements over competing methods.", "strengths": "1. The paper is nicely structured and easy to follow.\n\n2. The approach is well-motivated and yields substantial runtime benefits.\n\n3. The considered problem is important to the machine learning field.", "weaknesses": "1. Neither variant of the propsed method (CC / LF) consistently beats the ADMM baseline across all datasets. However, it is significantly faster in most cases.\n\n2. Only one baseline was considered. As neural-symbolic systems aren't my field I'll defer to other reviewers on whether this is adequate.", "questions": "1. In table 2, the LF D-BCD method performs poorly for the MNIST addition tasks. The authors note that this is due to the \"high number of tightly connected components.\" I didn't quite understand this explanation.\n\n2. Where does the value function introduced in (7) come from? Is it learned from data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work formulates neuro-symbolic learning as a bilevel optimization problem, using Moreau envelopes to smooth the NeSy energy function. This yields substantial runtime improvements over competing methods.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper is nicely structured and easy to follow.\n\n2. The approach is well-motivated and yields substantial runtime benefits.\n\n3. The considered problem is important to the machine learning field.", "weaknesses": "1. Neither variant of the propsed method (CC / LF) consistently beats the ADMM baseline across all datasets. However, it is significantly faster in most cases.\n\n2. Only one baseline was considered. As neural-symbolic systems aren't my field I'll defer to other reviewers on whether this is adequate.", "questions": "1. In table 2, the LF D-BCD method performs poorly for the MNIST addition tasks. The authors note that this is due to the \"high number of tightly connected components.\" I didn't quite understand this explanation.\n\n2. Where does the value function introduced in (7) come from? Is it learned from data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697735990298}], "openreview_url": "https://openreview.net/forum?id=uJPWeZffgl", "arxiv_id": "2401.09651", "paper_pdf": "papers/uJPWeZffgl.pdf", "paper_pdf_sha256": "0f8cea504930db62a2f94cd7d27380b8c58edd68dba8d58ee4171fa90ea4790c", "paper_pdf_bytes": 651254, "paper_pdf_source": "openreview", "code_url": "https://github.com/convexbilevelnesylearning/experimentscripts", "code_repository": "convexbilevelnesylearning/experimentscripts", "code_commit": "8d01210737c69c00ef1f867ead21dc22540a6244", "code_archive": "repos/uJPWeZffgl.zip", "code_archive_sha256": "6a641e389fee9fb08e464074fb1d0e105cb3142236d9bb5b1a862ee93ab7d05a", "code_archive_bytes": 83587, "code_file_count": 29, "code_extensions": {".py": 25, ".sh": 4}, "github_disk_usage_kb": 43, "github_languages": {"Python": 302497, "Shell": 13191}, "github_archived": false, "github_pushed_at": "2023-09-28T16:29:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/convex-and-bilevel-optimization-for-neuro"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SZYXyhE2c6f", "year": 2023, "status": "rejected", "title": "A Probabilistic Framework For Modular Continual Learning", "authors": ["Lazar Valkov", "Akash Srivastava", "Dipak Chaudhari", "Swarat Chaudhuri", "Charles Sutton"], "authorids": ["~Lazar_Valkov1", "~Akash_Srivastava1", "~Dipak_Chaudhari1", "~Swarat_Chaudhuri1", "~Charles_Sutton1"], "authors_source": "OpenReview API", "abstract": "Continual learning (CL) algorithms seek to accumulate and transfer knowledge across a sequence of tasks and achieve better performance on each successive task. Modular approaches, which use a different composition of modules for each task and avoid forgetting by design, have been shown to be a promising direction to CL. However, searching through the large space of possible module compositions remains a challenge. In this work, we develop a scalable probabilistic search framework as a solution to this challenge. Our framework has two distinct components. The first is designed to transfer knowledge across similar input domains. To this end, it models each module’s training input distribution and uses a Bayesian model to find the most promising module compositions for a new task. The second component targets transfer across tasks with disparate input distributions or different input spaces and uses Bayesian optimisation to explore the space of module compositions. We show that these two methods can be easily combined and evaluate the resulting approach on two benchmark suites designed to capture different desiderata of CL techniques. The experiments show that our framework offers superior performance compared to state-of-the-art CL baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Rdt30v_NJk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5838/Reviewer_E5fq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper introduces a probabilistic framework for modular continual learning.\nThey built on the idea in [1] that modules can be chosen locally based on their input distribution.\nContrarily to [1], this paper opts for a Bayesian treatment of the module composition problem.\n\nThe paper proposes two transfer learning settings for compositionality: perceptual vs non-perceptual transfer.\nIn the former, we assume that most of the drift in p(x,y) happens in p(x), whereas the in the latter, it occurs in p(y|x).\nThey prescribe one probabilistic method for both settings.\n\nSome experiments in CTRL [2] and a modified version show that the method can outperform some baselines.\n\n[1] Continual Learning via Local Module Composition. Oleksiy Ostapenko, Pau Rodríguez, Massimo Caccia, Laurent Charlin\n[2] Efficient Continual Learning with Modular Networks and Task-Driven Priors. Tom Veniat, Ludovic Denoyer, Marc'Aurelio Ranzato", "review_text": "I think the paper is worthy of a publication if the weakness mentioned above are addressed. ", "strengths": "**Strengths**\n\nThe proposed method is sensible and adequately scales w.r.t. the number of training modules, which was a limitation of [1].\nThis work can inspire future work in continual learning.\n\n**Weaknesses**\n\nI wish the paper offered a more profound contrast with [1], both technically and empirically.\n\nThe idea of using the modules' input distribution as a signal for activating the modules was proposed in [1], but section 4 does not acknowledge that. Section 4 should explain how previous work has solved the problem above and then explain how their Bayesian treatment relates. The Section should further explain the pros and cons of such treatment compared to the previous solution.\n\nThe paper could instead (or additionally) benchmark their method against [1], which is the most relevant baseline for the paper.\nI understand that running new experiments during rebuttal is not ideal. \nI'm not expecting the authors to do that.\nAt least the others might want to add the reported results of [1] for the CTRL benchmark while mentioning that [1] operates in a much more challenging setting, i.e., **class**-incremental learning and not task-incremental.\n\n  ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a probabilistic framework for modular continual learning.\nThey built on the idea in [1] that modules can be chosen locally based on their input distribution.\nContrarily to [1], this paper opts for a Bayesian treatment of the module composition problem.\n\nThe paper proposes two transfer learning settings for compositionality: perceptual vs non-perceptual transfer.\nIn the former, we assume that most of the drift in p(x,y) happens in p(x), whereas the in the latter, it occurs in p(y|x).\nThey prescribe one probabilistic method for both settings.\n\nSome experiments in CTRL [2] and a modified version show that the method can outperform some baselines.\n\n[1] Continual Learning via Local Module Composition. Oleksiy Ostapenko, Pau Rodríguez, Massimo Caccia, Laurent Charlin\n[2] Efficient Continual Learning with Modular Networks and Task-Driven Priors. Tom Veniat, Ludovic Denoyer, Marc'Aurelio Ranzato", "strength_and_weaknesses": "**Strengths**\n\nThe proposed method is sensible and adequately scales w.r.t. the number of training modules, which was a limitation of [1].\nThis work can inspire future work in continual learning.\n\n**Weaknesses**\n\nI wish the paper offered a more profound contrast with [1], both technically and empirically.\n\nThe idea of using the modules' input distribution as a signal for activating the modules was proposed in [1], but section 4 does not acknowledge that. Section 4 should explain how previous work has solved the problem above and then explain how their Bayesian treatment relates. The Section should further explain the pros and cons of such treatment compared to the previous solution.\n\nThe paper could instead (or additionally) benchmark their method against [1], which is the most relevant baseline for the paper.\nI understand that running new experiments during rebuttal is not ideal. \nI'm not expecting the authors to do that.\nAt least the others might want to add the reported results of [1] for the CTRL benchmark while mentioning that [1] operates in a much more challenging setting, i.e., **class**-incremental learning and not task-incremental.\n\n  ", "clarity,_quality,_novelty_and_reproducibility": "great on all accounts.", "summary_of_the_review": "I think the paper is worthy of a publication if the weakness mentioned above are addressed. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666780417301}, {"id": "NM-8OR716yz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5838/Reviewer_ARaq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper tackles the idea of doing inference over a set of modules such that the resulting model can generalize as well as adapt quickly. Different approaches differ as to how they do search over an ensemble of modules. The proposed paper deals with both \"perceptual\" transfer and \"non-perceptual\" transfer. They introduce a probabilistic framework for selection of pre-trained modules. The key idea for driving the selection (i.e., which module to select) is  how well the module can transform the input. Each module is parameterized as a neural network. The paper define a posterior and prior, where the posterior is a function of how well the module can explain the input, and prior gives preference to modules which help achieve a higher accuracy. ", "review_text": "The paper proposes a probabilistic framework to do inference over an ensemble of modules.", "strengths": "Strengths:\n\n-  The paper tackles an important problem i.e., exploring better ways of searching through possible combinations of modules.\n\nWeaknesses:\n\n- It would be useful to study how the proposed method scales as a function of number of modules.\n- It would also be useful how the proposed method compares to methods which does a \"soft\" selection of the output of modules (like in LMC). \n- The paper seems a bit hard to parse. It may be useful to give a high level \"overview\"of the proposed method before jumping to describe how the paper works. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper tackles the idea of doing inference over a set of modules such that the resulting model can generalize as well as adapt quickly. Different approaches differ as to how they do search over an ensemble of modules. The proposed paper deals with both \"perceptual\" transfer and \"non-perceptual\" transfer. They introduce a probabilistic framework for selection of pre-trained modules. The key idea for driving the selection (i.e., which module to select) is  how well the module can transform the input. Each module is parameterized as a neural network. The paper define a posterior and prior, where the posterior is a function of how well the module can explain the input, and prior gives preference to modules which help achieve a higher accuracy. ", "strength_and_weaknesses": "Strengths:\n\n-  The paper tackles an important problem i.e., exploring better ways of searching through possible combinations of modules.\n\nWeaknesses:\n\n- It would be useful to study how the proposed method scales as a function of number of modules.\n- It would also be useful how the proposed method compares to methods which does a \"soft\" selection of the output of modules (like in LMC). \n- The paper seems a bit hard to parse. It may be useful to give a high level \"overview\"of the proposed method before jumping to describe how the paper works. \n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: Different sections are well written.\n\nQuality: The paper tackles an important problem. \n\nNovelty: The underlying idea seems interesting, and to the best of my knowledge has not been explored in this context before. \n\n", "summary_of_the_review": "The paper proposes a probabilistic framework to do inference over an ensemble of modules.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666719738451}, {"id": "SlQl73QV6lp", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5838/Reviewer_LPZE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper extends prior work on Continual Learning in two directions: 1) It introduces new learning sequences designed to evaluate non-perceptual transfer (transfer the last l layers) and few-shot transfer (transfer all layers) as well as more challenging sequences for the evaluation of perpetual transfer (transfer the first l layers). 2) Building on modular approaches to CL, e.g., (Venital et al. 2020), a probabilistic framework is proposed to tackle the exponential space of module combinations. One probabilistic model is proposed for perceptual (and few-shot) transfer and a separate model is proposed for non-perceptual transfer.\n\nExperiments compare the proposed approach to some recent methods and the results are somewhat favorable to the proposed method.\n", "review_text": "The paper’s contribution is limited as outlined above and the quality of the presentation is very poor.", "strengths": "## Strengths\n\n**S1.** The paper identifies limitations of prior work and seeks to address them.\n\n## Weaknesses\n\n**W1.** The contributions of the submission are very incremental. I see it as an extension of the work of (Venital et al. 2020) in the two directions outlined in the summary. Further, the theoretical contributions are obfuscated by the rough presentation (more on this below) and the validation is, in part, designed for the proposed method (i.e., some of the evaluation sequences are designed to not be handled properly by prior work).\n\n**W2.** The presentation has many issues:\n\n1. often symbols are introduced without definition and sometimes never get defined, e.g., Z in eq. (10), d in eq. (10) does not match the d previously defined; A at the end of page 4;\n\n2. evaluation metrics like those in Table 1 are never defined; etc.\n\n3. some explanations do not make much sense, e.g., “gives preference to modules which helped achieve a higher accuracy, on the problems which they have been trained on,” “minimum number of transfer which need to be transferred,” “can be explored by applying this search strategy to each sequentially,” etc.\n\n4. the probabilistic models proposed are not explained in detail, e.g., eq. (5) is simply stated with no explanation and the Gaussian Process in section 5.1 is not even specified.\n\n**W3.** The paper disregards other benchmarks in CL like (Lin et al. NeurIPS 2021.) Such a benchmark should be discussed, and results on it should either be included or there should be a discussion explaining why those would not be relevant.\n\n## References\n\nLin et al. The CLEAR Benchmark: Continual LEArning on Real-World Imagery. NeurIPS 2021.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper extends prior work on Continual Learning in two directions: 1) It introduces new learning sequences designed to evaluate non-perceptual transfer (transfer the last l layers) and few-shot transfer (transfer all layers) as well as more challenging sequences for the evaluation of perpetual transfer (transfer the first l layers). 2) Building on modular approaches to CL, e.g., (Venital et al. 2020), a probabilistic framework is proposed to tackle the exponential space of module combinations. One probabilistic model is proposed for perceptual (and few-shot) transfer and a separate model is proposed for non-perceptual transfer.\n\nExperiments compare the proposed approach to some recent methods and the results are somewhat favorable to the proposed method.\n", "strength_and_weaknesses": "## Strengths\n\n**S1.** The paper identifies limitations of prior work and seeks to address them.\n\n## Weaknesses\n\n**W1.** The contributions of the submission are very incremental. I see it as an extension of the work of (Venital et al. 2020) in the two directions outlined in the summary. Further, the theoretical contributions are obfuscated by the rough presentation (more on this below) and the validation is, in part, designed for the proposed method (i.e., some of the evaluation sequences are designed to not be handled properly by prior work).\n\n**W2.** The presentation has many issues:\n\n1. often symbols are introduced without definition and sometimes never get defined, e.g., Z in eq. (10), d in eq. (10) does not match the d previously defined; A at the end of page 4;\n\n2. evaluation metrics like those in Table 1 are never defined; etc.\n\n3. some explanations do not make much sense, e.g., “gives preference to modules which helped achieve a higher accuracy, on the problems which they have been trained on,” “minimum number of transfer which need to be transferred,” “can be explored by applying this search strategy to each sequentially,” etc.\n\n4. the probabilistic models proposed are not explained in detail, e.g., eq. (5) is simply stated with no explanation and the Gaussian Process in section 5.1 is not even specified.\n\n**W3.** The paper disregards other benchmarks in CL like (Lin et al. NeurIPS 2021.) Such a benchmark should be discussed, and results on it should either be included or there should be a discussion explaining why those would not be relevant.\n\n## References\n\nLin et al. The CLEAR Benchmark: Continual LEArning on Real-World Imagery. NeurIPS 2021.\n", "clarity,_quality,_novelty_and_reproducibility": "In my view this work has potential but is work in progress, and is not quite ready for publication.", "summary_of_the_review": "The paper’s contribution is limited as outlined above and the quality of the presentation is very poor.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666660316588}, {"id": "SrlVtkg-pO9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5838/Reviewer_ifu1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a probabilistic framework to model modular architectures for the problem of continual learning (CL). They divide the problem of choosing modules in two parts: perceptual transfer (PT), and non-perceptual transfer (NT). In PT, the first l layers are assumed to be pre-trained and the model must choose between those modules to maximize performance when training new modules on top. NT works the other way around. From this, they derive 3 CL algorithms: (i) PeCL, which does perceptual and few-shot transfer, (ii) NoCL, which does non-perceptual transfer, and (iii) CCL, which combines both strategies. To evaluate their model, the authors introduce a compositional version of CTrL and evaluate plasticity, backward transfer, forward transfer, perceptual transfer, few-shot transfer, and non-perceptual transfer. In experiments they show that their models perform favourably compared to SA, MCL-RS, HOUDINI, MNTDP.", "review_text": "The proposed method is sound and it tackles an interesting problem for continual learning. However, the lack of clarity of the text makes it difficult to properly asses the quality of this work. In its current form, this work does not meet the standards of ICLR.", "strengths": "Strengths\n=======\n* The method is sound and it tries to approach modularity from first principles with a divide-and-conquer approach\n* The proposed method outperforms MNTDP\n\nWeaknesses\n=========\n* The model description is not clear, which makes it difficult to understand. Overall the submission looks unpolished. I highly encourage the authors to carefuly rework sections 3, 4, and 5, to make it clearer.\n* The authors cite LMC but do not compare with it, is there any reason?\n* The divide-and-conquer approach has a clear disadvantage: it introduces the additional complexity of establishing where to split the preceptual part from the non-perceptual part. Maybe the authors could include a limitations section to talk about that.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors propose a probabilistic framework to model modular architectures for the problem of continual learning (CL). They divide the problem of choosing modules in two parts: perceptual transfer (PT), and non-perceptual transfer (NT). In PT, the first l layers are assumed to be pre-trained and the model must choose between those modules to maximize performance when training new modules on top. NT works the other way around. From this, they derive 3 CL algorithms: (i) PeCL, which does perceptual and few-shot transfer, (ii) NoCL, which does non-perceptual transfer, and (iii) CCL, which combines both strategies. To evaluate their model, the authors introduce a compositional version of CTrL and evaluate plasticity, backward transfer, forward transfer, perceptual transfer, few-shot transfer, and non-perceptual transfer. In experiments they show that their models perform favourably compared to SA, MCL-RS, HOUDINI, MNTDP.", "strength_and_weaknesses": "Strengths\n=======\n* The method is sound and it tries to approach modularity from first principles with a divide-and-conquer approach\n* The proposed method outperforms MNTDP\n\nWeaknesses\n=========\n* The model description is not clear, which makes it difficult to understand. Overall the submission looks unpolished. I highly encourage the authors to carefuly rework sections 3, 4, and 5, to make it clearer.\n* The authors cite LMC but do not compare with it, is there any reason?\n* The divide-and-conquer approach has a clear disadvantage: it introduces the additional complexity of establishing where to split the preceptual part from the non-perceptual part. Maybe the authors could include a limitations section to talk about that.", "clarity,_quality,_novelty_and_reproducibility": "Clarity\n=====\nI found this work difficult to read and understand. Here are some unclear points I found while reviewing the submission:\n\n* To my understanding, only one module is chosen per layer. Then, at the beginning of Section 3, you mention that you split $\\Pi_t$ in subsets $\\Pi_t^i$. Does that mean that you choose modules layerwise? If so, why do you call $\\Pi_t^i$ a subset if it is a single module? Maybe I am missing something.\n* In equation 3, what is the j subscript after the parenthesis?\n* In equation 6, q is introduced without any definition. I could get it from the context but a more explicit definition would help. The same with $\\mathbf{A}$\n* In section 4.2: \"we freeze this selection and reuse the same l modules for PT paths which transfer more modules\" what do you mean?\n* \"We define the augmented search strategy\" why augmented?\n* It took some time for me to parse equation 7, I think some explanation or guidance to the reader (particularly about $\\pi^{*PT,l-1}[l-1]$) could make it much easier to understand.\n* \"The difference between two NT paths ... of length l is that their last l pre-trained modules compute different functions.\" If they are of length l, the last l pre-trained modules is the whole path no?\n* I understand equation 10 but how it is used in your algorithm is not clear to me.\n* \"number of transfer which need to be transferred in order to improve the performance.\" what do you mean by the transfer that needs to be transferred?\n* Eq 11. What is the $j$ after the parenthesis?\n* You introduce the term LCB but I could not find an explanation for what it is. The search brought me to the appendix, where I could find what the letters mean as well as many \"??\" for the references.\n\nBesides these points there are some minor typos like \"as it does not does not involve\" in Section 3.\n\nQuality\n=====\n* For the most part, the technical quality is good, however, the lack of clarity hinders my understanding of the method. In experiments,  it is not clear why the authors did not compare to LMC.\n* It is not clear whether the number of parameters of their method to put them in comparison to the other baselines (if it is the same I did not find it clearly stated in the text, maybe appendix C.2 last paragraph?)\n\nNovelty\n======\nThe proposed method is novel to the best of my knowledge.\n\nReproducibility\n===========\nThe authors did not provide the code neither a reproducibility statement, however, the appendix contain some implementation details.", "summary_of_the_review": "The proposed method is sound and it tackles an interesting problem for continual learning. However, the lack of clarity of the text makes it difficult to properly asses the quality of this work. In its current form, this work does not meet the standards of ICLR.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666643549523}], "openreview_url": "https://openreview.net/forum?id=SZYXyhE2c6f", "arxiv_id": "2306.06545", "paper_pdf": "papers/SZYXyhE2c6f.pdf", "paper_pdf_sha256": "2c0dbf34d6206d73e553d655ae2956328acf4aa5aa4c0120afbd97672444ea17", "paper_pdf_bytes": 1316703, "paper_pdf_source": "openreview", "code_url": "https://github.com/LazarValkov/PICLE", "code_repository": "LazarValkov/PICLE", "code_commit": "80b262825d39a8a54d8bc82bda3b89d8c0742eb2", "code_archive": "repos/SZYXyhE2c6f.zip", "code_archive_sha256": "2cd0485d172103c103bfc4cfa0f491b658b2ae3e09decf40067925e9c31ba7d1", "code_archive_bytes": 74043, "code_file_count": 44, "code_extensions": {".py": 44}, "github_disk_usage_kb": 62, "github_languages": {"Python": 253415}, "github_archived": false, "github_pushed_at": "2024-05-05T06:08:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-probabilistic-framework-for-modular"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LLHwQh9zEb", "year": 2022, "status": "rejected", "title": "Permutation invariant graph-to-sequence model for template-free retrosynthesis and reaction prediction", "authors": ["Zhengkai Tu", "Connor W. Coley"], "authorids": ["~Zhengkai_Tu1", "~Connor_W._Coley1"], "authors_source": "OpenReview API", "abstract": "Synthesis planning and reaction outcome prediction are two fundamental problems in computer-aided organic chemistry for which a variety of data-driven approaches have emerged. Natural language approaches that model each problem as a SMILES-to-SMILES translation lead to a simple end-to-end formulation, reduce the need for data preprocessing, and enable the use of well-optimized machine translation model architectures. However, SMILES representations are not an efficient representation for capturing information about molecular structure, as evidenced by the success of SMILES augmentation to boost empirical performance. Here, we describe a novel Graph2SMILES model that combines the power of Transformer models for text generation with the permutation invariance of molecular graph encoders. As an end-to-end architecture, Graph2SMILES can be used as a drop-in replacement for the Transformer in any task involving molecule(s)-to-molecule(s) transformations. In our encoder, an attention-augmented directed message passing neural network (D-MPNN) captures local chemical environments, and the global attention encoder allows for long-range and intermolecular interactions, enhanced by graph-aware positional embedding.  Graph2SMILES improves the top-1 accuracy of the Transformer baselines by $1.7\\%$ and $1.9\\%$ for reaction outcome prediction on USPTO_480k and USPTO_STEREO datasets respectively, and by $9.8\\%$ for one-step retrosynthesis on the USPTO_50k dataset.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "hNcuqExBE4k", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1963/Reviewer_jeTQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed a new method for retrosynthesis, which does not require the mapping numbers and extracting templates from the literature.\n\nBasically, the model consists of a graph-based encoder and a sequence based encoder. The encoder consists of local aggregation from neighbors and global attention using a new positional method. The decoder is a Transformer model with relative positional encoding.\n\nThe method achieved promising results on several retrosynthesis datasets.\n", "review_text": "1.\tFrom Eqn.(1) to Eqn.(5), you choose to use a complex gating mechanism to aggregate information. Is every component necessary? What if using a simple GCN or GAT?\n2.\tI think the model contains more parameters than conventional retrosynthesis models like conventional Transformer, GLN. Could you please show compare the number of parameters of different methods?\n3.\tThe authors should provide some real cases to show how the method outperforms previous baselines, and why the method can obtain good results without templates.\n\n\nMissing References:\n\n1.\tDual-view Molecule Pre-training, https://arxiv.org/abs/2106.10234, the authors also work on retrosynthesis using Transformer and GNN models. A comparison is necessary.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors proposed a new method for retrosynthesis, which does not require the mapping numbers and extracting templates from the literature.\n\nBasically, the model consists of a graph-based encoder and a sequence based encoder. The encoder consists of local aggregation from neighbors and global attention using a new positional method. The decoder is a Transformer model with relative positional encoding.\n\nThe method achieved promising results on several retrosynthesis datasets.\n", "main_review": "1.\tFrom Eqn.(1) to Eqn.(5), you choose to use a complex gating mechanism to aggregate information. Is every component necessary? What if using a simple GCN or GAT?\n2.\tI think the model contains more parameters than conventional retrosynthesis models like conventional Transformer, GLN. Could you please show compare the number of parameters of different methods?\n3.\tThe authors should provide some real cases to show how the method outperforms previous baselines, and why the method can obtain good results without templates.\n\n\nMissing References:\n\n1.\tDual-view Molecule Pre-training, https://arxiv.org/abs/2106.10234, the authors also work on retrosynthesis using Transformer and GNN models. A comparison is necessary.\n", "summary_of_the_review": "The results in this paper are good, although the method itself is not quite novel.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635843365817}, {"id": "yD3y7HieAoL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1963/Reviewer_3c52"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a graph-to-sequence architecture called Graph2SMILES for the retrosynthesis and the reaction outcome prediction. Graph2SMILES uses an attention-augmented D-MPNN encoder to capture the local information and a global attention encoder with graph-aware positional embeddings to capture the global information. Experiments show that Graph2SMILES is competitive with Transformer baselines but does not outperform state-of-the-art methods on tasks of the one-step retrosynthesis and the reaction outcome prediction. ", "review_text": "The main strengths of this paper are as follows.\n1.\tThis paper proposes Graph2SMILES, which is a graph-to-sequence architecture without using sequence representations of input SMILES. Therefore, Graph2SMILES is permutation invariant to the input and does not need the data augmentation.\n2.\tGraph2SMILES has a wide range of potential applications because it can serve as a drop-in replacement for Transformer in many tasks involving the molecule(s)-to-molecule(s) transformation.\n\nMy major concerns are as follows.\n1.\tThis paper states that Graph2SMILES achieves state-of-the-art top-1 accuracy on common benchmarks among methods that do not use reaction templates, atom mapping, pretraining, or data augmentation strategies. The authors claim that integrating the above features or techniques with Graph2SMILES could improve the performance. However, they do not conduct experiments to demonstrate their claim. Besides, as the aforementioned techniques are commonly seen in predictive chemistry tasks, the authors may want to explain why they do not equip Graph2SMILES with these techniques.\n2.\tD-GAT is a variant of D-GCN with attention-based message updates. However, D-GAT does not outperform D-GCN in terms of the top-1 accuracy, which is the basis for comparison throughout the discussion in this paper. According to Table 1, D-GAT has a small advantage over D-GCN only in terms of the top-5 and top-10 accuracies in the reaction outcome prediction.\n3.\tGraph2SMILES involves calculating pairwise shortest path lengths between atoms, which can be computationally prohibitive. The authors may want to compare Graph2SMILES against baselines in terms of the computational complexity.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a graph-to-sequence architecture called Graph2SMILES for the retrosynthesis and the reaction outcome prediction. Graph2SMILES uses an attention-augmented D-MPNN encoder to capture the local information and a global attention encoder with graph-aware positional embeddings to capture the global information. Experiments show that Graph2SMILES is competitive with Transformer baselines but does not outperform state-of-the-art methods on tasks of the one-step retrosynthesis and the reaction outcome prediction. ", "main_review": "The main strengths of this paper are as follows.\n1.\tThis paper proposes Graph2SMILES, which is a graph-to-sequence architecture without using sequence representations of input SMILES. Therefore, Graph2SMILES is permutation invariant to the input and does not need the data augmentation.\n2.\tGraph2SMILES has a wide range of potential applications because it can serve as a drop-in replacement for Transformer in many tasks involving the molecule(s)-to-molecule(s) transformation.\n\nMy major concerns are as follows.\n1.\tThis paper states that Graph2SMILES achieves state-of-the-art top-1 accuracy on common benchmarks among methods that do not use reaction templates, atom mapping, pretraining, or data augmentation strategies. The authors claim that integrating the above features or techniques with Graph2SMILES could improve the performance. However, they do not conduct experiments to demonstrate their claim. Besides, as the aforementioned techniques are commonly seen in predictive chemistry tasks, the authors may want to explain why they do not equip Graph2SMILES with these techniques.\n2.\tD-GAT is a variant of D-GCN with attention-based message updates. However, D-GAT does not outperform D-GCN in terms of the top-1 accuracy, which is the basis for comparison throughout the discussion in this paper. According to Table 1, D-GAT has a small advantage over D-GCN only in terms of the top-5 and top-10 accuracies in the reaction outcome prediction.\n3.\tGraph2SMILES involves calculating pairwise shortest path lengths between atoms, which can be computationally prohibitive. The authors may want to compare Graph2SMILES against baselines in terms of the computational complexity.\n\n", "summary_of_the_review": "This paper studies two important problems in the computer-aided organic chemistry and proposes a graph-to-sequence architecture called Graph2SMILES. However, the empirical results do not show a superior performance of Graph2SMILES to existing methods, and the technical contribution is incremental.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635818311801}, {"id": "AXWkgle9wkC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1963/Reviewer_6G2y"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a graph-to-SMILES framework, which incorporates several recently developed engineering techniques from the community, for synthesis planning and reaction outcome prediction tasks. The proposed method leverages graph neural networks and Transformer attention model to encode the graph inputs and then utilizes a Transformer decoder to generate the SMILES string as outputs. Experiments on benchmark retrosynthesis and reaction prediction tasks show that the proposed approach outperformed the vanilla SMILES-to-SMILES transformer baseline, but obtained inferior results than some other advanced methods. The paper is interesting, but both the technical novelty and the experimental studies are weak to me.", "review_text": "The proposed framework integrates several recently developed engineering techniques and empirically shows its superior performance over vanilla SMILES-to-SMILES transformer baseline. This paper provides another comparison baseline for research on retrosynthesis and reaction prediction. Nevertheless, I have the following concerns regarding the paper.\n\n1.\tThe proposed framework is similar to the NERF approach (Bi et al. ICML 2021) as cited by the authors. NERF formulates the reaction prediction problem as a graph-to-graph translation problem. Also, NERF first leverages graph neural networks to capture the local information in individual molecules and then utilizes a Transformer encoder to further models the intermolecular interactions between nodes from multiple molecules. Furthermore, NERF uses a Transformer decoder to decode the output as graph. These are almost the same as that in the method proposed in the paper. The only different to me is that the NERF uses a Transformer to decode the output into graph directly (in a non-autoregressive fashion), while the proposed method here uses the Transformer to decode the output into SMILES strings (in an autoregressive fashion). In this sense, the novelty of this paper is limited to me. Note: I think NERF can naturally apply to two or more molecules since the Transformer encoder is used by considering all node embeddings from multiple molecules as a node set. \n2.\tExperimentally, the proposed method is not directly compared with NERF. I think such comparison is necessary since as shown in the paper, NERF outperforms the SMILES-to-SMILES transformer baseline, and even the augmented version of it, to which the proposed method here obtained inferior performance. The two methods are similar and closely related. I would expect the paper to include NERF into the main results in Table 1. Also, for the USPTO_STEREO_mixed task, I wonder what the reason was for only comparing with the vanilla Transformer, and why not comparing with the Augmented Transformer or the state-of-the-art method Chemformer?\n3.\tResults in Table 1 show that, the proposed method is inferior to the Transformer baseline with simple augmentation. Also, 1% less than the tested method Chemformer, which makes the paper’s contribution less significant to me.\n4.\tThe claim in the Abstract “molecular graph encoders that mitigates the need for input data augmentation” is a strong claim to me. Nevertheless, there is no evidence to support that claim. The slightly better performance over the SMILES-to-SMILES transformer baseline is not a convincing evidence to me. Input data augmentation may play a significant role on regularizing the deep neural networks. I think better justification to support the claim is necessary. \n5.\tThe statement in the last sentence of the first paragraph on Page2:  “[SMILES augmentation]…be interpreted as evidence of the ineffectiveness of the SMILES representation itself.” I think this hypothesis may need better support and analysis. To me, the augmentation of SMILES strings can act as a model regularization method, which helps the trained model to generalize well to unseen data, and may not directly infer the ineffectiveness of the SMILES representation itself. \n6.\tI am not fully understand the claim in the second paragraph of Page2 “… we guarantee the permutation invariance of Graph2SMILES to the input, eliminating the need for input-side augmentation altogether.” I think it would be useful to specify how and why so.\n7.\tThe proposed method integrates several performance engineering techniques (such as attention weights and multi-headed attention in the graph encoder, integration of shortest path length in the positional embeddings etc.), so where the improvement is really coming from is not clear to me. In the ablation study in Table4, both the positional embedding and global attention are key to the Transformer’s performance, so the performance degradation is expected when remove them: Transformer expects a positional embedding to work, and without a global attention the encoder will not be able to capture information from multiple molecules (their graphs are disconnected). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a graph-to-SMILES framework, which incorporates several recently developed engineering techniques from the community, for synthesis planning and reaction outcome prediction tasks. The proposed method leverages graph neural networks and Transformer attention model to encode the graph inputs and then utilizes a Transformer decoder to generate the SMILES string as outputs. Experiments on benchmark retrosynthesis and reaction prediction tasks show that the proposed approach outperformed the vanilla SMILES-to-SMILES transformer baseline, but obtained inferior results than some other advanced methods. The paper is interesting, but both the technical novelty and the experimental studies are weak to me.", "main_review": "The proposed framework integrates several recently developed engineering techniques and empirically shows its superior performance over vanilla SMILES-to-SMILES transformer baseline. This paper provides another comparison baseline for research on retrosynthesis and reaction prediction. Nevertheless, I have the following concerns regarding the paper.\n\n1.\tThe proposed framework is similar to the NERF approach (Bi et al. ICML 2021) as cited by the authors. NERF formulates the reaction prediction problem as a graph-to-graph translation problem. Also, NERF first leverages graph neural networks to capture the local information in individual molecules and then utilizes a Transformer encoder to further models the intermolecular interactions between nodes from multiple molecules. Furthermore, NERF uses a Transformer decoder to decode the output as graph. These are almost the same as that in the method proposed in the paper. The only different to me is that the NERF uses a Transformer to decode the output into graph directly (in a non-autoregressive fashion), while the proposed method here uses the Transformer to decode the output into SMILES strings (in an autoregressive fashion). In this sense, the novelty of this paper is limited to me. Note: I think NERF can naturally apply to two or more molecules since the Transformer encoder is used by considering all node embeddings from multiple molecules as a node set. \n2.\tExperimentally, the proposed method is not directly compared with NERF. I think such comparison is necessary since as shown in the paper, NERF outperforms the SMILES-to-SMILES transformer baseline, and even the augmented version of it, to which the proposed method here obtained inferior performance. The two methods are similar and closely related. I would expect the paper to include NERF into the main results in Table 1. Also, for the USPTO_STEREO_mixed task, I wonder what the reason was for only comparing with the vanilla Transformer, and why not comparing with the Augmented Transformer or the state-of-the-art method Chemformer?\n3.\tResults in Table 1 show that, the proposed method is inferior to the Transformer baseline with simple augmentation. Also, 1% less than the tested method Chemformer, which makes the paper’s contribution less significant to me.\n4.\tThe claim in the Abstract “molecular graph encoders that mitigates the need for input data augmentation” is a strong claim to me. Nevertheless, there is no evidence to support that claim. The slightly better performance over the SMILES-to-SMILES transformer baseline is not a convincing evidence to me. Input data augmentation may play a significant role on regularizing the deep neural networks. I think better justification to support the claim is necessary. \n5.\tThe statement in the last sentence of the first paragraph on Page2:  “[SMILES augmentation]…be interpreted as evidence of the ineffectiveness of the SMILES representation itself.” I think this hypothesis may need better support and analysis. To me, the augmentation of SMILES strings can act as a model regularization method, which helps the trained model to generalize well to unseen data, and may not directly infer the ineffectiveness of the SMILES representation itself. \n6.\tI am not fully understand the claim in the second paragraph of Page2 “… we guarantee the permutation invariance of Graph2SMILES to the input, eliminating the need for input-side augmentation altogether.” I think it would be useful to specify how and why so.\n7.\tThe proposed method integrates several performance engineering techniques (such as attention weights and multi-headed attention in the graph encoder, integration of shortest path length in the positional embeddings etc.), so where the improvement is really coming from is not clear to me. In the ablation study in Table4, both the positional embedding and global attention are key to the Transformer’s performance, so the performance degradation is expected when remove them: Transformer expects a positional embedding to work, and without a global attention the encoder will not be able to capture information from multiple molecules (their graphs are disconnected). \n", "summary_of_the_review": "The proposed method is similar to NERF as proposed by Bi et al., so the technical novelty is limited. Also, the experiment study missed important comparison baselines. Furthermore, a more comprehensive ablation study is needed since several engineering techniques are employed, and it is difficult to tell where the performance improvement is really coming from when compared to the Transformer vanilla model.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635777062894}, {"id": "NWgRYLmj6HE", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1963/Reviewer_Wxfa"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a GNN-based extension of transformers, which have been shown to be effective for reaction prediction etc. before. In particular, the GNN-based embedding of molecules in the reaction embeddings overcomes the artificial bias inherent in the often applied sequence embeddings. The experiments show that there are sometimes increases of performance in reaction and retrosynthesis prediction - and the approach could be applied to similar problems.\n\n\n", "review_text": "(+)\n\nUsing a sequence-independent encoding in the easy-to-use transformer makes sense and is a research question which is interesting for the - I think mostly, AI in chemistry - community.\n\n(-)\n\n- As far as I can see, the technical novelty is limited. The proposal is a combination of rather well-known methods.\n\n D-MPNN (where attention is added) and\n\nre-parameterized, relative positional encodings \n\n (using 0 to represent atoms in different molecules)\n\nin transformers\n\n- It is unclear if the proposed attention in the GNN is useful: Since we only see \"For reaction outcome prediction, there is a small advantage of using D-GAT over D-GCN\", and the ablation results are missing.\n\n- The results in Table 1 are only convincing for USPTO_STEREO_mixed.\n\n- The experimental comparison for retrosynthesis compares to methods which use different forms of pretraining or augmentation, which makes sense. It shows that the augmentation still provides advantages. And, as far as I understand, the graph-based molecule embedding entails that Smiles augmentation would not improve your results further? Also, it would make sense to include the ablation results for Graph2Smiles just using the transformer into Table 2 directly. \n\n\n-------------------------------\nOther Comments\n\n\n- It is unclear to me what \"Our hyperparameters for D-GAT and D-GCN are adapted from GraphRetro\" means.\n- In (8), s_uv should be AttnSum(...)\n- How exactly is \\mathcal{B}_u,v used in the learnable \\tilde{r}_u,v? \n- Table 4: \nSince the paper proposes the attention-based GNN, the ablation should be provided for that model.\n- Table 4: What is \"no global attention encoder\"? Just the combination of GNN embeddings w/o transformer? Since transformer is the baseline, I would not consider this as an ablation setting.\n- Table 4: How do the results look on the other tasks? For retrosynthesis, all open existing systems that yield full retrosynthesis trees which I know use top50 (or similar), so the fact that the base model is better at top10 (already) renders the analysis questionable.\n- \"We include part of our code to reproduce some specific results\" - Why not all?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a GNN-based extension of transformers, which have been shown to be effective for reaction prediction etc. before. In particular, the GNN-based embedding of molecules in the reaction embeddings overcomes the artificial bias inherent in the often applied sequence embeddings. The experiments show that there are sometimes increases of performance in reaction and retrosynthesis prediction - and the approach could be applied to similar problems.\n\n\n", "main_review": "(+)\n\nUsing a sequence-independent encoding in the easy-to-use transformer makes sense and is a research question which is interesting for the - I think mostly, AI in chemistry - community.\n\n(-)\n\n- As far as I can see, the technical novelty is limited. The proposal is a combination of rather well-known methods.\n\n D-MPNN (where attention is added) and\n\nre-parameterized, relative positional encodings \n\n (using 0 to represent atoms in different molecules)\n\nin transformers\n\n- It is unclear if the proposed attention in the GNN is useful: Since we only see \"For reaction outcome prediction, there is a small advantage of using D-GAT over D-GCN\", and the ablation results are missing.\n\n- The results in Table 1 are only convincing for USPTO_STEREO_mixed.\n\n- The experimental comparison for retrosynthesis compares to methods which use different forms of pretraining or augmentation, which makes sense. It shows that the augmentation still provides advantages. And, as far as I understand, the graph-based molecule embedding entails that Smiles augmentation would not improve your results further? Also, it would make sense to include the ablation results for Graph2Smiles just using the transformer into Table 2 directly. \n\n\n-------------------------------\nOther Comments\n\n\n- It is unclear to me what \"Our hyperparameters for D-GAT and D-GCN are adapted from GraphRetro\" means.\n- In (8), s_uv should be AttnSum(...)\n- How exactly is \\mathcal{B}_u,v used in the learnable \\tilde{r}_u,v? \n- Table 4: \nSince the paper proposes the attention-based GNN, the ablation should be provided for that model.\n- Table 4: What is \"no global attention encoder\"? Just the combination of GNN embeddings w/o transformer? Since transformer is the baseline, I would not consider this as an ablation setting.\n- Table 4: How do the results look on the other tasks? For retrosynthesis, all open existing systems that yield full retrosynthesis trees which I know use top50 (or similar), so the fact that the base model is better at top10 (already) renders the analysis questionable.\n- \"We include part of our code to reproduce some specific results\" - Why not all?\n", "summary_of_the_review": "Overall, I think the authors' proposed model for reaction prediction makes sense. However, as mentioned above, the paper's writing could be improved, the technical contribution is limited, and the experiments also show only limited improvements. Altogether, I therefore suggest to reject the paper at the current moment. I am happy to adjust my score in case I missed critical parts.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635772844333}], "openreview_url": "https://openreview.net/forum?id=LLHwQh9zEb", "arxiv_id": "2110.09681", "paper_pdf": "papers/LLHwQh9zEb.pdf", "paper_pdf_sha256": "c6823d4c31d1c81225617915ad117900ed90fa877beac11bd15595ce4093ddc1", "paper_pdf_bytes": 577501, "paper_pdf_source": "openreview", "code_url": "https://github.com/coleygroup/Graph2SMILES", "code_repository": "coleygroup/Graph2SMILES", "code_commit": "1b87f9ffdb739346b94853dfe8e5d63ac0198a38", "code_archive": "repos/LLHwQh9zEb.zip", "code_archive_sha256": "9b8495d262054c4af0f5bd697d6b2232642db0c4db32ece1f2b55eed8d34556a", "code_archive_bytes": 50516, "code_file_count": 23, "code_extensions": {".py": 18, ".sh": 5}, "github_disk_usage_kb": 67, "github_languages": {"Python": 139440, "Shell": 4780}, "github_archived": false, "github_pushed_at": "2023-11-07T03:01:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/permutation-invariant-graph-to-sequence-model-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "I-VfjSBzi36", "year": 2021, "status": "rejected", "title": "EarlyBERT: Efficient BERT Training via Early-bird Lottery Tickets", "authors": ["Xiaohan Chen", "Yu Cheng", "Shuohang Wang", "Zhe Gan", "Zhangyang Wang", "Jingjing Liu"], "authorids": ["~Xiaohan_Chen1", "~Yu_Cheng1", "~Shuohang_Wang1", "~Zhe_Gan1", "~Zhangyang_Wang1", "~Jingjing_Liu2"], "authors_source": "OpenReview API", "abstract": "Deep, heavily overparameterized language models such as BERT, XLNet and T5 have achieved impressive success in many NLP tasks. However, their high model complexity requires enormous computation resources and extremely long training time for both pre-training and fine-tuning. Many works have studied model compression on large NLP models, but only focus on reducing inference cost/time, while still requiring expensive training process. Other works use extremely large batch sizes to shorten the pre-training time at the expense of high demand for computation resources. In this paper, inspired by the Early-Bird Lottery Tickets studied for computer vision tasks, we propose EarlyBERT, a general computationally-efficient training algorithm applicable to both pre-training and fine-tuning of large-scale language models. We are the first to identify structured winning tickets in the early stage of BERT training, and use them for efficient training. Comprehensive pre-training and fine-tuning experiments on GLUE and SQuAD downstream tasks show that EarlyBERT easily achieves comparable performance to standard BERT with 35~45% less training time.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "DfnIh_mRH-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2348/AnonReviewer5"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe authors propose a technique for reducing the computational requirements of training BERT early in training to reduce the overall amount of resources required.\n\nPros:\n\nThe paper is well written and clear for the most part. The authors do thorough experimental evaluation.\n\nCons:\n\nI have two primary concerns about the paper and the proposed technique.\n1. The positioning of the technique is not entirely clear to me. The authors pitch it as a technique for reducing the training time of BERT and use LayerDrop as a baseline technique that also removes network components. However, it feels like another baseline that should be considered is neural architecture search, which also seeks to automatically find a more efficient model to train. The difference here is that the authors find the model early in the training run, but it seems like the EarlyBERT procedure could be run once and the resulting model architecture could be saved and re-trained like NAS models are. \n2. I found the experimental results to be lacking detail and breadth necessary to establish the value of the technique. Firstly, the rough time estimates in Table 2 are very odd given the primary value of the proposed technique is to reduce training time. The accuracy of EarlyBERT is close enough to LayerDrop that accurate training cost numbers are needed to differentiate between the techniques. Secondly, quoting training time reductions over the dense baseline when the EarlyBERT mode does not achieve the same accuracy makes the comparison very difficult to make. This problem shows up quite commonly in the model compression literature [1] and I’d encourage the authors to show full accuracy-training time tradeoff curves so that the training time savings for a given accuracy can be more clearly established. Lastly, I found the use of reduced training epochs in EarlyBERT to be odd because you do not evaluate whether or not this can be done for the baseline models and there isn’t clear evidence as to why your model would be able to do this while others (e.g., DropLayer) cannot. Figure 2 also does not seem to corroborate that higher learning rates can be used with shorter training time to achieve better accuracy. The data in your figure shows that the best learning rate achieves the best model quality independent of the number of training epochs.\n\nReferences:\n1. https://arxiv.org/abs/2003.03033\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear experimental results", "review": "Summary:\n\nThe authors propose a technique for reducing the computational requirements of training BERT early in training to reduce the overall amount of resources required.\n\nPros:\n\nThe paper is well written and clear for the most part. The authors do thorough experimental evaluation.\n\nCons:\n\nI have two primary concerns about the paper and the proposed technique.\n1. The positioning of the technique is not entirely clear to me. The authors pitch it as a technique for reducing the training time of BERT and use LayerDrop as a baseline technique that also removes network components. However, it feels like another baseline that should be considered is neural architecture search, which also seeks to automatically find a more efficient model to train. The difference here is that the authors find the model early in the training run, but it seems like the EarlyBERT procedure could be run once and the resulting model architecture could be saved and re-trained like NAS models are. \n2. I found the experimental results to be lacking detail and breadth necessary to establish the value of the technique. Firstly, the rough time estimates in Table 2 are very odd given the primary value of the proposed technique is to reduce training time. The accuracy of EarlyBERT is close enough to LayerDrop that accurate training cost numbers are needed to differentiate between the techniques. Secondly, quoting training time reductions over the dense baseline when the EarlyBERT mode does not achieve the same accuracy makes the comparison very difficult to make. This problem shows up quite commonly in the model compression literature [1] and I’d encourage the authors to show full accuracy-training time tradeoff curves so that the training time savings for a given accuracy can be more clearly established. Lastly, I found the use of reduced training epochs in EarlyBERT to be odd because you do not evaluate whether or not this can be done for the baseline models and there isn’t clear evidence as to why your model would be able to do this while others (e.g., DropLayer) cannot. Figure 2 also does not seem to corroborate that higher learning rates can be used with shorter training time to achieve better accuracy. The data in your figure shows that the best learning rate achieves the best model quality independent of the number of training epochs.\n\nReferences:\n1. https://arxiv.org/abs/2003.03033\n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604701117603}, {"id": "_QHZA65mG4K", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2348/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an approach to sparsifying BERT. What sets this work apart from prior work on model compression is that while prior works attempt to compress a pre-trained BERT, the authors in this work attempt to learn a sparsified BERT for the purpose of speeding up pre-training. The method essentially involves learning an independent bernoulli mask for all BERT heads along with Bernoulli masks for some later intermediate neurons. This model is then trained, along with the masks, for a few epochs. Then, heads/neurons corresponding to a small value of mask are pruned out. This results in a sparser BERT model, leading to faster pre-training.\n\n#### Pros\nThe paper is well written and the presentation of the contribution is simple and well-motivated. \n\n#### Cons\n1. Since one of the main contributions of this work is to make progress on improving the training / inference speed of large transformers, the authors could spend more time going over how the “Time Saved” column is computed. As of now, it seems casual and hand-wavy.\n2. Experiment protocol:\n  * How were the hyperparameters decided? Could we have uncertainty estimates for all results by reporting mean/std dev. across 3-5 runs (atleast for fine-tuning if doing this for pre-training is too compute intensive). This is especially useful since some of the numbers between EarlyBERT and Random appear very close in Table-1.\n  * While a central argument is that most model distillation techniques still require expensive pre-training, it would still be useful to include some of those results in Table-2 since EarlyBERT is comparable to those techniques for the purpose of Table-2.\n  * One way to contextualize how much time is saved during pre-training would be to report total time required for fine-tuning+pretraining on the entire glue benchmark. This would involve computing the pre-training time (time to learn BERT parameters on Wikipedia) + total fine-tuning time across all datasets (QQP/CoLA/MNLI etc) considered. This could then be compared against alternatives (such as DistilBert and DeeBERT). \n3. Baselines: Experiments on pre-training compare with no baselines. Some possibilities:\n  * Using DeeBERT / other early exit approaches at pre-training time.\n  * Training a BERT model for some epochs, and then distilling it into a smaller network for the rest of the training.\n\n\n#### Misc Nitpicks\n1. The phrase “structured sparsity’ is used in multiple places, but never defined. \n2. You et al. pioneers $\\rightarrow$ You et al. pioneer\n3. Could Prasanna et al. 2020 be extended to sparsify pre-training?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach but needs some more experiments", "review": "This paper proposes an approach to sparsifying BERT. What sets this work apart from prior work on model compression is that while prior works attempt to compress a pre-trained BERT, the authors in this work attempt to learn a sparsified BERT for the purpose of speeding up pre-training. The method essentially involves learning an independent bernoulli mask for all BERT heads along with Bernoulli masks for some later intermediate neurons. This model is then trained, along with the masks, for a few epochs. Then, heads/neurons corresponding to a small value of mask are pruned out. This results in a sparser BERT model, leading to faster pre-training.\n\n#### Pros\nThe paper is well written and the presentation of the contribution is simple and well-motivated. \n\n#### Cons\n1. Since one of the main contributions of this work is to make progress on improving the training / inference speed of large transformers, the authors could spend more time going over how the “Time Saved” column is computed. As of now, it seems casual and hand-wavy.\n2. Experiment protocol:\n  * How were the hyperparameters decided? Could we have uncertainty estimates for all results by reporting mean/std dev. across 3-5 runs (atleast for fine-tuning if doing this for pre-training is too compute intensive). This is especially useful since some of the numbers between EarlyBERT and Random appear very close in Table-1.\n  * While a central argument is that most model distillation techniques still require expensive pre-training, it would still be useful to include some of those results in Table-2 since EarlyBERT is comparable to those techniques for the purpose of Table-2.\n  * One way to contextualize how much time is saved during pre-training would be to report total time required for fine-tuning+pretraining on the entire glue benchmark. This would involve computing the pre-training time (time to learn BERT parameters on Wikipedia) + total fine-tuning time across all datasets (QQP/CoLA/MNLI etc) considered. This could then be compared against alternatives (such as DistilBert and DeeBERT). \n3. Baselines: Experiments on pre-training compare with no baselines. Some possibilities:\n  * Using DeeBERT / other early exit approaches at pre-training time.\n  * Training a BERT model for some epochs, and then distilling it into a smaller network for the rest of the training.\n\n\n#### Misc Nitpicks\n1. The phrase “structured sparsity’ is used in multiple places, but never defined. \n2. You et al. pioneers $\\rightarrow$ You et al. pioneer\n3. Could Prasanna et al. 2020 be extended to sparsify pre-training?", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604310511021}, {"id": "WJCsET4sSBq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2348/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper uses the Lottery Ticket Hypothesis to compress BERT. More specifically, they adapt EarlyBird lottery tickets to the BERT setting in order to find winning configurations in early stages of training combine it with structured pruning methods to ensure the resulting network is more efficient to train. The method is a three-stage process: (1) searching - this phase involves training full BERT with some coefficient parameters to learn the mask, (2) drawing - use the mask to “draw a ticket” or select the sub-network to train, (3) training. Experiments show that performance isn’t that much worse when EarlyBERT is used for fine-tuning and for pre-training.\n\nThe goal of the paper is to find structured winning tickets for BERT in the early stages of training/fine-tuning.\n\nStrengths:\n1. Compressing BERT using the lottery ticket hypothesis has been getting a lot of attention recently, and doing so relatively efficiently is an exciting and interesting contribution.\n2. Compared to previous and contemporary work, the combination of EarlyBird lottery tickets (You et al., 2020) to detect tickets early and network slimming (Liu et al., 2017) for structured pruning, is an interesting one. Prasanna et al., (2020) are doing structured pruning too, but via an iterative pruning method. \n3. Experiments for both pre-training (the first of its kind) and fine-tuning show that performance does not drop all that much for GLUE and Squad which are the main set of tasks BERT is typically evaluated on.  \n\nWeaknesses:\n1. The paper isn’t very clear in some places. It starts out explaining things well but then it gets harder to follow the details. Eg: need to add more information on the mask - it’s binarized at some point? The distance metric is still Hadamard like in EarlyBird? \n2. (More of a question/nit) Is it possible to also show what happens when a winning ticket for BERT fine-tuning is selected based on the pre-training objective? Chen et al., (2020) showed that these make for better tickets that are performant on many of the downstream tasks. It would be interesting to see if this holds true for EarlyBert and would add another layer of fine-tuning efficiency.\n\nThe work isn't terribly novel, but it's still interesting.\n\nQuestions and comments:\n1. Why does the mask distance diverge for FC in pre-training (Figure 1b)? Does it somehow indicate that the training run is degenerate?\n2. The ablation on the regularization parameter didn’t give a clear indication of how important selecting this is. It seems to not be that important, but would using separate values for attention and FC make a difference?\n3. It would probably be helpful to expand the table/figure captions, make them a bit more detailed. \n4. Curious, why did you only use the largest tasks from GLUE?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice work", "review": "Summary:\nThis paper uses the Lottery Ticket Hypothesis to compress BERT. More specifically, they adapt EarlyBird lottery tickets to the BERT setting in order to find winning configurations in early stages of training combine it with structured pruning methods to ensure the resulting network is more efficient to train. The method is a three-stage process: (1) searching - this phase involves training full BERT with some coefficient parameters to learn the mask, (2) drawing - use the mask to “draw a ticket” or select the sub-network to train, (3) training. Experiments show that performance isn’t that much worse when EarlyBERT is used for fine-tuning and for pre-training.\n\nThe goal of the paper is to find structured winning tickets for BERT in the early stages of training/fine-tuning.\n\nStrengths:\n1. Compressing BERT using the lottery ticket hypothesis has been getting a lot of attention recently, and doing so relatively efficiently is an exciting and interesting contribution.\n2. Compared to previous and contemporary work, the combination of EarlyBird lottery tickets (You et al., 2020) to detect tickets early and network slimming (Liu et al., 2017) for structured pruning, is an interesting one. Prasanna et al., (2020) are doing structured pruning too, but via an iterative pruning method. \n3. Experiments for both pre-training (the first of its kind) and fine-tuning show that performance does not drop all that much for GLUE and Squad which are the main set of tasks BERT is typically evaluated on.  \n\nWeaknesses:\n1. The paper isn’t very clear in some places. It starts out explaining things well but then it gets harder to follow the details. Eg: need to add more information on the mask - it’s binarized at some point? The distance metric is still Hadamard like in EarlyBird? \n2. (More of a question/nit) Is it possible to also show what happens when a winning ticket for BERT fine-tuning is selected based on the pre-training objective? Chen et al., (2020) showed that these make for better tickets that are performant on many of the downstream tasks. It would be interesting to see if this holds true for EarlyBert and would add another layer of fine-tuning efficiency.\n\nThe work isn't terribly novel, but it's still interesting.\n\nQuestions and comments:\n1. Why does the mask distance diverge for FC in pre-training (Figure 1b)? Does it somehow indicate that the training run is degenerate?\n2. The ablation on the regularization parameter didn’t give a clear indication of how important selecting this is. It seems to not be that important, but would using separate values for attention and FC make a difference?\n3. It would probably be helpful to expand the table/figure captions, make them a bit more detailed. \n4. Curious, why did you only use the largest tasks from GLUE?\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603857256459}, {"id": "LxOVxNXxUNX", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2348/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The main contribution of this work is to use Early Bird Lottery Tickets to reduce pre-training and fine tuning time for BERT. In order to reduce computation, the authors propose pruning the number of attention heads and neurons in the fully connected layers. They demonstrate that the technique works for BERT-Base and BERT-Large for GLUE and SQUAD tasks. \n\nThe work is well-thought through and authors do a good job of explaining their approach and other existing works using lottery tickets.  The usage of lottery tickets during the pre training phase is the biggest strength of the paper since it can result in significant computational savings. The authors perform interesting ablation studies but they could be augmented by a few more experiments (see below). The paper could also be improved by comparison with relevant prior work. \n\nHere are some thoughts and questions that could help improve the paper: \n\n* Experiments in Section 4.2 are only during the fine tuning stage. How do these results compare with prior work for drawing lottery tickets in transfer learning for NLP (Chen et al, Prasanna et. al)? Particularly, it would be good to compare against Chen et al since their work is very relevant. \n* How long is the searching stage followed by the efficient training stage? Does this remain the same for all the tasks? Or did it require tuning? It would be interesting to see how soon we can switch from the searching stage to the efficient training stage.\n* The paper proposes two different approaches for pruning: pruning attention heads and removing neurons from fully connected layers. It would be interesting to see how the accuracy of the approach changes if only neurons or attention heads are pruned. \n* In Section 4.3, the authors discuss reducing training time by reducing training steps for EarlyBERT. I think a similar analysis for BERT would be helpful to understand if the training time can be reduced for the baseline model as well. Also, pre-training techniques have tended to show improvements in downstreams tasks with longer pre-training as the long as the pre-training dataset is large enough. So, perhaps this reduction in pre-training steps might not be applicable for larger pre-training datasets.\n\nOverall, I find the approach interesting and the authors show computational savings in the pre-training models. I recommend accepting the paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach to reduce pre-training approaches for large NLP models", "review": "The main contribution of this work is to use Early Bird Lottery Tickets to reduce pre-training and fine tuning time for BERT. In order to reduce computation, the authors propose pruning the number of attention heads and neurons in the fully connected layers. They demonstrate that the technique works for BERT-Base and BERT-Large for GLUE and SQUAD tasks. \n\nThe work is well-thought through and authors do a good job of explaining their approach and other existing works using lottery tickets.  The usage of lottery tickets during the pre training phase is the biggest strength of the paper since it can result in significant computational savings. The authors perform interesting ablation studies but they could be augmented by a few more experiments (see below). The paper could also be improved by comparison with relevant prior work. \n\nHere are some thoughts and questions that could help improve the paper: \n\n* Experiments in Section 4.2 are only during the fine tuning stage. How do these results compare with prior work for drawing lottery tickets in transfer learning for NLP (Chen et al, Prasanna et. al)? Particularly, it would be good to compare against Chen et al since their work is very relevant. \n* How long is the searching stage followed by the efficient training stage? Does this remain the same for all the tasks? Or did it require tuning? It would be interesting to see how soon we can switch from the searching stage to the efficient training stage.\n* The paper proposes two different approaches for pruning: pruning attention heads and removing neurons from fully connected layers. It would be interesting to see how the accuracy of the approach changes if only neurons or attention heads are pruned. \n* In Section 4.3, the authors discuss reducing training time by reducing training steps for EarlyBERT. I think a similar analysis for BERT would be helpful to understand if the training time can be reduced for the baseline model as well. Also, pre-training techniques have tended to show improvements in downstreams tasks with longer pre-training as the long as the pre-training dataset is large enough. So, perhaps this reduction in pre-training steps might not be applicable for larger pre-training datasets.\n\nOverall, I find the approach interesting and the authors show computational savings in the pre-training models. I recommend accepting the paper. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603830415080}, {"id": "LjymxTRiccT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2348/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tackles a very important and under-studied problem: reducing the cost of training NLP models. The authors present a method that builds on the lottery ticket hypothesis (LTH). The authors first identify redundant structures early during training, then prune these structures, which leads to faster training. The authors experiment both with pre-training and fine-tuning of contextual models (BERT-{base,large}) and claim large reduction in training time, with reasonable loss in performance. \nDespite very encouraging results, several important methodological questions about the source of the efficiency gains and other aspects of the paper are left unanswered. I cannot recommend accepting this paper in its current form, but am looking forward to reading the authors' response which might clarify things.\n\n\nDetailed comments:\n\nTraining (especially pretraining) costs have been going wild in AI and NLP more particularly, which leads to large costs ([1]) as well as potential environmental problems ([2]). Reducing these costs could have a very high impact on the field, allowing many more researchers to participate in state-of-the-art research [3].\nAs a result, this paper has a great potential, and its results seem very promising. \nNonetheless, the current paper leaves too many open questions regarding the validity of the experiments.\nFirst, much of the improvement (I think) comes from reducing the number of epochs and/or the number of steps. For fine-tuning, the authors run their model for 2.2 epochs, while their baseline model runs for 3 epochs, roughly 30% more which accounts for much of the reduction observed in Table 2. Similarly, for pretraining, the model runs 80% of the training steps (20% reduction), which accounts much of the training time reduction reported on section 4.3. Running a baseline model that runs *for the same amount of time* is essential to appreciate the contribution of this work (e.g., repeat the same analysis in Figure 3 for the vanilla BERT).\nSecond, the authors argue for a large reduction in runtime, but are very cryptic about how they actually measure this reduction (footnote 2), while reporting number in ranges of 5%. As the main contribution of this paper is the increased efficiency of the proposed approach, it must be clear how efficiency is measured. Finally, writing in general can be made clearer:\n\n1. Last sentence of intro: \"without scarifying accuracy\" seems like an inaccurate description of the results presented in this paper. around 1% might be reasonable for 30-40% reduction in training time, but it is certainly a reduction in accuracy.\n2. Some description of Network sliming would help\n3. The term \"intermediate neurons\" (section 3.2) was unclear to me.\n4. Section 3.3: how is mask difference defined?\n5. Figure 1 was unclear to me. What do the axis represent? The authors say \"the axes in the plots are the number of training steps finished.\" so why do you need two of them?\n6. The \"Non-trivial Sub-network\" paragraph feels like it should be part of the Experiments section.\n7. Implementation details are only given for the vanilla BERT Are they similar to the EarlyBERT model as well?\n8. \"Since we observe that the randomly pruned models do not competitive performance ...\": how uncompetitive? I would have liked to see these results (also, please fix grammar in this sentence)\n9. \"Reducing it to 80% seems to be a sweet point with the best balance between performance and efficiency.\" -> I would disagree. The graph indicates that for MNLI and QNLI 60% seems like a better choice. \n\nQuestions:\n1. \"We observe empirically that if pruned globally, the attention heads in some layers may be completely removed, making the network un-trainable.\": in this case, couldn't the authors remove a full layer? \n2. The difference in the ablation results seem quite small (tables 3 and 4). Are they statistically significant?\n\nMinor:\n1. Missing period at the end of the first paragraph in the related work section.\n2. Second paragraph of related work: McCarley et al. (2019) appears twice with different descriptions, is this intentional?\n3. The authors say \"we focus on larger datasets from GLUE (MNLI, QNLI, QQP and SST-2), as it is less meaningful to discuss efficient training\", but then report and analyze results from other GLUE datasets as well.\n4. \"Hon downstream tasks with smaller learning rate\" -> Do you mean smaller datasets?\n\n\n\n\n\nReferences:\n[1] Sharir, O., Peleg, B., and Shoham, Y. (2020). The cost of training NLP models: A concise overview. arXiv:2004.08900.\n[2] Strubell, E., Ganesh, A., and McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. In Proc. of ACL.\n[3] Schwartz, R., Dodge, J., Smith, N. A., and Etzioni, O. (2019). Green AI. arXiv:1907.10597.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and promising results, but important methodological unclarities.", "review": "This paper tackles a very important and under-studied problem: reducing the cost of training NLP models. The authors present a method that builds on the lottery ticket hypothesis (LTH). The authors first identify redundant structures early during training, then prune these structures, which leads to faster training. The authors experiment both with pre-training and fine-tuning of contextual models (BERT-{base,large}) and claim large reduction in training time, with reasonable loss in performance. \nDespite very encouraging results, several important methodological questions about the source of the efficiency gains and other aspects of the paper are left unanswered. I cannot recommend accepting this paper in its current form, but am looking forward to reading the authors' response which might clarify things.\n\n\nDetailed comments:\n\nTraining (especially pretraining) costs have been going wild in AI and NLP more particularly, which leads to large costs ([1]) as well as potential environmental problems ([2]). Reducing these costs could have a very high impact on the field, allowing many more researchers to participate in state-of-the-art research [3].\nAs a result, this paper has a great potential, and its results seem very promising. \nNonetheless, the current paper leaves too many open questions regarding the validity of the experiments.\nFirst, much of the improvement (I think) comes from reducing the number of epochs and/or the number of steps. For fine-tuning, the authors run their model for 2.2 epochs, while their baseline model runs for 3 epochs, roughly 30% more which accounts for much of the reduction observed in Table 2. Similarly, for pretraining, the model runs 80% of the training steps (20% reduction), which accounts much of the training time reduction reported on section 4.3. Running a baseline model that runs *for the same amount of time* is essential to appreciate the contribution of this work (e.g., repeat the same analysis in Figure 3 for the vanilla BERT).\nSecond, the authors argue for a large reduction in runtime, but are very cryptic about how they actually measure this reduction (footnote 2), while reporting number in ranges of 5%. As the main contribution of this paper is the increased efficiency of the proposed approach, it must be clear how efficiency is measured. Finally, writing in general can be made clearer:\n\n1. Last sentence of intro: \"without scarifying accuracy\" seems like an inaccurate description of the results presented in this paper. around 1% might be reasonable for 30-40% reduction in training time, but it is certainly a reduction in accuracy.\n2. Some description of Network sliming would help\n3. The term \"intermediate neurons\" (section 3.2) was unclear to me.\n4. Section 3.3: how is mask difference defined?\n5. Figure 1 was unclear to me. What do the axis represent? The authors say \"the axes in the plots are the number of training steps finished.\" so why do you need two of them?\n6. The \"Non-trivial Sub-network\" paragraph feels like it should be part of the Experiments section.\n7. Implementation details are only given for the vanilla BERT Are they similar to the EarlyBERT model as well?\n8. \"Since we observe that the randomly pruned models do not competitive performance ...\": how uncompetitive? I would have liked to see these results (also, please fix grammar in this sentence)\n9. \"Reducing it to 80% seems to be a sweet point with the best balance between performance and efficiency.\" -> I would disagree. The graph indicates that for MNLI and QNLI 60% seems like a better choice. \n\nQuestions:\n1. \"We observe empirically that if pruned globally, the attention heads in some layers may be completely removed, making the network un-trainable.\": in this case, couldn't the authors remove a full layer? \n2. The difference in the ablation results seem quite small (tables 3 and 4). Are they statistically significant?\n\nMinor:\n1. Missing period at the end of the first paragraph in the related work section.\n2. Second paragraph of related work: McCarley et al. (2019) appears twice with different descriptions, is this intentional?\n3. The authors say \"we focus on larger datasets from GLUE (MNLI, QNLI, QQP and SST-2), as it is less meaningful to discuss efficient training\", but then report and analyze results from other GLUE datasets as well.\n4. \"Hon downstream tasks with smaller learning rate\" -> Do you mean smaller datasets?\n\n\n\n\n\nReferences:\n[1] Sharir, O., Peleg, B., and Shoham, Y. (2020). The cost of training NLP models: A concise overview. arXiv:2004.08900.\n[2] Strubell, E., Ganesh, A., and McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. In Proc. of ACL.\n[3] Schwartz, R., Dodge, J., Smith, N. A., and Etzioni, O. (2019). Green AI. arXiv:1907.10597.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603546746552}], "openreview_url": "https://openreview.net/forum?id=I-VfjSBzi36", "arxiv_id": "2101.00063", "paper_pdf": "papers/I-VfjSBzi36.pdf", "paper_pdf_sha256": "50e16d340f463235f2dedd0a5f59f01464523ad0768145cc059c3878a530865a", "paper_pdf_bytes": 2139091, "paper_pdf_source": "openreview", "code_url": "https://github.com/VITA-Group/EarlyBERT", "code_repository": "VITA-Group/EarlyBERT", "code_commit": "593624240d17fb8e1dee9b7ccf83641d603d5ce9", "code_archive": "repos/I-VfjSBzi36.zip", "code_archive_sha256": "3ba5735d02c31421d99d77fa743c0e36e7dbb80bb211da08fbf52818fe06e08b", "code_archive_bytes": 94512, "code_file_count": 13, "code_extensions": {".py": 11, ".sh": 2}, "github_disk_usage_kb": 71, "github_languages": {"Python": 379037, "Shell": 5680}, "github_archived": false, "github_pushed_at": "2021-12-30T05:58:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/earlybert-efficient-bert-training-via-early-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1lsXREYvr", "year": 2020, "status": "rejected", "title": "One-Shot Neural Architecture Search via Compressive Sensing", "authors": ["Minsu Cho", "Mohammadreza Soltani", "Chinmay Hegde"], "authorids": ["chomd90@iastate.edu", "mohammadreza.soltani@duke.edu", "chinmay@iastate.edu"], "authors_source": "OpenReview API", "abstract": "Neural architecture search (NAS), or automated design of neural network models, remains a very challenging meta-learning problem. Several recent works (called \"one-shot\" approaches) have focused on dramatically reducing NAS running time by leveraging proxy models that still provide architectures with competitive performance. In our work, we propose a new meta-learning algorithm that we call CoNAS, or Compressive sensing-based Neural Architecture Search. Our approach merges ideas from one-shot NAS approaches with iterative techniques for learning low-degree sparse Boolean polynomial functions. We validate our approach on several standard test datasets, discover novel architectures hitherto unreported, and achieve competitive (or better) results in both performance and search time compared to existing NAS approaches. Further, we provide theoretical analysis via upper bounds on the number of validation error measurements needed to perform reliable meta-learning; to our knowledge, these analysis tools are novel to the NAS literature and may be of independent interest.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJxy8jjAcH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1050/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new algorithm for one-shot neural architecture search (NAS) via compressive sensing. The authors propose a new search strategy, as well as a slightly different search space compared to DARTS [1], ProxylessNAS [2], etc. They use architecture samples from the one-shot model evaluated with the search parameters as a surrogate of the true objective in order to speed-up the search. Afterwards, these surrogate function evaluations are used to compute Fourier coefficients which are eventually used to optimize the vector of binary parameters encoding the architecture. \n\nOverall, I think the proposed algorithm is interesting and of practical usefulness. However, in terms of novelty, this work seems more to be an application of Harmonica [6] to the NAS problem (with small modifications in order to make it applicable). In page 10 you state some of the differences of your method with Harmonica. I agree that the number of function evaluations you use (coming from the one-shot model) is larger and computationally less expensive to obtain, however this does not guarantee that these are a good surrogate of the true objective that NAS aims to minimize, i.e. the validation/test accuracy of final (stand-alone) architectures . The empirical evaluations of their algorithm seem to outperform/be competitive compared to other NAS methods on all benchmarks used in the paper, however only DARTS is evaluated on their search space and the other results are taken from the corresponding papers. The paper is well-written and -structured with the caveat of being more than the recommended 8 pages.\n\nI will adjust my score depending on the authors responses concerning the following questions/issues:\n\n1. The correlation between the architectures evaluated using the one-shot weights and retrained from scratch, seems to be of crucial importance in your method, since you directly use the one-shot weights to collect the measurements, similarly to Random Search with weight sharing [3], ENAS [4] or Bender et al. [5]. What is the correlation of these measurements with the stand-alone architectures trained from scratch using the final evaluation settings? How did you tune the p in the Bernoulli distribution during the one-shot weight updates. According to Bender et al. [5] the ScheduledDropPath probability is an important hyperparameter affecting the aforementioned correlation.\n\n2. What is the main motivation for using 5 operations in the operation set and not 8 as in DARTS [1] for example? Does the main contribution in the competitive results come from the different search space or the search method?\n\n3. Is there any reference or proof for the correctness of Theorem 3.2?\n\n4. I think there are some parts that can be moved in the Supplementary, such as the pseudocode for the proposed algorithm or Figure 3, and some other parts that can be compressed, such as the Related Work section.\n\t\nReferences\n[1] Hanxiao Liu, Karen Simonyan, and Yiming Yang.  DARTS: Differentiable architecture search.  In ICLR, 2019.\n[2] Han Cai, Ligeng Zhu, Song Han. ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware. In ICLR, 2019.\n[3] LIAM LI, AMEET TALWALKAR. Random Search and Reproducibility for Neural Architecture Search.\n[4] Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, Jeff Dean. Efficient Neural Architecture Search via Parameter Sharing. In ICML, 2018\n[5] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, Quoc Le. Understanding and Simplifying One-Shot Architecture Search. In ICML, 2018\n[6] Elad Hazan, Adam Klivans, Yang Yuan. Hyperparameter Optimization: A Spectral Approach\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a new algorithm for one-shot neural architecture search (NAS) via compressive sensing. The authors propose a new search strategy, as well as a slightly different search space compared to DARTS [1], ProxylessNAS [2], etc. They use architecture samples from the one-shot model evaluated with the search parameters as a surrogate of the true objective in order to speed-up the search. Afterwards, these surrogate function evaluations are used to compute Fourier coefficients which are eventually used to optimize the vector of binary parameters encoding the architecture. \n\nOverall, I think the proposed algorithm is interesting and of practical usefulness. However, in terms of novelty, this work seems more to be an application of Harmonica [6] to the NAS problem (with small modifications in order to make it applicable). In page 10 you state some of the differences of your method with Harmonica. I agree that the number of function evaluations you use (coming from the one-shot model) is larger and computationally less expensive to obtain, however this does not guarantee that these are a good surrogate of the true objective that NAS aims to minimize, i.e. the validation/test accuracy of final (stand-alone) architectures . The empirical evaluations of their algorithm seem to outperform/be competitive compared to other NAS methods on all benchmarks used in the paper, however only DARTS is evaluated on their search space and the other results are taken from the corresponding papers. The paper is well-written and -structured with the caveat of being more than the recommended 8 pages.\n\nI will adjust my score depending on the authors responses concerning the following questions/issues:\n\n1. The correlation between the architectures evaluated using the one-shot weights and retrained from scratch, seems to be of crucial importance in your method, since you directly use the one-shot weights to collect the measurements, similarly to Random Search with weight sharing [3], ENAS [4] or Bender et al. [5]. What is the correlation of these measurements with the stand-alone architectures trained from scratch using the final evaluation settings? How did you tune the p in the Bernoulli distribution during the one-shot weight updates. According to Bender et al. [5] the ScheduledDropPath probability is an important hyperparameter affecting the aforementioned correlation.\n\n2. What is the main motivation for using 5 operations in the operation set and not 8 as in DARTS [1] for example? Does the main contribution in the competitive results come from the different search space or the search method?\n\n3. Is there any reference or proof for the correctness of Theorem 3.2?\n\n4. I think there are some parts that can be moved in the Supplementary, such as the pseudocode for the proposed algorithm or Figure 3, and some other parts that can be compressed, such as the Related Work section.\n\t\nReferences\n[1] Hanxiao Liu, Karen Simonyan, and Yiming Yang.  DARTS: Differentiable architecture search.  In ICLR, 2019.\n[2] Han Cai, Ligeng Zhu, Song Han. ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware. In ICLR, 2019.\n[3] LIAM LI, AMEET TALWALKAR. Random Search and Reproducibility for Neural Architecture Search.\n[4] Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, Jeff Dean. Efficient Neural Architecture Search via Parameter Sharing. In ICML, 2018\n[5] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, Quoc Le. Understanding and Simplifying One-Shot Architecture Search. In ICML, 2018\n[6] Elad Hazan, Adam Klivans, Yang Yuan. Hyperparameter Optimization: A Spectral Approach\n"}, "tcdate": 1572940614724}, {"id": "rygy26eptr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1050/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Contributions:\nThis paper tackles the problem of One-shot Neural architecture search by proposing a new method. \nThe method consists mainly of new search strategy of the optimal architecture that is inspired by the recovery of boolean functions from their sparse Fourier expansions. As such, this work is an application of recent progress in the field of compressive sensing to One-shot neural architecture search. Given the problem formalism, the authors have also provides guarantee for the optimality of their method, i.e the method can recover the optimal sub-network of any given  a sufficient number of performance measurements.\n\nClarity\nOverall, the paper is well motivated and the technical content is good. That said the structure could be enormously  improved to ease the reading and the overall understanding. For example: better caption for Figure 2 explaining what is shown; presenting the pseudo-code directly in the method overview and spending the rest of the section explaining the method; showing the related work before the experiments; etc.\n\nNovelty\nThe main novelty in my opinion is the application of compressive sensing methods to One-shot NAS. This approach is significantly different from other One-shot NAS method that I am aware of mainly regarding the search strategy employed to find the best architecture. \nHowever, this work seems like an incremental improvement over Hazan et al 2018. To this regard, the only novelty that was the framing of One-shot NAS as a recovery of boolean functions from their sparse Fourier expansions is not new either.  That is said, I am open to be proven wrong on this point!\n\nResults\nThe experiment section is not self-content, the readers is refered a couple of times to other papers to get details that are  critical to reproducibility and understanding. \nOverall, the search strategy of CoNAS seems  parameter efficient, fast and competitive. Also, small ablation studies showing the effect of the different parameters of CoNAS were very informative and well-appreciated.\nHowever,  the search space is different between CoNAS and the others methods for some experiments, making it difficult to decide if the search strategy of CoNAS is definitely competitive compared to other methods or not.\n\nPoints of improvement:\n1 - Structure of the paper\n2 - Clarify novelty compared to HARMONICA ( not the application domain please)\n3 - demonstrate that with the same search space your method is competitive.\n4 - Does m=1000 in your experiments satisfies theorem 3.2? What is the value of d in your experiments? Can you provide supporting experiments that answer those questions?\n\nPreliminary decision:\nFor now, I will say *weak reject* \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Contributions:\nThis paper tackles the problem of One-shot Neural architecture search by proposing a new method. \nThe method consists mainly of new search strategy of the optimal architecture that is inspired by the recovery of boolean functions from their sparse Fourier expansions. As such, this work is an application of recent progress in the field of compressive sensing to One-shot neural architecture search. Given the problem formalism, the authors have also provides guarantee for the optimality of their method, i.e the method can recover the optimal sub-network of any given  a sufficient number of performance measurements.\n\nClarity\nOverall, the paper is well motivated and the technical content is good. That said the structure could be enormously  improved to ease the reading and the overall understanding. For example: better caption for Figure 2 explaining what is shown; presenting the pseudo-code directly in the method overview and spending the rest of the section explaining the method; showing the related work before the experiments; etc.\n\nNovelty\nThe main novelty in my opinion is the application of compressive sensing methods to One-shot NAS. This approach is significantly different from other One-shot NAS method that I am aware of mainly regarding the search strategy employed to find the best architecture. \nHowever, this work seems like an incremental improvement over Hazan et al 2018. To this regard, the only novelty that was the framing of One-shot NAS as a recovery of boolean functions from their sparse Fourier expansions is not new either.  That is said, I am open to be proven wrong on this point!\n\nResults\nThe experiment section is not self-content, the readers is refered a couple of times to other papers to get details that are  critical to reproducibility and understanding. \nOverall, the search strategy of CoNAS seems  parameter efficient, fast and competitive. Also, small ablation studies showing the effect of the different parameters of CoNAS were very informative and well-appreciated.\nHowever,  the search space is different between CoNAS and the others methods for some experiments, making it difficult to decide if the search strategy of CoNAS is definitely competitive compared to other methods or not.\n\nPoints of improvement:\n1 - Structure of the paper\n2 - Clarify novelty compared to HARMONICA ( not the application domain please)\n3 - demonstrate that with the same search space your method is competitive.\n4 - Does m=1000 in your experiments satisfies theorem 3.2? What is the value of d in your experiments? Can you provide supporting experiments that answer those questions?\n\nPreliminary decision:\nFor now, I will say *weak reject* \n\n"}, "tcdate": 1571782054786}, {"id": "HJxQXdh-Fr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1050/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors study Neural Architecture Search, which aims to automate design of neural network models. Their approach consists in using a two stage algorithm:\n-\tA first Neural Network $f$ is trained for predicting the performances of sub-architectures. Then binary sub-graphs coding the sub-architectures are uniformly sampled (Bernouilli(0.5)), and their performances y are evaluated thanks to the first Neural Network.\n-\tThe graph sampling matrix A, which is indexed by the m sampled architectures and the Fourier basis of size $O(n^d)$, is built. Then the optimization problem $x^*= \\arg\\min_x ||y – Ax||$ is solved using Lasso. The largest Fourier coefficients are chosen to build an estimate g of f. Finally, computing minimum of g, the architecture is generated. \n\nSince $m << n^d$, the optimization problem is ill-posed. Theorem 3.2 shows that if A satisfies the restricted isometry of order s, then the sparse coefficients x can be recovered.\nThe algorithm is evaluated and compared to the state-of-the-art on various image classification tasks and on RNN.\n\n\nMajor concern:\n\n1/ I did not find the proof of Theorem 3.2 in the main paper and in the appendix, so I do not buy it.\n\n2/ The authors claim that their algorithm performs better than the state-of-the-art, but according to tables 1,2,3, I did not find significant differences of performances in term of test errors.\n\n3/ The key idea of the algorithm is not well explained. A one-shot NAS f is pre-trained. f is assumed to be well-trained. Then it is approximated with a Fourier-sparse Boolean function. Why using an approximation if f is perfect? Do you expect to reduce the needed number of sampled architectures?\n\n\nMinor concerns:\n \nEquation 3.1 is not clear. $X_S (\\alpha_l)$ does not depend on k. So all rows seem to be the same.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1050", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "In this paper, the authors study Neural Architecture Search, which aims to automate design of neural network models. Their approach consists in using a two stage algorithm:\n-\tA first Neural Network $f$ is trained for predicting the performances of sub-architectures. Then binary sub-graphs coding the sub-architectures are uniformly sampled (Bernouilli(0.5)), and their performances y are evaluated thanks to the first Neural Network.\n-\tThe graph sampling matrix A, which is indexed by the m sampled architectures and the Fourier basis of size $O(n^d)$, is built. Then the optimization problem $x^*= \\arg\\min_x ||y – Ax||$ is solved using Lasso. The largest Fourier coefficients are chosen to build an estimate g of f. Finally, computing minimum of g, the architecture is generated. \n\nSince $m << n^d$, the optimization problem is ill-posed. Theorem 3.2 shows that if A satisfies the restricted isometry of order s, then the sparse coefficients x can be recovered.\nThe algorithm is evaluated and compared to the state-of-the-art on various image classification tasks and on RNN.\n\n\nMajor concern:\n\n1/ I did not find the proof of Theorem 3.2 in the main paper and in the appendix, so I do not buy it.\n\n2/ The authors claim that their algorithm performs better than the state-of-the-art, but according to tables 1,2,3, I did not find significant differences of performances in term of test errors.\n\n3/ The key idea of the algorithm is not well explained. A one-shot NAS f is pre-trained. f is assumed to be well-trained. Then it is approximated with a Fourier-sparse Boolean function. Why using an approximation if f is perfect? Do you expect to reduce the needed number of sampled architectures?\n\n\nMinor concerns:\n \nEquation 3.1 is not clear. $X_S (\\alpha_l)$ does not depend on k. So all rows seem to be the same.\n\n"}, "tcdate": 1571043354983}], "openreview_url": "https://openreview.net/forum?id=B1lsXREYvr", "arxiv_id": "1906.02869", "paper_pdf": "papers/B1lsXREYvr.pdf", "paper_pdf_sha256": "3138e5c9053cc6294421860b58a08677aba87909d765229335699d81757b80fa", "paper_pdf_bytes": 658471, "paper_pdf_source": "openreview", "code_url": "https://github.com/chomd90/CoNAS_release", "code_repository": "chomd90/CoNAS_release", "code_commit": "1df9f24f13fa5c8624f4fdcdf5e4e937851baf9d", "code_archive": "repos/B1lsXREYvr.zip", "code_archive_sha256": "a794179f5a79aee5af39cb2442fd9e707423a6e830fede32c8c0776c2f2b07de", "code_archive_bytes": 107890, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 105, "github_languages": {"Python": 64031}, "github_archived": false, "github_pushed_at": "2021-02-26T18:29:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/one-shot-neural-architecture-search-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJfRpoA9YX", "year": 2019, "status": "rejected", "title": "Adversarial Information Factorization", "authors": ["Antonia Creswell", "Yumnah Mohamied", "Biswa Sengupta", "Anil Bharath"], "authorids": ["ac2211@ic.ac.uk", "ym1008@ic.ac.uk", "biswasengupta@gmail.com", "aab01@ic.ac.uk"], "authors_source": "OpenReview API", "abstract": "We propose a novel generative model architecture designed to learn representations for images that factor out a single attribute from the rest of the representation. A single object may have many attributes which when altered do not change the identity of the object itself. Consider the human face; the identity of a particular person is independent of whether or not they happen to be wearing glasses. The attribute of wearing glasses can be changed without changing the identity of the person. However, the ability to manipulate and alter image attributes without altering the object identity is not a trivial task. Here, we are interested in learning a representation of the image that separates the identity of an object (such as a human face) from an attribute (such as 'wearing glasses'). We demonstrate the success of our factorization approach by using the learned representation to synthesize the same face with and without a chosen attribute. We refer to this specific synthesis process as image attribute manipulation. We further demonstrate that our model achieves competitive scores, with state of the art, on a facial attribute classification task.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "H1gUKn2d2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper851/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors introduce a neural network architecture that has three components.\nFirst a VAE is used to encode images in to two latent states \\hat{y} and \\hat{z}, with \\hat{z}\nintended to be class (e.g. face attribute) agnostic. The decoder reconstructs images from \\hat{y}\nand \\hat{z} concatenated together. A GAN style discriminator attempts to distinguish the \ndecoded image from the original input image as real or fake, allowing the decoder to produce \nhigher quality decoded images. An auxiliary network A attempts to classify the face attribute y\nfrom the class agnostic features \\hat{z}, with the idea being that the encoder should try to produce \n\\hat{z} vectors from which the class cannot be predicted. An additional classifier is trained\nusing a classification loss \\hat{L}_{class} on the encoded reconstructed image, the use of which \nI don't understand.\n\nI think additional work on section 2.5 through section 3 would be helpful to improve clarity.\nAs one example, \"y\" is unnecessarily overloaded: y denotes a specific attribute, \\hat{y}\ndenotes a latent vector that is intended to not be class agnostic, \\tilde{y} denotes the\nprediction of an auxiliary network on an intended class-agnostic latent vector \\hat{z} of\nthe presence of the original attribute y, and \\hat{\\hat{y}} denotes the non agnostic latent\nvector achieved by passing the decoded image back through the encoder.\n\nThis notational complexity is compounded by the fact that a number of steps in the method are\nnot well motivated in the text, and left to the reader to understand their purpose. For example,\nthe authors state that \"we incorporate a classification model into the encoder so that our model may\neasily be used to perform classification tasks.\" What does this mean? In the diagram (Figure 1),\nwhere is this classification model? Why in the GAN loss is there a term that compares the\nfake loss with the result of classifying a decoded z vector? Is this z \\hat{z}, or a latent vector\ndrawn from a distribution p(z)? If it is the former, how does this term differ from the second\nterm in the GAN loss. If it is the latter, then shouldn't it be concatenated with some y in order to\nbe used as input to the decoder D_{\\theta}?\n\nWhy is it important to extract \\hat{\\hat{y}} from \\hat{x}? In the paper you state that the loss\n\"provides a gradient containing label information to the decoder,\" but why can't we use the known label y\nof the original input x to ensure that the encoder and decoder preserve this information if it is used as \\hat{y}?\nLater in the paper, you explicitly state that \\hat{\\mathcal{L}_{class}} \"does not provide any clear benefit.\"\nIf that is the case, then you should ideally include it neither in the model nor in the paper. If it was\nincluded primarily because previous models included it, then I would recommend you introduce its use\nin a background section on Bao et al., 2017 rather than including it in your model description with an\nexplanation like \"so that our model may easily be used to perform classification tasks.\"\n\nUltimately, this last point brings us to a good summary of my concerns with the model: the inclusion\nof too many moving parts, some of which the authors explicitly say later on provide no benefit.\n\nMoving on to experimental results, I think this is another area where I have a few concerns. First, in\nFigure 2, the authors argue that your model is \"better for 6 out of 10 attributes\" and comparable results for most others. The authors include a gap of 0.1 in the \"Gray_hair\" category as \"better\" but label a gap of 0.5\nin the Black hair category as \"comparable.\" I think results in several of the categories are sufficiently close\nthat error bars would be necessary to draw actual conclusions. If \"better\" were to mean \"better by 0.5\" for example,\nthen the authors method is better on 4 tasks (smiling, blonde hair, heavy makeup, mustache) and worse on 3 (black hair, brown hair, wavy hair).\n\nWith respect to the actual attribute editing, my main concern here is a lack of comparison to models other than Bao et al., despite the fact that face attribute changing is an exhaustively studied task. A number of papers like Perarnau et al., 2016, Upchurch et al., 2017, Lample et al., 2017 and others study this task from machine learning perspectives, and in some cases can perform photorealistic image attribute editing without complicated machinery on megapixel face\nimages. At least the images in Figure 3 and 4 are substantially downsampled from the typical resolution found in the Celeba dataset, suggesting that there was some failure mode on full resolution images.\n\n----\n\nEdit: I've reviewed the authors' addressing my concerns in their paper and am happy to increase my rating as a result.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Method clarity can be improved, and lacks some key comparisons experimentally.", "review": "In this paper, the authors introduce a neural network architecture that has three components.\nFirst a VAE is used to encode images in to two latent states \\hat{y} and \\hat{z}, with \\hat{z}\nintended to be class (e.g. face attribute) agnostic. The decoder reconstructs images from \\hat{y}\nand \\hat{z} concatenated together. A GAN style discriminator attempts to distinguish the \ndecoded image from the original input image as real or fake, allowing the decoder to produce \nhigher quality decoded images. An auxiliary network A attempts to classify the face attribute y\nfrom the class agnostic features \\hat{z}, with the idea being that the encoder should try to produce \n\\hat{z} vectors from which the class cannot be predicted. An additional classifier is trained\nusing a classification loss \\hat{L}_{class} on the encoded reconstructed image, the use of which \nI don't understand.\n\nI think additional work on section 2.5 through section 3 would be helpful to improve clarity.\nAs one example, \"y\" is unnecessarily overloaded: y denotes a specific attribute, \\hat{y}\ndenotes a latent vector that is intended to not be class agnostic, \\tilde{y} denotes the\nprediction of an auxiliary network on an intended class-agnostic latent vector \\hat{z} of\nthe presence of the original attribute y, and \\hat{\\hat{y}} denotes the non agnostic latent\nvector achieved by passing the decoded image back through the encoder.\n\nThis notational complexity is compounded by the fact that a number of steps in the method are\nnot well motivated in the text, and left to the reader to understand their purpose. For example,\nthe authors state that \"we incorporate a classification model into the encoder so that our model may\neasily be used to perform classification tasks.\" What does this mean? In the diagram (Figure 1),\nwhere is this classification model? Why in the GAN loss is there a term that compares the\nfake loss with the result of classifying a decoded z vector? Is this z \\hat{z}, or a latent vector\ndrawn from a distribution p(z)? If it is the former, how does this term differ from the second\nterm in the GAN loss. If it is the latter, then shouldn't it be concatenated with some y in order to\nbe used as input to the decoder D_{\\theta}?\n\nWhy is it important to extract \\hat{\\hat{y}} from \\hat{x}? In the paper you state that the loss\n\"provides a gradient containing label information to the decoder,\" but why can't we use the known label y\nof the original input x to ensure that the encoder and decoder preserve this information if it is used as \\hat{y}?\nLater in the paper, you explicitly state that \\hat{\\mathcal{L}_{class}} \"does not provide any clear benefit.\"\nIf that is the case, then you should ideally include it neither in the model nor in the paper. If it was\nincluded primarily because previous models included it, then I would recommend you introduce its use\nin a background section on Bao et al., 2017 rather than including it in your model description with an\nexplanation like \"so that our model may easily be used to perform classification tasks.\"\n\nUltimately, this last point brings us to a good summary of my concerns with the model: the inclusion\nof too many moving parts, some of which the authors explicitly say later on provide no benefit.\n\nMoving on to experimental results, I think this is another area where I have a few concerns. First, in\nFigure 2, the authors argue that your model is \"better for 6 out of 10 attributes\" and comparable results for most others. The authors include a gap of 0.1 in the \"Gray_hair\" category as \"better\" but label a gap of 0.5\nin the Black hair category as \"comparable.\" I think results in several of the categories are sufficiently close\nthat error bars would be necessary to draw actual conclusions. If \"better\" were to mean \"better by 0.5\" for example,\nthen the authors method is better on 4 tasks (smiling, blonde hair, heavy makeup, mustache) and worse on 3 (black hair, brown hair, wavy hair).\n\nWith respect to the actual attribute editing, my main concern here is a lack of comparison to models other than Bao et al., despite the fact that face attribute changing is an exhaustively studied task. A number of papers like Perarnau et al., 2016, Upchurch et al., 2017, Lample et al., 2017 and others study this task from machine learning perspectives, and in some cases can perform photorealistic image attribute editing without complicated machinery on megapixel face\nimages. At least the images in Figure 3 and 4 are substantially downsampled from the typical resolution found in the Celeba dataset, suggesting that there was some failure mode on full resolution images.\n\n----\n\nEdit: I've reviewed the authors' addressing my concerns in their paper and am happy to increase my rating as a result.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541094526142}, {"id": "Hyecgv9dhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper851/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \n\nThis paper builds upon the work of Boa et al (2017 ) (Conditional VAE GAN) to allow attribute manipulation in the synthesis process. \n\nIn order to disentangle the identity information from the attributes the paper proposes adversarial information factorization : let z be the latent code and y be the attribute the paper proposes to have p(y) =  p(y|z= E_phi(x)), i.e to have z independent of y.  This disentanglement is implemented through a GAN on the variable y  min _phi Distance (p(y), p(y|z)), the distance is defined via a discriminator on y.  \n\nExperiments are presented on celeba dataset,  1) on attribute manipulation from smiling to non smiling for example, on 2) attribute classification results are presented , 3) ablation studies are given to study the effect of each component of the model highlighting the effect of the adversarial information factorization. \n\nOriginality Novelty: \n\nThere is a large body of work on disentanglement that the paper does not cite or compare to for instance, InfoGAN,  Beta- VAE https://openreview.net/pdf?id=Sy2fzU9gl and disentangled latent concepts https://arxiv.org/pdf/1711.00848.pdf\n\nNote that for example that in beta- VAE it is a similar idea where but it is on z and z|x and the distance used is KL (since it is has closed form with gaussian) , min_phi Loss+ beta KL (p(z), p(z|x)), a discussion of the previous related work in the paper is necessary.  \n\nThe work is also related to MINE https://arxiv.org/pdf/1801.04062.pdf where one would like to minimize the mutual information I(z;y)  this mutual information is estimated through a min/max game.\n \nQuestions: \n\n-  why is RMSprop used for optimization, your model and the Bao et al baseline might benefit from the use of Adam?\n\n- (Table 3 in appendix ) Have you tried higher values of alpha the weight of KL, with the model of Bao et al (it is recommended in beta VAE to have high value of what you call alpha)?\n\nOverall assessment: \n\nThe paper novelty is using min/max game to estimate the mutual information between y (attribute) and z (identity code). Disentanglement and use of min/max games for estimating mutual information has been explored before.  Further discussion and comparaison to previous work is needed. \n\n\n\n \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "missing references to previous work ", "review": "Summary: \n\nThis paper builds upon the work of Boa et al (2017 ) (Conditional VAE GAN) to allow attribute manipulation in the synthesis process. \n\nIn order to disentangle the identity information from the attributes the paper proposes adversarial information factorization : let z be the latent code and y be the attribute the paper proposes to have p(y) =  p(y|z= E_phi(x)), i.e to have z independent of y.  This disentanglement is implemented through a GAN on the variable y  min _phi Distance (p(y), p(y|z)), the distance is defined via a discriminator on y.  \n\nExperiments are presented on celeba dataset,  1) on attribute manipulation from smiling to non smiling for example, on 2) attribute classification results are presented , 3) ablation studies are given to study the effect of each component of the model highlighting the effect of the adversarial information factorization. \n\nOriginality Novelty: \n\nThere is a large body of work on disentanglement that the paper does not cite or compare to for instance, InfoGAN,  Beta- VAE https://openreview.net/pdf?id=Sy2fzU9gl and disentangled latent concepts https://arxiv.org/pdf/1711.00848.pdf\n\nNote that for example that in beta- VAE it is a similar idea where but it is on z and z|x and the distance used is KL (since it is has closed form with gaussian) , min_phi Loss+ beta KL (p(z), p(z|x)), a discussion of the previous related work in the paper is necessary.  \n\nThe work is also related to MINE https://arxiv.org/pdf/1801.04062.pdf where one would like to minimize the mutual information I(z;y)  this mutual information is estimated through a min/max game.\n \nQuestions: \n\n-  why is RMSprop used for optimization, your model and the Bao et al baseline might benefit from the use of Adam?\n\n- (Table 3 in appendix ) Have you tried higher values of alpha the weight of KL, with the model of Bao et al (it is recommended in beta VAE to have high value of what you call alpha)?\n\nOverall assessment: \n\nThe paper novelty is using min/max game to estimate the mutual information between y (attribute) and z (identity code). Disentanglement and use of min/max games for estimating mutual information has been explored before.  Further discussion and comparaison to previous work is needed. \n\n\n\n \n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541084914283}, {"id": "Syglr36U2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper851/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a generative model to learn the representation which can separates the identity of an object from an attribute. Authors extended the autoencoder adversarial by adding an auxiliary network. \n\nStrength\nThe motivation of adding this auxiliary network, which is to distinguish the information between latent code z and attribute vector y, is clean and clear.\nExperiments illustrate the advantage of using auxiliary network and demonstrating the role of classify. Experimental results also show the proposed model learning to factor attributes from identity on the face dataset.\n\nWeakness \nThe proposed model seem to be unnecessarily complex. For example, the loss of  in (6) actually includes 6 components (5 are from L_enc) and 4~5 tuning hyper-parameters. The L_gan also includes 3 parts. The reason of adding gan loss lacks either theoretical or empirical analysis. So as L_KL. In addition, the second term in L_gan is unnecessary since you already have a reconstruction loss. It also make it to be unclear what we obtain if the equilibrium of the GAN objective achieved.\n\nThe written of this paper can be improved to make it more clear. \nIt looks \\hat_y and \\tilde_y are same thing. \nHow do you get \\hat_z? Do you assume the posterior distribution is Gaussian and use the reparameterization trick? What are \\hat_y and \\hat_\\hat_y? Are they binary or a scalar between 0 and 1?  How do you generate \\hat_x? When generating \\hat_x, do you sample \\hat_z and \\hat_y? If so, how do treat the variance problem of \\hat_y? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, Too complex model", "review": "This paper proposed a generative model to learn the representation which can separates the identity of an object from an attribute. Authors extended the autoencoder adversarial by adding an auxiliary network. \n\nStrength\nThe motivation of adding this auxiliary network, which is to distinguish the information between latent code z and attribute vector y, is clean and clear.\nExperiments illustrate the advantage of using auxiliary network and demonstrating the role of classify. Experimental results also show the proposed model learning to factor attributes from identity on the face dataset.\n\nWeakness \nThe proposed model seem to be unnecessarily complex. For example, the loss of  in (6) actually includes 6 components (5 are from L_enc) and 4~5 tuning hyper-parameters. The L_gan also includes 3 parts. The reason of adding gan loss lacks either theoretical or empirical analysis. So as L_KL. In addition, the second term in L_gan is unnecessary since you already have a reconstruction loss. It also make it to be unclear what we obtain if the equilibrium of the GAN objective achieved.\n\nThe written of this paper can be improved to make it more clear. \nIt looks \\hat_y and \\tilde_y are same thing. \nHow do you get \\hat_z? Do you assume the posterior distribution is Gaussian and use the reparameterization trick? What are \\hat_y and \\hat_\\hat_y? Are they binary or a scalar between 0 and 1?  How do you generate \\hat_x? When generating \\hat_x, do you sample \\hat_z and \\hat_y? If so, how do treat the variance problem of \\hat_y? \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540967480189}], "openreview_url": "https://openreview.net/forum?id=BJfRpoA9YX", "arxiv_id": "1711.05175", "paper_pdf": "papers/BJfRpoA9YX.pdf", "paper_pdf_sha256": "e43c338d64b4ac679357d74af313f4938e2a2ee4229d90b82bb679e84b0fbb3f", "paper_pdf_bytes": 2942774, "paper_pdf_source": "openreview", "code_url": "https://github.com/ToniCreswell/attribute-cVAEGAN", "code_repository": "ToniCreswell/attribute-cVAEGAN", "code_commit": "b8d543a962e68ffb4f5153470c1778a15f5508e4", "code_archive": "repos/BJfRpoA9YX.zip", "code_archive_sha256": "bb39a5c192d789606a0eeb8c4b71902f2e29cf900ccf22a50138a3fb6bfb3c76", "code_archive_bytes": 498734, "code_file_count": 7, "code_extensions": {".py": 6, ".ipynb": 1}, "github_disk_usage_kb": 684, "github_languages": {"Jupyter Notebook": 253095, "Python": 38052}, "github_archived": false, "github_pushed_at": "2018-11-27T09:20:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-information-factorization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DBqOInhRkG", "year": 2026, "status": "rejected", "title": "RARE: Retrieval-Aware Robustness Evaluation for Retrieval-Augmented Generation Systems", "authors": ["Yixiao Zeng", "Tianyu Cao", "Danqing Wang", "Xinran Zhao", "Zimeng Qiu", "Morteza Ziyadi", "Tongshuang Wu", "Lei Li"], "authorids": ["~Yixiao_Zeng1", "~Tianyu_Cao2", "~Danqing_Wang1", "~Xinran_Zhao1", "~Zimeng_Qiu1", "~Morteza_Ziyadi2", "~Tongshuang_Wu1", "~Lei_Li11"], "authors_source": "OpenReview API", "abstract": "Retrieval-Augmented Generation (RAG) enhances recency and factuality in answers. However, existing evaluations rarely test how well these systems cope with real-world noise, conflicting between internal and external retrieved contexts, or fast-changing facts. \nWe introduce Retrieval-Aware Robustness Evaluation (RARE), a unified framework and large-scale benchmark that jointly stress-tests query and document perturbations over dynamic, time-sensitive corpora. One of the central features of RARE is a knowledge-graph-driven synthesis pipeline (RARE-Get) that automatically extracts single and multi-hop relations from the customized corpus and generates multi-level question sets without manual intervention. Leveraging this pipeline, we construct a dataset (RARE-Set) spanning 527 expert-level time-sensitive finance, economics, and policy documents and 48295 questions whose distribution evolves as the underlying sources change. To quantify resilience, we formalize retrieval-conditioned robustness metrics (RARE-Met) that capture a model’s ability to remain correct or recover when queries, documents, or real-world retrieval results are systematically altered. Our findings reveal that RAG systems are unexpectedly sensitive to perturbations. Moreover, they consistently demonstrate lower robustness on multi-hop queries compared to single-hop queries across all domains.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "xBZLZaxCKf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21066/Reviewer_nXfs"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper focuses on evaluating existing Retrieval-Augmented Generation (RAG) systems. It first identifies the limitations of current evaluation frameworks and then introduces a novel approach for automatically generating complex evaluation data. Specifically, the paper presents RARE-Get, a synthesis pipeline that constructs time-sensitive evaluation data via knowledge graphs; RARE-Set, a multi-domain dataset created using this pipeline; and RARE-Met, an evaluation metric designed to assess RAG performance under query and document perturbations. Experimental results demonstrate that current RAG systems exhibit fragility under certain perturbations and that robustness does not always increase proportionally with model size.", "review_text": "This paper focuses on evaluating existing Retrieval-Augmented Generation (RAG) systems. It first identifies the limitations of current evaluation frameworks and then introduces a novel approach for automatically generating complex evaluation data. Specifically, the paper presents RARE-Get, a synthesis pipeline that constructs time-sensitive evaluation data via knowledge graphs; RARE-Set, a multi-domain dataset created using this pipeline; and RARE-Met, an evaluation metric designed to assess RAG performance under query and document perturbations. Experimental results demonstrate that current RAG systems exhibit fragility under certain perturbations and that robustness does not always increase proportionally with model size.", "strengths": "- This paper identifies key limitations in existing RAG evaluation datasets and introduces a dynamic data generation pipeline that leverages knowledge graphs to automatically construct both single-hop and multi-hop questions. The proposed pipeline has strong potential value for the community, as it can generate diverse and complex datasets from unstructured data. In addition, the proposed dataset **RARE-Set** and the robustness evaluation metric contribute useful resources for assessing RAG systems.\n- The paper is well-written, and the analysis is thorough, supported by extensive experiments.", "weaknesses": "Although the paper includes experiments with various LLMs, the baseline methods are limited to standard RAG setups. Recent state-of-the-art RAG variants, such as adaptive or noise-robust models, are not included, which makes it difficult to fully assess the effectiveness of the proposed benchmark and evaluation metric. Incorporating results from these stronger baselines would significantly enhance the rigor and credibility of the evaluation.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on evaluating existing Retrieval-Augmented Generation (RAG) systems. It first identifies the limitations of current evaluation frameworks and then introduces a novel approach for automatically generating complex evaluation data. Specifically, the paper presents RARE-Get, a synthesis pipeline that constructs time-sensitive evaluation data via knowledge graphs; RARE-Set, a multi-domain dataset created using this pipeline; and RARE-Met, an evaluation metric designed to assess RAG performance under query and document perturbations. Experimental results demonstrate that current RAG systems exhibit fragility under certain perturbations and that robustness does not always increase proportionally with model size.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This paper identifies key limitations in existing RAG evaluation datasets and introduces a dynamic data generation pipeline that leverages knowledge graphs to automatically construct both single-hop and multi-hop questions. The proposed pipeline has strong potential value for the community, as it can generate diverse and complex datasets from unstructured data. In addition, the proposed dataset **RARE-Set** and the robustness evaluation metric contribute useful resources for assessing RAG systems.\n- The paper is well-written, and the analysis is thorough, supported by extensive experiments.", "weaknesses": "Although the paper includes experiments with various LLMs, the baseline methods are limited to standard RAG setups. Recent state-of-the-art RAG variants, such as adaptive or noise-robust models, are not included, which makes it difficult to fully assess the effectiveness of the proposed benchmark and evaluation metric. Incorporating results from these stronger baselines would significantly enhance the rigor and credibility of the evaluation.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761987629265}, {"id": "Jk1m1O2Q90", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21066/Reviewer_gD5y"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces a framework for evaluating RAG system robustness under realistic perturbations. RARE addresses critical gaps in existing benchmarks by simultaneously examining three key dimensions: dynamics (handling changing information), query complexity (single-hop to multi-hop reasoning), and robustness measurement (quantifying performance degradation). The framework consists of RARE-Get (automated KG-driven synthesis pipeline), RARE-Set (questions spanning multiple documents and complexity levels), and RARE-Met (retrieval-conditioned robustness metrics). Evaluation reveals fragility in current RAG systems.", "review_text": "This paper introduces a framework for evaluating RAG system robustness under realistic perturbations. RARE addresses critical gaps in existing benchmarks by simultaneously examining three key dimensions: dynamics (handling changing information), query complexity (single-hop to multi-hop reasoning), and robustness measurement (quantifying performance degradation). The framework consists of RARE-Get (automated KG-driven synthesis pipeline), RARE-Set (questions spanning multiple documents and complexity levels), and RARE-Met (retrieval-conditioned robustness metrics). Evaluation reveals fragility in current RAG systems.", "strengths": "- The proposed method is highly scalable: RARE-Get uses KG-driven synthesis to automatically generate multi-hop questions without manual curation, focusing on specialized, time-sensitive corpora appropriate for real-world RAG applications.\n- This paper proposes well-defined metrics: RARE-Met distinguishes memorization from retrieval-based reasoning and tests robustness across query perturbations (typos, paraphrasing) and document perturbations (lexical/answer variations). Moreover, its evaluation presents useful insights: model size doesn't guarantee robustness, multi-hop queries underperform single-hop, and domain factors significantly impact performance.", "weaknesses": "- The paper does not analyze error propagation which harms its reliability. RARE-Get chains multiple LLMs (GPT-4.1 for extraction and generation, Claude variants for filtering and evaluation), where errors compound at each stage. The paper omits failure rates, data discarded during quality checks, and how extraction errors corrupt downstream question generation.\n- The paper assumes generated questions require multi-hop reasoning without validation—questions may be answerable through single-chunk retrieval despite their structure. The robustness definition is debatable: it penalizes models that follow perturbed context, but robust systems could instead detect and reject unreliable retrieved information, yielding opposite interpretations of the same behavior.", "questions": "Could you comment on the failure rate or amount of data discarded during the LLM-based quality assurance step? What percentage of extracted triplets are factually correct? For instance, sample 100-200 triplets and verify against source documents can give a better sense of the raw quality of initial LLM-based generation and how errors propagate through the pipeline.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a framework for evaluating RAG system robustness under realistic perturbations. RARE addresses critical gaps in existing benchmarks by simultaneously examining three key dimensions: dynamics (handling changing information), query complexity (single-hop to multi-hop reasoning), and robustness measurement (quantifying performance degradation). The framework consists of RARE-Get (automated KG-driven synthesis pipeline), RARE-Set (questions spanning multiple documents and complexity levels), and RARE-Met (retrieval-conditioned robustness metrics). Evaluation reveals fragility in current RAG systems.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- The proposed method is highly scalable: RARE-Get uses KG-driven synthesis to automatically generate multi-hop questions without manual curation, focusing on specialized, time-sensitive corpora appropriate for real-world RAG applications.\n- This paper proposes well-defined metrics: RARE-Met distinguishes memorization from retrieval-based reasoning and tests robustness across query perturbations (typos, paraphrasing) and document perturbations (lexical/answer variations). Moreover, its evaluation presents useful insights: model size doesn't guarantee robustness, multi-hop queries underperform single-hop, and domain factors significantly impact performance.", "weaknesses": "- The paper does not analyze error propagation which harms its reliability. RARE-Get chains multiple LLMs (GPT-4.1 for extraction and generation, Claude variants for filtering and evaluation), where errors compound at each stage. The paper omits failure rates, data discarded during quality checks, and how extraction errors corrupt downstream question generation.\n- The paper assumes generated questions require multi-hop reasoning without validation—questions may be answerable through single-chunk retrieval despite their structure. The robustness definition is debatable: it penalizes models that follow perturbed context, but robust systems could instead detect and reject unreliable retrieved information, yielding opposite interpretations of the same behavior.", "questions": "Could you comment on the failure rate or amount of data discarded during the LLM-based quality assurance step? What percentage of extracted triplets are factually correct? For instance, sample 100-200 triplets and verify against source documents can give a better sense of the raw quality of initial LLM-based generation and how errors propagate through the pipeline.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761967076063}, {"id": "hHS4rCyBHa", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21066/Reviewer_a38J"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces Retrieval-Aware Robustness Evaluation (RARE), a comprehensive framework designed for evaluating retrieval-augmented generation systems. Without requiring manual curation, RARE automatically constructs time-sensitive benchmarks via knowledge graph extraction and traversal, and systematically assesses robustness under query, document, and real-world retrieval perturbations. Extensive experiments across financial, economic, and policy domains with over 48,000 queries demonstrate that RARE effectively reveals critical limitations in RAG systems, particularly in multi-hop and domain-specific scenarios.", "review_text": "This paper introduces Retrieval-Aware Robustness Evaluation (RARE), a comprehensive framework designed for evaluating retrieval-augmented generation systems. Without requiring manual curation, RARE automatically constructs time-sensitive benchmarks via knowledge graph extraction and traversal, and systematically assesses robustness under query, document, and real-world retrieval perturbations. Extensive experiments across financial, economic, and policy domains with over 48,000 queries demonstrate that RARE effectively reveals critical limitations in RAG systems, particularly in multi-hop and domain-specific scenarios.", "strengths": "1. The authors introduce Retrieval-Aware Robustness Evaluation (RARE), a comprehensive framework designed for evaluating retrieval-augmented generation systems.\n\n2. Extensive experiments across financial, economic, and policy domains with over 48,000 queries demonstrate that RARE effectively reveals critical limitations in RAG systems, particularly in multi-hop and domain-specific scenarios.\n\n3. The analysis dimensions are thorough and insightful.  The paper not only presents overall performance comparisons but also compares robustness under query perturbations, document perturbations, and real retrieval scenarios.", "weaknesses": "1.  Insufficient technical details in the knowledge graph construction process: The paper does not provide enough description of the core steps of knowledge graph construction, particularly the specific implementation of relation normalization (detailed process using E5-Mistral-7B-Instruct), the quality control mechanisms for handling extracted conflicting or incorrect triples, and how to manage entity alignment and conflict resolution during cross-document knowledge graph merging.\n\n2.  The comparison across unique dimensions is not clearly explained.  In terms of time-sensitivity (time-sens), multi-hop questions (MH), and cross-document retrieval (Dynamic), the improvements are not significantly different from existing datasets, and the innovation in these dimensions is insufficient.\n\n3.  Limited coverage of perturbation types: The current perturbation types mainly focus on surface-level and semantic-level changes but lack factual perturbations (e.g., inserting contradictory information into documents), structural perturbations (e.g., changing document order or format), and handling of multimodal documents (e.g., cases where tables and text are mixed in images).\n\n4.  Practical guidance could be strengthened: The experiments identified several important phenomena (e.g., the vulnerability of multi-hop queries), but there is a lack of in-depth analysis of the underlying causes of these phenomena and no specific system improvement suggestions based on the findings.", "questions": "1. The core innovation of the paper relies on large models like GPT-4 for knowledge graph triple extraction and QA pair generation.      Considering the large scale of the benchmark datasets, what are the specific computational costs and time overheads of this automated process?      Have more cost-effective alternatives been explored?      Providing estimates and discussion on this aspect would help the community replicate and extend this work.\n\n2. On the realism and intensity balance of perturbations: In document perturbations, \"lexically similar but answerless\" is achieved by directly deleting the answer sentence, which may be too extreme and easy to identify in real retrieval scenarios.      Have more natural and subtle perturbation methods been considered?      Additionally, how can it be ensured that different types of perturbations have roughly comparable \"interference intensity\" to avoid overly harsh or lenient evaluation of certain perturbation types?\n\n4. On the fairness of model comparison and prompt engineering: Experiments compared various open-source and closed-source models, but different models may have different sensitivities and optimization needs for prompts.      Were the unified prompts used in the paper sufficiently optimized or validated for each evaluated model?      If not specifically optimized, could this potentially lead to an underestimation of some models’ performance?\n\n5. On the phenomenon that \"model robustness does not strictly increase with scale\": The paper observes an interesting phenomenon where Qwen3-32B shows lower robustness than the smaller Qwen3-8B and 4B models, attributing it to \"architecture design and training methods.\"      This explanation could be more specific and detailed.      Can more nuanced hypotheses be proposed?\n\n6. On the basis for quality assurance threshold selection and data bias analysis: In the quality assurance stage, Claude 3.5 was used to score generated QA pairs, with a retention threshold set at \"all dimension scores above 6.\"      What considerations were behind this specific threshold?      Can the score distribution of filtered-out QA pairs be provided, analyzing which dimensions mainly had deficiencies?      This would help readers understand the quality boundaries of the dataset and potential biases.\n\nIf the authors can address all of the concerns, I am willing to consider raising my rating.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Retrieval-Aware Robustness Evaluation (RARE), a comprehensive framework designed for evaluating retrieval-augmented generation systems. Without requiring manual curation, RARE automatically constructs time-sensitive benchmarks via knowledge graph extraction and traversal, and systematically assesses robustness under query, document, and real-world retrieval perturbations. Extensive experiments across financial, economic, and policy domains with over 48,000 queries demonstrate that RARE effectively reveals critical limitations in RAG systems, particularly in multi-hop and domain-specific scenarios.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The authors introduce Retrieval-Aware Robustness Evaluation (RARE), a comprehensive framework designed for evaluating retrieval-augmented generation systems.\n\n2. Extensive experiments across financial, economic, and policy domains with over 48,000 queries demonstrate that RARE effectively reveals critical limitations in RAG systems, particularly in multi-hop and domain-specific scenarios.\n\n3. The analysis dimensions are thorough and insightful.  The paper not only presents overall performance comparisons but also compares robustness under query perturbations, document perturbations, and real retrieval scenarios.", "weaknesses": "1.  Insufficient technical details in the knowledge graph construction process: The paper does not provide enough description of the core steps of knowledge graph construction, particularly the specific implementation of relation normalization (detailed process using E5-Mistral-7B-Instruct), the quality control mechanisms for handling extracted conflicting or incorrect triples, and how to manage entity alignment and conflict resolution during cross-document knowledge graph merging.\n\n2.  The comparison across unique dimensions is not clearly explained.  In terms of time-sensitivity (time-sens), multi-hop questions (MH), and cross-document retrieval (Dynamic), the improvements are not significantly different from existing datasets, and the innovation in these dimensions is insufficient.\n\n3.  Limited coverage of perturbation types: The current perturbation types mainly focus on surface-level and semantic-level changes but lack factual perturbations (e.g., inserting contradictory information into documents), structural perturbations (e.g., changing document order or format), and handling of multimodal documents (e.g., cases where tables and text are mixed in images).\n\n4.  Practical guidance could be strengthened: The experiments identified several important phenomena (e.g., the vulnerability of multi-hop queries), but there is a lack of in-depth analysis of the underlying causes of these phenomena and no specific system improvement suggestions based on the findings.", "questions": "1. The core innovation of the paper relies on large models like GPT-4 for knowledge graph triple extraction and QA pair generation.      Considering the large scale of the benchmark datasets, what are the specific computational costs and time overheads of this automated process?      Have more cost-effective alternatives been explored?      Providing estimates and discussion on this aspect would help the community replicate and extend this work.\n\n2. On the realism and intensity balance of perturbations: In document perturbations, \"lexically similar but answerless\" is achieved by directly deleting the answer sentence, which may be too extreme and easy to identify in real retrieval scenarios.      Have more natural and subtle perturbation methods been considered?      Additionally, how can it be ensured that different types of perturbations have roughly comparable \"interference intensity\" to avoid overly harsh or lenient evaluation of certain perturbation types?\n\n4. On the fairness of model comparison and prompt engineering: Experiments compared various open-source and closed-source models, but different models may have different sensitivities and optimization needs for prompts.      Were the unified prompts used in the paper sufficiently optimized or validated for each evaluated model?      If not specifically optimized, could this potentially lead to an underestimation of some models’ performance?\n\n5. On the phenomenon that \"model robustness does not strictly increase with scale\": The paper observes an interesting phenomenon where Qwen3-32B shows lower robustness than the smaller Qwen3-8B and 4B models, attributing it to \"architecture design and training methods.\"      This explanation could be more specific and detailed.      Can more nuanced hypotheses be proposed?\n\n6. On the basis for quality assurance threshold selection and data bias analysis: In the quality assurance stage, Claude 3.5 was used to score generated QA pairs, with a retention threshold set at \"all dimension scores above 6.\"      What considerations were behind this specific threshold?      Can the score distribution of filtered-out QA pairs be provided, analyzing which dimensions mainly had deficiencies?      This would help readers understand the quality boundaries of the dataset and potential biases.\n\nIf the authors can address all of the concerns, I am willing to consider raising my rating.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761906070564}], "openreview_url": "https://openreview.net/forum?id=DBqOInhRkG", "arxiv_id": "2506.00789", "paper_pdf": "papers/DBqOInhRkG.pdf", "paper_pdf_sha256": "ff00589f9b45328c693d4e8a46e357683706ea6b24ed698a91a8b47cb1964591", "paper_pdf_bytes": 848980, "paper_pdf_source": "openreview", "code_url": "https://github.com/LeiLiLab/RARE", "code_repository": "LeiLiLab/RARE", "code_commit": "2496f95c695eb23dcb197191e72c692f14d53781", "code_archive": "repos/DBqOInhRkG.zip", "code_archive_sha256": "a3744cdb1785c5639a609d6613173aa0e57e5c4913c3445d92ac37a70ffc6932", "code_archive_bytes": 368880, "code_file_count": 21, "code_extensions": {".py": 19, ".ipynb": 2}, "github_disk_usage_kb": 441, "github_languages": {"Jupyter Notebook": 4133839, "Python": 362344}, "github_archived": false, "github_pushed_at": "2026-05-03T06:25:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rare-retrieval-aware-robustness-evaluation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "F9JZiGradI", "year": 2025, "status": "rejected", "title": "MLP-KAN: Unifying Deep Representation and Function Learning", "authors": ["Yunhong He", "Zhengqing Yuan", "Yifeng Xie", "Lichao Sun"], "authorids": ["~Yunhong_He1", "~Zhengqing_Yuan2", "~Yifeng_Xie1", "~Lichao_Sun1"], "authors_source": "OpenReview API", "abstract": "Recent advancements in both representation learning and function learning have demonstrated substantial promise across diverse domains of artificial intelligence. However, the effective integration of these paradigms poses a significant challenge, particularly in cases where users must manually decide whether to apply a representation learning or function learning model based on dataset characteristics. To address this issue, we introduce MLP-KAN, a unified method designed to eliminate the need for manual model selection. By integrating Multi-Layer Perceptrons (MLPs) for representation learning and Kolmogorov-Arnold Networks (KANs) for function learning within a Mixture-of-Experts (MoE) architecture, MLP-KAN dynamically adapts to the specific characteristics of the task at hand, ensuring optimal performance. Embedded within a transformer-based framework, our work achieves remarkable results on four widely-used datasets across diverse domains. Extensive experimental evaluation demonstrates its superior versatility, delivering competitive performance across both deep representation and function learning tasks. These findings highlight the potential of MLP-KAN to simplify the model selection process, offering a comprehensive, adaptable solution across various domains.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "56Ylcm7zin", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4126/Reviewer_YbF1"], "rating": 1, "soundness": 1, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The authors hypothesize that KAN networks and MLPs are effective for solving different types of problems: specifically, MLPs are good for representation learning, while KANs are good for function learning.  They propose a modeling strategy that includes both MLPs and KANs, which are adaptively selected based upon the problem setting.", "review_text": "The authors hypothesize that KAN networks and MLPs are effective for solving different types of problems: specifically, MLPs are good for representation learning, while KANs are good for function learning.  They propose a modeling strategy that includes both MLPs and KANs, which are adaptively selected based upon the problem setting.", "strengths": "1) The premise of the paper is potentially reasonable; there is evidence in the literature that KANs and MLPs are complementary, and developing a method that adaptively selects which modeling strategy is best for a given problem is a good idea.", "weaknesses": "1) The presentation of the paper is not sufficiently clear to facilitate review.   The paper is generally vague, contains many typos, and the methodology is ultimately unclear.  I provide examples\n\n(ii) Poor presentation.  In Line 47 in the 2nd paragraph of the paper, the authors define KAN as Kernel Attention Network, yet the whole paper seems to be about Kolmogorov-Arnold Network.  Line 51 and 73 in the introduction are nearly identical, and repeat the same idea. \n\n(ii) Incorrect/unclear methodological description.  For example, in Eq. (5) the dimensions of W and X are incompatible.  The entire architecture of the proposed model is unclear.  For example, it is unclear whether there are multiple MLP-KAN layers?  The authors have a section about \"Architecture\" which then, without motivation, discusses self attention.   The number of layers in the MLP-KAN are never provided (although it is unclear if there are multiple layers), nor is the overall size of the model discussed. \n\n(iii) Vague motivation.  The whole premise of the paper is not clearly explained.  While the idea of combining KANs and MLPs seems reasonable, the authors repeatedly argue a theoretical motive based upon MLPs being \"representation learning\" methods while KANs are \"function learning\".  This premise for the proposed approach is repeatedly mentioned throughout the paper, yet the difference between these two approaches is never precisely described.  \n \n2) Insufficient Experiments.  The experiments are insufficient to demonstrate the efficacy of the proposed approach.  To the best I can discern, the proposed method would significantly increase the number of modeling parameters because we now have multiple KANs and MLPs in a single model, along with some parameters to select among them.  However, the resulting model often performs similarly to other models only composed of a single KAN or MLP architecture (e.g., Table 3).  What would happen if we simply used a single MLP or single KAN model that is the same size (in terms of free parameters) to the MLP-KAN?  Or what if we made a simple fusion model where we interlaced MLP and KAN layers, or added a few KAN layers to the end of a standard MLP?  How do we know whether this more complex architecture proposed by the authors is superior to simpler and/or smaller models?", "questions": "I think the paper is insufficiently clear to support proper review, and therefore I don't have any questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors hypothesize that KAN networks and MLPs are effective for solving different types of problems: specifically, MLPs are good for representation learning, while KANs are good for function learning.  They propose a modeling strategy that includes both MLPs and KANs, which are adaptively selected based upon the problem setting.", "soundness": 1, "presentation": 1, "contribution": 2, "strengths": "1) The premise of the paper is potentially reasonable; there is evidence in the literature that KANs and MLPs are complementary, and developing a method that adaptively selects which modeling strategy is best for a given problem is a good idea.", "weaknesses": "1) The presentation of the paper is not sufficiently clear to facilitate review.   The paper is generally vague, contains many typos, and the methodology is ultimately unclear.  I provide examples\n\n(ii) Poor presentation.  In Line 47 in the 2nd paragraph of the paper, the authors define KAN as Kernel Attention Network, yet the whole paper seems to be about Kolmogorov-Arnold Network.  Line 51 and 73 in the introduction are nearly identical, and repeat the same idea. \n\n(ii) Incorrect/unclear methodological description.  For example, in Eq. (5) the dimensions of W and X are incompatible.  The entire architecture of the proposed model is unclear.  For example, it is unclear whether there are multiple MLP-KAN layers?  The authors have a section about \"Architecture\" which then, without motivation, discusses self attention.   The number of layers in the MLP-KAN are never provided (although it is unclear if there are multiple layers), nor is the overall size of the model discussed. \n\n(iii) Vague motivation.  The whole premise of the paper is not clearly explained.  While the idea of combining KANs and MLPs seems reasonable, the authors repeatedly argue a theoretical motive based upon MLPs being \"representation learning\" methods while KANs are \"function learning\".  This premise for the proposed approach is repeatedly mentioned throughout the paper, yet the difference between these two approaches is never precisely described.  \n \n2) Insufficient Experiments.  The experiments are insufficient to demonstrate the efficacy of the proposed approach.  To the best I can discern, the proposed method would significantly increase the number of modeling parameters because we now have multiple KANs and MLPs in a single model, along with some parameters to select among them.  However, the resulting model often performs similarly to other models only composed of a single KAN or MLP architecture (e.g., Table 3).  What would happen if we simply used a single MLP or single KAN model that is the same size (in terms of free parameters) to the MLP-KAN?  Or what if we made a simple fusion model where we interlaced MLP and KAN layers, or added a few KAN layers to the end of a standard MLP?  How do we know whether this more complex architecture proposed by the authors is superior to simpler and/or smaller models?", "questions": "I think the paper is insufficiently clear to support proper review, and therefore I don't have any questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731183356478}, {"id": "4QnDMfaso2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4126/Reviewer_ZQmM"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 4, "confidence": 4, "summary": "This manuscript introduces MLP-KAN, a unified block that combines representation and function learning within a single framework. By using MLPs for representation learning and KANs for function learning in a mixture-of-experts setup, MLP-KAN adapts to various tasks and data modalities. When integrated into a transformer-based architecture, MLP-KAN demonstrates both versatility and robustness across multiple data domains. The proposed method is evaluated on four datasets, CIFAR-10, CIFAR-100, mini-ImageNet, and SST2, showing strong performance in both image classification and natural language processing tasks.\n\n############################ Post Rebuttal ############################\n\nAll of my concerns have been addressed during rebuttal. I am happy to raise my score from 6 to 8.\n\n############################ Post Rebuttal ############################", "review_text": "This manuscript introduces MLP-KAN, a unified block that combines representation and function learning within a single framework. By using MLPs for representation learning and KANs for function learning in a mixture-of-experts setup, MLP-KAN adapts to various tasks and data modalities. When integrated into a transformer-based architecture, MLP-KAN demonstrates both versatility and robustness across multiple data domains. The proposed method is evaluated on four datasets, CIFAR-10, CIFAR-100, mini-ImageNet, and SST2, showing strong performance in both image classification and natural language processing tasks.\n\n############################ Post Rebuttal ############################\n\nAll of my concerns have been addressed during rebuttal. I am happy to raise my score from 6 to 8.\n\n############################ Post Rebuttal ############################", "strengths": "1. The manuscript is well-written, presenting clear motivations and providing step-by-step derivations of the proposed method.\n\n2. Combining MLPs and KANs within an MoE framework is interesting. Moreover, integrating this block into a transformer architecture develops a robust backbone that effectively extracts and integrates features across various data modalities.\n\n3. The ablation studies demonstrate that the proposed method can scale easily by increasing the number of experts, which enhances performance without introducing too much computational burdens.\n\n4. The proposed method is extensively evaluated on multiple CV and NLP datasets, demonstrating its versatility for diverse AI applications and practical values.", "weaknesses": "1. Although the proposed technique is shown to be generalizable to different tasks, its effectiveness in other types of tasks or with different types of data (e.g., time-series, reinforcement learning) remains unexplored.\n\n2. Using multiple experts in an MoE architecture, especially with higher Top-K values, can significantly increase computational resource requirements. I suggest that the authors conduct ablation studies on runtime complexity and compare their proposed method with the standard transformer architecture.", "questions": "1. How does MLP-KAN perform in the presence of noisy or adversarial inputs compared to other models? Are there any robustness benchmarks included in the evaluation?\n\n2. What optimization algorithms and strategies were employed to train MLP-KAN effectively? Were any specific techniques used to balance the training of multiple experts?\n\n3. Were all models, including baselines, trained under the same conditions to ensure fair comparisons? What datasets splits, augmentation techniques, and training epochs were used?\n\n4. How sensitive is MLP-KAN to changes in hyperparameters other than the number of experts and Top-K values? For example, how do variations in learning rates, network depth, or activation functions affect performance?\n\n5. Were there any challenges related to training stability when combining MLPs and KANs within the MoE framework? How were these challenges addressed?\n\n6. How does the inclusion of MLP-KAN affect the standard attention mechanisms within transformers? Are there any changes to how attention weights are applied?\n\n7. How effective is MLP-KAN in transfer learning where it is fine-tuned on different tasks after initial source pre-training?\n\n8. What is the computational cost of MLP-KAN compared to other architecture with MLPs (e.g., transfomer) or KANs alone? How does the addition of multiple experts affect training and inference times?\n\n9. How interpretable are the latent features generated by MLP-KAN? Are there any visualizations demonstrating the semantic captured by the model (e.g., t-SNE visualizations)?\n\n10. What future research directions do the authors suggest to address the current limitations or to further enhance the capabilities of MLP-KAN?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript introduces MLP-KAN, a unified block that combines representation and function learning within a single framework. By using MLPs for representation learning and KANs for function learning in a mixture-of-experts setup, MLP-KAN adapts to various tasks and data modalities. When integrated into a transformer-based architecture, MLP-KAN demonstrates both versatility and robustness across multiple data domains. The proposed method is evaluated on four datasets, CIFAR-10, CIFAR-100, mini-ImageNet, and SST2, showing strong performance in both image classification and natural language processing tasks.\n\n############################ Post Rebuttal ############################\n\nAll of my concerns have been addressed during rebuttal. I am happy to raise my score from 6 to 8.\n\n############################ Post Rebuttal ############################", "soundness": 4, "presentation": 4, "contribution": 4, "strengths": "1. The manuscript is well-written, presenting clear motivations and providing step-by-step derivations of the proposed method.\n\n2. Combining MLPs and KANs within an MoE framework is interesting. Moreover, integrating this block into a transformer architecture develops a robust backbone that effectively extracts and integrates features across various data modalities.\n\n3. The ablation studies demonstrate that the proposed method can scale easily by increasing the number of experts, which enhances performance without introducing too much computational burdens.\n\n4. The proposed method is extensively evaluated on multiple CV and NLP datasets, demonstrating its versatility for diverse AI applications and practical values.", "weaknesses": "1. Although the proposed technique is shown to be generalizable to different tasks, its effectiveness in other types of tasks or with different types of data (e.g., time-series, reinforcement learning) remains unexplored.\n\n2. Using multiple experts in an MoE architecture, especially with higher Top-K values, can significantly increase computational resource requirements. I suggest that the authors conduct ablation studies on runtime complexity and compare their proposed method with the standard transformer architecture.", "questions": "1. How does MLP-KAN perform in the presence of noisy or adversarial inputs compared to other models? Are there any robustness benchmarks included in the evaluation?\n\n2. What optimization algorithms and strategies were employed to train MLP-KAN effectively? Were any specific techniques used to balance the training of multiple experts?\n\n3. Were all models, including baselines, trained under the same conditions to ensure fair comparisons? What datasets splits, augmentation techniques, and training epochs were used?\n\n4. How sensitive is MLP-KAN to changes in hyperparameters other than the number of experts and Top-K values? For example, how do variations in learning rates, network depth, or activation functions affect performance?\n\n5. Were there any challenges related to training stability when combining MLPs and KANs within the MoE framework? How were these challenges addressed?\n\n6. How does the inclusion of MLP-KAN affect the standard attention mechanisms within transformers? Are there any changes to how attention weights are applied?\n\n7. How effective is MLP-KAN in transfer learning where it is fine-tuned on different tasks after initial source pre-training?\n\n8. What is the computational cost of MLP-KAN compared to other architecture with MLPs (e.g., transfomer) or KANs alone? How does the addition of multiple experts affect training and inference times?\n\n9. How interpretable are the latent features generated by MLP-KAN? Are there any visualizations demonstrating the semantic captured by the model (e.g., t-SNE visualizations)?\n\n10. What future research directions do the authors suggest to address the current limitations or to further enhance the capabilities of MLP-KAN?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730684596414}, {"id": "pnI2E92eOy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4126/Reviewer_HLoC"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes MLP-KAN, a mixture of experts method, to combine MLP and KAN. The authors claim their method can address the shortcomings of both KAN and MLP in one structure and solve the scalability problem of MLPs. Furthermore, they show the superior performance of their method across different tasks and datasets compared to other methods and investigate the effectiveness of various components of their method.", "review_text": "This paper proposes MLP-KAN, a mixture of experts method, to combine MLP and KAN. The authors claim their method can address the shortcomings of both KAN and MLP in one structure and solve the scalability problem of MLPs. Furthermore, they show the superior performance of their method across different tasks and datasets compared to other methods and investigate the effectiveness of various components of their method.", "strengths": "* The paper is well-written for the most part.\n* The motivation of the paper is valid and exciting. \n* The idea is simple but yet effective.", "weaknesses": "* The discussion part of the paper heavily relies on the description of Multi-expert and router gating, which is not the paper's contribution. More discussion or experiments are needed to show why combining MLP and KAN should improve the performance. \n* One of the paper's main points is scalability, but experiments about scalability, computation time, and memory are missing. \n\nMinor Weakness:\n\n* The numbers in all Tables don't have confidence intervals, so it is hard to grasp the significance of the differences.  The authors should include confidence intervals or standard deviations from multiple runs.\n* Details about competitors are missing in the experimental setup.", "questions": "* Is the number of tasks connected to optimal k for topk?\n* Is the number of experts for MLP and parameters the same across experiments with MLP-KAN?\n* Are all experiments for Tables 2 and 3 trained together as a multi-task scenario? Or are experiments for Tables 2 and 3 separated?\n\nSuggestion:\n\n* It would be great to add more details about experiments and summaries in section 4.1.  \n* I think it improves the justifiability of the paper if they provide an ablation on the router gating to show how it assigns tokens to each expert.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes MLP-KAN, a mixture of experts method, to combine MLP and KAN. The authors claim their method can address the shortcomings of both KAN and MLP in one structure and solve the scalability problem of MLPs. Furthermore, they show the superior performance of their method across different tasks and datasets compared to other methods and investigate the effectiveness of various components of their method.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* The paper is well-written for the most part.\n* The motivation of the paper is valid and exciting. \n* The idea is simple but yet effective.", "weaknesses": "* The discussion part of the paper heavily relies on the description of Multi-expert and router gating, which is not the paper's contribution. More discussion or experiments are needed to show why combining MLP and KAN should improve the performance. \n* One of the paper's main points is scalability, but experiments about scalability, computation time, and memory are missing. \n\nMinor Weakness:\n\n* The numbers in all Tables don't have confidence intervals, so it is hard to grasp the significance of the differences.  The authors should include confidence intervals or standard deviations from multiple runs.\n* Details about competitors are missing in the experimental setup.", "questions": "* Is the number of tasks connected to optimal k for topk?\n* Is the number of experts for MLP and parameters the same across experiments with MLP-KAN?\n* Are all experiments for Tables 2 and 3 trained together as a multi-task scenario? Or are experiments for Tables 2 and 3 separated?\n\nSuggestion:\n\n* It would be great to add more details about experiments and summaries in section 4.1.  \n* I think it improves the justifiability of the paper if they provide an ablation on the router gating to show how it assigns tokens to each expert.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730581899265}, {"id": "I8PDSmSMOv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4126/Reviewer_FkFz"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a method, KAN-MLP, for improving the performance of ViT on function learning and CIFAR classification tasks. In order to achieve that, they replaced the traditional MLP layers within the transformer architecture. In this method, experts are dynamically selected based on the input through a gating mechanism, ensuring efficient routing of tokens to the most relevant experts. Their main contribution is applying an explainable KAN architecture to the Vision Transformer model. Comparing KAN, MLP, and MLP-KAN on function learning tasks, they show that MLP-KAN performs better than the other architectures in certain cases. Additionally, they also perform an ablation study on the CIFAR dataset to demonstrate that their extended model outperforms the naive ViT.", "review_text": "This paper proposes a method, KAN-MLP, for improving the performance of ViT on function learning and CIFAR classification tasks. In order to achieve that, they replaced the traditional MLP layers within the transformer architecture. In this method, experts are dynamically selected based on the input through a gating mechanism, ensuring efficient routing of tokens to the most relevant experts. Their main contribution is applying an explainable KAN architecture to the Vision Transformer model. Comparing KAN, MLP, and MLP-KAN on function learning tasks, they show that MLP-KAN performs better than the other architectures in certain cases. Additionally, they also perform an ablation study on the CIFAR dataset to demonstrate that their extended model outperforms the naive ViT.", "strengths": "1. The paper proposes KAN-MLP, a novel approach to enhance Vision Transformers (ViT) for function learning and CIFAR classification tasks.\n\n2. It introduces the explainable KAN architecture into the ViT model, which could improve interoperability.\n\n3. The paper compares KAN, MLP, and MLP-KAN in function learning, showing potential advantages of MLP-KAN over other architectures.\n\n4. An ablation study on CIFAR demonstrates that the proposed model improves on the plain ViT model, highlighting the effectiveness of the enhancements.", "weaknesses": "1. The scaling law for MLP-KAN is missing, making it difficult to assess if MLP-KAN overcomes KAN’s limitations.\n\n2. The method may still suffer from the curse of dimensionality (COD), and the paper does not address how MLP-KAN performs for learning non-smooth and high-dimensional functions.\n\n3. There is a potential conflict between KAN’s suitability for low-dimensional tasks and ViT’s insuitability for small datasets. The experiments are limited to smooth functions and CIFAR datasets. Testing on larger datasets like ImageNet or MSCOCO would provide a more comprehensive view of the model’s performance relative to the Vision Transformer baseline.", "questions": "This paper contributes to advancing the KAN model, though it requires some clarifications on theory and experiments. Given these clarifications, if provided in an author's response, I would consider increasing the score. \n\nFor the theory, there are a few steps that need clarification and further clarification on novelty. \n\n1. KAN has advantages in model scaling under certain constraint conditions but not for all learning tasks. The KAN model obeys the Error Scaling formula $\\\\| f - (KAN) \\\\|\\_{C^m} \\le C G^{-k-1+m}$ and the scaling law $l \\\\propto N^{-\\\\alpha}$, but the author did not clarify the relationship between grid size $G$ and the input space dimension $n$. As a function approximation method, usually, the number of grid points $G$ is directly proportional to $G \\propto I^n$ ($I+1$ is the number of intervals on each dimension), which makes $\\\\| f - (KAN) \\\\|_{C^m} \\le C G^{(-k-1+m)/n}$ and triggers a serious curse of dimensionality (COD) problem (see KAN [1], Fig. 3.1, $f(x_1, \\cdots, x\\_{100}) = \\text{exp}(\\frac{1}{100} \\sum\\_{i=1}^{100} \\sin^2{\\frac{\\pi x_i}{2}})$). To address this problem, KAN authors hypothesize that the objective high-dimensional function is smooth and has sparse compositional structure to reduce the number of grid nodes $G \\ll I^n$. The authors did not provide the scaling law of MLP-KAN in the paper. Therefore, I don't know whether MLP-KAN overcomes the inherent limitations of KAN.\n\n2. This is particularly called into question due to the integration of KAN and ViT, since KAN and ViT usually exhibit different behavior on datasets of varying sizes. At present, KAN is suitable for low-dimensional function learning, and the dataset is generally small, with only a few thousand samples. However, ViTs are particularly powerful on large datasets (e.g., ImageNet) and tend to underperform relative to convolutional models on small datasets. Empirically, KAN and Transformer have potential conflicts.\n\nFor the experiments, the following should be addressed.\n\n1. It would have been better also to show the performance of learning the **non-smooth** or **high-dimensional functions**. The Feynman Equations may be too simple for conventional function approximation methods. You can try $f(x)=\\frac{1}{x} \\\\ \\sin{\\frac{1}{x}}$ and $f(x_1, \\cdots, x_{100}) = \\sum_{i=1}^{99} \\sin{(x_i + x\\_{100-i})}$; testing these functions would indicate if MLP-KAN overcomes certain limitations of KAN.\n\n2. In our testing, we found that KAN's training process differs from neural network models and is considerably slower. Comparing wall-clock training time could reveal any potential efficiency advantages.\n\n3. The central contribution focuses on enhancing Vision Transformer performance on CIFAR. It would be beneficial to compare with the Vision Transformer baseline on larger datasets like ImageNet-1K, which would add value.\n\n\n---\n\n[1] Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., ... & Tegmark, M. (2024). Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method, KAN-MLP, for improving the performance of ViT on function learning and CIFAR classification tasks. In order to achieve that, they replaced the traditional MLP layers within the transformer architecture. In this method, experts are dynamically selected based on the input through a gating mechanism, ensuring efficient routing of tokens to the most relevant experts. Their main contribution is applying an explainable KAN architecture to the Vision Transformer model. Comparing KAN, MLP, and MLP-KAN on function learning tasks, they show that MLP-KAN performs better than the other architectures in certain cases. Additionally, they also perform an ablation study on the CIFAR dataset to demonstrate that their extended model outperforms the naive ViT.", "soundness": 4, "presentation": 3, "contribution": 2, "strengths": "1. The paper proposes KAN-MLP, a novel approach to enhance Vision Transformers (ViT) for function learning and CIFAR classification tasks.\n\n2. It introduces the explainable KAN architecture into the ViT model, which could improve interoperability.\n\n3. The paper compares KAN, MLP, and MLP-KAN in function learning, showing potential advantages of MLP-KAN over other architectures.\n\n4. An ablation study on CIFAR demonstrates that the proposed model improves on the plain ViT model, highlighting the effectiveness of the enhancements.", "weaknesses": "1. The scaling law for MLP-KAN is missing, making it difficult to assess if MLP-KAN overcomes KAN’s limitations.\n\n2. The method may still suffer from the curse of dimensionality (COD), and the paper does not address how MLP-KAN performs for learning non-smooth and high-dimensional functions.\n\n3. There is a potential conflict between KAN’s suitability for low-dimensional tasks and ViT’s insuitability for small datasets. The experiments are limited to smooth functions and CIFAR datasets. Testing on larger datasets like ImageNet or MSCOCO would provide a more comprehensive view of the model’s performance relative to the Vision Transformer baseline.", "questions": "This paper contributes to advancing the KAN model, though it requires some clarifications on theory and experiments. Given these clarifications, if provided in an author's response, I would consider increasing the score. \n\nFor the theory, there are a few steps that need clarification and further clarification on novelty. \n\n1. KAN has advantages in model scaling under certain constraint conditions but not for all learning tasks. The KAN model obeys the Error Scaling formula $\\\\| f - (KAN) \\\\|\\_{C^m} \\le C G^{-k-1+m}$ and the scaling law $l \\\\propto N^{-\\\\alpha}$, but the author did not clarify the relationship between grid size $G$ and the input space dimension $n$. As a function approximation method, usually, the number of grid points $G$ is directly proportional to $G \\propto I^n$ ($I+1$ is the number of intervals on each dimension), which makes $\\\\| f - (KAN) \\\\|_{C^m} \\le C G^{(-k-1+m)/n}$ and triggers a serious curse of dimensionality (COD) problem (see KAN [1], Fig. 3.1, $f(x_1, \\cdots, x\\_{100}) = \\text{exp}(\\frac{1}{100} \\sum\\_{i=1}^{100} \\sin^2{\\frac{\\pi x_i}{2}})$). To address this problem, KAN authors hypothesize that the objective high-dimensional function is smooth and has sparse compositional structure to reduce the number of grid nodes $G \\ll I^n$. The authors did not provide the scaling law of MLP-KAN in the paper. Therefore, I don't know whether MLP-KAN overcomes the inherent limitations of KAN.\n\n2. This is particularly called into question due to the integration of KAN and ViT, since KAN and ViT usually exhibit different behavior on datasets of varying sizes. At present, KAN is suitable for low-dimensional function learning, and the dataset is generally small, with only a few thousand samples. However, ViTs are particularly powerful on large datasets (e.g., ImageNet) and tend to underperform relative to convolutional models on small datasets. Empirically, KAN and Transformer have potential conflicts.\n\nFor the experiments, the following should be addressed.\n\n1. It would have been better also to show the performance of learning the **non-smooth** or **high-dimensional functions**. The Feynman Equations may be too simple for conventional function approximation methods. You can try $f(x)=\\frac{1}{x} \\\\ \\sin{\\frac{1}{x}}$ and $f(x_1, \\cdots, x_{100}) = \\sum_{i=1}^{99} \\sin{(x_i + x\\_{100-i})}$; testing these functions would indicate if MLP-KAN overcomes certain limitations of KAN.\n\n2. In our testing, we found that KAN's training process differs from neural network models and is considerably slower. Comparing wall-clock training time could reveal any potential efficiency advantages.\n\n3. The central contribution focuses on enhancing Vision Transformer performance on CIFAR. It would be beneficial to compare with the Vision Transformer baseline on larger datasets like ImageNet-1K, which would add value.\n\n\n---\n\n[1] Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., ... & Tegmark, M. (2024). Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730484748603}], "openreview_url": "https://openreview.net/forum?id=F9JZiGradI", "arxiv_id": "2410.03027", "paper_pdf": "papers/F9JZiGradI.pdf", "paper_pdf_sha256": "0e8953a3e56c1180178dd9428e3f6f1373f75800ef86343fb5f0bad7943ea92a", "paper_pdf_bytes": 798828, "paper_pdf_source": "openreview", "code_url": "https://github.com/DLYuanGod/MLP-KAN", "code_repository": "DLYuanGod/MLP-KAN", "code_commit": "827569a7c04d4ce9d41e219e27d51b285735f165", "code_archive": "repos/F9JZiGradI.zip", "code_archive_sha256": "f768a03a32f83c52b5f9cd6957b977175a58e13bccc7c40138543133522d28eb", "code_archive_bytes": 57967, "code_file_count": 23, "code_extensions": {".py": 19, ".sh": 4}, "github_disk_usage_kb": 53, "github_languages": {"Python": 230546, "Shell": 2154}, "github_archived": false, "github_pushed_at": "2024-10-08T16:34:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mlp-kan-unifying-deep-representation-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "CXjz7p4qha", "year": 2024, "status": "rejected", "title": "Rotation Invariant Quantization for Model Compression", "authors": ["Joseph Kampeas", "Yury Nahshan", "Hanoch Kremer", "Gil Lederman", "Shira Zaloshinski", "Zheng Li", "Emir Haleva"], "authorids": ["~Joseph_Kampeas1", "~Yury_Nahshan1", "~Hanoch_Kremer1", "~Gil_Lederman1", "~Shira_Zaloshinski1", "~Zheng_Li16", "~Emir_Haleva1"], "authors_source": "OpenReview API", "abstract": "Post-training Neural Network (NN) model compression is an attractive approach for deploying large, memory-consuming models on devices with limited memory resources. \nIn this study, we investigate the rate-distortion tradeoff for NN model compression. First, we suggest a Rotation-Invariant Quantization (RIQ) technique that utilizes a single parameter to quantize the entire NN model, yielding a different rate at each layer, i.e., mixed-precision quantization. Then, we prove that our rotation-invariant approach is optimal in terms of compression. We rigorously evaluate RIQ and demonstrate its capabilities on various models and tasks. For example, RIQ facilitates $\\times 19.4$ and $\\times 52.9$ compression ratios on pre-trained VGG dense and pruned models, respectively,  with $<0.4\\%$ accuracy degradation.\nThe code is available in the supplementary material.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qGnYsR2SdN", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5056/Reviewer_N6Nd"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposes a new post-training quantization algorithm that given a neural network and some calibration images, produces mixed-precision quantized network. The key contribution of the paper is a new analysis technique motivated by the cosine-similarity based distortion measure between outputs quantized and unquantized network. The authors provide rate distortion analysis under the proposed measure and find the relation between the quantization bin width ($delta$) and layer's distortion (lemma 1) and the whole model's distortion (corollary 1). The authors then re-parametrize the search over $delta$ as search over parameter $k$ defined wrt distortion, and provide an efficient search algorithm (alg.1). Authors bound the possible values for optimal $k$ and make a heuristic search. The effectiveness of search algorithm, and of the proposed analysis is demonstrated on multiple networks and datasets.", "review_text": "The paper proposes a new post-training quantization algorithm that given a neural network and some calibration images, produces mixed-precision quantized network. The key contribution of the paper is a new analysis technique motivated by the cosine-similarity based distortion measure between outputs quantized and unquantized network. The authors provide rate distortion analysis under the proposed measure and find the relation between the quantization bin width ($delta$) and layer's distortion (lemma 1) and the whole model's distortion (corollary 1). The authors then re-parametrize the search over $delta$ as search over parameter $k$ defined wrt distortion, and provide an efficient search algorithm (alg.1). Authors bound the possible values for optimal $k$ and make a heuristic search. The effectiveness of search algorithm, and of the proposed analysis is demonstrated on multiple networks and datasets.", "strengths": "- The technique is well motivated and executed; even without the application to compression, the results and its analysis have their merits on its own.\n- The heuristic search over $k$ (alg 1), has good initial parameters and does not seem to require heavy tuning.\n- Very good presentation and flow.", "weaknesses": "- I found it a bit hard to understand how $k$ come into picture during the first readthrough of the paper. I feel like a little more work needed to introduce it. \n- There are a few arguable points that I count as a weakness, but they are easily addressable:\n  - I would like to understand how good is the search parameters wrt a synthetically created problem where (say weights sampled from Gaussian/Laplacian), single, layer and etc. How much the heuristics during the search may (or may not) miss the optimal bin width?\n  - Also, while I agree that in general setting search of $k$ is unbounded (as you write in the paper), practically speaking it is not the case: the weights are finite, and thus there are only certain number of  $delta$-s to check. This, has in fact been done in the work called \"Optimal quantization using scaled codebook\" (btw, you cite this paper but attribute it as QAT, which is not correct)\n- Results:\n  - I believe all compression results are given after ANS encoding; providing the compression ratio before ANS would be of great value (most papers report results before any additional encodings)", "questions": "Please see weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new post-training quantization algorithm that given a neural network and some calibration images, produces mixed-precision quantized network. The key contribution of the paper is a new analysis technique motivated by the cosine-similarity based distortion measure between outputs quantized and unquantized network. The authors provide rate distortion analysis under the proposed measure and find the relation between the quantization bin width ($delta$) and layer's distortion (lemma 1) and the whole model's distortion (corollary 1). The authors then re-parametrize the search over $delta$ as search over parameter $k$ defined wrt distortion, and provide an efficient search algorithm (alg.1). Authors bound the possible values for optimal $k$ and make a heuristic search. The effectiveness of search algorithm, and of the proposed analysis is demonstrated on multiple networks and datasets.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The technique is well motivated and executed; even without the application to compression, the results and its analysis have their merits on its own.\n- The heuristic search over $k$ (alg 1), has good initial parameters and does not seem to require heavy tuning.\n- Very good presentation and flow.", "weaknesses": "- I found it a bit hard to understand how $k$ come into picture during the first readthrough of the paper. I feel like a little more work needed to introduce it. \n- There are a few arguable points that I count as a weakness, but they are easily addressable:\n  - I would like to understand how good is the search parameters wrt a synthetically created problem where (say weights sampled from Gaussian/Laplacian), single, layer and etc. How much the heuristics during the search may (or may not) miss the optimal bin width?\n  - Also, while I agree that in general setting search of $k$ is unbounded (as you write in the paper), practically speaking it is not the case: the weights are finite, and thus there are only certain number of  $delta$-s to check. This, has in fact been done in the work called \"Optimal quantization using scaled codebook\" (btw, you cite this paper but attribute it as QAT, which is not correct)\n- Results:\n  - I believe all compression results are given after ANS encoding; providing the compression ratio before ANS would be of great value (most papers report results before any additional encodings)", "questions": "Please see weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698824491408}, {"id": "iNm1HvfT08", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5056/Reviewer_9KPJ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors proposed a post-training mixed-precision weight quantization technique for neural networks by optimization based on an information-theoretic paradigm.  They choose to minimize layer-wise bitwidth constrained by cosine distance.  They experimented with example models in comparison with other methods.", "review_text": "The authors proposed a post-training mixed-precision weight quantization technique for neural networks by optimization based on an information-theoretic paradigm.  They choose to minimize layer-wise bitwidth constrained by cosine distance.  They experimented with example models in comparison with other methods.", "strengths": "- The paradigm of optimization for mixed-precision quantization is novel.  \n- Theoretical results on optimization bounds are useful.", "weaknesses": "- How do layer-wise quantization errors accumilate?  The proposed algorithm does not seem to address this.   \n- In order for post-training quantization to be practical, activations are quantized too.  How activation quantization can be jointly done is not addressed.  \n- Experimental results did not show a definitive advantage over competing methods.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors proposed a post-training mixed-precision weight quantization technique for neural networks by optimization based on an information-theoretic paradigm.  They choose to minimize layer-wise bitwidth constrained by cosine distance.  They experimented with example models in comparison with other methods.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The paradigm of optimization for mixed-precision quantization is novel.  \n- Theoretical results on optimization bounds are useful.", "weaknesses": "- How do layer-wise quantization errors accumilate?  The proposed algorithm does not seem to address this.   \n- In order for post-training quantization to be practical, activations are quantized too.  How activation quantization can be jointly done is not addressed.  \n- Experimental results did not show a definitive advantage over competing methods.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698791238554}, {"id": "EGpyguTTRk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5056/Reviewer_4eGc"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes a post-training quantization algorithm named Rotation-Invariant Quantization (RIQ) to quantify the NN to mixed-precision, and the main approach is picking the quantization bin width to be proportional to the layers’ norm. Based on the rate-distortion theory, the proposed method searching for the optimal solution over the family of spherical distributions. Empirical results show the competitive performance of RIQ on several benchmarks.", "review_text": "This paper proposes a post-training quantization algorithm named Rotation-Invariant Quantization (RIQ) to quantify the NN to mixed-precision, and the main approach is picking the quantization bin width to be proportional to the layers’ norm. Based on the rate-distortion theory, the proposed method searching for the optimal solution over the family of spherical distributions. Empirical results show the competitive performance of RIQ on several benchmarks.", "strengths": "1.The authors provide a detailed analysis of the rate-distortion theory, which clarifies their research motivation well.\n\n2.This paper is well-written and organized, and the supplementary material is sufficiently detailed.", "weaknesses": "1. Mixed precision is difficult to apply in the industrial scenarios. It usually requires the design of specialized chips to achieve a slight increase in inference speed, so I have doubts about the impact of the proposed method. \n\n2. The experimental result lacks a comparison of inference speed of compressed model, between the proposed method and existing works.\n\n3. The experimental result lacks a comparison of lightweight structure including separable convolution (like mobilenetv2) with existing methods like AdaRound or BRECQ.", "questions": "1. I suggest the author provide more descriptions about the scenarios where the mixed-precision model can be applied. The advantages of mixed-precision models can be manifested in NLP-type structures.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a post-training quantization algorithm named Rotation-Invariant Quantization (RIQ) to quantify the NN to mixed-precision, and the main approach is picking the quantization bin width to be proportional to the layers’ norm. Based on the rate-distortion theory, the proposed method searching for the optimal solution over the family of spherical distributions. Empirical results show the competitive performance of RIQ on several benchmarks.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1.The authors provide a detailed analysis of the rate-distortion theory, which clarifies their research motivation well.\n\n2.This paper is well-written and organized, and the supplementary material is sufficiently detailed.", "weaknesses": "1. Mixed precision is difficult to apply in the industrial scenarios. It usually requires the design of specialized chips to achieve a slight increase in inference speed, so I have doubts about the impact of the proposed method. \n\n2. The experimental result lacks a comparison of inference speed of compressed model, between the proposed method and existing works.\n\n3. The experimental result lacks a comparison of lightweight structure including separable convolution (like mobilenetv2) with existing methods like AdaRound or BRECQ.", "questions": "1. I suggest the author provide more descriptions about the scenarios where the mixed-precision model can be applied. The advantages of mixed-precision models can be manifested in NLP-type structures.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698668589193}, {"id": "fhz0No7vIA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5056/Reviewer_qJtZ"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper examines post-training quantization of neural networks featuring linear layers, taking into account cosine distortion. The authors first demonstrate the rate-distortion trade-off between the original and quantized weights to establish the step size, $\\Delta_{\\ell}$, for each layer. Notably, all $\\Delta_{\\ell}$ values are governed by a singular parameter, $k$. Additionally, the paper investigates the rate-deviation analysis, wherein the deviation assesses the disparity in output. The authors introduce surrogate models in which the quantized weight is uniformly distributed across a space subject to random rotation, characterized by the angle $\\theta_{\\ell}$. Subsequently, it is proven that the mutual information is minimized when using the product distribution.", "review_text": "The paper examines post-training quantization of neural networks featuring linear layers, taking into account cosine distortion. The authors first demonstrate the rate-distortion trade-off between the original and quantized weights to establish the step size, $\\Delta_{\\ell}$, for each layer. Notably, all $\\Delta_{\\ell}$ values are governed by a singular parameter, $k$. Additionally, the paper investigates the rate-deviation analysis, wherein the deviation assesses the disparity in output. The authors introduce surrogate models in which the quantized weight is uniformly distributed across a space subject to random rotation, characterized by the angle $\\theta_{\\ell}$. Subsequently, it is proven that the mutual information is minimized when using the product distribution.", "strengths": "1. The paper is well-written and straightforward to understand.\n2. The concept of rotational invariance in neural network quantization is intriguing.", "weaknesses": "1. The central assumption, $\\||w\\|| = \\||\\hat{w}\\|| + o(\\||w\\||)$, appears to be contentious. For instance, with fixed-bit quantization, the discrepancy between $\\||w\\||$ and $\\||\\hat{w}\\||$ is proportional to $\\||w\\||$, or in other words, $O(\\||w\\||)$.\n\n2. Lemma 1 could benefit from a more rigorous presentation, especially concerning the $o(\\cdot)$. Additionally, both Lemma 1 and Corollary 1 do not appear to offer significant novel contributions.\n\n3. The surrogate model feels somewhat contrived. Moreover, the method for deriving $\\tilde{w}_{\\ell}$ is not clear. For instance, are the angles $\\theta_{\\ell}$ specified? The notion of being \"uniformly distributed on a cone\" is ambiguous due to its unbounded norm. As a result, Theorem 1 requires a more comprehensive problem definition.\n\n4. There seem to be some technical inaccuracies in the proof of Theorem 1. It is also essential to ensure that each mutual information, \n$I(w_{\\ell}, \\tilde{w}_{\\ell})$, adheres to the distortion criteria.\n\n5. The link between distortion (related to weights) and deviation (pertaining to output) is unclear.", "questions": "Please check Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper examines post-training quantization of neural networks featuring linear layers, taking into account cosine distortion. The authors first demonstrate the rate-distortion trade-off between the original and quantized weights to establish the step size, $\\Delta_{\\ell}$, for each layer. Notably, all $\\Delta_{\\ell}$ values are governed by a singular parameter, $k$. Additionally, the paper investigates the rate-deviation analysis, wherein the deviation assesses the disparity in output. The authors introduce surrogate models in which the quantized weight is uniformly distributed across a space subject to random rotation, characterized by the angle $\\theta_{\\ell}$. Subsequently, it is proven that the mutual information is minimized when using the product distribution.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is well-written and straightforward to understand.\n2. The concept of rotational invariance in neural network quantization is intriguing.", "weaknesses": "1. The central assumption, $\\||w\\|| = \\||\\hat{w}\\|| + o(\\||w\\||)$, appears to be contentious. For instance, with fixed-bit quantization, the discrepancy between $\\||w\\||$ and $\\||\\hat{w}\\||$ is proportional to $\\||w\\||$, or in other words, $O(\\||w\\||)$.\n\n2. Lemma 1 could benefit from a more rigorous presentation, especially concerning the $o(\\cdot)$. Additionally, both Lemma 1 and Corollary 1 do not appear to offer significant novel contributions.\n\n3. The surrogate model feels somewhat contrived. Moreover, the method for deriving $\\tilde{w}_{\\ell}$ is not clear. For instance, are the angles $\\theta_{\\ell}$ specified? The notion of being \"uniformly distributed on a cone\" is ambiguous due to its unbounded norm. As a result, Theorem 1 requires a more comprehensive problem definition.\n\n4. There seem to be some technical inaccuracies in the proof of Theorem 1. It is also essential to ensure that each mutual information, \n$I(w_{\\ell}, \\tilde{w}_{\\ell})$, adheres to the distortion criteria.\n\n5. The link between distortion (related to weights) and deviation (pertaining to output) is unclear.", "questions": "Please check Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698650404209}], "openreview_url": "https://openreview.net/forum?id=CXjz7p4qha", "arxiv_id": "2303.03106", "paper_pdf": "papers/CXjz7p4qha.pdf", "paper_pdf_sha256": "cf0709aeeac86d3c7658505f8a13909a4f76f1b6e146419a3bee84006bbf8b02", "paper_pdf_bytes": 559904, "paper_pdf_source": "openreview", "code_url": "https://github.com/ehaleva/RIQ", "code_repository": "ehaleva/RIQ", "code_commit": "2b4dfd28b8e1debe562e26cd11f612da8725d20a", "code_archive": "repos/CXjz7p4qha.zip", "code_archive_sha256": "6d1bf11d6979d940749180ab893a2ffbeb911106f91047a6cd157c21f74c1dd7", "code_archive_bytes": 24376, "code_file_count": 18, "code_extensions": {".py": 16, ".ipynb": 1, ".sh": 1}, "github_disk_usage_kb": 44, "github_languages": {"Python": 41133, "Jupyter Notebook": 8153, "Shell": 4482}, "github_archived": false, "github_pushed_at": "2024-01-23T17:24:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rotation-invariant-quantization-for-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ooqH4D9Xys", "year": 2023, "status": "rejected", "title": "LatentAugment: Dynamically Optimized Latent Probabilities of Data Augmentation", "authors": ["Koichi Kuriyama"], "authorids": ["~Koichi_Kuriyama1"], "authors_source": "OpenReview API", "abstract": "Although data augmentation is a powerful technique for improving the performance of image classification tasks, it is difficult to\nidentify the best augmentation policy. The optimal augmentation policy, which is the latent variable, cannot be directly observed. To address this problem, this study proposes \\textit{LatentAugment}, which estimates the latent probability of optimal augmentation. The proposed method is appealing in that it can dynamically optimize the augmentation strategies for each input and model parameter in\nlearning iterations. Theoretical analysis shows that LatentAugment is a general model that includes other augmentation methods as special cases, and it is simple and computationally efficient in comparison with existing augmentation methods. Experimental results show that the proposed LatentAugment has higher test accuracy than previous augmentation methods on the CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets. \n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "xvRNCU0neix", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1344/Reviewer_LaAK"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a new algorithm for data augmentation based on the same transformations used for autoaugment and follows-up papers. The algorithm considers policies that are composed of two sequential transformations selected from 16. Thus in total there are 256 different policies that can be selected. The probabilities of selecting those policies are latent variables that are estimated during training with a expectation maximization approach. In addition the algorithm can estimate also the probabilities of a policy conditioned to the given sample by normalizing the corresponding losses. Results show that the proposed algorithm outperforms previous approaches in most of the cases.", "review_text": "I consider this paper a valuable contribution for ICLR. However there are several points that the authors should improve for acceptance.\n\n\\- the writing the paper seems rushed, with some parts that are disconnected and other that should be improved and clarified (see clarity).\n\\- the differences with advAA seems minor. Authors should present and evaluate the differences with advAA.\n\n", "strengths": "\\+ The proposed approach seems to generate more meaningful augmentations than previous approaches as shown in the experimental results\n\n\\- Different parts of the paper are disconnected and difficult to follow. See clarity for more details.\n\n\\- The differences between this work and adversarial autoaugment (advAA) seem minimal. As shown by theorem 2.1, with uniform unconditional probabilities and $\\sigma \\rightarrow \\infty$ the proposed approach is advAA. As shown in table 3, the contribution of considering variable unconditional probabilities is quite low and in the order of the std. Not sure about the difference between $\\sigma = 1$ and a high value. It would be interesting to see table 3.\n\n\\- Differences in classification accuracy are relatively small. How can we verify that those differences are not due to a larger batch sizes or other implementation details", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents a new algorithm for data augmentation based on the same transformations used for autoaugment and follows-up papers. The algorithm considers policies that are composed of two sequential transformations selected from 16. Thus in total there are 256 different policies that can be selected. The probabilities of selecting those policies are latent variables that are estimated during training with a expectation maximization approach. In addition the algorithm can estimate also the probabilities of a policy conditioned to the given sample by normalizing the corresponding losses. Results show that the proposed algorithm outperforms previous approaches in most of the cases.", "strength_and_weaknesses": "\\+ The proposed approach seems to generate more meaningful augmentations than previous approaches as shown in the experimental results\n\n\\- Different parts of the paper are disconnected and difficult to follow. See clarity for more details.\n\n\\- The differences between this work and adversarial autoaugment (advAA) seem minimal. As shown by theorem 2.1, with uniform unconditional probabilities and $\\sigma \\rightarrow \\infty$ the proposed approach is advAA. As shown in table 3, the contribution of considering variable unconditional probabilities is quite low and in the order of the std. Not sure about the difference between $\\sigma = 1$ and a high value. It would be interesting to see table 3.\n\n\\- Differences in classification accuracy are relatively small. How can we verify that those differences are not due to a larger batch sizes or other implementation details", "clarity,_quality,_novelty_and_reproducibility": "\\- Clarity: the presentation of the paper should be improved in several aspects.\nIn the introduction I consider that the authors should not limit the presentation of previous methods to AutoAugment, even if the other approaches are mentioned in related work. I do not see many similarities between the proposed work and the Bayesian augmentations of Tran et al. (2017). Fig. 1 has a disconnect with the proposed formulation. For instance in Fig. 1 the E-step is applied on the multiplication of the conditional probabilities of the policies and the losses, while in Equ. 4 the conditional probabilities are multiplied by $P(y|o_z(x),\\theta)$. In method, in the third line the authors mention the use of random augmentations, but then $\\pi_z$ is used. In 2.1 first equation is presented as Bayes' rule, but in my understanding is just probability normalization. Algorithm 1 contains too much text and some definitions of variables are inaccurate. Equ. 3 seems a bit strange as it normalizes $h_z$ which is based on already normalized probabilities, thus twice exponentiation which might not be ideal for gradient propagation. \n\n\\-Quality: the method seems to outperform previous approaches, however the differences are relatively small. It would be important to provide the code in order to make sure that the improvements are due to the algorithm and not different and better hyperparameters.\n\n\\- Novelty: as mentioned previously, the paper seems very similar to advAA, thus the novelty is limited. Authors should explicitly present the improvements with respect advAA in related work as well with experiments.\n\n\\- Reproducibility: authors provide the hyperparameters for the trained model. However, it would be important also to provide the code and make sure that the obtained results are due to only the algorithm and not different hyper-parameters than previous work.", "summary_of_the_review": "I consider this paper a valuable contribution for ICLR. However there are several points that the authors should improve for acceptance.\n\n\\- the writing the paper seems rushed, with some parts that are disconnected and other that should be improved and clarified (see clarity).\n\\- the differences with advAA seems minor. Authors should present and evaluate the differences with advAA.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666668588898}, {"id": "kr9gC3rPxu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1344/Reviewer_AC1v"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposed a simple yet effective data augmentation method, which is called LatentAugment.", "review_text": "Although the paper propose a simple and effective method, I believe the paper needs improvement before getting accepted.", "strengths": "Strength:\n\nThe proposed LatentAugment is straightforward and reasonable;\n\nCompared to some previous methods, the proposed method is efficient in terms of computation cost;\n\nWeakness:\n\nThe performance gain is relatively small. According to Table 1, UBS is also an efficient method, while the proposed method is only slightly better than UBS with more computation cost. Given the results, I'm curious about result of K=1 for Table 1, from which we may better compare the training cost and performance of proposed method and UBS.\n\nGiven that the results are slightly better than previous method, more experiments are needed. For example, instead of vanilla fully-supervised learning, how will the proposed method benefit some settings where data augmentations might be really important, including few-shot learning, transfer learning, meta learning, etc. \n\nSome presentation needs to be improved. For instance, Figure 2 looks incomplete because some bars seems to be missing without clear explanation. I would suggest the authors either compared all the related methods under all the same settings, or explain clearly about the figure/results.\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposed a simple yet effective data augmentation method, which is called LatentAugment.", "strength_and_weaknesses": "Strength:\n\nThe proposed LatentAugment is straightforward and reasonable;\n\nCompared to some previous methods, the proposed method is efficient in terms of computation cost;\n\nWeakness:\n\nThe performance gain is relatively small. According to Table 1, UBS is also an efficient method, while the proposed method is only slightly better than UBS with more computation cost. Given the results, I'm curious about result of K=1 for Table 1, from which we may better compare the training cost and performance of proposed method and UBS.\n\nGiven that the results are slightly better than previous method, more experiments are needed. For example, instead of vanilla fully-supervised learning, how will the proposed method benefit some settings where data augmentations might be really important, including few-shot learning, transfer learning, meta learning, etc. \n\nSome presentation needs to be improved. For instance, Figure 2 looks incomplete because some bars seems to be missing without clear explanation. I would suggest the authors either compared all the related methods under all the same settings, or explain clearly about the figure/results.\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "I believe the proposed method has good reproducibility and some novelty, while the quality of the paper needs to be improved. Specifically, I believe more experiments are needed, and the presentation needs to be polished.", "summary_of_the_review": "Although the paper propose a simple and effective method, I believe the paper needs improvement before getting accepted.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666663148997}, {"id": "LaRDYLsEtQ1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1344/Reviewer_M9Bp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper addresses the learn optimal policy for data augmentation on image classification task.  Learning the optimal augmentation policy, which is the latent variable, interesting research problem. The paper proposes a simple via EM-based method (call LatentAugment) to estimate the probability of optimal policy via conditional probability of the policy customized for given input and network model. It also provides the theorical analysis to show some exisitng methods (AdvAA and UBS) as the special cases. Experiments are on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets using different network architectures showing the encouraging results that it outperforms most of existing methods.  ", "review_text": "Overall, I think the method is novel to me and most of claims are well-supported except one in the weakness. The paper is in the form can be published.  ", "strengths": "[Strength]\n\nThe paper is well-written and easy to follows. \n\nNovelty and with theoretical analysis provided.  \n\nThe model looks simple and computational efficiency, but would be good if with evidence supported.  \n\nThe results are encouraging with ablation study provided.  \n\n[Weakness]\n\nThe claim on computational efficiency needs to be supported with theoretical or empirical results. For example, it would improve a paper if the authors can provide the computational comparison to other methods.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper addresses the learn optimal policy for data augmentation on image classification task.  Learning the optimal augmentation policy, which is the latent variable, interesting research problem. The paper proposes a simple via EM-based method (call LatentAugment) to estimate the probability of optimal policy via conditional probability of the policy customized for given input and network model. It also provides the theorical analysis to show some exisitng methods (AdvAA and UBS) as the special cases. Experiments are on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets using different network architectures showing the encouraging results that it outperforms most of existing methods.  ", "strength_and_weaknesses": "[Strength]\n\nThe paper is well-written and easy to follows. \n\nNovelty and with theoretical analysis provided.  \n\nThe model looks simple and computational efficiency, but would be good if with evidence supported.  \n\nThe results are encouraging with ablation study provided.  \n\n[Weakness]\n\nThe claim on computational efficiency needs to be supported with theoretical or empirical results. For example, it would improve a paper if the authors can provide the computational comparison to other methods.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and the idea is novel to me.  Code is provided in the supplementary (but I do not have chance to test the code). ", "summary_of_the_review": "Overall, I think the method is novel to me and most of claims are well-supported except one in the weakness. The paper is in the form can be published.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666512865353}, {"id": "bkfpbw1x_39", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1344/Reviewer_AEKr"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a new framework for data augmentation that generalizes other existing data augmentation methods by incorporating Bayesian inference methods. Empirical performances on common benchmark datasets suggest that the LatentAugment methods outperforms other Adversarial benchmarks not only as measured by test accuracy in classification tasks, but also by training cost.", "review_text": "My personal evaluation would be a 5. I'd be happy to adjust my score, if the authors respond to the concerns accordingly.", "strengths": "Strengths:\n\n1. The paper is well-organized, with each section clearly stating what it does. The contribution of the paper has been clearly outlined.\n\n2. Experiments look sound and justify the main claim of the paper.\n\nWeaknesses:\n\n1. The level of novelty is relatively low. On a high-level, LatentAugment appears to be a straightforward application of the EM algorithm over the logarithmic losses defined with respect to conditional probability. The intuition behind the choice of loss function is also a bit vague. It would be helpful to explain why softmax is used and how the conditional probability relates to the goal of the LatentAugment algorithm.\n\n2. Except for the classification task, no other task has been used to test the performance of LatentAugment. It would be helpful to incorporate other relevant tasks where data augmentation is helpful (e.g. inpainting, image generation) to see how LatentAugment is useful in general.\n\n3. CIFAR, SVHN, and ImageNet are public datesets very commonly used across different image tasks. Has there been any tests on other less commonly used datasets where LatentAugment might be helpful? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new framework for data augmentation that generalizes other existing data augmentation methods by incorporating Bayesian inference methods. Empirical performances on common benchmark datasets suggest that the LatentAugment methods outperforms other Adversarial benchmarks not only as measured by test accuracy in classification tasks, but also by training cost.", "strength_and_weaknesses": "Strengths:\n\n1. The paper is well-organized, with each section clearly stating what it does. The contribution of the paper has been clearly outlined.\n\n2. Experiments look sound and justify the main claim of the paper.\n\nWeaknesses:\n\n1. The level of novelty is relatively low. On a high-level, LatentAugment appears to be a straightforward application of the EM algorithm over the logarithmic losses defined with respect to conditional probability. The intuition behind the choice of loss function is also a bit vague. It would be helpful to explain why softmax is used and how the conditional probability relates to the goal of the LatentAugment algorithm.\n\n2. Except for the classification task, no other task has been used to test the performance of LatentAugment. It would be helpful to incorporate other relevant tasks where data augmentation is helpful (e.g. inpainting, image generation) to see how LatentAugment is useful in general.\n\n3. CIFAR, SVHN, and ImageNet are public datesets very commonly used across different image tasks. Has there been any tests on other less commonly used datasets where LatentAugment might be helpful? ", "clarity,_quality,_novelty_and_reproducibility": "The paper is written very clearly. However, the originality of the work should be more elaborated. ", "summary_of_the_review": "My personal evaluation would be a 5. I'd be happy to adjust my score, if the authors respond to the concerns accordingly.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666506558272}], "openreview_url": "https://openreview.net/forum?id=ooqH4D9Xys", "arxiv_id": "2305.02668", "paper_pdf": "papers/ooqH4D9Xys.pdf", "paper_pdf_sha256": "42462877839a0a83bd1447d62726516b439641fe4ea1ef64c36de774430c2e96", "paper_pdf_bytes": 810362, "paper_pdf_source": "openreview", "code_url": "https://github.com/KoichiKuriyama/LatentAugment", "code_repository": "KoichiKuriyama/LatentAugment", "code_commit": "09179d794499fac8fc9de8584bda3ad3c0139da0", "code_archive": "repos/ooqH4D9Xys.zip", "code_archive_sha256": "5234e5ece051889c35f72bd72962b5e5f92bf61e8c7f6e28b5da87a2b4a2054d", "code_archive_bytes": 46775, "code_file_count": 21, "code_extensions": {".sh": 11, ".py": 10}, "github_disk_usage_kb": 63, "github_languages": {"Python": 65130, "Shell": 4156}, "github_archived": false, "github_pushed_at": "2023-05-05T01:07:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/latentaugment-dynamically-optimized-latent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zXne1klXIQ", "year": 2022, "status": "rejected", "title": "Improving Out-of-Distribution Robustness via Selective Augmentation", "authors": ["Huaxiu Yao", "Yu Wang", "Sai Li", "Linjun Zhang", "Weixin Liang", "James Zou", "Chelsea Finn"], "authorids": ["~Huaxiu_Yao1", "~Yu_Wang24", "saili@ruc.edu.cn", "~Linjun_Zhang1", "~Weixin_Liang1", "~James_Zou1", "~Chelsea_Finn1"], "authors_source": "OpenReview API", "abstract": "Machine learning algorithms typically assume that training and test examples are drawn from the same distribution. However, distribution shifts is a common problem in real-world applications and can cause models to perform dramatically worse at test time. In this paper, we specifically consider the problems of domain shifts and subpopulation shifts, where learning invariant representations by aligning domain-specific representations or balancing the risks across domains with regularizers are popular solutions. However, designing regularizers that are suitable for diverse real-world datasets is challenging. Instead, we shed new light on addressing distribution shifts by directly eliminating domain-related spurious correlations with augmentation, leading to a simple technique based on mixup, called LISA (Learning Invariant Representations via Selective Augmentation). LISA selectively interpolates samples either with the same labels but different domains or with the same domain but different labels. Empirically, we study the effectiveness of LISA on nine benchmarks ranging from subpopulation shifts to domain shifts. The results indicate that LISA consistently outperforms other state-of-the-art methods with superior invariant representations. The empirical findings are further strengthened by our theoretical analysis.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ZU9Dq3GC-Xp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper277/Reviewer_VX2N"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper propose a mixup-style data augmentation method under the data distribution shift context. In particular, data distributions are formulated as mixture of distributions (i.e., domains), and two distribution shift scenarios are considered: (1) domain shift, where the test domain and train domain are disjoint. (2) subpopulation shift, where test distribution has different mixture proportion than train distribution. It's  assumed that domain identification spuriously correlates with labels. To tackle this problem, this paper proposes two mixup strategies: (I) mixup two examples with same label but different domains; (II) mixup two examples with same domain but different labels. It's claimed that such mixup could cancel out the spurious correlations. Extensive experiments on a variety of datasets show its superiority compared to empirical risk minimization (ERM) and alternative data augmentation methods. The paper further provide theoretical analysis that under certain conditions, the proposed method has asymptotically smaller worst case classification errors than ERM and vanilla mixup. ", "review_text": "the idea is quite simple and intuitively reasonable, and empirical results seem extensive and significant, and theoretically justified to some extent\n\nsome of the results analysis is a bit confusing to me\n(1) in 4.1 \"evaluating robustness to domain shifts\", the best strategy was to always mixup same label with different domains, and the potential reason given is that the datasets actually have weak or even no spurious correlation between domain and label. Two questions follow:\n(a) From early text, seems both selection strategies are motivated by the spurious correlation, but here why do we still observe advantage over ERM or vanilla mixup? It would be great if you could clarify the different motivations (if any) of the two selection strategies\n(b) the reasoning about \"weak or no spurious correlation\" seems to be contradicting with claim in 4.3, where it's stated that \"compared with vanilla mixup, ...LISA...improve the OOD robustness by canceling out the spurious correlations..\". or did I misunderstand something? Is it easy to quantify such spurious correlation? if so, why not present the actual correlation metrics for these datasets? \n\n(2) for the ablation study, I think a more convincing way would be: first test LISA and mixup on a dataset that is known to have NO spurious correlations, then we expect neutral results; then test them on a data that is known to have spurious correlations, and we could give quantitative metrics of such correlations if possible, and show LISA is better than mixup; further more, the stronger the correlation, the more advantage LISA has. Is that what you're trying to demonstrate here? \n\n(3) In table 8, vanilla mixup shows much worse performance than ERM in terms of learning invariant representations under the defined metric, which doesn't seem quite reasonable to me. Shouldn't we expect the contrary? \n \n\n\nAnother minor question: In Theorem 1, is p the dimension of x? If so it's better to state that explicitly in the theorem instead of relying on readers to go back to text and only to find in the superscript notation. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper propose a mixup-style data augmentation method under the data distribution shift context. In particular, data distributions are formulated as mixture of distributions (i.e., domains), and two distribution shift scenarios are considered: (1) domain shift, where the test domain and train domain are disjoint. (2) subpopulation shift, where test distribution has different mixture proportion than train distribution. It's  assumed that domain identification spuriously correlates with labels. To tackle this problem, this paper proposes two mixup strategies: (I) mixup two examples with same label but different domains; (II) mixup two examples with same domain but different labels. It's claimed that such mixup could cancel out the spurious correlations. Extensive experiments on a variety of datasets show its superiority compared to empirical risk minimization (ERM) and alternative data augmentation methods. The paper further provide theoretical analysis that under certain conditions, the proposed method has asymptotically smaller worst case classification errors than ERM and vanilla mixup. ", "main_review": "the idea is quite simple and intuitively reasonable, and empirical results seem extensive and significant, and theoretically justified to some extent\n\nsome of the results analysis is a bit confusing to me\n(1) in 4.1 \"evaluating robustness to domain shifts\", the best strategy was to always mixup same label with different domains, and the potential reason given is that the datasets actually have weak or even no spurious correlation between domain and label. Two questions follow:\n(a) From early text, seems both selection strategies are motivated by the spurious correlation, but here why do we still observe advantage over ERM or vanilla mixup? It would be great if you could clarify the different motivations (if any) of the two selection strategies\n(b) the reasoning about \"weak or no spurious correlation\" seems to be contradicting with claim in 4.3, where it's stated that \"compared with vanilla mixup, ...LISA...improve the OOD robustness by canceling out the spurious correlations..\". or did I misunderstand something? Is it easy to quantify such spurious correlation? if so, why not present the actual correlation metrics for these datasets? \n\n(2) for the ablation study, I think a more convincing way would be: first test LISA and mixup on a dataset that is known to have NO spurious correlations, then we expect neutral results; then test them on a data that is known to have spurious correlations, and we could give quantitative metrics of such correlations if possible, and show LISA is better than mixup; further more, the stronger the correlation, the more advantage LISA has. Is that what you're trying to demonstrate here? \n\n(3) In table 8, vanilla mixup shows much worse performance than ERM in terms of learning invariant representations under the defined metric, which doesn't seem quite reasonable to me. Shouldn't we expect the contrary? \n \n\n\nAnother minor question: In Theorem 1, is p the dimension of x? If so it's better to state that explicitly in the theorem instead of relying on readers to go back to text and only to find in the superscript notation. ", "summary_of_the_review": "the idea is quite simple and intuitively reasonable, and empirical results seem extensive and significant, and theoretically justified to some extent, but the results analysis and some experimental design could be more insightful or improved. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635925777323}, {"id": "AIjTwOXgDz_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper277/Reviewer_iP5p"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers the model robustness under distribution shift brought by domains and subpopulations. Specifically, based on the interpolation scheme in mixup, the authors propose two selection strategies to perform data augmentation, aim at eliminating the spurious correlations and learning an invariant representation.", "review_text": "Strength:\n(1) The authors address a critical point that prevent models from generalization, namely spurious correlation.\n(2) The proposed method is simple and easy to implement, and the empirical results are within expectation.\n\nWeakness:\n(1) Maybe the biggest concern is the contribution over previous work. The proposed method can be seen as a heuristic extention of mixup. Though simple and easy to follow, the contribution is marginal. \n(2) There ara growing trends on investigating OOD generalization under missing domain label, it is better to at least include such work (e.g. [1]) for discussion.\n(3) There is a hyper-parameter p_sel that controls the probability of performing different strategy, is there a rule of thumb or we need to tune it for every task?\n[1] Qiao, F., Zhao, L., & Peng, X. (2020). Learning to Learn Single Domain Generalization. In CVPR.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the model robustness under distribution shift brought by domains and subpopulations. Specifically, based on the interpolation scheme in mixup, the authors propose two selection strategies to perform data augmentation, aim at eliminating the spurious correlations and learning an invariant representation.", "main_review": "Strength:\n(1) The authors address a critical point that prevent models from generalization, namely spurious correlation.\n(2) The proposed method is simple and easy to implement, and the empirical results are within expectation.\n\nWeakness:\n(1) Maybe the biggest concern is the contribution over previous work. The proposed method can be seen as a heuristic extention of mixup. Though simple and easy to follow, the contribution is marginal. \n(2) There ara growing trends on investigating OOD generalization under missing domain label, it is better to at least include such work (e.g. [1]) for discussion.\n(3) There is a hyper-parameter p_sel that controls the probability of performing different strategy, is there a rule of thumb or we need to tune it for every task?\n[1] Qiao, F., Zhao, L., & Peng, X. (2020). Learning to Learn Single Domain Generalization. In CVPR.", "summary_of_the_review": "This paper address the spurious correlation by augmenting the data via interpolation. Although intuitions are provided and empirical effectiveness is illustrated accordingly, the contributions over previous work (e.g. mixup) are marginal.\n\n=========After Response==========\n\nThe response from authors has addressed my major concerns, so I raise my score accordingly.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635873091123}, {"id": "_ChCHY2Eycq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper277/Reviewer_4YgJ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "Authors introduced approaches aimed at learning invariant predictors across data sources. Rather than using distribution/risk matching schemes as often done by previous work, they propose to train models against mixtures of data points as a means to avoid that models rely on spurious correlations between domain and class labels, since such correlations observed during training might not hold at testing time. The proposed setting uses the idea of mixup to combine data instances in two different schemes: I-combine data points from the same class but different domains, and II-combine data points from the same domain but from different classes.", "review_text": "Strengths:\n\n+ The proposed approach is simple and efficient, and can be directly incorporated in or combined with other invariance-inducing approaches;\n\n+ Prediction performance is shown to improve over a number of recent baselines under challenging benchmarks.\n\nWeaknesses/suggestions: \n\n- The proposal requires assumptions that are not discussed in the manuscript. In [1], it was shown that domain-invariant approaches can only improve out-of-distribution generalization if data-conditional label distributions P(y|x) are fixed across domains; i.e., observing x suffices in order to determine y, regardless of the domain according to which x was observed.\n\n- My main concern lies in the reported evaluation, which is focused on showing improvements in terms of downstream performance, and presented results consist of comparing the proposed approach with alternative methods. While improving downstream performance is of course our ultimate goal, and results are strong in this sense, doing so does not explain the sources of improvements. In particular, authors claim that the mixup strategies they introduce yield some type of domain invariance, which is not verified empirically. To verify that learned representations are invariant, that could be achieved via domain-prediction experiments; i.e., train domain classifiers on top of representations learned by different methods. The higher the accuracy of such a classifier, the less invariant are representations. For the case of prediction-level invariance, authors could perhaps evaluate the range of estimated risks across domains. Improvements on either one of these notions of invariance would then explain observed improvements in terms of prediction accuracy.\n\n- \"We argue that the failure of other methods in some datasets may be caused by their regularizers limiting model capacity to some extent\". That's another case where the evaluation lacks in supporting authors' claims. This hypothesis can be verified via in-domain prediction performance, i.e., overly regularized underfitting models should result in accuracy degradation in the training domains. Alternatively, one could rule out the underfitting effect by using higher capacity model classes.\n\n- Conclusions from the risk bounds provided in theorems 1 and 2 are a bit unrealistic given the strong assumptions that imply the results. In particular, the model assumed for the data generating process is overly simplified and, although it enables theoretical analysis, it's unclear to which extend the conclusions hold in practice.\n\n- Finally, regarding novelty, it seems different recent approaches introduce methods that use some sort of mixup across domains in similar settings. For the domain adaptation/generalization cases, there are, for instance, [2,3,4]. For the multi-domain case, cross-domain mixup was studied in [5]. Authors do compare results against [4], but it's unclear how the proposed approach differs from those other recent applications of mixup under similar settings, and the related work section should include such a discussion.\n\nOther comments:\n\n- On page 2, the setting described was not originally introduced by Koh et al. (2021). To my knowledge, it was first discussed in [6] and later on in [7].\n\n- The definition of mixup for labels in the rightmost term in eq. 2 seems to require labels y are one-hot encoded, which is not mentioned in the text.\n\n- While in the title authors claim to be improving out-of-distribution robustness, a large part of the work focus on the multi-domain learning (or fairness) setting, where one's goal is to find predictors with uniform risks across a set of domains that doesn't change from training to testing. Technically, that wouldn't be out-of-distribution. Perhaps there should be a sentence or two in the introduction indicating what authors refer to as out-of-distribution.\n\nReferences:\n\n[1] Zhao, Han, et al. \"On learning invariant representations for domain adaptation.\" International Conference on Machine Learning. PMLR, 2019.\n\n[2] Shu, Yang, et al. \"Open Domain Generalization with Domain-Augmented Meta-Learning.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.\n\n[3] Wang, Yufei, Haoliang Li, and Alex C. Kot. \"Heterogeneous domain generalization via domain mixup.\" ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.\n\n[4] Xu, Minghao, et al. \"Adversarial domain adaptation with domain mixup.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 04. 2020.\n\n[5] Chuang, Ching-Yao, and Youssef Mroueh. \"Fair mixup: Fairness via interpolation.\" arXiv preprint arXiv:2103.06503 (2021).\n\n[6] Muandet, Krikamol, David Balduzzi, and Bernhard Schölkopf. \"Domain generalization via invariant feature representation.\" International Conference on Machine Learning. PMLR, 2013.\n\n[7] Albuquerque, Isabela, et al. \"Generalizing to unseen domains via distribution matching.\" arXiv preprint arXiv:1911.00804 (2019).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Authors introduced approaches aimed at learning invariant predictors across data sources. Rather than using distribution/risk matching schemes as often done by previous work, they propose to train models against mixtures of data points as a means to avoid that models rely on spurious correlations between domain and class labels, since such correlations observed during training might not hold at testing time. The proposed setting uses the idea of mixup to combine data instances in two different schemes: I-combine data points from the same class but different domains, and II-combine data points from the same domain but from different classes.", "main_review": "Strengths:\n\n+ The proposed approach is simple and efficient, and can be directly incorporated in or combined with other invariance-inducing approaches;\n\n+ Prediction performance is shown to improve over a number of recent baselines under challenging benchmarks.\n\nWeaknesses/suggestions: \n\n- The proposal requires assumptions that are not discussed in the manuscript. In [1], it was shown that domain-invariant approaches can only improve out-of-distribution generalization if data-conditional label distributions P(y|x) are fixed across domains; i.e., observing x suffices in order to determine y, regardless of the domain according to which x was observed.\n\n- My main concern lies in the reported evaluation, which is focused on showing improvements in terms of downstream performance, and presented results consist of comparing the proposed approach with alternative methods. While improving downstream performance is of course our ultimate goal, and results are strong in this sense, doing so does not explain the sources of improvements. In particular, authors claim that the mixup strategies they introduce yield some type of domain invariance, which is not verified empirically. To verify that learned representations are invariant, that could be achieved via domain-prediction experiments; i.e., train domain classifiers on top of representations learned by different methods. The higher the accuracy of such a classifier, the less invariant are representations. For the case of prediction-level invariance, authors could perhaps evaluate the range of estimated risks across domains. Improvements on either one of these notions of invariance would then explain observed improvements in terms of prediction accuracy.\n\n- \"We argue that the failure of other methods in some datasets may be caused by their regularizers limiting model capacity to some extent\". That's another case where the evaluation lacks in supporting authors' claims. This hypothesis can be verified via in-domain prediction performance, i.e., overly regularized underfitting models should result in accuracy degradation in the training domains. Alternatively, one could rule out the underfitting effect by using higher capacity model classes.\n\n- Conclusions from the risk bounds provided in theorems 1 and 2 are a bit unrealistic given the strong assumptions that imply the results. In particular, the model assumed for the data generating process is overly simplified and, although it enables theoretical analysis, it's unclear to which extend the conclusions hold in practice.\n\n- Finally, regarding novelty, it seems different recent approaches introduce methods that use some sort of mixup across domains in similar settings. For the domain adaptation/generalization cases, there are, for instance, [2,3,4]. For the multi-domain case, cross-domain mixup was studied in [5]. Authors do compare results against [4], but it's unclear how the proposed approach differs from those other recent applications of mixup under similar settings, and the related work section should include such a discussion.\n\nOther comments:\n\n- On page 2, the setting described was not originally introduced by Koh et al. (2021). To my knowledge, it was first discussed in [6] and later on in [7].\n\n- The definition of mixup for labels in the rightmost term in eq. 2 seems to require labels y are one-hot encoded, which is not mentioned in the text.\n\n- While in the title authors claim to be improving out-of-distribution robustness, a large part of the work focus on the multi-domain learning (or fairness) setting, where one's goal is to find predictors with uniform risks across a set of domains that doesn't change from training to testing. Technically, that wouldn't be out-of-distribution. Perhaps there should be a sentence or two in the introduction indicating what authors refer to as out-of-distribution.\n\nReferences:\n\n[1] Zhao, Han, et al. \"On learning invariant representations for domain adaptation.\" International Conference on Machine Learning. PMLR, 2019.\n\n[2] Shu, Yang, et al. \"Open Domain Generalization with Domain-Augmented Meta-Learning.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.\n\n[3] Wang, Yufei, Haoliang Li, and Alex C. Kot. \"Heterogeneous domain generalization via domain mixup.\" ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.\n\n[4] Xu, Minghao, et al. \"Adversarial domain adaptation with domain mixup.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 04. 2020.\n\n[5] Chuang, Ching-Yao, and Youssef Mroueh. \"Fair mixup: Fairness via interpolation.\" arXiv preprint arXiv:2103.06503 (2021).\n\n[6] Muandet, Krikamol, David Balduzzi, and Bernhard Schölkopf. \"Domain generalization via invariant feature representation.\" International Conference on Machine Learning. PMLR, 2013.\n\n[7] Albuquerque, Isabela, et al. \"Generalizing to unseen domains via distribution matching.\" arXiv preprint arXiv:1911.00804 (2019).\n", "summary_of_the_review": "The paper is well-written, the approach is efficient and observed to work well on a number of benchmarks. However, experiments supporting key claims are lacking, and it's unclear whether the observed improvements in terms of invariance (either at feature- or prediction-level) hold true since no supporting experiments are reported. Contextualization of the proposal relative to past literature also needs improvements since cross-domain data mixup was studied in the past under both settings considered by the authors. The provided discussion on related work does not clarify what and how the authors' proposal improves upon previous work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1634589471055}], "openreview_url": "https://openreview.net/forum?id=zXne1klXIQ", "arxiv_id": "2201.00299", "paper_pdf": "papers/zXne1klXIQ.pdf", "paper_pdf_sha256": "e795abdd0efa5aee990b496a6fb2fe9938ec5e90cc79b2bbfa3cc552c1cc462b", "paper_pdf_bytes": 484630, "paper_pdf_source": "openreview", "code_url": "https://github.com/huaxiuyao/LISA", "code_repository": "huaxiuyao/LISA", "code_commit": "c857a6b296b5d130898f0d51a6d693411c39e651", "code_archive": "repos/zXne1klXIQ.zip", "code_archive_sha256": "2a456d77d55efb4562682f9d0c124a91477c8d41260a87bc2d4b8c9adc0f745d", "code_archive_bytes": 65661, "code_file_count": 38, "code_extensions": {".py": 37, ".sh": 1}, "github_disk_usage_kb": 72, "github_languages": {"Python": 196014, "Shell": 3428}, "github_archived": false, "github_pushed_at": "2023-04-12T19:08:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-out-of-distribution-robustness-via-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "le9LIliDOG", "year": 2021, "status": "rejected", "title": "Efficient Long-Range Convolutions for Point Clouds", "authors": ["Yifan Peng", "Lin Lin", "Lexing Ying", "Leonardo Zepeda-Nunez"], "authorids": ["pyf04142017@sjtu.edu.cn", "~Lin_Lin1", "~Lexing_Ying1", "~Leonardo_Zepeda-Nunez1"], "authors_source": "OpenReview API", "abstract": "The efficient treatment of long-range interactions for point clouds is a challenging problem in many scientific machine learning applications. To extract global information, one usually needs a large window size, a large number of layers, and/or a large number of channels. This can often significantly increase the computational cost. In this work, we present a novel neural network layer that directly incorporates long-range information for a point cloud. This layer, dubbed the long-range convolutional (LRC)-layer, leverages the convolutional theorem coupled with the non-uniform Fourier transform. In a nutshell, the LRC-layer mollifies the point cloud to an adequately sized regular grid, computes its Fourier transform, multiplies the result by a set of trainable Fourier multipliers, computes the inverse Fourier transform, and finally interpolates the result back to the point cloud. The resulting global all-to-all convolution operation can be performed in nearly-linear time asymptotically with respect to the number of input points. The LRC-layer is a particularly powerful tool when combined with local convolution as together they offer efficient and seamless treatment of both short and long range interactions. We showcase this framework by introducing a neural network architecture that combines LRC-layers with short-range convolutional layers to accurately learn the energy and force associated with a $N$-body potential.  We also exploit the induced two-level decomposition and propose an efficient strategy to train the combined architecture with a reduced number of samples.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0WQeS4qeZCk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1864/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is clearly written, and presents an approach to efficiently utilize long range convolutions through a nonuniform FFT in for coulomb particle configurations. \n\nPros: \n- Presents a long range convolution layer, with an efficient implementation so that a neural network model can benefit from both short and long range interactions among data points. \n\n- To model pairwise relations between points of the N-body potential, two short range descriptor networks are utilized, where the input relations are constructed both repulsively and attractively. \n\nCons:\nExperimentally, the new long-short range NN model is validated only for computing the energy and forces for a N-body model. There are no experiments with real-life point cloud data.\n\nIn the proposed algorithm, a sampling in a regular grid is required. A fixed size Cartesian grid is utilized in the experiments, therefore, the effect of the gridding resolution is not analyzed.\n\nThe quantitative experimental results in Tables 1, 2, and 3 dwell on varying the screening coefficient mu, hence investigates how the test error changes for point configurations with varying screening parameters, which signifies the amount of diffusion of the potential. It can be seen that the test error for the proposed approach falls as the screening gets higher, that is when we have more localized interactions between the particles. What could be the significance of this finding for a practical application is not discussed, therefore not made clear. \n\nThe authors state that the proposed method could be a useful tool in a wide range of Machine Learning tasks. However, in the paper, only estimation of potential energy for Coulomb particle configurations is demonstrated as an application. I am curious to see how the presented approach could find use in learning tasks of real point cloud configurations.\n\nThe paper presents a way to incorporate long range convolutions in neural networks, which could be beneficial. However, experimental evaluation is not satisfactory, as immediate implications in point cloud analysis is not obvious. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Utilizing the nonuniform FFT, a long range convolutional layer (LRC) is presented. A neural network which combines both LRC and short range conv layers are built. In a two-scale training strategy, first many small-scale data are trained without the LRC, then a small set of large-scale data are trained with both short and long range conv layers. The model is tested on 1D and 2D screened Coulomb particle configurations to estimate the overall potential energy and forces. ", "review": "The paper is clearly written, and presents an approach to efficiently utilize long range convolutions through a nonuniform FFT in for coulomb particle configurations. \n\nPros: \n- Presents a long range convolution layer, with an efficient implementation so that a neural network model can benefit from both short and long range interactions among data points. \n\n- To model pairwise relations between points of the N-body potential, two short range descriptor networks are utilized, where the input relations are constructed both repulsively and attractively. \n\nCons:\nExperimentally, the new long-short range NN model is validated only for computing the energy and forces for a N-body model. There are no experiments with real-life point cloud data.\n\nIn the proposed algorithm, a sampling in a regular grid is required. A fixed size Cartesian grid is utilized in the experiments, therefore, the effect of the gridding resolution is not analyzed.\n\nThe quantitative experimental results in Tables 1, 2, and 3 dwell on varying the screening coefficient mu, hence investigates how the test error changes for point configurations with varying screening parameters, which signifies the amount of diffusion of the potential. It can be seen that the test error for the proposed approach falls as the screening gets higher, that is when we have more localized interactions between the particles. What could be the significance of this finding for a practical application is not discussed, therefore not made clear. \n\nThe authors state that the proposed method could be a useful tool in a wide range of Machine Learning tasks. However, in the paper, only estimation of potential energy for Coulomb particle configurations is demonstrated as an application. I am curious to see how the presented approach could find use in learning tasks of real point cloud configurations.\n\nThe paper presents a way to incorporate long range convolutions in neural networks, which could be beneficial. However, experimental evaluation is not satisfactory, as immediate implications in point cloud analysis is not obvious. ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603972419339}, {"id": "K6oGTndTwG", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1864/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper concentrated on exploring how to efficiently extract the interaction information from the point clouds.  A key point lies in utilizing the non-uniform Fourier transform, rather than the regular Fourier transform. However, there exist some issues that need to be solved. \n\n+ves: \n+ The exploration of long-range interactions for point clouds is interesting.\n\n+ The paper is well written. The related work makes a clear description of many fields about point cloud.\n\nConcerns: \n1. In the introduction part, the authors describe many tasks that rely on point-cloud presentation. However, the authors ignore pointing out the existing issues. The authors should clearly present them.\n\n2. In the algorithmic section, the authors claim that \"NUFFT serves as the corner-stone of the LRC-layer\". So \"NUFFT\" is the unique solution? Whether or not some operations can replace FFT? \n\n3. Point-cloud is indeed important for many tasks as described in the introduction part, but the authors just explore the effects of the proposed algorithm in a \"synthetic\" experiment. The experimental results are not convincing for readers, the authors should conduct more real-world tasks to verify the effectiveness of the proposed method.\n\n4. The presentation of this work needs to improve, if possible, the authors should provide an intuitive schematic diagram to present the procedure of proposing this idea.\n\n5. In the experimental section, the author should replot Figure2-3 to ensure clear enough for a better read.  \n\n#########################################################################\n\nMinor Comments:  \n\n(1)  “N_sample” and \"N\" maybe exist the inclusion relation，it will be better to replace one of them with the other form;\n\n(2)  The formula system is a little vague, if possible, the authors can simplify them to clearly describe.\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper needs to make changes in some aspects.", "review": "This paper concentrated on exploring how to efficiently extract the interaction information from the point clouds.  A key point lies in utilizing the non-uniform Fourier transform, rather than the regular Fourier transform. However, there exist some issues that need to be solved. \n\n+ves: \n+ The exploration of long-range interactions for point clouds is interesting.\n\n+ The paper is well written. The related work makes a clear description of many fields about point cloud.\n\nConcerns: \n1. In the introduction part, the authors describe many tasks that rely on point-cloud presentation. However, the authors ignore pointing out the existing issues. The authors should clearly present them.\n\n2. In the algorithmic section, the authors claim that \"NUFFT serves as the corner-stone of the LRC-layer\". So \"NUFFT\" is the unique solution? Whether or not some operations can replace FFT? \n\n3. Point-cloud is indeed important for many tasks as described in the introduction part, but the authors just explore the effects of the proposed algorithm in a \"synthetic\" experiment. The experimental results are not convincing for readers, the authors should conduct more real-world tasks to verify the effectiveness of the proposed method.\n\n4. The presentation of this work needs to improve, if possible, the authors should provide an intuitive schematic diagram to present the procedure of proposing this idea.\n\n5. In the experimental section, the author should replot Figure2-3 to ensure clear enough for a better read.  \n\n#########################################################################\n\nMinor Comments:  \n\n(1)  “N_sample” and \"N\" maybe exist the inclusion relation，it will be better to replace one of them with the other form;\n\n(2)  The formula system is a little vague, if possible, the authors can simplify them to clearly describe.\n ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603890762754}, {"id": "zm2EgYTfCcp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1864/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an efficient long-range convolution method for point clouds by using the non-uniform Fourier transform. The long-range convolutional (LRC)-layer mollifies the point cloud to an adequately sized regular grid, computes its Fourier transform, multiplies the results by a set of trainable Fourier multipliers, computes the inverse Fourier transform, and finally interpolates the result back to the point cloud. The method is demonstrated to be effective by a N-body problem.\n\nOverall, the paper is clearly written with high quality. The originality of the paper seems to be not very strong since it directly adapts the NUFFT to this work. Besides that, there are several concerns about this paper that need to be addressed:\n\n1. How to choose the grid size L_{FFT} in the Fourier Space? How does this parameter affect the results?\n\n2. The global pooling layers in DNN can also capture the long-range information to some extent, and are also very efficient. How does the LRC compare with the global pooling layers?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting work", "review": "The paper proposes an efficient long-range convolution method for point clouds by using the non-uniform Fourier transform. The long-range convolutional (LRC)-layer mollifies the point cloud to an adequately sized regular grid, computes its Fourier transform, multiplies the results by a set of trainable Fourier multipliers, computes the inverse Fourier transform, and finally interpolates the result back to the point cloud. The method is demonstrated to be effective by a N-body problem.\n\nOverall, the paper is clearly written with high quality. The originality of the paper seems to be not very strong since it directly adapts the NUFFT to this work. Besides that, there are several concerns about this paper that need to be addressed:\n\n1. How to choose the grid size L_{FFT} in the Fourier Space? How does this parameter affect the results?\n\n2. The global pooling layers in DNN can also capture the long-range information to some extent, and are also very efficient. How does the LRC compare with the global pooling layers?\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603693984250}, {"id": "KthYYqQls1a", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1864/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Edit: I was unaware that papers could be submitted to arXiv simultaneously, I am sorry for that. Here is my (very late) review.\nI made it before reading other reviewers' reviews.\n\nAn efficient method for fitting long-range style interactions in point clouds is presented.\nIt makes use of NUFFT rather than convoluting a long-ranged (thus system-size and expensive) kernel directly in real space.\nAlong with this architecture, a method for training it efficiently is presented. This 2 steps strategy consists in training the short-range part of the kernels well with a lot of (supposedly inexpensive) short range data, and fitting the long-range kernels with less data in a second time.\n\nOverall, the paper is clearly written and clearly exposes the methods used.\nThe results are interesting for applications but do not seem ground-breaking (although I am not an expert of point clouds -specialized networks).\nIn terms of experimental results, I think a couple of points would deserved to be answered (see below).\n\nIn conclusion, I think the paper is marginally above acceptance level.\n\n\nIn terms of results, the paper clearly shows that NUFFT scales essentially like O(N) (with N the points number) whereas naive direct space convolution scales as O(N^2).\nHowever, when I read the algorithm (page 3), my main curiosity is : how well does the algorithm deals with large systems (large domain \\Omega), and the competing parameters seems to be the resolution of the function g_\\tau (which plays the role of mollified dirac) versus the system size \\Omega or L. Concretely, I expect that for too large \\tau/L (presented in appendix, Equation 20), the precision of the long range kernels will be poor and error will be large; and in the opposite case of very small \\tau/L, precision will be good but compute time will increase (I guess it would increase as FFT does, in O(L/\\tau log(L/\\tau))).\nProbably there is a regime where direct convolution (which does not need the approximation introduced by g_\\tau, as far as I understood), is better then the NUFFT approach introduced here. I would guess that in the large system size (large \\Omega=L^d) and small particles number N limit (i.e. the low density limit), the direct approach does well ?\nI think such a discussion (and possibly a couple of experiments) would improve the paper a lot, showing the limits of the method and letting the reader understand the reasons for its strengths (which are very real, I do believe it !)\n\n\nA bit more of commenting on the experiments' results would be appreciated. For instance, it seems that the 2-scale training strategy is especially efficient (not needing many samples) when the LRIs are sufficiently strong (figure 3, right). This is probably an effect of the \"screening\" of LRI by short-range Interactions when alpha_1 is tooo strong compared to alpha_2, making it harder to learn about LRIs.\nAlso, the fact that all curves essentially collapse beyond a given N_sample could receive a comment.\nA couple of remarks like that about the strengths and limits of the approach would be nice.\n\n\nAside from these 2 main comments, I have minor remarks of presentation:\n- figure 2 left and figure 3 both sides: poor choice of colors/linestyles. Curves come in pairs, and this should be suggested in the choice of display, using similar colors for pairs, and different linetyles in different pairs. It would improve readability (also, think of the color blind people!)\n- fig2 , right: the O(N), O(N^2) scalings are un-readable. make them parallel to the plots you do and simple black dashed to improve clarity.\n\nIn the tables, although here results are \"trivial\" (always the same line has the smallest error), you could use the convention of putting in bold the better numbers.\n\nTable 1, you say \"mu_2 can be arbitrary here\". What you mean is that it needs not be defined, because alpha_2=0 ?\nI found this sentence confusing (maybe it's just me).\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "good improvement of existing method is presented, experiments could be improved", "review": "Edit: I was unaware that papers could be submitted to arXiv simultaneously, I am sorry for that. Here is my (very late) review.\nI made it before reading other reviewers' reviews.\n\nAn efficient method for fitting long-range style interactions in point clouds is presented.\nIt makes use of NUFFT rather than convoluting a long-ranged (thus system-size and expensive) kernel directly in real space.\nAlong with this architecture, a method for training it efficiently is presented. This 2 steps strategy consists in training the short-range part of the kernels well with a lot of (supposedly inexpensive) short range data, and fitting the long-range kernels with less data in a second time.\n\nOverall, the paper is clearly written and clearly exposes the methods used.\nThe results are interesting for applications but do not seem ground-breaking (although I am not an expert of point clouds -specialized networks).\nIn terms of experimental results, I think a couple of points would deserved to be answered (see below).\n\nIn conclusion, I think the paper is marginally above acceptance level.\n\n\nIn terms of results, the paper clearly shows that NUFFT scales essentially like O(N) (with N the points number) whereas naive direct space convolution scales as O(N^2).\nHowever, when I read the algorithm (page 3), my main curiosity is : how well does the algorithm deals with large systems (large domain \\Omega), and the competing parameters seems to be the resolution of the function g_\\tau (which plays the role of mollified dirac) versus the system size \\Omega or L. Concretely, I expect that for too large \\tau/L (presented in appendix, Equation 20), the precision of the long range kernels will be poor and error will be large; and in the opposite case of very small \\tau/L, precision will be good but compute time will increase (I guess it would increase as FFT does, in O(L/\\tau log(L/\\tau))).\nProbably there is a regime where direct convolution (which does not need the approximation introduced by g_\\tau, as far as I understood), is better then the NUFFT approach introduced here. I would guess that in the large system size (large \\Omega=L^d) and small particles number N limit (i.e. the low density limit), the direct approach does well ?\nI think such a discussion (and possibly a couple of experiments) would improve the paper a lot, showing the limits of the method and letting the reader understand the reasons for its strengths (which are very real, I do believe it !)\n\n\nA bit more of commenting on the experiments' results would be appreciated. For instance, it seems that the 2-scale training strategy is especially efficient (not needing many samples) when the LRIs are sufficiently strong (figure 3, right). This is probably an effect of the \"screening\" of LRI by short-range Interactions when alpha_1 is tooo strong compared to alpha_2, making it harder to learn about LRIs.\nAlso, the fact that all curves essentially collapse beyond a given N_sample could receive a comment.\nA couple of remarks like that about the strengths and limits of the approach would be nice.\n\n\nAside from these 2 main comments, I have minor remarks of presentation:\n- figure 2 left and figure 3 both sides: poor choice of colors/linestyles. Curves come in pairs, and this should be suggested in the choice of display, using similar colors for pairs, and different linetyles in different pairs. It would improve readability (also, think of the color blind people!)\n- fig2 , right: the O(N), O(N^2) scalings are un-readable. make them parallel to the plots you do and simple black dashed to improve clarity.\n\nIn the tables, although here results are \"trivial\" (always the same line has the smallest error), you could use the convention of putting in bold the better numbers.\n\nTable 1, you say \"mu_2 can be arbitrary here\". What you mean is that it needs not be defined, because alpha_2=0 ?\nI found this sentence confusing (maybe it's just me).\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602606123324}], "openreview_url": "https://openreview.net/forum?id=le9LIliDOG", "arxiv_id": "2010.05295", "paper_pdf": "papers/le9LIliDOG.pdf", "paper_pdf_sha256": "5e8fbe55c69c7e2b9b31010936d602d3cb02ed5aa541d14ca0a730ef5dd6d702", "paper_pdf_bytes": 1514152, "paper_pdf_source": "openreview", "code_url": "https://github.com/Forgotten/Efficient_Long-Range_Convolutions_for_Point_Clouds", "code_repository": "Forgotten/Efficient_Long-Range_Convolutions_for_Point_Clouds", "code_commit": "1fe364052eca9330edeaeb32c59d0ec5195c12c4", "code_archive": "repos/le9LIliDOG.zip", "code_archive_sha256": "3134b04565d12849475e67c4b80bcfb0fd3a0d3f9ef0d61ca57acb91291e7ac5", "code_archive_bytes": 45981, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 73, "github_languages": {"Python": 49242}, "github_archived": false, "github_pushed_at": "2020-10-12T15:16:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-long-range-convolutions-for-point-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1e8WTEYPB", "year": 2020, "status": "rejected", "title": "Sparse and Structured Visual Attention", "authors": ["Pedro Henrique Martins", "Vlad Niculae", "Zita Marinho", "André F.T. Martins"], "authorids": ["pedrohenriqueamartins@gmail.com", "vlad@vene.ro", "zita.marinho@priberam.pt", "andre.martins@unbabel.com"], "authors_source": "OpenReview API", "abstract": "Visual attention mechanisms have been widely used in image captioning models. In this paper, to better link the image structure with the generated text, we replace the traditional softmax attention mechanism by two alternative sparsity-promoting transformations: sparsemax and Total-Variation Sparse Attention (TVmax). With sparsemax, we obtain sparse attention weights, selecting relevant features.  In order to promote sparsity and encourage fusing of the related adjacent spatial locations, we propose TVmax.  By selecting relevant groups of features, the TVmax transformation improves interpretability. We present results in the Microsoft COCO and Flickr30k datasets, obtaining gains in comparison to softmax.  TVmax outperforms the other compared attention mechanisms in terms of human-rated caption quality and attention relevance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hygts56c5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper376/AnonReviewer4"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper produces a new method call TVmax and presents that the selective visual attention could improve the score in the Image Captioning task. Different from the fusedmax[2] which fuses attention in one dimension, the proposed method encourages the sparse attention over contiguous 2D regions. Compared with the softmax function, the sparsemax[1]  and the TVmax are able to sparse the visual attention very well. The paper also evaluates the score in both automated metrics and the human rating. Experiments show that the sparse visual attention achieves higher performance with a little computational cost.\n\nOne problem in this paper is that the author applies their proposed TVmax on Image Captioning  task, however it only achieves a little improvement on the automated metrics compared with the baseline(softmax). I wonder whether there is a better task for evaluating the visual attention. \n\nAlthough the proposed method (TVmax) is slightly worse than the sparsemax in the automated metrics, it is still promising in multimodal problems. \n\nTherefore, My decision leans to a weak accept.\n\nSome questions:\n1.From the experiments, the proposed method achieved only a little higher performance than the baseline(softmax). Could you please show some reasons about that?\n2.Could you show some results of TVmax on the other task in order to show the effectiveness of the proposed method?\n\n\n[1]From Softmax to Sparsemax: A Sparse Model of Attention and Multi-Label Classification.      André F. T. Martins, Ramón Fernandez Astudillo\n[2]A Regularized Framework for Sparse and Structured Neural Attention .            Vlad Niculae, Mathieu Blondel", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #376", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper produces a new method call TVmax and presents that the selective visual attention could improve the score in the Image Captioning task. Different from the fusedmax[2] which fuses attention in one dimension, the proposed method encourages the sparse attention over contiguous 2D regions. Compared with the softmax function, the sparsemax[1]  and the TVmax are able to sparse the visual attention very well. The paper also evaluates the score in both automated metrics and the human rating. Experiments show that the sparse visual attention achieves higher performance with a little computational cost.\n\nOne problem in this paper is that the author applies their proposed TVmax on Image Captioning  task, however it only achieves a little improvement on the automated metrics compared with the baseline(softmax). I wonder whether there is a better task for evaluating the visual attention. \n\nAlthough the proposed method (TVmax) is slightly worse than the sparsemax in the automated metrics, it is still promising in multimodal problems. \n\nTherefore, My decision leans to a weak accept.\n\nSome questions:\n1.From the experiments, the proposed method achieved only a little higher performance than the baseline(softmax). Could you please show some reasons about that?\n2.Could you show some results of TVmax on the other task in order to show the effectiveness of the proposed method?\n\n\n[1]From Softmax to Sparsemax: A Sparse Model of Attention and Multi-Label Classification.      André F. T. Martins, Ramón Fernandez Astudillo\n[2]A Regularized Framework for Sparse and Structured Neural Attention .            Vlad Niculae, Mathieu Blondel"}, "tcdate": 1572686497268}, {"id": "SyeAYTPb9r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper376/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes two sparsifying methods of computing attention weights, dubbed sparsemax and TVmax, which appear to slightly improve objective and subjective image captioning scores.    The sparsifying projections are posed as optimization problems, and algorithms for their computation, along with formula for their gradients are given.  Proof of the optimality of these algorithms relies significantly on prior work, so could not be checked deeply without bringing in additional sources.  \n\nIt is not clear that the motivation for these sparsifying objectives is sound.   The conventional softmax approach to attention weights should be capable of producing attention weights near zero, which would be effectively sparse, especially if the pre-activations, z_i, in equation (1), are allowed to have a large enough range.   It's not clear why weights should need to be zero exactly in the ignored regions, since being near zero should be sufficient to contribute almost nothing to the subsequent weighted sum.    So it is also not clear why the strict sparsity itself, as opposed to the effective sparsity of the softmax, should explain the differences in Figure 1, and in the results.  In particular it is unclear why the strict sparsity should prevent repetition; when looking at the weight distributions in the two cases, a more likely story seems to be that the weight distributions don't repeat as much from one word to the next in the second case, but there is no clear reason to attribute this to sparsity.   The pictures of the attention weights are lacking a color scale, so it is impossible to see how close to zero it comes in the unattended regions, although the gray color values chosen for these regions might be misleading. \nThe TVmax approach, in addition to sparsity, also constrains the non-zero region to be contiguous.   To the extent that this improves performance, this presumably introduces an inductive bias that matches the data.   It is unclear why this fails to produce better objective scores than sparsemax, while producing better human ratings.   In any case it is not clear why this should necessarily be a good inductive bias for all images, although it is plausible that it helps in some cases.  \nIn many neural network problems, what makes a difference has more to do with the optimizability of the gradients, than the specific activations per se, and that might be the case here too, although the paper does not analyze this aspect of the proposed models.   \n\nOverall the paper is flawed by the lack of clarity in the motivation for the proposed methods, and the lack of retrospective analysis and understanding of why the proposed methods should improve results. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes two sparsifying methods of computing attention weights, dubbed sparsemax and TVmax, which appear to slightly improve objective and subjective image captioning scores.    The sparsifying projections are posed as optimization problems, and algorithms for their computation, along with formula for their gradients are given.  Proof of the optimality of these algorithms relies significantly on prior work, so could not be checked deeply without bringing in additional sources.  \n\nIt is not clear that the motivation for these sparsifying objectives is sound.   The conventional softmax approach to attention weights should be capable of producing attention weights near zero, which would be effectively sparse, especially if the pre-activations, z_i, in equation (1), are allowed to have a large enough range.   It's not clear why weights should need to be zero exactly in the ignored regions, since being near zero should be sufficient to contribute almost nothing to the subsequent weighted sum.    So it is also not clear why the strict sparsity itself, as opposed to the effective sparsity of the softmax, should explain the differences in Figure 1, and in the results.  In particular it is unclear why the strict sparsity should prevent repetition; when looking at the weight distributions in the two cases, a more likely story seems to be that the weight distributions don't repeat as much from one word to the next in the second case, but there is no clear reason to attribute this to sparsity.   The pictures of the attention weights are lacking a color scale, so it is impossible to see how close to zero it comes in the unattended regions, although the gray color values chosen for these regions might be misleading. \nThe TVmax approach, in addition to sparsity, also constrains the non-zero region to be contiguous.   To the extent that this improves performance, this presumably introduces an inductive bias that matches the data.   It is unclear why this fails to produce better objective scores than sparsemax, while producing better human ratings.   In any case it is not clear why this should necessarily be a good inductive bias for all images, although it is plausible that it helps in some cases.  \nIn many neural network problems, what makes a difference has more to do with the optimizability of the gradients, than the specific activations per se, and that might be the case here too, although the paper does not analyze this aspect of the proposed models.   \n\nOverall the paper is flawed by the lack of clarity in the motivation for the proposed methods, and the lack of retrospective analysis and understanding of why the proposed methods should improve results. \n"}, "tcdate": 1572072838225}, {"id": "HkxewMw0tH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper376/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the problem of applying attention to the problem of image captioning. To this end the authors first apply full attention where the probabilities are computed using softmax before applying the recently proposed Sparsemax - which essentially computes probabilities from scores by performing a projection on to the probability simplex. The authors then propose a variant of Sparsemax which they call TVMax, which has the property that it encourages assigning probability weights to contiguous regions in 2D space unlike Sparsemax which has no such incentive. The main idea is to augment the Sparsemax projection loss with a Lasso like penalty which penalizes assigning different attention probabilities to contiguous regions in the image. The authors then compare their TVMax approach with softmax and Sparsemax attention for image captioning and show improvements on the MSCOCO and Flickr datasets.\n\nThe idea of applying additional structural constraints on the sparsity structure induced by Sparsemax is a cool idea, and I like the idea of incentivizing contiguous pixels to have similar attention probabilities. Sparse attention patterns seems like an important direction of research with the motivation of either 1) improving generalization over full attention or 2) scaling to inputs of length where full attention is not feasible. This seems like a good progress in the first direction. The major weakness I see in this work is that the authors only limited their experiments to image captioning. It would be interesting to see if their approach could benefit other tasks such as machine translation, image generation etc. The other issue with their approach is that it doesn't seem to scale well - if I understand correctly their algorithm takes O(n^2logn) for sequence length n. The other issue potentially could be weak baselines since the authors use an LSTM for their caption generation network instead of Transformer.\n\n[Edit: After going through reviewer discussion, I updated my score to reject. I am not convinced of the motivation for sparse attention unless it is for long sequences, since otherwise the regular softmax should be able to assign 0's to the un-needed items. Moreover, for generalization one can use attention dropout which is simpler instead.]", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "This paper studies the problem of applying attention to the problem of image captioning. To this end the authors first apply full attention where the probabilities are computed using softmax before applying the recently proposed Sparsemax - which essentially computes probabilities from scores by performing a projection on to the probability simplex. The authors then propose a variant of Sparsemax which they call TVMax, which has the property that it encourages assigning probability weights to contiguous regions in 2D space unlike Sparsemax which has no such incentive. The main idea is to augment the Sparsemax projection loss with a Lasso like penalty which penalizes assigning different attention probabilities to contiguous regions in the image. The authors then compare their TVMax approach with softmax and Sparsemax attention for image captioning and show improvements on the MSCOCO and Flickr datasets.\n\nThe idea of applying additional structural constraints on the sparsity structure induced by Sparsemax is a cool idea, and I like the idea of incentivizing contiguous pixels to have similar attention probabilities. Sparse attention patterns seems like an important direction of research with the motivation of either 1) improving generalization over full attention or 2) scaling to inputs of length where full attention is not feasible. This seems like a good progress in the first direction. The major weakness I see in this work is that the authors only limited their experiments to image captioning. It would be interesting to see if their approach could benefit other tasks such as machine translation, image generation etc. The other issue with their approach is that it doesn't seem to scale well - if I understand correctly their algorithm takes O(n^2logn) for sequence length n. The other issue potentially could be weak baselines since the authors use an LSTM for their caption generation network instead of Transformer.\n\n[Edit: After going through reviewer discussion, I updated my score to reject. I am not convinced of the motivation for sparse attention unless it is for long sequences, since otherwise the regular softmax should be able to assign 0's to the un-needed items. Moreover, for generalization one can use attention dropout which is simpler instead.]", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571873367661}], "openreview_url": "https://openreview.net/forum?id=r1e8WTEYPB", "arxiv_id": "2002.05556", "paper_pdf": "papers/r1e8WTEYPB.pdf", "paper_pdf_sha256": "cfa8a50ed83609a2c05f43e6b7f0e9873bd3c591c15efa99edf78b2af89b535d", "paper_pdf_bytes": 1566959, "paper_pdf_source": "openreview", "code_url": "https://github.com/deep-spin/TVmax", "code_repository": "deep-spin/TVmax", "code_commit": "2b812dc1b78fe8869bbd2f11dca89aba5341bcc5", "code_archive": "repos/r1e8WTEYPB.zip", "code_archive_sha256": "fd95b4c318f14b6769d38fb43033436ebdc56d223eb10d73e16d61974efd2e10", "code_archive_bytes": 115789, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 113, "github_languages": {"Python": 90932}, "github_archived": false, "github_pushed_at": "2021-12-21T23:41:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sparse-and-structured-visual-attention-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1lfHhR9tm", "year": 2019, "status": "rejected", "title": "The Natural Language Decathlon: Multitask Learning as Question Answering", "authors": ["Bryan McCann", "Nitish Shirish Keskar", "Caiming Xiong", "Richard Socher"], "authorids": ["bmccann@salesforce.com", "nkeskar@salesforce.com", "cxiong@salesforce.com", "rsocher@salesforce.com"], "authors_source": "OpenReview API", "abstract": "Deep learning has improved performance on many natural language processing (NLP) tasks individually.\nHowever, general NLP models cannot emerge within a paradigm that focuses on the particularities of a single metric, dataset, and task.\nWe introduce the Natural Language Decathlon (decaNLP), a challenge that spans ten tasks:\nquestion answering, machine translation, summarization, natural language inference, sentiment analysis, semantic role labeling, relation extraction, goal-oriented dialogue, semantic parsing, and commonsense pronoun resolution.\nWe cast all tasks as question answering over a context.\nFurthermore, we present a new multitask question answering network (MQAN) that jointly learns all tasks in decaNLP without any task-specific modules or parameters more effectively than sequence-to-sequence and reading comprehension baselines.\nMQAN shows improvements in transfer learning for machine translation and named entity recognition, domain adaptation for sentiment analysis and natural language inference, and zero-shot capabilities for text classification.\nWe demonstrate that the MQAN's multi-pointer-generator decoder is key to this success and that performance further improves with an anti-curriculum training strategy.\nThough designed for decaNLP, MQAN also achieves state of the art results on the WikiSQL semantic parsing task in the single-task setting. \nWe also release code for procuring and processing data, training and evaluating models, and reproducing all experiments for decaNLP.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "B1xIPcoqh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1522/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper formulates several different NLP problems as Q&A problem and proposed a general deep learning architecture. All these tasks are trained together. \n\nIf the goal is to achieve general AI, the paper gives a good starting point. One technical novelty is the deep learning architecture for this general Q&A problem including the multi-pointer-generator. The paper presents an example of how to do a multi-task learning for 10 different tasks. It raises a very challenging problem or in some way release a new dataset.\n\nIf our goal is to optimize a single task, the usefulness of the method proposed by the paper is questionable. \nAs we know, multi-task learning works well if some important knowledge shared by different tasks can be learned and leveraged. From table 2, we see for many problems, the results of the single task training are better than the multi-task training, meaning that other tasks can't really help at least under this framework. This makes me doubt if this multi-task learning is useful if our goal is to optimize the performance of a single task. This general model also sacrifices some important prior knowledge of an individual task. For example, for the Squad, the prior that the answer is a continuous span. Ideally, the prior knowledge should be leveraged.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good example to treat different NLP problems as Q&A and trained together. Results for some problems are worse than their state-of-the-art.", "review": "The paper formulates several different NLP problems as Q&A problem and proposed a general deep learning architecture. All these tasks are trained together. \n\nIf the goal is to achieve general AI, the paper gives a good starting point. One technical novelty is the deep learning architecture for this general Q&A problem including the multi-pointer-generator. The paper presents an example of how to do a multi-task learning for 10 different tasks. It raises a very challenging problem or in some way release a new dataset.\n\nIf our goal is to optimize a single task, the usefulness of the method proposed by the paper is questionable. \nAs we know, multi-task learning works well if some important knowledge shared by different tasks can be learned and leveraged. From table 2, we see for many problems, the results of the single task training are better than the multi-task training, meaning that other tasks can't really help at least under this framework. This makes me doubt if this multi-task learning is useful if our goal is to optimize the performance of a single task. This general model also sacrifices some important prior knowledge of an individual task. For example, for the Squad, the prior that the answer is a continuous span. Ideally, the prior knowledge should be leveraged.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541220958111}, {"id": "Syx1siQK37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1522/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Update: I've updated my score based on the clarifications from the authors to some of my questions/concerns about the experimental set-up and multi-task/single-task differences.\n\nOriginal Review:\nThis paper provides a new framework for multitask learning in nlp by taking advantage of the similarities in 10 common NLP tasks. The modeling is building on pre-existing qa models but has some original aspects that were augmented to accommodate the various tasks.  The decaNLP framework could be a useful benchmark for other nlp researchers.  \n\nExperiments indicate that the multi-task set-up does worse on average than the single-task set-up.  I wish there was more analysis on why multi-task setups are helpful in some tasks and not others.  With a bit more fine-grained analysis, the experiments and framework in this paper could be very beneficial towards other researchers who want to experiment with multi-task learning or who want to use the decaNLP framework as a benchmark.\n\nI also found the adaptation to new tasks and zero-shot experiments very interesting but the set-up was not described very concretely: \n  -in the transfer learning section, I hope the writers will elaborate on whether the performance gain is coming from the model being pretrained on a multi-task objective or if there would still be performance gain by pretraining a model on only one of those tasks.  For example, would a model pre-trained solely on IWSLT see the same performance gain when transferred to English->Czech as in Figure 4? Or is it actually the multi-task training that is causing the improvement in transfer learning? \n  -Can you please add more detail about the setup for replacing +/- with happy/angry or supportive/unsupportive? What were the (empirical) results of that experiment?\n\nI think the paper doesn’t quite stand on its own without the appendix, which is a major weakness in terms of clarity.  The related work, for example, should really be included in the main body of the paper.  I also recommend that more of the original insights (such as the experimentation with curriculum learning) should be included in the body of the text to count towards original contributions.  \n\nAs a suggestion, the authors may be able to condense the discussion of the 10 tasks in order to make more room in the main text for a related work section plus more of their motivations and experimental results.  If necessary, the main paper *can* exceed 8 pages and still fit ICLR guidelines.\n\nVery minor detail: I noticed some inconsistency in the bibliography regarding full names vs. first initials only.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "New framework has a lot of potential, but the experiments, motivations, and related work are missing details", "review": "Update: I've updated my score based on the clarifications from the authors to some of my questions/concerns about the experimental set-up and multi-task/single-task differences.\n\nOriginal Review:\nThis paper provides a new framework for multitask learning in nlp by taking advantage of the similarities in 10 common NLP tasks. The modeling is building on pre-existing qa models but has some original aspects that were augmented to accommodate the various tasks.  The decaNLP framework could be a useful benchmark for other nlp researchers.  \n\nExperiments indicate that the multi-task set-up does worse on average than the single-task set-up.  I wish there was more analysis on why multi-task setups are helpful in some tasks and not others.  With a bit more fine-grained analysis, the experiments and framework in this paper could be very beneficial towards other researchers who want to experiment with multi-task learning or who want to use the decaNLP framework as a benchmark.\n\nI also found the adaptation to new tasks and zero-shot experiments very interesting but the set-up was not described very concretely: \n  -in the transfer learning section, I hope the writers will elaborate on whether the performance gain is coming from the model being pretrained on a multi-task objective or if there would still be performance gain by pretraining a model on only one of those tasks.  For example, would a model pre-trained solely on IWSLT see the same performance gain when transferred to English->Czech as in Figure 4? Or is it actually the multi-task training that is causing the improvement in transfer learning? \n  -Can you please add more detail about the setup for replacing +/- with happy/angry or supportive/unsupportive? What were the (empirical) results of that experiment?\n\nI think the paper doesn’t quite stand on its own without the appendix, which is a major weakness in terms of clarity.  The related work, for example, should really be included in the main body of the paper.  I also recommend that more of the original insights (such as the experimentation with curriculum learning) should be included in the body of the text to count towards original contributions.  \n\nAs a suggestion, the authors may be able to condense the discussion of the 10 tasks in order to make more room in the main text for a related work section plus more of their motivations and experimental results.  If necessary, the main paper *can* exceed 8 pages and still fit ICLR guidelines.\n\nVery minor detail: I noticed some inconsistency in the bibliography regarding full names vs. first initials only.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541122966701}, {"id": "S1lGsQAShm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1522/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I appreciate the work that went into creating this paper, but I'm afraid I see little justification for accepting it.  I have three major complaints with this paper:                                                                         \n                                                                                                     \n1. I think the framing of decaNLP presented in this paper does more harm than good, because it perpetuates a misguided view of question answering.\n                                                                                                     \nQuestion answering is not a unified phenomenon.  There is no such thing as \"general question answering\", not even for humans.  Consider \"What is 2 + 3?\", \"What's the terminal velocity of a rain drop?\", and \"What is the meaning of life?\"  All of these questions require very different systems to answer, and trying to pretend they are the same doesn't help anyone solve any problems.\n                                                                                                     \nQuestion answering is a _format_ for studying particular phenomena.  Sometimes it is useful to pose a task as QA, and sometimes it is not.  QA is not a useful format for studying problems when you only have a single question (like \"what is the sentiment?\" or \"what is the translation?\"), and there is no hope of transfer from a related task.  Posing translation or classification as QA serves no useful purpose and gives people the wrong impression about question answering as a format for studying problems.\n\nWe have plenty of work that studies multiple datasets at a time (including in the context of semi-supervised / transfer learning), without doing this misguided framing of all of them as QA (see, e.g., the ELMo and BERT papers, which evaluated on many separate tasks).  I don't see any compelling justification for setting things up this way.\n                                                                                                     \n2. One of the main claims of this paper is transfer from one task to another by posing them all as question answering.  There is nothing new in the transfer results that were presented here, however.  For QA-SRL / QA-ZRE, transfer from SQuAD / other QA tasks has already been shown by Luheng He (http://aclweb.org/anthology/N18-2089) and Omer Levy (that was the whole point of the QA-ZRE paper), so this is merely reproducing that result (without mentioning that they did it first).  For all other tasks, performance drops when you try to train all tasks together, sometimes significantly (as in translation, unsurprisingly).  For the Czech task, fine tuning a pre-trained model has already been shown to help.  Transfer from MNLI to SNLI is known already and not surprising - one of the main points of MNLI was domain transfer, so obviously this has been studied before.  The claims about transfer to new classification tasks are misleading, as you really have the _same_ classification task, you've just arbitrarily changed how you're encoding the class label.  It _might_ be the case that you still get transfer if you actually switch to a related classification task, but you haven't examined that case.\n                                                                                                     \n3. This paper tries to put three separate ideas into a single conference paper, and all three ideas suffer as a result, because there is not enough space to do any of them justice.  Giving 15 pages of appendix for an 8 page paper, where some of the main content of the paper is pushed to the appendix, is egregious.  Putting your work in the context of related work is not something that should be pushed into an appendix, and we should not encourage this behavior.\n                                                                                                     \nThe three ideas here seem to me to be (1) decaNLP, (2) the model architecture of MQAN, (3) transfer results.  Any of these three could have been a single conference paper, had it been done well.  As it stands, decaNLP isn't described or motivated well enough, and there isn't any space left in the paper to address my severe criticisms of it in my first point.  Perhaps if you had dedicated the paper to decaNLP, you could have given arguments that the framing is worthwhile, and described the tasks and their setup as QA sufficiently (as it is, I don't see any description anywhere of how the context is constructed for WikiSQL; did I miss it somewhere?).  For MQAN, there's more than a page of the core new architecture that's pushed into the appendix.  And for the transfer results, there is very little comparison to other transfer methods (e.g., ELMo, CoVe), or any deep analysis of what's going on - as I mentioned above, basically all of the results presented are just confirming what has already been done elsewhere.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Misguided and overcrowded", "review": "I appreciate the work that went into creating this paper, but I'm afraid I see little justification for accepting it.  I have three major complaints with this paper:                                                                         \n                                                                                                     \n1. I think the framing of decaNLP presented in this paper does more harm than good, because it perpetuates a misguided view of question answering.\n                                                                                                     \nQuestion answering is not a unified phenomenon.  There is no such thing as \"general question answering\", not even for humans.  Consider \"What is 2 + 3?\", \"What's the terminal velocity of a rain drop?\", and \"What is the meaning of life?\"  All of these questions require very different systems to answer, and trying to pretend they are the same doesn't help anyone solve any problems.\n                                                                                                     \nQuestion answering is a _format_ for studying particular phenomena.  Sometimes it is useful to pose a task as QA, and sometimes it is not.  QA is not a useful format for studying problems when you only have a single question (like \"what is the sentiment?\" or \"what is the translation?\"), and there is no hope of transfer from a related task.  Posing translation or classification as QA serves no useful purpose and gives people the wrong impression about question answering as a format for studying problems.\n\nWe have plenty of work that studies multiple datasets at a time (including in the context of semi-supervised / transfer learning), without doing this misguided framing of all of them as QA (see, e.g., the ELMo and BERT papers, which evaluated on many separate tasks).  I don't see any compelling justification for setting things up this way.\n                                                                                                     \n2. One of the main claims of this paper is transfer from one task to another by posing them all as question answering.  There is nothing new in the transfer results that were presented here, however.  For QA-SRL / QA-ZRE, transfer from SQuAD / other QA tasks has already been shown by Luheng He (http://aclweb.org/anthology/N18-2089) and Omer Levy (that was the whole point of the QA-ZRE paper), so this is merely reproducing that result (without mentioning that they did it first).  For all other tasks, performance drops when you try to train all tasks together, sometimes significantly (as in translation, unsurprisingly).  For the Czech task, fine tuning a pre-trained model has already been shown to help.  Transfer from MNLI to SNLI is known already and not surprising - one of the main points of MNLI was domain transfer, so obviously this has been studied before.  The claims about transfer to new classification tasks are misleading, as you really have the _same_ classification task, you've just arbitrarily changed how you're encoding the class label.  It _might_ be the case that you still get transfer if you actually switch to a related classification task, but you haven't examined that case.\n                                                                                                     \n3. This paper tries to put three separate ideas into a single conference paper, and all three ideas suffer as a result, because there is not enough space to do any of them justice.  Giving 15 pages of appendix for an 8 page paper, where some of the main content of the paper is pushed to the appendix, is egregious.  Putting your work in the context of related work is not something that should be pushed into an appendix, and we should not encourage this behavior.\n                                                                                                     \nThe three ideas here seem to me to be (1) decaNLP, (2) the model architecture of MQAN, (3) transfer results.  Any of these three could have been a single conference paper, had it been done well.  As it stands, decaNLP isn't described or motivated well enough, and there isn't any space left in the paper to address my severe criticisms of it in my first point.  Perhaps if you had dedicated the paper to decaNLP, you could have given arguments that the framing is worthwhile, and described the tasks and their setup as QA sufficiently (as it is, I don't see any description anywhere of how the context is constructed for WikiSQL; did I miss it somewhere?).  For MQAN, there's more than a page of the core new architecture that's pushed into the appendix.  And for the transfer results, there is very little comparison to other transfer methods (e.g., ELMo, CoVe), or any deep analysis of what's going on - as I mentioned above, basically all of the results presented are just confirming what has already been done elsewhere.", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540903833775}], "openreview_url": "https://openreview.net/forum?id=B1lfHhR9tm", "arxiv_id": "1806.08730", "paper_pdf": "papers/B1lfHhR9tm.pdf", "paper_pdf_sha256": "16f82a5a5549b2c93afc4664e34d376ae7cf874f7e74f326cc62b7d5c703b708", "paper_pdf_bytes": 1268248, "paper_pdf_source": "openreview", "code_url": "https://github.com/salesforce/decaNLP", "code_repository": "salesforce/decaNLP", "code_commit": "f1d474b0ff7c7c45a325177401d46b8d0dd16b38", "code_archive": "repos/B1lfHhR9tm.zip", "code_archive_sha256": "c3bad1fc03bc37bbe341c4f6c3b5998c60b9936470c8794526eb709b844c504e", "code_archive_bytes": 738498, "code_file_count": 59, "code_extensions": {".py": 56, ".sh": 3}, "github_disk_usage_kb": 849, "github_languages": {"Python": 381337, "Shell": 2794, "CSS": 2303, "Makefile": 1036, "Batchfile": 817}, "github_archived": true, "github_pushed_at": "2025-05-01T17:29:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-natural-language-decathlon-multitask"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dw2vxWVrA9", "year": 2026, "status": "rejected", "title": "Learning to Dissipate Energy in Oscillatory State-Space Models", "authors": ["Jared Boyer", "T. Konstantin Rusch", "Daniela Rus"], "authorids": ["~Jared_Boyer1", "~T._Konstantin_Rusch1", "~Daniela_Rus1"], "authors_source": "OpenReview API", "abstract": "State-space models (SSMs) are a class of networks for sequence learning that benefit from fixed state size and linear complexity with respect to sequence length, contrasting the quadratic scaling of typical attention mechanisms.  Inspired from observations in neuroscience, Linear Oscillatory State-Space models (LinOSS) are a recently proposed class of SSMs constructed from layers of discretized forced harmonic oscillators.  Although these models perform competitively, leveraging fast parallel scans over diagonal recurrent matrices and achieving state-of-the-art performance on tasks with sequence length up to 50k, LinOSS models rely on rigid energy dissipation (``forgetting'') mechanisms that are inherently coupled to the time scale of state evolution.  As forgetting is a crucial mechanism for long-range reasoning, we demonstrate the representational limitations of these models and introduce Damped Linear Oscillatory State-Space models (D-LinOSS), a more general class of oscillatory SSMs that learn to dissipate latent state energy on arbitrary time scales.  We analyze the spectral distribution of the model's recurrent matrices and prove that the SSM layers exhibit stable dynamics under a simple, flexible parameterization. Without additional complexity, D-LinOSS consistently outperforms previous LinOSS methods on long-range learning tasks, achieves faster convergence, and relinquishes the need for multiple discretization schemes.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "4orC16zPon", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22097/Reviewer_F45N"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "In their paper, the authors build upon LinOSS and extend the previous physics-inspired oscillatory state-space model by introducing learnable damping. They identify limitations in existing oscillatory SSMs and theoretically motivate learnable damping. The authors perform a small synthetic experiment to learn exponentially decaying functions, and show that D-LinOSS outperforms LinOSS across eight different real-world learning tasks. The authors derive theoretical proofs of the improved representational capacity of D-LinOSS, while maintaining computational efficiency.", "review_text": "In their paper, the authors build upon LinOSS and extend the previous physics-inspired oscillatory state-space model by introducing learnable damping. They identify limitations in existing oscillatory SSMs and theoretically motivate learnable damping. The authors perform a small synthetic experiment to learn exponentially decaying functions, and show that D-LinOSS outperforms LinOSS across eight different real-world learning tasks. The authors derive theoretical proofs of the improved representational capacity of D-LinOSS, while maintaining computational efficiency.", "strengths": "- The authors clearly outline the contributions of their paper, highlighting the limitations and improvements over existing oscillatory state-space models (LinOSS).\n- The authors conduct rigorous theoretical analysis of their method and provide a link between the new parameterization to improved representational capacity. They prove that D-LinOSS layers span the full unit disk in the complex plane. \n- The authors provide a controlled synthetic experiment (learning exponential decay) to justify D-LinOSS and an empirical evaluation on more complex, real-world learning tasks.\n- The authors present thorough empirical results outperforming LinOSS across eight datasets with details on hyperparameters, good reproducibility.\n- Figure 1 is very well-designed.", "weaknesses": "- Most reported improvements in the presented empirical results are marginal. I would like to see more ablation on model size/parameter count to be sure of the practical significance of D-LinOSS. \n- While D-LinOSS is well-motivated, the contribution feels like a simple extension of LinOSS. The contribution feels incremental in the broader landscape without additional analysis on the interpretation of learned weights.", "questions": "Are the learned G values interpretable? Do there exist any links to the underlying physical motivation? Can you show qualitative evidence that learned damping captures long-term versus short-term dependencies?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In their paper, the authors build upon LinOSS and extend the previous physics-inspired oscillatory state-space model by introducing learnable damping. They identify limitations in existing oscillatory SSMs and theoretically motivate learnable damping. The authors perform a small synthetic experiment to learn exponentially decaying functions, and show that D-LinOSS outperforms LinOSS across eight different real-world learning tasks. The authors derive theoretical proofs of the improved representational capacity of D-LinOSS, while maintaining computational efficiency.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "- The authors clearly outline the contributions of their paper, highlighting the limitations and improvements over existing oscillatory state-space models (LinOSS).\n- The authors conduct rigorous theoretical analysis of their method and provide a link between the new parameterization to improved representational capacity. They prove that D-LinOSS layers span the full unit disk in the complex plane. \n- The authors provide a controlled synthetic experiment (learning exponential decay) to justify D-LinOSS and an empirical evaluation on more complex, real-world learning tasks.\n- The authors present thorough empirical results outperforming LinOSS across eight datasets with details on hyperparameters, good reproducibility.\n- Figure 1 is very well-designed.", "weaknesses": "- Most reported improvements in the presented empirical results are marginal. I would like to see more ablation on model size/parameter count to be sure of the practical significance of D-LinOSS. \n- While D-LinOSS is well-motivated, the contribution feels like a simple extension of LinOSS. The contribution feels incremental in the broader landscape without additional analysis on the interpretation of learned weights.", "questions": "Are the learned G values interpretable? Do there exist any links to the underlying physical motivation? Can you show qualitative evidence that learned damping captures long-term versus short-term dependencies?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762389881831}, {"id": "AvOf9hOYEy", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22097/Reviewer_Uv1a"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces a new continuous-time recurrent network by extending Linear Oscillatory State-Space models (LinOSS). Specifically, the proposed Damped LinOSS (D-LinOSS) adds a damping term to the continous-time forced second-order ODEs underlying each layer. This damping term enables learnable energy dissipation, allowing the discrete-time model to represent a wider range of stable dynamics than previous discretzied LinOSS models. On average, D-LinOSS outperforms LinOSS (both the IM and IMEX discretizations) and reaches State-of-the-Art performance on the evaluated benchmark tasks.", "review_text": "The paper introduces a new continuous-time recurrent network by extending Linear Oscillatory State-Space models (LinOSS). Specifically, the proposed Damped LinOSS (D-LinOSS) adds a damping term to the continous-time forced second-order ODEs underlying each layer. This damping term enables learnable energy dissipation, allowing the discrete-time model to represent a wider range of stable dynamics than previous discretzied LinOSS models. On average, D-LinOSS outperforms LinOSS (both the IM and IMEX discretizations) and reaches State-of-the-Art performance on the evaluated benchmark tasks.", "strengths": "- The empirical results of D-LinOSS on several real-world datasets are significant, improving State-of-the-Art as well as the previous LinOSS models. \n- The addition of damping to D-LinOSS offers clear practical and theoretical advantages over LinOSS models, enabling eigenvalues that cover the entire unit disc, thus uncoupling frequency and energy dissipation. Table 1 and Figure 3 are particularly compelling illustrations of these advantages.", "weaknesses": "- Please clearly define (and if necessary, differentiate) the terms: energy dissipation, forgetting, damping, and their relation to real and imaginary eigenvalue components, eigenvalue magnitude, and oscillation frequency. Particularly energy dissipation, damping, and forgetting are often used loosely and somewhat interchangeably, which can be confusing. It may help to add a section defining important vocabulary at the beginning of Section 2. \n- The proof of bijectivity in Proposition A.3 is not sufficiently elaborated. Particularly, more details on the exact derivation of the inverse function from eigenvalues to A, G would help convince that it is the true inverse. In addition, classical injectivity-surjectivity arguments would aid in comprehension. \n- Figure 1 is somewhat difficult to parse. Using unnormalized time and different colors for each “frequency” as well as clearly labelling the z axis could go a long way. In addition, the caption should more closely explain the figure and directly reference it. Finally, it might be worth considering including Figure 3 in Figure 1 (or at least more prominently placing Figure 3 since it’s quite easy to understand and clarifies the main contribution of adding learnable damping). \n- The phrase \"reducing the hyperparameter search space by 50%” is a bit misleading (unclear whether you are referring to the number of hyperparameters which are eliminated or the entire search space covered by all hyperparameters...). Maybe rephrase this to clearly state that you are just eliminating the need to choose between LinOSS-IM and LinOSS-IMEX. \n- Even though the empirical results are strong, their interpretation is missing. Please elaborate on possible reasons why D-LinOSS is better at the UEA Motor task while lagging on the Heartbeat task. \n\nIf these weaknesses are sufficiently addressed, I am happy to increase my rating score.", "questions": "- What is meant by the trainability of the timestep parameters in line 124? In the original LinOSS, the timestep is not trainable and fixed. \n- Could you add some more intuition on why LinOSS-IMEX and LinOSS-IM can be universal while still failing to model exponential decay? \n- One of the main reasons to choose LinOSS-IMEX over LinOSS-IM is that it can model conservative systems. How does D-LinOSS compare to LinOSS-IMEX when modelling conservative systems? What solution does D-LinOSS reach (i.e., does it actually learn to set G=0, or does it find some other solution for conservative systems)? \n- Figure 3 gives a clear explanation of why D-LinOSS is better than LinOSS, theoretically. If we plot the actual eigenvalues after training on a real-world task (of your choice), is this behavior verified? This would give more body to the answer to “A natural question is whether or not a larger set of reachable eigenvalues is empirically useful” of line 250. \n- Figure 2 confirms the faster convergence on the synthetic task. Is this benefit also observed on the real-world datasets? A similar figure in the appendix would be a great addition.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new continuous-time recurrent network by extending Linear Oscillatory State-Space models (LinOSS). Specifically, the proposed Damped LinOSS (D-LinOSS) adds a damping term to the continous-time forced second-order ODEs underlying each layer. This damping term enables learnable energy dissipation, allowing the discrete-time model to represent a wider range of stable dynamics than previous discretzied LinOSS models. On average, D-LinOSS outperforms LinOSS (both the IM and IMEX discretizations) and reaches State-of-the-Art performance on the evaluated benchmark tasks.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The empirical results of D-LinOSS on several real-world datasets are significant, improving State-of-the-Art as well as the previous LinOSS models. \n- The addition of damping to D-LinOSS offers clear practical and theoretical advantages over LinOSS models, enabling eigenvalues that cover the entire unit disc, thus uncoupling frequency and energy dissipation. Table 1 and Figure 3 are particularly compelling illustrations of these advantages.", "weaknesses": "- Please clearly define (and if necessary, differentiate) the terms: energy dissipation, forgetting, damping, and their relation to real and imaginary eigenvalue components, eigenvalue magnitude, and oscillation frequency. Particularly energy dissipation, damping, and forgetting are often used loosely and somewhat interchangeably, which can be confusing. It may help to add a section defining important vocabulary at the beginning of Section 2. \n- The proof of bijectivity in Proposition A.3 is not sufficiently elaborated. Particularly, more details on the exact derivation of the inverse function from eigenvalues to A, G would help convince that it is the true inverse. In addition, classical injectivity-surjectivity arguments would aid in comprehension. \n- Figure 1 is somewhat difficult to parse. Using unnormalized time and different colors for each “frequency” as well as clearly labelling the z axis could go a long way. In addition, the caption should more closely explain the figure and directly reference it. Finally, it might be worth considering including Figure 3 in Figure 1 (or at least more prominently placing Figure 3 since it’s quite easy to understand and clarifies the main contribution of adding learnable damping). \n- The phrase \"reducing the hyperparameter search space by 50%” is a bit misleading (unclear whether you are referring to the number of hyperparameters which are eliminated or the entire search space covered by all hyperparameters...). Maybe rephrase this to clearly state that you are just eliminating the need to choose between LinOSS-IM and LinOSS-IMEX. \n- Even though the empirical results are strong, their interpretation is missing. Please elaborate on possible reasons why D-LinOSS is better at the UEA Motor task while lagging on the Heartbeat task. \n\nIf these weaknesses are sufficiently addressed, I am happy to increase my rating score.", "questions": "- What is meant by the trainability of the timestep parameters in line 124? In the original LinOSS, the timestep is not trainable and fixed. \n- Could you add some more intuition on why LinOSS-IMEX and LinOSS-IM can be universal while still failing to model exponential decay? \n- One of the main reasons to choose LinOSS-IMEX over LinOSS-IM is that it can model conservative systems. How does D-LinOSS compare to LinOSS-IMEX when modelling conservative systems? What solution does D-LinOSS reach (i.e., does it actually learn to set G=0, or does it find some other solution for conservative systems)? \n- Figure 3 gives a clear explanation of why D-LinOSS is better than LinOSS, theoretically. If we plot the actual eigenvalues after training on a real-world task (of your choice), is this behavior verified? This would give more body to the answer to “A natural question is whether or not a larger set of reachable eigenvalues is empirically useful” of line 250. \n- Figure 2 confirms the faster convergence on the synthetic task. Is this benefit also observed on the real-world datasets? A similar figure in the appendix would be a great addition.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761921749218}, {"id": "U0IocJbJhq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22097/Reviewer_nwX2"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "This work revisits a very well celebrated work from last year and identifies a shortcoming in the set of equations describing the linear oscillatory state space model (LinOSS). Namely, the set of equations did not involve a damping term, which the authors in this work introduce. They show that with the addition of this term enables faster, efficient, and effective training across various benchmarks.", "review_text": "This work revisits a very well celebrated work from last year and identifies a shortcoming in the set of equations describing the linear oscillatory state space model (LinOSS). Namely, the set of equations did not involve a damping term, which the authors in this work introduce. They show that with the addition of this term enables faster, efficient, and effective training across various benchmarks.", "strengths": "- The addition of the dampening term is a trivial, yet needed, addition that is well motivated and, in this reviewer's opinion, should have been part of the LinOSS to begin with. In this sense, authors have identified an important gap.\n\n- Benchmarks are comprehensive and convincing that the addition of the dampening term is broadly useful; though more is needed to support the specific claims (see below).", "weaknesses": "- The scientific novelty/contribution is quite limited. The main model, i.e. LinOSS, in my opinion is a very important direction that future research should build on. However, the addition here is trivial, and does not necessarily lead to new novel insights. To me, this paper feels like a comment on the original paper as opposed to having enough novelty for deserving a paper of its own.\n\n- I find the claim \"reducing the hyperparameter search space by 50%\" to be slightly misleading. Practically, one should have just simulated an oscillatory model as accurately as possible. In that sense, the choice between implicit or implicit-explicit integration method should not have been presented as an hyperparameter by the original work, since the former is effectively introducing decay that does not exist in real continuous dynamics. In my opinion, the fact that explicitly modeling decay corrects for this \"spurious capability\" resulting from incorrect discretization is a much more interesting contribution as opposed to \"taking away\" one hyperparameter. Thus, I would like to recommend emphasizing this contribution to the scientific methodology as opposed to reduction in hyperparameter count.", "questions": "- Could you please use the same symbols as the LinOSS paper for Eq. 1? It seems the variables x and y are switched, which leads to confusion on the part of the reader.\n\n- It seems in Tables 2 and 3 that LinOSS-Im is more or less at the same level as D-LinOSS? The same may be true for Table 4, though no error bars are reported (for a good reason due to data scarcity, so not authors' fault...).  In that sense, can a devil's advocate say learning to decay is not that important as long as some amount of decay is built-in? For instance, could you please add a baseline in which the decay terms are set equal to some value (defined by the task requirements, or some hyperparameter)? Will D-LinOSS still beat this baseline?\n\n- Looking at Figure 2, it seems that increasing the sequence length makes D-LinOSS more or less the same accuracy and speed as LinOSS-Im? Can you show training for longer sequences? Moreover, similar to above, can you fix the G entries and add this new model as a baseline too?\n\nOverall, my current assessment is as follows:  On the empirical side, more experiments are needed to fully understand when learning the decay timescales is necessary. Theoretically, this work makes an important addition to a very influential model that was given an oral presentation in last year's ICLR conference. However, this addition almost feels like a correction, as opposed to introducing a new capability of the model. In that sense, what makes this model very exciting is already published and very well celebrated by the community. In my assessment, this work could be published as a comment on the original work, for instance in TMLR; but does not satisfy the significance and novelty requirements of ICLR.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work revisits a very well celebrated work from last year and identifies a shortcoming in the set of equations describing the linear oscillatory state space model (LinOSS). Namely, the set of equations did not involve a damping term, which the authors in this work introduce. They show that with the addition of this term enables faster, efficient, and effective training across various benchmarks.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- The addition of the dampening term is a trivial, yet needed, addition that is well motivated and, in this reviewer's opinion, should have been part of the LinOSS to begin with. In this sense, authors have identified an important gap.\n\n- Benchmarks are comprehensive and convincing that the addition of the dampening term is broadly useful; though more is needed to support the specific claims (see below).", "weaknesses": "- The scientific novelty/contribution is quite limited. The main model, i.e. LinOSS, in my opinion is a very important direction that future research should build on. However, the addition here is trivial, and does not necessarily lead to new novel insights. To me, this paper feels like a comment on the original paper as opposed to having enough novelty for deserving a paper of its own.\n\n- I find the claim \"reducing the hyperparameter search space by 50%\" to be slightly misleading. Practically, one should have just simulated an oscillatory model as accurately as possible. In that sense, the choice between implicit or implicit-explicit integration method should not have been presented as an hyperparameter by the original work, since the former is effectively introducing decay that does not exist in real continuous dynamics. In my opinion, the fact that explicitly modeling decay corrects for this \"spurious capability\" resulting from incorrect discretization is a much more interesting contribution as opposed to \"taking away\" one hyperparameter. Thus, I would like to recommend emphasizing this contribution to the scientific methodology as opposed to reduction in hyperparameter count.", "questions": "- Could you please use the same symbols as the LinOSS paper for Eq. 1? It seems the variables x and y are switched, which leads to confusion on the part of the reader.\n\n- It seems in Tables 2 and 3 that LinOSS-Im is more or less at the same level as D-LinOSS? The same may be true for Table 4, though no error bars are reported (for a good reason due to data scarcity, so not authors' fault...).  In that sense, can a devil's advocate say learning to decay is not that important as long as some amount of decay is built-in? For instance, could you please add a baseline in which the decay terms are set equal to some value (defined by the task requirements, or some hyperparameter)? Will D-LinOSS still beat this baseline?\n\n- Looking at Figure 2, it seems that increasing the sequence length makes D-LinOSS more or less the same accuracy and speed as LinOSS-Im? Can you show training for longer sequences? Moreover, similar to above, can you fix the G entries and add this new model as a baseline too?\n\nOverall, my current assessment is as follows:  On the empirical side, more experiments are needed to fully understand when learning the decay timescales is necessary. Theoretically, this work makes an important addition to a very influential model that was given an oral presentation in last year's ICLR conference. However, this addition almost feels like a correction, as opposed to introducing a new capability of the model. In that sense, what makes this model very exciting is already published and very well celebrated by the community. In my assessment, this work could be published as a comment on the original work, for instance in TMLR; but does not satisfy the significance and novelty requirements of ICLR.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761887949442}, {"id": "RxokIsjjpH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22097/Reviewer_h2wS"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces the Damped Linear Oscillatory State-Space Model (D-LinOSS), an extension of the Linear Oscillatory State-Space Model (LinOSS), which models sequence data through layers of discretized forced harmonic oscillators. While previous LinOSS models coupled frequency and damping, restricting energy dissipation to a single time scale, D-LinOSS introduces learnable damping parameters that allow the system to adaptively dissipate energy on arbitrary time scales.", "review_text": "This paper introduces the Damped Linear Oscillatory State-Space Model (D-LinOSS), an extension of the Linear Oscillatory State-Space Model (LinOSS), which models sequence data through layers of discretized forced harmonic oscillators. While previous LinOSS models coupled frequency and damping, restricting energy dissipation to a single time scale, D-LinOSS introduces learnable damping parameters that allow the system to adaptively dissipate energy on arbitrary time scales.", "strengths": "This paper presents a solid and well-motivated contribution to the theory and design of oscillatory state-space models. By introducing learnable damping into the Linear Oscillatory State-Space Model (LinOSS), the proposed D-LinOSS enhances the model’s ability to control forgetting and energy dissipation over arbitrary time scales. This addresses a key limitation of previous LinOSS variants, where frequency and damping were rigidly coupled, restricting flexibility in long-range dynamics. The work is theoretically rigorous, providing a detailed spectral analysis and stability proof that clarifies how the new parameterization expands the set of representable stable systems.\n\nOn the empirical side, the paper supports its theoretical insights with well-designed numerical experiments on synthetic and real-world datasets such as PPG-DaLiA and the Weather forecasting benchmark. These results, though modest in scale, consistently demonstrate that D-LinOSS achieves faster convergence, improved stability, and slightly higher accuracy compared to prior oscillatory SSMs. Moreover, the authors’ discussion of universality—showing that D-LinOSS preserves the universal approximation property of LinOSS—adds conceptual depth and connects the work to broader theoretical frameworks in sequence modeling. Overall, the paper offers a meaningful and well-executed advancement in understanding and improving the forgetting mechanisms of state-space models.", "weaknesses": "The main limitation of the paper lies in its empirical evaluation, which appears relatively underdeveloped given the strength of its theoretical contributions. The experimental results seem under-trained and insufficiently tuned, particularly for the baseline comparisons—there is little evidence that the baselines were optimized to their best performance. This makes it difficult to assess the true effectiveness of the proposed model.\n\nFurthermore, the study is limited to small-scale numerical experiments, leaving open the question of whether the claimed properties—especially those related to stability and forgetting—would generalize to larger or more complex datasets. Without experiments at scale or more diverse benchmarks, the practical impact of the proposed approach remains uncertain. To strengthen the paper, the authors could include carefully tuned baselines, ablation studies, and larger-scale experiments that better validate the theoretical claims in real-world scenarios.", "questions": "Can you provide the detail about the tuning of the numerical experiments provided in the paper?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Damped Linear Oscillatory State-Space Model (D-LinOSS), an extension of the Linear Oscillatory State-Space Model (LinOSS), which models sequence data through layers of discretized forced harmonic oscillators. While previous LinOSS models coupled frequency and damping, restricting energy dissipation to a single time scale, D-LinOSS introduces learnable damping parameters that allow the system to adaptively dissipate energy on arbitrary time scales.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "This paper presents a solid and well-motivated contribution to the theory and design of oscillatory state-space models. By introducing learnable damping into the Linear Oscillatory State-Space Model (LinOSS), the proposed D-LinOSS enhances the model’s ability to control forgetting and energy dissipation over arbitrary time scales. This addresses a key limitation of previous LinOSS variants, where frequency and damping were rigidly coupled, restricting flexibility in long-range dynamics. The work is theoretically rigorous, providing a detailed spectral analysis and stability proof that clarifies how the new parameterization expands the set of representable stable systems.\n\nOn the empirical side, the paper supports its theoretical insights with well-designed numerical experiments on synthetic and real-world datasets such as PPG-DaLiA and the Weather forecasting benchmark. These results, though modest in scale, consistently demonstrate that D-LinOSS achieves faster convergence, improved stability, and slightly higher accuracy compared to prior oscillatory SSMs. Moreover, the authors’ discussion of universality—showing that D-LinOSS preserves the universal approximation property of LinOSS—adds conceptual depth and connects the work to broader theoretical frameworks in sequence modeling. Overall, the paper offers a meaningful and well-executed advancement in understanding and improving the forgetting mechanisms of state-space models.", "weaknesses": "The main limitation of the paper lies in its empirical evaluation, which appears relatively underdeveloped given the strength of its theoretical contributions. The experimental results seem under-trained and insufficiently tuned, particularly for the baseline comparisons—there is little evidence that the baselines were optimized to their best performance. This makes it difficult to assess the true effectiveness of the proposed model.\n\nFurthermore, the study is limited to small-scale numerical experiments, leaving open the question of whether the claimed properties—especially those related to stability and forgetting—would generalize to larger or more complex datasets. Without experiments at scale or more diverse benchmarks, the practical impact of the proposed approach remains uncertain. To strengthen the paper, the authors could include carefully tuned baselines, ablation studies, and larger-scale experiments that better validate the theoretical claims in real-world scenarios.", "questions": "Can you provide the detail about the tuning of the numerical experiments provided in the paper?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761015885707}], "openreview_url": "https://openreview.net/forum?id=dw2vxWVrA9", "arxiv_id": "2505.12171", "paper_pdf": "papers/dw2vxWVrA9.pdf", "paper_pdf_sha256": "b96d986d489a6dcacc251e02dede1519b338c4b5e3e1c0e3ed878bc364d26ff3", "paper_pdf_bytes": 5603773, "paper_pdf_source": "openreview", "code_url": "https://github.com/jaredbmit/damped-linoss", "code_repository": "jaredbmit/damped-linoss", "code_commit": "450b546f693918fe7cfe44082e88538fb29fbd64", "code_archive": "repos/dw2vxWVrA9.zip", "code_archive_sha256": "e5c7b629bccac05837c9e605269110141ae866b62ea8e491f110a2a2de1bbcc1", "code_archive_bytes": 136532, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 453, "github_languages": {"Python": 156887}, "github_archived": false, "github_pushed_at": "2026-02-13T19:58:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-dissipate-energy-in-oscillatory"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZtvRqm6oBu", "year": 2025, "status": "rejected", "title": "Applying Sparse Autoencoders to Unlearn Knowledge in Language Models", "authors": ["Eoin Farrell", "Yeu-Tong Lau", "Arthur Conmy"], "authorids": ["~Eoin_Farrell1", "~Yeu-Tong_Lau1", "~Arthur_Conmy1"], "authors_source": "OpenReview API", "abstract": "We investigate whether sparse autoencoders (SAEs) can be used to remove knowledge from language models. We use the biology subset of the Weapons of Mass Destruction Proxy dataset and test on the gemma-2b-it and gemma-2-2b-it language models. We demonstrate that individual interpretable biology-related SAE features can be used to unlearn a subset of WMDP-Bio questions with minimal side-effects in domains other than biology. Our results suggest that negative scaling of feature activations is necessary and that zero ablating features is ineffective. We find that intervening using multiple SAE features simultaneously can unlearn multiple different topics, but with similar or larger unwanted side-effects than the existing Representation Misdirection for Unlearning technique. Current SAE quality or intervention techniques would need to improve to make SAE-based unlearning comparable to the existing fine-tuning based techniques.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "IUYQE4iTjJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9994/Reviewer_C6qk"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper uses sparse autoencoders to identify features related to biology, which they isolate and clamp down to unlearn on a subset of the WMDP dataset. They use two versions of gemma-2b-it, intervening at an intermediate layer on specific bio-related features. They find that such an intervention can be successful in unlearning, while adding relatively small loss on a generic text dataset.", "review_text": "The paper uses sparse autoencoders to identify features related to biology, which they isolate and clamp down to unlearn on a subset of the WMDP dataset. They use two versions of gemma-2b-it, intervening at an intermediate layer on specific bio-related features. They find that such an intervention can be successful in unlearning, while adding relatively small loss on a generic text dataset.", "strengths": "I think applying SAEs to this task is useful – for us to do good unlearning we almost certainly want an interpretable method, so these are worthwhile first steps. I like the depth you went into in Section 3, as well as Figure 2. I think the methodology was clearly defined, as well as the metrics and tasks you were evaluating. I think there are some nice ablations as well, e.g., Section 4.2.", "weaknesses": "I think the messaging of the paper needs to change to increase the novelty by emphasizing how your work and existing work (RMU) differ. Specifically, how yours is more interpretable and why that’s a good thing. I know the latter is mentioned in the intro but that’s the most substantive discussion of this difference, which is the main reason right now people would want to use SAEs for unlearning.\n\nI think the experiments section should include both gemma models. Relatedly, I think Figure 4 is weak and makes the results hard to interpret. I also don’t really know why the added loss would matter much when you show such a relatively large deterioration on MMLU.\n\nI’d prefer you to show at least some results on even one other subset of WMDP just to see how this generalizes. I wonder if SAEs may be more helpful than fine-tuning if we are trying to unlearn a combination of separate tasks, e.g., both biology and chemistry.\n\nFinally, there is a lack of a related work section, which I feel is necessary as I think you could better contextualize your, e.g., by further showing how people are currently using SAEs to adjust features.", "questions": "-\tWhy did you select the hyperparameters you did for the clamped feature activations (1x, 10x, 50x, etc.)? Same for RMU hyperparameters? Specifically, the layers.\n-\tWhy did you use 300x in Figure 5 but not Figure 4?\n-\tOn page 2, “We trained SAEs at several intermediate layers of the residual stream” – which layers? Sometimes you use layer 3, sometimes layer 9.\n-\tWhy do you use gemma-2b-it for Section 3, but don’t present results for gemma-2b-it in the main body? There is a disconnect between this and Section 4, which uses gemma-2-2b-it.\n-\tOn page 5, there are experiments mentioned but I’d like to see the results somewhere: “To investigate the importance of the particular feature that we selected, we performed the same ablation on a variety of features that activate on this prompt, chosen at random.”\n-\tFigure 1 feels bare. You could consider including some results in half of this figure or some examples on the bio WMDP subset so it is more tailored to your goal instead of a somewhat bare depiction of a SAE.\n\nThese did not affect my review, but some smaller things to note:\n\n-\tCould you use different symbols to represent the feature number and question? You use “#” for both feature number and question number (e.g., “feature #9163” and “question #841”), which I think is distracting.\n-\tIn the conclusion you say intervening using “10-20” features – I would change it to 10 or 20 because you don’t use any intermediate values in the range.\n-\tOn page 4, “The model still provides “A” as the correct answer with probably” -> “probability”\n-\tOn page 6, I would explicitly define what $L0$ means in $L0\\approx 59$ for readers are less familiar with SAEs.\n-\tFigure 5 you should use a different color for the random decoder vector as in Figure 4 you use that some color for representing a 20 feature intervention.\n-\tOn page 8, you say “We propose four key directions for future research” but only provide three.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper uses sparse autoencoders to identify features related to biology, which they isolate and clamp down to unlearn on a subset of the WMDP dataset. They use two versions of gemma-2b-it, intervening at an intermediate layer on specific bio-related features. They find that such an intervention can be successful in unlearning, while adding relatively small loss on a generic text dataset.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "I think applying SAEs to this task is useful – for us to do good unlearning we almost certainly want an interpretable method, so these are worthwhile first steps. I like the depth you went into in Section 3, as well as Figure 2. I think the methodology was clearly defined, as well as the metrics and tasks you were evaluating. I think there are some nice ablations as well, e.g., Section 4.2.", "weaknesses": "I think the messaging of the paper needs to change to increase the novelty by emphasizing how your work and existing work (RMU) differ. Specifically, how yours is more interpretable and why that’s a good thing. I know the latter is mentioned in the intro but that’s the most substantive discussion of this difference, which is the main reason right now people would want to use SAEs for unlearning.\n\nI think the experiments section should include both gemma models. Relatedly, I think Figure 4 is weak and makes the results hard to interpret. I also don’t really know why the added loss would matter much when you show such a relatively large deterioration on MMLU.\n\nI’d prefer you to show at least some results on even one other subset of WMDP just to see how this generalizes. I wonder if SAEs may be more helpful than fine-tuning if we are trying to unlearn a combination of separate tasks, e.g., both biology and chemistry.\n\nFinally, there is a lack of a related work section, which I feel is necessary as I think you could better contextualize your, e.g., by further showing how people are currently using SAEs to adjust features.", "questions": "-\tWhy did you select the hyperparameters you did for the clamped feature activations (1x, 10x, 50x, etc.)? Same for RMU hyperparameters? Specifically, the layers.\n-\tWhy did you use 300x in Figure 5 but not Figure 4?\n-\tOn page 2, “We trained SAEs at several intermediate layers of the residual stream” – which layers? Sometimes you use layer 3, sometimes layer 9.\n-\tWhy do you use gemma-2b-it for Section 3, but don’t present results for gemma-2b-it in the main body? There is a disconnect between this and Section 4, which uses gemma-2-2b-it.\n-\tOn page 5, there are experiments mentioned but I’d like to see the results somewhere: “To investigate the importance of the particular feature that we selected, we performed the same ablation on a variety of features that activate on this prompt, chosen at random.”\n-\tFigure 1 feels bare. You could consider including some results in half of this figure or some examples on the bio WMDP subset so it is more tailored to your goal instead of a somewhat bare depiction of a SAE.\n\nThese did not affect my review, but some smaller things to note:\n\n-\tCould you use different symbols to represent the feature number and question? You use “#” for both feature number and question number (e.g., “feature #9163” and “question #841”), which I think is distracting.\n-\tIn the conclusion you say intervening using “10-20” features – I would change it to 10 or 20 because you don’t use any intermediate values in the range.\n-\tOn page 4, “The model still provides “A” as the correct answer with probably” -> “probability”\n-\tOn page 6, I would explicitly define what $L0$ means in $L0\\approx 59$ for readers are less familiar with SAEs.\n-\tFigure 5 you should use a different color for the random decoder vector as in Figure 4 you use that some color for representing a 20 feature intervention.\n-\tOn page 8, you say “We propose four key directions for future research” but only provide three.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731131139744}, {"id": "BYifGiXeWy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9994/Reviewer_tFqg"], "rating": 5, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "This paper tests the use of sparse autoencoders (SAEs) for unlearning specific knowledge in LLMs, their application of interest is the biosecurity-related Weapons of Mass Destruction Proxy dataset. The authors selectively suppresses SAE features that are highly used by the WMDP dataset (i.e. figure 2). Experiments compare SAE-based unlearning with Representation Misdirection for Unlearning methods, they test accuracy and side effect on the OpenWebText dataset. The authors report effective unlearning with some success in minimizing side effects, though challenges remain with both interpretability and the general applicability of SAEs for unlearning.", "review_text": "This paper tests the use of sparse autoencoders (SAEs) for unlearning specific knowledge in LLMs, their application of interest is the biosecurity-related Weapons of Mass Destruction Proxy dataset. The authors selectively suppresses SAE features that are highly used by the WMDP dataset (i.e. figure 2). Experiments compare SAE-based unlearning with Representation Misdirection for Unlearning methods, they test accuracy and side effect on the OpenWebText dataset. The authors report effective unlearning with some success in minimizing side effects, though challenges remain with both interpretability and the general applicability of SAEs for unlearning.", "strengths": "* Unlearning\n* Minimal side effects", "weaknesses": "* Not concise: I still don't fully understand what a SAE is. The entire paper proposes a new methodological framework without a single math equation. It is very hard to follow and reads like a conversation between LLM software engineers moreso than a technical report on a new methodology.\n* Plots everywhere: why is figure 10 cited an entire page before figure 2? The figures should be placed in close proximity to the text in which it is being discussed\n* What causes the drop on OpenWebText? What datapoints do you lose performance on? How many datapoints do you misclassify? What type of questions are they?\n* It might just be unlearning biology related content. There should be more focus on testing biology related content.", "questions": "Figure 2: the figure says it's a distribution, but it has counts on the y-axis. Also, it does not seem to be normalized per dataset, i.e. the OpenWebText is a lot smaller, why is that? Also, why is figure 2 on page 2 instead of next to the text it's mentioned in. Please fix that it makes it hard to follow the storyline.\n\n\"Interestingly, the model modified using RMU answers option “A” on 62% of questions,\ncompared to 25% for the base model.\" I dont understand what that means.\n\n\"2.4 HOW TO SELECT RELEVANT SAE FEATURES\" by this point in the paper I still have no idea what a SAE feature is and this section is unintuitive to me. In general, until this point, the paper has been very verbose and has had limited conciseness.\n\nSection 3:\nThe entire section has heavy use of DL/LLM lingo and explains the methodology using \"math-intuition\". This is, in fact, quite unintuitive to me. Please rewrite and be concise about the method you are proposing. I am sure you can do it in much less space.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tests the use of sparse autoencoders (SAEs) for unlearning specific knowledge in LLMs, their application of interest is the biosecurity-related Weapons of Mass Destruction Proxy dataset. The authors selectively suppresses SAE features that are highly used by the WMDP dataset (i.e. figure 2). Experiments compare SAE-based unlearning with Representation Misdirection for Unlearning methods, they test accuracy and side effect on the OpenWebText dataset. The authors report effective unlearning with some success in minimizing side effects, though challenges remain with both interpretability and the general applicability of SAEs for unlearning.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "* Unlearning\n* Minimal side effects", "weaknesses": "* Not concise: I still don't fully understand what a SAE is. The entire paper proposes a new methodological framework without a single math equation. It is very hard to follow and reads like a conversation between LLM software engineers moreso than a technical report on a new methodology.\n* Plots everywhere: why is figure 10 cited an entire page before figure 2? The figures should be placed in close proximity to the text in which it is being discussed\n* What causes the drop on OpenWebText? What datapoints do you lose performance on? How many datapoints do you misclassify? What type of questions are they?\n* It might just be unlearning biology related content. There should be more focus on testing biology related content.", "questions": "Figure 2: the figure says it's a distribution, but it has counts on the y-axis. Also, it does not seem to be normalized per dataset, i.e. the OpenWebText is a lot smaller, why is that? Also, why is figure 2 on page 2 instead of next to the text it's mentioned in. Please fix that it makes it hard to follow the storyline.\n\n\"Interestingly, the model modified using RMU answers option “A” on 62% of questions,\ncompared to 25% for the base model.\" I dont understand what that means.\n\n\"2.4 HOW TO SELECT RELEVANT SAE FEATURES\" by this point in the paper I still have no idea what a SAE feature is and this section is unintuitive to me. In general, until this point, the paper has been very verbose and has had limited conciseness.\n\nSection 3:\nThe entire section has heavy use of DL/LLM lingo and explains the methodology using \"math-intuition\". This is, in fact, quite unintuitive to me. Please rewrite and be concise about the method you are proposing. I am sure you can do it in much less space.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730664419729}, {"id": "3L2BUc8Wng", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9994/Reviewer_zKVH"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper explores the potential of using sparse autoencoders (SAEs) to remove specific types of knowledge from language models in an interpretable way. Specifically, it investigates whether SAEs can selectively \"unlearn\" harmful biology-related information from two language models, gemma-2b-it and gemma-2-2b-it, using a subset of Weapons of Mass Destruction Proxy dataset. The main findings suggest that adjusting (negatively scaling) certain biology-related feature activations is effective for unlearning this knowledge, whereas simply zeroing out features is not effective.\n\nKey insights include:\n1. Negative scaling of feature activations is necessary for unlearning specific topics, while zeroing features is ineffective.\n2. Multiple SAE features can be manipulated simultaneously to unlearn several topics, but this method has side effects similar to, or greater than, those of the existing fine-tuning-based Representation Misdirection for Unlearning technique.", "review_text": "The paper explores the potential of using sparse autoencoders (SAEs) to remove specific types of knowledge from language models in an interpretable way. Specifically, it investigates whether SAEs can selectively \"unlearn\" harmful biology-related information from two language models, gemma-2b-it and gemma-2-2b-it, using a subset of Weapons of Mass Destruction Proxy dataset. The main findings suggest that adjusting (negatively scaling) certain biology-related feature activations is effective for unlearning this knowledge, whereas simply zeroing out features is not effective.\n\nKey insights include:\n1. Negative scaling of feature activations is necessary for unlearning specific topics, while zeroing features is ineffective.\n2. Multiple SAE features can be manipulated simultaneously to unlearn several topics, but this method has side effects similar to, or greater than, those of the existing fine-tuning-based Representation Misdirection for Unlearning technique.", "strengths": "- The paper addresses an important and current issue in AI safety, focusing on controlled knowledge removal.\n- The authors present a thorough analysis of how individual SAE features can be targeted to unlearn specific knowledge, showcasing the possibility for precise, fine-grained control.", "weaknesses": "- Novelty: I am not sure about the difference between the paper’s method and negative activation in [1] and [2].\n- Unlearning Performance: The submission underperforms relative to existing unlearning methods, notably RMU, as benchmarked by the WMDP. While it proposes an innovative approach, it fails to deliver superior results compared to RMU across several metrics, raising concerns about its effectiveness and relevance in high-stakes applications.\n- Helpfulness: Furthermore, the exact MMLU accuracy of the Gemma models in absolute terms is unclear, although it appears to be close to random (approximately 25%). The authors should clarify the selection criteria, the number of questions in the MMLU subset, and whether the overall performance of the model improves or worsens.\n- Validity of Evaluation Method: The evaluation is limited to small subset of the WMDP dataset (300 questions, or less than 8% of the full dataset), which diminishes the credibility of its results. Expanding the evaluation to other subsets, such as Chemistry and Cybersecurity, would provide a more robust measure of generalizability. The choice of “Selected MMLU” for assessing unlearning is also problematic; the subset itself is potentially biased toward knowledge that is resistant to permutation, which complicates unlearning and evaluation.\n- Explainability: The negative scaling approach in the submission should be explained more. According to the monosemanticity principle, zeroing the feature activation should be enough to suppress targeted knowledge, since features are believed to be untangled. It’s unclear why stronger negative values would yield better suppression. The lack of transparency on what negative values signify and whether the feature activation is one-dimensional undermines the plausibility of the approach.\n\n\n[1] Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html\n\n[2] Scaling and evaluating sparse autoencoders https://cdn.openai.com/papers/sparse-autoencoders.pdf", "questions": "- **On \"Perfectly Unlearned Models\"**: The paper should clarify why a perfectly unlearned model should achieve a score below 6, as this criterion feels arbitrary without supporting rationale.\n- **On the Negative Value in Feature Suppression**: A deeper analysis of the meaning and impact of negative scaling in feature suppression would add clarity to the method's robustness claims.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores the potential of using sparse autoencoders (SAEs) to remove specific types of knowledge from language models in an interpretable way. Specifically, it investigates whether SAEs can selectively \"unlearn\" harmful biology-related information from two language models, gemma-2b-it and gemma-2-2b-it, using a subset of Weapons of Mass Destruction Proxy dataset. The main findings suggest that adjusting (negatively scaling) certain biology-related feature activations is effective for unlearning this knowledge, whereas simply zeroing out features is not effective.\n\nKey insights include:\n1. Negative scaling of feature activations is necessary for unlearning specific topics, while zeroing features is ineffective.\n2. Multiple SAE features can be manipulated simultaneously to unlearn several topics, but this method has side effects similar to, or greater than, those of the existing fine-tuning-based Representation Misdirection for Unlearning technique.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper addresses an important and current issue in AI safety, focusing on controlled knowledge removal.\n- The authors present a thorough analysis of how individual SAE features can be targeted to unlearn specific knowledge, showcasing the possibility for precise, fine-grained control.", "weaknesses": "- Novelty: I am not sure about the difference between the paper’s method and negative activation in [1] and [2].\n- Unlearning Performance: The submission underperforms relative to existing unlearning methods, notably RMU, as benchmarked by the WMDP. While it proposes an innovative approach, it fails to deliver superior results compared to RMU across several metrics, raising concerns about its effectiveness and relevance in high-stakes applications.\n- Helpfulness: Furthermore, the exact MMLU accuracy of the Gemma models in absolute terms is unclear, although it appears to be close to random (approximately 25%). The authors should clarify the selection criteria, the number of questions in the MMLU subset, and whether the overall performance of the model improves or worsens.\n- Validity of Evaluation Method: The evaluation is limited to small subset of the WMDP dataset (300 questions, or less than 8% of the full dataset), which diminishes the credibility of its results. Expanding the evaluation to other subsets, such as Chemistry and Cybersecurity, would provide a more robust measure of generalizability. The choice of “Selected MMLU” for assessing unlearning is also problematic; the subset itself is potentially biased toward knowledge that is resistant to permutation, which complicates unlearning and evaluation.\n- Explainability: The negative scaling approach in the submission should be explained more. According to the monosemanticity principle, zeroing the feature activation should be enough to suppress targeted knowledge, since features are believed to be untangled. It’s unclear why stronger negative values would yield better suppression. The lack of transparency on what negative values signify and whether the feature activation is one-dimensional undermines the plausibility of the approach.\n\n\n[1] Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html\n\n[2] Scaling and evaluating sparse autoencoders https://cdn.openai.com/papers/sparse-autoencoders.pdf", "questions": "- **On \"Perfectly Unlearned Models\"**: The paper should clarify why a perfectly unlearned model should achieve a score below 6, as this criterion feels arbitrary without supporting rationale.\n- **On the Negative Value in Feature Suppression**: A deeper analysis of the meaning and impact of negative scaling in feature suppression would add clarity to the method's robustness claims.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730599274362}, {"id": "4VHDgBV0Dq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9994/Reviewer_FKb1"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper investigates the use of Sparse Autoencoders (SAEs) to selectively unlearn specific knowledge within language models, using interpretable interventions. The study focuses on the Gemma-2b-it and Gemma-2-2b-it models, particularly on knowledge related to biosecurity from the Weapons of Mass Destruction Proxy Dataset (WMDP-bio).", "review_text": "This paper investigates the use of Sparse Autoencoders (SAEs) to selectively unlearn specific knowledge within language models, using interpretable interventions. The study focuses on the Gemma-2b-it and Gemma-2-2b-it models, particularly on knowledge related to biosecurity from the Weapons of Mass Destruction Proxy Dataset (WMDP-bio).", "strengths": "- Overall, applying SAE to unlearn specific knowledge in LLMs is an interesting and practical approach.\n- SAE interventions provide precise control over targeted knowledge, enhancing transparency.\n- This method avoids weight modification, offering a novel, activation-based approach to unlearning.", "weaknesses": "- It negatively impacts performance on unrelated domains.\n- The proposed method seems to require re-training each time a new domain needs to be unlearned, along with access to data from that domain.\n- More evaluation results beyond the WMDP-bio dataset would enhance the assessment.", "questions": "- The evaluation appears to be based on multiple-choice question answering. How would the method perform in open-ended question answering?\n- What are the additional training and inference costs of using SAE?\n- Can a trained SAE model be transferred to other large language models ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the use of Sparse Autoencoders (SAEs) to selectively unlearn specific knowledge within language models, using interpretable interventions. The study focuses on the Gemma-2b-it and Gemma-2-2b-it models, particularly on knowledge related to biosecurity from the Weapons of Mass Destruction Proxy Dataset (WMDP-bio).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Overall, applying SAE to unlearn specific knowledge in LLMs is an interesting and practical approach.\n- SAE interventions provide precise control over targeted knowledge, enhancing transparency.\n- This method avoids weight modification, offering a novel, activation-based approach to unlearning.", "weaknesses": "- It negatively impacts performance on unrelated domains.\n- The proposed method seems to require re-training each time a new domain needs to be unlearned, along with access to data from that domain.\n- More evaluation results beyond the WMDP-bio dataset would enhance the assessment.", "questions": "- The evaluation appears to be based on multiple-choice question answering. How would the method perform in open-ended question answering?\n- What are the additional training and inference costs of using SAE?\n- Can a trained SAE model be transferred to other large language models ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730263022775}], "openreview_url": "https://openreview.net/forum?id=ZtvRqm6oBu", "arxiv_id": "2410.19278", "paper_pdf": "papers/ZtvRqm6oBu.pdf", "paper_pdf_sha256": "f057083a13568fd4348195f56362be03be40e9af1e96bd669c22e14723800cf8", "paper_pdf_bytes": 2540107, "paper_pdf_source": "openreview", "code_url": "https://github.com/efarrell1/train_sparse_autoencoder", "code_repository": "efarrell1/train_sparse_autoencoder", "code_commit": "60c756d9bd2eb796736ab6c2d74ffc0acc65b717", "code_archive": "repos/ZtvRqm6oBu.zip", "code_archive_sha256": "ab947d64e46984af824033f39e5128c09850beb0ae9ad0a225a5e4c75415d00c", "code_archive_bytes": 65912, "code_file_count": 18, "code_extensions": {".ipynb": 11, ".py": 7}, "github_disk_usage_kb": 54, "github_languages": {"Jupyter Notebook": 153502, "Python": 131912}, "github_archived": false, "github_pushed_at": "2024-10-07T02:24:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/applying-sparse-autoencoders-to-unlearn"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ICuUgRLp4C", "year": 2024, "status": "rejected", "title": "Learning High-Order Relationships of Brain Regions", "authors": ["Weikang Qiu", "Huangrui Chu", "Selena Wang", "Haolan Zuo", "Xiaoxiao Li", "Yize Zhao", "Zhitao Ying"], "authorids": ["~Weikang_Qiu1", "~Huangrui_Chu1", "~Selena_Wang1", "~Haolan_Zuo1", "~Xiaoxiao_Li1", "~Yize_Zhao1", "~Zhitao_Ying1"], "authors_source": "OpenReview API", "abstract": "Discovering reliable and informative interactions among brain regions from functional magnetic resonance imaging (fMRI) signals is essential in neuroscientific predictions of cognition. Most of the current methods fail to accurately characterize those interactions because they only focus on pairwise connections and overlook the high-order relationships of brain regions. We delve into this problem and argue that these high-order relationships should be maximally informative and minimally redundant (MIMR). However, identifying such high-order relationships is challenging and highly under-explored. Methods that can be tailored to our context are also non-existent. In response to this gap, we propose a novel method named HyBRiD that aims to extract MIMR high-order relationships from fMRI data. HyBRiD employs a Constructor to identify hyperedge structures, and a Weighter to compute a weight for each hyperedge. HyBRiD achieves the MIMR objective through an innovative information bottleneck framework named multi-head drop-bottleneck with theoretical guarantees. Our comprehensive experiments demonstrate the effectiveness of our model. In terms of the quality of hyperedges measured by the CPM metric, our model outperforms the state-of-the-art predictive model by an average of 12.1%.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "gA7GRFf7gV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8333/Reviewer_JozW"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This work proposed a hypergraph inference method based on optimizing the predictive power (in the form of mutual information) of the selected (\"connected by hyperedge\") graph node towards certain labels and the redundancy term. The effectiveness of the inference method was evaluated based on an fMRI condition classification task with comparison to pair-wise connectivity estimation and connectivity-based cognition prediction methods and achieved superior performance.", "review_text": "This work proposed a hypergraph inference method based on optimizing the predictive power (in the form of mutual information) of the selected (\"connected by hyperedge\") graph node towards certain labels and the redundancy term. The effectiveness of the inference method was evaluated based on an fMRI condition classification task with comparison to pair-wise connectivity estimation and connectivity-based cognition prediction methods and achieved superior performance.", "strengths": "Eq. 9 and 10 provide a useful solution for the MRMR-like feature selection problem.", "weaknesses": "Formulating the hyperedge inference problem into a feature selection (MRMR-like) problem is interesting, yet not valid, at least in the context of functional connectivity analysis. Regions that are predictive together towards a certain cognitive condition do not imply that they are functionally connected. It's actually easy to construct a case where two regions have very similar fMRI signals, indicating potential strong functional connectivity, but will not be considered as \"hyperedge connected\" in the presented model as their information is redundant for the prediction.", "questions": "1) How is the p-value calculated for the hyperedge in Fig. 4?\n2) It is recommended to discuss how the DimReduction MLP could be trained with a large number of nodes.\n3) Are the linear-head Fl shared across hyperedges or trained separately for each hyperedge?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposed a hypergraph inference method based on optimizing the predictive power (in the form of mutual information) of the selected (\"connected by hyperedge\") graph node towards certain labels and the redundancy term. The effectiveness of the inference method was evaluated based on an fMRI condition classification task with comparison to pair-wise connectivity estimation and connectivity-based cognition prediction methods and achieved superior performance.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "Eq. 9 and 10 provide a useful solution for the MRMR-like feature selection problem.", "weaknesses": "Formulating the hyperedge inference problem into a feature selection (MRMR-like) problem is interesting, yet not valid, at least in the context of functional connectivity analysis. Regions that are predictive together towards a certain cognitive condition do not imply that they are functionally connected. It's actually easy to construct a case where two regions have very similar fMRI signals, indicating potential strong functional connectivity, but will not be considered as \"hyperedge connected\" in the presented model as their information is redundant for the prediction.", "questions": "1) How is the p-value calculated for the hyperedge in Fig. 4?\n2) It is recommended to discuss how the DimReduction MLP could be trained with a large number of nodes.\n3) Are the linear-head Fl shared across hyperedges or trained separately for each hyperedge?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699598957902}, {"id": "1BoUFvOO3j", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8333/Reviewer_eSJE"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper attempts at capturing multivariate relationships among a set or random variables as would be captured by edges in a hypergraph representation. For that, the paper constructs a differentiable regression model that first applies a set of learnable square (for the number of regions) linear projection with subsequent thresholding of the output - a mask, then applies an MLP to compute a scalar value (weight) for each of the masked subsets of the nodes. The weights are used to produce a scalar value (after an inner product with the output weight vector). This model is trained in a regularized regression manner and the produces features are evaluated in an acceptable feature selection evaluation pipeline using predictive strengths of the features as final evaluations. The approach is applied to a subset of the ABCD dataset.", "review_text": "The paper attempts at capturing multivariate relationships among a set or random variables as would be captured by edges in a hypergraph representation. For that, the paper constructs a differentiable regression model that first applies a set of learnable square (for the number of regions) linear projection with subsequent thresholding of the output - a mask, then applies an MLP to compute a scalar value (weight) for each of the masked subsets of the nodes. The weights are used to produce a scalar value (after an inner product with the output weight vector). This model is trained in a regularized regression manner and the produces features are evaluated in an acceptable feature selection evaluation pipeline using predictive strengths of the features as final evaluations. The approach is applied to a subset of the ABCD dataset.", "strengths": "The paper contains an interesting approach to model building, where an encoder builds clustering - a representation interpretable to human experts. A glass-layer with the partition clearly visible. Potentially, a rewrite of the paper could focus on this part, instead of the unsubstantiated claims about capturing high-order interations", "weaknesses": "1.  **The positioning of the paper is a problem.** The assertion that it captures high-order interaction is not substantiated, even though the feature selection model's entire motivation hinges on this claim. Certainly, the title emphasizes high-order interaction. However, the exact type of high-order interaction that the proposed model captures remains ambiguous. I would suggest considering the following papers, which were mistakenly overlooked. These papers seek to formally define what is being captured before attempting to estimate the interactions:\n    -   Rosas FE, Mediano PA, Gastpar M, Jensen HJ. [Quantifying high-order interdependencies via multivariate extensions of the mutual information](https://journals.aps.org/pre/abstract/10.1103/PhysRevE.100.032305). Physical Review E. 2019 Sep 13;100(3):032305.\n    -   Varley TF, Pope M, Faskowitz J, Sporns O. [Multivariate information theory uncovers synergistic subsystems of the human cerebral cortex](https://www.nature.com/articles/s42003-023-04843-w). Communications biology. 2023 Apr 24;6(1):451.\n    -   Santoro A, Battiston F, Petri G, Amico E. [Higher-order organization of multivariate time series](https://www.nature.com/articles/s41567-022-01852-0). Nature Physics. 2023 Feb;19(2):221-9.\n2.  **The clarity of the writing, with regard to the implementation**, is also inadequate in other sections. If we are estimating a hypergraph, then the edges, or node clusters, should form a cover rather than a partition. However, the regularization of the Mean Squared Error (MSE) used in Equation 11, as well as a preceding statement, both confirm the need for the edges to be disjoint, thereby suggesting a partition. This leads us back to the issue of positioning as it means that what is proposed is a clustering algorithm or the task of finding a partition. It's worth noting that node partitioning can still be conducted to recover high-order interactions, as has been exemplified in the neural imaging context, for instance, here:\n    -   Plis SM, Sui J, Lane T, Roy S, Clark VP, Potluru VK, Huster RJ, Michael A, Sponheim SR, Weisend MP, Calhoun VD. [High-order interactions observed in multi-task intrinsic networks are dominant indicators of aberrant brain function in schizophrenia.](https://www.sciencedirect.com/science/article/pii/S1053811913007970) NeuroImage. 2014 Nov 15;102:35-48.\n3.  Overall, **the approach feels like an ad hoc method** for grouping the feature vectors via their predictive potential for a dependent variable. Even the input features are correlation coefficients. That is, the initial input matrix for each subject is the correlation matrix of which the goal is to subselect rows (or columns, which is equivalent due to symmetry) into K different groups.\n4.  **Comparisons in Table 1 are highly problematic** as well.\n    1.  If the goal is to find high-order relations what does predictive quality of representations has to do with it? Table 1 in my opinion does not belong in a paper on high-order relations.\n    2.  However, if the paper would be rewritten to focus on feature grouping and clustering, this approach may potentially work although not without changes. In this case, the proposed model needs to be compared with other approaches that do clustering or partition of random variables. For example, it seems appropriate to consider comparison with Deep Clustering.\n5.  **Results are confusing.** They do not show individual \"hyperedges\" and analyze the ROIs that have grouped together and explain high-order interactions that grouped them. If the regularization enforces the partition, why are so many clusters overlap per my interpretation of Figures 5 and 4c?", "questions": "Note, I do not think answering my question below can change the problems with the way it is written which lead to experiments not supporting the claims. My comments below are to help future clarity of the work:\n\n1.  The abstract states that CRM measures quality of hyperedges, however, CRM looks like a feature selection protocol that also has recommendations for assessing predictivity of the features. This is a disconnect between what is claimed and what is presented.\n2.  The beginning of the second paragraph of the Introduction needs references to support the claim. The last phrase of that paragraph needs a rewrite \"of the intricate behind brain regions\"\n3.  Section 2.1 \"Inupt\" - Do you mean Input?\n4.  Section 2.2 describes feature selection - why is this tied to high-order interactions. Confusing.\n5.  Section 4.1 what do you mean by \"N-dimensional shallow embedding layer parameterized by\". It would be best if all operations and parameters were clearly defined. I assume this is a linear transformation but that is a guess.\n6.  Hyperedge weighting. MLP is not a sufficient description of the used model. Please also mention the activation function.\n7.  Baselines: \"standard\" method mentioned there is unclear. What is standard? How is it defined?\n8.  Why the model is compared with arbitrary models that solve problems different from the proposed method? How were models shown in the comparison selected?\n9.  Ktena, Li, and Kan papers do not construct hypergraphs and yet used as such in the paper for comparisons. Confusing.\n10. Consider fixing capitalization in your bib file. To do that, you can go over the cited papers in your .bib file and put all words you want to preserve capitalization of in additional curly braces. Like {fMRI}.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper attempts at capturing multivariate relationships among a set or random variables as would be captured by edges in a hypergraph representation. For that, the paper constructs a differentiable regression model that first applies a set of learnable square (for the number of regions) linear projection with subsequent thresholding of the output - a mask, then applies an MLP to compute a scalar value (weight) for each of the masked subsets of the nodes. The weights are used to produce a scalar value (after an inner product with the output weight vector). This model is trained in a regularized regression manner and the produces features are evaluated in an acceptable feature selection evaluation pipeline using predictive strengths of the features as final evaluations. The approach is applied to a subset of the ABCD dataset.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The paper contains an interesting approach to model building, where an encoder builds clustering - a representation interpretable to human experts. A glass-layer with the partition clearly visible. Potentially, a rewrite of the paper could focus on this part, instead of the unsubstantiated claims about capturing high-order interations", "weaknesses": "1.  **The positioning of the paper is a problem.** The assertion that it captures high-order interaction is not substantiated, even though the feature selection model's entire motivation hinges on this claim. Certainly, the title emphasizes high-order interaction. However, the exact type of high-order interaction that the proposed model captures remains ambiguous. I would suggest considering the following papers, which were mistakenly overlooked. These papers seek to formally define what is being captured before attempting to estimate the interactions:\n    -   Rosas FE, Mediano PA, Gastpar M, Jensen HJ. [Quantifying high-order interdependencies via multivariate extensions of the mutual information](https://journals.aps.org/pre/abstract/10.1103/PhysRevE.100.032305). Physical Review E. 2019 Sep 13;100(3):032305.\n    -   Varley TF, Pope M, Faskowitz J, Sporns O. [Multivariate information theory uncovers synergistic subsystems of the human cerebral cortex](https://www.nature.com/articles/s42003-023-04843-w). Communications biology. 2023 Apr 24;6(1):451.\n    -   Santoro A, Battiston F, Petri G, Amico E. [Higher-order organization of multivariate time series](https://www.nature.com/articles/s41567-022-01852-0). Nature Physics. 2023 Feb;19(2):221-9.\n2.  **The clarity of the writing, with regard to the implementation**, is also inadequate in other sections. If we are estimating a hypergraph, then the edges, or node clusters, should form a cover rather than a partition. However, the regularization of the Mean Squared Error (MSE) used in Equation 11, as well as a preceding statement, both confirm the need for the edges to be disjoint, thereby suggesting a partition. This leads us back to the issue of positioning as it means that what is proposed is a clustering algorithm or the task of finding a partition. It's worth noting that node partitioning can still be conducted to recover high-order interactions, as has been exemplified in the neural imaging context, for instance, here:\n    -   Plis SM, Sui J, Lane T, Roy S, Clark VP, Potluru VK, Huster RJ, Michael A, Sponheim SR, Weisend MP, Calhoun VD. [High-order interactions observed in multi-task intrinsic networks are dominant indicators of aberrant brain function in schizophrenia.](https://www.sciencedirect.com/science/article/pii/S1053811913007970) NeuroImage. 2014 Nov 15;102:35-48.\n3.  Overall, **the approach feels like an ad hoc method** for grouping the feature vectors via their predictive potential for a dependent variable. Even the input features are correlation coefficients. That is, the initial input matrix for each subject is the correlation matrix of which the goal is to subselect rows (or columns, which is equivalent due to symmetry) into K different groups.\n4.  **Comparisons in Table 1 are highly problematic** as well.\n    1.  If the goal is to find high-order relations what does predictive quality of representations has to do with it? Table 1 in my opinion does not belong in a paper on high-order relations.\n    2.  However, if the paper would be rewritten to focus on feature grouping and clustering, this approach may potentially work although not without changes. In this case, the proposed model needs to be compared with other approaches that do clustering or partition of random variables. For example, it seems appropriate to consider comparison with Deep Clustering.\n5.  **Results are confusing.** They do not show individual \"hyperedges\" and analyze the ROIs that have grouped together and explain high-order interactions that grouped them. If the regularization enforces the partition, why are so many clusters overlap per my interpretation of Figures 5 and 4c?", "questions": "Note, I do not think answering my question below can change the problems with the way it is written which lead to experiments not supporting the claims. My comments below are to help future clarity of the work:\n\n1.  The abstract states that CRM measures quality of hyperedges, however, CRM looks like a feature selection protocol that also has recommendations for assessing predictivity of the features. This is a disconnect between what is claimed and what is presented.\n2.  The beginning of the second paragraph of the Introduction needs references to support the claim. The last phrase of that paragraph needs a rewrite \"of the intricate behind brain regions\"\n3.  Section 2.1 \"Inupt\" - Do you mean Input?\n4.  Section 2.2 describes feature selection - why is this tied to high-order interactions. Confusing.\n5.  Section 4.1 what do you mean by \"N-dimensional shallow embedding layer parameterized by\". It would be best if all operations and parameters were clearly defined. I assume this is a linear transformation but that is a guess.\n6.  Hyperedge weighting. MLP is not a sufficient description of the used model. Please also mention the activation function.\n7.  Baselines: \"standard\" method mentioned there is unclear. What is standard? How is it defined?\n8.  Why the model is compared with arbitrary models that solve problems different from the proposed method? How were models shown in the comparison selected?\n9.  Ktena, Li, and Kan papers do not construct hypergraphs and yet used as such in the paper for comparisons. Confusing.\n10. Consider fixing capitalization in your bib file. To do that, you can go over the cited papers in your .bib file and put all words you want to preserve capitalization of in additional curly braces. Like {fMRI}.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698894068048}, {"id": "qcLMyTGP44", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8333/Reviewer_yeBt"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work proposes a principle to learn high-order relationships of brain regions -- high-order relationships should be maximally informative and minimally redundant (MIMR), and a method called Hypergraph of Brain Regions via multi-head Drop- bottleneck (HyBRiD) to learn such relationships from fMRI data. HyBRiD includes a constructor to identify hyperedge structures, and a weighter to compute a weight for each hyperedge. The results show that HyBRiD outperformed 8 baseline methods in 7 out of 8 fMRI datasets.", "review_text": "This work proposes a principle to learn high-order relationships of brain regions -- high-order relationships should be maximally informative and minimally redundant (MIMR), and a method called Hypergraph of Brain Regions via multi-head Drop- bottleneck (HyBRiD) to learn such relationships from fMRI data. HyBRiD includes a constructor to identify hyperedge structures, and a weighter to compute a weight for each hyperedge. The results show that HyBRiD outperformed 8 baseline methods in 7 out of 8 fMRI datasets.", "strengths": "1. The paper is well written with clear organization, detailed theoretical explanation, and comprehensive empirical evaluation. \n\n2. The proposed method is intuitively simple yet effective, and could be potentially applied to learn high-order relationships of brain regions with respect to different prediction targets. \n\n3. The neuroimaging experiments were comprehensive. A large sample size (8 datasets with 11875 subjects) was used to evaluate the models. The proposed method HyBRiD was compared to 3 types of baseline methods including 8 methods. HyBRiD outperformed 8 baseline methods in 7 out of 8 datasets. \n\n4. The hyperedge profile analysis indicates interactions of multiple brain regions are more important in cognition tasks. The region importance reveals reasonable task-related brain regions under different conditions.", "weaknesses": "1. The authors mentioned that \"Due to the data scarcity, training on individual datasets would result in serious overfitting.\" Each individual dataset includes at least 1000 subjects but the model performance is still not ideal. Can the model be applied to a dataset with less samples? Most clinical datasets are relatively small with < 1000 subjects. Is it applicable to apply the model on datasets with fewer samples? \n\n2. It would be helpful if the authors could discuss potential reasons why HyBRiD failed at Rest 1 dataset. \n\n3. Region importance. What about region importance for resting state data? If I understand it correctly, region importance is a metric for nodes. What about edges? Can you show how edges are connected under different conditions? \n\n4. I appreciate that the authors include the code in supplemental material, but a README file should be also included to explain how to replicate the results. \n\n5. The notations in Section 5.1 Metric are not very clear to me. \n\n\ti. The input of CPM should include the prediction target Y, right? Maybe use CPM(E, Y)? \n\n\tii. What is the dimension of E? Is E equivalent to H+$\\mathbf{w}$ in HyBRiD? \n\n\tiii. CPM is evaluated on each model separately, right? If so, I think Eq. 13 should be defined separately for HyBRiD.\n\n6. Minor: \n\n\ti. Typos: Section 2.1 \"Inupt\" should be \"Input\"; Section 5.2 \"conducte\" should be \"conduct\"; Figure 7: \"grpahical\" should be \"graphical\".\n\n\tii. CPM should be defined in the abstract and introduction upon its first occurrence.\n\n\tiii. In Section 5.1 Dataset, RS (resting state) should be defined.\n\n\tiv. In Table 5, $\\beta = 0.2$ instead of $0.3$ to be consistent with Section E.3?", "questions": "1. Does the model generalize well across conditions or out-of-sample data? For example, if a model is trained on resting state data, can it be applied to predict task data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes a principle to learn high-order relationships of brain regions -- high-order relationships should be maximally informative and minimally redundant (MIMR), and a method called Hypergraph of Brain Regions via multi-head Drop- bottleneck (HyBRiD) to learn such relationships from fMRI data. HyBRiD includes a constructor to identify hyperedge structures, and a weighter to compute a weight for each hyperedge. The results show that HyBRiD outperformed 8 baseline methods in 7 out of 8 fMRI datasets.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper is well written with clear organization, detailed theoretical explanation, and comprehensive empirical evaluation. \n\n2. The proposed method is intuitively simple yet effective, and could be potentially applied to learn high-order relationships of brain regions with respect to different prediction targets. \n\n3. The neuroimaging experiments were comprehensive. A large sample size (8 datasets with 11875 subjects) was used to evaluate the models. The proposed method HyBRiD was compared to 3 types of baseline methods including 8 methods. HyBRiD outperformed 8 baseline methods in 7 out of 8 datasets. \n\n4. The hyperedge profile analysis indicates interactions of multiple brain regions are more important in cognition tasks. The region importance reveals reasonable task-related brain regions under different conditions.", "weaknesses": "1. The authors mentioned that \"Due to the data scarcity, training on individual datasets would result in serious overfitting.\" Each individual dataset includes at least 1000 subjects but the model performance is still not ideal. Can the model be applied to a dataset with less samples? Most clinical datasets are relatively small with < 1000 subjects. Is it applicable to apply the model on datasets with fewer samples? \n\n2. It would be helpful if the authors could discuss potential reasons why HyBRiD failed at Rest 1 dataset. \n\n3. Region importance. What about region importance for resting state data? If I understand it correctly, region importance is a metric for nodes. What about edges? Can you show how edges are connected under different conditions? \n\n4. I appreciate that the authors include the code in supplemental material, but a README file should be also included to explain how to replicate the results. \n\n5. The notations in Section 5.1 Metric are not very clear to me. \n\n\ti. The input of CPM should include the prediction target Y, right? Maybe use CPM(E, Y)? \n\n\tii. What is the dimension of E? Is E equivalent to H+$\\mathbf{w}$ in HyBRiD? \n\n\tiii. CPM is evaluated on each model separately, right? If so, I think Eq. 13 should be defined separately for HyBRiD.\n\n6. Minor: \n\n\ti. Typos: Section 2.1 \"Inupt\" should be \"Input\"; Section 5.2 \"conducte\" should be \"conduct\"; Figure 7: \"grpahical\" should be \"graphical\".\n\n\tii. CPM should be defined in the abstract and introduction upon its first occurrence.\n\n\tiii. In Section 5.1 Dataset, RS (resting state) should be defined.\n\n\tiv. In Table 5, $\\beta = 0.2$ instead of $0.3$ to be consistent with Section E.3?", "questions": "1. Does the model generalize well across conditions or out-of-sample data? For example, if a model is trained on resting state data, can it be applied to predict task data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698812251391}, {"id": "2jCzZZg8SX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8333/Reviewer_Z7u6"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "This paper proposes a novel method named HYBRID for extracting maximally informative and minimally redundant high-order relationships from fMRI data. The authors argue that most current methods fail to accurately characterize interactions among brain regions because they only focus on pairwise connections and overlook high-order relationships. HYBRID addresses this limitation by constructing a hypergraph where hyperedges represent high-order relationships and their weights represent the strengths of those relationships. The authors demonstrate the effectiveness of HYBRID through comprehensive experiments, outperforming the state-of-the-art predictive model by an average of 12.1%. The contributions of this paper include a novel method for extracting high-order relationships from fMRI data and a comprehensive evaluation of the proposed method.", "review_text": "This paper proposes a novel method named HYBRID for extracting maximally informative and minimally redundant high-order relationships from fMRI data. The authors argue that most current methods fail to accurately characterize interactions among brain regions because they only focus on pairwise connections and overlook high-order relationships. HYBRID addresses this limitation by constructing a hypergraph where hyperedges represent high-order relationships and their weights represent the strengths of those relationships. The authors demonstrate the effectiveness of HYBRID through comprehensive experiments, outperforming the state-of-the-art predictive model by an average of 12.1%. The contributions of this paper include a novel method for extracting high-order relationships from fMRI data and a comprehensive evaluation of the proposed method.", "strengths": "Originality: The paper proposes a novel method named HYBRID for extracting maximally informative and minimally redundant high-order relationships from fMRI data. This is a significant contribution as most current methods focus on pairwise connections and overlook high-order relationships. HYBRID addresses this limitation by constructing a hypergraph where hyperedges represent high-order relationships and their weights represent the strengths of those relationships. The proposed method is original and creative, and the authors provide a comprehensive evaluation of the proposed method, demonstrating its effectiveness through experiments.\n\nQuality & clarity: The paper presents a clear problem formulation, a detailed description of the proposed method, and a comprehensive evaluation of the proposed method. The authors provide theoretical guarantees for the proposed method and demonstrate its effectiveness through experiments. The paper is well-written, with clear and concise language, making it easy to understand.\n\nSignificance: The paper addresses an important problem in neuroscience and machine learning. Discovering reliable and informative interactions among brain regions from fMRI signals is essential in neuroscientific predictions of cognition. Most of the current methods fail to accurately characterize those interactions because they only focus on pairwise connections and overlook the high-order relationships of brain regions. The proposed method addresses this limitation and provides a new approach for extracting high-order relationships from fMRI data.", "weaknesses": "The matrix notation in SEction 4 is confusing:\nSection 4.1, first line: $X = [X_1, X_2, \\dots, X_N]$ should be $X = [X_1, X_2, \\dots, X_N]^T$ (i.e., a transpose operator should be inserted).\nEq. (4): a transpose operator should be inserted after the square brackets.", "questions": "Can you briefly explain how to choose the number of hyperedges?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel method named HYBRID for extracting maximally informative and minimally redundant high-order relationships from fMRI data. The authors argue that most current methods fail to accurately characterize interactions among brain regions because they only focus on pairwise connections and overlook high-order relationships. HYBRID addresses this limitation by constructing a hypergraph where hyperedges represent high-order relationships and their weights represent the strengths of those relationships. The authors demonstrate the effectiveness of HYBRID through comprehensive experiments, outperforming the state-of-the-art predictive model by an average of 12.1%. The contributions of this paper include a novel method for extracting high-order relationships from fMRI data and a comprehensive evaluation of the proposed method.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Originality: The paper proposes a novel method named HYBRID for extracting maximally informative and minimally redundant high-order relationships from fMRI data. This is a significant contribution as most current methods focus on pairwise connections and overlook high-order relationships. HYBRID addresses this limitation by constructing a hypergraph where hyperedges represent high-order relationships and their weights represent the strengths of those relationships. The proposed method is original and creative, and the authors provide a comprehensive evaluation of the proposed method, demonstrating its effectiveness through experiments.\n\nQuality & clarity: The paper presents a clear problem formulation, a detailed description of the proposed method, and a comprehensive evaluation of the proposed method. The authors provide theoretical guarantees for the proposed method and demonstrate its effectiveness through experiments. The paper is well-written, with clear and concise language, making it easy to understand.\n\nSignificance: The paper addresses an important problem in neuroscience and machine learning. Discovering reliable and informative interactions among brain regions from fMRI signals is essential in neuroscientific predictions of cognition. Most of the current methods fail to accurately characterize those interactions because they only focus on pairwise connections and overlook the high-order relationships of brain regions. The proposed method addresses this limitation and provides a new approach for extracting high-order relationships from fMRI data.", "weaknesses": "The matrix notation in SEction 4 is confusing:\nSection 4.1, first line: $X = [X_1, X_2, \\dots, X_N]$ should be $X = [X_1, X_2, \\dots, X_N]^T$ (i.e., a transpose operator should be inserted).\nEq. (4): a transpose operator should be inserted after the square brackets.", "questions": "Can you briefly explain how to choose the number of hyperedges?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698762057860}], "openreview_url": "https://openreview.net/forum?id=ICuUgRLp4C", "arxiv_id": "2312.02203", "paper_pdf": "papers/ICuUgRLp4C.pdf", "paper_pdf_sha256": "55b8b34bd73c861cc819ee501e0a319a97770673ffa4a6b569d06de8a44f2bdf", "paper_pdf_bytes": 2038438, "paper_pdf_source": "openreview", "code_url": "https://github.com/Graph-and-Geometric-Learning/HyBRiD", "code_repository": "Graph-and-Geometric-Learning/HyBRiD", "code_commit": "973d937a38449409264218e380ec6593ee30568c", "code_archive": "repos/ICuUgRLp4C.zip", "code_archive_sha256": "6a588170a4f396e7d6efee6b063e9204381193925cdd11c05646a878545d0f74", "code_archive_bytes": 27411, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 44, "github_languages": {"Python": 35629}, "github_archived": false, "github_pushed_at": "2025-05-29T02:39:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-high-order-relationships-of-brain"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6dtI3qPVwVp", "year": 2023, "status": "rejected", "title": "Link Prediction without Graph Neural Networks", "authors": ["Zexi Huang", "Mert Kosan", "Arlei Lopes da Silva", "Ambuj Singh"], "authorids": ["~Zexi_Huang1", "~Mert_Kosan1", "~Arlei_Lopes_da_Silva1", "~Ambuj_Singh1"], "authors_source": "OpenReview API", "abstract": "Link prediction, which consists of predicting edges based on graph features, is a fundamental task in many graph applications. As for several related problems, Graph Neural Networks (GNNs), which are based on an attribute-centric message-passing paradigm, have become the predominant framework for link prediction. GNNs have consistently outperformed traditional topology-based heuristics, but what contributes to their performance? Are there simpler approaches that achieve comparable or better results? To answer these questions, we first identify important limitations in how GNN-based link prediction methods handle the intrinsic class imbalance of the problem---due to the graph sparsity---in their training and evaluation. Moreover, we propose Gelato, a novel topology-centric framework that applies a topological heuristic to a graph enhanced by attribute information via graph learning. Our model is trained end-to-end with an N-pair loss on an unbiased training set to address class imbalance. Experiments show that Gelato is 145% more accurate, trains 11 times faster, infers 6,000 times faster, and has less than half of the trainable parameters compared to state-of-the-art GNNs for link prediction.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "RAjCD5UCH-", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1014/Reviewer_p64G"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a model for link prediction without graph neural networks. The authors firstly examine existing benchmarks and argue they are biased towards positive links. Then, they propose an effective framework with graph structure learning with a topology heuristic. A series experiments show the superior performance of the proposed approach.", "review_text": "I think this work provides an important contribution to \"graph-less\" neural networks for link prediction, but I feel like the motivation of this work needs further elaboration. Some claims on existing \"biased\" models are too strong.", "strengths": "Strengths:\n+ This paper is easy to read and follow.\n+ The proposed strategy significantly outperforms existing benchmarks and especially improves efficiency.\n\nWeaknesses:\n- Several statements may be too strong to claim. For example: in Page 2 the authors claim existing approaches \"highly overestimate\" the ratio of positive pairs. I feels skeptical about this claim, as the negative pairs are randomly drawn rather than fixed. It is not clear to me why they are biased towards positive pairs. Similarly, I am not clear about why prior approaches need to construct a biased test dataset.\n- I in person believe there is a gap between the motivation and the model design. The only connection I can see is the objective Eq. (16) uses a full set of negative links, where prior approaches can do this as well.\n- Experimental results are not fully convincing. On several experiments (Cora, Citeseer, Pubmed) the improvements seem marginal. The overall dataset sizes are small. It seems that the model is very sensitive to the parameter $\\eta$. More explanations are needed. Also, more head-to-head ablation studies are needed. For example, I am curious whether the graph structure learning component (Attributes+Topology) can be used with conventional training schemes.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents a model for link prediction without graph neural networks. The authors firstly examine existing benchmarks and argue they are biased towards positive links. Then, they propose an effective framework with graph structure learning with a topology heuristic. A series experiments show the superior performance of the proposed approach.", "strength_and_weaknesses": "Strengths:\n+ This paper is easy to read and follow.\n+ The proposed strategy significantly outperforms existing benchmarks and especially improves efficiency.\n\nWeaknesses:\n- Several statements may be too strong to claim. For example: in Page 2 the authors claim existing approaches \"highly overestimate\" the ratio of positive pairs. I feels skeptical about this claim, as the negative pairs are randomly drawn rather than fixed. It is not clear to me why they are biased towards positive pairs. Similarly, I am not clear about why prior approaches need to construct a biased test dataset.\n- I in person believe there is a gap between the motivation and the model design. The only connection I can see is the objective Eq. (16) uses a full set of negative links, where prior approaches can do this as well.\n- Experimental results are not fully convincing. On several experiments (Cora, Citeseer, Pubmed) the improvements seem marginal. The overall dataset sizes are small. It seems that the model is very sensitive to the parameter $\\eta$. More explanations are needed. Also, more head-to-head ablation studies are needed. For example, I am curious whether the graph structure learning component (Attributes+Topology) can be used with conventional training schemes.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and easy to follow. The technical contribution could be improved and several claims need further justification. The authors provide anonymous code but the reviewer does not check the code.", "summary_of_the_review": "I think this work provides an important contribution to \"graph-less\" neural networks for link prediction, but I feel like the motivation of this work needs further elaboration. Some claims on existing \"biased\" models are too strong.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666907977014}, {"id": "H37WYohOFsi", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1014/Reviewer_L8KX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the authors propose an interesting and novel technique towards solving the problem of link prediction in graphs without using GNN based approach which happens to be the predominant methodology at present. The authors also demonstrate training their model using an N-pair loss (akin to metric learning based loss) to address class imbalance in real-world graph datasets. The authors demonstrate the efficacy of their approach via empirical results.", "review_text": "The authors present an interesting and lightweight approach towards solving the problem of link prediction. The main issues with the current work is the limited level of novelty and the heuristic aspects of the work which require more discussion and/or more work. ", "strengths": "Here below are some associated strengths of the paper :\n\n1) Even though graph augmentation based techniques previously exist in the literature, the authors present an interesting perspective of solving the link prediction problem without using GNN explicitly. \n\n2) The authors highlight and make an effort to solve the class imbalance problem in graph datasets which hampers model performance significantly. \n\n3) The approach proposed is a lightweight one which can easily be incorporated and would potentially help model performance in the link prediction problem.\n\n4) The empirical results demonstrate the efficacy of the approach effectively.\n\nHere below are some associated weaknesses of the paper :\n\n1) While there is some merit in the approach presented and the expectation from the reviewer is that this current work might usher in a new field of work, in general the level of novelty of the current approach is incremental in nature and requires more refinement. For example, graph augmentation/enrichment is known in literature as well as metric learning based loss functions as well as the auto covariance based similarity matrix which has demonstrated excellent performance for link prediction based tasks.\n\n2) The approach felt pretty heuristic. In particular the approach is strongly dependent on a similarity metric/function and an associated threshold. The paper does not contain any discussion as to how to choose this similarity metric/function and the associated threshold in general. The authors themselves choose cosine similarity but does this work well in all scenarios. The current work requires more discussion in general. \n\n3) The current work should have included more baselines to compare against. \n\n4) The current work does not list any limitations and/or potential directions for improvement which would have significantly added to the value associated with the work. As mentioned earlier, the current work requires more discussion to be included in general. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this work, the authors propose an interesting and novel technique towards solving the problem of link prediction in graphs without using GNN based approach which happens to be the predominant methodology at present. The authors also demonstrate training their model using an N-pair loss (akin to metric learning based loss) to address class imbalance in real-world graph datasets. The authors demonstrate the efficacy of their approach via empirical results.", "strength_and_weaknesses": "Here below are some associated strengths of the paper :\n\n1) Even though graph augmentation based techniques previously exist in the literature, the authors present an interesting perspective of solving the link prediction problem without using GNN explicitly. \n\n2) The authors highlight and make an effort to solve the class imbalance problem in graph datasets which hampers model performance significantly. \n\n3) The approach proposed is a lightweight one which can easily be incorporated and would potentially help model performance in the link prediction problem.\n\n4) The empirical results demonstrate the efficacy of the approach effectively.\n\nHere below are some associated weaknesses of the paper :\n\n1) While there is some merit in the approach presented and the expectation from the reviewer is that this current work might usher in a new field of work, in general the level of novelty of the current approach is incremental in nature and requires more refinement. For example, graph augmentation/enrichment is known in literature as well as metric learning based loss functions as well as the auto covariance based similarity matrix which has demonstrated excellent performance for link prediction based tasks.\n\n2) The approach felt pretty heuristic. In particular the approach is strongly dependent on a similarity metric/function and an associated threshold. The paper does not contain any discussion as to how to choose this similarity metric/function and the associated threshold in general. The authors themselves choose cosine similarity but does this work well in all scenarios. The current work requires more discussion in general. \n\n3) The current work should have included more baselines to compare against. \n\n4) The current work does not list any limitations and/or potential directions for improvement which would have significantly added to the value associated with the work. As mentioned earlier, the current work requires more discussion to be included in general. ", "clarity,_quality,_novelty_and_reproducibility": "The approach presented in clear in details. The level of novelty is somewhat limited and the proposed approach could be further improved/refined. The authors made publicly available their code which allowed for reproducibility checking.", "summary_of_the_review": "The authors present an interesting and lightweight approach towards solving the problem of link prediction. The main issues with the current work is the limited level of novelty and the heuristic aspects of the work which require more discussion and/or more work. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666852409238}, {"id": "4sh1sI4HzU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1014/Reviewer_SSk7"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors first the important limitations in how GNN-based link prediction methods handle the intrinsic class imbalance of the problem—due to the graph sparsity—in their training and evaluation. Moreover, the authors propose Gelato, a novel topology-centric framework that\napplies a topological heuristic to a graph enhanced by attribute information via graph learning. Gelato is trained end-to-end with an N-pair loss on an unbiased training set to address the class imbalance.", "review_text": "Please refer to the above review comments for improvements.", "strengths": "Pros: the authors scrutinize the training and evaluation of supervised link prediction methods and identify their limitations in handling class imbalance; the authors propose a simple, effective, and efficient framework to combine topological and attribute information for link prediction without using GNNs; and they introduce an N-pair link prediction loss combined with an unbiased set of training edges that we show to be more effective at addressing the class imbalance.\n\nCons: The authors only use some small datasets for evaluation, on which there do not exist standard splits for different methods to be evaluated in a standard manner. The authors are suggested to test their method on larger datasets such as OGBL ones to comprehensively evaluate their method.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors first the important limitations in how GNN-based link prediction methods handle the intrinsic class imbalance of the problem—due to the graph sparsity—in their training and evaluation. Moreover, the authors propose Gelato, a novel topology-centric framework that\napplies a topological heuristic to a graph enhanced by attribute information via graph learning. Gelato is trained end-to-end with an N-pair loss on an unbiased training set to address the class imbalance.", "strength_and_weaknesses": "Pros: the authors scrutinize the training and evaluation of supervised link prediction methods and identify their limitations in handling class imbalance; the authors propose a simple, effective, and efficient framework to combine topological and attribute information for link prediction without using GNNs; and they introduce an N-pair link prediction loss combined with an unbiased set of training edges that we show to be more effective at addressing the class imbalance.\n\nCons: The authors only use some small datasets for evaluation, on which there do not exist standard splits for different methods to be evaluated in a standard manner. The authors are suggested to test their method on larger datasets such as OGBL ones to comprehensively evaluate their method.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and somehow novel. There is a separate section to describe the resources for Reproducibility. ", "summary_of_the_review": "Please refer to the above review comments for improvements.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666670570317}, {"id": "r8S7k4rRcfI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1014/Reviewer_maA8"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper is a re-submission. I have previously reviewed and the authors made only minor changes without incorporating the feedback from their prior submission.\n\nThe paper rightly critises biased testing in homogenous graph link prediction and proposes to combine topological features with MLPs for this problem.", "review_text": "A re-submission that doesn't adress the weaknessess from the last round of feedback (see Weaknessess section). Clear reject.", "strengths": "Weaknesses:\n1. A low-effort re-submission!!\n2. Doesn't cite and compare to the relevant Knowledge graph link prediction literature, which for a long time now uses unbiased testing (i.e. the idea is not new) as well as failing to compare their loss function to negative sampling from the KG community.\n3. The benchmark datasets are simply not interesting anymore and are not sufficient to validate the empirical performance of their method, they should incorporate the homogenous graph link prediction datasets from OGB at least.\n\nStrengths:\n1. Method is efficient.\n2. Clear writing.\n3. Rightly criticises biased testing.\n4. Models that are compared against are representative.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper is a re-submission. I have previously reviewed and the authors made only minor changes without incorporating the feedback from their prior submission.\n\nThe paper rightly critises biased testing in homogenous graph link prediction and proposes to combine topological features with MLPs for this problem.", "strength_and_weaknesses": "Weaknesses:\n1. A low-effort re-submission!!\n2. Doesn't cite and compare to the relevant Knowledge graph link prediction literature, which for a long time now uses unbiased testing (i.e. the idea is not new) as well as failing to compare their loss function to negative sampling from the KG community.\n3. The benchmark datasets are simply not interesting anymore and are not sufficient to validate the empirical performance of their method, they should incorporate the homogenous graph link prediction datasets from OGB at least.\n\nStrengths:\n1. Method is efficient.\n2. Clear writing.\n3. Rightly criticises biased testing.\n4. Models that are compared against are representative.", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\n - Good\n\nQuality\n - Empirical evaluation datasets are poor, models, however, are well chosen\n\nNovelty\n - Low to Medium, the technical innovation is not super significant.\n\nReproducibility\n- Good\n\n", "summary_of_the_review": "A re-submission that doesn't adress the weaknessess from the last round of feedback (see Weaknessess section). Clear reject.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject"}, "tcdate": 1666657847467}], "openreview_url": "https://openreview.net/forum?id=6dtI3qPVwVp", "arxiv_id": "2305.13656", "paper_pdf": "papers/6dtI3qPVwVp.pdf", "paper_pdf_sha256": "7640862273a54a6ff4d9a4ca398f6944d02e81d9a177af508a80286c6bdc5b3a", "paper_pdf_bytes": 14523542, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/SEAL_OGB", "code_repository": "facebookresearch/SEAL_OGB", "code_commit": "ea01ef509df1a90bc246e4e9828e192eeec4289c", "code_archive": "repos/6dtI3qPVwVp.zip", "code_archive_sha256": "c46325538a3d8b48d413996fae2f332f952c816a209f4cf30c04f20c3ea47e67", "code_archive_bytes": 18330, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 65, "github_languages": {"Python": 55932}, "github_archived": true, "github_pushed_at": "2023-06-24T04:27:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/link-prediction-without-graph-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Kvbr8NicKq", "year": 2022, "status": "rejected", "title": "Fast and Reliable Evaluation of Adversarial Robustness with Minimum-Margin Attack", "authors": ["Ruize Gao", "Jiongxiao Wang", "Kaiwen Zhou", "Feng Liu", "Binghui Xie", "Gang Niu", "Bo Han", "James Cheng"], "authorids": ["~Ruize_Gao1", "~Jiongxiao_Wang1", "~Kaiwen_Zhou2", "~Feng_Liu2", "~Binghui_Xie1", "~Gang_Niu1", "~Bo_Han1", "~James_Cheng2"], "authors_source": "OpenReview API", "abstract": "The AutoAttack (AA) has been the most reliable method to evaluate adversarial robustness when considerable computational resources are available. However, the high computational cost (e.g., 100 times more than that of the project gradient descent attack) makes AA infeasible for practitioners with limited computational resources, and also hinders applications of AA in the adversarial training (AT). In this paper, we propose a novel method, minimum-margin (MM) attack, to fast and reliably evaluate adversarial robustness. Compared with AA, our method achieves comparable performance but only costs 3% of the computational time in extensive experiments. The reliability of our method lies in that we evaluate the quality of adversarial examples using the margin between two targets that can precisely identify the most adversarial example. The computational efficiency of our method lies in an effective Sequential TArget Ranking Selection (STARS) method, ensuring that the cost of the MM attack is independent of the number of classes. The MM attack opens a new way for evaluating adversarial robustness and contributes a feasible and reliable method to generate high-quality adversarial examples in AT.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JjjvbSLcC1j", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2778/Reviewer_s3fo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose minimum-margin (MM) attack to provide comparable performance with AutoAttack while significantly decreasing the computational cost. They propose Sequential TArget Ranking Selection (STARS) to make the computational cost independent of the number of classes. ", "review_text": "Strengths:\n\n-- The paper is well-written and the preliminaries are described clearly. \n\n-- The proposed method presents significantly low computational complexity. \n\nWeaknesses:\n\n-- The proposed method is only compared against PGD and CW. \n\n-- The authors have mentioned that: \"For reliability, we evaluate the quality of adversarial examples using the margin between two targets for precisely identifying the most adversarial example.\". Can you please explain about \"most adversarial example\" ?  \n\n-- $\\beta$ in Equation 9 is not defined.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose minimum-margin (MM) attack to provide comparable performance with AutoAttack while significantly decreasing the computational cost. They propose Sequential TArget Ranking Selection (STARS) to make the computational cost independent of the number of classes. ", "main_review": "Strengths:\n\n-- The paper is well-written and the preliminaries are described clearly. \n\n-- The proposed method presents significantly low computational complexity. \n\nWeaknesses:\n\n-- The proposed method is only compared against PGD and CW. \n\n-- The authors have mentioned that: \"For reliability, we evaluate the quality of adversarial examples using the margin between two targets for precisely identifying the most adversarial example.\". Can you please explain about \"most adversarial example\" ?  \n\n-- $\\beta$ in Equation 9 is not defined.", "summary_of_the_review": "Although, I believe the proposed method has the potential for a good publication, I do not recommend the acceptance of the paper in the current form.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636454430775}, {"id": "f_xT_PzdYef", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2778/Reviewer_pCJy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposed a strong adversarial attack, i.e., an attack that can generate strong adversarial examples and thus can better evaluate the adversarial robustness of given deep learning models. Compared with the SOTA attack, the proposed attack is much faster and thus easier to be applied in practice. The idea is novel and the results are solid.\n", "review_text": "Major contributions:\n\nThe main idea has been illustrated in Figure 2. Traditional PGD attack minimizes the probability of the true label (by maximizing the loss), and the proposed minimum margin attack minimizes the margin of the probabilities between the true label and the most confusing label. To the best of my knowledge, the idea is novel in adversarial machine learning.\n\nThe computational efficiency of the MM attack is amazing. Nowadays, researchers are still using mainly PGD for training but AA for evaluation because AA is more than 100 times slower than PGD. The MM attack is 20 or even 30 times faster than the AA attack, making it possible for training with stronger adversarial examples besides faster evaluation of the adversarial robustness of given deep learning models. In my opinion, the results are significant.\n\nConcerns:\n\nThe paper lacks some theoretical analysis, for example, how would the attack converge, how to guarantee the minimum probability margin example is stronger than the minimum probability example (and thus more informative for both evaluating and training), and whether the iterative MM attack algorithm is as stable as the PGD and AA algorithms.\n\nThe experiments mainly focused on the evaluation part and then the training part is quite weak. Although researchers believe stronger adversarial examples lead to more robust models, it is not always the case because adversarial examples can be generated by quite different underlying principles as minimum probability vs. minimum probability margin. It is better to concretely show that MM is almost as fast as PGD and almost as strong as AA for training besides for evaluation. This is quite critical for the significance of the paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposed a strong adversarial attack, i.e., an attack that can generate strong adversarial examples and thus can better evaluate the adversarial robustness of given deep learning models. Compared with the SOTA attack, the proposed attack is much faster and thus easier to be applied in practice. The idea is novel and the results are solid.\n", "main_review": "Major contributions:\n\nThe main idea has been illustrated in Figure 2. Traditional PGD attack minimizes the probability of the true label (by maximizing the loss), and the proposed minimum margin attack minimizes the margin of the probabilities between the true label and the most confusing label. To the best of my knowledge, the idea is novel in adversarial machine learning.\n\nThe computational efficiency of the MM attack is amazing. Nowadays, researchers are still using mainly PGD for training but AA for evaluation because AA is more than 100 times slower than PGD. The MM attack is 20 or even 30 times faster than the AA attack, making it possible for training with stronger adversarial examples besides faster evaluation of the adversarial robustness of given deep learning models. In my opinion, the results are significant.\n\nConcerns:\n\nThe paper lacks some theoretical analysis, for example, how would the attack converge, how to guarantee the minimum probability margin example is stronger than the minimum probability example (and thus more informative for both evaluating and training), and whether the iterative MM attack algorithm is as stable as the PGD and AA algorithms.\n\nThe experiments mainly focused on the evaluation part and then the training part is quite weak. Although researchers believe stronger adversarial examples lead to more robust models, it is not always the case because adversarial examples can be generated by quite different underlying principles as minimum probability vs. minimum probability margin. It is better to concretely show that MM is almost as fast as PGD and almost as strong as AA for training besides for evaluation. This is quite critical for the significance of the paper. ", "summary_of_the_review": "This is an overall well-executed paper, with good novelty and solid experiments. Some points should be clarified and stregnthened in the revision. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635815231439}, {"id": "RxbXZXcSIFI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2778/Reviewer_Zhx4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a minimum-margin (MM) attack to evaluate defenses. The authors report detailed results on the effects of different loss functions. Experiments are done on CIFAR-10/100 and SVHN, against the adversarially trained models.", "review_text": "Strengths:\n- This paper is well-written, especially with detailed descriptions and empirical results on the effects of different attacking loss functions.\n- The improvements shown in Figure 1 seems promising with significant saving on computation.\n\nWeaknesses:\n- Although several attacking baselines are considered, they are all only evaluated against PGD-AT (Madry et al. 2018). This could cause a biased evaluation of the attacking performance. Namely, as a potential substitute for AA, the proposed MM should be widely tested against different defenses, just as done in the AA paper (Croce & Hein, 2020). This should not be computationally hard, considering the efficiency of MM and many existing defenses (and their checkpoints) provided in, e.g., RobustBench.\n- The multi-target attacking strategy has already been proposed in [a], but I'm surprised that [a] is not even cited in this paper. For me, the proposed STAR strategy is just a top-K variant for the original multi-target attack. Besides, using logits rather than softmax outputs is also not a new discovery since Carlini & Wagner (2017b). Thus, the technical contribution and novelty of MM are quite limited.  \n\nMinors:\n- What is the definition of $l\\_{MM}$ in Algorithm 1?\n\nReferences:\n[a] Gowal et al. An alternative surrogate loss for pgd-based adversarial testing, 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a minimum-margin (MM) attack to evaluate defenses. The authors report detailed results on the effects of different loss functions. Experiments are done on CIFAR-10/100 and SVHN, against the adversarially trained models.", "main_review": "Strengths:\n- This paper is well-written, especially with detailed descriptions and empirical results on the effects of different attacking loss functions.\n- The improvements shown in Figure 1 seems promising with significant saving on computation.\n\nWeaknesses:\n- Although several attacking baselines are considered, they are all only evaluated against PGD-AT (Madry et al. 2018). This could cause a biased evaluation of the attacking performance. Namely, as a potential substitute for AA, the proposed MM should be widely tested against different defenses, just as done in the AA paper (Croce & Hein, 2020). This should not be computationally hard, considering the efficiency of MM and many existing defenses (and their checkpoints) provided in, e.g., RobustBench.\n- The multi-target attacking strategy has already been proposed in [a], but I'm surprised that [a] is not even cited in this paper. For me, the proposed STAR strategy is just a top-K variant for the original multi-target attack. Besides, using logits rather than softmax outputs is also not a new discovery since Carlini & Wagner (2017b). Thus, the technical contribution and novelty of MM are quite limited.  \n\nMinors:\n- What is the definition of $l\\_{MM}$ in Algorithm 1?\n\nReferences:\n[a] Gowal et al. An alternative surrogate loss for pgd-based adversarial testing, 2019.", "summary_of_the_review": "Limited technical contribution, lack of evaluations against more defenses.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635696390507}, {"id": "ayTXEckd31H", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2778/Reviewer_BW92"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes an attack for testing adversarial robustness that is reportedly faster than the state-of-the-art attacks but still produces reliable results. The advantage in speed is obtained by using a sequential target ranking selection method, while reliability is achieved by using a minimum-margin loss.\n", "review_text": "### Comments:\n\n- The threat model is not stated anywhere. There is no definition of adversarial robustness/robust accuracy. It is not clear then if (1) the attack is a minimum-distance or maximum-confidence attack, (2) it is a targeted or untargeted attack in the common sense used in the field (as opposed to the \"targeted\" version of APGD, that is instead just using the \"targets\" for reducing the number of adversarial classes to consider in the optimization), and (3) if the attack is only defined in the $\\ell_\\infty$ norm.\n\n\n### Flaws in the experimental evaluation\n\n- the evaluation does not consider a state-of-the-art attack such as [brendel2020]. This is a pity, as [brendel2020] is presented as a \"fast and reliable method\" for evaluating robustness, and has similar desired characteristics as this attack.\n- The parameters of the attacks used seem sub-optimal. There is no choice of the hyperparameters, and using 10 steps for PGD seems to be limiting the capabilities of the attack. The same can be true for the CW attack, especially as there is no mention on how many binary-search steps are being used. It is OK to test the attacks with limited resources, but a more detailed asymptotic analysis (e.g. with 1k steps) would concretely support the claims of the paper that the attack remains comparable to the other attacks while reducing computational time.\n- The runtimes are computed in an uneven scenario. The total time per-step, or better, per-query to the model should be used instead of the total cumulative time. This makes no sense. As an alternative, one should compare the capabilities of the PGD attack (depicted here as fast but not reliable) with a fixed computational time, i.e., by increasing the number of steps performed by PGD until it spends the same amount of time as the MM attack.\n- the authors did not state if they used some available implementations of the attack, or implemented their own versions. Since the computational time depends on the implementation, this might be a problem when using the runtime as a benchmark.\n\n### Incorrect statements and unsupported claims\n\n**Abstract**\n\n- there is no definition of the \"most adversarial example\", even though there are several references of this in the paper. This is also used in the abstract. Depending on the objective, a stong adversarial example can be seen in different ways. I suggest to expand this with a definition.\n- there is no evidence suggesting that the PGD attack is 100 times slower than AA. The comparison is performed in uneven scenario, where AA uses 100 iterations while PGD uses 10. Moreover, this is stated in the abstract, which makes the statement easy to take and quote, without knowing the context. This statement should be removed.\n\n\n**Introduction**\n\n- \"for practitioners who need real-time evaluation at each epoch of the training process of a robust model\". Is there real cases that require this kind of evaluation? This is missing a reference.\n- \"Unfortunately, PGD fails to reliably evaluate adversarial robustness of a robust DNN\". This sentence is over-generalistic and not true for the majority of the cases. PGD was succesfully used against many defenses, just by making it adaptive to the defense [tramer2020].\n- \"CE loss, which is based on the probability of the true label $p_y$, is not an appropriate measure to the quality of adversarial examples\". There is no definition in the paper for \"quality of adversarial examples\", which makes this statement very confusing.\n- \"Hence, the reliable method is to minimize $z_y - z_t$ for each $t \\neq y$ and take the most adversarial one, which is a widely used solution\". This is not widely-used, as for now it seems only used in [croce2020].\n\n**Preliminary**\n\n- \"$x^{(0)}$ refers to the starting point which corresponds to the natural example (or the natural example perturbed by a small Gaussian or uniformly random noise)\". The statement within parentheses makes the definition of the closed ball in eq. 2 makes the ball centered in $x^{(0)}$. This does not correspond to the adversarial robustness measured in the original clean sample.\n- many equations (see Eqs. 3-7) depend on f, x, y, but they often don't appear inside the equations.\n- \"They showed that using adaptive step size significantly improves the adversarial robustness\". Should be \"improves the adversarial examples\" or \"improves the adversarial evaluation\". The attacks are not improving robustness.\n\n**Realization**\n\n- Eqs. 8 and 9 use variables ($\\alpha$, $\\beta$) never introduced in the text.\n\n\n### Minor issues\n\n- the comparison with targeted-dlr loss in sect. 3 should be clarified. It is very difficult to read, and it does not really capture the advantage of using different methods for rescaling. This might be better supported by some evidence or toy example, and surely by adding some insight on which the hypothesis is based on. Moreover, the authors should then explain what is the difference from the CW loss, as it seems that they are using that one.\n- Figures and tables need descriptive captions that clarify what is being depicted. In particular, tables need improvements in the headers and some highlighting of the results. It is also a good practice to mention them in the text (Figure 1b). Figure 2 is difficult to understand, and contains a legend with unclear definitions (see \"classified area\"). In table 2, it is impossible to understand what are the values presented in the cells. The algorithm needs some hints/comments/description.\n\n\n### References:\n\n- [tramer2020] Tramer, Florian, et al. \"On Adaptive Attacks to Adversarial Example Defenses.\" Advances in Neural Information Processing Systems 33 (2020).\n\n- [croce2020] Croce, Francesco, and Matthias Hein. \"Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks.\" International conference on machine learning. PMLR, 2020.\n\n- [brendel2020] Brendel, W., et al. \"Accurate, reliable and fast robustness evaluation.\" Thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019). Curran, 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes an attack for testing adversarial robustness that is reportedly faster than the state-of-the-art attacks but still produces reliable results. The advantage in speed is obtained by using a sequential target ranking selection method, while reliability is achieved by using a minimum-margin loss.\n", "main_review": "### Comments:\n\n- The threat model is not stated anywhere. There is no definition of adversarial robustness/robust accuracy. It is not clear then if (1) the attack is a minimum-distance or maximum-confidence attack, (2) it is a targeted or untargeted attack in the common sense used in the field (as opposed to the \"targeted\" version of APGD, that is instead just using the \"targets\" for reducing the number of adversarial classes to consider in the optimization), and (3) if the attack is only defined in the $\\ell_\\infty$ norm.\n\n\n### Flaws in the experimental evaluation\n\n- the evaluation does not consider a state-of-the-art attack such as [brendel2020]. This is a pity, as [brendel2020] is presented as a \"fast and reliable method\" for evaluating robustness, and has similar desired characteristics as this attack.\n- The parameters of the attacks used seem sub-optimal. There is no choice of the hyperparameters, and using 10 steps for PGD seems to be limiting the capabilities of the attack. The same can be true for the CW attack, especially as there is no mention on how many binary-search steps are being used. It is OK to test the attacks with limited resources, but a more detailed asymptotic analysis (e.g. with 1k steps) would concretely support the claims of the paper that the attack remains comparable to the other attacks while reducing computational time.\n- The runtimes are computed in an uneven scenario. The total time per-step, or better, per-query to the model should be used instead of the total cumulative time. This makes no sense. As an alternative, one should compare the capabilities of the PGD attack (depicted here as fast but not reliable) with a fixed computational time, i.e., by increasing the number of steps performed by PGD until it spends the same amount of time as the MM attack.\n- the authors did not state if they used some available implementations of the attack, or implemented their own versions. Since the computational time depends on the implementation, this might be a problem when using the runtime as a benchmark.\n\n### Incorrect statements and unsupported claims\n\n**Abstract**\n\n- there is no definition of the \"most adversarial example\", even though there are several references of this in the paper. This is also used in the abstract. Depending on the objective, a stong adversarial example can be seen in different ways. I suggest to expand this with a definition.\n- there is no evidence suggesting that the PGD attack is 100 times slower than AA. The comparison is performed in uneven scenario, where AA uses 100 iterations while PGD uses 10. Moreover, this is stated in the abstract, which makes the statement easy to take and quote, without knowing the context. This statement should be removed.\n\n\n**Introduction**\n\n- \"for practitioners who need real-time evaluation at each epoch of the training process of a robust model\". Is there real cases that require this kind of evaluation? This is missing a reference.\n- \"Unfortunately, PGD fails to reliably evaluate adversarial robustness of a robust DNN\". This sentence is over-generalistic and not true for the majority of the cases. PGD was succesfully used against many defenses, just by making it adaptive to the defense [tramer2020].\n- \"CE loss, which is based on the probability of the true label $p_y$, is not an appropriate measure to the quality of adversarial examples\". There is no definition in the paper for \"quality of adversarial examples\", which makes this statement very confusing.\n- \"Hence, the reliable method is to minimize $z_y - z_t$ for each $t \\neq y$ and take the most adversarial one, which is a widely used solution\". This is not widely-used, as for now it seems only used in [croce2020].\n\n**Preliminary**\n\n- \"$x^{(0)}$ refers to the starting point which corresponds to the natural example (or the natural example perturbed by a small Gaussian or uniformly random noise)\". The statement within parentheses makes the definition of the closed ball in eq. 2 makes the ball centered in $x^{(0)}$. This does not correspond to the adversarial robustness measured in the original clean sample.\n- many equations (see Eqs. 3-7) depend on f, x, y, but they often don't appear inside the equations.\n- \"They showed that using adaptive step size significantly improves the adversarial robustness\". Should be \"improves the adversarial examples\" or \"improves the adversarial evaluation\". The attacks are not improving robustness.\n\n**Realization**\n\n- Eqs. 8 and 9 use variables ($\\alpha$, $\\beta$) never introduced in the text.\n\n\n### Minor issues\n\n- the comparison with targeted-dlr loss in sect. 3 should be clarified. It is very difficult to read, and it does not really capture the advantage of using different methods for rescaling. This might be better supported by some evidence or toy example, and surely by adding some insight on which the hypothesis is based on. Moreover, the authors should then explain what is the difference from the CW loss, as it seems that they are using that one.\n- Figures and tables need descriptive captions that clarify what is being depicted. In particular, tables need improvements in the headers and some highlighting of the results. It is also a good practice to mention them in the text (Figure 1b). Figure 2 is difficult to understand, and contains a legend with unclear definitions (see \"classified area\"). In table 2, it is impossible to understand what are the values presented in the cells. The algorithm needs some hints/comments/description.\n\n\n### References:\n\n- [tramer2020] Tramer, Florian, et al. \"On Adaptive Attacks to Adversarial Example Defenses.\" Advances in Neural Information Processing Systems 33 (2020).\n\n- [croce2020] Croce, Francesco, and Matthias Hein. \"Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks.\" International conference on machine learning. PMLR, 2020.\n\n- [brendel2020] Brendel, W., et al. \"Accurate, reliable and fast robustness evaluation.\" Thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019). Curran, 2020.", "summary_of_the_review": "### Strengths: \n- Tries to improve efficiency of adversarial attacks\n\n### Weaknesses:\n- the paper is missing some definition that should not be taken for granted\n- the evaluation is not entirely convincing and might be unfair\n- results should be presented better, as they are very difficult to read and understand", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635590646688}], "openreview_url": "https://openreview.net/forum?id=Kvbr8NicKq", "arxiv_id": "2206.07314", "paper_pdf": "papers/Kvbr8NicKq.pdf", "paper_pdf_sha256": "ea19558a51fceac6a52522dff9be361328184b52864424f76f0a8867be17caf3", "paper_pdf_bytes": 685991, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sjtubrian/MM-attack", "code_repository": "Sjtubrian/MM-attack", "code_commit": "0a9883eb3cf06b8800540b841f70424302f24012", "code_archive": "repos/Kvbr8NicKq.zip", "code_archive_sha256": "0429fc55ec909b316e9bd53c66c27ffbfa618a2c005f741d6b17a51cbcc561dc", "code_archive_bytes": 55767, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 76, "github_languages": {"Python": 215771}, "github_archived": false, "github_pushed_at": "2022-10-20T05:03:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fast-and-reliable-evaluation-of-adversarial-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "KJSC_AsN14", "year": 2021, "status": "rejected", "title": "Contrastive Learning with Stronger Augmentations", "authors": ["Xiao Wang", "Guo-Jun Qi"], "authorids": ["~Xiao_Wang6", "~Guo-Jun_Qi1"], "authors_source": "OpenReview API", "abstract": "Representation learning has been greatly improved with the advance of contrastive learning methods with the performance being closer to their supervised learning counterparts. Those methods have greatly benefited from various data augmentations that are carefully designated to maintain their identities so that the images transformed from the same instance can still be retrieved. Although stronger augmentations could expose novel patterns of representations to improve their generalizability, directly using stronger augmentations in instance discrimination-based contrastive learning may even deteriorate the performance, because the distortions induced from the stronger augmentations could ridiculously change the image structures and thus the transformed images cannot be viewed as the same as the original ones any more. Additional efforts are needed for us to explore the role of the stronger augmentations in further pushing the performance of unsupervised learning to the fully supervised upper bound. Instead of applying the stronger augmentations directly to minimize the contrastive loss, we propose to minimize the distribution divergence between the weakly and strongly augmented images over the representation bank to supervise the retrieval of strongly augmented queries from a pool of candidates. This avoids an overoptimistic assumption that could overfit the strongly augmented queries containing distorted visual structures into the positive targets in the representation bank, while still being able to distinguish them from the negative samples by leveraging the distributions of weakly augmented counterparts. The proposed method achieves top-1 accuracy of 76.2% on ImageNet with a standard ResNet-50 architecture with a single-layer classifier fine-tuned. This is almost the same as 76.5% of top-1 accuracy with a fully supervised ResNet-50. Moreover, it outperforms the previous self-supervised and supervised methods on both the transfer learning and object detection tasks.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "oeNKgjl6Kv6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1274/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper focuses on designing more effective ways for contrastive learning. The author claims that stronger augmentations are beneficial for better representation learning. Different from directly applying the stronger augmentations to minimize the contrastive loss, the author proposes to minimize the distribution divergence between the weakly and strongly augmented images. The experimental evaluations are conducted on ImageNet classification and related downstream tasks, and the results are promising.\n\nClarity:\n1. The method is very simple and straightforward. My main concern is the experimental comparisons. As we all know, contrastive learning algorithms like MOCO and SimCLR benefits from longer training epochs a lot (for example, training with 800 epochs is much better than with 400 epochs). Thus I think the comparisons in Table 2 are not convincing. From algorithm 1, we can find that the equivalent batch size of the proposed CLSA method is two times as classical MOCO method. Thus I would prefer to check the results of CLSA at epochs 100 and 400 for fair comparisons.\n2. What is the value of the balancing coefficient? It would be nice if some ablation results are provided.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The experimental evaluations are not convincing", "review": "This paper focuses on designing more effective ways for contrastive learning. The author claims that stronger augmentations are beneficial for better representation learning. Different from directly applying the stronger augmentations to minimize the contrastive loss, the author proposes to minimize the distribution divergence between the weakly and strongly augmented images. The experimental evaluations are conducted on ImageNet classification and related downstream tasks, and the results are promising.\n\nClarity:\n1. The method is very simple and straightforward. My main concern is the experimental comparisons. As we all know, contrastive learning algorithms like MOCO and SimCLR benefits from longer training epochs a lot (for example, training with 800 epochs is much better than with 400 epochs). Thus I think the comparisons in Table 2 are not convincing. From algorithm 1, we can find that the equivalent batch size of the proposed CLSA method is two times as classical MOCO method. Thus I would prefer to check the results of CLSA at epochs 100 and 400 for fair comparisons.\n2. What is the value of the balancing coefficient? It would be nice if some ablation results are provided.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604364413301}, {"id": "AMoGFJTWXDk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1274/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper proposes the better utilization of strong data augmentations for contrastive loss functions in unsupervised learning. In Moco set up, typically, weaker augmentations such as color jittering, cropping is applied to construct positive pairs from the same image. In this study, by proposing a modified objective, the authors leverage stronger data augmentations to construct more challenging positives and negatives pairs to improve the quality of the representations. The paper delivers a novel objective together with leveraging existing strong augmentations to improve downstream performance. The authors can find my questions/concerns listed below.\n\n1. The paper is overall well-written, however, it is disappointing to see many typos grammar mistakes throughout the paper. Some examples are in \"Thus we proposed the CLSA (Contrastive Learning with Stronger Augementations)\", \"to train an unsupervised representation\", \"The contrastive learning (Hadsell et al. (2006)) is a popular self-supervised idea\".\n\n2. In section 3.1, the authors mention that the keys in the memory bank is managed with first in first out method. Is it not supposed to be first in last out? I would like to see some clarification on this.\n\n3.  The numerator in Equation 3 should be z_i' vs. z_i not z_k.\n\n4. The authors claim that in He et al. an input image is resized and cropped to 224×224 pixels. It should be \"an image is first cropped from an input image and resized to 224x224 pixels.\"\n\n5. In the experiments section, the authors list other methods including MoCo, SimCLR, MoCo-v2, BYOL and compare to what they propose. As a baseline, it would be nice to directly use the stronger augmentations in MoCo-v2 objective and perform comparison to their method. Throughout the paper, the authors claim that strong augmentations hurt the learned representations due to distorted images. It would be meaningful to show this experimentally as well.\n\n6. The authors explain that they choose a strong transformation randomly from the given 14 transformations and repeat it 5 times to strongly augment an image. Is the sampling done without replacement? In other words, do the authors choose 5 unique transformations with the corresponding magnitude and apply those transformations to a single image?\n\n7. I like how the authors point the similarity of their objective to knowledge distillation. In this case, strong augmentations are assigned probability of being a positive pair from the positive pair constructed with weak augmentations. It helps to understand the full picture for the proposed method.\n\n8. Finally, I think the figure 3 is confusing rather than being helpful. Both weak and strong augmentations go to the memory bank and it looks like two distributions come out of nowhere in the figure. It would be more clear to point out that there is distribution of the representations from the strong augmentations and weak augmentations and they supervise the assignment for strong augmentations given predictions on the weak augmentations.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "CONTRASTIVE LEARNING WITH STRONGER AUGMENTATIONS", "review": "\nThis paper proposes the better utilization of strong data augmentations for contrastive loss functions in unsupervised learning. In Moco set up, typically, weaker augmentations such as color jittering, cropping is applied to construct positive pairs from the same image. In this study, by proposing a modified objective, the authors leverage stronger data augmentations to construct more challenging positives and negatives pairs to improve the quality of the representations. The paper delivers a novel objective together with leveraging existing strong augmentations to improve downstream performance. The authors can find my questions/concerns listed below.\n\n1. The paper is overall well-written, however, it is disappointing to see many typos grammar mistakes throughout the paper. Some examples are in \"Thus we proposed the CLSA (Contrastive Learning with Stronger Augementations)\", \"to train an unsupervised representation\", \"The contrastive learning (Hadsell et al. (2006)) is a popular self-supervised idea\".\n\n2. In section 3.1, the authors mention that the keys in the memory bank is managed with first in first out method. Is it not supposed to be first in last out? I would like to see some clarification on this.\n\n3.  The numerator in Equation 3 should be z_i' vs. z_i not z_k.\n\n4. The authors claim that in He et al. an input image is resized and cropped to 224×224 pixels. It should be \"an image is first cropped from an input image and resized to 224x224 pixels.\"\n\n5. In the experiments section, the authors list other methods including MoCo, SimCLR, MoCo-v2, BYOL and compare to what they propose. As a baseline, it would be nice to directly use the stronger augmentations in MoCo-v2 objective and perform comparison to their method. Throughout the paper, the authors claim that strong augmentations hurt the learned representations due to distorted images. It would be meaningful to show this experimentally as well.\n\n6. The authors explain that they choose a strong transformation randomly from the given 14 transformations and repeat it 5 times to strongly augment an image. Is the sampling done without replacement? In other words, do the authors choose 5 unique transformations with the corresponding magnitude and apply those transformations to a single image?\n\n7. I like how the authors point the similarity of their objective to knowledge distillation. In this case, strong augmentations are assigned probability of being a positive pair from the positive pair constructed with weak augmentations. It helps to understand the full picture for the proposed method.\n\n8. Finally, I think the figure 3 is confusing rather than being helpful. Both weak and strong augmentations go to the memory bank and it looks like two distributions come out of nowhere in the figure. It would be more clear to point out that there is distribution of the representations from the strong augmentations and weak augmentations and they supervise the assignment for strong augmentations given predictions on the weak augmentations.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603861004928}, {"id": "SYTwselb9w", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1274/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a method to incorporate stronger augmentations into the visual representation contrastive learning framework. Specifically, three correlated views of an image are first generated by using two weak and one strong augmentation operations on the same image. Then, the networks are trained to maximize the agreement between the two weak views and also to minimize the distribution divergence between a weak view and the strong view. The method is evaluated on several visual tasks including classification,  transfer learning, and object detection, with the standard evaluation protocol for self-supervised learning, and the results are promising.\n\nPros:\n\n1. This paper is well-structured and easy-to-follow.\n2. The idea of utilizing strong augmentations for contrastive learning is interesting and novel to me, and the results are promising.\n3. The proposed framework seems general which might be easily incorporated into the existing contrastive learning frameworks.\n\nCons:\n\n1. The motivation about using stronger augmentations is not well justified. Specifically, the authors propose to use stronger augmentations based on two reasons: (1) stronger augmentations can expose some novel useful patterns; (2) the effectiveness of stronger augmentations is proved in the semi-supervised learning and supervised learning field. However, no related papers are provided to support the first point, while the papers (Cubuk et al. (2018)); Qi et al. (2019); Wang et al. (2019)) that are cited to support the second point do not explicitly make relevant conclusions. (Chen et al. (2020a)) even demonstrate that when training supervised models, stronger color augmentation hurts their performance. I would like to see a more comprehensive review of related works to clarify the motivation.\n\n2. In addition, some important ablation studies are missing in the experiment. E.g.,  how does the performance change as the magnitude or usage times of stronger augmentations changes?\n\n3. The proposed DDM loss seems general for different contrastive learning frameworks. I would like to see if it still works when applied to other frameworks, e.g., SimCLR, InfoMin?\n\nOverall, given the novelty and strong results of the proposed framework, I remain positive towards this paper. I will be happy to increase my rating if my concerns are addressed in the rebuttal period.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #3", "review": "This paper presents a method to incorporate stronger augmentations into the visual representation contrastive learning framework. Specifically, three correlated views of an image are first generated by using two weak and one strong augmentation operations on the same image. Then, the networks are trained to maximize the agreement between the two weak views and also to minimize the distribution divergence between a weak view and the strong view. The method is evaluated on several visual tasks including classification,  transfer learning, and object detection, with the standard evaluation protocol for self-supervised learning, and the results are promising.\n\nPros:\n\n1. This paper is well-structured and easy-to-follow.\n2. The idea of utilizing strong augmentations for contrastive learning is interesting and novel to me, and the results are promising.\n3. The proposed framework seems general which might be easily incorporated into the existing contrastive learning frameworks.\n\nCons:\n\n1. The motivation about using stronger augmentations is not well justified. Specifically, the authors propose to use stronger augmentations based on two reasons: (1) stronger augmentations can expose some novel useful patterns; (2) the effectiveness of stronger augmentations is proved in the semi-supervised learning and supervised learning field. However, no related papers are provided to support the first point, while the papers (Cubuk et al. (2018)); Qi et al. (2019); Wang et al. (2019)) that are cited to support the second point do not explicitly make relevant conclusions. (Chen et al. (2020a)) even demonstrate that when training supervised models, stronger color augmentation hurts their performance. I would like to see a more comprehensive review of related works to clarify the motivation.\n\n2. In addition, some important ablation studies are missing in the experiment. E.g.,  how does the performance change as the magnitude or usage times of stronger augmentations changes?\n\n3. The proposed DDM loss seems general for different contrastive learning frameworks. I would like to see if it still works when applied to other frameworks, e.g., SimCLR, InfoMin?\n\nOverall, given the novelty and strong results of the proposed framework, I remain positive towards this paper. I will be happy to increase my rating if my concerns are addressed in the rebuttal period.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603764867780}, {"id": "EUgoVglDREZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1274/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\\\nThis work investigate the recent popular direction of unsupervised representation learning using contrastive loss between augmented images. Authors propose to minimize the divergence between the distributions of strongly augmented vs. weakly augmented images. The method reaches competitive performance in recognition and object detection.\n-----\n+Strengths\\\n+The main idea is well motivated: that strong augmentation reveal useful cues in visual representation learning but has not been successfully exploited in unsupervised learning.\\\n+The proposed solution is novel within contrastive learning to my best knowledge.\\\n+Results are extremely strong.\n-----\n-Concerns\\\n-The divergence between the two conditional distributions can be a moving target since they are trained jointly. It is not clear if this is will result in stable learning for the unsupervised setting, and what effect that may have on the performance and quality of the representations.\\\n-Evaluation only focus on final result and lacks analysis of the proposed method, especially when compared to recent paper of similar nature published in top conferences. For example, strong augmentation is a focus of this paper, but there are no ablation regarding the augmentations. Is the performance sensitive to the choice of strong augmentation?\\\n-The paper could also use some more theoretical analysis to address some of the weaknesses stated above.\n-----\nRecommendation\\\nI like the proposed idea. It is novel and interesting and seems to achieve good results. However the lack of both theoretical and empirical analysis beyond results on performance raises many questions. As a result I am on the fence but leaning towards accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "Summary:\\\nThis work investigate the recent popular direction of unsupervised representation learning using contrastive loss between augmented images. Authors propose to minimize the divergence between the distributions of strongly augmented vs. weakly augmented images. The method reaches competitive performance in recognition and object detection.\n-----\n+Strengths\\\n+The main idea is well motivated: that strong augmentation reveal useful cues in visual representation learning but has not been successfully exploited in unsupervised learning.\\\n+The proposed solution is novel within contrastive learning to my best knowledge.\\\n+Results are extremely strong.\n-----\n-Concerns\\\n-The divergence between the two conditional distributions can be a moving target since they are trained jointly. It is not clear if this is will result in stable learning for the unsupervised setting, and what effect that may have on the performance and quality of the representations.\\\n-Evaluation only focus on final result and lacks analysis of the proposed method, especially when compared to recent paper of similar nature published in top conferences. For example, strong augmentation is a focus of this paper, but there are no ablation regarding the augmentations. Is the performance sensitive to the choice of strong augmentation?\\\n-The paper could also use some more theoretical analysis to address some of the weaknesses stated above.\n-----\nRecommendation\\\nI like the proposed idea. It is novel and interesting and seems to achieve good results. However the lack of both theoretical and empirical analysis beyond results on performance raises many questions. As a result I am on the fence but leaning towards accept.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603637104452}], "openreview_url": "https://openreview.net/forum?id=KJSC_AsN14", "arxiv_id": "2104.07713", "paper_pdf": "papers/KJSC_AsN14.pdf", "paper_pdf_sha256": "fc0db3bc2416967f913beeaaed593dcbf48da5a26af2643d5727fe740381eb71", "paper_pdf_bytes": 1749257, "paper_pdf_source": "openreview", "code_url": "https://github.com/maple-research-lab/CLSA", "code_repository": "maple-research-lab/CLSA", "code_commit": "37df76cf5cb032683e57b70a3a4090f0d524c8fd", "code_archive": "repos/KJSC_AsN14.zip", "code_archive_sha256": "d6404572a58a7b1a5074853099730c7b61d2033524096619a3d70133ac1dc34c", "code_archive_bytes": 45948, "code_file_count": 21, "code_extensions": {".py": 19, ".sh": 2}, "github_disk_usage_kb": 80, "github_languages": {"Python": 117550, "Shell": 928}, "github_archived": false, "github_pushed_at": "2021-08-02T13:24:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contrastive-learning-with-stronger-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryga2CNKDH", "year": 2020, "status": "rejected", "title": "Evaluating Lossy Compression Rates of Deep Generative Models", "authors": ["Sicong Huang", "Alireza Makhzani", "Yanshuai Cao", "Roger Grosse"], "authorids": ["huang@cs.toronto.edu", "a.makhzani@gmail.com", "yanshuai.cao@borealisai.com", "rgrosse@cs.toronto.edu"], "authors_source": "OpenReview API", "abstract": "Deep generative models have achieved remarkable progress in recent years. Despite this progress, quantitative evaluation and comparison of generative models remains as one of the important challenges. One of the most popular metrics for evaluating generative models is the log-likelihood. While the direct computation of log-likelihood can be intractable, it has been recently shown that the log-likelihood of some of the most interesting generative models such as variational autoencoders (VAE) or generative adversarial networks (GAN) can be efficiently estimated using annealed importance sampling (AIS). In this work, we argue that the log-likelihood metric by itself cannot represent all the different performance characteristics of generative models, and propose to use rate distortion curves to evaluate and compare deep generative models. We show that we can approximate the entire rate distortion curve using one single run of AIS for roughly the same computational cost as a single log-likelihood estimate. We evaluate lossy compression rates of different deep generative models such as VAEs, GANs (and its variants) and adversarial autoencoders (AAE) on MNIST and CIFAR10, and arrive at a number of insights not obtainable from log-likelihoods alone.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bkeatl8PcB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1374/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a method for evaluating latent-variable generative models in terms of the rate-distortion curve that compares the number of bits needed to encode the representation with how well you can reconstruct an input under some distortion measure. To estimate this curve, the author’s use AIS and show how intermediate distributions in AIS can be used to bound and estimate rate and distortion. They apply their evaluation to GANs, VAEs, and AAEs trained on MNIST and CIFAR-10.\n\nI found this paper well written, with a number of interesting technical contributions, particularly how to leverage AIS to compute R-D curves for an individual model. However, the utility and interpretation of these R-D curves for single models remains confusing to me, and there is insufficient discussion and comparison to other joint diversity/sample quality metrics proposed in the GAN literature. The compute time required for evaluation may also limit the applicability: 4-7 hours for 50 images on MNIST, and 7 days for CIFAR-10. \n \nMajor comments:\n* How should we interpret rate-prior distortion for an individual model vs. rate-distortion where models are optimized for each rate? Past work in learned compression and generative models (Theis et al. 2016, Balle et al. 2016, Alemi et al., 2018) show that models must adjust their decoder (and prior) as a function of rate to be optimal in terms of distortion. For a fixed decoder, optimizing the prior may still be required to achieve low rate. Given that many of the models you compare are trained to do well at one point on the R-D curve, why does it make sense to evaluate them at other points? Additionally, you only evaluate models with deterministic decoders and many of the experimental conclusions are highly specific to this setting but not noted. \n* As you focus on general distortion metrics instead of NLL alone, it'd be interesting to compare curves under different distortion measures, e.g. MS-SSIM for images or L1 vs. L2. Right now there's not much experimental novelty vs. prior work that looked at rate-distortion curves with NLL distortion and Gaussian observation models.\n* It’d be useful to include experiments comparing Rate-Prior distortion curves and Rate-distortion curves where you a) optimize over the prior, b) optimize the decoder, fixing the prior, and c) optimize both the prior and decoder. \n* There’s no comparison of other approaches to generate the rate-prior distortion curve. For example, you could just use an amortized inference network like in VAE w/ a flexible variational family and anneal beta over time.\n* There are several related papers which should be discussed and contrasted, in particular https://arxiv.org/abs/1901.07821 which looks at rate-distortion-perception tradeoffs, and https://arxiv.org/abs/1806.00035 which presents precision-recall curves for diversity/quality metrics applied to implicit models. How do the insights gained from the rate-distortion curves relate to precision/recall and why should one be preferred over the other? https://arxiv.org/abs/1611.02163 also looked at distortion as a metric for GANs (equivalent to beta -> infinity in your framework).\n\nMinor comments:\n* Missing discussion of several related works: that presents a complexity measure for the latent space of GANs: https://arxiv.org/abs/1802.04874\n* “Wasserstein distance remains difficult to approximate…” - see https://openreview.net/forum?id=HkxKH2AcFm that advocates for evaluation with Wasserstein\n* Tightness of bound on simulated data (what BDMC provides) may not correspond to tightness of bound on real data (what you care about in practice). \n* The treatment of VAEs as implicit models only makes sense with location scale family p(x|z), thus the entire framework proposed here doesn’t make sense with e.g. autoregressive p(x|z), as used in PixelVAE and others.\n* Why focus on fixed prior p(z)? An alternative would be to optimize p(z), q(z|x) and fix p(x|z). How would this change the resulting rate-prior distortion curves?\n* “We can compute the R-D curve by sweeping over \\beta rather than by sweeping over D” - this is not the case when the R-D curve has linear segments, see e.g. Rezende & Viola 2018\n* Many of the properties and discussion around rate-prior distortion functions (especially w/NLL distortion) are also in Alemi et al. 2018 as their definition of “rate” is identical to your definition of “rate-prior”. Also many of these properties are specific to continuous latents which isn’t noted.\n* Should clarify that q_k(z|x) correspond to points along R_p(D)\n* The results in Eqn 14-17 showing you can tractably estimate distortion and get an upper bound on rate using the AIS-derived distributions are very cool!\n* “AIS variance is proportional to 1/MK” - this is for variance in the partition function? How does this translate to variance in estimates of rate/distortion?\n* “In the case of probabilistic decoders, …” -> need the caveat this is with NLL distortion\n* Validation on the linear VAE is great!It looks like some of the points for AIS at low distortion are below the analytic rate, but the proofs indicate the estimated rate should be greater than the analytic rate. Is this just noise?\n* Fig 4 and 5: hard to see difference between the dashed lines\n* VAE results would change drastically if you targeted them to different regimes (e.g. beta-VAE or constrained optimization like GECO)\n* Statements like “VAE is trained with the ELBO objective, which encourages good reconstructions” only make sense when the decoder is location-scale. VAEs w/rich autoregressive decoders typically do a horrible job reconstructing.\n* How robust are model differences across random initialization? It’d be great to add error bars to all these plots, especially given that GAN training can stochastically succeed.\n* Eqn 18/Fig6a: depending on the dataset, you could easily notice the difference of 4.6 nats and log-likelihood could still tell these two models apart. It’d be useful to add a line at beta=1 to show that the likelihood would be the same but the R-D curves are different.\n\n========================\nUpdate after rebuttal\n\nThank you to the authors for addressing a number of my concerns, adding additional text and experiments. However, my main concern remains: if rate-distortion in an individual model is a useful method for evaluating generative models, you should compare it empirically with other metrics that have been proposed for this purpose (e.g. precision-recall). Additionally, while the theoretical novelty of getting the full R-D curve from a single AIS run is very cool, I'm skeptical of the practical utility as a metric for generative models due to the computational costs of AIS (4-7 hours for 50 images on MNIST). The simple baseline I suggested of training an amortized inference network with beta annealed over time would not require training a separate encoder for each point in R-D, you could just start beta large, and anneal beta in steps to 0 over time, tracing out an R-D curve. Given the current experiments, it's not obvious if the win of AIS in terms of accurate posterior inference is worth the increased computational cost over a simple VI baseline.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This paper presents a method for evaluating latent-variable generative models in terms of the rate-distortion curve that compares the number of bits needed to encode the representation with how well you can reconstruct an input under some distortion measure. To estimate this curve, the author’s use AIS and show how intermediate distributions in AIS can be used to bound and estimate rate and distortion. They apply their evaluation to GANs, VAEs, and AAEs trained on MNIST and CIFAR-10.\n\nI found this paper well written, with a number of interesting technical contributions, particularly how to leverage AIS to compute R-D curves for an individual model. However, the utility and interpretation of these R-D curves for single models remains confusing to me, and there is insufficient discussion and comparison to other joint diversity/sample quality metrics proposed in the GAN literature. The compute time required for evaluation may also limit the applicability: 4-7 hours for 50 images on MNIST, and 7 days for CIFAR-10. \n \nMajor comments:\n* How should we interpret rate-prior distortion for an individual model vs. rate-distortion where models are optimized for each rate? Past work in learned compression and generative models (Theis et al. 2016, Balle et al. 2016, Alemi et al., 2018) show that models must adjust their decoder (and prior) as a function of rate to be optimal in terms of distortion. For a fixed decoder, optimizing the prior may still be required to achieve low rate. Given that many of the models you compare are trained to do well at one point on the R-D curve, why does it make sense to evaluate them at other points? Additionally, you only evaluate models with deterministic decoders and many of the experimental conclusions are highly specific to this setting but not noted. \n* As you focus on general distortion metrics instead of NLL alone, it'd be interesting to compare curves under different distortion measures, e.g. MS-SSIM for images or L1 vs. L2. Right now there's not much experimental novelty vs. prior work that looked at rate-distortion curves with NLL distortion and Gaussian observation models.\n* It’d be useful to include experiments comparing Rate-Prior distortion curves and Rate-distortion curves where you a) optimize over the prior, b) optimize the decoder, fixing the prior, and c) optimize both the prior and decoder. \n* There’s no comparison of other approaches to generate the rate-prior distortion curve. For example, you could just use an amortized inference network like in VAE w/ a flexible variational family and anneal beta over time.\n* There are several related papers which should be discussed and contrasted, in particular https://arxiv.org/abs/1901.07821 which looks at rate-distortion-perception tradeoffs, and https://arxiv.org/abs/1806.00035 which presents precision-recall curves for diversity/quality metrics applied to implicit models. How do the insights gained from the rate-distortion curves relate to precision/recall and why should one be preferred over the other? https://arxiv.org/abs/1611.02163 also looked at distortion as a metric for GANs (equivalent to beta -> infinity in your framework).\n\nMinor comments:\n* Missing discussion of several related works: that presents a complexity measure for the latent space of GANs: https://arxiv.org/abs/1802.04874\n* “Wasserstein distance remains difficult to approximate…” - see https://openreview.net/forum?id=HkxKH2AcFm that advocates for evaluation with Wasserstein\n* Tightness of bound on simulated data (what BDMC provides) may not correspond to tightness of bound on real data (what you care about in practice). \n* The treatment of VAEs as implicit models only makes sense with location scale family p(x|z), thus the entire framework proposed here doesn’t make sense with e.g. autoregressive p(x|z), as used in PixelVAE and others.\n* Why focus on fixed prior p(z)? An alternative would be to optimize p(z), q(z|x) and fix p(x|z). How would this change the resulting rate-prior distortion curves?\n* “We can compute the R-D curve by sweeping over \\beta rather than by sweeping over D” - this is not the case when the R-D curve has linear segments, see e.g. Rezende & Viola 2018\n* Many of the properties and discussion around rate-prior distortion functions (especially w/NLL distortion) are also in Alemi et al. 2018 as their definition of “rate” is identical to your definition of “rate-prior”. Also many of these properties are specific to continuous latents which isn’t noted.\n* Should clarify that q_k(z|x) correspond to points along R_p(D)\n* The results in Eqn 14-17 showing you can tractably estimate distortion and get an upper bound on rate using the AIS-derived distributions are very cool!\n* “AIS variance is proportional to 1/MK” - this is for variance in the partition function? How does this translate to variance in estimates of rate/distortion?\n* “In the case of probabilistic decoders, …” -> need the caveat this is with NLL distortion\n* Validation on the linear VAE is great!It looks like some of the points for AIS at low distortion are below the analytic rate, but the proofs indicate the estimated rate should be greater than the analytic rate. Is this just noise?\n* Fig 4 and 5: hard to see difference between the dashed lines\n* VAE results would change drastically if you targeted them to different regimes (e.g. beta-VAE or constrained optimization like GECO)\n* Statements like “VAE is trained with the ELBO objective, which encourages good reconstructions” only make sense when the decoder is location-scale. VAEs w/rich autoregressive decoders typically do a horrible job reconstructing.\n* How robust are model differences across random initialization? It’d be great to add error bars to all these plots, especially given that GAN training can stochastically succeed.\n* Eqn 18/Fig6a: depending on the dataset, you could easily notice the difference of 4.6 nats and log-likelihood could still tell these two models apart. It’d be useful to add a line at beta=1 to show that the likelihood would be the same but the R-D curves are different.\n\n========================\nUpdate after rebuttal\n\nThank you to the authors for addressing a number of my concerns, adding additional text and experiments. However, my main concern remains: if rate-distortion in an individual model is a useful method for evaluating generative models, you should compare it empirically with other metrics that have been proposed for this purpose (e.g. precision-recall). Additionally, while the theoretical novelty of getting the full R-D curve from a single AIS run is very cool, I'm skeptical of the practical utility as a metric for generative models due to the computational costs of AIS (4-7 hours for 50 images on MNIST). The simple baseline I suggested of training an amortized inference network with beta annealed over time would not require training a separate encoder for each point in R-D, you could just start beta large, and anneal beta in steps to 0 over time, tracing out an R-D curve. Given the current experiments, it's not obvious if the win of AIS in terms of accurate posterior inference is worth the increased computational cost over a simple VI baseline.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1572458629438}, {"id": "BygiiXj6KS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1374/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the rate-distortion tradeoffs of deep generative models such as variational autoencoders (VAEs) and generative adversarial networks (GANs).\nThe authors propose an annealed importance sampling (AIS) method to compute the rate-distortion curve efficiently. In experiments, the authors compare the rate-distortion curves for VAEs and GANs and discuss the properties of rate-distortion curves.\n\nThe method for computing the rate-distortion curves of deep generative models is interesting and the rate-prior distortion curve is promising as a performance measure. However, the main technical contribution of this work is the estimated AIS rate-prior distortion curve and it is based on a straight-forward application of AIS. \n\nIn fact, Sections 2 and 3 discuss already known result in literature although summarizing them in a paper is nice for readers.\n\nAlthough the findings in the experiments are interesting and insightful, they are still preliminary and further investigations are desirable.\n\nIn Section 5, the authors mention the consistency of their framework with Shannon’s rate distortion theorem. This seems to be a little overstatement because the authors discuss little about the optimization of the prior p(z).\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper considers the rate-distortion tradeoffs of deep generative models such as variational autoencoders (VAEs) and generative adversarial networks (GANs).\nThe authors propose an annealed importance sampling (AIS) method to compute the rate-distortion curve efficiently. In experiments, the authors compare the rate-distortion curves for VAEs and GANs and discuss the properties of rate-distortion curves.\n\nThe method for computing the rate-distortion curves of deep generative models is interesting and the rate-prior distortion curve is promising as a performance measure. However, the main technical contribution of this work is the estimated AIS rate-prior distortion curve and it is based on a straight-forward application of AIS. \n\nIn fact, Sections 2 and 3 discuss already known result in literature although summarizing them in a paper is nice for readers.\n\nAlthough the findings in the experiments are interesting and insightful, they are still preliminary and further investigations are desirable.\n\nIn Section 5, the authors mention the consistency of their framework with Shannon’s rate distortion theorem. This seems to be a little overstatement because the authors discuss little about the optimization of the prior p(z).\n\n"}, "tcdate": 1571824546575}, {"id": "ryldo2qvYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1374/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe paper proposes a new way to evaluate generative models that don't have tractable likelihoods, such as VAEs or GANs. Such generative models are composed of a prior over latent variables and a decoder that maps latent variables to data. The idea is to evaluate a trained model in terms of the best (lossy) compression rate that can be achieved by encoding a datapoint (e.g. an image) into the latent space, as a function of a permitted distortion between the datapoint and its reconstruction after decoding. The paper describes a method that estimates an upper bound on this rate-distortion curve using annealed importance sampling. The method is applied in evaluating and comparing a few VAE, GAN and AAE architectures on images (MNIST and CIFAR-10).\n\nOverall evaluation:\n\nThis is a very good paper, and I'm happy to recommend it for acceptance.\n\nThe problem considered (evaluating generative models with intractable likelihoods) is an interesting and important one. In general, such models are hard to evaluate and compare with each other. The paper proposes a new method for evaluating them, which can also improve our understanding of these models and potentially diagnose them in practice.\n\nThe method is well-motivated and backed by theoretical results. One clever aspect of the method is the way annealed importance sampling is used to approximate the rate-distortion curve: instead of sampling separately the rate for each distortion level with a different AIS run, a single AIS run is used to approximate the whole curve. This is done by taking the various points on the curve to correspond to intermediate distributions in AIS, which is quite clever.\n\nThe paper is well written, precise, and contains sufficient theoretical background to motivate the method.\n\nThe experiments are done carefully, and the results are interesting. I found particularly interesting the fact that VAEs behave differently to GANs (in terms of their rate-distortion tradeoff) when the dimensionality of the latent space is increased.\n\nSome discussion and critical feedback:\n\nI found the paper too long (10 full pages). I appreciate the detail, precision and depth of explanation, but I think it would be good to reduce the amount of text if possible.\n\nI though that the introduction was too specific to VAEs/GANs and to image modelling, which may give the impression that these are the main models/tasks of interest. I understand that these are the models and tasks that the paper is interested in, but I think it would be better if the introduction acknowledged the existence of other types of generative models (e.g. likelihood-based models such as autoregressive models and normalizing flows) and were less specific to image applications.\n\nProposition 3 is true almost by definition, since R_p is defined to be the minimum of all rate-distortion curves. I wonder if something more informative can be shown here. For example, my understanding is that the reason R^{AIS}_p is not optimal is due to the bias of the importance-sampling estimator in eq. (12). Since this bias asymptotically goes to zero, I suspect that R^{AIS}_p may become equal to R_p for M -> infinity, and perhaps the bound improves monotonically as M increases.\n\nIf my understanding is correct, the reason for the inequality in proposition 4 is that log\\hat{Z} is a biased estimate of logZ (due to Jensen's inequality) despite \\hat{Z} being an unbiased estimate of Z. If that's all there is to it, the proof in the appendix, although precise, is a bit of an overkill. It also seems to me that the log\\hat{Z} bias also approaches zero as M -> infinity, so this inequality maybe also becomes an equality asymptotically.\n\nIn future work, it would be interesting to also evaluate flow-based models (such as Glow). Since these models give exact likelihoods, it may be good to observe how the evaluation based on rate-distortion curves compares with a likelihood-based evaluation.\n\nSection 6 attributes the performance drop of VAEs in the low-rate regime to the \"holes problem\". If that's true, then I would expect the situation to be improved with more flexible prior / posterior models. What prior / posterior models were used in the experiments? If only diagonal Gaussians were used, then it would be interesting to see whether more flexible priors / posteriors such as normalizing flows would change the results.\n\nMinor corrections and suggestions for improvement:\n\n\"For continuous inputs, the metric is often dominated by the fine-grained distribution over pixels rather than the high-level structure\"\nThis statement is specific to images, not continuous inputs in general.\n\nOn the quantitative analysis of deep belief networks, https://dl.acm.org/citation.cfm?id=1390266\nis an older example of AIS used to evaluate generative models that could be cited.\n\nI think it would be better to drop z_0 and T_0 in equations (1) and (2), and have z_1 sampled from p_0 directly, to make the equations consistent with the equations that follow afterwards.\n\n\"using a latent variable z with a fixed prior distribution\"\nIn VAEs the prior can also be learned, it doesn't have to be fixed (this may in fact help alleviate the holes problem).\n\nIt could be mentioned that the objective in Eq. (9) has the same form as a generalized VAE objective, such as the one used by beta-VAE, https://openreview.net/forum?id=Sy2fzU9gl\n\nAt the bottom of page 4 and page 5, R_p(D) uses a different font for R than the rest of the paper.\n\nFig. 1 is too small to read on paper, I had to zoom in using the pdf in order to see it properly.\n\nLast paragraph of section 4 specifically mentions images, even though it doesn't need to (datapoints don't have to be images).\n\nThe last two paragraphs of page 7 contain a few grammatical mistakes:\n- fixing the architecture of neural network --> fixing the architecture of the neural network\n- as the result --> as a result\n- there exist a rate distortion code for any rate distortion pairs  -->  there exists a rate distortion code for any rate distortion pair\n- While in our definition of rate distortion -->Whereas, in out definition of rate distortion\n\nTop of page 8, use \\citet instead of \\citep where appropriate.\n\nCapitalize names and acronyms in references, such as ELBO, GAN, MMD, VAE, Bayes, Monte Carlo, etc.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Summary:\n\nThe paper proposes a new way to evaluate generative models that don't have tractable likelihoods, such as VAEs or GANs. Such generative models are composed of a prior over latent variables and a decoder that maps latent variables to data. The idea is to evaluate a trained model in terms of the best (lossy) compression rate that can be achieved by encoding a datapoint (e.g. an image) into the latent space, as a function of a permitted distortion between the datapoint and its reconstruction after decoding. The paper describes a method that estimates an upper bound on this rate-distortion curve using annealed importance sampling. The method is applied in evaluating and comparing a few VAE, GAN and AAE architectures on images (MNIST and CIFAR-10).\n\nOverall evaluation:\n\nThis is a very good paper, and I'm happy to recommend it for acceptance.\n\nThe problem considered (evaluating generative models with intractable likelihoods) is an interesting and important one. In general, such models are hard to evaluate and compare with each other. The paper proposes a new method for evaluating them, which can also improve our understanding of these models and potentially diagnose them in practice.\n\nThe method is well-motivated and backed by theoretical results. One clever aspect of the method is the way annealed importance sampling is used to approximate the rate-distortion curve: instead of sampling separately the rate for each distortion level with a different AIS run, a single AIS run is used to approximate the whole curve. This is done by taking the various points on the curve to correspond to intermediate distributions in AIS, which is quite clever.\n\nThe paper is well written, precise, and contains sufficient theoretical background to motivate the method.\n\nThe experiments are done carefully, and the results are interesting. I found particularly interesting the fact that VAEs behave differently to GANs (in terms of their rate-distortion tradeoff) when the dimensionality of the latent space is increased.\n\nSome discussion and critical feedback:\n\nI found the paper too long (10 full pages). I appreciate the detail, precision and depth of explanation, but I think it would be good to reduce the amount of text if possible.\n\nI though that the introduction was too specific to VAEs/GANs and to image modelling, which may give the impression that these are the main models/tasks of interest. I understand that these are the models and tasks that the paper is interested in, but I think it would be better if the introduction acknowledged the existence of other types of generative models (e.g. likelihood-based models such as autoregressive models and normalizing flows) and were less specific to image applications.\n\nProposition 3 is true almost by definition, since R_p is defined to be the minimum of all rate-distortion curves. I wonder if something more informative can be shown here. For example, my understanding is that the reason R^{AIS}_p is not optimal is due to the bias of the importance-sampling estimator in eq. (12). Since this bias asymptotically goes to zero, I suspect that R^{AIS}_p may become equal to R_p for M -> infinity, and perhaps the bound improves monotonically as M increases.\n\nIf my understanding is correct, the reason for the inequality in proposition 4 is that log\\hat{Z} is a biased estimate of logZ (due to Jensen's inequality) despite \\hat{Z} being an unbiased estimate of Z. If that's all there is to it, the proof in the appendix, although precise, is a bit of an overkill. It also seems to me that the log\\hat{Z} bias also approaches zero as M -> infinity, so this inequality maybe also becomes an equality asymptotically.\n\nIn future work, it would be interesting to also evaluate flow-based models (such as Glow). Since these models give exact likelihoods, it may be good to observe how the evaluation based on rate-distortion curves compares with a likelihood-based evaluation.\n\nSection 6 attributes the performance drop of VAEs in the low-rate regime to the \"holes problem\". If that's true, then I would expect the situation to be improved with more flexible prior / posterior models. What prior / posterior models were used in the experiments? If only diagonal Gaussians were used, then it would be interesting to see whether more flexible priors / posteriors such as normalizing flows would change the results.\n\nMinor corrections and suggestions for improvement:\n\n\"For continuous inputs, the metric is often dominated by the fine-grained distribution over pixels rather than the high-level structure\"\nThis statement is specific to images, not continuous inputs in general.\n\nOn the quantitative analysis of deep belief networks, https://dl.acm.org/citation.cfm?id=1390266\nis an older example of AIS used to evaluate generative models that could be cited.\n\nI think it would be better to drop z_0 and T_0 in equations (1) and (2), and have z_1 sampled from p_0 directly, to make the equations consistent with the equations that follow afterwards.\n\n\"using a latent variable z with a fixed prior distribution\"\nIn VAEs the prior can also be learned, it doesn't have to be fixed (this may in fact help alleviate the holes problem).\n\nIt could be mentioned that the objective in Eq. (9) has the same form as a generalized VAE objective, such as the one used by beta-VAE, https://openreview.net/forum?id=Sy2fzU9gl\n\nAt the bottom of page 4 and page 5, R_p(D) uses a different font for R than the rest of the paper.\n\nFig. 1 is too small to read on paper, I had to zoom in using the pdf in order to see it properly.\n\nLast paragraph of section 4 specifically mentions images, even though it doesn't need to (datapoints don't have to be images).\n\nThe last two paragraphs of page 7 contain a few grammatical mistakes:\n- fixing the architecture of neural network --> fixing the architecture of the neural network\n- as the result --> as a result\n- there exist a rate distortion code for any rate distortion pairs  -->  there exists a rate distortion code for any rate distortion pair\n- While in our definition of rate distortion -->Whereas, in out definition of rate distortion\n\nTop of page 8, use \\citet instead of \\citep where appropriate.\n\nCapitalize names and acronyms in references, such as ELBO, GAN, MMD, VAE, Bayes, Monte Carlo, etc."}, "tcdate": 1571429536007}], "openreview_url": "https://openreview.net/forum?id=ryga2CNKDH", "arxiv_id": "2008.06653", "paper_pdf": "papers/ryga2CNKDH.pdf", "paper_pdf_sha256": "4353dceb005eae52a14bd303432db1cbce657e8a2b93c598d3185dc925f9bd3b", "paper_pdf_bytes": 1447534, "paper_pdf_source": "openreview", "code_url": "https://github.com/BorealisAI/rate_distortion", "code_repository": "BorealisAI/rate_distortion", "code_commit": "13b60b94c65199bd228292f301151a3fd54dcd89", "code_archive": "repos/ryga2CNKDH.zip", "code_archive_sha256": "3dd1201d2cf753888f2d71c421fc7cd39a12fbb4baba26820f9b7279606678d7", "code_archive_bytes": 95209, "code_file_count": 57, "code_extensions": {".py": 49, ".sh": 8}, "github_disk_usage_kb": 122, "github_languages": {"Python": 275869, "Shell": 16034}, "github_archived": false, "github_pushed_at": "2021-01-20T07:31:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evaluating-lossy-compression-rates-of-deep"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0LwWImlpCP", "year": 2026, "status": "rejected", "title": "Electrostatics from Laplacian Eigenbasis for Neural Network Interatomic Potentials", "authors": ["Maksim Zhdanov", "Vladislav Kurenkov"], "authorids": ["~Maksim_Zhdanov2", "~Vladislav_Kurenkov1"], "authors_source": "OpenReview API", "abstract": "In this work, we introduce $\\Phi$-Module, a universal plugin module that enforces Poisson’s equation within the message-passing framework to learn electrostatic interactions in a self-supervised manner. Specifically, each atom-wise representation is encouraged to satisfy a discretized Poisson's equation, making it possible to acquire a potential $\\boldsymbol{\\phi}$ and a corresponding charges $\\boldsymbol{\\rho}$ linked to the learnable Laplacian eigenbasis coefficients of a given molecular graph. We then derive an electrostatic energy term, crucial for improved total energy predictions. This approach integrates seamlessly into any existing neural potential with insignificant computational overhead. Our results underscore how embedding a first-principles constraint in neural interatomic potentials can significantly improve performance while remaining hyperparameter-friendly, memory-efficient and lightweight in training.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "qzkP0wlSnn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18263/Reviewer_GKpD"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper introduces Φ-Module, a physics-informed plugin designed to integrate electrostatic interactions into graph neural networks in a self-supervised way. By enforcing Poisson’s equation within the message-passing framework, Φ-Module enables models to learn atomic potentials and charges represented in the Laplacian eigenbasis of molecular graphs. This approach captures long-range electrostatic effects that standard local message passing often misses. The module includes a lightweight subnetwork, α-Net, that predicts eigenbasis coefficients, allowing derivation of an electrostatic energy term that improves total energy predictions with minimal computational overhead. Experiments on the OE62 and MD22 benchmarks show accuracy gains across several neural potentials", "review_text": "The paper introduces Φ-Module, a physics-informed plugin designed to integrate electrostatic interactions into graph neural networks in a self-supervised way. By enforcing Poisson’s equation within the message-passing framework, Φ-Module enables models to learn atomic potentials and charges represented in the Laplacian eigenbasis of molecular graphs. This approach captures long-range electrostatic effects that standard local message passing often misses. The module includes a lightweight subnetwork, α-Net, that predicts eigenbasis coefficients, allowing derivation of an electrostatic energy term that improves total energy predictions with minimal computational overhead. Experiments on the OE62 and MD22 benchmarks show accuracy gains across several neural potentials", "strengths": "1. The paper proposes an integration of a first-principles physical law (Poisson’s equation) into GNN-based interatomic potentials. By embedding the Poisson constraint, the model learns to produce physically meaningful electrostatic potentials and charges in a self-supervised way, without requiring any ground-truth charges or external fields.\n2. Φ-Module is designed as a universal augmentation that can be attached to essentially any GNN architecture for molecules. The authors demonstrate this generality by incorporating Φ-Module into multiple established models \n3. A claimed advantage is that Φ-Module introduces very little overhead. The spectral α-Net and Laplacian eigenbasis computation are lightweight, adding roughly 5–10% to training time per epoch and a modest amount of memory usage.", "weaknesses": "1. The biggest weakness is generalizability and scalability. Experiments are restricted to OE62 and MD22. There’s no end-to-end training on truly realistic, large, diverse, million–sample datasets (e.g., OMol 25) or cross-dataset transfer demonstrating generalization across broader chemistries. As a result, it’s unclear how Φ-Module scales in sample size or generalizes to diverse systems.\n2. Although the overall results favor Φ-Module, the improvements are not uniformly overwhelming. In OE62, one baseline GNN saw only ~5% error improvement with Φ-Module, which is relatively modest. On MD22, the Φ-augmented model did not win on every single metric – the original ViSNet still had the best outcome on 2 of the 14 comparisons, and in 3 of 14 cases Φ-Module failed to set a new state-of-the-art. This indicates that the benefits, while present, can be incremental, task-dependent, or just random (no error bar / std is given)\n3. As a plug-in, Φ-Module adds four new hyperparameters and extra computations (solving for eigenvectors) to a model. While the authors argue this is lightweight, one might be concerned about implementation complexity.", "questions": "1. Could you clarify how the Laplacian eigenpairs are computed during training? The paper suggests using a fixed number k of eigenvectors and a batched eigendecomposition approach. Is this done via an iterative solver each message-passing step, or are eigenvalues computed once per epoch/structure and reused? \n2. The paper references alternatives like adding Ewald summation to GNNs or using pre-computed partial charges. Did you consider comparing Φ-Module’s performance to such methods? \n3. Beyond the tasks in this paper, how general is Φ-Module’s applicability? For instance, can it handle systems with periodic boundary conditions (common in materials simulations where Ewald is often needed)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Φ-Module, a physics-informed plugin designed to integrate electrostatic interactions into graph neural networks in a self-supervised way. By enforcing Poisson’s equation within the message-passing framework, Φ-Module enables models to learn atomic potentials and charges represented in the Laplacian eigenbasis of molecular graphs. This approach captures long-range electrostatic effects that standard local message passing often misses. The module includes a lightweight subnetwork, α-Net, that predicts eigenbasis coefficients, allowing derivation of an electrostatic energy term that improves total energy predictions with minimal computational overhead. Experiments on the OE62 and MD22 benchmarks show accuracy gains across several neural potentials", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper proposes an integration of a first-principles physical law (Poisson’s equation) into GNN-based interatomic potentials. By embedding the Poisson constraint, the model learns to produce physically meaningful electrostatic potentials and charges in a self-supervised way, without requiring any ground-truth charges or external fields.\n2. Φ-Module is designed as a universal augmentation that can be attached to essentially any GNN architecture for molecules. The authors demonstrate this generality by incorporating Φ-Module into multiple established models \n3. A claimed advantage is that Φ-Module introduces very little overhead. The spectral α-Net and Laplacian eigenbasis computation are lightweight, adding roughly 5–10% to training time per epoch and a modest amount of memory usage.", "weaknesses": "1. The biggest weakness is generalizability and scalability. Experiments are restricted to OE62 and MD22. There’s no end-to-end training on truly realistic, large, diverse, million–sample datasets (e.g., OMol 25) or cross-dataset transfer demonstrating generalization across broader chemistries. As a result, it’s unclear how Φ-Module scales in sample size or generalizes to diverse systems.\n2. Although the overall results favor Φ-Module, the improvements are not uniformly overwhelming. In OE62, one baseline GNN saw only ~5% error improvement with Φ-Module, which is relatively modest. On MD22, the Φ-augmented model did not win on every single metric – the original ViSNet still had the best outcome on 2 of the 14 comparisons, and in 3 of 14 cases Φ-Module failed to set a new state-of-the-art. This indicates that the benefits, while present, can be incremental, task-dependent, or just random (no error bar / std is given)\n3. As a plug-in, Φ-Module adds four new hyperparameters and extra computations (solving for eigenvectors) to a model. While the authors argue this is lightweight, one might be concerned about implementation complexity.", "questions": "1. Could you clarify how the Laplacian eigenpairs are computed during training? The paper suggests using a fixed number k of eigenvectors and a batched eigendecomposition approach. Is this done via an iterative solver each message-passing step, or are eigenvalues computed once per epoch/structure and reused? \n2. The paper references alternatives like adding Ewald summation to GNNs or using pre-computed partial charges. Did you consider comparing Φ-Module’s performance to such methods? \n3. Beyond the tasks in this paper, how general is Φ-Module’s applicability? For instance, can it handle systems with periodic boundary conditions (common in materials simulations where Ewald is often needed)?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762079319877}, {"id": "6Cys8Q4YDy", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18263/Reviewer_gZro"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper introduces a Phi module, a module that can be added to any MLIP architecture to introduce a long-range electro statics inductive bias. They imitate the computational structure of a poission equation to learn a sort of latent charges that can be trained self-consistently.", "review_text": "The paper introduces a Phi module, a module that can be added to any MLIP architecture to introduce a long-range electro statics inductive bias. They imitate the computational structure of a poission equation to learn a sort of latent charges that can be trained self-consistently.", "strengths": "-The method is well motivated and makes intuitive sense\n-The experimental evaluation is thorough\n-The improvements are very consistent", "weaknesses": "-The accuracy improvements are rather small\n-The spectral decomposition introduces an N^2 scaling operation, which could become problematic for larger-scale simulations. The paper only benchmarks memory, but not runtime with system size; this should be benchmarked and could change my opinion\n-There have been works before that incorporate explicit charge equilibrium/coulomb interactions, in particular https://www.nature.com/articles/s41467-020-20427-2 . A comparison would be appropriate\n- The phi module only targets electrostatics and not other long range effects, which may make fully learned approaches preferable", "questions": "- Do you use a dense connectivity graph L? And why is it weighted by d_ij, doenst this imply that far away atoms interact stronger than closer ones?\n- Why did you use VisNet in favour of a newer architecture?\n\nRemark:\n\"In quantum chemistry, the task of correct prediction of atomic energies is paramount, but stands\na great challenge…” Atomic energies are not a well defined concept, did you mean molecular energies?\n“Some of those require prior data in the form of partial charges or\ndipole moments, which is costly to retrieve using DFT”, If the DFT calculation is already converged to get the molecular energies, it is trivial to get dipoles or partial charges at negligible costs\n“...and gives the opportunity to process large macromolecules with the Φ-Module. This decision also keeps us away from the ambiguity of invariance and sorting of eigenvalues and eigenvectors during their computations - we strictly get k-selected eigenvalues and their corresponding eigenvectors without the need to sort them anyhow.” This is not convincing, if the Laplacina has a degenerate eigenspace the order will be arbitrary and specifics will be subject to numerical noise, a conjugated gradient solver doesnt change this", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a Phi module, a module that can be added to any MLIP architecture to introduce a long-range electro statics inductive bias. They imitate the computational structure of a poission equation to learn a sort of latent charges that can be trained self-consistently.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "-The method is well motivated and makes intuitive sense\n-The experimental evaluation is thorough\n-The improvements are very consistent", "weaknesses": "-The accuracy improvements are rather small\n-The spectral decomposition introduces an N^2 scaling operation, which could become problematic for larger-scale simulations. The paper only benchmarks memory, but not runtime with system size; this should be benchmarked and could change my opinion\n-There have been works before that incorporate explicit charge equilibrium/coulomb interactions, in particular https://www.nature.com/articles/s41467-020-20427-2 . A comparison would be appropriate\n- The phi module only targets electrostatics and not other long range effects, which may make fully learned approaches preferable", "questions": "- Do you use a dense connectivity graph L? And why is it weighted by d_ij, doenst this imply that far away atoms interact stronger than closer ones?\n- Why did you use VisNet in favour of a newer architecture?\n\nRemark:\n\"In quantum chemistry, the task of correct prediction of atomic energies is paramount, but stands\na great challenge…” Atomic energies are not a well defined concept, did you mean molecular energies?\n“Some of those require prior data in the form of partial charges or\ndipole moments, which is costly to retrieve using DFT”, If the DFT calculation is already converged to get the molecular energies, it is trivial to get dipoles or partial charges at negligible costs\n“...and gives the opportunity to process large macromolecules with the Φ-Module. This decision also keeps us away from the ambiguity of invariance and sorting of eigenvalues and eigenvectors during their computations - we strictly get k-selected eigenvalues and their corresponding eigenvectors without the need to sort them anyhow.” This is not convincing, if the Laplacina has a degenerate eigenspace the order will be arbitrary and specifics will be subject to numerical noise, a conjugated gradient solver doesnt change this", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762043218675}, {"id": "x7cdhiOgA5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18263/Reviewer_H2RH"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents a module for learning the electrostatic interactions for interatomic potentials based on sparse graph neural networks and Poisson equation.", "review_text": "This paper presents a module for learning the electrostatic interactions for interatomic potentials based on sparse graph neural networks and Poisson equation.", "strengths": "- The proposed method is simple yet performative.\n- The paper is very well presented with the advantage of the designed model clearly highlighted.\n- The tradeoff between efficiency and performance is discussed in detail.\n- The problem addressed in this paper is of importance to the molecular modeling community.", "weaknesses": "- There seems to be limited novelty in the proposed graph representation module.", "questions": "- Could you kindly elaborate more on the choice of using spectral graph neural networks, rather than the spatial ones, which are more popular?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a module for learning the electrostatic interactions for interatomic potentials based on sparse graph neural networks and Poisson equation.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The proposed method is simple yet performative.\n- The paper is very well presented with the advantage of the designed model clearly highlighted.\n- The tradeoff between efficiency and performance is discussed in detail.\n- The problem addressed in this paper is of importance to the molecular modeling community.", "weaknesses": "- There seems to be limited novelty in the proposed graph representation module.", "questions": "- Could you kindly elaborate more on the choice of using spectral graph neural networks, rather than the spatial ones, which are more popular?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762010766311}, {"id": "TNF7WGIToC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18263/Reviewer_oDat"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This work introduces $\\Phi$-Module, an extension to atom-based machine learning interatomic potentials that intends to resolve long-range energy contributions. $\\Phi$-Module is derived from the discretized Poisson equation for electrostatics and uses the graph Laplacian for efficient long-range propagation. In their experimental evaluation, the authors find the $\\Phi$-Module to be an effficient addition to MLIAP, reducing errors at low computational overheads.", "review_text": "This work introduces $\\Phi$-Module, an extension to atom-based machine learning interatomic potentials that intends to resolve long-range energy contributions. $\\Phi$-Module is derived from the discretized Poisson equation for electrostatics and uses the graph Laplacian for efficient long-range propagation. In their experimental evaluation, the authors find the $\\Phi$-Module to be an effficient addition to MLIAP, reducing errors at low computational overheads.", "strengths": "* The proposed $\\Phi$-module presents a novel addition to the field of MLFFs.\n* The experiments suggest a very valuable extensions with favorable runtime-accuracy tradeoff.\n* $\\Phi$-module appears quite modular and presents an easy integration.", "weaknesses": "1. The authors claim that their method is self-supervised and does not need external labeled data (l.50-51 + abstract). However, the losses introduced in l.174 and l.179 do not ensure that predictions improve, the method additionally **needs** the standard supervised learning loss.\n2. The use of 1D convolutions over the nuclei breaks the permutation invariance. I find the paper lacks to discuss this disadvantage. One could in general neglect permutation invariance in MLFFs, does that yield similar improvements?\n3. The authors leave a lot of questions open or are not specific in several key areas, see questions.\n\n\nMinor:\n* l.26-33: DFT and MLFFs are quite different things and the section suggest they accomplish the same.\n* l.34 which alternatives?\n* Table 1 is never referenced.\n* l.432 formatting", "questions": "1. What is used as graph connectivity for the graph laplacian? A radial cutoff or molecular bonds? In either case, doesn't bond-breaking or going out of the cutoff radius, introduces jumps in the energy surface?\n2. Why is the Laplacian weighted by the pairwise distance, intuitively, I'd expect it to be weighted by the inverse of the distance?\n3. l.173 is the same alpha-Net used for phi and rho?\n4. Couldn't one enforce the PDE and net zero loss analytically? How does that compare to training?\n5. What is the x-axis in Figure 4? (Also the figure is never referenced)\n6. Table 1: What about other MLFF+Phi module combinations?\n7. Are the hyperparameter optimization from l.684 used for all experiments? This seems quite excessive given that no hyperparameter optimization is done for any of the baseline models.\n8. What units are Table 4 and what target, which dataset, etc?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces $\\Phi$-Module, an extension to atom-based machine learning interatomic potentials that intends to resolve long-range energy contributions. $\\Phi$-Module is derived from the discretized Poisson equation for electrostatics and uses the graph Laplacian for efficient long-range propagation. In their experimental evaluation, the authors find the $\\Phi$-Module to be an effficient addition to MLIAP, reducing errors at low computational overheads.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "* The proposed $\\Phi$-module presents a novel addition to the field of MLFFs.\n* The experiments suggest a very valuable extensions with favorable runtime-accuracy tradeoff.\n* $\\Phi$-module appears quite modular and presents an easy integration.", "weaknesses": "1. The authors claim that their method is self-supervised and does not need external labeled data (l.50-51 + abstract). However, the losses introduced in l.174 and l.179 do not ensure that predictions improve, the method additionally **needs** the standard supervised learning loss.\n2. The use of 1D convolutions over the nuclei breaks the permutation invariance. I find the paper lacks to discuss this disadvantage. One could in general neglect permutation invariance in MLFFs, does that yield similar improvements?\n3. The authors leave a lot of questions open or are not specific in several key areas, see questions.\n\n\nMinor:\n* l.26-33: DFT and MLFFs are quite different things and the section suggest they accomplish the same.\n* l.34 which alternatives?\n* Table 1 is never referenced.\n* l.432 formatting", "questions": "1. What is used as graph connectivity for the graph laplacian? A radial cutoff or molecular bonds? In either case, doesn't bond-breaking or going out of the cutoff radius, introduces jumps in the energy surface?\n2. Why is the Laplacian weighted by the pairwise distance, intuitively, I'd expect it to be weighted by the inverse of the distance?\n3. l.173 is the same alpha-Net used for phi and rho?\n4. Couldn't one enforce the PDE and net zero loss analytically? How does that compare to training?\n5. What is the x-axis in Figure 4? (Also the figure is never referenced)\n6. Table 1: What about other MLFF+Phi module combinations?\n7. Are the hyperparameter optimization from l.684 used for all experiments? This seems quite excessive given that no hyperparameter optimization is done for any of the baseline models.\n8. What units are Table 4 and what target, which dataset, etc?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761917085975}, {"id": "1YWKCNP0ES", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18263/Reviewer_QJiW"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper introduces Φ-Module, a plug-and-play Laplacian-eigenbasis module that embeds the Poisson equation as a self-supervised physical constraint within graph neural network (GNN) interatomic potentials.", "review_text": "This paper introduces Φ-Module, a plug-and-play Laplacian-eigenbasis module that embeds the Poisson equation as a self-supervised physical constraint within graph neural network (GNN) interatomic potentials.", "strengths": "The proposed module can predict atomic potential ϕ and charge ρ from learned eigenbasis coefficients and adds an electrostatic energy term to improve total-energy and force predictions.", "weaknesses": "The paper repeatedly claims that Φ-Module captures non-local electrostatic interactions, yet no benchmark explicitly demonstrates this. Improvements in energy and force MAE alone are insufficient proof of non-local interaction modeling.\n\nFor MD22, the authors selected two of the smallest molecules, while the claim centers on modeling non-local interactions. Larger systems such as Ac-Ala15-NHMe or DHA are more appropriate to evaluate the alleged long-range capability. Without them, the argument remains speculative.\n\nOn the OE62 experiments, the baselines (SchNet, DimeNet++, PaiNN, etc.) are 2019–2021 architectures. Recent high-order equivariant models—MACE, eSCN, NequIP, Equiformer-V2, ViSNet—represent the current state of the field. Because the Φ-Module is advertised as a general plug-in, results on these modern architectures are essential for a credible evaluation. \n\nOn the MD22 experiments, the reported performance differences from the original ViSNet paper are unusually large for the same datasets. The authors should clarify training settings (data splits, learning rates, cutoff, etc.) and reproduce the original ViSNet baseline under identical conditions.", "questions": "See Section Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Φ-Module, a plug-and-play Laplacian-eigenbasis module that embeds the Poisson equation as a self-supervised physical constraint within graph neural network (GNN) interatomic potentials.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The proposed module can predict atomic potential ϕ and charge ρ from learned eigenbasis coefficients and adds an electrostatic energy term to improve total-energy and force predictions.", "weaknesses": "The paper repeatedly claims that Φ-Module captures non-local electrostatic interactions, yet no benchmark explicitly demonstrates this. Improvements in energy and force MAE alone are insufficient proof of non-local interaction modeling.\n\nFor MD22, the authors selected two of the smallest molecules, while the claim centers on modeling non-local interactions. Larger systems such as Ac-Ala15-NHMe or DHA are more appropriate to evaluate the alleged long-range capability. Without them, the argument remains speculative.\n\nOn the OE62 experiments, the baselines (SchNet, DimeNet++, PaiNN, etc.) are 2019–2021 architectures. Recent high-order equivariant models—MACE, eSCN, NequIP, Equiformer-V2, ViSNet—represent the current state of the field. Because the Φ-Module is advertised as a general plug-in, results on these modern architectures are essential for a credible evaluation. \n\nOn the MD22 experiments, the reported performance differences from the original ViSNet paper are unusually large for the same datasets. The authors should clarify training settings (data splits, learning rates, cutoff, etc.) and reproduce the original ViSNet baseline under identical conditions.", "questions": "See Section Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1760545341110}], "openreview_url": "https://openreview.net/forum?id=0LwWImlpCP", "arxiv_id": "2505.14606", "paper_pdf": "papers/0LwWImlpCP.pdf", "paper_pdf_sha256": "3e2df0f85baedac9ddc8c127d36eebcb1fde559e7d5c386ee4499806d100dfa5", "paper_pdf_bytes": 818079, "paper_pdf_source": "openreview", "code_url": "https://github.com/dunnolab/phi-module", "code_repository": "dunnolab/phi-module", "code_commit": "730c4ddb5be07d6f67a5a42db77f923e2de5fac0", "code_archive": "repos/0LwWImlpCP.zip", "code_archive_sha256": "47edc2cf26fa67e6c5a9e4f0d1470a53b56bba6c4d904a289c09cecc6a313bfb", "code_archive_bytes": 505301, "code_file_count": 66, "code_extensions": {".py": 66}, "github_disk_usage_kb": 473, "github_languages": {"Python": 546165, "Dockerfile": 831}, "github_archived": false, "github_pushed_at": "2025-06-12T23:37:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/electrostatics-from-laplacian-eigenbasis-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3Xfa63ggsq", "year": 2025, "status": "rejected", "title": "AlignIQL: Policy Alignment in Implicit Q-Learning through Constrained Optimization", "authors": ["Longxiang He", "Li Shen", "Junbo Tan", "Xueqian Wang"], "authorids": ["~Longxiang_He2", "~Li_Shen1", "~Junbo_Tan1", "~Xueqian_Wang1"], "authors_source": "OpenReview API", "abstract": "Implicit Q-learning (IQL) serves as a strong baseline for offline RL, which never needs to evaluate actions outside of the dataset through quantile regression. However, it is unclear how to recover the implicit policy from the learned implicit Q-function and whether IQL can utilize weighted regression for policy extraction. IDQL reinterprets IQL as an actor-critic method and gets weights of implicit policy, however, this weight only holds for the optimal value function under certain critic loss functions. In this work, we introduce a different way to solve the $\\textit{implicit policy-finding problem}$ (IPF) by formulating this problem as an optimization problem. Based on this optimization problem, we further propose two practical algorithms AlignIQL and AlignIQL-hard, which inherit the advantages of decoupling actor from critic in IQL and provide insights into why IQL can use weighted regression for policy extraction. Compared with IQL and IDQL, we find that our method keeps the simplicity of IQL and solves the implicit policy-finding problem.  Experimental results on D4RL datasets show that our method achieves competitive or superior results compared with other SOTA offline RL methods. Especially in complex sparse reward tasks like AntMaze and Adroit, our method outperforms IQL and IDQL by a significant margin.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "vXhgplioho", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1875/Reviewer_g3hi"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper considers policy extraction problem, where sometimes in offline RL, existing algorithms only learn value function, and policy extraction problem is to find a policy that coorespond to the policy and does not perform OOD actions. The paper considers distilling policies from value functions learned with IQL algorithm, and propose the implicit policy-finding problem. The solution of the IPF problem leads to the proposed AlignIQL algorithm, from a careful derivation of the IPF formulation. In the experiment, the proposed method is compared with several baselines with competitive performance.", "review_text": "The paper considers policy extraction problem, where sometimes in offline RL, existing algorithms only learn value function, and policy extraction problem is to find a policy that coorespond to the policy and does not perform OOD actions. The paper considers distilling policies from value functions learned with IQL algorithm, and propose the implicit policy-finding problem. The solution of the IPF problem leads to the proposed AlignIQL algorithm, from a careful derivation of the IPF formulation. In the experiment, the proposed method is compared with several baselines with competitive performance.", "strengths": "1. The proposed method is derived rigorously.\n2. The experiment shows that the proposed method has good empirical performance compared with other baselines on standard benchmarks.", "weaknesses": "1. The formulation aims to use a general regularization function $f$, which is a good attempt. However, the remaining results seems to rely on the case that $f(x) = \\log(x)$. Does the result generalize to any other regularization function?\n2. Remark 5.7 seems very hand-wavy. How does the algorithm ensure that the action with the positive advantage is chosen? It does not seem to be reflected in the loss function. \n3. While the result in table 1 looks impressive, I am not sure if this can serve a strong evidence that the proposed method is better than AWR. The proposed method is equipped with diffusion policies, but the IQL (AWR) baseline seem to only use MLP so it might not be a fair comparison. \n4. The result in table 1 is missing standard deviation. \n5. The goal of section 6.2 is unclear. What is the baseline that is compared against in this section? \n6. Some minor issues: a) in eq. 1, is a $\\pi(a \\mid s)$ missing? b) in eq. 2, where is the $Q_{\\theta}$ from? It does not appear eq. 1. c) in eq. 7, the notation $a$ is overloaded in $a' \\sim \\pi(a \\mid s)$.", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers policy extraction problem, where sometimes in offline RL, existing algorithms only learn value function, and policy extraction problem is to find a policy that coorespond to the policy and does not perform OOD actions. The paper considers distilling policies from value functions learned with IQL algorithm, and propose the implicit policy-finding problem. The solution of the IPF problem leads to the proposed AlignIQL algorithm, from a careful derivation of the IPF formulation. In the experiment, the proposed method is compared with several baselines with competitive performance.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The proposed method is derived rigorously.\n2. The experiment shows that the proposed method has good empirical performance compared with other baselines on standard benchmarks.", "weaknesses": "1. The formulation aims to use a general regularization function $f$, which is a good attempt. However, the remaining results seems to rely on the case that $f(x) = \\log(x)$. Does the result generalize to any other regularization function?\n2. Remark 5.7 seems very hand-wavy. How does the algorithm ensure that the action with the positive advantage is chosen? It does not seem to be reflected in the loss function. \n3. While the result in table 1 looks impressive, I am not sure if this can serve a strong evidence that the proposed method is better than AWR. The proposed method is equipped with diffusion policies, but the IQL (AWR) baseline seem to only use MLP so it might not be a fair comparison. \n4. The result in table 1 is missing standard deviation. \n5. The goal of section 6.2 is unclear. What is the baseline that is compared against in this section? \n6. Some minor issues: a) in eq. 1, is a $\\pi(a \\mid s)$ missing? b) in eq. 2, where is the $Q_{\\theta}$ from? It does not appear eq. 1. c) in eq. 7, the notation $a$ is overloaded in $a' \\sim \\pi(a \\mid s)$.", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730690495050}, {"id": "5SNjeeHFR9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1875/Reviewer_pv7F"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes AlignIQL (in two versions) to address the implicit policy-finding problem. The authors formulate it as a constrained optimization problem and derive a closed-form solution. The performance of AlignIQL is competitive compared to the baselines.", "review_text": "This paper proposes AlignIQL (in two versions) to address the implicit policy-finding problem. The authors formulate it as a constrained optimization problem and derive a closed-form solution. The performance of AlignIQL is competitive compared to the baselines.", "strengths": "- This paper introduces a new approach to tackle the implicit policy-finding problem, combining theoretical rigor with practical effectiveness in offline RL.\n- The proposed algorithm, AlignIQL, performs well across varied tasks, demonstrating versatility and effectiveness across different offline RL benchmarks.", "weaknesses": "- While AlignIQL is rigorous, it adds complexity to training by requiring additional multiplier networks and diffusion models, which may increase computational costs and sensitivity to hyperparameters. The scalability of the method is also a concern; can it be extended to image-based tasks?\n- The authors do not explain the use of diffusion modeling in the methods section.\n- The performance of AlignIQL raises some concerns:\n    - The authors argue that MuJoCo tasks are already saturated for offline RL, which I agree with. However, AlignIQL's performance is also considerably worse than Diffusion QL and even worse than IQL in 4 out of 9 tasks. Given that AlignIQL consumes more computational resources, this discrepancy is problematic.\n    - There is a significant performance difference between the authors' version and the original IDQL paper, which further leaves the reader uncertain about the supposed improvements in AlignIQL's performance.\n    - The results were obtained using inconsistent hyperparameters, yet the authors under-analyze the ablation study and hyperparameter sensitivity.\n    - The authors state, “Figure 2 shows that as training time increases, the performance of AlignIQL with different N converges to the same value, which shows that AlignIQL is insensitive to N.” This conclusion is not obvious from Figure 2. A clearer approach would be to report the mean and standard deviation of these scores.\n    - The optimization techniques may risk overfitting in AntMaze environments with sparse rewards, potentially reducing generalization to new scenarios. More testing on sparse reward tasks would benefit this submission.\n- In this submission, \"policy alignment\" is defined differently from its use in language models. A more formal definition of “policy alignment” should be provided, or the authors could consider renaming it.\n- A minor issue: multiple duplicate references appear in the bibliography (e.g., lines 568-573). Additionally, lines 916-917 may contain editing errors.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes AlignIQL (in two versions) to address the implicit policy-finding problem. The authors formulate it as a constrained optimization problem and derive a closed-form solution. The performance of AlignIQL is competitive compared to the baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- This paper introduces a new approach to tackle the implicit policy-finding problem, combining theoretical rigor with practical effectiveness in offline RL.\n- The proposed algorithm, AlignIQL, performs well across varied tasks, demonstrating versatility and effectiveness across different offline RL benchmarks.", "weaknesses": "- While AlignIQL is rigorous, it adds complexity to training by requiring additional multiplier networks and diffusion models, which may increase computational costs and sensitivity to hyperparameters. The scalability of the method is also a concern; can it be extended to image-based tasks?\n- The authors do not explain the use of diffusion modeling in the methods section.\n- The performance of AlignIQL raises some concerns:\n    - The authors argue that MuJoCo tasks are already saturated for offline RL, which I agree with. However, AlignIQL's performance is also considerably worse than Diffusion QL and even worse than IQL in 4 out of 9 tasks. Given that AlignIQL consumes more computational resources, this discrepancy is problematic.\n    - There is a significant performance difference between the authors' version and the original IDQL paper, which further leaves the reader uncertain about the supposed improvements in AlignIQL's performance.\n    - The results were obtained using inconsistent hyperparameters, yet the authors under-analyze the ablation study and hyperparameter sensitivity.\n    - The authors state, “Figure 2 shows that as training time increases, the performance of AlignIQL with different N converges to the same value, which shows that AlignIQL is insensitive to N.” This conclusion is not obvious from Figure 2. A clearer approach would be to report the mean and standard deviation of these scores.\n    - The optimization techniques may risk overfitting in AntMaze environments with sparse rewards, potentially reducing generalization to new scenarios. More testing on sparse reward tasks would benefit this submission.\n- In this submission, \"policy alignment\" is defined differently from its use in language models. A more formal definition of “policy alignment” should be provided, or the authors could consider renaming it.\n- A minor issue: multiple duplicate references appear in the bibliography (e.g., lines 568-573). Additionally, lines 916-917 may contain editing errors.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730386480280}, {"id": "zBHsmVNSJk", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1875/Reviewer_kMNe"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces AlignIQL, a novel approach to extracting implicit policies in offline reinforcement learning by formulating the implicit policy-finding problem as a constrained optimization problem. AlignIQL and its variant AlignIQL-hard leverage policy alignment constraints to ensure that the extracted policy reflects the learned Q-function while maintaining the advantages of decoupling the actor and critic in Implicit Q-Learning (IQL). The authors demonstrate that their method achieves competitive or superior performance on D4RL datasets, particularly excelling in complex sparse reward tasks like AntMaze and Adroit, while also being more robust to hyperparameter variations than existing methods. Additionally, the study provides theoretical insights into the conditions under which weighted regression can be effectively utilized for policy extraction in IQL. Overall, the proposed algorithms contribute to a better understanding of the bottlenecks in IQL-style methods and offer a more effective means for implicit policy extraction in offline reinforcement learning.", "review_text": "The paper introduces AlignIQL, a novel approach to extracting implicit policies in offline reinforcement learning by formulating the implicit policy-finding problem as a constrained optimization problem. AlignIQL and its variant AlignIQL-hard leverage policy alignment constraints to ensure that the extracted policy reflects the learned Q-function while maintaining the advantages of decoupling the actor and critic in Implicit Q-Learning (IQL). The authors demonstrate that their method achieves competitive or superior performance on D4RL datasets, particularly excelling in complex sparse reward tasks like AntMaze and Adroit, while also being more robust to hyperparameter variations than existing methods. Additionally, the study provides theoretical insights into the conditions under which weighted regression can be effectively utilized for policy extraction in IQL. Overall, the proposed algorithms contribute to a better understanding of the bottlenecks in IQL-style methods and offer a more effective means for implicit policy extraction in offline reinforcement learning.", "strengths": "- The introduction of AlignIQL as a constrained optimization approach represents a significant advancement in offline reinforcement learning, providing a fresh perspective on implicit policy extraction.\n- The empirical results demonstrate that AlignIQL and its variant achieve competitive performance across a variety of D4RL benchmarks, particularly in challenging tasks with sparse rewards, indicating the effectiveness of the proposed methods.\n- Theoretical Insights: The paper offers valuable theoretical analysis regarding the use of weighted regression for policy extraction, enhancing the understanding of the underlying mechanisms that contribute to the success of IQL methods.\n- By incorporating policy alignment constraints, the approach ensures that the extracted policies are both effective and representative of the learned Q-function, leading to improved stability and reliability in offline settings.\n- AlignIQL shows increased robustness to variations in hyperparameters compared to existing methods, which is crucial for practical applications where tuning can be challenging.", "weaknesses": "- While the experiments demonstrate competitive performance on specific D4RL benchmarks, the applicability of AlignIQL to other domains or more diverse environments may not be fully established, limiting its generalizability.\n- The proposed framework may introduce additional complexity in implementation compared to existing methods, which could deter practitioners who seek simpler solutions for offline reinforcement learning.\n- Although the paper includes comparisons with several baseline methods, it may benefit from a more comprehensive analysis against a broader range of state-of-the-art algorithms to fully contextualize its contributions.\n- The performance improvements may be contingent on the quality of the dataset used, raising concerns about the approach’s robustness in real-world scenarios where data can be noisy or incomplete.\n- The computational requirements for training AlignIQL could be higher than those of simpler methods, potentially limiting its scalability for larger-scale applications or real-time scenarios.", "questions": "- How does AlignIQL perform in real-world environments with noisy or incomplete datasets? The paper evaluates performance on D4RL benchmarks, but it would be interesting to see how the method handles imperfect data.\n- What are the key factors that affect the alignment between the Q-values and the learned policy in AlignIQL? Understanding the sensitivity of the method to different alignment parameters could help clarify its robustness.\n- How does the computational complexity of AlignIQL compare to other state-of-the-art offline reinforcement learning methods in terms of training time and resource usage? This would help evaluate the method’s scalability for larger or more complex tasks.\n- Is the approach compatible with more advanced neural network architectures, such as transformers, for offline reinforcement learning? Could integrating more modern architectures improve its performance?\n- What are the potential limitations of applying AlignIQL to tasks outside of continuous control, such as discrete action spaces or hierarchical tasks? This could provide insight into the method’s broader applicability.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces AlignIQL, a novel approach to extracting implicit policies in offline reinforcement learning by formulating the implicit policy-finding problem as a constrained optimization problem. AlignIQL and its variant AlignIQL-hard leverage policy alignment constraints to ensure that the extracted policy reflects the learned Q-function while maintaining the advantages of decoupling the actor and critic in Implicit Q-Learning (IQL). The authors demonstrate that their method achieves competitive or superior performance on D4RL datasets, particularly excelling in complex sparse reward tasks like AntMaze and Adroit, while also being more robust to hyperparameter variations than existing methods. Additionally, the study provides theoretical insights into the conditions under which weighted regression can be effectively utilized for policy extraction in IQL. Overall, the proposed algorithms contribute to a better understanding of the bottlenecks in IQL-style methods and offer a more effective means for implicit policy extraction in offline reinforcement learning.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The introduction of AlignIQL as a constrained optimization approach represents a significant advancement in offline reinforcement learning, providing a fresh perspective on implicit policy extraction.\n- The empirical results demonstrate that AlignIQL and its variant achieve competitive performance across a variety of D4RL benchmarks, particularly in challenging tasks with sparse rewards, indicating the effectiveness of the proposed methods.\n- Theoretical Insights: The paper offers valuable theoretical analysis regarding the use of weighted regression for policy extraction, enhancing the understanding of the underlying mechanisms that contribute to the success of IQL methods.\n- By incorporating policy alignment constraints, the approach ensures that the extracted policies are both effective and representative of the learned Q-function, leading to improved stability and reliability in offline settings.\n- AlignIQL shows increased robustness to variations in hyperparameters compared to existing methods, which is crucial for practical applications where tuning can be challenging.", "weaknesses": "- While the experiments demonstrate competitive performance on specific D4RL benchmarks, the applicability of AlignIQL to other domains or more diverse environments may not be fully established, limiting its generalizability.\n- The proposed framework may introduce additional complexity in implementation compared to existing methods, which could deter practitioners who seek simpler solutions for offline reinforcement learning.\n- Although the paper includes comparisons with several baseline methods, it may benefit from a more comprehensive analysis against a broader range of state-of-the-art algorithms to fully contextualize its contributions.\n- The performance improvements may be contingent on the quality of the dataset used, raising concerns about the approach’s robustness in real-world scenarios where data can be noisy or incomplete.\n- The computational requirements for training AlignIQL could be higher than those of simpler methods, potentially limiting its scalability for larger-scale applications or real-time scenarios.", "questions": "- How does AlignIQL perform in real-world environments with noisy or incomplete datasets? The paper evaluates performance on D4RL benchmarks, but it would be interesting to see how the method handles imperfect data.\n- What are the key factors that affect the alignment between the Q-values and the learned policy in AlignIQL? Understanding the sensitivity of the method to different alignment parameters could help clarify its robustness.\n- How does the computational complexity of AlignIQL compare to other state-of-the-art offline reinforcement learning methods in terms of training time and resource usage? This would help evaluate the method’s scalability for larger or more complex tasks.\n- Is the approach compatible with more advanced neural network architectures, such as transformers, for offline reinforcement learning? Could integrating more modern architectures improve its performance?\n- What are the potential limitations of applying AlignIQL to tasks outside of continuous control, such as discrete action spaces or hierarchical tasks? This could provide insight into the method’s broader applicability.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729737943167}], "openreview_url": "https://openreview.net/forum?id=3Xfa63ggsq", "arxiv_id": "2405.18187", "paper_pdf": "papers/3Xfa63ggsq.pdf", "paper_pdf_sha256": "990b2578d66c04189767d06767dd79e96efccb2253e5990c79f96c52c61846eb", "paper_pdf_bytes": 571327, "paper_pdf_source": "openreview", "code_url": "https://github.com/felix-thu/AlignIQL", "code_repository": "felix-thu/AlignIQL", "code_commit": "857f90b5fa355bb979f3ef6d997cbb20fc4b5ded", "code_archive": "repos/3Xfa63ggsq.zip", "code_archive_sha256": "e5081f67d032ef9b80ee2412f5109872c0406307f106f23e81d466b0c7bfe71e", "code_archive_bytes": 83667, "code_file_count": 71, "code_extensions": {".py": 71}, "github_disk_usage_kb": 54, "github_languages": {"Python": 168878}, "github_archived": false, "github_pushed_at": "2024-05-29T05:28:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/aligniql-policy-alignment-in-implicit-q"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sVl1KO5K76", "year": 2024, "status": "rejected", "title": "Momentum-SAM: Sharpness Aware Minimization without Computational Overhead", "authors": ["Marlon Becker", "Frederick Altrock", "Benjamin Risse"], "authorids": ["~Marlon_Becker1", "~Frederick_Altrock1", "~Benjamin_Risse1"], "authors_source": "OpenReview API", "abstract": "The recently proposed optimization algorithm for deep neural networks Sharpness Aware Minimization (SAM) suggests perturbing parameters before gradient calculation by a gradient ascent step to guide the optimization into parameter space regions of flat loss.\nWhile significant generalization improvements and thus reduction of overfitting could be demonstrated, the computational costs are doubled due to the additionally needed gradient calculation, making SAM unfeasible in case of limited computationally capacities.\nMotivated by Nesterov Accelerated Gradient (NAG) we propose Momentum-SAM (MSAM), which perturbs parameters in the direction of the accumulated momentum vector to achieve low sharpness without significant computational overhead or memory demands over SGD or Adam.\nWe evaluate MSAM in detail and reveal insights on separable mechanisms of NAG, SAM and MSAM regarding training optimization and generalization.\nCode is available at https://XXXXXXXX.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "VFz7PVssmn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5090/Reviewer_auP2"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposed an efficient sharpness-aware-minimization optimizer that does not need to calculate the inner optimization during the training which accelerates the training progress while keeping the model performance close to the original SAM. In order to obtain a good estimator of the gradient ascent direction in SAM, this paper uses the negative momentum direction as the estimation and shows that the model reaches a low sharpness and high performance. Since for most optimizers such as SGD and ADAM, momentum is used during training, there will be no computation overhead for the proposed method. MSAM achieves good empirical results on both cifar10 and ImageNet.", "review_text": "The paper proposed an efficient sharpness-aware-minimization optimizer that does not need to calculate the inner optimization during the training which accelerates the training progress while keeping the model performance close to the original SAM. In order to obtain a good estimator of the gradient ascent direction in SAM, this paper uses the negative momentum direction as the estimation and shows that the model reaches a low sharpness and high performance. Since for most optimizers such as SGD and ADAM, momentum is used during training, there will be no computation overhead for the proposed method. MSAM achieves good empirical results on both cifar10 and ImageNet.", "strengths": "The paper is well-written and easy to follow. The problem is important and the phenomenon that the direction of momentum is negatively correlated to the gradient direction of the current batch is very interesting. The empirical results are good at both cifar10 and ImageNet.", "weaknesses": "To me, the intuition of why the negative momentum direction is a good estimate of the gradient ascent direction is still not clear.  To be honest, the phenomenon why this happens is more important and interesting to me since there are quite a lot of methods that can efficiently simulate SAM. The paper does have some analysis showing that MSAM actually minimizes the sharpness, but I would love to have more intuitions or hypotheses. For example, is this dataset related? Is this task related, like only for classification tasks or only for vision problems?", "questions": "Can we have more experiments on finetuning results as the original SAM? \nCan we have some language model related experiments? \nAny hypotheses or intuitions of why the momentum direction is so different from the current batch? \nFor Section 4.3, can you make it clear why it is related to m-sharpness?\nCan we have some ImageNet NAG results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed an efficient sharpness-aware-minimization optimizer that does not need to calculate the inner optimization during the training which accelerates the training progress while keeping the model performance close to the original SAM. In order to obtain a good estimator of the gradient ascent direction in SAM, this paper uses the negative momentum direction as the estimation and shows that the model reaches a low sharpness and high performance. Since for most optimizers such as SGD and ADAM, momentum is used during training, there will be no computation overhead for the proposed method. MSAM achieves good empirical results on both cifar10 and ImageNet.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper is well-written and easy to follow. The problem is important and the phenomenon that the direction of momentum is negatively correlated to the gradient direction of the current batch is very interesting. The empirical results are good at both cifar10 and ImageNet.", "weaknesses": "To me, the intuition of why the negative momentum direction is a good estimate of the gradient ascent direction is still not clear.  To be honest, the phenomenon why this happens is more important and interesting to me since there are quite a lot of methods that can efficiently simulate SAM. The paper does have some analysis showing that MSAM actually minimizes the sharpness, but I would love to have more intuitions or hypotheses. For example, is this dataset related? Is this task related, like only for classification tasks or only for vision problems?", "questions": "Can we have more experiments on finetuning results as the original SAM? \nCan we have some language model related experiments? \nAny hypotheses or intuitions of why the momentum direction is so different from the current batch? \nFor Section 4.3, can you make it clear why it is related to m-sharpness?\nCan we have some ImageNet NAG results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698803855285}, {"id": "qg7qKDWlvt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5090/Reviewer_kkxn"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors propose Momentum-SAM (MSAM) that removes the computational overhead. The method utilizes momentum inspired by NAG instead of gradient ascent step, and therefore has almost the same speed as conventional optimizers, such as SGD and Adam. The experimental results show that the method yields comparable performance within the same computational budget.", "review_text": "The authors propose Momentum-SAM (MSAM) that removes the computational overhead. The method utilizes momentum inspired by NAG instead of gradient ascent step, and therefore has almost the same speed as conventional optimizers, such as SGD and Adam. The experimental results show that the method yields comparable performance within the same computational budget.", "strengths": "- The paper is easy to follow and provides clear motivation.\n- The proposed method does not introduce additional hyperparameter except the momentum coefficient.\n- The paper provides experimental results from various perspectives.", "weaknesses": "- From the results in Table 1, 2, 3, it can be seen that MSAM achieves relatively low performance that SAM except several cases. When the number of epochs of MSAM is doubled to make the same budget just like in A.9, can MSAM achieve higher performance than SAM? More generally, if MSAM yields higher final performance than that of SAM, the method could be more convincing. It would be nice to add more results to Figure A.8, for example, other methods like ESAM or other models/datasets.\n- The claims in chapter 4 are not convincing. Is there any theoretical reasons why positive momentum direction doesn't make the performance better? Also, it is unclear why $\\rho_0$ is close to $\\rho^{\\textrm{opt}}$.\n- Is the momentum applied to SAM as well in the paper? I understand that SAM has no momentum term.\n- \"efficient implementation\" of the method described in page 4 is quite straightforward, it seems to be unnecessary to describe it separately.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose Momentum-SAM (MSAM) that removes the computational overhead. The method utilizes momentum inspired by NAG instead of gradient ascent step, and therefore has almost the same speed as conventional optimizers, such as SGD and Adam. The experimental results show that the method yields comparable performance within the same computational budget.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is easy to follow and provides clear motivation.\n- The proposed method does not introduce additional hyperparameter except the momentum coefficient.\n- The paper provides experimental results from various perspectives.", "weaknesses": "- From the results in Table 1, 2, 3, it can be seen that MSAM achieves relatively low performance that SAM except several cases. When the number of epochs of MSAM is doubled to make the same budget just like in A.9, can MSAM achieve higher performance than SAM? More generally, if MSAM yields higher final performance than that of SAM, the method could be more convincing. It would be nice to add more results to Figure A.8, for example, other methods like ESAM or other models/datasets.\n- The claims in chapter 4 are not convincing. Is there any theoretical reasons why positive momentum direction doesn't make the performance better? Also, it is unclear why $\\rho_0$ is close to $\\rho^{\\textrm{opt}}$.\n- Is the momentum applied to SAM as well in the paper? I understand that SAM has no momentum term.\n- \"efficient implementation\" of the method described in page 4 is quite straightforward, it seems to be unnecessary to describe it separately.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698763062838}, {"id": "9go4ookWou", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5090/Reviewer_BYgZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors are striving to reduce the computational overhead incurred in the SAM algorithms, where they propose Momentum-SAM (MSAM). The proposed method utilizes the momentum direction to approximate sharpness computations. In such a manner, nearly no extra computation cost would be needed compared to the vanilla SAM. The authors empirically show the effectiveness of their method via experiments on CIFAR100 and ImageNet. The results show that the proposed method can reduce the computational overhead, but may harm the model performance.", "review_text": "The authors are striving to reduce the computational overhead incurred in the SAM algorithms, where they propose Momentum-SAM (MSAM). The proposed method utilizes the momentum direction to approximate sharpness computations. In such a manner, nearly no extra computation cost would be needed compared to the vanilla SAM. The authors empirically show the effectiveness of their method via experiments on CIFAR100 and ImageNet. The results show that the proposed method can reduce the computational overhead, but may harm the model performance.", "strengths": "**Strengths**\n1. The paper is clearly written and easy to follow. Also, the presentation of the paper is good.\n2. The paper present some interesting results, where it not only gives empirical results on the conventional benchmarks, also presents detailed study in regards to its properties.", "weaknesses": "**Weakness**\n\nThe paper is an empirical paper aiming to reducing the computational overhead incurred by the vanilla SAM algorithm. The proposed method is a quite heuristic method that substitutes the accent direction in SAM which is the positive gradient direction of the current weight model by the history gradient direction. In this way, the computational overhead could indeed be reduced. I would discuss the paper from several perspectives.\n\n1. I think the novelty of the paper is moderate. There are quite a number of papers targeting on the very topic. The core idea is somewhat quite similar to LookSAM. And the improvement of the proposed method is marginal. In fact, it is quite difficult to contribute further on this very topic, given there are a number of efficient methods proposed. And it may be more important to ensure the accuracy first. Given the results of the proposed method is good, I think the proposed method is acceptable. But the authors should present discussions in regards to the following paper, which contribute the topic from another perspective.\n\n    [1] Bahri, Dara, Hossein Mobahi, and Yi Tay. \"Sharpness-aware minimization improves language model generalization.\" arXiv preprint arXiv:2110.08529 (2021).\n\n    [2] Ni, Renkun, et al. \"K-SAM: Sharpness-Aware Minimization at the Speed of SGD.\" arXiv preprint arXiv:2210.12864 (2022).\n\n2. The significance of this paper is limited. To my understanding, the core of NAG is to accelerate convergence. It should be noted that here accelerating convergence does not mean to reduce the computation cost at each training batch. The authors have not provided any demonstration regarding the convergence of the proposed heuristic method. Given that the support of the proposed method is too weak, in my opinion, the minimum requirement of ICLR acceptance here is to discuss to what extent the proposed method can convergence from a theoretical perspective. See [3] and other related papers for reference.\n\n    [3] Andriushchenko, Maksym, and Nicolas Flammarion. \"Towards understanding sharpness-aware minimization.\" International Conference on Machine Learning. PMLR, 2022.\n\n3. The proposed method is not well supported. I could understand the authors' intention. However, the rationality of the proposed method is not clearly demonstrated. Regarding SAM, the ascent direction is solved explicitly and has a clear meaning. Minimization following the ascent direction can force the loss within the neighbourhood region to be low (i.e. flat). This is why SAM works. However, in my opinion, the proposed method cannot provide such a guarantee, given that the ascent direction could not characterize the maximum loss within the neighbourhood region. I notice that the authors give some simple empirical demonstration regarding the similarity between the SAM and MSAM. However, such a empirical demonstration could not give enough support and moreover, a similarity in direction could not give direct connection that the corresponding loss of the two models are close. It is not quite appropriate that one just give some other directions and claim that such directions could lead training to flat minimum.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors are striving to reduce the computational overhead incurred in the SAM algorithms, where they propose Momentum-SAM (MSAM). The proposed method utilizes the momentum direction to approximate sharpness computations. In such a manner, nearly no extra computation cost would be needed compared to the vanilla SAM. The authors empirically show the effectiveness of their method via experiments on CIFAR100 and ImageNet. The results show that the proposed method can reduce the computational overhead, but may harm the model performance.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "**Strengths**\n1. The paper is clearly written and easy to follow. Also, the presentation of the paper is good.\n2. The paper present some interesting results, where it not only gives empirical results on the conventional benchmarks, also presents detailed study in regards to its properties.", "weaknesses": "**Weakness**\n\nThe paper is an empirical paper aiming to reducing the computational overhead incurred by the vanilla SAM algorithm. The proposed method is a quite heuristic method that substitutes the accent direction in SAM which is the positive gradient direction of the current weight model by the history gradient direction. In this way, the computational overhead could indeed be reduced. I would discuss the paper from several perspectives.\n\n1. I think the novelty of the paper is moderate. There are quite a number of papers targeting on the very topic. The core idea is somewhat quite similar to LookSAM. And the improvement of the proposed method is marginal. In fact, it is quite difficult to contribute further on this very topic, given there are a number of efficient methods proposed. And it may be more important to ensure the accuracy first. Given the results of the proposed method is good, I think the proposed method is acceptable. But the authors should present discussions in regards to the following paper, which contribute the topic from another perspective.\n\n    [1] Bahri, Dara, Hossein Mobahi, and Yi Tay. \"Sharpness-aware minimization improves language model generalization.\" arXiv preprint arXiv:2110.08529 (2021).\n\n    [2] Ni, Renkun, et al. \"K-SAM: Sharpness-Aware Minimization at the Speed of SGD.\" arXiv preprint arXiv:2210.12864 (2022).\n\n2. The significance of this paper is limited. To my understanding, the core of NAG is to accelerate convergence. It should be noted that here accelerating convergence does not mean to reduce the computation cost at each training batch. The authors have not provided any demonstration regarding the convergence of the proposed heuristic method. Given that the support of the proposed method is too weak, in my opinion, the minimum requirement of ICLR acceptance here is to discuss to what extent the proposed method can convergence from a theoretical perspective. See [3] and other related papers for reference.\n\n    [3] Andriushchenko, Maksym, and Nicolas Flammarion. \"Towards understanding sharpness-aware minimization.\" International Conference on Machine Learning. PMLR, 2022.\n\n3. The proposed method is not well supported. I could understand the authors' intention. However, the rationality of the proposed method is not clearly demonstrated. Regarding SAM, the ascent direction is solved explicitly and has a clear meaning. Minimization following the ascent direction can force the loss within the neighbourhood region to be low (i.e. flat). This is why SAM works. However, in my opinion, the proposed method cannot provide such a guarantee, given that the ascent direction could not characterize the maximum loss within the neighbourhood region. I notice that the authors give some simple empirical demonstration regarding the similarity between the SAM and MSAM. However, such a empirical demonstration could not give enough support and moreover, a similarity in direction could not give direct connection that the corresponding loss of the two models are close. It is not quite appropriate that one just give some other directions and claim that such directions could lead training to flat minimum.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "I have not found any discussions about the limitations and potential negative societal impact. But in my opinion, this may not be a problem, since the work only focuses on the learning method in machine learning. Still, it is highly encouraged to add corresponding discussions.", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698673927197}, {"id": "HSk0ykmoIh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5090/Reviewer_R3HM"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the use of momentum in Sharpness Aware Minimization (SAM) is investigated and then a speed-up method is proposed. The main idea is to use the capability of momentum on lookahead to reduce one backward calculation in each iteration.", "review_text": "In this paper, the use of momentum in Sharpness Aware Minimization (SAM) is investigated and then a speed-up method is proposed. The main idea is to use the capability of momentum on lookahead to reduce one backward calculation in each iteration.", "strengths": "+ The computational time of SAM is almost twice of that of SGD, which is due to twice backward  calculation used in SAM. Thus, the idea of using momentum to estimate the gradient and to reduce one backward calculation is interesting. \n\n+ The paper is generally well organized and some explanation, e.g., Fig. 1, is clear.", "weaknesses": "- the mechanism of momentum has not been clearly presented.If we look two successive iterations together, we could find that using a momentum in the next iteration almost equals to use a larger learning rate in the previous iteration. Then I can understand that why the author claim that  \"the loss is expected to increase\" when \"a well-chosen learning rate\" is used, which has similar effect on attack used in SAM. However, if this is the reason, then there will be some other unclear points, see question 1 and 2.\n\n- the numerical experiments need improvements: (1) the general performance is not very good, comparing with the performance reported in recent papers. (2) as a method that focusing on speeding up SAM by cutting one backward  calculation, it is should be compared with random perturbation based methods, which have almost the same speed-up performance and mechanism.", "questions": "- Using momentum in the next iteration has similar performance as taking a larger learning rate in the previous iteration. But it is based on the fact that the \"a well-chosen learning rate\" is used. In practice, the learning rate is not necessarily optimal, thus I do not think increasing it can always increase the loss value, as attack procedure in SAM. In Section 4.1, the authors \"empirically show that the momentum descent step actually results in an ascent on the per-batch loss\", however, by showing the cos similarity (which is not directly related to value, due to the heavy non-linearity of DNN's training landscape). Can the author report the value change directly, also for different learning rate?\n\n- In a  paper, https://arxiv.org/pdf/2309.15639v2.pdf, recently accepted in NeurIPS2023, an opposite conclusion is given: one need to add momentum to enhance the attack in SAM. Following their idea, even when we do not want to perform backward calculation, we need to add a momentum to increase not to decrease the value. So can the author discuss the link to that paper? (generally I do not want to ask the authors to compare with recent papers, but they two are strongly linked, so I think it is better to discuss here)\n\n- Could the author provide a formulation similarly to NAG on which lookahead direction is calculated. \n\n- Could the author include comparison with other speed-up method for SAM and also please notice that the current reported performance for other method also the baseline SAM is not very good.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the use of momentum in Sharpness Aware Minimization (SAM) is investigated and then a speed-up method is proposed. The main idea is to use the capability of momentum on lookahead to reduce one backward calculation in each iteration.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "+ The computational time of SAM is almost twice of that of SGD, which is due to twice backward  calculation used in SAM. Thus, the idea of using momentum to estimate the gradient and to reduce one backward calculation is interesting. \n\n+ The paper is generally well organized and some explanation, e.g., Fig. 1, is clear.", "weaknesses": "- the mechanism of momentum has not been clearly presented.If we look two successive iterations together, we could find that using a momentum in the next iteration almost equals to use a larger learning rate in the previous iteration. Then I can understand that why the author claim that  \"the loss is expected to increase\" when \"a well-chosen learning rate\" is used, which has similar effect on attack used in SAM. However, if this is the reason, then there will be some other unclear points, see question 1 and 2.\n\n- the numerical experiments need improvements: (1) the general performance is not very good, comparing with the performance reported in recent papers. (2) as a method that focusing on speeding up SAM by cutting one backward  calculation, it is should be compared with random perturbation based methods, which have almost the same speed-up performance and mechanism.", "questions": "- Using momentum in the next iteration has similar performance as taking a larger learning rate in the previous iteration. But it is based on the fact that the \"a well-chosen learning rate\" is used. In practice, the learning rate is not necessarily optimal, thus I do not think increasing it can always increase the loss value, as attack procedure in SAM. In Section 4.1, the authors \"empirically show that the momentum descent step actually results in an ascent on the per-batch loss\", however, by showing the cos similarity (which is not directly related to value, due to the heavy non-linearity of DNN's training landscape). Can the author report the value change directly, also for different learning rate?\n\n- In a  paper, https://arxiv.org/pdf/2309.15639v2.pdf, recently accepted in NeurIPS2023, an opposite conclusion is given: one need to add momentum to enhance the attack in SAM. Following their idea, even when we do not want to perform backward calculation, we need to add a momentum to increase not to decrease the value. So can the author discuss the link to that paper? (generally I do not want to ask the authors to compare with recent papers, but they two are strongly linked, so I think it is better to discuss here)\n\n- Could the author provide a formulation similarly to NAG on which lookahead direction is calculated. \n\n- Could the author include comparison with other speed-up method for SAM and also please notice that the current reported performance for other method also the baseline SAM is not very good.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698631323459}], "openreview_url": "https://openreview.net/forum?id=sVl1KO5K76", "arxiv_id": "2401.12033", "paper_pdf": "papers/sVl1KO5K76.pdf", "paper_pdf_sha256": "760e595b58f33cad182cb99a9a98e1314b7592c41c258cd6ef7ad04b3ab52c33", "paper_pdf_bytes": 1820284, "paper_pdf_source": "openreview", "code_url": "https://github.com/MarlonBecker/MSAM", "code_repository": "MarlonBecker/MSAM", "code_commit": "780fa4f623fe0c99760fd7cfb1a94d14f00179ee", "code_archive": "repos/sVl1KO5K76.zip", "code_archive_sha256": "478747d058bad76ab7ad4e6ab9d2ad05aa9dddea508ac45d837c02ee309f5231", "code_archive_bytes": 57887, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 44, "github_languages": {"Python": 123241}, "github_archived": false, "github_pushed_at": "2024-01-23T15:00:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/momentum-sam-sharpness-aware-minimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VBB4fh45HF", "year": 2023, "status": "rejected", "title": "Learning Interpretable Dynamics from Images of a Freely Rotating 3D Rigid Body", "authors": ["Justice Mason", "Christine Allen-Blanchette", "Nicholas F Zolman", "Elizabeth Davison", "Naomi Leonard"], "authorids": ["~Justice_Mason1", "~Christine_Allen-Blanchette1", "~Nicholas_F_Zolman1", "~Elizabeth_Davison1", "~Naomi_Leonard1"], "authors_source": "OpenReview API", "abstract": "In many real-world settings, image observations of freely rotating 3D rigid bodies, such as satellites, may be available when low-dimensional measurements are not. However, the high-dimensionality of image data precludes the use of classical estimation techniques to learn the dynamics and a lack of interpretability reduces the usefulness of standard deep learning methods. In this work, we present a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. We achieve this using a multi-stage prediction pipeline that maps individual images to a latent representation homeomorphic to $\\mathbf{SO}(3)$, computes angular velocities from latent pairs, and predicts future latent states using the Hamiltonian equations of motion with a learned representation of the Hamiltonian. We demonstrate the efficacy of our approach on a new rotating rigid-body dataset with sequences of rotating cubes and rectangular prisms with uniform and non-uniform density.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "6H7J8UOBag", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5388/Reviewer_togy"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper propose a method to learn dynamics and make predictions from sequences of images of 3D rigid bodies. The method incorporates physics priors by using SO(3) as the latent space, and by translating the problem of learning the dynamics to that of learning a (reduced) Hamiltonian. Predictions are then made by and integrating the kinematic equations on SO(3), and Lie-Poisson equations on the Lie coalgebra of SO(3).\n\nThey perform experiments with a synthetic dataset consisting of sequences of rotating rigid bodies with different shapes and different, diagonal and non-diagonal, inertia tensors.", "review_text": "The paper propose a method to learn dynamics of freely rotating rigid bodies based on the simple idea that in the Hamiltonian framework, rigid bodies dynamics happens on $T^{*}SO(3)$. This makes learned dynamics interpretable and is a concrete step towards more trustworthy, less black-box machine learning models. Despite I asked for a few clarifications, I think the paper is overall clear.", "strengths": "The methods makes good use of well-known classical mechanics results about freely rotating rigid bodies giving good theoretical reasons for why the method should work. Interpretability of the learned dynamics is of utmost importance when using machine learning to model dynamical systems and the proposed algorithm gives very interpretable results constraining the dynamics on SO(3) and leaving all the unknown and irrelevant characteristics, such as color and shape of the modelled object, to the autoencoder.\n\nUnfortunately, as stated in the paper, there aren't many datasets to assess performances for the particular problem tackled in this paper. \nThe dataset could also represent a good benchmark for future works if made available, and therefore is a notable contribution.\nThe comparison with LSTMs and Neural ODEs is really meaningful as these methods are not as domain-specific as the proposed algorithms but certainly show how inductive biases like the one proposed in this paper are the way to go. Is there any \n\nApart from the dynamics, the SO(3) latent space seems to facilitate the reconstruction: perhaps a pre-training of just the encoder-decoder architecture, which is not unrealistic for real-world application, would boost performances. Or perhaps you already do that.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper propose a method to learn dynamics and make predictions from sequences of images of 3D rigid bodies. The method incorporates physics priors by using SO(3) as the latent space, and by translating the problem of learning the dynamics to that of learning a (reduced) Hamiltonian. Predictions are then made by and integrating the kinematic equations on SO(3), and Lie-Poisson equations on the Lie coalgebra of SO(3).\n\nThey perform experiments with a synthetic dataset consisting of sequences of rotating rigid bodies with different shapes and different, diagonal and non-diagonal, inertia tensors.", "strength_and_weaknesses": "The methods makes good use of well-known classical mechanics results about freely rotating rigid bodies giving good theoretical reasons for why the method should work. Interpretability of the learned dynamics is of utmost importance when using machine learning to model dynamical systems and the proposed algorithm gives very interpretable results constraining the dynamics on SO(3) and leaving all the unknown and irrelevant characteristics, such as color and shape of the modelled object, to the autoencoder.\n\nUnfortunately, as stated in the paper, there aren't many datasets to assess performances for the particular problem tackled in this paper. \nThe dataset could also represent a good benchmark for future works if made available, and therefore is a notable contribution.\nThe comparison with LSTMs and Neural ODEs is really meaningful as these methods are not as domain-specific as the proposed algorithms but certainly show how inductive biases like the one proposed in this paper are the way to go. Is there any \n\nApart from the dynamics, the SO(3) latent space seems to facilitate the reconstruction: perhaps a pre-training of just the encoder-decoder architecture, which is not unrealistic for real-world application, would boost performances. Or perhaps you already do that.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The background section is very concise but sufficiently clear for readers that are familiar with Hamiltonian formulation of rigid bodies dynamics.\nAs claimed by the authors, and to the best of my knowledge it is the first physics-informed deep learning algorithm for predicting image sequences of 3D rigid bodies.\n\nA few questions:\n\nHow figure 4 should be interpreted? I struggle to understand why the predicted $S^{2}\\times S^{2}$ parameterisation of SO(3) is so different from the ground truth. In this setting there is a lot of prior knowledge used in the algorithm, namely we know that the system is a freely rotating rigid body. Therefore, as you emphasise, \"learning dynamics\" boils down to making good predictions. In figure 3 I observe seemingly good predictions that I can't really conciliate with figure 4, where trajectories clearly differ. Section 6.2 does not provide sufficient explanation for why this is the case.\n\nBecause of the freedom os assigning an inertial frame Eq. (4) (and the corresponding Lie-Poisson) and (2) are independent. It is not very clear why you need the latter system of equations? Technically you could perform predictions using only the quaternion equations. Is it only because you want to enforce energy conservation? If so it would be interesting to know how the algorithm performs without this penalty in the loss.", "summary_of_the_review": "The paper propose a method to learn dynamics of freely rotating rigid bodies based on the simple idea that in the Hamiltonian framework, rigid bodies dynamics happens on $T^{*}SO(3)$. This makes learned dynamics interpretable and is a concrete step towards more trustworthy, less black-box machine learning models. Despite I asked for a few clarifications, I think the paper is overall clear.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667223735373}, {"id": "_hJfAXkNesh", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5388/Reviewer_RcCn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The manuscript proposes a method for estimating and predicting 3D rotation dynamics from image sequences of freely rotating 3D rigid bodies such as satellites. The proposed method maps individual images to latent representations isomorphic to SO(3) using auto encoder type neural networks, calculates angular velocities from the latent pairs, and then leams the moment of inertia tensor J in the Hamiltonian. Based on this prediction of future latent states, a predictive image is constructed using the auto encoder. After constructing a new rotating rigid body dataset consisting of a sequence of rotating cubes and rectangles, the authors show that the proposed method gives higher performance than LSTM and NeuralODE.", "review_text": "The importance of the research is high because it is a challenging task to realize a neural network method that learns a physical model capable of predicting video images. However, the performance of the method has not been compared to similar more general-purpose methods, which makes its effectiveness questionable. Therefore, I believe this concern needs to be resolved to accept this manuscript.", "strengths": "Strengths\n\nThis research is impressive because it is a challenging task to realize a neural network method that learns a physical model capable of predicting video images. The fact that they have succeeded in predicting and reconstructing video images to some extent by incorporating a physical inductive bias specific to rotational motion into the model is also highly novel.\n\n\n\nWeaknesses\n\nThe following two points are considered weaknesses\n\n\nThe advantage over previous research is unclear:\n\nHamiltonian Generative Networks (HGN) [1] has been shown to be able to learn up to 400 steps in the future for a two-body problem [2].\nHGN is also more general-purpose than the proposed method.\nThe authors need to show that the proposed method has better prediction performance for videos than HGN.\n\n[1] Peter Toth, Danilo J. Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, Irina Higgins, \"Hamiltonian Generative Networks,\" ICLR 2020, 2019.\n\n[2] Irina Higgins, Peter Wirnsberger, Andrew Jaegle, Aleksandar Botev, \"SyMetric: Measuring the Quality of Learnt Hamiltonian Dynamics Inferred from Vision,\" NeurIPS 2021, 2021.\n\n\n\nThe purpose of the method is unclear:.\n\nThe author writes that the application of the study is to estimate the orientation of satellites, but the training data in this study assigns a different color to each side of the object. I think it does not match the configuration of the satellites. If the differences between each side of a satellite can be determined, the orientation and motion of a satellite with a known geometry can be estimated analytically. Therefore, the proposed method is not considered necessary to achieve the objectives of the study.\n\nAlso, the work [Duong and Atanasov (2021)] cited by the authors in section \"2.3 LEARNING DYNAMICS FOR RIGID BODIES\" to achieve more general learning of rigid body motions by estimating inverse mass properties could be used for satellite orientation estimation, but no comparative experiments have been conducted. \n\nThese factors made me think that the relationship between the proposed method and the purpose of the study did not seem appropriate.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The manuscript proposes a method for estimating and predicting 3D rotation dynamics from image sequences of freely rotating 3D rigid bodies such as satellites. The proposed method maps individual images to latent representations isomorphic to SO(3) using auto encoder type neural networks, calculates angular velocities from the latent pairs, and then leams the moment of inertia tensor J in the Hamiltonian. Based on this prediction of future latent states, a predictive image is constructed using the auto encoder. After constructing a new rotating rigid body dataset consisting of a sequence of rotating cubes and rectangles, the authors show that the proposed method gives higher performance than LSTM and NeuralODE.", "strength_and_weaknesses": "Strengths\n\nThis research is impressive because it is a challenging task to realize a neural network method that learns a physical model capable of predicting video images. The fact that they have succeeded in predicting and reconstructing video images to some extent by incorporating a physical inductive bias specific to rotational motion into the model is also highly novel.\n\n\n\nWeaknesses\n\nThe following two points are considered weaknesses\n\n\nThe advantage over previous research is unclear:\n\nHamiltonian Generative Networks (HGN) [1] has been shown to be able to learn up to 400 steps in the future for a two-body problem [2].\nHGN is also more general-purpose than the proposed method.\nThe authors need to show that the proposed method has better prediction performance for videos than HGN.\n\n[1] Peter Toth, Danilo J. Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, Irina Higgins, \"Hamiltonian Generative Networks,\" ICLR 2020, 2019.\n\n[2] Irina Higgins, Peter Wirnsberger, Andrew Jaegle, Aleksandar Botev, \"SyMetric: Measuring the Quality of Learnt Hamiltonian Dynamics Inferred from Vision,\" NeurIPS 2021, 2021.\n\n\n\nThe purpose of the method is unclear:.\n\nThe author writes that the application of the study is to estimate the orientation of satellites, but the training data in this study assigns a different color to each side of the object. I think it does not match the configuration of the satellites. If the differences between each side of a satellite can be determined, the orientation and motion of a satellite with a known geometry can be estimated analytically. Therefore, the proposed method is not considered necessary to achieve the objectives of the study.\n\nAlso, the work [Duong and Atanasov (2021)] cited by the authors in section \"2.3 LEARNING DYNAMICS FOR RIGID BODIES\" to achieve more general learning of rigid body motions by estimating inverse mass properties could be used for satellite orientation estimation, but no comparative experiments have been conducted. \n\nThese factors made me think that the relationship between the proposed method and the purpose of the study did not seem appropriate.", "clarity,_quality,_novelty_and_reproducibility": "quality\n\nSome variables with hat are not defined.\n\n\nclarity\n\nSymbols and variables in chapter 3 were defined in chapter 4, which made the paper difficult for me to understand.\n\n\noriginality\n\nSince there are already several methods for learning Hamiltonian systems from video images[1,2], it is difficult to say that the proposed method is novel unless its performance is sufficiently high.", "summary_of_the_review": "The importance of the research is high because it is a challenging task to realize a neural network method that learns a physical model capable of predicting video images. However, the performance of the method has not been compared to similar more general-purpose methods, which makes its effectiveness questionable. Therefore, I believe this concern needs to be resolved to accept this manuscript.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667188127554}, {"id": "OizowNN51lB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5388/Reviewer_7a5H"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. The target is achieved by using a multi-stage prediction pipeline that maps individual images to a latent representation homeomorphic to SO(3), computes angular velocities from latent pairs, and predicts future latent states using the Hamiltonian equations of motion with a learned representation of the Hamiltonian.", "review_text": "The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. However, the current experimental evaluation and ablation studies are insufficient to justify the paper. The novelty/contribution of the proposed method over existing work has not been fully justified.\n", "strengths": "Strength: \n\n+ A physics-informed deep learning framework for predicting image sequences of 3D rigid-bodies by embedding the images as measurements in the configuration space SO(3) and propagating the Hamiltonian dynamics forward in time\n\n+ A new rotating rigid-body dataset with sequences of rotating cubes and rectangular prisms with uniform and non-uniform density.\n\nWeaknesses:\n\n- As the paper acknowledged, the paper only evaluated the proposed method in a rather idealized setting. To conclude the current experimental evaluation and ablation studies are insufficient to justify the paper.\n\n- The proposed approach currently relies on the ability to train the model for each system being examined.\n\n- The secod paragraph of the Summary section seems to be overclaiming its contributions without sufficient justification.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. The target is achieved by using a multi-stage prediction pipeline that maps individual images to a latent representation homeomorphic to SO(3), computes angular velocities from latent pairs, and predicts future latent states using the Hamiltonian equations of motion with a learned representation of the Hamiltonian.", "strength_and_weaknesses": "Strength: \n\n+ A physics-informed deep learning framework for predicting image sequences of 3D rigid-bodies by embedding the images as measurements in the configuration space SO(3) and propagating the Hamiltonian dynamics forward in time\n\n+ A new rotating rigid-body dataset with sequences of rotating cubes and rectangular prisms with uniform and non-uniform density.\n\nWeaknesses:\n\n- As the paper acknowledged, the paper only evaluated the proposed method in a rather idealized setting. To conclude the current experimental evaluation and ablation studies are insufficient to justify the paper.\n\n- The proposed approach currently relies on the ability to train the model for each system being examined.\n\n- The secod paragraph of the Summary section seems to be overclaiming its contributions without sufficient justification.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The introduction section should be improved with details of the proposed method. In Fig. 1, what are (a), (b) and (c) actually not clear.\n\nQuality & Novelty: As the paper acknowledged, the paper only evaluated the proposed method in a rather idealized setting. To conclude the current experimental evaluation and ablation studies are insufficient to justify the paper.\n\n\nReproducibility: Currently, more details are needed to achieve reproduction of the proposed method.", "summary_of_the_review": "The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. However, the current experimental evaluation and ablation studies are insufficient to justify the paper. The novelty/contribution of the proposed method over existing work has not been fully justified.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666796690304}, {"id": "nuUCV7SZTzH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5388/Reviewer_14VD"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper presents a method to estimate 3D rotations of an object from images using classical physics. The paper is focused on predicting  satellite rotations. It uses Hamiltonian representation of rotational motion. With this representation, motion trajectory is differentiable everywhere. This is not true for rotation parametrisation using exponential or quaternion maps as the inverses may be discontinuous. The paper provides a 3D dataset of uniform and non-uniform cube and prism to train the method.\nThe method uses a set of sequential images to estimate rotation and angular velocities. It encodes images to low-dimensional space representing rotations (learnt during training phase) which are used to estimate angular velocity and predicts next image using Hamiltonian dynamics.", "review_text": "The paper is definitely interesting. My only concern is the lack of evaluation on real data and the inability of method to deal with non-uniform data. \n\nAfter reading the rebuttal and discussion with other reviewers, I have changed my opinion (a bit unfavorably). Here are the main concerns:\n\n1. The experimentation is highly limited. The discussion on handling non-linear mass distribution is non-existent. The authors provide some explanation in the rebuttal but a lot questions remain. Given that the setup is difficult, perhaps the authors could explore non-linearity a bit more. For example, how does the method behave when the mass is non-uniform in a way that the centre of mass deviates strongly with respect to object centre. This discussion could have helped to understand the practical limitations of the method better. Plus, the authors must experiment with more data. Right now, the datasets are limited to a cube and a prism.\n\n2. There were some gaps in writing. The figure 2 in initial draft, which shows the trajectory to drift apart drastically, was later on removed. It is unclear to the readers how the registration works if the estimation of trajectory is so wrong. It is not clear whether the authors are performing a global alignment to fit into the projection. Perhaps the authors should have clarified.\n\nOverall, the paper has an important contribution over sota but the technical novelty is limited. Without a strong experimental evaluation, it was difficult to have an undoubted positive opinion.", "strengths": "Strengths:\n1. The use of hamiltonian dynamics is novel to the context of rotation estimation and trajectory prediction of satellites.\n2. A dataset has been proposed in order to learn 3D dynamics through images.\n3. The experiments show convincing results.\n\nWeakness:\n1. No evaluation on real data.\n2. In figure 4, it seems that the rotation estimation starts well and drifts away quickly. The rotation trajectory prediction is very off while the image prediction is quite decent. Do the images in fugure 3 (upto t=50) represent a small portion of the trajectory show on figure 4. More details are needed.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper presents a method to estimate 3D rotations of an object from images using classical physics. The paper is focused on predicting  satellite rotations. It uses Hamiltonian representation of rotational motion. With this representation, motion trajectory is differentiable everywhere. This is not true for rotation parametrisation using exponential or quaternion maps as the inverses may be discontinuous. The paper provides a 3D dataset of uniform and non-uniform cube and prism to train the method.\nThe method uses a set of sequential images to estimate rotation and angular velocities. It encodes images to low-dimensional space representing rotations (learnt during training phase) which are used to estimate angular velocity and predicts next image using Hamiltonian dynamics.", "strength_and_weaknesses": "Strengths:\n1. The use of hamiltonian dynamics is novel to the context of rotation estimation and trajectory prediction of satellites.\n2. A dataset has been proposed in order to learn 3D dynamics through images.\n3. The experiments show convincing results.\n\nWeakness:\n1. No evaluation on real data.\n2. In figure 4, it seems that the rotation estimation starts well and drifts away quickly. The rotation trajectory prediction is very off while the image prediction is quite decent. Do the images in fugure 3 (upto t=50) represent a small portion of the trajectory show on figure 4. More details are needed.", "clarity,_quality,_novelty_and_reproducibility": "I think that the writing is sufficiently comprehensible except in few sections. The main concepts are made quite clear. The limitations and failure cases have been discussed and the future works have been correctly identified. The authors will release the code upon acceptance to make the method reproducible. The main contribution is novel in the context of satellite trajectory estimation.", "summary_of_the_review": "The paper is definitely interesting. My only concern is the lack of evaluation on real data and the inability of method to deal with non-uniform data. \n\nAfter reading the rebuttal and discussion with other reviewers, I have changed my opinion (a bit unfavorably). Here are the main concerns:\n\n1. The experimentation is highly limited. The discussion on handling non-linear mass distribution is non-existent. The authors provide some explanation in the rebuttal but a lot questions remain. Given that the setup is difficult, perhaps the authors could explore non-linearity a bit more. For example, how does the method behave when the mass is non-uniform in a way that the centre of mass deviates strongly with respect to object centre. This discussion could have helped to understand the practical limitations of the method better. Plus, the authors must experiment with more data. Right now, the datasets are limited to a cube and a prism.\n\n2. There were some gaps in writing. The figure 2 in initial draft, which shows the trajectory to drift apart drastically, was later on removed. It is unclear to the readers how the registration works if the estimation of trajectory is so wrong. It is not clear whether the authors are performing a global alignment to fit into the projection. Perhaps the authors should have clarified.\n\nOverall, the paper has an important contribution over sota but the technical novelty is limited. Without a strong experimental evaluation, it was difficult to have an undoubted positive opinion.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666353799917}], "openreview_url": "https://openreview.net/forum?id=VBB4fh45HF", "arxiv_id": "2209.11355", "paper_pdf": "papers/VBB4fh45HF.pdf", "paper_pdf_sha256": "b8304673157153ef2d7945831ab4964519f371c2b6d6c55fa9eec6926b513386", "paper_pdf_bytes": 11568674, "paper_pdf_source": "openreview", "code_url": "https://github.com/jjmason687/LearningSO3fromImages", "code_repository": "jjmason687/LearningSO3fromImages", "code_commit": "357128e2eccd0dd1c1788152380e78b2ec4b4087", "code_archive": "repos/VBB4fh45HF.zip", "code_archive_sha256": "43f37f804d388398194758b13a4c514525c02a179841b67d0bcceefdac25d6de", "code_archive_bytes": 65302, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 71, "github_languages": {"Python": 245876}, "github_archived": false, "github_pushed_at": "2022-10-15T01:31:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-interpretable-dynamics-from-images"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TKrlyiqKWB", "year": 2022, "status": "rejected", "title": "Prototype Based Classification from Hierarchy to Fairness", "authors": ["Mycal Tucker", "Julie Shah"], "authorids": ["~Mycal_Tucker1", "~Julie_Shah1"], "authors_source": "OpenReview API", "abstract": "Artificial neural nets can represent and classify many types of high-dimensional data but are often tailored to particular applications -- e.g., for ``fair'' or ``hierarchical'' classification. Once an architecture has been selected, it is often difficult for humans to adjust models for a new task; for example, a hierarchical classifier cannot be easily transformed into a fair classifier that shields a protected field. Our contribution in this work is a new neural network architecture, the concept subspace network (CSN), which generalizes existing specialized classifiers to produce a unified model capable of learning a spectrum of multi-concept relationships. We demonstrate that CSNs reproduce state-of-the-art results in fair classification when enforcing concept independence, may be transformed into hierarchical classifiers, or may even reconcile fairness and hierarchy within a single classifier. The CSN is inspired by and matches the performance of existing prototype-based classifiers that promote interpretability.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "PSW6JOIhJJ7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper120/Reviewer_BhqD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper builds on prior work on prototypical classification networks (more specifically, the work of Li et al. 2018) and additionally tries to include criteria such as orthogonality to enable applications such as fair classification. An application to hierarchical networks is also described though the details are very hard to understand. Experiments show that the resulting models are able to achieve reasonable fairness accuracy tradeoffs.", "review_text": "The paper attempts interesting problems but lacks on 2 major fronts: (1) It is not clear what the improvement over the existing work is, and if it is significant enough to merit acceptance at ICLR (2) the writing needs a lot of work to bring out motivation for different choices.\n\nAbout the first point, the main contribution seems to lie in section 3.1, but most of the machinery here seems to be borrowed from the PCN work of Li et al. 2018. The additional contribution seems to be the concept subspace projection (if I understood correctly), whose motivation is not explained very well, and the addition of the alignment term in Eq. 1. The paper does not explain what is the additional value added by these terms over PCN.\n\nContinuing from the previous point, the paper is very hard to understand. In the second paragraph of Section 3.1, there is some departure from PCN where some combinations of $p1$ and other prototypes are taken. The process is defined in a very handwavy manner and not clear what it is the formal mathematical operation performed here. How is this different from PCN and why was this step needed? Talking about digits and concept subspaces, do we have as many concept subspaces as the number of classes? If not, how is this number picked?\n\nMoving on to the third paragraph of 3.1, the first few lines seem to be quite similar to PCN. However, at some point, a probability distribution is mentioned in connection with traditional softmax probabilities, but then yet another probability distribution is mentioned. It is not clear what the second distribution does. In the absence of formal equations, it is very difficult to understand what each component does. I would highly recommend describing each operation formally (in a sequential manner) and also adding a visualization like Figure 1 in the PCN paper to clearly convey the idea to the reader.\n\nFourth paragraph of 3.1 mentions two differences from PCN. Again, it is not clear what each of the differences achieves. Moreover, Figure 1 is neither described well in the main text nor in the caption, leaving the reader puzzled over what is happening in the figure. Instead of the regular autoencoder, a variational autoencoder is used, but again, the motivation is not clear. Other important details like the text above Equation 1, and the usage of KL diveregnce regularization term are skimmed over very quickly. The details of hierarchical classification setup in 3.4 are also glossed over quickly. The same happens in 4.2. For instance, what is meant by \"adopting the conditional probability training loss introduced by Hase et al\"?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper builds on prior work on prototypical classification networks (more specifically, the work of Li et al. 2018) and additionally tries to include criteria such as orthogonality to enable applications such as fair classification. An application to hierarchical networks is also described though the details are very hard to understand. Experiments show that the resulting models are able to achieve reasonable fairness accuracy tradeoffs.", "main_review": "The paper attempts interesting problems but lacks on 2 major fronts: (1) It is not clear what the improvement over the existing work is, and if it is significant enough to merit acceptance at ICLR (2) the writing needs a lot of work to bring out motivation for different choices.\n\nAbout the first point, the main contribution seems to lie in section 3.1, but most of the machinery here seems to be borrowed from the PCN work of Li et al. 2018. The additional contribution seems to be the concept subspace projection (if I understood correctly), whose motivation is not explained very well, and the addition of the alignment term in Eq. 1. The paper does not explain what is the additional value added by these terms over PCN.\n\nContinuing from the previous point, the paper is very hard to understand. In the second paragraph of Section 3.1, there is some departure from PCN where some combinations of $p1$ and other prototypes are taken. The process is defined in a very handwavy manner and not clear what it is the formal mathematical operation performed here. How is this different from PCN and why was this step needed? Talking about digits and concept subspaces, do we have as many concept subspaces as the number of classes? If not, how is this number picked?\n\nMoving on to the third paragraph of 3.1, the first few lines seem to be quite similar to PCN. However, at some point, a probability distribution is mentioned in connection with traditional softmax probabilities, but then yet another probability distribution is mentioned. It is not clear what the second distribution does. In the absence of formal equations, it is very difficult to understand what each component does. I would highly recommend describing each operation formally (in a sequential manner) and also adding a visualization like Figure 1 in the PCN paper to clearly convey the idea to the reader.\n\nFourth paragraph of 3.1 mentions two differences from PCN. Again, it is not clear what each of the differences achieves. Moreover, Figure 1 is neither described well in the main text nor in the caption, leaving the reader puzzled over what is happening in the figure. Instead of the regular autoencoder, a variational autoencoder is used, but again, the motivation is not clear. Other important details like the text above Equation 1, and the usage of KL diveregnce regularization term are skimmed over very quickly. The details of hierarchical classification setup in 3.4 are also glossed over quickly. The same happens in 4.2. For instance, what is meant by \"adopting the conditional probability training loss introduced by Hase et al\"?", "summary_of_the_review": "The contributions are not clear and the writing needs major work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636050275778}, {"id": "0TzXQw227Fv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper120/Reviewer_z2bK"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a novel model — called Concept Subspace Network (CSN) — for both hierarchical and fair classification. The idea behind the network is to use sets of prototypes to define concept subspaces in the latent space defined by the neural network itself. The relationships between the subspaces can be manipulated at training time to enforce concept relationships (i.e., two concept subspaces are orthogonal if the concepts they represent are independent, while they are parallel if the concepts they represent are hierarchically organised). ", "review_text": "In this paper, the ideas are quite novel and mostly well presented, and the problem handled is significant. \n\nI have though some questions and some minor comments that I hope will be addressed in the final version: \n1. The way in which concept subspaces are defined is not clear to me. In the paper, the authors write: “Given a set of k prototypes in $R^Z$ , the prototypes define a subspace generated by starting at the first prototype, $p_1$, and adding all linear combinations of vector differences from $p_1$ to all other $p_i$; $i \\in [2, k]$.” This is not clear to me, and it would be beneficial having a clearer example than the one given in the paper, in which it is not clear why we should obtain the plane $x-y$.\n2. Also, in order to get a subspace, do you need the assumption that $k < Z$? \n3. In equation (2) the term $PCN(\\cdot)$ is just defined as the loss introduced for PCN. Where is it defined?\n4. The random baseline seems to achieve very high performance in tables 1 and 2. \n5. At page 7 the authors mention a global ordering of the nodes, how was such ordering decided? \n\n\nMinor comments: \n1. $Z$ in the figure 1 instead of $z$\n2.  Add upward and downward arrows nearby the metric names to improve readability \n3. “Random” and “Rand” in Tables 1 and 2 ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a novel model — called Concept Subspace Network (CSN) — for both hierarchical and fair classification. The idea behind the network is to use sets of prototypes to define concept subspaces in the latent space defined by the neural network itself. The relationships between the subspaces can be manipulated at training time to enforce concept relationships (i.e., two concept subspaces are orthogonal if the concepts they represent are independent, while they are parallel if the concepts they represent are hierarchically organised). ", "main_review": "In this paper, the ideas are quite novel and mostly well presented, and the problem handled is significant. \n\nI have though some questions and some minor comments that I hope will be addressed in the final version: \n1. The way in which concept subspaces are defined is not clear to me. In the paper, the authors write: “Given a set of k prototypes in $R^Z$ , the prototypes define a subspace generated by starting at the first prototype, $p_1$, and adding all linear combinations of vector differences from $p_1$ to all other $p_i$; $i \\in [2, k]$.” This is not clear to me, and it would be beneficial having a clearer example than the one given in the paper, in which it is not clear why we should obtain the plane $x-y$.\n2. Also, in order to get a subspace, do you need the assumption that $k < Z$? \n3. In equation (2) the term $PCN(\\cdot)$ is just defined as the loss introduced for PCN. Where is it defined?\n4. The random baseline seems to achieve very high performance in tables 1 and 2. \n5. At page 7 the authors mention a global ordering of the nodes, how was such ordering decided? \n\n\nMinor comments: \n1. $Z$ in the figure 1 instead of $z$\n2.  Add upward and downward arrows nearby the metric names to improve readability \n3. “Random” and “Rand” in Tables 1 and 2 ", "summary_of_the_review": "The paper can be accepted if some clarifications are made", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635881571267}, {"id": "7FV9U6_l0YI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper120/Reviewer_aHhN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed a framework (that authors called the concept subspace network) using prototype-based representation controlling the alignment between two subspaces for the purpose of the classifier (fair or hierarch classification). ", "review_text": "Strength:\n- The paper proposed a new prototype-based approach considering the relationship between two concepts (classification tasks) for fairness and hierarchical classification. \n\nWeakness\n- The motivation of this paper is not clear to me. It would be helpful to understand the motivation by giving examples of major applications where both fairness and hierarchical classification should be considered. Also, is there any challenge when training a classifier using a regularization term regarding fairness in the existing hierarchy classifier training method?\n- It is not clear that why the two subspaces should be orthogonal and parallel for a fair and hierarchical classifier, respectively. Specifically, in section 3.4: \n1. For fair classification, what are the two subspaces? is it correct that the two subspaces are for label classification and sensitive attributes classification (e.g., male vs. female), respectively? Then, does the orthogonal relationship between the prototypes for estimating the sensitive attribute and label prediction guarantee the independence of the actual sensitive attribute and label prediction?\n2.  “In hierarchical classification, concepts are highly aligned and therefore parallel: the difference\nbetween a toucan and a Dalmatian is similar to the difference between a generic bird and\ndog.” --> In this example, why do parallel two concepts imply the difference between a toucan and a Dalmatian can be similar to the difference between a generic bird and dog? I think that the parallelism of two concepts is not related to the different relationships among prototypes. Then It is not clear that why two concepts should be parallel in the hierarchical classification. \n\nQuestions: \n- How does one train a hierarchical classifier with 3 or more concepts? In hierarchical classification, I think there are at least 3 concepts (e.g., dog vs bird, dog species classification, bird species classification)\n- In experiments, How was the parameter lambdas (in equation 2) chosen in the experiments?\n\nMinor feedback: \n- Adding a figure describing the proposed architecture will help readers understand the framework. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a framework (that authors called the concept subspace network) using prototype-based representation controlling the alignment between two subspaces for the purpose of the classifier (fair or hierarch classification). ", "main_review": "Strength:\n- The paper proposed a new prototype-based approach considering the relationship between two concepts (classification tasks) for fairness and hierarchical classification. \n\nWeakness\n- The motivation of this paper is not clear to me. It would be helpful to understand the motivation by giving examples of major applications where both fairness and hierarchical classification should be considered. Also, is there any challenge when training a classifier using a regularization term regarding fairness in the existing hierarchy classifier training method?\n- It is not clear that why the two subspaces should be orthogonal and parallel for a fair and hierarchical classifier, respectively. Specifically, in section 3.4: \n1. For fair classification, what are the two subspaces? is it correct that the two subspaces are for label classification and sensitive attributes classification (e.g., male vs. female), respectively? Then, does the orthogonal relationship between the prototypes for estimating the sensitive attribute and label prediction guarantee the independence of the actual sensitive attribute and label prediction?\n2.  “In hierarchical classification, concepts are highly aligned and therefore parallel: the difference\nbetween a toucan and a Dalmatian is similar to the difference between a generic bird and\ndog.” --> In this example, why do parallel two concepts imply the difference between a toucan and a Dalmatian can be similar to the difference between a generic bird and dog? I think that the parallelism of two concepts is not related to the different relationships among prototypes. Then It is not clear that why two concepts should be parallel in the hierarchical classification. \n\nQuestions: \n- How does one train a hierarchical classifier with 3 or more concepts? In hierarchical classification, I think there are at least 3 concepts (e.g., dog vs bird, dog species classification, bird species classification)\n- In experiments, How was the parameter lambdas (in equation 2) chosen in the experiments?\n\nMinor feedback: \n- Adding a figure describing the proposed architecture will help readers understand the framework. \n", "summary_of_the_review": "The motivation of the paper and the intuition of the proposed approach is not clear. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635863903097}, {"id": "jgTHq9RmVYb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper120/Reviewer_fPSK"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The present paper proposes a novel architecture for prototype-based classification to support class hierarchies and fairness. In particular, hierarchies are supported by training the model for multiple classification problems jointly, each in its own subspace of the feature space, spanned by the respective prototypes. For fairness, the paper proposes to make the subspace for the classification between subgroups orthogonal to all other subspaces, such that any change in subgroup membership does not influence any other classification. In a series of experiments, the paper evaluates hierarchical classification and fairness separately as well as jointly and demonstrates equal or superior results to a state-of-the-art approach from the literature.", "review_text": "The paper's main strengths are:\n* I found it particularly elegant to phrase both hierarchical classification and fairness in the same language, namely that of classification subspaces which are spanned by prototypes.\n* The paper connects to a wide range of concepts, namely hierarchical classification, interpretability, and fairness, such that it is of potential interest to a wide range of researchers.\n* The paper reports a wide range of experiments; so wide, indeed, that much experimental material had to be pushed to the appendix. I particularly appreciate the analysis of the hierarchies discovered by the prototype network and the comparison to the ground-truth hierarchy via edit distance.\n* The paper is clearly written. I, for one, had no problem following along and would feel well equipped to reproduce the reported results.\n\nThe paper's main weaknesses are:\n* The wide range of concepts discussed result in a certain lack of focus. Fairness, privacy, and causality are all mentioned but only discussed superficially. For fairness, this is particularly dangerous as readers may be mislead to believe that the proposed notion of orthogonality is sufficient for fairness. However, fairness has many meanings (as the paper acknowledges in the appendix) and only some of them are related to the proposed notion of orthogonality. Therefore, I would advise to revise references to fairness, privacy, and causality ind to mention explicitly that only a narrow notion of these terms is implemented by the proposed model.\n* The related work fails to mention the historic roots of the prototype concept. I understand that many recent works in prototype networks make the same mistake but I would still advise to not continue it. Prototype-based classification has - to my knowledge - been pioneered by Kohonen in the late 1980s/early 1990s with his work on Learning Vector Quantization (refer to the review by Nova and Estevez, 2014; doi: [10.1007/s00521-013-1535-3](https://doi.org/10.1007/s00521-013-1535-3) ) and has since been extended in many directions, such as metric learning (Schneider et al., 2009, doi:[10.1162/neco.2009.11-08-908](https://doi.org/10.1162/neco.2009.11-08-908)), or probabilistic classification (Seo and Obermayer, 2003, doi: [10.1162/089976603321891819](https://doi.org/10.1162/089976603321891819) ). The latter extension should be of particular interest because the classification scheme is very similar to the one proposed in this paper.\n* While the paper reports many different experiments, any single one seems relatively small with few data sets and (for hierarchical classification) few baselines. Further, I could not find information on the hyperparameter selection (e.g. how many prototypes and how strong the regularization strengths lambda were).\n\nOverall, my recommendation is to accept this paper. While the paper could be more focused, make its own contribution and limitations more clearly, and experiments could be extended, I still believe that most flaws could be addressed with minor adjustments and that the core contribution of the paper is interesting enough to a wide range of scholars that publication is warranted.\n\nNonetheless, I would appreciate if the authors could help me to deepen my understand of the work by responding to two questions:\n\n* I am not fully convinced that the projection into the plane spanned by the prototypes of a classification problem has any effect on the classification itself. If I understand correctly, the paper uses a softmax on the squared distances to all prototypes for classification (which is entirely reasonable). Now, let $D^2$ be the squared distance between a point $z$ and a prototype $p$, let $d^2$ be the squared distance between the projected point $\\tilde z$ and the same prototype $p$, and let $h^2$ be the squared distance between $z$ and $\\tilde z$. Since the distances form a right-angle triangle, we obtain $d^2 = D^2 - h^2$. This holds for any prototype in the same classification problem. Accordingly, all projected distances within one classification problem are merely the original distance minus a constant offset. This constant offset gets removed by softmax, anyways. So I would assume that the softmax probabilities are the same - no matter whether a point is projected or not.\n* Why was the parity hierarchy used as ground truth for MNIST? Garnot et al. use a hierarchy based on visual similarity of the digits (e.g. 3 and 8). Wouldn't that be more natural?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The present paper proposes a novel architecture for prototype-based classification to support class hierarchies and fairness. In particular, hierarchies are supported by training the model for multiple classification problems jointly, each in its own subspace of the feature space, spanned by the respective prototypes. For fairness, the paper proposes to make the subspace for the classification between subgroups orthogonal to all other subspaces, such that any change in subgroup membership does not influence any other classification. In a series of experiments, the paper evaluates hierarchical classification and fairness separately as well as jointly and demonstrates equal or superior results to a state-of-the-art approach from the literature.", "main_review": "The paper's main strengths are:\n* I found it particularly elegant to phrase both hierarchical classification and fairness in the same language, namely that of classification subspaces which are spanned by prototypes.\n* The paper connects to a wide range of concepts, namely hierarchical classification, interpretability, and fairness, such that it is of potential interest to a wide range of researchers.\n* The paper reports a wide range of experiments; so wide, indeed, that much experimental material had to be pushed to the appendix. I particularly appreciate the analysis of the hierarchies discovered by the prototype network and the comparison to the ground-truth hierarchy via edit distance.\n* The paper is clearly written. I, for one, had no problem following along and would feel well equipped to reproduce the reported results.\n\nThe paper's main weaknesses are:\n* The wide range of concepts discussed result in a certain lack of focus. Fairness, privacy, and causality are all mentioned but only discussed superficially. For fairness, this is particularly dangerous as readers may be mislead to believe that the proposed notion of orthogonality is sufficient for fairness. However, fairness has many meanings (as the paper acknowledges in the appendix) and only some of them are related to the proposed notion of orthogonality. Therefore, I would advise to revise references to fairness, privacy, and causality ind to mention explicitly that only a narrow notion of these terms is implemented by the proposed model.\n* The related work fails to mention the historic roots of the prototype concept. I understand that many recent works in prototype networks make the same mistake but I would still advise to not continue it. Prototype-based classification has - to my knowledge - been pioneered by Kohonen in the late 1980s/early 1990s with his work on Learning Vector Quantization (refer to the review by Nova and Estevez, 2014; doi: [10.1007/s00521-013-1535-3](https://doi.org/10.1007/s00521-013-1535-3) ) and has since been extended in many directions, such as metric learning (Schneider et al., 2009, doi:[10.1162/neco.2009.11-08-908](https://doi.org/10.1162/neco.2009.11-08-908)), or probabilistic classification (Seo and Obermayer, 2003, doi: [10.1162/089976603321891819](https://doi.org/10.1162/089976603321891819) ). The latter extension should be of particular interest because the classification scheme is very similar to the one proposed in this paper.\n* While the paper reports many different experiments, any single one seems relatively small with few data sets and (for hierarchical classification) few baselines. Further, I could not find information on the hyperparameter selection (e.g. how many prototypes and how strong the regularization strengths lambda were).\n\nOverall, my recommendation is to accept this paper. While the paper could be more focused, make its own contribution and limitations more clearly, and experiments could be extended, I still believe that most flaws could be addressed with minor adjustments and that the core contribution of the paper is interesting enough to a wide range of scholars that publication is warranted.\n\nNonetheless, I would appreciate if the authors could help me to deepen my understand of the work by responding to two questions:\n\n* I am not fully convinced that the projection into the plane spanned by the prototypes of a classification problem has any effect on the classification itself. If I understand correctly, the paper uses a softmax on the squared distances to all prototypes for classification (which is entirely reasonable). Now, let $D^2$ be the squared distance between a point $z$ and a prototype $p$, let $d^2$ be the squared distance between the projected point $\\tilde z$ and the same prototype $p$, and let $h^2$ be the squared distance between $z$ and $\\tilde z$. Since the distances form a right-angle triangle, we obtain $d^2 = D^2 - h^2$. This holds for any prototype in the same classification problem. Accordingly, all projected distances within one classification problem are merely the original distance minus a constant offset. This constant offset gets removed by softmax, anyways. So I would assume that the softmax probabilities are the same - no matter whether a point is projected or not.\n* Why was the parity hierarchy used as ground truth for MNIST? Garnot et al. use a hierarchy based on visual similarity of the digits (e.g. 3 and 8). Wouldn't that be more natural?", "summary_of_the_review": "Overall, my recommendation is to accept this paper. While the paper could be more focused, make its own contribution and limitations more clearly, and experiments could be extended, I still believe that most flaws could be addressed with minor adjustments and that the core contribution of the paper is interesting enough to a wide range of scholars that publication is warranted.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634726755627}], "openreview_url": "https://openreview.net/forum?id=TKrlyiqKWB", "arxiv_id": "2205.13997", "paper_pdf": "papers/TKrlyiqKWB.pdf", "paper_pdf_sha256": "901ff0aee3df2d2f350f3fbb77e2f901a9c809681ee2915534ced49e1f0a7f52", "paper_pdf_bytes": 963722, "paper_pdf_source": "openreview", "code_url": "https://github.com/mycal-tucker/csn", "code_repository": "mycal-tucker/csn", "code_commit": "0628a7c76d0d1a1f07b96d02d050780f2bd2cf06", "code_archive": "repos/TKrlyiqKWB.zip", "code_archive_sha256": "972660a1f3e6b1e752907f294fad9d16d93878a0499dfc6231f0d531c5413bc2", "code_archive_bytes": 91987, "code_file_count": 38, "code_extensions": {".py": 38}, "github_disk_usage_kb": 78, "github_languages": {"Python": 250389}, "github_archived": false, "github_pushed_at": "2022-12-05T19:06:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/prototype-based-classification-from-hierarchy-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PdauS7wZBfC", "year": 2021, "status": "rejected", "title": "Predictive Coding Approximates Backprop along Arbitrary Computation Graphs", "authors": ["Beren Millidge", "Alexander Tschantz", "Christopher Buckley"], "authorids": ["~Beren_Millidge1", "~Alexander_Tschantz1", "~Christopher_Buckley1"], "authors_source": "OpenReview API", "abstract": "The backpropagation of error (backprop) is a powerful algorithm for training machine learning architectures through end-to-end differentiation. Recently it has been shown that backprop in multilayer-perceptrons (MLPs) can be approximated using predictive coding, a biologically-plausible process theory of cortical computation which relies solely on local and Hebbian updates. The power of backprop, however, lies not in its instantiation in MLPs, but rather in the concept of automatic differentiation which allows for the optimisation of any differentiable program expressed as a computation graph. Here, we demonstrate that predictive coding converges asymptotically (and in practice rapidly) to exact backprop gradients on arbitrary computation graphs using only local learning rules. We apply this result to develop a straightforward strategy to translate core machine learning architectures into their predictive coding equivalents. We construct predictive coding CNNs, RNNs, and the more complex LSTMs, which include a non-layer-like branching internal graph structure and multiplicative interactions. Our models perform equivalently to backprop on challenging machine learning benchmarks, while utilising only local and (mostly) Hebbian plasticity. Our method raises the potential that standard machine learning algorithms could in principle be directly implemented in neural circuitry, and may also contribute to the development of completely distributed neuromorphic architectures.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "b9oWwaQTTlJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1764/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper extends prior work on equivalence between predictive coding and backprop in layered neural networks to arbitrary computation graphs. This is empirically tested first on a simple nonlinear scalar function, and then on a few commonly used architectures (CNNs, RNNs, LSTMs), confirming the theoretical results. The importance of this advance is highlighted by noting that the demonstrated equivalence shows how in principle modern architectures could be implemented in biological neural systems, and that the highly parallel nature of predictive coding could lead to efficient implementations in neuromorphic hardware.\n\nThe paper is very well written, easy to read, and includes a nice introduction section with a fairly comprehensive overview of backprop, and the problems related to its potential implementations in biological systems. I also appreciated the \"tension in a chain\" metaphor illustrating the dynamics of backprop and predictive coding. That the exact backprop gradients are computable in a fully local system with Hebbian plasticity for an arbitrary graph is an interesting and promising result.\n\nThroughout the text, predictive coding is quoted as biologically plausible. This isn't strictly true as noted already in (Whittington & Bogacz, 2017), as e.g. dedicated error nodes are not known to exist for every cell in the brain. I'd suggest calling this \"potentially biologically plausible\", and including a short discussion on how these plausibility concerns could be addressed.\n\nAll in all, the results are interesting, open up interesting directions for future work, and I recommend the acceptance of the paper.\n\nAdditional questions/suggestions:\n- Is the fixed-prediction assumption a limitation to biological plausibility?\n- Fig. 2 shows the model converging at high inference learning rates for the case of the scalar function. Is the 0.1 rate used for CNNs the max that was stable, or could higher values be used to reduce the computational overhead?\n- What convergence condition was used?\n- How does convergence speed depend on the size/diameter of the computational graph? This is similar to Fig. 9, but asking a slightly different question -- i.e. how many iterations are needed to reach convergence as a function of graph size.\n\nTypos:\nFig. 1 caption: \"backwawrds\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Prective coding shown to converge to backprop gradients for abritrary computational graphs", "review": "The paper extends prior work on equivalence between predictive coding and backprop in layered neural networks to arbitrary computation graphs. This is empirically tested first on a simple nonlinear scalar function, and then on a few commonly used architectures (CNNs, RNNs, LSTMs), confirming the theoretical results. The importance of this advance is highlighted by noting that the demonstrated equivalence shows how in principle modern architectures could be implemented in biological neural systems, and that the highly parallel nature of predictive coding could lead to efficient implementations in neuromorphic hardware.\n\nThe paper is very well written, easy to read, and includes a nice introduction section with a fairly comprehensive overview of backprop, and the problems related to its potential implementations in biological systems. I also appreciated the \"tension in a chain\" metaphor illustrating the dynamics of backprop and predictive coding. That the exact backprop gradients are computable in a fully local system with Hebbian plasticity for an arbitrary graph is an interesting and promising result.\n\nThroughout the text, predictive coding is quoted as biologically plausible. This isn't strictly true as noted already in (Whittington & Bogacz, 2017), as e.g. dedicated error nodes are not known to exist for every cell in the brain. I'd suggest calling this \"potentially biologically plausible\", and including a short discussion on how these plausibility concerns could be addressed.\n\nAll in all, the results are interesting, open up interesting directions for future work, and I recommend the acceptance of the paper.\n\nAdditional questions/suggestions:\n- Is the fixed-prediction assumption a limitation to biological plausibility?\n- Fig. 2 shows the model converging at high inference learning rates for the case of the scalar function. Is the 0.1 rate used for CNNs the max that was stable, or could higher values be used to reduce the computational overhead?\n- What convergence condition was used?\n- How does convergence speed depend on the size/diameter of the computational graph? This is similar to Fig. 9, but asking a slightly different question -- i.e. how many iterations are needed to reach convergence as a function of graph size.\n\nTypos:\nFig. 1 caption: \"backwawrds\"\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603887109408}, {"id": "lWegAXK_uVJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1764/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Update after author responses: \nWhile the author address some of my comments, I would have still liked to see a more detailed discussion of how the algorithm compares in terms of algorithmic scaling, which I think is relevant because it is a fundamental property of the algorithm, even if it is targeted towards understanding biology. So my score remains the same.\n\nSummary:\n\nThe authors extend recent work on MLPs to show that predictive coding converges asymptotically to exact backprop gradients on arbitrary computation graphs. They construct predictive coding networks for common architectures and show that it works well.\n\nOverall, I vote for an accept because I think the generalisation is quite useful and interesting for training deep networks with local learning rules. The authors demonstrate that this method works, but haven't demonstrated its computational advantages clearly enough. There are some issues of clarity that I have also outlined below.\n\nStrengths:\n+ The generalisation of the earlier MLP results to arbitrary computational graphs is quite powerful esp. since it can be applied to most deep learning architectures.\n+ The experimental evaluation includes all popular deep learning architecture, and it's impressive that this works on all of them. The experimental evaluation is also extensive.\n  \nWeaknesses:\n- The increase in computational cost (of 100x) is mentioned quite late and seems to be glossed over a bit.\n- Due to the potential for parallelisation in the predictive coding network, a comparison of wall-clock time for training on highly parallel setups might have been very interesting.\n- For RNNs and LSTMs, the equivalent predictive coding network is generated after unrolling the network, which means that the predictive coding network has no memory advantage over BPTT and a huge performance penalty. This is related to the previous point, where the utility of the predictive coding network is not demonstrated sufficiently.\n\nClarity:\n- Many of the figures are almost unreadable on paper. E.g. Fig. 3.\n- Algorithm 1 does't seem to be referenced anywhere.\n- In fig. 1 bottom, $\\delta$ missing in the denominators\n- Eqn. 3 is a bit sloppy, where derivative w.r.t $\\theta$ is suddenly equated to a derivative w.r.t $\\theta_i$.\n- If $\\epsilon_i = v_i - \\hat{v}_i$ then eqn. 7 is inconsistent with this, since $\\frac{d\\epsilon^*_i}{d\\theta}$ would be $-\\frac{d\\hat{v}_i}{d\\theta_i}$ (missing -ve sign). Unless I misunderstood something.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Generalisation of predictive coding approach to arbitrary computational graphs could be quite powerful", "review": "### Update after author responses: \nWhile the author address some of my comments, I would have still liked to see a more detailed discussion of how the algorithm compares in terms of algorithmic scaling, which I think is relevant because it is a fundamental property of the algorithm, even if it is targeted towards understanding biology. So my score remains the same.\n\nSummary:\n\nThe authors extend recent work on MLPs to show that predictive coding converges asymptotically to exact backprop gradients on arbitrary computation graphs. They construct predictive coding networks for common architectures and show that it works well.\n\nOverall, I vote for an accept because I think the generalisation is quite useful and interesting for training deep networks with local learning rules. The authors demonstrate that this method works, but haven't demonstrated its computational advantages clearly enough. There are some issues of clarity that I have also outlined below.\n\nStrengths:\n+ The generalisation of the earlier MLP results to arbitrary computational graphs is quite powerful esp. since it can be applied to most deep learning architectures.\n+ The experimental evaluation includes all popular deep learning architecture, and it's impressive that this works on all of them. The experimental evaluation is also extensive.\n  \nWeaknesses:\n- The increase in computational cost (of 100x) is mentioned quite late and seems to be glossed over a bit.\n- Due to the potential for parallelisation in the predictive coding network, a comparison of wall-clock time for training on highly parallel setups might have been very interesting.\n- For RNNs and LSTMs, the equivalent predictive coding network is generated after unrolling the network, which means that the predictive coding network has no memory advantage over BPTT and a huge performance penalty. This is related to the previous point, where the utility of the predictive coding network is not demonstrated sufficiently.\n\nClarity:\n- Many of the figures are almost unreadable on paper. E.g. Fig. 3.\n- Algorithm 1 does't seem to be referenced anywhere.\n- In fig. 1 bottom, $\\delta$ missing in the denominators\n- Eqn. 3 is a bit sloppy, where derivative w.r.t $\\theta$ is suddenly equated to a derivative w.r.t $\\theta_i$.\n- If $\\epsilon_i = v_i - \\hat{v}_i$ then eqn. 7 is inconsistent with this, since $\\frac{d\\epsilon^*_i}{d\\theta}$ would be $-\\frac{d\\hat{v}_i}{d\\theta_i}$ (missing -ve sign). Unless I misunderstood something.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603879511770}, {"id": "lOigDgNZdv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1764/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n \nAuthors propose that predictive coding gives similar convergence as backprop algorithms by extending the work of (Whittington &Bogacz 2017,  https://www.mrcbndu.ox.ac.uk/sites/default/files/pdf_files/Whittington%20Bogacz%202017_Neural%20Comput.pdf) to arbitrary graphs. This is an important topic both for the application of classical deep nets to neuromorphic hardware, but also for our understanding of computations in biological tissues.\n\n1. The work presented in this paper follows similar results by Amit (2019) or Lilicrap, and provides with numerical simulations comforting the theoretical predictions. \n2. Authors present an extension of the previous work to arbitrary graphs, and they apply their claims to LSTM and RNN models.\n3. It provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed framework. This makes the point of the paper more convincing and complementary to similar works.  \n\n##########################################################################\n\nConcern:\n \nA major issue of this paper is to not identify its originality. The extension to « arbitrary graphs » or to CNNs is straightforward in theory, While it is not explicitly stated in  (Whittington & Bogacz, 2017), an unwrapped RNN is by definition a feed-forward graph and is therefore a direct application of their work. \n\nConcerning novelty, the actual derivations (eg for the LSTM) are original and the simulations clearly support that original contributions. However this material is at the end of the paper or in appendices. Recentered on this original contributions and how this makes a suitable contribution to the community would make the paper acceptable to be accepted at ICLR.\n\n##########################################################################\nminor:\np.2 « backwawrds »\np7 « dataest » & p 14", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clear paper - marginal originality", "review": "##########################################################################\n\nSummary:\n \nAuthors propose that predictive coding gives similar convergence as backprop algorithms by extending the work of (Whittington &Bogacz 2017,  https://www.mrcbndu.ox.ac.uk/sites/default/files/pdf_files/Whittington%20Bogacz%202017_Neural%20Comput.pdf) to arbitrary graphs. This is an important topic both for the application of classical deep nets to neuromorphic hardware, but also for our understanding of computations in biological tissues.\n\n1. The work presented in this paper follows similar results by Amit (2019) or Lilicrap, and provides with numerical simulations comforting the theoretical predictions. \n2. Authors present an extension of the previous work to arbitrary graphs, and they apply their claims to LSTM and RNN models.\n3. It provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed framework. This makes the point of the paper more convincing and complementary to similar works.  \n\n##########################################################################\n\nConcern:\n \nA major issue of this paper is to not identify its originality. The extension to « arbitrary graphs » or to CNNs is straightforward in theory, While it is not explicitly stated in  (Whittington & Bogacz, 2017), an unwrapped RNN is by definition a feed-forward graph and is therefore a direct application of their work. \n\nConcerning novelty, the actual derivations (eg for the LSTM) are original and the simulations clearly support that original contributions. However this material is at the end of the paper or in appendices. Recentered on this original contributions and how this makes a suitable contribution to the community would make the paper acceptable to be accepted at ICLR.\n\n##########################################################################\nminor:\np.2 « backwawrds »\np7 « dataest » & p 14", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603836039260}, {"id": "9evPd1ch91R", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1764/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Summary of the paper\nIn their paper, the authors demonstrate that Predictive Coding (PC) is a local approximation of back-propagation and could then be interpreted with Hebbian learning rule (a neuro-plausible learning rule). This result has been first demonstrated by [1] with MLP network (on the MNIST dataset) and the presented paper extend this finding to CNNs (on CIFAR10, CIFAR100 and SVHN), RNN and LSTM. \n\n#### Pros\n* The authors provided experimental evidences on a wide variety of networks' type and databases.\n* The link between neuro-plausible learning rule and back propagation is interesting.\n* The paper is well situated in the literature.\n* The authors are providing the code for clean reproducibility\n\n#### Cons\n* The mathematical definition and notation of the paper are not rigorous enough. It makes the paper unclear and hard to follow.\n* Some crucial points would have deserved in-depth discussion and are just ignored (see below)\n* The paper is not well enough motivated: what’s the point of such a local approximation beside the neuro-plausibility (faster ? Consume less resources ? …)\n* The demonstration seems to include only one kind of loss function, which does not match the claim of the paper\n\n#### Recommendation\nGiven the limited impact and the lack of clarity of the paper, I would tend to reject the article.\n\n#### Detailed comments:\n* The gaussian parametrization used by the authors constrains the comparison between PC and backprop to networks with L2 loss function. One cannot claim to approximate arbitrary computational graph if one demonstrates the approximation on a specific loss function (which is known to poorly perform on classification problem). So could your framework be generalized to more effective loss function like cross entropy ? If yes, what would be the underlying probabilistic hypothesis ? This should be included in the paper, as it will strongly strengthen your claim.\n\n* The PC framework proposed by the authors propose a solution to the ’non locality’ of the back propagation (to be a bio-plausible mechanism). However the authors also raised the weight transport problem. On my understanding the proposed framework is still suffering from weight transport as the backward connection weights are the transpose of the feedforward one (due to the derivation of the forward operator). The paper would deserve an in-depth discussion concerning this point.\n\n* What is the computational advantage of local approximations ? Is it saving computational resources (computational time, memory…) ? A comparison of the algorithmic complexity between PC and back-propagation would be valuable to support your claim. In the discussion, the authors mention that their framework, being substantially more expensive than back-prop network, could be deeply parallelized across layer. The authors should provide experimental or theoretical evidences that such parallelization is enough to mitigate the higher number of inference steps (i.e. 100-200) needed by their PC framework.\n\n* The concept of ‘fixed-prediction assumption’ introduced by authors in the paper considers that each (v_i) are fixed to their feedforward value. Then what is the point of the Eq. 2, as you already know the value of the activity vector? On my mind this is here a crucial point, as this is dealing with the core principle of PC : an inner-loop (i.e. the expectation step) that find the most likely v_i, and an outer loop (i.e. the maximization step) that update the parameters. I have the intuition that this problem arises because the authors are tackling a discriminative problem (i.e. finding a mapping between inputs and labels) using a generative model (PC is a generative generative model as described by [2, 3]). Can you please clarify this point ?\n\n* What is the testing procedure of your network ? Is it a simple feedforward pass (which I suspect) or is it an Expectation-Maximization scheme. Your algorithm 1 shows only the training procedure (as you need the label information to perform the computation). If this is a simple feedforward pass, what would be the advantages of the inference process (more robustness ? Better prediction ?)\n\n## Typos and suggestions to improve the paper :\n* The authors state (3rd paragraph page 3) that they are considering a generative model. If it is the case, the formula p(v_i) = product(p(v_i | parent(v_i)) is inaccurate as the authors forgot the prior p(v_N).\n* Eq 1 : The derivation between the first and the second line of Eq.1 has to be demonstrated or referenced (at least in annex)…\n* The authors consider the posterior is a marginal probability (see eq1, and subsequent paragraph). In general the posterior is a conditional probability (this specific point makes your equation 1 hard to grasp because readers are not making the link with the classical ELBO, i.e. the negative free energy). In general, the probabilistic notations are not rigorous enough, and it makes the rest of the mathematical derivation complicated to follow. \n* The authors should reorganize the Annex to make sure it follows the reading order (Appendix D is cited first)\n* Caption of Figure 1 : backwawrds —> backwards\n* Page 3, 2 lines below eq 1 : as as —> as \n* Page 4, 4 lines below eq 3 : forwards —> forward\n* Figure 3, which is on my mind the most import one, is shown but not cited in the core text.\n\n[1] Whittington, J. C., & Bogacz, R. (2017). An approximation of the error backpropagation algorithm in a predictive coding network with local hebbian synaptic plasticity. Neural computation, 29(5), 1229-1262.\n\n[2] Rao, Rajesh PN, and Dana H. Ballard. \"Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.\" Nature neuroscience 2.1 (1999): 79-87.\n\n[3]Friston, Karl. \"A theory of cortical responses.\" Philosophical transactions of the Royal Society B: Biological sciences360.1456 (2005): 815-836.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is not clear enough", "review": "#### Summary of the paper\nIn their paper, the authors demonstrate that Predictive Coding (PC) is a local approximation of back-propagation and could then be interpreted with Hebbian learning rule (a neuro-plausible learning rule). This result has been first demonstrated by [1] with MLP network (on the MNIST dataset) and the presented paper extend this finding to CNNs (on CIFAR10, CIFAR100 and SVHN), RNN and LSTM. \n\n#### Pros\n* The authors provided experimental evidences on a wide variety of networks' type and databases.\n* The link between neuro-plausible learning rule and back propagation is interesting.\n* The paper is well situated in the literature.\n* The authors are providing the code for clean reproducibility\n\n#### Cons\n* The mathematical definition and notation of the paper are not rigorous enough. It makes the paper unclear and hard to follow.\n* Some crucial points would have deserved in-depth discussion and are just ignored (see below)\n* The paper is not well enough motivated: what’s the point of such a local approximation beside the neuro-plausibility (faster ? Consume less resources ? …)\n* The demonstration seems to include only one kind of loss function, which does not match the claim of the paper\n\n#### Recommendation\nGiven the limited impact and the lack of clarity of the paper, I would tend to reject the article.\n\n#### Detailed comments:\n* The gaussian parametrization used by the authors constrains the comparison between PC and backprop to networks with L2 loss function. One cannot claim to approximate arbitrary computational graph if one demonstrates the approximation on a specific loss function (which is known to poorly perform on classification problem). So could your framework be generalized to more effective loss function like cross entropy ? If yes, what would be the underlying probabilistic hypothesis ? This should be included in the paper, as it will strongly strengthen your claim.\n\n* The PC framework proposed by the authors propose a solution to the ’non locality’ of the back propagation (to be a bio-plausible mechanism). However the authors also raised the weight transport problem. On my understanding the proposed framework is still suffering from weight transport as the backward connection weights are the transpose of the feedforward one (due to the derivation of the forward operator). The paper would deserve an in-depth discussion concerning this point.\n\n* What is the computational advantage of local approximations ? Is it saving computational resources (computational time, memory…) ? A comparison of the algorithmic complexity between PC and back-propagation would be valuable to support your claim. In the discussion, the authors mention that their framework, being substantially more expensive than back-prop network, could be deeply parallelized across layer. The authors should provide experimental or theoretical evidences that such parallelization is enough to mitigate the higher number of inference steps (i.e. 100-200) needed by their PC framework.\n\n* The concept of ‘fixed-prediction assumption’ introduced by authors in the paper considers that each (v_i) are fixed to their feedforward value. Then what is the point of the Eq. 2, as you already know the value of the activity vector? On my mind this is here a crucial point, as this is dealing with the core principle of PC : an inner-loop (i.e. the expectation step) that find the most likely v_i, and an outer loop (i.e. the maximization step) that update the parameters. I have the intuition that this problem arises because the authors are tackling a discriminative problem (i.e. finding a mapping between inputs and labels) using a generative model (PC is a generative generative model as described by [2, 3]). Can you please clarify this point ?\n\n* What is the testing procedure of your network ? Is it a simple feedforward pass (which I suspect) or is it an Expectation-Maximization scheme. Your algorithm 1 shows only the training procedure (as you need the label information to perform the computation). If this is a simple feedforward pass, what would be the advantages of the inference process (more robustness ? Better prediction ?)\n\n## Typos and suggestions to improve the paper :\n* The authors state (3rd paragraph page 3) that they are considering a generative model. If it is the case, the formula p(v_i) = product(p(v_i | parent(v_i)) is inaccurate as the authors forgot the prior p(v_N).\n* Eq 1 : The derivation between the first and the second line of Eq.1 has to be demonstrated or referenced (at least in annex)…\n* The authors consider the posterior is a marginal probability (see eq1, and subsequent paragraph). In general the posterior is a conditional probability (this specific point makes your equation 1 hard to grasp because readers are not making the link with the classical ELBO, i.e. the negative free energy). In general, the probabilistic notations are not rigorous enough, and it makes the rest of the mathematical derivation complicated to follow. \n* The authors should reorganize the Annex to make sure it follows the reading order (Appendix D is cited first)\n* Caption of Figure 1 : backwawrds —> backwards\n* Page 3, 2 lines below eq 1 : as as —> as \n* Page 4, 4 lines below eq 3 : forwards —> forward\n* Figure 3, which is on my mind the most import one, is shown but not cited in the core text.\n\n[1] Whittington, J. C., & Bogacz, R. (2017). An approximation of the error backpropagation algorithm in a predictive coding network with local hebbian synaptic plasticity. Neural computation, 29(5), 1229-1262.\n\n[2] Rao, Rajesh PN, and Dana H. Ballard. \"Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.\" Nature neuroscience 2.1 (1999): 79-87.\n\n[3]Friston, Karl. \"A theory of cortical responses.\" Philosophical transactions of the Royal Society B: Biological sciences360.1456 (2005): 815-836.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603550212321}], "openreview_url": "https://openreview.net/forum?id=PdauS7wZBfC", "arxiv_id": "2006.04182", "paper_pdf": "papers/PdauS7wZBfC.pdf", "paper_pdf_sha256": "11083ec53b9e00b29830dc252523378b10b6ec0e4ce0bb743501352ee8726d4c", "paper_pdf_bytes": 2116262, "paper_pdf_source": "openreview", "code_url": "https://github.com/BerenMillidge/PredictiveCodingBackprop", "code_repository": "BerenMillidge/PredictiveCodingBackprop", "code_commit": "2154d7829a8f786e0a7a7b1bd8bb33c39d5af9ba", "code_archive": "repos/PdauS7wZBfC.zip", "code_archive_sha256": "09f15e47efddbe5b651f5077a36c3f86f2d6c98b0d98a280b4766cd5daf9c846", "code_archive_bytes": 34865, "code_file_count": 18, "code_extensions": {".py": 16, ".sh": 2}, "github_disk_usage_kb": 82, "github_languages": {"Python": 131154, "Shell": 3713}, "github_archived": false, "github_pushed_at": "2020-11-23T11:12:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/predictive-coding-approximates-backprop-along"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Skltqh4KvB", "year": 2020, "status": "rejected", "title": "Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?", "authors": ["Ella M. Gale", "Nicholas Martin", "Ryan Blything", "Anh Nguyen", "Jeffrey S. Bowers"], "authorids": ["ella.gale@bristol.ac.uk", "nm13850@bristol.ac.uk", "ryan.blything@bristol.ac.uk", "anhnguyen@auburn.edu", "j.bowers@bristol.ac.uk"], "authors_source": "OpenReview API", "abstract": "Various methods of measuring unit selectivity have been developed with the aim of better understanding how neural networks work.  But the different measures provide divergent estimates of selectivity, and this has led to different conclusions regarding the conditions in which selective object representations are learned and the functional relevance of these representations. In an attempt to better characterize object selectivity, we undertake a comparison of various selectivity measures on a large set of units in AlexNet, including localist selectivity, precision, class-conditional mean activity selectivity (CCMAS), network dissection, the human interpretation of activation maximization (AM) images, and standard signal-detection measures.  We find that the different measures provide different estimates of object selectivity, with precision and CCMAS measures providing misleadingly high estimates. Indeed, the most selective units had a poor hit-rate or a high false-alarm rate (or both) in object classification, making them poor object detectors.  We fail to find any units that are even remotely as selective as the 'grandmother cell' units reported in recurrent neural networks. In order to generalize these results, we compared selectivity measures on a few units in VGG-16 and GoogLeNet trained on the ImageNet or Places-365 datasets that have been described as 'object detectors'. Again, we find poor hit-rates and high false-alarm rates for object classification. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1x6uRTsYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper123/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work investigates the collection of methods that have been proposed to find units in neural networks that are selective for certain object classes.  Previous works have used different measures of selectivity (with sometimes contradictory results), and the authors investigate the degree to which these units qualify as “object detectors”.\n\nThis research area is important for understanding deep networks because claims have been made as to the relative importance (or lack thereof) of these individual units as identified by different measures vis-a-vis distributed representations -- the identification of such units would be interesting for understanding the predictions of classification networks.\n\nThe authors find that (1) different proposed measures of selectivity are not consistent and (2) units identified as selective cannot be considered object detectors due to the high false alarm / low hit rates, analyzing a large number of selectivity measures.  I would have liked to see experiments on more recent architectures (the focus of the paper is on a dated architecture (AlexNet)); there is analysis on units in GoogLeNet and VGG-16 but it would also be interesting to see results for more modern architectures (e.g. DenseNet and ResNet).\n\nOverall, I think that the authors have presented a strong meta-analysis and compelling argument for further study in rigorously identifying the presence (or lack thereof) of selective units in neural networks and the degree to which they may be considered \"object detectors.\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This work investigates the collection of methods that have been proposed to find units in neural networks that are selective for certain object classes.  Previous works have used different measures of selectivity (with sometimes contradictory results), and the authors investigate the degree to which these units qualify as “object detectors”.\n\nThis research area is important for understanding deep networks because claims have been made as to the relative importance (or lack thereof) of these individual units as identified by different measures vis-a-vis distributed representations -- the identification of such units would be interesting for understanding the predictions of classification networks.\n\nThe authors find that (1) different proposed measures of selectivity are not consistent and (2) units identified as selective cannot be considered object detectors due to the high false alarm / low hit rates, analyzing a large number of selectivity measures.  I would have liked to see experiments on more recent architectures (the focus of the paper is on a dated architecture (AlexNet)); there is analysis on units in GoogLeNet and VGG-16 but it would also be interesting to see results for more modern architectures (e.g. DenseNet and ResNet).\n\nOverall, I think that the authors have presented a strong meta-analysis and compelling argument for further study in rigorously identifying the presence (or lack thereof) of selective units in neural networks and the degree to which they may be considered \"object detectors.\""}, "tcdate": 1571704436874}, {"id": "B1e6QO3SKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper123/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper empirically studies the category selectivity of individual cells in hidden units of CNNs. It is a sort of \"meta-study\" and comparison of different metrics proposed to identify cells with a preference for a specific target category. The claimed finding is that there are no cells that are \"sufficiently\" selective to be called object detectors.\n\nThe paper is seemingly motivated by the authors' perceiving a contradiction: it is assumed that the power of neural networks is (among others) due to the distributed representation; whereas the presence of object detectors would, in the extreme case, mean that the representation is disentangled into a separate unit per category. It may be a matter of terminology, but this is where my disagreement with the authors start. I do not see a simplistic dichotomy, where one could or should determine which of the two interpretations is \"right\" or \"wrong\". In my view, which I believe is the mainstream interpretation, a distributed representation does not contradict the presence of specialised units. Some categories probably are easily identified by few distinctive features, so there will be more detector-like units; others are complex and hence more diffusely spread through the network; and of course there is no guarantee that the learned \"object detectors\" are tuned to exactly the target categories, after all it is the purpose of the network to gradually translate the data distribution to the label distribution - if the categories were directly apparent in the data, nearest-neighbour would be enough. So it is not only possible, but rather likely that the learned \"object detectors\" are to some degree driven by the statistics of the data, not the labels - e.g., there could be a highly selective \"bird\" unit which nevertheless has high false positive rate for any of the more specific bird species categories in the imageNet nomenclature. And vice versa, there could be a highly specific \"Ferrari\" detector that is so specialised that it has low recall for the \"sports car\" class (this case includes, among others, the case of viewpoint-specific detectors for certain categories). In the words of the paper, the \"selective units are sensitive to some feature that is frequently, but not exclusively associated with the class\" - I thought this is the standard majority view, not a surprising finding. In this context terminology matters: the study effectively tries to disprove that the network learns \"near-perfect single output-category detectors\", but who claims that it would do that?\n\nI agree with the authors that there is by now a zoo of selectivity metrics that are not always highly correlated. But is that a problem? We have a zoo of quality metrics for many machine learning problems - that is not necessarily a weakness, but simply reflects the obvious fact that a single number is not enough to characterise performance in a complex cognitive task. It is the job of the researcher/user to chose the metric that s most suitable for their specific question, and to correctly interpret its numerical value.\n\nRegarding the methodology, the paper did a lot of work to systematically crunch the numbers and analyse network units. It is a laudable effort that someone took on that job. A few technical decisions are unclear to me. Why analyse only some of the units? If one collects statistics over >2000 units of a fully connected layer, one might as well do the complete job and use all 4096 units. Similarly, why analyse only the correctly classified images? While it is clear that one must separately look at them, also the activations on incorrectly classified ones could provide valuable insights. E.g., do false positives of class X on average activate a certain \"class X detector\"? Why chose only the class with highest mean activation for CCMAS? That might be unrepresentative, e.g., a neuron might, for that particular class, always have high activation due to some very common background context, and still be not selective at all.\n\nRegarding the results, I find them much less clear-cut than the paper claims. For example, I find it quite remarkable that some unit has 8% recall at perfect precision. After all, only approximately 0.1% of the images are in the correct category, so a unit that flags 8% of them without making a mistake is a pretty good detector for (part of) the target class, cf Fig. 3. Also regarding Fig. 2 / maximum informedness, the statistics actually do not look bad. Of course false alarm proportion remains high - but the chance level here is 99.9%, so even a 99% false alarm rate means that your unit can, on its own, reject 90% of the true negatives. I find the proposed \"minimum condition\" for an object detector (>50% recall at >50% precision) unrealistic: the top-1 accuracy of AlexNet is, to my knowledge, <63%. Even the complete network probably never reaches 50% recall for most classes.\n\nEspecially the user study - which is again a commendable effort - in my view does not confirm the claims. According to that study, almost 60% 0f all fc8 units are \"object detectors\", with very high conherence between humans and selectivity metrics.\n\nOverall, while it is an interesting study, it remains unclear to me what I should learn from it. I don't see why different measures provide \"misleading conclusions\" that need to be rectified. Conclusions are the responsibility of the researcher interpreting the numbers, not of the formula to calculate some statistical performance metric. I am in a difficult situation here: the study is one of those things (like determining human performance on ImageNet, or re-coding some baseline where the original code is not available) where I find it valuable that someone did them in the community, but still I don't think they need a reviewed paper.  A note on the blog, or on arXiv, is enough.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper empirically studies the category selectivity of individual cells in hidden units of CNNs. It is a sort of \"meta-study\" and comparison of different metrics proposed to identify cells with a preference for a specific target category. The claimed finding is that there are no cells that are \"sufficiently\" selective to be called object detectors.\n\nThe paper is seemingly motivated by the authors' perceiving a contradiction: it is assumed that the power of neural networks is (among others) due to the distributed representation; whereas the presence of object detectors would, in the extreme case, mean that the representation is disentangled into a separate unit per category. It may be a matter of terminology, but this is where my disagreement with the authors start. I do not see a simplistic dichotomy, where one could or should determine which of the two interpretations is \"right\" or \"wrong\". In my view, which I believe is the mainstream interpretation, a distributed representation does not contradict the presence of specialised units. Some categories probably are easily identified by few distinctive features, so there will be more detector-like units; others are complex and hence more diffusely spread through the network; and of course there is no guarantee that the learned \"object detectors\" are tuned to exactly the target categories, after all it is the purpose of the network to gradually translate the data distribution to the label distribution - if the categories were directly apparent in the data, nearest-neighbour would be enough. So it is not only possible, but rather likely that the learned \"object detectors\" are to some degree driven by the statistics of the data, not the labels - e.g., there could be a highly selective \"bird\" unit which nevertheless has high false positive rate for any of the more specific bird species categories in the imageNet nomenclature. And vice versa, there could be a highly specific \"Ferrari\" detector that is so specialised that it has low recall for the \"sports car\" class (this case includes, among others, the case of viewpoint-specific detectors for certain categories). In the words of the paper, the \"selective units are sensitive to some feature that is frequently, but not exclusively associated with the class\" - I thought this is the standard majority view, not a surprising finding. In this context terminology matters: the study effectively tries to disprove that the network learns \"near-perfect single output-category detectors\", but who claims that it would do that?\n\nI agree with the authors that there is by now a zoo of selectivity metrics that are not always highly correlated. But is that a problem? We have a zoo of quality metrics for many machine learning problems - that is not necessarily a weakness, but simply reflects the obvious fact that a single number is not enough to characterise performance in a complex cognitive task. It is the job of the researcher/user to chose the metric that s most suitable for their specific question, and to correctly interpret its numerical value.\n\nRegarding the methodology, the paper did a lot of work to systematically crunch the numbers and analyse network units. It is a laudable effort that someone took on that job. A few technical decisions are unclear to me. Why analyse only some of the units? If one collects statistics over >2000 units of a fully connected layer, one might as well do the complete job and use all 4096 units. Similarly, why analyse only the correctly classified images? While it is clear that one must separately look at them, also the activations on incorrectly classified ones could provide valuable insights. E.g., do false positives of class X on average activate a certain \"class X detector\"? Why chose only the class with highest mean activation for CCMAS? That might be unrepresentative, e.g., a neuron might, for that particular class, always have high activation due to some very common background context, and still be not selective at all.\n\nRegarding the results, I find them much less clear-cut than the paper claims. For example, I find it quite remarkable that some unit has 8% recall at perfect precision. After all, only approximately 0.1% of the images are in the correct category, so a unit that flags 8% of them without making a mistake is a pretty good detector for (part of) the target class, cf Fig. 3. Also regarding Fig. 2 / maximum informedness, the statistics actually do not look bad. Of course false alarm proportion remains high - but the chance level here is 99.9%, so even a 99% false alarm rate means that your unit can, on its own, reject 90% of the true negatives. I find the proposed \"minimum condition\" for an object detector (>50% recall at >50% precision) unrealistic: the top-1 accuracy of AlexNet is, to my knowledge, <63%. Even the complete network probably never reaches 50% recall for most classes.\n\nEspecially the user study - which is again a commendable effort - in my view does not confirm the claims. According to that study, almost 60% 0f all fc8 units are \"object detectors\", with very high conherence between humans and selectivity metrics.\n\nOverall, while it is an interesting study, it remains unclear to me what I should learn from it. I don't see why different measures provide \"misleading conclusions\" that need to be rectified. Conclusions are the responsibility of the researcher interpreting the numbers, not of the formula to calculate some statistical performance metric. I am in a difficult situation here: the study is one of those things (like determining human performance on ImageNet, or re-coding some baseline where the original code is not available) where I find it valuable that someone did them in the community, but still I don't think they need a reviewed paper.  A note on the blog, or on arXiv, is enough.\n"}, "tcdate": 1571305508857}, {"id": "HJxqg9VQFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper123/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper makes an empirical claim that CNNs for object recognition do not contain hidden neuron which is highly selective to each class, mainly based on three aspects: (a) metrics related to the maximum informedness, (b) jitterplots of activation data, and (c) a user study assessing whether generated images maximizing a given unit is perceptible to the user. The paper point out these results are in contrast to the case of RNN, where it has been reported that many localist hidden units emerge. It is also noticed that the existing metrics for selectivity do not adequately discriminate highly selective units in CNN.  \n\nIn overall, the manuscript is well-written and easy-to-follow. I particularly appreciated a kindly presented overview on the literature and thoroughly conducted experiments including a complete user study. One of my key concerns, however, is that I am still not fully convinced whether the key finding in this paper - the lack of highly selective units in CNNs - is an indeed important problem for ICLR community: Personally, I feel the \"existence\" of selective units in RNN could be interesting, but the \"non-existence\" in the case of CNN is not that surprising for some readers, as it seems much likely (at least to me): The final layer of CNN would be surely selective across classes, but it may be not the case for the hidden layers - Nevertheless, some of the units may act selectively, not across classes but in some other concepts: e.g. stripes, orientations, etc. Therefore, I wonder if the paper could further provide a discussion on the importance of the key finding. \n\n- In Section 2 - Network and Dataset - \"... We selected 233 ...\" : Which criteria is actually used to choose the candidate units for the analysis?\n- Do the overall results mean that the \"maximum informedness\"-based metrics are superior to the others for assessing selectivity of a unit? Also, are those metrics original to this paper?\n- Currently, I feel the concept of \"selectivity\" is presented in somewhat subjectively: the definition could vary across the context. It would be nice if this could be more formalized to support the claims in the paper, e.g. the superiority of the proposed metrics.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper makes an empirical claim that CNNs for object recognition do not contain hidden neuron which is highly selective to each class, mainly based on three aspects: (a) metrics related to the maximum informedness, (b) jitterplots of activation data, and (c) a user study assessing whether generated images maximizing a given unit is perceptible to the user. The paper point out these results are in contrast to the case of RNN, where it has been reported that many localist hidden units emerge. It is also noticed that the existing metrics for selectivity do not adequately discriminate highly selective units in CNN.  \n\nIn overall, the manuscript is well-written and easy-to-follow. I particularly appreciated a kindly presented overview on the literature and thoroughly conducted experiments including a complete user study. One of my key concerns, however, is that I am still not fully convinced whether the key finding in this paper - the lack of highly selective units in CNNs - is an indeed important problem for ICLR community: Personally, I feel the \"existence\" of selective units in RNN could be interesting, but the \"non-existence\" in the case of CNN is not that surprising for some readers, as it seems much likely (at least to me): The final layer of CNN would be surely selective across classes, but it may be not the case for the hidden layers - Nevertheless, some of the units may act selectively, not across classes but in some other concepts: e.g. stripes, orientations, etc. Therefore, I wonder if the paper could further provide a discussion on the importance of the key finding. \n\n- In Section 2 - Network and Dataset - \"... We selected 233 ...\" : Which criteria is actually used to choose the candidate units for the analysis?\n- Do the overall results mean that the \"maximum informedness\"-based metrics are superior to the others for assessing selectivity of a unit? Also, are those metrics original to this paper?\n- Currently, I feel the concept of \"selectivity\" is presented in somewhat subjectively: the definition could vary across the context. It would be nice if this could be more formalized to support the claims in the paper, e.g. the superiority of the proposed metrics."}, "tcdate": 1571142129943}], "openreview_url": "https://openreview.net/forum?id=Skltqh4KvB", "arxiv_id": "2007.01062", "paper_pdf": "papers/Skltqh4KvB.pdf", "paper_pdf_sha256": "a67a7902d5d5014d891296dee8ea77c07dd0e98c152f59b755408ab77ef44ff3", "paper_pdf_bytes": 39476601, "paper_pdf_source": "openreview", "code_url": "https://github.com/ellagale/testing_object_detectors_in_deepCNNs", "code_repository": "ellagale/testing_object_detectors_in_deepCNNs", "code_commit": "fbc15e0b8f7e023c1e3e139feec185b36f68e6f6", "code_archive": "repos/Skltqh4KvB.zip", "code_archive_sha256": "67ce0cfe94cdf6de63c06bbfd04ceeee2492fb0cca729f5d6b39cb61d44327ef", "code_archive_bytes": 161205, "code_file_count": 36, "code_extensions": {".py": 36}, "github_disk_usage_kb": 125, "github_languages": {"Python": 671935}, "github_archived": false, "github_pushed_at": "2020-05-28T10:09:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/are-there-any-object-detectors-in-the-hidden-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1gR2sC9FX", "year": 2019, "status": "rejected", "title": "On the Spectral Bias of Neural Networks", "authors": ["Nasim Rahaman", "Aristide Baratin", "Devansh Arpit", "Felix Draxler", "Min Lin", "Fred Hamprecht", "Yoshua Bengio", "Aaron Courville"], "authorids": ["nasim.rahaman@iwr.uni-heidelberg.de", "aristidebaratin@hotmail.com", "devansharpit@gmail.com", "felix.draxler@iwr.uni-heidelberg.de", "mavenlin@gmail.com", "yoshua.umontreal@gmail.com", "aaron.courville@umontreal.ca"], "authors_source": "OpenReview API", "abstract": "Neural networks are known to be a class of highly expressive functions able to fit even random input-output mappings with 100% accuracy. In this work we present properties of neural networks that complement this aspect of expressivity. By using tools from Fourier analysis, we show that deep ReLU networks are biased towards low frequency functions, meaning that they cannot have local fluctuations without affecting their global behavior. Intuitively, this property is in line with the observation that over-parameterized networks find simple patterns that generalize across data samples. We also investigate how the shape of the data manifold affects expressivity by showing evidence that learning high frequencies gets easier with increasing manifold complexity, and present a theoretical understanding of this behavior. Finally, we study the robustness of the frequency components with respect to parameter perturbation, to develop the intuition that the parameters must be finely tuned to express high frequency functions.", "decision": null, "meta_review": "Reject", "num_reviews": 4, "reviews": [{"id": "ByeCM8C8a7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper757/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary.\n\nThis paper has theoretical and empirical contributions on topic of Fourier coefficients of neural networks.  First is upper bound on Fourier coefficients in terms of number of affine pieces and Lipschitz constant of network.  Second is collection of synthetic data and trained networks whereupon is argued that neural networks focus early effort upon low Fourier coefficients.\n\n\nBrief evaluation.\n\nPros:\n\n+ This paper attacks important and timely topic: identifying and analyzing implicit bias of neural networks paired with standard training methods.\n\nCons:\n\n- \"Implicit bias\" hypothesis has been put forth by many authors for many years, and this paper does not provide compelling argument that Fourier coefficients provide good characterization of this bias.\n\n- Regarding \"many authors for many years\", this paper fails to cite and utilize vast body of prior work, as detailed below.\n\n- Main theorem here is loose upper bound primarily derived from prior work, and no lower bounds are given.  Prior work does assess lower bounds.\n\n- Experiments are on synthetic data; prior work on implicit regularization does check real data.\n\n\nDetailed evaluation.\n\n* \"Implicit bias\" hypothesis appears in many places, for instance in work of Nati Srebro and colleagues (\"The Implicit Bias of Gradient Descent on Separable Data\" (and follow-ups), \"Exploring generalization in deep learning\" (and follow-ups), and others); it can also be found in variety of recent generalization papers, for instance again the work of Srebro et al, but also Bartlett et al, Arora et al.  E.g., Arora et al do detailed analysis of favorable biases in order to obtain refined generalization bound.  Consequently I expect this paper to argue to me, with strong theorems and experiments, that Fourier coefficients are a good way to assess implicit bias.\n\n* Theorem 1 is proved via bounds and tools on the Fourier spectra of indicators of polytopes due to Diaz et al, and linearity of the Fourier transform.  It is only upper bound (indeed one that makes no effort to deal with cancellations and thus become tight).  By contrast, the original proofs of depth separation for neural networks (e.g., Eldan and Shamir, or Telgarsky, both 2015), provide lower bounds and metric space separation.  Indeed, the work of Eldan&Shamir extensively uses Fourier analysis, and the proof develops a refined understanding of why it is hard for a ReLU network to approximate a Fourier transform of even simple functions: it has to approximate exponentially many tubes in Fourier space, which it can only do with exponentially many pieces.  While the present paper aims to cover some material not in Eldan&Shamir --- e.g., the bias with training --- this latter contribution is argued via synthetic data, and overall I feel the present work does not meet the (high) bar set by Eldan&Shamir.\n\n*  I will also point out that prior work of Barron, his \"superposition\" paper from 1993, is not cited. That paper presents upper bounds on approximation with neural networks which depends on the Fourier transform.  There is also follow-up by Arora et al with \"Barron functions\".\n\n* For experiments, I would really like to see experiment showing Fourier coefficients at various stages of training of standard network on standard data and standard data but with randomized labels (or different complexity in some other way).  These Fourier coefficients could also be compared to other \"implicit bias\" quantities; e.g., various norms and complexity measures.  In this way, it would be demonstrated that (a) spectral bias happens in practice, (b) spectral bias is a good way of measuring implicit bias.  Admittedly, this is computationally expensive experiment. \n\n* Regarding my claim that Theorem 1 is \"loose upper bound\": the slope of each piece is being upper bounded by Lipschitz constant, which will be far off in most regions.  Meanwhile, Lemma 1, \"exact characterization\", does not give any sense of how the slopes relate to weights of network.  Improving either issue would need to deal with \"cancellations\" I mention, and this is where it is hard to get upper and lower bounds to match.\n\nI feel this paper could be made much stronger by carefully using the results of all this prior work; these are not merely citation omissions, but indeed there is good understanding and progress in these papers.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "theoretical and empirical analysis of implicit bias in neural networks via Fourier coefficients.", "review": "Summary.\n\nThis paper has theoretical and empirical contributions on topic of Fourier coefficients of neural networks.  First is upper bound on Fourier coefficients in terms of number of affine pieces and Lipschitz constant of network.  Second is collection of synthetic data and trained networks whereupon is argued that neural networks focus early effort upon low Fourier coefficients.\n\n\nBrief evaluation.\n\nPros:\n\n+ This paper attacks important and timely topic: identifying and analyzing implicit bias of neural networks paired with standard training methods.\n\nCons:\n\n- \"Implicit bias\" hypothesis has been put forth by many authors for many years, and this paper does not provide compelling argument that Fourier coefficients provide good characterization of this bias.\n\n- Regarding \"many authors for many years\", this paper fails to cite and utilize vast body of prior work, as detailed below.\n\n- Main theorem here is loose upper bound primarily derived from prior work, and no lower bounds are given.  Prior work does assess lower bounds.\n\n- Experiments are on synthetic data; prior work on implicit regularization does check real data.\n\n\nDetailed evaluation.\n\n* \"Implicit bias\" hypothesis appears in many places, for instance in work of Nati Srebro and colleagues (\"The Implicit Bias of Gradient Descent on Separable Data\" (and follow-ups), \"Exploring generalization in deep learning\" (and follow-ups), and others); it can also be found in variety of recent generalization papers, for instance again the work of Srebro et al, but also Bartlett et al, Arora et al.  E.g., Arora et al do detailed analysis of favorable biases in order to obtain refined generalization bound.  Consequently I expect this paper to argue to me, with strong theorems and experiments, that Fourier coefficients are a good way to assess implicit bias.\n\n* Theorem 1 is proved via bounds and tools on the Fourier spectra of indicators of polytopes due to Diaz et al, and linearity of the Fourier transform.  It is only upper bound (indeed one that makes no effort to deal with cancellations and thus become tight).  By contrast, the original proofs of depth separation for neural networks (e.g., Eldan and Shamir, or Telgarsky, both 2015), provide lower bounds and metric space separation.  Indeed, the work of Eldan&Shamir extensively uses Fourier analysis, and the proof develops a refined understanding of why it is hard for a ReLU network to approximate a Fourier transform of even simple functions: it has to approximate exponentially many tubes in Fourier space, which it can only do with exponentially many pieces.  While the present paper aims to cover some material not in Eldan&Shamir --- e.g., the bias with training --- this latter contribution is argued via synthetic data, and overall I feel the present work does not meet the (high) bar set by Eldan&Shamir.\n\n*  I will also point out that prior work of Barron, his \"superposition\" paper from 1993, is not cited. That paper presents upper bounds on approximation with neural networks which depends on the Fourier transform.  There is also follow-up by Arora et al with \"Barron functions\".\n\n* For experiments, I would really like to see experiment showing Fourier coefficients at various stages of training of standard network on standard data and standard data but with randomized labels (or different complexity in some other way).  These Fourier coefficients could also be compared to other \"implicit bias\" quantities; e.g., various norms and complexity measures.  In this way, it would be demonstrated that (a) spectral bias happens in practice, (b) spectral bias is a good way of measuring implicit bias.  Admittedly, this is computationally expensive experiment. \n\n* Regarding my claim that Theorem 1 is \"loose upper bound\": the slope of each piece is being upper bounded by Lipschitz constant, which will be far off in most regions.  Meanwhile, Lemma 1, \"exact characterization\", does not give any sense of how the slopes relate to weights of network.  Improving either issue would need to deal with \"cancellations\" I mention, and this is where it is hard to get upper and lower bounds to match.\n\nI feel this paper could be made much stronger by carefully using the results of all this prior work; these are not merely citation omissions, but indeed there is good understanding and progress in these papers.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1542018582448}, {"id": "HygwdiS037", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper757/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Synopsis:\nThis paper analyzes deep Relu neural networks based on the Fourier decomposition of their input-output map. They show theoretically that the decomposition is biased towards low frequencies and give some support that low frequency components of a function are learned earlier under gradient descent training. \n\nPros:\n--Fourier decomposition is an important and (to the best of my knowledge) mostly original angle from which the authors analyze the input-output map governing neural networks. There is some neat mathematical analysis contained here based off of the piecewise-linearity of deep Relu nets and Fourier decomposition of polytopes in input space.\n\n--The setup in the toy experiments of Sec. 4 seems novel & thoughtful; the authors consider a lower-dimensional manifold embedded in a higher dimensional input space, and the Fourier decomposition of the composition of two functions is related to the decomposition of constituents.\n\nCons:\n--While this paper does a fairly good job establishing that NNs are spectrally biased towards low frequencies, I’m skeptical of its impact on our understanding of deep neural nets. Specifically, at a qualitative level it doesn’t seem very surprising: intuitively (as the authors write in Sec. 5), capturing higher frequencies in a function requires more fine tuning of the parameters.  At initialization, we don’t have such fine tuning (e.g. weights/biases drawn i.i.d Normal), and upon training it takes a certain amount of optimization time before we obtain greater “fine tuning.” At a quantitative level, these results would be more useful if (i) some insight could be gleaned from their dependence on the architectural choices of the network (in particular, depth) or (ii) some insight could be gained from how the spectral bias compares between deep NNs and other models (as is discussed briefly in the appendix -- for instance, kernel machines and K-NN classifiers). The primary dependence in the spectral decay (Theorem 1) seems to be that it (i) decays in a way which depends on the input dimensionality in most directions and (ii) it is highly anisotropic and decays more slowly in specific directions. The depth dependence seems to arise from the constants in the bound in Theorem 1 (see my comment below on the bound). \n\n--Relying on the growth of the weight norm to justify the network's bias towards learning lower frequencies earlier in training seems a bit tenuous to me. (I think the stronger evidence for learning lower frequencies comes from the experiments.) In particular, I'm not sure I would use the bound in Theorem 1 to conclude what would happen to actual Fourier components during training, since the bound may be far from being met. For instance, (1) the number of linear regions N_f changes during training -- what effect would this have? Also, (2) what if one were to use orthogonal weight matrices for training? Presumably the network would still train and generalize but the conclusions might be different (e.g. the idea that growth of weight norms is the cause of learning low frequency components earlier). \n\nMiscellaneous:\n--Would appreciate a greater discussion on the role of the cost function (MSE vs cross-entropy) in the analysis or experiments. Are the empirical conclusions mostly identical?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Intriguing topic and analysis, but its impact on understanding of neural nets seems limited", "review": "Synopsis:\nThis paper analyzes deep Relu neural networks based on the Fourier decomposition of their input-output map. They show theoretically that the decomposition is biased towards low frequencies and give some support that low frequency components of a function are learned earlier under gradient descent training. \n\nPros:\n--Fourier decomposition is an important and (to the best of my knowledge) mostly original angle from which the authors analyze the input-output map governing neural networks. There is some neat mathematical analysis contained here based off of the piecewise-linearity of deep Relu nets and Fourier decomposition of polytopes in input space.\n\n--The setup in the toy experiments of Sec. 4 seems novel & thoughtful; the authors consider a lower-dimensional manifold embedded in a higher dimensional input space, and the Fourier decomposition of the composition of two functions is related to the decomposition of constituents.\n\nCons:\n--While this paper does a fairly good job establishing that NNs are spectrally biased towards low frequencies, I’m skeptical of its impact on our understanding of deep neural nets. Specifically, at a qualitative level it doesn’t seem very surprising: intuitively (as the authors write in Sec. 5), capturing higher frequencies in a function requires more fine tuning of the parameters.  At initialization, we don’t have such fine tuning (e.g. weights/biases drawn i.i.d Normal), and upon training it takes a certain amount of optimization time before we obtain greater “fine tuning.” At a quantitative level, these results would be more useful if (i) some insight could be gleaned from their dependence on the architectural choices of the network (in particular, depth) or (ii) some insight could be gained from how the spectral bias compares between deep NNs and other models (as is discussed briefly in the appendix -- for instance, kernel machines and K-NN classifiers). The primary dependence in the spectral decay (Theorem 1) seems to be that it (i) decays in a way which depends on the input dimensionality in most directions and (ii) it is highly anisotropic and decays more slowly in specific directions. The depth dependence seems to arise from the constants in the bound in Theorem 1 (see my comment below on the bound). \n\n--Relying on the growth of the weight norm to justify the network's bias towards learning lower frequencies earlier in training seems a bit tenuous to me. (I think the stronger evidence for learning lower frequencies comes from the experiments.) In particular, I'm not sure I would use the bound in Theorem 1 to conclude what would happen to actual Fourier components during training, since the bound may be far from being met. For instance, (1) the number of linear regions N_f changes during training -- what effect would this have? Also, (2) what if one were to use orthogonal weight matrices for training? Presumably the network would still train and generalize but the conclusions might be different (e.g. the idea that growth of weight norms is the cause of learning low frequency components earlier). \n\nMiscellaneous:\n--Would appreciate a greater discussion on the role of the cost function (MSE vs cross-entropy) in the analysis or experiments. Are the empirical conclusions mostly identical?\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541458798819}, {"id": "SylaVMTh27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper757/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers the Fourier spectrum of functions represented by Deep ReLU networks, as well as the relationship to the training procedure by which the network weights can be learned. \n\nIt is well-known (and somewhat obvious) that deep neural networks with rectifier activations represent piecewise linear continuous function. Thus, the function can be written as a sum of the products of  indicators of various polytopes (which define the partition of R^d) and the linear function on that polytope. This allows the authors to compute the Fourier transform (cf. Thm. 1) and the magnitude of f(k) decays as k^{-i} where the i can depend on the polytope in some intricate fashion. Despite the remarks at the end of Thm 1, I found the result hard to interpret and relate to the rest of the paper. The appearance of N_f in the numerator (which can be exponentially large in the depth) may well make these bounds meaningless for any networks that are relevant in practice.\n\nThe main paper only has experiments on some synthetic data. \n\nSec 3: Does the MSE actually go to 0 in these experiments? Or are you observing that GD fits lower frequencies, because it has a hard time fitting things that oscillate frequently?\n\nSec 4: I would have liked to see a clearer explanation for example of why increasing L is better for regression, but not for classification. As it stands I can't read much from these experiments. \n\nOverall, I feel that there might be some interesting ideas in this paper, but the way it's currently written, I found it very hard to get a good \"picture\" of what the authors want to convey.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting ideas; message unclear", "review": "The paper considers the Fourier spectrum of functions represented by Deep ReLU networks, as well as the relationship to the training procedure by which the network weights can be learned. \n\nIt is well-known (and somewhat obvious) that deep neural networks with rectifier activations represent piecewise linear continuous function. Thus, the function can be written as a sum of the products of  indicators of various polytopes (which define the partition of R^d) and the linear function on that polytope. This allows the authors to compute the Fourier transform (cf. Thm. 1) and the magnitude of f(k) decays as k^{-i} where the i can depend on the polytope in some intricate fashion. Despite the remarks at the end of Thm 1, I found the result hard to interpret and relate to the rest of the paper. The appearance of N_f in the numerator (which can be exponentially large in the depth) may well make these bounds meaningless for any networks that are relevant in practice.\n\nThe main paper only has experiments on some synthetic data. \n\nSec 3: Does the MSE actually go to 0 in these experiments? Or are you observing that GD fits lower frequencies, because it has a hard time fitting things that oscillate frequently?\n\nSec 4: I would have liked to see a clearer explanation for example of why increasing L is better for regression, but not for classification. As it stands I can't read much from these experiments. \n\nOverall, I feel that there might be some interesting ideas in this paper, but the way it's currently written, I found it very hard to get a good \"picture\" of what the authors want to convey.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541358133016}, {"id": "BkgJFfPt37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper757/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Analysis of Spectral Bias of ReLU networks\n\nThe paper uses Fourier analysis to study ReLU network utilizing its continuous piecewise linear structure.\n\nMain finding is that these networks are biased towards learning low frequency which authors denote `spectral bias’.  This provides another theoretical perspective of neural networks preferring more smooth functions while being able to fit complicated function. Also shows that in terms of parameters networks representing lower frequency modes are more robust. \n\nPro: \n- Nice introduction to Fourier analysis providing non-trivial insights of ReLU networks.\n- Intuitive toy experiments to show spectral bias and its properties \n- Thorough theoretical analysis and empirical support\n\nCon: \n- The analysis is clearly for ReLU networks although the title may provide a false impression that it corresponds to general networks with other non-linearities. It is an interesting question whether the behaviour characterized by the authors are universal. \n- At least for me, Section 4 was not as clearly presented as other section. It takes more effort to parse what experiments were conducted and why such experiments are provided.\n- Although some experiments on real dataset are provided in the appendix, I personally could not read much intuition of theoretical findings to the networks used in practice. Does the spectral bias suggest better way of training or designing neural networks for example?\n\nComments/Questions:\n- In Figure 1, two experiments show different layerwise behaviour, i.e. equal amplitude experiment (a) shows spectral norm evolution for all the layers are almost identical whereas in increasing amplitude experiment (b) shows higher layer change spectral norm more than the lower layer. Do you understand why and does Fourier spectrum provide insights into layerwise behaviour?\n- Experiment 3 seems to perform binary classification using thresholding to the logits. But how do you find these results also hold for cross-entropy loss?\n“The results confirm the behaviour observed in Experiment 2, but in the case of classification tasks with categorical cross-entropy loss.”\n\n\nNit: p3 ReLu -> ReLU / p5 k \\in {50, 100, … 350, 400} (close bracket) / p5 in Experiment 2 and 3 descriptions the order of Figure appears flipped. Easier to read if the figure appears as the paper reads / p7 Equation 11 [0, 1]^m\n\n\n********* updated review *************\n\nBased on the issues raised from other reviewers and rebuttal from authors, I started to share some of the concerns on applicability of Thm 1 in obtaining information on low k Fourier coefficients. Although I empathize author's choice to mainly analyze synthetic data, I think it is critical to show the decays for moderately large k in realistic datasets. It will convince other reviewers of significance of main result of the paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Analysis of Spectral Bias of ReLU networks", "review": "Analysis of Spectral Bias of ReLU networks\n\nThe paper uses Fourier analysis to study ReLU network utilizing its continuous piecewise linear structure.\n\nMain finding is that these networks are biased towards learning low frequency which authors denote `spectral bias’.  This provides another theoretical perspective of neural networks preferring more smooth functions while being able to fit complicated function. Also shows that in terms of parameters networks representing lower frequency modes are more robust. \n\nPro: \n- Nice introduction to Fourier analysis providing non-trivial insights of ReLU networks.\n- Intuitive toy experiments to show spectral bias and its properties \n- Thorough theoretical analysis and empirical support\n\nCon: \n- The analysis is clearly for ReLU networks although the title may provide a false impression that it corresponds to general networks with other non-linearities. It is an interesting question whether the behaviour characterized by the authors are universal. \n- At least for me, Section 4 was not as clearly presented as other section. It takes more effort to parse what experiments were conducted and why such experiments are provided.\n- Although some experiments on real dataset are provided in the appendix, I personally could not read much intuition of theoretical findings to the networks used in practice. Does the spectral bias suggest better way of training or designing neural networks for example?\n\nComments/Questions:\n- In Figure 1, two experiments show different layerwise behaviour, i.e. equal amplitude experiment (a) shows spectral norm evolution for all the layers are almost identical whereas in increasing amplitude experiment (b) shows higher layer change spectral norm more than the lower layer. Do you understand why and does Fourier spectrum provide insights into layerwise behaviour?\n- Experiment 3 seems to perform binary classification using thresholding to the logits. But how do you find these results also hold for cross-entropy loss?\n“The results confirm the behaviour observed in Experiment 2, but in the case of classification tasks with categorical cross-entropy loss.”\n\n\nNit: p3 ReLu -> ReLU / p5 k \\in {50, 100, … 350, 400} (close bracket) / p5 in Experiment 2 and 3 descriptions the order of Figure appears flipped. Easier to read if the figure appears as the paper reads / p7 Equation 11 [0, 1]^m\n\n\n********* updated review *************\n\nBased on the issues raised from other reviewers and rebuttal from authors, I started to share some of the concerns on applicability of Thm 1 in obtaining information on low k Fourier coefficients. Although I empathize author's choice to mainly analyze synthetic data, I think it is critical to show the decays for moderately large k in realistic datasets. It will convince other reviewers of significance of main result of the paper.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541137014956}], "openreview_url": "https://openreview.net/forum?id=r1gR2sC9FX", "arxiv_id": "1806.08734", "paper_pdf": "papers/r1gR2sC9FX.pdf", "paper_pdf_sha256": "c9e0b9df78ebfe4c5e22e13a3fe6cd80635ff080ff569a0276ad24c94c3b2060", "paper_pdf_bytes": 2656003, "paper_pdf_source": "openreview", "code_url": "https://github.com/nasimrahaman/SpectralBias", "code_repository": "nasimrahaman/SpectralBias", "code_commit": "edfe5150c648414f6f4e7404626f834b26a0286c", "code_archive": "repos/r1gR2sC9FX.zip", "code_archive_sha256": "e8cef0b146e2214ab50b074c3244f74cb3683d554d8ed1b7caa54d1281edefe9", "code_archive_bytes": 944902, "code_file_count": 5, "code_extensions": {".ipynb": 5}, "github_disk_usage_kb": 922, "github_languages": {"Jupyter Notebook": 1311346}, "github_archived": false, "github_pushed_at": "2019-05-06T14:06:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-the-spectral-bias-of-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "AeHZWzDjTk", "year": 2026, "status": "rejected", "title": "Any-Order GPT as Masked Diffusion Model: Decoupling Formulation and Architecture", "authors": ["Shuchen Xue", "Tianyu Xie", "Tianyang Hu", "Zijin Feng", "Jiacheng Sun", "Kenji Kawaguchi", "Zhenguo Li", "Zhi-Ming Ma"], "authorids": ["~Shuchen_Xue1", "~Tianyu_Xie1", "~Tianyang_Hu1", "~Zijin_Feng1", "~Jiacheng_Sun1", "~Kenji_Kawaguchi1", "~Zhenguo_Li1", "~Zhi-Ming_Ma1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) predominantly use autoregressive (AR) approaches, but masked diffusion models (MDMs) are emerging as viable alternatives. A key challenge in comparing AR and MDM paradigms is their typical architectural difference: AR models are often decoder-only, while MDMs have largely been encoder-only. This practice of changing both the modeling paradigm and architecture simultaneously makes direct comparisons unfair, as it's hard to distinguish whether observed differences stem from the paradigm itself or the architectural shift. This research evaluates MDMs within a decoder-only framework to: (1) equitably compare MDM (as Any-Order AR, or AO-AR) and standard AR paradigms. Our investigation suggests that the standard AO-AR objective, which averages over all token permutations, may benefit from refinement, as many permutations appear less informative compared to the language's inherent left-to-right structure. (2) Investigate architectural influences (decoder-only vs. encoder-only) within MDMs. We demonstrate that while encoder-only MDMs model a simpler conditional probability space, decoder-only MDMs can achieve dramatic generation speedups ($\\sim25\\times$) and comparable perplexity with temperature annealing despite modeling a vastly larger space, highlighting key trade-offs. This work thus decouples core paradigm differences from architectural influences, offering insights for future model design.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "cPFZnwg00b", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13483/Reviewer_9dpA"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "The paper decouples formulation (autoregressive vs. masked diffusion / any‑order AR) from architecture (decoder‑only vs. encoder‑only) by building AO‑GPT, a decoder‑only masked‑diffusion/any‑order model. This lets the authors compare paradigms fairly within the same causal Transformer backbone. They (i) show the training objectives for masked diffusion (LMDM) and any‑order AR (LAO‑AR) are equivalent, (ii) analyze architectural trade‑offs—encoder‑only MDM vs. decoder‑only AO‑AR—in density estimation and generation complexity, and (iii) introduce practical ingredients (explicit target‑position injection via adaptive LayerNorm; EMA; a parallel multi‑mask attention mask) that enable fast, order‑agnostic decoding with competitive perplexity and reported large speedups. Experiments at GPT‑2‑scale compare orders, mixtures with a small fraction of left‑to‑right (L2R) updates, and encoder‑only vs. decoder‑only implementations.", "review_text": "The paper decouples formulation (autoregressive vs. masked diffusion / any‑order AR) from architecture (decoder‑only vs. encoder‑only) by building AO‑GPT, a decoder‑only masked‑diffusion/any‑order model. This lets the authors compare paradigms fairly within the same causal Transformer backbone. They (i) show the training objectives for masked diffusion (LMDM) and any‑order AR (LAO‑AR) are equivalent, (ii) analyze architectural trade‑offs—encoder‑only MDM vs. decoder‑only AO‑AR—in density estimation and generation complexity, and (iii) introduce practical ingredients (explicit target‑position injection via adaptive LayerNorm; EMA; a parallel multi‑mask attention mask) that enable fast, order‑agnostic decoding with competitive perplexity and reported large speedups. Experiments at GPT‑2‑scale compare orders, mixtures with a small fraction of left‑to‑right (L2R) updates, and encoder‑only vs. decoder‑only implementations.", "strengths": "++ By keeping the backbone decoder‑only and varying only the order/formulation, the paper cleanly separates effects of AO‑AR/MDM vs. AR. Section 2.2 crisply contrasts training signal density, density‑estimation, and generation complexity (O(n) for decoder‑only with KV cache vs. ~O(T·n) for encoder‑only MDM), avoiding the usual apples‑to‑oranges comparisons.\n\n++ The work unifies masked‑diffusion learning and any‑order AR by showing LMDM ≡ LAO‑AR, anchoring later design choices and analyses. This is a useful reference for the community.\n\n++ Mixing ~10% L2R with any‑order improves convergence/perplexity, and the authors report ~25× generation speedups for decoder‑only MDMs with temperature annealing while keeping perplexity competitive, clarifying the trade‑offs vs. encoder‑only MDMs.", "weaknesses": "-- Results are primarily at GPT‑2 small/medium scale; claims about competitiveness would be more convincing at ≥1B parameters and on stronger reasoning/benchmarks beyond perplexity.\n\n-- The ~25× speedup is promising but depends on decoder‑only specifics; head‑to‑head latency/throughput vs. tuned AR (Flash‑/paged‑KV, speculative decoding) and vs. well‑optimized encoder‑only MDMs (varying T) under identical hardware and sequence lengths would strengthen the claim.\n\n-- It’s unclear whether AO‑GPT and AR baselines matched total tokens/updates; the parameter and runtime overhead of target‑position encoders (AdaLN 128‑d) and EMA are not quantified.", "questions": "1. Were training tokens, optimizer schedules, and wall‑clock compute matched across AR and AO‑GPT? Please include FLOPs and training time comparisons.\n\n2. How is the ~25× speedup measured (batch size, length, KV cache policy, decoding temperature/annealing)? Can you add wall‑clock charts vs. a strong AR baseline with KV caching and against an encoder‑only MDM across T∈{8,16,32}?\n\n3. Beyond 10% L2R, did you try curriculum or entropy‑based order schedules? How sensitive are results to the order distribution?\n\n4. What is the parameter/latency overhead of AdaLN target‑positional encodings per layer? Any impact on memory footprint vs. a vanilla GPT‑2 of the same size?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper decouples formulation (autoregressive vs. masked diffusion / any‑order AR) from architecture (decoder‑only vs. encoder‑only) by building AO‑GPT, a decoder‑only masked‑diffusion/any‑order model. This lets the authors compare paradigms fairly within the same causal Transformer backbone. They (i) show the training objectives for masked diffusion (LMDM) and any‑order AR (LAO‑AR) are equivalent, (ii) analyze architectural trade‑offs—encoder‑only MDM vs. decoder‑only AO‑AR—in density estimation and generation complexity, and (iii) introduce practical ingredients (explicit target‑position injection via adaptive LayerNorm; EMA; a parallel multi‑mask attention mask) that enable fast, order‑agnostic decoding with competitive perplexity and reported large speedups. Experiments at GPT‑2‑scale compare orders, mixtures with a small fraction of left‑to‑right (L2R) updates, and encoder‑only vs. decoder‑only implementations.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "++ By keeping the backbone decoder‑only and varying only the order/formulation, the paper cleanly separates effects of AO‑AR/MDM vs. AR. Section 2.2 crisply contrasts training signal density, density‑estimation, and generation complexity (O(n) for decoder‑only with KV cache vs. ~O(T·n) for encoder‑only MDM), avoiding the usual apples‑to‑oranges comparisons.\n\n++ The work unifies masked‑diffusion learning and any‑order AR by showing LMDM ≡ LAO‑AR, anchoring later design choices and analyses. This is a useful reference for the community.\n\n++ Mixing ~10% L2R with any‑order improves convergence/perplexity, and the authors report ~25× generation speedups for decoder‑only MDMs with temperature annealing while keeping perplexity competitive, clarifying the trade‑offs vs. encoder‑only MDMs.", "weaknesses": "-- Results are primarily at GPT‑2 small/medium scale; claims about competitiveness would be more convincing at ≥1B parameters and on stronger reasoning/benchmarks beyond perplexity.\n\n-- The ~25× speedup is promising but depends on decoder‑only specifics; head‑to‑head latency/throughput vs. tuned AR (Flash‑/paged‑KV, speculative decoding) and vs. well‑optimized encoder‑only MDMs (varying T) under identical hardware and sequence lengths would strengthen the claim.\n\n-- It’s unclear whether AO‑GPT and AR baselines matched total tokens/updates; the parameter and runtime overhead of target‑position encoders (AdaLN 128‑d) and EMA are not quantified.", "questions": "1. Were training tokens, optimizer schedules, and wall‑clock compute matched across AR and AO‑GPT? Please include FLOPs and training time comparisons.\n\n2. How is the ~25× speedup measured (batch size, length, KV cache policy, decoding temperature/annealing)? Can you add wall‑clock charts vs. a strong AR baseline with KV caching and against an encoder‑only MDM across T∈{8,16,32}?\n\n3. Beyond 10% L2R, did you try curriculum or entropy‑based order schedules? How sensitive are results to the order distribution?\n\n4. What is the parameter/latency overhead of AdaLN target‑positional encodings per layer? Any impact on memory footprint vs. a vanilla GPT‑2 of the same size?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1762225706394}, {"id": "sxlGlDarPZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13483/Reviewer_cfK5"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper asks a good question: when people say “diffusion-style LMs are slower / worse than AR GPTs,” are we blaming the generative formulation (any-order / masked diffusion) or the architecture (encoder-only vs decoder-only)? To isolate this, the authors implement an any-order / masked-diffusion objective inside a decoder-only GPT—they call it Any-Order GPT (AO-GPT)—and compare it directly with standard left-to-right AR GPT under the same backbone. They show three main things: (i) true any-order training converges noticeably slower than L2R AR on the same model, confirming an optimization gap; (ii) adding a small fraction of L2R examples, plus stronger target-position injection (adaLN) and EMA, largely fixes this gap and brings AO-GPT close to encoder-only diffusion baselines such as MDLM/SEDD; and (iii) because AO-GPT is still decoder-only, it can decode in roughly linear time with KV-cache and even run ~25× faster than encoder-only masked diffusion samplers at equal length. The paper also analyzes why decoder-only any-order is harder—because it must model order-sensitive conditionals whose count grows like (e\\cdot n!), whereas encoder-only MDMs model order-invariant conditionals of size ($n2^{n-1}$).", "review_text": "This paper asks a good question: when people say “diffusion-style LMs are slower / worse than AR GPTs,” are we blaming the generative formulation (any-order / masked diffusion) or the architecture (encoder-only vs decoder-only)? To isolate this, the authors implement an any-order / masked-diffusion objective inside a decoder-only GPT—they call it Any-Order GPT (AO-GPT)—and compare it directly with standard left-to-right AR GPT under the same backbone. They show three main things: (i) true any-order training converges noticeably slower than L2R AR on the same model, confirming an optimization gap; (ii) adding a small fraction of L2R examples, plus stronger target-position injection (adaLN) and EMA, largely fixes this gap and brings AO-GPT close to encoder-only diffusion baselines such as MDLM/SEDD; and (iii) because AO-GPT is still decoder-only, it can decode in roughly linear time with KV-cache and even run ~25× faster than encoder-only masked diffusion samplers at equal length. The paper also analyzes why decoder-only any-order is harder—because it must model order-sensitive conditionals whose count grows like (e\\cdot n!), whereas encoder-only MDMs model order-invariant conditionals of size ($n2^{n-1}$).", "strengths": "1. Well-posed problem statement. The paper identifies a real confounder in current comparisons: AR↔MDM and decoder-only↔encoder-only are almost always changed together, so we don’t know which factor is responsible for the gap. Making MDM/AO run on a GPT-style decoder is a clean way to decouple these effects. This is genuinely useful for the community.\n2. Concrete, nontrivial engineering recipe. The combination “any-order objective + per-layer target-position injection (adaLN) + very slow EMA + 10% L2R mix” is not a cosmetic tweak; it is exactly what makes AO-GPT trainable at GPT-2 scale, and it empirically outperforms the more naïve σ-GPT-style injection on the same backbone. That’s a real contribution over earlier “just add an output position” works like σ-GPTs.\n3. Runtime story is compelling. Showing that a decoder-only instantiation of a diffusion/any-order objective can be linear-time and ~25× faster than encoder-only MDLM/SEDD on long sequences hits one of the most common complaints about masked diffusion LMs (“nice idea, too slow”). This moves diffusion-style LMs closer to being a practical alternative.", "weaknesses": "1. “Fair comparison” is only partially fair. The central claim is “we decouple formulation and architecture,” but the decoder-only any-order model gets a custom training recipe (adaLN, EMA=0.9999, 10% L2R mixing) that is not re-applied and re-tuned for the encoder-only diffusion baselines it is compared against. Yet we know from MDLM and EDLM that diffusion LMs are very sensitive to the exact denoising/weighting schedule and to Rao-Blackwellization tricks. If you let the baselines also adopt “a bit of L2R bias” or decoder-style conditioning, some of the reported gap reductions may disappear. Right now the story can still be read as “you improved your variant of any-order with several extra knobs” rather than “any-order itself is fine once you remove the architecture confounder.”\n2. The order-ensemble fix weakens the main narrative. A key observation in the paper is that decoder-only any-order must model vastly more order-sensitive conditionals ($(\\approx e\\cdot n!)$) than encoder-only ($(n2^{n-1})$), so its perplexity looks worse. The proposed remedy is to average over several context permutations at inference to wash out order bias. But once you need extra test-time ensembles to reach encoder-only quality, the big advantage “we’re linear-time and GPT-like” is qualified: you either pay extra forward passes, or you accept worse PPL. And the paper doesn’t show that a small ensemble (say, 2–4 orders) is always enough across datasets and lengths; the slides and OpenReview discussion indicate benefits keep increasing with more orders. That’s an unresolved algorithmic debt.\n3. Optimization diagnosis is shallow. The paper documents the symptom—pure any-order converges slower than L2R on the same GPT-2-style model—but the explanation stops at “language has a natural L2R bias, so mixing 10% L2R helps.” This is plausible but incomplete. For example, the work doesn’t separate: (i) gradient-noise increase from sampling uniformly over huge permutation spaces; (ii) mismatch between causal masking and “predict an earlier position”; and (iii) the fact that GPT’s rotary/absolute positions are themselves L2R-biased. Without ablations that fix the positional scheme or that use permutation-invariant encodings, we can’t tell which of these is the true bottleneck. Right now the fix is “add L2R,” which is very much a band-aid.\n4. Evidence does not scale. All empirical results are at roughly GPT-2-small / medium–like scale (the arXiv and OpenReview versions both state this) and on standard LM corpora (WikiText, OWT, LAMBADA). It is precisely at larger scales that (a) AR models start to benefit most from KV-cache amortization, and (b) discrete diffusion models in other papers start to fall behind when the number of sampling steps is restricted. Without at least a 1–7B run or a partial scaling curve, the claim “decoder-only MDM is a viable alternative to AR GPT” is not yet substantiated. This is especially important because contemporaneous MDLM work does report strong results at larger scales with optimized objectives.", "questions": "1. How essential is the 10% L2R mix, really? If you remove only the L2R mixing but keep adaLN and EMA, does AO-GPT still close the gap to encoder-only diffusion on WikiText103/1BW, or does it fall back to σ-GPT-like behavior? In other words, is the paper actually demonstrating “any-order works on decoder-only” or is it demonstrating “a mostly any-order objective regularized by a small AR prior works”? This matters for the main thesis.\n2. What is the inference-time story when you need order ensemble? The paper highlights a ~25× decoding speedup over encoder-only MDMs per order thanks to KV-cache, but to match their PPL they average over multiple permutations. For long sequences (1–2k tokens), how many permutations are actually needed before the curve saturates, and does the total wall-clock stay better than strong MDLM samplers like those in Sahoo et al. 2024? A plot of “#permutations vs PPL vs time” would make the runtime claim rock-solid.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper asks a good question: when people say “diffusion-style LMs are slower / worse than AR GPTs,” are we blaming the generative formulation (any-order / masked diffusion) or the architecture (encoder-only vs decoder-only)? To isolate this, the authors implement an any-order / masked-diffusion objective inside a decoder-only GPT—they call it Any-Order GPT (AO-GPT)—and compare it directly with standard left-to-right AR GPT under the same backbone. They show three main things: (i) true any-order training converges noticeably slower than L2R AR on the same model, confirming an optimization gap; (ii) adding a small fraction of L2R examples, plus stronger target-position injection (adaLN) and EMA, largely fixes this gap and brings AO-GPT close to encoder-only diffusion baselines such as MDLM/SEDD; and (iii) because AO-GPT is still decoder-only, it can decode in roughly linear time with KV-cache and even run ~25× faster than encoder-only masked diffusion samplers at equal length. The paper also analyzes why decoder-only any-order is harder—because it must model order-sensitive conditionals whose count grows like (e\\cdot n!), whereas encoder-only MDMs model order-invariant conditionals of size ($n2^{n-1}$).", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. Well-posed problem statement. The paper identifies a real confounder in current comparisons: AR↔MDM and decoder-only↔encoder-only are almost always changed together, so we don’t know which factor is responsible for the gap. Making MDM/AO run on a GPT-style decoder is a clean way to decouple these effects. This is genuinely useful for the community.\n2. Concrete, nontrivial engineering recipe. The combination “any-order objective + per-layer target-position injection (adaLN) + very slow EMA + 10% L2R mix” is not a cosmetic tweak; it is exactly what makes AO-GPT trainable at GPT-2 scale, and it empirically outperforms the more naïve σ-GPT-style injection on the same backbone. That’s a real contribution over earlier “just add an output position” works like σ-GPTs.\n3. Runtime story is compelling. Showing that a decoder-only instantiation of a diffusion/any-order objective can be linear-time and ~25× faster than encoder-only MDLM/SEDD on long sequences hits one of the most common complaints about masked diffusion LMs (“nice idea, too slow”). This moves diffusion-style LMs closer to being a practical alternative.", "weaknesses": "1. “Fair comparison” is only partially fair. The central claim is “we decouple formulation and architecture,” but the decoder-only any-order model gets a custom training recipe (adaLN, EMA=0.9999, 10% L2R mixing) that is not re-applied and re-tuned for the encoder-only diffusion baselines it is compared against. Yet we know from MDLM and EDLM that diffusion LMs are very sensitive to the exact denoising/weighting schedule and to Rao-Blackwellization tricks. If you let the baselines also adopt “a bit of L2R bias” or decoder-style conditioning, some of the reported gap reductions may disappear. Right now the story can still be read as “you improved your variant of any-order with several extra knobs” rather than “any-order itself is fine once you remove the architecture confounder.”\n2. The order-ensemble fix weakens the main narrative. A key observation in the paper is that decoder-only any-order must model vastly more order-sensitive conditionals ($(\\approx e\\cdot n!)$) than encoder-only ($(n2^{n-1})$), so its perplexity looks worse. The proposed remedy is to average over several context permutations at inference to wash out order bias. But once you need extra test-time ensembles to reach encoder-only quality, the big advantage “we’re linear-time and GPT-like” is qualified: you either pay extra forward passes, or you accept worse PPL. And the paper doesn’t show that a small ensemble (say, 2–4 orders) is always enough across datasets and lengths; the slides and OpenReview discussion indicate benefits keep increasing with more orders. That’s an unresolved algorithmic debt.\n3. Optimization diagnosis is shallow. The paper documents the symptom—pure any-order converges slower than L2R on the same GPT-2-style model—but the explanation stops at “language has a natural L2R bias, so mixing 10% L2R helps.” This is plausible but incomplete. For example, the work doesn’t separate: (i) gradient-noise increase from sampling uniformly over huge permutation spaces; (ii) mismatch between causal masking and “predict an earlier position”; and (iii) the fact that GPT’s rotary/absolute positions are themselves L2R-biased. Without ablations that fix the positional scheme or that use permutation-invariant encodings, we can’t tell which of these is the true bottleneck. Right now the fix is “add L2R,” which is very much a band-aid.\n4. Evidence does not scale. All empirical results are at roughly GPT-2-small / medium–like scale (the arXiv and OpenReview versions both state this) and on standard LM corpora (WikiText, OWT, LAMBADA). It is precisely at larger scales that (a) AR models start to benefit most from KV-cache amortization, and (b) discrete diffusion models in other papers start to fall behind when the number of sampling steps is restricted. Without at least a 1–7B run or a partial scaling curve, the claim “decoder-only MDM is a viable alternative to AR GPT” is not yet substantiated. This is especially important because contemporaneous MDLM work does report strong results at larger scales with optimized objectives.", "questions": "1. How essential is the 10% L2R mix, really? If you remove only the L2R mixing but keep adaLN and EMA, does AO-GPT still close the gap to encoder-only diffusion on WikiText103/1BW, or does it fall back to σ-GPT-like behavior? In other words, is the paper actually demonstrating “any-order works on decoder-only” or is it demonstrating “a mostly any-order objective regularized by a small AR prior works”? This matters for the main thesis.\n2. What is the inference-time story when you need order ensemble? The paper highlights a ~25× decoding speedup over encoder-only MDMs per order thanks to KV-cache, but to match their PPL they average over multiple permutations. For long sequences (1–2k tokens), how many permutations are actually needed before the curve saturates, and does the total wall-clock stay better than strong MDLM samplers like those in Sahoo et al. 2024? A plot of “#permutations vs PPL vs time” would make the runtime claim rock-solid.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762046831313}, {"id": "NtDbrum7VE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13483/Reviewer_fm5B"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This work is the first systematic and equitable comparison of masked-diffusion language models (MDMs) and standard autoregressive (AR) LMs by decoupling training paradigm from architecture. The authors show that when both are instantiated with the same decoder-only backbone, the core distinction reduces to the distribution over token orders. They therefore implement MDM as Any-Order AR within a decoder-only GPT (“AO-GPT”) and study: (1) how AR vs. AO-AR differ in modeling capacity and empirical behavior under an identical architecture; and (2) for MDMs, how encoder-only and decoder-only architectures compare theoretically and empirically. Experiments find AO-GPT converges notably more slowly early in training than left-to-right (L2R) GPT; a fixed block-wise random order interpolates between L2R and fully random in convergence; and mixing a small fraction (~10%) of L2R samples improves AO performance. Decoder-only AO-AR underperforms encoder-only variants on perplexity unless one ensembles across order contexts, which substantially narrows the gap. At the same time, decoder-only models enable linear-time generation with KV caching and deliver large speedups (up to 25×), whereas encoder-only MDMs occupy a simpler conditional space but require multi-step refinement. With careful temperature/annealing and order handling, decoder-only MDMs reach competitive perplexity, highlighting clear trade-offs between modeling space and efficiency. Overall, the work decouples paradigm from architecture to provide a fairer basis for evaluating AR and MDM and to guide future MDMs design.", "review_text": "This work is the first systematic and equitable comparison of masked-diffusion language models (MDMs) and standard autoregressive (AR) LMs by decoupling training paradigm from architecture. The authors show that when both are instantiated with the same decoder-only backbone, the core distinction reduces to the distribution over token orders. They therefore implement MDM as Any-Order AR within a decoder-only GPT (“AO-GPT”) and study: (1) how AR vs. AO-AR differ in modeling capacity and empirical behavior under an identical architecture; and (2) for MDMs, how encoder-only and decoder-only architectures compare theoretically and empirically. Experiments find AO-GPT converges notably more slowly early in training than left-to-right (L2R) GPT; a fixed block-wise random order interpolates between L2R and fully random in convergence; and mixing a small fraction (~10%) of L2R samples improves AO performance. Decoder-only AO-AR underperforms encoder-only variants on perplexity unless one ensembles across order contexts, which substantially narrows the gap. At the same time, decoder-only models enable linear-time generation with KV caching and deliver large speedups (up to 25×), whereas encoder-only MDMs occupy a simpler conditional space but require multi-step refinement. With careful temperature/annealing and order handling, decoder-only MDMs reach competitive perplexity, highlighting clear trade-offs between modeling space and efficiency. Overall, the work decouples paradigm from architecture to provide a fairer basis for evaluating AR and MDM and to guide future MDMs design.", "strengths": "(1). This work clearly and rigorously shows that the MDM loss function is mathematically equivalent to the AO-AR loss. This is important for putting AR and MDM on a common theoretical footing, isolating the true source of differences to the token-order distribution rather than architecture, and enabling apples-to-apples empirical comparisons that inform practical choices like order mixing, annealing, and ensembling. Moreover, the equivalence between the efficient sampling algorithm and Eq. (8) also makes the generation speedup for AO-GPT convincing and well-supported. \n\n(2). The experiments are thorough and well motivated. For example, to carefully answer the first research question, this work keeps the decoder-only backbone, datasets, and hyperparameters fixed, vary only the token-order distribution, and add targeted ablations (e.g., ~10% L2R mixing). It also reports both quality metrics like perplexity and convergence curves and systems metrics like generation time across steps. This tight isolation of variables enables causal conclusions and yields clear and actionable takeaways, underscoring the work’s practical significance.\n\n(3). This manuscript is well written. The concepts are introduced progressively, notation is consistent, experiments are well-motivated, making the findings and methodology easy to follow and build upon.", "weaknesses": "(1). The major concern is the scalability of the findings elaborated in this work. This work only tests small model with 350M parameters. It is unclear whether the findings like the observed convergence behavior and order-mixing benefits hold true for larger models. Even though the manuscript acknowledges this, but it does not provide confirming evidence, which is indeed a non-negligible weakness.\n\n(2). This work only tests on perplexity, with limited coverage of downstream conditional tasks like QA, summarization, and long-context retrieval/reasoning. Since token-order policies and annealing and ensembling may behave differently under conditioning and at much longer sequence lengths, the validity of the findings is quite constrained.\n\n(3). Even though Finding 8.2 is convincing given the experiment results in Table 2, selecting these settings is non-trivial, appears dataset- and model-dependent, and this work does not standardize or bound the tuning budget. This raises real concerns about the feasibility of this finding. In practice, extensive per-task or per-model sweeps could erode the claimed benefits.", "questions": "(1). It is interesting to know Finding2.2 that fixed block-wise random serves as an interpolation between Left-to-Right and purely random order in terms of convergence speed. How sensitive are these gains to the block size? Do you observe regimes where larger or smaller blocks hurt convergence or final perplexity?\n\n(2). Following Q(1), recent MDM works like Fast-dLLM-v2 [1] is also using this block-like design, achieving excellent performance. Do you think your findings on block-wise random can explain its superior performance?\n\n(3). For Finding 3.2, why do you only test 10% L2R data? What will the performance be like for other proportions?\n\n(4). Could the authors give some thoughts on a principled and low-overhead strategy to set and adapt annealing? \n\n[1]. Wu, C., Zhang, H., Xue, S., Diao, S., Fu, Y., Liu, Z., Molchanov, P., Luo, P., Han, S., & Xie, E. (2025). Fast-dLLM v2: Efficient Block-Diffusion LLM. arXiv preprint arXiv:2509.26328.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work is the first systematic and equitable comparison of masked-diffusion language models (MDMs) and standard autoregressive (AR) LMs by decoupling training paradigm from architecture. The authors show that when both are instantiated with the same decoder-only backbone, the core distinction reduces to the distribution over token orders. They therefore implement MDM as Any-Order AR within a decoder-only GPT (“AO-GPT”) and study: (1) how AR vs. AO-AR differ in modeling capacity and empirical behavior under an identical architecture; and (2) for MDMs, how encoder-only and decoder-only architectures compare theoretically and empirically. Experiments find AO-GPT converges notably more slowly early in training than left-to-right (L2R) GPT; a fixed block-wise random order interpolates between L2R and fully random in convergence; and mixing a small fraction (~10%) of L2R samples improves AO performance. Decoder-only AO-AR underperforms encoder-only variants on perplexity unless one ensembles across order contexts, which substantially narrows the gap. At the same time, decoder-only models enable linear-time generation with KV caching and deliver large speedups (up to 25×), whereas encoder-only MDMs occupy a simpler conditional space but require multi-step refinement. With careful temperature/annealing and order handling, decoder-only MDMs reach competitive perplexity, highlighting clear trade-offs between modeling space and efficiency. Overall, the work decouples paradigm from architecture to provide a fairer basis for evaluating AR and MDM and to guide future MDMs design.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "(1). This work clearly and rigorously shows that the MDM loss function is mathematically equivalent to the AO-AR loss. This is important for putting AR and MDM on a common theoretical footing, isolating the true source of differences to the token-order distribution rather than architecture, and enabling apples-to-apples empirical comparisons that inform practical choices like order mixing, annealing, and ensembling. Moreover, the equivalence between the efficient sampling algorithm and Eq. (8) also makes the generation speedup for AO-GPT convincing and well-supported. \n\n(2). The experiments are thorough and well motivated. For example, to carefully answer the first research question, this work keeps the decoder-only backbone, datasets, and hyperparameters fixed, vary only the token-order distribution, and add targeted ablations (e.g., ~10% L2R mixing). It also reports both quality metrics like perplexity and convergence curves and systems metrics like generation time across steps. This tight isolation of variables enables causal conclusions and yields clear and actionable takeaways, underscoring the work’s practical significance.\n\n(3). This manuscript is well written. The concepts are introduced progressively, notation is consistent, experiments are well-motivated, making the findings and methodology easy to follow and build upon.", "weaknesses": "(1). The major concern is the scalability of the findings elaborated in this work. This work only tests small model with 350M parameters. It is unclear whether the findings like the observed convergence behavior and order-mixing benefits hold true for larger models. Even though the manuscript acknowledges this, but it does not provide confirming evidence, which is indeed a non-negligible weakness.\n\n(2). This work only tests on perplexity, with limited coverage of downstream conditional tasks like QA, summarization, and long-context retrieval/reasoning. Since token-order policies and annealing and ensembling may behave differently under conditioning and at much longer sequence lengths, the validity of the findings is quite constrained.\n\n(3). Even though Finding 8.2 is convincing given the experiment results in Table 2, selecting these settings is non-trivial, appears dataset- and model-dependent, and this work does not standardize or bound the tuning budget. This raises real concerns about the feasibility of this finding. In practice, extensive per-task or per-model sweeps could erode the claimed benefits.", "questions": "(1). It is interesting to know Finding2.2 that fixed block-wise random serves as an interpolation between Left-to-Right and purely random order in terms of convergence speed. How sensitive are these gains to the block size? Do you observe regimes where larger or smaller blocks hurt convergence or final perplexity?\n\n(2). Following Q(1), recent MDM works like Fast-dLLM-v2 [1] is also using this block-like design, achieving excellent performance. Do you think your findings on block-wise random can explain its superior performance?\n\n(3). For Finding 3.2, why do you only test 10% L2R data? What will the performance be like for other proportions?\n\n(4). Could the authors give some thoughts on a principled and low-overhead strategy to set and adapt annealing? \n\n[1]. Wu, C., Zhang, H., Xue, S., Diao, S., Fu, Y., Liu, Z., Molchanov, P., Luo, P., Han, S., & Xie, E. (2025). Fast-dLLM v2: Efficient Block-Diffusion LLM. arXiv preprint arXiv:2509.26328.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761975077182}, {"id": "UBMoBzi7yL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13483/Reviewer_rwo7"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper aims to fairly compare autoregressive (AR) and masked diffusion (MDM) paradigms by decoupling formulation from architecture. Prior works conflated these by pairing AR with decoder-only causal attention and MDM with encoder-only full attention.\n\nThe authors introduce AO-GPT, a decoder-only masked diffusion model that implements the Any-Order Autoregressive (AO-AR) objective, equivalent to MDM but averaged over all token permutations. Using identical architectures, they reveal:\n\n* AO-GPT trains slower than left-to-right GPT because many random token orders are uninformative.\n\n* Injecting 10% left-to-right data sharply improves both convergence and perplexity, showing language’s inherent sequential bias.\n\nCompared with encoder-only MDMs, decoder-based AO-GPT models a vastly larger conditional space (~ e · n!) but achieves ≈ 25× faster generation with KV-cache and efficient sampling, approaching MDM perplexity after temperature annealing. Architecturally, AO-GPT employs adaptive LayerNorm and EMA to stabilize training, achieving near-encoder-level performance while preserving decoder efficiency.", "review_text": "This paper aims to fairly compare autoregressive (AR) and masked diffusion (MDM) paradigms by decoupling formulation from architecture. Prior works conflated these by pairing AR with decoder-only causal attention and MDM with encoder-only full attention.\n\nThe authors introduce AO-GPT, a decoder-only masked diffusion model that implements the Any-Order Autoregressive (AO-AR) objective, equivalent to MDM but averaged over all token permutations. Using identical architectures, they reveal:\n\n* AO-GPT trains slower than left-to-right GPT because many random token orders are uninformative.\n\n* Injecting 10% left-to-right data sharply improves both convergence and perplexity, showing language’s inherent sequential bias.\n\nCompared with encoder-only MDMs, decoder-based AO-GPT models a vastly larger conditional space (~ e · n!) but achieves ≈ 25× faster generation with KV-cache and efficient sampling, approaching MDM perplexity after temperature annealing. Architecturally, AO-GPT employs adaptive LayerNorm and EMA to stabilize training, achieving near-encoder-level performance while preserving decoder efficiency.", "strengths": "1.  The work’s exploratory framing, decoupling modeling formulation (AR vs. MDM) from architectural choice (encoder vs. decoder), is conceptually fresh and helps clarify long-standing confusions in diffusion-based language modeling. This perspective provides some insights into how generation order and attention structure affect learning and efficiency.\n\n2. Although limited in scale, the experiments are carefully controlled and serve as useful ablations for understanding the impact of causal vs. full attention and decoder vs. encoder designs. The results are interpretable with clear reasoning, making them valuable for guiding future model design. The finding that partial left-to-right ordering stabilizes any-order training offers a practical and actionable insight for improving masked diffusion LMs without heavy architectural change, echoing ideas from hybrid AR–diffusion works like block diffusion.", "weaknesses": "1. Lack of clear narrative and focus. The paper presents over eight separate findings spanning AR vs. AO-AR, encoder vs. decoder MDM, training dynamics, architectural ablations, and efficiency. This breadth makes the work feel more like a collection of exploratory observations than a cohesive study with a central takeaway. The main conceptual message, how “any-order” decoding interacts with causal attention, gets diluted by numerous side analyses and minor findings. The authors should consolidate around one or two key insights to strengthen the paper’s focus.\n\n2. Questionable fairness premise. The notion of “fair comparison” by forcing MDM into a decoder-only causal-attention setup is debatable. Changing an encoder-based model to match the AR architecture inherently biases the comparison toward AR-style inductive biases. A symmetric control, e.g., adapting AR to an encoder/full-attention variant, would make the fairness claim more credible. As written, the work could be seen as optimizing one paradigm (MDM) to fit another’s design, rather than evaluating them on equal footing.\n\n3. Limited novelty in methodology. The technical modifications (decoder reimplementation, order ensembling, adaptive LayerNorm, EMA) are individually incremental and borrowed from prior work (e.g., σ-GPT, diffusion model training). The novelty lies mainly in the analysis framework, but this could be emphasized more explicitly rather than presented as model contributions.\n\n4. Small-scale experiments and weak generalization claims. All experiments are conducted at ≤350M parameters. Given the known scale sensitivity of LLM training, these results may not extrapolate to the multi-billion-parameter regime. The claimed speed–quality trade-offs, convergence behaviors, and “10% L2R” benefit might differ under large-scale or instruction-tuned settings. The paper would benefit from at least partial large-scale validation or sensitivity analysis to hyperparameters.\n\n5. Underdeveloped connection to prior hybrid approaches. The observation that \"partial autoregressive ordering\" stabilizes MDM training aligns with prior “block diffusion” or “hybrid AR–diffusion” work, but this link is not acknowledged.", "questions": "The fairness claim is very questionable, converting MDMs into a decoder-only setup inherently imposes AR-style inductive biases, so it’s unclear why this one-directional adaptation is considered “fair,” or how conclusions might change if AR models were instead adapted to an encoder/full-attention configuration.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to fairly compare autoregressive (AR) and masked diffusion (MDM) paradigms by decoupling formulation from architecture. Prior works conflated these by pairing AR with decoder-only causal attention and MDM with encoder-only full attention.\n\nThe authors introduce AO-GPT, a decoder-only masked diffusion model that implements the Any-Order Autoregressive (AO-AR) objective, equivalent to MDM but averaged over all token permutations. Using identical architectures, they reveal:\n\n* AO-GPT trains slower than left-to-right GPT because many random token orders are uninformative.\n\n* Injecting 10% left-to-right data sharply improves both convergence and perplexity, showing language’s inherent sequential bias.\n\nCompared with encoder-only MDMs, decoder-based AO-GPT models a vastly larger conditional space (~ e · n!) but achieves ≈ 25× faster generation with KV-cache and efficient sampling, approaching MDM perplexity after temperature annealing. Architecturally, AO-GPT employs adaptive LayerNorm and EMA to stabilize training, achieving near-encoder-level performance while preserving decoder efficiency.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1.  The work’s exploratory framing, decoupling modeling formulation (AR vs. MDM) from architectural choice (encoder vs. decoder), is conceptually fresh and helps clarify long-standing confusions in diffusion-based language modeling. This perspective provides some insights into how generation order and attention structure affect learning and efficiency.\n\n2. Although limited in scale, the experiments are carefully controlled and serve as useful ablations for understanding the impact of causal vs. full attention and decoder vs. encoder designs. The results are interpretable with clear reasoning, making them valuable for guiding future model design. The finding that partial left-to-right ordering stabilizes any-order training offers a practical and actionable insight for improving masked diffusion LMs without heavy architectural change, echoing ideas from hybrid AR–diffusion works like block diffusion.", "weaknesses": "1. Lack of clear narrative and focus. The paper presents over eight separate findings spanning AR vs. AO-AR, encoder vs. decoder MDM, training dynamics, architectural ablations, and efficiency. This breadth makes the work feel more like a collection of exploratory observations than a cohesive study with a central takeaway. The main conceptual message, how “any-order” decoding interacts with causal attention, gets diluted by numerous side analyses and minor findings. The authors should consolidate around one or two key insights to strengthen the paper’s focus.\n\n2. Questionable fairness premise. The notion of “fair comparison” by forcing MDM into a decoder-only causal-attention setup is debatable. Changing an encoder-based model to match the AR architecture inherently biases the comparison toward AR-style inductive biases. A symmetric control, e.g., adapting AR to an encoder/full-attention variant, would make the fairness claim more credible. As written, the work could be seen as optimizing one paradigm (MDM) to fit another’s design, rather than evaluating them on equal footing.\n\n3. Limited novelty in methodology. The technical modifications (decoder reimplementation, order ensembling, adaptive LayerNorm, EMA) are individually incremental and borrowed from prior work (e.g., σ-GPT, diffusion model training). The novelty lies mainly in the analysis framework, but this could be emphasized more explicitly rather than presented as model contributions.\n\n4. Small-scale experiments and weak generalization claims. All experiments are conducted at ≤350M parameters. Given the known scale sensitivity of LLM training, these results may not extrapolate to the multi-billion-parameter regime. The claimed speed–quality trade-offs, convergence behaviors, and “10% L2R” benefit might differ under large-scale or instruction-tuned settings. The paper would benefit from at least partial large-scale validation or sensitivity analysis to hyperparameters.\n\n5. Underdeveloped connection to prior hybrid approaches. The observation that \"partial autoregressive ordering\" stabilizes MDM training aligns with prior “block diffusion” or “hybrid AR–diffusion” work, but this link is not acknowledged.", "questions": "The fairness claim is very questionable, converting MDMs into a decoder-only setup inherently imposes AR-style inductive biases, so it’s unclear why this one-directional adaptation is considered “fair,” or how conclusions might change if AR models were instead adapted to an encoder/full-attention configuration.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761963852965}], "openreview_url": "https://openreview.net/forum?id=AeHZWzDjTk", "arxiv_id": "2506.19935", "paper_pdf": "papers/AeHZWzDjTk.pdf", "paper_pdf_sha256": "41cd8050bfe9d7931934454d9e43bf2784ec214d5e789da4de0af6871149ec69", "paper_pdf_bytes": 680030, "paper_pdf_source": "openreview", "code_url": "https://github.com/scxue/AO-GPT-MDM", "code_repository": "scxue/AO-GPT-MDM", "code_commit": "3a02e3bc60c7514f93601de6a39034bbefa8555b", "code_archive": "repos/AeHZWzDjTk.zip", "code_archive_sha256": "4388e0819cb0b3932d7fd61ed825ae8d451a15f4faf4723c03f51d0c7cce46c8", "code_archive_bytes": 69704, "code_file_count": 30, "code_extensions": {".py": 25, ".sh": 5}, "github_disk_usage_kb": 485, "github_languages": {"Python": 197240, "Shell": 5594}, "github_archived": false, "github_pushed_at": "2025-06-23T20:52:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/any-order-gpt-as-masked-diffusion-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Ng1r9kTep4", "year": 2025, "status": "rejected", "title": "Inverted Activations: Reducing Memory Footprint in Neural Network Training", "authors": ["Georgii Sergeevich Novikov", "Ivan Oseledets"], "authorids": ["~Georgii_Sergeevich_Novikov1", "~Ivan_Oseledets1"], "authors_source": "OpenReview API", "abstract": "The scaling of neural networks with increasing data and model sizes necessitates the development of more efficient deep learning algorithms. \n    A significant challenge in neural network training is the memory footprint associated with activation tensors, particularly in pointwise nonlinearity layers that traditionally save the entire input tensor for the backward pass, leading to substantial memory consumption.\n    \n    In this paper, we propose a modification to the handling of activation tensors in pointwise nonlinearity layers. \n    Our method involves saving the output tensor instead of the input tensor during the forward pass. Since the subsequent layer typically also saves its input tensor, this approach reduces the total memory required by storing only one tensor between layers instead of two. This optimization is especially beneficial for transformer-based architectures like GPT, BERT, Mistral, and Llama.\n\n    To enable this approach, we utilize the inverse function of the nonlinearity during the backward pass. As the inverse cannot be computed analytically for most nonlinearities, we construct accurate approximations using simpler functions. \n    Experimental results demonstrate that our method significantly reduces memory usage without affecting training accuracy or computational performance.\n\n    Our implementation is provided as a drop-in replacement for standard nonlinearity layers in the PyTorch framework, facilitating easy adoption without requiring architectural modifications. The code is available at \\url{https://github.com/removed/for/anonimity}.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "e02xDYKiU4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9876/Reviewer_jQ9m"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper proposes a method, called InvAct, to reduce the memory footprint associated with activation tensors during neural network training by modifying how activation tensors are saved for the backward pass in pointwise nonlinearity layers like GELU and SiLU.\nConsidering a block of two layers, instead of saving the input tensor for each layer, the authors suggest storing the output tensor of layer 1, which is the input tensor of layer 2. Then during the backward pass the input layer of layer 1 is reconstructed using the inverse of the nonlinearity. Since the nonlinearities considered in the paper do not have straightforward analytical inverses, the paper approximates the inverse functions using simpler functions. The paper shows great deal of memory savings for Transformer based architecture, around 25%, without noticeable impact on accuracy or computational efficiency, and can be used as a drop-in replacement for existing nonlinearity layers in PyTorch.", "review_text": "This paper proposes a method, called InvAct, to reduce the memory footprint associated with activation tensors during neural network training by modifying how activation tensors are saved for the backward pass in pointwise nonlinearity layers like GELU and SiLU.\nConsidering a block of two layers, instead of saving the input tensor for each layer, the authors suggest storing the output tensor of layer 1, which is the input tensor of layer 2. Then during the backward pass the input layer of layer 1 is reconstructed using the inverse of the nonlinearity. Since the nonlinearities considered in the paper do not have straightforward analytical inverses, the paper approximates the inverse functions using simpler functions. The paper shows great deal of memory savings for Transformer based architecture, around 25%, without noticeable impact on accuracy or computational efficiency, and can be used as a drop-in replacement for existing nonlinearity layers in PyTorch.", "strengths": "- **Originality**: The approach introduces a novel solution to the memory bottleneck in neural networks training. The proposed solution is straightforward yet impactful adjustment for deep learning workflows.\n- **Clarity**: The method is clearly outlined, with pseudocode, detailed experimentation, and well written text.\n- **Significance**: The reduction of memory footprint in training aligns well with current trends in model scaling. The method's compatibility with widely-used architectures and frameworks adds further value, especially for resource-constrained training environments.", "weaknesses": "- **Experimental Scope**: The evaluation is limited to a few model architectures (BERT and LLama) and tasks (e.g., BERT fine-tuning on Yelp Reviews). Adding results from multiple tasks, such as [GLUE](https://huggingface.co/datasets/nyu-mll/glue), would strengthen claims of general applicability and robustness across varied scenarios.\n- **Reporting of Experimental Details**: \nIncluding the fine-tuning details for the results reported in Sections 3.2 and 3.3, perhaps in the appendix could benefit the readers and help future research with reproducing the results reported in the paper.\n- **Lack of Task-Specific Performance Metrics**: The results shown in Section 3.2 focus on training/validation loss without reporting task-specific metrics (e.g., accuracy, F1 score for Yelp Reviews). Including such metrics would better contextualize the impact of approximation errors on model quality.\n- **Lack of Performance Evaluation for Other Models**: Although memory savings are reported for models like the Audio Spectral Transformer, ViT, and CLIP, no experimental analysis is provided to assess if or how these savings impact model performance. For instance, similar to the BERT and Llama evaluations in Sections 3.1 and 3.2, experiments could show if the method affects performance metrics and how much computational overhead is introduced for these models. Furthermore, in the abstract, the authors mention models like *GPT* and *Mistral* as potential beneficiaries of the method, yet there is no experimental data on memory savings, time overhead, or performance impact for these models. Including analysis or removing these model names would provide a more accurate representation of the paper's scope.\n\n**P.S.**: I am more than willing to raise my score if the concerns mentioned in the **Weakness** and **Questions** section are addressed.", "questions": "1. **Sample Size and Standard Deviation**: Are the results in Table 1 averaged over multiple trials? If so, what is the sample size and standard deviation?  If the results show metrics over only a single run, then reporting the results over multiple runs with reports on the standard deviations would better highlight robustness and reproducibility of the method.\n2. **Precision Bit InvAct**: I am curious to know if you have performed any performance evaluation as seen in Sections 3.2 and 3.3 on Precision Bit InvAct. The results presented in Table 1 suggest this variant to have less overhead but it is not clear how much error it produces in downstream tasks.\n3. **Clarification on Experimental Consistency**: For the FewBit comparison in Figure 5, was the same dataset used as in Section 3.2? If different datasets were used, specifying this in Section 3.3 would enhance clarity.\n4. **Sequence Length in BERT Experiment**: In Appendix A.3, the experiment setup for BERT specifies a sequence length of 1024, which exceeds the model’s typical limitation of 512 tokens. Could you clarify how this was achieved? Was a specific technique used to expand the sequence limit to 1024? If so please share more details.\n## Minor suggestions:\n1. **Additional Details Needed in Listing 1**: In Listing 1, memory savings for the Llama model are not shown, despite Llama being used in Section 3 for performance analysis. Including this information would create a more comprehensive view of the memory savings achievable with this approach.\n2. **Sentence Rewrite in Line 464**: In the sentence, \"Former are used more in during inference,\" the word \"in\" is unnecessary. The corrected sentence should read:  \n  > \"Former are used more during inference.\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method, called InvAct, to reduce the memory footprint associated with activation tensors during neural network training by modifying how activation tensors are saved for the backward pass in pointwise nonlinearity layers like GELU and SiLU.\nConsidering a block of two layers, instead of saving the input tensor for each layer, the authors suggest storing the output tensor of layer 1, which is the input tensor of layer 2. Then during the backward pass the input layer of layer 1 is reconstructed using the inverse of the nonlinearity. Since the nonlinearities considered in the paper do not have straightforward analytical inverses, the paper approximates the inverse functions using simpler functions. The paper shows great deal of memory savings for Transformer based architecture, around 25%, without noticeable impact on accuracy or computational efficiency, and can be used as a drop-in replacement for existing nonlinearity layers in PyTorch.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "- **Originality**: The approach introduces a novel solution to the memory bottleneck in neural networks training. The proposed solution is straightforward yet impactful adjustment for deep learning workflows.\n- **Clarity**: The method is clearly outlined, with pseudocode, detailed experimentation, and well written text.\n- **Significance**: The reduction of memory footprint in training aligns well with current trends in model scaling. The method's compatibility with widely-used architectures and frameworks adds further value, especially for resource-constrained training environments.", "weaknesses": "- **Experimental Scope**: The evaluation is limited to a few model architectures (BERT and LLama) and tasks (e.g., BERT fine-tuning on Yelp Reviews). Adding results from multiple tasks, such as [GLUE](https://huggingface.co/datasets/nyu-mll/glue), would strengthen claims of general applicability and robustness across varied scenarios.\n- **Reporting of Experimental Details**: \nIncluding the fine-tuning details for the results reported in Sections 3.2 and 3.3, perhaps in the appendix could benefit the readers and help future research with reproducing the results reported in the paper.\n- **Lack of Task-Specific Performance Metrics**: The results shown in Section 3.2 focus on training/validation loss without reporting task-specific metrics (e.g., accuracy, F1 score for Yelp Reviews). Including such metrics would better contextualize the impact of approximation errors on model quality.\n- **Lack of Performance Evaluation for Other Models**: Although memory savings are reported for models like the Audio Spectral Transformer, ViT, and CLIP, no experimental analysis is provided to assess if or how these savings impact model performance. For instance, similar to the BERT and Llama evaluations in Sections 3.1 and 3.2, experiments could show if the method affects performance metrics and how much computational overhead is introduced for these models. Furthermore, in the abstract, the authors mention models like *GPT* and *Mistral* as potential beneficiaries of the method, yet there is no experimental data on memory savings, time overhead, or performance impact for these models. Including analysis or removing these model names would provide a more accurate representation of the paper's scope.\n\n**P.S.**: I am more than willing to raise my score if the concerns mentioned in the **Weakness** and **Questions** section are addressed.", "questions": "1. **Sample Size and Standard Deviation**: Are the results in Table 1 averaged over multiple trials? If so, what is the sample size and standard deviation?  If the results show metrics over only a single run, then reporting the results over multiple runs with reports on the standard deviations would better highlight robustness and reproducibility of the method.\n2. **Precision Bit InvAct**: I am curious to know if you have performed any performance evaluation as seen in Sections 3.2 and 3.3 on Precision Bit InvAct. The results presented in Table 1 suggest this variant to have less overhead but it is not clear how much error it produces in downstream tasks.\n3. **Clarification on Experimental Consistency**: For the FewBit comparison in Figure 5, was the same dataset used as in Section 3.2? If different datasets were used, specifying this in Section 3.3 would enhance clarity.\n4. **Sequence Length in BERT Experiment**: In Appendix A.3, the experiment setup for BERT specifies a sequence length of 1024, which exceeds the model’s typical limitation of 512 tokens. Could you clarify how this was achieved? Was a specific technique used to expand the sequence limit to 1024? If so please share more details.\n## Minor suggestions:\n1. **Additional Details Needed in Listing 1**: In Listing 1, memory savings for the Llama model are not shown, despite Llama being used in Section 3 for performance analysis. Including this information would create a more comprehensive view of the memory savings achievable with this approach.\n2. **Sentence Rewrite in Line 464**: In the sentence, \"Former are used more in during inference,\" the word \"in\" is unnecessary. The corrected sentence should read:  \n  > \"Former are used more during inference.\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731172577622}, {"id": "lDvumYyidQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9876/Reviewer_ZY8F"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This article presents a modification to the GELU and SiLU activations to save the output tensor instead of the input, recomputing the input to save approximately a quarter of the memory overhead in practice. They provide several variants of their method, all revolving around the idea of dividing the activation function into monotonous parts which can be inverted. They measure extensively that their approach does not change the computation time and doesn't affect performance.", "review_text": "This article presents a modification to the GELU and SiLU activations to save the output tensor instead of the input, recomputing the input to save approximately a quarter of the memory overhead in practice. They provide several variants of their method, all revolving around the idea of dividing the activation function into monotonous parts which can be inverted. They measure extensively that their approach does not change the computation time and doesn't affect performance.", "strengths": "- The method proposed is simple and only requires a drop-in replacement of the activation functions, providing a major memory improvement.\n- The proposed inversion methods are novel, notably the approximate inverses of the GELU and SiLU functions.\n- The computational efficiency of the method is considered on many different frameworks.\n- The paper is clear and well-presented.", "weaknesses": "- The main method proposed has very limited novelty outside of the consideration of GELU and SiLU. Other approaches have already proposed inverting the computations of the activation function such as [1] for similar reasons reducing the memory by half and with further benefits. \n- At no point comparison with activation/gradient checkpointing is considered, despite the method being extremely similar. Indeed, why not simply store the input tensor only and recompute the output of the activation for efficient gradient computation? This is the same method, but it doesn't require an approximate inversion like proposed here. An extensive literature on checkpointing is necessary to consider here to compare the proposed method to, such as [2-4]. Notably, [5] has a particular consideration for the forward and backward of GeLU. It seems very hard to justify this method compared to more standard checkpointing methods, which computes an optimal trade-off between activation and computation time.\n- The approximation error of the derivative of the inverse as represented in Figure 2 shows a non-negligible error for values of $x$ close to 0, which seems a bit worrying.\n- Other activation functions are not considered. How is ReLU supposed to work in this framework?\n\n[1] In-Place Activated BatchNorm for Memory-Optimized Training of DNNs, S. R. Bulò et al.\n\n[2] Reducing Activation Recomputation in Large Transformer Models, V. Korthikanti et al.\n\n[3] Efficient rematerialization for deep networks, Kumar et al. \n\n[4] Rockmate: an efficient, fast, automatic and generic tool for re-materialization in pytorch, Zhao et al.\n\n[5] Transcending Runtime-Memory Tradeoffs in Checkpointing by being Fusion Aware, S. Yu et al.", "questions": "- How were the coefficients $c_i$ obtained ?\n- Why is there an increase in execution time for InvAct only for the Plain GELU? What about the execution time for SiLU?\n- The Sign-bit Inverted activation should be explained more clearly.\n- Figure 3 requires labels for the rows and columns to be more clear.\n\n**Minor details**\n\n- The abstract in the openreview summary has a formatting error.\n- line 182 \"based\"\n- line 701 \"assstance\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article presents a modification to the GELU and SiLU activations to save the output tensor instead of the input, recomputing the input to save approximately a quarter of the memory overhead in practice. They provide several variants of their method, all revolving around the idea of dividing the activation function into monotonous parts which can be inverted. They measure extensively that their approach does not change the computation time and doesn't affect performance.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The method proposed is simple and only requires a drop-in replacement of the activation functions, providing a major memory improvement.\n- The proposed inversion methods are novel, notably the approximate inverses of the GELU and SiLU functions.\n- The computational efficiency of the method is considered on many different frameworks.\n- The paper is clear and well-presented.", "weaknesses": "- The main method proposed has very limited novelty outside of the consideration of GELU and SiLU. Other approaches have already proposed inverting the computations of the activation function such as [1] for similar reasons reducing the memory by half and with further benefits. \n- At no point comparison with activation/gradient checkpointing is considered, despite the method being extremely similar. Indeed, why not simply store the input tensor only and recompute the output of the activation for efficient gradient computation? This is the same method, but it doesn't require an approximate inversion like proposed here. An extensive literature on checkpointing is necessary to consider here to compare the proposed method to, such as [2-4]. Notably, [5] has a particular consideration for the forward and backward of GeLU. It seems very hard to justify this method compared to more standard checkpointing methods, which computes an optimal trade-off between activation and computation time.\n- The approximation error of the derivative of the inverse as represented in Figure 2 shows a non-negligible error for values of $x$ close to 0, which seems a bit worrying.\n- Other activation functions are not considered. How is ReLU supposed to work in this framework?\n\n[1] In-Place Activated BatchNorm for Memory-Optimized Training of DNNs, S. R. Bulò et al.\n\n[2] Reducing Activation Recomputation in Large Transformer Models, V. Korthikanti et al.\n\n[3] Efficient rematerialization for deep networks, Kumar et al. \n\n[4] Rockmate: an efficient, fast, automatic and generic tool for re-materialization in pytorch, Zhao et al.\n\n[5] Transcending Runtime-Memory Tradeoffs in Checkpointing by being Fusion Aware, S. Yu et al.", "questions": "- How were the coefficients $c_i$ obtained ?\n- Why is there an increase in execution time for InvAct only for the Plain GELU? What about the execution time for SiLU?\n- The Sign-bit Inverted activation should be explained more clearly.\n- Figure 3 requires labels for the rows and columns to be more clear.\n\n**Minor details**\n\n- The abstract in the openreview summary has a formatting error.\n- line 182 \"based\"\n- line 701 \"assstance\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730386654384}, {"id": "NL418BryOX", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9876/Reviewer_r3ne"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "This paper derives a method that reduces the amount of data that has to be stored in memory during the forward pass for the backpropagation. By doing so, it is possible to save a significant amount of memory during the training of a neural network. The proposed method is easily implementable and can be deployed in a training procedure with almost no impact\n+ on the training duration,\n+ on the overall performance of the model,\n\nas it is supported by experimental results.", "review_text": "This paper derives a method that reduces the amount of data that has to be stored in memory during the forward pass for the backpropagation. By doing so, it is possible to save a significant amount of memory during the training of a neural network. The proposed method is easily implementable and can be deployed in a training procedure with almost no impact\n+ on the training duration,\n+ on the overall performance of the model,\n\nas it is supported by experimental results.", "strengths": "+ The proposed method is derived from a very simple idea, yet it leads to substantial reduction in memory costs, without introducing significant drawbacks or trade-offs (in particular it has no significant impact on either the training time or the model's performance)\n+ The method is explained in a very clear way, as well as the research objectives of the paper\n+ Thanks to its simplicity, the method is likely implementable within existing frameworks, without changing their core functioning", "weaknesses": "The main weeknesses of the paper lie in its experimental section:\n+ Some claims of the paper are not supported by experiments, such as line 60: \"effectively reducing the memory footprint by nearly $25\\\\%$ in practice\", it would be nice to see empirical evidence of such a claim\n+ More generally, no experiment on the memory footprint achieved by the proposed method is shown, yet since the main objective of the paper is apparently to reduce the memory footprint during training, the paper would benefit greatly from such results\n+ In section 3.1, the authors use two different frameworks to assess the computational efficiency of their method, but it seems thus difficult to have a clear view on whether the observed gap in computational efficiency is due to the authors method, or to the fact that two different frameworks are being used\n+ Finally, could the authors add a theoretical example of how in practice the proposed method would allow to discard some tensors for back propagation computation would make the contribution of the paper more striking", "questions": "+ The authors claim that the studied activations are chosen because they are popular choices for transformers, did the authors try to apply their methods to other activations / nonlinearities ?\n+ There is a tipo at line 337 (\"implementatoins\")", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper derives a method that reduces the amount of data that has to be stored in memory during the forward pass for the backpropagation. By doing so, it is possible to save a significant amount of memory during the training of a neural network. The proposed method is easily implementable and can be deployed in a training procedure with almost no impact\n+ on the training duration,\n+ on the overall performance of the model,\n\nas it is supported by experimental results.", "soundness": 2, "presentation": 3, "contribution": 4, "strengths": "+ The proposed method is derived from a very simple idea, yet it leads to substantial reduction in memory costs, without introducing significant drawbacks or trade-offs (in particular it has no significant impact on either the training time or the model's performance)\n+ The method is explained in a very clear way, as well as the research objectives of the paper\n+ Thanks to its simplicity, the method is likely implementable within existing frameworks, without changing their core functioning", "weaknesses": "The main weeknesses of the paper lie in its experimental section:\n+ Some claims of the paper are not supported by experiments, such as line 60: \"effectively reducing the memory footprint by nearly $25\\\\%$ in practice\", it would be nice to see empirical evidence of such a claim\n+ More generally, no experiment on the memory footprint achieved by the proposed method is shown, yet since the main objective of the paper is apparently to reduce the memory footprint during training, the paper would benefit greatly from such results\n+ In section 3.1, the authors use two different frameworks to assess the computational efficiency of their method, but it seems thus difficult to have a clear view on whether the observed gap in computational efficiency is due to the authors method, or to the fact that two different frameworks are being used\n+ Finally, could the authors add a theoretical example of how in practice the proposed method would allow to discard some tensors for back propagation computation would make the contribution of the paper more striking", "questions": "+ The authors claim that the studied activations are chosen because they are popular choices for transformers, did the authors try to apply their methods to other activations / nonlinearities ?\n+ There is a tipo at line 337 (\"implementatoins\")", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729776714897}], "openreview_url": "https://openreview.net/forum?id=Ng1r9kTep4", "arxiv_id": "2407.15545", "paper_pdf": "papers/Ng1r9kTep4.pdf", "paper_pdf_sha256": "0792c8e015f8d0e201ea646afb644276c0b7cd8c497438d653370221d91e3421", "paper_pdf_bytes": 227385, "paper_pdf_source": "openreview", "code_url": "https://github.com/PgLoLo/optiacts", "code_repository": "PgLoLo/optiacts", "code_commit": "1d7a1e4173bf98f1db9fa11d6e2409b149491f9f", "code_archive": "repos/Ng1r9kTep4.zip", "code_archive_sha256": "5875d2117ed6bbabddf16ef0e8b6817fa73b6115193464907a639979ddd362a8", "code_archive_bytes": 57115, "code_file_count": 10, "code_extensions": {".py": 7, ".ipynb": 3}, "github_disk_usage_kb": 55, "github_languages": {"Python": 5973}, "github_archived": false, "github_pushed_at": "2024-07-19T08:10:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/inverted-activations"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WpQbM1kBuy", "year": 2024, "status": "rejected", "title": "Prodigy: An Expeditiously Adaptive Parameter-Free Learner", "authors": ["Konstantin Mishchenko", "Aaron Defazio"], "authorids": ["~Konstantin_Mishchenko1", "~Aaron_Defazio1"], "authors_source": "OpenReview API", "abstract": "We consider the problem of estimating the learning rate in adaptive methods, such as Adagrad and Adam. We describe two techniques, Prodigy and Resetting, to provably estimate the distance to the solution $D$, which is needed to set the learning rate optimally. Our techniques are modifications of the D-Adaptation method for learning-rate-free learning. Our methods improve upon the convergence rate of D-Adaptation by a factor of $\\mathcal{O}(\\sqrt{\\log(D/d_0)})$, where $d_0$ is the initial estimate of $D$. We test our methods on 12 common logistic-regression benchmark datasets, VGG11 and ResNet-50 training on CIFAR10, ViT training on Imagenet, LSTM training on IWSLT14, DLRM training on Criteo dataset, VarNet on Knee MRI dataset, as well as RoBERTa and GPT transformer training on BookWiki. Our experimental results show that our approaches consistently outperform D-Adaptation and reach test accuracy values close to that of hand-tuned Adam.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "DyxsvRHrcL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7643/Reviewer_JC31"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper provides a modification of D-Adaptation to improve its worst-case non-asymptotic convergence rate for a G-Lipschitz objective. D-Adaptation’s convergence rate scales with $\\frac{\\log(D/d_0)}{\\sqrt{n}}$. Prodigy (the paper’s modification of D-Adaptation) has a convergence rate that scales instead with $\\frac{\\log(n)\\sqrt{\\log(D/d_0)}}{\\sqrt{n}}$. Asymptotically, Prodigy is slower due to the additional $\\log(n)$ in the numerator, but faster in finite time due to the scaling of $\\log(D/d_0)$. \n\nThe main point is that Prodigy effectively uses larger step sizes. Prodigy replaces the D-Adaptation learning rate $\\frac{d_k}{\\sqrt{G^2 + \\sum_{i=0}^k \\|g_i\\|^2}}$ with $\\frac{d_k}{\\sqrt{\\frac{1}{\\lambda_k^2} G^2 + \\sum_{i=0}^k \\left[\\left(\\frac{d_i \\lambda_i}{d_k \\lambda_k}\\right)^2 \\|{g_i}\\|^2\\right]}}$. The estimates of D, $d_i$, are non-decreasing, so for the right choices of $\\lambda_i$, the learning rate is effectively larger. \n\nThey show improved training losses from D-Adaptation for deep learning tasks on various architectures, sometimes doing better than Adam.", "review_text": "This paper provides a modification of D-Adaptation to improve its worst-case non-asymptotic convergence rate for a G-Lipschitz objective. D-Adaptation’s convergence rate scales with $\\frac{\\log(D/d_0)}{\\sqrt{n}}$. Prodigy (the paper’s modification of D-Adaptation) has a convergence rate that scales instead with $\\frac{\\log(n)\\sqrt{\\log(D/d_0)}}{\\sqrt{n}}$. Asymptotically, Prodigy is slower due to the additional $\\log(n)$ in the numerator, but faster in finite time due to the scaling of $\\log(D/d_0)$. \n\nThe main point is that Prodigy effectively uses larger step sizes. Prodigy replaces the D-Adaptation learning rate $\\frac{d_k}{\\sqrt{G^2 + \\sum_{i=0}^k \\|g_i\\|^2}}$ with $\\frac{d_k}{\\sqrt{\\frac{1}{\\lambda_k^2} G^2 + \\sum_{i=0}^k \\left[\\left(\\frac{d_i \\lambda_i}{d_k \\lambda_k}\\right)^2 \\|{g_i}\\|^2\\right]}}$. The estimates of D, $d_i$, are non-decreasing, so for the right choices of $\\lambda_i$, the learning rate is effectively larger. \n\nThey show improved training losses from D-Adaptation for deep learning tasks on various architectures, sometimes doing better than Adam.", "strengths": "- Empirically, there are some observable improvements with Prodigy versus D-Adaptation.\n- The derivation/algorithm seems to be sound.\n- I am not an expert in this area, so leave no comments about the novelty.", "weaknesses": "I’m mostly concerned about some of the claims made in the non-convex deep learning section, beyond just the fact that what any of these algorithms are really doing deep learning is unclear. \n\n- In terms of practical impact, the improvements made by Prodigy reported Figure 1/2/3 are small. Often the differences between Adam, Prodigy, and D-Adaptation are much smaller than their standard errors. It doesn’t always do better either, superseded by Adam, and sometimes D-Adaptation. \n- Adam’s initial learning rate should be hyperparameter tuned. \n- They start to make claims about the test accuracy / generalization that I don’t think they should/need to include. Specifically, they make this claim about CIFAR10 on ResNet and ImageNet on Vision Transformers. In both scenarios, the models are overfitting, trained for hundreds of epochs on the training data. Lower training loss here does not necessarily mean higher test accuracy.\n    - Under Figure 1, they write “Prodigy estimates a larger step size than D-Adaptation, which helps it reach test accuracy closer to the one of Adam.” (They aren’t talking about  any implicit bias of large learning rate, just strictly that large learning rate leads to better training loss convergence.)\n    - Under paragraph “ViT training” they write “Prodigy almost closes the gap [of test performance] between tuned Adam and D-Adaptation.” \n\nSome other minor comments:\n- Can Figure 2 be plotted in log scale? It’s hard to see the difference between the lines. \n- Text spacing issue in Page 3", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper provides a modification of D-Adaptation to improve its worst-case non-asymptotic convergence rate for a G-Lipschitz objective. D-Adaptation’s convergence rate scales with $\\frac{\\log(D/d_0)}{\\sqrt{n}}$. Prodigy (the paper’s modification of D-Adaptation) has a convergence rate that scales instead with $\\frac{\\log(n)\\sqrt{\\log(D/d_0)}}{\\sqrt{n}}$. Asymptotically, Prodigy is slower due to the additional $\\log(n)$ in the numerator, but faster in finite time due to the scaling of $\\log(D/d_0)$. \n\nThe main point is that Prodigy effectively uses larger step sizes. Prodigy replaces the D-Adaptation learning rate $\\frac{d_k}{\\sqrt{G^2 + \\sum_{i=0}^k \\|g_i\\|^2}}$ with $\\frac{d_k}{\\sqrt{\\frac{1}{\\lambda_k^2} G^2 + \\sum_{i=0}^k \\left[\\left(\\frac{d_i \\lambda_i}{d_k \\lambda_k}\\right)^2 \\|{g_i}\\|^2\\right]}}$. The estimates of D, $d_i$, are non-decreasing, so for the right choices of $\\lambda_i$, the learning rate is effectively larger. \n\nThey show improved training losses from D-Adaptation for deep learning tasks on various architectures, sometimes doing better than Adam.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- Empirically, there are some observable improvements with Prodigy versus D-Adaptation.\n- The derivation/algorithm seems to be sound.\n- I am not an expert in this area, so leave no comments about the novelty.", "weaknesses": "I’m mostly concerned about some of the claims made in the non-convex deep learning section, beyond just the fact that what any of these algorithms are really doing deep learning is unclear. \n\n- In terms of practical impact, the improvements made by Prodigy reported Figure 1/2/3 are small. Often the differences between Adam, Prodigy, and D-Adaptation are much smaller than their standard errors. It doesn’t always do better either, superseded by Adam, and sometimes D-Adaptation. \n- Adam’s initial learning rate should be hyperparameter tuned. \n- They start to make claims about the test accuracy / generalization that I don’t think they should/need to include. Specifically, they make this claim about CIFAR10 on ResNet and ImageNet on Vision Transformers. In both scenarios, the models are overfitting, trained for hundreds of epochs on the training data. Lower training loss here does not necessarily mean higher test accuracy.\n    - Under Figure 1, they write “Prodigy estimates a larger step size than D-Adaptation, which helps it reach test accuracy closer to the one of Adam.” (They aren’t talking about  any implicit bias of large learning rate, just strictly that large learning rate leads to better training loss convergence.)\n    - Under paragraph “ViT training” they write “Prodigy almost closes the gap [of test performance] between tuned Adam and D-Adaptation.” \n\nSome other minor comments:\n- Can Figure 2 be plotted in log scale? It’s hard to see the difference between the lines. \n- Text spacing issue in Page 3", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699289431667}, {"id": "a1YvwVsm19", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7643/Reviewer_UUCM"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this submission, the authors improve on D-adaptation---a method for deterministic non-smooth convex optimization with promising performance in practice---by changing the step-sizes to be normalized not by the sum of the square-root of the norm squared of the gradients seen so far, but by the product of these norms squared with the current (lower) estimates of the distance of the initial point to the optimum. They also add extra parameters to allow for \"step-size schedules\", which do not affect the theoretical results but are important for the empirical performance. They show improved theoretical convergence rates (shaving off a square root log factor), and also show that a restarted version of D-adaptation also enjoy similar guarantees. Moreover, the authors also show a few lower-bounds for convex non-smooth optimization. Finally, they conclude by showing consistent improvement of Prodigy over D-adaptation over a large array of deep learning optimization tasks.", "review_text": "In this submission, the authors improve on D-adaptation---a method for deterministic non-smooth convex optimization with promising performance in practice---by changing the step-sizes to be normalized not by the sum of the square-root of the norm squared of the gradients seen so far, but by the product of these norms squared with the current (lower) estimates of the distance of the initial point to the optimum. They also add extra parameters to allow for \"step-size schedules\", which do not affect the theoretical results but are important for the empirical performance. They show improved theoretical convergence rates (shaving off a square root log factor), and also show that a restarted version of D-adaptation also enjoy similar guarantees. Moreover, the authors also show a few lower-bounds for convex non-smooth optimization. Finally, they conclude by showing consistent improvement of Prodigy over D-adaptation over a large array of deep learning optimization tasks.", "strengths": "The methods described do seem to improve over D-adaptation, either by changing the denominator in the step-sizes used or via resetting. The fact that these improve on the convergence guarantees from D-adaptation and show consistent improved performance on a variety of deep learning tasks is interesting.", "weaknesses": "The main weakness of the paper might boil down to presentation, but I am having a hard time correctly understanding many of the contributions. I am mostly knowledgeable on the theoretical aspects of optimization and online learning (although I am not closely acquainted with parameter-free methods in online learning). Maybe some of my worries are due to a lack of background from my part, At the same time, I consider myself a researcher with more knowledge in these fields than the average person in the community. So I do think these worries would be shared by many people that even work in topics related to this paper. I hope my questions and discussion with authors and reviewers help me arrive to a fair assessment of the paper. Of course, feel free to let me know if I am missing a major point of the paper or a big piece of the literature that was not mentioned in the prodigy paper due to space constraints. The main point is likely the first one, the other ones are minor points.\n\n**Prodigy convergence rates vs rates in more general settings**: Theorem 1 and 2 show that prodigy improved on the learning rate of D-adaptation by a factor of $\\sqrt{\\log(D/d_0)}$. It is not clear if this improvement should be expected to be reflected in empirical problems since in practice we see convergence rates that are much faster than $1/\\sqrt{t}$. In this case, if the rate does not reflect what happens empirically, I would guess that getting better rates is important for its novelty among other convergence rates. However, in the related work section the authors say that this new rate matches the one from online learning, which is a much more adversarial setting. In the stochastic case, which more closely matches the deep learning optimization case, the work of Carmon and Hinder (2022) have a $\\log \\log (D)$ factor in the convergence rate for the deterministic version of the algorithm if I am not mistaken, which is better than Prodigy's convergence rates. However, this is mostly a theoretical work, and the goal of Prodigy is to also be practical. On this line, the authors claim that they improve on the DoG convergence rates by shaving off a $\\sqrt{\\log D/d_0}$ factor. The DoG rates are for the stochastic setting with locally Lipschitz gradients, significantly more general than the deterministic setting where the rates of Prodigy hold.\n\nSo in the end I am not able to grasp what is the relevance of the improved theoretical convergence rates. It is in fact interesting that the improvement in the empirical performance might be connected to this slightly tighter convergence rate. But it is not clear how connected they are, if at all, and the paper does not discuss this connection. In the purely theoretical side, the rates at match (or slightly improve in the case of DoG) over other convergence rates, but on a more restricted setting.\n\n\n**Comparison with other algorithms**: Although the empirical results are definitely the strongest part of the paper, I did not understand why the only comparison point in the experiments is D-adaptation and DoG. Maybe this is discussed in a part of the appendix that I have not read (if so, please let me know), but is it the case that any other algorithms perform poorly enough that they are not worth considering in these comparisons?\n\n---\n\nHere are a few minor suggestions:\n\n**Lower bounds and the dependency of $n$ and $D$**: I thought the lower bounds were interesting, but the fact that the function $f$ (and, thus, $D$) depend on $n$ in such a way that $\\log \\log D/d_0$ is roughly of order $\\sqrt{n}$ seems very important and a big reason why these lower bounds do not rule out the possibility of improved asymptotic convergence rates, even in the stochastic case. Although this is mentioned by the authors before the lower bounds, adding that this is the case in the theorem statements themselves would make them more readily understandable by people. Mentioning that the function is of the form $f(x) = |x - x^*|$ could also be interesting, if space allows.\n\n\n\n**Typos and crammed equations**: As a last minor point, there are several equations that are crammed in the paper (sometimes with parts going on top of the text). I understand the submission process can be rushed, but if the authors could do a revision pass of the paper, it would be great.", "questions": "So here are my main question that I would appreciate if the authors could comment on.\n\n- Do you have some explanation of why/how the improved convergence rates should impact practical performance of the algorithm? In practice we see a convergence rate that is much faster than a $O(1/\\sqrt{n})$ rate, and at this point it is not clear why shaving off the $\\sqrt{\\log D / d_0}$ factor matters, even more so considering that these rates are only proven for the deterministic setting, while the empirical performance is studied in the stochastic case.\n- Could the authors expand on the relevance of the convergence rates given the comparisons I described with other works in more general settings? A big focus of the paper is put on the convergence rates, but the rates by themselves do not seem relevant if compared to other works, so I am probably misunderstanding something. Could the authors clarify this point? At this stage, it feels like this is a interesting contribution for deep learning optimization, but most of the paper was written trying to frame the contribution as a theoretical one.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this submission, the authors improve on D-adaptation---a method for deterministic non-smooth convex optimization with promising performance in practice---by changing the step-sizes to be normalized not by the sum of the square-root of the norm squared of the gradients seen so far, but by the product of these norms squared with the current (lower) estimates of the distance of the initial point to the optimum. They also add extra parameters to allow for \"step-size schedules\", which do not affect the theoretical results but are important for the empirical performance. They show improved theoretical convergence rates (shaving off a square root log factor), and also show that a restarted version of D-adaptation also enjoy similar guarantees. Moreover, the authors also show a few lower-bounds for convex non-smooth optimization. Finally, they conclude by showing consistent improvement of Prodigy over D-adaptation over a large array of deep learning optimization tasks.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The methods described do seem to improve over D-adaptation, either by changing the denominator in the step-sizes used or via resetting. The fact that these improve on the convergence guarantees from D-adaptation and show consistent improved performance on a variety of deep learning tasks is interesting.", "weaknesses": "The main weakness of the paper might boil down to presentation, but I am having a hard time correctly understanding many of the contributions. I am mostly knowledgeable on the theoretical aspects of optimization and online learning (although I am not closely acquainted with parameter-free methods in online learning). Maybe some of my worries are due to a lack of background from my part, At the same time, I consider myself a researcher with more knowledge in these fields than the average person in the community. So I do think these worries would be shared by many people that even work in topics related to this paper. I hope my questions and discussion with authors and reviewers help me arrive to a fair assessment of the paper. Of course, feel free to let me know if I am missing a major point of the paper or a big piece of the literature that was not mentioned in the prodigy paper due to space constraints. The main point is likely the first one, the other ones are minor points.\n\n**Prodigy convergence rates vs rates in more general settings**: Theorem 1 and 2 show that prodigy improved on the learning rate of D-adaptation by a factor of $\\sqrt{\\log(D/d_0)}$. It is not clear if this improvement should be expected to be reflected in empirical problems since in practice we see convergence rates that are much faster than $1/\\sqrt{t}$. In this case, if the rate does not reflect what happens empirically, I would guess that getting better rates is important for its novelty among other convergence rates. However, in the related work section the authors say that this new rate matches the one from online learning, which is a much more adversarial setting. In the stochastic case, which more closely matches the deep learning optimization case, the work of Carmon and Hinder (2022) have a $\\log \\log (D)$ factor in the convergence rate for the deterministic version of the algorithm if I am not mistaken, which is better than Prodigy's convergence rates. However, this is mostly a theoretical work, and the goal of Prodigy is to also be practical. On this line, the authors claim that they improve on the DoG convergence rates by shaving off a $\\sqrt{\\log D/d_0}$ factor. The DoG rates are for the stochastic setting with locally Lipschitz gradients, significantly more general than the deterministic setting where the rates of Prodigy hold.\n\nSo in the end I am not able to grasp what is the relevance of the improved theoretical convergence rates. It is in fact interesting that the improvement in the empirical performance might be connected to this slightly tighter convergence rate. But it is not clear how connected they are, if at all, and the paper does not discuss this connection. In the purely theoretical side, the rates at match (or slightly improve in the case of DoG) over other convergence rates, but on a more restricted setting.\n\n\n**Comparison with other algorithms**: Although the empirical results are definitely the strongest part of the paper, I did not understand why the only comparison point in the experiments is D-adaptation and DoG. Maybe this is discussed in a part of the appendix that I have not read (if so, please let me know), but is it the case that any other algorithms perform poorly enough that they are not worth considering in these comparisons?\n\n---\n\nHere are a few minor suggestions:\n\n**Lower bounds and the dependency of $n$ and $D$**: I thought the lower bounds were interesting, but the fact that the function $f$ (and, thus, $D$) depend on $n$ in such a way that $\\log \\log D/d_0$ is roughly of order $\\sqrt{n}$ seems very important and a big reason why these lower bounds do not rule out the possibility of improved asymptotic convergence rates, even in the stochastic case. Although this is mentioned by the authors before the lower bounds, adding that this is the case in the theorem statements themselves would make them more readily understandable by people. Mentioning that the function is of the form $f(x) = |x - x^*|$ could also be interesting, if space allows.\n\n\n\n**Typos and crammed equations**: As a last minor point, there are several equations that are crammed in the paper (sometimes with parts going on top of the text). I understand the submission process can be rushed, but if the authors could do a revision pass of the paper, it would be great.", "questions": "So here are my main question that I would appreciate if the authors could comment on.\n\n- Do you have some explanation of why/how the improved convergence rates should impact practical performance of the algorithm? In practice we see a convergence rate that is much faster than a $O(1/\\sqrt{n})$ rate, and at this point it is not clear why shaving off the $\\sqrt{\\log D / d_0}$ factor matters, even more so considering that these rates are only proven for the deterministic setting, while the empirical performance is studied in the stochastic case.\n- Could the authors expand on the relevance of the convergence rates given the comparisons I described with other works in more general settings? A big focus of the paper is put on the convergence rates, but the rates by themselves do not seem relevant if compared to other works, so I am probably misunderstanding something. Could the authors clarify this point? At this stage, it feels like this is a interesting contribution for deep learning optimization, but most of the paper was written trying to frame the contribution as a theoretical one.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698828226729}, {"id": "mzNegnIt5D", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7643/Reviewer_bqiE"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose two different improvements to the award-winning learning-rate-free D-adaptation algorithm, termed Prodigy (product of D and G), in turn, with GD, Dual Averaging and Adam versions, and D-adaptation with Resetting. \nBoth approaches improve the non-asymptotic bound for D-adaptation by removing a log factor, but seem to add other factors of their own.\nThe authors also prove a technical result that among exponentially bounded algorithms, a new characterization of algorithms, their D-adaptation variants are optimal. \nIn a series of convincing experiments, similar to the ones in the original D-adaptation paper, the Adam Prodigy variant is found to  be comparable (possibly slightly better in some cases) to the D-adaptation Adam variant.\nThe resetting approach is not experimentally evaluated as it isn't expected to outperform Prodigy.", "review_text": "The authors propose two different improvements to the award-winning learning-rate-free D-adaptation algorithm, termed Prodigy (product of D and G), in turn, with GD, Dual Averaging and Adam versions, and D-adaptation with Resetting. \nBoth approaches improve the non-asymptotic bound for D-adaptation by removing a log factor, but seem to add other factors of their own.\nThe authors also prove a technical result that among exponentially bounded algorithms, a new characterization of algorithms, their D-adaptation variants are optimal. \nIn a series of convincing experiments, similar to the ones in the original D-adaptation paper, the Adam Prodigy variant is found to  be comparable (possibly slightly better in some cases) to the D-adaptation Adam variant.\nThe resetting approach is not experimentally evaluated as it isn't expected to outperform Prodigy.", "strengths": "1. The paper is well-motivated and appears to be theoretically strong. (This reviewer didn't check the proofs though.)\n2. The experimental results confirm the theoretical guarantees for the convex logistic loss. \nThey also show that Prodigy and D-adaptation perform similarly, possibly slightly better in some cases, on small and large neural networks with non-convex losses despite the lack of theoretical guarantees.\n3. Experimental results in Fig. 1 seem to show the apparently new result that D-adaptation as well as Prodigy outperform the recently proposed DoG and L-DoG algorithms.", "weaknesses": "1. Both approaches appear to be more complex than the original D-adaptation approach, by introducing additional weights, \nand the practical benefit of the newer theoretical guarantees is not clear.\nAs mentioned, they remove one factor from the non-asymptotic bound of D-adaptation, but replace it with another.\nIt is not clear how much tuning was needed to obtain the small occasional improvements over D-adaptation in Figures 1-3.\n\n2. The paper is also marred by symbolic confusion and occasional grammatical errors, e.g., \n\n    2a. line 4 in Algorithm 2 mentions $\\lambda_k = d_k^2$, but the fifth line of Sec. 2 mentions $\\lambda_k = d_k$ \n\n    2b. Grammatical error or typo in Sec 5, page 7: \"...can divergence in theory...\"\n\n    2c. Grammatical error or typo in Sec 5, page 7: \"...initial sub-optimally of the D estimate ... \"", "questions": "Please note the inherent question underlying weakness 1. \nIt is mentioned below Theorem 1 regarding the careful setting of $\\lambda_k$, \"While it is not guaranteed to be better in theory, it is usually quite important to do so in practice.\" However, Sec 6 shows no great practical benefit from the apparent careful setting of $\\lambda_k$.\nThere is apparently some slight benefit in test losses for ViT for Resnet-50, but it is not clear how much tuning of $\\lambda_k$ was done to achieve this benefit.\nThe main promise of D-adaptation is that one does not need to perform much tuning of such optimization parameters.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose two different improvements to the award-winning learning-rate-free D-adaptation algorithm, termed Prodigy (product of D and G), in turn, with GD, Dual Averaging and Adam versions, and D-adaptation with Resetting. \nBoth approaches improve the non-asymptotic bound for D-adaptation by removing a log factor, but seem to add other factors of their own.\nThe authors also prove a technical result that among exponentially bounded algorithms, a new characterization of algorithms, their D-adaptation variants are optimal. \nIn a series of convincing experiments, similar to the ones in the original D-adaptation paper, the Adam Prodigy variant is found to  be comparable (possibly slightly better in some cases) to the D-adaptation Adam variant.\nThe resetting approach is not experimentally evaluated as it isn't expected to outperform Prodigy.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper is well-motivated and appears to be theoretically strong. (This reviewer didn't check the proofs though.)\n2. The experimental results confirm the theoretical guarantees for the convex logistic loss. \nThey also show that Prodigy and D-adaptation perform similarly, possibly slightly better in some cases, on small and large neural networks with non-convex losses despite the lack of theoretical guarantees.\n3. Experimental results in Fig. 1 seem to show the apparently new result that D-adaptation as well as Prodigy outperform the recently proposed DoG and L-DoG algorithms.", "weaknesses": "1. Both approaches appear to be more complex than the original D-adaptation approach, by introducing additional weights, \nand the practical benefit of the newer theoretical guarantees is not clear.\nAs mentioned, they remove one factor from the non-asymptotic bound of D-adaptation, but replace it with another.\nIt is not clear how much tuning was needed to obtain the small occasional improvements over D-adaptation in Figures 1-3.\n\n2. The paper is also marred by symbolic confusion and occasional grammatical errors, e.g., \n\n    2a. line 4 in Algorithm 2 mentions $\\lambda_k = d_k^2$, but the fifth line of Sec. 2 mentions $\\lambda_k = d_k$ \n\n    2b. Grammatical error or typo in Sec 5, page 7: \"...can divergence in theory...\"\n\n    2c. Grammatical error or typo in Sec 5, page 7: \"...initial sub-optimally of the D estimate ... \"", "questions": "Please note the inherent question underlying weakness 1. \nIt is mentioned below Theorem 1 regarding the careful setting of $\\lambda_k$, \"While it is not guaranteed to be better in theory, it is usually quite important to do so in practice.\" However, Sec 6 shows no great practical benefit from the apparent careful setting of $\\lambda_k$.\nThere is apparently some slight benefit in test losses for ViT for Resnet-50, but it is not clear how much tuning of $\\lambda_k$ was done to achieve this benefit.\nThe main promise of D-adaptation is that one does not need to perform much tuning of such optimization parameters.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698816286257}, {"id": "ru0d2ltT4q", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7643/Reviewer_gz9E"], "rating": "1: strong reject", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper develops an improvement to the D-Adaptation algorithm of Defazio & Mischenko (2023), improving the rate from $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}{\\sqrt{T}}\\right)$  to $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\sqrt{\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}}{\\sqrt{T}}\\right)$.", "review_text": "The paper develops an improvement to the D-Adaptation algorithm of Defazio & Mischenko (2023), improving the rate from $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}{\\sqrt{T}}\\right)$  to $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\sqrt{\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}}{\\sqrt{T}}\\right)$.", "strengths": "The paper is easy to read and mostly free of typos and grammatical errors.", "weaknesses": "## Summary\nOverall, I very strongly recommend rejection. The paper makes no new theoretical contributions and is seemingly unaware of large portions of the related literature. The results presented here have already been achieved more generally and in harder problem settings, and the lower bounds are both invalid. The experiments show some evidence of improvement over D-Adaptation, but makes no attempt to compare against the obvious existing baselines from the online learning literature. The main novelty is the improvement of the results from the D-Adaptation paper, which suffered from all of the same issues mentioned above, so I don't believe improving over this work warrants publication. Additional discussion on these points is provided below.\n\n## The Algorithm\nAlgorithms that attain the $\\\\|x_0-x^\\*\\\\|\\sqrt{\\log(\\\\|x_0-x^\\*\\\\|/d_0)}/\\sqrt{T}$ rate (or equivalently $\\\\|x_0-x^\\*\\\\|\\sqrt{T\\log(\\\\|x_0-x^\\*\\\\|/d_0)}$ regret) under Lipschitz losses have existed for almost a decade now, and they accomplish it in *strictly* harder problems (see e.g. McMahan & Orabona 2014, Orabona & Pal 2016, Cutkosky & Orabona 2018). In particular, there are countless algorithms from the  so-called \"parameter-free\" online learning literature which already achieve this result in the significantly harder adversarial online learning setting, in which the $\\\\|x_0-x^*\\\\|\\sqrt{T\\log(\\\\|x_0-x^\\*\\\\|/d_0)}$ regret is un-improvable. The methods presented here achieve this rate in the *easiest possible problem setting*: optimizing a *fixed* function, where this rate isn't even optimal. Non-trivial extensions have even been achieved in these harder problem settings, such as as scale-free learning (Mhammedi & Koolen 2020), dynamic regret (Jacobsen & Cutkosky 2022), learning with switching costs (Zhang & Cutkosky 2022b), and many more. Note that these results have been achieved using a variety of approaches: coin-betting, FTRL, mirror descent, and general potential-based approaches, while the discussion in Section 5 seems to only be aware of the coin-betting approach.\n\nThe paper also claims to improve on methods other than D-Adaptation, such as the T-DoG algorithm of Ivgi et al (2023). This is factually incorrect: the problem setting studied in this paper is strictly easier than all other comparable works in this field (besides D-Adaptation), so Prodigy actually has *no* guarantee in the problem setting studied by Ivgi et al. 2023. In this sense Prodigy is actually a *strict downgrade* of T-DoG, and likewise of any of the other existing works that can solve the problem studied here.\n\nSection 5 also implies that standard D-adaptation (and Prodigy by extension) actually improve over the existing online learning works, as \"Standard D-Adaptation obtains asymptotic rates without the log factor\". This is at least misleading: their asymptotic result holds under the condition that the **user** chooses $d_0\\le \\\\|x_0-x^\\*\\\\|$, which would only be possible to guarantee if you have prior knowledge of $\\\\|x_0-x^\\*\\\\|$. If you had this prior knowledge, you wouldn't need any special algorithm to achieve the $O(G\\\\|x_0-x^\\*\\\\|/\\sqrt{T})$ rate, you could accomplish this using gradient descent with step-sizes $\\eta_t = \\frac{\\\\|x_0-x^\\*\\\\|}{\\sqrt{G^2+\\sum_s^{t-1}\\\\|g_s\\\\|^2}}$. One might argue that you can just set $d_0$ to be very close to 0, but the smaller you set $d_0$, the longer it takes for the asymptotic result to kick in, making it easy to wind up with a result that holds only for some $\\tilde T$ such that $G\\\\|x_0-x^\\*\\\\|\\sqrt{\\tilde T}\\ge G\\\\|x_0-x^\\*\\\\|\\sqrt{T\\log(\\\\|x^\\*-x_0\\\\|/d_0)}$, making the result again redundant. So there is no meaningful improvement via the asymptotic result either.\n\nAn Adam-based variant is also proposed, but this leads to something of a contradiction: Adam itself isn't guaranteed to converge (Reddi et al. 2019), so Prodigy+Adam isn't guaranteed to converge either. How can we claim an algorithm as \"parameter-free\" if it makes zero performance guarantees and may diverge? The whole point is that the algorithm makes some performance *guarantee* without tuning hyperparameters. Moreover, it's unclear to me how interesting the global step-size of Adam even is --- the momentum parameters have far more impact on the behavior of Adam. The experiments claim that Prodigy+Adam performs roughly as well as hand-tuned Adam, but I would not be surprised if Adam with the default global step-size *also* performed similarly to hand-tuned Adam.\n\n## The Lower bounds\nNeither of the lower bounds are valid constructions. The lower bounds choose $D$ to be either $2^{2^T}x_1$ or $2^Tx_1$, neither of which are valid ways to construct the stated lower bound, because now the bound holds for only a specific class of comparators, rather than for all comparators simultaneously. In other words, the lower bounds should show that *for any* $D,G$, there is an $x^\\*$ such that $\\\\|x_0-x^\\*\\\\|= D$ and the stated lower bound holds. This is also what the theorem statement *implies* is happening, until you actually read the proof, which is misleading.\n\n## The Experiments\nThe experiments demonstrate some improvement over D-Adaptation, but are not particularly convincing aside from that. Notably, the experiments include *no* baselines from the many existing works from online learning. Given that there are several existing algorithms that already achieve stronger results than presented here, this paper should at the *very least* be justifying its existence by showing some improvement over these existing works. Yet not a single one is included as a baseline.\n\nThe experiments use other tricks on top of Prodigy such as a warm-up epoch, step-size annealing, and weight-decay, so I'm again unsure how the algorithm can be claimed to be an \"Expeditiously Adaptive Parameter-Free Learner\" --- these all involve some form of hyperparameter selection. I also don't understand why weight decay is necessary; adding an L2 penalty to the loss will implicitly constrain the algorithm to a ball of some radius, which shouldn't even be necessary for an algorithm that already adapts to $\\\\|x_0-x^\\*\\\\|$. The fact that this needed to be included suggests that the algorithm in fact does *not* attain this form of adaptivity in general, as one would hope to demonstrate in the experiments.\n\nThe Large-scale Adam experiments of Section 6.1 show evidence that the performance of Prodigy can be similar to that of hand-tuned Adam. However, as mentioned earlier, this is not particularly convincing on its own because Adam's global step-size has a relatively benign impact on performance. These experiments should at least include \"un-tuned Adam\" using the default step-size, which I suspect will also perform similarly to Prodigy.\n\n## References\n- Cutkosky, Ashok, and Francesco Orabona. \"Black-box reductions for parameter-free online learning in banach spaces.\" Conference On Learning Theory. PMLR, 2018.\n- Mhammedi, Zakaria, and Wouter M. Koolen. \"Lipschitz and comparator-norm adaptivity in online learning.\" Conference on Learning Theory. PMLR, 2020.\n- Jacobsen, Andrew, and Ashok Cutkosky. \"Parameter-free mirror descent.\" Conference on Learning Theory. PMLR, 2022.\n- Orabona, Francesco, and Dávid Pál. \"Parameter-free stochastic optimization of variationally coherent functions.\" arXiv preprint arXiv:2102.00236 (2021).\n- Orabona, Francesco. \"A modern introduction to online learning.\" arXiv preprint arXiv:1912.13213 (2019).\n- Orabona, Francesco, and Dávid Pál. \"Coin betting and parameter-free online learning.\" Advances in Neural Information Processing Systems 29 (2016).\n- McMahan, H. Brendan, and Francesco Orabona. \"Unconstrained online linear learning in hilbert spaces: Minimax algorithms and normal approximations.\" Conference on Learning Theory. PMLR, 2014.\n- McMahan, Brendan, and Matthew Streeter. \"No-regret algorithms for unconstrained online convex optimization.\" Advances in neural information processing systems 25 (2012).\n- Reddi, Sashank J., Satyen Kale, and Sanjiv Kumar. \"On the convergence of adam and beyond.\" arXiv preprint arXiv:1904.09237 (2019).\n- Zhang, Jiujia, and Ashok Cutkosky. \"Parameter-free regret in high probability with heavy tails.\" Advances in Neural Information Processing Systems 35 (2022a).\n- Zhang, Zhiyu, Ashok Cutkosky, and Yannis Paschalidis. \"Optimal Comparator Adaptive Online Learning with Switching Cost.\" Advances in Neural Information Processing Systems 35 (2022b)", "questions": "- What new contributions does this work make that haven't already been addressed in the online learning literature? \n- Why do the experiments include no baselines from the relevant online learning literature?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper develops an improvement to the D-Adaptation algorithm of Defazio & Mischenko (2023), improving the rate from $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}{\\sqrt{T}}\\right)$  to $O\\left(\\frac{\\\\|x_0-x^\\*\\\\|\\sqrt{\\log\\left(\\\\|x_0-x^\\*\\\\|/d_0\\right)}}{\\sqrt{T}}\\right)$.", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "strengths": "The paper is easy to read and mostly free of typos and grammatical errors.", "weaknesses": "## Summary\nOverall, I very strongly recommend rejection. The paper makes no new theoretical contributions and is seemingly unaware of large portions of the related literature. The results presented here have already been achieved more generally and in harder problem settings, and the lower bounds are both invalid. The experiments show some evidence of improvement over D-Adaptation, but makes no attempt to compare against the obvious existing baselines from the online learning literature. The main novelty is the improvement of the results from the D-Adaptation paper, which suffered from all of the same issues mentioned above, so I don't believe improving over this work warrants publication. Additional discussion on these points is provided below.\n\n## The Algorithm\nAlgorithms that attain the $\\\\|x_0-x^\\*\\\\|\\sqrt{\\log(\\\\|x_0-x^\\*\\\\|/d_0)}/\\sqrt{T}$ rate (or equivalently $\\\\|x_0-x^\\*\\\\|\\sqrt{T\\log(\\\\|x_0-x^\\*\\\\|/d_0)}$ regret) under Lipschitz losses have existed for almost a decade now, and they accomplish it in *strictly* harder problems (see e.g. McMahan & Orabona 2014, Orabona & Pal 2016, Cutkosky & Orabona 2018). In particular, there are countless algorithms from the  so-called \"parameter-free\" online learning literature which already achieve this result in the significantly harder adversarial online learning setting, in which the $\\\\|x_0-x^*\\\\|\\sqrt{T\\log(\\\\|x_0-x^\\*\\\\|/d_0)}$ regret is un-improvable. The methods presented here achieve this rate in the *easiest possible problem setting*: optimizing a *fixed* function, where this rate isn't even optimal. Non-trivial extensions have even been achieved in these harder problem settings, such as as scale-free learning (Mhammedi & Koolen 2020), dynamic regret (Jacobsen & Cutkosky 2022), learning with switching costs (Zhang & Cutkosky 2022b), and many more. Note that these results have been achieved using a variety of approaches: coin-betting, FTRL, mirror descent, and general potential-based approaches, while the discussion in Section 5 seems to only be aware of the coin-betting approach.\n\nThe paper also claims to improve on methods other than D-Adaptation, such as the T-DoG algorithm of Ivgi et al (2023). This is factually incorrect: the problem setting studied in this paper is strictly easier than all other comparable works in this field (besides D-Adaptation), so Prodigy actually has *no* guarantee in the problem setting studied by Ivgi et al. 2023. In this sense Prodigy is actually a *strict downgrade* of T-DoG, and likewise of any of the other existing works that can solve the problem studied here.\n\nSection 5 also implies that standard D-adaptation (and Prodigy by extension) actually improve over the existing online learning works, as \"Standard D-Adaptation obtains asymptotic rates without the log factor\". This is at least misleading: their asymptotic result holds under the condition that the **user** chooses $d_0\\le \\\\|x_0-x^\\*\\\\|$, which would only be possible to guarantee if you have prior knowledge of $\\\\|x_0-x^\\*\\\\|$. If you had this prior knowledge, you wouldn't need any special algorithm to achieve the $O(G\\\\|x_0-x^\\*\\\\|/\\sqrt{T})$ rate, you could accomplish this using gradient descent with step-sizes $\\eta_t = \\frac{\\\\|x_0-x^\\*\\\\|}{\\sqrt{G^2+\\sum_s^{t-1}\\\\|g_s\\\\|^2}}$. One might argue that you can just set $d_0$ to be very close to 0, but the smaller you set $d_0$, the longer it takes for the asymptotic result to kick in, making it easy to wind up with a result that holds only for some $\\tilde T$ such that $G\\\\|x_0-x^\\*\\\\|\\sqrt{\\tilde T}\\ge G\\\\|x_0-x^\\*\\\\|\\sqrt{T\\log(\\\\|x^\\*-x_0\\\\|/d_0)}$, making the result again redundant. So there is no meaningful improvement via the asymptotic result either.\n\nAn Adam-based variant is also proposed, but this leads to something of a contradiction: Adam itself isn't guaranteed to converge (Reddi et al. 2019), so Prodigy+Adam isn't guaranteed to converge either. How can we claim an algorithm as \"parameter-free\" if it makes zero performance guarantees and may diverge? The whole point is that the algorithm makes some performance *guarantee* without tuning hyperparameters. Moreover, it's unclear to me how interesting the global step-size of Adam even is --- the momentum parameters have far more impact on the behavior of Adam. The experiments claim that Prodigy+Adam performs roughly as well as hand-tuned Adam, but I would not be surprised if Adam with the default global step-size *also* performed similarly to hand-tuned Adam.\n\n## The Lower bounds\nNeither of the lower bounds are valid constructions. The lower bounds choose $D$ to be either $2^{2^T}x_1$ or $2^Tx_1$, neither of which are valid ways to construct the stated lower bound, because now the bound holds for only a specific class of comparators, rather than for all comparators simultaneously. In other words, the lower bounds should show that *for any* $D,G$, there is an $x^\\*$ such that $\\\\|x_0-x^\\*\\\\|= D$ and the stated lower bound holds. This is also what the theorem statement *implies* is happening, until you actually read the proof, which is misleading.\n\n## The Experiments\nThe experiments demonstrate some improvement over D-Adaptation, but are not particularly convincing aside from that. Notably, the experiments include *no* baselines from the many existing works from online learning. Given that there are several existing algorithms that already achieve stronger results than presented here, this paper should at the *very least* be justifying its existence by showing some improvement over these existing works. Yet not a single one is included as a baseline.\n\nThe experiments use other tricks on top of Prodigy such as a warm-up epoch, step-size annealing, and weight-decay, so I'm again unsure how the algorithm can be claimed to be an \"Expeditiously Adaptive Parameter-Free Learner\" --- these all involve some form of hyperparameter selection. I also don't understand why weight decay is necessary; adding an L2 penalty to the loss will implicitly constrain the algorithm to a ball of some radius, which shouldn't even be necessary for an algorithm that already adapts to $\\\\|x_0-x^\\*\\\\|$. The fact that this needed to be included suggests that the algorithm in fact does *not* attain this form of adaptivity in general, as one would hope to demonstrate in the experiments.\n\nThe Large-scale Adam experiments of Section 6.1 show evidence that the performance of Prodigy can be similar to that of hand-tuned Adam. However, as mentioned earlier, this is not particularly convincing on its own because Adam's global step-size has a relatively benign impact on performance. These experiments should at least include \"un-tuned Adam\" using the default step-size, which I suspect will also perform similarly to Prodigy.\n\n## References\n- Cutkosky, Ashok, and Francesco Orabona. \"Black-box reductions for parameter-free online learning in banach spaces.\" Conference On Learning Theory. PMLR, 2018.\n- Mhammedi, Zakaria, and Wouter M. Koolen. \"Lipschitz and comparator-norm adaptivity in online learning.\" Conference on Learning Theory. PMLR, 2020.\n- Jacobsen, Andrew, and Ashok Cutkosky. \"Parameter-free mirror descent.\" Conference on Learning Theory. PMLR, 2022.\n- Orabona, Francesco, and Dávid Pál. \"Parameter-free stochastic optimization of variationally coherent functions.\" arXiv preprint arXiv:2102.00236 (2021).\n- Orabona, Francesco. \"A modern introduction to online learning.\" arXiv preprint arXiv:1912.13213 (2019).\n- Orabona, Francesco, and Dávid Pál. \"Coin betting and parameter-free online learning.\" Advances in Neural Information Processing Systems 29 (2016).\n- McMahan, H. Brendan, and Francesco Orabona. \"Unconstrained online linear learning in hilbert spaces: Minimax algorithms and normal approximations.\" Conference on Learning Theory. PMLR, 2014.\n- McMahan, Brendan, and Matthew Streeter. \"No-regret algorithms for unconstrained online convex optimization.\" Advances in neural information processing systems 25 (2012).\n- Reddi, Sashank J., Satyen Kale, and Sanjiv Kumar. \"On the convergence of adam and beyond.\" arXiv preprint arXiv:1904.09237 (2019).\n- Zhang, Jiujia, and Ashok Cutkosky. \"Parameter-free regret in high probability with heavy tails.\" Advances in Neural Information Processing Systems 35 (2022a).\n- Zhang, Zhiyu, Ashok Cutkosky, and Yannis Paschalidis. \"Optimal Comparator Adaptive Online Learning with Switching Cost.\" Advances in Neural Information Processing Systems 35 (2022b)", "questions": "- What new contributions does this work make that haven't already been addressed in the online learning literature? \n- Why do the experiments include no baselines from the relevant online learning literature?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698612060486}], "openreview_url": "https://openreview.net/forum?id=WpQbM1kBuy", "arxiv_id": "2306.06101", "paper_pdf": "papers/WpQbM1kBuy.pdf", "paper_pdf_sha256": "0420c6fcd569f0a2c5eae1140832d961d449631b829204fb587fb32532947186", "paper_pdf_bytes": 2518028, "paper_pdf_source": "openreview", "code_url": "https://github.com/konstmish/prodigy", "code_repository": "konstmish/prodigy", "code_commit": "3efb213ee8af5a6bf76f28726398433a847b38e9", "code_archive": "repos/WpQbM1kBuy.zip", "code_archive_sha256": "3da66d295cfc3628cdd98ada112ecce4f333ce3a90cfa9f6ec892be111a5a0e0", "code_archive_bytes": 12537, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 47, "github_languages": {"Python": 21063}, "github_archived": false, "github_pushed_at": "2025-01-16T14:57:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/prodigy-an-expeditiously-adaptive-parameter"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UFKW7EVrJAm", "year": 2023, "status": "rejected", "title": "Generating Adversarial Examples with Task Oriented Multi-Objective Optimization", "authors": ["Anh Tuan Bui", "Trung Le", "He Zhao", "Quan Hung Tran", "Paul Montague", "Dinh Phung"], "authorids": ["~Anh_Tuan_Bui2", "~Trung_Le2", "~He_Zhao1", "~Quan_Hung_Tran1", "~Paul_Montague1", "~Dinh_Phung2"], "authors_source": "OpenReview API", "abstract": "Deep learning models, even the-state-of-the-art ones, are highly vulnerable to adversarial examples. Adversarial training is one of the most efficient methods to improve the model's robustness. The key factor for the success of adversarial training is the capability to generate qualified and divergent adversarial examples which satisfy some objectives/goals (e.g., finding adversarial examples that maximize the model losses for simultaneously attacking multiple models). Therefore, multi-objective optimization (MOO) is a natural tool for adversarial example generation, where we search adversarial examples simultaneously maximizing some objectives/goals. However, we observe that a naive application of MOO tends to maximize all objectives/goals equally, without caring if an objective/goal has been achieved yet. This leads to useless effort to further improve the goal-achieved tasks, while putting less focus on the goal-unachieved tasks. In this paper, we propose \\emph{Task Oriented MOO} to address this issue, in the context where we can explicitly define the goal achievement for a task. Our principle is to only maintain the goal-achieved tasks, while letting the optimizer spend more effort on improving the goal-unachieved tasks. We conduct comprehensive experiments for our Task Oriented MOO on various adversarial example generation schemes. The experimental results firmly demonstrate the merit of our proposed approach.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "DQ4xwTEgg7", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2011/Reviewer_aC9i"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a method for generating adversarial examples for satisfying multiple goals based on multi-objective optimisation (MOO). The paper can be classified into the family of adversarial attack algorithms. A multi-gradient decent algorithm is utilised to solve the optimisation problem by integrating all gradients w.r.t. different objectives into a single gradient direction, but the optimisation process of MOO always tends to treat all tasks equally. Therefore, a regularisation term is proposed to emphasise unachieved tasks while putting less weight on already satisfying tasks. The paper then gives four experiments, including adversarial attack and training over multiple models, universal perturbation, and adversarial attack over data transformations, to validate the utility of the proposed approach.", "review_text": "See Section Strength And Weaknesses\n", "strengths": "Strength:\n\n1. The proposed method is basically considering a generating attacks scenario where multiple domain losses are needed to be optimised together simultaneously. The insight behind the technique is to balance the weights of all tasks, and a regularizer is added to push greater weights on unsatisfied tasks.\n\n2. The paper is well-written, and the main idea is easy to follow.\n\n3. Comprehensive experiments show that the proposed method is better than the uniform case and minimax case.\n\n\nMain concerns:\n\n1. The novelty of such formulations is not clear to the reviewer. The main contribution claimed by the authors is that this is the first work that regards adversarial attacks as a multi-objective optimisation problem. But, there are some works in the literature which takes this formulation into consideration in adversarial example generation scenario (especially, e.g., [1], [2]).\n\n2. This problem and considered scenarios are similar in spirit to [2]. The proposed regularizer added to the optimisation objective function is a fairly basic and straightforward trick to weigh different domain losses. It cannot be considered a sufficient contribution for an ICLR paper.\n\n3. From the perspective of multi-objective optimisation theory, any analysis in terms of convergence and Pareto-optimal frontier is essential.\nReference \n\n[1] Qiu, H., Du, Y. and Lu, T., 2022. The Framework of Cross-Domain and Model Adversarial Attack against Deepfake. Future Internet, 14(2), p.46.\n\n[2] Wang, J., Zhang, T., Liu, S., Chen, P.Y., Xu, J., Fardad, M. and Li, B., 2021. Adversarial attack generation empowered by min-max optimization. Advances in Neural Information Processing Systems, 34, pp.16020-16033.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors propose a method for generating adversarial examples for satisfying multiple goals based on multi-objective optimisation (MOO). The paper can be classified into the family of adversarial attack algorithms. A multi-gradient decent algorithm is utilised to solve the optimisation problem by integrating all gradients w.r.t. different objectives into a single gradient direction, but the optimisation process of MOO always tends to treat all tasks equally. Therefore, a regularisation term is proposed to emphasise unachieved tasks while putting less weight on already satisfying tasks. The paper then gives four experiments, including adversarial attack and training over multiple models, universal perturbation, and adversarial attack over data transformations, to validate the utility of the proposed approach.", "strength_and_weaknesses": "Strength:\n\n1. The proposed method is basically considering a generating attacks scenario where multiple domain losses are needed to be optimised together simultaneously. The insight behind the technique is to balance the weights of all tasks, and a regularizer is added to push greater weights on unsatisfied tasks.\n\n2. The paper is well-written, and the main idea is easy to follow.\n\n3. Comprehensive experiments show that the proposed method is better than the uniform case and minimax case.\n\n\nMain concerns:\n\n1. The novelty of such formulations is not clear to the reviewer. The main contribution claimed by the authors is that this is the first work that regards adversarial attacks as a multi-objective optimisation problem. But, there are some works in the literature which takes this formulation into consideration in adversarial example generation scenario (especially, e.g., [1], [2]).\n\n2. This problem and considered scenarios are similar in spirit to [2]. The proposed regularizer added to the optimisation objective function is a fairly basic and straightforward trick to weigh different domain losses. It cannot be considered a sufficient contribution for an ICLR paper.\n\n3. From the perspective of multi-objective optimisation theory, any analysis in terms of convergence and Pareto-optimal frontier is essential.\nReference \n\n[1] Qiu, H., Du, Y. and Lu, T., 2022. The Framework of Cross-Domain and Model Adversarial Attack against Deepfake. Future Internet, 14(2), p.46.\n\n[2] Wang, J., Zhang, T., Liu, S., Chen, P.Y., Xu, J., Fardad, M. and Li, B., 2021. Adversarial attack generation empowered by min-max optimization. Advances in Neural Information Processing Systems, 34, pp.16020-16033.\n", "clarity,_quality,_novelty_and_reproducibility": "See Section Strength And Weaknesses\n", "summary_of_the_review": "See Section Strength And Weaknesses\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667084811940}, {"id": "Eqg7nruNt9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2011/Reviewer_Fq8c"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to use multi-objective optimization to generate adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation. Specifically, by setting the different models in the ensemble as different objectives, it turns the problem into multi-objective optimization. To further improve the performance, a regularizer is then proposed to suppress the successful attack and encourage the unsuccessful attack. Extensive experiments have been done in several cases including generating adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation to show the effectiveness of the proposed method. The experiment results show the proposed method could achieve a better performance compared to several baselines.", "review_text": "The paper proposes to use multi-objective optimization to generate adversarial examples in several settings including model ensemble, universal perturbation, etc.  The experiment shows the proposed method could achieve good results in several settings. However,  as all the settings use the max as its ensemble, it is not clear whether the proposed method could be used in other ensemble cases.", "strengths": "Pros:\n1. The paper is well-written and easy to follow.\n2. The experiment results show that the proposed method could achieve a good result in generating adversarial examples in several settings.\n\n\nCons:\n1. To my understanding, the proposed framework could only be applied into the ensemble that uses the max, min, and majority vote as its aggregation rule. It is not clear if the proposed method still works if the ensemble is taking the average of every model's output.\n2. The proposed settings are kind of similar to each other. A better setting will be generating adversarial examples with the different norms and proposing a defense to defend against several norms attacking at the same time. \n3. I am not quite sure if the novelty is strong or not since the proposed method just applies multi-objective optimization. Also, I am not sure whether the proposed regularizer is already proposed in the multi-objective optimization before or not.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes to use multi-objective optimization to generate adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation. Specifically, by setting the different models in the ensemble as different objectives, it turns the problem into multi-objective optimization. To further improve the performance, a regularizer is then proposed to suppress the successful attack and encourage the unsuccessful attack. Extensive experiments have been done in several cases including generating adversarial examples for model ensembles, adversarial examples against several transformations, and universal adversarial perturbation to show the effectiveness of the proposed method. The experiment results show the proposed method could achieve a better performance compared to several baselines.", "strength_and_weaknesses": "Pros:\n1. The paper is well-written and easy to follow.\n2. The experiment results show that the proposed method could achieve a good result in generating adversarial examples in several settings.\n\n\nCons:\n1. To my understanding, the proposed framework could only be applied into the ensemble that uses the max, min, and majority vote as its aggregation rule. It is not clear if the proposed method still works if the ensemble is taking the average of every model's output.\n2. The proposed settings are kind of similar to each other. A better setting will be generating adversarial examples with the different norms and proposing a defense to defend against several norms attacking at the same time. \n3. I am not quite sure if the novelty is strong or not since the proposed method just applies multi-objective optimization. Also, I am not sure whether the proposed regularizer is already proposed in the multi-objective optimization before or not.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is well-written and easy to follow.\n\nQuality: The experiment is extensive and detailed. \n\nNovelty: The novelty is limited if the proposed regularizer is already proposed in the multi-objective optimization before because the proposed method just applies multi-objective optimization to generating adversarial examples.\n\nReproducibility: Although the code is not provided, all hyperparameters are listed and I believe the proposed method should be reproducible.", "summary_of_the_review": "The paper proposes to use multi-objective optimization to generate adversarial examples in several settings including model ensemble, universal perturbation, etc.  The experiment shows the proposed method could achieve good results in several settings. However,  as all the settings use the max as its ensemble, it is not clear whether the proposed method could be used in other ensemble cases.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666800945257}, {"id": "PEiFAGevlMe", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2011/Reviewer_H7Re"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The other suggest to extend the multi-objective optimization (MOO) method. Their new proposed method is called Task Oriented MOO. They claim that naive MOO invest useless effort in trying to maximize already achieved goals, their method let the optimizer spend more effort on improving the goal-unachieved tasks. They formalize the adversarial generation task as a multi-objective optimization problem and apply their TA-MOO approach to the problem. The authors show comprehensive experiments in which their method is better than the current baseline.", "review_text": "The paper is original and improves the baselines. The paper is written in high quality fashion.", "strengths": "### Strength  \nThe paper is written in a clear and easy to follow fashion. The results seems to improve the baseline and the MOO method by a nice margin most of the time. The method and formulation, therefore, make sense to the adversarial generation task. The authors are the first to view the adversarial generation task as a MOO problem. The discussions and comprehensive experiment are clear and supporting the understanding of the method.\n\n### Weaknesses\nThe paper did not fully explained the evaluation metrics. Why does the A-All is the most important metric for this task?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The other suggest to extend the multi-objective optimization (MOO) method. Their new proposed method is called Task Oriented MOO. They claim that naive MOO invest useless effort in trying to maximize already achieved goals, their method let the optimizer spend more effort on improving the goal-unachieved tasks. They formalize the adversarial generation task as a multi-objective optimization problem and apply their TA-MOO approach to the problem. The authors show comprehensive experiments in which their method is better than the current baseline.", "strength_and_weaknesses": "### Strength  \nThe paper is written in a clear and easy to follow fashion. The results seems to improve the baseline and the MOO method by a nice margin most of the time. The method and formulation, therefore, make sense to the adversarial generation task. The authors are the first to view the adversarial generation task as a MOO problem. The discussions and comprehensive experiment are clear and supporting the understanding of the method.\n\n### Weaknesses\nThe paper did not fully explained the evaluation metrics. Why does the A-All is the most important metric for this task?", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and very informative to the reader and community. The paper show originality by casting the adversarial generation task as a MOO problem and applying MOO algorithms to the problem. The paper present promising results improving the baseline with a large margin most of the time.", "summary_of_the_review": "The paper is original and improves the baselines. The paper is written in high quality fashion.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "no ethic conerncs", "recommendation": "8: accept, good paper"}, "tcdate": 1666261838217}, {"id": "ApBB-wIU1hd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2011/Reviewer_g196"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a multi-objective optimization approach capable of dealing with objectives that converge at different rates. The authors then claim that three problems within the realm of generating adversarial examples can be cast as multi-objective optimization and apply the proposed Task Aware Multi-objective optimization (TA-MOO) algorithm for these problems. The main rationale behind TA-MOO is to augment the optimization problem to find a common descent direction within the standard multi gradient descent (MGD) with a regularization term that enforces the weights of the common descent direction to be larger for objectives with smaller losses (in the case of a problem of maximization of objectives). Experiments on two datasets, namely CIFAR-10 and CIFAR-100, considering the tasks of generating i- adversarial examples for an ensemble model, ii- universal perturbations, and iii- adversarial examples against transformations showed that TA-MOO outperformed the compared approaches in most of the cases.   ", "review_text": "The main contributions of this work are casting problems related to generating adversarial perturbations as multi-objective optimization and proposing a gradient-based multi-objective algorithm to solve the aforementioned problems. In the current version of the manuscript, the the major limitations of the work are lack of motivation for the use of multi-objective optimization (i.e. why using multi-objective optimization for this problem?), and limited scope of experiments both in terms of baselines and aspects considered in the analysis. In my review, I raised questions and provided suggestions that could address some of the limitations. At this point, I believe this contribution is not yet ready to be considered for publication.", "strengths": "Strengths:\n- The authors proposed a general multi-objective approach that showed promising results on three different (but related) problems related to adversarial examples generation;\n\n- The proposed algorithm, TA-MOO, has a mechanism to enforce the solutions of the multi-objective problem to focus on the central regions of the Pareto front, which is a desired feature in the case of the studied applications;\n\n- The experiments on generation of adversarial perturbations are broad in the sense that they account for different aspects of the considered problems. \n\n\nWeaknesses:\n\n- Lack of motivation for the contribution: one of the main claims of this work is “multi-objective optimization is a natural tool for adversarial example generation”, however, other than the fact that there are multiple objectives to be simultaneously satisfied, I am not able see why this is the case since there is no clear evidence that there is a tradeoff between the objectives when generating adversarial examples;\n\n- Limited evaluation: the main contribution of this work is proposing a variation of a gradient-based multi-objective optimization algorithm capable of dealing with objectives that converge at different rates. However, it is not clear to me whether the choice of problems for evaluation is in fact suffering from the issue TA-MOO is proposed to solve. Moreover, the experiments in the main paper do not analyze important aspects of the proposed algorithm, such as if it is able to reach Pareto stationary solutions in practical applications;  \n \n- Lack of strong baselines: I believe using unconstrained MGD (referred to in the manuscript as MOO) is not a fair choice of baseline in a case where it is known a priori that the desired solutions are in the central region of the Pareto front since MGD has no mechanism whatsoever to incorporate user preferences to guide the optimization. In the following section of the review I suggest stronger baselines to be considered. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes a multi-objective optimization approach capable of dealing with objectives that converge at different rates. The authors then claim that three problems within the realm of generating adversarial examples can be cast as multi-objective optimization and apply the proposed Task Aware Multi-objective optimization (TA-MOO) algorithm for these problems. The main rationale behind TA-MOO is to augment the optimization problem to find a common descent direction within the standard multi gradient descent (MGD) with a regularization term that enforces the weights of the common descent direction to be larger for objectives with smaller losses (in the case of a problem of maximization of objectives). Experiments on two datasets, namely CIFAR-10 and CIFAR-100, considering the tasks of generating i- adversarial examples for an ensemble model, ii- universal perturbations, and iii- adversarial examples against transformations showed that TA-MOO outperformed the compared approaches in most of the cases.   ", "strength_and_weaknesses": "Strengths:\n- The authors proposed a general multi-objective approach that showed promising results on three different (but related) problems related to adversarial examples generation;\n\n- The proposed algorithm, TA-MOO, has a mechanism to enforce the solutions of the multi-objective problem to focus on the central regions of the Pareto front, which is a desired feature in the case of the studied applications;\n\n- The experiments on generation of adversarial perturbations are broad in the sense that they account for different aspects of the considered problems. \n\n\nWeaknesses:\n\n- Lack of motivation for the contribution: one of the main claims of this work is “multi-objective optimization is a natural tool for adversarial example generation”, however, other than the fact that there are multiple objectives to be simultaneously satisfied, I am not able see why this is the case since there is no clear evidence that there is a tradeoff between the objectives when generating adversarial examples;\n\n- Limited evaluation: the main contribution of this work is proposing a variation of a gradient-based multi-objective optimization algorithm capable of dealing with objectives that converge at different rates. However, it is not clear to me whether the choice of problems for evaluation is in fact suffering from the issue TA-MOO is proposed to solve. Moreover, the experiments in the main paper do not analyze important aspects of the proposed algorithm, such as if it is able to reach Pareto stationary solutions in practical applications;  \n \n- Lack of strong baselines: I believe using unconstrained MGD (referred to in the manuscript as MOO) is not a fair choice of baseline in a case where it is known a priori that the desired solutions are in the central region of the Pareto front since MGD has no mechanism whatsoever to incorporate user preferences to guide the optimization. In the following section of the review I suggest stronger baselines to be considered. \n", "clarity,_quality,_novelty_and_reproducibility": "- The authors claim that in some cases multi-objective optimization does not work. However, it is not clear to me if this is indeed the case. In more details, the purpose of (unconstrained) multi-objective optimization is to find a point within the Pareto front, and, perhaps in the case of the reported results MOO was able to find point in a region of the Pareto front that does not yield equally good performance across all objectives, which means it is actually “working”. I would like to learn more about the authors’ opinion on this aspect and suggest a more detailed discussion is added to the manuscript.\n\n- Importantly, it is still unclear to me whether there are indeed conflicts between the objectives in the considered test cases within the paper. Notably, previous work showed evidence that adversarial attacks are transferable across different models and architectures [1, 2]. Therefore, I believe that in such cases the conflict between gradients should be negligible. Moreover, since there are no theoretical guarantees that TA-MOO finds Pareto stationary solutions, empirical evidence needs to be provided to illustrate why the proposed algorithm works. As an example, the authors can report quantities such as the norm of the common descent direction, since for Pareto stationary points this quantity should be close to zero. Finally, as I mentioned in my review, I strongly believe the authors should include in the empirical evaluation further gradient-based multi-objective approaches that aim at finding solutions in the central region of the Pareto front.\n\n- Since in the case of MOO / MGD there is no constraint to enforce the solutions to reach the desired preferred region of the Pareto front, could it be the case that a particular initialization of MOO would suffice to make it find the desired solutions? Also, could a previously proposed constrained version of MGD such as [3] be able to do so and potentially achieve a performance as good as TA-MOO?\n\n- Recent work [4, 5, 6] on multi-task learning has shown that well-tuned, simple baselines such as scalarizing the objectives with fixed weights equal to 1 random weights, yields solutions as good as the ones found by some multi-objective approaches, including MGDA, the base method for TA-MOO. Given that, and the fact that it seems that the tasks considered in this submission are similar to the considered multi-task settings mentioned in [4, 5, 6], I wonder if similar conclusions would be drawn in case extensive tuning of baselines were to be performed.\n\n- As mentioned in the previous section of the review, the manuscript lacks in the choice of multi-objective optimization baselines. For example, I believe multi-objective approaches that aim at finding solutions lying in the central region of the Pareto front [7, 8] should be included, as well as methods that have been shown to prevent the bias issue in MGD [9].\n\n[1] Liu et al. \"Delving into Transferable Adversarial Examples and Black-box Attacks\", 2016. \\\n[2] Che et al. \"A New Ensemble Adversarial Attack Powered by Long-term Gradient Memories\", 2019. \\\n[3] Lin et al. \"Pareto multi-task learning\", 2019. \\\n[4] Kurin et al. “In defense of the unitary scalarization for deep multi-task learning”, 2022. \\\n[5] Xin et al. \"Do Current Multi-Task Optimization Methods in Deep Learning Even Help?”, 2022. \\\n[6] Lin et al. “Reasonable Effectiveness of Random Weighting: A Litmus Test for Multi-Task Learning”, 2022. \\\n[7] Albuquerque et al. \"Multi-objective training of generative adversarial networks with multiple discriminators\", 2019. \\\n[8] Miranda et al. ”Single-solution hypervolume maximization and its use for improving generalization of neural networks”, 2016. \\\n[9] Navon et al. “Multi-Task Learning as a Bargaining Game”, 2022.\n", "summary_of_the_review": "The main contributions of this work are casting problems related to generating adversarial perturbations as multi-objective optimization and proposing a gradient-based multi-objective algorithm to solve the aforementioned problems. In the current version of the manuscript, the the major limitations of the work are lack of motivation for the use of multi-objective optimization (i.e. why using multi-objective optimization for this problem?), and limited scope of experiments both in terms of baselines and aspects considered in the analysis. In my review, I raised questions and provided suggestions that could address some of the limitations. At this point, I believe this contribution is not yet ready to be considered for publication.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666259727743}], "openreview_url": "https://openreview.net/forum?id=UFKW7EVrJAm", "arxiv_id": "2304.13229", "paper_pdf": "papers/UFKW7EVrJAm.pdf", "paper_pdf_sha256": "5f6bf05e4a5abac0e7f338d71b33698729b3abdf69a59c81b9d5644620ffce0c", "paper_pdf_bytes": 4862642, "paper_pdf_source": "openreview", "code_url": "https://github.com/tuananhbui89/TAMOO", "code_repository": "tuananhbui89/TAMOO", "code_commit": "1485ca64e879f9bf333957a4176929eacadc99ee", "code_archive": "repos/UFKW7EVrJAm.zip", "code_archive_sha256": "ac08f33003ebdc5ba0d8e5fb67f2df1ad78af5748185af2930780c0196d0b605", "code_archive_bytes": 99592, "code_file_count": 47, "code_extensions": {".py": 38, ".sh": 9}, "github_disk_usage_kb": 79, "github_languages": {"Python": 363104, "Shell": 46110}, "github_archived": false, "github_pushed_at": "2023-06-01T19:12:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generating-adversarial-examples-with-task"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B7O85qTDgU4", "year": 2022, "status": "rejected", "title": "Meta-Learning Dynamics Forecasting Using Task Inference", "authors": ["Rui Wang", "Robin Walters", "Rose Yu"], "authorids": ["~Rui_Wang11", "~Robin_Walters1", "~Rose_Yu1"], "authors_source": "OpenReview API", "abstract": "Current deep learning models for dynamics forecasting struggle with generalization. They can only forecast in a specific domain and fail when applied to systems with different parameters, external forces, or boundary conditions.  We propose a model-based meta-learning method called DyAd which can generalize across heterogeneous domains by partitioning them into different tasks.  DyAd has two parts: an encoder which infers the time-invariant hidden features of the task with weak supervision, and a forecaster which learns the shared dynamics of the entire domain. The encoder adapts and controls the forecaster during inference using adaptive instance normalization and adaptive padding.  Theoretically, we prove that the generalization error of such procedure is related to the task relatedness in the source domain, as well as the domain differences between source and target. Experimentally, we demonstrate that our model outperforms state-of-the-art approaches on both turbulent flow and real-world ocean data forecasting tasks.     ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "S-z0UjvO0Vk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2064/Reviewer_cEq7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work tackles the task of forecasting dynamics in different domains simultaneously. Using an encoder which is trained to determine the task, the inferred latent vector is then used to adapt a forecasting network to the task at hand. Experiments on three datasets linked to fluid dynamics are then conducted to assess the proposed model.", "review_text": "Pros :\n\n- This is an interesting problem which is quite timely given the development of the field of forecasting physical dynamics using neural networks. \n\n- The proposed solution seems sound and principled. Moreover, it is well motivated and the writing was quite clear.\n\n- The different additions made to the forecaster network are also quite interesting, I especially liked the AdaPad solution to deal with boundary conditions. Conducting an ablation study also considerably strengthens the paper. \n\nCons :\n\n- All experiments are conducted on somewhat similar datasets, which are based on fluid dynamics PDEs. It would be nice to see how the model deals with other families of dynamics. Especially given the fact that the contributions of this work seem geared towards practical considerations.\n\n- The setting of the experiments should be more precise and additional details should be given: how are the different datasets constructed, what supervision is there exactly regarding the different tasks, how many domains are there in each dataset and what are the differences, how is the balance between the different domains ect.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work tackles the task of forecasting dynamics in different domains simultaneously. Using an encoder which is trained to determine the task, the inferred latent vector is then used to adapt a forecasting network to the task at hand. Experiments on three datasets linked to fluid dynamics are then conducted to assess the proposed model.", "main_review": "Pros :\n\n- This is an interesting problem which is quite timely given the development of the field of forecasting physical dynamics using neural networks. \n\n- The proposed solution seems sound and principled. Moreover, it is well motivated and the writing was quite clear.\n\n- The different additions made to the forecaster network are also quite interesting, I especially liked the AdaPad solution to deal with boundary conditions. Conducting an ablation study also considerably strengthens the paper. \n\nCons :\n\n- All experiments are conducted on somewhat similar datasets, which are based on fluid dynamics PDEs. It would be nice to see how the model deals with other families of dynamics. Especially given the fact that the contributions of this work seem geared towards practical considerations.\n\n- The setting of the experiments should be more precise and additional details should be given: how are the different datasets constructed, what supervision is there exactly regarding the different tasks, how many domains are there in each dataset and what are the differences, how is the balance between the different domains ect.", "summary_of_the_review": "This is a good work on a timely subject. The contribution is not groundbreaking but should be significant enough to warrant acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635878899181}, {"id": "4lPHv-aEFVF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2064/Reviewer_SrGv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper addresses the problem of learning a deep learning model for dynamics forecasting which generalizes to changes in dynamics. These changes can be induced by different parameters, boundary conditions or external forces. The proposed model takes a meta-learning approach and proposes to partition data into different heterogeneous domains. It consists of two components: an encoder which infers time-invariant features given observed domain data and a forecaster which predicts the dynamics given these features. The paper evaluates the proposed approach on several datasets and provides some theoretical insights.", "review_text": "+\n\n* This paper addresses a new and interesting generalization problem for dynamics forecasting \n* It proposes a model to address different changes in the dynamics. \n* Evaluation is done on relevant datasets with several baselines and some ablation studies.\n\n-\n\n* The applicability of the proposed approach is restricted to problems where relevant weak supervision from task parameters is available. This seems like an important limitation in real-world applications. How valid is this scenario? The question of choosing relevant parameters for weak supervision is important for applying this model to other datasets, yet the definition of these parameters is unclear; how robust is the model when chosen parameters are not useful ? The performance of Wrong_enc (Table 2) tends to say that this model will then fail. \n* It is unclear why the model can adapt to changing boundary conditions with AdaPad as it generates them from features $\\hat{z}_c$ extracted from data inside the domain and weakly supervised by quantities unrelated to the boundary condition (e.g. mean vorticity or season). \n* The theoretical analysis, inspired by existing work in multi-task learning / domain adaptation, has some limitations and does not add much value to the paper. I have some concerns with the domain adaptation upper-bound to the target error in Theorem 3.4 and Proposition 3.5. This upper-bound is not minimized thus the target risk can be high i.e. the model is not guaranteed to adapt well. Moreover, the validity of the theoretical analysis is unclear as several assumptions may not be verified e.g. bounded loss in Theorem 3.1, Proposition 3.3; lipschitz continuity in Proposition 3.5. Theorem 3.4 requires that the assumptions in Theorem 2 in Redko et al 2017 are verified, yet these assumptions are not mentioned in the paper.\n* Some ablation studies are missing: 1) the contribution of each term in equation (2) and 2) the dimensionality of $\\hat{z}_c$ which is fixed arbitrarily.\n\nOther questions: \n* It would be good to better explain how the experiments include changing boundary conditions between domains. The testing scenarios only mention different initial conditions or external forces.\n* Why do the baselines ResNet-c and Unet-c not adapt well despite having access to relevant weak supervision (p8)? This is the same information used by the proposed model to adapt.\n* How redundant is the time invariance term (3rd term in equation (2)) with the invariances enforced in the architecture of the encoder?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper addresses the problem of learning a deep learning model for dynamics forecasting which generalizes to changes in dynamics. These changes can be induced by different parameters, boundary conditions or external forces. The proposed model takes a meta-learning approach and proposes to partition data into different heterogeneous domains. It consists of two components: an encoder which infers time-invariant features given observed domain data and a forecaster which predicts the dynamics given these features. The paper evaluates the proposed approach on several datasets and provides some theoretical insights.", "main_review": "+\n\n* This paper addresses a new and interesting generalization problem for dynamics forecasting \n* It proposes a model to address different changes in the dynamics. \n* Evaluation is done on relevant datasets with several baselines and some ablation studies.\n\n-\n\n* The applicability of the proposed approach is restricted to problems where relevant weak supervision from task parameters is available. This seems like an important limitation in real-world applications. How valid is this scenario? The question of choosing relevant parameters for weak supervision is important for applying this model to other datasets, yet the definition of these parameters is unclear; how robust is the model when chosen parameters are not useful ? The performance of Wrong_enc (Table 2) tends to say that this model will then fail. \n* It is unclear why the model can adapt to changing boundary conditions with AdaPad as it generates them from features $\\hat{z}_c$ extracted from data inside the domain and weakly supervised by quantities unrelated to the boundary condition (e.g. mean vorticity or season). \n* The theoretical analysis, inspired by existing work in multi-task learning / domain adaptation, has some limitations and does not add much value to the paper. I have some concerns with the domain adaptation upper-bound to the target error in Theorem 3.4 and Proposition 3.5. This upper-bound is not minimized thus the target risk can be high i.e. the model is not guaranteed to adapt well. Moreover, the validity of the theoretical analysis is unclear as several assumptions may not be verified e.g. bounded loss in Theorem 3.1, Proposition 3.3; lipschitz continuity in Proposition 3.5. Theorem 3.4 requires that the assumptions in Theorem 2 in Redko et al 2017 are verified, yet these assumptions are not mentioned in the paper.\n* Some ablation studies are missing: 1) the contribution of each term in equation (2) and 2) the dimensionality of $\\hat{z}_c$ which is fixed arbitrarily.\n\nOther questions: \n* It would be good to better explain how the experiments include changing boundary conditions between domains. The testing scenarios only mention different initial conditions or external forces.\n* Why do the baselines ResNet-c and Unet-c not adapt well despite having access to relevant weak supervision (p8)? This is the same information used by the proposed model to adapt.\n* How redundant is the time invariance term (3rd term in equation (2)) with the invariances enforced in the architecture of the encoder?\n", "summary_of_the_review": "This paper tackles a new generalization problem for dynamics forecasting and proposes a model supported by experimental results. However, this model can only be applied to problems with relevant weak supervision which may not always be available in practise. Moreover, the definition of relevant parameters is unclear and the robustness of the model to the choice of these parameters is not measured which may restrict its application to other datasets. There are also unclarities on the ability of the model to adapt to changing boundary conditions with AdaPad, some ablation studies are missing and I have concerns on the theoretical analysis which brings limited value to the paper. For this reason, I am giving this paper a weak reject.\n\n--- Post-Rebuttal comments ---\nI thank the authors for their response. After studying it, the theoretical results still have some major issues and feel disconnected from the model. In particular, key assumptions are not enforced in the model (e.g. lipschitz continuity) and the generalization error of the model in Th3.3 is uncontrolled as the upper-bound is not minimized by the model (the Wasserstein distance between domains is fixed and is high in all generality). Its use for the model is thus not very convincing. On practical aspects, the capability of handling boundary conditions should be better justified and evaluated. For this reason, I keep my score unchanged and recommend rejecting this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635876390304}, {"id": "FirkqBw1CnL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2064/Reviewer_4kg2"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper suggest a remediation for a common problem for dynamics forecasting which is the lack of generalization to other domains/tasks. The author suggest to tackle this with via a 2 component architecture, one for learning the task and one for forecasting. In empiricial experiments the authors show the practical feasibility of their approach.", "review_text": "As a caveat: I'm not an expert in the area, so my review remains on a superficial level consequently for which I apologize. I overall liked the paper quite a bit, the question discussed is relevant, the empirical evaluation is very good, the theoretical results seem as relevant as they would get and the related work discussed is crisply presented and relevant. \n\nOne question I would have is that results in Table 1 are overwhelmingly good with only UNET-c coming close. Do we know for these tasks what the \"theoretical\" upper bound (e.g. by the right PDE system) would be? Is it computationally even possible to compute this upper bound? I'm wondering how much of a gap there still is too close. \n\nIn a similar vein, what is the intuition behind DyAD + ResNet being better than DyAD + UNET mostly? Are there some complementary strengths between DyAD and ResNet that this combination can exploit better than DyAD + UNET?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper suggest a remediation for a common problem for dynamics forecasting which is the lack of generalization to other domains/tasks. The author suggest to tackle this with via a 2 component architecture, one for learning the task and one for forecasting. In empiricial experiments the authors show the practical feasibility of their approach.", "main_review": "As a caveat: I'm not an expert in the area, so my review remains on a superficial level consequently for which I apologize. I overall liked the paper quite a bit, the question discussed is relevant, the empirical evaluation is very good, the theoretical results seem as relevant as they would get and the related work discussed is crisply presented and relevant. \n\nOne question I would have is that results in Table 1 are overwhelmingly good with only UNET-c coming close. Do we know for these tasks what the \"theoretical\" upper bound (e.g. by the right PDE system) would be? Is it computationally even possible to compute this upper bound? I'm wondering how much of a gap there still is too close. \n\nIn a similar vein, what is the intuition behind DyAD + ResNet being better than DyAD + UNET mostly? Are there some complementary strengths between DyAD and ResNet that this combination can exploit better than DyAD + UNET?", "summary_of_the_review": "This is a good paper that I'd like to see accepted for its combination of theoretical results, empirical results and methodological novelty.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635808225870}, {"id": "H34jsXYS5IO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2064/Reviewer_nhax"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper is interested in learning general forecasting models for physical dynamical processes. The paper proposes a decomposition of such a model into an encoder that captures the innate properties of the system, and a forecaster that autoregressively makes predictions conditioned on the encoded properties. This is framed as a meta-learning approach, and is shown to substantially outperform single-task approaches and off-the-shell meta-learning approaches across multiple datasets. The paper provides some theoretical analysis, and qualitative analysis of what is learned. Overall, the paper shows that learning shared models across domains is an important and fruitful way forward for modeling physical processes with machine learning.", "review_text": "Strengths:\n- The problem statement is well-motivated. Learning generalizable deep learning models across diverse settings is an important open problem.\n- Experiments use interesting and real-world problems.\n- Results are strong and appear reliable.\n- AdaPad is an interesting idea specialized to the case of physical complex systems, since it is designed to address boundary condition issues.\n- Visualizations show the model is behaving essentially as expected.\n- Although there are many design choices that go in to the model, each such design choice is well-motivated.\n- Aside from some aspects of the theory section, the exposition is generally quite clear and well-organized.\n- Assumptions are made clear.\n- The fact that the encoder can be trained first and independently of the forecaster should be very useful for further rapid developments.\n- Great to see ESE metric used as a complement to raw error.\n- Table in Appendix showing alternatives to AdaIn is very useful in increasing confidence in AdaIn for this application.\n\n\nWeaknesses:\n- The biggest concern is the theory section. The multi-task learning and domain adaptation results are general results that are not adequately connected back to the specific model and problem the paper is considering. Yes, it is widely accepted that multi-task learning and domain adaptation can work well, especially when tasks are related in some measurable way, and it can be a useful exercise to restate existing theory in the language of your framework, but what (if any) novel claims is the theory implying? Are there any predictions the theory makes about the particular approach which can be validated in experiments?\n- The theoretical bound on error that decomposes the error of the encoder and forecaster is similarly lacking in its interpretation. Yes, it can be a useful exercise to show that the error can be decomposed along the lines of the model, but does this bound somehow suggest that the decomposition results in lower error than a monolithic model? Or is it showing that you can work independently on improving either part of the model and improve the overall error? Where is there potential for practical value in this theorem?\n- For example, one place there could be potential to validate the theory is to check in experiments that task pairs with lower Wasserstein distance actually support better domain adaptation. However, in the Introduction of the paper it acknowledges that “Even the slightest change in these features may lead to vastly different phenomena”, but doesn’t that suggest that Wasserstein distance may not be a useful metric here for measuring task similarity? Couldn't turbulence limit the usefulness of such a metric?\n- Proposition 3.3 says the bound is “strictly looser” than the bound in Theorem 3.1. For clarity, it would be very helpful to combine the bounds into an inequality showing this strictly-looser property. It is not immediately apparent from the statement of the theorems since the inequalities contain different terms.\n- As is, the theory doesn’t really hurt the paper, but, for the amount of space dedicated to it, it doesn’t add much. The paper could be substantially improved by either (1) adding interpretation/predictions/validation of the theory that connect it back to the approach in the paper, or (2) removing some of the less useful parts of the theory from the main paper to free up space for more of the interesting analysis of what the model actually learns.\n- Also, it is interesting but a bit counter-intuitive that the theory section relies on results in multi-task learning and domain adaptation, instead of theoretical results from the meta-learning literature. As is, since the paper relies on multi-task learning so much, it is missing references to related work in multi-task learning (i.e., related work outside of modeling physical dynamical systems).\n- Similarly, it would be helpful to mention why there are no comparisons to multi-task learning or domain adaptation methods in the experiments. Why do they not apply here?\n- The three terms in the loss function of the encoder are well-motivated, but it is not clear how important each term is. Ablations on these terms would be very informative for the reader to understand what’s generally required to train an encoder.\n- In Section 5 it says “VarSepNet employs separation of variables through different loss terms”. What are these loss terms and how are they different from the ones in the paper?\n- In the ablations with no encoder, how do AdaIn and AdaPad work? Don’t they require some z? Where does this come from if not from the encoder?\n- U-Net does seem it could be at a qualitative disadvantage compared to DyAd in terms on number of parameters, especially since U-Net c is one of the more competitive baselines. It would be useful to see results for a larger U-Net c, or at least some evidence that the U-Net is not underfitting the training data.\n\n\nAdditional question of interest:\n\nOverall, this is a very important a potentially deep line of research. The most exciting promise of such work is the potential of revealing shared regularities across vastly disparate dynamic systems, that is, across complex physical processes. And it seems the approach in the paper could be particularly well-suited to such research. For example, the authors could train a single encoder+forecaster model across all the datasets in the paper, and analyze relationships in the learned encodings across datasets. Training models across highly diverse domains have been tried in multi-task learning (e.g., \"Pretrained Transformers as Universal Computation Engines\" arxiv 2021, \"The Traveling Observer Model\" ICLR 2021, \"Modular Universal Reparameterization\" NeurIPS 2019, \"One Model to Learn Them All\" arxiv 2017). Is such a generalization part of the longer term vision for this line of work?\n\n\nMinor comments:\n- In Section 2.4, some references would be useful in the sentence ending with “…the combined force equation.”\n- There are several inconsistencies in the use of parentheses in citations throughout the paper. Correcting these would improve readability.\n- In last sentence of first paragraph of Section 4, the word “task” could be changed to something like “problem”, since “task” has another meaning in the paper.\n- Should the 7.26 for U-Net-c on Ocean Currents future be bolded?\n- In the last paragraph of Section 5.1: “We tried to vary…” -> “We tried varying…” or “We varied…”.\n- Appendix A.2.1: footnote for PhiFlow is on the wrong page.\n- Appendix A.2.1: The last paragraph seems like it should be the first paragraph of A.2.2.\n- In proof of Proposition B.5, there is an extra or missing set of norm bars in the first inequality.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper is interested in learning general forecasting models for physical dynamical processes. The paper proposes a decomposition of such a model into an encoder that captures the innate properties of the system, and a forecaster that autoregressively makes predictions conditioned on the encoded properties. This is framed as a meta-learning approach, and is shown to substantially outperform single-task approaches and off-the-shell meta-learning approaches across multiple datasets. The paper provides some theoretical analysis, and qualitative analysis of what is learned. Overall, the paper shows that learning shared models across domains is an important and fruitful way forward for modeling physical processes with machine learning.", "main_review": "Strengths:\n- The problem statement is well-motivated. Learning generalizable deep learning models across diverse settings is an important open problem.\n- Experiments use interesting and real-world problems.\n- Results are strong and appear reliable.\n- AdaPad is an interesting idea specialized to the case of physical complex systems, since it is designed to address boundary condition issues.\n- Visualizations show the model is behaving essentially as expected.\n- Although there are many design choices that go in to the model, each such design choice is well-motivated.\n- Aside from some aspects of the theory section, the exposition is generally quite clear and well-organized.\n- Assumptions are made clear.\n- The fact that the encoder can be trained first and independently of the forecaster should be very useful for further rapid developments.\n- Great to see ESE metric used as a complement to raw error.\n- Table in Appendix showing alternatives to AdaIn is very useful in increasing confidence in AdaIn for this application.\n\n\nWeaknesses:\n- The biggest concern is the theory section. The multi-task learning and domain adaptation results are general results that are not adequately connected back to the specific model and problem the paper is considering. Yes, it is widely accepted that multi-task learning and domain adaptation can work well, especially when tasks are related in some measurable way, and it can be a useful exercise to restate existing theory in the language of your framework, but what (if any) novel claims is the theory implying? Are there any predictions the theory makes about the particular approach which can be validated in experiments?\n- The theoretical bound on error that decomposes the error of the encoder and forecaster is similarly lacking in its interpretation. Yes, it can be a useful exercise to show that the error can be decomposed along the lines of the model, but does this bound somehow suggest that the decomposition results in lower error than a monolithic model? Or is it showing that you can work independently on improving either part of the model and improve the overall error? Where is there potential for practical value in this theorem?\n- For example, one place there could be potential to validate the theory is to check in experiments that task pairs with lower Wasserstein distance actually support better domain adaptation. However, in the Introduction of the paper it acknowledges that “Even the slightest change in these features may lead to vastly different phenomena”, but doesn’t that suggest that Wasserstein distance may not be a useful metric here for measuring task similarity? Couldn't turbulence limit the usefulness of such a metric?\n- Proposition 3.3 says the bound is “strictly looser” than the bound in Theorem 3.1. For clarity, it would be very helpful to combine the bounds into an inequality showing this strictly-looser property. It is not immediately apparent from the statement of the theorems since the inequalities contain different terms.\n- As is, the theory doesn’t really hurt the paper, but, for the amount of space dedicated to it, it doesn’t add much. The paper could be substantially improved by either (1) adding interpretation/predictions/validation of the theory that connect it back to the approach in the paper, or (2) removing some of the less useful parts of the theory from the main paper to free up space for more of the interesting analysis of what the model actually learns.\n- Also, it is interesting but a bit counter-intuitive that the theory section relies on results in multi-task learning and domain adaptation, instead of theoretical results from the meta-learning literature. As is, since the paper relies on multi-task learning so much, it is missing references to related work in multi-task learning (i.e., related work outside of modeling physical dynamical systems).\n- Similarly, it would be helpful to mention why there are no comparisons to multi-task learning or domain adaptation methods in the experiments. Why do they not apply here?\n- The three terms in the loss function of the encoder are well-motivated, but it is not clear how important each term is. Ablations on these terms would be very informative for the reader to understand what’s generally required to train an encoder.\n- In Section 5 it says “VarSepNet employs separation of variables through different loss terms”. What are these loss terms and how are they different from the ones in the paper?\n- In the ablations with no encoder, how do AdaIn and AdaPad work? Don’t they require some z? Where does this come from if not from the encoder?\n- U-Net does seem it could be at a qualitative disadvantage compared to DyAd in terms on number of parameters, especially since U-Net c is one of the more competitive baselines. It would be useful to see results for a larger U-Net c, or at least some evidence that the U-Net is not underfitting the training data.\n\n\nAdditional question of interest:\n\nOverall, this is a very important a potentially deep line of research. The most exciting promise of such work is the potential of revealing shared regularities across vastly disparate dynamic systems, that is, across complex physical processes. And it seems the approach in the paper could be particularly well-suited to such research. For example, the authors could train a single encoder+forecaster model across all the datasets in the paper, and analyze relationships in the learned encodings across datasets. Training models across highly diverse domains have been tried in multi-task learning (e.g., \"Pretrained Transformers as Universal Computation Engines\" arxiv 2021, \"The Traveling Observer Model\" ICLR 2021, \"Modular Universal Reparameterization\" NeurIPS 2019, \"One Model to Learn Them All\" arxiv 2017). Is such a generalization part of the longer term vision for this line of work?\n\n\nMinor comments:\n- In Section 2.4, some references would be useful in the sentence ending with “…the combined force equation.”\n- There are several inconsistencies in the use of parentheses in citations throughout the paper. Correcting these would improve readability.\n- In last sentence of first paragraph of Section 4, the word “task” could be changed to something like “problem”, since “task” has another meaning in the paper.\n- Should the 7.26 for U-Net-c on Ocean Currents future be bolded?\n- In the last paragraph of Section 5.1: “We tried to vary…” -> “We tried varying…” or “We varied…”.\n- Appendix A.2.1: footnote for PhiFlow is on the wrong page.\n- Appendix A.2.1: The last paragraph seems like it should be the first paragraph of A.2.2.\n- In proof of Proposition B.5, there is an extra or missing set of norm bars in the first inequality.", "summary_of_the_review": "Overall, this is very interesting and useful work. The problem is well-motivated, and the approach and experiments are carefully designed and generally convincing. If the concerns about the theory are addressed, I would be happy to increase my score. Adding the additional info and experiments requested could increase it further, and make this a particularly strong paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635544770201}], "openreview_url": "https://openreview.net/forum?id=B7O85qTDgU4", "arxiv_id": "2102.10271", "paper_pdf": "papers/B7O85qTDgU4.pdf", "paper_pdf_sha256": "8cdf1f89d50b36d2aaaa75ca6dd95d35ab4283ac0a33d8cf78fb8685595dccd2", "paper_pdf_bytes": 3999225, "paper_pdf_source": "openreview", "code_url": "https://github.com/Rose-STL-Lab/Dynamic-Adaptation-Network", "code_repository": "Rose-STL-Lab/Dynamic-Adaptation-Network", "code_commit": "2afd6e09db1cce65102a7b8d55c24668cd220ab6", "code_archive": "repos/B7O85qTDgU4.zip", "code_archive_sha256": "fe5fefec5fac4d63cf0cb23c16f4edadccb4512cce852011cbd1dc63671e27ae", "code_archive_bytes": 49545, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 78, "github_languages": {"Python": 149813}, "github_archived": false, "github_pushed_at": "2023-10-30T14:05:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/meta-learning-dynamics-forecasting-using-task"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Ip195saXqIX", "year": 2021, "status": "rejected", "title": "Knowledge Distillation By Sparse Representation Matching", "authors": ["Dat Thanh Tran", "Moncef Gabbouj", "Alexandros Iosifidis"], "authorids": ["~Dat_Thanh_Tran1", "~Moncef_Gabbouj1", "~Alexandros_Iosifidis2"], "authors_source": "OpenReview API", "abstract": "Knowledge Distillation refers to a class of methods that transfers the knowledge from a teacher network to a student network. In this paper, we propose Sparse Representation Matching (SRM), a method to transfer intermediate knowledge obtained from one Convolutional Neural Network (CNN) to another by utilizing sparse representation learning. SRM first extracts sparse representations of the hidden features of the teacher CNN, which are then used to generate both pixel-level and image-level labels for training intermediate feature maps of the student network. We formulate SRM as a neural processing block, which can be efficiently optimized using stochastic gradient descent and integrated into any CNN in a plug-and-play manner. Our experiments demonstrate that SRM is robust to architectural differences between the teacher and student networks, and outperforms other KD techniques across several datasets. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "EkDilRQvn3r", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1536/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Paper summary**\nThis paper proposes a knowledge distillation on the feature maps using sparse representation. The proposed method firstly constructs an over-complement dictionary to express the teacher's feature maps and learn sparse representation to express the teacher's feature map using the dictionary. Since directly utilizing the sparse representation is a too strong restriction for the student network, the loss function is designed to find the indices of sparse codes. The proposed distillation method is validated through several experiments. \n\n**Pros**\n1. This paper proposes a way of utilizing sparse representation for knowledge distillation.\n2. The algorithm is written in clear formulations. \n\n**Cons**\n1. The main idea of sparse representation matching (SRM) is the combination of sparse representation and knowledge distillation. However, the actual implementation of SRM does not transfer the sparse representation of the teacher to the student. Only the indices or the entire image's sparse code were transferred via knowledge distillation, so the feature map information of the teacher is not transferred. This looks counter-intuitive and weakens the arguments of the paper. Therefore, it is necessary to describe the information transferred by SRM in detail and the rationale for using this kind of information transferring.\n2. All experiments were conducted on All-CNN, but this network is not usually used by other knowledge distillation papers, so the experiments should be re-conducted on a more standard and efficient setting. For example, All-CNN got 74.7% accuracy using SRM with 2.2M parameters (in table 1), but a recent paper [1] (CRD) got 75.5% accuracy using WRN16-2 with 0.7M parameters, which uses only 1/3 number of parameters. In other words, All-CNN is not proper to compare with other distillation methods. Use popular network architectures (WRN, ResNet, VGG etc.) to get more reasonable performance. \n[1] Contrastive Representation Distillation (ICLR 2020)\n3. In the case of transfer learning, it is common to tune the learning rates and weight decays for each dataset and each network. However, the experiments in the paper consistently use the same learning hyper-parameters. As a result, the performance gap between the baseline and the proposed method seems to have greatly inflated. A great gap against baseline (about 10%~20%) seems to be very different from the results of the other distillation papers.  In short, to be fair comparisons with other distillation methods, learning parameter tuning is necessary. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of \"Knowledge Distillation By Sparse Representation Matching\"", "review": "**Paper summary**\nThis paper proposes a knowledge distillation on the feature maps using sparse representation. The proposed method firstly constructs an over-complement dictionary to express the teacher's feature maps and learn sparse representation to express the teacher's feature map using the dictionary. Since directly utilizing the sparse representation is a too strong restriction for the student network, the loss function is designed to find the indices of sparse codes. The proposed distillation method is validated through several experiments. \n\n**Pros**\n1. This paper proposes a way of utilizing sparse representation for knowledge distillation.\n2. The algorithm is written in clear formulations. \n\n**Cons**\n1. The main idea of sparse representation matching (SRM) is the combination of sparse representation and knowledge distillation. However, the actual implementation of SRM does not transfer the sparse representation of the teacher to the student. Only the indices or the entire image's sparse code were transferred via knowledge distillation, so the feature map information of the teacher is not transferred. This looks counter-intuitive and weakens the arguments of the paper. Therefore, it is necessary to describe the information transferred by SRM in detail and the rationale for using this kind of information transferring.\n2. All experiments were conducted on All-CNN, but this network is not usually used by other knowledge distillation papers, so the experiments should be re-conducted on a more standard and efficient setting. For example, All-CNN got 74.7% accuracy using SRM with 2.2M parameters (in table 1), but a recent paper [1] (CRD) got 75.5% accuracy using WRN16-2 with 0.7M parameters, which uses only 1/3 number of parameters. In other words, All-CNN is not proper to compare with other distillation methods. Use popular network architectures (WRN, ResNet, VGG etc.) to get more reasonable performance. \n[1] Contrastive Representation Distillation (ICLR 2020)\n3. In the case of transfer learning, it is common to tune the learning rates and weight decays for each dataset and each network. However, the experiments in the paper consistently use the same learning hyper-parameters. As a result, the performance gap between the baseline and the proposed method seems to have greatly inflated. A great gap against baseline (about 10%~20%) seems to be very different from the results of the other distillation papers.  In short, to be fair comparisons with other distillation methods, learning parameter tuning is necessary. \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603977598560}, {"id": "YDrIDnDu_N2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1536/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Strength and weaknesses:\n+\nThe method is reasonable and competitive. Especially for the task transfer scenario is seems to advance the state of the art\n\n-\nComparison is lacking as a main relevant competitor was not tested\nI believe the method can be much simplified by instead of using sparse decomposition using plain clustering of pixel columns. Experiments testing this option should be conducted\nThe task transfer experiments, whose results are the most impressive, are done without weight Initialization of the student networks. It is not clear why, and the results may be different if initialized networks are used. \nAt the bottom line: cleaner experiments (adding comparison to Jain et al, using plain clustering instead of sparse decomposition, using student networks with ImageNet initialized weights) are required to take this paper beyond reasonable doubt. \n\n\nDetailed comments:\n\n-\tPage 2: “However, we argue that the intermediate feature maps by themselves are not a good representation of the knowledge encoded in the teacher to teach the students” – This sentence states a claim, which is a main claim of this paper. However, the claim is not supported by any argument or justification in the introduction\n-\tPage 3: the subscript \\Tau in D_{\\Tau}^{(l)} seems to be meaningless (why do we need it?)\n\tLater is becomes clearer as D_S is introduced, but a note should be given before to make \\Tau meaningful\n-\tPage 4: the pixel-labels are based on the 1-nn dictionary item, and so the knowledge transfer proposed in actually based a simple clustering of the pixels (each pixel represented by a single cluster index), not on the sparse representation. First, this means that K=1 can be naturally used, as K>1 is not really adding information to be transferred. Second, the method is very similar to transfer by clustering, seemed to be proposed by (Jain et al.) (based on the description of this work in the relevant work – I am not familiar with it)\n-\tPage 4: the image labels are enforcing similarity of word histograms between teacher and student. It is not clear, tough, why the BCE loss is used and not some histogram distance loss (l_1, Xi square, or even l_2). BCE is appropriate when the entrees in the two input vectors are probabilities of binary classification for independent problems. This is not the case here.\n-\tPage 5: it seems that comparison to the most similar method of (Jain et al.,) is not conducted \n-\tPage 6: the experiments showing that SRM is not sensitive to \\mu and \\lambda increase the likelihood that simple clustering, where each pixel is hard-assigned to one of the clusters can be used instead of sparse code encoding (we know that the identity of the single first dictionary item is the only one transferred in the pixel-labels, so it is likely that similarity to other dictionary items is not relevant)\n-\tThe task transfer results (table 5) are the most impressive results of the paper, indicating clear advantage of SRM over competitors. However, two drawbacks:\no\tWhy are the student networks randomly initialized and not initialized with their trained ImageNet weights (at least for ResNet 18 weights are publically available, I believe). Using pre-trained weight is likely to improve accuracy, and it is not clear why they are avoided\no\tComaprison to Jain et al is missing\n-\tSection 4.3: the teacher accuracy is missing\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The method presented looks promising but cleaner experiments are required to clarify its working elements and provide comparison to a close aompetitor", "review": "Strength and weaknesses:\n+\nThe method is reasonable and competitive. Especially for the task transfer scenario is seems to advance the state of the art\n\n-\nComparison is lacking as a main relevant competitor was not tested\nI believe the method can be much simplified by instead of using sparse decomposition using plain clustering of pixel columns. Experiments testing this option should be conducted\nThe task transfer experiments, whose results are the most impressive, are done without weight Initialization of the student networks. It is not clear why, and the results may be different if initialized networks are used. \nAt the bottom line: cleaner experiments (adding comparison to Jain et al, using plain clustering instead of sparse decomposition, using student networks with ImageNet initialized weights) are required to take this paper beyond reasonable doubt. \n\n\nDetailed comments:\n\n-\tPage 2: “However, we argue that the intermediate feature maps by themselves are not a good representation of the knowledge encoded in the teacher to teach the students” – This sentence states a claim, which is a main claim of this paper. However, the claim is not supported by any argument or justification in the introduction\n-\tPage 3: the subscript \\Tau in D_{\\Tau}^{(l)} seems to be meaningless (why do we need it?)\n\tLater is becomes clearer as D_S is introduced, but a note should be given before to make \\Tau meaningful\n-\tPage 4: the pixel-labels are based on the 1-nn dictionary item, and so the knowledge transfer proposed in actually based a simple clustering of the pixels (each pixel represented by a single cluster index), not on the sparse representation. First, this means that K=1 can be naturally used, as K>1 is not really adding information to be transferred. Second, the method is very similar to transfer by clustering, seemed to be proposed by (Jain et al.) (based on the description of this work in the relevant work – I am not familiar with it)\n-\tPage 4: the image labels are enforcing similarity of word histograms between teacher and student. It is not clear, tough, why the BCE loss is used and not some histogram distance loss (l_1, Xi square, or even l_2). BCE is appropriate when the entrees in the two input vectors are probabilities of binary classification for independent problems. This is not the case here.\n-\tPage 5: it seems that comparison to the most similar method of (Jain et al.,) is not conducted \n-\tPage 6: the experiments showing that SRM is not sensitive to \\mu and \\lambda increase the likelihood that simple clustering, where each pixel is hard-assigned to one of the clusters can be used instead of sparse code encoding (we know that the identity of the single first dictionary item is the only one transferred in the pixel-labels, so it is likely that similarity to other dictionary items is not relevant)\n-\tThe task transfer results (table 5) are the most impressive results of the paper, indicating clear advantage of SRM over competitors. However, two drawbacks:\no\tWhy are the student networks randomly initialized and not initialized with their trained ImageNet weights (at least for ResNet 18 weights are publically available, I believe). Using pre-trained weight is likely to improve accuracy, and it is not clear why they are avoided\no\tComaprison to Jain et al is missing\n-\tSection 4.3: the teacher accuracy is missing\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603883254489}, {"id": "in-swVyEztP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1536/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "[Summary] \nThis paper proposes to do knowledge distillation via matching the coefficients of sparse representations. The sparse representation reconstructs deep neural networks’ intermediate representation via a set of over-complete dictionary and corresponding coefficients. The method matches the coefficients at two levels. One is at the pixel level, and the other is at the image level, which pools the coefficients from the pixel level. The experiments perform model compression for image classification tasks and show a performance gain over previous methods.\n\n[Strengths]\n1. The idea of using sparse representation is novel and interesting.\n2. The clarity is good and easy to follow. The related works are sufficiently discussed.\n\n[Weaknesses]\n1. The method lacks some details, such as how D_T and D_S are learned and optimized. Please consider using an algorithm block to describe the full procedure of the proposed method.\n2. The experiment part has a large room to be improved. It is known that KD methods are sensitive to the type of architectures used in the teacher or student model. Moreover, most of the methods are very sensitive to the choice of hyper-parameters. There are two concerns: (1) Does the proposed method generalize to a broader range of network architectures pairs? (ex: the analysis in Tian et al. 2020) (2) Do all methods in the table be given the same amount of tuning budget for the hyper-parameter? Based on the text, different methods are given a different tuning budget. For example, Tables 1 and 5 only have part of the methods that are tuned with a validation set while others use the default value from other papers, making the comparison unfair.\n3. There are already more than a dozen distillation methods try to match the intermediate representation in a wide variety of ways, while most of them seem to work. Maybe matching a random projection of intermediate representation between teacher and student will work as well. Then why we want to use a more complicated method? The work will be more valuable if it provides a more in-depth insight into why using sparse representation helps. The paper has some arguments in the introduction, but they are not convincing since the coefficients can be the same even though the dictionaries between teacher and student are very different. A mathematical explanation is preferred. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "[Summary] \nThis paper proposes to do knowledge distillation via matching the coefficients of sparse representations. The sparse representation reconstructs deep neural networks’ intermediate representation via a set of over-complete dictionary and corresponding coefficients. The method matches the coefficients at two levels. One is at the pixel level, and the other is at the image level, which pools the coefficients from the pixel level. The experiments perform model compression for image classification tasks and show a performance gain over previous methods.\n\n[Strengths]\n1. The idea of using sparse representation is novel and interesting.\n2. The clarity is good and easy to follow. The related works are sufficiently discussed.\n\n[Weaknesses]\n1. The method lacks some details, such as how D_T and D_S are learned and optimized. Please consider using an algorithm block to describe the full procedure of the proposed method.\n2. The experiment part has a large room to be improved. It is known that KD methods are sensitive to the type of architectures used in the teacher or student model. Moreover, most of the methods are very sensitive to the choice of hyper-parameters. There are two concerns: (1) Does the proposed method generalize to a broader range of network architectures pairs? (ex: the analysis in Tian et al. 2020) (2) Do all methods in the table be given the same amount of tuning budget for the hyper-parameter? Based on the text, different methods are given a different tuning budget. For example, Tables 1 and 5 only have part of the methods that are tuned with a validation set while others use the default value from other papers, making the comparison unfair.\n3. There are already more than a dozen distillation methods try to match the intermediate representation in a wide variety of ways, while most of them seem to work. Maybe matching a random projection of intermediate representation between teacher and student will work as well. Then why we want to use a more complicated method? The work will be more valuable if it provides a more in-depth insight into why using sparse representation helps. The paper has some arguments in the introduction, but they are not convincing since the coefficients can be the same even though the dictionaries between teacher and student are very different. A mathematical explanation is preferred. \n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603877258512}, {"id": "E1QfDE1aQAq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1536/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "<Paper summary>\nThis paper focuses on the problem of knowledge transfer between deep learning models, with the goal of transferring some of the information contained in a (typically large) teacher model to a smaller student network, improving performance of the latter. In particular, this work proposes to require the sparse representations of the activations of the teacher model to be 'similar' to that of the student model, as way of knowledge transfer. The method is presented and empirically evaluated.\n\n<Review summary>\nThis reviewer likes the general motivation of this work, proposing that in some cases it might be better to enforce similarity between activations (or features) in a different domain, and requiring these representations to be sparse under some transformation is natural. While the general idea is appealing, several of the motivating claims are vague and the way these ideas are implemented (e.g. via classification to enforce similarity between vectors) are questionable.\n\n<Details comments>\nStrengths:\n- the authors study an interest problem.\n- the proposed method obtains good empirical performance.\n\nWeaknesses:\n- The idea of matching the representations for two different data in order to enforce some similarity (or in this context, 'knowledge transfer') has been extensively used. However, this only makes sense if the dictionaries for one and other case are related (see [1,2,3]). Enforcing the representations to be similar (or, as it's done here, to use the same leading atom) is reasonable when the atoms from one dictionary share some properties (or 'code' for related things) in the other.  In this work, the authors learn dictionaries for the teacher and student networks (Ds and Dt) completely independently, as there's no connection between the atoms in one and other dictionaries. \n- The authors include an 'image-level labeling' which basically compares the mean value in the approximate features from the teacher network ($\\tilde{t}$) to that of the vector of similarities of the student one ($k_s$). 1) I do not see how this is informative of relevant information between models, but more importantly 2) Comparing these two real numbers with a logistic loss makes little (if any) sense to me.\n- The idea of employing sparse representations for data (in this case, the intermediate representations of networks) is natural. There is a large body of work that the authors seem to ignore. For example: on pg 2 they mention that \"[sparse representations learning] were not proposed to be jointly optimized with other objectives\". Please see refs [4-6] below for examples of this.\n- Representations under redundant dictionaries are not unique, and the problem of finding sparse representations is NP-hard (see e.g [7]). Certainly, one can propose relaxations of this problem and even heuristic approximations (see e.g. [8]) but this is never discussed, and it is unclear how the obtained representations in this work fit in this context.\n\nSmaller comments:\n- The proposed method seems to be a pre-processing step: firs train teacher dictionaries, then student dictionaries and student weights, and then employ the KD method from Hinton et al. The authors should consider making this more explicit, perhaps detailing the full algorithm in a formal way.\n- On page 4, the authors motivate the use of the sigmoid as activation function saying that the 'gradients in the backward pass are stable'. What does this mean? Would they be unstable if a ReLU was used instead (as used in most deep learning models and in the other models in this work)?\n- At the end of Pixel-level labeling, stating the definition for $c_{n,i,j}$, the authors take the argmax over m, but there's no m in the expression.\n- 'Relaxing' the problem of requiring the representations to be similar simply by turning a regression problem into a classification problem seems unfounded: why would the latter be easier than simply allowing the representations to be similar (say, with small L2 norm)?\n- In defining the similarity kernel $\\kappa$, I believe the $x$ and $y$ should be bold according to the authors notation.\n- On a subjective note, the notation is not standard and thus a bit confusing: calligraphic capital letters usually denote sets or distributions, whereas here they denote vectors (as do bold non-calligraphic letters).\n\nReferences:\n1] Wang, Shenlong, et al. \"Semi-coupled dictionary learning with applications to image super-resolution and photo-sketch synthesis.\" 2012 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2012.\n\n2] Qiu, Qiang, et al. \"Domain adaptive dictionary learning.\" European Conference on Computer Vision. Springer, Berlin, Heidelberg, 2012.\n\n3] Peleg, Tomer, and Michael Elad. \"A statistical prediction model based on sparse representations for single image super-resolution.\" IEEE transactions on image processing 23.6 (2014): 2569-2582.\n\n4] Mairal, Julien, Francis Bach, and Jean Ponce. \"Task-driven dictionary learning.\" IEEE transactions on pattern analysis and machine intelligence 34.4 (2011): 791-804.\n\n5] Sprechmann, Pablo, Alexander M. Bronstein, and Guillermo Sapiro. \"Learning efficient sparse and low rank models.\" IEEE transactions on pattern analysis and machine intelligence 37.9 (2015): 1821-1833.\n\n6] Monga, Vishal, Yuelong Li, and Yonina C. Eldar. \"Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing.\" arXiv preprint arXiv:1912.10557 (2019).\n\n7] Mairal, Julien, Francis Bach, and Jean Ponce. \"Sparse modeling for image and vision processing.\" \n\n8] Makhzani, Alireza, and Brendan Frey. \"K-sparse autoencoders.\" arXiv preprint arXiv:1312.5663 (2013).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Anonymous Review", "review": "<Paper summary>\nThis paper focuses on the problem of knowledge transfer between deep learning models, with the goal of transferring some of the information contained in a (typically large) teacher model to a smaller student network, improving performance of the latter. In particular, this work proposes to require the sparse representations of the activations of the teacher model to be 'similar' to that of the student model, as way of knowledge transfer. The method is presented and empirically evaluated.\n\n<Review summary>\nThis reviewer likes the general motivation of this work, proposing that in some cases it might be better to enforce similarity between activations (or features) in a different domain, and requiring these representations to be sparse under some transformation is natural. While the general idea is appealing, several of the motivating claims are vague and the way these ideas are implemented (e.g. via classification to enforce similarity between vectors) are questionable.\n\n<Details comments>\nStrengths:\n- the authors study an interest problem.\n- the proposed method obtains good empirical performance.\n\nWeaknesses:\n- The idea of matching the representations for two different data in order to enforce some similarity (or in this context, 'knowledge transfer') has been extensively used. However, this only makes sense if the dictionaries for one and other case are related (see [1,2,3]). Enforcing the representations to be similar (or, as it's done here, to use the same leading atom) is reasonable when the atoms from one dictionary share some properties (or 'code' for related things) in the other.  In this work, the authors learn dictionaries for the teacher and student networks (Ds and Dt) completely independently, as there's no connection between the atoms in one and other dictionaries. \n- The authors include an 'image-level labeling' which basically compares the mean value in the approximate features from the teacher network ($\\tilde{t}$) to that of the vector of similarities of the student one ($k_s$). 1) I do not see how this is informative of relevant information between models, but more importantly 2) Comparing these two real numbers with a logistic loss makes little (if any) sense to me.\n- The idea of employing sparse representations for data (in this case, the intermediate representations of networks) is natural. There is a large body of work that the authors seem to ignore. For example: on pg 2 they mention that \"[sparse representations learning] were not proposed to be jointly optimized with other objectives\". Please see refs [4-6] below for examples of this.\n- Representations under redundant dictionaries are not unique, and the problem of finding sparse representations is NP-hard (see e.g [7]). Certainly, one can propose relaxations of this problem and even heuristic approximations (see e.g. [8]) but this is never discussed, and it is unclear how the obtained representations in this work fit in this context.\n\nSmaller comments:\n- The proposed method seems to be a pre-processing step: firs train teacher dictionaries, then student dictionaries and student weights, and then employ the KD method from Hinton et al. The authors should consider making this more explicit, perhaps detailing the full algorithm in a formal way.\n- On page 4, the authors motivate the use of the sigmoid as activation function saying that the 'gradients in the backward pass are stable'. What does this mean? Would they be unstable if a ReLU was used instead (as used in most deep learning models and in the other models in this work)?\n- At the end of Pixel-level labeling, stating the definition for $c_{n,i,j}$, the authors take the argmax over m, but there's no m in the expression.\n- 'Relaxing' the problem of requiring the representations to be similar simply by turning a regression problem into a classification problem seems unfounded: why would the latter be easier than simply allowing the representations to be similar (say, with small L2 norm)?\n- In defining the similarity kernel $\\kappa$, I believe the $x$ and $y$ should be bold according to the authors notation.\n- On a subjective note, the notation is not standard and thus a bit confusing: calligraphic capital letters usually denote sets or distributions, whereas here they denote vectors (as do bold non-calligraphic letters).\n\nReferences:\n1] Wang, Shenlong, et al. \"Semi-coupled dictionary learning with applications to image super-resolution and photo-sketch synthesis.\" 2012 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2012.\n\n2] Qiu, Qiang, et al. \"Domain adaptive dictionary learning.\" European Conference on Computer Vision. Springer, Berlin, Heidelberg, 2012.\n\n3] Peleg, Tomer, and Michael Elad. \"A statistical prediction model based on sparse representations for single image super-resolution.\" IEEE transactions on image processing 23.6 (2014): 2569-2582.\n\n4] Mairal, Julien, Francis Bach, and Jean Ponce. \"Task-driven dictionary learning.\" IEEE transactions on pattern analysis and machine intelligence 34.4 (2011): 791-804.\n\n5] Sprechmann, Pablo, Alexander M. Bronstein, and Guillermo Sapiro. \"Learning efficient sparse and low rank models.\" IEEE transactions on pattern analysis and machine intelligence 37.9 (2015): 1821-1833.\n\n6] Monga, Vishal, Yuelong Li, and Yonina C. Eldar. \"Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing.\" arXiv preprint arXiv:1912.10557 (2019).\n\n7] Mairal, Julien, Francis Bach, and Jean Ponce. \"Sparse modeling for image and vision processing.\" \n\n8] Makhzani, Alireza, and Brendan Frey. \"K-sparse autoencoders.\" arXiv preprint arXiv:1312.5663 (2013).", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603853095467}], "openreview_url": "https://openreview.net/forum?id=Ip195saXqIX", "arxiv_id": "2103.17012", "paper_pdf": "papers/Ip195saXqIX.pdf", "paper_pdf_sha256": "447743ef5767e211c317c60c7869733051dbfe5a1b6064c519bac43d4d8f8b6c", "paper_pdf_bytes": 237873, "paper_pdf_source": "openreview", "code_url": "https://github.com/viebboy/SRM", "code_repository": "viebboy/SRM", "code_commit": "77a12ba790c799e64624d155c8faff3a5e47c477", "code_archive": "repos/Ip195saXqIX.zip", "code_archive_sha256": "935175d781aff22a6f1ba8328058f20dabdbad6e4be0109146055bf14e08814d", "code_archive_bytes": 109690, "code_file_count": 71, "code_extensions": {".py": 68, ".sh": 3}, "github_disk_usage_kb": 83, "github_languages": {"Python": 520853, "Shell": 3444}, "github_archived": false, "github_pushed_at": "2021-03-31T09:37:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/knowledge-distillation-by-sparse-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HyeqPJHYvH", "year": 2020, "status": "rejected", "title": "Stochastic Latent Residual Video Prediction", "authors": ["Jean-Yves Franceschi", "Edouard Delasalles", "Mickael Chen", "Sylvain Lamprier", "Patrick Gallinari"], "authorids": ["jean-yves.franceschi@lip6.fr", "edouard.delasalles@lip6.fr", "mickael.chen@lip6.fr", "sylvain.lamprier@lip6.fr", "patrick.gallinari@lip6.fr"], "authors_source": "OpenReview API", "abstract": "Video prediction is a challenging task: models have to account for the inherent uncertainty of the future. Most works in the literature are based on stochastic image-autoregressive recurrent networks, raising several performance and applicability issues. An alternative is to use fully latent temporal models which untie frame synthesis and dynamics. However, no such model for video prediction has been proposed in the literature yet, due to design and training difficulties. In this paper, we overcome these difficulties by introducing a novel stochastic temporal model. It is based on residual updates of a latent state, motivated by discretization schemes of differential equations. This first-order principle naturally models video dynamics as it allows our simpler, lightweight, interpretable, latent model to outperform prior state-of-the-art methods on challenging datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "S1efBB1AYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1777/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe paper proposes a video prediction method based on State-Space Models. The paper describes two main contributions:\n\n1. By learning dynamics in the latent state space, the method avoids the high computational cost and accumulating image reconstruction errors of autoregressive models that condition on generated frames.\n\n2. To model dynamics, the paper proposes a residual update rule inspired by Euler’s method to solve ODEs. According to this rule, the update to the latent state y_t is modeled as an additive residual f(y_t, z_{t+1}). This has the advantage that the step size of the discretization can be adjusted freely, e.g. between training and inference. \n\nThe paper provides extensive experimental comparison of their model to the SVG and SAVP models on several standard datasets. The paper further contains experiments illustrating features of the model such as disentangling dynamics and content, and interpolation of dynamics in the latent space.\n\nDecision:\n\nThe paper is written clearly and the mathematical treatment and experiments appear rigorous. The idea of predicting video using state-space models is interesting and promising. However, as described below, the paper overstates its novelty and falls short of showing the advantages of the method beyond incremental improvements on frame-wise image quality metrics. I therefore suggest rejection in its current version.\n\nSupporting arguments and suggestions:\n\n1. The idea to use fully latent models for video prediction, to untie frame synthesis and dynamics, is not new and the paper does not fully cite this literature. For example, [1] and [2] perform unsupervised, non-autoregressive video prediction. The differences to these models should be discussed.\n\n2. The advantages of the residual update rule are not made clear enough. The parallels to the ODE literature seem tenuous. The main advantage described in the paper is the ability to synthesize videos at different frame rates, but interpolation over such short time horizons is not a hard problem. At least, the paper should compare to existing methods for frame interpolation. Apart from interpolation (variable step size), it appears that the update rule could be changed from y_{t+1} = y_t + f(y_t, z_{t+1}) to y_{t+1} = f(y_t, z_{t+1}) without impact to the model. How is it different from the standard VRNN formulation [3]? More experiments to show the advantage of the proposed update rule would be helpful.\n\n3. Some of the experiments seem like interesting starting points but do not support general claims. For example, Fig 2 (b) shows that the proposed dynamics model is better than an MLP or GRU on deterministic Moving MNIST, but is this also true on real datasets, which have much more complex dynamics? Similarly, the interpolation in Figure 9 is intriguing, but it would be helpful to describe and test how this ability is useful for applications of the predictive model.\n\n4. The comparisons use frame-wise metrics of image quality (PSNR, SSIM, LPIPS). Even though they are common in the literature, these metrics are unsuitable for comparing long video sequences due to their stochasticity. The metrics are probably dominated by relatively uninteresting features such as the quality of the static background. Metrics for comparing entire videos exist (e.g. FVD [4]) and should be used. Even better, the paper should demonstrate the usefulness of the model for downstream tasks such as reinforcement learning, although I understand that this may be out of scope.\n\nMinor comments:\n\n- As far as I know, the correct term for error terms is residual, not residue.\n- What do the error bars in the figures show? Please add this information to the figure legends.\n\n[1] Wichers et al., 2018, https://arxiv.org/pdf/1806.04768.pdf\n[2] Minderer et al., 2019, https://arxiv.org/abs/1906.07889\n[3] Chung et al, 2015, https://arxiv.org/abs/1506.02216\n[4] Unterthiner et al., 2018, https://arxiv.org/abs/1812.01717", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary:\n\nThe paper proposes a video prediction method based on State-Space Models. The paper describes two main contributions:\n\n1. By learning dynamics in the latent state space, the method avoids the high computational cost and accumulating image reconstruction errors of autoregressive models that condition on generated frames.\n\n2. To model dynamics, the paper proposes a residual update rule inspired by Euler’s method to solve ODEs. According to this rule, the update to the latent state y_t is modeled as an additive residual f(y_t, z_{t+1}). This has the advantage that the step size of the discretization can be adjusted freely, e.g. between training and inference. \n\nThe paper provides extensive experimental comparison of their model to the SVG and SAVP models on several standard datasets. The paper further contains experiments illustrating features of the model such as disentangling dynamics and content, and interpolation of dynamics in the latent space.\n\nDecision:\n\nThe paper is written clearly and the mathematical treatment and experiments appear rigorous. The idea of predicting video using state-space models is interesting and promising. However, as described below, the paper overstates its novelty and falls short of showing the advantages of the method beyond incremental improvements on frame-wise image quality metrics. I therefore suggest rejection in its current version.\n\nSupporting arguments and suggestions:\n\n1. The idea to use fully latent models for video prediction, to untie frame synthesis and dynamics, is not new and the paper does not fully cite this literature. For example, [1] and [2] perform unsupervised, non-autoregressive video prediction. The differences to these models should be discussed.\n\n2. The advantages of the residual update rule are not made clear enough. The parallels to the ODE literature seem tenuous. The main advantage described in the paper is the ability to synthesize videos at different frame rates, but interpolation over such short time horizons is not a hard problem. At least, the paper should compare to existing methods for frame interpolation. Apart from interpolation (variable step size), it appears that the update rule could be changed from y_{t+1} = y_t + f(y_t, z_{t+1}) to y_{t+1} = f(y_t, z_{t+1}) without impact to the model. How is it different from the standard VRNN formulation [3]? More experiments to show the advantage of the proposed update rule would be helpful.\n\n3. Some of the experiments seem like interesting starting points but do not support general claims. For example, Fig 2 (b) shows that the proposed dynamics model is better than an MLP or GRU on deterministic Moving MNIST, but is this also true on real datasets, which have much more complex dynamics? Similarly, the interpolation in Figure 9 is intriguing, but it would be helpful to describe and test how this ability is useful for applications of the predictive model.\n\n4. The comparisons use frame-wise metrics of image quality (PSNR, SSIM, LPIPS). Even though they are common in the literature, these metrics are unsuitable for comparing long video sequences due to their stochasticity. The metrics are probably dominated by relatively uninteresting features such as the quality of the static background. Metrics for comparing entire videos exist (e.g. FVD [4]) and should be used. Even better, the paper should demonstrate the usefulness of the model for downstream tasks such as reinforcement learning, although I understand that this may be out of scope.\n\nMinor comments:\n\n- As far as I know, the correct term for error terms is residual, not residue.\n- What do the error bars in the figures show? Please add this information to the figure legends.\n\n[1] Wichers et al., 2018, https://arxiv.org/pdf/1806.04768.pdf\n[2] Minderer et al., 2019, https://arxiv.org/abs/1906.07889\n[3] Chung et al, 2015, https://arxiv.org/abs/1506.02216\n[4] Unterthiner et al., 2018, https://arxiv.org/abs/1812.01717"}, "tcdate": 1571841338374}, {"id": "r1xt2GQpFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1777/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a video prediction model which explicitly decouples frame synthesis and motion dynamics. This is a very subtle change (compared to the current models) that can result in higher quality predictions.\n\nFirst of all, the paper is extremely well written. It provides clear motivations and goals, as well as an impressively comprehensive related work that discusses their shortcomings. The experiments are comprehensive and provide good support for the claims. And finally, the appendix presents additional visualization and information. \n\nOn the main proposed method, it is a very subtle but reasonable change. Therefore, my suggestion to the authors is to provide a more thorough comparison with existing methods specifically SVG (Denton 2018) since the models share a lot of similarities. It is also quite similar to PlaNet (Hafner 2019). This is where the paper can be improved.\n\nFor the experiments, although they are quite comprehensive, there is still room for improvement. First, none of the metrics used are good evaluation metrics for frame prediction (I know they are quite common but that doesn't make them good) as they do not give us an objective evaluation in the sense of the semantic quality of predicted frames, specially for long videos. It really helps if the authors present additional quantitative evaluation to show that the predicted frames contain useful semantic information with metrics such as FVD and Inception score. Second, a pretribulation study is required to see where the improvements are coming from. Is it from a different architecture or the separation of dynamics?  Finally, a website with generated videos really helps for qualitative comparison! \n\nOverall, this is a well-written paper with clear motivations and goals. I find the impact of the paper to be marginal (given the quality difference with already existing models) which can be improved by emphasizing more on other aspects such as disentanglement.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper proposes a video prediction model which explicitly decouples frame synthesis and motion dynamics. This is a very subtle change (compared to the current models) that can result in higher quality predictions.\n\nFirst of all, the paper is extremely well written. It provides clear motivations and goals, as well as an impressively comprehensive related work that discusses their shortcomings. The experiments are comprehensive and provide good support for the claims. And finally, the appendix presents additional visualization and information. \n\nOn the main proposed method, it is a very subtle but reasonable change. Therefore, my suggestion to the authors is to provide a more thorough comparison with existing methods specifically SVG (Denton 2018) since the models share a lot of similarities. It is also quite similar to PlaNet (Hafner 2019). This is where the paper can be improved.\n\nFor the experiments, although they are quite comprehensive, there is still room for improvement. First, none of the metrics used are good evaluation metrics for frame prediction (I know they are quite common but that doesn't make them good) as they do not give us an objective evaluation in the sense of the semantic quality of predicted frames, specially for long videos. It really helps if the authors present additional quantitative evaluation to show that the predicted frames contain useful semantic information with metrics such as FVD and Inception score. Second, a pretribulation study is required to see where the improvements are coming from. Is it from a different architecture or the separation of dynamics?  Finally, a website with generated videos really helps for qualitative comparison! \n\nOverall, this is a well-written paper with clear motivations and goals. I find the impact of the paper to be marginal (given the quality difference with already existing models) which can be improved by emphasizing more on other aspects such as disentanglement.  "}, "tcdate": 1571791536871}, {"id": "r1gsleMpKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1777/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Contributions: this submission proposes a video pixel generation framework with the goal to decouple visual appearance and dynamics. The latent dynamics are modeled with a latent residual dynamics model. Empirical evaluations on moving MNIST show that the proposed residual dynamics model outperform MLP or GRU. On more challenging KTH and BAIR datasets, the proposed method achieves on par or better quantitative performance with previous methods, and have nice qualitative results on content \"swap\" and dynamics interpolation.\n\nAssessment:\n- To my knowledge, the proposed model is novel for video generation.\n- The proposed method is supported with strong quantitative results and qualitative analysis, ablation on Moving MNIST shows that the proposed latent residual dynamics model outperforms MLP and GRU baselines.\n- The authors might be interested in related work on video generation with decoupled appearance and dynamics models, such as [1]. It would also be interesting to see evaluation on more challenging datasets, such as Human3.6M.\n- Question: how does the proposed inference framework make sure to decouple appearance with dynamics? Can y_i not encode the appearance information?\n\n\n[1] Minderer et al., Unsupervised Learning of Object Structure and Dynamics from Videos. NeurIPS 2019.\n\n-----------------------------\nPost rebuttal:\nThank you for your answers to my questions and the updated manuscript. My questions have been addressed and the additional results further confirm the performance of the proposed method. Therefore I recommend weak accept of the submission.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "Contributions: this submission proposes a video pixel generation framework with the goal to decouple visual appearance and dynamics. The latent dynamics are modeled with a latent residual dynamics model. Empirical evaluations on moving MNIST show that the proposed residual dynamics model outperform MLP or GRU. On more challenging KTH and BAIR datasets, the proposed method achieves on par or better quantitative performance with previous methods, and have nice qualitative results on content \"swap\" and dynamics interpolation.\n\nAssessment:\n- To my knowledge, the proposed model is novel for video generation.\n- The proposed method is supported with strong quantitative results and qualitative analysis, ablation on Moving MNIST shows that the proposed latent residual dynamics model outperforms MLP and GRU baselines.\n- The authors might be interested in related work on video generation with decoupled appearance and dynamics models, such as [1]. It would also be interesting to see evaluation on more challenging datasets, such as Human3.6M.\n- Question: how does the proposed inference framework make sure to decouple appearance with dynamics? Can y_i not encode the appearance information?\n\n\n[1] Minderer et al., Unsupervised Learning of Object Structure and Dynamics from Videos. NeurIPS 2019.\n\n-----------------------------\nPost rebuttal:\nThank you for your answers to my questions and the updated manuscript. My questions have been addressed and the additional results further confirm the performance of the proposed method. Therefore I recommend weak accept of the submission.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571786739013}], "openreview_url": "https://openreview.net/forum?id=HyeqPJHYvH", "arxiv_id": "2002.09219", "paper_pdf": "papers/HyeqPJHYvH.pdf", "paper_pdf_sha256": "1c3f295587273043f7d8aea2ab845dded749683bdc1ab9cd67e7c09c66c8bd80", "paper_pdf_bytes": 5806910, "paper_pdf_source": "openreview", "code_url": "https://github.com/edouardelasalles/srvp", "code_repository": "edouardelasalles/srvp", "code_commit": "05661faf767cdb33d40fc328679bbe50c3a1f938", "code_archive": "repos/HyeqPJHYvH.zip", "code_archive_sha256": "63bdeb12c2fe10117b52e9613f4729d25a13e1990a4f2b3883bf08a2dcdeb43b", "code_archive_bytes": 77526, "code_file_count": 29, "code_extensions": {".py": 27, ".sh": 2}, "github_disk_usage_kb": 130, "github_languages": {"Python": 199873, "Shell": 551}, "github_archived": false, "github_pushed_at": "2022-03-29T09:25:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/stochastic-latent-residual-video-prediction-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkGpW3C5KX", "year": 2019, "status": "rejected", "title": "Heated-Up Softmax Embedding", "authors": ["Xu Zhang", "Felix Xinnan Yu", "Svebor Karaman", "Wei Zhang", "Shih-Fu Chang"], "authorids": ["xu.zhang@columbia.edu", "felixyu@google.com", "svebor.karaman@gmail.com", "wz2363@columbia.edu", "sc250@columbia.edu"], "authors_source": "OpenReview API", "abstract": "Metric learning aims at learning a distance which is consistent with the semantic meaning of the samples. The problem is generally solved by learning an embedding, such that the samples of the same category are close (compact) while samples from different categories are far away (spread-out) in the embedding space. One popular way of generating such embeddings is to use the second-to-last layer of a deep neural network trained as a classifier with the softmax cross-entropy loss. In this paper, we show that training classifiers with different temperatures of the softmax function lead to different distributions of the embedding space. And finding a balance between the compactness, 'spread-out' and the generalization ability of the feature is critical in metric learning. Leveraging these insights, we propose a 'heating-up' strategy to train a classifier with increasing temperatures. Extensive experiments show that the proposed method achieves state-of-the-art embeddings on a variety of metric learning benchmarks. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "ByeFu5_ahQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1215/AnonReviewer3"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents an interesting idea to improve the softmax embedding performance with heated-up strategy. It is well-written and the proposed method is easy to implement. Several experiments on metric learning datasets demonstrate the effectiveness of the proposed method.\n\nThe motivation to find a balance between the compactness and \"spread-out\" embedding is reasonable. The major weakness is the intermediate temperature selection, it might be a little tricky. How to generalize it to other applications?\n\nThe authors claim that \"heated-up\" strategy produces well generalized feature, but the rationale behind is unclear. And there is no quantitative analysis to support this point. \n\nThe starting temperature aims at pushing the “incorrect” samples to “boundary” samples and pushing the “boundary” samples to “centroid” samples. I would like to see the ratio of #incorrect/total and #boundary/total changed with different temperature in training process, i.e., alpha = 16, 4, 1. This experiment may help to verify the idea.\n\nAs mentioned in Section 3, multiple strategies could be defined to increase the temperature. It is interesting to design a multiple heat-up strategy. Does it help to improve the learning speed?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Heated-Up Softmax Embedding", "review": "This paper presents an interesting idea to improve the softmax embedding performance with heated-up strategy. It is well-written and the proposed method is easy to implement. Several experiments on metric learning datasets demonstrate the effectiveness of the proposed method.\n\nThe motivation to find a balance between the compactness and \"spread-out\" embedding is reasonable. The major weakness is the intermediate temperature selection, it might be a little tricky. How to generalize it to other applications?\n\nThe authors claim that \"heated-up\" strategy produces well generalized feature, but the rationale behind is unclear. And there is no quantitative analysis to support this point. \n\nThe starting temperature aims at pushing the “incorrect” samples to “boundary” samples and pushing the “boundary” samples to “centroid” samples. I would like to see the ratio of #incorrect/total and #boundary/total changed with different temperature in training process, i.e., alpha = 16, 4, 1. This experiment may help to verify the idea.\n\nAs mentioned in Section 3, multiple strategies could be defined to increase the temperature. It is interesting to design a multiple heat-up strategy. Does it help to improve the learning speed?\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541405297080}, {"id": "S1loo0Von7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1215/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The introduction and the title does not match. Metric learning does not require to specify the dimension; while the embedding has to specify the reduced dimension. I feel confused that the authors mix these two concepts.\n\nThe objective in (1) is very close to that of t-SNE[5], where it uses the KL as the objective. Then other update formula are similar.  \n\nThis paper facilitates the effect of temperature in the Softmax function to heuristically learn a compact and spread-out embedding. However, such an idea have been widely used and investigated in Reinforcement learning [1], Knowledge distillation [2], classification [3] and discrete variable optimization [4] and t-SNE visualization [5] etc. Thus, the insight about the temperature effect on the embedding from the second last layer, cannot be novel any more. Based on this, the proposed ``heating-up” strategy to leverage its effect on the embedding is heuristic, since the temperature parameter is manually set instead of automatically learning. In this case, I do expect the authors should provide more in-depth theoretical analysis. \n\nThe authors do not present more experimental results on the correlation between the final performance and this temperature setting. \n\nBesides, as the alpha increases or decreases, the side-effect on the learning rate setting for the optimization have not clearly analyzed, which leaves more concerns on tuning performance. \n\n\n[1] Sutton, R. S. and Barto A. G. Reinforcement Learning: An Introduction. The MIT Press, Cambridge, MA, 1998.\n[2] Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. NIPS 2015.\n[3] Guo, Chuan, et al. \"On calibration of modern neural networks.\" ICML 2017.\n[4] Jang E, Gu S, Poole B. Categorical reparameterization with gumbel-softmax. ICLR 2017.\n[5] Maaten L, Hinton G. Visualizing data using t-SNE[J]. Journal of machine learning research, 2008, 9(Nov): 2579-2605.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novelty", "review": "The introduction and the title does not match. Metric learning does not require to specify the dimension; while the embedding has to specify the reduced dimension. I feel confused that the authors mix these two concepts.\n\nThe objective in (1) is very close to that of t-SNE[5], where it uses the KL as the objective. Then other update formula are similar.  \n\nThis paper facilitates the effect of temperature in the Softmax function to heuristically learn a compact and spread-out embedding. However, such an idea have been widely used and investigated in Reinforcement learning [1], Knowledge distillation [2], classification [3] and discrete variable optimization [4] and t-SNE visualization [5] etc. Thus, the insight about the temperature effect on the embedding from the second last layer, cannot be novel any more. Based on this, the proposed ``heating-up” strategy to leverage its effect on the embedding is heuristic, since the temperature parameter is manually set instead of automatically learning. In this case, I do expect the authors should provide more in-depth theoretical analysis. \n\nThe authors do not present more experimental results on the correlation between the final performance and this temperature setting. \n\nBesides, as the alpha increases or decreases, the side-effect on the learning rate setting for the optimization have not clearly analyzed, which leaves more concerns on tuning performance. \n\n\n[1] Sutton, R. S. and Barto A. G. Reinforcement Learning: An Introduction. The MIT Press, Cambridge, MA, 1998.\n[2] Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. NIPS 2015.\n[3] Guo, Chuan, et al. \"On calibration of modern neural networks.\" ICML 2017.\n[4] Jang E, Gu S, Poole B. Categorical reparameterization with gumbel-softmax. ICLR 2017.\n[5] Maaten L, Hinton G. Visualizing data using t-SNE[J]. Journal of machine learning research, 2008, 9(Nov): 2579-2605.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541258915169}, {"id": "BJlWjOyc2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1215/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper proposes a novel optimization strategy regarding softmax cross-entropy loss, to extract the effective features of well generalization in the framework of metric learning.\nThe authors focus on the \"temperature\" parameter in the softmax and through analyzing the role of the temperature in terms of gradient, propose the approach of heating-up softmax in which the temperature is varied from low to high in training.\nAnd, the effects of normalization such as by l2 and BatchNorm are discussed in the framework of heated-up softmax.\nThe experimental results on metric learning tasks demonstrate the effectiveness of the proposed method in comparison with the other methods.\n\nComments:\nPros:\n+ The idea of heating up the temperature in softmax is interesting, and seems novel in the literature of metric learning.\n+ The performance improvement, especially produced by batchNorm-based normalization, is shown.\n\nCons:\n- The formulation of tempered softmax with normalization is already presented in [Wang et al., 2017].\n- The reason why the heating-up approach contributes to better metric learning is not clearly provided in a well convincing way.\n- It lacks an important ablation study to fairly validate the method.\n- The discussion/comparison is limited to the simple softmax function.\n\nAlthough the reviewer likes the idea of heating up softmax, this paper can be judged as a borderline slightly leaning toward reject, due to the above weak points, the details of which are explained as follows.\n\n- Formulation\nThe softmax equipped with temperature for the normalized features and weights are shown in [Wang et al., 2017]. The only difference from that work is the way to deal with temperature; in [Wang et al., 2017], the temperature is \"optimized\" as a trainable parameter, while it is dealt with in a hand-crafted way of heating up in this work. Honestly speaking, it is unclear which approach is better, though the optimization in [Wang et al., 2017] seems elegant as stated in that paper. The only way to validate this work compared to [Wang et al., 2017] is to empirically evaluate those two methods in the experiments. Such a comparison experiment is not found and it is a main flaw of this paper.\n\n- Justification of the method\nThe gradients of the softmax cross-entropy loss parameterized with a temperature T are well analyzed in Sections 3.1&3.2. But, in Section 3.3, the reviewer cannot find the clear and convincing explanation for why the temperature T should be increased in the training. My question is: why don't you use alpha=4 consistently throughout the training?\n It might be related to the process of simulated annealing (though \"temperature\" is usually cooled down in SA), and more interestingly, it would also be possible to find connection with the work of [Guo et al., 2017]. In [Guo et al., 2017], the temperature in the softmax is optimized as a post processing for calibrating the classifier outputs. Though the calibration task itself is a little bit apart from the metric learning of the authors' interest, we can find in that paper an interesting result that the temperature is heated up to increase the confidence of the classifier outputs, which is quite similar to the process of fine-tuning by heating up softmax as done in this work. Therefore, the reviewer guesses that the effectiveness of heating up softmax can also be interpreted from the viewpoint of [Guo et al., 2017].\n\nThere is also less description about Figure 1; in particular, the reviewer cannot understand what Figure 1(d) means.\n\n- Ablation study\nTo empirically resolve the above concerns, it is necessary to present the empirical comparison with the \"static\" softmax.\nNamely, the methods of HLN/HBN should be carefully compared to LN/BN of \"alpha=4\", not only those of alpha=16 shown in Table 1&2; the comparison in Table 3 seems unfair since the authors apply the static softmax without normalization.\nAnd, it would be better to show the performance of heated-up softmax \"without\" normalization to show the important role of the normalization, as done in [Wang et al., 2017].\nIn summary, since the proposed method is composed of a heating-up approach and feature normalization, the authors are required to validate the method from those two aspects, respectively, for increasing the significance of this paper.\n\n- Other loss function\nFor achieving a compactness in feature representation, the simple softmax requires both temperature and normalization. It, however, is also conceivable to employ the other types of loss function for that purpose, such as [a] which is based on the (Mahalanobis) distance with taking into account the margin between categories. The distance based loss also embeds features into localized clusters, which satisfies the authors' objective in this work. To validate the proposed method, it is required to compare the method with such a different types of loss function.\n\n[a] Wan, W., Zhong, Y., Li, T., & Chen, J. (2018). Rethinking Feature Distribution for Loss Functions in Image Classification, In CVPR2018, pp. 9117–9126.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "It is a simple and interesting method, but lacks discussions and/or empirical evaluation in comparison with the prior work.", "review": "Summary:\nThis paper proposes a novel optimization strategy regarding softmax cross-entropy loss, to extract the effective features of well generalization in the framework of metric learning.\nThe authors focus on the \"temperature\" parameter in the softmax and through analyzing the role of the temperature in terms of gradient, propose the approach of heating-up softmax in which the temperature is varied from low to high in training.\nAnd, the effects of normalization such as by l2 and BatchNorm are discussed in the framework of heated-up softmax.\nThe experimental results on metric learning tasks demonstrate the effectiveness of the proposed method in comparison with the other methods.\n\nComments:\nPros:\n+ The idea of heating up the temperature in softmax is interesting, and seems novel in the literature of metric learning.\n+ The performance improvement, especially produced by batchNorm-based normalization, is shown.\n\nCons:\n- The formulation of tempered softmax with normalization is already presented in [Wang et al., 2017].\n- The reason why the heating-up approach contributes to better metric learning is not clearly provided in a well convincing way.\n- It lacks an important ablation study to fairly validate the method.\n- The discussion/comparison is limited to the simple softmax function.\n\nAlthough the reviewer likes the idea of heating up softmax, this paper can be judged as a borderline slightly leaning toward reject, due to the above weak points, the details of which are explained as follows.\n\n- Formulation\nThe softmax equipped with temperature for the normalized features and weights are shown in [Wang et al., 2017]. The only difference from that work is the way to deal with temperature; in [Wang et al., 2017], the temperature is \"optimized\" as a trainable parameter, while it is dealt with in a hand-crafted way of heating up in this work. Honestly speaking, it is unclear which approach is better, though the optimization in [Wang et al., 2017] seems elegant as stated in that paper. The only way to validate this work compared to [Wang et al., 2017] is to empirically evaluate those two methods in the experiments. Such a comparison experiment is not found and it is a main flaw of this paper.\n\n- Justification of the method\nThe gradients of the softmax cross-entropy loss parameterized with a temperature T are well analyzed in Sections 3.1&3.2. But, in Section 3.3, the reviewer cannot find the clear and convincing explanation for why the temperature T should be increased in the training. My question is: why don't you use alpha=4 consistently throughout the training?\n It might be related to the process of simulated annealing (though \"temperature\" is usually cooled down in SA), and more interestingly, it would also be possible to find connection with the work of [Guo et al., 2017]. In [Guo et al., 2017], the temperature in the softmax is optimized as a post processing for calibrating the classifier outputs. Though the calibration task itself is a little bit apart from the metric learning of the authors' interest, we can find in that paper an interesting result that the temperature is heated up to increase the confidence of the classifier outputs, which is quite similar to the process of fine-tuning by heating up softmax as done in this work. Therefore, the reviewer guesses that the effectiveness of heating up softmax can also be interpreted from the viewpoint of [Guo et al., 2017].\n\nThere is also less description about Figure 1; in particular, the reviewer cannot understand what Figure 1(d) means.\n\n- Ablation study\nTo empirically resolve the above concerns, it is necessary to present the empirical comparison with the \"static\" softmax.\nNamely, the methods of HLN/HBN should be carefully compared to LN/BN of \"alpha=4\", not only those of alpha=16 shown in Table 1&2; the comparison in Table 3 seems unfair since the authors apply the static softmax without normalization.\nAnd, it would be better to show the performance of heated-up softmax \"without\" normalization to show the important role of the normalization, as done in [Wang et al., 2017].\nIn summary, since the proposed method is composed of a heating-up approach and feature normalization, the authors are required to validate the method from those two aspects, respectively, for increasing the significance of this paper.\n\n- Other loss function\nFor achieving a compactness in feature representation, the simple softmax requires both temperature and normalization. It, however, is also conceivable to employ the other types of loss function for that purpose, such as [a] which is based on the (Mahalanobis) distance with taking into account the margin between categories. The distance based loss also embeds features into localized clusters, which satisfies the authors' objective in this work. To validate the proposed method, it is required to compare the method with such a different types of loss function.\n\n[a] Wan, W., Zhong, Y., Li, T., & Chen, J. (2018). Rethinking Feature Distribution for Loss Functions in Image Classification, In CVPR2018, pp. 9117–9126.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541171353479}], "openreview_url": "https://openreview.net/forum?id=SkGpW3C5KX", "arxiv_id": "1809.04157", "paper_pdf": "papers/SkGpW3C5KX.pdf", "paper_pdf_sha256": "dce246b2e8149e0d5e2b825e6a6ecaa1211b18ee8c9d9ac76f9fdaf3cc8aaace", "paper_pdf_bytes": 642781, "paper_pdf_source": "openreview", "code_url": "https://github.com/ColumbiaDVMM/Heated_Up_Softmax_Embedding", "code_repository": "ColumbiaDVMM/Heated_Up_Softmax_Embedding", "code_commit": "cb62d28e5faaf7fdb134b31c461125e3fef50d06", "code_archive": "repos/SkGpW3C5KX.zip", "code_archive_sha256": "f2cd6c095403753c7de604ad92b602c7c1554615c0d59f73ad6f703364fdc33e", "code_archive_bytes": 765171, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 951, "github_languages": {"Python": 116371}, "github_archived": false, "github_pushed_at": "2018-09-13T15:13:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/heated-up-softmax-embedding"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vBZXJzFV6x", "year": 2026, "status": "rejected", "title": "Rethinking Gating Mechanism in Sparse MoE: Handling Arbitrary Modality Inputs with Confidence-Guided Gate", "authors": ["Liangwei Nathan Zheng", "Wei Emma Zhang", "Mingyu Guo", "Miao Xu", "Olaf Maennel", "Weitong Chen"], "authorids": ["~Liangwei_Nathan_Zheng1", "~Wei_Emma_Zhang1", "~Mingyu_Guo1", "~Miao_Xu3", "~Olaf_Maennel1", "~Weitong_Chen2"], "authors_source": "OpenReview API", "abstract": "Effectively managing missing modalities is a fundamental challenge in real-world multimodal learning scenarios, where data incompleteness often results from systematic collection errors or sensor failures. Sparse Mixture-of-Experts (SMoE) architectures have the potential to naturally handle multimodal data, with individual experts specializing in different modalities. However, existing SMoE approach often lacks proper ability to handle missing modality, leading to performance degradation and poor generalization in real-world applications. We propose ConfSMoE to introduce a two-stage imputation module to handle the missing modality problem for the SMoE architecture by taking the opinion of experts and reveal the insight of expert collapse from theoretical analysis with strong empirical evidence. Inspired by our theoretical analysis, ConfSMoE propose a novel expert gating mechanism by detaching the softmax routing score to task confidence score w.r.t ground truth signal. This naturally relieves expert collapse without introducing additional load balance loss function. We show that the insights of expert collapse aligns with other gating mechanism such as Gaussian and Laplacian gate. The proposed method is evaluated on four different real world dataset with three distinct experiment settings to conduct comprehensive analysis of ConfSMoE on resistance to missing modality and the impacts of proposed gating mechanism.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "RmDt28wHuH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3268/Reviewer_dEKy"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes ConfSMoE, a Sparse Mixture-of-Experts (SMoE) model designed to mitigate expert collapse and handle missing modalities.  \nThe method combines two components:\n1. Confidence-guided gating (ConfNet): replaces softmax routing with a supervised \"confidence\" score aligned with ground-truth labels, avoiding gradient conflicts from entropy-based load-balancing losses.  \n2. Two-stage imputation: imputes missing modalities first via intra-modality sampling (pre-imputation) and then refines them using sparse cross-attention over available modalities (post-imputation).  \n\nThe authors analyze the gradient cause of expert collapse, argue that load-balancing losses create conflicting gradients, and claim that confidence-guided routing both balances expert usage and improves robustness to incomplete multimodal data. Experiments on MIMIC-III/IV, CMU-MOSI, and CMU-MOSEI show moderate gains in F1/AUC over FuseMoE, FlexMoE, and other baselines.", "review_text": "This paper proposes ConfSMoE, a Sparse Mixture-of-Experts (SMoE) model designed to mitigate expert collapse and handle missing modalities.  \nThe method combines two components:\n1. Confidence-guided gating (ConfNet): replaces softmax routing with a supervised \"confidence\" score aligned with ground-truth labels, avoiding gradient conflicts from entropy-based load-balancing losses.  \n2. Two-stage imputation: imputes missing modalities first via intra-modality sampling (pre-imputation) and then refines them using sparse cross-attention over available modalities (post-imputation).  \n\nThe authors analyze the gradient cause of expert collapse, argue that load-balancing losses create conflicting gradients, and claim that confidence-guided routing both balances expert usage and improves robustness to incomplete multimodal data. Experiments on MIMIC-III/IV, CMU-MOSI, and CMU-MOSEI show moderate gains in F1/AUC over FuseMoE, FlexMoE, and other baselines.", "strengths": "- Provides a clear gradient-based explanation of expert collapse and why load-balance losses cause conflicting updates.  \n- Introduces a supervised confidence gate, which is a fresh idea compared to typical softmax or Laplacian routing.  \n- Includes comprehensive experiments across clinical and multimodal benchmarks.  \n- Visualization of expert usage and attention maps supports interpretability claims.", "weaknesses": "- **W1. Unclear conceptual linkage between collapse mitigation and missing-modality imputation.**  \n  The paper merges two largely independent problems--expert collapse (an optimization issue) and missing-modality imputation (a data-level issue)--without a convincing reason they must be solved together. Each part could stand as a separate paper, and the combined scope dilutes the narrative focus.\n\n- **W2. Limited novelty of the two-stage imputation module.**  \n  The \"pre-imputation + cross-attention refinement\" design closely resembles existing multimodal imputation methods like DrFuse [1], which also separate modality-specific and cross-modal reconstruction. The proposed variant adds little conceptual innovation beyond placing the procedure inside an MoE.  \n\n- **W3. Empirical improvements are modest.**  \n  Reported F1 scores (40-50 %) are low due to dataset imbalance, but even relative gains over baselines are small (1-4 %). Given the added architectural complexity, the empirical evidence feels insufficient to demonstrate clear practical benefit.\n\n- **W4. Missing comparison to closely related gating approaches.**  \n  The paper does not benchmark against recent load-balancing and uncertainty-aware routing methods that share similar motivations, such as Auxiliary-Loss-Free Load Balancing [2]. Without such baselines, the claimed advantages of the confidence-guided gate remain unverified.\n\n- **W5. Reliance on supervised confidence limits generality.**  \n  The gating network learns from ground-truth labels to estimate confidence, restricting applicability to supervised classification tasks. It is unclear how this method would extend to unsupervised or generative MoE settings where such supervision is unavailable.\n\n[1] DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency. https://arxiv.org/abs/2403.06197\n\n[2] Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts. https://arxiv.org/abs/2408.15664", "questions": "If the author could provide clarification on the points in the weakness section and the following additional questions, I will be happy to raise the score.\n1. Can ConfNet operate without ground-truth labels (e.g., using predictive uncertainty as confidence)?  \n2. What is the sensitivity analysis of the sparsity hyperparameter (B) in post-imputation, i.e., Line 300 how sensitive is the model performance to hyperparameter T for top-T entry?  \n3. Could the confidence gate improve MoE models without missing data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes ConfSMoE, a Sparse Mixture-of-Experts (SMoE) model designed to mitigate expert collapse and handle missing modalities.  \nThe method combines two components:\n1. Confidence-guided gating (ConfNet): replaces softmax routing with a supervised \"confidence\" score aligned with ground-truth labels, avoiding gradient conflicts from entropy-based load-balancing losses.  \n2. Two-stage imputation: imputes missing modalities first via intra-modality sampling (pre-imputation) and then refines them using sparse cross-attention over available modalities (post-imputation).  \n\nThe authors analyze the gradient cause of expert collapse, argue that load-balancing losses create conflicting gradients, and claim that confidence-guided routing both balances expert usage and improves robustness to incomplete multimodal data. Experiments on MIMIC-III/IV, CMU-MOSI, and CMU-MOSEI show moderate gains in F1/AUC over FuseMoE, FlexMoE, and other baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Provides a clear gradient-based explanation of expert collapse and why load-balance losses cause conflicting updates.  \n- Introduces a supervised confidence gate, which is a fresh idea compared to typical softmax or Laplacian routing.  \n- Includes comprehensive experiments across clinical and multimodal benchmarks.  \n- Visualization of expert usage and attention maps supports interpretability claims.", "weaknesses": "- **W1. Unclear conceptual linkage between collapse mitigation and missing-modality imputation.**  \n  The paper merges two largely independent problems--expert collapse (an optimization issue) and missing-modality imputation (a data-level issue)--without a convincing reason they must be solved together. Each part could stand as a separate paper, and the combined scope dilutes the narrative focus.\n\n- **W2. Limited novelty of the two-stage imputation module.**  \n  The \"pre-imputation + cross-attention refinement\" design closely resembles existing multimodal imputation methods like DrFuse [1], which also separate modality-specific and cross-modal reconstruction. The proposed variant adds little conceptual innovation beyond placing the procedure inside an MoE.  \n\n- **W3. Empirical improvements are modest.**  \n  Reported F1 scores (40-50 %) are low due to dataset imbalance, but even relative gains over baselines are small (1-4 %). Given the added architectural complexity, the empirical evidence feels insufficient to demonstrate clear practical benefit.\n\n- **W4. Missing comparison to closely related gating approaches.**  \n  The paper does not benchmark against recent load-balancing and uncertainty-aware routing methods that share similar motivations, such as Auxiliary-Loss-Free Load Balancing [2]. Without such baselines, the claimed advantages of the confidence-guided gate remain unverified.\n\n- **W5. Reliance on supervised confidence limits generality.**  \n  The gating network learns from ground-truth labels to estimate confidence, restricting applicability to supervised classification tasks. It is unclear how this method would extend to unsupervised or generative MoE settings where such supervision is unavailable.\n\n[1] DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency. https://arxiv.org/abs/2403.06197\n\n[2] Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts. https://arxiv.org/abs/2408.15664", "questions": "If the author could provide clarification on the points in the weakness section and the following additional questions, I will be happy to raise the score.\n1. Can ConfNet operate without ground-truth labels (e.g., using predictive uncertainty as confidence)?  \n2. What is the sensitivity analysis of the sparsity hyperparameter (B) in post-imputation, i.e., Line 300 how sensitive is the model performance to hyperparameter T for top-T entry?  \n3. Could the confidence gate improve MoE models without missing data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761698732896}, {"id": "rh5AjhxZat", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3268/Reviewer_8riB"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "In this paper, the authors propose ConfSMoE, a new sparse mixture-of-experts (SMoE) architecture designed to robustly handle missing modalities in multimodal learning. The paper identifies two major limitations of existing SMoE models:\n\n(i) poor robustness under incomplete modality conditions, and\n\n(ii) conventional load-balancing losses exacerbate expert collapse issue (caused by softmax-based router) by introducing gradient conflicts during optimization.\n\nTo address these problems, the authors introduce two key innovations:\n\n1. A confidence-guided gating mechanism (ConfNet) that replaces softmax gating with token-level confidence scores, thereby avoiding gradient conflicts and improving expert diversity without using load-balance losses.\n\n2. A two-stage modality imputation framework that reconstructs missing modalities using intra-modality sampling followed by instance-specific refinement through sparse cross-modal attention.\n\nLastly, the authors perform several experiments across three missing-modality settings to demonstrate the robustness and generalization of the proposed ConfSMoE. Furthermore, an ablation study is also provided to verify the theoretical analysis.", "review_text": "In this paper, the authors propose ConfSMoE, a new sparse mixture-of-experts (SMoE) architecture designed to robustly handle missing modalities in multimodal learning. The paper identifies two major limitations of existing SMoE models:\n\n(i) poor robustness under incomplete modality conditions, and\n\n(ii) conventional load-balancing losses exacerbate expert collapse issue (caused by softmax-based router) by introducing gradient conflicts during optimization.\n\nTo address these problems, the authors introduce two key innovations:\n\n1. A confidence-guided gating mechanism (ConfNet) that replaces softmax gating with token-level confidence scores, thereby avoiding gradient conflicts and improving expert diversity without using load-balance losses.\n\n2. A two-stage modality imputation framework that reconstructs missing modalities using intra-modality sampling followed by instance-specific refinement through sparse cross-modal attention.\n\nLastly, the authors perform several experiments across three missing-modality settings to demonstrate the robustness and generalization of the proposed ConfSMoE. Furthermore, an ablation study is also provided to verify the theoretical analysis.", "strengths": "1. Originality: The confidence-guided gate is a promising alternative for softmax routing. It reduces the “rich-get-richer” feedback loop responsible for expert collapse while remaining interpretable and supervision-aligned.\n\n2. Soundness: The paper provides theoretical derivation for illustrating the expert collapse and the gradient conflicts induced by existing load balancing losses.\n\n3. Relevance: The problem of gating design for alleviatiing expert collapse and handling missing modalities is of interest.", "weaknesses": "1. Dependence on supervision: ConfNet’s confidence gating relies on supervised signals (via MSE loss to ground-truth confidence). This limits applicability to unsupervised or weakly labeled multimodal scenarios.\n\n2. Training complexity and cost: The two-stage imputation introduces additional sampling and sparse cross-attention steps that may increase training time. Although Table 14 discusses FLOPs, a clearer runtime or memory comparison would strengthen claims of efficiency.\n\n3. Quality: The theoretical analysis in Section 2 is not good. \n\n- Firstly, in Section 2.1, the MoE in Eq. (1) is not defined well. In particular, the explicit form of softmax router $G(\\textbf{h})$ is not provided. Furthermore, how the top-K experts are selected is not specified, either. \n\n- Secondly, it seems that the Jacobian in Eq. (2) does not take into account the Top-K function. Additionally, sparse MoE is not differentiable in general due to the discontinuity of the Top-K function. Therefore, I'm quite concerned about the validity of Eq. (2).\n\n- Thirdly, the Jacobian of softmax is not accurate. More specifically, softmax weights in MoE should depend on the input $\\textbf{h}$ but the input is not considered when calculating the Jacobian. \n\n4. Clarity: \n\n- The maths presented in Section 3.2 is not rigorous and difficult to follow. In line 273, the notation $M_{m,i}$ is not defined precisely.\n\n- The Gaussian gating is discussed in the paper but the authors do not provide any math formulation or reference for that gating. \n\n- The authors argue that the ConfSMoE helps address the expert collapse but there are not any evidences provided in the paper.\n\n5. Minor issues:\n\n- In line 205, \"Confidence-Guidede\"\n\n- In Eq. (5), $\\sum^{D}y_t\\log(p_t)$", "questions": "1. How sensitive is the performance to the accuracy of confidence supervision? Could miscalibrated confidence estimates lead to suboptimal expert routing?\n\n2. Can the proposed ConfNet be adapted to unsupervised or self-supervised settings where no ground-truth confidence is available?\n\n3. How significant is the additional computational overhead introduced by the two-stage imputation compared to FuseMoE or FlexMoE?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors propose ConfSMoE, a new sparse mixture-of-experts (SMoE) architecture designed to robustly handle missing modalities in multimodal learning. The paper identifies two major limitations of existing SMoE models:\n\n(i) poor robustness under incomplete modality conditions, and\n\n(ii) conventional load-balancing losses exacerbate expert collapse issue (caused by softmax-based router) by introducing gradient conflicts during optimization.\n\nTo address these problems, the authors introduce two key innovations:\n\n1. A confidence-guided gating mechanism (ConfNet) that replaces softmax gating with token-level confidence scores, thereby avoiding gradient conflicts and improving expert diversity without using load-balance losses.\n\n2. A two-stage modality imputation framework that reconstructs missing modalities using intra-modality sampling followed by instance-specific refinement through sparse cross-modal attention.\n\nLastly, the authors perform several experiments across three missing-modality settings to demonstrate the robustness and generalization of the proposed ConfSMoE. Furthermore, an ablation study is also provided to verify the theoretical analysis.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Originality: The confidence-guided gate is a promising alternative for softmax routing. It reduces the “rich-get-richer” feedback loop responsible for expert collapse while remaining interpretable and supervision-aligned.\n\n2. Soundness: The paper provides theoretical derivation for illustrating the expert collapse and the gradient conflicts induced by existing load balancing losses.\n\n3. Relevance: The problem of gating design for alleviatiing expert collapse and handling missing modalities is of interest.", "weaknesses": "1. Dependence on supervision: ConfNet’s confidence gating relies on supervised signals (via MSE loss to ground-truth confidence). This limits applicability to unsupervised or weakly labeled multimodal scenarios.\n\n2. Training complexity and cost: The two-stage imputation introduces additional sampling and sparse cross-attention steps that may increase training time. Although Table 14 discusses FLOPs, a clearer runtime or memory comparison would strengthen claims of efficiency.\n\n3. Quality: The theoretical analysis in Section 2 is not good. \n\n- Firstly, in Section 2.1, the MoE in Eq. (1) is not defined well. In particular, the explicit form of softmax router $G(\\textbf{h})$ is not provided. Furthermore, how the top-K experts are selected is not specified, either. \n\n- Secondly, it seems that the Jacobian in Eq. (2) does not take into account the Top-K function. Additionally, sparse MoE is not differentiable in general due to the discontinuity of the Top-K function. Therefore, I'm quite concerned about the validity of Eq. (2).\n\n- Thirdly, the Jacobian of softmax is not accurate. More specifically, softmax weights in MoE should depend on the input $\\textbf{h}$ but the input is not considered when calculating the Jacobian. \n\n4. Clarity: \n\n- The maths presented in Section 3.2 is not rigorous and difficult to follow. In line 273, the notation $M_{m,i}$ is not defined precisely.\n\n- The Gaussian gating is discussed in the paper but the authors do not provide any math formulation or reference for that gating. \n\n- The authors argue that the ConfSMoE helps address the expert collapse but there are not any evidences provided in the paper.\n\n5. Minor issues:\n\n- In line 205, \"Confidence-Guidede\"\n\n- In Eq. (5), $\\sum^{D}y_t\\log(p_t)$", "questions": "1. How sensitive is the performance to the accuracy of confidence supervision? Could miscalibrated confidence estimates lead to suboptimal expert routing?\n\n2. Can the proposed ConfNet be adapted to unsupervised or self-supervised settings where no ground-truth confidence is available?\n\n3. How significant is the additional computational overhead introduced by the two-stage imputation compared to FuseMoE or FlexMoE?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761454195337}, {"id": "bq5HyJHqU1", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3268/Reviewer_q3sb"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces ConfSMoE, a confidence-guided sparse MoE designed to address the challenge of missing modalities in multimodal learning. Traditional SMoE models rely on softmax-based gating, which often causes expert collapse and poor generalization when inputs are incomplete. To overcome these limitations, ConfSMoE incorporates two key innovations: (1) it proposes a confidence-guided gating mechanism (ConfNet) that replaces softmax routing with token-level confidence scores, enabling stable and interpretable expert selection without requiring load-balancing losses; (2) it introduces a two-stage imputation strategy that reconstructs missing modalities through modality-specific inference followed by instance-level cross-modal refinement, preserving both structural and contextual fidelity. Experiments demonstrate that ConfSMoE outperforms existing methods in robustness and accuracy.", "review_text": "This paper introduces ConfSMoE, a confidence-guided sparse MoE designed to address the challenge of missing modalities in multimodal learning. Traditional SMoE models rely on softmax-based gating, which often causes expert collapse and poor generalization when inputs are incomplete. To overcome these limitations, ConfSMoE incorporates two key innovations: (1) it proposes a confidence-guided gating mechanism (ConfNet) that replaces softmax routing with token-level confidence scores, enabling stable and interpretable expert selection without requiring load-balancing losses; (2) it introduces a two-stage imputation strategy that reconstructs missing modalities through modality-specific inference followed by instance-level cross-modal refinement, preserving both structural and contextual fidelity. Experiments demonstrate that ConfSMoE outperforms existing methods in robustness and accuracy.", "strengths": "1. This paper tackles a crucial problem in multimodal learning: how to effectively handle missing or incomplete modalities in real-world data. This challenge is highly relevant for domains such as healthcare, where data from different sensors or sources are often unavailable or corrupted.\n\n2. The paper provides valuable insights into the root cause of expert collapse in SMoE models through a detailed gradient analysis, revealing that softmax-based gating and load-balancing losses induce conflicting optimization directions.\n\n3. The proposed ConfSMoE framework has made contributions in both confidence-guided gating mechanism and the two-stage imputation strategy, where they demonstrated their unique advantages.", "weaknesses": "The overall contribution of this paper is clear and nontrivial, so my comments on the weakness is relatively marginal.\n\n1. Based on my understanding, the framework does not fully address temporal or causal dependencies among modalities or expert activations, which is a common problem in healthcare/activity monitoring/autonomous driving etc. The current design assumes static modality relationships and focuses primarily on handling missing inputs at the feature level. In dynamic settings such as time-evolving multimodal data or streaming environments, confidence-guided gating and two-stage imputation may not adapt quickly to changing modality relevance. So it is interesting to see further extension of this work on these settings.\n\n2. Although the two-stage imputation is effective in handling inter-modality relationships, it may introduce additional computational overhead and stochasticity into the model training procedure.", "questions": "In line 244, it is not very clear how the “token-level confidence” is computed or propagated through the MoE.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ConfSMoE, a confidence-guided sparse MoE designed to address the challenge of missing modalities in multimodal learning. Traditional SMoE models rely on softmax-based gating, which often causes expert collapse and poor generalization when inputs are incomplete. To overcome these limitations, ConfSMoE incorporates two key innovations: (1) it proposes a confidence-guided gating mechanism (ConfNet) that replaces softmax routing with token-level confidence scores, enabling stable and interpretable expert selection without requiring load-balancing losses; (2) it introduces a two-stage imputation strategy that reconstructs missing modalities through modality-specific inference followed by instance-level cross-modal refinement, preserving both structural and contextual fidelity. Experiments demonstrate that ConfSMoE outperforms existing methods in robustness and accuracy.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper tackles a crucial problem in multimodal learning: how to effectively handle missing or incomplete modalities in real-world data. This challenge is highly relevant for domains such as healthcare, where data from different sensors or sources are often unavailable or corrupted.\n\n2. The paper provides valuable insights into the root cause of expert collapse in SMoE models through a detailed gradient analysis, revealing that softmax-based gating and load-balancing losses induce conflicting optimization directions.\n\n3. The proposed ConfSMoE framework has made contributions in both confidence-guided gating mechanism and the two-stage imputation strategy, where they demonstrated their unique advantages.", "weaknesses": "The overall contribution of this paper is clear and nontrivial, so my comments on the weakness is relatively marginal.\n\n1. Based on my understanding, the framework does not fully address temporal or causal dependencies among modalities or expert activations, which is a common problem in healthcare/activity monitoring/autonomous driving etc. The current design assumes static modality relationships and focuses primarily on handling missing inputs at the feature level. In dynamic settings such as time-evolving multimodal data or streaming environments, confidence-guided gating and two-stage imputation may not adapt quickly to changing modality relevance. So it is interesting to see further extension of this work on these settings.\n\n2. Although the two-stage imputation is effective in handling inter-modality relationships, it may introduce additional computational overhead and stochasticity into the model training procedure.", "questions": "In line 244, it is not very clear how the “token-level confidence” is computed or propagated through the MoE.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761411394983}], "openreview_url": "https://openreview.net/forum?id=vBZXJzFV6x", "arxiv_id": "2505.19525", "paper_pdf": "papers/vBZXJzFV6x.pdf", "paper_pdf_sha256": "2810fd185a75d35414c8ea9bab94e6f341db230cc633f923b5b271c5c80ac8ea", "paper_pdf_bytes": 1442306, "paper_pdf_source": "openreview", "code_url": "https://github.com/IcurasLW/ICML2026-Official-Repository-of-ConfSMoE", "code_repository": "IcurasLW/ICML2026-Official-Repository-of-ConfSMoE", "code_commit": "b73c5167aff2c6d53507ddfee2d1a98b749511e5", "code_archive": "repos/vBZXJzFV6x.zip", "code_archive_sha256": "ea4918a61f89f127285dfa85186390d5b43ca6962b71a8ebed0539349b162589", "code_archive_bytes": 496042, "code_file_count": 26, "code_extensions": {".py": 17, ".sh": 9}, "github_disk_usage_kb": 490, "github_languages": {"Python": 188459, "Shell": 6729}, "github_archived": false, "github_pushed_at": "2026-09-01T02:40:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rethinking-gating-mechanism-in-sparse-moe"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xSSo8kCA9G", "year": 2025, "status": "rejected", "title": "FRUGAL: Memory-Efficient Optimization by Reducing State Overhead for Scalable Training", "authors": ["Philip Zmushko", "Aleksandr Beznosikov", "Martin Takáč", "Samuel Horváth"], "authorids": ["~Philip_Zmushko1", "~Aleksandr_Beznosikov1", "~Martin_Takáč1", "~Samuel_Horváth1"], "authors_source": "OpenReview API", "abstract": "With the increase in the number of parameters in large language models, the process of pre-training and fine-tuning increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA (Hu et al., 2021)), low-rank gradient projection (GaLore (Zhao\net al., 2024)), and block-wise optimization (BAdam (Luo et al., 2024)) have been proposed. However, in all these algorithms, the effective rank of the weight updates remains low-rank, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce **FRUGAL**; (**F**ull-**R**ank **U**pdates with **G**r**A**dient sp**L**itting),, a new memory-efficient optimization framework. The framework leverages gradient splitting to perform low-rank updates using advanced optimization algorithms (such as Adam), while updates along the remaining directions are\nexecuted via state-free methods like SGD or signSGD. Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when\nusing SGDM for low-rank updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches across various fixed memory budgets, achieving state-of-the-art results in pre-training and fine-tuning\ntasks while balancing memory efficiency and perplexity targets.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "tAned2tIdE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1538/Reviewer_2Mhk"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The work proposes a memory efficient training method called FRUGAL which is essentially a combination of full-rank updates with gradient splitting. The authors partition the parameters and update using advanced optimizers (like Adam) for low-dimensional updates and state-free methods (like SGD or signSGD) for remaining directions. Additionally, the authors provide theoretical convergence guarantees and validate FRUGAL’s effectiveness through experiments on models like LLaMA.", "review_text": "The work proposes a memory efficient training method called FRUGAL which is essentially a combination of full-rank updates with gradient splitting. The authors partition the parameters and update using advanced optimizers (like Adam) for low-dimensional updates and state-free methods (like SGD or signSGD) for remaining directions. Additionally, the authors provide theoretical convergence guarantees and validate FRUGAL’s effectiveness through experiments on models like LLaMA.", "strengths": "1. The combination of state-free optimizers with advanced ones, like SGD and Adam, for memory efficient training is a novel idea.\n2. The  empirical results show that FRUGAL does better than other methods in terms of memory use and perplexity,\n3. The paper includes sufficient ablation studies and it helps to see how FRUGAL works in different situations and settings.", "weaknesses": "Line 249 introduces state-free and stateful parameters but could provide more explicit explanation on the selection criteria. Are parameters randomly selected to each category? In that case the assumption is all the parameters are equally important for that iteration. The work could benefit from more detailed study on how to choose the parameters for state free updates. \n\nThe purpose of the density parameter ($\\rho$) is not thoroughly explained, especially in relation to zero-density training. Please clarify whether zero-density training implies all parameters are state-free (i.e., trained exclusively with SGD). The selection of $\\rho$ is not mentioned in the algorithm as well.", "questions": "GaLore theoretically prove that gradient is low-rank and a study in BlockLLM (https://arxiv.org/pdf/2406.17296)  show that only a few parameters are updated during the training. A few other recent works also seem to suggest that the low rank structure exists in the network. But this paper seems to suggest the opposite. Do you see a space where these two ideas coexist? For example, low rank for certain tasks vs full rank for other tasks? \n\nMinor:\n- Introduce abbreviations for better readability. For example SGD as Stochastic Gradient Descent. \n- Missing references Adam-mini and BlockLLM", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work proposes a memory efficient training method called FRUGAL which is essentially a combination of full-rank updates with gradient splitting. The authors partition the parameters and update using advanced optimizers (like Adam) for low-dimensional updates and state-free methods (like SGD or signSGD) for remaining directions. Additionally, the authors provide theoretical convergence guarantees and validate FRUGAL’s effectiveness through experiments on models like LLaMA.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. The combination of state-free optimizers with advanced ones, like SGD and Adam, for memory efficient training is a novel idea.\n2. The  empirical results show that FRUGAL does better than other methods in terms of memory use and perplexity,\n3. The paper includes sufficient ablation studies and it helps to see how FRUGAL works in different situations and settings.", "weaknesses": "Line 249 introduces state-free and stateful parameters but could provide more explicit explanation on the selection criteria. Are parameters randomly selected to each category? In that case the assumption is all the parameters are equally important for that iteration. The work could benefit from more detailed study on how to choose the parameters for state free updates. \n\nThe purpose of the density parameter ($\\rho$) is not thoroughly explained, especially in relation to zero-density training. Please clarify whether zero-density training implies all parameters are state-free (i.e., trained exclusively with SGD). The selection of $\\rho$ is not mentioned in the algorithm as well.", "questions": "GaLore theoretically prove that gradient is low-rank and a study in BlockLLM (https://arxiv.org/pdf/2406.17296)  show that only a few parameters are updated during the training. A few other recent works also seem to suggest that the low rank structure exists in the network. But this paper seems to suggest the opposite. Do you see a space where these two ideas coexist? For example, low rank for certain tasks vs full rank for other tasks? \n\nMinor:\n- Introduce abbreviations for better readability. For example SGD as Stochastic Gradient Descent. \n- Missing references Adam-mini and BlockLLM", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731133913824}, {"id": "fOyVcjRTtc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1538/Reviewer_YtQB"], "rating": 5, "soundness": 3, "presentation": 1, "contribution": 3, "confidence": 3, "summary": "This paper introduces FRUGAL (Full-Rank Updates with GrAdient spLitting) that reduces memory consumption by splitting gradient updates into two subspaces. A *state-full* subspace is updated using advanced optimization algorithms like Adam, while a *state-free* subspace is updated using stateless and memory-efficient methods like SGD or signSGD. The framework allows for a flexible choice of optimizers and projection methods. FRUGAL achieves state-of-the-art results in pre-training and fine-tuning tasks, outperforming existing memory-efficient algorithms while maintaining a similar memory budget.", "review_text": "This paper introduces FRUGAL (Full-Rank Updates with GrAdient spLitting) that reduces memory consumption by splitting gradient updates into two subspaces. A *state-full* subspace is updated using advanced optimization algorithms like Adam, while a *state-free* subspace is updated using stateless and memory-efficient methods like SGD or signSGD. The framework allows for a flexible choice of optimizers and projection methods. FRUGAL achieves state-of-the-art results in pre-training and fine-tuning tasks, outperforming existing memory-efficient algorithms while maintaining a similar memory budget.", "strengths": "-1) The paper presents a novel approach to improving memory efficiency while performing updates using full-rank information. \n-2) The proposed method is flexible, supporting various choices for both stateful and stateless optimizers as well as different projection methods. It offers convergence guarantees for FRUGAL within a specified framework and consistently outperforms existing memory-efficient algorithms, such as GaLore and BAdam, achieving performance levels close to the memory-intensive Adam optimizer. \n-3) Additionally, the paper provides valuable insights into the learning dynamics of transformer models.", "weaknesses": "-1) The paper's structure would greatly benefit from a clearer organization. Currently, some analysis and experimental results appear within the Methods section, which disrupts the logical flow and makes it challenging for readers to follow the methodology. Reorganizing the paper and dedicating specific sections to distinct aspects of the research could significantly enhance readability and impact.\n\n-2) Several notations (e.g., g~) are introduced without proper definitions, which assumes too much prior knowledge from readers. Additionally, concepts like smoothness and unbiasedness are only vaguely referenced and would benefit from clearer definitions. The theory section should be expanded to explicitly define each notation and assumption, as well as to contextualize them within a more general setting relevant to the proposed method.\n\n-3) Including a full-parameter fine-tuning baseline in Table 4 would provide a valuable benchmark, offering a clearer context for evaluating the results.\n\n-4) Definitions for Full-Rank SVD/Random and Low-Rank SVD/Random are scattered across Table 1 and lack clear differentiation. Consolidating these explanations into a concise paragraph would improve clarity and reader comprehension.\n\n-5) Lastly, there are deviations from the primary algorithm, such as using column-wise projection instead of block-wise projection. For completeness, it would be beneficial to include results using the original proposed approach alongside the variations in the experiments.\n\n-6) By solving this issues in the revision, especially following a more structured writing style and lowering the jumps, the paper would definitely level up.", "questions": "- 1) Including more experiments comparing the method with various stateful and stateless optimizers would enhance the paper. \n\n- 2) Testing models with larger sizes (e.g., 3B and 7B) could further demonstrate the generalizability of the proposed method. \n\n- 3) Please clarify the reasons for selecting the specific optimizers in the theoretical section. They appear restrictive and differ from those used in the main algorithm. Additional details and guarantees would help generalize this proof. \n\n- 4) While it’s mentioned that stateless optimizers typically underperform with transformer architectures, the paper doesn’t explain why FRUGAL with $\\rho=0$ achieves optimal performance in certain scenarios. Providing more details and comparisons would clarify this.\nExpanding the dataset and incorporating diverse architectures could strengthen the argument for FRUGAL's superior characteristics.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces FRUGAL (Full-Rank Updates with GrAdient spLitting) that reduces memory consumption by splitting gradient updates into two subspaces. A *state-full* subspace is updated using advanced optimization algorithms like Adam, while a *state-free* subspace is updated using stateless and memory-efficient methods like SGD or signSGD. The framework allows for a flexible choice of optimizers and projection methods. FRUGAL achieves state-of-the-art results in pre-training and fine-tuning tasks, outperforming existing memory-efficient algorithms while maintaining a similar memory budget.", "soundness": 3, "presentation": 1, "contribution": 3, "strengths": "-1) The paper presents a novel approach to improving memory efficiency while performing updates using full-rank information. \n-2) The proposed method is flexible, supporting various choices for both stateful and stateless optimizers as well as different projection methods. It offers convergence guarantees for FRUGAL within a specified framework and consistently outperforms existing memory-efficient algorithms, such as GaLore and BAdam, achieving performance levels close to the memory-intensive Adam optimizer. \n-3) Additionally, the paper provides valuable insights into the learning dynamics of transformer models.", "weaknesses": "-1) The paper's structure would greatly benefit from a clearer organization. Currently, some analysis and experimental results appear within the Methods section, which disrupts the logical flow and makes it challenging for readers to follow the methodology. Reorganizing the paper and dedicating specific sections to distinct aspects of the research could significantly enhance readability and impact.\n\n-2) Several notations (e.g., g~) are introduced without proper definitions, which assumes too much prior knowledge from readers. Additionally, concepts like smoothness and unbiasedness are only vaguely referenced and would benefit from clearer definitions. The theory section should be expanded to explicitly define each notation and assumption, as well as to contextualize them within a more general setting relevant to the proposed method.\n\n-3) Including a full-parameter fine-tuning baseline in Table 4 would provide a valuable benchmark, offering a clearer context for evaluating the results.\n\n-4) Definitions for Full-Rank SVD/Random and Low-Rank SVD/Random are scattered across Table 1 and lack clear differentiation. Consolidating these explanations into a concise paragraph would improve clarity and reader comprehension.\n\n-5) Lastly, there are deviations from the primary algorithm, such as using column-wise projection instead of block-wise projection. For completeness, it would be beneficial to include results using the original proposed approach alongside the variations in the experiments.\n\n-6) By solving this issues in the revision, especially following a more structured writing style and lowering the jumps, the paper would definitely level up.", "questions": "- 1) Including more experiments comparing the method with various stateful and stateless optimizers would enhance the paper. \n\n- 2) Testing models with larger sizes (e.g., 3B and 7B) could further demonstrate the generalizability of the proposed method. \n\n- 3) Please clarify the reasons for selecting the specific optimizers in the theoretical section. They appear restrictive and differ from those used in the main algorithm. Additional details and guarantees would help generalize this proof. \n\n- 4) While it’s mentioned that stateless optimizers typically underperform with transformer architectures, the paper doesn’t explain why FRUGAL with $\\rho=0$ achieves optimal performance in certain scenarios. Providing more details and comparisons would clarify this.\nExpanding the dataset and incorporating diverse architectures could strengthen the argument for FRUGAL's superior characteristics.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730997029338}, {"id": "rWaZtP0nHq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1538/Reviewer_Cacw"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 2, "summary": "This paper introduces a novel memory-efficient optimization method. Unlike other state-of-the-art approaches, such as LoRA and GaLore, that have low-rank updates, this method maintains a full-rank update structure. The experimental results demonstrate its superior performance, highlighting its potential advantages in both efficiency and effectiveness over competing methods.", "review_text": "This paper introduces a novel memory-efficient optimization method. Unlike other state-of-the-art approaches, such as LoRA and GaLore, that have low-rank updates, this method maintains a full-rank update structure. The experimental results demonstrate its superior performance, highlighting its potential advantages in both efficiency and effectiveness over competing methods.", "strengths": "**Well-structured Presentation:** The paper is well-structured and easy to follow, with a clear presentation of concepts and methodology.\n\n**Practical Impact:** The method is straightforward to implement and has broad applicability, making it valuable for practical use in various settings.", "weaknesses": "**Lack of Discussion on Limitations:** The paper would benefit from a discussion of the method's limitations and potential failure modes. Addressing these aspects would provide a more balanced view of the approach's applicability and constraints.\n\n**Vague Terminology:** Given the importance of \"state-full\" and \"state-free\" in the proposed method, the paper should offer clearer definitions of these terms. Precise terminology is essential to fully understand the mechanics and implications of the approach.", "questions": "**Formal Definitions of Full and Free States:** Could the authors provide formal definitions of \"full\" and \"free\" states as used in the method? A clearer understanding of these terms would improve the paper’s theoretical foundation.\n\n**Main Limitations:** What are the primary limitations of this approach? A discussion on the constraints or situations where the method might be less effective would help clarify its scope and potential trade-offs.\n\n**Running Time Comparisons:** Beyond memory efficiency, how does the method’s running time compare to that of other baseline approaches? Performance in terms of speed is crucial for practical deployment, so direct comparisons would provide a more complete picture of the method’s efficiency.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel memory-efficient optimization method. Unlike other state-of-the-art approaches, such as LoRA and GaLore, that have low-rank updates, this method maintains a full-rank update structure. The experimental results demonstrate its superior performance, highlighting its potential advantages in both efficiency and effectiveness over competing methods.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "**Well-structured Presentation:** The paper is well-structured and easy to follow, with a clear presentation of concepts and methodology.\n\n**Practical Impact:** The method is straightforward to implement and has broad applicability, making it valuable for practical use in various settings.", "weaknesses": "**Lack of Discussion on Limitations:** The paper would benefit from a discussion of the method's limitations and potential failure modes. Addressing these aspects would provide a more balanced view of the approach's applicability and constraints.\n\n**Vague Terminology:** Given the importance of \"state-full\" and \"state-free\" in the proposed method, the paper should offer clearer definitions of these terms. Precise terminology is essential to fully understand the mechanics and implications of the approach.", "questions": "**Formal Definitions of Full and Free States:** Could the authors provide formal definitions of \"full\" and \"free\" states as used in the method? A clearer understanding of these terms would improve the paper’s theoretical foundation.\n\n**Main Limitations:** What are the primary limitations of this approach? A discussion on the constraints or situations where the method might be less effective would help clarify its scope and potential trade-offs.\n\n**Running Time Comparisons:** Beyond memory efficiency, how does the method’s running time compare to that of other baseline approaches? Performance in terms of speed is crucial for practical deployment, so direct comparisons would provide a more complete picture of the method’s efficiency.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730875016718}, {"id": "RG7lGVJTch", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1538/Reviewer_f3bb"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This work proposed a new memory-efficient training methods that allows part of the parameters being optimization with optimization states within a compact space while other parameters are optimizated in the original space without optimization states. Results on serveral pre-training and fine-tuning tasks demonstrates the effectiveness of the proposed methods.", "review_text": "This work proposed a new memory-efficient training methods that allows part of the parameters being optimization with optimization states within a compact space while other parameters are optimizated in the original space without optimization states. Results on serveral pre-training and fine-tuning tasks demonstrates the effectiveness of the proposed methods.", "strengths": "- Plenty of experiments are conducted to evaluate FRUGAL where FRUGAL demonstrates significant improvments against GaLore.\n\n- Both empirical and theoretically justification are provided to validate the effectiveness of FRUGAL.", "weaknesses": "- The GLUE benchmarks is little bit outdated, more recent tasks like common-sense reasoning, mt-bench would further improve this work.\n\n- Is there any explanations about which part of the parameters can be directly optimized with SGD type optimizer with other requires adam and why?\n\n- For $\\rho=0$ in Table 2, is it equals to fully optimized with SGD? Does it controdict with recent works that demonstrates that transformers can not be effectively optimzied with SGD? [1]\n\n- The concepts of state-full and state-free subspace in line80/82 is hard to understand, it's better to formally define these two concepts. \n\n- line 192: \"Surprisingly, we found that although SVD decomposition delivers an initial boost, subsequent training with random projection yields significant improvements\", this sequence make it a little bit confusing whether the \"Low-rank Random\" in Table 1 is training of entire random projection or first SVD and later random.\n\n- it's better to define the meaning of K in the inputs of algorithm 1, as well as s.\n\n\n[1] Why Transformers Need Adam: A Hessian Perspective", "questions": "Please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposed a new memory-efficient training methods that allows part of the parameters being optimization with optimization states within a compact space while other parameters are optimizated in the original space without optimization states. Results on serveral pre-training and fine-tuning tasks demonstrates the effectiveness of the proposed methods.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- Plenty of experiments are conducted to evaluate FRUGAL where FRUGAL demonstrates significant improvments against GaLore.\n\n- Both empirical and theoretically justification are provided to validate the effectiveness of FRUGAL.", "weaknesses": "- The GLUE benchmarks is little bit outdated, more recent tasks like common-sense reasoning, mt-bench would further improve this work.\n\n- Is there any explanations about which part of the parameters can be directly optimized with SGD type optimizer with other requires adam and why?\n\n- For $\\rho=0$ in Table 2, is it equals to fully optimized with SGD? Does it controdict with recent works that demonstrates that transformers can not be effectively optimzied with SGD? [1]\n\n- The concepts of state-full and state-free subspace in line80/82 is hard to understand, it's better to formally define these two concepts. \n\n- line 192: \"Surprisingly, we found that although SVD decomposition delivers an initial boost, subsequent training with random projection yields significant improvements\", this sequence make it a little bit confusing whether the \"Low-rank Random\" in Table 1 is training of entire random projection or first SVD and later random.\n\n- it's better to define the meaning of K in the inputs of algorithm 1, as well as s.\n\n\n[1] Why Transformers Need Adam: A Hessian Perspective", "questions": "Please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730702000350}, {"id": "zgYVX8LMTI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1538/Reviewer_yRQJ"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces FRUGAL. The fundamental idea is that during the backward pass, we will take a subset of parameters (a block) to perform stateful Adam updates and for the rest parameters (with blockwise selection) or the gradient residuals (with low-rank gradient projection), we use stateless signSGD updates. The memory efficiency of FRUGAL is achieved by reducing the optimizer states. The authors provide a convergence rate similar to SGD momentum's usual rate under nonconvex optimization. The authors also perform experiments with Llama pretraining on C4 and RoBerta-base fine-tuning on GLUE tasks. The baselines are primarily Galore and BAdam.", "review_text": "This paper introduces FRUGAL. The fundamental idea is that during the backward pass, we will take a subset of parameters (a block) to perform stateful Adam updates and for the rest parameters (with blockwise selection) or the gradient residuals (with low-rank gradient projection), we use stateless signSGD updates. The memory efficiency of FRUGAL is achieved by reducing the optimizer states. The authors provide a convergence rate similar to SGD momentum's usual rate under nonconvex optimization. The authors also perform experiments with Llama pretraining on C4 and RoBerta-base fine-tuning on GLUE tasks. The baselines are primarily Galore and BAdam.", "strengths": "1. FRUGAL's convergence rate is provided and it can recover the rate of standard SGD(M). \n\n2. The experiment execution is strong and the results are convincing. The hyperparameter details are well disclosed and the implementation is provided.", "weaknesses": "**Major concern**:\n\n1. The idea of FRUGAL is fairly simple (as a combination of signSGD, Adam, and a gradient projector) but the empirical and theoretical support behind FRUGAL is not solid enough. FRUGAL's stateful optimizer is basically either Galore or BAdam. The main contribution is therefore stateless optimizer part (signSGD), and such effectivenss relies on the finding that stateless optimizers are sufficient to optimize most parameters in LLM (linear weight matrices). The authors only provide a single ablation study in Table 3 without further empirical or theoretical insights on the stateless optimizer part. This evidence alone is not convincing enough on an assured generalization to other non-Llama architectures. So it appears to me that the contribution of this paper is insufficient for an ICLR paper. \n\n2. The motivation (Figure 2) of FRUGAL is that low-rank gradient projection is similar, and random or blockwise selection can cover the whole space. Figure 2 justifies that the top gradient directions across timestep is similar, but *is insufficient to show that random or blockwise selection is always/necessarily better. It is highly likely that after a certain threshold, the role of randomly selected parameters/blocks of parameters have worse performance than top gradient directions. An ablation study on projector type versus stateful optimization density $\\rho$ is definitely needed.\n\n**Minor concern**:\n\n1. The presentation of the Algorithm needs to be clearer. It is hard to understand the exact algorithm (which is actually simple) in the first time of reading Algorithm 1 and Section 3.\n\n\nI consider the first major weakness as critical and I would vote for a borderline reject score at this moment.", "questions": "I don't have other questions. All major weaknesses are listed above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces FRUGAL. The fundamental idea is that during the backward pass, we will take a subset of parameters (a block) to perform stateful Adam updates and for the rest parameters (with blockwise selection) or the gradient residuals (with low-rank gradient projection), we use stateless signSGD updates. The memory efficiency of FRUGAL is achieved by reducing the optimizer states. The authors provide a convergence rate similar to SGD momentum's usual rate under nonconvex optimization. The authors also perform experiments with Llama pretraining on C4 and RoBerta-base fine-tuning on GLUE tasks. The baselines are primarily Galore and BAdam.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. FRUGAL's convergence rate is provided and it can recover the rate of standard SGD(M). \n\n2. The experiment execution is strong and the results are convincing. The hyperparameter details are well disclosed and the implementation is provided.", "weaknesses": "**Major concern**:\n\n1. The idea of FRUGAL is fairly simple (as a combination of signSGD, Adam, and a gradient projector) but the empirical and theoretical support behind FRUGAL is not solid enough. FRUGAL's stateful optimizer is basically either Galore or BAdam. The main contribution is therefore stateless optimizer part (signSGD), and such effectivenss relies on the finding that stateless optimizers are sufficient to optimize most parameters in LLM (linear weight matrices). The authors only provide a single ablation study in Table 3 without further empirical or theoretical insights on the stateless optimizer part. This evidence alone is not convincing enough on an assured generalization to other non-Llama architectures. So it appears to me that the contribution of this paper is insufficient for an ICLR paper. \n\n2. The motivation (Figure 2) of FRUGAL is that low-rank gradient projection is similar, and random or blockwise selection can cover the whole space. Figure 2 justifies that the top gradient directions across timestep is similar, but *is insufficient to show that random or blockwise selection is always/necessarily better. It is highly likely that after a certain threshold, the role of randomly selected parameters/blocks of parameters have worse performance than top gradient directions. An ablation study on projector type versus stateful optimization density $\\rho$ is definitely needed.\n\n**Minor concern**:\n\n1. The presentation of the Algorithm needs to be clearer. It is hard to understand the exact algorithm (which is actually simple) in the first time of reading Algorithm 1 and Section 3.\n\n\nI consider the first major weakness as critical and I would vote for a borderline reject score at this moment.", "questions": "I don't have other questions. All major weaknesses are listed above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730346515003}], "openreview_url": "https://openreview.net/forum?id=xSSo8kCA9G", "arxiv_id": "2411.07837", "paper_pdf": "papers/xSSo8kCA9G.pdf", "paper_pdf_sha256": "a39a5fff17c0f5996b516fda3478e72295fd4ef145777e3ce6d50fdd4cc4f093", "paper_pdf_bytes": 608202, "paper_pdf_source": "openreview", "code_url": "https://github.com/fzmushko/FRUGAL", "code_repository": "fzmushko/FRUGAL", "code_commit": "c7ebe4c242f7908e036c287e3a10512a84634648", "code_archive": "repos/xSSo8kCA9G.zip", "code_archive_sha256": "b98977abc6b484b9a5bec1aa6ea04a19057d4b03e4c2b5192d35dacf9e3e5936", "code_archive_bytes": 74185, "code_file_count": 36, "code_extensions": {".sh": 21, ".py": 13, ".ipynb": 2}, "github_disk_usage_kb": 57, "github_languages": {"Python": 178820, "Shell": 17200, "Jupyter Notebook": 12467}, "github_archived": false, "github_pushed_at": "2024-11-13T11:07:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/frugal-memory-efficient-optimization-by"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vlQ56aWJhl", "year": 2024, "status": "rejected", "title": "S-TLLR: STDP-inspired Temporal Local Learning Rule for Spiking Neural Networks", "authors": ["Marco Paul E. Apolinario", "Kaushik Roy"], "authorids": ["~Marco_Paul_E._Apolinario1", "~Kaushik_Roy1"], "authors_source": "OpenReview API", "abstract": "Spiking Neural Networks (SNNs) are biologically plausible models that have been identified as potentially apt for deploying energy-efficient intelligence at the edge, particularly for sequential learning tasks. However, training of SNNs poses significant challenges due to the necessity for precise temporal and spatial credit assignments. Back-propagation through time (BPTT) algorithm, whilst the most widely used method for addressing these issues, incurs a high computational cost due to its temporal dependency. In this work, we propose S-TLLR, a novel three-factor temporal local learning rule inspired by the Spike-Timing Dependent Plasticity (STDP) mechanism, aimed at training deep SNNs on event-based learning tasks. Furthermore, S-TLLR is designed to have low memory and time complexities, which are independent of the number of time steps, rendering it suitable for online learning on low-power edge devices. To demonstrate the scalability of our proposed method, we have conducted extensive evaluations on event-based datasets spanning a wide range of applications, such as image and gesture recognition, audio classification, and optical flow estimation. In all the experiments, S-TLLR achieved high accuracy, comparable to BPTT, with reduction in the number of computations between $1.1-10\\times$.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "u0iEpeEAOG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7662/Reviewer_KFXk"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces a new learning rule for Spiking Neural Networks. This rule has low linear memory complexity and quadratic time complexity in terms of number of neurons. Moreover, the proposed learning algorithm incorporates a non-causal learning term,  inspired by Spike-Timing-Dependent Plasticity.", "review_text": "The paper introduces a new learning rule for Spiking Neural Networks. This rule has low linear memory complexity and quadratic time complexity in terms of number of neurons. Moreover, the proposed learning algorithm incorporates a non-causal learning term,  inspired by Spike-Timing-Dependent Plasticity.", "strengths": "1) Evaluation is done on variety of tasks;\n2) Paper is well-written and easy to follow.", "weaknesses": "My main concern is that the method considered in the paper (S-TLLR) is very similar to OTTT[1]: \n\n1) OTTT has the same learning rule as S-TLLR except that additionally S-TLLR leverages non-causality and few other minor differences. But this non-causal term doesn’t help S-TLLR consistently based on Fig. 2;\n2) S-TLLR has the same time and memory complexity;\n3) S-TLLR doesn’t outperform OTTT. \n\n[1] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online Training Through Time for Spiking Neural Networks, NeurIPS 2022", "questions": "1) Can the authors list all the differences between OTTT with S-TLLR methods?\n2) In the paper, it is mentioned that OTTT applies learning rules at each forward pass, whereas S-TLLR enforces the learning rule at every fourth forward step. Could the authors test the performance if S-TLLR's learning rule was applied at each forward pass, similar to OTTT?\n3) Can the authors do ablation study taking OTTT model as a reference starting point? The study would systematically integrate modifications that transition the model towards the S-TLLR approach.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new learning rule for Spiking Neural Networks. This rule has low linear memory complexity and quadratic time complexity in terms of number of neurons. Moreover, the proposed learning algorithm incorporates a non-causal learning term,  inspired by Spike-Timing-Dependent Plasticity.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1) Evaluation is done on variety of tasks;\n2) Paper is well-written and easy to follow.", "weaknesses": "My main concern is that the method considered in the paper (S-TLLR) is very similar to OTTT[1]: \n\n1) OTTT has the same learning rule as S-TLLR except that additionally S-TLLR leverages non-causality and few other minor differences. But this non-causal term doesn’t help S-TLLR consistently based on Fig. 2;\n2) S-TLLR has the same time and memory complexity;\n3) S-TLLR doesn’t outperform OTTT. \n\n[1] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online Training Through Time for Spiking Neural Networks, NeurIPS 2022", "questions": "1) Can the authors list all the differences between OTTT with S-TLLR methods?\n2) In the paper, it is mentioned that OTTT applies learning rules at each forward pass, whereas S-TLLR enforces the learning rule at every fourth forward step. Could the authors test the performance if S-TLLR's learning rule was applied at each forward pass, similar to OTTT?\n3) Can the authors do ablation study taking OTTT model as a reference starting point? The study would systematically integrate modifications that transition the model towards the S-TLLR approach.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "--", "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698877847564}, {"id": "67EJtdaPHR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7662/Reviewer_H5rp"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces S-TLLR, a novel learning rule for Spiking Neural Networks (SNNs) aimed at efficient online learning on resource-constrained edge devices. S-TLLR draws inspiration from Spike-Timing Dependent Plasticity (STDP) and utilizes both causal and non-causal relationships for synaptic weight updates, maintaining constant memory and time complexity. Through extensive experimentation, the authors demonstrate that S-TLLR achieves comparable accuracy to traditional methods like BPTT but with significantly lower computational demands. The paper's contributions are highlighted by the improved generalization and performance of SNNs on a variety of event-based tasks—including image and gesture recognition, audio classification, and optical flow estimation—and the validation of S-TLLR's efficacy across multiple network topologies, marking a step forward in deploying energy-efficient intelligence in real-world applications.", "review_text": "This paper introduces S-TLLR, a novel learning rule for Spiking Neural Networks (SNNs) aimed at efficient online learning on resource-constrained edge devices. S-TLLR draws inspiration from Spike-Timing Dependent Plasticity (STDP) and utilizes both causal and non-causal relationships for synaptic weight updates, maintaining constant memory and time complexity. Through extensive experimentation, the authors demonstrate that S-TLLR achieves comparable accuracy to traditional methods like BPTT but with significantly lower computational demands. The paper's contributions are highlighted by the improved generalization and performance of SNNs on a variety of event-based tasks—including image and gesture recognition, audio classification, and optical flow estimation—and the validation of S-TLLR's efficacy across multiple network topologies, marking a step forward in deploying energy-efficient intelligence in real-world applications.", "strengths": "1. S-TLLR is a groundbreaking approach that successfully trains SNNs with high efficiency, addressing the temporal and spatial credit assignment challenge that is inherent in such networks.\n2.  By incorporating principles from STDP, S-TLLR aligns closely with biological neural processes, potentially unlocking more natural learning patterns and efficiencies.\n3. S-TLLR successfully integrates both top-down modulation and the local algorithm.\n4. The proposed learning rule maintains constant time and memory complexity, which is a significant advancement for deploying SNNs on edge devices where resources are constrained.", "weaknesses": "1. The complexity was estimated, but the real energy consumption/efficiency haven't been calculated/tested.\n2. While BPTT could work on much deeper SNNs, how about S-TLLR? Could it be extended to larger models/datasets?", "questions": "Please see the weaknesses:\n1. Could the energy consumption/efficiency be calculated/tested.\n2. While BPTT could work on much deeper SNNs, how about S-TLLR? Could it be extended to larger models/datasets, such as CIFAR100?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces S-TLLR, a novel learning rule for Spiking Neural Networks (SNNs) aimed at efficient online learning on resource-constrained edge devices. S-TLLR draws inspiration from Spike-Timing Dependent Plasticity (STDP) and utilizes both causal and non-causal relationships for synaptic weight updates, maintaining constant memory and time complexity. Through extensive experimentation, the authors demonstrate that S-TLLR achieves comparable accuracy to traditional methods like BPTT but with significantly lower computational demands. The paper's contributions are highlighted by the improved generalization and performance of SNNs on a variety of event-based tasks—including image and gesture recognition, audio classification, and optical flow estimation—and the validation of S-TLLR's efficacy across multiple network topologies, marking a step forward in deploying energy-efficient intelligence in real-world applications.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. S-TLLR is a groundbreaking approach that successfully trains SNNs with high efficiency, addressing the temporal and spatial credit assignment challenge that is inherent in such networks.\n2.  By incorporating principles from STDP, S-TLLR aligns closely with biological neural processes, potentially unlocking more natural learning patterns and efficiencies.\n3. S-TLLR successfully integrates both top-down modulation and the local algorithm.\n4. The proposed learning rule maintains constant time and memory complexity, which is a significant advancement for deploying SNNs on edge devices where resources are constrained.", "weaknesses": "1. The complexity was estimated, but the real energy consumption/efficiency haven't been calculated/tested.\n2. While BPTT could work on much deeper SNNs, how about S-TLLR? Could it be extended to larger models/datasets?", "questions": "Please see the weaknesses:\n1. Could the energy consumption/efficiency be calculated/tested.\n2. While BPTT could work on much deeper SNNs, how about S-TLLR? Could it be extended to larger models/datasets, such as CIFAR100?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698832170323}, {"id": "ik2A66ns5H", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7662/Reviewer_FGQ4"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposed an STDP-based learning algorithm that focuses on SNN training from the memory efficient perspective. The proposed algorithm has shown reduced complexity on the event-based dataset.", "review_text": "This paper proposed an STDP-based learning algorithm that focuses on SNN training from the memory efficient perspective. The proposed algorithm has shown reduced complexity on the event-based dataset.", "strengths": "The proposed method shows reduced time complexity, it is natural that the STDP-based learning requires less memory compared to BPTT with gradient surrogation. The proposed algorithm seems hardware friendly with discrete operations.", "weaknesses": "**W1:** Insufficient experiments: I understand that the event-based computer vision tasks are suitable for spiking neural networks, but I think the dataset reported in this paper is not comprehensive enough. In addition to the popular DVS-CIFAR10 and DVS-Gesture, N-CalTech101, and NCARs are also adopted in prior works [R1] as benchmarks. However, these results are missing in the paper. \n\n[R1] AEGNN: Asynchronous Event-Based Graph Neural Networks, CVPR, 2022.\n\n\n**W2:** Since the proposed method claims that the conventional BNTT is memory expensive, it is important to demonstrate the memory-accuracy comparison between the proposed method and BNTT (e.g., GPU Memory) \n\n**W3:** Some recent papers and SoTA methods are not cited in this paper: \n\n[R2]: Temporal Efficient Training of Spiking Neural Network via Gradient Re-weighting\n\n[R2]: Differentiable Spike: Rethinking Gradient-Descent for Training Spiking Neural Networks, NeurIPS'21\n\n[R4]: Training High-Performance Low-Latency Spiking Neural Networks by Differentiation on Spike Representation CVPR 2022\n\n**W4:** The methodology section should be elaborated more. Based on Figure 1, S-TLLR introduces the incoming gradient $\\partial L / \\partial y$ on top of discrete STDP. What is the theoretical advantage (or intuition) of doing that?", "questions": "Please refer to Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed an STDP-based learning algorithm that focuses on SNN training from the memory efficient perspective. The proposed algorithm has shown reduced complexity on the event-based dataset.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The proposed method shows reduced time complexity, it is natural that the STDP-based learning requires less memory compared to BPTT with gradient surrogation. The proposed algorithm seems hardware friendly with discrete operations.", "weaknesses": "**W1:** Insufficient experiments: I understand that the event-based computer vision tasks are suitable for spiking neural networks, but I think the dataset reported in this paper is not comprehensive enough. In addition to the popular DVS-CIFAR10 and DVS-Gesture, N-CalTech101, and NCARs are also adopted in prior works [R1] as benchmarks. However, these results are missing in the paper. \n\n[R1] AEGNN: Asynchronous Event-Based Graph Neural Networks, CVPR, 2022.\n\n\n**W2:** Since the proposed method claims that the conventional BNTT is memory expensive, it is important to demonstrate the memory-accuracy comparison between the proposed method and BNTT (e.g., GPU Memory) \n\n**W3:** Some recent papers and SoTA methods are not cited in this paper: \n\n[R2]: Temporal Efficient Training of Spiking Neural Network via Gradient Re-weighting\n\n[R2]: Differentiable Spike: Rethinking Gradient-Descent for Training Spiking Neural Networks, NeurIPS'21\n\n[R4]: Training High-Performance Low-Latency Spiking Neural Networks by Differentiation on Spike Representation CVPR 2022\n\n**W4:** The methodology section should be elaborated more. Based on Figure 1, S-TLLR introduces the incoming gradient $\\partial L / \\partial y$ on top of discrete STDP. What is the theoretical advantage (or intuition) of doing that?", "questions": "Please refer to Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698641223483}, {"id": "wxPurMCUKL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7662/Reviewer_mZie"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a new learning rule to train spiking neural networks. The idea is based on a three-factor structure, but using BPTT and STDP as its components. The STDP-based eligibility trace function scales with $n$ and is temporally local (does not scale with $T$), which is an improvement over existing methods. This method, referred to by the authors as S-TLLR, has an additional non-causal component which scales with $n$, just like the causal component, and therefore does not affect its scaling with space and time. Experiments and benchmarks on numerous datasets reveal the advantage of this non-causal learning component.\n\nI am generally positive about this work in regards to the new proposed method and how it improves training of spiking networks. I hope that authors can clarify any misunderstandings I may have in the weaknesses section and I am very willing to adjust my score in the rebuttal.", "review_text": "This paper introduces a new learning rule to train spiking neural networks. The idea is based on a three-factor structure, but using BPTT and STDP as its components. The STDP-based eligibility trace function scales with $n$ and is temporally local (does not scale with $T$), which is an improvement over existing methods. This method, referred to by the authors as S-TLLR, has an additional non-causal component which scales with $n$, just like the causal component, and therefore does not affect its scaling with space and time. Experiments and benchmarks on numerous datasets reveal the advantage of this non-causal learning component.\n\nI am generally positive about this work in regards to the new proposed method and how it improves training of spiking networks. I hope that authors can clarify any misunderstandings I may have in the weaknesses section and I am very willing to adjust my score in the rebuttal.", "strengths": "The theoretical scaling advantage is highly relevant and important to the spiking neural network community. Using an STDP-based eligibility trace function also lends to biological plausibility, which has relevance to neuroscience audiences.", "weaknesses": "I am doubtful of both main claims, on (1) temporal locality and (2) improvements from non-causal terms.\n\n(1) The temporal locality property of this method is unconvincing. In Figure 1, my naive understanding is that it is possible to simply truncate both BPTT and STDP methods in the same way S-TLLR is truncated using equation (11). In other words, all methods can have temporal locality. The only way to truly claim that the proposed method does not scale with time, is by using both BPTT and STDP (and perhaps even other existing methods) with this truncation and see if S-TLLR learns faster or if other methods fail to learn the objective. \n\n(2) The improvement from non-causal terms is similarly highly confounded by the secondary activation functions in equations (14-17). Suggestions for fair experiments could be:\n- universally use the same secondary activation function across all tasks, or use all 4 activation functions for all tasks\n- apply the same activation functions to other methods  \n\nTo be very clear, I understand that S-TLLR is compared across different values of $\\alpha$ within the same secondary activation functions, but it is not clear if this behavior is task and function specific. For example, dataset A and secondary function X could give better results with non-zero $\\alpha$, while dataset B with secondary function X or dataset A with secondary function Y has better results with $\\alpha = 0$. \n\n(3) It is also not clear how the method works in the recurrent neural network task. If I were to incorporate causal recurrent gradients in Figure 1, that would correspond to red lines being drawn from $u[t]$ to $y[t-1]$ (and others), which means most terms with have red and blue lines in parallel.  \n\n(4) The recurrent term in equation (1), while true and makes the equation general, simply disappears and lacks coherence and continuity with all future equations where the narrative centers around a feedforward network. For example, equation (4) has no recurrent term. This should be stated in the text somewhere or removed.", "questions": "Should blue terms in Figure 1 also extend beyond $t-2$ (with three dots) just like the red terms?\n\nWhile theoretical scaling arguments are convincing, there are many factors underlying number of computations. How are the 1.1x, 4x and 10x claims actually made? Was it done by recording the number of floating point operations? More information is needed to substantiate these claims. The actual amount of time taken to train the networks is also an important metric to include as well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new learning rule to train spiking neural networks. The idea is based on a three-factor structure, but using BPTT and STDP as its components. The STDP-based eligibility trace function scales with $n$ and is temporally local (does not scale with $T$), which is an improvement over existing methods. This method, referred to by the authors as S-TLLR, has an additional non-causal component which scales with $n$, just like the causal component, and therefore does not affect its scaling with space and time. Experiments and benchmarks on numerous datasets reveal the advantage of this non-causal learning component.\n\nI am generally positive about this work in regards to the new proposed method and how it improves training of spiking networks. I hope that authors can clarify any misunderstandings I may have in the weaknesses section and I am very willing to adjust my score in the rebuttal.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "The theoretical scaling advantage is highly relevant and important to the spiking neural network community. Using an STDP-based eligibility trace function also lends to biological plausibility, which has relevance to neuroscience audiences.", "weaknesses": "I am doubtful of both main claims, on (1) temporal locality and (2) improvements from non-causal terms.\n\n(1) The temporal locality property of this method is unconvincing. In Figure 1, my naive understanding is that it is possible to simply truncate both BPTT and STDP methods in the same way S-TLLR is truncated using equation (11). In other words, all methods can have temporal locality. The only way to truly claim that the proposed method does not scale with time, is by using both BPTT and STDP (and perhaps even other existing methods) with this truncation and see if S-TLLR learns faster or if other methods fail to learn the objective. \n\n(2) The improvement from non-causal terms is similarly highly confounded by the secondary activation functions in equations (14-17). Suggestions for fair experiments could be:\n- universally use the same secondary activation function across all tasks, or use all 4 activation functions for all tasks\n- apply the same activation functions to other methods  \n\nTo be very clear, I understand that S-TLLR is compared across different values of $\\alpha$ within the same secondary activation functions, but it is not clear if this behavior is task and function specific. For example, dataset A and secondary function X could give better results with non-zero $\\alpha$, while dataset B with secondary function X or dataset A with secondary function Y has better results with $\\alpha = 0$. \n\n(3) It is also not clear how the method works in the recurrent neural network task. If I were to incorporate causal recurrent gradients in Figure 1, that would correspond to red lines being drawn from $u[t]$ to $y[t-1]$ (and others), which means most terms with have red and blue lines in parallel.  \n\n(4) The recurrent term in equation (1), while true and makes the equation general, simply disappears and lacks coherence and continuity with all future equations where the narrative centers around a feedforward network. For example, equation (4) has no recurrent term. This should be stated in the text somewhere or removed.", "questions": "Should blue terms in Figure 1 also extend beyond $t-2$ (with three dots) just like the red terms?\n\nWhile theoretical scaling arguments are convincing, there are many factors underlying number of computations. How are the 1.1x, 4x and 10x claims actually made? Was it done by recording the number of floating point operations? More information is needed to substantiate these claims. The actual amount of time taken to train the networks is also an important metric to include as well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698614422396}], "openreview_url": "https://openreview.net/forum?id=vlQ56aWJhl", "arxiv_id": "2306.15220", "paper_pdf": "papers/vlQ56aWJhl.pdf", "paper_pdf_sha256": "1ead20f31432e8866b4336f287e529f4617f55f405f0507643d6c912c1e116da", "paper_pdf_bytes": 1078947, "paper_pdf_source": "openreview", "code_url": "https://github.com/mapolinario94/S-TLLR", "code_repository": "mapolinario94/S-TLLR", "code_commit": "29c575529617f6c384b0a045e5f32ea01102845d", "code_archive": "repos/vlQ56aWJhl.zip", "code_archive_sha256": "a0aa92bb9fe816d8e0c18a2646cca5397b8cd0261b7fa7c5aac89e665acb1c44", "code_archive_bytes": 46660, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 50, "github_languages": {"Python": 169171}, "github_archived": false, "github_pushed_at": "2025-02-14T00:06:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/s-tllr-stdp-inspired-temporal-local-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9qgOs_IwRS3", "year": 2023, "status": "rejected", "title": "Neighborhood Gradient Clustering: An Efficient Decentralized Learning Method for Non-IID Data Distributions", "authors": ["Sai Aparna Aketi", "Sangamesh Kodge", "Kaushik Roy"], "authorids": ["~Sai_Aparna_Aketi1", "~Sangamesh_Kodge1", "~Kaushik_Roy1"], "authors_source": "OpenReview API", "abstract": "Decentralized learning algorithms enable the training of deep learning models over large distributed datasets generated at different devices and locations, without the need for a central server. In practical scenarios, the distributed datasets can have significantly different data distributions across the agents. The current state-of-the-art decentralized algorithms mostly assume the data distributions to be Independent and Identically Distributed (IID). This paper focuses on improving decentralized learning over non-IID data distributions with minimal compute and memory overheads. We propose Neighborhood Gradient Clustering (NGC), a novel decentralized learning algorithm that modifies the local gradients of each agent using self- and cross-gradient information. Cross-gradients for a pair of neighboring agents are the derivatives of the model parameters of an agent with respect to the dataset of the other agent. In particular, the proposed method replaces the local gradients of the model with the weighted mean of the self-gradients, model-variant cross-gradients (derivatives of the received neighbors’ model parameters with respect to the local dataset - computed locally), and data-variant cross-gradients (derivatives of the local model with respect to its neighbors’ datasets - received through communication). The data-variant cross-gradients are aggregated through an additional communication round without breaking the privacy constraints of the decentralized setting. Further, we present CompNGC, a compressed version of NGC that reduces the communication overhead by $32 \\times$ by compressing the cross-gradients. We demonstrate the empirical convergence and efficiency of the proposed technique over non-IID data distributions sampled from the CIFAR-10 dataset on various model architectures and graph topologies. Our experiments demonstrate that NGC and CompNGC outperform the existing state-of-the-art (SoTA) decentralized learning algorithm over non-IID data by $1-5\\%$ with significantly less compute and memory requirements. Further, we also show that the proposed NGC method outperforms the baseline by $5-40\\%$ with no additional communication. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "_tr9Zlred4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1992/Reviewer_y4qJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a new algorithm that utilizes groups of stochastic gradients collected from neighborhood nodes in a decentralized learning environment with heterogeneous data distributed in different locations. The proposed algorithm is communication efficient and achieves better empirical performance on non-IID data sampled from the CIFAR-10 Dataset.", "review_text": "The paper is very well written but the supporting evidence for the model performance is not sufficient.", "strengths": "Strength: the paper is quite well-written and easy to follow.\nWeakness: the working mechanism of the proposed method is not fully demonstrated. Other than the empirical performance on a benchmark data set, it is not clear to me why this method work or under what circumstances the proposed method will work best.\n\n1. The paper has emphasized the concept of \"non-IID\" data. But what types of data can this method handle? It would be very ambitious to claim that it will work for any type of setting. So it is necessary to clarify or at least give some concrete examples of the non-IID data types under consideration.\n\n2. What types of models are most suited for the proposed method? Regression problems? Classification problems? or other tasks? Different models have very different structures on the gradients and one needs to be more specific.\n\n3. The term \"clustering\" in the title implies some sort of unsupervised learning algorithms which I do not see in the paper.  To the best of my understanding, the paper seems to assume that the data distributions among the neighboring nodes are somehow \"homogeneous\". Otherwise, I do not understand why taking a weighted average of a bunch of inhomogeneous gradients will help improve the model's performance.  Please clarify.\n\n4. Since there are no theoretical justifications for the proposed method, I think carefully designed synthetic experiments are necessary to study the advantages and limitations of the proposed method. Otherwise, I have little confidence in the generalizability of the method.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed a new algorithm that utilizes groups of stochastic gradients collected from neighborhood nodes in a decentralized learning environment with heterogeneous data distributed in different locations. The proposed algorithm is communication efficient and achieves better empirical performance on non-IID data sampled from the CIFAR-10 Dataset.", "strength_and_weaknesses": "Strength: the paper is quite well-written and easy to follow.\nWeakness: the working mechanism of the proposed method is not fully demonstrated. Other than the empirical performance on a benchmark data set, it is not clear to me why this method work or under what circumstances the proposed method will work best.\n\n1. The paper has emphasized the concept of \"non-IID\" data. But what types of data can this method handle? It would be very ambitious to claim that it will work for any type of setting. So it is necessary to clarify or at least give some concrete examples of the non-IID data types under consideration.\n\n2. What types of models are most suited for the proposed method? Regression problems? Classification problems? or other tasks? Different models have very different structures on the gradients and one needs to be more specific.\n\n3. The term \"clustering\" in the title implies some sort of unsupervised learning algorithms which I do not see in the paper.  To the best of my understanding, the paper seems to assume that the data distributions among the neighboring nodes are somehow \"homogeneous\". Otherwise, I do not understand why taking a weighted average of a bunch of inhomogeneous gradients will help improve the model's performance.  Please clarify.\n\n4. Since there are no theoretical justifications for the proposed method, I think carefully designed synthetic experiments are necessary to study the advantages and limitations of the proposed method. Otherwise, I have little confidence in the generalizability of the method.\n", "clarity,_quality,_novelty_and_reproducibility": "The presentation is very clear and easy to follow. The proposed methodology is novel but needs a more comprehensive justification.", "summary_of_the_review": "The paper is very well written but the supporting evidence for the model performance is not sufficient.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666727890945}, {"id": "KNlfeQ20fRx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1992/Reviewer_LFjx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the decentralized distributed training under the non-iid data distribution setting. Different from the baseline method D-PSGD, it leverages self- and cross-gradient information to modify the gradient update for each agent. ", "review_text": "The paper proposed an interesting method for decentralized learning but it has several weaknesses. ", "strengths": "Strength\n1. The motivation of this paper is clear and interesting.\n2. The proposed method looks reasonable.\n3. Experiments have some promising results.\n\nWeaknesses\n1. As pointed out in Introduction section, the idea of leveraging cross-gradient information has already been used in Esfandiari et al. (2021). Therefore, the novelty of this paper is limited in this sense.\n2. It would be great to include some convergence analysis in this paper. Many decentralized learning methods have convergence analysis, such as D-PSGD. Actually, Esfandiari et al. (2021) which also uses cross-gradient information, has convergence analysis as well.\n3. The experiments are only conducted on one dataset, i.e., CIFAR10. It’s better to use more datasets. Decentralized learning algorithms are particularly useful for training large models on large distributed datasets. The authors are encouraged to use large datasets (such as ImageNet).\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the decentralized distributed training under the non-iid data distribution setting. Different from the baseline method D-PSGD, it leverages self- and cross-gradient information to modify the gradient update for each agent. ", "strength_and_weaknesses": "Strength\n1. The motivation of this paper is clear and interesting.\n2. The proposed method looks reasonable.\n3. Experiments have some promising results.\n\nWeaknesses\n1. As pointed out in Introduction section, the idea of leveraging cross-gradient information has already been used in Esfandiari et al. (2021). Therefore, the novelty of this paper is limited in this sense.\n2. It would be great to include some convergence analysis in this paper. Many decentralized learning methods have convergence analysis, such as D-PSGD. Actually, Esfandiari et al. (2021) which also uses cross-gradient information, has convergence analysis as well.\n3. The experiments are only conducted on one dataset, i.e., CIFAR10. It’s better to use more datasets. Decentralized learning algorithms are particularly useful for training large models on large distributed datasets. The authors are encouraged to use large datasets (such as ImageNet).\n", "clarity,_quality,_novelty_and_reproducibility": "Look good to me.", "summary_of_the_review": "The paper proposed an interesting method for decentralized learning but it has several weaknesses. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666684796547}, {"id": "y0N6pfW0CIW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1992/Reviewer_b2jc"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a new decentralized learning algorithm that bases on the proposed cross gradient aggregation (CGA) algorithm. Specifically, the authors leverage the existing self-gradient and cross-gradient concepts to develop the neighborhood gradient clustering (NGC) algorithm. This proposed method replaces the local gradients of the model with the weighted mean of the self-gradients, model-variant cross-gradients, and data-variant cross-gradients, which is better able to handle the parameter variations in non-IID scenarios. To reduce the communication bottleneck, the authors develop the compressed version of NGC, CompNGC. To validate both algorithms, the authors utilize a benchmark dataset with various model architectures and the results show the superiority of the proposed algorithms.", "review_text": "This paper presents a new decentralized learning algorithm called NGC to cope with the issues in decentralized learning settings. The authors provide algorithm design and empirical evidence to show case the superiority. This work seems incremental, failing to providing good novelties. The authors should analyze the proposed algorithms theoretically and present more experimental result to validate.", "strengths": "Strength: this paper presents a novel algorithm on top of an existing one and shows the better model performance compared to the existing baseline. The investigated topic is quite interesting as non-IID remains a challenging problem in the decentralized learning area. The paper is easy to follow and the presentation is clear.\n\nWeaknesses: the novelty in this work is quite marginal. It looks the only novelty is to replace the gradient calculation step in CGA with the weighted mean of self-gradients, model-variant cross-gradients, and data-variant cross-gradients. Such a change is just a minor incremental work, which fails to provide sufficient contributions. Also, concepts are just based on the existing works. Given this, if the authors could provide detailed theoretical analysis on NGC and CompNGC, that may make this paper technically solid and sound. However, the authors only provide empirical results, which to me is not promising as well. If this paper is posited as an applied paper, then comprehensive results are necessarily required, including diverse model architectures, datasets, tasks, topologies and numbers of agents. Through these, if NGC and CompNGC remain competitive and outperform baselines, then the work will look very technically solid and sound. However, the existing experimental results to me are not very comprehensive and promising. \n\n*******************************Post-rebuttal************************************\nI appreciate the detailed response and additional results in the revised paper from the authors. After carefully reviewing the responses, I think the additional results have made the work better. However, the overall novelties in this work is still low to me and it still has much room for improvement in terms of theory. Though I would raise my score given the much better empirical evidences for the proposed algorithm in the revised version, the paper with the current form is still marginally below the acceptance standard. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents a new decentralized learning algorithm that bases on the proposed cross gradient aggregation (CGA) algorithm. Specifically, the authors leverage the existing self-gradient and cross-gradient concepts to develop the neighborhood gradient clustering (NGC) algorithm. This proposed method replaces the local gradients of the model with the weighted mean of the self-gradients, model-variant cross-gradients, and data-variant cross-gradients, which is better able to handle the parameter variations in non-IID scenarios. To reduce the communication bottleneck, the authors develop the compressed version of NGC, CompNGC. To validate both algorithms, the authors utilize a benchmark dataset with various model architectures and the results show the superiority of the proposed algorithms.", "strength_and_weaknesses": "Strength: this paper presents a novel algorithm on top of an existing one and shows the better model performance compared to the existing baseline. The investigated topic is quite interesting as non-IID remains a challenging problem in the decentralized learning area. The paper is easy to follow and the presentation is clear.\n\nWeaknesses: the novelty in this work is quite marginal. It looks the only novelty is to replace the gradient calculation step in CGA with the weighted mean of self-gradients, model-variant cross-gradients, and data-variant cross-gradients. Such a change is just a minor incremental work, which fails to provide sufficient contributions. Also, concepts are just based on the existing works. Given this, if the authors could provide detailed theoretical analysis on NGC and CompNGC, that may make this paper technically solid and sound. However, the authors only provide empirical results, which to me is not promising as well. If this paper is posited as an applied paper, then comprehensive results are necessarily required, including diverse model architectures, datasets, tasks, topologies and numbers of agents. Through these, if NGC and CompNGC remain competitive and outperform baselines, then the work will look very technically solid and sound. However, the existing experimental results to me are not very comprehensive and promising. \n\n*******************************Post-rebuttal************************************\nI appreciate the detailed response and additional results in the revised paper from the authors. After carefully reviewing the responses, I think the additional results have made the work better. However, the overall novelties in this work is still low to me and it still has much room for improvement in terms of theory. Though I would raise my score given the much better empirical evidences for the proposed algorithm in the revised version, the paper with the current form is still marginally below the acceptance standard. ", "clarity,_quality,_novelty_and_reproducibility": "The clarity and quality of the presentation in this paper look good, but the big weakness as mentioned above is the novelty, which is quite marginal. The current form of the paper cannot provide sufficient contributions. ", "summary_of_the_review": "This paper presents a new decentralized learning algorithm called NGC to cope with the issues in decentralized learning settings. The authors provide algorithm design and empirical evidence to show case the superiority. This work seems incremental, failing to providing good novelties. The authors should analyze the proposed algorithms theoretically and present more experimental result to validate.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666476314578}], "openreview_url": "https://openreview.net/forum?id=9qgOs_IwRS3", "arxiv_id": "2209.14390", "paper_pdf": "papers/9qgOs_IwRS3.pdf", "paper_pdf_sha256": "43589f0371a8d45a93eac13d7e6cd8501bcd785a0b3d67df34fe76ac85c2405d", "paper_pdf_bytes": 696053, "paper_pdf_source": "openreview", "code_url": "https://github.com/aparna-aketi/neighborhood_gradient_clustering", "code_repository": "aparna-aketi/neighborhood_gradient_clustering", "code_commit": "2e6e3676f335b79dcb2cb53722c4f8cd80073862", "code_archive": "repos/9qgOs_IwRS3.zip", "code_archive_sha256": "9b401b95185efe31d3b0fe305134369596f881a1bc24e100c2306646816b19ec", "code_archive_bytes": 64907, "code_file_count": 38, "code_extensions": {".py": 37, ".sh": 1}, "github_disk_usage_kb": 82, "github_languages": {"Python": 181638, "Shell": 1825}, "github_archived": false, "github_pushed_at": "2023-11-15T01:31:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neighborhood-gradient-clustering-an-efficient"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WN2Sup7qLdw", "year": 2022, "status": "rejected", "title": "Multi-Resolution Continuous Normalizing Flows", "authors": ["Vikram Voleti", "Chris Finlay", "Adam M Oberman", "Christopher Pal"], "authorids": ["~Vikram_Voleti1", "~Chris_Finlay1", "~Adam_M_Oberman1", "~Christopher_Pal1"], "authors_source": "OpenReview API", "abstract": "Recent work has shown that Neural Ordinary Differential Equations (ODEs) can serve as generative models of images using the perspective of Continuous Normalizing Flows (CNFs). Such models offer exact likelihood calculation, and invertible generation/density estimation. In this work we introduce a Multi-Resolution variant of such models (MRCNF), by characterizing the conditional distribution over the additional information required to generate a fine image that is consistent with the coarse image. We introduce a transformation between resolutions that allows for no change in the log likelihood. We show that this approach yields comparable likelihood values for various image datasets, using orders of magnitude fewer parameters than the prior methods, in significantly less training time, using only one GPU.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "t9TLN3PPkpO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2172/Reviewer_hXQn"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper describes the novel approach to generate high-resolution images in progressive manner. The approach is based on set of conditional flows that take the image generated from previous stage as conditioning factor and generate the image with the higher resolution. The quality of the method is compared to the reference approaches, mainly focusing on WaveletFlow as the most similar method.", "review_text": "Strengths\n\nThe problem considered in this paper is challenging and the approach to solve it (upscaling via conditional flows) is adequate to the problem. The improvements compared to WaveletFlow seems to work in practice what can be observed while analysing the quantitative results (BPD). It is also very beneficial for that work that the authors analyse also training time and complexity of the model during the experimental stage. The OOD analysis (unfortunately provided in Supplementary) is also interesting in terms of possible other applications of the model. It is also important to mention, that the code is provided in the suplement, so reproducing the results should be easy. \n\nWeaknesses \n\nMy major concern is about the contribution of the paper. The first two points in introduction that point out the contribution are non-informative in my opinion. The first refers to the transformation that does not add the cost to the training objective. Why is it so important for the model to not modify the likelihood cost with that transformation? If the (log) cost is negative, it can reduce the likelihood value while testing. \n\nThe second contribution is simply the name of the model and should be removed. Even for the third contribution I have the problem of understanding what components of the model are responsible for time and model capacity reduction. In my opinion the contribution of the paper is based on the different base flow model (CNF) and using unimodular transformation in eq. (7).  This contribution is rather incremental with respect to the WaveletFlow. Using CNF seems to be a good direction but is not motivated taking into account that any conditional flow can be used instead. The new transformation function does not provide a significant improvement compared to WaveletFlow and is motivated by reduction of the cost of transformation not the quality of model itself. \n\nThe second concern is about the provided results. The Table 1 is incomplete, some results are missing for reference approaches, it is sparse, and therefore, difficult to follow and analyse. The authors do not provide the qualitative results for high-resolution images as in WaveletFlow - such empirical would be beneficial to show the gain of using the approach provided in the paper. The selection of BPD criterion for quantitative evaluation is ok, but the goal of multi-resolution generative model is to generate good-quality images for larger dimensions. Using some standard measures like FID or related and comparison to GAN models would be beneficial for this case.    \n\nThe third concern is about the quality of the paper - the provided manuscript is difficult to follow, the contribution is not expressed well and the description of the method should be more clear. I do not understand, why it is important to optimize BPD  directly instead of likelihood? For me it is just the scaling issue that does not affect the training procedure. Some interesting parts like OOD analysis are moved to suplement. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper describes the novel approach to generate high-resolution images in progressive manner. The approach is based on set of conditional flows that take the image generated from previous stage as conditioning factor and generate the image with the higher resolution. The quality of the method is compared to the reference approaches, mainly focusing on WaveletFlow as the most similar method.", "main_review": "Strengths\n\nThe problem considered in this paper is challenging and the approach to solve it (upscaling via conditional flows) is adequate to the problem. The improvements compared to WaveletFlow seems to work in practice what can be observed while analysing the quantitative results (BPD). It is also very beneficial for that work that the authors analyse also training time and complexity of the model during the experimental stage. The OOD analysis (unfortunately provided in Supplementary) is also interesting in terms of possible other applications of the model. It is also important to mention, that the code is provided in the suplement, so reproducing the results should be easy. \n\nWeaknesses \n\nMy major concern is about the contribution of the paper. The first two points in introduction that point out the contribution are non-informative in my opinion. The first refers to the transformation that does not add the cost to the training objective. Why is it so important for the model to not modify the likelihood cost with that transformation? If the (log) cost is negative, it can reduce the likelihood value while testing. \n\nThe second contribution is simply the name of the model and should be removed. Even for the third contribution I have the problem of understanding what components of the model are responsible for time and model capacity reduction. In my opinion the contribution of the paper is based on the different base flow model (CNF) and using unimodular transformation in eq. (7).  This contribution is rather incremental with respect to the WaveletFlow. Using CNF seems to be a good direction but is not motivated taking into account that any conditional flow can be used instead. The new transformation function does not provide a significant improvement compared to WaveletFlow and is motivated by reduction of the cost of transformation not the quality of model itself. \n\nThe second concern is about the provided results. The Table 1 is incomplete, some results are missing for reference approaches, it is sparse, and therefore, difficult to follow and analyse. The authors do not provide the qualitative results for high-resolution images as in WaveletFlow - such empirical would be beneficial to show the gain of using the approach provided in the paper. The selection of BPD criterion for quantitative evaluation is ok, but the goal of multi-resolution generative model is to generate good-quality images for larger dimensions. Using some standard measures like FID or related and comparison to GAN models would be beneficial for this case.    \n\nThe third concern is about the quality of the paper - the provided manuscript is difficult to follow, the contribution is not expressed well and the description of the method should be more clear. I do not understand, why it is important to optimize BPD  directly instead of likelihood? For me it is just the scaling issue that does not affect the training procedure. Some interesting parts like OOD analysis are moved to suplement. ", "summary_of_the_review": "Unfortunately, the paper is rather incremental with respect to WaveletFlow, the contribution is not expressed well and qualitative analysis is missing. I would like to encourage authors for more extensive empirical evaluation and highlighting the novelty of the approach. Therefore, my current recommendation is to reject this work. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635885343117}, {"id": "ViCLhHxnvzO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2172/Reviewer_K92k"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a multi-resolution variant of continuous normalizing flows for images. They show empirical results for image sizes upto 64 x 64 where they seem to be better than regular continuous normalizing flows. The key proposed benefit seems to be in number of parameters and training times.", "review_text": "Comments/ questions: \n- The paper is written well and easy to understand.\n- The contributions do not seem to be significant enough yet from either a theoretical or practical perspective. I would request the authors to investigate further and work with larger image sizes (256 x 256 Glow paper). Perhaps more insight into why flows are worse on OOD from a multiscale perspective would be a good addition. Can we check if the problem is worse or better at coarser scales? Are constant images consistently assigned higher likelihoods at all image scales? If so, can't we regularize for this? When shuffling images, again use the multi-scale ideas and check which scales are affected most and how that is related to the tile size used in shuffling. I think these would be interesting and insightful additions to the paper.\n- The main advantage of the method seems to be training times and number of parameters in a network which is good but there are other papers out there now that claim similar advantages with injective flows (Trumpets, UAI '21 Kothari et al, M-flows, NeurIPS' 20 Cranmer et al).  Rather than being yet another paper in the same direction, I think the author's would benefit greatly from investigating their key ideas a bit more. For example, a more in-depth analysis of noise injection at various scales and how that affects image generation would be good to have. A simple analysis could show how posterior samples of $x_s$ vary given, say $y_s, y_{s-1}$  and $x_{s}$. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a multi-resolution variant of continuous normalizing flows for images. They show empirical results for image sizes upto 64 x 64 where they seem to be better than regular continuous normalizing flows. The key proposed benefit seems to be in number of parameters and training times.", "main_review": "Comments/ questions: \n- The paper is written well and easy to understand.\n- The contributions do not seem to be significant enough yet from either a theoretical or practical perspective. I would request the authors to investigate further and work with larger image sizes (256 x 256 Glow paper). Perhaps more insight into why flows are worse on OOD from a multiscale perspective would be a good addition. Can we check if the problem is worse or better at coarser scales? Are constant images consistently assigned higher likelihoods at all image scales? If so, can't we regularize for this? When shuffling images, again use the multi-scale ideas and check which scales are affected most and how that is related to the tile size used in shuffling. I think these would be interesting and insightful additions to the paper.\n- The main advantage of the method seems to be training times and number of parameters in a network which is good but there are other papers out there now that claim similar advantages with injective flows (Trumpets, UAI '21 Kothari et al, M-flows, NeurIPS' 20 Cranmer et al).  Rather than being yet another paper in the same direction, I think the author's would benefit greatly from investigating their key ideas a bit more. For example, a more in-depth analysis of noise injection at various scales and how that affects image generation would be good to have. A simple analysis could show how posterior samples of $x_s$ vary given, say $y_s, y_{s-1}$  and $x_{s}$. \n\n", "summary_of_the_review": "I think that the current draft of this work makes incremental contributions to the area of normalizing flows for image synthesis. While I completely agree with the core motivations of bringing multi-resolution structures and scale-separated noise injection to CNFs, I believe the authors can make a much more significant contribution by analyzing how their core motivations play out in their proposed method. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635834522496}, {"id": "qnG0_vQfPjX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2172/Reviewer_cSSK"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors proposed a multi-resolution version continuous normalizing flows with the claim of faster training time resulting in comparable performance. The authors note a few key differences with previous work (WaveletFlow) - using CNF instead of realnvp / glow based architectures, modeling noise at multi-resolution, and utilizing a volume and range preserving formulation, ", "review_text": "Overall, my general take is that the paper is relatively weak on both conceptual novelty as well as application performance. My main findings are below.\n\nCons:\n\nConceptual Novelty:\nThough the paper tried to differentiate from prior work (WaveletFlow) in various aspects (using CNF rather than RealNVP, utilizing a volume and range preserving formulation), the main idea of multi-resolution conditional modeling of images at the next resolution based on the previous, is very much the same. In fact, I don't think the authors did a good job at justifying the few differentiating design factors from WaveletFlow. For instance, the authors claimed that the volume-preserving and range-preserving transformations are a key novelty but the effects of this transformation (as opposed to the wavelet transforms) are not studies in an experiment.\n\nExperiment Results:\nFrom the experiment results (Table 1), it's not clear that (1) the multiresolution formulation is more efficient, or (2) the generative modeling quality is improved, both of which I think are central to the claims of the paper. \n- First, the BPD metric seems to get worse with more resolutions, which seems contrary to the central story?\n- Second, even though the authors claimed in the paper that the model is \"significantly faster with on-par performance\", the performance seems quite a bit worse than even very early methods like RealNVP. While the model size seems smaller, that seems more likely be due to the use of FFJORD/CNF, rather than the multi-resolution formulation, which is an existing architecture the authors adopted, rather than a contribution of this paper. \n- Third, while model size seems to be a main selling point of the story, being \"smaller/cheaper with lower performance\" does not feel very compelling. Very likely than not, when the \"large\" models are reduced using simple methods like reducing the number of feature channels / number of stacked flows, they can slim down to the same size with comparable or even performance.\n\nPros:\n\nI do like that the authors are quite describing the mathematical details of the proposed multi-resolution transformations (Sec. 3.1, 3.2) which are easy to follow and enjoyable to read.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors proposed a multi-resolution version continuous normalizing flows with the claim of faster training time resulting in comparable performance. The authors note a few key differences with previous work (WaveletFlow) - using CNF instead of realnvp / glow based architectures, modeling noise at multi-resolution, and utilizing a volume and range preserving formulation, ", "main_review": "Overall, my general take is that the paper is relatively weak on both conceptual novelty as well as application performance. My main findings are below.\n\nCons:\n\nConceptual Novelty:\nThough the paper tried to differentiate from prior work (WaveletFlow) in various aspects (using CNF rather than RealNVP, utilizing a volume and range preserving formulation), the main idea of multi-resolution conditional modeling of images at the next resolution based on the previous, is very much the same. In fact, I don't think the authors did a good job at justifying the few differentiating design factors from WaveletFlow. For instance, the authors claimed that the volume-preserving and range-preserving transformations are a key novelty but the effects of this transformation (as opposed to the wavelet transforms) are not studies in an experiment.\n\nExperiment Results:\nFrom the experiment results (Table 1), it's not clear that (1) the multiresolution formulation is more efficient, or (2) the generative modeling quality is improved, both of which I think are central to the claims of the paper. \n- First, the BPD metric seems to get worse with more resolutions, which seems contrary to the central story?\n- Second, even though the authors claimed in the paper that the model is \"significantly faster with on-par performance\", the performance seems quite a bit worse than even very early methods like RealNVP. While the model size seems smaller, that seems more likely be due to the use of FFJORD/CNF, rather than the multi-resolution formulation, which is an existing architecture the authors adopted, rather than a contribution of this paper. \n- Third, while model size seems to be a main selling point of the story, being \"smaller/cheaper with lower performance\" does not feel very compelling. Very likely than not, when the \"large\" models are reduced using simple methods like reducing the number of feature channels / number of stacked flows, they can slim down to the same size with comparable or even performance.\n\nPros:\n\nI do like that the authors are quite describing the mathematical details of the proposed multi-resolution transformations (Sec. 3.1, 3.2) which are easy to follow and enjoyable to read.", "summary_of_the_review": "Overall, I enjoyed reading the paper and I really like the mutliresolution idea that leverages conditional probability modeling, but I think the claimed main conceptual novelties are weak and not well supported by experimental results. Therefore I cannot recommend acceptance at ICLR.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635826191622}, {"id": "bfRJNJIszai", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2172/Reviewer_efZN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the authors propose a multi-resolution strategy for continuous normalizing flows. The proposed approach consists of a general wavelet based decomposition/downsampling where the transformation obeys some useful mathematical properties such as volume and range preservation. Empirical results indicate likelihood estimation performance better or on par with benchmarked methods. The proposed model also trains significantly faster with fewer parameters.", "review_text": "Strengths:\n-The authors provide a faster multi-resolution strategy for normalizing flows.\n-The empirical results are better than the benchmarked methods, e.g. WaveletFlow [1], Glow [2], FFJORD [3].\n-The paper is clearly written and easy to read.\n-Authors provide detailed descriptions and the source code thus the work should be easier to reproduce.\n\nWeaknesses:\n-Comparisons to some of the benchmarked methods are not fair/relevant (details below).\n-Being faster to train, one would expect the method to be scaled up to larger images but no such experiments were reported. Thus, it is hard to evaluate how significant the training speedup is since the ODE based solutions are bottlenecked by training time [3], i.e. if the method does not address/alleviate the scalability issues, training faster on 64x64 images is less impactful.\n-Clarifications are needed for various parts of the submission (details below).\n\nRequired clarifications:\n-Section 3.2, the sentence: “This can be scaled up to larger spatial regions by performing the same calculation for each 2×2 patch.” I do not immediately see how. If “each” means a sliding window with stride 2, authors should explicitly write this. Otherwise some clarification is needed.\n-Section 3.2: Why does the logdet term being zero makes the training faster? More generally, the authors propose a new M matrix but other than looking nice, the value of the novelty is not clear (i.e., why should an additive log term be problematic for the training speed?). Maybe the authors can stress the training speedup a bit more and provide some intuition about the cause of this speedup.\n-Section 3.3, sentence with “...are generated conditioned on the entire coarser image...”: I do not see how. M is local, sliding it in stages (I assume) will still yield a (larger) local mapping. Could the authors expand their explanation? This is likely going to result in inconsistent textures, similar to WaveletFlow [1].\n-Section 4, “Comparison to WaveletFlow” part: It is not feasible to ask the authors to apply their innovations on top of a Glow [2] based architecture. However, as it stands, their comparison to WaveletFlow [1] does not seem fair. The ablations section does not seem relevant either (more on that below). The authors might want to add a disclaimer here that core architectures, training strategy, hardware/software are extremely different. If possible, I suggest the authors limit the qualitative comparisons to FFJORD [3] and leave the WaveletFlow comparisons to the empirical results section only. This would also inform the reader better since I believe this paper is a direct extension of FFJORD [3] with a generalized wavelet multi-resolution strategy. The other perspective of “applying CNF etc. to WaveletFlow [1]” would be misleading.\n-Table 3 and more generally Section 5.1: As I understand it, the comparisons are not between a WaveletFlow [1] modified to incorporate CNF vs. MRCNF but rather between MRCNF with Haar vs. MRCNF with the authors’ M matrix. This needs to be extremely clear.\nFurthermore, bullet point (1) is confusing to me. Do the authors replace the flow steps in WaveletFlow with a CNF? If so, what were the exact architectural changes?\nFinally, the last paragraph may be misleading (or at least unclear). The authors seem not to have modified (or even reproduced) WaveletFlow themselves, seeing as they report the exact same empirical results from the paper [1] and no mention of it in the provided code. Thus, the claim of  “MRCNF being more than WaveletFlow with CNF” is confusing - the experiments are just not there for proving/disproving this claim. I may be missing something, I ask the authors to please clarify.\n-Supplementary: Authors mention that gradient clipping for the adversarial loss being different. I could not find another mention of the adversarial loss in the paper.\n\nMinor points:\n-Introduction: I would urge the authors to abstain from blanket statements like “... notoriously difficult to train…”. With the newer normalization, loss function and architectural innovations GANs, at least in my experience, are easier to train.\n-Section 3.2, first paragraph: I understand that this is probably due to the page limit but if the authors can find the space, each bullet point should start as a new line.\n-Figure 2: I believe this figure can be cut without impacting the readability of the paper.\n\n\n[1] Jason Yu, Konstantinos Derpanis, and Marcus Brubaker. Wavelet flow: Fast training of high resolution normalizing flows. In Advances in Neural Information Processing Systems, 2020\n\n[2] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In Advances in neural information processing systems, pp. 10215–10224, 2018\n\n[3] Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. International Conference on Learning Representations, 2019\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this work, the authors propose a multi-resolution strategy for continuous normalizing flows. The proposed approach consists of a general wavelet based decomposition/downsampling where the transformation obeys some useful mathematical properties such as volume and range preservation. Empirical results indicate likelihood estimation performance better or on par with benchmarked methods. The proposed model also trains significantly faster with fewer parameters.", "main_review": "Strengths:\n-The authors provide a faster multi-resolution strategy for normalizing flows.\n-The empirical results are better than the benchmarked methods, e.g. WaveletFlow [1], Glow [2], FFJORD [3].\n-The paper is clearly written and easy to read.\n-Authors provide detailed descriptions and the source code thus the work should be easier to reproduce.\n\nWeaknesses:\n-Comparisons to some of the benchmarked methods are not fair/relevant (details below).\n-Being faster to train, one would expect the method to be scaled up to larger images but no such experiments were reported. Thus, it is hard to evaluate how significant the training speedup is since the ODE based solutions are bottlenecked by training time [3], i.e. if the method does not address/alleviate the scalability issues, training faster on 64x64 images is less impactful.\n-Clarifications are needed for various parts of the submission (details below).\n\nRequired clarifications:\n-Section 3.2, the sentence: “This can be scaled up to larger spatial regions by performing the same calculation for each 2×2 patch.” I do not immediately see how. If “each” means a sliding window with stride 2, authors should explicitly write this. Otherwise some clarification is needed.\n-Section 3.2: Why does the logdet term being zero makes the training faster? More generally, the authors propose a new M matrix but other than looking nice, the value of the novelty is not clear (i.e., why should an additive log term be problematic for the training speed?). Maybe the authors can stress the training speedup a bit more and provide some intuition about the cause of this speedup.\n-Section 3.3, sentence with “...are generated conditioned on the entire coarser image...”: I do not see how. M is local, sliding it in stages (I assume) will still yield a (larger) local mapping. Could the authors expand their explanation? This is likely going to result in inconsistent textures, similar to WaveletFlow [1].\n-Section 4, “Comparison to WaveletFlow” part: It is not feasible to ask the authors to apply their innovations on top of a Glow [2] based architecture. However, as it stands, their comparison to WaveletFlow [1] does not seem fair. The ablations section does not seem relevant either (more on that below). The authors might want to add a disclaimer here that core architectures, training strategy, hardware/software are extremely different. If possible, I suggest the authors limit the qualitative comparisons to FFJORD [3] and leave the WaveletFlow comparisons to the empirical results section only. This would also inform the reader better since I believe this paper is a direct extension of FFJORD [3] with a generalized wavelet multi-resolution strategy. The other perspective of “applying CNF etc. to WaveletFlow [1]” would be misleading.\n-Table 3 and more generally Section 5.1: As I understand it, the comparisons are not between a WaveletFlow [1] modified to incorporate CNF vs. MRCNF but rather between MRCNF with Haar vs. MRCNF with the authors’ M matrix. This needs to be extremely clear.\nFurthermore, bullet point (1) is confusing to me. Do the authors replace the flow steps in WaveletFlow with a CNF? If so, what were the exact architectural changes?\nFinally, the last paragraph may be misleading (or at least unclear). The authors seem not to have modified (or even reproduced) WaveletFlow themselves, seeing as they report the exact same empirical results from the paper [1] and no mention of it in the provided code. Thus, the claim of  “MRCNF being more than WaveletFlow with CNF” is confusing - the experiments are just not there for proving/disproving this claim. I may be missing something, I ask the authors to please clarify.\n-Supplementary: Authors mention that gradient clipping for the adversarial loss being different. I could not find another mention of the adversarial loss in the paper.\n\nMinor points:\n-Introduction: I would urge the authors to abstain from blanket statements like “... notoriously difficult to train…”. With the newer normalization, loss function and architectural innovations GANs, at least in my experience, are easier to train.\n-Section 3.2, first paragraph: I understand that this is probably due to the page limit but if the authors can find the space, each bullet point should start as a new line.\n-Figure 2: I believe this figure can be cut without impacting the readability of the paper.\n\n\n[1] Jason Yu, Konstantinos Derpanis, and Marcus Brubaker. Wavelet flow: Fast training of high resolution normalizing flows. In Advances in Neural Information Processing Systems, 2020\n\n[2] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In Advances in neural information processing systems, pp. 10215–10224, 2018\n\n[3] Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. International Conference on Learning Representations, 2019\n", "summary_of_the_review": "As the authors also point out, using multi-resolution strategies is not novel in itself. Additionally, the reported experiments indicate that the proposed method is not mature enough to be applied to larger images, limiting its usefulness. However, despite the incremental novelty of the approach and limited empirical results I believe the authors propose a method with potential to be widely adopted in the normalizing flows models. I would be willing to update my recommendation based on the authors’ response to my clarification requests listed above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["Yes, Other reasons (please specify below)"], "details_of_ethics_concerns": "I have communicated my concern to the AC.", "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635808032685}, {"id": "2pSMVzWT1Wq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2172/Reviewer_McBP"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new architecture for continuous normalizing flows that explicitly models images at multiple resolutions. The authors achieve this by learning an unconditional distribution of coarse images and then learning conditional distributions at progressively finer resolutions, each with a continuous normalizing flow. In contrast to the typical “squeeze” layers used in normalizing flows, where the noise is only passed at the first layer, the proposed model injects noise at several resolutions (which has already proved successful with other generative models). The authors introduce a new (invertible) transformation to go from a coarse image to a finer image which preserves the range of the image while having unit determinant (implying the transformation leaves the log likelihood unchanged). The whole model is then trained via maximum likelihood.\n\nThe authors evaluate their method on standard image datasets and perform several ablations to test the contributions of their model.\n\nThe main contributions of the paper in my eyes are then:\n- The introduction of a new architecture and layer for learning multi resolution normalizing flows\n- Demonstration that the proposed model can be trained with limited computational resources even on fairly large scale datasets (such as Imagenet128)\n- Ablation studies evaluating the improvements from each aspect of the proposed model\n", "review_text": "The motivation and execution of the paper are good and I appreciate the many experiments and ablations the authors have run. However, I have a few concerns about the discussion of some of the results in the paper as well as questions around some aspects of the model. The main strengths and weaknesses (in my eyes) are described below.\n\n**Strengths**:\n- The paper is well motivated and the ideas are sound - it makes sense to take advantage of the multi resolution structure of images.\n- The paper is well written and the figures are nice.\n- The authors make extensive ablation studies and compare their model to a large number of baselines (including reimplementations of baselines which is great). I appreciate the effort the authors have put in, in order to have thorough experiments on a variety of image datasets.\n- The results on Imagenet64 are impressive, getting low bpd with a low number of parameters and reduced training time.\n- The progressive training results on Imagenet128 are very nice. The authors are able to achieve fairly good results by only training a single refining step on top of an Imagenet64 model, which is impressive.\n- While the authors do not perform any experiments on this aspect of the model, the fact that the model can be trained in parallel is interesting.\n- Discussion of related work is thorough and the model is well situated in the literature.\n- The OoD experiments in the appendix are quite interesting.\n\n**Weaknesses**:\n- Some of the experimental results are slightly confusing. From both Table 1 and Table 3, it appears that adding more levels of resolutions (and hence more parameters) consistently _decreases_ performance, which seems very counterintuitive with the main point of the paper (that modeling multiple resolutions is important). This could suggest that it’s not the multi resolution aspect of the model that helps, but something else. I was disappointed that the authors do not seem to discuss this at all in the paper (in fact I had not realised this until I looked at the numbers in detail). This should be explicitly mentioned and discussed in the paper.\n- Experimental results on low resolution datasets are not very good. For example, on CIFAR10, for all tested resolutions (2 and 3) the model does worse than the RNODE and vanilla FFJORD baselines. Again this does not seem to be discussed or acknowledged anywhere in the paper. It is not a problem as such that this doesn’t work well for low res datasets given that it works well for high res datasets, but it would be good to discuss this and be more explicit about limitations in general.\n\n**Questions**:\n- I am confused about how the conditioning on the coarse image is done exactly. This feels like quite a significant part of the model but as far as I can tell is only described by a single sentence on page 6. How is the input image of the CNF concatenated with the coarser image? The dimension of the input image is equal to the dimension of the coarser image + the dimension of the noise variables if I understood correctly, so how can the input image and the coarse image be concatenated (they have different shapes)? Shouldn’t the noise and the coarse image be concatenated instead? It would be good to be more explicit about how this works.\n- In section 3.2, you choose $\\mathbf{y}_1$ such that it “contains information not present in $\\mathbf{x}_2$ such that $\\mathbf{x}_1$ is obtained when $\\mathbf{y}_1$ and $\\mathbf{x}_2$ are combined”. A very simple thing to do here that would satisfy this criterion is to set $\\mathbf{y}_1 = (x_1, x_2, x_3)$. Have you run this ablation?\n\n**Typos**:\n- Page 6: “we use to two regularization terms” -> “we use two regularization terms”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new architecture for continuous normalizing flows that explicitly models images at multiple resolutions. The authors achieve this by learning an unconditional distribution of coarse images and then learning conditional distributions at progressively finer resolutions, each with a continuous normalizing flow. In contrast to the typical “squeeze” layers used in normalizing flows, where the noise is only passed at the first layer, the proposed model injects noise at several resolutions (which has already proved successful with other generative models). The authors introduce a new (invertible) transformation to go from a coarse image to a finer image which preserves the range of the image while having unit determinant (implying the transformation leaves the log likelihood unchanged). The whole model is then trained via maximum likelihood.\n\nThe authors evaluate their method on standard image datasets and perform several ablations to test the contributions of their model.\n\nThe main contributions of the paper in my eyes are then:\n- The introduction of a new architecture and layer for learning multi resolution normalizing flows\n- Demonstration that the proposed model can be trained with limited computational resources even on fairly large scale datasets (such as Imagenet128)\n- Ablation studies evaluating the improvements from each aspect of the proposed model\n", "main_review": "The motivation and execution of the paper are good and I appreciate the many experiments and ablations the authors have run. However, I have a few concerns about the discussion of some of the results in the paper as well as questions around some aspects of the model. The main strengths and weaknesses (in my eyes) are described below.\n\n**Strengths**:\n- The paper is well motivated and the ideas are sound - it makes sense to take advantage of the multi resolution structure of images.\n- The paper is well written and the figures are nice.\n- The authors make extensive ablation studies and compare their model to a large number of baselines (including reimplementations of baselines which is great). I appreciate the effort the authors have put in, in order to have thorough experiments on a variety of image datasets.\n- The results on Imagenet64 are impressive, getting low bpd with a low number of parameters and reduced training time.\n- The progressive training results on Imagenet128 are very nice. The authors are able to achieve fairly good results by only training a single refining step on top of an Imagenet64 model, which is impressive.\n- While the authors do not perform any experiments on this aspect of the model, the fact that the model can be trained in parallel is interesting.\n- Discussion of related work is thorough and the model is well situated in the literature.\n- The OoD experiments in the appendix are quite interesting.\n\n**Weaknesses**:\n- Some of the experimental results are slightly confusing. From both Table 1 and Table 3, it appears that adding more levels of resolutions (and hence more parameters) consistently _decreases_ performance, which seems very counterintuitive with the main point of the paper (that modeling multiple resolutions is important). This could suggest that it’s not the multi resolution aspect of the model that helps, but something else. I was disappointed that the authors do not seem to discuss this at all in the paper (in fact I had not realised this until I looked at the numbers in detail). This should be explicitly mentioned and discussed in the paper.\n- Experimental results on low resolution datasets are not very good. For example, on CIFAR10, for all tested resolutions (2 and 3) the model does worse than the RNODE and vanilla FFJORD baselines. Again this does not seem to be discussed or acknowledged anywhere in the paper. It is not a problem as such that this doesn’t work well for low res datasets given that it works well for high res datasets, but it would be good to discuss this and be more explicit about limitations in general.\n\n**Questions**:\n- I am confused about how the conditioning on the coarse image is done exactly. This feels like quite a significant part of the model but as far as I can tell is only described by a single sentence on page 6. How is the input image of the CNF concatenated with the coarser image? The dimension of the input image is equal to the dimension of the coarser image + the dimension of the noise variables if I understood correctly, so how can the input image and the coarse image be concatenated (they have different shapes)? Shouldn’t the noise and the coarse image be concatenated instead? It would be good to be more explicit about how this works.\n- In section 3.2, you choose $\\mathbf{y}_1$ such that it “contains information not present in $\\mathbf{x}_2$ such that $\\mathbf{x}_1$ is obtained when $\\mathbf{y}_1$ and $\\mathbf{x}_2$ are combined”. A very simple thing to do here that would satisfy this criterion is to set $\\mathbf{y}_1 = (x_1, x_2, x_3)$. Have you run this ablation?\n\n**Typos**:\n- Page 6: “we use to two regularization terms” -> “we use two regularization terms”\n", "summary_of_the_review": "This paper introduces a new architecture for multi resolution continuous normalizing flows. The paper is well written and the model is sound and the authors achieve impressive results on Imagenet at high resolutions. However, as described in the weakness sections I believe there are still some aspects of the model and experiment discussion that need to be clarified. I therefore believe this paper is currently marginally above the acceptance threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635794125226}], "openreview_url": "https://openreview.net/forum?id=WN2Sup7qLdw", "arxiv_id": "2106.08462", "paper_pdf": "papers/WN2Sup7qLdw.pdf", "paper_pdf_sha256": "cc211cb96e7585a2094675b2f3533761e6afc94fbd0cc81660a069650045bb63", "paper_pdf_bytes": 1541495, "paper_pdf_source": "openreview", "code_url": "https://github.com/voletiv/mrcnf", "code_repository": "voletiv/mrcnf", "code_commit": "a564ca562909a0749a7c83f680a4a3b857b8961c", "code_archive": "repos/WN2Sup7qLdw.zip", "code_archive_sha256": "fd7b975d6cb916d3607a9960d38de8dc49868090cf39a794ebbbf2be0112677c", "code_archive_bytes": 89074, "code_file_count": 43, "code_extensions": {".py": 42, ".sh": 1}, "github_disk_usage_kb": 79, "github_languages": {"Python": 322462, "Shell": 1588}, "github_archived": false, "github_pushed_at": "2021-08-22T21:51:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-resolution-continuous-normalizing-flows"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iQxS0S9ir1a", "year": 2021, "status": "rejected", "title": "Distributional Generalization: A New Kind of Generalization", "authors": ["Preetum Nakkiran", "Yamini Bansal"], "authorids": ["~Preetum_Nakkiran1", "~Yamini_Bansal1"], "authors_source": "OpenReview API", "abstract": "We introduce a new notion of generalization--- Distributional Generalization--- which roughly states that outputs of a classifier at train and test time are close as distributions, as opposed to close in just their average error. For example, if we mislabel 30% of dogs as cats in the train set of CIFAR-10, then a ResNet trained to interpolation will in fact mislabel roughly 30% of dogs as cats on the test set as well, while leaving other classes unaffected. This behavior is not captured by classical generalization, which would only consider the average error and not the distribution of errors over the input domain. Our formal conjectures, which are much more general than this example, characterize the form of distributional generalization that can be expected in terms of problem parameters: model architecture, training procedure, number of samples, and data distribution. We give empirical evidence for these conjectures across a variety of domains in machine learning, including neural networks, kernel machines, and decision trees. Our results thus advance our understanding of interpolating classifiers.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "9YWrlrqZGX9", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1932/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new notion of \"distributional generalization\" as a tool to quantify the difference between the outputs from training and testing data sets using a certain machine learning algorithm. The authors formulate two conjectures for the so-called \"Interpolating classifiers\": the Feature Calibration Conjecture and Agreement Conjecture. The conjectures are supported numerically by some real data experiments and are proven for the 1-nearest-neighbor classifier under some technical conditions. \n\nStrengths: \n1. Two conjectures are interesting and have the potential to explain some unexplained phenomena for some existing machine learning algorithms.\n2. Numerical experiments are quite thorough, using different data sets and different machine learning algorithms.\n\nWeaknesses:\n1. The justifications of the conjectures are almost purely empirical (except in the 1-nearest-neighbor (1-N-N) classifier case). However, the 1-NN classifier is just a toy example that has limited use in machine learning literature. In fact, one even does not need to train the 1-NN classifier given a training data set. This is probably why it is easy to analyze 1-NN under the considered setting.\n2. While observations from data experiments are interesting and I enjoyed reading them, the paper fails to directly demonstrate how these observations can be used to improve performances or our understandings of existing algorithms. For example, how Conjecture 2 is useful? can we use it to evaluate the uncertainty (or confidence interval) of the test accuracy?\nHow can we find distinguishable features in Conjecture 1 in practice? using trees?\n\nConclusion:\nAlthough I find the proposed conjectures interesting and may have potential values, the current formulation, and justification for these conjectures are a bit too superficial and loose. It is difficult for me to envision a scenario where these conjectures can make a meaningful impact.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Recommendation to Reject", "review": "This paper introduces a new notion of \"distributional generalization\" as a tool to quantify the difference between the outputs from training and testing data sets using a certain machine learning algorithm. The authors formulate two conjectures for the so-called \"Interpolating classifiers\": the Feature Calibration Conjecture and Agreement Conjecture. The conjectures are supported numerically by some real data experiments and are proven for the 1-nearest-neighbor classifier under some technical conditions. \n\nStrengths: \n1. Two conjectures are interesting and have the potential to explain some unexplained phenomena for some existing machine learning algorithms.\n2. Numerical experiments are quite thorough, using different data sets and different machine learning algorithms.\n\nWeaknesses:\n1. The justifications of the conjectures are almost purely empirical (except in the 1-nearest-neighbor (1-N-N) classifier case). However, the 1-NN classifier is just a toy example that has limited use in machine learning literature. In fact, one even does not need to train the 1-NN classifier given a training data set. This is probably why it is easy to analyze 1-NN under the considered setting.\n2. While observations from data experiments are interesting and I enjoyed reading them, the paper fails to directly demonstrate how these observations can be used to improve performances or our understandings of existing algorithms. For example, how Conjecture 2 is useful? can we use it to evaluate the uncertainty (or confidence interval) of the test accuracy?\nHow can we find distinguishable features in Conjecture 1 in practice? using trees?\n\nConclusion:\nAlthough I find the proposed conjectures interesting and may have potential values, the current formulation, and justification for these conjectures are a bit too superficial and loose. It is difficult for me to envision a scenario where these conjectures can make a meaningful impact.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603888730425}, {"id": "oVm6p6iuAA3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1932/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new notion of generalization called distributional generalization which states that the outputs of the classifier for train and test are close as distributions not just their corresponding accuracy numbers. They propose conjectures about their the distributional closeness that they expect and how it depends on the model, number of data points and the algorithm. This paper gives experiments to support their conjectures for different model classes including neural networks, kernel methods and decision trees.\n\nI find the notion of distributional generalization interesting. The first experiment of this paper is that the label noise introduced in the training set in one subgroup of a particular class is also localized in the same subgroup in the test set although the classifier does not have any explicit information about the subgroups which is interesting. The paper formally states that the distribution of train and test outcomes are similar in distribution with respect to tests which themselves can learned by that model class with the current number of samples.\n\nI recommend accepting this paper because the notion of distributional generalization that this paper introduces is both interesting and surprising. Moreover, this paper has shown that it holds widely across different interpolating classifiers like decision trees, neural networks and kernel methods.\n\nThe paper includes an extensive set of experiments that are easy to follow.\n\nOne concern is as itself mentioned in the paper that this work does not talk about what conditions on the distribution, algorithm and the model class are necessary for this conjecture to hold either empirically or theoretically and the conjecture is not precisely defined. Any insights on this part would be good to include. \n\nAnother concern is that although this is an interesting observation, the paper does not talk about why and how studying this form of generalization would be useful for understanding interpolating classifiers and generalization in general. \n\nQuestions:\n1) The authors suspect locality to be the underlying reason behind distributional generalization and find this observation to be true for kernel methods also. Do the authors know if there is some work on the locality aspects of kernel methods arguing that some particular kernels are more local than others and how this relates to the observations in this paper?\n2) Regarding the agreement property in section 4, I find it a little surprising that the conjecture says that the correlation and the accuracy are the same. Does this relate to the number of classes present?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting observations", "review": "This paper proposes a new notion of generalization called distributional generalization which states that the outputs of the classifier for train and test are close as distributions not just their corresponding accuracy numbers. They propose conjectures about their the distributional closeness that they expect and how it depends on the model, number of data points and the algorithm. This paper gives experiments to support their conjectures for different model classes including neural networks, kernel methods and decision trees.\n\nI find the notion of distributional generalization interesting. The first experiment of this paper is that the label noise introduced in the training set in one subgroup of a particular class is also localized in the same subgroup in the test set although the classifier does not have any explicit information about the subgroups which is interesting. The paper formally states that the distribution of train and test outcomes are similar in distribution with respect to tests which themselves can learned by that model class with the current number of samples.\n\nI recommend accepting this paper because the notion of distributional generalization that this paper introduces is both interesting and surprising. Moreover, this paper has shown that it holds widely across different interpolating classifiers like decision trees, neural networks and kernel methods.\n\nThe paper includes an extensive set of experiments that are easy to follow.\n\nOne concern is as itself mentioned in the paper that this work does not talk about what conditions on the distribution, algorithm and the model class are necessary for this conjecture to hold either empirically or theoretically and the conjecture is not precisely defined. Any insights on this part would be good to include. \n\nAnother concern is that although this is an interesting observation, the paper does not talk about why and how studying this form of generalization would be useful for understanding interpolating classifiers and generalization in general. \n\nQuestions:\n1) The authors suspect locality to be the underlying reason behind distributional generalization and find this observation to be true for kernel methods also. Do the authors know if there is some work on the locality aspects of kernel methods arguing that some particular kernels are more local than others and how this relates to the observations in this paper?\n2) Regarding the agreement property in section 4, I find it a little surprising that the conjecture says that the correlation and the accuracy are the same. Does this relate to the number of classes present?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603874388077}, {"id": "vN3RJYOK_G", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1932/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "> Summarize what the paper claims to contribute\n\nThis paper generalizes the classical notion of generalization to the notion that the output of the classifier should be close when applied to the training data and the testing data. Two conjectures were made to predict the behavior of the distributional generalization for a few models and data. The first conjecture, Feature Calibration Conjecture, asserts that the distribution of the outputs are similar up to “distinguishable features”. The second conjecture, Agreement Conjecture, asserts that test accuracy matches the classification stability for interpolated classifiers. A number of experiments surrounding these conjectures were conducted to illustrate the Distributional Generalization\n\n> List strong and weak points of the paper. Be as comprehensive as possible.\n+ Great idea to look into distribution of the output beyond the error. I think this is long overdue and this paper has a lot of potential.\n+ Comprehensive numerical studies which show some interesting findings.\n- Many vague and ambiguous notations which make the paper hard to read.\n- Barely any theoretical justification. \n- \"Too many\" ideas. I would suggest the authors to divide the paper into 3 papers and dig deeper in each topic. \n\n> Provide supporting arguments for your recommendation.\n\n1. Notation. While the display on the bottom of page 2 is supposed to be an informal conjecture, it may be too informal. Does x∈TestSet indicate a distribution or a set? What does ≡ mean here? It was mentioned that the joint distribution of X and Y is D, and D^n denote n iid samples from D. But is D^n a sample of training data points, or a distribution? Later on it was mentioned that \"S ∼ D^n\" which does not make sense if D^n is a sample. \n2. There are only two theoretical results in the paper with rigorous proofs. But they are established for nearest neighbor classifier only and some assumptions are imposed which warrants the results straightforwardly. However, most of the conjectures are made for many classifiers beyond nearest neighbor. \n3. Indistinguishability and interpolated classifier: in the Indistinguishability meta-conjecture on page 2, the second equality is clearly due to interpolated classifier. I wonder if the closeness indicated in the first approximation (between trainset and testset) is a merit of the interpolated classifier, or the general good generalization performance of a classifier. It is not clear to me why interpolated classifiers will induce this closeness.\n4. The role of L. From what I can see, this intermediate feature L is introduced because \\mathcal{X} can be a large high-dimensional space which can make visualization or quantification of the closeness difficult. This function L is introduced so that it is easier to see how (L(x), f(x)) is distributed. For constant partition (L(x) = 0), the distributional generalization reduces to the marginal distribution of \"f(x)\". In my opinion, if a classifier is true generalizable, the distribution of (x, f(x)) itself, without L, should be the same between the test and training domains. From the statement of Theorem 1, it seems to me that  Conjecture 1 holds for nearest neighbor BECAUSE some distinguishable features exist, which may already be a strong assumption. So strong that the very existence of the distinguishable features warrants the generalizability. Unfortunately because there are only conjectures and no rigorous proofs, this part is still not clear to me.\n\n> Ask questions you would like answered by the authors to help you clarify your understanding of the paper and provide the additional evidence you need to be confident in your assessment. \n\nWhile the numerical studies have convinced me that distributional generalization is really a thing here, I am not sure that I am that impressed. After all, the precision was only two decimal points. Can the total-variation distance be calculated? Is there a counterexample in which one classifier does have classical generalizability but not distributional generalizability? Focus on a narrower set of topics but dig deeper may go a long way here.  \n\n> Minor comments\n\nPage 3: “do we this same procedure” is a typo\nPage 4: how is “learnable” defined?\nPage 5: f←Train F (D^n) is not defined, esp. given that it is unclear if D^n is a distribution or a set. Should D^n be replaced by S?\nPage 5: Indistinguishably is a typo\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of Distributional Generalization", "review": "> Summarize what the paper claims to contribute\n\nThis paper generalizes the classical notion of generalization to the notion that the output of the classifier should be close when applied to the training data and the testing data. Two conjectures were made to predict the behavior of the distributional generalization for a few models and data. The first conjecture, Feature Calibration Conjecture, asserts that the distribution of the outputs are similar up to “distinguishable features”. The second conjecture, Agreement Conjecture, asserts that test accuracy matches the classification stability for interpolated classifiers. A number of experiments surrounding these conjectures were conducted to illustrate the Distributional Generalization\n\n> List strong and weak points of the paper. Be as comprehensive as possible.\n+ Great idea to look into distribution of the output beyond the error. I think this is long overdue and this paper has a lot of potential.\n+ Comprehensive numerical studies which show some interesting findings.\n- Many vague and ambiguous notations which make the paper hard to read.\n- Barely any theoretical justification. \n- \"Too many\" ideas. I would suggest the authors to divide the paper into 3 papers and dig deeper in each topic. \n\n> Provide supporting arguments for your recommendation.\n\n1. Notation. While the display on the bottom of page 2 is supposed to be an informal conjecture, it may be too informal. Does x∈TestSet indicate a distribution or a set? What does ≡ mean here? It was mentioned that the joint distribution of X and Y is D, and D^n denote n iid samples from D. But is D^n a sample of training data points, or a distribution? Later on it was mentioned that \"S ∼ D^n\" which does not make sense if D^n is a sample. \n2. There are only two theoretical results in the paper with rigorous proofs. But they are established for nearest neighbor classifier only and some assumptions are imposed which warrants the results straightforwardly. However, most of the conjectures are made for many classifiers beyond nearest neighbor. \n3. Indistinguishability and interpolated classifier: in the Indistinguishability meta-conjecture on page 2, the second equality is clearly due to interpolated classifier. I wonder if the closeness indicated in the first approximation (between trainset and testset) is a merit of the interpolated classifier, or the general good generalization performance of a classifier. It is not clear to me why interpolated classifiers will induce this closeness.\n4. The role of L. From what I can see, this intermediate feature L is introduced because \\mathcal{X} can be a large high-dimensional space which can make visualization or quantification of the closeness difficult. This function L is introduced so that it is easier to see how (L(x), f(x)) is distributed. For constant partition (L(x) = 0), the distributional generalization reduces to the marginal distribution of \"f(x)\". In my opinion, if a classifier is true generalizable, the distribution of (x, f(x)) itself, without L, should be the same between the test and training domains. From the statement of Theorem 1, it seems to me that  Conjecture 1 holds for nearest neighbor BECAUSE some distinguishable features exist, which may already be a strong assumption. So strong that the very existence of the distinguishable features warrants the generalizability. Unfortunately because there are only conjectures and no rigorous proofs, this part is still not clear to me.\n\n> Ask questions you would like answered by the authors to help you clarify your understanding of the paper and provide the additional evidence you need to be confident in your assessment. \n\nWhile the numerical studies have convinced me that distributional generalization is really a thing here, I am not sure that I am that impressed. After all, the precision was only two decimal points. Can the total-variation distance be calculated? Is there a counterexample in which one classifier does have classical generalizability but not distributional generalizability? Focus on a narrower set of topics but dig deeper may go a long way here.  \n\n> Minor comments\n\nPage 3: “do we this same procedure” is a typo\nPage 4: how is “learnable” defined?\nPage 5: f←Train F (D^n) is not defined, esp. given that it is unclear if D^n is a distribution or a set. Should D^n be replaced by S?\nPage 5: Indistinguishably is a typo\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603855581991}, {"id": "_623CXN_77", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1932/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an extended notion of generalization. The new proposed notion asks that for a family of tests $T: X \\times Y → [0,1]$, $T(x, f(x))$  will be similar for train/test examples. The paper proposes three interesting conjectures that are related to distributional generalization. The paper proves Conjecture 1 for nearest-neighbor classifiers. The paper also gives empirical evidence supporting their conjectures.\n\nI think this is a nice contribution. It raises some new questions on generalization, and points to some interesting properties that interpolating classifiers satisfy (at least empirically) that deserve to be studied in more detail. In particular, Conjecture 1 seems to suggest that there is a form of hierarchical learning that interpolating classifiers are doing without being told to explicitly do so. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper proposes an extended notion of generalization. The new proposed notion asks that for a family of tests $T: X \\times Y → [0,1]$, $T(x, f(x))$  will be similar for train/test examples. The paper proposes three interesting conjectures that are related to distributional generalization. The paper proves Conjecture 1 for nearest-neighbor classifiers. The paper also gives empirical evidence supporting their conjectures.\n\nI think this is a nice contribution. It raises some new questions on generalization, and points to some interesting properties that interpolating classifiers satisfy (at least empirically) that deserve to be studied in more detail. In particular, Conjecture 1 seems to suggest that there is a form of hierarchical learning that interpolating classifiers are doing without being told to explicitly do so. \n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603834453035}], "openreview_url": "https://openreview.net/forum?id=iQxS0S9ir1a", "arxiv_id": "2009.08092", "paper_pdf": "papers/iQxS0S9ir1a.pdf", "paper_pdf_sha256": "ad589072d553ac2cf4b41f43ed6e426f0af7787bd5900a51b60a53a6f7c9d3bf", "paper_pdf_bytes": 5339815, "paper_pdf_source": "openreview", "code_url": "https://github.com/kuangliu/pytorch-cifar", "code_repository": "kuangliu/pytorch-cifar", "code_commit": "49b7aa97b0c12fe0d4054e670403a16b6b834ddd", "code_archive": "repos/iQxS0S9ir1a.zip", "code_archive_sha256": "9a807c8c6c68a1842346f41d7f0a18effa36a28179d86fc56dee8cf05f1fb41c", "code_archive_bytes": 26050, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 85, "github_languages": {"Python": 73999}, "github_archived": false, "github_pushed_at": "2023-02-24T10:06:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distributional-generalization-a-new-kind-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJgDb1SFwB", "year": 2020, "status": "rejected", "title": "MGP-AttTCN: An Interpretable Machine Learning Model for the Prediction of Sepsis", "authors": ["Margherita Rosnati", "Vincent Fortuin"], "authorids": ["mrosnati@ethz.ch", "fortuin@inf.ethz.ch"], "authors_source": "OpenReview API", "abstract": "With a mortality rate of 5.4 million lives worldwide every year and a healthcare cost of more than 16 billion dollars in the USA alone, sepsis is one of the leading causes of hospital mortality and an increasing concern in the ageing western world. Recently, medical and technological advances have helped re-define the illness criteria of this disease, which is otherwise poorly understood by the medical society. Together with the rise of widely accessible Electronic Health Records, the advances in data mining and complex nonlinear algorithms are a promising avenue for the early detection of sepsis. This work contributes to the research effort in the field of automated sepsis detection with an open-access labelling of the medical MIMIC-III data set. Moreover, we propose MGP-AttTCN: a joint multitask Gaussian Process and attention-based deep learning model to early predict the occurrence of sepsis in an interpretable manner. We show that our model outperforms the current state-of-the-art and present evidence that different labelling heuristics lead to discrepancies in task difficulty.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H1gqUrzD5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1546/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors consider a combination of an Gaussian model and neural network learned together to be able to find specific Gaussian features that would predict sepsis in a more intuitive, interpretable way for a medical experts. Also, they have provided a new labelling for the MIMIC-III dataset, which is of great value.\n\nUsually constraining the feature space reduces the accuracy, as one tends to miss important features. However, here, one is using a Gaussian model to generate a feature space that is easier to train with a neural network (by filling the sparse data by Gaussian process interpolation) but also augment the dataset.\n\nThe author report that his improves prediction. Unfortunately, I did not find tables of solutions where one could the actual impact. The authors should include a numerical table of their result comparison results.  Now there is only a narrative in section 5.3 and an image showing that at different covariance times, different feature groups starts to interpret the results. Obviously, a question arises if the model would perform even better with a combination of covariance times, or is there some covariance time range that is missing that would improve the result even more.\n\nThe Gaussian model creates smooth interpolation of data spaces and also forces the training to look at corresponding smoothened features - that are very good for human eyes. However, there are situations (like the detecting heart beat from a video from a head moving with a recoil from the blood rushing to the brain) in where the signal is too weak for human to see, but is definitely there for a computer.  I would state that this as interpretable,  as an explicitly visible signal would be. Even  shorter time constant signal might be valuable as well, but it would not be visible here...  It seems that in sepsis, it was a good idea, as it improves the result compared to the situation of not using the Gaussian model. \n\nOne has to be careful to state that this would address the interpretability of the results. Gaussian process by itself is not giving understanding nor interpretability as it is too general. But it can make the provided solution \"teachable\" to a human expert by showing what visible features one can track.\n\nCompare this to a situation where one is using physical model to regularize detection. It has the analogous two model structure like the one in the manuscript. In https://xbpeng.github.io/projects/SFV/index.html the authors of the paper report that that one can achieve a better pose estimation by constraining the pose to only those that are achievable by a physical model based policy trained by reinforcement learning. This one is able to \"interpret\" the pose.\n\nAs a summary, the authors have done solid and valuable work in improving the accuracy detection. They should have a more formal way to present the results and baseline comparisions as tables. On the explainability and interpretability, there remains a lot of interpretations and one has lots of explaining to do, even after this manuscript.\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The authors consider a combination of an Gaussian model and neural network learned together to be able to find specific Gaussian features that would predict sepsis in a more intuitive, interpretable way for a medical experts. Also, they have provided a new labelling for the MIMIC-III dataset, which is of great value.\n\nUsually constraining the feature space reduces the accuracy, as one tends to miss important features. However, here, one is using a Gaussian model to generate a feature space that is easier to train with a neural network (by filling the sparse data by Gaussian process interpolation) but also augment the dataset.\n\nThe author report that his improves prediction. Unfortunately, I did not find tables of solutions where one could the actual impact. The authors should include a numerical table of their result comparison results.  Now there is only a narrative in section 5.3 and an image showing that at different covariance times, different feature groups starts to interpret the results. Obviously, a question arises if the model would perform even better with a combination of covariance times, or is there some covariance time range that is missing that would improve the result even more.\n\nThe Gaussian model creates smooth interpolation of data spaces and also forces the training to look at corresponding smoothened features - that are very good for human eyes. However, there are situations (like the detecting heart beat from a video from a head moving with a recoil from the blood rushing to the brain) in where the signal is too weak for human to see, but is definitely there for a computer.  I would state that this as interpretable,  as an explicitly visible signal would be. Even  shorter time constant signal might be valuable as well, but it would not be visible here...  It seems that in sepsis, it was a good idea, as it improves the result compared to the situation of not using the Gaussian model. \n\nOne has to be careful to state that this would address the interpretability of the results. Gaussian process by itself is not giving understanding nor interpretability as it is too general. But it can make the provided solution \"teachable\" to a human expert by showing what visible features one can track.\n\nCompare this to a situation where one is using physical model to regularize detection. It has the analogous two model structure like the one in the manuscript. In https://xbpeng.github.io/projects/SFV/index.html the authors of the paper report that that one can achieve a better pose estimation by constraining the pose to only those that are achievable by a physical model based policy trained by reinforcement learning. This one is able to \"interpret\" the pose.\n\nAs a summary, the authors have done solid and valuable work in improving the accuracy detection. They should have a more formal way to present the results and baseline comparisions as tables. On the explainability and interpretability, there remains a lot of interpretations and one has lots of explaining to do, even after this manuscript.\n\n\n\n\n"}, "tcdate": 1572443473750}, {"id": "r1ge0ZnWcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1546/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present a Sepsis-3 compliant labeling of the MIMIC-III dataset and a sepsis-prediction model largely based on MGP-TCN that uses attention mechanisms to enable explainability.\n\nIt is not entirely clear what the authors mean by MC samples from Y_MGP, are these simply samples from the posterior in (6)?\n\nIf z_j and z'_j are M-dimensional, how does one apply (8) and (9) for W_{\\alpha,0}, W_{\\alpha,1} being (M+Q)-dimensional or W_{\\beta,0}, W_{\\beta,1} being matrices?\n\nThe labelling of the data, largely following Johnson & Pollard (2018) and Moor et al (2019), is only different to Moor et al (2019) in the assumption that in the SOFA calculation missing values have zero contribution. Unless the authors provide evidence that this is reasonable, it is not necessarily clear whether labels resulting from the proposed scheme will be biased and affected by differences in clinical practice at different sites or data collection practices. That being said, it is not clear whether the proposed labeling is a contribution from the work.\n\nThe fact that the proposed labels are harder to fit does not imply that the proposed labels are better or more reasonable. This provided that is difficult to know (without ground truth) whether the difficulty originates from a broader use case (not as easy as Moor et al (2019)) or labels being noisy, imperfect proxies for sepsis diagnosis. I understand the author's motivation for doing it, however, their approach is not sufficiently justified. I also agree that predicting sepsis in a realistic setting is harder than suggested in prior work, however, the proposed labeling does not necessarily yields evidence of that being the case.\n\nThe interpretation of the covariance matrices of the MGP is interesting, though not surprising considering that covariates in green are measured regularly while blue covariates are ordered sparingly.\n\nFigure 5 is interesting, though raises questions about of interpretability of the model. How should unobserved covariates be interpreted (INR in Figure 5)?\n\nIn summary, the contributions of the present work are not sufficiently justified (labeling), the novelty of the proposed model is minor, relative to MGP-TCN, and the added value of the attention mechanism as a means to interpret predictions in terms of the journey of a patient is not clear.\n\nMinor:\n- Figure 1 needs a better caption. Being in page 2 makes it very difficult to understand.\n- TCN is used before being defined.\n- In (1) it should be t_{p,i,k} not t_{p,k,i}", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The authors present a Sepsis-3 compliant labeling of the MIMIC-III dataset and a sepsis-prediction model largely based on MGP-TCN that uses attention mechanisms to enable explainability.\n\nIt is not entirely clear what the authors mean by MC samples from Y_MGP, are these simply samples from the posterior in (6)?\n\nIf z_j and z'_j are M-dimensional, how does one apply (8) and (9) for W_{\\alpha,0}, W_{\\alpha,1} being (M+Q)-dimensional or W_{\\beta,0}, W_{\\beta,1} being matrices?\n\nThe labelling of the data, largely following Johnson & Pollard (2018) and Moor et al (2019), is only different to Moor et al (2019) in the assumption that in the SOFA calculation missing values have zero contribution. Unless the authors provide evidence that this is reasonable, it is not necessarily clear whether labels resulting from the proposed scheme will be biased and affected by differences in clinical practice at different sites or data collection practices. That being said, it is not clear whether the proposed labeling is a contribution from the work.\n\nThe fact that the proposed labels are harder to fit does not imply that the proposed labels are better or more reasonable. This provided that is difficult to know (without ground truth) whether the difficulty originates from a broader use case (not as easy as Moor et al (2019)) or labels being noisy, imperfect proxies for sepsis diagnosis. I understand the author's motivation for doing it, however, their approach is not sufficiently justified. I also agree that predicting sepsis in a realistic setting is harder than suggested in prior work, however, the proposed labeling does not necessarily yields evidence of that being the case.\n\nThe interpretation of the covariance matrices of the MGP is interesting, though not surprising considering that covariates in green are measured regularly while blue covariates are ordered sparingly.\n\nFigure 5 is interesting, though raises questions about of interpretability of the model. How should unobserved covariates be interpreted (INR in Figure 5)?\n\nIn summary, the contributions of the present work are not sufficiently justified (labeling), the novelty of the proposed model is minor, relative to MGP-TCN, and the added value of the attention mechanism as a means to interpret predictions in terms of the journey of a patient is not clear.\n\nMinor:\n- Figure 1 needs a better caption. Being in page 2 makes it very difficult to understand.\n- TCN is used before being defined.\n- In (1) it should be t_{p,i,k} not t_{p,k,i}"}, "tcdate": 1572090311642}, {"id": "BylDbvpdtB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1546/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I've read the rebuttal and I'd like to keep my score as is. My main concern is the questionable role of attention in making the model more interpretable (which is the main contribution of the paper).\n\n###########################\n\nThe paper proposes a new model for automated sepsis detection using multitask GP and attention-based GP. The sepsis detection problem is of paramount importance in the clinical domain and the authors rightly emphasized that point. Also the paper tries to combine interpretability with prediction accuracy by using attention mechanism. \n\nThe paper is generally well written and well motivated; however, in terms of technical novelty and empirical evidence the paper can be further improved. \n\nThe MGP-AttnTCN model is mostly a minor modification of the model proposed by Moor et al. 2019 and has the additional attention element to be more interpretable. Unfortunately, it’s not easy for an ICLR reader without any medical background to evaluate the validity of the interpretability results provided in the paper. Furthermore, the recent works in NLP have agued against the value of attention for interpretability (see for instance “Attention is not Explanation” by Jain & Wallace 2019). That said, I believe the paper is probably a better fit for a machine learning in healthcare venue (such as MLHC). \n\nIn terms of empirical evidence of the prediction accuracy the paper only compares with Moor et al 2019 (which does not show a significant improvement in the realistic setting) and a much older InSight paper (2016). This would have been typically enough for a paper with major technical novelty; however, for this paper, I believe adding more recent baselines and discussing the advantages of the method over these baselines would be necessary. \n\nMinor:\nCaption in Figure 1 can be more informative and useful for the reader if you add more details on different parts of the model. \n“Graphically, once can observe” should be “one can observe” . ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "I've read the rebuttal and I'd like to keep my score as is. My main concern is the questionable role of attention in making the model more interpretable (which is the main contribution of the paper).\n\n###########################\n\nThe paper proposes a new model for automated sepsis detection using multitask GP and attention-based GP. The sepsis detection problem is of paramount importance in the clinical domain and the authors rightly emphasized that point. Also the paper tries to combine interpretability with prediction accuracy by using attention mechanism. \n\nThe paper is generally well written and well motivated; however, in terms of technical novelty and empirical evidence the paper can be further improved. \n\nThe MGP-AttnTCN model is mostly a minor modification of the model proposed by Moor et al. 2019 and has the additional attention element to be more interpretable. Unfortunately, it’s not easy for an ICLR reader without any medical background to evaluate the validity of the interpretability results provided in the paper. Furthermore, the recent works in NLP have agued against the value of attention for interpretability (see for instance “Attention is not Explanation” by Jain & Wallace 2019). That said, I believe the paper is probably a better fit for a machine learning in healthcare venue (such as MLHC). \n\nIn terms of empirical evidence of the prediction accuracy the paper only compares with Moor et al 2019 (which does not show a significant improvement in the realistic setting) and a much older InSight paper (2016). This would have been typically enough for a paper with major technical novelty; however, for this paper, I believe adding more recent baselines and discussing the advantages of the method over these baselines would be necessary. \n\nMinor:\nCaption in Figure 1 can be more informative and useful for the reader if you add more details on different parts of the model. \n“Graphically, once can observe” should be “one can observe” . ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571505918874}], "openreview_url": "https://openreview.net/forum?id=rJgDb1SFwB", "arxiv_id": "1909.12637", "paper_pdf": "papers/rJgDb1SFwB.pdf", "paper_pdf_sha256": "967ee4ec0bbe2f2f0019fe50f4bc7cd3d091ab0e3cc5d8c3fb58a13ceab7f466", "paper_pdf_bytes": 425124, "paper_pdf_source": "openreview", "code_url": "https://github.com/mmr12/MGP-AttTCN", "code_repository": "mmr12/MGP-AttTCN", "code_commit": "57c8f623d7deee4085a0ac58e795c0264d2f0e49", "code_archive": "repos/rJgDb1SFwB.zip", "code_archive_sha256": "275b19da07b5c4a0a463c1a6729ec9256acf6c26a676df94116fc25e119fd56f", "code_archive_bytes": 112897, "code_file_count": 59, "code_extensions": {".py": 31, ".sql": 28}, "github_disk_usage_kb": 130, "github_languages": {"Python": 174580}, "github_archived": false, "github_pushed_at": "2023-03-25T01:05:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mgp-atttcn-an-interpretable-machine-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJlMBjAcYX", "year": 2019, "status": "rejected", "title": "Optimizing for Generalization in Machine Learning with Cross-Validation Gradients", "authors": ["Barratt", "Shane", "Sharma", "Rishi"], "authorids": ["sbarratt@stanford.edu", "rsh@stanford.edu"], "authors_source": "OpenReview API", "abstract": "Cross-validation is the workhorse of modern applied statistics and machine learning, as it provides a principled framework for selecting the model that maximizes generalization performance. In this paper, we show that the cross-validation risk is differentiable with respect to the hyperparameters and training data for many common machine learning algorithms, including logistic regression, elastic-net regression, and support vector machines. Leveraging this property of differentiability, we propose a cross-validation gradient method (CVGM) for hyperparameter optimization. Our method enables efficient optimization in high-dimensional hyperparameter spaces of the cross-validation risk, the best surrogate of the true generalization ability of our learning algorithm.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HkgzyGoV2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper66/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes the so-called cross-validation gradient method (CVGM).\nThis is idea is to express the CV score as a differentiable function\nof the hyperparameters and then to update hyperparameters with gradient\ndescent. Derivations are provided with Logistic regression and Elastic-Net\nthanks to the sign splitting trick.\n\nOnce the problem is expressed as a QP, the work is mostly done\nby the qpth library that offers a differentiable layer for the QP solver.\n\nMajor points:\n\n- This idea has been around for quite some time but yes it is now certainly more\ntimely with the new DL tools such as pytorch. However the novelty is limited\nwhich means that numerical experiments should be quite extensive to\ndemonstrate a clear impact on the field. Unfortunately the experiments\nare very limited: data mostly simulated and very small. What is missing\nis a real evaluation of larger datasets and a demonstration that one can\noutperform the state of the art using CVGM. For example for the Elastic-Net\nit is unclear if CVGM is faster than glmnet that computes full grid search\nbut uses warm start so is very efficient.\n\n- Given a new dataset, how do you set step sizes? The purpose is to\nfind faster good hyperparameters than using Bayes Opt or random search\nbut if I need to fiddle with the choice of step size is it really worth it?\n\nMinor points:\n\nPlease proof read manuscript as there are a few typos.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper but with limited novelty and lacking convincing experiments", "review": "This paper proposes the so-called cross-validation gradient method (CVGM).\nThis is idea is to express the CV score as a differentiable function\nof the hyperparameters and then to update hyperparameters with gradient\ndescent. Derivations are provided with Logistic regression and Elastic-Net\nthanks to the sign splitting trick.\n\nOnce the problem is expressed as a QP, the work is mostly done\nby the qpth library that offers a differentiable layer for the QP solver.\n\nMajor points:\n\n- This idea has been around for quite some time but yes it is now certainly more\ntimely with the new DL tools such as pytorch. However the novelty is limited\nwhich means that numerical experiments should be quite extensive to\ndemonstrate a clear impact on the field. Unfortunately the experiments\nare very limited: data mostly simulated and very small. What is missing\nis a real evaluation of larger datasets and a demonstration that one can\noutperform the state of the art using CVGM. For example for the Elastic-Net\nit is unclear if CVGM is faster than glmnet that computes full grid search\nbut uses warm start so is very efficient.\n\n- Given a new dataset, how do you set step sizes? The purpose is to\nfind faster good hyperparameters than using Bayes Opt or random search\nbut if I need to fiddle with the choice of step size is it really worth it?\n\nMinor points:\n\nPlease proof read manuscript as there are a few typos.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1540825561668}, {"id": "HJeaamKV3Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper66/AnonReviewer3"], "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers the problem of automatic tuning of hyperparameters in machine learning models. To address this problem the authors propose to use the so called cross-validation gradients, which optimize a validation objective with respect to the hyperparameters of a model. This approach and the investigated setting falls into a class of optimization problems known as bilevel optimization. The main characteristic of this class of optimization problems is the nested structure, with an outer and inner optimization objectives/problems. The outer problem corresponds to the validation objective and it is defined via an optimal solution to the inner problem which corresponds to a training objective. The paper, however, fails to make a reference to a rather rich literature on bilevel optimization (e.g., see [3-5] and references therein).\n\nThe approach, presented as Algorithm 1, does not seem different from [1] and [2] where hyperparameter optimization was considered for (kernel) support vector machines and (kernel) ridge regression. The data is initially split into k-folds (not necessarily of identical size) and each fold is used exactly once to define a validation objective whereas the complementary folds act as training data. The validation gradient is obtained by averaging the gradients of the k validation folds. Essentially, the same algorithm with k-fold cross-validation was considered in [1]. Thus, for me there does not seem to be any novelty in this approach and the paper itself.\n\nThe experiments involve synthetic regression and classification datasets but there are no novel insights that advance what is already known about the hyperparameter optimization (e.g., see [3]). For example, there is no intuition on the geometry of the optimization problem and the optimality of the outer optimization problem which is non-convex (e.g., see [7]), or dependence of the outer solution on the accuracy of the inner solution.\n\nReferences:\n\n[1] S. Keerthi, V. Sindhwani, and O. Chapelle. An Efficient Method for Gradient-Based Adaptation of Hyperparameters. NIPS 2007.\n[2] O. Chapelle, V. Vapnik, O. Bousquet, and S. Mukherjee. Choosing Multiple Parameters for Support Vector Machines. Machine Learning, 2002.\n\n[-3] L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil. Bilevel Programming for Hyperparameter Optimization and Meta-Learning. ICML 2018.\n[-4] G. Kunapuli, K.P. Bennet, J. Hu, and J-S. Pang. Bilevel Model Selection for SVMs. American Mathematical Society, 2008.\n[-5] E.S.H. Neto and A.R. de Pierro. On Perturbed Steepest Descent Methods with Inexact Line Search for Bilevel Convex Optimization. Journal of Mathematical Programming and Operations Research, 2011.\n[-6] B. Colson, P. Marcotte, and G. Savard. A Trust-Region Method for Nonlinear Bilevel Programming: Algorithms and Computational Experience. Computational Optimization and Applications, 2005.\n[-7] M. Janzamin, H. Sedghi, and A. Anandkumar. Beating the Perils of Non-Convexity: Guaranteed Training of Neural Networks using Tensor Method. arXiv preprint arXiv:1506.08473v3, 2016.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "I think there is not enough novelty in this work to be considered for this conference.", "review": "The paper considers the problem of automatic tuning of hyperparameters in machine learning models. To address this problem the authors propose to use the so called cross-validation gradients, which optimize a validation objective with respect to the hyperparameters of a model. This approach and the investigated setting falls into a class of optimization problems known as bilevel optimization. The main characteristic of this class of optimization problems is the nested structure, with an outer and inner optimization objectives/problems. The outer problem corresponds to the validation objective and it is defined via an optimal solution to the inner problem which corresponds to a training objective. The paper, however, fails to make a reference to a rather rich literature on bilevel optimization (e.g., see [3-5] and references therein).\n\nThe approach, presented as Algorithm 1, does not seem different from [1] and [2] where hyperparameter optimization was considered for (kernel) support vector machines and (kernel) ridge regression. The data is initially split into k-folds (not necessarily of identical size) and each fold is used exactly once to define a validation objective whereas the complementary folds act as training data. The validation gradient is obtained by averaging the gradients of the k validation folds. Essentially, the same algorithm with k-fold cross-validation was considered in [1]. Thus, for me there does not seem to be any novelty in this approach and the paper itself.\n\nThe experiments involve synthetic regression and classification datasets but there are no novel insights that advance what is already known about the hyperparameter optimization (e.g., see [3]). For example, there is no intuition on the geometry of the optimization problem and the optimality of the outer optimization problem which is non-convex (e.g., see [7]), or dependence of the outer solution on the accuracy of the inner solution.\n\nReferences:\n\n[1] S. Keerthi, V. Sindhwani, and O. Chapelle. An Efficient Method for Gradient-Based Adaptation of Hyperparameters. NIPS 2007.\n[2] O. Chapelle, V. Vapnik, O. Bousquet, and S. Mukherjee. Choosing Multiple Parameters for Support Vector Machines. Machine Learning, 2002.\n\n[-3] L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil. Bilevel Programming for Hyperparameter Optimization and Meta-Learning. ICML 2018.\n[-4] G. Kunapuli, K.P. Bennet, J. Hu, and J-S. Pang. Bilevel Model Selection for SVMs. American Mathematical Society, 2008.\n[-5] E.S.H. Neto and A.R. de Pierro. On Perturbed Steepest Descent Methods with Inexact Line Search for Bilevel Convex Optimization. Journal of Mathematical Programming and Operations Research, 2011.\n[-6] B. Colson, P. Marcotte, and G. Savard. A Trust-Region Method for Nonlinear Bilevel Programming: Algorithms and Computational Experience. Computational Optimization and Applications, 2005.\n[-7] M. Janzamin, H. Sedghi, and A. Anandkumar. Beating the Perils of Non-Convexity: Guaranteed Training of Neural Networks using Tensor Method. arXiv preprint arXiv:1506.08473v3, 2016.\n", "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540817860734}, {"id": "Skl1tO4ZnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper66/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method to optimize for the cross-validation performance of a model by expressing said performance as a differentiable function of the model parameters and applying a gradient-based method. \n\nThe majority of the work is clear and well-written, and appears to be correct, but I find it lacking in originality. The main contribution over related work  (references cited under the \"Implicit Differentiation\" heading in section 2.2) appears to be that the hyperparameters are optimized with respect to a cross-validation loss rather than a held out validation set. That is, using K=1 in the CVGM (Algorithm 1) reduces to existing work. There is also a discussion of when the cross-validation loss will be differentiable, but no new results on this. I am not sure that these contributions justify the paper. \n\nThe experiments are also not particularly strong. Only synthetic data is considered. The logistic regression baseline for the classification application in section 5.2 is irrelevant, and the neural network baseline could be clarified. Does the baseline also use the optimal parameters in the last layer throughout training? If not, how much of the improvement of the CVGM over the baseline is due to this change? \n\nTo say that the CVGM is able to stably learn the “hyperparameters” of the network kernel in this setting seems like an exaggeration -- the neural network baseline also learns these “hyperparameters”. The difference is that they are optimized with respect to the CV loss rather than the training loss. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clearly written but incremental with respect to related work.", "review": "The paper proposes a method to optimize for the cross-validation performance of a model by expressing said performance as a differentiable function of the model parameters and applying a gradient-based method. \n\nThe majority of the work is clear and well-written, and appears to be correct, but I find it lacking in originality. The main contribution over related work  (references cited under the \"Implicit Differentiation\" heading in section 2.2) appears to be that the hyperparameters are optimized with respect to a cross-validation loss rather than a held out validation set. That is, using K=1 in the CVGM (Algorithm 1) reduces to existing work. There is also a discussion of when the cross-validation loss will be differentiable, but no new results on this. I am not sure that these contributions justify the paper. \n\nThe experiments are also not particularly strong. Only synthetic data is considered. The logistic regression baseline for the classification application in section 5.2 is irrelevant, and the neural network baseline could be clarified. Does the baseline also use the optimal parameters in the last layer throughout training? If not, how much of the improvement of the CVGM over the baseline is due to this change? \n\nTo say that the CVGM is able to stably learn the “hyperparameters” of the network kernel in this setting seems like an exaggeration -- the neural network baseline also learns these “hyperparameters”. The difference is that they are optimized with respect to the CV loss rather than the training loss. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540601974992}], "openreview_url": "https://openreview.net/forum?id=rJlMBjAcYX", "arxiv_id": "1805.07072", "paper_pdf": "papers/rJlMBjAcYX.pdf", "paper_pdf_sha256": "117605edac73b0b9385521ced3b7618a821c27299b58b0f2695af87dbf712967", "paper_pdf_bytes": 417525, "paper_pdf_source": "openreview", "code_url": "https://github.com/sbarratt/crossval", "code_repository": "sbarratt/crossval", "code_commit": "d275d40cd8c6109c844fdb9242bd588266971275", "code_archive": "repos/rJlMBjAcYX.zip", "code_archive_sha256": "9bade9e20b5f52f712df25fc8df3d65c88b0f0638cf82193274b593d96f877ca", "code_archive_bytes": 1105379, "code_file_count": 3, "code_extensions": {".ipynb": 3}, "github_disk_usage_kb": 1078, "github_languages": {"Jupyter Notebook": 1580579}, "github_archived": false, "github_pushed_at": "2018-05-21T00:53:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimizing-for-generalization-in-machine-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Sm8yTdWjai", "year": 2026, "status": "rejected", "title": "RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates", "authors": ["Ali Asad", "Stephen Obadinma", "Radin Shayanfar", "Xiaodan Zhu"], "authorids": ["~Ali_Asad1", "~Stephen_Obadinma1", "~Radin_Shayanfar1", "~Xiaodan_Zhu1"], "authors_source": "OpenReview API", "abstract": "We introduce RedDebate, a novel multi-agent debate framework that provides the foundation for Large Language Models (LLMs) to identify and mitigate their own unsafe behaviors. Existing AI safety approaches often rely on costly human evaluation or isolated single-model assessment, both constrained by scalability and prone to oversight failures. RedDebate employs collaborative argumentation among multiple LLMs across diverse debate scenarios, enabling them to critically evaluate one another’s reasoning and systematically uncover unsafe failure modes through fully automated red-teaming. We further integrate distinct long-term memory modules that preserve safety-relevant insights from debate interactions and leverage them during subsequent inference, facilitating continuous refinement of model behavior. Empirical evaluation on safety benchmarks across a diverse set of models demonstrates that RedDebate substantially reduces unsafe outputs. While debate alone allows LLMs to refine their behavior, the addition of memory modules yields further significant reductions. To the best of our knowledge, RedDebate is the first fully automated framework to unify multi-agent debate and red-teaming to progressively enhance LLM safety without human intervention.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Zp7AsgopE0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12258/Reviewer_vVnY"], "rating": 4, "soundness": 4, "presentation": 2, "contribution": 4, "confidence": 4, "summary": "This work proposes RedDebate, the first fully automated framework integrating **Multi-Agent Debate** with **Red Teaming** to enhance LLM safety without human intervention. Through collaborative debate among model agents, it systematically exposes and corrects potentially unsafe behaviors, significantly outperforming traditional single-agent self-critique or manual red teaming approaches.", "review_text": "This work proposes RedDebate, the first fully automated framework integrating **Multi-Agent Debate** with **Red Teaming** to enhance LLM safety without human intervention. Through collaborative debate among model agents, it systematically exposes and corrects potentially unsafe behaviors, significantly outperforming traditional single-agent self-critique or manual red teaming approaches.", "strengths": "1. First to propose and systematically implement the combination of \"multi-agent debate + automated red teaming\" for LLM safety alignment.\n2. Creative introduction of a **long- and short-term memory mechanism** to continuously accumulate safety experience, with experiments conducted on multiple memory variants.\n3. Compares multiple debate strategies and validates the corresponding effectiveness on standard safety benchmarks such as **HarmBench** and **CoSafe**.", "weaknesses": "1. Multi-agent debate and memory updates significantly increase inference cost (e.g., debate agents generate 1.3× more tokens per round than Self-Critique). While the authors argue that safety gains justify the cost, the approach may not be applicable in resource-constrained scenarios.\n2. Although the primary focus of this work is on improving safety, the experimental design might allow an overly \"safe\" agent to dominate the debate, potentially leading the entire debate process toward overly cautious conclusions. *Note: This is just a concern, not a confirmed issue.*\n\n**Minor Comments**\nOverall visual presentation (figures, diagrams) could be further improved.", "questions": "1. Has the framework been tested with agents of differing safety tendencies (e.g., pairing high-risk models with conservative models) to verify robustness in heterogeneous model settings?\n2. Have you considered possible approaches to reduce computation and deployment costs?\n3. How is safety ensured during the debate process? For example, a high-risk model may reveal harmful details during discussion — how is this controlled or prevented?\n\nI'm willing to increase my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes RedDebate, the first fully automated framework integrating **Multi-Agent Debate** with **Red Teaming** to enhance LLM safety without human intervention. Through collaborative debate among model agents, it systematically exposes and corrects potentially unsafe behaviors, significantly outperforming traditional single-agent self-critique or manual red teaming approaches.", "soundness": 4, "presentation": 2, "contribution": 4, "strengths": "1. First to propose and systematically implement the combination of \"multi-agent debate + automated red teaming\" for LLM safety alignment.\n2. Creative introduction of a **long- and short-term memory mechanism** to continuously accumulate safety experience, with experiments conducted on multiple memory variants.\n3. Compares multiple debate strategies and validates the corresponding effectiveness on standard safety benchmarks such as **HarmBench** and **CoSafe**.", "weaknesses": "1. Multi-agent debate and memory updates significantly increase inference cost (e.g., debate agents generate 1.3× more tokens per round than Self-Critique). While the authors argue that safety gains justify the cost, the approach may not be applicable in resource-constrained scenarios.\n2. Although the primary focus of this work is on improving safety, the experimental design might allow an overly \"safe\" agent to dominate the debate, potentially leading the entire debate process toward overly cautious conclusions. *Note: This is just a concern, not a confirmed issue.*\n\n**Minor Comments**\nOverall visual presentation (figures, diagrams) could be further improved.", "questions": "1. Has the framework been tested with agents of differing safety tendencies (e.g., pairing high-risk models with conservative models) to verify robustness in heterogeneous model settings?\n2. Have you considered possible approaches to reduce computation and deployment costs?\n3. How is safety ensured during the debate process? For example, a high-risk model may reveal harmful details during discussion — how is this controlled or prevented?\n\nI'm willing to increase my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762097976266}, {"id": "8gS9Lhy7H2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12258/Reviewer_bAEh"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces RedDebate, a novel framework designed to enhance LLM safety by automating the red-teaming process. Addressing the scalability limitations of human evaluation and the inherent blind spots of single-agent self-correction, the authors propose a multi-agent system where LLM agents collaboratively debate adversarial or unsafe prompts. This structured argumentation allows agents to critically evaluate one another's reasoning, systematically uncover unsafe failure modes, and iteratively refine their own responses. The paper's primary contributions are the fully automated framework itself, which unifies multi-agent debate with red-teaming; the exploration of different debate strategies (such as Socratic and Devil-Angel) to effectively elicit and correct unsafe behavior; and the integration of distinct long-term memory modules (including textual, parametric, and guardrail-based) that enable agents to learn persistently from previously identified failures. Empirical evaluations on safety benchmarks demonstrate that the RedDebate framework significantly reduces unsafe outputs without human intervention, with the guardrail-based memory approach yielding the most substantial safety improvements.", "review_text": "This paper introduces RedDebate, a novel framework designed to enhance LLM safety by automating the red-teaming process. Addressing the scalability limitations of human evaluation and the inherent blind spots of single-agent self-correction, the authors propose a multi-agent system where LLM agents collaboratively debate adversarial or unsafe prompts. This structured argumentation allows agents to critically evaluate one another's reasoning, systematically uncover unsafe failure modes, and iteratively refine their own responses. The paper's primary contributions are the fully automated framework itself, which unifies multi-agent debate with red-teaming; the exploration of different debate strategies (such as Socratic and Devil-Angel) to effectively elicit and correct unsafe behavior; and the integration of distinct long-term memory modules (including textual, parametric, and guardrail-based) that enable agents to learn persistently from previously identified failures. Empirical evaluations on safety benchmarks demonstrate that the RedDebate framework significantly reduces unsafe outputs without human intervention, with the guardrail-based memory approach yielding the most substantial safety improvements.", "strengths": "1. The paper addresses the well-motivated and critical problem of LLM safety, tackling the clear scalability and reliability limitations of existing human-led or single-agent evaluation methods.\n\n2. The manuscript is well-written, clearly articulating the proposed \"RedDebate\" framework, the experimental setup, and the subsequent analysis of the results.\n\n3. The work provides a great, novel perspective on AI safety by framing it as a learning problem solved through multi-agent interaction and, most notably, by integrating different long-term memory modules (textual, parametric, and guardrail-based) to ensure persistent safety improvements.", "weaknesses": "1. The technical novelty of the framework is somewhat limited, as it primarily integrates and applies existing concepts (multi-agent systems, red-teaming, and memory) rather than introducing entirely new techniques.\n\n2. The paper lacks sufficient baseline comparisons. While it includes a \"Self-Critique\" baseline, it would be strengthened by comparisons against other contemporary automated red-teaming or multi-agent debate frameworks.\n\n3. The evaluation is limited in its scope, focusing on a specific set of smaller-scale open-source models and two standard safety datasets. The findings' generalizability to larger, state-of-the-art models remains unclear.\n\n4. While some ablation studies are present (primarily in the appendix), the paper lacks a comprehensive ablation on the different components to clearly isolate their individual impact (e.g., the precise contribution of specific debate strategies versus the long-term memory).", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces RedDebate, a novel framework designed to enhance LLM safety by automating the red-teaming process. Addressing the scalability limitations of human evaluation and the inherent blind spots of single-agent self-correction, the authors propose a multi-agent system where LLM agents collaboratively debate adversarial or unsafe prompts. This structured argumentation allows agents to critically evaluate one another's reasoning, systematically uncover unsafe failure modes, and iteratively refine their own responses. The paper's primary contributions are the fully automated framework itself, which unifies multi-agent debate with red-teaming; the exploration of different debate strategies (such as Socratic and Devil-Angel) to effectively elicit and correct unsafe behavior; and the integration of distinct long-term memory modules (including textual, parametric, and guardrail-based) that enable agents to learn persistently from previously identified failures. Empirical evaluations on safety benchmarks demonstrate that the RedDebate framework significantly reduces unsafe outputs without human intervention, with the guardrail-based memory approach yielding the most substantial safety improvements.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper addresses the well-motivated and critical problem of LLM safety, tackling the clear scalability and reliability limitations of existing human-led or single-agent evaluation methods.\n\n2. The manuscript is well-written, clearly articulating the proposed \"RedDebate\" framework, the experimental setup, and the subsequent analysis of the results.\n\n3. The work provides a great, novel perspective on AI safety by framing it as a learning problem solved through multi-agent interaction and, most notably, by integrating different long-term memory modules (textual, parametric, and guardrail-based) to ensure persistent safety improvements.", "weaknesses": "1. The technical novelty of the framework is somewhat limited, as it primarily integrates and applies existing concepts (multi-agent systems, red-teaming, and memory) rather than introducing entirely new techniques.\n\n2. The paper lacks sufficient baseline comparisons. While it includes a \"Self-Critique\" baseline, it would be strengthened by comparisons against other contemporary automated red-teaming or multi-agent debate frameworks.\n\n3. The evaluation is limited in its scope, focusing on a specific set of smaller-scale open-source models and two standard safety datasets. The findings' generalizability to larger, state-of-the-art models remains unclear.\n\n4. While some ablation studies are present (primarily in the appendix), the paper lacks a comprehensive ablation on the different components to clearly isolate their individual impact (e.g., the precise contribution of specific debate strategies versus the long-term memory).", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761610158223}, {"id": "B9Ts7DNQ5O", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12258/Reviewer_LkJB"], "rating": 2, "soundness": 1, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces RedDebate, a multi-agent debate framework for LLM behavioral safety. The authors tested 3 main types of debating mechanisms: peer refinement, devil-angel, and Socratic - Socratic worked best among those. They also tested different mechanisms of long-term memory, which was integrated into multi-agent debate, and showed that incorporating LTM substantially improves behavioral safety. Among the LTM mechanisms, the guardrail approach was overall the most effective.", "review_text": "This paper introduces RedDebate, a multi-agent debate framework for LLM behavioral safety. The authors tested 3 main types of debating mechanisms: peer refinement, devil-angel, and Socratic - Socratic worked best among those. They also tested different mechanisms of long-term memory, which was integrated into multi-agent debate, and showed that incorporating LTM substantially improves behavioral safety. Among the LTM mechanisms, the guardrail approach was overall the most effective.", "strengths": "The idea of applying multi-agent debating is interesting and intuitive. The authors also thoroughly explored different mechanisms of long-term memory to augment debating, which is a novel approach. This paper is clearly written. The appendices provide helpful information that supplement the main text, such as the capability evaluation before vs after safety training and human validation of LlamaGuard.", "weaknesses": "I have a few major concerns about the evaluation, which prevent me from fully understanding the significance of the contribution. I'm open to raising my score if these concerns can be addressed during rebuttal.\n\n1. The current selection of benchmarks doesn't enable robust evaluation on the effectiveness of RedDebate. While HarmBench is a widely adopted benchmark in the safety literature, it's relatively small (with a few hundred examples) and potentially overfit by recent models. CoSafe doesn't seem to be an informative benchmark since the baseline error rate is already very low (7-8%) with little room for meaningful improvement. Including benchmarks that are bigger, more recent, and more able to distinguish different models/methods will provide substantially more information about how effective RedDebate is. nvidia/Aegis-AI-Content-Safety-Dataset-2.0 might be a good resource for this.\n\n2. The evaluation in Table 2 seems unfair for the Self-Critique baseline. Self-Critique outperforms SReD without LTM on HarmBench by a large margin - only when LTM mechanisms are included does SReD beat Self-Critique. My interpretation of this is that debating may not be more effective than self-reflection for mitigating safety, and self-reflection is potentially cheaper since it doesn't require multiple models, which undermines the contribution of RedDebate.", "questions": "1. Line 298: the term \"agreement rate\" is misleading when the metric quantifies the switch rate from unsafe to safe rather than agreement. Switching from unsafe to safe doesn't necessarily indicate agreement, and agreement doesn't always lead to unsafe to safe switches. Consider using a more accurate metric name.\n\n2. Any insights on why Self-Critique consistently outperforms SReD on HarmBench but underperforms on CoSafe?\n\n3. Table 2: Could you equip Self-Critique with LTM? I wonder if that would outperform SReD + LTM, at least on HarmBench\n\n4. How do different evaluation metrics change over early rounds? Table 7 only shows rounds 3-5, but how about rounds 1-2?\n\n5. What is the inference, computational, and time costs of the various debate methods evaluated in the main text? Is this framework realistic to be deployed at inference time?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces RedDebate, a multi-agent debate framework for LLM behavioral safety. The authors tested 3 main types of debating mechanisms: peer refinement, devil-angel, and Socratic - Socratic worked best among those. They also tested different mechanisms of long-term memory, which was integrated into multi-agent debate, and showed that incorporating LTM substantially improves behavioral safety. Among the LTM mechanisms, the guardrail approach was overall the most effective.", "soundness": 1, "presentation": 3, "contribution": 2, "strengths": "The idea of applying multi-agent debating is interesting and intuitive. The authors also thoroughly explored different mechanisms of long-term memory to augment debating, which is a novel approach. This paper is clearly written. The appendices provide helpful information that supplement the main text, such as the capability evaluation before vs after safety training and human validation of LlamaGuard.", "weaknesses": "I have a few major concerns about the evaluation, which prevent me from fully understanding the significance of the contribution. I'm open to raising my score if these concerns can be addressed during rebuttal.\n\n1. The current selection of benchmarks doesn't enable robust evaluation on the effectiveness of RedDebate. While HarmBench is a widely adopted benchmark in the safety literature, it's relatively small (with a few hundred examples) and potentially overfit by recent models. CoSafe doesn't seem to be an informative benchmark since the baseline error rate is already very low (7-8%) with little room for meaningful improvement. Including benchmarks that are bigger, more recent, and more able to distinguish different models/methods will provide substantially more information about how effective RedDebate is. nvidia/Aegis-AI-Content-Safety-Dataset-2.0 might be a good resource for this.\n\n2. The evaluation in Table 2 seems unfair for the Self-Critique baseline. Self-Critique outperforms SReD without LTM on HarmBench by a large margin - only when LTM mechanisms are included does SReD beat Self-Critique. My interpretation of this is that debating may not be more effective than self-reflection for mitigating safety, and self-reflection is potentially cheaper since it doesn't require multiple models, which undermines the contribution of RedDebate.", "questions": "1. Line 298: the term \"agreement rate\" is misleading when the metric quantifies the switch rate from unsafe to safe rather than agreement. Switching from unsafe to safe doesn't necessarily indicate agreement, and agreement doesn't always lead to unsafe to safe switches. Consider using a more accurate metric name.\n\n2. Any insights on why Self-Critique consistently outperforms SReD on HarmBench but underperforms on CoSafe?\n\n3. Table 2: Could you equip Self-Critique with LTM? I wonder if that would outperform SReD + LTM, at least on HarmBench\n\n4. How do different evaluation metrics change over early rounds? Table 7 only shows rounds 3-5, but how about rounds 1-2?\n\n5. What is the inference, computational, and time costs of the various debate methods evaluated in the main text? Is this framework realistic to be deployed at inference time?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761516022333}, {"id": "iLekZ3PQBd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12258/Reviewer_TaUG"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper introduces RedDebate, a MAS debate framework designed to identify and mitigate unsafe behaviors. RedDebate employs fully automated red-teaming to uncover unsafe patterns and uses long-term memory to preserve these insights, leveraging them in subsequent inference. Experiments demonstrate the effectiveness of RedDebate.", "review_text": "This paper introduces RedDebate, a MAS debate framework designed to identify and mitigate unsafe behaviors. RedDebate employs fully automated red-teaming to uncover unsafe patterns and uses long-term memory to preserve these insights, leveraging them in subsequent inference. Experiments demonstrate the effectiveness of RedDebate.", "strengths": "- The concept of a fully automated MAS safety enhancement framework is very interesting, and I believe this is a highly practical research direction.\n\n- The authors have conducted very detailed experiments, designing various debate strategies and memory modules. They also performed ablation studies on various hyperparameters to evaluate the effectiveness of RedDebate.", "weaknesses": "- The paper lacks a discussion on the additional time overhead introduced by RedDebate. Although this point is mentioned in the limitations section, I believe it is still necessary to measure and present the time and computational resources consumed, as efficiency and cost are critical in many real-world scenarios.\n\n- The paper lacks comparison with stronger baselines. The authors only compare RedDebate with Self-Critique, while potentially overlooking other work with similar objectives, such as [1] and [2].\n\n[1] arxiv.org/abs/2305.14325\n\n[2] arxiv.org/abs/2305.19118", "questions": "- I am curious whether larger-scale commercial models (e.g., gpt, claude) could benefit from this framework. Or would they simply refuse to answer harmful questions and fail to correct other agents' responses, similar to the behavior of gpt-oss as shown in Appendix B?\n\n- What is the intended attack/defense scenario? Is RedDebate meant to be a pre-processing or training stage for a MAS, where its safety is enhanced through many rounds of debate before being deployed on real-world tasks? Or is the debate itself the end task?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces RedDebate, a MAS debate framework designed to identify and mitigate unsafe behaviors. RedDebate employs fully automated red-teaming to uncover unsafe patterns and uses long-term memory to preserve these insights, leveraging them in subsequent inference. Experiments demonstrate the effectiveness of RedDebate.", "soundness": 4, "presentation": 3, "contribution": 2, "strengths": "- The concept of a fully automated MAS safety enhancement framework is very interesting, and I believe this is a highly practical research direction.\n\n- The authors have conducted very detailed experiments, designing various debate strategies and memory modules. They also performed ablation studies on various hyperparameters to evaluate the effectiveness of RedDebate.", "weaknesses": "- The paper lacks a discussion on the additional time overhead introduced by RedDebate. Although this point is mentioned in the limitations section, I believe it is still necessary to measure and present the time and computational resources consumed, as efficiency and cost are critical in many real-world scenarios.\n\n- The paper lacks comparison with stronger baselines. The authors only compare RedDebate with Self-Critique, while potentially overlooking other work with similar objectives, such as [1] and [2].\n\n[1] arxiv.org/abs/2305.14325\n\n[2] arxiv.org/abs/2305.19118", "questions": "- I am curious whether larger-scale commercial models (e.g., gpt, claude) could benefit from this framework. Or would they simply refuse to answer harmful questions and fail to correct other agents' responses, similar to the behavior of gpt-oss as shown in Appendix B?\n\n- What is the intended attack/defense scenario? Is RedDebate meant to be a pre-processing or training stage for a MAS, where its safety is enhanced through many rounds of debate before being deployed on real-world tasks? Or is the debate itself the end task?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1760640945593}], "openreview_url": "https://openreview.net/forum?id=Sm8yTdWjai", "arxiv_id": "2506.11083", "paper_pdf": "papers/Sm8yTdWjai.pdf", "paper_pdf_sha256": "a6a38a07c70025b7780108c8f02baaa92689d52bb5808415b672a888178dc690", "paper_pdf_bytes": 2279544, "paper_pdf_source": "openreview", "code_url": "https://github.com/aliasad059/RedDebate", "code_repository": "aliasad059/RedDebate", "code_commit": "713bcf09a7ea63247453d8df644b53c833ce13d2", "code_archive": "repos/Sm8yTdWjai.zip", "code_archive_sha256": "7fc7a97a6032d93a48e249cdeeabcb4f1dda18e13d268055f72eedc30df19170", "code_archive_bytes": 183869, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 491, "github_languages": {"Python": 148182}, "github_archived": false, "github_pushed_at": "2026-07-04T19:01:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/reddebate-safer-responses-through-multi-agent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nD5tbHBfut", "year": 2025, "status": "rejected", "title": "LRVS-Fashion: Extending Visual Search with Referring Instructions", "authors": ["Simon Lepage", "Jeremie Mary", "David Picard"], "authorids": ["~Simon_Lepage1", "~Jeremie_Mary1", "~David_Picard1"], "authors_source": "OpenReview API", "abstract": "This paper introduces a new challenge for image similarity search in the context of fashion, addressing the inherent ambiguity in this domain stemming from complex images. We present Referred Visual Search (RVS), a task allowing users to define more precisely the desired similarity, following recent interest in the industry. We release a new large public dataset, LRVS-Fashion, consisting of 272k fashion products with 842k images extracted from fashion catalogs, designed explicitly for this task. However, unlike traditional visual search methods in the industry, we demonstrate that superior performance can be achieved by bypassing explicit object detection and adopting weakly-supervised conditional contrastive learning on image tuples. Our method is lightweight and demonstrates robustness, reaching Recall at one superior to strong detection-based baselines against 2M distractors.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "3mro9k2rxp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7576/Reviewer_DWgA"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper introduces the Referred Visual Search (RVS) task, addressing the challenges of image similarity search in the fashion domain, particularly in scenarios where product images need to be retrieved from query images containing multiple products or from model pictures. The authors present a large dataset, LRVS-Fashion, containing 272k fashion products and 842k images, designed for this task. They propose a novel method based on weakly-supervised conditional contrastive learning, achieving superior performance without relying on explicit cropping or segmentation. The study demonstrates the effectiveness of their approach in comparison to a variety of baselines. The paper is overall well written.", "review_text": "This paper introduces the Referred Visual Search (RVS) task, addressing the challenges of image similarity search in the fashion domain, particularly in scenarios where product images need to be retrieved from query images containing multiple products or from model pictures. The authors present a large dataset, LRVS-Fashion, containing 272k fashion products and 842k images, designed for this task. They propose a novel method based on weakly-supervised conditional contrastive learning, achieving superior performance without relying on explicit cropping or segmentation. The study demonstrates the effectiveness of their approach in comparison to a variety of baselines. The paper is overall well written.", "strengths": "The introduction of RVS as a task for fashion images addresses a gap in the current literature and reveals significant value for practical applications such as e-commerce systems.\n\nThe LRVS-Fashion dataset is extensive in terms of its scale, organization, and cleaning process, providing valuable resources for future research and enhancing the field's accessibility.\n\nThe proposed approach, based on conditional contrastive learning, is intuitive yet elegant. By avoiding explicit object detection or segmentation, it is efficient and easier to implement in various applications, particularly by reducing the extra stages of data collection or model design. In addition, it demonstrates impressive robustness against a large number of distractors, which is critical in real-world applications.\n\nThe paper provides a thorough literature review and extensive comparisons with existing methods. For example, it compares with compositional image retrieval methods that share a conditional learning requirement but targeting at different application scenarios. The experiment setup is clear and the overall results are convincing in terms of accuracy and efficiency.", "weaknesses": "While the experimental results are solid, it could be more comprehensive if the analysis can go deeper into the impact of objects' visibility in query images considering object sizes, occlusions or viewpoints. Or, a simpler referring aspect could be segmentation mask area or its ratio to a detected object bounding boxes? One example is how the retrieval accuracy changes among shirts because they are often partially visible due to overlaid with outwears.", "questions": "The study focuses primarily on fashion images and leverages LAION-5B to create the benchmark dataset. While it serves as a valuable testbed for future research, it would be interesting to discuss how to construct larger datasets in real-world scenarios and the challenges associated with generalizability to close domains such as street outfit photos that have noisy backgrounds, or even beyond fashion such as furniture, landmarks, etc. This would further strengthen the impact of this work?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Referred Visual Search (RVS) task, addressing the challenges of image similarity search in the fashion domain, particularly in scenarios where product images need to be retrieved from query images containing multiple products or from model pictures. The authors present a large dataset, LRVS-Fashion, containing 272k fashion products and 842k images, designed for this task. They propose a novel method based on weakly-supervised conditional contrastive learning, achieving superior performance without relying on explicit cropping or segmentation. The study demonstrates the effectiveness of their approach in comparison to a variety of baselines. The paper is overall well written.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "The introduction of RVS as a task for fashion images addresses a gap in the current literature and reveals significant value for practical applications such as e-commerce systems.\n\nThe LRVS-Fashion dataset is extensive in terms of its scale, organization, and cleaning process, providing valuable resources for future research and enhancing the field's accessibility.\n\nThe proposed approach, based on conditional contrastive learning, is intuitive yet elegant. By avoiding explicit object detection or segmentation, it is efficient and easier to implement in various applications, particularly by reducing the extra stages of data collection or model design. In addition, it demonstrates impressive robustness against a large number of distractors, which is critical in real-world applications.\n\nThe paper provides a thorough literature review and extensive comparisons with existing methods. For example, it compares with compositional image retrieval methods that share a conditional learning requirement but targeting at different application scenarios. The experiment setup is clear and the overall results are convincing in terms of accuracy and efficiency.", "weaknesses": "While the experimental results are solid, it could be more comprehensive if the analysis can go deeper into the impact of objects' visibility in query images considering object sizes, occlusions or viewpoints. Or, a simpler referring aspect could be segmentation mask area or its ratio to a detected object bounding boxes? One example is how the retrieval accuracy changes among shirts because they are often partially visible due to overlaid with outwears.", "questions": "The study focuses primarily on fashion images and leverages LAION-5B to create the benchmark dataset. While it serves as a valuable testbed for future research, it would be interesting to discuss how to construct larger datasets in real-world scenarios and the challenges associated with generalizability to close domains such as street outfit photos that have noisy backgrounds, or even beyond fashion such as furniture, landmarks, etc. This would further strengthen the impact of this work?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730679164652}, {"id": "eXCYBkGJN0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7576/Reviewer_MVgE"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces Referred Visual Search (RVS) in the fashion domain. The paper makes two main contributions: (1) A new large-scale dataset called LRVS-Fashion and (2) A weakly-supervised method that outperforms detection-based approaches.", "review_text": "This paper introduces Referred Visual Search (RVS) in the fashion domain. The paper makes two main contributions: (1) A new large-scale dataset called LRVS-Fashion and (2) A weakly-supervised method that outperforms detection-based approaches.", "strengths": "1. The paper proposes a high-quality dataset that will be beneficial to the field\n2. The paper is well-written and easy to read.", "weaknesses": "1. The proposed task is like a special case for the task of composed image retrieval(CIR). However, the experiment is not compared with recent CIR methods, which makes the result not convincing. \n2. The proposed method shows high performance in LRVS-F, as shown in Table 2, nearly 99% in Cat@1. These superior performances raise the question of the proposed dataset's motivation. Is there any specific problem the dataset wants to diagnose?", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Referred Visual Search (RVS) in the fashion domain. The paper makes two main contributions: (1) A new large-scale dataset called LRVS-Fashion and (2) A weakly-supervised method that outperforms detection-based approaches.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper proposes a high-quality dataset that will be beneficial to the field\n2. The paper is well-written and easy to read.", "weaknesses": "1. The proposed task is like a special case for the task of composed image retrieval(CIR). However, the experiment is not compared with recent CIR methods, which makes the result not convincing. \n2. The proposed method shows high performance in LRVS-F, as shown in Table 2, nearly 99% in Cat@1. These superior performances raise the question of the proposed dataset's motivation. Is there any specific problem the dataset wants to diagnose?", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "no", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730650739774}, {"id": "uSn6PI6afI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7576/Reviewer_risT"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper proposes a new visual search task with referring instructions as the hints.\nIt also creates a new large public dataset, LRVS-Fashion, as the benchmark.\nThe authors also introduce a CLIP-like simple baseline.", "review_text": "This paper proposes a new visual search task with referring instructions as the hints.\nIt also creates a new large public dataset, LRVS-Fashion, as the benchmark.\nThe authors also introduce a CLIP-like simple baseline.", "strengths": "1. The efforts on collecting the benckmark is worthy of applause. This dataset could be potentially used for other fashion-related tasks.\n2. The writing of this paper is good and easy to follow.", "weaknesses": "1. I am not convinced on the motivation of the proposed task. I cannot fully imagine a real application scenario for the proposed Referred Visual Search. According to my experience and understanding on fashion/ads, the query image (ie, dressed human) and the target image (ie, garment, product and so on) shown in Figure 1 are normally from the same product set which provided by the advertiser (at least they share some common information, eg, brand name). In that case, we don't need to do large-scale image-base retrieval, which is implementation-wise expensive in the industry. The only case I understand is letting the user themselves upload their individual query images and do a cross-domain image retrieval. However, the target and query image used in this paper are from the same domain and thus    cannot model the real world problem, although I understand collecting cross-domain benchmark is even harder.\n2. I believe the proposed RVS is just a sub-task of Composed Image Retrieval (CIR), while their relationship is not clearly discussed (between L147 and L153) and the latest CIR related works were not cited. Regarding \"rather than modifying the image\" in L151, I think getting the feature of the ROI of the query image (Figure 4) is also kind of modifying the query image feature.\n3. The proposed model arch is okayish but not that innovative (L081) to me. I agree this can be regarded as a strong baseline but don't think this widely used arch can be claimed as innovative.\n4. I think some zero-shot methods (no training needed) should be considered as baselines too, eg, multimodal LLM that can fuse the query image feature and the instructions.", "questions": "Please refer to the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new visual search task with referring instructions as the hints.\nIt also creates a new large public dataset, LRVS-Fashion, as the benchmark.\nThe authors also introduce a CLIP-like simple baseline.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The efforts on collecting the benckmark is worthy of applause. This dataset could be potentially used for other fashion-related tasks.\n2. The writing of this paper is good and easy to follow.", "weaknesses": "1. I am not convinced on the motivation of the proposed task. I cannot fully imagine a real application scenario for the proposed Referred Visual Search. According to my experience and understanding on fashion/ads, the query image (ie, dressed human) and the target image (ie, garment, product and so on) shown in Figure 1 are normally from the same product set which provided by the advertiser (at least they share some common information, eg, brand name). In that case, we don't need to do large-scale image-base retrieval, which is implementation-wise expensive in the industry. The only case I understand is letting the user themselves upload their individual query images and do a cross-domain image retrieval. However, the target and query image used in this paper are from the same domain and thus    cannot model the real world problem, although I understand collecting cross-domain benchmark is even harder.\n2. I believe the proposed RVS is just a sub-task of Composed Image Retrieval (CIR), while their relationship is not clearly discussed (between L147 and L153) and the latest CIR related works were not cited. Regarding \"rather than modifying the image\" in L151, I think getting the feature of the ROI of the query image (Figure 4) is also kind of modifying the query image feature.\n3. The proposed model arch is okayish but not that innovative (L081) to me. I agree this can be regarded as a strong baseline but don't think this widely used arch can be claimed as innovative.\n4. I think some zero-shot methods (no training needed) should be considered as baselines too, eg, multimodal LLM that can fuse the query image feature and the instructions.", "questions": "Please refer to the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730576329658}, {"id": "Xe5VwIjmRl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7576/Reviewer_jqz1"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper introduces a new task, Referred Visual Search, which retrieves specified products within a given image based on user requirements, and constructs a new dataset for this task. Additionally, a CLIP-based model architecture is proposed and its performance is validated on the newly created dataset.", "review_text": "This paper introduces a new task, Referred Visual Search, which retrieves specified products within a given image based on user requirements, and constructs a new dataset for this task. Additionally, a CLIP-based model architecture is proposed and its performance is validated on the newly created dataset.", "strengths": "1.\tThe proposed new task has practical application value in the e-commerce domain.\n2.\tThe illustrations are clear and detailed, facilitating understanding.", "weaknesses": "1.\tThe significance of the proposed task needs further clarification. What distinguishes Referred Visual Search from text-guided visual localization or composed image retrieval? Can existing methods in these areas be adapted to this task? \n2.\tThe paper lacks novelty, as applying the CLIP architecture in the e-commerce domain has been previously explored [a-c]. The authors need to further explain the innovation of their proposed method compared to existing methods. \n3.\tThe experiments require further refinement. Given that the authors utilized methods from composed image retrieval for performance comparison, it is essential to validate the proposed method on the FashionIQ and Shoes datasets within this context to effectively demonstrate its efficacy.\n4.\tThe proposed method incorporates the Grounding DINO and SAM models; thus, a discussion on the fairness of performance comparisons with other methods is warranted. Additionally, the time and space complexity of the proposed method should be compared against other representative methods to provide a comprehensive evaluation.\n\n[a] Han Y, Zhang L, Chen Q, et al. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 15028-15038.\n\n[b] Zhou J, Zheng X, Lyu Y, et al. E-clip: Towards label-efficient event-based open-world understanding by clip[J]. arXiv preprint arXiv:2308.03135, 2023.\n\n[c] Han X, Zhu X, Yu L, et al. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 2669-2680.", "questions": "Please refer to the Weaknesses section and address my concerns regarding the task setup, methodological innovation, and performance of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new task, Referred Visual Search, which retrieves specified products within a given image based on user requirements, and constructs a new dataset for this task. Additionally, a CLIP-based model architecture is proposed and its performance is validated on the newly created dataset.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1.\tThe proposed new task has practical application value in the e-commerce domain.\n2.\tThe illustrations are clear and detailed, facilitating understanding.", "weaknesses": "1.\tThe significance of the proposed task needs further clarification. What distinguishes Referred Visual Search from text-guided visual localization or composed image retrieval? Can existing methods in these areas be adapted to this task? \n2.\tThe paper lacks novelty, as applying the CLIP architecture in the e-commerce domain has been previously explored [a-c]. The authors need to further explain the innovation of their proposed method compared to existing methods. \n3.\tThe experiments require further refinement. Given that the authors utilized methods from composed image retrieval for performance comparison, it is essential to validate the proposed method on the FashionIQ and Shoes datasets within this context to effectively demonstrate its efficacy.\n4.\tThe proposed method incorporates the Grounding DINO and SAM models; thus, a discussion on the fairness of performance comparisons with other methods is warranted. Additionally, the time and space complexity of the proposed method should be compared against other representative methods to provide a comprehensive evaluation.\n\n[a] Han Y, Zhang L, Chen Q, et al. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 15028-15038.\n\n[b] Zhou J, Zheng X, Lyu Y, et al. E-clip: Towards label-efficient event-based open-world understanding by clip[J]. arXiv preprint arXiv:2308.03135, 2023.\n\n[c] Han X, Zhu X, Yu L, et al. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 2669-2680.", "questions": "Please refer to the Weaknesses section and address my concerns regarding the task setup, methodological innovation, and performance of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730272927090}], "openreview_url": "https://openreview.net/forum?id=nD5tbHBfut", "arxiv_id": "2306.02928", "paper_pdf": "papers/nD5tbHBfut.pdf", "paper_pdf_sha256": "137cad4fbb0efdea90430a7f59ad782fb9dd7ed09d857176d74d330486912b72", "paper_pdf_bytes": 7803827, "paper_pdf_source": "openreview", "code_url": "https://github.com/Simon-Lepage/LRVSF-Benchmark", "code_repository": "Simon-Lepage/LRVSF-Benchmark", "code_commit": "46e5e943b3db0962daa843fc6dfa2a78c8182382", "code_archive": "repos/nD5tbHBfut.zip", "code_archive_sha256": "324495c2d7fa17e52e3893b48fdfa0061619e09f6d2b991e6b2216b76e631d93", "code_archive_bytes": 57740, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 58, "github_languages": {"Python": 14949, "Makefile": 382}, "github_archived": false, "github_pushed_at": "2024-05-15T15:04:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/weakly-supervised-conditional-embedding-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Tpk0p9QBM6", "year": 2024, "status": "rejected", "title": "Computing Low-Entropy Couplings for Large-Support Distributions", "authors": ["Samuel Sokota", "Dylan Sam", "Christian Schroeder de Witt", "Spencer Compton", "Jakob Nicolaus Foerster", "J Zico Kolter"], "authorids": ["~Samuel_Sokota1", "~Dylan_Sam1", "~Christian_Schroeder_de_Witt1", "~Spencer_Compton1", "~Jakob_Nicolaus_Foerster1", "~J_Zico_Kolter1"], "authors_source": "OpenReview API", "abstract": "A minimum-entropy coupling is a joint probability distribution having minimum joint entropy among all joint distributions with given pre-specified marginals. While provable approximation algorithms for a minimum-entropy coupling exist, they take log-linear time in the size of the support of the marginal distributions. Thus, applications involving very large-support distributions instead use a class of heuristic iterative minimum-entropy coupling (IMEC) algorithms. Unfortunately, existing IMEC algorithms are limited to specific classes of distributions, prohibiting applications involving general large-support distributions. In this work, we resolve this issue by making three main contributions: 1) We unify existing IMEC algorithms under a single formalism using sets of partitions. 2) We derive a new IMEC instance from this formalism, which we call ARIMEC, that, unlike existing IMEC algorithms, can be applied to arbitrary discrete distributions; furthermore, we introduce associated operations that make ARIMEC efficient in practice. 3) We empirically illustrate the utility of ARIMEC for both Markov coding games and steganography.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "lXqtTvvdsn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4090/Reviewer_caMu"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Minimum entropy coupling (MEC) is the following problem. Given marginal distributions of two finitely supported random variables X and Y, find a joint distribution \\gamma on X and Y (called a “coupling” of X and Y) s.t. the entropy of \\gamma is as small as possible. In addition to being a natural problem from a theoretical point of view, MEC also has various practical applications, including steganography (the art of hiding secret messages in plain sight).\n\nThe authors propose a unified framework that captures previous approaches to MEC and propose a novel *heuristic* algorithm for MEC called ARIMEC. That is, their method does not have provable guarantees regarding the entropy achieved by the output coupling. The reason for this is that MEC is NP-hard (where the instance size is defined to be the support size of X and Y) and provable approximation algorithms for MEC run in O(N log N) time, where N is the support size. In many interesting practical applications, the support size N is often extremely large so even linear-in-N time is considered too slow. Thus, the authors turn to heuristic algorithms and argue for the utility of their approach via empirical evaluation.", "review_text": "Minimum entropy coupling (MEC) is the following problem. Given marginal distributions of two finitely supported random variables X and Y, find a joint distribution \\gamma on X and Y (called a “coupling” of X and Y) s.t. the entropy of \\gamma is as small as possible. In addition to being a natural problem from a theoretical point of view, MEC also has various practical applications, including steganography (the art of hiding secret messages in plain sight).\n\nThe authors propose a unified framework that captures previous approaches to MEC and propose a novel *heuristic* algorithm for MEC called ARIMEC. That is, their method does not have provable guarantees regarding the entropy achieved by the output coupling. The reason for this is that MEC is NP-hard (where the instance size is defined to be the support size of X and Y) and provable approximation algorithms for MEC run in O(N log N) time, where N is the support size. In many interesting practical applications, the support size N is often extremely large so even linear-in-N time is considered too slow. Thus, the authors turn to heuristic algorithms and argue for the utility of their approach via empirical evaluation.", "strengths": "The unifying framework using collections of partitions is relatively simple and nicely captures previous approaches, though its exposition could be improved. In addition, the application to steganography is interesting. Heuristic MEC algorithms, despite their lack of rigorous guarantees, could lead to interesting future work on steganography using large language models (LLMs) since for LLMs we have full access to the generating distribution.", "weaknesses": "The paper's failure to address several important issues, in addition to the numerous ambiguities in the exposition, diminishes the paper's overall quality. Therefore, I am inclined to reject the current version of the paper. My concerns include the following.\n\n- **Basic MEC setup is unclear.** Throughout the paper, the authors seem to *implicitly* assume that X, Y are both vectors (or strings). In fact, the prefix tree in Section 4 does not even make sense unless elements of X are strings. This implicit vector assumption is also used in Section 2.3, Algorithm 1 and 2. However, this assumption on the structure of X and Y doesn’t seem to be stated explicitly anywhere in the paper.\n    \n    Also, what kind of access do we have to the marginals of X and Y? Do we get black-box queries to the probability mass evaluations? Is it a sampling oracle?\n    \n- **Notion of efficiency is never formally defined.** The main reason for using heuristic MEC algorithms instead of the provable approximation algorithms is to avoid the log-linear-in-N run time. Computational efficiency, at least in CS, is always be defined relative to some problem instance size (i.e., it is an asymptotic notion). If N is “too big”, then what is the right index for the instance size for MEC? This would make sense if one explicitly assumed that X is supported on {0,1}^n and Y is supported on {0,1}^m, and we call any quantity “too large” or “intractably large” if it grows superpolynomially in n or m.\n\n- **Small support size of conditional distributions does not imply efficient sampleability (Condition 4.6).** Suppose X is a random vector in {0,1}^n defined as the output distribution of a pseudorandom generator (PRG), i.e., the pushforward of the uniform distribution over {0,1}^s through the PRG. Clearly, each conditional distribution arising in the autoregressive form of this distribution is supported on {0,1}, which is of size 2. However, these conditional distributions are not even efficiently computable (since they are given by a PRG).\n\n- **Missing run-time analysis of subroutines.** In Algorithm 3 (IMEC), is it clear that the optimization over the collection of partitions U can be implemented efficiently?\n\n- **Motivation for ARIMEC.** In what sense is ARIMEC better than previous approaches? In the regimes where TIMEC and FIMEC performs well, is it expected that ARIMEC performs not worse than these two approaches? Also, what is the motivation behind using partitions defined using the prefix tree?\n\n- **Missing details on the steganography task.** Details of the steganography experiment are missing (even including Appendix D), which makes it hard to understand what experiment is exactly being conducted here. Could the authors please expand on the last paragraph of Section 5?", "questions": "- What kind of access do we have to the marginals of X and Y? Do we get black-box queries to the probability mass evaluations? Is it a sampling oracle?\n- In Condition 2.5, what qualifies as “small” support size and what quantifies as “intractably large”?\n- In Algorithm 3 (IMEC), is it clear that the optimization over the collection of partitions U can be implemented efficiently?\n- If the encoding is perfectly secure (i.e., encoding of ciphertext X into stegotext S is deterministic) and the distribution of stegotext S and covertext C is the same, doesn’t this mean that one can “hallucinate” secret messages from innocuous text?\n\n**Editorial comments**\n\n- The NP-hard reference is rather misleading. For the NP-hardness result, the instances are indexed by N, the support size. Even if MEC were in P, this would not suffice for the setting this paper is interested in.\n- In the Algorithm boxes, the subscript 1:j-1 doesn’t make sense for j = 1.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Minimum entropy coupling (MEC) is the following problem. Given marginal distributions of two finitely supported random variables X and Y, find a joint distribution \\gamma on X and Y (called a “coupling” of X and Y) s.t. the entropy of \\gamma is as small as possible. In addition to being a natural problem from a theoretical point of view, MEC also has various practical applications, including steganography (the art of hiding secret messages in plain sight).\n\nThe authors propose a unified framework that captures previous approaches to MEC and propose a novel *heuristic* algorithm for MEC called ARIMEC. That is, their method does not have provable guarantees regarding the entropy achieved by the output coupling. The reason for this is that MEC is NP-hard (where the instance size is defined to be the support size of X and Y) and provable approximation algorithms for MEC run in O(N log N) time, where N is the support size. In many interesting practical applications, the support size N is often extremely large so even linear-in-N time is considered too slow. Thus, the authors turn to heuristic algorithms and argue for the utility of their approach via empirical evaluation.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The unifying framework using collections of partitions is relatively simple and nicely captures previous approaches, though its exposition could be improved. In addition, the application to steganography is interesting. Heuristic MEC algorithms, despite their lack of rigorous guarantees, could lead to interesting future work on steganography using large language models (LLMs) since for LLMs we have full access to the generating distribution.", "weaknesses": "The paper's failure to address several important issues, in addition to the numerous ambiguities in the exposition, diminishes the paper's overall quality. Therefore, I am inclined to reject the current version of the paper. My concerns include the following.\n\n- **Basic MEC setup is unclear.** Throughout the paper, the authors seem to *implicitly* assume that X, Y are both vectors (or strings). In fact, the prefix tree in Section 4 does not even make sense unless elements of X are strings. This implicit vector assumption is also used in Section 2.3, Algorithm 1 and 2. However, this assumption on the structure of X and Y doesn’t seem to be stated explicitly anywhere in the paper.\n    \n    Also, what kind of access do we have to the marginals of X and Y? Do we get black-box queries to the probability mass evaluations? Is it a sampling oracle?\n    \n- **Notion of efficiency is never formally defined.** The main reason for using heuristic MEC algorithms instead of the provable approximation algorithms is to avoid the log-linear-in-N run time. Computational efficiency, at least in CS, is always be defined relative to some problem instance size (i.e., it is an asymptotic notion). If N is “too big”, then what is the right index for the instance size for MEC? This would make sense if one explicitly assumed that X is supported on {0,1}^n and Y is supported on {0,1}^m, and we call any quantity “too large” or “intractably large” if it grows superpolynomially in n or m.\n\n- **Small support size of conditional distributions does not imply efficient sampleability (Condition 4.6).** Suppose X is a random vector in {0,1}^n defined as the output distribution of a pseudorandom generator (PRG), i.e., the pushforward of the uniform distribution over {0,1}^s through the PRG. Clearly, each conditional distribution arising in the autoregressive form of this distribution is supported on {0,1}, which is of size 2. However, these conditional distributions are not even efficiently computable (since they are given by a PRG).\n\n- **Missing run-time analysis of subroutines.** In Algorithm 3 (IMEC), is it clear that the optimization over the collection of partitions U can be implemented efficiently?\n\n- **Motivation for ARIMEC.** In what sense is ARIMEC better than previous approaches? In the regimes where TIMEC and FIMEC performs well, is it expected that ARIMEC performs not worse than these two approaches? Also, what is the motivation behind using partitions defined using the prefix tree?\n\n- **Missing details on the steganography task.** Details of the steganography experiment are missing (even including Appendix D), which makes it hard to understand what experiment is exactly being conducted here. Could the authors please expand on the last paragraph of Section 5?", "questions": "- What kind of access do we have to the marginals of X and Y? Do we get black-box queries to the probability mass evaluations? Is it a sampling oracle?\n- In Condition 2.5, what qualifies as “small” support size and what quantifies as “intractably large”?\n- In Algorithm 3 (IMEC), is it clear that the optimization over the collection of partitions U can be implemented efficiently?\n- If the encoding is perfectly secure (i.e., encoding of ciphertext X into stegotext S is deterministic) and the distribution of stegotext S and covertext C is the same, doesn’t this mean that one can “hallucinate” secret messages from innocuous text?\n\n**Editorial comments**\n\n- The NP-hard reference is rather misleading. For the NP-hardness result, the instances are indexed by N, the support size. Even if MEC were in P, this would not suffice for the setting this paper is interested in.\n- In the Algorithm boxes, the subscript 1:j-1 doesn’t make sense for j = 1.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699399662335}, {"id": "FTl9scqE2B", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4090/Reviewer_dCYg"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper provides an efficient algorithm for min-entropy coupling that the authors call ARIMEC. The best previously known algorithm required one of the distributions to be  factorable into blocks with small supports, while their algorithm doesn't require that.", "review_text": "The paper provides an efficient algorithm for min-entropy coupling that the authors call ARIMEC. The best previously known algorithm required one of the distributions to be  factorable into blocks with small supports, while their algorithm doesn't require that.", "strengths": "The result is new and seems to be strong. It answers an open question from Sokota et al. (2022). The paper is easy to read, the main result is clear and its comparison with the state of the art is explained. The idea is nice and simple.", "weaknesses": "Even though I like the idea, the approach is not very sophisticated, it combines some standard (but non-trivial) combinatorial techniques with the results of prior works.\n\nAlso, even though the paper is easy to read, there are no rigorous formulations of the results in the main part of the paper, and some statements are not very precise (e.g. I find usage of terms like \"small\" in formal conditions not very nice).", "questions": "I do not have any specific questions, since the paper basically solves the problem as it was stated in Sokota et al. (2022). Maybe just a high-level question: Did you think of any potential future directions where your approach could be useful?\n\nSuggestions: As I said before, I would also write conditions in more formal manner (e.g. I recommend to replace \"small\" by something concrete and formal, and then add a high-level explanation below or above the formal definition). I also recommend to write your results as theorems (and keep current high-level explanations close to the formal statements).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper provides an efficient algorithm for min-entropy coupling that the authors call ARIMEC. The best previously known algorithm required one of the distributions to be  factorable into blocks with small supports, while their algorithm doesn't require that.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The result is new and seems to be strong. It answers an open question from Sokota et al. (2022). The paper is easy to read, the main result is clear and its comparison with the state of the art is explained. The idea is nice and simple.", "weaknesses": "Even though I like the idea, the approach is not very sophisticated, it combines some standard (but non-trivial) combinatorial techniques with the results of prior works.\n\nAlso, even though the paper is easy to read, there are no rigorous formulations of the results in the main part of the paper, and some statements are not very precise (e.g. I find usage of terms like \"small\" in formal conditions not very nice).", "questions": "I do not have any specific questions, since the paper basically solves the problem as it was stated in Sokota et al. (2022). Maybe just a high-level question: Did you think of any potential future directions where your approach could be useful?\n\nSuggestions: As I said before, I would also write conditions in more formal manner (e.g. I recommend to replace \"small\" by something concrete and formal, and then add a high-level explanation below or above the formal definition). I also recommend to write your results as theorems (and keep current high-level explanations close to the formal statements).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698766939105}, {"id": "RZ5MQ86XEk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4090/Reviewer_n51E"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper addresses the challenge of computing minimum-entropy couplings (MEC) for large-support distributions. \nMEC aims to find a joint distribution with the least joint entropy given two specific marginal distributions.\nExisting algorithms that find provable approximations for such couplings are unsuitable for very large-support distributions. \nCurrent heuristic methods, called Iterative Minimum-Entropy Coupling (IMEC), have limitations in handling these distributions.\n\nContributions:\n* Unified IMEC Algorithms: The authors provide a unified framework for IMEC algorithms using sets of partitions.\n* ARIMEC Introduction: Leveraging the unified view, a new method, ARIMEC, is introduced. It computes low-entropy couplings for any large-support distribution. Efficiency improvements, like lazy updates and entropy bounds, are incorporated.\n* Empirical Results: ARIMEC's effectiveness is showcased in Markov coding games and steganography, resulting in improved communication rates.\n\nThe paper presents a novel approach, ARIMEC, to compute low-entropy couplings for large-support distributions and validates its utility with real-world applications. \nThe authors promise to share their codebase with the community.", "review_text": "The paper addresses the challenge of computing minimum-entropy couplings (MEC) for large-support distributions. \nMEC aims to find a joint distribution with the least joint entropy given two specific marginal distributions.\nExisting algorithms that find provable approximations for such couplings are unsuitable for very large-support distributions. \nCurrent heuristic methods, called Iterative Minimum-Entropy Coupling (IMEC), have limitations in handling these distributions.\n\nContributions:\n* Unified IMEC Algorithms: The authors provide a unified framework for IMEC algorithms using sets of partitions.\n* ARIMEC Introduction: Leveraging the unified view, a new method, ARIMEC, is introduced. It computes low-entropy couplings for any large-support distribution. Efficiency improvements, like lazy updates and entropy bounds, are incorporated.\n* Empirical Results: ARIMEC's effectiveness is showcased in Markov coding games and steganography, resulting in improved communication rates.\n\nThe paper presents a novel approach, ARIMEC, to compute low-entropy couplings for large-support distributions and validates its utility with real-world applications. \nThe authors promise to share their codebase with the community.", "strengths": "As the author states, this paper has three contributions. Let's discuss the strengths of each in order.\n\n1. Unified IMEC Algorithms\nThe key strength of the unified IMEC framework, based on the description, seems to lie in its flexibility through the use of partitions. \nDepending on how these partitions are selected, different algorithms or strategies can be realized. \n\n2. ARIMEC Introduction\nThe proposed ARIMEC algorithm falls into the unified IMEC framework. \nIn the framework, irrespective of the specific autoregressive structure or the particularities of the marginal \\mu, \nthe algorithm can always satisfy the three conditions: Condition 3.1, 3.2 and 3.3. \nThis capability is significant because it allows ARIMEC to be applied to a wide range of problems or datasets \nthat have an autoregressive nature without the need for tweaking the algorithm for each specific case.\n\n3. Empirical Results\nilding on the work of Sokota et al. (2022), the authors explore an application of ARIMEC in the communication of messages \nthrough Markov decision processes. It's a unique application that hasn't been frequently touched upon in literature.", "weaknesses": "* The experiments seem to be heavily centered around specific domains like Markov coding games and steganography. \n  While these are valuable explorations, they might not give a complete picture of ARIMEC's versatility across other potential domains or applications.\n  I'd like a bit more discussion on the potential applications of ARIMEC.", "questions": "* In Section 3, a unified view of IMEC is proposed, but are there any existing algorithms to be unified other than TIMEC and FIMEC?\n* Figure 3, why do you compare Token-wise Error Rate?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of computing minimum-entropy couplings (MEC) for large-support distributions. \nMEC aims to find a joint distribution with the least joint entropy given two specific marginal distributions.\nExisting algorithms that find provable approximations for such couplings are unsuitable for very large-support distributions. \nCurrent heuristic methods, called Iterative Minimum-Entropy Coupling (IMEC), have limitations in handling these distributions.\n\nContributions:\n* Unified IMEC Algorithms: The authors provide a unified framework for IMEC algorithms using sets of partitions.\n* ARIMEC Introduction: Leveraging the unified view, a new method, ARIMEC, is introduced. It computes low-entropy couplings for any large-support distribution. Efficiency improvements, like lazy updates and entropy bounds, are incorporated.\n* Empirical Results: ARIMEC's effectiveness is showcased in Markov coding games and steganography, resulting in improved communication rates.\n\nThe paper presents a novel approach, ARIMEC, to compute low-entropy couplings for large-support distributions and validates its utility with real-world applications. \nThe authors promise to share their codebase with the community.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "As the author states, this paper has three contributions. Let's discuss the strengths of each in order.\n\n1. Unified IMEC Algorithms\nThe key strength of the unified IMEC framework, based on the description, seems to lie in its flexibility through the use of partitions. \nDepending on how these partitions are selected, different algorithms or strategies can be realized. \n\n2. ARIMEC Introduction\nThe proposed ARIMEC algorithm falls into the unified IMEC framework. \nIn the framework, irrespective of the specific autoregressive structure or the particularities of the marginal \\mu, \nthe algorithm can always satisfy the three conditions: Condition 3.1, 3.2 and 3.3. \nThis capability is significant because it allows ARIMEC to be applied to a wide range of problems or datasets \nthat have an autoregressive nature without the need for tweaking the algorithm for each specific case.\n\n3. Empirical Results\nilding on the work of Sokota et al. (2022), the authors explore an application of ARIMEC in the communication of messages \nthrough Markov decision processes. It's a unique application that hasn't been frequently touched upon in literature.", "weaknesses": "* The experiments seem to be heavily centered around specific domains like Markov coding games and steganography. \n  While these are valuable explorations, they might not give a complete picture of ARIMEC's versatility across other potential domains or applications.\n  I'd like a bit more discussion on the potential applications of ARIMEC.", "questions": "* In Section 3, a unified view of IMEC is proposed, but are there any existing algorithms to be unified other than TIMEC and FIMEC?\n* Figure 3, why do you compare Token-wise Error Rate?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698550852620}], "openreview_url": "https://openreview.net/forum?id=Tpk0p9QBM6", "arxiv_id": "2405.19540", "paper_pdf": "papers/Tpk0p9QBM6.pdf", "paper_pdf_sha256": "a15fd57eae9b2344d5edfa50c2b509f6944ad48244d0b95257eb9a3d4c44f48e", "paper_pdf_bytes": 1153300, "paper_pdf_source": "openreview", "code_url": "https://github.com/ssokota/mec", "code_repository": "ssokota/mec", "code_commit": "e56ad6619498914fe289d91516febb1fa3be050f", "code_archive": "repos/Tpk0p9QBM6.zip", "code_archive_sha256": "b9724a7ec296565ffe0ac486f6492d3ca314b52e8313366e650e954d7edc735e", "code_archive_bytes": 43750, "code_file_count": 27, "code_extensions": {".py": 25, ".rs": 2}, "github_disk_usage_kb": 50, "github_languages": {"Python": 86017, "Rust": 13793}, "github_archived": false, "github_pushed_at": "2026-01-06T05:19:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/computing-low-entropy-couplings-for-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7qSpaOSbRVO", "year": 2023, "status": "rejected", "title": "Data Poisoning Attacks Against Multimodal Encoders", "authors": ["Ziqing Yang", "Xinlei He", "Zheng Li", "Michael Backes", "Mathias Humbert", "Pascal Berrang", "Yang Zhang"], "authorids": ["~Ziqing_Yang3", "~Xinlei_He1", "~Zheng_Li17", "~Michael_Backes1", "~Mathias_Humbert2", "~Pascal_Berrang1", "~Yang_Zhang15"], "authors_source": "OpenReview API", "abstract": "Traditional machine learning (ML) models, e.g., image classifiers, usually rely on large-scale labeled datasets to achieve strong performance. However, such labeled datasets are often challenging and expensive to obtain. Also, the predefined categories limit the model's ability to generalize to other visual concepts as additional labeled data is required. On the contrary, the newly emerged multimodal model, which contains both visual and linguistic modalities, learns the concept of images from the raw text. It is a promising way to solve the above problems as it can use easy-to-collect image-text pairs to construct the training dataset and the raw texts contain almost unlimited categories according to their semantics. However, learning from a  large-scale unlabeled dataset also exposes the model to the risk of potential poisoning attacks, whereby the adversary aims to perturb the model's training dataset to trigger malicious behaviors in it. Previous work mainly focuses on the visual modality. In this paper, we instead focus on answering two questions: (1) Is the linguistic modality also vulnerable to poisoning attacks? and (2) Which modality is most vulnerable? To answer the two questions, we conduct three types of poisoning attacks against CLIP, the most representative multimodal contrastive learning framework. Extensive evaluations on different datasets and model architectures show that all three attacks can perform well on the linguistic modality with only a relatively low poisoning rate and limited epochs. Also, we observe that the poisoning effect differs between different modalities, i.e., with lower MinRank in the visual modality and with higher Hit@K when K is small in the linguistic modality. To mitigate the attacks, we propose both pre-training and post-training defenses. We empirically show that both defenses can significantly reduce the attack performance while preserving the model's utility.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "TQQTEPrCEp", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1976/Reviewer_KeYj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper examines data poisoning attacks and defense for joint language-vision model (CLIP) in a retrieval setting. Expanding on Carlini et al. 2022, the authors propose attacks and defenses on both modalities instead of just vision signal and show vulnerability in both.\n\nThe paper proposes three attacks, all try to lead the model to mis-associate the text signal to the image signal. The first attack targets a class of text, swapping out the image from an <image, text> pair for another image of a different class. The second attack generalizes the target of the first into a class of image, and the third attack generalize the second attack into multiple classes of images. All three attacks work well during the fine-tuning process on COCO and Flickr-PASCAL even with very low poisoning rate and with virtually no loss in term of original test data utility. \n\nThe authors then analyze the difference in the effect of data poisoning on the text/image encoders by freezing each and calculate the performance (Hit and Minimum rank) and show that the image and text encoder lead to slightly different forms of poisoned behavior. While the poisoned image encoder generally leads to worse behavior overall (generally higher rank of the poisoned image in retrieval), the poisoned text encoder leads to higher probability of top retrieval results being the poisoned class.\n\nIn the defense, the author proposes two methods, the first is data filtering with distance from the original encoders and the second is retraining with clean dataset. Both method improves the robustness substantially.", "review_text": "My main concern is about the potential impact of the paper. The previous work of Carlini has shown the vulnerability in CLIP model while the attack and defense method is not novel enough.", "strengths": "Strength:\n+ The paper is one of the first works to explore data security in both modalities of a joint vision-language model\n+ Both the attacks and defense methods are simple but effective, with strong experimental results.\n\nWeakness:\n+ There has been numerous works on adversarial attacks on text models, many of them also about data poisoning, so it has been quite clear that linguistic modality is also vulnerable to data poison attack [1,2]\n+ As Carlini et al. and others have shown in their previous work, CLIP is quite vulnerable to adversarial attacks, and they have also shown that CLIP can be poisoned with a very limited number of poisoned data points. \n+ The attacks and defense method proposed in the paper is different compared to previous work of Carlini, but the novelty is limited.\n\n[1] https://aclanthology.org/2021.naacl-main.13.pdf\n[2] https://aclanthology.org/2021.naacl-main.13.pdf", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper examines data poisoning attacks and defense for joint language-vision model (CLIP) in a retrieval setting. Expanding on Carlini et al. 2022, the authors propose attacks and defenses on both modalities instead of just vision signal and show vulnerability in both.\n\nThe paper proposes three attacks, all try to lead the model to mis-associate the text signal to the image signal. The first attack targets a class of text, swapping out the image from an <image, text> pair for another image of a different class. The second attack generalizes the target of the first into a class of image, and the third attack generalize the second attack into multiple classes of images. All three attacks work well during the fine-tuning process on COCO and Flickr-PASCAL even with very low poisoning rate and with virtually no loss in term of original test data utility. \n\nThe authors then analyze the difference in the effect of data poisoning on the text/image encoders by freezing each and calculate the performance (Hit and Minimum rank) and show that the image and text encoder lead to slightly different forms of poisoned behavior. While the poisoned image encoder generally leads to worse behavior overall (generally higher rank of the poisoned image in retrieval), the poisoned text encoder leads to higher probability of top retrieval results being the poisoned class.\n\nIn the defense, the author proposes two methods, the first is data filtering with distance from the original encoders and the second is retraining with clean dataset. Both method improves the robustness substantially.", "strength_and_weaknesses": "Strength:\n+ The paper is one of the first works to explore data security in both modalities of a joint vision-language model\n+ Both the attacks and defense methods are simple but effective, with strong experimental results.\n\nWeakness:\n+ There has been numerous works on adversarial attacks on text models, many of them also about data poisoning, so it has been quite clear that linguistic modality is also vulnerable to data poison attack [1,2]\n+ As Carlini et al. and others have shown in their previous work, CLIP is quite vulnerable to adversarial attacks, and they have also shown that CLIP can be poisoned with a very limited number of poisoned data points. \n+ The attacks and defense method proposed in the paper is different compared to previous work of Carlini, but the novelty is limited.\n\n[1] https://aclanthology.org/2021.naacl-main.13.pdf\n[2] https://aclanthology.org/2021.naacl-main.13.pdf", "clarity,_quality,_novelty_and_reproducibility": "To my knowledge, this is one of the first work to explore attack/defense from both modalities in joint vision-language model. However, the attack and defense methods are simple and not entirely novel. Previous works have also shown vulnerability of CLIP and text model in data poisoning attacks, which also reduce the novelty and potential impact of this work.", "summary_of_the_review": "My main concern is about the potential impact of the paper. The previous work of Carlini has shown the vulnerability in CLIP model while the attack and defense method is not novel enough.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666699323698}, {"id": "zKqKA-CRE5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1976/Reviewer_zCXF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper focuses on the data poisoning task on multimodal encoders. This paper investigates three types of poisoning attacks first. After that, it studies the effectiveness of attacking visual and linguistic features. In addition, it explores two types of defense mechanisms for defending against the attack on multimodal encoders.", "review_text": "In total, this paper is novel for providing a study on attacking linguistic modality for multimodal encoders. The experiment is comprehensive and provides an easy-to-follow structure for readers. However, this paper lacks an illustration of why linguistic modality matters for poisoning attacking and whether the used attack method is better than the baselines. Thus, the reasonableness for conducting this type of attack needs more explanation.", "strengths": "Strength:\n1. This paper investigates a new problem about whether attacking the linguistic modality is also effective for the poisoning attack task. \n2. The experiments are extensive and provide a deeper understanding of the vulnerability of multimodal encoders.\n\nWeaknesses: \n1. There is a lack of baselines in this paper. Although the paper proposes three types of attack, the comparison does not include other state-of-the-art methods. As mentioned by the author, there are existing methods in this domain, but they are not introduced in the experiment to show whether the attacking method is not redundant. \n2. There is a lack of discussion on whether it’s worth of attacking linguistic modality. Although it can be as effective as attacking linguistic modality, it is not clear whether it changes a lot of linguistic data compared to visual data. If not, based on the principle of not causing much difference in the data, it does not make sense to change the linguistic data.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper focuses on the data poisoning task on multimodal encoders. This paper investigates three types of poisoning attacks first. After that, it studies the effectiveness of attacking visual and linguistic features. In addition, it explores two types of defense mechanisms for defending against the attack on multimodal encoders.", "strength_and_weaknesses": "Strength:\n1. This paper investigates a new problem about whether attacking the linguistic modality is also effective for the poisoning attack task. \n2. The experiments are extensive and provide a deeper understanding of the vulnerability of multimodal encoders.\n\nWeaknesses: \n1. There is a lack of baselines in this paper. Although the paper proposes three types of attack, the comparison does not include other state-of-the-art methods. As mentioned by the author, there are existing methods in this domain, but they are not introduced in the experiment to show whether the attacking method is not redundant. \n2. There is a lack of discussion on whether it’s worth of attacking linguistic modality. Although it can be as effective as attacking linguistic modality, it is not clear whether it changes a lot of linguistic data compared to visual data. If not, based on the principle of not causing much difference in the data, it does not make sense to change the linguistic data.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written and easy to follow. As for its novelty, it is new to see work on the poisoning attack domain to conduct a study on the linguistic modality. As for reproducibility, although the authors provide a reproducibility statement, it is still better to provide codes for other readers to implement the experiments.", "summary_of_the_review": "In total, this paper is novel for providing a study on attacking linguistic modality for multimodal encoders. The experiment is comprehensive and provides an easy-to-follow structure for readers. However, this paper lacks an illustration of why linguistic modality matters for poisoning attacking and whether the used attack method is better than the baselines. Thus, the reasonableness for conducting this type of attack needs more explanation.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666661401503}, {"id": "Nvg5V5sq6D", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1976/Reviewer_ygKY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper explores the poisoning attacks on CLIP. Compared to the previous work that focused on pre-training attacks on the image-modal encoder, this paper focuses on attacking the fine-tuning process and evaluating it using text-image retrieval, a downstream task that shows the attack's effect on both image and text-modal encoder. The attack is also generalized to a class-class poison rather than a target-class poison. The paper analyzes the differences between poisoning effects on both modalities and proposes post- and pre-training defenses against these attacks. ", "review_text": "The paper gives many new insights into analyzing attacks on multi-modal models. The training scheme on fine-tuning tasks makes future experiments more manageable and easier to transfer. However, I think there is a lack of explanation for many of the findings. I will raise my rating if there are more explanations on the points I mentioned. ", "strengths": "Strength:\n1. To my knowledge, this is the first paper that tries poisoning attacks on a multimodal model during the fine-tuning phase. The scenario is realistic and easier to expand further for future experiments compared to training CLIP from the scratch. Using Image-Text retrieval is also a novel evaluation task for analyzing multi-modal attacks with more evaluation metrics like hit@k and minrank. \n2. This paper isolates the linguistic and visual modalities and analyzes the attack's effect on them separately, providing an interesting insight that may lead to effective attacks with more focusing goals (lower minrank or higher hit@k rate).\n3. The paper gives simple but effective defenses against the proposed attacks.\n4. The paper considers a more general attack scheme compared to the previous works.\n\nWeakness:\n1. I do not see significant differences between attack 2 and attack 3. both datasets seem to have consistent Hit@K ratios and MinRank values. The only values inconsistent are \"ours-2\" under the Flickr-PASCAL dataset, but I think it can be attributed to the specific dataset selections rather than the attack 3 itself being more harmful. There is an analysis in the ablation study on different combinations of classes during the attacks, but the consistency issue is not analyzed fully during the main experiment part.\n\n2. I also cast a doubt on the dataset used, specifically Flickr-PASCAL. Even though the combined dataset has a relatively large size, only 20 categories from the PASCAL datasets could be selected to be poisoned and that may influence the scope of the attack. Namely, even though the overall poisoning rate is low, the effect may be stronger due to the data imbalance issue. \n\n3. Combing the freezing experiment with the main experimental result, I can have a clear understanding of the linguistic modality being attacked. However, the analysis of which modality is most vulnerable is not satisfactory. From the experiment, each image is matched to more than one caption, rendering an imbalance in encoding representations. As stated in previous work, when there are more diverse caption sets for each image, the model is more likely to change the image encoder compared to the text encoder. Although the statement is unjustified theoretically, It makes sense intuitively and a more considerate experiment concerning this issue or an explanation of why this concern is unnecessary could be very helpful to the overall argument. There is also a lack of explanation and extension on their result with the difference in attacking different modalities. \n\n4. Both the pre-training defense (removing less correlated data pairs) and the post-training defense sound a bit weak. For the pre-training, it is possible to have harder training tasks with larger cosine distances, and setting a threshold is difficult to do and could be harmful to the performances. For the post-training, the paper does not explore the situation when the fine-tuning datasets are of different scopes.\n\n5. Overall, I think there is a lack of analysis and expansion on the experiments discussed", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper explores the poisoning attacks on CLIP. Compared to the previous work that focused on pre-training attacks on the image-modal encoder, this paper focuses on attacking the fine-tuning process and evaluating it using text-image retrieval, a downstream task that shows the attack's effect on both image and text-modal encoder. The attack is also generalized to a class-class poison rather than a target-class poison. The paper analyzes the differences between poisoning effects on both modalities and proposes post- and pre-training defenses against these attacks. ", "strength_and_weaknesses": "Strength:\n1. To my knowledge, this is the first paper that tries poisoning attacks on a multimodal model during the fine-tuning phase. The scenario is realistic and easier to expand further for future experiments compared to training CLIP from the scratch. Using Image-Text retrieval is also a novel evaluation task for analyzing multi-modal attacks with more evaluation metrics like hit@k and minrank. \n2. This paper isolates the linguistic and visual modalities and analyzes the attack's effect on them separately, providing an interesting insight that may lead to effective attacks with more focusing goals (lower minrank or higher hit@k rate).\n3. The paper gives simple but effective defenses against the proposed attacks.\n4. The paper considers a more general attack scheme compared to the previous works.\n\nWeakness:\n1. I do not see significant differences between attack 2 and attack 3. both datasets seem to have consistent Hit@K ratios and MinRank values. The only values inconsistent are \"ours-2\" under the Flickr-PASCAL dataset, but I think it can be attributed to the specific dataset selections rather than the attack 3 itself being more harmful. There is an analysis in the ablation study on different combinations of classes during the attacks, but the consistency issue is not analyzed fully during the main experiment part.\n\n2. I also cast a doubt on the dataset used, specifically Flickr-PASCAL. Even though the combined dataset has a relatively large size, only 20 categories from the PASCAL datasets could be selected to be poisoned and that may influence the scope of the attack. Namely, even though the overall poisoning rate is low, the effect may be stronger due to the data imbalance issue. \n\n3. Combing the freezing experiment with the main experimental result, I can have a clear understanding of the linguistic modality being attacked. However, the analysis of which modality is most vulnerable is not satisfactory. From the experiment, each image is matched to more than one caption, rendering an imbalance in encoding representations. As stated in previous work, when there are more diverse caption sets for each image, the model is more likely to change the image encoder compared to the text encoder. Although the statement is unjustified theoretically, It makes sense intuitively and a more considerate experiment concerning this issue or an explanation of why this concern is unnecessary could be very helpful to the overall argument. There is also a lack of explanation and extension on their result with the difference in attacking different modalities. \n\n4. Both the pre-training defense (removing less correlated data pairs) and the post-training defense sound a bit weak. For the pre-training, it is possible to have harder training tasks with larger cosine distances, and setting a threshold is difficult to do and could be harmful to the performances. For the post-training, the paper does not explore the situation when the fine-tuning datasets are of different scopes.\n\n5. Overall, I think there is a lack of analysis and expansion on the experiments discussed", "clarity,_quality,_novelty_and_reproducibility": "This work is easy to follow and explains its reasoning in a clear manner. The focus is interesting and the experiments are very thorough. ", "summary_of_the_review": "The paper gives many new insights into analyzing attacks on multi-modal models. The training scheme on fine-tuning tasks makes future experiments more manageable and easier to transfer. However, I think there is a lack of explanation for many of the findings. I will raise my rating if there are more explanations on the points I mentioned. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666653429300}], "openreview_url": "https://openreview.net/forum?id=7qSpaOSbRVO", "arxiv_id": "2209.15266", "paper_pdf": "papers/7qSpaOSbRVO.pdf", "paper_pdf_sha256": "614d4150779cd439bf5c9e517e4a48a35639143c870b6cc25bc541d988d4592a", "paper_pdf_bytes": 985510, "paper_pdf_source": "openreview", "code_url": "https://github.com/zqypku/mm_poison", "code_repository": "zqypku/mm_poison", "code_commit": "b11a503035db543cccd944c4b01efc6f9dcc5d11", "code_archive": "repos/7qSpaOSbRVO.zip", "code_archive_sha256": "a94727d488eab2fe75273783e10381eaf8e3187cd4f0f53fd49d7d8cb9ddd077", "code_archive_bytes": 86713, "code_file_count": 35, "code_extensions": {".py": 35}, "github_disk_usage_kb": 83, "github_languages": {"Python": 308268}, "github_archived": false, "github_pushed_at": "2023-10-25T12:28:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/data-poisoning-attacks-against-multimodal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6PTUd_zPdHL", "year": 2022, "status": "rejected", "title": "Differentiable Top-k Classification Learning", "authors": ["Felix Petersen", "Hilde Kuehne", "Christian Borgelt", "Oliver Deussen"], "authorids": ["~Felix_Petersen1", "~Hilde_Kuehne5", "~Christian_Borgelt1", "~Oliver_Deussen1"], "authors_source": "OpenReview API", "abstract": "The top-k classification accuracy is one of the core metrics in machine learning. Here, k is conventionally a positive integer, such as 1 or 5. In this work, we relax this assumption and propose to draw k from a probability distribution for training. Combining this with recent advances in differentiable sorting and ranking, we propose a new family of differentiable top-k cross-entropy classification losses. We find that relaxing k does not only produce better top-5 accuracies, but also makes models more robust, which leads to top-1 accuracy improvements. When fine-tuning publicly available ImageNet models, we achieve a new state-of-the-art on ImageNet for publicly available models with an 88.36% top-1 and a 98.71% top-5 accuracy. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "XY-jt234yxO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2716/Reviewer_575y"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a loss to relax the assumption of using a fixed k for top-k classification learning. The authors use the existing differentiable sorting and ranking operators. Experimental results also achieve a state-of-the-art on ImageNet.", "review_text": "Strengths: \n\tThe motivation of this paper is clear to draw k from a probability distribution for training. \n\tThe idea of this paper is pretty novel and exciting which makes the classification model robust.\n\tThe extensive experiments conducted on five data sets are sufficient to show the advantages of the proposed idea\nWeaknesses:\n\tThe details of the differentiable sorting networks is not represented. How to rank the predicted scores of the final classification layer and get the probability distribution?\n\tIn figure 1, the first row (rank1) are multiplied by 1 and the second row(rank2) are multiplied by 0.5. Please explain the reason.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a loss to relax the assumption of using a fixed k for top-k classification learning. The authors use the existing differentiable sorting and ranking operators. Experimental results also achieve a state-of-the-art on ImageNet.", "main_review": "Strengths: \n\tThe motivation of this paper is clear to draw k from a probability distribution for training. \n\tThe idea of this paper is pretty novel and exciting which makes the classification model robust.\n\tThe extensive experiments conducted on five data sets are sufficient to show the advantages of the proposed idea\nWeaknesses:\n\tThe details of the differentiable sorting networks is not represented. How to rank the predicted scores of the final classification layer and get the probability distribution?\n\tIn figure 1, the first row (rank1) are multiplied by 1 and the second row(rank2) are multiplied by 0.5. Please explain the reason.\n", "summary_of_the_review": "This paper derives a family of top-k cross entropy losses which is a novel practice. The experimental analysis on ImageNet including the impact of the distribution and ranking set size m, etc, is concrete and sufficient.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635936318311}, {"id": "gknz7esELlT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2716/Reviewer_Abaq"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper addresses Top-k classification learning. Based on the recent progress on differentiable sorting and ranking, the author proposes a loss function for top-k classification where the k is not fixed but follows a given probability distribution. To improve the efficiency, a splitter selection network is proposed so that fewer layers are required for the sorting network. \n\nThe proposed loss function can be combined with different sorting methods. In experiments, the loss function is shown to be effective in training a model from scratch on Cifar10. It can also be used in fine-tuning on ImageNet dataset and has performance gain. ", "review_text": "strengths\n1. The idea of using different probability distributions for k is interesting. The results also demonstrate the effectiveness of this idea.\n2. The experiments of incorporating different sorting methods are comprehensive.\n\n\nweaknesses\n1. In my opinion, the P_K is more like a set of weights, rather than a probability distribution. If it is the case, I recommend improving the descriptions to reduce confusion.\n2. It would be nice to present an experiment with conditional probability distributions for k of different classes based on their semantic meaning (like person, animal). I think it is also a significant contribution of this paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper addresses Top-k classification learning. Based on the recent progress on differentiable sorting and ranking, the author proposes a loss function for top-k classification where the k is not fixed but follows a given probability distribution. To improve the efficiency, a splitter selection network is proposed so that fewer layers are required for the sorting network. \n\nThe proposed loss function can be combined with different sorting methods. In experiments, the loss function is shown to be effective in training a model from scratch on Cifar10. It can also be used in fine-tuning on ImageNet dataset and has performance gain. ", "main_review": "strengths\n1. The idea of using different probability distributions for k is interesting. The results also demonstrate the effectiveness of this idea.\n2. The experiments of incorporating different sorting methods are comprehensive.\n\n\nweaknesses\n1. In my opinion, the P_K is more like a set of weights, rather than a probability distribution. If it is the case, I recommend improving the descriptions to reduce confusion.\n2. It would be nice to present an experiment with conditional probability distributions for k of different classes based on their semantic meaning (like person, animal). I think it is also a significant contribution of this paper.\n", "summary_of_the_review": "This paper proposes a flexible loss function for top-k classification, providing useful insights for image classification. So it is worth of reading for the researchers in this area.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635896267176}, {"id": "rOwo0LGbjx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2716/Reviewer_RKLW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a method to employ the benefits of differential sorting methods towards top-k classification learning. The presents several experiments with sampling weights for different ranks and presents the results. The loss is used for fine-tuning in most experiments (apart from the CIFAR100 case). The method seems to give minor improvements on the ResNeXt-101 32*48d baseline. ", "review_text": "Strengths:\n- The efforts on optimizing the top-k classification learning through differentiable sorting appear novel to me. The discussion on differential sorting is comprehensive. The paper specifically discusses each of the options and how it is optimized for the studied scenario. \n- Experiments are thorough \n- The work is also interesting because the performance gains come only because of fine-tuning.\n\n\nWeaknesses\n\n- The second term and eqn 2 would be constant with k=5 (if only the top five rows of the P matrix are constructed). Expanding eqn2 for the given example in Fig1, the loss would be: -log(0.5 *.03 + 0.5 (0.3+0.6)), assuming Panda is the ground truth class. Consider a case of P_K [0.5 0 0 0 0.5], the equation would be -log (0.5 top1 + 0.5 (top1+top2+top3+top4+top5)). If only five columns are reconstructed and if they are column stochastic, then the sum of top1 to top5 would always be 1. Then the second term will always give a constant value. Requesting the authors to clarify this aspect.\n\n- At first, it appears that the distribution would be a sample. However, fixed distribution is used for a set of experiments. For example, it is either [0.5 0 0 0 0.5] or [0.2 0.2 0.2 0.2 0.2] for the entire experiment. Hence, presenting it as \"sampled\" is confusing. The best results come when you have the top1 and the sum of the top five values. Hence, the initial discussion and intuition can be improved a bit.\n\n- The improvements on Noisy Student EfficientNet-L2 are negligible. 88.35 to 88.36 is certainly not statistically significant. Were experiments for Table1 were also ran 10 times (like table 2?).\n\n- Please mention the number of rows that were reconstructed for each experiment. The number of columns (m) is mentioned in the experiments but not the number of rows.\n\n- Berrada et al. was used to train the model from scratch. It would be worth comparing their loss for fine-tuning purposes as well. I think that would be a fairer comparison. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method to employ the benefits of differential sorting methods towards top-k classification learning. The presents several experiments with sampling weights for different ranks and presents the results. The loss is used for fine-tuning in most experiments (apart from the CIFAR100 case). The method seems to give minor improvements on the ResNeXt-101 32*48d baseline. ", "main_review": "Strengths:\n- The efforts on optimizing the top-k classification learning through differentiable sorting appear novel to me. The discussion on differential sorting is comprehensive. The paper specifically discusses each of the options and how it is optimized for the studied scenario. \n- Experiments are thorough \n- The work is also interesting because the performance gains come only because of fine-tuning.\n\n\nWeaknesses\n\n- The second term and eqn 2 would be constant with k=5 (if only the top five rows of the P matrix are constructed). Expanding eqn2 for the given example in Fig1, the loss would be: -log(0.5 *.03 + 0.5 (0.3+0.6)), assuming Panda is the ground truth class. Consider a case of P_K [0.5 0 0 0 0.5], the equation would be -log (0.5 top1 + 0.5 (top1+top2+top3+top4+top5)). If only five columns are reconstructed and if they are column stochastic, then the sum of top1 to top5 would always be 1. Then the second term will always give a constant value. Requesting the authors to clarify this aspect.\n\n- At first, it appears that the distribution would be a sample. However, fixed distribution is used for a set of experiments. For example, it is either [0.5 0 0 0 0.5] or [0.2 0.2 0.2 0.2 0.2] for the entire experiment. Hence, presenting it as \"sampled\" is confusing. The best results come when you have the top1 and the sum of the top five values. Hence, the initial discussion and intuition can be improved a bit.\n\n- The improvements on Noisy Student EfficientNet-L2 are negligible. 88.35 to 88.36 is certainly not statistically significant. Were experiments for Table1 were also ran 10 times (like table 2?).\n\n- Please mention the number of rows that were reconstructed for each experiment. The number of columns (m) is mentioned in the experiments but not the number of rows.\n\n- Berrada et al. was used to train the model from scratch. It would be worth comparing their loss for fine-tuning purposes as well. I think that would be a fairer comparison. \n\n", "summary_of_the_review": " Although the paper brings several novel perspectives, there remain several ambiguities as well. Some additional experiments, clarifications can also strengthen the draft. Overall, in the current form, the paper is a borderline one and the final decision will depend a lot on the discussion during the rebuttal phase. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635861422676}, {"id": "BI03bTZqSEF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2716/Reviewer_FWmd"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a differentiable loss for top-K classification based on differentiable sorting networks, i.e. sorting neural networks in which basic min/max operations are replaced by smoothed versions (i.e. softmax/softmin). The main principle is to use the sorting network to estimate the probability of the rank of each class and then filter only the top-k. An extension consists in considering that k can take several possible values at random (e.g. 50% chance of being 1 and  50% chance of being 5). \n\nThe resulting loss is experimented on three datasets (CIFAR-100, ImageNet-1K and ImageNet-21K-P) and with three existing sorting networks. Performances are mainly compared to cross-entropy showing low improvements. ", "review_text": "Strengths\n- Set-valued classification is an important topic to cope with class ambiguity. Few works (only one as far as I know [1]) have proposed a top-k loss for neural networks and there is room for improvements \n- The proposed approach is different from [1] as it relies on sorting networks to determine the set of the most likely classes rather than a purely top-k objective\n\nWeaknesses\n- A first weakness is that the contribution is quite incremental and not well justified from a theoretical point of view. Using sorting networks for top-k is an acceptable strategy from a practical point of view but a bit over-kill and not very new from a theoretical point of view. The proposal to use several values of K is also not really justified. The principle of top-K is to predict sets of fixed size contrary to other set-valued classification approaches that attempt to solve other objectives (e.g. adaptive set sizes but equal to K on average). We let the authors refer to [2] for a clear overview of the different objectives. Here, the objective is not really clear. If K is supposed to be a random variable (e.g. 50% chance of being 1 and  50% chance of being 5), that means that for the same image x, the classifier is supposed to return randomly either one class or 5 classes without any consideration with regard to the image content itself. \n- another main weakness is that no significant improvement of the proposed loss over cross-entropy is shown. The reported top-K accuracy gains are not systematic and so low that they may be not statistically significant. As a first step towards a better understanding of the results, the authors should first compute some significance tests (e.g. p-values on several runs and a clear cross-validation procedure for model selection among epochs). But even so, it won’t resolve the fact that the performance gain is observed only for some specific configurations (e.g. a specific sorting network and specific values of K probabilities) and remains very low even in such advantageous conditions. \n\n[1] Berrada, L., Zisserman, A., & Kumar, M. P. (2018). Smooth loss functions for deep top-k classification. arXiv preprint arXiv:1802.07595.\n[2] Chzhen, E., Denis, C., Hebiri, M., & Lorieul, T. (2021). Set-valued classification--overview via a unified framework. arXiv preprint arXiv:2102.12318.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a differentiable loss for top-K classification based on differentiable sorting networks, i.e. sorting neural networks in which basic min/max operations are replaced by smoothed versions (i.e. softmax/softmin). The main principle is to use the sorting network to estimate the probability of the rank of each class and then filter only the top-k. An extension consists in considering that k can take several possible values at random (e.g. 50% chance of being 1 and  50% chance of being 5). \n\nThe resulting loss is experimented on three datasets (CIFAR-100, ImageNet-1K and ImageNet-21K-P) and with three existing sorting networks. Performances are mainly compared to cross-entropy showing low improvements. ", "main_review": "Strengths\n- Set-valued classification is an important topic to cope with class ambiguity. Few works (only one as far as I know [1]) have proposed a top-k loss for neural networks and there is room for improvements \n- The proposed approach is different from [1] as it relies on sorting networks to determine the set of the most likely classes rather than a purely top-k objective\n\nWeaknesses\n- A first weakness is that the contribution is quite incremental and not well justified from a theoretical point of view. Using sorting networks for top-k is an acceptable strategy from a practical point of view but a bit over-kill and not very new from a theoretical point of view. The proposal to use several values of K is also not really justified. The principle of top-K is to predict sets of fixed size contrary to other set-valued classification approaches that attempt to solve other objectives (e.g. adaptive set sizes but equal to K on average). We let the authors refer to [2] for a clear overview of the different objectives. Here, the objective is not really clear. If K is supposed to be a random variable (e.g. 50% chance of being 1 and  50% chance of being 5), that means that for the same image x, the classifier is supposed to return randomly either one class or 5 classes without any consideration with regard to the image content itself. \n- another main weakness is that no significant improvement of the proposed loss over cross-entropy is shown. The reported top-K accuracy gains are not systematic and so low that they may be not statistically significant. As a first step towards a better understanding of the results, the authors should first compute some significance tests (e.g. p-values on several runs and a clear cross-validation procedure for model selection among epochs). But even so, it won’t resolve the fact that the performance gain is observed only for some specific configurations (e.g. a specific sorting network and specific values of K probabilities) and remains very low even in such advantageous conditions. \n\n[1] Berrada, L., Zisserman, A., & Kumar, M. P. (2018). Smooth loss functions for deep top-k classification. arXiv preprint arXiv:1802.07595.\n[2] Chzhen, E., Denis, C., Hebiri, M., & Lorieul, T. (2021). Set-valued classification--overview via a unified framework. arXiv preprint arXiv:2102.12318.\n", "summary_of_the_review": "An interesting attempt to improve the top-K classification but consistent limitations:  \n(I) an incremental contribution and no clear justification of considering k as a random variable\n(ii) no significant improvement of the proposed loss over cross-entropy\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635265708301}], "openreview_url": "https://openreview.net/forum?id=6PTUd_zPdHL", "arxiv_id": "2206.07290", "paper_pdf": "papers/6PTUd_zPdHL.pdf", "paper_pdf_sha256": "be2738bba9d8af9e65d0d7779cb95a2e42e7ea57b42ad9945900996fe99d627f", "paper_pdf_bytes": 548857, "paper_pdf_source": "openreview", "code_url": "https://github.com/Felix-Petersen/difftopk", "code_repository": "Felix-Petersen/difftopk", "code_commit": "76ef96db648058a73571628f1db5e6a9f4478bfd", "code_archive": "repos/6PTUd_zPdHL.zip", "code_archive_sha256": "70e8ef6e5492eea07f2be84ca12a5f9fdfc6806f83933992b3ae81314fe73d44", "code_archive_bytes": 76608, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 80, "github_languages": {"Python": 80416}, "github_archived": false, "github_pushed_at": "2023-01-03T11:59:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/differentiable-top-k-classification-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3Wp8HM2CNdR", "year": 2021, "status": "rejected", "title": "Whitening for Self-Supervised Representation Learning", "authors": ["Aleksandr Ermolov", "Aliaksandr Siarohin", "Enver Sangineto", "Nicu Sebe"], "authorids": ["~Aleksandr_Ermolov1", "~Aliaksandr_Siarohin1", "~Enver_Sangineto1", "~Nicu_Sebe1"], "authors_source": "OpenReview API", "abstract": "Most of the self-supervised representation learning methods are based on the contrastive loss and the instance-discrimination task, where augmented versions of the same image instance  (\"positives\") are contrasted with instances extracted from other images (\"negatives\"). For the learning to be effective, a lot of negatives should be compared with a positive pair, which is computationally demanding. In this paper, we propose a different direction and a new loss function for self-supervised representation learning which is based on the whitening of the latent-space features. The whitening operation has a \"scattering\" effect on the batch samples, which compensates the use of negatives, avoiding degenerate solutions where all the sample representations collapse to a single point. Our Whitening MSE (W-MSE) loss does not require special heuristics (e.g. additional networks) and it is conceptually simple. Since negatives are not needed, we can extract multiple positive pairs from the same image instance. We  empirically show that W-MSE is competitive with respect to popular, more complex self-supervised methods. The source code of the method and all the experiments is included in the Supplementary Material.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "svow4IP1Cl0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1203/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes the mean square error loss with whitening operation to project positive pairs closely to each others while projecting the different positive pairs far away from each other on a unit sphere. This way, similar to BYOL, this paper removes the construction of negative pairs while improving the MoCo-V2 slightly on not very challenging benchmarks. The authors can find my questions/comments in the list below.\n\n1. What does gray and blue squares represent in figure 3? I believe figure 3 can be revisited for the sake of better understanding the method. For example, I was not able to understand why v_i and v_i+8 represent the same image. In this current shape, it creates confusion rather than helping for clarity.\n\n2. In the Batch Slicing section, the authors mention that the size each of sub-batch should be close to the size of the embedding x 2. Does this make sense? Is the size of embedding 512? If yes, do you have 1024 samples in a sub-batch? I would be happy if the authors can comment on this.\n\n3. I think my biggest concern with the paper is the lack of extensive experiments. All the experiments are done on small-scale datasets which is highly questionable when they are used for unsupervised learning. It would be much more convincing to have experiments on ImageNet which is a standard experiment for unsupervised learning.\n\n4. The improvement on relatively less challenging benchmarks are very marginal. Even in some cases, CIFAR100, contrastive loss does better than the proposed one. And the proposed method only outperforms BYOL in CIFAR10.\n\n5. My impression is that the contrastive method in table 1 and 2 represent the MoCo-v2. I would replace it with MoCo-v2 in these tables as there are many other methods that uses contrastive loss functions.\n\n6. Another problem with the experiments is that, they lack experiments on different downstream tasks such as object detection. \n\n7. The authors try the Euclidean and MSE distance to project positives closely. Did they consider the cosine similarity loss? I did not understand the point of using Euclidean distance experiments without normalization? What do they exactly prove? And we can always normalize the embeddings as it is done in the other methods. I would be happy to receive some comments on this from the authors.\n\n8. Finally, it would be nice to the advantages of the proposed method (removal of negatives) in terms of training complexity. Does it, as expected, reduce the complexity in the training time? If yes, quantifying it would increase the strength of the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A well-established method for unsupervised learning with contrastive loss functions with not very impressive results.                                            ", "review": "This paper proposes the mean square error loss with whitening operation to project positive pairs closely to each others while projecting the different positive pairs far away from each other on a unit sphere. This way, similar to BYOL, this paper removes the construction of negative pairs while improving the MoCo-V2 slightly on not very challenging benchmarks. The authors can find my questions/comments in the list below.\n\n1. What does gray and blue squares represent in figure 3? I believe figure 3 can be revisited for the sake of better understanding the method. For example, I was not able to understand why v_i and v_i+8 represent the same image. In this current shape, it creates confusion rather than helping for clarity.\n\n2. In the Batch Slicing section, the authors mention that the size each of sub-batch should be close to the size of the embedding x 2. Does this make sense? Is the size of embedding 512? If yes, do you have 1024 samples in a sub-batch? I would be happy if the authors can comment on this.\n\n3. I think my biggest concern with the paper is the lack of extensive experiments. All the experiments are done on small-scale datasets which is highly questionable when they are used for unsupervised learning. It would be much more convincing to have experiments on ImageNet which is a standard experiment for unsupervised learning.\n\n4. The improvement on relatively less challenging benchmarks are very marginal. Even in some cases, CIFAR100, contrastive loss does better than the proposed one. And the proposed method only outperforms BYOL in CIFAR10.\n\n5. My impression is that the contrastive method in table 1 and 2 represent the MoCo-v2. I would replace it with MoCo-v2 in these tables as there are many other methods that uses contrastive loss functions.\n\n6. Another problem with the experiments is that, they lack experiments on different downstream tasks such as object detection. \n\n7. The authors try the Euclidean and MSE distance to project positives closely. Did they consider the cosine similarity loss? I did not understand the point of using Euclidean distance experiments without normalization? What do they exactly prove? And we can always normalize the embeddings as it is done in the other methods. I would be happy to receive some comments on this from the authors.\n\n8. Finally, it would be nice to the advantages of the proposed method (removal of negatives) in terms of training complexity. Does it, as expected, reduce the complexity in the training time? If yes, quantifying it would increase the strength of the paper.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603954009305}, {"id": "zjeBkkuNAcT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1203/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed an MSE loss function with whitening (W-MSE) for self-supervised representation learning. The motivation is to reduce the demand of negative examples in contrastive representation learning. The proposed W-MSE loss function is compared with popular contrastive loss on a few benchmarks.\n\nContrastive learning is a popular topic in the self-supervised learning domain. Most of the existing methods rely on negative samps to avoid trivial solutions. This paper proposed a simple and clean solution to tackle this problem, by using whitening in the loss term. This paper is also well explained and illustrated.\n\nThe main weakness of the paper is the experiments. Based on the results, I am not convinced that the proposed W-MSE is effective. Here are more comments:\n\n(1) The experimental results are pretty close to the existing contrastive/BYOL baselines. On CIFAR-10 and STL-10, results are saturated thus the diff is very minor. On more challenging CIFAR-100 and Tiny ImageNet, the results are mixed when compared to BYOL. \n\n(2) Given that the results are very close, what are the other benefits of using W-MSE loss? For example, is the training speed faster than the other methods (contrastive, BYOL) with negative sampling? It would be nice to include such results to demonstrate the effectiveness of W-MSE.\n\n(3) This paper claims that without negative sampling, W-MSE loss encourages the use of more positive pairs in the batch. I'm wondering if the authors have tried more positive pairs beyond 4 in the paper.\n\n(4) I cannot understand the motivation of ablating the popular methods with Euclidean distance. Cosine similarity is just simple and it costs nothing compared to Euclidean distance. I think ablating this from existing contrastive loss is not sufficient to show the effectiveness of BYOL. \n \n(5) It would be nice to include a few comparable numbers from literature directly, instead of reproducing their methods. For example, this paper used ResNet18 while the BYOL paper used ResNet50. This paper reports results on Tiny ImageNet while the existing methods report on ImageNet. Note that I do not penalize this paper for this point, as computational cost is high for using larger architecture and larger dataset. But it would be good to have a comparable baseline in the middle ground, e.g. ResNet50 architecture on CIFAR-100 dataset, where BYOL reports 78.4% top-1 accuracy with linear eval. \n\n======Post Rebuttal Update======\n\nI would like to thank the authors for their rebuttal, which has addressed part of my concerns. After reading the authors' rebuttal and other reviewers' comments, I'm still concerned on the weak baselines and mixed results in this paper. Unfortunately, I will keep my rating.\n\nConcrete suggestions to improve this paper in the future:\n\n(1) Strongly recommended: use ResNet50 instead of ResNet18 for the small scale experiments, in this way you get directly the numbers from the literature (e.g. BYOL on CIFAR-100);\n\n(2) Nice to have: for the expensive ImageNet experiments, it would be nice to get comparable results using the smallest comparable architecture (e.g. ResNet50) from the literature. SimCLR claims that \"With 128 TPU v3 cores, it takes ∼1.5 hours to train our ResNet-50 with a batch size of 4096 for 100 epochs\" and MoCo claims that \"For IN-1M, we use a mini-batch size of 256 in 8 GPUs, ..., train for 200 epochs ..., taking ∼53 hours training ResNet-50.\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A simple and clean method, with very weak experiments. ", "review": "This paper proposed an MSE loss function with whitening (W-MSE) for self-supervised representation learning. The motivation is to reduce the demand of negative examples in contrastive representation learning. The proposed W-MSE loss function is compared with popular contrastive loss on a few benchmarks.\n\nContrastive learning is a popular topic in the self-supervised learning domain. Most of the existing methods rely on negative samps to avoid trivial solutions. This paper proposed a simple and clean solution to tackle this problem, by using whitening in the loss term. This paper is also well explained and illustrated.\n\nThe main weakness of the paper is the experiments. Based on the results, I am not convinced that the proposed W-MSE is effective. Here are more comments:\n\n(1) The experimental results are pretty close to the existing contrastive/BYOL baselines. On CIFAR-10 and STL-10, results are saturated thus the diff is very minor. On more challenging CIFAR-100 and Tiny ImageNet, the results are mixed when compared to BYOL. \n\n(2) Given that the results are very close, what are the other benefits of using W-MSE loss? For example, is the training speed faster than the other methods (contrastive, BYOL) with negative sampling? It would be nice to include such results to demonstrate the effectiveness of W-MSE.\n\n(3) This paper claims that without negative sampling, W-MSE loss encourages the use of more positive pairs in the batch. I'm wondering if the authors have tried more positive pairs beyond 4 in the paper.\n\n(4) I cannot understand the motivation of ablating the popular methods with Euclidean distance. Cosine similarity is just simple and it costs nothing compared to Euclidean distance. I think ablating this from existing contrastive loss is not sufficient to show the effectiveness of BYOL. \n \n(5) It would be nice to include a few comparable numbers from literature directly, instead of reproducing their methods. For example, this paper used ResNet18 while the BYOL paper used ResNet50. This paper reports results on Tiny ImageNet while the existing methods report on ImageNet. Note that I do not penalize this paper for this point, as computational cost is high for using larger architecture and larger dataset. But it would be good to have a comparable baseline in the middle ground, e.g. ResNet50 architecture on CIFAR-100 dataset, where BYOL reports 78.4% top-1 accuracy with linear eval. \n\n======Post Rebuttal Update======\n\nI would like to thank the authors for their rebuttal, which has addressed part of my concerns. After reading the authors' rebuttal and other reviewers' comments, I'm still concerned on the weak baselines and mixed results in this paper. Unfortunately, I will keep my rating.\n\nConcrete suggestions to improve this paper in the future:\n\n(1) Strongly recommended: use ResNet50 instead of ResNet18 for the small scale experiments, in this way you get directly the numbers from the literature (e.g. BYOL on CIFAR-100);\n\n(2) Nice to have: for the expensive ImageNet experiments, it would be nice to get comparable results using the smallest comparable architecture (e.g. ResNet50) from the literature. SimCLR claims that \"With 128 TPU v3 cores, it takes ∼1.5 hours to train our ResNet-50 with a batch size of 4096 for 100 epochs\" and MoCo claims that \"For IN-1M, we use a mini-batch size of 256 in 8 GPUs, ..., train for 200 epochs ..., taking ∼53 hours training ResNet-50.\"", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603918202477}, {"id": "1h488jkojc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1203/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n\n \nThe paper provides a simple self-supervision loss function which is computed using only positive samples. In particular, the proposed loss function is based on the whitening of a latent space-features. Results show competitive but not better results than other methods. \n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for accepting. The paper proposes a simple loss function based on whitening and the MSE loss. The method does not need negative samples to contrast with the positives since the whitening process shrinks the overall samples to satisfy the spherical-distribution while the MSE loss attracts the positive ones. Although I have some concerns with the validation I really like the method.\n\n \n##########################################################################\n\nPros: \n\n \n1. Interesting and simple method.\n\n2. Clarity of the paper.\n \n \n##########################################################################\n\nCons: \n\n1. The validation \n \n-- The evaluation has been carried with small and simple datasets (CIFAR,STL and Tiny ImageNet). BYOL authors presented an extensive experiment setup with several datasets. Why not use the same (at least some) databases and compare them with their reported results on the paper (and not use your own implemented version of BYOL)? In fact, some of those datasets are already used in this paper but as you use ResNet-18 instead of ResNet-50 (used in BYOL paper) the results are not comparable. \n\n-- Why not ResNet-50 instead of ResNet-18. Validation would be easier with other published methods\n\n-- W-MSE with d=4 seems to work better than d=2. What about larger d?\n\n-- Although the authors point out that the proposed MSE loss does not need negative samples, negative samples are needed in the whitening transformation. \n\n-- It would be great to decouple the system performance. What is the system performance using Withening with the classical contrastive loss? Which is the improvement provided by the MSE loss over the classical contrastive loss?\n\n-- The authors say that Contrastive Loss needs large numbers of negative samples. How many negative samples were used in the experimental setup? The proposed method is evaluated with d=2 and d=4. Which is the positive/negative proportion of samples in the batches using the contrastive loss? How was this proportion selected?\n\nMinor Comments:\n- Please check: \"On the other hand, Hjelm et al. (2019) have shown that the contrastive loss needs a large number of negatives to be competitive\"\n\n\n \n##########################################################################\n\nQuestions during the rebuttal period: \n\n \nPlease address and clarify the cons above \n\n \n#########################################################################\n\nSome typos: \n\n(1) Page 4: matrix and (z_i,z_j) correspond -> corresponds\n(2) Page 7: jitterering --> jittering\n\n\nI think that the paper\n\n\nUPDATE AFTER REBUTTAL:\nThe authors have covered most of my concerns about the paper and I think that the paper has been substantially improved. However, my biggest concern was about the experiment results and  I think the paper still lacks on validation comparison with other methods. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method ", "review": "##########################################################################\n\nSummary:\n\n \nThe paper provides a simple self-supervision loss function which is computed using only positive samples. In particular, the proposed loss function is based on the whitening of a latent space-features. Results show competitive but not better results than other methods. \n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for accepting. The paper proposes a simple loss function based on whitening and the MSE loss. The method does not need negative samples to contrast with the positives since the whitening process shrinks the overall samples to satisfy the spherical-distribution while the MSE loss attracts the positive ones. Although I have some concerns with the validation I really like the method.\n\n \n##########################################################################\n\nPros: \n\n \n1. Interesting and simple method.\n\n2. Clarity of the paper.\n \n \n##########################################################################\n\nCons: \n\n1. The validation \n \n-- The evaluation has been carried with small and simple datasets (CIFAR,STL and Tiny ImageNet). BYOL authors presented an extensive experiment setup with several datasets. Why not use the same (at least some) databases and compare them with their reported results on the paper (and not use your own implemented version of BYOL)? In fact, some of those datasets are already used in this paper but as you use ResNet-18 instead of ResNet-50 (used in BYOL paper) the results are not comparable. \n\n-- Why not ResNet-50 instead of ResNet-18. Validation would be easier with other published methods\n\n-- W-MSE with d=4 seems to work better than d=2. What about larger d?\n\n-- Although the authors point out that the proposed MSE loss does not need negative samples, negative samples are needed in the whitening transformation. \n\n-- It would be great to decouple the system performance. What is the system performance using Withening with the classical contrastive loss? Which is the improvement provided by the MSE loss over the classical contrastive loss?\n\n-- The authors say that Contrastive Loss needs large numbers of negative samples. How many negative samples were used in the experimental setup? The proposed method is evaluated with d=2 and d=4. Which is the positive/negative proportion of samples in the batches using the contrastive loss? How was this proportion selected?\n\nMinor Comments:\n- Please check: \"On the other hand, Hjelm et al. (2019) have shown that the contrastive loss needs a large number of negatives to be competitive\"\n\n\n \n##########################################################################\n\nQuestions during the rebuttal period: \n\n \nPlease address and clarify the cons above \n\n \n#########################################################################\n\nSome typos: \n\n(1) Page 4: matrix and (z_i,z_j) correspond -> corresponds\n(2) Page 7: jitterering --> jittering\n\n\nI think that the paper\n\n\nUPDATE AFTER REBUTTAL:\nThe authors have covered most of my concerns about the paper and I think that the paper has been substantially improved. However, my biggest concern was about the experiment results and  I think the paper still lacks on validation comparison with other methods. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603820174253}, {"id": "TvReA43uQrQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1203/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to first do representation \"whitening\", so that the representations are scattered in the space and not collapsing to a single data point; then compute distance metric on top of that (e.g. Euclidean, cosine similarity). A nice thing about explicit scattering is that it does not require large numbers of negative examples to pull the features apart. Experiments are done on several toy datasets like CIFAR.\n\n+ The paper is a nice, alternative attempt to remove negative examples in contrastive learning. Indeed, large number of negative examples is annoying and this direction is both exciting and significant.\n+ The approach proposed in this paper intuitively makes sense. There are several works already explaining that contrastive learning is essentially doing some kind of scattering in the space. \n+ The proposed approach seems pretty simple. The whitening code is only a dozen lines in PyTorch. I haven't run the code to verify the results though.\n\nI think the experiments are not too satisfying though. \n\n- The comparison is less fair between BYOL and multi-crop version of W-MSE. In SwAV, it shows that with multiple crops, the performance can be boosted quite a bit. I haven't tried on BYOL but I believe it could also be helping there. So the most apple-to-apple comparison between BYOL and W-MSE is the 2-crop version (d=2). In this case, BYOL is outperforming in most entries. Though it can be viewed as concurrent work (I think W-MSE is actually even earlier than BYOL), but the experiment session in the paper is not clear about this.\n- Overall running experiments on these toy datasets are less satisfying, not only because it lacks comparison to other major approaches (like MoCo on ImageNet), but also because the signal we get from smaller datasets may not transfer well to more real-world images. \n- I would like to see a comparison in terms of timing -- maybe BYOL (because of its momentum encoder and it needs 4 forward pass of the network to compute a single pair of losses) is running much smaller in training. W-MSE can run much faster because it only needs 4 (or even 1?) forward pass. This is a potential advantage that W-MSE has, but it is not clear from the paper.\n\nOther than experiments, I am also not too satisfied with the writing. The paper mentions another paper when talking about the key method (how to do whitening, e.g. which is back propagated, which is not) when describing the central technique of the paper. I would like to see the paper more self-contained in the next version. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting attempt to remove negative examples by whitening, however experiments are not too satisfying", "review": "The paper proposes to first do representation \"whitening\", so that the representations are scattered in the space and not collapsing to a single data point; then compute distance metric on top of that (e.g. Euclidean, cosine similarity). A nice thing about explicit scattering is that it does not require large numbers of negative examples to pull the features apart. Experiments are done on several toy datasets like CIFAR.\n\n+ The paper is a nice, alternative attempt to remove negative examples in contrastive learning. Indeed, large number of negative examples is annoying and this direction is both exciting and significant.\n+ The approach proposed in this paper intuitively makes sense. There are several works already explaining that contrastive learning is essentially doing some kind of scattering in the space. \n+ The proposed approach seems pretty simple. The whitening code is only a dozen lines in PyTorch. I haven't run the code to verify the results though.\n\nI think the experiments are not too satisfying though. \n\n- The comparison is less fair between BYOL and multi-crop version of W-MSE. In SwAV, it shows that with multiple crops, the performance can be boosted quite a bit. I haven't tried on BYOL but I believe it could also be helping there. So the most apple-to-apple comparison between BYOL and W-MSE is the 2-crop version (d=2). In this case, BYOL is outperforming in most entries. Though it can be viewed as concurrent work (I think W-MSE is actually even earlier than BYOL), but the experiment session in the paper is not clear about this.\n- Overall running experiments on these toy datasets are less satisfying, not only because it lacks comparison to other major approaches (like MoCo on ImageNet), but also because the signal we get from smaller datasets may not transfer well to more real-world images. \n- I would like to see a comparison in terms of timing -- maybe BYOL (because of its momentum encoder and it needs 4 forward pass of the network to compute a single pair of losses) is running much smaller in training. W-MSE can run much faster because it only needs 4 (or even 1?) forward pass. This is a potential advantage that W-MSE has, but it is not clear from the paper.\n\nOther than experiments, I am also not too satisfied with the writing. The paper mentions another paper when talking about the key method (how to do whitening, e.g. which is back propagated, which is not) when describing the central technique of the paper. I would like to see the paper more self-contained in the next version. ", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603673032813}], "openreview_url": "https://openreview.net/forum?id=3Wp8HM2CNdR", "arxiv_id": "2007.06346", "paper_pdf": "papers/3Wp8HM2CNdR.pdf", "paper_pdf_sha256": "241282b4a3aed827fd56034b81ecb8cbe4e035d212a72e57b5461dd980d56cd3", "paper_pdf_bytes": 1720986, "paper_pdf_source": "openreview", "code_url": "https://github.com/htdt/self-supervised", "code_repository": "htdt/self-supervised", "code_commit": "d9662c8d07dafd194a9045375f4f6aa09f5b03e9", "code_archive": "repos/3Wp8HM2CNdR.zip", "code_archive_sha256": "7f8bdde022d80048d6fc16e7e693b44d3fb8c3bf19e6bf8a6d764026e1a0cb26", "code_archive_bytes": 20021, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 87, "github_languages": {"Python": 30903, "Dockerfile": 206}, "github_archived": false, "github_pushed_at": "2023-02-05T10:14:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/whitening-for-self-supervised-representation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Skl3SkSKDr", "year": 2020, "status": "rejected", "title": "Generating valid Euclidean distance matrices", "authors": ["Moritz Hoffmann", "Frank Noe"], "authorids": ["moritz.hoffmann@fu-berlin.de", "frank.noe@fu-berlin.de"], "authors_source": "OpenReview API", "abstract": "Generating point clouds, e.g., molecular structures, in arbitrary rotations, translations, and enumerations remains a challenging task. Meanwhile, neural networks\nutilizing symmetry invariant layers have been shown to be able to optimize their\ntraining objective in a data-efficient way. In this spirit, we present an architecture\nwhich allows to produce valid Euclidean distance matrices, which by construction are already invariant under rotation and translation of the described object.\nMotivated by the goal to generate molecular structures in Cartesian space, we use\nthis architecture to construct a Wasserstein GAN utilizing a permutation invariant critic network. This makes it possible to generate molecular structures in a\none-shot fashion by producing Euclidean distance matrices which have a three-\ndimensional embedding.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BkeMvX5TFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1707/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\t\nThis paper addresses the important problem of molecular structures generation, and more generally of efficient point-cloud distributions learning in d-dimensional space.\nThe paper is written clearly, the pseudo-code is presented in a clear way, striking a good balance between explicit-ness and concision.\n\nAfter my first reading I was disappointed in the experiments, because I read the paper quite quickly at first.  After a second more careful read I understood most of it and was much more enthusiastic. Disclaimer: I am inexperienced in this particular field (GANs and GANs for molecular generation) so I lack literature knowledge, and I may be over-evaluating the quality of performances (compared to recent results).\nBut currently I believe that the authors are a bit too humble about their results (a rather uncommon phenomenon). There is a part of the method that I would like to have clarified (about bond/atom types), but other than this clarification, I strongly recommend the paper to be accepted.\n\nThe paper deals with a couple of distinct, related problems.  One is that of sampling valid (Euclidean) distances matrices (EDM). This is done with algorithm 1.\nThe paper uses these EDM to train a generator G directly in EDM-space, against a Critic network C (the architecture of which is taken from existing literature).\nSome of the output configurations have to be discarded due to incorrect bond types assignment (this is the part that is still obscure to me).\nThe remaining outputs are chemically sensible in terms of bond types/valence/local chemistry/etc, but also, and quite impressively, have reasonable ground state energies !  In this sense, the paper produces samples that may deserve to be added into the QM9 data set (After some data augmentation using DFT or other physics-validated methods of course).\n\n\nI have mainly two requests:\n\n1. The bond type or atom assignment is not clearly explained. I ask further detailed questions below, but overall I think you should attempt to make a pedagogical explanation of how the atom types are assigned/learned, and how some of them are later discarded.\n\n'The  generator produces an additional type vector in a multi-task fashion which is checked against a  constant type reference with a cross-entropy loss.'\nCould you explain this better ? At least in the appendix. Otherwise this very dense sentence remains mysterious. Eq.11 is currently the single occurrence of the H, t, and tref terms.\n\nAbout the validity test using Open Babel. Again, what happens with bond types is not very clear. Are they set in stone at generation time, and then most of these affected types are 'wrong' and are discarded by Open Babel ?\nWhy don't you include this validation at training time (if it's very complicated to do, explain why) ?\nIn any case, please clarify this paragraph, as for now it is cryptic.\n\n2. Part of the paper goal is to achieve learning of point clouds distributions and not about chemistry.  However, discovery of new structures (and their conformations) is in itself a big topic.  I think it would be good to mention a couple of follow ups to your work in the conclusion. For instance, it would seem rather natural to me to include some equivalent of OpenBAbel and Energy-estimates within the learning loop, so that the generator directly generates valid (open-babel wise) and reasonable (energy level-wise) structures.\nBesides the conclusion, you should stress out better the significance of your results with a couple of comments here and there. (I have more precise suggestions below).\n\n\nSome other comments to improve the paper (in clarity or other):\n\nIn algorithm 1, you should precise that G() is your NN-based generator network. At this point of the paper G and z and N_z have never been mentioned. This is a problem when reading the paper for the first time.\nLine 5, you could explicitly precise, 'with (5)'.  Also I am puzzled by the use of the softplus here, but later in page 5 you say g() is softplus for the first three eigenvalues, and set to 0 for all others. Why not already anticipate and say this in section 2 ?\nLine 8, I would have written 'eigenvalues .... of D'  (what is the role of Eq. 1 here?)\n\nAt some point (i.e. around end of section 2), a comment would be welcomed, about whether you may sample the space of EDMs uniformly at random using algorithm 1 (I think you do not, and it is ok, but the question naturally arises and it's not addressed).\n\n´´which transforms a prior distribution into a target distribution´´\nyou could specify that this prior is the Gaussian N(0,1)^Nz in this case, to better connect with algorithm 1.\n\nAbout equation 6: it would be nice to have some intuitive explanation of what the output values of C(x) mean. They are scalars that represent the opinion of C on the molecule x, so they are the probability that the observed molecule is 'a true molecule x' ?  You should recall that for the inexperienced reader.\n\nStill about Eq.6 : could you quickly provide the motivation for demanding C to be L-Lipschit with  $L \\leq 1$ ?\n\nThe drift term Eq.9 seems to be some sort of regularization, maybe it could be mentioned once in the text for completeness?\n\nUsing Mopac Stewart and figure 4: you could insist more on this result. Up to that point I was very dubious about the usefulness of the whole work, because I would have expected many generated samples to be highly unrealistic, i.e. having huge energies, and so being extremely unstable. Instead here you show it is not the case, and even all your energies lie within the range of observed energies !  This is a very strong result, that is not obvious to expect, and is highly valuable (even after keeping only 7.5% of structures, this is still a strong result.)\nI think this test and the corresponding result should be emphasized more.\n\nYou observe new topological types and complain there are not many. I think ~20 new topologies is not a small number (for so few atoms), and you may be rather proud of it.  What is too bad is that you do not have a tool for differentiating between the very similar conformations (that correspond to thermal fluctuations around a given structure) and the conformations that encode a new structure (which does not necessarily means new topology). \nIf you had this tool, you could enrich figure 3 with a curve showing the new structures (not counting conformational variants).\n\nrefs 2017a and 2017b are the same.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "\t\nThis paper addresses the important problem of molecular structures generation, and more generally of efficient point-cloud distributions learning in d-dimensional space.\nThe paper is written clearly, the pseudo-code is presented in a clear way, striking a good balance between explicit-ness and concision.\n\nAfter my first reading I was disappointed in the experiments, because I read the paper quite quickly at first.  After a second more careful read I understood most of it and was much more enthusiastic. Disclaimer: I am inexperienced in this particular field (GANs and GANs for molecular generation) so I lack literature knowledge, and I may be over-evaluating the quality of performances (compared to recent results).\nBut currently I believe that the authors are a bit too humble about their results (a rather uncommon phenomenon). There is a part of the method that I would like to have clarified (about bond/atom types), but other than this clarification, I strongly recommend the paper to be accepted.\n\nThe paper deals with a couple of distinct, related problems.  One is that of sampling valid (Euclidean) distances matrices (EDM). This is done with algorithm 1.\nThe paper uses these EDM to train a generator G directly in EDM-space, against a Critic network C (the architecture of which is taken from existing literature).\nSome of the output configurations have to be discarded due to incorrect bond types assignment (this is the part that is still obscure to me).\nThe remaining outputs are chemically sensible in terms of bond types/valence/local chemistry/etc, but also, and quite impressively, have reasonable ground state energies !  In this sense, the paper produces samples that may deserve to be added into the QM9 data set (After some data augmentation using DFT or other physics-validated methods of course).\n\n\nI have mainly two requests:\n\n1. The bond type or atom assignment is not clearly explained. I ask further detailed questions below, but overall I think you should attempt to make a pedagogical explanation of how the atom types are assigned/learned, and how some of them are later discarded.\n\n'The  generator produces an additional type vector in a multi-task fashion which is checked against a  constant type reference with a cross-entropy loss.'\nCould you explain this better ? At least in the appendix. Otherwise this very dense sentence remains mysterious. Eq.11 is currently the single occurrence of the H, t, and tref terms.\n\nAbout the validity test using Open Babel. Again, what happens with bond types is not very clear. Are they set in stone at generation time, and then most of these affected types are 'wrong' and are discarded by Open Babel ?\nWhy don't you include this validation at training time (if it's very complicated to do, explain why) ?\nIn any case, please clarify this paragraph, as for now it is cryptic.\n\n2. Part of the paper goal is to achieve learning of point clouds distributions and not about chemistry.  However, discovery of new structures (and their conformations) is in itself a big topic.  I think it would be good to mention a couple of follow ups to your work in the conclusion. For instance, it would seem rather natural to me to include some equivalent of OpenBAbel and Energy-estimates within the learning loop, so that the generator directly generates valid (open-babel wise) and reasonable (energy level-wise) structures.\nBesides the conclusion, you should stress out better the significance of your results with a couple of comments here and there. (I have more precise suggestions below).\n\n\nSome other comments to improve the paper (in clarity or other):\n\nIn algorithm 1, you should precise that G() is your NN-based generator network. At this point of the paper G and z and N_z have never been mentioned. This is a problem when reading the paper for the first time.\nLine 5, you could explicitly precise, 'with (5)'.  Also I am puzzled by the use of the softplus here, but later in page 5 you say g() is softplus for the first three eigenvalues, and set to 0 for all others. Why not already anticipate and say this in section 2 ?\nLine 8, I would have written 'eigenvalues .... of D'  (what is the role of Eq. 1 here?)\n\nAt some point (i.e. around end of section 2), a comment would be welcomed, about whether you may sample the space of EDMs uniformly at random using algorithm 1 (I think you do not, and it is ok, but the question naturally arises and it's not addressed).\n\n´´which transforms a prior distribution into a target distribution´´\nyou could specify that this prior is the Gaussian N(0,1)^Nz in this case, to better connect with algorithm 1.\n\nAbout equation 6: it would be nice to have some intuitive explanation of what the output values of C(x) mean. They are scalars that represent the opinion of C on the molecule x, so they are the probability that the observed molecule is 'a true molecule x' ?  You should recall that for the inexperienced reader.\n\nStill about Eq.6 : could you quickly provide the motivation for demanding C to be L-Lipschit with  $L \\leq 1$ ?\n\nThe drift term Eq.9 seems to be some sort of regularization, maybe it could be mentioned once in the text for completeness?\n\nUsing Mopac Stewart and figure 4: you could insist more on this result. Up to that point I was very dubious about the usefulness of the whole work, because I would have expected many generated samples to be highly unrealistic, i.e. having huge energies, and so being extremely unstable. Instead here you show it is not the case, and even all your energies lie within the range of observed energies !  This is a very strong result, that is not obvious to expect, and is highly valuable (even after keeping only 7.5% of structures, this is still a strong result.)\nI think this test and the corresponding result should be emphasized more.\n\nYou observe new topological types and complain there are not many. I think ~20 new topologies is not a small number (for so few atoms), and you may be rather proud of it.  What is too bad is that you do not have a tool for differentiating between the very similar conformations (that correspond to thermal fluctuations around a given structure) and the conformations that encode a new structure (which does not necessarily means new topology). \nIf you had this tool, you could enrich figure 3 with a curve showing the new structures (not counting conformational variants).\n\nrefs 2017a and 2017b are the same.\n\n"}, "tcdate": 1571820377677}, {"id": "rkeVnp86YH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1707/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Authors of this paper present architecture to produce valid Euclidean distance matrices. A Wasserstein GAN is then constructed by utilizing a permutation invariant critic network based on the architecture. Generating molecular structures in a one-shot fashion is conducted using the produced distance matrices in 3-d embedding.\n\nIn Section 2, the constraint on L makes M symmetric and positive semi-definite. This seems to be equivalent to treating M as a kernel matrix, and D is the pairwise distance between the kernel function induced by M. So, learning a valid Euclidian distance matrix is same as learning a kernel function.\n\nAuthors need to provide the evidence that Equation (5) holds for all function g, especially for the used softplus activation function. As I know, it holds if function g is a polynomial matrix function.\n\nIn Algorithm 1, what is the meaning of step 14? There is no definition or discussion about G and the \\nabla L. The construction function is the key contribution of this paper, which is incorporated the existing SchNet, so it is better to show the advantage of the proposed construction method comparing with SchNet.\n\nAlthough this paper is an application-oriented paper, the comparisons with baseline methods are preferred, such as some simple and straightforward baselines. Due to the lack of comparisons, it is hard to understand the statement given by authors that “ the current performance… is not optimal and can likely be improved by a better hyperparameter selection”.\n\nIn Section 4, authors calculate the similarity of generated molecules by the closest respective matches, which is determined by the maximal atomic distance after assignment of atom identities and superposition. Is this the standard way to compute the similarity between two graph structures?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Authors of this paper present architecture to produce valid Euclidean distance matrices. A Wasserstein GAN is then constructed by utilizing a permutation invariant critic network based on the architecture. Generating molecular structures in a one-shot fashion is conducted using the produced distance matrices in 3-d embedding.\n\nIn Section 2, the constraint on L makes M symmetric and positive semi-definite. This seems to be equivalent to treating M as a kernel matrix, and D is the pairwise distance between the kernel function induced by M. So, learning a valid Euclidian distance matrix is same as learning a kernel function.\n\nAuthors need to provide the evidence that Equation (5) holds for all function g, especially for the used softplus activation function. As I know, it holds if function g is a polynomial matrix function.\n\nIn Algorithm 1, what is the meaning of step 14? There is no definition or discussion about G and the \\nabla L. The construction function is the key contribution of this paper, which is incorporated the existing SchNet, so it is better to show the advantage of the proposed construction method comparing with SchNet.\n\nAlthough this paper is an application-oriented paper, the comparisons with baseline methods are preferred, such as some simple and straightforward baselines. Due to the lack of comparisons, it is hard to understand the statement given by authors that “ the current performance… is not optimal and can likely be improved by a better hyperparameter selection”.\n\nIn Section 4, authors calculate the similarity of generated molecules by the closest respective matches, which is determined by the maximal atomic distance after assignment of atom identities and superposition. Is this the standard way to compute the similarity between two graph structures?\n"}, "tcdate": 1571806636080}, {"id": "Skx9wBpHFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1707/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is extremely interesting, solid and very well written. The idea is simple but nonetheless developed in a smart and effective fashion. The underlying theory is solid, even if some choices should have been discussed more deeply (e.g. the chosen loss function). Introduction and references are adequate, and the paper is readable by a quite broad audience, despite the detailed technical sections. The main issue related to the manuscript is very narrow target of the experimental part, limited to the isomers of a given compound - it would have been interesting to check its potentialities in generating more different structures and distance matrices, and thus to compare its effectiveness versus alternative generative approaches.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper is extremely interesting, solid and very well written. The idea is simple but nonetheless developed in a smart and effective fashion. The underlying theory is solid, even if some choices should have been discussed more deeply (e.g. the chosen loss function). Introduction and references are adequate, and the paper is readable by a quite broad audience, despite the detailed technical sections. The main issue related to the manuscript is very narrow target of the experimental part, limited to the isomers of a given compound - it would have been interesting to check its potentialities in generating more different structures and distance matrices, and thus to compare its effectiveness versus alternative generative approaches."}, "tcdate": 1571308898438}], "openreview_url": "https://openreview.net/forum?id=Skl3SkSKDr", "arxiv_id": "1910.03131", "paper_pdf": "papers/Skl3SkSKDr.pdf", "paper_pdf_sha256": "969e8ccab5a960b005d1db73848a6c3f0ac02a15094960d580f0183ac31572fe", "paper_pdf_bytes": 1475096, "paper_pdf_source": "openreview", "code_url": "https://github.com/noegroup/EDMnets", "code_repository": "noegroup/EDMnets", "code_commit": "c73eedd31248166ce739e9a196f08ea0ec23e54f", "code_archive": "repos/Skl3SkSKDr.zip", "code_archive_sha256": "21abb8ba0f7afb1c68382fb206a569b35d673b37eb57c8a5e8960c785af9617c", "code_archive_bytes": 131239, "code_file_count": 19, "code_extensions": {".py": 13, ".h": 5, ".cpp": 1}, "github_disk_usage_kb": 140, "github_languages": {"Python": 111314, "C++": 17508, "C": 87}, "github_archived": false, "github_pushed_at": "2019-10-18T16:52:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generating-valid-euclidean-distance-matrices-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1x33sC9KQ", "year": 2019, "status": "rejected", "title": "ACIQ: Analytical Clipping for Integer Quantization of neural networks", "authors": ["Ron Banner", "Yury Nahshan", "Elad Hoffer", "Daniel Soudry"], "authorids": ["ron.banner@intel.com", "yury.nahshan@intel.com", "daniel.soudry@gmail.com", "elad.hoffer@gmail.com"], "authors_source": "OpenReview API", "abstract": "We analyze the trade-off between quantization noise and clipping distortion in low precision networks. We identify the statistics of various tensors, and derive exact expressions for the mean-square-error degradation due to clipping. By optimizing these expressions, we show marked improvements over standard quantization schemes that normally avoid clipping. For example, just by choosing the accurate clipping values, more than 40\\% accuracy improvement is obtained for the quantization of VGG-16 to 4-bits of precision. Our results have many applications for the quantization of neural networks at both training and inference time. \n", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SylOokVw6m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper749/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes a clipping method to improve the performance of one particular type of quantization method that is naive clipping to closest \"bins\". The contribution of the paper is the (possibly incorrect) derivation of the clipping value that causes the least quantization error IF assumptions can be made about the distribution of the parameters (in a non-bayesian sense). Thus, the significance is low due to both reasons.\n\nOne conceptual issue is the assumed relationship between quantization error and classification accuracy. The literature has shown that high quantization error does not necessarily mean low classification accuracy when using non-uniform quantization. The proposed clipping does not account for classification accuracy (on training set), but I understand the motivation being that the training set is not available. \n\n1. There seems to be an error in derivation of Eq (3), the first term should be $(x-sgn(x).\\alpha) = x+\\alpha$ for $x$ negative. Please comment on this.\n\n2. When solving the integrals, the authors simply pull the solution \"out of the hat\" and show that the derivative is the integrand. This is a very opaque presentation that we cannot see how you solved the integral. What is C in $\\psi(x)$?\n\n3. The assumptions on the parameters are only valid for the particular model/dataset/precision. The assumption does not generalize arbitrarily. For example, models with quantized weights have bi-modal distributions. How would you clip the  activations after e.g. a ReLu? This is without going in to the weaknesses of the K-S test. \n\n4. Experiments do not show any comparison to the large body of prior work in this area. \n\n5. Page 4, para below (3), what is \"common additive orthogonal noise\"? You should explain or give intuition instead of simply referring to a different paper.\n\n6. In the uniform case, one would think f(x)=1/<range of the interval>=2\\alpha. Why is it 1/\\Delta?\n\n6. Section 4, range should be [-\\alpha, \\alpha] instead of [\\alpha, -\\alpha]? Since \\alpha is positive.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Errors and Contributions not significant ", "review": "The paper describes a clipping method to improve the performance of one particular type of quantization method that is naive clipping to closest \"bins\". The contribution of the paper is the (possibly incorrect) derivation of the clipping value that causes the least quantization error IF assumptions can be made about the distribution of the parameters (in a non-bayesian sense). Thus, the significance is low due to both reasons.\n\nOne conceptual issue is the assumed relationship between quantization error and classification accuracy. The literature has shown that high quantization error does not necessarily mean low classification accuracy when using non-uniform quantization. The proposed clipping does not account for classification accuracy (on training set), but I understand the motivation being that the training set is not available. \n\n1. There seems to be an error in derivation of Eq (3), the first term should be $(x-sgn(x).\\alpha) = x+\\alpha$ for $x$ negative. Please comment on this.\n\n2. When solving the integrals, the authors simply pull the solution \"out of the hat\" and show that the derivative is the integrand. This is a very opaque presentation that we cannot see how you solved the integral. What is C in $\\psi(x)$?\n\n3. The assumptions on the parameters are only valid for the particular model/dataset/precision. The assumption does not generalize arbitrarily. For example, models with quantized weights have bi-modal distributions. How would you clip the  activations after e.g. a ReLu? This is without going in to the weaknesses of the K-S test. \n\n4. Experiments do not show any comparison to the large body of prior work in this area. \n\n5. Page 4, para below (3), what is \"common additive orthogonal noise\"? You should explain or give intuition instead of simply referring to a different paper.\n\n6. In the uniform case, one would think f(x)=1/<range of the interval>=2\\alpha. Why is it 1/\\Delta?\n\n6. Section 4, range should be [-\\alpha, \\alpha] instead of [\\alpha, -\\alpha]? Since \\alpha is positive.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1542041503944}, {"id": "HygMWn2Dn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper749/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper derives a formula for finding the minimum and maximum clipping values for uniform quantization which minimize the square error resulting from quantization, for either a Laplace or Gaussian distribution over pre-quantized value. This seems like too small a contribution to warrant a paper. I wasn't convinced that appropriate baselines were used in experiments. There were a number of statements that I believed to be technically slightly incorrect. There were also some small language problems (though these didn't hinder understanding).\n\nmore specific comments:\n\nabstract:\n\"derive exact expressions\" -- these expressions aren't exact. they turn out to be based on a piecewise zeroth order Taylor approximation to the density.\n\nmain paper:\n\"allow fit bigger networks into\" -> \"allow bigger network to fit into\"\n\"that we are need\" -> \"that need\"\n\"introduces an additional\" -> \"introduces additional\"\nclippig -> clipping\n\nit's not clear a-priori that information loss is the property to minimize that maximizes performance of the quantized network.\n\n\"distributions of tensors\" -> \"distribution of tensor elements\"\nthis comment also applies in a number of other places, where the writing refers to the marginal distribution of values taken on by entries in a tensor as the distribution over the tensor. note that a distribution over tensors is a joint distribution over all entries in a tensor. e.g. it would capture things like eigenvalues, entry-entry covariance, rather than just marginal statistics.\n\n\"than they could have by working individually\" -> \"than could have been achieved by each individually\"\n\nWhy the focus on small activation bit depth? I would imagine weight bit-depth was more important than activation bit depth. Especially since you're using ?32-bit? precision in the weight/activations multiplications, so activations are computed at a high bit depth anyways.\n\nTable 1: Give absolute accuracies too! Improvement relative to what baseline?\n\nsec 2:\nsufficeint -> sufficient\n\\citep often used when it should instead be \\citet.\n\"As contrast\" -> \"In contrast\"\n\nsection 3:\nuniformity -> uniformly\n\nI don't believe the notion of p-value is being used correctly here w.r.t. the Kolmogorov-Smirnov test.\n\nFigure 1: The mean square error should never go to 0. This suggests something is wrong. If it's just a scaling issue, consider a semilogy plot.\n\nFigure 2: I'm unclear what baseline (no clipping) refers to in terms of clipping values. For uniform quantization there needs to be some min and max value.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "review": "This paper derives a formula for finding the minimum and maximum clipping values for uniform quantization which minimize the square error resulting from quantization, for either a Laplace or Gaussian distribution over pre-quantized value. This seems like too small a contribution to warrant a paper. I wasn't convinced that appropriate baselines were used in experiments. There were a number of statements that I believed to be technically slightly incorrect. There were also some small language problems (though these didn't hinder understanding).\n\nmore specific comments:\n\nabstract:\n\"derive exact expressions\" -- these expressions aren't exact. they turn out to be based on a piecewise zeroth order Taylor approximation to the density.\n\nmain paper:\n\"allow fit bigger networks into\" -> \"allow bigger network to fit into\"\n\"that we are need\" -> \"that need\"\n\"introduces an additional\" -> \"introduces additional\"\nclippig -> clipping\n\nit's not clear a-priori that information loss is the property to minimize that maximizes performance of the quantized network.\n\n\"distributions of tensors\" -> \"distribution of tensor elements\"\nthis comment also applies in a number of other places, where the writing refers to the marginal distribution of values taken on by entries in a tensor as the distribution over the tensor. note that a distribution over tensors is a joint distribution over all entries in a tensor. e.g. it would capture things like eigenvalues, entry-entry covariance, rather than just marginal statistics.\n\n\"than they could have by working individually\" -> \"than could have been achieved by each individually\"\n\nWhy the focus on small activation bit depth? I would imagine weight bit-depth was more important than activation bit depth. Especially since you're using ?32-bit? precision in the weight/activations multiplications, so activations are computed at a high bit depth anyways.\n\nTable 1: Give absolute accuracies too! Improvement relative to what baseline?\n\nsec 2:\nsufficeint -> sufficient\n\\citep often used when it should instead be \\citet.\n\"As contrast\" -> \"In contrast\"\n\nsection 3:\nuniformity -> uniformly\n\nI don't believe the notion of p-value is being used correctly here w.r.t. the Kolmogorov-Smirnov test.\n\nFigure 1: The mean square error should never go to 0. This suggests something is wrong. If it's just a scaling issue, consider a semilogy plot.\n\nFigure 2: I'm unclear what baseline (no clipping) refers to in terms of clipping values. For uniform quantization there needs to be some min and max value.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541028858076}, {"id": "HJegw7SEhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper749/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper empirically finds that the distribution of activations in quantized networks follow  Gaussian or Laplacian distribution, and proposes to determine the optimal clipping factor by minimizing the quantization error based on the distribution assumption.\n\nThe pros of the work are its simplicity, the proposed clipping and quantization does not need additional re-training. However, while the key of this paper is to determine a good clipping factor, the authors use uniform density function to represent the middle part of both Gaussian and Laplacian distributions where the majority of data points lie in, but exact computation for the tails of the distributions at both ends. Thus the computation of quantization error is not quite convincing. Moreover, the authors do not compare with the other recent works that also clip the activations, thus it is hard to validate the efficacy of the proposed method.\n\nFor the experiments, the authors mention that a look-up table can be pre-computed for fast retrieval of clipping factors given the mean and sigma of a distribution.  However, the mean and sigma are continuous numbers, how is the look-up table made?  Moreover, how is the mean and std estimated for each weight tensor and what is  the complexity?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A simple but not very convincing clipping method for activation quantization in deep networks", "review": "This paper empirically finds that the distribution of activations in quantized networks follow  Gaussian or Laplacian distribution, and proposes to determine the optimal clipping factor by minimizing the quantization error based on the distribution assumption.\n\nThe pros of the work are its simplicity, the proposed clipping and quantization does not need additional re-training. However, while the key of this paper is to determine a good clipping factor, the authors use uniform density function to represent the middle part of both Gaussian and Laplacian distributions where the majority of data points lie in, but exact computation for the tails of the distributions at both ends. Thus the computation of quantization error is not quite convincing. Moreover, the authors do not compare with the other recent works that also clip the activations, thus it is hard to validate the efficacy of the proposed method.\n\nFor the experiments, the authors mention that a look-up table can be pre-computed for fast retrieval of clipping factors given the mean and sigma of a distribution.  However, the mean and sigma are continuous numbers, how is the look-up table made?  Moreover, how is the mean and std estimated for each weight tensor and what is  the complexity?\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540801367941}], "openreview_url": "https://openreview.net/forum?id=B1x33sC9KQ", "arxiv_id": null, "paper_pdf": "papers/B1x33sC9KQ.pdf", "paper_pdf_sha256": "0ca44443e9315e4d24d17e48138263ad121803955018e1f33320b81ded23b959", "paper_pdf_bytes": 740828, "paper_pdf_source": "openreview", "code_url": "https://github.com/submission2019/AnalyticalScaleForIntegerQuantization", "code_repository": "submission2019/AnalyticalScaleForIntegerQuantization", "code_commit": "3246ee8cbfb747d7ef821c8cecc50283a73eaf92", "code_archive": "repos/B1x33sC9KQ.zip", "code_archive_sha256": "9a3fc3214025cd06281a04708bc55f868c06017f50d6be496a9c0d4b721f0847", "code_archive_bytes": 957719, "code_file_count": 22, "code_extensions": {".py": 18, ".sh": 1, ".cu": 1, ".cpp": 1, ".ipynb": 1}, "github_disk_usage_kb": 1120, "github_languages": {"Jupyter Notebook": 413088, "Python": 87304, "Cuda": 1661, "C++": 337, "Shell": 328}, "github_archived": false, "github_pushed_at": "2019-01-20T07:53:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/aciq-analytical-clipping-for-integer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oGlgHjYKBi", "year": 2026, "status": "rejected", "title": "Model-Preserving Adaptive Rounding", "authors": ["Albert Tseng", "Zhaofeng Sun", "Christopher De Sa"], "authorids": ["~Albert_Tseng1", "~Zhaofeng_Sun1", "~Christopher_De_Sa2"], "authors_source": "OpenReview API", "abstract": "The goal of quantization is to produce a compressed model whose output distribution is as close to the original model's as possible. \nTo do this tractably, most quantization algorithms minimize the immediate activation error of each layer as a proxy for the end-to-end error.\nHowever, this ignores the effect of future layers, making it a poor proxy.\nIn this work, we introduce Yet Another Quantization Algorithm (YAQA), an adaptive rounding algorithm that directly considers the error at the network's output.\nYAQA introduces a series of theoretical results that culminate in the first end-to-end error bounds for quantization algorithms.\nFirst, we characterize the convergence time of adaptive rounding algorithms via the structure of their Hessian approximations.\nWe then show that the end-to-end error can be bounded by the approximation's cosine similarity to the true Hessian.\nThis admits a natural Kronecker-factored approximation with corresponding near-optimal Hessian sketches.\nYAQA is provably better than GPTQ/LDLQ and empirically reduces the error by $\\approx 30\\%$ over these methods.\nYAQA even achieves a lower error than quantization aware training.\nThis translates to state of the art performance on downstream tasks, all while adding no inference overhead.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "nSCkf7ZsrR", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6450/Reviewer_EML8"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces PTQ approach called YAQA which solves the issue in GPTQ and LDLQ of layerwise activation error minimization as a proxy to end to end model error. YAQA minimized the KL divergence between original model and quantized model output distribution. The paper also provides the theoretical bound on end to end quantization error and shows the cosine similarity between a Kronecker factored Hessian approximation and true Hessian. The evaluations in the paper show that YAQA improves KL error by approx 30% compared to GPTQ and LDLQ.", "review_text": "The paper introduces PTQ approach called YAQA which solves the issue in GPTQ and LDLQ of layerwise activation error minimization as a proxy to end to end model error. YAQA minimized the KL divergence between original model and quantized model output distribution. The paper also provides the theoretical bound on end to end quantization error and shows the cosine similarity between a Kronecker factored Hessian approximation and true Hessian. The evaluations in the paper show that YAQA improves KL error by approx 30% compared to GPTQ and LDLQ.", "strengths": "- The paper is easy to read, well structured and clearly written. The main claims in the paper are well supported by rigorous theoretical proofs and extensive empirical evaluation.\n- The paper provides novel contributions to a highly impactful area of model compression.\n- The theoretical framework is strong and justifies the adoption of Kronecker factored Hessian that enables fast and symmetric I/O feedback during adaptive rounding. \n- The paper provably shows superiority over LDLQ under the low rank condition.\n- The empirical results are strong. YAQA has been shown to consistently outperform LDLQ by reduction in KL div by approximately 30% margin. YAQA also has been shown to achieve lower KLD than large scale QAT Gemma-3 12B.", "weaknesses": "- Its not clear how much is the actual quantization time cost. The paper states that YAQA adds no inference overhead and has the same asymptotic complexity as LDLQ. It would be good to show actual wall clock time comparison against LDLQ on large model since YAQA requires power iteration on model hessian.\n- YAQA is provable shown to be better than LDLQ based on the assumption H_O is low rank. Could the authors show empirical evidence to support the assumption?\n- The evaluations are mainly focused on W4A16 but it is not clear from the paper if YAQA can be used for 3 bit or 2 bit where rounding noise is more dominant and also more challenging setup of W4A4 where both weights and activations are low precision.\n- The paper contrasts YAQA with block diagonal approximations. Could the authors provide a more detailed theoretical comparison, within the SND framework, of why the Kronecker factored structure is better than the block diagonal structure, beyond the empirical observation?", "questions": "- LLMs are known to suffer from activation outliers in intermediate layers. Since YAQA is based on an objective to minimize KL divergence, does the inclusion of the Hessian factor (H_O​) implicitly mitigate the effect of these activation outliers. \n- Can the authors show via an ablation on if the performance gain in YAQA is due to advance power iteration or due to the symmetric form?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces PTQ approach called YAQA which solves the issue in GPTQ and LDLQ of layerwise activation error minimization as a proxy to end to end model error. YAQA minimized the KL divergence between original model and quantized model output distribution. The paper also provides the theoretical bound on end to end quantization error and shows the cosine similarity between a Kronecker factored Hessian approximation and true Hessian. The evaluations in the paper show that YAQA improves KL error by approx 30% compared to GPTQ and LDLQ.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper is easy to read, well structured and clearly written. The main claims in the paper are well supported by rigorous theoretical proofs and extensive empirical evaluation.\n- The paper provides novel contributions to a highly impactful area of model compression.\n- The theoretical framework is strong and justifies the adoption of Kronecker factored Hessian that enables fast and symmetric I/O feedback during adaptive rounding. \n- The paper provably shows superiority over LDLQ under the low rank condition.\n- The empirical results are strong. YAQA has been shown to consistently outperform LDLQ by reduction in KL div by approximately 30% margin. YAQA also has been shown to achieve lower KLD than large scale QAT Gemma-3 12B.", "weaknesses": "- Its not clear how much is the actual quantization time cost. The paper states that YAQA adds no inference overhead and has the same asymptotic complexity as LDLQ. It would be good to show actual wall clock time comparison against LDLQ on large model since YAQA requires power iteration on model hessian.\n- YAQA is provable shown to be better than LDLQ based on the assumption H_O is low rank. Could the authors show empirical evidence to support the assumption?\n- The evaluations are mainly focused on W4A16 but it is not clear from the paper if YAQA can be used for 3 bit or 2 bit where rounding noise is more dominant and also more challenging setup of W4A4 where both weights and activations are low precision.\n- The paper contrasts YAQA with block diagonal approximations. Could the authors provide a more detailed theoretical comparison, within the SND framework, of why the Kronecker factored structure is better than the block diagonal structure, beyond the empirical observation?", "questions": "- LLMs are known to suffer from activation outliers in intermediate layers. Since YAQA is based on an objective to minimize KL divergence, does the inclusion of the Hessian factor (H_O​) implicitly mitigate the effect of these activation outliers. \n- Can the authors show via an ablation on if the performance gain in YAQA is due to advance power iteration or due to the symmetric form?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762270867595}, {"id": "eiwD8xDKLj", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6450/Reviewer_u5sZ"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "I am not an expert of LLM quantization, so my comments or reviews might not be very useful.\n\nThe authors propose a new type of adaptive rounding algorithm for LLM quantization, which enables efficient adaptive rounding (line 179 of the original paper). The authors conduct some theoretical analysis to ground their proposed algorithms, with two Hessian sketches (line 257 of the original paper), and provided empirical results on popular LLM models such as LLaMA and Gemma (Table 1), demonstrating comparable or better performance in terms of KL divergence with the full-precision model or other downstream task performance such as PPL and zero-shot accuracy.", "review_text": "I am not an expert of LLM quantization, so my comments or reviews might not be very useful.\n\nThe authors propose a new type of adaptive rounding algorithm for LLM quantization, which enables efficient adaptive rounding (line 179 of the original paper). The authors conduct some theoretical analysis to ground their proposed algorithms, with two Hessian sketches (line 257 of the original paper), and provided empirical results on popular LLM models such as LLaMA and Gemma (Table 1), demonstrating comparable or better performance in terms of KL divergence with the full-precision model or other downstream task performance such as PPL and zero-shot accuracy.", "strengths": "- The experiments seem to be extensive, as in Tables 1-4.\n- The theoretical analysis seems to be interesting and solid, but I am not good at math and theorem proof, and I am not capable of carefully checking the theorems and their correctness.", "weaknesses": "- In the introduction and background, the authors first introduced PTQ and QAT, and somewhere suddenly convert to adaptive rounding algorithm. I am not sure, but it seems adaptive rounding algorithm is something between these two methods (see lines 095-099 of the original paper). It might be better to claim here or somewhere that adaptive rounding algorithm is a new type of method similar to PTQ without training on data, but optimize the quantized weights from the original full-precision model, and the proposed YAQA is a new method of adaptive rounding (see line 014 in the original paper). If possible, it might also be possible to state that adaptive rounding is a distillation method, but I am not sure if such claim is accurate enough.\n- In the experiment section, it can be better to provide the dataset for evaluation in the table captions, at least in Table 1, and explain what W2 and C4 mean, either in the caption or in the context.\n- The improvement shown in Table 1 seems not strong enough, and not consistent.", "questions": "I am not an expert in LLM quantization, so I do not have very suggestive questions. Also, I do not believe increasing my score will be useful or constructive, but if the weaknesses listed above could be improved, and if other reviewers who have more background in this area believe this work is good enough, I will be glad to change my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "I am not an expert of LLM quantization, so my comments or reviews might not be very useful.\n\nThe authors propose a new type of adaptive rounding algorithm for LLM quantization, which enables efficient adaptive rounding (line 179 of the original paper). The authors conduct some theoretical analysis to ground their proposed algorithms, with two Hessian sketches (line 257 of the original paper), and provided empirical results on popular LLM models such as LLaMA and Gemma (Table 1), demonstrating comparable or better performance in terms of KL divergence with the full-precision model or other downstream task performance such as PPL and zero-shot accuracy.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The experiments seem to be extensive, as in Tables 1-4.\n- The theoretical analysis seems to be interesting and solid, but I am not good at math and theorem proof, and I am not capable of carefully checking the theorems and their correctness.", "weaknesses": "- In the introduction and background, the authors first introduced PTQ and QAT, and somewhere suddenly convert to adaptive rounding algorithm. I am not sure, but it seems adaptive rounding algorithm is something between these two methods (see lines 095-099 of the original paper). It might be better to claim here or somewhere that adaptive rounding algorithm is a new type of method similar to PTQ without training on data, but optimize the quantized weights from the original full-precision model, and the proposed YAQA is a new method of adaptive rounding (see line 014 in the original paper). If possible, it might also be possible to state that adaptive rounding is a distillation method, but I am not sure if such claim is accurate enough.\n- In the experiment section, it can be better to provide the dataset for evaluation in the table captions, at least in Table 1, and explain what W2 and C4 mean, either in the caption or in the context.\n- The improvement shown in Table 1 seems not strong enough, and not consistent.", "questions": "I am not an expert in LLM quantization, so I do not have very suggestive questions. Also, I do not believe increasing my score will be useful or constructive, but if the weaknesses listed above could be improved, and if other reviewers who have more background in this area believe this work is good enough, I will be glad to change my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761977921207}, {"id": "wY3yD91Wmx", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6450/Reviewer_odoJ"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper proposes YAQA (Yet Another Quantization Algorithm), a quantization method that directly optimizes end-to-end model error via a Kronecker-factored Hessian approximation, providing the first theoretical bounds on quantization error while preserving model behavior without affecting inference efficiency.", "review_text": "This paper proposes YAQA (Yet Another Quantization Algorithm), a quantization method that directly optimizes end-to-end model error via a Kronecker-factored Hessian approximation, providing the first theoretical bounds on quantization error while preserving model behavior without affecting inference efficiency.", "strengths": "1. YAQA provides the first formal end-to-end quantization error bounds, linking quantization quality directly to the cosine similarity between the true and approximated Hessians.\n2. Built on a Kronecker-factored Hessian approximation, YAQA is compatible with various quantizers and maintains the same computational complexity as prior adaptive rounding methods\n3. It introduces symmetric input–output feedback and structured Hessian forms, ensuring faster convergence and greater rounding stability", "weaknesses": "1. Is the Kronecker-factored Hessian approximation transferable across different model architectures?\n2. Can the “structural nilpotence degree” quantitatively predict convergence speed in real applications?\n3. Can the proposed method be applied to mixed-precision quantization?\n4. The paper claims that the proposed method introduces no additional inference overhead — could the authors provide supporting evidence or justification for this claim?\n5. The advantage of YAQA appears to mainly arise from its more accurate Hessian approximation — does this suggest that its core contribution lies in the approximation technique rather than in the optimization principle itself?", "questions": "See the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes YAQA (Yet Another Quantization Algorithm), a quantization method that directly optimizes end-to-end model error via a Kronecker-factored Hessian approximation, providing the first theoretical bounds on quantization error while preserving model behavior without affecting inference efficiency.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. YAQA provides the first formal end-to-end quantization error bounds, linking quantization quality directly to the cosine similarity between the true and approximated Hessians.\n2. Built on a Kronecker-factored Hessian approximation, YAQA is compatible with various quantizers and maintains the same computational complexity as prior adaptive rounding methods\n3. It introduces symmetric input–output feedback and structured Hessian forms, ensuring faster convergence and greater rounding stability", "weaknesses": "1. Is the Kronecker-factored Hessian approximation transferable across different model architectures?\n2. Can the “structural nilpotence degree” quantitatively predict convergence speed in real applications?\n3. Can the proposed method be applied to mixed-precision quantization?\n4. The paper claims that the proposed method introduces no additional inference overhead — could the authors provide supporting evidence or justification for this claim?\n5. The advantage of YAQA appears to mainly arise from its more accurate Hessian approximation — does this suggest that its core contribution lies in the approximation technique rather than in the optimization principle itself?", "questions": "See the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761913156294}, {"id": "r4Mr4lBqB1", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6450/Reviewer_RfGj"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper addresses an important challenge in post-training quantization (PTQ) for large language models: conventional methods quantize each linear layer by minimizing the immediate activation error locally, without considering how that rounding propagates through subsequent layers and affects the overall model output distribution. The authors propose YAQA (Yet Another Quantization Algorithm), a two-part adaptive rounding method: (1) they compute Kronecker-factored approximations of each layer’s Hessian of the full-model Kullback-Leibler (KL) divergence loss; (2) they use these sketches in a rounding algorithm with theoretical guarantees. It combines A \"Kronecker-factored Hessian approximation\" to efficiently estimate model sensitivity, and  an \"adaptive rounding algorithm\" that minimizes quantization error with respect to the full-model KL divergence.  \n\nThe approach maintains model fidelity more effectively than standard PTQ, achieving up to 30% lower KL divergence and improved downstream performance on large language models. \n\nMy recommendation is to accept this paper provided Authors are able to demonstrate that this method can be extended to other important layers of transformers.", "review_text": "The paper addresses an important challenge in post-training quantization (PTQ) for large language models: conventional methods quantize each linear layer by minimizing the immediate activation error locally, without considering how that rounding propagates through subsequent layers and affects the overall model output distribution. The authors propose YAQA (Yet Another Quantization Algorithm), a two-part adaptive rounding method: (1) they compute Kronecker-factored approximations of each layer’s Hessian of the full-model Kullback-Leibler (KL) divergence loss; (2) they use these sketches in a rounding algorithm with theoretical guarantees. It combines A \"Kronecker-factored Hessian approximation\" to efficiently estimate model sensitivity, and  an \"adaptive rounding algorithm\" that minimizes quantization error with respect to the full-model KL divergence.  \n\nThe approach maintains model fidelity more effectively than standard PTQ, achieving up to 30% lower KL divergence and improved downstream performance on large language models. \n\nMy recommendation is to accept this paper provided Authors are able to demonstrate that this method can be extended to other important layers of transformers.", "strengths": "* Proposes a new quantization formulation that minimizes the KL divergence between the original and quantized model outputs, instead of per-layer reconstruction errors. This supposedly produces a better quantized model (also supported by experiments) \n* Introduces a Kronecker-factored Hessian approximation (\\(H_O \\otimes H_I\\)) to capture second-order information efficiently for large-scale networks.\n* Develops an adaptive rounding method that minimizes a Hessian-weighted quadratic error function, providing better alignment between quantized and original weights.\n* Derives theoretical error bounds for the rounding objective and proposes incoherence transformations (e.g., Hadamard transforms) to reduce approximation bias.\n* Ensures compatibility with multiple quantization schemes (INT4, INT8, mixed precision) without retraining or model-specific fine-tuning.\n* Demonstrates significant improvements in KL divergence and downstream metrics on large language models (e.g., LLaMA-family) compared to existing PTQ baselines (CBQ, LDLQ etc) .", "weaknesses": "* The method relies on the Kronecker-factored Hessian approximation \\(H_O \\otimes H_I\\) which is generally efficient, it is still an approximation and the performance depends heavily on how accurately it captures the true curvature of the loss.  The sensitivity of results to sketch quality, dataset size, or model architecture is underexplored.\n\n* Estimating Hessian sketches (via power iterations or data sampling) adds computational cost compared to simple rounding which can have implication on runtime and memory overhead, especially for very large LLMs and this needs to be quantified. \n* The algorithm mainly targets linear (fully-connected) layers.  It may not generalize directly to non-linear components such as attention modules, normalization layers, or activation quantization. Although linear layers are significant part of these models, attention needs to be tackled for a meaningful conclusion. \n* Experiments are focused on LLMs and it would have been better to add more tasks such as vision to see if the idea has benefit in other domains\n* While the paper focuses on post-training quantization (PTQ), comparing against quantization-aware training (QAT) baselines would clarify the performance gap and practical trade-offs. \n* Some critical implementation details — such as hyperparameter settings, sketching configurations, or data requirements — are not fully described.  Authors are encouraged to provide more details.", "questions": "* What are hyperparameter settings, sketching configurations etc.? /\n* Can Authors provide comparison of end to end model performance compared to any QAT method such as LSQ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses an important challenge in post-training quantization (PTQ) for large language models: conventional methods quantize each linear layer by minimizing the immediate activation error locally, without considering how that rounding propagates through subsequent layers and affects the overall model output distribution. The authors propose YAQA (Yet Another Quantization Algorithm), a two-part adaptive rounding method: (1) they compute Kronecker-factored approximations of each layer’s Hessian of the full-model Kullback-Leibler (KL) divergence loss; (2) they use these sketches in a rounding algorithm with theoretical guarantees. It combines A \"Kronecker-factored Hessian approximation\" to efficiently estimate model sensitivity, and  an \"adaptive rounding algorithm\" that minimizes quantization error with respect to the full-model KL divergence.  \n\nThe approach maintains model fidelity more effectively than standard PTQ, achieving up to 30% lower KL divergence and improved downstream performance on large language models. \n\nMy recommendation is to accept this paper provided Authors are able to demonstrate that this method can be extended to other important layers of transformers.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "* Proposes a new quantization formulation that minimizes the KL divergence between the original and quantized model outputs, instead of per-layer reconstruction errors. This supposedly produces a better quantized model (also supported by experiments) \n* Introduces a Kronecker-factored Hessian approximation (\\(H_O \\otimes H_I\\)) to capture second-order information efficiently for large-scale networks.\n* Develops an adaptive rounding method that minimizes a Hessian-weighted quadratic error function, providing better alignment between quantized and original weights.\n* Derives theoretical error bounds for the rounding objective and proposes incoherence transformations (e.g., Hadamard transforms) to reduce approximation bias.\n* Ensures compatibility with multiple quantization schemes (INT4, INT8, mixed precision) without retraining or model-specific fine-tuning.\n* Demonstrates significant improvements in KL divergence and downstream metrics on large language models (e.g., LLaMA-family) compared to existing PTQ baselines (CBQ, LDLQ etc) .", "weaknesses": "* The method relies on the Kronecker-factored Hessian approximation \\(H_O \\otimes H_I\\) which is generally efficient, it is still an approximation and the performance depends heavily on how accurately it captures the true curvature of the loss.  The sensitivity of results to sketch quality, dataset size, or model architecture is underexplored.\n\n* Estimating Hessian sketches (via power iterations or data sampling) adds computational cost compared to simple rounding which can have implication on runtime and memory overhead, especially for very large LLMs and this needs to be quantified. \n* The algorithm mainly targets linear (fully-connected) layers.  It may not generalize directly to non-linear components such as attention modules, normalization layers, or activation quantization. Although linear layers are significant part of these models, attention needs to be tackled for a meaningful conclusion. \n* Experiments are focused on LLMs and it would have been better to add more tasks such as vision to see if the idea has benefit in other domains\n* While the paper focuses on post-training quantization (PTQ), comparing against quantization-aware training (QAT) baselines would clarify the performance gap and practical trade-offs. \n* Some critical implementation details — such as hyperparameter settings, sketching configurations, or data requirements — are not fully described.  Authors are encouraged to provide more details.", "questions": "* What are hyperparameter settings, sketching configurations etc.? /\n* Can Authors provide comparison of end to end model performance compared to any QAT method such as LSQ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761497131624}], "openreview_url": "https://openreview.net/forum?id=oGlgHjYKBi", "arxiv_id": "2505.22988", "paper_pdf": "papers/oGlgHjYKBi.pdf", "paper_pdf_sha256": "e641e4e0e1cca650357d6e4ea5d1287d3a9fe06e4d46f78a4c91fd1a999e7215", "paper_pdf_bytes": 480975, "paper_pdf_source": "openreview", "code_url": "https://github.com/Cornell-RelaxML/yaqa-quantization", "code_repository": "Cornell-RelaxML/yaqa-quantization", "code_commit": "f9508723251ad839f0162326569f17fe70486fcc", "code_archive": "repos/oGlgHjYKBi.zip", "code_archive_sha256": "c98cf680caef3e7b48c4685f8e76f2e8b6f345d42389be6d42442b6a74e3ffa7", "code_archive_bytes": 379085, "code_file_count": 42, "code_extensions": {".py": 36, ".cu": 3, ".sh": 1, ".h": 1, ".cpp": 1}, "github_disk_usage_kb": 547, "github_languages": {"Python": 2045846, "Cuda": 45186, "C++": 26529, "Makefile": 432, "Shell": 328, "HTML": 53}, "github_archived": false, "github_pushed_at": "2025-06-20T14:03:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/model-preserving-adaptive-rounding"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WDheQxWAo4", "year": 2025, "status": "rejected", "title": "Simple synthetic data reduces sycophancy in large language models", "authors": ["Jerry Wei", "Da Huang", "Yifeng Lu", "Denny Zhou", "Quoc V Le"], "authorids": ["~Jerry_Wei1", "~Da_Huang2", "~Yifeng_Lu1", "~Denny_Zhou1", "~Quoc_V_Le1"], "authors_source": "OpenReview API", "abstract": "Sycophancy is an undesirable behavior where models tailor their responses to follow a human user's view even when that view is not objectively correct (e.g., adapting liberal views once a user reveals that they are liberal). In this paper, we study the prevalence of sycophancy in language models and propose a simple synthetic-data intervention to reduce this behavior.\n\nFirst, on a set of three sycophancy tasks where models are asked for an opinion on statements with no correct answers (e.g., politics), we observe that both model scaling and instruction tuning significantly increase sycophancy for large language models up to 540B parameters. Second, we extend sycophancy evaluations to simple addition statements that are objectively incorrect, finding that despite knowing that these statements are wrong, language models will still agree with them if the user does as well.\n\nTo reduce sycophancy, we present a straightforward synthetic-data intervention that takes public NLP tasks and encourages models to be robust to user opinions on these tasks. Adding these data in a lightweight finetuning step can significantly reduce sycophantic behavior on held-out prompts. Code for generating synthetic data for intervention can be found at https://anonymous.4open.science/r/sycophancy-intervention-F0D1/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "YI5mqL7K6k", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3235/Reviewer_LnsW"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "The authors aim to address the phenomenon of sycophancy in language models, where models tend to align with user opinions even when they are incorrect or subjective. This phenomenon is observed even for large language models up to 540B parameters. The authors present an approach to mitigate this by using synthetic data in a lightweight fine-tuning process. This synthetic data is generated by reformatting publicly available NLP tasks, intending to decouple truthfulness from user opinions in model responses. The intervention effectively reduces sycophancy, especially in scenarios involving incorrect statements, and shows generalization to models of varying sizes.", "review_text": "The authors aim to address the phenomenon of sycophancy in language models, where models tend to align with user opinions even when they are incorrect or subjective. This phenomenon is observed even for large language models up to 540B parameters. The authors present an approach to mitigate this by using synthetic data in a lightweight fine-tuning process. This synthetic data is generated by reformatting publicly available NLP tasks, intending to decouple truthfulness from user opinions in model responses. The intervention effectively reduces sycophancy, especially in scenarios involving incorrect statements, and shows generalization to models of varying sizes.", "strengths": "1. The paper is well-structured, with a clear explanation of sycophancy, its implications, and how the proposed intervention addresses this problem.\n2.  The fine-tuning process is lightweight, making this approach accessible and adaptable for large-scale language models with limited computational resources.\n3. The intervention's impact is demonstrated with comprehensive results across multiple models and tasks, showing clear reductions in sycophantic responses.", "weaknesses": "1. The sycophancy evaluations are primarily limited to multiple-choice tasks. It would be beneficial to explore if the intervention works in generative settings where response options are more diverse.\n2. The smallest model used (Flan-LLM-8B) did not respond well to the intervention, highlighting a potential limitation in the effectiveness of the approach for smaller models.", "questions": "1. How well does the intervention generalize to generative tasks where the model isn’t limited to choosing from predefined responses?\n2. Is there any evidence to suggest that the intervention could be adapted for smaller models to improve performance?\n3. Could this approach be extended to improve the model’s adherence to factual information when user opinions align with correct statements?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors aim to address the phenomenon of sycophancy in language models, where models tend to align with user opinions even when they are incorrect or subjective. This phenomenon is observed even for large language models up to 540B parameters. The authors present an approach to mitigate this by using synthetic data in a lightweight fine-tuning process. This synthetic data is generated by reformatting publicly available NLP tasks, intending to decouple truthfulness from user opinions in model responses. The intervention effectively reduces sycophancy, especially in scenarios involving incorrect statements, and shows generalization to models of varying sizes.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper is well-structured, with a clear explanation of sycophancy, its implications, and how the proposed intervention addresses this problem.\n2.  The fine-tuning process is lightweight, making this approach accessible and adaptable for large-scale language models with limited computational resources.\n3. The intervention's impact is demonstrated with comprehensive results across multiple models and tasks, showing clear reductions in sycophantic responses.", "weaknesses": "1. The sycophancy evaluations are primarily limited to multiple-choice tasks. It would be beneficial to explore if the intervention works in generative settings where response options are more diverse.\n2. The smallest model used (Flan-LLM-8B) did not respond well to the intervention, highlighting a potential limitation in the effectiveness of the approach for smaller models.", "questions": "1. How well does the intervention generalize to generative tasks where the model isn’t limited to choosing from predefined responses?\n2. Is there any evidence to suggest that the intervention could be adapted for smaller models to improve performance?\n3. Could this approach be extended to improve the model’s adherence to factual information when user opinions align with correct statements?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730721415654}, {"id": "CyIgLc5Jew", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3235/Reviewer_e8Yz"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper studies the prevalence of sycophancy in language models and puts for a synthetic data based approach to reduce sycophantic behavior in LLMs.", "review_text": "This paper studies the prevalence of sycophancy in language models and puts for a synthetic data based approach to reduce sycophantic behavior in LLMs.", "strengths": "- the synthetic data intervention step leverages openly available datasets, as well as a good variety of such datasets at 17 total\n- well fleshed out limitations section, indicating a paper that is grounded it what it purports to provide evidence for.", "weaknesses": "- the set of models that are used for experiments are quite limited.\n- the intervention to reduce sycophancy requires fine-tuning, which may not be feasible for all use-cases. For example, when access to the model is limited by openness or resource constraints.\n- single prompt format in all experiments", "questions": "1. is there any reason why larger or different models were not chosen? For example, the Mixtral models or Llama family?\n2. In cases where fine-tuning is not an option, what are other ways to reduce sycophantic outputs?\n3. were there attempts with different prompt formats, e.g., with some ablation studies? If so, what were the results? If not, why not?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the prevalence of sycophancy in language models and puts for a synthetic data based approach to reduce sycophantic behavior in LLMs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- the synthetic data intervention step leverages openly available datasets, as well as a good variety of such datasets at 17 total\n- well fleshed out limitations section, indicating a paper that is grounded it what it purports to provide evidence for.", "weaknesses": "- the set of models that are used for experiments are quite limited.\n- the intervention to reduce sycophancy requires fine-tuning, which may not be feasible for all use-cases. For example, when access to the model is limited by openness or resource constraints.\n- single prompt format in all experiments", "questions": "1. is there any reason why larger or different models were not chosen? For example, the Mixtral models or Llama family?\n2. In cases where fine-tuning is not an option, what are other ways to reduce sycophantic outputs?\n3. were there attempts with different prompt formats, e.g., with some ablation studies? If so, what were the results? If not, why not?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730665229280}, {"id": "qEQaDgCkiF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3235/Reviewer_AFnr"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper investigates sycophancy in LLMs, where models tend to tailor responses to align with user opinions, even when those statements are incorrect. To measure this behavior, the authors first examine sycophancy on three classification tasks with no definitive correct answers and one simple addition classification task involving incorrect statements. They find that both model scaling and instruction tuning amplify sycophantic tendencies, and sycophancy still exists even when the statements are apparently wrong. To mitigate this issue, the authors propose a synthetic data intervention designed to encourages more robust responses to potentially incorrect user beliefs. They demonstrate that simply finetuning on this dataset can significantly reduce sycophantic responses on held-out prompts. Further evaluations reveal that this intervention does not impact performance on established benchmarks, such as MMLU and Big-Bench Hard, indicating that the approach maintains model accuracy while reducing sycophancy.", "review_text": "The paper investigates sycophancy in LLMs, where models tend to tailor responses to align with user opinions, even when those statements are incorrect. To measure this behavior, the authors first examine sycophancy on three classification tasks with no definitive correct answers and one simple addition classification task involving incorrect statements. They find that both model scaling and instruction tuning amplify sycophantic tendencies, and sycophancy still exists even when the statements are apparently wrong. To mitigate this issue, the authors propose a synthetic data intervention designed to encourages more robust responses to potentially incorrect user beliefs. They demonstrate that simply finetuning on this dataset can significantly reduce sycophantic responses on held-out prompts. Further evaluations reveal that this intervention does not impact performance on established benchmarks, such as MMLU and Big-Bench Hard, indicating that the approach maintains model accuracy while reducing sycophancy.", "strengths": "This work effectively highlights the issue of sycophancy in LLMs, and conducts evaluations across three model sizes—8B, 62B, and 540B. This finding that sycophantic behavior becomes more pronounced as model size increases provides a valuable insight into how scaling influences sycophancy.\n\nThe synthetic data intervention method is straightforward and effective, making the intervention potentially easy to replicate across different models.\n\nThe proposed method is tested on two popular benchmarks, MMLU and Big-Bench Hard, showing the effectiveness without compromising the model’s performance on established tasks or affecting the existing learned knowledge.", "weaknesses": "While the paper offers insights about sycophancy in language models and a method for reducing it, further experiments could enhance the robustness and generalizability of the proposed finetuning method:\n1. Although three models of varying sizes were tested, the evaluation is limited to a single model type. It would be beneficial to examine sycophancy across a wider range of both open-source LLMs, such as LLaMA -- which has been widely studied in research and also offers multiple size options -- and closed-source models like GPT or Gemini. Expanding the evaluation to diverse model architectures would also help verify whether the synthetic data intervention remains effective across different types of models.\n2. The paper acknowledges limitations in prompt diversity and task scope, as the experiments are largely restricted to classification tasks with a narrow range of prompt formats. Testing the fine-tuned model’s performance on more varied, out-of-distribution samples would provide a stronger assessment of the intervention’s generalizability. For example, the user's belief is subtly embedded within a description of their actions, serving as contextual information.\n3. In Section 3, the study on objectively incorrect statements is limited to simple addition tasks, which constrains the ability to generalize the claim that “models are sycophantic for objectively wrong answers.” Expanding this evaluation to include a broader array of tasks involving objectively false statements—such as those related to misinformation (e.g., \"Apple cider vinegar is a miracle cure for cancer\"), fake news (e.g., a famous celebrity supports xxx but actually not), or conspiracy theories (e.g., \"the Earth is flat\")—would lend stronger support to this conclusion.\n\nWriting clarity:\nTable 1 and Figure 1 are nearly duplicated.", "questions": "What proportion of questions is filtered out for each model during the filtration process? Given the differences in model sizes, I suppose there would be a trend where larger models retain a greater number of synthesized training examples.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates sycophancy in LLMs, where models tend to tailor responses to align with user opinions, even when those statements are incorrect. To measure this behavior, the authors first examine sycophancy on three classification tasks with no definitive correct answers and one simple addition classification task involving incorrect statements. They find that both model scaling and instruction tuning amplify sycophantic tendencies, and sycophancy still exists even when the statements are apparently wrong. To mitigate this issue, the authors propose a synthetic data intervention designed to encourages more robust responses to potentially incorrect user beliefs. They demonstrate that simply finetuning on this dataset can significantly reduce sycophantic responses on held-out prompts. Further evaluations reveal that this intervention does not impact performance on established benchmarks, such as MMLU and Big-Bench Hard, indicating that the approach maintains model accuracy while reducing sycophancy.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "This work effectively highlights the issue of sycophancy in LLMs, and conducts evaluations across three model sizes—8B, 62B, and 540B. This finding that sycophantic behavior becomes more pronounced as model size increases provides a valuable insight into how scaling influences sycophancy.\n\nThe synthetic data intervention method is straightforward and effective, making the intervention potentially easy to replicate across different models.\n\nThe proposed method is tested on two popular benchmarks, MMLU and Big-Bench Hard, showing the effectiveness without compromising the model’s performance on established tasks or affecting the existing learned knowledge.", "weaknesses": "While the paper offers insights about sycophancy in language models and a method for reducing it, further experiments could enhance the robustness and generalizability of the proposed finetuning method:\n1. Although three models of varying sizes were tested, the evaluation is limited to a single model type. It would be beneficial to examine sycophancy across a wider range of both open-source LLMs, such as LLaMA -- which has been widely studied in research and also offers multiple size options -- and closed-source models like GPT or Gemini. Expanding the evaluation to diverse model architectures would also help verify whether the synthetic data intervention remains effective across different types of models.\n2. The paper acknowledges limitations in prompt diversity and task scope, as the experiments are largely restricted to classification tasks with a narrow range of prompt formats. Testing the fine-tuned model’s performance on more varied, out-of-distribution samples would provide a stronger assessment of the intervention’s generalizability. For example, the user's belief is subtly embedded within a description of their actions, serving as contextual information.\n3. In Section 3, the study on objectively incorrect statements is limited to simple addition tasks, which constrains the ability to generalize the claim that “models are sycophantic for objectively wrong answers.” Expanding this evaluation to include a broader array of tasks involving objectively false statements—such as those related to misinformation (e.g., \"Apple cider vinegar is a miracle cure for cancer\"), fake news (e.g., a famous celebrity supports xxx but actually not), or conspiracy theories (e.g., \"the Earth is flat\")—would lend stronger support to this conclusion.\n\nWriting clarity:\nTable 1 and Figure 1 are nearly duplicated.", "questions": "What proportion of questions is filtered out for each model during the filtration process? Given the differences in model sizes, I suppose there would be a trend where larger models retain a greater number of synthesized training examples.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730615134560}, {"id": "cFqQaSyF42", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3235/Reviewer_zVCa"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper investigates the prevalence of sycophancy in LLMs and discovers that both model scaling and instruction-tuning can exacerbate this issue. In response, the authors propose a straightforward synthetic data construction method designed to mitigate sycophancy in LLMs. Experimental results demonstrate that their training approach effectively generalizes to prompts that the models have not seen during training.", "review_text": "This paper investigates the prevalence of sycophancy in LLMs and discovers that both model scaling and instruction-tuning can exacerbate this issue. In response, the authors propose a straightforward synthetic data construction method designed to mitigate sycophancy in LLMs. Experimental results demonstrate that their training approach effectively generalizes to prompts that the models have not seen during training.", "strengths": "**Clarity:** The paper is very well-written and easy to follow, with a smooth flow that makes it enjoyable to read. The figures are clear and effectively help in understanding the problems being addressed and the approach employed.\n\n**Motivation:** The paper is well-motivated, addressing the critical issue of sycophancy in LLMs. This phenomenon poses a significant challenge, as it resembles reward hacking and may undermine the effectiveness of the RLHF process, potentially hindering the further development of LLMs.\n\n**Implications:** The paper highlights that both model scaling and instruction tuning can lead to increased sycophancy in LLMs. These important findings serve as a reminder of the problem's significance.\n\n**Soundness:** The proposed approach is logically sound, and the experimental results convincingly demonstrate the effectiveness of their training method, particularly in generalizing to prompts that the models have not seen during training.", "weaknesses": "**Comparison:** It would be helpful to compare the model's performance using alternative prompting techniques, such as system 2 attention [1]. If these prompting methods effectively reduce sycophancy, it could challenge the applicability of the current training approach, which appears to require more efforts. Additionally, it remains unclear how the current approach would generalize to scenarios involving more open-ended questions where there are no set of gold answers to choose from. In comparison, prompting techniques could generalize to those open-ended questions more easily.\n\n\n\n[1] System 2 Attention (is something you might need too)", "questions": "See Weaknesses part. Could you add a comparison to some prompting-based techniques to see whether the training based approach works better or not?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the prevalence of sycophancy in LLMs and discovers that both model scaling and instruction-tuning can exacerbate this issue. In response, the authors propose a straightforward synthetic data construction method designed to mitigate sycophancy in LLMs. Experimental results demonstrate that their training approach effectively generalizes to prompts that the models have not seen during training.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "**Clarity:** The paper is very well-written and easy to follow, with a smooth flow that makes it enjoyable to read. The figures are clear and effectively help in understanding the problems being addressed and the approach employed.\n\n**Motivation:** The paper is well-motivated, addressing the critical issue of sycophancy in LLMs. This phenomenon poses a significant challenge, as it resembles reward hacking and may undermine the effectiveness of the RLHF process, potentially hindering the further development of LLMs.\n\n**Implications:** The paper highlights that both model scaling and instruction tuning can lead to increased sycophancy in LLMs. These important findings serve as a reminder of the problem's significance.\n\n**Soundness:** The proposed approach is logically sound, and the experimental results convincingly demonstrate the effectiveness of their training method, particularly in generalizing to prompts that the models have not seen during training.", "weaknesses": "**Comparison:** It would be helpful to compare the model's performance using alternative prompting techniques, such as system 2 attention [1]. If these prompting methods effectively reduce sycophancy, it could challenge the applicability of the current training approach, which appears to require more efforts. Additionally, it remains unclear how the current approach would generalize to scenarios involving more open-ended questions where there are no set of gold answers to choose from. In comparison, prompting techniques could generalize to those open-ended questions more easily.\n\n\n\n[1] System 2 Attention (is something you might need too)", "questions": "See Weaknesses part. Could you add a comparison to some prompting-based techniques to see whether the training based approach works better or not?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730514305794}], "openreview_url": "https://openreview.net/forum?id=WDheQxWAo4", "arxiv_id": "2308.03958", "paper_pdf": "papers/WDheQxWAo4.pdf", "paper_pdf_sha256": "2cf4859df7cc557c2e3f823fbc909e9a25d61aed44635bee906329ecc22076ec", "paper_pdf_bytes": 426892, "paper_pdf_source": "openreview", "code_url": "https://github.com/google/sycophancy-intervention", "code_repository": "google/sycophancy-intervention", "code_commit": "fb75986c759abdff7fe4d4b09fd7102383dc71aa", "code_archive": "repos/WDheQxWAo4.zip", "code_archive_sha256": "45c7d65e00424e451f378cf5f63a43fd742a94d0d48a34c3d940f01bd17e5465", "code_archive_bytes": 61856, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 60, "github_languages": {"Python": 15465}, "github_archived": true, "github_pushed_at": "2023-08-16T17:34:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/simple-synthetic-data-reduces-sycophancy-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ro4CgvfUKy", "year": 2024, "status": "rejected", "title": "Latent Noise Segmentation: How Neural Noise Leads to the Emergence of Segmentation and Grouping", "authors": ["Ben Lonnqvist", "Frank Zhengqing Wu", "Michael Herzog"], "authorids": ["~Ben_Lonnqvist1", "~Frank_Zhengqing_Wu1", "~Michael_Herzog1"], "authors_source": "OpenReview API", "abstract": "Deep Neural Networks (DNNs) that achieve human-level performance in general tasks like object segmentation typically require supervised labels. In contrast, humans are able to perform these tasks effortlessly without supervision.\nTo accomplish this, the human visual system makes use of perceptual grouping: for example, the black and white stripes of a zebra are perceptually grouped together despite their vastly different colors.\nUnderstanding how perceptual grouping arises in an unsupervised manner is critical for improving both models of the visual system, and computer vision models. In this work, we propose a counterintuitive approach to unsupervised perceptual grouping and segmentation: that they arise because of neural noise, rather than in spite of it. We (1) mathematically demonstrate that under realistic assumptions, neural noise can be used to separate objects from each other, and (2) show that adding noise in a DNN enables the network to segment images even though it was never trained on any segmentation labels. Interestingly, we find that (3) segmenting objects using noise results in segmentation performance that aligns with the perceptual grouping phenomena observed in humans. We introduce the Good Gestalt (GG) datasets --- six datasets designed to specifically test perceptual grouping, and show that our DNN models reproduce many important phenomena in human perception, such as illusory contours, closure, continuity, proximity, and occlusion. Finally, we (4) demonstrate the ecological plausibility of the method by analyzing the sensitivity of the DNN to different magnitudes of noise. We find that some model variants consistently succeed with remarkably low levels of neural noise ($\\sigma<0.001$), and surprisingly, that segmenting this way requires as few as a handful of samples. Together, our results suggest a novel unsupervised segmentation method requiring few assumptions, a new explanation for the formation of perceptual grouping, and a potential benefit of neural noise in the visual system.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "pTcpg2JIlQ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7243/Reviewer_DELR"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper studies the role of neural noise in the formation of perceptual groups. They show how one can obtain segmentation maps from a (V)AE simply through the injection of noise in the latent space, without any supervision on a segmentation task. Concretely, N noisy versions of a latent vector are passed through the decoder to result in N slightly different outputs. The difference maps between two consecutive outputs are then turned into one segmentation map using a clustering algorithm. The paper posits that the reason this process reveals perceptual groups in a scene is because pixels belonging to the same group tend to co-vary. Indeed, an experiment using a novel dataset, Good Gestalt, reveals cognitively viable segmentation maps that seem to obey the Gestalt laws. The appendix also includes an experiment on natural images (CelebA). In the paper's conclusion, neural noise is put forward as a potential mechanism for perceptual grouping.", "review_text": "This paper studies the role of neural noise in the formation of perceptual groups. They show how one can obtain segmentation maps from a (V)AE simply through the injection of noise in the latent space, without any supervision on a segmentation task. Concretely, N noisy versions of a latent vector are passed through the decoder to result in N slightly different outputs. The difference maps between two consecutive outputs are then turned into one segmentation map using a clustering algorithm. The paper posits that the reason this process reveals perceptual groups in a scene is because pixels belonging to the same group tend to co-vary. Indeed, an experiment using a novel dataset, Good Gestalt, reveals cognitively viable segmentation maps that seem to obey the Gestalt laws. The appendix also includes an experiment on natural images (CelebA). In the paper's conclusion, neural noise is put forward as a potential mechanism for perceptual grouping.", "strengths": "Asking how perceptual grouping may occur without explicit supervision is important considering how many modern models rely on such supervision, whereas our own visual system arguably handles it differently. Moreover, I think the papers takes an interesting and fresh take on it by studying how noise might in fact be beneficial to visually separate objects. \n\nThe extent of the analyses (mathematical accounts, extra results on CelebA, noise sensitivity analysis) etc. is impressive. More than once I wrote down something I intended to inquire about, only to see that exact question already addressed a bit further down the paper.\n\nIt's a well-prepared manuscript, written with care.", "weaknesses": "The potential weaknesses I have spotted could very well rather be unclarities, so I'll save it for the \"Questions\" section.", "questions": "1)\nI'm unclear on the extent to which Latent Noise Segmentation is a method to reveal the perceptual groups already formed inside the network through some mechanism or another, versus a mechanism that gives rise to perceptual groups in its own right. Is the hypothesis that LNS is a way of \"doing\" segmentation or is it a way to \"show\" the result of segmentation, if that makes sense?\n\n2)\nThe training samples in the GG dataset often combine two Gestalt cues. For example, in Proximity, the parts closest to each other are also similar in color. Was that crucial to the results? Would the segmentations maps no longer obey the law of Proximity if there was no color cue during training?\n\nIn Section 2, does pretraining refer to training on GG before doing the noise injection, or was there any pretraining on natural images? If so, it would be interesting to see segmentation maps for GG test images before training on GG train images. Would it group by continuity just by learning from natural statistics, for example?\n\nI think it might be worthwhile to show examples of VAE outputs without noise (i.e., the actual reconstructions, not the segmentation maps). If it groups the pixels of the Kanizsa squares together supposedly \"perceives\" the square, does it output its illusionary contours? \n\n3) \n\"Time steps\" is used to refer to the number of noisy samples needed for a segmentation map. Does it bear any relation with actual time in the human visual system? \n\n4) \nThe paper refers to ecological validity and biological plausibility, but I'd love to see a little more elaboration on how exactly a biological visual system would potentially carry out the operations suggested here (e.g., how can we picture the 'clustering' being done)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the role of neural noise in the formation of perceptual groups. They show how one can obtain segmentation maps from a (V)AE simply through the injection of noise in the latent space, without any supervision on a segmentation task. Concretely, N noisy versions of a latent vector are passed through the decoder to result in N slightly different outputs. The difference maps between two consecutive outputs are then turned into one segmentation map using a clustering algorithm. The paper posits that the reason this process reveals perceptual groups in a scene is because pixels belonging to the same group tend to co-vary. Indeed, an experiment using a novel dataset, Good Gestalt, reveals cognitively viable segmentation maps that seem to obey the Gestalt laws. The appendix also includes an experiment on natural images (CelebA). In the paper's conclusion, neural noise is put forward as a potential mechanism for perceptual grouping.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Asking how perceptual grouping may occur without explicit supervision is important considering how many modern models rely on such supervision, whereas our own visual system arguably handles it differently. Moreover, I think the papers takes an interesting and fresh take on it by studying how noise might in fact be beneficial to visually separate objects. \n\nThe extent of the analyses (mathematical accounts, extra results on CelebA, noise sensitivity analysis) etc. is impressive. More than once I wrote down something I intended to inquire about, only to see that exact question already addressed a bit further down the paper.\n\nIt's a well-prepared manuscript, written with care.", "weaknesses": "The potential weaknesses I have spotted could very well rather be unclarities, so I'll save it for the \"Questions\" section.", "questions": "1)\nI'm unclear on the extent to which Latent Noise Segmentation is a method to reveal the perceptual groups already formed inside the network through some mechanism or another, versus a mechanism that gives rise to perceptual groups in its own right. Is the hypothesis that LNS is a way of \"doing\" segmentation or is it a way to \"show\" the result of segmentation, if that makes sense?\n\n2)\nThe training samples in the GG dataset often combine two Gestalt cues. For example, in Proximity, the parts closest to each other are also similar in color. Was that crucial to the results? Would the segmentations maps no longer obey the law of Proximity if there was no color cue during training?\n\nIn Section 2, does pretraining refer to training on GG before doing the noise injection, or was there any pretraining on natural images? If so, it would be interesting to see segmentation maps for GG test images before training on GG train images. Would it group by continuity just by learning from natural statistics, for example?\n\nI think it might be worthwhile to show examples of VAE outputs without noise (i.e., the actual reconstructions, not the segmentation maps). If it groups the pixels of the Kanizsa squares together supposedly \"perceives\" the square, does it output its illusionary contours? \n\n3) \n\"Time steps\" is used to refer to the number of noisy samples needed for a segmentation map. Does it bear any relation with actual time in the human visual system? \n\n4) \nThe paper refers to ecological validity and biological plausibility, but I'd love to see a little more elaboration on how exactly a biological visual system would potentially carry out the operations suggested here (e.g., how can we picture the 'clustering' being done)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699137079106}, {"id": "z3hOttR9Hs", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7243/Reviewer_zkrR"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors demonstrate that segmentation and grouping features emerge unsupervisedly by injecting iid noise in the latent space of VAE and AE. They show this by designing a simple algorithm that compute relative differences of reconstructed images from latent code corrupted by noise on top of which they stack et simple clustering algorithm. In addition, the author have built a large dataset consisting of different grouping/segmentation tasks corresponding to various Gestalt aspects of perception such as closure, continuity, proximity, etc.", "review_text": "The authors demonstrate that segmentation and grouping features emerge unsupervisedly by injecting iid noise in the latent space of VAE and AE. They show this by designing a simple algorithm that compute relative differences of reconstructed images from latent code corrupted by noise on top of which they stack et simple clustering algorithm. In addition, the author have built a large dataset consisting of different grouping/segmentation tasks corresponding to various Gestalt aspects of perception such as closure, continuity, proximity, etc.", "strengths": "- well-grounded in the vision science field, enough references\n- the idea is well motivated by the search of a role for neural noise and tested in artificial neural network\n- the performances of the proposed method are extensively tested on a relevant dataset that is build for this purpose (that one counts twice)\n- comparison of VAE and AE is provided together with a control of the idea of adding noise in the latent space\n- amount of added noise and step required in the algorithm are also evaluated", "weaknesses": "**Minor weaknesses** \n- The role of the post-processing step is not evaluated : does agglomerative clustering play a big role ? There are other standard clustering methods that could be tested.\n- Even if it's not designed for the segmentation of natural images and if it's likely to not perform very well compared to SOTA algorithms it is worth testing it. I have in mind a recent paper (Vacher et al 2022) in which deep neural network features are evaluated for segmenting natural images.\n- Latent space of VAE are known to enable appealing morphing between natural images (by linearly interpolating the latent code) so I am wondering what would be the segmentation related uncertainty that could be obtained with this method ... I guess this would require a more involved post-processing step.\n- Other neural network architectures are not tested (GANs, Normalizing flows, ...) \n\n**Minor remark**\n- In table 1, bold should be used for every best performing model, eg also for Continuity and Gradient Occ.\n\n\nRefs : \n- Vacher, J., Launay, C., & Coen-Cagli, R. (2022). Flexibly regularized mixture models and application to image segmentation. Neural Networks, 149, 107-123.\n\n\n**Post-response update**\nI thank the authors for their response and considering it together with the other reviews, I am increasing my score so that it is now a clear accept.", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors demonstrate that segmentation and grouping features emerge unsupervisedly by injecting iid noise in the latent space of VAE and AE. They show this by designing a simple algorithm that compute relative differences of reconstructed images from latent code corrupted by noise on top of which they stack et simple clustering algorithm. In addition, the author have built a large dataset consisting of different grouping/segmentation tasks corresponding to various Gestalt aspects of perception such as closure, continuity, proximity, etc.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "- well-grounded in the vision science field, enough references\n- the idea is well motivated by the search of a role for neural noise and tested in artificial neural network\n- the performances of the proposed method are extensively tested on a relevant dataset that is build for this purpose (that one counts twice)\n- comparison of VAE and AE is provided together with a control of the idea of adding noise in the latent space\n- amount of added noise and step required in the algorithm are also evaluated", "weaknesses": "**Minor weaknesses** \n- The role of the post-processing step is not evaluated : does agglomerative clustering play a big role ? There are other standard clustering methods that could be tested.\n- Even if it's not designed for the segmentation of natural images and if it's likely to not perform very well compared to SOTA algorithms it is worth testing it. I have in mind a recent paper (Vacher et al 2022) in which deep neural network features are evaluated for segmenting natural images.\n- Latent space of VAE are known to enable appealing morphing between natural images (by linearly interpolating the latent code) so I am wondering what would be the segmentation related uncertainty that could be obtained with this method ... I guess this would require a more involved post-processing step.\n- Other neural network architectures are not tested (GANs, Normalizing flows, ...) \n\n**Minor remark**\n- In table 1, bold should be used for every best performing model, eg also for Continuity and Gradient Occ.\n\n\nRefs : \n- Vacher, J., Launay, C., & Coen-Cagli, R. (2022). Flexibly regularized mixture models and application to image segmentation. Neural Networks, 149, 107-123.\n\n\n**Post-response update**\nI thank the authors for their response and considering it together with the other reviews, I am increasing my score so that it is now a clear accept.", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699019521086}, {"id": "j71w36y9gv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7243/Reviewer_ipWu"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper investigates noise in feature space as a means of segmenting images. Several datasets of simple artificial images are developed for this work, each of which is designed to test whether the method exhibits a different Gestalt property. Autoencoders are used to produce latent representations. Small independent noise is repeatedly added to the latent representations. Differences are calculated between pairs of noisy outputs to produce pixel-wise vectors of differences. These vectors are clustered to produce a segmentation. This often results in Gestalt-like segmentations, e.g. segmentation of Kanizsa squares from background.", "review_text": "This paper investigates noise in feature space as a means of segmenting images. Several datasets of simple artificial images are developed for this work, each of which is designed to test whether the method exhibits a different Gestalt property. Autoencoders are used to produce latent representations. Small independent noise is repeatedly added to the latent representations. Differences are calculated between pairs of noisy outputs to produce pixel-wise vectors of differences. These vectors are clustered to produce a segmentation. This often results in Gestalt-like segmentations, e.g. segmentation of Kanizsa squares from background.", "strengths": "The dataset is a nice contribution, particularly the test set with expected segmentations for images that are expected to elicit various Gestalt phenomena. \n\nThe segmentation method is novel (as far as I know) and creative.", "weaknesses": "The abstract claims that the results show, “potential benefit of neural noise in the visual system” and the paper repeatedly claims to study the ecologically plausibility of the method. Noise is certainly prevalent in biological vision, but I'm not convinced that there is a substantial connection between this method and biology. Issues include: 1) reliance of the method on small-amplitude noise, whereas spiking noise is closer to Poisson; 2) use of agglomerative clustering; 3) lack of comparison with neural data; 4) need for specialized training datasets to produce Gestalt phenomena; 5) poor performance on natural images. \n\nIf the goal is to explain something about biological visual systems, I think much clearer links to biology are needed. It seems to me that this would require substantial changes to the model and/or much more detailed justification of multiple model elements. \n\nTo elaborate on point 4 above, the method seems to rely on the design of the training datasets to work properly. For example, to segment Kanizsa squares, an autoencoder is first trained on stimuli that show the squares in a different color than the background color. This kind of dependence is claimed explicitly in the appendix: “If p1 and p2 are pixels belonging to the same object, the way the dataset is generated … dictates that the training samples projected onto the pixel value space will only stretch in a direction where there exists a strict linear ratio between the values of the two objects.” Humans don't require such specialized training to experience Gestalt perception.  \n\nTo elaborate on point 5 above, the method is not meant to be practical, but its low practicality is also a concern for biological plausibility. It is not extensively tested on natural or otherwise practical images but in addition to the Gestalt dataset images, it was tested on celebrity faces and the paper claims that the method “often finds a semantically meaningful segmentation of face-hair-background”. However, examination of the results (Appendix A.3) shows that the results are generally poor. According to Table 1, even in these artificial circumstances, the model only outperforms the control in 4/6 cases. The control is the same model with noise applied to the output rather than to the latent representation. It could be interesting to also contrast with established strong segmentation methods, particularly if this method agreed with humans in conditions where others don’t. However, some promise of strong performance would be needed in an convincing model of human vision.", "questions": "What size are the GG datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates noise in feature space as a means of segmenting images. Several datasets of simple artificial images are developed for this work, each of which is designed to test whether the method exhibits a different Gestalt property. Autoencoders are used to produce latent representations. Small independent noise is repeatedly added to the latent representations. Differences are calculated between pairs of noisy outputs to produce pixel-wise vectors of differences. These vectors are clustered to produce a segmentation. This often results in Gestalt-like segmentations, e.g. segmentation of Kanizsa squares from background.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "The dataset is a nice contribution, particularly the test set with expected segmentations for images that are expected to elicit various Gestalt phenomena. \n\nThe segmentation method is novel (as far as I know) and creative.", "weaknesses": "The abstract claims that the results show, “potential benefit of neural noise in the visual system” and the paper repeatedly claims to study the ecologically plausibility of the method. Noise is certainly prevalent in biological vision, but I'm not convinced that there is a substantial connection between this method and biology. Issues include: 1) reliance of the method on small-amplitude noise, whereas spiking noise is closer to Poisson; 2) use of agglomerative clustering; 3) lack of comparison with neural data; 4) need for specialized training datasets to produce Gestalt phenomena; 5) poor performance on natural images. \n\nIf the goal is to explain something about biological visual systems, I think much clearer links to biology are needed. It seems to me that this would require substantial changes to the model and/or much more detailed justification of multiple model elements. \n\nTo elaborate on point 4 above, the method seems to rely on the design of the training datasets to work properly. For example, to segment Kanizsa squares, an autoencoder is first trained on stimuli that show the squares in a different color than the background color. This kind of dependence is claimed explicitly in the appendix: “If p1 and p2 are pixels belonging to the same object, the way the dataset is generated … dictates that the training samples projected onto the pixel value space will only stretch in a direction where there exists a strict linear ratio between the values of the two objects.” Humans don't require such specialized training to experience Gestalt perception.  \n\nTo elaborate on point 5 above, the method is not meant to be practical, but its low practicality is also a concern for biological plausibility. It is not extensively tested on natural or otherwise practical images but in addition to the Gestalt dataset images, it was tested on celebrity faces and the paper claims that the method “often finds a semantically meaningful segmentation of face-hair-background”. However, examination of the results (Appendix A.3) shows that the results are generally poor. According to Table 1, even in these artificial circumstances, the model only outperforms the control in 4/6 cases. The control is the same model with noise applied to the output rather than to the latent representation. It could be interesting to also contrast with established strong segmentation methods, particularly if this method agreed with humans in conditions where others don’t. However, some promise of strong performance would be needed in an convincing model of human vision.", "questions": "What size are the GG datasets?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698812211614}, {"id": "tQFPrczjvM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7243/Reviewer_9tYc"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose a new unsupervised object discovery technique based on variational autoencoders with noise injected to the latent representation. The authors propose an unsupervised object segmentation algorithm that works by computing the difference between reconstructions of an input image from noisy latent representations followed by pixel-wise clustering. The proposed approach is tested on unsupervised perceptual grouping performance on the Good Gestalt (GG) dataset that the authors propose in this submission as well. Quantitative evaluation performed on the GG dataset shows promising evidence of Latent Noise Segmentation discovering objects while being trained on unsupervised image reconstruction. Further analyses on the sensitivity of LNS to latent noise parameters helps better understand the working of VAEs equipped with latent noise segmentation.", "review_text": "The authors propose a new unsupervised object discovery technique based on variational autoencoders with noise injected to the latent representation. The authors propose an unsupervised object segmentation algorithm that works by computing the difference between reconstructions of an input image from noisy latent representations followed by pixel-wise clustering. The proposed approach is tested on unsupervised perceptual grouping performance on the Good Gestalt (GG) dataset that the authors propose in this submission as well. Quantitative evaluation performed on the GG dataset shows promising evidence of Latent Noise Segmentation discovering objects while being trained on unsupervised image reconstruction. Further analyses on the sensitivity of LNS to latent noise parameters helps better understand the working of VAEs equipped with latent noise segmentation.", "strengths": "+ Very intriguing to see emergent perceptual grouping from ANNs trained on unsupervised image reconstruction objectives. Super interesting perspective that latent neuronal noise could lead to learning good gestalt priors.\n+ The approach is very straightforward and easy to understand. I think this simplicity is a big strength of the proposed work.\n+ Good Gestalt datasets are also a nice addition to the contributions from this work, I hope this benchmark can serve as a good way to measure perceptual grouping abilities of models in the community. \n+ I like the additional analyses performed on understanding how latent noise parameters affects emergent grouping.", "weaknesses": "- The focus of the paper feels a bit narrow in terms of the architecture / learning objective. Is there something special about VAEs trained with ELBO + LNS that makes them develop emergent grouping, or does LNS generalize across architectural choices and learning rules (say, diffusion-based or adversarial generative models)? Adding a discussion on this would add more value to the submission\n- I feel that GG's difficulty could be significantly improved by adding more distractors and/or noise to the background of images. Although the current emergence of grouping looks interesting, I would be even more surprised if the model is learning to discount background noise in its presence, i.e., currently the dataset makes figure-ground organization too simple by providing a largely low-frequency background and I believe GG can be solved merely by using simple rules on low-level feature detectors. \n- The authors have covered a variety of unsupervised object discovery approaches such as Slot Attention, Complex-valued autoencoders in related work but have not performed a direct comparison to these baselines in their reported experiments. This makes the paper weaker due to the absence of relevant baselines other than the VAE-based ones currently reported in this version of the paper.\n\nOverall, the strengths of this work marginally outweigh the weaknesses mentioned above, I would be inclined to further improving my score provided the authors convincingly rebut my concerns about this work.", "questions": "- Can the authors please comment on whether they experimented with harder versions of GG? Here are a few potential options: (1) Change the size of the square/circle between train and test splits for Kanisza Squares, Closure, Continuity, (2) Use different rotations (without overlap between train and test splits) of the Kanisza squares / Closure squares, (3) Use different colors / textures as backgrounds in all tasks. Any modification of the dataset in the spirit of the modifications suggested here will further strengthen the message that simple low-level statistics don't drive the emergence of perceptual grouping.\n- Adding stronger baselines such as the ones the authors have mentioned in related work will help improve my score further.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a new unsupervised object discovery technique based on variational autoencoders with noise injected to the latent representation. The authors propose an unsupervised object segmentation algorithm that works by computing the difference between reconstructions of an input image from noisy latent representations followed by pixel-wise clustering. The proposed approach is tested on unsupervised perceptual grouping performance on the Good Gestalt (GG) dataset that the authors propose in this submission as well. Quantitative evaluation performed on the GG dataset shows promising evidence of Latent Noise Segmentation discovering objects while being trained on unsupervised image reconstruction. Further analyses on the sensitivity of LNS to latent noise parameters helps better understand the working of VAEs equipped with latent noise segmentation.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "+ Very intriguing to see emergent perceptual grouping from ANNs trained on unsupervised image reconstruction objectives. Super interesting perspective that latent neuronal noise could lead to learning good gestalt priors.\n+ The approach is very straightforward and easy to understand. I think this simplicity is a big strength of the proposed work.\n+ Good Gestalt datasets are also a nice addition to the contributions from this work, I hope this benchmark can serve as a good way to measure perceptual grouping abilities of models in the community. \n+ I like the additional analyses performed on understanding how latent noise parameters affects emergent grouping.", "weaknesses": "- The focus of the paper feels a bit narrow in terms of the architecture / learning objective. Is there something special about VAEs trained with ELBO + LNS that makes them develop emergent grouping, or does LNS generalize across architectural choices and learning rules (say, diffusion-based or adversarial generative models)? Adding a discussion on this would add more value to the submission\n- I feel that GG's difficulty could be significantly improved by adding more distractors and/or noise to the background of images. Although the current emergence of grouping looks interesting, I would be even more surprised if the model is learning to discount background noise in its presence, i.e., currently the dataset makes figure-ground organization too simple by providing a largely low-frequency background and I believe GG can be solved merely by using simple rules on low-level feature detectors. \n- The authors have covered a variety of unsupervised object discovery approaches such as Slot Attention, Complex-valued autoencoders in related work but have not performed a direct comparison to these baselines in their reported experiments. This makes the paper weaker due to the absence of relevant baselines other than the VAE-based ones currently reported in this version of the paper.\n\nOverall, the strengths of this work marginally outweigh the weaknesses mentioned above, I would be inclined to further improving my score provided the authors convincingly rebut my concerns about this work.", "questions": "- Can the authors please comment on whether they experimented with harder versions of GG? Here are a few potential options: (1) Change the size of the square/circle between train and test splits for Kanisza Squares, Closure, Continuity, (2) Use different rotations (without overlap between train and test splits) of the Kanisza squares / Closure squares, (3) Use different colors / textures as backgrounds in all tasks. Any modification of the dataset in the spirit of the modifications suggested here will further strengthen the message that simple low-level statistics don't drive the emergence of perceptual grouping.\n- Adding stronger baselines such as the ones the authors have mentioned in related work will help improve my score further.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698797529099}, {"id": "xl8tfAlcki", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7243/Reviewer_cWdJ"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this manuscript the authors present a method to segment images by adding noise to the latent of a (variational) autoencoder and clustering based on the differences between reconstructions in pixel space. In a couple of controlled experiments on grouping stimuli, this method shows the groupings that are shown by humans.", "review_text": "In this manuscript the authors present a method to segment images by adding noise to the latent of a (variational) autoencoder and clustering based on the differences between reconstructions in pixel space. In a couple of controlled experiments on grouping stimuli, this method shows the groupings that are shown by humans.", "strengths": "It is an interesting observation that autoencoder objectives alone lead to a representation that separates objects in a similar way as grouping experiments in humans suggest. Also, I think it is an interesting hypothesis that noise at higher levels is translated into correlated noise at lower levels that supports segmentations into objects. The authors present their novel set of gestalt tests that test for the expected grouping results explicitly while being properly image computable.", "weaknesses": "I am not convinced by this paper for three reasons:\n\nFirst, the main evaluation is done on a self made grouping stimulus set, which trains the network quite explicitly to produce the grouping made by human observers as the intended objects are exactly the groups that vary separately in the stimulus generation, by switching color together for example. While this is clearly not direct supervision, it does create a statistical structure that strongly favours the representation of object centred dimensions while there are no variations within objects or over the whole scene. This makes it a lot less surprising that auto encoders find dimensions that create correlations within objects.\n\nSecond, the authors emphasise noise strongly in their arguments and go to some length to explain how iid noise might emphasise the local PCs around the stimulus. While I do not think these arguments are technically incorrect, I think they are besides the point. The main step to make this technique work seems to be the shape of the derivative of the decoder around the stimulus. The decoder seems to create correlated changes in the parts that belong to the same object, which yields high similarity for those pixels. It is interesting that this effect is strong enough that few samples are sufficient to separate objects successfully, but in principle any way of estimating this derivative should work. As the noise size here corresponds to the typical delta used to compute the approximate derivative, it is also not surprising that small noise works well. The interesting part is that the derivative creates correlated changes for each object, not that this can be estimated based on noise samples.\n\nThird, I think the connection to human or biological vision is weaker than suggested by the authors. In biological vision we need a segmentation of the internal layer representations, not of the pixels in the image. This requires that the encoder and decoder are in some way related that allows us to connect our noise reconstructions and the encoding elements. Additionally, we would need a biologically plausible clustering algorithm that is based on the found similarities in the noise. Both are not present in the model presented here. Thus, substantial revisions would be necessary to transform the method proposed here into a biologically plausible method for object segmentation.", "questions": "Perhaps my questions highlight that this manuscript seems not ideally placed at a machine learning conference:\n- I would like to understand how the authors imagine the noise based segmentation to work in biological vision: How does the autoencoder model map to the brain? And what evidence is there that anything like this might actually happen in the brain?\n- Does this work in any way for natural scenes? Testing this on existing autoencoders for natural scenes could avoid my concerns about the training data being very targeted to create the patterns observed in humans.\n- And what about comparisons to alternative methods? Do other methods for segmentation or grouping get the gestalt tests wrong?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this manuscript the authors present a method to segment images by adding noise to the latent of a (variational) autoencoder and clustering based on the differences between reconstructions in pixel space. In a couple of controlled experiments on grouping stimuli, this method shows the groupings that are shown by humans.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "It is an interesting observation that autoencoder objectives alone lead to a representation that separates objects in a similar way as grouping experiments in humans suggest. Also, I think it is an interesting hypothesis that noise at higher levels is translated into correlated noise at lower levels that supports segmentations into objects. The authors present their novel set of gestalt tests that test for the expected grouping results explicitly while being properly image computable.", "weaknesses": "I am not convinced by this paper for three reasons:\n\nFirst, the main evaluation is done on a self made grouping stimulus set, which trains the network quite explicitly to produce the grouping made by human observers as the intended objects are exactly the groups that vary separately in the stimulus generation, by switching color together for example. While this is clearly not direct supervision, it does create a statistical structure that strongly favours the representation of object centred dimensions while there are no variations within objects or over the whole scene. This makes it a lot less surprising that auto encoders find dimensions that create correlations within objects.\n\nSecond, the authors emphasise noise strongly in their arguments and go to some length to explain how iid noise might emphasise the local PCs around the stimulus. While I do not think these arguments are technically incorrect, I think they are besides the point. The main step to make this technique work seems to be the shape of the derivative of the decoder around the stimulus. The decoder seems to create correlated changes in the parts that belong to the same object, which yields high similarity for those pixels. It is interesting that this effect is strong enough that few samples are sufficient to separate objects successfully, but in principle any way of estimating this derivative should work. As the noise size here corresponds to the typical delta used to compute the approximate derivative, it is also not surprising that small noise works well. The interesting part is that the derivative creates correlated changes for each object, not that this can be estimated based on noise samples.\n\nThird, I think the connection to human or biological vision is weaker than suggested by the authors. In biological vision we need a segmentation of the internal layer representations, not of the pixels in the image. This requires that the encoder and decoder are in some way related that allows us to connect our noise reconstructions and the encoding elements. Additionally, we would need a biologically plausible clustering algorithm that is based on the found similarities in the noise. Both are not present in the model presented here. Thus, substantial revisions would be necessary to transform the method proposed here into a biologically plausible method for object segmentation.", "questions": "Perhaps my questions highlight that this manuscript seems not ideally placed at a machine learning conference:\n- I would like to understand how the authors imagine the noise based segmentation to work in biological vision: How does the autoencoder model map to the brain? And what evidence is there that anything like this might actually happen in the brain?\n- Does this work in any way for natural scenes? Testing this on existing autoencoders for natural scenes could avoid my concerns about the training data being very targeted to create the patterns observed in humans.\n- And what about comparisons to alternative methods? Do other methods for segmentation or grouping get the gestalt tests wrong?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698791209316}], "openreview_url": "https://openreview.net/forum?id=ro4CgvfUKy", "arxiv_id": "2309.16515", "paper_pdf": "papers/ro4CgvfUKy.pdf", "paper_pdf_sha256": "9475886aeba18c6f1930e9f9b7fc2aa86c1a3c683f895d8b20954915a4d64917", "paper_pdf_bytes": 10063675, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhengqingUUU/LatentNoiseSegmentation", "code_repository": "ZhengqingUUU/LatentNoiseSegmentation", "code_commit": "eb79511f66c6e2f32c80aa2a35ef804bc30637c3", "code_archive": "repos/ro4CgvfUKy.zip", "code_archive_sha256": "c88c03e654dd54a8240d843551190f2a429bda77cdec50368824bf4ff344ff7c", "code_archive_bytes": 52462, "code_file_count": 12, "code_extensions": {".py": 10, ".sh": 2}, "github_disk_usage_kb": 52, "github_languages": {"Python": 165266, "Shell": 578}, "github_archived": false, "github_pushed_at": "2024-05-14T16:37:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/latent-noise-segmentation-how-neural-noise"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "D4JQEKlTyG", "year": 2023, "status": "rejected", "title": "Individual Privacy Accounting for Differentially Private Stochastic Gradient Descent", "authors": ["Da Yu", "Gautam Kamath", "Janardhan Kulkarni", "Tie-Yan Liu", "Jian Yin", "Huishuai Zhang"], "authorids": ["~Da_Yu1", "~Gautam_Kamath1", "~Janardhan_Kulkarni1", "~Tie-Yan_Liu1", "~Jian_Yin3", "~Huishuai_Zhang3"], "authors_source": "OpenReview API", "abstract": "Differentially private stochastic gradient descent (DP-SGD) is the workhorse algorithm for recent advances in private deep learning. It provides a single privacy guarantee to all datapoints in the dataset. We propose an efficient algorithm to compute privacy guarantees for individual examples when releasing models trained by DP-SGD. We use our algorithm to investigate individual privacy parameters across a number of datasets. We find that most examples enjoy stronger privacy guarantees than the worst-case bound. We further discover that the training loss and the privacy parameter of an example are well-correlated. This implies groups that are underserved in terms of model utility are simultaneously underserved in terms of privacy guarantee. For example, on CIFAR-10, the average $\\varepsilon$ of the class with the lowest test accuracy is 43.6% higher than that of the class with the highest accuracy. We also run membership inference attacks to show this reflects disparate empirical privacy risks.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "NKd8qRVDru", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper164/Reviewer_rPUG"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors provide an efficient algorithm for individual privacy accounting when using DP-SGD. By checking the individual privacy parameters, the authors find that these parameters are highly correlated to individual training loss. The authors also verify the validity of their individual privacy parameters by the results of membership inference attacks.", "review_text": "This paper is novel in terms of finding a correlation between training loss and individual privacy. Although without any theoretical analysis of this phenomenon, this paper provides various experimental results to support their finding. \n\n---\n After discussion with Area Chair and other reviewers \n\n---\nIt would be great if the authors could explain directly how Definition 2 is used in considering the privacy guarantee for a group. Currently, there is a lack of evidence why Definition 2 and Theorem 3.1 is necessary for Section 5, which is the highlight of this paper.\n", "strengths": "Major strengths:\n1. This is the first result showing the correlation between individual training loss and individual privacy parameters. Although this idea is natural if we consider the influence of the data points in the tail of the data distribution on both loss and privacy guarantee, it is great to see such a detailed experimental analysis.\n\nMinor strengths:\n1. This paper is easy to follow.\n2. The visualization of the experimental results is good.\n\nMajor weaknesses:\n1. I see that the authors are aware of the work by Feldman and Zenic (2021) on the Renyi Filter. It is not explained in this paper why they choose to use the concept by Redberg and Wang (2021) instead. (The authors are not giving enough acknowledgment to Redberg and Wang (2021) in their Definitions 1 and 2.)\n2. There is no theoretical result on the underlying reason behind this correlation between training loss and privacy parameters.\n3. In Section 6, the correlation between the attack success rate and the average $\\varepsilon$ is from the non-private model. This does not support the relationship between the individual privacy parameters and the actual privacy risk. This part of the experiment needs to be removed or replaced.\n\nMinor weaknesses:\n1. In Figure 4, I do not understand why the Pearson correlation coefficient is not calculated with the original points but with the fitted curve.\n2. Algorithm 1 needs to be revised. I think $q$ should be $p$ which is the sampling ratio for each batch. The authors have not defined $I_j$, and I guess $m$ is the batch size.\n3. In Section 3.4, the authors propose to release $\\varepsilon_i$ to the owner of the $i$th example. I think there needs a theoretical proof of why this does not incur additional privacy loss. I think the calculation of $\\varepsilon_i$ depends on the differentially private model which depends on the other examples. Therefore, we should be careful about what has been released and how much of the privacy budget has been used up.\n4. The usage of individual clipping is discussed in Section C but I do not think this is a common practice. If we do not adopt this individual clipping but only global clipping, will the result still show the correlation between loss and privacy as strong as it is now?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors provide an efficient algorithm for individual privacy accounting when using DP-SGD. By checking the individual privacy parameters, the authors find that these parameters are highly correlated to individual training loss. The authors also verify the validity of their individual privacy parameters by the results of membership inference attacks.", "strength_and_weaknesses": "Major strengths:\n1. This is the first result showing the correlation between individual training loss and individual privacy parameters. Although this idea is natural if we consider the influence of the data points in the tail of the data distribution on both loss and privacy guarantee, it is great to see such a detailed experimental analysis.\n\nMinor strengths:\n1. This paper is easy to follow.\n2. The visualization of the experimental results is good.\n\nMajor weaknesses:\n1. I see that the authors are aware of the work by Feldman and Zenic (2021) on the Renyi Filter. It is not explained in this paper why they choose to use the concept by Redberg and Wang (2021) instead. (The authors are not giving enough acknowledgment to Redberg and Wang (2021) in their Definitions 1 and 2.)\n2. There is no theoretical result on the underlying reason behind this correlation between training loss and privacy parameters.\n3. In Section 6, the correlation between the attack success rate and the average $\\varepsilon$ is from the non-private model. This does not support the relationship between the individual privacy parameters and the actual privacy risk. This part of the experiment needs to be removed or replaced.\n\nMinor weaknesses:\n1. In Figure 4, I do not understand why the Pearson correlation coefficient is not calculated with the original points but with the fitted curve.\n2. Algorithm 1 needs to be revised. I think $q$ should be $p$ which is the sampling ratio for each batch. The authors have not defined $I_j$, and I guess $m$ is the batch size.\n3. In Section 3.4, the authors propose to release $\\varepsilon_i$ to the owner of the $i$th example. I think there needs a theoretical proof of why this does not incur additional privacy loss. I think the calculation of $\\varepsilon_i$ depends on the differentially private model which depends on the other examples. Therefore, we should be careful about what has been released and how much of the privacy budget has been used up.\n4. The usage of individual clipping is discussed in Section C but I do not think this is a common practice. If we do not adopt this individual clipping but only global clipping, will the result still show the correlation between loss and privacy as strong as it is now?\n\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written. The statements in the paper are clear and concise. The novelty of this paper is good (see the strengths). I believe that the experimental results are reproducible (also see arXiv 2209.15596).", "summary_of_the_review": "This paper is novel in terms of finding a correlation between training loss and individual privacy. Although without any theoretical analysis of this phenomenon, this paper provides various experimental results to support their finding. \n\n---\n After discussion with Area Chair and other reviewers \n\n---\nIt would be great if the authors could explain directly how Definition 2 is used in considering the privacy guarantee for a group. Currently, there is a lack of evidence why Definition 2 and Theorem 3.1 is necessary for Section 5, which is the highlight of this paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666669068485}, {"id": "uy-dBoukIQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper164/Reviewer_JYd4"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper presents an algorithm to accurately account for individual differential privacy guarantees in DPSGD. ", "review_text": "This paper provides an algorithm for individual privacy analysis in DPSGD. However, the proposed algorithm does not correctly analyze the individual privacy as defined in their paper. In addition, the techniques proposed are straightforward. Furthermore, no ground truth is given.\n", "strengths": "Strength: Compared with worst-case DP guarantees, the individual DP guarantee of a sample allows us to better understand the privacy guarantee of a specific data sample, and its impact over model parameters.\n\nWeaknesses: \n\nW1. The proposed algorithm does not correctly estimate individual-level privacy as in Definitions 1 and 2. Hence, the claimed contribution is not justified and the impact of this paper is questionable. \n\n(1) The proposed algorithm modifies the underlying DPSGD algorithm for training. In particular, the algorithm modifies clipping norms for individuals, whereas the original DPSGD clips all gradient samples using the same constant. This change would affect the privacy analysis in DPSGD, and hence, it is unclear whether the algorithm could still correctly estimate the individual privacy guarantee. \n\n(2) The authors claim that their techniques are used for the output-specific individual privacy of DPSGD. Hence, we expect the algorithm to take the output of DPSGD as the input, which is the noisy gradient sum.  However, in the proposed algorithm, only the DP estimates for individual clipping norms (which, again, is not a component of the original DPSGD) are taken into account. It is unclear how the noisy gradient sums in multiple iterations, as the ``trajectory`` of output, impact the output of the algorithm. \n\n(3) No ground truth about individual privacy is provided. Hence, it is confusing how the paper evaluates the accuracy/correctness of the proposed algorithm. \n\nW2. The proposed techniques are straightforward.\n\n(1) The claim that subsampling in DPSGD complicates the privacy analysis is not well justified. The worst-case privacy analysis of DPSGD already incorporates subsampling, as shown in the following papers: \"Rényi Differential Privacy of the Sampled Gaussian Mechanism\" by Zhang et al., and \"Subsampled Rényi Differential Privacy and Analytical Moments Accountant\" by Wang et al. Thus, it is why subsampling brings a challenge in the analysis of individual DP.\n\n(2) Overall, the algorithm simply computes the Rényi divergence based on different clipping norms, which is merely an application of existing analysis.\n\n(3) The computation cost of N (which is the size of the private dataset) does not seem to be a problem for individual privacy accounting by definition, since otherwise you would miss some data samples and the accounting is not individual-level anymore (instead it is group level).\n\n(4) The technique for reducing the computation cost is to simply group data samples by their estimates of clipping norms at each iteration. This technique is straightforward. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents an algorithm to accurately account for individual differential privacy guarantees in DPSGD. ", "strength_and_weaknesses": "Strength: Compared with worst-case DP guarantees, the individual DP guarantee of a sample allows us to better understand the privacy guarantee of a specific data sample, and its impact over model parameters.\n\nWeaknesses: \n\nW1. The proposed algorithm does not correctly estimate individual-level privacy as in Definitions 1 and 2. Hence, the claimed contribution is not justified and the impact of this paper is questionable. \n\n(1) The proposed algorithm modifies the underlying DPSGD algorithm for training. In particular, the algorithm modifies clipping norms for individuals, whereas the original DPSGD clips all gradient samples using the same constant. This change would affect the privacy analysis in DPSGD, and hence, it is unclear whether the algorithm could still correctly estimate the individual privacy guarantee. \n\n(2) The authors claim that their techniques are used for the output-specific individual privacy of DPSGD. Hence, we expect the algorithm to take the output of DPSGD as the input, which is the noisy gradient sum.  However, in the proposed algorithm, only the DP estimates for individual clipping norms (which, again, is not a component of the original DPSGD) are taken into account. It is unclear how the noisy gradient sums in multiple iterations, as the ``trajectory`` of output, impact the output of the algorithm. \n\n(3) No ground truth about individual privacy is provided. Hence, it is confusing how the paper evaluates the accuracy/correctness of the proposed algorithm. \n\nW2. The proposed techniques are straightforward.\n\n(1) The claim that subsampling in DPSGD complicates the privacy analysis is not well justified. The worst-case privacy analysis of DPSGD already incorporates subsampling, as shown in the following papers: \"Rényi Differential Privacy of the Sampled Gaussian Mechanism\" by Zhang et al., and \"Subsampled Rényi Differential Privacy and Analytical Moments Accountant\" by Wang et al. Thus, it is why subsampling brings a challenge in the analysis of individual DP.\n\n(2) Overall, the algorithm simply computes the Rényi divergence based on different clipping norms, which is merely an application of existing analysis.\n\n(3) The computation cost of N (which is the size of the private dataset) does not seem to be a problem for individual privacy accounting by definition, since otherwise you would miss some data samples and the accounting is not individual-level anymore (instead it is group level).\n\n(4) The technique for reducing the computation cost is to simply group data samples by their estimates of clipping norms at each iteration. This technique is straightforward. ", "clarity,_quality,_novelty_and_reproducibility": "The clarity and quality of the paper are low. \n\n\n", "summary_of_the_review": "This paper provides an algorithm for individual privacy analysis in DPSGD. However, the proposed algorithm does not correctly analyze the individual privacy as defined in their paper. In addition, the techniques proposed are straightforward. Furthermore, no ground truth is given.\n", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666619995300}, {"id": "whn1uh9hN8g", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper164/Reviewer_9tdY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThe paper studies \"individual differential privacy\", i.e., given a specific example x, to what extent can one determine the membership (in a dataset D) of that example x given the output of a given algorithm?\n\n", "review_text": "\nThe paper is interesting and with some more effort can be made sufficiently clear.", "strengths": "\nThe idea is interesting, it allows to evaluate to what extent an individual suffered privacy loss due to the publication of a learnt model.\n\nThe paper could be made clearer (avoiding ambiguity among others by using symbols in addition to text when confusion is possible) and more self-contained (reviewing briefly already existing concepts).\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\nThe paper studies \"individual differential privacy\", i.e., given a specific example x, to what extent can one determine the membership (in a dataset D) of that example x given the output of a given algorithm?\n\n", "strength_and_weaknesses": "\nThe idea is interesting, it allows to evaluate to what extent an individual suffered privacy loss due to the publication of a learnt model.\n\nThe paper could be made clearer (avoiding ambiguity among others by using symbols in addition to text when confusion is possible) and more self-contained (reviewing briefly already existing concepts).\n\n", "clarity,_quality,_novelty_and_reproducibility": "\n\nFigure 2 is not very surprising given that the groups are created based on gradient norm at some point.  It may be more interesting to study the difference in gradient norm across groups constructed in advance not considering any gradient norm.\n\nThe steps 3 and 4 in Algorithm 1 are unclear to me, e.g., how exactly does rounding happen, what does \"compute RDP\" in line 4 mean exactly, where are the results of the \"compute RDP\" operation in line 4 stored (and where are they used later on), ... ?  The same holds for the explanation in Section 3.3 of the rounding.  It may help to define variables rather than vaguely describe concepts in words such as \"different privacy costs\" (the privacy costs of what exactly is meant here?)", "summary_of_the_review": "\nThe paper is interesting and with some more effort can be made sufficiently clear.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666432559078}], "openreview_url": "https://openreview.net/forum?id=D4JQEKlTyG", "arxiv_id": "2206.02617", "paper_pdf": "papers/D4JQEKlTyG.pdf", "paper_pdf_sha256": "473fc3d8ec4c4b341e7b322369d2deb3615fbec51e9a1704f12829d0bec4fdbb", "paper_pdf_bytes": 1484045, "paper_pdf_source": "openreview", "code_url": "https://github.com/dayu11/individual_privacy_of_DPSGD", "code_repository": "dayu11/individual_privacy_of_DPSGD", "code_commit": "ba43551e4ec8673efeb6191f9b243bcae12ffbd2", "code_archive": "repos/D4JQEKlTyG.zip", "code_archive_sha256": "aa8ec89b88296fe1e1d7e0ac929c898188171b747f14af0cc267181395d14992", "code_archive_bytes": 58839, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 86, "github_languages": {"Python": 36419}, "github_archived": false, "github_pushed_at": "2023-09-04T21:52:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/per-instance-privacy-accounting-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IptBMO1AR5g", "year": 2022, "status": "rejected", "title": "Regularizing Deep Neural Networks with Stochastic Estimators of Hessian Trace", "authors": ["Yucong Liu", "Tong Lin"], "authorids": ["~Yucong_Liu1", "~Tong_Lin1"], "authors_source": "OpenReview API", "abstract": "In this paper we develop a novel regularization method for deep neural networks by penalizing the trace of Hessian. This regularizer is motivated by a recent guarantee bound of the generalization error. Hutchinson method is a classical unbiased estimator for the trace of a matrix, but it is very time-consuming on deep learning models. Hence a dropout scheme is proposed to efficiently implements the Hutchinson method. Then we discuss a connection to linear stability of a nonlinear dynamical system. Experiments demonstrate that our method outperforms existing regularizers such as Jacobian, confidence penalty, and label smoothing. Our regularization method is also orthogonal to data augmentation methods, achieving the best performance when our method is combined with data augmentation. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "b5EBP-PgpsB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1596/Reviewer_GDik"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors propose a regularizer for training DNNs - specifically adding the trace of the Hessian w.r.t. the parameters of the model.\nThey present two efficient stochastic estimators of the trace, based on the Hutchinson method and their own extension to it, specific to NNs. \nThe authors also provide a motivation of the regularizer through a known generalization bound and through the perspective of dynamical systems stability. \nIn three experiments - on CIFAR10, CIFAR100 and Wiki-Text they present results suggesting that the proposed method is performing better on average than other techniques. \n", "review_text": "The paper starts well, with clear and simple goal - to introduce a new regualrizer. It motivates it through an existing generalization bound on linear models. Although, this bound does not translate to DNN models, I think this as a motivation is clearly presented and makes sense.\nThe stochastic estimators of the Hessian through Hutchinson method are well known and used in many places in ML. \n\nEquation 11 is somewhat redundant, since anyone familiar with how modern autodiff packages perform efficient Hessian-vector products  would know that they are just restating well known facts. For the least they should be citing Perlmutter's paper on fast Hessian-vector products, which basically define this. \nIn this whole section on the Hessian estimator, I'm very unclear what is the benefit of the proposed extension SETH-D over SETH-H? Theoretically, due to dropping blocks of the whole network it should be more efficient. However, practically the way that modern autodiff packages work, it would be a significantly challenge to actually compute Hessian-vector products such that they benefit from this. E.g. if you just zero out those parts of `sigma` and perform still perform \"dense\" products on line 5 and 6 of Algorithm 2 it won't make any difference. Later in the results the authors do cite that based on the probability that they pick `p` they gain different run times. I think the authors should provide more details on how in practice they actually implement SEHT-D in order to take advantage of the structure of \\sigma? These details although might look like not so important, are probably the only reason why one would even use SEHT-D over SETH-H. \n\nI have quite a bit of issue with the whole section 3.4. The authors attempt to motivate the trace trough stability analysis. First, that they expand the idealized ODE, for which there is now some literature which uses backward analysis to derive more accurate ODEs and in practice show case that the idealized version is not well capturing what happens with the dynamical system. Second, the main point at the end of this whole section, talking about stability of equilibrium points, is to argue that the goal of the method is to have a \"more unstable dynamical system\". Why would we want that - I don't know. The last paragraph presents an argument that since standard training is a function only of the training data, we would like it to be unstable in order to avoid overfitting. However, the authors do not present any evidence or even elude to why that last statement would be true and I'm quite doubtful if it will at all. We can make training unstable in many other ways (injecting significant amount of noise for instance) which do not lead to better generalization. As a result I would advice the authors to either significantly change, or just remove this whole section, as it currently serves no purpose. \n\nFinally, on the results section I have a few major issue with some of the numbers I'm seeing, which make me a bit doubtful of the overall results.\n\nThe WRN-28-10 has a baseline with 73% top 1 accuracy. The defualt pytorch open-source results from https://github.com/meliketoy/wide-resnet.pytorch indicate a 80% accuracy (ZCA - no-dropout 5e-4 weight decay). Perhaps the authors are suing different type of preprocessing scheme, or there are other differences, but this should be cleared out as to why they get so much worse results with the same model.\n\nAnother issue I have with the presented results, is that the authors do not mention to have done any search over the regualrizer penalty for any of the competing methods (and for theirs they do only two values - 0.01 0.05). They must mention why this is done - if they pick the values from one of the papers, they should make that more clear and also should make it clear that problems considered are equivalent to those in the reference. For many regualrizers their optimal weight changes between different problems. This might or might not explain some of the degradation results we see for some of the other techniques. \n\nProbably, the final experiment is best execute, with significant hyper parameter search and good ablations, which makes it the only one that makes me trust its results, and unsurprisingly there the benefits are significantly more marginal then in the others. However, it is unclear what is Validation and Test Perplexy exactly? Do you train and cross-validate the hyper parameter to select them, and then retrain on train + validation and measure Test, or do you just measure the performance on the Test set of the best cross-validated model? Please clarify. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a regularizer for training DNNs - specifically adding the trace of the Hessian w.r.t. the parameters of the model.\nThey present two efficient stochastic estimators of the trace, based on the Hutchinson method and their own extension to it, specific to NNs. \nThe authors also provide a motivation of the regularizer through a known generalization bound and through the perspective of dynamical systems stability. \nIn three experiments - on CIFAR10, CIFAR100 and Wiki-Text they present results suggesting that the proposed method is performing better on average than other techniques. \n", "main_review": "The paper starts well, with clear and simple goal - to introduce a new regualrizer. It motivates it through an existing generalization bound on linear models. Although, this bound does not translate to DNN models, I think this as a motivation is clearly presented and makes sense.\nThe stochastic estimators of the Hessian through Hutchinson method are well known and used in many places in ML. \n\nEquation 11 is somewhat redundant, since anyone familiar with how modern autodiff packages perform efficient Hessian-vector products  would know that they are just restating well known facts. For the least they should be citing Perlmutter's paper on fast Hessian-vector products, which basically define this. \nIn this whole section on the Hessian estimator, I'm very unclear what is the benefit of the proposed extension SETH-D over SETH-H? Theoretically, due to dropping blocks of the whole network it should be more efficient. However, practically the way that modern autodiff packages work, it would be a significantly challenge to actually compute Hessian-vector products such that they benefit from this. E.g. if you just zero out those parts of `sigma` and perform still perform \"dense\" products on line 5 and 6 of Algorithm 2 it won't make any difference. Later in the results the authors do cite that based on the probability that they pick `p` they gain different run times. I think the authors should provide more details on how in practice they actually implement SEHT-D in order to take advantage of the structure of \\sigma? These details although might look like not so important, are probably the only reason why one would even use SEHT-D over SETH-H. \n\nI have quite a bit of issue with the whole section 3.4. The authors attempt to motivate the trace trough stability analysis. First, that they expand the idealized ODE, for which there is now some literature which uses backward analysis to derive more accurate ODEs and in practice show case that the idealized version is not well capturing what happens with the dynamical system. Second, the main point at the end of this whole section, talking about stability of equilibrium points, is to argue that the goal of the method is to have a \"more unstable dynamical system\". Why would we want that - I don't know. The last paragraph presents an argument that since standard training is a function only of the training data, we would like it to be unstable in order to avoid overfitting. However, the authors do not present any evidence or even elude to why that last statement would be true and I'm quite doubtful if it will at all. We can make training unstable in many other ways (injecting significant amount of noise for instance) which do not lead to better generalization. As a result I would advice the authors to either significantly change, or just remove this whole section, as it currently serves no purpose. \n\nFinally, on the results section I have a few major issue with some of the numbers I'm seeing, which make me a bit doubtful of the overall results.\n\nThe WRN-28-10 has a baseline with 73% top 1 accuracy. The defualt pytorch open-source results from https://github.com/meliketoy/wide-resnet.pytorch indicate a 80% accuracy (ZCA - no-dropout 5e-4 weight decay). Perhaps the authors are suing different type of preprocessing scheme, or there are other differences, but this should be cleared out as to why they get so much worse results with the same model.\n\nAnother issue I have with the presented results, is that the authors do not mention to have done any search over the regualrizer penalty for any of the competing methods (and for theirs they do only two values - 0.01 0.05). They must mention why this is done - if they pick the values from one of the papers, they should make that more clear and also should make it clear that problems considered are equivalent to those in the reference. For many regualrizers their optimal weight changes between different problems. This might or might not explain some of the degradation results we see for some of the other techniques. \n\nProbably, the final experiment is best execute, with significant hyper parameter search and good ablations, which makes it the only one that makes me trust its results, and unsurprisingly there the benefits are significantly more marginal then in the others. However, it is unclear what is Validation and Test Perplexy exactly? Do you train and cross-validate the hyper parameter to select them, and then retrain on train + validation and measure Test, or do you just measure the performance on the Test set of the best cross-validated model? Please clarify. ", "summary_of_the_review": "Good and simple idea. I have some issues with lacking implementation details of the method that the authors use in practice. One of the sections seems completely redundant to the narrative, unless the authors provide better explanation for their final statements there. Some of the experiments seem to have lower numbers than publicly known results and in general there is a lack of exploring hyper parameters for other regualrizers. The last experiment is probably the only one that is done really well.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635877221853}, {"id": "w9klzMtlzTq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1596/Reviewer_YAnV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this paper, the authors propose an efficient regularizer for the trace of the Hessian. To circumvent the extra computation needed to do the naive Hutchinson estimator (which pretty much amounts to taking *three* derivatives of the loss), the authors subsample the weights matrices, and then a second time internal to each weights matrix). Experimental results on CIFAR-10 and CIFAR-100 are presented: the proposed regularizer helps, but unfortunately not in a statistically significant way.", "review_text": "Strengths:\n\n* The motivation of using jacobian and hessian trace norm regularizations is drawn from existing theoretical work, it's quite a nice way of tying it together with the extensive literature on generalization bounds. One comment worth adding though is that (I believe) all of those bounds are unfortunately vacuous. But even in that light, I think the idea is natural/worth writing a paper about.\n* The initial downsampling (corresponding to p_1) is a clever way to make this work. Otherwise, instead of the 1.2-1.5x slowdowns, I would have expected to have seen at the very least a 2x slowdown (for two more derivatives), and practically speaking more like a 10-50x slowdown given how slow hessian vector products seem to be in practice.\n\nWeaknesses:\n\n* Unfortunately, the \"better\" results have heavily overlapping confidence intervals. In the top pane, the 2 std confidence interval is +/- 0.47/ Combined with say the seht-d + label smoothing, it's clear that none of these results are significant. I would suggest that the authors aggressively drive down the sizes of these confidence intervals. In the bottom pane, the comparison between mixup and seht-d + mixup suffers from the same problem. In fact, from these experiments, the only reasonable convlusion is that \"mixup works really well\".\n\n* For comparison's sake, I think the authors really need to try the full version of the regularizer (with multiple iterations and prob=1.0). It's obviously going to be much slower, but if that doesn't work then the existing seht-d then I think something is probably wrong i.e., if the base claim is that \"hessian regularization works\", it is much less convincing if you don't directly test that hypothesis and go directly to approximations. I'd even be interested in seeing how much variance is introduced by the trace estimator, the two rounds of downsampling etc.\n\n* I think there needs to be a careful study of hyperparameters of the regularizer i.e., prob and regularization coefficient etc.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose an efficient regularizer for the trace of the Hessian. To circumvent the extra computation needed to do the naive Hutchinson estimator (which pretty much amounts to taking *three* derivatives of the loss), the authors subsample the weights matrices, and then a second time internal to each weights matrix). Experimental results on CIFAR-10 and CIFAR-100 are presented: the proposed regularizer helps, but unfortunately not in a statistically significant way.", "main_review": "Strengths:\n\n* The motivation of using jacobian and hessian trace norm regularizations is drawn from existing theoretical work, it's quite a nice way of tying it together with the extensive literature on generalization bounds. One comment worth adding though is that (I believe) all of those bounds are unfortunately vacuous. But even in that light, I think the idea is natural/worth writing a paper about.\n* The initial downsampling (corresponding to p_1) is a clever way to make this work. Otherwise, instead of the 1.2-1.5x slowdowns, I would have expected to have seen at the very least a 2x slowdown (for two more derivatives), and practically speaking more like a 10-50x slowdown given how slow hessian vector products seem to be in practice.\n\nWeaknesses:\n\n* Unfortunately, the \"better\" results have heavily overlapping confidence intervals. In the top pane, the 2 std confidence interval is +/- 0.47/ Combined with say the seht-d + label smoothing, it's clear that none of these results are significant. I would suggest that the authors aggressively drive down the sizes of these confidence intervals. In the bottom pane, the comparison between mixup and seht-d + mixup suffers from the same problem. In fact, from these experiments, the only reasonable convlusion is that \"mixup works really well\".\n\n* For comparison's sake, I think the authors really need to try the full version of the regularizer (with multiple iterations and prob=1.0). It's obviously going to be much slower, but if that doesn't work then the existing seht-d then I think something is probably wrong i.e., if the base claim is that \"hessian regularization works\", it is much less convincing if you don't directly test that hypothesis and go directly to approximations. I'd even be interested in seeing how much variance is introduced by the trace estimator, the two rounds of downsampling etc.\n\n* I think there needs to be a careful study of hyperparameters of the regularizer i.e., prob and regularization coefficient etc.", "summary_of_the_review": "The idea is interesting, the approximation is nice, but the experimental results are too weak/incomplete to make this a convincing paper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635813416634}, {"id": "g9GveEhBHDp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1596/Reviewer_imd7"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper develops a new regularization method for deep neural networks.  The proposed methods penalizes the trace of the Hessian.  The paper adapts Hutchinson method for estimating trace of a positive semi-definite matrix for the purposes of estimating trace of the Hessian.  This adaptation uses a dropout mechanism to efficiently compute trace of the Hessian.   The paper also studies the effects of minimizing the trace of the Hessian using linear dynamical systems theory, and shows that lowering the trace of Hessian diminishes the stability of the equilibrium points in the parameter (i.e., weight) space.  Thereby, reducing overfitting and improving generalizability of the network.  The paper concludes with a set of experiments on standard deep learning benchmarks---CIFAR-10, CIFAR-100, and WIKI-TEXT2---and in almost every case the proposed SEHT-D scheme achieves best results.\n", "review_text": "The paper makes a strong case for the proposed regularization scheme: penalizing the trace of the Hessian.  The empirical results included in the paper seem to indicate that the proposed regularization scheme improves network performance (and generalizability).  The technical details are accessible, and it is easy to follow the mathematical steps presented in the paper that lead to the final \"efficient\" algorithm for estimating the trace of the Hessian.  Section 3.4 that describes the linear stability analysis cogently argues why minimizing the trace of the Hessian is one way to improve the generalizability of the network. \n\nThe paper also suffers from some shortcomings.  While reading the paper, I stumbled upon a number typos: misspellings, incorrect double quotes (very common in LaTeX users), etc.  These can be easily fixed.\n\nA more serious issue with the paper is its use of the word \"efficient,\" as in, \"to efficiently implements [sic] the Hutchinson method.\"  I did not see any experiments that showcase how efficient the proposed method is.  I would have liked to see some runtime figures that drive home the point about the efficiency of the proposed approach.  Is there a way to remedy this?\n\nThe paper actually presents two methods for estimating the trace of the Hessian.  The first method does not use dropout,  whereas the second, more efficient, method uses dropout.  The paper contains no results using the first method.  It would be nice to know the performance gap between the first and the second method.  In many situations performance and efficiency are inversely related.  Does that hold here too?  Alternately, may be deep neural networks upend this trend and the second method posts both better performance and better runtimes.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper develops a new regularization method for deep neural networks.  The proposed methods penalizes the trace of the Hessian.  The paper adapts Hutchinson method for estimating trace of a positive semi-definite matrix for the purposes of estimating trace of the Hessian.  This adaptation uses a dropout mechanism to efficiently compute trace of the Hessian.   The paper also studies the effects of minimizing the trace of the Hessian using linear dynamical systems theory, and shows that lowering the trace of Hessian diminishes the stability of the equilibrium points in the parameter (i.e., weight) space.  Thereby, reducing overfitting and improving generalizability of the network.  The paper concludes with a set of experiments on standard deep learning benchmarks---CIFAR-10, CIFAR-100, and WIKI-TEXT2---and in almost every case the proposed SEHT-D scheme achieves best results.\n", "main_review": "The paper makes a strong case for the proposed regularization scheme: penalizing the trace of the Hessian.  The empirical results included in the paper seem to indicate that the proposed regularization scheme improves network performance (and generalizability).  The technical details are accessible, and it is easy to follow the mathematical steps presented in the paper that lead to the final \"efficient\" algorithm for estimating the trace of the Hessian.  Section 3.4 that describes the linear stability analysis cogently argues why minimizing the trace of the Hessian is one way to improve the generalizability of the network. \n\nThe paper also suffers from some shortcomings.  While reading the paper, I stumbled upon a number typos: misspellings, incorrect double quotes (very common in LaTeX users), etc.  These can be easily fixed.\n\nA more serious issue with the paper is its use of the word \"efficient,\" as in, \"to efficiently implements [sic] the Hutchinson method.\"  I did not see any experiments that showcase how efficient the proposed method is.  I would have liked to see some runtime figures that drive home the point about the efficiency of the proposed approach.  Is there a way to remedy this?\n\nThe paper actually presents two methods for estimating the trace of the Hessian.  The first method does not use dropout,  whereas the second, more efficient, method uses dropout.  The paper contains no results using the first method.  It would be nice to know the performance gap between the first and the second method.  In many situations performance and efficiency are inversely related.  Does that hold here too?  Alternately, may be deep neural networks upend this trend and the second method posts both better performance and better runtimes.", "summary_of_the_review": "Overall the paper is well-written.  It proposes a new regularization scheme that seems to improve deep neural networks performance on CIFAR-10, CIFAR-100, and WIKI-TEXT2 datasets.  The proposed scheme improves performance for both convolutional and recurrent neural networks.  The paper uses linear dynamical systems theory to argue that minimizing the trace of the Hessian improves the generalizability of the network.  Lastly, the mathematical description for \"efficiently\" computing the trace of the Hessian is easy to follow.\n\nMy only issue with the paper is that it doesn't include any results to support the assertion that the algorithm for estimating trace of the Hessian is efficient.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635781645890}, {"id": "qUhS-ehLPe", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1596/Reviewer_CMzT"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a method for estimating the trace of the Hessian of the cross-entropy loss with respect to the weights of a neural network classifier. They suggest adding the trace estimator as a regularizing penalty term for training neural networks with improved generalization.\n\nThe authors give two theoretical motivations for regularizing the Hessian trace:\n- For linear models the trace of the Hessian of the cross-entropy loss with respect to the logits appears as a factor in a term for bounding generalization error. The authors draw a connection between the Hessian with respect to the weights and the Hessian with respect to the logits and argue that penalizing the trace of the Hessian with respect to the weights also leads to a tighter bound on the generalization error.\n- For the continuous gradient dynamics around the optimum the Hessian eigenspectrum is the negative eigenspectrum of the Jacobian of the gradient dynamics which determine the stability of the dynamic system. The authors argue that decreasing Hessian trace increases Jacobian trace and hence reduces data dependend stability which will avoid overfitting.\n\nThe authors' suggested method for estimating the trace uses the stochastic Hutchinson trace estimator which samples a random vector sigma with E(sigma sigma^T) = I to get E(sigma^T H sigma) = tr(H). The authors appear to use back-propagation through a back-propagation gradient approach to compute sigma^T H sigma by first computing the gradient g = dl/domega of the loss l with respect to the weights omega, and then computing the gradient of the inner product g^T sigma which is a scalar, with respect to the weights again. This gradient-over-gradient can be inner product multiplied again with sigma.\n\nIn order to save cost the authors suggest a drop out method setting a fraction of the terms of sigma to 0 which saves a fraction of the derivative computation and correspondingly drops terms of the trace sum.\n\nThe authors then present experimental results evaluating their suggested regularization method in combination with and compared to other regularization methods on\n- CIFAR10 / Resnet18\n- CIFAR100 / Wide Residual Network\n- WikiText-2 / 2-layer LSTM", "review_text": "Strengths:\n- The connection of Hessian spectrum and trace to generalization is an interesting and promising area of research and it is great to see results for a method regularizing the Hessian that seems to outperform other regularization methods\n- The structure of the paper is clear and the paper is understandable\n\nWeaknesses:\n- The connection between theoretical justification and practical method is not always clear and the relevant details could be expanded:\n\nMy feeling is that the connection between using the Hessian / Jacobian with respect to the logits vs with respect to the parameters is not as straight forward as it is made out to be in the paper and could be explained better. For example the paper states: \"Naturally, at the minimum the gradient is zero and the norm of the Jacobian matrix is small near minimum.\"\n\nBut if we have for e.g. a nonlinear classifier f(x) = W y(x) where W are the weights of the last layer then dl/df is what is meant as Jacobian in the bound but the gradient in optimization is with respect to parameters, e.g. dl(f(x))/dW = dl/df * df/dW = dl/df * y(x). Could there perhaps be a way for the gradient to be small if y(x) is small without dl/df being small? This question could be expanded upon.\n\nSimilarly, an explanation of the difference between the trace of the loss Hessian w.r.t. to logits vs w.r.t to the weights would benefit the paper. For example it is well-known that the Hessian w.r.t. to weights for a deep neural network can have negative eigenvalues but the Hessian w.r.t. to logits for a convex loss has only positive eigenvalues. Does this matter to the trace and associated bound somehow?\n\nRegarding the linear stability argument the paper could be improved by empirically demonstrating that regularizing the Hessian trace leads to an optimum with a Hessian that has \"less stability\", perhaps by analyzing the eigenvalues at the optimum. Furthermore, is there a connection between the idea that increasing instability helps generalization to the literature that says that flat optima help generalization?\n\n- There is no discussion of using the reverse-over-reverse autodiff method for computing the Hessian vector product instead of a forward-over-reverse autodiff method which can save memory and hence could also be faster on a GPU:\n\nIn principle you can get the directional derivative of the scalar g^T(omega) * sigma with respect to omega in the direction of sigma by evaluating\n\ndg/domega * sigma = (g(omega + eps * sigma) - g(omega)) / eps\n\nor similarly\n\nd(g*sigma)/domega * sigma = (g(omega + eps * sigma) * sigma - g(omega) * sigma) / eps\n\nat a factor two of the cost of evaluating only the gradient (gradient evaluated at omega). Alternatively, forward-mode autodiff can be used to the same effect and the dropout idea can still be applied by setting some of the sigma to zero and hence eliminating the cost further due to dropped nodes in the computational graph.\n\n- The experimental section does not analyse the regularizer itself or the connection to theory of the Hessian trace regularization and is instead focused on just demonstrating beating baselines:\n\nThere is currently no analysis of the trade-off of the bias of the trace estimator by using more dropout vs additional compute cost, just the statement \"Although increasing the probability to select parameters can improve test accuracy, the time consumption will increase a lot\". The paper could be improved by adding an empiricial analysis of the accuracy of the trace estimator.\n\nThere is also no analysis of how trace accuracy impacts generalization performance and trace itself, e.g. you could add plots showing how the true Hessian trace changes throughout training.\n\nIt would be good to see the training and cross validation curves for different hyperparams of overlayed penalty terms (like Hessian trace term, L2 norm term etc.) to see how they impact training and how stable training is to the corresponding hyperparameters. It would be good to have an analysis of how much does Hessian trace term add on top of hyper-param tuned other regularization, e.g. for Jacobian regularization the paper says just \"we set number of projections nproj = 1 and weight values λJR = 0.01\" which does not necessarily suggest that there was a sufficient hyperparameter tuning effort.\n\nAnother interesting question to analyze would be: How does the added regularizer impact the training dynamics, e.g. faster or slower convergence in the beginning vs end of training and why?\n\n\nSome more general recommendations to improve the paper:\n- Checking grammar and spelling typos will improve readability and quality of the paper\n- State the generalization bound as a theorem as in the original paper (Theorem 4.1)\n- In related work perhaps also connect to other works analyzing Hessian and Fisher matrix spectrum and related norms for generalization (e.g. \"Fisher-Rao Metric, Geometry, and Complexity of Neural Networks\")\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a method for estimating the trace of the Hessian of the cross-entropy loss with respect to the weights of a neural network classifier. They suggest adding the trace estimator as a regularizing penalty term for training neural networks with improved generalization.\n\nThe authors give two theoretical motivations for regularizing the Hessian trace:\n- For linear models the trace of the Hessian of the cross-entropy loss with respect to the logits appears as a factor in a term for bounding generalization error. The authors draw a connection between the Hessian with respect to the weights and the Hessian with respect to the logits and argue that penalizing the trace of the Hessian with respect to the weights also leads to a tighter bound on the generalization error.\n- For the continuous gradient dynamics around the optimum the Hessian eigenspectrum is the negative eigenspectrum of the Jacobian of the gradient dynamics which determine the stability of the dynamic system. The authors argue that decreasing Hessian trace increases Jacobian trace and hence reduces data dependend stability which will avoid overfitting.\n\nThe authors' suggested method for estimating the trace uses the stochastic Hutchinson trace estimator which samples a random vector sigma with E(sigma sigma^T) = I to get E(sigma^T H sigma) = tr(H). The authors appear to use back-propagation through a back-propagation gradient approach to compute sigma^T H sigma by first computing the gradient g = dl/domega of the loss l with respect to the weights omega, and then computing the gradient of the inner product g^T sigma which is a scalar, with respect to the weights again. This gradient-over-gradient can be inner product multiplied again with sigma.\n\nIn order to save cost the authors suggest a drop out method setting a fraction of the terms of sigma to 0 which saves a fraction of the derivative computation and correspondingly drops terms of the trace sum.\n\nThe authors then present experimental results evaluating their suggested regularization method in combination with and compared to other regularization methods on\n- CIFAR10 / Resnet18\n- CIFAR100 / Wide Residual Network\n- WikiText-2 / 2-layer LSTM", "main_review": "Strengths:\n- The connection of Hessian spectrum and trace to generalization is an interesting and promising area of research and it is great to see results for a method regularizing the Hessian that seems to outperform other regularization methods\n- The structure of the paper is clear and the paper is understandable\n\nWeaknesses:\n- The connection between theoretical justification and practical method is not always clear and the relevant details could be expanded:\n\nMy feeling is that the connection between using the Hessian / Jacobian with respect to the logits vs with respect to the parameters is not as straight forward as it is made out to be in the paper and could be explained better. For example the paper states: \"Naturally, at the minimum the gradient is zero and the norm of the Jacobian matrix is small near minimum.\"\n\nBut if we have for e.g. a nonlinear classifier f(x) = W y(x) where W are the weights of the last layer then dl/df is what is meant as Jacobian in the bound but the gradient in optimization is with respect to parameters, e.g. dl(f(x))/dW = dl/df * df/dW = dl/df * y(x). Could there perhaps be a way for the gradient to be small if y(x) is small without dl/df being small? This question could be expanded upon.\n\nSimilarly, an explanation of the difference between the trace of the loss Hessian w.r.t. to logits vs w.r.t to the weights would benefit the paper. For example it is well-known that the Hessian w.r.t. to weights for a deep neural network can have negative eigenvalues but the Hessian w.r.t. to logits for a convex loss has only positive eigenvalues. Does this matter to the trace and associated bound somehow?\n\nRegarding the linear stability argument the paper could be improved by empirically demonstrating that regularizing the Hessian trace leads to an optimum with a Hessian that has \"less stability\", perhaps by analyzing the eigenvalues at the optimum. Furthermore, is there a connection between the idea that increasing instability helps generalization to the literature that says that flat optima help generalization?\n\n- There is no discussion of using the reverse-over-reverse autodiff method for computing the Hessian vector product instead of a forward-over-reverse autodiff method which can save memory and hence could also be faster on a GPU:\n\nIn principle you can get the directional derivative of the scalar g^T(omega) * sigma with respect to omega in the direction of sigma by evaluating\n\ndg/domega * sigma = (g(omega + eps * sigma) - g(omega)) / eps\n\nor similarly\n\nd(g*sigma)/domega * sigma = (g(omega + eps * sigma) * sigma - g(omega) * sigma) / eps\n\nat a factor two of the cost of evaluating only the gradient (gradient evaluated at omega). Alternatively, forward-mode autodiff can be used to the same effect and the dropout idea can still be applied by setting some of the sigma to zero and hence eliminating the cost further due to dropped nodes in the computational graph.\n\n- The experimental section does not analyse the regularizer itself or the connection to theory of the Hessian trace regularization and is instead focused on just demonstrating beating baselines:\n\nThere is currently no analysis of the trade-off of the bias of the trace estimator by using more dropout vs additional compute cost, just the statement \"Although increasing the probability to select parameters can improve test accuracy, the time consumption will increase a lot\". The paper could be improved by adding an empiricial analysis of the accuracy of the trace estimator.\n\nThere is also no analysis of how trace accuracy impacts generalization performance and trace itself, e.g. you could add plots showing how the true Hessian trace changes throughout training.\n\nIt would be good to see the training and cross validation curves for different hyperparams of overlayed penalty terms (like Hessian trace term, L2 norm term etc.) to see how they impact training and how stable training is to the corresponding hyperparameters. It would be good to have an analysis of how much does Hessian trace term add on top of hyper-param tuned other regularization, e.g. for Jacobian regularization the paper says just \"we set number of projections nproj = 1 and weight values λJR = 0.01\" which does not necessarily suggest that there was a sufficient hyperparameter tuning effort.\n\nAnother interesting question to analyze would be: How does the added regularizer impact the training dynamics, e.g. faster or slower convergence in the beginning vs end of training and why?\n\n\nSome more general recommendations to improve the paper:\n- Checking grammar and spelling typos will improve readability and quality of the paper\n- State the generalization bound as a theorem as in the original paper (Theorem 4.1)\n- In related work perhaps also connect to other works analyzing Hessian and Fisher matrix spectrum and related norms for generalization (e.g. \"Fisher-Rao Metric, Geometry, and Complexity of Neural Networks\")\n", "summary_of_the_review": "I do not recommend to accept the paper in its current form since the connection between the theoretical justification and the empirical evidence is weak and the experiments in the paper are not sufficient for an empirical understanding of how the suggested regularizer really works.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635766678663}], "openreview_url": "https://openreview.net/forum?id=IptBMO1AR5g", "arxiv_id": "2208.05924", "paper_pdf": "papers/IptBMO1AR5g.pdf", "paper_pdf_sha256": "c2017a0a918c16c906d6f83c15f081d87afc2c55910b0d97c29e2c222abe7505", "paper_pdf_bytes": 256462, "paper_pdf_source": "openreview", "code_url": "https://github.com/ICLRsubmission1596/Regularizing-Deep-Neural-Networks-with-Stochastic-Estimators-of-Hessian-Trace", "code_repository": "ICLRsubmission1596/Regularizing-Deep-Neural-Networks-with-Stochastic-Estimators-of-Hessian-Trace", "code_commit": "cd721ae5c569bc9b9e73a35c5f10c87bbd97febb", "code_archive": "repos/IptBMO1AR5g.zip", "code_archive_sha256": "d69b21fd66585c6fc38344d391c8448955bf0a536023def040fa4f7cabc50492", "code_archive_bytes": 76671, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 86, "github_languages": {"Python": 235849}, "github_archived": false, "github_pushed_at": "2021-10-05T20:54:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/regularizing-deep-neural-networks-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Twm9LnWK-zt", "year": 2021, "status": "rejected", "title": "Searching towards Class-Aware Generators for Conditional Generative Adversarial Networks", "authors": ["Peng Zhou", "Lingxi Xie", "XIAOPENG ZHANG", "Bingbing Ni", "Qi Tian"], "authorids": ["~Peng_Zhou2", "~Lingxi_Xie1", "~XIAOPENG_ZHANG7", "~Bingbing_Ni3", "~Qi_Tian3"], "authors_source": "OpenReview API", "abstract": "Conditional Generative Adversarial Networks (cGAN) were designed to generate images based on the provided conditions, e.g., class-level distributions. However, existing methods have used the same generating architecture for all classes. This paper presents a novel idea that adopts NAS to find a distinct architecture for each class. The search space contains regular and class-modulated convolutions, where the latter is designed to introduce class-specific information while avoiding the reduction of training data for each class generator. The search algorithm follows a weight-sharing pipeline with mixed-architecture optimization so that the search cost does not grow with the number of classes. To learn the sampling policy, a Markov decision process is embedded into the search algorithm and a moving average is applied for better stability. We evaluate our approach on CIFAR10 and CIFAR100. Besides achieving better image generation quality in terms of FID scores, we discover several insights that are helpful in designing cGAN models.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "L8SETnSWTs3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2548/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an interesting idea that adopts NAS to find a distinct architecture for each class based on cGAN framework. Within the framework, the paper also proposes an operator, Class-Modulated convolution (CMconv), to allow the training data to be shared among different architectures, so as to balance the training data across classes. The proposed method leverages a Markov Decision Process (MDP) in the search algorithm, and learns the sampling policy for NAS. Comprehensive experiments demonstrate the class-aware NAS can outperform class-agnostic NAS.\n\n- The paper is well written with sufficient figures and plots. \n- The proposed idea is straightforward and convincing.\n- Rich experiments and analysis are conducted. Implementation details are clearly described in the appendix.\n- I like the figures architecture searched specifically for different classes.\n\nHowever, I still have some concerns:\n\n- From my point of view, the proposed CMconv has exactly the same architecture as the one in Karras et al., instead of using class embedding as input. Please clarify the difference and clearly point out in the paper.\n\n- Quantitative results in Table 2 seem not promising. The proposed method is not compatible with the SOTA. Although the paper claims that the proposed idea can be applied to the existing methods and performs better, there’s no evidence showing that.\n\n- It’d be better to also list infra-FIDs of one existing unconditioned GAN method in the Table 1, for the lower bound / baseline of the experiments.\n\n- For fair comparison, it’s better to have a table with quantitative results of IS/FID on CIFAR-10, listing the existing methods with conditioned/unconditioned, such as AutoGAN, style GAN etc. \n\nOverall, the proposed method is interesting, NAS by MDP with class-aware functionality, which can ideally outperform the class-agnostic based method. The experiments are comprehensive, with strong ablation study and analysis. However, it still requires some clarification and convincing experiments to demonstrate the performance. \n\nI'm willing to rate higher if the concerns are addressed.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Class-awareness mechanism involved in the NAS approach", "review": "This paper proposes an interesting idea that adopts NAS to find a distinct architecture for each class based on cGAN framework. Within the framework, the paper also proposes an operator, Class-Modulated convolution (CMconv), to allow the training data to be shared among different architectures, so as to balance the training data across classes. The proposed method leverages a Markov Decision Process (MDP) in the search algorithm, and learns the sampling policy for NAS. Comprehensive experiments demonstrate the class-aware NAS can outperform class-agnostic NAS.\n\n- The paper is well written with sufficient figures and plots. \n- The proposed idea is straightforward and convincing.\n- Rich experiments and analysis are conducted. Implementation details are clearly described in the appendix.\n- I like the figures architecture searched specifically for different classes.\n\nHowever, I still have some concerns:\n\n- From my point of view, the proposed CMconv has exactly the same architecture as the one in Karras et al., instead of using class embedding as input. Please clarify the difference and clearly point out in the paper.\n\n- Quantitative results in Table 2 seem not promising. The proposed method is not compatible with the SOTA. Although the paper claims that the proposed idea can be applied to the existing methods and performs better, there’s no evidence showing that.\n\n- It’d be better to also list infra-FIDs of one existing unconditioned GAN method in the Table 1, for the lower bound / baseline of the experiments.\n\n- For fair comparison, it’s better to have a table with quantitative results of IS/FID on CIFAR-10, listing the existing methods with conditioned/unconditioned, such as AutoGAN, style GAN etc. \n\nOverall, the proposed method is interesting, NAS by MDP with class-aware functionality, which can ideally outperform the class-agnostic based method. The experiments are comprehensive, with strong ablation study and analysis. However, it still requires some clarification and convincing experiments to demonstrate the performance. \n\nI'm willing to rate higher if the concerns are addressed.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604062119821}, {"id": "x-pL7mEIRe4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2548/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:  \nThis paper proposes to use neural architecture search (NAS) to automatically discover useful conditional generative adversarial network (cGAN) architectures. Specifically, this work aims to find a dedicated architecture for each class.  \n\nStrengths:  \n-Paper is well written.  \n-Demonstrates that optimizing architectures for each class yields some improvement over using a single architecture for all classes.  \n-Architectures learned by the NAS reveal insights about how to use existing building blocks, such as where best to place feature modulation layers in the network.  \n\nWeaknesses:  \n-No random baseline (see [1]).  \n-Given how close NAS-cGAN and NAS-caGAN are in terms of performance, confidence intervals should be reported to confirm that improvement is significant.  \n-Majority of improvement comes from fine-tuning on each class individually. It is unclear how much of this improvement is simply due to additional capacity.  \n-Complexity of the model appears to be disproportionate to the improvement in performance (lots of implementation effort for a somewhat small gain in performance).  \n\nRecommendation and Justification:  \nWhile I think this paper is well written, after reading it I am not convinced of the usefulness of the core idea, which is that generator architectures should be class-aware. It is never explained why it might be desirable for each class to have a distinct generator network, and I cannot think of any reason why this may be the case aside from increased model capacity. For this reason I think that this paper is currently marginally below the acceptance threshold, but look forward to the author's explanation.  \n\nClarifying Questions:  \n-How is calibration performed exactly? Is this procedure the same for NAS-cGAN and NAS-caGAN? Is a separate copy of weights fine-tuned for each class?  If a new copy of weights needs to be created for each class, proper comparison would be to a non-NAS model with up to n_classes times more model capacity than the base model.  \n\n[1] Li, Liam, and Ameet Talwalkar. \"Random search and reproducibility for neural architecture search.\" Uncertainty in Artificial Intelligence. PMLR, 2020.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "Summary:  \nThis paper proposes to use neural architecture search (NAS) to automatically discover useful conditional generative adversarial network (cGAN) architectures. Specifically, this work aims to find a dedicated architecture for each class.  \n\nStrengths:  \n-Paper is well written.  \n-Demonstrates that optimizing architectures for each class yields some improvement over using a single architecture for all classes.  \n-Architectures learned by the NAS reveal insights about how to use existing building blocks, such as where best to place feature modulation layers in the network.  \n\nWeaknesses:  \n-No random baseline (see [1]).  \n-Given how close NAS-cGAN and NAS-caGAN are in terms of performance, confidence intervals should be reported to confirm that improvement is significant.  \n-Majority of improvement comes from fine-tuning on each class individually. It is unclear how much of this improvement is simply due to additional capacity.  \n-Complexity of the model appears to be disproportionate to the improvement in performance (lots of implementation effort for a somewhat small gain in performance).  \n\nRecommendation and Justification:  \nWhile I think this paper is well written, after reading it I am not convinced of the usefulness of the core idea, which is that generator architectures should be class-aware. It is never explained why it might be desirable for each class to have a distinct generator network, and I cannot think of any reason why this may be the case aside from increased model capacity. For this reason I think that this paper is currently marginally below the acceptance threshold, but look forward to the author's explanation.  \n\nClarifying Questions:  \n-How is calibration performed exactly? Is this procedure the same for NAS-cGAN and NAS-caGAN? Is a separate copy of weights fine-tuned for each class?  If a new copy of weights needs to be created for each class, proper comparison would be to a non-NAS model with up to n_classes times more model capacity than the base model.  \n\n[1] Li, Liam, and Ameet Talwalkar. \"Random search and reproducibility for neural architecture search.\" Uncertainty in Artificial Intelligence. PMLR, 2020.  ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603853954356}, {"id": "me6tDJs9UK8", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2548/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n \nSummary:\n\nThis paper proposes a framework NAS-caGAN that adopts RL-based NAS to search the optimal class-aware generator architecture by directly optimizing the Inception Score (IS) using the  REINFORCE algorithm, and leverages the mixed-architecture optimization to mitigate the training data sparsity of each category. The authors design a Class-Modulated Convolution to allow for the weight-sharing among different searched architectures. The proposed NAS-caGAN outperforms the model that employs searched class-agnostic architecture on CIFAR 10 and achieves better results compared with cproj (Miyato & Koyama, 2018) on CIFAR 100. \n \n##########################################################################\n \nReasons for score: \n\nOverall, I vote for rejecting, but I am happy to modify the score if the authors could provide further details about my concerns. My major concerns are about the baseline choice, computational costs, and the evidence to support the method’s utility (see cons below).  Hopefully the authors can address my concern in the rebuttal period. \n \n##########################################################################\n \nPros: \n1. The idea of using an RL-based NAS to search for the optimal architecture of different categories is quite interesting and empirically demonstrates its superior effect when compared with the searched identical generator architecture regardless of the category class.\n2. Overall, the paper is well written and technically sound. \n3. Good ablation test that highlights the utility of the class-projection (cproj) discriminator over a standard GAN’s discriminator.\n4. Good illustration of the proportion of different operators (i.e., RConv and CMConv) in each layer of the class-aware architecture.\n \n##########################################################################\n \nCons: \n1. The key concern about the paper is the lack of experiments to validate the utility of the proposed method compared with the previous work. Despite the paper asserting that the usefulness of searching for distinct architecture for different categories, the paper only compares the **searched** class-agnostic model (NAS-cGAN) with the proposed class-aware one (NAS-caGAN), and ignores the comparison with the literature on CIFAR 10, such as (Brock et al., 2018; Kavalerov et al., 2019; Zhao et al., 2020) .\n2.  Another concern is the baseline presented. \n(1) Most recent works such as (Kavalerov et al., 2019; Zhao et al., 2020) selected BigGAN (Brock et al., 2018) as their baseline. Considering the limited results (not achieving the state-of-the-art on neither of the presented metrics), a deeper analysis and performance comparison of the proposed framework would have been helpful to argue for its effectiveness. \n(2) It is interesting to inject CMConv operation at the first layer of BigGAN, which achieves a better result than other operator settings. Why not use BigGAN as the baseline?\n3. Could the authors provide the details about computational and **time** costs of the proposed method? How much time/resources would this method take to search?  \n4. In the Sec.4.1, the paper claims that this method could benefit other works. Considering the lack of convincing experiments to support this argument, it is doubtful about its correctness.\n5. For the “moving average”, it would be better to provide more details about it. The paper claims the usage of it for training stability, but never mentions the details about it, which seems not very clear to me.\n6. In the Sec.2, it would have been nice to supplement the details compared with the recent literature of GANs with NAS.\n \n#########################################################################\n \nMinor comments: \n1. Table 3:  The reported evaluation score is not mentioned, which might be FID.\n2. The “moving average” has not been well addressed in the main paper. \n3. Section 3.1: it would be easier to follow if the authors could paraphrase the last two paragraphs. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting idea but may lack of some supportive experiments (see cons).", "review": "##########################################################################\n \nSummary:\n\nThis paper proposes a framework NAS-caGAN that adopts RL-based NAS to search the optimal class-aware generator architecture by directly optimizing the Inception Score (IS) using the  REINFORCE algorithm, and leverages the mixed-architecture optimization to mitigate the training data sparsity of each category. The authors design a Class-Modulated Convolution to allow for the weight-sharing among different searched architectures. The proposed NAS-caGAN outperforms the model that employs searched class-agnostic architecture on CIFAR 10 and achieves better results compared with cproj (Miyato & Koyama, 2018) on CIFAR 100. \n \n##########################################################################\n \nReasons for score: \n\nOverall, I vote for rejecting, but I am happy to modify the score if the authors could provide further details about my concerns. My major concerns are about the baseline choice, computational costs, and the evidence to support the method’s utility (see cons below).  Hopefully the authors can address my concern in the rebuttal period. \n \n##########################################################################\n \nPros: \n1. The idea of using an RL-based NAS to search for the optimal architecture of different categories is quite interesting and empirically demonstrates its superior effect when compared with the searched identical generator architecture regardless of the category class.\n2. Overall, the paper is well written and technically sound. \n3. Good ablation test that highlights the utility of the class-projection (cproj) discriminator over a standard GAN’s discriminator.\n4. Good illustration of the proportion of different operators (i.e., RConv and CMConv) in each layer of the class-aware architecture.\n \n##########################################################################\n \nCons: \n1. The key concern about the paper is the lack of experiments to validate the utility of the proposed method compared with the previous work. Despite the paper asserting that the usefulness of searching for distinct architecture for different categories, the paper only compares the **searched** class-agnostic model (NAS-cGAN) with the proposed class-aware one (NAS-caGAN), and ignores the comparison with the literature on CIFAR 10, such as (Brock et al., 2018; Kavalerov et al., 2019; Zhao et al., 2020) .\n2.  Another concern is the baseline presented. \n(1) Most recent works such as (Kavalerov et al., 2019; Zhao et al., 2020) selected BigGAN (Brock et al., 2018) as their baseline. Considering the limited results (not achieving the state-of-the-art on neither of the presented metrics), a deeper analysis and performance comparison of the proposed framework would have been helpful to argue for its effectiveness. \n(2) It is interesting to inject CMConv operation at the first layer of BigGAN, which achieves a better result than other operator settings. Why not use BigGAN as the baseline?\n3. Could the authors provide the details about computational and **time** costs of the proposed method? How much time/resources would this method take to search?  \n4. In the Sec.4.1, the paper claims that this method could benefit other works. Considering the lack of convincing experiments to support this argument, it is doubtful about its correctness.\n5. For the “moving average”, it would be better to provide more details about it. The paper claims the usage of it for training stability, but never mentions the details about it, which seems not very clear to me.\n6. In the Sec.2, it would have been nice to supplement the details compared with the recent literature of GANs with NAS.\n \n#########################################################################\n \nMinor comments: \n1. Table 3:  The reported evaluation score is not mentioned, which might be FID.\n2. The “moving average” has not been well addressed in the main paper. \n3. Section 3.1: it would be easier to follow if the authors could paraphrase the last two paragraphs. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603763471788}, {"id": "6YWwvVDgTj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2548/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an interesting method that adopts NAS to search multiple class-aware generator architectures for cGAN instead of class-agnostic type. A search space containing both normal and class-modulated convolutions are introduced to simplify the process of re-training. Besides, this paper design a mixed-architecture optimization to specifically address the computational burden issue under the setting of a multi-net search. The search results also give some insights about constructing cGAN models.\n\n\nStrengths:\n- The perspective of adopting the NAS method to explore the class-aware generator architectures for conditional GANs is relatively novel and interesting, although there are some related works about searching unconditional GANs. \n-The proposed flexibility and safety search space is effective to address the categories grow issue. \n- The developed mixed-architecture optimization is a clever way to improve the efficiency of the search and re-training process.\n- The authors conduct extensive experiments and give some interesting insight/discussion about the results.\n\nWeaknesses:\n- There are no innovative approaches toward neural architecture search are proposed, and this work only focuses on how to bring existing RL-based NAS methods to cGANs while overcoming some issues.\n- The motivation for using distinct architecture to generate images for each class instead of using one architecture is unclear. Most of the existing cGANs in noise-to-image and image-to-image translation settings employ CBN or AdaIN to embed conditional information to a unified generator. \n- The idea is similar to the dynamic routing/inference networks such as Blockdrop [1], and it needs a related discussion about the difference.\n- The experiments are only conducted on low-resolution datasets. In my view, the GANs search algorithm needs to be verified on high-resolution datasets, instead of still continuing to achieve marginal performance improvements on small datasets.\n\nWu, Zuxuan, et al. \"Blockdrop: Dynamic inference paths in residual networks.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall a good work, but the motivation is not quite convincing, and the novelty is somewhat limited. Experiments on higher-resolution datasets are expected.", "review": "This paper proposes an interesting method that adopts NAS to search multiple class-aware generator architectures for cGAN instead of class-agnostic type. A search space containing both normal and class-modulated convolutions are introduced to simplify the process of re-training. Besides, this paper design a mixed-architecture optimization to specifically address the computational burden issue under the setting of a multi-net search. The search results also give some insights about constructing cGAN models.\n\n\nStrengths:\n- The perspective of adopting the NAS method to explore the class-aware generator architectures for conditional GANs is relatively novel and interesting, although there are some related works about searching unconditional GANs. \n-The proposed flexibility and safety search space is effective to address the categories grow issue. \n- The developed mixed-architecture optimization is a clever way to improve the efficiency of the search and re-training process.\n- The authors conduct extensive experiments and give some interesting insight/discussion about the results.\n\nWeaknesses:\n- There are no innovative approaches toward neural architecture search are proposed, and this work only focuses on how to bring existing RL-based NAS methods to cGANs while overcoming some issues.\n- The motivation for using distinct architecture to generate images for each class instead of using one architecture is unclear. Most of the existing cGANs in noise-to-image and image-to-image translation settings employ CBN or AdaIN to embed conditional information to a unified generator. \n- The idea is similar to the dynamic routing/inference networks such as Blockdrop [1], and it needs a related discussion about the difference.\n- The experiments are only conducted on low-resolution datasets. In my view, the GANs search algorithm needs to be verified on high-resolution datasets, instead of still continuing to achieve marginal performance improvements on small datasets.\n\nWu, Zuxuan, et al. \"Blockdrop: Dynamic inference paths in residual networks.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603718113711}, {"id": "VPt_l9eCeut", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2548/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\nThis paper designs class-aware generators (increasing flexibility) for cGAN by RL-based neural architecture search (NAS) algorithm. GAN models with better flexibility are promising to yield better performances.\n\nClass-Modulated convolution (CMconv) is designed to increase model flexibility while enabling data sharing among classes, increasing the efficiency of training data.\n\nMixed-architecture optimization is presented to ease the training procedure of multiple class-aware generators.\n\nCons:\n\nThe proposed Class-Modulated Conv also inserts the class embedding to a common convolution, which is similar to a regular BN. The architectures for different categories are still weight-sharing, which is quite a common approach in NAS. Thus, this work may be treated as applying NAS to cGAN.\n\nFor searching network for GANs, the main challenge lies in how to provide stable and efficient supervision as a reward. Note that during training, the generator (G) and discriminator (D) play as rivals. Searching for the architectures of G and choosing IS as a reward only helps the G compete against D. I wonder whether it is the optimal choice, so the authors need to consider the settings in [1], i.e., updating the architectures for both G and D.\n\nMany NAS algorithms for GAN models are relevant to this work, but none of them is evaluated against the proposed method in the experiments. I think the authors should add competing results from AdversarialNAS [1] and AutoGAN. Specifically, the Intra FIDs on CIFAR10, FID, and IS scores on CIFAR100 of the two methods should be reported.\n\nNeither time complexity nor space complexity is reported. Considering that the resolution of outputs is only 32$\\times$32, I wonder whether the proposed method is prohibitively expensive to be applied to practical applications that require higher resolution. Note that the resolution for most generation tasks is at least 256$\\times$256 or 128$\\times$128. I suggest the authors report the training and testing times, as well as the consumed GPU resources.\n\n \n[1] Gao, Chen, Yunpeng Chen, Si Liu, Zhenxiong Tan, and Shuicheng Yan. \"Adversarialnas: Adversarial neural architecture search for gans.\" CVPR, pp. 5680-5689. 2020.\n\n[2] Gong, Xinyu, Shiyu Chang, Yifan Jiang, and Zhangyang Wang. \"Autogan: Neural architecture search for generative adversarial networks.\" ICCV, pp. 3224-3234. 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper applies Neural Architecture Search to Conditional GAN. The Class-Modulated Conv is proposed to condition the search space.", "review": "Pros:\nThis paper designs class-aware generators (increasing flexibility) for cGAN by RL-based neural architecture search (NAS) algorithm. GAN models with better flexibility are promising to yield better performances.\n\nClass-Modulated convolution (CMconv) is designed to increase model flexibility while enabling data sharing among classes, increasing the efficiency of training data.\n\nMixed-architecture optimization is presented to ease the training procedure of multiple class-aware generators.\n\nCons:\n\nThe proposed Class-Modulated Conv also inserts the class embedding to a common convolution, which is similar to a regular BN. The architectures for different categories are still weight-sharing, which is quite a common approach in NAS. Thus, this work may be treated as applying NAS to cGAN.\n\nFor searching network for GANs, the main challenge lies in how to provide stable and efficient supervision as a reward. Note that during training, the generator (G) and discriminator (D) play as rivals. Searching for the architectures of G and choosing IS as a reward only helps the G compete against D. I wonder whether it is the optimal choice, so the authors need to consider the settings in [1], i.e., updating the architectures for both G and D.\n\nMany NAS algorithms for GAN models are relevant to this work, but none of them is evaluated against the proposed method in the experiments. I think the authors should add competing results from AdversarialNAS [1] and AutoGAN. Specifically, the Intra FIDs on CIFAR10, FID, and IS scores on CIFAR100 of the two methods should be reported.\n\nNeither time complexity nor space complexity is reported. Considering that the resolution of outputs is only 32$\\times$32, I wonder whether the proposed method is prohibitively expensive to be applied to practical applications that require higher resolution. Note that the resolution for most generation tasks is at least 256$\\times$256 or 128$\\times$128. I suggest the authors report the training and testing times, as well as the consumed GPU resources.\n\n \n[1] Gao, Chen, Yunpeng Chen, Si Liu, Zhenxiong Tan, and Shuicheng Yan. \"Adversarialnas: Adversarial neural architecture search for gans.\" CVPR, pp. 5680-5689. 2020.\n\n[2] Gong, Xinyu, Shiyu Chang, Yifan Jiang, and Zhangyang Wang. \"Autogan: Neural architecture search for generative adversarial networks.\" ICCV, pp. 3224-3234. 2019.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603112769180}], "openreview_url": "https://openreview.net/forum?id=Twm9LnWK-zt", "arxiv_id": "2006.14208", "paper_pdf": "papers/Twm9LnWK-zt.pdf", "paper_pdf_sha256": "a63d0d3afd49b4abda88b0e68437a8f7730ea351ddfe82512a114c201942dc79", "paper_pdf_bytes": 1401870, "paper_pdf_source": "openreview", "code_url": "https://github.com/PeterouZh/NAS_cGAN", "code_repository": "PeterouZh/NAS_cGAN", "code_commit": "2f7819aee0ddc0f97e4d8084099b656b7c7a36a9", "code_archive": "repos/Twm9LnWK-zt.zip", "code_archive_sha256": "8938b4216b3254d10e47868e5abd08b9696dc038f30d4bbcd69826a4ac0d017e", "code_archive_bytes": 81268, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 92, "github_languages": {"Python": 305756}, "github_archived": false, "github_pushed_at": "2022-07-02T07:36:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/searching-towards-class-aware-generators-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1lXnhVKPr", "year": 2020, "status": "rejected", "title": "Variance Reduced Local SGD with Lower Communication Complexity", "authors": ["Xianfeng Liang", "Shuheng Shen", "Jingchang Liu", "Zhen Pan", "Yifei Cheng", "Enhong Chen"], "authorids": ["zeroxf@mail.ustc.edu.cn", "vaip@mail.ustc.edu.cn", "jliude@cse.ust.hk", "pzhen@mail.ustc.edu.cn", "chengyif@mail.ustc.edu.cn", "cheneh@ustc.edu.cn"], "authors_source": "OpenReview API", "abstract": "To accelerate the training of machine learning models, distributed stochastic gradient descent (SGD) and its variants have been widely adopted, which apply multiple workers in parallel to speed up training. Among them, Local SGD has gained much attention due to its lower communication cost. Nevertheless, when the data distribution on workers is non-identical, Local SGD requires $O(T^{\\frac{3}{4}} N^{\\frac{3}{4}})$ communications to maintain its \\emph{linear iteration speedup} property, where $T$ is the total number of iterations and $N$ is the number of workers. In this paper, we propose Variance Reduced Local SGD (VRL-SGD) to further reduce the communication complexity. Benefiting from eliminating the dependency on the gradient variance among workers, we theoretically prove that VRL-SGD achieves a \\emph{linear iteration speedup} with a lower communication complexity $O(T^{\\frac{1}{2}} N^{\\frac{3}{2}})$ even if workers access non-identical datasets. We conduct experiments on three machine learning tasks, and the experimental results demonstrate that VRL-SGD performs impressively better than Local SGD when the data among workers are quite diverse.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJelZJgJ9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper182/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tackles the problem of data-parallel (synchronous) distributed SGD, to optimize the (finite) sum of N non-convex, possibly different (in the so-called non-identical case), loss functions. This paper focuses on improving the communication efficiency compared to several existing methods tackling this problem.\n\nTo that end, the authors contribute:\n· A novel algorithm and its asymptotic communication complexity.\n· The proof that the common metric of the sum over the training steps of the expected squared norm of the gradient at the average of the N parameters is bounded above.\n· Experimental results (training loss function of epoch number) comparing this algorithm with 2 existing ones, solving 3 problems under reasonable settings.\n\n  The training time per epoch of VRL-SGD is claimed to be identical to the one of Local SGD, as the algorithm only have minor differences.\n\n- strengths of the paper: \n\n· The main paper is very easy to follow.\n· Good effort to give intuitions on why VLR-SGD can improve the convergence rate of existing algorithms.\n  Such an effort is to be highlighted.\n· No obvious mistake in the main.   I have not thoroughly checked the full proof though.\n\n-  weaknesses of the paper: \n\n· The algorithm, while having differences, is quite reminiscent of Elastic Averaging SGD (EASGD) [1].\n  Indeed in both algorithms the model update at the workers consists in both descending the local gradient plus descending toward some \"moving-average\"obtained through averaging all the local models.\n  In EASGD, this \"moving-average\" is common to every worker and the master, which updates it every k steps.\n  In this paper, each worker has its own \"moving-average\", which update computations are different than in EASGD as the use the instant average of the workers' models instead of the previous \"moving-average\".\n\n[1]Sixin Zhang, Anna Choromanska, Yann LeCun, Deep learning with Elastic Averaging SGD,  NeurIPS, 2015\n\n- Questions I would like the authors to respond to during the rebuttal:\n\n· Could Elastic Averaging SGD (in particular their fastest variant EAMSGD) be applied as-is to solve the non-identical, non-convex optimization problem at hand?\n  Despite the authors of EASGD not studying their algorithm in the non-identical case, following what is done in the intuition part of VRL-SGD (in particular Equation (8)), it seems that the update rule of the \"moving-average\" in EASGD is then equivalent to having a momentumSGD with dampening (instead of the \"generalized SGD form\" obtained with the approach of VRL-SGD).  Hence my question.\n\nI suggest acceptance. However I'm willing to change my opinion after reading other more qualified reviewers in the sub-area of variance-reduction techniques.\n\nnote: If EASGD was to be sound in the non-identical case as well, my decision would not change much.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper tackles the problem of data-parallel (synchronous) distributed SGD, to optimize the (finite) sum of N non-convex, possibly different (in the so-called non-identical case), loss functions. This paper focuses on improving the communication efficiency compared to several existing methods tackling this problem.\n\nTo that end, the authors contribute:\n· A novel algorithm and its asymptotic communication complexity.\n· The proof that the common metric of the sum over the training steps of the expected squared norm of the gradient at the average of the N parameters is bounded above.\n· Experimental results (training loss function of epoch number) comparing this algorithm with 2 existing ones, solving 3 problems under reasonable settings.\n\n  The training time per epoch of VRL-SGD is claimed to be identical to the one of Local SGD, as the algorithm only have minor differences.\n\n- strengths of the paper: \n\n· The main paper is very easy to follow.\n· Good effort to give intuitions on why VLR-SGD can improve the convergence rate of existing algorithms.\n  Such an effort is to be highlighted.\n· No obvious mistake in the main.   I have not thoroughly checked the full proof though.\n\n-  weaknesses of the paper: \n\n· The algorithm, while having differences, is quite reminiscent of Elastic Averaging SGD (EASGD) [1].\n  Indeed in both algorithms the model update at the workers consists in both descending the local gradient plus descending toward some \"moving-average\"obtained through averaging all the local models.\n  In EASGD, this \"moving-average\" is common to every worker and the master, which updates it every k steps.\n  In this paper, each worker has its own \"moving-average\", which update computations are different than in EASGD as the use the instant average of the workers' models instead of the previous \"moving-average\".\n\n[1]Sixin Zhang, Anna Choromanska, Yann LeCun, Deep learning with Elastic Averaging SGD,  NeurIPS, 2015\n\n- Questions I would like the authors to respond to during the rebuttal:\n\n· Could Elastic Averaging SGD (in particular their fastest variant EAMSGD) be applied as-is to solve the non-identical, non-convex optimization problem at hand?\n  Despite the authors of EASGD not studying their algorithm in the non-identical case, following what is done in the intuition part of VRL-SGD (in particular Equation (8)), it seems that the update rule of the \"moving-average\" in EASGD is then equivalent to having a momentumSGD with dampening (instead of the \"generalized SGD form\" obtained with the approach of VRL-SGD).  Hence my question.\n\nI suggest acceptance. However I'm willing to change my opinion after reading other more qualified reviewers in the sub-area of variance-reduction techniques.\n\nnote: If EASGD was to be sound in the non-identical case as well, my decision would not change much."}, "tcdate": 1571909368014}, {"id": "rJgPsOx6FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper182/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In the paper, the authors propose a variance reduced local SGD and prove its convergence rate. In the experiments, they show that the proposed method converges faster than local SGD.  \n\nThe following are my concerns:\n1) I am concerned about the convergence result, it shows that the convergence of the proposed method has nothing to do with the extent of non-iid. However, it is not correct intuitively. It is easy to imagine that non-iid data will converge slower than iid data. \n\n2) In Corollary 5.2, the convergence result is not related to k. It is false to me.\n\n3) It is not clear in algorithm 1 how the \\delta^{t''} is updated.\n\n4) The assumption in equation (11)  \"When all local model x^t, x\\tau and the average model \\hat x converge to the local minimum x∗\" is not correct when data is non-iid distributed. Suppose x^t and \\hat x is x^*,  and \\Delta^{t''} = 0.  Because data is non-iid, the solution of the local problem is not equal to the global problem, therefore, x^t will go away from x^*. \n\n5) In the experiment, the setting of k should affect the experiment. However, authors don't analyze this parameter.\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "In the paper, the authors propose a variance reduced local SGD and prove its convergence rate. In the experiments, they show that the proposed method converges faster than local SGD.  \n\nThe following are my concerns:\n1) I am concerned about the convergence result, it shows that the convergence of the proposed method has nothing to do with the extent of non-iid. However, it is not correct intuitively. It is easy to imagine that non-iid data will converge slower than iid data. \n\n2) In Corollary 5.2, the convergence result is not related to k. It is false to me.\n\n3) It is not clear in algorithm 1 how the \\delta^{t''} is updated.\n\n4) The assumption in equation (11)  \"When all local model x^t, x\\tau and the average model \\hat x converge to the local minimum x∗\" is not correct when data is non-iid distributed. Suppose x^t and \\hat x is x^*,  and \\Delta^{t''} = 0.  Because data is non-iid, the solution of the local problem is not equal to the global problem, therefore, x^t will go away from x^*. \n\n5) In the experiment, the setting of k should affect the experiment. However, authors don't analyze this parameter.\n\n\n\n"}, "tcdate": 1571780766526}, {"id": "HyxV4BPPtB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper182/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes the variance reduction to the local SGD algorithm and shows the proposed VRL-SGD can achieve better convergence than local SGD when the data are not identical over workers. The idea is interesting and the paper is easy to follow. However, the study does not go into depth. I do not support the acceptance for current situation. \n\n1. The paper does not show the convergence rate when the function is convex/strongly convex, which make it hard to compare with previous works e.g. EXTRA [Shi et al. 2015]. \n2. There are many alternatives to achieve communication efficiency. The paper does not argue why to choose variance reduced Local SGD. A natural idea to reduce the variance is to distribute the inner loop of SVRG to different workers as proposed in [Konecny et al. 2016]. Moreover, the analysis is given in [Lee et al. 2015] and [Cen et al. 2019]. Especially, [Cen et al. 2019] suggests using a regularization term to handle the data load is not balanced over the workers. \n3. The paper states that S-SGD cannot achieve linear iteration speedup due to communication bottleneck. Can we avoid the communication bottleneck by increasing the batch size? There are vast literatures on distributed training of deep neural network by using large batch size.  \n4. The experiment comparison is not complete given there are many related work in this area.\n\nQuestion: How to determine the number of local SGD steps for each communication round? In SVRG, the number of iterations in the inner loop is related to the condition number (strongly convex case). Does the number of local SGD steps have a similar correspondence\n\n[Jason Lee, et al.]  2015. Distributed Stochastic Variance Reduced Gradient Methods and A Lower Bound for Communication Complexity.\n[Shicong Cen, et al.] 2019 Convergence of Distributed Stochastic Variance Reduced Methods without Sampling Extra Data\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes the variance reduction to the local SGD algorithm and shows the proposed VRL-SGD can achieve better convergence than local SGD when the data are not identical over workers. The idea is interesting and the paper is easy to follow. However, the study does not go into depth. I do not support the acceptance for current situation. \n\n1. The paper does not show the convergence rate when the function is convex/strongly convex, which make it hard to compare with previous works e.g. EXTRA [Shi et al. 2015]. \n2. There are many alternatives to achieve communication efficiency. The paper does not argue why to choose variance reduced Local SGD. A natural idea to reduce the variance is to distribute the inner loop of SVRG to different workers as proposed in [Konecny et al. 2016]. Moreover, the analysis is given in [Lee et al. 2015] and [Cen et al. 2019]. Especially, [Cen et al. 2019] suggests using a regularization term to handle the data load is not balanced over the workers. \n3. The paper states that S-SGD cannot achieve linear iteration speedup due to communication bottleneck. Can we avoid the communication bottleneck by increasing the batch size? There are vast literatures on distributed training of deep neural network by using large batch size.  \n4. The experiment comparison is not complete given there are many related work in this area.\n\nQuestion: How to determine the number of local SGD steps for each communication round? In SVRG, the number of iterations in the inner loop is related to the condition number (strongly convex case). Does the number of local SGD steps have a similar correspondence\n\n[Jason Lee, et al.]  2015. Distributed Stochastic Variance Reduced Gradient Methods and A Lower Bound for Communication Complexity.\n[Shicong Cen, et al.] 2019 Convergence of Distributed Stochastic Variance Reduced Methods without Sampling Extra Data\n"}, "tcdate": 1571415339942}], "openreview_url": "https://openreview.net/forum?id=S1lXnhVKPr", "arxiv_id": "1912.12844", "paper_pdf": "papers/S1lXnhVKPr.pdf", "paper_pdf_sha256": "6dbdfdb83c30b49a906328f033c803db678cf03f290ed8f6c225c4fd54c585c2", "paper_pdf_bytes": 734475, "paper_pdf_source": "openreview", "code_url": "https://github.com/zerolxf/VRL-SGD", "code_repository": "zerolxf/VRL-SGD", "code_commit": "8fb56dad507239f760b2756887c23404f0985fca", "code_archive": "repos/S1lXnhVKPr.zip", "code_archive_sha256": "5eecbcc041643e2da825035a5729e9dba4d2ef37e07f52898acaac2f82dce6b4", "code_archive_bytes": 200321, "code_file_count": 12, "code_extensions": {".py": 10, ".sh": 2}, "github_disk_usage_kb": 142, "github_languages": {"Python": 60210, "Shell": 3657}, "github_archived": false, "github_pushed_at": "2020-05-20T03:30:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variance-reduced-local-sgd-with-lower-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJxfm2CqKm", "year": 2019, "status": "rejected", "title": "Discovering General-Purpose Active Learning Strategies", "authors": ["Ksenia Konyushkova", "Raphael Sznitman", "Pascal Fua"], "authorids": ["ksenia.konyushkova@epfl.ch", "raphael.sznitman@artorg.unibe.ch", "pascal.fua@epfl.ch"], "authors_source": "OpenReview API", "abstract": "We propose a general-purpose approach to discovering active learning (AL) strategies from data. These strategies are transferable from one domain to another and can be used in conjunction with many machine learning models. To this end, we formalize the annotation process as a Markov decision process, design universal state and action spaces and introduce a new reward function that precisely reflects the AL objective of minimizing the annotation cost We seek to find an optimal (non-myopic) AL strategy using reinforcement learning. We evaluate the learned strategies on multiple unrelated domains and show that they consistently outperform state-of-the-art baselines.", "decision": null, "meta_review": "Reject", "num_reviews": 4, "reviews": [{"id": "HkeE1fkTpQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1336/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper presents an RL approach to active learning that is generic across ML model being learned, and across dataset being used. The paper formulates the standard active learning problem as an MDP with the objective of minimizing the number of annotated labels required to meet a pre-specified prediction quality. \n\nThe MDP state proposed by this paper is the current performance score on each sample in a hold-out set. The actions are specified by selecting a datapoint from the set of all un-annotated datapoints. The action feature vector consists of the current performance score of the model on the datapoint, and the average distance of that datapoint from every datapoint in the labeled set and every datapoint in the unlabeled set.\n\nReview:\nI do not recommend this paper for publication in ICLR because I believe:\n1) the work is too incremental\n2) the comparison to baseline and competing methods is incomplete\n3) some design decisions of the proposed method are not well motivated.\n\nI appreciated the clarity of the writting, and the paper organization. I also believe that the proposed method is quite intuitive, and is a good addition to the field. Finally, I appreciate that sufficient experimental details are available within the paper to be able to easily reproduce the results.\n\nDetails:\nMy points (1) and (2) are highly related, so I will discuss both simultaneously. I find that this paper makes only incremental forward progress from the Pang 2018 paper and the Konyushkova 2017 paper. The methodology here looks very similar to the SingleRL method, which Pang 2018 notes can be considered a special case of Konyushkova 2017's method. I think that the work in this paper would be sufficient to stand on its own if it performed a convincing comparison to SingleRL and/or MLP-GAL from Pang 2018. I recognize that this paper references why no such comparison currently exists, but I think this comparison would be extremely valuable to the paper.\n\nA further comment on my point (2), I do not find the comparisons to baseline methods to be entirely convincing. Of note, only the average performance for each method is reported. I'm curious of the variance---and more specifically the standard error and number of independent runs---of each of the reported results. On many of the datasets, the performance difference between the proposed method and uncertainty sampling is quite small in table 1.\n\nA final comment on point (2): I would have liked to see more exploration of different models. I think table 2 is quite informative, showing notable differences between simple baseline AL methods. I would have liked to see table 2 with more classifiers and with more competing AL methods. Because logistic regression is a simple model, the differences between AL methods may be more subtle. Perhaps a more complex model (say a single hidden layer NN) would show more notable differences.\n\nFor point (3), I would have liked to see either an exploration of other design decisions or an explanation of given design decisions. For instance, why only use 30 hold-out samples for the state? I imagine the proposed method would be fairly sensitive to this choice.  Another unexplained design decision was using a maximum budget of 100 datapoints. Table 2 shows some extremely interesting interactions with this budget in its comparison between LogReg-100 and LogReg-200, and further explanation would have been useful. Finally, I would have liked to see some motivation for choice of stopping condition. Using the stopping condition of 98% of maximum performance may have some biasing effect of each method, and it would helpful to have some motivation behind this choice.\n\nQuestions:\n - Why did uncertainty sampling have such limited benefits on LogReg-200 in table 2? This was a surprising result to me, as uncertainty sampling consistently outperformed most other methods.\n - Why is there a disparity between the results for the SVM in table 2 and the discussion in the first paragraph of section 4.3?\n - How does choice of final performance metric affect all methods? Choosing final performance to be 98% of maximum performance could have a major effect on each method. Because the proposed method is non-myopic, I would expect that it performs well when this value is large but would perform poorly with a smaller percentage of maximum performance.\n - Is the proposed method sensitive to number of samples used to compute the state?\n - What does figure 1 show? Are the same 30 samples used for all three subfigures? Perhaps this would more interpretable if, instead of showing the predicted class, this figure showed the prediction error.\n\nMinor nitpicks (did not influence decision):\n - The datasets are 1-based indexed sometimes and 0-based indexed sometimes, even with disparities within a single paragraph.\n - Figure 1 appears a long time before it is discussed, which made it difficult to understand what was going on.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An intuitive combination of reinforcement learning and active learning.", "review": "Summary:\nThis paper presents an RL approach to active learning that is generic across ML model being learned, and across dataset being used. The paper formulates the standard active learning problem as an MDP with the objective of minimizing the number of annotated labels required to meet a pre-specified prediction quality. \n\nThe MDP state proposed by this paper is the current performance score on each sample in a hold-out set. The actions are specified by selecting a datapoint from the set of all un-annotated datapoints. The action feature vector consists of the current performance score of the model on the datapoint, and the average distance of that datapoint from every datapoint in the labeled set and every datapoint in the unlabeled set.\n\nReview:\nI do not recommend this paper for publication in ICLR because I believe:\n1) the work is too incremental\n2) the comparison to baseline and competing methods is incomplete\n3) some design decisions of the proposed method are not well motivated.\n\nI appreciated the clarity of the writting, and the paper organization. I also believe that the proposed method is quite intuitive, and is a good addition to the field. Finally, I appreciate that sufficient experimental details are available within the paper to be able to easily reproduce the results.\n\nDetails:\nMy points (1) and (2) are highly related, so I will discuss both simultaneously. I find that this paper makes only incremental forward progress from the Pang 2018 paper and the Konyushkova 2017 paper. The methodology here looks very similar to the SingleRL method, which Pang 2018 notes can be considered a special case of Konyushkova 2017's method. I think that the work in this paper would be sufficient to stand on its own if it performed a convincing comparison to SingleRL and/or MLP-GAL from Pang 2018. I recognize that this paper references why no such comparison currently exists, but I think this comparison would be extremely valuable to the paper.\n\nA further comment on my point (2), I do not find the comparisons to baseline methods to be entirely convincing. Of note, only the average performance for each method is reported. I'm curious of the variance---and more specifically the standard error and number of independent runs---of each of the reported results. On many of the datasets, the performance difference between the proposed method and uncertainty sampling is quite small in table 1.\n\nA final comment on point (2): I would have liked to see more exploration of different models. I think table 2 is quite informative, showing notable differences between simple baseline AL methods. I would have liked to see table 2 with more classifiers and with more competing AL methods. Because logistic regression is a simple model, the differences between AL methods may be more subtle. Perhaps a more complex model (say a single hidden layer NN) would show more notable differences.\n\nFor point (3), I would have liked to see either an exploration of other design decisions or an explanation of given design decisions. For instance, why only use 30 hold-out samples for the state? I imagine the proposed method would be fairly sensitive to this choice.  Another unexplained design decision was using a maximum budget of 100 datapoints. Table 2 shows some extremely interesting interactions with this budget in its comparison between LogReg-100 and LogReg-200, and further explanation would have been useful. Finally, I would have liked to see some motivation for choice of stopping condition. Using the stopping condition of 98% of maximum performance may have some biasing effect of each method, and it would helpful to have some motivation behind this choice.\n\nQuestions:\n - Why did uncertainty sampling have such limited benefits on LogReg-200 in table 2? This was a surprising result to me, as uncertainty sampling consistently outperformed most other methods.\n - Why is there a disparity between the results for the SVM in table 2 and the discussion in the first paragraph of section 4.3?\n - How does choice of final performance metric affect all methods? Choosing final performance to be 98% of maximum performance could have a major effect on each method. Because the proposed method is non-myopic, I would expect that it performs well when this value is large but would perform poorly with a smaller percentage of maximum performance.\n - Is the proposed method sensitive to number of samples used to compute the state?\n - What does figure 1 show? Are the same 30 samples used for all three subfigures? Perhaps this would more interpretable if, instead of showing the predicted class, this figure showed the prediction error.\n\nMinor nitpicks (did not influence decision):\n - The datasets are 1-based indexed sometimes and 0-based indexed sometimes, even with disparities within a single paragraph.\n - Figure 1 appears a long time before it is discussed, which made it difficult to understand what was going on.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1542414812471}, {"id": "rJxMVWxQTQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1336/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors suggest to model active learning (AL) as a Markov Decision Process to try to learn the best possible AL strategy across related domains. \n\nThe paper is well-written and structured -- although the background section could be expanded. Sec 3 presents the method in a clear and straightforward manner. \n\nMy main concern with regards to the paper is novelty. The authors mention two main contributions, the first one being to defined the AL objective to minimize the number of annotations required to achieve a given prediction quality, instead of maximizing performance given an annotation budget. There has been AL approaches from that perspective in the past (e.g., https://arxiv.org/pdf/1510.02847.pdf). \n\nThe second contribution has to do with a procedure to learn the AL strategy using data from different domains (with available labels). Again, the literature in transfer learning in Reinforcement Learning is extensive and should be discussed. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A Reinforcement Learning approach to Active Learning", "review": "The authors suggest to model active learning (AL) as a Markov Decision Process to try to learn the best possible AL strategy across related domains. \n\nThe paper is well-written and structured -- although the background section could be expanded. Sec 3 presents the method in a clear and straightforward manner. \n\nMy main concern with regards to the paper is novelty. The authors mention two main contributions, the first one being to defined the AL objective to minimize the number of annotations required to achieve a given prediction quality, instead of maximizing performance given an annotation budget. There has been AL approaches from that perspective in the past (e.g., https://arxiv.org/pdf/1510.02847.pdf). \n\nThe second contribution has to do with a procedure to learn the AL strategy using data from different domains (with available labels). Again, the literature in transfer learning in Reinforcement Learning is extensive and should be discussed. ", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541763370010}, {"id": "SJgx1F79hX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1336/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper studies the recently problem of learning active learning (LAL). It sets up a MDP where the the state is determined by the labeled, unlabelled datasets and classifier, the acton is to query a point, the reward is linked to classifier test set performance improvement and the transition is to update the base classifier. Recent Q-learning algorithms are used to perform the optimisation. The results show that it outperforms some classic handcrafted AL algorithms and some prior LAL algorithms. A feature of this paper is that the method is relatively simple compared to some prior LAL methods, and also that it learns policies that can transfer successfully across diverse heterogenous datasets.\n\nStrengths:\n+ Good results. \n+ Nice that it works well while being simpler and faster than prior transferrable method MLP-GAL.\n+ Generally well written.\n+ Fig 4 is interesting.\n\nWeaknesses:\n- Novelty/originality is rather incremental. \n- Experiments are still on toy datasets.\n\nSpecifics:\n1. Novelty: The concept of formulating AL as a MDP for optimisation is now a standard idea. The optimisers used are recent off-the-shelf Q-learners. The result is that this method is similar to a non-myopic extension of LAL (Konyushkova’17) but several papers already did non-myopic AL. In particular it’s very similar to the SingleRL method in (Pang’18). The only differences are smallish design parameters like: slightly different reward function definition, use Q-learning instead of policy-gradient optimiser, and slightly different state featurisation. The improved sample/speed-efficiency vs SingleRL is likely relatively automatic due to use of recent Q-learning optimisers, rather than vanilla PG optimiser of SingleRL. Not clear that benefit comes from something uniquely contributed here. Other limitations of various prior LAL work, such as binary classifier only, are not alleviated here.\n2. Experiments: The experiments are on toy datasets. Particularly given the small novelty, then evaluation should be much more. For example: 1. How well does it work when transferred to a relatively less toy dataset such as CIFAR. 2. To what extent can it transfer across classifiers rather than only across datasets? \n3. The state representation as a sorted list of scores is rather unintuitive. Is there any intuition on what smart decisions the model could be using this to make?\n4. The featurisations used are not very standard: Like the classifier state sorted score list, and the action featurisation (instance score, instance distance to class, instance distance to unlabelled). It would be good to evaluate this featurisation with a supervised active learner (like LAL), in order to disambiguate whether the good performance comes from these feature choices, or from the recent RL algorithms used to optimise. Similarly for the choice of reward function.\n5. How does the proposed method deal with a suite of training datasets for AL that are of greatly varying difficulty. A relatively very easy dataset needing << 100 examples to reach threshold would generate few AL training examples due to early stopping. A very hard dataset might use all 100 examples. Does it mean that easy datasets contribute less to training than hard ones? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "OK paper. Well written, but weak novelty.", "review": "Summary: This paper studies the recently problem of learning active learning (LAL). It sets up a MDP where the the state is determined by the labeled, unlabelled datasets and classifier, the acton is to query a point, the reward is linked to classifier test set performance improvement and the transition is to update the base classifier. Recent Q-learning algorithms are used to perform the optimisation. The results show that it outperforms some classic handcrafted AL algorithms and some prior LAL algorithms. A feature of this paper is that the method is relatively simple compared to some prior LAL methods, and also that it learns policies that can transfer successfully across diverse heterogenous datasets.\n\nStrengths:\n+ Good results. \n+ Nice that it works well while being simpler and faster than prior transferrable method MLP-GAL.\n+ Generally well written.\n+ Fig 4 is interesting.\n\nWeaknesses:\n- Novelty/originality is rather incremental. \n- Experiments are still on toy datasets.\n\nSpecifics:\n1. Novelty: The concept of formulating AL as a MDP for optimisation is now a standard idea. The optimisers used are recent off-the-shelf Q-learners. The result is that this method is similar to a non-myopic extension of LAL (Konyushkova’17) but several papers already did non-myopic AL. In particular it’s very similar to the SingleRL method in (Pang’18). The only differences are smallish design parameters like: slightly different reward function definition, use Q-learning instead of policy-gradient optimiser, and slightly different state featurisation. The improved sample/speed-efficiency vs SingleRL is likely relatively automatic due to use of recent Q-learning optimisers, rather than vanilla PG optimiser of SingleRL. Not clear that benefit comes from something uniquely contributed here. Other limitations of various prior LAL work, such as binary classifier only, are not alleviated here.\n2. Experiments: The experiments are on toy datasets. Particularly given the small novelty, then evaluation should be much more. For example: 1. How well does it work when transferred to a relatively less toy dataset such as CIFAR. 2. To what extent can it transfer across classifiers rather than only across datasets? \n3. The state representation as a sorted list of scores is rather unintuitive. Is there any intuition on what smart decisions the model could be using this to make?\n4. The featurisations used are not very standard: Like the classifier state sorted score list, and the action featurisation (instance score, instance distance to class, instance distance to unlabelled). It would be good to evaluate this featurisation with a supervised active learner (like LAL), in order to disambiguate whether the good performance comes from these feature choices, or from the recent RL algorithms used to optimise. Similarly for the choice of reward function.\n5. How does the proposed method deal with a suite of training datasets for AL that are of greatly varying difficulty. A relatively very easy dataset needing << 100 examples to reach threshold would generate few AL training examples due to early stopping. A very hard dataset might use all 100 examples. Does it mean that easy datasets contribute less to training than hard ones? ", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541187799693}, {"id": "SJxio9y9hX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1336/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a laudable attempt to generalize the learning of active learning strategies to learn general strategies that apply across many different datasets that have variables of different, not pre-determined, types, and apply the learned active learning strategies to datasets that are different from what they have been learned with. The paper is written quite clearly and is clear in its discussion of what its advance is beyond the current state of the art.\n\nUnfortunately, the motivation of the details of the algorithm and the experiment analysis leave the paper short of what is needed to truly assess the value of this area of work and; therefore, short of what is needed for publication in ICLR. The most notable shortcoming is on page 4, at the bottom, where the actions are described. Among the components of the actions are statistics related to the dataset---the average distance from the chosen point to all the labeled data, and the average distance from the chosen point to all the unlabeled data. The authors do not provide a motivation for the use of these particular statistics. Additionally, the authors did not explore any other statistics. I should think that statistics relevant to the sparsity of the data (e.g., how well they cluster). Additionally, what distance measure is being used? A variety of distance metrics should be explored, such as d-separation for continuous variables and Hamming distance for discrete variables, should be tested, as they intuitively seem likely to affect the results. Additionally, many values are chosen for the experiments without motivation and without testing a variety of values (e.g., 30 for the size of the dataset used to calculate the reward, 1000 RL iterations, and others).\n\nIn the experiments, there needs to be discussion of how much variety there is in the different datasets in terms of their statistical properties that are relevant to active learning, such as how well the data cluster? That would help in understanding why the new algorithm performs as it does relative to the baseline.\n\nOne relatively minor point: The authors state on page 3, \"For example, the probability that the classifier assigns to a datapoint suits this purpose because most classifiers estimate this value.\" This is a bit misleading---only generative classifiers would do this, not discriminative classifiers.\n\nPros:\n1. Very clear writing.\n2. Good motivation for the general problem.\n3. Precise description of algorithm.\n\nCons:\n1. Poor motivation for the particular algorithm implementation---features used in the actions, parameter values chosen.\n2. Lack of experiments with different choices for features and parameter values.\n3. Lack of assessment of the dataset characteristics and how they relate to algorithm performance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper describes the use of reinforcement learning to learn active learning strategies. This paper attempts to increase the scope of the learning of active learning strategies to transfer across very different datasets.", "review": "This paper presents a laudable attempt to generalize the learning of active learning strategies to learn general strategies that apply across many different datasets that have variables of different, not pre-determined, types, and apply the learned active learning strategies to datasets that are different from what they have been learned with. The paper is written quite clearly and is clear in its discussion of what its advance is beyond the current state of the art.\n\nUnfortunately, the motivation of the details of the algorithm and the experiment analysis leave the paper short of what is needed to truly assess the value of this area of work and; therefore, short of what is needed for publication in ICLR. The most notable shortcoming is on page 4, at the bottom, where the actions are described. Among the components of the actions are statistics related to the dataset---the average distance from the chosen point to all the labeled data, and the average distance from the chosen point to all the unlabeled data. The authors do not provide a motivation for the use of these particular statistics. Additionally, the authors did not explore any other statistics. I should think that statistics relevant to the sparsity of the data (e.g., how well they cluster). Additionally, what distance measure is being used? A variety of distance metrics should be explored, such as d-separation for continuous variables and Hamming distance for discrete variables, should be tested, as they intuitively seem likely to affect the results. Additionally, many values are chosen for the experiments without motivation and without testing a variety of values (e.g., 30 for the size of the dataset used to calculate the reward, 1000 RL iterations, and others).\n\nIn the experiments, there needs to be discussion of how much variety there is in the different datasets in terms of their statistical properties that are relevant to active learning, such as how well the data cluster? That would help in understanding why the new algorithm performs as it does relative to the baseline.\n\nOne relatively minor point: The authors state on page 3, \"For example, the probability that the classifier assigns to a datapoint suits this purpose because most classifiers estimate this value.\" This is a bit misleading---only generative classifiers would do this, not discriminative classifiers.\n\nPros:\n1. Very clear writing.\n2. Good motivation for the general problem.\n3. Precise description of algorithm.\n\nCons:\n1. Poor motivation for the particular algorithm implementation---features used in the actions, parameter values chosen.\n2. Lack of experiments with different choices for features and parameter values.\n3. Lack of assessment of the dataset characteristics and how they relate to algorithm performance.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541171875232}], "openreview_url": "https://openreview.net/forum?id=HJxfm2CqKm", "arxiv_id": "1810.04114", "paper_pdf": "papers/HJxfm2CqKm.pdf", "paper_pdf_sha256": "f51dad5ce5325b214273c3f0c8441de6c539d85451f575a00994ec574bd27783", "paper_pdf_bytes": 424128, "paper_pdf_source": "openreview", "code_url": "https://github.com/ksenia-konyushkova/LAL-RL", "code_repository": "ksenia-konyushkova/LAL-RL", "code_commit": "6377fb0e5b6b64b62b0c3bc6e26ed4459138e1b8", "code_archive": "repos/HJxfm2CqKm.zip", "code_archive_sha256": "773efb64e773d732bfd1138c5c13aae1a511440cb3daec5cbec8e5236312a67d", "code_archive_bytes": 1458041, "code_file_count": 59, "code_extensions": {".py": 51, ".c": 3, ".ipynb": 2, ".h": 2, ".cpp": 1}, "github_disk_usage_kb": 1398, "github_languages": {"Python": 208230, "Jupyter Notebook": 41641, "C++": 37716, "C": 24287, "Cython": 5194}, "github_archived": false, "github_pushed_at": "2022-05-17T21:20:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discovering-general-purpose-active-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "knc1y6jkTL", "year": 2026, "status": "rejected", "title": "Accelerating Newton-Schulz Iteration for Orthogonalization via Chebyshev-type Polynomials", "authors": ["Ekaterina Grishina", "Matvey Smirnov", "Maxim Rakhuba"], "authorids": ["~Ekaterina_Grishina1", "~Matvey_Smirnov1", "~Maxim_Rakhuba1"], "authors_source": "OpenReview API", "abstract": "The problem of computing optimal orthogonal approximation to a given matrix has attracted growing interest in machine learning. Notable applications include the recent Muon optimizer or Riemannian optimization on the Stiefel manifold. Among existing approaches, the Newton-Schulz iteration has emerged as a particularly effective solution, as it relies solely on matrix multiplications and thus achieves high computational efficiency on GPU hardware. Despite its efficiency, the method has inherent limitations—its coefficients are fixed and thus not optimized for a given matrix. In this paper we address this issue by proposing a Chebyshev-optimized version of Newton-Schulz (CANS). Based on the Chebyshev's alternance theorem, we theoretically derive optimal coefficients for the 3-rd order Newton-Schulz iteration and apply a Remez algorithm to compute optimal higher-degree polynomials. We leverage these polynomials to construct controlled approximate orthogonalization schemes, which is of interest in deep learning applications. Practically, we demonstrate the method's effectiveness in two key applications: orthogonalization in the Muon optimizer, and providing an efficient retraction alternative for Riemannian optimization on the Stiefel manifold.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "UFAkVWN7th", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19093/Reviewer_6zcm"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "Authors look into finding optimal polynomials for Newton-Schulz iterations that make matrices approximately orthogonal, which is a key step in the Muon optimizer for LLMs and is useful in optimization over the Stiefel manifold. They also propose to use an alternative cheap pre-normalization for the algorithm to work differing from the Frobenius norm. Experimental results show an improvement over previous approaches", "review_text": "Authors look into finding optimal polynomials for Newton-Schulz iterations that make matrices approximately orthogonal, which is a key step in the Muon optimizer for LLMs and is useful in optimization over the Stiefel manifold. They also propose to use an alternative cheap pre-normalization for the algorithm to work differing from the Frobenius norm. Experimental results show an improvement over previous approaches", "strengths": "Through rigorous theory, optimal polynomials are found for this task (sections 3 and 4)\n\nA trick is proposed for using Gelfand's formula with almost no computational overhead in order to get accurate upper bounds on the spectral norm.", "weaknesses": "The theory seems to me to be more like a corollary of prior studies, but this does not necessarily undermine the value of this approach in this context.\n\nPolynomials better fit to the task\n\nSee the questions/suggestions section for more.\n\ntypo in appendix K, it references Figure 1 instead of Figure 2", "questions": "Experiments with and without the Gelfand bound were performed, but I am missing info on directly what the Gelfand bound was versus the Frobenius norm approach to see directly what advantage it is providing for the initial estimate. \n\nAlso, it would have been good to isolate the effect of your different polynomials vs Gelfand's formula (you run Muon with the Frobenius norm bound and your approach is run using Gelfand's so both effects are mixed)\n\nThere could be more explanations on Figure 5. Several plots are show by number of steps despite that each line curve used a different number of matrix multiplications per step. The only comparable ones are Muon vs Cans order=5 iter=4, (this was mentioned in the text) and also Jiacheng iter 6 mm=18 vs CANS order 3 iter=9. This ploto should have been drawn with number of matrix multiplications in the x axis directly, not number of steps.\n\nAlso, a few more comparisons would be needed with Jiacheng work in order to see that this comparison was not hand picked", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors look into finding optimal polynomials for Newton-Schulz iterations that make matrices approximately orthogonal, which is a key step in the Muon optimizer for LLMs and is useful in optimization over the Stiefel manifold. They also propose to use an alternative cheap pre-normalization for the algorithm to work differing from the Frobenius norm. Experimental results show an improvement over previous approaches", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "Through rigorous theory, optimal polynomials are found for this task (sections 3 and 4)\n\nA trick is proposed for using Gelfand's formula with almost no computational overhead in order to get accurate upper bounds on the spectral norm.", "weaknesses": "The theory seems to me to be more like a corollary of prior studies, but this does not necessarily undermine the value of this approach in this context.\n\nPolynomials better fit to the task\n\nSee the questions/suggestions section for more.\n\ntypo in appendix K, it references Figure 1 instead of Figure 2", "questions": "Experiments with and without the Gelfand bound were performed, but I am missing info on directly what the Gelfand bound was versus the Frobenius norm approach to see directly what advantage it is providing for the initial estimate. \n\nAlso, it would have been good to isolate the effect of your different polynomials vs Gelfand's formula (you run Muon with the Frobenius norm bound and your approach is run using Gelfand's so both effects are mixed)\n\nThere could be more explanations on Figure 5. Several plots are show by number of steps despite that each line curve used a different number of matrix multiplications per step. The only comparable ones are Muon vs Cans order=5 iter=4, (this was mentioned in the text) and also Jiacheng iter 6 mm=18 vs CANS order 3 iter=9. This ploto should have been drawn with number of matrix multiplications in the x axis directly, not number of steps.\n\nAlso, a few more comparisons would be needed with Jiacheng work in order to see that this comparison was not hand picked", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761845173844}, {"id": "DmHwkQjvjB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19093/Reviewer_H3Bj"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper addresses the problem of computing optimal orthogonal approximations to matrices, a fundamental operation in machine learning applications such as the Muon optimizer and Riemannian optimization on the Stiefel manifold. The authors propose CANS (Chebyshev-accelerated Newton-Schulz), which optimizes the coefficients of Newton-Schulz iterations using Chebyshev's alternance theorem. For degree-3 polynomials, they derive explicit optimal formulas, and for higher degrees, they apply the Remez algorithm. The method is demonstrated on three applications: matrix orthogonalization, the Muon optimizer for neural network training, and retraction on the Stiefel manifold for Riemannian optimization.", "review_text": "This paper addresses the problem of computing optimal orthogonal approximations to matrices, a fundamental operation in machine learning applications such as the Muon optimizer and Riemannian optimization on the Stiefel manifold. The authors propose CANS (Chebyshev-accelerated Newton-Schulz), which optimizes the coefficients of Newton-Schulz iterations using Chebyshev's alternance theorem. For degree-3 polynomials, they derive explicit optimal formulas, and for higher degrees, they apply the Remez algorithm. The method is demonstrated on three applications: matrix orthogonalization, the Muon optimizer for neural network training, and retraction on the Stiefel manifold for Riemannian optimization.", "strengths": "1. The work presents a novel theoretical framework for optimizing Newton-Schulz iteration coefficients. While polynomial approximation theory is classical, its systematic application to this problem through Chebyshev's alternance theorem is creative.\n\n2. The theoretical contributions are rigorous with complete proofs in the appendices. Proposition 2 provides closed-form solutions for degree-3 polynomials, and the convergence analysis establishing quadratic convergence is solid. The experimental validation spans multiple domains with appropriate baselines. \n\n3. The paper is generally well-written with clear motivation.", "weaknesses": "1. The paper cites concurrent work by Amsel et al. (2025). A slightly more detailed comparison in the related work section could help readers more clearly understand the overlapping and distinct contributions of the two papers regarding the exact case.", "questions": "1. Section 3.3 states that Gelfand's formula to estimate $\\sigma_1$ can be implemented \"without introducing any extra matmuls\". However, normalization of matrix is applized before NS iteration, does that mean $(A^T A)^k $ must be saved? Clarifying this is important for understanding the practical implementation overhead.\n\n2. Regarding $\\delta$-orthogonalization, \n\n(1) could you elaborate more on why \"If we use distinct polynomials on each iteration, we can achieve more rapid convergence\"? Say, given arbitrary $\\delta \\in (0,1)$, since optimal $a = a(d, \\delta) \\in (0,1-\\delta)$, could we instead do binary search for $a(d, \\delta)$ in region $(0, 1-\\delta)$ and iterates the process $p, \\epsilon = remez(a, 1+\\delta, 2d-1)$ until $|\\epsilon - \\delta| \\le 1e-7$ as follows:\n\n$L = 0; R = 1-\\delta$\nrepeat:\n  a = (L+R)/2\n\n  p, ε = remez(a, 1+$\\delta$, 2d-1)\n\n  if $|\\epsilon-\\delta| ≤ tol$: break\n\n  if $\\epsilon$ > $\\delta$: L = a       \n\n  else:      R = a    \n\n\n(2) Why we need rapid convergence for $\\delta$-orthogonalization? Fast convergence is desirable for exact orthogonalization through optimal odd polynomials. However, the goal of $\\delta$-orthogonalization is to a find a polynomial to push singular values into interval $[1-\\delta, 1+\\delta]$ while with large derivative at the origin $0$. However, the rapid convergence through distinct polynomials over iterations would easily make $\\epsilon$ approaching 0 (as stated in proposition 3), which could easily make $\\epsilon < \\delta$, especially for high tolerance.\n\n3. A high-level question is do we really need optimal/exact orthogonalization in real world applications like Muon optimizer?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of computing optimal orthogonal approximations to matrices, a fundamental operation in machine learning applications such as the Muon optimizer and Riemannian optimization on the Stiefel manifold. The authors propose CANS (Chebyshev-accelerated Newton-Schulz), which optimizes the coefficients of Newton-Schulz iterations using Chebyshev's alternance theorem. For degree-3 polynomials, they derive explicit optimal formulas, and for higher degrees, they apply the Remez algorithm. The method is demonstrated on three applications: matrix orthogonalization, the Muon optimizer for neural network training, and retraction on the Stiefel manifold for Riemannian optimization.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The work presents a novel theoretical framework for optimizing Newton-Schulz iteration coefficients. While polynomial approximation theory is classical, its systematic application to this problem through Chebyshev's alternance theorem is creative.\n\n2. The theoretical contributions are rigorous with complete proofs in the appendices. Proposition 2 provides closed-form solutions for degree-3 polynomials, and the convergence analysis establishing quadratic convergence is solid. The experimental validation spans multiple domains with appropriate baselines. \n\n3. The paper is generally well-written with clear motivation.", "weaknesses": "1. The paper cites concurrent work by Amsel et al. (2025). A slightly more detailed comparison in the related work section could help readers more clearly understand the overlapping and distinct contributions of the two papers regarding the exact case.", "questions": "1. Section 3.3 states that Gelfand's formula to estimate $\\sigma_1$ can be implemented \"without introducing any extra matmuls\". However, normalization of matrix is applized before NS iteration, does that mean $(A^T A)^k $ must be saved? Clarifying this is important for understanding the practical implementation overhead.\n\n2. Regarding $\\delta$-orthogonalization, \n\n(1) could you elaborate more on why \"If we use distinct polynomials on each iteration, we can achieve more rapid convergence\"? Say, given arbitrary $\\delta \\in (0,1)$, since optimal $a = a(d, \\delta) \\in (0,1-\\delta)$, could we instead do binary search for $a(d, \\delta)$ in region $(0, 1-\\delta)$ and iterates the process $p, \\epsilon = remez(a, 1+\\delta, 2d-1)$ until $|\\epsilon - \\delta| \\le 1e-7$ as follows:\n\n$L = 0; R = 1-\\delta$\nrepeat:\n  a = (L+R)/2\n\n  p, ε = remez(a, 1+$\\delta$, 2d-1)\n\n  if $|\\epsilon-\\delta| ≤ tol$: break\n\n  if $\\epsilon$ > $\\delta$: L = a       \n\n  else:      R = a    \n\n\n(2) Why we need rapid convergence for $\\delta$-orthogonalization? Fast convergence is desirable for exact orthogonalization through optimal odd polynomials. However, the goal of $\\delta$-orthogonalization is to a find a polynomial to push singular values into interval $[1-\\delta, 1+\\delta]$ while with large derivative at the origin $0$. However, the rapid convergence through distinct polynomials over iterations would easily make $\\epsilon$ approaching 0 (as stated in proposition 3), which could easily make $\\epsilon < \\delta$, especially for high tolerance.\n\n3. A high-level question is do we really need optimal/exact orthogonalization in real world applications like Muon optimizer?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761833162561}, {"id": "4a6lXXku72", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19093/Reviewer_xmCb"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces the Chebyshev-Accelerated Newton-Schulz (CANS) framework to address the notable drawbacks of conventional matrix orthogonalization methods. The framework derives explicit formulas and analyzes the convergence of optimal odd polynomials, utilizing the Remez algorithm to compute their higher-order counterparts. Compared to classical methods like the Newton-Schulz iteration and Cayley retraction, the polynomials generated by CANS reduce computational complexity and accelerate convergence while maintaining accuracy. Furthermore, the authors validate the method's effectiveness through tasks such as accelerating parameter orthogonalization with the Muon optimizer, training NanoGPT, and performing fast retractions on the Stiefel manifold.", "review_text": "This paper introduces the Chebyshev-Accelerated Newton-Schulz (CANS) framework to address the notable drawbacks of conventional matrix orthogonalization methods. The framework derives explicit formulas and analyzes the convergence of optimal odd polynomials, utilizing the Remez algorithm to compute their higher-order counterparts. Compared to classical methods like the Newton-Schulz iteration and Cayley retraction, the polynomials generated by CANS reduce computational complexity and accelerate convergence while maintaining accuracy. Furthermore, the authors validate the method's effectiveness through tasks such as accelerating parameter orthogonalization with the Muon optimizer, training NanoGPT, and performing fast retractions on the Stiefel manifold.", "strengths": "1.This paper systematically applies Chebyshev approximation theory to optimize the coefficients of the Newton-Schulz iteration, proposing the Chebyshev-Accelerated Newton-Schulz (CANS) framework for finding \"provably optimal\" odd polynomials.\n\n2.The paper is built upon a solid mathematical theory, with detailed and rigorous proofs for each proposition and corollary provided in the appendix. This offers robust mathematical support for the uniqueness and key properties of the optimal odd polynomials.\n\n3.The overall structure of the paper is clear. It begins by stating the problem and progressively develops its main theory, supplemented by figures and textual explanations to help readers understand the meaning behind each proposition.", "weaknesses": "1.The typesetting for the proof of Proposition 1 in the appendix is slightly disorganized and could be improved. Additionally, the paper contains some errors; for instance, in the equation on line 50, the final exponent appears to be a transpose symbol 'T' when it should likely be 't'. The paper also mistakenly includes two distinct algorithms both labeled as \"Algorithm 1\".\n\n2.The experiments on the Stiefel manifold are conducted solely with a Wide ResNet-16-10 on the CIFAR-10 dataset. The evaluation does not cover other architectures  or larger datasets. This limited scope makes it difficult to assess the generalizability of the CANS retraction for deeper networks and larger-scale data, indicating a need for more extensive experiments.", "questions": "Regarding the core parameter $\\delta$, the paper only sets specific values experimentally and does not provide a quantitative guide for its selection. It fails to clarify how one might determine an appropriate δ based on the singular value distribution of a matrix or the specific requirements of a task.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Chebyshev-Accelerated Newton-Schulz (CANS) framework to address the notable drawbacks of conventional matrix orthogonalization methods. The framework derives explicit formulas and analyzes the convergence of optimal odd polynomials, utilizing the Remez algorithm to compute their higher-order counterparts. Compared to classical methods like the Newton-Schulz iteration and Cayley retraction, the polynomials generated by CANS reduce computational complexity and accelerate convergence while maintaining accuracy. Furthermore, the authors validate the method's effectiveness through tasks such as accelerating parameter orthogonalization with the Muon optimizer, training NanoGPT, and performing fast retractions on the Stiefel manifold.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.This paper systematically applies Chebyshev approximation theory to optimize the coefficients of the Newton-Schulz iteration, proposing the Chebyshev-Accelerated Newton-Schulz (CANS) framework for finding \"provably optimal\" odd polynomials.\n\n2.The paper is built upon a solid mathematical theory, with detailed and rigorous proofs for each proposition and corollary provided in the appendix. This offers robust mathematical support for the uniqueness and key properties of the optimal odd polynomials.\n\n3.The overall structure of the paper is clear. It begins by stating the problem and progressively develops its main theory, supplemented by figures and textual explanations to help readers understand the meaning behind each proposition.", "weaknesses": "1.The typesetting for the proof of Proposition 1 in the appendix is slightly disorganized and could be improved. Additionally, the paper contains some errors; for instance, in the equation on line 50, the final exponent appears to be a transpose symbol 'T' when it should likely be 't'. The paper also mistakenly includes two distinct algorithms both labeled as \"Algorithm 1\".\n\n2.The experiments on the Stiefel manifold are conducted solely with a Wide ResNet-16-10 on the CIFAR-10 dataset. The evaluation does not cover other architectures  or larger datasets. This limited scope makes it difficult to assess the generalizability of the CANS retraction for deeper networks and larger-scale data, indicating a need for more extensive experiments.", "questions": "Regarding the core parameter $\\delta$, the paper only sets specific values experimentally and does not provide a quantitative guide for its selection. It fails to clarify how one might determine an appropriate δ based on the singular value distribution of a matrix or the specific requirements of a task.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761662201205}, {"id": "y0LpAmVu5A", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19093/Reviewer_6mHR"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "This paper introduces the 3rd order Newton-Schulz method with optimal coefficients through Chebyshev’s alternance theorem, and higher order extension via Remez algorithm. The authors then apply it to Muon optimizer and to Riemannina optimization on the Stiefel manifold, and shows acceleration from the proposed algorithms on both cases.", "review_text": "This paper introduces the 3rd order Newton-Schulz method with optimal coefficients through Chebyshev’s alternance theorem, and higher order extension via Remez algorithm. The authors then apply it to Muon optimizer and to Riemannina optimization on the Stiefel manifold, and shows acceleration from the proposed algorithms on both cases.", "strengths": "This paper improves the classic Newton-Schultz algorithm and shows acceleration on Muon and Riemannian optimization settings both theoretically and empirically.", "weaknesses": "The paper reads more like a mathematics paper than a machine learning paper. It lacks high-level intuition and is overly detailed, making it difficult to follow. The presentation is not very well-organized.", "questions": "1.In the Muon application, only partial test loss results are shown. It would be helpful to also include training loss curves and full test loss trajectories to verify the theory. Also, please clarify what “optimal” means — is it optimal in the sense of training (optimization) performance?\n\n2.In section 4,\n\n2.1 Equations (5) and (6) are unclear and should be rewritten more rigorously.\n\n2.2 There are two “Algorithm 1” entries — please correct accordingly.\n\n2.3 Lines 276–285 are difficult to interpret. It seems you are discussing an extension that uses different polynomials at each iteration. This should either be formalized as a theorem, lemma, or corollary, or rewritten more intuitively as a high-level remark rather than a sequence of technical statements.\n\n2.4 How are d_i chosen at each step?\n\n3.What is the relationship between 1-\\delta and a? According to Proposition 4, a does not seem to be symmetric w.r.t. 1\n\n4.Please introduce Remez algorithm in the corresponding part. It’s unclear to me how the coefficients of higher-order (>3) polynomials are chosen (as well as the order in the above question).\n\n5.Please introduce the relationship Muon algorithm and Newton-Schultz more formally (at least you should state the whole algorithm of Muon), and then compare the corresponding part with your proposed optimal algorithm.\n\n6.In Figure 3 and 4, the authors show the comparison between their proposed polynomials and original Muon. It seems that whether the polynomial is closer to 1 is not important since Muon has smaller right end point and are closer to 1 in most of the points than the proposed polynomials, and only the derivative at zero matters. How should this be understood?\n\n\nMinor:\n\nLine184: “greater or equal than”\n\nLine 238: “On the other hand”", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the 3rd order Newton-Schulz method with optimal coefficients through Chebyshev’s alternance theorem, and higher order extension via Remez algorithm. The authors then apply it to Muon optimizer and to Riemannina optimization on the Stiefel manifold, and shows acceleration from the proposed algorithms on both cases.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "This paper improves the classic Newton-Schultz algorithm and shows acceleration on Muon and Riemannian optimization settings both theoretically and empirically.", "weaknesses": "The paper reads more like a mathematics paper than a machine learning paper. It lacks high-level intuition and is overly detailed, making it difficult to follow. The presentation is not very well-organized.", "questions": "1.In the Muon application, only partial test loss results are shown. It would be helpful to also include training loss curves and full test loss trajectories to verify the theory. Also, please clarify what “optimal” means — is it optimal in the sense of training (optimization) performance?\n\n2.In section 4,\n\n2.1 Equations (5) and (6) are unclear and should be rewritten more rigorously.\n\n2.2 There are two “Algorithm 1” entries — please correct accordingly.\n\n2.3 Lines 276–285 are difficult to interpret. It seems you are discussing an extension that uses different polynomials at each iteration. This should either be formalized as a theorem, lemma, or corollary, or rewritten more intuitively as a high-level remark rather than a sequence of technical statements.\n\n2.4 How are d_i chosen at each step?\n\n3.What is the relationship between 1-\\delta and a? According to Proposition 4, a does not seem to be symmetric w.r.t. 1\n\n4.Please introduce Remez algorithm in the corresponding part. It’s unclear to me how the coefficients of higher-order (>3) polynomials are chosen (as well as the order in the above question).\n\n5.Please introduce the relationship Muon algorithm and Newton-Schultz more formally (at least you should state the whole algorithm of Muon), and then compare the corresponding part with your proposed optimal algorithm.\n\n6.In Figure 3 and 4, the authors show the comparison between their proposed polynomials and original Muon. It seems that whether the polynomial is closer to 1 is not important since Muon has smaller right end point and are closer to 1 in most of the points than the proposed polynomials, and only the derivative at zero matters. How should this be understood?\n\n\nMinor:\n\nLine184: “greater or equal than”\n\nLine 238: “On the other hand”", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761164367081}], "openreview_url": "https://openreview.net/forum?id=knc1y6jkTL", "arxiv_id": "2506.10935", "paper_pdf": "papers/knc1y6jkTL.pdf", "paper_pdf_sha256": "3b4f2aaf95bf1eca282365aed1c87d4e1594ed87561f582fa2d1cfdfe6aeb67d", "paper_pdf_bytes": 847943, "paper_pdf_source": "openreview", "code_url": "https://github.com/GrishKate/accelerating_orthogonalization", "code_repository": "GrishKate/accelerating_orthogonalization", "code_commit": "3eca916ffa6f46ab3b313dd34c48235322747565", "code_archive": "repos/knc1y6jkTL.zip", "code_archive_sha256": "6e2f0ab727e6f821e077e4c557caa6d1646a2aa97433e7e5ba8e3215c81fb5a8", "code_archive_bytes": 599122, "code_file_count": 10, "code_extensions": {".py": 8, ".sh": 1, ".ipynb": 1}, "github_disk_usage_kb": 582, "github_languages": {"Jupyter Notebook": 791598, "Python": 54029, "Shell": 54}, "github_archived": false, "github_pushed_at": "2025-06-13T15:01:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerating-newton-schulz-iteration-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UFzE9njwMG", "year": 2025, "status": "rejected", "title": "Mitigating Time Discretization Challenges with WeatherODE: A Sandwich Physics-Driven Neural ODE for Weather Forecasting", "authors": ["Peiyuan Liu", "Tian Zhou", "Liang Sun", "Rong Jin"], "authorids": ["~Peiyuan_Liu1", "~Tian_Zhou2", "~Liang_Sun2", "~Rong_Jin1"], "authors_source": "OpenReview API", "abstract": "In the field of weather forecasting, traditional models often grapple with discretization errors and time-dependent source discrepancies, which limit their predictive performance. In this paper, we present WeatherODE, a novel one-stage, physics-driven ordinary differential equation (ODE) model designed to enhance weather forecasting accuracy. By leveraging wave equation theory and integrating a time-dependent source model, WeatherODE effectively addresses the challenges associated with time-discretization error and dynamic atmospheric processes. Moreover, we design a CNN-ViT-CNN sandwich structure, facilitating efficient learning dynamics tailored for distinct yet interrelated tasks with varying optimization biases in advection equation estimation. Through rigorous experiments, WeatherODE demonstrates superior performance in both global and regional weather forecasting tasks, outperforming recent state-of-the-art approaches by significant margins of over 40.0\\% and 31.8\\% in root mean square error (RMSE), respectively. The source code is available at https://anonymous.4open.science/r/WeatherODE-5C13/.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "EV8kXO7WVa", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1411/Reviewer_wKzX"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper is related to weather modeling on a 5.625° resolution using ERA-5 data as target. The solution is based on the NeuralODE model with several new features inside. A CNN-ViT-CNN sandwich architecture is proposed to model the right term of the ODE. The authors claim a strong improvement in simulation quality over the alternative approaches. Such an improvement is attributed to new scheme of velocity derivatives estimation based on the wave equation and to the CNN-based approximation of a source term in the advection equation", "review_text": "The paper is related to weather modeling on a 5.625° resolution using ERA-5 data as target. The solution is based on the NeuralODE model with several new features inside. A CNN-ViT-CNN sandwich architecture is proposed to model the right term of the ODE. The authors claim a strong improvement in simulation quality over the alternative approaches. Such an improvement is attributed to new scheme of velocity derivatives estimation based on the wave equation and to the CNN-based approximation of a source term in the advection equation", "strengths": "* An interesting wave-equation inspired model to capture time derivative of velocity\n* Comparison with several baselines showing better results of a proposed model\n* Ability to interpolate in time and produce a continuous dynamics\n* Ablation studies for the velocity derivative approximation and model architecture", "weaknesses": "* The CNN used in variable space scale loses part of its physical meaning\n* The wave equation doesn’t describe the atmospheric dynamics really well so the derivative approximation is not governed by physics\n* The model has low practical applicability\nProbably an interesting experiment will be to use the model in the autoregressive mode and to assess its quality for longer forecast horizons. The same can be done for regional models to assess the borders influence.", "questions": "* How the model will perform on poles where the horizontal resolution is very different from the equator?\n* Do your CNN implementation consider the discontinuity in latitudes (360=0)?\n* The border influences much the model during regional forecasts. How do you handle this problem? (if wind velocity is 100 km/h, in 24 hours it will travel 2400 km)\n* How your model can be used in practice? Consider that ERA-5 dataset is not available operationally.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper is related to weather modeling on a 5.625° resolution using ERA-5 data as target. The solution is based on the NeuralODE model with several new features inside. A CNN-ViT-CNN sandwich architecture is proposed to model the right term of the ODE. The authors claim a strong improvement in simulation quality over the alternative approaches. Such an improvement is attributed to new scheme of velocity derivatives estimation based on the wave equation and to the CNN-based approximation of a source term in the advection equation", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* An interesting wave-equation inspired model to capture time derivative of velocity\n* Comparison with several baselines showing better results of a proposed model\n* Ability to interpolate in time and produce a continuous dynamics\n* Ablation studies for the velocity derivative approximation and model architecture", "weaknesses": "* The CNN used in variable space scale loses part of its physical meaning\n* The wave equation doesn’t describe the atmospheric dynamics really well so the derivative approximation is not governed by physics\n* The model has low practical applicability\nProbably an interesting experiment will be to use the model in the autoregressive mode and to assess its quality for longer forecast horizons. The same can be done for regional models to assess the borders influence.", "questions": "* How the model will perform on poles where the horizontal resolution is very different from the equator?\n* Do your CNN implementation consider the discontinuity in latitudes (360=0)?\n* The border influences much the model during regional forecasts. How do you handle this problem? (if wind velocity is 100 km/h, in 24 hours it will travel 2400 km)\n* How your model can be used in practice? Consider that ERA-5 dataset is not available operationally.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730719516605}, {"id": "SndtUwlTlx", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1411/Reviewer_tB5a"], "rating": 1, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces a sandwich method whereby the atmospheric dynamics is represented by a simplified advection differentials and the components are separately learned (e.g., source, advection, initial condition) with a parameterized deep models, depending on whether they possess fast or slow convergence.", "review_text": "This paper introduces a sandwich method whereby the atmospheric dynamics is represented by a simplified advection differentials and the components are separately learned (e.g., source, advection, initial condition) with a parameterized deep models, depending on whether they possess fast or slow convergence.", "strengths": "Building physics-informed network for chaotic systems, such as weather dynamics, is crucial and this paper offers an interesting ODE-based approach to tackle the problem.", "weaknesses": "Let me start off by saying that using an advection equation is only valid under __very simplified scenario__, and weather dynamics is definitely not the case at all. Elaborations to follow:\n\n1. The paper assumes incompressible fluid, where fluid density remains static along the pressure level. This is a gross simplification and is unphysical as the weather dynamics is completely dependent on the fluid being compressible, to allow interesting processes to take place such as convection through buoyancy (cloud formation, precipitation, energy-water cycling), large-scale fluid motion (teleconnection), turbulence (boundary-layer interaction), etc. This very assumption, therefore, does not allow for many atmospheric phenomena, including but not limited to extreme events (ENSO, hurricanes, etc), or just the general circulation of the atmosphere. This is not including the paper's simplification of ignoring the spherical nature of the Earth that enables for a different set of dynamics enabled by e.g., coriolis force.\n\n2. The paper also ignores many important conservation constraints found in the classical dynamical core, such as the conservation of mass, energy, and momentum. \n\n3. As such, there are better approximation to the full Navier-Stokes equation, such as Shallow Water Equation or Quasi-Geostrophic flow, that attempts to capture some of the realism of the atmosphere, and is therefore a much better differentials than the advection equation. \n\nRegardless, there are additional weaknesses that warrant a reject rating:\n\n4. The limited number of variables used (5), coarse spatial resolution (5.625-degree vs 0.25; ~400x smaller horizontal resolution), and small forecasting lead-time (72-hour) are too unconvincing to test this Neural ODE formulation for real weather application. I suspect that given the gross oversimplification through the advection equation, the framework could not accurately evolve the full atmospheric state with significant vertical motion/interaction, and is therefore not useful in a short forecasting window, let alone over a longer rollout in the medium-weather and longer sub-seasonal scale. \n\n5.  This is echoed by the result in Appendix E Table 9 is inferior to ClimaX at 3-day lead time, with error growth much larger than the other model (also lacking IFS baseline here to see how the rate of error propagation). The ACC for for u10, for instance, deteriorates from 0.93 to 0.80 even at 36-hour difference! The fundamental error in the assumptions (which ClimODE is too, and the authors should therefore use stronger baselines such as GraphCast) might contribute to this result. \n\nOverall, the basic assumptions underpinning this ODE is over-simplistic and does not represent realistic atmosphere (e.g., no vertical motion, no conservation constraints). If the goal is to model a simplified flow, such as tracer dynamics where their evolution is passive e.g., does not have any feedback loop with other quantities, and local, then advection might be sufficient. But the atmosphere is definitely not the case. Also, the inferior results over long lead-time, weak baseline, and the limited scope of the problem setup, need more significant work.", "questions": "Similar as the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a sandwich method whereby the atmospheric dynamics is represented by a simplified advection differentials and the components are separately learned (e.g., source, advection, initial condition) with a parameterized deep models, depending on whether they possess fast or slow convergence.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Building physics-informed network for chaotic systems, such as weather dynamics, is crucial and this paper offers an interesting ODE-based approach to tackle the problem.", "weaknesses": "Let me start off by saying that using an advection equation is only valid under __very simplified scenario__, and weather dynamics is definitely not the case at all. Elaborations to follow:\n\n1. The paper assumes incompressible fluid, where fluid density remains static along the pressure level. This is a gross simplification and is unphysical as the weather dynamics is completely dependent on the fluid being compressible, to allow interesting processes to take place such as convection through buoyancy (cloud formation, precipitation, energy-water cycling), large-scale fluid motion (teleconnection), turbulence (boundary-layer interaction), etc. This very assumption, therefore, does not allow for many atmospheric phenomena, including but not limited to extreme events (ENSO, hurricanes, etc), or just the general circulation of the atmosphere. This is not including the paper's simplification of ignoring the spherical nature of the Earth that enables for a different set of dynamics enabled by e.g., coriolis force.\n\n2. The paper also ignores many important conservation constraints found in the classical dynamical core, such as the conservation of mass, energy, and momentum. \n\n3. As such, there are better approximation to the full Navier-Stokes equation, such as Shallow Water Equation or Quasi-Geostrophic flow, that attempts to capture some of the realism of the atmosphere, and is therefore a much better differentials than the advection equation. \n\nRegardless, there are additional weaknesses that warrant a reject rating:\n\n4. The limited number of variables used (5), coarse spatial resolution (5.625-degree vs 0.25; ~400x smaller horizontal resolution), and small forecasting lead-time (72-hour) are too unconvincing to test this Neural ODE formulation for real weather application. I suspect that given the gross oversimplification through the advection equation, the framework could not accurately evolve the full atmospheric state with significant vertical motion/interaction, and is therefore not useful in a short forecasting window, let alone over a longer rollout in the medium-weather and longer sub-seasonal scale. \n\n5.  This is echoed by the result in Appendix E Table 9 is inferior to ClimaX at 3-day lead time, with error growth much larger than the other model (also lacking IFS baseline here to see how the rate of error propagation). The ACC for for u10, for instance, deteriorates from 0.93 to 0.80 even at 36-hour difference! The fundamental error in the assumptions (which ClimODE is too, and the authors should therefore use stronger baselines such as GraphCast) might contribute to this result. \n\nOverall, the basic assumptions underpinning this ODE is over-simplistic and does not represent realistic atmosphere (e.g., no vertical motion, no conservation constraints). If the goal is to model a simplified flow, such as tracer dynamics where their evolution is passive e.g., does not have any feedback loop with other quantities, and local, then advection might be sufficient. But the atmosphere is definitely not the case. Also, the inferior results over long lead-time, weak baseline, and the limited scope of the problem setup, need more significant work.", "questions": "Similar as the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730122381115}, {"id": "PD7Y6wRyaw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1411/Reviewer_g7RP"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces WeatherODE, a novel weather forecasting model based on deep learning. Building on the recent ClimODE model, WeatherODE follows a similar methodology and models the evolution of atmospheric quantities through a learnable advection equation. In this framework, the authors propose new architectures for learning the advection equation. A CNN-based neural network is used to predict the initial velocity of the equation, leveraging insights from the wave equation, rather than relying on time discretization. Additionally, a Vision Transformer-based neural ODE is trained to model the evolution of velocity over time. Finally, a CNN-based neural network is introduced to learn the source term of the advection equation, aiming to reduce the propagation of numerical errors. The paper provides detailed information on the training process of this new architecture, along with extensive experiments comparing WeatherODE to several state-of-the-art baselines across global and regional weather forecasting tasks with varying lead times. Ablation studies on the model's architecture are also included.", "review_text": "This paper introduces WeatherODE, a novel weather forecasting model based on deep learning. Building on the recent ClimODE model, WeatherODE follows a similar methodology and models the evolution of atmospheric quantities through a learnable advection equation. In this framework, the authors propose new architectures for learning the advection equation. A CNN-based neural network is used to predict the initial velocity of the equation, leveraging insights from the wave equation, rather than relying on time discretization. Additionally, a Vision Transformer-based neural ODE is trained to model the evolution of velocity over time. Finally, a CNN-based neural network is introduced to learn the source term of the advection equation, aiming to reduce the propagation of numerical errors. The paper provides detailed information on the training process of this new architecture, along with extensive experiments comparing WeatherODE to several state-of-the-art baselines across global and regional weather forecasting tasks with varying lead times. Ablation studies on the model's architecture are also included.", "strengths": "The paper is well organized with a clear structure and informative figures and tables. The experimental work is extensive and technically well-documented.\n\nIncorporating a strong physical bias into learning models is of great importance, especially for climate-related problems where the chaotic nature of underlying physical processes can cause the system to deviate from the training distribution. The inclusion of physical priors is essential for ensuring that the model generalizes well.\n\nThe proposed method outperforms the other models by a significant amount, in both global and regional forecasts.\n\nThe proposed method significantly outperforms the baselines, demonstrating improvements in both global and regional forecasting tasks.\n\nAblation studies, accompanied by illustrative explanations of the core contributions, are provided. The paper, in particular, demonstrates the effect of time discretization and how the proposed architecture mitigates this issue, offering comparisons to previous methods.\n\nThe discussion on the convergence speed of various learning models is first introduced qualitatively in Section 3.3 and then quantified in Section 5.3, which offers valuable insights into the architecture's design choices.\n\nAlthough briefly mentioned, the paper also proposes a flexible inference model capable of producing forecasts for different lead times. The ability to make weather predictions at varying lead times is a significant advantage and a promising idea.", "weaknesses": "Despite the solid experimental results, I believe that the paper's explanation of the methodology lacks some clarity, context, and references. While the proposed architecture achieves strong performance, the scientific reasoning behind these improvements is somewhat incomplete.\n\nFirst, the authors claim that all physical variables follow an advection equation, which is learned. This assumption, initially made in the ClimODE model, forms the foundation of the WeatherODE architecture and is said to provide a strong physical bias. It would be beneficial if the paper provided more physical context regarding the advection equation and why it serves as a good prior for learning atmospheric dynamics. From a physical perspective, what are the consequences of assuming that all variables follow this equation with a learned velocity field? Can the authors connect this to known atmospheric processes?\n\n\n#### Section 3.1\n\nThere is some lack of clarity in the discussion of discretization between lines 192 and 204, particularly around Equation (3). It is unclear at which stage the ODE is being discretized. The problem of learning an ODE's flow using neural ODEs should, in theory, be independent of the discretization, as neural ODEs are designed to differentiate through the ODE in a solver-agnostic way.  Second, indices $t_0$ and $t_n$ seem to play the same role.  It would be clearer if a single notation were used for the initial condition, even if this involves only one discretization step, with the methodology then being extended to subsequent steps.\n\n####  Section 3.2\n\nThis section focuses on predicting the initial velocity $v(t_0)$. The authors provide physical insight on the computation of the state derivative $\\partial u/ \\partial t$. However, the link between the two quantities $v(t_0)$ and $\\partial u / \\partial t$ is not explicitly stated. Yet it is one of the paper's main objectives to improve the computation of the $v$ as a function of the state $u$, without resorting to time discretization. Indeed, it is stated that $v$ is estimated from $u$ in the ClimODE paper using time discretization, but the estimation is never mentioned. As a result, the section presents a new method based on the wave equation to estimate state derivatives, but the connection with the ultimate problem of estimating the velocity is missing. This omission impairs the clarity of the paper, as this step is one of the core contribution.\nI understand that the estimation method is framed as a variational inverse problem in the ClimODE paper, and that WeatherODE improves it by casting it as a learning problem instead. I believe that it should be stated in this paper as well for Section 3.2 to be relevant. \n\nAdditionally, the link between the wave equation and the advection equation is not clearly explained. The authors mention that the wave equation is commonly used in atmospheric dynamics, but they do not provide any references or detailed physical context. While the wave equation is indeed important in describing physical processes related to propagation, further clarification is needed regarding its purpose in this model.\n\nIf I understood correctly, the integrated wave equation (5) seems to show that the first-order derivative of the state $\\partial u / \\partial t (t_0)$ is linked to the state gradients $\\nabla u(t_0)$, which motivates predicting the velocity field should as a function of the state gradients rather than the state itself. However, assuming that the wave equation holds, the involved spatial derivatives are of second order rather than first order, and are integrated over a past time interval rather than evaluated only at the current time $t_0$. In my opinion, and if I understood correctly, equation (5) may not be entirely appropriate for predicting $v(t_0)$ as described.\n\nThis concern is compounded by the results of Figure 3, where the performance gap between the proposed method using the wave equation and other approximators is quite small. Couldn't it be that the performance gap comes from learning the initial velocity from a neural network conditioned on non-discretized state values instead of solving an inverse problem with discretized state derivatives, rather than from the wave equation-informed predictive structure of the neural network? \n\nFinally, the qualitative argument about spatial resolution at the end of this section is not particularly convincing, as it compares spatial resolution to time resolution in a way that seems dimensionally inconsistent. The statement on line 243,\n``\nthe spatial domain is nearly 100 times denser than the temporal domain.\n ``\n\nneeds further clarification, as its implications are unclear.", "questions": "Neural ODEs are known to be computationally demanding, yet the paper does not address computational resources or the back-propagation method used to train the neural ODE. Since computational power can often be a limiting factor, it would be valuable to compare the computational times across the different architectures, particularly for the neural ODE versus other feedforward models.\n\nThe WeatherODE* model appears to be an interesting research direction, offering flexibility in generating forecasts at different lead times. However, there is limited explanation as to why WeatherODE is capable of such flexibility. Could the authors clarify the connection between the following sentence:\n\n``\nby modeling the atmosphere as a physics-driven continuous process anddesigning a time-dependent source network to account for errors at each time step, WeatherODE can capture information across all intermediate time points\n``\n\n and the success of the 24-hour model WeatherODE*?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces WeatherODE, a novel weather forecasting model based on deep learning. Building on the recent ClimODE model, WeatherODE follows a similar methodology and models the evolution of atmospheric quantities through a learnable advection equation. In this framework, the authors propose new architectures for learning the advection equation. A CNN-based neural network is used to predict the initial velocity of the equation, leveraging insights from the wave equation, rather than relying on time discretization. Additionally, a Vision Transformer-based neural ODE is trained to model the evolution of velocity over time. Finally, a CNN-based neural network is introduced to learn the source term of the advection equation, aiming to reduce the propagation of numerical errors. The paper provides detailed information on the training process of this new architecture, along with extensive experiments comparing WeatherODE to several state-of-the-art baselines across global and regional weather forecasting tasks with varying lead times. Ablation studies on the model's architecture are also included.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper is well organized with a clear structure and informative figures and tables. The experimental work is extensive and technically well-documented.\n\nIncorporating a strong physical bias into learning models is of great importance, especially for climate-related problems where the chaotic nature of underlying physical processes can cause the system to deviate from the training distribution. The inclusion of physical priors is essential for ensuring that the model generalizes well.\n\nThe proposed method outperforms the other models by a significant amount, in both global and regional forecasts.\n\nThe proposed method significantly outperforms the baselines, demonstrating improvements in both global and regional forecasting tasks.\n\nAblation studies, accompanied by illustrative explanations of the core contributions, are provided. The paper, in particular, demonstrates the effect of time discretization and how the proposed architecture mitigates this issue, offering comparisons to previous methods.\n\nThe discussion on the convergence speed of various learning models is first introduced qualitatively in Section 3.3 and then quantified in Section 5.3, which offers valuable insights into the architecture's design choices.\n\nAlthough briefly mentioned, the paper also proposes a flexible inference model capable of producing forecasts for different lead times. The ability to make weather predictions at varying lead times is a significant advantage and a promising idea.", "weaknesses": "Despite the solid experimental results, I believe that the paper's explanation of the methodology lacks some clarity, context, and references. While the proposed architecture achieves strong performance, the scientific reasoning behind these improvements is somewhat incomplete.\n\nFirst, the authors claim that all physical variables follow an advection equation, which is learned. This assumption, initially made in the ClimODE model, forms the foundation of the WeatherODE architecture and is said to provide a strong physical bias. It would be beneficial if the paper provided more physical context regarding the advection equation and why it serves as a good prior for learning atmospheric dynamics. From a physical perspective, what are the consequences of assuming that all variables follow this equation with a learned velocity field? Can the authors connect this to known atmospheric processes?\n\n\n#### Section 3.1\n\nThere is some lack of clarity in the discussion of discretization between lines 192 and 204, particularly around Equation (3). It is unclear at which stage the ODE is being discretized. The problem of learning an ODE's flow using neural ODEs should, in theory, be independent of the discretization, as neural ODEs are designed to differentiate through the ODE in a solver-agnostic way.  Second, indices $t_0$ and $t_n$ seem to play the same role.  It would be clearer if a single notation were used for the initial condition, even if this involves only one discretization step, with the methodology then being extended to subsequent steps.\n\n####  Section 3.2\n\nThis section focuses on predicting the initial velocity $v(t_0)$. The authors provide physical insight on the computation of the state derivative $\\partial u/ \\partial t$. However, the link between the two quantities $v(t_0)$ and $\\partial u / \\partial t$ is not explicitly stated. Yet it is one of the paper's main objectives to improve the computation of the $v$ as a function of the state $u$, without resorting to time discretization. Indeed, it is stated that $v$ is estimated from $u$ in the ClimODE paper using time discretization, but the estimation is never mentioned. As a result, the section presents a new method based on the wave equation to estimate state derivatives, but the connection with the ultimate problem of estimating the velocity is missing. This omission impairs the clarity of the paper, as this step is one of the core contribution.\nI understand that the estimation method is framed as a variational inverse problem in the ClimODE paper, and that WeatherODE improves it by casting it as a learning problem instead. I believe that it should be stated in this paper as well for Section 3.2 to be relevant. \n\nAdditionally, the link between the wave equation and the advection equation is not clearly explained. The authors mention that the wave equation is commonly used in atmospheric dynamics, but they do not provide any references or detailed physical context. While the wave equation is indeed important in describing physical processes related to propagation, further clarification is needed regarding its purpose in this model.\n\nIf I understood correctly, the integrated wave equation (5) seems to show that the first-order derivative of the state $\\partial u / \\partial t (t_0)$ is linked to the state gradients $\\nabla u(t_0)$, which motivates predicting the velocity field should as a function of the state gradients rather than the state itself. However, assuming that the wave equation holds, the involved spatial derivatives are of second order rather than first order, and are integrated over a past time interval rather than evaluated only at the current time $t_0$. In my opinion, and if I understood correctly, equation (5) may not be entirely appropriate for predicting $v(t_0)$ as described.\n\nThis concern is compounded by the results of Figure 3, where the performance gap between the proposed method using the wave equation and other approximators is quite small. Couldn't it be that the performance gap comes from learning the initial velocity from a neural network conditioned on non-discretized state values instead of solving an inverse problem with discretized state derivatives, rather than from the wave equation-informed predictive structure of the neural network? \n\nFinally, the qualitative argument about spatial resolution at the end of this section is not particularly convincing, as it compares spatial resolution to time resolution in a way that seems dimensionally inconsistent. The statement on line 243,\n``\nthe spatial domain is nearly 100 times denser than the temporal domain.\n ``\n\nneeds further clarification, as its implications are unclear.", "questions": "Neural ODEs are known to be computationally demanding, yet the paper does not address computational resources or the back-propagation method used to train the neural ODE. Since computational power can often be a limiting factor, it would be valuable to compare the computational times across the different architectures, particularly for the neural ODE versus other feedforward models.\n\nThe WeatherODE* model appears to be an interesting research direction, offering flexibility in generating forecasts at different lead times. However, there is limited explanation as to why WeatherODE is capable of such flexibility. Could the authors clarify the connection between the following sentence:\n\n``\nby modeling the atmosphere as a physics-driven continuous process anddesigning a time-dependent source network to account for errors at each time step, WeatherODE can capture information across all intermediate time points\n``\n\n and the success of the 24-hour model WeatherODE*?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729862064122}, {"id": "UYeVcmuYwH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1411/Reviewer_dcuz"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The study proposes a neural ODE approach to data-driven NWP by using three different neural nets with three different rates of convergence to solve for three different terms in the system of ODEs obtained by applying the method of lines to the advection continuity equation. The three models correspond to (a) solving for the initial velocity estimate using the wave equation, (b) an advection model to compute the tendency of the velocity, and (c) to estimate the source term in the momentum equation. Respectively, a 2D CNN, a vision transformer and a 3D CNN is used to estimate the three terms which are then used to march the variables in time.\n\nThe model is compared with multiple other AI NWP models and this approach is found to substantialy reduce prediction errors in both global and regional contexts mst likely by reducing the errors in the estimation of the time derivative of velocities which might be quite erroneous due to a large (hourly) time step.\n\nThe ultimate approach is to the create a \"sandwich\"  physics driven ODE which uses the three architectures together to predict the future state of the atmosphere.", "review_text": "The study proposes a neural ODE approach to data-driven NWP by using three different neural nets with three different rates of convergence to solve for three different terms in the system of ODEs obtained by applying the method of lines to the advection continuity equation. The three models correspond to (a) solving for the initial velocity estimate using the wave equation, (b) an advection model to compute the tendency of the velocity, and (c) to estimate the source term in the momentum equation. Respectively, a 2D CNN, a vision transformer and a 3D CNN is used to estimate the three terms which are then used to march the variables in time.\n\nThe model is compared with multiple other AI NWP models and this approach is found to substantialy reduce prediction errors in both global and regional contexts mst likely by reducing the errors in the estimation of the time derivative of velocities which might be quite erroneous due to a large (hourly) time step.\n\nThe ultimate approach is to the create a \"sandwich\"  physics driven ODE which uses the three architectures together to predict the future state of the atmosphere.", "strengths": "The experimental setup is robust and it is nice to see the significant improvements in performance across all variables for the neural ODE sandwich architecture. The paper is well written and the hybrid approach in the paper could likely be valuable to the AI NWP community.\n\nI like the paper for its skillful approach of blending classical PDE theory with AI to improve errors in atmospheric state forecasting.", "weaknesses": "1) I don't think ClimaX is a good baseline for such comparisons, as it was itself trained on a wide range of CMIP6 models which are (a) not really tuned for weather forecasting and (b) due to model design have multiple biases. Finetuning can only fix so many of them. Thus, I would not put too much weight into how the WeatherODE performs against ClimaX as it is a weak baseline to begin with. Similarly, I would recommend comparing the method with the newer version of FourCastNet based on the SFNO architecture which is more stable for short and long term rollouts and is more topology-aware.\n\n2) A minor point but the paper tends to over-cite at places: the Evans (2022) reference, the Vaswani et al reference and the Biswas et al. (2013) reference are totally unnecessary. Vaswani et al. has not even been cited in the correct context as it has not realtion to weather forecasting. Similarly on Line 37-38.\n\n3) The figures' text could be increased. It is tough to interpret Figures 1 and 4 in print.\n\n4) It would have been interesting to see the time complexity of the sandwich model proposed in the study and how it compares with other models. Current AI weather forecasting models have been a bit over-glorified in a sense that they ignore baroclinic motions in the atmosphere and use data on very few vertical levels, and then claim to be as good as traditional NWP models - which are way more versatile thab AI NWPs in the problems they can be used to solve. Understandably, the only advantage then is the computational speed and the reduction in operational cost that they offer. Therefore, as more traditional numerical are embedded into pure data-driven architectures, it would be interesting to see the effect on more complicated hybrid architectures on the run speed.\n\n5) If the refined horizontal resolution is the key here (as it allows computing spatial gradients accurately), they can simply using higher spatial resolution training data (like 25 km x 25 km ERA5 data) lead to similar improvements in performance?\n\n6) One worry is that the achitecture is becoming too combersome to be practical: a unified model structure for past AI NWP models is one of the key cornerstones of its appeal. Traditional NWP model provide forecasts at a 9 km resolution. If the same models were to be run at 5.625 deg resolution, one can expect similar order run times between AI weather emulators and NWP models, ultimately leading one to question the central point of these models. So, having multiple ML models in sequence, such as those proposed in this study, can increase the complexity of the problem to the point that one begins to question the novelty of this approach. What would have been great could be to train one single neural net on atmospheric data sampled at a high frquency and use that for more accurate initializations. \n\n7) If i undestand correctly, the study employs a two dimensional equation. Have the authors considered using a three dimensional equation instead which considers the vertical verlocity into account as well? This could be important especially for tropical predictability ad thermodynamical processes like dry and moist convection evolve over sub-hourly timescales and can introduce notable errors into the equation. This could also affect longer lead time rollouts of the model. Or, if I undertand it correctly, the vertical velocity is simply treated within the source term (which might not be the best approach).\n\n8) Source term: If I understand correctly, and I could be wrong, the source model tend to learn all the other forcing terms of the advection equation. In the context of ERA5, this would not just contain other forcings, but also data assimilation errors. Since the DA errors are not bery systematics in nature, how can errors in predicting the terms influence the u_t+1 prediction obtained from the neural ODEs?", "questions": "Questions are combined with weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The study proposes a neural ODE approach to data-driven NWP by using three different neural nets with three different rates of convergence to solve for three different terms in the system of ODEs obtained by applying the method of lines to the advection continuity equation. The three models correspond to (a) solving for the initial velocity estimate using the wave equation, (b) an advection model to compute the tendency of the velocity, and (c) to estimate the source term in the momentum equation. Respectively, a 2D CNN, a vision transformer and a 3D CNN is used to estimate the three terms which are then used to march the variables in time.\n\nThe model is compared with multiple other AI NWP models and this approach is found to substantialy reduce prediction errors in both global and regional contexts mst likely by reducing the errors in the estimation of the time derivative of velocities which might be quite erroneous due to a large (hourly) time step.\n\nThe ultimate approach is to the create a \"sandwich\"  physics driven ODE which uses the three architectures together to predict the future state of the atmosphere.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The experimental setup is robust and it is nice to see the significant improvements in performance across all variables for the neural ODE sandwich architecture. The paper is well written and the hybrid approach in the paper could likely be valuable to the AI NWP community.\n\nI like the paper for its skillful approach of blending classical PDE theory with AI to improve errors in atmospheric state forecasting.", "weaknesses": "1) I don't think ClimaX is a good baseline for such comparisons, as it was itself trained on a wide range of CMIP6 models which are (a) not really tuned for weather forecasting and (b) due to model design have multiple biases. Finetuning can only fix so many of them. Thus, I would not put too much weight into how the WeatherODE performs against ClimaX as it is a weak baseline to begin with. Similarly, I would recommend comparing the method with the newer version of FourCastNet based on the SFNO architecture which is more stable for short and long term rollouts and is more topology-aware.\n\n2) A minor point but the paper tends to over-cite at places: the Evans (2022) reference, the Vaswani et al reference and the Biswas et al. (2013) reference are totally unnecessary. Vaswani et al. has not even been cited in the correct context as it has not realtion to weather forecasting. Similarly on Line 37-38.\n\n3) The figures' text could be increased. It is tough to interpret Figures 1 and 4 in print.\n\n4) It would have been interesting to see the time complexity of the sandwich model proposed in the study and how it compares with other models. Current AI weather forecasting models have been a bit over-glorified in a sense that they ignore baroclinic motions in the atmosphere and use data on very few vertical levels, and then claim to be as good as traditional NWP models - which are way more versatile thab AI NWPs in the problems they can be used to solve. Understandably, the only advantage then is the computational speed and the reduction in operational cost that they offer. Therefore, as more traditional numerical are embedded into pure data-driven architectures, it would be interesting to see the effect on more complicated hybrid architectures on the run speed.\n\n5) If the refined horizontal resolution is the key here (as it allows computing spatial gradients accurately), they can simply using higher spatial resolution training data (like 25 km x 25 km ERA5 data) lead to similar improvements in performance?\n\n6) One worry is that the achitecture is becoming too combersome to be practical: a unified model structure for past AI NWP models is one of the key cornerstones of its appeal. Traditional NWP model provide forecasts at a 9 km resolution. If the same models were to be run at 5.625 deg resolution, one can expect similar order run times between AI weather emulators and NWP models, ultimately leading one to question the central point of these models. So, having multiple ML models in sequence, such as those proposed in this study, can increase the complexity of the problem to the point that one begins to question the novelty of this approach. What would have been great could be to train one single neural net on atmospheric data sampled at a high frquency and use that for more accurate initializations. \n\n7) If i undestand correctly, the study employs a two dimensional equation. Have the authors considered using a three dimensional equation instead which considers the vertical verlocity into account as well? This could be important especially for tropical predictability ad thermodynamical processes like dry and moist convection evolve over sub-hourly timescales and can introduce notable errors into the equation. This could also affect longer lead time rollouts of the model. Or, if I undertand it correctly, the vertical velocity is simply treated within the source term (which might not be the best approach).\n\n8) Source term: If I understand correctly, and I could be wrong, the source model tend to learn all the other forcing terms of the advection equation. In the context of ERA5, this would not just contain other forcings, but also data assimilation errors. Since the DA errors are not bery systematics in nature, how can errors in predicting the terms influence the u_t+1 prediction obtained from the neural ODEs?", "questions": "Questions are combined with weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729402340376}, {"id": "fWUZEwG244", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1411/Reviewer_o4jc"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper extends a recent neural weather ODE model (ClimODE 2024) by using wave-equations for initial state estimation, adopting a more transformer based dynamics, and adding a CNN source field. The contributions range from incremental (transformers) to substantial (source field). The results are outstanding.\n\n------\n\nPost-response update. I'm decreasing the score to reject, which I understand is exceptional given my initial positive review. The authors have not been able to clarify or address my concerns regarding the role of the source term, and the role of wave term in the advective system. I believe using the wave term is unfounded in an advective system, while the source term seems not to be part of the PDE afterall (given that source gets as input future states), which violates eq 1. The paper is not ready for publication.", "review_text": "The paper extends a recent neural weather ODE model (ClimODE 2024) by using wave-equations for initial state estimation, adopting a more transformer based dynamics, and adding a CNN source field. The contributions range from incremental (transformers) to substantial (source field). The results are outstanding.\n\n------\n\nPost-response update. I'm decreasing the score to reject, which I understand is exceptional given my initial positive review. The authors have not been able to clarify or address my concerns regarding the role of the source term, and the role of wave term in the advective system. I believe using the wave term is unfounded in an advective system, while the source term seems not to be part of the PDE afterall (given that source gets as input future states), which violates eq 1. The paper is not ready for publication.", "strengths": "- The paper extends weather ODEs in sensible ways, and the overall model is sensible. Adding the source field is a major contribution, while the initial estimation and network searches seem useful (but more incremental).\n- The results are outstanding: they even beat IFS at times, which is an outstanding achievement!", "weaknesses": "- The motivations behind the model choices seem a bit weak\n- The ablations are very interesting, but have some issues. In the v0 ablation there is no ClimODE baseline, so it’s difficult to say if any improvement was actually gained here. The Fig4 is superficial and difficult to interpret. The stability analysis seems interesting, but I don’t think it provides much insights into why things sometimes fail. One would not really expect such a simple ODE system to fail in the first place.\n- The experiments make an unfair comparison to the ClimODE baseline, where ClimODE uses 5 data variables and weatherODE 48. Results lack standard deviations. Some results show that weatherODE beats IFS, and this is not elaborated further. This is a major achievement, and is now provided a bit too casually. There should be some discussion on how the results relate to the larger model (panguweather, gencast, graphcast, etc).\n- The clarity and text needs improvements.", "questions": "- I have hard time understanding the v0 estimation. First, a wave equation is introduced in general terms, and then a CNN appears. These two are not connected together, so I’m left wondering what do we do with the waves, and how do they relate to the CNN. Furthermore, the wave equation is poorly motivated, and it seems disconnected from the advection equation 3. To me this looks like a discrepancy: the system follows advection, but initial state assumes different kind of physics (ie. laplacians appear out of nowhere). Surely the initial state estimation needs to respect the chosen ODE model, and not introduce some physics effects that are not part of the ODE.\n- The arguments about spatial vs temporal resolution are not convincing. Low temporal sampling rate does not necessarily mean that the temporal resolution is low: this depends on how quickly the process varies. It also feels misguided to say that spatial resolution is 100x higher than temporal resolution: this is only true if you take 32*64, and I don’t think you should do this. Since the (t,x,y) axes have similar ranges (24,32,64), I would argue that there is no resolution gap in the sampling rate between space and time.\n- I couldn’t follow the CNN/ViT convergence arguments. I’m not sure what convergence even means here (of training..?, of rollouts..?). I think the paper is arguing that training CNNs is somehow instable, and thus ViT’s have to be used. This sounds implausible, and a poor motivation for choosing how the dynamics should evolve. Surely the networks need to be chosen such that they respect some physical system properties (eg. feature locality/globality).\n- I don’t understand the source model. It takes as input all 1…N simulated states u(t_n), but one needs the source states s(t_n) to do this simulation. This has to be a mistake: I assume that we instead take as inputs u(t_1 : t_n), so that the history grows along the ODE rollouts. How do you handle set inputs and set outputs? I don’t think a 3D CNN supports sets as inputs or outputs [maybe these are not sets, but just tensors of fixed size..].\n- At sec 3.5. I’m again confused what do you mean by “convergence”.\n- Sec 3.5. claims that advection dynamics involves long-range dependencies. I’m not sure I agree, and would even argue the opposite. PDE’s are by nature local models: why would the weather in new york affect the infinitesimal change of weather in london? There is no connection between them. Can you elaborate?\n- Sec 3.6. claims that earlier methods only train against final state, and ignore intermediate terms. I’m surprised by this statement. ClimODE trains with all intermediate points, as does the original neural ODE, and every other neural ODE/PDE model I’ve seen. Can you provide examples of models where this happens? Calling this “multi-task” learning is also wrong: intermediate points in an ODE are not different “tasks” (they are not even a single repeated task, since it’s not a “task” in the first place).\n- The benchmarks use 48 weather variables, but only evaluate 5 of them. The benchmarks also take ClimODE results as-is from the paper at least in Table 2. The results are then unfair to ClimODE: the ClimODE is using only 5 variables worth of information, while weatherODE is using 48 variables. The paper needs compare apple to apples by either running weatherODE results using only 5 variables, or running ClimODE with 48 variables.\n\nI'm looking forward to the responses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper extends a recent neural weather ODE model (ClimODE 2024) by using wave-equations for initial state estimation, adopting a more transformer based dynamics, and adding a CNN source field. The contributions range from incremental (transformers) to substantial (source field). The results are outstanding.\n\n------\n\nPost-response update. I'm decreasing the score to reject, which I understand is exceptional given my initial positive review. The authors have not been able to clarify or address my concerns regarding the role of the source term, and the role of wave term in the advective system. I believe using the wave term is unfounded in an advective system, while the source term seems not to be part of the PDE afterall (given that source gets as input future states), which violates eq 1. The paper is not ready for publication.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper extends weather ODEs in sensible ways, and the overall model is sensible. Adding the source field is a major contribution, while the initial estimation and network searches seem useful (but more incremental).\n- The results are outstanding: they even beat IFS at times, which is an outstanding achievement!", "weaknesses": "- The motivations behind the model choices seem a bit weak\n- The ablations are very interesting, but have some issues. In the v0 ablation there is no ClimODE baseline, so it’s difficult to say if any improvement was actually gained here. The Fig4 is superficial and difficult to interpret. The stability analysis seems interesting, but I don’t think it provides much insights into why things sometimes fail. One would not really expect such a simple ODE system to fail in the first place.\n- The experiments make an unfair comparison to the ClimODE baseline, where ClimODE uses 5 data variables and weatherODE 48. Results lack standard deviations. Some results show that weatherODE beats IFS, and this is not elaborated further. This is a major achievement, and is now provided a bit too casually. There should be some discussion on how the results relate to the larger model (panguweather, gencast, graphcast, etc).\n- The clarity and text needs improvements.", "questions": "- I have hard time understanding the v0 estimation. First, a wave equation is introduced in general terms, and then a CNN appears. These two are not connected together, so I’m left wondering what do we do with the waves, and how do they relate to the CNN. Furthermore, the wave equation is poorly motivated, and it seems disconnected from the advection equation 3. To me this looks like a discrepancy: the system follows advection, but initial state assumes different kind of physics (ie. laplacians appear out of nowhere). Surely the initial state estimation needs to respect the chosen ODE model, and not introduce some physics effects that are not part of the ODE.\n- The arguments about spatial vs temporal resolution are not convincing. Low temporal sampling rate does not necessarily mean that the temporal resolution is low: this depends on how quickly the process varies. It also feels misguided to say that spatial resolution is 100x higher than temporal resolution: this is only true if you take 32*64, and I don’t think you should do this. Since the (t,x,y) axes have similar ranges (24,32,64), I would argue that there is no resolution gap in the sampling rate between space and time.\n- I couldn’t follow the CNN/ViT convergence arguments. I’m not sure what convergence even means here (of training..?, of rollouts..?). I think the paper is arguing that training CNNs is somehow instable, and thus ViT’s have to be used. This sounds implausible, and a poor motivation for choosing how the dynamics should evolve. Surely the networks need to be chosen such that they respect some physical system properties (eg. feature locality/globality).\n- I don’t understand the source model. It takes as input all 1…N simulated states u(t_n), but one needs the source states s(t_n) to do this simulation. This has to be a mistake: I assume that we instead take as inputs u(t_1 : t_n), so that the history grows along the ODE rollouts. How do you handle set inputs and set outputs? I don’t think a 3D CNN supports sets as inputs or outputs [maybe these are not sets, but just tensors of fixed size..].\n- At sec 3.5. I’m again confused what do you mean by “convergence”.\n- Sec 3.5. claims that advection dynamics involves long-range dependencies. I’m not sure I agree, and would even argue the opposite. PDE’s are by nature local models: why would the weather in new york affect the infinitesimal change of weather in london? There is no connection between them. Can you elaborate?\n- Sec 3.6. claims that earlier methods only train against final state, and ignore intermediate terms. I’m surprised by this statement. ClimODE trains with all intermediate points, as does the original neural ODE, and every other neural ODE/PDE model I’ve seen. Can you provide examples of models where this happens? Calling this “multi-task” learning is also wrong: intermediate points in an ODE are not different “tasks” (they are not even a single repeated task, since it’s not a “task” in the first place).\n- The benchmarks use 48 weather variables, but only evaluate 5 of them. The benchmarks also take ClimODE results as-is from the paper at least in Table 2. The results are then unfair to ClimODE: the ClimODE is using only 5 variables worth of information, while weatherODE is using 48 variables. The paper needs compare apple to apples by either running weatherODE results using only 5 variables, or running ClimODE with 48 variables.\n\nI'm looking forward to the responses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729160895686}], "openreview_url": "https://openreview.net/forum?id=UFzE9njwMG", "arxiv_id": "2410.06560", "paper_pdf": "papers/UFzE9njwMG.pdf", "paper_pdf_sha256": "b6718d50007b5853993e21b0514a35227952333421ced7867644453b720b5636", "paper_pdf_bytes": 15165138, "paper_pdf_source": "openreview", "code_url": "https://github.com/DAMO-DI-ML/WeatherODE", "code_repository": "DAMO-DI-ML/WeatherODE", "code_commit": "056b961ea0bc3dbcae47b88c79cdbb257a8c8d47", "code_archive": "repos/UFzE9njwMG.zip", "code_archive_sha256": "4e025cbc7dc26fa56d55a1744075a50966487c9575df370d155026fb5dca2fe7", "code_archive_bytes": 83053, "code_file_count": 44, "code_extensions": {".py": 28, ".sh": 16}, "github_disk_usage_kb": 61, "github_languages": {"Python": 200463, "Shell": 17503}, "github_archived": false, "github_pushed_at": "2024-10-23T05:58:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mitigating-time-discretization-challenges"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sSWGqY2qNJ", "year": 2024, "status": "rejected", "title": "Indeterminate Probability Theory", "authors": ["Tao Yang", "Chuang Liu", "Xiaofeng Ma", "Weijia Lu", "Ning Wu", "Bingyang Li", "ZHIFEI YANG", "Peng Liu", "Lin Sun", "xiaodong Zhang", "Can Zhang"], "authorids": ["~Tao_Yang19", "~Chuang_Liu5", "~Xiaofeng_Ma1", "~Weijia_Lu1", "~Ning_Wu3", "~Bingyang_Li1", "~ZHIFEI_YANG2", "~Peng_Liu17", "~Lin_Sun11", "~xiaodong_Zhang6", "~Can_Zhang4"], "authors_source": "OpenReview API", "abstract": "Currently, there is no mathematical analytical form for a general posterior. We have discovered a new theory to address this issue, which is called Indeterminate Probability Theory. This is a big discovery in the field of probability, and it is an extension of classical probability theory, and makes classical probability theory a special case to our theory. In this paper, we propose a new perspective for understanding probability theory by introducing Observers and treating the outcome of each random experiment as an indeterminate probability distribution, which leads to probability calculations being a combination of ground truth and observation errors. We then discover three conditional mutual independent assumptions as Candidate Axioms and divide the probability process into two phase: observation phase and inference phase. In the observation phase, a general equation for any complex posterior is derived. In the inference phase, the inference probability equation with the posterior is derived. Base on this theory, we propose a new general model called IPNN - Indeterminate Probability Neural Network to validate our theory. Furthermore, in one of our another papers, this new theory is successfully applied to the task of multivariate time series (MTS) forecasting without relying on any neural models, and it outperforms LSTM models as well as some transformer-based models. In addition, further applications of this new theory are also discussed in this paper. Validations of this theory are reflected in experimental results.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "4DdfJbqlyA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4295/Reviewer_MWjN"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This is an extremely ambitious paper that attempts to construct a new theory called indeterminate probability theory. The key idea of Indeterminate probability theory is to introduce a new concept of auxiliary observers and to treat the results of each random experiment as an indeterminate probability distribution, while still preserving the assumption of mutual independence. As a result, the posterior probabilities of the system can be derived in a form that is easy to handle analytically, an important benefit in applications.\nThe authors demonstrate the applicability of this idea to regression and classification problems by combining it with neural networks.", "review_text": "This is an extremely ambitious paper that attempts to construct a new theory called indeterminate probability theory. The key idea of Indeterminate probability theory is to introduce a new concept of auxiliary observers and to treat the results of each random experiment as an indeterminate probability distribution, while still preserving the assumption of mutual independence. As a result, the posterior probabilities of the system can be derived in a form that is easy to handle analytically, an important benefit in applications.\nThe authors demonstrate the applicability of this idea to regression and classification problems by combining it with neural networks.", "strengths": "I am very grateful to the authors for sharing their novel attempt at this paper. I enjoyed reading this paper very much.\n- The paper devotes a great deal of effort in its presentation to illustrate new ideas that are outside of the conventional wisdom. The paper is very well written and its organization is designed to appeal to a diverse audience. In particular, it is designed to be easily understood by explaining the core ideas by means of toy examples.\n- The practical contribution of this paper is very significant. Traditionally, posterior probabilities in statistical machine learning have been approximated by some kind of approximation method (e.g., Markov chain Monte Carlo or variational methods), but the ideas in this paper have the potential to be a new option to add to that.", "weaknesses": "First of all, let me emphasize that I am trying to be very open minded in understanding the value of this paper. My grade on my first peer review may not be very high, but I am prepared to improve it as soon as I properly understand the value of this paper.\nMy concern is whether this paper could create a new system of probability theory (i.e., a major historical breakthrough) or whether it provides a new perspective on approximation and interpretation for the system in a form that is easy to handle in applications (i.e., a new alternative alongside MCMC and VB), a somewhat excessive Is it an appealing proposition? I would like to inquire in the question section for more details.", "questions": "My question can be summarized very simply as to whether or not indeterminate probability theory can be expressed in terms of a definition of probability space using abstract probability space.\n\nFirst of all, I understand this new insightful strategy of the authors as follows (Perhaps this understanding of mine is incorrect. If I am wrong, I would be very grateful if you could correct me.) \n- The authors' system introduces uncertainty as an auxiliary variable for observers. If this were to be expressed in the context of a conventional standard Bayesian analysis, the observer could be represented as making an observation error according to the auxiliary random variable.\n- Next, since this auxiliary random variable is not needed to describe the system, we will try to eliminate it in some way. In a conventional standard Bayesian analysis, this can be done by eliminating the auxiliary random variable by marginalization. However, a problem arises here. If the auxiliary random variable is shared by all observers, the system loses observer independence (Axiom 2 of the proposed probability theory) when it is eliminated.\n- Therefore, the proposed probability theory simply ignores the auxiliary random variable while simultaneously assuming Axiom 2.\n\nIf we were to use such a strategy, it would certainly seem that we could view the system as different from classical probability theory (as mentioned in the paper, we could of course make special cases that are equivalent to classical probability theory in special circumstances).\n\nFollowing this intuition, my interest is in what the authors' system would look like if it were represented in an abstract probability space. That is, a situation where all randomness in the world is governed by an abstract space $\\Theta$, where all randomness is lost if the abstract space is determined at a point $\\theta\\in\\Theta$, and where all variables can be described deterministically. In the abstract space, random variables are represented as a projection of the world as a map to an object, e.g., $Y(\\theta), X(\\theta), A(\\theta)$ can be uniquely determined for a given source $\\theta$. Can the authors' system be represented using such a conventional abstract probability space? Or is it a deviation from that rule?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This is an extremely ambitious paper that attempts to construct a new theory called indeterminate probability theory. The key idea of Indeterminate probability theory is to introduce a new concept of auxiliary observers and to treat the results of each random experiment as an indeterminate probability distribution, while still preserving the assumption of mutual independence. As a result, the posterior probabilities of the system can be derived in a form that is easy to handle analytically, an important benefit in applications.\nThe authors demonstrate the applicability of this idea to regression and classification problems by combining it with neural networks.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "I am very grateful to the authors for sharing their novel attempt at this paper. I enjoyed reading this paper very much.\n- The paper devotes a great deal of effort in its presentation to illustrate new ideas that are outside of the conventional wisdom. The paper is very well written and its organization is designed to appeal to a diverse audience. In particular, it is designed to be easily understood by explaining the core ideas by means of toy examples.\n- The practical contribution of this paper is very significant. Traditionally, posterior probabilities in statistical machine learning have been approximated by some kind of approximation method (e.g., Markov chain Monte Carlo or variational methods), but the ideas in this paper have the potential to be a new option to add to that.", "weaknesses": "First of all, let me emphasize that I am trying to be very open minded in understanding the value of this paper. My grade on my first peer review may not be very high, but I am prepared to improve it as soon as I properly understand the value of this paper.\nMy concern is whether this paper could create a new system of probability theory (i.e., a major historical breakthrough) or whether it provides a new perspective on approximation and interpretation for the system in a form that is easy to handle in applications (i.e., a new alternative alongside MCMC and VB), a somewhat excessive Is it an appealing proposition? I would like to inquire in the question section for more details.", "questions": "My question can be summarized very simply as to whether or not indeterminate probability theory can be expressed in terms of a definition of probability space using abstract probability space.\n\nFirst of all, I understand this new insightful strategy of the authors as follows (Perhaps this understanding of mine is incorrect. If I am wrong, I would be very grateful if you could correct me.) \n- The authors' system introduces uncertainty as an auxiliary variable for observers. If this were to be expressed in the context of a conventional standard Bayesian analysis, the observer could be represented as making an observation error according to the auxiliary random variable.\n- Next, since this auxiliary random variable is not needed to describe the system, we will try to eliminate it in some way. In a conventional standard Bayesian analysis, this can be done by eliminating the auxiliary random variable by marginalization. However, a problem arises here. If the auxiliary random variable is shared by all observers, the system loses observer independence (Axiom 2 of the proposed probability theory) when it is eliminated.\n- Therefore, the proposed probability theory simply ignores the auxiliary random variable while simultaneously assuming Axiom 2.\n\nIf we were to use such a strategy, it would certainly seem that we could view the system as different from classical probability theory (as mentioned in the paper, we could of course make special cases that are equivalent to classical probability theory in special circumstances).\n\nFollowing this intuition, my interest is in what the authors' system would look like if it were represented in an abstract probability space. That is, a situation where all randomness in the world is governed by an abstract space $\\Theta$, where all randomness is lost if the abstract space is determined at a point $\\theta\\in\\Theta$, and where all variables can be described deterministically. In the abstract space, random variables are represented as a projection of the world as a map to an object, e.g., $Y(\\theta), X(\\theta), A(\\theta)$ can be uniquely determined for a given source $\\theta$. Can the authors' system be represented using such a conventional abstract probability space? Or is it a deviation from that rule?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698772236660}, {"id": "ChRiYuwGYA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4295/Reviewer_Xi4x"], "rating": "1: strong reject", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposes to introduce a new theory of probability to cope with imperfect observations, in the sense that the reported value can be different from the true experimental result. This is done through the introduction of an \"observer\", which can be imperfect, in the sense that it is noisy. The theory is then applied to various case studies.", "review_text": "The paper proposes to introduce a new theory of probability to cope with imperfect observations, in the sense that the reported value can be different from the true experimental result. This is done through the introduction of an \"observer\", which can be imperfect, in the sense that it is noisy. The theory is then applied to various case studies.", "strengths": "I do not really perceive any strong point in the paper, other than the fact that modelling imperfect observational process is an interesting, yet arguably old topic.", "weaknesses": "This paper is puzzling me in more than one ways, and I will focus on the main ones (some for which the authors can offer a rebuttal, mostly when it concerns the content and not the form of the paper). \n\nA first thing is that the paper is written in a very unusual way, at least for a paper of computer science and/or machine learning. It is very rare to directly start with a mathematical formulation, without making first an introduction (and possibly related work) positioning the proposal and its originality. \n\nA second thing is that the paper is very quick on some technical details, while being very verbose on rather basing thing such as classical applications of probabilistic conditioning. It is also a bit cryptic in terms of language as well as bit naive about some aspects. For instance, P2 top, it is not true that one cannot apply Bayes rule in continuous setting, and it has been numerous, numerous times. At this point, what means indeterminate is also quite obscure. Similarly, it is not clear for the naive Bayes what exactly means $P(A^j=a^j_{i_j}|Y=y_l)$ being not solvable? It can certainly be estimated from data, even in case of noisy observations or untrue assumptions (potentially leading to biased estimates, but it can nonetheless be estimated). \n\nA third thing is that it is unclear what authors really understand by “indeterminate”: is it that the observational process is noisy, or that the obtained probabilities are ill-known and hence that one should consider sets of possible probabilities? The paper suggests the first case, yet in such a situation I really do not see what is different between what is proposed in the paper and the consideration of noisy data where one does know or can estimate the noise process? Given that there is a huge literature on learning from noisy (and/or imprecise) data, at least a positioning with respect to those should be done. Indeed, if the main idea of the paper is to have $P(y_{obs}=y|y_{true}=y)<1$ ($y$ here can be either the output value or a feature value) and then to proceed from that, then I would argue that considering such a situation is not new at all. Similarly, if indeterminate means ill-defined, then there is a whole literature about that (see, e.g., work following the book on Peter Walley on imprecise probabilities and similar). Claiming to build a new theory of probability should be backed up by being very precise about why previous theories do not answer the considered problem. \n\nA fourth thing is that it is really unclear to me why the current experiments, that merely show accuracy results for standard problems, do show that the theory is “valid”? I would equally question a statistical learning theory or more generally an uncertainty theory whose axioms cannot be the subject of tests and falsification? All theories of uncertainty I know of that are a bit serious in terms of operationally are subject to falsifiability, and this especially true for probabilistic theories (see the Ellsberg paradox for a good example of attempted falsification). Also, since Softmax does not enjoy peculiarly good properties from a theoretical perspective, I would not consider it as a strong baselines against which to test the axioms of a theory?", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes to introduce a new theory of probability to cope with imperfect observations, in the sense that the reported value can be different from the true experimental result. This is done through the introduction of an \"observer\", which can be imperfect, in the sense that it is noisy. The theory is then applied to various case studies.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "strengths": "I do not really perceive any strong point in the paper, other than the fact that modelling imperfect observational process is an interesting, yet arguably old topic.", "weaknesses": "This paper is puzzling me in more than one ways, and I will focus on the main ones (some for which the authors can offer a rebuttal, mostly when it concerns the content and not the form of the paper). \n\nA first thing is that the paper is written in a very unusual way, at least for a paper of computer science and/or machine learning. It is very rare to directly start with a mathematical formulation, without making first an introduction (and possibly related work) positioning the proposal and its originality. \n\nA second thing is that the paper is very quick on some technical details, while being very verbose on rather basing thing such as classical applications of probabilistic conditioning. It is also a bit cryptic in terms of language as well as bit naive about some aspects. For instance, P2 top, it is not true that one cannot apply Bayes rule in continuous setting, and it has been numerous, numerous times. At this point, what means indeterminate is also quite obscure. Similarly, it is not clear for the naive Bayes what exactly means $P(A^j=a^j_{i_j}|Y=y_l)$ being not solvable? It can certainly be estimated from data, even in case of noisy observations or untrue assumptions (potentially leading to biased estimates, but it can nonetheless be estimated). \n\nA third thing is that it is unclear what authors really understand by “indeterminate”: is it that the observational process is noisy, or that the obtained probabilities are ill-known and hence that one should consider sets of possible probabilities? The paper suggests the first case, yet in such a situation I really do not see what is different between what is proposed in the paper and the consideration of noisy data where one does know or can estimate the noise process? Given that there is a huge literature on learning from noisy (and/or imprecise) data, at least a positioning with respect to those should be done. Indeed, if the main idea of the paper is to have $P(y_{obs}=y|y_{true}=y)<1$ ($y$ here can be either the output value or a feature value) and then to proceed from that, then I would argue that considering such a situation is not new at all. Similarly, if indeterminate means ill-defined, then there is a whole literature about that (see, e.g., work following the book on Peter Walley on imprecise probabilities and similar). Claiming to build a new theory of probability should be backed up by being very precise about why previous theories do not answer the considered problem. \n\nA fourth thing is that it is really unclear to me why the current experiments, that merely show accuracy results for standard problems, do show that the theory is “valid”? I would equally question a statistical learning theory or more generally an uncertainty theory whose axioms cannot be the subject of tests and falsification? All theories of uncertainty I know of that are a bit serious in terms of operationally are subject to falsifiability, and this especially true for probabilistic theories (see the Ellsberg paradox for a good example of attempted falsification). Also, since Softmax does not enjoy peculiarly good properties from a theoretical perspective, I would not consider it as a strong baselines against which to test the axioms of a theory?", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethical concerns", "rating": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698589216177}, {"id": "jzU7qTbBWa", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4295/Reviewer_S3ce"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes \"Indeterminate Probability Theory\", which is claimed as an extension of classical probability theory. Based on the proposed theory, the authors derive an analytical expression of general posterior, which has some applications such as IPNN and CIPNN. Experimental results validate the proposed theory.", "review_text": "This paper proposes \"Indeterminate Probability Theory\", which is claimed as an extension of classical probability theory. Based on the proposed theory, the authors derive an analytical expression of general posterior, which has some applications such as IPNN and CIPNN. Experimental results validate the proposed theory.", "strengths": "This paper claims to extend classical probability theory, which is very ambitious and is definitely important if this is true.", "weaknesses": "The main contribution of this paper is the proposed \"Indeterministic Probability Theory\", but it is far from satisfaction to be a theory, especially when it is stated to be \"an extension of classical probability theory\". It is actually built on the axioms of classical probability theory, added with a specific generation process of random variables, and three proposed \"candidate axioms\", thus it at most becomes \"a special sub-field of classical probability theory\". \n\nEven worse, the paper claims that \"our most important contribution is that we propose a new **general analytical** and **tractable** probability equation\", but neither is theoretically validated: for **general analytical**, it is analytical, but it is not discussed enough why the proposed two-phase protocol is general; for **tractable**, it is also not verified the error of approximation via Monte Carlo methods. There are some experimental results to verify the effectiveness of Monte Carlo, but is over-simplified to validate it in such a general theory as is claimed, and more importantly, the effectiveness of Monte Carlo in this paper is not \"proved\" yet. \n\nTo be honest, section 3 is more like a section of \"problem formulation + proposed approach\": The two-phase protocol is more like the problem formulation, the axioms are more like some assumptions of independence, and the complexity reduction using Monte Carlo is more like the proposed approach.", "questions": "What do you want to say in Section 2? It seems that the example does not go beyond classical probability theory, i.e., all definitions, quantities and calculations are consistent with definitions and axioms in classical probability theory.\n\nWhat is new in your indeterminate probability theory? Specifically, I am confused why eq. (4) must be 0 or 1 in classical probability theory. Could the authors give some references? The authors should give references, clear derivations or rigorous counter-examples when refuting something in classical probability theory, as it is based on rigorous mathematics. \n\nHow general your proposed theory is? For example, does your theory enable A and Y to be any kind of random variables, and can the two-phase protocol in your theory model any data-generation process? If not, then the generality of your theory should be discussed. \n\nIn page 5 the authors say \"...Otherwise, Candidate Axiom 2 and Candidate Axiom 3 cannot both be true\". In my opinion, it is strange to discuss the soundness of an axiom once it is proposed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes \"Indeterminate Probability Theory\", which is claimed as an extension of classical probability theory. Based on the proposed theory, the authors derive an analytical expression of general posterior, which has some applications such as IPNN and CIPNN. Experimental results validate the proposed theory.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "strengths": "This paper claims to extend classical probability theory, which is very ambitious and is definitely important if this is true.", "weaknesses": "The main contribution of this paper is the proposed \"Indeterministic Probability Theory\", but it is far from satisfaction to be a theory, especially when it is stated to be \"an extension of classical probability theory\". It is actually built on the axioms of classical probability theory, added with a specific generation process of random variables, and three proposed \"candidate axioms\", thus it at most becomes \"a special sub-field of classical probability theory\". \n\nEven worse, the paper claims that \"our most important contribution is that we propose a new **general analytical** and **tractable** probability equation\", but neither is theoretically validated: for **general analytical**, it is analytical, but it is not discussed enough why the proposed two-phase protocol is general; for **tractable**, it is also not verified the error of approximation via Monte Carlo methods. There are some experimental results to verify the effectiveness of Monte Carlo, but is over-simplified to validate it in such a general theory as is claimed, and more importantly, the effectiveness of Monte Carlo in this paper is not \"proved\" yet. \n\nTo be honest, section 3 is more like a section of \"problem formulation + proposed approach\": The two-phase protocol is more like the problem formulation, the axioms are more like some assumptions of independence, and the complexity reduction using Monte Carlo is more like the proposed approach.", "questions": "What do you want to say in Section 2? It seems that the example does not go beyond classical probability theory, i.e., all definitions, quantities and calculations are consistent with definitions and axioms in classical probability theory.\n\nWhat is new in your indeterminate probability theory? Specifically, I am confused why eq. (4) must be 0 or 1 in classical probability theory. Could the authors give some references? The authors should give references, clear derivations or rigorous counter-examples when refuting something in classical probability theory, as it is based on rigorous mathematics. \n\nHow general your proposed theory is? For example, does your theory enable A and Y to be any kind of random variables, and can the two-phase protocol in your theory model any data-generation process? If not, then the generality of your theory should be discussed. \n\nIn page 5 the authors say \"...Otherwise, Candidate Axiom 2 and Candidate Axiom 3 cannot both be true\". In my opinion, it is strange to discuss the soundness of an axiom once it is proposed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697617155207}], "openreview_url": "https://openreview.net/forum?id=sSWGqY2qNJ", "arxiv_id": "2303.11536", "paper_pdf": "papers/sSWGqY2qNJ.pdf", "paper_pdf_sha256": "3b6ee3a5860119d31030839d96a66511286f45a5efda551896e11968e43a488d", "paper_pdf_bytes": 1180207, "paper_pdf_source": "openreview", "code_url": "https://github.com/Starfruit007/ipnn", "code_repository": "Starfruit007/ipnn", "code_commit": "987666de7d4bead745d498aa17a80f05bad3fe06", "code_archive": "repos/sSWGqY2qNJ.zip", "code_archive_sha256": "d3134f21e683002d35f966b950d36aca435fc77188cd44b5788cea9531912ce9", "code_archive_bytes": 68900, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 54, "github_languages": {"Python": 84780}, "github_archived": false, "github_pushed_at": "2023-10-05T05:37:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/indeterminate-probability-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dfPuLye6RvY", "year": 2023, "status": "rejected", "title": "Light-weight probing of unsupervised representations for Reinforcement Learning", "authors": ["Wancong Zhang", "Anthony GX-Chen", "Vlad Sobal", "Yann LeCun", "Nicolas Carion"], "authorids": ["~Wancong_Zhang1", "~Anthony_GX-Chen1", "~Vlad_Sobal1", "~Yann_LeCun1", "~Nicolas_Carion1"], "authors_source": "OpenReview API", "abstract": "Unsupervised visual representation learning offers the opportunity to leverage large corpora of unlabeled trajectories to form useful visual representations, which can benefit the training of reinforcement learning (RL) algorithms. However, evaluating the fitness of such representations requires training RL algorithms which is both computationally intensive and has high variance outcomes. To alleviate this issue, we design an evaluation protocol for unsupervised RL representations with lower variance and up to 600x lower computational cost. Inspired by the vision community, we propose two linear probing tasks: predicting the reward observed in a given state, and predicting the action of an expert in a given state. These two tasks are generally applicable to many RL domains, and we show through rigorous experimentation that they correlate strongly with the actual downstream control performance on the Atari100k Benchmark. This provides a better method for exploring the space of pretraining algorithms without the need of running RL evaluations for every setting. Leveraging this framework, we further improve existing self-supervised learning (SSL) recipes for RL, highlighting the importance of the forward model, the size of the visual backbone, and the precise formulation of the unsupervised objective.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0t8sTNccIZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1254/Reviewer_Vgoa"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method for evaluating unsupervised representation learning in reinforcement learning.  Using a linear probe on top of frozen, pretrained representations, the paper suggests learning to predict reward values from various states in downstream tasks.  Additionally, the paper uses a linear probe to predict expert actions from learned representations.  They authors show evidence that, for a selection of representation learning approaches, the F1 score of the linear probe correlates strongly with full reinforcement learning on the downstream task.", "review_text": "The paper presents an interesting approach to an interesting problem, with the promise of helping evaluate representations much more quickly.  Although the ablation studies are thorough, there could be a broader comparison to other styles of representation learning, which would significantly strengthen the claims that linear probing for reward prediction correlates well with downstream RL training.", "strengths": "The paper tackles a very difficult and relevant problem, that of evaluating self-supervised representations.  The paper shows evidence that linear probing can give strong indications of eventual RL training performance, which promises to shorten evaluation time and could be impactful in the representation learning for reinforcement learning field.  My main concern with the paper is the lack of diversity in methods used to assess the correlation between linear probes and RL training performance.  All methods compared are ablations of the self-predictive representation approach described in the paper.  While these are important and elucidating experiments, I would like to see a broader set of methods compared, like augmentation-based representations (DrQ or CURL).  Do these correlations hold in these cases as well?  Also, I'm curious about the noise in the linear probe F1 score.  Do the numbers reported in the tables stay the same regardless of random seed?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a method for evaluating unsupervised representation learning in reinforcement learning.  Using a linear probe on top of frozen, pretrained representations, the paper suggests learning to predict reward values from various states in downstream tasks.  Additionally, the paper uses a linear probe to predict expert actions from learned representations.  They authors show evidence that, for a selection of representation learning approaches, the F1 score of the linear probe correlates strongly with full reinforcement learning on the downstream task.", "strength_and_weaknesses": "The paper tackles a very difficult and relevant problem, that of evaluating self-supervised representations.  The paper shows evidence that linear probing can give strong indications of eventual RL training performance, which promises to shorten evaluation time and could be impactful in the representation learning for reinforcement learning field.  My main concern with the paper is the lack of diversity in methods used to assess the correlation between linear probes and RL training performance.  All methods compared are ablations of the self-predictive representation approach described in the paper.  While these are important and elucidating experiments, I would like to see a broader set of methods compared, like augmentation-based representations (DrQ or CURL).  Do these correlations hold in these cases as well?  Also, I'm curious about the noise in the linear probe F1 score.  Do the numbers reported in the tables stay the same regardless of random seed?", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, the experiments are carefully done and interesting ablations are conducted.  Although linear probing is common in computer vision representation evaluation, the generalization to RL and reward prediction is novel as far as I am aware.", "summary_of_the_review": "The paper presents an interesting approach to an interesting problem, with the promise of helping evaluate representations much more quickly.  Although the ablation studies are thorough, there could be a broader comparison to other styles of representation learning, which would significantly strengthen the claims that linear probing for reward prediction correlates well with downstream RL training.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666677810215}, {"id": "G3paIwoJ4X", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1254/Reviewer_LabE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper develops an evaluation protocol for unsupervised visual pretraining. They learn linear probes to predict expert agent actions and rewards from encoded states. These probes provide a more cost-efficient way of comparing visual representation learning methods for RL. The evaluation protocol is tested on a handful of Atari tasks and the authors show that the performance of networks on the linear probes well correlates with RL performance.", "review_text": "The paper tackles an interesting problem, but I think it should include evaluation over a larger set of ", "strengths": "This paper introduces an interesting idea for the important problem of cost-effective evaluations of visual representations. Currently results are limited to Atari and the evaluation is only performed for a small number of models + self-supervised losses. Because the main goal of this paper is to provide an evaluation protocol that can be used in place of downstream RL performance, it would be helpful to see a much broader range of losses and models as well as more difficult control tasks. It's also unclear how well these evaluation protocols predict downstream performance in the presence of task transfer: e.g., one goal in developing visual representation pre-training methods is to get a good generalizable encoder. Because the evaluation leverages reward information and information about the optimal policy, it doesn't seem like it would predict the fitness of an encoder for new tasks.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper develops an evaluation protocol for unsupervised visual pretraining. They learn linear probes to predict expert agent actions and rewards from encoded states. These probes provide a more cost-efficient way of comparing visual representation learning methods for RL. The evaluation protocol is tested on a handful of Atari tasks and the authors show that the performance of networks on the linear probes well correlates with RL performance.", "strength_and_weaknesses": "This paper introduces an interesting idea for the important problem of cost-effective evaluations of visual representations. Currently results are limited to Atari and the evaluation is only performed for a small number of models + self-supervised losses. Because the main goal of this paper is to provide an evaluation protocol that can be used in place of downstream RL performance, it would be helpful to see a much broader range of losses and models as well as more difficult control tasks. It's also unclear how well these evaluation protocols predict downstream performance in the presence of task transfer: e.g., one goal in developing visual representation pre-training methods is to get a good generalizable encoder. Because the evaluation leverages reward information and information about the optimal policy, it doesn't seem like it would predict the fitness of an encoder for new tasks.", "clarity,_quality,_novelty_and_reproducibility": "Overall I thought the presentation was straightforward save a few minor confusions:\nIs section 4.1 a contribution of this paper? From the introduction and abstract, I assumed that the paper's main contribution was the evaluation protocol, but it was unclear if the architecture in Figure 2 was adapted from past work or newly introduced for this task. I think this would be helpful to clarify because it would be useful to see the evaluation protocol on multiple kinds of models or on models developed in past work.", "summary_of_the_review": "The paper tackles an interesting problem, but I think it should include evaluation over a larger set of ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666664388955}, {"id": "FGDHxiYmX3x", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1254/Reviewer_etov"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper attempts to propose an evaluation protocol for lightweight probing of unsupervised representations and investigates the correlation between RL performance and linear probing from a pretrained representation. Authors are testing this on a very specific class of self-predictive (recurrent) representation models that are being trained with SSL objectives. The authors extensively analyse several design choices of their studied model regarding performance on the two probing tasks of predicting reward or action from a held-out labelled train set. Finally, the paper aims to present some correlation between probing and RL performance on 9 Atari games.\n", "review_text": "While this paper attempts to address an important challenge in RL and aims to propose a generally applicable evaluation protocol for unsupervised pretrained RL representations I am not fully convinced that the paper holds up to these promises and several claims in the paper. I am willing to change my opinion in case I missed central parts of the paper but want to encourage the authors to further improve the manuscript based on some suggestions and comments above and recommend rejection.", "strengths": "Strengths\n\n1. Identification of an important problem setup in RL, that is how to assess which representations and pretraining data is best suited for improving downstream RL performance.\n2. Extensive analysis of various design choices of the model architecture studied\n\nWeaknesses\n1. It is not very clear what the goal or objective of the paper is. Authors say they propose an evaluation protocol for unsupervised RL representations that saves up to 600x computation cost. However, this computational cost saving and protocol seems to be not addressed or described in detail in the main paper and doesn't appear to be the main focus of the paper. Instead, the authors analyse a very specific class of self-predictive (recurrent) representation models that are being trained with SSL objectives. I would have expected to see a more thorough coverage of different unsupervised representation learning methods and more empirical analysis to support these claims (see additional comments in point 3)\n2. It is not very clear what is actually novel about the proposed model and what is based on prior work. The listed contributions are very vague. I would hope the authors can clarify what exactly their contributions are. Also, there exist additional prior works that very extensively studied various light-weight probing tasks on unsupervised representations and how their performance relates to RL performances [1,2]\n3. I am a bit confused by the experiments, especially the correlative analysis and it is not very clear to me if this holds beyond the very particular method and environments. In particular Figure 3 is very confusing and I do not fully understand what the 7 representative setups are supposed to be; how this relates to the 9 rows/models presented in Figure 3 (where do they come from?); and how this relates to the 7 Atari games studied in Figure 1. Without this additional information, it is not clear to me if it is sound to draw a general conclusion about the correlation between probing task performance and RL performance or if there are any other confounders. Also, can authors provide error bars for results in Table 1-5?\n4. Minor: There appears to be a slight mismatch between the title/abstract and the presented experiments and contributions of the main paper. The authors study a very particular narrow setting but the title suggests a generally applicable evaluation protocol for different unsupervised representation learning methods. \n5. Regarding the two probing tasks I wonder how generally applicable they really are. E.g. the reward task seems to be constrained to have labelled data from a very early point in RL training (i.e. rather random policy), whereas the action prediction is limited to labelled data from close to expert trajectories. To be generally applicable I would have hoped to see both probes being trained on the same labelled dataset to compare apples to apples. \n\n[1] Dittadi, Andrea et al. “The Role of Pretrained Representations for the OOD Generalization of Reinforcement Learning Agents”, ICLR 2022.\n\n[2] Higgins, Irina, et al. \"Darla: Improving zero-shot transfer in reinforcement learning.\" International Conference on Machine Learning. PMLR, 2017.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper attempts to propose an evaluation protocol for lightweight probing of unsupervised representations and investigates the correlation between RL performance and linear probing from a pretrained representation. Authors are testing this on a very specific class of self-predictive (recurrent) representation models that are being trained with SSL objectives. The authors extensively analyse several design choices of their studied model regarding performance on the two probing tasks of predicting reward or action from a held-out labelled train set. Finally, the paper aims to present some correlation between probing and RL performance on 9 Atari games.\n", "strength_and_weaknesses": "Strengths\n\n1. Identification of an important problem setup in RL, that is how to assess which representations and pretraining data is best suited for improving downstream RL performance.\n2. Extensive analysis of various design choices of the model architecture studied\n\nWeaknesses\n1. It is not very clear what the goal or objective of the paper is. Authors say they propose an evaluation protocol for unsupervised RL representations that saves up to 600x computation cost. However, this computational cost saving and protocol seems to be not addressed or described in detail in the main paper and doesn't appear to be the main focus of the paper. Instead, the authors analyse a very specific class of self-predictive (recurrent) representation models that are being trained with SSL objectives. I would have expected to see a more thorough coverage of different unsupervised representation learning methods and more empirical analysis to support these claims (see additional comments in point 3)\n2. It is not very clear what is actually novel about the proposed model and what is based on prior work. The listed contributions are very vague. I would hope the authors can clarify what exactly their contributions are. Also, there exist additional prior works that very extensively studied various light-weight probing tasks on unsupervised representations and how their performance relates to RL performances [1,2]\n3. I am a bit confused by the experiments, especially the correlative analysis and it is not very clear to me if this holds beyond the very particular method and environments. In particular Figure 3 is very confusing and I do not fully understand what the 7 representative setups are supposed to be; how this relates to the 9 rows/models presented in Figure 3 (where do they come from?); and how this relates to the 7 Atari games studied in Figure 1. Without this additional information, it is not clear to me if it is sound to draw a general conclusion about the correlation between probing task performance and RL performance or if there are any other confounders. Also, can authors provide error bars for results in Table 1-5?\n4. Minor: There appears to be a slight mismatch between the title/abstract and the presented experiments and contributions of the main paper. The authors study a very particular narrow setting but the title suggests a generally applicable evaluation protocol for different unsupervised representation learning methods. \n5. Regarding the two probing tasks I wonder how generally applicable they really are. E.g. the reward task seems to be constrained to have labelled data from a very early point in RL training (i.e. rather random policy), whereas the action prediction is limited to labelled data from close to expert trajectories. To be generally applicable I would have hoped to see both probes being trained on the same labelled dataset to compare apples to apples. \n\n[1] Dittadi, Andrea et al. “The Role of Pretrained Representations for the OOD Generalization of Reinforcement Learning Agents”, ICLR 2022.\n\n[2] Higgins, Irina, et al. \"Darla: Improving zero-shot transfer in reinforcement learning.\" International Conference on Machine Learning. PMLR, 2017.\n", "clarity,_quality,_novelty_and_reproducibility": "I think the paper could benefit a lot from improving clarity and presentation. I would especially suggest that authors more explicitly specify the novel contributions and the scope of the work. Maybe the authors could provide more detail in section 5.4. It is possible that I missed parts when reading the manuscript but I believe novelty and originality are limited in light of my comments above.\n", "summary_of_the_review": "While this paper attempts to address an important challenge in RL and aims to propose a generally applicable evaluation protocol for unsupervised pretrained RL representations I am not fully convinced that the paper holds up to these promises and several claims in the paper. I am willing to change my opinion in case I missed central parts of the paper but want to encourage the authors to further improve the manuscript based on some suggestions and comments above and recommend rejection.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666625080078}, {"id": "y3peFpzM-w", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1254/Reviewer_vEah"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper investigates whether or not light-weight probings to the action between continuous states and reward can measure the pretrained encoder performance on RL tasks. To do this, they pretrained the encoders through self-supervised learning loss with a transition model implemented through recurrent module or recurrent state space modeling with several variations. They tested the pretrained encoder for linear probing and RL tasks. They showed the correlations between the performances on linear probings and RL tasks. They found the linear probing to reward is highly correlated with performance on RL tasks, while the relationship between action linear probing and RL performance is weaker.", "review_text": "This paper investigates the linear probing to the action, and reward could be an indicator of RL performance. Mainly they evaluated the SSL methods, BYOL, and Barlow Twins for 9 Atari tasks. It is interesting and could be helpful for others because applying the pretrained encoder to RL tasks is one of the ways to improve the sample efficiency on RL, but the evaluation is very expensive. However, I felt their writing is rushed, so I think minor revision is necessary. ", "strengths": "Strength\n- The motivation behind this paper is good. Applying the pretrained encoders to RL tasks has been investigated [1,2]. It requires lots of resources. This investigation could be a piece of good evidence to skip the costly evaluation.\n- The many variations and ablations are evaluated. For transition modeling, deterministic recurrent modules and RSSM are validated. For SSL, BYOL and Barlow Twins are tested with various configurations. The ablation studies are reported with and without each objective, such as inverse dynamics modeling and goal-conditioned RL.\n\nWeaknesses\n- They only investigated the SSL methods, not other unsupervised methods such as VAE. I expected they would cover the overall methods from their title, but it is not.\n- They only evaluated nine games. It could not be enough to back up the conclusion.\n- In the equation for reward-reg loss in Reward Probing in section 3.2, should the encoder get $o_{t+1}$ not $o_t$? Because the reward $r_t$ is given when the action is given on the observation $o_t$. \n- In the equation for action-classif loss in Action Prediction in section 3.2, shouldn't the input be $o_t$ and $o_{t+1}$? In Figure 2, the consecutive observations are given to the loss, but not in this equation.\n- In Figure 2, I cannot understand the below sentence. Why is the stacked observation related to data augmentation?\n    - > The observations consist of a stack of 4 frames, to which we apply data augmentation before passing them to a convolutional encoder.\n-  > The action is represented as a 2D one-hot vector and appended to the input to the first convolutional layer.\n    - Why is the action represented through a 2D one-hot vector? The action space is larger than the 2 Dimension.\n- For RSSM, you used the discrete latent. Why didn't you try the continuous latent version [3]? Perhaps, because the discrete version outperforms the continuous latent version in [4], but the discrete latent variable training is more unstable than the continuous latent variable training, so maybe the RSSM with continuous latent variable could outperform the discrete latent version.\n- In BYOL loss equation in section 4.2, should $q(\\hat{y}_{t+k})$ be $q(\\hat{e}_{t+k})$?\n- For Algorithm 1, why did you use the Pseudo code block? It is just a single equation.\n- For the goal-oriented RL loss, please introduce it roughly in the main paper even though the details are in Appendix.\n    - In A.7, there are typo $\\tilde{e}_t+1$.\n- > Similar to the BYOL model, the Barlow model can also be improved with inverse dynamics modeling, while the addition of goal loss has a slight negative impact.\n    - It is interesting. Could you analyze this?\n\n[1] Schwarzer, Max, et al. \"Pretraining representations for data-efficient reinforcement learning.\" Advances in Neural Information Processing Systems 34 (2021): 12686-12699.\n\n[2] Dittadi, Andrea, et al. \"The role of pretrained representations for the ood generalization of rl agents.\" arXiv preprint arXiv:2107.05686 (2021).\n\n[3] Hafner, Danijar, et al. \"Dream to control: Learning behaviors by latent imagination.\" arXiv preprint arXiv:1912.01603 (2019).\n\n[4] Hafner, Danijar, et al. \"Mastering atari with discrete world models.\" arXiv preprint arXiv:2010.02193 (2020).", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper investigates whether or not light-weight probings to the action between continuous states and reward can measure the pretrained encoder performance on RL tasks. To do this, they pretrained the encoders through self-supervised learning loss with a transition model implemented through recurrent module or recurrent state space modeling with several variations. They tested the pretrained encoder for linear probing and RL tasks. They showed the correlations between the performances on linear probings and RL tasks. They found the linear probing to reward is highly correlated with performance on RL tasks, while the relationship between action linear probing and RL performance is weaker.", "strength_and_weaknesses": "Strength\n- The motivation behind this paper is good. Applying the pretrained encoders to RL tasks has been investigated [1,2]. It requires lots of resources. This investigation could be a piece of good evidence to skip the costly evaluation.\n- The many variations and ablations are evaluated. For transition modeling, deterministic recurrent modules and RSSM are validated. For SSL, BYOL and Barlow Twins are tested with various configurations. The ablation studies are reported with and without each objective, such as inverse dynamics modeling and goal-conditioned RL.\n\nWeaknesses\n- They only investigated the SSL methods, not other unsupervised methods such as VAE. I expected they would cover the overall methods from their title, but it is not.\n- They only evaluated nine games. It could not be enough to back up the conclusion.\n- In the equation for reward-reg loss in Reward Probing in section 3.2, should the encoder get $o_{t+1}$ not $o_t$? Because the reward $r_t$ is given when the action is given on the observation $o_t$. \n- In the equation for action-classif loss in Action Prediction in section 3.2, shouldn't the input be $o_t$ and $o_{t+1}$? In Figure 2, the consecutive observations are given to the loss, but not in this equation.\n- In Figure 2, I cannot understand the below sentence. Why is the stacked observation related to data augmentation?\n    - > The observations consist of a stack of 4 frames, to which we apply data augmentation before passing them to a convolutional encoder.\n-  > The action is represented as a 2D one-hot vector and appended to the input to the first convolutional layer.\n    - Why is the action represented through a 2D one-hot vector? The action space is larger than the 2 Dimension.\n- For RSSM, you used the discrete latent. Why didn't you try the continuous latent version [3]? Perhaps, because the discrete version outperforms the continuous latent version in [4], but the discrete latent variable training is more unstable than the continuous latent variable training, so maybe the RSSM with continuous latent variable could outperform the discrete latent version.\n- In BYOL loss equation in section 4.2, should $q(\\hat{y}_{t+k})$ be $q(\\hat{e}_{t+k})$?\n- For Algorithm 1, why did you use the Pseudo code block? It is just a single equation.\n- For the goal-oriented RL loss, please introduce it roughly in the main paper even though the details are in Appendix.\n    - In A.7, there are typo $\\tilde{e}_t+1$.\n- > Similar to the BYOL model, the Barlow model can also be improved with inverse dynamics modeling, while the addition of goal loss has a slight negative impact.\n    - It is interesting. Could you analyze this?\n\n[1] Schwarzer, Max, et al. \"Pretraining representations for data-efficient reinforcement learning.\" Advances in Neural Information Processing Systems 34 (2021): 12686-12699.\n\n[2] Dittadi, Andrea, et al. \"The role of pretrained representations for the ood generalization of rl agents.\" arXiv preprint arXiv:2107.05686 (2021).\n\n[3] Hafner, Danijar, et al. \"Dream to control: Learning behaviors by latent imagination.\" arXiv preprint arXiv:1912.01603 (2019).\n\n[4] Hafner, Danijar, et al. \"Mastering atari with discrete world models.\" arXiv preprint arXiv:2010.02193 (2020).", "clarity,_quality,_novelty_and_reproducibility": "Their motivation, model design to evaluate, and evaluation are clearly written except for some minor things, such as the equation typos. Their investigation is novel and looks reproducible through the hyperparameters shared in Appendix.", "summary_of_the_review": "This paper investigates the linear probing to the action, and reward could be an indicator of RL performance. Mainly they evaluated the SSL methods, BYOL, and Barlow Twins for 9 Atari tasks. It is interesting and could be helpful for others because applying the pretrained encoder to RL tasks is one of the ways to improve the sample efficiency on RL, but the evaluation is very expensive. However, I felt their writing is rushed, so I think minor revision is necessary. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666495090012}], "openreview_url": "https://openreview.net/forum?id=dfPuLye6RvY", "arxiv_id": "2208.12345", "paper_pdf": "papers/dfPuLye6RvY.pdf", "paper_pdf_sha256": "71b212d02add11fa5b302f115318ea399f09a8404dfe9fac8943cdba9fbc3fae", "paper_pdf_bytes": 595693, "paper_pdf_source": "openreview", "code_url": "https://github.com/kevinghst/lightweight_rl_probe", "code_repository": "kevinghst/lightweight_rl_probe", "code_commit": "0467b1f0f5aecafa780a52c85be724b493fd6a4e", "code_archive": "repos/dfPuLye6RvY.zip", "code_archive_sha256": "267b2bdafe934d3e416b35e949ef679a24b09c356ca552af2bc636a2cf8c251a", "code_archive_bytes": 80419, "code_file_count": 37, "code_extensions": {".sh": 22, ".py": 15}, "github_disk_usage_kb": 93, "github_languages": {"Python": 182716, "Shell": 41459}, "github_archived": false, "github_pushed_at": "2022-08-13T15:24:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/light-weight-probing-of-unsupervised"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e_FK_rDajEv", "year": 2022, "status": "rejected", "title": "Learning Neural Causal Models with Active Interventions", "authors": ["Nino Scherrer", "Olexa Bilaniuk", "Yashas Annadani", "Anirudh Goyal", "Patrick Schwab", "Bernhard Schölkopf", "Michael Curtis Mozer", "Yoshua Bengio", "Stefan Bauer", "Nan Rosemary Ke"], "authorids": ["~Nino_Scherrer1", "~Olexa_Bilaniuk1", "~Yashas_Annadani1", "~Anirudh_Goyal1", "~Patrick_Schwab1", "~Bernhard_Schölkopf1", "~Michael_Curtis_Mozer1", "~Yoshua_Bengio1", "~Stefan_Bauer1", "~Nan_Rosemary_Ke1"], "authors_source": "OpenReview API", "abstract": "Discovering causal structures from data is a challenging inference problem of fundamental importance in all areas of science. The appealing scaling properties of neural networks have recently led to a surge of interest in differentiable neural network-based methods for learning causal structures from data. So far, differentiable causal discovery has focused on static datasets of observational or interventional origin. In this work, we introduce an active intervention-targeting mechanism which enables quick identification of the underlying causal structure of the data-generating process. Our method significantly reduces the required number of interactions compared with random intervention targeting and is applicable for both discrete and continuous optimization formulations of learning the underlying directed acyclic graph (DAG) from data. We examine the proposed method across multiple frameworks in a wide range of settings and demonstrate superior performance on multiple benchmarks from simulated to real-world data. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "OR0GJc4_jIA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1400/Reviewer_ma3K"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors take a further step by introducing active interventions in differentiable neural network-based methods. The intervention target is selected based on maximizing the disagreement between the post-interventional sample distributions under the hypothesis graphs. The experimental results verify the effectiveness of the active strategy compared to random selection.", "review_text": "Strengths: \n1. The paper is written clearly;\n2. the tackled problem is well-motivated;\n3. I think the idea that select the interventional variable by maximizing the disagreement between the post-interventional sample distributions under the hypothesis graphs is novel.\n\nWeaknesses:\nOne main aspect where some improvements could make the paper better is the experiment part. The comparison results between the proposed method and DCDI are supportive in only verifying that the active strategy is better than random selection. However, according to my understanding, although the novel point of this paper is to present an active causal discovery method based on continuous optimization framework, the tackled problem is still actively causal discovery. In the literature, there have been many classical methods regarding actively causal discovery (e.g. Active learning of causal networks with intervention experiments and optimal designs, Two optimal strategies for active learning of causal models from interventional data). Hence, in addition to illustrating that under continuous optimization framework the active strategy works, it is necessary to illustrate that the proposed method could preform better than the previous active causal discovery methods, which convince the readers that the proposed method is the best in this kind of tasks. In addition, could the authors give more details about the baselines? The baselines do not quite match the tasks in this paper, hence I want to know some execution details. For example, GIES is for obser. + passive inter., the active strategy is not involved. How do the authors select the intervention variable? ICP does not exploit the information about which variables are softly intervened (taken by dynamic environment). The comparison between the proposed method and ICP seems not quite fair. Do authors make some modifications? Towards active causal discovery, I think the two methods I mentioned above seem to be more suitable as baseline methods.\n\nI have one another question. The authors claim that the proposed method is applicable for multi-node intervention. I agree with that. However, in this case the complexity of Line 2 of Algorithm 1 is quite large. How do the authors address this problem? \n\nOther suggestions or typos: \n1. I think it will be better if the authors give more details about training functional parameters in the paper or appendix. The current version made me have to read paper DSDI.\n\n2. A missing full point in the paragraph of Assumptions (Page 3).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors take a further step by introducing active interventions in differentiable neural network-based methods. The intervention target is selected based on maximizing the disagreement between the post-interventional sample distributions under the hypothesis graphs. The experimental results verify the effectiveness of the active strategy compared to random selection.", "main_review": "Strengths: \n1. The paper is written clearly;\n2. the tackled problem is well-motivated;\n3. I think the idea that select the interventional variable by maximizing the disagreement between the post-interventional sample distributions under the hypothesis graphs is novel.\n\nWeaknesses:\nOne main aspect where some improvements could make the paper better is the experiment part. The comparison results between the proposed method and DCDI are supportive in only verifying that the active strategy is better than random selection. However, according to my understanding, although the novel point of this paper is to present an active causal discovery method based on continuous optimization framework, the tackled problem is still actively causal discovery. In the literature, there have been many classical methods regarding actively causal discovery (e.g. Active learning of causal networks with intervention experiments and optimal designs, Two optimal strategies for active learning of causal models from interventional data). Hence, in addition to illustrating that under continuous optimization framework the active strategy works, it is necessary to illustrate that the proposed method could preform better than the previous active causal discovery methods, which convince the readers that the proposed method is the best in this kind of tasks. In addition, could the authors give more details about the baselines? The baselines do not quite match the tasks in this paper, hence I want to know some execution details. For example, GIES is for obser. + passive inter., the active strategy is not involved. How do the authors select the intervention variable? ICP does not exploit the information about which variables are softly intervened (taken by dynamic environment). The comparison between the proposed method and ICP seems not quite fair. Do authors make some modifications? Towards active causal discovery, I think the two methods I mentioned above seem to be more suitable as baseline methods.\n\nI have one another question. The authors claim that the proposed method is applicable for multi-node intervention. I agree with that. However, in this case the complexity of Line 2 of Algorithm 1 is quite large. How do the authors address this problem? \n\nOther suggestions or typos: \n1. I think it will be better if the authors give more details about training functional parameters in the paper or appendix. The current version made me have to read paper DSDI.\n\n2. A missing full point in the paragraph of Assumptions (Page 3).", "summary_of_the_review": "The main reason that I give a slightly negative score in the current is that I think the experiments are not supportive enough. Different from some theoretical causality paper, this paper tackles a practical and widely researched problem. Hence I think the experiments need to be designed carefully. I look forward to more clues from the authors. If they could address my concerns and questions, I am very happy to increase my score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635733788199}, {"id": "2LCLBXMDKXt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1400/Reviewer_KxZA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a method to select interventions that enable efficient identification of the underlying causal structure. In particular, the method picks the intervention that exhibits the highest discrepancies between post-interventional sample distributions generated. The authors provide experiment results to demonstrate that the proposed method has a better performance over the other baselines, and also improves the sample efficiency.", "review_text": "Overall, the proposed method is interesting and novel. However, since the proposed method is based on a heuristic without theoretical analysis, as also acknowledged by the authors, I think it is important to demonstrate empirically with a fair experiment setup that the proposed method works well compared to the existing baselines. I am willing to increase my score if the authors are able to address my comments below.\n\nStrengths:\n- The paper is well written.\n- The proposed method is interesting and novel.\n\nWeaknesses:\n- The proposed method lacks theoretical analysis.\n- It is not surprising that GES, NOTEARS, and DAG-GNN do not perform well, as they are based on observational data. I would encourage the authors to include more baselines that handle interventional data, such as IGSP [1] or UT-IGSP [2], together with nonparametric test.\n- It would be better to provide what score is used for GIES and GES. Have the authors considered the generalized score [3] that better handles nonparametric relationship for a fairer comparison?\n- Based on my understanding, the authors used the original ICP [4] that is based on linear model. It may be better to also include its nonlinear variant [5].\n- Figure 23 seems to show that DCDI performs better than DCDI+AIT. It would be better to include similar experiments in Tables 1 and 2 for DCDI.\n- How did the authors ensure a fair comparison between the proposed method based on active intervention, and the baselines based on observational data and/or random interventions? E.g., what are the number of samples/interventions used for both cases?\n- The paper would be stronger if it includes the structural intervention distance (SID) [6] that may also be informative.\n\nMinor comments:\n- Based on my understanding, the experiments in Tables 1 and 2 are based on discrete data. It would be better to explain how NOTEARS handles discrete data.\n- For the real datasets considered in Section 4, did the authors experiment with the original datasets, or only adopt their ground truth causal graphs and generate synthetic data based on these graphs? It would be better to make this clear in the main paper. The current description sounds like it is the former (e.g., the last two sentences in the subsection \"Structure discovery: flow cytometry and asia dataset\"), although it seems unclear how active interventions could be applied there.\n\nReferences:\n1.  Permutation-based Causal Inference Algorithms with Interventions, 2017.\n2.  Permutation-Based Causal Structure Learning with Unknown Intervention Targets, 2020.\n3. Generalized Score Functions for Causal Discovery, 2018.\n4.  Causal inference using invariant prediction: identification and confidence intervals, 2015.\n5.  Invariant Causal Prediction for Nonlinear Models, 2018.\n6.  Structural Intervention Distance (SID) for Evaluating Causal Graphs, 2014.\n\n-----------\nAfter reading the authors' response and the other reviews, my concerns about the experiment setup persist. I agree with Reviewer ma3K that some part of the experiments could be further improved to ensure a fair experiment setup.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method to select interventions that enable efficient identification of the underlying causal structure. In particular, the method picks the intervention that exhibits the highest discrepancies between post-interventional sample distributions generated. The authors provide experiment results to demonstrate that the proposed method has a better performance over the other baselines, and also improves the sample efficiency.", "main_review": "Overall, the proposed method is interesting and novel. However, since the proposed method is based on a heuristic without theoretical analysis, as also acknowledged by the authors, I think it is important to demonstrate empirically with a fair experiment setup that the proposed method works well compared to the existing baselines. I am willing to increase my score if the authors are able to address my comments below.\n\nStrengths:\n- The paper is well written.\n- The proposed method is interesting and novel.\n\nWeaknesses:\n- The proposed method lacks theoretical analysis.\n- It is not surprising that GES, NOTEARS, and DAG-GNN do not perform well, as they are based on observational data. I would encourage the authors to include more baselines that handle interventional data, such as IGSP [1] or UT-IGSP [2], together with nonparametric test.\n- It would be better to provide what score is used for GIES and GES. Have the authors considered the generalized score [3] that better handles nonparametric relationship for a fairer comparison?\n- Based on my understanding, the authors used the original ICP [4] that is based on linear model. It may be better to also include its nonlinear variant [5].\n- Figure 23 seems to show that DCDI performs better than DCDI+AIT. It would be better to include similar experiments in Tables 1 and 2 for DCDI.\n- How did the authors ensure a fair comparison between the proposed method based on active intervention, and the baselines based on observational data and/or random interventions? E.g., what are the number of samples/interventions used for both cases?\n- The paper would be stronger if it includes the structural intervention distance (SID) [6] that may also be informative.\n\nMinor comments:\n- Based on my understanding, the experiments in Tables 1 and 2 are based on discrete data. It would be better to explain how NOTEARS handles discrete data.\n- For the real datasets considered in Section 4, did the authors experiment with the original datasets, or only adopt their ground truth causal graphs and generate synthetic data based on these graphs? It would be better to make this clear in the main paper. The current description sounds like it is the former (e.g., the last two sentences in the subsection \"Structure discovery: flow cytometry and asia dataset\"), although it seems unclear how active interventions could be applied there.\n\nReferences:\n1.  Permutation-based Causal Inference Algorithms with Interventions, 2017.\n2.  Permutation-Based Causal Structure Learning with Unknown Intervention Targets, 2020.\n3. Generalized Score Functions for Causal Discovery, 2018.\n4.  Causal inference using invariant prediction: identification and confidence intervals, 2015.\n5.  Invariant Causal Prediction for Nonlinear Models, 2018.\n6.  Structural Intervention Distance (SID) for Evaluating Causal Graphs, 2014.\n\n-----------\nAfter reading the authors' response and the other reviews, my concerns about the experiment setup persist. I agree with Reviewer ma3K that some part of the experiments could be further improved to ensure a fair experiment setup.\n", "summary_of_the_review": "Overall, the proposed method is interesting and novel. However, since the proposed method is based on a heuristic without theoretical analysis, as also acknowledged by the authors, I think it is important to demonstrate empirically with a fair experiment setup that the proposed method works well compared to the existing baselines.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635629984010}, {"id": "p92I4j9A8zv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1400/Reviewer_cZsR"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "Structure learning from observational data is a long-standing challenge that could benefit from creative uses of interventional data. The authors explore the use of active learning in a continuous optimization framework to better traverse and narrow down the space of potential graphs that describe the data. The contribution is a heuristic to choose the \"best\" intervention target based on a comparison between samples from an intervened SCM with the current best guess of an underlying causal discovery algorithm (in which the proposal is embedded).", "review_text": "Sampling from interventional distributions and quantifying its difference with the available data is creative as an approach to guide the selection of interventional targets. Unfortunately, the paper is very hard to follow and seems to rely excessively on DSDI, both to understand the proposed approach and in applications despite the authors claiming that the proposal can be used with any differentiable causal discovery algorithm (contributions). The exact procedure is complex, it involves sampling DAGs and parameters, and should be clarified. \n\nAm I correct in assuming that in every training iteration of the causal discovery algorithm, the proposed approach samples DAGs, interventional distributions, and samples from each interventional distribution to be compared using the proposed score? What is the computational complexity?\n\nAlong these lines, the exact procedure is unclear. Algorithm 1 vaguely mentions \"perform an intervention\" on line 3 without details of what kind of interventions are considered (e.g. do, soft, what exact values are set, etc.). Next, it is meant to draw samples from $\\mathcal P$ which is defined as a set of functional parameters which I understand are conditional distributions. How are the conditional distributions chosen? What is the functional relationship between parents and children?\n\nMy main concern is about the claim that one can meaningfully sample from constructed interventional distributions. For this to be valid one would need access to the underlying structural causal model in general, which is far beyond the DAG we seek to estimate in the first place - I don't believe that the constructed interventional distributions give any useful guidance to select interventions as interventional distributions are not identified in general. The use of the term \"interventional data\" I find misleading as nowhere in the algorithm description of experiments (unless I missed it) do the authors make use of a data source in which an experiment has been performed. \n\nOne argument repeatedly made in the paper is that existing structure learning algorithms are computationally expensive. Yet, experimental evaluations are made on graphs of less than 20 nodes, on relatively simple graphs, and without run time comparisons. This claim and others are far from justified.\n\nIf the main contribution is the score, this is not sufficiently discussed or justified. What makes this a good score? Do we have any guarantees that it favors good interventions?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Structure learning from observational data is a long-standing challenge that could benefit from creative uses of interventional data. The authors explore the use of active learning in a continuous optimization framework to better traverse and narrow down the space of potential graphs that describe the data. The contribution is a heuristic to choose the \"best\" intervention target based on a comparison between samples from an intervened SCM with the current best guess of an underlying causal discovery algorithm (in which the proposal is embedded).", "main_review": "Sampling from interventional distributions and quantifying its difference with the available data is creative as an approach to guide the selection of interventional targets. Unfortunately, the paper is very hard to follow and seems to rely excessively on DSDI, both to understand the proposed approach and in applications despite the authors claiming that the proposal can be used with any differentiable causal discovery algorithm (contributions). The exact procedure is complex, it involves sampling DAGs and parameters, and should be clarified. \n\nAm I correct in assuming that in every training iteration of the causal discovery algorithm, the proposed approach samples DAGs, interventional distributions, and samples from each interventional distribution to be compared using the proposed score? What is the computational complexity?\n\nAlong these lines, the exact procedure is unclear. Algorithm 1 vaguely mentions \"perform an intervention\" on line 3 without details of what kind of interventions are considered (e.g. do, soft, what exact values are set, etc.). Next, it is meant to draw samples from $\\mathcal P$ which is defined as a set of functional parameters which I understand are conditional distributions. How are the conditional distributions chosen? What is the functional relationship between parents and children?\n\nMy main concern is about the claim that one can meaningfully sample from constructed interventional distributions. For this to be valid one would need access to the underlying structural causal model in general, which is far beyond the DAG we seek to estimate in the first place - I don't believe that the constructed interventional distributions give any useful guidance to select interventions as interventional distributions are not identified in general. The use of the term \"interventional data\" I find misleading as nowhere in the algorithm description of experiments (unless I missed it) do the authors make use of a data source in which an experiment has been performed. \n\nOne argument repeatedly made in the paper is that existing structure learning algorithms are computationally expensive. Yet, experimental evaluations are made on graphs of less than 20 nodes, on relatively simple graphs, and without run time comparisons. This claim and others are far from justified.\n\nIf the main contribution is the score, this is not sufficiently discussed or justified. What makes this a good score? Do we have any guarantees that it favors good interventions?\n", "summary_of_the_review": "Ultimately my opinion is that without additional theoretical results or better justification and exposition of the proposed approach to advance the state of the art, the contribution is limited.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635606762676}], "openreview_url": "https://openreview.net/forum?id=e_FK_rDajEv", "arxiv_id": "2109.02429", "paper_pdf": "papers/e_FK_rDajEv.pdf", "paper_pdf_sha256": "fc766fc4f6b331923d57c293ffc6373f87655de76f179edcdb08931e7cd069ad", "paper_pdf_bytes": 3777007, "paper_pdf_source": "openreview", "code_url": "https://github.com/nke001/causal_learning_unknown_interventions", "code_repository": "nke001/causal_learning_unknown_interventions", "code_commit": "e41302aad36b0a5fedf6211d18cecca202c5e6b1", "code_archive": "repos/e_FK_rDajEv.zip", "code_archive_sha256": "4535c5d17fc2b03f7d2a9f9a5d8d02d4a96e2cc83e0ec9fdd979aa0b54367281", "code_archive_bytes": 61747, "code_file_count": 14, "code_extensions": {".py": 12, ".c": 1, ".m": 1}, "github_disk_usage_kb": 88, "github_languages": {"C": 95203, "Python": 91114, "MATLAB": 1633}, "github_archived": false, "github_pushed_at": "2020-07-08T01:59:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-neural-causal-models-with-active"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1x_DaVKwH", "year": 2020, "status": "rejected", "title": "Is Deep Reinforcement Learning Really Superhuman on Atari? Leveling the playing field", "authors": ["Marin Toromanoff", "Emilie Wirbel", "Fabien Moutarde"], "authorids": ["marin.toromanoff@mines-paristech.fr", "emilie.wirbel@valeo.com", "fabien.moutarde@mines-paristech.fr"], "authors_source": "OpenReview API", "abstract": "Consistent and reproducible evaluation of Deep Reinforcement Learning (DRL) is not straightforward. In the Arcade Learning Environment (ALE), small changes in environment parameters such as stochasticity or the maximum allowed play time can lead to very different performance. In this work, we discuss the difficulties of comparing different agents trained on ALE. In order to take a step further towards reproducible and comparable DRL, we introduce SABER, a Standardized Atari BEnchmark for general Reinforcement learning algorithms. Our methodology extends previous recommendations and contains a complete set of environment parameters as well as train and test procedures. We then use SABER to evaluate the current state of the art, Rainbow. Furthermore, we introduce a human world records baseline, and argue that previous claims of expert or superhuman performance of DRL might not be accurate. Finally, we propose Rainbow-IQN by extending Rainbow with Implicit Quantile Networks (IQN) leading to new state-of-the-art performance. Source code is available for reproducibility.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1gj4nbAKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper601/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper revisits the way RL algorithms are typically evaluated on the ALE benchmark, advocating for several key changes that contribute to more robust and reliable comparisons between algorithms. It also brings the following additional contributions: (1) a new measure of comparison to human performance based on actual human world records (which shows that RL algorithms are not as « super-human » as is generally believed), and (2) an evaluation (based on the proposed guidelines) of Rainbow as well as a Rainbow-IQN variant (replacing the C51 component of Rainbow with Implicit Quantile Networks), showing that the latter brings a significant improvement upon the original Rainbow algorithm.\n\nOverall I am leaning towards acceptance as I believe that such papers encouraging better benchmarking practice on Atari are definitely needed. Even if the technical contribution is limited, this paper could have a positive impact on the field by providing a clearer picture of the current state of deep RL algorithms on Atari (assuming that other researchers start following these recommendations -- and if that is not the case at least it will highlight issues with the way evaluation is currently done).\n\nI do have a few concerns / questions though:\n\n1.\tI am not convinced by the recommendation to use performance during training for evaluation purpose. In Machado et al. (2018) it is argued that « this better aligns the performance metric with the goal of continual learning », but most deep RL algorithms trained on Atari games have not been intended to be used in a continual learning setting. It definitely has the advantage of being simple, but it seems to me that it can cause some issues, like making it difficult to compare different exploration techniques for off-policy learning (the exploration may cause poor behavior during training even if it helps the agent learn a better greedy policy), and more generally not being representative of the common practical use case where the goal is to obtain the best agent possible to use in production (with no further learning). Finally, it could make results even harder to reproduce due to the potential high variance of an agent’s performance at a fixed # of timesteps (vs. considering the max performance it can reach over the whole period). As a result, I am currently reluctant to see the proposed performance measure become the standard evaluation metric on ALE, and I would appreciate some additional justification from the authors on this point.\n\n2.\tWhy not suggest to remove reward clipping in the recommendations? As mentioned in Section 6, reward clipping can prevent RL algorithms from properly playing some games, and thus in my opinion should be removed if the goal is to reach the highest score possible on all games. It seems to me that the choice of clipping the reward should be part of the algorithm (if it is not able to handle the high variety of « raw » rewards) and not of the benchmark environment, thus enabling further advancements towards algorithms that are robust to a wide range of rewards.\n\n3.\tWhy bother to keep the mean performance when, as mentioned, it is highly sensitive to outliers compared to the median?\n\n\nAdditional remarks:\n•\tI might have missed it but I do not see the link to the source code. Am I correct to assume it will be released, to help with reproducibility?\n•\tIt is not clear, when reading the paper, that the distributed version of Rainbow is actually constrained to mimic a single agent sequential algorithm in the experiments. I would suggest to remove mentions of the distributed version in the main text to avoid confusion, and mention it only in the Appendix section where it is used.\n•\tThe « infinite reward loop » point at the end of Section 6 does not seem relevant in the list of reasons why Deep RL algorithms are far from the best human performance, since with infinite playtime and an infinite reward loop, the algorithm should be guaranteed to outperform humans.\n•\tI would have appreciated an evaluation of Rainbow-IQN with the current most commonly used evaluation schemes (e.g. the one used in the original Rainbow paper), for comparison purpose (even if such an evaluation has flaws, it is often the only performance measure available for existing deep RL algorithms)\n\nReview update: thank you for the response, I am currently keeping my \"Weak accept\" rating because I agree it is important to highlight and (try to) fix the problems with the way algorithms are currently evaluated on ALE, in spite of the limited technical contributions (and the fact that I remained unconvinced regarding #1)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This paper revisits the way RL algorithms are typically evaluated on the ALE benchmark, advocating for several key changes that contribute to more robust and reliable comparisons between algorithms. It also brings the following additional contributions: (1) a new measure of comparison to human performance based on actual human world records (which shows that RL algorithms are not as « super-human » as is generally believed), and (2) an evaluation (based on the proposed guidelines) of Rainbow as well as a Rainbow-IQN variant (replacing the C51 component of Rainbow with Implicit Quantile Networks), showing that the latter brings a significant improvement upon the original Rainbow algorithm.\n\nOverall I am leaning towards acceptance as I believe that such papers encouraging better benchmarking practice on Atari are definitely needed. Even if the technical contribution is limited, this paper could have a positive impact on the field by providing a clearer picture of the current state of deep RL algorithms on Atari (assuming that other researchers start following these recommendations -- and if that is not the case at least it will highlight issues with the way evaluation is currently done).\n\nI do have a few concerns / questions though:\n\n1.\tI am not convinced by the recommendation to use performance during training for evaluation purpose. In Machado et al. (2018) it is argued that « this better aligns the performance metric with the goal of continual learning », but most deep RL algorithms trained on Atari games have not been intended to be used in a continual learning setting. It definitely has the advantage of being simple, but it seems to me that it can cause some issues, like making it difficult to compare different exploration techniques for off-policy learning (the exploration may cause poor behavior during training even if it helps the agent learn a better greedy policy), and more generally not being representative of the common practical use case where the goal is to obtain the best agent possible to use in production (with no further learning). Finally, it could make results even harder to reproduce due to the potential high variance of an agent’s performance at a fixed # of timesteps (vs. considering the max performance it can reach over the whole period). As a result, I am currently reluctant to see the proposed performance measure become the standard evaluation metric on ALE, and I would appreciate some additional justification from the authors on this point.\n\n2.\tWhy not suggest to remove reward clipping in the recommendations? As mentioned in Section 6, reward clipping can prevent RL algorithms from properly playing some games, and thus in my opinion should be removed if the goal is to reach the highest score possible on all games. It seems to me that the choice of clipping the reward should be part of the algorithm (if it is not able to handle the high variety of « raw » rewards) and not of the benchmark environment, thus enabling further advancements towards algorithms that are robust to a wide range of rewards.\n\n3.\tWhy bother to keep the mean performance when, as mentioned, it is highly sensitive to outliers compared to the median?\n\n\nAdditional remarks:\n•\tI might have missed it but I do not see the link to the source code. Am I correct to assume it will be released, to help with reproducibility?\n•\tIt is not clear, when reading the paper, that the distributed version of Rainbow is actually constrained to mimic a single agent sequential algorithm in the experiments. I would suggest to remove mentions of the distributed version in the main text to avoid confusion, and mention it only in the Appendix section where it is used.\n•\tThe « infinite reward loop » point at the end of Section 6 does not seem relevant in the list of reasons why Deep RL algorithms are far from the best human performance, since with infinite playtime and an infinite reward loop, the algorithm should be guaranteed to outperform humans.\n•\tI would have appreciated an evaluation of Rainbow-IQN with the current most commonly used evaluation schemes (e.g. the one used in the original Rainbow paper), for comparison purpose (even if such an evaluation has flaws, it is often the only performance measure available for existing deep RL algorithms)\n\nReview update: thank you for the response, I am currently keeping my \"Weak accept\" rating because I agree it is important to highlight and (try to) fix the problems with the way algorithms are currently evaluated on ALE, in spite of the limited technical contributions (and the fact that I remained unconvinced regarding #1)", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571851315120}, {"id": "HJxg5_Satr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper601/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an extension to the work of Machado et al. (2018) for standardizing training and evaluation procedures in the Arcade Learning Environment (ALE). It then introduces a collection of human world records for each Atari game to refute previous claims of superhuman performance, as well as recommend comparisons against these records. They proceed to evaluate Rainbow under their proposed evaluation procedures, as well as introduce a new algorithm, Rainbow-IQN, with similar evaluations made based on their proposal.\n\nI'm proposing a weak rejection as I feel some of the arguments made in the paper aren't very strong. In particular, I'd like the authors to comment on the following:\n\n1) The key difference between their evaluation benchmark and the recommendations in Machado et al. (2018) are that episodes should not have a time limit. The justification for this is that many algorithms might achieve practically optimal performance within this time limit, and so one wouldn't be able to compare algorithms on certain games within significance. They further emphasize that human high scores were achieved without limiting to 30 minutes of play. That said, several algorithms performing similarly within said limit can be instead interpreted as shifting emphasis toward comparing performance on the harder games. As the paper acknowledged, removing the maximum episode length ended up introducing more issues, such as the emulator never ending an episode (due to a supposed bug), as well as increasing the likelihood of the score overflowing. The paper suggested a trick of limiting how long an agent can go without receiving a reward, but it's unclear (1) if needing this fix is worth the proposed change, and (2) if the fix introduces additional game-specific nuances in evaluation; e.g., are there any situations where this can be detrimental to properly evaluating performance, or introduce biases based on a game's reward distribution?\n\n2) The paper gathered a list of human world records for the Atari games in the ALE. In my opinion, this is very valuable for the literature in terms of addressing prior work misrepresenting the competency of an algorithm relative to what humans are capable of; a professional game tester is supposed to be representative of the average game player, who is typically tasked with optimizing fun, whereas speedrunners and scorerunners of games are tasked with optimizing a comparable objective to an RL agent. Beyond this though, I think an alternative conclusion would be to use this information in support of not comparing results to human scores, and to focus on comparisons between algorithms. A considerable number of the human world records have reached the maximum allowable score, over drastically variable gameplay times to achieve these scores, that it might still not be that fair a comparison. Have the authors considered this possibility?\n\n3) Were any other algorithms evaluated on this benchmark, beyond Rainbow? While there are computational considerations, it seems odd for a benchmarking-focused paper to only evaluate one standard algorithm and slight modification of it.\n\nSuggestions\n\n1) The introduction of Rainbow-IQN in this paper feels a little random and out of place given the context created by the rest of the paper's contributions- I feel it might be more appropriate for a benchmarking paper to focus on a representative set of \"standard\" or relatively simple/trivial algorithms (Like Machado et al. (2018) did) to give a frame of reference for comparing novel ones.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper proposes an extension to the work of Machado et al. (2018) for standardizing training and evaluation procedures in the Arcade Learning Environment (ALE). It then introduces a collection of human world records for each Atari game to refute previous claims of superhuman performance, as well as recommend comparisons against these records. They proceed to evaluate Rainbow under their proposed evaluation procedures, as well as introduce a new algorithm, Rainbow-IQN, with similar evaluations made based on their proposal.\n\nI'm proposing a weak rejection as I feel some of the arguments made in the paper aren't very strong. In particular, I'd like the authors to comment on the following:\n\n1) The key difference between their evaluation benchmark and the recommendations in Machado et al. (2018) are that episodes should not have a time limit. The justification for this is that many algorithms might achieve practically optimal performance within this time limit, and so one wouldn't be able to compare algorithms on certain games within significance. They further emphasize that human high scores were achieved without limiting to 30 minutes of play. That said, several algorithms performing similarly within said limit can be instead interpreted as shifting emphasis toward comparing performance on the harder games. As the paper acknowledged, removing the maximum episode length ended up introducing more issues, such as the emulator never ending an episode (due to a supposed bug), as well as increasing the likelihood of the score overflowing. The paper suggested a trick of limiting how long an agent can go without receiving a reward, but it's unclear (1) if needing this fix is worth the proposed change, and (2) if the fix introduces additional game-specific nuances in evaluation; e.g., are there any situations where this can be detrimental to properly evaluating performance, or introduce biases based on a game's reward distribution?\n\n2) The paper gathered a list of human world records for the Atari games in the ALE. In my opinion, this is very valuable for the literature in terms of addressing prior work misrepresenting the competency of an algorithm relative to what humans are capable of; a professional game tester is supposed to be representative of the average game player, who is typically tasked with optimizing fun, whereas speedrunners and scorerunners of games are tasked with optimizing a comparable objective to an RL agent. Beyond this though, I think an alternative conclusion would be to use this information in support of not comparing results to human scores, and to focus on comparisons between algorithms. A considerable number of the human world records have reached the maximum allowable score, over drastically variable gameplay times to achieve these scores, that it might still not be that fair a comparison. Have the authors considered this possibility?\n\n3) Were any other algorithms evaluated on this benchmark, beyond Rainbow? While there are computational considerations, it seems odd for a benchmarking-focused paper to only evaluate one standard algorithm and slight modification of it.\n\nSuggestions\n\n1) The introduction of Rainbow-IQN in this paper feels a little random and out of place given the context created by the rest of the paper's contributions- I feel it might be more appropriate for a benchmarking paper to focus on a representative set of \"standard\" or relatively simple/trivial algorithms (Like Machado et al. (2018) did) to give a frame of reference for comparing novel ones."}, "tcdate": 1571801223757}, {"id": "rylUiKKiYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper601/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper proposes a new way to benchmark DRL algorithms using the Atari environment which is twofold, one part is a set of emulator recommendations, the other part is what quantity we should consider as a \"human reference\". The paper also compares Rainbow and Rainbow-IQN, where the IQN improvement matches the proposed human normalized score improvment.\n\nI'm not quite sure how to rate this paper, I have put weak-reject for now, as I don't strongly disagree with anything in the paper, but at the same time:\n- the difference to Machados et al. is marginal, but is a bit surprising\n- the Rainbow-IQN improvement is too incremental to be considered a significant contribution\n- there are some interesting remarks on why Atari is _not_ necessarily a good environment, e.g. most of Section 6, but this clashes with the paper's premise that we should be using Atari.\n\nIn a way, this paper reads like an interesting technical review of Atari, but I don't think it provides enough new knowledge to be a conference paper.\n\nDetailed comments:\n- I find it a bit weird that the many weaknesses of Atari as a platform are presented at DRL being bad at Atari. The line between environment design and algorithm design can be blurry, but in Atari's case, the weird peculiarities of each game are known to make it an inconvenient benchmark.\n- In the same vein, why is Atari+SABER better than other RL environments? This is rather crucial. We should only work on improving a benchmark if it is a useful benchmark, yet, we have many clues that Atari is not.\n- The link to TwinGalaxies should be a proper reference with the time of visit, especially if humans break new records in the future.\n- Why only compare Rainbow and a variant of Rainbow? I understand compute resources being a limitation, but at the same time, the reasoning behind having standardized testing is to be able to compare a wide variety of algorithms. This paper would be much stronger if it focused on a few representative games (e.g. one reflex game, one hard exploration game, etc.) and tested these games with a bunch of DRL algorithms. That ranking might reveal something very interesting.\n- The paper is easy to read, but there are a few grammar mistakes here and there.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "\nThis paper proposes a new way to benchmark DRL algorithms using the Atari environment which is twofold, one part is a set of emulator recommendations, the other part is what quantity we should consider as a \"human reference\". The paper also compares Rainbow and Rainbow-IQN, where the IQN improvement matches the proposed human normalized score improvment.\n\nI'm not quite sure how to rate this paper, I have put weak-reject for now, as I don't strongly disagree with anything in the paper, but at the same time:\n- the difference to Machados et al. is marginal, but is a bit surprising\n- the Rainbow-IQN improvement is too incremental to be considered a significant contribution\n- there are some interesting remarks on why Atari is _not_ necessarily a good environment, e.g. most of Section 6, but this clashes with the paper's premise that we should be using Atari.\n\nIn a way, this paper reads like an interesting technical review of Atari, but I don't think it provides enough new knowledge to be a conference paper.\n\nDetailed comments:\n- I find it a bit weird that the many weaknesses of Atari as a platform are presented at DRL being bad at Atari. The line between environment design and algorithm design can be blurry, but in Atari's case, the weird peculiarities of each game are known to make it an inconvenient benchmark.\n- In the same vein, why is Atari+SABER better than other RL environments? This is rather crucial. We should only work on improving a benchmark if it is a useful benchmark, yet, we have many clues that Atari is not.\n- The link to TwinGalaxies should be a proper reference with the time of visit, especially if humans break new records in the future.\n- Why only compare Rainbow and a variant of Rainbow? I understand compute resources being a limitation, but at the same time, the reasoning behind having standardized testing is to be able to compare a wide variety of algorithms. This paper would be much stronger if it focused on a few representative games (e.g. one reflex game, one hard exploration game, etc.) and tested these games with a bunch of DRL algorithms. That ranking might reveal something very interesting.\n- The paper is easy to read, but there are a few grammar mistakes here and there."}, "tcdate": 1571686814162}], "openreview_url": "https://openreview.net/forum?id=r1x_DaVKwH", "arxiv_id": "1908.04683", "paper_pdf": "papers/r1x_DaVKwH.pdf", "paper_pdf_sha256": "97bdef21666bc1ff43ec061d8272c8dc2122a19cfa1357789c228a96458f5d61", "paper_pdf_bytes": 685654, "paper_pdf_source": "openreview", "code_url": "https://github.com/valeoai/rainbow-iqn-apex", "code_repository": "valeoai/rainbow-iqn-apex", "code_commit": "7e71986703e112eecb943104fb2a9403c7cdb88a", "code_archive": "repos/r1x_DaVKwH.zip", "code_archive_sha256": "313b3551d5a10273acda2bb1779361df8a147600a3ae6c5d50dd9d263229bb06", "code_archive_bytes": 69382, "code_file_count": 17, "code_extensions": {".py": 15, ".sh": 2}, "github_disk_usage_kb": 149, "github_languages": {"Python": 114228, "Shell": 2656, "Dockerfile": 856}, "github_archived": false, "github_pushed_at": "2019-12-17T10:13:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/is-deep-reinforcement-learning-really"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BygMAiRqK7", "year": 2019, "status": "rejected", "title": "Entropic GANs meet VAEs: A Statistical Approach to Compute Sample Likelihoods in GANs", "authors": ["Yogesh Balaji", "Hamed Hasani", "Rama Chellappa", "Soheil Feizi"], "authorids": ["yogesh@cs.umd.edu", "hassani@seas.upenn.edu", "rama@umiacs.umd.edu", "sfeizi@cs.umd.edu"], "authors_source": "OpenReview API", "abstract": "Building on the success of deep learning, two modern approaches to learn a probability model of the observed data are Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs). VAEs consider an explicit probability model for the data and compute a generative distribution by maximizing a variational lower-bound on the log-likelihood function. GANs, however, compute a generative model by minimizing a distance between observed and generated probability distributions without considering an explicit model for the observed data. The lack of having explicit probability models in GANs prohibits computation of sample likelihoods in their frameworks and limits their use in statistical inference problems. In this work, we show that an optimal transport GAN with the entropy regularization can be viewed as a generative model that maximizes a lower-bound on average sample likelihoods, an approach that VAEs are based on. In particular, our proof constructs an explicit probability model for GANs that can be used to compute likelihood statistics within GAN's framework. Our numerical results on several datasets demonstrate consistent trends with the proposed theory. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SJeUt8H62X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper873/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\nThe authors notice that entropy regularized optimal transport produce an upper bound of a certain model likelihood. Then, the authors claim it is possible to leverage that upper bound to come up with a measure of 'sample likelihood', the probability of a certain sample under the model.\n\nEvaluation\nThe idea is certainly interesting and novel, as it allows to bridge two distinct worlds (VAE and GANs). However, I am concerned about the message (or lack of thereof) that is conveyed in the paper. Particularly, the following two points makes me be reluctant to recommend an acceptance:\n\n1)There is no measure on the tightness of the lower bound. How can we tell if this bound isnt tight? All results are dependent on the bound being close to the true value. No comments about this are given.\n2)The sample likelihoods are dependent on a certain \"model\". Here the nomenclature is confusing because I thought GANS were a probabilistic model, but now there is an additional model regarding a function f. How these two relate? What happens if I change f? to which extent the results depend on f?\n3)related to 2): the histograms in figure 2 are interesting, but they are not conclusive that the measure that is being proposed is a 'bona fide' sample likelihood.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting connection, but lacks clarity", "review": "Summary\nThe authors notice that entropy regularized optimal transport produce an upper bound of a certain model likelihood. Then, the authors claim it is possible to leverage that upper bound to come up with a measure of 'sample likelihood', the probability of a certain sample under the model.\n\nEvaluation\nThe idea is certainly interesting and novel, as it allows to bridge two distinct worlds (VAE and GANs). However, I am concerned about the message (or lack of thereof) that is conveyed in the paper. Particularly, the following two points makes me be reluctant to recommend an acceptance:\n\n1)There is no measure on the tightness of the lower bound. How can we tell if this bound isnt tight? All results are dependent on the bound being close to the true value. No comments about this are given.\n2)The sample likelihoods are dependent on a certain \"model\". Here the nomenclature is confusing because I thought GANS were a probabilistic model, but now there is an additional model regarding a function f. How these two relate? What happens if I change f? to which extent the results depend on f?\n3)related to 2): the histograms in figure 2 are interesting, but they are not conclusive that the measure that is being proposed is a 'bona fide' sample likelihood.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541391997710}, {"id": "HygcBfTO2m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper873/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n1. The assumption made by the authors that \"generator is injective\" is problematic or even wrong, as it is well known that GAN suffers from mode collapsing problem. \n\n2. It is very confusing when the authors mentioned the negative Shannon entropy. Because the equation the authors wrote is the Shannon entropy, not the negative version.\n\n3. In the 5th paragraph in the  introduction section, the paper (Cuturi, 2013) has nothing to do with \"improve computational aspect of GAN\", maybe the authors want to cite this paper \"Learning Generative Models with Sinkhorn Divergences\".\n\n4. The authors failed to discuss their paper with \"ON THE QUANTITATIVE ANALYSIS OF DECODERBASED GENERATIVE MODELS\", which uses AIS to estimate the likelihood.\n\nSuggestion:\n1. Please use \\cdot instead of , i.e. F(\\cdot) instead of F(.)\n2. Typo: in Appendix ?? and ??, in section 4", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, but need more polishing", "review": "\n1. The assumption made by the authors that \"generator is injective\" is problematic or even wrong, as it is well known that GAN suffers from mode collapsing problem. \n\n2. It is very confusing when the authors mentioned the negative Shannon entropy. Because the equation the authors wrote is the Shannon entropy, not the negative version.\n\n3. In the 5th paragraph in the  introduction section, the paper (Cuturi, 2013) has nothing to do with \"improve computational aspect of GAN\", maybe the authors want to cite this paper \"Learning Generative Models with Sinkhorn Divergences\".\n\n4. The authors failed to discuss their paper with \"ON THE QUANTITATIVE ANALYSIS OF DECODERBASED GENERATIVE MODELS\", which uses AIS to estimate the likelihood.\n\nSuggestion:\n1. Please use \\cdot instead of , i.e. F(\\cdot) instead of F(.)\n2. Typo: in Appendix ?? and ??, in section 4", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541096002256}, {"id": "Hkleh_K_3Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper873/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The contribution of the paper is to show that WGAN with entropic regularization maximize a lower bound on the likelihood of the observed data distribution. While the WGAN formulation minimizes the Wasserstein distance of the transformed latent distribution and the empirical distribution which is already a nice measure of \"progress\", having a bound on the likelihood can be interesting.\n\nPros:\n+ I like the entropic GAN formulation and believe it is very interesting as it gives access to the joint distribution of latent and observed variables. \n+ While there are some doubtful statements, overall the paper is well written and easy to read.\n\nCons:\n- The assumption of injectivity of the generator could be problematic, as it might not be fulfilled due to mode collapse.\n- I feel the theory is not very deep. Since one has a closed form of the transportation map (Eq. 3.7), the likelihood of the data is obtained by marginalizing out the latent space. However, this assumes that the inner dual maximization problem is solved to stationarity so that Eq 3.7 holds, which is not the case in practice (5 discriminator updates).\n- Thus in Sec. 4.1 for the likelihood at various points in training it is not clear what is actually happening.\n- Sec 4.3 for unregularized GANs might be problematic. In general, the transportation plan is not a density function, so I'm not certain whether Theorem 1 / Corollary 2 still hold. Furthermore, the heuristic for \"inverting\" G^* is very crude. \n\n- There are also some minor problematic statements in the paper. While they can be easily fixed, they give me doubts:\n  * The original VAE paper is not cited in the introduction for VAEs\n  * The 2013 paper by Cuturi cited on page 2 has nothing to do with \"computational aspects of GANs\". It is about fast computation of approximate OT between two discrete prob. measures. \n  * First-order / second-order Wasserstein distance is I think a bit unusual name for W_1, W_2\n  * On pg. 4, the point of the entropy term is to make the objective strongly convex. Strict convexity has no computational benefits.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting attempt on theory of entropic GANs ", "review": "The contribution of the paper is to show that WGAN with entropic regularization maximize a lower bound on the likelihood of the observed data distribution. While the WGAN formulation minimizes the Wasserstein distance of the transformed latent distribution and the empirical distribution which is already a nice measure of \"progress\", having a bound on the likelihood can be interesting.\n\nPros:\n+ I like the entropic GAN formulation and believe it is very interesting as it gives access to the joint distribution of latent and observed variables. \n+ While there are some doubtful statements, overall the paper is well written and easy to read.\n\nCons:\n- The assumption of injectivity of the generator could be problematic, as it might not be fulfilled due to mode collapse.\n- I feel the theory is not very deep. Since one has a closed form of the transportation map (Eq. 3.7), the likelihood of the data is obtained by marginalizing out the latent space. However, this assumes that the inner dual maximization problem is solved to stationarity so that Eq 3.7 holds, which is not the case in practice (5 discriminator updates).\n- Thus in Sec. 4.1 for the likelihood at various points in training it is not clear what is actually happening.\n- Sec 4.3 for unregularized GANs might be problematic. In general, the transportation plan is not a density function, so I'm not certain whether Theorem 1 / Corollary 2 still hold. Furthermore, the heuristic for \"inverting\" G^* is very crude. \n\n- There are also some minor problematic statements in the paper. While they can be easily fixed, they give me doubts:\n  * The original VAE paper is not cited in the introduction for VAEs\n  * The 2013 paper by Cuturi cited on page 2 has nothing to do with \"computational aspects of GANs\". It is about fast computation of approximate OT between two discrete prob. measures. \n  * First-order / second-order Wasserstein distance is I think a bit unusual name for W_1, W_2\n  * On pg. 4, the point of the entropy term is to make the objective strongly convex. Strict convexity has no computational benefits.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541081256273}], "openreview_url": "https://openreview.net/forum?id=BygMAiRqK7", "arxiv_id": "1810.04147", "paper_pdf": "papers/BygMAiRqK7.pdf", "paper_pdf_sha256": "6e4c3a3c5646afaf4c11523d5357c16ed99d3e338e7e43d15a12993e406440f5", "paper_pdf_bytes": 9986289, "paper_pdf_source": "openreview", "code_url": "https://github.com/yogeshbalaji/EntropicGANs_meet_VAEs", "code_repository": "yogeshbalaji/EntropicGANs_meet_VAEs", "code_commit": "8021a601fdbc3be04d2197b5b4201f75435ae739", "code_archive": "repos/BygMAiRqK7.zip", "code_archive_sha256": "d46c78704e653c6244741822a3951f05356978fcc55604dd340e6be45a67fbbf", "code_archive_bytes": 328019, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 1413, "github_languages": {"Python": 78205}, "github_archived": false, "github_pushed_at": "2019-11-22T22:38:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/entropic-gans-meet-vaes-a-statistical"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "97kU2E3UKU", "year": 2026, "status": "rejected", "title": "LLM-First Search: Self-Guided Exploration of the Solution Space", "authors": ["Nathan Herr", "Tim Rocktäschel", "Roberta Raileanu"], "authorids": ["~Nathan_Herr1", "~Tim_Rocktäschel1", "~Roberta_Raileanu2"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) have demonstrated remarkable improvements in reasoning and planning through increased test-time compute, often by framing problem-solving as a search process. While methods like Monte Carlo Tree Search (MCTS) have proven effective in some domains, their reliance on fixed exploration hyperparameters limits their adaptability across tasks of varying difficulty, rendering them impractical or expensive in certain settings. In this paper, we propose \\textbf{LLM-First Search (LFS)}, a novel \\textit{LLM Self-Guided Search} method that removes the need for pre-defined search strategies by empowering the LLM to autonomously control the search process via self-guided exploration. Rather than relying on external heuristics or hardcoded policies, the LLM evaluates whether to pursue the current search path or explore alternative branches based on its internal scoring mechanisms. This enables more flexible and context-sensitive reasoning without requiring manual tuning or task-specific adaptation. We evaluate LFS on Countdown and Sudoku against three classic widely-used search algorithms, Tree-of-Thoughts' Breadth First Search (ToT-BFS), Best First Search (BestFS), and MCTS, each of which have been used to achieve SotA results on a range of challenging reasoning tasks. We found that LFS (1) performs better on more challenging tasks without additional tuning, (2) is more computationally efficient compared to the other methods, especially when powered by a stronger model, (3) scales better with stronger models, due to its LLM-First design, and (4) scales better with increased compute budget. Our code will become publicly available upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "6WEMjTId6O", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19954/Reviewer_bV2f"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper proposes LLM-First Search (LFS), a self-guided test-time search method where the language model itself decides when to exploit the current path or explore alternatives, scoring actions and dynamically backtracking via a priority queue and it removes hand-tuned exploration schedules and heuristics common in ToT-BFS, BestFS, and MCTS. Evaluated on Countdown and Sudoku with GPT-4o and o3-mini, LFS achieves competitive or superior WinRate, better efficiency (wins per token), and stronger scaling with harder tasks, larger models, and higher compute budgets.", "review_text": "The paper proposes LLM-First Search (LFS), a self-guided test-time search method where the language model itself decides when to exploit the current path or explore alternatives, scoring actions and dynamically backtracking via a priority queue and it removes hand-tuned exploration schedules and heuristics common in ToT-BFS, BestFS, and MCTS. Evaluated on Countdown and Sudoku with GPT-4o and o3-mini, LFS achieves competitive or superior WinRate, better efficiency (wins per token), and stronger scaling with harder tasks, larger models, and higher compute budgets.", "strengths": "1. Choosing win rate to accommodate generation stochasticity is sensible. The paper also reports Wilson 95% CIs, efficiency (wins per token), and performance profiles (AUP), which improves statistical transparency.", "weaknesses": "1. Countdown and Sudoku are rigorous but synthetic; standing alone (without math/coding or other reasoning tasks) they have relatively limited action spaces, which narrows external validity. The authors themselves note the restricted scope. However, given that it is a paper that proposes a methodology, the limitation is substantial.   \n\n\n2. Limited ablations: there is no analysis of LFS internals to understand the effectiveness of the proposed pipeline. Component-wise ablations would help isolate what drives gains.    \n\n\n3. Baselines seem incomplete: ToT-BFS baseline only appear in GPT-4o but not in o3-mini. MCTS for o3-mini only uses c=0.5 and as authors pointed out this particular choice was seen only optimal for CountDown. It might lead to unintentionally choosing weak baseline due the incompleteness of experiments. Furthermore, authors only evaluated on two models, which might be limited to validate the methodology.", "questions": "1. In Table 1/Figs., why is 6×6 Sudoku so low (e.g., with o3-mini: 0–25% WinRate across methods)? Do you suspect artifacts of the setup, and why not 9×9 to better reflect standard Sudoku difficulty?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes LLM-First Search (LFS), a self-guided test-time search method where the language model itself decides when to exploit the current path or explore alternatives, scoring actions and dynamically backtracking via a priority queue and it removes hand-tuned exploration schedules and heuristics common in ToT-BFS, BestFS, and MCTS. Evaluated on Countdown and Sudoku with GPT-4o and o3-mini, LFS achieves competitive or superior WinRate, better efficiency (wins per token), and stronger scaling with harder tasks, larger models, and higher compute budgets.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "1. Choosing win rate to accommodate generation stochasticity is sensible. The paper also reports Wilson 95% CIs, efficiency (wins per token), and performance profiles (AUP), which improves statistical transparency.", "weaknesses": "1. Countdown and Sudoku are rigorous but synthetic; standing alone (without math/coding or other reasoning tasks) they have relatively limited action spaces, which narrows external validity. The authors themselves note the restricted scope. However, given that it is a paper that proposes a methodology, the limitation is substantial.   \n\n\n2. Limited ablations: there is no analysis of LFS internals to understand the effectiveness of the proposed pipeline. Component-wise ablations would help isolate what drives gains.    \n\n\n3. Baselines seem incomplete: ToT-BFS baseline only appear in GPT-4o but not in o3-mini. MCTS for o3-mini only uses c=0.5 and as authors pointed out this particular choice was seen only optimal for CountDown. It might lead to unintentionally choosing weak baseline due the incompleteness of experiments. Furthermore, authors only evaluated on two models, which might be limited to validate the methodology.", "questions": "1. In Table 1/Figs., why is 6×6 Sudoku so low (e.g., with o3-mini: 0–25% WinRate across methods)? Do you suspect artifacts of the setup, and why not 9×9 to better reflect standard Sudoku difficulty?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761968468513}, {"id": "UR99jRf53Q", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19954/Reviewer_KDbc"], "rating": 2, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper introduces LLM-First Search (LFS), a novel method for reasoning and planning in Large Language Models (LLMs). The core premise of LFS is to replace the fixed search strategies and pre-defined hyperparameters (such as the exploration constant $C$ in MCTS) with an autonomous control mechanism managed directly by the LLM itself.\n\nThe LFS method employs the LLM in two key operations: (1) 'Evaluate', where the LLM assesses the value of all available actions from a given state, and (2) 'Explore', where the LLM dynamically decides whether to continue exploiting the current path or to \"jump\" to an alternative, high-potential path. To facilitate this, LFS maintains a priority queue of all unelected, promising actions, which serves as a mechanism for backtracking.\n\nThe authors evaluate LFS on two benchmarks, Countdown and Sudoku. The results reportedly show that LFS outperforms baseline methods, including ToT-BFS, BestFS, and MCTS, particularly on more difficult tasks, while also demonstrating superior computational efficiency.", "review_text": "This paper introduces LLM-First Search (LFS), a novel method for reasoning and planning in Large Language Models (LLMs). The core premise of LFS is to replace the fixed search strategies and pre-defined hyperparameters (such as the exploration constant $C$ in MCTS) with an autonomous control mechanism managed directly by the LLM itself.\n\nThe LFS method employs the LLM in two key operations: (1) 'Evaluate', where the LLM assesses the value of all available actions from a given state, and (2) 'Explore', where the LLM dynamically decides whether to continue exploiting the current path or to \"jump\" to an alternative, high-potential path. To facilitate this, LFS maintains a priority queue of all unelected, promising actions, which serves as a mechanism for backtracking.\n\nThe authors evaluate LFS on two benchmarks, Countdown and Sudoku. The results reportedly show that LFS outperforms baseline methods, including ToT-BFS, BestFS, and MCTS, particularly on more difficult tasks, while also demonstrating superior computational efficiency.", "strengths": "1.  Well-Motivated Problem: The paper clearly identifies a significant limitation in existing search-augmented LLMs. It highlights the impracticality and sub-optimality of relying on fixed hyperparameters (like the MCTS exploration constant $C$), which require costly re-tuning for different tasks or models. The finding that MCTS performance can degrade when using a stronger model provides a compelling motivation for a more adaptive approach.\n\n2.  Strong Empirical Performance: Within the confines of the two evaluated benchmarks, LFS demonstrates superior performance. The method achieves a higher WinRate than all baselines, with the performance gap widening on the most challenging tasks (e.G., Countdown Diff=7).\n\n3.  Impressive Scalability and Efficiency: LFS shows strong evidence of scalability. It not only performs well but shows a greater performance improvement when paired with a stronger model (O3-mini) compared to baselines. Furthermore, the method is shown to be more computationally efficient, achieving a better Area Under Profile (AUP) for efficiency and generating smaller, more focused search trees than MCTS.", "weaknesses": "1.  Concerns Regarding Novelty: The proposed method appears to have significant overlap with the Tree-of-Thoughts (ToT) framework. The core mechanism—using an 'Evaluate' prompt for scoring and an 'Explore' prompt for decision-making, coupled with a priority queue—could be interpreted as a sophisticated form of prompt engineering built upon the ToT concept, rather than a fundamentally new search paradigm. The novelty beyond this implementation is not made sufficiently clear.\n\n\n2.  Insufficient Analysis of Mechanism: The paper provides extensive data on *what* LFS achieves (i.e., higher WinRates) but lacks a deep analysis of *how* it achieves this performance. The claim that \"self-guided\" exploration works better is a high-level description of the phenomenon, not an explanation. There is no analysis of the LLM's decision-making process during the 'Explore' step. What cues does the LLM use to decide to backtrack? How do these emergent heuristics qualitatively differ from, and improve upon, the MCTS formula? Without this analysis, the paper is reporting numbers, not mechanisms.\n\n3.  Limited Generalization: The validation is confined to two highly structured \"toy examples\" (Countdown and Sudoku), where states are clearly defined and rules are deterministic. No generalization is guaranteed beyond these benchmarks. The authors themselves acknowledge in the appendix that these tasks \"lack some complexities of real-world problems\". It is unclear if this \"LLM-First\" approach would be effective, or even feasible, in more open-ended, complex reasoning tasks where state representation is ambiguous.", "questions": "4.  Clarity and Presentation: The manuscript's structure is not always well-organized, making the central arguments difficult to follow at times. In particular, Figure 1 is hard to understand.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LLM-First Search (LFS), a novel method for reasoning and planning in Large Language Models (LLMs). The core premise of LFS is to replace the fixed search strategies and pre-defined hyperparameters (such as the exploration constant $C$ in MCTS) with an autonomous control mechanism managed directly by the LLM itself.\n\nThe LFS method employs the LLM in two key operations: (1) 'Evaluate', where the LLM assesses the value of all available actions from a given state, and (2) 'Explore', where the LLM dynamically decides whether to continue exploiting the current path or to \"jump\" to an alternative, high-potential path. To facilitate this, LFS maintains a priority queue of all unelected, promising actions, which serves as a mechanism for backtracking.\n\nThe authors evaluate LFS on two benchmarks, Countdown and Sudoku. The results reportedly show that LFS outperforms baseline methods, including ToT-BFS, BestFS, and MCTS, particularly on more difficult tasks, while also demonstrating superior computational efficiency.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "1.  Well-Motivated Problem: The paper clearly identifies a significant limitation in existing search-augmented LLMs. It highlights the impracticality and sub-optimality of relying on fixed hyperparameters (like the MCTS exploration constant $C$), which require costly re-tuning for different tasks or models. The finding that MCTS performance can degrade when using a stronger model provides a compelling motivation for a more adaptive approach.\n\n2.  Strong Empirical Performance: Within the confines of the two evaluated benchmarks, LFS demonstrates superior performance. The method achieves a higher WinRate than all baselines, with the performance gap widening on the most challenging tasks (e.G., Countdown Diff=7).\n\n3.  Impressive Scalability and Efficiency: LFS shows strong evidence of scalability. It not only performs well but shows a greater performance improvement when paired with a stronger model (O3-mini) compared to baselines. Furthermore, the method is shown to be more computationally efficient, achieving a better Area Under Profile (AUP) for efficiency and generating smaller, more focused search trees than MCTS.", "weaknesses": "1.  Concerns Regarding Novelty: The proposed method appears to have significant overlap with the Tree-of-Thoughts (ToT) framework. The core mechanism—using an 'Evaluate' prompt for scoring and an 'Explore' prompt for decision-making, coupled with a priority queue—could be interpreted as a sophisticated form of prompt engineering built upon the ToT concept, rather than a fundamentally new search paradigm. The novelty beyond this implementation is not made sufficiently clear.\n\n\n2.  Insufficient Analysis of Mechanism: The paper provides extensive data on *what* LFS achieves (i.e., higher WinRates) but lacks a deep analysis of *how* it achieves this performance. The claim that \"self-guided\" exploration works better is a high-level description of the phenomenon, not an explanation. There is no analysis of the LLM's decision-making process during the 'Explore' step. What cues does the LLM use to decide to backtrack? How do these emergent heuristics qualitatively differ from, and improve upon, the MCTS formula? Without this analysis, the paper is reporting numbers, not mechanisms.\n\n3.  Limited Generalization: The validation is confined to two highly structured \"toy examples\" (Countdown and Sudoku), where states are clearly defined and rules are deterministic. No generalization is guaranteed beyond these benchmarks. The authors themselves acknowledge in the appendix that these tasks \"lack some complexities of real-world problems\". It is unclear if this \"LLM-First\" approach would be effective, or even feasible, in more open-ended, complex reasoning tasks where state representation is ambiguous.", "questions": "4.  Clarity and Presentation: The manuscript's structure is not always well-organized, making the central arguments difficult to follow at times. In particular, Figure 1 is hard to understand.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761835265749}, {"id": "c3N9mAlbUD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19954/Reviewer_FYPb"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes LLM-First Search (LFS), a test-time LLM agentic framework where the LLM explore different decision nodes in a tree-based manner similar to prior works, but prompts LLM itself to give scores to every expandable node in the tree and decide whether to explore or exploit. If the LLM decides to explore and jump from the current reasoning trace to another trace, the LLM will switch to the most promising alternative node retrieved by a priority queue; otherwise, the LLM will continue diving into the current node and update the priority queue. The proposed method outperforms several baselines on two tasks, which are Sudoku and Countdown.", "review_text": "The paper proposes LLM-First Search (LFS), a test-time LLM agentic framework where the LLM explore different decision nodes in a tree-based manner similar to prior works, but prompts LLM itself to give scores to every expandable node in the tree and decide whether to explore or exploit. If the LLM decides to explore and jump from the current reasoning trace to another trace, the LLM will switch to the most promising alternative node retrieved by a priority queue; otherwise, the LLM will continue diving into the current node and update the priority queue. The proposed method outperforms several baselines on two tasks, which are Sudoku and Countdown.", "strengths": "1. The idea is simple, intuitive and clearly conveyed: to remove handcrafted rule for exploration-exploitation and let LLM take charge of which node to expand during tree search. \n\n2. The related work seems to be rather complete, which discusses different LLM test-time reasoning framework in details and clearly stated the difference between the proposed method and prior works.\n\n3. The paper has a very detailed appendix, which greatly increases the reproducibility of the paper by providing all the prompts and implementation details. The appendix also provides detailed experiment results which clearly show the advantage of LFS.", "weaknesses": "1. As mentioned in the limitation section, the proposed method highly relies on the LLM's base ability as the exploration-exploitation tradeoff is made by the LLM itself. However, if future long-context, thinking state-of-the-art models are strong enough, they may inherently possess the ability to jump between different branches of thoughts without adopting LFS (e.g. many papers [1] mention crucial tokens such as \"wait\" or \"however\" in solving complicated math problems, which can be seen as a variant of LFS). \n\n2. Following 1, the motivation of getting rid of \"fixed handcrafted exploration\" becomes less convincing as LFS also adopts somewhat handcrafted way of expanding nodes (priority queue and the explicit, explore-exploit judgment), and MCTS-based methods also involves LLM-guided evaluation of scores as mentioned in Alg. 4. The question is: since every method contains more or less degree of \"fixed handcrafted exploration\" compared to expecting the \"wait\" or \"however\" token from LLM itself, for state-of-the-art models, where should we stop and say \"this level of exploration-exploitation strategy contains neither too much nor too little handcrafted prior?\" \n\n**Minor Weakness**\n\n1. In line 652, \"LLM-First- Search\" -> \"LLM-First-Search\"\n\n2. In line 222, \"Three-of-Thoughts\" -> \"Tree-of-Thoughts\"\n\n3. The formatting in page 27 is strange; it only has one line saying \"Exploration Decision System Instruction and User Request\".\n\n\n**References**\n\n[1] S. Wang et al. Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning, 2025.", "questions": "I have several questions: \n\n1. The authors mention that the MCTS tree is noticably wider and over-explores. This is somewhat counterintuitive, as MCTS's rollout always goes on until the leaf / terminal (see Alg. 4 line 6-9), while in LFS the LLM can jump to another branch in the middle of a rollout (see Alg. 1); i.e, MCTS's rollouts are on average deeper. Can the author explain this?  \n\n2. Following 1, the authors also mention that the result shows that MCTS with a value of $C=0.5$ (i.e. fewest exploration) works better than higher $C$. The question is: will MCTS's performance grow further if $C$ is further decreased to, e.g., 0.1 or 0.2? Is 0.5 still \"too high an exploration constant\" as suggested in line 85?\n\n3. As there are many possible future states, the priority queue could save many unexplored nodes. However, the LLM ususally gives a discrete scale (e.g. 0.05, 0.1, 0.15, ..., 1) instead of a continuous value between 0 and 1. Does LFS often encounter nodes with the same weights in the priority queue? If so, how does LFS deal with this?", "flag_for_ethics_review": ["Yes, Research integrity issues (e.g., plagiarism, dual submission)"], "all_content": {"summary": "The paper proposes LLM-First Search (LFS), a test-time LLM agentic framework where the LLM explore different decision nodes in a tree-based manner similar to prior works, but prompts LLM itself to give scores to every expandable node in the tree and decide whether to explore or exploit. If the LLM decides to explore and jump from the current reasoning trace to another trace, the LLM will switch to the most promising alternative node retrieved by a priority queue; otherwise, the LLM will continue diving into the current node and update the priority queue. The proposed method outperforms several baselines on two tasks, which are Sudoku and Countdown.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The idea is simple, intuitive and clearly conveyed: to remove handcrafted rule for exploration-exploitation and let LLM take charge of which node to expand during tree search. \n\n2. The related work seems to be rather complete, which discusses different LLM test-time reasoning framework in details and clearly stated the difference between the proposed method and prior works.\n\n3. The paper has a very detailed appendix, which greatly increases the reproducibility of the paper by providing all the prompts and implementation details. The appendix also provides detailed experiment results which clearly show the advantage of LFS.", "weaknesses": "1. As mentioned in the limitation section, the proposed method highly relies on the LLM's base ability as the exploration-exploitation tradeoff is made by the LLM itself. However, if future long-context, thinking state-of-the-art models are strong enough, they may inherently possess the ability to jump between different branches of thoughts without adopting LFS (e.g. many papers [1] mention crucial tokens such as \"wait\" or \"however\" in solving complicated math problems, which can be seen as a variant of LFS). \n\n2. Following 1, the motivation of getting rid of \"fixed handcrafted exploration\" becomes less convincing as LFS also adopts somewhat handcrafted way of expanding nodes (priority queue and the explicit, explore-exploit judgment), and MCTS-based methods also involves LLM-guided evaluation of scores as mentioned in Alg. 4. The question is: since every method contains more or less degree of \"fixed handcrafted exploration\" compared to expecting the \"wait\" or \"however\" token from LLM itself, for state-of-the-art models, where should we stop and say \"this level of exploration-exploitation strategy contains neither too much nor too little handcrafted prior?\" \n\n**Minor Weakness**\n\n1. In line 652, \"LLM-First- Search\" -> \"LLM-First-Search\"\n\n2. In line 222, \"Three-of-Thoughts\" -> \"Tree-of-Thoughts\"\n\n3. The formatting in page 27 is strange; it only has one line saying \"Exploration Decision System Instruction and User Request\".\n\n\n**References**\n\n[1] S. Wang et al. Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning, 2025.", "questions": "I have several questions: \n\n1. The authors mention that the MCTS tree is noticably wider and over-explores. This is somewhat counterintuitive, as MCTS's rollout always goes on until the leaf / terminal (see Alg. 4 line 6-9), while in LFS the LLM can jump to another branch in the middle of a rollout (see Alg. 1); i.e, MCTS's rollouts are on average deeper. Can the author explain this?  \n\n2. Following 1, the authors also mention that the result shows that MCTS with a value of $C=0.5$ (i.e. fewest exploration) works better than higher $C$. The question is: will MCTS's performance grow further if $C$ is further decreased to, e.g., 0.1 or 0.2? Is 0.5 still \"too high an exploration constant\" as suggested in line 85?\n\n3. As there are many possible future states, the priority queue could save many unexplored nodes. However, the LLM ususally gives a discrete scale (e.g. 0.05, 0.1, 0.15, ..., 1) instead of a continuous value between 0 and 1. Does LFS often encounter nodes with the same weights in the priority queue? If so, how does LFS deal with this?", "flag_for_ethics_review": ["Yes, Research integrity issues (e.g., plagiarism, dual submission)"], "details_of_ethics_concerns": "The paper violates the double-blind policy. In page 10 of the paper, the authors include acknowledgement, which says: \"This work was supported by the UK Engineering and Physical Sciences Research Council (EPSRC) under grant number EP/S021566/1\", which gives away the identity of the authors.", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761350322545}], "openreview_url": "https://openreview.net/forum?id=97kU2E3UKU", "arxiv_id": "2506.05213", "paper_pdf": "papers/97kU2E3UKU.pdf", "paper_pdf_sha256": "fcf6ee94339c2ef0a81bb64bffdbda4b4d2740acc77a9691f54fab46a2662724", "paper_pdf_bytes": 1215370, "paper_pdf_source": "openreview", "code_url": "https://github.com/NathanHerr/LLM-First-Search", "code_repository": "NathanHerr/LLM-First-Search", "code_commit": "3025bdaa3add6f41388c1d5a6d354522489d312e", "code_archive": "repos/97kU2E3UKU.zip", "code_archive_sha256": "089f4ed6ec86e9455027f34e33e46a7b3f9e59af1fd230b84157b57ebb1ba822", "code_archive_bytes": 533322, "code_file_count": 21, "code_extensions": {".py": 18, ".sh": 3}, "github_disk_usage_kb": 616, "github_languages": {"Python": 337568, "Shell": 10071}, "github_archived": false, "github_pushed_at": "2025-06-09T15:11:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/llm-first-search-self-guided-exploration-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hxm0hOxph2", "year": 2025, "status": "rejected", "title": "On Provable Length and Compositional Generalization", "authors": ["Kartik Ahuja", "Amin Mansouri"], "authorids": ["~Kartik_Ahuja1", "~Amin_Mansouri1"], "authors_source": "OpenReview API", "abstract": "Out-of-distribution generalization capabilities of sequence-to-sequence models can be studied from the lens of two crucial forms of generalization: length generalization --  the ability to generalize to longer sequences than ones seen during training, and compositional generalization: the ability to generalize to token combinations not seen during training. In this work, we provide first provable guarantees on length and compositional generalization for common sequence-to-sequence models -- deep sets, transformers, state space models, and recurrent neural nets -- trained to minimize the prediction error. Taking a first principles perspective, we study the realizable case, i.e., the labeling function is realizable on the architecture.  We show that simple limited capacity versions of these different architectures achieve  both length and compositional generalization. Across different architectures, we also find that a linear relationship between the learned representation and the representation in the labeling function is necessary for length and compositional generalization.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "97oXncXANO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5063/Reviewer_4sDj"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper addresses two key challenges in out-of-distribution generalization for seq-to-seq models: length generalization and compositional generalization. While previous works have primarily focused on empirical approaches to these challenges, this paper aims to provide a theoretical, provable guarantee for achieving both types of generalization. For various architectures (deep sets, Transformers, SSMs, RNNs), the authors provide conditions on models for achieving length and compositional generalization, under the expected risk minimization setup and the realizable assumption. The paper validates their theoretical findings by presenting the experimental results on toy examples.", "review_text": "This paper addresses two key challenges in out-of-distribution generalization for seq-to-seq models: length generalization and compositional generalization. While previous works have primarily focused on empirical approaches to these challenges, this paper aims to provide a theoretical, provable guarantee for achieving both types of generalization. For various architectures (deep sets, Transformers, SSMs, RNNs), the authors provide conditions on models for achieving length and compositional generalization, under the expected risk minimization setup and the realizable assumption. The paper validates their theoretical findings by presenting the experimental results on toy examples.", "strengths": "- The paper addresses important challenges in seq-to-seq generalization, length generalization and compositional generalization, across multiple architectures.\n- The paper is well-structured and easy to follow.\n- This paper is the first work in the literature that presents provable guarantees for length and compositional generalization for seq-to-seq models. Also, the theoretical framework introduced in the paper is a pioneering contribution to the ML community.\n- I briefly read the proof, and the proof seems correct and solid.", "weaknesses": "1. The models investigated in the paper are highly simplified. For example, the authors describe the decoder-only Transformer model as $w(\\sum_{j=1}^i \\frac{1}{i} \\psi(x_i, x_j))$, where $w$ is a single-layer perception with a bijective activation. This model can not capture softmax operation (as explained in the paper), and apply only to single-layer Transformers. I understand that models with arbitrary capacity (no constraints on $w, \\psi$ ) will not exhibit any generalization (as explained in the paper), but the current restrictions are too strong. \n- Continuing from the first point, although the theoretical framework is novel, the derived results are not particularly surprising. Given that $w$ is a bijective single-layer perceptron, one can immediately derive $A\\sum_{j\\le T} \\psi(x_j) = B \\sum_{j \\le T} \\phi(x_j)$ (where the right-hand side corresponds to the realizable solution). While the paper presents rigorous proofs involving mathematical analysis techniques to establish the length and compositional generalization from this equation, the results feel somewhat expected, and I think this predictability diminishes the impact and novelty of the findings.\n2. The theorems are built under the expected risk minimization setup, which deviates from the practical scenario where we can only access a finite number of samples.\n3. While the paper validates their findings by providing experimental results, they are conducted on simplified models. Including experiments on more realistic scenarios would significantly strengthen the paper’s applicability.", "questions": "1. Would it be possible to incorporate the optimization process into the theoretical framework?\n2. The key point of the proof relies on showing that the Jacobians of the realizable solution f and an arbitrary expected risk minimization solution g are identical. On a related note, it seems that the approach in [1] also involves the use of Jacobians. Could you briefly explain the main differences between your proof and that of [1]? I ask this as I am not familiar with theoretical works on compositional generalization.\n\n[1] Wiedemer, Thaddäus, et al. \"Compositional generalization from first principles.\" *Advances in Neural Information Processing Systems* 36 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses two key challenges in out-of-distribution generalization for seq-to-seq models: length generalization and compositional generalization. While previous works have primarily focused on empirical approaches to these challenges, this paper aims to provide a theoretical, provable guarantee for achieving both types of generalization. For various architectures (deep sets, Transformers, SSMs, RNNs), the authors provide conditions on models for achieving length and compositional generalization, under the expected risk minimization setup and the realizable assumption. The paper validates their theoretical findings by presenting the experimental results on toy examples.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper addresses important challenges in seq-to-seq generalization, length generalization and compositional generalization, across multiple architectures.\n- The paper is well-structured and easy to follow.\n- This paper is the first work in the literature that presents provable guarantees for length and compositional generalization for seq-to-seq models. Also, the theoretical framework introduced in the paper is a pioneering contribution to the ML community.\n- I briefly read the proof, and the proof seems correct and solid.", "weaknesses": "1. The models investigated in the paper are highly simplified. For example, the authors describe the decoder-only Transformer model as $w(\\sum_{j=1}^i \\frac{1}{i} \\psi(x_i, x_j))$, where $w$ is a single-layer perception with a bijective activation. This model can not capture softmax operation (as explained in the paper), and apply only to single-layer Transformers. I understand that models with arbitrary capacity (no constraints on $w, \\psi$ ) will not exhibit any generalization (as explained in the paper), but the current restrictions are too strong. \n- Continuing from the first point, although the theoretical framework is novel, the derived results are not particularly surprising. Given that $w$ is a bijective single-layer perceptron, one can immediately derive $A\\sum_{j\\le T} \\psi(x_j) = B \\sum_{j \\le T} \\phi(x_j)$ (where the right-hand side corresponds to the realizable solution). While the paper presents rigorous proofs involving mathematical analysis techniques to establish the length and compositional generalization from this equation, the results feel somewhat expected, and I think this predictability diminishes the impact and novelty of the findings.\n2. The theorems are built under the expected risk minimization setup, which deviates from the practical scenario where we can only access a finite number of samples.\n3. While the paper validates their findings by providing experimental results, they are conducted on simplified models. Including experiments on more realistic scenarios would significantly strengthen the paper’s applicability.", "questions": "1. Would it be possible to incorporate the optimization process into the theoretical framework?\n2. The key point of the proof relies on showing that the Jacobians of the realizable solution f and an arbitrary expected risk minimization solution g are identical. On a related note, it seems that the approach in [1] also involves the use of Jacobians. Could you briefly explain the main differences between your proof and that of [1]? I ask this as I am not familiar with theoretical works on compositional generalization.\n\n[1] Wiedemer, Thaddäus, et al. \"Compositional generalization from first principles.\" *Advances in Neural Information Processing Systems* 36 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730699923783}, {"id": "bYuRNc5zi3", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5063/Reviewer_JxtZ"], "rating": 5, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper analyzes length generalization and compositional generalization in deep sets, transformers, SSMs, and RNNs, proving that these models achieve both types of generalization under certain assumptions. Additionally, the paper demonstrates that the representations learned by each model in these settings are linearly related to the true labeling function, indicating linear identification. Experimental results further confirm the occurrence of linear identification in real models.", "review_text": "This paper analyzes length generalization and compositional generalization in deep sets, transformers, SSMs, and RNNs, proving that these models achieve both types of generalization under certain assumptions. Additionally, the paper demonstrates that the representations learned by each model in these settings are linearly related to the true labeling function, indicating linear identification. Experimental results further confirm the occurrence of linear identification in real models.", "strengths": "- This paper provides a clear definition of length generalization and compositional generalization in sequence-to-sequence models and demonstrates valuable guidance on how to analyze these properties. This serves as a strong foundation for future work in this area.\n- The figures effectively illustrate the concepts of length generalization and compositional generalization, and the rationale behind focusing on simple limited capacity models is explained in a convincing manner.\n- The paper not only examines a range of models, but also extends its theorems to cover various scenarios, such as $C^1$-diffeomorphisms and discrete tokens.", "weaknesses": "- **Strong assumption in hypothesis selection**: As acknowledged by the authors in the discussion and limitations section, the assumption, stated as equation 1, that a hypothesis is chosen to minimize expected risk may be a bit strong.\n- **Repetitive proofs**: Some sections of the proofs appear to rely heavily on copy-pasting, leading to redundancy and potential errors. For example:\n  + On page 28, line 1483, the reference to “equation 31 of equation 30” seems incorrect and should likely be “equation 67 of equation 66.”\n  + At the beginning of Appendix C.3.1, “Assumption 2” appears, which may be intended to refer to “Assumption 15.”\n  \n  If possible, reducing redundancy by grouping similar arguments into a single lemma could help minimize errors and improve readability.\n- **Extension to discrete tokens**: While the extension of the theorems to discrete tokens is particularly intriguing, some aspects of the proof raise questions:\n  + This overlaps slightly with the second point. The phrase “regular closedness of the support” appears in Theorems 6,9,11, and 12, but for discrete tokens, the support is not regular closed, and the “equality almost everywhere -> equality everywhere” arguments may not be necessary after all.\n  + The reasoning from equation 30 to 33 is somewhat unclear; could you refine this explanation a bit? Is it not possible to derive equation 33 directly by setting $x_1 = x_2$ in equation 30?\n- **Presentation issues**:\n  + In the statements of Theorems 3 and 4, Assumptions 12 and 15 are referenced before being defined. It would be helpful to include these assumptions in the main text, or alternatively, to provide informal statements of Theorems 3 and 4 in the main text and formalize them in the Appendix.\n  + In equation 66, $\\psi$ and $\\phi$ are swapped.\n  + In the second line of equation 108, the left-hand side would be more readable if written as $\\left.\\frac{\\partial^p \\sigma(s)}{\\partial s^p}\\right|_{s=v^\\top y}$.\n  + The formula on page 37, line 1987, should have line breaks before and after it, with an added explanation to improve readability.", "questions": "- This question applies to the proofs of nearly all the theorems; as an example, let’s consider the case of deep sets: if we are considering a hypothesis that minimizes the expected risk $R(h;T) = \\sum_{t=1}^T \\mathbb{E}\\left[\\ell(h(x_{\\leq t},y_t)\\right]$, why is it not possible to directly derive $A\\psi(x) = B\\phi(x)$ for all $x \\in [0,1]^n$ from the equality $\\omega\\left(\\psi(x)\\right) = \\rho\\left(\\phi(x)\\right)$ for all $x \\in [0,1]^n$ in the case of $t=1$?\n- How realistic is Assumption 18? It would be helpful to see concrete examples that illustrate this assumption.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper analyzes length generalization and compositional generalization in deep sets, transformers, SSMs, and RNNs, proving that these models achieve both types of generalization under certain assumptions. Additionally, the paper demonstrates that the representations learned by each model in these settings are linearly related to the true labeling function, indicating linear identification. Experimental results further confirm the occurrence of linear identification in real models.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- This paper provides a clear definition of length generalization and compositional generalization in sequence-to-sequence models and demonstrates valuable guidance on how to analyze these properties. This serves as a strong foundation for future work in this area.\n- The figures effectively illustrate the concepts of length generalization and compositional generalization, and the rationale behind focusing on simple limited capacity models is explained in a convincing manner.\n- The paper not only examines a range of models, but also extends its theorems to cover various scenarios, such as $C^1$-diffeomorphisms and discrete tokens.", "weaknesses": "- **Strong assumption in hypothesis selection**: As acknowledged by the authors in the discussion and limitations section, the assumption, stated as equation 1, that a hypothesis is chosen to minimize expected risk may be a bit strong.\n- **Repetitive proofs**: Some sections of the proofs appear to rely heavily on copy-pasting, leading to redundancy and potential errors. For example:\n  + On page 28, line 1483, the reference to “equation 31 of equation 30” seems incorrect and should likely be “equation 67 of equation 66.”\n  + At the beginning of Appendix C.3.1, “Assumption 2” appears, which may be intended to refer to “Assumption 15.”\n  \n  If possible, reducing redundancy by grouping similar arguments into a single lemma could help minimize errors and improve readability.\n- **Extension to discrete tokens**: While the extension of the theorems to discrete tokens is particularly intriguing, some aspects of the proof raise questions:\n  + This overlaps slightly with the second point. The phrase “regular closedness of the support” appears in Theorems 6,9,11, and 12, but for discrete tokens, the support is not regular closed, and the “equality almost everywhere -> equality everywhere” arguments may not be necessary after all.\n  + The reasoning from equation 30 to 33 is somewhat unclear; could you refine this explanation a bit? Is it not possible to derive equation 33 directly by setting $x_1 = x_2$ in equation 30?\n- **Presentation issues**:\n  + In the statements of Theorems 3 and 4, Assumptions 12 and 15 are referenced before being defined. It would be helpful to include these assumptions in the main text, or alternatively, to provide informal statements of Theorems 3 and 4 in the main text and formalize them in the Appendix.\n  + In equation 66, $\\psi$ and $\\phi$ are swapped.\n  + In the second line of equation 108, the left-hand side would be more readable if written as $\\left.\\frac{\\partial^p \\sigma(s)}{\\partial s^p}\\right|_{s=v^\\top y}$.\n  + The formula on page 37, line 1987, should have line breaks before and after it, with an added explanation to improve readability.", "questions": "- This question applies to the proofs of nearly all the theorems; as an example, let’s consider the case of deep sets: if we are considering a hypothesis that minimizes the expected risk $R(h;T) = \\sum_{t=1}^T \\mathbb{E}\\left[\\ell(h(x_{\\leq t},y_t)\\right]$, why is it not possible to directly derive $A\\psi(x) = B\\phi(x)$ for all $x \\in [0,1]^n$ from the equality $\\omega\\left(\\psi(x)\\right) = \\rho\\left(\\phi(x)\\right)$ for all $x \\in [0,1]^n$ in the case of $t=1$?\n- How realistic is Assumption 18? It would be helpful to see concrete examples that illustrate this assumption.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730571546285}, {"id": "TCor8O93e9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5063/Reviewer_n34f"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper considers the problems of length/compositional generalization from a functional expressivity perspective. It considers a variety of similar architectures under assumptions of realizability, some intricate assumptions about the support of continuous tokens, and strongly limited capacity (limiting the interaction between tokens). In this setting it considers the ERM solutions and shows that they must recover the ground truth features up to a linear transformation which allows them to generalize. There are negative results presented for all of the architectures considered if the capacity is increased.", "review_text": "This paper considers the problems of length/compositional generalization from a functional expressivity perspective. It considers a variety of similar architectures under assumptions of realizability, some intricate assumptions about the support of continuous tokens, and strongly limited capacity (limiting the interaction between tokens). In this setting it considers the ERM solutions and shows that they must recover the ground truth features up to a linear transformation which allows them to generalize. There are negative results presented for all of the architectures considered if the capacity is increased.", "strengths": "1. The paper proposes a (to my knowledge) novel method to study length/compositional generalization. \n\n2. The paper does a good job of extending results to a family of related toy models of different types of architectures. This generality is a strength of the approach.\n\n3. The paper does a good job of presenting both negative and positive results. In particular the negative results actually seem to show that once the restrictive limits on the capacity in the architectures are weakened, then generalization can easily fail.", "weaknesses": "1. The perspective of the paper on the prospects of length/compositional generalization seems overly optimistic. In particular, in practice achieving length/compositional has proven very difficult. The papers cited on length generalization struggle to even achieve 2x length generalization despite using all sorts of unnatural tricks. This paper seems to approach the problem as \"length generalization happens and we want to explain it\", where the real situation appears to be \"length generalization is very difficult to achieve if even possible generally\". Even in the context of the paper, perhaps the more realistic results are the negative results for each architecture showing that if you allow for (still restrictive) higher capacity, the architectures are able to represent solutions that achieve zero in-distribution error but fail to generalize.\n\n2. It is not clear why we would expect the simplified models presented in the paper to bear any relation to things we may see in practice. Of course in theory we want to study simplified models, but the ultimate goal is for those models to bring insights to practical settings. The negative results indicate that as soon as we expand the very tight capacity restrictions, things can Simplified architectures seem to have no chance of applying to real settings.\n\n3. There is no consideration for statistical issues. Essentially the ERM approach is assuming infinite data. This is combined with the strong assumptions of full support over continuous variables to pin down the learned function to be similar to the realizable ground-truth solution. It is not clear if the arguments presented can apply at all to the case of finite datasets, even as those datasets become very large. \n\n4. The coverage assumptions and continuous support seem to be quite strong. It seems that the tools are very reliant on having full support over continuous variables. For transformers the assumption is even stronger than having full support over the marginals to full support over the joint distributions of each pair of tokens which is even stronger. It is not clear how these assumptions could be relaxed.\n\n5. The paper does not study standard attention, but instead a simplified version. This should be made more clear in the abstract/intro.\n\n6. The numerical experiments basically only verify the theory. Moreover the plots showing large trends in increasing or decreasing loss are a somewhat confusing way to claim that length generalization is always happening (because the y-axes are very small). Perhaps adding some baselines/alternatives that do not length generalize could make it more clear whether the ability of these architectures to generalize is interesting.", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper considers the problems of length/compositional generalization from a functional expressivity perspective. It considers a variety of similar architectures under assumptions of realizability, some intricate assumptions about the support of continuous tokens, and strongly limited capacity (limiting the interaction between tokens). In this setting it considers the ERM solutions and shows that they must recover the ground truth features up to a linear transformation which allows them to generalize. There are negative results presented for all of the architectures considered if the capacity is increased.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The paper proposes a (to my knowledge) novel method to study length/compositional generalization. \n\n2. The paper does a good job of extending results to a family of related toy models of different types of architectures. This generality is a strength of the approach.\n\n3. The paper does a good job of presenting both negative and positive results. In particular the negative results actually seem to show that once the restrictive limits on the capacity in the architectures are weakened, then generalization can easily fail.", "weaknesses": "1. The perspective of the paper on the prospects of length/compositional generalization seems overly optimistic. In particular, in practice achieving length/compositional has proven very difficult. The papers cited on length generalization struggle to even achieve 2x length generalization despite using all sorts of unnatural tricks. This paper seems to approach the problem as \"length generalization happens and we want to explain it\", where the real situation appears to be \"length generalization is very difficult to achieve if even possible generally\". Even in the context of the paper, perhaps the more realistic results are the negative results for each architecture showing that if you allow for (still restrictive) higher capacity, the architectures are able to represent solutions that achieve zero in-distribution error but fail to generalize.\n\n2. It is not clear why we would expect the simplified models presented in the paper to bear any relation to things we may see in practice. Of course in theory we want to study simplified models, but the ultimate goal is for those models to bring insights to practical settings. The negative results indicate that as soon as we expand the very tight capacity restrictions, things can Simplified architectures seem to have no chance of applying to real settings.\n\n3. There is no consideration for statistical issues. Essentially the ERM approach is assuming infinite data. This is combined with the strong assumptions of full support over continuous variables to pin down the learned function to be similar to the realizable ground-truth solution. It is not clear if the arguments presented can apply at all to the case of finite datasets, even as those datasets become very large. \n\n4. The coverage assumptions and continuous support seem to be quite strong. It seems that the tools are very reliant on having full support over continuous variables. For transformers the assumption is even stronger than having full support over the marginals to full support over the joint distributions of each pair of tokens which is even stronger. It is not clear how these assumptions could be relaxed.\n\n5. The paper does not study standard attention, but instead a simplified version. This should be made more clear in the abstract/intro.\n\n6. The numerical experiments basically only verify the theory. Moreover the plots showing large trends in increasing or decreasing loss are a somewhat confusing way to claim that length generalization is always happening (because the y-axes are very small). Perhaps adding some baselines/alternatives that do not length generalize could make it more clear whether the ability of these architectures to generalize is interesting.", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730560984653}, {"id": "ZIbZToNhwe", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5063/Reviewer_MsSC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper theoretically studies length generalization and compositional generalization in various simplified models: deep sets, simplified one-layer transformers and simple RNNs and state space models. All results are given in the realizable case, i.e. when the ground-truth labels are generated by a model from the hypothesis class, and assuming some capacity restriction on the class. The authors define length generalization to be a setting where the model minimizes the expected risk up to length T and achieves zero generalization error on longer lengths. Similarly, they define compositional generalization of a model minimizing the expected risk on some input distribution and generalizing to the Cartesian product of the support of the different variables. Then, the authors show that the simplified models achieve length and compositional generalization, under assumptions on the capacity of the models. Additionally, the authors complement their results with experiments.", "review_text": "The paper theoretically studies length generalization and compositional generalization in various simplified models: deep sets, simplified one-layer transformers and simple RNNs and state space models. All results are given in the realizable case, i.e. when the ground-truth labels are generated by a model from the hypothesis class, and assuming some capacity restriction on the class. The authors define length generalization to be a setting where the model minimizes the expected risk up to length T and achieves zero generalization error on longer lengths. Similarly, they define compositional generalization of a model minimizing the expected risk on some input distribution and generalizing to the Cartesian product of the support of the different variables. Then, the authors show that the simplified models achieve length and compositional generalization, under assumptions on the capacity of the models. Additionally, the authors complement their results with experiments.", "strengths": "To my knowledge, this is the first work defining and analyzing length generalization, and one of the first works to theoretically study compositional generalization. The authors study several different architectures for learning over sequences and sets that are used in practice and show positive results on length and compositional generalization under reasonable assumptions. Additionally, the authors show that without restricting the capacity of the models, length generalization and compositional generalization are not possible.", "weaknesses": "-\tWhile the paper shows some novel theoretical results, the setting analyzed is somewhat idealized. Primarily, to my understanding, the authors assume that the model achieves zero error on the training distribution. Typically, however, theory of learning focuses on achieving (arbitrarily) small error and using a finite sample. It is not clear how these results extend to a setting where the learner achieves small, but non-zero, error on the training distribution.\n-\tThe result on transformers extends only to positional encodings where tokens far enough from one another do not affect each other, which does not apply to positional encodings used in practice.\n-\tThe authors can be a bit more rigorous in stating the definitions and setting. For example, in Definition 1, it is not clear what “zero generalization error” means. Does this mean zero error on the target distribution? Additionally, when assuming “a model trained on sequences of length up to train T”, does this mean that the model achieves zero error on the distribution P(T)?\n-\tAssumption 15, used in Theorem 4, is deferred to the appendix, which makes reading and understanding the result harder. This assumption should be stated in the main paper.\n-\tIt is not clear what is the takeaway from the experimental section. For example, the authors plot the error of different sequence models displaying different behaviors when varying the sequence length, but then admit that all errors are very low so the conclusion from these plots is not clear. What are the questions or hypotheses that these experiments are suppose to prove or validate?", "questions": "-\tThe models studied in the paper are simplified, one layer, models. Can the results shown in the paper be extended to deep, multi-layer, models?\n-\tThe discussion on high-capacity models, specifically the “negative results” on length and compositional generalization, assumes no constraints at all on the model, which may be too permissive, in contrast to the “positive results” which analyze a specific family of functions. Can these results be extended to general classes of functions with some more generic capacity control (e.g., finite classes or classes with bounded VC dimension)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper theoretically studies length generalization and compositional generalization in various simplified models: deep sets, simplified one-layer transformers and simple RNNs and state space models. All results are given in the realizable case, i.e. when the ground-truth labels are generated by a model from the hypothesis class, and assuming some capacity restriction on the class. The authors define length generalization to be a setting where the model minimizes the expected risk up to length T and achieves zero generalization error on longer lengths. Similarly, they define compositional generalization of a model minimizing the expected risk on some input distribution and generalizing to the Cartesian product of the support of the different variables. Then, the authors show that the simplified models achieve length and compositional generalization, under assumptions on the capacity of the models. Additionally, the authors complement their results with experiments.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "To my knowledge, this is the first work defining and analyzing length generalization, and one of the first works to theoretically study compositional generalization. The authors study several different architectures for learning over sequences and sets that are used in practice and show positive results on length and compositional generalization under reasonable assumptions. Additionally, the authors show that without restricting the capacity of the models, length generalization and compositional generalization are not possible.", "weaknesses": "-\tWhile the paper shows some novel theoretical results, the setting analyzed is somewhat idealized. Primarily, to my understanding, the authors assume that the model achieves zero error on the training distribution. Typically, however, theory of learning focuses on achieving (arbitrarily) small error and using a finite sample. It is not clear how these results extend to a setting where the learner achieves small, but non-zero, error on the training distribution.\n-\tThe result on transformers extends only to positional encodings where tokens far enough from one another do not affect each other, which does not apply to positional encodings used in practice.\n-\tThe authors can be a bit more rigorous in stating the definitions and setting. For example, in Definition 1, it is not clear what “zero generalization error” means. Does this mean zero error on the target distribution? Additionally, when assuming “a model trained on sequences of length up to train T”, does this mean that the model achieves zero error on the distribution P(T)?\n-\tAssumption 15, used in Theorem 4, is deferred to the appendix, which makes reading and understanding the result harder. This assumption should be stated in the main paper.\n-\tIt is not clear what is the takeaway from the experimental section. For example, the authors plot the error of different sequence models displaying different behaviors when varying the sequence length, but then admit that all errors are very low so the conclusion from these plots is not clear. What are the questions or hypotheses that these experiments are suppose to prove or validate?", "questions": "-\tThe models studied in the paper are simplified, one layer, models. Can the results shown in the paper be extended to deep, multi-layer, models?\n-\tThe discussion on high-capacity models, specifically the “negative results” on length and compositional generalization, assumes no constraints at all on the model, which may be too permissive, in contrast to the “positive results” which analyze a specific family of functions. Can these results be extended to general classes of functions with some more generic capacity control (e.g., finite classes or classes with bounded VC dimension)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730319755139}], "openreview_url": "https://openreview.net/forum?id=Hxm0hOxph2", "arxiv_id": "2402.04875", "paper_pdf": "papers/Hxm0hOxph2.pdf", "paper_pdf_sha256": "97cb08d76fd8c04dcc5212dc922ab907556a628cd687f8a3c4e0e4f3d0e251b0", "paper_pdf_bytes": 1863816, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/Length-and-Compositional-Generalization", "code_repository": "facebookresearch/Length-and-Compositional-Generalization", "code_commit": "5a77eb14d7b378bce0d0c6c69716fea8a038de05", "code_archive": "repos/Hxm0hOxph2.zip", "code_archive_sha256": "1a82e6860437e1d53716d77968f3fd360506dc4be93c9ca97105341d45a9953a", "code_archive_bytes": 72778, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 63, "github_languages": {"Python": 93691}, "github_archived": true, "github_pushed_at": "2024-07-19T15:24:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-provable-length-and-compositional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MpWRCiw8g5", "year": 2024, "status": "rejected", "title": "JOSENet: A Joint Stream Embedding Network for Violence Detection in Surveillance Videos", "authors": ["Pietro Nardelli", "Danilo Comminiello"], "authorids": ["~Pietro_Nardelli1", "~Danilo_Comminiello1"], "authors_source": "OpenReview API", "abstract": "Due to the ever-increasing availability of video surveillance cameras and the growing need for crime prevention, the violence detection task is attracting greater attention from the research community. With respect to other action recognition tasks, violence detection in surveillance videos shows additional issues. Indeed, violence detection requires real-world fights from surveillance cameras and available datasets seem to be very small compared with other action recognition datasets. Moreover, in surveillance applications, people in the scenes always differ for each video and the background of the footage differs for each camera. Also, violent actions in real-life surveillance videos must be detected quickly to prevent unwanted consequences, thus models would definitely benefit from a reduction in memory usage and computational costs. Such problems make classical action recognition methods difficult to be adopted. To tackle such problems, we introduce JOSENet, a novel self-supervised framework that provides outstanding performance for violence detection in surveillance videos. The proposed model receives two spatiotemporal video streams, i.e., RGB frames and optical flows, and involves a new regularized self-supervised learning approach for videos. JOSENet provides improved performance while requiring one-fourth of the number of frames per video segment and a reduced frame rate compared to state-of-the-art methods.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "V8I3j2tdsk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5668/Reviewer_94ds"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Paper proposes a novel violence detection framework which combines 2 features, 2 spatiotemporal streams (RBG + optical flow) and self-supervised learning (SSL).\n\nThe design is more efficient in memory usage (75%) and inference speed (2-fold). For the SSL, the paper adopts the VICReg which is more memory efficient.\n\nEmpirical experiments were done with RWF-2000, HMDB51, UCF101 and UCFCrime. The proposed framework was compared against the SOTA SSL methods: InfoNCE, UberNCE and CoCLR for the UCF101 dataset.", "review_text": "Paper proposes a novel violence detection framework which combines 2 features, 2 spatiotemporal streams (RBG + optical flow) and self-supervised learning (SSL).\n\nThe design is more efficient in memory usage (75%) and inference speed (2-fold). For the SSL, the paper adopts the VICReg which is more memory efficient.\n\nEmpirical experiments were done with RWF-2000, HMDB51, UCF101 and UCFCrime. The proposed framework was compared against the SOTA SSL methods: InfoNCE, UberNCE and CoCLR for the UCF101 dataset.", "strengths": "1. Paper's proposed method is more efficient in memory and inference speed compared to the original baseline methods.\n2. The motivation for the design is well explained.", "weaknesses": "1. Novelty is highly limited. The combination of optical flow with RGB has been used in multiple prior work. See references.\nThe novelty of SSL is also limited as it is a direct implementation of VICReg.\n\n2. Experimental design is confusing and does not directly support the core claim of the paper. Only SSL-based SOTA algorithms were directly compared with the proposed method for one single dataset (UCF101). There were several experiments on JoseNet based methods. But these experiments are not relevant to demonstrate the core claim of the paper \"outstanding performance for violence detection\" against other SOTA methods.\n\n3. (minor) Writing style is informal and not well-structured. This is especially for the experiment section. E.g. \"We\nhave noticed that we do not reach the state-of-the-art performances for RWF-2000. However, this is not a big deal in a deployment application.\". There is no reference to which experiment this statement refers to (which Table).\n\nReferences\n\nDiba, A., Pazandeh, A. M., & Van Gool, L. (2016). Efficient two-stream motion and appearance 3d cnns for video classification. arXiv preprint arXiv:1608.08851.\n\nWang, G., Muhammad, A., Liu, C., Du, L., & Li, D. (2021). Automatic recognition of fish behavior with a fusion of RGB and optical flow data based on deep learning. Animals, 11(10), 2774.\n\nLi, S., Zhang, L., & Diao, X. (2020). Deep-learning-based human intention prediction using RGB images and optical flow. Journal of Intelligent & Robotic Systems, 97, 95-107.", "questions": "Why is the comparison against SOTA limited to SSL methods for a single UCF101? This is insufficient to show the generalization of the claim of superior performance of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Paper proposes a novel violence detection framework which combines 2 features, 2 spatiotemporal streams (RBG + optical flow) and self-supervised learning (SSL).\n\nThe design is more efficient in memory usage (75%) and inference speed (2-fold). For the SSL, the paper adopts the VICReg which is more memory efficient.\n\nEmpirical experiments were done with RWF-2000, HMDB51, UCF101 and UCFCrime. The proposed framework was compared against the SOTA SSL methods: InfoNCE, UberNCE and CoCLR for the UCF101 dataset.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1. Paper's proposed method is more efficient in memory and inference speed compared to the original baseline methods.\n2. The motivation for the design is well explained.", "weaknesses": "1. Novelty is highly limited. The combination of optical flow with RGB has been used in multiple prior work. See references.\nThe novelty of SSL is also limited as it is a direct implementation of VICReg.\n\n2. Experimental design is confusing and does not directly support the core claim of the paper. Only SSL-based SOTA algorithms were directly compared with the proposed method for one single dataset (UCF101). There were several experiments on JoseNet based methods. But these experiments are not relevant to demonstrate the core claim of the paper \"outstanding performance for violence detection\" against other SOTA methods.\n\n3. (minor) Writing style is informal and not well-structured. This is especially for the experiment section. E.g. \"We\nhave noticed that we do not reach the state-of-the-art performances for RWF-2000. However, this is not a big deal in a deployment application.\". There is no reference to which experiment this statement refers to (which Table).\n\nReferences\n\nDiba, A., Pazandeh, A. M., & Van Gool, L. (2016). Efficient two-stream motion and appearance 3d cnns for video classification. arXiv preprint arXiv:1608.08851.\n\nWang, G., Muhammad, A., Liu, C., Du, L., & Li, D. (2021). Automatic recognition of fish behavior with a fusion of RGB and optical flow data based on deep learning. Animals, 11(10), 2774.\n\nLi, S., Zhang, L., & Diao, X. (2020). Deep-learning-based human intention prediction using RGB images and optical flow. Journal of Intelligent & Robotic Systems, 97, 95-107.", "questions": "Why is the comparison against SOTA limited to SSL methods for a single UCF101? This is insufficient to show the generalization of the claim of superior performance of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "Not applicable.", "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698834663110}, {"id": "EhGA6Vt3Op", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5668/Reviewer_o7vs"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces JOSENet, a network for video violence detection. It contains a pretraining part and a detection part. Given the RGB and Flow inputs, a two-stream flow gated network (FGN) is firstly pretrained on UCF-101, HMDB-51 and UCF-Crime datasets using VICReg method. Then, the pretrained FGN weights are used to initialize the FGN in the detection part. In this way, the model requires less training data and generalizes better. In addition, some optimization of the network improves the efficiency of the model in terms of memory consumption and computation load. The proposed method is evaluated on RWF-2000 dataset.", "review_text": "This paper introduces JOSENet, a network for video violence detection. It contains a pretraining part and a detection part. Given the RGB and Flow inputs, a two-stream flow gated network (FGN) is firstly pretrained on UCF-101, HMDB-51 and UCF-Crime datasets using VICReg method. Then, the pretrained FGN weights are used to initialize the FGN in the detection part. In this way, the model requires less training data and generalizes better. In addition, some optimization of the network improves the efficiency of the model in terms of memory consumption and computation load. The proposed method is evaluated on RWF-2000 dataset.", "strengths": "1) The overall idea is easy to understand and makes sense.\n2) By efficient implementation, the model requires less memory and less frames for each segment.\n3) The model leverages self-supervised learning to improve the generalization of the model.", "weaknesses": "1) The goal of the paper is violence detection, but there is no related contents in the method part. Necessary components such as loss function of violence detection should be included. \n2) The proposed “computational enhancement” is just hyper-parameter tuning. N_s=7.5s is found to be the optimal. However, different datasets may have different optimal parameters. More justification are needed to demonstrate the generalization performance. \n3) The theoretical contribution is limited. The pretraining part is borrowed from VICReg and the detector is borrowed from FGN. \n4) The proposed method is only evaluated on RWF-2000 dataset, which is not enough. I suggest authors to include results on more datasets since you claim the proposed method generalizes better. \n5) Missing comparison with recent methods such as:\n[1] Islam, Zahidul, et al. \"Efficient two-stream network for violence detection using separable convolutional lstm.\" 2021 International Joint Conference on Neural Networks (IJCNN). IEEE, 2021.\n[2] Garcia-Cobo, Guillermo, and Juan C. SanMiguel. \"Human skeletons and change detection for efficient violence detection in surveillance videos.\" Computer Vision and Image Understanding 233 (2023): 103739.\n6) Compared with other methods, the proposed method uses addition training data (UCF-101, HMDB-51, and UCF-Crime). This may be a concern the comparison is not fair.   \n7) The related work of violence detection is incomplete, it should contains more recent methods and discussion. \n8) To demonstrate the efficiency, a comparison with other methods should be included.", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces JOSENet, a network for video violence detection. It contains a pretraining part and a detection part. Given the RGB and Flow inputs, a two-stream flow gated network (FGN) is firstly pretrained on UCF-101, HMDB-51 and UCF-Crime datasets using VICReg method. Then, the pretrained FGN weights are used to initialize the FGN in the detection part. In this way, the model requires less training data and generalizes better. In addition, some optimization of the network improves the efficiency of the model in terms of memory consumption and computation load. The proposed method is evaluated on RWF-2000 dataset.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1) The overall idea is easy to understand and makes sense.\n2) By efficient implementation, the model requires less memory and less frames for each segment.\n3) The model leverages self-supervised learning to improve the generalization of the model.", "weaknesses": "1) The goal of the paper is violence detection, but there is no related contents in the method part. Necessary components such as loss function of violence detection should be included. \n2) The proposed “computational enhancement” is just hyper-parameter tuning. N_s=7.5s is found to be the optimal. However, different datasets may have different optimal parameters. More justification are needed to demonstrate the generalization performance. \n3) The theoretical contribution is limited. The pretraining part is borrowed from VICReg and the detector is borrowed from FGN. \n4) The proposed method is only evaluated on RWF-2000 dataset, which is not enough. I suggest authors to include results on more datasets since you claim the proposed method generalizes better. \n5) Missing comparison with recent methods such as:\n[1] Islam, Zahidul, et al. \"Efficient two-stream network for violence detection using separable convolutional lstm.\" 2021 International Joint Conference on Neural Networks (IJCNN). IEEE, 2021.\n[2] Garcia-Cobo, Guillermo, and Juan C. SanMiguel. \"Human skeletons and change detection for efficient violence detection in surveillance videos.\" Computer Vision and Image Understanding 233 (2023): 103739.\n6) Compared with other methods, the proposed method uses addition training data (UCF-101, HMDB-51, and UCF-Crime). This may be a concern the comparison is not fair.   \n7) The related work of violence detection is incomplete, it should contains more recent methods and discussion. \n8) To demonstrate the efficiency, a comparison with other methods should be included.", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics concerns", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698631030270}, {"id": "7GWbfzAnc7", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5668/Reviewer_tKHs"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper describes an approach for performing the video task of violence detection in surveillance videos by employing a self-supervised learning network to help improve the primary supervised model. The core network to perform the primary task is based on flow gated network (FGN), by Cheng et al (2021). The semi-supervised learning block applies VICReg approach, by Bardes et al. (2021), to the two streams of RGB and optical flow. The results are reported on three datasets related to activity recognition with the comparison with multiple SOTA approaches and an ablation study.", "review_text": "This paper describes an approach for performing the video task of violence detection in surveillance videos by employing a self-supervised learning network to help improve the primary supervised model. The core network to perform the primary task is based on flow gated network (FGN), by Cheng et al (2021). The semi-supervised learning block applies VICReg approach, by Bardes et al. (2021), to the two streams of RGB and optical flow. The results are reported on three datasets related to activity recognition with the comparison with multiple SOTA approaches and an ablation study.", "strengths": "The paper describes an interesting idea that can leverage the strengths of semi-supervised learning in the domain of violence detection in surveillance videos where the rarity of the events poses challenges for obtaining a large volume of positive training samples and the need for a low false alarm rate. The proposed approach also has some interesting nuggets related to computational efficiency and reduced memory footprint. They have also studied the tradeoff between the size of the temporal window, framerate, and quality of results.", "weaknesses": "The problem, application, and the core part of the solution (FGN) is not new. However, the addition of SSL \n\nThe baseline model from `Sec. 4.2` should have been reported in the tabular form for a more effective presentation of material and instead of explaining the numerical differences in the narrative form as done in `Sec. 4.2` and other sections. It should be clear from ONE table the various variants, baselines, and the final version. Additionally, it is hard to follow this paper at times because the different tables are reporting results on different datasets. Are the results in this section reported on the exact same test data as that in Table 3? If so, then should we be comparing $F_1$ of $85.87$ (baseline) with $86.5$ (JOSENet)? i.e. improvement of $0.63$? Is it also fair to say that the baseline approach is very close to FGN, by Cheng et al (2021)? \n\nThe main result comparing JOSENet with SOTA in Table 3 has aspects that are not clear. I assumed this statement\n`We decide to take as reference the results obtained on the 15% subset of UCF-101 with JOSENet.` \nmeant that Table 3 results are on UCF101 but then `a pretraining obtained on a random 15% subset of UCF-101` suggests that it was used for pretraining. Is it a different 15%? More importantly, UCF101 does NOT have violent activities in surveillance scenes (to the best of my knowledge) in the way it has been portrayed in the motivation described in the paper. There are activities like Punch or Boxing Punching Bag, but not much else. Additionally, why stick with some *random* 15% split of UCF101 instead of using the standard test split that could be compared with the SOTA. \n\nThe writing quality of the paper can be improved significantly. There are several grammatical mistakes, a few long run-on sentences, unusual usage of some phrases, and confusing or inconsistent usage of citations that break the flow.", "questions": "1. It was surprising that results were not reported explicitly on the RWF-2000 dataset in the `4. Experimental Results`, as far as I could tell. In my opinion, it is unusual to make statements like this:\n`pg 8: We have noticed that we do not reach the state-of-the-art performances for RWF-2000.`\nand not provide the quantified numbers. The other statement (`To train and validate the model during supervised learning we use the RWF-2000 dataset`) was also noted. \n\n2. Is there a reason why Table 3 does not have a row with a comparison with FGN, by Cheng et al (2021)?\n\n3. Table 3, AUC column has numbers in [0,100] and [0,1.0] ranges. Are those just typos? \n\n4. Table 5, the use of temporal pooling is not clear as it makes things worse as reported by the scores. The explanation in `Sec. 5` is unclear. The table does not support this claim (if I am following it as intended):\n`To find a confirmation of this approach, using the zoom crop strategy, we apply the temporal pooling\nin the merging block, obtaining on the target task a very low value for most of the evaluation metrics\nused.`\n\n5. pg: 2, FGN was not defined or cited until pg 3 so it was confusing.\n\n6. pg: 2, should `contrastive learning (CT)` be `contrastive learning (CL)` ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper describes an approach for performing the video task of violence detection in surveillance videos by employing a self-supervised learning network to help improve the primary supervised model. The core network to perform the primary task is based on flow gated network (FGN), by Cheng et al (2021). The semi-supervised learning block applies VICReg approach, by Bardes et al. (2021), to the two streams of RGB and optical flow. The results are reported on three datasets related to activity recognition with the comparison with multiple SOTA approaches and an ablation study.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "The paper describes an interesting idea that can leverage the strengths of semi-supervised learning in the domain of violence detection in surveillance videos where the rarity of the events poses challenges for obtaining a large volume of positive training samples and the need for a low false alarm rate. The proposed approach also has some interesting nuggets related to computational efficiency and reduced memory footprint. They have also studied the tradeoff between the size of the temporal window, framerate, and quality of results.", "weaknesses": "The problem, application, and the core part of the solution (FGN) is not new. However, the addition of SSL \n\nThe baseline model from `Sec. 4.2` should have been reported in the tabular form for a more effective presentation of material and instead of explaining the numerical differences in the narrative form as done in `Sec. 4.2` and other sections. It should be clear from ONE table the various variants, baselines, and the final version. Additionally, it is hard to follow this paper at times because the different tables are reporting results on different datasets. Are the results in this section reported on the exact same test data as that in Table 3? If so, then should we be comparing $F_1$ of $85.87$ (baseline) with $86.5$ (JOSENet)? i.e. improvement of $0.63$? Is it also fair to say that the baseline approach is very close to FGN, by Cheng et al (2021)? \n\nThe main result comparing JOSENet with SOTA in Table 3 has aspects that are not clear. I assumed this statement\n`We decide to take as reference the results obtained on the 15% subset of UCF-101 with JOSENet.` \nmeant that Table 3 results are on UCF101 but then `a pretraining obtained on a random 15% subset of UCF-101` suggests that it was used for pretraining. Is it a different 15%? More importantly, UCF101 does NOT have violent activities in surveillance scenes (to the best of my knowledge) in the way it has been portrayed in the motivation described in the paper. There are activities like Punch or Boxing Punching Bag, but not much else. Additionally, why stick with some *random* 15% split of UCF101 instead of using the standard test split that could be compared with the SOTA. \n\nThe writing quality of the paper can be improved significantly. There are several grammatical mistakes, a few long run-on sentences, unusual usage of some phrases, and confusing or inconsistent usage of citations that break the flow.", "questions": "1. It was surprising that results were not reported explicitly on the RWF-2000 dataset in the `4. Experimental Results`, as far as I could tell. In my opinion, it is unusual to make statements like this:\n`pg 8: We have noticed that we do not reach the state-of-the-art performances for RWF-2000.`\nand not provide the quantified numbers. The other statement (`To train and validate the model during supervised learning we use the RWF-2000 dataset`) was also noted. \n\n2. Is there a reason why Table 3 does not have a row with a comparison with FGN, by Cheng et al (2021)?\n\n3. Table 3, AUC column has numbers in [0,100] and [0,1.0] ranges. Are those just typos? \n\n4. Table 5, the use of temporal pooling is not clear as it makes things worse as reported by the scores. The explanation in `Sec. 5` is unclear. The table does not support this claim (if I am following it as intended):\n`To find a confirmation of this approach, using the zoom crop strategy, we apply the temporal pooling\nin the merging block, obtaining on the target task a very low value for most of the evaluation metrics\nused.`\n\n5. pg: 2, FGN was not defined or cited until pg 3 so it was confusing.\n\n6. pg: 2, should `contrastive learning (CT)` be `contrastive learning (CL)` ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698625056499}, {"id": "00HPAwJYlu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5668/Reviewer_sND9"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces JOSENet, a novel framework designed for violence detection in surveillance videos. It aims to tackle the challenges of real-world surveillance, such as varying scenes, actors, and the need for real-time detection. The framework consists of a primary target model and an auxiliary self-supervised learning (SSL) model. It uses multiple datasets for training and validation, applies various preprocessing and data augmentation strategies, and evaluates the model using a comprehensive set of metrics.", "review_text": "The paper introduces JOSENet, a novel framework designed for violence detection in surveillance videos. It aims to tackle the challenges of real-world surveillance, such as varying scenes, actors, and the need for real-time detection. The framework consists of a primary target model and an auxiliary self-supervised learning (SSL) model. It uses multiple datasets for training and validation, applies various preprocessing and data augmentation strategies, and evaluates the model using a comprehensive set of metrics.", "strengths": "**Originality**\n\nThe paper is innovative in proposing a dual-model architecture, involving a primary target model and an auxiliary SSL model. It also introduces a new SSL algorithm based on VICReg and a novel data augmentation strategy called \"zoom crop.\"\n\n**Quality**\n\nThe research is thorough, with detailed experimental settings, multiple datasets, and a diverse set of evaluation metrics. The use of an auxiliary SSL model to achieve a balance between performance and computational resources is commendable.\n\n**Clarity**\n\nThe paper is well-structured and clear, with each section contributing to the reader's understanding of the proposed framework.\n\n**Significance**\n\nThe work addresses a vital real-world problem, that of violence detection in surveillance videos, and proposes a framework that seems both effective and efficient.", "weaknesses": "Lack of Details: Some sections could provide more implementation details, especially on how the VICReg loss and weight optimization between the two models are implemented.\n\nDataset Limitations: While multiple datasets are used, they are mostly centered around violence detection, which could limit the model's generalizability across domains.\n\nRobustness: The paper does not address how the model handles potential issues like occlusion, varying light conditions, or camera angles, which are common in real-world surveillance.\n\nHyperparameter Tuning: The paper doesn't discuss the process or criteria for hyperparameter selection, which could affect the model's performance.", "questions": "1.\tCould you provide more details on the \"zoom crop\" data augmentation strategy, specifically its effectiveness and efficiency?\n\n2.\tWhy were these particular datasets chosen, and have you considered using more diverse datasets to improve the model's generalizability?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces JOSENet, a novel framework designed for violence detection in surveillance videos. It aims to tackle the challenges of real-world surveillance, such as varying scenes, actors, and the need for real-time detection. The framework consists of a primary target model and an auxiliary self-supervised learning (SSL) model. It uses multiple datasets for training and validation, applies various preprocessing and data augmentation strategies, and evaluates the model using a comprehensive set of metrics.", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "strengths": "**Originality**\n\nThe paper is innovative in proposing a dual-model architecture, involving a primary target model and an auxiliary SSL model. It also introduces a new SSL algorithm based on VICReg and a novel data augmentation strategy called \"zoom crop.\"\n\n**Quality**\n\nThe research is thorough, with detailed experimental settings, multiple datasets, and a diverse set of evaluation metrics. The use of an auxiliary SSL model to achieve a balance between performance and computational resources is commendable.\n\n**Clarity**\n\nThe paper is well-structured and clear, with each section contributing to the reader's understanding of the proposed framework.\n\n**Significance**\n\nThe work addresses a vital real-world problem, that of violence detection in surveillance videos, and proposes a framework that seems both effective and efficient.", "weaknesses": "Lack of Details: Some sections could provide more implementation details, especially on how the VICReg loss and weight optimization between the two models are implemented.\n\nDataset Limitations: While multiple datasets are used, they are mostly centered around violence detection, which could limit the model's generalizability across domains.\n\nRobustness: The paper does not address how the model handles potential issues like occlusion, varying light conditions, or camera angles, which are common in real-world surveillance.\n\nHyperparameter Tuning: The paper doesn't discuss the process or criteria for hyperparameter selection, which could affect the model's performance.", "questions": "1.\tCould you provide more details on the \"zoom crop\" data augmentation strategy, specifically its effectiveness and efficiency?\n\n2.\tWhy were these particular datasets chosen, and have you considered using more diverse datasets to improve the model's generalizability?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697794896805}], "openreview_url": "https://openreview.net/forum?id=MpWRCiw8g5", "arxiv_id": "2405.02961", "paper_pdf": "papers/MpWRCiw8g5.pdf", "paper_pdf_sha256": "7b0208b7c9f39304499239d936852dfccbfedeae8904035f81d0232131213fda", "paper_pdf_bytes": 3995719, "paper_pdf_source": "openreview", "code_url": "https://github.com/ispamm/JOSENet", "code_repository": "ispamm/JOSENet", "code_commit": "3f9512e9e08e6c4e1cd983993fa57369996bc79e", "code_archive": "repos/MpWRCiw8g5.zip", "code_archive_sha256": "ec68abe8eb5ca4bf2b712adbff0531d15077181bb31dacaa846742f3d2c2888a", "code_archive_bytes": 28501, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 56, "github_languages": {"Python": 135093}, "github_archived": false, "github_pushed_at": "2025-06-05T08:15:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/josenet-a-joint-stream-embedding-network-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MCe881WzBr0", "year": 2023, "status": "rejected", "title": "Variational Classification", "authors": ["Shehzaad Zuzar Dhuliawala", "Mrinmaya Sachan", "Carl Allen"], "authorids": ["~Shehzaad_Zuzar_Dhuliawala3", "~Mrinmaya_Sachan3", "~Carl_Allen1"], "authors_source": "OpenReview API", "abstract": "Classification tasks, ubiquitous across machine learning, are commonly tackled by a suitably designed neural network with a softmax output layer, mapping each data point to a categorical distribution over class labels. \nWe extend this familiar model from a latent variable perspective to variational classification (VC), analogous to how the variational auto-encoder relates to its deterministic counterpart. We derive a training objective based on the ELBO together with an \\textit{adversarial} approach for optimising it.\n\nWithin this framework, we identify design choices made implicitly in off-the-shelf softmax functions and can instead include domain-specific assumptions, such as class-conditional latent priors. We demonstrate benefits of the VC model in image classification. We show on several standard datasets, that treating inputs to the softmax layer as latent variables under a mixture of Gaussians prior, improves several desirable aspects of a classifier, such as prediction accuracy, calibration, out-of-domain calibration and adversarial robustness.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "i7OcgwWqxW5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4907/Reviewer_Lvkv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces variational classification (VC) by treating the inputs to the softmax layer of a classification model as latent variables with a mixture of Gaussians prior. To achieve this probabilistic interpretation of the softmax classifier, they derived an objective to be minimized, which is similar to the ELBO used in VAEs. The resulting VC model generalizes the softmax classifier (similar to VAE vs deterministic autoencoder) and enables incorporating class-conditional priors. A derivation of the training objective is provided with some practical design choices to optimize it. The evaluations are performed on several datasets (CIFAR-10, CIFAR-100, SVHN, CelebA, MNIST) in terms of accuracy, calibration, OOD generalization, adversarial robustness, and performance on low data regime. ", "review_text": "While the paper makes some interesting points, I think it is still a borderline one as the performance is somewhat mixed and gains are mostly incremental. I would be happy to revise my score if the above concerns are addressed sufficiently.", "strengths": "Strengths:\n\n- The generalization of the softmax classifier with a Bayesian interpretation is interesting and it can potentially have impacts on the several desired properties for classification models such as calibration and OOD generalization. \n\n- The authors provide a derivation for the VC objective with nice connections to VAE. I also liked that they provided practical implementation points of the VC using the commonly employed tools like the reparameterization trick. However, I believe it could be useful to provide an algorithm table or pseudo-code for better clarity. Could the authors also clarify if there are any instability issues faced during the adversarial optimization?\n\n- The experiments were selected to validate different aspects of the proposed VC. More on that below.\n\nWeaknesses: My main concerns are about the experimental results.\n\n- From the results, it is hard to say VC is doing notably better than GM or CE methods, except model calibration (even for that, it is sometimes worse than CE, see Fig. 4. left). Could this be due to optimization difficulties, e.g. due to adversarial objective, or are there other reasons?\n\n- For calibration, adding temperature scaling as another baseline could be helpful to gauge the improvements.\n\n- Can this method be applicable to more complex tasks like ImageNet classification and more strong adversarial attacks like PGD? Also for low data regime, are there any benefits for datasets other than MNIST where the performance is more or less saturated? What happens to calibration under low data regime?\n\n- Demonstration of the experimental results are somewhat inconsistent. For example, why is there no standard deviation in Table 1 from multiple runs while they are provided for Table 4? What do the bars and whiskers represent in Fig. 4?\n\n- Is there a substantial increase in the training time compared to standard training? An analysis on this could be useful for practitioners.\n\n- Outlining the experimental protocol, e.g. used augmentations and hyperparameters, could increase the reproducibility of the work.\n\n- I found Sec. 3.3 to be somewhat confusing. Did the authors observe whether the model learned any disentangled and semantically meaningful representations? \n\nMinor: There are several typos throughout the paper. Also, the font size in figures are too small.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces variational classification (VC) by treating the inputs to the softmax layer of a classification model as latent variables with a mixture of Gaussians prior. To achieve this probabilistic interpretation of the softmax classifier, they derived an objective to be minimized, which is similar to the ELBO used in VAEs. The resulting VC model generalizes the softmax classifier (similar to VAE vs deterministic autoencoder) and enables incorporating class-conditional priors. A derivation of the training objective is provided with some practical design choices to optimize it. The evaluations are performed on several datasets (CIFAR-10, CIFAR-100, SVHN, CelebA, MNIST) in terms of accuracy, calibration, OOD generalization, adversarial robustness, and performance on low data regime. ", "strength_and_weaknesses": "Strengths:\n\n- The generalization of the softmax classifier with a Bayesian interpretation is interesting and it can potentially have impacts on the several desired properties for classification models such as calibration and OOD generalization. \n\n- The authors provide a derivation for the VC objective with nice connections to VAE. I also liked that they provided practical implementation points of the VC using the commonly employed tools like the reparameterization trick. However, I believe it could be useful to provide an algorithm table or pseudo-code for better clarity. Could the authors also clarify if there are any instability issues faced during the adversarial optimization?\n\n- The experiments were selected to validate different aspects of the proposed VC. More on that below.\n\nWeaknesses: My main concerns are about the experimental results.\n\n- From the results, it is hard to say VC is doing notably better than GM or CE methods, except model calibration (even for that, it is sometimes worse than CE, see Fig. 4. left). Could this be due to optimization difficulties, e.g. due to adversarial objective, or are there other reasons?\n\n- For calibration, adding temperature scaling as another baseline could be helpful to gauge the improvements.\n\n- Can this method be applicable to more complex tasks like ImageNet classification and more strong adversarial attacks like PGD? Also for low data regime, are there any benefits for datasets other than MNIST where the performance is more or less saturated? What happens to calibration under low data regime?\n\n- Demonstration of the experimental results are somewhat inconsistent. For example, why is there no standard deviation in Table 1 from multiple runs while they are provided for Table 4? What do the bars and whiskers represent in Fig. 4?\n\n- Is there a substantial increase in the training time compared to standard training? An analysis on this could be useful for practitioners.\n\n- Outlining the experimental protocol, e.g. used augmentations and hyperparameters, could increase the reproducibility of the work.\n\n- I found Sec. 3.3 to be somewhat confusing. Did the authors observe whether the model learned any disentangled and semantically meaningful representations? \n\nMinor: There are several typos throughout the paper. Also, the font size in figures are too small.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear. For reproducibility, I think more details should be provided.", "summary_of_the_review": "While the paper makes some interesting points, I think it is still a borderline one as the performance is somewhat mixed and gains are mostly incremental. I would be happy to revise my score if the above concerns are addressed sufficiently.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667418077833}, {"id": "GIDx7BGOdDk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4907/Reviewer_qvVd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a variational classifier (VC). Typical machine learning classifiers use sigmoid or softmax to deterministically map last layer feature vector to class label predictions. This paper revisits the MLE, MAP and Bayesian under a variational framework. This work designs a novel objective based on ELBO and adversarial/contrastive technique. Experiments show that VC has some desired properties in interpolation, prediction confidence, out-of-sample detection, etc.", "review_text": "This paper is a more theory-oriented paper with empirical experiments support. The idea is interesting, though resembling several prior works. This paper should better shape its relations with literature and elaborate on the model details and experiments. ", "strengths": "Pros:\n1. Compared to deterministic methods, probabilistic ones like variational model can also smooth the predictions and alleviate overfitting issues. That's why this method can potentially help with calibration, interpolation, OOD, etc.\n2. The paper explains the details very thoroughly, and carefully compare the difference and similarity between existing and proposed methods.\n3. The experiments explored multiple domains/tasks.\n\nCons:\n1. From a machine learning perspective, I think another view of understanding this paper is the embedding learning. Both input feature x and label y are mapped to a latent embedding z. The label learnt probabilistic space is the prior space and the input learnt probabilistic space is the posterior space. You kinda try to match the two spaces. While probabilistic method often brings smoothness, the performance loss is also concerned. In general, I feel this work still doesn't jump out of this framework and maybe suffer from the same issues?\n2. Though this type of variational learning is relatively novel, the use of VAE in classification has been quite ubiquitous. The latent variable models also learn and align latent subspaces [1, 2]. I think you could discuss these works.\n3. While this paper has experiments on multiple setups/scenarios, they look a bit thin to me. First, the major datasets are cifar-10, cifar-100, etc. In the computer vision domain, these are just too small. It's less convincing to only use these datasets. Also, you only tested WideResNet-28-10 and ResNet-50. More and larger models would give readers a more comprehensive view on the performance.\n4. You can give some derivation steps for Eq. 5 and 6, at least in Appendix. \n5. Figure 2: Variation Classier -> Variational Classifier\n\n[1] Bai, J., Kong, S. and Gomes, C.P., 2022, June. Gaussian mixture variational autoencoder with contrastive learning for multi-label classification. In ICML (pp. 1383-1398). PMLR.\n\n[2] Bai, J., Kong, S. and Gomes, C., 2021, January. Disentangled variational autoencoder based multi-label classification with covariance-aware multivariate probit model. In IJCAI (pp. 4313-4321).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a variational classifier (VC). Typical machine learning classifiers use sigmoid or softmax to deterministically map last layer feature vector to class label predictions. This paper revisits the MLE, MAP and Bayesian under a variational framework. This work designs a novel objective based on ELBO and adversarial/contrastive technique. Experiments show that VC has some desired properties in interpolation, prediction confidence, out-of-sample detection, etc.", "strength_and_weaknesses": "Pros:\n1. Compared to deterministic methods, probabilistic ones like variational model can also smooth the predictions and alleviate overfitting issues. That's why this method can potentially help with calibration, interpolation, OOD, etc.\n2. The paper explains the details very thoroughly, and carefully compare the difference and similarity between existing and proposed methods.\n3. The experiments explored multiple domains/tasks.\n\nCons:\n1. From a machine learning perspective, I think another view of understanding this paper is the embedding learning. Both input feature x and label y are mapped to a latent embedding z. The label learnt probabilistic space is the prior space and the input learnt probabilistic space is the posterior space. You kinda try to match the two spaces. While probabilistic method often brings smoothness, the performance loss is also concerned. In general, I feel this work still doesn't jump out of this framework and maybe suffer from the same issues?\n2. Though this type of variational learning is relatively novel, the use of VAE in classification has been quite ubiquitous. The latent variable models also learn and align latent subspaces [1, 2]. I think you could discuss these works.\n3. While this paper has experiments on multiple setups/scenarios, they look a bit thin to me. First, the major datasets are cifar-10, cifar-100, etc. In the computer vision domain, these are just too small. It's less convincing to only use these datasets. Also, you only tested WideResNet-28-10 and ResNet-50. More and larger models would give readers a more comprehensive view on the performance.\n4. You can give some derivation steps for Eq. 5 and 6, at least in Appendix. \n5. Figure 2: Variation Classier -> Variational Classifier\n\n[1] Bai, J., Kong, S. and Gomes, C.P., 2022, June. Gaussian mixture variational autoencoder with contrastive learning for multi-label classification. In ICML (pp. 1383-1398). PMLR.\n\n[2] Bai, J., Kong, S. and Gomes, C., 2021, January. Disentangled variational autoencoder based multi-label classification with covariance-aware multivariate probit model. In IJCAI (pp. 4313-4321).", "clarity,_quality,_novelty_and_reproducibility": "This paper introduces a lot of concepts, but in a relatively clear way. The quality of the paper can be improved by providing more experimental results. The paper does bring some novelty in addressing classification from a variational direction. The code is not provided and it seems non-trivial to implement the whole algorithm.", "summary_of_the_review": "This paper is a more theory-oriented paper with empirical experiments support. The idea is interesting, though resembling several prior works. This paper should better shape its relations with literature and elaborate on the model details and experiments. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666937388580}, {"id": "m7zbQ8n0sC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4907/Reviewer_aooA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a Variational Classification (VC), which generalizes the softmax classifier, and the authors claim that it mirrors the relationship between the variational auto-encoder and the deterministic auto-encoder. The main contribution is to include the latent variable to softmax classifier and then VC objective analogous to the ELBO is designed. The experimental results show that the VC classifier outperforms the standard softmax in several ways, in particular in terms of calibration, adversarial attacks and when data is scarce. \n", "review_text": "The paper adds a latent variable to classifier and impose constraints (class conditional priors) in the objective function.\nOverall it is a good paper, but not ready for publication.", "strengths": "Quality/Clarity: the paper is well written and the techniques presented are ok to follow. Its motivation is clear, which is to generalize softmax and design a better model: (1) modeling p(y|x) (2) measure of\nsimilarity between data samples. On a technical level, I do not see how VC objective in Eq. 6 related to ELBO. The authors need to give more details derivation from Eq.5 to Eq. 6.\n\nOriginality/significance: the idea is interesting, which introduces latent variable and adds more constraints to generalize the softmax function. However, it gives readers an impression that the VC objective in Eq. 6 is a heuristic result, not from ELBO. If the authors can provide the detailed derivation to bridge the gap as well as show more experimental results, it should be an accepted paper. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents a Variational Classification (VC), which generalizes the softmax classifier, and the authors claim that it mirrors the relationship between the variational auto-encoder and the deterministic auto-encoder. The main contribution is to include the latent variable to softmax classifier and then VC objective analogous to the ELBO is designed. The experimental results show that the VC classifier outperforms the standard softmax in several ways, in particular in terms of calibration, adversarial attacks and when data is scarce. \n", "strength_and_weaknesses": "Quality/Clarity: the paper is well written and the techniques presented are ok to follow. Its motivation is clear, which is to generalize softmax and design a better model: (1) modeling p(y|x) (2) measure of\nsimilarity between data samples. On a technical level, I do not see how VC objective in Eq. 6 related to ELBO. The authors need to give more details derivation from Eq.5 to Eq. 6.\n\nOriginality/significance: the idea is interesting, which introduces latent variable and adds more constraints to generalize the softmax function. However, it gives readers an impression that the VC objective in Eq. 6 is a heuristic result, not from ELBO. If the authors can provide the detailed derivation to bridge the gap as well as show more experimental results, it should be an accepted paper. ", "clarity,_quality,_novelty_and_reproducibility": "The idea to introduce the latent variable to classification task is interesting, but I think it will be better to add more theoretical analysis and experiments to support the claims. ", "summary_of_the_review": "The paper adds a latent variable to classifier and impose constraints (class conditional priors) in the objective function.\nOverall it is a good paper, but not ready for publication.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666730307225}], "openreview_url": "https://openreview.net/forum?id=MCe881WzBr0", "arxiv_id": "2305.10406", "paper_pdf": "papers/MCe881WzBr0.pdf", "paper_pdf_sha256": "c5b40d01fc7e7e96c8f49ffda7db42285d62f702c5ebdb946b102015af38851a", "paper_pdf_bytes": 851388, "paper_pdf_source": "openreview", "code_url": "https://github.com/shehzaadzd/variational-classification", "code_repository": "shehzaadzd/variational-classification", "code_commit": "e7463a33d97a60497671b06167d26f813b61d67d", "code_archive": "repos/MCe881WzBr0.zip", "code_archive_sha256": "b40f637c7cced1bf34a2bdfbacd66f69055fe89fb95ce226e5ae94554f8bdf0e", "code_archive_bytes": 22906, "code_file_count": 8, "code_extensions": {".py": 5, ".sh": 3}, "github_disk_usage_kb": 96, "github_languages": {"Python": 82521, "Shell": 269}, "github_archived": false, "github_pushed_at": "2023-12-22T18:11:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variational-classification"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7vXQJ2QW8hR", "year": 2022, "status": "rejected", "title": "Max-Affine Spline Insights Into Deep Network Pruning", "authors": ["Randall Balestriero", "Haoran You", "Zhihan Lu", "Yutong Kou", "Huihong Shi", "Yingyan Lin", "Richard Baraniuk"], "authorids": ["~Randall_Balestriero1", "~Haoran_You1", "~Zhihan_Lu1", "~Yutong_Kou1", "~Huihong_Shi1", "~Yingyan_Lin1", "~Richard_Baraniuk1"], "authors_source": "OpenReview API", "abstract": "State-of-the-art (SOTA) approaches to deep network (DN) training overparametrize the model and then prune a posteriori to obtain a \"winning ticket'' subnetwork that can be trained from scratch to achieve high accuracy. To date, the literature has remained largely empirical and hence provides little insights into how pruning affects a DN's decision boundary and no guidance regarding how to design a principled pruning technique. Using a recently developed spline interpretation of DNs, we develop new theory and visualization tools that provide new insights into how pruning DN nodes affects the decision boundary. We discover that a DN's spline mappings exhibit an early-bird (EB) phenomenon whereby the spline's partition converges at early training stages, bridging the recently developed max-affine spline theory and lottery ticket hypothesis of DNs. We leverage this new insight to develop a principled and efficient pruning strategy that focuses on a tiny fraction of DN nodes whose corresponding spline partition regions actually contribute to the final decision boundary. Extensive experiments on four networks and three datasets validate that our new spline-based DN pruning approach reduces training FLOPs by up to 3.5x while achieving similar or even better accuracy than state-of-the-art methods. All the codes will be released publicly upon acceptance.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "UF0JyvqqYHp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2134/Reviewer_2J1h"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper conducts a series of empirical studies, which aim at relating node/filter/channel pruning with the partition of input space by neural nets using spline operators as activation functions. In particular, the final decision boundary is only determined by a few splines defined by a few filters of the network. This is verified in the paper through case studies of both fully-connected networks and convolutional networks, which implies that pruning does not severely degrade the accuracy as long as the important splines are not removed by pruning. Another observation is the early-bird tickets: the paper argues that the important splines do not change too much after a few early epochs since the binary activation patterns of data converge rapidly, so prunning can be applied after only a few epochs. The paper then proposes a pruning strategy that sequentially finds the most similar filter pairs and removes them. In experiments, they show that the proposed prunning method is more efficient than some recent pruning methods to achieve similar accuracy with pruning ratio <=70% for CIFAR10/100 and <=50% for ImageNet. ", "review_text": "Please read the summary first. Here are detailed comments:\n\n(1) Section 3.1 shows a trend in case studies that the pruning tends to preserve subdivision splines useful to the decision boundary and remove the redundant ones. As long as the decision boundary is preserved, the performance will not degrade. This is interesting. However, the two toy datasets here are too synthetic and no quantitative analysis is presented, especially for more complicated datasets or neural nets. Moreover, given the previous success of pruning on DNNs with piecewise linear activation functions like ReLU, the phenomenon observed here is somewhat obvious. I expected to find some in-depth analysis of why the trained decision boundary only depends on a small subset of splines and why the applied pruning method here can preserve them, which could be an important novelty of this paper, but only the phenomenon is described. The splines also lack a formal math definition in terms of the network parameters/filters/units, which could be very helpful to understand their relationship in pruning.  \n\n(2) Section 3.2 claims that a novel metric based on those subdivision lines is developed to study the early-bird ticket phenomenon but the metric is simply the change of ReLU activation patterns between two consecutive epochs. It is obvious that if the activation patterns do not change too much, the decision boundary of the DNNs would not change too much, but this does not necessarily depend on the previous analysis of the spline partition of input space. The change of partition is only reflected via a toy example in Fig.4 without quantitative analysis and in-depth discussion, e.g., do all the splines or just the important ones (forming the boundary) not change since very early epochs, what happens to the partitioned regions that do not contain any training data, etc.\n\n(3) I also have a similar concern about Section 3.3. It proposes an intuitive pruning strategy, i.e., removing the redundant units that are too similar to each other in terms of cosine similarity of parameters and difference between bias terms. However, this heuristic does not necessarily depend on any previous observations of spline subdivisions and input-space partitions. To achieve a pruning ratio > 50%, this strategy needs to be applied sequentially and thus cannot leverage parallel computation as most existing pruning methods. In each round, its complexity is quadratic in the number of units since it needs to compute all pairwise similarities, which is also costly. Hence, I am worried about its efficiency in practice. \n\n(4) The experiments try to show the proposed strategy's advantage over existing pruning methods on efficiency (when achieving comparable accuracies). However, they did not try high enough pruning ratios. As reported, almost all the evaluated pruning ratios achieve similar accuracies, indicating that the pruned DNNs are still very redundant and the pruning does not enter a trade-off regime between accuracy and pruning ratio. In the current regime, \"better pruning strategies\" can not be well distinguished from \"worse pruning strategies\" and thus cannot provide strong evidence to show the advantage of the proposed method. \n\nAdditional questions:\n\n(1) Did you compare to any iterative pruning methods, which keep removing unimportant units/weights over epochs since the very beginning? They might be as efficient as or even more efficient than the early-bird metric used in this paper.\n\n(2) Did you consider the difference of layers on their contributions to the final decision boundary and their difference in the convergence of activation patterns? Will these differences be helpful to improve the pruning performance?\n\n(3) Instead of sequentially finding the redundant pairs of units, as proposed in Section 3.3, why not apply clustering based on the metric in Eq. (2)? It is not clear how to get the \"grouped partition trajectories\" without using any clustering-type method. \n\n(4) Though starting with max-affine spline DNs in Eq. (1), the paper only considers ReLU activations. Can you add empirical analysis and experiments for other activations belonging to this family?\n\n(5) The current related works lack discussion about the research area for linear regions of neural nets, which covers a great number of papers and is strongly related to the spline subdivisions in this paper and thus cannot be ignored.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper conducts a series of empirical studies, which aim at relating node/filter/channel pruning with the partition of input space by neural nets using spline operators as activation functions. In particular, the final decision boundary is only determined by a few splines defined by a few filters of the network. This is verified in the paper through case studies of both fully-connected networks and convolutional networks, which implies that pruning does not severely degrade the accuracy as long as the important splines are not removed by pruning. Another observation is the early-bird tickets: the paper argues that the important splines do not change too much after a few early epochs since the binary activation patterns of data converge rapidly, so prunning can be applied after only a few epochs. The paper then proposes a pruning strategy that sequentially finds the most similar filter pairs and removes them. In experiments, they show that the proposed prunning method is more efficient than some recent pruning methods to achieve similar accuracy with pruning ratio <=70% for CIFAR10/100 and <=50% for ImageNet. ", "main_review": "Please read the summary first. Here are detailed comments:\n\n(1) Section 3.1 shows a trend in case studies that the pruning tends to preserve subdivision splines useful to the decision boundary and remove the redundant ones. As long as the decision boundary is preserved, the performance will not degrade. This is interesting. However, the two toy datasets here are too synthetic and no quantitative analysis is presented, especially for more complicated datasets or neural nets. Moreover, given the previous success of pruning on DNNs with piecewise linear activation functions like ReLU, the phenomenon observed here is somewhat obvious. I expected to find some in-depth analysis of why the trained decision boundary only depends on a small subset of splines and why the applied pruning method here can preserve them, which could be an important novelty of this paper, but only the phenomenon is described. The splines also lack a formal math definition in terms of the network parameters/filters/units, which could be very helpful to understand their relationship in pruning.  \n\n(2) Section 3.2 claims that a novel metric based on those subdivision lines is developed to study the early-bird ticket phenomenon but the metric is simply the change of ReLU activation patterns between two consecutive epochs. It is obvious that if the activation patterns do not change too much, the decision boundary of the DNNs would not change too much, but this does not necessarily depend on the previous analysis of the spline partition of input space. The change of partition is only reflected via a toy example in Fig.4 without quantitative analysis and in-depth discussion, e.g., do all the splines or just the important ones (forming the boundary) not change since very early epochs, what happens to the partitioned regions that do not contain any training data, etc.\n\n(3) I also have a similar concern about Section 3.3. It proposes an intuitive pruning strategy, i.e., removing the redundant units that are too similar to each other in terms of cosine similarity of parameters and difference between bias terms. However, this heuristic does not necessarily depend on any previous observations of spline subdivisions and input-space partitions. To achieve a pruning ratio > 50%, this strategy needs to be applied sequentially and thus cannot leverage parallel computation as most existing pruning methods. In each round, its complexity is quadratic in the number of units since it needs to compute all pairwise similarities, which is also costly. Hence, I am worried about its efficiency in practice. \n\n(4) The experiments try to show the proposed strategy's advantage over existing pruning methods on efficiency (when achieving comparable accuracies). However, they did not try high enough pruning ratios. As reported, almost all the evaluated pruning ratios achieve similar accuracies, indicating that the pruned DNNs are still very redundant and the pruning does not enter a trade-off regime between accuracy and pruning ratio. In the current regime, \"better pruning strategies\" can not be well distinguished from \"worse pruning strategies\" and thus cannot provide strong evidence to show the advantage of the proposed method. \n\nAdditional questions:\n\n(1) Did you compare to any iterative pruning methods, which keep removing unimportant units/weights over epochs since the very beginning? They might be as efficient as or even more efficient than the early-bird metric used in this paper.\n\n(2) Did you consider the difference of layers on their contributions to the final decision boundary and their difference in the convergence of activation patterns? Will these differences be helpful to improve the pruning performance?\n\n(3) Instead of sequentially finding the redundant pairs of units, as proposed in Section 3.3, why not apply clustering based on the metric in Eq. (2)? It is not clear how to get the \"grouped partition trajectories\" without using any clustering-type method. \n\n(4) Though starting with max-affine spline DNs in Eq. (1), the paper only considers ReLU activations. Can you add empirical analysis and experiments for other activations belonging to this family?\n\n(5) The current related works lack discussion about the research area for linear regions of neural nets, which covers a great number of papers and is strongly related to the spline subdivisions in this paper and thus cannot be ignored.", "summary_of_the_review": "Although both the input space partition of neural nets with piecewise linear activations (sometimes called linear regions, convex polytopes, etc.) and node/structured pruning have been widely studied and well known in recent years, this paper provides an interesting and novel perspective to relate them, i.e., explaining why a high pruning ratio can still work, early-stopping for pruning, and pruning by removing redundant units. The case studies clearly explain the empirical observations and well motivate some conclusions. However, the main results are a little bit disappointing to me due to the lack of in-depth discussion of the phenomenons. Necessary math formulation of some important concepts (e.g., splines, grouped partition trajectories, etc) is missing. Both the early-bird metric and the pruning strategy lack a strong or insightful connection to the spline partition and decision boundary discussed in the first part of the paper. The experiments have not explored higher pruning ratio regimes, which are necessary since the reported accuracy does not change too much over the evaluated pruning ratios. \n\n-------------Update---------------\n\nThe authors address most of my major concerns in their new reported experiments and updated draft. Therefore, I raise my score to 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635926577133}, {"id": "tdNIcCkPPrz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2134/Reviewer_pmay"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper mainly proposes a new angle to understand the deep neural network pruning based on Spline theory and propose a new pruning algorithm approach. First, the author introduces the back ground of the Spline theory and current pruning methods. Second, the author analysis the different pruning methods from space partition perspective and introduce the proposed algorithm. Third, the author run experiments to evaluate the proposed algorithm in different aspects. ", "review_text": "Strength: \n1.the paper provides a new perspective on deep neural network pruning, \n2.the analysis through spline space partition is interesting. \n3.The paper has good written and visualization.\nWeakness: \n1.The author claim that ‘The observation consistently shows that only parts of subdivision splines are useful for decision boundary; and the goal of pruning is to remove those (redundant) subdivision splines and find winning tickets.’, however, in theoretical part, the author didn’t provide how the proposed algorithm in detail to remove the subdivision splines. Will the algorithm need extra computation cost for such space partition building? \n2. When the author introduces the proposed algorithm, the author didn’t analysis if such method has the same convergence guarantee as Lottery Ticket Hypothesis. If so, what is the bound of the error probability?\n3.In the experiment, the author didn’t consider Vision Transformer, which is an important SOTA model in image classification. And it is unsure if such technique is still working for larger image dataset such as ImageNet. Will the pruning strategy will be different in self attention layers?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper mainly proposes a new angle to understand the deep neural network pruning based on Spline theory and propose a new pruning algorithm approach. First, the author introduces the back ground of the Spline theory and current pruning methods. Second, the author analysis the different pruning methods from space partition perspective and introduce the proposed algorithm. Third, the author run experiments to evaluate the proposed algorithm in different aspects. ", "main_review": "Strength: \n1.the paper provides a new perspective on deep neural network pruning, \n2.the analysis through spline space partition is interesting. \n3.The paper has good written and visualization.\nWeakness: \n1.The author claim that ‘The observation consistently shows that only parts of subdivision splines are useful for decision boundary; and the goal of pruning is to remove those (redundant) subdivision splines and find winning tickets.’, however, in theoretical part, the author didn’t provide how the proposed algorithm in detail to remove the subdivision splines. Will the algorithm need extra computation cost for such space partition building? \n2. When the author introduces the proposed algorithm, the author didn’t analysis if such method has the same convergence guarantee as Lottery Ticket Hypothesis. If so, what is the bound of the error probability?\n3.In the experiment, the author didn’t consider Vision Transformer, which is an important SOTA model in image classification. And it is unsure if such technique is still working for larger image dataset such as ImageNet. Will the pruning strategy will be different in self attention layers?\n", "summary_of_the_review": "The idea and the observation is interesting, and most of the experiment results are promising. However, it is unclear how the author finds the subdivision spline to remove in experiment implementation. And in the theoretical part, I suggest the author to provide the convergence proof of proposed algorithm’s under Lottery Ticket hypothesis. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635899954231}, {"id": "KH7mhJIiBQH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2134/Reviewer_haRu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes to use the max-affine spline function to understand network pruning in the deep learning architecture. Specifically, it uses two examples (one in fully-connected NN and one in CNN) to show the splines visualization with different pruning rates. Inspired by the theory, it then proposes a policy that is based on calculating the cosine similarity between slope and biases to determine the importance. The experiment results show the proposed metric could achieve a similar or better accuracy with good energy efficiency in multiple datasets with both structured pruning and unstructured pruning.", "review_text": "Pros:\n1. The proposed idea to explain network pruning is quite interesting and could have a significant impact.\n2. The experiment results do show the proposed method could achieve similar or better accuracy with good energy efficiency. \n\nCons:\n1. While the paper claims it bridges the understanding between pruning and decision boundary, however, I find it not convincing. It is only through two simple examples to illustrate the relation. There lacks a systematic way how to derive which spline function could be affect the final decision boundary. \n2. Some of their main claims are not well-supported. Proposition 1 is supporting evidence to conduct their pruning however it does not clearly show why Proposition 1 is working and only gives a reference to the paper proposing the deep learning spline function. Also, the early bird detection is also through two simple trajectories which are not convincing to me either. \n3. To use the global spline pruning, it has to first use PCA to shrink the space. However, the dimension reduction method is crucial to calculating the cosine similarity score. It could be seen in the appendix the performance is fluctuating with a not very small margin, which makes the method not reliable.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to use the max-affine spline function to understand network pruning in the deep learning architecture. Specifically, it uses two examples (one in fully-connected NN and one in CNN) to show the splines visualization with different pruning rates. Inspired by the theory, it then proposes a policy that is based on calculating the cosine similarity between slope and biases to determine the importance. The experiment results show the proposed metric could achieve a similar or better accuracy with good energy efficiency in multiple datasets with both structured pruning and unstructured pruning.", "main_review": "Pros:\n1. The proposed idea to explain network pruning is quite interesting and could have a significant impact.\n2. The experiment results do show the proposed method could achieve similar or better accuracy with good energy efficiency. \n\nCons:\n1. While the paper claims it bridges the understanding between pruning and decision boundary, however, I find it not convincing. It is only through two simple examples to illustrate the relation. There lacks a systematic way how to derive which spline function could be affect the final decision boundary. \n2. Some of their main claims are not well-supported. Proposition 1 is supporting evidence to conduct their pruning however it does not clearly show why Proposition 1 is working and only gives a reference to the paper proposing the deep learning spline function. Also, the early bird detection is also through two simple trajectories which are not convincing to me either. \n3. To use the global spline pruning, it has to first use PCA to shrink the space. However, the dimension reduction method is crucial to calculating the cosine similarity score. It could be seen in the appendix the performance is fluctuating with a not very small margin, which makes the method not reliable.", "summary_of_the_review": "Although the paper has proposed a very interesting perspective on network pruning, it lacks supporting evidence on some of their main claims. Therefore, I vote for a borderline paper with a tendency towards reject. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635743206349}, {"id": "QG4vmcW5JrX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2134/Reviewer_yjUc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a novel methodology for deep network pruning from the perspective of the max-affine spline.\nThe key idea of the paper is to remove the redundant subdivision splines and find winning tickets.\nThis is an interesting and very nicely written paper that bridges the max-affine spline formulation and empirical pruning techniques.\nThe authors provide a sufficient and straightforward introduction of the behind motivation.\nBeyond these, the authors discuss robustness considerations and perform thorough experiments on various benchmarked models and datasets, showing the favorable performance of their method.", "review_text": "- The idea of pruning the deep network by removing the redundant subdivision splines is a very novel idea. What's more, the authors interpret this idea with many vivid figures which is very readable.\n\n- The authors first bridge the connection between deep network spline theory and deep network pruning. The authors provide new insights into how pruning deep network nodes affect the decision boundary. Furthermore, they leverage this new insight to develop a principled and efficient pruning strategy.\n\n- Training deep network is a remaining huge obstacle. The authors propose a new robust and efficient algorithm for deep network training by effective network pruning techniques, which outperforms state-of-the-art competitors.\n\n- There remains a lack of explicit understanding of network pruning impact on a deep network’s decision boundary. The authors illustrate the effectiveness of their method from many aspects.\n\n- I don't understand why the authors choose the pairwise redundancy measure in (2). They need to give readers some more motivation. Is there any other measure that may be better? Some simple analysis may help us recognize it intuitively. \n\n- In section 4.3, Global spline pruning, the big problem is the mismatch of the filter dimension in different layers impedes the cosine similarity calculation. The authors solve this problem by adopting PCA for reducing the feature dimensions to the same. Is there any reason to measure the similarity between different layers by PCA or FA?\nThey need to clarify it.\n\n- In practice, units with small enough but nonzero $N_\\rho^l(k,k')$ are also highly redundant and can be removed. How to decide the threshold? The authors need to give more details.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a novel methodology for deep network pruning from the perspective of the max-affine spline.\nThe key idea of the paper is to remove the redundant subdivision splines and find winning tickets.\nThis is an interesting and very nicely written paper that bridges the max-affine spline formulation and empirical pruning techniques.\nThe authors provide a sufficient and straightforward introduction of the behind motivation.\nBeyond these, the authors discuss robustness considerations and perform thorough experiments on various benchmarked models and datasets, showing the favorable performance of their method.", "main_review": "- The idea of pruning the deep network by removing the redundant subdivision splines is a very novel idea. What's more, the authors interpret this idea with many vivid figures which is very readable.\n\n- The authors first bridge the connection between deep network spline theory and deep network pruning. The authors provide new insights into how pruning deep network nodes affect the decision boundary. Furthermore, they leverage this new insight to develop a principled and efficient pruning strategy.\n\n- Training deep network is a remaining huge obstacle. The authors propose a new robust and efficient algorithm for deep network training by effective network pruning techniques, which outperforms state-of-the-art competitors.\n\n- There remains a lack of explicit understanding of network pruning impact on a deep network’s decision boundary. The authors illustrate the effectiveness of their method from many aspects.\n\n- I don't understand why the authors choose the pairwise redundancy measure in (2). They need to give readers some more motivation. Is there any other measure that may be better? Some simple analysis may help us recognize it intuitively. \n\n- In section 4.3, Global spline pruning, the big problem is the mismatch of the filter dimension in different layers impedes the cosine similarity calculation. The authors solve this problem by adopting PCA for reducing the feature dimensions to the same. Is there any reason to measure the similarity between different layers by PCA or FA?\nThey need to clarify it.\n\n- In practice, units with small enough but nonzero $N_\\rho^l(k,k')$ are also highly redundant and can be removed. How to decide the threshold? The authors need to give more details.\n", "summary_of_the_review": "The paper is well-written and the ideas are well-presented. However, some points listed above need to be clarified.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635728540766}], "openreview_url": "https://openreview.net/forum?id=7vXQJ2QW8hR", "arxiv_id": "2101.02338", "paper_pdf": "papers/7vXQJ2QW8hR.pdf", "paper_pdf_sha256": "c37f967fb01869cb33073086d253227ac408f24746f8219f15fb9e8c76c2ba6d", "paper_pdf_bytes": 2933925, "paper_pdf_source": "openreview", "code_url": "https://github.com/GATECH-EIC/Spline-EB", "code_repository": "GATECH-EIC/Spline-EB", "code_commit": "f26909ec902cbb752d1192ad6c7bcb8a3c835fb2", "code_archive": "repos/7vXQJ2QW8hR.zip", "code_archive_sha256": "e88575856ef856aab097fff70b6bb4b14fc673cdde371d7e94d4d81c9b30ab51", "code_archive_bytes": 105140, "code_file_count": 28, "code_extensions": {".py": 28}, "github_disk_usage_kb": 89, "github_languages": {"Python": 183056}, "github_archived": false, "github_pushed_at": "2022-11-12T03:22:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/max-affine-spline-insights-into-deep-network-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RgDq8-AwvtN", "year": 2021, "status": "rejected", "title": "Model-Based Robust Deep Learning: Generalizing to Natural, Out-of-Distribution Data", "authors": ["Alexander Robey", "Hamed Hassani", "George J. Pappas"], "authorids": ["~Alexander_Robey1", "~Hamed_Hassani2", "~George_J._Pappas1"], "authors_source": "OpenReview API", "abstract": "While deep learning (DL) has resulted in major breakthroughs in many applications, the frameworks commonly used in DL remain fragile to seemingly innocuous changes in the data. In response, adversarial training has emerged as a principled approach for improving the robustness of DL against norm-bounded perturbations.  Despite this progress, DL is also known to be fragile to unbounded shifts in the data distribution due to many forms of natural variation, including changes in weather or lighting in images.  However, there are remarkably few techniques that can address robustness to natural, out-of-distribution shifts in the data distribution in a general context.  To address this gap, we propose a paradigm shift from perturbation-based adversarial robustness to model-based robust deep learning.  Critical to our paradigm is to obtain models of natural variation, which vary data over a range of natural conditions.  Then by exploiting these models, we develop three novel model-based robust training algorithms that improve the robustness of DL with respect to natural variation. Our extensive experiments show that across a variety of natural conditions in twelve distinct datasets, classifiers trained with our algorithms significantly outperform classifiers trained via ERM, adversarial training, and domain adaptation techniques.  Specifically, when training on ImageNet and testing on various subsets of ImageNet-c, our algorithms improve over baseline methods by up to 30 percentage points in top-1 accuracy.  Further, we show that our methods provide robustness (1) against natural, out-of-distribution data, (2) against multiple simultaneous distributional shifts, and (3) to domains entirely unseen during training.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "uN7_9dG4MVD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper359/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper looks at the idea of building models of natural variation of an input and then using these models to develop robust training algorithms that are less susceptible to outliers ( testable for worst-cases via adversarial attacks ). \n\nStrength:\n\n+ The paper addresses a very important topic of adversarial robustness of DNN models and is accompanied by diligent evaluation over different datasets. \n\nWeakness:\n\n- The paper adopts an approach which is very reminiscent of Ajil Jalal, Andrew Ilyas, Constantinos Daskalakis, and Alexandros G Dimakis. The robust manifold defense: Adversarial training using generative models. arXiv preprint arXiv:1712.09196, 2017. The key difference seems to be adoption of the idea of using auxiliary transformations (called natural perturbations) which is also very well-studied in literature, for e.g. see https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhang_Auxiliary_Training_Towards_Accurate_and_Robust_Models_CVPR_2020_paper.pdf  So, the approach presented here is quite incremental for a premier venue such as ICLR.\n\n- The attack in Jalal et. al. paper can be launched against this approach once the attacker also access to this generative/natural model (which would be easy to build for an attacker)\n\n- The reviewer will strongly recommend reviewing the advices in https://arxiv.org/abs/1902.06705 for writing papers on defense and how to self-evaluate its robustness by suitably designing the attacker. If the defense approach uses some background/auxiliary knowledge, one must consider the attacker with this knowledge if it is accessible to the attacker. \n\nQuestions to authors:\n\n- Can authors explain why an attacker can't build similar natural model to defeat the proposed defense? \n- Any clarification on incremental novelty from  Jalal et. al. and Zhang et. al. (and references therein) would be also useful. \n\n\nAfter author's rebuttal:\n\n\"The other paper that the review points to ([4]) addresses a similar setting to our paper, but the approach is completely different. ... Moreover, [4] was published within three months of the ICLR submission deadline, meaning that it is essentially concurrent with our work.\"\n\nYes, the reviewer concurs that a work published so close to deadline should be treated as concurrent and will raise the score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear novelty and very similar to a broken robustness method ( defense )", "review": "The paper looks at the idea of building models of natural variation of an input and then using these models to develop robust training algorithms that are less susceptible to outliers ( testable for worst-cases via adversarial attacks ). \n\nStrength:\n\n+ The paper addresses a very important topic of adversarial robustness of DNN models and is accompanied by diligent evaluation over different datasets. \n\nWeakness:\n\n- The paper adopts an approach which is very reminiscent of Ajil Jalal, Andrew Ilyas, Constantinos Daskalakis, and Alexandros G Dimakis. The robust manifold defense: Adversarial training using generative models. arXiv preprint arXiv:1712.09196, 2017. The key difference seems to be adoption of the idea of using auxiliary transformations (called natural perturbations) which is also very well-studied in literature, for e.g. see https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhang_Auxiliary_Training_Towards_Accurate_and_Robust_Models_CVPR_2020_paper.pdf  So, the approach presented here is quite incremental for a premier venue such as ICLR.\n\n- The attack in Jalal et. al. paper can be launched against this approach once the attacker also access to this generative/natural model (which would be easy to build for an attacker)\n\n- The reviewer will strongly recommend reviewing the advices in https://arxiv.org/abs/1902.06705 for writing papers on defense and how to self-evaluate its robustness by suitably designing the attacker. If the defense approach uses some background/auxiliary knowledge, one must consider the attacker with this knowledge if it is accessible to the attacker. \n\nQuestions to authors:\n\n- Can authors explain why an attacker can't build similar natural model to defeat the proposed defense? \n- Any clarification on incremental novelty from  Jalal et. al. and Zhang et. al. (and references therein) would be also useful. \n\n\nAfter author's rebuttal:\n\n\"The other paper that the review points to ([4]) addresses a similar setting to our paper, but the approach is completely different. ... Moreover, [4] was published within three months of the ICLR submission deadline, meaning that it is essentially concurrent with our work.\"\n\nYes, the reviewer concurs that a work published so close to deadline should be treated as concurrent and will raise the score.", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1604651394394}, {"id": "4kddaVVuOQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper359/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a model-based framework for improving the robustness of image classifiers to average-case corruptions of varying severity. The proposed framework can be thought of as adversarial training where the perturbation is replaced by a function that transforms the image according to a specific corruption. A nuisance parameter controls the instantiation and severity of the corruption that is applied to the input. The paper compares baselines to different versions of this general model-based framework with experiments on several datasets.\n\nPros:\n- Focusing on average-case corruptions is an important and underexplored problem\n- Using image translation to learn the corruptions is an interesting proposal\n\nCons:\n- There isn't much novelty in the proposed technique. Adversarial training with 'natural' corruptions has been done several times before.\n- There isn't any discussion of how the nuisance parameter is used. I suspect it isn't used for ImageNet-C experiments, though I may be wrong. In this case, the method becomes data augmentation with image translation, which is not very novel either.\n- It sounds like the ImageNet-C experiments use the corruptions to train the image translation network. Even though these experiments use held-out images, the corruption types themselves are not being held out, which goes against the recommended methodology for that dataset.\n\n\nTypo: \"ImageNet-c\" should be \"ImageNet-C\"\n\n___________________________________________________________________________________\n\nUpdate after author feedback:\n\nI thank the authors for improving my understanding of the paper. I feel better about it after reading it again. One of the most interesting findings--that a translation network trained on one domain/set of classes generalizes to another--needs more discussion. Using this phenomenon to improve robustness is a good idea, and the MRT/MDA/MAT methods explored in this work are nice choices for this investigation. However, I agree with the other reviewers that it would be nice to include more baselines from other work where appropriate.\n\nThe results in Table 1 are also very interesting and deserve more discussion--perhaps an analysis of whether a translation network trained on weak corruptions can generalize to create something akin to the ground-truth stronger corruptions, as the results in the table for MRT imply.\n\nOverall, I think this paper has some strong experiments and investigates a good idea, but the claims of a \"paradigm shift\" are overly grandiose, and some of the most interesting experiments could use more analysis. Additionally, the writing clarity and presentation could be improved. My initial understanding of the paper was flawed in some places, so I'll raise my score to 5.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #3", "review": "This paper proposes a model-based framework for improving the robustness of image classifiers to average-case corruptions of varying severity. The proposed framework can be thought of as adversarial training where the perturbation is replaced by a function that transforms the image according to a specific corruption. A nuisance parameter controls the instantiation and severity of the corruption that is applied to the input. The paper compares baselines to different versions of this general model-based framework with experiments on several datasets.\n\nPros:\n- Focusing on average-case corruptions is an important and underexplored problem\n- Using image translation to learn the corruptions is an interesting proposal\n\nCons:\n- There isn't much novelty in the proposed technique. Adversarial training with 'natural' corruptions has been done several times before.\n- There isn't any discussion of how the nuisance parameter is used. I suspect it isn't used for ImageNet-C experiments, though I may be wrong. In this case, the method becomes data augmentation with image translation, which is not very novel either.\n- It sounds like the ImageNet-C experiments use the corruptions to train the image translation network. Even though these experiments use held-out images, the corruption types themselves are not being held out, which goes against the recommended methodology for that dataset.\n\n\nTypo: \"ImageNet-c\" should be \"ImageNet-C\"\n\n___________________________________________________________________________________\n\nUpdate after author feedback:\n\nI thank the authors for improving my understanding of the paper. I feel better about it after reading it again. One of the most interesting findings--that a translation network trained on one domain/set of classes generalizes to another--needs more discussion. Using this phenomenon to improve robustness is a good idea, and the MRT/MDA/MAT methods explored in this work are nice choices for this investigation. However, I agree with the other reviewers that it would be nice to include more baselines from other work where appropriate.\n\nThe results in Table 1 are also very interesting and deserve more discussion--perhaps an analysis of whether a translation network trained on weak corruptions can generalize to create something akin to the ground-truth stronger corruptions, as the results in the table for MRT imply.\n\nOverall, I think this paper has some strong experiments and investigates a good idea, but the claims of a \"paradigm shift\" are overly grandiose, and some of the most interesting experiments could use more analysis. Additionally, the writing clarity and presentation could be improved. My initial understanding of the paper was flawed in some places, so I'll raise my score to 5.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604558983077}, {"id": "oezlT96FUV8", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper359/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper extends current adversarial learning approaches beyond imperceptible L_p norm perturbations. The proposed approach can handle many models of natural variation, such as a change in brightness. The main idea behind the approach is to use unsupervised approaches such as GANs to model the natural variation. Given this model of natural variation, the paper replaces the adversarial learning objective of finding the worst example in an L_p norm ball around a data point to finding the worst example based on the model of natural variation. This is expensive, so the paper also proposes more computationally efficient approaches based on data augmentation. The experimental results demonstrate that the proposed approach performs well on a variety of tasks.\n\nStrengths:\n\n1. The question of training robust neural networks is clearly timely and well-motivated. A valid raised criticism of the current adversarial examples literature is that it is too tied to imperceptible or L_p norm perturbations, and this paper tries to go beyond them.\n2. There is a bit of a lack of relevant methods to compare against, but the experiments still show that the proposed approach does well. I also appreciated the comparisons to domain adaption, and the multiple experimental setups generally.\n\nWeaknesses:\n\n1. Though the paper suggests that the approach is a “paradigm shift” in robust deep learning, it appears to be more or less an extension of the current adversarial examples literature. The adversarial training approach has been quite successful in defending against a fixed adversary model, from various L_p norms to geometric transformations. This paper is essentially defining a new adversary model, given by natural variations. In fact it is simpler than this, since the goal is to only do well on \"random variations\" in the predefined test set, i.e. there is no adversary which will try and attack the network by using the class of variations. There is a novelty here though that the variation model is learnt rather than pre-specified. But given that learning transformations such as change in brightness etc. is quite straightforward for unsupervised models such as GANs, this is not too much to ask for. Therefore, I find it a bit unsurprising that the proposed method works well, given the success of adversarial training and GANs.\n2. Related to the above, in my view one main drawback of the current adversarial learning literature is that it is not difficult to get robustness to one pre-specified model of perturbation/variation, but in reality the space of possible variations is quite large. It is good that the proposed approach does seem to work on two simultaneous shifts, but the approach still seems restricted and not a substantial change in the current understanding. Note that some of the compared approaches are different in this regard: AugMix does not tailor the model to a particular variation (it explicitly excludes operations which overlap with ImageNet-C corruptions) and domain adaptation can handle quite large changes in the data distribution (which are not just transformations of the original data points). \n3. Perhaps less importantly, in the experiments the proposed MDA approach based on data augmentation does almost as well as the other ones, and data augmentation using unsupervised models is already studied in the literature.\n\nOverall, this is not a bad paper and I am not completely opposed to acceptance, but I am not sure I can argue for it.\n\nOther less important comments:\n\n1. Since there is a lack of baselines to compare against, the paper could benefit from a few ablation studies. For example, is there any difference on using other unsupervised approaches? \n2. I think it would be better to have a concise list of a few major contributions in Section 1, rather than the current rather long list.\n3. It would be nice to have representative images of the models of variation, perhaps in an appendix.  \n\n------Updates after author response------\n\nI thank the authors for the response and it helps clarify some points. However, I am still not unconvinced that the paper is a significant departure from current work on adversarial robustness (see weakness 1 and 2 above). I think it would be much more interesting if the approach could yield robustness against a wider/different class of shifts compared to what it was adversarially trained against. Therefore, I am keeping my score at 5 and will not advocate for acceptance, though I am not completely opposed to it.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Extends adversarial learning to natural variations using GANs, but appears to have somewhat limited contribution", "review": "The paper extends current adversarial learning approaches beyond imperceptible L_p norm perturbations. The proposed approach can handle many models of natural variation, such as a change in brightness. The main idea behind the approach is to use unsupervised approaches such as GANs to model the natural variation. Given this model of natural variation, the paper replaces the adversarial learning objective of finding the worst example in an L_p norm ball around a data point to finding the worst example based on the model of natural variation. This is expensive, so the paper also proposes more computationally efficient approaches based on data augmentation. The experimental results demonstrate that the proposed approach performs well on a variety of tasks.\n\nStrengths:\n\n1. The question of training robust neural networks is clearly timely and well-motivated. A valid raised criticism of the current adversarial examples literature is that it is too tied to imperceptible or L_p norm perturbations, and this paper tries to go beyond them.\n2. There is a bit of a lack of relevant methods to compare against, but the experiments still show that the proposed approach does well. I also appreciated the comparisons to domain adaption, and the multiple experimental setups generally.\n\nWeaknesses:\n\n1. Though the paper suggests that the approach is a “paradigm shift” in robust deep learning, it appears to be more or less an extension of the current adversarial examples literature. The adversarial training approach has been quite successful in defending against a fixed adversary model, from various L_p norms to geometric transformations. This paper is essentially defining a new adversary model, given by natural variations. In fact it is simpler than this, since the goal is to only do well on \"random variations\" in the predefined test set, i.e. there is no adversary which will try and attack the network by using the class of variations. There is a novelty here though that the variation model is learnt rather than pre-specified. But given that learning transformations such as change in brightness etc. is quite straightforward for unsupervised models such as GANs, this is not too much to ask for. Therefore, I find it a bit unsurprising that the proposed method works well, given the success of adversarial training and GANs.\n2. Related to the above, in my view one main drawback of the current adversarial learning literature is that it is not difficult to get robustness to one pre-specified model of perturbation/variation, but in reality the space of possible variations is quite large. It is good that the proposed approach does seem to work on two simultaneous shifts, but the approach still seems restricted and not a substantial change in the current understanding. Note that some of the compared approaches are different in this regard: AugMix does not tailor the model to a particular variation (it explicitly excludes operations which overlap with ImageNet-C corruptions) and domain adaptation can handle quite large changes in the data distribution (which are not just transformations of the original data points). \n3. Perhaps less importantly, in the experiments the proposed MDA approach based on data augmentation does almost as well as the other ones, and data augmentation using unsupervised models is already studied in the literature.\n\nOverall, this is not a bad paper and I am not completely opposed to acceptance, but I am not sure I can argue for it.\n\nOther less important comments:\n\n1. Since there is a lack of baselines to compare against, the paper could benefit from a few ablation studies. For example, is there any difference on using other unsupervised approaches? \n2. I think it would be better to have a concise list of a few major contributions in Section 1, rather than the current rather long list.\n3. It would be nice to have representative images of the models of variation, perhaps in an appendix.  \n\n------Updates after author response------\n\nI thank the authors for the response and it helps clarify some points. However, I am still not unconvinced that the paper is a significant departure from current work on adversarial robustness (see weakness 1 and 2 above). I think it would be much more interesting if the approach could yield robustness against a wider/different class of shifts compared to what it was adversarially trained against. Therefore, I am keeping my score at 5 and will not advocate for acceptance, though I am not completely opposed to it.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603928571713}, {"id": "ER1aRZ474Jg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper359/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary\nThis paper proposes a “paradigm shift” for augmenting datasets when training CNN-based image classifiers. On one side, traditional augmentations include blur, Gaussian noise, color distortions. On the other side, methods like adversarial training consider augmentations under norm bounds in the image space. The proposed method, instead, uses models of natural variation to augment the images. Increased out-of-domain accuracy is shown on ImageNet-C and several slices of the CURE-TSR dataset. \n\nCentral question: How to shift from perturbation-based robustness analysis to model-based robustness analysis?\n\n# Strong and Weak points\n\n## Strong points \n\n  * This paper introduces a “paradigm shift” from perturbation based to model-based augmentation, when training CNN based image classifiers. The existing perturbations only reach so far, and using models of natural variation provides a new avenue for augmenting data sets. \n  * The method is presented in clear pseudo code, and detailed in the 36 page counting appendix. \n\n## Weak points\n\n  * The paper claims a “paradigm shift”, but the proposed method of augmenting data has been used since 2012,  [1], and is used ever since [2]. Tables 2 and 3 provide key results for this method, but both were obtained with Model-based Data Augmentation, which is the common method before the \"paradigm shift\".\n  * Results outlined in Section 5.2 and Table 2 concerns scenarios where the test-time perturbation is known at training time. Table 2 then compares MDA to ERM using this a-priori knowledge, and shows improved accuracy. However, the comparison is unbalanced, as ERM does not benefit from the prior knowledge. If Table 2 would be about comparing methods under the assumption that the test-time perturbation would be known at training time, then the ERM should be trained with the same data augmentation.\n  * The paper misses comparisons against previous published work. Although Table 2 and 4 compare against another method, these results were not obtained by the respective authors, thus introducing a bias. The related work, Section 6, cites as many as 9 papers that addressed the same problem as this paper. The argument for the proposed method could be stronger when comparing to results from any of these papers. \nMoreover, Table 3 reports results on ImageNet-C, which multiple other research papers have published results on. See for example, [3] and [4]. Likewise, Table 4 reports results on MNIST-C, which have been reported in [5] and [6]. Omitting those comparisons hampers the argument for the proposed method. \n\n[1] Krizhevsky, 2012\n[2] Szegedy, 2015\n[3] https://paperswithcode.com/sota/domain-generalization-on-imagenet-c\n[4] https://arxiv.org/pdf/1903.12261.pdf \n[5] https://proceedings.icml.cc/paper/2020/file/20546457187cf3d52ea86538403e47cc-Paper.pdf \n[6] https://arxiv.org/pdf/1906.02337.pdf \n\n# Statement\n\nRecommendation: Reject\nReason:\nAlthough this paper provides four tables of extensive evaluation of the method, none of the numbers were obtained from previous literature. The related work section cites 9 papers on the same problem, but none were compared with in the results section. (See concrete details above)\n\n# Clarifying questions\n\n  * How to quantify the increase in computational complexity for the inner maximization in Equation (2)? As I understand, even evaluating one point requires a forward pass of the variation-model and the classifier-model. In case of iterative optimization, this would also require two backward passes(see line 5 in Algorithm (1)). \nIf the proposed method requires significantly more compute, the comparison should be made with a method that has access to the same amount of compute (for example a larger model, or longer trained model).\n  * Line 6 of Algorithm (1). I miss the definition of set $\\Delta$ and the projection operator in the text nor the appendix. Please explain?\n  * Section 5.2 speaks of ImageNet “classes”. I am confused if these “classes” are referring to the ImageNet classes (1000 in total) or the severity levels in ImageNet-C (75 in total). Could you please clarify?\n  Moreover, this argument implicitly assumes that the test-time perturbation is known at training-time. Could you name a scenario where this assumption is justified?\n  * To what extent is the accuracy on the clean data preserved by the proposed method? Table 3, for example, shows an increase in top-1 accuracy on ImageNet data under perturbation. However, I wonder if this model preserves accuracy on the clean data. We know from existing literature [1] that a trade off exists for adversarial training between accuracy on perturbed and clean data. Providing the clean accuracy would provide evidence that this method compares favourably on this trade off. \n\n[1] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" arXiv preprint arXiv:1901.08573 (2019).\n\n# Minor feedback\n\nThis minor feedback is not part of the assessment.\n\n  * “arguably more common” -> “arguably more important”\n  * “x with a corresponding label y” -> “x with a corresponding prediction y”\n  * Abbreviation “MBRDL” appears in Section 5 without any explanation or definition. \n  * Abbreviation “CURE-TSR” appears in Section 5.0 without any explanation or definition.\n  * The citation for CURE in Section 5.1 is wrong. The bibliography refers to “Temel, Dogancan, Min-Hung Chen, and Ghassan AlRegib. \"Traffic sign detection under challenging conditions: A deeper look into performance variations and spectral characteristics.\" IEEE Transactions on Intelligent Transportation Systems (2019).”, but the CURE TSR dataset was published in “Temel, Dogancan, et al. \"CURE-TSR: Challenging unreal and real environments for traffic sign recognition.\" arXiv preprint arXiv:1712.02463 (2017).”\n  * Caption of Table 4: “ Then, we use this model to perform model-based training on a new dataset D1” -> “ Then, we use this model to perform model-based testing on a new dataset D2”? \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of \"Model-Based Robust Deep Learning: Generalizing to Natural, Out-of-Distribution Data\"", "review": "# Summary\nThis paper proposes a “paradigm shift” for augmenting datasets when training CNN-based image classifiers. On one side, traditional augmentations include blur, Gaussian noise, color distortions. On the other side, methods like adversarial training consider augmentations under norm bounds in the image space. The proposed method, instead, uses models of natural variation to augment the images. Increased out-of-domain accuracy is shown on ImageNet-C and several slices of the CURE-TSR dataset. \n\nCentral question: How to shift from perturbation-based robustness analysis to model-based robustness analysis?\n\n# Strong and Weak points\n\n## Strong points \n\n  * This paper introduces a “paradigm shift” from perturbation based to model-based augmentation, when training CNN based image classifiers. The existing perturbations only reach so far, and using models of natural variation provides a new avenue for augmenting data sets. \n  * The method is presented in clear pseudo code, and detailed in the 36 page counting appendix. \n\n## Weak points\n\n  * The paper claims a “paradigm shift”, but the proposed method of augmenting data has been used since 2012,  [1], and is used ever since [2]. Tables 2 and 3 provide key results for this method, but both were obtained with Model-based Data Augmentation, which is the common method before the \"paradigm shift\".\n  * Results outlined in Section 5.2 and Table 2 concerns scenarios where the test-time perturbation is known at training time. Table 2 then compares MDA to ERM using this a-priori knowledge, and shows improved accuracy. However, the comparison is unbalanced, as ERM does not benefit from the prior knowledge. If Table 2 would be about comparing methods under the assumption that the test-time perturbation would be known at training time, then the ERM should be trained with the same data augmentation.\n  * The paper misses comparisons against previous published work. Although Table 2 and 4 compare against another method, these results were not obtained by the respective authors, thus introducing a bias. The related work, Section 6, cites as many as 9 papers that addressed the same problem as this paper. The argument for the proposed method could be stronger when comparing to results from any of these papers. \nMoreover, Table 3 reports results on ImageNet-C, which multiple other research papers have published results on. See for example, [3] and [4]. Likewise, Table 4 reports results on MNIST-C, which have been reported in [5] and [6]. Omitting those comparisons hampers the argument for the proposed method. \n\n[1] Krizhevsky, 2012\n[2] Szegedy, 2015\n[3] https://paperswithcode.com/sota/domain-generalization-on-imagenet-c\n[4] https://arxiv.org/pdf/1903.12261.pdf \n[5] https://proceedings.icml.cc/paper/2020/file/20546457187cf3d52ea86538403e47cc-Paper.pdf \n[6] https://arxiv.org/pdf/1906.02337.pdf \n\n# Statement\n\nRecommendation: Reject\nReason:\nAlthough this paper provides four tables of extensive evaluation of the method, none of the numbers were obtained from previous literature. The related work section cites 9 papers on the same problem, but none were compared with in the results section. (See concrete details above)\n\n# Clarifying questions\n\n  * How to quantify the increase in computational complexity for the inner maximization in Equation (2)? As I understand, even evaluating one point requires a forward pass of the variation-model and the classifier-model. In case of iterative optimization, this would also require two backward passes(see line 5 in Algorithm (1)). \nIf the proposed method requires significantly more compute, the comparison should be made with a method that has access to the same amount of compute (for example a larger model, or longer trained model).\n  * Line 6 of Algorithm (1). I miss the definition of set $\\Delta$ and the projection operator in the text nor the appendix. Please explain?\n  * Section 5.2 speaks of ImageNet “classes”. I am confused if these “classes” are referring to the ImageNet classes (1000 in total) or the severity levels in ImageNet-C (75 in total). Could you please clarify?\n  Moreover, this argument implicitly assumes that the test-time perturbation is known at training-time. Could you name a scenario where this assumption is justified?\n  * To what extent is the accuracy on the clean data preserved by the proposed method? Table 3, for example, shows an increase in top-1 accuracy on ImageNet data under perturbation. However, I wonder if this model preserves accuracy on the clean data. We know from existing literature [1] that a trade off exists for adversarial training between accuracy on perturbed and clean data. Providing the clean accuracy would provide evidence that this method compares favourably on this trade off. \n\n[1] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" arXiv preprint arXiv:1901.08573 (2019).\n\n# Minor feedback\n\nThis minor feedback is not part of the assessment.\n\n  * “arguably more common” -> “arguably more important”\n  * “x with a corresponding label y” -> “x with a corresponding prediction y”\n  * Abbreviation “MBRDL” appears in Section 5 without any explanation or definition. \n  * Abbreviation “CURE-TSR” appears in Section 5.0 without any explanation or definition.\n  * The citation for CURE in Section 5.1 is wrong. The bibliography refers to “Temel, Dogancan, Min-Hung Chen, and Ghassan AlRegib. \"Traffic sign detection under challenging conditions: A deeper look into performance variations and spectral characteristics.\" IEEE Transactions on Intelligent Transportation Systems (2019).”, but the CURE TSR dataset was published in “Temel, Dogancan, et al. \"CURE-TSR: Challenging unreal and real environments for traffic sign recognition.\" arXiv preprint arXiv:1712.02463 (2017).”\n  * Caption of Table 4: “ Then, we use this model to perform model-based training on a new dataset D1” -> “ Then, we use this model to perform model-based testing on a new dataset D2”? \n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603881189184}], "openreview_url": "https://openreview.net/forum?id=RgDq8-AwvtN", "arxiv_id": "2005.10247", "paper_pdf": "papers/RgDq8-AwvtN.pdf", "paper_pdf_sha256": "7446633b0f2e826d700e4ec66f4e3c49730bf220e37cd00790e6545678f86f96", "paper_pdf_bytes": 3137186, "paper_pdf_source": "openreview", "code_url": "https://github.com/arobey1/mbrdl", "code_repository": "arobey1/mbrdl", "code_commit": "f49bd0514850b109689a371334f585dc85d172a3", "code_archive": "repos/RgDq8-AwvtN.zip", "code_archive_sha256": "571fc65ef1ef57fd6a65adf6a858d5e424f5621197af99143f73260dab805925", "code_archive_bytes": 55987, "code_file_count": 28, "code_extensions": {".py": 24, ".sh": 4}, "github_disk_usage_kb": 113, "github_languages": {"Python": 139936, "Shell": 5775}, "github_archived": false, "github_pushed_at": "2021-01-22T08:14:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/model-based-robust-deep-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1x0CnEtvB", "year": 2020, "status": "rejected", "title": "AutoGrow: Automatic Layer Growing in Deep Convolutional Networks", "authors": ["Wei Wen", "Feng Yan", "Hai Li"], "authorids": ["wei.wen@duke.edu", "fyan@unr.edu", "hai.li@duke.edu"], "authors_source": "OpenReview API", "abstract": "Depth is a key component of Deep Neural Networks (DNNs), however, designing depth is heuristic and requires many human efforts. We propose AutoGrow to automate depth discovery in DNNs: starting from a shallow seed architecture, AutoGrow grows new layers if the growth improves the accuracy; otherwise, stops growing and thus discovers the depth. We propose robust growing and stopping policies to generalize to different network architectures and datasets. Our experiments show that by applying the same policy to different network architectures, AutoGrow can always discover near-optimal depth on various datasets of MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100 and ImageNet. For example, in terms of accuracy-computation trade-off, AutoGrow discovers a better depth combination in ResNets than human experts. Our AutoGrow is efficient. It discovers depth within similar time of training a single DNN.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HkgRdsvQqr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper284/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a meta-learning algorithm to automatically detemine the depth of neural network through a policy to add depth if this bring improvement on accuracy.\n\nI have conserved opinion based on the technique being used here is extremely simple, basically is an implementation of naive greedy algorithm in such a scenario, which implies the problem may not be intrinsically hard, or even useful. The paper consists of detailed narrative about how these procedure are conducted, but still, it is really hard for me to find the true merit to appreciate, and why this brings a nontrivial and usefull contribution. The tables, visualization figures also didnot imply too much about whether this is more than overfitting on previous works with hand-chosen depth. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper presents a meta-learning algorithm to automatically detemine the depth of neural network through a policy to add depth if this bring improvement on accuracy.\n\nI have conserved opinion based on the technique being used here is extremely simple, basically is an implementation of naive greedy algorithm in such a scenario, which implies the problem may not be intrinsically hard, or even useful. The paper consists of detailed narrative about how these procedure are conducted, but still, it is really hard for me to find the true merit to appreciate, and why this brings a nontrivial and usefull contribution. The tables, visualization figures also didnot imply too much about whether this is more than overfitting on previous works with hand-chosen depth. "}, "tcdate": 1572203382138}, {"id": "SkeAVHaRYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper284/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper's contribution is a method for automatically growing the depth of a neural network during training. It compares several heuristics that may be used to successfully achieve this goal and identifies a set of choices that work well together on multiple datasets.\n\nThe paper focuses on CNNs that conform to a popular design pattern where the network is organized into a series of sub-networks, each consisting of a series of sub-modules (sometimes called blocks) operating at the same resolution. To be precise, the proposed method aims to learn the length of each series of sub-modules. A main contribution of the paper is the demonstration that it is not necessary to train a network until convergence before adding new sub-modules as proposed in past work. Instead, it is better to grow the network after training for a short while.\n\nMy current decision for this paper is a weak rejection due to the points below. However, I am open to revising my opinion if these points are addressed satisfactorily.\n\n- The growing strategy identified in the paper as a superior alternative seems to be already known and used, at least in the speech recognition community. Seide et al. (2011) called it Discriminative Pre-training, and showed that it outperforms greedy layer-wise pretraining and DBN pre-training. Zeyer et al. (2017) reported that a similar method also enables the training of very deep LSTM networks which is otherwise notoriously hard. In general, the existence of prior work with the same ideas does not preclude acceptance, but the existence of this work needs to be clearly stated early on and the additional value of the current study sufficiently clarified.\n\n- I find it strange that the final networks found by the proposed method usually have the same/similar number of sub-modules per sub-network (Tables 4,5,6) on multiple datasets. The only exceptions appear to be Basic4ResNet/CIFAR100 in Table 6 and about 50% of ImageNet results in Table 7. This regularity suggests that either A) the proposed algorithm prefers to set same number of sub-modules per sub-network due to its design, or B) datasets except ImageNet have an inherent shared property that produces this result. Since option A suggests a bias in the algorithm, this peculiarity of the results needs to be investigated or explained further.\n\n- Figure 3 constitutes the main evidence that Autogrow finds approximately optimal depths as compared to manual searching, but it is not clear how the plot for baselines is obtained. For any given parameter budget, there are multiple baseline networks possible since the sub-networks can have different number of sub-modules (see previous point). This does not appear to be accounted for in Figure 3. Further, when dealing with CNNs, it would be more useful to have computation budget on the x-axis instead of the parameter budget. This would better account for the difference between increasing depth in an earlier sub-network vs. a later one.\n\n- The reported results appear to be for single trials throughout the paper. This does not seem sufficient especially for results in Tables 2 and 3 where many differences are rather small, and so drawing conclusions from these tables would be unscientific.\n\nReferences:\n\nSeide, Frank, et al. \"Feature engineering in context-dependent deep neural networks for conversational speech transcription.\" 2011 IEEE Workshop on Automatic Speech Recognition & Understanding. IEEE, 2011. https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/FeatureEngineeringInCD-DNN-ASRU2011-pub.pdf\n\nZeyer, Albert, et al. \"A comprehensive study of deep bidirectional LSTM RNNs for acoustic modeling in speech recognition.\" 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017. https://arxiv.org/abs/1606.06871\n\nUpdate after rebuttal\n-----------------------------\nI'm sympathetic to the unfortunate situation that the authors are in, since the underlying growing strategy has already been covered by prior work. As I mentioned earlier, a sufficient rewrite of the paper can clearly state what has been done already so as not to take credit from the earlier authors. A revised version of the paper has not been uploaded; I suggest that the authors do so for the future. \n\nI agree that the focus of this paper is learning the 'optimal' depth by using the growing strategy. But I am not convinced that the technique indeed finds optimal depths based on the regularity of the sub-network depths mentioned in my review. The rebuttal suggests reasons for the obtained regularity, but does not prove that these regular structures obtained are indeed optimal and not an artifact of the algorithm itself. The baselines are also using the same regular architectures, which distorts the overall picture because it is possible that a non-regular architecture provides a better trade-off.\n\nWhile my rating doesn't change, I do think that the work is in an interesting direction. My final suggestions for the future are:\n- Investigate where non-regular architectures (unequal sub-network depths) are in the trade-off between accuracy, flops and parameters.\n- Investigate whether the proposed algorithm can be modified to easily find non-regular architectures if they can yield equally good performance as regular ones at similar or lower cost.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "This paper's contribution is a method for automatically growing the depth of a neural network during training. It compares several heuristics that may be used to successfully achieve this goal and identifies a set of choices that work well together on multiple datasets.\n\nThe paper focuses on CNNs that conform to a popular design pattern where the network is organized into a series of sub-networks, each consisting of a series of sub-modules (sometimes called blocks) operating at the same resolution. To be precise, the proposed method aims to learn the length of each series of sub-modules. A main contribution of the paper is the demonstration that it is not necessary to train a network until convergence before adding new sub-modules as proposed in past work. Instead, it is better to grow the network after training for a short while.\n\nMy current decision for this paper is a weak rejection due to the points below. However, I am open to revising my opinion if these points are addressed satisfactorily.\n\n- The growing strategy identified in the paper as a superior alternative seems to be already known and used, at least in the speech recognition community. Seide et al. (2011) called it Discriminative Pre-training, and showed that it outperforms greedy layer-wise pretraining and DBN pre-training. Zeyer et al. (2017) reported that a similar method also enables the training of very deep LSTM networks which is otherwise notoriously hard. In general, the existence of prior work with the same ideas does not preclude acceptance, but the existence of this work needs to be clearly stated early on and the additional value of the current study sufficiently clarified.\n\n- I find it strange that the final networks found by the proposed method usually have the same/similar number of sub-modules per sub-network (Tables 4,5,6) on multiple datasets. The only exceptions appear to be Basic4ResNet/CIFAR100 in Table 6 and about 50% of ImageNet results in Table 7. This regularity suggests that either A) the proposed algorithm prefers to set same number of sub-modules per sub-network due to its design, or B) datasets except ImageNet have an inherent shared property that produces this result. Since option A suggests a bias in the algorithm, this peculiarity of the results needs to be investigated or explained further.\n\n- Figure 3 constitutes the main evidence that Autogrow finds approximately optimal depths as compared to manual searching, but it is not clear how the plot for baselines is obtained. For any given parameter budget, there are multiple baseline networks possible since the sub-networks can have different number of sub-modules (see previous point). This does not appear to be accounted for in Figure 3. Further, when dealing with CNNs, it would be more useful to have computation budget on the x-axis instead of the parameter budget. This would better account for the difference between increasing depth in an earlier sub-network vs. a later one.\n\n- The reported results appear to be for single trials throughout the paper. This does not seem sufficient especially for results in Tables 2 and 3 where many differences are rather small, and so drawing conclusions from these tables would be unscientific.\n\nReferences:\n\nSeide, Frank, et al. \"Feature engineering in context-dependent deep neural networks for conversational speech transcription.\" 2011 IEEE Workshop on Automatic Speech Recognition & Understanding. IEEE, 2011. https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/FeatureEngineeringInCD-DNN-ASRU2011-pub.pdf\n\nZeyer, Albert, et al. \"A comprehensive study of deep bidirectional LSTM RNNs for acoustic modeling in speech recognition.\" 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017. https://arxiv.org/abs/1606.06871\n\nUpdate after rebuttal\n-----------------------------\nI'm sympathetic to the unfortunate situation that the authors are in, since the underlying growing strategy has already been covered by prior work. As I mentioned earlier, a sufficient rewrite of the paper can clearly state what has been done already so as not to take credit from the earlier authors. A revised version of the paper has not been uploaded; I suggest that the authors do so for the future. \n\nI agree that the focus of this paper is learning the 'optimal' depth by using the growing strategy. But I am not convinced that the technique indeed finds optimal depths based on the regularity of the sub-network depths mentioned in my review. The rebuttal suggests reasons for the obtained regularity, but does not prove that these regular structures obtained are indeed optimal and not an artifact of the algorithm itself. The baselines are also using the same regular architectures, which distorts the overall picture because it is possible that a non-regular architecture provides a better trade-off.\n\nWhile my rating doesn't change, I do think that the work is in an interesting direction. My final suggestions for the future are:\n- Investigate where non-regular architectures (unequal sub-network depths) are in the trade-off between accuracy, flops and parameters.\n- Investigate whether the proposed algorithm can be modified to easily find non-regular architectures if they can yield equally good performance as regular ones at similar or lower cost.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571898677764}, {"id": "r1emxj7GKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper284/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Contributions:\n\tThis paper best fits in the literature that explores growing network depth.  The main framework here is to interleave training a shallower network and adding new layers.  This paper (their final algorithm) differs from existing methods in that they: 1) initialize the new layers using standard initialization as opposed to the commonly used zero-init in this literature, 2) grows at a fixed interval , and this interval is short (to avoid the shallower nets being  overly-trained)., 3) uses a large and constant learning rate during the growing phase.  \nEmpirically, they show competitive results on standard image benchmarks.  \nMore interestingly (to me), they provide interesting insights to this paradigm of ‘growing networks’.  \n\nComments/Questions:\nSection 2 of the paper describes the proposed method is good details.  \n\nSection 3 of the paper describes the experiments.  Since for now I see the contribution of this paper is mostly empirical, I will give my detailed feedback here.\n3.1 (Suboptimum of Network Morphism (NM) )\nTable 2 shows NM is worse than training from scratch, and this isn’t fixed by AdamInit.\nTable 3 shows c-AutoGrow (in between p-AutoGrow and NM) still does worse than from scratch, pinpoint the problem to converged subnetworks.\n3.2  (p-AutoGrow)\nTable 3 shows +Constant LR helps, then +RandomInit helps. \nTable 4, 5 shows +Periodic gets the best performance.  \n*Suggestion* The found net is Table 4,5  are significantly deeper than those in Table 2,3, also there are no \\Delta.  Also, although within this write-up those are the highest numbers, in the broader literature of NAS this doesn’t seem to be that good.  From a quick search, many methods in the Table 1 of [1] seems to give >96% accuracy on CIFAR10, some even close to 98%.  It might be good to at least discuss why this method is limited from achieving that.\nI do like the finding that ZeroInit is unnecessary, as reported in the rest of this subsection.  However, it is unsatisfying to me that many past works (as cited by the authors) required this ZeroInit without ever trying RandomInit.  \n*Suggestion* I would love to see a more thorough discussion on why GauInit is better than ZeroInit, not just more numbers.  For example, even just text description on why past works found ZeroInit useful, and countering some of those claims would be interesting.  A more controlled experiment rather than training 2 networks by swapping this would be interesting.  ZeroInit is used in more context than just NM.  For example, good flow models like Glow also uses such initialization, for likely a different reason, but I wonder if findings here have any implication for ZeroInit more generally.\n\n3.3 (Many datasets)\nTable 6 is a strong result.  One odd thing is how deep the found-net has to be for MNIST.  This actually suggest to me that AutoGrow does not have the ability to stop early when it can.  And in the discussion, the authors argue that by using a better sub-module like in NAS they can do better.  This raises the question why the authors did not choose to use it.  I would believe it if the proposed method has obvious reasons that it can transfer to different architecture, but for now I cannot jump to the conclusion that, say, p-AutoGrow with GauInit will necessarily work when using a different sub-module.  Perhaps, the reason past NM works didn’t use a GauInit was also due to the fact that past sub-modules didn’t work with GauInit.  \n\n3.4 (Scale to ImageNet) It’d be good to add reference results from other papers.  \n\nMinor details:\n\n\nThere are some good contents in this work, but for it to be a strong *empirical* contribution, perhaps it would be more useful to include experiments on other data modality where things are not so well tuned, and show state-of-the-art results.  For it to be a strong *analysis* paper, it should expanded, at least addressing some of the *suggestions* mentioned above. \nUnrelated to my evaluation of this work, reading this makes me think we should (and can) develop theoretical understanding to this paradigm of growing networks.\n\n\n\nReferences:\n[1] https://arxiv.org/pdf/1905.13360.pdf\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Contributions:\n\tThis paper best fits in the literature that explores growing network depth.  The main framework here is to interleave training a shallower network and adding new layers.  This paper (their final algorithm) differs from existing methods in that they: 1) initialize the new layers using standard initialization as opposed to the commonly used zero-init in this literature, 2) grows at a fixed interval , and this interval is short (to avoid the shallower nets being  overly-trained)., 3) uses a large and constant learning rate during the growing phase.  \nEmpirically, they show competitive results on standard image benchmarks.  \nMore interestingly (to me), they provide interesting insights to this paradigm of ‘growing networks’.  \n\nComments/Questions:\nSection 2 of the paper describes the proposed method is good details.  \n\nSection 3 of the paper describes the experiments.  Since for now I see the contribution of this paper is mostly empirical, I will give my detailed feedback here.\n3.1 (Suboptimum of Network Morphism (NM) )\nTable 2 shows NM is worse than training from scratch, and this isn’t fixed by AdamInit.\nTable 3 shows c-AutoGrow (in between p-AutoGrow and NM) still does worse than from scratch, pinpoint the problem to converged subnetworks.\n3.2  (p-AutoGrow)\nTable 3 shows +Constant LR helps, then +RandomInit helps. \nTable 4, 5 shows +Periodic gets the best performance.  \n*Suggestion* The found net is Table 4,5  are significantly deeper than those in Table 2,3, also there are no \\Delta.  Also, although within this write-up those are the highest numbers, in the broader literature of NAS this doesn’t seem to be that good.  From a quick search, many methods in the Table 1 of [1] seems to give >96% accuracy on CIFAR10, some even close to 98%.  It might be good to at least discuss why this method is limited from achieving that.\nI do like the finding that ZeroInit is unnecessary, as reported in the rest of this subsection.  However, it is unsatisfying to me that many past works (as cited by the authors) required this ZeroInit without ever trying RandomInit.  \n*Suggestion* I would love to see a more thorough discussion on why GauInit is better than ZeroInit, not just more numbers.  For example, even just text description on why past works found ZeroInit useful, and countering some of those claims would be interesting.  A more controlled experiment rather than training 2 networks by swapping this would be interesting.  ZeroInit is used in more context than just NM.  For example, good flow models like Glow also uses such initialization, for likely a different reason, but I wonder if findings here have any implication for ZeroInit more generally.\n\n3.3 (Many datasets)\nTable 6 is a strong result.  One odd thing is how deep the found-net has to be for MNIST.  This actually suggest to me that AutoGrow does not have the ability to stop early when it can.  And in the discussion, the authors argue that by using a better sub-module like in NAS they can do better.  This raises the question why the authors did not choose to use it.  I would believe it if the proposed method has obvious reasons that it can transfer to different architecture, but for now I cannot jump to the conclusion that, say, p-AutoGrow with GauInit will necessarily work when using a different sub-module.  Perhaps, the reason past NM works didn’t use a GauInit was also due to the fact that past sub-modules didn’t work with GauInit.  \n\n3.4 (Scale to ImageNet) It’d be good to add reference results from other papers.  \n\nMinor details:\n\n\nThere are some good contents in this work, but for it to be a strong *empirical* contribution, perhaps it would be more useful to include experiments on other data modality where things are not so well tuned, and show state-of-the-art results.  For it to be a strong *analysis* paper, it should expanded, at least addressing some of the *suggestions* mentioned above. \nUnrelated to my evaluation of this work, reading this makes me think we should (and can) develop theoretical understanding to this paradigm of growing networks.\n\n\n\nReferences:\n[1] https://arxiv.org/pdf/1905.13360.pdf\n\n"}, "tcdate": 1571072746921}], "openreview_url": "https://openreview.net/forum?id=S1x0CnEtvB", "arxiv_id": "1906.02909", "paper_pdf": "papers/S1x0CnEtvB.pdf", "paper_pdf_sha256": "f67b3aed2d1282528c779172ebe52e274de7fe0ca6c19144324bbec0476a7a06", "paper_pdf_bytes": 1987332, "paper_pdf_source": "openreview", "code_url": "https://github.com/wenwei202/autogrow", "code_repository": "wenwei202/autogrow", "code_commit": "4435ee057126bacfc85063791ee2008b32c75cad", "code_archive": "repos/S1x0CnEtvB.zip", "code_archive_sha256": "5047b39fe379f14078e06f2413dddf382b2d989a94f507599f883b804ba0b67e", "code_archive_bytes": 61234, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 152, "github_languages": {"Python": 202941}, "github_archived": false, "github_pushed_at": "2019-06-10T05:25:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/autogrow-automatic-layer-growing-in-deep"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkxxIs0qY7", "year": 2019, "status": "rejected", "title": "CoT: Cooperative Training for Generative Modeling of Discrete Data", "authors": ["Sidi Lu", "Lantao Yu", "Siyuan Feng", "Yaoming Zhu", "Weinan Zhang", "Yong Yu"], "authorids": ["steve_lu@apex.sjtu.edu.cn", "yulantao@apex.sjtu.edu.cn", "siyuanfeng@apex.sjtu.edu", "ymzhu@apex.sjtu.edu.cn", "wnzhang@apex.sjtu.edu.cn", "yyu@apex.sjtu.edu.cn"], "authors_source": "OpenReview API", "abstract": "We propose Cooperative Training (CoT) for training generative models that measure a tractable density for discrete data. CoT coordinately trains a generator G and an auxiliary predictive mediator M. The training target of M is to estimate a mixture density of the learned distribution G and the target distribution P, and that of G is to minimize the Jensen-Shannon divergence estimated through M. CoT achieves independent success without the necessity of pre-training via Maximum Likelihood Estimation or involving high-variance algorithms like REINFORCE. This low-variance algorithm is theoretically proved to be superior for both sample generation and likelihood prediction. We also theoretically and empirically show the superiority of CoT over most previous algorithms in terms of generative quality and diversity, predictive generalization ability and computational cost.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "r1lfAuiyaQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper145/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "*Summary*\nA clear an interresting presentation on learning sequences distributions. It achieve this objective by replacing the discriminator with a \"mediator\", a mixture between the training distribution and the target distribution which is estimated via maximum likelihood.\n\n*Pros*\n- Original idea for modelling distribution of sequence data\n- Theoretical convergence in the Jensen Shanon divergence sense\n- Promising experiments\n\n*Cons*\n- No major cons to the best of my knowledge\n\n*Typos*\n- It would be very nice to have black and white / color blind friendly graphs\n- Eq 10 too long\n- Introduce J_m & J_g in  sentence\n- Coma at the end of Eq 5, and maybe align Generator and Discriminator in some position (e.g. at the semi colon).\n- missing dot at Eq 8.\n\n*Question*\n- How would you ensure reproducibility (e.g. link to some code?)\n- Is there any hope to obtain consistency (convergence) wrt other metrics?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Original idea, clear presentation", "review": "*Summary*\nA clear an interresting presentation on learning sequences distributions. It achieve this objective by replacing the discriminator with a \"mediator\", a mixture between the training distribution and the target distribution which is estimated via maximum likelihood.\n\n*Pros*\n- Original idea for modelling distribution of sequence data\n- Theoretical convergence in the Jensen Shanon divergence sense\n- Promising experiments\n\n*Cons*\n- No major cons to the best of my knowledge\n\n*Typos*\n- It would be very nice to have black and white / color blind friendly graphs\n- Eq 10 too long\n- Introduce J_m & J_g in  sentence\n- Coma at the end of Eq 5, and maybe align Generator and Discriminator in some position (e.g. at the semi colon).\n- missing dot at Eq 8.\n\n*Question*\n- How would you ensure reproducibility (e.g. link to some code?)\n- Is there any hope to obtain consistency (convergence) wrt other metrics?", "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541548233790}, {"id": "HkgJzPJP2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper145/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an interesting method, where the discriminator is replaced by a component that estimates the density that is the mixture of the data and the generator's distributions. In a sense, that component is only a device that allows estimating a Jensen-Shannon divergence for the generator to then be optimized against. Other GAN papers have replaced their discriminator by a similar device (e.g., WGANs, ..), but the present formulation seems novel. The numerical experiments presented on a synthetic Turing test and text generation from EMNLP's 2017 news dataset appear promising. \n\nOverall, the mediator seems to allow to achieve lower Jensen-Shannon (JS) divergence values in the experiments (and is kind of designed for that). Although this may be an improvement with respect to existing methods for discrete sequential data, it may also be limited in that it may not easily extend to other types of divergences that have proved superior to JS in some continuous settings.\n\nThe paper is rather clear, although there are lots of small grammatical errors as well as odd formulations which end up being distracting or confusing. The language should be proof-read carefully. \n\nPros:\n- Generative modeling of sequence data still in its infancy\n- Potentially lower variance than policy gradient approaches\n- Experiments are promising\n\nCons:\n- Lots of grammatical errors and odd formulations\n\nQuestions:\n- Equation 14: what does it mean to find the \"maximum entropy solution\" for the given optimization problem?\n- Figure 2: how do (b) and (c) relate to each other?\n\nRemarks, small typos and odd formulations:\n- \"for measuring M_\\/phi\": what does measuring mean in this context?\n- What does small m refer to? Algorithm 1 says the total number of steps  but it is also used in the main text as an index for J and \\pi (for mediator?)\n- Equation block 8: J_m has not been defined yet\n- \"the supports of distributions G and P\"... -> G without subscript has now been defined in this context\n- \"if the training being perfect\"\n- \"tend to get stuck in some sub-optimals\"\n- the learned distribution \"collapseS\"\n- \"since  the data distribution is, thus ...\"\n- \"that measures a\" -> \"that estimates a ...\"?\n- \"a predictive module\": a bit unclear - generative v. discriminative is more usual terminology\n- \"is well ensured\"\n- \"with the cost of diversity\" -> \"at the cost of diversity\"?\n- \"has theoretical guarantee\"\n- in the references: \"ALIAS PARTH GOYAL\" (all caps)\n- \"let p denote the intermediate states\": I don't understand what this is. Where is \"p\" used? (proof of Theorem 3)\n- \"CoT theoretically guarantees the training effectiveness\": what does that mean?\n- Figure 3: \"epochs\" -> \"Epochs\"\n- Algorithm 1: what does \"mixed balanced samples\" mean? Make this more precise\n- \"wide-ranged\"\n- Equation 10 is too long and equation number is not properly formatted\n- Figures hard to read in black & white\n- Figure 2 doesn't use the same limits for the Y axis of the two NLL plots, making comparisons difficult. The two NLL plots are also not side-by-side", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and promising method for generative modelling of sequence data without policy gradient", "review": "The paper proposes an interesting method, where the discriminator is replaced by a component that estimates the density that is the mixture of the data and the generator's distributions. In a sense, that component is only a device that allows estimating a Jensen-Shannon divergence for the generator to then be optimized against. Other GAN papers have replaced their discriminator by a similar device (e.g., WGANs, ..), but the present formulation seems novel. The numerical experiments presented on a synthetic Turing test and text generation from EMNLP's 2017 news dataset appear promising. \n\nOverall, the mediator seems to allow to achieve lower Jensen-Shannon (JS) divergence values in the experiments (and is kind of designed for that). Although this may be an improvement with respect to existing methods for discrete sequential data, it may also be limited in that it may not easily extend to other types of divergences that have proved superior to JS in some continuous settings.\n\nThe paper is rather clear, although there are lots of small grammatical errors as well as odd formulations which end up being distracting or confusing. The language should be proof-read carefully. \n\nPros:\n- Generative modeling of sequence data still in its infancy\n- Potentially lower variance than policy gradient approaches\n- Experiments are promising\n\nCons:\n- Lots of grammatical errors and odd formulations\n\nQuestions:\n- Equation 14: what does it mean to find the \"maximum entropy solution\" for the given optimization problem?\n- Figure 2: how do (b) and (c) relate to each other?\n\nRemarks, small typos and odd formulations:\n- \"for measuring M_\\/phi\": what does measuring mean in this context?\n- What does small m refer to? Algorithm 1 says the total number of steps  but it is also used in the main text as an index for J and \\pi (for mediator?)\n- Equation block 8: J_m has not been defined yet\n- \"the supports of distributions G and P\"... -> G without subscript has now been defined in this context\n- \"if the training being perfect\"\n- \"tend to get stuck in some sub-optimals\"\n- the learned distribution \"collapseS\"\n- \"since  the data distribution is, thus ...\"\n- \"that measures a\" -> \"that estimates a ...\"?\n- \"a predictive module\": a bit unclear - generative v. discriminative is more usual terminology\n- \"is well ensured\"\n- \"with the cost of diversity\" -> \"at the cost of diversity\"?\n- \"has theoretical guarantee\"\n- in the references: \"ALIAS PARTH GOYAL\" (all caps)\n- \"let p denote the intermediate states\": I don't understand what this is. Where is \"p\" used? (proof of Theorem 3)\n- \"CoT theoretically guarantees the training effectiveness\": what does that mean?\n- Figure 3: \"epochs\" -> \"Epochs\"\n- Algorithm 1: what does \"mixed balanced samples\" mean? Make this more precise\n- \"wide-ranged\"\n- Equation 10 is too long and equation number is not properly formatted\n- Figures hard to read in black & white\n- Figure 2 doesn't use the same limits for the Y axis of the two NLL plots, making comparisons difficult. The two NLL plots are also not side-by-side", "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1540974342883}, {"id": "Bygvu1w2jm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper145/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\nThis paper is easy to follow. The idea is nice in three folds. \n1. By changing the auxiliary model's role from a discriminator to a mediator, it directly optimizes the JSD measure, which is a symmetrized and smoothed version of KL divergence.  \n2. Moreover, the mediator and the generator follow similar predictive goals, rather than the opposite  goals of G and D in GANs. \n3. For discrete sequential data, it avoids approximating expected rewards using Markov rollouts.  \n \nCons:\nSome details are missing in the experiments. \n1. In Table 2 of [A], LeakGAN, SeqGAN and RankGAN all show significantly better performances in terms of BLEU on EMNLP2017 WMT, compared to results reported in Table 3 of the submission. Any difference?\n2. The Word Mover Distance is computed by training a discriminator, which could be unstable. Could you provide other metrics to evaluate diveristy like self-bleu?\n\n[A] Guo, Jiaxian, et al. \"Long text generation via adversarial training with leaked information.\" arXiv preprint arXiv:1709.08624 (2017).\n\nMisc:\n1. How will the number of samples (i.e. batch size) affect CoT ?\n2. How is the applicability of CoT for continuous data? It seems to me there is no theoretical difficulties to apply CoT on continuous data.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "nice idea", "review": "Pros:\nThis paper is easy to follow. The idea is nice in three folds. \n1. By changing the auxiliary model's role from a discriminator to a mediator, it directly optimizes the JSD measure, which is a symmetrized and smoothed version of KL divergence.  \n2. Moreover, the mediator and the generator follow similar predictive goals, rather than the opposite  goals of G and D in GANs. \n3. For discrete sequential data, it avoids approximating expected rewards using Markov rollouts.  \n \nCons:\nSome details are missing in the experiments. \n1. In Table 2 of [A], LeakGAN, SeqGAN and RankGAN all show significantly better performances in terms of BLEU on EMNLP2017 WMT, compared to results reported in Table 3 of the submission. Any difference?\n2. The Word Mover Distance is computed by training a discriminator, which could be unstable. Could you provide other metrics to evaluate diveristy like self-bleu?\n\n[A] Guo, Jiaxian, et al. \"Long text generation via adversarial training with leaked information.\" arXiv preprint arXiv:1709.08624 (2017).\n\nMisc:\n1. How will the number of samples (i.e. batch size) affect CoT ?\n2. How is the applicability of CoT for continuous data? It seems to me there is no theoretical difficulties to apply CoT on continuous data.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540284271380}], "openreview_url": "https://openreview.net/forum?id=SkxxIs0qY7", "arxiv_id": "1804.03782", "paper_pdf": "papers/SkxxIs0qY7.pdf", "paper_pdf_sha256": "f5f03e901fe5739c0d7abbf8bce2874ba9af8ff7268313a74f4cf2c9e1b2b34c", "paper_pdf_bytes": 800823, "paper_pdf_source": "openreview", "code_url": "https://github.com/desire2020/CoT", "code_repository": "desire2020/CoT", "code_commit": "05f390817e11733b7e4328ed2677e63287ed1290", "code_archive": "repos/SkxxIs0qY7.zip", "code_archive_sha256": "91e3088ad8fbe773cd6a989db6135b4dd01ecc04c03047927d9bca56d3e8a2a2", "code_archive_bytes": 1279051, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 1513, "github_languages": {"Python": 37310}, "github_archived": false, "github_pushed_at": "2019-05-25T14:01:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cot-cooperative-training-for-generative"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QDLGcnYdxe", "year": 2026, "status": "rejected", "title": "Rrls: Robust reinforcement learning suite", "authors": ["Adil Zouitine", "David Bertoin", "Pierre Clavier", "Matthieu Geist", "Emmanuel Rachelson"], "authorids": ["~Adil_Zouitine1", "~David_Bertoin1", "~Pierre_Clavier1", "~Matthieu_Geist1", "~Emmanuel_Rachelson1"], "authors_source": "OpenReview API", "abstract": "Robust reinforcement learning is the problem of learning control policies that provide optimal worst-case performance against a span of adversarial environments. It is a crucial ingredient for deploying algorithms in real-world scenarios with prevalent environmental uncertainties and has been a long-standing object of attention in the community, without a standardized set of benchmarks. This contribution endeavors to fill this gap. We introduce the Robust Reinforcement Learning Suite (RRLS), a benchmark suite based on Mujoco environments. RRLS provides six continuous control tasks with two types of uncertainty sets for training and evaluation. Our benchmark aims to standardize robust reinforcement learning tasks, facilitating reproducible and comparable experiments, in particular those from recent state-of-the-art contributions, for which we demonstrate the use of RRLS. It is also designed to be easily expandable to new environments.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "DZWr6Lt3XW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16756/Reviewer_ayZ2"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "The paper proposes Robust Reinforcement Learning Suite (RRLS), a set of benchmark tasks to evaluate robustness of reinforcement learning (RL) algorithms. RRLS includes 6 continuous control MuJoCo tasks (Ant, HalfCheetah, Hopper, Humanoid Stand Up, Inverted Pendulum, Walker), with parametric uncertainty sets for robust evaluation on each task. The paper also provides experimental results for 5 algorithms (a non-robust RL algorithm, domain randomization, and 3 robust RL algorithms) across the 6 tasks for one of the proposed uncertainty sets.", "review_text": "The paper proposes Robust Reinforcement Learning Suite (RRLS), a set of benchmark tasks to evaluate robustness of reinforcement learning (RL) algorithms. RRLS includes 6 continuous control MuJoCo tasks (Ant, HalfCheetah, Hopper, Humanoid Stand Up, Inverted Pendulum, Walker), with parametric uncertainty sets for robust evaluation on each task. The paper also provides experimental results for 5 algorithms (a non-robust RL algorithm, domain randomization, and 3 robust RL algorithms) across the 6 tasks for one of the proposed uncertainty sets.", "strengths": "**[S1]** Robust RL is an important area of research that could benefit from standardized benchmarks to evaluate existing algorithms and accelerate research advancements. The performance of robust RL algorithms often depends heavily on the choice of uncertainty set, and there is significant variation in the tasks / uncertainty sets used for evaluation across the field.", "weaknesses": "**[W1]** The authors claim there is a gap in benchmarks designed for robust RL, but they do not cite or discuss existing benchmarks that were designed to test robustness such as Robust Gymnasium [1] (ICLR 2025) and Real-World RL (RWRL) Suite [2]. Robust Gymnasium, in particular, provides a much more extensive set of tasks and disturbances for testing robustness in RL than the proposed RRLS, including the 6 tasks considered in RRLS. Given the existence of Robust Gymnasium, I do not believe this work provides a novel contribution to the field of robust RL.\n\n**[W2]** The scope of the proposed benchmark set is very limited, containing only 6 basic continuous control MuJoCo tasks with parametric uncertainty sets for evaluation. It does not provide a wide range of tasks covering different domains, does not provide realistic robotics tasks such as the ones provided in Isaac Lab [3] and others, and does not provide a wide range of disturbance types for evaluation. \n\n**[W3]** There is no analysis done to understand the difficulty levels of each environment contained in the proposed uncertainty sets. A standard RL algorithm should be trained on each specific evaluation environment to understand if meaningful performance can be achieved and quantify the trade-off required to achieve robustness. If an uncertainty set contains environments where meaningful performance is not possible, then it would not be a good choice for evaluating the performance of robust RL algorithms. \n\n**[W4]** The paper only provides benchmark results for a single uncertainty set per task (the 3D uncertainty sets) under a single set of robustness hyperparameters for robust RL training. This provides minimal insight into the important performance vs. robustness tradeoff of robust RL algorithms under different training / evaluation setups. This limited experimental setup is insufficient for a paper focused on proposing a benchmark of tasks and uncertainty sets to advance robust RL research.\n\n**References:**\n\n[1] Gu et al., “Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning.” In ICLR 2025. https://github.com/SafeRL-Lab/Robust-Gymnasium\n\n[2] Dulac-Arnold et al., “Challenges of real-world reinforcement learning: definitions, benchmarks and analysis.” Machine Learning, 2021. https://github.com/google-research/realworldrl_suite\n\n[3] Mittal et al., “Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments.” IEEE RA-L, 2023. https://github.com/isaac-sim/IsaacLab", "questions": "**[Q1]** Please discuss the contribution of this work compared to existing robust RL benchmark sets including Robust Gymnasium [1] and Real-World RL (RWRL) Suite [2]. Robust Gymnasium, in particular, already contains the 6 tasks proposed in this work as part of a comprehensive set of robust RL benchmark tasks. These related benchmark sets should also be cited and discussed in the paper.\n\n**[Q2]** I recommend significantly expanding the experimental analysis to demonstrate why the provided choices of tasks and uncertainty sets represent a meaningful set of evaluation environments that will push forward the field of robust RL.\n\n**[Q3]** How were the robustness hyperparameters of the robust RL methods selected / tuned? Please include these implementation details in the paper, as the performance of robust RL algorithms often depends heavily on the choice of these hyperparameters.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Robust Reinforcement Learning Suite (RRLS), a set of benchmark tasks to evaluate robustness of reinforcement learning (RL) algorithms. RRLS includes 6 continuous control MuJoCo tasks (Ant, HalfCheetah, Hopper, Humanoid Stand Up, Inverted Pendulum, Walker), with parametric uncertainty sets for robust evaluation on each task. The paper also provides experimental results for 5 algorithms (a non-robust RL algorithm, domain randomization, and 3 robust RL algorithms) across the 6 tasks for one of the proposed uncertainty sets.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "**[S1]** Robust RL is an important area of research that could benefit from standardized benchmarks to evaluate existing algorithms and accelerate research advancements. The performance of robust RL algorithms often depends heavily on the choice of uncertainty set, and there is significant variation in the tasks / uncertainty sets used for evaluation across the field.", "weaknesses": "**[W1]** The authors claim there is a gap in benchmarks designed for robust RL, but they do not cite or discuss existing benchmarks that were designed to test robustness such as Robust Gymnasium [1] (ICLR 2025) and Real-World RL (RWRL) Suite [2]. Robust Gymnasium, in particular, provides a much more extensive set of tasks and disturbances for testing robustness in RL than the proposed RRLS, including the 6 tasks considered in RRLS. Given the existence of Robust Gymnasium, I do not believe this work provides a novel contribution to the field of robust RL.\n\n**[W2]** The scope of the proposed benchmark set is very limited, containing only 6 basic continuous control MuJoCo tasks with parametric uncertainty sets for evaluation. It does not provide a wide range of tasks covering different domains, does not provide realistic robotics tasks such as the ones provided in Isaac Lab [3] and others, and does not provide a wide range of disturbance types for evaluation. \n\n**[W3]** There is no analysis done to understand the difficulty levels of each environment contained in the proposed uncertainty sets. A standard RL algorithm should be trained on each specific evaluation environment to understand if meaningful performance can be achieved and quantify the trade-off required to achieve robustness. If an uncertainty set contains environments where meaningful performance is not possible, then it would not be a good choice for evaluating the performance of robust RL algorithms. \n\n**[W4]** The paper only provides benchmark results for a single uncertainty set per task (the 3D uncertainty sets) under a single set of robustness hyperparameters for robust RL training. This provides minimal insight into the important performance vs. robustness tradeoff of robust RL algorithms under different training / evaluation setups. This limited experimental setup is insufficient for a paper focused on proposing a benchmark of tasks and uncertainty sets to advance robust RL research.\n\n**References:**\n\n[1] Gu et al., “Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning.” In ICLR 2025. https://github.com/SafeRL-Lab/Robust-Gymnasium\n\n[2] Dulac-Arnold et al., “Challenges of real-world reinforcement learning: definitions, benchmarks and analysis.” Machine Learning, 2021. https://github.com/google-research/realworldrl_suite\n\n[3] Mittal et al., “Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments.” IEEE RA-L, 2023. https://github.com/isaac-sim/IsaacLab", "questions": "**[Q1]** Please discuss the contribution of this work compared to existing robust RL benchmark sets including Robust Gymnasium [1] and Real-World RL (RWRL) Suite [2]. Robust Gymnasium, in particular, already contains the 6 tasks proposed in this work as part of a comprehensive set of robust RL benchmark tasks. These related benchmark sets should also be cited and discussed in the paper.\n\n**[Q2]** I recommend significantly expanding the experimental analysis to demonstrate why the provided choices of tasks and uncertainty sets represent a meaningful set of evaluation environments that will push forward the field of robust RL.\n\n**[Q3]** How were the robustness hyperparameters of the robust RL methods selected / tuned? Please include these implementation details in the paper, as the performance of robust RL algorithms often depends heavily on the choice of these hyperparameters.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761920209120}, {"id": "w2X8AIGn8d", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16756/Reviewer_vEuz"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper presents the Robust Reinforcement Learning Suite (RRLS), a Gymnasium-compatible benchmark designed to standardize robustness evaluation in continuous-control tasks. RRLS provides a clean API for parameter perturbations and defines non-rectangular uncertainty sets with multiple difficulty levels across six standard MuJoCo environments. The authors aim to establish a reproducible testbed for robust RL evaluation and demonstrate its use through experiments comparing representative algorithms.", "review_text": "The paper presents the Robust Reinforcement Learning Suite (RRLS), a Gymnasium-compatible benchmark designed to standardize robustness evaluation in continuous-control tasks. RRLS provides a clean API for parameter perturbations and defines non-rectangular uncertainty sets with multiple difficulty levels across six standard MuJoCo environments. The authors aim to establish a reproducible testbed for robust RL evaluation and demonstrate its use through experiments comparing representative algorithms.", "strengths": "RRLS offers a practical and well-structured interface for parameter perturbations within Gymnasium, consolidating previously scattered robustness protocols into a single reproducible pipeline. The paper documents the uncertainty sets with explicit parameter listings and ranges, enhancing comparability and experimental transparency. The paper includes baseline experiments on several standard algorithms under the proposed uncertainty sets, reporting training curves and seed-level performance variance as an empirical characterization of robustness behavior.", "weaknesses": "Relative to the current state of robust RL benchmarking, the contribution is narrow in scope. RRLS only covers six MuJoCo control tasks with environment-parameter uncertainty and does not address observation, action, or reward-level disruptions, nor broader domains such as safety, multi-agent, or vision-based RL. The benchmark lacks the breadth, modularity, and disruption coverage demonstrated in recent work like Robust-Gymnasium, which already defines a unified disrupted-MDP abstraction across multiple modalities.\n\nThe empirical evaluation is modest, reporting only a few standard baselines and offering limited analytical depth. No large-scale stress testing, cross-domain comparisons, or leaderboard-style benchmarking are included, making the study closer to a small-scale technical experiment than a community-standard benchmark.\n\nFinally, the submission does not convincingly differentiate itself from existing platforms; its overlap with prior benchmarks undermines the novelty claim.\n\n[1] Gu S, Shi L, Wen M, et al. Robust gymnasium: A unified modular benchmark for robust reinforcement learning[J]. arXiv preprint arXiv:2502.19652, 2025.\n[2] Dulac-Arnold G, Levine N, Mankowitz D J, et al. An empirical investigation of the challenges of real-world reinforcement learning[J]. arXiv preprint arXiv:2003.11881, 2020.", "questions": "The paper does not clearly articulate its advantages over existing benchmark suites. What are the concrete improvements or distinctive features of RRLS, and is the code or framework publicly available?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents the Robust Reinforcement Learning Suite (RRLS), a Gymnasium-compatible benchmark designed to standardize robustness evaluation in continuous-control tasks. RRLS provides a clean API for parameter perturbations and defines non-rectangular uncertainty sets with multiple difficulty levels across six standard MuJoCo environments. The authors aim to establish a reproducible testbed for robust RL evaluation and demonstrate its use through experiments comparing representative algorithms.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "RRLS offers a practical and well-structured interface for parameter perturbations within Gymnasium, consolidating previously scattered robustness protocols into a single reproducible pipeline. The paper documents the uncertainty sets with explicit parameter listings and ranges, enhancing comparability and experimental transparency. The paper includes baseline experiments on several standard algorithms under the proposed uncertainty sets, reporting training curves and seed-level performance variance as an empirical characterization of robustness behavior.", "weaknesses": "Relative to the current state of robust RL benchmarking, the contribution is narrow in scope. RRLS only covers six MuJoCo control tasks with environment-parameter uncertainty and does not address observation, action, or reward-level disruptions, nor broader domains such as safety, multi-agent, or vision-based RL. The benchmark lacks the breadth, modularity, and disruption coverage demonstrated in recent work like Robust-Gymnasium, which already defines a unified disrupted-MDP abstraction across multiple modalities.\n\nThe empirical evaluation is modest, reporting only a few standard baselines and offering limited analytical depth. No large-scale stress testing, cross-domain comparisons, or leaderboard-style benchmarking are included, making the study closer to a small-scale technical experiment than a community-standard benchmark.\n\nFinally, the submission does not convincingly differentiate itself from existing platforms; its overlap with prior benchmarks undermines the novelty claim.\n\n[1] Gu S, Shi L, Wen M, et al. Robust gymnasium: A unified modular benchmark for robust reinforcement learning[J]. arXiv preprint arXiv:2502.19652, 2025.\n[2] Dulac-Arnold G, Levine N, Mankowitz D J, et al. An empirical investigation of the challenges of real-world reinforcement learning[J]. arXiv preprint arXiv:2003.11881, 2020.", "questions": "The paper does not clearly articulate its advantages over existing benchmark suites. What are the concrete improvements or distinctive features of RRLS, and is the code or framework publicly available?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No", "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761894650066}, {"id": "55gPpvdgxv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16756/Reviewer_QAB7"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper provided additional functionalities mostly needed by robust RL works, i.e., changing parameters of the environmental dynamics such as friction, inertia, on top of the existing Gymnasium benchmark which was originally designed for more generic RL approaches.", "review_text": "This paper provided additional functionalities mostly needed by robust RL works, i.e., changing parameters of the environmental dynamics such as friction, inertia, on top of the existing Gymnasium benchmark which was originally designed for more generic RL approaches.", "strengths": "* Paper is clearly written.\n* A few baselines were included and benchmarked using the Rrls suite.", "weaknesses": "* The major concern from the reviewer is that there seems to already exist robust RL benchmarks, Robust Gymnasium (RG) [1], that could cover the setup considered by Rrls. Specifically, Rrls considered a fixed set of environmental disturbances (or the so-called uncertainty set in the paper), while RG considered a various types of disturbance. \n\n* In general, Rrls maybe suited for a specific type of robust RL approaches, e.g., when the uncertainty set is fixed. However, there exist lines of robust RL research that are not specific to a pre-given set of uncertainties. For example, action perturbation [2, 3], risk-averse [4], or distribution robust [5].\n\n* RG also seems to support more environments beyond the 6 supported by Rrls, and benchmarked against a wider suite of baselines.\n\n[1] Gu, Shangding, et al. \"Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning.\" The Thirteenth International Conference on Learning Representations.\n\n[2] Stanton, Samuel, et al. \"Robust reinforcement learning for shifting dynamics during deployment.\" (2021).\n\n[3] Tessler, Chen, Yonathan Efroni, and Shie Mannor. \"Action robust reinforcement learning and applications in continuous control.\" International Conference on Machine Learning. PMLR, 2019.\n\n[4] Singh, Rahul, Qinsheng Zhang, and Yongxin Chen. \"Improving robustness via risk averse distributional reinforcement learning.\" Learning for Dynamics and Control. PMLR, 2020.\n\n[5] Liu, Zijian, et al. \"Distributionally Robust $ Q $-Learning.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "The reviewer is curious about the distinct advantages that RRLS offers compared to Robust-Gymnasium. Specifically, what unique benefits or experimental capabilities does RRLS provide that are not already covered by Robust-Gymnasium?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper provided additional functionalities mostly needed by robust RL works, i.e., changing parameters of the environmental dynamics such as friction, inertia, on top of the existing Gymnasium benchmark which was originally designed for more generic RL approaches.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "* Paper is clearly written.\n* A few baselines were included and benchmarked using the Rrls suite.", "weaknesses": "* The major concern from the reviewer is that there seems to already exist robust RL benchmarks, Robust Gymnasium (RG) [1], that could cover the setup considered by Rrls. Specifically, Rrls considered a fixed set of environmental disturbances (or the so-called uncertainty set in the paper), while RG considered a various types of disturbance. \n\n* In general, Rrls maybe suited for a specific type of robust RL approaches, e.g., when the uncertainty set is fixed. However, there exist lines of robust RL research that are not specific to a pre-given set of uncertainties. For example, action perturbation [2, 3], risk-averse [4], or distribution robust [5].\n\n* RG also seems to support more environments beyond the 6 supported by Rrls, and benchmarked against a wider suite of baselines.\n\n[1] Gu, Shangding, et al. \"Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning.\" The Thirteenth International Conference on Learning Representations.\n\n[2] Stanton, Samuel, et al. \"Robust reinforcement learning for shifting dynamics during deployment.\" (2021).\n\n[3] Tessler, Chen, Yonathan Efroni, and Shie Mannor. \"Action robust reinforcement learning and applications in continuous control.\" International Conference on Machine Learning. PMLR, 2019.\n\n[4] Singh, Rahul, Qinsheng Zhang, and Yongxin Chen. \"Improving robustness via risk averse distributional reinforcement learning.\" Learning for Dynamics and Control. PMLR, 2020.\n\n[5] Liu, Zijian, et al. \"Distributionally Robust $ Q $-Learning.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "The reviewer is curious about the distinct advantages that RRLS offers compared to Robust-Gymnasium. Specifically, what unique benefits or experimental capabilities does RRLS provide that are not already covered by Robust-Gymnasium?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761884832409}, {"id": "PazFG1Ud24", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16756/Reviewer_MtfD"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper introduces RRLS, a new benchmark suite for evaluating robust RL algorithms. RRLS provides six continuous-control tasks (Ant, HalfCheetah, Hopper, Walker, Humanoid Stand Up, and Inverted Pendulum) built on MuJoCo, each with two types of uncertainty sets (parameter perturbations and adversarial forces).", "review_text": "This paper introduces RRLS, a new benchmark suite for evaluating robust RL algorithms. RRLS provides six continuous-control tasks (Ant, HalfCheetah, Hopper, Walker, Humanoid Stand Up, and Inverted Pendulum) built on MuJoCo, each with two types of uncertainty sets (parameter perturbations and adversarial forces).", "strengths": "* The lack of standardized robust RL benchmarks is a well-known issue; this paper directly addresses it with a clear, well-scoped suite.\n\n* The wrapper-based architecture makes RRLS easy to integrate with existing Gymnasium pipelines.\n\n* The empirical section systematically compares multiple algorithms, highlighting trade-offs between robustness and average performance.", "weaknesses": "* There are some important related works providing robust RL benchmarks, that are missing in the related work and even the main motivations of the paper [1, 2]. \n* There exist other types of robustness perturbations, such as state, reward and action perturbations [3]. It would be helpful to claim the position of the paper compared with those works. \n\n[1] Dulac-Arnold G, Levine N, Mankowitz D J, et al. An empirical investigation of the challenges of real-world reinforcement learning[J]. arXiv preprint arXiv:2003.11881, 2020.\n\n[2] Gu S, Shi L, Wen M, et al. Robust gymnasium: A unified modular benchmark for robust reinforcement learning[J]. arXiv preprint arXiv:2502.19652, 2025. \n\n[3] Moos J, Hansel K, Abdulsamad H, et al. Robust reinforcement learning: A review of foundations and recent advances[J]. Machine Learning and Knowledge Extraction, 2022, 4(1): 276-315.", "questions": "* Are training and testing environments generated by the same sampling distribution?\n* Table 5: what is the meaning of the number after the robot name?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces RRLS, a new benchmark suite for evaluating robust RL algorithms. RRLS provides six continuous-control tasks (Ant, HalfCheetah, Hopper, Walker, Humanoid Stand Up, and Inverted Pendulum) built on MuJoCo, each with two types of uncertainty sets (parameter perturbations and adversarial forces).", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "* The lack of standardized robust RL benchmarks is a well-known issue; this paper directly addresses it with a clear, well-scoped suite.\n\n* The wrapper-based architecture makes RRLS easy to integrate with existing Gymnasium pipelines.\n\n* The empirical section systematically compares multiple algorithms, highlighting trade-offs between robustness and average performance.", "weaknesses": "* There are some important related works providing robust RL benchmarks, that are missing in the related work and even the main motivations of the paper [1, 2]. \n* There exist other types of robustness perturbations, such as state, reward and action perturbations [3]. It would be helpful to claim the position of the paper compared with those works. \n\n[1] Dulac-Arnold G, Levine N, Mankowitz D J, et al. An empirical investigation of the challenges of real-world reinforcement learning[J]. arXiv preprint arXiv:2003.11881, 2020.\n\n[2] Gu S, Shi L, Wen M, et al. Robust gymnasium: A unified modular benchmark for robust reinforcement learning[J]. arXiv preprint arXiv:2502.19652, 2025. \n\n[3] Moos J, Hansel K, Abdulsamad H, et al. Robust reinforcement learning: A review of foundations and recent advances[J]. Machine Learning and Knowledge Extraction, 2022, 4(1): 276-315.", "questions": "* Are training and testing environments generated by the same sampling distribution?\n* Table 5: what is the meaning of the number after the robot name?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761863474596}], "openreview_url": "https://openreview.net/forum?id=QDLGcnYdxe", "arxiv_id": "2406.08406", "paper_pdf": "papers/QDLGcnYdxe.pdf", "paper_pdf_sha256": "43a7f18574ae397f1b4dadb16f9f1273ef025c0d3668c35949ca9ce7bd5e2277", "paper_pdf_bytes": 1728928, "paper_pdf_source": "openreview", "code_url": "https://github.com/SuReLI/RRLS", "code_repository": "SuReLI/RRLS", "code_commit": "0c43dc981c0e892a604c596e2bc399df187a87ac", "code_archive": "repos/QDLGcnYdxe.zip", "code_archive_sha256": "44ddf88d1894f811c6988f16bfcdcfe26fae28feed725dd73c86d9126c29648a", "code_archive_bytes": 682944, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 736, "github_languages": {"Python": 142354}, "github_archived": false, "github_pushed_at": "2024-12-24T11:20:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rrls-robust-reinforcement-learning-suite"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nwZHFKrYTB", "year": 2025, "status": "rejected", "title": "How to Train Long-Context Language Models (Effectively)", "authors": ["Tianyu Gao", "Alexander Wettig", "Howard Yen", "Danqi Chen"], "authorids": ["~Tianyu_Gao1", "~Alexander_Wettig1", "~Howard_Yen1", "~Danqi_Chen1"], "authors_source": "OpenReview API", "abstract": "We study the problem of adapting a language model (LM) to make effective use of long-context information. We first establish a reliable evaluation protocol to guide model development---instead of perplexity, we use a broad set of long-context tasks, and we evaluate models after supervised fine-tuning (SFT) with instruction data as this better reveals long-context abilities. Supported by our robust evaluations, we run thorough experiments to decide the data mix for continued pre-training, the instruction tuning dataset, and other design choices such as position extrapolation. We find that (1) code repositories and books are excellent sources of long data, but it is crucial to combine them with high-quality short data; (2) training with a sequence length beyond the evaluation length boosts long-context performance; (3) for SFT, using only short instruction datasets yields strong performance on long-context tasks. Our final model, ProLong-8B, which is initialized from Llama-3 and trained on 40B tokens, demonstrates state-of-the-art long-context performance among similarly sized models at a length of 128K, outperforming Llama-3.1-8B on the majority of long-context tasks despite having seen 5% as many tokens during long-context training. Additionally, ProLong can effectively process up to 512K tokens, one of the longest context windows of publicly available LMs.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "X55krR8MMq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10781/Reviewer_ive5"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The main contributions of the paper are related to evaluations, data mix, and training recipe. \n\nEvaluation related contribution\n* Using HELMET (an existing benchmark) instead of perplexity and needle in the haystack for evaluating long context expansion.\n* The authors suggest evaluating LLMs for long context after SFT instead of after pretraining.\n\nData mix related contributions\n\n* Carefully designed data mixtures like high quality code and books are important. \n* Additionally, the authors ablate the ratio of short context and long context data in the data mix to show that a very high fraction of long context data is detrimental to both long context and short context tasks.\n\nTraining recipe: \n* Training on more tokens improves performance of long context tasks\n* Increasing pretraining context length beyond target context length can also help long context tasks.\n* They also show that synthetic data during SFT can be harmful to long context benchmarks\n\nThey combine these improvements to develop a new model called ProLong which achieves better results than prior models on HELMET and NoCha (another long context benchmark).", "review_text": "The main contributions of the paper are related to evaluations, data mix, and training recipe. \n\nEvaluation related contribution\n* Using HELMET (an existing benchmark) instead of perplexity and needle in the haystack for evaluating long context expansion.\n* The authors suggest evaluating LLMs for long context after SFT instead of after pretraining.\n\nData mix related contributions\n\n* Carefully designed data mixtures like high quality code and books are important. \n* Additionally, the authors ablate the ratio of short context and long context data in the data mix to show that a very high fraction of long context data is detrimental to both long context and short context tasks.\n\nTraining recipe: \n* Training on more tokens improves performance of long context tasks\n* Increasing pretraining context length beyond target context length can also help long context tasks.\n* They also show that synthetic data during SFT can be harmful to long context benchmarks\n\nThey combine these improvements to develop a new model called ProLong which achieves better results than prior models on HELMET and NoCha (another long context benchmark).", "strengths": "* Careful examination of individual changes (data mix, sft etc.)\n* Final recipe does provide a model that outperforms some of the existing models", "weaknesses": "The paper contributions are somewhat limited. The main contributions are around datamixing and training recipe. The authors make some interesting claims but unfortunately don't investigate these claims further. \n\n* The results on perplexity are a bit surprising. The authors show that increasing long context data results in perplexity improvements but long context tasks performance is lower. It would be really interesting to understand this further. Is the quality degraded due to the data mix? Or is the target eval task not correlated to the PG19 books dataset?\n* The authors show that using synthetic data during SFT can lead to poor performance. It would have been helpful if the authors had verified that the generated data is correct prior to dismissing it. \n* Additionally, using a large model (say Llama 70B) for synthetic data generation could have been more interesting. Perhaps, this larger model would have generated data that can be used for SFT.\n* The trends between SFT evaluation and pretrained model evaluation are not clear. In Figure 2, many of the benchmarks show similar trends before and after SFT, so it's hard to conclude that it is necessary to measure performance after SFT as the authors claim. \n\nThe paper would also be stronger if the authors show that their methods generalize beyond just Llama 3 8b. Would the same set of techniques also expand context lengths effectively for other open source models?", "questions": "* How does this recipe scale? Would this recipe work for larger models? \n* Since the model is trained for 512k context length, it would be very helpful to understand the performance of the model at 512k context length tasks. Do any of the evaluations have such long context tasks? \n* The authors mention that the performance on short context tasks is important but don't report performance on these tasks for their final ProLong model. How does ProLong perform on short context tasks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The main contributions of the paper are related to evaluations, data mix, and training recipe. \n\nEvaluation related contribution\n* Using HELMET (an existing benchmark) instead of perplexity and needle in the haystack for evaluating long context expansion.\n* The authors suggest evaluating LLMs for long context after SFT instead of after pretraining.\n\nData mix related contributions\n\n* Carefully designed data mixtures like high quality code and books are important. \n* Additionally, the authors ablate the ratio of short context and long context data in the data mix to show that a very high fraction of long context data is detrimental to both long context and short context tasks.\n\nTraining recipe: \n* Training on more tokens improves performance of long context tasks\n* Increasing pretraining context length beyond target context length can also help long context tasks.\n* They also show that synthetic data during SFT can be harmful to long context benchmarks\n\nThey combine these improvements to develop a new model called ProLong which achieves better results than prior models on HELMET and NoCha (another long context benchmark).", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* Careful examination of individual changes (data mix, sft etc.)\n* Final recipe does provide a model that outperforms some of the existing models", "weaknesses": "The paper contributions are somewhat limited. The main contributions are around datamixing and training recipe. The authors make some interesting claims but unfortunately don't investigate these claims further. \n\n* The results on perplexity are a bit surprising. The authors show that increasing long context data results in perplexity improvements but long context tasks performance is lower. It would be really interesting to understand this further. Is the quality degraded due to the data mix? Or is the target eval task not correlated to the PG19 books dataset?\n* The authors show that using synthetic data during SFT can lead to poor performance. It would have been helpful if the authors had verified that the generated data is correct prior to dismissing it. \n* Additionally, using a large model (say Llama 70B) for synthetic data generation could have been more interesting. Perhaps, this larger model would have generated data that can be used for SFT.\n* The trends between SFT evaluation and pretrained model evaluation are not clear. In Figure 2, many of the benchmarks show similar trends before and after SFT, so it's hard to conclude that it is necessary to measure performance after SFT as the authors claim. \n\nThe paper would also be stronger if the authors show that their methods generalize beyond just Llama 3 8b. Would the same set of techniques also expand context lengths effectively for other open source models?", "questions": "* How does this recipe scale? Would this recipe work for larger models? \n* Since the model is trained for 512k context length, it would be very helpful to understand the performance of the model at 512k context length tasks. Do any of the evaluations have such long context tasks? \n* The authors mention that the performance on short context tasks is important but don't report performance on these tasks for their final ProLong model. How does ProLong perform on short context tasks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730762380136}, {"id": "G4FU1wcsTV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10781/Reviewer_kqM8"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper proposes a continual training recipe, including data curation and multi-stage context extension, that successfully equips the llama-3-8B-instruct model with long context ability up to 512K tokens.", "review_text": "The paper proposes a continual training recipe, including data curation and multi-stage context extension, that successfully equips the llama-3-8B-instruct model with long context ability up to 512K tokens.", "strengths": "- The model trained with the final recipe (ProLong) shows strong performance on long context benchmarks.\n\n- The proposed data mixture can be a valuable contribution for the community.", "weaknesses": "- Missing crucial baselines: The author should at least include a very baseline that applies the recipe of Fu et al. (2024) on the llama-3-8B model with 40B tokens to provide a fair comparison between the ProLong recipe and the Fu et al. (2024) recipe on both short and long context tasks. The Table 10 and 12 is not very informative because of lacking both fair comparisons and the short-context results.\n\n- Misleading evaluation: As in L174, the short-context performance is evaluated before SFT, while the long context performance is after SFT (As in L149). However, what people really care about is if the final model after SFT can have both good short and long-context performances.\n\n- The paper claims counterintuitive conclusions but lacks in-depth analyses explaining why they happen. The author claims that training on longer than evaluation context can bring performance benefits (in L331), but it can be simply due to the fact of training on 4B more tokens, which can improve the instruction following ability because of the short data mixture. The author claims that the long synthetic SFT data does not improve performance but it can be simply because of using a weak 8B model for data generation. See more details in the Question section.", "questions": "- Why training on all long context data will harm the downstream long-context tasks? Is it because (1) the proposed ShortMix data actually contains lots of QA data from Fineweb-Edu and Stack Exchange which can improve instruction following (2) the curated long context data (code repos+books) does not have QA data, and training purely on it will harm the instruction following ability?\nIn Line 351, why use the llama-3-8B-instruct model to synthesize long context SFT data, instead of more powerful models such as lama-3-70B-instruct? I would assume an 8B model cannot accomplish these difficult tasks and can provide low-quality long-context SFT data.\n\n- Could you also evaluate ProLong on some popular synthetic long-context benchmarks such as RULER? In this way, we can have a better understanding of the limitations of ProLong.\n\n- The stage 2 training also doubled the batch size from 4M to 8M. Do you have an ablation study for it? Batch size ramp-up is believed to provide more accurate optimization trajectory and boost model convergences, therefore the performance gains from 512K length training may actually be from the batch size ramp-up instead of longer context.\n\n- Could you also evaluate the final ProLong model on short-context tasks, and add comparisons with a llama3-8B-instruct model SFTed with UltraChat?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a continual training recipe, including data curation and multi-stage context extension, that successfully equips the llama-3-8B-instruct model with long context ability up to 512K tokens.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The model trained with the final recipe (ProLong) shows strong performance on long context benchmarks.\n\n- The proposed data mixture can be a valuable contribution for the community.", "weaknesses": "- Missing crucial baselines: The author should at least include a very baseline that applies the recipe of Fu et al. (2024) on the llama-3-8B model with 40B tokens to provide a fair comparison between the ProLong recipe and the Fu et al. (2024) recipe on both short and long context tasks. The Table 10 and 12 is not very informative because of lacking both fair comparisons and the short-context results.\n\n- Misleading evaluation: As in L174, the short-context performance is evaluated before SFT, while the long context performance is after SFT (As in L149). However, what people really care about is if the final model after SFT can have both good short and long-context performances.\n\n- The paper claims counterintuitive conclusions but lacks in-depth analyses explaining why they happen. The author claims that training on longer than evaluation context can bring performance benefits (in L331), but it can be simply due to the fact of training on 4B more tokens, which can improve the instruction following ability because of the short data mixture. The author claims that the long synthetic SFT data does not improve performance but it can be simply because of using a weak 8B model for data generation. See more details in the Question section.", "questions": "- Why training on all long context data will harm the downstream long-context tasks? Is it because (1) the proposed ShortMix data actually contains lots of QA data from Fineweb-Edu and Stack Exchange which can improve instruction following (2) the curated long context data (code repos+books) does not have QA data, and training purely on it will harm the instruction following ability?\nIn Line 351, why use the llama-3-8B-instruct model to synthesize long context SFT data, instead of more powerful models such as lama-3-70B-instruct? I would assume an 8B model cannot accomplish these difficult tasks and can provide low-quality long-context SFT data.\n\n- Could you also evaluate ProLong on some popular synthetic long-context benchmarks such as RULER? In this way, we can have a better understanding of the limitations of ProLong.\n\n- The stage 2 training also doubled the batch size from 4M to 8M. Do you have an ablation study for it? Batch size ramp-up is believed to provide more accurate optimization trajectory and boost model convergences, therefore the performance gains from 512K length training may actually be from the batch size ramp-up instead of longer context.\n\n- Could you also evaluate the final ProLong model on short-context tasks, and add comparisons with a llama3-8B-instruct model SFTed with UltraChat?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730697674664}, {"id": "7cSmybifih", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10781/Reviewer_DH4H"], "rating": 6, "soundness": 2, "presentation": 4, "contribution": 3, "confidence": 5, "summary": "This paper investigates the optimal recipe of training a long-context LLM from the data engineering perspective. The authors investigate the effects of data source, mixing proportion of long and short data, etc, and evaluate the trained models on realistic tasks beyond perplexity.", "review_text": "This paper investigates the optimal recipe of training a long-context LLM from the data engineering perspective. The authors investigate the effects of data source, mixing proportion of long and short data, etc, and evaluate the trained models on realistic tasks beyond perplexity.", "strengths": "The paper gives a thorough investigation of data engineering for long context training across various aspects, covering long-short performance balance, SFT, and evaluation. This paper can potentially serve as a handbook for long context training. The paper is well-organized and easy to understand.", "weaknesses": "- The evaluation beyond ppl is good and there are some previous studies [1] proprose ppl is not proper for evaluating long-context performance. However, it is kind of confused if the evaluation can serve as a \"contribution\" of this paper? It seems the paper just adopt the HELMET benchmark (I don't think the decision is a contribution as most long-context works have adopted some realistic benchmarks like LongBench [2] and InfiniteBench [3]).\n\n- The paper proposed short-context instruction tuning is sufficient for achieving good performance. The study is conducted by mixing partial long synthetic instruction data. The authors found that the incorporation of long synthetic data would degrade the long-context performance.  As our understanding in short-context scenario, data quality is (most) essential for SFT [4]. However, the synthetic long-context data (generated by LLaMA-3-8B) is probably of low quality. So I think this study only demonstrates that long synthetic data is not good enough for improving long-context performance. BTW, have you compared the long-context performance before and after SFT? LLaMA 3.1 technical report [5] proposed the long-context performance would degenerate with short instruction data only after SFT. I wonder if this point holds here.\n\n- I believe summarization is a vital and basic long-context ability, however, there is a significant performance drop in the summarization task after ProLong's training (compared to naive LLaMA-3.1). Is this due to the lack of some long data types/sources in training data (books and code only)? The information in long books and codes may be distributed without summarizing. I feel like arxiv/wikipedia  (which you only use in shortmix) are good data sources for summarization, as they naively contain an abstract section. The authors can consider including more data sources with ablations (as that in Table 4).\n\n- Would you consider reporting results on more benchmarks like RULER [6] and InfiniteBench [3]? HELMET is a good but new benchmark while some experiences (e.g., the degeneration of long-context performance when using short instruction data only.) are acquired from evaluation on some older benchmark. More results would help us align the findings.\n\n\n[1] Can Perplexity Reflect Large Language Model’s Ability in Long Text Understanding?\n\n[2] LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding\n\n[3] ∞Bench: Extending Long Context Evaluation Beyond 100K Tokens\n\n[4] LIMA: Less Is More for Alignment\n\n[5] The Llama 3 Herd of Models\n\n[6] RULER: What's the Real Context Size of Your Long-Context Language Models?", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the optimal recipe of training a long-context LLM from the data engineering perspective. The authors investigate the effects of data source, mixing proportion of long and short data, etc, and evaluate the trained models on realistic tasks beyond perplexity.", "soundness": 2, "presentation": 4, "contribution": 3, "strengths": "The paper gives a thorough investigation of data engineering for long context training across various aspects, covering long-short performance balance, SFT, and evaluation. This paper can potentially serve as a handbook for long context training. The paper is well-organized and easy to understand.", "weaknesses": "- The evaluation beyond ppl is good and there are some previous studies [1] proprose ppl is not proper for evaluating long-context performance. However, it is kind of confused if the evaluation can serve as a \"contribution\" of this paper? It seems the paper just adopt the HELMET benchmark (I don't think the decision is a contribution as most long-context works have adopted some realistic benchmarks like LongBench [2] and InfiniteBench [3]).\n\n- The paper proposed short-context instruction tuning is sufficient for achieving good performance. The study is conducted by mixing partial long synthetic instruction data. The authors found that the incorporation of long synthetic data would degrade the long-context performance.  As our understanding in short-context scenario, data quality is (most) essential for SFT [4]. However, the synthetic long-context data (generated by LLaMA-3-8B) is probably of low quality. So I think this study only demonstrates that long synthetic data is not good enough for improving long-context performance. BTW, have you compared the long-context performance before and after SFT? LLaMA 3.1 technical report [5] proposed the long-context performance would degenerate with short instruction data only after SFT. I wonder if this point holds here.\n\n- I believe summarization is a vital and basic long-context ability, however, there is a significant performance drop in the summarization task after ProLong's training (compared to naive LLaMA-3.1). Is this due to the lack of some long data types/sources in training data (books and code only)? The information in long books and codes may be distributed without summarizing. I feel like arxiv/wikipedia  (which you only use in shortmix) are good data sources for summarization, as they naively contain an abstract section. The authors can consider including more data sources with ablations (as that in Table 4).\n\n- Would you consider reporting results on more benchmarks like RULER [6] and InfiniteBench [3]? HELMET is a good but new benchmark while some experiences (e.g., the degeneration of long-context performance when using short instruction data only.) are acquired from evaluation on some older benchmark. More results would help us align the findings.\n\n\n[1] Can Perplexity Reflect Large Language Model’s Ability in Long Text Understanding?\n\n[2] LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding\n\n[3] ∞Bench: Extending Long Context Evaluation Beyond 100K Tokens\n\n[4] LIMA: Less Is More for Alignment\n\n[5] The Llama 3 Herd of Models\n\n[6] RULER: What's the Real Context Size of Your Long-Context Language Models?", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730629127689}, {"id": "oLPFRiUouw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10781/Reviewer_WrRx"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 4, "confidence": 4, "summary": "This paper study the problem of how to effectively adapting a short-context language models to be long-context. The paper first starts with identifying deficiency in widely used perplexity and Needle-In-A-Heystack test and establishing a reliable evaluation protocol for long-context LLM by adopting HELMET, which covers diverse range of realistic long-context applications. \nThe authors also advocate for (1) checking model performance after supervised finetuning (2) ensuring that model performance on short-context tasks is not compromised after long context training. For long context data curation, the authors provide best practices for (1) data mixture in long-context data (2) quantity ratio between long and short context data (3) choosing high-quality short-context data. \nNext, the paper discuss the impact of scaling training tokens and context length on both long and short task performance. Finally, the recipe ends with exploration on the choice of SFT data, pointing out that short-context instruction data can yield strong long-context result. The resulting long-context model ProLong as well as associated code and data are open-sourced.", "review_text": "This paper study the problem of how to effectively adapting a short-context language models to be long-context. The paper first starts with identifying deficiency in widely used perplexity and Needle-In-A-Heystack test and establishing a reliable evaluation protocol for long-context LLM by adopting HELMET, which covers diverse range of realistic long-context applications. \nThe authors also advocate for (1) checking model performance after supervised finetuning (2) ensuring that model performance on short-context tasks is not compromised after long context training. For long context data curation, the authors provide best practices for (1) data mixture in long-context data (2) quantity ratio between long and short context data (3) choosing high-quality short-context data. \nNext, the paper discuss the impact of scaling training tokens and context length on both long and short task performance. Finally, the recipe ends with exploration on the choice of SFT data, pointing out that short-context instruction data can yield strong long-context result. The resulting long-context model ProLong as well as associated code and data are open-sourced.", "strengths": "1. The paper is very well structured and written, and is easy to follow.\n2. The paper thoroughly explored important components and provide useful observations for long-context language models development, including training recipe and evaluation protocol. From the evaluation perspective, the authors underscore the importance of using diverse and realistic long-context tasks rather than relying on conventional perplexity-based and needle-in-a-haystack-like benchmarks. The preservation of short-context performance is also highlighted. From the training perspective, the authors conducted experiments and found empirically effective choices of data mixture, data ratio, and data quality.\n3. The ultimately trained model, ProLong, achieves strong long context performance at 10B parameter level.\n4. The assets associated with this paper could serve as useful resources for the research community and are open-sourced.", "weaknesses": "1. The choice of long-context evaluation benchmark(HELMET) seems arbitrary. The authors should discuss the rationale behind selecting HELMET instead of other long-context benchmarks, such as LongBench, Infinite Bench, etc.\n2. The training-free length extrapolation methods compared in the paper is limited. The experiments would be more solid if more advanced training-free extension methods are incorporated, e.g., SelfExtend and ChunkLLaMa.\n3. There exists other community-tuned long-context LLMs initialized from LLaMa-3-8B-Instruct, e.g., gradientai/Llama-3-8B-Instruct-Gradient-1048k. As they claim their training process only consume 1.4B tokens, it would be more convincing to include such models as strong baselines.", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper study the problem of how to effectively adapting a short-context language models to be long-context. The paper first starts with identifying deficiency in widely used perplexity and Needle-In-A-Heystack test and establishing a reliable evaluation protocol for long-context LLM by adopting HELMET, which covers diverse range of realistic long-context applications. \nThe authors also advocate for (1) checking model performance after supervised finetuning (2) ensuring that model performance on short-context tasks is not compromised after long context training. For long context data curation, the authors provide best practices for (1) data mixture in long-context data (2) quantity ratio between long and short context data (3) choosing high-quality short-context data. \nNext, the paper discuss the impact of scaling training tokens and context length on both long and short task performance. Finally, the recipe ends with exploration on the choice of SFT data, pointing out that short-context instruction data can yield strong long-context result. The resulting long-context model ProLong as well as associated code and data are open-sourced.", "soundness": 3, "presentation": 4, "contribution": 4, "strengths": "1. The paper is very well structured and written, and is easy to follow.\n2. The paper thoroughly explored important components and provide useful observations for long-context language models development, including training recipe and evaluation protocol. From the evaluation perspective, the authors underscore the importance of using diverse and realistic long-context tasks rather than relying on conventional perplexity-based and needle-in-a-haystack-like benchmarks. The preservation of short-context performance is also highlighted. From the training perspective, the authors conducted experiments and found empirically effective choices of data mixture, data ratio, and data quality.\n3. The ultimately trained model, ProLong, achieves strong long context performance at 10B parameter level.\n4. The assets associated with this paper could serve as useful resources for the research community and are open-sourced.", "weaknesses": "1. The choice of long-context evaluation benchmark(HELMET) seems arbitrary. The authors should discuss the rationale behind selecting HELMET instead of other long-context benchmarks, such as LongBench, Infinite Bench, etc.\n2. The training-free length extrapolation methods compared in the paper is limited. The experiments would be more solid if more advanced training-free extension methods are incorporated, e.g., SelfExtend and ChunkLLaMa.\n3. There exists other community-tuned long-context LLMs initialized from LLaMa-3-8B-Instruct, e.g., gradientai/Llama-3-8B-Instruct-Gradient-1048k. As they claim their training process only consume 1.4B tokens, it would be more convincing to include such models as strong baselines.", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729172194170}, {"id": "hHKnu9NZ1L", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10781/Reviewer_GrHC"], "rating": 6, "soundness": 1, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "The paper addresses the challenges involved in training language models (LMs) to make effective use of long-context information. It proposes a systematic methodology that goes beyond traditional evaluation metrics, such as perplexity, to train models capable of handling long-context inputs more effectively. The authors introduce ProLong-8B, a model initialized from Llama-3 and trained on 40 billion tokens, demonstrating state-of-the-art performance for long-context tasks.", "review_text": "The paper addresses the challenges involved in training language models (LMs) to make effective use of long-context information. It proposes a systematic methodology that goes beyond traditional evaluation metrics, such as perplexity, to train models capable of handling long-context inputs more effectively. The authors introduce ProLong-8B, a model initialized from Llama-3 and trained on 40 billion tokens, demonstrating state-of-the-art performance for long-context tasks.", "strengths": "1.  The study highlights that training on a mix of both long and short-context data sources is key to achieving good performance across tasks. The authors show that using long-context data like code repositories and books, balanced with high-quality short-context data, preserves both long and short-context abilities of the model.\n2.  The extensive ablation experiments provide meaningful insights into effective training practices. For instance, the paper reveals that training only on long-context data negatively impacts overall performance and that including high-quality short-context data is essential for optimal results.\n3. The resulting ProLong-8B model achieves strong performance on long-context tasks compared to similarly sized models", "weaknesses": "1. **Misalignment Between Title and Content**: A significant portion of the paper focuses on data engineering rather than providing comprehensive training methodologies for long-context language models, which does not fully align with the promise of the title. To better align with the title, the paper could have included more discussions on training techniques, such as **pose: efficient context window extension of LLMs via positional skip-wise training** and **CLEX: Continuous Length Extrapolation for Large Language Models**. Additionally, incorporating experiments involving different rotary position encoding methods and synthetic data, such as **LongSkywork: A Training Recipe for Efficiently Extending Context Length in Large Language Models**, would provide a more thorough exploration of how to train long-context LLMs.\n\n2. **Benchmark Justification**: In Section 2, the necessity of introducing a new benchmark is not clearly justified, especially considering that many of the tasks overlap with existing benchmarks like **RULER**. Providing more rationale for the benchmark's uniqueness？\n\n3. **Lack of Explanation**: In Section 3.2, the authors note that previous techniques (using long text) deteriorate short-context performance, but they do not provide a satisfactory explanation for this phenomenon. Without a clear rationale, the motivation for mixing in short-context training data remains weak.\n\n4. **Insufficient Explanation of Data Mixture Ratios**: The paper proposes two mixture ratios—one for combining long-context data sources (books and repositories) and another for mixing long and short-context data. While the experiments show the effectiveness of these ratios, there is no clear explanation of why these particular proportions work. \n\n5. **Lack of Comparison with Short-Context Performance Task**: The final experiments lack a comparison with short-context performance tasks. Including such comparisons would provide a more holistic evaluation of the model's performance.", "questions": "The paper's contributions are not clearly defined in terms of their nature, like technical report？ If the paper aims to contribute as a methodological study on data engineering, the novelty seems insufficient, as it mainly provides data mixture ratios without in-depth justification or exploration. On the other hand, if it aims to be an experimental study, the lack of diverse experiments and limited comparisons with other training approaches weakens its claims of novelty.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenges involved in training language models (LMs) to make effective use of long-context information. It proposes a systematic methodology that goes beyond traditional evaluation metrics, such as perplexity, to train models capable of handling long-context inputs more effectively. The authors introduce ProLong-8B, a model initialized from Llama-3 and trained on 40 billion tokens, demonstrating state-of-the-art performance for long-context tasks.", "soundness": 1, "presentation": 4, "contribution": 2, "strengths": "1.  The study highlights that training on a mix of both long and short-context data sources is key to achieving good performance across tasks. The authors show that using long-context data like code repositories and books, balanced with high-quality short-context data, preserves both long and short-context abilities of the model.\n2.  The extensive ablation experiments provide meaningful insights into effective training practices. For instance, the paper reveals that training only on long-context data negatively impacts overall performance and that including high-quality short-context data is essential for optimal results.\n3. The resulting ProLong-8B model achieves strong performance on long-context tasks compared to similarly sized models", "weaknesses": "1. **Misalignment Between Title and Content**: A significant portion of the paper focuses on data engineering rather than providing comprehensive training methodologies for long-context language models, which does not fully align with the promise of the title. To better align with the title, the paper could have included more discussions on training techniques, such as **pose: efficient context window extension of LLMs via positional skip-wise training** and **CLEX: Continuous Length Extrapolation for Large Language Models**. Additionally, incorporating experiments involving different rotary position encoding methods and synthetic data, such as **LongSkywork: A Training Recipe for Efficiently Extending Context Length in Large Language Models**, would provide a more thorough exploration of how to train long-context LLMs.\n\n2. **Benchmark Justification**: In Section 2, the necessity of introducing a new benchmark is not clearly justified, especially considering that many of the tasks overlap with existing benchmarks like **RULER**. Providing more rationale for the benchmark's uniqueness？\n\n3. **Lack of Explanation**: In Section 3.2, the authors note that previous techniques (using long text) deteriorate short-context performance, but they do not provide a satisfactory explanation for this phenomenon. Without a clear rationale, the motivation for mixing in short-context training data remains weak.\n\n4. **Insufficient Explanation of Data Mixture Ratios**: The paper proposes two mixture ratios—one for combining long-context data sources (books and repositories) and another for mixing long and short-context data. While the experiments show the effectiveness of these ratios, there is no clear explanation of why these particular proportions work. \n\n5. **Lack of Comparison with Short-Context Performance Task**: The final experiments lack a comparison with short-context performance tasks. Including such comparisons would provide a more holistic evaluation of the model's performance.", "questions": "The paper's contributions are not clearly defined in terms of their nature, like technical report？ If the paper aims to contribute as a methodological study on data engineering, the novelty seems insufficient, as it mainly provides data mixture ratios without in-depth justification or exploration. On the other hand, if it aims to be an experimental study, the lack of diverse experiments and limited comparisons with other training approaches weakens its claims of novelty.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729171781907}], "openreview_url": "https://openreview.net/forum?id=nwZHFKrYTB", "arxiv_id": "2410.02660", "paper_pdf": "papers/nwZHFKrYTB.pdf", "paper_pdf_sha256": "9c00e32fddd31390bb4fc30c6486f51ac28d5821179e93d45213e07268aa9bb6", "paper_pdf_bytes": 412185, "paper_pdf_source": "openreview", "code_url": "https://github.com/princeton-nlp/ProLong", "code_repository": "princeton-nlp/ProLong", "code_commit": "499fa2994fa727b4c8b53b706ccd5c23f77d7f89", "code_archive": "repos/nwZHFKrYTB.zip", "code_archive_sha256": "ca1ffa9dbaa9002e596244eb11e481237cbfb1597e857a355173b14e901717f9", "code_archive_bytes": 43146, "code_file_count": 8, "code_extensions": {".py": 5, ".sh": 3}, "github_disk_usage_kb": 64, "github_languages": {"Python": 105806, "Shell": 14272}, "github_archived": false, "github_pushed_at": "2025-09-12T19:56:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-to-train-long-context-language-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5oJlyJXUxK", "year": 2024, "status": "rejected", "title": "Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable?", "authors": ["Ričards Marcinkevičs", "Sonia Laguna", "Moritz Vandenhirtz", "Julia E Vogt"], "authorids": ["~Ričards_Marcinkevičs1", "~Sonia_Laguna1", "~Moritz_Vandenhirtz1", "~Julia_E_Vogt1"], "authors_source": "OpenReview API", "abstract": "Recently, interpretable machine learning has re-explored concept bottleneck models (CBM), comprising step-by-step prediction of the high-level concepts from the raw features and the target variable from the predicted concepts. A compelling advantage of this model class is the user's ability to intervene on the predicted concept values, consequently affecting the model's downstream output. In this work, we introduce a method to perform such concept-based interventions on already-trained neural networks, which are not interpretable by design. Furthermore, we formalise the model's *intervenability* as a measure of the effectiveness of concept-based interventions and leverage this definition to fine-tune black-box models. Empirically, we explore the intervenability of black-box classifiers on synthetic tabular and natural image benchmarks.  We demonstrate that fine-tuning improves intervention effectiveness and often yields better-calibrated predictions. To showcase the practical utility of the proposed techniques, we apply them to deep chest X-ray classifiers and show that fine-tuned black boxes can be as intervenable and more performant than CBMs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "EQZgTuE1kn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5402/Reviewer_JHNX"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents a way to do concept interventions on standard neural networks without the need for a Concept Bottleneck Model. This can be done by learning classifiers to predict the presence of a concept based on the hidden layer representation, and then optimizing to find representation that is as close as possible to original but changes the concept classifier prediction. They also propose to improve the effectiveness of interventions by finetuning the model to get better results under intervention.", "review_text": "This paper presents a way to do concept interventions on standard neural networks without the need for a Concept Bottleneck Model. This can be done by learning classifiers to predict the presence of a concept based on the hidden layer representation, and then optimizing to find representation that is as close as possible to original but changes the concept classifier prediction. They also propose to improve the effectiveness of interventions by finetuning the model to get better results under intervention.", "strengths": "Clearly written. Interesting perspective highlighting that intervention could be useful even on standard networks.", "weaknesses": "I don't really see any use cases where we would want to intervene on standard models instead of just creating a CBM. \n- Intervention performance on models that weren't finetuned is quite poor, and still requires labeled concept data to learn the classifiers.\n- The performance of models (even finetuned) is worse than CBM on most datasets, while both require additional training and the same kind of data with dense concept labels, most cases it would be better to just learn a CBM\n- Intervening now requires solving an optimization problem making it more costly and harder to understand than original CBM interventions\n- CBMs have many interpretability benefits in addition to intervene-ability, which we lose when using standard architecture, such as predictions being simple functions of interpretable concepts. \n\nLacking evaluation:\n- I think improved performance on CheXpert is likely caused by the fact that the model can use information outside of concepts to make the prediction. This is similar to having residual as is done by Posthoc-CBM-h, and I think some comparison agaisnt that would be needed.\n- Choice of datasets is a little odd, should use at least some of the datasets original CBM was trained on such as CUB", "questions": "Looks like each intervention requires running gradient descent to minimize eq. 1, what is the computational cost of this? \nHow did you intervene on multiple concepts at once?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a way to do concept interventions on standard neural networks without the need for a Concept Bottleneck Model. This can be done by learning classifiers to predict the presence of a concept based on the hidden layer representation, and then optimizing to find representation that is as close as possible to original but changes the concept classifier prediction. They also propose to improve the effectiveness of interventions by finetuning the model to get better results under intervention.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "Clearly written. Interesting perspective highlighting that intervention could be useful even on standard networks.", "weaknesses": "I don't really see any use cases where we would want to intervene on standard models instead of just creating a CBM. \n- Intervention performance on models that weren't finetuned is quite poor, and still requires labeled concept data to learn the classifiers.\n- The performance of models (even finetuned) is worse than CBM on most datasets, while both require additional training and the same kind of data with dense concept labels, most cases it would be better to just learn a CBM\n- Intervening now requires solving an optimization problem making it more costly and harder to understand than original CBM interventions\n- CBMs have many interpretability benefits in addition to intervene-ability, which we lose when using standard architecture, such as predictions being simple functions of interpretable concepts. \n\nLacking evaluation:\n- I think improved performance on CheXpert is likely caused by the fact that the model can use information outside of concepts to make the prediction. This is similar to having residual as is done by Posthoc-CBM-h, and I think some comparison agaisnt that would be needed.\n- Choice of datasets is a little odd, should use at least some of the datasets original CBM was trained on such as CUB", "questions": "Looks like each intervention requires running gradient descent to minimize eq. 1, what is the computational cost of this? \nHow did you intervene on multiple concepts at once?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698817964801}, {"id": "4aiLRNX3sR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5402/Reviewer_39mA"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "- The paper proposes interventions on black-box models in CBM style.\n\n\n- Given a black-box model:\n     -  Train a probing network to extract concepts from an intermediate representation $z$\n     -  Now that they have the probing network for a given sample $x$ they can extract $z$ and concepts $c$, given the ground-truth concepts $c'$ they learn a new embedding $z'$ that should produce concepts $c'$ when given to probing network.\n     -  For interventions they replace the original $z$ with new $z'$\n\n- Given $z$ how do we calculate $z'$? (from the text is a bit unclear I had to look at the code to understand how this is done so please correct me if I am wrong)\n     -    Start from $z'==z$ but $z'$ is differentiable.\n     -    The probing network is frozen where you calculate concepts $q_\\xi(z')$\n     -    Update $z'$ based on the loss in equation 1 you repeat multiple times i.e for a few epochs.\n\n\n- The paper quantifies the effectiveness of interventions as the gap between the regular prediction loss and the loss attained after the intervention.\n- The paper proposes a fine-tuning strategy for intervention that can be summarized as follows:\n     -    Given a black-box model $f_\\theta$, we will look at the network as if its a cbm model such that we first have a network $h_\\phi(x)=z$ that gives us the intermediate representation used to train the probing network and $g_\\psi(z)=y$ that gives the final prediction.\n     -    The black box network is now trained end to end using loss in equation 4: the first term is regular black-box optimization the second term is optimizing the outer network subjected to the intervention and all of this is in addition to optimizing $z'$  as mentioned previously.\n     -    Equation 4 is then simplified to avoid trilevel optimizing the first loss is ignored so the feature extractor layer is basically frozen.\n-  Experiments:\n    -    The paper one synthetic tabular dataset, 3 image datasets.\n    -    The paper tested the proposed fine-tuning approach against a black-box NN a CBM and two other different fine-tuning approaches.", "review_text": "- The paper proposes interventions on black-box models in CBM style.\n\n\n- Given a black-box model:\n     -  Train a probing network to extract concepts from an intermediate representation $z$\n     -  Now that they have the probing network for a given sample $x$ they can extract $z$ and concepts $c$, given the ground-truth concepts $c'$ they learn a new embedding $z'$ that should produce concepts $c'$ when given to probing network.\n     -  For interventions they replace the original $z$ with new $z'$\n\n- Given $z$ how do we calculate $z'$? (from the text is a bit unclear I had to look at the code to understand how this is done so please correct me if I am wrong)\n     -    Start from $z'==z$ but $z'$ is differentiable.\n     -    The probing network is frozen where you calculate concepts $q_\\xi(z')$\n     -    Update $z'$ based on the loss in equation 1 you repeat multiple times i.e for a few epochs.\n\n\n- The paper quantifies the effectiveness of interventions as the gap between the regular prediction loss and the loss attained after the intervention.\n- The paper proposes a fine-tuning strategy for intervention that can be summarized as follows:\n     -    Given a black-box model $f_\\theta$, we will look at the network as if its a cbm model such that we first have a network $h_\\phi(x)=z$ that gives us the intermediate representation used to train the probing network and $g_\\psi(z)=y$ that gives the final prediction.\n     -    The black box network is now trained end to end using loss in equation 4: the first term is regular black-box optimization the second term is optimizing the outer network subjected to the intervention and all of this is in addition to optimizing $z'$  as mentioned previously.\n     -    Equation 4 is then simplified to avoid trilevel optimizing the first loss is ignored so the feature extractor layer is basically frozen.\n-  Experiments:\n    -    The paper one synthetic tabular dataset, 3 image datasets.\n    -    The paper tested the proposed fine-tuning approach against a black-box NN a CBM and two other different fine-tuning approaches.", "strengths": "Strength:\n- Originality: The paper is original, there has been work around post-hoc CBM (cited by the paper) but this intervention strategy is novel and very interesting.\n- Quality: The paper is evaluated against a reasonable baseline on multiple datasets.\n- Significance: The paper's contribution is significant, if we can intervene on model in test time we can get higher accuracy as shown on multiple datasets.", "weaknesses": "Method weakness:\n- There is no guarantee that a network has learned the desired concepts, i.e. there is a big probability that the probing network can not learn a concept you would want to intervene on.\n- The method is quite expensive optimizing to get the $z'$ and optimizing for fine-tuning on top of $z'$ can be quite costly.\n- Creating a probing network per concept can be costly when we have a large number of concepts, it is not clear if this can scale to hundreds or thousands of concepts.\n- It is not clear which concepts one should intervene on.\n- What if we can never intervene during test time (say we don't have ground-truth concepts or even ground-truth label at that point) it is unclear how this can be useful in that case.\n\n\nPaper quality:\n- How you get $z'$ and the intervening strategy is not very clear in the main paper I would strongly recommend moving Algorithm A.1 to the main text for clarity.", "questions": "- Which layer do you do probing on is it the last layer before the classifier?\n- Usually how many iterations do you need to extract a reasonable $z'$.\n- In the experiments, the accuracy on the test set is \"with\" interventions correct?\n- How do you select the concepts to intervene on?\n- How would this model be used practically? (I am assuming something similar to the following steps):\n   \n\n    - You have an example that is incorrect\n    - You calculate the concepts that affect that example.\n   - You show the concepts to a domain expert, and they propose different concepts.\n   - You calculate $z'$ using this new concept and make a new prediction.\n\nIf so how is the original model improved it?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "- The paper proposes interventions on black-box models in CBM style.\n\n\n- Given a black-box model:\n     -  Train a probing network to extract concepts from an intermediate representation $z$\n     -  Now that they have the probing network for a given sample $x$ they can extract $z$ and concepts $c$, given the ground-truth concepts $c'$ they learn a new embedding $z'$ that should produce concepts $c'$ when given to probing network.\n     -  For interventions they replace the original $z$ with new $z'$\n\n- Given $z$ how do we calculate $z'$? (from the text is a bit unclear I had to look at the code to understand how this is done so please correct me if I am wrong)\n     -    Start from $z'==z$ but $z'$ is differentiable.\n     -    The probing network is frozen where you calculate concepts $q_\\xi(z')$\n     -    Update $z'$ based on the loss in equation 1 you repeat multiple times i.e for a few epochs.\n\n\n- The paper quantifies the effectiveness of interventions as the gap between the regular prediction loss and the loss attained after the intervention.\n- The paper proposes a fine-tuning strategy for intervention that can be summarized as follows:\n     -    Given a black-box model $f_\\theta$, we will look at the network as if its a cbm model such that we first have a network $h_\\phi(x)=z$ that gives us the intermediate representation used to train the probing network and $g_\\psi(z)=y$ that gives the final prediction.\n     -    The black box network is now trained end to end using loss in equation 4: the first term is regular black-box optimization the second term is optimizing the outer network subjected to the intervention and all of this is in addition to optimizing $z'$  as mentioned previously.\n     -    Equation 4 is then simplified to avoid trilevel optimizing the first loss is ignored so the feature extractor layer is basically frozen.\n-  Experiments:\n    -    The paper one synthetic tabular dataset, 3 image datasets.\n    -    The paper tested the proposed fine-tuning approach against a black-box NN a CBM and two other different fine-tuning approaches.", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "strengths": "Strength:\n- Originality: The paper is original, there has been work around post-hoc CBM (cited by the paper) but this intervention strategy is novel and very interesting.\n- Quality: The paper is evaluated against a reasonable baseline on multiple datasets.\n- Significance: The paper's contribution is significant, if we can intervene on model in test time we can get higher accuracy as shown on multiple datasets.", "weaknesses": "Method weakness:\n- There is no guarantee that a network has learned the desired concepts, i.e. there is a big probability that the probing network can not learn a concept you would want to intervene on.\n- The method is quite expensive optimizing to get the $z'$ and optimizing for fine-tuning on top of $z'$ can be quite costly.\n- Creating a probing network per concept can be costly when we have a large number of concepts, it is not clear if this can scale to hundreds or thousands of concepts.\n- It is not clear which concepts one should intervene on.\n- What if we can never intervene during test time (say we don't have ground-truth concepts or even ground-truth label at that point) it is unclear how this can be useful in that case.\n\n\nPaper quality:\n- How you get $z'$ and the intervening strategy is not very clear in the main paper I would strongly recommend moving Algorithm A.1 to the main text for clarity.", "questions": "- Which layer do you do probing on is it the last layer before the classifier?\n- Usually how many iterations do you need to extract a reasonable $z'$.\n- In the experiments, the accuracy on the test set is \"with\" interventions correct?\n- How do you select the concepts to intervene on?\n- How would this model be used practically? (I am assuming something similar to the following steps):\n   \n\n    - You have an example that is incorrect\n    - You calculate the concepts that affect that example.\n   - You show the concepts to a domain expert, and they propose different concepts.\n   - You calculate $z'$ using this new concept and make a new prediction.\n\nIf so how is the original model improved it?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698528126488}, {"id": "JuwhnKjY00", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5402/Reviewer_icNv"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors introduce a method to perform concept-based interventions on pre-trained neural networks. They then formalize intervenability as a metric to measure concept-based interventions. Finally, they show that finetuning probes for intervenability can improve intervenability.", "review_text": "The authors introduce a method to perform concept-based interventions on pre-trained neural networks. They then formalize intervenability as a metric to measure concept-based interventions. Finally, they show that finetuning probes for intervenability can improve intervenability.", "strengths": "- The paper's proposed metric of intervenability is interesting and helps practically measure the utility of methods such as CBMs\n- The introduced methods are clear and intuitive\n- Strong comparisons are made to baselines from previous work", "weaknesses": "- Limited improvements over CBMs -- In most settings, CBMs seem to outperform the proposed Fine-tuned, I method while providing much greater interpretability (the main exception is in Fig 5)\n- Potential missing baseline: I am not sure if a simple convex combination of the predictions of the CBM and the black-box would outperform the proposed model", "questions": "- Note: would be nice for the paper to mention the link between intervenability and concept/feature importance - this should be straightforward based on interventions, as this is how many feature importance metrics (e.g. LIME/SHAP) are computed\n- Would be nice to describe the experimental setup in more detail, e.g. the three synthetic scenarios are only described in previous work\n- Minor: fig legends in Fig 4 are slightly difficult to read.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce a method to perform concept-based interventions on pre-trained neural networks. They then formalize intervenability as a metric to measure concept-based interventions. Finally, they show that finetuning probes for intervenability can improve intervenability.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The paper's proposed metric of intervenability is interesting and helps practically measure the utility of methods such as CBMs\n- The introduced methods are clear and intuitive\n- Strong comparisons are made to baselines from previous work", "weaknesses": "- Limited improvements over CBMs -- In most settings, CBMs seem to outperform the proposed Fine-tuned, I method while providing much greater interpretability (the main exception is in Fig 5)\n- Potential missing baseline: I am not sure if a simple convex combination of the predictions of the CBM and the black-box would outperform the proposed model", "questions": "- Note: would be nice for the paper to mention the link between intervenability and concept/feature importance - this should be straightforward based on interventions, as this is how many feature importance metrics (e.g. LIME/SHAP) are computed\n- Would be nice to describe the experimental setup in more detail, e.g. the three synthetic scenarios are only described in previous work\n- Minor: fig legends in Fig 4 are slightly difficult to read.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698249645489}, {"id": "XEgvCMYWcc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5402/Reviewer_Emzi"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a method to make conceptual interventions by modifying the activations of a pre-trained black-box model. Based on the popular counterfactual optimization approach by Wachter et al [1] for pixel-based counterfactuals, they propose to instead optimize for activations. That is, they apply the distance loss in the activation instead of pixel space and use the gradients obtained by an additionally trained concept detector ((non)-linear probe) instead of the gradients of the classifier (c.f., Eq. 1). Finally, the authors propose a fine-tuning scheme to make the classifier more reliant on the “concept activation vectors” based on a proposed notion of intervenability (Eq. 2 & 3), while keeping the feature extractor frozen. Through experiments on tabular and vision data the authors show the efficacy of their intervention as well as fine-tuning scheme.\n\n[1] Wachter, Sandra, et al. \"Counterfactual explanations without opening the black box: Automated decisions and the GDPR.\" Harv. JL & Tech. 31 2017: 841.", "review_text": "This paper proposes a method to make conceptual interventions by modifying the activations of a pre-trained black-box model. Based on the popular counterfactual optimization approach by Wachter et al [1] for pixel-based counterfactuals, they propose to instead optimize for activations. That is, they apply the distance loss in the activation instead of pixel space and use the gradients obtained by an additionally trained concept detector ((non)-linear probe) instead of the gradients of the classifier (c.f., Eq. 1). Finally, the authors propose a fine-tuning scheme to make the classifier more reliant on the “concept activation vectors” based on a proposed notion of intervenability (Eq. 2 & 3), while keeping the feature extractor frozen. Through experiments on tabular and vision data the authors show the efficacy of their intervention as well as fine-tuning scheme.\n\n[1] Wachter, Sandra, et al. \"Counterfactual explanations without opening the black box: Automated decisions and the GDPR.\" Harv. JL & Tech. 31 2017: 841.", "strengths": "* The work addresses an interesting problem to better understand the behavior of models through conceptual interventions.\n* The intervention strategy is simple yet effective.\n* The fine-tuning scheme is well-designed and yet simple.\n* Code is provided via an anonymized repository and supplementary material.", "weaknesses": "* The work seems to have missed the most relevant work on conceptual (interventional) counterfactuals, e.g., [1,2]. While there are some technical differences (usage of the gradients stemming from the linear probe instead of the classifier in the counterfactual optimization problem), there is still significant overlap. For example, Abid et al. [1] also use linear probes to identify concept activation vectors and use it to intervene on the features.\n\n* The work motivates their approach by stating that “concept labels are not required [during training]” (p. 2). While true, it is still required for the fine-tuning (as we need to train the probes initially; Sec. 3.3, and shown to be important in the experiments), which one could also see as part of model development. For the case that we would like to make the same conceptual interventions as for CBMs, this would result in a similar amount of annotation cost.\n\n* The paper makes the (implicit) assumption that the feature extractor $h_{\\phi}$ learns the concept; both in their intervention approach and fine-tuning (since $\\beta$ is set to 1). However, the authors provide no evidence that this is actually the case. The linear probes could just learn to predict from some correlated concept/feature. Further evidence would be required that the black-box feature extractor has actually learned the concept that is intervened on.\n\n* It is very unclear whether the intervention strategy actually results in *plausible* and not just *adversarial* changes of the activations. This is a prominent problem for pixel-spaced counterfactuals methods, where, e.g., different types of regularization or generative models are used to obtain plausible counterfactuals.\n\n* There are no comparisons to prior work that converted pretrained models into CBMs [3,4]. It would be interesting to show how the proposed method compares to them (using the proposed intervention strategy).\n\n[1] Abid, Abubakar et al. \"Meaningfully explaining model mistakes using conceptual counterfactuals.\" ICML 2022.\n\n[2] Kim, Siwon, et al. \"Grounding Counterfactual Explanation of Image Classifiers to Textual Concept Space.\" CVPR 2023.\n\n[3] Yuksekgonul, Mert et al. \"Post-hoc concept bottleneck models.\" ICLR 2023.\n\n[4] Oikarinen, Tuomas, et al. \"Label-Free Concept Bottleneck Models.\" ICLR 2023.", "questions": "* Why does Eq. 2 & 3 assume that c’ does not change y to y’? As is, it assumes that the concept does not change the class. Let’s say the concept c’ changes the fur texture of a cat, then the resulting class may also change. This seems not to be included in the current notion of intervenability and may be easily obtained by a suitable choice of the distribution $\\pi(c’,y’|x,\\hat{c},c,\\hat{y},y)$, which subsumes $\\pi(c’|x,\\hat{c},c,\\hat{y},y)$ when $y’=y$.\n\n* Why is the ResNet-18 architecture used with four fully-connected layers instead of the standard setting? Why are not just the bottleneck features used?\n\n* What happens for the baseline (fine-tuned, MT) if we interleave intervened activations $z’$ also during training? As is, it just may be that (fine-tuned, I) is more robust to the interventions.\n\n* Are AUROC and AUPR computed for the concepts or targets/classes? This may also change for the different experimental results (e.g., for the figures in the main paper this is unclear). Could the authors clarify this?\n\n## Suggestions\n\n* It’d have been good to discuss the data-generating mechanisms (bottleneck, confounder, incomplete) in the main text and not only supplemental.\n\n* Given the overlapping confidence bands in Fig. 3(c) it would be good to either run more simulations or reformulate the sentence since it is unclear if “black-box classifiers are expectedly superior to the CBM”.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to make conceptual interventions by modifying the activations of a pre-trained black-box model. Based on the popular counterfactual optimization approach by Wachter et al [1] for pixel-based counterfactuals, they propose to instead optimize for activations. That is, they apply the distance loss in the activation instead of pixel space and use the gradients obtained by an additionally trained concept detector ((non)-linear probe) instead of the gradients of the classifier (c.f., Eq. 1). Finally, the authors propose a fine-tuning scheme to make the classifier more reliant on the “concept activation vectors” based on a proposed notion of intervenability (Eq. 2 & 3), while keeping the feature extractor frozen. Through experiments on tabular and vision data the authors show the efficacy of their intervention as well as fine-tuning scheme.\n\n[1] Wachter, Sandra, et al. \"Counterfactual explanations without opening the black box: Automated decisions and the GDPR.\" Harv. JL & Tech. 31 2017: 841.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* The work addresses an interesting problem to better understand the behavior of models through conceptual interventions.\n* The intervention strategy is simple yet effective.\n* The fine-tuning scheme is well-designed and yet simple.\n* Code is provided via an anonymized repository and supplementary material.", "weaknesses": "* The work seems to have missed the most relevant work on conceptual (interventional) counterfactuals, e.g., [1,2]. While there are some technical differences (usage of the gradients stemming from the linear probe instead of the classifier in the counterfactual optimization problem), there is still significant overlap. For example, Abid et al. [1] also use linear probes to identify concept activation vectors and use it to intervene on the features.\n\n* The work motivates their approach by stating that “concept labels are not required [during training]” (p. 2). While true, it is still required for the fine-tuning (as we need to train the probes initially; Sec. 3.3, and shown to be important in the experiments), which one could also see as part of model development. For the case that we would like to make the same conceptual interventions as for CBMs, this would result in a similar amount of annotation cost.\n\n* The paper makes the (implicit) assumption that the feature extractor $h_{\\phi}$ learns the concept; both in their intervention approach and fine-tuning (since $\\beta$ is set to 1). However, the authors provide no evidence that this is actually the case. The linear probes could just learn to predict from some correlated concept/feature. Further evidence would be required that the black-box feature extractor has actually learned the concept that is intervened on.\n\n* It is very unclear whether the intervention strategy actually results in *plausible* and not just *adversarial* changes of the activations. This is a prominent problem for pixel-spaced counterfactuals methods, where, e.g., different types of regularization or generative models are used to obtain plausible counterfactuals.\n\n* There are no comparisons to prior work that converted pretrained models into CBMs [3,4]. It would be interesting to show how the proposed method compares to them (using the proposed intervention strategy).\n\n[1] Abid, Abubakar et al. \"Meaningfully explaining model mistakes using conceptual counterfactuals.\" ICML 2022.\n\n[2] Kim, Siwon, et al. \"Grounding Counterfactual Explanation of Image Classifiers to Textual Concept Space.\" CVPR 2023.\n\n[3] Yuksekgonul, Mert et al. \"Post-hoc concept bottleneck models.\" ICLR 2023.\n\n[4] Oikarinen, Tuomas, et al. \"Label-Free Concept Bottleneck Models.\" ICLR 2023.", "questions": "* Why does Eq. 2 & 3 assume that c’ does not change y to y’? As is, it assumes that the concept does not change the class. Let’s say the concept c’ changes the fur texture of a cat, then the resulting class may also change. This seems not to be included in the current notion of intervenability and may be easily obtained by a suitable choice of the distribution $\\pi(c’,y’|x,\\hat{c},c,\\hat{y},y)$, which subsumes $\\pi(c’|x,\\hat{c},c,\\hat{y},y)$ when $y’=y$.\n\n* Why is the ResNet-18 architecture used with four fully-connected layers instead of the standard setting? Why are not just the bottleneck features used?\n\n* What happens for the baseline (fine-tuned, MT) if we interleave intervened activations $z’$ also during training? As is, it just may be that (fine-tuned, I) is more robust to the interventions.\n\n* Are AUROC and AUPR computed for the concepts or targets/classes? This may also change for the different experimental results (e.g., for the figures in the main paper this is unclear). Could the authors clarify this?\n\n## Suggestions\n\n* It’d have been good to discuss the data-generating mechanisms (bottleneck, confounder, incomplete) in the main text and not only supplemental.\n\n* Given the overlapping confidence bands in Fig. 3(c) it would be good to either run more simulations or reformulate the sentence since it is unclear if “black-box classifiers are expectedly superior to the CBM”.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697982196422}], "openreview_url": "https://openreview.net/forum?id=5oJlyJXUxK", "arxiv_id": "2401.13544", "paper_pdf": "papers/5oJlyJXUxK.pdf", "paper_pdf_sha256": "3a61726319456c88ec4d3597a1501751c4203d32b858a5e7804ae5c79593b348", "paper_pdf_bytes": 12053856, "paper_pdf_source": "openreview", "code_url": "https://github.com/sonialagunac/Beyond-CBM", "code_repository": "sonialagunac/Beyond-CBM", "code_commit": "d1c24659496558c2a79b9b51bf9077a2a157f12e", "code_archive": "repos/5oJlyJXUxK.zip", "code_archive_sha256": "e854cc92cac9672fde19b5f0c07895d10ea6f8c0df5b10a27c115a1dcca3948f", "code_archive_bytes": 69894, "code_file_count": 25, "code_extensions": {".py": 25}, "github_disk_usage_kb": 56, "github_languages": {"Python": 179737, "Shell": 1928}, "github_archived": false, "github_pushed_at": "2024-10-26T12:15:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beyond-concept-bottleneck-models-how-to-make"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lYZZl2hp1gp", "year": 2023, "status": "rejected", "title": "Robustness of Unsupervised Representation Learning without Labels", "authors": ["Aleksandar Petrov", "Marta Kwiatkowska"], "authorids": ["~Aleksandar_Petrov1", "~Marta_Kwiatkowska1"], "authors_source": "OpenReview API", "abstract": "Unsupervised representation learning leverages large unlabeled datasets and is competitive with supervised learning. But non-robust encoders may affect downstream task robustness. Recently, robust representation encoders have become of interest. Still, all prior work evaluates robustness using a downstream classification task. Instead, we propose a family of unsupervised robustness measures, which are model- and task-agnostic and label-free. We benchmark state-of-the-art representation encoders and show that none dominates the rest. We offer unsupervised extensions to the FGSM and PGD attacks. When used in adversarial training, they improve most unsupervised robustness measures, including certified robustness. We validate our results against a linear probe and show that, for MOCOv2, adversarial training results in 3 times higher certified accuracy, a 2-fold decrease in impersonation attack success rate and considerable improvements in certified robustness.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "l902qlnji3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4058/Reviewer_rApq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes several evaluations of unsupervised robustness which are model agnostic and tasks agnostic. This paper examines recent unsupervised models with proposed unsupervised robustness evaluations. Specifically, the paper proposes universal quantiles for untargeted attacks and relative quantiles for targeted attacks as an evaluation metric for unsupervised robustness. The former metric represents the percentage of paired samples that are closer than the clean and adversarial pairs. The latter metric represents the ratio that the targeted attack moves the original image to the target image.", "review_text": "Overall, I recommend marginally below the acceptance threshold. Because whether the proposed evaluation metric could represent robustness seems unclear I hope the authors could resolve my concerns in the rebuttal.", "strengths": "**Strength**\n- This paper first tackles the problem that previous unsupervised adversarial methods only evaluated the classification task as a downstream task. To overcome such limitations, this paper proposes evaluation metrics that do not use any class labels.\n- The idea and the approach to making the evaluation metric for unsupervised robustness are quite novel to me.\n\n**Weakness**\n- I am not quite sure why proposed universal quantiles and relative quantiles are able to represent the robustness of the models. Intuitively, robust models could have a relatively low ratio of universal quantiles but these metrics could represent the vulnerability that induces wrong decisions or unintended actions in downstream tasks. I think the authors could describe how we can interpret these metrics in terms of robustness.\n- I think there is no big difference between U-PGD and L-PGD since the distance function is also a loss function. Further, KL divergence as a distance function could not be used in unsupervised models since there is no class probability in the unsupervised models. I might have missed but what kind of distance function is used for evaluation metric? Moreover, it seems that the L-PGD is already proposed in the previous unsupervised adversarial learning which seems to lack novelty.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes several evaluations of unsupervised robustness which are model agnostic and tasks agnostic. This paper examines recent unsupervised models with proposed unsupervised robustness evaluations. Specifically, the paper proposes universal quantiles for untargeted attacks and relative quantiles for targeted attacks as an evaluation metric for unsupervised robustness. The former metric represents the percentage of paired samples that are closer than the clean and adversarial pairs. The latter metric represents the ratio that the targeted attack moves the original image to the target image.", "strength_and_weaknesses": "**Strength**\n- This paper first tackles the problem that previous unsupervised adversarial methods only evaluated the classification task as a downstream task. To overcome such limitations, this paper proposes evaluation metrics that do not use any class labels.\n- The idea and the approach to making the evaluation metric for unsupervised robustness are quite novel to me.\n\n**Weakness**\n- I am not quite sure why proposed universal quantiles and relative quantiles are able to represent the robustness of the models. Intuitively, robust models could have a relatively low ratio of universal quantiles but these metrics could represent the vulnerability that induces wrong decisions or unintended actions in downstream tasks. I think the authors could describe how we can interpret these metrics in terms of robustness.\n- I think there is no big difference between U-PGD and L-PGD since the distance function is also a loss function. Further, KL divergence as a distance function could not be used in unsupervised models since there is no class probability in the unsupervised models. I might have missed but what kind of distance function is used for evaluation metric? Moreover, it seems that the L-PGD is already proposed in the previous unsupervised adversarial learning which seems to lack novelty.\n", "clarity,_quality,_novelty_and_reproducibility": "**Clarity:** The paper is easy to understand.\n\n**Quality:** The representation of the paper seems fine. However, I hope the authors describe which dataset is used in each table in the caption.\n\n**Novelty:** The problem and the metric that the paper demonstrates are novel to me.\n\n**Reproducibility:** The paper has well reproducibility which elaborates well on the details.", "summary_of_the_review": "Overall, I recommend marginally below the acceptance threshold. Because whether the proposed evaluation metric could represent robustness seems unclear I hope the authors could resolve my concerns in the rebuttal.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666699203900}, {"id": "6yX9Rfwa4jO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4058/Reviewer_v4fr"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper tackles the problem of quantifying and improving the robustness of representation models in a task-agnostic fashion. Being one of the first papers to approach the problem, the authors also motivate and provide a mathematical definition for unsupervised robustness. To evaluate the unsupervised robustness of the representation models, the authors propose a generalized attack framework and use it to propose a set of metrics representation-agnostic metrics that can be used to measure the relative robustness of various representation models. Using the attacks, the authors also train adversarially robust representation models and provide a detailed empirical analysis of the relative performance of different representation models with respect to each other and adversarially trained models.", "review_text": "The paper points out and addresses an important problem for representation models. The paper approaches the problem by considering two possible sources of risk: breakaway risk and overlap risk. Then the authors use this to motivate various metrics for measuring the unsupervised robustness of the classifier. However, the later analysis does not use the estimated values of the two risks. So, I feel some of the empirical investigations are not well motivated. Some of the presented empirical results also need further investigation. Especially the impersonation bounds need to be recalculated with better attacks. The paper has some great ideas which could benefit the community, but I think it could use some rewriting.", "strengths": "Strengths\n- The paper brings up an essential issue of the robustness of unsupervised representation models and identifies some novel task-independent metrics to measure their robustness.\n- The quantile-based metrics proposed in the paper give a representation-agnostic view of robustness that allows users to compare multiple representation models.\n- The authors also propose some adversarial training methods that improve the randomized smoothing-based certified robustness of the trained models. \n\nWeaknesses\n- Although the empirical attack evidence suggests that the adversarially trained representation models require more PGD iterations, this can also be explained by gradient masking. Moreover, these benefits are also not present for MOCOv3 models. Using a different attack method for the evaluation would resolve this issue.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper tackles the problem of quantifying and improving the robustness of representation models in a task-agnostic fashion. Being one of the first papers to approach the problem, the authors also motivate and provide a mathematical definition for unsupervised robustness. To evaluate the unsupervised robustness of the representation models, the authors propose a generalized attack framework and use it to propose a set of metrics representation-agnostic metrics that can be used to measure the relative robustness of various representation models. Using the attacks, the authors also train adversarially robust representation models and provide a detailed empirical analysis of the relative performance of different representation models with respect to each other and adversarially trained models.", "strength_and_weaknesses": "Strengths\n- The paper brings up an essential issue of the robustness of unsupervised representation models and identifies some novel task-independent metrics to measure their robustness.\n- The quantile-based metrics proposed in the paper give a representation-agnostic view of robustness that allows users to compare multiple representation models.\n- The authors also propose some adversarial training methods that improve the randomized smoothing-based certified robustness of the trained models. \n\nWeaknesses\n- Although the empirical attack evidence suggests that the adversarially trained representation models require more PGD iterations, this can also be explained by gradient masking. Moreover, these benefits are also not present for MOCOv3 models. Using a different attack method for the evaluation would resolve this issue.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper addresses an important problem and proposes a novel approach to it. As some of the metrics are presented in a new format, the authors need to take special care to remind the reader of the quantile version of the bounds. It is sometimes hard to follow the paper, especially the certified robustness bounds presented as quantiles in the tables. The authors do provide all the required details to reproduce the work. ", "summary_of_the_review": "The paper points out and addresses an important problem for representation models. The paper approaches the problem by considering two possible sources of risk: breakaway risk and overlap risk. Then the authors use this to motivate various metrics for measuring the unsupervised robustness of the classifier. However, the later analysis does not use the estimated values of the two risks. So, I feel some of the empirical investigations are not well motivated. Some of the presented empirical results also need further investigation. Especially the impersonation bounds need to be recalculated with better attacks. The paper has some great ideas which could benefit the community, but I think it could use some rewriting.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666679685345}, {"id": "jO5iawtqU1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4058/Reviewer_TA8W"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes two unsupervised adversarial risks, i.e., breakaway risk and overlap risk, that evaluate the adversarial robustness without requiring labels. Further, this paper generates unsupervised adversarial attacks via FGSM and PGD maximizes or minimizes the distance between benign and adversarial data. The authors also propose to use adversarial training based on the unsupervised FGSM or PGD to improve the robustness. Then, this paper provides a series of robustness measurements and empirically shows that adversarial-trained models are more robust than standard-trained models.", "review_text": "This paper proposes two unsupervised adversarial risks and realizes them based on PGD and conventional distance metrics. However, I have several concerns that have been illustrated in Weaknesses section to be solved. ", "strengths": "Strength\n+ This paper proposes two interesting unsupervised adversarial risks based on the distance between benign and adversarial data. The risks indeed solve the problem that what value of the distance being “small”.  \n+ This paper gives several robustness metrics, including quantiles for (un)targeted attacks, estimation of breakaway risk and overlap risk, nearest neighbour accuracy, adversarial margin, and certified robustness.\n\nWeaknesses:\n- This paper seems to have minor technical novelty. The main techniques of adversarial attacks and adversarial training are based on PGD or FGSM and even the loss function is simply the conventional distance.\n- The relationship (such as differences and similarities) between breakaway risk and overlap risk is not very clear. I am confused about which risk really evaluates unsupervised adversarial robustness. \n- It is somewhat difficult to understand how the metric “universal quantiles for untargeted attacks” can evaluate robustness. This metric closely depends on the sampling procedure. It is hard to say a model is non-robust if the fraction of distance between $x’$ and $x’’$ is smaller than the distance between $\\hat{x}$ and $x$ is smaller since it could be incurred by the sampling procedure.\n- The claim that “the breakaway risk can be very small” is somewhat confusing. Due to this claim, the author proposes nearest neighbour accuracy. However, Table 2 seems to show that breakaway risk can better differentiate the robustness of each model than nearest neighbour accuracy since nearest neighbour accuracy is almost very low and the same among different models.\n- It seems that different unsupervised adversarial metrics do not provide a consistent evaluation. For example, in Table 2, ResNet50 has a lower breakaway risk while a higher overlap risk and PixPro has a higher breakaway risk while a lower overlap risk. Therefore, it is hard to compare the unsupervised robustness between these models using these two proposed metrics.\n- It is not very clear the relationship between unsupervised robustness evaluation and supervised robustness evaluation. Will the unsupervised robust accuracy and supervised robust accuracy be positively correlated between them?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes two unsupervised adversarial risks, i.e., breakaway risk and overlap risk, that evaluate the adversarial robustness without requiring labels. Further, this paper generates unsupervised adversarial attacks via FGSM and PGD maximizes or minimizes the distance between benign and adversarial data. The authors also propose to use adversarial training based on the unsupervised FGSM or PGD to improve the robustness. Then, this paper provides a series of robustness measurements and empirically shows that adversarial-trained models are more robust than standard-trained models.", "strength_and_weaknesses": "Strength\n+ This paper proposes two interesting unsupervised adversarial risks based on the distance between benign and adversarial data. The risks indeed solve the problem that what value of the distance being “small”.  \n+ This paper gives several robustness metrics, including quantiles for (un)targeted attacks, estimation of breakaway risk and overlap risk, nearest neighbour accuracy, adversarial margin, and certified robustness.\n\nWeaknesses:\n- This paper seems to have minor technical novelty. The main techniques of adversarial attacks and adversarial training are based on PGD or FGSM and even the loss function is simply the conventional distance.\n- The relationship (such as differences and similarities) between breakaway risk and overlap risk is not very clear. I am confused about which risk really evaluates unsupervised adversarial robustness. \n- It is somewhat difficult to understand how the metric “universal quantiles for untargeted attacks” can evaluate robustness. This metric closely depends on the sampling procedure. It is hard to say a model is non-robust if the fraction of distance between $x’$ and $x’’$ is smaller than the distance between $\\hat{x}$ and $x$ is smaller since it could be incurred by the sampling procedure.\n- The claim that “the breakaway risk can be very small” is somewhat confusing. Due to this claim, the author proposes nearest neighbour accuracy. However, Table 2 seems to show that breakaway risk can better differentiate the robustness of each model than nearest neighbour accuracy since nearest neighbour accuracy is almost very low and the same among different models.\n- It seems that different unsupervised adversarial metrics do not provide a consistent evaluation. For example, in Table 2, ResNet50 has a lower breakaway risk while a higher overlap risk and PixPro has a higher breakaway risk while a lower overlap risk. Therefore, it is hard to compare the unsupervised robustness between these models using these two proposed metrics.\n- It is not very clear the relationship between unsupervised robustness evaluation and supervised robustness evaluation. Will the unsupervised robust accuracy and supervised robust accuracy be positively correlated between them?\n", "clarity,_quality,_novelty_and_reproducibility": "The organization of this paper is not very good. It is somewhat weird to suddenly illustrate adversarial training which is a defensive strategy in Section 4 of unsupervised attacks. This paper proposes two unsupervised adversarial risks and realizes them based on PGD and conventional distance metrics. But it seems that the metric is not consistent. Therefore, the technical quality seems to be not very good and the novelty is somewhat fair. The author provides algorithm and experimental details. Thus, this paper has good reproducibility.  ", "summary_of_the_review": "This paper proposes two unsupervised adversarial risks and realizes them based on PGD and conventional distance metrics. However, I have several concerns that have been illustrated in Weaknesses section to be solved. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666601076606}, {"id": "5ZiSYi3uJDu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4058/Reviewer_HYfP"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this work, the authors define what it means for an encoder to be robust and describe metrics for measuring robustness.  Using these metrics, they then evaluate existing SOTA representation encoders with and without adversarial finetuning (based on the proposed metrics).", "review_text": "I find this paper very interesting as it rigorously defines what it means for an encoder to be robust, and I think that the proposed metrics will be very useful to the adversarial ML research community especially since recently there have been works proposing methods for robust pretraining.  I am a little disappointed by the experiments though, not that they are bad or uninteresting, I just feel like the authors introduced a very powerful toolkit but did not use this toolkit to its full potential.", "strengths": "Strengths:\n- Definitions for breakaway risk and overlap risk are well-motivated and easy to understand.\n- Interesting problem setting (measuring robustness without labels)\n- Well-motivated metrics\n- Show that fine tuning based on some of the proposed metrics can greatly improve robustness of encoder\n- Writing is clear\n\nWeaknesses:\n- The proposed metrics could be used to analyze encoders trained using any technique: supervised learning, self-supervised learning, and unsupervised learning, but the experiments are restricted to unsupervised encoders.  I think it would be interesting to perform evaluations on encoders of adversarially trained models (supervised learning but drop the final classifier layer) and see if these learned representations actually satisfy the definitions proposed.  Additionally, it would be interesting to see whether the pretraining approaches proposed by works for adversarially robust self-supervised learning (ie. Kim et al 2020) lead to more robust encoders.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this work, the authors define what it means for an encoder to be robust and describe metrics for measuring robustness.  Using these metrics, they then evaluate existing SOTA representation encoders with and without adversarial finetuning (based on the proposed metrics).", "strength_and_weaknesses": "Strengths:\n- Definitions for breakaway risk and overlap risk are well-motivated and easy to understand.\n- Interesting problem setting (measuring robustness without labels)\n- Well-motivated metrics\n- Show that fine tuning based on some of the proposed metrics can greatly improve robustness of encoder\n- Writing is clear\n\nWeaknesses:\n- The proposed metrics could be used to analyze encoders trained using any technique: supervised learning, self-supervised learning, and unsupervised learning, but the experiments are restricted to unsupervised encoders.  I think it would be interesting to perform evaluations on encoders of adversarially trained models (supervised learning but drop the final classifier layer) and see if these learned representations actually satisfy the definitions proposed.  Additionally, it would be interesting to see whether the pretraining approaches proposed by works for adversarially robust self-supervised learning (ie. Kim et al 2020) lead to more robust encoders.", "clarity,_quality,_novelty_and_reproducibility": "Overall, I think the writing and presentation is very clear.  I also think that the work is novel as it is the first work to rigorously define what properties we would expect a robust encoder to follow and propose metrics for measuring the robustness of an encoder.  Overall, I think that the metrics are well-motivated and can be very useful for assessing the robustness of current training techniques for adversarial ML.", "summary_of_the_review": "I find this paper very interesting as it rigorously defines what it means for an encoder to be robust, and I think that the proposed metrics will be very useful to the adversarial ML research community especially since recently there have been works proposing methods for robust pretraining.  I am a little disappointed by the experiments though, not that they are bad or uninteresting, I just feel like the authors introduced a very powerful toolkit but did not use this toolkit to its full potential.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666297338066}], "openreview_url": "https://openreview.net/forum?id=lYZZl2hp1gp", "arxiv_id": "2210.04076", "paper_pdf": "papers/lYZZl2hp1gp.pdf", "paper_pdf_sha256": "372ad962d3643bff026649ca555969f5445934e26af1f142dfafe8f2cec8f3c0", "paper_pdf_bytes": 24651918, "paper_pdf_source": "openreview", "code_url": "https://github.com/AleksandarPetrov/unsupervised-robustness", "code_repository": "AleksandarPetrov/unsupervised-robustness", "code_commit": "b8606562c397516c2e68d9318861d0cd6dbbf312", "code_archive": "repos/lYZZl2hp1gp.zip", "code_archive_sha256": "127cb0ab9f20fcfa0c26c89f6a690946e420f5d7b0dd67681622f0533aeb938d", "code_archive_bytes": 117639, "code_file_count": 39, "code_extensions": {".py": 34, ".sh": 5}, "github_disk_usage_kb": 97, "github_languages": {"Python": 263016, "Shell": 7195}, "github_archived": false, "github_pushed_at": "2022-10-18T05:29:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/robustness-of-unsupervised-representation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tG8QrhMwEqS", "year": 2022, "status": "rejected", "title": "Adaptive Activation-based Structured Pruning", "authors": ["Kaiqi Zhao", "Animesh Jain", "Ming Zhao"], "authorids": ["~Kaiqi_Zhao2", "~Animesh_Jain1", "~Ming_Zhao2"], "authors_source": "OpenReview API", "abstract": "Pruning is a promising approach to compress complex deep learning models in order to deploy them on resource-constrained edge devices. However, many existing pruning solutions are based on unstructured pruning, which yield models that cannot efficiently run on commodity hardware, and require users to manually explore and tune the pruning process, which is time consuming and often leads to sub-optimal results. To address these limitations, this paper presents an adaptive, activation-based, structured pruning approach to automatically and efficiently generate small, accurate, and hardware-efficient models that meet user requirements. First, it proposes iterative structured pruning using activation-based attention feature maps to effectively identify and prune unimportant filters. Then, it proposes adaptive pruning policies for automatically meeting the pruning objectives of accuracy-critical, memory-constrained, and latency-sensitive tasks. A comprehensive evaluation shows that the proposed method can substantially outperform the state-of-the-art structured pruning works on CIFAR-10 and ImageNet datasets. For example, on ResNet-56 with CIFAR-10, without any accuracy drop, our method achieves the largest parameter reduction (79.11%), outperforming the related works by 22.81% to 66.07%, and the largest FLOPs reduction (70.13%), outperforming the related works by 14.13% to 26.53%.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "0edlJDXef-_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper61/Reviewer_EiLE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed an activation-based, adaptive threshold, iterative structured pruning method, combining several existing techniques together to perform a comprehensive pruning. ", "review_text": "Strengths:\nThe proposed method has relatively good results on cifar10 and imagenet.  The method can automatically meet users' several requirements generally, like accuracy, latency, memory, and so on. \n\nWeakness:\n1. Each part of the technique is not that novel. It is like a combination of several existing tricks to perform comprehensive pruning, not enough analysis of the convergence or correctness. For example, how to guarantee the convergence of algorithm 3? \n2. The comparisons on Imagenet are too weak. There are many SOTA pruning methods on Imagenet. The paper only lists four of them. I suggest comparing the paper with more recent SOTA pruning papers and comparing on more benchmark models besides Res50. \nfor example, DMCP: Differentiable Markov Channel Pruning for Neural Networks\neagleeye: fast sub-net evaluation for efficient neural network pruning. \ngdp: neural network pruning via gates with differentiable polarization.\nand so on. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed an activation-based, adaptive threshold, iterative structured pruning method, combining several existing techniques together to perform a comprehensive pruning. ", "main_review": "Strengths:\nThe proposed method has relatively good results on cifar10 and imagenet.  The method can automatically meet users' several requirements generally, like accuracy, latency, memory, and so on. \n\nWeakness:\n1. Each part of the technique is not that novel. It is like a combination of several existing tricks to perform comprehensive pruning, not enough analysis of the convergence or correctness. For example, how to guarantee the convergence of algorithm 3? \n2. The comparisons on Imagenet are too weak. There are many SOTA pruning methods on Imagenet. The paper only lists four of them. I suggest comparing the paper with more recent SOTA pruning papers and comparing on more benchmark models besides Res50. \nfor example, DMCP: Differentiable Markov Channel Pruning for Neural Networks\neagleeye: fast sub-net evaluation for efficient neural network pruning. \ngdp: neural network pruning via gates with differentiable polarization.\nand so on. \n\n", "summary_of_the_review": "Due to the concerns on the novelty and the weak comparisons on Imagenet. I temporarily think this paper is marginally below the acceptance threshold. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635878356906}, {"id": "747N5kDrz7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper61/Reviewer_YdeV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a technique for iterative structured pruning, without necessarily requiring too much manual human intervention. There are two parts to this paper that are important:\n\n1. It is argued that we should prune channels based on the activation maps generated, rather than focusing on the weights of the channel.\n2. They propose an iterative procedure that automatically backtracks if it has made a poor pruning decision.", "review_text": "I did not like the way that this paper is written. It is extremely waffley, and is not upfront about what is going on. It is not completely clear (to me at least) what the overall strategy for pruning will be in this work until half way through page 4 (!). The paper needs to be (completely) rewritten to match the expected format in the computer science literature.\n\n**Strengths**\n- I am not a total expert on the pruning literature, but I do find the idea of pruning based on the activation maps rather than the weights intuitive and sensible.\n- The results are compelling relative to other works they compare to. NOTE: again, I will not claim to be an absolute expert on the state of the pruning literature.\n- The automated approach is relatively simple, and far more interpretable than approaches such as AMC. This matters a lot in many real-world applications.\n- ImageNet results are included, which is relatively rare for pruning papers.\n\n**Weaknesses**:\n- The writing of this paper is unacceptable, as I mentioned earlier.\n- A few really important ablation studies are missing. Firstly, there is no direct reason to couple the pruning technique, with the method for iteratively pruning; can these not be compared separately. It is misleading to conflate the two. Secondly, there is no comment on runtime for this method, which is important if this is an automated technique.\n- Another important issue is that the FLOPs reduction doesn't seem that impressive. Works like Eigendamage (Wang et al.) achieve far more impressive reductions in FLOPs, which is arguably far more important than parameters in the real world. As an aside, would you be able to compare to Eigendamage?\n- The work only focuses on ResNets. I know that they're harder to prune, but I'd also like to see some experiments on VGG models, for example.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes a technique for iterative structured pruning, without necessarily requiring too much manual human intervention. There are two parts to this paper that are important:\n\n1. It is argued that we should prune channels based on the activation maps generated, rather than focusing on the weights of the channel.\n2. They propose an iterative procedure that automatically backtracks if it has made a poor pruning decision.", "main_review": "I did not like the way that this paper is written. It is extremely waffley, and is not upfront about what is going on. It is not completely clear (to me at least) what the overall strategy for pruning will be in this work until half way through page 4 (!). The paper needs to be (completely) rewritten to match the expected format in the computer science literature.\n\n**Strengths**\n- I am not a total expert on the pruning literature, but I do find the idea of pruning based on the activation maps rather than the weights intuitive and sensible.\n- The results are compelling relative to other works they compare to. NOTE: again, I will not claim to be an absolute expert on the state of the pruning literature.\n- The automated approach is relatively simple, and far more interpretable than approaches such as AMC. This matters a lot in many real-world applications.\n- ImageNet results are included, which is relatively rare for pruning papers.\n\n**Weaknesses**:\n- The writing of this paper is unacceptable, as I mentioned earlier.\n- A few really important ablation studies are missing. Firstly, there is no direct reason to couple the pruning technique, with the method for iteratively pruning; can these not be compared separately. It is misleading to conflate the two. Secondly, there is no comment on runtime for this method, which is important if this is an automated technique.\n- Another important issue is that the FLOPs reduction doesn't seem that impressive. Works like Eigendamage (Wang et al.) achieve far more impressive reductions in FLOPs, which is arguably far more important than parameters in the real world. As an aside, would you be able to compare to Eigendamage?\n- The work only focuses on ResNets. I know that they're harder to prune, but I'd also like to see some experiments on VGG models, for example.", "summary_of_the_review": "I think this work has some merit, but it is not ready for publication at this time. I can be swayed at rebuttal time if there is convincing new experimental evidence.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635854075906}, {"id": "cLdBWCliQcN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper61/Reviewer_XNxE"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes iterative structured pruning methods using activation-based attention feature maps and an adaptive threshold selection strategy. Inspired by attention transfer, Activation-based attention feature maps are constructed as the important evaluation of filters in each layer. Adaptive threshold selection strategy decides the number of removed filters, which satisfies one of three adaptive pruning policies. Experimental results on CIFAR-10 and ImageNet with ResNet architectures show performance gains over several state-of-the-art methods.", "review_text": "Strength: \n+ The proposed method is simple and easy to follow.\n+ This paper proposes three modes for structured pruning, including accuracy-guaranteed adaptive pruning, memory-constrained adaptive pruning, and FLOPs-constrained adaptive pruning. \n\nWeakness:\n- The contribution of the evaluation for the importance of filter, since activation-based attention maps are similar to attention score in [Sergey Zagoruyko, ICLR 16] and the activation-based score is also proposed in many related works, such as APoZ [1] and HRank. Please clarify the difference between these works and the proposed methods.\n- The rationale of calculating the threshold T in each layer. T in layer-aware threshold adjustment is directly related to the percentage of each layer on parameters or FLOPs. Moreover, the typo in line 3 of algorithm N^i[r]->T^i[r].\n- Lack of some important experimental evaluation: 1. The models selected to compress are all based resnet architectures, how about other architectures, especially light backbones (e.g. mobilenets); 2. The effect of attention score and pruning strategy should be discussed in ablation study; 3. The actual speedup of the pruned models should be evaluated if minimizing FLOPs.\n\n[1] Network trimming: A data-driven neuron pruning approach towards efficient deep architectures, arXiv preprint arXiv:1607.03250, 2016.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes iterative structured pruning methods using activation-based attention feature maps and an adaptive threshold selection strategy. Inspired by attention transfer, Activation-based attention feature maps are constructed as the important evaluation of filters in each layer. Adaptive threshold selection strategy decides the number of removed filters, which satisfies one of three adaptive pruning policies. Experimental results on CIFAR-10 and ImageNet with ResNet architectures show performance gains over several state-of-the-art methods.", "main_review": "Strength: \n+ The proposed method is simple and easy to follow.\n+ This paper proposes three modes for structured pruning, including accuracy-guaranteed adaptive pruning, memory-constrained adaptive pruning, and FLOPs-constrained adaptive pruning. \n\nWeakness:\n- The contribution of the evaluation for the importance of filter, since activation-based attention maps are similar to attention score in [Sergey Zagoruyko, ICLR 16] and the activation-based score is also proposed in many related works, such as APoZ [1] and HRank. Please clarify the difference between these works and the proposed methods.\n- The rationale of calculating the threshold T in each layer. T in layer-aware threshold adjustment is directly related to the percentage of each layer on parameters or FLOPs. Moreover, the typo in line 3 of algorithm N^i[r]->T^i[r].\n- Lack of some important experimental evaluation: 1. The models selected to compress are all based resnet architectures, how about other architectures, especially light backbones (e.g. mobilenets); 2. The effect of attention score and pruning strategy should be discussed in ablation study; 3. The actual speedup of the pruned models should be evaluated if minimizing FLOPs.\n\n[1] Network trimming: A data-driven neuron pruning approach towards efficient deep architectures, arXiv preprint arXiv:1607.03250, 2016.\n", "summary_of_the_review": "Although the proposed method is simple yet effective on ResNets, the contribution is limited and experimental evaluation is not comprehensive. I tend to reject this paper in this version.\n\n\n-------------------POST-REBUTTAL COMMENTS-------------------\n\nI thank the authors for the response and the efforts in the updated draft. After this rebuttal, the authors well answer my comments about the comparison to SOTA methods. However, I still believe that the proposed measurement of filter importance is a limited novelty, as it is a simple revision of attention score from [Sergey Zagoruyko, ICLR 16]. Moreover, the rationale of threshold T is based on the assumption that a layer is more likely to have redundant filters to prune if it contains more remaining parameters. I think it is not a correct assumption, as the entire filters of a layer with a smaller number of parameters may be redundant to be removed safely, especially for ResNets. Thus, I still keep my original rate to reject it.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635753449869}], "openreview_url": "https://openreview.net/forum?id=tG8QrhMwEqS", "arxiv_id": "2201.10520", "paper_pdf": "papers/tG8QrhMwEqS.pdf", "paper_pdf_sha256": "929e7656e34e04e658a6f1be45b178e0f979f6023e9e8d12ac6bafabd11d0d21", "paper_pdf_bytes": 3102856, "paper_pdf_source": "openreview", "code_url": "https://github.com/kaiqi123/Automatic-Attention-Pruning", "code_repository": "kaiqi123/Automatic-Attention-Pruning", "code_commit": "96f71d3eec651097b1c09dc0cb8412db94c3c8a6", "code_archive": "repos/tG8QrhMwEqS.zip", "code_archive_sha256": "ebf13c14f4ba78b14d0d12e3518b4a244bc6347eb361321e43c45b336952dce1", "code_archive_bytes": 56223, "code_file_count": 26, "code_extensions": {".py": 18, ".sh": 8}, "github_disk_usage_kb": 107, "github_languages": {"Python": 176097, "Shell": 9581, "Dockerfile": 228}, "github_archived": false, "github_pushed_at": "2023-10-25T18:07:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-activation-based-structured-pruning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "edku48LG0pT", "year": 2021, "status": "rejected", "title": "A Neural Network MCMC sampler that maximizes Proposal Entropy", "authors": ["ZENGYI LI", "Yubei Chen", "Friedrich Sommer"], "authorids": ["~ZENGYI_LI1", "~Yubei_Chen1", "~Friedrich_Sommer1"], "authors_source": "OpenReview API", "abstract": "Markov Chain Monte Carlo (MCMC) methods sample from unnormalized probability distributions and offer guarantees of exact sampling. However, in the continuous case, unfavorable geometry of the target distribution can greatly limit the efficiency of MCMC methods. Augmenting samplers with neural networks can potentially improve their efficiency. Previous neural network based samplers were trained with objectives that either did not explicitly encourage exploration, or used a L2 jump objective which could only be applied to well structured distributions. Thus it seems promising to instead maximize the proposal entropy for adapting the proposal to distributions of any shape. To allow direct optimization of the proposal entropy, we propose a neural network MCMC sampler that has a flexible and tractable proposal distribution. Specifically, our network architecture utilizes the gradient of the target distribution for generating proposals. Our model achieves significantly higher efficiency than previous neural network MCMC techniques in a variety of sampling tasks. Further, the sampler is applied on training of a convergent energy-based model of natural images. The learned sampler achieves significantly higher proposal entropy and sample quality compared to Langevin dynamics sampler.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1wc8pLPBa4z", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2093/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new MCMC transition kernel. This kernel is parameterized by neural networks and is optimized through an objective maximizing the proposal entropy. Specifically, the authors use a combination of a flow model and non-volume preserving flow in [Dinh et al., 2016] as the neural network parameterized kernel. Then they use the objective in [Titsias & Dellaportas, 2019] which maximizes the proposal entropy to optimize the kernel. The proposed method is tested on synthetic datasets, Bayesian logistic regression and a deep energy-based model.\n\nThe problem of improving the exploration efficiency of MCMC kernel is important. The proposed method is well-motivated. As far as I understand, the proposed method appears to be technically sound. \n\nHowever, I have the following concerns about the paper.\n\n-\tThe connection and the difference to previous work are not very clear. If I understand it correctly, the proposed method seems a combination of L2HMC and [Titsias & Dellaportas, 2019] with some slight modification (a flow model) since the naïve combination did not work well (as stated in Section 4). I think it would improve the clarity a lot if the authors explain more clearly how the proposed method differs from previous work. \n-\tSince the use of a flow model is the main difference compared to the naïve combination of two previous methods, the authors should explain more about this choice. Currently, it is not clear why this helps and there is no explanation on why the naïve combination fails. \n-\tThe proposed method seems to use more neural networks (e.g. an additional network R to handle gradient) than previous neural network MCMC. I wonder how the method performs if considering the cost. For example, the authors may instead show ESS per second on the experiments in Section 5.1. Though the authors mentioned the difficulty of computing ESS per second in the paper, I’m not entirely convinced. As the experiments in Section 5.1 are all very small-scale, is it really necessary to use GPUs? \n\n-\tThe baselines vary from experiments to experiments for no reason. For example, the authors compare their method to L2HMC on ill-conditioned Gaussian and strongly correlated Gaussian, to Neutra on Funnel distribution, and to MALA on EBM. I think this experiment design needs explanation. Also, there is no empirical comparison to [Titsias & Dellaportas, 2019] which is closely related to the proposed method.\n\nSome minor comments:\n\n-\tA.1 intends to show the benefit of using gradient information. But Variant 2 also uses gradient. What is the point of showing it? \n-\tIt is not clear to me how to interpret the empirical results in Section 5.2. For example, how does Figure 3 show that the proposed method needs less sampling steps? \n-\tThe color of points in figure 1 is hard to read.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper proposes a new MCMC transition kernel. This kernel is parameterized by neural networks and is optimized through an objective maximizing the proposal entropy. Specifically, the authors use a combination of a flow model and non-volume preserving flow in [Dinh et al., 2016] as the neural network parameterized kernel. Then they use the objective in [Titsias & Dellaportas, 2019] which maximizes the proposal entropy to optimize the kernel. The proposed method is tested on synthetic datasets, Bayesian logistic regression and a deep energy-based model.\n\nThe problem of improving the exploration efficiency of MCMC kernel is important. The proposed method is well-motivated. As far as I understand, the proposed method appears to be technically sound. \n\nHowever, I have the following concerns about the paper.\n\n-\tThe connection and the difference to previous work are not very clear. If I understand it correctly, the proposed method seems a combination of L2HMC and [Titsias & Dellaportas, 2019] with some slight modification (a flow model) since the naïve combination did not work well (as stated in Section 4). I think it would improve the clarity a lot if the authors explain more clearly how the proposed method differs from previous work. \n-\tSince the use of a flow model is the main difference compared to the naïve combination of two previous methods, the authors should explain more about this choice. Currently, it is not clear why this helps and there is no explanation on why the naïve combination fails. \n-\tThe proposed method seems to use more neural networks (e.g. an additional network R to handle gradient) than previous neural network MCMC. I wonder how the method performs if considering the cost. For example, the authors may instead show ESS per second on the experiments in Section 5.1. Though the authors mentioned the difficulty of computing ESS per second in the paper, I’m not entirely convinced. As the experiments in Section 5.1 are all very small-scale, is it really necessary to use GPUs? \n\n-\tThe baselines vary from experiments to experiments for no reason. For example, the authors compare their method to L2HMC on ill-conditioned Gaussian and strongly correlated Gaussian, to Neutra on Funnel distribution, and to MALA on EBM. I think this experiment design needs explanation. Also, there is no empirical comparison to [Titsias & Dellaportas, 2019] which is closely related to the proposed method.\n\nSome minor comments:\n\n-\tA.1 intends to show the benefit of using gradient information. But Variant 2 also uses gradient. What is the point of showing it? \n-\tIt is not clear to me how to interpret the empirical results in Section 5.2. For example, how does Figure 3 show that the proposed method needs less sampling steps? \n-\tThe color of points in figure 1 is hard to read.\n", "rating": "3: Clear rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603950074589}, {"id": "E1RuBmmtJHc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2093/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors describe an approach for adaptive MCMC which uses a proposal distribution parameterized by a neural network and which optimizes the entropy of the resulting proposal. Overall I found it to be an interesting algorithm, with fairly good results as compared to alternative methods, however the description of the algorithm itself was somewhat confusing and/or convoluted.\n\nThe introduction (including section 2) was great, and provided a concise but very readable introduction to recent approaches for adaptive MCMC. The beginning of section 3 was also quite well presented. However, when the authors turn to the parts of their approach inspired by HMC it gets a bit murkier. Part of the difficulty here is that at this point (the first paragraph of p4) the authors are describing HMC rather than their method, which is not entirely clear. Perhaps this could have been simplified by moving the \"related work\" section earlier in the paper and allowing for the description of HMC before diving into their own approach.\n\nThe relation between HMC and their approach which makes use of intermediate steps x+R could also have been more thoroughly explained and given more intuition. The link is there, but it doesn't seem to be as close to the leapfrog step as x+R doesn't correspond to an intermediate step (which would be x+z^n). Not that I'm saying in any way that this makes the algorithm incorrect, just that the connection isn't quite as clear cut.\n\nSimilarly I would also have liked to see a more clear description of the Q, T, and S matrices.\n\nThroughout this work the authors also refer to \"during training\". I assume they perform their adaptation steps during the sample process as is the case of the Titsias and Dellaportas work, however this could use some clarification. Overall, I think that this and other confusions cited above could have been done away with by including a clearer outline/overview of the algorithm as a whole.\n\nFinally, overall the results seem to be quite a bit better than competing methods (although I'm not an expert in this area). However, although the authors hasted to add that they do not compare \"exact computation time or ESS/second\" it would have been nice to see a more thorough discussion of the relative computational complexity of the alternatives. While ESS/grad does in some sense get close to this, the computation necessary for networks (depending on their size) could play a factor here. Similarly, I would like to see more discussion with regards to the choice of architecture for these networks---ie were they a simple MLP? It's possible I missed this, but do not think it was discussed.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper, but somewhat convoluted description of the algorithm", "review": "The authors describe an approach for adaptive MCMC which uses a proposal distribution parameterized by a neural network and which optimizes the entropy of the resulting proposal. Overall I found it to be an interesting algorithm, with fairly good results as compared to alternative methods, however the description of the algorithm itself was somewhat confusing and/or convoluted.\n\nThe introduction (including section 2) was great, and provided a concise but very readable introduction to recent approaches for adaptive MCMC. The beginning of section 3 was also quite well presented. However, when the authors turn to the parts of their approach inspired by HMC it gets a bit murkier. Part of the difficulty here is that at this point (the first paragraph of p4) the authors are describing HMC rather than their method, which is not entirely clear. Perhaps this could have been simplified by moving the \"related work\" section earlier in the paper and allowing for the description of HMC before diving into their own approach.\n\nThe relation between HMC and their approach which makes use of intermediate steps x+R could also have been more thoroughly explained and given more intuition. The link is there, but it doesn't seem to be as close to the leapfrog step as x+R doesn't correspond to an intermediate step (which would be x+z^n). Not that I'm saying in any way that this makes the algorithm incorrect, just that the connection isn't quite as clear cut.\n\nSimilarly I would also have liked to see a more clear description of the Q, T, and S matrices.\n\nThroughout this work the authors also refer to \"during training\". I assume they perform their adaptation steps during the sample process as is the case of the Titsias and Dellaportas work, however this could use some clarification. Overall, I think that this and other confusions cited above could have been done away with by including a clearer outline/overview of the algorithm as a whole.\n\nFinally, overall the results seem to be quite a bit better than competing methods (although I'm not an expert in this area). However, although the authors hasted to add that they do not compare \"exact computation time or ESS/second\" it would have been nice to see a more thorough discussion of the relative computational complexity of the alternatives. While ESS/grad does in some sense get close to this, the computation necessary for networks (depending on their size) could play a factor here. Similarly, I would like to see more discussion with regards to the choice of architecture for these networks---ie were they a simple MLP? It's possible I missed this, but do not think it was discussed.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603917033178}, {"id": "WIgnue3JOQb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2093/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe author proposes a novel MCMC sampler parametrized by the neural networks. In particular, the neural network is chosen to be a flow-based model that allows the exact evaluation of the proposal probability. \n\nThe update equations of MCMC mimic the dynamics of HMC by incorporating the gradient of the energy function into the flow. To ensure the invertibility, each update equation only depends on the other part of the variables. In addition, the author also designs another network $R$ to propose an evaluation location for the target gradient. This ensures the similarity between the proposed method and HMC. \n\nAs for the training objective, the author proposes to maximize the proposal entropy and acceptance rate, controlled by coefficient $\\beta$. Due to the tractability of the proposal density, this objective can be analytically computed. Stop gradient trick is used to stabilize the training and reduce the cost of back-propagation. \n\nEmpirically, the author evaluates the proposed sampler in some toy datasets, logistic regression, and deep energy-based models. The proposed sampler achieves higher ESS compared to baselines.\n\n---------\nReview:\nClarity: The paper is clearly written and easy to follow. The author also addresses the related work and mention the difference compared to previous baselines. \n\nTechnical soundness:\nI have a quick look at the details of the proposed method. It seems the derivation is correct. \n\nNovelty: \nAlthough the structure of the sampler is inspired by L2HMC, there are still some differences. L2HMC generalizes the HMC update equation by partitioning $\\pmb{x}$ into two parts, each part is parameterized by neural networks. The proposed sampler instead partitions the state $\\pmb{z}$ instead of parameter $\\pmb{x}$ with additional network $R$. The idea of using entropy as an objective to encourage exploration is not new.  However, the novelty lies in the usage of the flow model to allows tractability of proposal density. Overall, the proposed method is novel to some extent.  \n\nSignificance of the work:\nThe proposed method can be regarded as a variant of L2HMC with slightly different NN parameterization and training objectives. The advantage of the proposed method is the higher ESS compared to previous samplers. This may be helpful to some audiences.\n\nWeakness:\n1. Although the author demonstrates the better ESS can be obtained using the proposed method, I still prefer more analysis to understand the properties of this sampler. For example, I am curious to know how important the network $R$ is? If $R$ is removed, and the leapfrog integrator is used (i.e. update $z^{n'}$ followed by updating $x^{n}$ and finally $z^{n}$ to mimic HMC update rule). Then the gradient evaluation is also evaluated at different locations. What are the differences in terms of performances?\n2. The reason that the proposed method can have higher ESS is due to the maximization of the proposal entropy. Although the training objective has an acceptance rate term, I am curious to know: does the entropy term hurts the sample quality? For example, in logistic regression, the author only reports the ESS. What about the convergence speed of the sampler compared to others that use sample quality as the training objective? What about their performances in logistic regression?\n3. For training deep EBM, apart from ESS. I also want to know the convergence speed of the EBM training and the quality of the generated images compared to MALA. These are the standard evaluation metric for EBM.\n4. For EBM, why only compared to MALA. Any reasons why excludes other baselines?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting MCMC samplers but need some improvements in empirical analysis.", "review": "Summary:\nThe author proposes a novel MCMC sampler parametrized by the neural networks. In particular, the neural network is chosen to be a flow-based model that allows the exact evaluation of the proposal probability. \n\nThe update equations of MCMC mimic the dynamics of HMC by incorporating the gradient of the energy function into the flow. To ensure the invertibility, each update equation only depends on the other part of the variables. In addition, the author also designs another network $R$ to propose an evaluation location for the target gradient. This ensures the similarity between the proposed method and HMC. \n\nAs for the training objective, the author proposes to maximize the proposal entropy and acceptance rate, controlled by coefficient $\\beta$. Due to the tractability of the proposal density, this objective can be analytically computed. Stop gradient trick is used to stabilize the training and reduce the cost of back-propagation. \n\nEmpirically, the author evaluates the proposed sampler in some toy datasets, logistic regression, and deep energy-based models. The proposed sampler achieves higher ESS compared to baselines.\n\n---------\nReview:\nClarity: The paper is clearly written and easy to follow. The author also addresses the related work and mention the difference compared to previous baselines. \n\nTechnical soundness:\nI have a quick look at the details of the proposed method. It seems the derivation is correct. \n\nNovelty: \nAlthough the structure of the sampler is inspired by L2HMC, there are still some differences. L2HMC generalizes the HMC update equation by partitioning $\\pmb{x}$ into two parts, each part is parameterized by neural networks. The proposed sampler instead partitions the state $\\pmb{z}$ instead of parameter $\\pmb{x}$ with additional network $R$. The idea of using entropy as an objective to encourage exploration is not new.  However, the novelty lies in the usage of the flow model to allows tractability of proposal density. Overall, the proposed method is novel to some extent.  \n\nSignificance of the work:\nThe proposed method can be regarded as a variant of L2HMC with slightly different NN parameterization and training objectives. The advantage of the proposed method is the higher ESS compared to previous samplers. This may be helpful to some audiences.\n\nWeakness:\n1. Although the author demonstrates the better ESS can be obtained using the proposed method, I still prefer more analysis to understand the properties of this sampler. For example, I am curious to know how important the network $R$ is? If $R$ is removed, and the leapfrog integrator is used (i.e. update $z^{n'}$ followed by updating $x^{n}$ and finally $z^{n}$ to mimic HMC update rule). Then the gradient evaluation is also evaluated at different locations. What are the differences in terms of performances?\n2. The reason that the proposed method can have higher ESS is due to the maximization of the proposal entropy. Although the training objective has an acceptance rate term, I am curious to know: does the entropy term hurts the sample quality? For example, in logistic regression, the author only reports the ESS. What about the convergence speed of the sampler compared to others that use sample quality as the training objective? What about their performances in logistic regression?\n3. For training deep EBM, apart from ESS. I also want to know the convergence speed of the EBM training and the quality of the generated images compared to MALA. These are the standard evaluation metric for EBM.\n4. For EBM, why only compared to MALA. Any reasons why excludes other baselines?\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603887784052}, {"id": "IKM9Q_wQWKU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2093/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper argues that a better objective to train neural MCMC kernels is to maximize the proposal entropy (Titsias & Dellaportas, 2019) and demonstrate a method on doing so. The method shows improved sampling efficiency compared to previous method, especially one that optimize the alternative L2 expected jump. The novelty is not of the training objective but a neural instantiation with improved sampling efficiency.\n\n## Pros\n\n1. The method is well-motivated in Section 1 and clearly demonstrated by Figure 1.\n2. The neural instantiation consists a few clever tricks to make the network tractable and explore the space well.\n3. The method demonstrates improved sampling efficiency, measured by ESS.\n\n## Cons\n\n1. The paper may need more work on presentation.\n- Section 3 is hard to follow. It's very long and maybe adding some subsection would help.\n- I feel Section 4 should go before Section 3, given how it is currently presented.\n\n2. There is a false statements in the paper: \"one can easily make any transition probability satisfy it by including an additional Metropolis-Hastings accept-reject step (Hastings, 1970)\"\n- It's not true. One need to additionally ensure that the kernel is irreducible and aperiodic.\n- For the reason above, it's worth the mention that the proposed neural approach is irreducible and aperiodic, and give some justificaiton.\n\n3. No experiments for correctness check. I understand that showing better ESS is good, but even some biased sampler can lead to much improved ESS. I think there should be some results on comparing the proposed method against well-established HMC on a few Bayesian inference problem in terms of the posterior. Using some hypothesis testing methods to check if the posterior actually matches would give me more confidence that the method is sampling from the target; or at least, comparing a few moments.\n\n## Questions\n\n1. \"In (Titsias & Dellaportas, 2019), the authors avoid multiple back-propagation by stopping the derivative calculation at the density gradient term. In our experiment, we find it is necessary for good performance.\": does it mean you effectively using a biased gradient and it turns out to be better than an unbiased one? It looks quite weird to me as it means you are in fact optimizing some other objective which turns out to be better.\n\n2. I'm generally unsatisfied with the setup of Section 5.2.\n- \"which removes the need to choose the step size. The only tunable parameter is a target accept rate.\": I think vanilla HMC or MALA can do so by some online otpimization for step size as well (e.g. dual averaging as it's done in NUTS).\n- The unstable training of EBMs via contrastive divergence comes from the fact that the gradient estimated by a short-run MCMC is **unbiased**. The proposed method itself doesn't deal with the biasness directly, so the only plausible reason to explain why it improves the training is that the mixing is so good such that with the short chain, the bias of the gradient is so small. Is this what happening?\n- I wonder whether or not the simultaneous/interweaved training of samplers and EBMs would cause any issue. In particular, will it encourage that EBM only putting energy around data points and the sampler only effectively draw samples from some noisy distribution nearby? Are samples from EBMs too similar to data points.\n- For Figure 3.b, what's the convergence behaviour, i.e. do learnable sampler and MALA converge to a similar entropy or not?\n- For Figure 3.c, does long-run MALA produce sensible samples? If yes, it indicates that this way of training EBMs avoid some pathology which CD with short-run MCMC has; if no, then it may indicate that somehow the learnable sampler co-adapt with the EBM somehow to avoid the pathology. Is the latter we wanted?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper", "review": "The paper argues that a better objective to train neural MCMC kernels is to maximize the proposal entropy (Titsias & Dellaportas, 2019) and demonstrate a method on doing so. The method shows improved sampling efficiency compared to previous method, especially one that optimize the alternative L2 expected jump. The novelty is not of the training objective but a neural instantiation with improved sampling efficiency.\n\n## Pros\n\n1. The method is well-motivated in Section 1 and clearly demonstrated by Figure 1.\n2. The neural instantiation consists a few clever tricks to make the network tractable and explore the space well.\n3. The method demonstrates improved sampling efficiency, measured by ESS.\n\n## Cons\n\n1. The paper may need more work on presentation.\n- Section 3 is hard to follow. It's very long and maybe adding some subsection would help.\n- I feel Section 4 should go before Section 3, given how it is currently presented.\n\n2. There is a false statements in the paper: \"one can easily make any transition probability satisfy it by including an additional Metropolis-Hastings accept-reject step (Hastings, 1970)\"\n- It's not true. One need to additionally ensure that the kernel is irreducible and aperiodic.\n- For the reason above, it's worth the mention that the proposed neural approach is irreducible and aperiodic, and give some justificaiton.\n\n3. No experiments for correctness check. I understand that showing better ESS is good, but even some biased sampler can lead to much improved ESS. I think there should be some results on comparing the proposed method against well-established HMC on a few Bayesian inference problem in terms of the posterior. Using some hypothesis testing methods to check if the posterior actually matches would give me more confidence that the method is sampling from the target; or at least, comparing a few moments.\n\n## Questions\n\n1. \"In (Titsias & Dellaportas, 2019), the authors avoid multiple back-propagation by stopping the derivative calculation at the density gradient term. In our experiment, we find it is necessary for good performance.\": does it mean you effectively using a biased gradient and it turns out to be better than an unbiased one? It looks quite weird to me as it means you are in fact optimizing some other objective which turns out to be better.\n\n2. I'm generally unsatisfied with the setup of Section 5.2.\n- \"which removes the need to choose the step size. The only tunable parameter is a target accept rate.\": I think vanilla HMC or MALA can do so by some online otpimization for step size as well (e.g. dual averaging as it's done in NUTS).\n- The unstable training of EBMs via contrastive divergence comes from the fact that the gradient estimated by a short-run MCMC is **unbiased**. The proposed method itself doesn't deal with the biasness directly, so the only plausible reason to explain why it improves the training is that the mixing is so good such that with the short chain, the bias of the gradient is so small. Is this what happening?\n- I wonder whether or not the simultaneous/interweaved training of samplers and EBMs would cause any issue. In particular, will it encourage that EBM only putting energy around data points and the sampler only effectively draw samples from some noisy distribution nearby? Are samples from EBMs too similar to data points.\n- For Figure 3.b, what's the convergence behaviour, i.e. do learnable sampler and MALA converge to a similar entropy or not?\n- For Figure 3.c, does long-run MALA produce sensible samples? If yes, it indicates that this way of training EBMs avoid some pathology which CD with short-run MCMC has; if no, then it may indicate that somehow the learnable sampler co-adapt with the EBM somehow to avoid the pathology. Is the latter we wanted?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603823424974}], "openreview_url": "https://openreview.net/forum?id=edku48LG0pT", "arxiv_id": "2010.03587", "paper_pdf": "papers/edku48LG0pT.pdf", "paper_pdf_sha256": "c17fb80293c6a7fa758d1a6c6c6f9a393d3cfc3c51356feed80224f13c682c0b", "paper_pdf_bytes": 1161195, "paper_pdf_source": "openreview", "code_url": "https://github.com/NEM-MC/Entropy-NeuralMC", "code_repository": "NEM-MC/Entropy-NeuralMC", "code_commit": "067dee6795d8403df1e5214f8e859c7c90c21b65", "code_archive": "repos/edku48LG0pT.zip", "code_archive_sha256": "7a18c1816a9daa537c6fabc0662ff068ec81199761c0c9720fd871925979eea6", "code_archive_bytes": 62520, "code_file_count": 4, "code_extensions": {".py": 3, ".ipynb": 1}, "github_disk_usage_kb": 136, "github_languages": {"Jupyter Notebook": 77677, "Python": 34163}, "github_archived": false, "github_pushed_at": "2020-10-01T00:58:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-neural-network-mcmc-sampler-that-maximizes-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1x1IkHtPr", "year": 2020, "status": "rejected", "title": "A Generative Model for Molecular Distance Geometry", "authors": ["Gregor N. C. Simm", "José Miguel Hernández-Lobato"], "authorids": ["gncsimm@gmail.com", "jmh233@cam.ac.uk"], "authors_source": "OpenReview API", "abstract": "Computing equilibrium states for many-body systems, such as molecules, is a long-standing challenge. In the absence of methods for generating statistically independent samples, great computational effort is invested in simulating these systems using, for example, Markov chain Monte Carlo. We present a probabilistic model that generates such samples for molecules from their graph representations. Our model learns a low-dimensional manifold that preserves the geometry of local atomic neighborhoods through a principled learning representation that is based on Euclidean distance geometry. We create a new dataset for molecular conformation generation with which we show experimentally that our generative model achieves state-of-the-art accuracy. Finally, we show how to use our model as a proposal distribution in an importance sampling scheme to compute molecular properties.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyxOGi_D5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1714/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a generative model for generating molecule with three dimensional structure. Given a graph, it leverages a variational autoencoder to embed the distance between two atoms into latent vectors (encoder) and then generate distance between two atoms based on the latent vector. Even if the problem is well motivated, I have several concerns:\nI. What is the variable \"x\"? Is it possible to define a bit formally? It appears that x=(G,d) i.e. the graph with d.\nII. It would be nice to develop the generative model of (G,d) i.e. both G and d. Simply generating d looks a bit restrictive. \nIII. How do you model/learn \\mathcal{O}(x)?\n\nI would be curious to know the answer of these questions.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes a generative model for generating molecule with three dimensional structure. Given a graph, it leverages a variational autoencoder to embed the distance between two atoms into latent vectors (encoder) and then generate distance between two atoms based on the latent vector. Even if the problem is well motivated, I have several concerns:\nI. What is the variable \"x\"? Is it possible to define a bit formally? It appears that x=(G,d) i.e. the graph with d.\nII. It would be nice to develop the generative model of (G,d) i.e. both G and d. Simply generating d looks a bit restrictive. \nIII. How do you model/learn \\mathcal{O}(x)?\n\nI would be curious to know the answer of these questions."}, "tcdate": 1572469520330}, {"id": "HyxVxQy0FH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1714/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe authors propose a generative model designed for molecules, which is essentially a conditional variational auto-encoder. The model learns to generate, when conditioning on a molecule graph, the distribution of distances between each of the atoms and its second and third neighbour. Finally, using these information new molecules can be generated, which satisfy the generated distances. In the experiments the authors show the effectiveness of the method.\n\nGeneral Comments:\nThe paper is ok written, but as a non-expert of the field, I find it quite hard to follow. More specifically, I think that the main content aims to readers which are very familiar with the problem of molecule modeling, which is of course understandable. However, I find the technical (machine learning) content a bit unclear, because I think that the proposed method is not presented properly in details (see comments):\n\n1. I think that authors should specify the dimensionalities of each variable and function in the paper. Unfortunately, none of them is clearly defined, which makes the technical part a bit hard to follow.\n\n2. As a non-expert of the filed, I would like to know what exactly is a conformation x? Does it refer to the Euclidean coordinates of each node (atom) v of the graph? (Related to comment #1.)\n\n3. For each G, there are multiple possible conformations x_i, where each x_i specifies the Euclidean coordinates of all the nodes (atoms) v_j? Also, each node v_j has some attributes? So the goal is to represent each conformation x_i -implicitly- as a set of distances between the nodes v_j? I think that Sec 2.1. can be more clear about the definition of the data, such that to be accessible from non experts in molecule modeling. (Related to comment #1.)\n\n4. I think that the definition of the model in Sec 2.2. could have been better. As before, I believe that it is hard to understand explicitly what the functions input and output is, since the dimensionalities are not defined. In my opinion, it would have been better if instead of this high level text you could provide explicitly all the steps of the model e.g. the concatenations, the appends, etc.\n\n5. I could not understand if and how the model is able to handle graphs with different number of nodes, and different number of distances among them.\n\n6. The experimental section seems solid enough, and shows that the proposed approach works better than the other methods. Also, I like the fact that some of the limitations of the method are stated. Moreover, the related work seems to be properly included. However, since I am not an expert of the field, I am not able to provide precise feedback for these two parts.\n\nIn general, I think that the proposed model solves good enough the problem that is designed for. Also, in my opinion, the manuscript aims for specialized readers, which is definitely understandable. However, my main concern regarding the current version, is that the technical (machine learning) content is a bit unclear and probably too high level. In case the technical content was presented properly, I think that it would have been a much better fit for this community.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "title": "Official Blind Review #2", "review": "Summary:\nThe authors propose a generative model designed for molecules, which is essentially a conditional variational auto-encoder. The model learns to generate, when conditioning on a molecule graph, the distribution of distances between each of the atoms and its second and third neighbour. Finally, using these information new molecules can be generated, which satisfy the generated distances. In the experiments the authors show the effectiveness of the method.\n\nGeneral Comments:\nThe paper is ok written, but as a non-expert of the field, I find it quite hard to follow. More specifically, I think that the main content aims to readers which are very familiar with the problem of molecule modeling, which is of course understandable. However, I find the technical (machine learning) content a bit unclear, because I think that the proposed method is not presented properly in details (see comments):\n\n1. I think that authors should specify the dimensionalities of each variable and function in the paper. Unfortunately, none of them is clearly defined, which makes the technical part a bit hard to follow.\n\n2. As a non-expert of the filed, I would like to know what exactly is a conformation x? Does it refer to the Euclidean coordinates of each node (atom) v of the graph? (Related to comment #1.)\n\n3. For each G, there are multiple possible conformations x_i, where each x_i specifies the Euclidean coordinates of all the nodes (atoms) v_j? Also, each node v_j has some attributes? So the goal is to represent each conformation x_i -implicitly- as a set of distances between the nodes v_j? I think that Sec 2.1. can be more clear about the definition of the data, such that to be accessible from non experts in molecule modeling. (Related to comment #1.)\n\n4. I think that the definition of the model in Sec 2.2. could have been better. As before, I believe that it is hard to understand explicitly what the functions input and output is, since the dimensionalities are not defined. In my opinion, it would have been better if instead of this high level text you could provide explicitly all the steps of the model e.g. the concatenations, the appends, etc.\n\n5. I could not understand if and how the model is able to handle graphs with different number of nodes, and different number of distances among them.\n\n6. The experimental section seems solid enough, and shows that the proposed approach works better than the other methods. Also, I like the fact that some of the limitations of the method are stated. Moreover, the related work seems to be properly included. However, since I am not an expert of the field, I am not able to provide precise feedback for these two parts.\n\nIn general, I think that the proposed model solves good enough the problem that is designed for. Also, in my opinion, the manuscript aims for specialized readers, which is definitely understandable. However, my main concern regarding the current version, is that the technical (machine learning) content is a bit unclear and probably too high level. In case the technical content was presented properly, I think that it would have been a much better fit for this community.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1571840748327}, {"id": "SkgzPOBuYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1714/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nIn this paper, a new method for estimating the equilibrium measure of molecular compounds is proposed and tested on a new (enriched) version of an existing data set (CONF17 is derived from the pre-existing ISO17, itself extracted from QM9). \nThe method takes as input a collection of graph structures of a given stoichiometry mixture, and learns the corresponding conformations.\nThis results in a generative model, taking as input a graph structure (or actually a single conformation instance, thus setting the stoichiometry AND the isomer/graph structure), to generate (in a one-shot fashion) other conformations of that isomer.\nThe distribution of conformations sampled in this way compare well with those of an extensive training set (CONF17, set of isomers and conformations of a single compound), and in particular they perform better than 2 alternative methods published recently (2015 and 2019, the former performing globally much better than the latter!).\n\n\nConcretely, the model/method/algorithm consists in a combination of existing network architectures.  The main original contribution is to map the Cartesian coordinates of atoms into a pairwise distance space, which is naturally translationally and rotationally invariant.  Some of the second and third neighbors distances in the graph structure are considered, such that the conformation is completely set (this would not be the case if considering only 1st nearest neighbors distances).\nThe precise combination of architectures is explained in section 2.2. This explanation is a little fast for me (but I am new to this field). \nIn that section, it could be nice to precise that the job of the CVAE is to learn (among others) the parameters mu,sigma of q_phi(z|d,G), using a batch of conformations of a given isomer (and then generalize to other isomers, each time guessing the correct mu,sigma parameters, even for a new isomer).\nThis is not obvious for the inexperienced reader (I hope I \"guessed\" correctly).\n\nThe \"new dataset\" shared in this work is indeed useful for reproducing results and for later comparison with other works which aim at sampling conformations.\nHowever, it derives directly from ISO17, which is itself an extraction from QM9. ISO17 lists the Cartesian coordinates of 127 isomers of C7,H10,O2 (each having ~3000 conformations).  The \"new\" CONF17 is novel only in that it provides these conformations in distance-matrix space, i.e. it sets which pairs of atoms need to have their distances computed.  This is very useful because this choice of pairs is not unique, and so setting it once is necessary for comparison.  I would not stress too much that this is a new dataset: at first it actually sounds like authors chose a dataset such that their method outperforms others'!  Which turns out not to be the case (they chose C7,H10,O2, but this doesn't look like fine-tuning).  Indeed, their sharing of this augmented version of ISO17 is absolutely necessary: I would simply put it differently, stating from the beginning that they share an augmented version of a pre-existing data set.\n\nAbout the experimental protocol/methodology:\nthe work carried out seems quite impressive. Indeed, the paper presents a rather extensive comparison of the new model with 2 other ones, on a large data set.\nHowever sometimes the presentation rushes a bit towards results, in the sense that it is not easy to figure out over which ensemble the mean/median/max/min has been computed. This is related to the intrinsic complexity of estimating accuracy, given the encapsulation (compound-> isomer-> conformations-> each bond-> distribution over bond properties).  I think however it could be made clearer, and this would benefit the paper greatly.\nA deeper comment: in the field of conformation sampling, there are two issues. One is to correctly sample states in each local minimum of the free energy. This is addressed in the paper. The second issue is to hunt for all possible local minima of the isomer, i.e. sample even unlikely but crucial conformations. This is not addressed here.  For computing average energy, dipole moment etc, the most likely conformations will dominate, most of the time.  However given the topic I think it would be fair to mention this point, and explain why in this specific case, the second issue is not crucial.\nBecause of this, also, I think \n\nOverall, the paper is well written and is very well situated within the literature (including the very recent literature).\nIt states clearly its contributions and results, and provides detailed presentation of experimental results.\nI think it contributes to the field and provides an interesting new way of dealing with the problem of molecular conformations sampling.  The results obtained are much better than those of the compared ML method (RDkit).\nHowever, some claims (in introduction or in section 3) are stated more strongly than they are supported by direct experimental evidence.\n- ''... the first work on sampling molecular conformations for molecules of arbitrary size and shape based on deep learning.''\nHowever, the experiments have been performed only on isomers of  C7,H10,O2.\nFurthermore, there is no discussion on the computational cost of the method at all, so its extendibility to larger shapes is not obvious.\n- '' We create a new, challenging benchmark dataset CONF 17 for conformation generation, which is made publicly available.''\nI made remarks on this claim above.'' \n- '' We develop a rigorous experimental approach for evaluating and comparing the accuracy of conformation generation methods based on the mean maximum deviation distance metric.''\nHowever the paper's methodology doesn't address the issue of ''hunting for new local minima''. \nThe data set could be enriched also with a labeling of ''which conformation is in which local minimum of free energy''\n- ''It is a fast alternative to resource-intensive approaches based on MCMC or MD.''\nAs there was no mention of speed of computation in the paper (although we can guess it is faster), this is unsupported. However it is easy to remedy, simply give an order of magnitude of sampling time (for thousands of conformations of 1 isomer).\n- ''Our principled representation based on pair-wise distances does not restrict our approach to any type of system (e.g., proteins) or any particular graph structure. In addition, it is extendable in a systematic fashion.''\nThis suggests that the method would also work for dense packings. However this is not shown in the experiments.\nProbably I misunderstood the claim, and it only refers to the ability to tackle general molecular compounds, not just linear ones?  However proteins are not always linear, but can present cycles, and other branch-style patterns, so the claim is not very clear to me.\n\nBecause some of the important claims are only weakly supported, and because of the relatively small novelty of the method, I lean on rejecting the paper.\n\n\n\nAdditional comments (for improving the paper):\n\nThere is no discussion of hyper-parameter tuning, or any detail of the intermediate GNN that constitute the encoder and decoder in the CVAE. \nSo it was either not done at all, and then it should be mentioned; or it was done and it should be discussed quickly.\nIn particular, in section 2.2 there is no discussion on the impact of the latent space' dimension, here set to Nv, i.e. 7+2+10=19 (correct?).  I guess there is a balance between accuracy and generalizability, a short discussion on this point would be interesting.\n\nThe paper should clarify the hierarchy of molecular configurations:\nfor a given composition (%C, %O, %H, etc) (here, fixed throughout the paper at C7,H10,O2)\nthere are several structures possible (graphs or isomers),\nfor a given graph (isomer), there can be a couple of chiral variants (mirroring of (pieces of) the molecule,\nfor each chiral variant, there are conformations, which correspond to rather weak variations of the Cartesian coordinates, due to thermal fluctuations. Some of these variations can be important, e.g. when a bond can freely rotate. Some correspond to bending or torsion, etc, of the molecule.\nThe algorithm presented in the paper attempts to sample appropriately the equilibrium distribution of conformations for each given graph structure (at fixed chirality, if I understood correctly).\n\nAdd a comment on why (in your opinion) the distributions are not exactly correctly sampled, which after all is the goal (since the initial graph structure is given as input here).\n\n\nAdditional question, for curiosity. In section 2.3: only a single molecule output is taken from each distance matrix.\nCould you comment on the possibility to sample several Cartesian coordinates, for each distance matrix ?  It seems like several solutions should (often) exist. \nSampling continuously from that distribution would surely provide a more continuous (rich) output (am I correct) ?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "\nIn this paper, a new method for estimating the equilibrium measure of molecular compounds is proposed and tested on a new (enriched) version of an existing data set (CONF17 is derived from the pre-existing ISO17, itself extracted from QM9). \nThe method takes as input a collection of graph structures of a given stoichiometry mixture, and learns the corresponding conformations.\nThis results in a generative model, taking as input a graph structure (or actually a single conformation instance, thus setting the stoichiometry AND the isomer/graph structure), to generate (in a one-shot fashion) other conformations of that isomer.\nThe distribution of conformations sampled in this way compare well with those of an extensive training set (CONF17, set of isomers and conformations of a single compound), and in particular they perform better than 2 alternative methods published recently (2015 and 2019, the former performing globally much better than the latter!).\n\n\nConcretely, the model/method/algorithm consists in a combination of existing network architectures.  The main original contribution is to map the Cartesian coordinates of atoms into a pairwise distance space, which is naturally translationally and rotationally invariant.  Some of the second and third neighbors distances in the graph structure are considered, such that the conformation is completely set (this would not be the case if considering only 1st nearest neighbors distances).\nThe precise combination of architectures is explained in section 2.2. This explanation is a little fast for me (but I am new to this field). \nIn that section, it could be nice to precise that the job of the CVAE is to learn (among others) the parameters mu,sigma of q_phi(z|d,G), using a batch of conformations of a given isomer (and then generalize to other isomers, each time guessing the correct mu,sigma parameters, even for a new isomer).\nThis is not obvious for the inexperienced reader (I hope I \"guessed\" correctly).\n\nThe \"new dataset\" shared in this work is indeed useful for reproducing results and for later comparison with other works which aim at sampling conformations.\nHowever, it derives directly from ISO17, which is itself an extraction from QM9. ISO17 lists the Cartesian coordinates of 127 isomers of C7,H10,O2 (each having ~3000 conformations).  The \"new\" CONF17 is novel only in that it provides these conformations in distance-matrix space, i.e. it sets which pairs of atoms need to have their distances computed.  This is very useful because this choice of pairs is not unique, and so setting it once is necessary for comparison.  I would not stress too much that this is a new dataset: at first it actually sounds like authors chose a dataset such that their method outperforms others'!  Which turns out not to be the case (they chose C7,H10,O2, but this doesn't look like fine-tuning).  Indeed, their sharing of this augmented version of ISO17 is absolutely necessary: I would simply put it differently, stating from the beginning that they share an augmented version of a pre-existing data set.\n\nAbout the experimental protocol/methodology:\nthe work carried out seems quite impressive. Indeed, the paper presents a rather extensive comparison of the new model with 2 other ones, on a large data set.\nHowever sometimes the presentation rushes a bit towards results, in the sense that it is not easy to figure out over which ensemble the mean/median/max/min has been computed. This is related to the intrinsic complexity of estimating accuracy, given the encapsulation (compound-> isomer-> conformations-> each bond-> distribution over bond properties).  I think however it could be made clearer, and this would benefit the paper greatly.\nA deeper comment: in the field of conformation sampling, there are two issues. One is to correctly sample states in each local minimum of the free energy. This is addressed in the paper. The second issue is to hunt for all possible local minima of the isomer, i.e. sample even unlikely but crucial conformations. This is not addressed here.  For computing average energy, dipole moment etc, the most likely conformations will dominate, most of the time.  However given the topic I think it would be fair to mention this point, and explain why in this specific case, the second issue is not crucial.\nBecause of this, also, I think \n\nOverall, the paper is well written and is very well situated within the literature (including the very recent literature).\nIt states clearly its contributions and results, and provides detailed presentation of experimental results.\nI think it contributes to the field and provides an interesting new way of dealing with the problem of molecular conformations sampling.  The results obtained are much better than those of the compared ML method (RDkit).\nHowever, some claims (in introduction or in section 3) are stated more strongly than they are supported by direct experimental evidence.\n- ''... the first work on sampling molecular conformations for molecules of arbitrary size and shape based on deep learning.''\nHowever, the experiments have been performed only on isomers of  C7,H10,O2.\nFurthermore, there is no discussion on the computational cost of the method at all, so its extendibility to larger shapes is not obvious.\n- '' We create a new, challenging benchmark dataset CONF 17 for conformation generation, which is made publicly available.''\nI made remarks on this claim above.'' \n- '' We develop a rigorous experimental approach for evaluating and comparing the accuracy of conformation generation methods based on the mean maximum deviation distance metric.''\nHowever the paper's methodology doesn't address the issue of ''hunting for new local minima''. \nThe data set could be enriched also with a labeling of ''which conformation is in which local minimum of free energy''\n- ''It is a fast alternative to resource-intensive approaches based on MCMC or MD.''\nAs there was no mention of speed of computation in the paper (although we can guess it is faster), this is unsupported. However it is easy to remedy, simply give an order of magnitude of sampling time (for thousands of conformations of 1 isomer).\n- ''Our principled representation based on pair-wise distances does not restrict our approach to any type of system (e.g., proteins) or any particular graph structure. In addition, it is extendable in a systematic fashion.''\nThis suggests that the method would also work for dense packings. However this is not shown in the experiments.\nProbably I misunderstood the claim, and it only refers to the ability to tackle general molecular compounds, not just linear ones?  However proteins are not always linear, but can present cycles, and other branch-style patterns, so the claim is not very clear to me.\n\nBecause some of the important claims are only weakly supported, and because of the relatively small novelty of the method, I lean on rejecting the paper.\n\n\n\nAdditional comments (for improving the paper):\n\nThere is no discussion of hyper-parameter tuning, or any detail of the intermediate GNN that constitute the encoder and decoder in the CVAE. \nSo it was either not done at all, and then it should be mentioned; or it was done and it should be discussed quickly.\nIn particular, in section 2.2 there is no discussion on the impact of the latent space' dimension, here set to Nv, i.e. 7+2+10=19 (correct?).  I guess there is a balance between accuracy and generalizability, a short discussion on this point would be interesting.\n\nThe paper should clarify the hierarchy of molecular configurations:\nfor a given composition (%C, %O, %H, etc) (here, fixed throughout the paper at C7,H10,O2)\nthere are several structures possible (graphs or isomers),\nfor a given graph (isomer), there can be a couple of chiral variants (mirroring of (pieces of) the molecule,\nfor each chiral variant, there are conformations, which correspond to rather weak variations of the Cartesian coordinates, due to thermal fluctuations. Some of these variations can be important, e.g. when a bond can freely rotate. Some correspond to bending or torsion, etc, of the molecule.\nThe algorithm presented in the paper attempts to sample appropriately the equilibrium distribution of conformations for each given graph structure (at fixed chirality, if I understood correctly).\n\nAdd a comment on why (in your opinion) the distributions are not exactly correctly sampled, which after all is the goal (since the initial graph structure is given as input here).\n\n\nAdditional question, for curiosity. In section 2.3: only a single molecule output is taken from each distance matrix.\nCould you comment on the possibility to sample several Cartesian coordinates, for each distance matrix ?  It seems like several solutions should (often) exist. \nSampling continuously from that distribution would surely provide a more continuous (rich) output (am I correct) ?\n"}, "tcdate": 1571473497606}], "openreview_url": "https://openreview.net/forum?id=S1x1IkHtPr", "arxiv_id": "1909.11459", "paper_pdf": "papers/S1x1IkHtPr.pdf", "paper_pdf_sha256": "69b96be331dcd0003e3063de83e05c5c16edeb6c710ab5cc23d6e43c8e96f01d", "paper_pdf_bytes": 5856860, "paper_pdf_source": "openreview", "code_url": "https://github.com/gncs/graphdg", "code_repository": "gncs/graphdg", "code_commit": "a138448de24fcc1ead3284e380ace5e88a2c6643", "code_archive": "repos/S1x1IkHtPr.zip", "code_archive_sha256": "f7b02985374294fa06c3269ebf1790feb2d8a311c64efbfdf29dab302d8a3428", "code_archive_bytes": 164480, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 169, "github_languages": {"Python": 101020}, "github_archived": false, "github_pushed_at": "2023-03-24T22:10:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-generative-model-for-molecular-distance-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkxAUjRqY7", "year": 2019, "status": "rejected", "title": "An Information-Theoretic Metric of Transferability for Task Transfer Learning", "authors": ["Yajie Bao", "Yang Li", "Shao-Lun Huang", "Lin Zhang", "Amir R. Zamir", "Leonidas J. Guibas"], "authorids": ["byjem123@163.com", "tori2011@gmail.com", "shaolun.huang@sz.tsinghua.edu.cn", "linzhang@tsinghua.edu.cn", "zamir@cs.stanford.edu", "guibas@cs.stanford.edu"], "authors_source": "OpenReview API", "abstract": "An important question in task transfer learning is to determine task transferability, i.e. given a common input domain, estimating to what extent representations learned from a source task can help in learning a target task. Typically, transferability is either measured experimentally or inferred through task relatedness, which is often defined without a clear operational meaning. In this paper, we present a novel metric, H-score, an easily-computable evaluation function that estimates the performance of transferred representations from one task to another in classification problems. Inspired by a principled information theoretic approach, H-score has a direct connection to the asymptotic error probability of the decision function based on the transferred feature. This formulation of transferability can further be used to select a suitable set of source tasks in task transfer learning problems or to devise efficient transfer learning policies. Experiments using both synthetic and real image data show that not only our formulation of transferability is meaningful in practice, but also it can generalize to inference problems beyond classification, such as recognition tasks for 3D indoor-scene understanding.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rJl5qONCnX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper224/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose an information theoretic metric to determine a priori if a representation learned from a source task can be useful when learning another task. This is a timely topic and being able to compute such as metric between tasks upfront would be of great interest.\n\nWhile the motivation and contributions of the paper are clearly stated, a number of questions remain. Central concepts like error exponent or maximal HGR correlation are mentioned in passing. The former is important to understand whether (1) makes sense and the latter is a fundamental quantity for this work. The surprising fact is that HGR is not discussed at all, nor is there any intuition provided of what this quantity represents.\n\nWhile the proposed H-score is attractive when a source task has to be selected, as the denominator of the task transferability itself does not need to be computed, it is not straightforward to compute this quantity in general. The authors suggest to use the alternative conditional expectation (ACE) algorithm in order to optimize (2). ACE is not discussed in any detail (e.g., how accurate is this procedure; what are the choices and/or trade-offs if any?) and (2) is not justified. Overall, I felt section 4.2 was relatively inaccessible, but important as it indicates whether the proposed metric is of any use in practice.\n\nFinally, I did not understand the argument that says the H-score is to be prepared over the mutual information. To me the mutual information is still the golden standard. It was also not clear why the proposed approach would not straightforwardly apply to non classification problems as suggested in the future work.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Essential concepts and quantities undefined", "review": "The authors propose an information theoretic metric to determine a priori if a representation learned from a source task can be useful when learning another task. This is a timely topic and being able to compute such as metric between tasks upfront would be of great interest.\n\nWhile the motivation and contributions of the paper are clearly stated, a number of questions remain. Central concepts like error exponent or maximal HGR correlation are mentioned in passing. The former is important to understand whether (1) makes sense and the latter is a fundamental quantity for this work. The surprising fact is that HGR is not discussed at all, nor is there any intuition provided of what this quantity represents.\n\nWhile the proposed H-score is attractive when a source task has to be selected, as the denominator of the task transferability itself does not need to be computed, it is not straightforward to compute this quantity in general. The authors suggest to use the alternative conditional expectation (ACE) algorithm in order to optimize (2). ACE is not discussed in any detail (e.g., how accurate is this procedure; what are the choices and/or trade-offs if any?) and (2) is not justified. Overall, I felt section 4.2 was relatively inaccessible, but important as it indicates whether the proposed metric is of any use in practice.\n\nFinally, I did not understand the argument that says the H-score is to be prepared over the mutual information. To me the mutual information is still the golden standard. It was also not clear why the proposed approach would not straightforwardly apply to non classification problems as suggested in the future work.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541453970081}, {"id": "BklF6M1c2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper224/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes the H score H(f), a quantity that measure the goodness of feature f(x) for predicting some target y. This heavily builds on Makur et al. (2015) who introduce information vectors, the error exponent, and the DTM matrix. The paper connects H(f) with these quantities to justify the proposal (e.g., H(f) is proportional to the error exponent (Theorem 1)). The actual transferability is measured by the ratio between H(f) and H(f_opt) where the latter can be computed using the approach of Makur et al. \n\nThe question of how to determine the relevance of a source task for a target task without learning is an important one with lots of previous work. The proposed method seems to be novel and brings many interesting ideas in Makur et al. with empirical validation. One comment is that the paper imports heavily from Makur et al. but does not make the imported definitions and results as clear as they can be. I am still unsure of what exactly the error exponent is: its definition should probably be defined in the main paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "new measure of transferability ", "review": "The paper proposes the H score H(f), a quantity that measure the goodness of feature f(x) for predicting some target y. This heavily builds on Makur et al. (2015) who introduce information vectors, the error exponent, and the DTM matrix. The paper connects H(f) with these quantities to justify the proposal (e.g., H(f) is proportional to the error exponent (Theorem 1)). The actual transferability is measured by the ratio between H(f) and H(f_opt) where the latter can be computed using the approach of Makur et al. \n\nThe question of how to determine the relevance of a source task for a target task without learning is an important one with lots of previous work. The proposed method seems to be novel and brings many interesting ideas in Makur et al. with empirical validation. One comment is that the paper imports heavily from Makur et al. but does not make the imported definitions and results as clear as they can be. I am still unsure of what exactly the error exponent is: its definition should probably be defined in the main paper. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541169857455}, {"id": "BygTt4TK2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper224/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nsummary:\nIn this paper, the authors considered a source domain selection problem in transfer learning. Given a feature representation function f, the H-score is defined as the normalized correlation between the output f(X) and the label Y. The transferability is then measured by the ratio of H-score on the target domain and the optimal one. The authors introduced the information-theoretic and statistical meaning of the H-score. Validation of H-score was confirmed by numerical experimenters with image data. \n\n\ncomments:\nApplication of H-score to transfer learning interesting. Numerical experiments using relatively large dataset were convincing to show the validity of the proposed method. The following is some minor comments. \n\n* Some notations and terms in section 2.1 were a bit hard to understand. e.g. what does \"the error exponent corresponding to f(x)\" mean? In particular, \"corresponding to f(x)\" was not clear for me. \n\n* The authors showed some relationship between H-score and error of the statistical test. It would be nice to show a more direct relation between H-score and test accuracy in transfer learning. Typically, the risk in the target domain (T) is bounded above by the risk in the source domain (S) plus some dispersion between T and S as shown in the following paper: \nShen, et al., Wasserstein Distance Guided Representation Learning for Domain Adaptation, AAAI (2018). \n\n* In the higher order transfer of numerical experiments, the concatenated features are employed. Showing a theoretical justification of such a concatenation would be nice. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "In this paper, the authors considered source domain selection problem in transfer learning.", "review": "\nsummary:\nIn this paper, the authors considered a source domain selection problem in transfer learning. Given a feature representation function f, the H-score is defined as the normalized correlation between the output f(X) and the label Y. The transferability is then measured by the ratio of H-score on the target domain and the optimal one. The authors introduced the information-theoretic and statistical meaning of the H-score. Validation of H-score was confirmed by numerical experimenters with image data. \n\n\ncomments:\nApplication of H-score to transfer learning interesting. Numerical experiments using relatively large dataset were convincing to show the validity of the proposed method. The following is some minor comments. \n\n* Some notations and terms in section 2.1 were a bit hard to understand. e.g. what does \"the error exponent corresponding to f(x)\" mean? In particular, \"corresponding to f(x)\" was not clear for me. \n\n* The authors showed some relationship between H-score and error of the statistical test. It would be nice to show a more direct relation between H-score and test accuracy in transfer learning. Typically, the risk in the target domain (T) is bounded above by the risk in the source domain (S) plus some dispersion between T and S as shown in the following paper: \nShen, et al., Wasserstein Distance Guided Representation Learning for Domain Adaptation, AAAI (2018). \n\n* In the higher order transfer of numerical experiments, the concatenated features are employed. Showing a theoretical justification of such a concatenation would be nice. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541162116824}], "openreview_url": "https://openreview.net/forum?id=BkxAUjRqY7", "arxiv_id": null, "paper_pdf": "papers/BkxAUjRqY7.pdf", "paper_pdf_sha256": "1b1717abdb36ac9eb325635441a0321092c26d241d2b3962d330c95959188195", "paper_pdf_bytes": 3466947, "paper_pdf_source": "openreview", "code_url": "https://github.com/YaojieBao/An-Information-theoretic-Metric-of-Transferability", "code_repository": "YaojieBao/An-Information-theoretic-Metric-of-Transferability", "code_commit": "9a3f2d3f2511b6825a9d18c5fd7becbb06cae0b1", "code_archive": "repos/BkxAUjRqY7.zip", "code_archive_sha256": "834820c1d2f042baf175e0d3140190a8dd160583ca325d69a57922c1590f8648", "code_archive_bytes": 1793811, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 1752, "github_languages": {"Python": 47878}, "github_archived": false, "github_pushed_at": "2019-03-01T06:38:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/an-information-theoretic-metric-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MYqAKKsjF9", "year": 2026, "status": "rejected", "title": "LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners", "authors": ["Junhao Zheng", "Xidi Cai", "Qiuke Li", "Duzhen Zhang", "Zhong-Zhi Li", "Yingying Zhang", "Le Song", "Qianli Ma"], "authorids": ["~Junhao_Zheng3", "~Xidi_Cai1", "~Qiuke_Li1", "~Duzhen_Zhang1", "~Zhong-Zhi_Li1", "~Yingying_Zhang2", "~Le_Song1", "~Qianli_Ma3"], "authors_source": "OpenReview API", "abstract": "Lifelong learning is essential for intelligent agents operating in dynamic environments. Current large language model (LLM)-based agents, however, remain stateless and unable to accumulate or transfer knowledge over time. Existing benchmarks treat agents as static systems and fail to evaluate lifelong learning capabilities. We present LifelongAgentBench, the first unified benchmark designed to systematically assess the lifelong learning ability of LLM agents. It provides skill-grounded, interdependent tasks across three interactive environments—Database, Operating System, and Knowledge Graph—with automatic label verification, reproducibility, and modular extensibility. Extensive experiments reveal that conventional experience replay has limited effectiveness for LLM agents due to irrelevant information and context length constraints. We further introduce a group self-consistency mechanism that significantly improves lifelong learning performance. We hope LifelongAgentBench will advance the development of adaptive, memory-capable LLM agents.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "49lwiGbZjH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24174/Reviewer_eERx"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 5, "summary": "This work introduces LifelongAgentBench, the first unified benchmark for evaluating LLM agents’ lifelong learning across databases, operating systems, and knowledge graphs. It features task dependency, verifiable labels, reproducibility, and modularity. Experiments show traditional experience replay is limited, while a grouped self-consistency mechanism boosts performance. Experience quality matters more than quantity, and model architecture and task difficulty strongly affect replay effectiveness.", "review_text": "This work introduces LifelongAgentBench, the first unified benchmark for evaluating LLM agents’ lifelong learning across databases, operating systems, and knowledge graphs. It features task dependency, verifiable labels, reproducibility, and modularity. Experiments show traditional experience replay is limited, while a grouped self-consistency mechanism boosts performance. Experience quality matters more than quantity, and model architecture and task difficulty strongly affect replay effectiveness.", "strengths": "1. The benchmark is highly reliable and flexible, easy to use, and readily extensible.\n2. The grouped self-consistency mechanism effectively mitigates memory and inference overhead in large-scale experience replay.", "weaknesses": "1. The paper shows limited novelty. Among the four claimed innovations, Task Dependency is a common method for constructing tasks and does not clearly differ from prior work. Label Verifiability and Reproducibility are basic requirements for a benchmark, while Modularity relates to usability. Only Task Dependency contains some technical content, and the others cannot be considered true innovations.\n2. The evaluation of lifelong learning is incomplete because it only considers rapid adaptation to new tasks and does not assess whether new experiences cause forgetting on previously learned tasks. LifelongAgentBench does not measure the impact of new experiences on old tasks.\n3. The benchmark includes too few agent environments, which limits the generality of the conclusions.\n4. Many tasks in the database and operating system environments are generated by DeepSeek-R1, making them synthetic and potentially misaligned with real-world human task distributions.\n5. The paper defines lifelong learning narrowly, essentially by adding past experiences to the context, which resembles few-shot learning. Observed results, such as small gains for strong base models or performance improvement with more experiences, are common across tasks and not specific to agent settings.\n6. The writing and focus of the paper are problematic because it emphasizes lifelong learning while devoting most of the content to engineering details rather than conceptual or methodological contributions.", "questions": "The conclusion mentions that adding experience to a high-performing base model yields little improvement, and can even be detrimental. What, then, are the challenges faced by such strong base models in lifelong learning, and how can they be addressed?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces LifelongAgentBench, the first unified benchmark for evaluating LLM agents’ lifelong learning across databases, operating systems, and knowledge graphs. It features task dependency, verifiable labels, reproducibility, and modularity. Experiments show traditional experience replay is limited, while a grouped self-consistency mechanism boosts performance. Experience quality matters more than quantity, and model architecture and task difficulty strongly affect replay effectiveness.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "1. The benchmark is highly reliable and flexible, easy to use, and readily extensible.\n2. The grouped self-consistency mechanism effectively mitigates memory and inference overhead in large-scale experience replay.", "weaknesses": "1. The paper shows limited novelty. Among the four claimed innovations, Task Dependency is a common method for constructing tasks and does not clearly differ from prior work. Label Verifiability and Reproducibility are basic requirements for a benchmark, while Modularity relates to usability. Only Task Dependency contains some technical content, and the others cannot be considered true innovations.\n2. The evaluation of lifelong learning is incomplete because it only considers rapid adaptation to new tasks and does not assess whether new experiences cause forgetting on previously learned tasks. LifelongAgentBench does not measure the impact of new experiences on old tasks.\n3. The benchmark includes too few agent environments, which limits the generality of the conclusions.\n4. Many tasks in the database and operating system environments are generated by DeepSeek-R1, making them synthetic and potentially misaligned with real-world human task distributions.\n5. The paper defines lifelong learning narrowly, essentially by adding past experiences to the context, which resembles few-shot learning. Observed results, such as small gains for strong base models or performance improvement with more experiences, are common across tasks and not specific to agent settings.\n6. The writing and focus of the paper are problematic because it emphasizes lifelong learning while devoting most of the content to engineering details rather than conceptual or methodological contributions.", "questions": "The conclusion mentions that adding experience to a high-performing base model yields little improvement, and can even be detrimental. What, then, are the challenges faced by such strong base models in lifelong learning, and how can they be addressed?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761763387731}, {"id": "PuTIAeUjm2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24174/Reviewer_PjtQ"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 3, "summary": "This paper introduces LifelongAgentBench, a benchmark intended to evaluate the lifelong learning capabilities of  LLM agents. The authors posit that existing benchmarks fail to assess knowledge accumulation over time. The benchmark provides a sequence of skill-grounded tasks in three environments (Database, OS, Knowledge Graph). The paper's primary evaluation focuses on in-context experience replay (ER), finding that replaying relevant experiences is superior to replaying recent ones. It also proposes \"group self-consistency\", a voting method, to manage the context-length limitations of this replay strategy.", "review_text": "This paper introduces LifelongAgentBench, a benchmark intended to evaluate the lifelong learning capabilities of  LLM agents. The authors posit that existing benchmarks fail to assess knowledge accumulation over time. The benchmark provides a sequence of skill-grounded tasks in three environments (Database, OS, Knowledge Graph). The paper's primary evaluation focuses on in-context experience replay (ER), finding that replaying relevant experiences is superior to replaying recent ones. It also proposes \"group self-consistency\", a voting method, to manage the context-length limitations of this replay strategy.", "strengths": "Important Problem: The paper's core motivation is strong. Evaluating the ability of agents to learn continuously is a critical, timely, and under-studied problem in the field of LLM-based agents.\n\nBenchmark Artifact: The creation of a dedicated, open-source benchmark with containerized environments and automatic verification is a non-trivial engineering effort. This infrastructure could, in principle, be a useful tool for the community.", "weaknesses": "Unclear Definition of \"Lifelong Learning\": The paper fails to provide a precise and operational definition of lifelong learning. In Section 3, the problem formulation is presented as a generic sequential POMDP, which does not capture any distinctive characteristics of lifelong tasks. No explicit statement is given to clarify what “lifelong” means in this context, and how it affects the benchmark design.\n\nOverstated Novelty and Weak Analysis: The claimed methodological contribution appears minor and is overstated. The concept of group self-consistency (Section 6.5) seems to be a straightforward rebranding of standard self-consistency methods (e.g., Wang et al., 2023), without a clear theoretical or methodological differentiation. The “systematic analysis” is limited, focusing only on the comparison between experience replay and group self-consistency, which does not provide a comprehensive understanding of the proposed approach in relation to the broader body of methods.\n\nPoor Clarity and Presentation: The paper is difficult to follow due to unclear exposition of core concepts. Terminology such as \"skill concurrency\", \"skill-grounded\", \"label verification\" (is this equivalent to \"label validation\"?), \"parallel execution\", etc., is introduced without proper definitions or contextual examples. Figures and tables suffer from poor readability (extremely small font sizes; Figure 1 is overly cluttered). The paper does not convincingly justify what makes its task dependencies uniquely \"lifelong\", as similar setups could be replicated using existing benchmarks with replay-based agents. Table 1 lists differences, but it is unclear why these differences provide specific advantages under a lifelong learning scenario. Further explanation is necessary. Also, Table 1 is inconsistent with the “four key innovations” described later in the text, very confusing. Several citations are incorrect (e.g., VisualWebArena, AgentBench, wrong authors, wrong links), are they AI generated?", "questions": "I highly doubt that some major parts are written by llms without careful checking. I strongly recommend that the authors carefully review these paragraphs and thoroughly refine the wording to improve clarity and precision. Due to the poor readability of many parts, I may have overlooked some of the paper’s potential contributions. A significant improvement in writing quality would positively influence my evaluation and may result in a higher score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LifelongAgentBench, a benchmark intended to evaluate the lifelong learning capabilities of  LLM agents. The authors posit that existing benchmarks fail to assess knowledge accumulation over time. The benchmark provides a sequence of skill-grounded tasks in three environments (Database, OS, Knowledge Graph). The paper's primary evaluation focuses on in-context experience replay (ER), finding that replaying relevant experiences is superior to replaying recent ones. It also proposes \"group self-consistency\", a voting method, to manage the context-length limitations of this replay strategy.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "Important Problem: The paper's core motivation is strong. Evaluating the ability of agents to learn continuously is a critical, timely, and under-studied problem in the field of LLM-based agents.\n\nBenchmark Artifact: The creation of a dedicated, open-source benchmark with containerized environments and automatic verification is a non-trivial engineering effort. This infrastructure could, in principle, be a useful tool for the community.", "weaknesses": "Unclear Definition of \"Lifelong Learning\": The paper fails to provide a precise and operational definition of lifelong learning. In Section 3, the problem formulation is presented as a generic sequential POMDP, which does not capture any distinctive characteristics of lifelong tasks. No explicit statement is given to clarify what “lifelong” means in this context, and how it affects the benchmark design.\n\nOverstated Novelty and Weak Analysis: The claimed methodological contribution appears minor and is overstated. The concept of group self-consistency (Section 6.5) seems to be a straightforward rebranding of standard self-consistency methods (e.g., Wang et al., 2023), without a clear theoretical or methodological differentiation. The “systematic analysis” is limited, focusing only on the comparison between experience replay and group self-consistency, which does not provide a comprehensive understanding of the proposed approach in relation to the broader body of methods.\n\nPoor Clarity and Presentation: The paper is difficult to follow due to unclear exposition of core concepts. Terminology such as \"skill concurrency\", \"skill-grounded\", \"label verification\" (is this equivalent to \"label validation\"?), \"parallel execution\", etc., is introduced without proper definitions or contextual examples. Figures and tables suffer from poor readability (extremely small font sizes; Figure 1 is overly cluttered). The paper does not convincingly justify what makes its task dependencies uniquely \"lifelong\", as similar setups could be replicated using existing benchmarks with replay-based agents. Table 1 lists differences, but it is unclear why these differences provide specific advantages under a lifelong learning scenario. Further explanation is necessary. Also, Table 1 is inconsistent with the “four key innovations” described later in the text, very confusing. Several citations are incorrect (e.g., VisualWebArena, AgentBench, wrong authors, wrong links), are they AI generated?", "questions": "I highly doubt that some major parts are written by llms without careful checking. I strongly recommend that the authors carefully review these paragraphs and thoroughly refine the wording to improve clarity and precision. Due to the poor readability of many parts, I may have overlooked some of the paper’s potential contributions. A significant improvement in writing quality would positively influence my evaluation and may result in a higher score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761638545723}, {"id": "Wqod02lcIQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24174/Reviewer_mS98"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes LifelongAgentBench, the first unified benchmark framework specifically designed to evaluate the capabilities of LLM-based agents in lifelong learning scenarios. Unlike previous evaluation methods that treat agents as static systems, this framework emphasizes assessing an agent’s ability to accumulate, retain, and transfer knowledge within continuous, interdependent task sequences.\n\nMain contributions include:\n- Systematic analysis of the effects of experience replay: Identifies limitations of traditional approaches in LLM-based agents, such as interference from irrelevant information and context length constraints.\n- Proposing the group self-consistency mechanism: Improves decision quality by grouping historical experiences and using voting, significantly alleviating memory and reasoning overhead issues.", "review_text": "This paper proposes LifelongAgentBench, the first unified benchmark framework specifically designed to evaluate the capabilities of LLM-based agents in lifelong learning scenarios. Unlike previous evaluation methods that treat agents as static systems, this framework emphasizes assessing an agent’s ability to accumulate, retain, and transfer knowledge within continuous, interdependent task sequences.\n\nMain contributions include:\n- Systematic analysis of the effects of experience replay: Identifies limitations of traditional approaches in LLM-based agents, such as interference from irrelevant information and context length constraints.\n- Proposing the group self-consistency mechanism: Improves decision quality by grouping historical experiences and using voting, significantly alleviating memory and reasoning overhead issues.", "strengths": "- This work is the first to propose a benchmark specifically targeting the lifelong learning capability of LLM-based agents, with a novel problem definition that fills a gap in existing evaluation frameworks.\n- The proposed grouped self-consistency mechanism represents an improvement over traditional experience replay methods, demonstrating methodological innovation.\n- The work offers an extensive suite of well-defined and verifiable agent tasks, enabling performance evaluation and experiments.", "weaknesses": "There are flaws in the experimental aspect:\n1. On line 054, table 1 only includes a few agent-related benchmarks for comparison. Examples like osworld and browsecomp were not taken into consideration.\n2. On line 328, table 2 intends to express the effectiveness of replay, but it only uses one model.\n3. Line 435, Table 3 only measured DB and KG. Additionally, the number of models used for DB and KG was different. If a model fails in KG, then DB should not be included either, as it has no significance.\n4. On line 270, fig 3 is the only experiment that used a closed-source model. Why wasn't it presented in a table?\n\nOverall, as a benchmark, it fails to provide sufficient evaluation results using both open-source and closed-source models. The types and quantities of models used in each experiment are very arbitrary. There was no appropriate ablation study for the proposed replay and vote methods.", "questions": "See the \"Weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes LifelongAgentBench, the first unified benchmark framework specifically designed to evaluate the capabilities of LLM-based agents in lifelong learning scenarios. Unlike previous evaluation methods that treat agents as static systems, this framework emphasizes assessing an agent’s ability to accumulate, retain, and transfer knowledge within continuous, interdependent task sequences.\n\nMain contributions include:\n- Systematic analysis of the effects of experience replay: Identifies limitations of traditional approaches in LLM-based agents, such as interference from irrelevant information and context length constraints.\n- Proposing the group self-consistency mechanism: Improves decision quality by grouping historical experiences and using voting, significantly alleviating memory and reasoning overhead issues.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- This work is the first to propose a benchmark specifically targeting the lifelong learning capability of LLM-based agents, with a novel problem definition that fills a gap in existing evaluation frameworks.\n- The proposed grouped self-consistency mechanism represents an improvement over traditional experience replay methods, demonstrating methodological innovation.\n- The work offers an extensive suite of well-defined and verifiable agent tasks, enabling performance evaluation and experiments.", "weaknesses": "There are flaws in the experimental aspect:\n1. On line 054, table 1 only includes a few agent-related benchmarks for comparison. Examples like osworld and browsecomp were not taken into consideration.\n2. On line 328, table 2 intends to express the effectiveness of replay, but it only uses one model.\n3. Line 435, Table 3 only measured DB and KG. Additionally, the number of models used for DB and KG was different. If a model fails in KG, then DB should not be included either, as it has no significance.\n4. On line 270, fig 3 is the only experiment that used a closed-source model. Why wasn't it presented in a table?\n\nOverall, as a benchmark, it fails to provide sufficient evaluation results using both open-source and closed-source models. The types and quantities of models used in each experiment are very arbitrary. There was no appropriate ablation study for the proposed replay and vote methods.", "questions": "See the \"Weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761290302700}], "openreview_url": "https://openreview.net/forum?id=MYqAKKsjF9", "arxiv_id": "2505.11942", "paper_pdf": "papers/MYqAKKsjF9.pdf", "paper_pdf_sha256": "d829274949b3bf7f51b3ae449c7a186f2323f4626418df4a24ea06d59a19a8b4", "paper_pdf_bytes": 6526450, "paper_pdf_source": "openreview", "code_url": "https://github.com/caixd-220529/LifelongAgentBench", "code_repository": "caixd-220529/LifelongAgentBench", "code_commit": "d6f19b42eb358d9150379f0c68c2985c5a867520", "code_archive": "repos/MYqAKKsjF9.zip", "code_archive_sha256": "29a0a57ec64dbe6913cc80bc5ba601031165205f0a93e492733399f0876fd1bd", "code_archive_bytes": 563670, "code_file_count": 99, "code_extensions": {".py": 95, ".sh": 4}, "github_disk_usage_kb": 739, "github_languages": {"Python": 685603, "Shell": 9205}, "github_archived": false, "github_pushed_at": "2025-05-30T02:43:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lifelongagentbench-evaluating-llm-agents-as"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2L1OxhQCwS", "year": 2025, "status": "rejected", "title": "Transformers versus LSTMs for electronic trading", "authors": ["Paul Alexander Bilokon", "Yitao Qiu"], "authorids": ["~Paul_Alexander_Bilokon1", "~Yitao_Qiu1"], "authors_source": "OpenReview API", "abstract": "The rapid advancement of artificial intelligence has seen widespread application of long short-term memory (LSTM), a type of recurrent neural network (RNN), in time series forecasting. Despite the success of Transformers in natural language processing (NLP), which prompted interest in their efficacy for time series prediction, their application in financial time series forecasting is less explored compared to the dominant LSTM models. This study investigates whether Transformer-based models can outperform LSTMs in financial time series forecasting. It involves a comparative analysis of various LSTM-based and Transformer-based models on multiple financial prediction tasks using high-frequency limit order book data. A novel LSTM-based model named DLSTM is introduced alongside a newly designed Transformer-based model tailored for financial predictions. The findings indicate that Transformer-based models exhibit only a marginal advantage in predicting absolute price sequences, whereas LSTM-based models demonstrate superior and more consistent performance in predicting differential sequences such as price differences and movements.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "flDuAZNkXt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_riPs"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This research compares the effectiveness of Transformer and LSTM architectures in financial forecasting. The study examines both model types using high-frequency trading data and introduces DLSTM and a finance-specific Transformer. Results show that Transformers only slightly outperform in absolute price predictions, while LSTMs showing more reliable performance overall.", "review_text": "This research compares the effectiveness of Transformer and LSTM architectures in financial forecasting. The study examines both model types using high-frequency trading data and introduces DLSTM and a finance-specific Transformer. Results show that Transformers only slightly outperform in absolute price predictions, while LSTMs showing more reliable performance overall.", "strengths": "1. The paper addresses a relevant and significant question by comparing LSTM and Transformer models in financial time series forecasting.\n2. The experimental setup is extensive and provides substantial data.", "weaknesses": "1. The paper lacks code and detailed implementation information for both the Transformer and LSTM models, which limits reproducibility.\n2. The novelty of the proposed approach is limited. While the authors introduce a DLSTM model to improve performance, the idea of decomposition was previously explored in models like DLinear [1], diminishing the originality of the contribution. Beyond the comparative analysis, additional innovation is also limited.\n3. The decomposition strategy appears to be applied only to the LSTM model. For a fair comparison, a decomposition approach for the Transformer model should also be included. In Table 3, DLSTM significantly outperforms LSTM, which suggests that a decomposed Transformer might also show improved results.\n4. The paper does not include several state-of-the-art (SOTA) Transformer-based models, such as PatchTST [2], Crossformer [3], and iTransformer [4], in the comparison, which limits the comprehensiveness of the analysis.\n5. The statement \"Transformer-based models exhibit only a marginal advantage in predicting absolute price sequences, whereas LSTM-based models demonstrate superior and more consistent performance in predicting differential sequences such as price differences and movements\" requires further investigation. A deeper analysis into the underlying causes of this observed difference is missing, which weakens the interpretability of the results.\n\n[1] Zeng, Ailing, et al. \"Are transformers effective for time series forecasting?.\" Proceedings of the AAAI conference on artificial intelligence. Vol. 37. No. 9. 2023.\n\n[2] Nie, Yuqi, et al. \"A Time Series is Worth 64 Words: Long-term Forecasting with Transformers.\" The Eleventh International Conference on Learning Representations.\n\n[3] Zhang, Yunhao, and Junchi Yan. \"Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting.\" The eleventh international conference on learning representations. 2023.\n\n[4] Liu, Yong, et al. \"iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.\" The Twelfth International Conference on Learning Representations.", "questions": "1. What are the fundamental architectural characteristics that make LSTM models more effective for differential sequences compared to Transformers?\n2. Can you provide deeper analysis to support the generalizability of your findings of LSTM vs Transformer?\n3. How does financial time series forecasting different from other time series forecasting (like weather, traffic, etc.?)\n4. To address the remaining limitations identified in *Weaknesses*: a) Could you provide detailed model implementations and hyperparameter configurations? b) How would decomposition techniques benefit Transformer architectures? c) Please include comparisons with state-of-the-art models", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This research compares the effectiveness of Transformer and LSTM architectures in financial forecasting. The study examines both model types using high-frequency trading data and introduces DLSTM and a finance-specific Transformer. Results show that Transformers only slightly outperform in absolute price predictions, while LSTMs showing more reliable performance overall.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "1. The paper addresses a relevant and significant question by comparing LSTM and Transformer models in financial time series forecasting.\n2. The experimental setup is extensive and provides substantial data.", "weaknesses": "1. The paper lacks code and detailed implementation information for both the Transformer and LSTM models, which limits reproducibility.\n2. The novelty of the proposed approach is limited. While the authors introduce a DLSTM model to improve performance, the idea of decomposition was previously explored in models like DLinear [1], diminishing the originality of the contribution. Beyond the comparative analysis, additional innovation is also limited.\n3. The decomposition strategy appears to be applied only to the LSTM model. For a fair comparison, a decomposition approach for the Transformer model should also be included. In Table 3, DLSTM significantly outperforms LSTM, which suggests that a decomposed Transformer might also show improved results.\n4. The paper does not include several state-of-the-art (SOTA) Transformer-based models, such as PatchTST [2], Crossformer [3], and iTransformer [4], in the comparison, which limits the comprehensiveness of the analysis.\n5. The statement \"Transformer-based models exhibit only a marginal advantage in predicting absolute price sequences, whereas LSTM-based models demonstrate superior and more consistent performance in predicting differential sequences such as price differences and movements\" requires further investigation. A deeper analysis into the underlying causes of this observed difference is missing, which weakens the interpretability of the results.\n\n[1] Zeng, Ailing, et al. \"Are transformers effective for time series forecasting?.\" Proceedings of the AAAI conference on artificial intelligence. Vol. 37. No. 9. 2023.\n\n[2] Nie, Yuqi, et al. \"A Time Series is Worth 64 Words: Long-term Forecasting with Transformers.\" The Eleventh International Conference on Learning Representations.\n\n[3] Zhang, Yunhao, and Junchi Yan. \"Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting.\" The eleventh international conference on learning representations. 2023.\n\n[4] Liu, Yong, et al. \"iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.\" The Twelfth International Conference on Learning Representations.", "questions": "1. What are the fundamental architectural characteristics that make LSTM models more effective for differential sequences compared to Transformers?\n2. Can you provide deeper analysis to support the generalizability of your findings of LSTM vs Transformer?\n3. How does financial time series forecasting different from other time series forecasting (like weather, traffic, etc.?)\n4. To address the remaining limitations identified in *Weaknesses*: a) Could you provide detailed model implementations and hyperparameter configurations? b) How would decomposition techniques benefit Transformer architectures? c) Please include comparisons with state-of-the-art models", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730706699209}, {"id": "ubPvMvNBRF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_t3XX"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "5348_Transformers_versus_LSTMs\npdf\nLT\nHere is a new paper needs to be reviewed. Summary*\nBriefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\n*\n\nSummary\n\nThis paper conducts a comparative study between LSTM-based and Transformer-based models for financial time series forecasting, specifically in the context of electronic trading using high-frequency limit order book (LOB) data. The authors investigate the performance of these models across three prediction tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction.\n\nFor the mid-price prediction task, the study finds that Transformer-based models like FEDformer and Autoformer achieve lower prediction errors than LSTM-based models. However, the authors note that the practical utility of these results for high-frequency trading is limited due to insufficient prediction quality.\n\nIn the mid-price difference prediction task, LSTM-based models demonstrate superior performance and robustness compared to Transformer-based models. The canonical LSTM achieves the highest R^2 of around 11.5% within about 10 prediction steps, while state-of-the-art Transformer models struggle to effectively process difference sequences.\n\nThe paper's main contribution lies in the mid-price movement prediction task, where the authors introduce a novel LSTM-based model called DLSTM. This model integrates LSTM with a time series decomposition approach inspired by the Autoformer architecture. DLSTM significantly outperforms all other models in classification metrics and proves its effectiveness in trading simulations, particularly when transaction costs are considered.\n\nAdditionally, the authors adapt the architecture of existing Transformer-based models to better suit the demands of the movement prediction task. They incorporate both past and projected mid-price data, followed by a linear layer and softmax activation, to determine price movements.\n\nOverall, the study highlights that while Transformer-based models may excel in certain aspects of mid-price prediction, LSTM-based models, particularly the proposed DLSTM, demonstrate consistent superiority and practicality in financial time series prediction for electronic trading.", "review_text": "5348_Transformers_versus_LSTMs\npdf\nLT\nHere is a new paper needs to be reviewed. Summary*\nBriefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\n*\n\nSummary\n\nThis paper conducts a comparative study between LSTM-based and Transformer-based models for financial time series forecasting, specifically in the context of electronic trading using high-frequency limit order book (LOB) data. The authors investigate the performance of these models across three prediction tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction.\n\nFor the mid-price prediction task, the study finds that Transformer-based models like FEDformer and Autoformer achieve lower prediction errors than LSTM-based models. However, the authors note that the practical utility of these results for high-frequency trading is limited due to insufficient prediction quality.\n\nIn the mid-price difference prediction task, LSTM-based models demonstrate superior performance and robustness compared to Transformer-based models. The canonical LSTM achieves the highest R^2 of around 11.5% within about 10 prediction steps, while state-of-the-art Transformer models struggle to effectively process difference sequences.\n\nThe paper's main contribution lies in the mid-price movement prediction task, where the authors introduce a novel LSTM-based model called DLSTM. This model integrates LSTM with a time series decomposition approach inspired by the Autoformer architecture. DLSTM significantly outperforms all other models in classification metrics and proves its effectiveness in trading simulations, particularly when transaction costs are considered.\n\nAdditionally, the authors adapt the architecture of existing Transformer-based models to better suit the demands of the movement prediction task. They incorporate both past and projected mid-price data, followed by a linear layer and softmax activation, to determine price movements.\n\nOverall, the study highlights that while Transformer-based models may excel in certain aspects of mid-price prediction, LSTM-based models, particularly the proposed DLSTM, demonstrate consistent superiority and practicality in financial time series prediction for electronic trading.", "strengths": "Originality: The study offers a novel perspective on the application of LSTM-based and Transformer-based models in financial time series forecasting, specifically in the context of electronic trading using high-frequency LOB data. The authors introduce a new LSTM-based model, DLSTM, which creatively combines LSTM with a time series decomposition approach inspired by the Autoformer architecture. This innovative integration of existing ideas allows DLSTM to outperform other models in the mid-price movement prediction task.\nQuality: The paper demonstrates a high level of quality in its experimental design and analysis. The authors conduct a comprehensive comparative study across three prediction tasks (mid-price prediction, mid-price difference prediction, and mid-price movement prediction), using a diverse range of LSTM-based and Transformer-based models. The experiments are well-structured, and the results are thoroughly analyzed, providing valuable insights into the performance of different models in each task.\nClarity: The paper is well-written and easy to follow. The authors provide clear explanations of the problem formulation, the proposed DLSTM model, and the experimental setup. The use of tables and figures enhances the clarity of the results, making it easy for readers to compare the performance of different models across various metrics and prediction horizons.\nSignificance: The findings of this study have significant implications for the application of deep learning models in financial time series forecasting, particularly in the context of electronic trading. The authors demonstrate that while Transformer-based models may excel in certain aspects of mid-price prediction, LSTM-based models, especially the proposed DLSTM, exhibit superior and more consistent performance in tasks such as mid-price difference prediction and mid-price movement prediction. The incorporation of trading simulations with and without transaction costs further highlights the practical significance of the proposed DLSTM model for real-world trading scenarios.\n\nMoreover, the paper's adaptation of existing Transformer-based models' architecture to better suit the demands of the movement prediction task showcases the potential for further improvements in this domain. By incorporating both past and projected mid-price data, followed by a linear layer and softmax activation, the authors demonstrate a creative approach to enhancing the performance of Transformer-based models in financial time series forecasting.\nIn summary, the paper's originality, quality, clarity, and significance make it a valuable contribution to the field of financial time series forecasting using deep learning models, offering new insights and directions for future research in this domain.", "weaknesses": "While the paper presents valuable insights and contributions, there are a few areas that could be improved or require further clarification:\n\nLimited dataset diversity: The experiments in this study are conducted using LOB data from a single cryptocurrency pair (BTC-USDT or ETH-USDT) on one exchange (Binance). To demonstrate the generalizability of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models, it would be beneficial to include a wider range of financial instruments, such as stocks, forex, or other cryptocurrencies, as well as data from multiple exchanges. This would strengthen the paper's conclusions and provide a more comprehensive assessment of the models' performance across diverse financial time series.\nLack of ablation studies: While the paper introduces the novel DLSTM model, which integrates LSTM with a time series decomposition approach, there is a lack of ablation studies to investigate the individual contributions of each component. For example, the authors could compare the performance of DLSTM with and without the time series decomposition to assess the impact of this specific modification. Additionally, a more detailed analysis of the adapted Transformer-based models' architecture for the movement prediction task would provide valuable insights into the effectiveness of the proposed changes.\nLimited discussion on model interpretability: Interpretability is a crucial aspect of financial time series forecasting models, especially in the context of electronic trading, where understanding the factors driving the model's predictions is essential for risk management and decision-making. The paper could benefit from a more in-depth discussion on the interpretability of the proposed DLSTM model and the adapted Transformer-based models, as well as a comparison with the interpretability of other LSTM-based and Transformer-based models.\nHyperparameter tuning and model selection: The paper does not provide a detailed description of the hyperparameter tuning process and model selection criteria for the various models used in the experiments. It is essential to discuss the approach used for hyperparameter optimization, such as grid search, random search, or Bayesian optimization, and the specific hyperparameters tuned for each model. Additionally, the authors could provide more information on the model selection process, such as the use of validation sets or cross-validation techniques.\nRobustness to market conditions: The experiments in this study are conducted using LOB data from a specific time period (e.g., July 2022). To demonstrate the robustness of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models, it would be valuable to evaluate the models' performance under different market conditions, such as periods of high volatility, market crashes, or significant news events. This would provide a more comprehensive assessment of the models' ability to generalize and adapt to various market scenarios.\n\nAddressing these weaknesses would further strengthen the paper's contributions and provide a more comprehensive and robust analysis of the proposed DLSTM model and the comparative study between LSTM-based and Transformer-based models in financial time series forecasting for electronic trading.", "questions": "Dataset diversity and generalizability: Can you provide more insights into the choice of using only Binance LOB data for a single cryptocurrency pair in your experiments? How do you expect the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models to perform on a wider range of financial instruments, such as stocks, forex, or other cryptocurrencies, as well as data from multiple exchanges? Providing results on more diverse datasets could strengthen the claims of generalizability and robustness of the findings.\nAblation studies and component contributions: Can you conduct ablation studies to investigate the individual contributions of the time series decomposition approach in the proposed DLSTM model? It would be helpful to compare the performance of DLSTM with and without this specific modification to assess its impact on the model's effectiveness. Additionally, can you provide a more detailed analysis of the adapted Transformer-based models' architecture for the movement prediction task, highlighting the importance of each proposed change?\nModel interpretability: Can you elaborate on the interpretability of the proposed DLSTM model and the adapted Transformer-based models? How do these models compare with other LSTM-based and Transformer-based models in terms of interpretability? Providing insights into the factors driving the models' predictions and their relative importance could be valuable for understanding the models' decision-making process and enhancing trust in their applications for electronic trading.\nHyperparameter tuning and model selection: Can you provide more details on the hyperparameter tuning process and model selection criteria used for the various models in your experiments? Specifically, what approach was used for hyperparameter optimization (e.g., grid search, random search, Bayesian optimization), and which hyperparameters were tuned for each model? Additionally, how were the validation sets or cross-validation techniques employed in the model selection process?\nRobustness to market conditions: Have you considered evaluating the performance of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models under different market conditions, such as periods of high volatility, market crashes, or significant news events? Demonstrating the models' ability to generalize and adapt to various market scenarios could provide a more comprehensive assessment of their robustness and practical applicability in electronic trading.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "5348_Transformers_versus_LSTMs\npdf\nLT\nHere is a new paper needs to be reviewed. Summary*\nBriefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary.\n*\n\nSummary\n\nThis paper conducts a comparative study between LSTM-based and Transformer-based models for financial time series forecasting, specifically in the context of electronic trading using high-frequency limit order book (LOB) data. The authors investigate the performance of these models across three prediction tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction.\n\nFor the mid-price prediction task, the study finds that Transformer-based models like FEDformer and Autoformer achieve lower prediction errors than LSTM-based models. However, the authors note that the practical utility of these results for high-frequency trading is limited due to insufficient prediction quality.\n\nIn the mid-price difference prediction task, LSTM-based models demonstrate superior performance and robustness compared to Transformer-based models. The canonical LSTM achieves the highest R^2 of around 11.5% within about 10 prediction steps, while state-of-the-art Transformer models struggle to effectively process difference sequences.\n\nThe paper's main contribution lies in the mid-price movement prediction task, where the authors introduce a novel LSTM-based model called DLSTM. This model integrates LSTM with a time series decomposition approach inspired by the Autoformer architecture. DLSTM significantly outperforms all other models in classification metrics and proves its effectiveness in trading simulations, particularly when transaction costs are considered.\n\nAdditionally, the authors adapt the architecture of existing Transformer-based models to better suit the demands of the movement prediction task. They incorporate both past and projected mid-price data, followed by a linear layer and softmax activation, to determine price movements.\n\nOverall, the study highlights that while Transformer-based models may excel in certain aspects of mid-price prediction, LSTM-based models, particularly the proposed DLSTM, demonstrate consistent superiority and practicality in financial time series prediction for electronic trading.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Originality: The study offers a novel perspective on the application of LSTM-based and Transformer-based models in financial time series forecasting, specifically in the context of electronic trading using high-frequency LOB data. The authors introduce a new LSTM-based model, DLSTM, which creatively combines LSTM with a time series decomposition approach inspired by the Autoformer architecture. This innovative integration of existing ideas allows DLSTM to outperform other models in the mid-price movement prediction task.\nQuality: The paper demonstrates a high level of quality in its experimental design and analysis. The authors conduct a comprehensive comparative study across three prediction tasks (mid-price prediction, mid-price difference prediction, and mid-price movement prediction), using a diverse range of LSTM-based and Transformer-based models. The experiments are well-structured, and the results are thoroughly analyzed, providing valuable insights into the performance of different models in each task.\nClarity: The paper is well-written and easy to follow. The authors provide clear explanations of the problem formulation, the proposed DLSTM model, and the experimental setup. The use of tables and figures enhances the clarity of the results, making it easy for readers to compare the performance of different models across various metrics and prediction horizons.\nSignificance: The findings of this study have significant implications for the application of deep learning models in financial time series forecasting, particularly in the context of electronic trading. The authors demonstrate that while Transformer-based models may excel in certain aspects of mid-price prediction, LSTM-based models, especially the proposed DLSTM, exhibit superior and more consistent performance in tasks such as mid-price difference prediction and mid-price movement prediction. The incorporation of trading simulations with and without transaction costs further highlights the practical significance of the proposed DLSTM model for real-world trading scenarios.\n\nMoreover, the paper's adaptation of existing Transformer-based models' architecture to better suit the demands of the movement prediction task showcases the potential for further improvements in this domain. By incorporating both past and projected mid-price data, followed by a linear layer and softmax activation, the authors demonstrate a creative approach to enhancing the performance of Transformer-based models in financial time series forecasting.\nIn summary, the paper's originality, quality, clarity, and significance make it a valuable contribution to the field of financial time series forecasting using deep learning models, offering new insights and directions for future research in this domain.", "weaknesses": "While the paper presents valuable insights and contributions, there are a few areas that could be improved or require further clarification:\n\nLimited dataset diversity: The experiments in this study are conducted using LOB data from a single cryptocurrency pair (BTC-USDT or ETH-USDT) on one exchange (Binance). To demonstrate the generalizability of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models, it would be beneficial to include a wider range of financial instruments, such as stocks, forex, or other cryptocurrencies, as well as data from multiple exchanges. This would strengthen the paper's conclusions and provide a more comprehensive assessment of the models' performance across diverse financial time series.\nLack of ablation studies: While the paper introduces the novel DLSTM model, which integrates LSTM with a time series decomposition approach, there is a lack of ablation studies to investigate the individual contributions of each component. For example, the authors could compare the performance of DLSTM with and without the time series decomposition to assess the impact of this specific modification. Additionally, a more detailed analysis of the adapted Transformer-based models' architecture for the movement prediction task would provide valuable insights into the effectiveness of the proposed changes.\nLimited discussion on model interpretability: Interpretability is a crucial aspect of financial time series forecasting models, especially in the context of electronic trading, where understanding the factors driving the model's predictions is essential for risk management and decision-making. The paper could benefit from a more in-depth discussion on the interpretability of the proposed DLSTM model and the adapted Transformer-based models, as well as a comparison with the interpretability of other LSTM-based and Transformer-based models.\nHyperparameter tuning and model selection: The paper does not provide a detailed description of the hyperparameter tuning process and model selection criteria for the various models used in the experiments. It is essential to discuss the approach used for hyperparameter optimization, such as grid search, random search, or Bayesian optimization, and the specific hyperparameters tuned for each model. Additionally, the authors could provide more information on the model selection process, such as the use of validation sets or cross-validation techniques.\nRobustness to market conditions: The experiments in this study are conducted using LOB data from a specific time period (e.g., July 2022). To demonstrate the robustness of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models, it would be valuable to evaluate the models' performance under different market conditions, such as periods of high volatility, market crashes, or significant news events. This would provide a more comprehensive assessment of the models' ability to generalize and adapt to various market scenarios.\n\nAddressing these weaknesses would further strengthen the paper's contributions and provide a more comprehensive and robust analysis of the proposed DLSTM model and the comparative study between LSTM-based and Transformer-based models in financial time series forecasting for electronic trading.", "questions": "Dataset diversity and generalizability: Can you provide more insights into the choice of using only Binance LOB data for a single cryptocurrency pair in your experiments? How do you expect the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models to perform on a wider range of financial instruments, such as stocks, forex, or other cryptocurrencies, as well as data from multiple exchanges? Providing results on more diverse datasets could strengthen the claims of generalizability and robustness of the findings.\nAblation studies and component contributions: Can you conduct ablation studies to investigate the individual contributions of the time series decomposition approach in the proposed DLSTM model? It would be helpful to compare the performance of DLSTM with and without this specific modification to assess its impact on the model's effectiveness. Additionally, can you provide a more detailed analysis of the adapted Transformer-based models' architecture for the movement prediction task, highlighting the importance of each proposed change?\nModel interpretability: Can you elaborate on the interpretability of the proposed DLSTM model and the adapted Transformer-based models? How do these models compare with other LSTM-based and Transformer-based models in terms of interpretability? Providing insights into the factors driving the models' predictions and their relative importance could be valuable for understanding the models' decision-making process and enhancing trust in their applications for electronic trading.\nHyperparameter tuning and model selection: Can you provide more details on the hyperparameter tuning process and model selection criteria used for the various models in your experiments? Specifically, what approach was used for hyperparameter optimization (e.g., grid search, random search, Bayesian optimization), and which hyperparameters were tuned for each model? Additionally, how were the validation sets or cross-validation techniques employed in the model selection process?\nRobustness to market conditions: Have you considered evaluating the performance of the proposed DLSTM model and the comparative analysis between LSTM-based and Transformer-based models under different market conditions, such as periods of high volatility, market crashes, or significant news events? Demonstrating the models' ability to generalize and adapt to various market scenarios could provide a more comprehensive assessment of their robustness and practical applicability in electronic trading.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730701992965}, {"id": "HRWTuL26v1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_5BdM"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper explores the use of Transformer and LSTM-based models for financial time series forecasting tasks using high-frequency limit order book (LOB) data. A new LSTM-based model, DLSTM, is proposed alongside a modified Transformer architecture tailored for financial predictions. The study compares these models across three tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction. Results suggest that Transformer-based models offer only marginal improvements in specific tasks, while LSTM models, particularly DLSTM, are more reliable in predicting mid-price differences and movements.", "review_text": "The paper explores the use of Transformer and LSTM-based models for financial time series forecasting tasks using high-frequency limit order book (LOB) data. A new LSTM-based model, DLSTM, is proposed alongside a modified Transformer architecture tailored for financial predictions. The study compares these models across three tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction. Results suggest that Transformer-based models offer only marginal improvements in specific tasks, while LSTM models, particularly DLSTM, are more reliable in predicting mid-price differences and movements.", "strengths": "Relevant Application: The use of LSTM and Transformer models for financial predictions on LOB data is timely and relevant given the growing interest in high-frequency trading and predictive models in finance.\n\nComparative Scope: The study covers multiple models and tasks, providing a broad comparison between LSTM- and Transformer-based architectures on real-world financial data.", "weaknesses": "Unconvincing Novelty: The paper lacks substantial novelty. The DLSTM model is essentially a combination of existing methods, such as time series decomposition and LSTM layers, without a clear innovation. Similarly, the Transformer modifications are incremental and do not provide a compelling improvement. As a result, the contributions seem incremental and insufficiently distinct from existing work in financial time series forecasting.\n\nInterpretability Issues: The added complexity of Transformer-based models raises interpretability concerns, especially given the unclear benefit over simpler LSTM-based models. Without a more interpretable mechanism or explanation for its performance gains, the model’s added complexity appears unnecessary.\n\nInsufficient Performance Gain for Complexity: The study demonstrates only marginal improvements from the proposed Transformer modifications over traditional LSTMs, particularly in mid-price prediction. Despite the significant computational complexity introduced by Transformer-based models, the improvements are minimal and do not convincingly justify their adoption for practical trading applications.", "questions": "What specific modifications were made to the Transformer architecture to adapt it to financial prediction tasks?\n\nCan the authors elaborate on the metrics used to evaluate the models' performance? What criteria were significant in determining the practical utility of the models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores the use of Transformer and LSTM-based models for financial time series forecasting tasks using high-frequency limit order book (LOB) data. A new LSTM-based model, DLSTM, is proposed alongside a modified Transformer architecture tailored for financial predictions. The study compares these models across three tasks: mid-price prediction, mid-price difference prediction, and mid-price movement prediction. Results suggest that Transformer-based models offer only marginal improvements in specific tasks, while LSTM models, particularly DLSTM, are more reliable in predicting mid-price differences and movements.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "Relevant Application: The use of LSTM and Transformer models for financial predictions on LOB data is timely and relevant given the growing interest in high-frequency trading and predictive models in finance.\n\nComparative Scope: The study covers multiple models and tasks, providing a broad comparison between LSTM- and Transformer-based architectures on real-world financial data.", "weaknesses": "Unconvincing Novelty: The paper lacks substantial novelty. The DLSTM model is essentially a combination of existing methods, such as time series decomposition and LSTM layers, without a clear innovation. Similarly, the Transformer modifications are incremental and do not provide a compelling improvement. As a result, the contributions seem incremental and insufficiently distinct from existing work in financial time series forecasting.\n\nInterpretability Issues: The added complexity of Transformer-based models raises interpretability concerns, especially given the unclear benefit over simpler LSTM-based models. Without a more interpretable mechanism or explanation for its performance gains, the model’s added complexity appears unnecessary.\n\nInsufficient Performance Gain for Complexity: The study demonstrates only marginal improvements from the proposed Transformer modifications over traditional LSTMs, particularly in mid-price prediction. Despite the significant computational complexity introduced by Transformer-based models, the improvements are minimal and do not convincingly justify their adoption for practical trading applications.", "questions": "What specific modifications were made to the Transformer architecture to adapt it to financial prediction tasks?\n\nCan the authors elaborate on the metrics used to evaluate the models' performance? What criteria were significant in determining the practical utility of the models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730699762533}, {"id": "Xrf6Fykthl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_bpRu"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper compares the performance of LSTM and Transformer models in financial time series forecasting (limit order book data). They compared with FEDformer, Autoformer, Informer, Reformer, Transformer and LSTM. The main results show that Transformer has a slight advantage in predicting absolute price series, but the LSTM model performs more consistently and accurately in the prediction of price changes and price movements. In addition, the paper introduces DLSTM inspired by DLinear and Autoformer.", "review_text": "This paper compares the performance of LSTM and Transformer models in financial time series forecasting (limit order book data). They compared with FEDformer, Autoformer, Informer, Reformer, Transformer and LSTM. The main results show that Transformer has a slight advantage in predicting absolute price series, but the LSTM model performs more consistently and accurately in the prediction of price changes and price movements. In addition, the paper introduces DLSTM inspired by DLinear and Autoformer.", "strengths": "Even relatively simple LSTM models perform well in financial time series forecasting tasks, compared with transformer-based model.", "weaknesses": "1. The writing quality of the paper is low, especially the format of literature citation is not uniform and some of the citations are not standardized in formatting and arrangement. \n2. The experimental setup lacks comparison with the frameworks and standards widely used in the current research field and fails to demonstrate the advantages of the selected model. For example, the authors failed to cite and use the latest limit order book (LOB) benchmark frameworks, such as LOBFrame (https://github.com/FinancialComputingUCL/LOBFrame) and LOBCAST (https://arxiv.org/abs/2308.01915), both of which are open source frameworks currently widely used for Limit Order Book Forecasting. In addition, the authors did not include some of the latest Transformer based models (e.g., iTransformer and PatchTST), which have demonstrated advantages in terms of performance and efficiency in time series forecasting. Comparing these latest models would make the experimental results more convincing and practical.\n3. The experimental data used in this paper is limit order book data from three cryptocurrencies, which, although suitable for high-frequency forecasting tests, is not representative of the financial market, and the volatility and noise characteristics of the cryptocurrency market are quite different from those of the traditional financial market. Data from LOBSTER (https://lobsterdata.com/) are more common and widely used in the literature currently.", "questions": "If possible, include LOB data from the LOBSTER dataset, to increase the generalizability of the experiment. If possible, include latest transformer based model (e.g. iTransformer, PatchTST). Recommend to use benchmarking frameworks such as LOBFrame or LOBCAST  in the experimental design to ensure that the results can be more comparable to existing studies. A more detailed discussion of the specific differences and advantages of DLSTM over other temporal decomposition methods (e.g., DLinear) could be added. could also include some ablation studies. include code for reproducibility.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper compares the performance of LSTM and Transformer models in financial time series forecasting (limit order book data). They compared with FEDformer, Autoformer, Informer, Reformer, Transformer and LSTM. The main results show that Transformer has a slight advantage in predicting absolute price series, but the LSTM model performs more consistently and accurately in the prediction of price changes and price movements. In addition, the paper introduces DLSTM inspired by DLinear and Autoformer.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "Even relatively simple LSTM models perform well in financial time series forecasting tasks, compared with transformer-based model.", "weaknesses": "1. The writing quality of the paper is low, especially the format of literature citation is not uniform and some of the citations are not standardized in formatting and arrangement. \n2. The experimental setup lacks comparison with the frameworks and standards widely used in the current research field and fails to demonstrate the advantages of the selected model. For example, the authors failed to cite and use the latest limit order book (LOB) benchmark frameworks, such as LOBFrame (https://github.com/FinancialComputingUCL/LOBFrame) and LOBCAST (https://arxiv.org/abs/2308.01915), both of which are open source frameworks currently widely used for Limit Order Book Forecasting. In addition, the authors did not include some of the latest Transformer based models (e.g., iTransformer and PatchTST), which have demonstrated advantages in terms of performance and efficiency in time series forecasting. Comparing these latest models would make the experimental results more convincing and practical.\n3. The experimental data used in this paper is limit order book data from three cryptocurrencies, which, although suitable for high-frequency forecasting tests, is not representative of the financial market, and the volatility and noise characteristics of the cryptocurrency market are quite different from those of the traditional financial market. Data from LOBSTER (https://lobsterdata.com/) are more common and widely used in the literature currently.", "questions": "If possible, include LOB data from the LOBSTER dataset, to increase the generalizability of the experiment. If possible, include latest transformer based model (e.g. iTransformer, PatchTST). Recommend to use benchmarking frameworks such as LOBFrame or LOBCAST  in the experimental design to ensure that the results can be more comparable to existing studies. A more detailed discussion of the specific differences and advantages of DLSTM over other temporal decomposition methods (e.g., DLinear) could be added. could also include some ablation studies. include code for reproducibility.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730667380510}, {"id": "ZZn9fgzPuo", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_2dpN"], "rating": 3, "soundness": 1, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "In this paper, the authors conduct a comparative analysis of various LSTM-based and Transformer-based models for multiple financial prediction tasks using high-frequency limit order book data. They introduce a novel LSTM-based model called DLSTM and a newly designed Transformer-based model specifically tailored for financial predictions. Their results reveal that Transformer-based models offer a slight advantage in predicting absolute price sequences. However, LSTM-based models show superior and more consistent performance in predicting differential sequences, such as price differences and movements.", "review_text": "In this paper, the authors conduct a comparative analysis of various LSTM-based and Transformer-based models for multiple financial prediction tasks using high-frequency limit order book data. They introduce a novel LSTM-based model called DLSTM and a newly designed Transformer-based model specifically tailored for financial predictions. Their results reveal that Transformer-based models offer a slight advantage in predicting absolute price sequences. However, LSTM-based models show superior and more consistent performance in predicting differential sequences, such as price differences and movements.", "strengths": "The structure and logic of the paper is well organized. \n\nThe experimental setup, description, and analysis are clearly stated with sufficient detail.", "weaknesses": "1. The authors compare Transformers and LSTMs, concluding that LSTMs have advantages in multiple electronic trading tasks. However, the selection of Transformer-based models is limited to earlier studies (prior to 2023) and does not include recent state-of-the-art (SOTA) works, such as those mentioned in references [1], [2], and [3]. Notably, Liu et al. [2] claim significant improvements on similar tasks. Excluding these recent studies makes it premature to conclude that Transformer-based models underperform compared to LSTMs. Additionally, there is insufficient evidence to assert that the authors' proposed DLSTM model is the optimal choice for this application. Could you please include comparisons with some of these SOTA results to more robustly justify the conclusion?\n\n[1] Garza, A., Challu, C., & Mergenthaler-Canseco, M. (2023). TimeGPT-1. arXiv preprint arXiv:2310.03589.\n\n[2] Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., & Long, M. (2023). itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625.\n\n[3] Das, A., Kong, W., Sen, R., & Zhou, Y. (2023). A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688.\n\n2. The authors' conclusion lacks novelty and largely aligns with the findings and conclusions of Zeng et al. [4] It appears to apply established approaches and conclusions to domain-specific practices. While retaining empirical relevance, the study does not offer methodological breakthroughs.\n\n[4] Zeng, A., Chen, M., Zhang, L., & Xu, Q. (2023, June). Are transformers effective for time series forecasting?. In Proceedings of the AAAI conference on artificial intelligence (Vol. 37, No. 9, pp. 11121-11128).\n\n3. The experimental setup could be made more representative by incorporating additional metrics such as Mean Absolute Scaled Error and Relative Mean Absolute Error.", "questions": "Please refer to questions to be addressed, the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors conduct a comparative analysis of various LSTM-based and Transformer-based models for multiple financial prediction tasks using high-frequency limit order book data. They introduce a novel LSTM-based model called DLSTM and a newly designed Transformer-based model specifically tailored for financial predictions. Their results reveal that Transformer-based models offer a slight advantage in predicting absolute price sequences. However, LSTM-based models show superior and more consistent performance in predicting differential sequences, such as price differences and movements.", "soundness": 1, "presentation": 3, "contribution": 1, "strengths": "The structure and logic of the paper is well organized. \n\nThe experimental setup, description, and analysis are clearly stated with sufficient detail.", "weaknesses": "1. The authors compare Transformers and LSTMs, concluding that LSTMs have advantages in multiple electronic trading tasks. However, the selection of Transformer-based models is limited to earlier studies (prior to 2023) and does not include recent state-of-the-art (SOTA) works, such as those mentioned in references [1], [2], and [3]. Notably, Liu et al. [2] claim significant improvements on similar tasks. Excluding these recent studies makes it premature to conclude that Transformer-based models underperform compared to LSTMs. Additionally, there is insufficient evidence to assert that the authors' proposed DLSTM model is the optimal choice for this application. Could you please include comparisons with some of these SOTA results to more robustly justify the conclusion?\n\n[1] Garza, A., Challu, C., & Mergenthaler-Canseco, M. (2023). TimeGPT-1. arXiv preprint arXiv:2310.03589.\n\n[2] Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., & Long, M. (2023). itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625.\n\n[3] Das, A., Kong, W., Sen, R., & Zhou, Y. (2023). A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688.\n\n2. The authors' conclusion lacks novelty and largely aligns with the findings and conclusions of Zeng et al. [4] It appears to apply established approaches and conclusions to domain-specific practices. While retaining empirical relevance, the study does not offer methodological breakthroughs.\n\n[4] Zeng, A., Chen, M., Zhang, L., & Xu, Q. (2023, June). Are transformers effective for time series forecasting?. In Proceedings of the AAAI conference on artificial intelligence (Vol. 37, No. 9, pp. 11121-11128).\n\n3. The experimental setup could be made more representative by incorporating additional metrics such as Mean Absolute Scaled Error and Relative Mean Absolute Error.", "questions": "Please refer to questions to be addressed, the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730420715328}, {"id": "kYCjigTamA", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5348/Reviewer_Drb3"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "This research examines the performance differences between Transformer-based models and LSTMs across three cryptocurrency limit order book data prediction tasks. It also introduces DLSTM, a LSTM-based model, and a Transformer-based model redesigned for financial forecasting.", "review_text": "This research examines the performance differences between Transformer-based models and LSTMs across three cryptocurrency limit order book data prediction tasks. It also introduces DLSTM, a LSTM-based model, and a Transformer-based model redesigned for financial forecasting.", "strengths": "(1) This research presents findings that compare the performance of two types of models.\n(2) It successfully highlights the weaknesses in current measurement metrics.\n(3) Interesting task definition.", "weaknesses": "1.Baseline Selection Rationale: The paper does not clearly explain why specific Transformer and LSTM variants, such as Autoformer and FEDformer, were chosen in the comparison. It remains unclear if these variants have unique advantages for financial time series forecasting. Providing additional theoretical support or rationale for model selection would enhance the scientific basis of this choice.\n\n2.Data Risk: The study only tests on a single asset (BTC-USDT), lacking a broader dataset. This limited scope may mean the model’s performance does not generalize well to other financial data. Testing on a single asset is insufficient to comprehensively assess the model’s generalizability.\n\n3.Lack of Experimental Details: The paper lacks adequate details on the experimental setup, especially regarding hyperparameter settings and baseline model architectures. This omission makes replication challenging and affects the reliability of the results. Sufficient information is not provided to ensure a fair comparison among baseline models.\n\n4.Unclear Result Interpretation: The paper does not adequately explain the significant differences in performance between experiments with and without transaction costs. Lacking theoretical support or data analysis, it's hard for me to understand the causes behind these variations under different settings.\n\n5.Limited Community Contribution: Time series decomposition, used in this study, appears to be a common approach, closely resembling classical time series decomposition methods. It is unclear how this study provides any specific advantage over the standard decomposition methods.\n\n6.Although the paper points out shortcomings in MSE and MAE metrics, it fails to propose a robust method to address these deficiencies.\n\n7.Some capitalization inconsistencies, eg. in line 034  Self-attention mechanism.", "questions": "1.Given the limited dataset used and the lack of detailed experimental information (settings of baselines), I am very concerned about the reliability of this paper's conclusions. How would you address or demonstrate the robustness of your findings under these limitations?\n\n2.How do you explain the significant differences in experimental results with and without transaction costs? What factors contribute to this discrepancy?\n\n3.What are the specific advantages of your time series decomposition method compared to other decomposition approaches, and why do these advantages arise?\n\n4.Other questions can refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This research examines the performance differences between Transformer-based models and LSTMs across three cryptocurrency limit order book data prediction tasks. It also introduces DLSTM, a LSTM-based model, and a Transformer-based model redesigned for financial forecasting.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "(1) This research presents findings that compare the performance of two types of models.\n(2) It successfully highlights the weaknesses in current measurement metrics.\n(3) Interesting task definition.", "weaknesses": "1.Baseline Selection Rationale: The paper does not clearly explain why specific Transformer and LSTM variants, such as Autoformer and FEDformer, were chosen in the comparison. It remains unclear if these variants have unique advantages for financial time series forecasting. Providing additional theoretical support or rationale for model selection would enhance the scientific basis of this choice.\n\n2.Data Risk: The study only tests on a single asset (BTC-USDT), lacking a broader dataset. This limited scope may mean the model’s performance does not generalize well to other financial data. Testing on a single asset is insufficient to comprehensively assess the model’s generalizability.\n\n3.Lack of Experimental Details: The paper lacks adequate details on the experimental setup, especially regarding hyperparameter settings and baseline model architectures. This omission makes replication challenging and affects the reliability of the results. Sufficient information is not provided to ensure a fair comparison among baseline models.\n\n4.Unclear Result Interpretation: The paper does not adequately explain the significant differences in performance between experiments with and without transaction costs. Lacking theoretical support or data analysis, it's hard for me to understand the causes behind these variations under different settings.\n\n5.Limited Community Contribution: Time series decomposition, used in this study, appears to be a common approach, closely resembling classical time series decomposition methods. It is unclear how this study provides any specific advantage over the standard decomposition methods.\n\n6.Although the paper points out shortcomings in MSE and MAE metrics, it fails to propose a robust method to address these deficiencies.\n\n7.Some capitalization inconsistencies, eg. in line 034  Self-attention mechanism.", "questions": "1.Given the limited dataset used and the lack of detailed experimental information (settings of baselines), I am very concerned about the reliability of this paper's conclusions. How would you address or demonstrate the robustness of your findings under these limitations?\n\n2.How do you explain the significant differences in experimental results with and without transaction costs? What factors contribute to this discrepancy?\n\n3.What are the specific advantages of your time series decomposition method compared to other decomposition approaches, and why do these advantages arise?\n\n4.Other questions can refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730196782734}], "openreview_url": "https://openreview.net/forum?id=2L1OxhQCwS", "arxiv_id": "2309.11400", "paper_pdf": "papers/2L1OxhQCwS.pdf", "paper_pdf_sha256": "a7d60bfaa0fcc03d864026a9baa0bd1a09d2d9f07d2757fb456b335c114be055", "paper_pdf_bytes": 2383241, "paper_pdf_source": "openreview", "code_url": "https://github.com/772435284/transformers_versus_lstms_for_electronic_trading", "code_repository": "772435284/transformers_versus_lstms_for_electronic_trading", "code_commit": "65d691a52c4ebc5646048cb1a183e2b1748efedd", "code_archive": "repos/2L1OxhQCwS.zip", "code_archive_sha256": "a8ce17af9be0073966659adf9dce10f9649b73bc51be7c5bb0efbd80be93a09e", "code_archive_bytes": 73701, "code_file_count": 52, "code_extensions": {".py": 31, ".sh": 18, ".ipynb": 3}, "github_disk_usage_kb": 65, "github_languages": {"Python": 144742, "Jupyter Notebook": 49889, "Shell": 42520}, "github_archived": false, "github_pushed_at": "2023-12-04T13:14:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transformers-versus-lstms-for-electronic-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e1IMBXiDhW", "year": 2024, "status": "rejected", "title": "Matrix Information Theory for Self-Supervised Learning", "authors": ["Yifan Zhang", "Zhiquan Tan", "Jingqin Yang", "Weiran Huang", "Yang Yuan"], "authorids": ["~Yifan_Zhang16", "~Zhiquan_Tan1", "~Jingqin_Yang2", "~Weiran_Huang1", "~Yang_Yuan4"], "authors_source": "OpenReview API", "abstract": "Contrastive learning often relies on comparing a single positive anchor sample with multiple negative samples to perform Self-Supervised Learning (SSL). However, non-contrastive approaches like BYOL, SimSiam, and Barlow Twins achieve SSL without explicit negative samples. In this paper, we introduce a unified matrix information-theoretic framework that explains many contrastive and non-contrastive learning methods. We then propose a novel method Matrix-SSL based on matrix information theory. Experimental results reveal that Matrix-SSL significantly outperforms state-of-the-art methods on the ImageNet dataset under linear evaluation settings and on MS-COCO for transfer learning tasks. Specifically, when performing 100 epochs pre-training, our method outperforms SimCLR by 4.6\\%, and when performing transfer learning tasks on MS-COCO, our method outperforms previous SOTA methods such as MoCo v2 and BYOL up to 3.3\\% with only 400 epochs compared to 800 epochs pre-training.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "AYp3V8kFh9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission524/Reviewer_CjGu"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors introduce Matrix-SSL, an approach grounded in matrix\ninformation theory, to improve current SSL methods.  The approach is\nmotivated through theory and the experiments show improved accuracy.", "review_text": "The authors introduce Matrix-SSL, an approach grounded in matrix\ninformation theory, to improve current SSL methods.  The approach is\nmotivated through theory and the experiments show improved accuracy.", "strengths": "Casting various contrastive methods in a unifying notation and\nframework is helpful and shows the similarity.\n\nThe related work and cited literature is extensive and I could not\nmake out any significant missing literature.\n\nThe findings are clearly presented.", "weaknesses": "Table 1 only reports the accuracy of up to 400 epochs.  It would be\ninteresting to see the dynamics of all approaches after 800 epochs,\nare they closer to Matrix-SSL?  It also does not report any mean +-\nstd over multiple runs.\n\nWhile I find the experiments convincing, it could reproduce\nstate-of-the-art better with other methods.  E.g. SimCLR is usually\ntrained for 1000 epochs, but this is not done in this paper.", "questions": "Why do you think that the method works best for gamma = 1?  \n\nPerhaps the authors could comment on the computational aspect of the method?  Does it slow down training?  If yes, why?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce Matrix-SSL, an approach grounded in matrix\ninformation theory, to improve current SSL methods.  The approach is\nmotivated through theory and the experiments show improved accuracy.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "Casting various contrastive methods in a unifying notation and\nframework is helpful and shows the similarity.\n\nThe related work and cited literature is extensive and I could not\nmake out any significant missing literature.\n\nThe findings are clearly presented.", "weaknesses": "Table 1 only reports the accuracy of up to 400 epochs.  It would be\ninteresting to see the dynamics of all approaches after 800 epochs,\nare they closer to Matrix-SSL?  It also does not report any mean +-\nstd over multiple runs.\n\nWhile I find the experiments convincing, it could reproduce\nstate-of-the-art better with other methods.  E.g. SimCLR is usually\ntrained for 1000 epochs, but this is not done in this paper.", "questions": "Why do you think that the method works best for gamma = 1?  \n\nPerhaps the authors could comment on the computational aspect of the method?  Does it slow down training?  If yes, why?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698761000444}, {"id": "QsTweZK4hp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission524/Reviewer_zEzQ"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The article aims to introduce a unifying information-theoretic framework for self-supervised learning (SSL). For this purpose, the article\n -  Surveys some established SSL methodologies.,\n - Introduces matrix entropy, matrix KL divergence and matrix cross entropy measures,\n - Expresses certain existing SSL loss functions using the matrix cross entropy measure,\n - Proposes a new SSL loss function derived from matrix cross-entropy,\n - Conducts numerical experiments, demonstrating the enhanced performance of the proposed method compared to select existing approaches,\n - Draws a connection of matrix cross entropy with the effective rank.", "review_text": "The article aims to introduce a unifying information-theoretic framework for self-supervised learning (SSL). For this purpose, the article\n -  Surveys some established SSL methodologies.,\n - Introduces matrix entropy, matrix KL divergence and matrix cross entropy measures,\n - Expresses certain existing SSL loss functions using the matrix cross entropy measure,\n - Proposes a new SSL loss function derived from matrix cross-entropy,\n - Conducts numerical experiments, demonstrating the enhanced performance of the proposed method compared to select existing approaches,\n - Draws a connection of matrix cross entropy with the effective rank.", "strengths": "The article's pursuit of a unifying framework offers a commendable approach. Strategy to employ  matrix information measures to achieve this is intriguing. Moreover, the numerical examples showcase marked enhancements over certain existing methods, underscoring the efficacy of the algorithm derived from this framework.", "weaknesses": "The article lacks a clear organizational structure and consistent notation, making it challenging to follow. Concepts are introduced without adequate explanation or clarity. Additionally, the matrix information measures employed are not innovative; similar methods have been previously applied in the SSL context. The attempt to frame existing methods as special cases within this framework falls short of being convincing and satisfactory. Please see Questions section for details.", "questions": "## INTRODUCTION\n\n- The following reference,\n\n[a] Ozsoy S, Hamdan S, Arik S, Yuret D, Erdogan A. Self-supervised learning with an information maximization criterion. Advances in Neural Information Processing Systems. 2022 Dec 6;35:35240-53,\n\nproposes utilizing \"correlative information\" maximization for self-supervised learning. Analyzing this paper within the context of the proposed matrix information framework would be interesting, especially since the authors of [a] assert that maximizing correlative information between the representations of augmentations establishes a linear dependence rather than an arbitrary nonlinear one.\n\n- Figure 1: The citation for Coco is absent. It would be beneficial to compare the performances of  Vicreg, [a] (Bardes et. al, 2021), and (Tong et. al, 2023).\n\n\n### 2.1 CONTRASTIVE AND NON-CONTRASTIVE SELF-SUPERVISED LEARNING\n\n- First paragraph: The discussion here is based on the SimCLR and SimSiam, however, the authors introduce a generic SSL architecture. Furthermore, what is meant by dual networks is not clear at this point.\n\n- Would categorizing this section into subheadings like \"Contrastive SSL Approaches\" and \"Non-Contrastive SSL Approaches\" enhance clarity?\n\n- There seems to be inconsistency in the conventions used for sample and augmentation indices in Equations (2) and (3)?\n\n- For Equation (3), is there an underlying assumption about the normalization of the encoder,  such as  $\\||z_i^{(k)}\\||_2=1$ ?\n\n\n### 2.2 MATRIX INFORMATION-THEORETIC QUANTITIES\n- Could you provide a citation detailing Matrix Entropy. Furthermore, could you discuss its interpretation, and perhaps its existing applications especially within the context of machine learning/SSL?\n\n- Regarding the Matrix Entropy definition, is there a specific assumption about the trace of $\\mathbf{A}$ ensuring its eigenvalues form a probability mass function?\n\n- Can you provide interpretations and potential applications of Matrix KL Divergence and MCE?\n\n- Bach 2022's KL divergence doesn't seem to incorporate the  $-\\mathbf{P}+\\mathbf{Q}$ terms within trace? Could this discrepancy be addressed?\n\n- From the presented definitions, it appears that, unlike Shannon Entropy, MCE does not equate to the sum of Matrix KL and Matrix Entropy. Should there be a $\\text{Tr}(\\mathbf{P})$ term included in the matrix entropy definition?\n\n- A brief discussion explaining the relevance of these definitions to the SSL problem would be insightful.\n\n\n### 3 MATRIX INFORMATION-THEORETIC PERSPECTIVES OF SELF-SUPERVISED LEARNING\n\n- The assertion about proofs should be positioned adjacent to the first proposition, i.e., Proposition 3.1.\n\n- Proposition 3.3: The statement of the proposition is  ambigious:  InfoNCE cost in (1) is the MCE of which matrices? This should be clearly stated. Are InfNCE cost and SimCLR cost identical? The cost function obtained in the proof in terms of MCE does not match (1)?\n- The paragraph after Proposition 3.3: This part appears convoluted: Is $p_{data}=p_{\\mathbf{x}}$ ? and $\\mathcal{X}$ is the support set of $p_{\\mathbf{x}}$? I guess there is no clear definition of $f$ before. (there was $f_\\Theta$ and $f_\\phi$ without clear definitions before). $f$ is sometimes unbold and sometimes bold? Since $f$ is not prespecified, the assumption that there is a layer normalization $\\||f(\\mathbf{x})\\||_2^2=1$ is also not clear. Instead of stating \"straightforward calculation\", it is better to provide a proof of Lemma 3.4 with proper notation in Appendix A. In my opinion, both this paragraph and Lemma 3.4 is not properly motivated.\n\n- Lemma 3.4: Suggestion \"Let $\\sigma$ represent the uniform distribution on $S^{d-1}$. ...\".\n\n- The paragraph after Lemma 3.4: Change of variables formula? (Inverse image rule?). Why \"auto-correlation\" matrix for $q$ but \"covariance\" for $p_{data}$?  It is better for the authors to clearly state the uniformity principle. Can't we just say that we would like features $\\mathbf{z}$ to be uncorrelated? do we need the notation for $p_{data}$, $f^{-1}$.\n\n- Quoting the sentence: \"From Proposition 3.3, we find that SimCLR (InfoNCE) loss is not canonical for achieving matrix information-theoretic uniformity unless the covariance matrix is diagonal\". What do we mean by \"loss being not canonical\"? Has matrix information-theoretic uniformity been defined yet? Is this statement simply saying that  SimCLR or InfoNCE does not enforce feature whitening?\n\n- MCE-based decorrelation objective: why do we have a $\\lambda \\mathbf{I}_d$  perturbation for the desired $\\mathbf{I}_d$ matrix? it is already perfectly conditioned. This perturbation on the first argument of the MCE does not reflect on the right side of \n\n$$\\operatorname{MCE}\\left(\\frac{1}{d} \\mathbf{I}_d+\\lambda \\mathbf{I}_d, \\frac{1}{B} \\mathbf{Z} \\mathbf{Z}^{\\top}+\\lambda \\mathbf{I}_d\\right)=-\\operatorname{tr}\\left(\\log \\left(\\frac{1}{B} \\mathbf{Z Z}^{\\top}+\\lambda \\mathbf{I}_d\\right)\\right)+1+d \\lambda$$\n\nShouldn't there be a multiplier $\\frac{1}{d}$ or $\\frac{1}{d}+\\lambda$ in front of the trace term? I suggest that $\\mathcal{L}_{UMCE}$ should be defined at this stage.\n\n- The paragraph before Theorem 3.5: Suggestion \"This MCE based uniformity loss definition (or $\\mathcal{L}_{UMCE}$ ) and its Matrix-KL divergence based counterpart are closely related.... as outlined by the following theorem:\"\n\n- Theorem 3.6: Suggestion: .... under the constraint $\\||\\mathbf{z}_i\\||_2^2=1$, for $i=1, \\ldots, n$. The proof of Theorem 3.6 better be provided in Appendix.\n\n- Suggestion: The sentence \"Our formulation interestingly recovers the Maximal Entropy Coding (MEC) loss...\" can be written as \"The Maximal Entropy Coding (MEC) loss in ... can be formulated in terms of Matrix MCE..\" as\n\n$$ \\mathcal{L}_{MEC}=-\\mu \\log\\det(\\mathbf{I}_d+\\frac{d}{B\\epsilon^2}\\mathbf{Z}_1\\mathbf{Z}_2^T)$$\n\n$$ =MCE(...., .....)$$\n\n- $\\mathcal{L}_{EMP-TCR}$: $\\bar{\\mathbf{Z}}$ is not defined. Only $\\bar{\\mathbf{Z}}_i$ is defined. Again there is a confusion of index representations relative to  Equation (1). It is understood from this statement that $\\mathbf{z}_k^{i}$ vectors were defined as row vectors. The article should set up the proper data model and notation at the beginning an should stick with that throughout the article.\n\n- Can we also have MCE based representation for the Corinfomax SSL provided in [a] above?\n\n- Overall suggestion: I suggest that the article defines all SSL-related loss functions in Section 2.1, instead of introducing some in Section 2.1 and some in  Section 3.1. Furthermore,  In Section  3.1, the article can clearly write each SSL loss function in the form \nMCE(... , ...) to show that they can be put in the form of matrix cross entropies.\n\n#### 4 MATRIX INFORMATION THEORETIC UNIFORMITY AND ALIGNMENT FOR SELF-SUPERVISED LEARNING\n\n- First sentence: .... we would like embeddings to have zero mean and covariance ....\n\n- Sentence before Theorem 4.1: optimizing covariance matrix uniformity: is this maximizing $\\mathcal{L}_{UMCE}$ or $\\mathcal{L}_{UKL}$. This should be clarified.\n\n- Theorem 4.1. This needs to be clearly reworded with proper references to the objective function and constraints. What is \"effective rank\", how is it different than rank? If this is a constraint how do you pose it?  Is the argument of the MCE in the uniformity-MCE loss sample correlation or sample covariance? Is this theorem about  the following optimization?:\n\n$$ \\text{maximize } \\mathcal{L}_{UMCE}(\\frac{1}{B}\\mathbf{ZZ}^T)$$\n$$ \\text{ subject to } \\text{tr}(\\frac{1}{B}\\mathbf{ZZ}^T)=1$$\n\nThe proof of Theorem 4.1 in the appendix requires a rewrite: Dote (Typo?)  Denote? $\\mathbf{Z}$ can be confused as a matrix due to earlier notation. I guess the first sentence states that Let $\\mathbf{x}$ be a random vector, whose distribution has support $S^{d-1}$. Again what is effective rank? This proof needs to be in the form of a series of explicit mathematical assertions referring to a clearly stated optimization problem.\n\n- Lemma 4.2 is typically well known.\n\n- For $\\mathcal{L}_{Matrix-KL-uniformity}$,  $MCE$ is used not Matrix-KL measure. Why is it called this way?\n\n5 MATRIX-SSL: UNIFORMITY AND ALIGNMENT\n\n- Regarding the alignment cost based on Matrix: \n\n1. Again it is based on MCE rather than Matrix-KL. In fact after (11), it is stated that KL versions can also be considered. So why do you call it $\\mathcal{L}_{Matrix-KL-allignment}$ ?\n\n2. The fact that covariance matrices of two matrices are aligned with respect to MCE or KL does not necessarily imply that representations for the same image are aligned in the direction, where as euclidian distance based or cosine angle based approaches try to ensure that they are sample wise aligned. So why should $\\mathcal{L}_{Matrix-KL-allignment}$ be a better choice?\n\n### 5 EFFECTIVE RANK AND RANK INCREASING PHENOMENON\n\n- It is indeed surprising that effective rank is properly defined and connected to the framework of the article much later than it is already referred. \n\n### 6 EXPERIMENTS\n\n- It would be interesting to include Tong et.al, 2023 and [a] in the experiments for comparison.\n\n- Interestingly, the proposed Matrix-SSL method provides superior performance in experimental results. A natural question to ask if the authors reproduced the accuracy of other algorithms to calibrate their simulation and evaluation models.\n\n### 7 RELATED WORK\n\nThis section typically follows  the Introduction section. Furthermore, it should not be only stating the summary of literature but it should state the contributions of the article relative to these works.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The article aims to introduce a unifying information-theoretic framework for self-supervised learning (SSL). For this purpose, the article\n -  Surveys some established SSL methodologies.,\n - Introduces matrix entropy, matrix KL divergence and matrix cross entropy measures,\n - Expresses certain existing SSL loss functions using the matrix cross entropy measure,\n - Proposes a new SSL loss function derived from matrix cross-entropy,\n - Conducts numerical experiments, demonstrating the enhanced performance of the proposed method compared to select existing approaches,\n - Draws a connection of matrix cross entropy with the effective rank.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The article's pursuit of a unifying framework offers a commendable approach. Strategy to employ  matrix information measures to achieve this is intriguing. Moreover, the numerical examples showcase marked enhancements over certain existing methods, underscoring the efficacy of the algorithm derived from this framework.", "weaknesses": "The article lacks a clear organizational structure and consistent notation, making it challenging to follow. Concepts are introduced without adequate explanation or clarity. Additionally, the matrix information measures employed are not innovative; similar methods have been previously applied in the SSL context. The attempt to frame existing methods as special cases within this framework falls short of being convincing and satisfactory. Please see Questions section for details.", "questions": "## INTRODUCTION\n\n- The following reference,\n\n[a] Ozsoy S, Hamdan S, Arik S, Yuret D, Erdogan A. Self-supervised learning with an information maximization criterion. Advances in Neural Information Processing Systems. 2022 Dec 6;35:35240-53,\n\nproposes utilizing \"correlative information\" maximization for self-supervised learning. Analyzing this paper within the context of the proposed matrix information framework would be interesting, especially since the authors of [a] assert that maximizing correlative information between the representations of augmentations establishes a linear dependence rather than an arbitrary nonlinear one.\n\n- Figure 1: The citation for Coco is absent. It would be beneficial to compare the performances of  Vicreg, [a] (Bardes et. al, 2021), and (Tong et. al, 2023).\n\n\n### 2.1 CONTRASTIVE AND NON-CONTRASTIVE SELF-SUPERVISED LEARNING\n\n- First paragraph: The discussion here is based on the SimCLR and SimSiam, however, the authors introduce a generic SSL architecture. Furthermore, what is meant by dual networks is not clear at this point.\n\n- Would categorizing this section into subheadings like \"Contrastive SSL Approaches\" and \"Non-Contrastive SSL Approaches\" enhance clarity?\n\n- There seems to be inconsistency in the conventions used for sample and augmentation indices in Equations (2) and (3)?\n\n- For Equation (3), is there an underlying assumption about the normalization of the encoder,  such as  $\\||z_i^{(k)}\\||_2=1$ ?\n\n\n### 2.2 MATRIX INFORMATION-THEORETIC QUANTITIES\n- Could you provide a citation detailing Matrix Entropy. Furthermore, could you discuss its interpretation, and perhaps its existing applications especially within the context of machine learning/SSL?\n\n- Regarding the Matrix Entropy definition, is there a specific assumption about the trace of $\\mathbf{A}$ ensuring its eigenvalues form a probability mass function?\n\n- Can you provide interpretations and potential applications of Matrix KL Divergence and MCE?\n\n- Bach 2022's KL divergence doesn't seem to incorporate the  $-\\mathbf{P}+\\mathbf{Q}$ terms within trace? Could this discrepancy be addressed?\n\n- From the presented definitions, it appears that, unlike Shannon Entropy, MCE does not equate to the sum of Matrix KL and Matrix Entropy. Should there be a $\\text{Tr}(\\mathbf{P})$ term included in the matrix entropy definition?\n\n- A brief discussion explaining the relevance of these definitions to the SSL problem would be insightful.\n\n\n### 3 MATRIX INFORMATION-THEORETIC PERSPECTIVES OF SELF-SUPERVISED LEARNING\n\n- The assertion about proofs should be positioned adjacent to the first proposition, i.e., Proposition 3.1.\n\n- Proposition 3.3: The statement of the proposition is  ambigious:  InfoNCE cost in (1) is the MCE of which matrices? This should be clearly stated. Are InfNCE cost and SimCLR cost identical? The cost function obtained in the proof in terms of MCE does not match (1)?\n- The paragraph after Proposition 3.3: This part appears convoluted: Is $p_{data}=p_{\\mathbf{x}}$ ? and $\\mathcal{X}$ is the support set of $p_{\\mathbf{x}}$? I guess there is no clear definition of $f$ before. (there was $f_\\Theta$ and $f_\\phi$ without clear definitions before). $f$ is sometimes unbold and sometimes bold? Since $f$ is not prespecified, the assumption that there is a layer normalization $\\||f(\\mathbf{x})\\||_2^2=1$ is also not clear. Instead of stating \"straightforward calculation\", it is better to provide a proof of Lemma 3.4 with proper notation in Appendix A. In my opinion, both this paragraph and Lemma 3.4 is not properly motivated.\n\n- Lemma 3.4: Suggestion \"Let $\\sigma$ represent the uniform distribution on $S^{d-1}$. ...\".\n\n- The paragraph after Lemma 3.4: Change of variables formula? (Inverse image rule?). Why \"auto-correlation\" matrix for $q$ but \"covariance\" for $p_{data}$?  It is better for the authors to clearly state the uniformity principle. Can't we just say that we would like features $\\mathbf{z}$ to be uncorrelated? do we need the notation for $p_{data}$, $f^{-1}$.\n\n- Quoting the sentence: \"From Proposition 3.3, we find that SimCLR (InfoNCE) loss is not canonical for achieving matrix information-theoretic uniformity unless the covariance matrix is diagonal\". What do we mean by \"loss being not canonical\"? Has matrix information-theoretic uniformity been defined yet? Is this statement simply saying that  SimCLR or InfoNCE does not enforce feature whitening?\n\n- MCE-based decorrelation objective: why do we have a $\\lambda \\mathbf{I}_d$  perturbation for the desired $\\mathbf{I}_d$ matrix? it is already perfectly conditioned. This perturbation on the first argument of the MCE does not reflect on the right side of \n\n$$\\operatorname{MCE}\\left(\\frac{1}{d} \\mathbf{I}_d+\\lambda \\mathbf{I}_d, \\frac{1}{B} \\mathbf{Z} \\mathbf{Z}^{\\top}+\\lambda \\mathbf{I}_d\\right)=-\\operatorname{tr}\\left(\\log \\left(\\frac{1}{B} \\mathbf{Z Z}^{\\top}+\\lambda \\mathbf{I}_d\\right)\\right)+1+d \\lambda$$\n\nShouldn't there be a multiplier $\\frac{1}{d}$ or $\\frac{1}{d}+\\lambda$ in front of the trace term? I suggest that $\\mathcal{L}_{UMCE}$ should be defined at this stage.\n\n- The paragraph before Theorem 3.5: Suggestion \"This MCE based uniformity loss definition (or $\\mathcal{L}_{UMCE}$ ) and its Matrix-KL divergence based counterpart are closely related.... as outlined by the following theorem:\"\n\n- Theorem 3.6: Suggestion: .... under the constraint $\\||\\mathbf{z}_i\\||_2^2=1$, for $i=1, \\ldots, n$. The proof of Theorem 3.6 better be provided in Appendix.\n\n- Suggestion: The sentence \"Our formulation interestingly recovers the Maximal Entropy Coding (MEC) loss...\" can be written as \"The Maximal Entropy Coding (MEC) loss in ... can be formulated in terms of Matrix MCE..\" as\n\n$$ \\mathcal{L}_{MEC}=-\\mu \\log\\det(\\mathbf{I}_d+\\frac{d}{B\\epsilon^2}\\mathbf{Z}_1\\mathbf{Z}_2^T)$$\n\n$$ =MCE(...., .....)$$\n\n- $\\mathcal{L}_{EMP-TCR}$: $\\bar{\\mathbf{Z}}$ is not defined. Only $\\bar{\\mathbf{Z}}_i$ is defined. Again there is a confusion of index representations relative to  Equation (1). It is understood from this statement that $\\mathbf{z}_k^{i}$ vectors were defined as row vectors. The article should set up the proper data model and notation at the beginning an should stick with that throughout the article.\n\n- Can we also have MCE based representation for the Corinfomax SSL provided in [a] above?\n\n- Overall suggestion: I suggest that the article defines all SSL-related loss functions in Section 2.1, instead of introducing some in Section 2.1 and some in  Section 3.1. Furthermore,  In Section  3.1, the article can clearly write each SSL loss function in the form \nMCE(... , ...) to show that they can be put in the form of matrix cross entropies.\n\n#### 4 MATRIX INFORMATION THEORETIC UNIFORMITY AND ALIGNMENT FOR SELF-SUPERVISED LEARNING\n\n- First sentence: .... we would like embeddings to have zero mean and covariance ....\n\n- Sentence before Theorem 4.1: optimizing covariance matrix uniformity: is this maximizing $\\mathcal{L}_{UMCE}$ or $\\mathcal{L}_{UKL}$. This should be clarified.\n\n- Theorem 4.1. This needs to be clearly reworded with proper references to the objective function and constraints. What is \"effective rank\", how is it different than rank? If this is a constraint how do you pose it?  Is the argument of the MCE in the uniformity-MCE loss sample correlation or sample covariance? Is this theorem about  the following optimization?:\n\n$$ \\text{maximize } \\mathcal{L}_{UMCE}(\\frac{1}{B}\\mathbf{ZZ}^T)$$\n$$ \\text{ subject to } \\text{tr}(\\frac{1}{B}\\mathbf{ZZ}^T)=1$$\n\nThe proof of Theorem 4.1 in the appendix requires a rewrite: Dote (Typo?)  Denote? $\\mathbf{Z}$ can be confused as a matrix due to earlier notation. I guess the first sentence states that Let $\\mathbf{x}$ be a random vector, whose distribution has support $S^{d-1}$. Again what is effective rank? This proof needs to be in the form of a series of explicit mathematical assertions referring to a clearly stated optimization problem.\n\n- Lemma 4.2 is typically well known.\n\n- For $\\mathcal{L}_{Matrix-KL-uniformity}$,  $MCE$ is used not Matrix-KL measure. Why is it called this way?\n\n5 MATRIX-SSL: UNIFORMITY AND ALIGNMENT\n\n- Regarding the alignment cost based on Matrix: \n\n1. Again it is based on MCE rather than Matrix-KL. In fact after (11), it is stated that KL versions can also be considered. So why do you call it $\\mathcal{L}_{Matrix-KL-allignment}$ ?\n\n2. The fact that covariance matrices of two matrices are aligned with respect to MCE or KL does not necessarily imply that representations for the same image are aligned in the direction, where as euclidian distance based or cosine angle based approaches try to ensure that they are sample wise aligned. So why should $\\mathcal{L}_{Matrix-KL-allignment}$ be a better choice?\n\n### 5 EFFECTIVE RANK AND RANK INCREASING PHENOMENON\n\n- It is indeed surprising that effective rank is properly defined and connected to the framework of the article much later than it is already referred. \n\n### 6 EXPERIMENTS\n\n- It would be interesting to include Tong et.al, 2023 and [a] in the experiments for comparison.\n\n- Interestingly, the proposed Matrix-SSL method provides superior performance in experimental results. A natural question to ask if the authors reproduced the accuracy of other algorithms to calibrate their simulation and evaluation models.\n\n### 7 RELATED WORK\n\nThis section typically follows  the Introduction section. Furthermore, it should not be only stating the summary of literature but it should state the contributions of the article relative to these works.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698690319445}, {"id": "wxsREwohua", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission524/Reviewer_CuLc"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors investigate self-supervised learning through the lens of matrix information theory. They present a unified theoretical framework for analyzing both contrastive and non-contrastive learning methods. Specifically, they employ matrix cross-entropy as the training objective to enhance uniformity and alignment, thereby improving self-supervised learning. Experiments conducted on ImageNet and COCO datasets demonstrate that the proposed method outperforms existing classical approaches.", "review_text": "In this paper, the authors investigate self-supervised learning through the lens of matrix information theory. They present a unified theoretical framework for analyzing both contrastive and non-contrastive learning methods. Specifically, they employ matrix cross-entropy as the training objective to enhance uniformity and alignment, thereby improving self-supervised learning. Experiments conducted on ImageNet and COCO datasets demonstrate that the proposed method outperforms existing classical approaches.", "strengths": "1. This paper studies self-supervised learning through a matrix information-theoretic framework. The analysis presented in this paper is particularly intriguing and I find it quite appealing. \n\n2. The authors further introduce a Matrix-SSL scheme based on matrix cross-entropy, which consists of matrix uniformity and matrix alignment. \n\n3. The experiments on the ImageNet and COCO datasets not only show that the proposed method surpasses state-of-the-art methods but also highlight its robustness in transfer learning tasks.", "weaknesses": "1. There are some issues with the mathematical symbol definitions in this paper, such as inconsistency in the usage of symbols, missing definitions for certain symbols, and incorrect usage of mathematical symbols. For example, on the second page, the lowercase letter \"z\" represents features, and in subsequent chapters, the bolded lowercase letter \"**z**\" also represents features. In the part of the definition of matrix entropy, the definition of the lowercase letter $\\lambda$ is missing. In the proof of proposition 3.3, there is something wrong with the infoNCE loss. I suggest that the authors follow the definitions provided by the original authors in their arXiv paper.\n2. How was Lemma 3.4 obtained? I understand the purpose of this Lemma, but it's better to give the proof or the corresponding reference. On the other hand, Lemma 3.4 shows that minimizing matrix cross-entropy between the Identity diagonal matrix and the covariance matrix can achieve a uniformity target. However, starting from the fourth page, the zero-mean assumption is disregarded. Does this have any impact on the theoretical analysis results? \n3. Starting from the third page, the authors consistently assume that the feature matrix is positive semi-definite. However, can this constraint be maintained in practice？\n4. In section 3, the authors analyze that matrix information theory could provide a unified framework for many existing SSL methods. Then, according to Theorem 3.5, Uniformity-MCE loss is equal to MEC loss. The experiments in Table 3 can verify this, where the result of Matrix-SSL (when $\\gamma=0$ ) is equal to that of MEC (70.6%). With an increase in $\\gamma$, the results will improve. This means that matrix alignment is indeed helpful for final performance improvements. Therefore, if we consider the alignment term along with the MEC loss, what will be the results? I suggest the authors conduct a detailed analysis of the differences between the MEC loss and the Uniformity-MCE loss, especially from an experimental perspective. I wonder if the gradient computation for the Uniformity-MCE loss is easier compared to the MEC loss.\n5. Although matrix-KL and matrix-CE share similar optimization properties and theoretical results, are they consistent in practical experiments? I recommend that the authors conduct a set of experiments to validate this.", "questions": "Please check the questions in the Weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors investigate self-supervised learning through the lens of matrix information theory. They present a unified theoretical framework for analyzing both contrastive and non-contrastive learning methods. Specifically, they employ matrix cross-entropy as the training objective to enhance uniformity and alignment, thereby improving self-supervised learning. Experiments conducted on ImageNet and COCO datasets demonstrate that the proposed method outperforms existing classical approaches.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. This paper studies self-supervised learning through a matrix information-theoretic framework. The analysis presented in this paper is particularly intriguing and I find it quite appealing. \n\n2. The authors further introduce a Matrix-SSL scheme based on matrix cross-entropy, which consists of matrix uniformity and matrix alignment. \n\n3. The experiments on the ImageNet and COCO datasets not only show that the proposed method surpasses state-of-the-art methods but also highlight its robustness in transfer learning tasks.", "weaknesses": "1. There are some issues with the mathematical symbol definitions in this paper, such as inconsistency in the usage of symbols, missing definitions for certain symbols, and incorrect usage of mathematical symbols. For example, on the second page, the lowercase letter \"z\" represents features, and in subsequent chapters, the bolded lowercase letter \"**z**\" also represents features. In the part of the definition of matrix entropy, the definition of the lowercase letter $\\lambda$ is missing. In the proof of proposition 3.3, there is something wrong with the infoNCE loss. I suggest that the authors follow the definitions provided by the original authors in their arXiv paper.\n2. How was Lemma 3.4 obtained? I understand the purpose of this Lemma, but it's better to give the proof or the corresponding reference. On the other hand, Lemma 3.4 shows that minimizing matrix cross-entropy between the Identity diagonal matrix and the covariance matrix can achieve a uniformity target. However, starting from the fourth page, the zero-mean assumption is disregarded. Does this have any impact on the theoretical analysis results? \n3. Starting from the third page, the authors consistently assume that the feature matrix is positive semi-definite. However, can this constraint be maintained in practice？\n4. In section 3, the authors analyze that matrix information theory could provide a unified framework for many existing SSL methods. Then, according to Theorem 3.5, Uniformity-MCE loss is equal to MEC loss. The experiments in Table 3 can verify this, where the result of Matrix-SSL (when $\\gamma=0$ ) is equal to that of MEC (70.6%). With an increase in $\\gamma$, the results will improve. This means that matrix alignment is indeed helpful for final performance improvements. Therefore, if we consider the alignment term along with the MEC loss, what will be the results? I suggest the authors conduct a detailed analysis of the differences between the MEC loss and the Uniformity-MCE loss, especially from an experimental perspective. I wonder if the gradient computation for the Uniformity-MCE loss is easier compared to the MEC loss.\n5. Although matrix-KL and matrix-CE share similar optimization properties and theoretical results, are they consistent in practical experiments? I recommend that the authors conduct a set of experiments to validate this.", "questions": "Please check the questions in the Weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698652226588}, {"id": "U7hXuhxOhM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission524/Reviewer_m2RW"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper presents Matrix-SSL, a joint-embedding SSL method based on matrix information theory.\nSpecifically, the uniformity and alignment framework is implemented using principles from matrix information theory.\nThe results of this study demonstrate that Matrix-SSL surpasses prior state-of-the-art (SOTA) SSL methods.", "review_text": "This paper presents Matrix-SSL, a joint-embedding SSL method based on matrix information theory.\nSpecifically, the uniformity and alignment framework is implemented using principles from matrix information theory.\nThe results of this study demonstrate that Matrix-SSL surpasses prior state-of-the-art (SOTA) SSL methods.", "strengths": "This paper introduces a matrix-based information-theoretic framework that provides a comprehensive explanation for self-supervised learning methods, including both contrastive learning and non-contrastive learning.", "weaknesses": "- According to Propositions 3.1 and 5.2, it can be established that the Matrix-KL-uniformity loss is synonymous with the von Neumann entropy loss of I-VNE+ as proposed in [1]. This similarity diminishes the novelty of this paper. Therefore, it is imperative to substantiate, either through theoretical or empirical means, the superiority of Matrix-SSL in comparison to I-VNE+.\n- The results presented in this paper are not significant to substantiate the effectiveness of Matrix-SSL. Notably, Table 1 does not incorporate the official performance metrics of SwAV, as reported in [2], which report values of 71.99, 73.85, and 74.81 for 100, 200, and 400 training epochs, respectively. Additionally, Table 2 lacks the inclusion of performance data as reported in [1]. When both Table 1 and Table 2 are appropriately updated, it becomes evident that the performance of Matrix-SSL is not state-of-the-art.\nFurthermore, it is important to note that this paper does not provide comprehensive benchmark tables, including but not limited to \"Semi-supervised learning on ImageNet\" and \"Transfer learning: image classification,\" as elaborated in Table 2 and Table 3 of [3].\n- In Section 5, this paper demonstrates that enhancing uniformity leads to an increased effective rank through matrix entropy. However, this result is not groundbreaking. In [1], the authors have previously presented these mathematical findings and have empirically shown that von Neumann entropy regulates uniformity, thereby influencing the effective rank.\n\n[1] VNE: An Effective Method for Improving Deep Representation by Manipulating Eigenvalue Distribution, CVPR 2023.\n\n[2] https://github.com/facebookresearch/vissl/blob/main/MODEL_ZOO.md\n\n[3] Barlow Twins: Self-Supervised Learning via Redundancy Reduction, ICML 2021.", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Matrix-SSL, a joint-embedding SSL method based on matrix information theory.\nSpecifically, the uniformity and alignment framework is implemented using principles from matrix information theory.\nThe results of this study demonstrate that Matrix-SSL surpasses prior state-of-the-art (SOTA) SSL methods.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This paper introduces a matrix-based information-theoretic framework that provides a comprehensive explanation for self-supervised learning methods, including both contrastive learning and non-contrastive learning.", "weaknesses": "- According to Propositions 3.1 and 5.2, it can be established that the Matrix-KL-uniformity loss is synonymous with the von Neumann entropy loss of I-VNE+ as proposed in [1]. This similarity diminishes the novelty of this paper. Therefore, it is imperative to substantiate, either through theoretical or empirical means, the superiority of Matrix-SSL in comparison to I-VNE+.\n- The results presented in this paper are not significant to substantiate the effectiveness of Matrix-SSL. Notably, Table 1 does not incorporate the official performance metrics of SwAV, as reported in [2], which report values of 71.99, 73.85, and 74.81 for 100, 200, and 400 training epochs, respectively. Additionally, Table 2 lacks the inclusion of performance data as reported in [1]. When both Table 1 and Table 2 are appropriately updated, it becomes evident that the performance of Matrix-SSL is not state-of-the-art.\nFurthermore, it is important to note that this paper does not provide comprehensive benchmark tables, including but not limited to \"Semi-supervised learning on ImageNet\" and \"Transfer learning: image classification,\" as elaborated in Table 2 and Table 3 of [3].\n- In Section 5, this paper demonstrates that enhancing uniformity leads to an increased effective rank through matrix entropy. However, this result is not groundbreaking. In [1], the authors have previously presented these mathematical findings and have empirically shown that von Neumann entropy regulates uniformity, thereby influencing the effective rank.\n\n[1] VNE: An Effective Method for Improving Deep Representation by Manipulating Eigenvalue Distribution, CVPR 2023.\n\n[2] https://github.com/facebookresearch/vissl/blob/main/MODEL_ZOO.md\n\n[3] Barlow Twins: Self-Supervised Learning via Redundancy Reduction, ICML 2021.", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698316464492}], "openreview_url": "https://openreview.net/forum?id=e1IMBXiDhW", "arxiv_id": "2305.17326", "paper_pdf": "papers/e1IMBXiDhW.pdf", "paper_pdf_sha256": "f4bf75d7b9f0d1c4242d007bb875987b2189287c8ede929e33d813b38e68a51e", "paper_pdf_bytes": 977142, "paper_pdf_source": "openreview", "code_url": "https://github.com/yifanzhang-pro/Matrix-SSL", "code_repository": "yifanzhang-pro/Matrix-SSL", "code_commit": "b8164dfcb279f673654d94d544f0b73e26f152d8", "code_archive": "repos/e1IMBXiDhW.zip", "code_archive_sha256": "6b12b87d4db0a36a697e74aa18e9297f71ad4bd90d28365675b0815d0c55bf6b", "code_archive_bytes": 19843, "code_file_count": 6, "code_extensions": {".py": 4, ".sh": 2}, "github_disk_usage_kb": 57, "github_languages": {"Python": 47611, "Shell": 383}, "github_archived": false, "github_pushed_at": "2025-09-21T05:02:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/kernel-ssl-kernel-kl-divergence-for-self"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zfiYcbeQkH", "year": 2023, "status": "rejected", "title": "SciRepEval: A Multi-Format Benchmark for Scientific Document Representations", "authors": ["Amanpreet Singh", "Mike D'Arcy", "Arman Cohan", "Doug Downey", "Sergey Feldman"], "authorids": ["~Amanpreet_Singh2", "~Mike_D'Arcy1", "~Arman_Cohan1", "~Doug_Downey1", "~Sergey_Feldman1"], "authors_source": "OpenReview API", "abstract": "Learned representations of scientific documents can serve as valuable input features for downstream tasks, without the need for further fine-tuning. However, existing benchmarks for evaluating these representations fail to capture the diversity of relevant tasks. In response, we introduce SciRepEval, the first comprehensive benchmark for training and evaluating scientific document representations. It includes 25 challenging and realistic tasks across four formats: classification, regression, ranking and search. We then use the benchmark to study and improve the generalization ability of scientific document representation models.  We show how state-of-the-art models struggle to generalize across task formats, and that simple multi-task training fails to improve them.  However, a new approach that learns multiple embeddings per document, each tailored to a different task format, can improve performance.\nWe experiment with task-format-specific control codes and adapters in a multi-task setting and find that they outperform the existing single-embedding state-of-the-art by up to 1.5 points absolute. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AKYRUOCmpRh", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6140/Reviewer_uueE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new benchmark (SciRepEval) for scientific representation learning consisting of 25 tasks in 4 formats (classification, regression, ranking, and search). It shows that learning a separate document representation for each task format would improve the task performance compared to learning a single representation for all tasks.", "review_text": "This paper is written in above-average quality and provides a new benchmark with a comprehensive analysis", "strengths": "Strength\n- The authors provide a comprehensive analysis on 25 tasks in 4 formats. \n- The authors use strong baseline models to make the results more convincing.\n\nWeakness\n- There is not enough information on the tasks in the main content. Would be better to provide some high-level info there and left the majority in the Appendix.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a new benchmark (SciRepEval) for scientific representation learning consisting of 25 tasks in 4 formats (classification, regression, ranking, and search). It shows that learning a separate document representation for each task format would improve the task performance compared to learning a single representation for all tasks.", "strength_and_weaknesses": "Strength\n- The authors provide a comprehensive analysis on 25 tasks in 4 formats. \n- The authors use strong baseline models to make the results more convincing.\n\nWeakness\n- There is not enough information on the tasks in the main content. Would be better to provide some high-level info there and left the majority in the Appendix.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written with high clarity and above-average quality.\nSome concerns on the Novelty, since the paper doesn't go deep on the technique.", "summary_of_the_review": "This paper is written in above-average quality and provides a new benchmark with a comprehensive analysis", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1672985070771}, {"id": "196nCVxVANN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6140/Reviewer_RNDk"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new benchmark for scientific document representations, including 25 tasks across classification, regression, ranking and search. The paper also includes investigations on performances of existing models on the introduced tasks. ", "review_text": "While the paper provides a useful benchmark, there is a strong component of novelty missing from the paper. ICLR does not seem like the correct venue for this paper to be presented. ", "strengths": "Strengths: \n- The paper consolidates useful tasks to make one benchmark for scientific document representations\n\nWeaknesses: \n- The models mentioned in the paper are mostly created by others in the community\n- The benchmark, while useful, is a consolidation of tasks that were already existing \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces a new benchmark for scientific document representations, including 25 tasks across classification, regression, ranking and search. The paper also includes investigations on performances of existing models on the introduced tasks. ", "strength_and_weaknesses": "Strengths: \n- The paper consolidates useful tasks to make one benchmark for scientific document representations\n\nWeaknesses: \n- The models mentioned in the paper are mostly created by others in the community\n- The benchmark, while useful, is a consolidation of tasks that were already existing \n", "clarity,_quality,_novelty_and_reproducibility": "I did not find the paper to be novel, but the quality and clarity of the writing was sufficient", "summary_of_the_review": "While the paper provides a useful benchmark, there is a strong component of novelty missing from the paper. ICLR does not seem like the correct venue for this paper to be presented. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667264560068}, {"id": "POSOW6sPRC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6140/Reviewer_d6LN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors of this paper introduce a benchmark for scientific representation learning consisting of 25 tasks in 4 formats (classificaiton, regression, ranking, search) and evaluate multiple general-purpose scientific representation learning methods (SciBERt, SPECTER, SciNL) alongside adaptation methods to learn task-specific and format-specific representations. The results suggest that fine-grained scientific document representations significantly out-perform general-purpose representations.", "review_text": "The authors identify an area of opportunity relating to current datasets and methods to evaluate scientific representation learning models and propose a well-grounded, larger benchmark in SciRepEval. The ablation study is comprehensive and there is some interesting discussion of cross-task analysis, although I would have liked to see that section expanded.", "strengths": "Strengths:\n- Introduces a more comprehensive benchmark for scientific document representation learning\n- Clearly distinguishes SciRepEval from previous benchmarks (SciDocs) and justifies why a new benchmark is needed (low-power of the recommendation task, a need for more realistic task settings, and lack of training corpus)\n- Comprehensive ablation setup targeting fine-grained representation learning, multi-task learning with a strong base model\n\nWeaknesses:\n- It would have been nice to see an explicit section detailing what additional challenges remain in benchmarking scientific representation learning and how these limitations apply to the results seen with multi-task approaches.\n- Would have liked to see more discussion about task relatedness when discussing benchmarks, especially since SciRepEval seems to be designed in order to improve breadth and coverage of common real-world applications of the representations. For example, it would be interesting to see the relationship as applied to the task of choosing which sub-tasks to group for multi-task learning (see Efficiently Identifying Task Groupings for Multi-Task Learning; Fifty et al. NeurIPS 2021).", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors of this paper introduce a benchmark for scientific representation learning consisting of 25 tasks in 4 formats (classificaiton, regression, ranking, search) and evaluate multiple general-purpose scientific representation learning methods (SciBERt, SPECTER, SciNL) alongside adaptation methods to learn task-specific and format-specific representations. The results suggest that fine-grained scientific document representations significantly out-perform general-purpose representations.", "strength_and_weaknesses": "Strengths:\n- Introduces a more comprehensive benchmark for scientific document representation learning\n- Clearly distinguishes SciRepEval from previous benchmarks (SciDocs) and justifies why a new benchmark is needed (low-power of the recommendation task, a need for more realistic task settings, and lack of training corpus)\n- Comprehensive ablation setup targeting fine-grained representation learning, multi-task learning with a strong base model\n\nWeaknesses:\n- It would have been nice to see an explicit section detailing what additional challenges remain in benchmarking scientific representation learning and how these limitations apply to the results seen with multi-task approaches.\n- Would have liked to see more discussion about task relatedness when discussing benchmarks, especially since SciRepEval seems to be designed in order to improve breadth and coverage of common real-world applications of the representations. For example, it would be interesting to see the relationship as applied to the task of choosing which sub-tasks to group for multi-task learning (see Efficiently Identifying Task Groupings for Multi-Task Learning; Fifty et al. NeurIPS 2021).", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is written clearly and the logic flows well.\n\nQuality: I believe this is a high quality paper with interesting results and strong justification.\n\nNovelty: The authors identify a challenge with existing benchmarks and release a larger and more diverse benchmark. This is an important and novel contribution.\n\nReproducibility: Data is promised to be released; no mention of code release.", "summary_of_the_review": "The authors identify an area of opportunity relating to current datasets and methods to evaluate scientific representation learning models and propose a well-grounded, larger benchmark in SciRepEval. The ablation study is comprehensive and there is some interesting discussion of cross-task analysis, although I would have liked to see that section expanded.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666536187881}, {"id": "NZac4_IDVY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6140/Reviewer_tJad"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces SciRepEval, which is a novel benchmark for training and evaluating scientific document representations. Additionally, the struggles of multi-task learning with regards to generalization in these tasks is explored. An alternative that explicitly encodes the task type is offered and evaluated.\n", "review_text": "This paper extends an existing benchmark substantially, and offers a new approach to multi-task learning meant to deal with the multitude of tasks in a way that is not wholly novel, and beats the baseline, although only slightly. \n", "strengths": "Strengths\n\n- Not only are we treated to a single task in this benchmark, but 25, including classification, regression, and recommendation, each of which have appropriate measures for evaluation (though other measures may also have been appropriate).\n- Given the variety of tasks, a ‘multi-format’ method of representation learning is proposed which is not entirely novel on its own, but is additively useful within this paper more ostensibly about the benchmarking suite.\n- Suitable baselines (SciBERT, SPECTER) and several extensions (e.g., CTRL-code and embedding adapters) provide a suitable evaluation suite.\n\n\nWeaknesses\n\n- Although much of the information is found in the Appendices, more detail on the tasks themselves are expected within the main body, especially with regards to the source of the data, which could likely be added to Table 1, in some form.\n- Tables 2 and 3 and their discussion in the text require some additional context to evaluate these results, especially with regards to the units of measure and highlighting that the first two columns are not exactly comparable with the third (SciDocs). Given the relative similarities of the performance between SPECTER and SCiNCL (especially considering the variants of each), statistical significance tests should be performed.  \n\n\nMinor\n\n- The use of task-specific control codes is reminiscent (somewhat) of ELMo, and a comparison or acknowledgement should be made to some of the literature around that model.\n- Some further explanation around the deficiencies of SciDocs in Sec 2 (especially with regards to the correlations and easiness of negative candidates) really should be provided, especially as SciRepEval’s superiority is meant to be evaluated in contrast to the former.\n- There is a higher-than-usual amount of repetitiveness, especially around the 25 tasks in SciRepVal in the first few pages, but this is a very minor complaint\n- Some deeper dive into how performance is affected by covariates such as the field of study (Table 8) after one filters out the effect of data set size would be interesting.\n- Be sure to check the formatting of your references\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces SciRepEval, which is a novel benchmark for training and evaluating scientific document representations. Additionally, the struggles of multi-task learning with regards to generalization in these tasks is explored. An alternative that explicitly encodes the task type is offered and evaluated.\n", "strength_and_weaknesses": "Strengths\n\n- Not only are we treated to a single task in this benchmark, but 25, including classification, regression, and recommendation, each of which have appropriate measures for evaluation (though other measures may also have been appropriate).\n- Given the variety of tasks, a ‘multi-format’ method of representation learning is proposed which is not entirely novel on its own, but is additively useful within this paper more ostensibly about the benchmarking suite.\n- Suitable baselines (SciBERT, SPECTER) and several extensions (e.g., CTRL-code and embedding adapters) provide a suitable evaluation suite.\n\n\nWeaknesses\n\n- Although much of the information is found in the Appendices, more detail on the tasks themselves are expected within the main body, especially with regards to the source of the data, which could likely be added to Table 1, in some form.\n- Tables 2 and 3 and their discussion in the text require some additional context to evaluate these results, especially with regards to the units of measure and highlighting that the first two columns are not exactly comparable with the third (SciDocs). Given the relative similarities of the performance between SPECTER and SCiNCL (especially considering the variants of each), statistical significance tests should be performed.  \n\n\nMinor\n\n- The use of task-specific control codes is reminiscent (somewhat) of ELMo, and a comparison or acknowledgement should be made to some of the literature around that model.\n- Some further explanation around the deficiencies of SciDocs in Sec 2 (especially with regards to the correlations and easiness of negative candidates) really should be provided, especially as SciRepEval’s superiority is meant to be evaluated in contrast to the former.\n- There is a higher-than-usual amount of repetitiveness, especially around the 25 tasks in SciRepVal in the first few pages, but this is a very minor complaint\n- Some deeper dive into how performance is affected by covariates such as the field of study (Table 8) after one filters out the effect of data set size would be interesting.\n- Be sure to check the formatting of your references\n", "clarity,_quality,_novelty_and_reproducibility": "- The paper is well-written. It does not go into very much technical depth, though it doesn’t need to as a benchmarking paper, although the nature of the data could have been explored more fully. It is an extension of a previous benchmark in this space; while it is a substantial extension, this still limits its novelty. \n", "summary_of_the_review": "This paper extends an existing benchmark substantially, and offers a new approach to multi-task learning meant to deal with the multitude of tasks in a way that is not wholly novel, and beats the baseline, although only slightly. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666473746927}], "openreview_url": "https://openreview.net/forum?id=zfiYcbeQkH", "arxiv_id": "2211.13308", "paper_pdf": "papers/zfiYcbeQkH.pdf", "paper_pdf_sha256": "a87de0a9e40d306f5de4efa34c156738b197d2317a007df350b26f76a0a74a2e", "paper_pdf_bytes": 616170, "paper_pdf_source": "openreview", "code_url": "https://github.com/allenai/SPECTER2", "code_repository": "allenai/SPECTER2", "code_commit": "fac1cb0940fe7bd3c0db00973f6889ca0b481b63", "code_archive": "repos/zfiYcbeQkH.zip", "code_archive_sha256": "3ab2e8bb6096640cf33859f6ba3ec89290360540653a644a337b511acf1e8f39", "code_archive_bytes": 59393, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 97, "github_languages": {"Python": 12946}, "github_archived": false, "github_pushed_at": "2026-02-24T21:56:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/scirepeval-a-multi-format-benchmark-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MTsBazXmX00", "year": 2022, "status": "rejected", "title": "Target Propagation via Regularized Inversion", "authors": ["Vincent Roulet", "Zaid Harchaoui"], "authorids": ["~Vincent_Roulet1", "~Zaid_Harchaoui1"], "authors_source": "OpenReview API", "abstract": "Target Propagation (TP) algorithms compute targets instead of gradients along neural networks and propagate them backward in a way that is similar yet different than gradient back-propagation (BP). The idea was first presented as a perturbative alternative to back-propagation that may achieve greater accuracy in gradient evaluation when training multi-layer neural networks (LeCun et al., 1989). However, TP has remained more of a template algorithm with many variations than a well-identified algorithm. Revisiting insights of LeCun et al., (1989) and more recently of Lee et al. (2015), we present a simple version of target propagation based on regularized inversion of network layers, easily implementable in a differentiable programming framework. We compare its computational complexity to the one of BP and delineate the regimes in which TP can be attractive compared to BP. We show how our TP can be used to train recurrent neural networks with long sequences on various sequence modeling problems. The experimental results underscore the importance of regularization in TP in practice.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "YHy6Rls_bf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4414/Reviewer_NQCB"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors study a variant of target propagation in which targets are computed by solving a sequence of minimization problems. Instead of resorting to iterative methods the authors propose to use an analytical solution. The algorithm is investigated as a recurrent neural network learning algorithm in a number of experiments. A relationship between the proposed descent direction and the loss function gradient is developed.", "review_text": "While I find target propagation (TP) interesting, I am not very enthusiastic about the current paper. As detailed below, the main aspects that attract me towards TP are absent from the proposed algorithm, and its merits as a generic drop-in replacement for backpropagation are not strong enough or at least not yet well argued for.\n\nThe authors propose a new prescription to compute targets in TP. Unlike much of the recent work on TP, this prescription leads to a parameter update that requires weight transport. This limits its interest and narrows the scope of the contributions, as optimization without weight transport is of great interest not just in theoretical neuroscience but also for neuromorphic hardware design and large-scale distributed network implementations. \n\nAs noted by the authors, the proposed algorithm is also strictly costlier than backpropagation, although the relative difference in cost becomes smaller as sequence length grows. It also does not allow for decoupled updates, or at least this is not commented in the paper. Its applicability to decentralized optimization in distributed settings (another setting of potential interest for TP algorithms) is therfore not clear.\n\nFor the proposed algorithm to be interesting, it therefore has to produce descent directions that lead to faster optimization than gradients. The paper does not provide any theoretical guarantees that this is the case. The merits of the new algorithm must then be judged by the experimental results. This is a very difficult question to answer experimentally, and I did not find the results extremely convincing, but I am admittedly not an expert in sequence learning problems.\n\nI have a number of questions for the authors:\n- In page 3, below eq. 2, it is mentioned that DTP is a 0-th order approximation. Can the authors clarify this remark? In my understanding, the difference correction of DTP stems from a first-order expansion of $f$, see Appendix B of Lee et al. (2015). In what sense is it a 0-th order correction?\n\n- What are the new insights provided in section 2.2 compared to what is already used by Lee et al. to arrive at DTP (and perhaps more clearly restated in Lemma 1 of Meulemans et al. 2020)?\n\n-  I'm having trouble with eq. 4 and the text that follows. What happens when there are multiple minimizers of the objective presented in (4), for example when $a$ is the ReLU activation? How should the next unnumbered equation appearing below \"Formally, for [...] the ReLU, their inverse can be obtained analytically\" be understood in this case?\n\n- There are numerous remarks on the vanishing/exploding gradient problem and allusions to the promise of solving this problem with TP. But since the proposed parameter update relies on products of Jacobian inverses, isn't it also subject to vanishing and exploding gradient problems?\n\n- Why is the relationship between TP and Gauss-Newton optimization pointed out by Bengio (2020) and Meulemans et al. (2020) called into question due to the difficulties observed when setting the regularizer to zero? \n\n- Is GEMINI (LeCun et al., 1989), cited already in the abstract, the desired reference? To me, GEMINI is more obviously related to synthetic gradients than target propagation, and it is not usually cited as LeCun's original TP paper. Reference [1] below is perhaps more appropriate.\n\n[1] LeCun (1986) \"Learning processes in an asymmetric threshold network\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors study a variant of target propagation in which targets are computed by solving a sequence of minimization problems. Instead of resorting to iterative methods the authors propose to use an analytical solution. The algorithm is investigated as a recurrent neural network learning algorithm in a number of experiments. A relationship between the proposed descent direction and the loss function gradient is developed.", "main_review": "While I find target propagation (TP) interesting, I am not very enthusiastic about the current paper. As detailed below, the main aspects that attract me towards TP are absent from the proposed algorithm, and its merits as a generic drop-in replacement for backpropagation are not strong enough or at least not yet well argued for.\n\nThe authors propose a new prescription to compute targets in TP. Unlike much of the recent work on TP, this prescription leads to a parameter update that requires weight transport. This limits its interest and narrows the scope of the contributions, as optimization without weight transport is of great interest not just in theoretical neuroscience but also for neuromorphic hardware design and large-scale distributed network implementations. \n\nAs noted by the authors, the proposed algorithm is also strictly costlier than backpropagation, although the relative difference in cost becomes smaller as sequence length grows. It also does not allow for decoupled updates, or at least this is not commented in the paper. Its applicability to decentralized optimization in distributed settings (another setting of potential interest for TP algorithms) is therfore not clear.\n\nFor the proposed algorithm to be interesting, it therefore has to produce descent directions that lead to faster optimization than gradients. The paper does not provide any theoretical guarantees that this is the case. The merits of the new algorithm must then be judged by the experimental results. This is a very difficult question to answer experimentally, and I did not find the results extremely convincing, but I am admittedly not an expert in sequence learning problems.\n\nI have a number of questions for the authors:\n- In page 3, below eq. 2, it is mentioned that DTP is a 0-th order approximation. Can the authors clarify this remark? In my understanding, the difference correction of DTP stems from a first-order expansion of $f$, see Appendix B of Lee et al. (2015). In what sense is it a 0-th order correction?\n\n- What are the new insights provided in section 2.2 compared to what is already used by Lee et al. to arrive at DTP (and perhaps more clearly restated in Lemma 1 of Meulemans et al. 2020)?\n\n-  I'm having trouble with eq. 4 and the text that follows. What happens when there are multiple minimizers of the objective presented in (4), for example when $a$ is the ReLU activation? How should the next unnumbered equation appearing below \"Formally, for [...] the ReLU, their inverse can be obtained analytically\" be understood in this case?\n\n- There are numerous remarks on the vanishing/exploding gradient problem and allusions to the promise of solving this problem with TP. But since the proposed parameter update relies on products of Jacobian inverses, isn't it also subject to vanishing and exploding gradient problems?\n\n- Why is the relationship between TP and Gauss-Newton optimization pointed out by Bengio (2020) and Meulemans et al. (2020) called into question due to the difficulties observed when setting the regularizer to zero? \n\n- Is GEMINI (LeCun et al., 1989), cited already in the abstract, the desired reference? To me, GEMINI is more obviously related to synthetic gradients than target propagation, and it is not usually cited as LeCun's original TP paper. Reference [1] below is perhaps more appropriate.\n\n[1] LeCun (1986) \"Learning processes in an asymmetric threshold network\"", "summary_of_the_review": "The algorithm loses most of the features that make target propagation attractive as an alternative to backpropagation. As an optimization algorithm, its guarantees are not sufficiently strong.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635870561562}, {"id": "48Kkr44yY1O", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4414/Reviewer_Q1ay"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors provide a modified version of target propagation that is easily implementable via autograd system. This paper also compares the computation complexity of the algorithm with backpropagation and shows empirically that the algorithm yields good performance on recurrent neural networks with long sequences.", "review_text": "**Strengths:**\n- The proposed scheme features approximating the inverse of the layer by a regularized inverse that facilitates the use of autograd.\n- Linearized inversion offers a natural interpretation as an approximation of difference target propagation.\n- The cost of matrix inversion can be amortized by the length of sequence.\n- There is theoretical guarantee available that bounds the difference between target propagation and backpropagation.\n\n**Weaknesses:**\n- The process seems to have large variation on training loss and accuracy compared to back propagation, which might be an indication that the scheme is less amendable to theoretical analysis.\n\n**Correctness:**\n- There is no false claims to the best of my knowledge\n\n**Clarity:**\n- The paper is clear and structured\n- A typo in the paragraph under equation (2): \"allow us tp interpret\" ---> \"allow us to interpret\"\n\n**Additional Comments:**\n- Is there any convergence result possible for the proposed target propagation scheme given that the backward updates are more tractable compared to the original version? For example, there is a convergence analysis for feedback alignment by Song et al.. Anything similar possible for target propagation?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors provide a modified version of target propagation that is easily implementable via autograd system. This paper also compares the computation complexity of the algorithm with backpropagation and shows empirically that the algorithm yields good performance on recurrent neural networks with long sequences.", "main_review": "**Strengths:**\n- The proposed scheme features approximating the inverse of the layer by a regularized inverse that facilitates the use of autograd.\n- Linearized inversion offers a natural interpretation as an approximation of difference target propagation.\n- The cost of matrix inversion can be amortized by the length of sequence.\n- There is theoretical guarantee available that bounds the difference between target propagation and backpropagation.\n\n**Weaknesses:**\n- The process seems to have large variation on training loss and accuracy compared to back propagation, which might be an indication that the scheme is less amendable to theoretical analysis.\n\n**Correctness:**\n- There is no false claims to the best of my knowledge\n\n**Clarity:**\n- The paper is clear and structured\n- A typo in the paragraph under equation (2): \"allow us tp interpret\" ---> \"allow us to interpret\"\n\n**Additional Comments:**\n- Is there any convergence result possible for the proposed target propagation scheme given that the backward updates are more tractable compared to the original version? For example, there is a convergence analysis for feedback alignment by Song et al.. Anything similar possible for target propagation?", "summary_of_the_review": "The target propagation scheme proposed in this paper look novel to me. Even though the numerical result is a bit confusing, the algorithm has natural interpretation and more computationally tractable.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635816425328}, {"id": "bEvaAbboOG2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4414/Reviewer_Dczs"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes the train the target propagation algorithm (as an alternative to gradient back-propagation algorithm) through regularized inversion. Specifically, they discussed two techniques including regularized inversion and linearize propagation, and demonstrate their effectiveness (via convergence speed and accuracy) for learning the RNN model on synthetic and image benchmarks.", "review_text": "The paper presents their model quite clear. Demonstration and comparison with traditional BP algorithm is illustrative. The target propagation algorithm is an interesting new directions that worth exploring. \n\nThe main downside of the paper would be experimental demonstration. \n\n-- It's good to have experiments on realistic image dataset such as cifar10, however, either the baseline as well as proposed model seems to be quite weak (their accuracy is quite low, fig 4, less than 20%). I understand the authors focused on the illustrative purpose, but that also makes it unclear how the model could generalize,  scale and apply to the realistic datasets. Since the current literature has quite a few competitive RNN models on image classification, why not directly take their models and replace the BP component to see how it works? I'm not asking for the SOTA performance, only the competitive results could be good enough to show the potential of the proposed work. \n\n-- I assume the framework is targeted on recurrent model training, perhaps the authors would make it more clear throughout the paper. Some of the places claim such as \" We have used target propagation within a stochastic gradient outer loop to train neural networks for a fair compar- ison to stochastic gradient using gradient backpropagation.\" (in the conclusion section), which might be misleading or over claimed. If in general the proposed model is flexible enough to handle all kinds of neural networks, perhaps experiments should also include convolution neural net etc for the illustration. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes the train the target propagation algorithm (as an alternative to gradient back-propagation algorithm) through regularized inversion. Specifically, they discussed two techniques including regularized inversion and linearize propagation, and demonstrate their effectiveness (via convergence speed and accuracy) for learning the RNN model on synthetic and image benchmarks.", "main_review": "The paper presents their model quite clear. Demonstration and comparison with traditional BP algorithm is illustrative. The target propagation algorithm is an interesting new directions that worth exploring. \n\nThe main downside of the paper would be experimental demonstration. \n\n-- It's good to have experiments on realistic image dataset such as cifar10, however, either the baseline as well as proposed model seems to be quite weak (their accuracy is quite low, fig 4, less than 20%). I understand the authors focused on the illustrative purpose, but that also makes it unclear how the model could generalize,  scale and apply to the realistic datasets. Since the current literature has quite a few competitive RNN models on image classification, why not directly take their models and replace the BP component to see how it works? I'm not asking for the SOTA performance, only the competitive results could be good enough to show the potential of the proposed work. \n\n-- I assume the framework is targeted on recurrent model training, perhaps the authors would make it more clear throughout the paper. Some of the places claim such as \" We have used target propagation within a stochastic gradient outer loop to train neural networks for a fair compar- ison to stochastic gradient using gradient backpropagation.\" (in the conclusion section), which might be misleading or over claimed. If in general the proposed model is flexible enough to handle all kinds of neural networks, perhaps experiments should also include convolution neural net etc for the illustration. ", "summary_of_the_review": "The paper explore an interesting alternative to BP. However, the current form didn't fully convince me its superiority over bp. I lean towards the rejection, but willing to adjust my rating based on the authors' feedbacks. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635805078801}, {"id": "QLb9l6RPlsW", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4414/Reviewer_239T"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This manuscript proposes a new way of approaching optimization in RNNs based on Target Propagation.  It develops a linearized approximation to the method to make it straightforward to implement within a standard computational graph, and develops some theory about the approximations, and shows improvement compared to gradient descent in a few experiments.", "review_text": "The primary strengths of this manuscript are that it is easy to read, the proposed algorithm is straightforward to implement in the current computational setup, and the improvements versus gradient descent.\n\nIn my opinion, there are two main weaknesses of the manuscript.\n\n(1) The theory needs to be more fully explored and discussed.  For example, Lemma 3.1 bounds the different between the direction computed by TP and BP. First, it should be noted that this bound is quite loose in the case where the constants $a$ or $b$ are greater than 1. Second, there needs to be more complete discussion of this meaning of this Lemma, as the argument that TP is better than BP depends on the fact that TP calculates a \\textit{different} direction than BP.  I would appreciate greater follow-up and discussion in the analysis on how TP and BP differ rather than on how similar they are. Next, I find the relationship to Gauss-Newton methods quite interesting, but that section is barely sketched out and does not seem fully derived.  How close is TP to a Gauss-Newton method?  Also, I found the phrase \"In our experiments, we have not been able to get a non-diverging sequence of iterates if the regularization was set to 0, which questions the interpretation of target propagation as a Gauss-Newton method\" peculiar.  In a Gauss-Newton method, you need more samples than parameters to get the Jacobian to be well-conditioned and calculate a good direction, or it has to be regularized somehow.  The fact that you have to regularize the TP technique seems consistent with a Gauss-Newton method.  Am I missing something?  I believe that fully-fleshing out this relationship, if indeed they are close, would help provide rationale for your method.\n\n(2) The empirical evidence, in my opinion, is lacking.  There are few experiments and runs here.  When does TP outperform GD?  Why does it outperform GP?  This manuscript needs to explain more about what is different in the optimization to make it work better, and evaluate how  robust those improvements are with much more detailed and extensive experimentation, especially since the improvements are not fully explained from the theoretical side.  Additionally, there are many, many, learning algorithms designed to improve learning in RNNs.  There needs to be more comparison and discussion of these algorithms.  For example, how does this compare to the algorithm proposed in Lee et al 2015?  How does this relate to Hessian Free-Optimization (e.g., [1]), which is somewhat alluded to with the relationship to Gauss-Newton?  There must be a more complete literature search and discussion on optimizing RNNs as a whole, rather than the relatively narrow focus on exclusively TP.\n\n[1] Martens, James, and Ilya Sutskever. \"Learning recurrent neural networks with hessian-free optimization.\" ICML. 2011.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This manuscript proposes a new way of approaching optimization in RNNs based on Target Propagation.  It develops a linearized approximation to the method to make it straightforward to implement within a standard computational graph, and develops some theory about the approximations, and shows improvement compared to gradient descent in a few experiments.", "main_review": "The primary strengths of this manuscript are that it is easy to read, the proposed algorithm is straightforward to implement in the current computational setup, and the improvements versus gradient descent.\n\nIn my opinion, there are two main weaknesses of the manuscript.\n\n(1) The theory needs to be more fully explored and discussed.  For example, Lemma 3.1 bounds the different between the direction computed by TP and BP. First, it should be noted that this bound is quite loose in the case where the constants $a$ or $b$ are greater than 1. Second, there needs to be more complete discussion of this meaning of this Lemma, as the argument that TP is better than BP depends on the fact that TP calculates a \\textit{different} direction than BP.  I would appreciate greater follow-up and discussion in the analysis on how TP and BP differ rather than on how similar they are. Next, I find the relationship to Gauss-Newton methods quite interesting, but that section is barely sketched out and does not seem fully derived.  How close is TP to a Gauss-Newton method?  Also, I found the phrase \"In our experiments, we have not been able to get a non-diverging sequence of iterates if the regularization was set to 0, which questions the interpretation of target propagation as a Gauss-Newton method\" peculiar.  In a Gauss-Newton method, you need more samples than parameters to get the Jacobian to be well-conditioned and calculate a good direction, or it has to be regularized somehow.  The fact that you have to regularize the TP technique seems consistent with a Gauss-Newton method.  Am I missing something?  I believe that fully-fleshing out this relationship, if indeed they are close, would help provide rationale for your method.\n\n(2) The empirical evidence, in my opinion, is lacking.  There are few experiments and runs here.  When does TP outperform GD?  Why does it outperform GP?  This manuscript needs to explain more about what is different in the optimization to make it work better, and evaluate how  robust those improvements are with much more detailed and extensive experimentation, especially since the improvements are not fully explained from the theoretical side.  Additionally, there are many, many, learning algorithms designed to improve learning in RNNs.  There needs to be more comparison and discussion of these algorithms.  For example, how does this compare to the algorithm proposed in Lee et al 2015?  How does this relate to Hessian Free-Optimization (e.g., [1]), which is somewhat alluded to with the relationship to Gauss-Newton?  There must be a more complete literature search and discussion on optimizing RNNs as a whole, rather than the relatively narrow focus on exclusively TP.\n\n[1] Martens, James, and Ilya Sutskever. \"Learning recurrent neural networks with hessian-free optimization.\" ICML. 2011.", "summary_of_the_review": "This manuscript proposes an interesting and potentially useful algorithm, but needs to more fully explain its theoretical properties and provide more detailed experimentation and comparisons.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635776359959}], "openreview_url": "https://openreview.net/forum?id=MTsBazXmX00", "arxiv_id": "2112.01453", "paper_pdf": "papers/MTsBazXmX00.pdf", "paper_pdf_sha256": "9dd652243c6f080ea6441f3048c1715b9ec1c3585052fdd33b062310a17ffbfa", "paper_pdf_bytes": 1097086, "paper_pdf_source": "openreview", "code_url": "https://github.com/vroulet/tpri", "code_repository": "vroulet/tpri", "code_commit": "c7e7c10e3382773b634c183a977f125147aa9419", "code_archive": "repos/MTsBazXmX00.zip", "code_archive_sha256": "cb33997af6569f6e38029d1e5ed6c9e952325379fda7b7f346268173bbfe726b", "code_archive_bytes": 56869, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 109, "github_languages": {"Python": 133096}, "github_archived": false, "github_pushed_at": "2023-01-09T22:06:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/target-propagation-via-regularized-inversion-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Rld-9OxQ6HU", "year": 2021, "status": "rejected", "title": "MC-LSTM: Mass-conserving LSTM", "authors": ["Pieter-Jan Hoedt", "Frederik Kratzert", "Daniel Klotz", "Christina Halmich", "Markus Holzleitner", "Grey Nearing", "Sepp Hochreiter", "Günter Klambauer"], "authorids": ["~Pieter-Jan_Hoedt1", "~Frederik_Kratzert1", "~Daniel_Klotz1", "halmich@ml.jku.at", "~Markus_Holzleitner1", "gsnearing@google.com", "~Sepp_Hochreiter1", "~Günter_Klambauer1"], "authors_source": "OpenReview API", "abstract": "The success of Convolutional Neural Networks (CNNs) in computer vision is mainly driven by their strong inductive bias,  which is strong enough to allow CNNs to solve vision-related tasks with random weights, meaning without learning. Similarly, Long Short-Term Memory (LSTM) has a strong inductive bias towards storing information over time. However, many real-world systems are governed by conservation laws, which lead to the redistribution of particular quantities —e.g. in physical and economical systems.   Our novel Mass-Conserving LSTM (MC-LSTM) adheres to these conservation laws by extending the inductive bias of LSTM to model the redistribution of those stored quantities. MC-LSTMs set a new state-of-the-art for neural arithmetic units at learning arithmetic operations, such as addition tasks, which have a strong conservation law, as the sum is constant overtime. Further, MC-LSTM is applied to traffic forecasting, modeling a pendulum, and a large benchmark dataset in hydrology, where it sets a new state-of-the-art for predicting peak flows. In the hydrology example, we show that MC-LSTM states correlate with real world processes and are therefore interpretable.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Qab9cJgkvZa", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3049/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper the authors propose a novel architecture, called Mass-Conserving LSTM (MC-LSTM) based on LSTM. The authors base their work over the hypothesis that the real world is based over conservation laws related to mass, energy, etc. Thus, they propose that also the quantities involved in deep learning models should be conserved. To do so, they aim at exploiting the memory cells of the LSTM as mass accumulators and then force the conservation laws via the model equations. The authors finally show successfully the potential of this novel network into three experimental settings where several types of “conservation” are required (e.g. mass conservation, energy conservation, etc).\n\nPros:\n+ they deal with a problem which can arise in non-laboratory scenarios in a novel way. Moreover, their results show that their method is great at dealing with situations where conservation is required, which may be particularly important for real-world scenarios\n+ the paper explains in-depth all the decisions made, and it is well written. Moreover, I found really interesting how they deal with the related work and special cases. It shows a really in-depth understanding of up-to-date literature in the field. \n\nCons:\n- (minor) the authors focus their experimental section to settings where mass (or energy, etc) conservation is required. It would be interesting to see how it performs also in settings where it is not required as well, thus showing whether this method also generalizes to different settings. \n\n------\n\nAuthors' response addressed properly my request.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The experimental evaluation shows that the proposed method is great at dealing with situations where conservation is required, which may be particularly important for real-world scenarios", "review": "In this paper the authors propose a novel architecture, called Mass-Conserving LSTM (MC-LSTM) based on LSTM. The authors base their work over the hypothesis that the real world is based over conservation laws related to mass, energy, etc. Thus, they propose that also the quantities involved in deep learning models should be conserved. To do so, they aim at exploiting the memory cells of the LSTM as mass accumulators and then force the conservation laws via the model equations. The authors finally show successfully the potential of this novel network into three experimental settings where several types of “conservation” are required (e.g. mass conservation, energy conservation, etc).\n\nPros:\n+ they deal with a problem which can arise in non-laboratory scenarios in a novel way. Moreover, their results show that their method is great at dealing with situations where conservation is required, which may be particularly important for real-world scenarios\n+ the paper explains in-depth all the decisions made, and it is well written. Moreover, I found really interesting how they deal with the related work and special cases. It shows a really in-depth understanding of up-to-date literature in the field. \n\nCons:\n- (minor) the authors focus their experimental section to settings where mass (or energy, etc) conservation is required. It would be interesting to see how it performs also in settings where it is not required as well, thus showing whether this method also generalizes to different settings. \n\n------\n\nAuthors' response addressed properly my request.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603925196861}, {"id": "Nb_DnMAQGgu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3049/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper provides an interesting and novel LSTM structure named MC-LSTM, which extends the inductive bias of LSTM to deal with some real-world problems limited by conservation laws. The authors do some experiments related to traffic forecasting and hydrology to illustrate the effectiveness of MC-LSTM.  The new architecture is well-suited for predicting some physical systems, which is valuable.\n##########################################################################\n\nReasons for score: \n\nOverall, I vote for accepting. \nI deem that the novel architecture based on LSTM that conserves quantites is useful and interesting.\nMy major concern is about some explanation about definitions and some additional ablation models (see cons below). \nHopefully the authors can address my concern in the rebuttal period. \n\n##########################################################################Pros: \n\n1. The paper takes one of the most important issue of some real-world systems:  conservation laws, which is important and\nshould be expressed through LSTM units.\n\n2.  This paper provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed LSTM. \nThe entire structure is organized well and the formulas are very detailed.\n\n##########################################################################\n\nCons: \n\n1. In Basic gating, the formula about input, why do you use softmax operator? Because in basic LSTM, there is sigmoid.\n It would be better to provide more details about it.\n\n2. What are the advantages of MC-LSTM in terms of speed and resource consumption?\nIt would be more convincing if the authors can provide more cases in the rebuttal period. \n\n##########################################################################\n\nQuestions during rebuttal period: \n\nPlease address and clarify the cons above \n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Recommendation to Accept", "review": "The paper provides an interesting and novel LSTM structure named MC-LSTM, which extends the inductive bias of LSTM to deal with some real-world problems limited by conservation laws. The authors do some experiments related to traffic forecasting and hydrology to illustrate the effectiveness of MC-LSTM.  The new architecture is well-suited for predicting some physical systems, which is valuable.\n##########################################################################\n\nReasons for score: \n\nOverall, I vote for accepting. \nI deem that the novel architecture based on LSTM that conserves quantites is useful and interesting.\nMy major concern is about some explanation about definitions and some additional ablation models (see cons below). \nHopefully the authors can address my concern in the rebuttal period. \n\n##########################################################################Pros: \n\n1. The paper takes one of the most important issue of some real-world systems:  conservation laws, which is important and\nshould be expressed through LSTM units.\n\n2.  This paper provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed LSTM. \nThe entire structure is organized well and the formulas are very detailed.\n\n##########################################################################\n\nCons: \n\n1. In Basic gating, the formula about input, why do you use softmax operator? Because in basic LSTM, there is sigmoid.\n It would be better to provide more details about it.\n\n2. What are the advantages of MC-LSTM in terms of speed and resource consumption?\nIt would be more convincing if the authors can provide more cases in the rebuttal period. \n\n##########################################################################\n\nQuestions during rebuttal period: \n\nPlease address and clarify the cons above \n ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603700173393}, {"id": "He8rYglDaY7", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3049/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "<Summary>\n1. Many real-world systems follow conservation laws.\n2. This paper incorporates conservation laws into the RNN as an inductive bias.\n3. This paper proposes MC-LSTM and shows that MC-LSTM follows conservation laws empirically and theoretically.\n4. MC-LSTM utilizes a positive left-stochastic matrix to redistribute mass.\n5. This paper validates the MC-LSTM on arithmetic tasks, traffic forecasting, pendulum, rainfall tasks.\n\n<Strengths>\n1. Incorporating mass conservation laws into the neural network is important.\n2. This paper proposes MC-LSTM, a simple and effective mass-conserving model.\n3. The motivation of the research and the proposed methods are straightforward.\n4. Related works section provides a comparison between MC-LSTM and others such as the Markov chain, and it is also very interesting.\n\n<Weaknesses>\n1. This paper shows extensive experimental results, but there are less qualitative results. \n2. This paper missing some important related works, such as a physics-guided recurrent neural network model (PGRNN)\n\n<Questions and Additional Feedback>\n1. Is there empirical (or theoretical) analysis for long-term gradient vanishing?\n2. I wonder how MC-LSTM actually works. Is there qualitative analysis for a, i, o, and R?\n3. What is the reason that Eq. (5) and Eq. (6) utilize L1 norm?\n4. Does MC-LSTM can be extended to represent (2*mass) or (mass^2)-conserving properties?\n5. MC-Transformer will be one of a good extension.\n\n<Missing Reference>\n1. Lagrangian Neural Networks\n2. Physics-Guided Machine Learning for Scientific Discovery: An Application in Simulating Lake Temperature Profiles\n3. Discovering physical concepts with neural networks\n\n<Typos>\n1. One of the greatest success stories of deep learning are => One of the greatest success stories of deep learning is\n2. The former represent => The former represents\n\n<After Rebuttal>\nThank you for your detailed response.\nI will keep my positive score because my concerns are resolved partially.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "MC-LSTM is an interesting work.", "review": "<Summary>\n1. Many real-world systems follow conservation laws.\n2. This paper incorporates conservation laws into the RNN as an inductive bias.\n3. This paper proposes MC-LSTM and shows that MC-LSTM follows conservation laws empirically and theoretically.\n4. MC-LSTM utilizes a positive left-stochastic matrix to redistribute mass.\n5. This paper validates the MC-LSTM on arithmetic tasks, traffic forecasting, pendulum, rainfall tasks.\n\n<Strengths>\n1. Incorporating mass conservation laws into the neural network is important.\n2. This paper proposes MC-LSTM, a simple and effective mass-conserving model.\n3. The motivation of the research and the proposed methods are straightforward.\n4. Related works section provides a comparison between MC-LSTM and others such as the Markov chain, and it is also very interesting.\n\n<Weaknesses>\n1. This paper shows extensive experimental results, but there are less qualitative results. \n2. This paper missing some important related works, such as a physics-guided recurrent neural network model (PGRNN)\n\n<Questions and Additional Feedback>\n1. Is there empirical (or theoretical) analysis for long-term gradient vanishing?\n2. I wonder how MC-LSTM actually works. Is there qualitative analysis for a, i, o, and R?\n3. What is the reason that Eq. (5) and Eq. (6) utilize L1 norm?\n4. Does MC-LSTM can be extended to represent (2*mass) or (mass^2)-conserving properties?\n5. MC-Transformer will be one of a good extension.\n\n<Missing Reference>\n1. Lagrangian Neural Networks\n2. Physics-Guided Machine Learning for Scientific Discovery: An Application in Simulating Lake Temperature Profiles\n3. Discovering physical concepts with neural networks\n\n<Typos>\n1. One of the greatest success stories of deep learning are => One of the greatest success stories of deep learning is\n2. The former represent => The former represents\n\n<After Rebuttal>\nThank you for your detailed response.\nI will keep my positive score because my concerns are resolved partially.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603604211906}, {"id": "s18NnymVhw", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3049/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Summary\n\nThe authors proposed a new family of LSTMs (i.e. MC-LSTMS) which can be shown to have a mass conservation property for the LSTM memory cells. They have shown in various applications that these models lead to on par or better performance compared to state of art approaches. However, after examining the paper, I am not yet fully convinced that 1) a general unified MC-LSTMs can achieve good performance in all the mentioned cases 2) the conserved mass is interpretable (and corresponds to the problem’s invariant, conserved quantity). I have listed those concerns in the question sections as authors might have precise answers. \n\n#### Pros\n\nThe authors in this paper have proposed a new family of LSTMs (i.e. MC-LSTMs) which can be shown to have a conservation property for the mass held by the LSTM memory cells. The authors have applied their models in three different prediction tasks involving quantity conservation, namingly summation(and subtraction), traffic prediction and rainfall runoff modelling. The proposed models (with some variants) achieve better or at least on par with current state of art models.\n\nThese different applications illustrate the usefulness of the models; the authors also reference appropriately to current state of art approaches, helping reviewers to well situate the paper’s contribution.\n\n#### Questions:\n\nWell I understand there are differences for each application (motivating to choose different neural architectures), it would be very interesting to know the performance of one quite general architecture (compared to say standard LSTM implemented in torch). In the paper, we have seen in the experiments:\n- time - independent R^t used in arithmetics\n- time - dependent R^t used in hydrology experiments\n- Hypernetwork based R^t used in pendulum experiments\nAnd also they vary on the auxiliary/mass inputs choices. \nWhat would be the performance if the authors use a quite general architecture (for example the second setting) for all the experiments please?\n\nJust for confirmation, in the experiments, the LSTMs involve forgetting gate right (contrary to what described in eq (1))\n\nI don’t find explanations for the r value for the hydrology experiments. Meanwhile the r value for the traditional r value is different from what is chosen for the MC-LSTM. Why is this please?\n\n#### Minor issues:\n\nThere seems some formatting issues at the end of “Mechanisms beyond storing are required for real-world applications” in Introduction. \n\nIn related work, Hamiltonian approaches (Greydanus et al. 2019) don’t seem to assume access knowledge to time derivatives w.r.t. Inputs. Rather, it favours the conservation parametrised by the Hamiltonian.  \n\n#### Minor suggestions:\n\nIn abstract, “expressed through continuity equations”, these aspects don’t seem to be addressed by MC-LSTMs, so I propose to not include this phrase in the abstract. \n\nI suggest adding some citations In the introduction where the authors said “LSTM to excel at speech, text, and language tasks” http://nlpprogress.com/. \n\nThe authors mention in the abstract and introduction about the interpretability but only show the mass interpretation in the hydrology experiments. For which I would suggest the authors show it in at least one other experiments (e.g. traffic) to 1) be more convincing 2) highlight this property (which may reveal to be very beneficial for production systems)\n\nI would be good to mention forgetting gate around equation (1) as those are used for later experiments in the paper. \n\n#### Thank you for the authors to having answered my questions and having addressed all my comments. My main concern was about the generalisation of the proposed architecture and the results at the current revision are convincing to me. I believe this direction of the research together with the approach taken make a nice contribution to the conference. In consequence, I have raised my score from 5 to 7.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice idea, extensive experiments with clear references; having concerns about the architecture generalisation.", "review": "#### Summary\n\nThe authors proposed a new family of LSTMs (i.e. MC-LSTMS) which can be shown to have a mass conservation property for the LSTM memory cells. They have shown in various applications that these models lead to on par or better performance compared to state of art approaches. However, after examining the paper, I am not yet fully convinced that 1) a general unified MC-LSTMs can achieve good performance in all the mentioned cases 2) the conserved mass is interpretable (and corresponds to the problem’s invariant, conserved quantity). I have listed those concerns in the question sections as authors might have precise answers. \n\n#### Pros\n\nThe authors in this paper have proposed a new family of LSTMs (i.e. MC-LSTMs) which can be shown to have a conservation property for the mass held by the LSTM memory cells. The authors have applied their models in three different prediction tasks involving quantity conservation, namingly summation(and subtraction), traffic prediction and rainfall runoff modelling. The proposed models (with some variants) achieve better or at least on par with current state of art models.\n\nThese different applications illustrate the usefulness of the models; the authors also reference appropriately to current state of art approaches, helping reviewers to well situate the paper’s contribution.\n\n#### Questions:\n\nWell I understand there are differences for each application (motivating to choose different neural architectures), it would be very interesting to know the performance of one quite general architecture (compared to say standard LSTM implemented in torch). In the paper, we have seen in the experiments:\n- time - independent R^t used in arithmetics\n- time - dependent R^t used in hydrology experiments\n- Hypernetwork based R^t used in pendulum experiments\nAnd also they vary on the auxiliary/mass inputs choices. \nWhat would be the performance if the authors use a quite general architecture (for example the second setting) for all the experiments please?\n\nJust for confirmation, in the experiments, the LSTMs involve forgetting gate right (contrary to what described in eq (1))\n\nI don’t find explanations for the r value for the hydrology experiments. Meanwhile the r value for the traditional r value is different from what is chosen for the MC-LSTM. Why is this please?\n\n#### Minor issues:\n\nThere seems some formatting issues at the end of “Mechanisms beyond storing are required for real-world applications” in Introduction. \n\nIn related work, Hamiltonian approaches (Greydanus et al. 2019) don’t seem to assume access knowledge to time derivatives w.r.t. Inputs. Rather, it favours the conservation parametrised by the Hamiltonian.  \n\n#### Minor suggestions:\n\nIn abstract, “expressed through continuity equations”, these aspects don’t seem to be addressed by MC-LSTMs, so I propose to not include this phrase in the abstract. \n\nI suggest adding some citations In the introduction where the authors said “LSTM to excel at speech, text, and language tasks” http://nlpprogress.com/. \n\nThe authors mention in the abstract and introduction about the interpretability but only show the mass interpretation in the hydrology experiments. For which I would suggest the authors show it in at least one other experiments (e.g. traffic) to 1) be more convincing 2) highlight this property (which may reveal to be very beneficial for production systems)\n\nI would be good to mention forgetting gate around equation (1) as those are used for later experiments in the paper. \n\n#### Thank you for the authors to having answered my questions and having addressed all my comments. My main concern was about the generalisation of the proposed architecture and the results at the current revision are convincing to me. I believe this direction of the research together with the approach taken make a nice contribution to the conference. In consequence, I have raised my score from 5 to 7.  ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1602712096387}], "openreview_url": "https://openreview.net/forum?id=Rld-9OxQ6HU", "arxiv_id": "2101.05186", "paper_pdf": "papers/Rld-9OxQ6HU.pdf", "paper_pdf_sha256": "04c9f65f5c64ac5258fdf5b37e4adc380f86ad0780ac4430cf0203c658b0b46a", "paper_pdf_bytes": 2045609, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-jku/mc-lstm", "code_repository": "ml-jku/mc-lstm", "code_commit": "8bbaece3ecb4187a76c6318d4c6e40c1dcc71303", "code_archive": "repos/Rld-9OxQ6HU.zip", "code_archive_sha256": "6e1b4ff407b3d943a3080eafc41417841e47421767bfa02467af1aba5f588ea6", "code_archive_bytes": 100843, "code_file_count": 36, "code_extensions": {".py": 29, ".sh": 6, ".ipynb": 1}, "github_disk_usage_kb": 137, "github_languages": {"Python": 187499, "Jupyter Notebook": 42130, "Shell": 9385}, "github_archived": false, "github_pushed_at": "2021-07-20T14:47:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mc-lstm-mass-conserving-lstm-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ByxCrerKvS", "year": 2020, "status": "rejected", "title": "Set Functions for Time Series", "authors": ["Max Horn", "Michael Moor", "Christian Bock", "Bastian Rieck", "Karsten Borgwardt"], "authorids": ["max.horn@bsse.ethz.ch", "michael.moor@bsse.ethz.ch", "christian.bock@bsse.ethz.ch", "bastian.rieck@bsse.ethz.ch", "karsten.borgwardt@bsse.ethz.ch"], "authors_source": "OpenReview API", "abstract": "Despite the eminent successes of deep neural networks, many architectures are often hard to transfer to irregularly-sampled and asynchronous time series that occur in many real-world datasets, such as healthcare applications. This paper proposes a novel framework for classifying irregularly sampled time series with unaligned measurements, focusing on high scalability and data efficiency.\nOur method SeFT (Set Functions for Time Series) is based on recent advances in differentiable set function learning, extremely parallelizable, and scales well to very large datasets and online monitoring scenarios.\nWe extensively compare our method to competitors on multiple healthcare time series datasets and show that it performs competitively whilst significantly reducing runtime.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hygu7HYM9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2305/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of supervised classification of time-series data that are irregularly sampled and asynchronous, with a special focus on the healthcare applications in the experiments. Inspired by the recent progress on differentiable set function learning, the paper proposes an approach called Set Functions for Time Series (SEFT), which views the time series as sets, and use a parametrized sum-decomposing function f as the model for representing the probabilities of different classes, with the sets as the inputs. The problem then reduces to learning the finite dimensional parametrization of the function f under a given loss, which is a differentiable optimization problem that can be learned via standard optimization methods. Together with a positional embedding of the timestamps and an attention-based aggregation, the paper reports improved performance of the proposed approach on a few healthcare time series with asynchronous and irregularly sampled data. In particular, the runtime is largely shortened, while the final accuracy remains competitive to other methods compared in the paper. \n\nThe idea of SEFT is novel and the results are also showing its promise. In addition, the interpretability shown in section 4.3 is also attractive. However, there are several issues that limit the contribution and maturity of this paper. \n\nFirstly, the paper proposes to model time series as a set. But this loses the information of the order of the time series, which can be extremely important in those datasets with long history dependence. In such cases, I'm not convinced that the set modeling would work. The authors should double check the characteristics of the datasets that are used, and see if they lack long history dependence properties in intuition. If so, this should be mentioned clearly. The authors should also make a more fair comparison with other approaches (like those based on RNN) on datasets with strong history dependence, e.g., Memetracker datasets of web postings and limit-order books datasets. Otherwise, it would be not clear whether this set modeling is generally applicable for general time series data.\n\nSecondly, the authors missed a large amount of related literature for approaching asynchronous and irregularly sampled time series, namely (marked) point-process based approaches. See papers like [1, 2, 3], to name just a few. The authors should at least include some of the recent approaches in this direction for comparison before claiming the superiority of SEFT.\n\nThirdly, there are a few parts that are not very clear. 1) The discussion about complexity (order m and m\\log m) at the bottom of page 1 is weird -- what does this complexity refer to? Does it include the learning of the unknown parameters in the models (like training of the neural networks in this paper)? 2) The loss function in formula (5) is not specified later in the paper (at least hard to find). 3) The Table 1 should be explained in much more details. In particular, why don't we include SEFT-ATTN for H-MNIST? The comment after * is also not clear to me -- is it relevant to why SEFT-ATTN is not included? And what are MICRO/MACRO/WEIGHTED AUC? And why are we using different sets of performance criteria for the first two and last two datasets?\n\nFinally, some minor comments: 1) On page 2, \"the following methods\" should be \"the above methods\"; 2) on page 3, the meaning of \"channels\" should be specified clearer; 3) on page 4, in formulae (3) and (4), should there be \\pi or 2\\pi in the formula?\n\n[1] Mei, Hongyuan, and Jason M. Eisner. \"The neural hawkes process: A neurally self-modulating multivariate point process.\" Advances in Neural Information Processing Systems. 2017.\n[2] Xiao, Shuai, et al. \"Joint modeling of event sequence and time series with attentional twin recurrent neural networks.\" arXiv preprint arXiv:1703.08524 (2017).\n[3] Yang, Yingxiang, et al. \"Online learning for multivariate Hawkes processes.\" Advances in Neural Information Processing Systems. 2017.\n\n############## post rebuttal ###############\nAfter reading the authors' rebuttal, I decide to improve the rating to 5 (reflected as 6 due to the ICLR rating system limitation this year).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This paper considers the problem of supervised classification of time-series data that are irregularly sampled and asynchronous, with a special focus on the healthcare applications in the experiments. Inspired by the recent progress on differentiable set function learning, the paper proposes an approach called Set Functions for Time Series (SEFT), which views the time series as sets, and use a parametrized sum-decomposing function f as the model for representing the probabilities of different classes, with the sets as the inputs. The problem then reduces to learning the finite dimensional parametrization of the function f under a given loss, which is a differentiable optimization problem that can be learned via standard optimization methods. Together with a positional embedding of the timestamps and an attention-based aggregation, the paper reports improved performance of the proposed approach on a few healthcare time series with asynchronous and irregularly sampled data. In particular, the runtime is largely shortened, while the final accuracy remains competitive to other methods compared in the paper. \n\nThe idea of SEFT is novel and the results are also showing its promise. In addition, the interpretability shown in section 4.3 is also attractive. However, there are several issues that limit the contribution and maturity of this paper. \n\nFirstly, the paper proposes to model time series as a set. But this loses the information of the order of the time series, which can be extremely important in those datasets with long history dependence. In such cases, I'm not convinced that the set modeling would work. The authors should double check the characteristics of the datasets that are used, and see if they lack long history dependence properties in intuition. If so, this should be mentioned clearly. The authors should also make a more fair comparison with other approaches (like those based on RNN) on datasets with strong history dependence, e.g., Memetracker datasets of web postings and limit-order books datasets. Otherwise, it would be not clear whether this set modeling is generally applicable for general time series data.\n\nSecondly, the authors missed a large amount of related literature for approaching asynchronous and irregularly sampled time series, namely (marked) point-process based approaches. See papers like [1, 2, 3], to name just a few. The authors should at least include some of the recent approaches in this direction for comparison before claiming the superiority of SEFT.\n\nThirdly, there are a few parts that are not very clear. 1) The discussion about complexity (order m and m\\log m) at the bottom of page 1 is weird -- what does this complexity refer to? Does it include the learning of the unknown parameters in the models (like training of the neural networks in this paper)? 2) The loss function in formula (5) is not specified later in the paper (at least hard to find). 3) The Table 1 should be explained in much more details. In particular, why don't we include SEFT-ATTN for H-MNIST? The comment after * is also not clear to me -- is it relevant to why SEFT-ATTN is not included? And what are MICRO/MACRO/WEIGHTED AUC? And why are we using different sets of performance criteria for the first two and last two datasets?\n\nFinally, some minor comments: 1) On page 2, \"the following methods\" should be \"the above methods\"; 2) on page 3, the meaning of \"channels\" should be specified clearer; 3) on page 4, in formulae (3) and (4), should there be \\pi or 2\\pi in the formula?\n\n[1] Mei, Hongyuan, and Jason M. Eisner. \"The neural hawkes process: A neurally self-modulating multivariate point process.\" Advances in Neural Information Processing Systems. 2017.\n[2] Xiao, Shuai, et al. \"Joint modeling of event sequence and time series with attentional twin recurrent neural networks.\" arXiv preprint arXiv:1703.08524 (2017).\n[3] Yang, Yingxiang, et al. \"Online learning for multivariate Hawkes processes.\" Advances in Neural Information Processing Systems. 2017.\n\n############## post rebuttal ###############\nAfter reading the authors' rebuttal, I decide to improve the rating to 5 (reflected as 6 due to the ICLR rating system limitation this year).\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1572144416160}, {"id": "BJeBhVSAYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2305/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe work is focused on classification of irregularly sampled and unaligned multi-modal time series. Prior work has primarily focused on imputation methods, either end-to-end or otherwise. This paper approaches the problem as a set function mapping between the time-series tuples to the class label. The proposed method is uses a set encoding of a multi-modal time series input, followed by mode-specific encoding of the tuples which are then aggregated in multiple ways prior to classification. An attention mechanism is attached in order for the model to automatically weigh the relevance of tuples for the classification. The model is compared to imputation based baselines on clinical ICU time series classification tasks. The performance mostly appears comparable across baselines but the proposed method has much better run-times. \n\nThe paper is for the most part well written, and related work well characterized. The formulation is interesting and clinically relevant as well so the choice of data-sets makes some sense. I have a few concerns about the architecture formulation and lack of clarification and intuition in what appears to be the main contribution of the paper (Sec 3.2 and 3.3) which I will detail below:\n\na. In the evaluation, I really want to see a decoupling between the \"time encoding step\" and \"attention based aggregation\" on the performance to figure out to isolate different sources of performance improvements. That is can there be a SEFT without time encoding? If not, why not? I encourage more ablation like studies that look at different sources of performance gains and demonstrate them in experiments.\n\nb. The description of Sec 3.3. is really missing key motivation for the choices made around how the attention formulation is designed. For example why does the dot produce include the set elements? What if it doesn't? What is Q supposed to capture? \n\nc. Is a_{j,i} shared across instances? Then irrespective of the number of observations per instance, the $j^{th}$ tuple gets similar weights? If not appropriate indexing will help clarify this.\n\nd. It would be useful to provide how exactly a label is inferred for a *new* test instance. \n\nI have some minor additional feedback (just for presentation and motivation purposes):\n\n1. Authors make a claim in the introduction which should likely be qualified with a citation - \"Furthermore, even though a decoupled imputation scheme followed by classification is generally more scalable, it may lose information that is relevant for prediction tasks\". How does decoupled imputation imply loss of relevant information? By losing information about which observations are missing and relying on that for prediction? Does this clinically make sense? Or even generally? \n\n2. In Sec 3.3, you probably mean $W_i \\in R^{(im(f') + |s_j|) \\times d}$. That is parenthesis are missing?\n\n3. What are the +- std errors indicating? Is it cross validation error on a held-out test set? \n\n4. Initially $i$ is indexing samples and by equation (3), (4) $i$ indexes time(?) and in Sec 3.3 $i$ indexes observations? How are observations defined here? is it measurement of specific modality at a specific time instance? Can you clear this in the introduction itself? \n\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------------\nI have read the authors updated draft and response. The experiments section looks much better now. \n\n1. The overall contribution has less clinical utility in my opinion as generally a patient likely deteriorates over time before an adverse outcome and therefore -- to give the model too much flexibility w.r.t. time ordering doesn't make quite as much sense. This is reflected in the fact that experimental results are not drastically better than other baselines. The authors might be able to show the utility of the method on other time series classification datasets where this is not a limitation of the data itself. However in those settings, it may be a bit hard to beat transformers. Do the authors have a sense of where the benefits of this method really are?\n\n2. Mortality tasks are generally on the simpler side of clinical prediction problems as well. Nonetheless I think the contribution has some utility to the community. I do encourage the authors to try non--clinical datasets for a comparison\n\n3. Please have a discussion that includes limitations and to discuss where the benefits of your methods really lie. A clear and thoughtful discussion is currently missing in your conclusions. \n\nWith that said, I am updating my score to a 6.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "Summary:\nThe work is focused on classification of irregularly sampled and unaligned multi-modal time series. Prior work has primarily focused on imputation methods, either end-to-end or otherwise. This paper approaches the problem as a set function mapping between the time-series tuples to the class label. The proposed method is uses a set encoding of a multi-modal time series input, followed by mode-specific encoding of the tuples which are then aggregated in multiple ways prior to classification. An attention mechanism is attached in order for the model to automatically weigh the relevance of tuples for the classification. The model is compared to imputation based baselines on clinical ICU time series classification tasks. The performance mostly appears comparable across baselines but the proposed method has much better run-times. \n\nThe paper is for the most part well written, and related work well characterized. The formulation is interesting and clinically relevant as well so the choice of data-sets makes some sense. I have a few concerns about the architecture formulation and lack of clarification and intuition in what appears to be the main contribution of the paper (Sec 3.2 and 3.3) which I will detail below:\n\na. In the evaluation, I really want to see a decoupling between the \"time encoding step\" and \"attention based aggregation\" on the performance to figure out to isolate different sources of performance improvements. That is can there be a SEFT without time encoding? If not, why not? I encourage more ablation like studies that look at different sources of performance gains and demonstrate them in experiments.\n\nb. The description of Sec 3.3. is really missing key motivation for the choices made around how the attention formulation is designed. For example why does the dot produce include the set elements? What if it doesn't? What is Q supposed to capture? \n\nc. Is a_{j,i} shared across instances? Then irrespective of the number of observations per instance, the $j^{th}$ tuple gets similar weights? If not appropriate indexing will help clarify this.\n\nd. It would be useful to provide how exactly a label is inferred for a *new* test instance. \n\nI have some minor additional feedback (just for presentation and motivation purposes):\n\n1. Authors make a claim in the introduction which should likely be qualified with a citation - \"Furthermore, even though a decoupled imputation scheme followed by classification is generally more scalable, it may lose information that is relevant for prediction tasks\". How does decoupled imputation imply loss of relevant information? By losing information about which observations are missing and relying on that for prediction? Does this clinically make sense? Or even generally? \n\n2. In Sec 3.3, you probably mean $W_i \\in R^{(im(f') + |s_j|) \\times d}$. That is parenthesis are missing?\n\n3. What are the +- std errors indicating? Is it cross validation error on a held-out test set? \n\n4. Initially $i$ is indexing samples and by equation (3), (4) $i$ indexes time(?) and in Sec 3.3 $i$ indexes observations? How are observations defined here? is it measurement of specific modality at a specific time instance? Can you clear this in the introduction itself? \n\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------------\nI have read the authors updated draft and response. The experiments section looks much better now. \n\n1. The overall contribution has less clinical utility in my opinion as generally a patient likely deteriorates over time before an adverse outcome and therefore -- to give the model too much flexibility w.r.t. time ordering doesn't make quite as much sense. This is reflected in the fact that experimental results are not drastically better than other baselines. The authors might be able to show the utility of the method on other time series classification datasets where this is not a limitation of the data itself. However in those settings, it may be a bit hard to beat transformers. Do the authors have a sense of where the benefits of this method really are?\n\n2. Mortality tasks are generally on the simpler side of clinical prediction problems as well. Nonetheless I think the contribution has some utility to the community. I do encourage the authors to try non--clinical datasets for a comparison\n\n3. Please have a discussion that includes limitations and to discuss where the benefits of your methods really lie. A clear and thoughtful discussion is currently missing in your conclusions. \n\nWith that said, I am updating my score to a 6.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571865772852}, {"id": "H1l8MVt9FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2305/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper idea of this paper is straightforward and clear: treat the irregular time series as a bag of events, augment them with time information using positional encoding, and process the events in parallel. The idea is certainly faster than the sequential algorithms such as RNNs and their extensions. However, as shown in the experiments, because it does not encode the \"sequential-ness prior\" to the model, it is less accurate. Compared to RNNs, the proposed model has better access to the entire length of sequences and does not suffer from the limited memory issues of RNNs and variants.\n\nThe proposed idea in this paper can be considered a simplified version of the Transformers. Like transformers, the time and order are only provided to the model using the positional encoding and attention is central to aggregation over the sequence. Realizing the relationship with the Transformers not only decreases the novelty degree for this paper but also requires the authors to include the Transformers in the baselines.\n\nFinally, the results reported in the experiments are nice, especially for the baseline GRU-D! However, the MIMIC-III Mortality benchmark has a lot more than 21,000 stays to the best of my recollection. Can you please elaborate on how the number of data points has decreased?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper idea of this paper is straightforward and clear: treat the irregular time series as a bag of events, augment them with time information using positional encoding, and process the events in parallel. The idea is certainly faster than the sequential algorithms such as RNNs and their extensions. However, as shown in the experiments, because it does not encode the \"sequential-ness prior\" to the model, it is less accurate. Compared to RNNs, the proposed model has better access to the entire length of sequences and does not suffer from the limited memory issues of RNNs and variants.\n\nThe proposed idea in this paper can be considered a simplified version of the Transformers. Like transformers, the time and order are only provided to the model using the positional encoding and attention is central to aggregation over the sequence. Realizing the relationship with the Transformers not only decreases the novelty degree for this paper but also requires the authors to include the Transformers in the baselines.\n\nFinally, the results reported in the experiments are nice, especially for the baseline GRU-D! However, the MIMIC-III Mortality benchmark has a lot more than 21,000 stays to the best of my recollection. Can you please elaborate on how the number of data points has decreased?"}, "tcdate": 1571619853997}], "openreview_url": "https://openreview.net/forum?id=ByxCrerKvS", "arxiv_id": "1909.12064", "paper_pdf": "papers/ByxCrerKvS.pdf", "paper_pdf_sha256": "e9e77a4880f1eda442783a8068787ec52908d19e08990a45bb1282f6cdd09729", "paper_pdf_bytes": 308714, "paper_pdf_source": "openreview", "code_url": "https://github.com/BorgwardtLab/Set_Functions_for_Time_Series", "code_repository": "BorgwardtLab/Set_Functions_for_Time_Series", "code_commit": "6abd69c265b719f9703902af2c6a3254b6d2a779", "code_archive": "repos/ByxCrerKvS.zip", "code_archive_sha256": "21e0f85de43ece10e491631b0bd11250b7547df7824c615bb9516663b8b469f0", "code_archive_bytes": 87547, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 171, "github_languages": {"Python": 173486, "Makefile": 3709}, "github_archived": false, "github_pushed_at": "2021-02-25T09:32:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/set-functions-for-time-series-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1gFuiA9KX", "year": 2019, "status": "rejected", "title": "Skip-gram word embeddings in hyperbolic space", "authors": ["Matthias Leimeister", "Benjamin J. Wilson"], "authorids": ["matthias@lateral.io", "benjamin@lateral.io"], "authors_source": "OpenReview API", "abstract": "Embeddings of tree-like graphs in hyperbolic space were recently shown to surpass their Euclidean counterparts in performance by a large margin.\nInspired by these results, we present an algorithm for learning word embeddings in hyperbolic space from free text. An objective function based on the hyperbolic distance is derived and included in the skip-gram negative-sampling architecture from word2vec. The hyperbolic word embeddings are then evaluated on word similarity and analogy benchmarks. The results demonstrate the potential of hyperbolic word embeddings, particularly in low dimensions, though without clear superiority over their Euclidean counterparts. We further discuss subtleties in the formulation of the analogy task in curved spaces.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rJlysr_ka7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper371/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The present paper aims to apply recent developments in hyperbolic embeddings of graphs to the area of word embeddings.\n\nThe paper is relatively clearly written and looks technically correct. The main contribution of the paper is in suggesting the usage of Minkowski dot-product instead of Eucledian dot-product in skip-gram model and derivation of corresponding weight update formulas taking into account the peculiarities of hyperbolic space. The suggested aproach is realtively simple, though requiring adding additional bias parameter, which doesn't look entirely natural for the problem considered. I should note, that all the update equations are relatively straitforward given the results of recent papers on hyperbolic embeddings for graphs. Experimental results show some mild to no improvement over classical skip gram model in word similarity and word analogy problems.\n\nMy main concern about the paper is that it is not entirely clear throughout the text why the proposed model should be better than any of the baselines. Currently it lookss like the paper merging 2 ideas without any clear expalanation why they should work well together. I believe that the proposed approach (or similar one) might be useful for practice of natural language processing, but to asses that one would need to base on clear motivation and support this motivation with some examples showing that hyperbolicity indeed helps to capture semantics better (like famous world analogy examples for word2vec).\n\nPros:\n- clearly written\n- technically correct\nCons:\n- technically straightforward\n- not convincing experiments\n- unclear, why the approach should work", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "skip-gram + Minkowski distance -> unclear why it should work", "review": "The present paper aims to apply recent developments in hyperbolic embeddings of graphs to the area of word embeddings.\n\nThe paper is relatively clearly written and looks technically correct. The main contribution of the paper is in suggesting the usage of Minkowski dot-product instead of Eucledian dot-product in skip-gram model and derivation of corresponding weight update formulas taking into account the peculiarities of hyperbolic space. The suggested aproach is realtively simple, though requiring adding additional bias parameter, which doesn't look entirely natural for the problem considered. I should note, that all the update equations are relatively straitforward given the results of recent papers on hyperbolic embeddings for graphs. Experimental results show some mild to no improvement over classical skip gram model in word similarity and word analogy problems.\n\nMy main concern about the paper is that it is not entirely clear throughout the text why the proposed model should be better than any of the baselines. Currently it lookss like the paper merging 2 ideas without any clear expalanation why they should work well together. I believe that the proposed approach (or similar one) might be useful for practice of natural language processing, but to asses that one would need to base on clear motivation and support this motivation with some examples showing that hyperbolicity indeed helps to capture semantics better (like famous world analogy examples for word2vec).\n\nPros:\n- clearly written\n- technically correct\nCons:\n- technically straightforward\n- not convincing experiments\n- unclear, why the approach should work", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541535126641}, {"id": "HyxrmoZ5nX", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper371/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an algorithm that learns word embeddings in hyperbolic space. It adopts the Skip-Gram objective from Word2Vec on the hyperboloid model and derives the update equations for gradients accordingly. \nThe authors also propose to compute the word analogy by parallel transport along geodesics on the hyperboloid.\n\nStrength: The paper is well written, both the background geometry and the derived update method are clearly explained. The novelty and theoretical contribution are adequate. \n\nWeakness: My main concern is the lack of motivation for embedding words on the hyperboloid and the choice of evaluation metrics. For Poincare embeddings, the disc area and circle length grow exponentially with their radius, and the distances on the Poincare disk/ball reflect well the hierarchical structure of symbolic data, which make it natural to embed a graph in this space and lead to great evaluation results. The geometric property of the hyperboloid model, however, does not seem to in favor of encoding non-hierarchical semantics of words and the evaluation on word similarity/analogy tasks. The evaluation results in Table 1 and Table 2 show that the hyperbolic embeddings only performs better than the Euclidean embeddings in low dimensions but worse on higher dimensions (>50), while higher dimension embeddings generally encode more semantics and thus are used in downstream tasks. It will be great if the authors could elaborate on the advantages of learning word embeddings in hyperbolic space and evaluate accordingly. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good theoretical contribution but lack of motivation ", "review": "The paper proposes an algorithm that learns word embeddings in hyperbolic space. It adopts the Skip-Gram objective from Word2Vec on the hyperboloid model and derives the update equations for gradients accordingly. \nThe authors also propose to compute the word analogy by parallel transport along geodesics on the hyperboloid.\n\nStrength: The paper is well written, both the background geometry and the derived update method are clearly explained. The novelty and theoretical contribution are adequate. \n\nWeakness: My main concern is the lack of motivation for embedding words on the hyperboloid and the choice of evaluation metrics. For Poincare embeddings, the disc area and circle length grow exponentially with their radius, and the distances on the Poincare disk/ball reflect well the hierarchical structure of symbolic data, which make it natural to embed a graph in this space and lead to great evaluation results. The geometric property of the hyperboloid model, however, does not seem to in favor of encoding non-hierarchical semantics of words and the evaluation on word similarity/analogy tasks. The evaluation results in Table 1 and Table 2 show that the hyperbolic embeddings only performs better than the Euclidean embeddings in low dimensions but worse on higher dimensions (>50), while higher dimension embeddings generally encode more semantics and thus are used in downstream tasks. It will be great if the authors could elaborate on the advantages of learning word embeddings in hyperbolic space and evaluate accordingly. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541180189307}, {"id": "HJxfpQCOh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper371/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a technique for embedding words in hyperbolic space, which extends previous non-euclidean methods to non-structured data like free text. The authors provide a new gradient based method for creating the embeddings and then evaluate them on standard word embedding benchmarks. Overall the paper is very well written and well executed. They find that in the low dimensions the approach outperforms standard Euclidean space methods while in higher dimensions this advantage disappears.\n\nThe results do not try to claim state of the art on all benchmarks, which I find refreshing and I appreciate the authors candor in giving an honest presentation of their results. Overall, I enjoyed this paper and am eager to see how the authors develop the approach further. \n\nHowever, along these same lines it would be great to have the authors provide more discussion about the next steps and potential applications for this approach. Is the interest here purely methodological? Are there potential use cases where they believe this approach might be superior to Euclidean approaches? More detail in the discussion and intro about the trajectory of this work would help the reader understand the methodological and application-specific implications. \n\nPros:\n- Clearly written and results are presented in a straightforward manner. \n- Extension of analogy reasoning to non-euclidean spaces.\n\nCons:\n- Lack of clear motivation and compelling use case. \n- It would be nice to have a visualization of the approach in 2-dimensions. While Figure 3 is instructive for how analogies work in this space, it would be great to visualize an entire dataset. I'm sure that the proposed embeddings would result in  a very different space than euclidean embeddings (as the Poincare embedding paper showed), so it would be great to have at least one visualization of an embedded dataset. Presumably this would play to the strengths of the approach as it excels in lower dimensions. \n-  The largest of embedding dimension tested was 100, and it is common to use much larger embeddings of 500-d. Do the trends they observe continue to larger dimensions, e.g. is the performance gap even larger in higher dimensions? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "word2vec style embeddings in hyperbolic space", "review": "This paper presents a technique for embedding words in hyperbolic space, which extends previous non-euclidean methods to non-structured data like free text. The authors provide a new gradient based method for creating the embeddings and then evaluate them on standard word embedding benchmarks. Overall the paper is very well written and well executed. They find that in the low dimensions the approach outperforms standard Euclidean space methods while in higher dimensions this advantage disappears.\n\nThe results do not try to claim state of the art on all benchmarks, which I find refreshing and I appreciate the authors candor in giving an honest presentation of their results. Overall, I enjoyed this paper and am eager to see how the authors develop the approach further. \n\nHowever, along these same lines it would be great to have the authors provide more discussion about the next steps and potential applications for this approach. Is the interest here purely methodological? Are there potential use cases where they believe this approach might be superior to Euclidean approaches? More detail in the discussion and intro about the trajectory of this work would help the reader understand the methodological and application-specific implications. \n\nPros:\n- Clearly written and results are presented in a straightforward manner. \n- Extension of analogy reasoning to non-euclidean spaces.\n\nCons:\n- Lack of clear motivation and compelling use case. \n- It would be nice to have a visualization of the approach in 2-dimensions. While Figure 3 is instructive for how analogies work in this space, it would be great to visualize an entire dataset. I'm sure that the proposed embeddings would result in  a very different space than euclidean embeddings (as the Poincare embedding paper showed), so it would be great to have at least one visualization of an embedded dataset. Presumably this would play to the strengths of the approach as it excels in lower dimensions. \n-  The largest of embedding dimension tested was 100, and it is common to use much larger embeddings of 500-d. Do the trends they observe continue to larger dimensions, e.g. is the performance gap even larger in higher dimensions? ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541100474328}], "openreview_url": "https://openreview.net/forum?id=H1gFuiA9KX", "arxiv_id": "1809.01498", "paper_pdf": "papers/H1gFuiA9KX.pdf", "paper_pdf_sha256": "087364bb6cb2edb131a8f25fc2a36f5bd23302d2cc95142433f34a0116171e69", "paper_pdf_bytes": 351264, "paper_pdf_source": "openreview", "code_url": "https://github.com/lateral/minkowski", "code_repository": "lateral/minkowski", "code_commit": "7d312b750dc7104a1c7699336903f541956e3aaf", "code_archive": "repos/H1gFuiA9KX.zip", "code_archive_sha256": "39b2a907e1e630538b71abc7399f777cbce2f42cb3177194bf00ce17aac3d2d0", "code_archive_bytes": 1903752, "code_file_count": 151, "code_extensions": {".cc": 77, ".h": 41, ".py": 30, ".sh": 2, ".cmake": 1}, "github_disk_usage_kb": 1770, "github_languages": {"C++": 2169503, "Python": 282420, "CMake": 24793, "M4": 19093, "Shell": 15028, "Makefile": 14479, "C": 13067}, "github_archived": false, "github_pushed_at": "2018-09-28T09:33:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/skip-gram-word-embeddings-in-hyperbolic-space"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "REthH7hinP", "year": 2026, "status": "rejected", "title": "On Finetuning Tabular Foundation Models", "authors": ["Ivan Rubachev", "Akim Kotelnikov", "Nikolay Kartashev", "Artem Babenko"], "authorids": ["~Ivan_Rubachev1", "~Akim_Kotelnikov1", "~Nikolay_Kartashev1", "~Artem_Babenko1"], "authors_source": "OpenReview API", "abstract": "Foundation models are an emerging research direction in tabular deep learning. Notably, TabPFNv2 recently claimed superior performance over traditional GBDT-based methods on small-scale datasets using an in-context learning paradigm, which does not adapt model parameters to target datasets. However, the optimal finetuning approach for adapting tabular foundational models, and how this adaptation reshapes their internal mechanisms, remains underexplored. While prior works studied finetuning for earlier foundational models, inconsistent findings and TabPFNv2's unique architecture necessitate fresh investigation. To address these questions, we first systematically evaluate various finetuning strategies on diverse datasets. Our findings establish full finetuning as the most practical solution for TabPFNv2 in terms of time-efficiency and effectiveness. We then investigate how finetuning alters TabPFNv2's inner mechanisms, drawing an analogy to retrieval-augmented models. We reveal that the success of finetuning stems from the fact that after gradient-based adaptation, the dot products of the query-representations of test objects and the key-representations of in-context training objects more accurately reflect their target similarity. This improved similarity allows finetuned TabPFNv2 to better approximate target dependency by appropriately weighting relevant in-context samples, improving the retrieval-based prediction logic. From the practical perspective, we managed to finetune TabPFNv2 on datasets with up to 50K objects, observing performance improvements on almost all tasks. More precisely, on academic datasets with I.I.D. splits, finetuning allows TabPFNv2 to achieve state-of-the-art results, while on datasets with gradual temporal shifts and rich feature sets, TabPFNv2 is less stable and prior methods remain better.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "pXTGP4C8xm", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5248/Reviewer_eYTT"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "The paper explores fine tuning for tabular foundational models. Specifically, authors explore different fine tuning methods for TabPFNv2 and analyze the impact of fine tuning on attention.", "review_text": "The paper explores fine tuning for tabular foundational models. Specifically, authors explore different fine tuning methods for TabPFNv2 and analyze the impact of fine tuning on attention.", "strengths": "The paper is well written and explores an important aspect of foundation models. Analysis is anchored in real data and clearly shows that fine tuning is beneficial for model performance. Extensive empirical evaluation is conducted with diverse datasets and leading baselines.", "weaknesses": "I think the paper lacks novelty and provides fairly obvious conclusions that 1) fine tuning helps 2) full fine tuning is better than PEFT 3) fine tuning makes attention better (more peaked) at selecting the relevant information from the input context. All of these conclusion have been shown multiple times before on LLMs and other foundation models. So I think more is need to differentiate on novelty.", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores fine tuning for tabular foundational models. Specifically, authors explore different fine tuning methods for TabPFNv2 and analyze the impact of fine tuning on attention.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "The paper is well written and explores an important aspect of foundation models. Analysis is anchored in real data and clearly shows that fine tuning is beneficial for model performance. Extensive empirical evaluation is conducted with diverse datasets and leading baselines.", "weaknesses": "I think the paper lacks novelty and provides fairly obvious conclusions that 1) fine tuning helps 2) full fine tuning is better than PEFT 3) fine tuning makes attention better (more peaked) at selecting the relevant information from the input context. All of these conclusion have been shown multiple times before on LLMs and other foundation models. So I think more is need to differentiate on novelty.", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761876432891}, {"id": "lf8kZYJfSg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5248/Reviewer_WYkp"], "rating": 4, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 5, "summary": "This paper investigates different versions of fine-tuning of TabPFN v2. It finds simple fine-tuning being the strongest and most stable version of fine-tuning of TabPFN v2 comparing to other forms of fine-tuning strategies. In addition, this paper examines transformer attention map in details and finds the attention entropy decreases for a fine-tuned model comparing to a non-fine tuned version. Another contribution of this paper is the comparison between different models such as MLP-PLR, MNCA together with the fine-tuned and the non-fine-tuned versions of TabPFN v2.", "review_text": "This paper investigates different versions of fine-tuning of TabPFN v2. It finds simple fine-tuning being the strongest and most stable version of fine-tuning of TabPFN v2 comparing to other forms of fine-tuning strategies. In addition, this paper examines transformer attention map in details and finds the attention entropy decreases for a fine-tuned model comparing to a non-fine tuned version. Another contribution of this paper is the comparison between different models such as MLP-PLR, MNCA together with the fine-tuned and the non-fine-tuned versions of TabPFN v2.", "strengths": "### Originality and Significance\nThere has been many works examining fine-tuning of tabular foundation models. However, the previous works did not have strong conclusions. The author of this paper offers a slightly stronger conclusion, i.e. \"simple full fine-tuning is a strong and stable baseline for TabPFN v2 adaptation\" although it is still not very strong. The rigorous comparison with MNCA, MLP-PLR and other models are definitely also original. \n\nI believe the most original and significant contribution comes from the analyses of the attention map and the entropy of the attention scores. The paper finds there is a correlation between the decrease of entropy and the performance improvement. The paper makes analogy of the phenomenon with retrieval, a technique that has been proven very useful in tabular foundation models.\n\n### Quality and Clarity\nThe paper overall is very polished and very clear with some unclear details (see in weaknesses). The paper chooses established benchmarks and established baselines and the use of confidence intervals makes the results more conclusive.", "weaknesses": "### Contribution and Significance\nI believe the biggest weakness of the paper is the lack of clear and conclusive contribution. The paper finds full fine-tuning being the strongest and most stable baseline over other versions of fine-tuning, but it does not give any guidance or rule for when to use full fine-tuning vs other forms of fine-tuning or no fine-tuning. \n\nThe paper finds there is a correlation between the key/query alignment (or the entropy of the attention scores) and the performance improvement. However, it is still to be investigated what caused both phenomenon. Is the decrease of entropy inevitable after fine-tuning or not? Why are they correlated? \n\nInstead of analyzing different datasets separately (as in figure 4), I believe an aggregated graph of change in entropy vs performance would give a better view of the overall relationship between entropy and performance. \n\n### Choice of Models\nAnother reason the contribution is limited is because only TabPFN v2 model is used for all experiments for fine-tuning. It would be interesting to see whether TabPFN v1, TabICL, TabDPT also have similar phenomena.\n\n### Minor\nHighlights of the best model scores in table 3 would make it much more clear.", "questions": "1. The paper compares early on to prior works that uses context optimization PEFT method [line 78]. However, I believe these methods were never compared to later on in the experiments? \n2. In table 1, full fine-tuning appears to be faster than other PEFT methods, why is this? Shouldn't PEFT methods be faster?\n3. In table 2, are the Pred. lengths indicating the batch size or the sequence length? The term \"object\" in the caption is a bit confusing because it can mean either one.\n4. What happens if we take kNN from different layers instead of the last layer in table 3? Do we see entropy decreasing over the layers?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates different versions of fine-tuning of TabPFN v2. It finds simple fine-tuning being the strongest and most stable version of fine-tuning of TabPFN v2 comparing to other forms of fine-tuning strategies. In addition, this paper examines transformer attention map in details and finds the attention entropy decreases for a fine-tuned model comparing to a non-fine tuned version. Another contribution of this paper is the comparison between different models such as MLP-PLR, MNCA together with the fine-tuned and the non-fine-tuned versions of TabPFN v2.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "### Originality and Significance\nThere has been many works examining fine-tuning of tabular foundation models. However, the previous works did not have strong conclusions. The author of this paper offers a slightly stronger conclusion, i.e. \"simple full fine-tuning is a strong and stable baseline for TabPFN v2 adaptation\" although it is still not very strong. The rigorous comparison with MNCA, MLP-PLR and other models are definitely also original. \n\nI believe the most original and significant contribution comes from the analyses of the attention map and the entropy of the attention scores. The paper finds there is a correlation between the decrease of entropy and the performance improvement. The paper makes analogy of the phenomenon with retrieval, a technique that has been proven very useful in tabular foundation models.\n\n### Quality and Clarity\nThe paper overall is very polished and very clear with some unclear details (see in weaknesses). The paper chooses established benchmarks and established baselines and the use of confidence intervals makes the results more conclusive.", "weaknesses": "### Contribution and Significance\nI believe the biggest weakness of the paper is the lack of clear and conclusive contribution. The paper finds full fine-tuning being the strongest and most stable baseline over other versions of fine-tuning, but it does not give any guidance or rule for when to use full fine-tuning vs other forms of fine-tuning or no fine-tuning. \n\nThe paper finds there is a correlation between the key/query alignment (or the entropy of the attention scores) and the performance improvement. However, it is still to be investigated what caused both phenomenon. Is the decrease of entropy inevitable after fine-tuning or not? Why are they correlated? \n\nInstead of analyzing different datasets separately (as in figure 4), I believe an aggregated graph of change in entropy vs performance would give a better view of the overall relationship between entropy and performance. \n\n### Choice of Models\nAnother reason the contribution is limited is because only TabPFN v2 model is used for all experiments for fine-tuning. It would be interesting to see whether TabPFN v1, TabICL, TabDPT also have similar phenomena.\n\n### Minor\nHighlights of the best model scores in table 3 would make it much more clear.", "questions": "1. The paper compares early on to prior works that uses context optimization PEFT method [line 78]. However, I believe these methods were never compared to later on in the experiments? \n2. In table 1, full fine-tuning appears to be faster than other PEFT methods, why is this? Shouldn't PEFT methods be faster?\n3. In table 2, are the Pred. lengths indicating the batch size or the sequence length? The term \"object\" in the caption is a bit confusing because it can mean either one.\n4. What happens if we take kNN from different layers instead of the last layer in table 3? Do we see entropy decreasing over the layers?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761507006861}, {"id": "ZGTpwHp992", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5248/Reviewer_hPwk"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper studies single dataset adaptation of TabPFNv2. It compares full fine tuning to several parameter efficient methods (e.g. LoRA), reporting that full fine tuning reaches comparable or better performance with lower wall clock time. The analysis contrasts full fine tuning with training from scratch, examines dataset size effects, and argues that fine tuning improves in context retrieval behavior through better attention to relevant context. Additional experiments on TabRED (temporal shift) suggest TabPFNv2 is less stable than some alternatives. Overall, the empirical message is that simple full fine tuning is a strong and efficient method for adapting TabPFNv2 to a real dataset.", "review_text": "The paper studies single dataset adaptation of TabPFNv2. It compares full fine tuning to several parameter efficient methods (e.g. LoRA), reporting that full fine tuning reaches comparable or better performance with lower wall clock time. The analysis contrasts full fine tuning with training from scratch, examines dataset size effects, and argues that fine tuning improves in context retrieval behavior through better attention to relevant context. Additional experiments on TabRED (temporal shift) suggest TabPFNv2 is less stable than some alternatives. Overall, the empirical message is that simple full fine tuning is a strong and efficient method for adapting TabPFNv2 to a real dataset.", "strengths": "- The paper connects fine tuning outcomes to retrieval like behavior via kNN style analyses of representations, which is novel and potentially useful.\n- There is a clear claim that full fine tuning is an efficient and competitive choice for TabPFNv2 which is supported through many experiments.\n- Comparison to training from scratch and to dataset size scaling is provided that shows fine tuning is superior across many settings.", "weaknesses": "- The paper does not provide comparisons to retrieval based adaptation and sub sampling e.g. LocalPFN style retrieve + fine tune, despite being cited and working efficiently with larger datasets and being aligned with the paper's retrieval hypothesis.\n- The scope is limited to TabPFNv2, and the conclusions are not validated on other tabular foundation models such as TabICL [1] or TabDPT [2] (both more efficient to fine tune), even though the paper itself notes that findings from TabPFN to TabPFNv2 do not always transfer.\n- For Figure 3, it is unclear which attention weights are analyzed in a cell based model and the results are shown for a few datasets and two show minimal effect. There is a potential confound due to context length and attention temperature. Prior work [3] on attention temperature shows that when evaluating on longer context, the attention scores are not as sharp (higher entropy), and therefore a temperature controlled ablation is needed to show that fine tuning differs from simply temperature scaling to adapt the model to a longer context.\n\n\n[1] Qu, Jingang, et al. \"TabICL: A Tabular Foundation Model for In-Context Learning on Large Data.\" ICML 2025-Forty-Second International Conference on Machine Learning. 2025.\n\n[2] Ma, Junwei, et al. \"Tabdpt: Scaling tabular foundation models.\" arXiv preprint arXiv:2410.18164 (2024).\n\n[3] Veličković, Petar, et al. \"Softmax is not Enough (for Sharp Size Generalisation).\" Forty-second International Conference on Machine Learning.", "questions": "- What is the metric for each dataset in Table 2 (log loss, AUC, MSE, accuracy)?\n- For all settings in Table 1, what is the total number of optimization steps until early stop, the average time per step and peak memory usage? Answering this clarifies the reason for speed advantage for full fine tuning.\n- What is the representation used for kNN precisely? TabPFNv2 is cell based, how are row level similarities measured?\n- Same question as above for Figure 3. How are attention weights across samples (i.e. rows as opposed to cells) calculated?\n- Minor text editing suggestions:\n\n    - For Figure 4, the main text should preview the key definition (now in the appendix) where it is stated that the x axis is the sample index sorted by difference of entropy.\n  \n    - The paper should explicitly mention overfitting as the cause of failure in the main text (explained in Appendix E), as it is important enough to appear in the main section.\n  \n    - Real-TabPFN [4] seems directly relevant and should be cited as it fine tunes TabPFNv2 on real data (although not single dataset).\n\n[4] Garg, Anurag, et al. \"Real-tabpfn: Improving tabular foundation models via continued pre-training with real-world data.\" arXiv preprint arXiv:2507.03971 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies single dataset adaptation of TabPFNv2. It compares full fine tuning to several parameter efficient methods (e.g. LoRA), reporting that full fine tuning reaches comparable or better performance with lower wall clock time. The analysis contrasts full fine tuning with training from scratch, examines dataset size effects, and argues that fine tuning improves in context retrieval behavior through better attention to relevant context. Additional experiments on TabRED (temporal shift) suggest TabPFNv2 is less stable than some alternatives. Overall, the empirical message is that simple full fine tuning is a strong and efficient method for adapting TabPFNv2 to a real dataset.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper connects fine tuning outcomes to retrieval like behavior via kNN style analyses of representations, which is novel and potentially useful.\n- There is a clear claim that full fine tuning is an efficient and competitive choice for TabPFNv2 which is supported through many experiments.\n- Comparison to training from scratch and to dataset size scaling is provided that shows fine tuning is superior across many settings.", "weaknesses": "- The paper does not provide comparisons to retrieval based adaptation and sub sampling e.g. LocalPFN style retrieve + fine tune, despite being cited and working efficiently with larger datasets and being aligned with the paper's retrieval hypothesis.\n- The scope is limited to TabPFNv2, and the conclusions are not validated on other tabular foundation models such as TabICL [1] or TabDPT [2] (both more efficient to fine tune), even though the paper itself notes that findings from TabPFN to TabPFNv2 do not always transfer.\n- For Figure 3, it is unclear which attention weights are analyzed in a cell based model and the results are shown for a few datasets and two show minimal effect. There is a potential confound due to context length and attention temperature. Prior work [3] on attention temperature shows that when evaluating on longer context, the attention scores are not as sharp (higher entropy), and therefore a temperature controlled ablation is needed to show that fine tuning differs from simply temperature scaling to adapt the model to a longer context.\n\n\n[1] Qu, Jingang, et al. \"TabICL: A Tabular Foundation Model for In-Context Learning on Large Data.\" ICML 2025-Forty-Second International Conference on Machine Learning. 2025.\n\n[2] Ma, Junwei, et al. \"Tabdpt: Scaling tabular foundation models.\" arXiv preprint arXiv:2410.18164 (2024).\n\n[3] Veličković, Petar, et al. \"Softmax is not Enough (for Sharp Size Generalisation).\" Forty-second International Conference on Machine Learning.", "questions": "- What is the metric for each dataset in Table 2 (log loss, AUC, MSE, accuracy)?\n- For all settings in Table 1, what is the total number of optimization steps until early stop, the average time per step and peak memory usage? Answering this clarifies the reason for speed advantage for full fine tuning.\n- What is the representation used for kNN precisely? TabPFNv2 is cell based, how are row level similarities measured?\n- Same question as above for Figure 3. How are attention weights across samples (i.e. rows as opposed to cells) calculated?\n- Minor text editing suggestions:\n\n    - For Figure 4, the main text should preview the key definition (now in the appendix) where it is stated that the x axis is the sample index sorted by difference of entropy.\n  \n    - The paper should explicitly mention overfitting as the cause of failure in the main text (explained in Appendix E), as it is important enough to appear in the main section.\n  \n    - Real-TabPFN [4] seems directly relevant and should be cited as it fine tunes TabPFNv2 on real data (although not single dataset).\n\n[4] Garg, Anurag, et al. \"Real-tabpfn: Improving tabular foundation models via continued pre-training with real-world data.\" arXiv preprint arXiv:2507.03971 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761265351659}, {"id": "DG9houZjb5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5248/Reviewer_r8Z8"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper investigates the finetuning of tabular foundation model TabPFNv2, confirming full finetuning as the optimal strategy , revealing it enhances the model by improving query-key dot product alignment with target similarity and concentrating attention, while noting finetuned TabPFNv2 achieves SOTA on academic datasets but lacks stability on real-world datasets with shifts.", "review_text": "This paper investigates the finetuning of tabular foundation model TabPFNv2, confirming full finetuning as the optimal strategy , revealing it enhances the model by improving query-key dot product alignment with target similarity and concentrating attention, while noting finetuned TabPFNv2 achieves SOTA on academic datasets but lacks stability on real-world datasets with shifts.", "strengths": "1. The paper is clear and easy to understand, with intuitive tables/figures aiding comprehension.\n\n2. It provides a relatively comprehensive investigation into the finetuning of the specific model TabPFNv2, covering strategy evaluation, mechanism analysis and performance comparison.", "weaknesses": "1. There are typos that require correction. For example, Line 662 has \"concetrated\" which should be \"concentrated\", Line 693 has \"intermidiate\" which should be \"intermediate\", and Line 726 has \"perfomance\" which should be \"performance\". Though these errors do not affect the understanding of core content, they may reduce the sense of rigor, so a full text check is recommended to fix similar issues.\n\n2. The research contributions are relatively limited. The study only focuses on the finetuning evaluation of TabPFNv2, without exploring the universality of the proposed finetuning strategies on other tabular foundation models or putting forward innovative finetuning theories/methods.\n\n3. The applicable scope of the studied finetuning methods is narrow. They only work for small and medium-sized datasets and lack scalability to larger datasets. Additionally, on TabReD datasets with temporal shifts , finetuned TabPFNv2 shows reduced stability (even performance degradation), failing to adapt to complex real-world scenarios.\n\n4. The baseline comparison is insufficient. The paper does not include mainstream tabular deep learning models like ExcelFormer and RealMLP, nor does it compare with AutoGluon (mentioned in the original TabPFN paper).", "questions": "Refer to the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the finetuning of tabular foundation model TabPFNv2, confirming full finetuning as the optimal strategy , revealing it enhances the model by improving query-key dot product alignment with target similarity and concentrating attention, while noting finetuned TabPFNv2 achieves SOTA on academic datasets but lacks stability on real-world datasets with shifts.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "1. The paper is clear and easy to understand, with intuitive tables/figures aiding comprehension.\n\n2. It provides a relatively comprehensive investigation into the finetuning of the specific model TabPFNv2, covering strategy evaluation, mechanism analysis and performance comparison.", "weaknesses": "1. There are typos that require correction. For example, Line 662 has \"concetrated\" which should be \"concentrated\", Line 693 has \"intermidiate\" which should be \"intermediate\", and Line 726 has \"perfomance\" which should be \"performance\". Though these errors do not affect the understanding of core content, they may reduce the sense of rigor, so a full text check is recommended to fix similar issues.\n\n2. The research contributions are relatively limited. The study only focuses on the finetuning evaluation of TabPFNv2, without exploring the universality of the proposed finetuning strategies on other tabular foundation models or putting forward innovative finetuning theories/methods.\n\n3. The applicable scope of the studied finetuning methods is narrow. They only work for small and medium-sized datasets and lack scalability to larger datasets. Additionally, on TabReD datasets with temporal shifts , finetuned TabPFNv2 shows reduced stability (even performance degradation), failing to adapt to complex real-world scenarios.\n\n4. The baseline comparison is insufficient. The paper does not include mainstream tabular deep learning models like ExcelFormer and RealMLP, nor does it compare with AutoGluon (mentioned in the original TabPFN paper).", "questions": "Refer to the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761209528067}], "openreview_url": "https://openreview.net/forum?id=REthH7hinP", "arxiv_id": "2506.08982", "paper_pdf": "papers/REthH7hinP.pdf", "paper_pdf_sha256": "d713ba870923c2e49284b08db68677610dbc9597979271c8d3f994e1e2715762", "paper_pdf_bytes": 1179675, "paper_pdf_source": "openreview", "code_url": "https://github.com/yandex-research/tabpfn-finetuning", "code_repository": "yandex-research/tabpfn-finetuning", "code_commit": "c3ce0f74b8501de2c5f52abe2c03f5b4b382c628", "code_archive": "repos/REthH7hinP.zip", "code_archive_sha256": "b61ab048b1e3a66cd90ed0e741ff90002d49240004aa7d4b0a73e1da94dfe312", "code_archive_bytes": 1595905, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 845, "github_languages": {"Python": 541084}, "github_archived": false, "github_pushed_at": "2025-09-03T10:39:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-finetuning-tabular-foundation-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UfczlMudN6", "year": 2025, "status": "rejected", "title": "GRAM: Generalization in Deep RL with a Robust Adaptation Module", "authors": ["James Queeney", "Xiaoyi Cai", "Mouhacine Benosman", "JONATHAN P HOW"], "authorids": ["~James_Queeney1", "~Xiaoyi_Cai2", "~Mouhacine_Benosman1", "~JONATHAN_P_HOW1"], "authors_source": "OpenReview API", "abstract": "The reliable deployment of deep reinforcement learning in real-world settings requires the ability to generalize across a variety of conditions, including both in-distribution scenarios seen during training as well as novel out-of-distribution scenarios. In this work, we present a framework for dynamics generalization in deep reinforcement learning that unifies these two distinct types of generalization within a single architecture. We introduce a robust adaptation module that provides a mechanism for identifying and reacting to both in-distribution and out-of-distribution environment dynamics, along with a joint training pipeline that combines the goals of in-distribution adaptation and out-of-distribution robustness. Our algorithm GRAM achieves strong generalization performance across in-distribution and out-of-distribution scenarios upon deployment, which we demonstrate on a variety of realistic simulated locomotion tasks with a quadruped robot.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NqWpZ5LaQl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7421/Reviewer_E5e4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The authors propose a novel method (GRAM) to learn RL policies that exhibit both adaptive and robust behavior through a unified training pipeline, specifically in the context of deployment environments that vary in dynamics at test time. GRAM employs context-conditioned policies where the context is automatically learned through a history of past observations and actions. Crucially, when the history falls OOD or is unable to provide context point estimates, the learned context converges to a special null context vector, so that robust behavior is triggered.", "review_text": "The authors propose a novel method (GRAM) to learn RL policies that exhibit both adaptive and robust behavior through a unified training pipeline, specifically in the context of deployment environments that vary in dynamics at test time. GRAM employs context-conditioned policies where the context is automatically learned through a history of past observations and actions. Crucially, when the history falls OOD or is unable to provide context point estimates, the learned context converges to a special null context vector, so that robust behavior is triggered.", "strengths": "- The manuscript is extremely well written and provides clear statements to understand the proposed method.\n- The authors demonstrate a successful implementation of recently introduced Epistemic Neural Networks for a relevant and important open problem in robot learning (i.e. generalization over unobservable environment contexts)", "weaknesses": "- Missing recent related works: the authors should consider citing and discussing recent works [1] and [2] as they present themselves as state-of-the-art methods in Robust RL and Domain Randomization as of 2024. Specifically, DORAEMON [2] tackles the same problem setting as in this work where privileged information is available at training time, and a history of previous observations and actions is used to allow implicit system identification at test time and promote adaptive behavior.\n\n- Limited experimental evaluation:\n  - While Domain Randomization is an extremely popular baseline for learning robust/adaptive behavior in sim2real settings, the authors provide little information to the implementation of this baseline and no comparison of GRAM vs. DR in fig. 4 and 5.\n  - Implementation details of DR: I highly suggest the authors to compare GRAM against a DR baseline which also uses a history of previous state and actions, as this is the notorious way to implement DR [3]. Furthermore, DR may and should also leverage privileged information at training time, by using the notorious asymmetric actor-critic paradigm [2, 3, 4]. Conditioning the critic on the known context often drastically affects the results of DR methods vs. unprivileged critics, and does not require further assumptions.\n  - The analysis is carried out on locomotion environments only, and simulated environments only. The generalization problem under unobservable dynamics is likely exacerbated and more challenging for manipulation and contact-rich settings, which would make the experimental evaluation more significant and relevant.\n\n[1] Reddi, A. et al. \"Robust Adversarial Reinforcement Learning via Bounded Rationality Curricula.\" ICLR 2024.\n\n[2] Tiboni, G. et al. \"Domain Randomization via Entropy Maximization.\" ICLR 2024.\n\n[3] Peng, Xue Bin, et al. \"Sim-to-real transfer of robotic control with dynamics randomization.\" 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 2018.\n\n[4] Handa, Ankur, et al. \"Dextreme: Transfer of agile in-hand manipulation from simulation to reality.\" 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023.", "questions": "- Why is the performance of all baselines better in Fig. 3 (right) for the \"Base ID + frozen joints\" vs. the \"Base ID\" counterpart?\n\n- Regarding the statement at the end of Sec. 5 \"as the goal is to train the epinet in (5) to output estimates with low variance in ID contexts and high variance in OOD contexts.\". Does GRAM promote the learned encoder to output high variance in OOD contexts at training time? Judging from Eq. (7) alone, it seems to me that the encoder is only trained to provide low variance on ID contexts, whereas a higher OOD variance is a spontaneous effect that occurs at test time only.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a novel method (GRAM) to learn RL policies that exhibit both adaptive and robust behavior through a unified training pipeline, specifically in the context of deployment environments that vary in dynamics at test time. GRAM employs context-conditioned policies where the context is automatically learned through a history of past observations and actions. Crucially, when the history falls OOD or is unable to provide context point estimates, the learned context converges to a special null context vector, so that robust behavior is triggered.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The manuscript is extremely well written and provides clear statements to understand the proposed method.\n- The authors demonstrate a successful implementation of recently introduced Epistemic Neural Networks for a relevant and important open problem in robot learning (i.e. generalization over unobservable environment contexts)", "weaknesses": "- Missing recent related works: the authors should consider citing and discussing recent works [1] and [2] as they present themselves as state-of-the-art methods in Robust RL and Domain Randomization as of 2024. Specifically, DORAEMON [2] tackles the same problem setting as in this work where privileged information is available at training time, and a history of previous observations and actions is used to allow implicit system identification at test time and promote adaptive behavior.\n\n- Limited experimental evaluation:\n  - While Domain Randomization is an extremely popular baseline for learning robust/adaptive behavior in sim2real settings, the authors provide little information to the implementation of this baseline and no comparison of GRAM vs. DR in fig. 4 and 5.\n  - Implementation details of DR: I highly suggest the authors to compare GRAM against a DR baseline which also uses a history of previous state and actions, as this is the notorious way to implement DR [3]. Furthermore, DR may and should also leverage privileged information at training time, by using the notorious asymmetric actor-critic paradigm [2, 3, 4]. Conditioning the critic on the known context often drastically affects the results of DR methods vs. unprivileged critics, and does not require further assumptions.\n  - The analysis is carried out on locomotion environments only, and simulated environments only. The generalization problem under unobservable dynamics is likely exacerbated and more challenging for manipulation and contact-rich settings, which would make the experimental evaluation more significant and relevant.\n\n[1] Reddi, A. et al. \"Robust Adversarial Reinforcement Learning via Bounded Rationality Curricula.\" ICLR 2024.\n\n[2] Tiboni, G. et al. \"Domain Randomization via Entropy Maximization.\" ICLR 2024.\n\n[3] Peng, Xue Bin, et al. \"Sim-to-real transfer of robotic control with dynamics randomization.\" 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 2018.\n\n[4] Handa, Ankur, et al. \"Dextreme: Transfer of agile in-hand manipulation from simulation to reality.\" 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023.", "questions": "- Why is the performance of all baselines better in Fig. 3 (right) for the \"Base ID + frozen joints\" vs. the \"Base ID\" counterpart?\n\n- Regarding the statement at the end of Sec. 5 \"as the goal is to train the epinet in (5) to output estimates with low variance in ID contexts and high variance in OOD contexts.\". Does GRAM promote the learned encoder to output high variance in OOD contexts at training time? Judging from Eq. (7) alone, it seems to me that the encoder is only trained to provide low variance on ID contexts, whereas a higher OOD variance is a spontaneous effect that occurs at test time only.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730718561973}, {"id": "lSP8f84e3C", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7421/Reviewer_zHHm"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper develops an approach for generalization in reinforcement learning by tackling the problems of in-distribution generalization through contextual RL, and out-of-distribution generalization through robust RL. The authors develop a unified approach, called \"Robust Adaptation Module\" that relies of teacher-student training for tackling these challenges of generalization. The teacher policy is trained through RL in simulation by using privileged information of the environment that is not accessible at deployment. The student policy is trained to imitate rollouts from the teacher policy and leverage an adaptation module that predicts \"context\" from history of interactions. Experiments with a simulated quadruped robot in fixed velocity locomotion demonstrates the applicability of the approach in generalizing to different dynamics variations like mass and friction.", "review_text": "The paper develops an approach for generalization in reinforcement learning by tackling the problems of in-distribution generalization through contextual RL, and out-of-distribution generalization through robust RL. The authors develop a unified approach, called \"Robust Adaptation Module\" that relies of teacher-student training for tackling these challenges of generalization. The teacher policy is trained through RL in simulation by using privileged information of the environment that is not accessible at deployment. The student policy is trained to imitate rollouts from the teacher policy and leverage an adaptation module that predicts \"context\" from history of interactions. Experiments with a simulated quadruped robot in fixed velocity locomotion demonstrates the applicability of the approach in generalizing to different dynamics variations like mass and friction.", "strengths": "- The paper is interesting and the key ideas of adaptive RL and robust RL are easy to follow \n\n- To the best of my knowledge the robust adaptation module relying on context identification from history and epistemic uncertainty of the policy is novel in the context of generalization in RL. \n\n- The use of a teacher-student training paradigm is nice, and since the student doesn't require privileged information, the approach can be potentially scaled to real world tasks via sim2real. \n\n- The experiments are on a non-trivial simulated quadruped locomotion task and the generalization to dynamcis variations corresponds to a practical use-case in controls and robotics.", "weaknesses": "- It is unclear how general is the proposed approach in dealing with different types of variations. The paper is motivated from the perspective of very generic generalization in RL but the experiments are limited to a single task/environment and correspond to only dynamics variations. What about variations in visual scenes? generalization to different tasks with the same robot?\n\n- The in-distribution and out-of distribution generalization motivation is a bit confusing in the introduction. The authors should clarify scenarios where out-of-distribution generalization is possible without in-distribution generalization. The approach is based on the former not necessarily implying the latter and so this is important to motivate and explain with concrete examples. \n\n- It is a bit confusing that context prediction module is only trained from on-policy data. Can't off-policy and generic offline data with the same robot/agent be used for learning this context prediction? This could potentially alleviate some of the challenges with out-of-distribution generalization as well and remove the need for the second part of the module. \n\n- Some prior works that also do context prediction and demonstrate real-world quadruped locomotion through student-teaching training are not cited and discussed. For example the RMA paper below: \n\nKumar, Ashish, Zipeng Fu, Deepak Pathak, and Jitendra Malik. \"Rma: Rapid motor adaptation for legged robots.\"", "questions": "Refer to the weaknesses above. \n\n- It is unclear how general is the proposed approach in dealing with different types of variations. What about variations in visual scenes? generalization to different tasks with the same robot?\n\n- The in-distribution and out-of distribution generalization motivation is a bit confusing in the introduction. What are examples of scenarios where out-of-distribution generalization is possible without in-distribution generalization?\n\n- It is a bit confusing that context prediction module is only trained from on-policy data. Can't off-policy and generic offline data with the same robot/agent be used for learning this context prediction?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper develops an approach for generalization in reinforcement learning by tackling the problems of in-distribution generalization through contextual RL, and out-of-distribution generalization through robust RL. The authors develop a unified approach, called \"Robust Adaptation Module\" that relies of teacher-student training for tackling these challenges of generalization. The teacher policy is trained through RL in simulation by using privileged information of the environment that is not accessible at deployment. The student policy is trained to imitate rollouts from the teacher policy and leverage an adaptation module that predicts \"context\" from history of interactions. Experiments with a simulated quadruped robot in fixed velocity locomotion demonstrates the applicability of the approach in generalizing to different dynamics variations like mass and friction.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper is interesting and the key ideas of adaptive RL and robust RL are easy to follow \n\n- To the best of my knowledge the robust adaptation module relying on context identification from history and epistemic uncertainty of the policy is novel in the context of generalization in RL. \n\n- The use of a teacher-student training paradigm is nice, and since the student doesn't require privileged information, the approach can be potentially scaled to real world tasks via sim2real. \n\n- The experiments are on a non-trivial simulated quadruped locomotion task and the generalization to dynamcis variations corresponds to a practical use-case in controls and robotics.", "weaknesses": "- It is unclear how general is the proposed approach in dealing with different types of variations. The paper is motivated from the perspective of very generic generalization in RL but the experiments are limited to a single task/environment and correspond to only dynamics variations. What about variations in visual scenes? generalization to different tasks with the same robot?\n\n- The in-distribution and out-of distribution generalization motivation is a bit confusing in the introduction. The authors should clarify scenarios where out-of-distribution generalization is possible without in-distribution generalization. The approach is based on the former not necessarily implying the latter and so this is important to motivate and explain with concrete examples. \n\n- It is a bit confusing that context prediction module is only trained from on-policy data. Can't off-policy and generic offline data with the same robot/agent be used for learning this context prediction? This could potentially alleviate some of the challenges with out-of-distribution generalization as well and remove the need for the second part of the module. \n\n- Some prior works that also do context prediction and demonstrate real-world quadruped locomotion through student-teaching training are not cited and discussed. For example the RMA paper below: \n\nKumar, Ashish, Zipeng Fu, Deepak Pathak, and Jitendra Malik. \"Rma: Rapid motor adaptation for legged robots.\"", "questions": "Refer to the weaknesses above. \n\n- It is unclear how general is the proposed approach in dealing with different types of variations. What about variations in visual scenes? generalization to different tasks with the same robot?\n\n- The in-distribution and out-of distribution generalization motivation is a bit confusing in the introduction. What are examples of scenarios where out-of-distribution generalization is possible without in-distribution generalization?\n\n- It is a bit confusing that context prediction module is only trained from on-policy data. Can't off-policy and generic offline data with the same robot/agent be used for learning this context prediction?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730662724763}, {"id": "n7rEIftgZt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7421/Reviewer_GgR5"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes a new method for robust in-context adaptation in reinforcement learning. To remedy the OOD context distribution in the inference, the GRAM method proposes to adapt the the latent vector with a pre-trained context encoder that estimate the uncertainty of the context and bias the latent context into a confident region (i.e., 0). Empirical results show that the adaptation yields robust performance w.r.t. other baselines in the frozen joint OOD test cases.", "review_text": "This paper proposes a new method for robust in-context adaptation in reinforcement learning. To remedy the OOD context distribution in the inference, the GRAM method proposes to adapt the the latent vector with a pre-trained context encoder that estimate the uncertainty of the context and bias the latent context into a confident region (i.e., 0). Empirical results show that the adaptation yields robust performance w.r.t. other baselines in the frozen joint OOD test cases.", "strengths": "1. The method is simple, novel and easy to understand. It does not involve too much code change with vinilla in-context adaptation.\n2. The empirical results is comprehensive in quadraped OOD generalization case, which is persuasive.", "weaknesses": "1. The paper does not consider other easy to implement baselines, e.g. domain invariance prediction, adding noise to the context, or use a larger replay buffer for the context encoder training. I think to show that this specific design is useful, one should also consider other easy-to-implement baselines.\n2. GRAM biases towards 0, which is the mean at the beginning of the training. It is unclear if this is the best choice. How about bias toward the mean at the end of the training of the context encoder?", "questions": "See the two points in weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new method for robust in-context adaptation in reinforcement learning. To remedy the OOD context distribution in the inference, the GRAM method proposes to adapt the the latent vector with a pre-trained context encoder that estimate the uncertainty of the context and bias the latent context into a confident region (i.e., 0). Empirical results show that the adaptation yields robust performance w.r.t. other baselines in the frozen joint OOD test cases.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The method is simple, novel and easy to understand. It does not involve too much code change with vinilla in-context adaptation.\n2. The empirical results is comprehensive in quadraped OOD generalization case, which is persuasive.", "weaknesses": "1. The paper does not consider other easy to implement baselines, e.g. domain invariance prediction, adding noise to the context, or use a larger replay buffer for the context encoder training. I think to show that this specific design is useful, one should also consider other easy-to-implement baselines.\n2. GRAM biases towards 0, which is the mean at the beginning of the training. It is unclear if this is the best choice. How about bias toward the mean at the end of the training of the context encoder?", "questions": "See the two points in weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730662248065}, {"id": "VdsQtB3qqO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7421/Reviewer_BE6V"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces GRAM (Generalization in Deep RL with a Robust Adaptation Module), a deep RL framework designed to enhance generalization performance in both in-distribution (ID) and out-of-distribution (OOD) scenarios. GRAM aims to unify both generalizations within a single architecture.\n\nThe core of GRAM is the robust adaptation module. This module utilizes an epistemic neural network to estimate the current context and the uncertainty associated with that estimation.  This uncertainty measure allows the system to identify OOD situations.  When high uncertainty is detected, the module defaults to a pre-defined \"robust\" latent feature, triggering a robust control policy specifically trained for such scenarios. This module is trained by teacher-student distillation. A \"teacher\" policy is trained with access to contextual information, learning to adapt to different environments.  A \"student\" policy then learns to mimic the teacher's behavior based only on observed history, enabling deployment without needing explicit contextual data. \n\nGRAM's training pipeline also combines standard data collection for ID adaptation with adversarial training for OOD robustness.  An \"adversary\" agent introduces disturbances during training, forcing the robust policy to learn to handle unexpected perturbations. \n\nExperiments conducted on simulated quadruped robot locomotion tasks demonstrate GRAM's effectiveness.  Compared to traditional contextual and robust RL methods, GRAM achieves stronger performance in both ID and OOD environments.  It demonstrates adaptive behavior in familiar terrains while maintaining robust locomotion in challenging, unseen terrains like rough surfaces.  The experiments also confirm that GRAM effectively identifies OOD situations, automatically adjusting its behavior based on the uncertainty level of its context estimation.", "review_text": "This paper introduces GRAM (Generalization in Deep RL with a Robust Adaptation Module), a deep RL framework designed to enhance generalization performance in both in-distribution (ID) and out-of-distribution (OOD) scenarios. GRAM aims to unify both generalizations within a single architecture.\n\nThe core of GRAM is the robust adaptation module. This module utilizes an epistemic neural network to estimate the current context and the uncertainty associated with that estimation.  This uncertainty measure allows the system to identify OOD situations.  When high uncertainty is detected, the module defaults to a pre-defined \"robust\" latent feature, triggering a robust control policy specifically trained for such scenarios. This module is trained by teacher-student distillation. A \"teacher\" policy is trained with access to contextual information, learning to adapt to different environments.  A \"student\" policy then learns to mimic the teacher's behavior based only on observed history, enabling deployment without needing explicit contextual data. \n\nGRAM's training pipeline also combines standard data collection for ID adaptation with adversarial training for OOD robustness.  An \"adversary\" agent introduces disturbances during training, forcing the robust policy to learn to handle unexpected perturbations. \n\nExperiments conducted on simulated quadruped robot locomotion tasks demonstrate GRAM's effectiveness.  Compared to traditional contextual and robust RL methods, GRAM achieves stronger performance in both ID and OOD environments.  It demonstrates adaptive behavior in familiar terrains while maintaining robust locomotion in challenging, unseen terrains like rough surfaces.  The experiments also confirm that GRAM effectively identifies OOD situations, automatically adjusting its behavior based on the uncertainty level of its context estimation.", "strengths": "- GRAM tackles an important challenge of generalization in deep RL by proposing a unified architecture that effectively combines adaptation and robustness. This is an important perspective as previous methods often focused on one at the expense of the other.\n\n- The introduction of the robust adaptation module with its integrated epistemic neural network is a novel mechanism. It allows the system to not only adapt to different contexts but also quantify the uncertainty of its estimations, enabling switching between ID and OOD modes.\n\n- The paper is well-written and easy to follow.", "weaknesses": "### Major\n\n- The assumption underlying this paper is notably strong. While assuming access to context variable values within the simulation environment is reasonable, *assuming the knowledge of which parameters constitute the context is a strong assumption*. In their experiments, contexts include friction, added base mass, motor strength, and joint angle bias - effectively acknowledging a priori which parameters will vary in test environments. This raises the question: if these variation sources are known beforehand, why not just randomly sample these parameters from a wider range during training?\n\n- The adversary policy design similarly relies on prior knowledge. In their experiments, the adversary policy applies external forces to the robot's body. This is clearly tailored to address specific anticipated environmental variations. Such an adversary policy would be useless if the primary differences between training and test environments involved different factors, such as lighting conditions or camera poses.\n\n- It seems the motivation of this paper is to deploy deep RL policies in real-world settings reliably (the 1st sentence in abstract). However, alll experiments are conducted in simulation.  While simulation allows for controlled experiments and extensive testing, demonstrating GRAM's effectiveness on real-world hardware is crucial for validating its practical applicability.  Real-world deployments often introduce unforeseen challenges that simulations may not fully capture.\n\n\n### Minor\n\n- The OOD experiments primarily focus on variations in terrain roughness and robot parameters. While relevant, exploring a wider range of OOD scenarios, such as unexpected obstacles or changes in task objectives, would provide a more comprehensive evaluation of GRAM's robustness.\n\n- The paper mentions fine-tuning the parameters for the uncertainty hyperparameter ($\\alpha_t$) but doesn't provide a detailed analysis of its sensitivity.  Understanding how different threshold values impact performance in various ID and OOD scenarios would be beneficial.", "questions": "- It is still unclear to me why the $z_{rob}$ is set to 0. Why should it output 0 for context embedding if uncertainty is high?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces GRAM (Generalization in Deep RL with a Robust Adaptation Module), a deep RL framework designed to enhance generalization performance in both in-distribution (ID) and out-of-distribution (OOD) scenarios. GRAM aims to unify both generalizations within a single architecture.\n\nThe core of GRAM is the robust adaptation module. This module utilizes an epistemic neural network to estimate the current context and the uncertainty associated with that estimation.  This uncertainty measure allows the system to identify OOD situations.  When high uncertainty is detected, the module defaults to a pre-defined \"robust\" latent feature, triggering a robust control policy specifically trained for such scenarios. This module is trained by teacher-student distillation. A \"teacher\" policy is trained with access to contextual information, learning to adapt to different environments.  A \"student\" policy then learns to mimic the teacher's behavior based only on observed history, enabling deployment without needing explicit contextual data. \n\nGRAM's training pipeline also combines standard data collection for ID adaptation with adversarial training for OOD robustness.  An \"adversary\" agent introduces disturbances during training, forcing the robust policy to learn to handle unexpected perturbations. \n\nExperiments conducted on simulated quadruped robot locomotion tasks demonstrate GRAM's effectiveness.  Compared to traditional contextual and robust RL methods, GRAM achieves stronger performance in both ID and OOD environments.  It demonstrates adaptive behavior in familiar terrains while maintaining robust locomotion in challenging, unseen terrains like rough surfaces.  The experiments also confirm that GRAM effectively identifies OOD situations, automatically adjusting its behavior based on the uncertainty level of its context estimation.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- GRAM tackles an important challenge of generalization in deep RL by proposing a unified architecture that effectively combines adaptation and robustness. This is an important perspective as previous methods often focused on one at the expense of the other.\n\n- The introduction of the robust adaptation module with its integrated epistemic neural network is a novel mechanism. It allows the system to not only adapt to different contexts but also quantify the uncertainty of its estimations, enabling switching between ID and OOD modes.\n\n- The paper is well-written and easy to follow.", "weaknesses": "### Major\n\n- The assumption underlying this paper is notably strong. While assuming access to context variable values within the simulation environment is reasonable, *assuming the knowledge of which parameters constitute the context is a strong assumption*. In their experiments, contexts include friction, added base mass, motor strength, and joint angle bias - effectively acknowledging a priori which parameters will vary in test environments. This raises the question: if these variation sources are known beforehand, why not just randomly sample these parameters from a wider range during training?\n\n- The adversary policy design similarly relies on prior knowledge. In their experiments, the adversary policy applies external forces to the robot's body. This is clearly tailored to address specific anticipated environmental variations. Such an adversary policy would be useless if the primary differences between training and test environments involved different factors, such as lighting conditions or camera poses.\n\n- It seems the motivation of this paper is to deploy deep RL policies in real-world settings reliably (the 1st sentence in abstract). However, alll experiments are conducted in simulation.  While simulation allows for controlled experiments and extensive testing, demonstrating GRAM's effectiveness on real-world hardware is crucial for validating its practical applicability.  Real-world deployments often introduce unforeseen challenges that simulations may not fully capture.\n\n\n### Minor\n\n- The OOD experiments primarily focus on variations in terrain roughness and robot parameters. While relevant, exploring a wider range of OOD scenarios, such as unexpected obstacles or changes in task objectives, would provide a more comprehensive evaluation of GRAM's robustness.\n\n- The paper mentions fine-tuning the parameters for the uncertainty hyperparameter ($\\alpha_t$) but doesn't provide a detailed analysis of its sensitivity.  Understanding how different threshold values impact performance in various ID and OOD scenarios would be beneficial.", "questions": "- It is still unclear to me why the $z_{rob}$ is set to 0. Why should it output 0 for context embedding if uncertainty is high?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730273031991}], "openreview_url": "https://openreview.net/forum?id=UfczlMudN6", "arxiv_id": "2412.04323", "paper_pdf": "papers/UfczlMudN6.pdf", "paper_pdf_sha256": "68bc51b302afc7118dc7bc9edc60b5b762bc646f0398ee2d02b7943780b0952b", "paper_pdf_bytes": 1208532, "paper_pdf_source": "openreview", "code_url": "https://github.com/merlresearch/gram", "code_repository": "merlresearch/gram", "code_commit": "ea3f26ce898d7fdae4cc51497b22e8361a1c3f02", "code_archive": "repos/UfczlMudN6.zip", "code_archive_sha256": "30487eeff83c146517db59b94b0d706ff64ddccd3858fedd05d7d64621e942e8", "code_archive_bytes": 56237, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 70, "github_languages": {"Python": 209080}, "github_archived": false, "github_pushed_at": "2026-07-16T20:57:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gram-generalization-in-deep-rl-with-a-robust"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pm7O7gJObtk", "year": 2023, "status": "rejected", "title": "NIERT: Accurate Numerical Interpolation through Unifying Scattered Data Representations using Transformer Encoder", "authors": ["Shizhe Ding", "Dongbo Bu"], "authorids": ["~Shizhe_Ding2", "~Dongbo_Bu1"], "authors_source": "OpenReview API", "abstract": "Numerical interpolation for scattered data, i.e., estimating values for target points based on those of some observed points, is widely used in  computational science and engineering. The existing approaches either require explicitly pre-defined basis functions, which makes them inflexible and limits their performance in practical scenarios, or train neural networks as interpolators, which still have limited interpolation accuracy as they treat observed and target points separately and cannot effectively exploit the correlations among data points. Here, we present a learning-based approach to numerical interpolation for scattered data using encoder representation of Transformers (called NIERT). Unlike the recent learning-based approaches, NIERT treats observed and target points in a unified fashion through embedding them into the same representation space, thus gaining the advantage of  effectively exploiting the correlations among them. The specially-designed partial self-attention mechanism used by NIERT makes it escape from the unexpected interference of target points on observed points. We further show that the partial self-attention is essentially a learnable interpolation module combining multiple neural basis functions, which provides interpretability of NIERT. Through pre-training on large-scale synthetic datasets,  NIERT achieves considerable improvement in interpolation accuracy for practical tasks. On both synthetic and real-world datasets, NIERT  outperforms the existing approaches, e.g., on the TFRD-ADlet dataset for temperature field reconstruction, NIERT achieves an MAE of $1.897\\times 10^{-3}$, substantially better than the state-of-the-art  approach (MAE: $27.074\\times 10^{-3}$).  The source code of NIERT is available at  https://anonymous.4open.science/r/NIERT-2BCF.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "i9SBSKzVrs4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper863/Reviewer_rgW7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper introduces a new partial self-attention layer to improve numerical interpolation for scattered data. The module was used in a modified Transformer to solve the interpolation task. The proposed approach treats observed and target points in a unified way by embedding them in the same representation space.", "review_text": "In terms of text quality, the paper is easy to read. There are some minor issues with the text, but nothing major, except that the sentence abruptly ended at the end of page 4.\n\nExperiments and evaluation design are done nicely; the authors conducted many experiments and ablation studies. \n\nHowever, there are some issues with the paper:\n\nThe paper claims that one of the reasons why other methods perform poorly is because they process observed and target points separately. This fact prevents them from exploiting correlations between observed and target points (see page 2, end of paragraph 1, and page 5 before the experiments section). I do not see this claim proven in the paper, although I may be wrong.\n\nThere are a lot of variables (for example, different network parameter sizes) that can influence the outcome of this kind of experiment, so careful experiment design is needed to prove this.\n\nIn section 4.4 in figure 5, the fact that attention maps are imbalanced does not mean much. In my opinion, its area of affection for each point depends on its neighbors' count and their proximity placement. For example, if you have a lot of points in one area, it is natural that each point will contribute, or can either contribute a little to a wider area, or can contribute a lot to a local area. Which option is better is a good question. So the claim in the paper at the end of this section that NIERT can exploit the correlation between points observed and target points more effectively, in my opinion, is not well-supported. \n\nFinally, the simple use of the pre-training technique is not a contribution. \n\nQuestions:\nIs it correct that partly self-attention is just a particular case of self-attention, where observed and target points can’t “look” at the target points(wij=0)? \nPlease clarify, what is input in each batch for your neural network in evaluation time? \n\n\nSmall suggestions:\n1. Section 3.4 is a good addition to the text but can be moved to supplementary.\n2. It would be nice to see a comparison of several hyperparameters for each model for the methods you compare with.\n", "strengths": "Strength:\nA new mechanism to mask points in self-attention was introduced. \nThe proposed model achieves a better result on the synthetic and real-world dataset from the paper compared to the current state-of-the-art methods.\nExtensive analysis and ablation studies were conducted.\nThe Paper is written in an easy to read manner\nStrong evaluation and comparisons.\n\nWeakness:\nMistypes in the text.\nSome claims of the paper were not well-supported (see summary for more details).", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper introduces a new partial self-attention layer to improve numerical interpolation for scattered data. The module was used in a modified Transformer to solve the interpolation task. The proposed approach treats observed and target points in a unified way by embedding them in the same representation space.", "strength_and_weaknesses": "Strength:\nA new mechanism to mask points in self-attention was introduced. \nThe proposed model achieves a better result on the synthetic and real-world dataset from the paper compared to the current state-of-the-art methods.\nExtensive analysis and ablation studies were conducted.\nThe Paper is written in an easy to read manner\nStrong evaluation and comparisons.\n\nWeakness:\nMistypes in the text.\nSome claims of the paper were not well-supported (see summary for more details).", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written, and the results are novel to the best of my knowledge. An anonymous link to the code is provided. There are a lot of experiments and ablation studies conducted both on synthetic and real-world datasets.", "summary_of_the_review": "In terms of text quality, the paper is easy to read. There are some minor issues with the text, but nothing major, except that the sentence abruptly ended at the end of page 4.\n\nExperiments and evaluation design are done nicely; the authors conducted many experiments and ablation studies. \n\nHowever, there are some issues with the paper:\n\nThe paper claims that one of the reasons why other methods perform poorly is because they process observed and target points separately. This fact prevents them from exploiting correlations between observed and target points (see page 2, end of paragraph 1, and page 5 before the experiments section). I do not see this claim proven in the paper, although I may be wrong.\n\nThere are a lot of variables (for example, different network parameter sizes) that can influence the outcome of this kind of experiment, so careful experiment design is needed to prove this.\n\nIn section 4.4 in figure 5, the fact that attention maps are imbalanced does not mean much. In my opinion, its area of affection for each point depends on its neighbors' count and their proximity placement. For example, if you have a lot of points in one area, it is natural that each point will contribute, or can either contribute a little to a wider area, or can contribute a lot to a local area. Which option is better is a good question. So the claim in the paper at the end of this section that NIERT can exploit the correlation between points observed and target points more effectively, in my opinion, is not well-supported. \n\nFinally, the simple use of the pre-training technique is not a contribution. \n\nQuestions:\nIs it correct that partly self-attention is just a particular case of self-attention, where observed and target points can’t “look” at the target points(wij=0)? \nPlease clarify, what is input in each batch for your neural network in evaluation time? \n\n\nSmall suggestions:\n1. Section 3.4 is a good addition to the text but can be moved to supplementary.\n2. It would be nice to see a comparison of several hyperparameters for each model for the methods you compare with.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667508458713}, {"id": "Aws1d8DLZn", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper863/Reviewer_qiuo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a learning-based approach using the encoder representations of Transformers for numerical interpolation of scattered data.", "review_text": "This paper combines existing techniques with trivial modification to achieve good results for numerical interpolation. Due to lack of novelty, I believe this is a good technical report for practitioners to apply deep learning methods for numerical interpolation, but by its current form it is not ready to be published a paper in a top machine learning venue.", "strengths": "Strength: this paper combines a set of existing (and indeed very popular) deep learning methods to deal with the numerical interpolation problem, and it achieves good results by experiments on real data sets.\n\n\nWeakness: there are several concerns, among which lack of novelty is the major concern.\n\n1. This paper combines a set of very popular techniques in deep learning to the numerical interpolation problem, such as transformer with self-attention and enhancement with pre-training. Every component of the proposed method is either existing or trivial, so it is not clear what is the novel method proposed by this paper.\n\n2. The claimed contribution, the partial self-attention mechanism, is mostly trivial which sets zero correlation weights to target points (eq. (2)).  While the authors argued that such partial self-attention mechanism is connected to Radial Basis Function (RBF interpolation), it is really trivial to me: if one sets zero weights to target points, then it is naturally written as a combination of functions on the observed points. In addition, while observed and target points are decoupled, I believe it should be helpful to boost the performance by modeling the correlation among target points by common principles of semi-supervised learning (the target points are given).\n\n3. Why are different metrics applied to different data sets, such as MSE for the NeSymReS, D30 and PhysioNet datasets, and MAE (CMAE,BMAE) for the TFRD-ADlet dataset? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a learning-based approach using the encoder representations of Transformers for numerical interpolation of scattered data.", "strength_and_weaknesses": "Strength: this paper combines a set of existing (and indeed very popular) deep learning methods to deal with the numerical interpolation problem, and it achieves good results by experiments on real data sets.\n\n\nWeakness: there are several concerns, among which lack of novelty is the major concern.\n\n1. This paper combines a set of very popular techniques in deep learning to the numerical interpolation problem, such as transformer with self-attention and enhancement with pre-training. Every component of the proposed method is either existing or trivial, so it is not clear what is the novel method proposed by this paper.\n\n2. The claimed contribution, the partial self-attention mechanism, is mostly trivial which sets zero correlation weights to target points (eq. (2)).  While the authors argued that such partial self-attention mechanism is connected to Radial Basis Function (RBF interpolation), it is really trivial to me: if one sets zero weights to target points, then it is naturally written as a combination of functions on the observed points. In addition, while observed and target points are decoupled, I believe it should be helpful to boost the performance by modeling the correlation among target points by common principles of semi-supervised learning (the target points are given).\n\n3. Why are different metrics applied to different data sets, such as MSE for the NeSymReS, D30 and PhysioNet datasets, and MAE (CMAE,BMAE) for the TFRD-ADlet dataset? ", "clarity,_quality,_novelty_and_reproducibility": "This paper is mostly clear and well written. Source code is provided for reproducibility. Please refer to my above comments for the major concern about lack of novelty.", "summary_of_the_review": "This paper combines existing techniques with trivial modification to achieve good results for numerical interpolation. Due to lack of novelty, I believe this is a good technical report for practitioners to apply deep learning methods for numerical interpolation, but by its current form it is not ready to be published a paper in a top machine learning venue.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666973968790}, {"id": "pXmgmzh2a_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper863/Reviewer_uN6v"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper introduces NIERT which is a new framework for numerical interpolation for scattered data. The key idea of the framework is to use Transformers as a representation encoder and treats observed and target points in a unified fashion by embedding them into the same representation space. In the paper, the authors also propose a new partial self-attention mechanism that could escape from the unexpected interference of target points on observed points. Combining the partial self-attention mechanism,  masking mechanism, and pre-training technique, NIERT outperforms the existing approaches on both real-world and synthetic datasets. For example, NIERT gives the best mean absolute error on the NeSymReS dataset, the TFRD-ADlet dataset, the D30 dataset, and the PhysioNet. The authors also conduct ablation studies to understand the role of partial self-attention and the pre-training technique.", "review_text": "The paper proposes the partial self-attention mechanism and applies it to the numerical interpolation problem. The framework achieves great prediction results however the performance in downstream tasks is missing.", "strengths": "## Strength\n* The proposed framework exploits the power of Transform in the numerical interpolation where traditional interpolation approaches for scattered data use explicitly pre-defined basis functions to construct interpolation functions.\n* The partial self-attention mechanism is new to the best of my knowledge.\n* The authors also discuss the connection between NIERT and the traditional approach by showing that partial self-attention is a general form of the RBF interpolation function.\n* The proposed framework beats all baselines on all datasets.\n\n## Weaknesses\n* The computational trade-off should be discussed e.g, the training time, and the interpolation time.\n* The authors should discuss the statistics about the number of observed and target points in each dataset. I wonder how the proposed framework performs in different settings of sequence lengths.\n* The authors only focus on the mean absolute error. There is no downstream task for showing that NIERT is better than previous frameworks.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper introduces NIERT which is a new framework for numerical interpolation for scattered data. The key idea of the framework is to use Transformers as a representation encoder and treats observed and target points in a unified fashion by embedding them into the same representation space. In the paper, the authors also propose a new partial self-attention mechanism that could escape from the unexpected interference of target points on observed points. Combining the partial self-attention mechanism,  masking mechanism, and pre-training technique, NIERT outperforms the existing approaches on both real-world and synthetic datasets. For example, NIERT gives the best mean absolute error on the NeSymReS dataset, the TFRD-ADlet dataset, the D30 dataset, and the PhysioNet. The authors also conduct ablation studies to understand the role of partial self-attention and the pre-training technique.", "strength_and_weaknesses": "## Strength\n* The proposed framework exploits the power of Transform in the numerical interpolation where traditional interpolation approaches for scattered data use explicitly pre-defined basis functions to construct interpolation functions.\n* The partial self-attention mechanism is new to the best of my knowledge.\n* The authors also discuss the connection between NIERT and the traditional approach by showing that partial self-attention is a general form of the RBF interpolation function.\n* The proposed framework beats all baselines on all datasets.\n\n## Weaknesses\n* The computational trade-off should be discussed e.g, the training time, and the interpolation time.\n* The authors should discuss the statistics about the number of observed and target points in each dataset. I wonder how the proposed framework performs in different settings of sequence lengths.\n* The authors only focus on the mean absolute error. There is no downstream task for showing that NIERT is better than previous frameworks.", "clarity,_quality,_novelty_and_reproducibility": "* Clarity: The paper is well-written and easy to follow.\n* Quality: The proposed framework performs well in practice.\n* Novelty: Based on the claims of the authors, this is the first work that applies deep learning (Transformer) to the numerical interpolation problem and the partial self-attention mechanism is new.\n*  Reproducibility: The code is submitted. The experimental settings are reported in detail.", "summary_of_the_review": "The paper proposes the partial self-attention mechanism and applies it to the numerical interpolation problem. The framework achieves great prediction results however the performance in downstream tasks is missing.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666581620468}], "openreview_url": "https://openreview.net/forum?id=pm7O7gJObtk", "arxiv_id": "2209.09078", "paper_pdf": "papers/pm7O7gJObtk.pdf", "paper_pdf_sha256": "cee365554b4ab7b58dee3d6bf866b9a12639926f582dae8587ea3e12a4286129", "paper_pdf_bytes": 4251714, "paper_pdf_source": "openreview", "code_url": "https://github.com/DingShizhe/NIERT", "code_repository": "DingShizhe/NIERT", "code_commit": "053e7edd4ff0270a2205a78ba07ef1460ea8a354", "code_archive": "repos/pm7O7gJObtk.zip", "code_archive_sha256": "c61a0033d69fdc9bdcf68322f485a3fcf93154550e2e9ccec2676bfd649a6639", "code_archive_bytes": 77290, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 101, "github_languages": {"Python": 250843}, "github_archived": false, "github_pushed_at": "2024-11-14T11:08:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/niert-accurate-numerical-interpolation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vyn49BUAkoD", "year": 2022, "status": "rejected", "title": "Bayesian Active Learning with Fully Bayesian Gaussian Processes", "authors": ["Christoffer Riis", "Francisco Antunes", "Frederik Boe Hüttel", "Carlos Lima Azevedo", "Francisco C. Pereira"], "authorids": ["~Christoffer_Riis1", "~Francisco_Antunes1", "~Frederik_Boe_Hüttel1", "~Carlos_Lima_Azevedo1", "~Francisco_C._Pereira1"], "authors_source": "OpenReview API", "abstract": "The bias-variance trade-off is a well-known problem in machine learning that only gets more pronounced the less available data there is. When data is scarce, such as in metamodeling, active learning, and Bayesian optimization, neglecting this trade-off can cause inefficient and non-optimal querying, leading to unnecessary data labeling. In this paper, we focus on metamodeling with active learning and the canonical Gaussian Process (GP). We recognize that, for the GP, the bias-variance trade-off regulation is made by optimization of the two hyperparameters: the length scale and noise-term. Considering that the optimal mode of the joint posterior of the hyperparameters is equivalent to the optimal bias-variance trade-off, we approximate this joint posterior and utilize it to design two new acquisition functions. The first one is a mode-seeking Bayesian variant of Query-by-Committee (B-QBC), and the second is simultaneously mode-seeking and minimizing the predictive variance through a Query by Mixture Gaussian Processes (QB-MGP) formulation. Across seven simulators, we empirically show that B-QBC outperforms the benchmark functions, whereas QB-MGP is the most robust acquisition function and achieves the best accuracy with the fewest iterations. We generally show that incorporating the bias-variance trade-off in the acquisition functions mitigates unnecessary and expensive data labeling.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "YkxrVVHzBVF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2862/Reviewer_hDWW"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper takes a fully Bayesian approach to Gaussian Process (GP) active learning by using MCMC sampling to consider multiple model hypotheses from a full posterior, from which it selects examples for GP regression using two new active learning strategies --- Bayesian Query-by-Committee (B-QBC) and Query by Mixture of Gaussian Processes (QB-MGP). B-QBC queries the data point that maximizes the disagreement between the sampled models’ mean values. The idea is that this should provide the highest information about the optimal posterior mode. QB-MGP uses a combination of B-QBC with entropy sampling. The authors evaluate these strategies against several baseline active learning methods on several simulators.", "review_text": "####################################\n\nStrengths:\n- The themes discussed in this paper, including the full posterior having modes corresponding to either signal or noise which should be discriminated between with active learning, are very interesting. In particular, Figures 1 and 2 were interesting and illustrative.\n- The related work section and description of Gaussian Processes is mostly satisfactory\n- The experiments compare against a diverse set of active learning acquisition functions and experimental simulators\n\n####################################\n\nWeaknesses:\n\nWhile this paper presents some interesting ideas, overall I feel that certain major elements need to be significantly improved before it is ready for publication.\n\nWhen I first read the abstract, I was intrigued by the idea of the bias-variance tradeoff being neglected in active learning, how in GP this tradeoff corresponds to the length scale and noise term, how “the optimal mode of the joint posterior of the hyperparameters is equivalent to the optimal bias-variance tradeoff”, and how B-QBC and QB-MGP directly incorporate this tradeoff in their acquisition functions to mitigate bias-variance issues. However, in my opinion, these points are not strongly reflected in the paper itself:\n\n*** Introduction ***\n\nStarting in the introduction, I had expected a cohesive theme illustrating the issues presented in the abstract: how the bias-variance tradeoff is neglected in active learning, a strong argument for bias-variance correspondence to length scale and noise term, and strong connections with the hyperparameters posterior. Instead, I found the topics in the introduction only very loosely coupled: specifically, I think the connections between the predictor vs guide, bias vs variance, and signal vs noise in GPs need to be strengthened and more direct arguments made. For instance, statements such as “The problem is directly related to the well-known bias-variance trade-off, although we reformulate it as a balance between modeling the data as signal or noise” or “However, none of\nthese approaches directly address the core problem: which mode to choose? Ideally, this should be\nanswered with prior information about the problem, although typically that is not available, making\nthese approaches less practical.” seem extremely important, but most of the time I found such critical statements vague and unsupported. While Figure 1-2 were interesting, I did not find myself convinced that the bimodal behavior in Figure 2 is a general phenomenon and found myself unsure how it connected to bias and variance. Is this multimodal behavior between signal and noise an established fact in the literature, or is this perspective a major contribution of this paper? Either way, I believe statements around this point need to be strengthened. Overall, after reading the introduction I found myself unconvinced about the strong points presented in the abstract. I also expected that I would find more details supporting these points in Section 3, which seem to be missing there (e.g., establishing connections between bias-variance and specific GP hyperparameters).\n\n*** Sections 2-3 ***\n\nOtherwise, I was mostly happy with Sections 2-3. One point in Section 3 I’d like to bring up is that the authors claim earlier in the paper that “In this paper, we follow a general approach and assume no prior knowledge about the kernel or hyperparameters.” However, the authors state in Section 3 that they use an RBF kernel. Can the authors comment about this? Is this specific choice not in fact an assumption about the kernel?\n\n*** Section 4 ***\n\nIn Section 4, B-QBC seems reasonable, but I was not convinced that it somehow addressed the key bias-variance tradeoff and/or mode selection problem in GP active learning described earlier in the paper. The idea of discriminating between modes seems reasonable, but I believe that in order to motivate B-QBC, a much stronger argument (stronger than the illustrative example in Figures 1-2) needs to be presented about why the hyperparameter posterior is multimodal, and how B-QBC will specifically select the correct one. In terms of QB-MGP, I found myself confused by its presentation. I was not convinced that equation (11) is maximizing the predictive variance of the mixture model --- I can imagine that it comes directly from a variance decomposition, but I believe the derivation of this fact should be presented. Moreover, as with B-QBC, why should maximizing the predictive variance of this mixture model solve the mode-selection problem and tradeoff between bias and variance? Furthermore, how do the authors expect B-QBC and QB-MGP to compare qualitatively? Since QB-MGP is equivalent to B-QBC plus a variance term, how do the authors expect this variance term to impact the performance of the acquisition function? Overall, B-QBC and QB-MGP seem to me to be the two fundamental contributions of the paper, and I believe more discussion is needed here to derive and motivate their use and strongly connect them to resolving bias-variance/signal-noise/mode selection tradeoffs.\n\n*** Section 5 ***\n\nWhile I appreciate the multitude of simulators tested in the experiments, I am not convinced by the experimental results. The performance curves seem to have high variance, and I am not convinced by the summary statistics and takeaways presented. It is not at all clear to me from Figure 4-5 and Tables 2-3 that “B-QBC outperforms the benchmark functions, whereas QB-MGP is the most robust acquisition function and achieves the best accuracy with the fewest iterations.” \n\nSpecifically, analyzing performance at a single iteration number based on a converged reference curve only captures a small part of the overall picture. I believe it would be better to examine summary statistics that better capture the overall performance of each method, such as evaluating multiple iteration numbers or looking at summary statistics such as AUC. The authors make a comment in section 5.3 attempting to justify this single iteration point evaluation, but I did not find it convincing. I am also unsure if the likelihood ratio and error quotients in Table 2 provides an adequate summary of the data. When using the raw results in Table 3, any sort of benefits for B-QBC and QB-MGP in Figure 5 become much less clear to me when looking at raw data instead of quotients. Can the authors comment about this choice of presenting summary statistics in terms of relative performance rather than raw values? Regardless, given the not insignificant error bars, I would be more convinced by these results if they were accompanied by statistical testing against either the best performing method or random sampling (see other active learning meta-analyses such as “A benchmark and comparison of active learning for logistic regression”, Yang and Loog 2018).\n\nIn Figure 4, it seems difficult to take away anything very useful from these experiments, due to the overlapping error bars in each trial and the fact that in some experiments and methods (including QB-MGP and B-QBC) the error is in fact increasing. It seems that the performance differences can be made more clear by plotting error bars for summary statistics (e.g., standard error or confidence intervals) and increasing the number of trials beyond only 10. I do not look at Figure 4 and have confidence that I should use QB-MGP and B-QBC over the other methods. In my opinion, the takeaways stated in the text are overstated based on these curves, such as QB-MGP being the “best choice” from the Sine RMSE plot, when it seems to be that multiple methods tie QB-MGP in performance, or that B-QBC could even be considered better than the other methods, given the overlapping error bars. Furthermore, there are analytical takeaways that to me seem vague and unsupported, such as “This agrees with our expectation that QB-MGP seems to be more robust than B-QBC to different types of noise.” Why is this an expectation? Statements such as “QB-MGP is achieving the highest RMSE due to favoring to capture the small region with a non-linear signal compared to the vast linear region” or “Looking at all the active learning iterations, the GP fitted with MAP seems to give the best trade-off between the NLML and RMSE, indicating that the regular GP with ALM is the best for smooth problems” are in my opinion speculative takeaways that seem unsupported. Could QB-MGP favoring the non-linear portions be shown empirically, for instance?\n\nAlthough it’s possible that I missed these details in my reading (although I went back again to check), I have some reproducibility concerns. It is unclear how “convergence” is calculated for selecting each plot reference as well as the examined iteration number. How exactly are NLML and RMSE calculated, do they utilize the full posterior, and which data are they calculated with respect to --- the sampled data points, or a held-out set of randomly sampled data? Are the distributions in Figure 5 at the simulator level, or is all data pooled together from all simulators? In Table 2, if different trials' references converged at different times, how was a single iteration value selected? Low-level details such as these are important and seem to be missing (or should at least be included in an appendix), which prevents a full understanding of the experimental results.\n\nFurthermore, I believe additional experiments should be run to support the main claims made in the paper. A major claim is that the full posterior is multi-modal, motivating the proposed active learning experiments. I think an experiment that shows this empirically through some metric would strengthen this major claim. I also think that additional experiments are needed to characterize the performance of the proposed schemes beyond performance alone. Specifically, my reading of the paper is that the proposed methods are designed to select the correct full posterior mode as quickly as possible. It would be useful to show during trials of B-QBC and/or QB-MGP how a single (correct) mode of the full posterior emerges during MCMC sampling, and the others are suppressed (and possibly comparing this metric against baseline methods).\n\nIn general, I did not find myself convinced by Section 5.3 and found it lacking. It discusses two important details --- why a single iteration slice is sufficient for performance analysis, and where specifically in the acquisition and model evaluation pipeline the full posterior was used rather than the hyperparameter MAP (and why this choice was made for every step) --- but I found the first discussion unconvincing and the second discussion a bit  confusing. Specifically, I found this important sentence confusing “However, since we focus on benchmarking the acquisition functions and not the models, we believe that using the mode is more accurate when comparing it to the regular GP since we only change the fitting procedure.”\n\n######################################\n\nOther points *not* factoring into decision, but serving as feedback for the authors:\n\nIt has been pointed out before (Settles, B. (2012). Active learning: Synthesis lectures on artificial intelligence and machine learning. Long Island, NY: Morgan &  Clay Pool. --- Section 3.5) that information gain maximization can be interpreted as a type of query by committee, where KL divergence on the label is used as a notion of disagreement. Specifically, if the BALD acquisition function is equivalently written as $\\mathbb{E}_{\\theta} [\\mathrm{KL}(p(Y \\mid \\theta) \\Vert p(Y))]$, then this is comparable to B-QBC, which uses a mean discrepancy instead of a KL-divergence. I think this is an interesting connection that would strengthen the work in terms of why B-QBC might perform better than BALD. What is mean discrepancy doing well that KL divergence isn't?\n\nTo me, the use of \"GMM\" seems unusual here to describe a mixture of Gaussian Processes. I checked the literature cited by the authors, and I did not see “Gaussian Mixture Model” or GMM being used as the authors do here to describe a mixture of Gaussian Processes (as opposed to the classical GMM notion of a mixture of multivariate Gaussians in a vector space). Personally, I think using the phrase \"GMM\" runs the risk of confusing the model with a classical GMM framework (as opposed to a Gaussian Process). However, this may be a matter of personal opinion.\n\nThe first paragraph of Section 5 seems out of place to me. It discusses important points, but I think Section 5 should start with a more general description of the experiments instead of just the Sine simulator. Also, I think the experiments could be strengthened by comparison against random (passive) sampling.\n\nMinor points:\n- in the abstract, Query by Mixture Gaussian Processes -> Query by Mixture *of* Gaussian Processes\n- the sentence “The literature suggests handling this problem with clever initializations...either always initializing…” seems redundant to me\n- the location of Gramacy2d and Motorcycle should probably be switched in Figure 4 since the text seems to follow a row-first description of Figure 4\n- I think the right subplot of Figure 5 should show a zoomed-in region to better visualize the methods besides BALD\n- Personally, I don’t think the future work paragraph adds much to the paper, and I don’t understand what this sentence proposes: “Eventually, our ultimate goal is to provide practitioners with an auxiliary tool to map the simulators’ output behaviors in a more efficient manner.”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper takes a fully Bayesian approach to Gaussian Process (GP) active learning by using MCMC sampling to consider multiple model hypotheses from a full posterior, from which it selects examples for GP regression using two new active learning strategies --- Bayesian Query-by-Committee (B-QBC) and Query by Mixture of Gaussian Processes (QB-MGP). B-QBC queries the data point that maximizes the disagreement between the sampled models’ mean values. The idea is that this should provide the highest information about the optimal posterior mode. QB-MGP uses a combination of B-QBC with entropy sampling. The authors evaluate these strategies against several baseline active learning methods on several simulators.", "main_review": "####################################\n\nStrengths:\n- The themes discussed in this paper, including the full posterior having modes corresponding to either signal or noise which should be discriminated between with active learning, are very interesting. In particular, Figures 1 and 2 were interesting and illustrative.\n- The related work section and description of Gaussian Processes is mostly satisfactory\n- The experiments compare against a diverse set of active learning acquisition functions and experimental simulators\n\n####################################\n\nWeaknesses:\n\nWhile this paper presents some interesting ideas, overall I feel that certain major elements need to be significantly improved before it is ready for publication.\n\nWhen I first read the abstract, I was intrigued by the idea of the bias-variance tradeoff being neglected in active learning, how in GP this tradeoff corresponds to the length scale and noise term, how “the optimal mode of the joint posterior of the hyperparameters is equivalent to the optimal bias-variance tradeoff”, and how B-QBC and QB-MGP directly incorporate this tradeoff in their acquisition functions to mitigate bias-variance issues. However, in my opinion, these points are not strongly reflected in the paper itself:\n\n*** Introduction ***\n\nStarting in the introduction, I had expected a cohesive theme illustrating the issues presented in the abstract: how the bias-variance tradeoff is neglected in active learning, a strong argument for bias-variance correspondence to length scale and noise term, and strong connections with the hyperparameters posterior. Instead, I found the topics in the introduction only very loosely coupled: specifically, I think the connections between the predictor vs guide, bias vs variance, and signal vs noise in GPs need to be strengthened and more direct arguments made. For instance, statements such as “The problem is directly related to the well-known bias-variance trade-off, although we reformulate it as a balance between modeling the data as signal or noise” or “However, none of\nthese approaches directly address the core problem: which mode to choose? Ideally, this should be\nanswered with prior information about the problem, although typically that is not available, making\nthese approaches less practical.” seem extremely important, but most of the time I found such critical statements vague and unsupported. While Figure 1-2 were interesting, I did not find myself convinced that the bimodal behavior in Figure 2 is a general phenomenon and found myself unsure how it connected to bias and variance. Is this multimodal behavior between signal and noise an established fact in the literature, or is this perspective a major contribution of this paper? Either way, I believe statements around this point need to be strengthened. Overall, after reading the introduction I found myself unconvinced about the strong points presented in the abstract. I also expected that I would find more details supporting these points in Section 3, which seem to be missing there (e.g., establishing connections between bias-variance and specific GP hyperparameters).\n\n*** Sections 2-3 ***\n\nOtherwise, I was mostly happy with Sections 2-3. One point in Section 3 I’d like to bring up is that the authors claim earlier in the paper that “In this paper, we follow a general approach and assume no prior knowledge about the kernel or hyperparameters.” However, the authors state in Section 3 that they use an RBF kernel. Can the authors comment about this? Is this specific choice not in fact an assumption about the kernel?\n\n*** Section 4 ***\n\nIn Section 4, B-QBC seems reasonable, but I was not convinced that it somehow addressed the key bias-variance tradeoff and/or mode selection problem in GP active learning described earlier in the paper. The idea of discriminating between modes seems reasonable, but I believe that in order to motivate B-QBC, a much stronger argument (stronger than the illustrative example in Figures 1-2) needs to be presented about why the hyperparameter posterior is multimodal, and how B-QBC will specifically select the correct one. In terms of QB-MGP, I found myself confused by its presentation. I was not convinced that equation (11) is maximizing the predictive variance of the mixture model --- I can imagine that it comes directly from a variance decomposition, but I believe the derivation of this fact should be presented. Moreover, as with B-QBC, why should maximizing the predictive variance of this mixture model solve the mode-selection problem and tradeoff between bias and variance? Furthermore, how do the authors expect B-QBC and QB-MGP to compare qualitatively? Since QB-MGP is equivalent to B-QBC plus a variance term, how do the authors expect this variance term to impact the performance of the acquisition function? Overall, B-QBC and QB-MGP seem to me to be the two fundamental contributions of the paper, and I believe more discussion is needed here to derive and motivate their use and strongly connect them to resolving bias-variance/signal-noise/mode selection tradeoffs.\n\n*** Section 5 ***\n\nWhile I appreciate the multitude of simulators tested in the experiments, I am not convinced by the experimental results. The performance curves seem to have high variance, and I am not convinced by the summary statistics and takeaways presented. It is not at all clear to me from Figure 4-5 and Tables 2-3 that “B-QBC outperforms the benchmark functions, whereas QB-MGP is the most robust acquisition function and achieves the best accuracy with the fewest iterations.” \n\nSpecifically, analyzing performance at a single iteration number based on a converged reference curve only captures a small part of the overall picture. I believe it would be better to examine summary statistics that better capture the overall performance of each method, such as evaluating multiple iteration numbers or looking at summary statistics such as AUC. The authors make a comment in section 5.3 attempting to justify this single iteration point evaluation, but I did not find it convincing. I am also unsure if the likelihood ratio and error quotients in Table 2 provides an adequate summary of the data. When using the raw results in Table 3, any sort of benefits for B-QBC and QB-MGP in Figure 5 become much less clear to me when looking at raw data instead of quotients. Can the authors comment about this choice of presenting summary statistics in terms of relative performance rather than raw values? Regardless, given the not insignificant error bars, I would be more convinced by these results if they were accompanied by statistical testing against either the best performing method or random sampling (see other active learning meta-analyses such as “A benchmark and comparison of active learning for logistic regression”, Yang and Loog 2018).\n\nIn Figure 4, it seems difficult to take away anything very useful from these experiments, due to the overlapping error bars in each trial and the fact that in some experiments and methods (including QB-MGP and B-QBC) the error is in fact increasing. It seems that the performance differences can be made more clear by plotting error bars for summary statistics (e.g., standard error or confidence intervals) and increasing the number of trials beyond only 10. I do not look at Figure 4 and have confidence that I should use QB-MGP and B-QBC over the other methods. In my opinion, the takeaways stated in the text are overstated based on these curves, such as QB-MGP being the “best choice” from the Sine RMSE plot, when it seems to be that multiple methods tie QB-MGP in performance, or that B-QBC could even be considered better than the other methods, given the overlapping error bars. Furthermore, there are analytical takeaways that to me seem vague and unsupported, such as “This agrees with our expectation that QB-MGP seems to be more robust than B-QBC to different types of noise.” Why is this an expectation? Statements such as “QB-MGP is achieving the highest RMSE due to favoring to capture the small region with a non-linear signal compared to the vast linear region” or “Looking at all the active learning iterations, the GP fitted with MAP seems to give the best trade-off between the NLML and RMSE, indicating that the regular GP with ALM is the best for smooth problems” are in my opinion speculative takeaways that seem unsupported. Could QB-MGP favoring the non-linear portions be shown empirically, for instance?\n\nAlthough it’s possible that I missed these details in my reading (although I went back again to check), I have some reproducibility concerns. It is unclear how “convergence” is calculated for selecting each plot reference as well as the examined iteration number. How exactly are NLML and RMSE calculated, do they utilize the full posterior, and which data are they calculated with respect to --- the sampled data points, or a held-out set of randomly sampled data? Are the distributions in Figure 5 at the simulator level, or is all data pooled together from all simulators? In Table 2, if different trials' references converged at different times, how was a single iteration value selected? Low-level details such as these are important and seem to be missing (or should at least be included in an appendix), which prevents a full understanding of the experimental results.\n\nFurthermore, I believe additional experiments should be run to support the main claims made in the paper. A major claim is that the full posterior is multi-modal, motivating the proposed active learning experiments. I think an experiment that shows this empirically through some metric would strengthen this major claim. I also think that additional experiments are needed to characterize the performance of the proposed schemes beyond performance alone. Specifically, my reading of the paper is that the proposed methods are designed to select the correct full posterior mode as quickly as possible. It would be useful to show during trials of B-QBC and/or QB-MGP how a single (correct) mode of the full posterior emerges during MCMC sampling, and the others are suppressed (and possibly comparing this metric against baseline methods).\n\nIn general, I did not find myself convinced by Section 5.3 and found it lacking. It discusses two important details --- why a single iteration slice is sufficient for performance analysis, and where specifically in the acquisition and model evaluation pipeline the full posterior was used rather than the hyperparameter MAP (and why this choice was made for every step) --- but I found the first discussion unconvincing and the second discussion a bit  confusing. Specifically, I found this important sentence confusing “However, since we focus on benchmarking the acquisition functions and not the models, we believe that using the mode is more accurate when comparing it to the regular GP since we only change the fitting procedure.”\n\n######################################\n\nOther points *not* factoring into decision, but serving as feedback for the authors:\n\nIt has been pointed out before (Settles, B. (2012). Active learning: Synthesis lectures on artificial intelligence and machine learning. Long Island, NY: Morgan &  Clay Pool. --- Section 3.5) that information gain maximization can be interpreted as a type of query by committee, where KL divergence on the label is used as a notion of disagreement. Specifically, if the BALD acquisition function is equivalently written as $\\mathbb{E}_{\\theta} [\\mathrm{KL}(p(Y \\mid \\theta) \\Vert p(Y))]$, then this is comparable to B-QBC, which uses a mean discrepancy instead of a KL-divergence. I think this is an interesting connection that would strengthen the work in terms of why B-QBC might perform better than BALD. What is mean discrepancy doing well that KL divergence isn't?\n\nTo me, the use of \"GMM\" seems unusual here to describe a mixture of Gaussian Processes. I checked the literature cited by the authors, and I did not see “Gaussian Mixture Model” or GMM being used as the authors do here to describe a mixture of Gaussian Processes (as opposed to the classical GMM notion of a mixture of multivariate Gaussians in a vector space). Personally, I think using the phrase \"GMM\" runs the risk of confusing the model with a classical GMM framework (as opposed to a Gaussian Process). However, this may be a matter of personal opinion.\n\nThe first paragraph of Section 5 seems out of place to me. It discusses important points, but I think Section 5 should start with a more general description of the experiments instead of just the Sine simulator. Also, I think the experiments could be strengthened by comparison against random (passive) sampling.\n\nMinor points:\n- in the abstract, Query by Mixture Gaussian Processes -> Query by Mixture *of* Gaussian Processes\n- the sentence “The literature suggests handling this problem with clever initializations...either always initializing…” seems redundant to me\n- the location of Gramacy2d and Motorcycle should probably be switched in Figure 4 since the text seems to follow a row-first description of Figure 4\n- I think the right subplot of Figure 5 should show a zoomed-in region to better visualize the methods besides BALD\n- Personally, I don’t think the future work paragraph adds much to the paper, and I don’t understand what this sentence proposes: “Eventually, our ultimate goal is to provide practitioners with an auxiliary tool to map the simulators’ output behaviors in a more efficient manner.”\n", "summary_of_the_review": "In my opinion, this paper is ultimately making strides at an interesting and important topic in GP active learning --- what role does the full posterior have in managing the bias-variance tradeoff, and how can these insights be used to design better active learning acquisition functions? I think the proposition of QB-MGP and B-QBC is an important stride in this direction. However, I believe the paper is not ready for publication. In terms of presentation, I found the arguments made about bias-variance in its relation to the full posterior and the motivation for QB-MGP and B-QBC unconvincing. I also found some statements made in the paper to be confusing or unsupported. In terms of experiments, I found myself unconvinced by the presented analysis (specifically due to the high variance in the results and the lack of more general summary statistics such as AUC rather than a single iteration slice) and am concerned with some aspects of reproducibility. Therefore, I would recommend **rejection** for this paper and encourage the authors to revise the writing to strengthen the presented arguments, better motivate their active learning strategies, improve reproducibility, and conduct a stronger analysis and visualization of the experimental results.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635896444128}, {"id": "Mjb6yavGDB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2862/Reviewer_H5XZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper considers Bayesian treatment of the hyperparameters in Gaussian process based active learning, based on whichtwo novel acquisition functions are proposed.", "review_text": "Strengths: This work is interesting and well-motivated. Bayesian Gaussian process-based active learning has practical benefits as demonstrated in numerical tests.\n\nWeaknesses:\n1. It is well known that active learning tends to explore the function space. The proposed B-QBC seems to exploit more of the available information. Is there any intuition why it works well in practice?\n2. It is not clear what the performance metrics (i.e., NLML and NMSE) are in the experiments. How is the experimental validation different from Bayesian optimization?\n3. The experiments are based on synthetic functions. It would be better if practical problems are considered.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper considers Bayesian treatment of the hyperparameters in Gaussian process based active learning, based on whichtwo novel acquisition functions are proposed.", "main_review": "Strengths: This work is interesting and well-motivated. Bayesian Gaussian process-based active learning has practical benefits as demonstrated in numerical tests.\n\nWeaknesses:\n1. It is well known that active learning tends to explore the function space. The proposed B-QBC seems to exploit more of the available information. Is there any intuition why it works well in practice?\n2. It is not clear what the performance metrics (i.e., NLML and NMSE) are in the experiments. How is the experimental validation different from Bayesian optimization?\n3. The experiments are based on synthetic functions. It would be better if practical problems are considered.", "summary_of_the_review": "This is an interesting work that considered Bayesian Gaussian processes for active learning. But the merits of the paper can be better enhanced by addressing the aforementioned comments.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635887571503}, {"id": "RKRlsfREzii", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2862/Reviewer_ssFN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "It seems that the authors introduces two novel acquisitions functions for Gaussian Process models:\nthe first one is a mode-seeking Bayesian variant of Query-by- Committee (B-QBC), and the second is simultaneously mode-seeking and minimizing the predictive variance through a Query by Mixture Gaussian Processes (QB-MGP) formulation.  The authors also present several simulations.", "review_text": "The paper seems to be interesting. \nThe part that I most enjoy to read is Section 2 of \"related works\" and Section 4.1, which are also the clearer parts of the paper. \n\n- The rest of the paper, including abstract and introduction, is quite confusing in my opinion. The concepts of \"bias-variance trade-off\" and \"active learning\" are often referred but it is very difficult to understand the connection.\n\n- Some sentences like:\n\n\"In this paper, we follow a general approach and assume no prior knowledge about the kernel or\nhyperparameters. We consider multiple model hypotheses by replacing the fitting procedure of the\nmarginal likelihood with Markov Chain Monte Carlo (MCMC) sampling. Lalchand & Rasmussen\n(2020) show that with fixed, medium-sized data sets and carefully chosen kernels, it is beneficial\nto fit GPs with MCMC instead of by maximizing the marginal likelihood.\" \n\nare not very precise. I assume that you compare different kernels approximating the marginal likelihood ? \nOr maybe you refers considering a mixture of models with different kernels. In any case, we are assuming some kind of kernel functions and you have to learn their parameters. Please, improve your explanations in all the work.\n\n- Regarding the state of the art discussion in Sections 2 and 4.1 about active learning. Novel more robust (with less numerical problems) and more advanced (considering the gradient of the solutions) acquisition functions has been considered in the literature\n\nD. H. Svendsen et al, Active Emulation of Computer Codes with Gaussian Processes - Application to Remote Sensing, Pattern Recognition Volume 100, 2020, \n\nF. Llorente et al., \"Adaptive quadrature schemes for Bayesian inference via active learning\", IEEE Access, Volume 8, 2020,\n\nM. Kanagawa and P. Hennig, “Convergence Guarantees for Adaptive Bayesian Quadrature Methods,” in Advances in Neural Information Processing Systems, 2019, pp. 6234–6245.\n\nThey have been successfully applied in regression and to build adaptive quadrature schemes (all of them in the context of active learning). Please, include these kind of acquisition functions in your discussion.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "It seems that the authors introduces two novel acquisitions functions for Gaussian Process models:\nthe first one is a mode-seeking Bayesian variant of Query-by- Committee (B-QBC), and the second is simultaneously mode-seeking and minimizing the predictive variance through a Query by Mixture Gaussian Processes (QB-MGP) formulation.  The authors also present several simulations.", "main_review": "The paper seems to be interesting. \nThe part that I most enjoy to read is Section 2 of \"related works\" and Section 4.1, which are also the clearer parts of the paper. \n\n- The rest of the paper, including abstract and introduction, is quite confusing in my opinion. The concepts of \"bias-variance trade-off\" and \"active learning\" are often referred but it is very difficult to understand the connection.\n\n- Some sentences like:\n\n\"In this paper, we follow a general approach and assume no prior knowledge about the kernel or\nhyperparameters. We consider multiple model hypotheses by replacing the fitting procedure of the\nmarginal likelihood with Markov Chain Monte Carlo (MCMC) sampling. Lalchand & Rasmussen\n(2020) show that with fixed, medium-sized data sets and carefully chosen kernels, it is beneficial\nto fit GPs with MCMC instead of by maximizing the marginal likelihood.\" \n\nare not very precise. I assume that you compare different kernels approximating the marginal likelihood ? \nOr maybe you refers considering a mixture of models with different kernels. In any case, we are assuming some kind of kernel functions and you have to learn their parameters. Please, improve your explanations in all the work.\n\n- Regarding the state of the art discussion in Sections 2 and 4.1 about active learning. Novel more robust (with less numerical problems) and more advanced (considering the gradient of the solutions) acquisition functions has been considered in the literature\n\nD. H. Svendsen et al, Active Emulation of Computer Codes with Gaussian Processes - Application to Remote Sensing, Pattern Recognition Volume 100, 2020, \n\nF. Llorente et al., \"Adaptive quadrature schemes for Bayesian inference via active learning\", IEEE Access, Volume 8, 2020,\n\nM. Kanagawa and P. Hennig, “Convergence Guarantees for Adaptive Bayesian Quadrature Methods,” in Advances in Neural Information Processing Systems, 2019, pp. 6234–6245.\n\nThey have been successfully applied in regression and to build adaptive quadrature schemes (all of them in the context of active learning). Please, include these kind of acquisition functions in your discussion.\n\n", "summary_of_the_review": "The paper contains interesting material, but specially abstract and introduction are very confused.\nThe state-of-the-art discussion must be also improved.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635849699232}, {"id": "o1RCOkKscQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2862/Reviewer_uC7q"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces two new active-learning strategies for Gaussian process regression, based on a fully-Bayesian treatment of GPs. Instead of using fixed hyperparamers obtained by maximizing the marginal log-likelihood of the GP model, the authors suggest learning a posterior belief over GP hyperparameters, and make use of the full distribution to design more effective active-learning methods.\n\nIn practice, this results in two concrete proposals:\n\n1) a strategy that is similar to query-by-committee, where the committee is formed of multiple GPs sampled from the hyper-posterior (B-QBC).\n2) a variant that combines a notion of disagreement with the predictive uncertainty (QB-MGP)\n\nThese methods are then compared to others on 8 synthetic datasets.", "review_text": "**Strengths**\n\n- the contributions of this paper are interesting and appear to be original\n- the approaches are based on sound probabilistic foundations\n- the experiments on synthetic data provide intuitive insights into the relative merits of different approaches.\n\n**Weaknesses**\n\n- connections and comparison to prior work is incomplete\n- experimental results are not very conclusive; no experiments on real data\n- parts of the paper are unnecessarily convoluted and unclear\n\nThe problem of overfitting to the marginal likelihood problem is well-known in the GP community. In the context of Bayesian optimization for example, some authors have proposed a similar method that takes advantage of a mixture of GPs [1]. How does this paper's approach differ?\n\nThe experimental results on synthetic data are quite comprehensive and insightful. One relevant baseline that appears to be missing is Zhao et al (2020). How does their approach compare empirically?\n\nIt is difficult to reconcile the results on the negative log-marginal likelihood (NLML) with the RMSE. The squared error is identical to the negative log-likelihood (under a Gaussian model) and should thus closely match the NLML results - how do you explain the difference in relative ranking and the fact that many of the RMSE curves slope upwards? How is the RMSE in fact evaluated—is there a hold-out set?\n\nAdditional experiments on real-world datasets & finding one realistic application where the proposed approach is clearly superior would significantly strengthen the paper.\n\nFinally, I encourage the authors to improve the clarity of the paper: introducing all concepts progressively, and making statements precise. Examples:\n\n- canonical Gaussian Process: to the best of my knowledge, this is not a widely used terminology for the RBF kernel\n- metamodeling is never formally introduced\n- \"given the data (x, y), ..., a GP is typically denoted as GP(...)\" -> the data does not appear to be relevant to the way the GP is denoted.\n- the definition of the RBF kernel is incorrect (there is a scalar quantity, the squared norm, that is divided by a vector)\n- I don't think the reference to the \"no free lunch\" theorem is appropriate in the context of the discussion on page 4\n- \"the inner integral is intractable and is therefore optimized with approximate inference\": what exactly is optimized?\n- section 5: how is convergence declared?\n\n[1]: An Empirical Bayes Approach to Optimizing Machine Learning Algorithms, J. McInerney, NIPS 2017", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces two new active-learning strategies for Gaussian process regression, based on a fully-Bayesian treatment of GPs. Instead of using fixed hyperparamers obtained by maximizing the marginal log-likelihood of the GP model, the authors suggest learning a posterior belief over GP hyperparameters, and make use of the full distribution to design more effective active-learning methods.\n\nIn practice, this results in two concrete proposals:\n\n1) a strategy that is similar to query-by-committee, where the committee is formed of multiple GPs sampled from the hyper-posterior (B-QBC).\n2) a variant that combines a notion of disagreement with the predictive uncertainty (QB-MGP)\n\nThese methods are then compared to others on 8 synthetic datasets.", "main_review": "**Strengths**\n\n- the contributions of this paper are interesting and appear to be original\n- the approaches are based on sound probabilistic foundations\n- the experiments on synthetic data provide intuitive insights into the relative merits of different approaches.\n\n**Weaknesses**\n\n- connections and comparison to prior work is incomplete\n- experimental results are not very conclusive; no experiments on real data\n- parts of the paper are unnecessarily convoluted and unclear\n\nThe problem of overfitting to the marginal likelihood problem is well-known in the GP community. In the context of Bayesian optimization for example, some authors have proposed a similar method that takes advantage of a mixture of GPs [1]. How does this paper's approach differ?\n\nThe experimental results on synthetic data are quite comprehensive and insightful. One relevant baseline that appears to be missing is Zhao et al (2020). How does their approach compare empirically?\n\nIt is difficult to reconcile the results on the negative log-marginal likelihood (NLML) with the RMSE. The squared error is identical to the negative log-likelihood (under a Gaussian model) and should thus closely match the NLML results - how do you explain the difference in relative ranking and the fact that many of the RMSE curves slope upwards? How is the RMSE in fact evaluated—is there a hold-out set?\n\nAdditional experiments on real-world datasets & finding one realistic application where the proposed approach is clearly superior would significantly strengthen the paper.\n\nFinally, I encourage the authors to improve the clarity of the paper: introducing all concepts progressively, and making statements precise. Examples:\n\n- canonical Gaussian Process: to the best of my knowledge, this is not a widely used terminology for the RBF kernel\n- metamodeling is never formally introduced\n- \"given the data (x, y), ..., a GP is typically denoted as GP(...)\" -> the data does not appear to be relevant to the way the GP is denoted.\n- the definition of the RBF kernel is incorrect (there is a scalar quantity, the squared norm, that is divided by a vector)\n- I don't think the reference to the \"no free lunch\" theorem is appropriate in the context of the discussion on page 4\n- \"the inner integral is intractable and is therefore optimized with approximate inference\": what exactly is optimized?\n- section 5: how is convergence declared?\n\n[1]: An Empirical Bayes Approach to Optimizing Machine Learning Algorithms, J. McInerney, NIPS 2017", "summary_of_the_review": "I think there is potential in this paper: the problem is important and the method is sound. Improving the clarity of the paper, comparing against McInerney (2017) and Zhao et al. (2020) and applying the method on real-world data would make this paper significantly stronger.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635538333484}], "openreview_url": "https://openreview.net/forum?id=vyn49BUAkoD", "arxiv_id": "2205.10186", "paper_pdf": "papers/vyn49BUAkoD.pdf", "paper_pdf_sha256": "641da463c7ee1ac220fffa83cb7eda2a7866921cdc2845457129df78789db8d6", "paper_pdf_bytes": 910925, "paper_pdf_source": "openreview", "code_url": "https://github.com/CoRiis/active-learning-fbgp", "code_repository": "CoRiis/active-learning-fbgp", "code_commit": "3dbf09f2ec709981036f2ea7d018401ff9dac5be", "code_archive": "repos/vyn49BUAkoD.zip", "code_archive_sha256": "56da7e1247bb56f27d14fbea4adfe9e5cfafef70f818b69b65e632c0cd177a0b", "code_archive_bytes": 113716, "code_file_count": 26, "code_extensions": {".py": 25, ".sh": 1}, "github_disk_usage_kb": 110, "github_languages": {"Python": 90217, "Shell": 1735}, "github_archived": false, "github_pushed_at": "2022-09-29T20:45:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bayesian-active-learning-with-fully-bayesian-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TJSOfuZEd1B", "year": 2021, "status": "rejected", "title": "GeDi: Generative Discriminator Guided Sequence Generation", "authors": ["Ben Krause", "Akhilesh Deepak Gotmare", "Bryan McCann", "Nitish Shirish Keskar", "Shafiq Joty", "richard socher", "Nazneen Rajani"], "authorids": ["~Ben_Krause1", "~Akhilesh_Deepak_Gotmare1", "~Bryan_McCann1", "~Nitish_Shirish_Keskar1", "~Shafiq_Joty1", "~richard_socher1", "~Nazneen_Rajani1"], "authors_source": "OpenReview API", "abstract": "While large-scale language models (LMs) are able to imitate the distribution of natural language well enough to generate realistic text, it is difficult to control which regions of the distribution they generate. This is especially problematic because datasets used for training large LMs usually contain significant toxicity, hate, bias, and negativity. We propose GeDi as an efficient method for using smaller LMs as generative discriminators to guide generation from large LMs to make them safer and more controllable. GeDi guides generation at each step by computing classification probabilities for all possible next tokens via Bayes rule by normalizing over two class-conditional distributions; one conditioned on the desired attribute, or control code, and another conditioned on the undesired attribute, or anti control code. We find that GeDi gives controllability on par with or better than the state of the art method in a variety of settings, while also achieving generation speeds more than $30$ times faster. Additionally, training GeDi on only three topics allows us to controllably generate new topics zero-shot from just a keyword. Lastly, we show that GeDi can make GPT-2 and GPT-3 significantly less toxic without sacrificing on linguistic fluency, making it by far the most practical existing method for detoxifying large language models while maintaining a fast generation speed.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "Z51fEqNVNcQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2536/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\nThe paper considers the problem of attribute-based sequence generation, particularly in language models. Authors propose a framework “GeDi” which learns a generative classifier for controlling generation from a large language model. With experiments on publicly available datasets and models, and including human-evaluation, the authors empirically demonstrate that the algorithm is computationally efficient and is competitive against strong baseline algorithms like CTRL, Plug&Play language models (PPLM).\n\nReason for the score\nI vote for rejecting the current version of the paper (marginally below acceptance threshold). While the premise of the problem is well motivated, I think several sections of the paper are difficult to follow. I would strongly encourage the authors to include a pseudo code of the algorithm to improve the presentation of the central idea. The paper includes several experiments, though I think some critical ablations are missing.\n\nStrengths\n+ The problem is practically well motivated and is very relevant to the language learning community. \n+ The proposed algorithm of using generative classifiers is computationally efficient compared to strong baselines like CTRL and PPLM. The experimental results suggest that the algorithm also allows for better control over generation from a LM while maintaining linguistic quality.\n\nWeaknesses\n- Several sections of the paper are hard to follow. To improve the presentation of the idea, I would encourage the authors to distill the central idea into a pseudo code which goes along with Section 3.\n- The experiments on detoxification are critical to the thesis of the paper, however it seems that experiments in Section 5.2 consider only GPT-2 baselines? I think a strong baseline based on prior-work, like a CTRL generator conditioned on the positive label (as mentioned in Introduction), would help evaluating the gap between proposed approach and current algorithms.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Proposes algorithm for controllable sequence generation", "review": "Summary\nThe paper considers the problem of attribute-based sequence generation, particularly in language models. Authors propose a framework “GeDi” which learns a generative classifier for controlling generation from a large language model. With experiments on publicly available datasets and models, and including human-evaluation, the authors empirically demonstrate that the algorithm is computationally efficient and is competitive against strong baseline algorithms like CTRL, Plug&Play language models (PPLM).\n\nReason for the score\nI vote for rejecting the current version of the paper (marginally below acceptance threshold). While the premise of the problem is well motivated, I think several sections of the paper are difficult to follow. I would strongly encourage the authors to include a pseudo code of the algorithm to improve the presentation of the central idea. The paper includes several experiments, though I think some critical ablations are missing.\n\nStrengths\n+ The problem is practically well motivated and is very relevant to the language learning community. \n+ The proposed algorithm of using generative classifiers is computationally efficient compared to strong baselines like CTRL and PPLM. The experimental results suggest that the algorithm also allows for better control over generation from a LM while maintaining linguistic quality.\n\nWeaknesses\n- Several sections of the paper are hard to follow. To improve the presentation of the idea, I would encourage the authors to distill the central idea into a pseudo code which goes along with Section 3.\n- The experiments on detoxification are critical to the thesis of the paper, however it seems that experiments in Section 5.2 consider only GPT-2 baselines? I think a strong baseline based on prior-work, like a CTRL generator conditioned on the positive label (as mentioned in Introduction), would help evaluating the gap between proposed approach and current algorithms.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604775272302}, {"id": "fmdR4Q9luZZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2536/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposed a method —- GeDi — to generate guided and controlled text from a large language model (LM). The method utilizes smaller LMs as generative discriminators to guide generation from large LMs to make them safer and more controllable. By safer and controllable they emphasis on the toxicity, hate, bias, and negativity contains in the training of the large LM. The proposed method guides generation at each time step by computing classification probabilities for all possible next tokens via Bayes rule by normalizing over two class-conditional distributions (i.e. contrastive discrimination); one conditioned on the desired attribute, or control code, and another conditioned on the undesired attribute (i.e. contrastive attribute), or anti-control code. \n\nThe paper explores ways to increase generation speed and claimed that with the proposed techniques the generation speeds more than 30 times faster compared to PPLM model. The paper explores different heuristics to impose the guided generation including bias parameter, weighted decoding and filtering heuristics. The findings are that GeDi gives stronger controllability than the state of the art method (i.e.PPLM, CC-LM, CTRL).  \n\nExperiments show that, training GeDi on four topics (i.e. Business, Science/Tech, Sports, World ) allows the controlled generation of new topics zero-shot from just a keyword. They also demonstrate that GeDi can make GPT-2 (1.5B parameters) significantly less toxic without sacrificing linguistic quality.\n\n\nRe: “so long as the LM and GeDi share the same tokenization”: can you please elaborate the constraint on ‘same tokenization’?\n\nRe: “If the GeDi was trained on movie reviews for sentiment control, its direct class-conditional predictions will be biased towards predicting movie review words (illustrated by next word prediction of “cinematic”). However, by contrasting the predictions of opposing control codes via Bayes rule, the bias towards movie reviews can be cancelled out.”: The word cinematic can reveal a neutral/negative sentiment, is there any possibility that pushing the sentiment towards positive might degrade the accuracy of the overall generation?\n\nRe: GeDi training (λ < 1 in Equation (10)) and standard generative training(λ = 1 in Equation (10)). : How the value for λ = 0.6 was chosen? What is the impact of other values for this hyper-parameter? \n  \nRe: “In order to have prompts that are more likely to trigger aggressive generations but less likely to be explicitly toxic, we pass candidate prompts through a RoBERTa (Liu et al., 2019) model trained to classify toxicity, and only kept prompts where RoBERTa was less confident about the toxicity label.“: how did you measure model confidence about the toxicity label?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper is written well", "review": "The paper proposed a method —- GeDi — to generate guided and controlled text from a large language model (LM). The method utilizes smaller LMs as generative discriminators to guide generation from large LMs to make them safer and more controllable. By safer and controllable they emphasis on the toxicity, hate, bias, and negativity contains in the training of the large LM. The proposed method guides generation at each time step by computing classification probabilities for all possible next tokens via Bayes rule by normalizing over two class-conditional distributions (i.e. contrastive discrimination); one conditioned on the desired attribute, or control code, and another conditioned on the undesired attribute (i.e. contrastive attribute), or anti-control code. \n\nThe paper explores ways to increase generation speed and claimed that with the proposed techniques the generation speeds more than 30 times faster compared to PPLM model. The paper explores different heuristics to impose the guided generation including bias parameter, weighted decoding and filtering heuristics. The findings are that GeDi gives stronger controllability than the state of the art method (i.e.PPLM, CC-LM, CTRL).  \n\nExperiments show that, training GeDi on four topics (i.e. Business, Science/Tech, Sports, World ) allows the controlled generation of new topics zero-shot from just a keyword. They also demonstrate that GeDi can make GPT-2 (1.5B parameters) significantly less toxic without sacrificing linguistic quality.\n\n\nRe: “so long as the LM and GeDi share the same tokenization”: can you please elaborate the constraint on ‘same tokenization’?\n\nRe: “If the GeDi was trained on movie reviews for sentiment control, its direct class-conditional predictions will be biased towards predicting movie review words (illustrated by next word prediction of “cinematic”). However, by contrasting the predictions of opposing control codes via Bayes rule, the bias towards movie reviews can be cancelled out.”: The word cinematic can reveal a neutral/negative sentiment, is there any possibility that pushing the sentiment towards positive might degrade the accuracy of the overall generation?\n\nRe: GeDi training (λ < 1 in Equation (10)) and standard generative training(λ = 1 in Equation (10)). : How the value for λ = 0.6 was chosen? What is the impact of other values for this hyper-parameter? \n  \nRe: “In order to have prompts that are more likely to trigger aggressive generations but less likely to be explicitly toxic, we pass candidate prompts through a RoBERTa (Liu et al., 2019) model trained to classify toxicity, and only kept prompts where RoBERTa was less confident about the toxicity label.“: how did you measure model confidence about the toxicity label?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603980105279}, {"id": "giSNeP00gTD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2536/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Summary\n\nThe authors propose a method for controlling attributes of generated text (sentiment, topic, toxicity, etc) by reweighting a base language model's token-level distributions with auxiliary conditional language models, bayes rule, and additional heuristics.\n\nThe core idea is simple - which is a strength in my view - and does not require retraining the base language model, which could be important as language models become more expensive to train. However, the clarity and experiments in this paper fall short: the experimental setup has issues (detailed below), the effect on perplexity is quite large but relegated to the Appendix, several claims are speculative and lacking corresponding experimental evidence, and it is unclear how the additional heuristics affect performance.\n\nThe method seems promising, but with the current experiments it is difficult to draw conclusions about how the method affects performance and which parts of it are necessary; given that this is an empirical paper, I would therefore not recommend acceptance in its current form.\n\n#### Experimental setup\n\n- **Human evaluation**. There is little information given about how the human evaluation prompts are written (and why they are written that way), how many evaluators are used, and there are no significance tests (e.g. \"We run human evaluation to measure toxicity\" does not given enough details). This is concerning since all of the results in the main text use human evaluation, sometimes with small differences between methods.\n\n- **Automatic eval/perplexity**. The authors only measure perplexity in one set of experiments in the appendix (and it gets worse by introducing GeDI training). It would be good to have perplexity and automatic metrics to compare against the human evaluation (e.g. see Table 4 in the PPLM paper).\n\n- **Decoding algorithms**. The authors only show results for greedy decoding with a repetition penalty (with no ablation on the choice of penalty parameter). Results with a sampling method (e.g. nucleus) are needed for this open-ended setting, or an argument for why these aren't considered.\n\n#### Effect of the method\n- **Effect on perplexity**. The perplexity gets much worse as the gedi training is introduced (i.e. $\\lambda$ decreases), e.g. going from 25 to 45 on IMDb. This result is in the Appendix, and perplexity is never evaluated/reported in the other experiments.\n\n- **Gedi training**. It's unclear whether the gedi training (i.e. the $\\mathcal{L}_d$ loss) is beneficial: in some experiments $\\lambda=1.0$ performs similarly, and on the IMDb/MNLI/QNLI experiments decreasing $\\lambda$ either hurts, has no effect, or improves performance (i.e. no consistent trend).\n\n- **Detoxifying**. It's unclear how significant the results in Table 3 are, and there are no baselines; in general it it difficult to draw conclusions from these results.\n\n#### Speculative claims or conclusions\n- **Domain transfer**. Figure 1 gives an intuition for why domain transfer might be possible, but only an anecdote (first paragraph of 5.1, \"we noticed that\") and a single experiment is done, where the method performs similarly to PPLM. Crucially, the GEDI training does not appear to help over just re-weighting with the conditional LM ($\\lambda=1.0$ vs. $\\lambda=0.6$). Could the authors comment on this result? How well does domain transfer work for less similar domains? How is perplexity affected for the models reported in Table 2?\n\n- **Zero-shot control codes**. The authors only provide anecdotes for evaluating the Zero-shot control codes. Based on the evaluation it's quite speculative to say \"GeDi’s ability to generalize to new control codes zero-shot gives the ability to generate text corresponding to many topics and subtopics.\".\n\n- **Smaller language models guiding larger language models**. To be fair, the authors use GPT-2 medium as the conditional language model, and GPT-2 XL as the base language model, which is larger, but there was no investigation of this aspect of size difference. How small can the conditional LM be? Why was medium used instead of small? What if large was used? Does the conditional LM need to be a large-scale pretrained model (it would be nice to see a baseline of a simpler conditional LM)?\n\n#### Heuristics\n- Several heuristics are used: $\\alpha/T_i$ weighting, $\\omega$ weighting, nucleus filtering, keeping tokens over a threshold, repetition penalty, and rescaling the logits to positive (used in only one experiment).\n- How does each of these affect performance? There are no ablations, and given the small differences in some of the experiments it is unclear whether performance would actually be worse if we changed one of the heuristics. One outcome may be that the method only works for a careful balance of hyperparameters, which could be fine, but we don't have a sense of the variation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review: GeDi: Generative Discriminator Guided Sequence Generation", "review": "#### Summary\n\nThe authors propose a method for controlling attributes of generated text (sentiment, topic, toxicity, etc) by reweighting a base language model's token-level distributions with auxiliary conditional language models, bayes rule, and additional heuristics.\n\nThe core idea is simple - which is a strength in my view - and does not require retraining the base language model, which could be important as language models become more expensive to train. However, the clarity and experiments in this paper fall short: the experimental setup has issues (detailed below), the effect on perplexity is quite large but relegated to the Appendix, several claims are speculative and lacking corresponding experimental evidence, and it is unclear how the additional heuristics affect performance.\n\nThe method seems promising, but with the current experiments it is difficult to draw conclusions about how the method affects performance and which parts of it are necessary; given that this is an empirical paper, I would therefore not recommend acceptance in its current form.\n\n#### Experimental setup\n\n- **Human evaluation**. There is little information given about how the human evaluation prompts are written (and why they are written that way), how many evaluators are used, and there are no significance tests (e.g. \"We run human evaluation to measure toxicity\" does not given enough details). This is concerning since all of the results in the main text use human evaluation, sometimes with small differences between methods.\n\n- **Automatic eval/perplexity**. The authors only measure perplexity in one set of experiments in the appendix (and it gets worse by introducing GeDI training). It would be good to have perplexity and automatic metrics to compare against the human evaluation (e.g. see Table 4 in the PPLM paper).\n\n- **Decoding algorithms**. The authors only show results for greedy decoding with a repetition penalty (with no ablation on the choice of penalty parameter). Results with a sampling method (e.g. nucleus) are needed for this open-ended setting, or an argument for why these aren't considered.\n\n#### Effect of the method\n- **Effect on perplexity**. The perplexity gets much worse as the gedi training is introduced (i.e. $\\lambda$ decreases), e.g. going from 25 to 45 on IMDb. This result is in the Appendix, and perplexity is never evaluated/reported in the other experiments.\n\n- **Gedi training**. It's unclear whether the gedi training (i.e. the $\\mathcal{L}_d$ loss) is beneficial: in some experiments $\\lambda=1.0$ performs similarly, and on the IMDb/MNLI/QNLI experiments decreasing $\\lambda$ either hurts, has no effect, or improves performance (i.e. no consistent trend).\n\n- **Detoxifying**. It's unclear how significant the results in Table 3 are, and there are no baselines; in general it it difficult to draw conclusions from these results.\n\n#### Speculative claims or conclusions\n- **Domain transfer**. Figure 1 gives an intuition for why domain transfer might be possible, but only an anecdote (first paragraph of 5.1, \"we noticed that\") and a single experiment is done, where the method performs similarly to PPLM. Crucially, the GEDI training does not appear to help over just re-weighting with the conditional LM ($\\lambda=1.0$ vs. $\\lambda=0.6$). Could the authors comment on this result? How well does domain transfer work for less similar domains? How is perplexity affected for the models reported in Table 2?\n\n- **Zero-shot control codes**. The authors only provide anecdotes for evaluating the Zero-shot control codes. Based on the evaluation it's quite speculative to say \"GeDi’s ability to generalize to new control codes zero-shot gives the ability to generate text corresponding to many topics and subtopics.\".\n\n- **Smaller language models guiding larger language models**. To be fair, the authors use GPT-2 medium as the conditional language model, and GPT-2 XL as the base language model, which is larger, but there was no investigation of this aspect of size difference. How small can the conditional LM be? Why was medium used instead of small? What if large was used? Does the conditional LM need to be a large-scale pretrained model (it would be nice to see a baseline of a simpler conditional LM)?\n\n#### Heuristics\n- Several heuristics are used: $\\alpha/T_i$ weighting, $\\omega$ weighting, nucleus filtering, keeping tokens over a threshold, repetition penalty, and rescaling the logits to positive (used in only one experiment).\n- How does each of these affect performance? There are no ablations, and given the small differences in some of the experiments it is unclear whether performance would actually be worse if we changed one of the heuristics. One outcome may be that the method only works for a careful balance of hyperparameters, which could be fine, but we don't have a sense of the variation.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603896406178}, {"id": "54vRiyEa7kF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2536/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\nSummary:\nThis paper proposed using small-sized LM as a generative discriminator to guide large-sized LM for better controllability and decoding efficiency. \n \n##########################################################################\nReasons for score: \n \nMy score is marginally below the acceptance threshold. \n \nPros:\n\n1. The most exciting part of this paper is to factor out the opposite labels of each token (e.g., positive and negative, or toxic or non-toxic) using Bayes rule in generative models. \n\n \nCons: \n\n1. The contribution of controllability and producing safe output should be separated out. Safer LMs seem to be the outcome of the controllability of the LM by canceling out the opposite part of the target label using the Bayes rule. Highlighting this point might be helpful for showing out the novelty and value of this work. The current frame of the work seems quite distributed between applications and architectural contributions. \n \n2. The main concern of this work is the lack of focused contributions and their validations. The authors claim that this model is good at almost everything; efficiency, controllability while maintaining linguistic quality, reducing the toxicity of GPT2, zero-shot topical generation, etc. However, in fact, most of the experiments in Section 5 are very shallow, uncontrolled, and lack statistical significance. I appreciate the general effectiveness of the model and I don’t doubt it. However, as a conference paper with limited pages, it would be better to make one or two points among them and providing more in-depth with valid setups of experiments. I put additional notes about the experiments below. \n \n3. As mentioned above, the novelty of this comes from using the Bayes rule to make positive and negative labels far from each other. The authors also mention that this is a sort of contrastive learning, and they used contrastive generation. However, contrastive learning is often used to refer to learning by separating two different instances out far each other. In this work, there is no such auxiliary optimization during training time, but the posteriors are re-weighted using the Bayes rule. I would recommend using a more exact term to describe this rather than contrastive generation. \n \n4. In Section 3.1.1, various heuristics are used. I expected to see the effect or ablation of each heuristic and how important each of them is in terms of generation quality. Also, the baseline models such as CTRL, CC-LM, and PPLM seem to be not using the same heuristics, which seems to be not fair. \n \n5. The output in Table 6 makes me doubt how the experiments are badly controlled. The outputs from positive and negative sentiment are totally different and almost random text, meaning that the content of the generators is not controlled properly. In preparation for prompts for GeDis (5.2) or for measurement of label fidelity (5.3), authors used the pre-trained BERT or RoBERTa on the target attribute like toxicity and topics. As far as I know, these automatic classifiers are not correlated with human judgment, in fact, leading to huge wrongly-predicted labels. I wonder why human annotations are not used here. \n\n6. In Table 2, I don’t see any significant improvements of GeDi against PPLM in its attribution score (i.e., positivity) and transferability to the target domain. Similar to the comment above, none of the experiments are controlled in content. Measuring how the text is similar to the domain (e.g., book-like) sounds interesting but there are no further details of how the human evaluation is studied, what kinds of guidelines are provided to annotators, how the output looks like, etc. \n\n7. In Table 3 and 4, have you performed the same experiments with PPLM and CTRL? \n \n##########################################################################\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A novel approach but limited experiments with a lack of valid experimental setups. ", "review": "##########################################################################\nSummary:\nThis paper proposed using small-sized LM as a generative discriminator to guide large-sized LM for better controllability and decoding efficiency. \n \n##########################################################################\nReasons for score: \n \nMy score is marginally below the acceptance threshold. \n \nPros:\n\n1. The most exciting part of this paper is to factor out the opposite labels of each token (e.g., positive and negative, or toxic or non-toxic) using Bayes rule in generative models. \n\n \nCons: \n\n1. The contribution of controllability and producing safe output should be separated out. Safer LMs seem to be the outcome of the controllability of the LM by canceling out the opposite part of the target label using the Bayes rule. Highlighting this point might be helpful for showing out the novelty and value of this work. The current frame of the work seems quite distributed between applications and architectural contributions. \n \n2. The main concern of this work is the lack of focused contributions and their validations. The authors claim that this model is good at almost everything; efficiency, controllability while maintaining linguistic quality, reducing the toxicity of GPT2, zero-shot topical generation, etc. However, in fact, most of the experiments in Section 5 are very shallow, uncontrolled, and lack statistical significance. I appreciate the general effectiveness of the model and I don’t doubt it. However, as a conference paper with limited pages, it would be better to make one or two points among them and providing more in-depth with valid setups of experiments. I put additional notes about the experiments below. \n \n3. As mentioned above, the novelty of this comes from using the Bayes rule to make positive and negative labels far from each other. The authors also mention that this is a sort of contrastive learning, and they used contrastive generation. However, contrastive learning is often used to refer to learning by separating two different instances out far each other. In this work, there is no such auxiliary optimization during training time, but the posteriors are re-weighted using the Bayes rule. I would recommend using a more exact term to describe this rather than contrastive generation. \n \n4. In Section 3.1.1, various heuristics are used. I expected to see the effect or ablation of each heuristic and how important each of them is in terms of generation quality. Also, the baseline models such as CTRL, CC-LM, and PPLM seem to be not using the same heuristics, which seems to be not fair. \n \n5. The output in Table 6 makes me doubt how the experiments are badly controlled. The outputs from positive and negative sentiment are totally different and almost random text, meaning that the content of the generators is not controlled properly. In preparation for prompts for GeDis (5.2) or for measurement of label fidelity (5.3), authors used the pre-trained BERT or RoBERTa on the target attribute like toxicity and topics. As far as I know, these automatic classifiers are not correlated with human judgment, in fact, leading to huge wrongly-predicted labels. I wonder why human annotations are not used here. \n\n6. In Table 2, I don’t see any significant improvements of GeDi against PPLM in its attribution score (i.e., positivity) and transferability to the target domain. Similar to the comment above, none of the experiments are controlled in content. Measuring how the text is similar to the domain (e.g., book-like) sounds interesting but there are no further details of how the human evaluation is studied, what kinds of guidelines are provided to annotators, how the output looks like, etc. \n\n7. In Table 3 and 4, have you performed the same experiments with PPLM and CTRL? \n \n##########################################################################\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603675807363}, {"id": "OdilMghmAcE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2536/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "[Summary]\nIn this paper, the authors propose an efficient method for controllable language generation of large pre-trained LMs (e.g., GPT2). The main idea is to use a smaller, compared to the LM to control, language model trained with control code (Keskar et al., 2019) to generate a per-token score $P(c|x_{1:t}$) to steer the original Language Model distribution. The author proposes two ways, contrastive and discriminative, to approximate $P(c|x_{1:t})$ using bayesian rules. Experiments on open-ended language generation have been shown for positive/negative, detoxification, and topic-control, including a zero-shot topic-control. \n\n[Pros]\n- the proposed methodology is novel for the task and it is effective in controlling the desired attributes. \n- the proposed method is more efficient than WD (Ghazvininejad et al., 2017) since it does not require a forward to the discriminator for each to token in the vocabulary, and computationally more efficient then PPLM (Dathathri et al. 2020) which requires several updates per token. \n- the paper is well-written and easy to follow, except for some minor (later for more info). To the best of my knowledge, the paper is technically correct and reproducible.\n\n[Cons/Question for the authors]\n- I have read through the paper, but I could not find any significant test (e.g., annotator agreement, t-test etc.), are the reported human evaluation results significant? could you provide p-values for the results? \n- Both detoxification and topic control has no baselines to compare with. For instance, CC-LM-non-tox, CTRL, PPLM could have been used to detoxify the generation. For detox, why not using Universal Triggers (Eric et.al. 2019) for making the model generate toxic text, instead of using prompt from the dev set of the same dataset used for training GEDI? \n- Zero-Shot Topics: why PPLM and CTRL cannot do zero-shot on a topic (from the conclusion)? PPLM can use a bag-of-word discriminator, so no training required and generate any kind of topics, and CTRL can use different link/prompt to generate unseen topics?\n- Greedy decoding: why using greedy decoding for a language generation task? it is well known that top-p and top-k greatly improve the model generation, are there performance drop if using top-p? how GEDI compare to CTRL, PPLM  in this setting? \n\n[Reason to accept]\nThe proposed method is a simple and effective way to control the generation of large language models. This is an important and timely problem, especially for language detoxification.\n\n[Reason to reject]\nThe experiments are a bit unclear, looking forward to the author response\n\n[Suggestions and some more questions]\n- With reference to the sentence: \" In addition to class-conditional generation, CC-LMs can be used as generative classifiers by applying Bayes rule to compute $P(c|x_{1:T})$, as is done by Keskar et al. (2019) for source attribution.\" Could you please add the inline formula, $p_θ(c|x) \\approx p_θ(x|c)p(c)$, it saves one jump to the paper and makes the paper more readable :)\n- Could you please elaborate on why GEDI would be 10k fold less computation as compared with a unidirectional classifier? Could you include a more detailed computational cost analysis?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "[Summary]\nIn this paper, the authors propose an efficient method for controllable language generation of large pre-trained LMs (e.g., GPT2). The main idea is to use a smaller, compared to the LM to control, language model trained with control code (Keskar et al., 2019) to generate a per-token score $P(c|x_{1:t}$) to steer the original Language Model distribution. The author proposes two ways, contrastive and discriminative, to approximate $P(c|x_{1:t})$ using bayesian rules. Experiments on open-ended language generation have been shown for positive/negative, detoxification, and topic-control, including a zero-shot topic-control. \n\n[Pros]\n- the proposed methodology is novel for the task and it is effective in controlling the desired attributes. \n- the proposed method is more efficient than WD (Ghazvininejad et al., 2017) since it does not require a forward to the discriminator for each to token in the vocabulary, and computationally more efficient then PPLM (Dathathri et al. 2020) which requires several updates per token. \n- the paper is well-written and easy to follow, except for some minor (later for more info). To the best of my knowledge, the paper is technically correct and reproducible.\n\n[Cons/Question for the authors]\n- I have read through the paper, but I could not find any significant test (e.g., annotator agreement, t-test etc.), are the reported human evaluation results significant? could you provide p-values for the results? \n- Both detoxification and topic control has no baselines to compare with. For instance, CC-LM-non-tox, CTRL, PPLM could have been used to detoxify the generation. For detox, why not using Universal Triggers (Eric et.al. 2019) for making the model generate toxic text, instead of using prompt from the dev set of the same dataset used for training GEDI? \n- Zero-Shot Topics: why PPLM and CTRL cannot do zero-shot on a topic (from the conclusion)? PPLM can use a bag-of-word discriminator, so no training required and generate any kind of topics, and CTRL can use different link/prompt to generate unseen topics?\n- Greedy decoding: why using greedy decoding for a language generation task? it is well known that top-p and top-k greatly improve the model generation, are there performance drop if using top-p? how GEDI compare to CTRL, PPLM  in this setting? \n\n[Reason to accept]\nThe proposed method is a simple and effective way to control the generation of large language models. This is an important and timely problem, especially for language detoxification.\n\n[Reason to reject]\nThe experiments are a bit unclear, looking forward to the author response\n\n[Suggestions and some more questions]\n- With reference to the sentence: \" In addition to class-conditional generation, CC-LMs can be used as generative classifiers by applying Bayes rule to compute $P(c|x_{1:T})$, as is done by Keskar et al. (2019) for source attribution.\" Could you please add the inline formula, $p_θ(c|x) \\approx p_θ(x|c)p(c)$, it saves one jump to the paper and makes the paper more readable :)\n- Could you please elaborate on why GEDI would be 10k fold less computation as compared with a unidirectional classifier? Could you include a more detailed computational cost analysis?\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603598131317}], "openreview_url": "https://openreview.net/forum?id=TJSOfuZEd1B", "arxiv_id": "2009.06367", "paper_pdf": "papers/TJSOfuZEd1B.pdf", "paper_pdf_sha256": "f576582972d25ff322637851c708384dde33ff02b57dcfc28b8156c0d6299c7b", "paper_pdf_bytes": 1190201, "paper_pdf_source": "openreview", "code_url": "https://github.com/salesforce/GeDi", "code_repository": "salesforce/GeDi", "code_commit": "2346c7ee99cd9b2f434b088b98803a90e370ca49", "code_archive": "repos/TJSOfuZEd1B.zip", "code_archive_sha256": "b29f9f4e251b421afe9d00f2ccc4cb0b418ccdf025fccbe15f948aa1180a27b0", "code_archive_bytes": 119113, "code_file_count": 11, "code_extensions": {".py": 5, ".sh": 5, ".ipynb": 1}, "github_disk_usage_kb": 150, "github_languages": {"Python": 178636, "Jupyter Notebook": 10407, "Shell": 2192}, "github_archived": true, "github_pushed_at": "2025-06-16T14:46:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gedi-generative-discriminator-guided-sequence"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJeHwJSYvH", "year": 2020, "status": "rejected", "title": "Learning De-biased Representations with Biased Representations", "authors": ["Hyojin Bahng", "Sanghyuk Chun", "Sangdoo Yun", "Jaegul Choo", "Seong Joon Oh"], "authorids": ["hjj552@korea.ac.kr", "sanghyuk.c@navercorp.com", "sangdoo.yun@navercorp.com", "jchoo@korea.ac.kr", "coallaoh@linecorp.com"], "authors_source": "OpenReview API", "abstract": "Many machine learning algorithms are trained and evaluated by splitting data from a single source into training and test sets. While such focus on in-distribution learning scenarios has led interesting advances, it has not been able to tell if models are relying on dataset biases as shortcuts for successful prediction (e.g., using snow cues for recognising snowmobiles). Such biased models fail to generalise when the bias shifts to a different class. The cross-bias generalisation problem has been addressed by de-biasing training data through augmentation or re-sampling, which are often prohibitive due to the data collection cost (e.g., collecting images of snowmobile on a desert) and the difficulty of quantifying or expressing biases in the first place. In this work, we propose a novel framework to train a de-biased representation by encouraging it to be different from a set of representations that are biased by design. This tactic is feasible in many scenarios where it is much easier to define a set of biased representations than to define and quantify bias. Our experiments and analyses show that our method discourages models from taking bias shortcuts, resulting in improved performances on de-biased test data.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1gpPXAK5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1765/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This manuscript discusses the problem of bias shortcut employed by many machine learning algorithms (due to dataset problems or underlying effects of any algorithmic bias within an application). The authors argue that models tend to underutilize their capacities to extract non-bias signals when bias shortcuts provide enough cues for recognition. This is an interesting and important aspect of machine learning models neglected by many recent developments. \n\nThe only problem is that the paper seems to be a bit immature as the exemplar application is too naive for illustrating the idea. The authors’ idea is to assume that\n‘there is a family of feature extractors, such that features learned by any of these extractors would correspond to pure bias. Then in order to learn unbiased features, the goal is to construct a feature extractor that is ''as different as this family\" as possible’\nin practice, their claim is texture and color are biases; one should learn shape instead of texture and color. So the family of feature extractors are the ones with small receptive field that can only capture texture and color. Therefore, what they eventually achieved is the unbiased feature extractor only learns the shape of object and avoids learning any texture and color. \n\nSo, the problem is, in practice, it is very hard to define the family of biased feature extractors. It really depends on the dataset and the goal. Texture and color, in general, are still important cues for object recognition, removing this information is NOT equivalent to removing bias. Just as a suggestion, the background scene might be a better definition of bias. However, with the proposal in this paper, it would be unclear how to define the family of feature extractors for describing background. Therefore, the solution given for this important problem seems to be too ad-hoc and not generalizable. \n\nThe second example (that does not have an experiments on) is action recognition; the family of biased feature extractors is 2D-frame-wise CNNs (object recognition). The authors claim that objects are biases for action recognition systems, but again a large part of action recognition is indeed object recognition. Many actions are defined based interaction of humans with objects (e.g., opening bottle or pouring water from bottle). Some objects may be instroducing bias in the task, but not all. Again, the proposed solution in this paper cannot disentangle this. \n\nThe authors need to survey previous texture-shape disentanglement works and then compare with those methods.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This manuscript discusses the problem of bias shortcut employed by many machine learning algorithms (due to dataset problems or underlying effects of any algorithmic bias within an application). The authors argue that models tend to underutilize their capacities to extract non-bias signals when bias shortcuts provide enough cues for recognition. This is an interesting and important aspect of machine learning models neglected by many recent developments. \n\nThe only problem is that the paper seems to be a bit immature as the exemplar application is too naive for illustrating the idea. The authors’ idea is to assume that\n‘there is a family of feature extractors, such that features learned by any of these extractors would correspond to pure bias. Then in order to learn unbiased features, the goal is to construct a feature extractor that is ''as different as this family\" as possible’\nin practice, their claim is texture and color are biases; one should learn shape instead of texture and color. So the family of feature extractors are the ones with small receptive field that can only capture texture and color. Therefore, what they eventually achieved is the unbiased feature extractor only learns the shape of object and avoids learning any texture and color. \n\nSo, the problem is, in practice, it is very hard to define the family of biased feature extractors. It really depends on the dataset and the goal. Texture and color, in general, are still important cues for object recognition, removing this information is NOT equivalent to removing bias. Just as a suggestion, the background scene might be a better definition of bias. However, with the proposal in this paper, it would be unclear how to define the family of feature extractors for describing background. Therefore, the solution given for this important problem seems to be too ad-hoc and not generalizable. \n\nThe second example (that does not have an experiments on) is action recognition; the family of biased feature extractors is 2D-frame-wise CNNs (object recognition). The authors claim that objects are biases for action recognition systems, but again a large part of action recognition is indeed object recognition. Many actions are defined based interaction of humans with objects (e.g., opening bottle or pouring water from bottle). Some objects may be instroducing bias in the task, but not all. Again, the proposed solution in this paper cannot disentangle this. \n\nThe authors need to survey previous texture-shape disentanglement works and then compare with those methods.\n\n\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572623205448}, {"id": "ByewI_D15B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1765/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes a methodology for reducing model dependance on bias by specifying a model family of biases (i.e. conv nets with only 1x1 convs to model color biases), and then forcing independence between feature representations of the bias model and the a full model (i.e. conv nets with 3x3 convs to also model edges). \n\nOverall the method is very interesting. A few very related recent works were missing: https://arxiv.org/pdf/1908.10763.pdf and https://arxiv.org/abs/1909.03683. These works provide different formulations for factoring out know \"bias oriented\" models (and focus more on NLP, although the second is applied to VQA). Although, I do appreciate that the bias models used in this work are slightly more general in terms of family than those studied before, they do encode significant intuition about the target bias to be removed and perhaps in this context, the methods cited above could also be compared. That being said, I don't feel the paper needs to provide this comparison given how recent those works are, but it would improve the quality of the paper. \n\nOne aspect that worries me about this paper is that most of the biases studied are synthetic, so specification of the bias family is trivial, in contrast to the works I mentioned above (where the bias model is potentially somewhat misspecified and needed to be discovered by different researchers). But I do really like the experiments on Imagenet-a. \n\nOverall the paper presents an interesting contribution that would be useful for future study in reducing bias dependance in ml.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper describes a methodology for reducing model dependance on bias by specifying a model family of biases (i.e. conv nets with only 1x1 convs to model color biases), and then forcing independence between feature representations of the bias model and the a full model (i.e. conv nets with 3x3 convs to also model edges). \n\nOverall the method is very interesting. A few very related recent works were missing: https://arxiv.org/pdf/1908.10763.pdf and https://arxiv.org/abs/1909.03683. These works provide different formulations for factoring out know \"bias oriented\" models (and focus more on NLP, although the second is applied to VQA). Although, I do appreciate that the bias models used in this work are slightly more general in terms of family than those studied before, they do encode significant intuition about the target bias to be removed and perhaps in this context, the methods cited above could also be compared. That being said, I don't feel the paper needs to provide this comparison given how recent those works are, but it would improve the quality of the paper. \n\nOne aspect that worries me about this paper is that most of the biases studied are synthetic, so specification of the bias family is trivial, in contrast to the works I mentioned above (where the bias model is potentially somewhat misspecified and needed to be discovered by different researchers). But I do really like the experiments on Imagenet-a. \n\nOverall the paper presents an interesting contribution that would be useful for future study in reducing bias dependance in ml.\n\n"}, "tcdate": 1571940430714}, {"id": "S1g5ZEg6KB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1765/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "###  Summary \n\nThe paper proposes a method for regularizing neural networks to mitigate certain known biases from the representations learned by CNNs. The authors look at the setting in which the distribution of biases in the train and test set remains the same, but the distribution of targets given biases changes.\n\nTo remove the bias, they propose hand-designing a set of models that rely exclusively on these biases to make predictions. They then train a new unbiased model by making sure that this new model uses sufficiently different information than the biased models to make predictions (By adding a regularization term)\n\n### Decision and reasons\n\nI vote for a weak accept. \n\nStrengths: \n1: The paper is well written with a clearly defined problem-setting. The proposed method is sound and interesting, and the empirical results, thorough. \n\nWeaknesses:\n2: The solution just pushes the problem of 'learning a model that only uses the true signal to make predictions' to 'learning a set of models that only use the noise to make predictions.' It's not clear why the latter is easier than the former. \n\n3: The paper does not have any baselines that directly try to remove the bias (Instead of using the two-step process). As a result, it's hard to judge how meaningful the improvements are. \n\n\n### Supporting arguments for the reasons for the decision.\n\nStrengths: \n1: The paper does a really good job of defining the problem-setting. Contrasting cross-bias with cross-domain and in-distribution makes the goal of the paper very clear. The notation used to formalize the problem setting in Section 2.1 is also clear and concise. Moreover, the experiments on the toy dataset help clarify the proposed solution whereas experiments on Biased MNIST and Imagenet show that it successfully mitigates the bias. Finally, the authors show the importance of each component of the proposed solution by factor analysis.\n\nWeaknesses:\n2: In the most general case, it is not obvious why it is easier to define and learn a set of models that only use noise to make predictions (which is the required first step for their proposed solution) as opposed to learning a model that only uses signal (which is the goal of the problem). These two problems seem equally hard. The paper builds on the premise that in some cases former is easier (i.e. in some cases, it is easier to learn a set of models that use only noise as opposed to learning a debiased model directly). The authors give two such examples (They only explore the first experimentally.) \n1. Learning a model that relies on local texture. They achieve this by limiting the receptive field of the features. \n2. Learning a model that relies on static images to make predictions about actions in videos. \nI feel that the two given examples are very narrow. It would be nice if the authors could identify a broader class of problems for which G is given or can easily be defined. Moreover, even in these two examples, I'm not convinced that the proposed biased models only use B for making predictions. For example, for some classification tasks, the local texture could be part of the signal and not just the bias. Similarly, static images from a video do contain important information for making the prediction. \n\n3: The authors only compare their method to a baseline that does nothing to debias the representations. Even though this is an important comparison (as it shows that the proposed method can debias the representations), it does not tell the reader how effective the proposed method is compared to other possible solutions. The results would be more meaningful if the authors could include at-least a simple baseline that tries to remove the bias in other ways (For example, they could use the style-transfer baseline used by Geirhos et al., 2019). \n\nI vote for accepting the paper as a poster. It introduces an interesting approach for debiasing representations. However, due to its narrow scope and missing baselines, I would not recommend the paper for an oral presentation. \n\n\n### Questions\n\n1. What are some broader class of problems for which defining G is easier than directly regularizing for debiased representations?\n\n2. How well do Bagnets alone perform on the benchmarks in Table 3? I would expect to see that Bagnets alone do worse than vanilla Resnets on Unbiased and IN-A. Is that so? \n\n3. How well do other methods do in these domains? (Such as methods that directly debias the training data against texture by applying style-transfer). \n\n\n### Update after Author's response\n\nThe author's response has clarified the motivation behind the proposed approach to an extent. They have also added a comparison with a method that directly promotes learning the shape as opposed to the texture ( by training on stylized Imagenet) \n\nI agree with R1 on all accounts (i.e. it is very hard to define the family of biased feature extractors, the proposed approach is ad-hoc, the authors need to compare to texture-shape disentanglement methods, etc), however at the same time, I can see that proposed approach can act as a useful heuristic for regularizing neural networks to pay attention to certain kind of information. \n\nAn interesting use-case of the proposed method (which the authors indirectly mentioned in their response to my review) is in a multi-modal setting. It's not trivial to enforce deep learning systems to utilize all data modalities in a multi-modal setting. By defining G to be models trained on individual modalities, it would be possible to nudge our models to pay attention to the information in all modalities. \n\nSome important results in deep learning have been ad-hoc (For example skip connections in deep networks, ReLUs) and have nonetheless progressed the field. This work is not as widely applicable as skip connections or ReLUs, but it is, nonetheless, providing a heuristic for solving an important problem. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "###  Summary \n\nThe paper proposes a method for regularizing neural networks to mitigate certain known biases from the representations learned by CNNs. The authors look at the setting in which the distribution of biases in the train and test set remains the same, but the distribution of targets given biases changes.\n\nTo remove the bias, they propose hand-designing a set of models that rely exclusively on these biases to make predictions. They then train a new unbiased model by making sure that this new model uses sufficiently different information than the biased models to make predictions (By adding a regularization term)\n\n### Decision and reasons\n\nI vote for a weak accept. \n\nStrengths: \n1: The paper is well written with a clearly defined problem-setting. The proposed method is sound and interesting, and the empirical results, thorough. \n\nWeaknesses:\n2: The solution just pushes the problem of 'learning a model that only uses the true signal to make predictions' to 'learning a set of models that only use the noise to make predictions.' It's not clear why the latter is easier than the former. \n\n3: The paper does not have any baselines that directly try to remove the bias (Instead of using the two-step process). As a result, it's hard to judge how meaningful the improvements are. \n\n\n### Supporting arguments for the reasons for the decision.\n\nStrengths: \n1: The paper does a really good job of defining the problem-setting. Contrasting cross-bias with cross-domain and in-distribution makes the goal of the paper very clear. The notation used to formalize the problem setting in Section 2.1 is also clear and concise. Moreover, the experiments on the toy dataset help clarify the proposed solution whereas experiments on Biased MNIST and Imagenet show that it successfully mitigates the bias. Finally, the authors show the importance of each component of the proposed solution by factor analysis.\n\nWeaknesses:\n2: In the most general case, it is not obvious why it is easier to define and learn a set of models that only use noise to make predictions (which is the required first step for their proposed solution) as opposed to learning a model that only uses signal (which is the goal of the problem). These two problems seem equally hard. The paper builds on the premise that in some cases former is easier (i.e. in some cases, it is easier to learn a set of models that use only noise as opposed to learning a debiased model directly). The authors give two such examples (They only explore the first experimentally.) \n1. Learning a model that relies on local texture. They achieve this by limiting the receptive field of the features. \n2. Learning a model that relies on static images to make predictions about actions in videos. \nI feel that the two given examples are very narrow. It would be nice if the authors could identify a broader class of problems for which G is given or can easily be defined. Moreover, even in these two examples, I'm not convinced that the proposed biased models only use B for making predictions. For example, for some classification tasks, the local texture could be part of the signal and not just the bias. Similarly, static images from a video do contain important information for making the prediction. \n\n3: The authors only compare their method to a baseline that does nothing to debias the representations. Even though this is an important comparison (as it shows that the proposed method can debias the representations), it does not tell the reader how effective the proposed method is compared to other possible solutions. The results would be more meaningful if the authors could include at-least a simple baseline that tries to remove the bias in other ways (For example, they could use the style-transfer baseline used by Geirhos et al., 2019). \n\nI vote for accepting the paper as a poster. It introduces an interesting approach for debiasing representations. However, due to its narrow scope and missing baselines, I would not recommend the paper for an oral presentation. \n\n\n### Questions\n\n1. What are some broader class of problems for which defining G is easier than directly regularizing for debiased representations?\n\n2. How well do Bagnets alone perform on the benchmarks in Table 3? I would expect to see that Bagnets alone do worse than vanilla Resnets on Unbiased and IN-A. Is that so? \n\n3. How well do other methods do in these domains? (Such as methods that directly debias the training data against texture by applying style-transfer). \n\n\n### Update after Author's response\n\nThe author's response has clarified the motivation behind the proposed approach to an extent. They have also added a comparison with a method that directly promotes learning the shape as opposed to the texture ( by training on stylized Imagenet) \n\nI agree with R1 on all accounts (i.e. it is very hard to define the family of biased feature extractors, the proposed approach is ad-hoc, the authors need to compare to texture-shape disentanglement methods, etc), however at the same time, I can see that proposed approach can act as a useful heuristic for regularizing neural networks to pay attention to certain kind of information. \n\nAn interesting use-case of the proposed method (which the authors indirectly mentioned in their response to my review) is in a multi-modal setting. It's not trivial to enforce deep learning systems to utilize all data modalities in a multi-modal setting. By defining G to be models trained on individual modalities, it would be possible to nudge our models to pay attention to the information in all modalities. \n\nSome important results in deep learning have been ad-hoc (For example skip connections in deep networks, ReLUs) and have nonetheless progressed the field. This work is not as widely applicable as skip connections or ReLUs, but it is, nonetheless, providing a heuristic for solving an important problem. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1571779585879}], "openreview_url": "https://openreview.net/forum?id=SJeHwJSYvH", "arxiv_id": "1910.02806", "paper_pdf": "papers/SJeHwJSYvH.pdf", "paper_pdf_sha256": "2e49b9447992376b1fd2a1ca431054322c0e8d3bd6b03ca647d7152841183470", "paper_pdf_bytes": 5120080, "paper_pdf_source": "openreview", "code_url": "https://github.com/clovaai/rebias", "code_repository": "clovaai/rebias", "code_commit": "79d1d1d892412a5e3b07fbea8b776461bed0e096", "code_archive": "repos/SJeHwJSYvH.zip", "code_archive_sha256": "4b4ebd6391c239dce9907b2403a0d2791a651e27bcf062a1dddb8877963ab1c4", "code_archive_bytes": 171862, "code_file_count": 36, "code_extensions": {".py": 36}, "github_disk_usage_kb": 171, "github_languages": {"Python": 208634}, "github_archived": false, "github_pushed_at": "2024-07-25T10:17:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-de-biased-representations-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Sy4lojC9tm", "year": 2019, "status": "rejected", "title": "Dataset Distillation", "authors": ["Tongzhou Wang", "Jun-Yan Zhu", "Antonio Torralba", "Alexei A. Efros"], "authorids": ["tongzhou.wang.1994@gmail.com", "junyanz@mit.edu", "torralba@mit.edu", "efros@eecs.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Model distillation aims to distill the knowledge of a complex model into a simpler one.  In this paper, we consider an alternative formulation called {\\em dataset distillation}: we keep the model fixed and instead attempt to distill the knowledge from a large training dataset into a small one.  The idea is to {\\em synthesize} a small number of data points that do not need to come from the correct data distribution, but will, when given to the learning algorithm as training data, approximate the model trained on the original data.  For example, we show that it is possible to compress $60,000$ MNIST training images into just $10$ synthetic {\\em distilled images} (one per class) and achieve close to original performance with only a few steps of gradient descent, given a particular fixed network initialization. We evaluate our method in a wide range of initialization settings and with different learning objectives.  Experiments on multiple datasets show the advantage of our approach compared to alternative methods in most settings. ", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1gcr_G5hm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper593/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an algorithm for compressing the size of entire training data into a few synthetic training samples. The method is based on neural networks and is applied on image datsets. The authors comment two possible applications of their method domain adaptation and effective data poisoning attack.\n\nThe proposed technique seems to be limited to neural networks since it seems that is linked to the initialization of the networks. In this aspect, it could be interesting to have a more general method.\n\nThere are related works that are not commented, for instance :\n\nOlvera-López, J. Arturo, et al. \"A review of instance selection methods.\" Artificial Intelligence Review 34.2 (2010): 133-143.\n\n\nExperimental section is weak. Few datasets are considered, other problems should be added. Additionally, related methods should be included to  compare the performance of the proposal. Some comments about the computational cost should be inserted. In this aspect, the experimental section should be improved following these recommendations.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An algorithm to reduce datatset size for  NNs", "review": "The paper presents an algorithm for compressing the size of entire training data into a few synthetic training samples. The method is based on neural networks and is applied on image datsets. The authors comment two possible applications of their method domain adaptation and effective data poisoning attack.\n\nThe proposed technique seems to be limited to neural networks since it seems that is linked to the initialization of the networks. In this aspect, it could be interesting to have a more general method.\n\nThere are related works that are not commented, for instance :\n\nOlvera-López, J. Arturo, et al. \"A review of instance selection methods.\" Artificial Intelligence Review 34.2 (2010): 133-143.\n\n\nExperimental section is weak. Few datasets are considered, other problems should be added. Additionally, related methods should be included to  compare the performance of the proposal. Some comments about the computational cost should be inserted. In this aspect, the experimental section should be improved following these recommendations.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541183554133}, {"id": "S1xVvuOLh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper593/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n=== Post rebuttal update ===\n\nThanks to the addition of better baselines, I've increased my score for this paper. While I'm still not super convinced of its potential for application, I find the idea original and worth discussing at the conference.\n\n=== Pre-rebuttal review ===\nThis paper presents an approach to compress a dataset into a much smaller number of synthetic samples that are optimized to yield as good performance as possible when a given model is trained on that smaller dataset. This is done by unrolling the gradient descent procedure of training such a model to allow for gradient-based optimization of synthetic samples themselves as well as the used learning rates.\n\nIn summary, my evaluation is as follow:\n\n*Pros*\n- Pretty original problem formulation\n- Generally well written paper\n\n*Cons*\n- Lack of comparison with simple baselines in basic dataset distillation setting\n- Use in practical applications (domain adaptation, data poisoning) yet to be convincingly demonstrated\n- Possibly a mistake in the theoretical analysis of the linear case\n\nIndeed, I found the paper to be generally quite clear and enjoyed reading it. One minor thing I struggled a bit with is the distinction between \"SG steps\" and \"Epochs\" (I believe the former corresponds to when the synthetic samples are different between GD steps, whereas the later corresponds to the number of times the method repeatedly cycles over these samples) so I would perhaps encourage the authors to emphasize that difference. \n\nI also find the problem statement that proposed to be interesting and thought provoking, and the solution that's proposed seems quite appropriate and well thought out.\n\nThat said, I'm worried about the following:\n\n- Unless I misunderstood, in the basic dataset distillation setting a comparison is never provided with training on a randomly selected subset of the training set. Presumably the results are worse, but I think these results should be in the paper. I would also argue for having another baseline, which would try to (approximately) optimize the choice of which training examples are put in the subset. A very simple approach would be to take the 200 runs already performed for the random selection and select the subset providing the best accuracy on the full training set and only report the performance of that subset (instead of the mean and std of all 200 runs). In short, this would help determine to what extent there is value to synthesizing entirely new samples. Moreover, I think a simple alternative baseline for creating synthesized samples should be considered. Specifically, I'd personally would like to know the performance of using per-class k-means clustering and training on the cluster centroids as the distilled dataset.\n\n- While I appreciate that the authors identify potential applications and report some results on them, I think they currently fall short of convincing the reader of the potential of dataset distillation for these applications. For domain adaptation, no actual domain adaptation baseline is compared against (a good candidate would be method from Daume III 2007, at the very least). For data poisoning, I find that the assumptions for attacks are pretty strong, i.e. that a) you have access to the parameters of the pre-trained models to attack and b) that the model is doing additional updates *only* on the synthesized data. If there are reasons to think that such assumptions are reasonable, I'd at least expect the paper to motivate why that is. \n\n- In the analysis of the simple linear case, in Equation 7, there appears to be a mistake, specifically some missing parentheses:\n\nd^Td( (I-\\eta/M d^Td)\\theta_0 + \\eta/M \\tilde{d}^T\\tilde{t}) = d^T t\n\ni.e. there should be parentheses right after \"d^Td\" and right before \"=\". This is from replacing \\theta^* by the expression for \\theta_1 in Equation 6, which is what I think Equation 7 is supposed to be doing. This possibly doesn't affect some of the conclusions taken from this section, but I'd like to see this potential mistake discussed/addressed.\n\n\nThat said, if the authors can sufficiently address the 3 points above, I'd be willing to increase my rating for this paper.\n\nFinally, I have a few other more minor (nice-to-have) points:\n- Having in the related work a discussion on the relationship with coreset methods would be nice\n- Experiments showing how well the distilled datasets transfer to different network architectures than those used in training would be interesting? Even other ML algorithms would be quite interesting?\n- \"We often find that the number of distilled images required to achieve good performance is an informative indicator of the dataset diversity\" => I'm not sure what in the paper actually justifies / demonstrates this statement.\n- Figure 4 is presented as an \"Ablation study\", but an ablation study is where you remove certains parts of a model or algorithm and see what happens, which isn't the case here. I think it's better described as a hyper-parameter sensitivity study. \n- Some typos:\n   * the below objective => the objective below\n   * w.r.t. to => w.r.t\n   * the discrete part rather => the discrete parts rather\n   * necesary => necessary\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Original problem and well written paper, but that lacks comparisons to baselines", "review": "\n=== Post rebuttal update ===\n\nThanks to the addition of better baselines, I've increased my score for this paper. While I'm still not super convinced of its potential for application, I find the idea original and worth discussing at the conference.\n\n=== Pre-rebuttal review ===\nThis paper presents an approach to compress a dataset into a much smaller number of synthetic samples that are optimized to yield as good performance as possible when a given model is trained on that smaller dataset. This is done by unrolling the gradient descent procedure of training such a model to allow for gradient-based optimization of synthetic samples themselves as well as the used learning rates.\n\nIn summary, my evaluation is as follow:\n\n*Pros*\n- Pretty original problem formulation\n- Generally well written paper\n\n*Cons*\n- Lack of comparison with simple baselines in basic dataset distillation setting\n- Use in practical applications (domain adaptation, data poisoning) yet to be convincingly demonstrated\n- Possibly a mistake in the theoretical analysis of the linear case\n\nIndeed, I found the paper to be generally quite clear and enjoyed reading it. One minor thing I struggled a bit with is the distinction between \"SG steps\" and \"Epochs\" (I believe the former corresponds to when the synthetic samples are different between GD steps, whereas the later corresponds to the number of times the method repeatedly cycles over these samples) so I would perhaps encourage the authors to emphasize that difference. \n\nI also find the problem statement that proposed to be interesting and thought provoking, and the solution that's proposed seems quite appropriate and well thought out.\n\nThat said, I'm worried about the following:\n\n- Unless I misunderstood, in the basic dataset distillation setting a comparison is never provided with training on a randomly selected subset of the training set. Presumably the results are worse, but I think these results should be in the paper. I would also argue for having another baseline, which would try to (approximately) optimize the choice of which training examples are put in the subset. A very simple approach would be to take the 200 runs already performed for the random selection and select the subset providing the best accuracy on the full training set and only report the performance of that subset (instead of the mean and std of all 200 runs). In short, this would help determine to what extent there is value to synthesizing entirely new samples. Moreover, I think a simple alternative baseline for creating synthesized samples should be considered. Specifically, I'd personally would like to know the performance of using per-class k-means clustering and training on the cluster centroids as the distilled dataset.\n\n- While I appreciate that the authors identify potential applications and report some results on them, I think they currently fall short of convincing the reader of the potential of dataset distillation for these applications. For domain adaptation, no actual domain adaptation baseline is compared against (a good candidate would be method from Daume III 2007, at the very least). For data poisoning, I find that the assumptions for attacks are pretty strong, i.e. that a) you have access to the parameters of the pre-trained models to attack and b) that the model is doing additional updates *only* on the synthesized data. If there are reasons to think that such assumptions are reasonable, I'd at least expect the paper to motivate why that is. \n\n- In the analysis of the simple linear case, in Equation 7, there appears to be a mistake, specifically some missing parentheses:\n\nd^Td( (I-\\eta/M d^Td)\\theta_0 + \\eta/M \\tilde{d}^T\\tilde{t}) = d^T t\n\ni.e. there should be parentheses right after \"d^Td\" and right before \"=\". This is from replacing \\theta^* by the expression for \\theta_1 in Equation 6, which is what I think Equation 7 is supposed to be doing. This possibly doesn't affect some of the conclusions taken from this section, but I'd like to see this potential mistake discussed/addressed.\n\n\nThat said, if the authors can sufficiently address the 3 points above, I'd be willing to increase my rating for this paper.\n\nFinally, I have a few other more minor (nice-to-have) points:\n- Having in the related work a discussion on the relationship with coreset methods would be nice\n- Experiments showing how well the distilled datasets transfer to different network architectures than those used in training would be interesting? Even other ML algorithms would be quite interesting?\n- \"We often find that the number of distilled images required to achieve good performance is an informative indicator of the dataset diversity\" => I'm not sure what in the paper actually justifies / demonstrates this statement.\n- Figure 4 is presented as an \"Ablation study\", but an ablation study is where you remove certains parts of a model or algorithm and see what happens, which isn't the case here. I think it's better described as a hyper-parameter sensitivity study. \n- Some typos:\n   * the below objective => the objective below\n   * w.r.t. to => w.r.t\n   * the discrete part rather => the discrete parts rather\n   * necesary => necessary\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1540946011783}, {"id": "S1lPGgNVs7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper593/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper addresses the interesting problem of generating a small number of synthetic examples that can be used to train a classifier, replacing a larger dataset. \n\nThe paper is clearly written, the approach makes sense, and the experiments are interesting. \nMy major concerns are regarding to previous literature, analysis of the algorithm, and details of the experiment. Overall, I expect an ICLR paper to go deeper (rather than wide). I recommend presenting strong convincing evidence on one front. \n\n\nSpecific comments: \n(1)  I'm missing analysis of the proposed procedure. It wasn't fully clear which Loss it minimizes and if it indeed guaranteed to converge to the minimum of that loss. \n\n(2)  The topic of learning from few samples is presented as completely new. It is well known that for classical linear algorithms like the Perceptron and SVM, the weights are a weighted sum of (label-weighted) samples, hence by definition of these algorithm, there is a single sample that can be used to \"train\" the model in one step. I'd expect some discussion of how the proposed approach relate to these classical approaches. \nThere is also existing literature on a related problem of selecting samples (Teaching dimension Goldman&Kearns) that could be somewhat relevant here. \n\n(3) Motivation. The paper provide several motivations for dataset distillation. I support the first motivation of scientific understanding what data is actually needed for a classifier, and this means that deeper analysis is needed. The practical motivations are less convincing, because (a) domain adaptation experiments are not compared with real baselines (b) robustness of poisoning with a single sample is not studied/discussed.\n\n(4) experiments: The intro states that training with 10 images reaches 94% accuracy, but this does not seem consistent with the results in Table 1. The caption of figure 2 suggests that accuracy is between 12% and 94% which means the stated 94% is not representative or typical. Could you clarify? \nFor domain adaptation. The baseline (random images) are very weak, and still perform almost   comparably to the proposed approach. More robust experiments are needed here: stronger baselines, decent hyper-parameter search etc.\n\n(5) Writing and exposition: The paper addresses two issues: (a) learning with few synthetic samples, and (b) learning with few gradient steps. The intro tends to mix the two, and it is not clear why learning with a single gradient step is important. I recommend to separate the two topics more clearly. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach, not fully mature yet", "review": "The paper addresses the interesting problem of generating a small number of synthetic examples that can be used to train a classifier, replacing a larger dataset. \n\nThe paper is clearly written, the approach makes sense, and the experiments are interesting. \nMy major concerns are regarding to previous literature, analysis of the algorithm, and details of the experiment. Overall, I expect an ICLR paper to go deeper (rather than wide). I recommend presenting strong convincing evidence on one front. \n\n\nSpecific comments: \n(1)  I'm missing analysis of the proposed procedure. It wasn't fully clear which Loss it minimizes and if it indeed guaranteed to converge to the minimum of that loss. \n\n(2)  The topic of learning from few samples is presented as completely new. It is well known that for classical linear algorithms like the Perceptron and SVM, the weights are a weighted sum of (label-weighted) samples, hence by definition of these algorithm, there is a single sample that can be used to \"train\" the model in one step. I'd expect some discussion of how the proposed approach relate to these classical approaches. \nThere is also existing literature on a related problem of selecting samples (Teaching dimension Goldman&Kearns) that could be somewhat relevant here. \n\n(3) Motivation. The paper provide several motivations for dataset distillation. I support the first motivation of scientific understanding what data is actually needed for a classifier, and this means that deeper analysis is needed. The practical motivations are less convincing, because (a) domain adaptation experiments are not compared with real baselines (b) robustness of poisoning with a single sample is not studied/discussed.\n\n(4) experiments: The intro states that training with 10 images reaches 94% accuracy, but this does not seem consistent with the results in Table 1. The caption of figure 2 suggests that accuracy is between 12% and 94% which means the stated 94% is not representative or typical. Could you clarify? \nFor domain adaptation. The baseline (random images) are very weak, and still perform almost   comparably to the proposed approach. More robust experiments are needed here: stronger baselines, decent hyper-parameter search etc.\n\n(5) Writing and exposition: The paper addresses two issues: (a) learning with few synthetic samples, and (b) learning with few gradient steps. The intro tends to mix the two, and it is not clear why learning with a single gradient step is important. I recommend to separate the two topics more clearly. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1539747855330}], "openreview_url": "https://openreview.net/forum?id=Sy4lojC9tm", "arxiv_id": "1811.10959", "paper_pdf": "papers/Sy4lojC9tm.pdf", "paper_pdf_sha256": "ad01e938ef806e9031e08a809eaab74354f85cc023c0f93b19c7cd9ef1c24608", "paper_pdf_bytes": 2555599, "paper_pdf_source": "openreview", "code_url": "https://github.com/ssnl/dataset-distillation", "code_repository": "ssnl/dataset-distillation", "code_commit": "98e4894d108df23f386303e0b98fe4e0b3d15ba8", "code_archive": "repos/Sy4lojC9tm.zip", "code_archive_sha256": "53634ef2e4fb1355e3212125a2635ca0f1600684a8aec1ab7381f4152ad6c7b1", "code_archive_bytes": 1862739, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 1845, "github_languages": {"Python": 132281}, "github_archived": false, "github_pushed_at": "2025-06-17T22:44:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dataset-distillation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XeZ5WBIRvz", "year": 2026, "status": "rejected", "title": "When Ethics and Payoffs Diverge: LLM Agents in Morally Charged Social Dilemmas", "authors": ["Steffen Backmann", "David Guzman Piedrahita", "Emanuel Tewolde", "Rada Mihalcea", "Bernhard Schölkopf", "Zhijing Jin"], "authorids": ["~Steffen_Backmann1", "~David_Guzman_Piedrahita1", "~Emanuel_Tewolde1", "~Rada_Mihalcea1", "~Bernhard_Schölkopf1", "~Zhijing_Jin1"], "authors_source": "OpenReview API", "abstract": "Recent advances in large language models (LLMs) have enabled their use in complex agentic roles, involving decision-making with humans or other agents, making ethical alignment a key AI safety concern. While prior work has examined both LLMs' moral judgment and strategic behavior in social dilemmas, there is limited understanding of how they act when moral imperatives directly conflict with rewards or incentives. To investigate this, we introduce MORAL Behavior in Social Dilemma SIMulation (MoralSim) and evaluate how LLMs behave in the prisoner's dilemma and public goods game with morally charged contexts. In MoralSim, we test a range of frontier models across both game structures and three distinct moral framings, enabling a systematic examination of how LLMs navigate social dilemmas in which ethical norms conflict with payoff-maximizing strategies. Our results show substantial variation across models in both their general tendency to act morally and the consistency of their behavior across game types, the specific moral framing, and situational factors such as opponent behavior and survival risks. Crucially, no model exhibits consistently moral behavior in MoralSim, highlighting the need for caution when deploying LLMs in agentic roles where the agent's \"self-interest\" may conflict with ethical expectations.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "JVzgiDT3SH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18093/Reviewer_VVKc"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper investigates the behavior of large language models (LLMs) as agents in social dilemmas that include moral constraints. Specifically, it explores how these models make decisions when moral principles conflict with strategies that yield higher rewards. To address this, the authors propose a new evaluation framework called MoralSim (Moral Behavior Social Dilemma Simulation), which integrates classic game-theoretic environments (such as repeated Prisoner's Dilemma and Public Goods Game) with real-world moral scenarios.", "review_text": "The paper investigates the behavior of large language models (LLMs) as agents in social dilemmas that include moral constraints. Specifically, it explores how these models make decisions when moral principles conflict with strategies that yield higher rewards. To address this, the authors propose a new evaluation framework called MoralSim (Moral Behavior Social Dilemma Simulation), which integrates classic game-theoretic environments (such as repeated Prisoner's Dilemma and Public Goods Game) with real-world moral scenarios.", "strengths": "- The MoralSim framework combines classic game-theoretic scenarios with real-world moral dilemmas, enabling a systematic and comprehensive evaluation. This framework is more structured than previous fragmented tests. For example, compared to the MACHIAVELLI benchmark—which primarily uses text adventure games to assess agent behaviorarxiv.org—MoralSim employs formal game structures integrated with moral contexts, allowing for more controlled and easily quantifiable comparisons. This could be better than just simple QA.", "weaknesses": "- Scientificness of the evaluation of this assumption. Although multi-agent framework could provide a more vivid setting for revoke LLM's decision under certain scenarios, it could still be a question of how real and how consistent these evaluations are. According to my experience these testings are easily be changed by small parts of prompts. Yet, this paper don't provide a convincing enough evidence to illustrate the scintificness of this testing. \n\n- The novelty is somewhat incremental. Although the MoralSim framework’s integration of morality and game theory is commendable, its concepts overlap with some existing work [1,2,3]. Also, please discuss these papers in the main paper more to let audiance familiar with context and existing research as well as the differences. Especially [2,3] already reported the betray behavior of LLMs\n\n[1] Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark\n\n[2] Moral Alignment for LLM Agents \n\n[3] Survival Games: Human-LLM Strategic Showdowns under Severe Resource Scarcity", "questions": "- How authors evaluate the testing is stable instead of impacted by small part of prompt design? How often do models change archetype between Goal-oriented vs Neutral prompts? Provide a confusion matrix of archetypes across prompts.\n\n- Please include a comparison table vs MACHIAVELLI, GovSim, Survival Games, and moral-reward alignment detailing the unique “human-harm resource” dimension and survival horizon.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates the behavior of large language models (LLMs) as agents in social dilemmas that include moral constraints. Specifically, it explores how these models make decisions when moral principles conflict with strategies that yield higher rewards. To address this, the authors propose a new evaluation framework called MoralSim (Moral Behavior Social Dilemma Simulation), which integrates classic game-theoretic environments (such as repeated Prisoner's Dilemma and Public Goods Game) with real-world moral scenarios.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The MoralSim framework combines classic game-theoretic scenarios with real-world moral dilemmas, enabling a systematic and comprehensive evaluation. This framework is more structured than previous fragmented tests. For example, compared to the MACHIAVELLI benchmark—which primarily uses text adventure games to assess agent behaviorarxiv.org—MoralSim employs formal game structures integrated with moral contexts, allowing for more controlled and easily quantifiable comparisons. This could be better than just simple QA.", "weaknesses": "- Scientificness of the evaluation of this assumption. Although multi-agent framework could provide a more vivid setting for revoke LLM's decision under certain scenarios, it could still be a question of how real and how consistent these evaluations are. According to my experience these testings are easily be changed by small parts of prompts. Yet, this paper don't provide a convincing enough evidence to illustrate the scintificness of this testing. \n\n- The novelty is somewhat incremental. Although the MoralSim framework’s integration of morality and game theory is commendable, its concepts overlap with some existing work [1,2,3]. Also, please discuss these papers in the main paper more to let audiance familiar with context and existing research as well as the differences. Especially [2,3] already reported the betray behavior of LLMs\n\n[1] Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark\n\n[2] Moral Alignment for LLM Agents \n\n[3] Survival Games: Human-LLM Strategic Showdowns under Severe Resource Scarcity", "questions": "- How authors evaluate the testing is stable instead of impacted by small part of prompt design? How often do models change archetype between Goal-oriented vs Neutral prompts? Provide a confusion matrix of archetypes across prompts.\n\n- Please include a comparison table vs MACHIAVELLI, GovSim, Survival Games, and moral-reward alignment detailing the unique “human-harm resource” dimension and survival horizon.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762069492648}, {"id": "od26TzKlI5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18093/Reviewer_8FDD"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The authors introduce MoralSim, a benchmark that focuses on studying LLM behaviors in scenarios where explicitly forced to trade-off between moral behavior and rewards. The authors encase these scenarios in realistic settings to provide greater fidelity to underlying LLM behavior and perform analysis of different behavioral questions in LLMs to find that", "review_text": "The authors introduce MoralSim, a benchmark that focuses on studying LLM behaviors in scenarios where explicitly forced to trade-off between moral behavior and rewards. The authors encase these scenarios in realistic settings to provide greater fidelity to underlying LLM behavior and perform analysis of different behavioral questions in LLMs to find that", "strengths": "- The paper is well-written and the experiments are well-designed. \n- The settings developed by the paper more clearly expose trade-offs in moral behavior compared to prior work. \n- There are clear behavioral takeaways from the paper, e.g., opponent behavior meaningfully steers LLM actions or that moral context improves the morality of LLM behaviors.", "weaknesses": "While the games offer a step towards realism, they still do not capture the full nuances of reality. In particular:\n- All the games are two-player. Many real-world settings involve multiple players with different levels of power and interlocking incentives.\n- The text of the games themselves is still somewhat unrealistic. There is an explicit payoff structure described in the system prompt (e.g., Figure 2), whereas in the real world an LLM agent would need to uncover those trade-offs themselves.\nHowever, I realize that these nuances are quite tricky to incorporate and would make some of the analysis less clean, so I don't think that should hold the paper back.\n\nThere is no discussion or analysis of whether or not the LLMs participating in the games recognize that they are in the game. Recent research into scheming (https://www.antischeming.ai/) suggests that frontier LLMs may recognize that they are being evaluated, which could question the validity of the research results. However, it is difficult to assess scheming without access to the full reasoning trace, so I also don't count this as a strong weakness.", "questions": "- Would it be possible at all to analyze whether the LLMs recognize that they are playing a game and if that would alter their behavior in any sense? Would it be possible to attempt to induce \"evaluation-awareness\" into the LLMs (e.g., by being more suggestive with the wording that the environment is a game) and see how that affects the results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce MoralSim, a benchmark that focuses on studying LLM behaviors in scenarios where explicitly forced to trade-off between moral behavior and rewards. The authors encase these scenarios in realistic settings to provide greater fidelity to underlying LLM behavior and perform analysis of different behavioral questions in LLMs to find that", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- The paper is well-written and the experiments are well-designed. \n- The settings developed by the paper more clearly expose trade-offs in moral behavior compared to prior work. \n- There are clear behavioral takeaways from the paper, e.g., opponent behavior meaningfully steers LLM actions or that moral context improves the morality of LLM behaviors.", "weaknesses": "While the games offer a step towards realism, they still do not capture the full nuances of reality. In particular:\n- All the games are two-player. Many real-world settings involve multiple players with different levels of power and interlocking incentives.\n- The text of the games themselves is still somewhat unrealistic. There is an explicit payoff structure described in the system prompt (e.g., Figure 2), whereas in the real world an LLM agent would need to uncover those trade-offs themselves.\nHowever, I realize that these nuances are quite tricky to incorporate and would make some of the analysis less clean, so I don't think that should hold the paper back.\n\nThere is no discussion or analysis of whether or not the LLMs participating in the games recognize that they are in the game. Recent research into scheming (https://www.antischeming.ai/) suggests that frontier LLMs may recognize that they are being evaluated, which could question the validity of the research results. However, it is difficult to assess scheming without access to the full reasoning trace, so I also don't count this as a strong weakness.", "questions": "- Would it be possible at all to analyze whether the LLMs recognize that they are playing a game and if that would alter their behavior in any sense? Would it be possible to attempt to induce \"evaluation-awareness\" into the LLMs (e.g., by being more suggestive with the wording that the environment is a game) and see how that affects the results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761985026513}, {"id": "wTVAHNP5Q4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18093/Reviewer_mAm7"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 5, "summary": "The paper presents MORALSIM, a new framework for evaluating the moral behavior of LLM agents in situations where ethical norms conflict with personal interests and benefits. The authors investigate the behavior of 9 modern language models (including Claude-3.7-Sonnet, GPT-4o, Deepseek-R1, o3-mini, and others) in two classic game scenarios: Prisoner's Dilemma and the Public Goods Game. Each game is embedded in three different moral contexts: Contractual Reporting, Privacy Protection, and Green Production. The full factorial design of the experiment is used, varying the type of game, the moral context, the opponent's behavior (always cooperating / always betraying) and the risk of survival. The key result: none of the tested models demonstrates consistent moral behavior in all scenarios. The proportion of morally oriented actions varies from 7.9% (Qwen-3) to 76.3% (GPT-4o-mini). The authors use causal analysis (Average Treatment Effects) to determine the factors influencing moral decisions, and show that the structure of the game, the specific moral context, and the opponent's behavior have the greatest impact.", "review_text": "The paper presents MORALSIM, a new framework for evaluating the moral behavior of LLM agents in situations where ethical norms conflict with personal interests and benefits. The authors investigate the behavior of 9 modern language models (including Claude-3.7-Sonnet, GPT-4o, Deepseek-R1, o3-mini, and others) in two classic game scenarios: Prisoner's Dilemma and the Public Goods Game. Each game is embedded in three different moral contexts: Contractual Reporting, Privacy Protection, and Green Production. The full factorial design of the experiment is used, varying the type of game, the moral context, the opponent's behavior (always cooperating / always betraying) and the risk of survival. The key result: none of the tested models demonstrates consistent moral behavior in all scenarios. The proportion of morally oriented actions varies from 7.9% (Qwen-3) to 76.3% (GPT-4o-mini). The authors use causal analysis (Average Treatment Effects) to determine the factors influencing moral decisions, and show that the structure of the game, the specific moral context, and the opponent's behavior have the greatest impact.", "strengths": "- A full-factor design with clear manipulations (type of game, moral context, survival risk, opponent behavior) allows you to isolate the effects of each factor\n- Quality of analysis. Analysis of ~3500 reflections of agents reveals decision-making mechanisms. Causal assessment through ATEs yields quantitative effects with confidence intervals. It is shown that profit maximizer models (Deepseek-R1, Qwen) rely on profit maximization, while more cooperative ones (Claude, GPT-4o) more often take into account moral considerations.\n- All models show a decrease in morale precisely when the user is most vulnerable (at risk of bankruptcy), which raises important issues of AI security.\n- The code is open, the prompts are documented in detail.", "weaknesses": "1. PD and PGG only. Other structures (for example, Trust Game, Stag Hunt) could reveal other patterns of moral behavior. An extension to asymmetric games would be especially valuable.\n2. In the real world, agents can often negotiate, which significantly changes the dynamics of cooperation. The authors acknowledge this, but do not investigate it.\n3. For some models (Claude-3.7-Sonnet, Gemini-2.5-Flash), versions without reasoning mode were used for cost reasons. Given that the analysis has shown the importance of reasoning, this limits the conclusions.\n4. Multi-agent scenarios (N>2) could better reflect social dynamics and collective responsibility.", "questions": "1. You have shown invariance to paraphrases, but how sensitive are the results to more fundamental changes in the presentation of the problem? For example, what if we present the same dilemmas through different metaphors or change the order in which options are presented?\n2. Does the moral behavior of agents change as they gain experience in repetitive games? Are there signs of \"moral learning\" or adaptation of strategies over time?\n3. Moral norms vary between cultures. Do you plan to investigate how different cultural contexts affect the moral behavior of LLM agents?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents MORALSIM, a new framework for evaluating the moral behavior of LLM agents in situations where ethical norms conflict with personal interests and benefits. The authors investigate the behavior of 9 modern language models (including Claude-3.7-Sonnet, GPT-4o, Deepseek-R1, o3-mini, and others) in two classic game scenarios: Prisoner's Dilemma and the Public Goods Game. Each game is embedded in three different moral contexts: Contractual Reporting, Privacy Protection, and Green Production. The full factorial design of the experiment is used, varying the type of game, the moral context, the opponent's behavior (always cooperating / always betraying) and the risk of survival. The key result: none of the tested models demonstrates consistent moral behavior in all scenarios. The proportion of morally oriented actions varies from 7.9% (Qwen-3) to 76.3% (GPT-4o-mini). The authors use causal analysis (Average Treatment Effects) to determine the factors influencing moral decisions, and show that the structure of the game, the specific moral context, and the opponent's behavior have the greatest impact.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "- A full-factor design with clear manipulations (type of game, moral context, survival risk, opponent behavior) allows you to isolate the effects of each factor\n- Quality of analysis. Analysis of ~3500 reflections of agents reveals decision-making mechanisms. Causal assessment through ATEs yields quantitative effects with confidence intervals. It is shown that profit maximizer models (Deepseek-R1, Qwen) rely on profit maximization, while more cooperative ones (Claude, GPT-4o) more often take into account moral considerations.\n- All models show a decrease in morale precisely when the user is most vulnerable (at risk of bankruptcy), which raises important issues of AI security.\n- The code is open, the prompts are documented in detail.", "weaknesses": "1. PD and PGG only. Other structures (for example, Trust Game, Stag Hunt) could reveal other patterns of moral behavior. An extension to asymmetric games would be especially valuable.\n2. In the real world, agents can often negotiate, which significantly changes the dynamics of cooperation. The authors acknowledge this, but do not investigate it.\n3. For some models (Claude-3.7-Sonnet, Gemini-2.5-Flash), versions without reasoning mode were used for cost reasons. Given that the analysis has shown the importance of reasoning, this limits the conclusions.\n4. Multi-agent scenarios (N>2) could better reflect social dynamics and collective responsibility.", "questions": "1. You have shown invariance to paraphrases, but how sensitive are the results to more fundamental changes in the presentation of the problem? For example, what if we present the same dilemmas through different metaphors or change the order in which options are presented?\n2. Does the moral behavior of agents change as they gain experience in repetitive games? Are there signs of \"moral learning\" or adaptation of strategies over time?\n3. Moral norms vary between cultures. Do you plan to investigate how different cultural contexts affect the moral behavior of LLM agents?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761979177800}, {"id": "r2vhE31bfT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18093/Reviewer_w88c"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "This work studies how LLMs behave when moral obligations conflict with incentives in social dilemmas. The authors present a simulation framework called MoralSim that essentially is a collection of situations that reflect social dilemmas expressed in natural language. The authors evaluate several LLMs considering Prisoner's Dilemma and Public Good games. The findings show that models vary widely in their tendency to act morally and lack consistent ethical behavior across situations, suggesting that LLMs may be unreliable for decision-making roles requiring alignment with ethical norms.\n\n**General note**: I reviewed this manuscript for a previous conference. I would like to note that I went through the paper carefully, noting the main differences. The core contributions are essentially the same apart from the causality analysis; hence, I find myself in a position to repeat my critiques. Please note that the list of weaknesses below takes into consideration the revised version of the paper with respect to my initial review. I would also like to note that my comment about the naming of the “morality score” was addressed, and in this version, the authors talk about “cooperation score”. I noted the following changes with respect to the version I reviewed below (these are the main ones I was able to clearly identify):\n\n- The morality score has been renamed cooperation score.\n- A section about causal effect estimation has been added (plus a discussion of the treatment effects later in the paepr).\n- The discussion of the results regarding Q3 and Q4 has been extended.\n- There is a new section about reasoning trace analysis in the appendix.\n\nThese points/observations did not affect my review. I treated this paper completely *as new*. I believe that these are independent conferences and we should consider the manuscripts in their own right each time. I added this note explicitly in order to avoid potential criticisms about the fact I raised again some of the concerns of my previous review (since those parts are unchanged).", "review_text": "This work studies how LLMs behave when moral obligations conflict with incentives in social dilemmas. The authors present a simulation framework called MoralSim that essentially is a collection of situations that reflect social dilemmas expressed in natural language. The authors evaluate several LLMs considering Prisoner's Dilemma and Public Good games. The findings show that models vary widely in their tendency to act morally and lack consistent ethical behavior across situations, suggesting that LLMs may be unreliable for decision-making roles requiring alignment with ethical norms.\n\n**General note**: I reviewed this manuscript for a previous conference. I would like to note that I went through the paper carefully, noting the main differences. The core contributions are essentially the same apart from the causality analysis; hence, I find myself in a position to repeat my critiques. Please note that the list of weaknesses below takes into consideration the revised version of the paper with respect to my initial review. I would also like to note that my comment about the naming of the “morality score” was addressed, and in this version, the authors talk about “cooperation score”. I noted the following changes with respect to the version I reviewed below (these are the main ones I was able to clearly identify):\n\n- The morality score has been renamed cooperation score.\n- A section about causal effect estimation has been added (plus a discussion of the treatment effects later in the paepr).\n- The discussion of the results regarding Q3 and Q4 has been extended.\n- There is a new section about reasoning trace analysis in the appendix.\n\nThese points/observations did not affect my review. I treated this paper completely *as new*. I believe that these are independent conferences and we should consider the manuscripts in their own right each time. I added this note explicitly in order to avoid potential criticisms about the fact I raised again some of the concerns of my previous review (since those parts are unchanged).", "strengths": "- This is indeed a very interesting topic.\n- The paper is well written and very easy to follow. The evaluation of the work is conducted.", "weaknesses": "- This is a well-written paper, but it is difficult to identify a substantial contribution with respect to the state of the art. In fact, the papers listed under “LLMs in Game Theory Settings” focus directly or indirectly on decision-making problems in social dilemmas. In fact, the authors do not really consider actual moral frameworks in their analysis as these works do.  The authors claim that various prior research has already explored LLMs’ moral reasoning and strategic behaviour separately. The authors of [Tennant et al., 2024] (actually published and presented at ICLR 2025) essentially not only present a variety of games (including the Iterated Prisoner's Dilemma) but also study how to run a fine-tuning procedure to examine the effects of moral decision-making.\n- The authors present very limited analysis in terms of sensitivity to variations of the prompts, including the agent setup that is discussed in Section 3.3 (e.g., descriptions of the setting, personal memory, and current task). These might have substantial effects and should be discussed by the authors in my opinion.\n- It is quite surprising that the authors considered moral dilemmas, but the actual dynamics of the responses are only partially analysed/considered by the authors. In fact, the authors mainly focus on the choice of the single agents. The authors also consider the opponent alignment, but this is quite confusing since it appears to the reviewer quite orthogonal to the problem of acting morally. The authors consider the relative payoff, but considering the fact that these are classic (repeated) games, the analysis of the actual cumulative payoff would have probably been more informative from a game theory point of view.\n- The survival rate score is interesting, but it appears rather disjointed considering the core topic of the paper. With respect to the statistical validity of the simulations, it is unclear how different repetitions of the games have been implemented. \n- The paper essentially lacks a related work section. The authors moved it to the Appendix, but that is outside the 10 pages of the main body of the paper. It seems to me a way for going above the 10 page limit: this is not fair towards the other ICLR authors in my opinion.", "questions": "I do not have specific questions for the authors. My core concern is about the very limited contribution of this work with respect to the state of the art.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studies how LLMs behave when moral obligations conflict with incentives in social dilemmas. The authors present a simulation framework called MoralSim that essentially is a collection of situations that reflect social dilemmas expressed in natural language. The authors evaluate several LLMs considering Prisoner's Dilemma and Public Good games. The findings show that models vary widely in their tendency to act morally and lack consistent ethical behavior across situations, suggesting that LLMs may be unreliable for decision-making roles requiring alignment with ethical norms.\n\n**General note**: I reviewed this manuscript for a previous conference. I would like to note that I went through the paper carefully, noting the main differences. The core contributions are essentially the same apart from the causality analysis; hence, I find myself in a position to repeat my critiques. Please note that the list of weaknesses below takes into consideration the revised version of the paper with respect to my initial review. I would also like to note that my comment about the naming of the “morality score” was addressed, and in this version, the authors talk about “cooperation score”. I noted the following changes with respect to the version I reviewed below (these are the main ones I was able to clearly identify):\n\n- The morality score has been renamed cooperation score.\n- A section about causal effect estimation has been added (plus a discussion of the treatment effects later in the paepr).\n- The discussion of the results regarding Q3 and Q4 has been extended.\n- There is a new section about reasoning trace analysis in the appendix.\n\nThese points/observations did not affect my review. I treated this paper completely *as new*. I believe that these are independent conferences and we should consider the manuscripts in their own right each time. I added this note explicitly in order to avoid potential criticisms about the fact I raised again some of the concerns of my previous review (since those parts are unchanged).", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- This is indeed a very interesting topic.\n- The paper is well written and very easy to follow. The evaluation of the work is conducted.", "weaknesses": "- This is a well-written paper, but it is difficult to identify a substantial contribution with respect to the state of the art. In fact, the papers listed under “LLMs in Game Theory Settings” focus directly or indirectly on decision-making problems in social dilemmas. In fact, the authors do not really consider actual moral frameworks in their analysis as these works do.  The authors claim that various prior research has already explored LLMs’ moral reasoning and strategic behaviour separately. The authors of [Tennant et al., 2024] (actually published and presented at ICLR 2025) essentially not only present a variety of games (including the Iterated Prisoner's Dilemma) but also study how to run a fine-tuning procedure to examine the effects of moral decision-making.\n- The authors present very limited analysis in terms of sensitivity to variations of the prompts, including the agent setup that is discussed in Section 3.3 (e.g., descriptions of the setting, personal memory, and current task). These might have substantial effects and should be discussed by the authors in my opinion.\n- It is quite surprising that the authors considered moral dilemmas, but the actual dynamics of the responses are only partially analysed/considered by the authors. In fact, the authors mainly focus on the choice of the single agents. The authors also consider the opponent alignment, but this is quite confusing since it appears to the reviewer quite orthogonal to the problem of acting morally. The authors consider the relative payoff, but considering the fact that these are classic (repeated) games, the analysis of the actual cumulative payoff would have probably been more informative from a game theory point of view.\n- The survival rate score is interesting, but it appears rather disjointed considering the core topic of the paper. With respect to the statistical validity of the simulations, it is unclear how different repetitions of the games have been implemented. \n- The paper essentially lacks a related work section. The authors moved it to the Appendix, but that is outside the 10 pages of the main body of the paper. It seems to me a way for going above the 10 page limit: this is not fair towards the other ICLR authors in my opinion.", "questions": "I do not have specific questions for the authors. My core concern is about the very limited contribution of this work with respect to the state of the art.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761948740192}], "openreview_url": "https://openreview.net/forum?id=XeZ5WBIRvz", "arxiv_id": "2505.19212", "paper_pdf": "papers/XeZ5WBIRvz.pdf", "paper_pdf_sha256": "b496485458c8df743bc31687d488b21db87e0c702ad7deea0c88a0a83c595e4a", "paper_pdf_bytes": 787306, "paper_pdf_source": "openreview", "code_url": "https://github.com/sbackmann/moralsim", "code_repository": "sbackmann/moralsim", "code_commit": "6e7daa2fd393ebee7563de2944442353fa04414e", "code_archive": "repos/XeZ5WBIRvz.zip", "code_archive_sha256": "04aef98cdcd59817b61e0ebc3cd357dc1b35e4f9ef799e4aacfed35c6187309e", "code_archive_bytes": 942025, "code_file_count": 61, "code_extensions": {".py": 56, ".ipynb": 5}, "github_disk_usage_kb": 866, "github_languages": {"Jupyter Notebook": 838848, "Python": 207797}, "github_archived": false, "github_pushed_at": "2025-06-03T17:19:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/when-ethics-and-payoffs-diverge-llm-agents-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "CSZKElOtG5", "year": 2025, "status": "rejected", "title": "MeanSparse: Post-Training Robustness Enhancement Through Mean-Centered Feature Sparsification", "authors": ["Sajjad Amini", "Mohammadreza Teymoorianfard", "Shiqing Ma", "Amir Houmansadr"], "authorids": ["~Sajjad_Amini1", "~Mohammadreza_Teymoorianfard1", "~Shiqing_Ma2", "~Amir_Houmansadr1"], "authors_source": "OpenReview API", "abstract": "We present a simple yet effective method to improve the robustness of both Convolutional and attention-based Neural Networks against adversarial examples by post-processing an adversarially trained model. \nOur technique, MeanSparse,  cascades the activation functions of a trained model with novel operators that sparsify mean-centered feature vectors. \nThis is equivalent to reducing feature variations around the mean, and we show that such reduced variations merely affect the model's utility, yet they strongly attenuate the adversarial perturbations and decrease the attacker's success rate.\nOur experiments show that, when applied to the top models in the RobustBench leaderboard, MeanSparse achieves a new robustness record of $75.28$% (from $73.71$%), $44.78$% (from $42.67$%) and $62.12$% (from $59.56$%) on CIFAR-10, CIFAR-100 and ImageNet, respectively, in terms of AutoAttack accuracy. \nCode: https://anonymous.4open.science/r/MeanSparse-84B0/", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "CvlyXY766u", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11250/Reviewer_ik2X"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces a method called MEANSPARSE, which enhances the adversarial robustness of trained neural networks in a post-processing manner without compromising clean accuracy. The idea behind this method is to attenuate the reliance on non-robust features by modifying activation functions. Their experiments demonstrate a significant increase in the adversarial robustness of neural networks.", "review_text": "This paper introduces a method called MEANSPARSE, which enhances the adversarial robustness of trained neural networks in a post-processing manner without compromising clean accuracy. The idea behind this method is to attenuate the reliance on non-robust features by modifying activation functions. Their experiments demonstrate a significant increase in the adversarial robustness of neural networks.", "strengths": "1. The idea of manipulating activation functions to enhance adversarial robustness is quite interesting.\n2. Extensive experiments have been conducted to support the proposed method.\n3. The intuition behind the idea is provided, making it easier to follow.\n4. The paper includes a discussion of its limitations.", "weaknesses": "1. The explanation of the method in the Introduction is vague. On page 1, lines 37 and 44, you state: \"In this work,...\" twice, but the first sentence mentions \"design of activation functions,\" while the second shifts to \"sparsifying features to enhance robustness against adversarial attacks.\" These two sentences make it unclear what the key points of your work are and how investigating activation functions is related to sparsifying features. It would be better to provide a consistent expression for the main points and briefly describe the connection between them. I suggest a clear statement about your key contributions.\n\n2. Lack of intuitive explanation of the proposed concept. On page 1, line 48, the proposed sparsity method \"Mean-based Sparsification\" is not well explained, and it’s followed only by a brief description of Figure 1. Without a clear explanation of the \"sparsification operator\", it is difficult to follow. I suggest adding a brief explanation of the \"sparsification operation\" you mentioned in line 90 and providing a concise mathematical definition or detailed description of how the sparsification operator works, perhaps with a simple example.\n\n3. There are some writing errors in the paper. In Figure 2, \"equation 3\" appears at the end of the title, which is confusing. It would be better to remove it or rephrase the title.", "questions": "1. On page 6, in the first and second paragraphs, you mention two approaches and choose the second one. I am curious why you selected the second approach as your method. Is it superior? Can you provide evidence to support your choice and a brief comparison of the two approaches, highlighting the advantages and disadvantages of each?\n\n2. In Section 4 on page 7, I did not find the experimental settings. You only mentioned that the experiments were conducted using an NVIDIA A100 GPU. For better reproducibility, it would be helpful to include detailed experimental settings. I suggest providing details such as the software versions used, hyperparameters, data preprocessing steps, and any other relevant configuration details that would allow others to replicate their experiments", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a method called MEANSPARSE, which enhances the adversarial robustness of trained neural networks in a post-processing manner without compromising clean accuracy. The idea behind this method is to attenuate the reliance on non-robust features by modifying activation functions. Their experiments demonstrate a significant increase in the adversarial robustness of neural networks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The idea of manipulating activation functions to enhance adversarial robustness is quite interesting.\n2. Extensive experiments have been conducted to support the proposed method.\n3. The intuition behind the idea is provided, making it easier to follow.\n4. The paper includes a discussion of its limitations.", "weaknesses": "1. The explanation of the method in the Introduction is vague. On page 1, lines 37 and 44, you state: \"In this work,...\" twice, but the first sentence mentions \"design of activation functions,\" while the second shifts to \"sparsifying features to enhance robustness against adversarial attacks.\" These two sentences make it unclear what the key points of your work are and how investigating activation functions is related to sparsifying features. It would be better to provide a consistent expression for the main points and briefly describe the connection between them. I suggest a clear statement about your key contributions.\n\n2. Lack of intuitive explanation of the proposed concept. On page 1, line 48, the proposed sparsity method \"Mean-based Sparsification\" is not well explained, and it’s followed only by a brief description of Figure 1. Without a clear explanation of the \"sparsification operator\", it is difficult to follow. I suggest adding a brief explanation of the \"sparsification operation\" you mentioned in line 90 and providing a concise mathematical definition or detailed description of how the sparsification operator works, perhaps with a simple example.\n\n3. There are some writing errors in the paper. In Figure 2, \"equation 3\" appears at the end of the title, which is confusing. It would be better to remove it or rephrase the title.", "questions": "1. On page 6, in the first and second paragraphs, you mention two approaches and choose the second one. I am curious why you selected the second approach as your method. Is it superior? Can you provide evidence to support your choice and a brief comparison of the two approaches, highlighting the advantages and disadvantages of each?\n\n2. In Section 4 on page 7, I did not find the experimental settings. You only mentioned that the experiments were conducted using an NVIDIA A100 GPU. For better reproducibility, it would be helpful to include detailed experimental settings. I suggest providing details such as the software versions used, hyperparameters, data preprocessing steps, and any other relevant configuration details that would allow others to replicate their experiments", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730740302563}, {"id": "T3lY2JtqrD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11250/Reviewer_8mZY"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper proposes a post-training method for enhancing the robustness of adversarially trained models. In particular, it inserts, at various layers, the MeanSparse modules which project the features (before activation) onto their mean computed over the training set if they are closer than a threshold to the mean itself. This has the goal of preventing an attacker from exploiting non-informative features to change the predicted class. In the experiments, several SOTA robust models are equipped with MeanSparse, which leaves clean performance unchanged while improving robustness.", "review_text": "The paper proposes a post-training method for enhancing the robustness of adversarially trained models. In particular, it inserts, at various layers, the MeanSparse modules which project the features (before activation) onto their mean computed over the training set if they are closer than a threshold to the mean itself. This has the goal of preventing an attacker from exploiting non-informative features to change the predicted class. In the experiments, several SOTA robust models are equipped with MeanSparse, which leaves clean performance unchanged while improving robustness.", "strengths": "- The proposed method is efficient, can be applied to any pre-trained model without additional training, and incurs in limited additional inference cost.\n\n- The experiments include several architectures, datasets and threat models.", "weaknesses": "- The main concern is about the possible presence of gradient masking [A]. In fact, as mentioned in the paper, the MeanSparse operator induces zero gradients for the features which are projected. Since these are supposed to be the most common ones, and MeanSparse is applied in multiple points in the network, one can expect that the computed gradient might contain limited information, and thus the gradient-based attacks not work properly. While the paper tests black-box attacks, in this case the base model is (highly) robust, and the improvements given by MeanSparse are in the order of 1-3%, which might be of the same order or even smaller than the gap between white- and black-box (with standard query budget) attacks for the base model: then it is not clear that in this case this is sufficient to exclude gradient masking. A simple adaptive attack would consist in removing the projection operation when computing the gradient in the attacks, which would modify the backward pass of the model while preserving its predictions (a similar approach to BPDA [A]).\n\n- The discussion in Sec. 3.1 seems a bit disconnected from the final approach (also, $z_{k-1}$ in Eq. (7) is not defined).\n\n[A] https://arxiv.org/abs/1802.00420", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a post-training method for enhancing the robustness of adversarially trained models. In particular, it inserts, at various layers, the MeanSparse modules which project the features (before activation) onto their mean computed over the training set if they are closer than a threshold to the mean itself. This has the goal of preventing an attacker from exploiting non-informative features to change the predicted class. In the experiments, several SOTA robust models are equipped with MeanSparse, which leaves clean performance unchanged while improving robustness.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The proposed method is efficient, can be applied to any pre-trained model without additional training, and incurs in limited additional inference cost.\n\n- The experiments include several architectures, datasets and threat models.", "weaknesses": "- The main concern is about the possible presence of gradient masking [A]. In fact, as mentioned in the paper, the MeanSparse operator induces zero gradients for the features which are projected. Since these are supposed to be the most common ones, and MeanSparse is applied in multiple points in the network, one can expect that the computed gradient might contain limited information, and thus the gradient-based attacks not work properly. While the paper tests black-box attacks, in this case the base model is (highly) robust, and the improvements given by MeanSparse are in the order of 1-3%, which might be of the same order or even smaller than the gap between white- and black-box (with standard query budget) attacks for the base model: then it is not clear that in this case this is sufficient to exclude gradient masking. A simple adaptive attack would consist in removing the projection operation when computing the gradient in the attacks, which would modify the backward pass of the model while preserving its predictions (a similar approach to BPDA [A]).\n\n- The discussion in Sec. 3.1 seems a bit disconnected from the final approach (also, $z_{k-1}$ in Eq. (7) is not defined).\n\n[A] https://arxiv.org/abs/1802.00420", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730724897796}, {"id": "fJ6oSPygLH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11250/Reviewer_ry3U"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper seeks to improve adversarial training by inducing sparsity in the features of an adversarially-trained model. To achieve this, MeanSparse blocks variations in features within a given distance of the mean, which is intended to lessen the importance of non-robust features during training. Before describing their sparsification method, the authors provide intuition as to how non-robust features can be removed during training through regularizad optimization. Here, a regularization term is included in the training objective which penalizes the $\\ell_0$ norm of mean-centered learned features. In theory, this would remove small deviations from the mean which don't result in significant reductions in loss. Inspired by this intuition, the authors then propose a sparsification operator, which is applied during forward propagation and explicitly blocks variation within a given distance of the feature mean. Severel implementation challenges are then addressed. Experimental results show that integrating MeanSparse into state of the art image classifiers can significantly improve robustness with negligible impact on clean performance.", "review_text": "This paper seeks to improve adversarial training by inducing sparsity in the features of an adversarially-trained model. To achieve this, MeanSparse blocks variations in features within a given distance of the mean, which is intended to lessen the importance of non-robust features during training. Before describing their sparsification method, the authors provide intuition as to how non-robust features can be removed during training through regularizad optimization. Here, a regularization term is included in the training objective which penalizes the $\\ell_0$ norm of mean-centered learned features. In theory, this would remove small deviations from the mean which don't result in significant reductions in loss. Inspired by this intuition, the authors then propose a sparsification operator, which is applied during forward propagation and explicitly blocks variation within a given distance of the feature mean. Severel implementation challenges are then addressed. Experimental results show that integrating MeanSparse into state of the art image classifiers can significantly improve robustness with negligible impact on clean performance.", "strengths": "- The results show meaningful improvement in the robustness of SOTA adversarially trained models. Since this method can be applied post-training, it has the potential to lead to a significant jump across the board in standards for robustness to adversarial examples.\n- The experimental section appears well designed. Results are presented for CIFAR-10, CIFAR-100, and Imagenet, and a representative variety of model architectures and adversarial attacks are tested.\n- The method is also shown to be robust against adaptive attacks, as shown in Appendix A.4.\n- I quite like the visualizations provided in Figure 1, they simply and effectively convey how MeanSparse operates.", "weaknesses": "- The description of the MeanSparse technique is somewhat ambiguous to me. The term \"feature\" is often used, but never specifically defined. What features are being used here? Are you referring to input features? Activations of the final layer? Activations of some internal layer?\n- I don't entirely agree with the provided intuition for blocking minor variations around the feature mean. Based off of the provided explanation, I would expect this approach to work when minor variations are blocked around high-probability feature values. However, it's not clear to me that the feature mean would always be a high-probability point.\n- Appendix A.2 does provide results looking at $\\ell_2$ bounded attacks, rather than $\\ell_\\infty$ bounded attacks. However, I think the topic of how MeanSparse performs against different threat models does warrant additional study. Threat models have been studied in which the adversary can introduce perturbations that are unbounded in $\\ell_p$ space (i.e. [1]), and it is not obvious to me whether these types of attacks would have similar impacts on the distributions of features. If the claims made in this paper are limited to $\\ell_p$ bounded attacks, I think that should be made explicit.\n\n[1] Xiao, Chaowei, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. \"Spatially transformed adversarial examples.\" arXiv preprint arXiv:1801.02612 (2018).", "questions": "- Is the feature mean always a high-probability point? I would imagine that in certain multimodal distributions the mean is actually unlikely to occur. Might this occur in practice?\n- One component of MeanSparse involves calculating the mean and standard deviations of features. Can you provide more information regarding what data is used to calculate these values? Specifically, are these statistics computed on benign or adversarially perturbed data? \n- How sensitive is the feature mean to class imbalances in the training set? If classes aren't evenly balanced (assuming feature distributions are different for different classes), could that result in changes in feature means?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper seeks to improve adversarial training by inducing sparsity in the features of an adversarially-trained model. To achieve this, MeanSparse blocks variations in features within a given distance of the mean, which is intended to lessen the importance of non-robust features during training. Before describing their sparsification method, the authors provide intuition as to how non-robust features can be removed during training through regularizad optimization. Here, a regularization term is included in the training objective which penalizes the $\\ell_0$ norm of mean-centered learned features. In theory, this would remove small deviations from the mean which don't result in significant reductions in loss. Inspired by this intuition, the authors then propose a sparsification operator, which is applied during forward propagation and explicitly blocks variation within a given distance of the feature mean. Severel implementation challenges are then addressed. Experimental results show that integrating MeanSparse into state of the art image classifiers can significantly improve robustness with negligible impact on clean performance.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The results show meaningful improvement in the robustness of SOTA adversarially trained models. Since this method can be applied post-training, it has the potential to lead to a significant jump across the board in standards for robustness to adversarial examples.\n- The experimental section appears well designed. Results are presented for CIFAR-10, CIFAR-100, and Imagenet, and a representative variety of model architectures and adversarial attacks are tested.\n- The method is also shown to be robust against adaptive attacks, as shown in Appendix A.4.\n- I quite like the visualizations provided in Figure 1, they simply and effectively convey how MeanSparse operates.", "weaknesses": "- The description of the MeanSparse technique is somewhat ambiguous to me. The term \"feature\" is often used, but never specifically defined. What features are being used here? Are you referring to input features? Activations of the final layer? Activations of some internal layer?\n- I don't entirely agree with the provided intuition for blocking minor variations around the feature mean. Based off of the provided explanation, I would expect this approach to work when minor variations are blocked around high-probability feature values. However, it's not clear to me that the feature mean would always be a high-probability point.\n- Appendix A.2 does provide results looking at $\\ell_2$ bounded attacks, rather than $\\ell_\\infty$ bounded attacks. However, I think the topic of how MeanSparse performs against different threat models does warrant additional study. Threat models have been studied in which the adversary can introduce perturbations that are unbounded in $\\ell_p$ space (i.e. [1]), and it is not obvious to me whether these types of attacks would have similar impacts on the distributions of features. If the claims made in this paper are limited to $\\ell_p$ bounded attacks, I think that should be made explicit.\n\n[1] Xiao, Chaowei, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. \"Spatially transformed adversarial examples.\" arXiv preprint arXiv:1801.02612 (2018).", "questions": "- Is the feature mean always a high-probability point? I would imagine that in certain multimodal distributions the mean is actually unlikely to occur. Might this occur in practice?\n- One component of MeanSparse involves calculating the mean and standard deviations of features. Can you provide more information regarding what data is used to calculate these values? Specifically, are these statistics computed on benign or adversarially perturbed data? \n- How sensitive is the feature mean to class imbalances in the training set? If classes aren't evenly balanced (assuming feature distributions are different for different classes), could that result in changes in feature means?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730708461127}, {"id": "8CdWMaASFm", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11250/Reviewer_bqUf"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper introduces MeanSparse, a novel post-training method designed to enhance the robustness of convolutional and attention-based neural networks. MeanSparse works by sparsifying mean-centered feature vectors. This technique effectively decreases the success rate of adversarial attacks by minimizing the exploitable feature variations. The paper demonstrates that MEANSPARSE improves the robustness on datasets such as CIFAR-10, CIFAR-100, and ImageNet.", "review_text": "The paper introduces MeanSparse, a novel post-training method designed to enhance the robustness of convolutional and attention-based neural networks. MeanSparse works by sparsifying mean-centered feature vectors. This technique effectively decreases the success rate of adversarial attacks by minimizing the exploitable feature variations. The paper demonstrates that MEANSPARSE improves the robustness on datasets such as CIFAR-10, CIFAR-100, and ImageNet.", "strengths": "1. The application of sparsification to mean-centered features is an innovative approach, as it targets non-robust features in a unique way.\n\n2. Authors provide comprehensive results across multiple datasets and models\n\n3. The paper is well-organized and clearly describes the underlying motivation, methodology, and results", "weaknesses": "1. whether its effectiveness generalizes across different types of attention-based models, such as ViT?\n\n2. Could you provide provide an analysis of how MEANSPARSE's effectiveness scales with model size? Authors only apply their method on large networks, like wrs-70. How about the performance of the method on smaller network, like ResNet-18, Swin-Small?\n\n3. After reading the codes provided, I found that the MeadSquare will calculate the mean and var of the input data to get running_mean and running_var. I do no think the model should change its any parameters according to the test data.\n\n4. The number of baselines in this paper is too few. Authors do not compare their methods with other sparsity methods.", "questions": "1. Could the method be used for adversarial training and accumulate the running mean and running var during training? Then during testing, fix the running mean and var.\n\n2. Could you provide a detailed analysis of how MEANSPARSE's effectiveness changes as the attack strength increases, like 16/255?", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "all_content": {"summary": "The paper introduces MeanSparse, a novel post-training method designed to enhance the robustness of convolutional and attention-based neural networks. MeanSparse works by sparsifying mean-centered feature vectors. This technique effectively decreases the success rate of adversarial attacks by minimizing the exploitable feature variations. The paper demonstrates that MEANSPARSE improves the robustness on datasets such as CIFAR-10, CIFAR-100, and ImageNet.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The application of sparsification to mean-centered features is an innovative approach, as it targets non-robust features in a unique way.\n\n2. Authors provide comprehensive results across multiple datasets and models\n\n3. The paper is well-organized and clearly describes the underlying motivation, methodology, and results", "weaknesses": "1. whether its effectiveness generalizes across different types of attention-based models, such as ViT?\n\n2. Could you provide provide an analysis of how MEANSPARSE's effectiveness scales with model size? Authors only apply their method on large networks, like wrs-70. How about the performance of the method on smaller network, like ResNet-18, Swin-Small?\n\n3. After reading the codes provided, I found that the MeadSquare will calculate the mean and var of the input data to get running_mean and running_var. I do no think the model should change its any parameters according to the test data.\n\n4. The number of baselines in this paper is too few. Authors do not compare their methods with other sparsity methods.", "questions": "1. Could the method be used for adversarial training and accumulate the running mean and running var during training? Then during testing, fix the running mean and var.\n\n2. Could you provide a detailed analysis of how MEANSPARSE's effectiveness changes as the attack strength increases, like 16/255?", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729277508089}], "openreview_url": "https://openreview.net/forum?id=CSZKElOtG5", "arxiv_id": "2406.05927", "paper_pdf": "papers/CSZKElOtG5.pdf", "paper_pdf_sha256": "8f6f42c37182e448b8f35d5c830d235b7e3189c51d791cc83316f9b5357c69c9", "paper_pdf_bytes": 2082384, "paper_pdf_source": "openreview", "code_url": "https://github.com/SPIN-UMass/MeanSparse", "code_repository": "SPIN-UMass/MeanSparse", "code_commit": "4d7eb1a15e39bcfd88236b379e2496a289f9e66c", "code_archive": "repos/CSZKElOtG5.zip", "code_archive_sha256": "f660aa76974b4eb2aaaeb827dd7c014856aa2d2e0a37ecb3821547f5235031df", "code_archive_bytes": 106941, "code_file_count": 40, "code_extensions": {".py": 40}, "github_disk_usage_kb": 70, "github_languages": {"Python": 351891}, "github_archived": false, "github_pushed_at": "2025-10-16T00:36:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/meansparse-post-training-robustness"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TYMeXb6PAw", "year": 2024, "status": "rejected", "title": "Adaptive Compression of the Latent Space in Variational Autoencoders", "authors": ["Gabriela Sejnova", "Michal Vavrecka", "Karla Stepanova"], "authorids": ["~Gabriela_Sejnova1", "~Michal_Vavrecka1", "~Karla_Stepanova1"], "authors_source": "OpenReview API", "abstract": "Variational Autoencoders (VAEs) are powerful generative models that have been widely used in various fields, including image and text generation. However, one of the known challenges in using VAEs is the model's sensitivity to its hyperparameters, such as the latent space size. This paper presents a simple extension of VAEs for automatically determining the optimal latent space size during the training process by gradually decreasing the latent size through neuron removal and observing the model performance.  The proposed method is compared to traditional hyperparameter grid search and is shown to be significantly faster while still achieving the best optimal dimensionality on four image datasets. Furthermore, we show that the final performance of our method is comparable to training on the optimal latent size from scratch, and might thus serve as a convenient substitute.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jtIBXYeR8w", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5445/Reviewer_iAX2"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Variational AutoEncoders (VAEs) are powerful generative models, but the optimal latent space size is difficult to set. This paper proposes a heuristic method called ALD-VAE capable of finding the optimal latent dimension by gradually decreasing the latent size and evaluating the model performance during the training process. The method is compared to grid search and is shown to be faster and achieve the better latent space sizes on four image datasets: MNIST, FashionMNIST, SPRITES and EuroSAT.", "review_text": "Variational AutoEncoders (VAEs) are powerful generative models, but the optimal latent space size is difficult to set. This paper proposes a heuristic method called ALD-VAE capable of finding the optimal latent dimension by gradually decreasing the latent size and evaluating the model performance during the training process. The method is compared to grid search and is shown to be faster and achieve the better latent space sizes on four image datasets: MNIST, FashionMNIST, SPRITES and EuroSAT.", "strengths": "- This paper proposes an heuristic approach to find the optimal latent dimension for VAEs. During VAE training, the method gradually decrease the latent space size by evaluating the model performances in FID score, Silhouette score, and reconstruction error. When the slopes of all these scores are positive, then the pruning of the latent space size stops and the method returns the optimal latent space size. \n\n- The empirical results shown in Fig. 2 show that the method is able to find the latent space sizes that produce almost the lowest FID scores, the highest Silhouette scores, and the lowest reconstruction errors on four datasets.", "weaknesses": "- The algorithm is quite heuristic and straightforward. There is no theoretical justification of the method. \n\n- There are too many hyper-parameters in this method, such as p = 5, n = 5, window w, number of clusters, latent_decrease and so on. The authors need to investigate how sensitive are those parameters and how different values of the hyper-parameters affect the final performance. \n\n- The datasets used in this paper are small datasets, in terms of both number of images and image resolution. The authors need to do experiments on other large-scale high resolution image datasets such as CelebA-HQ, and ImageNet to evaluate the proposed method. The large-scale high resolution image datasets can also be used to evaluate how sensitive are those hyper-parameters.", "questions": "- The number of clusters in Silhouette score should also be chosen or optimized. How does the number of clusters affect the Silhouette score? \n\n- During the pruning process, how do you decide which $n$ neurons are pruned? If you randomly pick $n$ neurons and remove them, will you remove the neurons that are more important than the remaining neurons? \n\n- The stopping criteria is that \"the slopes of the observed metrics calculated over the last k=20 epochs are all positive\". I think this may not be robust. Could it be possible that in future epochs, the FIDs and reconstruction loss be decreasing again and the the Silhouette score be increasing? A related question is that the optimality in Eq. 2 is a multi-objective optimization problem. How to balance the three different objectives? \n\n- There are several important parts not clear. FID'_r and FID'_g are not defined. What are polynomial scores? $p$ in page 5 two lines below Alg. 1 is a scalar 5, but in Eq. 5 is a function.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Variational AutoEncoders (VAEs) are powerful generative models, but the optimal latent space size is difficult to set. This paper proposes a heuristic method called ALD-VAE capable of finding the optimal latent dimension by gradually decreasing the latent size and evaluating the model performance during the training process. The method is compared to grid search and is shown to be faster and achieve the better latent space sizes on four image datasets: MNIST, FashionMNIST, SPRITES and EuroSAT.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- This paper proposes an heuristic approach to find the optimal latent dimension for VAEs. During VAE training, the method gradually decrease the latent space size by evaluating the model performances in FID score, Silhouette score, and reconstruction error. When the slopes of all these scores are positive, then the pruning of the latent space size stops and the method returns the optimal latent space size. \n\n- The empirical results shown in Fig. 2 show that the method is able to find the latent space sizes that produce almost the lowest FID scores, the highest Silhouette scores, and the lowest reconstruction errors on four datasets.", "weaknesses": "- The algorithm is quite heuristic and straightforward. There is no theoretical justification of the method. \n\n- There are too many hyper-parameters in this method, such as p = 5, n = 5, window w, number of clusters, latent_decrease and so on. The authors need to investigate how sensitive are those parameters and how different values of the hyper-parameters affect the final performance. \n\n- The datasets used in this paper are small datasets, in terms of both number of images and image resolution. The authors need to do experiments on other large-scale high resolution image datasets such as CelebA-HQ, and ImageNet to evaluate the proposed method. The large-scale high resolution image datasets can also be used to evaluate how sensitive are those hyper-parameters.", "questions": "- The number of clusters in Silhouette score should also be chosen or optimized. How does the number of clusters affect the Silhouette score? \n\n- During the pruning process, how do you decide which $n$ neurons are pruned? If you randomly pick $n$ neurons and remove them, will you remove the neurons that are more important than the remaining neurons? \n\n- The stopping criteria is that \"the slopes of the observed metrics calculated over the last k=20 epochs are all positive\". I think this may not be robust. Could it be possible that in future epochs, the FIDs and reconstruction loss be decreasing again and the the Silhouette score be increasing? A related question is that the optimality in Eq. 2 is a multi-objective optimization problem. How to balance the three different objectives? \n\n- There are several important parts not clear. FID'_r and FID'_g are not defined. What are polynomial scores? $p$ in page 5 two lines below Alg. 1 is a scalar 5, but in Eq. 5 is a function.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698879977308}, {"id": "8RNucdtWqu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5445/Reviewer_ksU9"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In the paper, the authors propose a method to automatically find the optimal latent dimension during the training of VAEs. This technique is based on the continuous analysis of the slopes of some metrics: the Silhouette score, the reconstruction loss, and the FID (both for reconstructed and generated data). The experiments performed on 4 real-world datasets (SPRITES, MNIST, FMNIST, and EuroSAT) show that the resulting ALD-VAE model (with latent dimension adjustment), compared to the classical VAE (with a fixed optimal size of the latent), provides (at convergence) comparable results for the mentioned metrics. In addition, training ALD-VAE on MNIST and FMNIST to find the optimal latent dimension is robust to different initial latent sizes and data seeds.", "review_text": "In the paper, the authors propose a method to automatically find the optimal latent dimension during the training of VAEs. This technique is based on the continuous analysis of the slopes of some metrics: the Silhouette score, the reconstruction loss, and the FID (both for reconstructed and generated data). The experiments performed on 4 real-world datasets (SPRITES, MNIST, FMNIST, and EuroSAT) show that the resulting ALD-VAE model (with latent dimension adjustment), compared to the classical VAE (with a fixed optimal size of the latent), provides (at convergence) comparable results for the mentioned metrics. In addition, training ALD-VAE on MNIST and FMNIST to find the optimal latent dimension is robust to different initial latent sizes and data seeds.", "strengths": "(1) The paper is clearly written and easy to understand.\n\n(2) The proposed ALD-VAE model seems to be faster compared to VAE with grid search for optimal latent dimension.", "weaknesses": "(1) The significance of the proposed solution (although somewhat novel) was not sufficiently justified.\n\n(2) Limited experimental setup (lack of experiments on more complicated large-scale datasets, e.g., CelebA).\n\n(3) The authors emphasize that their solution is significantly faster than those using a grid search procedure (which is reliable), but they do not provide any experimental evidence for this claim.\n\n(4) As the authors claim, the optimal latent dimensions obtained by ALD-VAE trained on the SPRITES and EuroSAT datasets are not robust to different data seeds.", "questions": "Minor comment: wrong sign '¿' in line 23 of Alg. 1.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In the paper, the authors propose a method to automatically find the optimal latent dimension during the training of VAEs. This technique is based on the continuous analysis of the slopes of some metrics: the Silhouette score, the reconstruction loss, and the FID (both for reconstructed and generated data). The experiments performed on 4 real-world datasets (SPRITES, MNIST, FMNIST, and EuroSAT) show that the resulting ALD-VAE model (with latent dimension adjustment), compared to the classical VAE (with a fixed optimal size of the latent), provides (at convergence) comparable results for the mentioned metrics. In addition, training ALD-VAE on MNIST and FMNIST to find the optimal latent dimension is robust to different initial latent sizes and data seeds.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "(1) The paper is clearly written and easy to understand.\n\n(2) The proposed ALD-VAE model seems to be faster compared to VAE with grid search for optimal latent dimension.", "weaknesses": "(1) The significance of the proposed solution (although somewhat novel) was not sufficiently justified.\n\n(2) Limited experimental setup (lack of experiments on more complicated large-scale datasets, e.g., CelebA).\n\n(3) The authors emphasize that their solution is significantly faster than those using a grid search procedure (which is reliable), but they do not provide any experimental evidence for this claim.\n\n(4) As the authors claim, the optimal latent dimensions obtained by ALD-VAE trained on the SPRITES and EuroSAT datasets are not robust to different data seeds.", "questions": "Minor comment: wrong sign '¿' in line 23 of Alg. 1.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698779051908}, {"id": "5wxbtsWxdD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5445/Reviewer_fzax"], "rating": "1: strong reject", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The author proposes a method that automatically determines the optimal latent space dimensionality in Variational Autoencoders", "review_text": "The author proposes a method that automatically determines the optimal latent space dimensionality in Variational Autoencoders", "strengths": "The author proposes a method that automatically determines the optimal latent space dimensionality in Variational Autoencoders", "weaknesses": "1. This paper falls short of the standards required for publication. The novelty presented in the work is notably limited, with the proposed algorithm that render the research somewhat trivial. The method is basically about:\n- train a model\n- evaluate the model (use eq.4)\n- reduce the dimension when the evaluation result does not converge.\n- do everything again if the dimension is reduced.\n\n  This dimension-reduction pipeline looks like some simple hyper-parameter(the dimension of the latent) search loop. How about just launching a dozen of training processes with different dimensionalities? Also, the evaluation methods seem trivial as well: FID, reconstruction loss, K-means. All of them were sufficiently discussed in Machine Learning communities, simply aggregating them yields limited novelty.\n\n2. Furthermore, the absence of a baseline comparison is a big problem. Even some very basic method such as PCA is not included. \n\n3. The paper's focus on dimension reduction **alone**, as a research topic in representation learning, might not be impactful. A more relevant and insightful approach would be to evaluate representations in terms of \"bits per dimension\", \"ELBO\", or \"rate-distortion,\" rather than purely dimensionality. For example, [Alemi et al.](https://proceedings.mlr.press/v80/alemi18a/alemi18a.pdf) had some theoretical analysis of VAE's rate-distortion curve. [Higgins et al.](https://openreview.net/pdf?id=Sy2fzU9gl) discussed the disentanglement of the representation. [Balle et al.](https://arxiv.org/abs/1611.01704) proposed a method can use VAE for image compression, whose representation can be quantized and entropy-coded as bitstream. I also recommend referring to the paper at [Yang et al.](https://www.nowpublishers.com/article/Details/CGV-107) for a comprehensive introduction to data compression, which could provide valuable insights and context. Actually none of these work cares about the dimensionality of the representation, what people really want to know is \"how much or what kind of information the representation can carry\".\n\nIn conclusion, this paper falls short of the quality standards expected for a publication in ICLR. At least, more analysis and more baselines are needed.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The author proposes a method that automatically determines the optimal latent space dimensionality in Variational Autoencoders", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "strengths": "The author proposes a method that automatically determines the optimal latent space dimensionality in Variational Autoencoders", "weaknesses": "1. This paper falls short of the standards required for publication. The novelty presented in the work is notably limited, with the proposed algorithm that render the research somewhat trivial. The method is basically about:\n- train a model\n- evaluate the model (use eq.4)\n- reduce the dimension when the evaluation result does not converge.\n- do everything again if the dimension is reduced.\n\n  This dimension-reduction pipeline looks like some simple hyper-parameter(the dimension of the latent) search loop. How about just launching a dozen of training processes with different dimensionalities? Also, the evaluation methods seem trivial as well: FID, reconstruction loss, K-means. All of them were sufficiently discussed in Machine Learning communities, simply aggregating them yields limited novelty.\n\n2. Furthermore, the absence of a baseline comparison is a big problem. Even some very basic method such as PCA is not included. \n\n3. The paper's focus on dimension reduction **alone**, as a research topic in representation learning, might not be impactful. A more relevant and insightful approach would be to evaluate representations in terms of \"bits per dimension\", \"ELBO\", or \"rate-distortion,\" rather than purely dimensionality. For example, [Alemi et al.](https://proceedings.mlr.press/v80/alemi18a/alemi18a.pdf) had some theoretical analysis of VAE's rate-distortion curve. [Higgins et al.](https://openreview.net/pdf?id=Sy2fzU9gl) discussed the disentanglement of the representation. [Balle et al.](https://arxiv.org/abs/1611.01704) proposed a method can use VAE for image compression, whose representation can be quantized and entropy-coded as bitstream. I also recommend referring to the paper at [Yang et al.](https://www.nowpublishers.com/article/Details/CGV-107) for a comprehensive introduction to data compression, which could provide valuable insights and context. Actually none of these work cares about the dimensionality of the representation, what people really want to know is \"how much or what kind of information the representation can carry\".\n\nIn conclusion, this paper falls short of the quality standards expected for a publication in ICLR. At least, more analysis and more baselines are needed.", "questions": "see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "1: strong reject", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698754682395}, {"id": "qMNyGHf5uf", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5445/Reviewer_g7w5"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a method that adaptively decreases the latent space size during the VAE training procedure. The stopping mechanism is based on Silhouette score, reconstruction loss, and FID.", "review_text": "The paper proposes a method that adaptively decreases the latent space size during the VAE training procedure. The stopping mechanism is based on Silhouette score, reconstruction loss, and FID.", "strengths": "- The writing is easy to follow and understand.", "weaknesses": "- Lack of comparison with other SOTA methods. e.g. MaskAAE(Mondal, Arnab Kumar, et al.), GECO(De Boom, et al.), \n- Lack of ablation study. It might be worth ablating the contribution of e.g. Silhouette score for the decision rule. Are all of them really useful?\n- There are no clear quantitative results. All comparisons of the four metrics are shown in Figure 3, and it’s not always easy to compare the two curves. (e.g. FID gen in MNIST, FashionMNIST, reconstruction loss in EUROSAT, etc.) A table might be more clear.", "questions": "The paper claims that the proposed method is significantly faster, but there is no analysis for time. Why is the proposed method faster? Can the proposed method converge with less epochs? The grid search might need to be run multiple times but it can be run in parallel, while the proposed method has to run clustering and needs to run in sequential order.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method that adaptively decreases the latent space size during the VAE training procedure. The stopping mechanism is based on Silhouette score, reconstruction loss, and FID.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The writing is easy to follow and understand.", "weaknesses": "- Lack of comparison with other SOTA methods. e.g. MaskAAE(Mondal, Arnab Kumar, et al.), GECO(De Boom, et al.), \n- Lack of ablation study. It might be worth ablating the contribution of e.g. Silhouette score for the decision rule. Are all of them really useful?\n- There are no clear quantitative results. All comparisons of the four metrics are shown in Figure 3, and it’s not always easy to compare the two curves. (e.g. FID gen in MNIST, FashionMNIST, reconstruction loss in EUROSAT, etc.) A table might be more clear.", "questions": "The paper claims that the proposed method is significantly faster, but there is no analysis for time. Why is the proposed method faster? Can the proposed method converge with less epochs? The grid search might need to be run multiple times but it can be run in parallel, while the proposed method has to run clustering and needs to run in sequential order.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698697893943}], "openreview_url": "https://openreview.net/forum?id=TYMeXb6PAw", "arxiv_id": "2312.06280", "paper_pdf": "papers/TYMeXb6PAw.pdf", "paper_pdf_sha256": "6066150ba43be18e301ea40665e9877eb83066f153ca2312d6dddcf73f30097d", "paper_pdf_bytes": 1013259, "paper_pdf_source": "openreview", "code_url": "https://github.com/gabinsane/ald-vae", "code_repository": "gabinsane/ald-vae", "code_commit": "cdf2e8b71470092761daae01adabe3fac781f42c", "code_archive": "repos/TYMeXb6PAw.zip", "code_archive_sha256": "5c2b785cef8c31b4c8a8f88f70db28b46499a47d811df0c244afa02bea5deec5", "code_archive_bytes": 63873, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 59, "github_languages": {"Python": 196713}, "github_archived": false, "github_pushed_at": "2024-10-13T17:52:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-compression-of-the-latent-space-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "i2e2wqt0nAI", "year": 2023, "status": "rejected", "title": "Rethinking Symbolic Regression Datasets and Benchmarks for Scientific Discovery", "authors": ["Yoshitomo Matsubara", "Naoya Chiba", "Ryo Igarashi", "Tatsunori Taniai", "Yoshitaka Ushiku"], "authorids": ["~Yoshitomo_Matsubara1", "~Naoya_Chiba1", "~Ryo_Igarashi1", "~Tatsunori_Taniai4", "~Yoshitaka_Ushiku3"], "authors_source": "OpenReview API", "abstract": "This paper revisits datasets and evaluation criteria for Symbolic Regression, a task of expressing given data using mathematical equations, specifically focused on its potential for scientific discovery. Focused on a set of formulas used in the existing datasets based on Feynman Lectures on Physics, we recreate 120 datasets to discuss the performance of symbolic regression for scientific discovery (SRSD). For each of the 120 SRSD datasets, we carefully review the properties of the formula and its variables to design reasonably realistic sampling range of values so that our new SRSD datasets can be used for evaluating the potential of SRSD such as whether or not an SR method can (re)discover physical laws from such datasets. As an evaluation metric, we also propose to use normalized edit distances between a predicted equation and the ground-truth equation trees. While existing metrics are either binary or errors between the target values and an SR model's predicted values for a given input, normalized edit distances evaluate a sort of similarity between the ground-truth and predicted equation trees. We have conducted experiments on our new SRSD datasets using five state-of-the-art SR methods in SRBench and a simple baseline based on a recent Transformer architecture. The results show that we provide a more realistic performance evaluation and open up a new machine learning-based approach for scientific discovery. We provide our datasets and code as part of the supplementary material.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BA8Fp-rQ6Z", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3665/Reviewer_V7wv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations for existing datasets and\ndesign their own dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.", "review_text": "The main contribution of this work lies in the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations fo existing datasets and\ndesign their owm dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.\n\nWhere is I in the examples of Table 1? (It is mentioned in the caption\nof this table)\n\n\"we preprocess equations by 1) substituting constant values e.g., π and\nPlanck constant to the expression...\" --> isn't the idea to learn these\nexpressions from data? Do you mean that you *postprocess* the learned\nformulas?\n\n\"coefficient values themselves (e.g., value of C1 in Fig. 2) should\nnot be important\" --> I don't get the idea here. I'd think that\nconstants can differentiate one expression from another when modeling\nreal systems (e.g., physics, math etc).\n\n\"such true equations will not be available in practice\" --> but don't\nyou generate data from these equations in order to create the\ndatasets? How do you expect to validate learning models if you don't\nhave the ground truth? Besides, don't Feynman lectures present the\nequations?\n\nDo you have any explanation of why ST results in 0% R2 for all\nproblems? It seems strange.\n\nOther minor comments:\nboth the metrics --> both metrics\nTable 4 show --> Table 4 shows\n", "strengths": "The main contribution of this work lies in the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations fo existing datasets and\ndesign their owm dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.\n\nWhere is I in the examples of Table 1? (It is mentioned in the caption\nof this table)\n\n\"we preprocess equations by 1) substituting constant values e.g., π and\nPlanck constant to the expression...\" --> isn't the idea to learn these\nexpressions from data? Do you mean that you *postprocess* the learned\nformulas?\n\n\"coefficient values themselves (e.g., value of C1 in Fig. 2) should\nnot be important\" --> I don't get the idea here. I'd think that\nconstants can differentiate one expression from another when modeling\nreal systems (e.g., physics, math etc).\n\n\"such true equations will not be available in practice\" --> but don't\nyou generate data from these equations in order to create the\ndatasets? How do you expect to validate learning models if you don't\nhave the ground truth? Besides, don't Feynman lectures present the\nequations?\n\nDo you have any explanation of why ST results in 0% R2 for all\nproblems? It seems strange.\n\nOther minor comments:\nboth the metrics --> both metrics\nTable 4 show --> Table 4 shows\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations for existing datasets and\ndesign their own dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.", "strength_and_weaknesses": "The main contribution of this work lies in the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations fo existing datasets and\ndesign their owm dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.\n\nWhere is I in the examples of Table 1? (It is mentioned in the caption\nof this table)\n\n\"we preprocess equations by 1) substituting constant values e.g., π and\nPlanck constant to the expression...\" --> isn't the idea to learn these\nexpressions from data? Do you mean that you *postprocess* the learned\nformulas?\n\n\"coefficient values themselves (e.g., value of C1 in Fig. 2) should\nnot be important\" --> I don't get the idea here. I'd think that\nconstants can differentiate one expression from another when modeling\nreal systems (e.g., physics, math etc).\n\n\"such true equations will not be available in practice\" --> but don't\nyou generate data from these equations in order to create the\ndatasets? How do you expect to validate learning models if you don't\nhave the ground truth? Besides, don't Feynman lectures present the\nequations?\n\nDo you have any explanation of why ST results in 0% R2 for all\nproblems? It seems strange.\n\nOther minor comments:\nboth the metrics --> both metrics\nTable 4 show --> Table 4 shows\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: good\nNovelty: good\nReproducibility: authors provide various items that may allow the work to be reproduced", "summary_of_the_review": "The main contribution of this work lies in the design of\ndatasets and benchmark of symbolic regression for scientific\ndiscovery. Authors point our some limitations fo existing datasets and\ndesign their owm dataset categorizing subsets of small, medium and\nlarge complexity and develop a new metric based on tree distance\nbetween the true and the predicted equations.\n\nWhere is I in the examples of Table 1? (It is mentioned in the caption\nof this table)\n\n\"we preprocess equations by 1) substituting constant values e.g., π and\nPlanck constant to the expression...\" --> isn't the idea to learn these\nexpressions from data? Do you mean that you *postprocess* the learned\nformulas?\n\n\"coefficient values themselves (e.g., value of C1 in Fig. 2) should\nnot be important\" --> I don't get the idea here. I'd think that\nconstants can differentiate one expression from another when modeling\nreal systems (e.g., physics, math etc).\n\n\"such true equations will not be available in practice\" --> but don't\nyou generate data from these equations in order to create the\ndatasets? How do you expect to validate learning models if you don't\nhave the ground truth? Besides, don't Feynman lectures present the\nequations?\n\nDo you have any explanation of why ST results in 0% R2 for all\nproblems? It seems strange.\n\nOther minor comments:\nboth the metrics --> both metrics\nTable 4 show --> Table 4 shows\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666984492665}, {"id": "lHkZ-6R5AY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3665/Reviewer_2CZT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a new benchmark set for symbolic regression. Starting from the Feynman symbolic regression dataset (FSRD, Udrescu et al), they propose a new dataset of 120 problems, together with rules for sampling variables and parameters. They introduce a new metric for comparing symbolic functions, based on the edit distance between the trees representing the simplified skeletons of the functions. Finally, they provide baseline comparisons between different popular models (and one introduced by the authors).", "review_text": "The dataset proposed is interesting and welcome addition to current benchmarks and test sets. However, it is very small, and mostly amounts to a new curation of an existing dataset. The edit distance is an interesting idea, but would need significant work before it can be used as a measure of accuracy of symbolic regression. \n\nOverall, my feeling is that the contribution is very marginal, and not novel enough to feature in a venue like ICLR. ", "strengths": "**Strengths**\n\nThe discussion of the limitations of current benchmarks is interesting, and the improvements proposed by the authors (e.g. keeping constants constant, introducing physics-informed ranges for the parameters and variables) makes a lot of sense. Baseline evaluations are provided.\n\n**Weaknesses**\n\nThe dataset remains very small, and most if it is a new curation/annotation of an existing benchmark (FRSD). This limits the novelty and impact of this paper. To me, the main problem of symbolic regression benchmarks is the small size of the datasets, and their lack of diversity. The paper does not address these questions.\n\nThe edit distance is an interesting idea, but needs a lot of improvements before it can be used as a reliable criterion. Two limitations spring to mind.\n\n* Some weighting of the operators is probably needed. Take f(x) = cos(ax+b), g(x)=sin(ax+b), h(x)=cos(ax), k(x)=exp(ax+b). The edit distances are edit(f, g)=1 edit(f, k)=1 and edit(f, h)=2, yet g is equivalent to f, h is a phase shift away from f (i.e. very close), but k has very different mathematical properties from the three others.\n* The magnitude of the constants should be taken into account. For small a, sin(ax) is essentially the same as ax, for large a, the functions are very different. Also, suppose you compare functions of the form f(x)+a.g(x), and u(x)+b.v(x). If a<<1 and b<<1, the edit distance between f and u is probably a much better metric than the edit distance between f+ag and u+bv. This problem will appear every time a function is represented as a sum of terms of decreasing magnitude (a very common situation in science).\n\n ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose a new benchmark set for symbolic regression. Starting from the Feynman symbolic regression dataset (FSRD, Udrescu et al), they propose a new dataset of 120 problems, together with rules for sampling variables and parameters. They introduce a new metric for comparing symbolic functions, based on the edit distance between the trees representing the simplified skeletons of the functions. Finally, they provide baseline comparisons between different popular models (and one introduced by the authors).", "strength_and_weaknesses": "**Strengths**\n\nThe discussion of the limitations of current benchmarks is interesting, and the improvements proposed by the authors (e.g. keeping constants constant, introducing physics-informed ranges for the parameters and variables) makes a lot of sense. Baseline evaluations are provided.\n\n**Weaknesses**\n\nThe dataset remains very small, and most if it is a new curation/annotation of an existing benchmark (FRSD). This limits the novelty and impact of this paper. To me, the main problem of symbolic regression benchmarks is the small size of the datasets, and their lack of diversity. The paper does not address these questions.\n\nThe edit distance is an interesting idea, but needs a lot of improvements before it can be used as a reliable criterion. Two limitations spring to mind.\n\n* Some weighting of the operators is probably needed. Take f(x) = cos(ax+b), g(x)=sin(ax+b), h(x)=cos(ax), k(x)=exp(ax+b). The edit distances are edit(f, g)=1 edit(f, k)=1 and edit(f, h)=2, yet g is equivalent to f, h is a phase shift away from f (i.e. very close), but k has very different mathematical properties from the three others.\n* The magnitude of the constants should be taken into account. For small a, sin(ax) is essentially the same as ax, for large a, the functions are very different. Also, suppose you compare functions of the form f(x)+a.g(x), and u(x)+b.v(x). If a<<1 and b<<1, the edit distance between f and u is probably a much better metric than the edit distance between f+ag and u+bv. This problem will appear every time a function is represented as a sum of terms of decreasing magnitude (a very common situation in science).\n\n ", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, and its quality is good.\n\nThe novelty is very limited (see above). ", "summary_of_the_review": "The dataset proposed is interesting and welcome addition to current benchmarks and test sets. However, it is very small, and mostly amounts to a new curation of an existing dataset. The edit distance is an interesting idea, but would need significant work before it can be used as a measure of accuracy of symbolic regression. \n\nOverall, my feeling is that the contribution is very marginal, and not novel enough to feature in a venue like ICLR. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666271486539}, {"id": "8CQY6yPcDT", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3665/Reviewer_CRUL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper propose a new symbolic regression (SR) dataset for scientific discovery (SRSD) that is a modification of the Feynman symbolic regression database (FSRD) that fixes the following issues present in FSRD: 1) no physical meaning for a number of the datasets, 2) an oversimplified sampling process, 3) distinct equations that are marked as duplicates, and 4) inappropriate or incorrect formulas. The proposed SRSD datasets treats constants (e.g., lightspeed) as constants, samples values for variables with appropriate ranges using experimental observations, and partitions the 120 datasets into easy, medium, and hard sets based on formula operation length and variable domain range. The authors also propose using normalized edit distance to assess the similarity of the structures between the predicted equation models and their respective ground-truth models.\n\nThe authors evaluate 5 established and well-known SR methods and one transformer-based SR model on the proposed SRSD datasets and measure performance using the proposed normalized edit distance, accuracy (which captures instances where the predicted and true model differ by a constant or scalar), and solution rate. Overall, the authors find the transformer-based approach outperforms all other methods on normalized edit distance, but performs worse than all other methods in terms of accuracy and solution rate; this suggests the transformer-based approach provides more structurally similar equation models than existing methods.", "review_text": "Although the paper proposes a potentially valuable collection of SRSD datasets and a new evaluation metric measuring normalized edit distance, the dataset is not well-documented and it is unclear whether the new evaluation metric provides significantly meaningful insights in terms of scientific discovery compared to existing evaluation metrics.", "strengths": "Strengths\n---\nThe proposed SRSD datasets address issues present in the FSRD database in order to provide datasets that better align with the goals of symbolic regression for scientific discovery.\n\nThe proposed SRSD contains 120 datasets, partitioned into three different sets of varying difficulty level.\n\nThe proposed metric of normalized edit distance allows for a more fine-grained analysis into the structural differences between the predicted and true equation models.\n\nThe paper is relatively well-written and easy to read and follow.\n\nWeaknesses\n---\nIt is not clear to me how much insight is gained using the normalized edit distance (NED) over existing metrics measuring the success of symbolic regression models. The results suggest that the transformer-based model achieves slightly better NED over existing methods, but it is difficult to judge if this improvement is meaningful from a scientific discovery point of view. Even if two equation models are mostly structurally similar (as measured by NED), they may differ by one or two elements that could perceivably result in a very different equation model not representative of the actual ground truth equation.\n\nThe experimental results are not very thorough, and only provide results for a small number of metrics for a small number of baselines. Due to the difficulty evaluating the merit of equation-based scientific discovery, perhaps a metric based on manual inspection with domain experts could give some insight into whether or not the proposed metric (and existing metrics) are well-aligned with how humans judge how close a generated equation is to representing an actual scientific equation.\n\nWhen proposing a new dataset, the authors should consider documenting their dataset thoroughly using, for instance, \"Datasheets for Datasets\" (Gebru et al. 2021) or a dataset nutrition label (Holland et al. 2018).\n\nMinor Weaknesses\n---\nConsider using \"\\citet\" when using citations as a noun.\n\nThe solution rate metric is not clearly described.\n\nWhy is accuracy and NED not shown in the comparison between the FSRD and SRSD dataset (Table S16)?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper propose a new symbolic regression (SR) dataset for scientific discovery (SRSD) that is a modification of the Feynman symbolic regression database (FSRD) that fixes the following issues present in FSRD: 1) no physical meaning for a number of the datasets, 2) an oversimplified sampling process, 3) distinct equations that are marked as duplicates, and 4) inappropriate or incorrect formulas. The proposed SRSD datasets treats constants (e.g., lightspeed) as constants, samples values for variables with appropriate ranges using experimental observations, and partitions the 120 datasets into easy, medium, and hard sets based on formula operation length and variable domain range. The authors also propose using normalized edit distance to assess the similarity of the structures between the predicted equation models and their respective ground-truth models.\n\nThe authors evaluate 5 established and well-known SR methods and one transformer-based SR model on the proposed SRSD datasets and measure performance using the proposed normalized edit distance, accuracy (which captures instances where the predicted and true model differ by a constant or scalar), and solution rate. Overall, the authors find the transformer-based approach outperforms all other methods on normalized edit distance, but performs worse than all other methods in terms of accuracy and solution rate; this suggests the transformer-based approach provides more structurally similar equation models than existing methods.", "strength_and_weaknesses": "Strengths\n---\nThe proposed SRSD datasets address issues present in the FSRD database in order to provide datasets that better align with the goals of symbolic regression for scientific discovery.\n\nThe proposed SRSD contains 120 datasets, partitioned into three different sets of varying difficulty level.\n\nThe proposed metric of normalized edit distance allows for a more fine-grained analysis into the structural differences between the predicted and true equation models.\n\nThe paper is relatively well-written and easy to read and follow.\n\nWeaknesses\n---\nIt is not clear to me how much insight is gained using the normalized edit distance (NED) over existing metrics measuring the success of symbolic regression models. The results suggest that the transformer-based model achieves slightly better NED over existing methods, but it is difficult to judge if this improvement is meaningful from a scientific discovery point of view. Even if two equation models are mostly structurally similar (as measured by NED), they may differ by one or two elements that could perceivably result in a very different equation model not representative of the actual ground truth equation.\n\nThe experimental results are not very thorough, and only provide results for a small number of metrics for a small number of baselines. Due to the difficulty evaluating the merit of equation-based scientific discovery, perhaps a metric based on manual inspection with domain experts could give some insight into whether or not the proposed metric (and existing metrics) are well-aligned with how humans judge how close a generated equation is to representing an actual scientific equation.\n\nWhen proposing a new dataset, the authors should consider documenting their dataset thoroughly using, for instance, \"Datasheets for Datasets\" (Gebru et al. 2021) or a dataset nutrition label (Holland et al. 2018).\n\nMinor Weaknesses\n---\nConsider using \"\\citet\" when using citations as a noun.\n\nThe solution rate metric is not clearly described.\n\nWhy is accuracy and NED not shown in the comparison between the FSRD and SRSD dataset (Table S16)?", "clarity,_quality,_novelty_and_reproducibility": "The proposed SRSD datasets expand the FSRD database and appear novel and potentially useful to the scientific discovery via symbolic regression community, however the clarity of the paper could be improved somewhat.", "summary_of_the_review": "Although the paper proposes a potentially valuable collection of SRSD datasets and a new evaluation metric measuring normalized edit distance, the dataset is not well-documented and it is unclear whether the new evaluation metric provides significantly meaningful insights in terms of scientific discovery compared to existing evaluation metrics.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666037836220}], "openreview_url": "https://openreview.net/forum?id=i2e2wqt0nAI", "arxiv_id": "2206.10540", "paper_pdf": "papers/i2e2wqt0nAI.pdf", "paper_pdf_sha256": "e5c84c0fd791d9ffb63c348cf04e7aafca6d7dcf0afe4b9bc64b1620eb1bfbd5", "paper_pdf_bytes": 665887, "paper_pdf_source": "openreview", "code_url": "https://github.com/omron-sinicx/srsd-benchmark", "code_repository": "omron-sinicx/srsd-benchmark", "code_commit": "7d00b45d56250717ac08a256dc9f79b836d61027", "code_archive": "repos/i2e2wqt0nAI.zip", "code_archive_sha256": "eb46e4fcdb853a2741d424d1c1303462cc2b1e7e0b16b005b6eaad6e05063287", "code_archive_bytes": 102944, "code_file_count": 31, "code_extensions": {".py": 30, ".ipynb": 1}, "github_disk_usage_kb": 109, "github_languages": {"Python": 220721, "Jupyter Notebook": 14834}, "github_archived": false, "github_pushed_at": "2024-03-05T07:26:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rethinking-symbolic-regression-datasets-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Le8fg2ppDSv", "year": 2022, "status": "rejected", "title": "HydraSum - Disentangling Stylistic Features in Text Summarization using Multi-Decoder Models", "authors": ["Tanya Goyal", "Nazneen Rajani", "Wenhao Liu", "Wojciech Maciej Kryscinski"], "authorids": ["~Tanya_Goyal1", "~Nazneen_Rajani1", "~Wenhao_Liu1", "~Wojciech_Maciej_Kryscinski1"], "authors_source": "OpenReview API", "abstract": "Existing abstractive summarization models lack explicit control mechanisms that would allow users to influence the stylistic features of the model outputs. This results in generating generic summaries that do not cater to the users needs or preferences. To address this issue we introduce HydraSum, a new summarization architecture that extends the single decoder framework of current models, e.g. BART, to a mixture-of-experts version consisting of multiple decoders. Our proposed model encourages each expert, i.e. decoder, to learn and generate stylistically-distinct summaries along dimensions such as abstractiveness, length, specificity, and others. At each time step, HydraSum employs a gating mechanism that decides the contribution of each individual decoder to the next token's output probability distribution. Through experiments on three summarization datasets (CNN, Newsroom, XSum), we demonstrate that this gating mechanism automatically learns to assign contrasting summary styles to different HydraSum decoders under the standard training objective without the need for additional supervision. We further show that a guided version of the training process can explicitly govern which summary style is partitioned between decoders, e.g. high abstractiveness vs. low abstractiveness or high specificity vs. low specificity, and also increase the stylistic-difference between individual decoders. Finally, our experiments demonstrate that our decoder framework is highly flexible: during inference, we can sample from individual decoders or mixtures of different subsets of the decoders to yield a diverse set of summaries and enforce single- and multi-style control over summary generation.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "NcHVy6Pcb4o", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2373/Reviewer_3AAA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper propose using a multidecoder architecture to to control the style of output summarization. In the designed model, each decoder learns and generates stylisticall-distinct summaries. The contribution of each experts are controlled by a gating mechanism. This multidecoder model can be trained without supervision or with guidance. This model is tested on CNN, NEWSROOM and XSum. When trained unsupervised, different decoders learn to produce summaries with different abstractiveness and specificity. The guided training shows the capability of asigning certain styles to a specific encoder. ", "review_text": "==== strengths ==== \n\nThe task of controllable generation for summarization is well motivated. The finding that different decoders indeed capture different level of abstractiveness and specificity is interesting. The paper is clearly written and the methods are experiments are easy to understand. \n\n==== weaknesses ==== \n\nMy biggest concern with this work is lacking of good evaluation metrics. For the same text input, the model is supposed to control all high-quality summaries with distinct styles. To achieve that, a summarization dataset with multiple stylistically distinct references are needed. Lacking such a dataset, human evaluation would shine some light, which is missing in this work. In the guided training session, what is the rouge score of generated summaries? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper propose using a multidecoder architecture to to control the style of output summarization. In the designed model, each decoder learns and generates stylisticall-distinct summaries. The contribution of each experts are controlled by a gating mechanism. This multidecoder model can be trained without supervision or with guidance. This model is tested on CNN, NEWSROOM and XSum. When trained unsupervised, different decoders learn to produce summaries with different abstractiveness and specificity. The guided training shows the capability of asigning certain styles to a specific encoder. ", "main_review": "==== strengths ==== \n\nThe task of controllable generation for summarization is well motivated. The finding that different decoders indeed capture different level of abstractiveness and specificity is interesting. The paper is clearly written and the methods are experiments are easy to understand. \n\n==== weaknesses ==== \n\nMy biggest concern with this work is lacking of good evaluation metrics. For the same text input, the model is supposed to control all high-quality summaries with distinct styles. To achieve that, a summarization dataset with multiple stylistically distinct references are needed. Lacking such a dataset, human evaluation would shine some light, which is missing in this work. In the guided training session, what is the rouge score of generated summaries? ", "summary_of_the_review": "The paper presented a novel architecture for style control in abstractive summarization. However, without evaluating the model on a dataset where each input is associated with multiple references of different styles or running human evaluation, the effectiveness of the style control and the quality of generated summaries are not clear. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635931968137}, {"id": "KKdrrm7-45h", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2373/Reviewer_5qbJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a neural sequence-to-sequence (Seq2Seq) model, called HYDRASUM for text summarization. HYDRASUM incorporates multiple decoders where each decoder learns and generates stylistically distinct summaries along dimensions such as abstractiveness, specificity, and others. HYDRASUM is fundamentally built upon the notion of mixture-of-experts (MoE), where a gating mechanism decides the contribution of each expert (individual decoder) to the next token's output probability distribution. The paper demonstrates that HYDRASUM automatically learns to generate contrasting summary styles using each decoder. The paper further proposes a \"guided\" training scheme that explicitly govern which summary style is partitioned between decoders, e.g. high abstractiveness vs. low abstractiveness or high specificity vs. low specificity, and also increase the stylistic differences between individual decoders. The paper also show that HYDRASUM is flexible, during inference, it can be controlled to generate a diverse set of summaries by sampling from individual decoders or mixtures of different subsets of the decoders.", "review_text": "**Novelty**\n\nThe fundamental idea of the paper to use multiple decoders and a gating mechanism to improve abstractiveness in model generated summaries is previously explored in [1]. However, the \"guided\" training scheme is interesting, specially the way the training data is partitioned based on different features.\n\n[1] Improving Abstraction in Text Summarization, EMNLP 2018.\n\n**Strengths**\n\n- Good writing. I enjoyed reading the paper.\n- Thorough experiments; the paper covered a broad range of aspects to evaluate the proposed model.\n- Overall the findings of the paper is interesting, and it will help push future works in this direction.\n\n**Weaknessess**\n\n- Limited technical novelty.\n- As the Table 1 suggests, the overall performance of the proposed model does not surpass the baseline in 2 out of 3 evaluation dataset. This raises a key question, in what scenarios, the proposed model would be valuable? Are there any particular use cases of this model? I do not see any discussion on that.\n\n**Questions**\n\n- Why 2-gram overlapping is used as a measure for abstractiveness? What is the motivation?\n- If the feature refers to specificity (as in section 3.2), what is the criteria to split the training dataset?\n- Is it possible to perform human evaluation to judge the diversity of the generated summaries?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a neural sequence-to-sequence (Seq2Seq) model, called HYDRASUM for text summarization. HYDRASUM incorporates multiple decoders where each decoder learns and generates stylistically distinct summaries along dimensions such as abstractiveness, specificity, and others. HYDRASUM is fundamentally built upon the notion of mixture-of-experts (MoE), where a gating mechanism decides the contribution of each expert (individual decoder) to the next token's output probability distribution. The paper demonstrates that HYDRASUM automatically learns to generate contrasting summary styles using each decoder. The paper further proposes a \"guided\" training scheme that explicitly govern which summary style is partitioned between decoders, e.g. high abstractiveness vs. low abstractiveness or high specificity vs. low specificity, and also increase the stylistic differences between individual decoders. The paper also show that HYDRASUM is flexible, during inference, it can be controlled to generate a diverse set of summaries by sampling from individual decoders or mixtures of different subsets of the decoders.", "main_review": "**Novelty**\n\nThe fundamental idea of the paper to use multiple decoders and a gating mechanism to improve abstractiveness in model generated summaries is previously explored in [1]. However, the \"guided\" training scheme is interesting, specially the way the training data is partitioned based on different features.\n\n[1] Improving Abstraction in Text Summarization, EMNLP 2018.\n\n**Strengths**\n\n- Good writing. I enjoyed reading the paper.\n- Thorough experiments; the paper covered a broad range of aspects to evaluate the proposed model.\n- Overall the findings of the paper is interesting, and it will help push future works in this direction.\n\n**Weaknessess**\n\n- Limited technical novelty.\n- As the Table 1 suggests, the overall performance of the proposed model does not surpass the baseline in 2 out of 3 evaluation dataset. This raises a key question, in what scenarios, the proposed model would be valuable? Are there any particular use cases of this model? I do not see any discussion on that.\n\n**Questions**\n\n- Why 2-gram overlapping is used as a measure for abstractiveness? What is the motivation?\n- If the feature refers to specificity (as in section 3.2), what is the criteria to split the training dataset?\n- Is it possible to perform human evaluation to judge the diversity of the generated summaries?\n", "summary_of_the_review": "The novelty of the proposed method is thin. However, the paper presents rigorous experiments and evaluation to validate the main claim. Performing human evaluation to judge the quality of the generated summaries would be beneficial.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635892370777}, {"id": "Uwe8EccWmu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2373/Reviewer_kxHp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new architecture for controllable text summarization, leveraging multiple decoders (with the bottom decoder layers  shared across decoders) with the contributions of each decoder controlled by a gating mechanism. This model is studied in two gating settings: unguided (where the network learns the gating weights) and guided (where gating is controlled manually). The networks are shown to allow some stylistic controllability, most notably in terms of abstractiveness and specificity.\n", "review_text": "Strengths:\n- the paper is well motivated and tackles a hard and long-standing problem with seq2seq models: diversity and controllability.\n- the approach is fairly simple, interesting and to the best of my knowledge, novel. \n- the paper is well structured, clearly explained and free of orthographic/grammatical issues.\n\nWeaknesses:\n- It is surprising to use beam search as the only baseline algorithm for generating >1 summaries as it is a poor way to generate diverse outputs. It would have been more interesting to compare with diverse beam search, nucleus sampling, etc, possibly in addition to beam search. Can these algorithms be added to the analysis? Even if metric increases are more modest, the proposed approach still provides better controllability which is interesting in its own right.\n- Can the guided setting allow the model to manage gating? Otherwise, this makes the guided setting dependent on manual inputs while not letting the model learn dataset-specific gating. Given the authors used heuristics to partition the datasets, could these partitions be used as a supervision signal for gating? If not, why not?\n- The proposed approach has a non-negligible computational cost, compared to using a single decoder, that it would have been interesting to discuss. \n- The literature review is fairly short and to the point. Diversity and controllability have been large research areas in the past few years and it would have been helpful to, at the very least, provide more details on the cited papers and contrast them with the proposed approach.\n\nNits:\n- In Figure 1, the baseline summaries are likely generated using beam search. It would be good to mention it explicitly.\n- In Table 2, it would be clearer to mention R1/R2/RL as was done in Table 1.\n- In Table 2, R2 for XSum is lower than that of the baseline, so it should probably not be written in bold.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new architecture for controllable text summarization, leveraging multiple decoders (with the bottom decoder layers  shared across decoders) with the contributions of each decoder controlled by a gating mechanism. This model is studied in two gating settings: unguided (where the network learns the gating weights) and guided (where gating is controlled manually). The networks are shown to allow some stylistic controllability, most notably in terms of abstractiveness and specificity.\n", "main_review": "Strengths:\n- the paper is well motivated and tackles a hard and long-standing problem with seq2seq models: diversity and controllability.\n- the approach is fairly simple, interesting and to the best of my knowledge, novel. \n- the paper is well structured, clearly explained and free of orthographic/grammatical issues.\n\nWeaknesses:\n- It is surprising to use beam search as the only baseline algorithm for generating >1 summaries as it is a poor way to generate diverse outputs. It would have been more interesting to compare with diverse beam search, nucleus sampling, etc, possibly in addition to beam search. Can these algorithms be added to the analysis? Even if metric increases are more modest, the proposed approach still provides better controllability which is interesting in its own right.\n- Can the guided setting allow the model to manage gating? Otherwise, this makes the guided setting dependent on manual inputs while not letting the model learn dataset-specific gating. Given the authors used heuristics to partition the datasets, could these partitions be used as a supervision signal for gating? If not, why not?\n- The proposed approach has a non-negligible computational cost, compared to using a single decoder, that it would have been interesting to discuss. \n- The literature review is fairly short and to the point. Diversity and controllability have been large research areas in the past few years and it would have been helpful to, at the very least, provide more details on the cited papers and contrast them with the proposed approach.\n\nNits:\n- In Figure 1, the baseline summaries are likely generated using beam search. It would be good to mention it explicitly.\n- In Table 2, it would be clearer to mention R1/R2/RL as was done in Table 1.\n- In Table 2, R2 for XSum is lower than that of the baseline, so it should probably not be written in bold.", "summary_of_the_review": "Overall this is an interesting paper proposing a new approach to stylistic controllability for neural text summarization. To the best of my knowledge, this approach is novel. I believe this work is interesting as is, but were surprised by some of the experimental settings that I would like to see clarified by the authors.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635786702509}, {"id": "ZYMTd6DmTRU", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2373/Reviewer_aQkL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes an architecture for controlled summarization, allowing some type of aspect-based summarization, as in - for example - \"I want long/detailed summary\". The control proposed here is on what the authors call _stylistic_ features (which also includes _quality_, arguably not a stylistic aspect) and not other aspect (as in for example: give a summary of Darius the Great focusing on his earlier years).\n\nThey obtain this by using a Transformer (BART-like) seq2seq model, but where the last layers of the decoder are multiplied and considered as independent expert which are combined in the end. That combination consists in a weighted average (no non-linearity) of the last softmax layers.\n\nThe proposed solution is simple (good!) although not very elegant. The analysis focuses on evaluating the control capacity and diversity with respect to other (smaller) architectures.", "review_text": "Strengths:\n1. This provides an alternative to controlled generation. Standard solutions so far include control tokens (not cited, see eg [1]) or a-posteriori filtering (cited) which have drawbacks (less flexibility for the former, very expensive for the latter). \n2. The idea is simple (Methodology section takes just over 1 page)\n3. The paper has an in-depth analysis part, showing convincingly that stylistic features are indeed captured and can be used for control. Most importantly (I consider this the main novelty of this approach), they can be combined. Unfortunately, there is not much space left for detailed analysis on this generalization capacity of the proposed approach.\n\nWeaknesses.\n1. The proposed solution only acts on the single token level, as it intervenes by combining experts that predict individual tokens. This is an inherent weakness of auto-regressive models, although recent alternatives have been proposed (see [2] and references therein)\n2. The compared baselines do not include any of the existing approaches (like those mentioned in Strength1). Comparing one controlled model against other non-controllable models and not considering how the problem was tackled so far is a pity. The compared models are (probably) much smaller as they do not incur in the extra cost of the additional adapter layers. For fair comparisons, the model sizes should be equivalent\n\n\nAs a potential upper bound for ROUGE (page 6), instead of picking topK, one possibility would be to select for each document the decoder which obtains the best summary.\n\n[1] Self-Supervised and Controlled Multi-Document Opinion Summarization. EACL 2021\n[2] A Step-Wise Weighting Approach for Controllable Text Generation. ICLR 2022 submission (K8HF8tTQ-4i)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an architecture for controlled summarization, allowing some type of aspect-based summarization, as in - for example - \"I want long/detailed summary\". The control proposed here is on what the authors call _stylistic_ features (which also includes _quality_, arguably not a stylistic aspect) and not other aspect (as in for example: give a summary of Darius the Great focusing on his earlier years).\n\nThey obtain this by using a Transformer (BART-like) seq2seq model, but where the last layers of the decoder are multiplied and considered as independent expert which are combined in the end. That combination consists in a weighted average (no non-linearity) of the last softmax layers.\n\nThe proposed solution is simple (good!) although not very elegant. The analysis focuses on evaluating the control capacity and diversity with respect to other (smaller) architectures.", "main_review": "Strengths:\n1. This provides an alternative to controlled generation. Standard solutions so far include control tokens (not cited, see eg [1]) or a-posteriori filtering (cited) which have drawbacks (less flexibility for the former, very expensive for the latter). \n2. The idea is simple (Methodology section takes just over 1 page)\n3. The paper has an in-depth analysis part, showing convincingly that stylistic features are indeed captured and can be used for control. Most importantly (I consider this the main novelty of this approach), they can be combined. Unfortunately, there is not much space left for detailed analysis on this generalization capacity of the proposed approach.\n\nWeaknesses.\n1. The proposed solution only acts on the single token level, as it intervenes by combining experts that predict individual tokens. This is an inherent weakness of auto-regressive models, although recent alternatives have been proposed (see [2] and references therein)\n2. The compared baselines do not include any of the existing approaches (like those mentioned in Strength1). Comparing one controlled model against other non-controllable models and not considering how the problem was tackled so far is a pity. The compared models are (probably) much smaller as they do not incur in the extra cost of the additional adapter layers. For fair comparisons, the model sizes should be equivalent\n\n\nAs a potential upper bound for ROUGE (page 6), instead of picking topK, one possibility would be to select for each document the decoder which obtains the best summary.\n\n[1] Self-Supervised and Controlled Multi-Document Opinion Summarization. EACL 2021\n[2] A Step-Wise Weighting Approach for Controllable Text Generation. ICLR 2022 submission (K8HF8tTQ-4i)", "summary_of_the_review": "A valuable contribution to controlled generation. The evaluation focuses on the strong point of the models (more diversity) and is somehow unfair (against non-controllable models that are smaller). In this sense it is not clear if it is better or just different than existing methods.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635781635054}], "openreview_url": "https://openreview.net/forum?id=Le8fg2ppDSv", "arxiv_id": "2110.04400", "paper_pdf": "papers/Le8fg2ppDSv.pdf", "paper_pdf_sha256": "fec894c1926bfd28787f3706a089341c9e3cb01edabc31e630aea07246e1ce04", "paper_pdf_bytes": 2375098, "paper_pdf_source": "openreview", "code_url": "https://github.com/salesforce/hydra-sum", "code_repository": "salesforce/hydra-sum", "code_commit": "0cb4ea1be9a18fc227406e3a23a41f8c05853267", "code_archive": "repos/Le8fg2ppDSv.zip", "code_archive_sha256": "9688b9d5fc8313897cdfd1a973ed2f6b63dba0e421110d33526ba1d24bf94fc3", "code_archive_bytes": 74103, "code_file_count": 16, "code_extensions": {".py": 15, ".sh": 1}, "github_disk_usage_kb": 120, "github_languages": {"Python": 391463, "Shell": 860}, "github_archived": true, "github_pushed_at": "2025-05-01T17:29:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hydrasum-disentangling-stylistic-features-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gW8n0uD6rl", "year": 2021, "status": "rejected", "title": "Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data", "authors": ["Sindy Löwe", "David Madras", "Richard Zemel", "Max Welling"], "authorids": ["~Sindy_Löwe1", "~David_Madras1", "~Richard_Zemel1", "~Max_Welling1"], "authors_source": "OpenReview API", "abstract": "Standard causal discovery methods must fit a new model whenever they encounter samples from a new underlying causal graph. However, these samples often share relevant information - for instance, the dynamics describing the effects of causal relations - which is lost when following this approach. We propose Amortized Causal Discovery, a novel framework that leverages such shared dynamics to learn to infer causal relations from time-series data. This enables us to train a single, amortized model that infers causal relations across samples with different underlying causal graphs, and thus makes use of the information that is shared. We demonstrate experimentally that this approach, implemented as a variational model, leads to significant improvements in causal discovery performance, and show how it can be extended to perform well under hidden confounding.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "QYQca3YOlC4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper851/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper makes an observation that signal dynamics common to a class of causal systems may contain strong information to enable the use of the encoder of the Neural Relational Inference (2018) for extracting (Granger) causal graphs. It minimally extends the NRI model with an empty edge type and demonstrates that the observation holds in a few cases of dynamical systems. Additionally, to handle the unobserved common causes the paper shows improvements in ROCAUC when the encoder is modified to model them directly.\n\nThe paper discusses a potentially interesting use of the Neural Relational Inference for (Granger) causal discovery in multivariate time series. Although lacking in detail, it is very clearly written. The proposed method of causal discovery for Markov order one system does demonstrate impressive performance on smooth dynamical systems.\n\nPossibly, the greatest weakness of the NRI approach in this formulation is the seeming reliance on relatively smooth noiseless dynamics. While for the NRI in the original paper (2018) it was within the proposed scope, the claimed use of NRI in this paper is not fully supported by experiments. All of them are from a smooth dynamical system. The Particles experiment is no different from NRI (2018) with the same reported AUC result. Even the most realistic dynamical system from Smith et al. 2011, which is also quite simplistic by the way, still model hemodynamic lag and thus a relatively smooth signal. Notably, this last less smooth of all presented signal is where the NRI model performs the worst.\n\n-   How would NRI in this proposed application behave if we use a still simplistic but potentially less smooth VAR model? Would be good to test it with different SNR.\n-   In general, even in dynamical systems used for method demonstration, what happens at various noise levels?\n-   How would it behave when the assumption of the model is slightly violated and the signal is generated by an SVAR model?\n\nGiven the above limitations, it is unclear if the paper presents something substantially novel/interesting. Technically it is still the NRI model of 2018, so not much contribution there. That would not be a problem if the paper would manage to indeed demonstrate that the model is useful for causal discovery in multivariate time-series. The experiments are limited to very specific kinds of signals and thus not convincing of the approach generality.\n\nIt is not immediately clear from preliminaries, whether the proposed model is capable to recover relations of a Markov order higher than one. The described \"summary graph\" as a compressed representation only holds true to the Markov order one underlying dynamic Bayesian network-like structure (see D. Danks and S. Plis. Learning causal structure from undersampled time series. In NIPS 2013 Workshop on Causality, 2013.), yet the description of the model is a bit vague and may indicate that the method works with the full unrolled graph and then folds it into Markov order one. This part, despite the extensive use of notation, is not clear. It would be simpler to say that indeed an unrolled representation of a graph of limited Markov order is used. It is not clear though if there are any restrictions on the Markov order as discussed since the encoded graph holds just Markov order one.\n\nWhile the problem of unobserved common causes is discussed, only a single type is considered - fully missing random variables (or particles). Yet, a simple mismatch in measurement and causal time scales will violate some of the assumptions (e.g. generate samples from the VAR model and drop every other time sample for all variables and the model can be fit with SVAR and not VAR). See:\n\n-   M. Gong, K. Zhang, B. Schoelkopf, D. Tao, and P. Geiger. Discovering temporal causal relations from subsampled data. In Proc. ICML, volume 37 of JMLR W&CP, pages 1898–1906. JMLR.org, 2015\n-   S. Plis, D. Danks, C. Freeman, and V. Calhoun. Rate-agnostic (causal) structure learning. In Proc. NIPS, pages 3285–3293. Curran Associates, Inc., 2015\n\nFrom the paper it is unclear what encoder and decoder models were used and can be used to perform the experiments or use the approach. How does the graph enter the decoder together with a variable-length input? If the Decoder is MLP the input layer has a fixed size. Possibly, the MLP is fed one sample at a time with a vector representation of the encoded graph, but that contradicts the paper description. Generally unclear. Supplement mentions that the models are exactly the same as for the NRI paper, but this is still not enough information and does not make the paper self-contained.\n\nFrom the description of experiments it is unclear how many random variables were used and how many can the NRI handle.\n\nWhat was the experiment in Figure 3? How many particles/random variables? Was the training done on a single causal generative graph while testing performed on a variety of different unseen graphs? Was it a random sample of a large number of different graphs for both testing and training? Is NRI even capable of learning transferable information from data generated from a single graph?\n\n-   Note, reference P. Spirtes, C. N. Glymour, R. Scheines, and D. Heckerman. Causation, prediction, and search. MIT Press, 2000. - Heckerman is *not* a co-author of [this book](https://mitpress.mit.edu/books/causation-prediction-and-search-second-edition)\n-   Last paragraph of page 3 $x\\in \\mathbb{X}$, the $\\mathbb{X}$ is not defined.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Causal discovery based on shared dynamics is exciting but does this method work on real(istic) data in the presence of noise?", "review": "The paper makes an observation that signal dynamics common to a class of causal systems may contain strong information to enable the use of the encoder of the Neural Relational Inference (2018) for extracting (Granger) causal graphs. It minimally extends the NRI model with an empty edge type and demonstrates that the observation holds in a few cases of dynamical systems. Additionally, to handle the unobserved common causes the paper shows improvements in ROCAUC when the encoder is modified to model them directly.\n\nThe paper discusses a potentially interesting use of the Neural Relational Inference for (Granger) causal discovery in multivariate time series. Although lacking in detail, it is very clearly written. The proposed method of causal discovery for Markov order one system does demonstrate impressive performance on smooth dynamical systems.\n\nPossibly, the greatest weakness of the NRI approach in this formulation is the seeming reliance on relatively smooth noiseless dynamics. While for the NRI in the original paper (2018) it was within the proposed scope, the claimed use of NRI in this paper is not fully supported by experiments. All of them are from a smooth dynamical system. The Particles experiment is no different from NRI (2018) with the same reported AUC result. Even the most realistic dynamical system from Smith et al. 2011, which is also quite simplistic by the way, still model hemodynamic lag and thus a relatively smooth signal. Notably, this last less smooth of all presented signal is where the NRI model performs the worst.\n\n-   How would NRI in this proposed application behave if we use a still simplistic but potentially less smooth VAR model? Would be good to test it with different SNR.\n-   In general, even in dynamical systems used for method demonstration, what happens at various noise levels?\n-   How would it behave when the assumption of the model is slightly violated and the signal is generated by an SVAR model?\n\nGiven the above limitations, it is unclear if the paper presents something substantially novel/interesting. Technically it is still the NRI model of 2018, so not much contribution there. That would not be a problem if the paper would manage to indeed demonstrate that the model is useful for causal discovery in multivariate time-series. The experiments are limited to very specific kinds of signals and thus not convincing of the approach generality.\n\nIt is not immediately clear from preliminaries, whether the proposed model is capable to recover relations of a Markov order higher than one. The described \"summary graph\" as a compressed representation only holds true to the Markov order one underlying dynamic Bayesian network-like structure (see D. Danks and S. Plis. Learning causal structure from undersampled time series. In NIPS 2013 Workshop on Causality, 2013.), yet the description of the model is a bit vague and may indicate that the method works with the full unrolled graph and then folds it into Markov order one. This part, despite the extensive use of notation, is not clear. It would be simpler to say that indeed an unrolled representation of a graph of limited Markov order is used. It is not clear though if there are any restrictions on the Markov order as discussed since the encoded graph holds just Markov order one.\n\nWhile the problem of unobserved common causes is discussed, only a single type is considered - fully missing random variables (or particles). Yet, a simple mismatch in measurement and causal time scales will violate some of the assumptions (e.g. generate samples from the VAR model and drop every other time sample for all variables and the model can be fit with SVAR and not VAR). See:\n\n-   M. Gong, K. Zhang, B. Schoelkopf, D. Tao, and P. Geiger. Discovering temporal causal relations from subsampled data. In Proc. ICML, volume 37 of JMLR W&CP, pages 1898–1906. JMLR.org, 2015\n-   S. Plis, D. Danks, C. Freeman, and V. Calhoun. Rate-agnostic (causal) structure learning. In Proc. NIPS, pages 3285–3293. Curran Associates, Inc., 2015\n\nFrom the paper it is unclear what encoder and decoder models were used and can be used to perform the experiments or use the approach. How does the graph enter the decoder together with a variable-length input? If the Decoder is MLP the input layer has a fixed size. Possibly, the MLP is fed one sample at a time with a vector representation of the encoded graph, but that contradicts the paper description. Generally unclear. Supplement mentions that the models are exactly the same as for the NRI paper, but this is still not enough information and does not make the paper self-contained.\n\nFrom the description of experiments it is unclear how many random variables were used and how many can the NRI handle.\n\nWhat was the experiment in Figure 3? How many particles/random variables? Was the training done on a single causal generative graph while testing performed on a variety of different unseen graphs? Was it a random sample of a large number of different graphs for both testing and training? Is NRI even capable of learning transferable information from data generated from a single graph?\n\n-   Note, reference P. Spirtes, C. N. Glymour, R. Scheines, and D. Heckerman. Causation, prediction, and search. MIT Press, 2000. - Heckerman is *not* a co-author of [this book](https://mitpress.mit.edu/books/causation-prediction-and-search-second-edition)\n-   Last paragraph of page 3 $x\\in \\mathbb{X}$, the $\\mathbb{X}$ is not defined.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604013414745}, {"id": "lH959uMqOtu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper851/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors proposed a framework called Amortized Causal Discovery (ACD) for recovering causal relationships in time series where samples are generated from models with different underlying causal graphs but shared dynamics. This framework is applicable in settings such as modeling neural spiking trains where the dynamics of how neurons react to the activities of other neurons remain the same. In the proposed framework, they considered a causal discovery encoder $f_{\\theta}$, which tries to extract causal relationships and map it to a latent space. Moreover, there is a dynamic decoder $f_{\\phi}$, which provides one-step predictions. The proposed architecture is similar to a variational auto-encoder and the encoder part is based on graph neural networks.\n\nAlthough experimental results show that a considerable improvement in some dataset (even in the presence of a latent variable), there are some concerns about the contributions of this work and also its presentation:\n\n1- About the two main assumptions: The authors considered two assumptions on the causal model of time series (no instantaneous effect and invariant causal graph in time). We know that both assumptions are not satisfied in many practical applications. For instance, we have instantaneous effects if the sampling rate is low. Moreover, most causal mechanisms are changing in time.  As an example, in (Pfister et al., 2019), this problem has been addressed by detecting change points. I am not sure whether that approach is promising or not but it is expected from an experimental work to handle a more general setting.\n\n2- About the shared dynamic: Is it possible to check whether the assumption of shared dynamic is valid in the observed data? What if there are some small changes in the dynamics across different training samples.\n\n3- In Eq. (12-14), it seems that we are using the values of other time series at only time $t$ to give a one-step prediction of a particular time series. Why do not we consider the history of previous values from time $1$ to $t$?\n\n4- About handling latent confounders: As noted by the authors, recovering causal relationships is challenging in the presence of latent variables. Although experimental results show that the proposed method has better performance than baselines, it is still unclear whether there are some theoretical guarantees in recovering the correct causal graph using the proposed method.\n\n5- The presentation of the paper could be improved:\n*Please provide the exact definition of $\\hat{\\mathcal{G}}_x$, $\\mathbb{G}$, $\\mathbb{X}$, $r$,....\n*Please explain the two methods (the amortized encoder and TTA) in more detail on page 4.\n\n(Pfister et al., 2019) Pfister, Niklas, Peter Bühlmann, and Jonas Peters. \"Invariant causal prediction for sequential data.\" Journal of the American Statistical Association 114.527 (2019): 1264-1276.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting experimental work but lacking some implementation details and theoretical supports", "review": "The authors proposed a framework called Amortized Causal Discovery (ACD) for recovering causal relationships in time series where samples are generated from models with different underlying causal graphs but shared dynamics. This framework is applicable in settings such as modeling neural spiking trains where the dynamics of how neurons react to the activities of other neurons remain the same. In the proposed framework, they considered a causal discovery encoder $f_{\\theta}$, which tries to extract causal relationships and map it to a latent space. Moreover, there is a dynamic decoder $f_{\\phi}$, which provides one-step predictions. The proposed architecture is similar to a variational auto-encoder and the encoder part is based on graph neural networks.\n\nAlthough experimental results show that a considerable improvement in some dataset (even in the presence of a latent variable), there are some concerns about the contributions of this work and also its presentation:\n\n1- About the two main assumptions: The authors considered two assumptions on the causal model of time series (no instantaneous effect and invariant causal graph in time). We know that both assumptions are not satisfied in many practical applications. For instance, we have instantaneous effects if the sampling rate is low. Moreover, most causal mechanisms are changing in time.  As an example, in (Pfister et al., 2019), this problem has been addressed by detecting change points. I am not sure whether that approach is promising or not but it is expected from an experimental work to handle a more general setting.\n\n2- About the shared dynamic: Is it possible to check whether the assumption of shared dynamic is valid in the observed data? What if there are some small changes in the dynamics across different training samples.\n\n3- In Eq. (12-14), it seems that we are using the values of other time series at only time $t$ to give a one-step prediction of a particular time series. Why do not we consider the history of previous values from time $1$ to $t$?\n\n4- About handling latent confounders: As noted by the authors, recovering causal relationships is challenging in the presence of latent variables. Although experimental results show that the proposed method has better performance than baselines, it is still unclear whether there are some theoretical guarantees in recovering the correct causal graph using the proposed method.\n\n5- The presentation of the paper could be improved:\n*Please provide the exact definition of $\\hat{\\mathcal{G}}_x$, $\\mathbb{G}$, $\\mathbb{X}$, $r$,....\n*Please explain the two methods (the amortized encoder and TTA) in more detail on page 4.\n\n(Pfister et al., 2019) Pfister, Niklas, Peter Bühlmann, and Jonas Peters. \"Invariant causal prediction for sequential data.\" Journal of the American Statistical Association 114.527 (2019): 1264-1276.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603701061381}, {"id": "9m1jKWy51v-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper851/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the problem of learning causal graphs from time-series data. A new framework, called Amortized Causal Discovery (ACD), is proposed based on the observation that samples often share identical or similar dynamics. ACD is then implemented with a variational model, mostly following existing works. Overall, I think it is a very interesting idea to consider shared dynamics and, in some sense, to 'learn a score function' from data, compared with existing score-based methods that usually need to put assumptions on data generations to make the score function sound. However, there is also a key problem, identifibility and consistency, to ACD, for which I do not recommend an acceptance for now. I would be happy to increase my score if authors can address my concerns. \n \nIdentifiability and Consistency: (a) identifiability is an important problem to causal discovery from observational data. I find that only Peters et al, 2013 is cited for such an issue. However, in that work, the identifibility result requires a much stronger condition than the assumption used in this paper. It would be good to discuss this issue in more details. (b) as stated in Section, 4, the proposed approach can be seen a score based method with learned score functions. However, it is unclear when this score function could be 'consistent', i.e., maximizing the score function would lead to a consistent estimate of the true graph (possibly under some model assumptions). (c) Another part, related to this learned score function, is whether the overall approach is consistent. In particular, Eq. (5) has a regularization term (which is often a sparsity term, as stated in the paper). The question is how to pick this penalty term, including its weight, in practice? This is crucial for causal discovery and I do not find that this issue is explained in the current version. \n \nWriting: the overall writing is very good, but could be improved for some places. (a) Eq (12) uses $z_{ij, 0}$ which was not introduced until the next paragraph; (b) it seems that the problem size (how many variables of the problem) is small. It would be good to introduce the problem size explicitly in the main content. (c) as mentioned above, the identifiably and regularization term are not clearly discussed; they are however important to score based causal discovery.\n \nOther questions: (a) why the BIC score used by Chickering 2012 is called 'heuristic score'? (b) Gumbel technique may not give exactly zero or one, so how was the 'no edge' type processed? This part was somewhat omitted. (c) what if there are instantaneous effects? I do not see why ACD does not apply to this case.\n\n*********after reading rebuttal *******\nI have increased my rating, but may still have some concerns. See below.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, but lacking investigation into identifiablity and consistency", "review": "This paper studies the problem of learning causal graphs from time-series data. A new framework, called Amortized Causal Discovery (ACD), is proposed based on the observation that samples often share identical or similar dynamics. ACD is then implemented with a variational model, mostly following existing works. Overall, I think it is a very interesting idea to consider shared dynamics and, in some sense, to 'learn a score function' from data, compared with existing score-based methods that usually need to put assumptions on data generations to make the score function sound. However, there is also a key problem, identifibility and consistency, to ACD, for which I do not recommend an acceptance for now. I would be happy to increase my score if authors can address my concerns. \n \nIdentifiability and Consistency: (a) identifiability is an important problem to causal discovery from observational data. I find that only Peters et al, 2013 is cited for such an issue. However, in that work, the identifibility result requires a much stronger condition than the assumption used in this paper. It would be good to discuss this issue in more details. (b) as stated in Section, 4, the proposed approach can be seen a score based method with learned score functions. However, it is unclear when this score function could be 'consistent', i.e., maximizing the score function would lead to a consistent estimate of the true graph (possibly under some model assumptions). (c) Another part, related to this learned score function, is whether the overall approach is consistent. In particular, Eq. (5) has a regularization term (which is often a sparsity term, as stated in the paper). The question is how to pick this penalty term, including its weight, in practice? This is crucial for causal discovery and I do not find that this issue is explained in the current version. \n \nWriting: the overall writing is very good, but could be improved for some places. (a) Eq (12) uses $z_{ij, 0}$ which was not introduced until the next paragraph; (b) it seems that the problem size (how many variables of the problem) is small. It would be good to introduce the problem size explicitly in the main content. (c) as mentioned above, the identifiably and regularization term are not clearly discussed; they are however important to score based causal discovery.\n \nOther questions: (a) why the BIC score used by Chickering 2012 is called 'heuristic score'? (b) Gumbel technique may not give exactly zero or one, so how was the 'no edge' type processed? This part was somewhat omitted. (c) what if there are instantaneous effects? I do not see why ACD does not apply to this case.\n\n*********after reading rebuttal *******\nI have increased my rating, but may still have some concerns. See below.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603610891657}, {"id": "MC51xMN_aiC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper851/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Introduction\n\nYou should have praise for your first sentence in your second paragraph. It is excellent. A neat summary indeed. Emphasise it if you can.\n\nRefer to figure one in paragraph one to allow the novice reader to get a quick understanding of what you are trying to do. Better yet, move figure one into the first page and put it at the bottom, it is a great eyecatcher.\n\n# Background\n\nN should be italicised in definition 2.1.\n\n# Amortized causal discovery\n\nPresumably like all work of this kind you assume, there are no cycles or self-cycles in the SCM. That is pretty common fare for causal work (and a huge limitation) but it should still be noted at the very top of this section or indeed expanded upon in the preliminaries. I take it thus that you cannot handle SCMs with self-cycles and cycles in the endogenous variables? A far harder problem of course, but just to be clear from this reviewer's point of view.\n\nWhat your second point is saying in the preliminaries is that you assume that your data is sampled from a stationary process. You may want to add that key-word as that says a lot to people from outside this field, and again, emphasises the limitations of this causal discovery (and most in fact) method. It also gives you plenty of scope for future work! Indeed you go on to talk about the summary graph which is by definition only defined for stationary processes.\n\nA common theme throughout this paper is that you just insert figures without referring to them beforehand. This is very confusing. Please stop doing this and re-write the paper so that your figures follow the discussion of them. You introduce an unnecessary break in the flow of the reading otherwise, like between pages 1 and 2 and 3 and 4. \n\nPlease explain what you mean by this: \"Note, that we do not apply a supervised loss on the graph prediction\"\n\nUse \\text{} in your subscripts in your equations, it reads better i.e. $\\hat{\\mathcal{G}}_{\\text{TTA}}$ is preferable.\n\nPlease explain what you mean by \"expresses that there is a directed edge of type e from time-series i to j\" -- what is type e?\n\n## Hidden confounding\n\nThis section could do with a diagram. I find it very confusing. I think you are saying that you allow the decoder to predict the additional variable (in the summary graph?) which confounds _all_ endogenous variables in the summary? Or some? or one? Which is it? What if the confounding variable does not affect all endogenous variables?\n\n# Related work\n\nSolid section. Well done.\n\n# Experiments\n\n## 5.1\n\nIt would be helpful if you could show the inferred adjacency matrix of the particle dataset, alongside the GT adjacency matrix.\n\nThese experiments feel very selective. Your goal is to demonstrate ACD's superiority over the other methods, granted, that's the process and we all know it. However, this does seem like an odd comparison. Take Khanna and Tan, 2020 (not 2019 like you've quoted it): on a Lorenz-96 simulation, they reach an average AUROC of 0.95. Yet here they get 0.607, and this ought to be because you have generated asymmetric connectivity matrices which renders a structure which eSRU cannot handle. But how we do know? You haven't run ACD on say the Lorenz-96 system which is pretty common fare in causal discovery. If your model does not outperform SOTA methods in that space you may have to relax some of your claims. What is more, Tank et al do very well on Lorenz-96 too. They are not particularly good at the DREAM-3IN SILICO NETWORK INFERENCE CHALLENGE, and that's where ACD could shine through, rather than taking these models 'out of context' as it were. Point being, please demonstrate ACD on the _simpler_ operating space where the likes of NGC and eSRU live. You appear to have started that process with Table 2 but that is a complex setting, perhaps too complex to understand where ACD is failing. Again, I advise starting simple with the standard Lorenz-96 model just to ensure that your method performs well in that easy space.\n\n# 5.1\n\nIt is not okay for you to bold ACD's performance value in table 2. Remove it. MI is clearly higher and putting anything else is bold is confusing.\n\n## 5.2.2\n\nThis experiment is very interesting. The figures are good too. My advice would be to make the axis text larger, and set alpha=0.75 or something when you plot the prediction -- that way the reader can better see how the prediction tracks the GT.\n\nHowever, I am a bit confused about the deterministic chaos that you briefly mention at the start of 5.2.1 -- the parameter setting you use in this section (5.2.2) presumably does not give rise to deterministic chaos?\n\n---\n\nNone of your experiments contains fully identifiable noise models from what I can tell? You mention \"noise\" twice in the manuscript. None is in the experimental section. Now since you are only using simulations, I find it odd that you do not discuss this nor ACD's robustness against noise. As you are fully aware of course, most of these models break-down in the presence of a low SNR. And for this to be useful at all in the 'real' world, how does ACD fare?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors leverage the potential information shared across samples to illicit causal discovery.", "review": "# Introduction\n\nYou should have praise for your first sentence in your second paragraph. It is excellent. A neat summary indeed. Emphasise it if you can.\n\nRefer to figure one in paragraph one to allow the novice reader to get a quick understanding of what you are trying to do. Better yet, move figure one into the first page and put it at the bottom, it is a great eyecatcher.\n\n# Background\n\nN should be italicised in definition 2.1.\n\n# Amortized causal discovery\n\nPresumably like all work of this kind you assume, there are no cycles or self-cycles in the SCM. That is pretty common fare for causal work (and a huge limitation) but it should still be noted at the very top of this section or indeed expanded upon in the preliminaries. I take it thus that you cannot handle SCMs with self-cycles and cycles in the endogenous variables? A far harder problem of course, but just to be clear from this reviewer's point of view.\n\nWhat your second point is saying in the preliminaries is that you assume that your data is sampled from a stationary process. You may want to add that key-word as that says a lot to people from outside this field, and again, emphasises the limitations of this causal discovery (and most in fact) method. It also gives you plenty of scope for future work! Indeed you go on to talk about the summary graph which is by definition only defined for stationary processes.\n\nA common theme throughout this paper is that you just insert figures without referring to them beforehand. This is very confusing. Please stop doing this and re-write the paper so that your figures follow the discussion of them. You introduce an unnecessary break in the flow of the reading otherwise, like between pages 1 and 2 and 3 and 4. \n\nPlease explain what you mean by this: \"Note, that we do not apply a supervised loss on the graph prediction\"\n\nUse \\text{} in your subscripts in your equations, it reads better i.e. $\\hat{\\mathcal{G}}_{\\text{TTA}}$ is preferable.\n\nPlease explain what you mean by \"expresses that there is a directed edge of type e from time-series i to j\" -- what is type e?\n\n## Hidden confounding\n\nThis section could do with a diagram. I find it very confusing. I think you are saying that you allow the decoder to predict the additional variable (in the summary graph?) which confounds _all_ endogenous variables in the summary? Or some? or one? Which is it? What if the confounding variable does not affect all endogenous variables?\n\n# Related work\n\nSolid section. Well done.\n\n# Experiments\n\n## 5.1\n\nIt would be helpful if you could show the inferred adjacency matrix of the particle dataset, alongside the GT adjacency matrix.\n\nThese experiments feel very selective. Your goal is to demonstrate ACD's superiority over the other methods, granted, that's the process and we all know it. However, this does seem like an odd comparison. Take Khanna and Tan, 2020 (not 2019 like you've quoted it): on a Lorenz-96 simulation, they reach an average AUROC of 0.95. Yet here they get 0.607, and this ought to be because you have generated asymmetric connectivity matrices which renders a structure which eSRU cannot handle. But how we do know? You haven't run ACD on say the Lorenz-96 system which is pretty common fare in causal discovery. If your model does not outperform SOTA methods in that space you may have to relax some of your claims. What is more, Tank et al do very well on Lorenz-96 too. They are not particularly good at the DREAM-3IN SILICO NETWORK INFERENCE CHALLENGE, and that's where ACD could shine through, rather than taking these models 'out of context' as it were. Point being, please demonstrate ACD on the _simpler_ operating space where the likes of NGC and eSRU live. You appear to have started that process with Table 2 but that is a complex setting, perhaps too complex to understand where ACD is failing. Again, I advise starting simple with the standard Lorenz-96 model just to ensure that your method performs well in that easy space.\n\n# 5.1\n\nIt is not okay for you to bold ACD's performance value in table 2. Remove it. MI is clearly higher and putting anything else is bold is confusing.\n\n## 5.2.2\n\nThis experiment is very interesting. The figures are good too. My advice would be to make the axis text larger, and set alpha=0.75 or something when you plot the prediction -- that way the reader can better see how the prediction tracks the GT.\n\nHowever, I am a bit confused about the deterministic chaos that you briefly mention at the start of 5.2.1 -- the parameter setting you use in this section (5.2.2) presumably does not give rise to deterministic chaos?\n\n---\n\nNone of your experiments contains fully identifiable noise models from what I can tell? You mention \"noise\" twice in the manuscript. None is in the experimental section. Now since you are only using simulations, I find it odd that you do not discuss this nor ACD's robustness against noise. As you are fully aware of course, most of these models break-down in the presence of a low SNR. And for this to be useful at all in the 'real' world, how does ACD fare?\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603101862259}], "openreview_url": "https://openreview.net/forum?id=gW8n0uD6rl", "arxiv_id": "2006.10833", "paper_pdf": "papers/gW8n0uD6rl.pdf", "paper_pdf_sha256": "ae0d46e9fccc33f0ffa29757c4805699a176c44b366639e1730c2b92c9478475", "paper_pdf_bytes": 547682, "paper_pdf_source": "openreview", "code_url": "https://github.com/loeweX/AmortizedCausalDiscovery", "code_repository": "loeweX/AmortizedCausalDiscovery", "code_commit": "4d1661118c9fcfbefbbd00226bc41f7a76743e11", "code_archive": "repos/gW8n0uD6rl.zip", "code_archive_sha256": "f213392592897a594b0a2f35cba2c9a7d86ad2ab66832a0cc3331923d6956dc8", "code_archive_bytes": 127313, "code_file_count": 24, "code_extensions": {".py": 23, ".sh": 1}, "github_disk_usage_kb": 156, "github_languages": {"Python": 142492, "Shell": 155}, "github_archived": false, "github_pushed_at": "2022-03-25T13:03:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/amortized-causal-discovery-learning-to-infer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJlnmgrFvS", "year": 2020, "status": "rejected", "title": "BAIL: Best-Action Imitation Learning for Batch Deep Reinforcement Learning", "authors": ["Xinyue Chen", "Zijian Zhou", "Zheng Wang", "Che Wang", "Yanqiu Wu", "Qing Deng", "Keith Ross"], "authorids": ["xc1305@nyu.edu", "zz1435@nyu.edu", "zw1454@nyu.edu", "cw1681@nyu.edu", "yanqiu.wu@nyu.edu", "qd319@nyu.edu", "keithwross@nyu.edu"], "authors_source": "OpenReview API", "abstract": "The field of Deep Reinforcement Learning (DRL) has recently seen a surge in research in batch reinforcement learning, which aims for sample-efficient learning from a given data set without additional interactions with the environment. In the batch DRL setting, commonly employed off-policy DRL algorithms can perform poorly and sometimes even fail to learn altogether.  In this paper we propose anew algorithm, Best-Action Imitation Learning (BAIL), which unlike many off-policy DRL algorithms does not involve maximizing Q functions over the action space. Striving for simplicity as well as performance, BAIL first selects from the batch the actions it believes to be high-performing actions for their corresponding states; it then uses those state-action pairs to train a policy network using imitation learning.  Although BAIL is simple, we demonstrate that BAIL achieves state of the art performance on the Mujoco benchmark, typically outperforming BatchConstrained deep Q-Learning (BCQ) by a wide margin.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJelB0Pj5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2225/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper studies the problem of learning a policy from a fixed dataset. The authors propose to estimate a smooth upper envelope of the episodic returns from the dataset as a state-value function. The policy is then learned by imitating the state action pairs from the dataset whose actual episodic return is close to the estimated envelope.\n\nRecommended decision:\nThe direction of imitating \"good\" actions from the dataset is interesting. The intuition of estimating an upper envelope of the value function seems reasonable. However, I feel like this paper is not ready to be published in terms of its overall quality, mainly due to the lack of correctness, rigorousness and justification in statements and approaches.\n\nMajor comments:\n\n- On the top of page 4:  \"Because the Mujoco environments are continuing tasks, it is desirable to approximate the return over the infinite horizon, particularly for i values that are close to the (artificial) end of an episode. To do this, we note that the data-generation policy from one episode to the next typically changes slowly. We therefore apply a simple augmentation heuristic of concatenating the subsequent episode to the current episode, and running the sum in (1) to infinity.\" I cannot see how this approach is validated. The reset of initial state makes cross-episode cumulative reward from a state s not an approximation to the real return from state s. Estimating the infinite horizon return from finite horizon data is indeed a challenge here and simply cut the return at the end of an episode is be problematic. But the solution proposed by the authors is wrong in principle and cannot be simply justified by \"good empirical performance\". I feel hard to regard this choice a valid part of an algorithm unless further justification can be provided.\n\n- Statements of theorems (4.1 and 4.2) are non-rigorous and contain irrelevant information: \"lambda-smooth\" is not an appropriate terminology when lambda is the weight of the regularizer. The actual \"smoothness\" also depends on the other term in the loss (same lambda does not indicate same smoothness in different objectives). For the same reason, Theorem 4.2 is wrong as changing K also changes the smoothness of the learned function. Proof of Theorem 4.2 in appendix is wrong as the authors ignore the coefficients in the last equation. Theorem 4.1-(1) cannot be true unless how V_\\phi is parameterized is given: e.g. if there is no bias term or the regularization is applies to the bias term V will always output 0 as lambda 0-> \\infty. The \"2m+d\" in Theorem 4.1-(2) is irrelevant to this work and cannot be justified without more detailed statements about how the network is parameterized. I appreciate the motivation that the authors try to validate the use of their objective to learn a \"smooth upper envelope\" but most of these statements are somewhat trivial and/or wrong section 4.1 does not actually deliver a valid justification.\n\n- The use of \"smooth upper envelope\" itself can bring both over-estimation and under-estimation. For example, if one can concatenate different parts from different episodes to get a trajectory with higher return, the episodic return for the states along this trajectory is an under-estimate. Although it is fine to use a conservative estimate it would be better to be explicit about this and explain why this may not be a concern. On the other hand, it can bring over estimation to the state-values due to the smoothness enhanced to the fitted V. It would be better to see e.g. when these concerns do not matter (theoretically) or they are not real concerns in practice (by further inspecting the experiments). \n\n- Regarding Experiments: Why Hopper, Walker, HalfCheetah are trained with DDPG while Ant is trained by SAC? The performance of Final-DDPG/SAC after training for 1m steps looks way below what SAC and TD3 can get. Is it because they are just partially trained or noise is added to them? The baseline online-trained policy should not contain noise for a fair comparison. That said, in batch RL setting it is not necessary to compare to online-trained policy because it is a different setting. But if the authors want to compare to those, choice of baseline should be careful. An important baseline which is missing is to run vanilla DDPG/TD3/SAC as a batch-mode algorithm.  \n\n\nMinor comments:\n\n- Section 3, first paragraph: It is not very meaningful to say \"simulators are deterministic so deterministic environments are important\". Simulators are made by humans so they can be either deterministic or stochastic. \"many robotic tasks are expected to be deterministic environments\" is probably not true. I do not view \"assuming deterministic envs\" as a major limitation but I do not find these statements convincing as well. Similarly, the argument for studying non-stationary policy seems unsupportive: if the dataset comes from training a policy online then why do we care about learning another offline policy rather than just use or continue training the online policy. One argument I can see is that the online policy is worse. But the fact that these policies are worst than running e.g. SAC for a million steps makes the motivation questionable. Again, I do not view \"choice of setting\" as a limitation but I just find these statements a bit unsupportive. \n\n\nPotential directions for improvement:\n\nTo me the main part of the paper that looks problematic is Section 4.1 (both the approximation of infinite horizon returns and the theorems). It would be better to see a more rigorous and coherent justification of this approach (or some improved version), e.g. by either presenting analysis that is rigorous, correct and actually relevant or leave the space for more detailed empirical justification (e.g. whether potential over/under-estimating happens or not, comparing the estimated V to real episodic return of the learned policy).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "Summary:\nThis paper studies the problem of learning a policy from a fixed dataset. The authors propose to estimate a smooth upper envelope of the episodic returns from the dataset as a state-value function. The policy is then learned by imitating the state action pairs from the dataset whose actual episodic return is close to the estimated envelope.\n\nRecommended decision:\nThe direction of imitating \"good\" actions from the dataset is interesting. The intuition of estimating an upper envelope of the value function seems reasonable. However, I feel like this paper is not ready to be published in terms of its overall quality, mainly due to the lack of correctness, rigorousness and justification in statements and approaches.\n\nMajor comments:\n\n- On the top of page 4:  \"Because the Mujoco environments are continuing tasks, it is desirable to approximate the return over the infinite horizon, particularly for i values that are close to the (artificial) end of an episode. To do this, we note that the data-generation policy from one episode to the next typically changes slowly. We therefore apply a simple augmentation heuristic of concatenating the subsequent episode to the current episode, and running the sum in (1) to infinity.\" I cannot see how this approach is validated. The reset of initial state makes cross-episode cumulative reward from a state s not an approximation to the real return from state s. Estimating the infinite horizon return from finite horizon data is indeed a challenge here and simply cut the return at the end of an episode is be problematic. But the solution proposed by the authors is wrong in principle and cannot be simply justified by \"good empirical performance\". I feel hard to regard this choice a valid part of an algorithm unless further justification can be provided.\n\n- Statements of theorems (4.1 and 4.2) are non-rigorous and contain irrelevant information: \"lambda-smooth\" is not an appropriate terminology when lambda is the weight of the regularizer. The actual \"smoothness\" also depends on the other term in the loss (same lambda does not indicate same smoothness in different objectives). For the same reason, Theorem 4.2 is wrong as changing K also changes the smoothness of the learned function. Proof of Theorem 4.2 in appendix is wrong as the authors ignore the coefficients in the last equation. Theorem 4.1-(1) cannot be true unless how V_\\phi is parameterized is given: e.g. if there is no bias term or the regularization is applies to the bias term V will always output 0 as lambda 0-> \\infty. The \"2m+d\" in Theorem 4.1-(2) is irrelevant to this work and cannot be justified without more detailed statements about how the network is parameterized. I appreciate the motivation that the authors try to validate the use of their objective to learn a \"smooth upper envelope\" but most of these statements are somewhat trivial and/or wrong section 4.1 does not actually deliver a valid justification.\n\n- The use of \"smooth upper envelope\" itself can bring both over-estimation and under-estimation. For example, if one can concatenate different parts from different episodes to get a trajectory with higher return, the episodic return for the states along this trajectory is an under-estimate. Although it is fine to use a conservative estimate it would be better to be explicit about this and explain why this may not be a concern. On the other hand, it can bring over estimation to the state-values due to the smoothness enhanced to the fitted V. It would be better to see e.g. when these concerns do not matter (theoretically) or they are not real concerns in practice (by further inspecting the experiments). \n\n- Regarding Experiments: Why Hopper, Walker, HalfCheetah are trained with DDPG while Ant is trained by SAC? The performance of Final-DDPG/SAC after training for 1m steps looks way below what SAC and TD3 can get. Is it because they are just partially trained or noise is added to them? The baseline online-trained policy should not contain noise for a fair comparison. That said, in batch RL setting it is not necessary to compare to online-trained policy because it is a different setting. But if the authors want to compare to those, choice of baseline should be careful. An important baseline which is missing is to run vanilla DDPG/TD3/SAC as a batch-mode algorithm.  \n\n\nMinor comments:\n\n- Section 3, first paragraph: It is not very meaningful to say \"simulators are deterministic so deterministic environments are important\". Simulators are made by humans so they can be either deterministic or stochastic. \"many robotic tasks are expected to be deterministic environments\" is probably not true. I do not view \"assuming deterministic envs\" as a major limitation but I do not find these statements convincing as well. Similarly, the argument for studying non-stationary policy seems unsupportive: if the dataset comes from training a policy online then why do we care about learning another offline policy rather than just use or continue training the online policy. One argument I can see is that the online policy is worse. But the fact that these policies are worst than running e.g. SAC for a million steps makes the motivation questionable. Again, I do not view \"choice of setting\" as a limitation but I just find these statements a bit unsupportive. \n\n\nPotential directions for improvement:\n\nTo me the main part of the paper that looks problematic is Section 4.1 (both the approximation of infinite horizon returns and the theorems). It would be better to see a more rigorous and coherent justification of this approach (or some improved version), e.g. by either presenting analysis that is rigorous, correct and actually relevant or leave the space for more detailed empirical justification (e.g. whether potential over/under-estimating happens or not, comparing the estimated V to real episodic return of the learned policy).\n"}, "tcdate": 1572728376044}, {"id": "SJlbdX86tr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2225/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe paper tries to solve a batch reinforcement learning problem with a very simple but efficient algorithm. It first learns a smooth upper bound of Monte Carlo returns in the batch data (called the \"upper envelope\"). Then, the algorithm chooses state action pairs of the batch data that have returns larger than constant times the upper envelope. It lowers the constant until the algorithm gets 25% of the data. Then the algorithm trains the policy on chosen state-action pairs. The algorithm is shown to outperform BCQ in experiments.\n\nAlthough I like the idea of the paper, I vote for rejection. While there is no theoretical guarantee on the performance of the algorithm, the design of the algorithm does not follow the usual design the other researchers follow. The way of reporting the experiment results does not seem very professional. I recommend the authors to consult with some other researchers who have publication experience. In the current form, the paper is very poor in detail that makes readers hard to be convinced with the results.\n\nThese are some points that I could not understand:\n\n1. Why do you fix K=10000 on modified loss instead of dual gradient descent for constrained optimization?\n2. How do you guarantee that choosing (s,a) such that G>xV gives you good samples? Since mean returns are not zero, it won't pick the top 25% actions for all states. States with the high mean return will have all of its samples included, while states with the low mean return will have all of its samples excluded. Although the authors concatenated all the experiences to compute returns (which is ad-hoc as well), the initial states will have a lower return than other states. This means that most of the actions of the initial states will be excluded in the training set while more actions of the other states will be included, which does not seem desirable. (e.g. in Figure 1 Ant. If we set x=0 (extreme case), states of timestep >600000 will be all included where t<600000 will be partially excluded. )\n3. In the explanation of Figure 2, it is written as \"standard deviation confidence interval\". Is it standard deviation, or confidence interval? Also, why are the standard deviation in the Figure 2 and the Table 1 so different? How do you compute Improvement in the Table 1? What happens if the environment gives negative returns only (i.e. Pendulum), such that BCQ gives you -1000 and BAIL gives you -500?\n4. As claimed in theorems, V=max(G) if lambda->infinity. This means that the \"Highest Returns\" in figure 3 is also one specific hyperparameter choice of the suggested algorithm. There might be a better choice of regularization that outperforms both BAIL and Highest Returns as early-stopping done in the paper is just one random amount of regularization. What was the early-stopping criterion and how is it chosen? How do we know it is the best regularization option?\n5. Is the final DDPG or final SAC evaluated with a deterministic policy? According to the paper, I assume that it was not. Those algorithms usually add large noise while training for exploration, and such noise is removed while in evaluation. In Bear Q learning, better action selection technic is used, which chooses the action sample that maximizes critic Q. Is the evaluations really fair for all algorithms? As far as I know, Mujoco environments are deterministic except the initial state sampling, and there should only be very small variance.\n\nAlso, I believe the paper should be compared to Bear Q learning as well, as it is very easy to implement and outperforms BCQ by a large margin. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "\nThe paper tries to solve a batch reinforcement learning problem with a very simple but efficient algorithm. It first learns a smooth upper bound of Monte Carlo returns in the batch data (called the \"upper envelope\"). Then, the algorithm chooses state action pairs of the batch data that have returns larger than constant times the upper envelope. It lowers the constant until the algorithm gets 25% of the data. Then the algorithm trains the policy on chosen state-action pairs. The algorithm is shown to outperform BCQ in experiments.\n\nAlthough I like the idea of the paper, I vote for rejection. While there is no theoretical guarantee on the performance of the algorithm, the design of the algorithm does not follow the usual design the other researchers follow. The way of reporting the experiment results does not seem very professional. I recommend the authors to consult with some other researchers who have publication experience. In the current form, the paper is very poor in detail that makes readers hard to be convinced with the results.\n\nThese are some points that I could not understand:\n\n1. Why do you fix K=10000 on modified loss instead of dual gradient descent for constrained optimization?\n2. How do you guarantee that choosing (s,a) such that G>xV gives you good samples? Since mean returns are not zero, it won't pick the top 25% actions for all states. States with the high mean return will have all of its samples included, while states with the low mean return will have all of its samples excluded. Although the authors concatenated all the experiences to compute returns (which is ad-hoc as well), the initial states will have a lower return than other states. This means that most of the actions of the initial states will be excluded in the training set while more actions of the other states will be included, which does not seem desirable. (e.g. in Figure 1 Ant. If we set x=0 (extreme case), states of timestep >600000 will be all included where t<600000 will be partially excluded. )\n3. In the explanation of Figure 2, it is written as \"standard deviation confidence interval\". Is it standard deviation, or confidence interval? Also, why are the standard deviation in the Figure 2 and the Table 1 so different? How do you compute Improvement in the Table 1? What happens if the environment gives negative returns only (i.e. Pendulum), such that BCQ gives you -1000 and BAIL gives you -500?\n4. As claimed in theorems, V=max(G) if lambda->infinity. This means that the \"Highest Returns\" in figure 3 is also one specific hyperparameter choice of the suggested algorithm. There might be a better choice of regularization that outperforms both BAIL and Highest Returns as early-stopping done in the paper is just one random amount of regularization. What was the early-stopping criterion and how is it chosen? How do we know it is the best regularization option?\n5. Is the final DDPG or final SAC evaluated with a deterministic policy? According to the paper, I assume that it was not. Those algorithms usually add large noise while training for exploration, and such noise is removed while in evaluation. In Bear Q learning, better action selection technic is used, which chooses the action sample that maximizes critic Q. Is the evaluations really fair for all algorithms? As far as I know, Mujoco environments are deterministic except the initial state sampling, and there should only be very small variance.\n\nAlso, I believe the paper should be compared to Bear Q learning as well, as it is very easy to implement and outperforms BCQ by a large margin. \n"}, "tcdate": 1571804008794}, {"id": "S1gZB2I2YB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2225/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary of Claims:\n\nThe paper proposes a batch RL method that they claim is simpler than most existing methods that try to avoid the extrapolation error that is prevalent among batch RL methods. They do this by completely avoiding the minimization/maximization (cost/reward) of the approximate value function that is fit to the batch transitions. Instead, they train an approximation for the state value function's tight upper bound (which they refer to as the upper-envelope) by using their monte-carlo returns. By fitting such an approximator, they sample the state-action pairs that are close to the envelope (thus have high values/returns), and use behavioral cloning to fit a parameterized policy to those state-action pairs.\n\nDecision:\n\nWeak Reject.\nMy decision is influenced by two main reasons:\n\n(1) Although the simplicity of the method is apparent and a very desirable feature, the authors don't highlight situations where this can lead to bad policies. For example, consider that there are two pairs (s, a_1, s') and (s, a_2, s') in the batch that are close to the upper-envelope, and hence will both be used for training the policy. Using Behavioral cloning, the policy would regress to the mean of a_1 and a_2, which could be a terrible action altogether. The issue here is that only one of these two pairs has higher return and our policy needs to only predict that action (or in the case of tie, either one.) This can be really bad in situations where two very different actions can lead to same returns (e.g. in a reacher-like task the arm can reach a goal in two different rotations.) Even though I pointed out a very specific case, one could think of many other cases where the proposed approach might result in a bad policy. \n\nHaving said all of this, it might be true that such cases do not appear in practice (which I highly doubt) but its the authors job to raise and clarify that. The current set of experimental setups (mujoco locomotion problems) are not good enough evidence for that and they need experiments where optimal-policies can be multi-modal or have diverse experimental setups (manipulation etc.)\n\n(2) Experimental results are a little unsettling. The primary reason is that in all of the plots, BCQ, BAIL, BC aren't starting from the same test return at 0 parameter updates! In most plots BAIL starts off way higher in return than BCQ, BC with no parameter updates yet, which suggests that the experiments were not setup well. Maybe, they didn't initialize the policy in the same way for all the approaches, maybe the random seeds were not the same for all approaches, or maybe BAIL had some sort of pretraining for the policy that was not accounted for in the parameter updates. In any way, this needs to be addressed. This is also highlighted by the fact that the learning curves for BAIL are almost always flat across a million parameter updates! If you are starting off with a random initialization, there should be an upwards slope for the learning curve. Also, as raised in the previous point I think using these Mujoco locomotion environments is not convincing enough to claim that BAIL is a viable competitive batch RL approach.\n\nComments and Questions:\n\n(1) I like the simplicity of the approach and the fact that it is much more easier to understand than existing works like BCQ\n\n(2) Paper is well-written. It was clear, lucid and descriptive.\n\n(3) Why is the deterministic dynamics assumption needed? I am curious\n\n(4) The paper makes some subjective statements such as \"BEAR is also complex\", which is not substantiated well enough. Refrain from making such statements\n\n(5) Not comparing to BEAR because their code is not publicly available is a contentious reason. I personally feel that the authors could have reimplemented it and compared but I am not sure what the community feels about that\n\n(6) Is there any reason why REM cannot be applied to mujoco environments? If it can be, then why did the authors not compare to REM as well?\n\n(7) Another subjective statement (that is clearly wrong) \"many robotic tasks are expected to be deterministic environments\" - although this is slightly true, the reason we model environments to be stochastic is not because there is inherent randomness in them but because our state descriptions are never complete. The state descriptors are always partial and we account for them by assuming stochasticity in the dynamics. For example, consider a robotic manipulation task where if you know all the environmental factors as part of your state space(such as the friction coefficients) you can assume deterministic dynamics, else you are better off assuming stochastic dynamics because the same actuation might not result in the same motion every time (because of varying friction)\n\n(8) Concatenating subsequent episodes in a batch only makes sense (as the authors point out) if the policy doesn't change much across episodes. But this is not true of current off-policy RL methods like DDPG, SAC. You either need very small learning rate or a trust-region constraint to ensure that the policy doesn't change much across episodes. \n\n(9) Why do different batches with different seeds and the same algorithm lead to widely different results for batch RL? There is clearly something fishy here. Is it because of the off-policy RL methods used to collect the data, is it due to the batch RL method used? More investigation needed", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Summary of Claims:\n\nThe paper proposes a batch RL method that they claim is simpler than most existing methods that try to avoid the extrapolation error that is prevalent among batch RL methods. They do this by completely avoiding the minimization/maximization (cost/reward) of the approximate value function that is fit to the batch transitions. Instead, they train an approximation for the state value function's tight upper bound (which they refer to as the upper-envelope) by using their monte-carlo returns. By fitting such an approximator, they sample the state-action pairs that are close to the envelope (thus have high values/returns), and use behavioral cloning to fit a parameterized policy to those state-action pairs.\n\nDecision:\n\nWeak Reject.\nMy decision is influenced by two main reasons:\n\n(1) Although the simplicity of the method is apparent and a very desirable feature, the authors don't highlight situations where this can lead to bad policies. For example, consider that there are two pairs (s, a_1, s') and (s, a_2, s') in the batch that are close to the upper-envelope, and hence will both be used for training the policy. Using Behavioral cloning, the policy would regress to the mean of a_1 and a_2, which could be a terrible action altogether. The issue here is that only one of these two pairs has higher return and our policy needs to only predict that action (or in the case of tie, either one.) This can be really bad in situations where two very different actions can lead to same returns (e.g. in a reacher-like task the arm can reach a goal in two different rotations.) Even though I pointed out a very specific case, one could think of many other cases where the proposed approach might result in a bad policy. \n\nHaving said all of this, it might be true that such cases do not appear in practice (which I highly doubt) but its the authors job to raise and clarify that. The current set of experimental setups (mujoco locomotion problems) are not good enough evidence for that and they need experiments where optimal-policies can be multi-modal or have diverse experimental setups (manipulation etc.)\n\n(2) Experimental results are a little unsettling. The primary reason is that in all of the plots, BCQ, BAIL, BC aren't starting from the same test return at 0 parameter updates! In most plots BAIL starts off way higher in return than BCQ, BC with no parameter updates yet, which suggests that the experiments were not setup well. Maybe, they didn't initialize the policy in the same way for all the approaches, maybe the random seeds were not the same for all approaches, or maybe BAIL had some sort of pretraining for the policy that was not accounted for in the parameter updates. In any way, this needs to be addressed. This is also highlighted by the fact that the learning curves for BAIL are almost always flat across a million parameter updates! If you are starting off with a random initialization, there should be an upwards slope for the learning curve. Also, as raised in the previous point I think using these Mujoco locomotion environments is not convincing enough to claim that BAIL is a viable competitive batch RL approach.\n\nComments and Questions:\n\n(1) I like the simplicity of the approach and the fact that it is much more easier to understand than existing works like BCQ\n\n(2) Paper is well-written. It was clear, lucid and descriptive.\n\n(3) Why is the deterministic dynamics assumption needed? I am curious\n\n(4) The paper makes some subjective statements such as \"BEAR is also complex\", which is not substantiated well enough. Refrain from making such statements\n\n(5) Not comparing to BEAR because their code is not publicly available is a contentious reason. I personally feel that the authors could have reimplemented it and compared but I am not sure what the community feels about that\n\n(6) Is there any reason why REM cannot be applied to mujoco environments? If it can be, then why did the authors not compare to REM as well?\n\n(7) Another subjective statement (that is clearly wrong) \"many robotic tasks are expected to be deterministic environments\" - although this is slightly true, the reason we model environments to be stochastic is not because there is inherent randomness in them but because our state descriptions are never complete. The state descriptors are always partial and we account for them by assuming stochasticity in the dynamics. For example, consider a robotic manipulation task where if you know all the environmental factors as part of your state space(such as the friction coefficients) you can assume deterministic dynamics, else you are better off assuming stochastic dynamics because the same actuation might not result in the same motion every time (because of varying friction)\n\n(8) Concatenating subsequent episodes in a batch only makes sense (as the authors point out) if the policy doesn't change much across episodes. But this is not true of current off-policy RL methods like DDPG, SAC. You either need very small learning rate or a trust-region constraint to ensure that the policy doesn't change much across episodes. \n\n(9) Why do different batches with different seeds and the same algorithm lead to widely different results for batch RL? There is clearly something fishy here. Is it because of the off-policy RL methods used to collect the data, is it due to the batch RL method used? More investigation needed"}, "tcdate": 1571740728734}], "openreview_url": "https://openreview.net/forum?id=BJlnmgrFvS", "arxiv_id": "1910.12179", "paper_pdf": "papers/BJlnmgrFvS.pdf", "paper_pdf_sha256": "ed7f835dd2972eb29461e62ffe7b28a7d7cea605fb8f5b596e1f1c961b89e307", "paper_pdf_bytes": 20584286, "paper_pdf_source": "openreview", "code_url": "https://github.com/lanyavik/BAIL", "code_repository": "lanyavik/BAIL", "code_commit": "ca1cbc6ff60368a8cb1734a0a67e5478dceba732", "code_archive": "repos/BJlnmgrFvS.zip", "code_archive_sha256": "357547d03c47e73ee2d21ef8d9d7b1cf0cf8f715b3e083ccd9cb0020853db54c", "code_archive_bytes": 168052, "code_file_count": 33, "code_extensions": {".py": 33}, "github_disk_usage_kb": 186, "github_languages": {"Python": 196841}, "github_archived": false, "github_pushed_at": "2022-07-13T12:59:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bail-best-action-imitation-learning-for-batch-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJxbYoC9FQ", "year": 2019, "status": "rejected", "title": "Classifier-agnostic saliency map extraction", "authors": ["Konrad Zolna", "Krzysztof J. Geras", "Kyunghyun Cho"], "authorids": ["konrad.zolna@gmail.com", "k.j.geras@nyu.edu", "kyunghyun.cho@nyu.edu"], "authors_source": "OpenReview API", "abstract": "Extracting saliency maps, which indicate parts of the image important to classification, requires many tricks to achieve satisfactory performance when using classifier-dependent methods. Instead, we propose classifier-agnostic saliency map extraction, which finds all parts of the image that any classifier could use, not just one given in advance. We observe that the proposed approach extracts higher quality saliency maps and outperforms existing weakly-supervised localization techniques, setting the new state of the art result on the ImageNet dataset.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "HyxqVGlT37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper418/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a classifier-agnostic method for saliency map extraction. In order to address the dependence of saliency map extraction on the classifier, the authors propose to learn a saliency mapping by considering all possible classifiers (i.e., a certain classifier structure w.r.t. the space of all its parameters). The goal is to find the relevant features in the data that work with all possible classifiers. The proposed framework is formulated as a min-max game between two players: a mask m corresponding to the saliency mapping, and a function f sampled from a set of classifiers with the same structure but different parameters. The mapping m is optimized to maximize the masked-out classification error (such that m captures all relevant features whose removal can maximally confuse the classifier), while f is optimized to minimize the mask-out classification error.\n\nThe idea of how to formulate the classifier set and how to sample from the set is interesting. However, I have some concerns regarding the overall model:\n\n1) It seems not quite convincing to me why the model should involve an adversarial game. In particular, why f should be optimized to minimize the masked-out classification error? I understand that by doing this, f has an opposite goal with m so as to force m to capture as many relevant features as possible. However, I do not think f has a natural motivation to minimize the maxed-out error. In my opinion, it seems more convincing if f is optimized to minimize the masked-in classification error, but not necessarily the masked-out one. I think this also explains why the model works better by adding the classification loss over original images in Eq. (8). Would it be more natural if we optimize both f and m to minimize the masked-in classification error? And it would easier to train compared with the min-max model. Maybe some more explanation on the motivation of such an adversarial game can be helpful.\n\n2) I was curious whether the alternating optimization of m and f would cause the cumulation of errors? I mean, if in some iteration m just captures irrelevant features, f will still be optimized to accomodate to such a bad mapping. Would such kind of error accumulate during training?\n\n3) In Algorithm 1, after \\theta_m is learned, how is m is determined?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "classifier-agnostic method for object localization", "review": "This paper proposes a classifier-agnostic method for saliency map extraction. In order to address the dependence of saliency map extraction on the classifier, the authors propose to learn a saliency mapping by considering all possible classifiers (i.e., a certain classifier structure w.r.t. the space of all its parameters). The goal is to find the relevant features in the data that work with all possible classifiers. The proposed framework is formulated as a min-max game between two players: a mask m corresponding to the saliency mapping, and a function f sampled from a set of classifiers with the same structure but different parameters. The mapping m is optimized to maximize the masked-out classification error (such that m captures all relevant features whose removal can maximally confuse the classifier), while f is optimized to minimize the mask-out classification error.\n\nThe idea of how to formulate the classifier set and how to sample from the set is interesting. However, I have some concerns regarding the overall model:\n\n1) It seems not quite convincing to me why the model should involve an adversarial game. In particular, why f should be optimized to minimize the masked-out classification error? I understand that by doing this, f has an opposite goal with m so as to force m to capture as many relevant features as possible. However, I do not think f has a natural motivation to minimize the maxed-out error. In my opinion, it seems more convincing if f is optimized to minimize the masked-in classification error, but not necessarily the masked-out one. I think this also explains why the model works better by adding the classification loss over original images in Eq. (8). Would it be more natural if we optimize both f and m to minimize the masked-in classification error? And it would easier to train compared with the min-max model. Maybe some more explanation on the motivation of such an adversarial game can be helpful.\n\n2) I was curious whether the alternating optimization of m and f would cause the cumulation of errors? I mean, if in some iteration m just captures irrelevant features, f will still be optimized to accomodate to such a bad mapping. Would such kind of error accumulate during training?\n\n3) In Algorithm 1, after \\theta_m is learned, how is m is determined?", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541370417941}, {"id": "Sklh9zM237", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper418/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new saliency map extractor that seems to improve state-of-the-art results. Saliency maps are tools that can be useful to understand the decision-making of deep networks for object recognition; advances in this research topic may lead to a better understanding of the functioning of deep networks.\n\nThe paper is based on the approach of Fan et al. (2017). The improvements over Fan et al. seem to be mainly technical: the objective function and the optimization procedure. This makes the novelty of the paper quite thin, as the main underlying idea of the paper was previously introduced. \n\nThe algorithm is introduced without motivating well the different choices. What are the intuitions and evidence that guided the design of the algorithm? How are the technical choices made? Answers to these questions may help to understand how this paper builds on Fan et al. and other previous works.\n\nThe experiments compare a comprehensive set of algorithms and show an improvement over previous works. Yet, the OM metric is only compared for a few of these algorithms and it is unclear why is so. Also, the qualitative examples do not clarify how the proposed method improves over state-of-the-art (it would be useful to compare with qualitative examples of previous work). It remains unclear how much of an improvement over state-of-the-art there is.\n\nIn summary, I think the paper could be valuable as the proposed algorithm may improve state-of-the-art results. Yet, these results and the novelty of the paper are not entirely clear.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Improvement (maybe) of saliency maps by introducing technical improvements over previous work", "review": "This paper introduces a new saliency map extractor that seems to improve state-of-the-art results. Saliency maps are tools that can be useful to understand the decision-making of deep networks for object recognition; advances in this research topic may lead to a better understanding of the functioning of deep networks.\n\nThe paper is based on the approach of Fan et al. (2017). The improvements over Fan et al. seem to be mainly technical: the objective function and the optimization procedure. This makes the novelty of the paper quite thin, as the main underlying idea of the paper was previously introduced. \n\nThe algorithm is introduced without motivating well the different choices. What are the intuitions and evidence that guided the design of the algorithm? How are the technical choices made? Answers to these questions may help to understand how this paper builds on Fan et al. and other previous works.\n\nThe experiments compare a comprehensive set of algorithms and show an improvement over previous works. Yet, the OM metric is only compared for a few of these algorithms and it is unclear why is so. Also, the qualitative examples do not clarify how the proposed method improves over state-of-the-art (it would be useful to compare with qualitative examples of previous work). It remains unclear how much of an improvement over state-of-the-art there is.\n\nIn summary, I think the paper could be valuable as the proposed algorithm may improve state-of-the-art results. Yet, these results and the novelty of the paper are not entirely clear.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541313171754}, {"id": "r1gvfkii27", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper418/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper focuses on the extraction of high-quality model-agnostic saliency maps. The authors argue that when an extracted saliency map is directly dependent on a model, then it might not be useful for a different classifier and thus not general enough. To overcome this problem, they consider all the possible classifiers weighted by their posterior probabilities. This problem cannot be solved explicitly, and the authors suggest a scheme to approximate the solution using two networks. That is, pretrain an initial classifier and then, following an adversarial training procedure, one network is trying to confuse the classifier and the other one to maximize its accuracy. Using this formulation, the authors report state-of-the-art results for salience map extraction.\n\nSUMMARY/OVERALL COMMENTS\nThe authors present a simple and effective way to produce classifier-agnostic saliency maps. The argument for the approach is well justified and the results seem convincing on a first read. However, the novelty of the method is a concern given the previous work of Fan et al. (2017), and the manuscript is not upfront about the differences between the two works. The experiments are another cause for concern: Fan et al. should have been tested as a baseline with similar implementation (controlling for architecture and \\lambda), and implementation differences in prior works of Table 1 make it difficult to draw conclusions. \n\n\nRELATED WORKS\n* In the introduction, the authors mention related works but fail to mention the work of Fan et al. (2017) which is clearly the most relevant. The first mention of Fan et al is on page 4 in a very specific discussion the regularization coefficient. The problem formulation in Section 2 and the approach is Section 3 is largely borrowed from Fan et al but not acknowledged until the last page. This introduces bias and confusion to the reader in regards to the novelty of the approach. Please, mention the work of Fan et al. (2017) in the introduction and clearly delineate the differences in the works earlier in the text. (--)\n\n* Du et al. (2018), “Towards Explanation of DNN-based Prediction with Guided Feature Inversion”, use the VGG models for saliency map extraction and achieve a LE of 38.2. Note that Du et al. (2018), suggest that this modification could lead to SOTA results. I would like to see a comparison with this method. (-)\n\n* The work of Kindermans, et al. (2017), “Learning how to explain neural networks: PatternNet and Pattern Attribution”, although they do not aim for weakly supervised localization and thus, they do not present the LE, they produce saliency maps. I would like to see a LE comparison with that method. (minor -)\n\n* In the introduction, p1 (last paragraph) other methods are briefly mentioned (Extracted saliency maps show all the evidence….superpixels), etc.) without references. Please add references when needed. (-)\n\n* The framework presented in this paper was first proposed by Fan et al. (2017). The authors claim four main differences in their approach. In my eyes, not all of them are major or novel - probably the most impactful is removing superpixels as it simplifies the problem and implementation. (+)\n\n\nAPPROACH\n* The authors aim for simplicity (strong +)\n\n* The authors justify their approach and present their arguments clearly (strong ++)\n\n* In the algorithm section, the authors first mention the sampling procedure and then their motivation. Please alter the ordering of these to be conceptually easier to understand your approach\n\n* In the first sentence after the equation 6 I guess that “(cf. Alg. 1)” is a typo and should be modified to (Alg. 1)”\n\n* After Equation 6 it is argued that the method resembles the training procedure of GANs (Godfellow et al., 2014) but not the work of Fan et al., 2017. (--)\n\n\nEXPERIMENTS\n* The authors define as their baseline the F thinning strategy (i.e. use only the first classifier) which is a model dependent salience map. While this is a useful comparison against the classifier dependent methods, given the similarity to the work of Fan et al. (2017), experiments comparing the proposed model to Fan et al. are necessary. It is important to control for network architecture (ResNet-50) and choice of \\lambda to properly determine if the four changes outlined in Section 6 result in any real improvement over Fan et al. (strong --)\n\n* The authors use the Table 1 (borrowed from Fong and Vedaldi (2017)) to compare their results against other methods. This comparison is problematic as different approaches are using different models as classifiers which may lead to increase or decrease of the LE. (--)\n\n* In Table 2 the authors do not report how many times they run the same experiments to get these values. They also run less experiments with non-shared weights and they report only the LE. In my eyes it looks that the authors are trying to force their argument that the sharing weights helps (probably because it is one of their novelties). Please report the statistics of your experiments and fill the empty entries in the table. (--)\n\n* In Table 3, what does the last row represent?\n\n* Table 1 errors: (1) You write “Localization evaluation using OM, LE and F1 scores”. Please remove the F1 score as you do not report it. Also, correct the first sentence of the “Localization” subsection which states that you use three different metrics to “two different metrics”. (2) The LE from Fong and Vedaldi (2017) should be 43.2 and not 43.1.\n\n* Regarding the unseen classes (section 5): (1) Please report in the appendix the classes that you are using in each subset. Are there classes correlated? (-)  (2) I see that there is a strong correlation between the LE on subset A and E. It looks like you are training on E and you generalize on A.\n\n\nNOVELTY/IMPACT\n* Novelty is a strong concern, given the work of Fan et al. (2017) (strong --). Nevertheless, the authors propose some changes that can be seen as more general, but the effectiveness of the changes is clearly established.\n\n* This paper’s strongest point is the simplicity (conceptually and implementation-wise) of the method, an advantage over previous works (+)\n\n\nOTHER COMMENTS\n* Fan et al. (2017), use an adaptive λ that pushes the mask to 10% of the image whereas you are using a fixed one that pushes the mask to approximately 50% of the image. How can you make sure that this is not the reason that you are getting better results?\n\nIf the authors can clearly and fairly demonstrate that the changes they propose over Fan et al (2017) result in improved performance, and the manuscript is adjusted to be more upfront about this prior work, I would consider increasing my rating.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Issues with novelty and improvements in relation to prior works", "review": "This paper focuses on the extraction of high-quality model-agnostic saliency maps. The authors argue that when an extracted saliency map is directly dependent on a model, then it might not be useful for a different classifier and thus not general enough. To overcome this problem, they consider all the possible classifiers weighted by their posterior probabilities. This problem cannot be solved explicitly, and the authors suggest a scheme to approximate the solution using two networks. That is, pretrain an initial classifier and then, following an adversarial training procedure, one network is trying to confuse the classifier and the other one to maximize its accuracy. Using this formulation, the authors report state-of-the-art results for salience map extraction.\n\nSUMMARY/OVERALL COMMENTS\nThe authors present a simple and effective way to produce classifier-agnostic saliency maps. The argument for the approach is well justified and the results seem convincing on a first read. However, the novelty of the method is a concern given the previous work of Fan et al. (2017), and the manuscript is not upfront about the differences between the two works. The experiments are another cause for concern: Fan et al. should have been tested as a baseline with similar implementation (controlling for architecture and \\lambda), and implementation differences in prior works of Table 1 make it difficult to draw conclusions. \n\n\nRELATED WORKS\n* In the introduction, the authors mention related works but fail to mention the work of Fan et al. (2017) which is clearly the most relevant. The first mention of Fan et al is on page 4 in a very specific discussion the regularization coefficient. The problem formulation in Section 2 and the approach is Section 3 is largely borrowed from Fan et al but not acknowledged until the last page. This introduces bias and confusion to the reader in regards to the novelty of the approach. Please, mention the work of Fan et al. (2017) in the introduction and clearly delineate the differences in the works earlier in the text. (--)\n\n* Du et al. (2018), “Towards Explanation of DNN-based Prediction with Guided Feature Inversion”, use the VGG models for saliency map extraction and achieve a LE of 38.2. Note that Du et al. (2018), suggest that this modification could lead to SOTA results. I would like to see a comparison with this method. (-)\n\n* The work of Kindermans, et al. (2017), “Learning how to explain neural networks: PatternNet and Pattern Attribution”, although they do not aim for weakly supervised localization and thus, they do not present the LE, they produce saliency maps. I would like to see a LE comparison with that method. (minor -)\n\n* In the introduction, p1 (last paragraph) other methods are briefly mentioned (Extracted saliency maps show all the evidence….superpixels), etc.) without references. Please add references when needed. (-)\n\n* The framework presented in this paper was first proposed by Fan et al. (2017). The authors claim four main differences in their approach. In my eyes, not all of them are major or novel - probably the most impactful is removing superpixels as it simplifies the problem and implementation. (+)\n\n\nAPPROACH\n* The authors aim for simplicity (strong +)\n\n* The authors justify their approach and present their arguments clearly (strong ++)\n\n* In the algorithm section, the authors first mention the sampling procedure and then their motivation. Please alter the ordering of these to be conceptually easier to understand your approach\n\n* In the first sentence after the equation 6 I guess that “(cf. Alg. 1)” is a typo and should be modified to (Alg. 1)”\n\n* After Equation 6 it is argued that the method resembles the training procedure of GANs (Godfellow et al., 2014) but not the work of Fan et al., 2017. (--)\n\n\nEXPERIMENTS\n* The authors define as their baseline the F thinning strategy (i.e. use only the first classifier) which is a model dependent salience map. While this is a useful comparison against the classifier dependent methods, given the similarity to the work of Fan et al. (2017), experiments comparing the proposed model to Fan et al. are necessary. It is important to control for network architecture (ResNet-50) and choice of \\lambda to properly determine if the four changes outlined in Section 6 result in any real improvement over Fan et al. (strong --)\n\n* The authors use the Table 1 (borrowed from Fong and Vedaldi (2017)) to compare their results against other methods. This comparison is problematic as different approaches are using different models as classifiers which may lead to increase or decrease of the LE. (--)\n\n* In Table 2 the authors do not report how many times they run the same experiments to get these values. They also run less experiments with non-shared weights and they report only the LE. In my eyes it looks that the authors are trying to force their argument that the sharing weights helps (probably because it is one of their novelties). Please report the statistics of your experiments and fill the empty entries in the table. (--)\n\n* In Table 3, what does the last row represent?\n\n* Table 1 errors: (1) You write “Localization evaluation using OM, LE and F1 scores”. Please remove the F1 score as you do not report it. Also, correct the first sentence of the “Localization” subsection which states that you use three different metrics to “two different metrics”. (2) The LE from Fong and Vedaldi (2017) should be 43.2 and not 43.1.\n\n* Regarding the unseen classes (section 5): (1) Please report in the appendix the classes that you are using in each subset. Are there classes correlated? (-)  (2) I see that there is a strong correlation between the LE on subset A and E. It looks like you are training on E and you generalize on A.\n\n\nNOVELTY/IMPACT\n* Novelty is a strong concern, given the work of Fan et al. (2017) (strong --). Nevertheless, the authors propose some changes that can be seen as more general, but the effectiveness of the changes is clearly established.\n\n* This paper’s strongest point is the simplicity (conceptually and implementation-wise) of the method, an advantage over previous works (+)\n\n\nOTHER COMMENTS\n* Fan et al. (2017), use an adaptive λ that pushes the mask to 10% of the image whereas you are using a fixed one that pushes the mask to approximately 50% of the image. How can you make sure that this is not the reason that you are getting better results?\n\nIf the authors can clearly and fairly demonstrate that the changes they propose over Fan et al (2017) result in improved performance, and the manuscript is adjusted to be more upfront about this prior work, I would consider increasing my rating.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541283598532}], "openreview_url": "https://openreview.net/forum?id=BJxbYoC9FQ", "arxiv_id": "1805.08249", "paper_pdf": "papers/BJxbYoC9FQ.pdf", "paper_pdf_sha256": "9af0cffc8eaccd2205d4ea4726ae42fbdb450b7342c33c2dfdcd339f54ad35e5", "paper_pdf_bytes": 19268307, "paper_pdf_source": "openreview", "code_url": "https://github.com/kondiz/casme", "code_repository": "kondiz/casme", "code_commit": "52cd5769bb7c1089934dd528a331afaf3882eff2", "code_archive": "repos/BJxbYoC9FQ.zip", "code_archive_sha256": "325307e1e60da717ae00008d137f9c7165c88aff5f7e759a1244df48e777c0c4", "code_archive_bytes": 1881246, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 1848, "github_languages": {"Python": 52297}, "github_archived": false, "github_pushed_at": "2020-05-12T20:40:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/classifier-agnostic-saliency-map-extraction"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4jMeUvcO26", "year": 2026, "status": "rejected", "title": "Surrogate Modeling of 3D Rayleigh-Bénard Convection with Equivariant Autoencoders", "authors": ["Fynn Fromme", "Hans Harder", "Christine Allen-Blanchette", "Sebastian Peitz"], "authorids": ["~Fynn_Fromme1", "~Hans_Harder1", "~Christine_Allen-Blanchette1", "~Sebastian_Peitz1"], "authors_source": "OpenReview API", "abstract": "The use of machine learning for modeling, understanding, and controlling large-scale physics systems is quickly gaining in popularity, with examples ranging from electromagnetism over nuclear fusion reactors and magneto-hydrodynamics to fluid mechanics and climate modeling. These systems — governed by partial differential equations — present unique challenges regarding the large number of degrees of freedom and the complex dynamics over many scales both in space and time, and additional measures to improve accuracy and sample efficiency are highly desirable. We present an end-to-end equivariant surrogate model consisting of an equivariant convolutional autoencoder and an equivariant convolutional LSTM using $G$-steerable kernels. As a case study, we consider the three-dimensional Rayleigh-Bénard convection, which describes the buoyancy-driven fluid flow between a heated bottom and a cooled top plate. While the system is E(2)-equivariant in the horizontal plane, the boundary conditions break the translational equivariance in the vertical direction. Our architecture leverages vertically stacked layers of $D_4$-steerable kernels, with additional partial kernel sharing in the vertical direction for further efficiency improvement. We demonstrate significant gains in sample and parameter efficiency, as well as a better scaling to more complex dynamics.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "r8VJqH6z8a", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19000/Reviewer_V5wy"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper presents a surrogate modeling approach for three-dimensional Rayleigh-Bénard convection (RBC) using deep learning methods that respect spatial symmetries. The authors propose a two-stage architecture consisting of an equivariant convolutional autoencoder (CAE) and an equivariant convolutional LSTM. The CAE compresses high-dimensional flow fields into a structured latent space, while preserving horizontal rotational and reflectional symmetries through D4-steerable convolutions. The latent dynamics are then modeled with a convolutional LSTM that predicts temporal evolution in this compressed space. The model aims to approximate the underlying partial differential equations governing the RBC system, reducing computational cost relative to full numerical simulation.\n\nThe model operates on three-dimensional fields that contain temperature and velocity information across space. These fields are first passed through an encoder, which compresses them into a lower-dimensional latent representation while preserving spatial structure. A decoder then reconstructs the original field from this compressed form. For modeling time evolution, a recurrent neural network, specifically a convolutional LSTM, is used to predict future latent states based on previous ones. The system is trained in two separate phases: first, the autoencoder is optimized to minimize reconstruction error between the input and the decoded output; second, the LSTM is trained to forecast latent states over time. The full pipeline is evaluated using data from high-fidelity simulations of Rayleigh-Bénard convection at different Rayleigh numbers. According to the results, the proposed method achieves better data efficiency, lower reconstruction and forecasting errors, and requires fewer parameters than both standard convolutional neural networks and Fourier Neural Operator baselines.", "review_text": "This paper presents a surrogate modeling approach for three-dimensional Rayleigh-Bénard convection (RBC) using deep learning methods that respect spatial symmetries. The authors propose a two-stage architecture consisting of an equivariant convolutional autoencoder (CAE) and an equivariant convolutional LSTM. The CAE compresses high-dimensional flow fields into a structured latent space, while preserving horizontal rotational and reflectional symmetries through D4-steerable convolutions. The latent dynamics are then modeled with a convolutional LSTM that predicts temporal evolution in this compressed space. The model aims to approximate the underlying partial differential equations governing the RBC system, reducing computational cost relative to full numerical simulation.\n\nThe model operates on three-dimensional fields that contain temperature and velocity information across space. These fields are first passed through an encoder, which compresses them into a lower-dimensional latent representation while preserving spatial structure. A decoder then reconstructs the original field from this compressed form. For modeling time evolution, a recurrent neural network, specifically a convolutional LSTM, is used to predict future latent states based on previous ones. The system is trained in two separate phases: first, the autoencoder is optimized to minimize reconstruction error between the input and the decoded output; second, the LSTM is trained to forecast latent states over time. The full pipeline is evaluated using data from high-fidelity simulations of Rayleigh-Bénard convection at different Rayleigh numbers. According to the results, the proposed method achieves better data efficiency, lower reconstruction and forecasting errors, and requires fewer parameters than both standard convolutional neural networks and Fourier Neural Operator baselines.", "strengths": "The methodology is technically consistent and leverages symmetry-aware design principles grounded in group representation theory. The explicit use of D4-steerable convolutions ensures equivariance to horizontal rotations and reflections, which is physically appropriate for Rayleigh-Bénard convection where boundary conditions break vertical but not horizontal symmetry. The decision to apply height-dependent kernels is a reasonable approximation to accommodate vertical heterogeneity in the flow. The decomposition of spatial compression and temporal evolution into separate modules (CAE and LSTM) leads to a well-structured and interpretable model pipeline. This separation also reduces computational complexity, as the temporal predictor operates on latent tensors rather than full-resolution fields.\n\nEmpirical evaluation includes relevant comparisons with non-equivariant models and established operator learning baselines. The reported improvement in reconstruction error (approximately 40\\% reduction in RMSE) and parameter efficiency (roughly an order of magnitude fewer parameters) is quantitatively clear. The experiments also cover different Rayleigh numbers, providing some evidence that the model generalizes to flows of varying complexity. The framework demonstrates that enforcing group-theoretic structure can lead to better inductive bias for spatiotemporal PDE systems.", "weaknesses": "The experimental validation remains limited in scope and does not fully establish the method’s robustness. All results are obtained on a single physical system (Rayleigh-Bénard convection) with idealized, noise-free data, which restricts the conclusions about general applicability. A joint end-to-end optimization might yield a more physically coherent latent space, though at higher cost. The study does not include ablations quantifying how much each design element (e.g., D4-steerable filters, local vertical parameter sharing) contributes to performance gains. Additionally, while the model is termed “equivariant,” it is only partially so: symmetry constraints are applied in horizontal planes but not in the vertical dimension, meaning the full $E(3)$ group is not represented. The reported efficiency gains are primarily relative to baseline models that do not exploit symmetries or that are trained under different regimes, which makes direct fairness of comparison uncertain. Finally, the paper does not explore stability or error accumulation over very long autoregressive rollouts, which is a key concern for temporal surrogate models.", "questions": "1. The model is trained in two separate phases: first the autoencoder, then the temporal predictor. While this improves training efficiency, it may introduce a mismatch between the latent representations learned by the encoder and those needed for accurate forecasting. Did the authors attempt any form of joint fine-tuning or end-to-end training, even partially? Can the authors comment on whether this decomposition leads to artifacts or long-horizon degradation in practice?\n\n2. The proposed model architecture includes several components motivated by physical and architectural reasoning (e.g., D4-steerable convolutions, height-dependent filters, vertical parameter sharing). However, no ablation studies are provided to assess their relative contribution. Can the authors provide experiments or discussion that isolate the impact of these design choices?\n\n3. Several baselines, such as U-Net and FNO, are included in the comparison. Were these models adapted for 3D input and trained with similar levels of supervision and data? FNO in particular is known to have strong performance in 2D PDE forecasting; was it adapted in a memory-efficient manner for 3D, or was the comparison constrained by hardware? \n\n4. Does the learned latent space exhibit any physical interpretability? For instance, do certain channels correspond to flow structures such as rolls or plumes? Could latent vectors be interpolated or manipulated to generate physically meaningful transitions in the decoded space?\n\n5. Although different Rayleigh numbers are used in training and evaluation, the setup assumes a fixed domain and set of boundary conditions. Would the model generalize across different physical configurations, such as changes in domain size, aspect ratio, or boundary heating profiles? If not, what would be required to make the surrogate model adaptive to such changes?\n\n6. All training data appears to be generated from numerical simulations. How robust is the model to noise or distributional shifts, as would be expected in real-world experimental measurements or lower-fidelity simulations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a surrogate modeling approach for three-dimensional Rayleigh-Bénard convection (RBC) using deep learning methods that respect spatial symmetries. The authors propose a two-stage architecture consisting of an equivariant convolutional autoencoder (CAE) and an equivariant convolutional LSTM. The CAE compresses high-dimensional flow fields into a structured latent space, while preserving horizontal rotational and reflectional symmetries through D4-steerable convolutions. The latent dynamics are then modeled with a convolutional LSTM that predicts temporal evolution in this compressed space. The model aims to approximate the underlying partial differential equations governing the RBC system, reducing computational cost relative to full numerical simulation.\n\nThe model operates on three-dimensional fields that contain temperature and velocity information across space. These fields are first passed through an encoder, which compresses them into a lower-dimensional latent representation while preserving spatial structure. A decoder then reconstructs the original field from this compressed form. For modeling time evolution, a recurrent neural network, specifically a convolutional LSTM, is used to predict future latent states based on previous ones. The system is trained in two separate phases: first, the autoencoder is optimized to minimize reconstruction error between the input and the decoded output; second, the LSTM is trained to forecast latent states over time. The full pipeline is evaluated using data from high-fidelity simulations of Rayleigh-Bénard convection at different Rayleigh numbers. According to the results, the proposed method achieves better data efficiency, lower reconstruction and forecasting errors, and requires fewer parameters than both standard convolutional neural networks and Fourier Neural Operator baselines.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The methodology is technically consistent and leverages symmetry-aware design principles grounded in group representation theory. The explicit use of D4-steerable convolutions ensures equivariance to horizontal rotations and reflections, which is physically appropriate for Rayleigh-Bénard convection where boundary conditions break vertical but not horizontal symmetry. The decision to apply height-dependent kernels is a reasonable approximation to accommodate vertical heterogeneity in the flow. The decomposition of spatial compression and temporal evolution into separate modules (CAE and LSTM) leads to a well-structured and interpretable model pipeline. This separation also reduces computational complexity, as the temporal predictor operates on latent tensors rather than full-resolution fields.\n\nEmpirical evaluation includes relevant comparisons with non-equivariant models and established operator learning baselines. The reported improvement in reconstruction error (approximately 40\\% reduction in RMSE) and parameter efficiency (roughly an order of magnitude fewer parameters) is quantitatively clear. The experiments also cover different Rayleigh numbers, providing some evidence that the model generalizes to flows of varying complexity. The framework demonstrates that enforcing group-theoretic structure can lead to better inductive bias for spatiotemporal PDE systems.", "weaknesses": "The experimental validation remains limited in scope and does not fully establish the method’s robustness. All results are obtained on a single physical system (Rayleigh-Bénard convection) with idealized, noise-free data, which restricts the conclusions about general applicability. A joint end-to-end optimization might yield a more physically coherent latent space, though at higher cost. The study does not include ablations quantifying how much each design element (e.g., D4-steerable filters, local vertical parameter sharing) contributes to performance gains. Additionally, while the model is termed “equivariant,” it is only partially so: symmetry constraints are applied in horizontal planes but not in the vertical dimension, meaning the full $E(3)$ group is not represented. The reported efficiency gains are primarily relative to baseline models that do not exploit symmetries or that are trained under different regimes, which makes direct fairness of comparison uncertain. Finally, the paper does not explore stability or error accumulation over very long autoregressive rollouts, which is a key concern for temporal surrogate models.", "questions": "1. The model is trained in two separate phases: first the autoencoder, then the temporal predictor. While this improves training efficiency, it may introduce a mismatch between the latent representations learned by the encoder and those needed for accurate forecasting. Did the authors attempt any form of joint fine-tuning or end-to-end training, even partially? Can the authors comment on whether this decomposition leads to artifacts or long-horizon degradation in practice?\n\n2. The proposed model architecture includes several components motivated by physical and architectural reasoning (e.g., D4-steerable convolutions, height-dependent filters, vertical parameter sharing). However, no ablation studies are provided to assess their relative contribution. Can the authors provide experiments or discussion that isolate the impact of these design choices?\n\n3. Several baselines, such as U-Net and FNO, are included in the comparison. Were these models adapted for 3D input and trained with similar levels of supervision and data? FNO in particular is known to have strong performance in 2D PDE forecasting; was it adapted in a memory-efficient manner for 3D, or was the comparison constrained by hardware? \n\n4. Does the learned latent space exhibit any physical interpretability? For instance, do certain channels correspond to flow structures such as rolls or plumes? Could latent vectors be interpolated or manipulated to generate physically meaningful transitions in the decoded space?\n\n5. Although different Rayleigh numbers are used in training and evaluation, the setup assumes a fixed domain and set of boundary conditions. Would the model generalize across different physical configurations, such as changes in domain size, aspect ratio, or boundary heating profiles? If not, what would be required to make the surrogate model adaptive to such changes?\n\n6. All training data appears to be generated from numerical simulations. How robust is the model to noise or distributional shifts, as would be expected in real-world experimental measurements or lower-fidelity simulations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761934531414}, {"id": "S2GACkyk8k", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19000/Reviewer_Uzfy"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper designs an equivariant neural architecture for learning the phenomenon of buoyancy-driven fluid flow between a heated bottom and a cooled top plate involved in climate simulation, incorporating equivariant properties of the corresponding PDE solution directly inside the neural architecture. The authors validate this approach by comparing it to non equivariant surrogate models, FNO and U-Net, on the learning problem at hand.", "review_text": "The paper designs an equivariant neural architecture for learning the phenomenon of buoyancy-driven fluid flow between a heated bottom and a cooled top plate involved in climate simulation, incorporating equivariant properties of the corresponding PDE solution directly inside the neural architecture. The authors validate this approach by comparing it to non equivariant surrogate models, FNO and U-Net, on the learning problem at hand.", "strengths": "- The paper introduce a novel architecture component by tailoring equivariant layers to their test case, with several innovating technical contributions.\n- The presented architecture, $D_4$- steerable, outperforms other baselines with far fewer model paramaters for both short and long time horizons.", "weaknesses": "**W1** The obtained architecture seems to be applicable to more diverse test cases that the one presented: the paper would benefit from experiments on other datasets, all the more since the presented dataset is not standard and of moderate scale.", "questions": "**Q1** In what other types of simulation could the equivariant CNN be useful?\n\n**Q2** l.321: \"We generated a dataset of 100 randomly initialized 3D Rayleigh-Bénard convection simulations with Ra = 2500 and P r = 0.7,\" Are you sampling only initial conditions ? If Ra is set to 2500, how can different value of Ra be tested in Figure 4 (rightmost plot) ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper designs an equivariant neural architecture for learning the phenomenon of buoyancy-driven fluid flow between a heated bottom and a cooled top plate involved in climate simulation, incorporating equivariant properties of the corresponding PDE solution directly inside the neural architecture. The authors validate this approach by comparing it to non equivariant surrogate models, FNO and U-Net, on the learning problem at hand.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper introduce a novel architecture component by tailoring equivariant layers to their test case, with several innovating technical contributions.\n- The presented architecture, $D_4$- steerable, outperforms other baselines with far fewer model paramaters for both short and long time horizons.", "weaknesses": "**W1** The obtained architecture seems to be applicable to more diverse test cases that the one presented: the paper would benefit from experiments on other datasets, all the more since the presented dataset is not standard and of moderate scale.", "questions": "**Q1** In what other types of simulation could the equivariant CNN be useful?\n\n**Q2** l.321: \"We generated a dataset of 100 randomly initialized 3D Rayleigh-Bénard convection simulations with Ra = 2500 and P r = 0.7,\" Are you sampling only initial conditions ? If Ra is set to 2500, how can different value of Ra be tested in Figure 4 (rightmost plot) ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761731273103}, {"id": "loE1S7iApP", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19000/Reviewer_gjPD"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors present a method to obtain a surrogate model of the 3D Rayleigh-Bénard convection equation which is equivariant over rotations and translations in the horizontal 2D direction. They use a encoder-process-decoder architecture: (1) encoder based on steerable kernels, (2) LSTM to predict rollout predictions whose MLPs are substituted with equivariant kernels, and (3) decoder with steerable kernels and trilinear interpolation upsampling. The composition of equivariant components make the whole model to be E(2) equivariant in the horizontal dimension and height-dependant learnable kernels in the vertical direction, where the symmetry is broken due to the buoyancy dynamics. The model is tested over the mentioned equation resulting in lower errors than the non-equivariant counterpart with significantly less parameters. A long-term forecasting test also shows that the method outperforms FNO and U-Net baselines.", "review_text": "The authors present a method to obtain a surrogate model of the 3D Rayleigh-Bénard convection equation which is equivariant over rotations and translations in the horizontal 2D direction. They use a encoder-process-decoder architecture: (1) encoder based on steerable kernels, (2) LSTM to predict rollout predictions whose MLPs are substituted with equivariant kernels, and (3) decoder with steerable kernels and trilinear interpolation upsampling. The composition of equivariant components make the whole model to be E(2) equivariant in the horizontal dimension and height-dependant learnable kernels in the vertical direction, where the symmetry is broken due to the buoyancy dynamics. The model is tested over the mentioned equation resulting in lower errors than the non-equivariant counterpart with significantly less parameters. A long-term forecasting test also shows that the method outperforms FNO and U-Net baselines.", "strengths": "* The compression ratio of the autoencoder is very high and the results outperform the baselines by a decent margin.\n* The local vertical parameter sharing is a smart solution to the vertical loss of symmetriy.\n* The paper treats equivariance with mathematical rigour and proofs.", "weaknesses": "* The paper focuses primarily on a single specific equation and lacks validation on more complex phenomena or dynamical scenarios. One would expect a method designed for equivariant modeling to demonstrate broader applicability to a variety of PDEs exhibiting such symmetries and symmetry-breaking, but this is not explored in the paper.\n* The contribution of the paper is quite limited. Equivariant autoencoders have been extensively explored in the literature, and the replacement of LSTM forward networks with an equivariant version represents only a modest contribution. The use of height-dependent kernels, however, is an interesting aspect.\n* The code is not ready for review.", "questions": "* Lines 324-325: Are all 900 timesteps within the interval [100, 1000] included in the training data? Does the model perform any extrapolation to snapshots outside this range?\n* Section 3.2: The autoencoder is evaluated across multiple Rayleigh numbers, which is appropriate given their strong influence on the system’s dynamics. However, the long-term forecasting experiments appear to be conducted for a single Rayleigh number, though this is not explicitly stated. The results would be more convincing if a similar multi-Ra number analysis were included for the forecasting stage as well.\n* Section 4.1.2: One of the main contributions of the paper is the introduction of height-dependent kernels, yet this aspect remains relatively underexplored. How sensitive are the results to the choice of kernel size or vertical resolution?\n* Line 965: The text mentions that the non-equivariant model was trained with data augmentation. Can the authors clarify whether similar augmentations (rotations, translations) were also applied to the long-term forecasting baselines (FNO, U-Net)? This would help ensure a fair comparison and clarify whether the performance gap stems primarily from architectural equivariance or from differences in training data diversity.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present a method to obtain a surrogate model of the 3D Rayleigh-Bénard convection equation which is equivariant over rotations and translations in the horizontal 2D direction. They use a encoder-process-decoder architecture: (1) encoder based on steerable kernels, (2) LSTM to predict rollout predictions whose MLPs are substituted with equivariant kernels, and (3) decoder with steerable kernels and trilinear interpolation upsampling. The composition of equivariant components make the whole model to be E(2) equivariant in the horizontal dimension and height-dependant learnable kernels in the vertical direction, where the symmetry is broken due to the buoyancy dynamics. The model is tested over the mentioned equation resulting in lower errors than the non-equivariant counterpart with significantly less parameters. A long-term forecasting test also shows that the method outperforms FNO and U-Net baselines.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* The compression ratio of the autoencoder is very high and the results outperform the baselines by a decent margin.\n* The local vertical parameter sharing is a smart solution to the vertical loss of symmetriy.\n* The paper treats equivariance with mathematical rigour and proofs.", "weaknesses": "* The paper focuses primarily on a single specific equation and lacks validation on more complex phenomena or dynamical scenarios. One would expect a method designed for equivariant modeling to demonstrate broader applicability to a variety of PDEs exhibiting such symmetries and symmetry-breaking, but this is not explored in the paper.\n* The contribution of the paper is quite limited. Equivariant autoencoders have been extensively explored in the literature, and the replacement of LSTM forward networks with an equivariant version represents only a modest contribution. The use of height-dependent kernels, however, is an interesting aspect.\n* The code is not ready for review.", "questions": "* Lines 324-325: Are all 900 timesteps within the interval [100, 1000] included in the training data? Does the model perform any extrapolation to snapshots outside this range?\n* Section 3.2: The autoencoder is evaluated across multiple Rayleigh numbers, which is appropriate given their strong influence on the system’s dynamics. However, the long-term forecasting experiments appear to be conducted for a single Rayleigh number, though this is not explicitly stated. The results would be more convincing if a similar multi-Ra number analysis were included for the forecasting stage as well.\n* Section 4.1.2: One of the main contributions of the paper is the introduction of height-dependent kernels, yet this aspect remains relatively underexplored. How sensitive are the results to the choice of kernel size or vertical resolution?\n* Line 965: The text mentions that the non-equivariant model was trained with data augmentation. Can the authors clarify whether similar augmentations (rotations, translations) were also applied to the long-term forecasting baselines (FNO, U-Net)? This would help ensure a fair comparison and clarify whether the performance gap stems primarily from architectural equivariance or from differences in training data diversity.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "I have no ethics concerns.", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761249262050}], "openreview_url": "https://openreview.net/forum?id=4jMeUvcO26", "arxiv_id": "2505.13569", "paper_pdf": "papers/4jMeUvcO26.pdf", "paper_pdf_sha256": "078526e509a839ac74972091b87b293a136c5b5abc3cab09602c766d8588b7b0", "paper_pdf_bytes": 6681465, "paper_pdf_source": "openreview", "code_url": "https://github.com/FynnFromme/equivariant-rb-forecasting", "code_repository": "FynnFromme/equivariant-rb-forecasting", "code_commit": "9a5b424834ab345e6e7654580930f7e99688930c", "code_archive": "repos/4jMeUvcO26.zip", "code_archive_sha256": "d9d38455987d615ab0f88fa48d63b52023bd764dad8b7cfb41357975e09fbaf3", "code_archive_bytes": 1110099, "code_file_count": 25, "code_extensions": {".py": 23, ".ipynb": 1, ".jl": 1}, "github_disk_usage_kb": 877, "github_languages": {"Python": 328557, "Jupyter Notebook": 33220, "Julia": 12727}, "github_archived": false, "github_pushed_at": "2025-05-21T11:54:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/surrogate-modeling-of-3d-rayleigh-benard"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8r8H4gbFXf", "year": 2025, "status": "rejected", "title": "Uncertainty Quantification in Retrieval Augmented Question Answering", "authors": ["Laura Perez-Beltrachini", "Mirella Lapata"], "authorids": ["~Laura_Perez-Beltrachini1", "~Mirella_Lapata1"], "authors_source": "OpenReview API", "abstract": "Retrieval augmented Question Answering (QA) enables QA models to overcome knowledge gaps when answering questions at test time by taking as input the question together with retrieved evidence, that is usually a set of passages.  Previous studies show that this approach has numerous benefits such as improving QA performance and reducing hallucinations, without, however, qualifying whether the retrieved passages are indeed useful at answering correctly. In this work, we evaluate existing uncertainty quantification approaches and propose an approach that predicts answer correctness based on utility judgements on individual input passages. We train a small neural model that predicts passage utility for a target QA model. We find that simple information theoretic metrics can predict answer correctness up to a certain extent, more expensive sampling based approaches perform better, while our lightweight approach can efficiently approximate or improve upon sampling-based approaches.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "T9KudUfymP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11054/Reviewer_zmTA"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes an approach to measure the uncertainty in Retrieval Augmented Question Answering tasks. \nConcretely, they train a small neural network called utility ranker which assigns a score for each retrieved passage from a given retriever to judge if the retrieved passage is useful for the answer generated by some QA model. \nThe authors show that this approach is on par or better than existing error prediction approaches while being light-weight at the same time.", "review_text": "This paper proposes an approach to measure the uncertainty in Retrieval Augmented Question Answering tasks. \nConcretely, they train a small neural network called utility ranker which assigns a score for each retrieved passage from a given retriever to judge if the retrieved passage is useful for the answer generated by some QA model. \nThe authors show that this approach is on par or better than existing error prediction approaches while being light-weight at the same time.", "strengths": "The authors ran experiments with 2 QA models and for a lot of the settings the utility ranker outperforms existing metrics in terms of uncertainty estimations of the retrieved passages. \nExperiments results also suggest that the method they proposed is also robust to OOD datasets where the ranker is not trained on.", "weaknesses": "- Many of the notations are unclear. See in Questions. \n- QA models used for evaluation only limit to Gemma2-9b-instruct and Llama3.1-8b-instruct which are of similar size. More experiments should be done using models with various sizes to see if similar conclusions still hold. \n- For Llama3.1-8b-instruct, results from table3 seems to suggest that Utility Ranker is not doing better than just looking at the probability of generating the next token to be \"True\". Is the training of this ranker really necessary?", "questions": "1. in (1), is m some hyper-parameter introduced in the model? If it was taken from other works, where did it come from? If it is optimized for this task, how did you optimize? \n2. the i and js from L_{rank} are never summed up in the total loss term. But I assume you do this for each retrieved passage pair, is that the case? \n3. How is the accuracy a defined at the bottom of page 3?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an approach to measure the uncertainty in Retrieval Augmented Question Answering tasks. \nConcretely, they train a small neural network called utility ranker which assigns a score for each retrieved passage from a given retriever to judge if the retrieved passage is useful for the answer generated by some QA model. \nThe authors show that this approach is on par or better than existing error prediction approaches while being light-weight at the same time.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The authors ran experiments with 2 QA models and for a lot of the settings the utility ranker outperforms existing metrics in terms of uncertainty estimations of the retrieved passages. \nExperiments results also suggest that the method they proposed is also robust to OOD datasets where the ranker is not trained on.", "weaknesses": "- Many of the notations are unclear. See in Questions. \n- QA models used for evaluation only limit to Gemma2-9b-instruct and Llama3.1-8b-instruct which are of similar size. More experiments should be done using models with various sizes to see if similar conclusions still hold. \n- For Llama3.1-8b-instruct, results from table3 seems to suggest that Utility Ranker is not doing better than just looking at the probability of generating the next token to be \"True\". Is the training of this ranker really necessary?", "questions": "1. in (1), is m some hyper-parameter introduced in the model? If it was taken from other works, where did it come from? If it is optimized for this task, how did you optimize? \n2. the i and js from L_{rank} are never summed up in the total loss term. But I assume you do this for each retrieved passage pair, is that the case? \n3. How is the accuracy a defined at the bottom of page 3?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731101476532}, {"id": "5Y2BYtHrnV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11054/Reviewer_NaAL"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes a novel approach for answer error prediction in retrieval augmented question answering. The premise is that the the retrieved passages and their interaction with the QA model’s parametric knowledge is a strong indicator of answer correctness. To measure this as a utility score for each passage, a small neural network is trained using a ranking loss - where the maximum utility score for each passage is the estimate for answer error prediction. On a few existing QA benchmarks (Natural Questions, TriviaQA, WebQuestions, SQuAD), this is shown to to be better than existing error prediction approaches based on entropy and resampling, while being more compute efficient.", "review_text": "The paper proposes a novel approach for answer error prediction in retrieval augmented question answering. The premise is that the the retrieved passages and their interaction with the QA model’s parametric knowledge is a strong indicator of answer correctness. To measure this as a utility score for each passage, a small neural network is trained using a ranking loss - where the maximum utility score for each passage is the estimate for answer error prediction. On a few existing QA benchmarks (Natural Questions, TriviaQA, WebQuestions, SQuAD), this is shown to to be better than existing error prediction approaches based on entropy and resampling, while being more compute efficient.", "strengths": "- The approach to train a separate smaller neural network to predict passage utility scores is novel. The construction of the data and loss for the scoring model, using entailment and accuracy is also intersting and original. \n- The paper provides an efficient way to predict the error rate at an example level, which could be very useful for latency sensitive systems in order to make a triggering decision for question answering.\n- The overall flow of the paper is good, it is succinctly written, and the experimental results are compelling and clearly presented. \n- The paper also touches upon the reranking approach to improve the performance of QA model using their utility scoring model, which seems potentially useful for some applications.", "weaknesses": "- One strong shortcoming of this approach is where multiple passages are needed to correctly answer the question, i.e. using multihop reasoning. In such cases, the utility both each of the passages in isolation could be low, and hurt the error prediction. Most of the baselines that use the entire passage set would be robust to this. \n- The modeling utility scores used to create  the ranking dataset has room for improvement. The scores could have smoother accuracy or entailment values instead of the binary values. And other, more principled aggregation functions could be explored instead of a simple average. \n- The evaluation for the utility ranker seems weak. The baseline in table 5 is not reranking at all. A better baseline could be a different utility ranker trained using the neural network, possibly with a simple objective such as predicting the error rate of the neutral network given x and p.", "questions": "- In line 200, why is e arg max? Could be a typo.\n- Did you compare the inference time difference between your approach and the baseline? It would be useful to see that comparison as well, since that's one of the key claims made.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel approach for answer error prediction in retrieval augmented question answering. The premise is that the the retrieved passages and their interaction with the QA model’s parametric knowledge is a strong indicator of answer correctness. To measure this as a utility score for each passage, a small neural network is trained using a ranking loss - where the maximum utility score for each passage is the estimate for answer error prediction. On a few existing QA benchmarks (Natural Questions, TriviaQA, WebQuestions, SQuAD), this is shown to to be better than existing error prediction approaches based on entropy and resampling, while being more compute efficient.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The approach to train a separate smaller neural network to predict passage utility scores is novel. The construction of the data and loss for the scoring model, using entailment and accuracy is also intersting and original. \n- The paper provides an efficient way to predict the error rate at an example level, which could be very useful for latency sensitive systems in order to make a triggering decision for question answering.\n- The overall flow of the paper is good, it is succinctly written, and the experimental results are compelling and clearly presented. \n- The paper also touches upon the reranking approach to improve the performance of QA model using their utility scoring model, which seems potentially useful for some applications.", "weaknesses": "- One strong shortcoming of this approach is where multiple passages are needed to correctly answer the question, i.e. using multihop reasoning. In such cases, the utility both each of the passages in isolation could be low, and hurt the error prediction. Most of the baselines that use the entire passage set would be robust to this. \n- The modeling utility scores used to create  the ranking dataset has room for improvement. The scores could have smoother accuracy or entailment values instead of the binary values. And other, more principled aggregation functions could be explored instead of a simple average. \n- The evaluation for the utility ranker seems weak. The baseline in table 5 is not reranking at all. A better baseline could be a different utility ranker trained using the neural network, possibly with a simple objective such as predicting the error rate of the neutral network given x and p.", "questions": "- In line 200, why is e arg max? Could be a typo.\n- Did you compare the inference time difference between your approach and the baseline? It would be useful to see that comparison as well, since that's one of the key claims made.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731095556172}, {"id": "oSI7BlzXBo", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11054/Reviewer_trrB"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This work looks at the task of uncertainty estimation in retrieval-augmented, open-domain question answering. The frame their task as predicting confidence estimates of a base retrieval-based QA model's predictions (experiments with Llama + Gemma) based on the set of retrieved passages and the input query. Their proposed approach is based on approach is based on training a small, separate *utility prediction* model that estimates the confidence in a base QA model prediction based on the input query and a single retrieved passage. To estimate confidence in a final prediction using a set of retrieved passages, they take the max predicted utility over all passages as the final confidence estimate.\n\nTo train this *utility prediction* model, the authors average (? -- see question below) the binary correctness score of the base QA model's prediction on a given question + retrieved passage and the predicted probability of question and QA model's predicted answer being entailed by the retrieved passage, treating this as a gold \"utility value\". The authors then train their smaller *utility prediction* to predict these utility values by summing two losses: (1) the BCE loss of the predicted utility against the gold utility and (2) a ranking loss between passage rankings obtained from the gold and predicted utility values.\n\nIn their experiments, the authors compare against calibration baselines (all are based primarily on using only the base LLM with sampling, prompting, and analyzing its predicted distributions). They train and evaluate on a variety of QA datasets using Gemma2, and see minor gains/losses when evaluating on NQ, TriviaQA, and Webquestions and a significant improvement when evaluating on SQUAD. The authors also repeat these experiments using LLAMA3 as the base QA model, and observe more mixed gains/losses over the baselines.", "review_text": "This work looks at the task of uncertainty estimation in retrieval-augmented, open-domain question answering. The frame their task as predicting confidence estimates of a base retrieval-based QA model's predictions (experiments with Llama + Gemma) based on the set of retrieved passages and the input query. Their proposed approach is based on approach is based on training a small, separate *utility prediction* model that estimates the confidence in a base QA model prediction based on the input query and a single retrieved passage. To estimate confidence in a final prediction using a set of retrieved passages, they take the max predicted utility over all passages as the final confidence estimate.\n\nTo train this *utility prediction* model, the authors average (? -- see question below) the binary correctness score of the base QA model's prediction on a given question + retrieved passage and the predicted probability of question and QA model's predicted answer being entailed by the retrieved passage, treating this as a gold \"utility value\". The authors then train their smaller *utility prediction* to predict these utility values by summing two losses: (1) the BCE loss of the predicted utility against the gold utility and (2) a ranking loss between passage rankings obtained from the gold and predicted utility values.\n\nIn their experiments, the authors compare against calibration baselines (all are based primarily on using only the base LLM with sampling, prompting, and analyzing its predicted distributions). They train and evaluate on a variety of QA datasets using Gemma2, and see minor gains/losses when evaluating on NQ, TriviaQA, and Webquestions and a significant improvement when evaluating on SQUAD. The authors also repeat these experiments using LLAMA3 as the base QA model, and observe more mixed gains/losses over the baselines.", "strengths": "This work presents a method for uncertainty estimation in retrieval-based QA. Their method trains a separate smaller LM to estimate uncertainty in the base QA system's predictions based on a passage, question, and predicted answer. This system is trained on", "weaknesses": "## Related Work + Baselines\nSimilar methods that use small, additional trained models to estimate uncertainty have been proposed by [1] and [2] ([1]  is referenced in related work, but not compared against). Additionally, [3] has also noted the overlap between this passage utility / calibration task and similarly uses pretrained NLI models to verify / estimate uncertainty in QA system predictions. Given the similarity of these methods, they are important points of comparison to understand how this method differs and how it affects performance. See point below.\n\n[1] Selective question answering under domain shift\nAmita Kamath, Robin Jia, Percy Liang\n\n[2] Knowing More About Questions Can Help: Improving Calibration in Question Answering\nShujian Zhang, Chenyue Gong, Eunsol Choi\n\n[3] Can NLI Models Verify QA Systems' Predictions?\nJifan Chen, Eunsol Choi, Greg Durrett \n\n## Evaluating role of the ranking loss and entailment score\nA significant novelty from this work from the related works above is the usage of (1) an additional passage ranking loss (in addition to standard BCE loss) and (2) using entailment score in addition to answer correctness as a gold label to train the \"passage utility predictor\"; however, the role and usefulness of these changes are unclear. Additional ablation experiments would be helpful for understanding the impact of these changes and their benefits.\n\n## Poor generalization to LLAMA3\nWhile the results on Gemma2 seem promising, results using LLAMA3 as the base QA system are generally mixed/negative. Experimenting with more base QA systems and performing significance testing may help bolster these results.", "questions": "(Note in Summary) In L162, is this in-line equation supposed to be the average of accuracy and entailment score?\n\nWhy are generalization experiments were limited to only GEMMA and training on NQ and evaluating on SQuAD, PopQA, RefuNQ? It would be interesting to see the performance using LLAMA (especially givent he negative results here) and training + evaluation on a greater number of dataset combinations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work looks at the task of uncertainty estimation in retrieval-augmented, open-domain question answering. The frame their task as predicting confidence estimates of a base retrieval-based QA model's predictions (experiments with Llama + Gemma) based on the set of retrieved passages and the input query. Their proposed approach is based on approach is based on training a small, separate *utility prediction* model that estimates the confidence in a base QA model prediction based on the input query and a single retrieved passage. To estimate confidence in a final prediction using a set of retrieved passages, they take the max predicted utility over all passages as the final confidence estimate.\n\nTo train this *utility prediction* model, the authors average (? -- see question below) the binary correctness score of the base QA model's prediction on a given question + retrieved passage and the predicted probability of question and QA model's predicted answer being entailed by the retrieved passage, treating this as a gold \"utility value\". The authors then train their smaller *utility prediction* to predict these utility values by summing two losses: (1) the BCE loss of the predicted utility against the gold utility and (2) a ranking loss between passage rankings obtained from the gold and predicted utility values.\n\nIn their experiments, the authors compare against calibration baselines (all are based primarily on using only the base LLM with sampling, prompting, and analyzing its predicted distributions). They train and evaluate on a variety of QA datasets using Gemma2, and see minor gains/losses when evaluating on NQ, TriviaQA, and Webquestions and a significant improvement when evaluating on SQUAD. The authors also repeat these experiments using LLAMA3 as the base QA model, and observe more mixed gains/losses over the baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "This work presents a method for uncertainty estimation in retrieval-based QA. Their method trains a separate smaller LM to estimate uncertainty in the base QA system's predictions based on a passage, question, and predicted answer. This system is trained on", "weaknesses": "## Related Work + Baselines\nSimilar methods that use small, additional trained models to estimate uncertainty have been proposed by [1] and [2] ([1]  is referenced in related work, but not compared against). Additionally, [3] has also noted the overlap between this passage utility / calibration task and similarly uses pretrained NLI models to verify / estimate uncertainty in QA system predictions. Given the similarity of these methods, they are important points of comparison to understand how this method differs and how it affects performance. See point below.\n\n[1] Selective question answering under domain shift\nAmita Kamath, Robin Jia, Percy Liang\n\n[2] Knowing More About Questions Can Help: Improving Calibration in Question Answering\nShujian Zhang, Chenyue Gong, Eunsol Choi\n\n[3] Can NLI Models Verify QA Systems' Predictions?\nJifan Chen, Eunsol Choi, Greg Durrett \n\n## Evaluating role of the ranking loss and entailment score\nA significant novelty from this work from the related works above is the usage of (1) an additional passage ranking loss (in addition to standard BCE loss) and (2) using entailment score in addition to answer correctness as a gold label to train the \"passage utility predictor\"; however, the role and usefulness of these changes are unclear. Additional ablation experiments would be helpful for understanding the impact of these changes and their benefits.\n\n## Poor generalization to LLAMA3\nWhile the results on Gemma2 seem promising, results using LLAMA3 as the base QA system are generally mixed/negative. Experimenting with more base QA systems and performing significance testing may help bolster these results.", "questions": "(Note in Summary) In L162, is this in-line equation supposed to be the average of accuracy and entailment score?\n\nWhy are generalization experiments were limited to only GEMMA and training on NQ and evaluating on SQuAD, PopQA, RefuNQ? It would be interesting to see the performance using LLAMA (especially givent he negative results here) and training + evaluation on a greater number of dataset combinations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731000770004}, {"id": "lkaRACDySq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11054/Reviewer_gKP3"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes a straightforward approach of using a small passage utility model to improve the calibration of larger LLM-based QA models; i.e. it proposes a method to predict the reliability of the LLM answer based on the utility of the retrieved passages.\n\nFor a question $q$, the set of retrieved passages $R$ = $[p_1, p_2, ..., p_{|R|}]$, and a QA model $M$, the utility of a passage $p \\in R$ is given by: $$ u = (a + e) / 2$$ where $a$ is the accuracy of the $M$ in predicting the ground-truth answer given passage $p$ and $e$ is the NLI entailment score of the question and predicted answer given the passage. A distillRoBERTa-based LM is trained to fit the utility scores. At inference time, the utility predictor assigns a score to each retrieved passage (given the question). The maximum utility score over all passages is used as the heuristic to abstain from answering.\n\nThe quality of different calibration techniques is compared on 4 QA datasets: NaturalQuestions, TriviaQA, WebQuestions, and SQuAD. The calibration techniques are compared on area under the rejection accuracy curve (calibration of abstaining) and AUROC of detecting incorrect answers. For two QA models, the trained utility predictor matches or improves over the performance of simple answer entropy-based heuristics. It is shown to be competitive with more complicated calibration techniques that rely on resampling multiple answers from the QA model.\n\nAn experiment is conducted to show that the utility predictor trained on NQ can be generalized out of distribution to SQuAD, PopQA, and RefuNQ. Moreover, the utility predictor can be used to rerank documents and improve QA accuracy.", "review_text": "The paper proposes a straightforward approach of using a small passage utility model to improve the calibration of larger LLM-based QA models; i.e. it proposes a method to predict the reliability of the LLM answer based on the utility of the retrieved passages.\n\nFor a question $q$, the set of retrieved passages $R$ = $[p_1, p_2, ..., p_{|R|}]$, and a QA model $M$, the utility of a passage $p \\in R$ is given by: $$ u = (a + e) / 2$$ where $a$ is the accuracy of the $M$ in predicting the ground-truth answer given passage $p$ and $e$ is the NLI entailment score of the question and predicted answer given the passage. A distillRoBERTa-based LM is trained to fit the utility scores. At inference time, the utility predictor assigns a score to each retrieved passage (given the question). The maximum utility score over all passages is used as the heuristic to abstain from answering.\n\nThe quality of different calibration techniques is compared on 4 QA datasets: NaturalQuestions, TriviaQA, WebQuestions, and SQuAD. The calibration techniques are compared on area under the rejection accuracy curve (calibration of abstaining) and AUROC of detecting incorrect answers. For two QA models, the trained utility predictor matches or improves over the performance of simple answer entropy-based heuristics. It is shown to be competitive with more complicated calibration techniques that rely on resampling multiple answers from the QA model.\n\nAn experiment is conducted to show that the utility predictor trained on NQ can be generalized out of distribution to SQuAD, PopQA, and RefuNQ. Moreover, the utility predictor can be used to rerank documents and improve QA accuracy.", "strengths": "1. The contribution is intuitive and straightforward. The utility predictor is shown to be a low-cost mechanism for improving the calibration of larger, more expensive QA models.\n2. The chosen experiments are appropriate. The experiment testing generalization across datasets is valuable and adds to the strength of the approach.\n    - Some concerns about \"completeness\" of experiments are raised int he next section\n3. The connection of utility prediction to passage reranking should be studied further. This is especially important since the utility predictor is shown to be an effect reranker.\n    - Can we use utility score as a stronger signal for rankers in general (not just for calibration)?", "weaknesses": "1. Several small details are missing in parts of the paper. Detailed questions are in the next section.\n2. One big issue with the experiment set-up is that the QA models are not instructed to abstain. Thus, it is unclear how any of the calibration methods improve over the inherent ability of the QA models to abstain.\n    - Under the current setup, even if the QA model abstains, it would be treated as \"incorrect\".\n3. Please include a discussion of connections to \"Evaluating Retrieval Quality in Retrieval-Augmented Generation\" (SIGIR 2024)\n    - They utilize a similar (query, passage) utility score for ranking retrieval systems", "questions": "1. What is the range of the utility scores? Based on the definition in Line 162, the value should be between [0, 1]. If so, then:\n    1. Why do you need a sigmoid in Eq (2)?\n    2. How are predicted utility scores $< 0$ or $> 1$ in Figure 1? \n2. Sec 3.1: Is $y_M$ the predicted model answer given just passage $p$ or given the full set $R$?\n3. Eq 1: Shouldn't the equation for margin loss be $max(0, m -y(u_i - u_j))$? i.e. the margin does not depend on $y$. What is actually implemented?\n4. Line 181: It is unclear how important the BCE loss is. Please report the results of ablating the BCE loss.\n5. Line 183: Is distillRoBERTa used as the utility predictor? How do you predict the utility score from the model? Please clarify the language in this line.\n6. Line 200: Notation of utility predictor $v$ is misleading since $v$ does not depend on the model $M$ after it is trained. If I am misunderstanding this, please clarify.\n7. Please include the ROC and RAC curves in the Appendix for completeness.\n8. Line 259: How is the manual inspection performed? Over how many samples?\n9. Line 269: It is unclear what you mean by \"levels\" here. Moreover, since all datasets are short-form QA, why do you believe clustering is affected?\n10. Table 4: Please report OOD evaluation results of Utility Ranker (NQ) on TriviaQA and WebQuestions. The reported results of SQuAD are important, but it seems to be a setting where the utility ranker performed significantly better than all baselines. The distinction between different calibration approaches on the two other datasets is less clear.\n\nTypos\n---\n- Line 134: Repeated \"between\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a straightforward approach of using a small passage utility model to improve the calibration of larger LLM-based QA models; i.e. it proposes a method to predict the reliability of the LLM answer based on the utility of the retrieved passages.\n\nFor a question $q$, the set of retrieved passages $R$ = $[p_1, p_2, ..., p_{|R|}]$, and a QA model $M$, the utility of a passage $p \\in R$ is given by: $$ u = (a + e) / 2$$ where $a$ is the accuracy of the $M$ in predicting the ground-truth answer given passage $p$ and $e$ is the NLI entailment score of the question and predicted answer given the passage. A distillRoBERTa-based LM is trained to fit the utility scores. At inference time, the utility predictor assigns a score to each retrieved passage (given the question). The maximum utility score over all passages is used as the heuristic to abstain from answering.\n\nThe quality of different calibration techniques is compared on 4 QA datasets: NaturalQuestions, TriviaQA, WebQuestions, and SQuAD. The calibration techniques are compared on area under the rejection accuracy curve (calibration of abstaining) and AUROC of detecting incorrect answers. For two QA models, the trained utility predictor matches or improves over the performance of simple answer entropy-based heuristics. It is shown to be competitive with more complicated calibration techniques that rely on resampling multiple answers from the QA model.\n\nAn experiment is conducted to show that the utility predictor trained on NQ can be generalized out of distribution to SQuAD, PopQA, and RefuNQ. Moreover, the utility predictor can be used to rerank documents and improve QA accuracy.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The contribution is intuitive and straightforward. The utility predictor is shown to be a low-cost mechanism for improving the calibration of larger, more expensive QA models.\n2. The chosen experiments are appropriate. The experiment testing generalization across datasets is valuable and adds to the strength of the approach.\n    - Some concerns about \"completeness\" of experiments are raised int he next section\n3. The connection of utility prediction to passage reranking should be studied further. This is especially important since the utility predictor is shown to be an effect reranker.\n    - Can we use utility score as a stronger signal for rankers in general (not just for calibration)?", "weaknesses": "1. Several small details are missing in parts of the paper. Detailed questions are in the next section.\n2. One big issue with the experiment set-up is that the QA models are not instructed to abstain. Thus, it is unclear how any of the calibration methods improve over the inherent ability of the QA models to abstain.\n    - Under the current setup, even if the QA model abstains, it would be treated as \"incorrect\".\n3. Please include a discussion of connections to \"Evaluating Retrieval Quality in Retrieval-Augmented Generation\" (SIGIR 2024)\n    - They utilize a similar (query, passage) utility score for ranking retrieval systems", "questions": "1. What is the range of the utility scores? Based on the definition in Line 162, the value should be between [0, 1]. If so, then:\n    1. Why do you need a sigmoid in Eq (2)?\n    2. How are predicted utility scores $< 0$ or $> 1$ in Figure 1? \n2. Sec 3.1: Is $y_M$ the predicted model answer given just passage $p$ or given the full set $R$?\n3. Eq 1: Shouldn't the equation for margin loss be $max(0, m -y(u_i - u_j))$? i.e. the margin does not depend on $y$. What is actually implemented?\n4. Line 181: It is unclear how important the BCE loss is. Please report the results of ablating the BCE loss.\n5. Line 183: Is distillRoBERTa used as the utility predictor? How do you predict the utility score from the model? Please clarify the language in this line.\n6. Line 200: Notation of utility predictor $v$ is misleading since $v$ does not depend on the model $M$ after it is trained. If I am misunderstanding this, please clarify.\n7. Please include the ROC and RAC curves in the Appendix for completeness.\n8. Line 259: How is the manual inspection performed? Over how many samples?\n9. Line 269: It is unclear what you mean by \"levels\" here. Moreover, since all datasets are short-form QA, why do you believe clustering is affected?\n10. Table 4: Please report OOD evaluation results of Utility Ranker (NQ) on TriviaQA and WebQuestions. The reported results of SQuAD are important, but it seems to be a setting where the utility ranker performed significantly better than all baselines. The distinction between different calibration approaches on the two other datasets is less clear.\n\nTypos\n---\n- Line 134: Repeated \"between\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730925944444}, {"id": "rAxpAJRX6Z", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11054/Reviewer_575d"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper evaluates existing uncertainty quantification approaches that are used to quantify whether the retrieved passages at test time are useful to answer the question correctly. This paper also propose a neural-model based utility ranker that predicts answer correctness based on utility judgements on individual input passage, which boost the accuracy by 4% for NQ, and 2% for WebQ. The utility ranker trumps some uncertainty detection method for some datasets for the Gemma model. However, more analysis and explanations could be done.", "review_text": "This paper evaluates existing uncertainty quantification approaches that are used to quantify whether the retrieved passages at test time are useful to answer the question correctly. This paper also propose a neural-model based utility ranker that predicts answer correctness based on utility judgements on individual input passage, which boost the accuracy by 4% for NQ, and 2% for WebQ. The utility ranker trumps some uncertainty detection method for some datasets for the Gemma model. However, more analysis and explanations could be done.", "strengths": "- Previous work for QA error detection are either expensive to run for in-production QA systems, or rely on model’s internal confidence, and are for closed-book QA, so they are not applicable to retrieval augmented setup. The proposed utility ranker,\n- The utility ranker differs from Asai et al. (2024) because Asai et al. (2024) uses external critic model to judge, while the proposed method is target QA model based.\n- The baseline methods are thoroughly tested in both table 2, table 3 and table 4.\n- After applying utility ranker to filter out irrelevant passages, the accuracy and accuracy LM increases for both model on NQ and WebQ.", "weaknesses": "- Section 3 overcomplicates the method, and some of the math definitions are confusing instead of explaining the details. If I understand it correctly, the usefulness of a passage $p$ is defined on whether the model can correctly answer the question with the passage. The utility score is defined as the mean of accuracy and entailment score, where both scores are binary, so the only possible value for $u$ is 0, 0.5, or 1. Then it is combined with a binary cross entropy objective, and train a Siamese network that uses DistilRoBERTa to encode text pairs, and then use the ALBERT-xlarge model, trained on MNLI and VitaminC, is used to determine the entailment.\n- Result analysis could be done for section 5.3. Although both Acc and AccLM is improved by utility ranker, some explanations is appreciated. Why is the analysis only done for NQ and WebQ? Both Accuracy and AccuracyLM increases the same amount, is it a coincidence or one metric is enough.", "questions": "- Is the $m$ in equation 1 a hyper-parameter?\n- In equation 2, the summation is over $u_i$ and $u_j$, but they didn’t show up in the equation. You also mention $p(y) = \\sigmoid(u)$, but is it $u_i$ or $u_j$?\n- The citation format needs fixing, not limited to:\n    - Line 208: (PMI; Takayama & Arase, 2019), as well as the citation for p(true) on line 211 needs fixing.\n    - Line 218: Holtzman et al. 2020.\n    - Line 224, Gemma2-9B-Instruct (Riviere et al., 2024), and line 227 for contriever.\n    - Line 232: You can use \\citep for multiple citations, and use comma to separate each citation.\n- Missing citation: Top-k sampling is from Fan et al. (2018).\n- Why do you select |R| = 3 for table 5 rather than |R| = 5?\n- Is there analysis about when and what |R| people should use? Does the effect enhance or decrease when |R| change? Is the method still relevant if there are more than X number of passages?\n- It seems like utility ranker works better on Gemma rather than Llama, if experiment could be run on other models to confirm that utility ranker does work for most models, that would be wonderful.\n\n## Reference\n\nFan, Angela, Mike Lewis, and Yann Dauphin. \"Hierarchical neural story generation.\" *arXiv preprint arXiv:1805.04833*(2018).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper evaluates existing uncertainty quantification approaches that are used to quantify whether the retrieved passages at test time are useful to answer the question correctly. This paper also propose a neural-model based utility ranker that predicts answer correctness based on utility judgements on individual input passage, which boost the accuracy by 4% for NQ, and 2% for WebQ. The utility ranker trumps some uncertainty detection method for some datasets for the Gemma model. However, more analysis and explanations could be done.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Previous work for QA error detection are either expensive to run for in-production QA systems, or rely on model’s internal confidence, and are for closed-book QA, so they are not applicable to retrieval augmented setup. The proposed utility ranker,\n- The utility ranker differs from Asai et al. (2024) because Asai et al. (2024) uses external critic model to judge, while the proposed method is target QA model based.\n- The baseline methods are thoroughly tested in both table 2, table 3 and table 4.\n- After applying utility ranker to filter out irrelevant passages, the accuracy and accuracy LM increases for both model on NQ and WebQ.", "weaknesses": "- Section 3 overcomplicates the method, and some of the math definitions are confusing instead of explaining the details. If I understand it correctly, the usefulness of a passage $p$ is defined on whether the model can correctly answer the question with the passage. The utility score is defined as the mean of accuracy and entailment score, where both scores are binary, so the only possible value for $u$ is 0, 0.5, or 1. Then it is combined with a binary cross entropy objective, and train a Siamese network that uses DistilRoBERTa to encode text pairs, and then use the ALBERT-xlarge model, trained on MNLI and VitaminC, is used to determine the entailment.\n- Result analysis could be done for section 5.3. Although both Acc and AccLM is improved by utility ranker, some explanations is appreciated. Why is the analysis only done for NQ and WebQ? Both Accuracy and AccuracyLM increases the same amount, is it a coincidence or one metric is enough.", "questions": "- Is the $m$ in equation 1 a hyper-parameter?\n- In equation 2, the summation is over $u_i$ and $u_j$, but they didn’t show up in the equation. You also mention $p(y) = \\sigmoid(u)$, but is it $u_i$ or $u_j$?\n- The citation format needs fixing, not limited to:\n    - Line 208: (PMI; Takayama & Arase, 2019), as well as the citation for p(true) on line 211 needs fixing.\n    - Line 218: Holtzman et al. 2020.\n    - Line 224, Gemma2-9B-Instruct (Riviere et al., 2024), and line 227 for contriever.\n    - Line 232: You can use \\citep for multiple citations, and use comma to separate each citation.\n- Missing citation: Top-k sampling is from Fan et al. (2018).\n- Why do you select |R| = 3 for table 5 rather than |R| = 5?\n- Is there analysis about when and what |R| people should use? Does the effect enhance or decrease when |R| change? Is the method still relevant if there are more than X number of passages?\n- It seems like utility ranker works better on Gemma rather than Llama, if experiment could be run on other models to confirm that utility ranker does work for most models, that would be wonderful.\n\n## Reference\n\nFan, Angela, Mike Lewis, and Yann Dauphin. \"Hierarchical neural story generation.\" *arXiv preprint arXiv:1805.04833*(2018).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730594371235}], "openreview_url": "https://openreview.net/forum?id=8r8H4gbFXf", "arxiv_id": "2502.18108", "paper_pdf": "papers/8r8H4gbFXf.pdf", "paper_pdf_sha256": "9814bc713f59f0fd4ce2ca8fe33a803161dc7fff07f8aab156f4535c41988759", "paper_pdf_bytes": 467337, "paper_pdf_source": "openreview", "code_url": "https://github.com/lauhaide/ragu", "code_repository": "lauhaide/ragu", "code_commit": "15c317f8a070ac9073c850bda09d181aaf7be1f8", "code_archive": "repos/8r8H4gbFXf.zip", "code_archive_sha256": "7cbf7a31393b4b32bd32155a7462b3aea24ea19d56ecd7139195190f685f5222", "code_archive_bytes": 85301, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 74, "github_languages": {"Python": 259896, "Dockerfile": 506}, "github_archived": false, "github_pushed_at": "2025-02-25T10:36:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/uncertainty-quantification-in-retrieval"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "20KYsQ8Q4Z", "year": 2024, "status": "rejected", "title": "High-dimensional Bayesian Optimization with Group Testing", "authors": ["Erik Orm Hellsten", "Carl Hvarfner", "Leonard Papenmeier", "Luigi Nardi"], "authorids": ["~Erik_Orm_Hellsten1", "~Carl_Hvarfner1", "~Leonard_Papenmeier1", "~Luigi_Nardi1"], "authors_source": "OpenReview API", "abstract": "Bayesian optimization is an effective method for optimizing expensive-to-evaluate black-box functions. \nHigh-dimensional problems are particularly challenging as the surrogate model of the objective suffers from the curse of dimensionality, which makes accurate modeling difficult. \nWe propose a group testing approach to identify active variables to facilitate efficient optimization in these domains. \nThe proposed algorithm, Group Testing Bayesian Optimization (GTBO), first runs a testing phase where groups of variables are systematically selected and tested on whether they influence the objective. \nTo that end, we extend the well-established theory of group testing to functions of continuous ranges.\nIn the second phase, GTBO guides optimization by placing more importance on the active dimensions.\nBy exploiting the axis-aligned subspace assumption, GTBO is competitive against state-of-the-art methods on several synthetic and real-world high-dimensional optimization tasks. \nFurthermore, GTBO aids in the discovery of active parameters in applications, thereby enhancing practitioners' understanding of the problem at hand.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "vRAgNCEE2a", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7837/Reviewer_9ChC"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces GTBO which introduces ideas from feature selection literature and Group testing theory to the problem of selecting relevant features to reduce dimensionality of high dimensional BO problems and reduce the effect of pathology of curse of dimensionality for such problems. The total budget is divided into two halves, first the relevant features are identified using first set of function evaluations to create an active subspace of features, and then by assigning different priors on lengthscales for each set of relevant and irrelevant features the remaining budget is used for BO. Experiments are carried out on simulated popular datasets and two real world datasets and the performance of proposed method is compared against existing algorithms. The proposed model is probabilistic and can handle noisy observations. The method seems to do well and with the inherent advantage being that it is more interpretable than projection based dimenionality reduction methods.", "review_text": "The paper introduces GTBO which introduces ideas from feature selection literature and Group testing theory to the problem of selecting relevant features to reduce dimensionality of high dimensional BO problems and reduce the effect of pathology of curse of dimensionality for such problems. The total budget is divided into two halves, first the relevant features are identified using first set of function evaluations to create an active subspace of features, and then by assigning different priors on lengthscales for each set of relevant and irrelevant features the remaining budget is used for BO. Experiments are carried out on simulated popular datasets and two real world datasets and the performance of proposed method is compared against existing algorithms. The proposed model is probabilistic and can handle noisy observations. The method seems to do well and with the inherent advantage being that it is more interpretable than projection based dimenionality reduction methods.", "strengths": "1. The paper is mostly well written.\n2. The baselines and relevant literature is well covered and duly introduced to the readers.\n3. The results from the experiments suggest that the nmethod works well as compared to baselines both on simulated and real world datasets.\n4. I think it is a strenght of the method that it combines the advantages of interpretibility which come along with feature selection compared to projection based approaches so the user gets to understand his data as he is performing BO. \n5. The math and equations look fine to me. \n6. It is great the authors carry out and report sensitivity analysis and ablation study in Appendix and main paper.", "weaknesses": "1. The model makes many assumptions that the features are relatively independent, since for highly correlated features, it might not be possible to break them into sets of active and inactive features without knowing their correlations beforehand. Assuming that the probabilities of dimensions to be active are independent, is rather a strong simplifying assumption and will not hold in many practical datasets and situations. \n2. The paper does not list its own limitations properly. \n3. Certain choice of parameters for instance :  $\\sqrt(D)$ to be the value of active dimensions, choice of prior: logNormal with particular values of location and scale parameters can be better motivated. Why logNormal and not Gamma for instance, which is common hyperprior for lengthscale ? \n4. Maybe have one more real world dataset.\n5. The writing of the Experiment section can be improved, because somehow the flow of information is not good, as the authors introduce the figures in a weird order (Minor) and do a bit of back and forth. \n6.. Maybe make the lines in plots thicker.", "questions": "Minor comments and questions:\n1. What is the value of $C_{lower}$ and $C_{upper}$, apologies if I missed it.\n2.  Why did the authors use the particular acquistion function which they did ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces GTBO which introduces ideas from feature selection literature and Group testing theory to the problem of selecting relevant features to reduce dimensionality of high dimensional BO problems and reduce the effect of pathology of curse of dimensionality for such problems. The total budget is divided into two halves, first the relevant features are identified using first set of function evaluations to create an active subspace of features, and then by assigning different priors on lengthscales for each set of relevant and irrelevant features the remaining budget is used for BO. Experiments are carried out on simulated popular datasets and two real world datasets and the performance of proposed method is compared against existing algorithms. The proposed model is probabilistic and can handle noisy observations. The method seems to do well and with the inherent advantage being that it is more interpretable than projection based dimenionality reduction methods.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The paper is mostly well written.\n2. The baselines and relevant literature is well covered and duly introduced to the readers.\n3. The results from the experiments suggest that the nmethod works well as compared to baselines both on simulated and real world datasets.\n4. I think it is a strenght of the method that it combines the advantages of interpretibility which come along with feature selection compared to projection based approaches so the user gets to understand his data as he is performing BO. \n5. The math and equations look fine to me. \n6. It is great the authors carry out and report sensitivity analysis and ablation study in Appendix and main paper.", "weaknesses": "1. The model makes many assumptions that the features are relatively independent, since for highly correlated features, it might not be possible to break them into sets of active and inactive features without knowing their correlations beforehand. Assuming that the probabilities of dimensions to be active are independent, is rather a strong simplifying assumption and will not hold in many practical datasets and situations. \n2. The paper does not list its own limitations properly. \n3. Certain choice of parameters for instance :  $\\sqrt(D)$ to be the value of active dimensions, choice of prior: logNormal with particular values of location and scale parameters can be better motivated. Why logNormal and not Gamma for instance, which is common hyperprior for lengthscale ? \n4. Maybe have one more real world dataset.\n5. The writing of the Experiment section can be improved, because somehow the flow of information is not good, as the authors introduce the figures in a weird order (Minor) and do a bit of back and forth. \n6.. Maybe make the lines in plots thicker.", "questions": "Minor comments and questions:\n1. What is the value of $C_{lower}$ and $C_{upper}$, apologies if I missed it.\n2.  Why did the authors use the particular acquistion function which they did ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698773566943}, {"id": "n56ozyxk6X", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7837/Reviewer_yfZG"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces an algorithm for high-dimensional Bayesian optimization (BO), called GTBO (Group Testing Bayesian Optimization). The algorithm explicitly divides high-dimensional BO into two steps: in the first, a set of group testing experiments are run to probabilistically identify inactive input dimensions. In the second, BO is run with a relatively standard method (Matern 5/2 with qLogNEI acquisition function), while applying different length-scale priors for active vs. inactive dimensions.\n\n------ AFTER AUTHOR RESPONSE -----\n\nThanks for these clarifications. I definitely misunderstood how the GT iterations were treated in the computational results, which the authors have pointed out. I have raised my Contribution and Overall scores to account for this mis-judging of the results.\n\nThere are still some points I believe could be improved--the authors also note both points in their response. Firstly, the presentation could focus on this as a feature selection method rather than a BO algorithm. It is a great feature that GTBO actively selects sampling points in contrast to \"traditional feature selection methods,\" but the entire GT step is still de-coupled from the latter BO step, unless I am mistaken here. Secondly, it would be nice to see some more realistic benchmark problem(s).", "review_text": "This paper introduces an algorithm for high-dimensional Bayesian optimization (BO), called GTBO (Group Testing Bayesian Optimization). The algorithm explicitly divides high-dimensional BO into two steps: in the first, a set of group testing experiments are run to probabilistically identify inactive input dimensions. In the second, BO is run with a relatively standard method (Matern 5/2 with qLogNEI acquisition function), while applying different length-scale priors for active vs. inactive dimensions.\n\n------ AFTER AUTHOR RESPONSE -----\n\nThanks for these clarifications. I definitely misunderstood how the GT iterations were treated in the computational results, which the authors have pointed out. I have raised my Contribution and Overall scores to account for this mis-judging of the results.\n\nThere are still some points I believe could be improved--the authors also note both points in their response. Firstly, the presentation could focus on this as a feature selection method rather than a BO algorithm. It is a great feature that GTBO actively selects sampling points in contrast to \"traditional feature selection methods,\" but the entire GT step is still de-coupled from the latter BO step, unless I am mistaken here. Secondly, it would be nice to see some more realistic benchmark problem(s).", "strengths": "1. The presentation of the group testing methodology clearly explains the mathematical foundation and practical implementation details of the proposed algorithm.\n1. The computational tests are evaluated on several synthetic problems and also real-world benchmarks.", "weaknesses": "1. While this work is presented as a new BO algorithm, it is effectively a feature selection algorithm. The proposed group testing algorithm could select active dimensions to be optimized by any standard BO algorithm (e.g., ignoring the inactive dimensions). Likewise, a different feature selection method could be followed by the employed BO, which is relatively standard.\n1. The motivation for the benchmark problems needs to be strengthened. The synthetic benchmarks have 2-8 active dimensions and approximately 300 active dimensions without justification for this setting. The real-world benchmarks have no noise, again without justification.\n1. From what I understand, the comparisons for GTBO do not include the 39-112 iterations for group testing, which are used as the initial sample points for BO. Therefore, in Figs 3-4, where the other algorithms are starting from 0 function evaluations, GTBO is already starting at many.", "questions": "1. The description of batch evaluations on pg 6 needs significant clarification, i.e., how many is “several,” and how close is “close”? If batch sampling is available, why doesn’t the user just use the maximum number of batches for every group?\n1. How are the log-normal length scales for inactive dimensions determined?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an algorithm for high-dimensional Bayesian optimization (BO), called GTBO (Group Testing Bayesian Optimization). The algorithm explicitly divides high-dimensional BO into two steps: in the first, a set of group testing experiments are run to probabilistically identify inactive input dimensions. In the second, BO is run with a relatively standard method (Matern 5/2 with qLogNEI acquisition function), while applying different length-scale priors for active vs. inactive dimensions.\n\n------ AFTER AUTHOR RESPONSE -----\n\nThanks for these clarifications. I definitely misunderstood how the GT iterations were treated in the computational results, which the authors have pointed out. I have raised my Contribution and Overall scores to account for this mis-judging of the results.\n\nThere are still some points I believe could be improved--the authors also note both points in their response. Firstly, the presentation could focus on this as a feature selection method rather than a BO algorithm. It is a great feature that GTBO actively selects sampling points in contrast to \"traditional feature selection methods,\" but the entire GT step is still de-coupled from the latter BO step, unless I am mistaken here. Secondly, it would be nice to see some more realistic benchmark problem(s).", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The presentation of the group testing methodology clearly explains the mathematical foundation and practical implementation details of the proposed algorithm.\n1. The computational tests are evaluated on several synthetic problems and also real-world benchmarks.", "weaknesses": "1. While this work is presented as a new BO algorithm, it is effectively a feature selection algorithm. The proposed group testing algorithm could select active dimensions to be optimized by any standard BO algorithm (e.g., ignoring the inactive dimensions). Likewise, a different feature selection method could be followed by the employed BO, which is relatively standard.\n1. The motivation for the benchmark problems needs to be strengthened. The synthetic benchmarks have 2-8 active dimensions and approximately 300 active dimensions without justification for this setting. The real-world benchmarks have no noise, again without justification.\n1. From what I understand, the comparisons for GTBO do not include the 39-112 iterations for group testing, which are used as the initial sample points for BO. Therefore, in Figs 3-4, where the other algorithms are starting from 0 function evaluations, GTBO is already starting at many.", "questions": "1. The description of batch evaluations on pg 6 needs significant clarification, i.e., how many is “several,” and how close is “close”? If batch sampling is available, why doesn’t the user just use the maximum number of batches for every group?\n1. How are the log-normal length scales for inactive dimensions determined?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698687476890}, {"id": "6VRlm1kVdA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7837/Reviewer_fy4F"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes GTBO, a high-dimensional Bayesian optimization method which first uses adaptive group testing to identify active dimensions and then optimize over the active variables. The authors extend the binary group testing method to continuous space via maximizing the multual information estimated from Gaussian process. Experiment on synthetic functions and two real-world benchmarks demonstrate the efficiency of GTBO.", "review_text": "This paper proposes GTBO, a high-dimensional Bayesian optimization method which first uses adaptive group testing to identify active dimensions and then optimize over the active variables. The authors extend the binary group testing method to continuous space via maximizing the multual information estimated from Gaussian process. Experiment on synthetic functions and two real-world benchmarks demonstrate the efficiency of GTBO.", "strengths": "1. The proposed group testing idea is clear and easy to follow.\n\n2. The experiment shows that using group testing can efficiently identify active variables.", "weaknesses": "1. I think the experiment comparison is not fair, where the best initial point of GTBO is always better than baselines, which gives additional advantage to GTBO.\n\n2. I think the work of MCTS-VS[1] is also a HDBO method using variable selection, which is a similar work as GTBO and should be added into the baselines. \n\n[1] Song, Lei, et al. \"Monte carlo tree search based variable selection for high dimensional bayesian optimization.\" Advances in Neural Information Processing Systems 35 (2022): 28488-28501.", "questions": "1. As mentioned in weakness part, the initial points are different between GTBO and other baselines. Can you show the result when using same initial points in baselines?\n\n2. Why choosing the search center as the default configuration? What is the performance of GTBO when choosing default configuration as other position (e.g. best point in the random sampled initial dataset)?\n\n3. What is the search bound in the benchmark you used? Does GTBO utilize the advantage of symmetric search space as BAxUS?\n\n4. What is the batch size used in the experiment? The paper mention that \"GTBO integrates well with batch BO pipelines with little to no performance degradation\". Is there any experiment result to support this statement?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes GTBO, a high-dimensional Bayesian optimization method which first uses adaptive group testing to identify active dimensions and then optimize over the active variables. The authors extend the binary group testing method to continuous space via maximizing the multual information estimated from Gaussian process. Experiment on synthetic functions and two real-world benchmarks demonstrate the efficiency of GTBO.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The proposed group testing idea is clear and easy to follow.\n\n2. The experiment shows that using group testing can efficiently identify active variables.", "weaknesses": "1. I think the experiment comparison is not fair, where the best initial point of GTBO is always better than baselines, which gives additional advantage to GTBO.\n\n2. I think the work of MCTS-VS[1] is also a HDBO method using variable selection, which is a similar work as GTBO and should be added into the baselines. \n\n[1] Song, Lei, et al. \"Monte carlo tree search based variable selection for high dimensional bayesian optimization.\" Advances in Neural Information Processing Systems 35 (2022): 28488-28501.", "questions": "1. As mentioned in weakness part, the initial points are different between GTBO and other baselines. Can you show the result when using same initial points in baselines?\n\n2. Why choosing the search center as the default configuration? What is the performance of GTBO when choosing default configuration as other position (e.g. best point in the random sampled initial dataset)?\n\n3. What is the search bound in the benchmark you used? Does GTBO utilize the advantage of symmetric search space as BAxUS?\n\n4. What is the batch size used in the experiment? The paper mention that \"GTBO integrates well with batch BO pipelines with little to no performance degradation\". Is there any experiment result to support this statement?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698639102655}, {"id": "DJVlWFaJej", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7837/Reviewer_JQFf"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This submission tackles the case of high-dimensional optimization using Bayesian Optimization (BO), a sample-efficient method that has shown great success for hyperparameter tuning of large deep learning models, or in more concrete applications such as recommender systems tuning. It is therefore relevant for a venue such as ICLR, in my opinion. More particularly, the authors propose to tailor the so-called *group testing* theory to the BO setting. This is achieved by 1) extending group testing to real-valued functions and 2) dividing a usual BO run into two steps, a first step of relevant variable identification through group testing, and a second step of conventional BO, with the learned variable relevance being encoded in the Gaussian process surrogate lengthscales. The proposed method, *GTBO*, not only achieves state-of-the-art results of synthetic and real-world experiments but also enlights the practitioner with a ranking of the relevant variables, thus enhancing its understanding of the problem.", "review_text": "This submission tackles the case of high-dimensional optimization using Bayesian Optimization (BO), a sample-efficient method that has shown great success for hyperparameter tuning of large deep learning models, or in more concrete applications such as recommender systems tuning. It is therefore relevant for a venue such as ICLR, in my opinion. More particularly, the authors propose to tailor the so-called *group testing* theory to the BO setting. This is achieved by 1) extending group testing to real-valued functions and 2) dividing a usual BO run into two steps, a first step of relevant variable identification through group testing, and a second step of conventional BO, with the learned variable relevance being encoded in the Gaussian process surrogate lengthscales. The proposed method, *GTBO*, not only achieves state-of-the-art results of synthetic and real-world experiments but also enlights the practitioner with a ranking of the relevant variables, thus enhancing its understanding of the problem.", "strengths": "- The paper is well-written and organized in an easy-to-follow manner.\n- The approach is simple and works remarkably well.\n- The benchmarks include many competing methods, although some recent ones could have been considered as well, e.g. [1,2]\n\n\n[1] Sparse Bayesian optimization. AISTATS 2023.\n[2] Are Random Decompositions all we need in High Dimensional Bayesian Optimisation? ICML 2023.", "weaknesses": "- As often, the approach involves several hyperparameters, to determine whether a variable is deemed as relevant or not, and then incorporating this information in the BO statistical surrogate using carefully-designed priors. \n\nOther than that, I have to say I cannot really spot any weakness here. I am not very familiar with the group testing framework.", "questions": "I am genuinely surprised by how good *GTBO* is compared to other competitors, given that the proposed approach feels suboptimal. It performs in a sequential manner, where one first does not care about function maximization, only about variable relevance, even though finding out this information is costly, and then classical BO is performed.\nAs the group testing phase does not care about high function values, the design evaluated in this process can be associated with low function values. When this happens, the budget has been spent on a design that does not yield a high function value, and variable relevance is learned on a part of the space we do not really care about, as it does not yield a high function value.\nAny insights as to why *GTBO* seems to work despite that? Perhaps the initial starting points provided by group testing provide an accurate approximation of the function. A more relevant one than that usually obtained by uniform/Sobol sampling?\n\nSmall typo:\n\nI would write the r.h.s. of Eq. 4 as $H(Z_t) - H(Z_t|\\boldsymbol{\\xi}) = H(Z_t) - \\sum_{\\xi \\in \\{0,1\\}^D} p(\\xi)H(Z_t|\\boldsymbol{\\xi}=\\xi)$ instead of $H(Z_t) - H(Z_t|\\boldsymbol{\\xi}) = H(Z_t) - \\sum_{\\boldsymbol{\\xi} \\in \\{0,1\\}^D} p(\\xi)H(Z_t|\\boldsymbol{\\xi})$ (as is done in [1]).\n\n[1] Noisy Adaptive Group Testing using Bayesian Sequential Experimental Design, arXiv.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This submission tackles the case of high-dimensional optimization using Bayesian Optimization (BO), a sample-efficient method that has shown great success for hyperparameter tuning of large deep learning models, or in more concrete applications such as recommender systems tuning. It is therefore relevant for a venue such as ICLR, in my opinion. More particularly, the authors propose to tailor the so-called *group testing* theory to the BO setting. This is achieved by 1) extending group testing to real-valued functions and 2) dividing a usual BO run into two steps, a first step of relevant variable identification through group testing, and a second step of conventional BO, with the learned variable relevance being encoded in the Gaussian process surrogate lengthscales. The proposed method, *GTBO*, not only achieves state-of-the-art results of synthetic and real-world experiments but also enlights the practitioner with a ranking of the relevant variables, thus enhancing its understanding of the problem.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is well-written and organized in an easy-to-follow manner.\n- The approach is simple and works remarkably well.\n- The benchmarks include many competing methods, although some recent ones could have been considered as well, e.g. [1,2]\n\n\n[1] Sparse Bayesian optimization. AISTATS 2023.\n[2] Are Random Decompositions all we need in High Dimensional Bayesian Optimisation? ICML 2023.", "weaknesses": "- As often, the approach involves several hyperparameters, to determine whether a variable is deemed as relevant or not, and then incorporating this information in the BO statistical surrogate using carefully-designed priors. \n\nOther than that, I have to say I cannot really spot any weakness here. I am not very familiar with the group testing framework.", "questions": "I am genuinely surprised by how good *GTBO* is compared to other competitors, given that the proposed approach feels suboptimal. It performs in a sequential manner, where one first does not care about function maximization, only about variable relevance, even though finding out this information is costly, and then classical BO is performed.\nAs the group testing phase does not care about high function values, the design evaluated in this process can be associated with low function values. When this happens, the budget has been spent on a design that does not yield a high function value, and variable relevance is learned on a part of the space we do not really care about, as it does not yield a high function value.\nAny insights as to why *GTBO* seems to work despite that? Perhaps the initial starting points provided by group testing provide an accurate approximation of the function. A more relevant one than that usually obtained by uniform/Sobol sampling?\n\nSmall typo:\n\nI would write the r.h.s. of Eq. 4 as $H(Z_t) - H(Z_t|\\boldsymbol{\\xi}) = H(Z_t) - \\sum_{\\xi \\in \\{0,1\\}^D} p(\\xi)H(Z_t|\\boldsymbol{\\xi}=\\xi)$ instead of $H(Z_t) - H(Z_t|\\boldsymbol{\\xi}) = H(Z_t) - \\sum_{\\boldsymbol{\\xi} \\in \\{0,1\\}^D} p(\\xi)H(Z_t|\\boldsymbol{\\xi})$ (as is done in [1]).\n\n[1] Noisy Adaptive Group Testing using Bayesian Sequential Experimental Design, arXiv.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698489081211}], "openreview_url": "https://openreview.net/forum?id=20KYsQ8Q4Z", "arxiv_id": "2310.03515", "paper_pdf": "papers/20KYsQ8Q4Z.pdf", "paper_pdf_sha256": "e6170406d850d4ddfc14c4bae8100bf1629f1c09624c3689da86367b44fe8751", "paper_pdf_bytes": 2957796, "paper_pdf_source": "openreview", "code_url": "https://github.com/gtboauthors/gtbo", "code_repository": "gtboauthors/gtbo", "code_commit": "fe3dca9d179e46b4d9cc7bc6954b09e1078919e8", "code_archive": "repos/20KYsQ8Q4Z.zip", "code_archive_sha256": "8b7954aefdeb96db9318d460e31ca1ce54c917d58583c965c5a8d6a6de039bd7", "code_archive_bytes": 67575, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 59, "github_languages": {"Python": 117055}, "github_archived": false, "github_pushed_at": "2023-10-04T08:05:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/high-dimensional-bayesian-optimization-with-6"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e1u9PVnwNr", "year": 2023, "status": "rejected", "title": "BiBench: Benchmarking and Analyzing Network Binarization", "authors": ["Haotong Qin", "Mingyuan Zhang", "Yifu Ding", "Aoyu Li", "Zhongang Cai", "Ziwei Liu", "Fisher Yu", "Xianglong Liu"], "authorids": ["~Haotong_Qin1", "~Mingyuan_Zhang1", "~Yifu_Ding2", "s220059@e.ntu.edu.sg", "~Zhongang_Cai1", "~Ziwei_Liu1", "~Fisher_Yu2", "~Xianglong_Liu3"], "authors_source": "OpenReview API", "abstract": "Neural network binarization emerges as one of the most promising compression approaches with extraordinary computation and memory savings by minimizing the bit-width of weight and activation. However, despite being a generic technique, recent works reveal that applying binarization in a wide range of realistic scenarios involving diverse tasks, architectures, and hardware is not trivial. Moreover, common challenges, such as severe degradation in accuracy and limited efficiency gains, suggest that specific attributes of binarization are not thoroughly studied and adequately understood. To close this gap, we present BiBench, a rigorously designed benchmark with in-depth analysis for network binarization. We first carefully scrutinize the requirements of binarization in the actual production setting. We thus define the evaluation tracks and metrics for a fair and systematic investigation. We then perform a comprehensive evaluation with a rich collection of milestone binarization algorithms. Our benchmark results show binarization still faces severe accuracy challenges but diminishing improvements brought by newer state-of-the-art binarization algorithms, even at the expense of efficiency. Moreover, the actual deployment of certain binarization operations reveals a surprisingly large deviation from their theoretical consumption. Finally, we provide suggestions based on our benchmark results and analysis, devoted to establishing a paradigm for accurate and efficient binarization among existing techniques. We hope BiBench paves the way towards more extensive adoption of network binarization and serves as a foundation for future research.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "h2EgHANBHk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper17/Reviewer_LQy6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a benchmark to evaluate Network binarization. Network binarization is a compression method for neural networks that transforms the layers into binary vectors. Given the lack of a comprehensive evaluation and benchmarking methodology, the paper proposes sets of tasks, measures, hardware, and methods to validate new algorithms. Finally, the paper shows how network binarization is not a method that can seemingly apply to any network with no tuning.", "review_text": "Although the paper is interesting and the problem relevant, the presentation is sloppy and most of the decisions and claims not fully motivated. For this reason, I believe the paper is not ready for publication at ICLR. ", "strengths": "Strengths:\n\n(S1) The paper sheds a light on important issues related to network binarization, such as robustness and empirical time improvement.\n\n(S2) The benchmark is extensive and covers a number of architectures, learning tasks, hardware, and robustness evaluation.\n\n(S3) The paper evaluates 8 different network binarization methods, from older to more recent ones.\n\n(S4) The proposed benchmark BiBench is useful for the ML community.\n\nWeaknesses:\n\n(W1) The paper is poorly written: there is a large number of typos, broken sentences and bad use of English. The paper would benefit from thorough proofreading and rewriting. Some sentences make no sense or are hard to understand. For reference, here is an incomplete list of sentences and typos:\n\n- severe accuracy challenges **but** diminishing\n- 1 corruption benchmarks\n- pushing network binarization research to be accurate and efficient\n- Binarization ~~technology~~ compresses\n- to fully exert the generic of binarization technology.\n- We train the 1× number of training epochs\n- which requires specifically studied in binarization research.\n\n(W2) Clarity: The paper should introduce the concepts more smoothly and walk the reader through them. Many concepts are assumed to be known and dropped without reference. The paper is therefore hard to read. Examples of unclear/unexplained concepts or claims are:\n\n- “The most aggressive quantization technology” according to whom?\n- “imaging modality task”: what is that? how is it defined?\n- the 1-bit specialization of quantization” - missing reference\n- “requiring specified local structures” such as?\n- “model is exported in the ONNX” → Is this a common format? Where does it come from?\n- What do the colours in Table 2-3 represent?\n\n(W3) The paper would benefit from more rigour. E.g.,\n\n- Popcount is never defined\n- Examples of $\\alpha$ and $w$ should be provided\n- “The quadratic mean form is uniformly applied in BiBench to unify all metrics.” Why is this a good choice?\n\n(W3) The intuition behind the evaluation metrics in Eq. 2 - 9 is unclear. Why not consider standard deviation as well?\n\n(W4) The related work analysis should be more extensive and explain what are the choices determining the current selection of algorithms. A concurrent work [1] presents other models for binarization in its related work section. How are the algorithms in Table 1 chosen? Why are they representative?\n\n[1] Shang, Y., Xu, D., Zong, Z. and Yan, Y., 2022. Network Binarization via Contrastive Learning. ECCV\n\n(W5) As a benchmark, it should probably compare with other, simpler strategies to make the network more compact. For instance, one such strategy could be dropout or model quantization. There is no need to be exhaustive there, but there should be the possibility for researchers working on Network binarization to assess their methods on more traditional techniques.\n\n(W5) It is not clear to what extent the chosen datasets and tasks are challenging or representative. The paper should elaborate more on why some of the tasks have been chosen. The description in Section 3.1 assumes the reader knows the tasks but does not provide additional information on why they are representative.\n\n(W6) Code: The implementation is not available. A benchmark should provide the code (in this case anonymous) for reproducibility. Moreover, the code has to be clear, well-documented, and easy to run.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a benchmark to evaluate Network binarization. Network binarization is a compression method for neural networks that transforms the layers into binary vectors. Given the lack of a comprehensive evaluation and benchmarking methodology, the paper proposes sets of tasks, measures, hardware, and methods to validate new algorithms. Finally, the paper shows how network binarization is not a method that can seemingly apply to any network with no tuning.", "strength_and_weaknesses": "Strengths:\n\n(S1) The paper sheds a light on important issues related to network binarization, such as robustness and empirical time improvement.\n\n(S2) The benchmark is extensive and covers a number of architectures, learning tasks, hardware, and robustness evaluation.\n\n(S3) The paper evaluates 8 different network binarization methods, from older to more recent ones.\n\n(S4) The proposed benchmark BiBench is useful for the ML community.\n\nWeaknesses:\n\n(W1) The paper is poorly written: there is a large number of typos, broken sentences and bad use of English. The paper would benefit from thorough proofreading and rewriting. Some sentences make no sense or are hard to understand. For reference, here is an incomplete list of sentences and typos:\n\n- severe accuracy challenges **but** diminishing\n- 1 corruption benchmarks\n- pushing network binarization research to be accurate and efficient\n- Binarization ~~technology~~ compresses\n- to fully exert the generic of binarization technology.\n- We train the 1× number of training epochs\n- which requires specifically studied in binarization research.\n\n(W2) Clarity: The paper should introduce the concepts more smoothly and walk the reader through them. Many concepts are assumed to be known and dropped without reference. The paper is therefore hard to read. Examples of unclear/unexplained concepts or claims are:\n\n- “The most aggressive quantization technology” according to whom?\n- “imaging modality task”: what is that? how is it defined?\n- the 1-bit specialization of quantization” - missing reference\n- “requiring specified local structures” such as?\n- “model is exported in the ONNX” → Is this a common format? Where does it come from?\n- What do the colours in Table 2-3 represent?\n\n(W3) The paper would benefit from more rigour. E.g.,\n\n- Popcount is never defined\n- Examples of $\\alpha$ and $w$ should be provided\n- “The quadratic mean form is uniformly applied in BiBench to unify all metrics.” Why is this a good choice?\n\n(W3) The intuition behind the evaluation metrics in Eq. 2 - 9 is unclear. Why not consider standard deviation as well?\n\n(W4) The related work analysis should be more extensive and explain what are the choices determining the current selection of algorithms. A concurrent work [1] presents other models for binarization in its related work section. How are the algorithms in Table 1 chosen? Why are they representative?\n\n[1] Shang, Y., Xu, D., Zong, Z. and Yan, Y., 2022. Network Binarization via Contrastive Learning. ECCV\n\n(W5) As a benchmark, it should probably compare with other, simpler strategies to make the network more compact. For instance, one such strategy could be dropout or model quantization. There is no need to be exhaustive there, but there should be the possibility for researchers working on Network binarization to assess their methods on more traditional techniques.\n\n(W5) It is not clear to what extent the chosen datasets and tasks are challenging or representative. The paper should elaborate more on why some of the tasks have been chosen. The description in Section 3.1 assumes the reader knows the tasks but does not provide additional information on why they are representative.\n\n(W6) Code: The implementation is not available. A benchmark should provide the code (in this case anonymous) for reproducibility. Moreover, the code has to be clear, well-documented, and easy to run.", "clarity,_quality,_novelty_and_reproducibility": "Although the paper proposes a novel benchmark, the clarity of the explanation is not sufficient (see W2). Moreover, a number of typos and bad sentences render the reading even more strenuous. In terms of reproducibility, I could not find the code. For a benchmark, the presence of well-documented and easy-to-use code is paramount.", "summary_of_the_review": "Although the paper is interesting and the problem relevant, the presentation is sloppy and most of the decisions and claims not fully motivated. For this reason, I believe the paper is not ready for publication at ICLR. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667474611193}, {"id": "cx2caK2ajuZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper17/Reviewer_Byew"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a benchmark for binary neural networks. The perspectives of the proposed benchmark include BNN methods, architectures, multiple tasks, and inference tests on various hardware. The authors summarize some insights and offer practical guidance.", "review_text": "Rather than saying that this is a benchmark paper, it is actually more like evaluating some selected BNNs from different perspectives. The author's original intention is constructive. However, the selected work has certain limitations, with a large number of ancient BNNs, some newer architecture designs, and optimization methods that are missing. In addition, it is not very reasonable to directly use the architectures of the fp-network for architecture benchmarking. The inference speed tests are greatly limited by the libs used, and their results provide very limited guidance. It is difficult for me to get a clear takeaway from a lot of experiments and data.", "strengths": "Although I very much agree with the motivation of this work, I also see very interesting settings and conclusions. However, I still have the following concerns:\n\n- Why I think the evaluation of BNN on many downstream vision tasks does not make much sense now: The classification task model is usually used as the Backbone of other downstream tasks. For example, the imagenet pre-trained full precision ResNet backbone is widely used in a large number of vision tasks. Therefore, for a binary model with a serious accuracy degradation problem, it is of practical significance to first achieve sufficient accuracy on ImageNet, and then verify it in more downstream tasks. Therefore, the robustness of BNN does not seem to be an immediate concern until a feasible level of accuracy is reached.\n\n-  There is certain doubt about the significance of the assessment. For example, the authors concluded that \"There is no obvious difference in the theoretical complexity among binarization algorithms. \". I think this is mainly because similar architectural bases have been selected in the comparison. Many earlier methods such as XNOR, DoReFa, BiReal, xnor++ etc., are based on ResNet18 Backbone. More recent works selected in this paper as ReCU and FDA are based on the ReActNet network structure. So it is not surprising that there is no difference in theoretical complexity and inference time. Furthermore, the acc improvements of ReCU and FDA are about within 1% compared to ReActNet. Therefore, using the same Backbone with only marginal acc improvement brings little meaning to their benchmarking.\nMore BNN works should be included, especially those having different architectural designs, e.g., [1,2,3,4], etc.\n\n- At present, the optimization of BNN has the problem of objective mismatch, and we usually need to use latent weight for training. Therefore, from the shape of the model parameters (only two states +1 and -1), the parameter update method (sign flipping), and the information flow path mode are significantly different from the fp model, why the author only considers the use of the full-precision model for the Architecture benchmarking? Why do the authors think that full-precision architectures are naturally suitable for BNN? In this work [3], the author verified that many full-precision model designs, such as bottleneck convolution, ConvBlock without shortcut connections (e.g., VGG, inceptionNet) are actually completely unsuitable for BNN.\n\n- The results of the robustness analysis are nice. However, they are not new as well. [5,6] show that discrete BNNs exhibit superior stability and robustness against adversarial attacks.\n\n- An important work [7] is missing in the training consumption track. [7] considers sign flipping without using the latent weight in the backward pass, which could significantly reduce the training epochs. The most compared optimizers are from the same type relying on the fp latent weights. Most of those proposed gradient approximation methods only got marginal contributions in their improvements. The most acc gains were from the architecture design.\n\n- The inference speed test was originally the part I was most expected in this paper. However, as the authors said, extremely limited inference libraries make the contribution of this part minimal, and basically no new insights. Since most of the inference libs used by the authors have ceased development and maintenance, most of them do not have any performance optimizations for recent BNN models, so the evaluation results provided may not be the best performance. For example, ReActNet-A is based on the MobileNet-V1 Backbone. Neither larq nor DaBNN provide an implementation based on this Backbone and thus no corresponding instruction-set optimization and assembly optimization for the model on Arm. Without specific optimization, the implementation of a new model architecture usually does not have a significant efficiency gain.\n\n\n[1] Zhang, Yichi, Zhiru Zhang, and Lukasz Lew. \"PokeBNN: A Binary Pursuit of Lightweight Accuracy.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[2] Bethge, Joseph, et al. \"Meliusnet: Can binary neural networks achieve mobilenet-level accuracy?.\" arXiv preprint arXiv:2001.05936 (2020).\n\n[3] Bethge, J., Yang, H., Bornstein, M., & Meinel, C. (2019). Back to simplicity: How to train accurate bnns from scratch?. arXiv preprint arXiv:1906.08637.\n\n[4] Martinez, B., Yang, J., Bulat, A., & Tzimiropoulos, G. (2020). Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535.\n\n[5] A. Galloway et al. Attacking binarized neural networks. arXiv:1711.00449, 2017. \n\n[6] E. Khalil et al. Combinatorial attacks on binarized neural networks. arXiv:1810.03538, 2018.\n\n[7] Helwegen, Koen, et al. \"Latent weights do not exist: Rethinking binarized neural network optimization.\" Advances in neural information processing systems 32 (2019).", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents a benchmark for binary neural networks. The perspectives of the proposed benchmark include BNN methods, architectures, multiple tasks, and inference tests on various hardware. The authors summarize some insights and offer practical guidance.", "strength_and_weaknesses": "Although I very much agree with the motivation of this work, I also see very interesting settings and conclusions. However, I still have the following concerns:\n\n- Why I think the evaluation of BNN on many downstream vision tasks does not make much sense now: The classification task model is usually used as the Backbone of other downstream tasks. For example, the imagenet pre-trained full precision ResNet backbone is widely used in a large number of vision tasks. Therefore, for a binary model with a serious accuracy degradation problem, it is of practical significance to first achieve sufficient accuracy on ImageNet, and then verify it in more downstream tasks. Therefore, the robustness of BNN does not seem to be an immediate concern until a feasible level of accuracy is reached.\n\n-  There is certain doubt about the significance of the assessment. For example, the authors concluded that \"There is no obvious difference in the theoretical complexity among binarization algorithms. \". I think this is mainly because similar architectural bases have been selected in the comparison. Many earlier methods such as XNOR, DoReFa, BiReal, xnor++ etc., are based on ResNet18 Backbone. More recent works selected in this paper as ReCU and FDA are based on the ReActNet network structure. So it is not surprising that there is no difference in theoretical complexity and inference time. Furthermore, the acc improvements of ReCU and FDA are about within 1% compared to ReActNet. Therefore, using the same Backbone with only marginal acc improvement brings little meaning to their benchmarking.\nMore BNN works should be included, especially those having different architectural designs, e.g., [1,2,3,4], etc.\n\n- At present, the optimization of BNN has the problem of objective mismatch, and we usually need to use latent weight for training. Therefore, from the shape of the model parameters (only two states +1 and -1), the parameter update method (sign flipping), and the information flow path mode are significantly different from the fp model, why the author only considers the use of the full-precision model for the Architecture benchmarking? Why do the authors think that full-precision architectures are naturally suitable for BNN? In this work [3], the author verified that many full-precision model designs, such as bottleneck convolution, ConvBlock without shortcut connections (e.g., VGG, inceptionNet) are actually completely unsuitable for BNN.\n\n- The results of the robustness analysis are nice. However, they are not new as well. [5,6] show that discrete BNNs exhibit superior stability and robustness against adversarial attacks.\n\n- An important work [7] is missing in the training consumption track. [7] considers sign flipping without using the latent weight in the backward pass, which could significantly reduce the training epochs. The most compared optimizers are from the same type relying on the fp latent weights. Most of those proposed gradient approximation methods only got marginal contributions in their improvements. The most acc gains were from the architecture design.\n\n- The inference speed test was originally the part I was most expected in this paper. However, as the authors said, extremely limited inference libraries make the contribution of this part minimal, and basically no new insights. Since most of the inference libs used by the authors have ceased development and maintenance, most of them do not have any performance optimizations for recent BNN models, so the evaluation results provided may not be the best performance. For example, ReActNet-A is based on the MobileNet-V1 Backbone. Neither larq nor DaBNN provide an implementation based on this Backbone and thus no corresponding instruction-set optimization and assembly optimization for the model on Arm. Without specific optimization, the implementation of a new model architecture usually does not have a significant efficiency gain.\n\n\n[1] Zhang, Yichi, Zhiru Zhang, and Lukasz Lew. \"PokeBNN: A Binary Pursuit of Lightweight Accuracy.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[2] Bethge, Joseph, et al. \"Meliusnet: Can binary neural networks achieve mobilenet-level accuracy?.\" arXiv preprint arXiv:2001.05936 (2020).\n\n[3] Bethge, J., Yang, H., Bornstein, M., & Meinel, C. (2019). Back to simplicity: How to train accurate bnns from scratch?. arXiv preprint arXiv:1906.08637.\n\n[4] Martinez, B., Yang, J., Bulat, A., & Tzimiropoulos, G. (2020). Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535.\n\n[5] A. Galloway et al. Attacking binarized neural networks. arXiv:1711.00449, 2017. \n\n[6] E. Khalil et al. Combinatorial attacks on binarized neural networks. arXiv:1810.03538, 2018.\n\n[7] Helwegen, Koen, et al. \"Latent weights do not exist: Rethinking binarized neural network optimization.\" Advances in neural information processing systems 32 (2019).", "clarity,_quality,_novelty_and_reproducibility": "The paper is easy to follow. Novelty is very limited. Reproducibility seems to be not a problem.", "summary_of_the_review": "Rather than saying that this is a benchmark paper, it is actually more like evaluating some selected BNNs from different perspectives. The author's original intention is constructive. However, the selected work has certain limitations, with a large number of ancient BNNs, some newer architecture designs, and optimization methods that are missing. In addition, it is not very reasonable to directly use the architectures of the fp-network for architecture benchmarking. The inference speed tests are greatly limited by the libs used, and their results provide very limited guidance. It is difficult for me to get a clear takeaway from a lot of experiments and data.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666623518507}, {"id": "aa15wdnXrl5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper17/Reviewer_kQWB"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides one benchmark named BiBench which is desifned to test the efficiency, robustness and generalization of binary algorithms. It mainly contains two parts, one for accuracy and another is for efficiency, and each part has several components. The authors build quantitative indicators to measure the binary algorithms, which is very significative. The BiBench is the first benchmark in the area of binary models.", "review_text": "The BiBench explores the advantages and disadvantages of different binary algorithms, and it shed a light to furture methods.", "strengths": "Strengths:    \n1. This paper provides the first benchmark in the area of binary algorithms. It consides the accuracy and efficiency of binary models. Specially, the authors explore the track of hardware inference, which is very meaningful.    \n2. With very much exiperiments, the paper analyzes the pros and cons of existing methods.    \n\nWeakness:   \n1. As a benchmark, more principles about binary algorithms are expected, such as its formula mode in mathmatics and softwares, its compatibility with current hardware.   \n2. Some abbreviation of letter are not explained. In the XNOR-Net of Appendix, sub-tensors in 'I', then what if the meaning of 'I'?    \n3. Because there are many experiments, many details are unclear in the paper. The open source code of this benchmark is very helpful to this area.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper provides one benchmark named BiBench which is desifned to test the efficiency, robustness and generalization of binary algorithms. It mainly contains two parts, one for accuracy and another is for efficiency, and each part has several components. The authors build quantitative indicators to measure the binary algorithms, which is very significative. The BiBench is the first benchmark in the area of binary models.", "strength_and_weaknesses": "Strengths:    \n1. This paper provides the first benchmark in the area of binary algorithms. It consides the accuracy and efficiency of binary models. Specially, the authors explore the track of hardware inference, which is very meaningful.    \n2. With very much exiperiments, the paper analyzes the pros and cons of existing methods.    \n\nWeakness:   \n1. As a benchmark, more principles about binary algorithms are expected, such as its formula mode in mathmatics and softwares, its compatibility with current hardware.   \n2. Some abbreviation of letter are not explained. In the XNOR-Net of Appendix, sub-tensors in 'I', then what if the meaning of 'I'?    \n3. Because there are many experiments, many details are unclear in the paper. The open source code of this benchmark is very helpful to this area.", "clarity,_quality,_novelty_and_reproducibility": "The writing of this paper is very good, its easy to understand. As a benchmrah, its novelty is less important, its main contribution is that it can help newcoming methods to be more value. Its reproducibility needs the open code.", "summary_of_the_review": "The BiBench explores the advantages and disadvantages of different binary algorithms, and it shed a light to furture methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666352747676}], "openreview_url": "https://openreview.net/forum?id=e1u9PVnwNr", "arxiv_id": "2301.11233", "paper_pdf": "papers/e1u9PVnwNr.pdf", "paper_pdf_sha256": "ebb42417997fc95d53977274f363d91e041a5891188722fc30a0542360aa02f4", "paper_pdf_bytes": 701606, "paper_pdf_source": "openreview", "code_url": "https://github.com/AI-Efficiency/BiBench", "code_repository": "AI-Efficiency/BiBench", "code_commit": "a6662179aff8051a3605e771fc076b35080b5be7", "code_archive": "repos/e1u9PVnwNr.zip", "code_archive_sha256": "0b60acd60517ffecdb7ddbf1ab0177973c1fe4062c93c0a88405962a0146cbbb", "code_archive_bytes": 245865, "code_file_count": 369, "code_extensions": {".py": 367, ".sh": 2}, "github_disk_usage_kb": 113, "github_languages": {"Python": 400727, "Shell": 1265}, "github_archived": false, "github_pushed_at": "2024-03-04T15:37:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bibench-benchmarking-and-analyzing-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rryJiPXifr", "year": 2021, "status": "rejected", "title": "Optimization Planning for 3D ConvNets", "authors": ["Zhaofan Qiu", "Ting Yao", "Chong-wah Ngo", "Tao Mei"], "authorids": ["~Zhaofan_Qiu2", "~Ting_Yao1", "~Chong-wah_Ngo2", "~Tao_Mei3"], "authors_source": "OpenReview API", "abstract": "3D Convolutional Neural Networks (3D ConvNets) have been regarded as a powerful class of models for video recognition. Nevertheless, it is not trivial to optimally learn a 3D ConvNets due to high complexity and various options of the training scheme. The most common hand-tuning process starts from learning 3D ConvNets using short video clips and then is followed by learning long-term temporal dependency using lengthy clips, while gradually decaying the learning rate from high to low as training progresses. The fact that such process comes along with several heuristic settings motivates the study to seek an optimal ``path'' to automate the entire training. In this paper, we decompose the path into a series of training ``states'' and specify the hyper-parameters, e.g., learning rate and the length of input clips, in each state. The estimation of the knee point on the performance-epoch curve triggers the transition from one state to another. We perform dynamic programming over all the candidate states to plan the optimal permutation of states, i.e., optimization path. Furthermore, we devise a new 3D ConvNets with a unique design of dual-head classifier to improve the spatial and temporal discrimination. Extensive experiments conducted on seven public video recognition benchmarks demonstrate the advantages of our proposal. With the optimization planning, our 3D ConvNets achieves superior results when comparing to the state-of-the-art video recognition approaches. More remarkably, we obtain the top-1 accuracy of 82.5% and 84.3% on the large-scale Kinetics-400 and Kinetics-600 datasets, respectively.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "zPVNza_QRK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1358/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- The overall quality is good. It addresses the training issue of 3D convolutional network in a novel perspective and shows promising results. The presentation is clear except for some grammar mistakes. \n\nDetailed Comments:\n\n- In Sec. 3.2, how is the transition performance evaluated is not clear. What is the cost / gain from i to j? Is it the same for all states?\n\n- Ablation experiments should be provided to show the performance difference using different paths. Is the adopted path actually the optimal path? And how about if some intermediate states are skipped?\n\n- The proposed optimization planning doesn't seem to work specifically for the 3D convnets and for the action recognition task. As long as we need to train a network in stages with multiple hyper-parameters, this should apply theoretically. \n\n- Abstract: 'follows' --> 'is followed'", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposes an optimization planning strategy to systematically train a 3D network. It shows promising results on multiple datasets for the tasks of action recognition. ", "review": "- The overall quality is good. It addresses the training issue of 3D convolutional network in a novel perspective and shows promising results. The presentation is clear except for some grammar mistakes. \n\nDetailed Comments:\n\n- In Sec. 3.2, how is the transition performance evaluated is not clear. What is the cost / gain from i to j? Is it the same for all states?\n\n- Ablation experiments should be provided to show the performance difference using different paths. Is the adopted path actually the optimal path? And how about if some intermediate states are skipped?\n\n- The proposed optimization planning doesn't seem to work specifically for the 3D convnets and for the action recognition task. As long as we need to train a network in stages with multiple hyper-parameters, this should apply theoretically. \n\n- Abstract: 'follows' --> 'is followed'", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604052555247}, {"id": "hZuFQf1rEaL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1358/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overview:\nThe paper proposes a novel way to automatically tune 3D ConvNet hyper-parameters (learning rate, input clip length, sampling way). This is achieved by decomposing the optimization path into several states and the state transition is triggered when the knee-point on the performance-epoch curve is met. Extensive experiments are conducted on popular video benchmarks and show that the optimization planning is effective to improve the accuracy and requires less time time compared to the hand-tuned procedure.\n\n\nStrengths:\n\n++ The paper provides a novel perspective of designing the hyper-parameters for 3D ConvNets. The automatic planning through the optimization path alleviate the tedious work of hand tuning and save a lot of times: Once the knee-point on the performance-epoch curve is achieved, we can easily transit to the next optimization state (with another set of hyper-parameters).\n\n\n++ The experiments are thorough. The authors conduct experiments on all popular video recognition benchmarks. Also the experiment results corroborate the effectiveness of optimization planning. \n\nWeaknesses:\n\n-- The construction of transition graph has some strict principles that could be relaxed. \n  (1) For scheduling learning rate, It only considers that \"training starts from high learning rate\". However, previous works (e.g. ICLR2017: https://arxiv.org/abs/1608.03983) have shown that cyclic cosine learning schedule might benefit training. Therefore, I don't see the necessity that we put this restriction.\n  (2) For scheduling input length, this is not the first time that people do this. For example, Wang et. al. (CVPR2018, https://arxiv.org/abs/1711.07971) first train using 64 frames as input and then finetune on 128 frames. Also, Wu et. al. (CVPR2020, https://arxiv.org/abs/1912.00998) proposes another way of adjust the sampling strategy.\n\n\n-- The comparison is mostly conducted on the newly proposed DG-P3D family architectures. It would be better if there is an additional row of results on the standard I3D/SlowFast network in Table 3. (this is somewhat minor and not required considering the time limit)\n\n\n-- \"82.5%\" (with flow stream) is not the \"new record\" for Kinetics-400. ir-CSN-152 (ICCV2019, https://arxiv.org/abs/1904.02811) can achieve 82.6% on Kinetics-400 using **RGB only**. OmniSource (ECCV2020, https://arxiv.org/abs/2003.13042) improves it to 83.6% using additional data.\n\n\n\n\nI have some additional questions:\n\n-- In Figure 4, it shows that SS-V1/2 would prefer uniform sampling however Kinetics would prefer consecutive sampling. However, this seems somewhat counter-intuitive to me. We know that Kinetics have videos with static scenes or slow motion while SS-V1/2 have some videos with more intensive motion. Does it make more sense to use dense frames (consecutive) to recognize intensive motion?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Reviewer #2", "review": "Overview:\nThe paper proposes a novel way to automatically tune 3D ConvNet hyper-parameters (learning rate, input clip length, sampling way). This is achieved by decomposing the optimization path into several states and the state transition is triggered when the knee-point on the performance-epoch curve is met. Extensive experiments are conducted on popular video benchmarks and show that the optimization planning is effective to improve the accuracy and requires less time time compared to the hand-tuned procedure.\n\n\nStrengths:\n\n++ The paper provides a novel perspective of designing the hyper-parameters for 3D ConvNets. The automatic planning through the optimization path alleviate the tedious work of hand tuning and save a lot of times: Once the knee-point on the performance-epoch curve is achieved, we can easily transit to the next optimization state (with another set of hyper-parameters).\n\n\n++ The experiments are thorough. The authors conduct experiments on all popular video recognition benchmarks. Also the experiment results corroborate the effectiveness of optimization planning. \n\nWeaknesses:\n\n-- The construction of transition graph has some strict principles that could be relaxed. \n  (1) For scheduling learning rate, It only considers that \"training starts from high learning rate\". However, previous works (e.g. ICLR2017: https://arxiv.org/abs/1608.03983) have shown that cyclic cosine learning schedule might benefit training. Therefore, I don't see the necessity that we put this restriction.\n  (2) For scheduling input length, this is not the first time that people do this. For example, Wang et. al. (CVPR2018, https://arxiv.org/abs/1711.07971) first train using 64 frames as input and then finetune on 128 frames. Also, Wu et. al. (CVPR2020, https://arxiv.org/abs/1912.00998) proposes another way of adjust the sampling strategy.\n\n\n-- The comparison is mostly conducted on the newly proposed DG-P3D family architectures. It would be better if there is an additional row of results on the standard I3D/SlowFast network in Table 3. (this is somewhat minor and not required considering the time limit)\n\n\n-- \"82.5%\" (with flow stream) is not the \"new record\" for Kinetics-400. ir-CSN-152 (ICCV2019, https://arxiv.org/abs/1904.02811) can achieve 82.6% on Kinetics-400 using **RGB only**. OmniSource (ECCV2020, https://arxiv.org/abs/2003.13042) improves it to 83.6% using additional data.\n\n\n\n\nI have some additional questions:\n\n-- In Figure 4, it shows that SS-V1/2 would prefer uniform sampling however Kinetics would prefer consecutive sampling. However, this seems somewhat counter-intuitive to me. We know that Kinetics have videos with static scenes or slow motion while SS-V1/2 have some videos with more intensive motion. Does it make more sense to use dense frames (consecutive) to recognize intensive motion?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603930629980}, {"id": "_BiKiPSl17q", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1358/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- The paper proposed a Dual-head Global-contextual Pseudo-3D (DG-P3D) network and an automated optimization path to train 3D ConvNet for action recognition. The proposed method is evaluated on various of action recognition datasets, and achieved convincing results.\n\n- Another contribution of the paper is its decreased cost. It mentioned that the time cost for grid optimization planning with 8 NVidia Titan V GPUs reduced the running time from 4057h to 288h on Kinetics.\n\n- The optimization method is proposed for 3D ConvNet, However, the authors only benchmarked it with the new proposed DG-P3D network. It could be interesting to know what is the performance applying it to other basic 3D ConvNets. It is also interesting to know what is the performance of DG-P3D without this new optimization strategy.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper about automate the optimization process of 3D ConvNets", "review": "- The paper proposed a Dual-head Global-contextual Pseudo-3D (DG-P3D) network and an automated optimization path to train 3D ConvNet for action recognition. The proposed method is evaluated on various of action recognition datasets, and achieved convincing results.\n\n- Another contribution of the paper is its decreased cost. It mentioned that the time cost for grid optimization planning with 8 NVidia Titan V GPUs reduced the running time from 4057h to 288h on Kinetics.\n\n- The optimization method is proposed for 3D ConvNet, However, the authors only benchmarked it with the new proposed DG-P3D network. It could be interesting to know what is the performance applying it to other basic 3D ConvNets. It is also interesting to know what is the performance of DG-P3D without this new optimization strategy.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess"}, "tcdate": 1603929949368}, {"id": "YsZuEHtGnK7", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1358/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n  This paper proposed two things:  \n    1. hyper-parameter planning for training action recognition models.   \n    2. a new 3D net architecture for action recognition. \n\n  The results show with the planning, authors can reduce training time significantly. and the proposed DG-P3D also is good for action recognition\n\n\nStrong points:\n\n  1. This paper explores a new approach for hyperparameter tuning to get a better model quickly.\n\n  2. This paper performs extensive experiments to validate their proposed methods.\n\nWeak points:\n\n  1. It misses one very important reference on the same topic, \"Multigrid Method for Efficiently Training Video Models\" CVPR 2020. This paper also proposes a new way to efficiently train action recognition models. \n\n  2. The way authors compared to others are mixed, making that it is a little bit difficult to compare the contribution of the paper. \n  \n    2.1 To be more fair compare to SOTA approaches in Table 4 and Table 5 with respect to the contributions of DG-P3D, the author should include the results without planning approach and use a similar training setting to the other SOTA approaches. Or apply the proposed optimization planning on those SOTA approaches to see the improvement. In this case, the advantages of each component will be more clear. \n\n    2.2 When comparing to other approaches, authors should either include the number of frames used or computational load (like FLOPs) into the consideration, instead of performance only. E.g., for Kinetics400, slowfast-ResNet-50 (77.0% top-1 acc) only use 32 frames while the proposed models for Kinetics400 use 128 frames (If I understand correctly); thus, simply comparing the accuracy is unfair. On the other hand, If authors want to compare the performance only, authors might want to include the model with different backbones, e.g. bLVNet-TAM has models achieved 53.1% on SS-V1 and 65.2% on SS-V2 with RGB only.\n\n\nQuestions:\n\n  1. If a node has multiple destinations, e.g. node S5 in Figure 1a, how do the author select S7 or S8? In my understanding, the proposed algorithm will train the model at S5 with the parameters at S7 and S8 independently and then select the better one. Is it correct?\n\n  2. Followed by 1, when encountering multiple destinations, if all destination nodes need to be explored before making the transition, do the author takes those time into account as well in Table 4? I do not find those in Figure 4, either.\n\n  3. What is the delay parameter T in practical? Is it a dataset-dependent parameter?\n\n  4. What is the inference protocol used in the evaluation? The authors mentioned there are two common approaches but do not mention which one is used.\n\nOthers:\n\n 1. When comparing to SOTA, I think authors miss \"X3D: Expanding Architectures for Efficient Video Recognition\" CVPR 2020.\n\nAfter rebuttal:\n\nThe authors addressed some of my concerns, and the proposed optimization plan is interested.\n\n1. I still think that the proposed architecture does not outperform SOTA architecture, like SlowFast. \n  - First, in the rebuttal, the authors mention that with hand-crafted strategies, the proposed model is 0.4% better than SlowFast; nonetheless, the hand-crafted strategies train longer epochs than SlowFast. (SlowFast is trained with 196 epochs.) As authors propose new strategies, they know better how different strategies could affect. \n  - Second, in the rebuttal, the authors improve I3D by 1.7% with optimization planning, which means, SlowFast might outperform the proposed model if involving optimization planning. Moreover, in the revised Table 7, the SlowFast does not have more FLOPs than the proposed network under the same backbone. Simply checking Table 8 might be confused. \n\n2. About the optimization planning, it is still confused about how many epochs spent on the explored nodes. E.g. in Figure 4, for the Kinetics dataset, I do not understand why there is a black edge between S4 and S5 as S4 is not reached. (Authors noted that the black edges mean the explored strategies.) And why not there is no number on those explored edges? Another example, if we look at Table 3 and Figure 4 together, and again for the Kinetics dataset, the summation of those red numbers is 248, but this number is larger than any number in Table 3. As the authors mention that they included the epochs of the explored strategies in Table 3, that means they only spend 29 epochs in the exploration (if we treat Figure 4 for DG-P3D in Table 3.). For me, it seems too good to be true.\n\nThus, I keep my rating. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Reviews", "review": "Summary:\n  This paper proposed two things:  \n    1. hyper-parameter planning for training action recognition models.   \n    2. a new 3D net architecture for action recognition. \n\n  The results show with the planning, authors can reduce training time significantly. and the proposed DG-P3D also is good for action recognition\n\n\nStrong points:\n\n  1. This paper explores a new approach for hyperparameter tuning to get a better model quickly.\n\n  2. This paper performs extensive experiments to validate their proposed methods.\n\nWeak points:\n\n  1. It misses one very important reference on the same topic, \"Multigrid Method for Efficiently Training Video Models\" CVPR 2020. This paper also proposes a new way to efficiently train action recognition models. \n\n  2. The way authors compared to others are mixed, making that it is a little bit difficult to compare the contribution of the paper. \n  \n    2.1 To be more fair compare to SOTA approaches in Table 4 and Table 5 with respect to the contributions of DG-P3D, the author should include the results without planning approach and use a similar training setting to the other SOTA approaches. Or apply the proposed optimization planning on those SOTA approaches to see the improvement. In this case, the advantages of each component will be more clear. \n\n    2.2 When comparing to other approaches, authors should either include the number of frames used or computational load (like FLOPs) into the consideration, instead of performance only. E.g., for Kinetics400, slowfast-ResNet-50 (77.0% top-1 acc) only use 32 frames while the proposed models for Kinetics400 use 128 frames (If I understand correctly); thus, simply comparing the accuracy is unfair. On the other hand, If authors want to compare the performance only, authors might want to include the model with different backbones, e.g. bLVNet-TAM has models achieved 53.1% on SS-V1 and 65.2% on SS-V2 with RGB only.\n\n\nQuestions:\n\n  1. If a node has multiple destinations, e.g. node S5 in Figure 1a, how do the author select S7 or S8? In my understanding, the proposed algorithm will train the model at S5 with the parameters at S7 and S8 independently and then select the better one. Is it correct?\n\n  2. Followed by 1, when encountering multiple destinations, if all destination nodes need to be explored before making the transition, do the author takes those time into account as well in Table 4? I do not find those in Figure 4, either.\n\n  3. What is the delay parameter T in practical? Is it a dataset-dependent parameter?\n\n  4. What is the inference protocol used in the evaluation? The authors mentioned there are two common approaches but do not mention which one is used.\n\nOthers:\n\n 1. When comparing to SOTA, I think authors miss \"X3D: Expanding Architectures for Efficient Video Recognition\" CVPR 2020.\n\nAfter rebuttal:\n\nThe authors addressed some of my concerns, and the proposed optimization plan is interested.\n\n1. I still think that the proposed architecture does not outperform SOTA architecture, like SlowFast. \n  - First, in the rebuttal, the authors mention that with hand-crafted strategies, the proposed model is 0.4% better than SlowFast; nonetheless, the hand-crafted strategies train longer epochs than SlowFast. (SlowFast is trained with 196 epochs.) As authors propose new strategies, they know better how different strategies could affect. \n  - Second, in the rebuttal, the authors improve I3D by 1.7% with optimization planning, which means, SlowFast might outperform the proposed model if involving optimization planning. Moreover, in the revised Table 7, the SlowFast does not have more FLOPs than the proposed network under the same backbone. Simply checking Table 8 might be confused. \n\n2. About the optimization planning, it is still confused about how many epochs spent on the explored nodes. E.g. in Figure 4, for the Kinetics dataset, I do not understand why there is a black edge between S4 and S5 as S4 is not reached. (Authors noted that the black edges mean the explored strategies.) And why not there is no number on those explored edges? Another example, if we look at Table 3 and Figure 4 together, and again for the Kinetics dataset, the summation of those red numbers is 248, but this number is larger than any number in Table 3. As the authors mention that they included the epochs of the explored strategies in Table 3, that means they only spend 29 epochs in the exploration (if we treat Figure 4 for DG-P3D in Table 3.). For me, it seems too good to be true.\n\nThus, I keep my rating. \n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603859908137}], "openreview_url": "https://openreview.net/forum?id=rryJiPXifr", "arxiv_id": "2201.04021", "paper_pdf": "papers/rryJiPXifr.pdf", "paper_pdf_sha256": "ff22c8572b0beef4bd3caf469e03d23c9cd3e7344e040c4b6936613887438600", "paper_pdf_bytes": 6527274, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhaofanQiu/Optimization-Planning-for-3D-ConvNets", "code_repository": "ZhaofanQiu/Optimization-Planning-for-3D-ConvNets", "code_commit": "d9f1b777811ca0d8f462798ca2efcea39b96fcc5", "code_archive": "repos/rryJiPXifr.zip", "code_archive_sha256": "ccf1dd4f51f10d8c6e75241cd2aef37df1093ffe020b08d1a0c3042441a6e029", "code_archive_bytes": 218724, "code_file_count": 46, "code_extensions": {".py": 45, ".sh": 1}, "github_disk_usage_kb": 156, "github_languages": {"Python": 467759, "Shell": 585}, "github_archived": false, "github_pushed_at": "2022-01-12T02:05:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimization-planning-for-3d-convnets-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJlbo6VtDH", "year": 2020, "status": "rejected", "title": "A Generalized Framework of Sequence Generation with Application to Undirected Sequence Models", "authors": ["Elman Mansimov", "Alex Wang", "Kyunghyun Cho"], "authorids": ["elman.mansimov@gmail.com", "wangalexc@gmail.com", "kyunghyun.cho@nyu.edu"], "authors_source": "OpenReview API", "abstract": "Undirected neural sequence models such as BERT (Devlin et al., 2019) have received renewed interest due to their success on discriminative natural language understanding tasks such as question-answering and natural language inference. \nThe problem of generating sequences directly from these models has received relatively little attention, in part because generating from such models departs significantly from the conventional approach of monotonic generation in directed sequence models. We investigate this problem by first proposing a generalized model of sequence generation that unifies decoding in directed and undirected models. The proposed framework models the process of generation rather than a resulting sequence, and under this framework, we derive various neural sequence models as special cases, such as autoregressive, semi-autoregressive, and refinement-based non-autoregressive models. This unification enables us to adapt decoding algorithms originally developed for directed sequence models to undirected models. We demonstrate this by evaluating various decoding strategies for a cross-lingual masked translation model (Lample and Conneau, 2019). Our experiments show that generation from undirected sequence models, under our framework, is competitive with the state of the art on WMT'14 English-German translation. We also demonstrate that the proposed approach enables constant-time translation with similar performance to linear-time translation from the same model by rescoring hypotheses with an autoregressive model.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ryeBboGP9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper734/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a generalized framework for sequence generation that can be applied to both directed and undirected sequence models. The framework generates the final label sequence through generating a sequence of steps, where each step generates a coordinate sequence and an intermediate label sequence. This procedure is probabilistically modeled by length prediction, coordinate selection and symbol replacement. For inference, instead of the intractable naive approach based on Gibbs sampling to marginalize out all generation paths, the paper proposes a heuristic approach using length-conditioned beam search to generate the most likely final sequence. With the proposed framework, the paper shows that masked language models like BERT, even though they are undirected sequence models, can be used for sequence generation, which obtains close performance to the traditional left-to-right autoregressive models on the task of machine translation.\n\nOverall the paper has significant contributions in the following aspects:\n1. It enables undirected sequence models, like BERT, to perform decoding or sequence generation directly, instead of just serving as model pre-training.\n2. The proposed framework unifies directed and undirected sequence models decoding, and it can represent a few existing sequence model decoding as special cases.\n3. The coordinate sequence selection function in the framework can be dependent on the intermediate label sequence. A few simple selection approaches proposed in the paper are shown to be effective. It could be further extended. \n4. The analysis of the coordinate selection order is interesting and helpful for understanding the algorithm.\n5. The experiment results for decoding masked language models on machine translation are promising. It also provides the comparison to recent related work on non-autoregressive approaches.\n\nThe presentation of the paper is also clear. I am leaning towards accepting the paper.\n\nHowever, there are some weaknesses:\n\n1. It should be analyzed more why different coordinate selection approaches perform differently in linear-time decoding vs. constant-time decoding. Even in constant-time decoding, the conclusion varies in different decoding setting, easy-first is the worst for the L->1 case, but the best for the L/T case, why is that?\n\n2. What is the motivation for \"hard-first\"?\n\n3. The setting of \"least2most\" with L->1 is similar to Ghazvininejad et al. 2019. But Table 4 in the appendix shows the result in this paper is still worse (21.98 vs. 24.61, when both systems use 10 iterations without AR). Also, the gap from the AR baseline is larger than that in Ghazvininejad et al. 2019. Given the two systems are considered similar, it should be explained in the paper the possible reasons for these discrepancies in results.\n\nAdditional minor comments for improving the paper:\n\n1. In the introduction, it mentions the baseline AR is (Vaswani et al. 2017), while in the experimental settings, it mentions the baseline AR is (Bahdanau et al. 2015). Please clarify which one is used.\n\n2. In Table 1, how does T = 2L work for the \"Uniform\" case while the target sequence length is only T, since it is mentioned the positions are sampled without replacement. Similarly, how does T = 2L work for the \"Left2Right\" case? Is it just always choosing the last position when L < t <= 2L? In these two cases, it seems T > L is not needed.\n\n3. In Table 3, the header for the 2nd column should be o_t, as defined in Section 4 - \"Decoding scenarios\". What is the actual value of K and K'' for the constant-time machine translation experiments in the paper?\n\n4. \"Rescoring adds minimal overhead as it is run in parallel\" - it still needs to run left-to-right in sequence since it is auto-regressive. Please clarify what it means by \"in parallel\" here.\n\n5. What is the range and average for the target sentence length? How is T = 20 for constant-time decoding compared to linear-time decoding in terms of speed?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a generalized framework for sequence generation that can be applied to both directed and undirected sequence models. The framework generates the final label sequence through generating a sequence of steps, where each step generates a coordinate sequence and an intermediate label sequence. This procedure is probabilistically modeled by length prediction, coordinate selection and symbol replacement. For inference, instead of the intractable naive approach based on Gibbs sampling to marginalize out all generation paths, the paper proposes a heuristic approach using length-conditioned beam search to generate the most likely final sequence. With the proposed framework, the paper shows that masked language models like BERT, even though they are undirected sequence models, can be used for sequence generation, which obtains close performance to the traditional left-to-right autoregressive models on the task of machine translation.\n\nOverall the paper has significant contributions in the following aspects:\n1. It enables undirected sequence models, like BERT, to perform decoding or sequence generation directly, instead of just serving as model pre-training.\n2. The proposed framework unifies directed and undirected sequence models decoding, and it can represent a few existing sequence model decoding as special cases.\n3. The coordinate sequence selection function in the framework can be dependent on the intermediate label sequence. A few simple selection approaches proposed in the paper are shown to be effective. It could be further extended. \n4. The analysis of the coordinate selection order is interesting and helpful for understanding the algorithm.\n5. The experiment results for decoding masked language models on machine translation are promising. It also provides the comparison to recent related work on non-autoregressive approaches.\n\nThe presentation of the paper is also clear. I am leaning towards accepting the paper.\n\nHowever, there are some weaknesses:\n\n1. It should be analyzed more why different coordinate selection approaches perform differently in linear-time decoding vs. constant-time decoding. Even in constant-time decoding, the conclusion varies in different decoding setting, easy-first is the worst for the L->1 case, but the best for the L/T case, why is that?\n\n2. What is the motivation for \"hard-first\"?\n\n3. The setting of \"least2most\" with L->1 is similar to Ghazvininejad et al. 2019. But Table 4 in the appendix shows the result in this paper is still worse (21.98 vs. 24.61, when both systems use 10 iterations without AR). Also, the gap from the AR baseline is larger than that in Ghazvininejad et al. 2019. Given the two systems are considered similar, it should be explained in the paper the possible reasons for these discrepancies in results.\n\nAdditional minor comments for improving the paper:\n\n1. In the introduction, it mentions the baseline AR is (Vaswani et al. 2017), while in the experimental settings, it mentions the baseline AR is (Bahdanau et al. 2015). Please clarify which one is used.\n\n2. In Table 1, how does T = 2L work for the \"Uniform\" case while the target sequence length is only T, since it is mentioned the positions are sampled without replacement. Similarly, how does T = 2L work for the \"Left2Right\" case? Is it just always choosing the last position when L < t <= 2L? In these two cases, it seems T > L is not needed.\n\n3. In Table 3, the header for the 2nd column should be o_t, as defined in Section 4 - \"Decoding scenarios\". What is the actual value of K and K'' for the constant-time machine translation experiments in the paper?\n\n4. \"Rescoring adds minimal overhead as it is run in parallel\" - it still needs to run left-to-right in sequence since it is auto-regressive. Please clarify what it means by \"in parallel\" here.\n\n5. What is the range and average for the target sentence length? How is T = 20 for constant-time decoding compared to linear-time decoding in terms of speed?"}, "tcdate": 1572444924694}, {"id": "B1gITGseqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper734/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper focuses on decoding/generation in neural sequence models (specifically machine translation) in a non-autoregressive manner that instead of generating in a left-to-right manner, focuses on generating sequences by picking a length first ,and then indices to replace in a deterministic or random scheme and, finally using a context sensitive distribution over vocabulary (using BERT-like masked LM scheme)  to pick the word to replace. In practice, this procedure of picking indices and words to replace is repeated T number of times and hence the final sequence is obtained by this iterative refinement procedure. This is an interesting and important research direction because not only would it result in better and context sensitive greedy/approximate-MAP decoded outputs, but also opens up opportunities for parallelization of the decoding procedure which is difficult to achieve with left-to-right decoders.\nThat said, the results are fairly inconclusive and the practical implementation does leave things desired for a practical decoder. As observed by the authors, different deterministic strategies for choosing T results in very different performances among the variants of the proposed approach. Besides among the variants, one clear pattern is that uniformly random picking of indices is worse than other schemes (left-to-right, least-to-most, easy-first) which is not unexpected but no conclusive empirical evidence can be found for relative differences between the performances of other 3 schemes. Moreover, the proposed decoding variants generally perform worse than or at best similarly to standard autoregressive baselines. As authors note, this is due to the mismatch between the method in which the model was trained and the decoding procedure which is not surprising, but at the same time this does not give insight into the effectiveness of the proposed decoding objective. The central question is: if the training prefers left-to-right generation then how valuable is it to device  more reasonable but incompatible decoding procedures?\n\nAlso, authors also note that index picking schemes investigated in the paper are heuristic based and a more interesting decoder could be learned if index selection procedure itself was learned with features depending on the previous index selection states, decoded states Y, and other relevant quantities. They attribute poor performance of the proposed decoder to the nature of index selection approaches investigated in the paper. I think the paper would be strengthened with results with a more sophisticated learned index selection procedure in addition to the heuristics investigated in this paper.\n\nOverall, while the idea and motivation behind this work is exciting, the inconclusive results and the approaches for practical implementation leave open significant room for improvement. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper focuses on decoding/generation in neural sequence models (specifically machine translation) in a non-autoregressive manner that instead of generating in a left-to-right manner, focuses on generating sequences by picking a length first ,and then indices to replace in a deterministic or random scheme and, finally using a context sensitive distribution over vocabulary (using BERT-like masked LM scheme)  to pick the word to replace. In practice, this procedure of picking indices and words to replace is repeated T number of times and hence the final sequence is obtained by this iterative refinement procedure. This is an interesting and important research direction because not only would it result in better and context sensitive greedy/approximate-MAP decoded outputs, but also opens up opportunities for parallelization of the decoding procedure which is difficult to achieve with left-to-right decoders.\nThat said, the results are fairly inconclusive and the practical implementation does leave things desired for a practical decoder. As observed by the authors, different deterministic strategies for choosing T results in very different performances among the variants of the proposed approach. Besides among the variants, one clear pattern is that uniformly random picking of indices is worse than other schemes (left-to-right, least-to-most, easy-first) which is not unexpected but no conclusive empirical evidence can be found for relative differences between the performances of other 3 schemes. Moreover, the proposed decoding variants generally perform worse than or at best similarly to standard autoregressive baselines. As authors note, this is due to the mismatch between the method in which the model was trained and the decoding procedure which is not surprising, but at the same time this does not give insight into the effectiveness of the proposed decoding objective. The central question is: if the training prefers left-to-right generation then how valuable is it to device  more reasonable but incompatible decoding procedures?\n\nAlso, authors also note that index picking schemes investigated in the paper are heuristic based and a more interesting decoder could be learned if index selection procedure itself was learned with features depending on the previous index selection states, decoded states Y, and other relevant quantities. They attribute poor performance of the proposed decoder to the nature of index selection approaches investigated in the paper. I think the paper would be strengthened with results with a more sophisticated learned index selection procedure in addition to the heuristics investigated in this paper.\n\nOverall, while the idea and motivation behind this work is exciting, the inconclusive results and the approaches for practical implementation leave open significant room for improvement. "}, "tcdate": 1572020925904}, {"id": "BklHgEkAFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper734/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a general framework for sentence generation using a BERT-like model. The authors decompose the problem of sentence generation into two problems. One is selecting the positions at which changes should be made, and the other is actually replacing the current word with a new word. This framework enables them to represent many decoding strategies including that of Ghazvininejas et al. (2019) in a unified manner, and they propose a new decoding strategy that considers the prediction confidence of the current and the new word. The paper also presents a heuristic algorithm for beam search decoding to find the most likely generation path. Their experimental results on the WMT14 English-German dataset suggest that the proposed approach could achieve translation quality comparable to that of the standard autoregressive approach under a constant-time translation setting.\n\nIt is nice to see existing decoding strategies represented in a generalized framework, but I was a bit disappointed that the authors do not seem to address the most critical problem in using a BERT-like model for sentence generation, namely, how to find the most likely sentence in a probabilistically sound way. It seems to me that the authors rely on at least two approximations. One is using pseudo-likelihood and the other is using the most likely generation path instead of performing marginalization. It is fine that the authors focus on empirical results of translation quality but then I would like to see more strong and extensive evidence that supports the use of such approximation.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper presents a general framework for sentence generation using a BERT-like model. The authors decompose the problem of sentence generation into two problems. One is selecting the positions at which changes should be made, and the other is actually replacing the current word with a new word. This framework enables them to represent many decoding strategies including that of Ghazvininejas et al. (2019) in a unified manner, and they propose a new decoding strategy that considers the prediction confidence of the current and the new word. The paper also presents a heuristic algorithm for beam search decoding to find the most likely generation path. Their experimental results on the WMT14 English-German dataset suggest that the proposed approach could achieve translation quality comparable to that of the standard autoregressive approach under a constant-time translation setting.\n\nIt is nice to see existing decoding strategies represented in a generalized framework, but I was a bit disappointed that the authors do not seem to address the most critical problem in using a BERT-like model for sentence generation, namely, how to find the most likely sentence in a probabilistically sound way. It seems to me that the authors rely on at least two approximations. One is using pseudo-likelihood and the other is using the most likely generation path instead of performing marginalization. It is fine that the authors focus on empirical results of translation quality but then I would like to see more strong and extensive evidence that supports the use of such approximation.\n"}, "tcdate": 1571841005135}], "openreview_url": "https://openreview.net/forum?id=BJlbo6VtDH", "arxiv_id": "1905.12790", "paper_pdf": "papers/BJlbo6VtDH.pdf", "paper_pdf_sha256": "59fdc2a025a91b4a470be9bcc1e797969e5309b334bb6fa7e0dc13cc8a25c7ee", "paper_pdf_bytes": 644135, "paper_pdf_source": "openreview", "code_url": "https://github.com/nyu-dl/dl4mt-seqgen", "code_repository": "nyu-dl/dl4mt-seqgen", "code_commit": "dd5a08182f1a32386e6203cbeaa1cee5a0b83994", "code_archive": "repos/BJlbo6VtDH.zip", "code_archive_sha256": "a396533a744ac9329db47336b6b2909a4f22168ee41009a2af87414c9ecb24cb", "code_archive_bytes": 210726, "code_file_count": 26, "code_extensions": {".py": 19, ".sh": 7}, "github_disk_usage_kb": 189, "github_languages": {"Python": 275812, "Shell": 18347, "Perl": 5233}, "github_archived": false, "github_pushed_at": "2019-06-13T21:40:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-generalized-framework-of-sequence"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bye9LiR9YX", "year": 2019, "status": "rejected", "title": "Remember and Forget for Experience Replay", "authors": ["Guido Novati", "Petros Koumoutsakos"], "authorids": ["novatig@ethz.ch", "petros@ethz.ch"], "authors_source": "OpenReview API", "abstract": "Experience replay (ER) is crucial for attaining high data-efficiency in off-policy deep reinforcement learning (RL).  ER entails the recall of experiences obtained in past iterations to compute gradient estimates for the current policy. However, the accuracy of such updates may deteriorate when the policy diverges from past behaviors, possibly undermining the effectiveness of ER. Previous off-policy RL algorithms mitigated this issue by tuning their hyper-parameters in order to abate policy changes. We propose ReF-ER, a method for active management of experiences in the Replay Memory (RM). ReF-ER forgets experiences that would be too unlikely with the current policy and constrains policy changes within a trust region of the behaviors in the RM. We couple ReF-ER with Q-learning, deterministic policy gradient and off-policy gradient methods to show that ReF-ER reliably improves the performance of continuous-action off-policy RL. We complement ReF-ER with a novel off-policy actor-critic algorithm (RACER) for continuous-action control. RACER employs a computationally efficient closed-form approximation of the action values and is shown to be highly competitive with state-of-the-art algorithms on benchmark problems, while being robust to large hyper-parameter variations.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "ryeO5Fgbam", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper198/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors introduce two new algorithms: remember and forget experience replay (ReF-ER), and an actor-critic architecture for continuous-action problems which is significantly more computationally efficient than previous approaches (RACER). ReF-ER manages the experience in the replay memory more directly and removes trajectories (episodes) that follow policies less related to the current parameterized policy (based on the importance weights). RACER's main contribution is provides a closed form approximation of the action values, enabling significant gains computationally. They provide several empirical studies in benchmark domains showing the competitiveness of their approach, and the provided more stability to various continuous control algorithms (NAF, PG, u-DDPG).\n\nOverall, I think it is a nicely written paper with a lot of empirical evidence of the usefulness of ReF-ER. I am quite interested in this algorithm specifically, as the active management of experience in the replay memory is an important step towards the ER acting as a proxy to short term memory. To my knowledge this algorithm is novel, and performs admirably. I'm less clear of the main benefits of RACER over previous approaches, except for better computational complexity. This primarily comes from a lack of empirical comparison, and not much explanation as to why key competitors were excluded. The inclusion of RACER seems to muddy the message of the paper, and a much stronger and deeper look at ReF-ER would have made for a stronger submission.\n\nI have several questions for clarity and more comments below, but overall I think the paper is quite useful for the community and contains interesting insight into active management of transitions in an experience replay buffer.\n\nPros:\n------\n\nLots of empirical studies. And a lot of details to impart intuition of the new experience replay.\n\nInteresting take on experience replay.\n\nConvincing results in many simulation benchmark domains (even though the competitors are sparse).\n\nCons:\n------\n\nThere is some ambiguity and maybe some confusion about the difference between control and off-policy learning. While I agree you are learning off-policy for control (due to the experience replay buffer containing old data), the terms off-policy and on-policy seem overused here. Statements such as \"ER has become one of the mainstay techniques to improve the sample-efficiency of off-policy RL\" aren't entirely correct as the experience replay buffer is primarily used in deep reinforcement learning to improve sample-efficiency, not off-policy reinforcement learning as a whole.\n\nThe RACER algorithm seems to muddy up the message of the paper quite a bit. I would have much preferred an in-depth look at ReF-ER here, rather than the introduction of two algorithms. And I think your paper would have been stronger for it. That being said, the RACER algorithm seems incomplete. While it is an improvement over prior approachers (ACER) computationally, the need to use ReF-ER is concerning. I'm also a bit confused why ACER isn't used as a competitor against RACER? Even if you aren't outperforming the other approach on all benchmarks, the improved computational complexity is still a worthwhile improvement.\n\nNo confidence bounds in the results, although these are somewhat shown in the appendix (without the competitors shown!!). I'm curious at the significance of the different parameter settings.\n\nQuestions:\n----------\n\nI'm curious as to how this is related to something like rejection sampling? Or other importance sampling approaches more directly? How does your method compare with using retrace or some other off-policy algorithms? I'm unclear on the reasons why these types of comparisons aren't made empirically, could you clarify more directly?\n\nDoes your algorithm help with variance issues of other off-policy algorithms? Such as just using importance weights instead of retrace? How would it effect tree backup or just the usual importance sampling? It seems likely that this would help here, as you are limiting the amount of data with high importance weights, although this might also add bias.\n\nHave you removed the target network in your experiments? This detail is not obvious in the paper currently and when you introduce ReF-ER you seem to be leading to this, but never say explicitly.\n\nYou claim that ReF-ER \"reduces the sensitivity on the network architecture and training hyper-parameters.\" I'm unclear how you show this in the results with the current paper. You do some hyperparameter studies in the appendix, but don't compare against other algorithms here. Could you share a bit further how you are measuring the sensitivities of your algorithm against the competitors?\n\nDo you need to anneal the cmax? What are the effects if this is set to some constant?\n\nCould you expand on the results of HumanoidStandup-v2? Why do you believe your approach does significantly worse than the baselines here?\n\nFor DDPG, what happens if you change the bounds instead of removing them entirely? Also how does your method compare on a domain without unbounded actions?\n\nIt is unclear why RACER does not work with ER/PER. Do you have any intuition here? Could this be fixed through means other than ReF-ER? \n\n\nOther minor comments (not taken into consideration for the review):\n-------\n\nPseudo code: It is a bit unclear what algorithm 1 is supposed to be, I'm assuming ReF-ER? \n\n\nBegin revision comments:\n-----\n\nGiven the revisions to this paper, I am more confident that it will be of interest to the community. The major contributions here I see is the removal of target networks given their approach. Given this I still have concerns on clarity and still am unhappy with the lack of confidence intervals in the main experimental section. I've increased my score to 7 to reflect my increase in confidence.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The authors introduce two new algorithms: remember and forget experience replay (ReF-ER), and an actor-critic architecture for continuous-action problems which is significantly more computationally efficient than previous approaches (RACER). ReF-ER manages the experience in the replay memory more directly and removes trajectories (episodes) that follow policies less related to the current parameterized policy (based on the importance weights). RACER's main contribution is provides a closed form approximation of the action values, enabling significant gains computationally. They provide several empirical studies in benchmark domains showing the competitiveness of their approach, and the provided more stability to various continuous control algorithms (NAF, PG, u-DDPG).\n\nOverall, I think it is a nicely written paper with a lot of empirical evidence of the usefulness of ReF-ER. I am quite interested in this algorithm specifically, as the active management of experience in the replay memory is an important step towards the ER acting as a proxy to short term memory. To my knowledge this algorithm is novel, and performs admirably. I'm less clear of the main benefits of RACER over previous approaches, except for better computational complexity. This primarily comes from a lack of empirical comparison, and not much explanation as to why key competitors were excluded. The inclusion of RACER seems to muddy the message of the paper, and a much stronger and deeper look at ReF-ER would have made for a stronger submission.\n\nI have several questions for clarity and more comments below, but overall I think the paper is quite useful for the community and contains interesting insight into active management of transitions in an experience replay buffer.\n\nPros:\n------\n\nLots of empirical studies. And a lot of details to impart intuition of the new experience replay.\n\nInteresting take on experience replay.\n\nConvincing results in many simulation benchmark domains (even though the competitors are sparse).\n\nCons:\n------\n\nThere is some ambiguity and maybe some confusion about the difference between control and off-policy learning. While I agree you are learning off-policy for control (due to the experience replay buffer containing old data), the terms off-policy and on-policy seem overused here. Statements such as \"ER has become one of the mainstay techniques to improve the sample-efficiency of off-policy RL\" aren't entirely correct as the experience replay buffer is primarily used in deep reinforcement learning to improve sample-efficiency, not off-policy reinforcement learning as a whole.\n\nThe RACER algorithm seems to muddy up the message of the paper quite a bit. I would have much preferred an in-depth look at ReF-ER here, rather than the introduction of two algorithms. And I think your paper would have been stronger for it. That being said, the RACER algorithm seems incomplete. While it is an improvement over prior approachers (ACER) computationally, the need to use ReF-ER is concerning. I'm also a bit confused why ACER isn't used as a competitor against RACER? Even if you aren't outperforming the other approach on all benchmarks, the improved computational complexity is still a worthwhile improvement.\n\nNo confidence bounds in the results, although these are somewhat shown in the appendix (without the competitors shown!!). I'm curious at the significance of the different parameter settings.\n\nQuestions:\n----------\n\nI'm curious as to how this is related to something like rejection sampling? Or other importance sampling approaches more directly? How does your method compare with using retrace or some other off-policy algorithms? I'm unclear on the reasons why these types of comparisons aren't made empirically, could you clarify more directly?\n\nDoes your algorithm help with variance issues of other off-policy algorithms? Such as just using importance weights instead of retrace? How would it effect tree backup or just the usual importance sampling? It seems likely that this would help here, as you are limiting the amount of data with high importance weights, although this might also add bias.\n\nHave you removed the target network in your experiments? This detail is not obvious in the paper currently and when you introduce ReF-ER you seem to be leading to this, but never say explicitly.\n\nYou claim that ReF-ER \"reduces the sensitivity on the network architecture and training hyper-parameters.\" I'm unclear how you show this in the results with the current paper. You do some hyperparameter studies in the appendix, but don't compare against other algorithms here. Could you share a bit further how you are measuring the sensitivities of your algorithm against the competitors?\n\nDo you need to anneal the cmax? What are the effects if this is set to some constant?\n\nCould you expand on the results of HumanoidStandup-v2? Why do you believe your approach does significantly worse than the baselines here?\n\nFor DDPG, what happens if you change the bounds instead of removing them entirely? Also how does your method compare on a domain without unbounded actions?\n\nIt is unclear why RACER does not work with ER/PER. Do you have any intuition here? Could this be fixed through means other than ReF-ER? \n\n\nOther minor comments (not taken into consideration for the review):\n-------\n\nPseudo code: It is a bit unclear what algorithm 1 is supposed to be, I'm assuming ReF-ER? \n\n\nBegin revision comments:\n-----\n\nGiven the revisions to this paper, I am more confident that it will be of interest to the community. The major contributions here I see is the removal of target networks given their approach. Given this I still have concerns on clarity and still am unhappy with the lack of confidence intervals in the main experimental section. I've increased my score to 7 to reflect my increase in confidence.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541634448315}, {"id": "B1xun4Xqhm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper198/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a method for forgetting and re-weighting experiences from a buffer during updates. It is well quantified experimentally and has some interesting tricks to improve performance in DDPG and other methods in continuous control which make use of a replay buffer. The authors also present another method “RACER” which makes use of this. \n\nI would like to see this published at some point, particularly because of the interesting results on DDPG. However, while it is interesting and useful, I do I have concerns both on the novelty and experimental comparisons in the current version. For example, RACER seems similar to ACER, yet doesn’t compare to it, making it difficult to understand what is its benefit other than its use of REFER. Moreover, the authors state that without the REFER part (with PER instead), RACER doesn’t work well at all, making it difficult to assess the RACER algorithm on its own. I would suggest if the authors claim that the contribution is the REFER algorithm they assess REFER in ACER on its own to make the main contribution stronger. \n\nRegarding novelty, I suggest that more of the paper can be spent situating the work in the broader scope of experience selection. There were several other methods that could have been compared against — for example (de Bruin et al., 2015) —which also presents a forgetting method similar to this one. While that work is cited, I don't believe it is sufficiently contrasted against this work.\n\nBelow I will examine various points/thoughts that came up.\n\n+ well experimented, appreciated the use of confidence intervals in the appendix and extensive ablation. However, I’d like to point out that the confidence intervals for some tasks spanned anything from 0 to the max, which did not inspire confidence. However, this may be a problem with the task and not the method, so not a significant problem \n+ clearly a lot of effort went into getting all these experiments and architecting the system which is well appreciated, great job there.\n+ DDPG results are promising and may indicate the problem with DDPG is its off-policy-ness. Nice results there.\n+ For the Re-Fer part, it was a bit unclear why it is 1/c_max < p_T <c_max rather than 0 < p_T < c_max? I suppose this is because you still want to update even if your current policy has not likelihood of that action? It would be nice to point to an explanation from that part of that text even if the intuition is in the appendix, otherwise it’s a bit unclear as to why this is chosen to be the acquisition function. \n+ Along these lines it would be good to see more theoretical examination of on-policiness, rather than a binary threshold of the importance weight. \n+ This paper seemed somewhat unfocused and packed with stuff, almost like two papers together which made things a bit difficult to follow as to what the main contribution is. I believe this detracted from both methods. For example, it was unclear what the benefit of using RACER was vs. say any other method which makes use of REFER. As the authors state, RACER without the ReFer part seems to not really work well at all, which makes me question this part of the contribution. It seems like a more interesting experiment would be to update importance weighted off-policy PG algorithms with the REFER part. This would hone the message which seems to be the main contribution of the paper.\n+ I find it surprising that the authors compared PPO against RACER rather than using ACER which seems like the nearest analogue to this algorithm or IMPALA which seems to have a similar parallelized architecture.\n+ More work could have been cited on experience selection selection, for example:\n\nIsele, D., & Cosgun, A. (2018). Selective Experience Replay for Lifelong Learning. arXiv preprint arXiv:1802.10269.\nPan, Yangchen, Muhammad Zaheer, Adam White, Andrew Patterson, and Martha White. \"Organizing Experience: A Deeper Look at Replay Mechanisms for Sample-based Planning in Continuous State Domains.\" arXiv preprint arXiv:1806.04624 (2018).\n\n(I am aware that these are relatively new works, but after looking at the posting timestamps, I believe the original versions were posted several months at least prior to this publication.)\n\n+ Along these lines I have concerns about the novelty since de Bruin 2015 even uses a similar off-policy metric for forgetting already. There are several differences here, but I’m not sure if they’re significantly novel for publication in its current state. \n\nTypos/Grammar Issues Found:\n\n“However, the information contained in consecutive steps is highly correlated, worsening the quality of the gradient estimate, and episodes can be composed of thousands of time step.” —> “However, the information contained in consecutive steps is highly correlated, worsening the quality of the gradient estimate, and episodes can be composed of thousands of time step(s).”", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "REMEMBER AND FORGET FOR EXPERIENCE REPLAY REVIEW", "review": "This paper presents a method for forgetting and re-weighting experiences from a buffer during updates. It is well quantified experimentally and has some interesting tricks to improve performance in DDPG and other methods in continuous control which make use of a replay buffer. The authors also present another method “RACER” which makes use of this. \n\nI would like to see this published at some point, particularly because of the interesting results on DDPG. However, while it is interesting and useful, I do I have concerns both on the novelty and experimental comparisons in the current version. For example, RACER seems similar to ACER, yet doesn’t compare to it, making it difficult to understand what is its benefit other than its use of REFER. Moreover, the authors state that without the REFER part (with PER instead), RACER doesn’t work well at all, making it difficult to assess the RACER algorithm on its own. I would suggest if the authors claim that the contribution is the REFER algorithm they assess REFER in ACER on its own to make the main contribution stronger. \n\nRegarding novelty, I suggest that more of the paper can be spent situating the work in the broader scope of experience selection. There were several other methods that could have been compared against — for example (de Bruin et al., 2015) —which also presents a forgetting method similar to this one. While that work is cited, I don't believe it is sufficiently contrasted against this work.\n\nBelow I will examine various points/thoughts that came up.\n\n+ well experimented, appreciated the use of confidence intervals in the appendix and extensive ablation. However, I’d like to point out that the confidence intervals for some tasks spanned anything from 0 to the max, which did not inspire confidence. However, this may be a problem with the task and not the method, so not a significant problem \n+ clearly a lot of effort went into getting all these experiments and architecting the system which is well appreciated, great job there.\n+ DDPG results are promising and may indicate the problem with DDPG is its off-policy-ness. Nice results there.\n+ For the Re-Fer part, it was a bit unclear why it is 1/c_max < p_T <c_max rather than 0 < p_T < c_max? I suppose this is because you still want to update even if your current policy has not likelihood of that action? It would be nice to point to an explanation from that part of that text even if the intuition is in the appendix, otherwise it’s a bit unclear as to why this is chosen to be the acquisition function. \n+ Along these lines it would be good to see more theoretical examination of on-policiness, rather than a binary threshold of the importance weight. \n+ This paper seemed somewhat unfocused and packed with stuff, almost like two papers together which made things a bit difficult to follow as to what the main contribution is. I believe this detracted from both methods. For example, it was unclear what the benefit of using RACER was vs. say any other method which makes use of REFER. As the authors state, RACER without the ReFer part seems to not really work well at all, which makes me question this part of the contribution. It seems like a more interesting experiment would be to update importance weighted off-policy PG algorithms with the REFER part. This would hone the message which seems to be the main contribution of the paper.\n+ I find it surprising that the authors compared PPO against RACER rather than using ACER which seems like the nearest analogue to this algorithm or IMPALA which seems to have a similar parallelized architecture.\n+ More work could have been cited on experience selection selection, for example:\n\nIsele, D., & Cosgun, A. (2018). Selective Experience Replay for Lifelong Learning. arXiv preprint arXiv:1802.10269.\nPan, Yangchen, Muhammad Zaheer, Adam White, Andrew Patterson, and Martha White. \"Organizing Experience: A Deeper Look at Replay Mechanisms for Sample-based Planning in Continuous State Domains.\" arXiv preprint arXiv:1806.04624 (2018).\n\n(I am aware that these are relatively new works, but after looking at the posting timestamps, I believe the original versions were posted several months at least prior to this publication.)\n\n+ Along these lines I have concerns about the novelty since de Bruin 2015 even uses a similar off-policy metric for forgetting already. There are several differences here, but I’m not sure if they’re significantly novel for publication in its current state. \n\nTypos/Grammar Issues Found:\n\n“However, the information contained in consecutive steps is highly correlated, worsening the quality of the gradient estimate, and episodes can be composed of thousands of time step.” —> “However, the information contained in consecutive steps is highly correlated, worsening the quality of the gradient estimate, and episodes can be composed of thousands of time step(s).”", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541186736216}, {"id": "Bklm4Oxqh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper198/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper first proposed a variant of experience replay to achieve better data efficiency in off-policy RL. The RACER algorithm was then developed, by modifying the approximated advantage function in the NAF algorithm. The proposed methods were finally tested on the MuJoCo environment to show the competitive performance.\n\nThis paper is in general well written. The ideas look interesting, even though they are mostly small modification of the previous works. The experiments also show the promise of the proposed methods. One of my concerns is regarding the generality of ReF-ER. I am wondering if it can be also applied to the Atari domain to boost the performance there, similar to the prioritized experience replay paper. I understand that the requirement of GPUs is beyond the hardware configuration in this work, but that would be an important contribution to the community. My other questions and comments are as follows.\n- Regarding the parametric form of f^w in Eq. (7), what are the definitions for L_+ and L_-? What are the benefits of introducing min and max there, compared with the form in Eq. (11), as used in NAF? Does it cause any problems during optimization?\n- The y axis in Figure 3 is for KL (\\pi || \\mu), while the text below used KL(\\mu || \\pi) and the description regarding the change of C also seems to be inaccurate. \n- In Figure 4, do you have any explanation why using PER leads to worse performance for NAF?\n- For the implementation, did you use any parallelization to speed up the algorithm?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper first proposed a variant of experience replay to achieve better data efficiency in off-policy RL. The RACER algorithm was then developed, by modifying the approximated advantage function in the NAF algorithm. The proposed methods were finally tested on the MuJoCo environment to show the competitive performance.\n\nThis paper is in general well written. The ideas look interesting, even though they are mostly small modification of the previous works. The experiments also show the promise of the proposed methods. One of my concerns is regarding the generality of ReF-ER. I am wondering if it can be also applied to the Atari domain to boost the performance there, similar to the prioritized experience replay paper. I understand that the requirement of GPUs is beyond the hardware configuration in this work, but that would be an important contribution to the community. My other questions and comments are as follows.\n- Regarding the parametric form of f^w in Eq. (7), what are the definitions for L_+ and L_-? What are the benefits of introducing min and max there, compared with the form in Eq. (11), as used in NAF? Does it cause any problems during optimization?\n- The y axis in Figure 3 is for KL (\\pi || \\mu), while the text below used KL(\\mu || \\pi) and the description regarding the change of C also seems to be inaccurate. \n- In Figure 4, do you have any explanation why using PER leads to worse performance for NAF?\n- For the implementation, did you use any parallelization to speed up the algorithm?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541175338963}], "openreview_url": "https://openreview.net/forum?id=Bye9LiR9YX", "arxiv_id": "1807.05827", "paper_pdf": "papers/Bye9LiR9YX.pdf", "paper_pdf_sha256": "9c3c123f4725c967172069e8397b923fa06a0d26a532a67c59c2be0e14d94adb", "paper_pdf_bytes": 3524359, "paper_pdf_source": "openreview", "code_url": "https://github.com/cselab/smarties", "code_repository": "cselab/smarties", "code_commit": "34c53d1ef03a324b20b418eaac6343ba44b6c966", "code_archive": "repos/Bye9LiR9YX.zip", "code_archive_sha256": "b808af657a87bf4813004836622e52ecec212b130f7c9209faf464cbc2e57040", "code_archive_bytes": 1463248, "code_file_count": 223, "code_extensions": {".h": 80, ".cpp": 62, ".sh": 47, ".py": 21, ".m": 5, ".f90": 3, ".hpp": 3, ".c": 2}, "github_disk_usage_kb": 2535, "github_languages": {"C++": 810341, "C": 136407, "Python": 103875, "Fortran": 22040, "Shell": 18886, "MATLAB": 13301, "Makefile": 12975, "CMake": 7959}, "github_archived": false, "github_pushed_at": "2023-12-20T16:53:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/remember-and-forget-for-experience-replay"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "z55Fut1CRw", "year": 2026, "status": "rejected", "title": "UniFork: Exploring Modality Alignment for Unified Multimodal Understanding and Generation", "authors": ["Teng Li", "Quanfeng Lu", "Lirui Zhao", "Hao Li", "Xizhou Zhu", "Yu Qiao", "Jun Zhang", "Wenqi Shao"], "authorids": ["~Teng_Li5", "~Quanfeng_Lu1", "~Lirui_Zhao1", "~Hao_Li13", "~Xizhou_Zhu1", "~Yu_Qiao1", "~Jun_Zhang25", "~Wenqi_Shao2"], "authors_source": "OpenReview API", "abstract": "Unified image understanding and generation has emerged as a promising paradigm in multimodal artificial intelligence. Despite recent progress, the optimal architectural design for such unified models remains an open challenge. In this work, we start by analyzing the modality alignment behaviors of task-specific expert models for understanding and generation, as well as current unified models. Our analysis reveals a crucial observation: understanding tasks benefit from a progressively increasing modality alignment across network depth, which helps build up semantic information for better comprehension; In contrast, generation tasks follow a different trend—modality alignment increases in the early layers but decreases in the deep layers to recover spatial details. These divergent alignment patterns create a fundamental conflict in fully shared Transformer backbones, where a uniform representational flow often leads to performance compromises across two tasks. Motivated by this finding, we introduce UniFork, a novel Y-shaped architecture that shares the shallow layers for cross-task representation learning, while employing task-specific branches in deeper layers to avoid task interference. This design effectively balances shared learning and task specialization. Through extensive ablation experiments, we demonstrate that Unifork consistently outperforms conventional fully shared Transformer architectures, and achieves performance on par with or better than task-specific models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "DBPXEqf1BJ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission924/Reviewer_a4kZ"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This paper introduces UniFork, a Y-shaped Transformer model designed to handle both image understanding and image generation in one model. It works by sharing early layers for general learning and splitting into task-specific branches later. UniFork achieves strong results on both understanding and generation tasks, outperforming several understanding-only and unified models, despite of being much smaller in size.", "review_text": "This paper introduces UniFork, a Y-shaped Transformer model designed to handle both image understanding and image generation in one model. It works by sharing early layers for general learning and splitting into task-specific branches later. UniFork achieves strong results on both understanding and generation tasks, outperforming several understanding-only and unified models, despite of being much smaller in size.", "strengths": "- The paper is mostly well written and structured. The core idea—the conflict in alignment patterns and the Y-shaped solution—is easy to follow.\n- The alignment pattern analysis is insightful and seems new. \n- On the chosen evaluation benchmarks, UniFork has shown stronger performance than the ablated instances and prior models.", "weaknesses": "- Limited Understanding Benchmarks. The understanding benchmarks tested in this paper is quite outdated (e.g., VQAv2 and GQA). For visual perception, consider including benchmarks like MMBench, BLINK, CVBench, MM-VET, MMVP. \n- Limited baseline comparison, although Janus Pro and Bagel are cited, they are being compared against UniFork.\n- The paper fixes the split point between shared and task-specific layers, but doesn’t show what happens when you change how many layers are shared (e.g., early split vs. late split). This would test how sensitive UniFork is to the choice of fork depth and whether the proposed setting is optimal.\n\nMinor Presentation Issues\n- Figure 7 caption: clouseup -> closeup\n- the GenEval benchmark is spelled “GenEval” in some places and “Geneval” in others", "questions": "In addition to the weakness above, please find additional questions below.\n\n- Is symmetric branch really necessary? As it is task-specific head anyways, would asymmetric branches (e.g., even more branches for generation) work just as well or better?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces UniFork, a Y-shaped Transformer model designed to handle both image understanding and image generation in one model. It works by sharing early layers for general learning and splitting into task-specific branches later. UniFork achieves strong results on both understanding and generation tasks, outperforming several understanding-only and unified models, despite of being much smaller in size.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper is mostly well written and structured. The core idea—the conflict in alignment patterns and the Y-shaped solution—is easy to follow.\n- The alignment pattern analysis is insightful and seems new. \n- On the chosen evaluation benchmarks, UniFork has shown stronger performance than the ablated instances and prior models.", "weaknesses": "- Limited Understanding Benchmarks. The understanding benchmarks tested in this paper is quite outdated (e.g., VQAv2 and GQA). For visual perception, consider including benchmarks like MMBench, BLINK, CVBench, MM-VET, MMVP. \n- Limited baseline comparison, although Janus Pro and Bagel are cited, they are being compared against UniFork.\n- The paper fixes the split point between shared and task-specific layers, but doesn’t show what happens when you change how many layers are shared (e.g., early split vs. late split). This would test how sensitive UniFork is to the choice of fork depth and whether the proposed setting is optimal.\n\nMinor Presentation Issues\n- Figure 7 caption: clouseup -> closeup\n- the GenEval benchmark is spelled “GenEval” in some places and “Geneval” in others", "questions": "In addition to the weakness above, please find additional questions below.\n\n- Is symmetric branch really necessary? As it is task-specific head anyways, would asymmetric branches (e.g., even more branches for generation) work just as well or better?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761934568297}, {"id": "vGEbSRnd08", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission924/Reviewer_k4Sk"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper analyzes the variation curves of modality alignment scores in the intermediate layers of understanding models and generative models from the perspective of modality alignment. Based on the differences in the score variation curves between understanding and generative models, the paper proposes UniFork, a novel Y-shaped architecture that shares shallow layers for cross-task representation learning while employing task-specific branches in deeper layers to avoid task interference. The proposed model shows a significant improvement compared to baseline models.", "review_text": "The paper analyzes the variation curves of modality alignment scores in the intermediate layers of understanding models and generative models from the perspective of modality alignment. Based on the differences in the score variation curves between understanding and generative models, the paper proposes UniFork, a novel Y-shaped architecture that shares shallow layers for cross-task representation learning while employing task-specific branches in deeper layers to avoid task interference. The proposed model shows a significant improvement compared to baseline models.", "strengths": "1. This paper innovatively analyzes unified generative large models from a modality alignment perspective. Based on this analysis, the proposed Y-shape structure is experimentally tested and demonstrates good performance\n2. The paper's presentation is good; the figures clearly convey the conclusions and experimental results of the paper.", "weaknesses": "1. The scale of the experiment is insufficient. The experimental model in the paper is only 0.5B~0.76B in size, which is too small compared to other existing unified understanding-generation models. I personally believe that it is necessary to further expand the model scale and verify the effectiveness of the method with a larger number of parameters.\n2. The other methods compared in the paper are somewhat outdated. For example, widely accepted and published unified large model papers such as show-o2, tokenflow, and unitok were not compared or discussed\n3. The paper did not conduct experiments on the number of layers for shared and unshared layers.", "questions": "Please see weakness\n1.  Regarding the curve of intermediate layer alignment scores in generative models, it shows a distribution of first rising and then falling. However, the author only conducted comparative experiments on generative models using a pixel tokenizer and neglected the unified tokenizer. If semantic information is introduced in the tokenizer, would such an alignment score curve still hold?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper analyzes the variation curves of modality alignment scores in the intermediate layers of understanding models and generative models from the perspective of modality alignment. Based on the differences in the score variation curves between understanding and generative models, the paper proposes UniFork, a novel Y-shaped architecture that shares shallow layers for cross-task representation learning while employing task-specific branches in deeper layers to avoid task interference. The proposed model shows a significant improvement compared to baseline models.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper innovatively analyzes unified generative large models from a modality alignment perspective. Based on this analysis, the proposed Y-shape structure is experimentally tested and demonstrates good performance\n2. The paper's presentation is good; the figures clearly convey the conclusions and experimental results of the paper.", "weaknesses": "1. The scale of the experiment is insufficient. The experimental model in the paper is only 0.5B~0.76B in size, which is too small compared to other existing unified understanding-generation models. I personally believe that it is necessary to further expand the model scale and verify the effectiveness of the method with a larger number of parameters.\n2. The other methods compared in the paper are somewhat outdated. For example, widely accepted and published unified large model papers such as show-o2, tokenflow, and unitok were not compared or discussed\n3. The paper did not conduct experiments on the number of layers for shared and unshared layers.", "questions": "Please see weakness\n1.  Regarding the curve of intermediate layer alignment scores in generative models, it shows a distribution of first rising and then falling. However, the author only conducted comparative experiments on generative models using a pixel tokenizer and neglected the unified tokenizer. If semantic information is introduced in the tokenizer, would such an alignment score curve still hold?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761814888385}, {"id": "1YR6U5LJQ3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission924/Reviewer_5cnP"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper introduces UniFork, a novel Y-shaped architecture that shares the shallow layers for cross-task representation learning for unified model. The authors diagnose that image understanding and image generation tasks demand different cross-modal feature alignment behaviours in transformer layers. Experiments demonstrate performance gains relative to fully shared architectures, achieving similar or better results compared to task-specific expert models.", "review_text": "The paper introduces UniFork, a novel Y-shaped architecture that shares the shallow layers for cross-task representation learning for unified model. The authors diagnose that image understanding and image generation tasks demand different cross-modal feature alignment behaviours in transformer layers. Experiments demonstrate performance gains relative to fully shared architectures, achieving similar or better results compared to task-specific expert models.", "strengths": "1. Insightful analysis of modality alignment patterns: The paper provides a systematic investigation into the alignment dynamics between text and image representations across different architectures and tasks (understanding vs. generation). By empirically identifying distinct alignment patterns and connecting them to architectural design choices, the work offers valuable conceptual insights for building more principled and efficient unified multimodal models.\n2. Well-documented training and implementation details: The paper reports detailed information on training strategy, data composition ratios, and optimization settings, which facilitates reproducibility.", "weaknesses": "1. Outdated and weak baselines: The compared baselines (e.g., MobileVLM, Emu, LaVIT, LDM, LWM) are mostly early-generation unified or multimodal models whose performance and architecture are now considerably behind state-of-the-art models such as Emu3, Bagel, or Janus-Pro. Even when compared to these relatively weak baselines, UniFork shows only marginal or inconsistent improvements across several benchmarks. This weakens the empirical strength of the claimed performance gains and raises questions about whether the proposed method would still hold advantages under stronger baselines.\n2. Lack of ablation or sensitivity analysis on branch duplication and split ratio: The proposed UniFork architecture duplicates the latter half of the Transformer backbone (based on Qwen2.5) to construct two independent branches for understanding and generation. However, there is no thorough analysis or comparison regarding this specific design choice—e.g., why the fork point is placed in the latter half, and how sensitive the results are to the split ratio between shared and task-specific layers.\nAs described in §3.2 (“Given a Transformer of (M + N) total layers”), the determination of M and N (the numbers of shared and branch-specific layers, respectively) seems largely heuristic. A systematic study or empirical justification of these hyperparameters would substantially strengthen the paper’s credibility and generality.", "questions": "1. Choice of fork depth (M, N): In §3.2 the authors mention that the Transformer backbone is divided into M shared and N task-specific layers (“Given a Transformer of (M + N) total layers”). How was the specific fork point determined in practice?\n  - Was it based on alignment measurements, empirical tuning, or a fixed heuristic?\n  - Have you explored sensitivity analysis on different M/N splits (e.g., forking earlier vs. later), and how does this affect performance or modality alignment?\n2. Scalability and future experiments: The current experiments are conducted on relatively moderate-scale models. Do the authors plan to extend UniFork to larger-scale settings to further validate the scalability and general effectiveness of the proposed architecture?\nIt would be valuable to know whether the observed alignment patterns and performance trends remain consistent at higher scales.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces UniFork, a novel Y-shaped architecture that shares the shallow layers for cross-task representation learning for unified model. The authors diagnose that image understanding and image generation tasks demand different cross-modal feature alignment behaviours in transformer layers. Experiments demonstrate performance gains relative to fully shared architectures, achieving similar or better results compared to task-specific expert models.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Insightful analysis of modality alignment patterns: The paper provides a systematic investigation into the alignment dynamics between text and image representations across different architectures and tasks (understanding vs. generation). By empirically identifying distinct alignment patterns and connecting them to architectural design choices, the work offers valuable conceptual insights for building more principled and efficient unified multimodal models.\n2. Well-documented training and implementation details: The paper reports detailed information on training strategy, data composition ratios, and optimization settings, which facilitates reproducibility.", "weaknesses": "1. Outdated and weak baselines: The compared baselines (e.g., MobileVLM, Emu, LaVIT, LDM, LWM) are mostly early-generation unified or multimodal models whose performance and architecture are now considerably behind state-of-the-art models such as Emu3, Bagel, or Janus-Pro. Even when compared to these relatively weak baselines, UniFork shows only marginal or inconsistent improvements across several benchmarks. This weakens the empirical strength of the claimed performance gains and raises questions about whether the proposed method would still hold advantages under stronger baselines.\n2. Lack of ablation or sensitivity analysis on branch duplication and split ratio: The proposed UniFork architecture duplicates the latter half of the Transformer backbone (based on Qwen2.5) to construct two independent branches for understanding and generation. However, there is no thorough analysis or comparison regarding this specific design choice—e.g., why the fork point is placed in the latter half, and how sensitive the results are to the split ratio between shared and task-specific layers.\nAs described in §3.2 (“Given a Transformer of (M + N) total layers”), the determination of M and N (the numbers of shared and branch-specific layers, respectively) seems largely heuristic. A systematic study or empirical justification of these hyperparameters would substantially strengthen the paper’s credibility and generality.", "questions": "1. Choice of fork depth (M, N): In §3.2 the authors mention that the Transformer backbone is divided into M shared and N task-specific layers (“Given a Transformer of (M + N) total layers”). How was the specific fork point determined in practice?\n  - Was it based on alignment measurements, empirical tuning, or a fixed heuristic?\n  - Have you explored sensitivity analysis on different M/N splits (e.g., forking earlier vs. later), and how does this affect performance or modality alignment?\n2. Scalability and future experiments: The current experiments are conducted on relatively moderate-scale models. Do the authors plan to extend UniFork to larger-scale settings to further validate the scalability and general effectiveness of the proposed architecture?\nIt would be valuable to know whether the observed alignment patterns and performance trends remain consistent at higher scales.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761741820036}, {"id": "jNeLS5r4I3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission924/Reviewer_P6ZU"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "Empirical characterization of divergent modality-alignment patterns for understanding vs generation and evidence of representational compromise in fully shared NTP models.\n\nUniFork Y-architecture that shares early layers and decouples late layers into task-specific branches while keeping a simple AR/NTP interface.\n\nSimple, effective training recipe (three stages, no loss reweighting) that allows independent branch fine-tuning without delicate data balancing.", "review_text": "Empirical characterization of divergent modality-alignment patterns for understanding vs generation and evidence of representational compromise in fully shared NTP models.\n\nUniFork Y-architecture that shares early layers and decouples late layers into task-specific branches while keeping a simple AR/NTP interface.\n\nSimple, effective training recipe (three stages, no loss reweighting) that allows independent branch fine-tuning without delicate data balancing.", "strengths": "The paper’s exploration of modality alignment patterns in understanding versus generation tasks fills a gap in prior research by investigating how these tasks differ at a deep level. The observation that understanding benefits from progressively increasing alignment, while generation requires an initial rise and later drop, is a novel and insightful contribution to the multimodal field.\n\nThe presentation of the Y-shaped Transformer architecture is clear, and the distinction between the shared early layers and task-specific branches is well-articulated. The decision to share the first half of the layers and split the latter half for task-specific learning is explained in an understandable manner, making the technical concept more accessible without oversimplifying it.", "weaknesses": "The paper correctly notes that when M=0, UniFork is structurally similar to Mixture-of-Transformers designs like BAGEL, and its dual-pathway approach is conceptually related to the Janus series. However, the results tables lack a direct, quantitative comparison with these specific models under similar training budgets. This omission makes it difficult for the reader to gauge whether the Y-shaped architecture provides a tangible advantage over other state-of-the-art parameter-efficient unification strategies.\n\nActionable Improvement: Include direct comparisons with BAGEL and/or a Janus-series model in Tables 2, 3, and 4. This would precisely delineate UniFork's contribution within the existing landscape of unified model architectures.\n\nThe performance jump from the 0.57B to 0.76B ablation variant is promising, but it's a single data point. A critical unanswered question is how the UniFork architecture behaves at different scales (e.g., with a 3B or 7B backbone). Does the optimal ratio of shared-to-task-specific layers (M and N) change with model size?\n\nActionable Improvement: Conduct a basic scaling study. For instance, implement UniFork on two different backbone sizes (e.g., 0.5B and 2B) and show that the architecture's benefits are consistent. This would significantly bolster the claim that it is a general-purpose design.\n\n\nThe generative performance is assessed primarily on GenEval (text-image alignment) and FID (general quality). This misses important dimensions like compositional reasoning, complex scene rendering, and aesthetic quality.\n\nActionable Improvement: Supplement the evaluation with benchmarks like T2I-CompBench， WISE for compositionality and, ideally, a human evaluation study to assess image realism and prompt adherence on more complex, creative prompts.\n\nThe \"Where to Fork\" Ablation is Missing: The analysis reveals that the alignment trends diverge, but the choice to decouple specifically in the later layers is a design decision. The paper lacks an ablation study that justifies this specific choice. Is forking in the final third of the network truly optimal? What is the performance impact of forking earlier (e.g., after the first third) or later?\n\nActionable Improvement: Introduce a critical ablation experiment that varies the forking point M within the Transformer stack while holding total depth constant. A plot showing task performance versus forking depth would provide direct, empirical evidence that the chosen configuration is optimal, transforming a design choice into a data-driven conclusion.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Empirical characterization of divergent modality-alignment patterns for understanding vs generation and evidence of representational compromise in fully shared NTP models.\n\nUniFork Y-architecture that shares early layers and decouples late layers into task-specific branches while keeping a simple AR/NTP interface.\n\nSimple, effective training recipe (three stages, no loss reweighting) that allows independent branch fine-tuning without delicate data balancing.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The paper’s exploration of modality alignment patterns in understanding versus generation tasks fills a gap in prior research by investigating how these tasks differ at a deep level. The observation that understanding benefits from progressively increasing alignment, while generation requires an initial rise and later drop, is a novel and insightful contribution to the multimodal field.\n\nThe presentation of the Y-shaped Transformer architecture is clear, and the distinction between the shared early layers and task-specific branches is well-articulated. The decision to share the first half of the layers and split the latter half for task-specific learning is explained in an understandable manner, making the technical concept more accessible without oversimplifying it.", "weaknesses": "The paper correctly notes that when M=0, UniFork is structurally similar to Mixture-of-Transformers designs like BAGEL, and its dual-pathway approach is conceptually related to the Janus series. However, the results tables lack a direct, quantitative comparison with these specific models under similar training budgets. This omission makes it difficult for the reader to gauge whether the Y-shaped architecture provides a tangible advantage over other state-of-the-art parameter-efficient unification strategies.\n\nActionable Improvement: Include direct comparisons with BAGEL and/or a Janus-series model in Tables 2, 3, and 4. This would precisely delineate UniFork's contribution within the existing landscape of unified model architectures.\n\nThe performance jump from the 0.57B to 0.76B ablation variant is promising, but it's a single data point. A critical unanswered question is how the UniFork architecture behaves at different scales (e.g., with a 3B or 7B backbone). Does the optimal ratio of shared-to-task-specific layers (M and N) change with model size?\n\nActionable Improvement: Conduct a basic scaling study. For instance, implement UniFork on two different backbone sizes (e.g., 0.5B and 2B) and show that the architecture's benefits are consistent. This would significantly bolster the claim that it is a general-purpose design.\n\n\nThe generative performance is assessed primarily on GenEval (text-image alignment) and FID (general quality). This misses important dimensions like compositional reasoning, complex scene rendering, and aesthetic quality.\n\nActionable Improvement: Supplement the evaluation with benchmarks like T2I-CompBench， WISE for compositionality and, ideally, a human evaluation study to assess image realism and prompt adherence on more complex, creative prompts.\n\nThe \"Where to Fork\" Ablation is Missing: The analysis reveals that the alignment trends diverge, but the choice to decouple specifically in the later layers is a design decision. The paper lacks an ablation study that justifies this specific choice. Is forking in the final third of the network truly optimal? What is the performance impact of forking earlier (e.g., after the first third) or later?\n\nActionable Improvement: Introduce a critical ablation experiment that varies the forking point M within the Transformer stack while holding total depth constant. A plot showing task performance versus forking depth would provide direct, empirical evidence that the chosen configuration is optimal, transforming a design choice into a data-driven conclusion.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760591513388}], "openreview_url": "https://openreview.net/forum?id=z55Fut1CRw", "arxiv_id": "2506.17202", "paper_pdf": "papers/z55Fut1CRw.pdf", "paper_pdf_sha256": "26bcac7a02ff6bd6d51ba9eb1a14ee0c2d8bd003f82b4e84ee53468af0f6067a", "paper_pdf_bytes": 12781869, "paper_pdf_source": "openreview", "code_url": "https://github.com/tliby/UniFork", "code_repository": "tliby/UniFork", "code_commit": "2c5b75750e880589aacf2ab5f25efafb37477272", "code_archive": "repos/z55Fut1CRw.zip", "code_archive_sha256": "fd6e1de5f4bae70d0c6fa01daf42bb00ebd51d68380ce9be30308683258da5ee", "code_archive_bytes": 684083, "code_file_count": 77, "code_extensions": {".py": 66, ".sh": 11}, "github_disk_usage_kb": 915, "github_languages": {"Python": 599051, "Shell": 17112}, "github_archived": false, "github_pushed_at": "2025-08-26T03:58:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unifork-exploring-modality-alignment-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pbJMPo4HwR", "year": 2025, "status": "rejected", "title": "Learning Monotonic Attention in Transducer for Streaming Generation", "authors": ["Zhengrui Ma", "Yang Feng", "Min zhang"], "authorids": ["~Zhengrui_Ma1", "~Yang_Feng4", "~Min_zhang14"], "authors_source": "OpenReview API", "abstract": "Streaming generation models are increasingly utilized across various fields, with the Transducer architecture being particularly popular in industrial applications. However, its input-synchronous decoding mechanism presents challenges in tasks requiring non-monotonic alignments, such as simultaneous translation, leading to suboptimal performance in these contexts. In this research, we address this issue by tightly integrating Transducer's decoding with the history of input stream via a learnable monotonic attention mechanism. Our approach leverages the forward-backward algorithm to infer the posterior probability of alignments between the predictor states and input timestamps, which is then used to estimate the context representations of monotonic attention in training. This allows Transducer models to adaptively adjust the scope of attention based on their predictions, avoiding the need to enumerate the exponentially large alignment space. Extensive experiments demonstrate that our MonoAttn-Transducer significantly enhances the handling of non-monotonic alignments in streaming generation, offering a robust solution for Transducer-based frameworks to tackle more complex streaming generation tasks. Codes are publicly available in supplementary materials.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "X8MuB7t5so", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission350/Reviewer_zyS6"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper discusses the challenges faced by Transducer-based streaming generation models, particularly in tasks requiring non-monotonic alignments, such as simultaneous translation. These challenges arise from the input-synchronous decoding mechanism of Transducers, which can result in suboptimal performance. To address this, the authors propose integrating a learnable monotonic attention mechanism within the Transducer architecture. This mechanism uses a forward-backward algorithm to calculate the posterior probability of alignments between predictor states and input timestamps. Consequently, it allows for the estimation of context representations of monotonic attention during training, enabling the model to adaptively adjust its attention scope based on predictions. This innovative approach eliminates the need to explore the exponentially large alignment space. Experiments reveal that the proposed MonoAttn-Transducer significantly improves performance in streaming generation tasks dealing with non-monotonic alignments. The codes for this study are available in the supplementary materials.", "review_text": "This paper discusses the challenges faced by Transducer-based streaming generation models, particularly in tasks requiring non-monotonic alignments, such as simultaneous translation. These challenges arise from the input-synchronous decoding mechanism of Transducers, which can result in suboptimal performance. To address this, the authors propose integrating a learnable monotonic attention mechanism within the Transducer architecture. This mechanism uses a forward-backward algorithm to calculate the posterior probability of alignments between predictor states and input timestamps. Consequently, it allows for the estimation of context representations of monotonic attention during training, enabling the model to adaptively adjust its attention scope based on predictions. This innovative approach eliminates the need to explore the exponentially large alignment space. Experiments reveal that the proposed MonoAttn-Transducer significantly improves performance in streaming generation tasks dealing with non-monotonic alignments. The codes for this study are available in the supplementary materials.", "strengths": "1. The authors propose a learnable monotonic attention mechanism within the Transducer architecture to solve the non-monotonic alignments in streaming generation.\n\n2. The propsoed method is effectiveness and easy to reproduce.", "weaknesses": "1. I am unable to discern the difference between the proposed monotonic attention mechanism and the cross-attention mechanism.\n\n2. Furthermore, the authors should compare the cross-attention approach (Liu et al., 2021; Tang et al., 2023) with other methods that utilize input history.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper discusses the challenges faced by Transducer-based streaming generation models, particularly in tasks requiring non-monotonic alignments, such as simultaneous translation. These challenges arise from the input-synchronous decoding mechanism of Transducers, which can result in suboptimal performance. To address this, the authors propose integrating a learnable monotonic attention mechanism within the Transducer architecture. This mechanism uses a forward-backward algorithm to calculate the posterior probability of alignments between predictor states and input timestamps. Consequently, it allows for the estimation of context representations of monotonic attention during training, enabling the model to adaptively adjust its attention scope based on predictions. This innovative approach eliminates the need to explore the exponentially large alignment space. Experiments reveal that the proposed MonoAttn-Transducer significantly improves performance in streaming generation tasks dealing with non-monotonic alignments. The codes for this study are available in the supplementary materials.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The authors propose a learnable monotonic attention mechanism within the Transducer architecture to solve the non-monotonic alignments in streaming generation.\n\n2. The propsoed method is effectiveness and easy to reproduce.", "weaknesses": "1. I am unable to discern the difference between the proposed monotonic attention mechanism and the cross-attention mechanism.\n\n2. Furthermore, the authors should compare the cross-attention approach (Liu et al., 2021; Tang et al., 2023) with other methods that utilize input history.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730718981175}, {"id": "wyIAh8vkmj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission350/Reviewer_uLi8"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper proposes MonoAttn-Transducer, a novel approach to enhance Transducer models with learnable monotonic attention for streaming generation tasks. The key contributions are:\n\n- A method to integrate monotonic attention into Transducer's architecture while maintaining its efficient training through the forward-backward algorithm.\n- A training algorithm that uses posterior alignment probabilities to estimate context representations for monotonic attention.\n- A chunk synchronization mechanism to bridge the gap between training and inference.\n- Extensive experiments demonstrating improved performance on simultaneous translation tasks.", "review_text": "This paper proposes MonoAttn-Transducer, a novel approach to enhance Transducer models with learnable monotonic attention for streaming generation tasks. The key contributions are:\n\n- A method to integrate monotonic attention into Transducer's architecture while maintaining its efficient training through the forward-backward algorithm.\n- A training algorithm that uses posterior alignment probabilities to estimate context representations for monotonic attention.\n- A chunk synchronization mechanism to bridge the gap between training and inference.\n- Extensive experiments demonstrating improved performance on simultaneous translation tasks.", "strengths": "Technical Innovation:\n\n- The proposed method cleverly solves the exponential state space problem by using posterior alignments to estimate context representations\n- The solution maintains the same computational complexity as vanilla Transducer\n- The chunk synchronization mechanism shows thoughtful consideration of practical deployment\n\nTheoretical Foundation:\n\n- The approach is well-grounded in probability theory and previous work on Transducers\n- The mathematical derivations are sound and clearly explained\n- The relationship between prior and posterior alignments is well-analyzed\n\nEmpirical Results:\n\n- Comprehensive experiments on MuST-C dataset\n- Strong improvements over baseline Transducer (+0.75-1.0 BLEU, +0.95-2.06 COMET)\n- Thorough ablation studies and analysis\n- Competitive performance against SOTA methods", "weaknesses": "Limited Experimental Scope:\n\n- Experiments focus only on speech-to-text translation.\n- Only two language pairs (En->De, En->Es) are tested.\n- No experiments on other streaming generation tasks like ASR or TTS.\n\nTraining Efficiency:\n\n- The paper doesn't discuss training time comparison with baseline Transducer.\n- Memory usage analysis could be more detailed.\n- No discussion of potential overhead from posterior alignment calculation.\n\n\nAlgorithmic Limitations:\n\n- The method still requires chunk-based processing.\n- The impact of chunk size on performance could be better theoretically explained.\n- The choice of diagonal prior distribution seems somewhat arbitrary.", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes MonoAttn-Transducer, a novel approach to enhance Transducer models with learnable monotonic attention for streaming generation tasks. The key contributions are:\n\n- A method to integrate monotonic attention into Transducer's architecture while maintaining its efficient training through the forward-backward algorithm.\n- A training algorithm that uses posterior alignment probabilities to estimate context representations for monotonic attention.\n- A chunk synchronization mechanism to bridge the gap between training and inference.\n- Extensive experiments demonstrating improved performance on simultaneous translation tasks.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "Technical Innovation:\n\n- The proposed method cleverly solves the exponential state space problem by using posterior alignments to estimate context representations\n- The solution maintains the same computational complexity as vanilla Transducer\n- The chunk synchronization mechanism shows thoughtful consideration of practical deployment\n\nTheoretical Foundation:\n\n- The approach is well-grounded in probability theory and previous work on Transducers\n- The mathematical derivations are sound and clearly explained\n- The relationship between prior and posterior alignments is well-analyzed\n\nEmpirical Results:\n\n- Comprehensive experiments on MuST-C dataset\n- Strong improvements over baseline Transducer (+0.75-1.0 BLEU, +0.95-2.06 COMET)\n- Thorough ablation studies and analysis\n- Competitive performance against SOTA methods", "weaknesses": "Limited Experimental Scope:\n\n- Experiments focus only on speech-to-text translation.\n- Only two language pairs (En->De, En->Es) are tested.\n- No experiments on other streaming generation tasks like ASR or TTS.\n\nTraining Efficiency:\n\n- The paper doesn't discuss training time comparison with baseline Transducer.\n- Memory usage analysis could be more detailed.\n- No discussion of potential overhead from posterior alignment calculation.\n\n\nAlgorithmic Limitations:\n\n- The method still requires chunk-based processing.\n- The impact of chunk size on performance could be better theoretically explained.\n- The choice of diagonal prior distribution seems somewhat arbitrary.", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730713045541}, {"id": "mxfep6JZEf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission350/Reviewer_ak6X"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper describes an extension for Transducers for tasks where the input and output are not monotonically aligned, such as speech translation. The method is evaluated on two languages of the MuST-C benchmark and it is compared to existing methods.", "review_text": "This paper describes an extension for Transducers for tasks where the input and output are not monotonically aligned, such as speech translation. The method is evaluated on two languages of the MuST-C benchmark and it is compared to existing methods.", "strengths": "* The paper addresses an interesting setting which has practical implications\n* The method is described in detail and a large portion of the paper is dedicated to this description.", "weaknesses": "* The method is fairly complex, requiring a range of steps in addition to Transducer training as shown in Algorithm 1 (which is already complex).\n* The description of the method is sometimes not very clear. It would be helpful to provide intuition in addition to equations.\n* The evaluation was done on languages with relatively similar word order: En-Es and En-De differ in their word order but much less so, then say En-Chinese or English and any non-European language. There are speech translation benchmarks which enable these settings, e.g., FLEURS and which would provide more interesting results.\n* It seems a bit surprising that the BLEU improvements for En-Es and En-De between Transducer and the new method seem fairly similar across most chunk sizes (about 1 BLEU, Table 2). I would have expected En-De to benefit more the new method than En-Es given that the word order of En-De is more different.", "questions": "* Did you consider evaluating on other settings than En-De/En-Es where word order is more different?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper describes an extension for Transducers for tasks where the input and output are not monotonically aligned, such as speech translation. The method is evaluated on two languages of the MuST-C benchmark and it is compared to existing methods.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* The paper addresses an interesting setting which has practical implications\n* The method is described in detail and a large portion of the paper is dedicated to this description.", "weaknesses": "* The method is fairly complex, requiring a range of steps in addition to Transducer training as shown in Algorithm 1 (which is already complex).\n* The description of the method is sometimes not very clear. It would be helpful to provide intuition in addition to equations.\n* The evaluation was done on languages with relatively similar word order: En-Es and En-De differ in their word order but much less so, then say En-Chinese or English and any non-European language. There are speech translation benchmarks which enable these settings, e.g., FLEURS and which would provide more interesting results.\n* It seems a bit surprising that the BLEU improvements for En-Es and En-De between Transducer and the new method seem fairly similar across most chunk sizes (about 1 BLEU, Table 2). I would have expected En-De to benefit more the new method than En-Es given that the word order of En-De is more different.", "questions": "* Did you consider evaluating on other settings than En-De/En-Es where word order is more different?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730533606041}, {"id": "laBiPZlV5q", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission350/Reviewer_2LWH"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents a solution to the challenges faced by Transducer-based streaming generation models in handling non-monotonic alignments, particularly in simultaneous translation tasks. The authors introduce a learnable monotonic attention mechanism that integrates with Transducer decoding. Utilizing the forward-backward algorithm, they infer alignment probabilities between predictor states and input timestamps, allowing adaptive adjustments in attention scope. Experimental results show significant performance improvements of the MonoAttn-Transducer, offering a robust method for complex streaming generation tasks. Their code is publicly available.", "review_text": "This paper presents a solution to the challenges faced by Transducer-based streaming generation models in handling non-monotonic alignments, particularly in simultaneous translation tasks. The authors introduce a learnable monotonic attention mechanism that integrates with Transducer decoding. Utilizing the forward-backward algorithm, they infer alignment probabilities between predictor states and input timestamps, allowing adaptive adjustments in attention scope. Experimental results show significant performance improvements of the MonoAttn-Transducer, offering a robust method for complex streaming generation tasks. Their code is publicly available.", "strengths": "1. This paper improves the Transducer architecture, making it more suitable for streaming generation tasks.  \n2. The experiments demonstrate that the proposed method achieves commendable performance in the Speech-to-text Simultaneous Translation task.", "weaknesses": "1. The experiments in this paper only provide results on the Speech-to-Text Simultaneous Translation task, and do not validate the method on other streaming generation tasks.  \n2. Under lower latency conditions (about 1000 ms) for the EnDe and EnEs tasks, the performance of MA-T appears to be inferior to that of CAAT, another transducer-based method. This may indicate that the proposed method in this paper is not highly effective.  \n3. The proposed method in this paper shares a very similar model structure with both CAAT (Liu et al. 2021) and  MILK (Arivazhagan et al 2019), which raises concerns about the lack of novelty in the article.\n\n**Reference**\n\n[1]Dan Liu, Mengge Du, Xiaoxi Li, Ya Li, and Enhong Chen. Cross attention augmented transducer networks for simultaneous translation. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.\n\n[2]Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. Monotonic infinite lookback attention for simultaneous machine translation. In Anna Korhonen, David Traum, and Llu´ıs Marquez (eds.), ` Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.2019.", "questions": "1. After incorporating a unidirectional encoder, the model structures of MA-T and CAAT are very similar. What are the specific differences between the two? How does MA-T manage to perform attention computations while ensuring that Memory Overload remains O(1)?  \n2. Using AL as a latency metric may not accurately assess the phenomenon of over-generation. Could you provide alternative latency metrics, such as LAAL or LAAL-CA?  \n3. Regarding the comparison between MA-T and TAED, could you provide more experimental results to demonstrate that your method truly achieves better performance than MA-T? From the perspective of AL and BLEU, while TAED has a larger computational overhead, it does achieve higher translation quality at a lower latency and the overhead is still acceptable.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a solution to the challenges faced by Transducer-based streaming generation models in handling non-monotonic alignments, particularly in simultaneous translation tasks. The authors introduce a learnable monotonic attention mechanism that integrates with Transducer decoding. Utilizing the forward-backward algorithm, they infer alignment probabilities between predictor states and input timestamps, allowing adaptive adjustments in attention scope. Experimental results show significant performance improvements of the MonoAttn-Transducer, offering a robust method for complex streaming generation tasks. Their code is publicly available.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper improves the Transducer architecture, making it more suitable for streaming generation tasks.  \n2. The experiments demonstrate that the proposed method achieves commendable performance in the Speech-to-text Simultaneous Translation task.", "weaknesses": "1. The experiments in this paper only provide results on the Speech-to-Text Simultaneous Translation task, and do not validate the method on other streaming generation tasks.  \n2. Under lower latency conditions (about 1000 ms) for the EnDe and EnEs tasks, the performance of MA-T appears to be inferior to that of CAAT, another transducer-based method. This may indicate that the proposed method in this paper is not highly effective.  \n3. The proposed method in this paper shares a very similar model structure with both CAAT (Liu et al. 2021) and  MILK (Arivazhagan et al 2019), which raises concerns about the lack of novelty in the article.\n\n**Reference**\n\n[1]Dan Liu, Mengge Du, Xiaoxi Li, Ya Li, and Enhong Chen. Cross attention augmented transducer networks for simultaneous translation. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.\n\n[2]Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. Monotonic infinite lookback attention for simultaneous machine translation. In Anna Korhonen, David Traum, and Llu´ıs Marquez (eds.), ` Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.2019.", "questions": "1. After incorporating a unidirectional encoder, the model structures of MA-T and CAAT are very similar. What are the specific differences between the two? How does MA-T manage to perform attention computations while ensuring that Memory Overload remains O(1)?  \n2. Using AL as a latency metric may not accurately assess the phenomenon of over-generation. Could you provide alternative latency metrics, such as LAAL or LAAL-CA?  \n3. Regarding the comparison between MA-T and TAED, could you provide more experimental results to demonstrate that your method truly achieves better performance than MA-T? From the perspective of AL and BLEU, while TAED has a larger computational overhead, it does achieve higher translation quality at a lower latency and the overhead is still acceptable.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730454741492}], "openreview_url": "https://openreview.net/forum?id=pbJMPo4HwR", "arxiv_id": "2411.17170", "paper_pdf": "papers/pbJMPo4HwR.pdf", "paper_pdf_sha256": "2353813ef1371aa5a8b8eb4d5ed58806cecde4a3471fcc6e219278409fa85873", "paper_pdf_bytes": 3022089, "paper_pdf_source": "openreview", "code_url": "https://github.com/ictnlp/MonoAttn-Transducer", "code_repository": "ictnlp/MonoAttn-Transducer", "code_commit": "518d0555d020bfff543409bd1806541b40a6b496", "code_archive": "repos/pbJMPo4HwR.zip", "code_archive_sha256": "709a6cf896e6c2edff27670ef746f118dd8516ec1b6b65a7b0230ba60e422c3d", "code_archive_bytes": 96407, "code_file_count": 37, "code_extensions": {".py": 37}, "github_disk_usage_kb": 77, "github_languages": {"Python": 329821}, "github_archived": false, "github_pushed_at": "2025-05-19T17:59:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-monotonic-attention-in-transducer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OlwW4ZG3Ta", "year": 2024, "status": "rejected", "title": "Reflective Policy Optimization", "authors": ["Yaozhong Gan", "Renye Yan", "Zhe Wu", "Junliang Xing"], "authorids": ["~Yaozhong_Gan1", "~Renye_Yan1", "~Zhe_Wu6", "~Junliang_Xing1"], "authors_source": "OpenReview API", "abstract": "On-policy reinforcement learning methods, such as Trust Region Policy Optimization (TRPO) and Proximal Policy Optimization (PPO), often require significant data to be collected at each update, giving rise to issues of sample inefficiency. This paper introduces a novel extension to on-policy methods called Reflective Policy Optimization (RPO). RPO's fundamental objective is amalgamating prior and subsequent state and action information from trajectory data to optimize the current policy. This approach empowers the agent to engage in introspection and introduce modifications to its actions within the current state to a certain degree. Furthermore, theoretical analyses substantiate that our proposed method not only upholds the crucial property of monotonically improving policy performance but also adeptly contracts the solution space of the optimized policy, consequently expediting the training procedure. We empirically demonstrate the feasibility and efficacy of our approach in reinforcement learning benchmarks, culminating in superior performance in terms of sample efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "i1Qky1ED9f", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6789/Reviewer_kAky"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a method called Reflexive Policy Optimization, which the authors claim uses information “from trajectory data” more efficiently to optimize a policy. The authors claim that agents have “introspection” and “reflect on prior experience”. The method is claimed to have guaranteed monotonic progress improvement on the original policy performance objective and that it “contracts the solution space of the optimized policy”, thus “expediting” the training procedure. Moreover, the authors claim the method is superior to baselines on Mujoco tasks.\n\nI didn’t really understand the motivation for the paper, and I found it very confusing and difficult to follow. I expand in the following sections.", "review_text": "The paper proposes a method called Reflexive Policy Optimization, which the authors claim uses information “from trajectory data” more efficiently to optimize a policy. The authors claim that agents have “introspection” and “reflect on prior experience”. The method is claimed to have guaranteed monotonic progress improvement on the original policy performance objective and that it “contracts the solution space of the optimized policy”, thus “expediting” the training procedure. Moreover, the authors claim the method is superior to baselines on Mujoco tasks.\n\nI didn’t really understand the motivation for the paper, and I found it very confusing and difficult to follow. I expand in the following sections.", "strengths": "I believe the authors want to say that their method performs a kind of hindsight credit assignment \\citep{harutyunyan2019}, but this is just a hunch. I didn't really understand from the text.\nThe paper appears to be backed by theory and support their superior performance claims with empirical results. Unfortunately, I did not really understand the motivation and method to assess these  properly. Please see below my points of confusion. I am happy to revise my score if the authors the motivation and goal of the paper, and expand on the points below.\n\n\n@article{harutyunyan2019,\n  author       = {Anna Harutyunyan and\n                  Will Dabney and\n                  Thomas Mesnard and\n                  Mohammad Gheshlaghi Azar and\n                  Bilal Piot and\n                  Nicolas Heess and\n                  Hado van Hasselt and\n                  Greg Wayne and\n                  Satinder Singh and\n                  Doina Precup and\n                  R{\\'{e}}mi Munos},\n  title        = {Hindsight Credit Assignment},\n  journal      = {CoRR},\n  volume       = {abs/1912.02503},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1912.02503},\n  eprinttype    = {arXiv},\n  eprint       = {1912.02503},\n  timestamp    = {Wed, 20 Apr 2022 07:47:18 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1912-02503.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}", "weaknesses": "First, I do not understand why the authors believe proximal methods do not already account for the true gradient of the policy which also considers the contribution through the stationary distribution.  Indeed, in the original CPI paper \\citep{kakade}, the authors first derive the form of the bound that contains the stationary distribution with the current policy, and not a prior policy, then use a mixture policy to ensure that the new policy is “close” to the prior one and justify replacing the current stationary distribution with the previous one. However, a lot of work has been done since, and it is known that proximal PG methods \\citep{bhandari21, shani2019, vaswani2021} (and algorithmic implementation \\citep{li2022analytical}, PPO, TRPO, MDPO \\citep{tomar2020}, MPO \\citet{abdolmaleki2018} etc.) are functional-gradient methods \\citep{vaswani2021}, in the sense that the lower bound the algorithms optimize are linearizations of the policy performance objective in the direct policy representation. The functional gradient is $d_{\\pi_t}^\\top Q_{\\pi_t}$. In that sense the stationary distribution that the algorithms use is w.r.t. (with respect to) the previous policy (see \\citet{bandhari2019, vaswani2021, sutton2000, agarwal2019}). As they follow the true policy gradient, these methods are convergent \\citep{bhandari21, vaswani2021, xiao2022, johnson2023}, at least their theoretical versions, under adaptive step sizes. The true policy gradient does take into account the gradient through the stationary distribution.\nI did not understand at all the examples they give, which seem to say that “0” has some special meaning when it comes to the reward. \n\nThe text is very unclear, vague, and grammatically incorrect, which makes it very difficult to follow the authors’ arguments.\nIn the motivating introduction, the authors discuss multi-step methods, but this is completely out of context, and use words which have not been defined and are not scientifically rigorous, like “we give a nice theory.”, “the policy can be reflective”, “ this empowers the agent to engage in introspection and introduce modifications to its actions within the current state to a certain degree”, “some admirable algorithms”, “direct optimization of this generalized surrogate objective function may have to be done very carefully”, “RPO efficiently utilizes ‘good’ experiences, makes adjustments based on ‘bad’ experiences’”\n\nWith respect to scientific correctness, at some point the authors completely remove a term from the bound saying that “this term “1” may adversely affect policy optimization. I did not understand that at all. The authors say “for $k = 1$, the $l_1$ norm constraints are replaced by KL constraints”, but this to me makes no sense at all. \n\nThe notation is also confusing, as it seems the authors reuse “r” to represent the importance sampling ratio between the current policy and a prior one. It is unclear what “k” represents as it is not defined, just implied, and seems very important throughout the paper, with multiple reference and impacting the algorithms’ empirical performance.\n\nThen, I also did not really understand the emphasis on the exploration-exploitation coupling on on-policy methods since the whole point of the paper is to improve off-policy methods like PPO and TRPO. These proximal methods are off-policy algorithms, which is the entire point of the efficiency of the methods, the fact that it can reuse prior experience (from the previous policy) multiple times to minimize the lower bound. Otherwise, on-policy policy gradient methods like \\citet{sutton2000} (or algorithmic implementations with parallel streams of experiences like A3C \\citep{mnih16}, Impala \\citep{espeholt18}) directly use the PGT \\citep{sutton2000} w.r.t the parameter vector and rebuild the linear lower bound at each time-step so they don’t need a lower bound surrogate objective that is quasi-concave and can be optimized for multiple updates. \n\n\nThe authors claim the method maintains mototonic improvement, but it is known that proximal-gradient methods, with adaptive step size have this property \\citep{alfano2023, chen2022, xiao2022, johnson2023}, since they are functional-gradient methods, at least for tabular direct or softmax parametrizations.\n\n@misc{li2022analytical,\n      title={An Analytical Update Rule for General Policy Optimization}, \n      author={Hepeng Li and Nicholas Clavette and Haibo He},\n      year={2022},\n      eprint={2112.02045},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI}\n}\n\n@InProceedings{bhandari21,\n  title = \t { On the Linear Convergence of Policy Gradient Methods for Finite MDPs },\n  author =       {Bhandari, Jalaj and Russo, Daniel},\n  booktitle = \t {Proceedings of The 24th International Conference on Artificial Intelligence and Statistics},\n  pages = \t {2386--2394},\n  year = \t {2021},\n  editor = \t {Banerjee, Arindam and Fukumizu, Kenji},\n  volume = \t {130},\n  series = \t {Proceedings of Machine Learning Research},\n  month = \t {13--15 Apr},\n  publisher =    {PMLR},\n  pdf = \t {http://proceedings.mlr.press/v130/bhandari21a/bhandari21a.pdf},\n  url = \t {https://proceedings.mlr.press/v130/bhandari21a.html},\n  abstract = \t { We revisit the finite time analysis of policy gradient methods in the one of the simplest settings: finite state and action MDPs with a policy class consisting of all stochastic policies and with exact gradient evaluations. There has been some recent work viewing this setting as an instance of smooth non-linear optimization problems, to show sub-linear convergence rates with small step-sizes. Here, we take a completely different perspective based on illuminating connections with policy iteration, to show how many variants of policy gradient algorithms succeed with large step-sizes and attain a linear rate of convergence. }\n}\n@article{mnih16,\n  author       = {Volodymyr Mnih and\n                  Adri{\\`{a}} Puigdom{\\`{e}}nech Badia and\n                  Mehdi Mirza and\n                  Alex Graves and\n                  Timothy P. Lillicrap and\n                  Tim Harley and\n                  David Silver and\n                  Koray Kavukcuoglu},\n  title        = {Asynchronous Methods for Deep Reinforcement Learning},\n  journal      = {CoRR},\n  volume       = {abs/1602.01783},\n  year         = {2016},\n  url          = {http://arxiv.org/abs/1602.01783},\n  eprinttype    = {arXiv},\n  eprint       = {1602.01783},\n  timestamp    = {Mon, 13 Aug 2018 16:47:40 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/MnihBMGLHSK16.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@InProceedings{espeholt18,\n  title = \t {{IMPALA}: Scalable Distributed Deep-{RL} with Importance Weighted Actor-Learner Architectures},\n  author =       {Espeholt, Lasse and Soyer, Hubert and Munos, Remi and Simonyan, Karen and Mnih, Vlad and Ward, Tom and Doron, Yotam and Firoiu, Vlad and Harley, Tim and Dunning, Iain and Legg, Shane and Kavukcuoglu, Koray},\n  booktitle = \t {Proceedings of the 35th International Conference on Machine Learning},\n  pages = \t {1407--1416},\n  year = \t {2018},\n  editor = \t {Dy, Jennifer and Krause, Andreas},\n  volume = \t {80},\n  series = \t {Proceedings of Machine Learning Research},\n  month = \t {10--15 Jul},\n  publisher =    {PMLR},\n  pdf = \t {http://proceedings.mlr.press/v80/espeholt18a/espeholt18a.pdf},\n  url = \t {https://proceedings.mlr.press/v80/espeholt18a.html},\n  abstract = \t {In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in single-machine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach.}\n}\n\n@article{vaswani2021,\n  author       = {Sharan Vaswani and\n                  Olivier Bachem and\n                  Simone Totaro and\n                  Robert Mueller and\n                  Matthieu Geist and\n                  Marlos C. Machado and\n                  Pablo Samuel Castro and\n                  Nicolas Le Roux},\n  title        = {A functional mirror ascent view of policy gradient methods with function\n                  approximation},\n  journal      = {CoRR},\n  volume       = {abs/2108.05828},\n  year         = {2021},\n  url          = {https://arxiv.org/abs/2108.05828},\n  eprinttype    = {arXiv},\n  eprint       = {2108.05828},\n  timestamp    = {Wed, 18 Aug 2021 19:45:42 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-2108-05828.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{agarwal2019,\n  author       = {Alekh Agarwal and\n                  Sham M. Kakade and\n                  Jason D. Lee and\n                  Gaurav Mahajan},\n  title        = {Optimality and Approximation with Policy Gradient Methods in Markov\n                  Decision Processes},\n  journal      = {CoRR},\n  volume       = {abs/1908.00261},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1908.00261},\n  eprinttype    = {arXiv},\n  eprint       = {1908.00261},\n  timestamp    = {Fri, 09 Aug 2019 12:15:56 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1908-00261.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{shani2019,\n  author       = {Lior Shani and\n                  Yonathan Efroni and\n                  Shie Mannor},\n  title        = {Adaptive Trust Region Policy Optimization: Global Convergence and\n                  Faster Rates for Regularized MDPs},\n  journal      = {CoRR},\n  volume       = {abs/1909.02769},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1909.02769},\n  eprinttype    = {arXiv},\n  eprint       = {1909.02769},\n  timestamp    = {Mon, 16 Sep 2019 17:27:14 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1909-02769.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{abdolmaleki2018,\n  author       = {Abbas Abdolmaleki and\n                  Jost Tobias Springenberg and\n                  Yuval Tassa and\n                  R{\\'{e}}mi Munos and\n                  Nicolas Heess and\n                  Martin A. Riedmiller},\n  title        = {Maximum a Posteriori Policy Optimisation},\n  journal      = {CoRR},\n  volume       = {abs/1806.06920},\n  year         = {2018},\n  url          = {http://arxiv.org/abs/1806.06920},\n  eprinttype    = {arXiv},\n  eprint       = {1806.06920},\n  timestamp    = {Mon, 13 Aug 2018 16:48:15 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1806-06920.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@inproceedings{sutton2000,\n author = {Sutton, Richard S and McAllester, David and Singh, Satinder and Mansour, Yishay},\n booktitle = {Advances in Neural Information Processing Systems},\n editor = {S. Solla and T. Leen and K. M\\\"{u}ller},\n pages = {},\n publisher = {MIT Press},\n title = {Policy Gradient Methods for Reinforcement Learning with Function Approximation},\n url = {https://proceedings.neurips.cc/paper_files/paper/1999/file/464d828b85b0bed98e80ade0a5c43b0f-Paper.pdf},\n volume = {12},\n year = {1999}\n}\n\n@misc{johnson2023,\n      title={Optimal Convergence Rate for Exact Policy Mirror Descent in Discounted Markov Decision Processes}, \n      author={Emmeran Johnson and Ciara Pike-Burke and Patrick Rebeschini},\n      year={2023},\n      eprint={2302.11381},\n      archivePrefix={arXiv},\n      primaryClass={math.OC}\n}\n@misc{xiao2022,\n      title={On the Convergence Rates of Policy Gradient Methods}, \n      author={Lin Xiao},\n      year={2022},\n      eprint={2201.07443},\n      archivePrefix={arXiv},\n      primaryClass={math.OC}\n}\n@article{tomar2020,\n  author       = {Manan Tomar and\n                  Lior Shani and\n                  Yonathan Efroni and\n                  Mohammad Ghavamzadeh},\n  title        = {Mirror Descent Policy Optimization},\n  journal      = {CoRR},\n  volume       = {abs/2005.09814},\n  year         = {2020},\n  url          = {https://arxiv.org/abs/2005.09814},\n  eprinttype    = {arXiv},\n  eprint       = {2005.09814},\n  timestamp    = {Fri, 22 May 2020 16:21:28 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-2005-09814.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@misc{alfano2023,\n\ttitle        = {Linear Convergence for Natural Policy Gradient with Log-linear Policy Parametrization},\n\tauthor       = {Carlo Alfano and Patrick Rebeschini},\n\tyear         = 2023,\n\teprint       = {2209.15382},\n\tarchiveprefix = {arXiv},\n\tprimaryclass = {cs.LG}\n}\n@inproceedings{chen2022,\n\ttitle        = {Sample Complexity of Policy-Based Methods under Off-Policy Sampling and Linear Function Approximation},\n\tauthor       = {Chen, Zaiwei and Theja Maguluri, Siva},\n\tyear         = 2022,\n\tmonth        = {28--30 Mar},\n\tbooktitle    = {Proceedings of The 25th International Conference on Artificial Intelligence and Statistics},\n\tpublisher    = {PMLR},\n\tseries       = {Proceedings of Machine Learning Research},\n\tvolume       = 151,\n\tpages        = {11195--11214},\n\turl          = {https://proceedings.mlr.press/v151/chen22i.html},\n\teditor       = {Camps-Valls, Gustau and Ruiz, Francisco J. R. and Valera, Isabel},\n\tpdf          = {https://proceedings.mlr.press/v151/chen22i/chen22i.pdf},\n\tabstract     = {In this work, we study policy-based methods for solving the reinforcement learning problem, where off-policy sampling and linear function approximation are employed for policy evaluation, and various policy update rules (including natural policy gradient) are considered for policy improvement. To solve the policy evaluation sub-problem in the presence of the deadly triad, we propose a generic algorithm framework of multi-step TD-learning with generalized importance sampling ratios, which includes two specific algorithms: the $\\lambda$-averaged $Q$-trace and the two-sided $Q$-trace. The generic algorithm is single time-scale, has provable finite-sample guarantees, and overcomes the high variance issue in off-policy learning. As for the policy improvement, we provide a universal analysis that establishes geometric convergence of various policy update rules, which leads to an overall $\\Tilde{\\mathcal{O}}(\\epsilon^{-2})$ sample complexity.}\n}", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method called Reflexive Policy Optimization, which the authors claim uses information “from trajectory data” more efficiently to optimize a policy. The authors claim that agents have “introspection” and “reflect on prior experience”. The method is claimed to have guaranteed monotonic progress improvement on the original policy performance objective and that it “contracts the solution space of the optimized policy”, thus “expediting” the training procedure. Moreover, the authors claim the method is superior to baselines on Mujoco tasks.\n\nI didn’t really understand the motivation for the paper, and I found it very confusing and difficult to follow. I expand in the following sections.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "I believe the authors want to say that their method performs a kind of hindsight credit assignment \\citep{harutyunyan2019}, but this is just a hunch. I didn't really understand from the text.\nThe paper appears to be backed by theory and support their superior performance claims with empirical results. Unfortunately, I did not really understand the motivation and method to assess these  properly. Please see below my points of confusion. I am happy to revise my score if the authors the motivation and goal of the paper, and expand on the points below.\n\n\n@article{harutyunyan2019,\n  author       = {Anna Harutyunyan and\n                  Will Dabney and\n                  Thomas Mesnard and\n                  Mohammad Gheshlaghi Azar and\n                  Bilal Piot and\n                  Nicolas Heess and\n                  Hado van Hasselt and\n                  Greg Wayne and\n                  Satinder Singh and\n                  Doina Precup and\n                  R{\\'{e}}mi Munos},\n  title        = {Hindsight Credit Assignment},\n  journal      = {CoRR},\n  volume       = {abs/1912.02503},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1912.02503},\n  eprinttype    = {arXiv},\n  eprint       = {1912.02503},\n  timestamp    = {Wed, 20 Apr 2022 07:47:18 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1912-02503.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}", "weaknesses": "First, I do not understand why the authors believe proximal methods do not already account for the true gradient of the policy which also considers the contribution through the stationary distribution.  Indeed, in the original CPI paper \\citep{kakade}, the authors first derive the form of the bound that contains the stationary distribution with the current policy, and not a prior policy, then use a mixture policy to ensure that the new policy is “close” to the prior one and justify replacing the current stationary distribution with the previous one. However, a lot of work has been done since, and it is known that proximal PG methods \\citep{bhandari21, shani2019, vaswani2021} (and algorithmic implementation \\citep{li2022analytical}, PPO, TRPO, MDPO \\citep{tomar2020}, MPO \\citet{abdolmaleki2018} etc.) are functional-gradient methods \\citep{vaswani2021}, in the sense that the lower bound the algorithms optimize are linearizations of the policy performance objective in the direct policy representation. The functional gradient is $d_{\\pi_t}^\\top Q_{\\pi_t}$. In that sense the stationary distribution that the algorithms use is w.r.t. (with respect to) the previous policy (see \\citet{bandhari2019, vaswani2021, sutton2000, agarwal2019}). As they follow the true policy gradient, these methods are convergent \\citep{bhandari21, vaswani2021, xiao2022, johnson2023}, at least their theoretical versions, under adaptive step sizes. The true policy gradient does take into account the gradient through the stationary distribution.\nI did not understand at all the examples they give, which seem to say that “0” has some special meaning when it comes to the reward. \n\nThe text is very unclear, vague, and grammatically incorrect, which makes it very difficult to follow the authors’ arguments.\nIn the motivating introduction, the authors discuss multi-step methods, but this is completely out of context, and use words which have not been defined and are not scientifically rigorous, like “we give a nice theory.”, “the policy can be reflective”, “ this empowers the agent to engage in introspection and introduce modifications to its actions within the current state to a certain degree”, “some admirable algorithms”, “direct optimization of this generalized surrogate objective function may have to be done very carefully”, “RPO efficiently utilizes ‘good’ experiences, makes adjustments based on ‘bad’ experiences’”\n\nWith respect to scientific correctness, at some point the authors completely remove a term from the bound saying that “this term “1” may adversely affect policy optimization. I did not understand that at all. The authors say “for $k = 1$, the $l_1$ norm constraints are replaced by KL constraints”, but this to me makes no sense at all. \n\nThe notation is also confusing, as it seems the authors reuse “r” to represent the importance sampling ratio between the current policy and a prior one. It is unclear what “k” represents as it is not defined, just implied, and seems very important throughout the paper, with multiple reference and impacting the algorithms’ empirical performance.\n\nThen, I also did not really understand the emphasis on the exploration-exploitation coupling on on-policy methods since the whole point of the paper is to improve off-policy methods like PPO and TRPO. These proximal methods are off-policy algorithms, which is the entire point of the efficiency of the methods, the fact that it can reuse prior experience (from the previous policy) multiple times to minimize the lower bound. Otherwise, on-policy policy gradient methods like \\citet{sutton2000} (or algorithmic implementations with parallel streams of experiences like A3C \\citep{mnih16}, Impala \\citep{espeholt18}) directly use the PGT \\citep{sutton2000} w.r.t the parameter vector and rebuild the linear lower bound at each time-step so they don’t need a lower bound surrogate objective that is quasi-concave and can be optimized for multiple updates. \n\n\nThe authors claim the method maintains mototonic improvement, but it is known that proximal-gradient methods, with adaptive step size have this property \\citep{alfano2023, chen2022, xiao2022, johnson2023}, since they are functional-gradient methods, at least for tabular direct or softmax parametrizations.\n\n@misc{li2022analytical,\n      title={An Analytical Update Rule for General Policy Optimization}, \n      author={Hepeng Li and Nicholas Clavette and Haibo He},\n      year={2022},\n      eprint={2112.02045},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI}\n}\n\n@InProceedings{bhandari21,\n  title = \t { On the Linear Convergence of Policy Gradient Methods for Finite MDPs },\n  author =       {Bhandari, Jalaj and Russo, Daniel},\n  booktitle = \t {Proceedings of The 24th International Conference on Artificial Intelligence and Statistics},\n  pages = \t {2386--2394},\n  year = \t {2021},\n  editor = \t {Banerjee, Arindam and Fukumizu, Kenji},\n  volume = \t {130},\n  series = \t {Proceedings of Machine Learning Research},\n  month = \t {13--15 Apr},\n  publisher =    {PMLR},\n  pdf = \t {http://proceedings.mlr.press/v130/bhandari21a/bhandari21a.pdf},\n  url = \t {https://proceedings.mlr.press/v130/bhandari21a.html},\n  abstract = \t { We revisit the finite time analysis of policy gradient methods in the one of the simplest settings: finite state and action MDPs with a policy class consisting of all stochastic policies and with exact gradient evaluations. There has been some recent work viewing this setting as an instance of smooth non-linear optimization problems, to show sub-linear convergence rates with small step-sizes. Here, we take a completely different perspective based on illuminating connections with policy iteration, to show how many variants of policy gradient algorithms succeed with large step-sizes and attain a linear rate of convergence. }\n}\n@article{mnih16,\n  author       = {Volodymyr Mnih and\n                  Adri{\\`{a}} Puigdom{\\`{e}}nech Badia and\n                  Mehdi Mirza and\n                  Alex Graves and\n                  Timothy P. Lillicrap and\n                  Tim Harley and\n                  David Silver and\n                  Koray Kavukcuoglu},\n  title        = {Asynchronous Methods for Deep Reinforcement Learning},\n  journal      = {CoRR},\n  volume       = {abs/1602.01783},\n  year         = {2016},\n  url          = {http://arxiv.org/abs/1602.01783},\n  eprinttype    = {arXiv},\n  eprint       = {1602.01783},\n  timestamp    = {Mon, 13 Aug 2018 16:47:40 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/MnihBMGLHSK16.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@InProceedings{espeholt18,\n  title = \t {{IMPALA}: Scalable Distributed Deep-{RL} with Importance Weighted Actor-Learner Architectures},\n  author =       {Espeholt, Lasse and Soyer, Hubert and Munos, Remi and Simonyan, Karen and Mnih, Vlad and Ward, Tom and Doron, Yotam and Firoiu, Vlad and Harley, Tim and Dunning, Iain and Legg, Shane and Kavukcuoglu, Koray},\n  booktitle = \t {Proceedings of the 35th International Conference on Machine Learning},\n  pages = \t {1407--1416},\n  year = \t {2018},\n  editor = \t {Dy, Jennifer and Krause, Andreas},\n  volume = \t {80},\n  series = \t {Proceedings of Machine Learning Research},\n  month = \t {10--15 Jul},\n  publisher =    {PMLR},\n  pdf = \t {http://proceedings.mlr.press/v80/espeholt18a/espeholt18a.pdf},\n  url = \t {https://proceedings.mlr.press/v80/espeholt18a.html},\n  abstract = \t {In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in single-machine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach.}\n}\n\n@article{vaswani2021,\n  author       = {Sharan Vaswani and\n                  Olivier Bachem and\n                  Simone Totaro and\n                  Robert Mueller and\n                  Matthieu Geist and\n                  Marlos C. Machado and\n                  Pablo Samuel Castro and\n                  Nicolas Le Roux},\n  title        = {A functional mirror ascent view of policy gradient methods with function\n                  approximation},\n  journal      = {CoRR},\n  volume       = {abs/2108.05828},\n  year         = {2021},\n  url          = {https://arxiv.org/abs/2108.05828},\n  eprinttype    = {arXiv},\n  eprint       = {2108.05828},\n  timestamp    = {Wed, 18 Aug 2021 19:45:42 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-2108-05828.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{agarwal2019,\n  author       = {Alekh Agarwal and\n                  Sham M. Kakade and\n                  Jason D. Lee and\n                  Gaurav Mahajan},\n  title        = {Optimality and Approximation with Policy Gradient Methods in Markov\n                  Decision Processes},\n  journal      = {CoRR},\n  volume       = {abs/1908.00261},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1908.00261},\n  eprinttype    = {arXiv},\n  eprint       = {1908.00261},\n  timestamp    = {Fri, 09 Aug 2019 12:15:56 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1908-00261.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{shani2019,\n  author       = {Lior Shani and\n                  Yonathan Efroni and\n                  Shie Mannor},\n  title        = {Adaptive Trust Region Policy Optimization: Global Convergence and\n                  Faster Rates for Regularized MDPs},\n  journal      = {CoRR},\n  volume       = {abs/1909.02769},\n  year         = {2019},\n  url          = {http://arxiv.org/abs/1909.02769},\n  eprinttype    = {arXiv},\n  eprint       = {1909.02769},\n  timestamp    = {Mon, 16 Sep 2019 17:27:14 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1909-02769.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@article{abdolmaleki2018,\n  author       = {Abbas Abdolmaleki and\n                  Jost Tobias Springenberg and\n                  Yuval Tassa and\n                  R{\\'{e}}mi Munos and\n                  Nicolas Heess and\n                  Martin A. Riedmiller},\n  title        = {Maximum a Posteriori Policy Optimisation},\n  journal      = {CoRR},\n  volume       = {abs/1806.06920},\n  year         = {2018},\n  url          = {http://arxiv.org/abs/1806.06920},\n  eprinttype    = {arXiv},\n  eprint       = {1806.06920},\n  timestamp    = {Mon, 13 Aug 2018 16:48:15 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-1806-06920.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@inproceedings{sutton2000,\n author = {Sutton, Richard S and McAllester, David and Singh, Satinder and Mansour, Yishay},\n booktitle = {Advances in Neural Information Processing Systems},\n editor = {S. Solla and T. Leen and K. M\\\"{u}ller},\n pages = {},\n publisher = {MIT Press},\n title = {Policy Gradient Methods for Reinforcement Learning with Function Approximation},\n url = {https://proceedings.neurips.cc/paper_files/paper/1999/file/464d828b85b0bed98e80ade0a5c43b0f-Paper.pdf},\n volume = {12},\n year = {1999}\n}\n\n@misc{johnson2023,\n      title={Optimal Convergence Rate for Exact Policy Mirror Descent in Discounted Markov Decision Processes}, \n      author={Emmeran Johnson and Ciara Pike-Burke and Patrick Rebeschini},\n      year={2023},\n      eprint={2302.11381},\n      archivePrefix={arXiv},\n      primaryClass={math.OC}\n}\n@misc{xiao2022,\n      title={On the Convergence Rates of Policy Gradient Methods}, \n      author={Lin Xiao},\n      year={2022},\n      eprint={2201.07443},\n      archivePrefix={arXiv},\n      primaryClass={math.OC}\n}\n@article{tomar2020,\n  author       = {Manan Tomar and\n                  Lior Shani and\n                  Yonathan Efroni and\n                  Mohammad Ghavamzadeh},\n  title        = {Mirror Descent Policy Optimization},\n  journal      = {CoRR},\n  volume       = {abs/2005.09814},\n  year         = {2020},\n  url          = {https://arxiv.org/abs/2005.09814},\n  eprinttype    = {arXiv},\n  eprint       = {2005.09814},\n  timestamp    = {Fri, 22 May 2020 16:21:28 +0200},\n  biburl       = {https://dblp.org/rec/journals/corr/abs-2005-09814.bib},\n  bibsource    = {dblp computer science bibliography, https://dblp.org}\n}\n@misc{alfano2023,\n\ttitle        = {Linear Convergence for Natural Policy Gradient with Log-linear Policy Parametrization},\n\tauthor       = {Carlo Alfano and Patrick Rebeschini},\n\tyear         = 2023,\n\teprint       = {2209.15382},\n\tarchiveprefix = {arXiv},\n\tprimaryclass = {cs.LG}\n}\n@inproceedings{chen2022,\n\ttitle        = {Sample Complexity of Policy-Based Methods under Off-Policy Sampling and Linear Function Approximation},\n\tauthor       = {Chen, Zaiwei and Theja Maguluri, Siva},\n\tyear         = 2022,\n\tmonth        = {28--30 Mar},\n\tbooktitle    = {Proceedings of The 25th International Conference on Artificial Intelligence and Statistics},\n\tpublisher    = {PMLR},\n\tseries       = {Proceedings of Machine Learning Research},\n\tvolume       = 151,\n\tpages        = {11195--11214},\n\turl          = {https://proceedings.mlr.press/v151/chen22i.html},\n\teditor       = {Camps-Valls, Gustau and Ruiz, Francisco J. R. and Valera, Isabel},\n\tpdf          = {https://proceedings.mlr.press/v151/chen22i/chen22i.pdf},\n\tabstract     = {In this work, we study policy-based methods for solving the reinforcement learning problem, where off-policy sampling and linear function approximation are employed for policy evaluation, and various policy update rules (including natural policy gradient) are considered for policy improvement. To solve the policy evaluation sub-problem in the presence of the deadly triad, we propose a generic algorithm framework of multi-step TD-learning with generalized importance sampling ratios, which includes two specific algorithms: the $\\lambda$-averaged $Q$-trace and the two-sided $Q$-trace. The generic algorithm is single time-scale, has provable finite-sample guarantees, and overcomes the high variance issue in off-policy learning. As for the policy improvement, we provide a universal analysis that establishes geometric convergence of various policy update rules, which leads to an overall $\\Tilde{\\mathcal{O}}(\\epsilon^{-2})$ sample complexity.}\n}", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698790624335}, {"id": "GSp4rPiYgj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6789/Reviewer_KkoJ"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper considers policy optimization from the trajectory perspective and proposes to optimize polices with the information from both the current state-action and the subsequent state-action pairs. To achieve this, the paper unfolds the performance improvement lemma over time steps to get the unrolled surrogate objective, i.e., the generalized surrogate, and a “residual” term. The paper then shows the “residual term” can be well bounded, yielding a policy improvement strategy: by iteratively pushing up the lower bound. The paper then use the same clipping scheme from the PPO and forms the Reflective Policy Optimization (RPO). This method has been evaluated on Mujoco control benchmarks and showed good performance.", "review_text": "The paper considers policy optimization from the trajectory perspective and proposes to optimize polices with the information from both the current state-action and the subsequent state-action pairs. To achieve this, the paper unfolds the performance improvement lemma over time steps to get the unrolled surrogate objective, i.e., the generalized surrogate, and a “residual” term. The paper then shows the “residual term” can be well bounded, yielding a policy improvement strategy: by iteratively pushing up the lower bound. The paper then use the same clipping scheme from the PPO and forms the Reflective Policy Optimization (RPO). This method has been evaluated on Mujoco control benchmarks and showed good performance.", "strengths": "### Originality\nThe idea of consider the subsequent state-action pairs in the policy optimization sounds interesting and novel. The theory (if correct) can add new insights into the policy gradient methods when considering unrolling it for a long horizon.", "weaknesses": "### Quality & significance\n**There can be some technical misstatement & errors in the main contributions**.\n\nFirst, the following statement can be erroneous: “If you recombine the above equation $(r_0-1)r_1 A^{\\hat{\\pi}}(s_1, a_1)]\\cdot r_1 >0$, optimizing it will be found to increase the probability of $a_1$ . However, $A_{\\hat{\\pi}}(s_1 , a_1) < 0$, we should decrease the probability of $a_1$ . This would present a contradiction.” **Optimizing this objective can result in a bit more complicated consequence than what’s been stated here**: since the policy in both $r_0$ and $r_1$ is parameterized with a same set of parameters, optimizing the objective may increase the probability of $a_0$ and thus tip over the sign of $r_0-1$.  The problem is $a_0$ and $a_1$ are from the same parametrized policy. \n\n\nSecond, **the monotonic improvement statement of Theorem 4.1 can be erroneous**. Consider a special case of $k=1$, which yields the TRPO bound as suggested by the paper (assume this statement is correct), then the lower bound will be $\\alpha_0 \\hat{L}_0(\\pi, \\hat{\\pi}) - \\hat{C}_1(\\pi, \\hat{\\pi})$, \n\nwhich equals to $E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] - \\frac{\\gamma R_{max}}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$. This is the surrogate objective implied by Theorem 4.1 and should be improved upon. However, **this surrogate objective is strictly negative for $\\gamma=0.995$ used by this paper**: consider the first term\n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] = E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [ (r_0-1) A^{\\hat{\\pi}}(s_0, a_0)] $ (this is because the advantage is defined on $\\hat{\\pi}$ and the integral over $\\hat{\\pi}$ will be zero. Hence, \n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [ (r_0-1) A^{\\hat{\\pi}}(s_0, a_0)] $ \n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} $ (this is a direct result from the proof of Corollary A.2 in appendix, i.e., $G_1$). \n\nThus, \n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] - \\frac{\\gamma R_{max}}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$\n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} - \\frac{\\gamma Rmax}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$\n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} [ 1- \\frac{\\gamma}{(1-\\gamma)^2}] < 0$ (for $\\gamma =0.995)$\n\nSo, the maximum value of the lower bound is strictly negative. **It turns out to be impossible to get \"a monotonically improving sequence of policies $\\{\\pi_i\\}_{i=1}^{\\infty}$ satisfying $\\eta(\\pi_0) \\leq \\eta(\\pi_1) \\leq ...$\" for this case $k=1$**.\n\nPlease correct me if my understanding is wrong. \n\n### Clarity\nThere are some confusing points and unclear definitions. See in my questions.", "questions": "1. I don’t see the point in the examples of “cliff” and “treasure” states in the second paragraph of Introduction Section. Would the value function just suffice whether a state is favourable?\n\n2. Most of the citations are using a wrong format. It seems that the author misused the citep & citet. \n\n3. there is a notation issue in the second line of $\\eta(\\pi) - \\eta(\\hat{\\pi})$. The first term on RHS should be with (s_0, a_0).\n\n4. The $||\\pi - \\hat{\\pi}||_1$ has not been defined in the paper. \n\n5. The definition of the conditional occupancy measure is a bit confusing. From the proofs in the appendix, this conditional occupancy measure is a distribution with the conditions on the initial state-action distribution, i.e., the stationary state distribution induced from a predefined initial state distribution. Then I’m not sure if $(s_0, a_0)\\sim\\rho$ and $(s_i, a_i)\\sim\\rho(\\cdot|s_{i-1}, a_{i-1})$ can be defined coherently. \n\nSome language issues (those do not affect my assessment)\n* “between the performance of \\pi and \\hat{\\pi} from a trajectory-based.”", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers policy optimization from the trajectory perspective and proposes to optimize polices with the information from both the current state-action and the subsequent state-action pairs. To achieve this, the paper unfolds the performance improvement lemma over time steps to get the unrolled surrogate objective, i.e., the generalized surrogate, and a “residual” term. The paper then shows the “residual term” can be well bounded, yielding a policy improvement strategy: by iteratively pushing up the lower bound. The paper then use the same clipping scheme from the PPO and forms the Reflective Policy Optimization (RPO). This method has been evaluated on Mujoco control benchmarks and showed good performance.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "1 poor", "strengths": "### Originality\nThe idea of consider the subsequent state-action pairs in the policy optimization sounds interesting and novel. The theory (if correct) can add new insights into the policy gradient methods when considering unrolling it for a long horizon.", "weaknesses": "### Quality & significance\n**There can be some technical misstatement & errors in the main contributions**.\n\nFirst, the following statement can be erroneous: “If you recombine the above equation $(r_0-1)r_1 A^{\\hat{\\pi}}(s_1, a_1)]\\cdot r_1 >0$, optimizing it will be found to increase the probability of $a_1$ . However, $A_{\\hat{\\pi}}(s_1 , a_1) < 0$, we should decrease the probability of $a_1$ . This would present a contradiction.” **Optimizing this objective can result in a bit more complicated consequence than what’s been stated here**: since the policy in both $r_0$ and $r_1$ is parameterized with a same set of parameters, optimizing the objective may increase the probability of $a_0$ and thus tip over the sign of $r_0-1$.  The problem is $a_0$ and $a_1$ are from the same parametrized policy. \n\n\nSecond, **the monotonic improvement statement of Theorem 4.1 can be erroneous**. Consider a special case of $k=1$, which yields the TRPO bound as suggested by the paper (assume this statement is correct), then the lower bound will be $\\alpha_0 \\hat{L}_0(\\pi, \\hat{\\pi}) - \\hat{C}_1(\\pi, \\hat{\\pi})$, \n\nwhich equals to $E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] - \\frac{\\gamma R_{max}}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$. This is the surrogate objective implied by Theorem 4.1 and should be improved upon. However, **this surrogate objective is strictly negative for $\\gamma=0.995$ used by this paper**: consider the first term\n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] = E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [ (r_0-1) A^{\\hat{\\pi}}(s_0, a_0)] $ (this is because the advantage is defined on $\\hat{\\pi}$ and the integral over $\\hat{\\pi}$ will be zero. Hence, \n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [ (r_0-1) A^{\\hat{\\pi}}(s_0, a_0)] $ \n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} $ (this is a direct result from the proof of Corollary A.2 in appendix, i.e., $G_1$). \n\nThus, \n\n$E_{s_0, a_0\\sim\\rho^{\\hat{\\pi}}} [r_0 A^{\\hat{\\pi}}(s_0, a_0)] - \\frac{\\gamma R_{max}}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$\n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} - \\frac{\\gamma Rmax}{(1-\\gamma)^3}||\\pi - \\hat{\\pi}||_1^2$\n\n$\\leq \\frac{  ||\\pi - \\hat{\\pi}||_1^2 Rmax }{1-\\gamma} [ 1- \\frac{\\gamma}{(1-\\gamma)^2}] < 0$ (for $\\gamma =0.995)$\n\nSo, the maximum value of the lower bound is strictly negative. **It turns out to be impossible to get \"a monotonically improving sequence of policies $\\{\\pi_i\\}_{i=1}^{\\infty}$ satisfying $\\eta(\\pi_0) \\leq \\eta(\\pi_1) \\leq ...$\" for this case $k=1$**.\n\nPlease correct me if my understanding is wrong. \n\n### Clarity\nThere are some confusing points and unclear definitions. See in my questions.", "questions": "1. I don’t see the point in the examples of “cliff” and “treasure” states in the second paragraph of Introduction Section. Would the value function just suffice whether a state is favourable?\n\n2. Most of the citations are using a wrong format. It seems that the author misused the citep & citet. \n\n3. there is a notation issue in the second line of $\\eta(\\pi) - \\eta(\\hat{\\pi})$. The first term on RHS should be with (s_0, a_0).\n\n4. The $||\\pi - \\hat{\\pi}||_1$ has not been defined in the paper. \n\n5. The definition of the conditional occupancy measure is a bit confusing. From the proofs in the appendix, this conditional occupancy measure is a distribution with the conditions on the initial state-action distribution, i.e., the stationary state distribution induced from a predefined initial state distribution. Then I’m not sure if $(s_0, a_0)\\sim\\rho$ and $(s_i, a_i)\\sim\\rho(\\cdot|s_{i-1}, a_{i-1})$ can be defined coherently. \n\nSome language issues (those do not affect my assessment)\n* “between the performance of \\pi and \\hat{\\pi} from a trajectory-based.”", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698761722788}, {"id": "bO69IrF0Q7", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6789/Reviewer_WKkT"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, authors propose a Reflective Policy Optimization(RPO) that considers state-action pairs for n-steps in the policy optimization algorithm. The paper theoretically shows that the relationship between the performance of two policies depends on the next state-action pair and performance’s lower bound can be similar to Trust Region Policy Optimization(TRPO). Authors prove that as n increases, the solution space shrinks and can be contained in the solution space of the (n-1) steps. The practical implementation is based on Proximal Policy Optimization(PPO), and experimental results show that it has better convergence speed and average return performance than the baseline algorithm in cliff-walking and mujoco environments.", "review_text": "In this paper, authors propose a Reflective Policy Optimization(RPO) that considers state-action pairs for n-steps in the policy optimization algorithm. The paper theoretically shows that the relationship between the performance of two policies depends on the next state-action pair and performance’s lower bound can be similar to Trust Region Policy Optimization(TRPO). Authors prove that as n increases, the solution space shrinks and can be contained in the solution space of the (n-1) steps. The practical implementation is based on Proximal Policy Optimization(PPO), and experimental results show that it has better convergence speed and average return performance than the baseline algorithm in cliff-walking and mujoco environments.", "strengths": "1) For two policies, authors propose to maximize a generalized lower bound that directly takes into account that policy performance is related to the next state-action pair. In particular, they interestingly prove that the optimized policy is reflective through some theory and show that TRPO is a special case of this method, which allows the policy to be monotonically improved.\n2) In this paper, authors consider multi-step RL directly from the policy optimization perspective, which is different from the previously proposed multi-step value estimation. In addition, for the practical implementation of the proposed algorithm, the objective function based on PPO was proposed and the correlation between hyperparameters was explained.\n3) In the cliff walking environment, the proposed algorithm was experimentally shown to fall off the cliff less and reach the goal faster than the existing algorithms, which is consistent with the performance expected by the authors at the beginning of the paper. It also improved the average return and convergence speed compared to the existing algorithms in the mujoco environment, which is commonly tested in policy optimization.", "weaknesses": "1) The theorems and proofs proposed in the paper are interesting, but the performance shown in the experiments does not seem to be much different from the performance of the baselines. In particular, there are many mentions of convergence speed, but the results shown do not show a significant improvement over existing methods. It is necessary to show the performance improvement in an environment where the subsequent state-action can be better considered.\n2) The clipping working environment that the authors consider seems a little less relevant to the general field of policy optimization. The proposed environment is more appropriate to be considered in a constrained reinforcement learning or safe reinforcement learning environment where dangerous situations should be avoided. Therefore, the comparison target and baseline algorithm should be the constrained reinforcement learning or safety reinforcement learning algorithm.\n3) In the introduction section of the paper, it is argued that existing policy optimization algorithms do not directly consider the impact of subsequent state-actions in the trajectories. Also, in the example, if the agent repeatedly visits the cliff state, it is dangerous because agent is likely to act to fall off the cliff. However, in policy optimization, since there is a discounted sum of reward term, it is possible to provide feedback on the cliff state and the falling action. Therefore, the motivation of the proposed method is difficult to understand, as it allows us to consider the impact of subsequent state action under the influence of rewards.\n4) The proposed algorithm increases the number of hyperparameters as the value of k increases. The authors claim that k=2,3 is suitable because of the stability of learning, but even with k=3, it requires 5 hyperparameters to adjust the clipping and learning rate. In addition, the experiments show the case of k=2, and the best performing hyperparameters are not consistent across experimental environments. Therefore, it can be said that the algorithm is sensitive to hyperparameters.", "questions": "1) Please answer the questions posed in the weaknesses\n2) In Section 4, a new generalized lower bound is defined via Theorem 4.1. In this part, the authors introduce a new surrogate objective function, which has a slightly different semantics than the surrogate objective function in Theorem 3.1, just minus one. Also, they say that they directly optimize the information of current and future state-action pairs, can you explain what the difference is, or can you compare the difference experimentally?\n3) The paper states that it applied multi-step RL directly from a policy optimization point of view, but it seems that Generalized Adversarial Estimation (GAE) was also used when training. could you tell me if there was any correlation between the GAE parameters and multi-step k in experiments?\n4) The paper shows that the convergence speed increases with the value of K, but the performance is only shown in three environments. Can you explain whether the performance varies significantly depending on the difficulty of the environment or the dimensionality of the state and action?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, authors propose a Reflective Policy Optimization(RPO) that considers state-action pairs for n-steps in the policy optimization algorithm. The paper theoretically shows that the relationship between the performance of two policies depends on the next state-action pair and performance’s lower bound can be similar to Trust Region Policy Optimization(TRPO). Authors prove that as n increases, the solution space shrinks and can be contained in the solution space of the (n-1) steps. The practical implementation is based on Proximal Policy Optimization(PPO), and experimental results show that it has better convergence speed and average return performance than the baseline algorithm in cliff-walking and mujoco environments.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1) For two policies, authors propose to maximize a generalized lower bound that directly takes into account that policy performance is related to the next state-action pair. In particular, they interestingly prove that the optimized policy is reflective through some theory and show that TRPO is a special case of this method, which allows the policy to be monotonically improved.\n2) In this paper, authors consider multi-step RL directly from the policy optimization perspective, which is different from the previously proposed multi-step value estimation. In addition, for the practical implementation of the proposed algorithm, the objective function based on PPO was proposed and the correlation between hyperparameters was explained.\n3) In the cliff walking environment, the proposed algorithm was experimentally shown to fall off the cliff less and reach the goal faster than the existing algorithms, which is consistent with the performance expected by the authors at the beginning of the paper. It also improved the average return and convergence speed compared to the existing algorithms in the mujoco environment, which is commonly tested in policy optimization.", "weaknesses": "1) The theorems and proofs proposed in the paper are interesting, but the performance shown in the experiments does not seem to be much different from the performance of the baselines. In particular, there are many mentions of convergence speed, but the results shown do not show a significant improvement over existing methods. It is necessary to show the performance improvement in an environment where the subsequent state-action can be better considered.\n2) The clipping working environment that the authors consider seems a little less relevant to the general field of policy optimization. The proposed environment is more appropriate to be considered in a constrained reinforcement learning or safe reinforcement learning environment where dangerous situations should be avoided. Therefore, the comparison target and baseline algorithm should be the constrained reinforcement learning or safety reinforcement learning algorithm.\n3) In the introduction section of the paper, it is argued that existing policy optimization algorithms do not directly consider the impact of subsequent state-actions in the trajectories. Also, in the example, if the agent repeatedly visits the cliff state, it is dangerous because agent is likely to act to fall off the cliff. However, in policy optimization, since there is a discounted sum of reward term, it is possible to provide feedback on the cliff state and the falling action. Therefore, the motivation of the proposed method is difficult to understand, as it allows us to consider the impact of subsequent state action under the influence of rewards.\n4) The proposed algorithm increases the number of hyperparameters as the value of k increases. The authors claim that k=2,3 is suitable because of the stability of learning, but even with k=3, it requires 5 hyperparameters to adjust the clipping and learning rate. In addition, the experiments show the case of k=2, and the best performing hyperparameters are not consistent across experimental environments. Therefore, it can be said that the algorithm is sensitive to hyperparameters.", "questions": "1) Please answer the questions posed in the weaknesses\n2) In Section 4, a new generalized lower bound is defined via Theorem 4.1. In this part, the authors introduce a new surrogate objective function, which has a slightly different semantics than the surrogate objective function in Theorem 3.1, just minus one. Also, they say that they directly optimize the information of current and future state-action pairs, can you explain what the difference is, or can you compare the difference experimentally?\n3) The paper states that it applied multi-step RL directly from a policy optimization point of view, but it seems that Generalized Adversarial Estimation (GAE) was also used when training. could you tell me if there was any correlation between the GAE parameters and multi-step k in experiments?\n4) The paper shows that the convergence speed increases with the value of K, but the performance is only shown in three environments. Can you explain whether the performance varies significantly depending on the difficulty of the environment or the dimensionality of the state and action?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698649849527}, {"id": "1TJuXDaykM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6789/Reviewer_X1ce"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposed a new policy optimization method named reflective policy optimization based on expanding surrogate function via condition state-action visitation distributions. This paper also provided a clipped surrogate function(following PPO) for efficient calculating.\nCompared with current policy optimization methods, this reflective policy optimization performs better on many benchmarks.", "review_text": "This paper proposed a new policy optimization method named reflective policy optimization based on expanding surrogate function via condition state-action visitation distributions. This paper also provided a clipped surrogate function(following PPO) for efficient calculating.\nCompared with current policy optimization methods, this reflective policy optimization performs better on many benchmarks.", "strengths": "This paper proposes a novel method with a tighter bound than TRPO. The clipped optimization objective provided by this method is simple and general enough and can be directly applied to many existing methods to replace the PPO objective. It shows improved or comparable performance in the experiments, however again considering the simplicity and generality of the method, this method is valuable.\n\nOverall, the paper is well-written and easy to follow. Adding a detailed expression of the loss function for k=2 before Theorem 4.2 would have made it easier for me to understand.", "weaknesses": "It is better to add some experiments on the environments with discrete action space, e.g. MinAtar[1].\n\nThe gap between the original reflective policy optimization and the clipped version cannot be ignored.\n\nThe method is sensitive to some important hyperparameters, as shown in Figure 4, in Swimmer and Walker2d.\n\n\n[1] Young, K. Tian, T. (2019). MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments. arXiv preprint arXiv:1903.03176.", "questions": "In the ablation experiment for k, when k increases the variance of the objective function will increase significantly, which puts forward a high requirement for fine adjustment of $\\epsilon$ and $\\beta$. So I want to know what is the setting of these hyperparameters in the ablation experiment. How to set such hyperparameters to get a reliable result, choosing the right k?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a new policy optimization method named reflective policy optimization based on expanding surrogate function via condition state-action visitation distributions. This paper also provided a clipped surrogate function(following PPO) for efficient calculating.\nCompared with current policy optimization methods, this reflective policy optimization performs better on many benchmarks.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "This paper proposes a novel method with a tighter bound than TRPO. The clipped optimization objective provided by this method is simple and general enough and can be directly applied to many existing methods to replace the PPO objective. It shows improved or comparable performance in the experiments, however again considering the simplicity and generality of the method, this method is valuable.\n\nOverall, the paper is well-written and easy to follow. Adding a detailed expression of the loss function for k=2 before Theorem 4.2 would have made it easier for me to understand.", "weaknesses": "It is better to add some experiments on the environments with discrete action space, e.g. MinAtar[1].\n\nThe gap between the original reflective policy optimization and the clipped version cannot be ignored.\n\nThe method is sensitive to some important hyperparameters, as shown in Figure 4, in Swimmer and Walker2d.\n\n\n[1] Young, K. Tian, T. (2019). MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments. arXiv preprint arXiv:1903.03176.", "questions": "In the ablation experiment for k, when k increases the variance of the objective function will increase significantly, which puts forward a high requirement for fine adjustment of $\\epsilon$ and $\\beta$. So I want to know what is the setting of these hyperparameters in the ablation experiment. How to set such hyperparameters to get a reliable result, choosing the right k?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698426914199}], "openreview_url": "https://openreview.net/forum?id=OlwW4ZG3Ta", "arxiv_id": "2406.03678", "paper_pdf": "papers/OlwW4ZG3Ta.pdf", "paper_pdf_sha256": "4eddac48353f68653a656e39937acb8bf4975e311a8ed37795ee398175b678cc", "paper_pdf_bytes": 7292594, "paper_pdf_source": "openreview", "code_url": "https://github.com/Edgargan/RPO", "code_repository": "Edgargan/RPO", "code_commit": "3c43816ec410cf396a61fa2cedd83dda4ebc7907", "code_archive": "repos/OlwW4ZG3Ta.zip", "code_archive_sha256": "f3ec2bb286eed2569267182f92eca8cb9d883603a2ba4fe33f2858de3c868cdb", "code_archive_bytes": 49394, "code_file_count": 35, "code_extensions": {".py": 35}, "github_disk_usage_kb": 60, "github_languages": {"Python": 145153}, "github_archived": false, "github_pushed_at": "2024-05-31T03:02:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/reflective-policy-optimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-SBZ8c356Oc", "year": 2023, "status": "rejected", "title": "Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples", "authors": ["Dongyoon Yang", "Insung Kong", "Yongdai Kim"], "authorids": ["~Dongyoon_Yang2", "~Insung_Kong1", "~Yongdai_Kim1"], "authors_source": "OpenReview API", "abstract": "\nAdversarial training, which is to enhance robustness against adversarial attacks, has received much attention because it is easy to generate human-imperceptible perturbations of data to deceive a given deep neural network. In this paper, we propose a new adversarial training algorithm that is theoretically well motivated and empirically superior to other existing algorithms. A novel feature of the proposed algorithm is to apply more regularization to data vulnerable to adversarial attacks than other existing regularization algorithms do. Theoretically, we show that our algorithm can be understood as an algorithm of minimizing a newly derived upper bound of the robust risk. Numerical experiments illustrate that our proposed algorithm improves the generalization (accuracy on examples) and robustness (accuracy on adversarial attacks) simultaneously to achieve the state-of-the-art performance.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rOJDrGHZfe", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6254/Reviewer_E21G"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper a new adversarial training method to improve the robustness of deep learning classifiers in the field of computer vision. To do so, the authors derived a new loss function, which is the surrogate of an upper bound of the robust risk. Specifically, they point out the differences and connections between the proposed method and previous adversarial training method. Experiments on three datasets (CIFAR10, F-MINIST, SVHN) show that the proposed method outperforms other baselines in terms of clean classification accuracy and robust accuracy. Ablation studies are done to show the effect of different parts of the loss function. They also show that the proposed method can be combined with other adversarial training techniques, such as extra data, to further improve the performance. Finally, experiments on CIFAR10 show that the proposed method is helpful to improve the fairness of the classifier compared to TRADES (which is an important baseline).", "review_text": "Overall it's a good paper with with theoretical justification and experimental support. Though, there are some weakness in terms of novelty and the limitation of the datasets used, it is a paper above the margin.", "strengths": "Strength:\n\n- Adversarial robustness is an important security issue in the field of deep learning. The proposed method pushes the SOTA performance of adversarial training to a new level.\n- The proposed loss function is derived with a theoretical support.\n- Extensive experiments show the empirical advantages of the proposed method from different aspects.\n\nWeakness:\n\n- The novelty of the method is ok, but it is similar to previous methods like MART. The author does point out the differences between the proposed method and other methods, so should not be a big problem.\n- It's good that for table 1 and table 2 the results are based on 3 runs with standard errors given, but for most results, the improvements seem marginal.\n- CIFAR10, F-MNIST, SVHN are all relatively small datasets. Does the method also perform well on larger dataset?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper a new adversarial training method to improve the robustness of deep learning classifiers in the field of computer vision. To do so, the authors derived a new loss function, which is the surrogate of an upper bound of the robust risk. Specifically, they point out the differences and connections between the proposed method and previous adversarial training method. Experiments on three datasets (CIFAR10, F-MINIST, SVHN) show that the proposed method outperforms other baselines in terms of clean classification accuracy and robust accuracy. Ablation studies are done to show the effect of different parts of the loss function. They also show that the proposed method can be combined with other adversarial training techniques, such as extra data, to further improve the performance. Finally, experiments on CIFAR10 show that the proposed method is helpful to improve the fairness of the classifier compared to TRADES (which is an important baseline).", "strength_and_weaknesses": "Strength:\n\n- Adversarial robustness is an important security issue in the field of deep learning. The proposed method pushes the SOTA performance of adversarial training to a new level.\n- The proposed loss function is derived with a theoretical support.\n- Extensive experiments show the empirical advantages of the proposed method from different aspects.\n\nWeakness:\n\n- The novelty of the method is ok, but it is similar to previous methods like MART. The author does point out the differences between the proposed method and other methods, so should not be a big problem.\n- It's good that for table 1 and table 2 the results are based on 3 runs with standard errors given, but for most results, the improvements seem marginal.\n- CIFAR10, F-MNIST, SVHN are all relatively small datasets. Does the method also perform well on larger dataset?\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\n\n- The paper is clearly written and easy to follow.\n\nQuality:\n\n- The quality is good. The claims are clear and experiments are solid.\n\nNovelty:\n\n- Somewhat novel though similar to some previous methods.\n\nReproducibility:\n\n- Code and implementation details are provided, so the reproducibility should be good. Though, I did not run the code to test.", "summary_of_the_review": "Overall it's a good paper with with theoretical justification and experimental support. Though, there are some weakness in terms of novelty and the limitation of the datasets used, it is a paper above the margin.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667318671870}, {"id": "DWMKwb6PWc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6254/Reviewer_hmqr"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a simple, effective, and theoretically inspired regularization to enhance the robustness of DNNs agains adversarial attacks.\nExtensive experimental results were carried out showing the effectiveness of the proposed approach in providing robustness enhancements. ", "review_text": "There several aspects that I like about this work such as the theoretical motivation, the extensive experimental evaluation, and the writing.\nHowever, there are two concerns that  I hope to be addressed in the discussion period.", "strengths": "This appear has several strengths:\n\n- The paper is is well motivated and the added regularizer is theoretically inspired.\n\n- The experimental analysis show the consistent improvement of the proposed method over previous art.\n\n- The paper is well-written. Further, the contributions of this work is placed properly within the literature.\n\n- The wide broad of the empirical results shown in this paper covers many interesting aspects such as combining the proposed approach with AWP, and increasing the fairness.\n\n\nThere are few weaknesses that I hope to be addressed during the discussion:\n\n- While the paper is generally well-written, there few parts that require small adjustments. For example,\nIn caption of figure 1: “ We exclude MART from the figures because its performance is too bad”\n\n- Generally, the robustness improvements that ARoW provide is marginal. Would the proposed method improve the state-of-the-art model from [A]?\n\n[A] Uncovering the Limits of Adversarial Training against Norm-Bounded Adversarial Examples, 2021.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes a simple, effective, and theoretically inspired regularization to enhance the robustness of DNNs agains adversarial attacks.\nExtensive experimental results were carried out showing the effectiveness of the proposed approach in providing robustness enhancements. ", "strength_and_weaknesses": "This appear has several strengths:\n\n- The paper is is well motivated and the added regularizer is theoretically inspired.\n\n- The experimental analysis show the consistent improvement of the proposed method over previous art.\n\n- The paper is well-written. Further, the contributions of this work is placed properly within the literature.\n\n- The wide broad of the empirical results shown in this paper covers many interesting aspects such as combining the proposed approach with AWP, and increasing the fairness.\n\n\nThere are few weaknesses that I hope to be addressed during the discussion:\n\n- While the paper is generally well-written, there few parts that require small adjustments. For example,\nIn caption of figure 1: “ We exclude MART from the figures because its performance is too bad”\n\n- Generally, the robustness improvements that ARoW provide is marginal. Would the proposed method improve the state-of-the-art model from [A]?\n\n[A] Uncovering the Limits of Adversarial Training against Norm-Bounded Adversarial Examples, 2021.", "clarity,_quality,_novelty_and_reproducibility": "The paper clearly states its motivation, contributions, and places itself within prior art.", "summary_of_the_review": "There several aspects that I like about this work such as the theoretical motivation, the extensive experimental evaluation, and the writing.\nHowever, there are two concerns that  I hope to be addressed in the discussion period.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667290638758}, {"id": "-huIYeCaYsk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6254/Reviewer_cZx4"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a modification of TRADES, one of the most popular algorithms to obtain adversarially robust classifiers, to improve its performance: in particular, the regularization term to achieve robustness is weighted to penalize more the training examples which are less robust. In the experimental evaluation on several datasets, the proposed method, ARoW, is shown to achieve better robustness that existing methods, while preserving higher standard accuracy.", "review_text": "The proposed method is reasonable and shows promising results. However, clarifications about the discrepancy of the results for some baselines to the original ones are needed, as well as adding the missing baseline.\n\n---\nUpdate after rebuttal\n\nGiven the additional results and clarifications provided during the rebuttal, I increase the initial score to 6.", "strengths": "Strengths\n- The proposed modification of the TRADES loss, while small, is theoretically justified. The paper clearly presents the new scheme and its differences to existing algorithms.\n\n- The experimental results support ARoW in comparison to existing methods. Most of the relevant baselines are included, and several ablation studies are added to analyze the effect of different components of the training algorithm e.g. label smoothing.\n\n- The paper is well written, and it clearly presents the new method, the baselines and the experimental setup. The experiments include different architectures and datasets, and the case of using extra data for training.\n\nWeaknesses\n- The main concern is about the results reported for the baselines, especially for the case of additional data. For example, in Table 3, for the case of ResNet-18, HAT attains 56.40% and 55.44% of robust accuracy with the 500k extra images from 80M-TI and DDPM synthetic images respectively, while it is reported to get, for the same setups, 57.67% and 57.09% on [RobustBench](https://robustbench.github.io/index.html). Similarly, for WRN-28-10 with extra data, the model from [A] achieves 62.76%\tof robust accuracy, higher than any method in Table 3. Then, it is not clear whether the baselines are optimally tuned.\n\n- In general, the improvements over TRADES and HAT in terms of robustness are quite small, although consistent.\n\n[A] https://arxiv.org/abs/2010.03593", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposes a modification of TRADES, one of the most popular algorithms to obtain adversarially robust classifiers, to improve its performance: in particular, the regularization term to achieve robustness is weighted to penalize more the training examples which are less robust. In the experimental evaluation on several datasets, the proposed method, ARoW, is shown to achieve better robustness that existing methods, while preserving higher standard accuracy.", "strength_and_weaknesses": "Strengths\n- The proposed modification of the TRADES loss, while small, is theoretically justified. The paper clearly presents the new scheme and its differences to existing algorithms.\n\n- The experimental results support ARoW in comparison to existing methods. Most of the relevant baselines are included, and several ablation studies are added to analyze the effect of different components of the training algorithm e.g. label smoothing.\n\n- The paper is well written, and it clearly presents the new method, the baselines and the experimental setup. The experiments include different architectures and datasets, and the case of using extra data for training.\n\nWeaknesses\n- The main concern is about the results reported for the baselines, especially for the case of additional data. For example, in Table 3, for the case of ResNet-18, HAT attains 56.40% and 55.44% of robust accuracy with the 500k extra images from 80M-TI and DDPM synthetic images respectively, while it is reported to get, for the same setups, 57.67% and 57.09% on [RobustBench](https://robustbench.github.io/index.html). Similarly, for WRN-28-10 with extra data, the model from [A] achieves 62.76%\tof robust accuracy, higher than any method in Table 3. Then, it is not clear whether the baselines are optimally tuned.\n\n- In general, the improvements over TRADES and HAT in terms of robustness are quite small, although consistent.\n\n[A] https://arxiv.org/abs/2010.03593", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the paper is well written and clearly presents the method and the results.\n\nQuality: the proposed method is well justified, and the set of experiments is reasonable.\n\nNovelty: the modification to the TRADES loss is quite small, but relevant.\n\nReproducibility: sufficient experimental details and code are provided.", "summary_of_the_review": "The proposed method is reasonable and shows promising results. However, clarifications about the discrepancy of the results for some baselines to the original ones are needed, as well as adding the missing baseline.\n\n---\nUpdate after rebuttal\n\nGiven the additional results and clarifications provided during the rebuttal, I increase the initial score to 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666544041622}, {"id": "YcHNygUOr9q", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6254/Reviewer_m348"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work tries to develop an upper bound of the robust risk and design a new algorithm for\nadversarial training called ARoW which minimizes a surrogate version of the developed upper bound.", "review_text": "I believe this paper is not good enough to publish in ICLR for the following reasons.  \n1. (main) Overclaim in theoretical part. Ref weakness 1, 2. I do not think it has sufficient theoretical contribution and its theorem can support the method.\n2. (secondary) Weak experiments.", "strengths": "Strength:\n1. Well-written\n2. Claim both theoretical and empirical contributions\n\nWeakness:\n1. **(1)** I guess Equation A.1 should be '$=$' and Throem 1 (and the proof) should be '$=$' (rather than '$\\le$') for **binary classification problems**. Hence, Throem 1 is just another expression of 'R_rob = R_nat + R_bdy'. The second term in Theorm 1 is just another expression of boundary error rather than the upper bound of boundary error.     \n**(2)** For **multi-classification problems**, I guess Equation A.1 is still $=$ rather than $\\le$ as $x'$ includes $z(x)$. **Could authors clarify why use $\\le$ rather than $=$ here and what's the meaning of Lemma 2?**      \n**(3)** I also think the last line of last equation in Page 13 is $=$ (rather than $\\le$) for binary classification problems, and $\\le$ for **multi-classification problem** is meaningless, for me it seems like **loosing the equation and artificially creating a very loose bound** . I.e., use **($Y\\ne F(z(x))$ is the sufficient and unnecessary condition of $p(Y|z(x))<0.5$ for multi-classification problem)** to bound the equation. This is the only technical part in this theorem, could we call this a new theorem and claim this a theoretical contribution (theorem 1 is even a simple expansion equation for binary case)?  \n**(4) In conclusion, for me, Theorem 1 seems like an equation for binary classification problem and an artificially-created loose bound for multi-classification problem. That is,  \n$R_{bdy}=\\mathbb{E}[\\mathbb{1}(F(X)\\ne F(z(X)))\\mathbb{1}(Y\\ne F(z(X)))] \\le \\mathbb{E}[\\mathbb{1}(F(X)\\ne F(z(X))) \\mathbb{1} (p(Y|z(x))<0.5)]$, which uses a loose and simple bound to bound the original equation. In my opinion, the $\\le$ here is the only technical part in Theorem 1.   \nI cannot get the hidden meaning and theoretical contribution of this bound and think it is essentially meaningless. Could authors clarify what is the specific and profound meaning of Theorem 1?**    \n**Please point out my mistakes if I understand it incorrectly.**\n2. In Sec. 3.2, the term 1 {pθ(Y |z(X)) < 1/2} is replaced by its convex upper bound 2(1 − pθ(Y | \\hat X pgd)). This upper bound is loose, especially in a multi-classification problem (like CIFAR-10 in the experiments), this bound is meaningless as it is too loose.  \n**Thus, for me, the theoretical part in this manuscript seems like a created loose bound plus a created loose bound.**\n3. Weak experiments and only a little improvement, need more empirical resuts (e.g., CIFAR-100, compare with AWP-TRADES)  \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work tries to develop an upper bound of the robust risk and design a new algorithm for\nadversarial training called ARoW which minimizes a surrogate version of the developed upper bound.", "strength_and_weaknesses": "Strength:\n1. Well-written\n2. Claim both theoretical and empirical contributions\n\nWeakness:\n1. **(1)** I guess Equation A.1 should be '$=$' and Throem 1 (and the proof) should be '$=$' (rather than '$\\le$') for **binary classification problems**. Hence, Throem 1 is just another expression of 'R_rob = R_nat + R_bdy'. The second term in Theorm 1 is just another expression of boundary error rather than the upper bound of boundary error.     \n**(2)** For **multi-classification problems**, I guess Equation A.1 is still $=$ rather than $\\le$ as $x'$ includes $z(x)$. **Could authors clarify why use $\\le$ rather than $=$ here and what's the meaning of Lemma 2?**      \n**(3)** I also think the last line of last equation in Page 13 is $=$ (rather than $\\le$) for binary classification problems, and $\\le$ for **multi-classification problem** is meaningless, for me it seems like **loosing the equation and artificially creating a very loose bound** . I.e., use **($Y\\ne F(z(x))$ is the sufficient and unnecessary condition of $p(Y|z(x))<0.5$ for multi-classification problem)** to bound the equation. This is the only technical part in this theorem, could we call this a new theorem and claim this a theoretical contribution (theorem 1 is even a simple expansion equation for binary case)?  \n**(4) In conclusion, for me, Theorem 1 seems like an equation for binary classification problem and an artificially-created loose bound for multi-classification problem. That is,  \n$R_{bdy}=\\mathbb{E}[\\mathbb{1}(F(X)\\ne F(z(X)))\\mathbb{1}(Y\\ne F(z(X)))] \\le \\mathbb{E}[\\mathbb{1}(F(X)\\ne F(z(X))) \\mathbb{1} (p(Y|z(x))<0.5)]$, which uses a loose and simple bound to bound the original equation. In my opinion, the $\\le$ here is the only technical part in Theorem 1.   \nI cannot get the hidden meaning and theoretical contribution of this bound and think it is essentially meaningless. Could authors clarify what is the specific and profound meaning of Theorem 1?**    \n**Please point out my mistakes if I understand it incorrectly.**\n2. In Sec. 3.2, the term 1 {pθ(Y |z(X)) < 1/2} is replaced by its convex upper bound 2(1 − pθ(Y | \\hat X pgd)). This upper bound is loose, especially in a multi-classification problem (like CIFAR-10 in the experiments), this bound is meaningless as it is too loose.  \n**Thus, for me, the theoretical part in this manuscript seems like a created loose bound plus a created loose bound.**\n3. Weak experiments and only a little improvement, need more empirical resuts (e.g., CIFAR-100, compare with AWP-TRADES)  \n", "clarity,_quality,_novelty_and_reproducibility": "The manuscript is clear.", "summary_of_the_review": "I believe this paper is not good enough to publish in ICLR for the following reasons.  \n1. (main) Overclaim in theoretical part. Ref weakness 1, 2. I do not think it has sufficient theoretical contribution and its theorem can support the method.\n2. (secondary) Weak experiments.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666158207144}], "openreview_url": "https://openreview.net/forum?id=-SBZ8c356Oc", "arxiv_id": "2206.03353", "paper_pdf": "papers/-SBZ8c356Oc.pdf", "paper_pdf_sha256": "e547f032e74fb64bd976eca36f7fa2c17d516d8350ef71944bf9c27087b6c0f3", "paper_pdf_bytes": 368162, "paper_pdf_source": "openreview", "code_url": "https://github.com/dyoony/ARoW", "code_repository": "dyoony/ARoW", "code_commit": "0039411e16d1721b28aaf618e075a82cd74624a6", "code_archive": "repos/-SBZ8c356Oc.zip", "code_archive_sha256": "0fe876e06ee9199694b621af4376ca994b32ab92f213f6d92b56f97c8ed54c42", "code_archive_bytes": 129299, "code_file_count": 57, "code_extensions": {".py": 57}, "github_disk_usage_kb": 115, "github_languages": {"Python": 419030, "Jinja": 1720}, "github_archived": false, "github_pushed_at": "2024-09-14T13:20:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-regularization-for-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6PahjGFjVG-", "year": 2022, "status": "rejected", "title": "Secure Distributed Training at Scale", "authors": ["Eduard Gorbunov", "Alexander Borzunov", "Michael Diskin", "Max Ryabinin"], "authorids": ["~Eduard_Gorbunov1", "~Alexander_Borzunov1", "~Michael_Diskin1", "~Max_Ryabinin1"], "authors_source": "OpenReview API", "abstract": "Some of the hardest problems in deep learning can be solved via pooling together computational resources of many independent parties, as is the case for scientific collaborations and volunteer computing. Unfortunately, any single participant in such systems can jeopardize the entire training run by sending incorrect updates, whether deliberately or by mistake. Training in presence of such peers requires specialized distributed training algorithms with Byzantine tolerance. These algorithms often sacrifice efficiency by introducing redundant communication or passing all updates through a trusted server. As a result, it can be infeasible to apply such algorithms to large-scale distributed deep learning, where models can have billions of parameters. In this work, we propose a novel protocol for secure (Byzantine-tolerant) decentralized training that emphasizes communication efficiency. We rigorously analyze this protocol: in particular, we provide theoretical bounds for its resistance against Byzantine and Sybil attacks and show that it has a marginal communication overhead. To demonstrate its practical effectiveness, we conduct large-scale experiments on image classification and language modeling in presence of Byzantine attackers.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "CbD8YgYB0i-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper369/Reviewer_yAgj"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of decentralized training in the presence of Byzantine peers. The paper proposes a Byzantine-tolerant algorithm which uses randomly chosen peers as validators. A validator recomputes the gradient and ‘accuses’ a peer if the computation is not correct. Towards this end, the paper proposes to leverage two primitives from secure Multi-Party Computation (MPC) literature: secure broadcast and multi-party random number generator. For robust aggregation, the paper relies on Centered Clipping proposed recently in [Karimireddy et al. 2020].", "review_text": "While there are a large number of papers on Byzantine robustness in the parameter server setting, there are relatively few papers for the decentralized setting. In this sense, the paper is timely and relevant. However, the paper lacks in giving sufficient details, which makes it difficult to understand the contributions and novelty. Detailed comments are as follows.\n\nMajor comments:\n\n1. How can one ensure that the policy on page 6 where a worker can eliminate one other worker along with themselves, cannot be exploited by Byzantine workers? Can it happen that a Byzantine worker b eliminates an honest worker h, and ensures that another honest worker h’ eliminates b. In this case, (b,h) and (b,h’) will be eliminated. In other words, Byzantine worker b could eliminate two honest workers. It needs a proof that this type of attacks are not possible. Moreover, since the setup is decentralized how consensus happens to determine if a pair of workers to be eliminated. It will be important to give more details.\n\n2. What is the computation and communication complexity at each peer for BTARD? Validators need to recompute the gradients, and in addition, in case if a peer accuses another peer, all other peers need to recompute the gradient. This suggests huge computation overhead. Would this recalculation of gradients outweigh the benefits of parallelism? It is crucial to quantify computation and communication costs explicitly. \n\nMoreover, how much additional communication cost is incurred by using secure broadcast and MPRNG? Appendices A.2.1 and A.2.2 treat these primitives at a very high level. It would be helpful to select a suitable protocol from the literature, and give enough details.\n\n3. In experiments, baseline methods include coordinate-wise median and geometric median, How are these computed in the decentralized setting? It will be important to give more details. \n\n4. How much improvement the proposed BTARD algorithm gives as compared to baseline methods? From Figures 2 and 3, it seems like the improvements are very marginal. Can the authors quantify the improvements in terms of percentage increase in accuracy (or decrease in loss), and add a discussion on comparison?\n\n5. The authors mention that they omit common (and powerful) attacks in [Baruch et al. 2019, Xie et al. 2020, Allen-Zhu et al. 2021] as these attacks are designed to bypass certain checks, which are not used by the proposed algorithm. The proposed algorithm does not use variance and magnitude checks does not mean that these attacks will fail. These attacks were proposed to show inefficiency of several well-known Byzantine robust schemes such as Krum and coordinate-wise median. So, it will be important to test BTARD against some of these attacks.\n\nOther comments:\n\n1. The paper assumes that training is performed on a public dataset, which can be accessed by all peers. It would be helpful to justify this assumption. In the Byzantine robustness literature, many papers consider data parallelism, where data is distributed across peers. In such a setup, it would not be possible to validate by recalculating gradient. It would be helpful to add a discussion on this assumption. \n\n2. The paper does not formally introduce the decentralized training setup being considered. How many peers in total, and how many are Byzantine? Can any peer talk with any other peer? Can Byzantine peers collude? A few details are given in Sec. 3.2 after the proposed methods. It will be important to describe these details up front.\n\n3. The authors do not give any description or details about Algorithms 1-3 in the main text. It is difficult to understand the algorithms, and takes a long time for the reader. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of decentralized training in the presence of Byzantine peers. The paper proposes a Byzantine-tolerant algorithm which uses randomly chosen peers as validators. A validator recomputes the gradient and ‘accuses’ a peer if the computation is not correct. Towards this end, the paper proposes to leverage two primitives from secure Multi-Party Computation (MPC) literature: secure broadcast and multi-party random number generator. For robust aggregation, the paper relies on Centered Clipping proposed recently in [Karimireddy et al. 2020].", "main_review": "While there are a large number of papers on Byzantine robustness in the parameter server setting, there are relatively few papers for the decentralized setting. In this sense, the paper is timely and relevant. However, the paper lacks in giving sufficient details, which makes it difficult to understand the contributions and novelty. Detailed comments are as follows.\n\nMajor comments:\n\n1. How can one ensure that the policy on page 6 where a worker can eliminate one other worker along with themselves, cannot be exploited by Byzantine workers? Can it happen that a Byzantine worker b eliminates an honest worker h, and ensures that another honest worker h’ eliminates b. In this case, (b,h) and (b,h’) will be eliminated. In other words, Byzantine worker b could eliminate two honest workers. It needs a proof that this type of attacks are not possible. Moreover, since the setup is decentralized how consensus happens to determine if a pair of workers to be eliminated. It will be important to give more details.\n\n2. What is the computation and communication complexity at each peer for BTARD? Validators need to recompute the gradients, and in addition, in case if a peer accuses another peer, all other peers need to recompute the gradient. This suggests huge computation overhead. Would this recalculation of gradients outweigh the benefits of parallelism? It is crucial to quantify computation and communication costs explicitly. \n\nMoreover, how much additional communication cost is incurred by using secure broadcast and MPRNG? Appendices A.2.1 and A.2.2 treat these primitives at a very high level. It would be helpful to select a suitable protocol from the literature, and give enough details.\n\n3. In experiments, baseline methods include coordinate-wise median and geometric median, How are these computed in the decentralized setting? It will be important to give more details. \n\n4. How much improvement the proposed BTARD algorithm gives as compared to baseline methods? From Figures 2 and 3, it seems like the improvements are very marginal. Can the authors quantify the improvements in terms of percentage increase in accuracy (or decrease in loss), and add a discussion on comparison?\n\n5. The authors mention that they omit common (and powerful) attacks in [Baruch et al. 2019, Xie et al. 2020, Allen-Zhu et al. 2021] as these attacks are designed to bypass certain checks, which are not used by the proposed algorithm. The proposed algorithm does not use variance and magnitude checks does not mean that these attacks will fail. These attacks were proposed to show inefficiency of several well-known Byzantine robust schemes such as Krum and coordinate-wise median. So, it will be important to test BTARD against some of these attacks.\n\nOther comments:\n\n1. The paper assumes that training is performed on a public dataset, which can be accessed by all peers. It would be helpful to justify this assumption. In the Byzantine robustness literature, many papers consider data parallelism, where data is distributed across peers. In such a setup, it would not be possible to validate by recalculating gradient. It would be helpful to add a discussion on this assumption. \n\n2. The paper does not formally introduce the decentralized training setup being considered. How many peers in total, and how many are Byzantine? Can any peer talk with any other peer? Can Byzantine peers collude? A few details are given in Sec. 3.2 after the proposed methods. It will be important to describe these details up front.\n\n3. The authors do not give any description or details about Algorithms 1-3 in the main text. It is difficult to understand the algorithms, and takes a long time for the reader. \n", "summary_of_the_review": "The paper considers a challenging problem of Byzantine-tolerant learning in the decentralized setup (without a parameter server). However, the paper lacks in many details and does not quantify costs and gains of the proposed algorithm.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636678556003}, {"id": "Sjng72CqEct", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper369/Reviewer_w1BH"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper develops a byzantine resilient algorithm for distributed training via SGD without a trusted centralized node. The key idea is that it improves upon recent decentralized algorithms by using cryptographic primitives. In particular, a multiparty random number generator is used to randomly verify subsets of devices in each iteration, enabling the algorithm to learn byzantine behavior that deviates sufficiently from the protocol. The paper provides rigorous theoretical result in terms of convergence guarantee, with a rate that depends on the number of byzantine adversaries. The derived rate depends on the fraction of nodes that are byzantine (capped by 50%, which is standard in this literature). In cases where no node is adversarial, the derived convergence rate matches that of the standard algorithms.", "review_text": "+The paper develops a careful protocol utilizing cryptographic primitives (MPC), unlike similar works that mainly aim to eliminate byzantine behavior by comparing the gradients provided by various nodes. The utilization of  MPC is novel, and the provable convergence guarantee is a strength of the paper. \n+/- While the paper's protocols work only in the case where nodes are able to sample i.i.d. from the overall data set (the \"homogenous\" setting), the paper refers to recent lower bounds that preclude byzantine tolerant algorithms in the heterogenous setting (at least in the worst-case).   \n- The implementation of the cryptographic primitives to keep the byzantine behavior in check requires a number of all-to-all communication exchange. While the lack of a centralized trusted node is expected to make the protocols more expensive, a key missing ingredient of the paper is an explicit accounting for this communication cost. I believe it will greatly improve the contribution if this cost is explicitly characterized and reported.\n- The algorithm explanation could improve in terms of clarity. For instance, the paper refer's to a butterfly all reduce procedure, which seems to suggest that each node receives the sum of all the numbers (gradients) that are transmitted in the step. However, the descriptions in Algorithms 1,2, only include \"Send\" and \"Receive\" primitives, and the algorithms in the appendix do not seem to do aggregation as a part of the communication primitives - it is done separately and explicitly in the centeredclip procedure. This can be explained more clearly. Note that the peer-to-peer communication (i.e., n^2 different messages being sent) also has an impact on the communication cost (referred to above). \n- The paper seems to provide distinct converge behavior for the case where the number of byzantine adversaries is explicitly known. A clear/intuitive explanation of why the behavior is distinct can enhance the paper.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper develops a byzantine resilient algorithm for distributed training via SGD without a trusted centralized node. The key idea is that it improves upon recent decentralized algorithms by using cryptographic primitives. In particular, a multiparty random number generator is used to randomly verify subsets of devices in each iteration, enabling the algorithm to learn byzantine behavior that deviates sufficiently from the protocol. The paper provides rigorous theoretical result in terms of convergence guarantee, with a rate that depends on the number of byzantine adversaries. The derived rate depends on the fraction of nodes that are byzantine (capped by 50%, which is standard in this literature). In cases where no node is adversarial, the derived convergence rate matches that of the standard algorithms.", "main_review": "+The paper develops a careful protocol utilizing cryptographic primitives (MPC), unlike similar works that mainly aim to eliminate byzantine behavior by comparing the gradients provided by various nodes. The utilization of  MPC is novel, and the provable convergence guarantee is a strength of the paper. \n+/- While the paper's protocols work only in the case where nodes are able to sample i.i.d. from the overall data set (the \"homogenous\" setting), the paper refers to recent lower bounds that preclude byzantine tolerant algorithms in the heterogenous setting (at least in the worst-case).   \n- The implementation of the cryptographic primitives to keep the byzantine behavior in check requires a number of all-to-all communication exchange. While the lack of a centralized trusted node is expected to make the protocols more expensive, a key missing ingredient of the paper is an explicit accounting for this communication cost. I believe it will greatly improve the contribution if this cost is explicitly characterized and reported.\n- The algorithm explanation could improve in terms of clarity. For instance, the paper refer's to a butterfly all reduce procedure, which seems to suggest that each node receives the sum of all the numbers (gradients) that are transmitted in the step. However, the descriptions in Algorithms 1,2, only include \"Send\" and \"Receive\" primitives, and the algorithms in the appendix do not seem to do aggregation as a part of the communication primitives - it is done separately and explicitly in the centeredclip procedure. This can be explained more clearly. Note that the peer-to-peer communication (i.e., n^2 different messages being sent) also has an impact on the communication cost (referred to above). \n- The paper seems to provide distinct converge behavior for the case where the number of byzantine adversaries is explicitly known. A clear/intuitive explanation of why the behavior is distinct can enhance the paper.  \n", "summary_of_the_review": "Overall solid piece of work. It is worth considering the paper modulo addressing the comments above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636425105265}, {"id": "hhU6-odV74A", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper369/Reviewer_63QJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes BTARD-SGD, which is a Byzantine-tolerant training algorithm (via stochastic gradient descent) in decentralized environments (i.e., peer-to-peer architectures). BTARD-SGD does not require the presence of a trusted central server and introduces only a marginal overhead compared to a vanilla All-Reduce SGD algorithm. BTARD-SGD relies on an existing clipping algorithm in the literature called CenteredClip in addition to a series of checks to catch Byzantine nodes which might not follow the algorithm. The authors provide theoretical analysis in addition to a fair amount of experiments.", "review_text": "Thanks for submitting your work to ICLR2022. The paper addresses a very important and timely problem. While there is a lot of work addressing Byzantine resilience in ML, we need works that focus on their practicality in terms of performance and assumptions about the environment. I also like the motivating example of crowdsourcing computational resources from (possibly) untrusted parties. I appreciate the convergence rates that the authors derive in different scenarios (strongly convex, convex, and non-convex functions), and I like the fact that the overhead of the proposed algorithm does not depend on the model size. We, as a community, need to always think of such solutions (whose overhead is decoupled from the model size).\n\nYet, I believe the paper cannot be accepted as-is as it requires to be revised in a few aspects. In summary, (1) the authors make several assumptions that seem unrealistic and unjustified. (2) The most important parts of the paper are either unclear or deferred to the appendices; this makes it hard to understand the algorithm and to see its benefits. (3) As the main aim of the paper is to provide a computationally efficient algorithm, the paper should give clearly the computation complexity of all parts of the algorithm.\n\nThe first area of improvement is to clarify and back up the assumptions. Specifically, I'd like to point out 3 assumptions: The first is the public dataset assumption. Doesn't that make the problem trivial? Can you give examples in which this assumption is practical? Furthermore, many parts of the main algorithm require synchronous networks. It's important that the authors clearly state this in their assumptions. Synchronous networks are easy to achieve in a controlled environment like a datacenter. Yet, in a more challenging environment that the authors consider, I'm not sure how synchronicity can be achieved. The third uncommon assumption is about adversary capabilities. I do not understand the need for the second paragraph on page 6 (which describes the possible attacks). Byzantines should be able to do whatever they want [1,2]. Putting them in categories in which they can do only limited things raises the question of whether the considered model is really for Byzantine attacks or not.\n[1] Lamport, Leslie. \"The weak Byzantine generals problem.\" Journal of the ACM (JACM) 30.3 (1983): 668-676.\n[2] Castro, Miguel, and Barbara Liskov. \"Practical byzantine fault tolerance.\" OSDI. Vol. 99. No. 1999. 1999.\n\nThere are essential parts in the paper that are not described clearly and left to the appendices. One example is how the broadcast channels work in Section 2.3. The authors seem to describe consensus. Is that it? If yes, this concept is orthogonal to multi-party computation (MPC). If not, then what is the difference between the suggested method and consensus? The same comment goes for the main algorithm in the paper. Essentially, there is no appropriate (even high-level, intuitive) explanation of why it works, with all the proofs deferred to the appendices. The most important example is the CenteredClip algorithm, which is the main tool to tolerate the Byzantines. For example (more examples below in the additional comments/questions), from the description, I understand that the input to this algorithm is two gradients: the local one and the one gathered from all peers (by merging a small part of the gradient from each peer). If my understanding is correct, I do not see how clipping (based on two inputs only) can be effective in tolerating Byzantine behavior. It's crucial to understand at least intuition without looking at complex mathematical proofs. I also expected to see even proof sketches in Sections 3.2 and 3.3. The absence of these parts makes it very hard to judge the effectiveness of the proposed work.\n\nParts of the algorithm are not clear and the computation complexity is not rigorously quantified. For example, what is the cost of the validation step? If we need a lot of validators, why do we need distributed learning in the first place? If peers can compute their gradients and the others' gradients (for validation), why bother distributing? In addition, I believe many parts of the algorithm rely on consensus (even if that was never explicitly mentioned by the authors). Consensus is costly, and a careful analysis of its complexity is required. Along the same direction, I have two issues with the \"accuse\" procedure: (1) it's very costly because all peers have to compute the gradient of the accused peer; a Byzantine can use this trick to make learning very slow. (2) The procedure relies on consensus, which is also costly. Another unclear part is the cost of sampling the random direction z among all peers.\n\nAdditional comments/questions:\n1. The authors mentioned that existing solutions either require a trusted server or have prohibitive communication costs. I know at least one paper [3] that does not require either. Can the author comment on that?\n[3] El-Mhamdi, E. M., Guerraoui, R., Guirguis, A., Hoang, L. N., & Rouault, S. (2020, July). Genuinely distributed byzantine machine learning. In Proceedings of the 39th Symposium on Principles of Distributed Computing (pp. 355-364).\n\n2. The third paragraph on page 6 is problematic. The authors mentioned an attack in which the system removes one Byzantine and one benign node. In this case, the Byzantines can do this attack until 2f nodes are removed (with f being the original number of Byzantines). Since the algorithm requires n > 2f+1 then, in extreme cases, this attack leaves the system with only 1 node! In other words, the authors seem to acknowledge that a Byzantine node can defeat the distribution of the algorithm, making it behave as if there is only one central node.\n\n3. I do not understand the last line of Section 2.1. Why does **each** peer require O(dn) in the PS architecture? I believe each one requires only 2d: 1d to upload the gradient and 1d to download the model. Can you please clarify this?\n\n4. The first sentence of Section 3.1 (\"The core design principle behind our algorithm is that all types of Byzantine faults must have limited effect and a chance of being discovered.\") is a bit weird. Usually, Byzantine attacks are posed as the strongest attack to diverge learning. It is strong because Byzantine peers can do whatever they want (i.e., it's also called: behave arbitrarily).\n\n5. It's not clear if clipping is effective against random attacks? I understand it will be robust against gradients with big magnitudes (like the attacks in the evaluation section) but what about the other attacks?\n\n6. It's not clear how the butterfly method works? If you only broadcast a part of the gradient, then what is aggregated in this case? Is it only the local part of a peer and a corresponding part received from another peer?\n\n7. In Algorithm 2, a Byzantine can equivocate as follows. It can send different hashes (for the same part of the gradient) to different peers. It can then send different values for that part to different peers so that the hashes match the received gradients on each peer. Is that a problem for your algorithm?\n\n8. The validation step is not rigorous because you might be unlucky and get the Byzantine node to validate another Byzantine node in one iteration. Clearly, Byzantines can get away with their attack (without being caught) in this iteration. What are the practical/theoretical results of such a situation?\n\n9. The solution to Sybil attacks (in Section 3.2) is not convincing. A Byzantine node can prove its honesty only to enter the system and then it can do whatever it wants. This problem (i.e., tolerating Sybil attacks) is interesting and I encourage the authors to give it more discussion in the main paper.\n\n10. The authors seem to rely on (Karimireddy et al., 2020) a lot. What are the main differences between this work compared to this work? Is it only PS architecture vs. decentralized architecture? If so, what are the additional challenges that you solved?\n\n11. The federated learning (FL) example in the motivation does not fit nicely. As mentioned in the paper, one important aspect of FL is data privacy yet, the solution in the paper relies on the assumption that the dataset is public and can be accessible by anybody. I advise the authors to remove FL from their motivating examples.\n\n12. The authors chose the clipping value (\\tau) based on out-of-band experiments. Can the authors add a discussion on how to choose a clipping value practically (without initial experiments)?\n\n13. What are the coordinate-wise trimmed median and geometric median baselines? It's clear how to use these aggregators with a central server yet, they are not well-defined in the decentralized case. If the authors are referring to some work in the literature, please add the necessary citations.\n\n14. As the authors mentioned, they use only specific parts of MPC rather than a fully-fledged MPC protocol. I suggest the authors do not use this term and use specific terms of the specific algorithms they use instead.\n\n15. The expression \"zero trust assumption\" is a bit of a stretch. For example, you need at least 50% of honest peers.\n\n16. What is the percentage of the validators (i.e., what is the inequality between m and n)? How does this percentage affect the correctness of the algorithm?\n\n17. The authors mentioned that m validators will validate the work of m peers. It seems from the evaluation section that there is a one-to-one mapping, in this case, i.e., each validator validates one peer. Please make this clear while describing the algorithm.\n\n18. How can one report a Byzantine node? This again requires consensus, right?\n\n19. It will be great if the authors can add to the paper an explanation of how CenteredClip works.\n\n20. I suggest you better explain \\tau in Equation (2). Currently, it is not clear in the text; I understood what it means only in the evaluation section.\n\n21. In Algorithm 2, Step 7: you either mean \"broadcast\" instead of \"send\" or you mean g_i[j] instead of g_i. The same fro Step 8. I believe you mean the latter though (given Step 9 is \"MERGE\" rather than \"AGGREGATE\").\n\n22. A few important function definitions are missing: CenteredClip, ValidatePeer, VoteForBan, and Ban. Other functions are also missing, but they are less important.\n\n23. Footnote 4 is a bit off because the paper does not consider the heterogeneous case.\n\n24. I do not buy the omission of \"common low-magnitude attacks (Baruch et al., 2019; Xie et al., 2020; Allen-Zhu et al., 2021)\". These are the SOTA attacks, and it is important to show how your algorithm behaves under these attacks.\n\n25. What is the attack in Figure 2 (first two figures)?\n\n26. The upper-middle plot of Figure 2 shows that the geometric median is almost as good as your algorithm. What is the benefit of your algorithm in this case?\n\n27. Why does the coordinate-wise median diverge in the upper-left figure in Fig. 2?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes BTARD-SGD, which is a Byzantine-tolerant training algorithm (via stochastic gradient descent) in decentralized environments (i.e., peer-to-peer architectures). BTARD-SGD does not require the presence of a trusted central server and introduces only a marginal overhead compared to a vanilla All-Reduce SGD algorithm. BTARD-SGD relies on an existing clipping algorithm in the literature called CenteredClip in addition to a series of checks to catch Byzantine nodes which might not follow the algorithm. The authors provide theoretical analysis in addition to a fair amount of experiments.", "main_review": "Thanks for submitting your work to ICLR2022. The paper addresses a very important and timely problem. While there is a lot of work addressing Byzantine resilience in ML, we need works that focus on their practicality in terms of performance and assumptions about the environment. I also like the motivating example of crowdsourcing computational resources from (possibly) untrusted parties. I appreciate the convergence rates that the authors derive in different scenarios (strongly convex, convex, and non-convex functions), and I like the fact that the overhead of the proposed algorithm does not depend on the model size. We, as a community, need to always think of such solutions (whose overhead is decoupled from the model size).\n\nYet, I believe the paper cannot be accepted as-is as it requires to be revised in a few aspects. In summary, (1) the authors make several assumptions that seem unrealistic and unjustified. (2) The most important parts of the paper are either unclear or deferred to the appendices; this makes it hard to understand the algorithm and to see its benefits. (3) As the main aim of the paper is to provide a computationally efficient algorithm, the paper should give clearly the computation complexity of all parts of the algorithm.\n\nThe first area of improvement is to clarify and back up the assumptions. Specifically, I'd like to point out 3 assumptions: The first is the public dataset assumption. Doesn't that make the problem trivial? Can you give examples in which this assumption is practical? Furthermore, many parts of the main algorithm require synchronous networks. It's important that the authors clearly state this in their assumptions. Synchronous networks are easy to achieve in a controlled environment like a datacenter. Yet, in a more challenging environment that the authors consider, I'm not sure how synchronicity can be achieved. The third uncommon assumption is about adversary capabilities. I do not understand the need for the second paragraph on page 6 (which describes the possible attacks). Byzantines should be able to do whatever they want [1,2]. Putting them in categories in which they can do only limited things raises the question of whether the considered model is really for Byzantine attacks or not.\n[1] Lamport, Leslie. \"The weak Byzantine generals problem.\" Journal of the ACM (JACM) 30.3 (1983): 668-676.\n[2] Castro, Miguel, and Barbara Liskov. \"Practical byzantine fault tolerance.\" OSDI. Vol. 99. No. 1999. 1999.\n\nThere are essential parts in the paper that are not described clearly and left to the appendices. One example is how the broadcast channels work in Section 2.3. The authors seem to describe consensus. Is that it? If yes, this concept is orthogonal to multi-party computation (MPC). If not, then what is the difference between the suggested method and consensus? The same comment goes for the main algorithm in the paper. Essentially, there is no appropriate (even high-level, intuitive) explanation of why it works, with all the proofs deferred to the appendices. The most important example is the CenteredClip algorithm, which is the main tool to tolerate the Byzantines. For example (more examples below in the additional comments/questions), from the description, I understand that the input to this algorithm is two gradients: the local one and the one gathered from all peers (by merging a small part of the gradient from each peer). If my understanding is correct, I do not see how clipping (based on two inputs only) can be effective in tolerating Byzantine behavior. It's crucial to understand at least intuition without looking at complex mathematical proofs. I also expected to see even proof sketches in Sections 3.2 and 3.3. The absence of these parts makes it very hard to judge the effectiveness of the proposed work.\n\nParts of the algorithm are not clear and the computation complexity is not rigorously quantified. For example, what is the cost of the validation step? If we need a lot of validators, why do we need distributed learning in the first place? If peers can compute their gradients and the others' gradients (for validation), why bother distributing? In addition, I believe many parts of the algorithm rely on consensus (even if that was never explicitly mentioned by the authors). Consensus is costly, and a careful analysis of its complexity is required. Along the same direction, I have two issues with the \"accuse\" procedure: (1) it's very costly because all peers have to compute the gradient of the accused peer; a Byzantine can use this trick to make learning very slow. (2) The procedure relies on consensus, which is also costly. Another unclear part is the cost of sampling the random direction z among all peers.\n\nAdditional comments/questions:\n1. The authors mentioned that existing solutions either require a trusted server or have prohibitive communication costs. I know at least one paper [3] that does not require either. Can the author comment on that?\n[3] El-Mhamdi, E. M., Guerraoui, R., Guirguis, A., Hoang, L. N., & Rouault, S. (2020, July). Genuinely distributed byzantine machine learning. In Proceedings of the 39th Symposium on Principles of Distributed Computing (pp. 355-364).\n\n2. The third paragraph on page 6 is problematic. The authors mentioned an attack in which the system removes one Byzantine and one benign node. In this case, the Byzantines can do this attack until 2f nodes are removed (with f being the original number of Byzantines). Since the algorithm requires n > 2f+1 then, in extreme cases, this attack leaves the system with only 1 node! In other words, the authors seem to acknowledge that a Byzantine node can defeat the distribution of the algorithm, making it behave as if there is only one central node.\n\n3. I do not understand the last line of Section 2.1. Why does **each** peer require O(dn) in the PS architecture? I believe each one requires only 2d: 1d to upload the gradient and 1d to download the model. Can you please clarify this?\n\n4. The first sentence of Section 3.1 (\"The core design principle behind our algorithm is that all types of Byzantine faults must have limited effect and a chance of being discovered.\") is a bit weird. Usually, Byzantine attacks are posed as the strongest attack to diverge learning. It is strong because Byzantine peers can do whatever they want (i.e., it's also called: behave arbitrarily).\n\n5. It's not clear if clipping is effective against random attacks? I understand it will be robust against gradients with big magnitudes (like the attacks in the evaluation section) but what about the other attacks?\n\n6. It's not clear how the butterfly method works? If you only broadcast a part of the gradient, then what is aggregated in this case? Is it only the local part of a peer and a corresponding part received from another peer?\n\n7. In Algorithm 2, a Byzantine can equivocate as follows. It can send different hashes (for the same part of the gradient) to different peers. It can then send different values for that part to different peers so that the hashes match the received gradients on each peer. Is that a problem for your algorithm?\n\n8. The validation step is not rigorous because you might be unlucky and get the Byzantine node to validate another Byzantine node in one iteration. Clearly, Byzantines can get away with their attack (without being caught) in this iteration. What are the practical/theoretical results of such a situation?\n\n9. The solution to Sybil attacks (in Section 3.2) is not convincing. A Byzantine node can prove its honesty only to enter the system and then it can do whatever it wants. This problem (i.e., tolerating Sybil attacks) is interesting and I encourage the authors to give it more discussion in the main paper.\n\n10. The authors seem to rely on (Karimireddy et al., 2020) a lot. What are the main differences between this work compared to this work? Is it only PS architecture vs. decentralized architecture? If so, what are the additional challenges that you solved?\n\n11. The federated learning (FL) example in the motivation does not fit nicely. As mentioned in the paper, one important aspect of FL is data privacy yet, the solution in the paper relies on the assumption that the dataset is public and can be accessible by anybody. I advise the authors to remove FL from their motivating examples.\n\n12. The authors chose the clipping value (\\tau) based on out-of-band experiments. Can the authors add a discussion on how to choose a clipping value practically (without initial experiments)?\n\n13. What are the coordinate-wise trimmed median and geometric median baselines? It's clear how to use these aggregators with a central server yet, they are not well-defined in the decentralized case. If the authors are referring to some work in the literature, please add the necessary citations.\n\n14. As the authors mentioned, they use only specific parts of MPC rather than a fully-fledged MPC protocol. I suggest the authors do not use this term and use specific terms of the specific algorithms they use instead.\n\n15. The expression \"zero trust assumption\" is a bit of a stretch. For example, you need at least 50% of honest peers.\n\n16. What is the percentage of the validators (i.e., what is the inequality between m and n)? How does this percentage affect the correctness of the algorithm?\n\n17. The authors mentioned that m validators will validate the work of m peers. It seems from the evaluation section that there is a one-to-one mapping, in this case, i.e., each validator validates one peer. Please make this clear while describing the algorithm.\n\n18. How can one report a Byzantine node? This again requires consensus, right?\n\n19. It will be great if the authors can add to the paper an explanation of how CenteredClip works.\n\n20. I suggest you better explain \\tau in Equation (2). Currently, it is not clear in the text; I understood what it means only in the evaluation section.\n\n21. In Algorithm 2, Step 7: you either mean \"broadcast\" instead of \"send\" or you mean g_i[j] instead of g_i. The same fro Step 8. I believe you mean the latter though (given Step 9 is \"MERGE\" rather than \"AGGREGATE\").\n\n22. A few important function definitions are missing: CenteredClip, ValidatePeer, VoteForBan, and Ban. Other functions are also missing, but they are less important.\n\n23. Footnote 4 is a bit off because the paper does not consider the heterogeneous case.\n\n24. I do not buy the omission of \"common low-magnitude attacks (Baruch et al., 2019; Xie et al., 2020; Allen-Zhu et al., 2021)\". These are the SOTA attacks, and it is important to show how your algorithm behaves under these attacks.\n\n25. What is the attack in Figure 2 (first two figures)?\n\n26. The upper-middle plot of Figure 2 shows that the geometric median is almost as good as your algorithm. What is the benefit of your algorithm in this case?\n\n27. Why does the coordinate-wise median diverge in the upper-left figure in Fig. 2?", "summary_of_the_review": "Though the paper addresses a very important and timely problem, I believe it cannot be accepted as-is as it requires to be revised in a few aspects. In summary, (1) the authors make several assumptions that seem unrealistic and unjustified. (2) The most important parts of the paper are deferred to the appendices; this makes it hard to understand the algorithm and to see its benefits. (3) As the main aim of the paper is to provide a computationally efficient algorithm, the paper should give clearly the computation complexity of all parts of the algorithm. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635840999614}, {"id": "KbttF_2rWrx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper369/Reviewer_ojVo"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper seems to propose a new use-case of an existing Byzantine fault-tolerant protocol called CenteredClip in a federated learning setting. Although the authors state that their approach fits the decentralized learning setting, the protocol relies on some sort of leader selection.\n \nThe authors bring CenterClip's equation without explaining the problem setting to a reasonable clarity. It seems the use of multi-party random number generation for Eq.(1) is the main novelty, but its context is not clear.\n \nThe authors make two assumptions on the objective function on which distributed members strive to compute the gradient, but the novelty and the context are unclear.", "review_text": "This paper addresses a very interesting task of machine learning under a decentralized setting. As the authors mention, ensuring Byzantine fault tolerance in the learning system is very hard under a general learning setting. Some of the requirements will need to be relaxed, as is the case in recent decentralized learning works (just a few examples: I know they do not discuss Bizantine):\n\n- Tran et al., \"An efficient approach for privacy preserving decentralized deep learning models based on secure multi-party computation,\" Neurocomputing 422(2021) 245-262\n- Ide et al., \"Efficient Protocol for Collaborative Dictionary Learning in Decentralized Networks,\" Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI 19,  August 10-16, Macao, China), pp.2585-2591\n \nIn this paper, the authors use leader selection schemes for \"validator\" and \"accuse\", which does not make the protocol entirely decentralized even if the selection is random. The authors need to clarify the main challenge of the existing methods and the key achievement of the paper. The first half of the paper does not look very successful in clarifying research motivation.\n\nQuestions \np.3\n\"that permutation-invariant algorithms cannot converge to any predefined accuracy of the solution\" sounds unreasonable because the original aggregation is permutation symmetric. Please elaborate. \n \np.4\nIs there any risk assessment of the validator selection step?\n \np.5\nNo definition of Split.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper seems to propose a new use-case of an existing Byzantine fault-tolerant protocol called CenteredClip in a federated learning setting. Although the authors state that their approach fits the decentralized learning setting, the protocol relies on some sort of leader selection.\n \nThe authors bring CenterClip's equation without explaining the problem setting to a reasonable clarity. It seems the use of multi-party random number generation for Eq.(1) is the main novelty, but its context is not clear.\n \nThe authors make two assumptions on the objective function on which distributed members strive to compute the gradient, but the novelty and the context are unclear.", "main_review": "This paper addresses a very interesting task of machine learning under a decentralized setting. As the authors mention, ensuring Byzantine fault tolerance in the learning system is very hard under a general learning setting. Some of the requirements will need to be relaxed, as is the case in recent decentralized learning works (just a few examples: I know they do not discuss Bizantine):\n\n- Tran et al., \"An efficient approach for privacy preserving decentralized deep learning models based on secure multi-party computation,\" Neurocomputing 422(2021) 245-262\n- Ide et al., \"Efficient Protocol for Collaborative Dictionary Learning in Decentralized Networks,\" Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI 19,  August 10-16, Macao, China), pp.2585-2591\n \nIn this paper, the authors use leader selection schemes for \"validator\" and \"accuse\", which does not make the protocol entirely decentralized even if the selection is random. The authors need to clarify the main challenge of the existing methods and the key achievement of the paper. The first half of the paper does not look very successful in clarifying research motivation.\n\nQuestions \np.3\n\"that permutation-invariant algorithms cannot converge to any predefined accuracy of the solution\" sounds unreasonable because the original aggregation is permutation symmetric. Please elaborate. \n \np.4\nIs there any risk assessment of the validator selection step?\n \np.5\nNo definition of Split.\n", "summary_of_the_review": "- unclear description\n- limited novelty", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635824335279}, {"id": "Uzr-r89Xc2L", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper369/Reviewer_38p5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper introduces BTARD-SGD as a new algorithm for distributed training among untrusted Byzantine nodes. The underlying all reduce architecture makes the proposal appealing for large scale applications due to the reduction in the communication cost. The paper contains a convergence analysis for convex and non-convex loss functions and a proof for resilience against Byzantine and Sybil attacks. The authors verify the efficacy of their proposal on image classification and NLP tasks.", "review_text": "### Strengths\n\n[Significance]\n\nAll reduce is a very popular architecture for large scale distributed training with the case of Byzantine tolerance being an open problem. The results of this paper are important.\n\n[Novelty] \n\nThe proposal builds on multiple existing schemes that come from both the ML optimization literature (centered clip, butterfly all reduce) and distributed algorithms (e.g., DHT, MPRNG).\n\n[Quality] \n\nThere is sufficient technical depth for the main claims of the paper (both for theoretical guarantees and for system design choices). The authors discuss their main assumptions and clarify cases in which they don't apply. \n\n[Presentation / Clarity]\n\nThe paper is well written and easy to follow. Parameter choices are provided and the extensive supplementary material contain the necessary details for the proofs and complementary explanations of the method.\n\n### Weaknesses\n\n[Significance]\n\n- The assumptions needed for the algorithm, prohibit an extension of the work to a setup with decentralized data (such as in federated learning). The problem of Byzantine tolerance is less motivated in the case where each node can access the entire training dataset as in this case it is often preferable to scale out in a more controlled environment (e.g., data centers) where crash failures are present but Byzantine ones are much less likely to occur. The strict IID assumptions for the paper are limiting. Convergence guarantees can still be achieved with some bound on the dissimilarity among the datasets as in practice the datasets are not arbitrarily different. Byzantine-resilient learning on decentralized data is a challenging yet open problem. For reference see the submission and discussion here: https://openreview.net/forum?id=7JSTDTZtn7-\n\n- The reputation system for nodes that join and leave introduces another limitation for the proposal, as it requires an additional application-dependent hyperparameter (T).\n\n[Evaluation] \n\n- The paper targets the large-scale setup. However, there is no evaluation of the scalability of the proposal and only experiments with up to 16 participants are shown. \n\n- Missing experimental results for the case of nodes joining.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces BTARD-SGD as a new algorithm for distributed training among untrusted Byzantine nodes. The underlying all reduce architecture makes the proposal appealing for large scale applications due to the reduction in the communication cost. The paper contains a convergence analysis for convex and non-convex loss functions and a proof for resilience against Byzantine and Sybil attacks. The authors verify the efficacy of their proposal on image classification and NLP tasks.", "main_review": "### Strengths\n\n[Significance]\n\nAll reduce is a very popular architecture for large scale distributed training with the case of Byzantine tolerance being an open problem. The results of this paper are important.\n\n[Novelty] \n\nThe proposal builds on multiple existing schemes that come from both the ML optimization literature (centered clip, butterfly all reduce) and distributed algorithms (e.g., DHT, MPRNG).\n\n[Quality] \n\nThere is sufficient technical depth for the main claims of the paper (both for theoretical guarantees and for system design choices). The authors discuss their main assumptions and clarify cases in which they don't apply. \n\n[Presentation / Clarity]\n\nThe paper is well written and easy to follow. Parameter choices are provided and the extensive supplementary material contain the necessary details for the proofs and complementary explanations of the method.\n\n### Weaknesses\n\n[Significance]\n\n- The assumptions needed for the algorithm, prohibit an extension of the work to a setup with decentralized data (such as in federated learning). The problem of Byzantine tolerance is less motivated in the case where each node can access the entire training dataset as in this case it is often preferable to scale out in a more controlled environment (e.g., data centers) where crash failures are present but Byzantine ones are much less likely to occur. The strict IID assumptions for the paper are limiting. Convergence guarantees can still be achieved with some bound on the dissimilarity among the datasets as in practice the datasets are not arbitrarily different. Byzantine-resilient learning on decentralized data is a challenging yet open problem. For reference see the submission and discussion here: https://openreview.net/forum?id=7JSTDTZtn7-\n\n- The reputation system for nodes that join and leave introduces another limitation for the proposal, as it requires an additional application-dependent hyperparameter (T).\n\n[Evaluation] \n\n- The paper targets the large-scale setup. However, there is no evaluation of the scalability of the proposal and only experiments with up to 16 participants are shown. \n\n- Missing experimental results for the case of nodes joining.\n\n", "summary_of_the_review": "Although the paper has some significant weaknesses, I tend to recommend an acceptance given the technical depth and significance of the problem addressed.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635082891950}], "openreview_url": "https://openreview.net/forum?id=6PahjGFjVG-", "arxiv_id": "2106.11257", "paper_pdf": "papers/6PahjGFjVG-.pdf", "paper_pdf_sha256": "f06b41fcde5eca3b4cff6d651cf777df7658462825fac1b3c06bf84ebb5bec51", "paper_pdf_bytes": 1546801, "paper_pdf_source": "openreview", "code_url": "https://github.com/icml-2022-paper/BTARD", "code_repository": "icml-2022-paper/BTARD", "code_commit": "a0e758ae5c5e405775144bc3f2376bfc14d007f8", "code_archive": "repos/6PahjGFjVG-.zip", "code_archive_sha256": "1b49a38c2b3f599a2d91c14fbbf589457dd06e4a1c1646b18e1e5d5e96506aec", "code_archive_bytes": 138960, "code_file_count": 52, "code_extensions": {".py": 50, ".ipynb": 2}, "github_disk_usage_kb": 123, "github_languages": {"Python": 391126, "Jupyter Notebook": 22854}, "github_archived": false, "github_pushed_at": "2021-10-07T21:15:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/secure-distributed-training-at-scale"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "N07ebsD-lHp", "year": 2021, "status": "rejected", "title": "Defending against black-box adversarial attacks with gradient-free trained sign activation neural networks", "authors": ["Yunzhe Xue", "Meiyan Xie", "Zhibo Yang", "Usman Roshan"], "authorids": ["yx277@njit.edu", "mx42@njit.edu", "zy328@njit.edu", "~Usman_Roshan1"], "authors_source": "OpenReview API", "abstract": "While machine learning models today can achieve high accuracies on classification tasks, they can be deceived by minor imperceptible distortions to the data. These are known as adversarial attacks and can be lethal in the black-box setting which does not require knowledge of the target model type or its parameters. Binary neural networks that have sign activation and are trained with gradient descent have been shown to be harder to attack than conventional sigmoid activation networks but their improvements are marginal. We instead train sign activation networks with a novel gradient-free stochastic coordinate descent algorithm and propose an ensemble of such networks as a defense model. We evaluate the robustness of our model (a hard problem in itself) on image, text, and medical ECG data and find it to be more robust than ensembles of binary, full precision, and convolutional neural networks, and than random forests while attaining comparable clean test accuracy. In order to explain our model's robustness we show that an adversary targeting a single network in our ensemble fails to attack (and thus non-transferable to) other networks in the ensemble. Thus a datapoint requires a large distortion to fool the majority of networks in our ensemble and is likely to be detected in advance. This property of non-transferability arises naturally from the non-convexity of sign activation networks and randomization in our gradient-free training algorithm without any adversarial defense effort.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "TCDCUkt2ufH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1250/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new method to defend black-box attack based on an ensemble of sign activation neural networks. The authors demonstrate their method has much higher minimum distortion using HopSkipJump attack.\n\nHowever, I have many concerns regarding this paper:\n\n-The paper organization is very weird and confusing. The author did not put their main algorithm into the main paper. Instead, they put many unimportant results (e.g. Table 1) into the main paper. If there is not enough space the author should use simpler sentences to describe their algorithm and put some results into supp. Also, the figure and table style (i.e., unbounded table, screenshot figures) makes me feel this is an undergrad project report instead of an ICLR submission.\n\n-Besides transfer attack, the authors only evaluate the black-block robustness using HopSkipJump attack. Their claim is \"Compared to other boundary attack methods it is known to give the best estimate of a datapoint’s minimum adversarial distortion.\" \nIs there any paper support this claim? I don't believe one attack method is universally better than other method among all datasets.\nI think the authors should evaluate a set of attack methods instead of only one method otherwise the results are not convincing to me.\n\n-What is the purpose of providing detailed results of 10 random images in Table 2, 3? Those results not only occupied a lot of space but also did not provide any useful insight. An average number of the entire dataset is enough.\n\n-Citing issue: when referencing a paper the author should use the published version not the arxiv version if the cited paper is published.\nE.g., Angus Galloway, Graham W Taylor, and Medhat Moussa. Attacking binarized neural networks. arXiv preprint arXiv:1711.00449, 2017.\nshould be\nGalloway, Angus, Graham W. Taylor, and Medhat Moussa. Attacking Binarized Neural Networks. International Conference on Learning Representations. 2018.\n\n\n-The baseline comparison are all undefended networks. The author should compare to some other blackbox defense methods. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unconvincing evaluation", "review": "This paper presents a new method to defend black-box attack based on an ensemble of sign activation neural networks. The authors demonstrate their method has much higher minimum distortion using HopSkipJump attack.\n\nHowever, I have many concerns regarding this paper:\n\n-The paper organization is very weird and confusing. The author did not put their main algorithm into the main paper. Instead, they put many unimportant results (e.g. Table 1) into the main paper. If there is not enough space the author should use simpler sentences to describe their algorithm and put some results into supp. Also, the figure and table style (i.e., unbounded table, screenshot figures) makes me feel this is an undergrad project report instead of an ICLR submission.\n\n-Besides transfer attack, the authors only evaluate the black-block robustness using HopSkipJump attack. Their claim is \"Compared to other boundary attack methods it is known to give the best estimate of a datapoint’s minimum adversarial distortion.\" \nIs there any paper support this claim? I don't believe one attack method is universally better than other method among all datasets.\nI think the authors should evaluate a set of attack methods instead of only one method otherwise the results are not convincing to me.\n\n-What is the purpose of providing detailed results of 10 random images in Table 2, 3? Those results not only occupied a lot of space but also did not provide any useful insight. An average number of the entire dataset is enough.\n\n-Citing issue: when referencing a paper the author should use the published version not the arxiv version if the cited paper is published.\nE.g., Angus Galloway, Graham W Taylor, and Medhat Moussa. Attacking binarized neural networks. arXiv preprint arXiv:1711.00449, 2017.\nshould be\nGalloway, Angus, Graham W. Taylor, and Medhat Moussa. Attacking Binarized Neural Networks. International Conference on Learning Representations. 2018.\n\n\n-The baseline comparison are all undefended networks. The author should compare to some other blackbox defense methods. ", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603897867927}, {"id": "361i90E3fF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1250/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an architecture (ensemble of networks) aiming at being robust against black-box attacks, based on the idea that crafting an adversarial example able to fool enough individual networks such that the majority vote changes is a more difficult task.  The paper presents ways of training such ensembles and provides several sets of experiments showing the advantage of the approach. It also contains an observation on \"non-transferability\", counting how many co-networks are fooled when only one is targetted by the blackbox attack. It turns out that this amount is lower for the proposed scheme. \n\nThe algorithms are postponed to supplementary material and the paper itself mainly report the experimental part. It concerns : \n-> training time\n-> minimum adversarial distorsion (l2 and l_inf)\n-> transferabiliy\n-> different tasks (images, text and ecg)\n\nAll tested algorithms are ensembles and dataset are subparts of actual datasets.\n\nComments: \n* Table 1 : training time of a single run : should not it be avaraged on several runs somehow?  \n* Table 2, 3, 5 : I would find it more readable if some remarkable values were bold for instance\n* Table 4 : since all models don't have the same number of weak learners, I suggest that the presentation of the table recalls it, it * has a great impact on values here\n* (a verb is missing on top of page 6. )\nFigure 2 : hard to read. Top lines appears to be  some kind of reference to be compared to and the bottom lines the proposed methods. I suggest that one color is assigned to each algorithm and different line style for with or without GN. Also I think that there's a typo in the legend (SCDCE-BNN-> SCDCE-GN?) \n* biblio : I did not check each paper on arxiv but I'm quite sure that a significative part of arvix references have some published references.  Citing preprints is ok for recent work only. The reference by Alex Krizhesky seems to lack information too. \n\nOverall, I'm not really convinced by the paper in its present form, although I recognize that the results show some interesting properties for robustness.  Experiment on running time seems a little misleading, and shows that the most efficient variant is also significantly slower. I'm actually not against slower methods, but when I read this part I have a feeling of \"is it right?\" \nI also think that there would be enough space to provide more information on algorithms in the main paper. \nThe transferability idea seems to be an interesting point, but it has to be further developped. How to apply this observation to the actual blackbox attack, which, as far as I understand, does not attack weak learners one by one?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "an experimental paper that could be improved", "review": "The paper proposes an architecture (ensemble of networks) aiming at being robust against black-box attacks, based on the idea that crafting an adversarial example able to fool enough individual networks such that the majority vote changes is a more difficult task.  The paper presents ways of training such ensembles and provides several sets of experiments showing the advantage of the approach. It also contains an observation on \"non-transferability\", counting how many co-networks are fooled when only one is targetted by the blackbox attack. It turns out that this amount is lower for the proposed scheme. \n\nThe algorithms are postponed to supplementary material and the paper itself mainly report the experimental part. It concerns : \n-> training time\n-> minimum adversarial distorsion (l2 and l_inf)\n-> transferabiliy\n-> different tasks (images, text and ecg)\n\nAll tested algorithms are ensembles and dataset are subparts of actual datasets.\n\nComments: \n* Table 1 : training time of a single run : should not it be avaraged on several runs somehow?  \n* Table 2, 3, 5 : I would find it more readable if some remarkable values were bold for instance\n* Table 4 : since all models don't have the same number of weak learners, I suggest that the presentation of the table recalls it, it * has a great impact on values here\n* (a verb is missing on top of page 6. )\nFigure 2 : hard to read. Top lines appears to be  some kind of reference to be compared to and the bottom lines the proposed methods. I suggest that one color is assigned to each algorithm and different line style for with or without GN. Also I think that there's a typo in the legend (SCDCE-BNN-> SCDCE-GN?) \n* biblio : I did not check each paper on arxiv but I'm quite sure that a significative part of arvix references have some published references.  Citing preprints is ok for recent work only. The reference by Alex Krizhesky seems to lack information too. \n\nOverall, I'm not really convinced by the paper in its present form, although I recognize that the results show some interesting properties for robustness.  Experiment on running time seems a little misleading, and shows that the most efficient variant is also significantly slower. I'm actually not against slower methods, but when I read this part I have a feeling of \"is it right?\" \nI also think that there would be enough space to provide more information on algorithms in the main paper. \nThe transferability idea seems to be an interesting point, but it has to be further developped. How to apply this observation to the actual blackbox attack, which, as far as I understand, does not attack weak learners one by one?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603825914390}, {"id": "KxwP4xZcVQj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1250/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n\nThis paper proposes training an ensemble of binary neural networks with sign activations using gradient-free stochastic coordinate descent algorithm. \nThe nature of the training method and binary networks leads to robust models with non-transferable attacks.\n\n----\n*Pros*:\nI appreciated \n- the non-transferability experiment which demonstrated that the trained ensemble is diverse and sort-of \"orthogonal\".\n- the use of minimum distortion as a measure of the model's robustness.\n\n*Cons*:\n- It looks like the method is not scalable. Experiments were carried out on a single-hidden-layer network of 20 nodes (would really like to see experiments on bigger models)\n- Only one black-box attack is used to demonstrate the effectiveness of the method and that too under one perturbation threat (l2-norm). There are multiple black-box attacks that exploit different aspects of model vulnerabilities. It cant be claimed that the method is robust based on a single-attack single-threat.\n- The scale of experiments (10 images) does not provide significant evidence that this method is effective and the authors acknowledge that in Section 3.6\n\n\nRemarks/Comments:\n- Page 2, Line2: \"by with gradient\" -> \"with gradient\".\n- Sec 3.2: \"divide by 9900\" -> \"divide by 99\".\n- Table 1: What does \"single run\" mean in the caption? Is it a single forward-pass?\n- Fig 2: please make it more legible with markers (colors are not enough).\n\n\nBased on the above, I suggest that the authors redesign their experiment to support the paper's proposition in terms of the method scalability and robustness against more attacks & perturbation settings. Perhaps making the experiments more scalable by reducing the compute budget of HopSkipJump (1000 random init x 10 times x 100 iterations)\n\n**Post-Rebuttal**: No change in my score, please read my comments in the thread below.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lots of work is needed to improve the paper", "review": "\n\nThis paper proposes training an ensemble of binary neural networks with sign activations using gradient-free stochastic coordinate descent algorithm. \nThe nature of the training method and binary networks leads to robust models with non-transferable attacks.\n\n----\n*Pros*:\nI appreciated \n- the non-transferability experiment which demonstrated that the trained ensemble is diverse and sort-of \"orthogonal\".\n- the use of minimum distortion as a measure of the model's robustness.\n\n*Cons*:\n- It looks like the method is not scalable. Experiments were carried out on a single-hidden-layer network of 20 nodes (would really like to see experiments on bigger models)\n- Only one black-box attack is used to demonstrate the effectiveness of the method and that too under one perturbation threat (l2-norm). There are multiple black-box attacks that exploit different aspects of model vulnerabilities. It cant be claimed that the method is robust based on a single-attack single-threat.\n- The scale of experiments (10 images) does not provide significant evidence that this method is effective and the authors acknowledge that in Section 3.6\n\n\nRemarks/Comments:\n- Page 2, Line2: \"by with gradient\" -> \"with gradient\".\n- Sec 3.2: \"divide by 9900\" -> \"divide by 99\".\n- Table 1: What does \"single run\" mean in the caption? Is it a single forward-pass?\n- Fig 2: please make it more legible with markers (colors are not enough).\n\n\nBased on the above, I suggest that the authors redesign their experiment to support the paper's proposition in terms of the method scalability and robustness against more attacks & perturbation settings. Perhaps making the experiments more scalable by reducing the compute budget of HopSkipJump (1000 random init x 10 times x 100 iterations)\n\n**Post-Rebuttal**: No change in my score, please read my comments in the thread below.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603512756168}], "openreview_url": "https://openreview.net/forum?id=N07ebsD-lHp", "arxiv_id": null, "paper_pdf": "papers/N07ebsD-lHp.pdf", "paper_pdf_sha256": "aaa493edcbf910bf4c7472f4ed0dbe1ab8710dce10a77c03dc46b5610aeb8099", "paper_pdf_bytes": 284323, "paper_pdf_source": "openreview", "code_url": "https://github.com/zero-one-loss/scd_github", "code_repository": "zero-one-loss/scd_github", "code_commit": "c41fee1dd4d86f152748590887f7e4d0a95c89c8", "code_archive": "repos/N07ebsD-lHp.zip", "code_archive_sha256": "3a562a85154be834cd871ab4d9fab374a390087f20ceaddeb160f770d99ffe19", "code_archive_bytes": 211424, "code_file_count": 33, "code_extensions": {".py": 32, ".sh": 1}, "github_disk_usage_kb": 176, "github_languages": {"Python": 372079, "Shell": 40733}, "github_archived": false, "github_pushed_at": "2020-10-15T02:18:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/defending-against-black-box-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryxF80NYwS", "year": 2020, "status": "rejected", "title": "Neural Clustering Processes", "authors": ["Ari Pakman", "Yueqi Wang", "Catalin Mitelut", "JinHyung Lee", "Liam Paninski"], "authorids": ["aripakman@gmail.com", "yueqi.wang.pku@gmail.com", "mitelutco@gmail.com", "jl4303@columbia.edu", "liam@stat.columbia.edu"], "authors_source": "OpenReview API", "abstract": "Mixture models, a basic building block in countless statistical models, involve latent random variables over discrete spaces, and existing posterior inference methods can be inaccurate and/or very slow.  In this work we introduce a novel deep learning architecture for efficient amortized Bayesian inference over mixture models. While previous approaches to amortized clustering assumed a fixed or maximum number of mixture components and only amortized over the continuous parameters of each mixture component, our method amortizes over the local discrete labels of all the data points, and performs inference over an unbounded number of mixture components. The latter property makes our method natural for the challenging case of nonparametric Bayesian models, where the number of mixture components grows with the dataset. Our approach exploits the exchangeability of the generative models and is based on mapping distributed, permutation-invariant representations of discrete  arrangements into varying-size multinomial conditional probabilities. The resulting algorithm parallelizes easily, yields iid samples from the approximate posteriors along with a normalized probability estimate of each sample (a quantity generally unavailable using Markov Chain Monte Carlo) and can easily be applied to both conjugate and non-conjugate models, as training only requires samples from the generative model. We also present an extension of the method to models of random communities (such as infinite relational or stochastic block models). As a scientific application, we present a novel approach to neural spike sorting for high-density multielectrode arrays. \n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJxZrE8BcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1147/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nIn this paper, the authors consider the neural amortized inference for clustering processes, in which the number of cluster can be automatically adapted based on the observed samples. The proposed algorithm largely follows the standard variational auto-encoder. The major contribution of the paper is the design of the posterior parametrization so that the posterior satisfies the permutation invariant within a cluster, between clusters, and unassigned data, based on the DeepSet method. The model can be incorporated into random communities models. Finally, the authors apply the algorithm for neural spike sorting problem. \n\n\nThe paper is well-organized and easy to follow. However, there are two issues should be addressed:\n\n1, The novelty of the proposed algorithm might not enough. The two major components in this paper, i.e., VAE and DeepSet, are all carefully investigated before. This paper applies the DeepSet parameterization in the VAE framework.\n\n2, The details of the amortized inference training is not clearly explained. It is well-known that gradient through the discrete random variable is quite difficult. How the gradient for the parameter of the proposed model is calculated should be carefully discussed. The REINFORCE gradient in this model, whose support of c can be as large as the number of samples, can be quite huge. \n\n3, In the empirical evaluation, I was curious why the mean-field and MCMC have not been considered in the spike sorting problem.\n\nI am expecting the authors can address my concerns during rebuttal. \n\n=======================================================================\n\nThanks for the responses to clarify my concerns. \n\nThe learning procedure and experiments are clear now.  Indeed, as a purely variational inference paper,  the discrete variable problem is absent, since the model is always *fixed* and not updated. \n\nThe major contribution of the paper becomes the design of the posterior parametrization by DeepSets, which I think still not enough. \n\nI will keep my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "\nIn this paper, the authors consider the neural amortized inference for clustering processes, in which the number of cluster can be automatically adapted based on the observed samples. The proposed algorithm largely follows the standard variational auto-encoder. The major contribution of the paper is the design of the posterior parametrization so that the posterior satisfies the permutation invariant within a cluster, between clusters, and unassigned data, based on the DeepSet method. The model can be incorporated into random communities models. Finally, the authors apply the algorithm for neural spike sorting problem. \n\n\nThe paper is well-organized and easy to follow. However, there are two issues should be addressed:\n\n1, The novelty of the proposed algorithm might not enough. The two major components in this paper, i.e., VAE and DeepSet, are all carefully investigated before. This paper applies the DeepSet parameterization in the VAE framework.\n\n2, The details of the amortized inference training is not clearly explained. It is well-known that gradient through the discrete random variable is quite difficult. How the gradient for the parameter of the proposed model is calculated should be carefully discussed. The REINFORCE gradient in this model, whose support of c can be as large as the number of samples, can be quite huge. \n\n3, In the empirical evaluation, I was curious why the mean-field and MCMC have not been considered in the spike sorting problem.\n\nI am expecting the authors can address my concerns during rebuttal. \n\n=======================================================================\n\nThanks for the responses to clarify my concerns. \n\nThe learning procedure and experiments are clear now.  Indeed, as a purely variational inference paper,  the discrete variable problem is absent, since the model is always *fixed* and not updated. \n\nThe major contribution of the paper becomes the design of the posterior parametrization by DeepSets, which I think still not enough. \n\nI will keep my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572328505205}, {"id": "Hyemu9Byqr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1147/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a neural network based clustering process where the number of clusters is not known a-priori.  Proposed approach requires conjecturing a generative process (where number of clusters/classes is a random variable) and the model learns to uncover the posterior distribution over clusters given the observed samples.\nOverall I think it is a valuable contribution, well written paper with good results.  \nSpecific comments:\na) Even though the model allows for variable number of clusters, I feel there may be a strong dependence between the number of clusters the model can hypothesize and the number of clusters in the training data.  It will be useful to give further insights into this.  For instance, if an MNIST model is trained with only digits 0-5 training data, how well would it perform in detecting all 10 clusters at test time?  Understanding model’s biases based on training data is one area I feel is important and the paper could add to.\nb) The neural clustering process could potentially be viewed as a transductive inference model for classification of test data.  Typically at test time classification is done for each test sample independently, and the clustering process allows one to bring in other similar test samples to help with classification.  Have the authors considered this and have any comments on potential value / feasibility of this?\nc) In the examples presented in Section 2.3, please clarify how training & testing was done.  Specifically what training data was used (all of MNIST training data?), and the test time clustering was done on a subset of MNIST test data?\nd) Use of ‘q’ for a neural-network and ‘q_\\theta’ for posterior distribution is a little confusing, will be better to have different notation for these.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper presents a neural network based clustering process where the number of clusters is not known a-priori.  Proposed approach requires conjecturing a generative process (where number of clusters/classes is a random variable) and the model learns to uncover the posterior distribution over clusters given the observed samples.\nOverall I think it is a valuable contribution, well written paper with good results.  \nSpecific comments:\na) Even though the model allows for variable number of clusters, I feel there may be a strong dependence between the number of clusters the model can hypothesize and the number of clusters in the training data.  It will be useful to give further insights into this.  For instance, if an MNIST model is trained with only digits 0-5 training data, how well would it perform in detecting all 10 clusters at test time?  Understanding model’s biases based on training data is one area I feel is important and the paper could add to.\nb) The neural clustering process could potentially be viewed as a transductive inference model for classification of test data.  Typically at test time classification is done for each test sample independently, and the clustering process allows one to bring in other similar test samples to help with classification.  Have the authors considered this and have any comments on potential value / feasibility of this?\nc) In the examples presented in Section 2.3, please clarify how training & testing was done.  Specifically what training data was used (all of MNIST training data?), and the test time clustering was done on a subset of MNIST test data?\nd) Use of ‘q’ for a neural-network and ‘q_\\theta’ for posterior distribution is a little confusing, will be better to have different notation for these.\n"}, "tcdate": 1571932778841}, {"id": "H1g5dPCpFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1147/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper introduces a novel deep learning architecture for efficient amortized Bayesian inference over mixture models. Unlike previous approaches to amortized clustering, the proposed method allows us to treat local discrete labels of data points and infer the unbounded number of mixture components, making it more flexible as in the case of Bayesian nonparametrics. It is shown that the resulting algorithm can be parallelized and applied to both conjugate and non-conjugate models. The authors also suggest an extension to models of random communities and a novel approach to neural spike sorting for high-density multielectrode arrays based on the proposed method.\n\nStrengths:\nThe paper is generally well written and the relationship to previous works is well described. Empirical results seem quite convincing, for example, the clustering results presented in Fig. 2 and Fig. 3 clearly show not only the inferred number of clusters, but also the posterior probability which indicates that reasonable samples are assigned higher probability.\n\nWeaknesses:\n- Overall, the idea looks very original and promising, but I find some technical details are not easy to understand under the current form, especially for non-experts in this domain. I would recommend the authors to elaborate a bit more on the proposed architecture and the variable-input soft-max function in Sect. 2.1.\n- On page 8, the authors mention that the NCP is much more efficient compared with MCMC, for example, in the Gaussian 2D example. However, regarding the DPMM clustering model, it is known that MCMC methods are generally slower compared with variational inference, which is computationally faster. I think it would be interesting to add a discussion or comparison with variational inference in terms of computational efficiency.\n- If I understand correctly, the NCP is essentially based on a sequential sampling procedure. The authors claim that the proposed method is easily parallelized using a GPU, but there does not seem to be sufficient details on the GPU-parallelization of sequential sampling.\n\nMinor comments:\nThe size of some figures appears too small, for example Fig. 6 and Fig. 10, which may hinder readability.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary:\nThis paper introduces a novel deep learning architecture for efficient amortized Bayesian inference over mixture models. Unlike previous approaches to amortized clustering, the proposed method allows us to treat local discrete labels of data points and infer the unbounded number of mixture components, making it more flexible as in the case of Bayesian nonparametrics. It is shown that the resulting algorithm can be parallelized and applied to both conjugate and non-conjugate models. The authors also suggest an extension to models of random communities and a novel approach to neural spike sorting for high-density multielectrode arrays based on the proposed method.\n\nStrengths:\nThe paper is generally well written and the relationship to previous works is well described. Empirical results seem quite convincing, for example, the clustering results presented in Fig. 2 and Fig. 3 clearly show not only the inferred number of clusters, but also the posterior probability which indicates that reasonable samples are assigned higher probability.\n\nWeaknesses:\n- Overall, the idea looks very original and promising, but I find some technical details are not easy to understand under the current form, especially for non-experts in this domain. I would recommend the authors to elaborate a bit more on the proposed architecture and the variable-input soft-max function in Sect. 2.1.\n- On page 8, the authors mention that the NCP is much more efficient compared with MCMC, for example, in the Gaussian 2D example. However, regarding the DPMM clustering model, it is known that MCMC methods are generally slower compared with variational inference, which is computationally faster. I think it would be interesting to add a discussion or comparison with variational inference in terms of computational efficiency.\n- If I understand correctly, the NCP is essentially based on a sequential sampling procedure. The authors claim that the proposed method is easily parallelized using a GPU, but there does not seem to be sufficient details on the GPU-parallelization of sequential sampling.\n\nMinor comments:\nThe size of some figures appears too small, for example Fig. 6 and Fig. 10, which may hinder readability."}, "tcdate": 1571837809529}], "openreview_url": "https://openreview.net/forum?id=ryxF80NYwS", "arxiv_id": "1901.00409", "paper_pdf": "papers/ryxF80NYwS.pdf", "paper_pdf_sha256": "eb34cc78bf2fde50e3030a759de695867ef15ddf2e4f34afdc93ed8c78cabdfb", "paper_pdf_bytes": 2799517, "paper_pdf_source": "openreview", "code_url": "https://github.com/aripakman/neural_clustering_process", "code_repository": "aripakman/neural_clustering_process", "code_commit": "dd54a8793398f7cb2dc44365dcb3d9507b89ab34", "code_archive": "repos/ryxF80NYwS.zip", "code_archive_sha256": "03d604a55480e607e696db490e8ac339b8ca16776282b93f71cd8350d2f67410", "code_archive_bytes": 99309, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 190, "github_languages": {"Python": 30222}, "github_archived": false, "github_pushed_at": "2020-07-19T12:45:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discrete-neural-processes"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkNSehA9FQ", "year": 2019, "status": "rejected", "title": "Open Vocabulary Learning on Source Code with a Graph-Structured Cache", "authors": ["Milan Cvitkovic", "Badal Singh", "Anima Anandkumar"], "authorids": ["mcvitkov@caltech.edu", "sbadal@amazon.com", "anima@caltech.edu"], "authors_source": "OpenReview API", "abstract": "Machine learning models that take computer program source code as input typically use Natural Language Processing (NLP) techniques. However, a major challenge is that code is written using an open, rapidly changing vocabulary due to, e.g., the coinage of new variable and method names.  Reasoning over such a vocabulary is not something for which most NLP methods are designed.  We introduce a Graph-Structured Cache to address this problem; this cache contains a node for each new word the model encounters with edges connecting each word to its occurrences in the code.  We find that combining this graph-structured cache strategy with recent Graph-Neural-Network-based models for supervised learning on code improves the models' performance on a code completion task and a variable naming task --- with over 100\\% relative improvement on the latter --- at the cost of a moderate increase in computation time.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rklbFlAO2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1075/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "(updated with some summaries from discussion over the initial review)\n\nThe paper discusses the topics of predicting out-of-vocabulary tokens in programs abstract syntax trees. This could have application in code completion and more concretely two tasks are evaluated:\n - predicting a missing reference to a variable (called FillInTheBlank)\n - predicting a name of a variable (NameMe)\n\nUnfortunately, the paper proposes overly complex and strange formulations of these tasks, heavy implementation with unnecessary (non-motivated) neural architectures and as a result, does not demonstrate state-of-the-art performance or precision on comparable tasks. Figure 1 shows the complexity of the approach, with multiple steps of building a graph, introducing the vocabulary cache to then produce a vector at every node of the input tree of the program (instead of creating architecture for a given task), yet simple analysis over which variables can be chosen is missing.\n\nThe FillInTheBlank task is badly defined already on the running example. The goal is to select a variable to fill in a blank and already in the example on Figure 2, one of the candidate variables is out of scope at the location to fill. The motivation for the proposed formulation with building a graph and then computing attention over nodes in that graph is unclear and experiments do not help it. For example, [1] (also cited in the paper) solves the same problem more cleanly by considering only the variables in the scope*. There is no good experimental comparison to that work, but it is unlikely it will perform worse. Also [1] does not suffer from vocabulary problems for that task.\n\nSummary discussion below: the experiments here are incomparable on many levels with prior works: different architecture details, different even smaller dataset than from [1]. There is a third-party claim that on a full system, the general idea improves performance, but I take it with a grain of salt as no clean experiment was yet done. The reviewer notes that the authors disagree the baselines are not meaningful.\n\nThe NameMe tasks also shows the weakness of the proposed architectures. This work proposes to compute vectors at every node where a variable occurs and then to average them and decode the variable name to predict. In comparison, several prior works introduce one node per variable (not per occurrence), essentially removing the long distance relationships between occurrences of the same variable variables and removing the need to average vectors and enforcing the same name representation at every occurrence of the variable [name]. The setup here is incomparable to specialized naming prior works, one feature (a node per variable) is replaced with another (a node per subtoken), but for baselines authors choose to only to be similar to [1]. Also, while not on the same dataset, [2,3] consistently get higher accuracy on a related and more complicated task of predicting multiple names at the same time over multiple programming languages and with much simpler linear models. This is not surprising, because they propose simpler architectures better suited for the NameMe task.\n\n[1] Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs\nwith graphs. ICLR 2017\n[2] Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from Big\nCode\n[3] Uri Alon, Meital Zilberstein, Omer Levy, Eran Yahav. A General Path-Based Representation for Predicting Program\n\n* corrected text\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overly complicated techniques for previously well-addressed tasks in literature", "review": "(updated with some summaries from discussion over the initial review)\n\nThe paper discusses the topics of predicting out-of-vocabulary tokens in programs abstract syntax trees. This could have application in code completion and more concretely two tasks are evaluated:\n - predicting a missing reference to a variable (called FillInTheBlank)\n - predicting a name of a variable (NameMe)\n\nUnfortunately, the paper proposes overly complex and strange formulations of these tasks, heavy implementation with unnecessary (non-motivated) neural architectures and as a result, does not demonstrate state-of-the-art performance or precision on comparable tasks. Figure 1 shows the complexity of the approach, with multiple steps of building a graph, introducing the vocabulary cache to then produce a vector at every node of the input tree of the program (instead of creating architecture for a given task), yet simple analysis over which variables can be chosen is missing.\n\nThe FillInTheBlank task is badly defined already on the running example. The goal is to select a variable to fill in a blank and already in the example on Figure 2, one of the candidate variables is out of scope at the location to fill. The motivation for the proposed formulation with building a graph and then computing attention over nodes in that graph is unclear and experiments do not help it. For example, [1] (also cited in the paper) solves the same problem more cleanly by considering only the variables in the scope*. There is no good experimental comparison to that work, but it is unlikely it will perform worse. Also [1] does not suffer from vocabulary problems for that task.\n\nSummary discussion below: the experiments here are incomparable on many levels with prior works: different architecture details, different even smaller dataset than from [1]. There is a third-party claim that on a full system, the general idea improves performance, but I take it with a grain of salt as no clean experiment was yet done. The reviewer notes that the authors disagree the baselines are not meaningful.\n\nThe NameMe tasks also shows the weakness of the proposed architectures. This work proposes to compute vectors at every node where a variable occurs and then to average them and decode the variable name to predict. In comparison, several prior works introduce one node per variable (not per occurrence), essentially removing the long distance relationships between occurrences of the same variable variables and removing the need to average vectors and enforcing the same name representation at every occurrence of the variable [name]. The setup here is incomparable to specialized naming prior works, one feature (a node per variable) is replaced with another (a node per subtoken), but for baselines authors choose to only to be similar to [1]. Also, while not on the same dataset, [2,3] consistently get higher accuracy on a related and more complicated task of predicting multiple names at the same time over multiple programming languages and with much simpler linear models. This is not surprising, because they propose simpler architectures better suited for the NameMe task.\n\n[1] Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs\nwith graphs. ICLR 2017\n[2] Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from Big\nCode\n[3] Uri Alon, Meital Zilberstein, Omer Levy, Eran Yahav. A General Path-Based Representation for Predicting Program\n\n* corrected text\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541099641305}, {"id": "HkgDMtLd37", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1075/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces a new way to use a subword embedding model 2 tasks related to codes: fill-in-blank and variable naming. \n\n* pros: \n- the paper is very well written. \n- the model is easily to reimplement. \n- the experiments are solid and the results are convincing. \n\n* cons: \n- the title is very misleading. In fact, what the paper does is to use a very shallow subword embedding method for names. This approach is widely used in NLP, especially in machine translation. \n- the work is progressing, meaning that most of it is based on another work (i.e. Allamanis et al 2018). \n\n* questions: \n- how to build the (subword) vocabulary? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A subword embedding model for codes. What's new?", "review": "The paper introduces a new way to use a subword embedding model 2 tasks related to codes: fill-in-blank and variable naming. \n\n* pros: \n- the paper is very well written. \n- the model is easily to reimplement. \n- the experiments are solid and the results are convincing. \n\n* cons: \n- the title is very misleading. In fact, what the paper does is to use a very shallow subword embedding method for names. This approach is widely used in NLP, especially in machine translation. \n- the work is progressing, meaning that most of it is based on another work (i.e. Allamanis et al 2018). \n\n* questions: \n- how to build the (subword) vocabulary? \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541069071307}, {"id": "rJxfsg4Dh7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1075/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The submission presents an extension to the Allamanis et al ICLR'18 paper on learning from programs as graphs. The core contribution is the idea of introducing extra nodes and edges into the graph that correspond to (potentially rare) subwords used in the analyzed program code. Experiments show that this extended graph leads to better performance on two tasks, compared to a wide range of baseline methods.\n\nOverall, this is a nice paper with a small, incremental idea and substantial experiments that show its practical value. I only have minor comments / questions on the actual core content. However, the contribution is very incremental and of interest to a specialized subsegment of the ICLR audience, so it may be appropriate to reject the paper and redirect the authors to a more specialized venue.\n\nMinor comments:\n- There's a bunch of places where \\citep/\\citet are mixed up (e.g., second to last paragraph of page 2). It would make sense to go through the paper one more time to clean this up.\n- Sect. 4: I understand the need to introduce context, but it feels that more space should be spent on the actual contribution here (step 3). For example, it remains unclear why this extra nodes / edges are only introduced for subwords appearing in variables - why not also for field names / method names?\n- Sect. 5: It would be helpful if the authors would explicitly handle the code duplication problem (Lopes et al., OOPSLA'17), or discuss how they avoided these problems. Duplicated data files occurring in several folds are a significant risk to the validity of their experimental findings, and very common in code corpora.\n- Table 1: It is unclear to me what the \"Pointer Sentinel\" model can achieve. Without edges connecting the additional words to where they occur, it seems that this should not be performing different than \"Closed Vocab\", apart from noise introduced by additional nodes.\n- Table 1: Do Pointer Sentinel/GSC use a CharCNN to embed node labels of nodes that are not part of the \"cache\", or a closed vocabulary? [i.e., what's the embedding of a variable \"foo\"?] If not, what is the performance of the GSC model with CharCNN-embeddings everywhere? That would be architecturally simpler than the split variant, and so may be of interest.\n- Page 6: When truncating to 500 nodes per graph: How many graphs in your dataset are larger than that?\n- Page 7: Do you really use attention over all nodes, instead of only nodes corresponding to variables? How do you deal with results where the model picks a non-variable (e.g., a corresponding cache node)? Does this happen?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Improved graph representation for learning from programs", "review": "The submission presents an extension to the Allamanis et al ICLR'18 paper on learning from programs as graphs. The core contribution is the idea of introducing extra nodes and edges into the graph that correspond to (potentially rare) subwords used in the analyzed program code. Experiments show that this extended graph leads to better performance on two tasks, compared to a wide range of baseline methods.\n\nOverall, this is a nice paper with a small, incremental idea and substantial experiments that show its practical value. I only have minor comments / questions on the actual core content. However, the contribution is very incremental and of interest to a specialized subsegment of the ICLR audience, so it may be appropriate to reject the paper and redirect the authors to a more specialized venue.\n\nMinor comments:\n- There's a bunch of places where \\citep/\\citet are mixed up (e.g., second to last paragraph of page 2). It would make sense to go through the paper one more time to clean this up.\n- Sect. 4: I understand the need to introduce context, but it feels that more space should be spent on the actual contribution here (step 3). For example, it remains unclear why this extra nodes / edges are only introduced for subwords appearing in variables - why not also for field names / method names?\n- Sect. 5: It would be helpful if the authors would explicitly handle the code duplication problem (Lopes et al., OOPSLA'17), or discuss how they avoided these problems. Duplicated data files occurring in several folds are a significant risk to the validity of their experimental findings, and very common in code corpora.\n- Table 1: It is unclear to me what the \"Pointer Sentinel\" model can achieve. Without edges connecting the additional words to where they occur, it seems that this should not be performing different than \"Closed Vocab\", apart from noise introduced by additional nodes.\n- Table 1: Do Pointer Sentinel/GSC use a CharCNN to embed node labels of nodes that are not part of the \"cache\", or a closed vocabulary? [i.e., what's the embedding of a variable \"foo\"?] If not, what is the performance of the GSC model with CharCNN-embeddings everywhere? That would be architecturally simpler than the split variant, and so may be of interest.\n- Page 6: When truncating to 500 nodes per graph: How many graphs in your dataset are larger than that?\n- Page 7: Do you really use attention over all nodes, instead of only nodes corresponding to variables? How do you deal with results where the model picks a non-variable (e.g., a corresponding cache node)? Does this happen?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1540993178220}], "openreview_url": "https://openreview.net/forum?id=SkNSehA9FQ", "arxiv_id": "1810.08305", "paper_pdf": "papers/SkNSehA9FQ.pdf", "paper_pdf_sha256": "04e967642c6def63ebea8e6c13e7cd2e54b2476abd66bc0f0877ad682f8b30c4", "paper_pdf_bytes": 1201854, "paper_pdf_source": "openreview", "code_url": "https://github.com/mwcvitkovic/Open-Vocabulary-Learning-on-Source-Code-with-a-Graph-Structured-Cache--Code-Preprocessor", "code_repository": "mwcvitkovic/Open-Vocabulary-Learning-on-Source-Code-with-a-Graph-Structured-Cache--Code-Preprocessor", "code_commit": "9ff7c8caa7eb536bb6c36602424820e095a0f838", "code_archive": "repos/SkNSehA9FQ.zip", "code_archive_sha256": "9d4466364cc4db425b0925b77524db5237f1283af34029cbe75a455857ce42a7", "code_archive_bytes": 4360675, "code_file_count": 1207, "code_extensions": {".java": 1202, ".sh": 4, ".js": 1}, "github_disk_usage_kb": 2846, "github_languages": {"Java": 6667739, "JavaScript": 4889, "Shell": 3280, "HTML": 1149}, "github_archived": false, "github_pushed_at": "2018-10-22T15:48:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/open-vocabulary-learning-on-source-code-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yAns9ibCaE", "year": 2026, "status": "rejected", "title": "ByteMorph: Benchmarking Instruction-Guided Image Editing with Non-Rigid Motions", "authors": ["Di Chang", "Mingdeng Cao", "Yichun Shi", "Bo Liu", "Shengqu Cai", "Shijie Zhou", "Weilin Huang", "Gordon Wetzstein", "Mohammad Soleymani", "Peng Wang"], "authorids": ["~Di_Chang1", "~Mingdeng_Cao1", "~Yichun_Shi1", "~Bo_Liu50", "~Shengqu_Cai1", "~Shijie_Zhou1", "~Weilin_Huang1", "~Gordon_Wetzstein3", "~Mohammad_Soleymani2", "~Peng_Wang2"], "authors_source": "OpenReview API", "abstract": "Editing images with instructions to reflect non-rigid motions—camera viewpoint shifts, object deformations, human articulations, and complex interactions—poses a challenging yet underexplored problem in computer vision. Existing approaches and datasets predominantly focus on static scenes or rigid transformations, limiting their capacity to handle expressive edits involving dynamic motion. To address this gap, we introduce ByteMorph, a comprehensive framework for instruction-based image editing with an emphasis on non-rigid motions. ByteMorph comprises a large-scale dataset, ByteMorph-6M, and a strong baseline model built upon the Diffusion Transformer (DiT), named ByteMorpher. ByteMorph-6M includes over 6 million high-resolution image editing pairs for training, along with a carefully curated evaluation benchmark ByteMorph-Bench. Both capture a wide variety of non-rigid motion types across diverse environments, human figures, and object categories. The dataset is constructed using motion-guided data generation, layered compositing techniques, and automated captioning to ensure diversity, realism, and semantic coherence. We further conduct a comprehensive evaluation of recent instruction-based image editing methods from both academic and commercial domains.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "QjcH0wAWX0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4196/Reviewer_S7Wu"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces ByteMorph, a comprehensive framework for instruction-guided image editing with a specific and novel focus on non-rigid motions. The authors identify a significant gap in existing research, where current datasets and models primarily excel at static, appearance-centric edits but fail to handle dynamic transformations such as camera viewpoint shifts, object deformations, and human articulations.", "review_text": "This paper introduces ByteMorph, a comprehensive framework for instruction-guided image editing with a specific and novel focus on non-rigid motions. The authors identify a significant gap in existing research, where current datasets and models primarily excel at static, appearance-centric edits but fail to handle dynamic transformations such as camera viewpoint shifts, object deformations, and human articulations.", "strengths": "1. The paper tackles a well-defined and high-impact limitation of current image editing models. Non-rigid motion is a fundamental aspect of the visual world, and enabling instruction-based control over it is very important.\n\n2.  Experimental results are comprehensive, which make this paper very solid.", "weaknesses": "My main concern is about the dataset construction. The entire training set (ByteMorph-6M) relies entirely on a synthetic pipeline (ChatGPT-4o and Seaweed).  I mean, 6.4M data comes from a single video generation model. This directly results in the dataset's quality being limited by the capabilities of this specific video model. It seems like that training on ByteMorph-6M is actually distilling Seawead. Specifically, if the Seawead model has systematic limitations in generating certain types of motion, the ByteMorpher model will inevitably learn these same flaws.", "questions": "1. About Table 4 of the Ablation Study. The authors claim notable gains after fine-tuning on the ByteMorph-6M, but this point is not reflected in the Table.  Could the authors give more explanation?\n\n2.  During the dataset construction, how to handle the generation failures? I mean, what if the video generation model (Seaweed) fails to follow the motion caption $C_m$? Is there any filter strategy?\n\n3. Seed Data Diversity. The pipeline in Figure 2 begins with a frame from real video, which acts like a seed for the dataset generation. Therefore, the diversity of the entire 6.4M dataset (like the number of different objects) is highly dependent on the diversity of this initial seed set. Could the authors provide more details on these initial real-world frames (like the number, variety)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ByteMorph, a comprehensive framework for instruction-guided image editing with a specific and novel focus on non-rigid motions. The authors identify a significant gap in existing research, where current datasets and models primarily excel at static, appearance-centric edits but fail to handle dynamic transformations such as camera viewpoint shifts, object deformations, and human articulations.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper tackles a well-defined and high-impact limitation of current image editing models. Non-rigid motion is a fundamental aspect of the visual world, and enabling instruction-based control over it is very important.\n\n2.  Experimental results are comprehensive, which make this paper very solid.", "weaknesses": "My main concern is about the dataset construction. The entire training set (ByteMorph-6M) relies entirely on a synthetic pipeline (ChatGPT-4o and Seaweed).  I mean, 6.4M data comes from a single video generation model. This directly results in the dataset's quality being limited by the capabilities of this specific video model. It seems like that training on ByteMorph-6M is actually distilling Seawead. Specifically, if the Seawead model has systematic limitations in generating certain types of motion, the ByteMorpher model will inevitably learn these same flaws.", "questions": "1. About Table 4 of the Ablation Study. The authors claim notable gains after fine-tuning on the ByteMorph-6M, but this point is not reflected in the Table.  Could the authors give more explanation?\n\n2.  During the dataset construction, how to handle the generation failures? I mean, what if the video generation model (Seaweed) fails to follow the motion caption $C_m$? Is there any filter strategy?\n\n3. Seed Data Diversity. The pipeline in Figure 2 begins with a frame from real video, which acts like a seed for the dataset generation. Therefore, the diversity of the entire 6.4M dataset (like the number of different objects) is highly dependent on the diversity of this initial seed set. Could the authors provide more details on these initial real-world frames (like the number, variety)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761924637220}, {"id": "79LgqTK4tl", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4196/Reviewer_2iZC"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper presents ByteMorph, a framework for instruction-based image editing with focus on non-rigid motions. The work introduces: (1) ByteMorph-6M - a large-scale dataset of 6.4 million image pairs covering with five motion categories, using an automated pipeline combining video generation and VLM-based annotation; (2) ByteMorph-Bench - a curated evaluation benchmark; and (3) ByteMorpher - a DiT-based model fine-tuned on FLUX.1-dev. The method demonstrates superior performance in dynamic editing tasks compared to existing approaches.", "review_text": "This paper presents ByteMorph, a framework for instruction-based image editing with focus on non-rigid motions. The work introduces: (1) ByteMorph-6M - a large-scale dataset of 6.4 million image pairs covering with five motion categories, using an automated pipeline combining video generation and VLM-based annotation; (2) ByteMorph-Bench - a curated evaluation benchmark; and (3) ByteMorpher - a DiT-based model fine-tuned on FLUX.1-dev. The method demonstrates superior performance in dynamic editing tasks compared to existing approaches.", "strengths": "1. The dataset is well-motivated. ByteMorph-6M effectively addresses the gap in motion-focused editing data with comprehensive coverage of non-rigid transformations. The release of datasets, benchmarks, and code provides significant value to the community.\n2. The automated construction using video generation and VLMs ensures scalability while maintaining quality and semantic coherence.", "weaknesses": "1. The technical contribution is limited, as the methodology relies purely on fine-tuning a pre-existing DiT model (FLUX.1-dev) on ByteMorph-6M without introducing architectural innovations or tailored optimizations.\n2. The experimental analysis does not sufficiently address the trade-offs of fine-tuning on ByteMorph-6M. While it is intuitive that specialization improves motion-specific performance, the paper omits evaluation of the model's original capabilities on standard instruction-based benchmarks. This raises concerns about potential performance degradation in general editing tasks, such as stylization or attribute modification, which could limit the model's applicability in broader contexts. A comparative analysis on conventional benchmarks would offer a more balanced perspective on robustness and generalization.\n3. Experiments are primarily conducted on curated synthetic data (ByteMorph-Bench), with insufficient validation on real images contains real-world complexities such as occlusions and lighting variations.", "questions": "1. Does fine-tuning on ByteMorph-6M compromise performance on standard instruction-based benchmarks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents ByteMorph, a framework for instruction-based image editing with focus on non-rigid motions. The work introduces: (1) ByteMorph-6M - a large-scale dataset of 6.4 million image pairs covering with five motion categories, using an automated pipeline combining video generation and VLM-based annotation; (2) ByteMorph-Bench - a curated evaluation benchmark; and (3) ByteMorpher - a DiT-based model fine-tuned on FLUX.1-dev. The method demonstrates superior performance in dynamic editing tasks compared to existing approaches.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The dataset is well-motivated. ByteMorph-6M effectively addresses the gap in motion-focused editing data with comprehensive coverage of non-rigid transformations. The release of datasets, benchmarks, and code provides significant value to the community.\n2. The automated construction using video generation and VLMs ensures scalability while maintaining quality and semantic coherence.", "weaknesses": "1. The technical contribution is limited, as the methodology relies purely on fine-tuning a pre-existing DiT model (FLUX.1-dev) on ByteMorph-6M without introducing architectural innovations or tailored optimizations.\n2. The experimental analysis does not sufficiently address the trade-offs of fine-tuning on ByteMorph-6M. While it is intuitive that specialization improves motion-specific performance, the paper omits evaluation of the model's original capabilities on standard instruction-based benchmarks. This raises concerns about potential performance degradation in general editing tasks, such as stylization or attribute modification, which could limit the model's applicability in broader contexts. A comparative analysis on conventional benchmarks would offer a more balanced perspective on robustness and generalization.\n3. Experiments are primarily conducted on curated synthetic data (ByteMorph-Bench), with insufficient validation on real images contains real-world complexities such as occlusions and lighting variations.", "questions": "1. Does fine-tuning on ByteMorph-6M compromise performance on standard instruction-based benchmarks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761894855596}, {"id": "v6RKqyYTBD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4196/Reviewer_KJoM"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces ByteMorph, a motion-centric framework for instruction-guided image editing with non-rigid motions. It contributes:\n- ByteMorph-6M: a large training set (~6.4M source–target pairs) synthesized from video generations and re-labeled with motion-focused instructions.\n- ByteMorph-Bench: a curated evaluation set (613 hard cases) covering five motion categories—camera zoom, camera move, object motion, human motion, and human–object interaction.\n- ByteMorpher: a DiT-based baseline initialized from FLUX.1-dev; training concatenates noisy source/target latents along the sequence dimension with shared positional encodings.\n\nFor evaluation, the paper proposes $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $, which compares *difference vectors* in CLIP space to better capture edit quality under camera changes.\n\nExperiments benchmark against open-source and commercial systems with both VLM-based and human evaluations, showing that ByteMorpher achieves competitive or superior performance, especially on motion-driven edits.", "review_text": "This paper introduces ByteMorph, a motion-centric framework for instruction-guided image editing with non-rigid motions. It contributes:\n- ByteMorph-6M: a large training set (~6.4M source–target pairs) synthesized from video generations and re-labeled with motion-focused instructions.\n- ByteMorph-Bench: a curated evaluation set (613 hard cases) covering five motion categories—camera zoom, camera move, object motion, human motion, and human–object interaction.\n- ByteMorpher: a DiT-based baseline initialized from FLUX.1-dev; training concatenates noisy source/target latents along the sequence dimension with shared positional encodings.\n\nFor evaluation, the paper proposes $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $, which compares *difference vectors* in CLIP space to better capture edit quality under camera changes.\n\nExperiments benchmark against open-source and commercial systems with both VLM-based and human evaluations, showing that ByteMorpher achieves competitive or superior performance, especially on motion-driven edits.", "strengths": "- Elevates motion (non-rigid, articulation, camera pose) as a first-class editing dimension; introduces $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $ to evaluate edits as *changes* rather than absolute content similarity.\n- Large-scale training set + curated hard benchmark; broad comparisons (open-source & industrial), repeated sampling, and both human and VLM judgments.\n- Clear problem framing and taxonomy; tables position prior editing datasets/methods effectively; training details (backbone, losses) are given.\n- Bridges a gap between instruction-guided appearance edits and motion-centric edits; likely to become a default benchmark for this subarea if maintained.", "weaknesses": "- Training relies on synthesized videos; even with filtering, motion/texture statistics may diverge from real-world photos. Add more purely real frame-pair data or zero/low-shot tests on real-image edit benchmarks to quantify the gap.\n- Provide a larger rank-correlation study (Spearman/Kendall) per category (camera/human/object/interaction) and analyze systematic failure modes (e.g., composite camera + articulation edits).\n- Since training uses latent concatenation (source, target), include a concise inference-time diagram/pseudocode showing where source features and text are injected when only (source, instruction) are available.\n- Report sensitivity to API sampling settings (resolution, steps, guidance/temperature, resampling) to demonstrate robustness of conclusions.\n- More explicitly connect with camera/motion-controlled T2V (e.g., camera trajectory parameterizations) to transfer motion priors into the image-editing setting.", "questions": "1. What are the Spearman/Kendall correlations between $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $ and human preferences across the five categories? Any systematic mismatches (e.g., simultaneous camera + expression changes)?\n2. How does the model perform on purely real editing datasets (e.g., MagicBrush/HQ-Edit) without synthetic video support?\n3. Sensitivity to the instruction templating and VLM labeling settings (temperature/prompt variants) during data construction?\n4. Interplay with InstructMove: Does joint training with real video-frame pairs improve generalization on ByteMorph-Bench and real-image edits?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ByteMorph, a motion-centric framework for instruction-guided image editing with non-rigid motions. It contributes:\n- ByteMorph-6M: a large training set (~6.4M source–target pairs) synthesized from video generations and re-labeled with motion-focused instructions.\n- ByteMorph-Bench: a curated evaluation set (613 hard cases) covering five motion categories—camera zoom, camera move, object motion, human motion, and human–object interaction.\n- ByteMorpher: a DiT-based baseline initialized from FLUX.1-dev; training concatenates noisy source/target latents along the sequence dimension with shared positional encodings.\n\nFor evaluation, the paper proposes $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $, which compares *difference vectors* in CLIP space to better capture edit quality under camera changes.\n\nExperiments benchmark against open-source and commercial systems with both VLM-based and human evaluations, showing that ByteMorpher achieves competitive or superior performance, especially on motion-driven edits.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Elevates motion (non-rigid, articulation, camera pose) as a first-class editing dimension; introduces $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $ to evaluate edits as *changes* rather than absolute content similarity.\n- Large-scale training set + curated hard benchmark; broad comparisons (open-source & industrial), repeated sampling, and both human and VLM judgments.\n- Clear problem framing and taxonomy; tables position prior editing datasets/methods effectively; training details (backbone, losses) are given.\n- Bridges a gap between instruction-guided appearance edits and motion-centric edits; likely to become a default benchmark for this subarea if maintained.", "weaknesses": "- Training relies on synthesized videos; even with filtering, motion/texture statistics may diverge from real-world photos. Add more purely real frame-pair data or zero/low-shot tests on real-image edit benchmarks to quantify the gap.\n- Provide a larger rank-correlation study (Spearman/Kendall) per category (camera/human/object/interaction) and analyze systematic failure modes (e.g., composite camera + articulation edits).\n- Since training uses latent concatenation (source, target), include a concise inference-time diagram/pseudocode showing where source features and text are injected when only (source, instruction) are available.\n- Report sensitivity to API sampling settings (resolution, steps, guidance/temperature, resampling) to demonstrate robustness of conclusions.\n- More explicitly connect with camera/motion-controlled T2V (e.g., camera trajectory parameterizations) to transfer motion priors into the image-editing setting.", "questions": "1. What are the Spearman/Kendall correlations between $ \\mathrm{CLIP\\text{-}D}_{\\text{img}} $ and human preferences across the five categories? Any systematic mismatches (e.g., simultaneous camera + expression changes)?\n2. How does the model perform on purely real editing datasets (e.g., MagicBrush/HQ-Edit) without synthetic video support?\n3. Sensitivity to the instruction templating and VLM labeling settings (temperature/prompt variants) during data construction?\n4. Interplay with InstructMove: Does joint training with real video-frame pairs improve generalization on ByteMorph-Bench and real-image edits?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761712659460}, {"id": "K9NV9DbT8M", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4196/Reviewer_8HRk"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper addresses the underexplored challenge of non-rigid motion editing in instruction-guided image manipulation by introducing ByteMorph, a comprehensive framework. Its core contributions include: a large-scale training dataset (ByteMorph-6M) with over 6 million high-quality editing pairs, a carefully curated evaluation benchmark (ByteMorph-Bench), and a strong DiT-based baseline model (ByteMorpher). Using motion-guided generation and automated captioning pipelines, the dataset captures diverse dynamic scenarios including camera motion, object deformation and human-object interactions. Experiments demonstrate that ByteMorpher finetuned on this dataset significantly outperforms existing open-source methods, while also revealing the limitations of current industrial foundation models in handling such complex motion-aware edits.", "review_text": "This paper addresses the underexplored challenge of non-rigid motion editing in instruction-guided image manipulation by introducing ByteMorph, a comprehensive framework. Its core contributions include: a large-scale training dataset (ByteMorph-6M) with over 6 million high-quality editing pairs, a carefully curated evaluation benchmark (ByteMorph-Bench), and a strong DiT-based baseline model (ByteMorpher). Using motion-guided generation and automated captioning pipelines, the dataset captures diverse dynamic scenarios including camera motion, object deformation and human-object interactions. Experiments demonstrate that ByteMorpher finetuned on this dataset significantly outperforms existing open-source methods, while also revealing the limitations of current industrial foundation models in handling such complex motion-aware edits.", "strengths": "- This paper tackles the challenging problem of spatial manipulation in image editing, and the model trained on the proposed dataset demonstrates convincing performance in handling such tasks, making the research direction highly meaningful.\n- The authors have constructed an exceptionally large-scale dataset of 6M samples and developed a comprehensive and well-designed benchmark, demonstrating substantial research effort.\n- The paper is well-written and clearly structured, with professionally designed figures and illustrations.", "weaknesses": "- While the use of video generation models for dataset construction is understandable, the reliance on Seaweed's generative capabilities raises concerns about whether the editing performance can generalize effectively to other editing tasks beyond the ByteMorph-Bench evaluation.\n- Beyond fine-tuning the FLUX model to demonstrate the dataset's effectiveness, for such a challenging task, it would be valuable to additionally validate the dataset's utility by fine-tuning unified models such as Bagel, UniWorld, or Qwen-Image-Edit. Furthermore, it remains unclear how effectively the 6M dataset would support more complex instruction-based editing scenarios. The authors are encouraged to provide preliminary experimental results addressing these aspects.\n- Regarding related work, the authors may consider citing \"Unireal: Universal image generation and editing via learning real-world dynamics\" as additional relevant literature.", "questions": "My current concerns primarily stem from the weaknesses outlined above. Additionally, I have reviewed 3-4 contemporaneous papers following a similar paradigm - focusing on complex image editing and utilizing video models for data construction. While the existence of parallel work alone does not justify rejection, the weaknesses identified prevent me from assigning a higher score at this stage. I will reconsider my rating based on the authors' rebuttal and other reviewers' comments.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the underexplored challenge of non-rigid motion editing in instruction-guided image manipulation by introducing ByteMorph, a comprehensive framework. Its core contributions include: a large-scale training dataset (ByteMorph-6M) with over 6 million high-quality editing pairs, a carefully curated evaluation benchmark (ByteMorph-Bench), and a strong DiT-based baseline model (ByteMorpher). Using motion-guided generation and automated captioning pipelines, the dataset captures diverse dynamic scenarios including camera motion, object deformation and human-object interactions. Experiments demonstrate that ByteMorpher finetuned on this dataset significantly outperforms existing open-source methods, while also revealing the limitations of current industrial foundation models in handling such complex motion-aware edits.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- This paper tackles the challenging problem of spatial manipulation in image editing, and the model trained on the proposed dataset demonstrates convincing performance in handling such tasks, making the research direction highly meaningful.\n- The authors have constructed an exceptionally large-scale dataset of 6M samples and developed a comprehensive and well-designed benchmark, demonstrating substantial research effort.\n- The paper is well-written and clearly structured, with professionally designed figures and illustrations.", "weaknesses": "- While the use of video generation models for dataset construction is understandable, the reliance on Seaweed's generative capabilities raises concerns about whether the editing performance can generalize effectively to other editing tasks beyond the ByteMorph-Bench evaluation.\n- Beyond fine-tuning the FLUX model to demonstrate the dataset's effectiveness, for such a challenging task, it would be valuable to additionally validate the dataset's utility by fine-tuning unified models such as Bagel, UniWorld, or Qwen-Image-Edit. Furthermore, it remains unclear how effectively the 6M dataset would support more complex instruction-based editing scenarios. The authors are encouraged to provide preliminary experimental results addressing these aspects.\n- Regarding related work, the authors may consider citing \"Unireal: Universal image generation and editing via learning real-world dynamics\" as additional relevant literature.", "questions": "My current concerns primarily stem from the weaknesses outlined above. Additionally, I have reviewed 3-4 contemporaneous papers following a similar paradigm - focusing on complex image editing and utilizing video models for data construction. While the existence of parallel work alone does not justify rejection, the weaknesses identified prevent me from assigning a higher score at this stage. I will reconsider my rating based on the authors' rebuttal and other reviewers' comments.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761645139905}], "openreview_url": "https://openreview.net/forum?id=yAns9ibCaE", "arxiv_id": "2506.03107", "paper_pdf": "papers/yAns9ibCaE.pdf", "paper_pdf_sha256": "8b1a59096f47982b6b5b64a3fdb0a552a6f7ea60ce7add91aba240d8658f0fa1", "paper_pdf_bytes": 24503894, "paper_pdf_source": "openreview", "code_url": "https://github.com/ByteDance-Seed/BM-code", "code_repository": "ByteDance-Seed/BM-code", "code_commit": "594d54346db33fbce09bd435e77232a32b26e616", "code_archive": "repos/yAns9ibCaE.zip", "code_archive_sha256": "c34a6d6f7ed9cacfd8741b27bb143cffae761bb49875b67229d1dfcfa8e8de86", "code_archive_bytes": 1232953, "code_file_count": 154, "code_extensions": {".py": 110, ".java": 19, ".sh": 13, ".swift": 11, ".cpp": 1}, "github_disk_usage_kb": 1076, "github_languages": {"Python": 721377, "Java": 113288, "Swift": 58465, "C++": 9891, "CMake": 6607, "Shell": 5731, "Dockerfile": 922, "Ruby": 408}, "github_archived": false, "github_pushed_at": "2025-06-11T06:14:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bytemorph-benchmarking-instruction-guided"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cho9iE9POr", "year": 2025, "status": "rejected", "title": "Low-Budget Simulation-Based Inference with Bayesian Neural Networks", "authors": ["Arnaud Delaunoy", "Maxence de la Brassinne Bonardeaux", "Siddharth Mishra-Sharma", "Gilles Louppe"], "authorids": ["~Arnaud_Delaunoy1", "~Maxence_de_la_Brassinne_Bonardeaux1", "~Siddharth_Mishra-Sharma1", "~Gilles_Louppe1"], "authors_source": "OpenReview API", "abstract": "Simulation-based inference methods have been shown to be inaccurate in the data-poor regime, when training simulations are limited or expensive.\nUnder these circumstances, the inference network is particularly prone to overfitting, and using it without accounting for the computational uncertainty arising from the lack of identifiability of the network weights can lead to unreliable results.\nTo address this issue, we propose using Bayesian neural networks in low-budget simulation-based inference, thereby explicitly accounting for the computational uncertainty of the posterior approximation.\nWe design a family of Bayesian neural network priors that are tailored for inference and show that they lead to well-calibrated posteriors on tested benchmarks, even when as few as $O(10)$ simulations are available.\nThis opens up the possibility of performing reliable simulation-based inference using very expensive simulators, as we demonstrate on a problem from the field of cosmology where single simulations are computationally expensive. We show that Bayesian neural networks produce informative and well-calibrated posterior estimates with only a few hundred simulations.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "OIAmH6yyQL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7101/Reviewer_qwfa"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "The paper intodruces a new method for simulation-based inference (SBI) that uses\nBayesian neural networks (BNN) to better account for uncertainty in posterior estimates\ndue to limited training data. The authors demonstrate that previous attempts of\ncombining BNN with SBI showed limited success due to the choice of prior on the BNN\nweights and they propose an alternative prior more suitable for the SBI setting. To\nevaluate the proposed method, the authors show on four benchmarking tasks and on a SBI\nuse-case from cosmology that the resulting SBI posterior estimates are well-calibrated\neven in the low-data regime.", "review_text": "The paper intodruces a new method for simulation-based inference (SBI) that uses\nBayesian neural networks (BNN) to better account for uncertainty in posterior estimates\ndue to limited training data. The authors demonstrate that previous attempts of\ncombining BNN with SBI showed limited success due to the choice of prior on the BNN\nweights and they propose an alternative prior more suitable for the SBI setting. To\nevaluate the proposed method, the authors show on four benchmarking tasks and on a SBI\nuse-case from cosmology that the resulting SBI posterior estimates are well-calibrated\neven in the low-data regime.", "strengths": "### Originality\n\nThe idea of using BNN for SBI is not new (as discussed in the paper). However, showing\nwhy previous approaches did not work so well and the proposal of a new type of BNN prior\nmore suitable for the SBI setting is a valuable contribution. Most of the previous work\non better uncertainty quantification in SBI is discussed adequately in the introduction.\nThe only paper with a related approach that seems to be missing in the discussion is [Lueckmann et al.\n2017, section\n2.2](https://proceedings.neurips.cc/paper/2017/hash/addfa9b7e234254d26e9c7f2af1005cb-Abstract.html),\nwho propose performing SVI on the neural network weights for continual learning across\nSNPE rounds. Also, it seems a bit odd to me that the benchmark used are not cited\n(except for the spatial SIR by Hermans et al.). The SLCP benchmark was introduced by\nPapamakarios et al. (SNLE paper), two moons by Greenberg et al.; and the overall set of\nSBI benchmarking tasks was introduced in Lueckmann et al. 2021.\n\n### Quality\n\nThe technical contributions of the paper appear as sound and the selected methods for\napplying BNN to SBI are appropriate. The experimental results tend to support the\ninitial claim of obtaining well-calibrated posteriors in low-data regime. However, the\nnumber of performed experiments is quite low (three) and the results appear quite noisy.\n\nIn general, it should be made clearer how strong the trade-off is between posterior\ncalibration and posterior accuracy. At the moment it seems that the BNN approach tends\nto be quite underconfident in some scenarios (e.g., BNN-NRE on all benchmarks). I\ntherefore suggest that more experiments should be run and results should be reported\nwith error bars (e.g., 5-10 runs with error bars showing the standard error of the mean\nperformance). Additionally, I think it would be good to also check the posterior\npredictive distribution for the cases where the BNN has low nominal log posterior values\neven in high-data regimes.\n\n### Clarity\n\nOverall, the paper is well-written and clearly-structured. I have a couple of remarks\nthat should be addressed to improve the clarity.\n\n- In section 3, especially in section 3.2, it should be explained more clearly why and\n  how the prior has to be tuned to be used for the BNN approach. In the classical SBI\n  setting, the prior is usually set a priori using expert knowledge. This step usually\n  does not involve using simulated data for an optimization procedure. In the BNN\n  approach, it seems that the prior has to be tuned with actual simulations (e.g., lines\n  254-261). Do these simulations have to be run in addition to the training data? Are\n  they accounted for the in the budgets in the benchmarks?\n- Figure 1 needs a couple of clarifications. The caption says it's a visualization of\n  the tuned prior. Then, the next sentence says the left panels show posterior functions\n  sampled from the *tuned prior prior over the neural network's weights*. How are these\n  samples obtained and what are thet supposed to show? Are they obtained before or after\n  SBI training? Similarly question for the last panel: is the calibration calculated\n  after the full SBI training or only after the prior tuning?\n\nIn general, it seems that a better explanation of the steps involved in setting up and\ntraining the BNN-NPE (NRE) approach. I suggest adding more details in the next and\nadding an algorithm scheme in the text or appendix.\n\n### Significance\n\nUncertainty quantification in neural SBI methods on the posterior approximation itself\nis an important and timely problem. Especially in low-data regimes, common neural SBI\nmethods like NPE struggle. The proposed BNN approach with a Gaussian process prior\nas proposed here appears as an important contribution for addressing this issue.", "weaknesses": "I outlined several concerns and questions above. To summarize to most important points:\n\n- the experimental results are difficult to interpret because they show the media of\n  only three repetitions. More repetitions and error bars would be better here.\n- the high underconfidence of the BNN approach in how-data regimes is concerning.\n  Additional evaluation of the posterior predictive distributions would be appropriate.\n- the choice and construction of the prior from simulated should be explained more\n  clearly. A better explanation ideally will resolve the questions on the general\n  procedure and on Figure 1 above.", "questions": "1) Do these simulations for prior tuning have to be run in addition to the training data?\n2) Are they accounted for the in the budgets in the benchmarks?\n3) NPE / NRE ensembles seems to perform similarly well compared to the BNN approach. How\n  does it perform on the cosmology example? How does it compare to the BNN approach in\n  terms of computational budget? Do you think ensembles could be a good alternative when\n  computational budgets are limited? What are the disadvantages of ensembles compared to\n  BNN for SBI?\n4) More generally, what is the computational complexity of the BNN approach compared to\n  NPE (ensembles)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper intodruces a new method for simulation-based inference (SBI) that uses\nBayesian neural networks (BNN) to better account for uncertainty in posterior estimates\ndue to limited training data. The authors demonstrate that previous attempts of\ncombining BNN with SBI showed limited success due to the choice of prior on the BNN\nweights and they propose an alternative prior more suitable for the SBI setting. To\nevaluate the proposed method, the authors show on four benchmarking tasks and on a SBI\nuse-case from cosmology that the resulting SBI posterior estimates are well-calibrated\neven in the low-data regime.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "### Originality\n\nThe idea of using BNN for SBI is not new (as discussed in the paper). However, showing\nwhy previous approaches did not work so well and the proposal of a new type of BNN prior\nmore suitable for the SBI setting is a valuable contribution. Most of the previous work\non better uncertainty quantification in SBI is discussed adequately in the introduction.\nThe only paper with a related approach that seems to be missing in the discussion is [Lueckmann et al.\n2017, section\n2.2](https://proceedings.neurips.cc/paper/2017/hash/addfa9b7e234254d26e9c7f2af1005cb-Abstract.html),\nwho propose performing SVI on the neural network weights for continual learning across\nSNPE rounds. Also, it seems a bit odd to me that the benchmark used are not cited\n(except for the spatial SIR by Hermans et al.). The SLCP benchmark was introduced by\nPapamakarios et al. (SNLE paper), two moons by Greenberg et al.; and the overall set of\nSBI benchmarking tasks was introduced in Lueckmann et al. 2021.\n\n### Quality\n\nThe technical contributions of the paper appear as sound and the selected methods for\napplying BNN to SBI are appropriate. The experimental results tend to support the\ninitial claim of obtaining well-calibrated posteriors in low-data regime. However, the\nnumber of performed experiments is quite low (three) and the results appear quite noisy.\n\nIn general, it should be made clearer how strong the trade-off is between posterior\ncalibration and posterior accuracy. At the moment it seems that the BNN approach tends\nto be quite underconfident in some scenarios (e.g., BNN-NRE on all benchmarks). I\ntherefore suggest that more experiments should be run and results should be reported\nwith error bars (e.g., 5-10 runs with error bars showing the standard error of the mean\nperformance). Additionally, I think it would be good to also check the posterior\npredictive distribution for the cases where the BNN has low nominal log posterior values\neven in high-data regimes.\n\n### Clarity\n\nOverall, the paper is well-written and clearly-structured. I have a couple of remarks\nthat should be addressed to improve the clarity.\n\n- In section 3, especially in section 3.2, it should be explained more clearly why and\n  how the prior has to be tuned to be used for the BNN approach. In the classical SBI\n  setting, the prior is usually set a priori using expert knowledge. This step usually\n  does not involve using simulated data for an optimization procedure. In the BNN\n  approach, it seems that the prior has to be tuned with actual simulations (e.g., lines\n  254-261). Do these simulations have to be run in addition to the training data? Are\n  they accounted for the in the budgets in the benchmarks?\n- Figure 1 needs a couple of clarifications. The caption says it's a visualization of\n  the tuned prior. Then, the next sentence says the left panels show posterior functions\n  sampled from the *tuned prior prior over the neural network's weights*. How are these\n  samples obtained and what are thet supposed to show? Are they obtained before or after\n  SBI training? Similarly question for the last panel: is the calibration calculated\n  after the full SBI training or only after the prior tuning?\n\nIn general, it seems that a better explanation of the steps involved in setting up and\ntraining the BNN-NPE (NRE) approach. I suggest adding more details in the next and\nadding an algorithm scheme in the text or appendix.\n\n### Significance\n\nUncertainty quantification in neural SBI methods on the posterior approximation itself\nis an important and timely problem. Especially in low-data regimes, common neural SBI\nmethods like NPE struggle. The proposed BNN approach with a Gaussian process prior\nas proposed here appears as an important contribution for addressing this issue.", "weaknesses": "I outlined several concerns and questions above. To summarize to most important points:\n\n- the experimental results are difficult to interpret because they show the media of\n  only three repetitions. More repetitions and error bars would be better here.\n- the high underconfidence of the BNN approach in how-data regimes is concerning.\n  Additional evaluation of the posterior predictive distributions would be appropriate.\n- the choice and construction of the prior from simulated should be explained more\n  clearly. A better explanation ideally will resolve the questions on the general\n  procedure and on Figure 1 above.", "questions": "1) Do these simulations for prior tuning have to be run in addition to the training data?\n2) Are they accounted for the in the budgets in the benchmarks?\n3) NPE / NRE ensembles seems to perform similarly well compared to the BNN approach. How\n  does it perform on the cosmology example? How does it compare to the BNN approach in\n  terms of computational budget? Do you think ensembles could be a good alternative when\n  computational budgets are limited? What are the disadvantages of ensembles compared to\n  BNN for SBI?\n4) More generally, what is the computational complexity of the BNN approach compared to\n  NPE (ensembles)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730712824466}, {"id": "ywXzK4wUV3", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7101/Reviewer_VQN7"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The authors propose a novel approach for simulation-based inference for low-budget problems using Bayesian neural networks. They additionally propose to use tailored priors for the neural network weights which generally leads to more conservative posterior estimates. They evaluate their method on several experimental models against multiple baselines demonstrating promising results for SBI when the computational budget is limited.", "review_text": "The authors propose a novel approach for simulation-based inference for low-budget problems using Bayesian neural networks. They additionally propose to use tailored priors for the neural network weights which generally leads to more conservative posterior estimates. They evaluate their method on several experimental models against multiple baselines demonstrating promising results for SBI when the computational budget is limited.", "strengths": "- The paper proposes a novel contribution for simulation-based inference. The method is in my opinion very relevant to the community and interesting as a low-budget solution for inferential problems.\n- The motivation and derivation of the tailored prior is in my opinion particularly well done and convincing.\n- Empirically, the method seems to achieve the motivated goal: having calibrated posteriors for low simulation budgets.\n- The paper is well written and easy to follow.", "weaknesses": "The evaluations and presentations of the results could in my opinion be improved.\n\n- The authors evaluate their method with the expected coverage (EC). In Figure 2, it is in my opinion very difficult to draw conclusions which methods works best. This is likely due to the fact that only 3 runs have been evaluated. Given the complexity of SBI and the high variance of the inferential results, I think they should do at least 10 evaluations and report these.\n- The authors use as a second evaluation metric the expected posterior log density (EPLD). It is not clear what a good value should look like here or even if a high log density should be desired. When the goal is to have a conservative posterior, is a extremely high EPLD a good measure? The authors should, despite possible drawbacks, have used some divergence to the true posterior, such as MMD, in addition to the other two and report this.\n- It is not clear how crucial hyperparameters such as the temperature $T$ or the covariance functions of the reference GP are chosen in practice. \n- The authors propose to use cold posteriors, e.g., using a temperature of $T=0.01$. The motivation of this is not clear, given that the entire idea is to use a prior that enforces calibration. Why would it be desirable to in fact reduce the impact of the prior? \n\n#### Minor\n- Figures 2 should show the standard errors.\n- Figure 3 should show the reference posterior and not only the true parameter value.", "questions": "- The results in Figure 3 seem to indicate that NPE is indeed outperforming BNN-NPE here. Would the authors argue that BNN-NPE is preferable over NPE for this task?\n- What do the numbers in the legend in Figure 3 mean?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a novel approach for simulation-based inference for low-budget problems using Bayesian neural networks. They additionally propose to use tailored priors for the neural network weights which generally leads to more conservative posterior estimates. They evaluate their method on several experimental models against multiple baselines demonstrating promising results for SBI when the computational budget is limited.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper proposes a novel contribution for simulation-based inference. The method is in my opinion very relevant to the community and interesting as a low-budget solution for inferential problems.\n- The motivation and derivation of the tailored prior is in my opinion particularly well done and convincing.\n- Empirically, the method seems to achieve the motivated goal: having calibrated posteriors for low simulation budgets.\n- The paper is well written and easy to follow.", "weaknesses": "The evaluations and presentations of the results could in my opinion be improved.\n\n- The authors evaluate their method with the expected coverage (EC). In Figure 2, it is in my opinion very difficult to draw conclusions which methods works best. This is likely due to the fact that only 3 runs have been evaluated. Given the complexity of SBI and the high variance of the inferential results, I think they should do at least 10 evaluations and report these.\n- The authors use as a second evaluation metric the expected posterior log density (EPLD). It is not clear what a good value should look like here or even if a high log density should be desired. When the goal is to have a conservative posterior, is a extremely high EPLD a good measure? The authors should, despite possible drawbacks, have used some divergence to the true posterior, such as MMD, in addition to the other two and report this.\n- It is not clear how crucial hyperparameters such as the temperature $T$ or the covariance functions of the reference GP are chosen in practice. \n- The authors propose to use cold posteriors, e.g., using a temperature of $T=0.01$. The motivation of this is not clear, given that the entire idea is to use a prior that enforces calibration. Why would it be desirable to in fact reduce the impact of the prior? \n\n#### Minor\n- Figures 2 should show the standard errors.\n- Figure 3 should show the reference posterior and not only the true parameter value.", "questions": "- The results in Figure 3 seem to indicate that NPE is indeed outperforming BNN-NPE here. Would the authors argue that BNN-NPE is preferable over NPE for this task?\n- What do the numbers in the legend in Figure 3 mean?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730477681165}, {"id": "GQoR2yNRPc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7101/Reviewer_9hZV"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "The authors aim to produce well-calibrated posterior in simulation-based inference, in particular in cases where the simulator is very expensive and, therefore, few simulations can be used as training data. To achieve this, the authors propose to use Bayesian neural networks. They develop a novel prior which leads to a-priori well-calibrated posteriors. They apply their method to several toy simulators and to a physics simulator.", "review_text": "The authors aim to produce well-calibrated posterior in simulation-based inference, in particular in cases where the simulator is very expensive and, therefore, few simulations can be used as training data. To achieve this, the authors propose to use Bayesian neural networks. They develop a novel prior which leads to a-priori well-calibrated posteriors. They apply their method to several toy simulators and to a physics simulator.", "strengths": "The idea to generate neural network weight priors which lead to a-priori well-calibrated posteriors is novel, interesting, and potentially impactful. The methodology which the authors employ to achieve this is novel, elegant, and rigorous. I thoroughly enjoyed reading these parts of the paper. In addition, the authors demonstrate that the method can be applied across methods (NPE & NRE) and they evaluate the method on a series of useful tasks.", "weaknesses": "(1) The paper overstates its claims.\nMy main issue with this paper is that it overstates its claims and does not acknowledge the weakness of empirical results. Looking at Figure 2, and in particular for NRE: BNN-NRE _never_ reaches the log posterior of the other methods (even for 1M simulations). Yet the authors state that `the nominal log posterior density is on par with other methods for very high simulation budgets`. Where does this claim come from. Similarly, in Figure 2 (NRE), the authors do not acknowledge that the AUC is not particularly much higher for BNN-NRE as compared to NRE, even for 10 simulations. Indeed, for three of the four tasks NRE has a higher coverage AUC than BNN-NRE for 10 simulations. On top of this, for many tasks and simulation budgets, BNN-NRE does have a negative AUC, but the authors simply claim that they `show positive coverage AUC`. Finally, \n\n(2) The empirical results are weak.\nSecond, to me, the empirical results are difficult to interpret. (A) Many of the curves shown in Figure 2 are very noisy and irregular. It might be beneficial to average across more seeds to observe clear trends. (B) As a potential user, I would find it very worrisome that BNN-NRE has _significantly_ lower nominal log-posterior than standard methods. Indeed, for some tasks, it seems to require 5 orders of magnitude (for many other tasks 3 orders of magnitude) more simulations than standard methods.", "questions": "No questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors aim to produce well-calibrated posterior in simulation-based inference, in particular in cases where the simulator is very expensive and, therefore, few simulations can be used as training data. To achieve this, the authors propose to use Bayesian neural networks. They develop a novel prior which leads to a-priori well-calibrated posteriors. They apply their method to several toy simulators and to a physics simulator.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The idea to generate neural network weight priors which lead to a-priori well-calibrated posteriors is novel, interesting, and potentially impactful. The methodology which the authors employ to achieve this is novel, elegant, and rigorous. I thoroughly enjoyed reading these parts of the paper. In addition, the authors demonstrate that the method can be applied across methods (NPE & NRE) and they evaluate the method on a series of useful tasks.", "weaknesses": "(1) The paper overstates its claims.\nMy main issue with this paper is that it overstates its claims and does not acknowledge the weakness of empirical results. Looking at Figure 2, and in particular for NRE: BNN-NRE _never_ reaches the log posterior of the other methods (even for 1M simulations). Yet the authors state that `the nominal log posterior density is on par with other methods for very high simulation budgets`. Where does this claim come from. Similarly, in Figure 2 (NRE), the authors do not acknowledge that the AUC is not particularly much higher for BNN-NRE as compared to NRE, even for 10 simulations. Indeed, for three of the four tasks NRE has a higher coverage AUC than BNN-NRE for 10 simulations. On top of this, for many tasks and simulation budgets, BNN-NRE does have a negative AUC, but the authors simply claim that they `show positive coverage AUC`. Finally, \n\n(2) The empirical results are weak.\nSecond, to me, the empirical results are difficult to interpret. (A) Many of the curves shown in Figure 2 are very noisy and irregular. It might be beneficial to average across more seeds to observe clear trends. (B) As a potential user, I would find it very worrisome that BNN-NRE has _significantly_ lower nominal log-posterior than standard methods. Indeed, for some tasks, it seems to require 5 orders of magnitude (for many other tasks 3 orders of magnitude) more simulations than standard methods.", "questions": "No questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729850658529}, {"id": "VUI3FskLui", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7101/Reviewer_WoUh"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "The authors wish to provide a method for simulation-based inference for expensive simulators that accounts for the epistemic uncertainty, which tends to be high when available data (simulations) is low. They wish to do so by using a Bayesian neural network in combination with the SBI methods NRE and NPE. They note issue with current priors for Bayesian neural networks in simulation based inference, stating they are not calibrated a priori. They note that using a Gaussian process, with the prior distribution as its mean, can be used to define a prior which is well-calibrated a priori. By optimizing the Bayesian networks prior on the weights to approximate the Gaussian process prior, the Bayesian networks prior becomes approximately well-calibrated. They perform experiments, and show this often leads to better calibrated posteriors in the low-data (low simulation budget) regime.", "review_text": "The authors wish to provide a method for simulation-based inference for expensive simulators that accounts for the epistemic uncertainty, which tends to be high when available data (simulations) is low. They wish to do so by using a Bayesian neural network in combination with the SBI methods NRE and NPE. They note issue with current priors for Bayesian neural networks in simulation based inference, stating they are not calibrated a priori. They note that using a Gaussian process, with the prior distribution as its mean, can be used to define a prior which is well-calibrated a priori. By optimizing the Bayesian networks prior on the weights to approximate the Gaussian process prior, the Bayesian networks prior becomes approximately well-calibrated. They perform experiments, and show this often leads to better calibrated posteriors in the low-data (low simulation budget) regime.", "strengths": "Being able to perform reliable simulation-based inference with fewer simulations would be broadly useful to the scientific community. The idea to use a prior over the weights of a BNN that is compatible with the simulators prior distribution is sensible, and using a Gaussian process to achieve this is to the best of my knowledge novel and clever. The paper is well structured overall (except for where noted below) and includes some interesting experiments, using a good choice of metrics assessing reliability of the posteriors.", "weaknesses": "Structure/Presentation:\n- Lack of method summary: There's no clear summary of the method, e.g. in the contributions section, an algorithm, or the start of section 3. A brief overview of 3-4 sentences would be very beneficial for readers. It's only towards the end of section 3.2, that the overarching method begins to come together on a first read. In my opinion, the abstract also does not include enough information on the method.\n- Lack of clear definitions: The concept of an \"a priori-calibrated Bayesian model\" should be defined more clearly when it is first introduced.  I presume that an example of a well calibrated a priori model, would be one for which the Bayesian model average at initialization is equal to the simulator parameters prior $p(\\theta)$ for any $x \\in \\mathcal{X}$? Presumably this includes a large variety of useful and non-useful models for modelling epistemic uncertainty (as suggested in equation 8).\n- Figures: The credibility figure in Figure 1 is confusing, either the standard deviations are not in order, or there is a typo one of the standard deviations. Figure 3 should have a labeled legend. Many of the font sizes are too small.\n\nExperiments:\n- Limited experimental runs: Only 3 runs are performed for each experiment, and only the median is reported. This makes it hard to assess if differences between methods are significant. Repeating with more runs would be beneficial, although I am sympathetic to limitations in computational budget.\n- The BNN-NRE method appears to perform worse than NRE in terms of mass placed on true parameters, even for low simulation budgets (100-1000 in Figure 2), which is the domain where the method is proposed to be beneficial. Similarly, the results for the NPE case are not particularly convincing in terms of the mass placed on the true parameters. The coverage properties do appear to have improved, but coverage alone is not indicative of a good posterior estimate.\n- The results don't align with an intuitive understanding of epistemic uncertainty. For example, in figure 3, in NPE, we can see 4000 simulations produces a reasonable posterior estimate, whereas BNN-NPE is still massively conservative, even with 65536 simulations, suggesting overestimation of epistemic uncertainty. This suggests simple alternative approaches such as training posterior estimates on subsets of the data with standard methods, such as NPE, and mixing the resulting posteriors, would likely be more effective. I understand the temperature parameter is introduced to limit this problem, but this then introduces another hard to choose hyperparameter (along with the Gaussian process parameters introduced).", "questions": "How can we be sure that any benefits are from the BNN modelling epistemic uncertainty, and not the altered initialization? For example, if we \"pretrained\" NPE/NRE to prior samples (and $x$ simulated from some noise distribution), such that at \"initialization\" the posterior estimate would be approximately equal to the prior for any $x$, it would be interesting to see if there is still be any benefit to the introduced method.\n\nMost the benefits seem to be in the very small data regime (<100 simulations). I am not convinced that this scenario is of particular interest the scientific community, could you give an example of a practical case where this is useful?\n\nWhy is the convergence to the NPE solution so slow in terms of the number of simulations? Can the prior be altered to avoid this problem, rather than introducing a temperature parameter?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors wish to provide a method for simulation-based inference for expensive simulators that accounts for the epistemic uncertainty, which tends to be high when available data (simulations) is low. They wish to do so by using a Bayesian neural network in combination with the SBI methods NRE and NPE. They note issue with current priors for Bayesian neural networks in simulation based inference, stating they are not calibrated a priori. They note that using a Gaussian process, with the prior distribution as its mean, can be used to define a prior which is well-calibrated a priori. By optimizing the Bayesian networks prior on the weights to approximate the Gaussian process prior, the Bayesian networks prior becomes approximately well-calibrated. They perform experiments, and show this often leads to better calibrated posteriors in the low-data (low simulation budget) regime.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "Being able to perform reliable simulation-based inference with fewer simulations would be broadly useful to the scientific community. The idea to use a prior over the weights of a BNN that is compatible with the simulators prior distribution is sensible, and using a Gaussian process to achieve this is to the best of my knowledge novel and clever. The paper is well structured overall (except for where noted below) and includes some interesting experiments, using a good choice of metrics assessing reliability of the posteriors.", "weaknesses": "Structure/Presentation:\n- Lack of method summary: There's no clear summary of the method, e.g. in the contributions section, an algorithm, or the start of section 3. A brief overview of 3-4 sentences would be very beneficial for readers. It's only towards the end of section 3.2, that the overarching method begins to come together on a first read. In my opinion, the abstract also does not include enough information on the method.\n- Lack of clear definitions: The concept of an \"a priori-calibrated Bayesian model\" should be defined more clearly when it is first introduced.  I presume that an example of a well calibrated a priori model, would be one for which the Bayesian model average at initialization is equal to the simulator parameters prior $p(\\theta)$ for any $x \\in \\mathcal{X}$? Presumably this includes a large variety of useful and non-useful models for modelling epistemic uncertainty (as suggested in equation 8).\n- Figures: The credibility figure in Figure 1 is confusing, either the standard deviations are not in order, or there is a typo one of the standard deviations. Figure 3 should have a labeled legend. Many of the font sizes are too small.\n\nExperiments:\n- Limited experimental runs: Only 3 runs are performed for each experiment, and only the median is reported. This makes it hard to assess if differences between methods are significant. Repeating with more runs would be beneficial, although I am sympathetic to limitations in computational budget.\n- The BNN-NRE method appears to perform worse than NRE in terms of mass placed on true parameters, even for low simulation budgets (100-1000 in Figure 2), which is the domain where the method is proposed to be beneficial. Similarly, the results for the NPE case are not particularly convincing in terms of the mass placed on the true parameters. The coverage properties do appear to have improved, but coverage alone is not indicative of a good posterior estimate.\n- The results don't align with an intuitive understanding of epistemic uncertainty. For example, in figure 3, in NPE, we can see 4000 simulations produces a reasonable posterior estimate, whereas BNN-NPE is still massively conservative, even with 65536 simulations, suggesting overestimation of epistemic uncertainty. This suggests simple alternative approaches such as training posterior estimates on subsets of the data with standard methods, such as NPE, and mixing the resulting posteriors, would likely be more effective. I understand the temperature parameter is introduced to limit this problem, but this then introduces another hard to choose hyperparameter (along with the Gaussian process parameters introduced).", "questions": "How can we be sure that any benefits are from the BNN modelling epistemic uncertainty, and not the altered initialization? For example, if we \"pretrained\" NPE/NRE to prior samples (and $x$ simulated from some noise distribution), such that at \"initialization\" the posterior estimate would be approximately equal to the prior for any $x$, it would be interesting to see if there is still be any benefit to the introduced method.\n\nMost the benefits seem to be in the very small data regime (<100 simulations). I am not convinced that this scenario is of particular interest the scientific community, could you give an example of a practical case where this is useful?\n\nWhy is the convergence to the NPE solution so slow in terms of the number of simulations? Can the prior be altered to avoid this problem, rather than introducing a temperature parameter?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1729084300586}], "openreview_url": "https://openreview.net/forum?id=cho9iE9POr", "arxiv_id": "2408.15136", "paper_pdf": "papers/cho9iE9POr.pdf", "paper_pdf_sha256": "b2e269634625109e5d3db155a64f8922c43bab9237b589e7500334aeb9361f66", "paper_pdf_bytes": 3258977, "paper_pdf_source": "openreview", "code_url": "https://github.com/ADelau/low_budget_sbi_with_bnn", "code_repository": "ADelau/low_budget_sbi_with_bnn", "code_commit": "413556a988453845bcc80c91ec959bf73dfc99db", "code_archive": "repos/cho9iE9POr.zip", "code_archive_sha256": "051aa5a0049272d3e830990fa13d000ae32f76a254df36bf836a982d42eaeff1", "code_archive_bytes": 87102, "code_file_count": 26, "code_extensions": {".py": 26}, "github_disk_usage_kb": 78, "github_languages": {"Python": 258561}, "github_archived": false, "github_pushed_at": "2024-09-05T11:27:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/low-budget-simulation-based-inference-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xuxYaBMd9F", "year": 2024, "status": "rejected", "title": "Efficient Long Sequence Modeling via State Space Augmented Transformer", "authors": ["Simiao Zuo", "Xiaodong Liu", "Jian Jiao", "Denis X Charles", "Eren Manavoglu", "Tuo Zhao", "Jianfeng Gao"], "authorids": ["~Simiao_Zuo1", "~Xiaodong_Liu1", "~Jian_Jiao2", "~Denis_X_Charles1", "~Eren_Manavoglu1", "~Tuo_Zhao1", "~Jianfeng_Gao1"], "authors_source": "OpenReview API", "abstract": "Transformer models have achieved superior performance in various natural language processing tasks. However, the quadratic computational cost of the attention mechanism limits its practicality for long sequences. There are existing attention variants that improve the computational efficiency, but they have limited ability to effectively compute global information. In parallel to Transformer models, state space models (SSMs) are tailored for long sequences, but they are not flexible enough to capture complicated local information. We propose SPADE, short for State Space Augmented Transformer. Specifically, we augment a SSM into the bottom layer of SPADE, and we employ efficient local attention methods for the other layers. The SSM augments global information, which complements the lack of long-range dependency issue in local attention methods. Experimental results on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method. To further demonstrate the scalability of SPADE, we pre-train large encoder-decoder models and present fine-tuning results on natural language understanding and natural language generation tasks. Our code and pre-trained model checkpoints will be publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "sUeg9bWvuU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4247/Reviewer_QKEi"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes the SPADE (State sPace AugmenteD TransformEr) model, which augments a State Space Model (SSM) to a transformer model to effectively capture global information from long sequences. It also leverages local attention modules to capture local information. Both SSN and local attention can be computed efficiently compared with full attention mechanisms. The paper presents extensive experimental results and conducts ablation studies to show the effectiveness of the proposed approach.", "review_text": "The paper proposes the SPADE (State sPace AugmenteD TransformEr) model, which augments a State Space Model (SSM) to a transformer model to effectively capture global information from long sequences. It also leverages local attention modules to capture local information. Both SSN and local attention can be computed efficiently compared with full attention mechanisms. The paper presents extensive experimental results and conducts ablation studies to show the effectiveness of the proposed approach.", "strengths": "- The paper is well written and easy to understand.\n- The proposed architecture strikes a balance between simplicity and complexity, demonstrating strong performance on sequences of varying lengths while incurring lower computational expenses compared to full attention.\n- Extensive experiment results across diverse datasets and tasks, along with ablation studies are provided.", "weaknesses": "- The paper primarily consolidates existing concepts, such as SSM and local attention, and offers limited novelty in terms of new methodologies.\n- Aside from the experimental results, it falls short in providing a comprehensive understanding of why this architecture is effective, and more crucially, in identifying scenarios where this approach may not be as effective.", "questions": "- In Figure 3, you simply concatenate SSM and local attention output, and apply a weight $\\bf W$, did you try any other method to fuse them?\n- It would be beneficial to include a relatively rigorous complexity analysis of SPADE comparing with various other methods, potentially in the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes the SPADE (State sPace AugmenteD TransformEr) model, which augments a State Space Model (SSM) to a transformer model to effectively capture global information from long sequences. It also leverages local attention modules to capture local information. Both SSN and local attention can be computed efficiently compared with full attention mechanisms. The paper presents extensive experimental results and conducts ablation studies to show the effectiveness of the proposed approach.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The paper is well written and easy to understand.\n- The proposed architecture strikes a balance between simplicity and complexity, demonstrating strong performance on sequences of varying lengths while incurring lower computational expenses compared to full attention.\n- Extensive experiment results across diverse datasets and tasks, along with ablation studies are provided.", "weaknesses": "- The paper primarily consolidates existing concepts, such as SSM and local attention, and offers limited novelty in terms of new methodologies.\n- Aside from the experimental results, it falls short in providing a comprehensive understanding of why this architecture is effective, and more crucially, in identifying scenarios where this approach may not be as effective.", "questions": "- In Figure 3, you simply concatenate SSM and local attention output, and apply a weight $\\bf W$, did you try any other method to fuse them?\n- It would be beneficial to include a relatively rigorous complexity analysis of SPADE comparing with various other methods, potentially in the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699034292763}, {"id": "fn1LipzVnP", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4247/Reviewer_6SGN"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents a novel model architecture designed to handle the complexities of both long-sequence processing. This is achieved by integrating S4 global attention with local attention mechanisms to form a hierarchical structure. The S4 layer is utilized at the bottom to capture long dependencies, while the local attention layers above aim to simplify attention complexity and expedite computation. The model outperforms the S4 alone and traditional Transformers in specific tasks.", "review_text": "This paper presents a novel model architecture designed to handle the complexities of both long-sequence processing. This is achieved by integrating S4 global attention with local attention mechanisms to form a hierarchical structure. The S4 layer is utilized at the bottom to capture long dependencies, while the local attention layers above aim to simplify attention complexity and expedite computation. The model outperforms the S4 alone and traditional Transformers in specific tasks.", "strengths": "Strengths:\n1. This paper describes a novel model that integrates S4 global attention and local (window-based or chunk-based) attention to address both long-range dependencies in language modeling. The hierarchical structure, with S4 at the bottom and local attention on top, aims to balance complexity and computation speed. \n2. The model has shown improvements over S4 and traditional Transformers in long-range and text generation tasks.", "weaknesses": "Weaknesses:\n1. S4 and Transformer are widely-used models, combining them together brings somehow incremental novelty contributions. \n2. The comparisons with alternative methods for capturing global context, such as RNN-like mechanisms or other efficient attention mechanisms, are incomplete and lack essential details. Only S4 is compared in table 1.  Some recent long-sequence modeling studies:\n  a. LongNet: Scaling Transformers to 1,000,000,000 Tokens\n  b. Long Range Language Modeling via Gated State Spaces", "questions": "Questions:\n1. T5-base is not implemented with the same setting with the proposed model. It would be more convincing to report apple-to-apple comparisons with the same setting, like the same pre-trained datasets.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel model architecture designed to handle the complexities of both long-sequence processing. This is achieved by integrating S4 global attention with local attention mechanisms to form a hierarchical structure. The S4 layer is utilized at the bottom to capture long dependencies, while the local attention layers above aim to simplify attention complexity and expedite computation. The model outperforms the S4 alone and traditional Transformers in specific tasks.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "Strengths:\n1. This paper describes a novel model that integrates S4 global attention and local (window-based or chunk-based) attention to address both long-range dependencies in language modeling. The hierarchical structure, with S4 at the bottom and local attention on top, aims to balance complexity and computation speed. \n2. The model has shown improvements over S4 and traditional Transformers in long-range and text generation tasks.", "weaknesses": "Weaknesses:\n1. S4 and Transformer are widely-used models, combining them together brings somehow incremental novelty contributions. \n2. The comparisons with alternative methods for capturing global context, such as RNN-like mechanisms or other efficient attention mechanisms, are incomplete and lack essential details. Only S4 is compared in table 1.  Some recent long-sequence modeling studies:\n  a. LongNet: Scaling Transformers to 1,000,000,000 Tokens\n  b. Long Range Language Modeling via Gated State Spaces", "questions": "Questions:\n1. T5-base is not implemented with the same setting with the proposed model. It would be more convincing to report apple-to-apple comparisons with the same setting, like the same pre-trained datasets.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698941440862}, {"id": "as3Lq5302B", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4247/Reviewer_o9d9"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces State sPace AugmenteD TransformEr (SPADE) which combines S4 with local attention to achieve and transformer model that avoids the quadratic sequence length scaling.\n\nThe authors note that local attention usually hurts the model's ability to attend to long-range dependencies and state-space models (SSM) usually do poorly in tasks where local information is important like language modeling. By having the bottom layer be a SSM and the rest of the layers perform local attention, they get the best of both worlds.\n\nThis is evidenced by performance in Long Range Arena (LRA), language modeling, and GLUE.", "review_text": "The paper introduces State sPace AugmenteD TransformEr (SPADE) which combines S4 with local attention to achieve and transformer model that avoids the quadratic sequence length scaling.\n\nThe authors note that local attention usually hurts the model's ability to attend to long-range dependencies and state-space models (SSM) usually do poorly in tasks where local information is important like language modeling. By having the bottom layer be a SSM and the rest of the layers perform local attention, they get the best of both worlds.\n\nThis is evidenced by performance in Long Range Arena (LRA), language modeling, and GLUE.", "strengths": "The empirical results are both thorough and impressive. The authors did many ablations and on many different tasks and the model performs well on all of them.\n\nThe paper is clear, and the idea is simple and intuitive. The paper is timed well to capture interest in large language models and context length.\n\nSection 6.3 regrading the location and number of global layers anticipates many questions about the justification for the experiment setup.", "weaknesses": "Perhaps one ablation that wasn't done is length generalization. The authors claim that \"our pre-trained model can extrapolate to any sequence length\", and theoretical justification is sound, but it would be good to see empirical evidence.\n\nPerhaps more configurations could be tried like different attention mechanisms for different heads or alternating layers.", "questions": "Does the method scale further? Is it easy to parallelize on multiple devices or even longer sequence lengths?\n\nWere there any investigations into what type of global information is being propagated. Perhaps by looking at attention scores?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces State sPace AugmenteD TransformEr (SPADE) which combines S4 with local attention to achieve and transformer model that avoids the quadratic sequence length scaling.\n\nThe authors note that local attention usually hurts the model's ability to attend to long-range dependencies and state-space models (SSM) usually do poorly in tasks where local information is important like language modeling. By having the bottom layer be a SSM and the rest of the layers perform local attention, they get the best of both worlds.\n\nThis is evidenced by performance in Long Range Arena (LRA), language modeling, and GLUE.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "The empirical results are both thorough and impressive. The authors did many ablations and on many different tasks and the model performs well on all of them.\n\nThe paper is clear, and the idea is simple and intuitive. The paper is timed well to capture interest in large language models and context length.\n\nSection 6.3 regrading the location and number of global layers anticipates many questions about the justification for the experiment setup.", "weaknesses": "Perhaps one ablation that wasn't done is length generalization. The authors claim that \"our pre-trained model can extrapolate to any sequence length\", and theoretical justification is sound, but it would be good to see empirical evidence.\n\nPerhaps more configurations could be tried like different attention mechanisms for different heads or alternating layers.", "questions": "Does the method scale further? Is it easy to parallelize on multiple devices or even longer sequence lengths?\n\nWere there any investigations into what type of global information is being propagated. Perhaps by looking at attention scores?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698878881045}, {"id": "bgyOAN31X9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4247/Reviewer_iRvY"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Summary:\n\nThe paper proposes SPADE (State sPace AugmenteD TransformEr), a novel approach for efficient long sequence modeling. SPADE integrates a state space model (SSM) into the bottom layer of a Transformer architecture, which enhances global information processing. This is complemented by local attention methods in the upper layers to handle local dependencies. The proposed method addresses the limitations of existing attention variants that struggle with long-range dependencies and computational inefficiency. SPADE demonstrates improved performance on the Long Range Arena benchmark and various language modeling tasks. The architecture allows SPADE to scale efficiently and outperform baselines in both natural language understanding and generation tasks, with the additional benefit of being able to handle longer sequences than it was trained on due to the SSM's extrapolation capabilities.", "review_text": "Summary:\n\nThe paper proposes SPADE (State sPace AugmenteD TransformEr), a novel approach for efficient long sequence modeling. SPADE integrates a state space model (SSM) into the bottom layer of a Transformer architecture, which enhances global information processing. This is complemented by local attention methods in the upper layers to handle local dependencies. The proposed method addresses the limitations of existing attention variants that struggle with long-range dependencies and computational inefficiency. SPADE demonstrates improved performance on the Long Range Arena benchmark and various language modeling tasks. The architecture allows SPADE to scale efficiently and outperform baselines in both natural language understanding and generation tasks, with the additional benefit of being able to handle longer sequences than it was trained on due to the SSM's extrapolation capabilities.", "strengths": "Advantages:\n\n - Integration of State Space Models: SPADE incorporates a state space model into the bottom layer of the architecture, providing a strong structural bias for augmenting global information and addressing long-range dependency issues present in local attention methods​.\n\n - Performance on Benchmarks: SPADE outperforms existing approaches (arguably marginally) on the Long Range Arena benchmark, specifically designed to assess models' ability to handle long sequences​.\n\n - Efficiency and Speed: In autoregressive language modeling tasks, SPADE is significantly faster and more performant than the vanilla Transformer model​.", "weaknesses": "Disadvantages:\n \n - Incremental Performance Gain: In the experiments, the performance gain compared to previous (truncated) transformer approaches are quite limited. I view the proposed method as a novel position encoding mechanism, expecting to see the comparison of it against vanilla/truncated Transformers with more advanced position encodings, such as Rotary embeddings, ALiBi and/or Transformer-XL.", "questions": "I feel more confused than amazed about the fact that while SSM alone cannot build a successful language model, using it (as a replacement of the position encoding) along with truncated (windowed/chunked) Transformer will simply result in an efficient long-term dependency capturing mechanism. I would appreciate it if the authors can conduct some ablation study to show that the inferior versions of SSM are, while still computationally efficient, not capable enough to support the dependency-capturing capabilities in language modeling, compared to the proposed SPADE.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Summary:\n\nThe paper proposes SPADE (State sPace AugmenteD TransformEr), a novel approach for efficient long sequence modeling. SPADE integrates a state space model (SSM) into the bottom layer of a Transformer architecture, which enhances global information processing. This is complemented by local attention methods in the upper layers to handle local dependencies. The proposed method addresses the limitations of existing attention variants that struggle with long-range dependencies and computational inefficiency. SPADE demonstrates improved performance on the Long Range Arena benchmark and various language modeling tasks. The architecture allows SPADE to scale efficiently and outperform baselines in both natural language understanding and generation tasks, with the additional benefit of being able to handle longer sequences than it was trained on due to the SSM's extrapolation capabilities.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Advantages:\n\n - Integration of State Space Models: SPADE incorporates a state space model into the bottom layer of the architecture, providing a strong structural bias for augmenting global information and addressing long-range dependency issues present in local attention methods​.\n\n - Performance on Benchmarks: SPADE outperforms existing approaches (arguably marginally) on the Long Range Arena benchmark, specifically designed to assess models' ability to handle long sequences​.\n\n - Efficiency and Speed: In autoregressive language modeling tasks, SPADE is significantly faster and more performant than the vanilla Transformer model​.", "weaknesses": "Disadvantages:\n \n - Incremental Performance Gain: In the experiments, the performance gain compared to previous (truncated) transformer approaches are quite limited. I view the proposed method as a novel position encoding mechanism, expecting to see the comparison of it against vanilla/truncated Transformers with more advanced position encodings, such as Rotary embeddings, ALiBi and/or Transformer-XL.", "questions": "I feel more confused than amazed about the fact that while SSM alone cannot build a successful language model, using it (as a replacement of the position encoding) along with truncated (windowed/chunked) Transformer will simply result in an efficient long-term dependency capturing mechanism. I would appreciate it if the authors can conduct some ablation study to show that the inferior versions of SSM are, while still computationally efficient, not capable enough to support the dependency-capturing capabilities in language modeling, compared to the proposed SPADE.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698815553142}, {"id": "2o1oPQEJaG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4247/Reviewer_w8HW"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Authors propose to augment the Transformer with SSM models to achieve better performance from computational and memory consumption perspectives. \n\nExperiments show that such architecture outperforms selected baselines on language modeling tasks and on Long Range Arena datasets. Furthermore, the authors showed that such a model could be successfully utilized for pre-training and fine-tuning, as shown in Section 5.", "review_text": "Authors propose to augment the Transformer with SSM models to achieve better performance from computational and memory consumption perspectives. \n\nExperiments show that such architecture outperforms selected baselines on language modeling tasks and on Long Range Arena datasets. Furthermore, the authors showed that such a model could be successfully utilized for pre-training and fine-tuning, as shown in Section 5.", "strengths": "- The paper is well-motivated and solved task is important for the field.\n\n- The paper is mostly well-written except for some flaws described in the weaknesses Section.", "weaknesses": "- My main concern for this paper is the lack of baselines. Including a comparison with other recent SSMs, such as S5 [1] or Hyena [2], would be highly beneficial.\n- The abovementioned models could be incorporated with SPADE by replacing the SSM module with any other model. However, it would be beneficial to understand whether the performance gap between SPADE and S4 is caused by adding Transformer blocks on top of the SSM module. Training SPADE with Hyena may not improve performance over Hyena while being better than SPADE with S4. In this case, the motivation for the paper will disappear.\n- I struggled to understand which SSM was used in SPADE (did I miss it?). It should be S4 based on the text. However, it would be helpful to name it in Section 3.2 explicitly.\n- The paper lacks reproducibility despite an extensive description of training within the text since no supplementary material with source code was released.\n\n[1] https://arxiv.org/pdf/2208.04933.pdf\n\n[2] https://arxiv.org/pdf/2302.10866.pdf\n\nWhile at this point, I voted for rejecting this paper, I believe that flaws with baselines could be fixed. Then, I would happily increase my score.", "questions": "Please refer to the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors propose to augment the Transformer with SSM models to achieve better performance from computational and memory consumption perspectives. \n\nExperiments show that such architecture outperforms selected baselines on language modeling tasks and on Long Range Arena datasets. Furthermore, the authors showed that such a model could be successfully utilized for pre-training and fine-tuning, as shown in Section 5.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is well-motivated and solved task is important for the field.\n\n- The paper is mostly well-written except for some flaws described in the weaknesses Section.", "weaknesses": "- My main concern for this paper is the lack of baselines. Including a comparison with other recent SSMs, such as S5 [1] or Hyena [2], would be highly beneficial.\n- The abovementioned models could be incorporated with SPADE by replacing the SSM module with any other model. However, it would be beneficial to understand whether the performance gap between SPADE and S4 is caused by adding Transformer blocks on top of the SSM module. Training SPADE with Hyena may not improve performance over Hyena while being better than SPADE with S4. In this case, the motivation for the paper will disappear.\n- I struggled to understand which SSM was used in SPADE (did I miss it?). It should be S4 based on the text. However, it would be helpful to name it in Section 3.2 explicitly.\n- The paper lacks reproducibility despite an extensive description of training within the text since no supplementary material with source code was released.\n\n[1] https://arxiv.org/pdf/2208.04933.pdf\n\n[2] https://arxiv.org/pdf/2302.10866.pdf\n\nWhile at this point, I voted for rejecting this paper, I believe that flaws with baselines could be fixed. Then, I would happily increase my score.", "questions": "Please refer to the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697281873241}], "openreview_url": "https://openreview.net/forum?id=xuxYaBMd9F", "arxiv_id": "2212.08136", "paper_pdf": "papers/xuxYaBMd9F.pdf", "paper_pdf_sha256": "e79869a8ec6bb64a4c2e49734bd2e3e5e4789efc9a95d424320c3c498c71acb9", "paper_pdf_bytes": 1419772, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/EfficientLongSequenceModeling", "code_repository": "microsoft/EfficientLongSequenceModeling", "code_commit": "0819294c9fbfb335eef08d54849aa8cb201de130", "code_archive": "repos/xuxYaBMd9F.zip", "code_archive_sha256": "ddbc311fe9b5ebc2417cd9cd7b0f77f04705721774599c7c4303cc55c5258aa5", "code_archive_bytes": 55936, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 60, "github_languages": {"Python": 179500, "Shell": 1079}, "github_archived": true, "github_pushed_at": "2023-01-19T02:15:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-long-sequence-modeling-via-state"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UERcQuXlwy", "year": 2023, "status": "rejected", "title": "Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding", "authors": ["Kenton Lee", "Mandar Joshi", "Iulia Raluca Turc", "Hexiang Hu", "Fangyu Liu", "Julian Martin Eisenschlos", "Urvashi Khandelwal", "Peter Shaw", "Ming-Wei Chang", "Kristina Toutanova"], "authorids": ["~Kenton_Lee1", "~Mandar_Joshi1", "~Iulia_Raluca_Turc1", "~Hexiang_Hu1", "~Fangyu_Liu1", "~Julian_Martin_Eisenschlos1", "~Urvashi_Khandelwal1", "~Peter_Shaw1", "~Ming-Wei_Chang3", "~Kristina_Toutanova1"], "authors_source": "OpenReview API", "abstract": "Visually-situated language is ubiquitous---sources range from textbooks with diagrams to web pages with images and tables, to mobile apps with buttons and forms. Perhaps due to this diversity, previous work has typically relied on domain-specific recipes with limited sharing of the underlying data, model architectures, and objectives. We present Pix2Struct, a pretrained image-to-text model for purely visual language understanding, which can be finetuned on tasks containing visually-situated language. Pix2Struct is pretrained by learning to parse masked screenshots of web pages into simplified HTML. The web, with its richness of visual elements cleanly reflected in the HTML structure, provides a large source of pretraining data well suited to the diversity of downstream tasks. Intuitively, this objective subsumes common pretraining signals such as OCR, language modeling, image captioning. In addition to the novel pretraining strategy, we introduce a variable-resolution input representation and a more flexible integration of language and vision inputs, where language prompts such as questions are rendered directly on top of the input image. For the first time, we show that a single pretrained model can achieve state-of-the-art results in six out of nine tasks across four domains: documents, illustrations, user interfaces, and natural images. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "nTX6MEbqgNW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5452/Reviewer_Apqp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes screenshot parsing as a simple yet effective pre-training task for visually-situated language modeling (e.g., document QA). The model gets rid of the OCR and directly accepts an image (with rich layouts and rendered text) as input. The empirical results show that the pre-training outperforms prior pixel-only methods by a large margin.", "review_text": "Overall, the paper shows that screen parsing is an effective pre-training task for visually-situated language modeling. I am on the fence about the design choice of using a pure pixel-only transformer but that does not undermine the core argument for screen parsing. ", "strengths": "The paper makes a welcomed and impressive empirical contribution, with several nice discussions on certain design choices. I will note 3 concerning points:\n\n1. OCR or no OCR?\n\nThe paper is in line with the trend of replacing domain-specific components with a general pre-trained neural model, which is generally welcomed. However, in domains where OCR is cheap and adequate, I am not sure if it is worthwhile to spend so much compute on replacing the OCR with a pixel-only LM. \n\nAn ideal solution would be an adaptive model which is pre-trained on such general screenshot parsing task but when adapted to a downstream task, could be fine-tuned to accept domain-specific inputs when there are cheap and easy-to-use solutions. \n\n2. Scaling up the model parameters seems to only bring marginal performance improvement.\n\nIs there any observation about this? Is it because the pre-training task is too simple? It would also be good if the authors could provide evaluations on the pre-training task itself, using some simple metrics such as blue or exact match. \n\n3. More details on how the pre-training data are extracted and cleaned should be provided. \n\nThe authors mentioned that the data are from C4. Does C4 already provide the cleaned data in HTML? My impression of C4 was that it was primarily a web-text corpus.\n\nStatistics on the pre-training data should also be provided. E.g., what’s the lengths distribution of the output strings?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes screenshot parsing as a simple yet effective pre-training task for visually-situated language modeling (e.g., document QA). The model gets rid of the OCR and directly accepts an image (with rich layouts and rendered text) as input. The empirical results show that the pre-training outperforms prior pixel-only methods by a large margin.", "strength_and_weaknesses": "The paper makes a welcomed and impressive empirical contribution, with several nice discussions on certain design choices. I will note 3 concerning points:\n\n1. OCR or no OCR?\n\nThe paper is in line with the trend of replacing domain-specific components with a general pre-trained neural model, which is generally welcomed. However, in domains where OCR is cheap and adequate, I am not sure if it is worthwhile to spend so much compute on replacing the OCR with a pixel-only LM. \n\nAn ideal solution would be an adaptive model which is pre-trained on such general screenshot parsing task but when adapted to a downstream task, could be fine-tuned to accept domain-specific inputs when there are cheap and easy-to-use solutions. \n\n2. Scaling up the model parameters seems to only bring marginal performance improvement.\n\nIs there any observation about this? Is it because the pre-training task is too simple? It would also be good if the authors could provide evaluations on the pre-training task itself, using some simple metrics such as blue or exact match. \n\n3. More details on how the pre-training data are extracted and cleaned should be provided. \n\nThe authors mentioned that the data are from C4. Does C4 already provide the cleaned data in HTML? My impression of C4 was that it was primarily a web-text corpus.\n\nStatistics on the pre-training data should also be provided. E.g., what’s the lengths distribution of the output strings?\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and makes a novel empirical contribution.", "summary_of_the_review": "Overall, the paper shows that screen parsing is an effective pre-training task for visually-situated language modeling. I am on the fence about the design choice of using a pure pixel-only transformer but that does not undermine the core argument for screen parsing. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666857396515}, {"id": "x9b9XESt2G", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5452/Reviewer_GpRK"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a pretrained image-to-text model for visual language understanding, which has a wide range of application in sources such as textbooks with diagrams, webpages with images and tables, and mobile apps with buttons and forms. The model has a simple and general architecture which only takes an image as the input. It is pretrained by parsing masked screenshots of webpages into simplified HTML, which subsumes common pretraining signals. This paper also proposes an integration of language and vision inputs for fine-tuning the model. The model achieves SOTA results in six out of nine tasks across four domains.\n", "review_text": "This paper proposes a simple pretraining framework for visual language understanding. The simplicity of the model, taking only image as the input, enables it to apply to diverse tasks and achieve supreme results. It is an exciting innovation in this area, and I believe it would greatly benefit the community if the data/model can be made public.\n", "strengths": "Strength:\n\n1. Pix2Struct uses a general-purpose pixel-to-text design, which simplifies the model architecture and can be easily applied to multiple domains.\n2. The masked screenshot parsing objective is intuitive and effective. The pretraining data is easy to obtain from the web. The warmup stage with an image-to-text curriculum further improves the pretrained model.\n3. The proposed fine-tuning strategy seamlessly integrate language and vision inputs by rendering language prompts on the image, without changing the architecture.\n4. A single pretrained model achieves strong performance on multiple visual language understanding tasks from diverse domains.\n\nWeakness:\n\n1. The presentation of this paper is easy to follow, but not clear enough. Many important techniques are described at a high level, for example \n\n    (1) How the proposed variable-resolution input representation is implemented, how to determine the number of patches for width and height, do small and large inputs have the same number of patches?\n\n    (2) How the pretraining data is collected and processed, especially for the warmup stage. \n\n    (3) How the masked parts are selected?\n\n2. The authors do not mention if the data/code/model will be publicly available. It would be difficult for the community to reproduce or follow if they cannot be released.\n\n3. Some of the designs lack empirical justification. For example, are there alternative ways to fine-tune the model, such as concatenating image and text? Does the location/size/font of where the prompt is rendered affect the performance? \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes a pretrained image-to-text model for visual language understanding, which has a wide range of application in sources such as textbooks with diagrams, webpages with images and tables, and mobile apps with buttons and forms. The model has a simple and general architecture which only takes an image as the input. It is pretrained by parsing masked screenshots of webpages into simplified HTML, which subsumes common pretraining signals. This paper also proposes an integration of language and vision inputs for fine-tuning the model. The model achieves SOTA results in six out of nine tasks across four domains.\n", "strength_and_weaknesses": "Strength:\n\n1. Pix2Struct uses a general-purpose pixel-to-text design, which simplifies the model architecture and can be easily applied to multiple domains.\n2. The masked screenshot parsing objective is intuitive and effective. The pretraining data is easy to obtain from the web. The warmup stage with an image-to-text curriculum further improves the pretrained model.\n3. The proposed fine-tuning strategy seamlessly integrate language and vision inputs by rendering language prompts on the image, without changing the architecture.\n4. A single pretrained model achieves strong performance on multiple visual language understanding tasks from diverse domains.\n\nWeakness:\n\n1. The presentation of this paper is easy to follow, but not clear enough. Many important techniques are described at a high level, for example \n\n    (1) How the proposed variable-resolution input representation is implemented, how to determine the number of patches for width and height, do small and large inputs have the same number of patches?\n\n    (2) How the pretraining data is collected and processed, especially for the warmup stage. \n\n    (3) How the masked parts are selected?\n\n2. The authors do not mention if the data/code/model will be publicly available. It would be difficult for the community to reproduce or follow if they cannot be released.\n\n3. Some of the designs lack empirical justification. For example, are there alternative ways to fine-tune the model, such as concatenating image and text? Does the location/size/font of where the prompt is rendered affect the performance? \n", "clarity,_quality,_novelty_and_reproducibility": "The idea presented in this paper is novel and the quality of the proposed method is good. \n\nI have concerns about this paper’s presentation and reproducibility. As mentioned before, many technical details are missing. The paper uses a lot of space to explain the downstream datasets and previous models for them, which seems to me unnecessary. I would suggest to use a figure for each dataset to explain the task and how Pix2Struct handles it, move the details to appendix, and use more space to describe the data processing or modeling of Pix2Struct in more detail. \n\nThis paper can be easily understood, but the model cannot be easily reproduced. I appreciate the authors’ effort for processing the web data and designing the model, which requires a lot of engineering. However, it is hard to follow this research if they cannot be made public.", "summary_of_the_review": "This paper proposes a simple pretraining framework for visual language understanding. The simplicity of the model, taking only image as the input, enables it to apply to diverse tasks and achieve supreme results. It is an exciting innovation in this area, and I believe it would greatly benefit the community if the data/model can be made public.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666805727210}, {"id": "FEImxIqens", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5452/Reviewer_dECs"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents Pix2Struct, a pretrained image-to-text model. The model is pretrained to parse masked screenshots of web pages and the authors show that it can be fine-tuned on multiple tasks containing visually-situated language (VQA, Image captioning, Infographic VQA). The fine-tuned models achieve state-of-the-art results in six tasks, while using only pixel-level inputs. The authors also introduce a variable-resolution input representation to prevent the distortion of input images.\n", "review_text": "Although the proposed method for pre-training a model from web page screenshots is interesting, the conditions of the experiment, with very large amounts of data and mobilizing unreasonable computational resources, do not allow to highlight its interest, its applicability and the reproducibility of the results. ", "strengths": "Although interesting and very well written, this paper does not make major contributions to the field. \n\nThe first contribution concerning the pre-training strategy is quite interesting. The authors propose a strategy based on web pages screenshots that allows to collect quickly and easily a very large amount of data. Moreover, these data can be quite varied and contain a large variety of elements, which allows to train a model that can work on many tasks. \n\nThe second contribution consists in adding modifications to the transformer inputs to handle variable aspect ratios and resolutions. This strategy, although interesting, brings little gain to the results. Moreover, it does not constitute a major contribution to the paper.\n\nFor the experiments, the following comments should be addressed.\n\n1. The idea of simplifying the multi-modalities problem is interesting. However, it could have been nice to compare the method to a standard approach with a modality combination.\n2. The input masking is done using crossed-out opaque bounding boxes. There is no justification about this choice. I wonder if this choice has an impact on the results compared to other approaches like removing the boxes or using white bounding boxes.\n3. The authors claim that the changes they added to the standard ViT inputs provide major advantages in terms of robustness to extreme aspect ratios and on-the-fly changes to the sequence length. However, except for the results presented on Figure 4, there is no experiments showing these advantages.\n4. It is not clear whether the modifications applied to the screenshots and HTML only consist in masking some texts or if other modifications are applied.\n5. I would have appreciated to have the inference times of the base and large models.\n6. The method of Wang et al (2021a) for the Screen2Words task could be more detailed.\n7. It is not said on which task and dataset the ablation study on the inputs resolution is carried on.\n8. 13 articles are cited as ArXiv preprints, please update the citations for articles that have been published.\ninor comments:\n\n1. In Table 2, it would make the table clearer to add a line with the metrics.\n2. In Table 2, the value 160.4 should be highlighted instead of 145.0.\n3. The values should be given with the same number of significative numbers:\n    - 11.27 in AI2D\n    - 145 in TextCaps\n    - 40 and 81 in InfographiVQA\n4. Figure 3 is a table, it should be renamed Table 3.\n5. Typo:\n    - Figure 4 caption: \"Our variable-resolution inputs prevent**s**...\"\n6. In the appendix\n    - In the fine-tuning section, it is said \"Tables 4 and 4\", it should be corrected to \"Tables 4 and 5\". By the way, tables 4 and 5 are quite similar, they could be merged into a single table.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents Pix2Struct, a pretrained image-to-text model. The model is pretrained to parse masked screenshots of web pages and the authors show that it can be fine-tuned on multiple tasks containing visually-situated language (VQA, Image captioning, Infographic VQA). The fine-tuned models achieve state-of-the-art results in six tasks, while using only pixel-level inputs. The authors also introduce a variable-resolution input representation to prevent the distortion of input images.\n", "strength_and_weaknesses": "Although interesting and very well written, this paper does not make major contributions to the field. \n\nThe first contribution concerning the pre-training strategy is quite interesting. The authors propose a strategy based on web pages screenshots that allows to collect quickly and easily a very large amount of data. Moreover, these data can be quite varied and contain a large variety of elements, which allows to train a model that can work on many tasks. \n\nThe second contribution consists in adding modifications to the transformer inputs to handle variable aspect ratios and resolutions. This strategy, although interesting, brings little gain to the results. Moreover, it does not constitute a major contribution to the paper.\n\nFor the experiments, the following comments should be addressed.\n\n1. The idea of simplifying the multi-modalities problem is interesting. However, it could have been nice to compare the method to a standard approach with a modality combination.\n2. The input masking is done using crossed-out opaque bounding boxes. There is no justification about this choice. I wonder if this choice has an impact on the results compared to other approaches like removing the boxes or using white bounding boxes.\n3. The authors claim that the changes they added to the standard ViT inputs provide major advantages in terms of robustness to extreme aspect ratios and on-the-fly changes to the sequence length. However, except for the results presented on Figure 4, there is no experiments showing these advantages.\n4. It is not clear whether the modifications applied to the screenshots and HTML only consist in masking some texts or if other modifications are applied.\n5. I would have appreciated to have the inference times of the base and large models.\n6. The method of Wang et al (2021a) for the Screen2Words task could be more detailed.\n7. It is not said on which task and dataset the ablation study on the inputs resolution is carried on.\n8. 13 articles are cited as ArXiv preprints, please update the citations for articles that have been published.\ninor comments:\n\n1. In Table 2, it would make the table clearer to add a line with the metrics.\n2. In Table 2, the value 160.4 should be highlighted instead of 145.0.\n3. The values should be given with the same number of significative numbers:\n    - 11.27 in AI2D\n    - 145 in TextCaps\n    - 40 and 81 in InfographiVQA\n4. Figure 3 is a table, it should be renamed Table 3.\n5. Typo:\n    - Figure 4 caption: \"Our variable-resolution inputs prevent**s**...\"\n6. In the appendix\n    - In the fine-tuning section, it is said \"Tables 4 and 4\", it should be corrected to \"Tables 4 and 5\". By the way, tables 4 and 5 are quite similar, they could be merged into a single table.", "clarity,_quality,_novelty_and_reproducibility": " I have a major concern regarding the applicability and the reproducibility of the results. The base and large models contain many parameters and the pre-training requires very important resources which represent a huge financial and ecological cost. This could be an argument for not using this pre-training method in other applications. Moreover, few organisations or companies have the capacity or even the desire to devote so much money and resources to such pre-training, which makes the results non-reproducible.", "summary_of_the_review": "Although the proposed method for pre-training a model from web page screenshots is interesting, the conditions of the experiment, with very large amounts of data and mobilizing unreasonable computational resources, do not allow to highlight its interest, its applicability and the reproducibility of the results. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["Yes, Other reasons (please specify below)"], "details_of_ethics_concerns": "Considering the ecological crisis, is it ethical to train a 1.3B parameters model on a 80M examples dataset during 170K steps on 128 Google Cloud TPUs ?", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666722446682}], "openreview_url": "https://openreview.net/forum?id=UERcQuXlwy", "arxiv_id": "2210.03347", "paper_pdf": "papers/UERcQuXlwy.pdf", "paper_pdf_sha256": "bf3b428b75eb08cb36871c0e22e98c459858f89954d2186480352db57b104642", "paper_pdf_bytes": 5865162, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/pix2struct", "code_repository": "google-research/pix2struct", "code_commit": "6fe25c1dc8151823ee3b479519d8d5948812fee4", "code_archive": "repos/UERcQuXlwy.zip", "code_archive_sha256": "a7c676f1991d4f20a1117451df7359ab684cb2e71ab8f5fbf1d605a3fb02a7a7", "code_archive_bytes": 69627, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 118, "github_languages": {"Python": 97547, "HTML": 2552, "CSS": 253}, "github_archived": false, "github_pushed_at": "2026-09-02T05:50:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pix2struct-screenshot-parsing-as-pretraining"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "s5lIqsrOu3Z", "year": 2022, "status": "rejected", "title": "Closed-Loop Data Transcription to an LDR via Minimaxing Rate Reduction", "authors": ["Xili Dai", "Shengbang Tong", "Mingyang Li", "Ziyang Wu", "Kwan Ho Ryan Chan", "Pengyuan Zhai", "Yaodong Yu", "Michael Psenka", "Xiaojun Yuan", "Heung-Yeung Shum", "Yi Ma"], "authorids": ["~Xili_Dai2", "~Shengbang_Tong1", "~Mingyang_Li3", "~Ziyang_Wu1", "~Kwan_Ho_Ryan_Chan1", "~Pengyuan_Zhai1", "~Yaodong_Yu4", "psenka@berkeley.edu", "xjyuan@uestc.edu.cn", "msraharry@hotmail.com", "~Yi_Ma4"], "authors_source": "OpenReview API", "abstract": "This work proposes a new computational framework for automatically learning a closed-loop transcription between multi-class multi-dimensional data and a linear discriminative representation (LDR) that consists of multiple multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as the equilibrium point of a two-player minimax game between the encoder and decoder. A natural  utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation avoids expensive evaluating and minimizing approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, conceptually and computationally this new formulation unifies the benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for complex multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark datasets demonstrate tremendous potential of this framework: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and often better than existing methods based on GAN, VAE or a combination of both. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "9b2BOqH2UbU", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper788/Reviewer_beUC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The manuscript considers learning a closed-loop auto-encoder between multi-class multi-dimension data and LDR. It also provides various experiments to verify the proposed method. ", "review_text": "The idea of viewing the encoder and the decoder in a close-loop auto-eoncoder as generator and discriminator is quite interesting. \n\nOne of the critical parts of the proposed method is viewing the encoder as a discriminator and measuring the distance between X and \\hat{X} as the maximization of the rate reduction.  The authors explain that the maximization can avoid that g\\circ f is not an auto-encoding map. However, it is still not super clear that why such maximization is reasonable to make g(f(x)) = x. In other words, why not use eqn (7) together with conventional reconstruction loss. This seems like also provide desired maps between multi-class data and LDR. It is a bit hard for me to see the advantages of using maximization there. \n\nSome minor issues:\nIn eqns (2), Z maps to \\hat{X}, it is a bit strange to write g maps Z to (\\hat(X),X). Similar issue for eqn (5). \n\nIn Section 1.2, X has been amused for different sets. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The manuscript considers learning a closed-loop auto-encoder between multi-class multi-dimension data and LDR. It also provides various experiments to verify the proposed method. ", "main_review": "The idea of viewing the encoder and the decoder in a close-loop auto-eoncoder as generator and discriminator is quite interesting. \n\nOne of the critical parts of the proposed method is viewing the encoder as a discriminator and measuring the distance between X and \\hat{X} as the maximization of the rate reduction.  The authors explain that the maximization can avoid that g\\circ f is not an auto-encoding map. However, it is still not super clear that why such maximization is reasonable to make g(f(x)) = x. In other words, why not use eqn (7) together with conventional reconstruction loss. This seems like also provide desired maps between multi-class data and LDR. It is a bit hard for me to see the advantages of using maximization there. \n\nSome minor issues:\nIn eqns (2), Z maps to \\hat{X}, it is a bit strange to write g maps Z to (\\hat(X),X). Similar issue for eqn (5). \n\nIn Section 1.2, X has been amused for different sets. \n\n", "summary_of_the_review": "The explanation of the maximization part or two-player game part is not super clear to me. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635997242940}, {"id": "wSxRc7geXGt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper788/Reviewer_ykKP"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes to learn auto-encoder using an adversarial objective based on rate reduction. The experiments show that the proposed method achieves competitive performance on reconstruction, generation, and discrimination. ", "review_text": "Strengths: \n\n(1) The proposed adversarial objective function is novel and interesting. \n\n(2) The experimental results are solid. \n\nWeakness: \n\n(1) No background is given on rate reduction and MCR2. As a result, the paper is not self-contained, and is not easy to understand. \n\n(2) The objective function is less simple and clean than VAE and GAN loss functions, which are closely related to log-likelihood of a generative model or a discriminative model. The objective function in this paper is quite complex. \n\n(3) No theoretical analysis of the proposed method is given. \n\n(4) When generating a new x, is z sampled from N(0,I)? The probabilisitic generative model is not very explicit in this paper. A probabilistic generative/decoder model is a natural representation of the observed data, with encoder being interpreted as approximated posterior inference. The auto-encoding point of view appears rather limited in comparison. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to learn auto-encoder using an adversarial objective based on rate reduction. The experiments show that the proposed method achieves competitive performance on reconstruction, generation, and discrimination. ", "main_review": "Strengths: \n\n(1) The proposed adversarial objective function is novel and interesting. \n\n(2) The experimental results are solid. \n\nWeakness: \n\n(1) No background is given on rate reduction and MCR2. As a result, the paper is not self-contained, and is not easy to understand. \n\n(2) The objective function is less simple and clean than VAE and GAN loss functions, which are closely related to log-likelihood of a generative model or a discriminative model. The objective function in this paper is quite complex. \n\n(3) No theoretical analysis of the proposed method is given. \n\n(4) When generating a new x, is z sampled from N(0,I)? The probabilisitic generative model is not very explicit in this paper. A probabilistic generative/decoder model is a natural representation of the observed data, with encoder being interpreted as approximated posterior inference. The auto-encoding point of view appears rather limited in comparison. \n", "summary_of_the_review": "The paper is an extension of MCR. It appears a bit complex. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635919419415}, {"id": "yGtE5V9mJuk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper788/Reviewer_hk2J"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper \"Closed-Loop Data Transcription To An LDR via Minimaxing Rate Reduction\" introduces a two-player minimax game between an encoder and a decoder to yield a linear discriminative representation (LDR). It achieves this by building on top of the recently proposed MCR2 rate reduction principle, and then motivates a contractive and contrastive measure to yield a minimax game. The paper shows generative empirical results on MNIST, CIFAR-10 and ImageNet, as well as discriminative classification accuracies on MNIST.", "review_text": "Strengths.\n- the core idea of the paper is novel as far as I can tell. I'd argue it has some resemblance to CycleGan (Zhu et al. 2017), but the setting on top of rate reduction and motivation is significantly different.\n- the experiments demonstrate that the method can work well as a generative method, even though it formulates its losses only in latent space.\n\nWeaknesses. \n- The paper is not as self-contained as it could be, i.e. it is near impossible to understand without reading the core referred literature first. It would help readability a lot if core intuitions and concepts would be briefly discussed when introduced. E.g. when introducing MCR2 in equation 4, it would help the reader a lot what the target of each of the two terms is. Another example, after equation 6 it is stated that this measures the volume - why? (if this is not easy to explain, then there should be a reference to where this is explained). The math also needs more clarification, for example what is the union in equation 6 (Z and \\hat Z are both \\in Rˆ{dxn}, or)? Is the \\Delta R in equation 4 and equation 6 really the same (or is it semantically overloaded)?  writing (and logic of the writing) needs to be worked on.\n- Experiments. The comparisons on generative models in Table 1 are done with rather old baselines (all being at least 4 years old). E.g. DCGAN (Radford 2015) could be replaced by StyleGAN2/3 etc (similar for the VAE methods). It should also be reported what the training time for the method on the datasets is, and how it scales with dataset size and resolution. Also, it should be discussed that the model only roughly encodes the semantics, and sometimes disregards even color (e.g. as seen in Figure 5 where the red car turns yellow). In Figure 14 in the abstract it can be seen that the the model seems to sometimes collapse inputs into the same output? If so this should be discussed.\n- Already in the Abstract it is claimed that the model learns a discriminative representation, and it is shown in Table 3 to perform in the ballpark of VAE methods on MNIST data. As the model is also trained on ImageNet (see Fig5), why is no discrimnative comparison performed on ImageNet? It'd be interesting to see if the model can scale up to more classes and more complex datasets (what is the scaling behaviour theoretically of the model in the latent space? Does the dimensionality need to be proportional to the number of classes to ensure the possibility of orthogonality?) . \n\nMinor Issues.\n- The Introduction should start with a section contextualizing the work, i.e. review of context and background, what has been done by others broadly in the field that led to the current work. \n- Equation 2,3 etc. - the result of function g is not a tuple (X, \\hat X). The figure should be changed.\n- After eq. 3 it is written \"later studies [...] surrogate to earth mover's distance\". I think this is wrong. The original GAN paper showed that GAN's minimize Jensen-Shannon distance, and Arjovsky et al showed how to build the WGAN that operates on earth mover's distance.\n- Also on pg 2. - Combination of AE and GAN: \".. started with somewhat different motivation, they have evolved...\". While true that both methods are quite different, as both are generative models, it was this motivation (generative modeling) that they become successful in modeling of real-world data. So, the logic behind this sentence is confusing. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper \"Closed-Loop Data Transcription To An LDR via Minimaxing Rate Reduction\" introduces a two-player minimax game between an encoder and a decoder to yield a linear discriminative representation (LDR). It achieves this by building on top of the recently proposed MCR2 rate reduction principle, and then motivates a contractive and contrastive measure to yield a minimax game. The paper shows generative empirical results on MNIST, CIFAR-10 and ImageNet, as well as discriminative classification accuracies on MNIST.", "main_review": "Strengths.\n- the core idea of the paper is novel as far as I can tell. I'd argue it has some resemblance to CycleGan (Zhu et al. 2017), but the setting on top of rate reduction and motivation is significantly different.\n- the experiments demonstrate that the method can work well as a generative method, even though it formulates its losses only in latent space.\n\nWeaknesses. \n- The paper is not as self-contained as it could be, i.e. it is near impossible to understand without reading the core referred literature first. It would help readability a lot if core intuitions and concepts would be briefly discussed when introduced. E.g. when introducing MCR2 in equation 4, it would help the reader a lot what the target of each of the two terms is. Another example, after equation 6 it is stated that this measures the volume - why? (if this is not easy to explain, then there should be a reference to where this is explained). The math also needs more clarification, for example what is the union in equation 6 (Z and \\hat Z are both \\in Rˆ{dxn}, or)? Is the \\Delta R in equation 4 and equation 6 really the same (or is it semantically overloaded)?  writing (and logic of the writing) needs to be worked on.\n- Experiments. The comparisons on generative models in Table 1 are done with rather old baselines (all being at least 4 years old). E.g. DCGAN (Radford 2015) could be replaced by StyleGAN2/3 etc (similar for the VAE methods). It should also be reported what the training time for the method on the datasets is, and how it scales with dataset size and resolution. Also, it should be discussed that the model only roughly encodes the semantics, and sometimes disregards even color (e.g. as seen in Figure 5 where the red car turns yellow). In Figure 14 in the abstract it can be seen that the the model seems to sometimes collapse inputs into the same output? If so this should be discussed.\n- Already in the Abstract it is claimed that the model learns a discriminative representation, and it is shown in Table 3 to perform in the ballpark of VAE methods on MNIST data. As the model is also trained on ImageNet (see Fig5), why is no discrimnative comparison performed on ImageNet? It'd be interesting to see if the model can scale up to more classes and more complex datasets (what is the scaling behaviour theoretically of the model in the latent space? Does the dimensionality need to be proportional to the number of classes to ensure the possibility of orthogonality?) . \n\nMinor Issues.\n- The Introduction should start with a section contextualizing the work, i.e. review of context and background, what has been done by others broadly in the field that led to the current work. \n- Equation 2,3 etc. - the result of function g is not a tuple (X, \\hat X). The figure should be changed.\n- After eq. 3 it is written \"later studies [...] surrogate to earth mover's distance\". I think this is wrong. The original GAN paper showed that GAN's minimize Jensen-Shannon distance, and Arjovsky et al showed how to build the WGAN that operates on earth mover's distance.\n- Also on pg 2. - Combination of AE and GAN: \".. started with somewhat different motivation, they have evolved...\". While true that both methods are quite different, as both are generative models, it was this motivation (generative modeling) that they become successful in modeling of real-world data. So, the logic behind this sentence is confusing. \n", "summary_of_the_review": "While the paper proposes an interesting and promising method to learn an encoder-decoder with a novel two-player minimax game, the paper is not yet well enough written and lacks some experimental results that allow for comparison with state of the art methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635888164236}], "openreview_url": "https://openreview.net/forum?id=s5lIqsrOu3Z", "arxiv_id": "2111.06636", "paper_pdf": "papers/s5lIqsrOu3Z.pdf", "paper_pdf_sha256": "14357277906801d25c6b68c075eec61eee7f7b857fa4e4fc669c99f0faa5d5d0", "paper_pdf_bytes": 46561409, "paper_pdf_source": "openreview", "code_url": "https://github.com/Delay-Xili/LDR", "code_repository": "Delay-Xili/LDR", "code_commit": "4c6098a4ab03a67be108d4567709cd8ac013dfea", "code_archive": "repos/s5lIqsrOu3Z.zip", "code_archive_sha256": "eeb94b9216cb353c7c39a1471bc0b583d2e3cd2e7a317d551e30927cd8f04638", "code_archive_bytes": 101370, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 149, "github_languages": {"Python": 106107}, "github_archived": false, "github_pushed_at": "2022-11-03T08:35:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/closed-loop-data-transcription-to-an-ldr-via-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "D1E1h-K3jso", "year": 2021, "status": "rejected", "title": "Learning from Noisy Data with Robust Representation Learning", "authors": ["Junnan Li", "Caiming Xiong", "Steven Hoi"], "authorids": ["~Junnan_Li2", "~Caiming_Xiong1", "~Steven_Hoi2"], "authors_source": "OpenReview API", "abstract": "Learning from noisy data has attracted much attention, where most methods focus on label noise. In this work, we propose a new framework which simultaneously addresses three types of noise commonly seen in real-world data: label noise, out-of-distribution input, and input corruption. In contrast to most existing methods, we combat noise by learning robust representation. Specifically, we embed images into a low-dimensional subspace by training an autoencoder on the deep features. We regularize the geometric structure of the subspace with robust contrastive learning, which includes an unsupervised consistency loss and a supervised mixup prototypical loss. Furthermore, we leverage the structure of the learned subspace for noise cleaning, by aggregating information from neighboring samples. Experiments on multiple benchmarks demonstrate state-of-the-art performance of our method and robustness of the learned representation. Our code will be released.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "qm79EGHvQ2_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper321/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new methodology for noisy-robust learning by augmenting the simlr methodology with a Mixup-style augmentation and noise-cleaning.\n\nThe paper proposes a new methodology for training models in the presence of label noise. The method is a combination of standard SimCLR: using standard image augmentations and enforcing a contrastive consistence loss between the emeddings of the weakly augmented and strongly augmented versions of the same image. In addition, to the standard augmentation, the method also uses a Mixup style augmentation for the strongly augmented images: that is the projection of a randomly selected image is combined (using a convex combination) with the original image and then another contrastive loss is enforced for the class prototyle.\nThe high dimensional embedding of the weakly augmented image is trained with the (noisy) label using a cross-entropy loss.\nIn addition, the high dimensional embedding is also reconstructed from the low-dimensional projection.\n\nThe final evaluation also contains a noise-cleaning step by generating pseudo-labels from the smooth-neighborhoods using the above embeddings.\n\nThe quality of the writing is high and the paper presents a plausible combination of several strong methods. The paper is nicely written, well motivated and can be followed easily.\n\nThe paper also presents significant improvement of artificially noisy versions of CIFAR-10 and CIFAR-100. Also it is also measured on versions on which the images are noised by using extra datasets (SVHN) to create more interesting augmentations of the datasets.\n\nThe weakness of the paper is that it is only measured on cifar-10/100 and these datasest are often not very representative of real noisy datasets with label noise of complicated structure.\n\nAnother weakness is that this paper presents a relatively complex approach composed of 4 different methods, each of them are well studied and with very limited ablation analyses. It is plausible that the combined approach should do well, still it has limited novelty as it is the straightforward combination of four known approaches also the paper does not give a disciplined overview and study of the contributions of the different components.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting combination of Mixup, SimCLR,  denoising autoencoding and noise-cleaning.", "review": "This paper proposes a new methodology for noisy-robust learning by augmenting the simlr methodology with a Mixup-style augmentation and noise-cleaning.\n\nThe paper proposes a new methodology for training models in the presence of label noise. The method is a combination of standard SimCLR: using standard image augmentations and enforcing a contrastive consistence loss between the emeddings of the weakly augmented and strongly augmented versions of the same image. In addition, to the standard augmentation, the method also uses a Mixup style augmentation for the strongly augmented images: that is the projection of a randomly selected image is combined (using a convex combination) with the original image and then another contrastive loss is enforced for the class prototyle.\nThe high dimensional embedding of the weakly augmented image is trained with the (noisy) label using a cross-entropy loss.\nIn addition, the high dimensional embedding is also reconstructed from the low-dimensional projection.\n\nThe final evaluation also contains a noise-cleaning step by generating pseudo-labels from the smooth-neighborhoods using the above embeddings.\n\nThe quality of the writing is high and the paper presents a plausible combination of several strong methods. The paper is nicely written, well motivated and can be followed easily.\n\nThe paper also presents significant improvement of artificially noisy versions of CIFAR-10 and CIFAR-100. Also it is also measured on versions on which the images are noised by using extra datasets (SVHN) to create more interesting augmentations of the datasets.\n\nThe weakness of the paper is that it is only measured on cifar-10/100 and these datasest are often not very representative of real noisy datasets with label noise of complicated structure.\n\nAnother weakness is that this paper presents a relatively complex approach composed of 4 different methods, each of them are well studied and with very limited ablation analyses. It is plausible that the combined approach should do well, still it has limited novelty as it is the straightforward combination of four known approaches also the paper does not give a disciplined overview and study of the contributions of the different components.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603921243367}, {"id": "HwPVdjHOvMc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper321/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#######################################################################\n\nSummary:\n \nThe paper proposes noise-robust contrastive learning to combat label noise, out-of-distribution input and input corruption simultaneously. In particular, this paper embeds images into low-dimensional representations by training an autoencoder, and regularizes the geometric structure of the representations by contrastive learning. Furthermore, this paper introduces a new noise cleaning method based on the structure of the representations. Training samples with confident pseudo-labels are selected for supervised learning to clean both label noise and out-of-distribution noise. The effectiveness of the proposed method has been evaluated on multiple simulated and real-world noisy datasets. \n\n#######################################################################\n\nReasons for score: \n \nOverall, I vote for a weak acceptance. The proposed noise-robust contrastive learning introduces two contrastive losses: unsupervised consistency loss and supervised mixup prototypical loss. My major concern is about the clarity of the paper and some additional issues (see cons below). Hopefully the authors can address my concern in the rebuttal period. \n \n#######################################################################\n\nPros: \n \n1. The paper takes one import issue in deep learning: learning from noisy data.  \n \n2. For me, the proposed supervised mixup prototypical contrastive loss is novel for learning with noisy data. Specifically, it injects structure knowledge of classes into embedding space by combining the mixup technique and prototypical contrastive loss. The design is reasonable and interesting. \n \n3. This paper provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed framework. In particular, the proposed method outperforms several state-of-the-art robust learning methods in learning with label noise, out-of-distribution input and input corruption. \n\n#######################################################################\n\nCons: \n  \n1. For the motivation, it would be better to provide more details about it, which seems not very clear to me. Particularly, it is unclear why the contrastive loss is duly used in the paper. Will the functionality of the unsupervised contrastive loss be achieved in the supervised prototypical loss? Additionally, the prototypical contrastive loss in equation (1) is an InfoNCE with normalized mean embeddings as the prototypes, which seems different from the ProtoNCE in the original paper [1]. It is better to clarify the the differences of the formulation and training strategy, and the reason of design of supervised prototypical loss in this paper. \n\n[1] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C.H. Hoi. Prototypical Contrastive Learning of Unsupervised Representations, In ICLR, 2020. \n\n2. As the key contribution of this paper, mixup prototypical contrastive loss combines mixup technique and prototypical contrastive loss. In the appendix, the authors have provided ablation study to show the effect of proposed losses, and it shows that the mixup prototypical loss is most crucial to the model’s performance. Here, the authors utilizes mixup in two ways: first is to create virtual training samples, and the other is to define the mixup version of prototypical contrastive loss as an weighted combination of two prototypical contrastive loss with respect to true class and virtual class. However, the effect of mixup augmentation, prototypical contrastive loss and mixup prototypical contrastive loss are unclear.  Since mixup has been shown to be an effective method against label noise, it would be more convincing if the authors can study the individual effect of each components of the proposed loss in the rebuttal period. \n \n3. The proposed method uses many data augmentations, e.g., standard crop and horizontal flip as weak augmentation and AugMix as strong augmentation in the unsupervised consistency contrastive loss, and mixup technique in the supervised prototypical contrastive loss. I am concerning about the fairness in the experimental comparison. It is unclear to me if the authors have applied the same data augmentation to all the compared methods. \n \n#######################################################################\n\nQuestions during rebuttal period: \n \nPlease address and clarify the cons above. \n \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An extention of contrastive learning for training noise-robust deep networks. ", "review": "#######################################################################\n\nSummary:\n \nThe paper proposes noise-robust contrastive learning to combat label noise, out-of-distribution input and input corruption simultaneously. In particular, this paper embeds images into low-dimensional representations by training an autoencoder, and regularizes the geometric structure of the representations by contrastive learning. Furthermore, this paper introduces a new noise cleaning method based on the structure of the representations. Training samples with confident pseudo-labels are selected for supervised learning to clean both label noise and out-of-distribution noise. The effectiveness of the proposed method has been evaluated on multiple simulated and real-world noisy datasets. \n\n#######################################################################\n\nReasons for score: \n \nOverall, I vote for a weak acceptance. The proposed noise-robust contrastive learning introduces two contrastive losses: unsupervised consistency loss and supervised mixup prototypical loss. My major concern is about the clarity of the paper and some additional issues (see cons below). Hopefully the authors can address my concern in the rebuttal period. \n \n#######################################################################\n\nPros: \n \n1. The paper takes one import issue in deep learning: learning from noisy data.  \n \n2. For me, the proposed supervised mixup prototypical contrastive loss is novel for learning with noisy data. Specifically, it injects structure knowledge of classes into embedding space by combining the mixup technique and prototypical contrastive loss. The design is reasonable and interesting. \n \n3. This paper provides comprehensive experiments, including both qualitative analysis and quantitative results, to show the effectiveness of the proposed framework. In particular, the proposed method outperforms several state-of-the-art robust learning methods in learning with label noise, out-of-distribution input and input corruption. \n\n#######################################################################\n\nCons: \n  \n1. For the motivation, it would be better to provide more details about it, which seems not very clear to me. Particularly, it is unclear why the contrastive loss is duly used in the paper. Will the functionality of the unsupervised contrastive loss be achieved in the supervised prototypical loss? Additionally, the prototypical contrastive loss in equation (1) is an InfoNCE with normalized mean embeddings as the prototypes, which seems different from the ProtoNCE in the original paper [1]. It is better to clarify the the differences of the formulation and training strategy, and the reason of design of supervised prototypical loss in this paper. \n\n[1] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C.H. Hoi. Prototypical Contrastive Learning of Unsupervised Representations, In ICLR, 2020. \n\n2. As the key contribution of this paper, mixup prototypical contrastive loss combines mixup technique and prototypical contrastive loss. In the appendix, the authors have provided ablation study to show the effect of proposed losses, and it shows that the mixup prototypical loss is most crucial to the model’s performance. Here, the authors utilizes mixup in two ways: first is to create virtual training samples, and the other is to define the mixup version of prototypical contrastive loss as an weighted combination of two prototypical contrastive loss with respect to true class and virtual class. However, the effect of mixup augmentation, prototypical contrastive loss and mixup prototypical contrastive loss are unclear.  Since mixup has been shown to be an effective method against label noise, it would be more convincing if the authors can study the individual effect of each components of the proposed loss in the rebuttal period. \n \n3. The proposed method uses many data augmentations, e.g., standard crop and horizontal flip as weak augmentation and AugMix as strong augmentation in the unsupervised consistency contrastive loss, and mixup technique in the supervised prototypical contrastive loss. I am concerning about the fairness in the experimental comparison. It is unclear to me if the authors have applied the same data augmentation to all the compared methods. \n \n#######################################################################\n\nQuestions during rebuttal period: \n \nPlease address and clarify the cons above. \n \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603872284189}, {"id": "cnSpEnOcavL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper321/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors of the paper propose to use the contrastive loss, the mixup prototypical loss, and a reconstruction loss to regularize the learned representation in order to achieve robustness under various kinds of noise like label noise, out-of-distribution input, and input corruption. A noise-cleaning process based on the learned representation is also introduced to further enhance the results. Extensive experiments were conducted to demonstrate the effectiveness of the method. \n\nAll in all, the paper is well written and easy to understand. The proposed method is well justified and seems to be technically sound. I also find the idea of using neighboring samples to perform noise-cleaning intuitive and interesting. A major complaint I have regarding this submission is the lack of novelty. While the empirical results are encouraging, the proposed method is merely a combination of various previously proposed methods. Moreover, the final objective used (Equation 8) also involves 3 hyper-parameters that can be hard to tune. How sensitive are the results when different weights are used? Why is the $w_{pc}$ changed across different datasets? Due to the limitation addressed above, I think the paper a very borderline submission.\n\nQuestions: \n1. It was demonstrated in the appendix that the prototypical loss is the most critical of all the losses used in this paper. However, no ablation study was done with respect to the effect of the mixup on the prototypical loss. How would regular prototypical loss perform without mixup in the context of these experiments?\n2. It was stated that the labels for the pseudo-labeled samples were converted into hard labels. Does this lead to better performance than if soft labels are used? \n----------------------------------------------------------------------\nAfter Rebuttal: \n\nI would like to thank the authors for answering my questions and addressing my concerns on hyperparameters. I also appreciate the authors' efforts in the additional ablation study conducted. While I still do think that the novelty of the paper is a little bit lacking, I think the experiments are carefully conducted and the empirical results seem to be encouraging. As such, I have raised my score to 6.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for \"Noise-Robust Contrastive Learning\"", "review": "The authors of the paper propose to use the contrastive loss, the mixup prototypical loss, and a reconstruction loss to regularize the learned representation in order to achieve robustness under various kinds of noise like label noise, out-of-distribution input, and input corruption. A noise-cleaning process based on the learned representation is also introduced to further enhance the results. Extensive experiments were conducted to demonstrate the effectiveness of the method. \n\nAll in all, the paper is well written and easy to understand. The proposed method is well justified and seems to be technically sound. I also find the idea of using neighboring samples to perform noise-cleaning intuitive and interesting. A major complaint I have regarding this submission is the lack of novelty. While the empirical results are encouraging, the proposed method is merely a combination of various previously proposed methods. Moreover, the final objective used (Equation 8) also involves 3 hyper-parameters that can be hard to tune. How sensitive are the results when different weights are used? Why is the $w_{pc}$ changed across different datasets? Due to the limitation addressed above, I think the paper a very borderline submission.\n\nQuestions: \n1. It was demonstrated in the appendix that the prototypical loss is the most critical of all the losses used in this paper. However, no ablation study was done with respect to the effect of the mixup on the prototypical loss. How would regular prototypical loss perform without mixup in the context of these experiments?\n2. It was stated that the labels for the pseudo-labeled samples were converted into hard labels. Does this lead to better performance than if soft labels are used? \n----------------------------------------------------------------------\nAfter Rebuttal: \n\nI would like to thank the authors for answering my questions and addressing my concerns on hyperparameters. I also appreciate the authors' efforts in the additional ablation study conducted. While I still do think that the novelty of the paper is a little bit lacking, I think the experiments are carefully conducted and the empirical results seem to be encouraging. As such, I have raised my score to 6.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603855728411}, {"id": "SxFvTUl4uRx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper321/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors have conducted a range of experiments to validate the performance of the proposed method. And according to the results, it is appealing that the proposed method achieves relative improvement compared to current SOTAs . However, there are some concerns as follows.\n\n1) It is quite ad-hoc about the proposed method, which consists of multiple losses and two phases in the training progresses. The potential hyperparameter space is relative black-box to understand where is the bottleneck of learning with noisy labels in this method. It also requires careful hyperparameter-setting as shown in the implementation details.\n\n2) Some contributions are over-claimed since many components have also used in previous works for learning with noisy labels, e.g., the idea of consistency loss and autoencoder in Bootstrapping [1].  Besides, the title of the paper is also misleading since it is about learning with noisy labels instead of contrastive learning.\n\n\n[1] S. Reed et al. Training deep neural networks on noisy labels with bootstrapping. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper introduces a noise-robust contrastive learning loss. The authors first learn robust representation to regularize the learning procedure with label noise. Then, a noise cleaning process is implemented to select some confident samples for partially finetuning. ", "review": "The authors have conducted a range of experiments to validate the performance of the proposed method. And according to the results, it is appealing that the proposed method achieves relative improvement compared to current SOTAs . However, there are some concerns as follows.\n\n1) It is quite ad-hoc about the proposed method, which consists of multiple losses and two phases in the training progresses. The potential hyperparameter space is relative black-box to understand where is the bottleneck of learning with noisy labels in this method. It also requires careful hyperparameter-setting as shown in the implementation details.\n\n2) Some contributions are over-claimed since many components have also used in previous works for learning with noisy labels, e.g., the idea of consistency loss and autoencoder in Bootstrapping [1].  Besides, the title of the paper is also misleading since it is about learning with noisy labels instead of contrastive learning.\n\n\n[1] S. Reed et al. Training deep neural networks on noisy labels with bootstrapping. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603805612138}], "openreview_url": "https://openreview.net/forum?id=D1E1h-K3jso", "arxiv_id": null, "paper_pdf": "papers/D1E1h-K3jso.pdf", "paper_pdf_sha256": "98f2bfb372584218b2c0e92bde106169f820c26b5c296d6457148bd4515b29bc", "paper_pdf_bytes": 2071851, "paper_pdf_source": "openreview", "code_url": "https://github.com/salesforce/RRL", "code_repository": "salesforce/RRL", "code_commit": "a0d8a8fca05d040aef15b82f4870d33b65e59ce7", "code_archive": "repos/D1E1h-K3jso.zip", "code_archive_sha256": "4a9da6528cdfcf66ce4d0a678d027ff9473da7cad0c7d96068f36df3efad77fe", "code_archive_bytes": 178117, "code_file_count": 22, "code_extensions": {".py": 21, ".ipynb": 1}, "github_disk_usage_kb": 177, "github_languages": {"Python": 106457, "Jupyter Notebook": 31049}, "github_archived": true, "github_pushed_at": "2025-05-01T17:27:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-from-noisy-data-with-robust"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJlXgkHYvS", "year": 2020, "status": "rejected", "title": "Information-Theoretic Local Minima Characterization and Regularization", "authors": ["Zhiwei Jia", "Hao Su"], "authorids": ["zjia@ucsd.edu", "haosu@eng.ucsd.edu"], "authors_source": "OpenReview API", "abstract": "Recent advances in deep learning theory have evoked the study of generalizability across different local minima of deep neural networks (DNNs). While current work focused on either discovering properties of good local minima or developing regularization techniques to induce good local minima, no approach exists that can tackle both problems. We achieve these two goals successfully in a unified manner. Specifically, based on the Fisher information we propose a metric both strongly indicative of generalizability of local minima and effectively applied as a practical regularizer. We provide theoretical analysis including a generalization bound and empirically demonstrate the success of our approach in both capturing and improving the generalizability of DNNs. Experiments are performed on CIFAR-10 and CIFAR-100 for various network architectures.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJeaomh79S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1500/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper provides a metric to characterize local minima of deep network loss landscapes based on the Fisher information matrix of the model parameterized by the deep network. The authors connect the Fisher information to the curvature of the loss landscape (the loss considered is the negative loss likelihood) and obtain generalization bounds through PAC Bayes analysis. They further propose regularizing the training of deep networks using the local curvature of the loss as a regularizer. In the final experimental section of the paper, the relationship between the empirical measures and generalization is shown on a variety of networks.\n\nThis is an interesting paper, but I have a few concerns.\n\n1. The information-theoretic measure that is proposed is essentially the (log) determinant of the hessian of the loss function.  If there are degenerate eigendirections (zero eigenvalues) then the proposed measure would not be able to distinguish between minima with different numbers of degenerate directions / same number of degenerate directions but different spectral norms of the hessians. If the authors contention is that there will be no zero eigenvalues, that suggests that local minima of deep networks are all strict, isolated minima, contrary to recent work on connected solutions (See Draxler et. al. 2018, Essentially No Barriers in Neural Network Energy Landscapes, ICML 2018).\n\n2. I would like to see how the authors believe their measure deals with rescalings layer parameters in deep networks, ie the issue brought up by Dinh et. al. in \"Sharp Minima can Generalize for Deep Networks\" ICML 2017. While I can see that the log determinant is invariant, it is not clear that the proposed approximation will be invariant to rescaling of deep network layer parameters. If the parameters corresponding to the eigenvalues sampled in the approximation are rescaled, I believe the proposed measure will not be invariant.\n\n3. The experiments regarding the local minima characterization are well constructed, though some details are missing such as how the authors decided that training had converged to a local minimum. As far as regularization based on the local curvature is concerned, I would like to see some more experiments that compare the proposed technique to adagrad/adam and other techniques that purport to condition the gradient based on local curvature. It would also be interesting to see whether the regularization indeed converges to flatter minima characterized by the proposed flatness measure. Since the claim is that the regularizer gets you flatter solutions, that information is important to decide whether the proposed technique is performing as advertised.\n\nI am willing to update my score based on responses to these concerns.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper provides a metric to characterize local minima of deep network loss landscapes based on the Fisher information matrix of the model parameterized by the deep network. The authors connect the Fisher information to the curvature of the loss landscape (the loss considered is the negative loss likelihood) and obtain generalization bounds through PAC Bayes analysis. They further propose regularizing the training of deep networks using the local curvature of the loss as a regularizer. In the final experimental section of the paper, the relationship between the empirical measures and generalization is shown on a variety of networks.\n\nThis is an interesting paper, but I have a few concerns.\n\n1. The information-theoretic measure that is proposed is essentially the (log) determinant of the hessian of the loss function.  If there are degenerate eigendirections (zero eigenvalues) then the proposed measure would not be able to distinguish between minima with different numbers of degenerate directions / same number of degenerate directions but different spectral norms of the hessians. If the authors contention is that there will be no zero eigenvalues, that suggests that local minima of deep networks are all strict, isolated minima, contrary to recent work on connected solutions (See Draxler et. al. 2018, Essentially No Barriers in Neural Network Energy Landscapes, ICML 2018).\n\n2. I would like to see how the authors believe their measure deals with rescalings layer parameters in deep networks, ie the issue brought up by Dinh et. al. in \"Sharp Minima can Generalize for Deep Networks\" ICML 2017. While I can see that the log determinant is invariant, it is not clear that the proposed approximation will be invariant to rescaling of deep network layer parameters. If the parameters corresponding to the eigenvalues sampled in the approximation are rescaled, I believe the proposed measure will not be invariant.\n\n3. The experiments regarding the local minima characterization are well constructed, though some details are missing such as how the authors decided that training had converged to a local minimum. As far as regularization based on the local curvature is concerned, I would like to see some more experiments that compare the proposed technique to adagrad/adam and other techniques that purport to condition the gradient based on local curvature. It would also be interesting to see whether the regularization indeed converges to flatter minima characterized by the proposed flatness measure. Since the claim is that the regularizer gets you flatter solutions, that information is important to decide whether the proposed technique is performing as advertised.\n\nI am willing to update my score based on responses to these concerns."}, "tcdate": 1572221860932}, {"id": "H1xq1PY2tr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1500/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Post-rebuttal update: I have just noticed the authors modified their summary post below and claimed \"[my concerns] are all minor or resolved\". This is not true. Here is my summary of unresolved concerns written after the discussion period.\n\nThis work has been substantially improved during the rebuttal process, and some of my concerns are addressed. But there are still major issues, as raised in my [last comment]( https://openreview.net/forum?id=BJlXgkHYvS&noteId=r1xAnokijS ), that remains unanswered. Specifically,\n\n(A) the relation between this work and information theory\n\nIn the revision, the authors have make it very clear that the relation between FIA and their proposed regularized objective is very vague, relying on the crude approximation of expected Fisher information with observed Fisher information. Therefore the \"information-theoretic\" part in the title seems awkward and to some extent, misleading.\n\nAs Reviewer 1 has pointed out, it would have been better if the authors relate their theory and method to the observed FIM, instead of information theory, from the beginning. Since the observed FIM and the neural tangent kernel (NTK) share the same eigenspectrum, it would also be interesting to relate this work to the NTK.\n\n(B) the different behavior of the proposed regularization (log det(I)) and its bound that is actually implemented (log tr(I))\n\nThis is the more important issue. My concern is that the observed FIM (or the NTK) is known to have fast decaying spectrum; (Karakida et al) has shown empirically that the decay can be exponential. Thus log det(I) would be dominated by the long tail (since after taking logarithm it is the sum (or average) of an arithmetic sequence), while log tr(I) would be dominated by the first largest few values. \n\nThe authors claim that this is not an issue since they replaced the observed FIM with a subsampled, low-rank (<=10), version. It corresponds to consider a small submatrix of (the gram matrix of) the NTK. Denote this matrix as .\n(a) This does not help with the problem, since we now have no chance of recovering the smaller eigenvalues that would have dominated log(det(I)), and it is impossible that the proposed regularizer has a similar behavior to log(det(I)).\n(b) One could verify easily, using small feed-forward networks (or even simpler, computing a gram matrix using RBF kernels, since the FIM shares its eigenspectrum with NTK which is a p.d. kernel), that the new matrix  still has a fast-decaying eigenspectrum, so the behavior of  and \\tilde{I} are still significantly different, even though this cannot be established by concentration bounds as the authors argue. While FFN and modern deep architectures can have different behaviors, I believe the above evidence suggests that a numerical experiment comparing the behavior of the two bounds is a must.\n\nFollowing this argument we can see *another issue* of this work, namely the proposed generalization bound will be vacuous given the fast-decaying spectrum of the FIM, since it contains gamma=log(det(I)).\n\nReviewer 1 mentioned this work could enlighten future discussions on this subject. While I agree this paper presents interesting empirical observations (namely its final algorithm, which is vaguely connected to the proposed objective, leads to improved performance on CV tasks), I think this submission in its current form is a bit too misleading to serve this purpose well, and overall I believe it would be better to go through another round of revision.\n\n\nOriginal Review\n============================================\n\nThis paper presents a generalization bound based on Fisher information at a local optima, and proposes to optimize (an approximation to) it to get better generalization guarantees. There are issues in both parts, and I don't think it should be accepted. Specifically,\n\n1. The definition of Fisher information is incorrect (for almost every parameter). The expectation should be taken w.r.t the model distribution p(c_x|x;w), instead of the data distribution S. \n2. Assumption (1) (loss locally quadratic) is not reasonable for DNNs, since local optimas will not be unique in their neighborhoods. See e.g. Section 12.2.2, \"Information Geometry and Its Applications\".\n3. Regarding the approximation to the bound, approximating log det(I) with log trace(I) is not a good idea: adding a very small eigenvalue will lead to noticeable change in the former, but negligible change in the latter. This is particularly problematic for DNNs, since the spectrum of their Fisher information matrix varies in a wide range: see \"Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach\". \n\n(Edit 11.8:\n* regarding point (1), there is a quantity called observed Fisher information in e.g. Grunwald (2007) that coincide with Eq (1) in the paper, but it is a function of the dataset instead of the model parameter, and can only used to study model parameters near the *global optima* (as it is applied in Grunwald (2007)); it cannot help with choosing between different local optimas as this work claims. Additionally, the FIA criterion, which is used in this paper to devleop the generalization bound, is defined using the standard form of Fisher information (i.e. taking expectation w.r.t model distribution), see Rissanen (1996). These facts lead me to believe this is a confusion on the authors' part.\n* in point (2) I was referring to the authors' argument \" Since L(S,w) is analytic and w_0 is the *only local minimum* of L(S,w) in M(w_0)\", which is incorrect.)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "Post-rebuttal update: I have just noticed the authors modified their summary post below and claimed \"[my concerns] are all minor or resolved\". This is not true. Here is my summary of unresolved concerns written after the discussion period.\n\nThis work has been substantially improved during the rebuttal process, and some of my concerns are addressed. But there are still major issues, as raised in my [last comment]( https://openreview.net/forum?id=BJlXgkHYvS&noteId=r1xAnokijS ), that remains unanswered. Specifically,\n\n(A) the relation between this work and information theory\n\nIn the revision, the authors have make it very clear that the relation between FIA and their proposed regularized objective is very vague, relying on the crude approximation of expected Fisher information with observed Fisher information. Therefore the \"information-theoretic\" part in the title seems awkward and to some extent, misleading.\n\nAs Reviewer 1 has pointed out, it would have been better if the authors relate their theory and method to the observed FIM, instead of information theory, from the beginning. Since the observed FIM and the neural tangent kernel (NTK) share the same eigenspectrum, it would also be interesting to relate this work to the NTK.\n\n(B) the different behavior of the proposed regularization (log det(I)) and its bound that is actually implemented (log tr(I))\n\nThis is the more important issue. My concern is that the observed FIM (or the NTK) is known to have fast decaying spectrum; (Karakida et al) has shown empirically that the decay can be exponential. Thus log det(I) would be dominated by the long tail (since after taking logarithm it is the sum (or average) of an arithmetic sequence), while log tr(I) would be dominated by the first largest few values. \n\nThe authors claim that this is not an issue since they replaced the observed FIM with a subsampled, low-rank (<=10), version. It corresponds to consider a small submatrix of (the gram matrix of) the NTK. Denote this matrix as .\n(a) This does not help with the problem, since we now have no chance of recovering the smaller eigenvalues that would have dominated log(det(I)), and it is impossible that the proposed regularizer has a similar behavior to log(det(I)).\n(b) One could verify easily, using small feed-forward networks (or even simpler, computing a gram matrix using RBF kernels, since the FIM shares its eigenspectrum with NTK which is a p.d. kernel), that the new matrix  still has a fast-decaying eigenspectrum, so the behavior of  and \\tilde{I} are still significantly different, even though this cannot be established by concentration bounds as the authors argue. While FFN and modern deep architectures can have different behaviors, I believe the above evidence suggests that a numerical experiment comparing the behavior of the two bounds is a must.\n\nFollowing this argument we can see *another issue* of this work, namely the proposed generalization bound will be vacuous given the fast-decaying spectrum of the FIM, since it contains gamma=log(det(I)).\n\nReviewer 1 mentioned this work could enlighten future discussions on this subject. While I agree this paper presents interesting empirical observations (namely its final algorithm, which is vaguely connected to the proposed objective, leads to improved performance on CV tasks), I think this submission in its current form is a bit too misleading to serve this purpose well, and overall I believe it would be better to go through another round of revision.\n\n\nOriginal Review\n============================================\n\nThis paper presents a generalization bound based on Fisher information at a local optima, and proposes to optimize (an approximation to) it to get better generalization guarantees. There are issues in both parts, and I don't think it should be accepted. Specifically,\n\n1. The definition of Fisher information is incorrect (for almost every parameter). The expectation should be taken w.r.t the model distribution p(c_x|x;w), instead of the data distribution S. \n2. Assumption (1) (loss locally quadratic) is not reasonable for DNNs, since local optimas will not be unique in their neighborhoods. See e.g. Section 12.2.2, \"Information Geometry and Its Applications\".\n3. Regarding the approximation to the bound, approximating log det(I) with log trace(I) is not a good idea: adding a very small eigenvalue will lead to noticeable change in the former, but negligible change in the latter. This is particularly problematic for DNNs, since the spectrum of their Fisher information matrix varies in a wide range: see \"Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach\". \n\n(Edit 11.8:\n* regarding point (1), there is a quantity called observed Fisher information in e.g. Grunwald (2007) that coincide with Eq (1) in the paper, but it is a function of the dataset instead of the model parameter, and can only used to study model parameters near the *global optima* (as it is applied in Grunwald (2007)); it cannot help with choosing between different local optimas as this work claims. Additionally, the FIA criterion, which is used in this paper to devleop the generalization bound, is defined using the standard form of Fisher information (i.e. taking expectation w.r.t model distribution), see Rissanen (1996). These facts lead me to believe this is a confusion on the authors' part.\n* in point (2) I was referring to the authors' argument \" Since L(S,w) is analytic and w_0 is the *only local minimum* of L(S,w) in M(w_0)\", which is incorrect.)", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571751649854}, {"id": "rJgVQyi9FB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1500/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper contributes to the deep learning generalization theory, mainly from the theoretical perspective with experimental verifications. The key proposition is given by the unnumbered simple equation in the middle of page 4 (please number it), where \\mathcal{I} is the Fisher information matrix. According to the authors, this simple metric, which is the log-determinant of the Fisher information matrix, can characterize the generalization of a DNN.\n\nRemarkably, this piece of work is well written in terms of English and formulations, and complete, with a rigorous theoretical analysis (section 5.1, 5.2), practical approximations (section 5.3) and empirical verifications (section 6).\n\nOn the theoretical side, this work builds upon Rissanen's formulation of the MDL principle, which has two parts (describing data given the model as well as the model complexity). Under rough approximations, the complexity term becomes the log-determinant of the Fisher information matrix evaluated at the local (global) optimum. This simple approximation is further proved to upper-bounds the generalization error as stated in theorem 1.\n\nTo make the criterion to be practically useful, the author used the Jensen inequality so that the metric simply depends on the trace of the Fisher information matrix.\n\nThe empirical study showed the usefulness of the proposed metric which can well approximate the testing error and a regularization term (based on the trace of the Fisher information matrix) that can improve generalization on real DNN experiments.\n\nThe reviewer has the following minor comments to further improve this contribution:\n\nsection 5.1, explain the abbreviation FIA\n\nRegarding the choice of the neighborhood \\mathcal{M}(w_0), what is the reason to define the model (neighbourhood of w_0) based on the loss? Why not simply take a coordinate neighborhood?\n\nAccording to your metric, the smaller the scale of the Fisher information matrix, the better the generalization. In section 5.1, there has to be some remarks on the intuition and related works on the flatness of the local minimum that is related to generalization.\n\nAs this contribution is related to the spectral properties of the Fisher information matrix, the reviewer points the authors to \"Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach. Karakida et al. 2018.\" and \"Lightlike Neuromanifolds, Occam's Razor and Deep Learning. Sun and Nielsen. 2019\", which deals with asymptotic cases and have similar MDL formulations expressed in terms of the spectrum of the Fisher information matrix.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper contributes to the deep learning generalization theory, mainly from the theoretical perspective with experimental verifications. The key proposition is given by the unnumbered simple equation in the middle of page 4 (please number it), where \\mathcal{I} is the Fisher information matrix. According to the authors, this simple metric, which is the log-determinant of the Fisher information matrix, can characterize the generalization of a DNN.\n\nRemarkably, this piece of work is well written in terms of English and formulations, and complete, with a rigorous theoretical analysis (section 5.1, 5.2), practical approximations (section 5.3) and empirical verifications (section 6).\n\nOn the theoretical side, this work builds upon Rissanen's formulation of the MDL principle, which has two parts (describing data given the model as well as the model complexity). Under rough approximations, the complexity term becomes the log-determinant of the Fisher information matrix evaluated at the local (global) optimum. This simple approximation is further proved to upper-bounds the generalization error as stated in theorem 1.\n\nTo make the criterion to be practically useful, the author used the Jensen inequality so that the metric simply depends on the trace of the Fisher information matrix.\n\nThe empirical study showed the usefulness of the proposed metric which can well approximate the testing error and a regularization term (based on the trace of the Fisher information matrix) that can improve generalization on real DNN experiments.\n\nThe reviewer has the following minor comments to further improve this contribution:\n\nsection 5.1, explain the abbreviation FIA\n\nRegarding the choice of the neighborhood \\mathcal{M}(w_0), what is the reason to define the model (neighbourhood of w_0) based on the loss? Why not simply take a coordinate neighborhood?\n\nAccording to your metric, the smaller the scale of the Fisher information matrix, the better the generalization. In section 5.1, there has to be some remarks on the intuition and related works on the flatness of the local minimum that is related to generalization.\n\nAs this contribution is related to the spectral properties of the Fisher information matrix, the reviewer points the authors to \"Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach. Karakida et al. 2018.\" and \"Lightlike Neuromanifolds, Occam's Razor and Deep Learning. Sun and Nielsen. 2019\", which deals with asymptotic cases and have similar MDL formulations expressed in terms of the spectrum of the Fisher information matrix.\n"}, "tcdate": 1571626780489}], "openreview_url": "https://openreview.net/forum?id=BJlXgkHYvS", "arxiv_id": "1911.08192", "paper_pdf": "papers/BJlXgkHYvS.pdf", "paper_pdf_sha256": "334434fbf38a1c4fcb92a8a2c4b98e3ed73b6f5362c93cd1543e7fe26efaf219", "paper_pdf_bytes": 949579, "paper_pdf_source": "openreview", "code_url": "https://github.com/SeanJia/InfoMCR", "code_repository": "SeanJia/InfoMCR", "code_commit": "2b4760ad6ffdd98859fea1967eb1b8aa7e51be52", "code_archive": "repos/BJlXgkHYvS.zip", "code_archive_sha256": "dbc42beb7a7019f2a888cee260a8069ed4a9a34eef9a6eab33bbcf00465d4a33", "code_archive_bytes": 152435, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 193, "github_languages": {"Python": 17713}, "github_archived": false, "github_pushed_at": "2020-07-16T06:28:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/information-theoretic-local-minima-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bke96sC5tm", "year": 2019, "status": "rejected", "title": "SOLAR: Deep Structured Representations for Model-Based Reinforcement Learning", "authors": ["Marvin Zhang*", "Sharad Vikram*", "Laura Smith", "Pieter Abbeel", "Matthew Johnson", "Sergey Levine"], "authorids": ["marvin@cs.berkeley.edu", "svikram@cs.ucsd.edu", "smithlaura@berkeley.edu", "pabbeel@cs.berkeley.edu", "mattjj@google.com", "svlevine@cs.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Model-based reinforcement learning (RL) methods can be broadly categorized as global model methods, which depend on learning models that provide sensible predictions in a wide range of states, or local model methods, which iteratively refit simple models that are used for policy improvement. While predicting future states that will result from the current actions is difficult, local model methods only attempt to understand system dynamics in the neighborhood of the current policy, making it possible to produce local improvements without ever learning to predict accurately far into the future. The main idea in this paper is that we can learn representations that make it easy to retrospectively infer simple dynamics given the data from the current policy, thus enabling local models to be used for policy learning in complex systems. We evaluate our approach against other model-based and model-free RL methods on a suite of robotics tasks, including manipulation tasks on a real Sawyer robotic arm directly from camera images.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "B1g3stit6Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper831/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \n\nThe paper proposes SOLAR a model based RL algorithm that learns a low dimensional embedding such that the dynamics within the latent space are linear. Within this latent space the linear dynamics  are learned using a Bayesian regression. In addition, a quadratic cost function is approximated. The learned dynamics and the cost function are used to update the policy, while simultaneously bounding the change in policy by a KL-bound.  In contrast to other model-based RL algorithms the learned dynamics are not used for planning or imaginary roll-outs and are only used to improve the policy.\n\nReview:\nThe introduction and experiment section is clearly written but the algorithm description lacks clarity and details, which hinder the understanding of the complete algorithm. One understands the motivation and the main approach but lacks a detailed understanding. For my personal taste the detailed description of learning the embedding is missing. I personally would prefer the statement of the cost-functions and the optimisation problem within the paper and not the appendix. The same holds true for the policy improvement. Therefore, I do not fully understand the approach without extensively studying the appendix or the references. Especially the contribution remains unclear. I am not aware how much the previous work had to be extended.\n\nThe experimental evaluation focuses on learning control signal to achieve certain trajectories, where the observations are high-dimensional images rather than low-dimensional representations. I personally think that these tasks are unnecessarily made more complex to incorporate high-dimensional images. Especially, the Sawyer experiment throws away all joint information even though the reward function is solely defined in joint/end-effector position. However, I am aware that this is general practice in the RL community. From the learning curves it seems that the approach is working and achieving good sample complexity compared to model free approaches. However, the improvement over the naive VAE approach remains unclear. I would like to see more comparisons to other model-based approaches. In addition, I am missing qualitative comparisons as the learning curves can be misleading. Especially, the videos on the homepage are really short and do not provide a good overview about the actual performance. Furthermore, you are not providing videos for all models in comparison. The 1s video of a single episode on the reacher task make me wonder what happens in the other episodes. Could you please add longer videos for all comparisons. Furthermore, it would be interesting how the trajectories evolve over time. Could you plot these trajectories? \n\nFurthermore, it would be really interesting to try your approach on breakout. And test if your approach is learning the actual game dynamics and does not overfit to the block configuration. \n\nFurther minor comments:\n- \"This shifts our problem setting to that of a partially observed MDP, as we do not observe the latent state\"\nYou are mentioning that you are solving a POMDP. Could you elaborate how you exploit the POMDP formulation and relate your work to POMDP algorithms. In addition, how do you define partial observability? \n\n- You claim \"our method is also successful at handling the complex, contact-rich dynamics of block stacking, which poses a significant challenge compared to the other contactfree tasks.\" I am quite doubt-full about the claim. Is the dynamics model really modelling contacts and is your policy really reacting to these contacts? Or is your policy just tying to follow a trajectory? From your current evaluation and the videos, I personally wouldn't conclude this. Could you elaborate how you come to this conclusion and provide additional evaluations to solidify your argument? \n\n- You are not describing the action space for the Sawyer experiment. Are you using torques, velocities or positions? Can you guarantee that the control sequence is smooth? If not how do you ensure that the policy does not harm the robot? \n \n- Could you please incorporate the exact  reward functions for each experiment within the appendix.\n\n- Figure 4. Thanks a lot for including the additional model free baselines and adding all learning curves. However, the learning curves raise multiple questions:\n\n(1) The Global Model Ablation, i.e. the MPC in latent space, works well in the the navigation and car experiment however fails \nto achieve a meaning-full policy within the reacher task. Even though the initial performance is significantly better than \nthe other policies. Do you have an explanation for this failure?\n\n(2) The LDS SVAE and VAE Solar version on the reacher task experiences jumps in performance even though the change between policies is bounded by a KL-Bound and the cost function is smooth. How do you explain these jumps? Furthermore, why are these jumps only occurring within the reacher tasks and not the other experiments. \n\n(3) You are still missing the PPO baselines for the reacher and car experiment. Could you further explain the qualitative difference between the model-free and model-based policies. The difference in learning curves can be misleading. \n\n(4) What is the unit of \"Average Distance to Final Goal\"? Is this measured in pixel or a different unit? \n\n- Figure 5: You are plotting the distance to the goal as performance measure for the Sawyer experiment. The final policy has an approximate error of 2.5 cm. From just the learning curve I cannot conclude that the robot actually learns the task successfully. Is the block really stacked or can it also be wedged? Could you please provide image overlays of the last 10 episodes such that one can evaluate the qualitative performance? \n\n \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting model based RL approach that needs additional evaluation and clearer algorithm description", "review": "Summary: \n\nThe paper proposes SOLAR a model based RL algorithm that learns a low dimensional embedding such that the dynamics within the latent space are linear. Within this latent space the linear dynamics  are learned using a Bayesian regression. In addition, a quadratic cost function is approximated. The learned dynamics and the cost function are used to update the policy, while simultaneously bounding the change in policy by a KL-bound.  In contrast to other model-based RL algorithms the learned dynamics are not used for planning or imaginary roll-outs and are only used to improve the policy.\n\nReview:\nThe introduction and experiment section is clearly written but the algorithm description lacks clarity and details, which hinder the understanding of the complete algorithm. One understands the motivation and the main approach but lacks a detailed understanding. For my personal taste the detailed description of learning the embedding is missing. I personally would prefer the statement of the cost-functions and the optimisation problem within the paper and not the appendix. The same holds true for the policy improvement. Therefore, I do not fully understand the approach without extensively studying the appendix or the references. Especially the contribution remains unclear. I am not aware how much the previous work had to be extended.\n\nThe experimental evaluation focuses on learning control signal to achieve certain trajectories, where the observations are high-dimensional images rather than low-dimensional representations. I personally think that these tasks are unnecessarily made more complex to incorporate high-dimensional images. Especially, the Sawyer experiment throws away all joint information even though the reward function is solely defined in joint/end-effector position. However, I am aware that this is general practice in the RL community. From the learning curves it seems that the approach is working and achieving good sample complexity compared to model free approaches. However, the improvement over the naive VAE approach remains unclear. I would like to see more comparisons to other model-based approaches. In addition, I am missing qualitative comparisons as the learning curves can be misleading. Especially, the videos on the homepage are really short and do not provide a good overview about the actual performance. Furthermore, you are not providing videos for all models in comparison. The 1s video of a single episode on the reacher task make me wonder what happens in the other episodes. Could you please add longer videos for all comparisons. Furthermore, it would be interesting how the trajectories evolve over time. Could you plot these trajectories? \n\nFurthermore, it would be really interesting to try your approach on breakout. And test if your approach is learning the actual game dynamics and does not overfit to the block configuration. \n\nFurther minor comments:\n- \"This shifts our problem setting to that of a partially observed MDP, as we do not observe the latent state\"\nYou are mentioning that you are solving a POMDP. Could you elaborate how you exploit the POMDP formulation and relate your work to POMDP algorithms. In addition, how do you define partial observability? \n\n- You claim \"our method is also successful at handling the complex, contact-rich dynamics of block stacking, which poses a significant challenge compared to the other contactfree tasks.\" I am quite doubt-full about the claim. Is the dynamics model really modelling contacts and is your policy really reacting to these contacts? Or is your policy just tying to follow a trajectory? From your current evaluation and the videos, I personally wouldn't conclude this. Could you elaborate how you come to this conclusion and provide additional evaluations to solidify your argument? \n\n- You are not describing the action space for the Sawyer experiment. Are you using torques, velocities or positions? Can you guarantee that the control sequence is smooth? If not how do you ensure that the policy does not harm the robot? \n \n- Could you please incorporate the exact  reward functions for each experiment within the appendix.\n\n- Figure 4. Thanks a lot for including the additional model free baselines and adding all learning curves. However, the learning curves raise multiple questions:\n\n(1) The Global Model Ablation, i.e. the MPC in latent space, works well in the the navigation and car experiment however fails \nto achieve a meaning-full policy within the reacher task. Even though the initial performance is significantly better than \nthe other policies. Do you have an explanation for this failure?\n\n(2) The LDS SVAE and VAE Solar version on the reacher task experiences jumps in performance even though the change between policies is bounded by a KL-Bound and the cost function is smooth. How do you explain these jumps? Furthermore, why are these jumps only occurring within the reacher tasks and not the other experiments. \n\n(3) You are still missing the PPO baselines for the reacher and car experiment. Could you further explain the qualitative difference between the model-free and model-based policies. The difference in learning curves can be misleading. \n\n(4) What is the unit of \"Average Distance to Final Goal\"? Is this measured in pixel or a different unit? \n\n- Figure 5: You are plotting the distance to the goal as performance measure for the Sawyer experiment. The final policy has an approximate error of 2.5 cm. From just the learning curve I cannot conclude that the robot actually learns the task successfully. Is the block really stacked or can it also be wedged? Could you please provide image overlays of the last 10 episodes such that one can evaluate the qualitative performance? \n\n \n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1542203812425}, {"id": "SklP9FFR3Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper831/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a model-based reinforcement learning approach, called SOLAR, \nwhich consists of mapping complex, high-dimensional observations to low-dimensional \nrepresentations where transition dynamics between consecutive states are approximately linear. \nIn this low-dimensional space, local models can easily be fit in closed form and then used to optimize a policy, using a similar method to Guided Policy Search (GPS). The method is evaluated in 4 different settings (3 simulated, 1 on a real robot). \n\n*Quality: the method seems to work well in the experiments. However, there are issues with the experimental evaluation (detailed below) which make it unclear whether the method is better than standard baselines.\n\n*Clarity: the paper is well-written and clear overall.  \n\n*Originality: the paper proposes an extension of GPS, which to my knowledge is novel. \n\n*Significance: the idea of learning representations where transitions are linear seems well-founded and potentially useful. However the merits of this method are not yet clear from the experiments. \n\n\nSpecific Comments:\n\n- Please include an illustration of the 2D navigation task in Figure 3a\n- I'm confused by the poor performance of E2C in the 2D navigation task. \nThe previous works of [Watter et. al, 2015] and [Banijalami et. al, 2017] report close to 100% accuracy using similar methods. Is the task formulated differently here? \n- I would think a global action-conditional forward model (represented as convnet+deconvnet, and trained unrolled on its own predictions to reduce model errors) would perform quite well on the 2D navigation task, and possibly on the reacher task. Even though these are represented as images, they are very simple images with little distracting information, no changes in illumination/perspective, etc. It seems the model essentially just needs to learn a pixel translation for each action for the navigation task, and some rotations for the reacher. It already seems to work quite well for the non-holonomic car, which requires learning similar transformations. This baseline should be included for all the tasks. \n- Although it does seem that the method performs well on the stacking tasks for the real robot, there are no baselines included. However, there are many works which have explored representation learning and control for robotics using neural networks. A couple examples (+see references within):\n\n\"Learning to poke by poking: Experiential learning of intuitive physics\" Pulkit Agrawal, Ashvin V Nair, Pieter Abbeel, Jitendra Malik, Sergey Levine. NIPS 2016\n\"Deep Visual Foresight for Planning Robot Motion\" Chelsea Finn, Sergey Levine ICRA 2017\n\nAt the very least, the method should be compared to pixel-based global models and representations learned with some kind of autoencoder or forward model for the robot task. \n\nThe paper proposes what seems to be a good idea, but it is not yet demonstrated by the current experiments. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, but insufficient comparison to baselines", "review": "This paper proposes a model-based reinforcement learning approach, called SOLAR, \nwhich consists of mapping complex, high-dimensional observations to low-dimensional \nrepresentations where transition dynamics between consecutive states are approximately linear. \nIn this low-dimensional space, local models can easily be fit in closed form and then used to optimize a policy, using a similar method to Guided Policy Search (GPS). The method is evaluated in 4 different settings (3 simulated, 1 on a real robot). \n\n*Quality: the method seems to work well in the experiments. However, there are issues with the experimental evaluation (detailed below) which make it unclear whether the method is better than standard baselines.\n\n*Clarity: the paper is well-written and clear overall.  \n\n*Originality: the paper proposes an extension of GPS, which to my knowledge is novel. \n\n*Significance: the idea of learning representations where transitions are linear seems well-founded and potentially useful. However the merits of this method are not yet clear from the experiments. \n\n\nSpecific Comments:\n\n- Please include an illustration of the 2D navigation task in Figure 3a\n- I'm confused by the poor performance of E2C in the 2D navigation task. \nThe previous works of [Watter et. al, 2015] and [Banijalami et. al, 2017] report close to 100% accuracy using similar methods. Is the task formulated differently here? \n- I would think a global action-conditional forward model (represented as convnet+deconvnet, and trained unrolled on its own predictions to reduce model errors) would perform quite well on the 2D navigation task, and possibly on the reacher task. Even though these are represented as images, they are very simple images with little distracting information, no changes in illumination/perspective, etc. It seems the model essentially just needs to learn a pixel translation for each action for the navigation task, and some rotations for the reacher. It already seems to work quite well for the non-holonomic car, which requires learning similar transformations. This baseline should be included for all the tasks. \n- Although it does seem that the method performs well on the stacking tasks for the real robot, there are no baselines included. However, there are many works which have explored representation learning and control for robotics using neural networks. A couple examples (+see references within):\n\n\"Learning to poke by poking: Experiential learning of intuitive physics\" Pulkit Agrawal, Ashvin V Nair, Pieter Abbeel, Jitendra Malik, Sergey Levine. NIPS 2016\n\"Deep Visual Foresight for Planning Robot Motion\" Chelsea Finn, Sergey Levine ICRA 2017\n\nAt the very least, the method should be compared to pixel-based global models and representations learned with some kind of autoencoder or forward model for the robot task. \n\nThe paper proposes what seems to be a good idea, but it is not yet demonstrated by the current experiments. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541474703366}, {"id": "S1gFTfz537", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper831/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this work the authors propose an end to end approach for model based reinforcement learning from images, where the main building blocks are locally-linear dynamical systems and variational auto-encoders (VAE). Specifically, it is assumed that the input features (i.e., the images) are generated from a low dimensional latent representation mapped through parametric random functions; the latter are modeled via neural networks. A recognition model based on convolutional neural networks operates on the reverse way and is responsible for projecting the input features to the latent space, in order to proceed with the reinforcement learning task. The variational framework is employed in order to jointly learn the VAE and the linear dynamics on the latent state. As a final step, once the model is fitted a linear quadratic system (LQS) is solved in order to learn the cost function and the optimal policy. \n\n* The paper is well motivated and tries to solve an interesting problem, that of data-efficient reinforcement learning. The experiments are well picked and demonstrate the advantages of the proposed approach towards solving the task, however, the method is only evaluated on few environments and compared against only a couple of other methods. I would expect a broader evaluation and/or comparison against more methods. Since the model is able to reach TRPO’s performance in much less steps it would be nice to see how it performs against PPO from [Schulman et al. 2017] (at least on the simulated environments). Also, would it make sense to compare against [Levine et al. 2016] that has been evaluated on similar tasks?\n\n[Schulman et al 2017] “Proximal Policy Optimization Algorithms”.\n[Levine et al. 2016] “End-to-End Training of Deep Visuomotor Policies”.\n\n* Methodologically, the paper is sound. The model part (as the authors point out) is based on [Johnson et al. 2016] and is well explained. On the other hand, the policy part, and in particular the policy update in Section 4.2 has some issues regarding readability. There is a strong interplay between Section 4.2, Section 2.1 and Appendix D and the authors did not manage to nicely explain what exactly is happening during the update phase. In the beginning the reader has the impression that we are finding the optimal policy via the closed-form LQS. Later on we switch to constrained optimisation for the cost by accounting for the KL divergence between the policy on two episodes. Finally, in the appendix we are back to the original quadratic cost. The authors need to clarify all the above. Also, they need to explicitly mention why they opt for stochastic optimisation (is it because of minibatching?)\n\n* To continue with the policy, in Section 4.2 the authors argue that although the optimal policy can be found in closed form this is not desirable because the policy will overfit the model and will not generalise well in the real environment. I disagree with this statement. If this happens it effectively means that the learned model or the assumption/learning of the linear dynamics is not right. The authors seem to also agree with this since they clearly state in the the experimental section that “... our method does not heavily rely on an accurate model...”. To my understanding, this means that we need to refine the modelling strategy and not learn a sub-optimal policy. I am really interested in the authors opinion on that.\n\n* The above argument is also directly related to the recognition model and learning of the policy in the latent state (I completely agree with that). The recognition network, which in this case is a convolutional neural network, is used as an inference mechanism to project the observations to the latent space. We learn the (variational) parameters of the recognition model by optimising the likelihood’s lower bound. This means that we are “allowed” to overfit the variational parameters as long as the bound gets tighter. This can possibly result in degraded performance during the policy update. Furthermore, the variational distribution of the latent state, i.e., q(z_t | s_t) is assumed to be mean field across time (independent z’s), while clearly this is not the case in the posterior. You somehow mitigate that by augmenting the observed state (feeding consecutive frames to the network), but still this is not ideal. Finally, is there a reason why we only use the mean of the recognition model to fit the cost on the projected latent states? Why are we throwing away the uncertainty? Especially since you do not use an exact solver and follow a stochastic gradient.\n\n* In the end of Section 2.1, the authors argue regarding the fact that the prior work assumes access to a compact low-dimensional representation which does not allow them to perform well on images. Reference is needed.\n\n* In the related work the authors mention modelling bias as a downside of prior work. Can you please elaborate on that? Where does the bias come from and, more importantly, how does your approach overcome this issue?\n\n* In the experiment and specifically in Figure 4 am I right in assuming that the distance to target is measured in actual pixels? Furthermore, why the relevant plot for the reacher task is depicting rewards instead of the distance to target. To me this suggests that the task is not solved. In general what I find very upsetting in the field are plots that only depict accumulated reward for a specific task. There are many situations where the agent learns a weird behaviour that happens to give good rewards (e.g., spinning around the cart-pole), and unfortunately such behaviours are not spotted on the reward plots.\n\nOverall, the paper is nicely presented and definitely an interesting work. However, given the fact that methodologically we have not learned anything new from this paper and in combination with the not satisfying experimental evaluation I warrant for rejection.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice work but methodology is not clear enough in some parts. It can also benefit from a broader experimental evaluation.", "review": "In this work the authors propose an end to end approach for model based reinforcement learning from images, where the main building blocks are locally-linear dynamical systems and variational auto-encoders (VAE). Specifically, it is assumed that the input features (i.e., the images) are generated from a low dimensional latent representation mapped through parametric random functions; the latter are modeled via neural networks. A recognition model based on convolutional neural networks operates on the reverse way and is responsible for projecting the input features to the latent space, in order to proceed with the reinforcement learning task. The variational framework is employed in order to jointly learn the VAE and the linear dynamics on the latent state. As a final step, once the model is fitted a linear quadratic system (LQS) is solved in order to learn the cost function and the optimal policy. \n\n* The paper is well motivated and tries to solve an interesting problem, that of data-efficient reinforcement learning. The experiments are well picked and demonstrate the advantages of the proposed approach towards solving the task, however, the method is only evaluated on few environments and compared against only a couple of other methods. I would expect a broader evaluation and/or comparison against more methods. Since the model is able to reach TRPO’s performance in much less steps it would be nice to see how it performs against PPO from [Schulman et al. 2017] (at least on the simulated environments). Also, would it make sense to compare against [Levine et al. 2016] that has been evaluated on similar tasks?\n\n[Schulman et al 2017] “Proximal Policy Optimization Algorithms”.\n[Levine et al. 2016] “End-to-End Training of Deep Visuomotor Policies”.\n\n* Methodologically, the paper is sound. The model part (as the authors point out) is based on [Johnson et al. 2016] and is well explained. On the other hand, the policy part, and in particular the policy update in Section 4.2 has some issues regarding readability. There is a strong interplay between Section 4.2, Section 2.1 and Appendix D and the authors did not manage to nicely explain what exactly is happening during the update phase. In the beginning the reader has the impression that we are finding the optimal policy via the closed-form LQS. Later on we switch to constrained optimisation for the cost by accounting for the KL divergence between the policy on two episodes. Finally, in the appendix we are back to the original quadratic cost. The authors need to clarify all the above. Also, they need to explicitly mention why they opt for stochastic optimisation (is it because of minibatching?)\n\n* To continue with the policy, in Section 4.2 the authors argue that although the optimal policy can be found in closed form this is not desirable because the policy will overfit the model and will not generalise well in the real environment. I disagree with this statement. If this happens it effectively means that the learned model or the assumption/learning of the linear dynamics is not right. The authors seem to also agree with this since they clearly state in the the experimental section that “... our method does not heavily rely on an accurate model...”. To my understanding, this means that we need to refine the modelling strategy and not learn a sub-optimal policy. I am really interested in the authors opinion on that.\n\n* The above argument is also directly related to the recognition model and learning of the policy in the latent state (I completely agree with that). The recognition network, which in this case is a convolutional neural network, is used as an inference mechanism to project the observations to the latent space. We learn the (variational) parameters of the recognition model by optimising the likelihood’s lower bound. This means that we are “allowed” to overfit the variational parameters as long as the bound gets tighter. This can possibly result in degraded performance during the policy update. Furthermore, the variational distribution of the latent state, i.e., q(z_t | s_t) is assumed to be mean field across time (independent z’s), while clearly this is not the case in the posterior. You somehow mitigate that by augmenting the observed state (feeding consecutive frames to the network), but still this is not ideal. Finally, is there a reason why we only use the mean of the recognition model to fit the cost on the projected latent states? Why are we throwing away the uncertainty? Especially since you do not use an exact solver and follow a stochastic gradient.\n\n* In the end of Section 2.1, the authors argue regarding the fact that the prior work assumes access to a compact low-dimensional representation which does not allow them to perform well on images. Reference is needed.\n\n* In the related work the authors mention modelling bias as a downside of prior work. Can you please elaborate on that? Where does the bias come from and, more importantly, how does your approach overcome this issue?\n\n* In the experiment and specifically in Figure 4 am I right in assuming that the distance to target is measured in actual pixels? Furthermore, why the relevant plot for the reacher task is depicting rewards instead of the distance to target. To me this suggests that the task is not solved. In general what I find very upsetting in the field are plots that only depict accumulated reward for a specific task. There are many situations where the agent learns a weird behaviour that happens to give good rewards (e.g., spinning around the cart-pole), and unfortunately such behaviours are not spotted on the reward plots.\n\nOverall, the paper is nicely presented and definitely an interesting work. However, given the fact that methodologically we have not learned anything new from this paper and in combination with the not satisfying experimental evaluation I warrant for rejection.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541182145184}], "openreview_url": "https://openreview.net/forum?id=Bke96sC5tm", "arxiv_id": "1808.09105", "paper_pdf": "papers/Bke96sC5tm.pdf", "paper_pdf_sha256": "65447d057605cd4d9d71bf56795f9b7a786b6fa8b383ad976d440a77f01376bf", "paper_pdf_bytes": 1703454, "paper_pdf_source": "openreview", "code_url": "https://github.com/sharadmv/parasol", "code_repository": "sharadmv/parasol", "code_commit": "493d87829691428b7183bc6432200032ac86f835", "code_archive": "repos/Bke96sC5tm.zip", "code_archive_sha256": "cc76e2cb8a3aecd19765bcf5bdebc5f48d4be5f6170400a9b2171dc10a593d4b", "code_archive_bytes": 2603953, "code_file_count": 44, "code_extensions": {".py": 44}, "github_disk_usage_kb": 2883, "github_languages": {"Python": 133967}, "github_archived": false, "github_pushed_at": "2021-09-08T00:46:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/solar-deep-structured-representations-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6hdzuMTNmX", "year": 2026, "status": "rejected", "title": "SIMSHIFT: A Benchmark for Adapting Neural Surrogates to Distribution Shifts", "authors": ["Paul Setinek", "Gianluca Galletti", "Thomas Gross", "Dominik Schnürer", "Johannes Brandstetter", "Werner Zellinger"], "authorids": ["~Paul_Setinek1", "~Gianluca_Galletti1", "~Thomas_Gross2", "~Dominik_Schnürer1", "~Johannes_Brandstetter1", "~Werner_Zellinger1"], "authors_source": "OpenReview API", "abstract": "Neural surrogates for Partial Differential Equations (PDEs) often suffer significant performance degradation when evaluated on unseen problem configurations, such as new initial conditions or structural dimensions.\nMeanwhile, Domain Adaptation (DA) techniques have been widely used in vision and language processing to generalize from limited information about unseen configurations.\nIn this work, we address this gap through two focused contributions.\nFirst, we introduce SIMSHIFT, a novel benchmark dataset and evaluation suite composed of four industrial simulation tasks spanning diverse processes and physics: _hot rolling_, _sheet metal forming_, _electric motor design_ and _heatsink design_.\nSecond, we extend established DA methods to state-of-the-art neural surrogates and systematically evaluate them.\nThese approaches use parametric descriptions and ground truth simulations from multiple source configurations, together with only parametric descriptions from target configurations.\nThe goal is to accurately predict target simulations without access to ground truth simulation data.\nExtensive experiments on SIMSHIFT highlight the challenges of out of distribution neural surrogate modeling, demonstrate the potential of DA in simulation, and reveal open problems in achieving robust neural surrogates under distribution shifts in industrially relevant scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "1rssc4bWqC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16896/Reviewer_BN5C"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper tackles the problem of unsupervised domain adaptation (UDA) in the scientific machine learning context, for modeling PDE-governed processes. Specifically, the models evaluated in the paper aim to perform UDA under the covariate-shift assumption and employing the Radon-Nikodym derivative for learning distributional transfer from a source distribution to a target distribution where ground-truth solutions are only present in the source distribution. To demonstrate UDA capability, the scientific processes the paper focuses on are self-generated datasets of steady-state processes with applications in metallurgy, manufacturing, machinery and electronics. To evaluate the UDA capability of the proposed methods, the source and target domains are created by splitting the input parameter configurations in a disjoint manner across sorce and target domains. Further, the paper also details performance of state of the art neural surrogates and unsupervised domain adaptation (UDA) techniques on the proposed dataset. Overall, the dataset and the proposed model benchmark evaluations are a good contribution to push research on the critical problem of UDA in the scientific machine learning context.", "review_text": "This paper tackles the problem of unsupervised domain adaptation (UDA) in the scientific machine learning context, for modeling PDE-governed processes. Specifically, the models evaluated in the paper aim to perform UDA under the covariate-shift assumption and employing the Radon-Nikodym derivative for learning distributional transfer from a source distribution to a target distribution where ground-truth solutions are only present in the source distribution. To demonstrate UDA capability, the scientific processes the paper focuses on are self-generated datasets of steady-state processes with applications in metallurgy, manufacturing, machinery and electronics. To evaluate the UDA capability of the proposed methods, the source and target domains are created by splitting the input parameter configurations in a disjoint manner across sorce and target domains. Further, the paper also details performance of state of the art neural surrogates and unsupervised domain adaptation (UDA) techniques on the proposed dataset. Overall, the dataset and the proposed model benchmark evaluations are a good contribution to push research on the critical problem of UDA in the scientific machine learning context.", "strengths": "1. The investigated context of unsupervised domain adaptation (UDA) is a critical requirement in the scientific machine learning context because, for a neural surrogate to be useful, it needs to be able to \"extrapolate\" to unseen parameter configurations. This important problem has (unfortunately) not been investigated in depth as yet by the scientific machine learning community. Hence, this paper is a good start towards proposing a benchmark dataset as well as quantitative evaluations of state-of-the-art surrogates,  model selection techniques to perform the UDA task. \n\n2. The four problems investigated by the paper are (i) diverse (ii) each challenging in their own right (iii) well-motivated for the UDA problem as they each require costly computational simulation for data generation.", "weaknesses": "1. The paper requires better organization as there are multiple critical sections where the appendix is referred to, interrupting the flow of the paper and preventing the full understanding of the problem. One such point is the lack of clear demarcation of source and target domains for the four datasets in the main paper (this is done in the appendix in Appendices F1 - F4 where actual parameter ranges are discussed and Appendix C:Table 11 where actual source and target distribution splits are presented).  \n\n2. Full cost of the evaluated UDA methods (i.e., surrogate training with UDA + model selection) is unclear from the main paper as presented. For a holistic presentation as a benchmark, this is necessary. Understanding how the training cost scales with (i) number of data points (ii) mesh size, is crucial for independent adaptation of the proposed methods by interested readers to other UDA contexts.", "questions": "1. How are the specific parameter ranges for `source configurations` and `target configurations` arrived at for each of the four datasets? Also, how are the \"easy\", \"medium\" and \"hard\" configurations decided? A more pronounced and cohesive description in the main paper of the actual configurations (i.e., parameter ranges), and the classification into \"easy\", \"medium\" and hard\" would help to clarify the problem design.\n\n2. Are there examples in any of the datasets investigated, where a significant distributional shift occurs in the underlying dynamics? Meaning, even if P_s(Y|X) = P_t(Y'|X'), are there any examples where P_t(Y') is itself out-of-distribution relative to P_s(Y)? Reporting model performance (or identifying any failures) in such cases would also be insightful.\n\n3. Could you please provide numbers for (or pointer to a part of the paper that discusses) total training time of the proposed methods (i.e., surrogate training with UDA, model selection). Specifically, it would be useful to understand how total training time scales with training data size as well as the mesh size (i.e., number of mesh nodes)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the problem of unsupervised domain adaptation (UDA) in the scientific machine learning context, for modeling PDE-governed processes. Specifically, the models evaluated in the paper aim to perform UDA under the covariate-shift assumption and employing the Radon-Nikodym derivative for learning distributional transfer from a source distribution to a target distribution where ground-truth solutions are only present in the source distribution. To demonstrate UDA capability, the scientific processes the paper focuses on are self-generated datasets of steady-state processes with applications in metallurgy, manufacturing, machinery and electronics. To evaluate the UDA capability of the proposed methods, the source and target domains are created by splitting the input parameter configurations in a disjoint manner across sorce and target domains. Further, the paper also details performance of state of the art neural surrogates and unsupervised domain adaptation (UDA) techniques on the proposed dataset. Overall, the dataset and the proposed model benchmark evaluations are a good contribution to push research on the critical problem of UDA in the scientific machine learning context.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The investigated context of unsupervised domain adaptation (UDA) is a critical requirement in the scientific machine learning context because, for a neural surrogate to be useful, it needs to be able to \"extrapolate\" to unseen parameter configurations. This important problem has (unfortunately) not been investigated in depth as yet by the scientific machine learning community. Hence, this paper is a good start towards proposing a benchmark dataset as well as quantitative evaluations of state-of-the-art surrogates,  model selection techniques to perform the UDA task. \n\n2. The four problems investigated by the paper are (i) diverse (ii) each challenging in their own right (iii) well-motivated for the UDA problem as they each require costly computational simulation for data generation.", "weaknesses": "1. The paper requires better organization as there are multiple critical sections where the appendix is referred to, interrupting the flow of the paper and preventing the full understanding of the problem. One such point is the lack of clear demarcation of source and target domains for the four datasets in the main paper (this is done in the appendix in Appendices F1 - F4 where actual parameter ranges are discussed and Appendix C:Table 11 where actual source and target distribution splits are presented).  \n\n2. Full cost of the evaluated UDA methods (i.e., surrogate training with UDA + model selection) is unclear from the main paper as presented. For a holistic presentation as a benchmark, this is necessary. Understanding how the training cost scales with (i) number of data points (ii) mesh size, is crucial for independent adaptation of the proposed methods by interested readers to other UDA contexts.", "questions": "1. How are the specific parameter ranges for `source configurations` and `target configurations` arrived at for each of the four datasets? Also, how are the \"easy\", \"medium\" and \"hard\" configurations decided? A more pronounced and cohesive description in the main paper of the actual configurations (i.e., parameter ranges), and the classification into \"easy\", \"medium\" and hard\" would help to clarify the problem design.\n\n2. Are there examples in any of the datasets investigated, where a significant distributional shift occurs in the underlying dynamics? Meaning, even if P_s(Y|X) = P_t(Y'|X'), are there any examples where P_t(Y') is itself out-of-distribution relative to P_s(Y)? Reporting model performance (or identifying any failures) in such cases would also be insightful.\n\n3. Could you please provide numbers for (or pointer to a part of the paper that discusses) total training time of the proposed methods (i.e., surrogate training with UDA, model selection). Specifically, it would be useful to understand how total training time scales with training data size as well as the mesh size (i.e., number of mesh nodes)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762114202072}, {"id": "YYymihESiG", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16896/Reviewer_FxNP"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper introduces SIMSHIFT, a benchmark dataset and evaluation suite covering four industrial simulation tasks to study UDA for neural surrogates on unstructured meshes. The framework explicitly conditions neural operators on configuration parameters via a sinusoidal (sin–cos) encoding followed by a shallow MLP, producing an 8-dimensional latent vector used for conditioning; baselines include PointNet, GraphSAGE (FiLM), Transolver (DiT), and UPT selected per task/scale. The study benchmarks three UDA algorithms (Deep CORAL, CMD, DANN) together with unsupervised model selection and reports metrics such as (N)RMSE and custom engineering metrics. All datasets are steady-state by design, and the heatsink meshes are subsampled to one quarter, with experiments showing that UDA+ unsupervised selection usually reduces target error but no single method dominates.", "review_text": "The paper introduces SIMSHIFT, a benchmark dataset and evaluation suite covering four industrial simulation tasks to study UDA for neural surrogates on unstructured meshes. The framework explicitly conditions neural operators on configuration parameters via a sinusoidal (sin–cos) encoding followed by a shallow MLP, producing an 8-dimensional latent vector used for conditioning; baselines include PointNet, GraphSAGE (FiLM), Transolver (DiT), and UPT selected per task/scale. The study benchmarks three UDA algorithms (Deep CORAL, CMD, DANN) together with unsupervised model selection and reports metrics such as (N)RMSE and custom engineering metrics. All datasets are steady-state by design, and the heatsink meshes are subsampled to one quarter, with experiments showing that UDA+ unsupervised selection usually reduces target error but no single method dominates.", "strengths": "1. All architectures share the same conditioning network (sin–cos + shallow MLP, 8-dim latent), making the parametric setup consistent across models. \n\n2. PointNet (global pooling), GraphSAGE with FiLM (local message passing), Transolver with DiT (attention with learned slicing), and UPT (latent field modeling for very large meshes) are adapted to conditional mesh regression and chosen to match scale constraints.", "weaknesses": "1. The work extends established DA methods to neural surrogates and benchmarks them, it does not introduce a new neural operator or a new UDA/model-selection algorithm (conditioning, FiLM/DiT choices are standardized, not novel). \n\n2. While the paper reviews neural operator literature, the baseline lineup focuses on PointNet/ GNN/ Transformer/ UPT variants; adding FNO/Geo-FNO/GKN comparisons would better situate results, especially in 3D-large-mesh regimes.\n\n3. The datasets are steady-state only, which may limit claims about transient dynamics and extreme-scale fidelity.", "questions": "1. To sharpen the model-side contribution, could you add and ablate conditioning variants (e.g., learned frequency embeddings, FiLM/DiT) and report stability/accuracy sensitivity?\n\n2. Can you provide ablations that incorporate physics-informed constraints (residuals/conservation/BCs) into the loss within the same training–selection pipeline to assess reliability for industrial deployment?\n\n3. Will you include operator baselines (e.g., FNO/Geo-FNO/GKN/GNO)—ideally on the 3D heatsink—to clarify accuracy–efficiency trade-offs against widely used PDE surrogates?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SIMSHIFT, a benchmark dataset and evaluation suite covering four industrial simulation tasks to study UDA for neural surrogates on unstructured meshes. The framework explicitly conditions neural operators on configuration parameters via a sinusoidal (sin–cos) encoding followed by a shallow MLP, producing an 8-dimensional latent vector used for conditioning; baselines include PointNet, GraphSAGE (FiLM), Transolver (DiT), and UPT selected per task/scale. The study benchmarks three UDA algorithms (Deep CORAL, CMD, DANN) together with unsupervised model selection and reports metrics such as (N)RMSE and custom engineering metrics. All datasets are steady-state by design, and the heatsink meshes are subsampled to one quarter, with experiments showing that UDA+ unsupervised selection usually reduces target error but no single method dominates.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. All architectures share the same conditioning network (sin–cos + shallow MLP, 8-dim latent), making the parametric setup consistent across models. \n\n2. PointNet (global pooling), GraphSAGE with FiLM (local message passing), Transolver with DiT (attention with learned slicing), and UPT (latent field modeling for very large meshes) are adapted to conditional mesh regression and chosen to match scale constraints.", "weaknesses": "1. The work extends established DA methods to neural surrogates and benchmarks them, it does not introduce a new neural operator or a new UDA/model-selection algorithm (conditioning, FiLM/DiT choices are standardized, not novel). \n\n2. While the paper reviews neural operator literature, the baseline lineup focuses on PointNet/ GNN/ Transformer/ UPT variants; adding FNO/Geo-FNO/GKN comparisons would better situate results, especially in 3D-large-mesh regimes.\n\n3. The datasets are steady-state only, which may limit claims about transient dynamics and extreme-scale fidelity.", "questions": "1. To sharpen the model-side contribution, could you add and ablate conditioning variants (e.g., learned frequency embeddings, FiLM/DiT) and report stability/accuracy sensitivity?\n\n2. Can you provide ablations that incorporate physics-informed constraints (residuals/conservation/BCs) into the loss within the same training–selection pipeline to assess reliability for industrial deployment?\n\n3. Will you include operator baselines (e.g., FNO/Geo-FNO/GKN/GNO)—ideally on the 3D heatsink—to clarify accuracy–efficiency trade-offs against widely used PDE surrogates?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761998122977}, {"id": "3aMsWm37lh", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16896/Reviewer_uay5"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper provides a benchmark dataset with 4 industry relevant simulations with the goal of solving the unsupervised domain adaptation (UDA) problem for neural surrogates. They provide evaluate error across multiple baseline models, UDA methods, and unsupevised model selction methods.", "review_text": "The paper provides a benchmark dataset with 4 industry relevant simulations with the goal of solving the unsupervised domain adaptation (UDA) problem for neural surrogates. They provide evaluate error across multiple baseline models, UDA methods, and unsupevised model selction methods.", "strengths": "- The paper provides code and data for industry relevant applications.\n- Multiple models, UDA methods, and unsupervised model selection methods are tested.\n- The supplied data and experimental details are thoroughly explained and provide a nice test bed for evaluating steady state problems with UDA for other researchers.", "weaknesses": "- It would be interesting to see the author's hypotheses on why certain UDA mehods and unsupervised model selection methods worked well in specific datasets.\n- Steady state only problems limit the scope of this benchmark, adding time dependent problems would also be useful.", "questions": "- For the unsupervised domain adaptation methods, which distance metric $d$ do you use?\n- In Figure 1 of the paper the caption refers to two loss terms $\\mathcal{L}_{recon}$ and $\\mathcal{L}_{DA}$ but these terms are not defined in the main text as far as I could tell.\n- There are a few typos in the paper which should be corrected:\n  - Line 105: \"Numerous different Numerous benchmark datasets and evaluation protocols have been established\" \"Numerous\" appears twice.\n  - Line \"Detailed and descriptions of the parameter sampling ranges can be found in Appendix F.\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper provides a benchmark dataset with 4 industry relevant simulations with the goal of solving the unsupervised domain adaptation (UDA) problem for neural surrogates. They provide evaluate error across multiple baseline models, UDA methods, and unsupevised model selction methods.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The paper provides code and data for industry relevant applications.\n- Multiple models, UDA methods, and unsupervised model selection methods are tested.\n- The supplied data and experimental details are thoroughly explained and provide a nice test bed for evaluating steady state problems with UDA for other researchers.", "weaknesses": "- It would be interesting to see the author's hypotheses on why certain UDA mehods and unsupervised model selection methods worked well in specific datasets.\n- Steady state only problems limit the scope of this benchmark, adding time dependent problems would also be useful.", "questions": "- For the unsupervised domain adaptation methods, which distance metric $d$ do you use?\n- In Figure 1 of the paper the caption refers to two loss terms $\\mathcal{L}_{recon}$ and $\\mathcal{L}_{DA}$ but these terms are not defined in the main text as far as I could tell.\n- There are a few typos in the paper which should be corrected:\n  - Line 105: \"Numerous different Numerous benchmark datasets and evaluation protocols have been established\" \"Numerous\" appears twice.\n  - Line \"Detailed and descriptions of the parameter sampling ranges can be found in Appendix F.\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761953129533}, {"id": "4iwRbBHxqn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16896/Reviewer_hbC7"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces SIMSHIFT, a benchmark and dataset suite designed to evaluate unsupervised domain adaptation (UDA) methods for neural surrogates trained on physics-based simulations. It provides four industrially relevant simulation datasets (rolling, forming, motor, and heatsink design) and evaluates classical UDA algorithms (CMD, Deep CORAL, and DANN) across several neural surrogate architectures (PointNet, GraphSAGE, Transolver, UPT). The goal is to quantify performance degradation under parametric distribution shifts and assess whether domain adaptation can recover generalization in unseen configurations.", "review_text": "The paper introduces SIMSHIFT, a benchmark and dataset suite designed to evaluate unsupervised domain adaptation (UDA) methods for neural surrogates trained on physics-based simulations. It provides four industrially relevant simulation datasets (rolling, forming, motor, and heatsink design) and evaluates classical UDA algorithms (CMD, Deep CORAL, and DANN) across several neural surrogate architectures (PointNet, GraphSAGE, Transolver, UPT). The goal is to quantify performance degradation under parametric distribution shifts and assess whether domain adaptation can recover generalization in unseen configurations.", "strengths": "- While neural surrogates themselves are not new, the authors are arguably the first to systematically benchmark domain adaptation methods in this context, bridging simulation modeling and UDA research.\n- The proposed datasets are realistic, diverse, and physically grounded, covering a range of FEM and CFD problems from different industries.\n- The benchmark is well-documented, publicly released, and accompanied by detailed simulation parameters, architectures, and hyperparameter sweeps.\n- The study runs a large number of controlled experiments with multiple architectures, adaptation methods, and model-selection strategies, highlighting consistent trends and challenges in out-of-distribution generalization.", "weaknesses": "- The benchmark primarily relies on pre-2019 UDA methods (CMD, Deep CORAL, DANN). More modern approaches, e.g., adversarial information bottleneck, self-supervised alignment, or diffusion-based DA, could offer stronger baselines and better contextualize the results.\n- All benchmarked shifts are parametric (covariate) shifts, while the introduction also mentions “instrument shifts” such as mesh geometry or solver discretization changes, which are not actually tested. Including such structural or concept-level shifts would significantly increase the benchmark’s value.\n- The current metrics focus on RMSE or custom geometric error terms. The benchmark does not assess physical validity (e.g., conservation of energy, boundary-condition satisfaction, equilibrium consistency) of adapted surrogate predictions, which is critical in engineering contexts.\n- The work is primarily a benchmark contribution, not a conceptual advance in neural surrogate modeling or domain adaptation algorithms.\n- The distinction between “neural surrogates” and generic regression surrogates (MLPs) seems somewhat artificial; many of these architectures are used interchangeably in most industrial ML contexts, whether simulation-based or data-driven. The framing could overstate novelty relative to standard supervised regression with UDA.", "questions": "- Could the authors benchmark recent UDA or domain generalization methods, such as SWAD, Tent, AdaBN, or contrastive/self-supervised approaches, to contextualize performance trends?\n- Can non-parametric shifts (e.g., mesh geometry changes, discretization refinements, or modified boundary conditions) be introduced to better reflect the “instrument shift” mentioned in the motivation?\n- Would it be possible to integrate physics-based evaluation criteria, such as PDE residuals or constraint violation penalties, into the assessment?\n- Are all datasets equally sensitive to domain gaps? A cross-dataset or per-task ablation might reveal which physics or modeling types pose the greatest adaptation challenges.\n- Finally, could the authors clarify the extent to which the surrogates incorporate physics priors (operator learning, inductive biases) versus being purely data-driven regressors?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SIMSHIFT, a benchmark and dataset suite designed to evaluate unsupervised domain adaptation (UDA) methods for neural surrogates trained on physics-based simulations. It provides four industrially relevant simulation datasets (rolling, forming, motor, and heatsink design) and evaluates classical UDA algorithms (CMD, Deep CORAL, and DANN) across several neural surrogate architectures (PointNet, GraphSAGE, Transolver, UPT). The goal is to quantify performance degradation under parametric distribution shifts and assess whether domain adaptation can recover generalization in unseen configurations.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- While neural surrogates themselves are not new, the authors are arguably the first to systematically benchmark domain adaptation methods in this context, bridging simulation modeling and UDA research.\n- The proposed datasets are realistic, diverse, and physically grounded, covering a range of FEM and CFD problems from different industries.\n- The benchmark is well-documented, publicly released, and accompanied by detailed simulation parameters, architectures, and hyperparameter sweeps.\n- The study runs a large number of controlled experiments with multiple architectures, adaptation methods, and model-selection strategies, highlighting consistent trends and challenges in out-of-distribution generalization.", "weaknesses": "- The benchmark primarily relies on pre-2019 UDA methods (CMD, Deep CORAL, DANN). More modern approaches, e.g., adversarial information bottleneck, self-supervised alignment, or diffusion-based DA, could offer stronger baselines and better contextualize the results.\n- All benchmarked shifts are parametric (covariate) shifts, while the introduction also mentions “instrument shifts” such as mesh geometry or solver discretization changes, which are not actually tested. Including such structural or concept-level shifts would significantly increase the benchmark’s value.\n- The current metrics focus on RMSE or custom geometric error terms. The benchmark does not assess physical validity (e.g., conservation of energy, boundary-condition satisfaction, equilibrium consistency) of adapted surrogate predictions, which is critical in engineering contexts.\n- The work is primarily a benchmark contribution, not a conceptual advance in neural surrogate modeling or domain adaptation algorithms.\n- The distinction between “neural surrogates” and generic regression surrogates (MLPs) seems somewhat artificial; many of these architectures are used interchangeably in most industrial ML contexts, whether simulation-based or data-driven. The framing could overstate novelty relative to standard supervised regression with UDA.", "questions": "- Could the authors benchmark recent UDA or domain generalization methods, such as SWAD, Tent, AdaBN, or contrastive/self-supervised approaches, to contextualize performance trends?\n- Can non-parametric shifts (e.g., mesh geometry changes, discretization refinements, or modified boundary conditions) be introduced to better reflect the “instrument shift” mentioned in the motivation?\n- Would it be possible to integrate physics-based evaluation criteria, such as PDE residuals or constraint violation penalties, into the assessment?\n- Are all datasets equally sensitive to domain gaps? A cross-dataset or per-task ablation might reveal which physics or modeling types pose the greatest adaptation challenges.\n- Finally, could the authors clarify the extent to which the surrogates incorporate physics priors (operator learning, inductive biases) versus being purely data-driven regressors?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761845790880}], "openreview_url": "https://openreview.net/forum?id=6hdzuMTNmX", "arxiv_id": "2506.12007", "paper_pdf": "papers/6hdzuMTNmX.pdf", "paper_pdf_sha256": "16b963705e3f90de7912fe3357b27b807f5110e86b308150eee45a2b84628e09", "paper_pdf_bytes": 11056205, "paper_pdf_source": "openreview", "code_url": "https://github.com/psetinek/simshift", "code_repository": "psetinek/simshift", "code_commit": "d30db3951e8aae422aeba4329781c9424745db37", "code_archive": "repos/6hdzuMTNmX.zip", "code_archive_sha256": "3adbd7373b26ae27f514f15c957cbcfba2f40719356afc161b28be5de77d187f", "code_archive_bytes": 880923, "code_file_count": 55, "code_extensions": {".py": 53, ".ipynb": 2}, "github_disk_usage_kb": 1095, "github_languages": {"Python": 294481, "Jupyter Notebook": 203201}, "github_archived": false, "github_pushed_at": "2026-05-19T13:44:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/simshift-a-benchmark-for-adapting-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4Po8d9GAfQ", "year": 2025, "status": "rejected", "title": "Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding", "authors": ["Haolin Chen", "Yihao Feng", "Zuxin Liu", "Weiran Yao", "Akshara Prabhakar", "Shelby Heinecke", "Ricky Ho", "Phil L Mui", "Silvio Savarese", "Caiming Xiong", "Huan Wang"], "authorids": ["~Haolin_Chen2", "~Yihao_Feng1", "~Zuxin_Liu1", "~Weiran_Yao1", "~Akshara_Prabhakar1", "~Shelby_Heinecke1", "~Ricky_Ho2", "~Phil_L_Mui1", "~Silvio_Savarese1", "~Caiming_Xiong1", "~Huan_Wang1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have shown impressive capabilities, but still struggle with complex reasoning tasks requiring multiple steps. While prompt-based methods like Chain-of-Thought (CoT) can improve LLM reasoning at inference time, optimizing reasoning capabilities during training remains challenging. We introduce LaTent Reasoning Optimization (LaTRO), a principled framework that formulates reasoning as sampling from a latent distribution and optimizes it via variational approaches. LaTRO enables LLMs to concurrently improve both their reasoning process and ability to evaluate reasoning quality, without requiring external feedback or reward models. We validate LaTRO through experiments on GSM8K and ARC-Challenge datasets using multiple model architectures. On GSM8K, LaTRO improves zero-shot accuracy by an average of 12.5\\% over base models and 9.6\\% over supervised fine-tuning across Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B. Our findings suggest that pre-trained LLMs possess latent reasoning capabilities that can be unlocked and enhanced through our proposed optimization approach in a self-improvement manner.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "tzaJhYzJNn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9035/Reviewer_b1ci"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The work focuses on enhancing the reasoning abilities of large language models (LLMs) during the training phase without relying on external feedback. The authors motivate the work by raising an important question on improving the reasoning ability of LLMs during training phase since most prior works have focussed at achieving this at inference time. To do so, the authors propose to sample diverse reasoning rationales from latent distribution and optimize the LLM through a variational framework using the sampled rationales. The intuition behind application of the variational framework is well motivated through the objective of self-consistency based chain-of-thought prompting. Further, the work proposes to usethe  likelihood of generating correct answer conditioned on a rationale as a proxy to explicit reward models to optimise the LLM towards generating better reasoning chains. Results demonstrate that the proposed LaTRO method helps in improving the accuracy achieved for multiple LLMs such as Mistral-7B, Llama-3.1-8B etc. on GSM8K and ARC-Challenge datasets compared to the corresponding base and the SFT versions of the LLMs.", "review_text": "The work focuses on enhancing the reasoning abilities of large language models (LLMs) during the training phase without relying on external feedback. The authors motivate the work by raising an important question on improving the reasoning ability of LLMs during training phase since most prior works have focussed at achieving this at inference time. To do so, the authors propose to sample diverse reasoning rationales from latent distribution and optimize the LLM through a variational framework using the sampled rationales. The intuition behind application of the variational framework is well motivated through the objective of self-consistency based chain-of-thought prompting. Further, the work proposes to usethe  likelihood of generating correct answer conditioned on a rationale as a proxy to explicit reward models to optimise the LLM towards generating better reasoning chains. Results demonstrate that the proposed LaTRO method helps in improving the accuracy achieved for multiple LLMs such as Mistral-7B, Llama-3.1-8B etc. on GSM8K and ARC-Challenge datasets compared to the corresponding base and the SFT versions of the LLMs.", "strengths": "1. Enhancing the reasoning ability of LLMs without relying on feedback from any external LLM during the training phase is an important research question.\n2. Suitability of applying the variational objective is well motivated, explained and justified in Sections 3 and 4.1.\n3. Accuracy gains over the base and the SFT version of the LLMs are shown with some ablation studies on greedy decoding vs. self consistency based sampling. Further, the authors show that inference time scaling of number of reasoning samples obtained using LaTRO can additionally enhance the accuracy.", "weaknesses": "1. Presentation of the introduction section of the paper needs improvement - It should better motivate the need behind reducing the reliance on external LLMs/feedback for improving a given LLM. Further, it should provide information about why the variational framework is suitable and elaborate some details about the proposed method in the introduction itself.\n2. Discussion on related work is very limited. Even though the proposed method is an effort to reduce reliance on external LLMs, it should still discuss and contrast against existing methods that leverage external LLMs (for example, [1,2]) as well as employ self-play based techniques to improve an LLM by itself with requiring help from any other LLM (for example, [3, 4]). The example references are only indicative (but not exhaustive) of the type of citations that should be included.\n3. Lack of appropriate baselines - Even though the work claims and focusses at improving the base/SFT version of the LLM through the proposed LaTRO framework, it should still compare against some of the existing self-play trainable methods (by adapting them for 7B LLMs) as baselines (eg. [3] - on datasets where the ground-truth rationale is available and [4]). Such methods improve the rationales generated by an LLM by exploring the generation space and discriminating better rationales from the inferior ones.\n4. More reasoning datasets such as CSQA, Hellaswag etc. should be considered for evaluation studies.\n\n[1] Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017, Toronto, Canada. Association for Computational Linguistics.\n\n[2] Zephyr: Direct Distillation of LM Alignment. Lewis Tunstall and Edward Emanuel Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro Von Werra and Clementine Fourrier and Nathan Habib and Nathan Sarrazin and Omar Sanseviero and Alexander M Rush and Thomas Wolf. First Conference on Language Modeling, 2024.\n\n[3] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models. Zixiang Chen and Yihe Deng and Huizhuo Yuan and Kaixuan Ji and Quanquan Gu. ICML 2024.\n\n[4] Self-Rewarding Language Models. Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, Jason Weston", "questions": "1. Line 290: It is not clear about how the first gradient term in Eq. 4 would lead to optimising the LLM policy to generate higher quality rationales. Further elaboration is needed as to why generating/optimising the likelihood of input question/instruction conditioned on the rationale would lead to better reasoning.\n2. Lines 414-415: Please support with examples about why improvements on ARC-Challenge are relatively lesser. The magnitude of improvements is not at all an issue, however, it should be better demonstrated that why better reasoning chains are not leading to higher improvements. Does this make ARC-Challenge ill-suited for this study since it involves limited reasoning scope?\n3. Line 430: Why is it needed to generate rationales that are as long as 500 tokens? It would be good to show the usefulness of the information contained in the longer rationales and why does it lead to better performance before saturating.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work focuses on enhancing the reasoning abilities of large language models (LLMs) during the training phase without relying on external feedback. The authors motivate the work by raising an important question on improving the reasoning ability of LLMs during training phase since most prior works have focussed at achieving this at inference time. To do so, the authors propose to sample diverse reasoning rationales from latent distribution and optimize the LLM through a variational framework using the sampled rationales. The intuition behind application of the variational framework is well motivated through the objective of self-consistency based chain-of-thought prompting. Further, the work proposes to usethe  likelihood of generating correct answer conditioned on a rationale as a proxy to explicit reward models to optimise the LLM towards generating better reasoning chains. Results demonstrate that the proposed LaTRO method helps in improving the accuracy achieved for multiple LLMs such as Mistral-7B, Llama-3.1-8B etc. on GSM8K and ARC-Challenge datasets compared to the corresponding base and the SFT versions of the LLMs.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. Enhancing the reasoning ability of LLMs without relying on feedback from any external LLM during the training phase is an important research question.\n2. Suitability of applying the variational objective is well motivated, explained and justified in Sections 3 and 4.1.\n3. Accuracy gains over the base and the SFT version of the LLMs are shown with some ablation studies on greedy decoding vs. self consistency based sampling. Further, the authors show that inference time scaling of number of reasoning samples obtained using LaTRO can additionally enhance the accuracy.", "weaknesses": "1. Presentation of the introduction section of the paper needs improvement - It should better motivate the need behind reducing the reliance on external LLMs/feedback for improving a given LLM. Further, it should provide information about why the variational framework is suitable and elaborate some details about the proposed method in the introduction itself.\n2. Discussion on related work is very limited. Even though the proposed method is an effort to reduce reliance on external LLMs, it should still discuss and contrast against existing methods that leverage external LLMs (for example, [1,2]) as well as employ self-play based techniques to improve an LLM by itself with requiring help from any other LLM (for example, [3, 4]). The example references are only indicative (but not exhaustive) of the type of citations that should be included.\n3. Lack of appropriate baselines - Even though the work claims and focusses at improving the base/SFT version of the LLM through the proposed LaTRO framework, it should still compare against some of the existing self-play trainable methods (by adapting them for 7B LLMs) as baselines (eg. [3] - on datasets where the ground-truth rationale is available and [4]). Such methods improve the rationales generated by an LLM by exploring the generation space and discriminating better rationales from the inferior ones.\n4. More reasoning datasets such as CSQA, Hellaswag etc. should be considered for evaluation studies.\n\n[1] Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017, Toronto, Canada. Association for Computational Linguistics.\n\n[2] Zephyr: Direct Distillation of LM Alignment. Lewis Tunstall and Edward Emanuel Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro Von Werra and Clementine Fourrier and Nathan Habib and Nathan Sarrazin and Omar Sanseviero and Alexander M Rush and Thomas Wolf. First Conference on Language Modeling, 2024.\n\n[3] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models. Zixiang Chen and Yihe Deng and Huizhuo Yuan and Kaixuan Ji and Quanquan Gu. ICML 2024.\n\n[4] Self-Rewarding Language Models. Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, Jason Weston", "questions": "1. Line 290: It is not clear about how the first gradient term in Eq. 4 would lead to optimising the LLM policy to generate higher quality rationales. Further elaboration is needed as to why generating/optimising the likelihood of input question/instruction conditioned on the rationale would lead to better reasoning.\n2. Lines 414-415: Please support with examples about why improvements on ARC-Challenge are relatively lesser. The magnitude of improvements is not at all an issue, however, it should be better demonstrated that why better reasoning chains are not leading to higher improvements. Does this make ARC-Challenge ill-suited for this study since it involves limited reasoning scope?\n3. Line 430: Why is it needed to generate rationales that are as long as 500 tokens? It would be good to show the usefulness of the information contained in the longer rationales and why does it lead to better performance before saturating.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730740051800}, {"id": "8pMPi3D3Ns", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9035/Reviewer_n3ec"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper proposesLaTRO, a novel framework aimed at enhancing the reasoning capabilities of LLMs. LaTRO addresses the challenge of improving reasoning during the training phase by formulating reasoning as sampling from a latent distribution and optimizing it through variational approaches. This method allows LLMs to self-improve their reasoning process and ability to evaluate the quality of reasoning without external feedback or reward models. The paper validates LaTRO's effectiveness through experiments on GSM8K and ARC-Challenge datasets, demonstrating significant improvements over base models and supervised fine-tuning approaches.", "review_text": "This paper proposesLaTRO, a novel framework aimed at enhancing the reasoning capabilities of LLMs. LaTRO addresses the challenge of improving reasoning during the training phase by formulating reasoning as sampling from a latent distribution and optimizing it through variational approaches. This method allows LLMs to self-improve their reasoning process and ability to evaluate the quality of reasoning without external feedback or reward models. The paper validates LaTRO's effectiveness through experiments on GSM8K and ARC-Challenge datasets, demonstrating significant improvements over base models and supervised fine-tuning approaches.", "strengths": "- The writing is clear and easy to follow.\n- The discussed topic and motivation are both innovative and significant.", "weaknesses": "- Although I'm not familiar with the topic discussed in this article, I believe the experiments presented are too few and not comprehensive enough. Moreover, the datasets considered are limited to only GSM8K and ARC-Challenge, which lacks persuasiveness.\n- The number of case study examples is too limited, with only a few instances in Figure 4 and the appendix, which is not convincing.\n- The proposed method, LaTRO, especially the stability of the Self-reward component, has not been adequately considered.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposesLaTRO, a novel framework aimed at enhancing the reasoning capabilities of LLMs. LaTRO addresses the challenge of improving reasoning during the training phase by formulating reasoning as sampling from a latent distribution and optimizing it through variational approaches. This method allows LLMs to self-improve their reasoning process and ability to evaluate the quality of reasoning without external feedback or reward models. The paper validates LaTRO's effectiveness through experiments on GSM8K and ARC-Challenge datasets, demonstrating significant improvements over base models and supervised fine-tuning approaches.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The writing is clear and easy to follow.\n- The discussed topic and motivation are both innovative and significant.", "weaknesses": "- Although I'm not familiar with the topic discussed in this article, I believe the experiments presented are too few and not comprehensive enough. Moreover, the datasets considered are limited to only GSM8K and ARC-Challenge, which lacks persuasiveness.\n- The number of case study examples is too limited, with only a few instances in Figure 4 and the appendix, which is not convincing.\n- The proposed method, LaTRO, especially the stability of the Self-reward component, has not been adequately considered.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730430503530}, {"id": "zMl0wOOr27", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9035/Reviewer_dFSv"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces LaTRO, a novel approach that formulates Chain-of-Thought (CoT) reasoning as sampling from a latent distribution, optimized through variational techniques. By leveraging the probability of generating the correct answer as an implicit reward, LaTRO unifies the learning of both the policy and reward models, allowing large language models to refine reasoning paths in a self-rewarding manner. The authors demonstrate LaTRO’s effectiveness through experiments on the GSM8K and ARC-Challenge datasets across various model architectures. Their results indicate that latent reasoning capabilities within pre-trained language models can be unlocked and enhanced using this self-improvement framework.", "review_text": "This paper introduces LaTRO, a novel approach that formulates Chain-of-Thought (CoT) reasoning as sampling from a latent distribution, optimized through variational techniques. By leveraging the probability of generating the correct answer as an implicit reward, LaTRO unifies the learning of both the policy and reward models, allowing large language models to refine reasoning paths in a self-rewarding manner. The authors demonstrate LaTRO’s effectiveness through experiments on the GSM8K and ARC-Challenge datasets across various model architectures. Their results indicate that latent reasoning capabilities within pre-trained language models can be unlocked and enhanced using this self-improvement framework.", "strengths": "1. The paper offers a novel perspective by framing reasoning as a process of sampling from a latent distribution and addressing it through variational methods.\n\n2. The paper leverages the model's own probability estimates as an implicit reward, unifying the training of the policy and reward models.\n\n3. The paper is well-organized and easy to follow.", "weaknesses": "1. The experimental setup lacks sufficient strong baselines, which are essential for a robust evaluation. Two key baselines to consider are:\n  - A stronger SFT Baseline: Fine-tuning the policy model with correct reasoning paths. Given the availability of ground truth answers, multiple reasoning paths could be sampled, retaining only those that align with the ground truth. This baseline would provide a more rigorous comparison for evaluating LaTRO’s effectiveness in reasoning.\n  - DPO Baseline: The authors could further fine-tune the policy model using the DPO algorithm, incorporating both correct and incorrect reasoning paths. Actually, the DPO algorithm aligns closely with LaTRO in its approach, as both methods aim to avoid training an explicit reward model. Including DPO as a baseline would highlight LaTRO’s strengths relative to an approach that similarly leverages implicit reward mechanisms.\n\n2. The experimental scope is limited, as only two datasets and small models were tested. \n - Expanding the experiments to include a wider range of reasoning datasets would better assess the model's reasoning capabilities. Standard practice for evaluating reasoning in large language models includes diverse datasets that cover arithmetic reasoning (only GSM8K is not enough for arithmetic reasoning evaluation), commonsense reasoning, symbolic reasoning, and other reasoning types. Incorporating these would provide a more comprehensive evaluation.\n - Testing across varying model scales, especially with larger models, could provide insights into how the approach scales with model size and whether larger models yield better reasoning performance.\n\n3. Although the authors claim that training did not rely on external feedback, the ground truth answers effectively serve as a form of implicit external feedback or reward.", "questions": "1. In Proposition 1, $p(y|x) = \\int p(y|z,x) p(z|x) dz$ holds for any CoT-based method. Why, then, is CoT-SC introduced here?\n\n2. What is the definition of \"golden rationales\", and why can’t ARC-Challenge have golden rationales?\n\n3. What can the experimental results on ARC-Challenge demonstrate? The two baselines are too weak, as the SFT baseline did not utilize rationales.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LaTRO, a novel approach that formulates Chain-of-Thought (CoT) reasoning as sampling from a latent distribution, optimized through variational techniques. By leveraging the probability of generating the correct answer as an implicit reward, LaTRO unifies the learning of both the policy and reward models, allowing large language models to refine reasoning paths in a self-rewarding manner. The authors demonstrate LaTRO’s effectiveness through experiments on the GSM8K and ARC-Challenge datasets across various model architectures. Their results indicate that latent reasoning capabilities within pre-trained language models can be unlocked and enhanced using this self-improvement framework.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper offers a novel perspective by framing reasoning as a process of sampling from a latent distribution and addressing it through variational methods.\n\n2. The paper leverages the model's own probability estimates as an implicit reward, unifying the training of the policy and reward models.\n\n3. The paper is well-organized and easy to follow.", "weaknesses": "1. The experimental setup lacks sufficient strong baselines, which are essential for a robust evaluation. Two key baselines to consider are:\n  - A stronger SFT Baseline: Fine-tuning the policy model with correct reasoning paths. Given the availability of ground truth answers, multiple reasoning paths could be sampled, retaining only those that align with the ground truth. This baseline would provide a more rigorous comparison for evaluating LaTRO’s effectiveness in reasoning.\n  - DPO Baseline: The authors could further fine-tune the policy model using the DPO algorithm, incorporating both correct and incorrect reasoning paths. Actually, the DPO algorithm aligns closely with LaTRO in its approach, as both methods aim to avoid training an explicit reward model. Including DPO as a baseline would highlight LaTRO’s strengths relative to an approach that similarly leverages implicit reward mechanisms.\n\n2. The experimental scope is limited, as only two datasets and small models were tested. \n - Expanding the experiments to include a wider range of reasoning datasets would better assess the model's reasoning capabilities. Standard practice for evaluating reasoning in large language models includes diverse datasets that cover arithmetic reasoning (only GSM8K is not enough for arithmetic reasoning evaluation), commonsense reasoning, symbolic reasoning, and other reasoning types. Incorporating these would provide a more comprehensive evaluation.\n - Testing across varying model scales, especially with larger models, could provide insights into how the approach scales with model size and whether larger models yield better reasoning performance.\n\n3. Although the authors claim that training did not rely on external feedback, the ground truth answers effectively serve as a form of implicit external feedback or reward.", "questions": "1. In Proposition 1, $p(y|x) = \\int p(y|z,x) p(z|x) dz$ holds for any CoT-based method. Why, then, is CoT-SC introduced here?\n\n2. What is the definition of \"golden rationales\", and why can’t ARC-Challenge have golden rationales?\n\n3. What can the experimental results on ARC-Challenge demonstrate? The two baselines are too weak, as the SFT baseline did not utilize rationales.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730217656700}, {"id": "uO3weKqead", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9035/Reviewer_XUrD"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces LaTRO, a framework that enhances LLM’s reasoning abilities by treating reasoning as sampling from a latent distribution and optimizing it with variational methods. LaTRO allows LLMs to improve their reasoning process and evaluation of reasoning quality simultaneously, without external feedback. Experiments on GSM8K and ARC-Challenge datasets demonstrate the effectiveness of LaTRO compared with the SFT training method.", "review_text": "This paper introduces LaTRO, a framework that enhances LLM’s reasoning abilities by treating reasoning as sampling from a latent distribution and optimizing it with variational methods. LaTRO allows LLMs to improve their reasoning process and evaluation of reasoning quality simultaneously, without external feedback. Experiments on GSM8K and ARC-Challenge datasets demonstrate the effectiveness of LaTRO compared with the SFT training method.", "strengths": "1. LaTRO regards the reasoning process as sampling from a latent distribution and optimizes it using a variational method. This approach is different from prompt-based methods such as CoT and is closer to unsupervised learning. Besides, the feasibility of LaTRO is verified by mathematical proof.\n\n2. This paper focuses on a very interesting topic, which enables an LLM to improve itself through a self-reward mechanism without external supervision and feedback signals.", "weaknesses": "1. Request a step-by-step description or flowchart of the LaTRO pipeline. Although a large number of formulas are used in this paper to explain each part，but it lacks a detailed description of the proposed method, which makes it difficult for readers to understand the complete pipeline of the proposed method and reproduce it. \n\n2. As far as I know, there are some works to gradually enhance the capabilities of LLM (not limited to reasoning task), including prompt-based [1][2][3] and training-based methods [4][5][6], some of which do not use any feedback information to enhance the capabilities of LLM [7][8]. The author should discuss the differences between this work and these works and compare the performance of these works. It is necessary to add a dedicated subsection in the Related Work section discussing these specific works and their methods. If possible, include performance comparisons with these methods in the experimental section.\n\n[1] When can llms actually correct their own mistakes? A critical survey of self-correction of llms\n\n[2] Learning From Mistakes Makes LLM Better Reasoner\n\n[3] Mirror: A Multiple-perspective Self-Reflection Method for Knowledge-rich Reasoning\n\n[4] REFINER: Reasoning Feedback on Intermediate Representations\n\n[5] CRYSTAL: Introspective Reasoners Reinforced with Self-Feedback\n\n[6] SELF: Language-driven Self-evolution for Large Language model\n\n[7] Small language modes can self-correct\n\n[8] Think Thrice Before You Act: Progressive Thought Refinement in Large Language Models\n \n3. In the experiments, the author only employs SFT training method and the base model as baselines, without comparing the performance with COT-SC mentioned in Section 3 and the self-improvement classic work mentioned above, making it difficult to demonstrate the advantages of the proposed LaTRO. So, please provide a more comprehensive analysis of how LaTRO performs relative to these additional baselines across different tasks and metrics.", "questions": "1. About self-evaluation task in LaTRO: how does LaTRO evaluate the probability of each reasoning path producing the correct answer? What does the conditional probability mention in Section 4.1 mean? Is there a task restriction for this evaluation method? What is the accuracy of self-evaluation? The authors did not discuss these questions in the paper, nor did they explore them in depth in the experiments.\n\n2.  The authors only used a formula to explain the use of self-reward signals to achieve parameter updates, so what exactly does this update parameter refer to during the training phase? How is it implemented?  Suggest provide more detailed information.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LaTRO, a framework that enhances LLM’s reasoning abilities by treating reasoning as sampling from a latent distribution and optimizing it with variational methods. LaTRO allows LLMs to improve their reasoning process and evaluation of reasoning quality simultaneously, without external feedback. Experiments on GSM8K and ARC-Challenge datasets demonstrate the effectiveness of LaTRO compared with the SFT training method.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. LaTRO regards the reasoning process as sampling from a latent distribution and optimizes it using a variational method. This approach is different from prompt-based methods such as CoT and is closer to unsupervised learning. Besides, the feasibility of LaTRO is verified by mathematical proof.\n\n2. This paper focuses on a very interesting topic, which enables an LLM to improve itself through a self-reward mechanism without external supervision and feedback signals.", "weaknesses": "1. Request a step-by-step description or flowchart of the LaTRO pipeline. Although a large number of formulas are used in this paper to explain each part，but it lacks a detailed description of the proposed method, which makes it difficult for readers to understand the complete pipeline of the proposed method and reproduce it. \n\n2. As far as I know, there are some works to gradually enhance the capabilities of LLM (not limited to reasoning task), including prompt-based [1][2][3] and training-based methods [4][5][6], some of which do not use any feedback information to enhance the capabilities of LLM [7][8]. The author should discuss the differences between this work and these works and compare the performance of these works. It is necessary to add a dedicated subsection in the Related Work section discussing these specific works and their methods. If possible, include performance comparisons with these methods in the experimental section.\n\n[1] When can llms actually correct their own mistakes? A critical survey of self-correction of llms\n\n[2] Learning From Mistakes Makes LLM Better Reasoner\n\n[3] Mirror: A Multiple-perspective Self-Reflection Method for Knowledge-rich Reasoning\n\n[4] REFINER: Reasoning Feedback on Intermediate Representations\n\n[5] CRYSTAL: Introspective Reasoners Reinforced with Self-Feedback\n\n[6] SELF: Language-driven Self-evolution for Large Language model\n\n[7] Small language modes can self-correct\n\n[8] Think Thrice Before You Act: Progressive Thought Refinement in Large Language Models\n \n3. In the experiments, the author only employs SFT training method and the base model as baselines, without comparing the performance with COT-SC mentioned in Section 3 and the self-improvement classic work mentioned above, making it difficult to demonstrate the advantages of the proposed LaTRO. So, please provide a more comprehensive analysis of how LaTRO performs relative to these additional baselines across different tasks and metrics.", "questions": "1. About self-evaluation task in LaTRO: how does LaTRO evaluate the probability of each reasoning path producing the correct answer? What does the conditional probability mention in Section 4.1 mean? Is there a task restriction for this evaluation method? What is the accuracy of self-evaluation? The authors did not discuss these questions in the paper, nor did they explore them in depth in the experiments.\n\n2.  The authors only used a formula to explain the use of self-reward signals to achieve parameter updates, so what exactly does this update parameter refer to during the training phase? How is it implemented?  Suggest provide more detailed information.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730119256094}, {"id": "KNuduTBvGX", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9035/Reviewer_WHVo"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper proposes a principled framework, LaTRO, to treat the reasoning process as sampling from latent distribution and enable LLMs themselves as reward models to evaluate the quality of reasoning rationales. The proposed LaTRO outperforms supervised fine-tuning baselines on 2 datasets.", "review_text": "This paper proposes a principled framework, LaTRO, to treat the reasoning process as sampling from latent distribution and enable LLMs themselves as reward models to evaluate the quality of reasoning rationales. The proposed LaTRO outperforms supervised fine-tuning baselines on 2 datasets.", "strengths": "1. The structure of this paper is clear, and it's easy to follow the main idea and the contribution.\n2. The research problem this paper targets is very important to the community.\n3. The motivation is sound and the benefits of the approach are clear.", "weaknesses": "1. Although modeling the rationales as latent variables to sample is well defined and proposed, the paper lacks a discussion with previously proposed reasoning methods that formulate the reasoning process as latent variables [1, 2] as well. Even though they are cited in the related work. Specifically, [1] proposes to sample diverse reasoning rationales to improve the prediction performance and also proposes an EM-like algorithm to improve the reward model and LLM alternatively. It would be great to have a discussion and comparison with these methods.\n2. The paper does not compare with various prompting-based reasoning baselines mentioned in the related work section, such as tree-of-thought[3], and RAP[4], as well as fine-tuning baselines such as STaR[5], which is a missed opportunity to demonstrate its effectiveness. It would be better to compare them with metrics like training / inference computational cost and accuracy.\n3. The paper does not provide a confidence interval, leading to unawareness of how the proposed LaTRO is robust to the initialization and randomness. It would be great to report the results from at least 3 times repetitions.\n4. As the proposed LaTRO is a general approach, it would be great to evaluate it on more benchmark datasets to verify its effectiveness, such as HumanEval[6] and MBPP [7], which are popular in the current LLM reasoning community.\n\n\n[1] Hu, Edward J., et al. \"Amortizing intractable inference in large language models.\" arXiv preprint arXiv:2310.04363 (2023).\n\n[2] Hoffman, Matthew Douglas, et al. \"Training chain-of-thought via latent-variable inference.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[3] Yao, Shunyu, et al. \"Tree of thoughts: Deliberate problem solving with large language models.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[4] Hao, Shibo, et al. \"Reasoning with language model is planning with world model.\" arXiv preprint arXiv:2305.14992 (2023).\n\n[5] Zelikman, Eric, et al. \"Star: Bootstrapping reasoning with reasoning.\" Advances in Neural Information Processing Systems 35 (2022): 15476-15488.\n\n[6] Chen, Mark, et al. \"Evaluating large language models trained on code.\" arXiv preprint arXiv:2107.03374 (2021).\n\n[7] Austin, Jacob, et al. \"Program synthesis with large language models.\" arXiv preprint arXiv:2108.07732 (2021).", "questions": "See \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a principled framework, LaTRO, to treat the reasoning process as sampling from latent distribution and enable LLMs themselves as reward models to evaluate the quality of reasoning rationales. The proposed LaTRO outperforms supervised fine-tuning baselines on 2 datasets.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The structure of this paper is clear, and it's easy to follow the main idea and the contribution.\n2. The research problem this paper targets is very important to the community.\n3. The motivation is sound and the benefits of the approach are clear.", "weaknesses": "1. Although modeling the rationales as latent variables to sample is well defined and proposed, the paper lacks a discussion with previously proposed reasoning methods that formulate the reasoning process as latent variables [1, 2] as well. Even though they are cited in the related work. Specifically, [1] proposes to sample diverse reasoning rationales to improve the prediction performance and also proposes an EM-like algorithm to improve the reward model and LLM alternatively. It would be great to have a discussion and comparison with these methods.\n2. The paper does not compare with various prompting-based reasoning baselines mentioned in the related work section, such as tree-of-thought[3], and RAP[4], as well as fine-tuning baselines such as STaR[5], which is a missed opportunity to demonstrate its effectiveness. It would be better to compare them with metrics like training / inference computational cost and accuracy.\n3. The paper does not provide a confidence interval, leading to unawareness of how the proposed LaTRO is robust to the initialization and randomness. It would be great to report the results from at least 3 times repetitions.\n4. As the proposed LaTRO is a general approach, it would be great to evaluate it on more benchmark datasets to verify its effectiveness, such as HumanEval[6] and MBPP [7], which are popular in the current LLM reasoning community.\n\n\n[1] Hu, Edward J., et al. \"Amortizing intractable inference in large language models.\" arXiv preprint arXiv:2310.04363 (2023).\n\n[2] Hoffman, Matthew Douglas, et al. \"Training chain-of-thought via latent-variable inference.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[3] Yao, Shunyu, et al. \"Tree of thoughts: Deliberate problem solving with large language models.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[4] Hao, Shibo, et al. \"Reasoning with language model is planning with world model.\" arXiv preprint arXiv:2305.14992 (2023).\n\n[5] Zelikman, Eric, et al. \"Star: Bootstrapping reasoning with reasoning.\" Advances in Neural Information Processing Systems 35 (2022): 15476-15488.\n\n[6] Chen, Mark, et al. \"Evaluating large language models trained on code.\" arXiv preprint arXiv:2107.03374 (2021).\n\n[7] Austin, Jacob, et al. \"Program synthesis with large language models.\" arXiv preprint arXiv:2108.07732 (2021).", "questions": "See \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729870612154}], "openreview_url": "https://openreview.net/forum?id=4Po8d9GAfQ", "arxiv_id": "2411.04282", "paper_pdf": "papers/4Po8d9GAfQ.pdf", "paper_pdf_sha256": "6e4e45ee93db74ef1508625f12bb06e1064518bc109de3daf9124b4ccf17c059", "paper_pdf_bytes": 436976, "paper_pdf_source": "openreview", "code_url": "https://github.com/SalesforceAIResearch/LaTRO", "code_repository": "SalesforceAIResearch/LaTRO", "code_commit": "54f9356974cc138cd39ef79ec93eb9a421d5976f", "code_archive": "repos/4Po8d9GAfQ.zip", "code_archive_sha256": "a12ae43dd908bd38d7a25e132d2738013646df2fa1519032f53249f15c2a6f66", "code_archive_bytes": 44741, "code_file_count": 16, "code_extensions": {".py": 13, ".sh": 3}, "github_disk_usage_kb": 82, "github_languages": {"Python": 81899, "Shell": 3499}, "github_archived": false, "github_pushed_at": "2026-06-02T18:54:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/language-models-are-hidden-reasoners"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pK7V0glCdj", "year": 2024, "status": "rejected", "title": "BOtied: Multi-objective Bayesian optimization with tied multivariate ranks", "authors": ["Ji Won Park", "Natasa Tagasovska", "Michael Maser", "Stephen Ra", "Kyunghyun Cho"], "authorids": ["~Ji_Won_Park1", "~Natasa_Tagasovska2", "~Michael_Maser1", "~Stephen_Ra1", "~Kyunghyun_Cho1"], "authors_source": "OpenReview API", "abstract": "Many scientific and industrial applications require the joint optimization of multiple, potentially competing objectives. Multi-objective Bayesian optimization (MOBO) is a sample-efficient framework for identifying Pareto-optimal solutions. At the heart of MOBO is the acquisition function, which determines the next candidate to evaluate by navigating the best compromises among the objectives. Multi-objective acquisition functions that rely on box decomposition of the objective space, such as the expected hypervolume improvement (EHVI) and entropy search, scale poorly to a large number of objectives. We begin by showing a natural connection between non-dominated solutions and the highest multivariate rank, which coincides with the outermost level line of the joint cumulative distribution function (CDF). Motivated by this link, we propose the CDF indicator, a Pareto-compliant metric for evaluating the quality of approximate Pareto sets that complements the popular hypervolume indicator. We then propose an acquisition function based on the CDF indicator, called BOtied. BOtied can be implemented efficiently with copulas, a statistical tool for modeling complex, high-dimensional distributions. We benchmark BOtied against common acquisition functions, including EHVI, entropy search, and random scalarization, in a series of synthetic and real-data experiments. BOtied performs on par with the baselines across datasets and metrics while being computationally efficient.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "aKjkVJXxpD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6099/Reviewer_8cpR"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper solves a multi-objective Bayesian optimization problem, which attempts to find a Pareto frontier on a metric space.  Since identifying a Pareto frontier and calculating a hypervolume are time-consuming, this research proposes a novel method using the highest multivariate rank, which is the outermost level line of the joint cumulative distribution function.  Finally the authors show some experimental results on several benchmarks and wall-clock time.", "review_text": "This paper solves a multi-objective Bayesian optimization problem, which attempts to find a Pareto frontier on a metric space.  Since identifying a Pareto frontier and calculating a hypervolume are time-consuming, this research proposes a novel method using the highest multivariate rank, which is the outermost level line of the joint cumulative distribution function.  Finally the authors show some experimental results on several benchmarks and wall-clock time.", "strengths": "* Multi-objective optimization is an interesting topic in Bayesian optimization.\n* The proposed method based on multivariate ranking seems interesting.", "weaknesses": "* The authors should improve writing and presentation more.  For example, figures are too small.  Also, there are some typos or grammar issues.  For example, in Theorem 2, there exist should be there exists, and in Section 3.4, Aas (2016) propose should be Aas (2016) proposes.\n* Experimental results are not promising.\n* According to description on experiments, the authors repeated the experiments 5 times, but variances are not reported.\n* Five repetitions are not enough to validate the proposed algorithm.\n* Based on the motivation, the authors argued that the use of multivariate rank and its distribution can accelerate the process of multi-objective Bayesian optimization.  However, Figure 5 does not seem to support this motivation.  The proposed methods should be faster than the other algorithms, but some results are comparable to ones of some algorithms.", "questions": "* In Section 2.1, I think the sentence \"Often the integral is approximated by Monte Carlo ...\" is not correct.  In Bayesian optimization, we often use the statistics of the posterior predictive distribution calculated directly, instead of the samples of the distribution.\n* I agree that the use of ranking can be better than absolute metric values.  However, the consideration of absolute metric values is sometimes important.  What do you think of this issue?\n* BOtied v1 is always worse than Botied v2 if I understand correctly.  Why do you add BOtied v1?  Is it necessary to have it?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper solves a multi-objective Bayesian optimization problem, which attempts to find a Pareto frontier on a metric space.  Since identifying a Pareto frontier and calculating a hypervolume are time-consuming, this research proposes a novel method using the highest multivariate rank, which is the outermost level line of the joint cumulative distribution function.  Finally the authors show some experimental results on several benchmarks and wall-clock time.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* Multi-objective optimization is an interesting topic in Bayesian optimization.\n* The proposed method based on multivariate ranking seems interesting.", "weaknesses": "* The authors should improve writing and presentation more.  For example, figures are too small.  Also, there are some typos or grammar issues.  For example, in Theorem 2, there exist should be there exists, and in Section 3.4, Aas (2016) propose should be Aas (2016) proposes.\n* Experimental results are not promising.\n* According to description on experiments, the authors repeated the experiments 5 times, but variances are not reported.\n* Five repetitions are not enough to validate the proposed algorithm.\n* Based on the motivation, the authors argued that the use of multivariate rank and its distribution can accelerate the process of multi-objective Bayesian optimization.  However, Figure 5 does not seem to support this motivation.  The proposed methods should be faster than the other algorithms, but some results are comparable to ones of some algorithms.", "questions": "* In Section 2.1, I think the sentence \"Often the integral is approximated by Monte Carlo ...\" is not correct.  In Bayesian optimization, we often use the statistics of the posterior predictive distribution calculated directly, instead of the samples of the distribution.\n* I agree that the use of ranking can be better than absolute metric values.  However, the consideration of absolute metric values is sometimes important.  What do you think of this issue?\n* BOtied v1 is always worse than Botied v2 if I understand correctly.  Why do you add BOtied v1?  Is it necessary to have it?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699566685410}, {"id": "BWVaetu5zB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6099/Reviewer_vKP1"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In multi-objective Bayesian optimization, existing acquisition functions scale poorly to a large number of objectives. To address this, the paper introduces the CDF indicator as a Pareto-compliant performance criterion to measure the quality of Pareto sets and proposes an acquisition function called BOTIED, which can be implemented efficiently using copulas.", "review_text": "In multi-objective Bayesian optimization, existing acquisition functions scale poorly to a large number of objectives. To address this, the paper introduces the CDF indicator as a Pareto-compliant performance criterion to measure the quality of Pareto sets and proposes an acquisition function called BOTIED, which can be implemented efficiently using copulas.", "strengths": "1. This paper focuses on improving the computational efficiency of the acquisition function, which is an important issue in multi-objective optimization. \n2.The idea of incorporating domain knowledge and utilizing dependency structures of objectives in optimization seems new.", "weaknesses": "1. Experimental results do not fully demonstrate the advantage of BOTIED. As for optimization results, in table I, in DTLZ(M=6) and DTLZ(M=8), BOTIED underperforms NParEGO on HV. BOTIED even achieves lower HV than Random in DTLZ(M=8). The authors claim that BOTIED is computational efficient. However, according to Figure 5, ParEGO requires similar computational time to BOTIED.\n2. How to estimate high-dimensional CDFs with copulas is critical in BOTIED. However, there is a lack of description on this issue in the paper.", "questions": "1. Why is the HV and CDF data inconsistent in Figure 4 and Table 1 in BC(M=2) and DTLZ(M=4)?\n2. The authors mention that the CDF indicator does not discriminate sets with the same best element. This seems questionable because the purpose of multi-objective optimization is to identify a solution set instead of a single best element.\n3. The authors argue that ‘the CDF indicator is invariant to arbitrary monotonic transformations of the objectives, whereas the HV indicator is highly sensitive to them.’ However, I believe it is the HV indicator for a set instead of the HV for a single point that matters in multi-objective optimization. What is the problem of the HV indicator being sensitive to monotonic transformations of the objectives?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In multi-objective Bayesian optimization, existing acquisition functions scale poorly to a large number of objectives. To address this, the paper introduces the CDF indicator as a Pareto-compliant performance criterion to measure the quality of Pareto sets and proposes an acquisition function called BOTIED, which can be implemented efficiently using copulas.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. This paper focuses on improving the computational efficiency of the acquisition function, which is an important issue in multi-objective optimization. \n2.The idea of incorporating domain knowledge and utilizing dependency structures of objectives in optimization seems new.", "weaknesses": "1. Experimental results do not fully demonstrate the advantage of BOTIED. As for optimization results, in table I, in DTLZ(M=6) and DTLZ(M=8), BOTIED underperforms NParEGO on HV. BOTIED even achieves lower HV than Random in DTLZ(M=8). The authors claim that BOTIED is computational efficient. However, according to Figure 5, ParEGO requires similar computational time to BOTIED.\n2. How to estimate high-dimensional CDFs with copulas is critical in BOTIED. However, there is a lack of description on this issue in the paper.", "questions": "1. Why is the HV and CDF data inconsistent in Figure 4 and Table 1 in BC(M=2) and DTLZ(M=4)?\n2. The authors mention that the CDF indicator does not discriminate sets with the same best element. This seems questionable because the purpose of multi-objective optimization is to identify a solution set instead of a single best element.\n3. The authors argue that ‘the CDF indicator is invariant to arbitrary monotonic transformations of the objectives, whereas the HV indicator is highly sensitive to them.’ However, I believe it is the HV indicator for a set instead of the HV for a single point that matters in multi-objective optimization. What is the problem of the HV indicator being sensitive to monotonic transformations of the objectives?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699193699568}, {"id": "AGM1jZA0Me", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6099/Reviewer_x952"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a CDF based acquisition function for black-box multi-objective optimization (MOO). The basic idea is to use the CDF function as a criterion for the next point selection and the authors discuss a connection between CDF and multivariate rank. To evaluate the CDF, a copula based approach is introduced by which the authors claim scalability and invariance properties are obtained.", "review_text": "The paper proposes a CDF based acquisition function for black-box multi-objective optimization (MOO). The basic idea is to use the CDF function as a criterion for the next point selection and the authors discuss a connection between CDF and multivariate rank. To evaluate the CDF, a copula based approach is introduced by which the authors claim scalability and invariance properties are obtained.", "strengths": "Experimental results seemingly show a good result.\n\nIntroducing copula-based computations into multi-objective BO is seemingly novel.", "weaknesses": "Overall, the description is unclear. For example, although the CDF plays a key role, the distribution of CDF is not explicitly written in the methodology section. In appendix, I found the authors used multi-task GP. Another example is \\hat{f} in (3.8) suddenly appear without explanation (explained in the experimental section), by which the definition of the acquisition function becomes unclear at that point. \n\nAs mentioned above, the authors used multi-task GP, but for GP based MOO (i.e., BO), except for a few studies (e.g., Shah and Ghahraman 2016), typically uses independent GPs for multiple objectives. When the independent GPs are used, CDF is quite easy to evaluate (just by multiplying one dimensional CDFs). However, the author does not mention importance of modeling correlation among objectives. In practice, independent GPs often show sufficient (or better) performance compared with multi-task GP (for which task-correlation must be carefully tuned to achieve good performance). \n\nRationale behind the CDF based criterion is unclear. For simplicity, consider the case of independent Gaussian for each objective, which would be the simplest special case. Then, CDF becomes just a multiplication of each dimension of CDF, which intensely seeks a `specific direction' of the output space though, in MOO, the Pareto frontier should exist in a variety of direction of the output space. In this sense, in my current understanding, the proposed acquisition function is not appropriate for exploring the entire Pareto frontier that can be widely distributed in the output space. Even when correlated model is used, this problem would not be avoided.", "questions": "In the experiment section, the authors mention the predictive mean is used for CDF calculation. Does that mean \\hat{f} is set as the posterior mean of GPs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a CDF based acquisition function for black-box multi-objective optimization (MOO). The basic idea is to use the CDF function as a criterion for the next point selection and the authors discuss a connection between CDF and multivariate rank. To evaluate the CDF, a copula based approach is introduced by which the authors claim scalability and invariance properties are obtained.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "Experimental results seemingly show a good result.\n\nIntroducing copula-based computations into multi-objective BO is seemingly novel.", "weaknesses": "Overall, the description is unclear. For example, although the CDF plays a key role, the distribution of CDF is not explicitly written in the methodology section. In appendix, I found the authors used multi-task GP. Another example is \\hat{f} in (3.8) suddenly appear without explanation (explained in the experimental section), by which the definition of the acquisition function becomes unclear at that point. \n\nAs mentioned above, the authors used multi-task GP, but for GP based MOO (i.e., BO), except for a few studies (e.g., Shah and Ghahraman 2016), typically uses independent GPs for multiple objectives. When the independent GPs are used, CDF is quite easy to evaluate (just by multiplying one dimensional CDFs). However, the author does not mention importance of modeling correlation among objectives. In practice, independent GPs often show sufficient (or better) performance compared with multi-task GP (for which task-correlation must be carefully tuned to achieve good performance). \n\nRationale behind the CDF based criterion is unclear. For simplicity, consider the case of independent Gaussian for each objective, which would be the simplest special case. Then, CDF becomes just a multiplication of each dimension of CDF, which intensely seeks a `specific direction' of the output space though, in MOO, the Pareto frontier should exist in a variety of direction of the output space. In this sense, in my current understanding, the proposed acquisition function is not appropriate for exploring the entire Pareto frontier that can be widely distributed in the output space. Even when correlated model is used, this problem would not be avoided.", "questions": "In the experiment section, the authors mention the predictive mean is used for CDF calculation. Does that mean \\hat{f} is set as the posterior mean of GPs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699005695125}, {"id": "66HGP4fVKH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6099/Reviewer_3FiX"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors consider concurrent optimization, with multiple objectives, within a Bayesian optimization framework. Their primary emphasis lies on introducing a novel acquisition function rooted in multivariate ranks. This innovative approach aims to alleviate the computational burden tied to the hypervolume or entropy calculations in established criteria. The process entails sampling from the posterior distribution to estimate uniform marginals, followed by rank assignment post a copula transformation. The study includes a comprehensive evaluation across various simplified scenarios, along with a selection of more complex test cases.", "review_text": "The authors consider concurrent optimization, with multiple objectives, within a Bayesian optimization framework. Their primary emphasis lies on introducing a novel acquisition function rooted in multivariate ranks. This innovative approach aims to alleviate the computational burden tied to the hypervolume or entropy calculations in established criteria. The process entails sampling from the posterior distribution to estimate uniform marginals, followed by rank assignment post a copula transformation. The study includes a comprehensive evaluation across various simplified scenarios, along with a selection of more complex test cases.", "strengths": "- Detailed description of the method.\n- Summary of relationships with the state of the art", "weaknesses": "- Mostly toy examples are provided, e.g., DTLZ test function with a simple Pareto front.\n- Only discrete inputs are considered", "questions": "Figure 3a is too small.\n\nThe method only seems to work on discrete sets, according to Algorithm 1. But some problems are continuous, like Branin-Currin or the DTLZ test problems. The adaptation is not clearly discussed.\n\nIt is preferable to show progress curves over iterations rather than fixed snapshots.\n\nTypos\nP7: being the whether the", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors consider concurrent optimization, with multiple objectives, within a Bayesian optimization framework. Their primary emphasis lies on introducing a novel acquisition function rooted in multivariate ranks. This innovative approach aims to alleviate the computational burden tied to the hypervolume or entropy calculations in established criteria. The process entails sampling from the posterior distribution to estimate uniform marginals, followed by rank assignment post a copula transformation. The study includes a comprehensive evaluation across various simplified scenarios, along with a selection of more complex test cases.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Detailed description of the method.\n- Summary of relationships with the state of the art", "weaknesses": "- Mostly toy examples are provided, e.g., DTLZ test function with a simple Pareto front.\n- Only discrete inputs are considered", "questions": "Figure 3a is too small.\n\nThe method only seems to work on discrete sets, according to Algorithm 1. But some problems are continuous, like Branin-Currin or the DTLZ test problems. The adaptation is not clearly discussed.\n\nIt is preferable to show progress curves over iterations rather than fixed snapshots.\n\nTypos\nP7: being the whether the", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698315394322}], "openreview_url": "https://openreview.net/forum?id=pK7V0glCdj", "arxiv_id": "2306.00344", "paper_pdf": "papers/pK7V0glCdj.pdf", "paper_pdf_sha256": "c11503408aacc3eb4a34c674606adf53d59f31bfa6f344f0a23098cab7efa6fc", "paper_pdf_bytes": 1949876, "paper_pdf_source": "openreview", "code_url": "https://github.com/jiwoncpark/botied", "code_repository": "jiwoncpark/botied", "code_commit": "1d8eeb17b632a3b97a3a30519a939070178b4575", "code_archive": "repos/pK7V0glCdj.zip", "code_archive_sha256": "4d670880077cdee682595e67b6ab02f8d01e5dfb6e50a138a93917eae0dc1378", "code_archive_bytes": 71689, "code_file_count": 44, "code_extensions": {".py": 38, ".sh": 6}, "github_disk_usage_kb": 61, "github_languages": {"Python": 140924, "Shell": 3938}, "github_archived": false, "github_pushed_at": "2024-06-04T05:00:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/botied-multi-objective-bayesian-optimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vDY5Y8HMNxO", "year": 2023, "status": "rejected", "title": "GMML is All you Need", "authors": ["Sara Atito", "Muhammad Awais", "Josef Kittler"], "authorids": ["~Sara_Atito1", "muhammad.awais@surrey.ac.uk", "~Josef_Kittler1"], "authors_source": "OpenReview API", "abstract": "Vision transformers have generated significant interest in the computer vision (CV) community because of their flexibility in exploiting contextual information, whether it is sharply confined local, or long range global. However, they are known to be data hungry. This has motivated the research in self-supervised transformer pretraining, which does not need to decode the semantic information conveyed by labels to link it to the image properties, but rather focuses directly on extracting a concise representation of the image data that reflects the notion of similarity and is invariant to nuisance factors. The key vehicle for the self-learning process used by the majority of self-learning methods is the generation of multiple views of the training data and the creation of pretext tasks which use these views to define the notion of image similarity and data integrity. However, this approach lacks the natural propensity to extract contextual information. We propose group mask model learning (GMML), a self-supervised learning (SSL) mechanism for pretraining vision transformers with the ability to extract the contextual information present in all the concepts in an image. GMML achieves this by manipulating random groups of connected tokens, ensuingly covering a meaningful part of a semantic concept, and then recovering the hidden semantic information from the visible part of the concept. GMML implicitly introduces a novel data augmentation process. Unlike most of the existing SSL approaches, GMML does not require momentum encoder, nor rely on careful implementation details such as large batches and gradient stopping, which are all artefacts of most of the current self-supervised learning techniques. Since its conception at the beginning of 2021, GMML maintains itself as unbeaten SSL method with several desirable benefits and marked a significant milestone in computer vision by being one of the first self-supervised pretraining methods which outperform supervised pretraining consistently with a large margin. GMML is simple, elegant, and currently the best mechanism to extract information from a given dataset and instil this information into transformer's weights. The code will be made publicly available for the community to train on bigger corpora.   ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "wj7g5snkdh", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4529/Reviewer_JV27"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a self supervised learning framework GMML for vision transformers. The idea is essentially the same as MAE and SimMIM (the authors claim that MAE and SimMIM are extensions of GMML). When masking the inputs, they use noise or alien patches from other images, rather than just zeros. In this way, the initial blocks of the networks have to model the alien and the non-alien concepts, and thus can better model the context.", "review_text": "Overall, the paper propose an interesting masking method using alien patches. However, the experiments are weak and the relationship with MAE and SimMIM is unclear.", "strengths": "Strength\n\n1. The benefit from using noise or alien patches from other images, rather than just zeros when masking inputs is clear as shown in Figure 4(a).\n2. On small datasets, GMML converge faster and perform better (Table 1)\n\nWeakness\n\nThere are several critical weaknesses in experiments. \n\n1. Firstly, in the experiments, the authors only use ViT-tiny and ViT-small backbones. While it is more common to use a larger backbone for comparison (e.g. B, L, H) as in MAE and SimMIM.\n2. Secondly, the authors missed the comparison of ImageNet-1k in Table 1, which is probably the most important comparison with other baselines.\n3. Thirdly, some important baselines are missing for comparison, e.g. SimMIM, BEIT", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a self supervised learning framework GMML for vision transformers. The idea is essentially the same as MAE and SimMIM (the authors claim that MAE and SimMIM are extensions of GMML). When masking the inputs, they use noise or alien patches from other images, rather than just zeros. In this way, the initial blocks of the networks have to model the alien and the non-alien concepts, and thus can better model the context.", "strength_and_weaknesses": "Strength\n\n1. The benefit from using noise or alien patches from other images, rather than just zeros when masking inputs is clear as shown in Figure 4(a).\n2. On small datasets, GMML converge faster and perform better (Table 1)\n\nWeakness\n\nThere are several critical weaknesses in experiments. \n\n1. Firstly, in the experiments, the authors only use ViT-tiny and ViT-small backbones. While it is more common to use a larger backbone for comparison (e.g. B, L, H) as in MAE and SimMIM.\n2. Secondly, the authors missed the comparison of ImageNet-1k in Table 1, which is probably the most important comparison with other baselines.\n3. Thirdly, some important baselines are missing for comparison, e.g. SimMIM, BEIT", "clarity,_quality,_novelty_and_reproducibility": "\nThe authors claim that GMML is proposed “at the beginning of 2021” with no citation. They also claim that SimMIM and MAE are extensions of GMML. So, I have no choice but to search for GMML and found an unpublished paper on ArXiv [1].\n\nIgnore [1], in my opinion, the novelty of this paper mainly lies in using alien patches when masking inputs.\n\n[1] SiT: Self-supervised vIsion Transformer. 2104.03602", "summary_of_the_review": "Overall, the paper propose an interesting masking method using alien patches. However, the experiments are weak and the relationship with MAE and SimMIM is unclear.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666673235975}, {"id": "anPIM4ztSq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4529/Reviewer_4sPn"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper delivers a method called group mask model learning (GMML). GMML learns a representation of an image by partially corrupting it and reconstructing it with a transformer encoder-decoder structure. The authors conducted experiments that show GMML’s superiority in learning representations from small-scale datasets with small-sized models.", "review_text": "Due to the various weaknesses described above (see C1-C5) I cannot give a high recommendation to this paper.", "strengths": "- (C1) Models those capable of dealing with small-scale datasets are always welcome. However, assessing the model’s performance on large-scale datasets (at least ImageNet-1K) is needed to determine whether the model can also be used for pre-training those datasets. In that sense, the paper has its weakness in only delivering experiments on small-scale datasets and small-sized models.\n- (C2) GMML is almost identical to SimMIM. The authors go on to argue that structures like SimMIM all stem from the GMML structures presented in 2021, but it is questionable why this paper submitted to ICLR 2023 should take the contribution of GMML presented in 2021. Since SimMIM is already presented in CVPR 2022, it is appropriate to address that this paper is conveying the variant of SimMIM, regardless of the existence of GMML architecture, that I failed to find its reference in the paper.\n- (C3) Table 1 in the paper omits several papers, including SimMIM [1], BEiT [2], and SplitMask [3]. Especially, SplitMask shares topics with this paper: “SSL method that is capable of being pre-trained on small-scale dataset” and reports small datasets’ metrics seemingly on par with this paper’s result.\n- (C4) The authors said that most non-GMML approaches suffer from trivial constant solutions, but the representation collapsing problem only occurs for positive (BYOL-like) methods. Contrastive learning does not suffer from the collapse by design and the number of works adopting contrastive learning is not neglectable.\n- (C5) The manuscript has a self-contradicting statement that they have already suggested GMML in 2021 and picked a proposal of GMML as one of the paper's contributions.\n\n[1] Xie, Zhenda, et al. \"Simmim: A simple framework for masked image modeling.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2022.\n[2] Bao, Hangbo, et al. \"BEiT: BERT Pre-Training of Image Transformers.\" *International Conference on Learning Representations*. 2021.\n[3] El-Nouby, Alaaeldin, et al. \"Are Large-scale Datasets Necessary for Self-Supervised Pre-training?.\" *arXiv preprint arXiv:2112.10740* (2021).", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper delivers a method called group mask model learning (GMML). GMML learns a representation of an image by partially corrupting it and reconstructing it with a transformer encoder-decoder structure. The authors conducted experiments that show GMML’s superiority in learning representations from small-scale datasets with small-sized models.", "strength_and_weaknesses": "- (C1) Models those capable of dealing with small-scale datasets are always welcome. However, assessing the model’s performance on large-scale datasets (at least ImageNet-1K) is needed to determine whether the model can also be used for pre-training those datasets. In that sense, the paper has its weakness in only delivering experiments on small-scale datasets and small-sized models.\n- (C2) GMML is almost identical to SimMIM. The authors go on to argue that structures like SimMIM all stem from the GMML structures presented in 2021, but it is questionable why this paper submitted to ICLR 2023 should take the contribution of GMML presented in 2021. Since SimMIM is already presented in CVPR 2022, it is appropriate to address that this paper is conveying the variant of SimMIM, regardless of the existence of GMML architecture, that I failed to find its reference in the paper.\n- (C3) Table 1 in the paper omits several papers, including SimMIM [1], BEiT [2], and SplitMask [3]. Especially, SplitMask shares topics with this paper: “SSL method that is capable of being pre-trained on small-scale dataset” and reports small datasets’ metrics seemingly on par with this paper’s result.\n- (C4) The authors said that most non-GMML approaches suffer from trivial constant solutions, but the representation collapsing problem only occurs for positive (BYOL-like) methods. Contrastive learning does not suffer from the collapse by design and the number of works adopting contrastive learning is not neglectable.\n- (C5) The manuscript has a self-contradicting statement that they have already suggested GMML in 2021 and picked a proposal of GMML as one of the paper's contributions.\n\n[1] Xie, Zhenda, et al. \"Simmim: A simple framework for masked image modeling.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2022.\n[2] Bao, Hangbo, et al. \"BEiT: BERT Pre-Training of Image Transformers.\" *International Conference on Learning Representations*. 2021.\n[3] El-Nouby, Alaaeldin, et al. \"Are Large-scale Datasets Necessary for Self-Supervised Pre-training?.\" *arXiv preprint arXiv:2112.10740* (2021).", "clarity,_quality,_novelty_and_reproducibility": "- (Clarity) I couldn't quite understand why the GMML model presented in 2021 should be a contribution to this paper. (see C2, C5)\n- (Quality) The experiment presented in this paper is only a classification experiment on small-scale datasets, and it is too insufficient to claim the title \"GMML is All you Need\". (see C1) Even in the experiments performed, there are papers omitted. (see C3)\n- (Novelty) I consider this model to be about the same as SimMIM.\n- (Reproducibility) Source code is not provided, but implementation details in the appendix provide reproducibility.", "summary_of_the_review": "Due to the various weaknesses described above (see C1-C5) I cannot give a high recommendation to this paper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666672050115}, {"id": "ZUW3wqQsWY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4529/Reviewer_rp7V"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The work proposes Group Masked Model Learning framework for self supervised learning.", "review_text": "-", "strengths": "Strength:\nThe work reports promising results for self supervised learning in low data regime. \n\n\n\nWeaknesses:\n1. The novelty of the work is confusing as the work looks quite similar to other popular approaches for self supervised learning (such as MAE). However, the authors actually claim that the already published MAE (and also SIMMIM) extended actually the proposed work. \n\n2. There are other strong frameworks for self supervised learning that are not referred to in this work, for instance, data2vec (Baevski, Alexei, et al. \"Data2vec: A general framework for self-supervised learning in speech, vision and language.\" arXiv preprint arXiv:2202.03555 (2022).) \n\n3. It is difficult to understand the contribution of the proposed approach on larger datasets. For instance, how this work results are compared with the state-of-the-art on the full ImageNet. Particularly, where the proposed approach fits in the Table 1 of the  data2vec paper.\n\n4. Some parts are unclear, for instance, why the imput image is corrupted up to 70% in the case of “zeros” and “noise”, while in the case of “replace” it is corrupted up to 35%. Also not sure why the combination of “replace” and “noise” works better.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The work proposes Group Masked Model Learning framework for self supervised learning.", "strength_and_weaknesses": "Strength:\nThe work reports promising results for self supervised learning in low data regime. \n\n\n\nWeaknesses:\n1. The novelty of the work is confusing as the work looks quite similar to other popular approaches for self supervised learning (such as MAE). However, the authors actually claim that the already published MAE (and also SIMMIM) extended actually the proposed work. \n\n2. There are other strong frameworks for self supervised learning that are not referred to in this work, for instance, data2vec (Baevski, Alexei, et al. \"Data2vec: A general framework for self-supervised learning in speech, vision and language.\" arXiv preprint arXiv:2202.03555 (2022).) \n\n3. It is difficult to understand the contribution of the proposed approach on larger datasets. For instance, how this work results are compared with the state-of-the-art on the full ImageNet. Particularly, where the proposed approach fits in the Table 1 of the  data2vec paper.\n\n4. Some parts are unclear, for instance, why the imput image is corrupted up to 70% in the case of “zeros” and “noise”, while in the case of “replace” it is corrupted up to 35%. Also not sure why the combination of “replace” and “noise” works better.\n", "clarity,_quality,_novelty_and_reproducibility": "-", "summary_of_the_review": "-", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666631666905}, {"id": "0N-Jgl4pwWf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4529/Reviewer_8Ka6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a Group Masked Model Learning (GMML), a Self-Supervised Learning (SSL) mechanism\nfor pretraining vision transformers with the ability to extract the contextual information\npresent in all the concepts in an image. This is achieved by manipulating\nrandomly groups of connected tokens, ensuingly covering a meaningful part of\na semantic concept, and then recovering the hidden semantic information from\nthe visible part of the concept. The paper suggests that GMML implicitly introduces a novel data augmentation\nprocess.\n\nExperimental results are shown for Image Classification, DA in Transfer Learning, and a few types of ablations studies.\n\nAt the end of the manuscript, the authors claim the following - \nGMML based pretraining makes the vision transformers data-efficient. Also, GMML is the first\nself-supervised pretraining work which consistently outperformed supervised pretraining for any\npretraining and finetuning dataset, regardless of their sizes.\n\nThe key impact of the proposed GMML is that it makes it possible for transformers\nto train on small and medium size datasets. It is not only data efficient, but its outstanding information\nextraction ability enables it to outperform state-of-the-art supervised and self-supervised\nmethods with large margins.\n", "review_text": "\nI am not convinced with the paper.\nThere appears to be some contribution - but not substantial.\n\n\nNote for Area Chairs and PC chairs:\nThis paper is available in\nhttps://arxiv.org/abs/2205.14986\nin which the author names are revealed.\n\nThe authors of above are a subset of those in [2] in this submitted manuscript.\n", "strengths": "Pros:\nTransformers are hard to train, and providing any means to make them work for SSL\ntasks, for DA/TL jobs is even harder.\nAuthors appears to address that problem to some extent,\nNumerous results have shown the potential.\n\nCons\nAlthough convinced by elaborate experimentations and ablations studies,\nits difficult to get convinced due to lack of sufficient analytical proofs and justifications.\n\nThe manuscript has just one Eqn - the loss function used - with just the L1 vs L2 norm being discussed.\nWhat about use of other loss functions - KLD and variants, Softmax, CE with variants etc. ?\n \n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a Group Masked Model Learning (GMML), a Self-Supervised Learning (SSL) mechanism\nfor pretraining vision transformers with the ability to extract the contextual information\npresent in all the concepts in an image. This is achieved by manipulating\nrandomly groups of connected tokens, ensuingly covering a meaningful part of\na semantic concept, and then recovering the hidden semantic information from\nthe visible part of the concept. The paper suggests that GMML implicitly introduces a novel data augmentation\nprocess.\n\nExperimental results are shown for Image Classification, DA in Transfer Learning, and a few types of ablations studies.\n\nAt the end of the manuscript, the authors claim the following - \nGMML based pretraining makes the vision transformers data-efficient. Also, GMML is the first\nself-supervised pretraining work which consistently outperformed supervised pretraining for any\npretraining and finetuning dataset, regardless of their sizes.\n\nThe key impact of the proposed GMML is that it makes it possible for transformers\nto train on small and medium size datasets. It is not only data efficient, but its outstanding information\nextraction ability enables it to outperform state-of-the-art supervised and self-supervised\nmethods with large margins.\n", "strength_and_weaknesses": "Pros:\nTransformers are hard to train, and providing any means to make them work for SSL\ntasks, for DA/TL jobs is even harder.\nAuthors appears to address that problem to some extent,\nNumerous results have shown the potential.\n\nCons\nAlthough convinced by elaborate experimentations and ablations studies,\nits difficult to get convinced due to lack of sufficient analytical proofs and justifications.\n\nThe manuscript has just one Eqn - the loss function used - with just the L1 vs L2 norm being discussed.\nWhat about use of other loss functions - KLD and variants, Softmax, CE with variants etc. ?\n \n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Paper quite clearly written.\nNot much issues with the quality of presentation.\n\nEqn. (1) does not appear to be novel;\nNo other analytics provided for strengthening the proposed concept - convergence, impact of bias,..etc\n\nExperimentations appear to be reproducible, although not much time to explore myself\nand verify the same.", "summary_of_the_review": "\nI am not convinced with the paper.\nThere appears to be some contribution - but not substantial.\n\n\nNote for Area Chairs and PC chairs:\nThis paper is available in\nhttps://arxiv.org/abs/2205.14986\nin which the author names are revealed.\n\nThe authors of above are a subset of those in [2] in this submitted manuscript.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666586654925}], "openreview_url": "https://openreview.net/forum?id=vDY5Y8HMNxO", "arxiv_id": "2205.14986", "paper_pdf": "papers/vDY5Y8HMNxO.pdf", "paper_pdf_sha256": "f07fe9b92278b39fe118f4556af1996ea075966c1e7d8161e155158e60001517", "paper_pdf_bytes": 2678807, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sara-Ahmed/GMML", "code_repository": "Sara-Ahmed/GMML", "code_commit": "1882712745e87166c26dd77c1dd346f84fde6064", "code_archive": "repos/vDY5Y8HMNxO.zip", "code_archive_sha256": "3d0fed49275870d485c00b45dd2ba30cf136ffb351d879e03f641a5288c5e5c6", "code_archive_bytes": 105705, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 118, "github_languages": {"Python": 76742}, "github_archived": false, "github_pushed_at": "2022-08-12T10:43:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gmml-is-all-you-need"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RNf9AgtRtL", "year": 2022, "status": "rejected", "title": "Continuous Control With Ensemble Deep Deterministic Policy Gradients", "authors": ["Piotr Januszewski", "Mateusz Olko", "Michał Królikowski", "Jakub Swiatkowski", "Marcin Andrychowicz", "Łukasz Kuciński", "Piotr Miłoś"], "authorids": ["~Piotr_Januszewski1", "~Mateusz_Olko1", "~Michał_Królikowski1", "~Jakub_Swiatkowski1", "~Marcin_Andrychowicz1", "~Łukasz_Kuciński1", "~Piotr_Miłoś1"], "authors_source": "OpenReview API", "abstract": "The growth of deep reinforcement learning (RL) has brought multiple exciting tools and methods to the field. This rapid expansion makes it important to understand the interplay between individual elements of the RL toolbox. We approach this task from an empirical perspective by conducting a study in the continuous control setting. We present multiple insights of fundamental nature, including: a commonly used additive action noise is not required for effective exploration and can even hinder training; the performance of policies trained using existing methods varies significantly across training runs, epochs of training, and evaluation runs; the critics' initialization plays the major role in ensemble-based actor-critic exploration, while the training is mostly invariant to the actors' initialization; a strategy based on posterior sampling explores better than the approximated UCB combined with the weighted Bellman backup; the weighted Bellman backup alone cannot replace the clipped double Q-Learning. As a conclusion, we show how existing tools can be brought together in a novel way, giving rise to the Ensemble Deep Deterministic Policy Gradients (ED2) method, to yield state-of-the-art results on continuous control tasks from \\mbox{OpenAI Gym MuJoCo}. From the practical side, ED2 is conceptually straightforward, easy to code, and does not require knowledge outside of the existing RL toolbox.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "pDiASP6k0H", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3656/Reviewer_thDH"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper presents an empirical study evaluating the commonly accepted design choice in off-policy Deep RL algorithms in continuous control settings. The use of additive exploration noise, initialization choices, update frequency, and precision for retraining are tested empirically highlighting some interesting results. The paper also introduces ED2 - an ensemble method utilizing the design choices from the study which is demonstrated to achieve SOTA results on Mujoco benchmarks.", "review_text": "The paper is very well-written, flows smoothly and is a pleasure to read. The ideas are well articulated and clear.\n\nResult in Fig 2 is surprising as it suggests that the additive normal action noise is entirely unnecessary. This has been a fixture is most DRL algorithms. However, looking in Appendix B.2, the authors did not test Ornstein-Uhlenbeck noise (DDPG paper) as one of the baselines. Adding this popular choice would complete this empirical evaluation. I do not think that OU noise would change the conclusions, but it would be nice to include for the sake of completeness - considering that it is used in some of the seminal works in the field.\n\nThe results for the stability experiments in Fig 6-8 on inference stability, asymptotic performance stability and training stability do not seem that surprising to me. Wouldn’t an ensemble method fundamentally lead to more stable learning? SUNRISE seems to be an exception here for some domains. However, ED2 being more ‘stable’ that a lone network algorithm (like SAC baseline) run seems to be rather obvious to me. An ensemble method should fundamentally be more stable as it has the advantage of N=5 random initializations. Would perhaps a better comparison be against an ensemble of SAC runs? \n\nThe main contribution of the paper are (1) the empirical study into the various design choices within DRL algorithms used for continuous control settings and (2) an ensemble approach that integrates these learnings. While the ensemble method achieves SOTA results on many tasks and the empirical study presents some riveting results, the novelty in the paper is quite limited. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents an empirical study evaluating the commonly accepted design choice in off-policy Deep RL algorithms in continuous control settings. The use of additive exploration noise, initialization choices, update frequency, and precision for retraining are tested empirically highlighting some interesting results. The paper also introduces ED2 - an ensemble method utilizing the design choices from the study which is demonstrated to achieve SOTA results on Mujoco benchmarks.", "main_review": "The paper is very well-written, flows smoothly and is a pleasure to read. The ideas are well articulated and clear.\n\nResult in Fig 2 is surprising as it suggests that the additive normal action noise is entirely unnecessary. This has been a fixture is most DRL algorithms. However, looking in Appendix B.2, the authors did not test Ornstein-Uhlenbeck noise (DDPG paper) as one of the baselines. Adding this popular choice would complete this empirical evaluation. I do not think that OU noise would change the conclusions, but it would be nice to include for the sake of completeness - considering that it is used in some of the seminal works in the field.\n\nThe results for the stability experiments in Fig 6-8 on inference stability, asymptotic performance stability and training stability do not seem that surprising to me. Wouldn’t an ensemble method fundamentally lead to more stable learning? SUNRISE seems to be an exception here for some domains. However, ED2 being more ‘stable’ that a lone network algorithm (like SAC baseline) run seems to be rather obvious to me. An ensemble method should fundamentally be more stable as it has the advantage of N=5 random initializations. Would perhaps a better comparison be against an ensemble of SAC runs? \n\nThe main contribution of the paper are (1) the empirical study into the various design choices within DRL algorithms used for continuous control settings and (2) an ensemble approach that integrates these learnings. While the ensemble method achieves SOTA results on many tasks and the empirical study presents some riveting results, the novelty in the paper is quite limited. \n", "summary_of_the_review": "The paper presents some interesting learnings via a diligent evaluation of the design choices used in Deep RL algorithms for continuous control settings. The paper presents an ensemble method using these learnings and posts some SOTA results. I believe the community would benefit from these learnings and the resulting ED2 approach. However, I do note that the novelty for the method is quite limited as ED2 is fundamentally an ensemble of previous method (SOP) with well-evaluated design parameters for the target task of Mujoco benchmarks. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636347527837}, {"id": "KL2jvGzFgMO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3656/Reviewer_2e3U"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper presents an empirical study between different commonly-used tricks and elements of off-policy RL algorithms and tries to understand the interplay between those elements. The authors propose a new method called ED2 that utilizes their insights from the empirical study.", "review_text": "strengths\n1. The paper is also well-written and easy to understand\n2. I believe the paper studies a problem that's important and significant in the rl community and very relevant to the venue\n\nweaknesses/questions\nIn general, I think the results should be treated more carefully and some theoretical motivation is needed. Following are some of my questions. \n1.in section 4.1, the conclusion of \"This result shows that no additional exploration mechanism, often in a form of an exploration noise (Lillicrap et al., 2016; Fujimoto et al., 2018; Wang et al., 2020), is required for the diverse data collection and it can even hinder training.\" seems rather strong judging from fig.2 (I would say they are roughly the same). \n2. \"Figure 3 shows that ED2 magnifies the beneficial effect coming from the deterministic exploration.\", wouldn't it make better sense if you compare against the ones that are also ensemble, e.g. ensemble SOP instead of SOP? How do you know what brings the performance, is it the ensemble or the determinist action, or both? \n3. fig.4 again, why not compare the ensemble baselines? I think it's important to ablate the design choices that are actually important\n\nSome minor problems. \n1. maybe more introduction for SOP (e.g. what's the ere replay buffer?) \n2.sec.3, \"These two choices ensure coherent and temporally-extended exploration similarly to Osband et al. (2016).\", I do not understand why is the exploration \"coherent and temporally-extended\". \n3.in sec.3, I guess the used and not used should be comparable and matched, but why is action normalization compared with \"observations and rewards normalization\"? \n4.fig.5, what does it mean for the Humanoid velocities? could you elaborate?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents an empirical study between different commonly-used tricks and elements of off-policy RL algorithms and tries to understand the interplay between those elements. The authors propose a new method called ED2 that utilizes their insights from the empirical study.", "main_review": "strengths\n1. The paper is also well-written and easy to understand\n2. I believe the paper studies a problem that's important and significant in the rl community and very relevant to the venue\n\nweaknesses/questions\nIn general, I think the results should be treated more carefully and some theoretical motivation is needed. Following are some of my questions. \n1.in section 4.1, the conclusion of \"This result shows that no additional exploration mechanism, often in a form of an exploration noise (Lillicrap et al., 2016; Fujimoto et al., 2018; Wang et al., 2020), is required for the diverse data collection and it can even hinder training.\" seems rather strong judging from fig.2 (I would say they are roughly the same). \n2. \"Figure 3 shows that ED2 magnifies the beneficial effect coming from the deterministic exploration.\", wouldn't it make better sense if you compare against the ones that are also ensemble, e.g. ensemble SOP instead of SOP? How do you know what brings the performance, is it the ensemble or the determinist action, or both? \n3. fig.4 again, why not compare the ensemble baselines? I think it's important to ablate the design choices that are actually important\n\nSome minor problems. \n1. maybe more introduction for SOP (e.g. what's the ere replay buffer?) \n2.sec.3, \"These two choices ensure coherent and temporally-extended exploration similarly to Osband et al. (2016).\", I do not understand why is the exploration \"coherent and temporally-extended\". \n3.in sec.3, I guess the used and not used should be comparable and matched, but why is action normalization compared with \"observations and rewards normalization\"? \n4.fig.5, what does it mean for the Humanoid velocities? could you elaborate?", "summary_of_the_review": "I believe the paper should present a more careful analysis of these design elements and put more effort into justifying their choices used for comparison.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635947833100}, {"id": "VhX9kO8aP3I", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3656/Reviewer_ZQ4d"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a deep reinforcement algorithm, Ensemble Deep Deterministic Policy Gradients (ED2), for continuous control tasks. The algorithm is empirically derived and is claimed to represent SotA performance on several tasks and while providing more stable results. These claims are justified based primarily on the (reward and stability) results on 4 MuJoCo environments. \n", "review_text": "Pros:\n- The code is open-sourced, which is excellent for reproducibility \n- There is significant focus on the stability of RL algorithms, which is great to see. Often RL papers just present the reward curves and information on the stability of these algorithms is important. \n- Using 30 random seeds (instead of the usual 3-6) is great and strengthens any empirical claims\n\nCons:\n- The empirical results don’t seem to justify the algorithm. Figure 4 represents the main empirical argument for ED2 but the empirical gains shown are minor. On Walker it appears as though SUNRISE actually outperforms ED2. Humanoid and Hopper both have ED2 on top, but only by a very slim margin (one that has overlapping confidence intervals with other algorithms). \n- Small number of environments. The algorithms are only compared on 4 environments (Ant, Hopper, Humanoid, Walker) when MuJoCo/OpenAI Gym offers many more (e.g. HalfCheetah, InvertedDoublePendulum, Swimmer, Reacher, etc.). The empirical results could be more compelling if the scope was increased. If you expanded the results presented in Table 1 for all algorithms, this could be a more compelling argument. \n- Much of the paper is dedicated to information that would be better in the appendix. Information about non-essential ablation studies (e.g. Figure 2 shows the impacts are pretty minor) are not necessary in the main body. These figures could be replaced with more directly relevant information, like to aid claims of stability, e.g. by injecting varying degrees of noise/randomness into the environment and evaluating relative performances, or evaluating sensitivity to other hyperparameters (network structure, learning rates, etc.) \n- Lack of theoretical understanding. I am well aware of the problems of equation packing and the disconnect that often occurs in RL papers between the algorithm and the theoretical justification, but at least providing some insight into comparisons of UCB, posterior sampling, priors, etc. would be nice. \n- The empirical comparisons don’t seem to offer an even playing field for all algorithms. Based on (Haarnoja et al., 2019), SAC should reach ~8000 on Humanoid-v2 (after 10e6 steps). However, the experiments are only conducted till 3e6 steps with multiple algorithms appearing to still have upward trajectories. There seems to be no mention of a focus on the speed of learning/early learning performance vs final performance. \n- The relationship between Figure 4 and Figure 6 is not clear. Figure 6 seems to indicate that (e.g.) pretty much all algorithms have >1,000 STD on Humanoid and Ant early on in training. However, the bootstrapped CI in Figure 4 is substantially smaller. This discrepancy is uncommon and should be explained. \n- Reinforcement learning is notoriously brittle and I would encourage references to the previous work done evaluating the effect of random seeds/initialization on agent performance e.g. the famous (Henderson et al., 2019).\n- Figure 5 does not provide any support or meaningful insight. The velocity of humanoid is not what the algorithm is optimizing. It is included for “completeness”, but this seems to be a cherry-picked statistic that doesn’t convey anything meaningful (while the speed is incorporated in the reward function, since the rewards are much closer than the velocities we don’t see what the tradeoff is). \n- Figure 1 does not really help clarify anything. If one already knows the environments, then the figure is unnecessary, if one is unfamiliar with them, the figure doesn’t show what actually transpires in the environment and doesn’t clear anything up.\n- To claim ED2 is really SotA, further analysis is necessary across environments like Agarwal et al. 2021, Barreto et al. 2010, Jordan et al. 2020, etc. \n- There are minor typographical inconsistencies, e.g. differing usages of \\cite{} and \\citep{}\n- Inconsistent background information. The paper provides a definition of standard deviation, an extremely common statistical measurement, but not for much more niche terms such as approximated UCB. To be clear, I have no problem with giving the formula for STD, but not giving definitions for much less widely known terms is something that could be fixed. \n\nMisc/Note:\nThis paper seems like it’s trying to do two things at once: (1) provide a review of RL techniques (e.g. exploitation techniques), evaluate their impact and report on the key takeaways of this empirical analysis, and (2) introduce a novel RL algorithm and justify it’s empirical construction and performance. Both of these are papers that are perfectly fine, but by trying to do both it leaves something lacking from each of them. If this was a review of techniques, I would like to see more continuous control algorithms evaluated and more techniques experimented with. If this is just introducing a novel algorithm, I would like to see more theoretical explanations of the techniques (e.g. theoretical derivations and insights into the effect of K)  and more extensive ablation studies (e.g. evaluating on more MuJoCo tasks or other continuous control environments that have different properties such as RLBench, Industrial control benchmark, assistive gym, DM Control suite, etc.). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a deep reinforcement algorithm, Ensemble Deep Deterministic Policy Gradients (ED2), for continuous control tasks. The algorithm is empirically derived and is claimed to represent SotA performance on several tasks and while providing more stable results. These claims are justified based primarily on the (reward and stability) results on 4 MuJoCo environments. \n", "main_review": "Pros:\n- The code is open-sourced, which is excellent for reproducibility \n- There is significant focus on the stability of RL algorithms, which is great to see. Often RL papers just present the reward curves and information on the stability of these algorithms is important. \n- Using 30 random seeds (instead of the usual 3-6) is great and strengthens any empirical claims\n\nCons:\n- The empirical results don’t seem to justify the algorithm. Figure 4 represents the main empirical argument for ED2 but the empirical gains shown are minor. On Walker it appears as though SUNRISE actually outperforms ED2. Humanoid and Hopper both have ED2 on top, but only by a very slim margin (one that has overlapping confidence intervals with other algorithms). \n- Small number of environments. The algorithms are only compared on 4 environments (Ant, Hopper, Humanoid, Walker) when MuJoCo/OpenAI Gym offers many more (e.g. HalfCheetah, InvertedDoublePendulum, Swimmer, Reacher, etc.). The empirical results could be more compelling if the scope was increased. If you expanded the results presented in Table 1 for all algorithms, this could be a more compelling argument. \n- Much of the paper is dedicated to information that would be better in the appendix. Information about non-essential ablation studies (e.g. Figure 2 shows the impacts are pretty minor) are not necessary in the main body. These figures could be replaced with more directly relevant information, like to aid claims of stability, e.g. by injecting varying degrees of noise/randomness into the environment and evaluating relative performances, or evaluating sensitivity to other hyperparameters (network structure, learning rates, etc.) \n- Lack of theoretical understanding. I am well aware of the problems of equation packing and the disconnect that often occurs in RL papers between the algorithm and the theoretical justification, but at least providing some insight into comparisons of UCB, posterior sampling, priors, etc. would be nice. \n- The empirical comparisons don’t seem to offer an even playing field for all algorithms. Based on (Haarnoja et al., 2019), SAC should reach ~8000 on Humanoid-v2 (after 10e6 steps). However, the experiments are only conducted till 3e6 steps with multiple algorithms appearing to still have upward trajectories. There seems to be no mention of a focus on the speed of learning/early learning performance vs final performance. \n- The relationship between Figure 4 and Figure 6 is not clear. Figure 6 seems to indicate that (e.g.) pretty much all algorithms have >1,000 STD on Humanoid and Ant early on in training. However, the bootstrapped CI in Figure 4 is substantially smaller. This discrepancy is uncommon and should be explained. \n- Reinforcement learning is notoriously brittle and I would encourage references to the previous work done evaluating the effect of random seeds/initialization on agent performance e.g. the famous (Henderson et al., 2019).\n- Figure 5 does not provide any support or meaningful insight. The velocity of humanoid is not what the algorithm is optimizing. It is included for “completeness”, but this seems to be a cherry-picked statistic that doesn’t convey anything meaningful (while the speed is incorporated in the reward function, since the rewards are much closer than the velocities we don’t see what the tradeoff is). \n- Figure 1 does not really help clarify anything. If one already knows the environments, then the figure is unnecessary, if one is unfamiliar with them, the figure doesn’t show what actually transpires in the environment and doesn’t clear anything up.\n- To claim ED2 is really SotA, further analysis is necessary across environments like Agarwal et al. 2021, Barreto et al. 2010, Jordan et al. 2020, etc. \n- There are minor typographical inconsistencies, e.g. differing usages of \\cite{} and \\citep{}\n- Inconsistent background information. The paper provides a definition of standard deviation, an extremely common statistical measurement, but not for much more niche terms such as approximated UCB. To be clear, I have no problem with giving the formula for STD, but not giving definitions for much less widely known terms is something that could be fixed. \n\nMisc/Note:\nThis paper seems like it’s trying to do two things at once: (1) provide a review of RL techniques (e.g. exploitation techniques), evaluate their impact and report on the key takeaways of this empirical analysis, and (2) introduce a novel RL algorithm and justify it’s empirical construction and performance. Both of these are papers that are perfectly fine, but by trying to do both it leaves something lacking from each of them. If this was a review of techniques, I would like to see more continuous control algorithms evaluated and more techniques experimented with. If this is just introducing a novel algorithm, I would like to see more theoretical explanations of the techniques (e.g. theoretical derivations and insights into the effect of K)  and more extensive ablation studies (e.g. evaluating on more MuJoCo tasks or other continuous control environments that have different properties such as RLBench, Industrial control benchmark, assistive gym, DM Control suite, etc.). \n", "summary_of_the_review": "Given that the algorithm is entirely empirically derived and the empirical results are not compelling, I have given this paper a reject. I appreciate that there are important ideas represented here, but it currently isn’t up to the ICLR standard. \n\nAfter author rebuttal, the score has moved from 3 to 5 (see comment for more information). \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635892255028}, {"id": "Ue9AbLA4V-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3656/Reviewer_M5Fw"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper has two main contributions: it introduces an ensemble-based actor-critic method, and it answers some pertinent questions in policy optimization by focusing on its different components. The ensemble is different from multi-actor learners that interact with multiple environments simultaneously, violating the standard RL setup. Instead, the learner of this paper maintains multiple actors and critics but uses only a single actor at a time to interact with the environment. All actors and critics are trained on a common replay buffer. The base method is the streamlined off-policy (SOP) method, which unlike soft actor-critic (SAC) doesn’t use an entropy bonus. Additionally, no exploration noise is added, resulting in their Ensemble Deep Deterministic (ED2) method.\n\nThe proposed algorithm ED2 is shown to be superior and more stable in performance according to different measures compared to existing methods. It is also revealed that actor initialization affects performance less than critic initialization. ED2 uses deterministic actors, and its exploration comes from sampling among the actors. Such a form of exploration is also shown to be superior to UCB-style exploration.\n", "review_text": "Strength:\n\nThe main strength of the work is introducing a straightforward extension of an existing base actor-critic method that substantially outperforms existing algorithms in standard benchmark tasks. The new algorithm also has some desirable properties for policy optimization such as not having random additive noise and providing stable performance.\n\nMoreover, some key insights on deep policy gradient methods are presented such as the contribution of actor and critic initialization.\n\nAnother strength of the paper is its focus on various details such as ideas that didn’t work as well as ablative studies in various manners.\n\nWeakness:\n\nThe main weakness of the work is the fairness of the experiments. This deficiency is common in many papers including those that get published in top conferences but acceptance doesn’t justify wrong choices. I would be willing to know from the authors their thoughts on it.\n\nThe first issue with such experiments is that only a single hyper-parameters are used for the methods to show comparison. However, the claims concluded from such results are that one algorithm outperformed the others. However, to make such a claim, different hyper-parameter values should be tried for all algorithms. Otherwise, the claim should be humbler such as our method outperforms the competing methods with default choices of hyper-parameters and such. Such hyper-parameter search is also necessary for ablative studies. When we are removing one component at a time, we cannot assume that the default hyper-parameter configuration of the original method will still be effective for the subsequent variants.\n\nI understand that it will make deep RL experiments much more expensive. However, it isn’t necessary to perform a grid search over hyper-parameters. It has been shown before that random search can give close to the best performance within a handful of trials of configurations, which will considerably reduce the search cost.\n\nIn a similar vein, it has been common to compare algorithms with different computational profiles. However, when a new algorithm is computationally way more expensive than the competitors, is it fair to compare them with such computational disparity and claim one algorithm is better than the other?\n\nDetails on the computational expense of ED2 are not given. How many actors are used for ED2? How much more expensive ED2 is compared to its competitors such as SAC or SOP?\n\nOther comments:\n- To understand more clearly, det. SOP uses no exploration whatsoever and still performs well on these tasks?\n\n- Both the action averaging and greedy choice among actors yielded similar results. This leads me to suspect whether the actors either converged to similar performant behavior or stationary behavior with zero torque, which upon averaging gives a behavior similar to the greedy one.\n\n- Considering the performance and the computational expense compared to ED2, det. SOP seems a strong contender. Why is it not added to Figures 6, 7, or 8?\n\n- Figure 9 somewhat makes sense except that there is a puzzle. When ED2 is reduced to single critic, the diversity is reduced considerably, which hurts possibly the exploration and reduces performance substantially. But if we reduce ED2 single critic further by also having a single actor, then don’t we get det. SOP, which wasn’t doing as badly as ED2 single critic? How can that be explained?\n\n- What's really the motivation behind having a separate evaluation phase just to measure performance for plot when learning online? If these algorithms are deployed to learn online say on a robot, their online performance is the actual evaluation. Creating an additional evaluation phase to measure performance will only delay its learning in real-time. In what case, such a separate evaluation is useful other than because many other works repeat it? Even if there is a case, isn't it quite restrictive? Wouldn't it be important to see the online performance of ED2 as it randomly draws actors to interact? If that performance is also good, it would be a more interesting and stronger result.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper has two main contributions: it introduces an ensemble-based actor-critic method, and it answers some pertinent questions in policy optimization by focusing on its different components. The ensemble is different from multi-actor learners that interact with multiple environments simultaneously, violating the standard RL setup. Instead, the learner of this paper maintains multiple actors and critics but uses only a single actor at a time to interact with the environment. All actors and critics are trained on a common replay buffer. The base method is the streamlined off-policy (SOP) method, which unlike soft actor-critic (SAC) doesn’t use an entropy bonus. Additionally, no exploration noise is added, resulting in their Ensemble Deep Deterministic (ED2) method.\n\nThe proposed algorithm ED2 is shown to be superior and more stable in performance according to different measures compared to existing methods. It is also revealed that actor initialization affects performance less than critic initialization. ED2 uses deterministic actors, and its exploration comes from sampling among the actors. Such a form of exploration is also shown to be superior to UCB-style exploration.\n", "main_review": "Strength:\n\nThe main strength of the work is introducing a straightforward extension of an existing base actor-critic method that substantially outperforms existing algorithms in standard benchmark tasks. The new algorithm also has some desirable properties for policy optimization such as not having random additive noise and providing stable performance.\n\nMoreover, some key insights on deep policy gradient methods are presented such as the contribution of actor and critic initialization.\n\nAnother strength of the paper is its focus on various details such as ideas that didn’t work as well as ablative studies in various manners.\n\nWeakness:\n\nThe main weakness of the work is the fairness of the experiments. This deficiency is common in many papers including those that get published in top conferences but acceptance doesn’t justify wrong choices. I would be willing to know from the authors their thoughts on it.\n\nThe first issue with such experiments is that only a single hyper-parameters are used for the methods to show comparison. However, the claims concluded from such results are that one algorithm outperformed the others. However, to make such a claim, different hyper-parameter values should be tried for all algorithms. Otherwise, the claim should be humbler such as our method outperforms the competing methods with default choices of hyper-parameters and such. Such hyper-parameter search is also necessary for ablative studies. When we are removing one component at a time, we cannot assume that the default hyper-parameter configuration of the original method will still be effective for the subsequent variants.\n\nI understand that it will make deep RL experiments much more expensive. However, it isn’t necessary to perform a grid search over hyper-parameters. It has been shown before that random search can give close to the best performance within a handful of trials of configurations, which will considerably reduce the search cost.\n\nIn a similar vein, it has been common to compare algorithms with different computational profiles. However, when a new algorithm is computationally way more expensive than the competitors, is it fair to compare them with such computational disparity and claim one algorithm is better than the other?\n\nDetails on the computational expense of ED2 are not given. How many actors are used for ED2? How much more expensive ED2 is compared to its competitors such as SAC or SOP?\n\nOther comments:\n- To understand more clearly, det. SOP uses no exploration whatsoever and still performs well on these tasks?\n\n- Both the action averaging and greedy choice among actors yielded similar results. This leads me to suspect whether the actors either converged to similar performant behavior or stationary behavior with zero torque, which upon averaging gives a behavior similar to the greedy one.\n\n- Considering the performance and the computational expense compared to ED2, det. SOP seems a strong contender. Why is it not added to Figures 6, 7, or 8?\n\n- Figure 9 somewhat makes sense except that there is a puzzle. When ED2 is reduced to single critic, the diversity is reduced considerably, which hurts possibly the exploration and reduces performance substantially. But if we reduce ED2 single critic further by also having a single actor, then don’t we get det. SOP, which wasn’t doing as badly as ED2 single critic? How can that be explained?\n\n- What's really the motivation behind having a separate evaluation phase just to measure performance for plot when learning online? If these algorithms are deployed to learn online say on a robot, their online performance is the actual evaluation. Creating an additional evaluation phase to measure performance will only delay its learning in real-time. In what case, such a separate evaluation is useful other than because many other works repeat it? Even if there is a case, isn't it quite restrictive? Wouldn't it be important to see the online performance of ED2 as it randomly draws actors to interact? If that performance is also good, it would be a more interesting and stronger result.", "summary_of_the_review": "The paper’s strength is a straightforward performant policy optimization method and the insights developed through experiments. However, hyper-parameter search isn’t performed to substantiate the strong claims and it isn’t clear how much more computationally expensive ED2 is compared to its competitors.\n\n*** updated ***\n\nI will update later. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635739680061}, {"id": "2oYcEsHZsC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3656/Reviewer_hF9t"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper conducted an experimental study over a range of tricks that are often exploited to facilitate ensemble deep reinforcement learning. The experiment results show several interesting findings. For example, it was found that commonly used additive action noise may not be necessary for effective exploration. Meanwhile, experiments show that the initialization of critics perhaps has a higher impact on learning performance than the initialization methods adopted for actors. These findings can be quite important to guide future design of more effective ensemble reinforce learning algorithms.", "review_text": "While this paper seems to show some interesting new results related to ensemble reinforcement learning, there are several issues with this paper in this current shape:\n\n1. The technical innovation of this paper remains largely unclear. It was claimed by the authors in this paper that ED2 brings together existing RL tools in a novel way. However, it is unclear which part of the design of ED2 is truly novel. As far as I am aware, ED2 mainly used existing training techniques and ensemble tricks. A series of experiments were carried out to justify the use of several different tricks in ED2. While the combined use of these tricks might be new in ED2, it is not clear why such a combination is potentially more superior than other possible combinations. Furthermore, since the experiments focus mainly on four benchmark problems, it is questionable whether ED2 can achieve clearly better performance over other ensemble baseline algorithms on a much wider range of reinforcement learning problems. Hence the novelty and technical contribution of this paper may need to be improved.\n\n2. This paper lacks theoretical depth. The experiment results in the paper only revealed some insights. However, no further theoretical analysis was conducted to verify (or at least partially explain) the experimental observations. For example, on page 6, the authors conjectured that good exploration may come more from the critic. While this sounds interesting, it is unclear why the critic will play such a critical role to induce effective exploration and what the corresponding conditions are for this to happen.\n\n3. Some experiment findings and the corresponding claims do not appear to be consistent. For example, on page 4, the authors found experimentally that additive normal action noise can substantially improve the Ant performance. They subsequently concluded that additive noise is not required for effective learning. These two claims do not sound consistent. Accordingly, the main findings discovered in the paper may need to be further verified.\n\n4. Some experiment findings appear to be well-known a priori in the literature. For example, as acknowledged by the authors, posterior sampling techniques can be more effective than the OFU strategy for action selection. Consequently, the technical contribution of the corresponding experiment results does not seem to be sufficiently strong.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper conducted an experimental study over a range of tricks that are often exploited to facilitate ensemble deep reinforcement learning. The experiment results show several interesting findings. For example, it was found that commonly used additive action noise may not be necessary for effective exploration. Meanwhile, experiments show that the initialization of critics perhaps has a higher impact on learning performance than the initialization methods adopted for actors. These findings can be quite important to guide future design of more effective ensemble reinforce learning algorithms.", "main_review": "While this paper seems to show some interesting new results related to ensemble reinforcement learning, there are several issues with this paper in this current shape:\n\n1. The technical innovation of this paper remains largely unclear. It was claimed by the authors in this paper that ED2 brings together existing RL tools in a novel way. However, it is unclear which part of the design of ED2 is truly novel. As far as I am aware, ED2 mainly used existing training techniques and ensemble tricks. A series of experiments were carried out to justify the use of several different tricks in ED2. While the combined use of these tricks might be new in ED2, it is not clear why such a combination is potentially more superior than other possible combinations. Furthermore, since the experiments focus mainly on four benchmark problems, it is questionable whether ED2 can achieve clearly better performance over other ensemble baseline algorithms on a much wider range of reinforcement learning problems. Hence the novelty and technical contribution of this paper may need to be improved.\n\n2. This paper lacks theoretical depth. The experiment results in the paper only revealed some insights. However, no further theoretical analysis was conducted to verify (or at least partially explain) the experimental observations. For example, on page 6, the authors conjectured that good exploration may come more from the critic. While this sounds interesting, it is unclear why the critic will play such a critical role to induce effective exploration and what the corresponding conditions are for this to happen.\n\n3. Some experiment findings and the corresponding claims do not appear to be consistent. For example, on page 4, the authors found experimentally that additive normal action noise can substantially improve the Ant performance. They subsequently concluded that additive noise is not required for effective learning. These two claims do not sound consistent. Accordingly, the main findings discovered in the paper may need to be further verified.\n\n4. Some experiment findings appear to be well-known a priori in the literature. For example, as acknowledged by the authors, posterior sampling techniques can be more effective than the OFU strategy for action selection. Consequently, the technical contribution of the corresponding experiment results does not seem to be sufficiently strong.", "summary_of_the_review": "This paper conducted an experimental study over a range of tricks that are often exploited to facilitate ensemble deep reinforcement learning. The corresponding empirical findings can be quite important to guide future design of more effective ensemble reinforce learning algorithms. Meanwhile, the technical novelty and theoretical depth of this paper may need to be strengthened.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635201843301}], "openreview_url": "https://openreview.net/forum?id=RNf9AgtRtL", "arxiv_id": "2111.15382", "paper_pdf": "papers/RNf9AgtRtL.pdf", "paper_pdf_sha256": "93c3ca8bcb21ff96bcfde5c2e39a47af731f5f260a87c10485e7dd3b9e6e95f6", "paper_pdf_bytes": 3435979, "paper_pdf_source": "openreview", "code_url": "https://github.com/ed2-paper/ED2", "code_repository": "ed2-paper/ED2", "code_commit": "884e5428fd47d63266dbdd6100d13235d9322595", "code_archive": "repos/RNf9AgtRtL.zip", "code_archive_sha256": "cbfd2edd960aa3c1c93be7e465aafd513148429afdcd7c4ba83bcda9d23e0692", "code_archive_bytes": 167634, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 149, "github_languages": {"Python": 105353}, "github_archived": false, "github_pushed_at": "2021-06-09T15:38:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continuous-control-with-ensemble-deep-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7qmQNB6Wn_B", "year": 2021, "status": "rejected", "title": "Diversity Actor-Critic: Sample-Aware Entropy Regularization for Sample-Efficient Exploration", "authors": ["Seungyul Han", "Youngchul Sung"], "authorids": ["~Seungyul_Han1", "~Youngchul_Sung1"], "authors_source": "OpenReview API", "abstract": "Policy entropy regularization is commonly used for better exploration in deep reinforcement learning (RL). However, policy entropy regularization is sample-inefficient in off-policy learning since it does not take the distribution of previous samples stored in the replay buffer into account. In order to take advantage of the previous sample distribution from the replay buffer for sample-efficient exploration, we propose sample-aware entropy regularization which maximizes the entropy of weighted sum of the policy action distribution and the sample action distribution from the replay buffer. We formulate the problem of sample-aware entropy regularized policy iteration,  prove its convergence, and provide a practical algorithm named diversity actor-critic (DAC) which is a generalization of soft actor-critic (SAC). Numerical results show that DAC significantly outperforms SAC baselines and other state-of-the-art RL algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "IXFJcbejoak", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1185/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the exploration efficiency issues in off-policy deep reinforcement learning (DRL). The authors identify a sample efficiency limitation in the classical entropy regularization, which does not take into account the existing samples in the replay buffer. To avoid repeated sampling of previously seen scenarios/actions, the authors propose to replace the current policy in the entropy term with a mixture of the empirical policy estimation from the replay buffer and the current policy, and term this approach as sample-aware entropy regularization. The authors then propose a theoretical algorithm called sample-aware entropy regularized policy iteration, which is a generalization of the soft policy iteration (SPI) algorithm, and show that it converges assuming that the empirical policy estimation is fixed. A practical algorithm based on the sample-aware entropy regularized policy iteration, called Diversity Actor-Critic (DAC), is then proposed. This algorithm is a generalization of the well-known soft actor-critic (SAC) algorithm. Finally, numerical experiments show that DAC outperforms SAC and other SOTA RL algorithms, and some ablation studies are also provided to demonstrate the effect of hyper-parameter choices in DAC.\n\nIn general, the approach is novel to my knowledge and the high level idea of using mixed policies in the entropy regularization to avoid repeated sampling and encourage unseen scenarios/actions is also interesting and reasonable. However, there are some clarity and technical issues that should be addressed and improved, as listed below:\n1. The authors study finite horizon MDPs, for which the optimal policy should be non-stationary in general. However, the authors only consider stationary policies. Instead, the authors should either change the underlying setting to infinite horizon MDPs or consider non-stationary policies.  \n2. In (2), $s_t$ should be replaced by an arbitrary $s$ in the state space. Otherwise there may be contradicting definitions of the policy $q$ if $s_t$ and $s_{t’}$ are equal for some two different timestamps $t$ and $t’$. And in (3), it is better to write the $q_{\\rm target}^{\\pi,\\alpha}$ in the entropy term as $q_{\\rm target}^{\\pi,\\alpha}(\\cdot|s_t)$, to be consistent with (1). \n3. It’s not very clear why the authors propose to estimate $R^{\\pi,\\alpha}$ with some (neural network) parametrized $R^{\\alpha}$. The authors mention that one can only estimate $R^{\\pi_{\\rm old},\\alpha}$ for the previous policy $\\pi_{\\rm old}$ in practice. However, since in $R^{\\pi,\\alpha}$, all the quantities including $\\pi$, $q$ and $\\alpha$ are known, I’m confused why one cannot evaluate it directly. On a related point, it’s not very clear why the estimation procedure for $\\eta$ (the parameter of $R^{\\alpha}$) using hat $J_{R^{\\alpha}}(\\eta)$ makes sense. The form of hat $J_{R^{\\alpha}}(\\eta)$ looks like an entropy term extracted from the $J_{\\pi_{\\rm old}}$ function, but it’s unclear why maximizing it gives a good estimation of $R^{\\pi,\\alpha}$. Some more explanations are needed. \n4. There seem to be several errors (at least inaccuracies) in the proof of Theorem 1 (in the Appendix). Firstly, in the proof of Lemma 1, the term “correctly estimates” is not very accurate, and should be simply stated as something like “equals”. Also, it’s not very clear when the assumption $R^{\\alpha}\\in(0,1)$ can be guaranteed (e.g., using Gaussian/soft-max policies?). Secondly, in the main proof of Theorem 1, convergence of $Q^{\\pi_i}$ to some $Q^{\\star}$ is correct, but this does not immediately imply convergence of $J_{\\pi_i}$, let alone the convergence of $\\pi_i$ to some policy $\\pi^\\star$. On a related point, the proof for the optimality of $\\pi^\\star$ in terms of $J$ is not clear. In particular, it is not clear why (7) and Lemma 2 implies the chained inequality $J_{\\pi_{\\rm new}}(\\pi_{\\rm new})\\geq J_{\\pi_{\\rm old}}(\\pi_{\\rm new})\\geq J_{\\pi_{\\rm old}}(\\pi_{\\rm old})$. I understand that the authors may feel that the proofs are similar to that of SPI, but indeed there are several significant differences (e.g., the definitions of $\\pi_{\\rm new}$ and $J_{\\pi}$). More rigorous proofs are needed for these claims. \n5. In Section 5, it is unclear why the authors need to include the parameter $c$, how to choose it and what it serves for. Some additional explanations are needed. \n6. On a high level, the eventual goal of the paper is not clearly stated. From the experiments, it seems that the average episode reward is the actual goal of concern. However, the problem setting and the theoretical results (Theorem 1) seem to indicate that the problem of concern is the discounted entropy regularized reward. Some discussion about this is needed. \n\nFinally, here are some more minor comments and suggestions:\n1. In the analysis of the sample-aware entropy regularized policy iteration, the authors assume that $q$ is fixed. However, in practice, especially in the long run (as concerned in the analysis), such an assumption will not hold (even in just an approximate sense). Can you still obtain some sort of convergence when taking into account the $q$ changes?\n2. Why do you need to divide the reward and entropy regularization term in $Q^{\\pi}$ by $\\beta$? \n3. It’s better to write out the “binary entropy function $H$\" explicitly for clarity. \n4. At the beginning of Section 4.3, “propoed” should be “proposed”, and In Section 5, “a function $s_t$” should be “a function of $s_t$”. \n5. Some high level explanations on why the $(1-\\alpha)$ term can also be dropped in (8) will be helpful.\n6. The theoretical results only show that the algorithm converges, which is already guaranteed by SPI. Is there any possibility to show that there is also some theoretical improvement?\n\nSo in short, the paper proposes an interesting modification of the max-entropy regularization framework, but contains several technical and clarity issues. Hence I think it is not yet ready for publication in its current form. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel idea with some clarity and technical issues", "review": "This paper considers the exploration efficiency issues in off-policy deep reinforcement learning (DRL). The authors identify a sample efficiency limitation in the classical entropy regularization, which does not take into account the existing samples in the replay buffer. To avoid repeated sampling of previously seen scenarios/actions, the authors propose to replace the current policy in the entropy term with a mixture of the empirical policy estimation from the replay buffer and the current policy, and term this approach as sample-aware entropy regularization. The authors then propose a theoretical algorithm called sample-aware entropy regularized policy iteration, which is a generalization of the soft policy iteration (SPI) algorithm, and show that it converges assuming that the empirical policy estimation is fixed. A practical algorithm based on the sample-aware entropy regularized policy iteration, called Diversity Actor-Critic (DAC), is then proposed. This algorithm is a generalization of the well-known soft actor-critic (SAC) algorithm. Finally, numerical experiments show that DAC outperforms SAC and other SOTA RL algorithms, and some ablation studies are also provided to demonstrate the effect of hyper-parameter choices in DAC.\n\nIn general, the approach is novel to my knowledge and the high level idea of using mixed policies in the entropy regularization to avoid repeated sampling and encourage unseen scenarios/actions is also interesting and reasonable. However, there are some clarity and technical issues that should be addressed and improved, as listed below:\n1. The authors study finite horizon MDPs, for which the optimal policy should be non-stationary in general. However, the authors only consider stationary policies. Instead, the authors should either change the underlying setting to infinite horizon MDPs or consider non-stationary policies.  \n2. In (2), $s_t$ should be replaced by an arbitrary $s$ in the state space. Otherwise there may be contradicting definitions of the policy $q$ if $s_t$ and $s_{t’}$ are equal for some two different timestamps $t$ and $t’$. And in (3), it is better to write the $q_{\\rm target}^{\\pi,\\alpha}$ in the entropy term as $q_{\\rm target}^{\\pi,\\alpha}(\\cdot|s_t)$, to be consistent with (1). \n3. It’s not very clear why the authors propose to estimate $R^{\\pi,\\alpha}$ with some (neural network) parametrized $R^{\\alpha}$. The authors mention that one can only estimate $R^{\\pi_{\\rm old},\\alpha}$ for the previous policy $\\pi_{\\rm old}$ in practice. However, since in $R^{\\pi,\\alpha}$, all the quantities including $\\pi$, $q$ and $\\alpha$ are known, I’m confused why one cannot evaluate it directly. On a related point, it’s not very clear why the estimation procedure for $\\eta$ (the parameter of $R^{\\alpha}$) using hat $J_{R^{\\alpha}}(\\eta)$ makes sense. The form of hat $J_{R^{\\alpha}}(\\eta)$ looks like an entropy term extracted from the $J_{\\pi_{\\rm old}}$ function, but it’s unclear why maximizing it gives a good estimation of $R^{\\pi,\\alpha}$. Some more explanations are needed. \n4. There seem to be several errors (at least inaccuracies) in the proof of Theorem 1 (in the Appendix). Firstly, in the proof of Lemma 1, the term “correctly estimates” is not very accurate, and should be simply stated as something like “equals”. Also, it’s not very clear when the assumption $R^{\\alpha}\\in(0,1)$ can be guaranteed (e.g., using Gaussian/soft-max policies?). Secondly, in the main proof of Theorem 1, convergence of $Q^{\\pi_i}$ to some $Q^{\\star}$ is correct, but this does not immediately imply convergence of $J_{\\pi_i}$, let alone the convergence of $\\pi_i$ to some policy $\\pi^\\star$. On a related point, the proof for the optimality of $\\pi^\\star$ in terms of $J$ is not clear. In particular, it is not clear why (7) and Lemma 2 implies the chained inequality $J_{\\pi_{\\rm new}}(\\pi_{\\rm new})\\geq J_{\\pi_{\\rm old}}(\\pi_{\\rm new})\\geq J_{\\pi_{\\rm old}}(\\pi_{\\rm old})$. I understand that the authors may feel that the proofs are similar to that of SPI, but indeed there are several significant differences (e.g., the definitions of $\\pi_{\\rm new}$ and $J_{\\pi}$). More rigorous proofs are needed for these claims. \n5. In Section 5, it is unclear why the authors need to include the parameter $c$, how to choose it and what it serves for. Some additional explanations are needed. \n6. On a high level, the eventual goal of the paper is not clearly stated. From the experiments, it seems that the average episode reward is the actual goal of concern. However, the problem setting and the theoretical results (Theorem 1) seem to indicate that the problem of concern is the discounted entropy regularized reward. Some discussion about this is needed. \n\nFinally, here are some more minor comments and suggestions:\n1. In the analysis of the sample-aware entropy regularized policy iteration, the authors assume that $q$ is fixed. However, in practice, especially in the long run (as concerned in the analysis), such an assumption will not hold (even in just an approximate sense). Can you still obtain some sort of convergence when taking into account the $q$ changes?\n2. Why do you need to divide the reward and entropy regularization term in $Q^{\\pi}$ by $\\beta$? \n3. It’s better to write out the “binary entropy function $H$\" explicitly for clarity. \n4. At the beginning of Section 4.3, “propoed” should be “proposed”, and In Section 5, “a function $s_t$” should be “a function of $s_t$”. \n5. Some high level explanations on why the $(1-\\alpha)$ term can also be dropped in (8) will be helpful.\n6. The theoretical results only show that the algorithm converges, which is already guaranteed by SPI. Is there any possibility to show that there is also some theoretical improvement?\n\nSo in short, the paper proposes an interesting modification of the max-entropy regularization framework, but contains several technical and clarity issues. Hence I think it is not yet ready for publication in its current form. ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1605130549975}, {"id": "o2VsOTAJEOC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1185/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes diversity actor-critic (DAC) for exploration in reinforcement learning. The main idea of the proposed algorithm is to take advantage of the previous sample distribution from the replay buffer for sample-efficient exploration. The authors provide convergence analysis of DAC and conduct empirical investigations on several benchmarks. \n\nPros \n\nThe idea of using previous sample distribution from the replay buffer for better exploration seems interesting. The proposed exploration bonus $\\mathcal{H}(q^{\\pi, \\alpha}_{\\text{target}})$ can be decomposed into three terms as shown in (4). Since the last term does not depend on $\\pi$, intuitively this exploration bonus encourages the exploration of $\\pi$ (first term), and tries to make $\\pi$ different with previous policies approximated by the replay buffer (second term). The authors provide a reasonable method to optimized the proposed objective, which can be naturally combined with state-of-the-art algorithms like SAC. \n\nCons\n\n1. Theorem 1 seems misleading. The diverse policy iteration can only guarantee the converge to the optimal policy with respect to the regularized value function, not the optimal policy of the original problem. The authors should make the definition of $\\pi^*$ clear. \n\n2. It’s hard to see the motivation of using a mixture of $q$ and $\\pi$. Could you explain more about this choice?  \n\n3. It’s worth to provide the results of SAC-div with JS divergence as it’s more similar to the proposed objective (4).\n\n4. The experiment results are not convincing enough as some important baselines are missing. For example, [1] also uses a mixture of previous polices to encourage exploration with strong theoretical guarantees. I believe this is closely related to the proposed algorithms. \nAlso, the experiment results are not very promising compared with the baseline algorithms based on SAC. \n\n[1] Hazan, E., Kakade, S., Singh, K. and Van Soest, A., 2019, May. Provably efficient maximum entropy exploration. In International Conference on Machine Learning (pp. 2681-2691). \n\nOther suggestions\n\nThe main idea of the proposed method is to make the current policy different with previous policies. The paper uses a nonparametric method  (2) to approximate the previous policies. I think it’s also worth to try parametric $q$. For example, $q$ could be learned by fitting the replay buffer, or use a moving average of previous policies. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Diversity Actor-Critic: Sample-Aware Entropy Regularization for Sample-Efficient Exploration", "review": "This paper proposes diversity actor-critic (DAC) for exploration in reinforcement learning. The main idea of the proposed algorithm is to take advantage of the previous sample distribution from the replay buffer for sample-efficient exploration. The authors provide convergence analysis of DAC and conduct empirical investigations on several benchmarks. \n\nPros \n\nThe idea of using previous sample distribution from the replay buffer for better exploration seems interesting. The proposed exploration bonus $\\mathcal{H}(q^{\\pi, \\alpha}_{\\text{target}})$ can be decomposed into three terms as shown in (4). Since the last term does not depend on $\\pi$, intuitively this exploration bonus encourages the exploration of $\\pi$ (first term), and tries to make $\\pi$ different with previous policies approximated by the replay buffer (second term). The authors provide a reasonable method to optimized the proposed objective, which can be naturally combined with state-of-the-art algorithms like SAC. \n\nCons\n\n1. Theorem 1 seems misleading. The diverse policy iteration can only guarantee the converge to the optimal policy with respect to the regularized value function, not the optimal policy of the original problem. The authors should make the definition of $\\pi^*$ clear. \n\n2. It’s hard to see the motivation of using a mixture of $q$ and $\\pi$. Could you explain more about this choice?  \n\n3. It’s worth to provide the results of SAC-div with JS divergence as it’s more similar to the proposed objective (4).\n\n4. The experiment results are not convincing enough as some important baselines are missing. For example, [1] also uses a mixture of previous polices to encourage exploration with strong theoretical guarantees. I believe this is closely related to the proposed algorithms. \nAlso, the experiment results are not very promising compared with the baseline algorithms based on SAC. \n\n[1] Hazan, E., Kakade, S., Singh, K. and Van Soest, A., 2019, May. Provably efficient maximum entropy exploration. In International Conference on Machine Learning (pp. 2681-2691). \n\nOther suggestions\n\nThe main idea of the proposed method is to make the current policy different with previous policies. The paper uses a nonparametric method  (2) to approximate the previous policies. I think it’s also worth to try parametric $q$. For example, $q$ could be learned by fitting the replay buffer, or use a moving average of previous policies. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603963591905}, {"id": "hEe7fmcWT3y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1185/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n\nThis paper proposes a novel exploration method in off-policy learning. Compared to previous methods which do not take care into account the distribution of the samples in the replay buffer, the proposed method maximizes the entropy of the mixture of the policy distribution and the distribution of the samples in the replay buffer, hereby making exploration efficient.\n\nReasons for score\n\nI vote for accepting the paper. The paper proposes an intuitive and efficient exploration method that generalizes existing methods, including them as special cases. The authors provide a theoretical guarantee (Theorem 1) that the policy obtained from the iteration of evaluation and improvement under this new regime converges to the optimal policy.  The presentation is clear and concrete, and the experiments are convincing.\n\nPros\n\nThe experiment results are not limited to just showing that the proposed method achieves higher reward than state of the art methods, but they also address important questions such as \n(i) the pure exploration when rewards are assumed to be 0\n(i) the necessity of the adaptation of alpha, the parameter that controls the ratio of the current policy to the sample distribution in the target distribution.\n(ii) the effect of controlling alpha, the entropy weighting factor beta, and the control coefficient c (required for adapting alpha), and also, the robustness of the proposed method to these parameters.\n\nThe authors have stated the experiment details clearly and the results are convincing.\n\nCons\n\nThe methodology part in Section 3 and 4 could be improved. Some notations are confusing.\n(a) In Section 3, the policy \\pi is defined as a function from S to A. It looks like it is a  fixed function over time.\n(b) An explanation on the definition of J_{pi 1}(pi 2) would be helpful,e.g.,  J_{pi 1}(pi 2) is value of J(pi_2) computed under pi_1.\n\nMinor Comments\n\nIt would be good to add the line of SAC and SAC-Div in Figure 5 (c ) to show that the performance of DAC with adaptive alpha is robust to control coefficient c. For now, one has to go back to Figure 4 (b) to check that most of the case (when c is not 0), DAC with adaptive alpha performs better than SAC and SAC-Div.  \n\nIn Section 6 in the 5th line, J(\\pi) should be specified as “J(\\pi) in (1)”. It is done in the next sentence, but I prefer that it is done when it first appears. It was confusing\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Efficient exploration for offline policy learning", "review": "Summary\n\nThis paper proposes a novel exploration method in off-policy learning. Compared to previous methods which do not take care into account the distribution of the samples in the replay buffer, the proposed method maximizes the entropy of the mixture of the policy distribution and the distribution of the samples in the replay buffer, hereby making exploration efficient.\n\nReasons for score\n\nI vote for accepting the paper. The paper proposes an intuitive and efficient exploration method that generalizes existing methods, including them as special cases. The authors provide a theoretical guarantee (Theorem 1) that the policy obtained from the iteration of evaluation and improvement under this new regime converges to the optimal policy.  The presentation is clear and concrete, and the experiments are convincing.\n\nPros\n\nThe experiment results are not limited to just showing that the proposed method achieves higher reward than state of the art methods, but they also address important questions such as \n(i) the pure exploration when rewards are assumed to be 0\n(i) the necessity of the adaptation of alpha, the parameter that controls the ratio of the current policy to the sample distribution in the target distribution.\n(ii) the effect of controlling alpha, the entropy weighting factor beta, and the control coefficient c (required for adapting alpha), and also, the robustness of the proposed method to these parameters.\n\nThe authors have stated the experiment details clearly and the results are convincing.\n\nCons\n\nThe methodology part in Section 3 and 4 could be improved. Some notations are confusing.\n(a) In Section 3, the policy \\pi is defined as a function from S to A. It looks like it is a  fixed function over time.\n(b) An explanation on the definition of J_{pi 1}(pi 2) would be helpful,e.g.,  J_{pi 1}(pi 2) is value of J(pi_2) computed under pi_1.\n\nMinor Comments\n\nIt would be good to add the line of SAC and SAC-Div in Figure 5 (c ) to show that the performance of DAC with adaptive alpha is robust to control coefficient c. For now, one has to go back to Figure 4 (b) to check that most of the case (when c is not 0), DAC with adaptive alpha performs better than SAC and SAC-Div.  \n\nIn Section 6 in the 5th line, J(\\pi) should be specified as “J(\\pi) in (1)”. It is done in the next sentence, but I prefer that it is done when it first appears. It was confusing\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603903367403}, {"id": "k4QaQQz9Cdn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1185/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary\n\nThe paper proposes DAC, an actor-critic method exploiting the replay buffer to do policy entropy regularisation. The main idea of DAC is to use the data from the replay buffer to induce a distribution  $q(\\cdot, s_t)$ and replace the entropy part of the Soft Actor-Critic objective with a convex combination of $q$ and $\\pi$. This results positively on exploration properties and leads to sample-efficiency gains on some of the considered MuJoCo benchmarks.\n\n### Pros\n- Formulating the diversity using the entropy of the replay buffer frequences is an interesting idea.\n- Using the convex combination of $q$ and $\\pi$ for entropy regularisation is a nice way of generalising SAC for the considered purpose.\n- The paper shows the convergence of their method to an optimal policy and derives a surrogate objective whose gradient direction coincides with the original one, but which can be practically used. (However, I have not checked the proofs which are in the appendix).\n\n### Cons\n- It is not clear, what is the problem the paper tackles. Is it exploration? Is it a generic RL setup? What kind of problems is DAC good for?\n- If DAC is for improving exploration, then it should be compared with other exploration methods, not with vanilla SAC. Comparison with RND should not be in the appendix and there should be more details on this. Related work in this case should have a paragraph on exploration methods in RL.\n- The paper is based on assumptions not challenged/tested by the authors, e.g. policy entropy regularisation is inefficient, because it does not take the distribution of the samples into account.\n- The paper focuses more on the technical details of the solution rather than justifying the assumptions and making the research question clear.\n\n### Reasoning behind the score\n\nI believe, the paper has a great potential. However, at the moment I vote for rejection. The paper has to have a clear research question and its motivation. This should define the experimental part of the work. Lack of a clear positioning makes it unclear if the baselines of the experimental sections are the right ones and whether the claims have been properly supported by the results.\n\n### Questions to the authors\n- Can you formulate the exact problem you are solving?\n- How can you justify the claim that 'entropy regularization is sample inefficient in off-policy learning since it does not take the distribution of previous samples stored in the replay buffer into account.\n- \"it is preferable that the old sample distribution in the replay buffer is uniformly distributed\". Why is it true? Doesn't prioritized experience replay refute this claim?\n- You define $\\beta$ in Equation 1 in $(0, \\infty)$, can it really be infinite?\n- \"The rationale behind this is that it is preferable to have as diverse actions stored in the replay buffer as possible for better Q estimation in off-policy learning.\" What are the assumptions for this? Do you care more about better Q estimates or finding an better policy faster? How can you support your rationale?\n- In section 4.1. you define the target distribution as a convex combination of $\\pi$ and $q$. You assume that the buffer is generated by $q$. Does such a policy always exist? What are the assumptions for this?\n- You prove the convergence of your algorithm (I did not check the proof in the appendix), what are the assumptions for which the convergence is guaranteed?\n- Why do you use sparse/delayed MuJoCo benchmarks, but not the original ones?\n- The variance across different seeds seems to be huge for your method (as well as for the others). What do you think is the reason behind this? This also happens for the pure exploration task in 6.1, why do you think it happens?\n- For the adaptive $\\alpha$ case, you restrict the range of possible values, what is the reasoning behind the left boundary?\n- I think your paper can find an important application in Imitation Learning or off-line RL. Have you considered this? Are you aware of works which do something similar in those subfields?\n\n### Additional feedback not affecting the score\n- \"Reinforcement learning aims to maximize the discounted sum of rewards...'. Should be 'expected discounted sum'.\n- There should be a distribution over initial states under the expectation sign in 3.1.\n- 'A is the continuous action space'. This is not true for the general MDP definition, specify that this is specific for your paper.\n- Section 3.1, a policy is a mapping from states to distribution over actions, not to actions.\n- In off-policy, we can learn from any other samples, not only from 'previous samples' from our policy.\n- typo \"propoed\" at the bottom of page 4.\n- Equation 9 does not have a left hand side.\n- DAC acronym has been used in RL. I would choose a different one to avoid confusion.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting, but poorly motivated and positioned", "review": "### Summary\n\nThe paper proposes DAC, an actor-critic method exploiting the replay buffer to do policy entropy regularisation. The main idea of DAC is to use the data from the replay buffer to induce a distribution  $q(\\cdot, s_t)$ and replace the entropy part of the Soft Actor-Critic objective with a convex combination of $q$ and $\\pi$. This results positively on exploration properties and leads to sample-efficiency gains on some of the considered MuJoCo benchmarks.\n\n### Pros\n- Formulating the diversity using the entropy of the replay buffer frequences is an interesting idea.\n- Using the convex combination of $q$ and $\\pi$ for entropy regularisation is a nice way of generalising SAC for the considered purpose.\n- The paper shows the convergence of their method to an optimal policy and derives a surrogate objective whose gradient direction coincides with the original one, but which can be practically used. (However, I have not checked the proofs which are in the appendix).\n\n### Cons\n- It is not clear, what is the problem the paper tackles. Is it exploration? Is it a generic RL setup? What kind of problems is DAC good for?\n- If DAC is for improving exploration, then it should be compared with other exploration methods, not with vanilla SAC. Comparison with RND should not be in the appendix and there should be more details on this. Related work in this case should have a paragraph on exploration methods in RL.\n- The paper is based on assumptions not challenged/tested by the authors, e.g. policy entropy regularisation is inefficient, because it does not take the distribution of the samples into account.\n- The paper focuses more on the technical details of the solution rather than justifying the assumptions and making the research question clear.\n\n### Reasoning behind the score\n\nI believe, the paper has a great potential. However, at the moment I vote for rejection. The paper has to have a clear research question and its motivation. This should define the experimental part of the work. Lack of a clear positioning makes it unclear if the baselines of the experimental sections are the right ones and whether the claims have been properly supported by the results.\n\n### Questions to the authors\n- Can you formulate the exact problem you are solving?\n- How can you justify the claim that 'entropy regularization is sample inefficient in off-policy learning since it does not take the distribution of previous samples stored in the replay buffer into account.\n- \"it is preferable that the old sample distribution in the replay buffer is uniformly distributed\". Why is it true? Doesn't prioritized experience replay refute this claim?\n- You define $\\beta$ in Equation 1 in $(0, \\infty)$, can it really be infinite?\n- \"The rationale behind this is that it is preferable to have as diverse actions stored in the replay buffer as possible for better Q estimation in off-policy learning.\" What are the assumptions for this? Do you care more about better Q estimates or finding an better policy faster? How can you support your rationale?\n- In section 4.1. you define the target distribution as a convex combination of $\\pi$ and $q$. You assume that the buffer is generated by $q$. Does such a policy always exist? What are the assumptions for this?\n- You prove the convergence of your algorithm (I did not check the proof in the appendix), what are the assumptions for which the convergence is guaranteed?\n- Why do you use sparse/delayed MuJoCo benchmarks, but not the original ones?\n- The variance across different seeds seems to be huge for your method (as well as for the others). What do you think is the reason behind this? This also happens for the pure exploration task in 6.1, why do you think it happens?\n- For the adaptive $\\alpha$ case, you restrict the range of possible values, what is the reasoning behind the left boundary?\n- I think your paper can find an important application in Imitation Learning or off-line RL. Have you considered this? Are you aware of works which do something similar in those subfields?\n\n### Additional feedback not affecting the score\n- \"Reinforcement learning aims to maximize the discounted sum of rewards...'. Should be 'expected discounted sum'.\n- There should be a distribution over initial states under the expectation sign in 3.1.\n- 'A is the continuous action space'. This is not true for the general MDP definition, specify that this is specific for your paper.\n- Section 3.1, a policy is a mapping from states to distribution over actions, not to actions.\n- In off-policy, we can learn from any other samples, not only from 'previous samples' from our policy.\n- typo \"propoed\" at the bottom of page 4.\n- Equation 9 does not have a left hand side.\n- DAC acronym has been used in RL. I would choose a different one to avoid confusion.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603886743364}], "openreview_url": "https://openreview.net/forum?id=7qmQNB6Wn_B", "arxiv_id": "2006.01419", "paper_pdf": "papers/7qmQNB6Wn_B.pdf", "paper_pdf_sha256": "95b008894bc06eb7c06ee22a414224b5c2141f25039856f3f0c7f5500b8cafcc", "paper_pdf_bytes": 5198691, "paper_pdf_source": "openreview", "code_url": "https://github.com/seungyulhan/dac", "code_repository": "seungyulhan/dac", "code_commit": "3595d8724ac47486812c7067bce3459f65268af7", "code_archive": "repos/7qmQNB6Wn_B.zip", "code_archive_sha256": "945a0228cdd566c07bbfb407f5ff217e939fb6680f9b72c38ed0673313100ebe", "code_archive_bytes": 188305, "code_file_count": 36, "code_extensions": {".py": 36}, "github_disk_usage_kb": 180, "github_languages": {"Python": 103683}, "github_archived": false, "github_pushed_at": "2022-08-17T04:17:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/diversity-actor-critic-sample-aware-entropy"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HklliySFDS", "year": 2020, "status": "rejected", "title": "Continual Learning with Gated Incremental Memories for Sequential Data Processing", "authors": ["Andrea Cossu", "Antonio Carta", "Davide Bacciu"], "authorids": ["cossu48@gmail.com", "antonio.carta@di.unipi.it", "bacciu@di.unipi.it"], "authors_source": "OpenReview API", "abstract": "The ability to learn over changing task distributions without forgetting previous knowledge, also known as continual learning, is a key enabler for scalable and trustworthy deployments of adaptive solutions. While the importance of continual learning is largely acknowledged in machine vision and reinforcement learning problems, this is mostly under-documented for sequence processing tasks. This work focuses on characterizing and quantitatively assessing the impact of catastrophic forgetting and task interference when dealing with sequential data in recurrent neural networks. We also introduce a general architecture, named Gated Incremental Memory, for augmenting recurrent models with continual learning skills, whose effectiveness is demonstrated through the benchmarks introduced in this paper.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1xv8YTatr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1902/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposed an interesting continual learning approach for sequential data processing with recurrent neural network architecture. \nThe authors provide a general application on sequential data for continual learning, and show their proposed model outperforms baseline.\n\nIt is natural that their naive baseline shows poor performance since they do not consider any continual learning issues like the catastrophic forgetting problem. Then, I hesitate to evaluate the model in terms of performance. In that sense, it would be much crucial to show more meaningful ablation studies and analysis for proposed model. However, there is a few of thing about them. \n\nThen, I decide to give a lower score that even the authors suggest that the main contribution is a definition of problem setting. It requires more detailed and sophisticated analysis.\n          \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper proposed an interesting continual learning approach for sequential data processing with recurrent neural network architecture. \nThe authors provide a general application on sequential data for continual learning, and show their proposed model outperforms baseline.\n\nIt is natural that their naive baseline shows poor performance since they do not consider any continual learning issues like the catastrophic forgetting problem. Then, I hesitate to evaluate the model in terms of performance. In that sense, it would be much crucial to show more meaningful ablation studies and analysis for proposed model. However, there is a few of thing about them. \n\nThen, I decide to give a lower score that even the authors suggest that the main contribution is a definition of problem setting. It requires more detailed and sophisticated analysis.\n          \n"}, "tcdate": 1571834190919}, {"id": "r1esT_GTKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1902/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The goal of this work is to best understand the performance and benchmarking of continual learning algorithms when applied to sequential data processing problems like language or sequence data sets. The contributions of the paper are 3 fold - new benchmarks for CL with sequential data for RNN processing, new architecture introduced for more effective processing and a thorough empirical evaluation. \n\nIntroduction: \nI think a little more insight into why the sequential data processing CL scenario is any different than the vision scenario would be quite helpful. Specifically, it would be quite impactful to tell us more about what the additional challenges with RNNs for CL vs feedforward for CL are in the intro. \n\nThe paper is written as if the benchmark is the main contribution and the architecture improvement is just a delta on top of this, but it gets confusing when the methods section starts off with just directly stating the new architecture. \n\nThe algorithm seems like a straightforward combination of recurrent progressive nets and gated autoencoders for CL. Can the authors provide more justification if that is the contribution or there is more to the insight than has been previously suggested in prior work?\n\nFigure 1 has a very uninformative caption. It also doesn’t show how modules feed into one another properly. \n\nThe motivation for why one needs GIM after one already has A-LSTM or A-LMN is not very clear?\n\nOverall the contribution does seem a bit incremental based on prior work and the description lacks enough detail to properly indicate why this is a very important contribution?\n\nExperiments:\nWhat does it mean to be application agnostic but restricted to particular datasets and losses? This doesn’t quite parse to me. \n\nThe description of the tasks is very informal and hard to follow. It’s not clear what exactly the tasks and datasets look like \n\n“using morehidden units can bridge this gap” -> why not just do it? Its a benchmark after all. \n\nOverall the task descriptions should be in a separate section where the setup is described in a lot of detail and motivated properly. \n\nThe results in the experiments section are very hard to parse. The captions need much more detail for eg Table 2. \n\nCould we also possibly have more baselines from continual learning? For instance EWC (Kirkpatrick) or generative replay might be competitive baselines. \n\nOverall I think that the GIM and A-LMN and A-LSTM methods are reasonable although somewhat incremental. But the proposed benchmarks are pretty unclear and the results are a bit hard to really interpret well. It would also be important to run comparisons with more baselines and to provide more ablation/analysis experiments to really see the benefit of GIM/A-LMN or A-LSTM. I also think that the task descriptions should be much earlier in the paper and desribed in much more rigorous detail. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The goal of this work is to best understand the performance and benchmarking of continual learning algorithms when applied to sequential data processing problems like language or sequence data sets. The contributions of the paper are 3 fold - new benchmarks for CL with sequential data for RNN processing, new architecture introduced for more effective processing and a thorough empirical evaluation. \n\nIntroduction: \nI think a little more insight into why the sequential data processing CL scenario is any different than the vision scenario would be quite helpful. Specifically, it would be quite impactful to tell us more about what the additional challenges with RNNs for CL vs feedforward for CL are in the intro. \n\nThe paper is written as if the benchmark is the main contribution and the architecture improvement is just a delta on top of this, but it gets confusing when the methods section starts off with just directly stating the new architecture. \n\nThe algorithm seems like a straightforward combination of recurrent progressive nets and gated autoencoders for CL. Can the authors provide more justification if that is the contribution or there is more to the insight than has been previously suggested in prior work?\n\nFigure 1 has a very uninformative caption. It also doesn’t show how modules feed into one another properly. \n\nThe motivation for why one needs GIM after one already has A-LSTM or A-LMN is not very clear?\n\nOverall the contribution does seem a bit incremental based on prior work and the description lacks enough detail to properly indicate why this is a very important contribution?\n\nExperiments:\nWhat does it mean to be application agnostic but restricted to particular datasets and losses? This doesn’t quite parse to me. \n\nThe description of the tasks is very informal and hard to follow. It’s not clear what exactly the tasks and datasets look like \n\n“using morehidden units can bridge this gap” -> why not just do it? Its a benchmark after all. \n\nOverall the task descriptions should be in a separate section where the setup is described in a lot of detail and motivated properly. \n\nThe results in the experiments section are very hard to parse. The captions need much more detail for eg Table 2. \n\nCould we also possibly have more baselines from continual learning? For instance EWC (Kirkpatrick) or generative replay might be competitive baselines. \n\nOverall I think that the GIM and A-LMN and A-LSTM methods are reasonable although somewhat incremental. But the proposed benchmarks are pretty unclear and the results are a bit hard to really interpret well. It would also be important to run comparisons with more baselines and to provide more ablation/analysis experiments to really see the benefit of GIM/A-LMN or A-LSTM. I also think that the task descriptions should be much earlier in the paper and desribed in much more rigorous detail. \n"}, "tcdate": 1571788995269}, {"id": "r1lsUfWpKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1902/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nIn this paper, the authors propose a new method to apply continual learning on sequential data. The model is constructed by combining an Autoencoder and LSTM/LMN for each task. The experiments on several datasets show the proposed model outperforms basic LSTM/LMN.\n\n\nStrength:\n\n+ Sequential data widely exist in the real world, e.g., text, health records. Thus, It is interesting to see that continual learning is used in sequential data. \n\n+ The motivation of the proposed model is clear. The authors save the learned knowledge in the hidden representation of LSTM/LMN.\n\nWeakness:\n- In this paper, the model size linearly increases since the number of LSTM/LMN and AE increases when a new task comes in. Thus, if the number of tasks is too large, the model size is quite big. In traditional continual learning settings, researchers may not always increase the model size for overcoming catastrophic forgetting. For example, if task 1 and task 2 sample from the same distribution, they can share the same LSTM/LMN and AE. Thus, it would be better if the authors can consider how to reduce the model size in the future version.\n\n- In the experiments, the authors only compare the proposed model with simple LSTM or LMN. However, most continual learning methods can still be applied in this scenario, at least regularization based methods [1,2] can be simply applied in this scenario. The authors may need to compare the proposed method with them in the future version.\n\n- It is better to compare it with a larger dataset. For example, in the natural language processing field, we can regard sentiment analysis on one language as one task. Then, we can construct the continual learning dataset for sentiment analysis.\n\nMinor Comments:\nIt is better to improve Figure 3 by adding the x-axis label and y-axis label.\n\n\n[1] Kirkpatrick, James, et al. \"Overcoming catastrophic forgetting in neural networks.\" Proceedings of the national academy of sciences 114.13 (2017): 3521-3526.\n[2] Zenke, Friedemann, Ben Poole, and Surya Ganguli. \"Continual learning through synaptic intelligence.\" Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "Summary:\n\nIn this paper, the authors propose a new method to apply continual learning on sequential data. The model is constructed by combining an Autoencoder and LSTM/LMN for each task. The experiments on several datasets show the proposed model outperforms basic LSTM/LMN.\n\n\nStrength:\n\n+ Sequential data widely exist in the real world, e.g., text, health records. Thus, It is interesting to see that continual learning is used in sequential data. \n\n+ The motivation of the proposed model is clear. The authors save the learned knowledge in the hidden representation of LSTM/LMN.\n\nWeakness:\n- In this paper, the model size linearly increases since the number of LSTM/LMN and AE increases when a new task comes in. Thus, if the number of tasks is too large, the model size is quite big. In traditional continual learning settings, researchers may not always increase the model size for overcoming catastrophic forgetting. For example, if task 1 and task 2 sample from the same distribution, they can share the same LSTM/LMN and AE. Thus, it would be better if the authors can consider how to reduce the model size in the future version.\n\n- In the experiments, the authors only compare the proposed model with simple LSTM or LMN. However, most continual learning methods can still be applied in this scenario, at least regularization based methods [1,2] can be simply applied in this scenario. The authors may need to compare the proposed method with them in the future version.\n\n- It is better to compare it with a larger dataset. For example, in the natural language processing field, we can regard sentiment analysis on one language as one task. Then, we can construct the continual learning dataset for sentiment analysis.\n\nMinor Comments:\nIt is better to improve Figure 3 by adding the x-axis label and y-axis label.\n\n\n[1] Kirkpatrick, James, et al. \"Overcoming catastrophic forgetting in neural networks.\" Proceedings of the national academy of sciences 114.13 (2017): 3521-3526.\n[2] Zenke, Friedemann, Ben Poole, and Surya Ganguli. \"Continual learning through synaptic intelligence.\" Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017."}, "tcdate": 1571783250659}], "openreview_url": "https://openreview.net/forum?id=HklliySFDS", "arxiv_id": "2004.04077", "paper_pdf": "papers/HklliySFDS.pdf", "paper_pdf_sha256": "9c13a3a03a54690dc8556b27004b38ef002231d99b468a620c27bb82bf854036", "paper_pdf_bytes": 331239, "paper_pdf_source": "openreview", "code_url": "https://github.com/AndreaCossu/ContinualLearning-SequentialProcessing", "code_repository": "AndreaCossu/ContinualLearning-SequentialProcessing", "code_commit": "d1b1b6a99a0f5b294d37bca5e82cbbd80ad7722d", "code_archive": "repos/HklliySFDS.zip", "code_archive_sha256": "a639ee189c1ff0ef0d78f5d7ed79402516abe9ca6dc17a1cffa3884a8e6890ec", "code_archive_bytes": 176926, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 198, "github_languages": {"Python": 121707}, "github_archived": false, "github_pushed_at": "2021-10-13T07:35:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continual-learning-with-gated-incremental-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJMnG2C9YX", "year": 2019, "status": "rejected", "title": "Complementary-label learning for arbitrary losses and models", "authors": ["Takashi Ishida", "Gang Niu", "Aditya Krishna Menon", "Masashi Sugiyama"], "authorids": ["ishida@ms.k.u-tokyo.ac.jp", "gang.niu@riken.jp", "aditya.menon@anu.edu.au", "sugi@k.u-tokyo.ac.jp"], "authors_source": "OpenReview API", "abstract": "In contrast to the standard classification paradigm where the true (or possibly noisy) class is given to each training pattern, complementary-label learning only uses training patterns each equipped with a complementary label. This only specifies one of the classes that the pattern does not belong to. The seminal paper on complementary-label learning proposed an unbiased estimator of the classification risk that can be computed only from complementarily labeled data. How- ever, it required a restrictive condition on the loss functions, making it impossible to use popular losses such as the softmax cross-entropy loss. Recently, another formulation with the softmax cross-entropy loss was proposed with consistency guarantee. However, this formulation does not explicitly involve a risk estimator. Thus model/hyper-parameter selection is not possible by cross-validation— we may need additional ordinarily labeled data for validation purposes, which is not available in the current setup. In this paper, we give a novel general framework of complementary-label learning, and derive an unbiased risk estimator for arbitrary losses and models. We further improve the risk estimator by non-negative correction and demonstrate its superiority through experiments.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "SkeJ7R_m6X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1305/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\n- The authors consider an interesting problem of learning from complementary labels\n- They propose an approach that, assuming that the complementary label is selected uniformly at random, provides an unbiased estimate for any loss function, which is an improvement over the previous work. \n- Experiments show promising results for modifications of the proposed estimate\n\nCons:\n- Having an unbiased estimate doesn't imply that its minimisation is a successful learning strategy. Indeed, the authors show that minimising their original estimate for the cross-entropy loss leads to overfitting. While the authors attribute this behaviour to the fact that the estimate can be negative, I believe the loss being negative is not problem per se (for example, substituting 0/1 loss with -100/-99 loss would not change the learning; similarly, this is not a problem for the losses considered in [Ishida'17]). I would rather attribute the problem to the fact that the proposed estimate is unbounded from below and there are no generalisation guarantees for it. Indeed, assuming there exists a training example that appears in the training set only once, with one complementary label, estimate (8) can be made arbitrary small by just training to predict probability 0 for the provided complementary label on that example ( and any non-zero probability for other classes). \n- to cope with the above mentioned problem, the authors propose two heuristic-based modifications of the estimate, which are potentially biased. This weakens the initial motivation for finding an unbiased estimate and shifts the focus towards the experimental evaluation\n- one of the mentioned motivations for unbiased estimates - being able to perform model selection on complementary labeled validation set - is not illustrated in the experiments\n\nQuestions:\n- I believe 1/(K-1) normalisation factor in (5) is not needed\n- there seems to be a mistake in (9) (and its modifications later on) - I would expect either the subscript $j$ of the probability distribution in the last summand to be exchanged with $k$ in the loss, or a factor $\\pi_j/\\pi_k$ added\n- also, I think there are some mistakes in subscripts in (11)\n- what loss is the method from [Ishida'17] optimising in the experiments?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting setting, but problems with the original estimator and limited experimental evaluation weaken the claims", "review": "Pros:\n- The authors consider an interesting problem of learning from complementary labels\n- They propose an approach that, assuming that the complementary label is selected uniformly at random, provides an unbiased estimate for any loss function, which is an improvement over the previous work. \n- Experiments show promising results for modifications of the proposed estimate\n\nCons:\n- Having an unbiased estimate doesn't imply that its minimisation is a successful learning strategy. Indeed, the authors show that minimising their original estimate for the cross-entropy loss leads to overfitting. While the authors attribute this behaviour to the fact that the estimate can be negative, I believe the loss being negative is not problem per se (for example, substituting 0/1 loss with -100/-99 loss would not change the learning; similarly, this is not a problem for the losses considered in [Ishida'17]). I would rather attribute the problem to the fact that the proposed estimate is unbounded from below and there are no generalisation guarantees for it. Indeed, assuming there exists a training example that appears in the training set only once, with one complementary label, estimate (8) can be made arbitrary small by just training to predict probability 0 for the provided complementary label on that example ( and any non-zero probability for other classes). \n- to cope with the above mentioned problem, the authors propose two heuristic-based modifications of the estimate, which are potentially biased. This weakens the initial motivation for finding an unbiased estimate and shifts the focus towards the experimental evaluation\n- one of the mentioned motivations for unbiased estimates - being able to perform model selection on complementary labeled validation set - is not illustrated in the experiments\n\nQuestions:\n- I believe 1/(K-1) normalisation factor in (5) is not needed\n- there seems to be a mistake in (9) (and its modifications later on) - I would expect either the subscript $j$ of the probability distribution in the last summand to be exchanged with $k$ in the loss, or a factor $\\pi_j/\\pi_k$ added\n- also, I think there are some mistakes in subscripts in (11)\n- what loss is the method from [Ishida'17] optimising in the experiments?", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541799447112}, {"id": "B1gDnQhq3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1305/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an improved approach to the \"complementary-label\" form of weak supervision, in which a label that is *not* the true label is marked. Specifically, this paper proposes an unbiased estimator that accepts arbitrary loss functions and models. Noting that this proposed estimator can suffer from overfitting due to unbounded negative loss, a lower-bounded estimator is proposed. Experiments are then performed on several image classification datasets.\n\nPros:\n- This paper addresses a creative form of weak supervision, proposed by prior work, in which labels that are *not* the true label are labeled, in a clear fashion.\n\n- The first proposed estimator is unbiased, as shown by a proof, and accepts arbitrary losses, an improvement over prior approaches\n\n- The overall presentation is clear and clean\n\nCons:\n- One of the main claims of the paper is the proposal of an unbiased estimator. However, this estimator then does not seem to work well enough due to degenerate negative loss.  So then a modified version is proposed- which does not appear to be unbiased?  Either way, no assertion or proof of it being unbiased is given.  So then presumably this also reverses the claim of being able to cross-validate?  This seems like a major weakening of the paper's contributions\n\n- Since the unbiased estimator does not appear to work well, two implementations of a corrected one are proposed, using heuristic approaches without explicit theoretical guarantees.  This shifts the burden to the experimental studies.  These are somewhat thorough, but not extremely so: for example, one set of hyperparameters were used for all of the methods?  This seems like it could implicitly handicap / favor some over others?\n\n- The proposed estimator is based on the assumption that the probability of classes in the complement set (the set of labels other than the one marked as incorrect) is uniformly distributed (e.g. see beginning of Proof of thm 1).  However, this seems like a potentially naive assumption. Indeed, in the related work section, it is mentioned that work in 2018 already considered the case where this uniformity assumption does not hold.\n\n- More broadly, but following from the above: The paper does not provide any real world examples, real or hypothetical, to give the reader an idea of whether the above uniformity assumption---or really any of these assumptions---are well-motivated or empirically justified.  At the bottom of page 3 in the related work, a concrete application used in prior work is mentioned---where crowd workers are shown single labels and vote Y/N, leading to a mix of standard (if Y) and complement-labeled (if N) data---however this mixed setting is not considered explicitly in this paper.  So, how is the reader supposed to get any idea of whether the assumed setup is motivated or justified?  The experiments do not provide this, because the complementary labels are synthetically generated according to the model assumed in the paper.  Additionally, it is briefly mentioned that collecting complementary labeled data is faster, but again no concrete examples are given to support this.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clear paper on interesting setup, but claims are undermined by issues with first estimator + lack of motivation for assumptions", "review": "This paper proposes an improved approach to the \"complementary-label\" form of weak supervision, in which a label that is *not* the true label is marked. Specifically, this paper proposes an unbiased estimator that accepts arbitrary loss functions and models. Noting that this proposed estimator can suffer from overfitting due to unbounded negative loss, a lower-bounded estimator is proposed. Experiments are then performed on several image classification datasets.\n\nPros:\n- This paper addresses a creative form of weak supervision, proposed by prior work, in which labels that are *not* the true label are labeled, in a clear fashion.\n\n- The first proposed estimator is unbiased, as shown by a proof, and accepts arbitrary losses, an improvement over prior approaches\n\n- The overall presentation is clear and clean\n\nCons:\n- One of the main claims of the paper is the proposal of an unbiased estimator. However, this estimator then does not seem to work well enough due to degenerate negative loss.  So then a modified version is proposed- which does not appear to be unbiased?  Either way, no assertion or proof of it being unbiased is given.  So then presumably this also reverses the claim of being able to cross-validate?  This seems like a major weakening of the paper's contributions\n\n- Since the unbiased estimator does not appear to work well, two implementations of a corrected one are proposed, using heuristic approaches without explicit theoretical guarantees.  This shifts the burden to the experimental studies.  These are somewhat thorough, but not extremely so: for example, one set of hyperparameters were used for all of the methods?  This seems like it could implicitly handicap / favor some over others?\n\n- The proposed estimator is based on the assumption that the probability of classes in the complement set (the set of labels other than the one marked as incorrect) is uniformly distributed (e.g. see beginning of Proof of thm 1).  However, this seems like a potentially naive assumption. Indeed, in the related work section, it is mentioned that work in 2018 already considered the case where this uniformity assumption does not hold.\n\n- More broadly, but following from the above: The paper does not provide any real world examples, real or hypothetical, to give the reader an idea of whether the above uniformity assumption---or really any of these assumptions---are well-motivated or empirically justified.  At the bottom of page 3 in the related work, a concrete application used in prior work is mentioned---where crowd workers are shown single labels and vote Y/N, leading to a mix of standard (if Y) and complement-labeled (if N) data---however this mixed setting is not considered explicitly in this paper.  So, how is the reader supposed to get any idea of whether the assumed setup is motivated or justified?  The experiments do not provide this, because the complementary labels are synthetically generated according to the model assumed in the paper.  Additionally, it is briefly mentioned that collecting complementary labeled data is faster, but again no concrete examples are given to support this.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541223343307}, {"id": "rJedhpHq3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1305/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "pros:\n\n- Clearly written and sound paper.\n- Addresses interesting problem. \n- Improves existing methods used for this learning scenario.  \n\ncons:\n\n- The core contribution is a special case of previously published more general framework which is not cited in the paper.\n\nIt is clearly written paper with a good motivation. The major problem is that the core contribution, namely, the risk reformulation in Theorem 1 and the derived loss (6), are special cases of more general framework published in \n   Jesus Cid-Sueiro et al. Consistency of Losses for Learning from Weak Labels. ECML 2014.\n\nThe work of [Cid-Sueiro2014] proposes a general way how to construct losses for learning from weak labels. They require that the distribution of weak labels is a linear transformation of the true label distribution, i.e. the assumption (3) of the paper under review. According to [Cid-Sueiro2014], the loss on weak labels is constructed by $weak_loss = L*original_loss$, where $L$ is the left inversion of the \"mixing matrix\" $T$ in (3). [Cid-Sueiro2014] also shows that such weak loss is classification calibrated which implies statistical consistency of the method. \n\nLearning from complementary labels is a special case when the mixing matrix is $T=(E-I)/(K-1)$ (E is unitary matrix, I is matrix of ones, K is number of labels). In this case, the left inversion of $T$ is simply $L=- E*(K-1) + I$ and so the weak loss is $weak_loss=L*loss$ which corresponds to the loss (5) proposed in the paper under review (in fact, the loss (5) also adds a constant term (Y-2)/(Y-1) which however has no effect on the minimizer). \n\nThe novel part of the paper is the non-negative risk estimator proposed in sec 3.3 and the online optimization methods addressed in sec 3.4. These extensions, although relatively straightforward, are empirically shown to significantly improve the results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well written paper about an interesting problem. The major problem is that the core part of the contribution is a special case of previously published more general framework not cited in the paper. ", "review": "pros:\n\n- Clearly written and sound paper.\n- Addresses interesting problem. \n- Improves existing methods used for this learning scenario.  \n\ncons:\n\n- The core contribution is a special case of previously published more general framework which is not cited in the paper.\n\nIt is clearly written paper with a good motivation. The major problem is that the core contribution, namely, the risk reformulation in Theorem 1 and the derived loss (6), are special cases of more general framework published in \n   Jesus Cid-Sueiro et al. Consistency of Losses for Learning from Weak Labels. ECML 2014.\n\nThe work of [Cid-Sueiro2014] proposes a general way how to construct losses for learning from weak labels. They require that the distribution of weak labels is a linear transformation of the true label distribution, i.e. the assumption (3) of the paper under review. According to [Cid-Sueiro2014], the loss on weak labels is constructed by $weak_loss = L*original_loss$, where $L$ is the left inversion of the \"mixing matrix\" $T$ in (3). [Cid-Sueiro2014] also shows that such weak loss is classification calibrated which implies statistical consistency of the method. \n\nLearning from complementary labels is a special case when the mixing matrix is $T=(E-I)/(K-1)$ (E is unitary matrix, I is matrix of ones, K is number of labels). In this case, the left inversion of $T$ is simply $L=- E*(K-1) + I$ and so the weak loss is $weak_loss=L*loss$ which corresponds to the loss (5) proposed in the paper under review (in fact, the loss (5) also adds a constant term (Y-2)/(Y-1) which however has no effect on the minimizer). \n\nThe novel part of the paper is the non-negative risk estimator proposed in sec 3.3 and the online optimization methods addressed in sec 3.4. These extensions, although relatively straightforward, are empirically shown to significantly improve the results.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541197232476}], "openreview_url": "https://openreview.net/forum?id=SJMnG2C9YX", "arxiv_id": "1810.04327", "paper_pdf": "papers/SJMnG2C9YX.pdf", "paper_pdf_sha256": "7cdefe80564d02407c5993c74ed1cf34abf6f01d5db745ef6770d76041006673", "paper_pdf_bytes": 1861159, "paper_pdf_source": "openreview", "code_url": "https://github.com/takashiishida/comp", "code_repository": "takashiishida/comp", "code_commit": "aab868d49eefe4443865686574684253fe9b2487", "code_archive": "repos/SJMnG2C9YX.zip", "code_archive_sha256": "071487c7d6a897d3070780267d0faf0a2c54634dd4e6185952204176776365e5", "code_archive_bytes": 6265, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 3370, "github_languages": {"Python": 10504}, "github_archived": false, "github_pushed_at": "2025-08-17T02:03:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/complementary-label-learning-for-arbitrary"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GhM63V7z6v", "year": 2025, "status": "rejected", "title": "Temporal Source Recovery for Time-Series Source-Free Unsupervised Domain Adaptation", "authors": ["Yucheng Wang", "Peiliang Gong", "Min Wu", "Felix Ott", "Xiaoli Li", "Lihua Xie", "Zhenghua Chen"], "authorids": ["~Yucheng_Wang10", "~Peiliang_Gong1", "~Min_Wu2", "~Felix_Ott1", "~Xiaoli_Li1", "~Lihua_Xie2", "~Zhenghua_Chen2"], "authors_source": "OpenReview API", "abstract": "Source-Free Unsupervised Domain Adaptation (SFUDA) has gained popularity for its ability to adapt pretrained models to target domains without accessing source domains, ensuring source data privacy. While SFUDA is well-developed in visual tasks, its application to Time-Series SFUDA (TS-SFUDA) remains limited due to the challenge of transferring crucial temporal dependencies across domains. Although a few researchers begin to explore this area, they rely on specific source domain designs, which are impractical as source data owners cannot be expected to follow particular pretraining protocols. To solve this, we propose Temporal Source Recovery (TemSR), a framework that transfers temporal dependencies for effective TS-SFUDA without requiring source-specific designs. TemSR features a recovery process that leverages masking, recovery, and optimization to generate a source-like distribution with recovered source temporal dependencies. To ensure effective recovery, we further design segment-based regularization to restore local dependencies and anchor-based recovery diversity maximization to enhance the diversity of the source-like distribution. The source-like distribution is then adapted to the target domain using traditional UDA techniques. Extensive experiments across multiple TS tasks demonstrate the effectiveness of TemSR, even surpassing existing TS-SFUDA method that requires source domain designs.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "lXb2poNAol", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9789/Reviewer_gLQz"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper addresses the time-series source-free domain adaptation problem. It proposes a source recovery method that recovers target data to a source-like distribution and introduces anchor-based recovery diversity maximization to enhance the diversity of the source-like distribution.", "review_text": "This paper addresses the time-series source-free domain adaptation problem. It proposes a source recovery method that recovers target data to a source-like distribution and introduces anchor-based recovery diversity maximization to enhance the diversity of the source-like distribution.", "strengths": "1. The overall idea of source distribution recovery is reasonable.\n2. The method section is detailed and clearly presented.", "weaknesses": "1. The ablation study could be more comprehensive, as there are additional specific designs with intuitive explanations mentioned in the method section beyond the three variants in Table 4, such as the consistent segment entropy loss and different components in the ARDM loss.\n2. The two key requirements stated at the beginning of Section 3.3 are intuitive. More explanation or support from the literature would be beneficial.", "questions": "In Section 4.5, could the authors explain why minimizing the domain discrepancy between the source-like and target domains is required for domain adaptation? This point does not appear to be discussed in previous sections.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the time-series source-free domain adaptation problem. It proposes a source recovery method that recovers target data to a source-like distribution and introduces anchor-based recovery diversity maximization to enhance the diversity of the source-like distribution.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The overall idea of source distribution recovery is reasonable.\n2. The method section is detailed and clearly presented.", "weaknesses": "1. The ablation study could be more comprehensive, as there are additional specific designs with intuitive explanations mentioned in the method section beyond the three variants in Table 4, such as the consistent segment entropy loss and different components in the ARDM loss.\n2. The two key requirements stated at the beginning of Section 3.3 are intuitive. More explanation or support from the literature would be beneficial.", "questions": "In Section 4.5, could the authors explain why minimizing the domain discrepancy between the source-like and target domains is required for domain adaptation? This point does not appear to be discussed in previous sections.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730651710108}, {"id": "8EVIgYPwX9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9789/Reviewer_5jCn"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "In this paper, the authors present a temporal source recovery (TemSR) framework for source-free unsupervised domain adaptation. Specifically, a source-like distribution is generated with the source temporal dependencies recovered. A segment-based regularization and an anchor-based recovery diversity maximization are developed to enhance the source recovery. Experiments on three datasets are done to verify the proposed TemSR.", "review_text": "In this paper, the authors present a temporal source recovery (TemSR) framework for source-free unsupervised domain adaptation. Specifically, a source-like distribution is generated with the source temporal dependencies recovered. A segment-based regularization and an anchor-based recovery diversity maximization are developed to enhance the source recovery. Experiments on three datasets are done to verify the proposed TemSR.", "strengths": "1. The paper proposes to deal with a practical and meaningful adaptation scenario, that is source-free unsupervised domain adaptation.\n2. The paper presents a temporal source recovery (TemSR) framework for source-free unsupervised domain adaptation.\n3. The paper conduct experiments on three practical datasets, and demonstrate the effectiveness of the proposed TemSR  compared with given baselines.", "weaknesses": "1. One of my major concerns of the work is on the technical novelty. Using mask and then recovery has been shown the adaptation and transfer capability in generation tasks, e.g., masked autoencoder and videoMAE. The application of such an idea to source-free UDA setting makes the technical significance limited although it may not be explicitly done before.\n2. The paper misses some important video-based UDA related works. Video is a popular time-series data, and there are quite a lot of video-based UDA research. The reviewer suggests the authors to discuss them in related work section. \n3. For time-series data as stated in the setting with N channels and L time points, existing UDA works usually handle both the spatial and temporal divergence, e.g., [ref1,2]. However, the paper seems to overlook the spatial divergence and only focus on the temporal one.\n4. The experimental studies are not very convincing. The author may refer to my detailed comments on the below questions.", "questions": "1. Figure 2 is not very informative. It is inadequate to use two ellipses to represent complex temporal and spatial divergence across TS data. Moreover, what does “target distribution” mean here? Are you trying to explain why using masked target data? \n2. Regarding the recovery component, is a reconstruction loss included? Or using a pretrained reconstruction model as backbone? If not, is it better to initialize the recovery model by a reconstruction model pretrained on the target unsupervised data? \n3. Regarding the segment regularization line 248-250, low entropy difference between segments does not indicate the smooth dependencies among segments. The authors may need to present ablation studies on this term (Eq. (2)) to show its effectiveness. Moreover, what does {C,E,L,R} mean? As in Eq. (2), (k, s) belongs to {C,E,L,R}, is it trying to minimize the entropy difference between arbitrary segment pairs? \n4. Can you elaborate more on the test flow? Are you using the pretrained source classifier for final prediction?\n5. Regarding the experiments parts: (1) more complex video benchmark datasets are encouraged to be tested; (2) why using the same CNN backbone? For different tasks, different backbones should be used. For instance, for action recognition tasks, there are more advanced backbones, e.g., I3D, videomae; (3) Model details should be presented in the main contents. Please elaborate more on the recovery model as it is the main contribution. Are you using LSTM or Bi-LSTM? Why not considering Transformer structure? (4) The results are not very convincing to show the performance superiority of TemSR as TemSR is the winner in 2 of 5, 1 of 5, and 3 of 5 tasks in the three datasets. (5) Sensitivity analysis shows that the optimal hyper-parameter value varies across tasks. It is not convincing to conclude the best value range is 1 to 10 by only using 3 datasets. The hyper-parameter generalization ability of the proposed TemSR is also an issue when facing new tasks or datasets. Moreover, what is “EEG” in figures 4,5,6?\n\n[ref1] Contrast and mix: Temporal contrastive video domain adaptation with background mixing\n\n[ref2] Unsupervised video domain adaptation for action recognition: A disentanglement perspective", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors present a temporal source recovery (TemSR) framework for source-free unsupervised domain adaptation. Specifically, a source-like distribution is generated with the source temporal dependencies recovered. A segment-based regularization and an anchor-based recovery diversity maximization are developed to enhance the source recovery. Experiments on three datasets are done to verify the proposed TemSR.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper proposes to deal with a practical and meaningful adaptation scenario, that is source-free unsupervised domain adaptation.\n2. The paper presents a temporal source recovery (TemSR) framework for source-free unsupervised domain adaptation.\n3. The paper conduct experiments on three practical datasets, and demonstrate the effectiveness of the proposed TemSR  compared with given baselines.", "weaknesses": "1. One of my major concerns of the work is on the technical novelty. Using mask and then recovery has been shown the adaptation and transfer capability in generation tasks, e.g., masked autoencoder and videoMAE. The application of such an idea to source-free UDA setting makes the technical significance limited although it may not be explicitly done before.\n2. The paper misses some important video-based UDA related works. Video is a popular time-series data, and there are quite a lot of video-based UDA research. The reviewer suggests the authors to discuss them in related work section. \n3. For time-series data as stated in the setting with N channels and L time points, existing UDA works usually handle both the spatial and temporal divergence, e.g., [ref1,2]. However, the paper seems to overlook the spatial divergence and only focus on the temporal one.\n4. The experimental studies are not very convincing. The author may refer to my detailed comments on the below questions.", "questions": "1. Figure 2 is not very informative. It is inadequate to use two ellipses to represent complex temporal and spatial divergence across TS data. Moreover, what does “target distribution” mean here? Are you trying to explain why using masked target data? \n2. Regarding the recovery component, is a reconstruction loss included? Or using a pretrained reconstruction model as backbone? If not, is it better to initialize the recovery model by a reconstruction model pretrained on the target unsupervised data? \n3. Regarding the segment regularization line 248-250, low entropy difference between segments does not indicate the smooth dependencies among segments. The authors may need to present ablation studies on this term (Eq. (2)) to show its effectiveness. Moreover, what does {C,E,L,R} mean? As in Eq. (2), (k, s) belongs to {C,E,L,R}, is it trying to minimize the entropy difference between arbitrary segment pairs? \n4. Can you elaborate more on the test flow? Are you using the pretrained source classifier for final prediction?\n5. Regarding the experiments parts: (1) more complex video benchmark datasets are encouraged to be tested; (2) why using the same CNN backbone? For different tasks, different backbones should be used. For instance, for action recognition tasks, there are more advanced backbones, e.g., I3D, videomae; (3) Model details should be presented in the main contents. Please elaborate more on the recovery model as it is the main contribution. Are you using LSTM or Bi-LSTM? Why not considering Transformer structure? (4) The results are not very convincing to show the performance superiority of TemSR as TemSR is the winner in 2 of 5, 1 of 5, and 3 of 5 tasks in the three datasets. (5) Sensitivity analysis shows that the optimal hyper-parameter value varies across tasks. It is not convincing to conclude the best value range is 1 to 10 by only using 3 datasets. The hyper-parameter generalization ability of the proposed TemSR is also an issue when facing new tasks or datasets. Moreover, what is “EEG” in figures 4,5,6?\n\n[ref1] Contrast and mix: Temporal contrastive video domain adaptation with background mixing\n\n[ref2] Unsupervised video domain adaptation for action recognition: A disentanglement perspective", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730275102262}, {"id": "fEsp0KR6SK", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9789/Reviewer_9pzb"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper a noval framework, TemSR, for Time-Series Source-Free Unsupervised Domain Adaptation (TS-SFUDA), designed to recover and transfer temporal dependencies without access to source data. TemSR employs a masking, recovery, and optimization process to generate a source-like distribution with restored temporal dependencies. To improve the recovery process, segment-based regularization is introduced to capture local dependencies and maximize anchor-based recovery diversity to ensure diversity in the recovered distribution. Extensive experiments across various time-series tasks demonstrate the effectiveness of TemSR, surpassing existing methods that rely on source-specific pretraining designs while maintaining source data privacy.", "review_text": "The paper a noval framework, TemSR, for Time-Series Source-Free Unsupervised Domain Adaptation (TS-SFUDA), designed to recover and transfer temporal dependencies without access to source data. TemSR employs a masking, recovery, and optimization process to generate a source-like distribution with restored temporal dependencies. To improve the recovery process, segment-based regularization is introduced to capture local dependencies and maximize anchor-based recovery diversity to ensure diversity in the recovered distribution. Extensive experiments across various time-series tasks demonstrate the effectiveness of TemSR, surpassing existing methods that rely on source-specific pretraining designs while maintaining source data privacy.", "strengths": "1. The proposed task ensures source data privacy by performing domain adaptation without needing the source data.\n2. The paper is logically structured, with a smooth connection between the motivation and the proposed method. Mathematical proofs are provided.\n3. The introduction of the anchor-based recovery diversity maximization helps generate more diverse and realistic source-like distributions, improving adaptation performance.\n4. 3 Benchmark datasets have been applied for evaluations.", "weaknesses": "1. The notation in Figure 1 is inconsistent with the notation in the text. Specifically, is the **L_sse** in Figure 1 the same as **L_seg** mentioned in line 259?\n2. The non-source-free baseline models used are outdated, with CoDATS being published four years ago. It would be beneficial to include more recent baseline models from 2023, such as CLUDA (\"Contrastive Learning in Unsupervised Domain Adaptation for Semantic Segmentation\", Neurips 2023) and RainCoat (\"Domain Adaptation for Time Series Under Feature and Label Shifts\", ICML 2023).\n3. Although using three benchmark datasets is sufficient, most previous MTS-UDA tasks typically select 7-10 source-target pairs (as the source-pair combination for HAR and SSC datasets can be hundreds) and report the average and standard deviation for comparisons, which provides a more balanced and fair evaluation. Moreover, the standard deviation across all pairs can demonstrate the model's robustness by highlighting its performance consistency across different cases. Having only five pairs per dataset is insufficient for a fair comparison.", "questions": "Methodologies:\n1. In Section 3.3 on Recovery-initialization, could you further explain which parts are initialized in Figure.1 framework? Furthermore, could you provide further clarification on why initializing with the target distribution satisfies the first requirement:  \"The initialized distribution should be close to the source distribution; otherwise, obtaining an effective source-like distribution is difficult\". If possible, could you add an ablation study by initializing randomly? Could you further explain why GAN does not meet the two requirement?\n2. Targeting to the \"masking\", as there is no \"dimension related\" information provided, could you give more details on the masking strategy? Will the masked parts be replaced with 0 or some noise? Or, you apply a similar masking strategy as MAE (Masked Autoencoders Are Scalable Vision Learners)?\n\nExperiments:\n1. In Section 4.1 (Datasets and Settings), you state that you followed the \"identical training configurations as Ragab et al. (2023b),\" referring to the paper titled \"Source-Free Domain Adaptation with Temporal Imputation for Time Series Data. (MAPU)\" However, after reviewing both the cited paper and the appendix of your work, I noticed that the datasets were split into training and testing sets only, with no mention of a validation set. If the best model is selected based on the test dataset, this could lead to biased evaluations, which raises concerns about the fairness of the evaluation process. In addition, For reproducibility, could you please include a table in the appendix listing key hyperparameters (e.g. learning rates, batch sizes, number of epochs) for all models, including baselines, to ensure fair comparison which could be easier for reading?\n2. In the L_seg, four masking positions are utilized under the assumption that they can effectively capture local dependencies. Could you conduct some ablation study experiments to demonstrate that all four types of masking positions are significant, or how does each position affects the overall performance?\n3. Could you add a computational complexity analysis with the proposed model and the baseline models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper a noval framework, TemSR, for Time-Series Source-Free Unsupervised Domain Adaptation (TS-SFUDA), designed to recover and transfer temporal dependencies without access to source data. TemSR employs a masking, recovery, and optimization process to generate a source-like distribution with restored temporal dependencies. To improve the recovery process, segment-based regularization is introduced to capture local dependencies and maximize anchor-based recovery diversity to ensure diversity in the recovered distribution. Extensive experiments across various time-series tasks demonstrate the effectiveness of TemSR, surpassing existing methods that rely on source-specific pretraining designs while maintaining source data privacy.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The proposed task ensures source data privacy by performing domain adaptation without needing the source data.\n2. The paper is logically structured, with a smooth connection between the motivation and the proposed method. Mathematical proofs are provided.\n3. The introduction of the anchor-based recovery diversity maximization helps generate more diverse and realistic source-like distributions, improving adaptation performance.\n4. 3 Benchmark datasets have been applied for evaluations.", "weaknesses": "1. The notation in Figure 1 is inconsistent with the notation in the text. Specifically, is the **L_sse** in Figure 1 the same as **L_seg** mentioned in line 259?\n2. The non-source-free baseline models used are outdated, with CoDATS being published four years ago. It would be beneficial to include more recent baseline models from 2023, such as CLUDA (\"Contrastive Learning in Unsupervised Domain Adaptation for Semantic Segmentation\", Neurips 2023) and RainCoat (\"Domain Adaptation for Time Series Under Feature and Label Shifts\", ICML 2023).\n3. Although using three benchmark datasets is sufficient, most previous MTS-UDA tasks typically select 7-10 source-target pairs (as the source-pair combination for HAR and SSC datasets can be hundreds) and report the average and standard deviation for comparisons, which provides a more balanced and fair evaluation. Moreover, the standard deviation across all pairs can demonstrate the model's robustness by highlighting its performance consistency across different cases. Having only five pairs per dataset is insufficient for a fair comparison.", "questions": "Methodologies:\n1. In Section 3.3 on Recovery-initialization, could you further explain which parts are initialized in Figure.1 framework? Furthermore, could you provide further clarification on why initializing with the target distribution satisfies the first requirement:  \"The initialized distribution should be close to the source distribution; otherwise, obtaining an effective source-like distribution is difficult\". If possible, could you add an ablation study by initializing randomly? Could you further explain why GAN does not meet the two requirement?\n2. Targeting to the \"masking\", as there is no \"dimension related\" information provided, could you give more details on the masking strategy? Will the masked parts be replaced with 0 or some noise? Or, you apply a similar masking strategy as MAE (Masked Autoencoders Are Scalable Vision Learners)?\n\nExperiments:\n1. In Section 4.1 (Datasets and Settings), you state that you followed the \"identical training configurations as Ragab et al. (2023b),\" referring to the paper titled \"Source-Free Domain Adaptation with Temporal Imputation for Time Series Data. (MAPU)\" However, after reviewing both the cited paper and the appendix of your work, I noticed that the datasets were split into training and testing sets only, with no mention of a validation set. If the best model is selected based on the test dataset, this could lead to biased evaluations, which raises concerns about the fairness of the evaluation process. In addition, For reproducibility, could you please include a table in the appendix listing key hyperparameters (e.g. learning rates, batch sizes, number of epochs) for all models, including baselines, to ensure fair comparison which could be easier for reading?\n2. In the L_seg, four masking positions are utilized under the assumption that they can effectively capture local dependencies. Could you conduct some ablation study experiments to demonstrate that all four types of masking positions are significant, or how does each position affects the overall performance?\n3. Could you add a computational complexity analysis with the proposed model and the baseline models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729770365944}], "openreview_url": "https://openreview.net/forum?id=GhM63V7z6v", "arxiv_id": "2409.19635", "paper_pdf": "papers/GhM63V7z6v.pdf", "paper_pdf_sha256": "87906c60a1cecfd7adf8ac59fa022e93883416d799e03474b292b9333385bb46", "paper_pdf_bytes": 766888, "paper_pdf_source": "openreview", "code_url": "https://github.com/Frank-Wang-oss/TemSR", "code_repository": "Frank-Wang-oss/TemSR", "code_commit": "21e14339b14868faf270fcc10046d9fc138bf2e4", "code_archive": "repos/GhM63V7z6v.zip", "code_archive_sha256": "526e1afa913d20ab7b0486034ccbe5e3136f05e7d4fdb339ff67050da286ebfe", "code_archive_bytes": 66508, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 84, "github_languages": {"Python": 91336}, "github_archived": false, "github_pushed_at": "2024-10-02T12:39:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/temporal-source-recovery-for-time-series"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "23OEmHVkpq", "year": 2024, "status": "rejected", "title": "Disentanglement Learning via Topology", "authors": ["Nikita Balabin", "Daria Voronkova", "Ilya Trofimov", "Evgeny Burnaev", "Serguei Barannikov"], "authorids": ["~Nikita_Balabin1", "~Daria_Voronkova1", "~Ilya_Trofimov1", "~Evgeny_Burnaev1", "~Serguei_Barannikov1"], "authors_source": "OpenReview API", "abstract": "We propose TopDis (Topological Disentanglement), a method for learning disentangled representations via adding a multi-scale topological loss term. Disentanglement is a crucial property of data representations substantial for the explainability and robustness of deep learning models and a step towards high-level cognition. The state-of-the-art methods are based on VAE and encourage the joint distribution of latent variables to be factorized. We take a different perspective on disentanglement by analyzing topological properties of data manifolds. In particular, we optimize the topological similarity for data manifolds traversals. To the best of our knowledge, our paper is the first one to propose a differentiable topological loss for disentanglement learning. Our experiments have shown that the proposed TopDis loss improves disentanglement scores such as MIG, FactorVAE score, SAP score and DCI disentanglement score with respect to state-of-the-art results while preserving the reconstruction quality. Our method works in an unsupervised manner, permitting to apply it for problems without labeled factors of variation. The TopDis loss works even when factors of variation are correlated. Additionally, we show how  to use the proposed topological loss to find disentangled directions in a trained GAN.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "8CzSIMq0YG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9426/Reviewer_1sM8"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors of this paper present TopDis, which is a regularizer based on Representation Topology Divergence (RTD). In this approach, the objective to be optimized is a combination of “classic” VAE loss and TopDis loss. Unlike the preceding approaches, topDis does not assume statistical independence between the factors of variations. Generally, introducing this loss term appears to further improve the current SOTA values for several disentanglement metrics (FactorVAE, MIG, SAP and DCI) across several different datasets (dSprites, 3D Shapes, 3D Faces, MPI 3D).", "review_text": "The authors of this paper present TopDis, which is a regularizer based on Representation Topology Divergence (RTD). In this approach, the objective to be optimized is a combination of “classic” VAE loss and TopDis loss. Unlike the preceding approaches, topDis does not assume statistical independence between the factors of variations. Generally, introducing this loss term appears to further improve the current SOTA values for several disentanglement metrics (FactorVAE, MIG, SAP and DCI) across several different datasets (dSprites, 3D Shapes, 3D Faces, MPI 3D).", "strengths": "1. The paper is clearly written and easy to follow. In detail:\n\na. The authors explain the task of disentanglement rather clearly by providing a succinct overview of previous works.\n\nb. The motivation and contribution of the paper are also clearly defined with an intuitive explanation of the designed methodology.\n\n2. The authors provide a variety of experiments and ablations, helping to evaluate their proposed disentanglement regularization loss practically. In detail:\n\na. The experiments (Table 1) appear comprehensive (except for the vanilla VAE; we will explain in the weakness section our concerns).\n\nb. The authors also provide enough qualitative examples, comparing models trained with TopDis regularizer and without.\n\nc. The architecture is succinctly described in the Appendix\n\n3. Computational complexity is also discussed in the Appendix, which is crucial for ML algorithms nowadays.", "weaknesses": "1. One of the contributions the authors mention is: “We improve the reconstruction quality by applying gradient orthogonalization;” - however, this contribution is only briefly mentioned in the conclusion and analyzed in the Appendix in greater detail. We suggest the authors to “move” the gradient orthogonalization part to the main paper.\n\n2. As the authors explained, the RTD was defined in a previous work, but we believe it is important to be defined in the main paper.\n\n3. In section 4.1, bullets (2-4). In (2), g\\inG appears to be applied to both pixel and latent space. Later in (3,4), where decomposition G is defined, it seems that it can be applied only in the latent space. We believe the authors should re-write this part, clarifying how G can be applied in the pixel space or, if that is not the case remove from (2) the application of g in the pixel space.\n\n4. In equation (4) regularization parameter /gamma is defined. Later in the appendix Q, \\gamma_1, and \\gamma_2 are used in the ablation table. Does this correspond, instead, to the loss: \\gamma_1 L_{VAE-based} + \\gamma_2 L_{TD}.\n\n5. In page 5 footnote, the authors state that RPT can be computed in latent space instead of pixel space. Can the authors provide ablations in the appendix exploring this direction? Do the authors have insights into how this change can affect the final trained model?\n\n6. Finally, our main concern is whether the proposed regularizer contributes to the learning of the disentangled representation or the used base models (i.e., \\beta-VAE, Factor-VAE). Since, in the main paper, only the models with already disentanglement remedies are explored and not the vanilla VAE. More concerning in the ablation, VAE+TopDis is explored, but it seems that the training is not the same as the VAE reported in the main paper. Our guess is that the models in the ablation were trained for less number of iterations. We encourage the authors to include in the main paper VAE+TopDis trained under the same conditions (i.e. same number of iterations) as the reported VAE in Table 1. This will help readers understand to what extent the TopDis regularizer helps learn disentangled representations", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors of this paper present TopDis, which is a regularizer based on Representation Topology Divergence (RTD). In this approach, the objective to be optimized is a combination of “classic” VAE loss and TopDis loss. Unlike the preceding approaches, topDis does not assume statistical independence between the factors of variations. Generally, introducing this loss term appears to further improve the current SOTA values for several disentanglement metrics (FactorVAE, MIG, SAP and DCI) across several different datasets (dSprites, 3D Shapes, 3D Faces, MPI 3D).", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is clearly written and easy to follow. In detail:\n\na. The authors explain the task of disentanglement rather clearly by providing a succinct overview of previous works.\n\nb. The motivation and contribution of the paper are also clearly defined with an intuitive explanation of the designed methodology.\n\n2. The authors provide a variety of experiments and ablations, helping to evaluate their proposed disentanglement regularization loss practically. In detail:\n\na. The experiments (Table 1) appear comprehensive (except for the vanilla VAE; we will explain in the weakness section our concerns).\n\nb. The authors also provide enough qualitative examples, comparing models trained with TopDis regularizer and without.\n\nc. The architecture is succinctly described in the Appendix\n\n3. Computational complexity is also discussed in the Appendix, which is crucial for ML algorithms nowadays.", "weaknesses": "1. One of the contributions the authors mention is: “We improve the reconstruction quality by applying gradient orthogonalization;” - however, this contribution is only briefly mentioned in the conclusion and analyzed in the Appendix in greater detail. We suggest the authors to “move” the gradient orthogonalization part to the main paper.\n\n2. As the authors explained, the RTD was defined in a previous work, but we believe it is important to be defined in the main paper.\n\n3. In section 4.1, bullets (2-4). In (2), g\\inG appears to be applied to both pixel and latent space. Later in (3,4), where decomposition G is defined, it seems that it can be applied only in the latent space. We believe the authors should re-write this part, clarifying how G can be applied in the pixel space or, if that is not the case remove from (2) the application of g in the pixel space.\n\n4. In equation (4) regularization parameter /gamma is defined. Later in the appendix Q, \\gamma_1, and \\gamma_2 are used in the ablation table. Does this correspond, instead, to the loss: \\gamma_1 L_{VAE-based} + \\gamma_2 L_{TD}.\n\n5. In page 5 footnote, the authors state that RPT can be computed in latent space instead of pixel space. Can the authors provide ablations in the appendix exploring this direction? Do the authors have insights into how this change can affect the final trained model?\n\n6. Finally, our main concern is whether the proposed regularizer contributes to the learning of the disentangled representation or the used base models (i.e., \\beta-VAE, Factor-VAE). Since, in the main paper, only the models with already disentanglement remedies are explored and not the vanilla VAE. More concerning in the ablation, VAE+TopDis is explored, but it seems that the training is not the same as the VAE reported in the main paper. Our guess is that the models in the ablation were trained for less number of iterations. We encourage the authors to include in the main paper VAE+TopDis trained under the same conditions (i.e. same number of iterations) as the reported VAE in Table 1. This will help readers understand to what extent the TopDis regularizer helps learn disentangled representations", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698783831422}, {"id": "JJaPaIjUli", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9426/Reviewer_VGnt"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposed a disentanglement regularization term based on topology, to constrain the manifold relation between the latent points of original images and shifted images. The authors provided extensive experiments on VAE-based methods and showed the effectiveness of the proposed methods.", "review_text": "The paper proposed a disentanglement regularization term based on topology, to constrain the manifold relation between the latent points of original images and shifted images. The authors provided extensive experiments on VAE-based methods and showed the effectiveness of the proposed methods.", "strengths": "1.\tIt is important to explore the constrain in the manifold of latent space for disentanglement, due to the statistical arguments of Locatello et al. (2019). The paper explored a way from topology and proposed a regularization term, which can be easily optimized. \n\n2.\tThe paper provided a good formulation of the TopDis loss and how to optimize it in the VAE framework.", "weaknesses": "1.\tThe relation between the constrain on latent space and disentanglement is still unclear, the TopDis is based on VAE-framework, which is based on Probability, and the paper referred to the definition of disentanglement based Group. And the paper failed to connect the above two framework, and making the proposed TopDis only kind of an intuitive necessary condition, as shown in Figure 3. \n\n2.\tFrom Appendix L, the best performance hyperparameters are quite different across different methods and different datasets, is there any guidance or criterion to choose the hyper-parameter?", "questions": "1.\tMy main concern is the relation between the proposed TopDis and disentanglement, is there any theoretical guarantee or deduction?  \n2.\tThe authors applied the proposed TopDis to infer disentangled directions in a pretrained style-GAN, is there some quantitative results? Then dose the method can be applied to other disentangled methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed a disentanglement regularization term based on topology, to constrain the manifold relation between the latent points of original images and shifted images. The authors provided extensive experiments on VAE-based methods and showed the effectiveness of the proposed methods.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1.\tIt is important to explore the constrain in the manifold of latent space for disentanglement, due to the statistical arguments of Locatello et al. (2019). The paper explored a way from topology and proposed a regularization term, which can be easily optimized. \n\n2.\tThe paper provided a good formulation of the TopDis loss and how to optimize it in the VAE framework.", "weaknesses": "1.\tThe relation between the constrain on latent space and disentanglement is still unclear, the TopDis is based on VAE-framework, which is based on Probability, and the paper referred to the definition of disentanglement based Group. And the paper failed to connect the above two framework, and making the proposed TopDis only kind of an intuitive necessary condition, as shown in Figure 3. \n\n2.\tFrom Appendix L, the best performance hyperparameters are quite different across different methods and different datasets, is there any guidance or criterion to choose the hyper-parameter?", "questions": "1.\tMy main concern is the relation between the proposed TopDis and disentanglement, is there any theoretical guarantee or deduction?  \n2.\tThe authors applied the proposed TopDis to infer disentangled directions in a pretrained style-GAN, is there some quantitative results? Then dose the method can be applied to other disentangled methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698720685980}, {"id": "OtNzsmJtSr", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9426/Reviewer_2nX5"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposed a novel Topological Disentanglement loss (TopDis loss) that can be added to any VAE-type loss to improve the disentanglement by encouraging the preservation of topological similarity in the generated samples with shifted latent space. Experiments demonstrated the proposed TopDis loss increases the disentanglement performance of several SOTA methods for various disentanglement metrics and datasets.", "review_text": "This paper proposed a novel Topological Disentanglement loss (TopDis loss) that can be added to any VAE-type loss to improve the disentanglement by encouraging the preservation of topological similarity in the generated samples with shifted latent space. Experiments demonstrated the proposed TopDis loss increases the disentanglement performance of several SOTA methods for various disentanglement metrics and datasets.", "strengths": "(1) Inspired by [1], the proposed differentiable Representation Topology Divergence (RTD) as a loss for the VAE-framework looks promising to improve the disentanglement.\n\n(2) Rich experiments are conducted to evaluate the performance of the proposed TopDis loss for various VAE-based methods.\n\n[1] Barannikov, Serguei, et al. \"Representation topology divergence: A method for comparing neural network representations.\" ICML 2022.", "weaknesses": "(1) It is unclear how the hyper-parameters in Eqn (4) affect the performance. There are γ_1 and γ_2 in Table 9 (appendix N), but there is only one γ in Eqn (4). \n\n(2) In Table 1, it seems that some advanced disentanglement methods performed significantly worse than the vanilla VAE (e.g. FactorVAE on 3dshapes, and β-TCVAE on MPI3D, etc), making it a little suspicious for the experimental results and/or the model selections of baselines. Besides, two important evaluations of VAE+TopDis and β-TCVAE+TopDis are missing. \n\n(3) The evaluation of how the proposed methods handle the tradeoff between disentanglement and reconstruction is limited. Besides Table 4 and Table 8, the authors are encouraged to report the reconstruction errors of the proposed method with and without \"gradient orthogonalization\" for a complete comparison with the baselines. Did the \"gradient orthogonalization\" apply to the baselines as well?", "questions": "(1) The authors are encouraged to respond to the concerns above.\n\n(2) How the γ should be selected for different VAE-based methods? Does TopDis improve disentanglement when β is already very large? How does the TopDis loss affect the optimization of the original disentanglement loss in those baselines (like the total correction in TC-VAE and FactorVAE)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a novel Topological Disentanglement loss (TopDis loss) that can be added to any VAE-type loss to improve the disentanglement by encouraging the preservation of topological similarity in the generated samples with shifted latent space. Experiments demonstrated the proposed TopDis loss increases the disentanglement performance of several SOTA methods for various disentanglement metrics and datasets.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "(1) Inspired by [1], the proposed differentiable Representation Topology Divergence (RTD) as a loss for the VAE-framework looks promising to improve the disentanglement.\n\n(2) Rich experiments are conducted to evaluate the performance of the proposed TopDis loss for various VAE-based methods.\n\n[1] Barannikov, Serguei, et al. \"Representation topology divergence: A method for comparing neural network representations.\" ICML 2022.", "weaknesses": "(1) It is unclear how the hyper-parameters in Eqn (4) affect the performance. There are γ_1 and γ_2 in Table 9 (appendix N), but there is only one γ in Eqn (4). \n\n(2) In Table 1, it seems that some advanced disentanglement methods performed significantly worse than the vanilla VAE (e.g. FactorVAE on 3dshapes, and β-TCVAE on MPI3D, etc), making it a little suspicious for the experimental results and/or the model selections of baselines. Besides, two important evaluations of VAE+TopDis and β-TCVAE+TopDis are missing. \n\n(3) The evaluation of how the proposed methods handle the tradeoff between disentanglement and reconstruction is limited. Besides Table 4 and Table 8, the authors are encouraged to report the reconstruction errors of the proposed method with and without \"gradient orthogonalization\" for a complete comparison with the baselines. Did the \"gradient orthogonalization\" apply to the baselines as well?", "questions": "(1) The authors are encouraged to respond to the concerns above.\n\n(2) How the γ should be selected for different VAE-based methods? Does TopDis improve disentanglement when β is already very large? How does the TopDis loss affect the optimization of the original disentanglement loss in those baselines (like the total correction in TC-VAE and FactorVAE)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698690713594}, {"id": "xGgF6MfB2r", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9426/Reviewer_T3Gf"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this paper, the authors propose a method, named TopDis (Topological Disentanglement), for learning disentangled representations via adding a multi-scale topological loss term. The experiments results show that the proposed TopDis loss improves disentanglement scores such as MIG, FactorVAE score, SAP score and DCI disentanglement score with respect to state-of-the-art results while preserving the reconstruction quality.", "review_text": "In this paper, the authors propose a method, named TopDis (Topological Disentanglement), for learning disentangled representations via adding a multi-scale topological loss term. The experiments results show that the proposed TopDis loss improves disentanglement scores such as MIG, FactorVAE score, SAP score and DCI disentanglement score with respect to state-of-the-art results while preserving the reconstruction quality.", "strengths": "- This paper is the first to introduce the use of a topological regularization term in the field of disentangled representation learning.\n- The topological regularization term is shown to be effective across multiple VAE models and metrics.\n- The regularization term proposed in this paper is also demonstrated to be effective for discovering pre-trained StyleGAN models.", "weaknesses": "- The paper lacks a clear reasonable explanation as to why topological constraints are meaningful/effective for disentanglement representation learning.\n- The new loss function was already proposed in a 2022 ICML paper [a]. The main contribution of this work is applying it to disentanglement, making the explanation of the above issue crucial for this paper.\n- The experiments focus on models with some disentanglement capabilities, but the effectiveness of this regularization term on vanilla VAEs has not been studied.\n- The performance of vanilla VAEs presented in this paper show high DCI performance, but other papers [b] report poor performance instead. A reasonable explanation is needed, and it would be helpful to include evaluation code in the supplementary materials.\n\n[a] Representation Topology Divergence: A method for comparing neural network representations.\n\n[b] β-VAE: LEARNING BASIC VISUAL CONCEPTS WITH A CONSTRAINED VARIATIONAL FRAMEWORK\n\n[c] Disentangling by Factorising", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors propose a method, named TopDis (Topological Disentanglement), for learning disentangled representations via adding a multi-scale topological loss term. The experiments results show that the proposed TopDis loss improves disentanglement scores such as MIG, FactorVAE score, SAP score and DCI disentanglement score with respect to state-of-the-art results while preserving the reconstruction quality.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- This paper is the first to introduce the use of a topological regularization term in the field of disentangled representation learning.\n- The topological regularization term is shown to be effective across multiple VAE models and metrics.\n- The regularization term proposed in this paper is also demonstrated to be effective for discovering pre-trained StyleGAN models.", "weaknesses": "- The paper lacks a clear reasonable explanation as to why topological constraints are meaningful/effective for disentanglement representation learning.\n- The new loss function was already proposed in a 2022 ICML paper [a]. The main contribution of this work is applying it to disentanglement, making the explanation of the above issue crucial for this paper.\n- The experiments focus on models with some disentanglement capabilities, but the effectiveness of this regularization term on vanilla VAEs has not been studied.\n- The performance of vanilla VAEs presented in this paper show high DCI performance, but other papers [b] report poor performance instead. A reasonable explanation is needed, and it would be helpful to include evaluation code in the supplementary materials.\n\n[a] Representation Topology Divergence: A method for comparing neural network representations.\n\n[b] β-VAE: LEARNING BASIC VISUAL CONCEPTS WITH A CONSTRAINED VARIATIONAL FRAMEWORK\n\n[c] Disentangling by Factorising", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698576965814}], "openreview_url": "https://openreview.net/forum?id=23OEmHVkpq", "arxiv_id": "2308.12696", "paper_pdf": "papers/23OEmHVkpq.pdf", "paper_pdf_sha256": "ea7e940dcbdccafa20685b3c0b38b6817e970c5e959d8b3b4d2d7bea1b79dc83", "paper_pdf_bytes": 21582958, "paper_pdf_source": "openreview", "code_url": "https://github.com/nikitabalabin/TopDis", "code_repository": "nikitabalabin/TopDis", "code_commit": "2ca9d2e7979eae4e5efed378b791091dae27f90b", "code_archive": "repos/23OEmHVkpq.zip", "code_archive_sha256": "6c2b5bc6d9cad429ac839b2f60922868915769ebdfcc560382dc8a927ce12284", "code_archive_bytes": 97068, "code_file_count": 50, "code_extensions": {".py": 45, ".sh": 5}, "github_disk_usage_kb": 61, "github_languages": {"Python": 291123, "Shell": 2335}, "github_archived": false, "github_pushed_at": "2024-05-30T16:29:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/disentanglement-learning-via-topology"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cS45VNtZLW", "year": 2023, "status": "rejected", "title": "Traversing Between Modes in Function Space for Fast Ensembling", "authors": ["Eunggu Yun", "Hyungi Lee", "Giung Nam", "Juho Lee"], "authorids": ["~Eunggu_Yun1", "~Hyungi_Lee1", "~Giung_Nam1", "~Juho_Lee2"], "authors_source": "OpenReview API", "abstract": "Deep ensemble is a simple yet powerful way to improve the performance of deep neural networks. Under this motivation, recent works on mode connectivity have shown that parameters of ensembles are connected by low-loss subspaces, and one can efficiently collect ensemble parameters in those subspaces. While this provides a way to efficiently train ensembles, for inference, one should still execute multiple forward passes using all the ensemble parameters, which often becomes a serious bottleneck for real-world deployment. In this work, we propose a novel framework to reduce such costs. Given a low-loss subspace connecting two modes of a neural network, we build an additional neural network predicting outputs of the original neural network evaluated at a certain point in the low-loss subspace. The additional neural network, what we call a “ bridge”, is a lightweight network taking minimal features from the original network, and predicting outputs for the low-loss subspace without forward passes through the original network. We empirically demonstrate that we can indeed train such bridge networks and significantly reduce inference costs with the help of the bridge networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Ig5isFeo4T", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2725/Reviewer_iSwW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to address an important drawback of Deep Ensemble, the inference cost of executing multiple models. The intuition behind the proposed method is that the outputs in the function subspace can be estimated from the modes without having to forward the actual parameters on the subspace. Based on this intuition, an additional lightweight network is trained as a bridge network to predict the outputs from the connecting subspace. ", "review_text": "I have some concerns about \n1) fundamental difference between type I and II bridge models\n2) Why number of bridge models impact \n3) generalization of this method to the large-scale datasets\n\nI am willing to raise my score if the above concerns are addressed. ", "strengths": "### Strength \n\n* Reducing the inference cost of DE is an important topic. This paper is well-motivated. \n* The presentation of this paper is clear and easy to follow. \n* Extensive experiments demonstrate the effectiveness of the proposed two types of bridge networks on TinyImageNet and Cifar. \n\n### Weaknesses\n* In Table 4, why do more bridge models lead to better results? \n* What's the functional difference between type I and II bridge models according to theoretical and empirical results? Is there any conclusion we can reach about how to choose types I and II?\n* Are there any important/sensitive hyper-parameters during the training of bridge models? How to determine the optimal number of bridge models in practice? \n* I am curious about how the proposed methods perform on large-scale datasets like ImageNet. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper aims to address an important drawback of Deep Ensemble, the inference cost of executing multiple models. The intuition behind the proposed method is that the outputs in the function subspace can be estimated from the modes without having to forward the actual parameters on the subspace. Based on this intuition, an additional lightweight network is trained as a bridge network to predict the outputs from the connecting subspace. ", "strength_and_weaknesses": "### Strength \n\n* Reducing the inference cost of DE is an important topic. This paper is well-motivated. \n* The presentation of this paper is clear and easy to follow. \n* Extensive experiments demonstrate the effectiveness of the proposed two types of bridge networks on TinyImageNet and Cifar. \n\n### Weaknesses\n* In Table 4, why do more bridge models lead to better results? \n* What's the functional difference between type I and II bridge models according to theoretical and empirical results? Is there any conclusion we can reach about how to choose types I and II?\n* Are there any important/sensitive hyper-parameters during the training of bridge models? How to determine the optimal number of bridge models in practice? \n* I am curious about how the proposed methods perform on large-scale datasets like ImageNet. ", "clarity,_quality,_novelty_and_reproducibility": "see above", "summary_of_the_review": "I have some concerns about \n1) fundamental difference between type I and II bridge models\n2) Why number of bridge models impact \n3) generalization of this method to the large-scale datasets\n\nI am willing to raise my score if the above concerns are addressed. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666708470490}, {"id": "iIxJpheDIF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2725/Reviewer_bBzc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes two methods to reduce the inference cost of Deep Ensemble (a collection of ensemble models of the same model structure). Bézier Curve is fitted on the extracted features by a neural network (bridge neural network), so that \"interpolated\" inference results can be cheaply approximated by the bridge NN.", "review_text": "The paper is well motivated and easy to follow, but still lacks solid grounding for its claims. ", "strengths": "Strength\nThe work is well motivated and the two proposed bridge networks method are well crafted.\n\nWeakness\n1. there is a key alternative not compared, in Learning Neural Network Subspaces a method is proposed to train models that can be linearly interpolated. Though it is not easy to compare the approaches, one is altering the training process and the other is more about adding an ancillary equipment, the two approaches can be evaluated in the same coordinate system with inference cost and accuracy on two axes.\n1. the method has not been verified on large scale datasets, though it criticized literature with \"these methods do not scale well for complex large-scale datasets\"", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes two methods to reduce the inference cost of Deep Ensemble (a collection of ensemble models of the same model structure). Bézier Curve is fitted on the extracted features by a neural network (bridge neural network), so that \"interpolated\" inference results can be cheaply approximated by the bridge NN.", "strength_and_weaknesses": "Strength\nThe work is well motivated and the two proposed bridge networks method are well crafted.\n\nWeakness\n1. there is a key alternative not compared, in Learning Neural Network Subspaces a method is proposed to train models that can be linearly interpolated. Though it is not easy to compare the approaches, one is altering the training process and the other is more about adding an ancillary equipment, the two approaches can be evaluated in the same coordinate system with inference cost and accuracy on two axes.\n1. the method has not been verified on large scale datasets, though it criticized literature with \"these methods do not scale well for complex large-scale datasets\"", "clarity,_quality,_novelty_and_reproducibility": "Proper illustrations has been given for the effects of the method. It would be appreciated if heatmap (loss landscape) can be given to facilitated comparisons in this line of work.", "summary_of_the_review": "The paper is well motivated and easy to follow, but still lacks solid grounding for its claims. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666675995132}, {"id": "NQIk5pN8arg", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2725/Reviewer_tqBD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new ensembling framework to reduce inference cost and boost the performance. The proposed method build additional light-weight head as bridges to ensemble different runs. The experiments show the effectiveness of the proposed method on CIFAR-10, CIFAR-100, and tiny-ImageNet.", "review_text": "(1)The idea of the proposed deep ensembling is novel and interesting.\n(2) The experiments on CIFAR-10, CIFAR-100, and tiny-ImageNet shows the improvement, but do not scale up to large-scale datasets and large-scale backbones (or SOTA backbones) .", "strengths": "Strengths:\n(1)\tThe idea of building bridges between different models for ensembling is interesting. It re-uses the features and only focuses on classification heads. Therefore, it would reduce the inference cost compared with directly ensembling different models.\n(2)\tThe experiments on several datasets show the effectiveness of the proposed framework.\n\n\nWeaknesses:\n(1)\tI concern only considering classification would not lead to diverse solutions to ensembling. As described in the paper, the bridges are light-weight MLPs.\n(2)\tThe paper pointed out “these methods (the existing methods) do not scale well for complex large-scale datasets or require network capacity”. However, this paper also DO NOT scale up on complex large-scale datasets, such as ImageNet-1k. Since the experimental platform is on 8 TPUs, it is expected to conduct on large-scale datasets such as ImageNet-1K, large-scale backbones (or SOTA backbones) such as transformers. The experiments on CIFAR-10, CIFAR-100 or tiny ImageNet are not convincing to me.\n(3)\tMinor comments, Page 1: the modes, three models, two modes? Mode or models?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new ensembling framework to reduce inference cost and boost the performance. The proposed method build additional light-weight head as bridges to ensemble different runs. The experiments show the effectiveness of the proposed method on CIFAR-10, CIFAR-100, and tiny-ImageNet.", "strength_and_weaknesses": "Strengths:\n(1)\tThe idea of building bridges between different models for ensembling is interesting. It re-uses the features and only focuses on classification heads. Therefore, it would reduce the inference cost compared with directly ensembling different models.\n(2)\tThe experiments on several datasets show the effectiveness of the proposed framework.\n\n\nWeaknesses:\n(1)\tI concern only considering classification would not lead to diverse solutions to ensembling. As described in the paper, the bridges are light-weight MLPs.\n(2)\tThe paper pointed out “these methods (the existing methods) do not scale well for complex large-scale datasets or require network capacity”. However, this paper also DO NOT scale up on complex large-scale datasets, such as ImageNet-1k. Since the experimental platform is on 8 TPUs, it is expected to conduct on large-scale datasets such as ImageNet-1K, large-scale backbones (or SOTA backbones) such as transformers. The experiments on CIFAR-10, CIFAR-100 or tiny ImageNet are not convincing to me.\n(3)\tMinor comments, Page 1: the modes, three models, two modes? Mode or models?\n\n", "clarity,_quality,_novelty_and_reproducibility": "The idea of the proposed deep ensembling is novel and the clarity and originality of the work are good. However, the experiments are conducted on small-scale datasets, which is not convincing to the community. ", "summary_of_the_review": "(1)The idea of the proposed deep ensembling is novel and interesting.\n(2) The experiments on CIFAR-10, CIFAR-100, and tiny-ImageNet shows the improvement, but do not scale up to large-scale datasets and large-scale backbones (or SOTA backbones) .", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666606619069}, {"id": "fsF66LiejrM", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2725/Reviewer_Kw57"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new way of using an ensemble of neural networks to construct a new classifier that achieves better accuracy at a lower number of FLOPS (at inference time). This is achieved by creating so-called \"bridge-classifiers\" that approximate the output of an interpolating classifier between pairs of ensemble-base classifiers. The interpolating classifiers are picked from the low-loss Bezier curve in parameter space between the pair of ensemble members. The approximation classifiers are faster than computing more ensemble members because they are smaller classifiers trained on intermediate features that have already been computed in the ensemble. The paper provides experiments on (small) image datasets (CIFAR and Tiny-ImageNet) and (small) classifiers of the ResNet family. ", "review_text": "The paper contains interesting approaches and results. However, there are also some weaknesses and limitations. I therefore think that the paper in its current form is marginally below the acceptance threshold, but I would be happy to increase my score during the discussion period if some of the concerns can be addressed by the authors.", "strengths": "**Strengths**\n* The paper proposes a novel idea for approximating gains of ensembles at lower computational cost (at inference time).\n* The presented results are interesting and show that the presented method works (for small data set and classifier sizes).\n* The paper is generally easy to follow. \n\n**Weaknesses**\n* The method is relatively complex in relation to the performance gains that can be achieved.\n* A direct comparison to other methods in terms of the accuracy/computation frontier is missing. It seems that the numbers are reasonable, but e.g. the range of 63% to 67% on Tiny-INet (Table 4) is in the area that can be also reached by a single ResNet-18 according to [Papers with Code](https://paperswithcode.com/sota/image-classification-on-tiny-imagenet-1). Also, there are no direct comparisons to other \"efficient ensemble\" approaches.\n* FLOPS alone are often not a sufficient measure of model efficiency, see e.g. [The Efficiency Misnomer](https://arxiv.org/pdf/2110.12894.pdf)\n\nFurther comments:\n\nIt was not clear to me after reading the paper at which point in the base networks the z_{i} are extracted. I think this should be clarified (or maybe I missed it?)\n\nI did not see an explanation for the term \"Bridge-S\" in the paper. (Maybe I missed it?) I guess it refers to a \"small\" Bridge model, but I did not see what exact configuration \"Bridge\" and \"Bridge-S\" refer to.\n\nI found the discussion of knowledge distillation a bit distracting from the main idea. Especially the experimental setting described in the last paragraph of Sec 5.2 has a form that is not super-convincing in my opinion.\n\nIt seems to me that the most interesting setting among those discussed is the DE-1 + n bridges. In that case it would actually be possible to try to approximate the r=1 responses of the ensemble (i.e. approximate the outputs of the other ensemble members from the intermediate feature). If this worked similarly well, the whole construction of the Bezier interpolation could be avoided. I think this might be an interesting ablation experiment, but maybe not feasible to perform during the rebuttal.\n\nOverall, I found Sec.4 rather short, given that this area of research is currently quite active. For example, the following papers could be considered relevant here. (I'm not requesting that all of these should be cited, just to give a few examples).\n* [Batch Ensembles](https://arxiv.org/pdf/2002.06715.pdf)\n* [Ensembles with Shared Representations](https://arxiv.org/pdf/2103.03934.pdf)\n* [Evaluating Scalable BDL](https://arxiv.org/pdf/1906.01620.pdf)\n\nMinor points or typos not affecting the evaluation:\n* It seems that Eq. (3) and (1) can be combined into a shorter version of the two equations.\n* p.3 line ~5: parameters ... achieves -> achieve\n* p.3 line ~5: add them to [the/an] ensemble\n* p.3 line ~8: \"these strategy provide\" -> this strategy provides\n* p.4 Eq. (9): last \"i\" should be a \"j\"\n* Algorithm 1 looks like a fairly standard training loop, unless I overlooked something. Maybe removing it would enable moving parts of the Appendix into the main paper. \n* p.4 bottom: \"when the bridge network does not step toward the Bezier predictions\" - I did not fully understand what this means.\n* Table 2: The FLOPS are given as a multiple. I assume this is as a multiple of a base network? \n* p.7 ~center: \"achieve decent R^2 score\" -> \"a score\" or \"scores\"", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new way of using an ensemble of neural networks to construct a new classifier that achieves better accuracy at a lower number of FLOPS (at inference time). This is achieved by creating so-called \"bridge-classifiers\" that approximate the output of an interpolating classifier between pairs of ensemble-base classifiers. The interpolating classifiers are picked from the low-loss Bezier curve in parameter space between the pair of ensemble members. The approximation classifiers are faster than computing more ensemble members because they are smaller classifiers trained on intermediate features that have already been computed in the ensemble. The paper provides experiments on (small) image datasets (CIFAR and Tiny-ImageNet) and (small) classifiers of the ResNet family. ", "strength_and_weaknesses": "**Strengths**\n* The paper proposes a novel idea for approximating gains of ensembles at lower computational cost (at inference time).\n* The presented results are interesting and show that the presented method works (for small data set and classifier sizes).\n* The paper is generally easy to follow. \n\n**Weaknesses**\n* The method is relatively complex in relation to the performance gains that can be achieved.\n* A direct comparison to other methods in terms of the accuracy/computation frontier is missing. It seems that the numbers are reasonable, but e.g. the range of 63% to 67% on Tiny-INet (Table 4) is in the area that can be also reached by a single ResNet-18 according to [Papers with Code](https://paperswithcode.com/sota/image-classification-on-tiny-imagenet-1). Also, there are no direct comparisons to other \"efficient ensemble\" approaches.\n* FLOPS alone are often not a sufficient measure of model efficiency, see e.g. [The Efficiency Misnomer](https://arxiv.org/pdf/2110.12894.pdf)\n\nFurther comments:\n\nIt was not clear to me after reading the paper at which point in the base networks the z_{i} are extracted. I think this should be clarified (or maybe I missed it?)\n\nI did not see an explanation for the term \"Bridge-S\" in the paper. (Maybe I missed it?) I guess it refers to a \"small\" Bridge model, but I did not see what exact configuration \"Bridge\" and \"Bridge-S\" refer to.\n\nI found the discussion of knowledge distillation a bit distracting from the main idea. Especially the experimental setting described in the last paragraph of Sec 5.2 has a form that is not super-convincing in my opinion.\n\nIt seems to me that the most interesting setting among those discussed is the DE-1 + n bridges. In that case it would actually be possible to try to approximate the r=1 responses of the ensemble (i.e. approximate the outputs of the other ensemble members from the intermediate feature). If this worked similarly well, the whole construction of the Bezier interpolation could be avoided. I think this might be an interesting ablation experiment, but maybe not feasible to perform during the rebuttal.\n\nOverall, I found Sec.4 rather short, given that this area of research is currently quite active. For example, the following papers could be considered relevant here. (I'm not requesting that all of these should be cited, just to give a few examples).\n* [Batch Ensembles](https://arxiv.org/pdf/2002.06715.pdf)\n* [Ensembles with Shared Representations](https://arxiv.org/pdf/2103.03934.pdf)\n* [Evaluating Scalable BDL](https://arxiv.org/pdf/1906.01620.pdf)\n\nMinor points or typos not affecting the evaluation:\n* It seems that Eq. (3) and (1) can be combined into a shorter version of the two equations.\n* p.3 line ~5: parameters ... achieves -> achieve\n* p.3 line ~5: add them to [the/an] ensemble\n* p.3 line ~8: \"these strategy provide\" -> this strategy provides\n* p.4 Eq. (9): last \"i\" should be a \"j\"\n* Algorithm 1 looks like a fairly standard training loop, unless I overlooked something. Maybe removing it would enable moving parts of the Appendix into the main paper. \n* p.4 bottom: \"when the bridge network does not step toward the Bezier predictions\" - I did not fully understand what this means.\n* Table 2: The FLOPS are given as a multiple. I assume this is as a multiple of a base network? \n* p.7 ~center: \"achieve decent R^2 score\" -> \"a score\" or \"scores\"", "clarity,_quality,_novelty_and_reproducibility": "Clarity: High; the paper is overall well-written and easy to understand.\n\nQuality: See strengths and weaknesses above.\n\nNovelty: OK; the main idea is novel and interesting, but it is a rather complex approach for which it is not clear to me whether it would be widely applicable in practice among the large number of approaches that aim to improve the accuracy/efficiency ratio.\n\nReproducibility: High; the authors provide the training code and configurations used.", "summary_of_the_review": "The paper contains interesting approaches and results. However, there are also some weaknesses and limitations. I therefore think that the paper in its current form is marginally below the acceptance threshold, but I would be happy to increase my score during the discussion period if some of the concerns can be addressed by the authors.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666275731834}], "openreview_url": "https://openreview.net/forum?id=cS45VNtZLW", "arxiv_id": "2306.11304", "paper_pdf": "papers/cS45VNtZLW.pdf", "paper_pdf_sha256": "16581ec478df829ef557c1bc56e957a93f7835e6d839c74a895ac1d5ea9b8570", "paper_pdf_bytes": 555293, "paper_pdf_source": "openreview", "code_url": "https://github.com/yuneg11/Bridge-Network", "code_repository": "yuneg11/Bridge-Network", "code_commit": "cba2980a512b55f8a0e171e6d8b1413c4715a038", "code_archive": "repos/cS45VNtZLW.zip", "code_archive_sha256": "663bc6781d1f38ffe332038aecb6d8b8e8bfd3b59dac0611d859ee2adae51170", "code_archive_bytes": 186666, "code_file_count": 25, "code_extensions": {".py": 25}, "github_disk_usage_kb": 126, "github_languages": {"Python": 239446}, "github_archived": false, "github_pushed_at": "2023-09-04T01:57:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/traversing-between-modes-in-function-space"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DIsWHvtU7lF", "year": 2022, "status": "rejected", "title": "Composing Partial Differential Equations with Physics-Aware Neural Networks", "authors": ["Matthias Karlbauer", "Timothy Praditia", "Sebastian Otte", "Sergey Oladyshkin", "Wolfgang Nowak", "Martin V. Butz"], "authorids": ["~Matthias_Karlbauer1", "~Timothy_Praditia1", "~Sebastian_Otte1", "~Sergey_Oladyshkin1", "~Wolfgang_Nowak1", "~Martin_V._Butz2"], "authors_source": "OpenReview API", "abstract": "We introduce a compositional physics-aware neural network (FINN) for learning spatiotemporal advection-diffusion processes. FINN implements a new way of combining the learning abilities of artificial neural networks with physical and structural knowledge from numerical simulation by modeling the constituents of partial differential equations (PDEs) in a compositional manner. Results on both one- and two-dimensional PDEs (Burger's, diffusion-sorption, diffusion-reaction) demonstrate FINN's superior process modeling accuracy and excellent out-of-distribution generalization ability beyond initial and boundary conditions. With only one tenth of the number of parameters on average, FINN outperforms pure machine learning and other state-of-the-art physics-aware models in all cases---often even by multiple orders of magnitude. Moreover, FINN outperforms a calibrated physical model when approximating sparse real-world data in a diffusion-sorption scenario, confirming its generalization abilities and showing explanatory potential by revealing the unknown retardation factor of the observed process.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "X3UUAOSIRSb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper699/Reviewer_j4js"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the authors propose the finite volume neural network to solve advection-diffusion partial differential equations. The authors define a flux kernel as the sum of sub-kernel f_i on each element. And the authors define specific modules \\phi_D, \\phi_A, \\phi_N according to the form of the advection-diffusion equation. The model requires very few parameters to achieve the state of art results.", "review_text": "This is a very concrete paper with solid experiments. \nTo better assess this work, I would like to ask several questions:\n\n1. I find it misleading to call this model \"compositional neural network\", which usually refers to composing neural network as f \\cdot g = f(g). If I understand correctly, it is not what the paper does.\n\n2. If I understand correctly, the FINN model computes the derivative using the finite-difference method. It can be understood as a learn FDM/FVM stencil. I wonder how it compared to other learned FDM methods such as Machine learning–accelerated computational fluid dynamics by Brenner et. al.?\n\n3. If there is a spectrum of ML models vs conventional numerical solvers, I think this FINN method lies closer to the end of numerical solvers (which is nothing bad). Solver-like methods usually require fewer parameters and generalize better. Therefore, I think it could be valuable if the authors can add some numerical experiments comparing against the numerical solver. Are there any advantages (speed, accuracy, etc) to using the learn FVM vs the original FVM/FDM?\n\n4. Standard advection-diffusion equations are relatively easy to solve. I wonder if the idea proposed in this paper can be generalized and transferred to other PDEs?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this work, the authors propose the finite volume neural network to solve advection-diffusion partial differential equations. The authors define a flux kernel as the sum of sub-kernel f_i on each element. And the authors define specific modules \\phi_D, \\phi_A, \\phi_N according to the form of the advection-diffusion equation. The model requires very few parameters to achieve the state of art results.", "main_review": "This is a very concrete paper with solid experiments. \nTo better assess this work, I would like to ask several questions:\n\n1. I find it misleading to call this model \"compositional neural network\", which usually refers to composing neural network as f \\cdot g = f(g). If I understand correctly, it is not what the paper does.\n\n2. If I understand correctly, the FINN model computes the derivative using the finite-difference method. It can be understood as a learn FDM/FVM stencil. I wonder how it compared to other learned FDM methods such as Machine learning–accelerated computational fluid dynamics by Brenner et. al.?\n\n3. If there is a spectrum of ML models vs conventional numerical solvers, I think this FINN method lies closer to the end of numerical solvers (which is nothing bad). Solver-like methods usually require fewer parameters and generalize better. Therefore, I think it could be valuable if the authors can add some numerical experiments comparing against the numerical solver. Are there any advantages (speed, accuracy, etc) to using the learn FVM vs the original FVM/FDM?\n\n4. Standard advection-diffusion equations are relatively easy to solve. I wonder if the idea proposed in this paper can be generalized and transferred to other PDEs?\n", "summary_of_the_review": "I think the paper is above the threshold. If the authors can provide evidence that the proposed method outperforms (or has some relative advantage) compared to standard FVM solvers, I will be happy to raise the score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635901874950}, {"id": "ab3WGs3GXoh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper699/Reviewer_LWqu"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents a compositional physics-aware neural network (FINN) for learning spatiotemporal advection-diffusion processes. It claims that the FINN outperforms pure machine learning and other state-of-the-art physics-aware models in all cases—often even by multiple orders of magnitude. However, the design of the network depends too much on the form of the equation, which leads to a very narrow application of the method.", "review_text": "Strengths\n\nThis method can not only deal with smooth solutions of general PDE but also can deal with weak solutions of hyperbolic problems, which is a good point; It is also compared with many recent SOTA methods and is significantly ahead of recently published methods\n \nWeaknesses\n\nThe method limits the form of the equation, which will greatly increase the training complexity when it is extended to higher-dimensional problems\n\nThe problems studied by this method are relatively simple and are the fitting of some linear problems or low-order nonlinear problems. I am wondering if polynomial fitting will get better results.\n\nSome test examples can be supplemented. Such as equations containing an exponential function, the fitted coefficient containing singularity, or can this method be used for fluid equations, etc. \n\nThe robustness test of this method is absent. Can this method be used for noisy data?\n\nThe discussion of method limitation is absent.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a compositional physics-aware neural network (FINN) for learning spatiotemporal advection-diffusion processes. It claims that the FINN outperforms pure machine learning and other state-of-the-art physics-aware models in all cases—often even by multiple orders of magnitude. However, the design of the network depends too much on the form of the equation, which leads to a very narrow application of the method.", "main_review": "Strengths\n\nThis method can not only deal with smooth solutions of general PDE but also can deal with weak solutions of hyperbolic problems, which is a good point; It is also compared with many recent SOTA methods and is significantly ahead of recently published methods\n \nWeaknesses\n\nThe method limits the form of the equation, which will greatly increase the training complexity when it is extended to higher-dimensional problems\n\nThe problems studied by this method are relatively simple and are the fitting of some linear problems or low-order nonlinear problems. I am wondering if polynomial fitting will get better results.\n\nSome test examples can be supplemented. Such as equations containing an exponential function, the fitted coefficient containing singularity, or can this method be used for fluid equations, etc. \n\nThe robustness test of this method is absent. Can this method be used for noisy data?\n\nThe discussion of method limitation is absent.\n", "summary_of_the_review": "The demonstration of this method is pretty good, but there are too many restrictions on the problem. If the form of the equation is fixed, a simple polynomial fitting may achieve a better fitting effect than the neural networks. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635821695331}, {"id": "tWoHtF6XqOG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper699/Reviewer_rFkv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper introduces Finite Volume neural network for modelling fluid dynamics inspired by the finite volume method. The paper models the velocity and the spacial derivatives as the neural networks. The work demonstrates more precise fluid simulation than the related physics-inspired models.", "review_text": "**Strong points:**\n\nThe paper provides significant improvement in performance in comparison to other models in a few orders of magnitude using 3-5 times less parameters than related models.\n\nThe results also suggest that modelling the spacial derivatives du / dx with the neural network directly results in better generalisation than modelling u(x,t and taking the derivative with respect to x (similar to PINN and other similar works). This is an interesting and counter-intuitive conclusion and worth of further exploration.\n\n**Weak points:**\n\nThe paper requires more explanation on the structure of the model. What is the input to phi_N, phi_A  and phi_D? How is R computed (it is not part of the equation in figure 1)? In Figure 1, why does phi_N signify both the first and the second spacial derivatives of u?\n\nIf the neural networks phi take only the values u, how are u(x,t) computed at the first time step? \n\nBased on figure 1, the right-hand side of the equation is a combination of four neural network outputs. Do you think it is equivalent to replacing phi_D phi_N - phi_A phi_N with a single neutral network taking all u_{i-1}, u_i, u_{I+1} (which essentially turns it into a Neural ODE model)?\n\nIt would be helpful to emphasise the differences between FINN, PINN and PhyDNet and other models in the methods section. Some of this explanation is provided in section 4.1, but it is worth emphasising and explaining more how the models differ, what exactly is modelled by a neural network and the motivation why FINN performs better than other models.\n\n\n**Questions:**\n\nPage 7: “PINN requires complete knowledge of the modelled system in form of the equation”. In this case, I am puzzled by a worse performance of PINN in figures 6 and 7 compared to FINN, if PINN closely follows the PDE that generated the data. Can you provide some insight why FINN performs better than PINN in this case? \n\nSimilarly, can you clarify why PINN requires more parameters than FINN (table 1)? Is it due to the fact that PINN also models the function u(x, t) or because of the bigger network?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces Finite Volume neural network for modelling fluid dynamics inspired by the finite volume method. The paper models the velocity and the spacial derivatives as the neural networks. The work demonstrates more precise fluid simulation than the related physics-inspired models.", "main_review": "**Strong points:**\n\nThe paper provides significant improvement in performance in comparison to other models in a few orders of magnitude using 3-5 times less parameters than related models.\n\nThe results also suggest that modelling the spacial derivatives du / dx with the neural network directly results in better generalisation than modelling u(x,t and taking the derivative with respect to x (similar to PINN and other similar works). This is an interesting and counter-intuitive conclusion and worth of further exploration.\n\n**Weak points:**\n\nThe paper requires more explanation on the structure of the model. What is the input to phi_N, phi_A  and phi_D? How is R computed (it is not part of the equation in figure 1)? In Figure 1, why does phi_N signify both the first and the second spacial derivatives of u?\n\nIf the neural networks phi take only the values u, how are u(x,t) computed at the first time step? \n\nBased on figure 1, the right-hand side of the equation is a combination of four neural network outputs. Do you think it is equivalent to replacing phi_D phi_N - phi_A phi_N with a single neutral network taking all u_{i-1}, u_i, u_{I+1} (which essentially turns it into a Neural ODE model)?\n\nIt would be helpful to emphasise the differences between FINN, PINN and PhyDNet and other models in the methods section. Some of this explanation is provided in section 4.1, but it is worth emphasising and explaining more how the models differ, what exactly is modelled by a neural network and the motivation why FINN performs better than other models.\n\n\n**Questions:**\n\nPage 7: “PINN requires complete knowledge of the modelled system in form of the equation”. In this case, I am puzzled by a worse performance of PINN in figures 6 and 7 compared to FINN, if PINN closely follows the PDE that generated the data. Can you provide some insight why FINN performs better than PINN in this case? \n\nSimilarly, can you clarify why PINN requires more parameters than FINN (table 1)? Is it due to the fact that PINN also models the function u(x, t) or because of the bigger network?\n\n", "summary_of_the_review": "The results in comparison to other models are compelling, particularly the model comparison on figures 6 and 7. However, the methods section of the paper needs more clarification and justification of the modelling choices in comparison to other papers. As the paper requires a significant re-write, I suggest a reject.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635501532130}, {"id": "ZrtunbMNsg6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper699/Reviewer_TVaa"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose to model advection-diffusion partial differential equations as a composition of multiple neural networks. According to the authors, this leads to better generalization over different initial and boundary conditions for the physics system, and also the ability to learn different factors of the process as modeled by the terms in the PDE. Extensive experimental results are presented to support the author’s claims that their framework performs better than state-of-the-art.", "review_text": "Strong points:\n1.\tThe authors propose very detailed experimental results, with both synthetic and real data\n2.\tThe background is provided in detail as well\n\nWeak points:\n1.\tThe authors claim in section 3 that convolution layers can only accommodate single types of boundary conditions. However, this is not correct since convolution layers at present can also handle cases I.e. periodic boundary conditions, this sentence may need rewriting\n2.\tIn figure 1, phi_N is shown to be both first and second-order derivatives, which may create confusion for the reader.\n3.\tThe training process is not completely clear, the authors describe the forward process, one would assume backpropagation of error for updating the neural network, but this is not explicit in the text. A figure demonstrating the whole training process, including the feedback loop might be helpful\n4.\tAt the end of section 3, the authors claim to use NODE in place of Euler for the reason of numerical stability, however, NODE also used Euler and does not mention anything about adaptive time-stepping, more details on this part would help to clear things up\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose to model advection-diffusion partial differential equations as a composition of multiple neural networks. According to the authors, this leads to better generalization over different initial and boundary conditions for the physics system, and also the ability to learn different factors of the process as modeled by the terms in the PDE. Extensive experimental results are presented to support the author’s claims that their framework performs better than state-of-the-art.", "main_review": "Strong points:\n1.\tThe authors propose very detailed experimental results, with both synthetic and real data\n2.\tThe background is provided in detail as well\n\nWeak points:\n1.\tThe authors claim in section 3 that convolution layers can only accommodate single types of boundary conditions. However, this is not correct since convolution layers at present can also handle cases I.e. periodic boundary conditions, this sentence may need rewriting\n2.\tIn figure 1, phi_N is shown to be both first and second-order derivatives, which may create confusion for the reader.\n3.\tThe training process is not completely clear, the authors describe the forward process, one would assume backpropagation of error for updating the neural network, but this is not explicit in the text. A figure demonstrating the whole training process, including the feedback loop might be helpful\n4.\tAt the end of section 3, the authors claim to use NODE in place of Euler for the reason of numerical stability, however, NODE also used Euler and does not mention anything about adaptive time-stepping, more details on this part would help to clear things up\n\n", "summary_of_the_review": "The paper has very thorough experimental results and good descriptions of the methods. However, the claims made in the paper may need a second look/ need some rewriting.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635438870672}], "openreview_url": "https://openreview.net/forum?id=DIsWHvtU7lF", "arxiv_id": "2111.11798", "paper_pdf": "papers/DIsWHvtU7lF.pdf", "paper_pdf_sha256": "a7f1e0de8ca5a64de2980c6e91197acd1d9d247fdc6e0667042bbed878a6f3b0", "paper_pdf_bytes": 2876471, "paper_pdf_source": "openreview", "code_url": "https://github.com/CognitiveModeling/finn", "code_repository": "CognitiveModeling/finn", "code_commit": "82b4eac6b55be3b44e736f139ed50256c4415f0c", "code_archive": "repos/DIsWHvtU7lF.zip", "code_archive_sha256": "e0cdc90bba3bff323ecb79e876d3a2376a0ea798773688677381c1797fd02526", "code_archive_bytes": 215241, "code_file_count": 68, "code_extensions": {".py": 68}, "github_disk_usage_kb": 155, "github_languages": {"Python": 635460}, "github_archived": false, "github_pushed_at": "2022-10-28T08:45:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/composing-partial-differential-equations-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cgRzg1V9su", "year": 2021, "status": "rejected", "title": "Inner Ensemble Networks: Average Ensemble as an Effective Regularizer", "authors": ["Abduallah Mohamed", "Muhammed Mohaimin Sadiq", "Ehab AlBadawy", "Mohamed Elhoseiny", "Christian Claudel"], "authorids": ["~Abduallah_Mohamed1", "m.mohaiminsadiq@utexas.edu", "ealbadawy@albany.edu", "~Mohamed_Elhoseiny1", "~Christian_Claudel1"], "authors_source": "OpenReview API", "abstract": "We introduce Inner Ensemble Networks (IENs) which reduce the variance within the neural network itself without an increase in the model complexity. IENs utilize ensemble parameters during the training phase to reduce the network variance. While in the testing phase, these parameters are removed without a change in the enhanced performance. IENs reduce the variance of an ordinary deep model by a factor of $1/m^{L-1}$, where $m$ is the number of inner ensembles and $L$ is the depth of the model. Also, we show empirically and theoretically that IENs lead to a greater variance reduction in comparison with other similar approaches such as dropout and maxout. Our results show a decrease of error rates between 1.7\\% and 17.3\\% in comparison with an ordinary deep model. We also show that IEN was preferred by Neural Architecture Search (NAS) methods over prior approaches.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "M2IQzEejCQM", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1815/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes replacing each linear layer with a linear ensemble (average) of m (m>1) linear layers.  Authors argue that this helps to reduce network internal variance and achieve better accuracy.\n\nI think it is possible to see that the proposed model is equivalent to the baseline model with a different hyper parameter setting (specifically the random initialization and learning rate), and therefore the observed variance reduction and improved performance are a result of better hyper-parameter choice in the proposed model.  The equivalence between the baseline model (with single linear transformation in every layer) and proposed model (m linear transforms in every layer whose output is averaged) can be seen by observing that:\n\nBaseline:  y = w.x\nProposed: y = (1/m) sum_j w_j x = (sum_j (1/m)w_j) x\n\nWe see the following relation between initial weights of baseline and proposed model: w = \\sum_j (1/m)w_j.   Furthermore, the gradients of objective w.r.t. w and w_j also follow a linear relationship: grad_{w_j} = (1/m)*grad_w, and due to this gradient scaling, the effective learning rate of the proposed model is smaller by a factor (1/m) as compared to the learning rate of the baseline model. Gradient of objective w.r.t. layer input x is same in the two models as long as the gradients w.r.t. output y are same and weights follow the relationship w = \\sum_j (1/m)w_j.\n\nThus I would think that if the baseline model follows hyper-parameter settings that are equivalent to the proposed model — i.e. weight initialization such that w = \\sum_j w_j (leading to reduced variance of initial weights), and a learning rate that is smaller by factor (1/m) — then it should see variances and performance that is akin to that of the proposed model.\n\nI'd like to hear author's response and discuss if my reasoning is flawed.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "proposed model equivalent to baseline model with different hyper-parameters?", "review": "The paper proposes replacing each linear layer with a linear ensemble (average) of m (m>1) linear layers.  Authors argue that this helps to reduce network internal variance and achieve better accuracy.\n\nI think it is possible to see that the proposed model is equivalent to the baseline model with a different hyper parameter setting (specifically the random initialization and learning rate), and therefore the observed variance reduction and improved performance are a result of better hyper-parameter choice in the proposed model.  The equivalence between the baseline model (with single linear transformation in every layer) and proposed model (m linear transforms in every layer whose output is averaged) can be seen by observing that:\n\nBaseline:  y = w.x\nProposed: y = (1/m) sum_j w_j x = (sum_j (1/m)w_j) x\n\nWe see the following relation between initial weights of baseline and proposed model: w = \\sum_j (1/m)w_j.   Furthermore, the gradients of objective w.r.t. w and w_j also follow a linear relationship: grad_{w_j} = (1/m)*grad_w, and due to this gradient scaling, the effective learning rate of the proposed model is smaller by a factor (1/m) as compared to the learning rate of the baseline model. Gradient of objective w.r.t. layer input x is same in the two models as long as the gradients w.r.t. output y are same and weights follow the relationship w = \\sum_j (1/m)w_j.\n\nThus I would think that if the baseline model follows hyper-parameter settings that are equivalent to the proposed model — i.e. weight initialization such that w = \\sum_j w_j (leading to reduced variance of initial weights), and a learning rate that is smaller by factor (1/m) — then it should see variances and performance that is akin to that of the proposed model.\n\nI'd like to hear author's response and discuss if my reasoning is flawed.", "rating": "3: Clear rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603914716428}, {"id": "lMrSYsHxTjM", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1815/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a drop-in replacement for CNN and FC layers that use multiple instances of the same layer and apply the average operation to those layers’ output. For inference, the weights are averaged themselves, so the params count in inference is the same as without the inner ensembles. The authors show that IEN ensembles of size m decrease the overall variance of a deep network by a factor of $\\frac{1}{m^{L-1}}$. They also perform experiments with various image models and finally show theoretically and empirically that their method leads to a greater decrease in variance and error rate than similar methods (maxout, dropout).\n\nPros:\n+ A simple and practically viable variance reduction scheme\n+ Good experimental design for the main set of experiments that shows clear advantages of the proposed scheme\n\nCons:\n- The theoretical analysis lacks rigor. There are hidden  assumptions in the methods used to achieve the conclusions. Please see the comment section below for more details.\n- The proposed method being simple means a lot of related works used methods similar to the proposed one. The related work section lacks a clear explanation for what distinguishes the paper from the others and makes it novel (for example, comparing to Abbasian et al., 2013)\n\nGeneral comments/questions:\n* (Section 3) It seems that $x_l$ should be $n_l$ dimensional and $y_l$ should be $d_l$ dimensional for variance in (4) to be correct\n* (Section 4.1) The derivation from (6) to (7) is too important to be hidden inside the appendix as it hides the assumptions made in A.2. Better move it to the main part of the paper.\n* (A.2) There are two parts here aiming at one goal: one is (24)-(29) and one from (sgK). Better keep only one of them for clarity.\n* (A.2) The equations in (24)-(29) are not generalized. In (24) and in (sgK) x is assumed to have zero mean. The authors have to prove (24)-(29) without this assumption. This can be done for specific activation functions, for example, RELU, but for RELU (24) doesn’t stand true without any additional assumptions.   \n* (Section 4.1) Equation (11) should probably use n_l\n* (Table 1) What is $Base_{\\widetilde{w}}$?\n* (Table 1) This table has the $IEN+FC_{\\tilde{w}}$ results but no $IEN+FC$ results. Can those be added for the sake of completeness? \n* (Table 1, Figure 2) Nit: the extensive use of colors makes the paper harder to process for colorblind people \n* (Section 2) (Abblasian et al., 2013) seems to be the first paper introducing Inner Ensembles. The authors should have a more extensive explanation of what is new in their paper and what distinguishes it from that existing one.\n* (Section 2) (Opitz et al., 2017) does not force the use of a special loss function instead of a main loss but instead suggests using additional loss function to achieve better performance through making weights more diverse. Potentially Inner Ensemble Networks may benefit from using a similar loss (but maybe not the $IEN_{\\widetilde{w}}$ versions). Adding those additional losses may be an interesting new set of experiments to run. \n* (Table 2) It is unclear why the outer ensemble outperforms single model IEN (not the $IEN_{\\widetilde{w}}$). The paper will benefit from a discussion on why this may be happening. Is it because IEN makes model weights too close to each other?  \n* (G) The dropout results are very poor. It suggests the models were not operating in a normal state. This may be the same issue as discussed in https://arxiv.org/abs/1603.05027. The https://arxiv.org/abs/1605.07146 suggests that proper placement of dropout can fix this problem.. In any case, comparing Dropout+IEN and Dropout is misleading as all Dropout experiments look catastrophically broken\n* (Section 5) A more detailed description of CNN setup with IEN would be for the paper’s benefit.   \n* (Table 1) Was maxout also used as a drop-in replacement without hyperparams tuning? It may be beneficial to also tune both maxout and IEN to distinguish IEN advantage of being a drop in replacement and possible IEN advantage when both methods are tuned to their maximum potential\n\n\nOn rating:\nThe Inner Ensemble Networks method looks like a simple, clean and easily reproducible way to reduce variance within a network. The authors extensively test their method empirically and show that this method also decreases error rates for various image-based models. However, the presented theoretical analysis lacks rigor and has to be updated to be formally correct.\n\nUpdate: taking authors comments and improvements into account I've updated the rating from 6 to 7.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A practical way to reduce the variance within the neural network but lacks theoretical rigor", "review": "The authors propose a drop-in replacement for CNN and FC layers that use multiple instances of the same layer and apply the average operation to those layers’ output. For inference, the weights are averaged themselves, so the params count in inference is the same as without the inner ensembles. The authors show that IEN ensembles of size m decrease the overall variance of a deep network by a factor of $\\frac{1}{m^{L-1}}$. They also perform experiments with various image models and finally show theoretically and empirically that their method leads to a greater decrease in variance and error rate than similar methods (maxout, dropout).\n\nPros:\n+ A simple and practically viable variance reduction scheme\n+ Good experimental design for the main set of experiments that shows clear advantages of the proposed scheme\n\nCons:\n- The theoretical analysis lacks rigor. There are hidden  assumptions in the methods used to achieve the conclusions. Please see the comment section below for more details.\n- The proposed method being simple means a lot of related works used methods similar to the proposed one. The related work section lacks a clear explanation for what distinguishes the paper from the others and makes it novel (for example, comparing to Abbasian et al., 2013)\n\nGeneral comments/questions:\n* (Section 3) It seems that $x_l$ should be $n_l$ dimensional and $y_l$ should be $d_l$ dimensional for variance in (4) to be correct\n* (Section 4.1) The derivation from (6) to (7) is too important to be hidden inside the appendix as it hides the assumptions made in A.2. Better move it to the main part of the paper.\n* (A.2) There are two parts here aiming at one goal: one is (24)-(29) and one from (sgK). Better keep only one of them for clarity.\n* (A.2) The equations in (24)-(29) are not generalized. In (24) and in (sgK) x is assumed to have zero mean. The authors have to prove (24)-(29) without this assumption. This can be done for specific activation functions, for example, RELU, but for RELU (24) doesn’t stand true without any additional assumptions.   \n* (Section 4.1) Equation (11) should probably use n_l\n* (Table 1) What is $Base_{\\widetilde{w}}$?\n* (Table 1) This table has the $IEN+FC_{\\tilde{w}}$ results but no $IEN+FC$ results. Can those be added for the sake of completeness? \n* (Table 1, Figure 2) Nit: the extensive use of colors makes the paper harder to process for colorblind people \n* (Section 2) (Abblasian et al., 2013) seems to be the first paper introducing Inner Ensembles. The authors should have a more extensive explanation of what is new in their paper and what distinguishes it from that existing one.\n* (Section 2) (Opitz et al., 2017) does not force the use of a special loss function instead of a main loss but instead suggests using additional loss function to achieve better performance through making weights more diverse. Potentially Inner Ensemble Networks may benefit from using a similar loss (but maybe not the $IEN_{\\widetilde{w}}$ versions). Adding those additional losses may be an interesting new set of experiments to run. \n* (Table 2) It is unclear why the outer ensemble outperforms single model IEN (not the $IEN_{\\widetilde{w}}$). The paper will benefit from a discussion on why this may be happening. Is it because IEN makes model weights too close to each other?  \n* (G) The dropout results are very poor. It suggests the models were not operating in a normal state. This may be the same issue as discussed in https://arxiv.org/abs/1603.05027. The https://arxiv.org/abs/1605.07146 suggests that proper placement of dropout can fix this problem.. In any case, comparing Dropout+IEN and Dropout is misleading as all Dropout experiments look catastrophically broken\n* (Section 5) A more detailed description of CNN setup with IEN would be for the paper’s benefit.   \n* (Table 1) Was maxout also used as a drop-in replacement without hyperparams tuning? It may be beneficial to also tune both maxout and IEN to distinguish IEN advantage of being a drop in replacement and possible IEN advantage when both methods are tuned to their maximum potential\n\n\nOn rating:\nThe Inner Ensemble Networks method looks like a simple, clean and easily reproducible way to reduce variance within a network. The authors extensively test their method empirically and show that this method also decreases error rates for various image-based models. However, the presented theoretical analysis lacks rigor and has to be updated to be formally correct.\n\nUpdate: taking authors comments and improvements into account I've updated the rating from 6 to 7.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603903700423}, {"id": "o6h7qB9rDdu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1815/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nIn this paper, the authors proposed Inner Ensembe Networks (IEN), which trains an ensemble of layers and therefore reduces model predictive variance. Experiments on image classification demonstrate that the proposed method is effective in reducing model variance and it also leads to an improved accuracy, compared to other variance reduction baselines such as maxout and dropout.\n\nPros:\n\nThe theoretical analysis (under some assumptions) in Section 4.3 and 4.4 illustrates the connection between IEN and dropout & maxout. Therefore, the improvement of the proposed method (IEN) on variance reduction over dropout and maxout is theoretically justified. The authors also show the boundary case where IEN achieves the same variance reduction as dropout and maxout. \n\nThe empirical evaluation matches the theoretical analysis. Figure 2 demonstrates one of the core contributions of this work: the proposed method IEN reduces the model variance even further than dropout and maxout.\n\nFigure 3 demonstrates how ensemble size affects the ranking of all methods and all architectures considered in this work. It shows that the improvement of IEN is consistent among all network selections and ensemble sizes. The authors also showcase the benefit of IEN can be carried to deep ensembles (outer-ensemble defined in this paper) and NAS cells.\n\nCons:\n\nMy concern is the motivation of why variance reduction is needed in deep neural networks is not highlighted. This paper implicitly assumes that lower model variance leads to better performance (in terms of accuracy metrics). However, the connection between model predictive variance and averaged model accuracy remains unknown (especially in deep ensembles). Recently many papers aim to improve ensemble diversity. It improves the ensemble performance but it also leads to larger model variance. On the other hand, regularization can be connected to better generalization performance, but this paper spends more paragraphs on the variance reduction effect. \n\nThe authors mentioned that reducing the variance in the prediction helps the ensemble of networks generalize better to unseen data. This implies that the proposed method IEN is supposed to outperform other baselines on out-of-distribution dataset (CIFAR-10C and CIFAR-100C). It would be more convincing if the authors can add empirical evaluations on out-of-distribution datasets.\n\nThe proposed method shares some similarities to Stochastic Weight Averaging (SWA) [1]. SWA also leverages the mean of multiple copies of the network weights, which also reduces the model variance. One main difference is SWA is a global mean of network weights while IEN takes the local (layer) averaged weights. The paper would be more inspiring if the authors can discuss the connection to SWA and make a comparison in the empirical evaluations.\n\nOverall, the paper proposed an interesting method to efficiently reduce model variance, which leads to an improved performance. But the cons outweight the pros in its current version.\n\n[1]: Izmailov, Pavel et al. “Averaging Weights Leads to Wider Optima and Better Generalization.” UAI (2018).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An efficient to reduce model variance but some modificiations are needed to be more convincing. ", "review": "Summary:\n\nIn this paper, the authors proposed Inner Ensembe Networks (IEN), which trains an ensemble of layers and therefore reduces model predictive variance. Experiments on image classification demonstrate that the proposed method is effective in reducing model variance and it also leads to an improved accuracy, compared to other variance reduction baselines such as maxout and dropout.\n\nPros:\n\nThe theoretical analysis (under some assumptions) in Section 4.3 and 4.4 illustrates the connection between IEN and dropout & maxout. Therefore, the improvement of the proposed method (IEN) on variance reduction over dropout and maxout is theoretically justified. The authors also show the boundary case where IEN achieves the same variance reduction as dropout and maxout. \n\nThe empirical evaluation matches the theoretical analysis. Figure 2 demonstrates one of the core contributions of this work: the proposed method IEN reduces the model variance even further than dropout and maxout.\n\nFigure 3 demonstrates how ensemble size affects the ranking of all methods and all architectures considered in this work. It shows that the improvement of IEN is consistent among all network selections and ensemble sizes. The authors also showcase the benefit of IEN can be carried to deep ensembles (outer-ensemble defined in this paper) and NAS cells.\n\nCons:\n\nMy concern is the motivation of why variance reduction is needed in deep neural networks is not highlighted. This paper implicitly assumes that lower model variance leads to better performance (in terms of accuracy metrics). However, the connection between model predictive variance and averaged model accuracy remains unknown (especially in deep ensembles). Recently many papers aim to improve ensemble diversity. It improves the ensemble performance but it also leads to larger model variance. On the other hand, regularization can be connected to better generalization performance, but this paper spends more paragraphs on the variance reduction effect. \n\nThe authors mentioned that reducing the variance in the prediction helps the ensemble of networks generalize better to unseen data. This implies that the proposed method IEN is supposed to outperform other baselines on out-of-distribution dataset (CIFAR-10C and CIFAR-100C). It would be more convincing if the authors can add empirical evaluations on out-of-distribution datasets.\n\nThe proposed method shares some similarities to Stochastic Weight Averaging (SWA) [1]. SWA also leverages the mean of multiple copies of the network weights, which also reduces the model variance. One main difference is SWA is a global mean of network weights while IEN takes the local (layer) averaged weights. The paper would be more inspiring if the authors can discuss the connection to SWA and make a comparison in the empirical evaluations.\n\nOverall, the paper proposed an interesting method to efficiently reduce model variance, which leads to an improved performance. But the cons outweight the pros in its current version.\n\n[1]: Izmailov, Pavel et al. “Averaging Weights Leads to Wider Optima and Better Generalization.” UAI (2018).\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603867814985}, {"id": "YxOo_BCTGVI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1815/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**\nThis paper proposes \"Inner Ensemble Network\" that ensembles intermediate layers in a deep neural network, which would be trained simultaneously during the training. On the inference time, the ensembled portion of the network would be substituted by the averaged layer, enabling lighter-weighted inference while preserving the benefit from the ensemble. Similar to the classic ensemble methods, this method can reduce the variance of the training and achieve better than the models without the ensemble.\n\n**Originality and significance aspect**\nThis is probably the most weak aspect of this paper. The proposed methodology, Inner Ensemble Networks (or IEN) is very similar to the idea of average pooling layers that are widely used in computer vision architectures.  It is also fairly similar to some variants of online ensemble methods such as [1, 2, 3]. [2] is actually already cited in this paper; however, my point is that there is not enough novelty on this paper compared to these prior works. Authors claim this paper is simpler to apply than [2], but they do not have any empirical evidence that their proposed method is better than [2]. What if [2] works much better because their *sophisticated* loss function actually works out nicely?\n\nFrom my understanding, IEN works just as the average pooling layers of duplicate layers and probably only novelty here is that they substitute the ensembled layers with the averaged layer similar to drop-out technique. I feel this is not enough for a full conference paper even though this paper may be still useful in practice.\n\n[1] Liu, Yong, and Xin Yao. \"Simultaneous training of negatively correlated neural networks in an ensemble.\" IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 29.6 (1999): 716-725.\n[2] Opitz, Michael, Horst Possegger, and Horst Bischof. \"Efficient model averaging for deep neural networks.\" Asian Conference on Computer Vision. Springer, Cham, 2016.\n[3] Lee, Stefan, et al. \"Why M heads are better than one: Training a diverse ensemble of deep networks.\" arXiv preprint arXiv:1511.06314 (2015\n\n**Quality aspect**\nThe paper could be polished further in their figures and equations. For example the bottom line of the right side table of Figure 1 is missing and the left picture is not that informative. Many of the equations are not typesetted correctly. I think authors may consider using LaTex for tye type setting. The parenthesis in the equations (e.g. eq 4-11) should match the inner element’s size.\n\nFor the experiments, I think authors should actually use the classic version of ensemble (call outer-ensemble in Section 5.2 or in Table 2) as their baseline instead of a single model. It is because IEN costs almost 2x during the training and that is similar to the classic ensemble. If authors are concerned about the inference computation time, they can consider a distilled model from the classic ensemble method, which is a fairly standard technique.\n\n**Clarity aspect**\nThe paper is mostly clear. \n\n**Recommendation**\nOverall, the paper is interesting; however, I cannot recommend this paper mainly due to weak novelty and weak empirical support.\n\n----\n**Post rebuttal comment**\nI thank the authors for detailed rebuttal and new empirical results. I also have read other reviewer's comment, and decided to keep my original score. The main concern of this paper, the weak novelty, still remains (also pointed out by Reviewer#1/4). \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Concerns on the limited novelty of this paper", "review": "**Summary**\nThis paper proposes \"Inner Ensemble Network\" that ensembles intermediate layers in a deep neural network, which would be trained simultaneously during the training. On the inference time, the ensembled portion of the network would be substituted by the averaged layer, enabling lighter-weighted inference while preserving the benefit from the ensemble. Similar to the classic ensemble methods, this method can reduce the variance of the training and achieve better than the models without the ensemble.\n\n**Originality and significance aspect**\nThis is probably the most weak aspect of this paper. The proposed methodology, Inner Ensemble Networks (or IEN) is very similar to the idea of average pooling layers that are widely used in computer vision architectures.  It is also fairly similar to some variants of online ensemble methods such as [1, 2, 3]. [2] is actually already cited in this paper; however, my point is that there is not enough novelty on this paper compared to these prior works. Authors claim this paper is simpler to apply than [2], but they do not have any empirical evidence that their proposed method is better than [2]. What if [2] works much better because their *sophisticated* loss function actually works out nicely?\n\nFrom my understanding, IEN works just as the average pooling layers of duplicate layers and probably only novelty here is that they substitute the ensembled layers with the averaged layer similar to drop-out technique. I feel this is not enough for a full conference paper even though this paper may be still useful in practice.\n\n[1] Liu, Yong, and Xin Yao. \"Simultaneous training of negatively correlated neural networks in an ensemble.\" IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 29.6 (1999): 716-725.\n[2] Opitz, Michael, Horst Possegger, and Horst Bischof. \"Efficient model averaging for deep neural networks.\" Asian Conference on Computer Vision. Springer, Cham, 2016.\n[3] Lee, Stefan, et al. \"Why M heads are better than one: Training a diverse ensemble of deep networks.\" arXiv preprint arXiv:1511.06314 (2015\n\n**Quality aspect**\nThe paper could be polished further in their figures and equations. For example the bottom line of the right side table of Figure 1 is missing and the left picture is not that informative. Many of the equations are not typesetted correctly. I think authors may consider using LaTex for tye type setting. The parenthesis in the equations (e.g. eq 4-11) should match the inner element’s size.\n\nFor the experiments, I think authors should actually use the classic version of ensemble (call outer-ensemble in Section 5.2 or in Table 2) as their baseline instead of a single model. It is because IEN costs almost 2x during the training and that is similar to the classic ensemble. If authors are concerned about the inference computation time, they can consider a distilled model from the classic ensemble method, which is a fairly standard technique.\n\n**Clarity aspect**\nThe paper is mostly clear. \n\n**Recommendation**\nOverall, the paper is interesting; however, I cannot recommend this paper mainly due to weak novelty and weak empirical support.\n\n----\n**Post rebuttal comment**\nI thank the authors for detailed rebuttal and new empirical results. I also have read other reviewer's comment, and decided to keep my original score. The main concern of this paper, the weak novelty, still remains (also pointed out by Reviewer#1/4). \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603858126097}], "openreview_url": "https://openreview.net/forum?id=cgRzg1V9su", "arxiv_id": "2006.08305", "paper_pdf": "papers/cgRzg1V9su.pdf", "paper_pdf_sha256": "b33d0c266351bda9939e3e573cbe66e5f5ed3a5ea6b414f30bf4a7d7c9e015d2", "paper_pdf_bytes": 390622, "paper_pdf_source": "openreview", "code_url": "https://github.com/abduallahmohamed/inner_ensemble_nets", "code_repository": "abduallahmohamed/inner_ensemble_nets", "code_commit": "56b06e33c74e49ffe8ab89af64322a01d5ee92ad", "code_archive": "repos/cgRzg1V9su.zip", "code_archive_sha256": "4e227e9fe064f7ead67222ff037ab9af94572d47290ed3dd87c0f66e42f87a28", "code_archive_bytes": 73683, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 194, "github_languages": {"Python": 5953}, "github_archived": false, "github_pushed_at": "2020-10-09T06:01:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/inner-ensemble-nets"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJx-akSKPS", "year": 2020, "status": "rejected", "title": "Neural Subgraph Isomorphism Counting", "authors": ["Xin Liu", "Haojie Pan", "Mutian He", "Yangqiu Song", "Xin Jiang"], "authorids": ["xliucr@cse.ust.hk", "hpanad@cse.ust.hk", "mhear@cse.ust.hk", "yqsong@cse.ust.hk", "jiang.xin@huawei.com"], "authors_source": "OpenReview API", "abstract": "In this paper, we study a new graph learning problem: learning to count subgraph isomorphisms. Although the learning based approach is inexact, we are able to generalize to count large patterns and data graphs in polynomial time compared to the exponential time of the original NP-complete problem. Different from other traditional graph learning problems such as node classification and link prediction, subgraph isomorphism counting requires more global inference to oversee the whole graph. To tackle this problem, we propose a dynamic intermedium attention memory network (DIAMNet) which augments different representation learning architectures and iteratively attends pattern and target data graphs to memorize different subgraph isomorphisms for the global counting. We develop both small graphs (<= 1,024 subgraph isomorphisms in each) and large graphs (<= 4,096 subgraph isomorphisms in each) sets to evaluate different models. Experimental results show that learning based subgraph isomorphism counting can help reduce the time complexity with acceptable accuracy. Our DIAMNet can further improve existing representation learning models for this more global problem.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "B1lIjMJPjr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1979/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a dynamic inter-medium attention memory network and model the sub-graph isomorphism counting problem as a learning problem with both polynomial training and prediction time complexities.\nSince the testing time is reported in this paper, and the time complexity is one of the main contribution of this paper. The hardware and software used to run the algorithm should be reported in the main article.\n\nThe author argues that if we use neural networks to learn distributed representations for V_G and V_p or \\xi_G and \\xi_P without self-attention, the computational cost will acceptable for large graphs, but the missing of self-attention will hurt the performance. It’s encouraged to do corresponding experiments to compare it with the proposed method and better support the algorithm.\n\nOne of the main advantages of this paper is that the proposed method can efficiently deal with large graph tasks, so the model behaviors of different models in large dataset similar to Figure 5 is encouraged to be given.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This paper proposes a dynamic inter-medium attention memory network and model the sub-graph isomorphism counting problem as a learning problem with both polynomial training and prediction time complexities.\nSince the testing time is reported in this paper, and the time complexity is one of the main contribution of this paper. The hardware and software used to run the algorithm should be reported in the main article.\n\nThe author argues that if we use neural networks to learn distributed representations for V_G and V_p or \\xi_G and \\xi_P without self-attention, the computational cost will acceptable for large graphs, but the missing of self-attention will hurt the performance. It’s encouraged to do corresponding experiments to compare it with the proposed method and better support the algorithm.\n\nOne of the main advantages of this paper is that the proposed method can efficiently deal with large graph tasks, so the model behaviors of different models in large dataset similar to Figure 5 is encouraged to be given.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1573479070256}, {"id": "Hkeg_NnxoB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1979/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studied how to leverage the power of graph neural networks for counting subgraph isomorphism. The motivation is that the current subgraph isomorphism detection is NP-complete problem and a proposed approach based on GNN could approximately solve the counting problem in polynomial time. Then they relaxed original subgraph isomorphism (which is equivalent to the exact subgraph matching problem)  and proposed the problem of doing subgraph isomorphism counting task. The GNN and sequence modeling methods are discussed for solving this problem. The experimental results confirmed the effectiveness of these methods. \n\nAlthough I found the subgraph isomorphism counting problem is an interesting problem, I did not know how much practical usefulness of this task. More practical use case would be search for the matched subgraphs given the sub-graph query using subgraph isomorphism detection. \n\nAlso, although authors mentioned some approximation systems/methods in graph database community such as TurboISO (Han et al., 2013), VF3 (Carletti et al., 2018), and other approximation techniques [1][2], authors did not consider them as baselines to compare. These methods may also have limitations to deal with real-large graph but for the graph size that this paper studied I think they are fine to deal with. A parallel issue is that GNN also has scalability issues as well when dealing with large graphs [3]. Without comparing these existing fast (approximation) methods, it is really unfair to compare with only non-DL baseline VF2, which seems served as ground-truth as well. \n\n[1]  A Neural Graph Isomorphism Algorithm Based on Local Invariants, ESANN'2003\n[2] Subgraph Isomorphism in Polynomial Time\n[3] FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling\n\nIn terms of technical contributions, they leverage some existing sequence models (CNN, RNN and so on) and graph models (RGNN) and the whole framework is similar to doing a graph matching networks (without considering node alignment) for a regression task. The DYNAMIC INTERMEDIUM ATTENTION MEMORY NETWORK is interesting yet simple. I am not entirely clear what's the output  of this interactional module. The figure 4 shows the overall architecture of subgraph isomorphism counting model, which needs better descriptions to understand exact input and output for each module. In general, the novelty of this part is incremental. \n\nFinally, this subgraph isomorphism counting problem is closely related to graphlet counting problem. In the paper, the subgraph pattern considered seems like almost identical to graphlets the previous research extensively studied. I did not see any discussion about the connection of these two tasks either. \n\nMinor comments:\n\n|V_G| is is the number of pattern nodes -> |V_p| is is the number of pattern nodes", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #4", "review": "This paper studied how to leverage the power of graph neural networks for counting subgraph isomorphism. The motivation is that the current subgraph isomorphism detection is NP-complete problem and a proposed approach based on GNN could approximately solve the counting problem in polynomial time. Then they relaxed original subgraph isomorphism (which is equivalent to the exact subgraph matching problem)  and proposed the problem of doing subgraph isomorphism counting task. The GNN and sequence modeling methods are discussed for solving this problem. The experimental results confirmed the effectiveness of these methods. \n\nAlthough I found the subgraph isomorphism counting problem is an interesting problem, I did not know how much practical usefulness of this task. More practical use case would be search for the matched subgraphs given the sub-graph query using subgraph isomorphism detection. \n\nAlso, although authors mentioned some approximation systems/methods in graph database community such as TurboISO (Han et al., 2013), VF3 (Carletti et al., 2018), and other approximation techniques [1][2], authors did not consider them as baselines to compare. These methods may also have limitations to deal with real-large graph but for the graph size that this paper studied I think they are fine to deal with. A parallel issue is that GNN also has scalability issues as well when dealing with large graphs [3]. Without comparing these existing fast (approximation) methods, it is really unfair to compare with only non-DL baseline VF2, which seems served as ground-truth as well. \n\n[1]  A Neural Graph Isomorphism Algorithm Based on Local Invariants, ESANN'2003\n[2] Subgraph Isomorphism in Polynomial Time\n[3] FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling\n\nIn terms of technical contributions, they leverage some existing sequence models (CNN, RNN and so on) and graph models (RGNN) and the whole framework is similar to doing a graph matching networks (without considering node alignment) for a regression task. The DYNAMIC INTERMEDIUM ATTENTION MEMORY NETWORK is interesting yet simple. I am not entirely clear what's the output  of this interactional module. The figure 4 shows the overall architecture of subgraph isomorphism counting model, which needs better descriptions to understand exact input and output for each module. In general, the novelty of this part is incremental. \n\nFinally, this subgraph isomorphism counting problem is closely related to graphlet counting problem. In the paper, the subgraph pattern considered seems like almost identical to graphlets the previous research extensively studied. I did not see any discussion about the connection of these two tasks either. \n\nMinor comments:\n\n|V_G| is is the number of pattern nodes -> |V_p| is is the number of pattern nodes", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573074023767}, {"id": "SylbC532YB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1979/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method called Dynamic Intermedium Attention Memory Network (DIAMNet) to learn the subgraph isomorphism counting for a given pattern graph P and target graph G. This requires global information unlike usual GNN cases such as node classification, link prediction, community detection. First, input graphs P and G are converted embedding vectors through sequence models (CNN, RNN, Transformer-XL) or graph models (RGCN), and fed into their DIAMNet that uses an external memory as an intermedium to attend both the pattern and the graph. The external memory is updated based on multi-head attention as in Transformer. The output of DIAMNet is passed to FC that outputs 'count' directly. The training is based on minimizing MSE loss as a regression problem. Extensive experimental evaluations report that DIAMNet showed superior performance over competing methods and baselines.\n\nThis paper targets subgraph isomorphism counting as a learning problem for the first time I guess, and the proposed method combined with both graph- and sequence-based encoding is technically interesting. However, there are still two major issues of 1) why counting? 2) the RMSE loss for regression on counts 3) baseline of 'Zero'.\n\n1) the most unclear point is 'why counting?'. If I understand it, this method can be applied to subgraph isomorphism (NP-hard) or graph isomorphism (unknown complexity) as binary classification, and experimental evaluations can use the datasets used in evaluating VF2 or Naughty. It would be better to start this fundamental problem that would have many clear applications. Compared to subgraph isomorphism or graph isomorphism, the need for knowing accurate 'counts' of subgraph isomorphisms is unconvincing (given that we cannot explicitly obtains all subgraph matchings). Note that there is some existing research on GNNs targeted 'graph matching' and 'graph similarity'. \n\nAlso, the used datasets intentionally restrict the possible values for the number of subgraph isomorphisms, but the counts would be exponentially large if we consider practical (dense) graphs. \n\n2) the method fits the model using (R)MSE loss, but minimizing log errors ((R)MSLE) would be better considering distributions of response values (counts) of the used datasets in Figure 6. Fitting the MSE loss is not good for such highly skewed cases, and for example, might focus only on the few instances having very large count values. Or, if such instances are very small, training ignores all such extreme instances. Either way would be questionable when we consider learning 'subgraph isomorphism counting' in general. \n\nAlso, the error of counts by MSE or MAE would be less informative and it would be unclear how much errors are tolerant in practical use cases of this method. \n\n3) To interpret the RMSE and MAE values, Table 2 has the value for 'Zero'. This is for a constant predictor always returning zeros for any inputs. However, given that the loss is MSE, constant prediction values should be the average counts in the training data, not zero. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper proposes a method called Dynamic Intermedium Attention Memory Network (DIAMNet) to learn the subgraph isomorphism counting for a given pattern graph P and target graph G. This requires global information unlike usual GNN cases such as node classification, link prediction, community detection. First, input graphs P and G are converted embedding vectors through sequence models (CNN, RNN, Transformer-XL) or graph models (RGCN), and fed into their DIAMNet that uses an external memory as an intermedium to attend both the pattern and the graph. The external memory is updated based on multi-head attention as in Transformer. The output of DIAMNet is passed to FC that outputs 'count' directly. The training is based on minimizing MSE loss as a regression problem. Extensive experimental evaluations report that DIAMNet showed superior performance over competing methods and baselines.\n\nThis paper targets subgraph isomorphism counting as a learning problem for the first time I guess, and the proposed method combined with both graph- and sequence-based encoding is technically interesting. However, there are still two major issues of 1) why counting? 2) the RMSE loss for regression on counts 3) baseline of 'Zero'.\n\n1) the most unclear point is 'why counting?'. If I understand it, this method can be applied to subgraph isomorphism (NP-hard) or graph isomorphism (unknown complexity) as binary classification, and experimental evaluations can use the datasets used in evaluating VF2 or Naughty. It would be better to start this fundamental problem that would have many clear applications. Compared to subgraph isomorphism or graph isomorphism, the need for knowing accurate 'counts' of subgraph isomorphisms is unconvincing (given that we cannot explicitly obtains all subgraph matchings). Note that there is some existing research on GNNs targeted 'graph matching' and 'graph similarity'. \n\nAlso, the used datasets intentionally restrict the possible values for the number of subgraph isomorphisms, but the counts would be exponentially large if we consider practical (dense) graphs. \n\n2) the method fits the model using (R)MSE loss, but minimizing log errors ((R)MSLE) would be better considering distributions of response values (counts) of the used datasets in Figure 6. Fitting the MSE loss is not good for such highly skewed cases, and for example, might focus only on the few instances having very large count values. Or, if such instances are very small, training ignores all such extreme instances. Either way would be questionable when we consider learning 'subgraph isomorphism counting' in general. \n\nAlso, the error of counts by MSE or MAE would be less informative and it would be unclear how much errors are tolerant in practical use cases of this method. \n\n3) To interpret the RMSE and MAE values, Table 2 has the value for 'Zero'. This is for a constant predictor always returning zeros for any inputs. However, given that the loss is MSE, constant prediction values should be the average counts in the training data, not zero. \n"}, "tcdate": 1571764937340}, {"id": "r1x1gwAcYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1979/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposed NN based subgraph counting. By using synthetically generated graphs, NN learns the number of occurences of a given queried graph called 'pattern'. The author proposes a specific architecture for learning the count based on the multi-head attention method. The authors empirically evaluated the performance on the synthetic dataset.\n\nThe problem setting would be interesting. Applying NN to counting a subgraph is novel as far as I know. My current concerns are mainly on the appropriateness of the experimental evaluation. \n\nIn the tables, the trivial baseline 'Zero' is shown as F1_zero = 0, but is this correct? I think this should be non-zero. If zero is 'positive' in F1_zero, recall is 1 and precision is 0.75 (because the author set 75% of data as zero). F-score is harmonic mean of them, which is 0.86.\n\nRMSE and MAE of the Zero prediction is shown, but the more standard baseline of the error would be a constant prediction (e.g., the average of test points is often used, which can evaluate how much variance can be explained by the model).\n\nWhy were 75 percent of countings set as 0 in the evaluation dataset? This rate is seemingly a bit large for the evaluation purpose. I guess that when this percentage is much more smaller, MSE would increase. In other words, current MSE/MAE values might be underestiamted compared with when all the test points have non-zero countings.\n\nThe evaluation is only for synthetic dataset for which generating process is designed by the authors. If possible, evaluation on benchmark graph datasets would be convincing though creating the ground truth might be difficult for larger graphs.\n\nMinor comment:\nAt the third line of Sec 3.2: '|V_G| is the number of pattern nodes' should be |V_P|.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposed NN based subgraph counting. By using synthetically generated graphs, NN learns the number of occurences of a given queried graph called 'pattern'. The author proposes a specific architecture for learning the count based on the multi-head attention method. The authors empirically evaluated the performance on the synthetic dataset.\n\nThe problem setting would be interesting. Applying NN to counting a subgraph is novel as far as I know. My current concerns are mainly on the appropriateness of the experimental evaluation. \n\nIn the tables, the trivial baseline 'Zero' is shown as F1_zero = 0, but is this correct? I think this should be non-zero. If zero is 'positive' in F1_zero, recall is 1 and precision is 0.75 (because the author set 75% of data as zero). F-score is harmonic mean of them, which is 0.86.\n\nRMSE and MAE of the Zero prediction is shown, but the more standard baseline of the error would be a constant prediction (e.g., the average of test points is often used, which can evaluate how much variance can be explained by the model).\n\nWhy were 75 percent of countings set as 0 in the evaluation dataset? This rate is seemingly a bit large for the evaluation purpose. I guess that when this percentage is much more smaller, MSE would increase. In other words, current MSE/MAE values might be underestiamted compared with when all the test points have non-zero countings.\n\nThe evaluation is only for synthetic dataset for which generating process is designed by the authors. If possible, evaluation on benchmark graph datasets would be convincing though creating the ground truth might be difficult for larger graphs.\n\nMinor comment:\nAt the third line of Sec 3.2: '|V_G| is the number of pattern nodes' should be |V_P|.\n"}, "tcdate": 1571641062579}], "openreview_url": "https://openreview.net/forum?id=HJx-akSKPS", "arxiv_id": "1912.11589", "paper_pdf": "papers/HJx-akSKPS.pdf", "paper_pdf_sha256": "d2addbe7c82fc089459e7d507bbfe69d99f78e4782849341d2fb67f83fb1cbc7", "paper_pdf_bytes": 2885761, "paper_pdf_source": "openreview", "code_url": "https://github.com/HKUST-KnowComp/NeuralSubgraphCounting", "code_repository": "HKUST-KnowComp/NeuralSubgraphCounting", "code_commit": "22df62d59e112716a22f80db408fcddfc95da5c8", "code_archive": "repos/HJx-akSKPS.zip", "code_archive_sha256": "dbafb1d45ff98e07f711c6b11635b2b6bafac00b43736e61d2a502d9f02a47a4", "code_archive_bytes": 211240, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 199, "github_languages": {"Python": 318050}, "github_archived": false, "github_pushed_at": "2024-07-25T10:59:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-subgraph-isomorphism-counting-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkxn7nR5KX", "year": 2019, "status": "rejected", "title": "Incremental Few-Shot Learning with Attention Attractor Networks", "authors": ["Mengye Ren", "Renjie Liao", "Ethan Fetaya", "Richard S. Zemel"], "authorids": ["mren@cs.toronto.edu", "rjliao@cs.toronto.edu", "ethanf@cs.toronto.edu", "zemel@cs.toronto.edu"], "authors_source": "OpenReview API", "abstract": "Machine learning classifiers are often trained to recognize a set of pre-defined classes. However,\nin many real applications, it is often desirable to have the flexibility of learning additional\nconcepts, without re-training on the full training set. This paper addresses this problem,\nincremental few-shot learning, where a regular classification network has already been trained to\nrecognize a set of base classes; and several extra novel classes are being considered, each with\nonly a few labeled examples. After learning the novel classes, the model is then evaluated on the\noverall performance of both base and novel classes. To this end, we propose a meta-learning model,\nthe Attention Attractor Network, which regularizes the learning of novel classes. In each episode,\nwe train a set of new weights to recognize novel classes until they converge, and we show that the\ntechnique of recurrent back-propagation can back-propagate through the optimization process and\nfacilitate the learning of the attractor network regularizer. We demonstrate that the learned\nattractor network can recognize novel classes while remembering old classes without the need to\nreview the original training set, outperforming baselines that do not rely on an iterative\noptimization process.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "B1lJPA1hhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1396/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work addresses incremental few-shot learning that learns novel classes without forgetting old classes, which is interesting and different from conventional few-shot learning that considers only the few-shot learning task of interest. This problem is also related closely to the important problem of life-long learning. \n\nThis work presents an interesting framework based on meta-learning by learning to learn how to attend to the old classes using an attention mechanism. Experimental results also show improvement over two related works on incremental few-shot learning. The writing is quite clear. Some concerns, especially its novelty, are listed below.  \n\n1. The novelty appears to be limited. The presented framework looks quite similar to the recent work \n\nSpyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. CVPR'18\n\nthat addresses the same problem in a similar manner: 1) learn a base feature extractor and classifier; and then 2) attend to old classes also via meta-learning and attention mechanism.  \nAs mentioned by the authors, \"The main difference to this work is that we use an iterative optimization to compute W_b\". More discussions on the iterative optimization and why it matters may be helpful.\n\nAnother related work is \"Deep Meta-Learning: Learning to Learn in the Concept Space\", Arxiv'18, that also relies on an external base classes for few-shot learning. Similar to the proposed research, it also learns a feature extractor and a classifier from the base classes, which are used to regularize the learning of novel classes, in an end-to-end meta-learning manner. Extending it for the incremental setting seems natural. \n\n2. To learn a few novel classes, all U_k on old classes are relearned, which seems quite time-consuming with a large vocabulary of base classes.\n\n3. To learn a few novel classes, old data on base classes are still required, which seems different from how humans learn -- humans learn novel concepts solely from a few examples without forgetting old concepts, without requiring examples on old concepts.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The problem of incremental few-shot learning is interesting and the presented meta-learning method seems to be effective, but the novelty is limited.  ", "review": "This work addresses incremental few-shot learning that learns novel classes without forgetting old classes, which is interesting and different from conventional few-shot learning that considers only the few-shot learning task of interest. This problem is also related closely to the important problem of life-long learning. \n\nThis work presents an interesting framework based on meta-learning by learning to learn how to attend to the old classes using an attention mechanism. Experimental results also show improvement over two related works on incremental few-shot learning. The writing is quite clear. Some concerns, especially its novelty, are listed below.  \n\n1. The novelty appears to be limited. The presented framework looks quite similar to the recent work \n\nSpyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. CVPR'18\n\nthat addresses the same problem in a similar manner: 1) learn a base feature extractor and classifier; and then 2) attend to old classes also via meta-learning and attention mechanism.  \nAs mentioned by the authors, \"The main difference to this work is that we use an iterative optimization to compute W_b\". More discussions on the iterative optimization and why it matters may be helpful.\n\nAnother related work is \"Deep Meta-Learning: Learning to Learn in the Concept Space\", Arxiv'18, that also relies on an external base classes for few-shot learning. Similar to the proposed research, it also learns a feature extractor and a classifier from the base classes, which are used to regularize the learning of novel classes, in an end-to-end meta-learning manner. Extending it for the incremental setting seems natural. \n\n2. To learn a few novel classes, all U_k on old classes are relearned, which seems quite time-consuming with a large vocabulary of base classes.\n\n3. To learn a few novel classes, old data on base classes are still required, which seems different from how humans learn -- humans learn novel concepts solely from a few examples without forgetting old concepts, without requiring examples on old concepts.  ", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1541303895069}, {"id": "BJlnllqsn7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1396/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a novel few-shot learning method that achieves better overall accuracies on base and novel classes. The key idea is to regularize the learning of novel classes such that base classes are not forgotten. \n\nI mainly have the following two concerns. \n\n-In Table 2, I observe that performance on novel classess is actually not improved. The main improvement lies in overall accuracy. As numbers of training samples between base and novel classes are not balanced, there must be some trade-off between  obtaining better performance on base or novel classes. For instance, stopping early when training on novel classes would result in high base accuracy but low novel accuracy. Fine-tuning on novel classes for more iterations would lead to high novel accuracy but  low base accuracy. Such trade-off can be also controlled by simply over-sampling novel or base classes.  I would suggest the authors to study more on understanding this trade-off. In addition, another naive baseline is to train a softmax classifier at the second stage on both base and novel class training samples and sample mini-batch by uniformly sampling over novel and base classes.  \n\n-The following two papers extensively studied the problem of achieving better overall accuracies on base and novel classes. Including comparison and discussion with those two papers will enhance this paper further. \nLow-Shot Learning from Imaginary Data\nlow-shot visual recognition by shrinking and hallucinating features", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Import discussions missing", "review": "This paper proposes a novel few-shot learning method that achieves better overall accuracies on base and novel classes. The key idea is to regularize the learning of novel classes such that base classes are not forgotten. \n\nI mainly have the following two concerns. \n\n-In Table 2, I observe that performance on novel classess is actually not improved. The main improvement lies in overall accuracy. As numbers of training samples between base and novel classes are not balanced, there must be some trade-off between  obtaining better performance on base or novel classes. For instance, stopping early when training on novel classes would result in high base accuracy but low novel accuracy. Fine-tuning on novel classes for more iterations would lead to high novel accuracy but  low base accuracy. Such trade-off can be also controlled by simply over-sampling novel or base classes.  I would suggest the authors to study more on understanding this trade-off. In addition, another naive baseline is to train a softmax classifier at the second stage on both base and novel class training samples and sample mini-batch by uniformly sampling over novel and base classes.  \n\n-The following two papers extensively studied the problem of achieving better overall accuracies on base and novel classes. Including comparison and discussion with those two papers will enhance this paper further. \nLow-Shot Learning from Imaginary Data\nlow-shot visual recognition by shrinking and hallucinating features", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541279732333}, {"id": "Hyeike3c3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1396/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper addresses the incremental few-shot learning problem where a model starts with base network and then introduces the novel classes, building a connection between novel and base classes via an attention module.\n\nStrengths:\n+ clear writing. \n+ the experiments are compared with related work and the ablation studies can verify the effectiveness of the proposed (or \"introduced\" would be a precise term) recurrent BP.\n\nWeakness:\n\n- [Novelty]\nThe paper title is called attention attractor network, which shares very relevance to previous CVPR work (Gidaris & Komodakis, 2018). So the first thing I was looking for is the clear description of the difference between these two. Unfortunately, in related work, authors mention the CVPR work without stating the difference (last few lines in Section 2). As such, I don't see much novelty in the paper compared with previous work. Eqn. (7)-(10) explicitly describes the attention formula. What's the distinction from the CVPR work?\n\n- [Motivation of the regularizer using Recurrent BP is not clear]\nThe use of recurrent BP is probably the most distinction from previous work. However, I don't see a clear description on why such a technique is necessary.\n\nStarting from the first line in Section 3.3, \"since there is no closed-form of the regularizer in Eqn (13)\", E needs BPTT or the introduced recurrent BP. This part is simply a re-adaption of other algorithms. A very simple question is, how about use other regularizers to replace Eqn (13)? \n\n- [Some experiments missing]\nThe experiments section 4.6 uses a case of None and \"best WD\" to address some of my concerns. This is good. Does the \"gamma random\" indicates only E is used without the ||W||^2? why the best WD for one-shot is zero? This implies the model is best for applying no weight decay?\n\nWhat's the effect of using the recurrent BP technique to the CVPR work? Is there some similar improvement? If yes, then the paper makes some contribution by the regularization. If not, what's the reason?\n\nHow about using the truncated BPTT with a larger T?\n\nIn general, I think the recurrent BP part should be the highlight of the paper and yet authors fail to spread such a spirit in the abstract or title. And there are some experiments missed as I mentioned above.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Limited novelty and unclear motivation", "review": "The paper addresses the incremental few-shot learning problem where a model starts with base network and then introduces the novel classes, building a connection between novel and base classes via an attention module.\n\nStrengths:\n+ clear writing. \n+ the experiments are compared with related work and the ablation studies can verify the effectiveness of the proposed (or \"introduced\" would be a precise term) recurrent BP.\n\nWeakness:\n\n- [Novelty]\nThe paper title is called attention attractor network, which shares very relevance to previous CVPR work (Gidaris & Komodakis, 2018). So the first thing I was looking for is the clear description of the difference between these two. Unfortunately, in related work, authors mention the CVPR work without stating the difference (last few lines in Section 2). As such, I don't see much novelty in the paper compared with previous work. Eqn. (7)-(10) explicitly describes the attention formula. What's the distinction from the CVPR work?\n\n- [Motivation of the regularizer using Recurrent BP is not clear]\nThe use of recurrent BP is probably the most distinction from previous work. However, I don't see a clear description on why such a technique is necessary.\n\nStarting from the first line in Section 3.3, \"since there is no closed-form of the regularizer in Eqn (13)\", E needs BPTT or the introduced recurrent BP. This part is simply a re-adaption of other algorithms. A very simple question is, how about use other regularizers to replace Eqn (13)? \n\n- [Some experiments missing]\nThe experiments section 4.6 uses a case of None and \"best WD\" to address some of my concerns. This is good. Does the \"gamma random\" indicates only E is used without the ||W||^2? why the best WD for one-shot is zero? This implies the model is best for applying no weight decay?\n\nWhat's the effect of using the recurrent BP technique to the CVPR work? Is there some similar improvement? If yes, then the paper makes some contribution by the regularization. If not, what's the reason?\n\nHow about using the truncated BPTT with a larger T?\n\nIn general, I think the recurrent BP part should be the highlight of the paper and yet authors fail to spread such a spirit in the abstract or title. And there are some experiments missed as I mentioned above.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541222371244}], "openreview_url": "https://openreview.net/forum?id=rkxn7nR5KX", "arxiv_id": "1810.07218", "paper_pdf": "papers/rkxn7nR5KX.pdf", "paper_pdf_sha256": "2557a3b43bffa15fb16c8f1a6bb470d9e846d7fa02a65e15254a0d12d0220dba", "paper_pdf_bytes": 316347, "paper_pdf_source": "openreview", "code_url": "https://github.com/renmengye/inc-few-shot-attractor-public", "code_repository": "renmengye/inc-few-shot-attractor-public", "code_commit": "c560d5a81480cb22d903fa746ab0cfc2eb964e4c", "code_archive": "repos/rkxn7nR5KX.zip", "code_archive_sha256": "9f7bf7f8ecdcbdb5e124444a9ca7f97858d430b700f15735ba0a8851066ee39c", "code_archive_bytes": 5738270, "code_file_count": 38, "code_extensions": {".py": 37, ".sh": 1}, "github_disk_usage_kb": 5572, "github_languages": {"Python": 316748, "Shell": 278, "Makefile": 136}, "github_archived": false, "github_pushed_at": "2020-03-19T18:52:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/incremental-few-shot-learning-with-attention"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tDNr5rt47g", "year": 2026, "status": "rejected", "title": "Computational Reasoning of Large Language Models", "authors": ["Haitao Wu", "Zongbo Han", "Joey Tianyi Zhou", "Huaxi Huang", "Changqing Zhang"], "authorids": ["~Haitao_Wu1", "~Zongbo_Han1", "~Joey_Tianyi_Zhou1", "~Huaxi_Huang1", "~Changqing_Zhang1"], "authors_source": "OpenReview API", "abstract": "With the rapid development and widespread application of Large Language Models (LLMs), multidimensional evaluation has become increasingly critical. However, current evaluations are often domain-specific and overly complex, limiting their effectiveness as cross-domain proxies for core capabilities. To address these limitations and enable a unified and simple evaluation framework, an ideal proxy task should target a basic capability that generalizes across tasks and is independent of domain-specific knowledge. Turing machine provides a powerful theoretical lens by reducing complex processes to basic, domain-agnostic computational operations. This perspective offers a principled framework for evaluating foundational computational abilities essential to a wide range of tasks, particularly those involving complex, multi-step reasoning such as mathematics.\nMotivated by this abstraction, we introduce \\textbf{Turing Machine Bench}, a benchmark designed to assess the ability of LLMs to \\textbf{strictly follow rules} and \\textbf{accurately manage internal states} for multi-step, referred to as \\textbf{computational reasoning}.\nTMBench incorporates four key features: self-contained and knowledge-agnostic reasoning, a minimalistic multi-step structure, controllable difficulty, and a solid theoretical foundation based on Turing machine.\nEmpirical results demonstrate that TMBench serves as an effective proxy for evaluating computational reasoning on representative LLMs. It produces clear step-wise accuracy curves, revealing LLMs' ability to execute multi-step reasoning processes. \nBy analyzing performance trends across TMBench and established reasoning benchmarks, we find strong correlations with real-world tasks, bridging real-task evaluation with basic ability assessment. \nThese findings suggest that TMBench holds potential as a cross-domain dimension for evaluating reasoning in LLMs.\nCode and data are available at \\href{https://anonymous.4open.science/r/Turing-Machine-Bench-Anonymous-EBF4/}{Repo}.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "uLEijLgExM", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5606/Reviewer_m5Wv"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes TMBench, a benchmark for evaluation the \"computational reasoning\" ability of LLMs in a domain-independent and universal way. The authors make an argument for why this benchmark adds value to already existing set of benchmarks assessing various reasoning strengths of LLMs. They do an extensive evaluation of various LLMs and also compare correlation between TMBench scores and scores on other, as they call it \"real-world\", benchmarks.", "review_text": "This paper proposes TMBench, a benchmark for evaluation the \"computational reasoning\" ability of LLMs in a domain-independent and universal way. The authors make an argument for why this benchmark adds value to already existing set of benchmarks assessing various reasoning strengths of LLMs. They do an extensive evaluation of various LLMs and also compare correlation between TMBench scores and scores on other, as they call it \"real-world\", benchmarks.", "strengths": "* TMBench is a novel benchmark for assessing the reasoning abilities of models in a domain-independent fashion.\n\n* The authors show that performance on TMBench correlates with exisitng benchmarks for math reasoning, etc.\n\n* The synthetic nature of TMBench allows one to generate instances of varying difficulty, similar to some earlier works such as on grid logic puzzles and datasets inspired by constraint satisfaction problems. This allows the method to be \"future proof\" in the sense that if the latest LLMs solve the current instances, one can easily generate harder ones, again challenging the models further.\n\n* It is useful to know that changing the alphabet (Roman, numeral, Greek, etc.) doesn't really affect the performance of the tested LLMs, suggesting that they aren't solving these tasks via pattern matching and memorization.", "weaknesses": "* I am not convinced by the authors' pitch that this is the first benchmark that attempts to measure the computational reasoning ability of LLMs. There have been many attempts, some specifically targeting domain-agnostic assessments, such as multiple papers proposing various kinds of grid logic puzzles as benchmarks. \n\n* While I see that being able to simulate a Turing Machine is an interesting ability to measure, the authors miss an important distinction -- namely that *simulating* a given Turing Machine is different from *learning* (from data) what the rules of a machine should be and then being able to execute that specific Turing Machine, which is how machine learning models operate. So, for example, even if LLMs do well on the TMBench benchmark, it does not mean they can learn to (or be parameterized to) solve a specific task such as math reasoning or program execution. Additionally, in some ways, specific tasks such as solving math problems are much closer to real needs than simulating a (given) abstract Turing Machine.\n\n* The authors also don't clearly distinguish between chain-of-thought \"reasoning\" LLMs vs. non-reasoning LLMs. This distinction is crucial as, from a formal standpoint, it is now understood that basic transformers can only solve a subset of polynomial-time problems (namely, those that are highly parallelizable) [1,2], and that with chain-of-thought, they become universal [3,4]. The paper does add a complementary angle, namely a systematic empirical assessment of transformer-based LLMs. It would be more valuable if they could frame the discussion and empirical findings in light of what's known theoretically about these models.\n\n* I think the title is just too generic -- it makes the paper sound much broader than what its main contribution is, which is to propose a new new benchmark (TMBench) and report LLM performance results on it. At first glance, it also makes one think that this might be the first paper addressing the computational reasoning ability of LLMs, which clearly isn't the case.\n\n* Rule following ability of LLMs has been explored via other benchmarks as well, such as RuleTaker [5], PRover [6], and even extensions of the bAbI benchmark from many years ago. Since the author emphasize rule following ability, placing TMBench in the context of these earlier works on rule following would be appropriate.\n\n* Some of the empirical results are confusing or confusingly presented. E.g., at the end of section 4.2, it's noted that about Fig 1 that models larger than 14B size drop accuracy linearly as the number of simulated steps grows. However, (a) of the 12 models plotted in fig 1, only 1 model has size under 14B, and (b) unless the other plots are extended much farther to the right (till the accuracy drops close to zero), it isn't really possible to tell if the drop is linear or not (e.g., the drop for Qwen2.5-32B looks linear for the first 10 steps, but after that we see it isn't really linear). As another example, the choice of a column plot for Fig 3(a) is odd; I would have chosen some way to illustrate the distribution (e.g., with max correct steps on the x-axis, or a sorted table).\n\n* The dataset itself is very succinctly described (in one paragraph in section 4.1). It's unclear whether the generated problems are hard or easy, whether they correspond to any naturally interesting problems or not, etc. As such, it's not totally clear how the results should be interpreted -- what does it mean to be able to do well (or do poorly) on these 100 problem instances?\n\n[1] The Parallelism Tradeoff: Limitations of Log-Precision Transformers, Merrill et al., 2022\n\n[2] Transformers in Uniform TC0, Chiang, 2024\n\n[3] The Expressive Power of Transformers with Chain of Thought, Merrill et al., 2024\n\n[4] Chain of Thought Empowers Transformers to Solve Inherently Serial Problems, Li et al., 2024\n\n[5] Transformers as Soft Reasoners over Language, Clark et al., 2020\n\n[6] PRover: Proof Generation for Interpretable Reasoning over Rules, Saha et al., 2020", "questions": "I don't have specific questions, but there are some included in the \"weaknesses\" section that the authors might want to address.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes TMBench, a benchmark for evaluation the \"computational reasoning\" ability of LLMs in a domain-independent and universal way. The authors make an argument for why this benchmark adds value to already existing set of benchmarks assessing various reasoning strengths of LLMs. They do an extensive evaluation of various LLMs and also compare correlation between TMBench scores and scores on other, as they call it \"real-world\", benchmarks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "* TMBench is a novel benchmark for assessing the reasoning abilities of models in a domain-independent fashion.\n\n* The authors show that performance on TMBench correlates with exisitng benchmarks for math reasoning, etc.\n\n* The synthetic nature of TMBench allows one to generate instances of varying difficulty, similar to some earlier works such as on grid logic puzzles and datasets inspired by constraint satisfaction problems. This allows the method to be \"future proof\" in the sense that if the latest LLMs solve the current instances, one can easily generate harder ones, again challenging the models further.\n\n* It is useful to know that changing the alphabet (Roman, numeral, Greek, etc.) doesn't really affect the performance of the tested LLMs, suggesting that they aren't solving these tasks via pattern matching and memorization.", "weaknesses": "* I am not convinced by the authors' pitch that this is the first benchmark that attempts to measure the computational reasoning ability of LLMs. There have been many attempts, some specifically targeting domain-agnostic assessments, such as multiple papers proposing various kinds of grid logic puzzles as benchmarks. \n\n* While I see that being able to simulate a Turing Machine is an interesting ability to measure, the authors miss an important distinction -- namely that *simulating* a given Turing Machine is different from *learning* (from data) what the rules of a machine should be and then being able to execute that specific Turing Machine, which is how machine learning models operate. So, for example, even if LLMs do well on the TMBench benchmark, it does not mean they can learn to (or be parameterized to) solve a specific task such as math reasoning or program execution. Additionally, in some ways, specific tasks such as solving math problems are much closer to real needs than simulating a (given) abstract Turing Machine.\n\n* The authors also don't clearly distinguish between chain-of-thought \"reasoning\" LLMs vs. non-reasoning LLMs. This distinction is crucial as, from a formal standpoint, it is now understood that basic transformers can only solve a subset of polynomial-time problems (namely, those that are highly parallelizable) [1,2], and that with chain-of-thought, they become universal [3,4]. The paper does add a complementary angle, namely a systematic empirical assessment of transformer-based LLMs. It would be more valuable if they could frame the discussion and empirical findings in light of what's known theoretically about these models.\n\n* I think the title is just too generic -- it makes the paper sound much broader than what its main contribution is, which is to propose a new new benchmark (TMBench) and report LLM performance results on it. At first glance, it also makes one think that this might be the first paper addressing the computational reasoning ability of LLMs, which clearly isn't the case.\n\n* Rule following ability of LLMs has been explored via other benchmarks as well, such as RuleTaker [5], PRover [6], and even extensions of the bAbI benchmark from many years ago. Since the author emphasize rule following ability, placing TMBench in the context of these earlier works on rule following would be appropriate.\n\n* Some of the empirical results are confusing or confusingly presented. E.g., at the end of section 4.2, it's noted that about Fig 1 that models larger than 14B size drop accuracy linearly as the number of simulated steps grows. However, (a) of the 12 models plotted in fig 1, only 1 model has size under 14B, and (b) unless the other plots are extended much farther to the right (till the accuracy drops close to zero), it isn't really possible to tell if the drop is linear or not (e.g., the drop for Qwen2.5-32B looks linear for the first 10 steps, but after that we see it isn't really linear). As another example, the choice of a column plot for Fig 3(a) is odd; I would have chosen some way to illustrate the distribution (e.g., with max correct steps on the x-axis, or a sorted table).\n\n* The dataset itself is very succinctly described (in one paragraph in section 4.1). It's unclear whether the generated problems are hard or easy, whether they correspond to any naturally interesting problems or not, etc. As such, it's not totally clear how the results should be interpreted -- what does it mean to be able to do well (or do poorly) on these 100 problem instances?\n\n[1] The Parallelism Tradeoff: Limitations of Log-Precision Transformers, Merrill et al., 2022\n\n[2] Transformers in Uniform TC0, Chiang, 2024\n\n[3] The Expressive Power of Transformers with Chain of Thought, Merrill et al., 2024\n\n[4] Chain of Thought Empowers Transformers to Solve Inherently Serial Problems, Li et al., 2024\n\n[5] Transformers as Soft Reasoners over Language, Clark et al., 2020\n\n[6] PRover: Proof Generation for Interpretable Reasoning over Rules, Saha et al., 2020", "questions": "I don't have specific questions, but there are some included in the \"weaknesses\" section that the authors might want to address.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762188003856}, {"id": "V0TBKo8lkn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5606/Reviewer_wEq4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces the Turing Machine Benchmark (TMBench), which evaluates the ‘computational reasoning’ capabilities of large language models by simulating simplified Turing machines (m-Tag systems). Each task requires the model to incrementally update a symbol queue according to fixed rules. The authors define granular metrics—step accuracy, step-weighted accuracy, and pass rate—to quantify correctness per step and overall success. Key findings include: accuracy exhibits a systematic decline with increasing steps, and TMBench pass rates demonstrate strong empirical correlation with standard reasoning benchmarks.", "review_text": "This paper introduces the Turing Machine Benchmark (TMBench), which evaluates the ‘computational reasoning’ capabilities of large language models by simulating simplified Turing machines (m-Tag systems). Each task requires the model to incrementally update a symbol queue according to fixed rules. The authors define granular metrics—step accuracy, step-weighted accuracy, and pass rate—to quantify correctness per step and overall success. Key findings include: accuracy exhibits a systematic decline with increasing steps, and TMBench pass rates demonstrate strong empirical correlation with standard reasoning benchmarks.", "strengths": "1.Introducing TMBench as a domain-agnostic proxy task appears novel. By basing it on m-Tag (a simple Turing machine variant), the authors tie LLM evaluation to well-understood computational primitives. This abstraction is a strong point, as it isolates rule-following ability from background knowledge.\n\n2.A wide array of models (from small open models to top proprietary ones) are evaluated under consistent settings. The experiments cover not only basic performance but also varied conditions (unlimited steps, temperature changes, different alphabets, task difficulty via deletion count). This thoroughness lends credibility to the claims.\n\n3.Finding a high Pearson correlation (≈0.882) between TMBench pass rate and averaged reasoning scores is a compelling result. It suggests that this abstract benchmark may indeed reflect “real” reasoning ability, thus strengthening the paper’s significance. The authors also check residuals (Q-Q plots) to validate regression assumptions.", "weaknesses": "1.TMBench tasks are highly artificial (m-Tag simulations) and quite different from natural language problems. It is unclear how well success on these tasks truly reflects broader reasoning skills beyond what is measured. The paper posits “computational generality”, but only one type of formalism is used.\n\n2. The correlation claims are based on 12 data points (one per model), so even though p<0.05 is reported, this may not be robust. A single outlier model could sway the result. The analysis might benefit from confidence intervals or non-parametric checks.\n\n3. The paper lacks simple baselines. For example, it would strengthen conclusions to compare LLMs against trivial strategies (like a rule-based simulator) or to random guessing. Also, only one prompting style (1-shot example) is used; variations (0-shot, few-shot with explanations) are not explored.", "questions": "1.How exactly is a prediction counted as “correct” at each step? Does the model have to output the entire queue state string perfectly, including brackets and commas?\n\n2. Given only 12 models in the correlation, have you tested the robustness of the Pearson r? For example, what is the confidence interval of the correlation, or how would results change if one model is removed?\n\n3. You evaluate up to 30 steps by default. How did you choose the 30-step limit? If tasks ran longer (or shorter), would model rankings or metric behavior change qualitatively?\n\n4. In the SFT experiment (Sec.4.4/Table 10), how was the synthetic fine-tuning data generated? Is there a risk that models learn just the training tasks rather than true reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Turing Machine Benchmark (TMBench), which evaluates the ‘computational reasoning’ capabilities of large language models by simulating simplified Turing machines (m-Tag systems). Each task requires the model to incrementally update a symbol queue according to fixed rules. The authors define granular metrics—step accuracy, step-weighted accuracy, and pass rate—to quantify correctness per step and overall success. Key findings include: accuracy exhibits a systematic decline with increasing steps, and TMBench pass rates demonstrate strong empirical correlation with standard reasoning benchmarks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.Introducing TMBench as a domain-agnostic proxy task appears novel. By basing it on m-Tag (a simple Turing machine variant), the authors tie LLM evaluation to well-understood computational primitives. This abstraction is a strong point, as it isolates rule-following ability from background knowledge.\n\n2.A wide array of models (from small open models to top proprietary ones) are evaluated under consistent settings. The experiments cover not only basic performance but also varied conditions (unlimited steps, temperature changes, different alphabets, task difficulty via deletion count). This thoroughness lends credibility to the claims.\n\n3.Finding a high Pearson correlation (≈0.882) between TMBench pass rate and averaged reasoning scores is a compelling result. It suggests that this abstract benchmark may indeed reflect “real” reasoning ability, thus strengthening the paper’s significance. The authors also check residuals (Q-Q plots) to validate regression assumptions.", "weaknesses": "1.TMBench tasks are highly artificial (m-Tag simulations) and quite different from natural language problems. It is unclear how well success on these tasks truly reflects broader reasoning skills beyond what is measured. The paper posits “computational generality”, but only one type of formalism is used.\n\n2. The correlation claims are based on 12 data points (one per model), so even though p<0.05 is reported, this may not be robust. A single outlier model could sway the result. The analysis might benefit from confidence intervals or non-parametric checks.\n\n3. The paper lacks simple baselines. For example, it would strengthen conclusions to compare LLMs against trivial strategies (like a rule-based simulator) or to random guessing. Also, only one prompting style (1-shot example) is used; variations (0-shot, few-shot with explanations) are not explored.", "questions": "1.How exactly is a prediction counted as “correct” at each step? Does the model have to output the entire queue state string perfectly, including brackets and commas?\n\n2. Given only 12 models in the correlation, have you tested the robustness of the Pearson r? For example, what is the confidence interval of the correlation, or how would results change if one model is removed?\n\n3. You evaluate up to 30 steps by default. How did you choose the 30-step limit? If tasks ran longer (or shorter), would model rankings or metric behavior change qualitatively?\n\n4. In the SFT experiment (Sec.4.4/Table 10), how was the synthetic fine-tuning data generated? Is there a risk that models learn just the training tasks rather than true reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761828953966}, {"id": "z2dY2PgWWi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5606/Reviewer_crJ8"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper presents a new benchmark for testing the in-context reasoning capabilities of large language models, using recursively enumerable languages (those accepted by a Turing Machine). They use their benchmark to evaluate 35 large language models, finding that Gemini-2.5-Pro is the best performer.", "review_text": "This paper presents a new benchmark for testing the in-context reasoning capabilities of large language models, using recursively enumerable languages (those accepted by a Turing Machine). They use their benchmark to evaluate 35 large language models, finding that Gemini-2.5-Pro is the best performer.", "strengths": "The paper has the following strengths:\n* A commendably large set of language models are evaluated on the proposed benchmark.\n* The methodological decisions, such as focusing on m-tag systems, are defensible, given the limited space in a conference paper.", "weaknesses": "I am recommending reject for this paper simply because it does not engage with the state of the art when it comes to (a) in-context learning in language models and (b) using formal languages/automata theory to study the capabilities of these models. Essentially, I believe that most of what the authors propose has already been done, with a greater degree of formal rigour. While I agree that no-one has proposed quite the same benchmark as the authors do here, I don't think that what they offer is altogether that novel. I provide references to the relevant literature below:\n\n### In-Context Learning\n\nThe authors are essentially proposing a benchmark of what pre-trained transformers models can learn in-context, which is a live and active research area that the authors do not engage with at all:\n\nAkyürek, E., Wang, B., Kim, Y., & Andreas, J. (2024). In-context language learning: Architectures and algorithms. arXiv preprint arXiv:2401.12973.\n\nCoda-Forno, J., Binz, M., Akata, Z., Botvinick, M., Wang, J., & Schulz, E. (2023). Meta-in-context learning in large language models. Advances in Neural Information Processing Systems, 36, 65189-65201.\n\nLampinen, A. K., Chan, S. C., Singh, A. K., & Shanahan, M. (2024). The broader spectrum of in-context learning. arXiv preprint arXiv:2412.03782.\n\nVon Oswald, J., Niklasson, E., Randazzo, E., Sacramento, J., Mordvintsev, A., Zhmoginov, A., & Vladymyrov, M. (2023, July). Transformers learn in-context by gradient descent. In International Conference on Machine Learning (pp. 35151-35174). PMLR.\n\nXie, S. M., Raghunathan, A., Liang, P., & Ma, T. (2021). An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080.\n\n### Formal Automata Theory\n\nAutomata theory has been used extensively to study the capabilities of language models. I encourage the authors to thoroughly review the following papers and the references that they cite:\n\nAckerman, J., & Cybenko, G. (2020). A survey of neural networks and formal languages. arXiv preprint arXiv:2006.01338.\n\nButoi, A., Khalighinejad, G., Svete, A., Valvoda, J., Cotterell, R., & DuSell, B. (2024). Training neural networks as recognizers of formal languages. arXiv preprint arXiv:2411.07107.\n\nBorenstein, N., Svete, A., Chan, R., Valvoda, J., Nowak, F., Augenstein, I., ... & Cotterell, R. (2024). What languages are easy to language-model? a perspective from learning probabilistic regular languages. arXiv preprint arXiv:2406.04289.\n\nNowak, F., Svete, A., Butoi, A., & Cotterell, R. (2024). On the representational capacity of neural language models with chain-of-thought reasoning. arXiv preprint arXiv:2406.14197.\n\nLi, J., White, J. C., Sachan, M., & Cotterell, R. (2024). A transformer with stack attention. arXiv preprint arXiv:2405.04515.\n\nSvete, A., Chan, R. S. M., & Cotterell, R. (2024). On efficiently representing regular languages as RNNs. arXiv preprint arXiv:2402.15814.\n\nVoudouris, K., Barron, A., Halina, M., Klein, C., & Patel, M. (2025). Exploring Major Transitions in the Evolution of Biological Cognition With Artificial Neural Networks. arXiv preprint arXiv:2509.13968.\n\nWeiss, G., Goldberg, Y., & Yahav, E. (2018). On the practical computational power of finite precision RNNs for language recognition. arXiv preprint arXiv:1805.04908.", "questions": "* Do the authors expect any capability boosts from conducting post-training (e.g., supervised fine-tuning) on held-out m-tag problems? I would be interested to see how fine-tuning boosts in-context learning performance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a new benchmark for testing the in-context reasoning capabilities of large language models, using recursively enumerable languages (those accepted by a Turing Machine). They use their benchmark to evaluate 35 large language models, finding that Gemini-2.5-Pro is the best performer.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "The paper has the following strengths:\n* A commendably large set of language models are evaluated on the proposed benchmark.\n* The methodological decisions, such as focusing on m-tag systems, are defensible, given the limited space in a conference paper.", "weaknesses": "I am recommending reject for this paper simply because it does not engage with the state of the art when it comes to (a) in-context learning in language models and (b) using formal languages/automata theory to study the capabilities of these models. Essentially, I believe that most of what the authors propose has already been done, with a greater degree of formal rigour. While I agree that no-one has proposed quite the same benchmark as the authors do here, I don't think that what they offer is altogether that novel. I provide references to the relevant literature below:\n\n### In-Context Learning\n\nThe authors are essentially proposing a benchmark of what pre-trained transformers models can learn in-context, which is a live and active research area that the authors do not engage with at all:\n\nAkyürek, E., Wang, B., Kim, Y., & Andreas, J. (2024). In-context language learning: Architectures and algorithms. arXiv preprint arXiv:2401.12973.\n\nCoda-Forno, J., Binz, M., Akata, Z., Botvinick, M., Wang, J., & Schulz, E. (2023). Meta-in-context learning in large language models. Advances in Neural Information Processing Systems, 36, 65189-65201.\n\nLampinen, A. K., Chan, S. C., Singh, A. K., & Shanahan, M. (2024). The broader spectrum of in-context learning. arXiv preprint arXiv:2412.03782.\n\nVon Oswald, J., Niklasson, E., Randazzo, E., Sacramento, J., Mordvintsev, A., Zhmoginov, A., & Vladymyrov, M. (2023, July). Transformers learn in-context by gradient descent. In International Conference on Machine Learning (pp. 35151-35174). PMLR.\n\nXie, S. M., Raghunathan, A., Liang, P., & Ma, T. (2021). An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080.\n\n### Formal Automata Theory\n\nAutomata theory has been used extensively to study the capabilities of language models. I encourage the authors to thoroughly review the following papers and the references that they cite:\n\nAckerman, J., & Cybenko, G. (2020). A survey of neural networks and formal languages. arXiv preprint arXiv:2006.01338.\n\nButoi, A., Khalighinejad, G., Svete, A., Valvoda, J., Cotterell, R., & DuSell, B. (2024). Training neural networks as recognizers of formal languages. arXiv preprint arXiv:2411.07107.\n\nBorenstein, N., Svete, A., Chan, R., Valvoda, J., Nowak, F., Augenstein, I., ... & Cotterell, R. (2024). What languages are easy to language-model? a perspective from learning probabilistic regular languages. arXiv preprint arXiv:2406.04289.\n\nNowak, F., Svete, A., Butoi, A., & Cotterell, R. (2024). On the representational capacity of neural language models with chain-of-thought reasoning. arXiv preprint arXiv:2406.14197.\n\nLi, J., White, J. C., Sachan, M., & Cotterell, R. (2024). A transformer with stack attention. arXiv preprint arXiv:2405.04515.\n\nSvete, A., Chan, R. S. M., & Cotterell, R. (2024). On efficiently representing regular languages as RNNs. arXiv preprint arXiv:2402.15814.\n\nVoudouris, K., Barron, A., Halina, M., Klein, C., & Patel, M. (2025). Exploring Major Transitions in the Evolution of Biological Cognition With Artificial Neural Networks. arXiv preprint arXiv:2509.13968.\n\nWeiss, G., Goldberg, Y., & Yahav, E. (2018). On the practical computational power of finite precision RNNs for language recognition. arXiv preprint arXiv:1805.04908.", "questions": "* Do the authors expect any capability boosts from conducting post-training (e.g., supervised fine-tuning) on held-out m-tag problems? I would be interested to see how fine-tuning boosts in-context learning performance.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761385293382}, {"id": "NF7TKIo2Ks", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5606/Reviewer_b1Yx"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "The paper proposes the Turing Machine Bench (TMBench), which uses an m-Tag system to simulate a Turing machine. It uses self-consistent, knowledge-independent, multi-step rule execution to evaluate the computational reasoning capabilities of large models, and provides controllable difficulty and step-by-step metrics (SWA and Pass Rate). On a variety of open-source and closed-source models, TMBench demonstrates that top models can stably execute dozens of steps while still accumulating error with the number of steps. It also significantly correlates with high-level reasoning benchmarks such as AIME, MATH, and GPQA. These results demonstrate that TMBench can serve as a cross-domain, simple, and unified proxy for reasoning capabilities.", "review_text": "The paper proposes the Turing Machine Bench (TMBench), which uses an m-Tag system to simulate a Turing machine. It uses self-consistent, knowledge-independent, multi-step rule execution to evaluate the computational reasoning capabilities of large models, and provides controllable difficulty and step-by-step metrics (SWA and Pass Rate). On a variety of open-source and closed-source models, TMBench demonstrates that top models can stably execute dozens of steps while still accumulating error with the number of steps. It also significantly correlates with high-level reasoning benchmarks such as AIME, MATH, and GPQA. These results demonstrate that TMBench can serve as a cross-domain, simple, and unified proxy for reasoning capabilities.", "strengths": "1.This paper proposes TMBench based on m-Tag Turing machines as a self-consistent, knowledge-independent unified reasoning evaluation framework, which solves the problem that existing benchmarks mix knowledge and reasoning and lack simple generalizable proxy tasks.\n2. This article provides step-by-step interpretable curves and adjustable difficulty, and empirically proves a strong correlation with AIME/MATH/GPQA, solving the problem that multi-step reasoning processes are difficult to quantify error propagation and difficult to align evaluation with real-world tasks.", "weaknesses": "1. Are 100 samples too few for a general inference benchmark？ Maybe it will limit statistical robustness and extrapolation. Although each sample contains multiple steps of trajectory to increase the data points, the diversity is still limited.\n2. There are some writing problems in the article：\n\n **line 178** “…has gain significant attention…”→“…has gained significant attention…”\n\n**line 458** \"Figure Figure 3c\" --> \"Figure 3c\"\n\n**line 1238** “token distrubition”-->“token distribution”\n\nInconsistent terminology: \"computation reasoning\" and \"computational reasoning\" are used interchangeably. It is recommended to unify them into \"computational reasoning\"", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes the Turing Machine Bench (TMBench), which uses an m-Tag system to simulate a Turing machine. It uses self-consistent, knowledge-independent, multi-step rule execution to evaluate the computational reasoning capabilities of large models, and provides controllable difficulty and step-by-step metrics (SWA and Pass Rate). On a variety of open-source and closed-source models, TMBench demonstrates that top models can stably execute dozens of steps while still accumulating error with the number of steps. It also significantly correlates with high-level reasoning benchmarks such as AIME, MATH, and GPQA. These results demonstrate that TMBench can serve as a cross-domain, simple, and unified proxy for reasoning capabilities.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1.This paper proposes TMBench based on m-Tag Turing machines as a self-consistent, knowledge-independent unified reasoning evaluation framework, which solves the problem that existing benchmarks mix knowledge and reasoning and lack simple generalizable proxy tasks.\n2. This article provides step-by-step interpretable curves and adjustable difficulty, and empirically proves a strong correlation with AIME/MATH/GPQA, solving the problem that multi-step reasoning processes are difficult to quantify error propagation and difficult to align evaluation with real-world tasks.", "weaknesses": "1. Are 100 samples too few for a general inference benchmark？ Maybe it will limit statistical robustness and extrapolation. Although each sample contains multiple steps of trajectory to increase the data points, the diversity is still limited.\n2. There are some writing problems in the article：\n\n **line 178** “…has gain significant attention…”→“…has gained significant attention…”\n\n**line 458** \"Figure Figure 3c\" --> \"Figure 3c\"\n\n**line 1238** “token distrubition”-->“token distribution”\n\nInconsistent terminology: \"computation reasoning\" and \"computational reasoning\" are used interchangeably. It is recommended to unify them into \"computational reasoning\"", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761223786137}], "openreview_url": "https://openreview.net/forum?id=tDNr5rt47g", "arxiv_id": "2504.20771", "paper_pdf": "papers/tDNr5rt47g.pdf", "paper_pdf_sha256": "27d4feece55de04c66fa97dba7f59380e9b72a28c6865e53e746d0523f5c14b1", "paper_pdf_bytes": 488760, "paper_pdf_source": "openreview", "code_url": "https://github.com/HaitaoWuTJU/Turing-Machine-Bench", "code_repository": "HaitaoWuTJU/Turing-Machine-Bench", "code_commit": "9e04fcaec05664aa28fc86f20d9a2bb0e17551ea", "code_archive": "repos/tDNr5rt47g.zip", "code_archive_sha256": "bf46074567abebd4586abad67ee3badd9b914c55488bfa5be3e751b6ca6c5695", "code_archive_bytes": 1212682, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 1174, "github_languages": {"Python": 30570, "Shell": 1705}, "github_archived": false, "github_pushed_at": "2025-05-16T14:14:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/turing-machine-evaluation-for-large-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gDWkImLIKd", "year": 2025, "status": "rejected", "title": "Large Language Model Critics for Execution-Free Evaluation of Code Changes", "authors": ["Aashish Yadavally", "Hoan Nguyen", "Laurent Callot", "Gauthier Guinet"], "authorids": ["~Aashish_Yadavally1", "~Hoan_Nguyen1", "~Laurent_Callot1", "~Gauthier_Guinet1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) offer a promising way forward for automating software engineering tasks, such as bug fixes, feature additions, etc., via multi-step LLM-based agentic workflows. However, existing metrics for evaluating such workflows, mainly  build status and occasionally log analysis, are too sparse and limited in providing the information needed to assess the quality of changes made. In this work, we designed LLM-based critics to derive well-structured and rigorous intermediate/step-level, execution-free evaluation proxies for repo-level code changes. Importantly, we assume access to the gold patch for the problem (i.e., reference-aware) to assess both semantics and executability of generated patches. With the gold test patch as a reference, we predict executability of all editing locations with an accuracy of 91.6%, aggregating which, we can predict the build status in 82.1% of the instances in SWE-bench. In particular, such an execution-focused LLM critic outperforms other reference-free and reference-aware LLM critics by 38.9% to 72.5%. Moreover, we demonstrate the usefulness of such a reference-aware framework in comparing patches generated by different agentic workflows. Finally, we open-source the library developed for this project, which allow further usage for either other agentic workflows or other benchmarks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sn2lDyC42H", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11365/Reviewer_g42S"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes an execution-free, test-aware, LLM-based metric for evaluating code edits. Essentially, the authors prompt an LLM (claude-3-opus) with a candidate patch and an individual test in the test suite, and then the LLM will predict whether or not that patch will pass the given test. Then, they aggregate predictions across all tests in the test suit to assign a final build label. For experiments, the authors rely on SWE-Bench Lite, and trajectories form factory-code-droid, sweagent-gpt4, Gru, and code-story-aide-mixed. At the macro-level, their best approach achieves 71.4% accuracy (72.1 precision, 95.4 recall) with respect to predicting the build status outcome.", "review_text": "This paper proposes an execution-free, test-aware, LLM-based metric for evaluating code edits. Essentially, the authors prompt an LLM (claude-3-opus) with a candidate patch and an individual test in the test suite, and then the LLM will predict whether or not that patch will pass the given test. Then, they aggregate predictions across all tests in the test suit to assign a final build label. For experiments, the authors rely on SWE-Bench Lite, and trajectories form factory-code-droid, sweagent-gpt4, Gru, and code-story-aide-mixed. At the macro-level, their best approach achieves 71.4% accuracy (72.1 precision, 95.4 recall) with respect to predicting the build status outcome.", "strengths": "- The motivation for this work is quite nice and very important. An execution-free metric is definitely useful for fast iteration and in scenarios in which a test environment is not available.\n- The finding that aggregating individual test results works better than having a holistic evaluation across all tests is quite interesting. This could possibly extend to other types of LLM-based evaluations as well (e.g., rather than evaluating across multiple dimensions in a single call, evaluate across each dimension independently and then aggregate results). \n- The analysis with the model's self-reported confidence and test complexity is interesting.", "weaknesses": "- I believe an execution-free metric is most useful in situations in which you do not have a test suite at all or when the existing test suite has low coverage. However, this work requires having a high-quality test suite. The only dataset that the authors evaluate on is SWE-Bench, which comes with Docker images corresponding to the test environments, and so it seems like it is rather straightforward to just execute the tests in the test suite. Therefore, it seems that the impact of this work will be fairly limited. \n- Additionally, from Table 2, the best accuracy that is attained is 71.4% (which is incorrectly claimed as 82.1% in the abstract). From the paper alone, I am not convinced that we can simply replace the execution-based metric with this. Perhaps if the authors had demonstrated that the rankings of the top ~10 models on the SWE-Bench leaderboard remained identical when using the LLM-based metric, it would have been more convincing. Currently, the best approach nearly matches the random baseline in terms of precision (i.e., the LLM-based approach will often say the patch passes tests when it actually does not). And if it is possible to obtain the execution-based score, then this LLM-based metric will likely not be needed at all since it serves to approximate the execution-based metric.\n- A lot of details seem to be missing, misleading, inconsistent, and sometimes incorrect. This makes it difficult to follow and at times even assess the paper:\n\n1. In the abstract, the authors claim an accuracy of 91.6% at the micro-level and 82.1% at the macro-level. However, these are F1 scores, and not accuracy, based on Tables 1 and 2. \n2. For the majority of the first half of the paper, it seems that the authors use the gold code patch which resolves the issue (L017-018, L080-081, L137-141). However, it becomes clear later that the authors only consider the gold *test* patch and the gold code patch is not used in their approach and only in their baselines. \n3. The prompts that the authors use are not given. Additionally, not a single example is provided.\n4. The notion of \"function-level\" is in multiple tables but this is never explained. This also seems to be the best performing, so it is not clear what the best-performing method is actually. Namely, Table 1 has only two rows for \"Isolated, Test-Aware\", one corresponding to \"post-commit functions\" and one augmenting that method with \"function-level.\" However, the two methods that are introduced in the main text are \"context enhancement\" and \"source code\" (which is also referred to as post-commit functions). In L259, it says \"We can see that the LLM critic with context-enhanced patches performs the best, outperforming other baselines by 7.4% and 12.7%\". But \"context-enhanced\" is not in Table 1 and there is only 1 baseline. \n5. In the abstract, the authors motivate this work as an \"intermediate/step-level\" evaluation methodology. This suggests that the intermediate steps in the LLM-based agentic framework could be evaluated. However, this does not seem to be demonstrated in the paper. It is not obvious, but if Figure 4 (right) was intended to demonstrate this, it is not clear. Additionally, it is not clear why this was presented as a plot rather than an aggregate spearman's ranking coefficient. It looks like there is a decent chunk of the density in the negative range here too.", "questions": "- Please consider addressing the points raised above.\n- Have you considered using both the gold test patch and gold code patch together?\n- Did you consider a baseline which just uses the candidate code patch, with no tests at all?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an execution-free, test-aware, LLM-based metric for evaluating code edits. Essentially, the authors prompt an LLM (claude-3-opus) with a candidate patch and an individual test in the test suite, and then the LLM will predict whether or not that patch will pass the given test. Then, they aggregate predictions across all tests in the test suit to assign a final build label. For experiments, the authors rely on SWE-Bench Lite, and trajectories form factory-code-droid, sweagent-gpt4, Gru, and code-story-aide-mixed. At the macro-level, their best approach achieves 71.4% accuracy (72.1 precision, 95.4 recall) with respect to predicting the build status outcome.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The motivation for this work is quite nice and very important. An execution-free metric is definitely useful for fast iteration and in scenarios in which a test environment is not available.\n- The finding that aggregating individual test results works better than having a holistic evaluation across all tests is quite interesting. This could possibly extend to other types of LLM-based evaluations as well (e.g., rather than evaluating across multiple dimensions in a single call, evaluate across each dimension independently and then aggregate results). \n- The analysis with the model's self-reported confidence and test complexity is interesting.", "weaknesses": "- I believe an execution-free metric is most useful in situations in which you do not have a test suite at all or when the existing test suite has low coverage. However, this work requires having a high-quality test suite. The only dataset that the authors evaluate on is SWE-Bench, which comes with Docker images corresponding to the test environments, and so it seems like it is rather straightforward to just execute the tests in the test suite. Therefore, it seems that the impact of this work will be fairly limited. \n- Additionally, from Table 2, the best accuracy that is attained is 71.4% (which is incorrectly claimed as 82.1% in the abstract). From the paper alone, I am not convinced that we can simply replace the execution-based metric with this. Perhaps if the authors had demonstrated that the rankings of the top ~10 models on the SWE-Bench leaderboard remained identical when using the LLM-based metric, it would have been more convincing. Currently, the best approach nearly matches the random baseline in terms of precision (i.e., the LLM-based approach will often say the patch passes tests when it actually does not). And if it is possible to obtain the execution-based score, then this LLM-based metric will likely not be needed at all since it serves to approximate the execution-based metric.\n- A lot of details seem to be missing, misleading, inconsistent, and sometimes incorrect. This makes it difficult to follow and at times even assess the paper:\n\n1. In the abstract, the authors claim an accuracy of 91.6% at the micro-level and 82.1% at the macro-level. However, these are F1 scores, and not accuracy, based on Tables 1 and 2. \n2. For the majority of the first half of the paper, it seems that the authors use the gold code patch which resolves the issue (L017-018, L080-081, L137-141). However, it becomes clear later that the authors only consider the gold *test* patch and the gold code patch is not used in their approach and only in their baselines. \n3. The prompts that the authors use are not given. Additionally, not a single example is provided.\n4. The notion of \"function-level\" is in multiple tables but this is never explained. This also seems to be the best performing, so it is not clear what the best-performing method is actually. Namely, Table 1 has only two rows for \"Isolated, Test-Aware\", one corresponding to \"post-commit functions\" and one augmenting that method with \"function-level.\" However, the two methods that are introduced in the main text are \"context enhancement\" and \"source code\" (which is also referred to as post-commit functions). In L259, it says \"We can see that the LLM critic with context-enhanced patches performs the best, outperforming other baselines by 7.4% and 12.7%\". But \"context-enhanced\" is not in Table 1 and there is only 1 baseline. \n5. In the abstract, the authors motivate this work as an \"intermediate/step-level\" evaluation methodology. This suggests that the intermediate steps in the LLM-based agentic framework could be evaluated. However, this does not seem to be demonstrated in the paper. It is not obvious, but if Figure 4 (right) was intended to demonstrate this, it is not clear. Additionally, it is not clear why this was presented as a plot rather than an aggregate spearman's ranking coefficient. It looks like there is a decent chunk of the density in the negative range here too.", "questions": "- Please consider addressing the points raised above.\n- Have you considered using both the gold test patch and gold code patch together?\n- Did you consider a baseline which just uses the candidate code patch, with no tests at all?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730754956610}, {"id": "l0FDDhcneG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11365/Reviewer_JASr"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "This paper proposes a method for evaluating code changes made by agentic workflows in software engineering. The authors introduce LLM-based critics that utilize a test-centric, execution-free framework for assessing candidate patches, using a \"gold patch\" as a reference to predict the build status. The framework is evaluated on SWE-bench, demonstrating acceptable performance over baseline and alternative evaluation methods in predicting code executability and build outcomes.\n\nHowever, this paper lacks a clear algorithm/overview flow chart to show their methodological contributions - or, in other words, there is no eye-catching innovation except for the relative rigor of the indicators and experiments used in the comparison between LLM evaluation and human evaluation. This is more like an Empirical Study to reveal that LLM critic can do a good job in patch code generation. But if that's all, exploring the code volume and domain scope of the patch should be fully discussed. The granular discussion based only on SWE-bench is far from sufficient.", "review_text": "This paper proposes a method for evaluating code changes made by agentic workflows in software engineering. The authors introduce LLM-based critics that utilize a test-centric, execution-free framework for assessing candidate patches, using a \"gold patch\" as a reference to predict the build status. The framework is evaluated on SWE-bench, demonstrating acceptable performance over baseline and alternative evaluation methods in predicting code executability and build outcomes.\n\nHowever, this paper lacks a clear algorithm/overview flow chart to show their methodological contributions - or, in other words, there is no eye-catching innovation except for the relative rigor of the indicators and experiments used in the comparison between LLM evaluation and human evaluation. This is more like an Empirical Study to reveal that LLM critic can do a good job in patch code generation. But if that's all, exploring the code volume and domain scope of the patch should be fully discussed. The granular discussion based only on SWE-bench is far from sufficient.", "strengths": "1. **Evaluation Framework**: The introduction of LLM-based critics for execution-free assessment of code patches trys to fill a gap in software engineering evaluation, as traditional methods like build status or log analysis require execution environments. This framework's approach to utilizing a \"test-centric\" structure is commendable, as it allows a fine-grained evaluation of patch quality at a functional level, aligning with real-world scenarios where compiling and testing environments may be costly or unavailable. \n\n2. **Comparative Rigor**: The authors offer an extensive experimental setup, comparing the LLM critic-based framework against multiple baselines. Their approach includes micro-evaluations for individual test predictions and macro-evaluation aggregations for build predictions, which demonstrates the framework's flexibility and adaptability across multiple agentic workflows and LLM models.\n\n3. **Performance on SWE-bench**: The proposed framework outperforms reference-free and some reference-aware baselines on SWE-bench, achieving a prediction accuracy of 91.5% for executability and 82.1% for build status. The improvements in prediction accuracy, particularly compared to edit-distance or change-aware baselines, are notable and validate the effectiveness of the LLM.", "weaknesses": "1. **Technical Description**: The technical pipeline of the test-centric framework is intricate, and certain steps could be explained with more clarity and visualization. For example, the mechanisms by which LLM critics predict test pass/fail outcomes based on context enhancement are briefly mentioned but should benefit from further elaboration and details, especially regarding how these critics handle complex test cases, also the correctness of CoT of LLM when evaluating the code patch.\n\n2. **Limited Discussion on Scalability Constraints**: While the authors demonstrate strong results on SWE-bench, the generalizability of the approach to a broader set of software repositories and programming languages is not deeply explored. This leaves open questions regarding the framework’s ability to adapt to repositories that may require unique dependencies or multilingual support. Discussions on granularity (explanatory power), code size (upper limit of capability), and multi-domain (characterizing domain performance of LLM) of LLM in reviewing code patches are crucial but missing in this paper.\n\n3. **Reliance on Gold Patches**: The framework's dependence on a reference (gold patch) for optimal accuracy raises potential issues in scenarios where a ground-truth patch may not exist. Although the authors attempt to address this with reference-free baselines, the performance drop observed here indicates that further research may be needed to refine reference-free evaluation methods.", "questions": "1. **Expand Baseline Comparisons**: The evaluation would be strengthened by including a broader set of baselines, potentially from benchmarks outside of SWE-bench. This could provide additional context for how well the framework generalizes across varied code types and patching workflows.\n\n2. **Provide Example Outputs**: Including specific examples of code patches, along with LLM critics' predictions, would illustrate the system’s inner workings more vividly and clarify the practical implications of each evaluation level (e.g., context-enhanced vs. source code-based evaluations).\n\n3. **In-depth Scalability Discussion**: Expanding the discussion on scalability, particularly in handling repositories with extensive interdependencies or complex testing setups, would provide insight into the framework’s applicability in large-scale industrial settings.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method for evaluating code changes made by agentic workflows in software engineering. The authors introduce LLM-based critics that utilize a test-centric, execution-free framework for assessing candidate patches, using a \"gold patch\" as a reference to predict the build status. The framework is evaluated on SWE-bench, demonstrating acceptable performance over baseline and alternative evaluation methods in predicting code executability and build outcomes.\n\nHowever, this paper lacks a clear algorithm/overview flow chart to show their methodological contributions - or, in other words, there is no eye-catching innovation except for the relative rigor of the indicators and experiments used in the comparison between LLM evaluation and human evaluation. This is more like an Empirical Study to reveal that LLM critic can do a good job in patch code generation. But if that's all, exploring the code volume and domain scope of the patch should be fully discussed. The granular discussion based only on SWE-bench is far from sufficient.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "1. **Evaluation Framework**: The introduction of LLM-based critics for execution-free assessment of code patches trys to fill a gap in software engineering evaluation, as traditional methods like build status or log analysis require execution environments. This framework's approach to utilizing a \"test-centric\" structure is commendable, as it allows a fine-grained evaluation of patch quality at a functional level, aligning with real-world scenarios where compiling and testing environments may be costly or unavailable. \n\n2. **Comparative Rigor**: The authors offer an extensive experimental setup, comparing the LLM critic-based framework against multiple baselines. Their approach includes micro-evaluations for individual test predictions and macro-evaluation aggregations for build predictions, which demonstrates the framework's flexibility and adaptability across multiple agentic workflows and LLM models.\n\n3. **Performance on SWE-bench**: The proposed framework outperforms reference-free and some reference-aware baselines on SWE-bench, achieving a prediction accuracy of 91.5% for executability and 82.1% for build status. The improvements in prediction accuracy, particularly compared to edit-distance or change-aware baselines, are notable and validate the effectiveness of the LLM.", "weaknesses": "1. **Technical Description**: The technical pipeline of the test-centric framework is intricate, and certain steps could be explained with more clarity and visualization. For example, the mechanisms by which LLM critics predict test pass/fail outcomes based on context enhancement are briefly mentioned but should benefit from further elaboration and details, especially regarding how these critics handle complex test cases, also the correctness of CoT of LLM when evaluating the code patch.\n\n2. **Limited Discussion on Scalability Constraints**: While the authors demonstrate strong results on SWE-bench, the generalizability of the approach to a broader set of software repositories and programming languages is not deeply explored. This leaves open questions regarding the framework’s ability to adapt to repositories that may require unique dependencies or multilingual support. Discussions on granularity (explanatory power), code size (upper limit of capability), and multi-domain (characterizing domain performance of LLM) of LLM in reviewing code patches are crucial but missing in this paper.\n\n3. **Reliance on Gold Patches**: The framework's dependence on a reference (gold patch) for optimal accuracy raises potential issues in scenarios where a ground-truth patch may not exist. Although the authors attempt to address this with reference-free baselines, the performance drop observed here indicates that further research may be needed to refine reference-free evaluation methods.", "questions": "1. **Expand Baseline Comparisons**: The evaluation would be strengthened by including a broader set of baselines, potentially from benchmarks outside of SWE-bench. This could provide additional context for how well the framework generalizes across varied code types and patching workflows.\n\n2. **Provide Example Outputs**: Including specific examples of code patches, along with LLM critics' predictions, would illustrate the system’s inner workings more vividly and clarify the practical implications of each evaluation level (e.g., context-enhanced vs. source code-based evaluations).\n\n3. **In-depth Scalability Discussion**: Expanding the discussion on scalability, particularly in handling repositories with extensive interdependencies or complex testing setups, would provide insight into the framework’s applicability in large-scale industrial settings.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730721036870}, {"id": "hEbAyqry4l", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11365/Reviewer_3RwT"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper studies the problem of performing execution-free evaluation of repository-level code changes made by programming agents. Assuming access to ground truth code changes, the paper uses LLM critics (akin to generative verifiers) to assess the correctness of the proposed code changes without actually executing the code. It adopts a \"test-centric\" granular framework comprising micro-evaluation per test case aggregated to provide a macro \"problem-level\" evaluation.", "review_text": "The paper studies the problem of performing execution-free evaluation of repository-level code changes made by programming agents. Assuming access to ground truth code changes, the paper uses LLM critics (akin to generative verifiers) to assess the correctness of the proposed code changes without actually executing the code. It adopts a \"test-centric\" granular framework comprising micro-evaluation per test case aggregated to provide a macro \"problem-level\" evaluation.", "strengths": "* Execution-free evaluation for code generation is an important and under-explored problem. Execution-based evaluation is often a challenging engineering problem, and costly. Using gold-patch-guided neural verifiers addresses attempts to address this. Notably, the problem (and proposed approach) has applications beyond benchmarking; it can be used to collect inexpensive 'reward-model' feedback for arbitrary GitHub repositories.\n* Intuitive and novel approach to using LLM workflow for a test-driven evaluation of code changes. The micro-evaluation aggregation-based method is more sensible, provided that the evaluations are well-calibrated.", "weaknesses": "* **Accuracy of LLM critics.** The paper evaluates gold-patch-guided LLM critics aggregated over test cases; however, LLM-based judges and verifiers are usually quite inaccurate and miscalibrated, even for simple programming problems like those in HumanEval or LeetCode. For example, [1] reported about 50% accuracy for open-source models serving as LLM critics, while GPT-4 achieves only 70-80% accuracy. This raises doubts about the feasibility of execution-free approaches to more complex software code changes studied in this paper. \n\n* **Baseline experiment on standard programming evaluations.** Building on the previous point, standard programming benchmarks like HumanEval or programming contest problems can be formulated as 'code change' problems—for example, given a function with a docstring, insert the necessary code. Understanding the effectiveness of the approach on such \"simpler\" settings might provide a more grounded understanding of the strength of the approach.\n \n* **Overfitting to SWEBench.** The authors use SWE-Bench-Lite as the sole evaluation benchmark; however, as they acknowledge, SWE-Bench-Lite is imbalanced, with the majority of tests passing. This raises concerns about the generalization of the approach. Specifically, in the micro-evaluations, authors observe a 98% recall -- potentially due to bias in LLMs to respond correct [1]. This aligns with the evaluation suite with high positives potentially inflating the results. It is unclear if this approach will generalize to more challenging benchmarks where a smaller fraction of tests pass and should be evaluated (say on SWEBench full suite).\n\n* **Calibration on the test benchmark.** Authors recalibrated model confidences -- capping a 65% confidence on model responding YES. However, this number uses private knowledge about the correctness of the patches during micro-evaluation recalibration on the SWEBench test set. This raises concerns about the potential for data leakage and the validity of the evaluation results.\n\n* **Pass to Fail tests.** In many cases, a programming agent solution can fail if it introduces a bug in an already passing testcase. It seems this is not handled since the approach only handled \"newly introduced\" tests in the PR  \n\n* **Access to clean pull requests (PRs) is assumed.** SWE-Bench, problem instances (PRs) are cleaned via execution to map the tests and code changes using a Fail-to-Pass strategy. From my understanding, the authors assume access to this information for setting up LLM critics. While this approach works for benchmarks, real-world PRs can be messy, containing unrelated changes to tests and code. This complexity may require at least one execution round to collect the necessary information, limiting the real-world applicability of the work focusing on execution-free nature.\n\n* **Undefined or inconsistent Terminology.** \n  * 'build status' is never defined and it is unclear if authors mean simply \"all tests passing\" or something beyond that\n  * In Table 1, settings in the candidate patch column are not properly explained -- `+- function-level` supposed to mean context-enhanced patches is not clear from the table.\n  * In Table 2, change-aware vs test-aware is not immediately clear from the description and could be explained with examples.\n  * 'code changes,' 'patches,' and 'edits' are used interchangeably without clarification and should be normalized to one aspect\n \n\nReferences.\n\n1. The Counterfeit Conundrum: Can Code Language Models Grasp the Nuances of Their Incorrect Generations?", "questions": "* I would recommend evaluating the approach over more benchmarks ranging from simple coding problems, competition problems, and harder software engineering problems (SWEBench full set) to reinforce the results.\n\nMinor:\n\n* The paper mentions that \"build status does not provide insights into functional correctness or performance under various conditions,\" which is confusing since build status often involves passing the test suite, which assesses functional correctness. Could you clarify what you mean by \"build status\" and how it relates to functional correctness in your context? Additionally, since the approach is compared against test success rate on SWEBench, it is not clear how the proposed approach performed better than functional correctness\n\n* The authors mention \"modified code may not even compile, or pass unit tests or the integration test. In such cases, traditional metrics are not sufficiently available to an agent to improve the patch\". However, the agent cannot access the ground truth test cases and should not use the signals from evaluation proxies to improve the patch. Therefore, I do not follow this argument.\n\n* Typos:\n  * Line 80: Important -> Importantly\n  * Line 274: differ -> defer\n  * Some sentences in the paper can be restructured to flow better.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of performing execution-free evaluation of repository-level code changes made by programming agents. Assuming access to ground truth code changes, the paper uses LLM critics (akin to generative verifiers) to assess the correctness of the proposed code changes without actually executing the code. It adopts a \"test-centric\" granular framework comprising micro-evaluation per test case aggregated to provide a macro \"problem-level\" evaluation.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "* Execution-free evaluation for code generation is an important and under-explored problem. Execution-based evaluation is often a challenging engineering problem, and costly. Using gold-patch-guided neural verifiers addresses attempts to address this. Notably, the problem (and proposed approach) has applications beyond benchmarking; it can be used to collect inexpensive 'reward-model' feedback for arbitrary GitHub repositories.\n* Intuitive and novel approach to using LLM workflow for a test-driven evaluation of code changes. The micro-evaluation aggregation-based method is more sensible, provided that the evaluations are well-calibrated.", "weaknesses": "* **Accuracy of LLM critics.** The paper evaluates gold-patch-guided LLM critics aggregated over test cases; however, LLM-based judges and verifiers are usually quite inaccurate and miscalibrated, even for simple programming problems like those in HumanEval or LeetCode. For example, [1] reported about 50% accuracy for open-source models serving as LLM critics, while GPT-4 achieves only 70-80% accuracy. This raises doubts about the feasibility of execution-free approaches to more complex software code changes studied in this paper. \n\n* **Baseline experiment on standard programming evaluations.** Building on the previous point, standard programming benchmarks like HumanEval or programming contest problems can be formulated as 'code change' problems—for example, given a function with a docstring, insert the necessary code. Understanding the effectiveness of the approach on such \"simpler\" settings might provide a more grounded understanding of the strength of the approach.\n \n* **Overfitting to SWEBench.** The authors use SWE-Bench-Lite as the sole evaluation benchmark; however, as they acknowledge, SWE-Bench-Lite is imbalanced, with the majority of tests passing. This raises concerns about the generalization of the approach. Specifically, in the micro-evaluations, authors observe a 98% recall -- potentially due to bias in LLMs to respond correct [1]. This aligns with the evaluation suite with high positives potentially inflating the results. It is unclear if this approach will generalize to more challenging benchmarks where a smaller fraction of tests pass and should be evaluated (say on SWEBench full suite).\n\n* **Calibration on the test benchmark.** Authors recalibrated model confidences -- capping a 65% confidence on model responding YES. However, this number uses private knowledge about the correctness of the patches during micro-evaluation recalibration on the SWEBench test set. This raises concerns about the potential for data leakage and the validity of the evaluation results.\n\n* **Pass to Fail tests.** In many cases, a programming agent solution can fail if it introduces a bug in an already passing testcase. It seems this is not handled since the approach only handled \"newly introduced\" tests in the PR  \n\n* **Access to clean pull requests (PRs) is assumed.** SWE-Bench, problem instances (PRs) are cleaned via execution to map the tests and code changes using a Fail-to-Pass strategy. From my understanding, the authors assume access to this information for setting up LLM critics. While this approach works for benchmarks, real-world PRs can be messy, containing unrelated changes to tests and code. This complexity may require at least one execution round to collect the necessary information, limiting the real-world applicability of the work focusing on execution-free nature.\n\n* **Undefined or inconsistent Terminology.** \n  * 'build status' is never defined and it is unclear if authors mean simply \"all tests passing\" or something beyond that\n  * In Table 1, settings in the candidate patch column are not properly explained -- `+- function-level` supposed to mean context-enhanced patches is not clear from the table.\n  * In Table 2, change-aware vs test-aware is not immediately clear from the description and could be explained with examples.\n  * 'code changes,' 'patches,' and 'edits' are used interchangeably without clarification and should be normalized to one aspect\n \n\nReferences.\n\n1. The Counterfeit Conundrum: Can Code Language Models Grasp the Nuances of Their Incorrect Generations?", "questions": "* I would recommend evaluating the approach over more benchmarks ranging from simple coding problems, competition problems, and harder software engineering problems (SWEBench full set) to reinforce the results.\n\nMinor:\n\n* The paper mentions that \"build status does not provide insights into functional correctness or performance under various conditions,\" which is confusing since build status often involves passing the test suite, which assesses functional correctness. Could you clarify what you mean by \"build status\" and how it relates to functional correctness in your context? Additionally, since the approach is compared against test success rate on SWEBench, it is not clear how the proposed approach performed better than functional correctness\n\n* The authors mention \"modified code may not even compile, or pass unit tests or the integration test. In such cases, traditional metrics are not sufficiently available to an agent to improve the patch\". However, the agent cannot access the ground truth test cases and should not use the signals from evaluation proxies to improve the patch. Therefore, I do not follow this argument.\n\n* Typos:\n  * Line 80: Important -> Importantly\n  * Line 274: differ -> defer\n  * Some sentences in the paper can be restructured to flow better.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730650765136}, {"id": "0F6yga7udL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11365/Reviewer_EhhE"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces a novel approach for evaluating code patches made by Large Language Model (LLM) agents using LLM-based critics. The key innovation is developing execution-free evaluation proxies that can assess repository-level code patches without requiring actual compilation or test execution. The authors propose a test-centric framework that leverages gold test patches as references to predict both individual test outcomes and overall build status. Their method achieves 91.6% accuracy in predicting test executability and 82.1% accuracy in predicting build status on the SWE-bench dataset, significantly outperforming reference-free and other reference-aware approaches.", "review_text": "This paper introduces a novel approach for evaluating code patches made by Large Language Model (LLM) agents using LLM-based critics. The key innovation is developing execution-free evaluation proxies that can assess repository-level code patches without requiring actual compilation or test execution. The authors propose a test-centric framework that leverages gold test patches as references to predict both individual test outcomes and overall build status. Their method achieves 91.6% accuracy in predicting test executability and 82.1% accuracy in predicting build status on the SWE-bench dataset, significantly outperforming reference-free and other reference-aware approaches.", "strengths": "- The studied problem of evaluating code changes without execution is interesting.\n- The paper is well-written and easy to follow.\n- The proposed method is effective.\n- Thorough ablation studies and analysis of different components.", "weaknesses": "There are several major weaknesses in this paper:\n\n1. The requirement for gold patches in the proposed method limits its applicability in practice. Although it also discusses a reference-free situation, the accuracies below 0.5 are far from practical use.\n2. The proposed method is only compared to a simple Edit Distance baseline in the experiments. If given the gold patches, various metrics like CodeBLEU and ROUGE could also provide a way to evaluate the similarity between the deleted and inserted code snippets in the generated output and gold patches. However, such baselines are missing from the experiments. Including these comparisons in the experiments and discussing the results would provide a more comprehensive evaluation of the proposed method's advantages.\n3. To evaluate the proposed method across different LLMs, this paper uses three variants of Claude models. However, the choice of using models from the same series may limit the generalizability of the proposed method. It would be beneficial to evaluate the method on more diverse LLMs, such as GPT-4 or other open-source models, to demonstrate its effectiveness.\n4. In what scenarios would the proposed method fail? There appears to be a lack of case studies or discussions on the success and failure cases of the proposed method. It would be helpful to provide examples that illustrate the benefits and limitations of the proposed method compared to other baselines. Additionally, I look forward to seeing example inputs and outputs of the LLM evaluator for code patches to better understand the proposed method.\n\nAnd some minor issues:\n\n1. The organization of Section 2 is somewhat confusing and lacks a figure or algorithm to illustrate the proposed method, making it difficult for readers to understand. There is only one subsubsection under each subsection, which obscures the structure of the section. I recommend that the authors consider adding a figure or algorithm table to help readers better grasp the proposed method and reorganize the section for improved clarity.\n2. The authors claim to open-source the code in Lines 25 and 111 as an important contribution. However, the code is not provided in the supplementary material or the paper with anonymous links, making it difficult for me to evaluate the artifacts.\n3. As claimed in Line 192, the proposed method prompts LLMs to generate evaluations of code patches. However, the details of how the LLMs are prompted are not clear. It would be helpful to provide the exact prompt in the Appendix.\n4. The left figure in Figure 4 is too small, and there are timestamps in the caption that seem unnecessary. It is also recommended to use a larger font size for the text in Figures 3 and 4.", "questions": "Please address the concerns in the \"Weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel approach for evaluating code patches made by Large Language Model (LLM) agents using LLM-based critics. The key innovation is developing execution-free evaluation proxies that can assess repository-level code patches without requiring actual compilation or test execution. The authors propose a test-centric framework that leverages gold test patches as references to predict both individual test outcomes and overall build status. Their method achieves 91.6% accuracy in predicting test executability and 82.1% accuracy in predicting build status on the SWE-bench dataset, significantly outperforming reference-free and other reference-aware approaches.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The studied problem of evaluating code changes without execution is interesting.\n- The paper is well-written and easy to follow.\n- The proposed method is effective.\n- Thorough ablation studies and analysis of different components.", "weaknesses": "There are several major weaknesses in this paper:\n\n1. The requirement for gold patches in the proposed method limits its applicability in practice. Although it also discusses a reference-free situation, the accuracies below 0.5 are far from practical use.\n2. The proposed method is only compared to a simple Edit Distance baseline in the experiments. If given the gold patches, various metrics like CodeBLEU and ROUGE could also provide a way to evaluate the similarity between the deleted and inserted code snippets in the generated output and gold patches. However, such baselines are missing from the experiments. Including these comparisons in the experiments and discussing the results would provide a more comprehensive evaluation of the proposed method's advantages.\n3. To evaluate the proposed method across different LLMs, this paper uses three variants of Claude models. However, the choice of using models from the same series may limit the generalizability of the proposed method. It would be beneficial to evaluate the method on more diverse LLMs, such as GPT-4 or other open-source models, to demonstrate its effectiveness.\n4. In what scenarios would the proposed method fail? There appears to be a lack of case studies or discussions on the success and failure cases of the proposed method. It would be helpful to provide examples that illustrate the benefits and limitations of the proposed method compared to other baselines. Additionally, I look forward to seeing example inputs and outputs of the LLM evaluator for code patches to better understand the proposed method.\n\nAnd some minor issues:\n\n1. The organization of Section 2 is somewhat confusing and lacks a figure or algorithm to illustrate the proposed method, making it difficult for readers to understand. There is only one subsubsection under each subsection, which obscures the structure of the section. I recommend that the authors consider adding a figure or algorithm table to help readers better grasp the proposed method and reorganize the section for improved clarity.\n2. The authors claim to open-source the code in Lines 25 and 111 as an important contribution. However, the code is not provided in the supplementary material or the paper with anonymous links, making it difficult for me to evaluate the artifacts.\n3. As claimed in Line 192, the proposed method prompts LLMs to generate evaluations of code patches. However, the details of how the LLMs are prompted are not clear. It would be helpful to provide the exact prompt in the Appendix.\n4. The left figure in Figure 4 is too small, and there are timestamps in the caption that seem unnecessary. It is also recommended to use a larger font size for the text in Figures 3 and 4.", "questions": "Please address the concerns in the \"Weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730647226382}], "openreview_url": "https://openreview.net/forum?id=gDWkImLIKd", "arxiv_id": "2501.16655", "paper_pdf": "papers/gDWkImLIKd.pdf", "paper_pdf_sha256": "e1a32aa93d8e717baae677abb34fd45407acdfaa22f9a28020d21cd7c093dc90", "paper_pdf_bytes": 1207802, "paper_pdf_source": "openreview", "code_url": "https://github.com/amazon-science/code-agent-eval", "code_repository": "amazon-science/code-agent-eval", "code_commit": "947415316064212917033e6d06df20f2b1aaf743", "code_archive": "repos/gDWkImLIKd.zip", "code_archive_sha256": "b8dacdb2ab672a983332c7bd706dfc05857b41d4f63ea3bea58a29f9da199790", "code_archive_bytes": 94586, "code_file_count": 36, "code_extensions": {".py": 32, ".sh": 4}, "github_disk_usage_kb": 85, "github_languages": {"Python": 230439, "Shell": 5729}, "github_archived": false, "github_pushed_at": "2025-01-29T02:22:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/large-language-model-critics-for-execution"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hgrZluxFC7", "year": 2024, "status": "rejected", "title": "Adversarial Machine Learning in Latent Representations of Neural Networks", "authors": ["Milin Zhang", "Mohammad Abdi", "Francesco Restuccia"], "authorids": ["~Milin_Zhang1", "~Mohammad_Abdi1", "~Francesco_Restuccia2"], "authors_source": "OpenReview API", "abstract": "Distributed deep neural networks (DNNs) have been shown to reduce the computational burden of mobile devices and decrease the end-to-end inference latency in edge computing scenarios. While distributed DNNs have been studied, to the best of our knowledge the resilience of distributed DNNs to adversarial action still remains an open problem. In this paper, we fill the existing research gap by rigorously analyzing the robustness of distributed DNNs against adversarial action. We cast this problem in the context of information theory and introduce two new measurements for distortion and robustness. Our theoretical findings indicate that (i) assuming the same level of information distortion, latent features are always more robust than input representations; (ii) the adversarial robustness is jointly determined by the feature dimension and the generalization capability of the DNN. To test our theoretical findings, we perform extensive experimental analysis by considering 6 different DNN architectures, 6 different approaches for distributed DNN and 10 different adversarial attacks to the ImageNet-1K dataset. Our experimental results support our theoretical findings by showing that the compressed latent representations can reduce the success rate of adversarial attacks by 88% in the best case and by 57% on the average compared to attacks to the input space.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "gj4mn00z9J", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5558/Reviewer_CQZD"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work aims to evaluate the adversarial robustness of distributed DNNs. In this setting, latent representations of the DNNs are communicated among devices, and thus, an adversary could perturb the latent representations instead of the model's input. However, the work claims that attacks on latent representations are less effective than perturbing the input and presents theoretical information bounds supporting this claim. Standard adversarial attacks are then used to test this bound empirically over several distributed DNN architectures, and the results show that the attacks are less successful when employed on latent spaces than on the model's input.", "review_text": "This work aims to evaluate the adversarial robustness of distributed DNNs. In this setting, latent representations of the DNNs are communicated among devices, and thus, an adversary could perturb the latent representations instead of the model's input. However, the work claims that attacks on latent representations are less effective than perturbing the input and presents theoretical information bounds supporting this claim. Standard adversarial attacks are then used to test this bound empirically over several distributed DNN architectures, and the results show that the attacks are less successful when employed on latent spaces than on the model's input.", "strengths": "The suggested information bound entails that standard adversarial attacks targeting only the latent representation of DNNs would be less effective than those targeting the input. This is relevant not only in distributed DNN but for side-channel attacks as well. Moreover, it greatly aids in evaluating the robustness of distributed DNN, as such attacks arise naturally in this setting.", "weaknesses": "1. The experimental setup is lacking and insufficient to support the authors' claims. Only attacks on several distributed DNN architectures were reported. Not only is this insufficient to support the claim that attacks on latent spaces are generally less effective, but it does not explain the phenomenon or the behavior of the suggested information bound.\n2. No novel attacks targeting latent spaces were suggested, or even settings in which both the input and latent space are attacked. Without testing such attacks and settings, the robustness of distributed DNN cannot be correctly evaluated.\n3. The second key finding of \"DNN robustness is intrinsically related to the cardinality of the latent space\" is not a phenomenon exclusive to latent spaces. There are several examples of attacks working better on larger input samples such as Imagnet, compared to CIFAR10/100. In addition, the effect of the l_inf norm bound is highly dependent on the input size.", "questions": "1. For a correct evaluation of the suggested bound on a given DNN architecture, the experimental settings should present attacks on all the latent spaces in the architecture and not only those available in specific distributed DNN settings. The results should be compared for the depth of the latent spaces in the network and their cardinality. Such experiments consider side-channel attacks on DNN and not only the distributed DNN setting.\n2. Attacks targeting explicitly latent spaces should be considered; such attacks should be aware of the specifics of the latent spaces (e.g., depth and cardinality) and make use of them to improve the efficiency of the attack.\n3. Adversarial attacks targeting input and latent representations should be considered to evaluate if such a setting presents a greater risk to distributed DNN. \n4. As the effectiveness of perturbations depends on the input size, the results should be normalized accordingly.\n\nPost-rebuttal feedback\nI am satisfied with authors' clarifications and provided additional evaluation. Hence, my score to 6", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work aims to evaluate the adversarial robustness of distributed DNNs. In this setting, latent representations of the DNNs are communicated among devices, and thus, an adversary could perturb the latent representations instead of the model's input. However, the work claims that attacks on latent representations are less effective than perturbing the input and presents theoretical information bounds supporting this claim. Standard adversarial attacks are then used to test this bound empirically over several distributed DNN architectures, and the results show that the attacks are less successful when employed on latent spaces than on the model's input.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The suggested information bound entails that standard adversarial attacks targeting only the latent representation of DNNs would be less effective than those targeting the input. This is relevant not only in distributed DNN but for side-channel attacks as well. Moreover, it greatly aids in evaluating the robustness of distributed DNN, as such attacks arise naturally in this setting.", "weaknesses": "1. The experimental setup is lacking and insufficient to support the authors' claims. Only attacks on several distributed DNN architectures were reported. Not only is this insufficient to support the claim that attacks on latent spaces are generally less effective, but it does not explain the phenomenon or the behavior of the suggested information bound.\n2. No novel attacks targeting latent spaces were suggested, or even settings in which both the input and latent space are attacked. Without testing such attacks and settings, the robustness of distributed DNN cannot be correctly evaluated.\n3. The second key finding of \"DNN robustness is intrinsically related to the cardinality of the latent space\" is not a phenomenon exclusive to latent spaces. There are several examples of attacks working better on larger input samples such as Imagnet, compared to CIFAR10/100. In addition, the effect of the l_inf norm bound is highly dependent on the input size.", "questions": "1. For a correct evaluation of the suggested bound on a given DNN architecture, the experimental settings should present attacks on all the latent spaces in the architecture and not only those available in specific distributed DNN settings. The results should be compared for the depth of the latent spaces in the network and their cardinality. Such experiments consider side-channel attacks on DNN and not only the distributed DNN setting.\n2. Attacks targeting explicitly latent spaces should be considered; such attacks should be aware of the specifics of the latent spaces (e.g., depth and cardinality) and make use of them to improve the efficiency of the attack.\n3. Adversarial attacks targeting input and latent representations should be considered to evaluate if such a setting presents a greater risk to distributed DNN. \n4. As the effectiveness of perturbations depends on the input size, the results should be normalized accordingly.\n\nPost-rebuttal feedback\nI am satisfied with authors' clarifications and provided additional evaluation. Hence, my score to 6", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699195219615}, {"id": "7gfuPBGCXx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5558/Reviewer_an5M"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper studies the robustness of distributed DNN against adversarial attack theoretically and experimentally. The authors analyze the robustness of latent features using information bottleneck theory and prove that latent space perturbations are always less effective than input space perturbations. Empirically, the authors conduct extensive experiments to verify their theoretic findings from multiple perspectives.", "review_text": "This paper studies the robustness of distributed DNN against adversarial attack theoretically and experimentally. The authors analyze the robustness of latent features using information bottleneck theory and prove that latent space perturbations are always less effective than input space perturbations. Empirically, the authors conduct extensive experiments to verify their theoretic findings from multiple perspectives.", "strengths": "The paper is well written in general. The theoretical analysis from information bottleneck perspective is interesting and solid. Most of the claims are supported by ample experimental analysis.", "weaknesses": "1. In this paper, the attacker only has access to the latent representation provided by the mobile DNN from a mobile phone. While the authors successfully demonstrate that attacks on these latent representations exhibit a lower Attack Success Rate (ASR) than those on raw images—given the same level of information distortion—the appropriateness of imposing an identical distortion level in this context needs more clarification. In discussions around the adversarial robustness of image input, constraints are placed on distortion levels to ensure modifications are imperceptible to the human eye, yet potent enough to deceive the classifier [1]. However, when considering distortions applied to latent features, the rationale for enforcing the same constraint level is less clear. Given that the transmission in this scenario occurs between a mobile device and a cloud computer, with no human observer in the loop, the attacker might as well nullify all latent features, potentially achieving an ASR close to 100%. Comparing the ASR between attacks on raw input images (which are observable by humans) and attacks on latent features (processed by a cloud computer) under an equivalent distortion level seems illogical.\n\n2. Although the authors posit that this study on the robustness of distributed DNNs in the face of adversarial actions is novel, I find the concept markedly similar to existing works on attacks targeting intermediate layers or latent features [2]. I would appreciate it if the authors could highlight the distinctions between their work and prior research, and attempt to apply the attack methods delineated in [2] where feasible.\n\n3. I am confused about two terminologies in the paper: feature compression and bottleneck. Their relationship is not clear to me. In Table 1, it seems that only the first two feature compression methods contain a bottleneck layer. However, in Section 4.2 - DNN architectures, the authors write, '...the feature compression layer (i.e., the 'bottleneck').' Additionally, the authors use the same bottleneck design as Matsubara et al. (2022b) and denote the new architectures with feature compression as Resnet50-fc, etc. However, the specific design mentioned in Matsubara et al. (2022b) does not seem to belong to any of the six feature compression approaches in Table 1. Table A-3 is also confusing, as I cannot understand what the authors mean by 'JC and QT are DNNs without a bottleneck,' while JC and QT are feature compression approaches, and the authors claim that feature compression is the same as bottleneck.\n\n4. The experimental results in Section 5.3 need further explanation. ResNet152-fc with 12 channels achieves a validation accuracy of 77.47%, while ResNet152-fc with 3 channels achieves 70.01% accuracy. On the middle of page 9, the statement 'decreases to 7.47%' should be corrected to 'decreases by 7.46%.' However, the fact that I∗(Y ; T) decreases by 7.46% cannot explain why the ASR increases by a much larger percentage than 7.46% in Table A-4. For example, when \\epsilon=0.003, the ASR of PGD_2 increases by 19.9% when transitioning from 12 channels to 3 channels. According to the inequality in Key Theoretical Finding #1, since O(|T||Y|/√n) is smaller when transitioning from 12 channels to 3 channels, the ASR difference should be less than the difference in I∗(Y ; T), which is 7.46%. Please provide a detailed explanation.\n\n[1] Kurakin, Alexey, Ian J. Goodfellow, and Samy Bengio. \"Adversarial examples in the physical world.\" Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 99-112.\n[2] Yu, Yunrui, Xitong Gao, and Cheng-Zhong Xu. \"Lafeat: Piercing through adversarial defenses with latent features.\" CVPR. 2019.", "questions": "My primary concern is related to the validity of the problem setting presented in this paper (See weakness 1). Although the theoretical findings are intriguing and the experimental data is comprehensive, there is still uncertainty regarding the significance of defining the robustness of distributed Deep Neural Networks (DNNs) in the proposed manner.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the robustness of distributed DNN against adversarial attack theoretically and experimentally. The authors analyze the robustness of latent features using information bottleneck theory and prove that latent space perturbations are always less effective than input space perturbations. Empirically, the authors conduct extensive experiments to verify their theoretic findings from multiple perspectives.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well written in general. The theoretical analysis from information bottleneck perspective is interesting and solid. Most of the claims are supported by ample experimental analysis.", "weaknesses": "1. In this paper, the attacker only has access to the latent representation provided by the mobile DNN from a mobile phone. While the authors successfully demonstrate that attacks on these latent representations exhibit a lower Attack Success Rate (ASR) than those on raw images—given the same level of information distortion—the appropriateness of imposing an identical distortion level in this context needs more clarification. In discussions around the adversarial robustness of image input, constraints are placed on distortion levels to ensure modifications are imperceptible to the human eye, yet potent enough to deceive the classifier [1]. However, when considering distortions applied to latent features, the rationale for enforcing the same constraint level is less clear. Given that the transmission in this scenario occurs between a mobile device and a cloud computer, with no human observer in the loop, the attacker might as well nullify all latent features, potentially achieving an ASR close to 100%. Comparing the ASR between attacks on raw input images (which are observable by humans) and attacks on latent features (processed by a cloud computer) under an equivalent distortion level seems illogical.\n\n2. Although the authors posit that this study on the robustness of distributed DNNs in the face of adversarial actions is novel, I find the concept markedly similar to existing works on attacks targeting intermediate layers or latent features [2]. I would appreciate it if the authors could highlight the distinctions between their work and prior research, and attempt to apply the attack methods delineated in [2] where feasible.\n\n3. I am confused about two terminologies in the paper: feature compression and bottleneck. Their relationship is not clear to me. In Table 1, it seems that only the first two feature compression methods contain a bottleneck layer. However, in Section 4.2 - DNN architectures, the authors write, '...the feature compression layer (i.e., the 'bottleneck').' Additionally, the authors use the same bottleneck design as Matsubara et al. (2022b) and denote the new architectures with feature compression as Resnet50-fc, etc. However, the specific design mentioned in Matsubara et al. (2022b) does not seem to belong to any of the six feature compression approaches in Table 1. Table A-3 is also confusing, as I cannot understand what the authors mean by 'JC and QT are DNNs without a bottleneck,' while JC and QT are feature compression approaches, and the authors claim that feature compression is the same as bottleneck.\n\n4. The experimental results in Section 5.3 need further explanation. ResNet152-fc with 12 channels achieves a validation accuracy of 77.47%, while ResNet152-fc with 3 channels achieves 70.01% accuracy. On the middle of page 9, the statement 'decreases to 7.47%' should be corrected to 'decreases by 7.46%.' However, the fact that I∗(Y ; T) decreases by 7.46% cannot explain why the ASR increases by a much larger percentage than 7.46% in Table A-4. For example, when \\epsilon=0.003, the ASR of PGD_2 increases by 19.9% when transitioning from 12 channels to 3 channels. According to the inequality in Key Theoretical Finding #1, since O(|T||Y|/√n) is smaller when transitioning from 12 channels to 3 channels, the ASR difference should be less than the difference in I∗(Y ; T), which is 7.46%. Please provide a detailed explanation.\n\n[1] Kurakin, Alexey, Ian J. Goodfellow, and Samy Bengio. \"Adversarial examples in the physical world.\" Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 99-112.\n[2] Yu, Yunrui, Xitong Gao, and Cheng-Zhong Xu. \"Lafeat: Piercing through adversarial defenses with latent features.\" CVPR. 2019.", "questions": "My primary concern is related to the validity of the problem setting presented in this paper (See weakness 1). Although the theoretical findings are intriguing and the experimental data is comprehensive, there is still uncertainty regarding the significance of defining the robustness of distributed Deep Neural Networks (DNNs) in the proposed manner.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698843063111}, {"id": "DyrUDxGQ15", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5558/Reviewer_z2Hg"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "A distributed DNN can be regarded as a combination of two parts, namely, a mobile DNN and a local DNN, respectively. The mobile DNN is trained to learn the latent representations, which can reduce the amount of data that will be transmitted but suffers the risk of being attacked. Along this side, this paper investigates the robustness of the latent representations. Based on information theory, this paper claims that: 1) latent features are always more robust than input representations, and 2) the feature dimensions and the generalization capability of the DNN determine the adversarial robustness. Extensive experiments on ImageNet-1K are conducted to support the claims, considering 6 different DNN architectures, 6 different ways for distributed DNN, and 10 different adversarial attacks.", "review_text": "A distributed DNN can be regarded as a combination of two parts, namely, a mobile DNN and a local DNN, respectively. The mobile DNN is trained to learn the latent representations, which can reduce the amount of data that will be transmitted but suffers the risk of being attacked. Along this side, this paper investigates the robustness of the latent representations. Based on information theory, this paper claims that: 1) latent features are always more robust than input representations, and 2) the feature dimensions and the generalization capability of the DNN determine the adversarial robustness. Extensive experiments on ImageNet-1K are conducted to support the claims, considering 6 different DNN architectures, 6 different ways for distributed DNN, and 10 different adversarial attacks.", "strengths": "1. This paper is well-written, with clear explanations and illustrations. Section 2 is comprehensive, and someone interested in those related topics can learn from the article.\n2. Based on Information theory, Section 3 provides a thorough theoretical analysis. Besides, the theoretical conclusions are supported by the experimental results in Section 4 with detailed experimental settings.", "weaknesses": "In the Conclusion section, this article claims that ``This paper has investigated adversarial attacks to latent representations of DNNs for the first time``. Through the lens of distributed DNNs, this work may be the first one, as it claims. However, I am concerned about whether the distributed DNNs are a necessary background as the motivation to investigate the problem. Since I think the local DNN and mobile DNN are very like the architecture of an autoencoder, and there are many works about the adversarial robustness of autoencoders.", "questions": "As mentioned in the Weaknesses, my questions/concerns are mainly about the differences between the distributed DNNs and the autoencoders.\n1. If we compare the architecture between a distributed DNN and an autoencoder, I think the local DNN is very much like the encoder part, and the mobile DNN is very much like the decoder part. Can I compare them like this?\n2. If yes, I think some works have studied the adversarial robustness of the latent features, e.g., [1]. \n3. Therefore, I am a bit curious about whether distributed DNNs are a necessary background as the motivation to investigate the adversarial robustness of the latent features.\n\n---\n[1] Espinoza-Cuadros, F. M., Perero-Codosero, J. M., Antón-Martín, J., & Hernández-Gómez, L. A. (2020). Speaker de-identification system using autoencoders and adversarial training. arXiv preprint arXiv:2011.04696.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "A distributed DNN can be regarded as a combination of two parts, namely, a mobile DNN and a local DNN, respectively. The mobile DNN is trained to learn the latent representations, which can reduce the amount of data that will be transmitted but suffers the risk of being attacked. Along this side, this paper investigates the robustness of the latent representations. Based on information theory, this paper claims that: 1) latent features are always more robust than input representations, and 2) the feature dimensions and the generalization capability of the DNN determine the adversarial robustness. Extensive experiments on ImageNet-1K are conducted to support the claims, considering 6 different DNN architectures, 6 different ways for distributed DNN, and 10 different adversarial attacks.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper is well-written, with clear explanations and illustrations. Section 2 is comprehensive, and someone interested in those related topics can learn from the article.\n2. Based on Information theory, Section 3 provides a thorough theoretical analysis. Besides, the theoretical conclusions are supported by the experimental results in Section 4 with detailed experimental settings.", "weaknesses": "In the Conclusion section, this article claims that ``This paper has investigated adversarial attacks to latent representations of DNNs for the first time``. Through the lens of distributed DNNs, this work may be the first one, as it claims. However, I am concerned about whether the distributed DNNs are a necessary background as the motivation to investigate the problem. Since I think the local DNN and mobile DNN are very like the architecture of an autoencoder, and there are many works about the adversarial robustness of autoencoders.", "questions": "As mentioned in the Weaknesses, my questions/concerns are mainly about the differences between the distributed DNNs and the autoencoders.\n1. If we compare the architecture between a distributed DNN and an autoencoder, I think the local DNN is very much like the encoder part, and the mobile DNN is very much like the decoder part. Can I compare them like this?\n2. If yes, I think some works have studied the adversarial robustness of the latent features, e.g., [1]. \n3. Therefore, I am a bit curious about whether distributed DNNs are a necessary background as the motivation to investigate the adversarial robustness of the latent features.\n\n---\n[1] Espinoza-Cuadros, F. M., Perero-Codosero, J. M., Antón-Martín, J., & Hernández-Gómez, L. A. (2020). Speaker de-identification system using autoencoders and adversarial training. arXiv preprint arXiv:2011.04696.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698770191987}, {"id": "EfCVGsgeqf", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5558/Reviewer_nH51"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper studies the adversarial robustness of deep neural networks (DNNs) in latent space and compares it with the more common adversarial attacks on DNN inputs. The target models are distributed DNNs where the network is splitted in two parts, one is running on a mobile device and sends the output features to other part of the network which is running in cloud. The authors build on top of Information Bottleneck (IB) theory and show that 1) latent features of DNNs are more robust to adversarial attacks than inputs and 2) the smaller the latent dimension, the more difficult it is for an adversary to attack the network successfully. Their results on a wide variety of attacks and networks support the theoretical hypothesis.", "review_text": "The paper studies the adversarial robustness of deep neural networks (DNNs) in latent space and compares it with the more common adversarial attacks on DNN inputs. The target models are distributed DNNs where the network is splitted in two parts, one is running on a mobile device and sends the output features to other part of the network which is running in cloud. The authors build on top of Information Bottleneck (IB) theory and show that 1) latent features of DNNs are more robust to adversarial attacks than inputs and 2) the smaller the latent dimension, the more difficult it is for an adversary to attack the network successfully. Their results on a wide variety of attacks and networks support the theoretical hypothesis.", "strengths": "The paper is easy to follow, has good background sections and exploits the work of Shamir et al. 2010 (equation 3) in a natural way. The experiments are performed on a wide range of adversarial attacks, such as gradient-based, score-based and decision based, as well as white-box and black-box attacks. The results are clear and in line with the theoretical hypothesis.", "weaknesses": "There are a few weaknesses that I would like to point out:\n1. the paper does not take into consideration attacking the latent representations of a DNN that was adversarially trained. To the best of my knowledge, the adversarial training [1] has an impact over the latent representations of a neural network and under certain settings (a perturbation budget $\\epsilon$ not too large) the network is robust to adversarial input perturbations. I would be interested in the robustness of adversarially trained networks, would it be possible to perform such an experiment? I believe the paper is not complete without this experiment and I would really appreciate if you could include this. \n\nReferences:\n\n[1] **TOWARDS DEEP LEARNING MODELS RESISTANT TO ADVERSARIAL ATTACKS**, available at **https://openreview.net/forum?id=rJzIBfZAb**", "questions": "1. Related to the weakness point: how do you think the adversarial training would change the success rate of these attacks?\n2. recently there was another adversarial attack introduced for slightly more particular DNNs architectures with multi-exits (or early-exits), called DeepSloth [2], which aims to make the early-exits ineffective. They show that this attack changes the latent representation of DNNs (for example, Figure 3 in the paper) to actually create delay.\n3. Did think about analyzing the latent features robustness in LLMs? What do you think the challenges would be?\n4. In Section 6 you mention about defense mechanisms for adversarial attack on latent features. What would be the key element in designing defenses for this attack?\n\nReferences:\n\n[2] **A Panda? No, It's a Sloth: Slowdown Attacks on Adaptive Multi-Exit Neural Network Inference**, available at **https://openreview.net/forum?id=9xC2tWEwBD**", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the adversarial robustness of deep neural networks (DNNs) in latent space and compares it with the more common adversarial attacks on DNN inputs. The target models are distributed DNNs where the network is splitted in two parts, one is running on a mobile device and sends the output features to other part of the network which is running in cloud. The authors build on top of Information Bottleneck (IB) theory and show that 1) latent features of DNNs are more robust to adversarial attacks than inputs and 2) the smaller the latent dimension, the more difficult it is for an adversary to attack the network successfully. Their results on a wide variety of attacks and networks support the theoretical hypothesis.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper is easy to follow, has good background sections and exploits the work of Shamir et al. 2010 (equation 3) in a natural way. The experiments are performed on a wide range of adversarial attacks, such as gradient-based, score-based and decision based, as well as white-box and black-box attacks. The results are clear and in line with the theoretical hypothesis.", "weaknesses": "There are a few weaknesses that I would like to point out:\n1. the paper does not take into consideration attacking the latent representations of a DNN that was adversarially trained. To the best of my knowledge, the adversarial training [1] has an impact over the latent representations of a neural network and under certain settings (a perturbation budget $\\epsilon$ not too large) the network is robust to adversarial input perturbations. I would be interested in the robustness of adversarially trained networks, would it be possible to perform such an experiment? I believe the paper is not complete without this experiment and I would really appreciate if you could include this. \n\nReferences:\n\n[1] **TOWARDS DEEP LEARNING MODELS RESISTANT TO ADVERSARIAL ATTACKS**, available at **https://openreview.net/forum?id=rJzIBfZAb**", "questions": "1. Related to the weakness point: how do you think the adversarial training would change the success rate of these attacks?\n2. recently there was another adversarial attack introduced for slightly more particular DNNs architectures with multi-exits (or early-exits), called DeepSloth [2], which aims to make the early-exits ineffective. They show that this attack changes the latent representation of DNNs (for example, Figure 3 in the paper) to actually create delay.\n3. Did think about analyzing the latent features robustness in LLMs? What do you think the challenges would be?\n4. In Section 6 you mention about defense mechanisms for adversarial attack on latent features. What would be the key element in designing defenses for this attack?\n\nReferences:\n\n[2] **A Panda? No, It's a Sloth: Slowdown Attacks on Adaptive Multi-Exit Neural Network Inference**, available at **https://openreview.net/forum?id=9xC2tWEwBD**", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698602388842}, {"id": "gx6mkSLz4P", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5558/Reviewer_4MyR"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper uses information bottleneck (IB) analysis to study adversarial robustness in the split learning setting, consisting of a data encoder that gives the latent representations, followed by a local deep neural network (DNN) that takes the latent presentation for subsequent inference. The results show that the compressed latent representations can reduce the success rate of adversarial attacks, as also indicated by the theory.", "review_text": "This paper uses information bottleneck (IB) analysis to study adversarial robustness in the split learning setting, consisting of a data encoder that gives the latent representations, followed by a local deep neural network (DNN) that takes the latent presentation for subsequent inference. The results show that the compressed latent representations can reduce the success rate of adversarial attacks, as also indicated by the theory.", "strengths": "1. Use IB to study adversarial robustness is a good angle.\n2. The paper is well-written and easy to read", "weaknesses": "I have several major concerns about the technical novelty and empirical evaluation.\n\n1. On the claim that \"assuming the same level of information distortion, latent features are always more robust than input representations\", does this hold even if the latent embedding dimension is larger than the input dimension?  If yes, why it can be more robust? Moreover, if the latent embedding dimension is lower than the input dimension, then it has been proven in the ICML 2019 paper \"First-order Adversarial Vulnerability of Neural Networks and Input Dimension\" that the minimal adversarial perturbation scales inversely with the data dimension. So if we treat the latent representations as the new \"data\" and they have lower dimensions than the raw inputs, the new insights are not clear.\n\n2. The assumption on the same level of information distortion seems very strong and lacks justification. It's also not clear what is the threat model (what the attacker can do) that leads to a latent representation $T_{adv}$.\n\n3. The evaluated attacks are naive input perturbation attacks, and no adaptive attacks that take into account modifying the latent representations were studied. It should be easy to add a regularizer to attack objectives to encourage finding adversarial examples that share very similar (or very different) latent representations as the original data, and therefore the claim on improved robustness may not hold.", "questions": "1. W.r.t. to W1, is there any implicit assumption that the latent embedding dimension is smaller than the input dimension? If so, the improved robustness is a direct consequence of the ICML 2019 result.\n\n2. How to justify the assumption of \"the same level of information distortion\"? Does $T_{adv}$ hold for any arbitrary threat model? \n\n3. Does the result of improved robustness still hold against adaptive attacks, where the attacker can have access to the latent representations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper uses information bottleneck (IB) analysis to study adversarial robustness in the split learning setting, consisting of a data encoder that gives the latent representations, followed by a local deep neural network (DNN) that takes the latent presentation for subsequent inference. The results show that the compressed latent representations can reduce the success rate of adversarial attacks, as also indicated by the theory.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. Use IB to study adversarial robustness is a good angle.\n2. The paper is well-written and easy to read", "weaknesses": "I have several major concerns about the technical novelty and empirical evaluation.\n\n1. On the claim that \"assuming the same level of information distortion, latent features are always more robust than input representations\", does this hold even if the latent embedding dimension is larger than the input dimension?  If yes, why it can be more robust? Moreover, if the latent embedding dimension is lower than the input dimension, then it has been proven in the ICML 2019 paper \"First-order Adversarial Vulnerability of Neural Networks and Input Dimension\" that the minimal adversarial perturbation scales inversely with the data dimension. So if we treat the latent representations as the new \"data\" and they have lower dimensions than the raw inputs, the new insights are not clear.\n\n2. The assumption on the same level of information distortion seems very strong and lacks justification. It's also not clear what is the threat model (what the attacker can do) that leads to a latent representation $T_{adv}$.\n\n3. The evaluated attacks are naive input perturbation attacks, and no adaptive attacks that take into account modifying the latent representations were studied. It should be easy to add a regularizer to attack objectives to encourage finding adversarial examples that share very similar (or very different) latent representations as the original data, and therefore the claim on improved robustness may not hold.", "questions": "1. W.r.t. to W1, is there any implicit assumption that the latent embedding dimension is smaller than the input dimension? If so, the improved robustness is a direct consequence of the ICML 2019 result.\n\n2. How to justify the assumption of \"the same level of information distortion\"? Does $T_{adv}$ hold for any arbitrary threat model? \n\n3. Does the result of improved robustness still hold against adaptive attacks, where the attacker can have access to the latent representations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698368366567}], "openreview_url": "https://openreview.net/forum?id=hgrZluxFC7", "arxiv_id": "2309.17401", "paper_pdf": "papers/hgrZluxFC7.pdf", "paper_pdf_sha256": "56afaa049267e6e038178780a4adbb91309ebc28b8c48bd25ec4eb981e82b68c", "paper_pdf_bytes": 1759573, "paper_pdf_source": "openreview", "code_url": "https://github.com/asdfqwezxcf/AdvLatent", "code_repository": "asdfqwezxcf/AdvLatent", "code_commit": "f34e9a81fc4beebdc17f70f085140082fb3e3cfd", "code_archive": "repos/hgrZluxFC7.zip", "code_archive_sha256": "75a8992e535a23de27267bfa4df630d9202fd110194f5656a83dad598c668e0c", "code_archive_bytes": 53433, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 62, "github_languages": {"Python": 217090}, "github_archived": false, "github_pushed_at": "2024-01-26T18:19:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-machine-learning-in-latent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UyC1dXUA-n", "year": 2023, "status": "rejected", "title": "Design of the topology for contrastive visual-textual alignment", "authors": ["Zhun Sun"], "authorids": ["~Zhun_Sun1"], "authors_source": "OpenReview API", "abstract": "Pre-training weakly related image-text pairs in the contrastive style shows great power in learning semantic aligning cross-modal models. The common choice to measure the distance between the feature representations of the image-text pairs is the cosine similarity, which can be considered as the negative inner product distance of features embedded on a sphere, mathematically. However, empirically, aligning image-text pairs on the spherical topology is vulnerable to the semantic ambiguity phenomenon resulting from the noise in the pre-training datasets. Specifically, under the noisy training data, instead of the optimal alignment-uniformity solution, the system would achieve an equilibrium (a gap between distances of positive and negative pairs), when the gradients for attraction and repulsion are neutralized.  Although intuitively, the model should always find this equilibrium given a sufficiently long training scheme, its numerical values might be out of the distance range (e.g. [-1, 1] for the cosine similarity). In the practice of former studies, this problem is partly tackled by introducing a learnable softmax temperature parameter, in other words, by explicitly scaling the range of the distance function.  In this work, we alternatively design the topology of embedding space and its endowed distance function. Motivated by studies that make use of Riemannian geometry for visual tasks, we propose a rather simple solution to address the aforementioned equilibrium problem. That is, we map the feature representations onto the oblique manifold endowed with the negative inner product as the distance function. In the experimental analysis, we show that we can improve the baseline performance by a large margin (e.g. 4\\% in the zero-shot image to text retrieval task) by changing only two lines of the training codes.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gyx_Dc999FE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper131/Reviewer_NGQV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to employ the oblique manifold as the embedding space for contrastive visual-textual alignment learning, where the feature representations are mapped onto the oblique manifold endowed with the negative inner product as the distance function. Discussions and experiments on multiple different topologies endowed with different distances demonstrates the superiority and effectiveness of the proposed approach.", "review_text": "Overall, this paper utilized a new distance function for contrastive visual-textual learning, although simple and effective, this paper does not provide new things up with Oblique Manifold, and it lacks deeper analysis/demonstration on how the method solves existing problems. ", "strengths": "Strengths:\n- The proposed algorithm is easy to implement and proved effective with multiple datasets.\n- This paper provides detailed discussion on the properties of the embedding topology and make fair comparison with other topologies including Sphere, Euclidean, and Oblique.\n\nWeaknesses:\n- In Introduction, the authors discussed the noise problem in contrastive alignment learning, and it seems that the proposed embedding space would handle the problem better than others. However, it may lack the relevant discussions in the body part or experiments.\n- This paper claims that “the sphere and oblique manifold have a proper uniform distribution defined as the surface area measure”, which might be the reason why Sphere and Oblique performs better than Euclidean. However in Table 2, it seems that the conclusion would be largely affected by the Temperature init. \n- It may lack the discussion/visualization on the comparison of the learned distribution with Sphere and Oblique, which should give a more clear explanation on why Oblique performs better than Sphere.\n- Oblique Manifold has been utilized in other feature learning tasks, such as [i] and more discussions should be added.\n[i] Transductive Few-Shot Classification on the Oblique Manifold, ICCV 2021.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes to employ the oblique manifold as the embedding space for contrastive visual-textual alignment learning, where the feature representations are mapped onto the oblique manifold endowed with the negative inner product as the distance function. Discussions and experiments on multiple different topologies endowed with different distances demonstrates the superiority and effectiveness of the proposed approach.", "strength_and_weaknesses": "Strengths:\n- The proposed algorithm is easy to implement and proved effective with multiple datasets.\n- This paper provides detailed discussion on the properties of the embedding topology and make fair comparison with other topologies including Sphere, Euclidean, and Oblique.\n\nWeaknesses:\n- In Introduction, the authors discussed the noise problem in contrastive alignment learning, and it seems that the proposed embedding space would handle the problem better than others. However, it may lack the relevant discussions in the body part or experiments.\n- This paper claims that “the sphere and oblique manifold have a proper uniform distribution defined as the surface area measure”, which might be the reason why Sphere and Oblique performs better than Euclidean. However in Table 2, it seems that the conclusion would be largely affected by the Temperature init. \n- It may lack the discussion/visualization on the comparison of the learned distribution with Sphere and Oblique, which should give a more clear explanation on why Oblique performs better than Sphere.\n- Oblique Manifold has been utilized in other feature learning tasks, such as [i] and more discussions should be added.\n[i] Transductive Few-Shot Classification on the Oblique Manifold, ICCV 2021.\n\n", "clarity,_quality,_novelty_and_reproducibility": "This Paper is clearly written with good quality. The originality might be minor with using the Oblique Manifold in alignment learning. From the provided pseudo-code, this paper should be easy to reproduce.", "summary_of_the_review": "Overall, this paper utilized a new distance function for contrastive visual-textual learning, although simple and effective, this paper does not provide new things up with Oblique Manifold, and it lacks deeper analysis/demonstration on how the method solves existing problems. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666669751898}, {"id": "pnWtjW67WWW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper131/Reviewer_bLt8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work strives to learn a visual-textual embedding space from weakly label image-text pairs from the internet. \nIn a typical contrastive learning with cosine-similarity, the noise in training data prevents the deep network to produce a solution with alignment-uniformity, and the network achieves a suboptimal system of equilibrium instead. This work resorts to oblique manifold as embedding space that tackles the equilibrium problem. Empirical improvement on zero-shot image to text retrieval task is observed.", "review_text": "This paper proposes to learn visual-textual embedding space on oblique manifold. The idea is interesting and novel, and the paper is well written. The reviewer feels pleasant to read the paper. Waiting to see the clarification of motivations and some statements.", "strengths": "+ In general, this paper is easy to read. Formulation of the problem and description of the method are clear and rigorous.\n+ The idea of using oblique manifold to learn a visual-textual embedding space is novel.\n+ The finding on learnable softmax temperature is a distance scaling factor of contrastive learning on noisy dataset is interesting.\n\n- Missing some related literatures such as hyperbolic embedding space.\n\nThe motivation in choosing oblique manifold is a bit weak to some extent. The main arguments are 1) its geometry is spherical, where proper definition of uniformity is available, and 2) inner product saves computational cost.\n- Why the uniformity in embedding space is a valid goal to embed text, given the structure of text can be hierarchical?\n- While it is true that unbounded Euclidean space does not have proper uniform distribution defined (Sec 3.2 ii), no neural network is able to learn in unbounded space. We can simply clip to bounded range where uniform distribution is available.\n- In terms of computational cost, do authors take into consideration l2 normalization steps in their algorithm?  \n- Also, is the computation of distance function a bottleneck for optimizing underlying deep learning models? It sounds like a these difference in computational cost are marginal.\n- The statement of \"numerical values of distances at equilibrium might be out of the distance range (e.g. [-1,1] for the cosine similarity)\" is somehow unclear for me. Any two vectors naturally have a cosine similarity value constrained in [-1,1]. Why it might be out of range?\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work strives to learn a visual-textual embedding space from weakly label image-text pairs from the internet. \nIn a typical contrastive learning with cosine-similarity, the noise in training data prevents the deep network to produce a solution with alignment-uniformity, and the network achieves a suboptimal system of equilibrium instead. This work resorts to oblique manifold as embedding space that tackles the equilibrium problem. Empirical improvement on zero-shot image to text retrieval task is observed.", "strength_and_weaknesses": "+ In general, this paper is easy to read. Formulation of the problem and description of the method are clear and rigorous.\n+ The idea of using oblique manifold to learn a visual-textual embedding space is novel.\n+ The finding on learnable softmax temperature is a distance scaling factor of contrastive learning on noisy dataset is interesting.\n\n- Missing some related literatures such as hyperbolic embedding space.\n\nThe motivation in choosing oblique manifold is a bit weak to some extent. The main arguments are 1) its geometry is spherical, where proper definition of uniformity is available, and 2) inner product saves computational cost.\n- Why the uniformity in embedding space is a valid goal to embed text, given the structure of text can be hierarchical?\n- While it is true that unbounded Euclidean space does not have proper uniform distribution defined (Sec 3.2 ii), no neural network is able to learn in unbounded space. We can simply clip to bounded range where uniform distribution is available.\n- In terms of computational cost, do authors take into consideration l2 normalization steps in their algorithm?  \n- Also, is the computation of distance function a bottleneck for optimizing underlying deep learning models? It sounds like a these difference in computational cost are marginal.\n- The statement of \"numerical values of distances at equilibrium might be out of the distance range (e.g. [-1,1] for the cosine similarity)\" is somehow unclear for me. Any two vectors naturally have a cosine similarity value constrained in [-1,1]. Why it might be out of range?\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "\nThe paper is clear written and easy to follow. The problem and methodology are nicely formulated.\nIf I have to name a few flaws in presentation, there are some symbols do not can clear explanations. For example:\n- Eq. (2)  H(.|.) denotes cross-entropy loss.\n- [CLS] in section 3.1.\n\nThe idea sound novel, but the motivation needs to be strengthened.\nThe reviewer trust the reproducibility of this work.", "summary_of_the_review": "This paper proposes to learn visual-textual embedding space on oblique manifold. The idea is interesting and novel, and the paper is well written. The reviewer feels pleasant to read the paper. Waiting to see the clarification of motivations and some statements.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666669605251}, {"id": "okKSr0h6AXH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper131/Reviewer_W5p4"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper designs a new distance metric called Oblique(d/m, m) instead of the fashionable cosine similarity for the contrastive learning paradigm. To verify the effectiveness of the proposed simple method, the authors perform extensive experiments. ", "review_text": "The paper is not well written and hard to follow. The experiments are not solid for comprehensively evaluating the proposed method.", "strengths": "##Strength：\nThe proposed new distance metric is neat and could be easily used for the contrastive learning paradigm.\n\n##Weaknesses:\nAlthough the proposed method could be easily generalized to the existing multi-modal pretraining works using the contrastive learning paradigm, the effects cannot be verified well in the provided experiment results. As shown in Table 4, using the proposed metric only brings limited retrieval performance improvement compared to the vanilla baseline. The authors could provide more results to verify and defend the superiority of the proposed method.\n\nAlthough the technical implementation of the proposed method is clear, there are some statements and arguments needed to be further claimed. \n- “Here, the problem is that the numerical values of distances at equilibrium might be out of the distance range (e.g. [-1, 1] for the cosine similarity).” Why the numerical values of distances could be out of the range of [-1, 1] for the cosine similarity?\n- “For instance, if there is a reasonable amount of false negative samples, then the model would learn a smaller negative distance for not being punished too hard when encountering false negative samples.”\n\n- The reviewer finds that the font of the paper is different from other submission manuscripts. Does the paper use the proper template?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper designs a new distance metric called Oblique(d/m, m) instead of the fashionable cosine similarity for the contrastive learning paradigm. To verify the effectiveness of the proposed simple method, the authors perform extensive experiments. ", "strength_and_weaknesses": "##Strength：\nThe proposed new distance metric is neat and could be easily used for the contrastive learning paradigm.\n\n##Weaknesses:\nAlthough the proposed method could be easily generalized to the existing multi-modal pretraining works using the contrastive learning paradigm, the effects cannot be verified well in the provided experiment results. As shown in Table 4, using the proposed metric only brings limited retrieval performance improvement compared to the vanilla baseline. The authors could provide more results to verify and defend the superiority of the proposed method.\n\nAlthough the technical implementation of the proposed method is clear, there are some statements and arguments needed to be further claimed. \n- “Here, the problem is that the numerical values of distances at equilibrium might be out of the distance range (e.g. [-1, 1] for the cosine similarity).” Why the numerical values of distances could be out of the range of [-1, 1] for the cosine similarity?\n- “For instance, if there is a reasonable amount of false negative samples, then the model would learn a smaller negative distance for not being punished too hard when encountering false negative samples.”\n\n- The reviewer finds that the font of the paper is different from other submission manuscripts. Does the paper use the proper template?\n", "clarity,_quality,_novelty_and_reproducibility": "Quality: The paper is likely to have a modest impact on the community. Clarity: The paper is well organized but the presentation has minor details that could be improved. Originality: The main ideas of the paper are not novel or have limited novelty. Please see the weaknesses.", "summary_of_the_review": "The paper is not well written and hard to follow. The experiments are not solid for comprehensively evaluating the proposed method.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666605008510}, {"id": "pIcmH-fepPU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper131/Reviewer_YGrh"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposed a strategy to make the temperature learnable. ", "review_text": "This method is too simple and lacks novelty. The authors impose the proposed method related to manifold learning, but it is not well proven and discussed theoretically. The experiments cannot well demonstrate the effectiveness of the proposed method.", "strengths": "Strength:\nThis paper proposed a simple strategy to make the temperature learnable. The proposed method simply changes two lines of the training codes to improve the performance.\n\nWeakness:\n1. This paper is badly written and hard to follow.\n2. The key idea of the paper is to make the temperature learnable, which is straight and lacks novelty.\n3. The proposed method is pretended too much. The authors impose the proposed method related to manifold learning, but it is not well proven and discussed theoretically.\n4. The experiments cannot comprehensively demonstrate the effectiveness of the proposed method. From the experimental results, one could see that the proposed method could only improve slightly.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed a strategy to make the temperature learnable. ", "strength_and_weaknesses": "Strength:\nThis paper proposed a simple strategy to make the temperature learnable. The proposed method simply changes two lines of the training codes to improve the performance.\n\nWeakness:\n1. This paper is badly written and hard to follow.\n2. The key idea of the paper is to make the temperature learnable, which is straight and lacks novelty.\n3. The proposed method is pretended too much. The authors impose the proposed method related to manifold learning, but it is not well proven and discussed theoretically.\n4. The experiments cannot comprehensively demonstrate the effectiveness of the proposed method. From the experimental results, one could see that the proposed method could only improve slightly.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is badly written and hard to follow.  The idea is straightforward and lacks novelty. The authors impose the proposed method related to manifold learning, but it is not well proven and discussed theoretically.", "summary_of_the_review": "This method is too simple and lacks novelty. The authors impose the proposed method related to manifold learning, but it is not well proven and discussed theoretically. The experiments cannot well demonstrate the effectiveness of the proposed method.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666336660992}], "openreview_url": "https://openreview.net/forum?id=UyC1dXUA-n", "arxiv_id": "2209.02127", "paper_pdf": "papers/UyC1dXUA-n.pdf", "paper_pdf_sha256": "2a4b6dac049ff385c47a041b62279cc45fe2ed2df44ecda6d322ddc28b29384a", "paper_pdf_bytes": 904016, "paper_pdf_source": "openreview", "code_url": "https://github.com/minogame/clip-mtob", "code_repository": "minogame/clip-mtob", "code_commit": "15fa1ec4fac43e1dd15a5040aaba60ca2e513caa", "code_archive": "repos/UyC1dXUA-n.zip", "code_archive_sha256": "003bd6baed66626366e297c2f0b05d5eac66854b00ec5be90eb5e9423689865e", "code_archive_bytes": 137574, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 131, "github_languages": {"Python": 41882}, "github_archived": false, "github_pushed_at": "2023-10-09T04:52:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/design-of-the-topology-for-contrastive-visual"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QkfMWTl520U", "year": 2022, "status": "rejected", "title": "When do Convolutional Neural Networks Stop Learning?", "authors": ["SAHAN AHMAD", "Aminul Islam"], "authorids": ["~SAHAN_AHMAD1", "aminul@louisiana.edu"], "authors_source": "OpenReview API", "abstract": "Convolutional Neural Networks (CNNs) is one of the most essential architectures that has shown impressive performance in computer vision tasks such as image classification, detection, and segmentation. In training phase of CNN, an arbitrary number of epochs is used to train the neural networks. In a single epoch, the entire training data---divided by batch size---are fed to the network. However, the optimal number of epochs required to train a neural network is not well established. In practice, validation data is used to identify the generalization gap. To avoid overfitting, it is recommended to stop training when the generalization gap increases. However, this is a trial and error based approach. This raises a critical question: Is it possible to estimate when neural networks stop learning based on only the training data? In this research work, we introduce the stability property of data in layers and based on this property, we predict the near optimal epoch number of a CNN. We do not use any validation data to predict the near optimal epoch number. We experiment our hypothesis on six different CNN models and on three different datasets (CIFIR 10, CIFIR 100, SVHN). We save on average 58.49\\% computational time to train a CNN model. Our code is available at https://github.com/PaperUnderReviewDeepLearning/Optimization.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "85KLrfBVqcx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1162/Reviewer_P5xn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a data stability measure for early stopping in training various CNN architectures (RestNet18, VGG16), and a shallow CNN. Experiments demonstrate that comparing the means of layer weights across epochs to a certain number of decimal places allows for a significant (~30%-~70%) computational time savings with a relatively small drop in accuracy. Methodology is thoroughly quantitatively evaluation on classification networks, and a detailed ablation study in presented. ", "review_text": "Strengths:\n- A simple method for early stopping based on network layer weight averages\n- Offers a significant computational time savings\n- Technique can be used with any network architecture\n- Code is provided\n \nWeaknesses:\n- Evaluation is performed only on classification networks. How well would this approach for detection, segmentation? How does the number of layers affect performance?\n- It is surprising that precision of only two decimal places is needed, given that performance is averaged over such a large number of parameters. \n- As neural networks have a large number of parameters and are ever growing, what is the memory and computational cost of the proposed approach? How much does it slow down training\n- In Table 3, using the proposed method mainly makes results worse for CIFAR10, but improves results for CIFAR100, SVHN in some cases. A discussion of these findings would be helpful. A\n- Number of grammatical errors make the paper difficult to follow\n- Limited discussion of previous works, especially in early stopping and regularization methods, and no comparison to previous works, e.g., the basic validation set strategy.\n\nVery similar ideas are found in the following papers:\nMahsereci, M., Balles, L., Lassner, C., & Hennig, P. (2017). Early stopping without a validation set. arXiv preprint arXiv:1703.09580.\n\nBonet, D., Ortega, A., Ruiz-Hidalgo, J., & Shekkizhar, S. (2021). Channel-Wise Early Stopping without a Validation Set via NNK Polytope Interpolation. arXiv e-prints, APSIPA 2021\n\nI would also suggest to include some references to uncertainty estimation in neural networks, e.g., \nMaddox, W. J., Izmailov, P., Garipov, T., Vetrov, D. P., & Wilson, A. G. (2019). A simple baseline for bayesian uncertainty in deep learning. Advances in Neural Information Processing Systems, 32, 13153-13164.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a data stability measure for early stopping in training various CNN architectures (RestNet18, VGG16), and a shallow CNN. Experiments demonstrate that comparing the means of layer weights across epochs to a certain number of decimal places allows for a significant (~30%-~70%) computational time savings with a relatively small drop in accuracy. Methodology is thoroughly quantitatively evaluation on classification networks, and a detailed ablation study in presented. ", "main_review": "Strengths:\n- A simple method for early stopping based on network layer weight averages\n- Offers a significant computational time savings\n- Technique can be used with any network architecture\n- Code is provided\n \nWeaknesses:\n- Evaluation is performed only on classification networks. How well would this approach for detection, segmentation? How does the number of layers affect performance?\n- It is surprising that precision of only two decimal places is needed, given that performance is averaged over such a large number of parameters. \n- As neural networks have a large number of parameters and are ever growing, what is the memory and computational cost of the proposed approach? How much does it slow down training\n- In Table 3, using the proposed method mainly makes results worse for CIFAR10, but improves results for CIFAR100, SVHN in some cases. A discussion of these findings would be helpful. A\n- Number of grammatical errors make the paper difficult to follow\n- Limited discussion of previous works, especially in early stopping and regularization methods, and no comparison to previous works, e.g., the basic validation set strategy.\n\nVery similar ideas are found in the following papers:\nMahsereci, M., Balles, L., Lassner, C., & Hennig, P. (2017). Early stopping without a validation set. arXiv preprint arXiv:1703.09580.\n\nBonet, D., Ortega, A., Ruiz-Hidalgo, J., & Shekkizhar, S. (2021). Channel-Wise Early Stopping without a Validation Set via NNK Polytope Interpolation. arXiv e-prints, APSIPA 2021\n\nI would also suggest to include some references to uncertainty estimation in neural networks, e.g., \nMaddox, W. J., Izmailov, P., Garipov, T., Vetrov, D. P., & Wilson, A. G. (2019). A simple baseline for bayesian uncertainty in deep learning. Advances in Neural Information Processing Systems, 32, 13153-13164.", "summary_of_the_review": "The paper proposes a simple technique for early stopping in neural networks. However, limited discussion and comparisons to previous methods makes it challenging to understand how this technique can be used in practice. I would suggest the authors increase the discussion of related work, compare existing early stopping methods stated above, and explain the computational complexity and memory overhead required. I would also like to see an evaluation on non-classification tasks. \n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethics concerns.", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635886096144}, {"id": "YSLwiuZDtR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1162/Reviewer_QVY8"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The authors propose a method to perform early stopping in CNNs based on the training data alone, by measuring when the mean of the variance of the convolutional layers does not change much across several epochs. They show that for several architectures and datasets their method could be used to save computation time, in the cost of minor changes in the accuracy.", "review_text": "The paper in its current state feels like a very rough draft. As a non-English speaker, I prefer not to judge a work by the level of its English, but in this case, it is extreme. The paper is full of grammatical errors, spelling errors, and even misuses the names of commonly used datasets (CIFIR instead of CIFAR, for example).\n\nThe introduction of the paper is long and repetitive and goes over some points several times. The related work section cites works which I do not understand why should be relevant to this analysis.\n\nThe suggested method is also problematic. The authors introduce a new hyper-parameter, \"r\", which decides when the variance of each convolutional layer did not change \"much\", so learning should seize. I would expect this hyper-parameter to drastically change when increasing the learning rate and is highly dependant on the chosen architecture and the training data. Tuning this hyper-parameter will require some validation dataset, hence undermining the biggest contribution the paper claims to have - that its proposed method can do not require any additional validation dataset.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a method to perform early stopping in CNNs based on the training data alone, by measuring when the mean of the variance of the convolutional layers does not change much across several epochs. They show that for several architectures and datasets their method could be used to save computation time, in the cost of minor changes in the accuracy.", "main_review": "The paper in its current state feels like a very rough draft. As a non-English speaker, I prefer not to judge a work by the level of its English, but in this case, it is extreme. The paper is full of grammatical errors, spelling errors, and even misuses the names of commonly used datasets (CIFIR instead of CIFAR, for example).\n\nThe introduction of the paper is long and repetitive and goes over some points several times. The related work section cites works which I do not understand why should be relevant to this analysis.\n\nThe suggested method is also problematic. The authors introduce a new hyper-parameter, \"r\", which decides when the variance of each convolutional layer did not change \"much\", so learning should seize. I would expect this hyper-parameter to drastically change when increasing the learning rate and is highly dependant on the chosen architecture and the training data. Tuning this hyper-parameter will require some validation dataset, hence undermining the biggest contribution the paper claims to have - that its proposed method can do not require any additional validation dataset.\n", "summary_of_the_review": "At the current state of this work, I do not think it meets the standards of a Tier 1 conference.\nI highly recommend rejecting this paper at the current state.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635884163530}, {"id": "GOzn6MIRA6n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1162/Reviewer_sbHo"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper presents a method to measure the stability of vectors obtaining during training of a CNN as proxy measure for convergence of the model. This is used in a early-stopping fashion to allow interruption of the training procedure, instead of training for a fixed number of iterations/epochs. In comparison with training for a fixed number of epochs and using Curriculum by Smoothing (CBS), the proposed method is competitive in terms of final accuracy, while reducing the number of total epochs. Also, stability measured in 3 benchmark datasets behave similarly.", "review_text": "Strengths\n- The idea of computing stability of training per-layer and use it to stop learning has merit and it is worth pursuing.\n- There is empirical evidence that support the hypothesis of stability as a proxy measure for the rate of learning/changes in the model which can be helpful not only by reducing number of epochs and saving computational time, but also potential to prevent overtraining\n- I liked the analysis of initial-curved-stable phases, which may help evaluating convergence in practice\n\nWeaknesses\n- My main concern with the paper is the comparison and baselines used. Although CBS and vanilla-learning approaches are possible baselines, I missed the use of a regular early-stopping approach using the difference of the losses of two consecutive iterations as a stopping criterion, and a small validation set to avoid overtraining. Although the proposed method do not use/require a validation set, for the proposed benchmark datasets, it is feasible to separate at least 5% of the data for this purpose. Note the authors do not work on the premise of small data.\n- Comparing 200 epochs vs a reduced number of epochs only by its absolute value results in a poor discussion. The authors could have reported for example the \"estimated generalization\" which is the difference between the training and validation/test loss.  The final loss value of all models should also be reported so that some conclusions could be drawn from this.\n- The ablation study does not look like an \"ablation\", but just a combination of the proposed method with CBS. I recommend the authors to look for the definition of ablation.\n- It is true that a fixed \"safe\" number of epochs is used, but this is because this approach allows inspecting for previous epoch and selecting the model that better fits some criterion (e.g. training vs validation loss). By stopping the training using the proposed method, maybe we could not know what could happen in future iterations\n- Including a joint analysis with learning rate scheduling would also be important. For example, in cosine annealing, the premise of the phases that appear using a fixed learning rate can drastically change.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents a method to measure the stability of vectors obtaining during training of a CNN as proxy measure for convergence of the model. This is used in a early-stopping fashion to allow interruption of the training procedure, instead of training for a fixed number of iterations/epochs. In comparison with training for a fixed number of epochs and using Curriculum by Smoothing (CBS), the proposed method is competitive in terms of final accuracy, while reducing the number of total epochs. Also, stability measured in 3 benchmark datasets behave similarly.", "main_review": "Strengths\n- The idea of computing stability of training per-layer and use it to stop learning has merit and it is worth pursuing.\n- There is empirical evidence that support the hypothesis of stability as a proxy measure for the rate of learning/changes in the model which can be helpful not only by reducing number of epochs and saving computational time, but also potential to prevent overtraining\n- I liked the analysis of initial-curved-stable phases, which may help evaluating convergence in practice\n\nWeaknesses\n- My main concern with the paper is the comparison and baselines used. Although CBS and vanilla-learning approaches are possible baselines, I missed the use of a regular early-stopping approach using the difference of the losses of two consecutive iterations as a stopping criterion, and a small validation set to avoid overtraining. Although the proposed method do not use/require a validation set, for the proposed benchmark datasets, it is feasible to separate at least 5% of the data for this purpose. Note the authors do not work on the premise of small data.\n- Comparing 200 epochs vs a reduced number of epochs only by its absolute value results in a poor discussion. The authors could have reported for example the \"estimated generalization\" which is the difference between the training and validation/test loss.  The final loss value of all models should also be reported so that some conclusions could be drawn from this.\n- The ablation study does not look like an \"ablation\", but just a combination of the proposed method with CBS. I recommend the authors to look for the definition of ablation.\n- It is true that a fixed \"safe\" number of epochs is used, but this is because this approach allows inspecting for previous epoch and selecting the model that better fits some criterion (e.g. training vs validation loss). By stopping the training using the proposed method, maybe we could not know what could happen in future iterations\n- Including a joint analysis with learning rate scheduling would also be important. For example, in cosine annealing, the premise of the phases that appear using a fixed learning rate can drastically change.\n\n", "summary_of_the_review": "The paper presents interesting ideas and show experiments with promising results. However, better baselines should be evaluated and more discussion around the effects of early stopping vs stability is important before this paper is ready for publication. I believe the claims are not well supported without comparing with basic stopping criteria (with or without a validation set) and discussing the results. In summary, I believe it is a good initial set of results, but not a work that is ready/mature enough for publication.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635818593216}, {"id": "p5j6cuqg5dW", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1162/Reviewer_mbyZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper describes a method to define the optimal number of epochs for training process without using validation set.\n\nThe work proposes to compute, for each epoch and for each layer, the stability vector defined as the vector containing the standard deviation of the activations for each CNN layer, for all iterations of a particular epoch.\n\nAt each epoch, authors compute the mean of the stability vector for each layer (average over the iterations). If the difference between the mean of the stability vectors of two successive epochs for a layer is small, we can assume that the layer is stable for the considered epochs. If this happens for all layers, we can stop the training process. \n\nThe proposed approach is of simple implementation and the authors tested it on 6 different CNN based architectures for classification on CIFAR10, CIFAR100 and SVHN datasets.", "review_text": "The approach is simple, explained quite clearly and its implementation is straightforward.\n\nThe main assumption of the work is that the average (over the iteration) of the standard deviation of the activations of the layers is representative of the status of the learning process.\n\nI have some concerns about this hypothesis:\n1. The standard deviation is a measure of dispersion around the mean but the histograms of the activations are not provided to understand if there is a measure that better takes in account the variability inside the activations (e.g. Inter-quartile range could be more precise to capture the effective variability)\n2. The average (eq. 2) can mask some fluctuations that can occur in different iterations. In this way if there are similar changes in two iterations but with opposite sign, this information is lose and we can conclude that, for that epoch, there is a sort of stability but this is not the case. \n\nMoreover, in my opinions, other weakness are:\n- This process depends also on the selected optimization algorithm and/or initial seed and/or some tricks that can be implemented (e.g. reduce learning rate on plateau) but these situations are not considered neither tested. For example the different initial seeds are not properly considered in the experiments.\n- The limitations of the approach should be clearly described (e.g. is it applicable only to Conv Layers or also to Fully connected layers? Is it applicable in presence of Batch Normalization? etc.)\n- Also Batch Normalization involves the estimation of the standard deviation of the activations. Even though its role is different, introduce regularity during the learning, it is neither described nor simply cited.\n- The authors described the impact of the value of r on the selection of the \"optimal\" epoch and it emerges that increasing the value of r doesn't lead to a good stopping criterion. This high sensitivity to the number of significative digits seems to suggest that the proposed approach isn't enough robust. \n- The authors tested their approach comparing it with the fixed epochs defined in some works (i.e. 200). Since the best practice, as also authors describe in the introduction, is the early stopping, a comparison with this approach should be done and the Computational Time Saving should be tested respect to the early stopping situation. \n- The authors cite the Double Descent phenomenon but the epoch-wise double descent is not cited and, in this case, it is important.\n- The \"Related work\" section should be increased since several works that try to assess the learning status are not cited (e.g. loss landscape, relation with the flatness/curvature of the loss, weight watchers etc.) \n\nOther observations:\n- More details should be provided about the optimization algorithms used in tested architecture (e.g. type, learning rate, weights' initialization, etc.)\n- From Figure 1 looks like the Stability Vector is computed only for Conv Layers (also the caption confirms that) but in the introduction the authors wrote that \"we examine the data variation across all layers of a CNN architecture\") \n- In the Figure 2, it is used $S_i$ instead of $S_{ie}$\n- The $\\alpha$ should contain also the information about the related epoch (e.g. equation 2 or Figure 3), for instance $\\alpha_{nt}^e$ \n- The authors should clarify that the operation used in eq. 1 performed by CNNs is different by the mathematical convolution operation, it is a correlation.\n- $X_{nt}$ is not properly defined\n- In the paper often CIFIR is used instead of correct CIFAR\n- Figure 5 would be clearer if the authors presented the complete plot for each layer differentiating the several regions by colors.\n- The title, captions, legends and descriptions of **all** figures should be improved.\n- Page 9: dataser --> dataset\n- Page 3: convectional --> do you mean: conventional ?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper describes a method to define the optimal number of epochs for training process without using validation set.\n\nThe work proposes to compute, for each epoch and for each layer, the stability vector defined as the vector containing the standard deviation of the activations for each CNN layer, for all iterations of a particular epoch.\n\nAt each epoch, authors compute the mean of the stability vector for each layer (average over the iterations). If the difference between the mean of the stability vectors of two successive epochs for a layer is small, we can assume that the layer is stable for the considered epochs. If this happens for all layers, we can stop the training process. \n\nThe proposed approach is of simple implementation and the authors tested it on 6 different CNN based architectures for classification on CIFAR10, CIFAR100 and SVHN datasets.", "main_review": "The approach is simple, explained quite clearly and its implementation is straightforward.\n\nThe main assumption of the work is that the average (over the iteration) of the standard deviation of the activations of the layers is representative of the status of the learning process.\n\nI have some concerns about this hypothesis:\n1. The standard deviation is a measure of dispersion around the mean but the histograms of the activations are not provided to understand if there is a measure that better takes in account the variability inside the activations (e.g. Inter-quartile range could be more precise to capture the effective variability)\n2. The average (eq. 2) can mask some fluctuations that can occur in different iterations. In this way if there are similar changes in two iterations but with opposite sign, this information is lose and we can conclude that, for that epoch, there is a sort of stability but this is not the case. \n\nMoreover, in my opinions, other weakness are:\n- This process depends also on the selected optimization algorithm and/or initial seed and/or some tricks that can be implemented (e.g. reduce learning rate on plateau) but these situations are not considered neither tested. For example the different initial seeds are not properly considered in the experiments.\n- The limitations of the approach should be clearly described (e.g. is it applicable only to Conv Layers or also to Fully connected layers? Is it applicable in presence of Batch Normalization? etc.)\n- Also Batch Normalization involves the estimation of the standard deviation of the activations. Even though its role is different, introduce regularity during the learning, it is neither described nor simply cited.\n- The authors described the impact of the value of r on the selection of the \"optimal\" epoch and it emerges that increasing the value of r doesn't lead to a good stopping criterion. This high sensitivity to the number of significative digits seems to suggest that the proposed approach isn't enough robust. \n- The authors tested their approach comparing it with the fixed epochs defined in some works (i.e. 200). Since the best practice, as also authors describe in the introduction, is the early stopping, a comparison with this approach should be done and the Computational Time Saving should be tested respect to the early stopping situation. \n- The authors cite the Double Descent phenomenon but the epoch-wise double descent is not cited and, in this case, it is important.\n- The \"Related work\" section should be increased since several works that try to assess the learning status are not cited (e.g. loss landscape, relation with the flatness/curvature of the loss, weight watchers etc.) \n\nOther observations:\n- More details should be provided about the optimization algorithms used in tested architecture (e.g. type, learning rate, weights' initialization, etc.)\n- From Figure 1 looks like the Stability Vector is computed only for Conv Layers (also the caption confirms that) but in the introduction the authors wrote that \"we examine the data variation across all layers of a CNN architecture\") \n- In the Figure 2, it is used $S_i$ instead of $S_{ie}$\n- The $\\alpha$ should contain also the information about the related epoch (e.g. equation 2 or Figure 3), for instance $\\alpha_{nt}^e$ \n- The authors should clarify that the operation used in eq. 1 performed by CNNs is different by the mathematical convolution operation, it is a correlation.\n- $X_{nt}$ is not properly defined\n- In the paper often CIFIR is used instead of correct CIFAR\n- Figure 5 would be clearer if the authors presented the complete plot for each layer differentiating the several regions by colors.\n- The title, captions, legends and descriptions of **all** figures should be improved.\n- Page 9: dataser --> dataset\n- Page 3: convectional --> do you mean: conventional ?", "summary_of_the_review": "The work describes a simple criterion to decide when stop the learning procedure.\nEven though there are some interesting aspects, this work looks like an initial investigation but more experimental and/or theoretical demonstrations are required.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635169935085}], "openreview_url": "https://openreview.net/forum?id=QkfMWTl520U", "arxiv_id": "2403.02473", "paper_pdf": "papers/QkfMWTl520U.pdf", "paper_pdf_sha256": "df8543ffc73d6cdfe712e9ffe9bc108d01595f401f986a0bd399d96265f6ecda", "paper_pdf_bytes": 4504983, "paper_pdf_source": "openreview", "code_url": "https://github.com/PaperUnderReviewDeepLearning/Optimization", "code_repository": "PaperUnderReviewDeepLearning/Optimization", "code_commit": "a1029d56c090003289270c382ea81c82166ae22a", "code_archive": "repos/QkfMWTl520U.zip", "code_archive_sha256": "44b597c662f50b3001b1df408a5296496866856f016bba1760ad5b5900afa316", "code_archive_bytes": 155703, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 178, "github_languages": {"Python": 83145}, "github_archived": false, "github_pushed_at": "2022-10-04T15:30:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/when-do-convolutional-neural-networks-stop-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xyGFYKIPTDJ", "year": 2021, "status": "rejected", "title": "Learning Causal Semantic Representation for Out-of-Distribution Prediction", "authors": ["Chang Liu", "Xinwei Sun", "Jindong Wang", "Tao Li", "Tao Qin", "Wei Chen", "Tie-Yan Liu"], "authorids": ["~Chang_Liu10", "~Xinwei_Sun1", "~Jindong_Wang1", "~Tao_Li9", "~Tao_Qin1", "~Wei_Chen1", "~Tie-Yan_Liu1"], "authors_source": "OpenReview API", "abstract": "Conventional supervised learning methods, especially deep ones, are found to be sensitive to out-of-distribution (OOD) examples, largely because the learned representation mixes the semantic factor with the variation factor due to their domain-specific correlation, while only the semantic factor causes the output. To address the problem, we propose a Causal Semantic Generative model (CSG) based on causality to model the two factors separately, and learn it on a single training domain for prediction without (OOD generalization) or with unsupervised data (domain adaptation) in a test domain. We prove that CSG identifies the semantic factor on the training domain, and the invariance principle of causality subsequently guarantees the boundedness of OOD generalization error and the success of adaptation. We also design novel and delicate learning methods for both effective learning and easy prediction, following the first principle of variational Bayes and the graphical structure of CSG. Empirical study demonstrates the effect of our methods to improve test accuracy for OOD generalization and domain adaptation.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "vkOkjAaNV_q", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1774/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**\nThe paper focuses on the causal perspective of domain-generalization and domain adaptation setup for images. I.e. classifying an image under some distribution shift at test time. Similar to previous work [1-4], it assumes that some latent semantic-object representation (s) and semantic-domain representation (v) cause the image, and that these causal (generative) mechanisms (s,v ->x ) are stable, while their prior p(s,v) is prone to change at test time. It develops a new variational approach to estimate the generative distributions, and test the approach on two datasets for domain-generalization and domain-adaptation.\n\nOverall, the paper suggests a novel approach and theory to an important problem. Its major weaknesses are in its clarity and the experimental part. \n\n**Strong points**\nNovelty: The paper provides a novel approach for estimating the likelihood of p(class|image), by developing a new variational approach for modelling the causal direction (s,v->x).\nCorrectness: Although I didn’t verify the details of the proofs, the approach seems technically correct. Note that I was not convinced that s->y (see weakness)\n\n**Weak points**\nExperiments and Reproducibility:\nThe experiments show some signal, but are not through enough:\n• shifted-MNIST: it is not clear why shift=0 is much better than shift~$N(0,\\sigma^2)$, since both cases incorporate a domain shift\n• It would be useful to show the performance the model and baselines on test samples from the observational (in) distribution.\n• Missing details about evaluation split for shifted-MNIST: Did the experiments used a validation set for hyper-param search with shifted-MNIST and ImageCLEF? Was it based on in-distribution data or OOD data?\n• It would be useful to provide an ablation study, since the approach has a lot of \"moving parts\".\n• It would be useful to have an experiment on an additional dataset, maybe more controlled than ImageCLEF, but less artificial than shifted-MNIST.\n• What were the ranges used for hyper-param search? What was the search protocol?\n\nClarity:\n• The parts describing the method are hard to follow, it will be useful to improve their clarity.\n• It will be beneficial to explicitly state which are the learned parametrized distributions, and how inference is applied with them.\n• What makes the VAE inference mappings (x->s,v) stable to domain shift? E.g. [1] showed that correlated latent properties in VAEs are not robust to such domain shifts.\n• What makes v distinctive of s? Is it because y only depends on s?\n• Does the approach uses any information on the labels of the domain?\n\nCorrectness: I was not convinced about the causal relation s->y. I.e. that the semantic concept cause the label, independently of the image. I do agree that there is a semantic concept (e.g. s) that cause the image. But then, as explained by [Arjovsky 2019] the labelling process is caused by the image. I.e. s->image->y, and not as argued by the paper. The way I see it, is like a communication channel: y_tx -> s -> image -> y_rx. Could the authors elaborate how the model will change if replacing s->y by y_tx->s ?\n\n\n**Other comments:**\n• I suggest discussing [2,3,4], which learned similar stable mechanisms in images.\n• I am not sure about the statement that this work is the \"first to identify the semantic factor and leverage causal invariance for OOD prediction\" e.g. see [3,4]\n• The title may be confusing. OOD usually refers to anomaly-detection, while this paper relates to domain-generalization and domain-adaptation.\n• It will be useful to clarify that the approach doesn't use any external-semantic-knowledge.\n• Section 3.2 - I suggest to add a first sentence to introduce what this section is about.\n• About remark in page 6: (1) what is a deterministic s-v relation? (2) chairs can also appear in a workspace, and it may help to disentangle the desks from workspaces.\n\t\n[1] Suter et al. 2018, Robustly Disentangled Causal Mechanisms: Validating Deep Representations for Interventional Robustness\n[2] Besserve et al. 2020, Counterfactuals uncover the modular structure of deep generative models\n[3] Heinze-Deml et al. 2017, Conditional Variance Penalties and Domain Shift Robustness\n[4] Atzmon et al. 2020, A causal view of compositional zero-shot recognition\n\n\n\n\n\n**EDIT: Post rebuttal**\n\nI thank the authors for their reply. Although the authors answered most of my questions, I decided to keep the score as is, because I share similar concerns with R2 about the presentation, and because experiments are still lacking. \n\nAdditionally, I am concerned with one of the author's replies saying *All methods achieve accuracy 1 ... on the training distribution*, because usually there is a trade-off between accuracy on the observational distribution versus the shifted distribution (discussed by Rothenhäusler, 2018 [Anchor regression]): Achieving perfect accuracy on the observational distribution, usually means relying on the spurious correlations. And under domain-shift scenarios, this would hinder the performance on the shifted-distribution.\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Strengths: Novel approach and theory. Weaknesses: Clarity and experiments", "review": "**Summary**\nThe paper focuses on the causal perspective of domain-generalization and domain adaptation setup for images. I.e. classifying an image under some distribution shift at test time. Similar to previous work [1-4], it assumes that some latent semantic-object representation (s) and semantic-domain representation (v) cause the image, and that these causal (generative) mechanisms (s,v ->x ) are stable, while their prior p(s,v) is prone to change at test time. It develops a new variational approach to estimate the generative distributions, and test the approach on two datasets for domain-generalization and domain-adaptation.\n\nOverall, the paper suggests a novel approach and theory to an important problem. Its major weaknesses are in its clarity and the experimental part. \n\n**Strong points**\nNovelty: The paper provides a novel approach for estimating the likelihood of p(class|image), by developing a new variational approach for modelling the causal direction (s,v->x).\nCorrectness: Although I didn’t verify the details of the proofs, the approach seems technically correct. Note that I was not convinced that s->y (see weakness)\n\n**Weak points**\nExperiments and Reproducibility:\nThe experiments show some signal, but are not through enough:\n• shifted-MNIST: it is not clear why shift=0 is much better than shift~$N(0,\\sigma^2)$, since both cases incorporate a domain shift\n• It would be useful to show the performance the model and baselines on test samples from the observational (in) distribution.\n• Missing details about evaluation split for shifted-MNIST: Did the experiments used a validation set for hyper-param search with shifted-MNIST and ImageCLEF? Was it based on in-distribution data or OOD data?\n• It would be useful to provide an ablation study, since the approach has a lot of \"moving parts\".\n• It would be useful to have an experiment on an additional dataset, maybe more controlled than ImageCLEF, but less artificial than shifted-MNIST.\n• What were the ranges used for hyper-param search? What was the search protocol?\n\nClarity:\n• The parts describing the method are hard to follow, it will be useful to improve their clarity.\n• It will be beneficial to explicitly state which are the learned parametrized distributions, and how inference is applied with them.\n• What makes the VAE inference mappings (x->s,v) stable to domain shift? E.g. [1] showed that correlated latent properties in VAEs are not robust to such domain shifts.\n• What makes v distinctive of s? Is it because y only depends on s?\n• Does the approach uses any information on the labels of the domain?\n\nCorrectness: I was not convinced about the causal relation s->y. I.e. that the semantic concept cause the label, independently of the image. I do agree that there is a semantic concept (e.g. s) that cause the image. But then, as explained by [Arjovsky 2019] the labelling process is caused by the image. I.e. s->image->y, and not as argued by the paper. The way I see it, is like a communication channel: y_tx -> s -> image -> y_rx. Could the authors elaborate how the model will change if replacing s->y by y_tx->s ?\n\n\n**Other comments:**\n• I suggest discussing [2,3,4], which learned similar stable mechanisms in images.\n• I am not sure about the statement that this work is the \"first to identify the semantic factor and leverage causal invariance for OOD prediction\" e.g. see [3,4]\n• The title may be confusing. OOD usually refers to anomaly-detection, while this paper relates to domain-generalization and domain-adaptation.\n• It will be useful to clarify that the approach doesn't use any external-semantic-knowledge.\n• Section 3.2 - I suggest to add a first sentence to introduce what this section is about.\n• About remark in page 6: (1) what is a deterministic s-v relation? (2) chairs can also appear in a workspace, and it may help to disentangle the desks from workspaces.\n\t\n[1] Suter et al. 2018, Robustly Disentangled Causal Mechanisms: Validating Deep Representations for Interventional Robustness\n[2] Besserve et al. 2020, Counterfactuals uncover the modular structure of deep generative models\n[3] Heinze-Deml et al. 2017, Conditional Variance Penalties and Domain Shift Robustness\n[4] Atzmon et al. 2020, A causal view of compositional zero-shot recognition\n\n\n\n\n\n**EDIT: Post rebuttal**\n\nI thank the authors for their reply. Although the authors answered most of my questions, I decided to keep the score as is, because I share similar concerns with R2 about the presentation, and because experiments are still lacking. \n\nAdditionally, I am concerned with one of the author's replies saying *All methods achieve accuracy 1 ... on the training distribution*, because usually there is a trade-off between accuracy on the observational distribution versus the shifted distribution (discussed by Rothenhäusler, 2018 [Anchor regression]): Achieving perfect accuracy on the observational distribution, usually means relying on the spurious correlations. And under domain-shift scenarios, this would hinder the performance on the shifted-distribution.\n\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604594450854}, {"id": "n8An03Lahjr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1774/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The Causal Semantic Generative Model (CSG) presents an approach for learning both semantic and diverse latent causal variables in supervised settings using variational Bayes.  Contrary to many similar approaches which assume the label as the causal variable, this assumes a hidden causal variable which produces both the labels and observed features. Furthermore, this approach separates this latent causal variable into two components one for semantics, which impacts label generation and one for the diversity/variation which in combination with the semantic variable generates the observed features.\n\nThe authors present variations for learning this model which account for correlation/independence between semantic and diversity latent variables (CSG and CSG-ind) and also extend to settings where some data does contain labels (CSG-DA). The authors also show under which data generating assumptions the modeling holds and show how changes to the prior distribution of semantic and diversity variable distributions impact Out-of-Distribution Generalization error and Domain Adaptation Error.\n\nThe work presents empirical results demonstrating the effectiveness of this approach in both OOD settings (no test adaptation) and domain adaptation settings (test adaptation). For the presented experiments the authors show superior results in OOD generalization and competitive results in the domain adaptation setting. While the experiments present a compelling proof of concept, the tasks Shifted MNIST and ImageCLEF-DA are not the most representative challenges in their respective domains.  Would be interested to see performance in ColoredMNIST task for causal identification and OOD generalization as the generative structure is well understood as well as performance capabilities. The same could be said for the domain adaptation task, with Office-Home, VisDA-17, DomainNet or a variety of more challenging and representative tasks giving more empirical credibly to the experiments performed.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "By learning separate latent causal variables for both diversity and semantics, this work presents a tractable method to learn supervised models invariant to distributional shift.", "review": "The Causal Semantic Generative Model (CSG) presents an approach for learning both semantic and diverse latent causal variables in supervised settings using variational Bayes.  Contrary to many similar approaches which assume the label as the causal variable, this assumes a hidden causal variable which produces both the labels and observed features. Furthermore, this approach separates this latent causal variable into two components one for semantics, which impacts label generation and one for the diversity/variation which in combination with the semantic variable generates the observed features.\n\nThe authors present variations for learning this model which account for correlation/independence between semantic and diversity latent variables (CSG and CSG-ind) and also extend to settings where some data does contain labels (CSG-DA). The authors also show under which data generating assumptions the modeling holds and show how changes to the prior distribution of semantic and diversity variable distributions impact Out-of-Distribution Generalization error and Domain Adaptation Error.\n\nThe work presents empirical results demonstrating the effectiveness of this approach in both OOD settings (no test adaptation) and domain adaptation settings (test adaptation). For the presented experiments the authors show superior results in OOD generalization and competitive results in the domain adaptation setting. While the experiments present a compelling proof of concept, the tasks Shifted MNIST and ImageCLEF-DA are not the most representative challenges in their respective domains.  Would be interested to see performance in ColoredMNIST task for causal identification and OOD generalization as the generative structure is well understood as well as performance capabilities. The same could be said for the domain adaptation task, with Office-Home, VisDA-17, DomainNet or a variety of more challenging and representative tasks giving more empirical credibly to the experiments performed.  ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603918475855}, {"id": "DOZBr3QfkR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1774/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a Causal Semantic Generative model (CSG) to model semantic and variation factors separately and also provides a variational approach to learn the model.  It is proved that under some (perhaps strong) assumptions, CSG identifies the semantic factor on the training domain. Authors further show the boundedness of OOD generalization error based on the above result. Two  experiments are presented to validate the proposed method.\n\nWhile the paper seems to have many interesting ideas and theoretic results, it is poorly presented and not well prepared, making it hard to evaluate the technical correctness. I also notice that the paper changes the original latex formatting by reducing much vertical space for almost all section titles, theorems, and equations. Also, the top margin of page 8 is heavily reduced. There is approximately 2/3 - 1 page more content than that of the original format with required page limit. Given this consideration, I also lower my score. \n\nIn summary, I recommend a rejection for the present version. I highly suggest the authors revise the paper to make it more readable, self-contained, and concise, and resubmit the paper to another top conference/journal.\n\nPlease find my questions/suggestions below.\n\n1. The variables $s$, $v$, $x$, $y$ are introduced in the introduction part but then not mentioned when they are actually used in Section 3 to describe CSG. I suggest introducing the variables in Section 3.\n\n2. Authors use many italic words, and more crucially, lots of mixed uses of italic and normal fonts, particularly in Section 3. It really affects the reading flow. For introducing the CSG part, I recommend firstly introduce the model and then use another few paragraphs to present the examples for illustration.\n\n3. For the invariance principle 3.2, $p(s,v)$ is the only source of domain shift. Does it allow $p(v)$ and $p(s|v)$ (or $p(s)$ and $p(v|s)$) change while keeping $p(s,v)$ unchanged?\n\n3. (a) Sometimes you use 'Theorem' and also 'Thm.' in the main text, please be consistent. (b) The places for citation also affects the reading flow. For example, right after Assumption 5.2,  'It is a common (Janzing et al., 2009; Shalit et al., 2017; Khemakhem et al., 2019; Lee et al., 2019) sufficient condition for the fundamental (Peters et al., 2014, Prop. 7) requirement of causal minimality\n(Peters et al., 2014; 2017) for identifiability.' (c)  Some notations are not introduced, e.g.,  $q^{\\indep}(x)$ in CSG-ind. \n\n4. Why is it necessary to introduce CSG-ind? And in what case shall we consider CSG-ind? I did not find it well explained.\n\n5.  Also for Assumption 5.2, 'It is a common sufficient condition for the fundamental requirement of causal minimality for identifiability.' However, causal minimality is not equivalent to that $f$ being bijective. \n\n6. The paper should provide more details in comparing with IRM, and in fact I didn't get much information about the difference with IRM from Section 3.2.  Can the authors give more details about 'For noisy or degenerate mechanisms, ambiguity\noccurs during inference (Fig. 2), and the inferred result notably relies on the prior.'? Also, IRM is not compared in the experiments.\n\n7. In my opinion, the biggest contribution of this paper is to propose CSG, within which several theoretic results (identifiability, OOD error bound, etc.) are established under some necessary but maybe strong conditions. However, there is much less content on verifying that CSG does bring many benefits. One way is to conduct extensive experiments to verify so, but I find the experiments are not sufficient and also several methods (like IRM) are not included for comparison.\n\n** after reading rebuttal**\n\nThanks for clarifications and an improved version. I decide to increase my evaluation to 5.\n\nHowever, I think that the paper needs to take more content to illustrate the practical benefits of the proposed CSG frameworks, for the following reasons: \n- the framework is proposed based on empirical observations like 'intervening an image by e.g. breaking a camera sensor unit when taking the image, does not change how the photographer labels it', which is not mathematically rigorous; \n- the principles and assumptions are rather strong (though I understand one generally has to make assumptions in causality), and in practice it is not clear when such assumptions hold and how many applications satisfy these assumptions; \n- the interesting derivations and theorems are also based on the CSG framework, which means, if the framework is incorrect, then these results may fail; \n- the experiment settings are rather limited in the current version. I hope the authors to add further content in their next version, regardless of whether the paper gets accepted or rejected.\n\nLastly, I still feel it a bit tricky to change the original formatting in the previous submission.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting causality based model and several theoretic results. Insufficient experiments and poor presentation", "review": "This paper proposes a Causal Semantic Generative model (CSG) to model semantic and variation factors separately and also provides a variational approach to learn the model.  It is proved that under some (perhaps strong) assumptions, CSG identifies the semantic factor on the training domain. Authors further show the boundedness of OOD generalization error based on the above result. Two  experiments are presented to validate the proposed method.\n\nWhile the paper seems to have many interesting ideas and theoretic results, it is poorly presented and not well prepared, making it hard to evaluate the technical correctness. I also notice that the paper changes the original latex formatting by reducing much vertical space for almost all section titles, theorems, and equations. Also, the top margin of page 8 is heavily reduced. There is approximately 2/3 - 1 page more content than that of the original format with required page limit. Given this consideration, I also lower my score. \n\nIn summary, I recommend a rejection for the present version. I highly suggest the authors revise the paper to make it more readable, self-contained, and concise, and resubmit the paper to another top conference/journal.\n\nPlease find my questions/suggestions below.\n\n1. The variables $s$, $v$, $x$, $y$ are introduced in the introduction part but then not mentioned when they are actually used in Section 3 to describe CSG. I suggest introducing the variables in Section 3.\n\n2. Authors use many italic words, and more crucially, lots of mixed uses of italic and normal fonts, particularly in Section 3. It really affects the reading flow. For introducing the CSG part, I recommend firstly introduce the model and then use another few paragraphs to present the examples for illustration.\n\n3. For the invariance principle 3.2, $p(s,v)$ is the only source of domain shift. Does it allow $p(v)$ and $p(s|v)$ (or $p(s)$ and $p(v|s)$) change while keeping $p(s,v)$ unchanged?\n\n3. (a) Sometimes you use 'Theorem' and also 'Thm.' in the main text, please be consistent. (b) The places for citation also affects the reading flow. For example, right after Assumption 5.2,  'It is a common (Janzing et al., 2009; Shalit et al., 2017; Khemakhem et al., 2019; Lee et al., 2019) sufficient condition for the fundamental (Peters et al., 2014, Prop. 7) requirement of causal minimality\n(Peters et al., 2014; 2017) for identifiability.' (c)  Some notations are not introduced, e.g.,  $q^{\\indep}(x)$ in CSG-ind. \n\n4. Why is it necessary to introduce CSG-ind? And in what case shall we consider CSG-ind? I did not find it well explained.\n\n5.  Also for Assumption 5.2, 'It is a common sufficient condition for the fundamental requirement of causal minimality for identifiability.' However, causal minimality is not equivalent to that $f$ being bijective. \n\n6. The paper should provide more details in comparing with IRM, and in fact I didn't get much information about the difference with IRM from Section 3.2.  Can the authors give more details about 'For noisy or degenerate mechanisms, ambiguity\noccurs during inference (Fig. 2), and the inferred result notably relies on the prior.'? Also, IRM is not compared in the experiments.\n\n7. In my opinion, the biggest contribution of this paper is to propose CSG, within which several theoretic results (identifiability, OOD error bound, etc.) are established under some necessary but maybe strong conditions. However, there is much less content on verifying that CSG does bring many benefits. One way is to conduct extensive experiments to verify so, but I find the experiments are not sufficient and also several methods (like IRM) are not included for comparison.\n\n** after reading rebuttal**\n\nThanks for clarifications and an improved version. I decide to increase my evaluation to 5.\n\nHowever, I think that the paper needs to take more content to illustrate the practical benefits of the proposed CSG frameworks, for the following reasons: \n- the framework is proposed based on empirical observations like 'intervening an image by e.g. breaking a camera sensor unit when taking the image, does not change how the photographer labels it', which is not mathematically rigorous; \n- the principles and assumptions are rather strong (though I understand one generally has to make assumptions in causality), and in practice it is not clear when such assumptions hold and how many applications satisfy these assumptions; \n- the interesting derivations and theorems are also based on the CSG framework, which means, if the framework is incorrect, then these results may fail; \n- the experiment settings are rather limited in the current version. I hope the authors to add further content in their next version, regardless of whether the paper gets accepted or rejected.\n\nLastly, I still feel it a bit tricky to change the original formatting in the previous submission.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603868815824}], "openreview_url": "https://openreview.net/forum?id=xyGFYKIPTDJ", "arxiv_id": "2011.01681", "paper_pdf": "papers/xyGFYKIPTDJ.pdf", "paper_pdf_sha256": "e48d9262f64427ff7c7a31b4dd9aae93b6deae69abbe3b1bde980595343245b3", "paper_pdf_bytes": 508925, "paper_pdf_source": "openreview", "code_url": "https://github.com/changliu00/causal-semantic-generative-model", "code_repository": "changliu00/causal-semantic-generative-model", "code_commit": "05c10d6790f3db8d4847efd98d18a6ecafb469cb", "code_archive": "repos/xyGFYKIPTDJ.zip", "code_archive_sha256": "580cde18a3378f13598d3820d20aebb7b52f2edea53350f48280b43f75f21874", "code_archive_bytes": 205268, "code_file_count": 39, "code_extensions": {".py": 28, ".sh": 10, ".ipynb": 1}, "github_disk_usage_kb": 198, "github_languages": {"Python": 212959, "Shell": 20528, "Jupyter Notebook": 10218}, "github_archived": false, "github_pushed_at": "2022-04-18T15:12:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-causal-semantic-representation-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1em9h4KDS", "year": 2020, "status": "rejected", "title": "Generative Imputation and Stochastic Prediction", "authors": ["Mohammad Kachuee", "Kimmo Kärkkäinen", "Orpaz Goldstein", "Sajad Darabi", "Majid Sarrafzadeh"], "authorids": ["mkachuee@ucla.edu", "kimmo@cs.ucla.edu", "orpgol@cs.ucla.edu", "sajad.darabi@cs.ucla.edu", "majid@cs.ucla.edu"], "authors_source": "OpenReview API", "abstract": "In many machine learning applications, we are faced with incomplete datasets. In the literature, missing data imputation techniques have been mostly concerned with filling missing values. However, the existence of missing values is synonymous with uncertainties not only over the distribution of missing values but also over target class assignments that require careful consideration. In this paper, we propose a simple and effective method for imputing missing features and estimating the distribution of target assignments given incomplete data. In order to make imputations, we train a simple and effective generator network to generate imputations that a discriminator network is tasked to distinguish. Following this, a predictor network is trained using the imputed samples from the generator network to capture the classification uncertainties and make predictions accordingly. The proposed method is evaluated on CIFAR-10 image dataset as well as three real-world tabular classification datasets, under different missingness rates and structures. Our experimental results show the effectiveness of the proposed method in generating imputations as well as providing estimates for the class uncertainties in a classification task when faced with missing values.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hyxvxaag5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper110/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method to impute missing features using a generative model and train a predictive model on top of imputed dataset to improve classification results. They first train a GAN model where the generator outputs an imputed representation of the input and discriminator is trained to predict if an individual features (such as a pixel) is imputed or not. Given the generator and incomplete sample, they train a predictor using the output of the generator, imputed sample, as input. Their main contribution is using a MC averaging to compute the prediction by repetitively sampling from the noise variable, z, and generating different imputations from generator. They show that the proposed model improves upon the previous SOTA on final classification performance.\n\nOverall the paper is clearly written. But I do feel it is a bit incremental over the GAIN approach. The overall GAN architecture is very similar to GAIN's and although stochastic prediction shows clear improvements it is a bit straightforward. However, I think the uncertainty of the imputations and its effect on the final prediction is interesting. I suggest the authors to extend this part with more detailed analysis.\n\nThere are several parts that are confusing/missing in the paper:\n\n- In GAIN, they use a hint vector as an input to the discriminator. They show that without the hint vector, there is no unique solution (this is shown without the MSE loss). The authors do not use this vector in their approach (as in Figure 1) and it is not clear to me if it causes any instabilities or if multiple experiments yield similar results or if the stochastic prediction benefits from this.\n- On what type of examples GI is more accurate than other models? Since stochastic prediction is the main difference from GAIN, is this related to the multi-modality of the noisy examples?\n- Can you explain the difference between the results in Figure-7 and Table-2? Results between the two mismatch.\n- I think the statement in the first paragraph in Section 4.4 that \"MSE loss term would act as a denoising loss smoothing noisy missing pixels\" could be misleading. MSE is used with mask in GAIN, hence it only applies to the observed features during training. Its effect on smoothing noisy missing pixels is not clear.\n\n\nI think the paper would benefit if the authors could explain/show:\n- Increasing the missing rate would also increase the possibility that the ground truth be a more multi-modal distribution. Especially in rectangular generation part where it can remove a complete object. Does stochastic averaging benefit more in this case?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper proposes a method to impute missing features using a generative model and train a predictive model on top of imputed dataset to improve classification results. They first train a GAN model where the generator outputs an imputed representation of the input and discriminator is trained to predict if an individual features (such as a pixel) is imputed or not. Given the generator and incomplete sample, they train a predictor using the output of the generator, imputed sample, as input. Their main contribution is using a MC averaging to compute the prediction by repetitively sampling from the noise variable, z, and generating different imputations from generator. They show that the proposed model improves upon the previous SOTA on final classification performance.\n\nOverall the paper is clearly written. But I do feel it is a bit incremental over the GAIN approach. The overall GAN architecture is very similar to GAIN's and although stochastic prediction shows clear improvements it is a bit straightforward. However, I think the uncertainty of the imputations and its effect on the final prediction is interesting. I suggest the authors to extend this part with more detailed analysis.\n\nThere are several parts that are confusing/missing in the paper:\n\n- In GAIN, they use a hint vector as an input to the discriminator. They show that without the hint vector, there is no unique solution (this is shown without the MSE loss). The authors do not use this vector in their approach (as in Figure 1) and it is not clear to me if it causes any instabilities or if multiple experiments yield similar results or if the stochastic prediction benefits from this.\n- On what type of examples GI is more accurate than other models? Since stochastic prediction is the main difference from GAIN, is this related to the multi-modality of the noisy examples?\n- Can you explain the difference between the results in Figure-7 and Table-2? Results between the two mismatch.\n- I think the statement in the first paragraph in Section 4.4 that \"MSE loss term would act as a denoising loss smoothing noisy missing pixels\" could be misleading. MSE is used with mask in GAIN, hence it only applies to the observed features during training. Its effect on smoothing noisy missing pixels is not clear.\n\n\nI think the paper would benefit if the authors could explain/show:\n- Increasing the missing rate would also increase the possibility that the ground truth be a more multi-modal distribution. Especially in rectangular generation part where it can remove a complete object. Does stochastic averaging benefit more in this case?"}, "tcdate": 1572031727445}, {"id": "B1xLXZ319r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper110/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This is a nice piece of incremental work on top of previously published GAN imputation methods. It seems to work well in the limited evaluation and is at least claimed to be easier to use for practitioners. This paper could benefit tremendously from both better evaluation and discussion. The paper would be much clearer if GI contextualized itself relative to GAIN on the one hand (which is the most similar GAN method) and multiple imputation on the other hand (of which this is almost, but not quite, an instance of). \n\n\nSuggestions for improving the introduction & discussion: \n* The purpose of this paper is to model uncertainties about missing values — you really should say more about probabilistic methods than \"A few exceptions exist such as Bayesian models”.  At least give some motivation for why certain imputations problems couldn’t be feasibly solved by modeling the missing values in a probabilistic programming framework. \n* Other GAN methods for imputation (GAIN and MisGAN) are dismissed as \"often very complicated to be applied in practical setups by practitioners”. Given that the described method resembles GAIN, is it really much simpler? If so, can you be more specific when characterizing related work?\n* \"This is different from approaches such as multiple imputation where several predictors are trained on different imputed versions of a dataset.” — the main difference between this approach and MI is that you’re interleaving imputation and training a downstream model. Emphasize this earlier on, since it will make the whole technique easier to understand.\n\nSuggestions for improving the evaluation:\n* You’re imputing missing rectangles from an image dataset — please show us the resulting images. I would greatly prefer this to Section 4.5 — which is a very low sample size, low dimensionality example and it’s really unclear how well it generalizes to real data. \n* \"We also considered using root means squared error (RMSE); however, we decided not to use this measure as we observed an inconsistent behavior using RMSE in our comparisons as RMSE favors methods that show less variance rather than realistic and sharp samples from the distribution.” — I think this is a mistake, likely motivated by the proposed method doing worse under the RMSE metric. Show us several relevant metrics and then discuss their tradeoffs afterward. \n* \"We run each experiment multiple times (at least 4)” — please report how often each experiment was run, even better if you standardize this number. \n* For Table 2, please provide accuracy without missing values as a baseline.\n* Add MICE or some other “standard” imputation method as a baseline. \n\nSuggestions for improving readability: \n* Many sentences start with “in this” (e.g. “in this case”, “in this setting”, &c). Sometimes these sentences even co-occur within the same paragraph. Try to switch up the phrasing and move away from repetition. \n* Not a complete sentence: \"For instance, jointly training multiple generator/discriminator networks, tuning objective functions with multiple hyper-parameters, etc.\"\n\nUpdate: I think the latest draft of the paper is a big improvement, the inclusion of a \"classical\" baseline, improved language and additional appendices are all welcome. I'm leaving the rating as a \"weak accept\" since the paper still feels rough and could use additional editing/streamlining. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "This is a nice piece of incremental work on top of previously published GAN imputation methods. It seems to work well in the limited evaluation and is at least claimed to be easier to use for practitioners. This paper could benefit tremendously from both better evaluation and discussion. The paper would be much clearer if GI contextualized itself relative to GAIN on the one hand (which is the most similar GAN method) and multiple imputation on the other hand (of which this is almost, but not quite, an instance of). \n\n\nSuggestions for improving the introduction & discussion: \n* The purpose of this paper is to model uncertainties about missing values — you really should say more about probabilistic methods than \"A few exceptions exist such as Bayesian models”.  At least give some motivation for why certain imputations problems couldn’t be feasibly solved by modeling the missing values in a probabilistic programming framework. \n* Other GAN methods for imputation (GAIN and MisGAN) are dismissed as \"often very complicated to be applied in practical setups by practitioners”. Given that the described method resembles GAIN, is it really much simpler? If so, can you be more specific when characterizing related work?\n* \"This is different from approaches such as multiple imputation where several predictors are trained on different imputed versions of a dataset.” — the main difference between this approach and MI is that you’re interleaving imputation and training a downstream model. Emphasize this earlier on, since it will make the whole technique easier to understand.\n\nSuggestions for improving the evaluation:\n* You’re imputing missing rectangles from an image dataset — please show us the resulting images. I would greatly prefer this to Section 4.5 — which is a very low sample size, low dimensionality example and it’s really unclear how well it generalizes to real data. \n* \"We also considered using root means squared error (RMSE); however, we decided not to use this measure as we observed an inconsistent behavior using RMSE in our comparisons as RMSE favors methods that show less variance rather than realistic and sharp samples from the distribution.” — I think this is a mistake, likely motivated by the proposed method doing worse under the RMSE metric. Show us several relevant metrics and then discuss their tradeoffs afterward. \n* \"We run each experiment multiple times (at least 4)” — please report how often each experiment was run, even better if you standardize this number. \n* For Table 2, please provide accuracy without missing values as a baseline.\n* Add MICE or some other “standard” imputation method as a baseline. \n\nSuggestions for improving readability: \n* Many sentences start with “in this” (e.g. “in this case”, “in this setting”, &c). Sometimes these sentences even co-occur within the same paragraph. Try to switch up the phrasing and move away from repetition. \n* Not a complete sentence: \"For instance, jointly training multiple generator/discriminator networks, tuning objective functions with multiple hyper-parameters, etc.\"\n\nUpdate: I think the latest draft of the paper is a big improvement, the inclusion of a \"classical\" baseline, improved language and additional appendices are all welcome. I'm leaving the rating as a \"weak accept\" since the paper still feels rough and could use additional editing/streamlining. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571959070132}, {"id": "HyeAuZ3DtS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper110/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The uncertainty of having a missing value is investigated on the prediction by not assigning a single imputed value but N different values generated via an imputer network (based on GAIN). Unlike old-fashion multiple imputation techniques, one predictor is trained on different samples and it induces the uncertainty. The experiments on number of datasets show the proposed predictor is capable of having a fairly better performance.\nOverall, this paper raises an interesting point about missing data imputation via generative models, and well-written; however, there are number of concerns:\n1-\tThe predictor is trained on different version of imputed samples (imputed via the generator); this equates to making noisy version of the real samples, where noise is applied to the missing variables. A side effect of this is generalization of the predictor; thus, have you been careful that the improved accuracy is not due to generalization? In other words, if we imposed the generalization via adding gaussian noise to the imputed samples by GAIN for example, would we get improved accuracy too?\n2-\tYour method known as GI is a modified version of GAIN. You could also use MisGAN, and I am wondering if the results would have been different if generator in MisGAN was used in GI in Figure 2 (b), as MisGAN works better than GAIN.\n3-\tI am also wondering how the GAIN imputation changes by removing the MSE term? GI discards the MSE term in GAIN, and it changes the distribution of the imputed variables by GAIN. Could you maybe fit a Normal distribution on a chosen imputed variable  (N=128) and visualize how different it is from the distribution of imputed variables with GAIN with MSE term. \n4-\tIn justification for claim 1, it is said “This is equivalent to training models using noisy labels”. This is not accurate: in noisy label prediction, we have one (noisy) y corresponding to each x, in your case there are multiple ys for one sample.\n5-\tIn the implementation details, I cannot fully wrap my head around the part “z vector of size 1/8”; how did you choose this 1/8?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The uncertainty of having a missing value is investigated on the prediction by not assigning a single imputed value but N different values generated via an imputer network (based on GAIN). Unlike old-fashion multiple imputation techniques, one predictor is trained on different samples and it induces the uncertainty. The experiments on number of datasets show the proposed predictor is capable of having a fairly better performance.\nOverall, this paper raises an interesting point about missing data imputation via generative models, and well-written; however, there are number of concerns:\n1-\tThe predictor is trained on different version of imputed samples (imputed via the generator); this equates to making noisy version of the real samples, where noise is applied to the missing variables. A side effect of this is generalization of the predictor; thus, have you been careful that the improved accuracy is not due to generalization? In other words, if we imposed the generalization via adding gaussian noise to the imputed samples by GAIN for example, would we get improved accuracy too?\n2-\tYour method known as GI is a modified version of GAIN. You could also use MisGAN, and I am wondering if the results would have been different if generator in MisGAN was used in GI in Figure 2 (b), as MisGAN works better than GAIN.\n3-\tI am also wondering how the GAIN imputation changes by removing the MSE term? GI discards the MSE term in GAIN, and it changes the distribution of the imputed variables by GAIN. Could you maybe fit a Normal distribution on a chosen imputed variable  (N=128) and visualize how different it is from the distribution of imputed variables with GAIN with MSE term. \n4-\tIn justification for claim 1, it is said “This is equivalent to training models using noisy labels”. This is not accurate: in noisy label prediction, we have one (noisy) y corresponding to each x, in your case there are multiple ys for one sample.\n5-\tIn the implementation details, I cannot fully wrap my head around the part “z vector of size 1/8”; how did you choose this 1/8?\n"}, "tcdate": 1571434870386}], "openreview_url": "https://openreview.net/forum?id=B1em9h4KDS", "arxiv_id": "1905.09340", "paper_pdf": "papers/B1em9h4KDS.pdf", "paper_pdf_sha256": "1f7c067789a0a37109bcc21d360087f10ddf331f214527b045ac3fc44fe7e17a", "paper_pdf_bytes": 4615662, "paper_pdf_source": "openreview", "code_url": "https://github.com/mkachuee/GenerativeImputationStochasticPrediction", "code_repository": "mkachuee/GenerativeImputationStochasticPrediction", "code_commit": "f62e8d30e8d2d42088255da0d754ed6b7d75dfdf", "code_archive": "repos/B1em9h4KDS.zip", "code_archive_sha256": "004803c161cf288f4a94784ee3d75c811a889db57a700919d761684e2ef7dcaa", "code_archive_bytes": 227195, "code_file_count": 31, "code_extensions": {".py": 30, ".sh": 1}, "github_disk_usage_kb": 200, "github_languages": {"Python": 213002, "Shell": 741}, "github_archived": false, "github_pushed_at": "2020-09-04T04:26:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generative-imputation-and-stochastic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJl-HsR9KX", "year": 2019, "status": "rejected", "title": "Discriminative Active Learning", "authors": ["Daniel Gissin", "Shai Shalev-Shwartz"], "authorids": ["daniel.gissin@mail.huji.ac.il", "shais@cs.huji.ac.il"], "authors_source": "OpenReview API", "abstract": "We propose a new batch mode active learning algorithm designed for neural networks and large query batch sizes. The method, Discriminative Active Learning (DAL), poses active learning as a binary classification task, attempting to choose examples to label in such a way as to make the labeled set and the unlabeled pool indistinguishable. Experimenting on image classification tasks, we empirically show our method to be on par with state of the art methods in medium and large query batch sizes, while being simple to implement and also extend to other domains besides classification tasks. Our experiments also show that none of the state of the art methods of today are clearly better than uncertainty sampling, negating some of the reported results in the recent literature.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1lASPI9nm", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper65/AnonReviewer1"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Thank you for this enjoyable paper. \n\nSummary: The authors propose a novel approach to active learning as follows. At each iteration they develop a classifier that can discriminate between the samples in the labeled and unlabeled sets; they select the top few samples that are most likely to be from the unlabeled set as per this classifier, and request the oracle to provide labels for this batch next. This simple idea is shown to have a principled basis and theoretical background, related to GANs and to previous results from the literature. They provide clear algorithms and open source code for easy verification, and public testing. They provide good experimental verification on CIFAR-10 and MNIST benchmarks. I personally look at new papers more for novel ideas and good intuition/theoretical justification than an immediate improvement in benchmark results, so I enjoyed this paper thoroughly. \n\nResults: Among other things they show that their algorithms ranks the samples to be next labeled quite differently than uncertainty sampling based approaches; that their method is at least as accurate/sample-efficient as the state of the art ; and that some previously published experimental results are incorrect(!). As the authors will probably agree I am not convinced the proposed method is better than previous algorithms in any statistically significant way, but the novel idea itself is worth publishing even if it is just as good as the state of the art. \n\nNovelty: I liked the paper very much because it provides quite an innovative new approach to look at active learning, which resembles GANs and Core set ideas in some ways, yet differs in significant ways that are critical for active learning. I've been working and publishing in related areas for a long time so I genuinely found your central idea refreshing and new.\n\nRelevance: The paper is very relevant to the ICLR community and addresses critical questions. \n\nQuestion:\nMy intuition as a Bayesian is that we most need to find labels that maximize the mutual information I(y,w) where w are the weights of the neural net. In practice this corresponds to the samples x which have the maximum class uncertainty, but for which the parameters under the posterior disagree about the outcome the most, eg see discussion below equation 2 for  Bayesian Active Learning by Disagreement (BALD) in this paper https://arxiv.org/pdf/1112.5745.pdf . In essence: The above means that the labels that provide most information about the classification model are most valuable for active learning. \n\nHowever, your approach intuitively ignores the conditional distribution(ie py(|x)), and instead tries to make the original unconditional distribution p(x) between the labeled and unlabeled sets similar. Yet, it works beautifully. So: Why does this work? What is the intuition?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "very nice paper with neat idea, good theoretical intuition/justification, clear transparent code/algorithm, and good experimental results", "review": "Thank you for this enjoyable paper. \n\nSummary: The authors propose a novel approach to active learning as follows. At each iteration they develop a classifier that can discriminate between the samples in the labeled and unlabeled sets; they select the top few samples that are most likely to be from the unlabeled set as per this classifier, and request the oracle to provide labels for this batch next. This simple idea is shown to have a principled basis and theoretical background, related to GANs and to previous results from the literature. They provide clear algorithms and open source code for easy verification, and public testing. They provide good experimental verification on CIFAR-10 and MNIST benchmarks. I personally look at new papers more for novel ideas and good intuition/theoretical justification than an immediate improvement in benchmark results, so I enjoyed this paper thoroughly. \n\nResults: Among other things they show that their algorithms ranks the samples to be next labeled quite differently than uncertainty sampling based approaches; that their method is at least as accurate/sample-efficient as the state of the art ; and that some previously published experimental results are incorrect(!). As the authors will probably agree I am not convinced the proposed method is better than previous algorithms in any statistically significant way, but the novel idea itself is worth publishing even if it is just as good as the state of the art. \n\nNovelty: I liked the paper very much because it provides quite an innovative new approach to look at active learning, which resembles GANs and Core set ideas in some ways, yet differs in significant ways that are critical for active learning. I've been working and publishing in related areas for a long time so I genuinely found your central idea refreshing and new.\n\nRelevance: The paper is very relevant to the ICLR community and addresses critical questions. \n\nQuestion:\nMy intuition as a Bayesian is that we most need to find labels that maximize the mutual information I(y,w) where w are the weights of the neural net. In practice this corresponds to the samples x which have the maximum class uncertainty, but for which the parameters under the posterior disagree about the outcome the most, eg see discussion below equation 2 for  Bayesian Active Learning by Disagreement (BALD) in this paper https://arxiv.org/pdf/1112.5745.pdf . In essence: The above means that the labels that provide most information about the classification model are most valuable for active learning. \n\nHowever, your approach intuitively ignores the conditional distribution(ie py(|x)), and instead tries to make the original unconditional distribution p(x) between the labeled and unlabeled sets similar. Yet, it works beautifully. So: Why does this work? What is the intuition?\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541199686340}, {"id": "SJgxZFHc2m", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper65/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is proposing a distribution matching as a metric for active learning. Basic intuition is: if we can make the distribution of labelled and unlabelled examples similar to each other, training error in one will approximate the training error in the other. Hence, a model learned using labelled ones will do well in unlabelled ones. The main tool to enforce this distributional distance is using adversarial learning similar to GANs or gradient reversal network for domain adaptation.\n\nThe idea is definitely interesting. I am not sure about why should it work (I explained in detail later), but it does work well empirically. Moreover, it is very easy to implement. Given any learned or hard-coded features, learning a simple binary classifier is sufficient to implement the method. The mini-queries idea in 4.1 is especially interesting. Handling large batches in active learning is always a problem but this neat trick make it much easier.\n\nI think the proposed method is counter intuitive as the discussion does not explain why should it work better than random sampling. Clearly if labelled samples are randomly sampled, labelled and unlabelled data is coming from the exactly same distribution. Hence, the distance (H-divergence, TV-distance etc.) between them is 0. My main question to authors is why does this method work better than random sampling? A similar question is; since they are coming from the exact same distribution, what is the meaning of minimizing empirical H-divergence? I think a more detailed study on a toy problem could potentially explain this. Authors can generate 1-D or 2D samples from a well defined distribution (eg. Gaussians with different means/variances for each class) and visualize what is the algorithm actually doing. \n\nConsidering my point that these data points are actually coming from the same distribution, discussion in Section 3.1 is rather unjustified. Most of the entities discussed in that section are probabilistic entities (generally speaking expected values) and does not differ between labelled and unlabelled case since they have same underlying distribution. Their empirical values are different but this is beyond the study of Ben-David(2010). Therefore, I am not sure does the Section 3.1 is contributing to the paper without any explicit connection to the empirical divergence minimization. More importantly a much similar work from domain adaptation is [Unsupervised domain adaptation by backpropagation, ICML 2015] and it should also be discussed in the paper.\n\nSome minor issues:\n- Are the hyper-parameters kept fixed for all experiments. In other words, does the training size of 5k and 15k share hyperparameters? Which might be sub-optimal.\n- The experiments use very large batch sizes. A smaller batch sizes might separate the algorithms better.\n- References in the text have some issues. There are missing commas between references in the text. There are also some cases where \\citep should have been used but \\citet is used. A careful pass over them might be beneficial.\n\nIn summary, I think the paper is interesting, easy to implement and possibly useful to the large part of the community since active learning is very important problem. I think the major weakness of the paper is the fact that authors did not give a clear explain why does it actually work. I think it is crucial for authors to provide a theoretical or an empirical study which answer this question.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting but counter-intuitive idea which works well in practice. It needs a better motivation/explanation.", "review": "The paper is proposing a distribution matching as a metric for active learning. Basic intuition is: if we can make the distribution of labelled and unlabelled examples similar to each other, training error in one will approximate the training error in the other. Hence, a model learned using labelled ones will do well in unlabelled ones. The main tool to enforce this distributional distance is using adversarial learning similar to GANs or gradient reversal network for domain adaptation.\n\nThe idea is definitely interesting. I am not sure about why should it work (I explained in detail later), but it does work well empirically. Moreover, it is very easy to implement. Given any learned or hard-coded features, learning a simple binary classifier is sufficient to implement the method. The mini-queries idea in 4.1 is especially interesting. Handling large batches in active learning is always a problem but this neat trick make it much easier.\n\nI think the proposed method is counter intuitive as the discussion does not explain why should it work better than random sampling. Clearly if labelled samples are randomly sampled, labelled and unlabelled data is coming from the exactly same distribution. Hence, the distance (H-divergence, TV-distance etc.) between them is 0. My main question to authors is why does this method work better than random sampling? A similar question is; since they are coming from the exact same distribution, what is the meaning of minimizing empirical H-divergence? I think a more detailed study on a toy problem could potentially explain this. Authors can generate 1-D or 2D samples from a well defined distribution (eg. Gaussians with different means/variances for each class) and visualize what is the algorithm actually doing. \n\nConsidering my point that these data points are actually coming from the same distribution, discussion in Section 3.1 is rather unjustified. Most of the entities discussed in that section are probabilistic entities (generally speaking expected values) and does not differ between labelled and unlabelled case since they have same underlying distribution. Their empirical values are different but this is beyond the study of Ben-David(2010). Therefore, I am not sure does the Section 3.1 is contributing to the paper without any explicit connection to the empirical divergence minimization. More importantly a much similar work from domain adaptation is [Unsupervised domain adaptation by backpropagation, ICML 2015] and it should also be discussed in the paper.\n\nSome minor issues:\n- Are the hyper-parameters kept fixed for all experiments. In other words, does the training size of 5k and 15k share hyperparameters? Which might be sub-optimal.\n- The experiments use very large batch sizes. A smaller batch sizes might separate the algorithms better.\n- References in the text have some issues. There are missing commas between references in the text. There are also some cases where \\citep should have been used but \\citet is used. A careful pass over them might be beneficial.\n\nIn summary, I think the paper is interesting, easy to implement and possibly useful to the large part of the community since active learning is very important problem. I think the major weakness of the paper is the fact that authors did not give a clear explain why does it actually work. I think it is crucial for authors to provide a theoretical or an empirical study which answer this question.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541196024247}, {"id": "BJx-6PRtnQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper65/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new approach to an active learning problem where the idea is to train a classifier to distinguish labeled and unlabeled datapoints and select those that look the most like unlabeled.\n\nThe paper is clearly written and easy to follow. The idea is quite novel and evokes interesting thoughts. I appreciated that the authors provide links and connections to other problems. Another positive aspect is that evaluation methodology is quite sound and includes comparison to many recent algorithms for AL with neural networks. The analysis of Section 5.5 is quite interesting.\nHowever, I have a few concerns regarding the methodology. First of all, I am not completely convinced by the fact that selecting the samples that resemble the most unlabeled data is beneficial for the classifier. It seem that in this case just the data from under-explored regions will be selected at every new iteration. If this is the purpose, some simpler methods, for example, relying on density sampling, can be used. Could you elaborate how you method would compare to them? I can see this method as a way to measure the representativeness of datapoints, but I would see it as a component of AL, not an AL alone. What would happen it is combined with Uncertainty and you use it to labeled the points that are both uncertain and resemble unlabeled data? \nBesides, the proposed approach does not take the advantage of all the information that is available to AL, in particular, it does not use at the information about labels. I believe that labels contain a lot of useful information for making an informed selection decision and ignoring it when it is available is not rational.  \nNext, I have conceptual difficulties understanding what would happen to a classifier at next iteration when it is trained on the data that was determined by the previous classifier. Seems that the training data is non-iid and might cause some strange bias. In addition to this, it sounds a bit strange to use classification where overfitting is acceptable.\nFinally, the results of the experimental evaluation do not demonstrate a significant advantage of the proposed method and thus it is unclear is there is a benefit of using this method in practice. \n\nQuestions:\n- Could you elaborate why DAL strategy does not end up doing just random sampling?\n- Nothing restrict DAL from being applied with classifiers other than neural networks and smaller problems. How do you think DAL would work on simpler datasets and classifiers?\n- How does the classifier (that distinguished between labeled and unlabeled data) deal with very unbalanced classes? I suppose that normally unlabeled set is much bigger than labeled. What does 98% accuracy mean in this case?\n- How many experiments were run to produce each figure? Are error bars of most experiments so small that are almost invisible?\n\nSmall comments:\n- I think in many cases citep command should be used instead of cite. \n- Can you explain more about the paragraph 3 of related work where you say that uncertainty-based approach would be different from margin-based approach if the classifier is neural network?\n- Last sentence before 3.1: how do you guarantee in this case that the selected examples are not similar to each other (that was mentioned as a limitation for batch uncertainty selection, last paragraph on page 1)?\n- It was hard to understand the beginning of 5.5, at first it sounds like the ranking of methods is going to be analysed.\n- I am not sure \"discriminative\" is a good name for this algorithm. It suggested that is it opposite to \"generative\" (query synthesis?), but then all AL that rank datapoints with some scoring function are \"discriminative\".", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Method motivation and experimental results are not very convincing", "review": "This paper presents a new approach to an active learning problem where the idea is to train a classifier to distinguish labeled and unlabeled datapoints and select those that look the most like unlabeled.\n\nThe paper is clearly written and easy to follow. The idea is quite novel and evokes interesting thoughts. I appreciated that the authors provide links and connections to other problems. Another positive aspect is that evaluation methodology is quite sound and includes comparison to many recent algorithms for AL with neural networks. The analysis of Section 5.5 is quite interesting.\nHowever, I have a few concerns regarding the methodology. First of all, I am not completely convinced by the fact that selecting the samples that resemble the most unlabeled data is beneficial for the classifier. It seem that in this case just the data from under-explored regions will be selected at every new iteration. If this is the purpose, some simpler methods, for example, relying on density sampling, can be used. Could you elaborate how you method would compare to them? I can see this method as a way to measure the representativeness of datapoints, but I would see it as a component of AL, not an AL alone. What would happen it is combined with Uncertainty and you use it to labeled the points that are both uncertain and resemble unlabeled data? \nBesides, the proposed approach does not take the advantage of all the information that is available to AL, in particular, it does not use at the information about labels. I believe that labels contain a lot of useful information for making an informed selection decision and ignoring it when it is available is not rational.  \nNext, I have conceptual difficulties understanding what would happen to a classifier at next iteration when it is trained on the data that was determined by the previous classifier. Seems that the training data is non-iid and might cause some strange bias. In addition to this, it sounds a bit strange to use classification where overfitting is acceptable.\nFinally, the results of the experimental evaluation do not demonstrate a significant advantage of the proposed method and thus it is unclear is there is a benefit of using this method in practice. \n\nQuestions:\n- Could you elaborate why DAL strategy does not end up doing just random sampling?\n- Nothing restrict DAL from being applied with classifiers other than neural networks and smaller problems. How do you think DAL would work on simpler datasets and classifiers?\n- How does the classifier (that distinguished between labeled and unlabeled data) deal with very unbalanced classes? I suppose that normally unlabeled set is much bigger than labeled. What does 98% accuracy mean in this case?\n- How many experiments were run to produce each figure? Are error bars of most experiments so small that are almost invisible?\n\nSmall comments:\n- I think in many cases citep command should be used instead of cite. \n- Can you explain more about the paragraph 3 of related work where you say that uncertainty-based approach would be different from margin-based approach if the classifier is neural network?\n- Last sentence before 3.1: how do you guarantee in this case that the selected examples are not similar to each other (that was mentioned as a limitation for batch uncertainty selection, last paragraph on page 1)?\n- It was hard to understand the beginning of 5.5, at first it sounds like the ranking of methods is going to be analysed.\n- I am not sure \"discriminative\" is a good name for this algorithm. It suggested that is it opposite to \"generative\" (query synthesis?), but then all AL that rank datapoints with some scoring function are \"discriminative\".", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541167033089}], "openreview_url": "https://openreview.net/forum?id=rJl-HsR9KX", "arxiv_id": "1907.06347", "paper_pdf": "papers/rJl-HsR9KX.pdf", "paper_pdf_sha256": "dfd6410a588accfe02faa0673439712494bf64f4a4d5f3f2ad534ea5afcc1a1b", "paper_pdf_bytes": 1414381, "paper_pdf_source": "openreview", "code_url": "https://github.com/dsgissin/DiscriminativeActiveLearning", "code_repository": "dsgissin/DiscriminativeActiveLearning", "code_commit": "625e9a183b9c14356c81b9073456d8142f25f906", "code_archive": "repos/rJl-HsR9KX.zip", "code_archive_sha256": "60bab3afbed30a4934bc3e5c597cf2a90a40cde369403ecfd171ba959d3a5294", "code_archive_bytes": 15207, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 6859, "github_languages": {"Python": 67802}, "github_archived": false, "github_pushed_at": "2019-09-26T11:48:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discriminative-active-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uhVlvT3Pk1", "year": 2026, "status": "rejected", "title": "Probing Human Visual Robustness with Neurally-Guided Deep Neural Networks", "authors": ["Zhenan Shao", "Linjian Ma", "Yiqing Zhou", "Yibo Jacky Zhang", "Sanmi Koyejo", "Bo Li", "Diane Beck"], "authorids": ["~Zhenan_Shao1", "~Linjian_Ma1", "~Yiqing_Zhou1", "~Yibo_Jacky_Zhang1", "~Sanmi_Koyejo1", "~Bo_Li19", "~Diane_Beck1"], "authors_source": "OpenReview API", "abstract": "Humans effortlessly navigate the visual world, yet deep neural networks (DNNs), despite excelling at many visual tasks, are surprisingly vulnerable to minor image perturbations. Past theories suggest human visual robustness arises from a representational space that evolves along the ventral visual stream (VVS) of the brain to increasingly tolerate object transformations. To test whether robustness is supported by such progression as opposed to being confined to specialized higher-order regions, we trained DNNs to align their representations with human neural responses from consecutive VVS regions during visual tasks. We demonstrate a hierarchical improvement in DNN robustness: alignment to higher-order VVS regions yields greater gains. To investigate the mechanism behind this improvement, we test a prominent hypothesis that attributes human visual robustness to the unique geometry of neural category manifolds in the VVS. We show that desirable manifold properties, specifically, smaller extent and better linear separability, emerge across the human VVS. These properties are inherited by DNNs via neural guidance and can predict their subsequent robustness gains. Further, we show that supervision from neural manifolds alone, via manifold guidance, suffices to qualitatively reproduce the hierarchical robustness improvements. Together, our results highlight the evolving VVS representational space as critical for robust visual inference, with the more linearly separable category manifolds as one potential mechanism, offering insights for building more resilient AI systems.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "il4qid0a89", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10486/Reviewer_tPLc"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper studies whether aligning DNNs to human neural responses with specifically-designed techniques can improve adversarial robustness. The authors proposed a dual-head architecture to train ResNet models with “neural guidance” by simultaneously optimizing classification tasks and alignment to fMRI responses from seven VVS regions. The study shows that guidance from higher-order VVS regions leads to greater robustness against adversarial attacks. To analyze and explain this phenomenon, they apply statistical modeling, Mean-Field Theoretic Manifold Analysis, and show that human neural category manifolds become less diffuse and more linearly separable along the VVS hierarchy. Finally, as opposed to point-to-point guidance, the authors introduce “manifold guidance” to match coarse geometric properties and effectively reproduce the hierarchical robustness effect.", "review_text": "This paper studies whether aligning DNNs to human neural responses with specifically-designed techniques can improve adversarial robustness. The authors proposed a dual-head architecture to train ResNet models with “neural guidance” by simultaneously optimizing classification tasks and alignment to fMRI responses from seven VVS regions. The study shows that guidance from higher-order VVS regions leads to greater robustness against adversarial attacks. To analyze and explain this phenomenon, they apply statistical modeling, Mean-Field Theoretic Manifold Analysis, and show that human neural category manifolds become less diffuse and more linearly separable along the VVS hierarchy. Finally, as opposed to point-to-point guidance, the authors introduce “manifold guidance” to match coarse geometric properties and effectively reproduce the hierarchical robustness effect.", "strengths": "- **Clear motivation and novel systematic investigation.** The paper studies and improves modes’ adversarial robustness via the alignment between DNNs and human visual systems. It provides a systematic examination of how robustness evolves across multiple consecutive human VVS regions, which is different from prior work that focuses on isolated areas such as just V1 or IT. This provides valuable insights into the hierarchical nature of visual robustness.\n- **Rigorous experimental design and statistical analysis.** I find this paper especially interesting because it goes beyond only demonstrating the phenomenon but analyze how it occurs with the MFTMA analysis and further claims the importance of geometric structure with the manifold guidance experiments. The evaluation includes multiple baseline conditions, various NSD subjects (Figure 7), adversarial attacks and configurations (Figure 8), and tasks (Figure 9). \n- **Clear Presentation.** The paper is well-written with clear communication of the methodology and evaluation. Well-designed visualizations help make the concepts of the alignment between human visual system and DNNs accessible.", "weaknesses": "We thank the authors for submitting the paper to ICLR 2026! There are a few weaknesses listed below which I believe can make the paper better.\n- **Limited absolute robustness gain.** While the hierarchical pattern is consistent, with results from different NSD subjects, the absolute robustness improvements of TO-guided models are small. They still show substantial vulnerability to adversarial perturbations which are imperceptible to humans. The authors already acknowledge this limitation but more further insights on future approaches which can scale to human-level robustness or if there is an existence of fundamental barriers would make the work more significant to the field. \n- **Manifold guidance limitations.** While manifold guidance reproduces the hierarchical pattern, the absolute improvements are very limited (Figure 4B). The linear approximation of manifold geometry seem to be too coarse. It would be good to see what additional information beyond manifold structure contributes to adversarial robustness in the full neural guidance condition. \n- **Architecture generalization and other related work.** The paper focuses on ResNet architectures, it is unclear whether the findings can generalize to widely-used architectures such as vision transformers, other CNN architectures, and biologically-inspired architectures such as CORnet. Also, many defense mechanisms seem related but not discussed in the paper such as biologically-inspired defenses beyond neural alignment, adversarial training which also improves manifold geometric, etc.", "questions": "- Does the manifold geometry cause robustness, or do both come from some other property of neural representations? Could you test this by techniques such as directly controlling manifold properties independent of neural guidance?\n- What if you train a model to align to different VVS regions simultaneously? Would this provide the model with more information and help improve robustness? \n- There are many other robustness measures which come from human-model misalignment, such as common corruptions and out-of-distribution scenarios. Would the findings generalize to those scenarios?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies whether aligning DNNs to human neural responses with specifically-designed techniques can improve adversarial robustness. The authors proposed a dual-head architecture to train ResNet models with “neural guidance” by simultaneously optimizing classification tasks and alignment to fMRI responses from seven VVS regions. The study shows that guidance from higher-order VVS regions leads to greater robustness against adversarial attacks. To analyze and explain this phenomenon, they apply statistical modeling, Mean-Field Theoretic Manifold Analysis, and show that human neural category manifolds become less diffuse and more linearly separable along the VVS hierarchy. Finally, as opposed to point-to-point guidance, the authors introduce “manifold guidance” to match coarse geometric properties and effectively reproduce the hierarchical robustness effect.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- **Clear motivation and novel systematic investigation.** The paper studies and improves modes’ adversarial robustness via the alignment between DNNs and human visual systems. It provides a systematic examination of how robustness evolves across multiple consecutive human VVS regions, which is different from prior work that focuses on isolated areas such as just V1 or IT. This provides valuable insights into the hierarchical nature of visual robustness.\n- **Rigorous experimental design and statistical analysis.** I find this paper especially interesting because it goes beyond only demonstrating the phenomenon but analyze how it occurs with the MFTMA analysis and further claims the importance of geometric structure with the manifold guidance experiments. The evaluation includes multiple baseline conditions, various NSD subjects (Figure 7), adversarial attacks and configurations (Figure 8), and tasks (Figure 9). \n- **Clear Presentation.** The paper is well-written with clear communication of the methodology and evaluation. Well-designed visualizations help make the concepts of the alignment between human visual system and DNNs accessible.", "weaknesses": "We thank the authors for submitting the paper to ICLR 2026! There are a few weaknesses listed below which I believe can make the paper better.\n- **Limited absolute robustness gain.** While the hierarchical pattern is consistent, with results from different NSD subjects, the absolute robustness improvements of TO-guided models are small. They still show substantial vulnerability to adversarial perturbations which are imperceptible to humans. The authors already acknowledge this limitation but more further insights on future approaches which can scale to human-level robustness or if there is an existence of fundamental barriers would make the work more significant to the field. \n- **Manifold guidance limitations.** While manifold guidance reproduces the hierarchical pattern, the absolute improvements are very limited (Figure 4B). The linear approximation of manifold geometry seem to be too coarse. It would be good to see what additional information beyond manifold structure contributes to adversarial robustness in the full neural guidance condition. \n- **Architecture generalization and other related work.** The paper focuses on ResNet architectures, it is unclear whether the findings can generalize to widely-used architectures such as vision transformers, other CNN architectures, and biologically-inspired architectures such as CORnet. Also, many defense mechanisms seem related but not discussed in the paper such as biologically-inspired defenses beyond neural alignment, adversarial training which also improves manifold geometric, etc.", "questions": "- Does the manifold geometry cause robustness, or do both come from some other property of neural representations? Could you test this by techniques such as directly controlling manifold properties independent of neural guidance?\n- What if you train a model to align to different VVS regions simultaneously? Would this provide the model with more information and help improve robustness? \n- There are many other robustness measures which come from human-model misalignment, such as common corruptions and out-of-distribution scenarios. Would the findings generalize to those scenarios?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762130596444}, {"id": "P1454w9114", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10486/Reviewer_GQbY"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 5, "summary": "The paper investigates the neural mechanisms underlying human visual robustness and explores how these can be transferred to deep neural networks (DNNs). Specifically, the authors test the hypothesis that robustness in human vision arises progressively along the ventral visual stream (VVS)—from early visual areas (V1, V2) to higher-order regions (e.g., TO, PHC)—rather than being localized solely in specialized cortical areas.\nTo evaluate this, the authors train DNNs to align their internal representations with human neural responses from consecutive VVS regions using data from the Natural Scenes Dataset (NSD). They demonstrate that DNNs aligned to later visual areas exhibit greater adversarial robustness than those aligned to early areas, establishing a clear hierarchy. The study further connects this effect to the geometry of category manifolds in neural representational space—showing that smaller manifold extent and greater linear separability in higher-order regions predict improved robustness. Finally, the authors introduce “manifold guidance”, aligning DNN category manifold geometry (rather than individual activations) with human neural manifolds, and find that it qualitatively reproduces the hierarchical robustness pattern.", "review_text": "The paper investigates the neural mechanisms underlying human visual robustness and explores how these can be transferred to deep neural networks (DNNs). Specifically, the authors test the hypothesis that robustness in human vision arises progressively along the ventral visual stream (VVS)—from early visual areas (V1, V2) to higher-order regions (e.g., TO, PHC)—rather than being localized solely in specialized cortical areas.\nTo evaluate this, the authors train DNNs to align their internal representations with human neural responses from consecutive VVS regions using data from the Natural Scenes Dataset (NSD). They demonstrate that DNNs aligned to later visual areas exhibit greater adversarial robustness than those aligned to early areas, establishing a clear hierarchy. The study further connects this effect to the geometry of category manifolds in neural representational space—showing that smaller manifold extent and greater linear separability in higher-order regions predict improved robustness. Finally, the authors introduce “manifold guidance”, aligning DNN category manifold geometry (rather than individual activations) with human neural manifolds, and find that it qualitatively reproduces the hierarchical robustness pattern.", "strengths": "Strong experimental validation: multiple controls, cross-subject replication, and attack-type diversity.\n\n\nNovel conceptual framing: robustness as a byproduct of manifold geometry inherited from neural data.\n\n\nElegant theoretical integration: combines neuroscience principles with modern robustness evaluation.\n\n\nHigh reproducibility: clear methodological exposition, use of open datasets (NSD), and code availability.\n\n\nMechanistic insight: links representational geometry to robustness, moving beyond surface-level performance metrics.", "weaknesses": "The reliance on fMRI-based predictors limits representational precision; neural alignment fidelity could improve with higher-resolution data (e.g., intracranial recordings).\n\n\nThe scope of tasks (primarily object classification) is limited; extending analyses to temporal or contextual visual understanding would strengthen generalization claims.\n\n\nManifold guidance, while promising, remains a coarse approximation; incorporating nonlinear manifold constraints could further clarify its efficacy.\n\n\nThe current framework does not model feedback and recurrence, known to be crucial for human visual robustness.", "questions": "Could the authors quantify how fMRI signal noise or predictor quality influences the strength of the robustness hierarchy?\n\n\nWould the same hierarchical effect persist if neural alignment were performed using multi-subject averaged representations rather than subject-specific ROIs?\n\n\nCan the manifold guidance loss be extended to self-supervised or contrastive settings to assess scalability beyond supervised classification?\n\n\nHow do the authors interpret the trade-off between accuracy and robustness observed in later VVS-guided models?\n\nThe authors have a few missing citations: In line 40 around the unnatural image degradations, the authors should missed Extreme Image Transforms (EITs) [Crowder et al., 2022; Malik et al., 2023, Biol Cybernetics, Malik et al., 2023, arXiv] which present a novel view of structural changes in the input images. \n\nWhen comparing previous work of Dapello et al., the authors should consider comparing their work to VOneNet [Dapello et al., 2020, NeurIPS], which fails under similar circumstances as EITs and is not widely tested on datasets like ImageNet-C [Hendrycks et al, 2019, ICLR]. \n\nThe authors have uploaded a zip file for the code, but are encouraged to release it online on a publicly accessible platform.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates the neural mechanisms underlying human visual robustness and explores how these can be transferred to deep neural networks (DNNs). Specifically, the authors test the hypothesis that robustness in human vision arises progressively along the ventral visual stream (VVS)—from early visual areas (V1, V2) to higher-order regions (e.g., TO, PHC)—rather than being localized solely in specialized cortical areas.\nTo evaluate this, the authors train DNNs to align their internal representations with human neural responses from consecutive VVS regions using data from the Natural Scenes Dataset (NSD). They demonstrate that DNNs aligned to later visual areas exhibit greater adversarial robustness than those aligned to early areas, establishing a clear hierarchy. The study further connects this effect to the geometry of category manifolds in neural representational space—showing that smaller manifold extent and greater linear separability in higher-order regions predict improved robustness. Finally, the authors introduce “manifold guidance”, aligning DNN category manifold geometry (rather than individual activations) with human neural manifolds, and find that it qualitatively reproduces the hierarchical robustness pattern.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "Strong experimental validation: multiple controls, cross-subject replication, and attack-type diversity.\n\n\nNovel conceptual framing: robustness as a byproduct of manifold geometry inherited from neural data.\n\n\nElegant theoretical integration: combines neuroscience principles with modern robustness evaluation.\n\n\nHigh reproducibility: clear methodological exposition, use of open datasets (NSD), and code availability.\n\n\nMechanistic insight: links representational geometry to robustness, moving beyond surface-level performance metrics.", "weaknesses": "The reliance on fMRI-based predictors limits representational precision; neural alignment fidelity could improve with higher-resolution data (e.g., intracranial recordings).\n\n\nThe scope of tasks (primarily object classification) is limited; extending analyses to temporal or contextual visual understanding would strengthen generalization claims.\n\n\nManifold guidance, while promising, remains a coarse approximation; incorporating nonlinear manifold constraints could further clarify its efficacy.\n\n\nThe current framework does not model feedback and recurrence, known to be crucial for human visual robustness.", "questions": "Could the authors quantify how fMRI signal noise or predictor quality influences the strength of the robustness hierarchy?\n\n\nWould the same hierarchical effect persist if neural alignment were performed using multi-subject averaged representations rather than subject-specific ROIs?\n\n\nCan the manifold guidance loss be extended to self-supervised or contrastive settings to assess scalability beyond supervised classification?\n\n\nHow do the authors interpret the trade-off between accuracy and robustness observed in later VVS-guided models?\n\nThe authors have a few missing citations: In line 40 around the unnatural image degradations, the authors should missed Extreme Image Transforms (EITs) [Crowder et al., 2022; Malik et al., 2023, Biol Cybernetics, Malik et al., 2023, arXiv] which present a novel view of structural changes in the input images. \n\nWhen comparing previous work of Dapello et al., the authors should consider comparing their work to VOneNet [Dapello et al., 2020, NeurIPS], which fails under similar circumstances as EITs and is not widely tested on datasets like ImageNet-C [Hendrycks et al, 2019, ICLR]. \n\nThe authors have uploaded a zip file for the code, but are encouraged to release it online on a publicly accessible platform.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762007530660}, {"id": "lBkdSisrOd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10486/Reviewer_zmbi"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper is motivated by the robstness of the human visual system to adversarial examples, which have been studied extensively in the computer vision literature. To better understand this phenomenon, the VVS in human brains is modeled using deep neural networks. Using fMRI data from humans looking at images, neural predictors are traubed to model the responses of regions of interest in the brain to unseen images. These predictions are then used to align the outputs of neural heads, which are added to traditional ResNet image classifiers. A related training method where representations are aligned with compact manifolds is also defined in order to test the manifold disentanglement hypothesis. Experiments show that neural guidance does in fact increase robustness, and the hierarchy of robustness when aligning models with different regions of interest aligns with that of the human visual system. Further experiments explore the learned representation spaces of the NG-models and demonstrate the effectiveness of the manifold guidance method.", "review_text": "This paper is motivated by the robstness of the human visual system to adversarial examples, which have been studied extensively in the computer vision literature. To better understand this phenomenon, the VVS in human brains is modeled using deep neural networks. Using fMRI data from humans looking at images, neural predictors are traubed to model the responses of regions of interest in the brain to unseen images. These predictions are then used to align the outputs of neural heads, which are added to traditional ResNet image classifiers. A related training method where representations are aligned with compact manifolds is also defined in order to test the manifold disentanglement hypothesis. Experiments show that neural guidance does in fact increase robustness, and the hierarchy of robustness when aligning models with different regions of interest aligns with that of the human visual system. Further experiments explore the learned representation spaces of the NG-models and demonstrate the effectiveness of the manifold guidance method.", "strengths": "* The results in this paper are surprising : simply training a model to mimic human brain activity confers adversarial robustness. Not only could this provide insight into the robustness of the human visual system to image perturbations, it could also inform our knowledge of machine learning robustness more generally.\n* The results provide compelling evidence for the manifold disentanglement hypothesis and may be of note to cognitive scientists.\n* The novel manifold guidance loss term that is defined in this paper may inspire future research in training models that are robust to image corruptions.", "weaknesses": "* The model presented doesn't engage much with the existing literature on adversarial robustness. I think that additional space should be devoted to the relationship between this paper and the adversarial training literature in the related work. It would be informative to know whether adversarially trained models also conform to the manifold disentanglement hypothesis.\n* The robustness results would be more compelling if they weren't restricted to $\\ell_p$ bounded adversarial corruptions. There are many types of common (non-adversarial) corruptions that humans are robust to which we want our models to be invariant to (such as those included in Imagenet-C [1]). These corruptions can also be used to define adversarial bounds which are distinct from $\\ell_p$ (i.e. [2]). It's not clear from the results presented that these models are robust to anything other than $\\ell_p$ imperceptible corruptions.\n* Relatedly, I think it's important to note that just because a model is more robust to a specific adversarial attack, it is not necessarily more robust to adversarial attacks in general (as noted in [3]). It is possible that an informed adversary could design an attack that effectively targets neurally-guided models. I think that this should be noted as a limitation.\n* Experiments were only run using fMRI data from four individuals. This low sample size may be unavoidable due to difficulties in collecting data.\n\n[1] Hendrycks, Dan, and Thomas Dietterich. \"Benchmarking neural network robustness to common corruptions and perturbations.\" arXiv preprint arXiv:1903.12261 (2019).\n\n[2] Kaufmann, Max, Daniel Kang, Yi Sun, Steven Basart, Xuwang Yin, Mantas Mazeika, Akul Arora et al. \"Testing robustness against unforeseen adversaries.\" arXiv preprint arXiv:1908.08016 (2019).\n\n[3] Tramer, Florian, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. \"On adaptive attacks to adversarial example defenses.\" Advances in neural information processing systems 33 (2020): 1633-1645.", "questions": "* How does the performance of a DNN trained with neural guidance compare to an adversarially trained model in terms of robustness to adversarial examples?\n* Are the results architecture dependent? Did you run experiments on anything other than ResNet architectures?\n* Is the manifold guidance method of training useful outside of the neural guidance framework? Would you expect to see similar improvements in robustness if the manifolds for each class were not derived from neural data and instead chosen in some other (possibly unsupervised) way?\n* How do your neural and manifold guidance loss terms impact the time complexity of model training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper is motivated by the robstness of the human visual system to adversarial examples, which have been studied extensively in the computer vision literature. To better understand this phenomenon, the VVS in human brains is modeled using deep neural networks. Using fMRI data from humans looking at images, neural predictors are traubed to model the responses of regions of interest in the brain to unseen images. These predictions are then used to align the outputs of neural heads, which are added to traditional ResNet image classifiers. A related training method where representations are aligned with compact manifolds is also defined in order to test the manifold disentanglement hypothesis. Experiments show that neural guidance does in fact increase robustness, and the hierarchy of robustness when aligning models with different regions of interest aligns with that of the human visual system. Further experiments explore the learned representation spaces of the NG-models and demonstrate the effectiveness of the manifold guidance method.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "* The results in this paper are surprising : simply training a model to mimic human brain activity confers adversarial robustness. Not only could this provide insight into the robustness of the human visual system to image perturbations, it could also inform our knowledge of machine learning robustness more generally.\n* The results provide compelling evidence for the manifold disentanglement hypothesis and may be of note to cognitive scientists.\n* The novel manifold guidance loss term that is defined in this paper may inspire future research in training models that are robust to image corruptions.", "weaknesses": "* The model presented doesn't engage much with the existing literature on adversarial robustness. I think that additional space should be devoted to the relationship between this paper and the adversarial training literature in the related work. It would be informative to know whether adversarially trained models also conform to the manifold disentanglement hypothesis.\n* The robustness results would be more compelling if they weren't restricted to $\\ell_p$ bounded adversarial corruptions. There are many types of common (non-adversarial) corruptions that humans are robust to which we want our models to be invariant to (such as those included in Imagenet-C [1]). These corruptions can also be used to define adversarial bounds which are distinct from $\\ell_p$ (i.e. [2]). It's not clear from the results presented that these models are robust to anything other than $\\ell_p$ imperceptible corruptions.\n* Relatedly, I think it's important to note that just because a model is more robust to a specific adversarial attack, it is not necessarily more robust to adversarial attacks in general (as noted in [3]). It is possible that an informed adversary could design an attack that effectively targets neurally-guided models. I think that this should be noted as a limitation.\n* Experiments were only run using fMRI data from four individuals. This low sample size may be unavoidable due to difficulties in collecting data.\n\n[1] Hendrycks, Dan, and Thomas Dietterich. \"Benchmarking neural network robustness to common corruptions and perturbations.\" arXiv preprint arXiv:1903.12261 (2019).\n\n[2] Kaufmann, Max, Daniel Kang, Yi Sun, Steven Basart, Xuwang Yin, Mantas Mazeika, Akul Arora et al. \"Testing robustness against unforeseen adversaries.\" arXiv preprint arXiv:1908.08016 (2019).\n\n[3] Tramer, Florian, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. \"On adaptive attacks to adversarial example defenses.\" Advances in neural information processing systems 33 (2020): 1633-1645.", "questions": "* How does the performance of a DNN trained with neural guidance compare to an adversarially trained model in terms of robustness to adversarial examples?\n* Are the results architecture dependent? Did you run experiments on anything other than ResNet architectures?\n* Is the manifold guidance method of training useful outside of the neural guidance framework? Would you expect to see similar improvements in robustness if the manifolds for each class were not derived from neural data and instead chosen in some other (possibly unsupervised) way?\n* How do your neural and manifold guidance loss terms impact the time complexity of model training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761949352870}, {"id": "58KVuxH5Fi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10486/Reviewer_gx3G"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors co-train DNNs for image classification and alignment with fMRI recordings from the human ventral visual stream. They found that alignment to higher-order regions of the visual system (i.e., LO vs. V1) yielded more linearly separable neural category manifolds and, potentially as a consequence, larger adversarial robustness gains in the DNNs. These results suggest that the representational geometry shaped through transformations applied by the ventral visual system are a key reason why humans do not experience the same sensitivity to adversarial perturbations as DNNs.", "review_text": "The authors co-train DNNs for image classification and alignment with fMRI recordings from the human ventral visual stream. They found that alignment to higher-order regions of the visual system (i.e., LO vs. V1) yielded more linearly separable neural category manifolds and, potentially as a consequence, larger adversarial robustness gains in the DNNs. These results suggest that the representational geometry shaped through transformations applied by the ventral visual system are a key reason why humans do not experience the same sensitivity to adversarial perturbations as DNNs.", "strengths": "- Lots of interesting ideas and I really like the manifold analyses. It's an elegant way of comparing human and machine vision. \n- Well done adding the weight decay controls.\n- Very creative work.", "weaknesses": "- I'm struggling with the logic. Why would training on higher order layers lead to lower clean accuracy? I would actually guess the opposite. Higher order regions should encode more invariant object representations, which should train models to similarly have more tolerance to transformations as well as noise.\n- No discussion about the inherent limitations of doing this with fMRI. How could that be a limiting factor for this work?\n- \"Past theories suggest human visual robustness arises from a\nrepresentational space that evolves along the ventral visual stream (VVS) of the\nbrain to increasingly tolerate object transformations\" This is a weak sentence. And the follow up is at best underspecified, even for the abstract \"To test whether robustness is\nsupported by such progression as opposed to being confined to specialized higher-\norder regions...\"\n\nMinor: Line 220: \"We first applied this analysis to human neural representations in each of the seven ROIs using NSD.\" I would use something like voxel activity patterns instead of human neural representations here. The latter is some latent property of the patterns that you hope to extract from the former.", "questions": "Questions:\n- It would be good to know if any of the attack strengths are perceivable. Can you add that info?\n- In the MDS: Any idea of what the dimensions capture?\n- Why use regression slopes in Fig 3 but spearman in Fig 4? We should use spearman in both.\n- \"To further test the manifold hypothesis (DiCarlo & Cox, 2007) that identity-preserving transformations should be treated as part of the same category manifold, we enriched each category with adversarially perturbed variants of the test images, thus obtaining 100 samples in total per category, similar to previous work.\" The focus of the DiCarlo and Cox work was on identity-preserving transformations like translation and scaling. What about using these types of transformations? The more I think about it, the more confused I am that the authors focused on adversarial perturbations instead of naturalistic transformations for which there are a variety of popular benchmarks in computer vision (objectnet, imagenet-A, etc.).\n- What about an adversarially trained model as a control?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors co-train DNNs for image classification and alignment with fMRI recordings from the human ventral visual stream. They found that alignment to higher-order regions of the visual system (i.e., LO vs. V1) yielded more linearly separable neural category manifolds and, potentially as a consequence, larger adversarial robustness gains in the DNNs. These results suggest that the representational geometry shaped through transformations applied by the ventral visual system are a key reason why humans do not experience the same sensitivity to adversarial perturbations as DNNs.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Lots of interesting ideas and I really like the manifold analyses. It's an elegant way of comparing human and machine vision. \n- Well done adding the weight decay controls.\n- Very creative work.", "weaknesses": "- I'm struggling with the logic. Why would training on higher order layers lead to lower clean accuracy? I would actually guess the opposite. Higher order regions should encode more invariant object representations, which should train models to similarly have more tolerance to transformations as well as noise.\n- No discussion about the inherent limitations of doing this with fMRI. How could that be a limiting factor for this work?\n- \"Past theories suggest human visual robustness arises from a\nrepresentational space that evolves along the ventral visual stream (VVS) of the\nbrain to increasingly tolerate object transformations\" This is a weak sentence. And the follow up is at best underspecified, even for the abstract \"To test whether robustness is\nsupported by such progression as opposed to being confined to specialized higher-\norder regions...\"\n\nMinor: Line 220: \"We first applied this analysis to human neural representations in each of the seven ROIs using NSD.\" I would use something like voxel activity patterns instead of human neural representations here. The latter is some latent property of the patterns that you hope to extract from the former.", "questions": "Questions:\n- It would be good to know if any of the attack strengths are perceivable. Can you add that info?\n- In the MDS: Any idea of what the dimensions capture?\n- Why use regression slopes in Fig 3 but spearman in Fig 4? We should use spearman in both.\n- \"To further test the manifold hypothesis (DiCarlo & Cox, 2007) that identity-preserving transformations should be treated as part of the same category manifold, we enriched each category with adversarially perturbed variants of the test images, thus obtaining 100 samples in total per category, similar to previous work.\" The focus of the DiCarlo and Cox work was on identity-preserving transformations like translation and scaling. What about using these types of transformations? The more I think about it, the more confused I am that the authors focused on adversarial perturbations instead of naturalistic transformations for which there are a variety of popular benchmarks in computer vision (objectnet, imagenet-A, etc.).\n- What about an adversarially trained model as a control?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761874641452}, {"id": "S4bYoexcg9", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10486/Reviewer_8mca"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 4, "confidence": 3, "summary": "This paper explores the connection between robustness and human neural (fMRI) responses, demonstrating that guiding model representations towards neural responses improves robustness. Furthermore, the authors expand on this result, creating a supervision metric that encourages the learned manifold to have human-inspired beneficial geometric properties. The authors demonstrate that this manifold guidance also improves robustness without the need for human data.", "review_text": "This paper explores the connection between robustness and human neural (fMRI) responses, demonstrating that guiding model representations towards neural responses improves robustness. Furthermore, the authors expand on this result, creating a supervision metric that encourages the learned manifold to have human-inspired beneficial geometric properties. The authors demonstrate that this manifold guidance also improves robustness without the need for human data.", "strengths": "This paper is an important computational exploration of an outstanding hypothesis in visual neuroscience (that robustness is an emergent property of human vision and the VVS).\nThe authors have demonstrated exceptional rigor with comprehensive analyses for each experiment with all appropriate controls for each of the experiments presented.\nThe paper goes beyond demonstrating correlation of neural guidance with robustness to introduce \"Manifold Guidance\" as a form of supervision that isolates this geometric constraint, confirming that representational geometry in a model is sufficient to transfer the robustness principle. A new tool in the toolbox for inducing robustness will be useful, and it will be interesting to see what other aspects of human alignment emerge from this type of supervision.\nWhile the main results are for a single subject, the authors analyze 2 subjects in the supplemental demonstrating generality.\nThe authors test the results over multiple attack types demonstrating generality.\nThe authors clearly state (all but one of) the limitations of their experiments\nI find the paper overall exceptionally well written and generally easy to follow. The discussion in particular is interesting and contextualizes the results well.", "weaknesses": "The main weakness of the paper is in overstatement of implied architecture generality of the robustness findings reported. The evidence is based solely on results from ResNet-18. While ResNet is an obvious choice, adversarial robustness is highly dependent on architecture, so the claim that this is a result applying to DNNs broadly is unsubstantiated. If the authors wish to make this claim broadly, they should demonstrate if this result holds for other DNN architectures such as VIT, bio-inspired CORnet (which the authors mention in related work), and a simpler architecture such as VGG-16. Alternatively, the authors could restrict their claims (including in the title) to be regarding only ResNet-18 \nWhile the ordering of the robustness results (V1 to TO) is highly compelling and supports the central hypothesis, the magnitude of the quantitative improvement for manifold guidance (Figure 4) appears modest. The core finding currently rests on the trend of the means, not the statistical significance of the differences. This could be addressed by measuring variability across initialization seeds, and/or a statistical test for linear trend across regions.\n\nMinor Points:\nFormatting figures 5, 10", "questions": "I suggest the authors either:\n1) Reduce the scope of the claim including in the title to claim these results only for ResNet, not DNNs generally\nor\n2) Test the results presented on a set of other model architectures including VIT, bio-inspired, and a simpler CNN model (no skip-connections).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the connection between robustness and human neural (fMRI) responses, demonstrating that guiding model representations towards neural responses improves robustness. Furthermore, the authors expand on this result, creating a supervision metric that encourages the learned manifold to have human-inspired beneficial geometric properties. The authors demonstrate that this manifold guidance also improves robustness without the need for human data.", "soundness": 3, "presentation": 4, "contribution": 4, "strengths": "This paper is an important computational exploration of an outstanding hypothesis in visual neuroscience (that robustness is an emergent property of human vision and the VVS).\nThe authors have demonstrated exceptional rigor with comprehensive analyses for each experiment with all appropriate controls for each of the experiments presented.\nThe paper goes beyond demonstrating correlation of neural guidance with robustness to introduce \"Manifold Guidance\" as a form of supervision that isolates this geometric constraint, confirming that representational geometry in a model is sufficient to transfer the robustness principle. A new tool in the toolbox for inducing robustness will be useful, and it will be interesting to see what other aspects of human alignment emerge from this type of supervision.\nWhile the main results are for a single subject, the authors analyze 2 subjects in the supplemental demonstrating generality.\nThe authors test the results over multiple attack types demonstrating generality.\nThe authors clearly state (all but one of) the limitations of their experiments\nI find the paper overall exceptionally well written and generally easy to follow. The discussion in particular is interesting and contextualizes the results well.", "weaknesses": "The main weakness of the paper is in overstatement of implied architecture generality of the robustness findings reported. The evidence is based solely on results from ResNet-18. While ResNet is an obvious choice, adversarial robustness is highly dependent on architecture, so the claim that this is a result applying to DNNs broadly is unsubstantiated. If the authors wish to make this claim broadly, they should demonstrate if this result holds for other DNN architectures such as VIT, bio-inspired CORnet (which the authors mention in related work), and a simpler architecture such as VGG-16. Alternatively, the authors could restrict their claims (including in the title) to be regarding only ResNet-18 \nWhile the ordering of the robustness results (V1 to TO) is highly compelling and supports the central hypothesis, the magnitude of the quantitative improvement for manifold guidance (Figure 4) appears modest. The core finding currently rests on the trend of the means, not the statistical significance of the differences. This could be addressed by measuring variability across initialization seeds, and/or a statistical test for linear trend across regions.\n\nMinor Points:\nFormatting figures 5, 10", "questions": "I suggest the authors either:\n1) Reduce the scope of the claim including in the title to claim these results only for ResNet, not DNNs generally\nor\n2) Test the results presented on a set of other model architectures including VIT, bio-inspired, and a simpler CNN model (no skip-connections).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761751967685}], "openreview_url": "https://openreview.net/forum?id=uhVlvT3Pk1", "arxiv_id": "2405.02564", "paper_pdf": "papers/uhVlvT3Pk1.pdf", "paper_pdf_sha256": "31b1720b3facb27d4a5369b0b58d35617ccbb4a5edddce288d55b44e5b90ab54", "paper_pdf_bytes": 7409759, "paper_pdf_source": "openreview", "code_url": "https://github.com/shaox192/NeuralGuidance", "code_repository": "shaox192/NeuralGuidance", "code_commit": "6536eafe51d049b3d5d5f35659c2d9c51f10ec8d", "code_archive": "repos/uhVlvT3Pk1.zip", "code_archive_sha256": "aceaa4d0d83ddab4e5761e7813639f0906f27b92d23321dc0b2e18d4cbd391d8", "code_archive_bytes": 1053556, "code_file_count": 29, "code_extensions": {".py": 26, ".sh": 2, ".ipynb": 1}, "github_disk_usage_kb": 1240, "github_languages": {"Jupyter Notebook": 480570, "Python": 161971, "Shell": 2677}, "github_archived": false, "github_pushed_at": "2026-08-18T22:23:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/leveraging-the-human-ventral-visual-stream-to"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "63r2sTjkCv", "year": 2025, "status": "rejected", "title": "KinDEL: DNA-Encoded Library Dataset for Kinase Inhibitors", "authors": ["Benson Chen", "Tomasz Danel", "Patrick J. McEnaney", "Nikhil Jain", "Kirill Novikov", "Spurti Umesh Akki", "Joshua L. Turnbull", "Virja Atul Pandya", "Boris P. Belotserkovskii", "Jared Bryce Weaver", "Ankita Biswas", "Dat Nguyen", "Gabriel H. S. Dreiman", "Mohammad Sultan", "Nathaniel Stanley", "Daniel M Whalen", "Divya Kanichar", "Christoph Klein", "Emily Fox", "R. Edward Watts"], "authorids": ["~Benson_Chen1", "~Tomasz_Danel1", "~Patrick_J._McEnaney1", "~Nikhil_Jain1", "~Kirill_Novikov1", "~Spurti_Umesh_Akki1", "~Joshua_L._Turnbull1", "~Virja_Atul_Pandya1", "~Boris_P._Belotserkovskii1", "~Jared_Bryce_Weaver1", "~Ankita_Biswas1", "~Dat_Nguyen4", "~Gabriel_H._S._Dreiman1", "~Mohammad_Sultan1", "~Nathaniel_Stanley1", "~Daniel_M_Whalen1", "~Divya_Kanichar1", "~Christoph_Klein1", "~Emily_Fox2", "~R._Edward_Watts1"], "authors_source": "OpenReview API", "abstract": "DNA-Encoded Libraries (DEL) are combinatorial small molecule libraries that offer an efficient way to characterize diverse chemical spaces. Selection experiments using DELs are pivotal to drug discovery efforts, enabling high-throughput hit finding screens. However, limited availability of public DEL datasets hinders the advancement of computational techniques designed to utilize such data. To bridge this gap, we present KinDEL, one of the first large, publicly available DEL datasets on two kinases: Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1). Interest in this data modality is growing due to its ability to generate extensive supervised chemical data that densely samples around select molecular structures. Demonstrating one such application of the data, we benchmark different machine learning techniques to develop predictive models for hit identification; in particular, we highlight recent structure-based probabilistic approaches. Finally, we provide biophysical assay data, both on- and off-DNA, to validate our models on a smaller subset of molecules. Data and code for our benchmarks can be found at: https://kin-del-2024.s3.us-west-2.amazonaws.com/kindel.zip", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "8yMp6iWN9I", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8443/Reviewer_Z3Ms"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The authors have released a new dataset, KinDEL, based on DNA-encoded library (DEL) testing, specifically targeting two kinases, MAPK14 and DDR1. They conducted experiments on this dataset to test model performance.", "review_text": "The authors have released a new dataset, KinDEL, based on DNA-encoded library (DEL) testing, specifically targeting two kinases, MAPK14 and DDR1. They conducted experiments on this dataset to test model performance.", "strengths": "The authors provide a substantial amount of new data.\nThe structure of the article is clear and easy to follow.", "weaknesses": "1. The evaluation is limited to only two kinase targets, MAPK14 and DDR1. Given that both targets are kinases, this dataset may have limited generalizability as a benchmark for models applied to broader, non-kinase targets.\n2. The data-splitting method may ensure that compounds with the same disynthon do not end up in the same split, but it doesn’t fully prevent similar compounds from being grouped together, as disynthons do not necessarily represent the core structure of small molecules. Other approaches, such as scaffold-based or overall molecular similarity-based splits, may yield a more robust assessment.\n3. Presentation Improvements: The table headers are somewhat confusing, making it unclear what the numbers in the table represent without reading the text. In Figure 3, \"SP³\" should be corrected to \"sp³\" for accuracy.", "questions": "1. Why use AI to reduce data noise? If DEL diverges significantly from reality, it may indicate instability or unsuitability for the current task. AI-based discriminative models are inherently inaccurate to some extent, so how effective is it to use one inaccurate method to adjust for another?\n2. More explanation is needed for why certain chemical properties in Figure 3 are considered \"drug-like\". For instance, the molecular weight peak exceeds the traditional threshold of 500, and the QED values are relatively low.\n3. Why is on-DNA data significant here, when off-DNA structures are more relevant for practical applications like drug development? On-DNA structures are unlikely to be developed as drugs.\n4. Why didn’t the authors test more end-to-end models? Also, why did they use Morgan fingerprints as input instead of molecular representations like SMILES strings (1D), molecular graphs (2D), or atomic coordinates (3D)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors have released a new dataset, KinDEL, based on DNA-encoded library (DEL) testing, specifically targeting two kinases, MAPK14 and DDR1. They conducted experiments on this dataset to test model performance.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "The authors provide a substantial amount of new data.\nThe structure of the article is clear and easy to follow.", "weaknesses": "1. The evaluation is limited to only two kinase targets, MAPK14 and DDR1. Given that both targets are kinases, this dataset may have limited generalizability as a benchmark for models applied to broader, non-kinase targets.\n2. The data-splitting method may ensure that compounds with the same disynthon do not end up in the same split, but it doesn’t fully prevent similar compounds from being grouped together, as disynthons do not necessarily represent the core structure of small molecules. Other approaches, such as scaffold-based or overall molecular similarity-based splits, may yield a more robust assessment.\n3. Presentation Improvements: The table headers are somewhat confusing, making it unclear what the numbers in the table represent without reading the text. In Figure 3, \"SP³\" should be corrected to \"sp³\" for accuracy.", "questions": "1. Why use AI to reduce data noise? If DEL diverges significantly from reality, it may indicate instability or unsuitability for the current task. AI-based discriminative models are inherently inaccurate to some extent, so how effective is it to use one inaccurate method to adjust for another?\n2. More explanation is needed for why certain chemical properties in Figure 3 are considered \"drug-like\". For instance, the molecular weight peak exceeds the traditional threshold of 500, and the QED values are relatively low.\n3. Why is on-DNA data significant here, when off-DNA structures are more relevant for practical applications like drug development? On-DNA structures are unlikely to be developed as drugs.\n4. Why didn’t the authors test more end-to-end models? Also, why did they use Morgan fingerprints as input instead of molecular representations like SMILES strings (1D), molecular graphs (2D), or atomic coordinates (3D)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730721212754}, {"id": "rDIpe45Kqh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8443/Reviewer_BHqP"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This study introduces a dataset of DNA-Encoded Libraries (DEL) focused on two specific kinases: Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1). Although the DEL datasets have proven valuable in drug discovery, they are relatively scarce for public use. The introduced dataset, named KinDEL (Kinase Inhibitor DNA-Encoded Library), comprises 81 million small molecules tested against MAPK14 and DDR1 kinases. An experimental evaluation is provided, comparing the performance of the proposed method in both on-DNA and off-DNA scenarios.", "review_text": "This study introduces a dataset of DNA-Encoded Libraries (DEL) focused on two specific kinases: Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1). Although the DEL datasets have proven valuable in drug discovery, they are relatively scarce for public use. The introduced dataset, named KinDEL (Kinase Inhibitor DNA-Encoded Library), comprises 81 million small molecules tested against MAPK14 and DDR1 kinases. An experimental evaluation is provided, comparing the performance of the proposed method in both on-DNA and off-DNA scenarios.", "strengths": "+ The availability of a well-curated and publicly accessible dataset is a notable contribution on its own, making this work a valuable resource for the research community.\n\n+ The authors have conducted a thorough review of the relevant literature, effectively establishing the originality and motivation of the study, and demonstrating a clear understanding of the current state of the field.", "weaknesses": "- Given that the primary contribution of this work is a dataset, it would be beneficial to evaluate state-of-the-art (SOTA) methods on it, both to assess their performance in a new context and to demonstrate the dataset's comprehensiveness. The methods tested seem mostly old ones.", "questions": "- Referring to Section 5.1 Current Datasets, there have been, especially recent, efforts of providing DEL datasets even though they don't exactly match the features offered by the present study. However, it might possible that these existing datasets could be adapted to resemble KinDEL.Could you elaborate on whether KinDEL is novel in that sense, i.e., for instance, enhancing the dataset from [Iqbal et al. (2024)] by incorporating on-DNA synthesis can be challenging?\n\nSumaiya Iqbal, Wei Jiang, Eric Hansen, Tonia Aristotelous, Shuang Liu, Andrew Reidenbach, Cerise Raffier, Alison Leed, Chengkuan Chen, Lawrence Chung, et al. DEL+ ML paradigm for actionable hit discovery–a cross DEL and cross ML model assessment. ChemRxiv\ndoi:10.26434/chemrxiv-2024-2xrx4, 2024.\n\n- Furthermore, how does KinDEL compare to the existing DEL datasets in terms of diversity, and how well does it reflect the performance of existing methods in predicting Poisson enrichment? Does KinDEL offer a more comprehensive or representative testbed for evaluating these methods?\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces a dataset of DNA-Encoded Libraries (DEL) focused on two specific kinases: Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1). Although the DEL datasets have proven valuable in drug discovery, they are relatively scarce for public use. The introduced dataset, named KinDEL (Kinase Inhibitor DNA-Encoded Library), comprises 81 million small molecules tested against MAPK14 and DDR1 kinases. An experimental evaluation is provided, comparing the performance of the proposed method in both on-DNA and off-DNA scenarios.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "+ The availability of a well-curated and publicly accessible dataset is a notable contribution on its own, making this work a valuable resource for the research community.\n\n+ The authors have conducted a thorough review of the relevant literature, effectively establishing the originality and motivation of the study, and demonstrating a clear understanding of the current state of the field.", "weaknesses": "- Given that the primary contribution of this work is a dataset, it would be beneficial to evaluate state-of-the-art (SOTA) methods on it, both to assess their performance in a new context and to demonstrate the dataset's comprehensiveness. The methods tested seem mostly old ones.", "questions": "- Referring to Section 5.1 Current Datasets, there have been, especially recent, efforts of providing DEL datasets even though they don't exactly match the features offered by the present study. However, it might possible that these existing datasets could be adapted to resemble KinDEL.Could you elaborate on whether KinDEL is novel in that sense, i.e., for instance, enhancing the dataset from [Iqbal et al. (2024)] by incorporating on-DNA synthesis can be challenging?\n\nSumaiya Iqbal, Wei Jiang, Eric Hansen, Tonia Aristotelous, Shuang Liu, Andrew Reidenbach, Cerise Raffier, Alison Leed, Chengkuan Chen, Lawrence Chung, et al. DEL+ ML paradigm for actionable hit discovery–a cross DEL and cross ML model assessment. ChemRxiv\ndoi:10.26434/chemrxiv-2024-2xrx4, 2024.\n\n- Furthermore, how does KinDEL compare to the existing DEL datasets in terms of diversity, and how well does it reflect the performance of existing methods in predicting Poisson enrichment? Does KinDEL offer a more comprehensive or representative testbed for evaluating these methods?\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730719105590}, {"id": "TZcDYfIXvf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8443/Reviewer_6jAS"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper presents KinDEL, one of the first large publicly available DNA-Encoded Library (DEL) datasets focused on kinase inhibitors, specifically targeting MAPK14 and DDR1 kinases. The papers benchmarks different machine learning methods for binding prediction.", "review_text": "The paper presents KinDEL, one of the first large publicly available DNA-Encoded Library (DEL) datasets focused on kinase inhibitors, specifically targeting MAPK14 and DDR1 kinases. The papers benchmarks different machine learning methods for binding prediction.", "strengths": "1. the paper is clearly written and contains the experimental details in the appendix.\n\n2. The dataset includes off-DNA data. The benchmark used different data split strategies.", "weaknesses": "1. This paper is a valuable contribution to DEL-based drug discovery. It may serve as a good resource for computational drug discovery. But I am not sure about the importance of this paper for the ICLR community. \n\n2. The dataset has only two targets and both are kinase. The biophysical assay validation set is relatively small.\n3. The authors did not mention the library size or sequence depth of the DEL dataset. Does it have an effect on the dataset?\n4. The authors show that DEL-Compose performs better for off-DNA data. It would be helpful to discuss the potential biases due to the DNA barcode.\n5. Why does the RF method become worse for the disynthon split?", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents KinDEL, one of the first large publicly available DNA-Encoded Library (DEL) datasets focused on kinase inhibitors, specifically targeting MAPK14 and DDR1 kinases. The papers benchmarks different machine learning methods for binding prediction.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. the paper is clearly written and contains the experimental details in the appendix.\n\n2. The dataset includes off-DNA data. The benchmark used different data split strategies.", "weaknesses": "1. This paper is a valuable contribution to DEL-based drug discovery. It may serve as a good resource for computational drug discovery. But I am not sure about the importance of this paper for the ICLR community. \n\n2. The dataset has only two targets and both are kinase. The biophysical assay validation set is relatively small.\n3. The authors did not mention the library size or sequence depth of the DEL dataset. Does it have an effect on the dataset?\n4. The authors show that DEL-Compose performs better for off-DNA data. It would be helpful to discuss the potential biases due to the DNA barcode.\n5. Why does the RF method become worse for the disynthon split?", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730289511064}, {"id": "T76hjnb8GV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8443/Reviewer_63rq"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper propose a new open-source dataset as well as related benchmark for the DEL community. The main motivation behind this paper are: \n\n1. DEL community lacks a large, publicly available DEL dataset to benchmarking tasks. \n2. Current DEL dataset contains large bias and noise, and the existing methods cannot greatly address this issue. \n\nSo, the authors propose an open-source dataset and a related enhancement approach to address these challenges. In details, these improvements and contributions are: \n\n1. KinDEL: a library of 81 million small molecules tested against two kinase targets, MAPK14 and DDR1, which is novel and with large amount. \n2. A comprehensive benchmark tested on current computational methods with on both on-DNA and off-DNA settings.\n\nThe proposed dataset and benchmark have great potentials to stimulate the development of the community.", "review_text": "This paper propose a new open-source dataset as well as related benchmark for the DEL community. The main motivation behind this paper are: \n\n1. DEL community lacks a large, publicly available DEL dataset to benchmarking tasks. \n2. Current DEL dataset contains large bias and noise, and the existing methods cannot greatly address this issue. \n\nSo, the authors propose an open-source dataset and a related enhancement approach to address these challenges. In details, these improvements and contributions are: \n\n1. KinDEL: a library of 81 million small molecules tested against two kinase targets, MAPK14 and DDR1, which is novel and with large amount. \n2. A comprehensive benchmark tested on current computational methods with on both on-DNA and off-DNA settings.\n\nThe proposed dataset and benchmark have great potentials to stimulate the development of the community.", "strengths": "1. The dataset constuction process is reasonable, sound, and comprehensive. Also, the related process is explained clearly.\n2. The corresponding evaluation to the proposed dataset is comprehensive, and the two types of splits \"random\" and \"disynthon\" enhance the perspective of the benchmark. \n3. The evaluation is clear and straightforward (Table1, Table2, and Figure6). \n4. The comparison to the current datasets is well-written and shows the valuable insights from authors and their advanced understanding to the DEL-related tasks.", "weaknesses": "1. While there are very interesting performance comparison in the shown Table (Table1 and Table2), the explaination of the experiment results are expected, such as \"why RF method performs the best on on-DNA set (Line 335)\", \"why there are different performance rankings on on-DNA and off-DNA settings?\" and \"why DEL-Compose(M) and DEL-Compose(S) performs differently on on-DNA and off-DNA settings?\". I believe the insight provided by the authors would make the benchmark more solid and comprehensive. \n2. The proposed experiments including general performance (Table1, and Table2), and the visualization of experimental replicates are somehow not comprehensive enough. Serving as a benchmark for the DEL-community, more views and new settings are required, such as case study of top-ranking candidates (which can be potential candidates for real-world application), subset-Spearman coefficient (which is proposed in the paper \"DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries\" [1]), and potential effect of chemical properties and data source selections (building blocks) to the method performances on the KinDEL dataset. Then, the multi-view perspective of the benchmark could lead to larger contribution to the community. \n\nReference:\n1. Shmilovich K, Chen B, Karaletsos T, et al. DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries[J]. Journal of Chemical Information and Modeling, 2023, 63(9): 2719-2727.", "questions": "1. According to Table1 and Table2, the SOTA performance w.r.t. Spearman coefficient can reach over 0.7, which is a very promising result,  but in paper \"DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries\" [1], there's an another proposed dataset containing Fingerprint and docking poses, where existing methods can only achieve relatively poor performance (around 0.30 w.r.t. negative Spearman coefficient) compared to KinDEL dataset. May the authors explain what is the inner differeneces between two datasets that leads to the obvious differences between two datasets? I also believe this comparison can make this work more solid as the contributing benchmark. \n2. Is it possible to provide an advanced version of KinDEL dataset with machine-aided of molecule docking poses. I understand it's very time-consuming and CPU-resource costly, but the DEL dataset with 1D and 3D modalities would lead to wider applications and evaluations to this community. Considering the potential cost, I believe it is also great to have dataset with only fingerprint (1D) information of the molecules. \n\nReference:\n1. Shmilovich K, Chen B, Karaletsos T, et al. DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries[J]. Journal of Chemical Information and Modeling, 2023, 63(9): 2719-2727.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose a new open-source dataset as well as related benchmark for the DEL community. The main motivation behind this paper are: \n\n1. DEL community lacks a large, publicly available DEL dataset to benchmarking tasks. \n2. Current DEL dataset contains large bias and noise, and the existing methods cannot greatly address this issue. \n\nSo, the authors propose an open-source dataset and a related enhancement approach to address these challenges. In details, these improvements and contributions are: \n\n1. KinDEL: a library of 81 million small molecules tested against two kinase targets, MAPK14 and DDR1, which is novel and with large amount. \n2. A comprehensive benchmark tested on current computational methods with on both on-DNA and off-DNA settings.\n\nThe proposed dataset and benchmark have great potentials to stimulate the development of the community.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "1. The dataset constuction process is reasonable, sound, and comprehensive. Also, the related process is explained clearly.\n2. The corresponding evaluation to the proposed dataset is comprehensive, and the two types of splits \"random\" and \"disynthon\" enhance the perspective of the benchmark. \n3. The evaluation is clear and straightforward (Table1, Table2, and Figure6). \n4. The comparison to the current datasets is well-written and shows the valuable insights from authors and their advanced understanding to the DEL-related tasks.", "weaknesses": "1. While there are very interesting performance comparison in the shown Table (Table1 and Table2), the explaination of the experiment results are expected, such as \"why RF method performs the best on on-DNA set (Line 335)\", \"why there are different performance rankings on on-DNA and off-DNA settings?\" and \"why DEL-Compose(M) and DEL-Compose(S) performs differently on on-DNA and off-DNA settings?\". I believe the insight provided by the authors would make the benchmark more solid and comprehensive. \n2. The proposed experiments including general performance (Table1, and Table2), and the visualization of experimental replicates are somehow not comprehensive enough. Serving as a benchmark for the DEL-community, more views and new settings are required, such as case study of top-ranking candidates (which can be potential candidates for real-world application), subset-Spearman coefficient (which is proposed in the paper \"DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries\" [1]), and potential effect of chemical properties and data source selections (building blocks) to the method performances on the KinDEL dataset. Then, the multi-view perspective of the benchmark could lead to larger contribution to the community. \n\nReference:\n1. Shmilovich K, Chen B, Karaletsos T, et al. DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries[J]. Journal of Chemical Information and Modeling, 2023, 63(9): 2719-2727.", "questions": "1. According to Table1 and Table2, the SOTA performance w.r.t. Spearman coefficient can reach over 0.7, which is a very promising result,  but in paper \"DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries\" [1], there's an another proposed dataset containing Fingerprint and docking poses, where existing methods can only achieve relatively poor performance (around 0.30 w.r.t. negative Spearman coefficient) compared to KinDEL dataset. May the authors explain what is the inner differeneces between two datasets that leads to the obvious differences between two datasets? I also believe this comparison can make this work more solid as the contributing benchmark. \n2. Is it possible to provide an advanced version of KinDEL dataset with machine-aided of molecule docking poses. I understand it's very time-consuming and CPU-resource costly, but the DEL dataset with 1D and 3D modalities would lead to wider applications and evaluations to this community. Considering the potential cost, I believe it is also great to have dataset with only fingerprint (1D) information of the molecules. \n\nReference:\n1. Shmilovich K, Chen B, Karaletsos T, et al. DEL-Dock: Molecular Docking-Enabled Modeling of DNA-Encoded Libraries[J]. Journal of Chemical Information and Modeling, 2023, 63(9): 2719-2727.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729232345127}], "openreview_url": "https://openreview.net/forum?id=63r2sTjkCv", "arxiv_id": "2410.08938", "paper_pdf": "papers/63r2sTjkCv.pdf", "paper_pdf_sha256": "784a61f90a2283fc60b66738522c314bee3d834ee2f8d2d023da7e0930f1df08", "paper_pdf_bytes": 10432627, "paper_pdf_source": "openreview", "code_url": "https://github.com/insitro/kindel", "code_repository": "insitro/kindel", "code_commit": "93bcc6f64735160a486a0ea3e69c2f95c71dd2c4", "code_archive": "repos/63r2sTjkCv.zip", "code_archive_sha256": "c67351d69995d74474f718c6d28d37cf68e812feed2800601a6faaeb62fb4b83", "code_archive_bytes": 84208, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 86, "github_languages": {"Python": 73422}, "github_archived": false, "github_pushed_at": "2025-09-02T15:52:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/kindel-dna-encoded-library-dataset-for-kinase"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Zr96FfaUGR", "year": 2024, "status": "rejected", "title": "ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews", "authors": ["Mike D'Arcy", "Alexis Ross", "Erin Bransom", "Bailey Kuehl", "Jonathan Bragg", "Tom Hope", "Doug Downey"], "authorids": ["~Mike_D'Arcy1", "~Alexis_Ross1", "~Erin_Bransom1", "~Bailey_Kuehl1", "~Jonathan_Bragg1", "~Tom_Hope2", "~Doug_Downey1"], "authors_source": "OpenReview API", "abstract": "Revising scientific papers based on peer feedback is a challenging task that requires not only deep scientific knowledge and reasoning, but also the ability to recognize the implicit requests in high-level feedback and to choose the best of many possible ways to update the manuscript in response.  We introduce this task for large language models and release ARIES, a dataset of review comments and their corresponding paper edits, to enable training and evaluating models.  We study two versions of the task: comment-edit alignment and edit generation, and evaluate several baselines, including GPT-4.  We find that models struggle even to identify the edits that correspond to a comment, especially in cases where the comment is phrased in an indirect way or where the edit addresses the spirit of a comment but not the precise request.  When tasked with generating edits, GPT-4 often succeeds in addressing comments on a surface level, but it rigidly follows the wording of the feedback rather than the underlying intent, and includes fewer technical details than human-written edits.  We hope that our formalization, dataset, and analysis will form a foundation for future work in this area.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "gLAHmlvQLc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8084/Reviewer_fRaq"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a new data set named ARIES with scientific reviews, where each comment is matched to a paper edit. This data set could be used to detect which edits correspond to the request and for an edit generation from comment task. The data set consists of 196 human annotations across 42 reviews, judgements if statements are actionable and 3900 comments automatically matched with high precision and low recall. The paper present several approaches to align comments to edits and studies GPT-4 for generating the edits.", "review_text": "This paper introduces a new data set named ARIES with scientific reviews, where each comment is matched to a paper edit. This data set could be used to detect which edits correspond to the request and for an edit generation from comment task. The data set consists of 196 human annotations across 42 reviews, judgements if statements are actionable and 3900 comments automatically matched with high precision and low recall. The paper present several approaches to align comments to edits and studies GPT-4 for generating the edits.", "strengths": "Good annotation task setup and agreement metrics obtained on these annotations.", "weaknesses": "The data set that is manually annotated is small (196 annotations across 42 reviews) and may not be diverse enough.\n\nThe generation task setup is very challenging and includes multiple aspect that would not be available to a predictive model: the context of research at that point, historical information that may be relevant or the goals of the authors which for example would not like to admit a weakness in their publication. Because of this, I think potential applications and extensions would be important to mention.\n\nExperiments in Section 5 require more information about how the models were trained, especially how the negatives are selected. The results currently show better results on Macro F1 for the non fine-tuned model, and worse results when using cross-encoders as compared to bi-encoders, which is unintuitive and looks quite suspicious. Another detail that needs to be mentioned are how the units of text that are matched are computed (a sentence, a paragraph?)\n\nThere are several key assumption made in the data set construction that I think should be better highlighted or organized, such as that responses can be presented in the forum rather than in edits to the paper.\n\nThe experiments in section 6 are only performed using an off the shelf GPT model.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new data set named ARIES with scientific reviews, where each comment is matched to a paper edit. This data set could be used to detect which edits correspond to the request and for an edit generation from comment task. The data set consists of 196 human annotations across 42 reviews, judgements if statements are actionable and 3900 comments automatically matched with high precision and low recall. The paper present several approaches to align comments to edits and studies GPT-4 for generating the edits.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "Good annotation task setup and agreement metrics obtained on these annotations.", "weaknesses": "The data set that is manually annotated is small (196 annotations across 42 reviews) and may not be diverse enough.\n\nThe generation task setup is very challenging and includes multiple aspect that would not be available to a predictive model: the context of research at that point, historical information that may be relevant or the goals of the authors which for example would not like to admit a weakness in their publication. Because of this, I think potential applications and extensions would be important to mention.\n\nExperiments in Section 5 require more information about how the models were trained, especially how the negatives are selected. The results currently show better results on Macro F1 for the non fine-tuned model, and worse results when using cross-encoders as compared to bi-encoders, which is unintuitive and looks quite suspicious. Another detail that needs to be mentioned are how the units of text that are matched are computed (a sentence, a paragraph?)\n\nThere are several key assumption made in the data set construction that I think should be better highlighted or organized, such as that responses can be presented in the forum rather than in edits to the paper.\n\nThe experiments in section 6 are only performed using an off the shelf GPT model.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698992309920}, {"id": "zpsVXqVirg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8084/Reviewer_nrwT"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces the challenging task of revising scientific papers based on peer feedback. The authors present ARIES, a valuable dataset of review comments and corresponding paper edits, to facilitate training and evaluation of models. The study focuses on two subtasks: comment-edit alignment and edit generation. Experiments reveal that existing models, including GPT-4, struggle with these tasks. While GPT-4 can generate edits on a surface level, it rigidly follows the wording of feedback rather than capturing the underlying intent and tends to include fewer technical details than human-written edits. The findings suggest the need for further research in this area, emphasizing the complexities of reasoning about scientific text and addressing challenges in aligning feedback to edits and generating meaningful revisions.", "review_text": "The paper introduces the challenging task of revising scientific papers based on peer feedback. The authors present ARIES, a valuable dataset of review comments and corresponding paper edits, to facilitate training and evaluation of models. The study focuses on two subtasks: comment-edit alignment and edit generation. Experiments reveal that existing models, including GPT-4, struggle with these tasks. While GPT-4 can generate edits on a surface level, it rigidly follows the wording of feedback rather than capturing the underlying intent and tends to include fewer technical details than human-written edits. The findings suggest the need for further research in this area, emphasizing the complexities of reasoning about scientific text and addressing challenges in aligning feedback to edits and generating meaningful revisions.", "strengths": "+ Automatically revising scientific papers based on peer feedback is a meaningful task. Decomposing the task into comment-edit alignment and edit generation is well-motivated.\n\n+ The constructed ARIES dataset has its great practical values in improving NLP systems for not only scientific paper revision tasks but also other tasks during the peer review process.\n\n+ The empirical analyses are comprehensive, with meaningful case studies and error analyses. Different model architectures (BM25, BERT, GPT-4; bi-encoder, cross-encoder) are examined. In particular, the observations of GPT-4's performance on edit generation are inspiring.", "weaknesses": "- The technical novelty is somehow limited, although I understand that the main contribution of this submission is on the dataset and benchmark. All compared approaches are existing models. After the authors obtain meaningful observations from empirical studies, they do not further design an effective method based on their observations to achieve better performance.\n\n- This may be a common criticism for any paper showing that LLMs do not perform well on a certain task: It is possible that the poor performance of GPT-4 is due to inappropriate instructions or prompts. More analyses are needed on the effect of instructions. For example, if chain-of-thoughts prompting is used, would GPT-4 generate \"deeper\" edits?", "questions": "- Could you try chain-of-thoughts prompting or other more advanced techniques to see whether the performance of GPT-4 can be improved?\n\n- I would suggest directly writing \"SPECTER2\" rather than \"SPECTER\" in Table 1. Also, could you specify which adapter is used for SPECTER2?\n\n- The following reference may be very relevant to this paper, considering using GPT-4 for writing peer reviews.\n\n[1] Can large language models provide useful feedback on research papers? A large-scale empirical analysis. arXiv 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the challenging task of revising scientific papers based on peer feedback. The authors present ARIES, a valuable dataset of review comments and corresponding paper edits, to facilitate training and evaluation of models. The study focuses on two subtasks: comment-edit alignment and edit generation. Experiments reveal that existing models, including GPT-4, struggle with these tasks. While GPT-4 can generate edits on a surface level, it rigidly follows the wording of feedback rather than capturing the underlying intent and tends to include fewer technical details than human-written edits. The findings suggest the need for further research in this area, emphasizing the complexities of reasoning about scientific text and addressing challenges in aligning feedback to edits and generating meaningful revisions.", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "strengths": "+ Automatically revising scientific papers based on peer feedback is a meaningful task. Decomposing the task into comment-edit alignment and edit generation is well-motivated.\n\n+ The constructed ARIES dataset has its great practical values in improving NLP systems for not only scientific paper revision tasks but also other tasks during the peer review process.\n\n+ The empirical analyses are comprehensive, with meaningful case studies and error analyses. Different model architectures (BM25, BERT, GPT-4; bi-encoder, cross-encoder) are examined. In particular, the observations of GPT-4's performance on edit generation are inspiring.", "weaknesses": "- The technical novelty is somehow limited, although I understand that the main contribution of this submission is on the dataset and benchmark. All compared approaches are existing models. After the authors obtain meaningful observations from empirical studies, they do not further design an effective method based on their observations to achieve better performance.\n\n- This may be a common criticism for any paper showing that LLMs do not perform well on a certain task: It is possible that the poor performance of GPT-4 is due to inappropriate instructions or prompts. More analyses are needed on the effect of instructions. For example, if chain-of-thoughts prompting is used, would GPT-4 generate \"deeper\" edits?", "questions": "- Could you try chain-of-thoughts prompting or other more advanced techniques to see whether the performance of GPT-4 can be improved?\n\n- I would suggest directly writing \"SPECTER2\" rather than \"SPECTER\" in Table 1. Also, could you specify which adapter is used for SPECTER2?\n\n- The following reference may be very relevant to this paper, considering using GPT-4 for writing peer reviews.\n\n[1] Can large language models provide useful feedback on research papers? A large-scale empirical analysis. arXiv 2023.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698955390116}, {"id": "zokSPyEIQu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8084/Reviewer_Lkvy"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents newly constructed data of review comments and their corresponding edit parts in the original and revised versions of scientific papers taken from OpenReview. They introduced two tasks related to this corpus data: The comment-edit alignment task is to identify the correspondence between a comment and an edit, and the edit generation task is to generate an edit given a reviewer's comment.\nFor the former task, they applied a number of binary classification models, evaluated the results, and discussed the causes of errors. For the latter task, they applied GPT-4 to generate edit responses given a comment, and evaluated and discussed the difference between human and GPT-4 generated edit responses.", "review_text": "This paper presents newly constructed data of review comments and their corresponding edit parts in the original and revised versions of scientific papers taken from OpenReview. They introduced two tasks related to this corpus data: The comment-edit alignment task is to identify the correspondence between a comment and an edit, and the edit generation task is to generate an edit given a reviewer's comment.\nFor the former task, they applied a number of binary classification models, evaluated the results, and discussed the causes of errors. For the latter task, they applied GPT-4 to generate edit responses given a comment, and evaluated and discussed the difference between human and GPT-4 generated edit responses.", "strengths": "- The paper provides unique data that includes reviewers' comments on scientific papers and their corresponding edits obtained from the original and revised versions of the papers.\n- They applied a recent generative language model, GPT-4, to tackle two problems they defined, comment-edit alignment and edit generation tasks, and gave a detailed analysis of the results.", "weaknesses": "- The motivation and the usefulness of the proposed tasks are not clear. For the first task, comment-edit alignment, there are a number of papers that require the authors' answer letter to the editors that describe how the authors have responded to each of the comments raised by the reviewers. For such journal papers, there is no need to find correspondence between the edited parts of the revision. The second task is more puzzling. As the authors claim, GPT-4 often addresses responses on a superficial level without including technical details. It is totally unclear why such superficial or pretending responses are necessary.\n- Even though the comment-edit alignment task is a binary classification task, the overall results are very low, much lower than 50%, which may imply that this data or this task is an ill-formed one.\n- The details of the data construction and the task description are not understandable without referring to appendices.", "questions": "- Why are the results of micro scores for the comment-edit alignment so low, even though it is a binary classification task? What is the proportion between the positive and negative pairs, and what scores will be obtained under random guess?\n- In the macro evaluation, what is the reason that F1 scores are lower than both precision and recall? How are those F1 scores calculated?\n- According to the precision and recall scores for the macro evaluation, the performance of BM25 looks better than or at least competitive with those of GPT-4 multi-edit. It would be better to discuss the differences between the results of those models.\n- What is the motivation behind setting the edit generation task? This looks to be an unanswerable question by other than the authors. The motivation as well as the usefulness of the task should be described.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents newly constructed data of review comments and their corresponding edit parts in the original and revised versions of scientific papers taken from OpenReview. They introduced two tasks related to this corpus data: The comment-edit alignment task is to identify the correspondence between a comment and an edit, and the edit generation task is to generate an edit given a reviewer's comment.\nFor the former task, they applied a number of binary classification models, evaluated the results, and discussed the causes of errors. For the latter task, they applied GPT-4 to generate edit responses given a comment, and evaluated and discussed the difference between human and GPT-4 generated edit responses.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The paper provides unique data that includes reviewers' comments on scientific papers and their corresponding edits obtained from the original and revised versions of the papers.\n- They applied a recent generative language model, GPT-4, to tackle two problems they defined, comment-edit alignment and edit generation tasks, and gave a detailed analysis of the results.", "weaknesses": "- The motivation and the usefulness of the proposed tasks are not clear. For the first task, comment-edit alignment, there are a number of papers that require the authors' answer letter to the editors that describe how the authors have responded to each of the comments raised by the reviewers. For such journal papers, there is no need to find correspondence between the edited parts of the revision. The second task is more puzzling. As the authors claim, GPT-4 often addresses responses on a superficial level without including technical details. It is totally unclear why such superficial or pretending responses are necessary.\n- Even though the comment-edit alignment task is a binary classification task, the overall results are very low, much lower than 50%, which may imply that this data or this task is an ill-formed one.\n- The details of the data construction and the task description are not understandable without referring to appendices.", "questions": "- Why are the results of micro scores for the comment-edit alignment so low, even though it is a binary classification task? What is the proportion between the positive and negative pairs, and what scores will be obtained under random guess?\n- In the macro evaluation, what is the reason that F1 scores are lower than both precision and recall? How are those F1 scores calculated?\n- According to the precision and recall scores for the macro evaluation, the performance of BM25 looks better than or at least competitive with those of GPT-4 multi-edit. It would be better to discuss the differences between the results of those models.\n- What is the motivation behind setting the edit generation task? This looks to be an unanswerable question by other than the authors. The motivation as well as the usefulness of the task should be described.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "The paper is not quite related to representation learning.", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698807477586}, {"id": "SfZ954ccRo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8084/Reviewer_md3r"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This presents a dataset of paper review comments aligned to paper edits. It extracts an \"edits\" set somewhat cleverly by just getting different paper versions from openreview, doing pdf-parsing, and looking at differences, and then compares those changes to comments in reviews of the paper to detect when changes seem to be entailed by a specific change in the paper.   They provided a small dataset of full manual alignment between those edits and the reviews and a larger low-recall high-precision automatically evaluated dataset using an additional signal from rebuttals. Models are then presented both for generating those edits in response to review comments and for locating them within a text.", "review_text": "This presents a dataset of paper review comments aligned to paper edits. It extracts an \"edits\" set somewhat cleverly by just getting different paper versions from openreview, doing pdf-parsing, and looking at differences, and then compares those changes to comments in reviews of the paper to detect when changes seem to be entailed by a specific change in the paper.   They provided a small dataset of full manual alignment between those edits and the reviews and a larger low-recall high-precision automatically evaluated dataset using an additional signal from rebuttals. Models are then presented both for generating those edits in response to review comments and for locating them within a text.", "strengths": "- The biggest strength is that this paper provides an extensive analysis of the contours of their task and analysis of the underlying phenomena.   \n-They treat the task with relevant amounts of delicacy, by being very explicit about models lacking the lab notes/data to make many of these edits and doing labeling of actionability to study the issue. This adds nuance to what might otherwise be a problematic task with factuality issues. \n- The model work on comment-edit alignment seems relatively rigorous, using domain-relevant models like SPECTRE2.", "weaknesses": "1) The core manual data is very tiny (196 alignments, 42 reviews total) and so some of the value of the dataset really rests on the quality of the synthetic alignment work detailed in A.5. \n2) It's hard to have a good intuition about the quality of their IAA with Jaccard overlap of 65% in the reviews, but it does call into question whether the segmentation/identification of relevant review comments is clear.  Insofar as there is the whole field of peer review segmentation and typing  (e.g. Xua et al. 2021, Cheng et al. 2021, Kennard et al. 2021, Dycke et al. 2023), it might be worth checking if any of that is relevant for use (some of those may also be relevant to the analysis types).", "questions": "I appreciated the \"action class\" analysis the authors provided over the manual data, and am curious whether there is data (or even impressions) regarding whether that distribution of types is actually the same in the synthetic data. Wouldn't some types of actions be more likely to have their corresponding edits repeated in the rebuttal (e.g., explain) and some types very unlikely to be repeated in the rebuttal (e.g., remove)? \nIf this is envisioned as part of a workflow, does it really make sense for the model to be freely determining in effect whether to comply, disagree, make promises, etc.?  Wouldn't those components ideally make sense as a starting assumption for an edit?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This presents a dataset of paper review comments aligned to paper edits. It extracts an \"edits\" set somewhat cleverly by just getting different paper versions from openreview, doing pdf-parsing, and looking at differences, and then compares those changes to comments in reviews of the paper to detect when changes seem to be entailed by a specific change in the paper.   They provided a small dataset of full manual alignment between those edits and the reviews and a larger low-recall high-precision automatically evaluated dataset using an additional signal from rebuttals. Models are then presented both for generating those edits in response to review comments and for locating them within a text.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- The biggest strength is that this paper provides an extensive analysis of the contours of their task and analysis of the underlying phenomena.   \n-They treat the task with relevant amounts of delicacy, by being very explicit about models lacking the lab notes/data to make many of these edits and doing labeling of actionability to study the issue. This adds nuance to what might otherwise be a problematic task with factuality issues. \n- The model work on comment-edit alignment seems relatively rigorous, using domain-relevant models like SPECTRE2.", "weaknesses": "1) The core manual data is very tiny (196 alignments, 42 reviews total) and so some of the value of the dataset really rests on the quality of the synthetic alignment work detailed in A.5. \n2) It's hard to have a good intuition about the quality of their IAA with Jaccard overlap of 65% in the reviews, but it does call into question whether the segmentation/identification of relevant review comments is clear.  Insofar as there is the whole field of peer review segmentation and typing  (e.g. Xua et al. 2021, Cheng et al. 2021, Kennard et al. 2021, Dycke et al. 2023), it might be worth checking if any of that is relevant for use (some of those may also be relevant to the analysis types).", "questions": "I appreciated the \"action class\" analysis the authors provided over the manual data, and am curious whether there is data (or even impressions) regarding whether that distribution of types is actually the same in the synthetic data. Wouldn't some types of actions be more likely to have their corresponding edits repeated in the rebuttal (e.g., explain) and some types very unlikely to be repeated in the rebuttal (e.g., remove)? \nIf this is envisioned as part of a workflow, does it really make sense for the model to be freely determining in effect whether to comply, disagree, make promises, etc.?  Wouldn't those components ideally make sense as a starting assumption for an edit?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698728266216}], "openreview_url": "https://openreview.net/forum?id=Zr96FfaUGR", "arxiv_id": "2306.12587", "paper_pdf": "papers/Zr96FfaUGR.pdf", "paper_pdf_sha256": "f9949decc1aa5348284c77b11fed9ac2eb0a560a711f8ec4cac48e1d6eb359a7", "paper_pdf_bytes": 318531, "paper_pdf_source": "openreview", "code_url": "https://github.com/allenai/aries", "code_repository": "allenai/aries", "code_commit": "5691bae71a101225ed345d0ffc42e47609f03bbb", "code_archive": "repos/Zr96FfaUGR.zip", "code_archive_sha256": "2ebf1fa1ed59315916dc5a937262a322e4f7eb3711c7e1495fd2e7ee07b5ec85", "code_archive_bytes": 80400, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 65, "github_languages": {"Python": 174978, "Dockerfile": 700}, "github_archived": false, "github_pushed_at": "2023-07-05T18:05:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/aries-a-corpus-of-scientific-paper-edits-made"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "i1Z_VysEgu8", "year": 2023, "status": "rejected", "title": "Time-Myopic Go-Explore: Learning A State Representation for the Go-Explore Paradigm", "authors": ["Marc Höftmann", "Jan Robine", "Stefan Harmeling"], "authorids": ["~Marc_Höftmann1", "~Jan_Robine1", "~Stefan_Harmeling1"], "authors_source": "OpenReview API", "abstract": "Very large state spaces with a sparse reward signal are difficult to explore. The lack of a sophisticated guidance results in a poor performance for numerous reinforcement learning algorithms. In these cases, the commonly used random exploration is often not helpful. The literature shows that this kind of environments require enormous efforts to systematically explore large chunks of the state space. Learned state representations can help here to improve the search by providing semantic context and build a structure on top of the raw observations. In this work we introduce a novel time-myopic state representation that clusters temporal close states together while providing a time prediction capability between them. By adapting this model to the Go-Explore paradigm (Ecoffet et al., 2021b), we demonstrate the first learned state representation that reliably estimates novelty instead of using the hand-crafted representation heuristic. Our method shows an improved solution for the detachment problem which still remains an issue at the Go-Explore Exploration Phase. We provide evidence that our proposed method covers the entire state space with respect to all possible time trajectories — without causing disadvantageous conflict-overlaps in the cell archive. Analogous to native Go-Explore, our approach is evaluated on the hard exploration environments MontezumaRevenge, Gravitar and Frostbite (Atari) in order to validate its capabilities on difficult tasks. Our experiments show that time-myopic Go-Explore is an effective alternative for the domain-engineered heuristic while also being more general. The source code of the method is available on GitHub.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "O4ZKoVC3WNO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4293/Reviewer_Ze5a"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a modification of the Go-Explore algorithm to remove the need for hardcoded state representations (which are used in the original version of the algorithm). Specifically, this work proposes to learn an embedding using an auxiliary loss based on predicting the number of timesteps connecting two observations. This timestep prediction function is also used to populate Go-Explores's goal archive: specifically, states are only added to the archive if they are at least a certain distance from all other states in the archive. This is to ensure that new states/goals are only added if they expand the frontier of explored states. \n\nThe proposed method is evaluated on 3 hard exploration Atari games: Montezuma's Revenge, Gravitar and Frostbite. It doesn't quite match the original Go-Explore, but it does decently well compared to other baselines in terms of performance and sample complexity. ", "review_text": "Overall, I think that the paper's goal of making Go-Explore more general is a good one, and there are some interesting ideas in this paper. However, I believe it needs a good amount of work before it's ready for publication, specifically in terms of its clarity and its experimental results. ", "strengths": "## Strengths\n- Although the original Go-Explore algorithm achieved impressive performance, it used several cheats like using handcrafted state abstractions (based on downsampling images), resetting the environment (or equivalently, exploiting the determinism of the simulator) which are not generally applicable. It would be great to have a version of Go-Explore which keeps the good qualities and performance but does away with these hacks, which this paper tries to make progress towards. \n\n## Weaknesses\n- As far as i can tell, this paper still assumes a deterministic simulator or the ability to reset to previously visited states, which is still a very strong assumption. So the version of Go-Explore it proposes is still not generally applicable. \n- The paper is not very clearly written - see my comments below\n- The results are decent, but it is currently unclear how the proposed algorithm scales with more data (only performance for 1M and 5M are reported). Please report performance at least until 50M so we can compare to the other model-free approaches. \n\nDetailed comments:\n- Sections 1 and 2 are confusing. It was not clear to me until Section 4.1 what the point of predicting the number of timesteps k connecting two different states was. Section 2 is very confusing, with a lot of vague statements such as \"The conflict solver must resolve illogical and non-intuitive conflicts\" or \"The down-scaling method is compressing the image information without considering its semantic content\". Couldn't this be said about many different unsupervised learning methods? In the first paragraph, it seems to be saying that state abstraction is a bad thing, and that the fact that many states are grouped to the same representation is a bad thing. Can't this also be a good thing in that it filters out irrelevant details? It all depends on _how_ the grouping is done. I would suggest providing a detailed motivating example (from Montezuma's Revenge perhaps) that illustrates all of these claimed problems with Go-Explore - having a concrete example illustrating these statements would make this part much clearer. \n\n- It's not clear to me whether the timestep predictor is even necessary. Since you have access to the actual time step, and you are restarting exploration from different states in the archive, you should at least have the number of timesteps between a given state and the state the trajectory was restarted from. What happens if you use this as your archive insertion criterion? It might not work but should be a baseline to justify the need for the learned predictor. A simple baseline could be: record the timestep t of each state, then use the difference between the true timesteps for the archive insertion criterion. \n\n\n\nThere are a number of typos/grammar mistakes - please proofread carefully. A few examples:\n- Abstract: \"clusters temporal close states\" -> \"clusters temporally close states\"\n- Section 3.2: \"a pair of observation encodings is allow us to estimate\" -> \"a pair of observation encodings allows us to estimate\"\n- Section 4.1: \"An short example\" -> \"A short example\"\n- Section 4.1: \"is shown Table 1\" -> \"is shown in Table 1\"\n- Section 6.2: \"shows the practicability\" -> \"shows the practicality\"\n- Section 6.2: \"It is also providing a good performance\" -> \"It also provides a good performance\"\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a modification of the Go-Explore algorithm to remove the need for hardcoded state representations (which are used in the original version of the algorithm). Specifically, this work proposes to learn an embedding using an auxiliary loss based on predicting the number of timesteps connecting two observations. This timestep prediction function is also used to populate Go-Explores's goal archive: specifically, states are only added to the archive if they are at least a certain distance from all other states in the archive. This is to ensure that new states/goals are only added if they expand the frontier of explored states. \n\nThe proposed method is evaluated on 3 hard exploration Atari games: Montezuma's Revenge, Gravitar and Frostbite. It doesn't quite match the original Go-Explore, but it does decently well compared to other baselines in terms of performance and sample complexity. ", "strength_and_weaknesses": "## Strengths\n- Although the original Go-Explore algorithm achieved impressive performance, it used several cheats like using handcrafted state abstractions (based on downsampling images), resetting the environment (or equivalently, exploiting the determinism of the simulator) which are not generally applicable. It would be great to have a version of Go-Explore which keeps the good qualities and performance but does away with these hacks, which this paper tries to make progress towards. \n\n## Weaknesses\n- As far as i can tell, this paper still assumes a deterministic simulator or the ability to reset to previously visited states, which is still a very strong assumption. So the version of Go-Explore it proposes is still not generally applicable. \n- The paper is not very clearly written - see my comments below\n- The results are decent, but it is currently unclear how the proposed algorithm scales with more data (only performance for 1M and 5M are reported). Please report performance at least until 50M so we can compare to the other model-free approaches. \n\nDetailed comments:\n- Sections 1 and 2 are confusing. It was not clear to me until Section 4.1 what the point of predicting the number of timesteps k connecting two different states was. Section 2 is very confusing, with a lot of vague statements such as \"The conflict solver must resolve illogical and non-intuitive conflicts\" or \"The down-scaling method is compressing the image information without considering its semantic content\". Couldn't this be said about many different unsupervised learning methods? In the first paragraph, it seems to be saying that state abstraction is a bad thing, and that the fact that many states are grouped to the same representation is a bad thing. Can't this also be a good thing in that it filters out irrelevant details? It all depends on _how_ the grouping is done. I would suggest providing a detailed motivating example (from Montezuma's Revenge perhaps) that illustrates all of these claimed problems with Go-Explore - having a concrete example illustrating these statements would make this part much clearer. \n\n- It's not clear to me whether the timestep predictor is even necessary. Since you have access to the actual time step, and you are restarting exploration from different states in the archive, you should at least have the number of timesteps between a given state and the state the trajectory was restarted from. What happens if you use this as your archive insertion criterion? It might not work but should be a baseline to justify the need for the learned predictor. A simple baseline could be: record the timestep t of each state, then use the difference between the true timesteps for the archive insertion criterion. \n\n\n\nThere are a number of typos/grammar mistakes - please proofread carefully. A few examples:\n- Abstract: \"clusters temporal close states\" -> \"clusters temporally close states\"\n- Section 3.2: \"a pair of observation encodings is allow us to estimate\" -> \"a pair of observation encodings allows us to estimate\"\n- Section 4.1: \"An short example\" -> \"A short example\"\n- Section 4.1: \"is shown Table 1\" -> \"is shown in Table 1\"\n- Section 6.2: \"shows the practicability\" -> \"shows the practicality\"\n- Section 6.2: \"It is also providing a good performance\" -> \"It also provides a good performance\"\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: \n- Low, see my detailed comments above. \n\nQuality:\n- Medium. Results are ok but it would be nice to have additional results for different sample complexity budgets and the ablation described above. \n\nNovelty: \n- Medium. The algorithm proposes a novel modification to an existing algorithm. \n\nReproducibility:\n- High. The authors promise to open source their code. ", "summary_of_the_review": "Overall, I think that the paper's goal of making Go-Explore more general is a good one, and there are some interesting ideas in this paper. However, I believe it needs a good amount of work before it's ready for publication, specifically in terms of its clarity and its experimental results. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666753377275}, {"id": "EhXACWKYRYV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4293/Reviewer_RwfW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a learned novelty estimator, based on the temporal distance between states, to be used in the Go-Explore architecture as a basis for memory writing decisions. The proposed method is evaluated in the low-data regime on three hard exploration Atari games, where most baselines are outperformed. The effects of using the proposed novelty estimator are investigated in some detail, and compare favorably to the native Go-Explore one.", "review_text": "The idea behind the paper, of modeling time distances for exploration, is a good one in principle. However, the lack of any evidence that the proposed method generalizes is a major gap that renders the paper unsuitable for publication. I recommend rejection.", "strengths": "- Major weakness: the paper only contains results on Montezuma’s Revenge, Gravitar, and Frostbite, over very few frames. There is no evidence in the paper that the method generalizes beyond this regime.\n- Strength: the idea to use time difference predictions as a basis for novelty estimations is in principle sound.\n- Strength: the paper is well-written and easy to read.\n- Strength: the proposed novelty estimator is studied in some detail in the paper, and shown to produce a good distribution of states entered into memory.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a learned novelty estimator, based on the temporal distance between states, to be used in the Go-Explore architecture as a basis for memory writing decisions. The proposed method is evaluated in the low-data regime on three hard exploration Atari games, where most baselines are outperformed. The effects of using the proposed novelty estimator are investigated in some detail, and compare favorably to the native Go-Explore one.", "strength_and_weaknesses": "- Major weakness: the paper only contains results on Montezuma’s Revenge, Gravitar, and Frostbite, over very few frames. There is no evidence in the paper that the method generalizes beyond this regime.\n- Strength: the idea to use time difference predictions as a basis for novelty estimations is in principle sound.\n- Strength: the paper is well-written and easy to read.\n- Strength: the proposed novelty estimator is studied in some detail in the paper, and shown to produce a good distribution of states entered into memory.\n", "clarity,_quality,_novelty_and_reproducibility": "- Clarity: The paper is generally well-written and quite clear. Some details could be improved, e.g. in section 4.1, “the agent samples trajectories from c_1” - is this done by resetting the environment state, or by getting the agent to return?\n- Quality: The explanation and inspection of the proposed mechanism are well done. The main quality issue is the experimental evaluation, which is seriously lacking in width (only three Atari games) and depth (only the initial phases of those games). Questions about the detailed choices of the time distance modeling are a bit moot because of the limited evaluation, but it would have been interesting to discuss things like the time scale introduced through the cutoff L, and the fact that only the difference between the states is fed into the time predictor - wouldn’t there be settings where the constant part is useful when trying to predict the time difference between the states?\n- Novelty: The notion of modelling a difference between two states to obtain an exploration signal is not new, but the specific version here of modelling time difference and using that model output directly as a novelty signal, in Go-Explore, is new as far as I’m aware.\n- Reproducibility: no concerns.\n", "summary_of_the_review": "The idea behind the paper, of modeling time distances for exploration, is a good one in principle. However, the lack of any evidence that the proposed method generalizes is a major gap that renders the paper unsuitable for publication. I recommend rejection.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666697774319}, {"id": "tKM_dv010z2", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4293/Reviewer_7Zd7"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper proposes new learned state representations that can improve exploration by introducing time predictive semantics into the state representation. The authors propose a new notion of novelty based on time prediction representations and use this to create a new insertion criterion, a new cell count strategy and a new restart strategy for exploration in the settings of Go-Explore.", "review_text": "This paper introduces time predictive state representations and uses them as notion of novelty using in the go-explore context to store states and as different visitation counts, and new cell restarts, the paper provides an interesting perspective but lacks in clarity and scientific rigour. Further experiments would be needed to support the claims and provide a scientific understanding of the contributions proposed.\n", "strengths": "Strengths:\nThis paper focuses on a very relevant problem in exploration of state representation and introduces a novel and original way of measuring novelty.\nWeaknesses:\nThe paper lacks scientific rigour. The authors show experiments in table 2 with different frame counts making it very hard to compare among results. And even for the 1M frame budget the Go-Explore native seems to perform better. It is also unclear how precise the time predictive network is for each k, under each game. Is the time prediction correct on new seeds? One of the main arguments for the inadequacy of Go-Explore is that the overlapping conflicts lead to states being abandoned and not further pursued for exploration, how much conflicts is taking place quantitatively and does this lead to a significant performance drop, with the results presented so far it is still unclear this is true and time prediction mitigates this.\nThe paper is also not very clear, for example for the cell selection criteria Cvisits and Cscores is not defined, not even in the text. (see questions below for more)\nThe motivation for starting in proxy cell states is not clearly explained as well, how and where would this help/impact.?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes new learned state representations that can improve exploration by introducing time predictive semantics into the state representation. The authors propose a new notion of novelty based on time prediction representations and use this to create a new insertion criterion, a new cell count strategy and a new restart strategy for exploration in the settings of Go-Explore.", "strength_and_weaknesses": "Strengths:\nThis paper focuses on a very relevant problem in exploration of state representation and introduces a novel and original way of measuring novelty.\nWeaknesses:\nThe paper lacks scientific rigour. The authors show experiments in table 2 with different frame counts making it very hard to compare among results. And even for the 1M frame budget the Go-Explore native seems to perform better. It is also unclear how precise the time predictive network is for each k, under each game. Is the time prediction correct on new seeds? One of the main arguments for the inadequacy of Go-Explore is that the overlapping conflicts lead to states being abandoned and not further pursued for exploration, how much conflicts is taking place quantitatively and does this lead to a significant performance drop, with the results presented so far it is still unclear this is true and time prediction mitigates this.\nThe paper is also not very clear, for example for the cell selection criteria Cvisits and Cscores is not defined, not even in the text. (see questions below for more)\nThe motivation for starting in proxy cell states is not clearly explained as well, how and where would this help/impact.?\n", "clarity,_quality,_novelty_and_reproducibility": "This paper can be greatly improved by providing a more self-contained description of the problems and explicitly state the issues and contributions of the work, such as how does the time predictive semantics help the detachment part? These results have not been provided. \nComments/Suggestions on paper improvement:\nWhy would myopic prediction be good ? isn’t this going against the idea of go explore? \nHow is the policy affecting the time predictions? It is unclear from the text.\nWhere are the weights  in eq. 5 being used? How are Cscore and Cvisit being counted, in particular Cscore? (a more self-contained description would benefit clarity of the paper)\nWhat do the authors mean by extrapolation in 6.1? Of what?\n\nLimitations of the current approach have not been addressed either, I would suggest addressing the difficulties of estimating time predictive representations and how the current approach limits exploration, is spatial information still captured? To what extent?.\n", "summary_of_the_review": "This paper introduces time predictive state representations and uses them as notion of novelty using in the go-explore context to store states and as different visitation counts, and new cell restarts, the paper provides an interesting perspective but lacks in clarity and scientific rigour. Further experiments would be needed to support the claims and provide a scientific understanding of the contributions proposed.\n", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666607438396}, {"id": "7pAnDVtIgzK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4293/Reviewer_ALFv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a time-myopic state representation that estimates the temporal difference between two observations. The time prediction is trained by MSE loss with the normalized temporal difference between two observations from the sampled trajectories. The authors apply this time prediction to Go-Explore (Ecoffet et al. 2021) for achiving more temporally distant (abstracted) observations and reducing the number of elements in the archive. Consequently, the proposed Time-Myoptic Go-Explore achieves better performance than vanilla Go-Explore in Gravitar although it has much worse performance in Frostbite.\n", "review_text": "Although considering temporal distance for storing cells in Go-Explore is interesting and has reasonable improvement in terms of memory efficiency, the performance is generally degraded and requires more frames (sample-inefficient) compared to vanilla Go-Explore. Moreover, the authors only conducted experiments on three environments, and the results are not consistent. Consequently, I gave a reject in the initial review.\n\n**Update**\nAfter reading the authors' responses and other reviews, several concerns are still unsolved and the authors did not respond to the concerns. Hence, I maintain the rating.", "strengths": "**[Strength]**\n\n1. The proposed method that considers temporal difference for storing the observations for Go-Explore is straightforward and seems effective in terms of reducing the temporal redundancy and diversifying the archived observations.\n\n2. The paper is generally well-written and easy to understand.\n\n3. The visualization from an example trajectory (Figure 3) makes a better understanding of time prediction.  \n\n**[Weakness]**\n\n1. According to Table 3, L and Td hugely affect the size of the archive. The sensitivity analysis of the hyperparameters is needed.\n\n2. The description of Eq. (5) is insufficient. How to define the score weight term, especially, how to choose 0.075?\n\n3. The performance in Frostbite is notoriously worse compared to vanilla Go-Explore. Moreover, considering the performance in Montezuma's revenge, it seems that it requires 5 times more frames to achieve a slightly worse performance than Go-Explore (sample inefficient). The authors did not deal with these performance gaps.\n\n4. In Section 5.3, the authors mention that they added “some small local dataset for each cell” but it is not well described.\n\n5. It will be better if the authors can present the comparison with Go-Explore in many different Atari games and other tasks, and analyze when the proposed method is effective and when it is not.\n\n**[minor]**\n\n1. It is good to unify the citation for Go-Explore into one (Nature version).\n\n2. In Table 1, it is better to change ‘Td > 13’ to ‘Td > 0.65’ for consistency.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a time-myopic state representation that estimates the temporal difference between two observations. The time prediction is trained by MSE loss with the normalized temporal difference between two observations from the sampled trajectories. The authors apply this time prediction to Go-Explore (Ecoffet et al. 2021) for achiving more temporally distant (abstracted) observations and reducing the number of elements in the archive. Consequently, the proposed Time-Myoptic Go-Explore achieves better performance than vanilla Go-Explore in Gravitar although it has much worse performance in Frostbite.\n", "strength_and_weaknesses": "**[Strength]**\n\n1. The proposed method that considers temporal difference for storing the observations for Go-Explore is straightforward and seems effective in terms of reducing the temporal redundancy and diversifying the archived observations.\n\n2. The paper is generally well-written and easy to understand.\n\n3. The visualization from an example trajectory (Figure 3) makes a better understanding of time prediction.  \n\n**[Weakness]**\n\n1. According to Table 3, L and Td hugely affect the size of the archive. The sensitivity analysis of the hyperparameters is needed.\n\n2. The description of Eq. (5) is insufficient. How to define the score weight term, especially, how to choose 0.075?\n\n3. The performance in Frostbite is notoriously worse compared to vanilla Go-Explore. Moreover, considering the performance in Montezuma's revenge, it seems that it requires 5 times more frames to achieve a slightly worse performance than Go-Explore (sample inefficient). The authors did not deal with these performance gaps.\n\n4. In Section 5.3, the authors mention that they added “some small local dataset for each cell” but it is not well described.\n\n5. It will be better if the authors can present the comparison with Go-Explore in many different Atari games and other tasks, and analyze when the proposed method is effective and when it is not.\n\n**[minor]**\n\n1. It is good to unify the citation for Go-Explore into one (Nature version).\n\n2. In Table 1, it is better to change ‘Td > 13’ to ‘Td > 0.65’ for consistency.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written but several details such as the local dataset and explanation of worse results are missing. ", "summary_of_the_review": "Although considering temporal distance for storing cells in Go-Explore is interesting and has reasonable improvement in terms of memory efficiency, the performance is generally degraded and requires more frames (sample-inefficient) compared to vanilla Go-Explore. Moreover, the authors only conducted experiments on three environments, and the results are not consistent. Consequently, I gave a reject in the initial review.\n\n**Update**\nAfter reading the authors' responses and other reviews, several concerns are still unsolved and the authors did not respond to the concerns. Hence, I maintain the rating.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666133537409}], "openreview_url": "https://openreview.net/forum?id=i1Z_VysEgu8", "arxiv_id": "2301.05635", "paper_pdf": "papers/i1Z_VysEgu8.pdf", "paper_pdf_sha256": "7a738b15f7bed7bc9451b720c715e19aa9f774c286b74de5c518d57b3c78624d", "paper_pdf_bytes": 1191256, "paper_pdf_source": "openreview", "code_url": "https://github.com/Hauf3n/Time-Myopic-Go-Explore", "code_repository": "Hauf3n/Time-Myopic-Go-Explore", "code_commit": "4f116d4af7ce48f7b544b6b812428cc9473adcb1", "code_archive": "repos/i1Z_VysEgu8.zip", "code_archive_sha256": "4d0994fa4d419b7682181deae5bb8c37778a6846743133d2e900de35c7d1f27c", "code_archive_bytes": 135719, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 134, "github_languages": {"Python": 43198}, "github_archived": false, "github_pushed_at": "2024-04-16T12:33:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/time-myopic-go-explore-learning-a-state"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hxitw01k_Ql", "year": 2022, "status": "rejected", "title": "How memory architecture affects learning in a simple POMDP: the two-hypothesis testing problem", "authors": ["Mario Geiger", "Christophe Eloy", "Matthieu Wyart"], "authorids": ["~Mario_Geiger1", "~Christophe_Eloy1", "~Matthieu_Wyart2"], "authors_source": "OpenReview API", "abstract": "Reinforcement learning is generally difficult for partially observable Markov decision processes (POMDPs), which occurs when the agent's observation is partial or noisy. To seek good performance in POMDPs, one strategy is to endow the agent with a finite memory, whose update is governed by the policy. However, policy optimization is non-convex in that case and can lead to poor training performance for random initialization. The performance can be empirically improved by constraining the memory architecture, then sacrificing optimality to facilitate training. Here we study this trade-off in a two-hypothesis testing problem, akin to the two-arm bandit problem. We compare two extreme cases: (i) the random access memory where any transitions between $M$ memory states are allowed and (ii) a fixed memory where the agent can access its last $m$ actions and rewards. For (i), the probability $q$ to play the worst arm is known to be exponentially small in $M$ for the optimal policy. Our main result is to show that similar performance can be reached for (ii) as well, despite the simplicity of the memory architecture: using a conjecture on Gray-ordered binary necklaces, we find policies for which $q$ is exponentially small in $2^m$, i.e. $q\\sim\\alpha^{2^m}$ with $\\alpha < 1$. In addition, we observe empirically that training from random initialization leads to very poor results for (i), and significantly better results for (ii) thanks to the constraints on the memory architecture.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ZTW6uq3SM9", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2411/Reviewer_42t7"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper attempts to investigate how memory architecture affects learning performance of POMDP agents. It focuses on a very simple two-arm bandit problem with two hypotheses for their probabilities. Two memory structures are considered: random access memory and memento memory. For each memory structure, one policy is provided with a kind of asymptotic optimal performance. Simulation results for simple gradient based learning algorithms are shown under the two memory structures. ", "review_text": "How memory structure affects POMDP learning is very interesting, and it could be useful for future design of learning algorithm under uncertainties. But the paper only focuses on a very simple problem with two possible hypotheses for the unknown POMDP model. Though  insights gained from studying a simple problem could be useful for general problems, the paper doesn't provide discussion on how the ideas from the simple problem could be applied in general problems. Given the policy and analysis for the simple two-hypothesis problem, is there any insight which could be useful for the general problem? Could an agent with RAM or memento memory in a general problem learn the unknown parameter with some simulations? Any feature of the proposed policies may be extended to general POMDP problems? \n\nFor the simple two-armed problem, two memory structures are considered: random access memory and memento memory. For each memory structure, one policy is proposed with performance analysis. The proposed policy itself is not particularly novel since they are  available in the literature. The analysis is limited that the provided performance results are only in the asymptotic region when the discount factor converges to one. Given that the analysis focuses on the limit when the discount factor approaches one, why not considering the average performance formulation? Another import analysis would be the performance of the algorithm in a finite time horizon. Would it be possible to have some finite time performance analysis?\n\nThere are also some issues in the presentation and lack of details. Section 3 states that empirical results for local search are shown in Figure 2; however, Figure 2 only shows the results under different mu and r, and no results for the suggested local optimization. For policy optimization with different initialization schemes, there is no details on how each initialization scheme is implemented. Are the memory schemes only different in their initialization or actually different in their update dynamics? Is the same learning algorithm is used with different memory initialization or different policies are used to adapted to the memory structure?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper attempts to investigate how memory architecture affects learning performance of POMDP agents. It focuses on a very simple two-arm bandit problem with two hypotheses for their probabilities. Two memory structures are considered: random access memory and memento memory. For each memory structure, one policy is provided with a kind of asymptotic optimal performance. Simulation results for simple gradient based learning algorithms are shown under the two memory structures. ", "main_review": "How memory structure affects POMDP learning is very interesting, and it could be useful for future design of learning algorithm under uncertainties. But the paper only focuses on a very simple problem with two possible hypotheses for the unknown POMDP model. Though  insights gained from studying a simple problem could be useful for general problems, the paper doesn't provide discussion on how the ideas from the simple problem could be applied in general problems. Given the policy and analysis for the simple two-hypothesis problem, is there any insight which could be useful for the general problem? Could an agent with RAM or memento memory in a general problem learn the unknown parameter with some simulations? Any feature of the proposed policies may be extended to general POMDP problems? \n\nFor the simple two-armed problem, two memory structures are considered: random access memory and memento memory. For each memory structure, one policy is proposed with performance analysis. The proposed policy itself is not particularly novel since they are  available in the literature. The analysis is limited that the provided performance results are only in the asymptotic region when the discount factor converges to one. Given that the analysis focuses on the limit when the discount factor approaches one, why not considering the average performance formulation? Another import analysis would be the performance of the algorithm in a finite time horizon. Would it be possible to have some finite time performance analysis?\n\nThere are also some issues in the presentation and lack of details. Section 3 states that empirical results for local search are shown in Figure 2; however, Figure 2 only shows the results under different mu and r, and no results for the suggested local optimization. For policy optimization with different initialization schemes, there is no details on how each initialization scheme is implemented. Are the memory schemes only different in their initialization or actually different in their update dynamics? Is the same learning algorithm is used with different memory initialization or different policies are used to adapted to the memory structure?", "summary_of_the_review": "The considered problem is a bit too simple to have meaningful insights, and the paper provides almost no discussion on how the analysis on this simple problem could be useful for general POMDP problems. Even for the simple problem setting, the paper doesn't have performance guarantees for the proposed policy in the non-asymptotic region. \n\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636309755816}, {"id": "dvN9wErhQvq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2411/Reviewer_7ywF"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper tackles a two-armed bandit problem (of means $Ber(1/2+\\mu)$ and $Ber(1/2-\\mu)$, respectively) when memory is limited. Specifically, they model this problem as a POMDP (or, alternatively, hypothesis testing), where the hidden state determines the mean of each arm. The author claim that this simple model might provide some insights on memory architecture in general POMDPs (namely, how the history representation affects the learning process). The authors describe two memory architectures:\n\n(i) A general finite-state machine ('RAM'), which transitions based on the played arm and its reward. Specifically, they focus on the column of confidence policy (CCP). The memory structure - a chain where each end of the chain represents an arm. $+1$ reward moves the state towards an arm and $-1$ moves it away, except for the end states that have low escape probability. The policy - playing the arm that its end-state closer to the current state. \n\n(ii) History is presented as a memory buffer with the last arm plays and their rewards('Memento'). In this case, the authors present a policy that plays the same action sequence cyclically ('necklaces'). These cycles are chained in a Gray-ordering and the policy moves between cycles (change one action) only if a full cycle indicates that this switch is beneficial.\n\nThe authors analyze the (stationary) probability of playing the worse actions and show that the RAM model enjoys better (lower) probability. Finally, the authors try to learn a policy when relying on either RAM or Memento architectures. They showed that a random initialization, the Memento memory performs better, while imposing structural constraints that correspond with the suggested policies in (i) and (ii) make RAM memory perform better.", "review_text": "First, even though the ideas in this paper are fairly intuitive, I admit I found this paper very hard to read. Many of the explanations are not clear and the overall writing can be substantially improved.\n\nThe main novelty, in my opinion, is the presentation of the Gray-ordered necklaces, which I find to be very nice and quite original (while the CCP policy is well studied). However, the theoretical analysis of this structure is lacking, in the sense that the authors only analyzed this specific structure, but gave no theoretical guarantees on its optimality in the class of the Memento-based memories. At most, the authors conjecture optimality based on simulations that might only imply a local optimum. From a practical perspective, I think that this idea does not scale. First, it's not clear how to scale it to more than two arms, nor to non-binary rewards. Second, as a concept, playing actions in a cyclic manner can be disastrous in dynamic environments, so the usefulness of this idea beyond bandit problems is unclear.\n\nMoreover, the connection of the analyzed model to general POMDPs is a bit artificial, and I don't really see what insights the paper gives on memory architectures in POMDP. If I understand correctly, the claim is that Memento works better than RAM with random initialization, even though some RAM structures are better than Memento. This claim is mainly based on the empirical evaluation, and for it to be meaningful, it should be methodologically tested on various POMDP of different characteristics (e.g., such that require long/short memory, fast/slow mixing, etc.). Moreover, the application of both memory architectures is somewhat naive, and as both were extensively studied, I think that this does not suffice to prove this claim.\n\nOther comments\n- p.g. 2 - \"the value is a non-convex function of policy for POMDPs... This problem is even more acute when memory is large or when\nall transitions between memory states are allowed\" - please refrain from such comments. The value in nonconvex in the policy also in MDPs, but it is still possible to find an optimal solution - it's all about the structure of the problem and the chosen representation, so using it to claim that 'more memory is bad' is not really meaningful. For example - what if some RAM structure effectively reconstructs the real state?\n- POMDP formulation - the formulation in this paper feels very tailored for the specific problem and algorithms, but it's not the standard formulation. Specifically, it's not very natural to put the memory state inside the state representation - the memory is a characteristic of the policy, not of the environment, and for some representations, it unnecessarily increases the environment exponentially. \n- Honestly, I didn't understand section 2.2 and the related appendix (B). Please improve the explanations and add more details about how you calculate everything.\n- In Fig. 4, right column, it seems like the RAM-random option did not converge yet - does it converge to a worse value or just converge more slowly?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper tackles a two-armed bandit problem (of means $Ber(1/2+\\mu)$ and $Ber(1/2-\\mu)$, respectively) when memory is limited. Specifically, they model this problem as a POMDP (or, alternatively, hypothesis testing), where the hidden state determines the mean of each arm. The author claim that this simple model might provide some insights on memory architecture in general POMDPs (namely, how the history representation affects the learning process). The authors describe two memory architectures:\n\n(i) A general finite-state machine ('RAM'), which transitions based on the played arm and its reward. Specifically, they focus on the column of confidence policy (CCP). The memory structure - a chain where each end of the chain represents an arm. $+1$ reward moves the state towards an arm and $-1$ moves it away, except for the end states that have low escape probability. The policy - playing the arm that its end-state closer to the current state. \n\n(ii) History is presented as a memory buffer with the last arm plays and their rewards('Memento'). In this case, the authors present a policy that plays the same action sequence cyclically ('necklaces'). These cycles are chained in a Gray-ordering and the policy moves between cycles (change one action) only if a full cycle indicates that this switch is beneficial.\n\nThe authors analyze the (stationary) probability of playing the worse actions and show that the RAM model enjoys better (lower) probability. Finally, the authors try to learn a policy when relying on either RAM or Memento architectures. They showed that a random initialization, the Memento memory performs better, while imposing structural constraints that correspond with the suggested policies in (i) and (ii) make RAM memory perform better.", "main_review": "First, even though the ideas in this paper are fairly intuitive, I admit I found this paper very hard to read. Many of the explanations are not clear and the overall writing can be substantially improved.\n\nThe main novelty, in my opinion, is the presentation of the Gray-ordered necklaces, which I find to be very nice and quite original (while the CCP policy is well studied). However, the theoretical analysis of this structure is lacking, in the sense that the authors only analyzed this specific structure, but gave no theoretical guarantees on its optimality in the class of the Memento-based memories. At most, the authors conjecture optimality based on simulations that might only imply a local optimum. From a practical perspective, I think that this idea does not scale. First, it's not clear how to scale it to more than two arms, nor to non-binary rewards. Second, as a concept, playing actions in a cyclic manner can be disastrous in dynamic environments, so the usefulness of this idea beyond bandit problems is unclear.\n\nMoreover, the connection of the analyzed model to general POMDPs is a bit artificial, and I don't really see what insights the paper gives on memory architectures in POMDP. If I understand correctly, the claim is that Memento works better than RAM with random initialization, even though some RAM structures are better than Memento. This claim is mainly based on the empirical evaluation, and for it to be meaningful, it should be methodologically tested on various POMDP of different characteristics (e.g., such that require long/short memory, fast/slow mixing, etc.). Moreover, the application of both memory architectures is somewhat naive, and as both were extensively studied, I think that this does not suffice to prove this claim.\n\nOther comments\n- p.g. 2 - \"the value is a non-convex function of policy for POMDPs... This problem is even more acute when memory is large or when\nall transitions between memory states are allowed\" - please refrain from such comments. The value in nonconvex in the policy also in MDPs, but it is still possible to find an optimal solution - it's all about the structure of the problem and the chosen representation, so using it to claim that 'more memory is bad' is not really meaningful. For example - what if some RAM structure effectively reconstructs the real state?\n- POMDP formulation - the formulation in this paper feels very tailored for the specific problem and algorithms, but it's not the standard formulation. Specifically, it's not very natural to put the memory state inside the state representation - the memory is a characteristic of the policy, not of the environment, and for some representations, it unnecessarily increases the environment exponentially. \n- Honestly, I didn't understand section 2.2 and the related appendix (B). Please improve the explanations and add more details about how you calculate everything.\n- In Fig. 4, right column, it seems like the RAM-random option did not converge yet - does it converge to a worse value or just converge more slowly?", "summary_of_the_review": "The paper presents an interesting algorithm for two-armed bandit problems with finite history, but the ideas do not scale up to larger problems, and the empirical evaluation is not enough to draw any conclusion on more general POMDPs.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636064560062}, {"id": "oxhONg98Kpb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2411/Reviewer_PEeX"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper compares two memory structures for policies for POMDPs through theoretical and empirical analysis in a relatively simple problem.", "review_text": "The problem studied in the paper is quite a bit of interest for strategy learning and synthesis for POMDPs where, unlike MDPs, memory is known to be essential. \n\nWhile the methodology in the paper is rigorous and explained well, the scope is very narrow to derive any general insight. The comparison is limited to a single problem. \n\nWhile the results are stated relatively carefully, there is room for improvement and better precision. For example, what is the meaning of \"<<\" in Theorem 4.3?\n\nMemory in POMDPs has attracted interest in neighboring contexts and from a different perspective in RL recently. The paper seems to be missing that literature. \nhttps://arxiv.org/pdf/2007.08351.pdf \nhttps://arxiv.org/pdf/2009.11459.pdf\nhttps://arxiv.org/abs/2105.14073", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper compares two memory structures for policies for POMDPs through theoretical and empirical analysis in a relatively simple problem.", "main_review": "The problem studied in the paper is quite a bit of interest for strategy learning and synthesis for POMDPs where, unlike MDPs, memory is known to be essential. \n\nWhile the methodology in the paper is rigorous and explained well, the scope is very narrow to derive any general insight. The comparison is limited to a single problem. \n\nWhile the results are stated relatively carefully, there is room for improvement and better precision. For example, what is the meaning of \"<<\" in Theorem 4.3?\n\nMemory in POMDPs has attracted interest in neighboring contexts and from a different perspective in RL recently. The paper seems to be missing that literature. \nhttps://arxiv.org/pdf/2007.08351.pdf \nhttps://arxiv.org/pdf/2009.11459.pdf\nhttps://arxiv.org/abs/2105.14073", "summary_of_the_review": "An interesting yet very limited paper.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635796488141}, {"id": "NbtZAZt8TpZ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2411/Reviewer_BH7m"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors compare two different memory architectures for the two-hypothesis testing problem. They show that a simple fixed memory where the agent can only access the previous m environment interactions is more robust to random initializations than a more flexible RAM-setup while still maintaining similar performance.", "review_text": "Strengths:\n\n1. The work is in an interesting direction and points out potential issues with current empirical approaches to learning memory allocations.\n\n2. The explanation of optimal policies (such as the necklace policy) and their proofs were well-done and insightful.\n\nWeaknesses:\n\n1. While the authors cite shortcomings in recent works when it comes to learning memory allocations, it would be nice to see the ideas explored here scaled up to similar-sized problems. If the principle of \"restricted memory architectures are easier to learn with and more robust to random initializations\" is true, this should be testable in larger and more complex POMDP's (and policies) even without knowing the optimal policies. \n\nA simple/naive way to do this would be to compare something like RNN's to something like an attention-based architecture that can only observe its last m interactions with the environment (or any other restrictive memory architecture). These could be trained with reinforcement learning in a visual navigation environment, for example. The size of m or the recurrent state could then be played with. \n\nPerforming experiments like this would not be particularly difficult and would easily make the insights of the paper applicable to modern techniques. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors compare two different memory architectures for the two-hypothesis testing problem. They show that a simple fixed memory where the agent can only access the previous m environment interactions is more robust to random initializations than a more flexible RAM-setup while still maintaining similar performance.", "main_review": "Strengths:\n\n1. The work is in an interesting direction and points out potential issues with current empirical approaches to learning memory allocations.\n\n2. The explanation of optimal policies (such as the necklace policy) and their proofs were well-done and insightful.\n\nWeaknesses:\n\n1. While the authors cite shortcomings in recent works when it comes to learning memory allocations, it would be nice to see the ideas explored here scaled up to similar-sized problems. If the principle of \"restricted memory architectures are easier to learn with and more robust to random initializations\" is true, this should be testable in larger and more complex POMDP's (and policies) even without knowing the optimal policies. \n\nA simple/naive way to do this would be to compare something like RNN's to something like an attention-based architecture that can only observe its last m interactions with the environment (or any other restrictive memory architecture). These could be trained with reinforcement learning in a visual navigation environment, for example. The size of m or the recurrent state could then be played with. \n\nPerforming experiments like this would not be particularly difficult and would easily make the insights of the paper applicable to modern techniques. \n", "summary_of_the_review": "This work points out potential issues with current empirical approaches to learning memory allocations and suggests that restricted memory architectures could be easier to use. While the authors do show this to be the case in one very simple environment, they do not show that this concept scales to other larger or more complex environments, such as visual navigation, that are currently used. As-is, the work is simply too narrow in scope and experiments to be a significant contribution to modern techniques and understanding.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635530457571}], "openreview_url": "https://openreview.net/forum?id=hxitw01k_Ql", "arxiv_id": "2106.08849", "paper_pdf": "papers/hxitw01k_Ql.pdf", "paper_pdf_sha256": "7492df3a27ba397b12970c33e3d7e6043420500b38a93cf660b6b3874bdc5623", "paper_pdf_bytes": 550874, "paper_pdf_source": "openreview", "code_url": "https://github.com/pcsl-epfl/two-hypothesis", "code_repository": "pcsl-epfl/two-hypothesis", "code_commit": "4199d45977b8207f4e34f84fa6832df6534aa284", "code_archive": "repos/hxitw01k_Ql.zip", "code_archive_sha256": "c4e259865a3db0673fa648cf774362dc1d84c6d5d51e2f277cb1ae04d5cfc01a", "code_archive_bytes": 36752, "code_file_count": 4, "code_extensions": {".py": 3, ".nb": 1}, "github_disk_usage_kb": 179, "github_languages": {"Mathematica": 167858, "Python": 34821}, "github_archived": false, "github_pushed_at": "2021-10-05T10:43:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-memory-architecture-affects-performance"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "l3YcqzaPlx0", "year": 2021, "status": "rejected", "title": "Neighbor2Seq: Deep Learning on Massive Graphs by Transforming Neighbors to Sequences", "authors": ["Meng Liu", "Shuiwang Ji"], "authorids": ["~Meng_Liu3", "~Shuiwang_Ji1"], "authors_source": "OpenReview API", "abstract": "Modern Graph Neural Networks (GNNs) follow a recursive neighbor-wise message passing scheme and have achieved great success in many fields. However, this recursive design brings expensive computation and huge memory usage, making it difficult to deploy on large-scale graphs. In this work, we propose Neighbor2Seq, which transforms the hierarchical neighborhood of each node into an ordered sequence and enables the subsequent utilization of general deep learning operations, such as convolution and attention. Neighbor2Seq grants our proposed models, i.e., Neighbor2Seq-Conv and Neighbor2Seq-Attn, the ability to learn on arbitrarily large graphs as long as the Neighbor2Seq step can be precomputed. Another potential advantage obtained by the way is that Neighbor2Seq can alleviate the over-squashing issue existing in modern GNNs. We conduct thorough experiments on a massive graph with more than 111 million nodes and 1.6 billion edges, as well as several medium-scale graphs, to evaluate our proposed method. Experimental results demonstrate that our proposed method is scalable to the massive graph and achieves superior performance across datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "b56qZbP3WGh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper383/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n# Summary\n\nThis paper proposed a simple graph neural network architecture that is easy to scale up and perform stochastic training. Instead of performing message passing as commonly used GNN, this paper first performs weighted combinations of node features per each hop of the neighbors of a center node, and then performs either CNN or attention mechanism to aggregate the features and obtain center node embedding. Since the feature aggregation can be performed offline, and the computation can easily be decomposed and stochastic training is straightforward, the method can easily scale up to graphs with 10M nodes. Experiments on median size or large size graphs show the comparable or better performance than alternatives. \n\n# Pros\n\n- The idea is simple and easy to implement, while being effective at the same time. \n- The new design of the architecture \n- Experiments on large scale graphs are convincing.\n\n# Cons\n\n- some comparisons are incomplete\n- Not sure how general this approach would be\n- time/memory cost can be reported and compared\n\n\n# Details\n\nOverall I lean towards accepting the paper. \n\nThis paper provides a simple yet efficient and effective approach for graph node embedding calculation. It enjoys similar computation efficiency as the SGC, but is a bit more expressive in the design, where the CNN or attention is introduced on top of the per-hop embeddings. I like such a simple design that adds the expressiveness without too much additional cost. \n\nIt is also good to see the large scale experiments on OGB graphs. \n\nThere are several aspects that can potentially be improved:\n\n1. The SGC results are not presented in Table 4 or Table 5. It is necessary to include, as this might be the most relevant/comparable baseline.\n\n2.  Would this approach be useful for graph classification? I understand that the main purpose of this approach is scalability, but it would also be good to know the potential limitation on its parameterization. Also it would be more comprehensive to see the results on small benchmarks like Cora, Pubmed, etc,. This is mainly to polish the paper and get a better understanding for users. It would be fine if the results are worse on these small graphs. \n\n3. The runtime (during preprocessing, stochastic training, etc) and memory cost can be reported.\n\n# Questions\n\nI’d like to see the replies to my questions above. \n\n# Improvement\n\nIt would be good to include additional experiments as mentioned above, to make the paper more comprehensive. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review for \"Neighbor2Seq\"", "review": "\n# Summary\n\nThis paper proposed a simple graph neural network architecture that is easy to scale up and perform stochastic training. Instead of performing message passing as commonly used GNN, this paper first performs weighted combinations of node features per each hop of the neighbors of a center node, and then performs either CNN or attention mechanism to aggregate the features and obtain center node embedding. Since the feature aggregation can be performed offline, and the computation can easily be decomposed and stochastic training is straightforward, the method can easily scale up to graphs with 10M nodes. Experiments on median size or large size graphs show the comparable or better performance than alternatives. \n\n# Pros\n\n- The idea is simple and easy to implement, while being effective at the same time. \n- The new design of the architecture \n- Experiments on large scale graphs are convincing.\n\n# Cons\n\n- some comparisons are incomplete\n- Not sure how general this approach would be\n- time/memory cost can be reported and compared\n\n\n# Details\n\nOverall I lean towards accepting the paper. \n\nThis paper provides a simple yet efficient and effective approach for graph node embedding calculation. It enjoys similar computation efficiency as the SGC, but is a bit more expressive in the design, where the CNN or attention is introduced on top of the per-hop embeddings. I like such a simple design that adds the expressiveness without too much additional cost. \n\nIt is also good to see the large scale experiments on OGB graphs. \n\nThere are several aspects that can potentially be improved:\n\n1. The SGC results are not presented in Table 4 or Table 5. It is necessary to include, as this might be the most relevant/comparable baseline.\n\n2.  Would this approach be useful for graph classification? I understand that the main purpose of this approach is scalability, but it would also be good to know the potential limitation on its parameterization. Also it would be more comprehensive to see the results on small benchmarks like Cora, Pubmed, etc,. This is mainly to polish the paper and get a better understanding for users. It would be fine if the results are worse on these small graphs. \n\n3. The runtime (during preprocessing, stochastic training, etc) and memory cost can be reported.\n\n# Questions\n\nI’d like to see the replies to my questions above. \n\n# Improvement\n\nIt would be good to include additional experiments as mentioned above, to make the paper more comprehensive. \n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604032623261}, {"id": "zeZAcIWQZgk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper383/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors proposed a neighborhood to sequence construction & pre-training approach to handling graph representation learning on large graphs.\n\nThe merits of this work include:\n\n1. provide a decent attempt toward solving the computation bottleneck for representation learning on large graphs;\n2. discusses and shows the benefits of the pre-training based on (unsupervised) sequence learning.\n\nThe limitations, on the other hand, is obvious:\n\n1. Pre-training is a relatively standard technique for representation learning on large graphs, and the schema proposed in this paper has very limited novelty;\n2. The application of attention mechanism for sequence learning is also a standard practice that has been widely adopted in this domain;\n3. The only part where this paper may distinguish itself from the previous work is how the sequence is constructed from the neighborhood. However, minimal theoretical discussion and empirical ablation study are provided to reveal the guarantees & benefits of the proposed method.\n\nTherefore, the limited novelty and contribution of this paper clearly outweigh its merits.\n\n=======================================\n\nAfter reviewing the response from the authors, I decide to change my evaluation score and confidence score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Limited novelty and contribution", "review": "The authors proposed a neighborhood to sequence construction & pre-training approach to handling graph representation learning on large graphs.\n\nThe merits of this work include:\n\n1. provide a decent attempt toward solving the computation bottleneck for representation learning on large graphs;\n2. discusses and shows the benefits of the pre-training based on (unsupervised) sequence learning.\n\nThe limitations, on the other hand, is obvious:\n\n1. Pre-training is a relatively standard technique for representation learning on large graphs, and the schema proposed in this paper has very limited novelty;\n2. The application of attention mechanism for sequence learning is also a standard practice that has been widely adopted in this domain;\n3. The only part where this paper may distinguish itself from the previous work is how the sequence is constructed from the neighborhood. However, minimal theoretical discussion and empirical ablation study are provided to reveal the guarantees & benefits of the proposed method.\n\nTherefore, the limited novelty and contribution of this paper clearly outweigh its merits.\n\n=======================================\n\nAfter reviewing the response from the authors, I decide to change my evaluation score and confidence score.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604003557621}, {"id": "1E8Ky5POIc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper383/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method called neighbor2seq that converts the hierarchical structure of the center node to a sequence during message passing in graph neural networks. The proposed method aims to mitigate the issue of excessive computation and memory requirement of training graph neural networks. The proposed models Neighbor2Seq+Conv and Neighbor2Seq+Attn are tested on several datasets including a large scale benchmark dataset (ogbn-papers100M). The result shows some improvement especially on ogbn-papers100M while the improvement is not very obvious on other datasets.\n\nStrength:\n1. The idea of converting the graph topology to a sequence that could be tackled with methods designed for grid-like data is good. \n2. The proposed method shows improvement on a large scale dataset. There are also ablation studies that are useful to understand different components in the method, for example, removing the sequence information (position encoding here) clearly hurts the model performance. But it's surprising to see that Neighbor2Seq+Conv is better than Neighbor2Seq+Attn. I think there could be more information/experiment/insights added to explain why this is the case.\n\nWeakness:\n1. Since the proposed model is trying to mitigate the excessive computation and memory requirement issues, make it more efficient, and alleviate over-squashing, I'm expecting to see more in-depth analysis and empirical/theoretical justification with regard to those issues that the method is trying to solve. However, there's only some time complexity comparison and some argument about over-squashing. Can you provide the comparison of space complexity since memory requirement mitigation is one of your motivation? Also, Neighbor2Seq does not seem to improve upon the computation as compare with most of the methods listed in table 1? For the argument of alleviating over-squashing, can you provide more details? Why is converting to a sequence alleviating over-squashing? It seems that some sequence models also suffer from over-squashing, isn't it (trying to squash the whole sequence information into a single vector)? Please provide more explanation/analytical results, etc.\n2. The experiment results do not seem to improve a lot compared with existing models for some datasets and tasks. It seems that Neighbor2Seq+Conv consistently performs better than Neighbor2Seq+Attn in all experiment. Can you provide analysis of the number of parameters of your proposed model (2 variants) compared with existing models? Since people usually see that attention based approach performs better, is it because conv based one has more parameters. Does the proposed method have more parameters than existing models in general? Can you also provide some empirical comparison on the training time of your model compared with baseline models?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "need more clarification on the connection of the method and the issues it's trying to mitigate", "review": "The paper proposes a method called neighbor2seq that converts the hierarchical structure of the center node to a sequence during message passing in graph neural networks. The proposed method aims to mitigate the issue of excessive computation and memory requirement of training graph neural networks. The proposed models Neighbor2Seq+Conv and Neighbor2Seq+Attn are tested on several datasets including a large scale benchmark dataset (ogbn-papers100M). The result shows some improvement especially on ogbn-papers100M while the improvement is not very obvious on other datasets.\n\nStrength:\n1. The idea of converting the graph topology to a sequence that could be tackled with methods designed for grid-like data is good. \n2. The proposed method shows improvement on a large scale dataset. There are also ablation studies that are useful to understand different components in the method, for example, removing the sequence information (position encoding here) clearly hurts the model performance. But it's surprising to see that Neighbor2Seq+Conv is better than Neighbor2Seq+Attn. I think there could be more information/experiment/insights added to explain why this is the case.\n\nWeakness:\n1. Since the proposed model is trying to mitigate the excessive computation and memory requirement issues, make it more efficient, and alleviate over-squashing, I'm expecting to see more in-depth analysis and empirical/theoretical justification with regard to those issues that the method is trying to solve. However, there's only some time complexity comparison and some argument about over-squashing. Can you provide the comparison of space complexity since memory requirement mitigation is one of your motivation? Also, Neighbor2Seq does not seem to improve upon the computation as compare with most of the methods listed in table 1? For the argument of alleviating over-squashing, can you provide more details? Why is converting to a sequence alleviating over-squashing? It seems that some sequence models also suffer from over-squashing, isn't it (trying to squash the whole sequence information into a single vector)? Please provide more explanation/analytical results, etc.\n2. The experiment results do not seem to improve a lot compared with existing models for some datasets and tasks. It seems that Neighbor2Seq+Conv consistently performs better than Neighbor2Seq+Attn in all experiment. Can you provide analysis of the number of parameters of your proposed model (2 variants) compared with existing models? Since people usually see that attention based approach performs better, is it because conv based one has more parameters. Does the proposed method have more parameters than existing models in general? Can you also provide some empirical comparison on the training time of your model compared with baseline models?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603868378699}, {"id": "djSub53-m6g", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper383/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overall, the paper propose an interesting approach for computing node embeddings in a scalable way.\nHowever, the contribution is incremental as the idea of embedding nodes using sequences is not new; moreover, the discussion with prior works is very weak.\n\nConcretely, important related works are missing. \nThere is an ICML 2018 paper, \"Anonymous Walk Embeddings\" (https://arxiv.org/pdf/1805.11921.pdf). There are many following up papers as well. The idea of this line of works is to embedding node/graphs by random walk sequences. While the approaches are not exactly the same as  Neighbor2Seq, the idea of embedding nodes using sequences is not new. However, none of these works are mentioned in this paper. The authors should compare these papers in the experiments as the baselines. Without such comparison, it is hard to evaluate the performance gain of the proposed Neighbor2Seq.\n\nAdditionally, the discussion with prior works is very weak.\nThe paper misleads the readers by arguing \"However, they have inherent difficulties when applied on large graphs due to their excessive computation and memory requirements\" in Section 2.1. The paper refers \nHowever, existing approaches can scale to huge graphs. For example in the PinSage paper (https://arxiv.org/pdf/1806.01973.pdf), they can scale their GNN model to 3 billion nodes. Failing to mentioning this point is misleading for the readers, and is exaggerating the motivation of this paper.\n\nMore comments:\n1 The exact configuration of the Neighbor2Seq model is not mentioned. Providing an algorithm will greatly help.\n\n2 I can imagine multiple ways for sampling sequences from a node's neighborhood. However, they are not discussed in the paper.\n\n3 How does the expressive power of Neighbor2Seq compares with WL test? I assume Neighbor2Seq is theoretically less expressive than WL test. \n\n4 I guess there will be failure cases of this kind of Neighbor2Seq, when the topology of node neighborhood structure matters for the final performance. For example, Neighbor2Seq probably won't perform well for graph isomorphism tests. I suggest mentioning potential limitations in the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lacks discussions with prior works", "review": "Overall, the paper propose an interesting approach for computing node embeddings in a scalable way.\nHowever, the contribution is incremental as the idea of embedding nodes using sequences is not new; moreover, the discussion with prior works is very weak.\n\nConcretely, important related works are missing. \nThere is an ICML 2018 paper, \"Anonymous Walk Embeddings\" (https://arxiv.org/pdf/1805.11921.pdf). There are many following up papers as well. The idea of this line of works is to embedding node/graphs by random walk sequences. While the approaches are not exactly the same as  Neighbor2Seq, the idea of embedding nodes using sequences is not new. However, none of these works are mentioned in this paper. The authors should compare these papers in the experiments as the baselines. Without such comparison, it is hard to evaluate the performance gain of the proposed Neighbor2Seq.\n\nAdditionally, the discussion with prior works is very weak.\nThe paper misleads the readers by arguing \"However, they have inherent difficulties when applied on large graphs due to their excessive computation and memory requirements\" in Section 2.1. The paper refers \nHowever, existing approaches can scale to huge graphs. For example in the PinSage paper (https://arxiv.org/pdf/1806.01973.pdf), they can scale their GNN model to 3 billion nodes. Failing to mentioning this point is misleading for the readers, and is exaggerating the motivation of this paper.\n\nMore comments:\n1 The exact configuration of the Neighbor2Seq model is not mentioned. Providing an algorithm will greatly help.\n\n2 I can imagine multiple ways for sampling sequences from a node's neighborhood. However, they are not discussed in the paper.\n\n3 How does the expressive power of Neighbor2Seq compares with WL test? I assume Neighbor2Seq is theoretically less expressive than WL test. \n\n4 I guess there will be failure cases of this kind of Neighbor2Seq, when the topology of node neighborhood structure matters for the final performance. For example, Neighbor2Seq probably won't perform well for graph isomorphism tests. I suggest mentioning potential limitations in the paper.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603757846542}], "openreview_url": "https://openreview.net/forum?id=l3YcqzaPlx0", "arxiv_id": "2202.03341", "paper_pdf": "papers/l3YcqzaPlx0.pdf", "paper_pdf_sha256": "114e531c2632b2810e988fe59a24bf2c6adb977df5f362979ed2f92fc787cdfe", "paper_pdf_bytes": 306965, "paper_pdf_source": "openreview", "code_url": "https://github.com/divelab/Neighbor2Seq", "code_repository": "divelab/Neighbor2Seq", "code_commit": "327d73d872497eb3c7f70ee27a0ea696db30cb22", "code_archive": "repos/l3YcqzaPlx0.zip", "code_archive_sha256": "8af829b2c85caa513b307a8a0c69e5d0d106b8205ded88b04a0be19bbe367f81", "code_archive_bytes": 201897, "code_file_count": 8, "code_extensions": {".py": 7, ".sh": 1}, "github_disk_usage_kb": 206, "github_languages": {"Python": 36934, "Shell": 3874}, "github_archived": false, "github_pushed_at": "2022-02-08T03:20:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neighbor2seq-deep-learning-on-massive-graphs-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HklRKpEKDr", "year": 2020, "status": "rejected", "title": "Deep Coordination Graphs", "authors": ["Wendelin Boehmer", "Vitaly Kurin", "Shimon Whiteson"], "authorids": ["wendelin.boehmer@cs.ox.ac.uk", "vitaly.kurin@cs.ox.ac.uk", "shimon.whiteson@cs.ox.ac.uk"], "authors_source": "OpenReview API", "abstract": "This paper introduces the deep coordination graph (DCG) for collaborative multi-agent reinforcement learning. DCG strikes a flexible trade-off between representational capacity and generalization by factorizing the joint value function of all agents according to a coordination graph into payoffs between pairs of agents. The value can be maximized by local message passing along the graph, which allows training of the value function end-to-end with Q-learning. Payoff functions are approximated with deep neural networks and parameter sharing improves generalization over the state-action space. We show that DCG can solve challenging predator-prey tasks that are vulnerable to the relative overgeneralization pathology and in which all other known value factorization approaches fail.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rkgKE6E0FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper689/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary\nThis paper proposes a pairwise communication between agents using a shared neural network. The idea is to define the joint action-value function as the sum of individual agent's values + pair-wise payoff between agents, which is based on the prior work [Castellini et al.]. In particular, this paper proposes to share the parameters of the pairwise payoff function to improve efficiency. The result on a grid-world domain shows that the proposed method performs better than baselines that either do not have pairwise communication (VDN) or learn fully joint action value function (QTRAN). \n\n# Originality\nThe main novelty seems to be coming from the idea of parameter sharing between pairwise payoffs, but the overall architecture seems to be the same as [Castellini et al.]. \n\n# Quality\n- This paper is missing an important baseline which is DCG without parameter sharing. Given that parameter sharing is the main new component from [Castellini et al.], it would be important to show the benefit of parameter sharing.\n- Although the result against several baselines looks good, it would be much more convincing to show some qualitative analysis of the proposed method. For example, showing that the learned payoff captures reasonable and intuitive knowledge would strengthen the paper.\n\n# Clarity\n- The paper is well-written, and the figures are very clear. \n- It would be better to show a figure that illustrates the domain and task. \n\n# Significance\n- Although the paper presents a new idea very well, the overall idea seems a bit incremental. This paper overall looks like a straightforward extension of [Castellini et al.] by adding parameter sharing and evaluating it on a more complex domain. In addition, showing more in-depth analysis would make the paper stronger. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "# Summary\nThis paper proposes a pairwise communication between agents using a shared neural network. The idea is to define the joint action-value function as the sum of individual agent's values + pair-wise payoff between agents, which is based on the prior work [Castellini et al.]. In particular, this paper proposes to share the parameters of the pairwise payoff function to improve efficiency. The result on a grid-world domain shows that the proposed method performs better than baselines that either do not have pairwise communication (VDN) or learn fully joint action value function (QTRAN). \n\n# Originality\nThe main novelty seems to be coming from the idea of parameter sharing between pairwise payoffs, but the overall architecture seems to be the same as [Castellini et al.]. \n\n# Quality\n- This paper is missing an important baseline which is DCG without parameter sharing. Given that parameter sharing is the main new component from [Castellini et al.], it would be important to show the benefit of parameter sharing.\n- Although the result against several baselines looks good, it would be much more convincing to show some qualitative analysis of the proposed method. For example, showing that the learned payoff captures reasonable and intuitive knowledge would strengthen the paper.\n\n# Clarity\n- The paper is well-written, and the figures are very clear. \n- It would be better to show a figure that illustrates the domain and task. \n\n# Significance\n- Although the paper presents a new idea very well, the overall idea seems a bit incremental. This paper overall looks like a straightforward extension of [Castellini et al.] by adding parameter sharing and evaluating it on a more complex domain. In addition, showing more in-depth analysis would make the paper stronger. "}, "tcdate": 1571863856655}, {"id": "BJe9asNRYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper689/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. Summary\n\nTeh authors propose to learn value functions that are a sum of utility (single-agent) and payoff (2-agent) components. The weights between all function are shared (common RNN). This stands in contrast to VDNs (only uses utility functions or centralized value functions. The authors evaluate on predator-prey, where they argue that e.g. VDN fails to learn.\n\n(Note that other work has looked at reward shaping to learn *decentralized* agents for that problem.)\n\n1. Decision (accept or reject) with one or two key reasons for this choice.\n\nWeak reject.\n\n- The comparison between different topologies is nice, but implies that the structure of the graph has to be fixed manually. This seems to be a severe and unscalable constraint. It would be better if the authors would propose and evaluate a method to determine / learn what the right graph structure should be.\n- Weight sharing between the various agent components makes the problem closer to a single-agent problem. What happens if the agents are decentralized and the (shared-weight) pairwise functions are separate?\n- Authors only evaluate on a predator-prey problem.\n\n4. Supporting arguments\n\nN/A\n\n5. Additional feedback with the aim to improve the paper. Make it clear that these points are here to help, and not necessarily part of your decision assessment.\n\n6. Questions", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "1. Summary\n\nTeh authors propose to learn value functions that are a sum of utility (single-agent) and payoff (2-agent) components. The weights between all function are shared (common RNN). This stands in contrast to VDNs (only uses utility functions or centralized value functions. The authors evaluate on predator-prey, where they argue that e.g. VDN fails to learn.\n\n(Note that other work has looked at reward shaping to learn *decentralized* agents for that problem.)\n\n1. Decision (accept or reject) with one or two key reasons for this choice.\n\nWeak reject.\n\n- The comparison between different topologies is nice, but implies that the structure of the graph has to be fixed manually. This seems to be a severe and unscalable constraint. It would be better if the authors would propose and evaluate a method to determine / learn what the right graph structure should be.\n- Weight sharing between the various agent components makes the problem closer to a single-agent problem. What happens if the agents are decentralized and the (shared-weight) pairwise functions are separate?\n- Authors only evaluate on a predator-prey problem.\n\n4. Supporting arguments\n\nN/A\n\n5. Additional feedback with the aim to improve the paper. Make it clear that these points are here to help, and not necessarily part of your decision assessment.\n\n6. Questions"}, "tcdate": 1571863489521}, {"id": "HkxPtjf2FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper689/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces the deep coordination graph for collaborative multi-agent reinforcement learning aimed to solve predator-prey tasks by preventing relative overgeneralization during the exploration of agents. \n\nIn general, this paper gives a detailed and comprehensible depiction of the Introduction, Related work and Background section. However, I have two concerns about their method and experiments. \n\nThe presentation of the “Method” section is not clear enough to evaluate the contribution of this paper. Specifically: \n(1). In the \"Method\" section, your method incorporates three ideas: 1. restricting the payoffs; 2. sharing parameters; 3. allowing generalization; \nIn my understanding, both idea 1 and idea 2 come from VDN [4]. Idea 3 is not implemented in this work. Then, what is your contribution?\n\n(2). According to the announcement, the key benefit for your work is \"prevent relative overgeneralization during exploration of agents\". It is hard to access if the proposed method can prevent overgeneralization. Do you have some theoretical or empirical justifications?\n\n(3). I think it would be helpful if you could give a description according to your algorithm in the appendix. It would be better to draw a diagram to show what is your model.\n\nI think the experiments are too simple and not convincing. \n(1). I notice recent multi-agent reinforcement papers ([1], [2] closely related to your work) evaluate their work on the challenging set of StarCraft II micromanagement tasks and achieved the evident result. \n \n(2). In fig.2, why did you only compare your model with VDN [4]?\n\n(3). In fig.3, why does the return value of QTRAN first decrease and then increase? From [3], it seems their return curve continuously grows. \n\n(4). In fig.4, why QTRAN fail to this task?\n\n[1]. Rashid, Tabish, et al. \"QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning.\" arXiv preprint arXiv:1803.11485 (2018).\n[2]. Wang, Tonghan, et al. \"Learning Nearly Decomposable Value Functions Via Communication Minimization.\" arXiv preprint arXiv:1910.05366 (2019).\n[3]. Son, Kyunghwan, et al. \"QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning.\" arXiv preprint arXiv:1905.05408 (2019).\n[4]. Sunehag, Peter, et al. \"Value-decomposition networks for cooperative multi-agent learning.\" arXiv preprint arXiv:1706.05296 (2017).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "This paper introduces the deep coordination graph for collaborative multi-agent reinforcement learning aimed to solve predator-prey tasks by preventing relative overgeneralization during the exploration of agents. \n\nIn general, this paper gives a detailed and comprehensible depiction of the Introduction, Related work and Background section. However, I have two concerns about their method and experiments. \n\nThe presentation of the “Method” section is not clear enough to evaluate the contribution of this paper. Specifically: \n(1). In the \"Method\" section, your method incorporates three ideas: 1. restricting the payoffs; 2. sharing parameters; 3. allowing generalization; \nIn my understanding, both idea 1 and idea 2 come from VDN [4]. Idea 3 is not implemented in this work. Then, what is your contribution?\n\n(2). According to the announcement, the key benefit for your work is \"prevent relative overgeneralization during exploration of agents\". It is hard to access if the proposed method can prevent overgeneralization. Do you have some theoretical or empirical justifications?\n\n(3). I think it would be helpful if you could give a description according to your algorithm in the appendix. It would be better to draw a diagram to show what is your model.\n\nI think the experiments are too simple and not convincing. \n(1). I notice recent multi-agent reinforcement papers ([1], [2] closely related to your work) evaluate their work on the challenging set of StarCraft II micromanagement tasks and achieved the evident result. \n \n(2). In fig.2, why did you only compare your model with VDN [4]?\n\n(3). In fig.3, why does the return value of QTRAN first decrease and then increase? From [3], it seems their return curve continuously grows. \n\n(4). In fig.4, why QTRAN fail to this task?\n\n[1]. Rashid, Tabish, et al. \"QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning.\" arXiv preprint arXiv:1803.11485 (2018).\n[2]. Wang, Tonghan, et al. \"Learning Nearly Decomposable Value Functions Via Communication Minimization.\" arXiv preprint arXiv:1910.05366 (2019).\n[3]. Son, Kyunghwan, et al. \"QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning.\" arXiv preprint arXiv:1905.05408 (2019).\n[4]. Sunehag, Peter, et al. \"Value-decomposition networks for cooperative multi-agent learning.\" arXiv preprint arXiv:1706.05296 (2017).", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571724159415}], "openreview_url": "https://openreview.net/forum?id=HklRKpEKDr", "arxiv_id": "1910.00091", "paper_pdf": "papers/HklRKpEKDr.pdf", "paper_pdf_sha256": "35e1bbc38eb69de2fda18bff97e384b31a92cc5930124973a32b30eccee157a6", "paper_pdf_bytes": 609522, "paper_pdf_source": "openreview", "code_url": "https://github.com/wendelinboehmer/dcg", "code_repository": "wendelinboehmer/dcg", "code_commit": "4de100cddf7c3a7035cd89a47d7c1b8a878e7428", "code_archive": "repos/HklRKpEKDr.zip", "code_archive_sha256": "5fd82dadfa13181010958dfb3c6ab2ba068ab2db82bceb427f850ab45564a0cf", "code_archive_bytes": 263363, "code_file_count": 41, "code_extensions": {".py": 38, ".sh": 3}, "github_disk_usage_kb": 219, "github_languages": {"Python": 230756, "Dockerfile": 2141, "Shell": 1439}, "github_archived": false, "github_pushed_at": "2024-06-02T10:01:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-coordination-graphs"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1l8iiA9tQ", "year": 2019, "status": "rejected", "title": "Backdrop: Stochastic Backpropagation", "authors": ["Siavash Golkar", "Kyle Cranmer"], "authorids": ["siavash.golkar@gmail.com", "kyle.cranmer@nyu.edu"], "authors_source": "OpenReview API", "abstract": "We introduce backdrop, a flexible and simple-to-implement method, intuitively described as dropout acting only along the backpropagation pipeline. Backdrop is implemented via one or more masking layers which are inserted at specific points along the network. Each backdrop masking layer acts as the identity in the forward pass, but randomly masks parts of the backward gradient propagation.  Intuitively, inserting a backdrop layer after any  convolutional layer leads to stochastic  gradients corresponding to features of that scale.  Therefore, backdrop is well suited for problems in which the data have a multi-scale, hierarchical structure. Backdrop can also be applied to problems with non-decomposable loss functions where standard SGD methods are not well suited. We perform a number of experiments and demonstrate that backdrop leads to significant improvements in generalization.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "S1lt8Yuan7", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper624/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a stochastic based method, namely Backdrop, for updating the network structures via backpropagation type methods.  Backdrop inserts masking layers along the network; it acts as the identity in the forward pass, but as randomly masks parts of the backward gradient propagation. The paper claims this approach can significantly improves the overall generalization performance. \n\nAlthough some difference to Dropout is summarized in Section 2, I still feel these two methods have almost the same idea, with just different implementation. Actually this Backdrop seems to have one more limitation in the parameter complexity, as it introduces several mask layers but keep the dense structures from other intermediate layers. \n\nThe proposed Backdrop uses Bernoulli distribution to select active variables. This is the very fundamental way in the conventional Dropout method. On the other hand, the authors do not  provide convincing justification how this can guarantee the improvement in subsequent generalization. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposed Backdrop is similar to the traditional Dropout method. Overall this paper lacks of novelty and the observed generalization performance does not have convincing justification. ", "review": "This paper proposes a stochastic based method, namely Backdrop, for updating the network structures via backpropagation type methods.  Backdrop inserts masking layers along the network; it acts as the identity in the forward pass, but as randomly masks parts of the backward gradient propagation. The paper claims this approach can significantly improves the overall generalization performance. \n\nAlthough some difference to Dropout is summarized in Section 2, I still feel these two methods have almost the same idea, with just different implementation. Actually this Backdrop seems to have one more limitation in the parameter complexity, as it introduces several mask layers but keep the dense structures from other intermediate layers. \n\nThe proposed Backdrop uses Bernoulli distribution to select active variables. This is the very fundamental way in the conventional Dropout method. On the other hand, the authors do not  provide convincing justification how this can guarantee the improvement in subsequent generalization. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541405009030}, {"id": "Bke37P8q2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper624/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose to apply Dropout only in the backward pass, by applying a mask sampled from a Bernoulli distribution. They claim that this method can help in situations like optimizing non-decomposable losses where minibatch SGD is not viable. \n\nFirst and foremost, the paper has an acknowledgement paragraph that gives information violating, in my sense, the anonymity requirement. \n\nThis being said, I have other concerns with the paper, and this possible violation didn't effect much my rating. \n\nFirst, the authors claim that the proposed method \"is a flexible strategy for introducing data-dependent stochasticity into the gradient\". However, it doesn't seem to me that the sampled dropped nodes are data-dependent. \n\nIt is also not clear to me why the proposed method is better suited to non-decomposable losses and hierarchically structured data than the classical Dropout.\n\nMoreover, while the method is clearly related to Dropout, the paper lacks of comparison to this regularizer. \n\nThis being said, the idea is sound, and can have a good impact in for example combining the good aspects of batch-normalization and dropout. However, the authors structured the paper on a completely different argument that doesn't convince me for the reasons cited above.     ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Small modification and not enough comparison to other methods", "review": "The authors propose to apply Dropout only in the backward pass, by applying a mask sampled from a Bernoulli distribution. They claim that this method can help in situations like optimizing non-decomposable losses where minibatch SGD is not viable. \n\nFirst and foremost, the paper has an acknowledgement paragraph that gives information violating, in my sense, the anonymity requirement. \n\nThis being said, I have other concerns with the paper, and this possible violation didn't effect much my rating. \n\nFirst, the authors claim that the proposed method \"is a flexible strategy for introducing data-dependent stochasticity into the gradient\". However, it doesn't seem to me that the sampled dropped nodes are data-dependent. \n\nIt is also not clear to me why the proposed method is better suited to non-decomposable losses and hierarchically structured data than the classical Dropout.\n\nMoreover, while the method is clearly related to Dropout, the paper lacks of comparison to this regularizer. \n\nThis being said, the idea is sound, and can have a good impact in for example combining the good aspects of batch-normalization and dropout. However, the authors structured the paper on a completely different argument that doesn't convince me for the reasons cited above.     ", "rating": "3: Clear rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541199651825}, {"id": "HkgFP83Y3X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper624/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a data dependent strategy to mask parts of the partial derivatives in the chain rule computation. \n\nTypically with papers proposing modifications of the training regime of the neural network one would expect one of three outcomes:\n - a well justified, mathematically sound method, well tested in simple cases and with some proof of concept results on proper tasks\n - a more heuristic, empirical driven research, where strong results on proper tasks\n - method, however justified, allows us to do something previously impossible, removing some limitations/constraints (like biologically plausible learning etc.)\n\nIn its current form paper seems to lack any of these characteristics. On one hand method lacks any guarantees and on the other paper does not present significant improvements under any approved metrics, nor it introduces new which can be properly quantified. In fact, authors explicitly claim that empirical section \"Note that in these experiments, the purpose is not to achieve state of the art performance, but to exemplify how backdrop can be used and what measure of performance gains one can expect.\".  \n\nWith methods like this it is almost obvious that resulting update is not an unbiased gradient estimator of any function. Consequently convergence/learning guarantees that we have for GD or SGD no longer apply. Do authors have any thoughts on how bad can it get? As noted in the text, other methods of \"dropping\" data (such as dropout) don't have this issue as they still estimate proper gradients. Here, since dropping is done inside the network only on backwards pass, resulting estimates could, in principle, lead to oscilations, divergence and other issues. If these are not encountered in practice it might be interesting to understand why. \n\nIf authors prefer to go through more empirical path, one would expect at least to see some baselines for tasks proposed, rather than comparing Backdrop to SGD. There are many methods that could be applied in scenarios like this, including dozens forms of dropout (which, as authors note, is not aimed at the same goals, but this does not mean that it will not shine under the metrics introduced, as they are non-standard and so - noone tested them in this exact regime).\n\nI am happy to revisit my rating given authors restructure paper towards one of these paths (or other one which is not listed here).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, lacking proper improvements and/or applications provided", "review": "This paper introduces a data dependent strategy to mask parts of the partial derivatives in the chain rule computation. \n\nTypically with papers proposing modifications of the training regime of the neural network one would expect one of three outcomes:\n - a well justified, mathematically sound method, well tested in simple cases and with some proof of concept results on proper tasks\n - a more heuristic, empirical driven research, where strong results on proper tasks\n - method, however justified, allows us to do something previously impossible, removing some limitations/constraints (like biologically plausible learning etc.)\n\nIn its current form paper seems to lack any of these characteristics. On one hand method lacks any guarantees and on the other paper does not present significant improvements under any approved metrics, nor it introduces new which can be properly quantified. In fact, authors explicitly claim that empirical section \"Note that in these experiments, the purpose is not to achieve state of the art performance, but to exemplify how backdrop can be used and what measure of performance gains one can expect.\".  \n\nWith methods like this it is almost obvious that resulting update is not an unbiased gradient estimator of any function. Consequently convergence/learning guarantees that we have for GD or SGD no longer apply. Do authors have any thoughts on how bad can it get? As noted in the text, other methods of \"dropping\" data (such as dropout) don't have this issue as they still estimate proper gradients. Here, since dropping is done inside the network only on backwards pass, resulting estimates could, in principle, lead to oscilations, divergence and other issues. If these are not encountered in practice it might be interesting to understand why. \n\nIf authors prefer to go through more empirical path, one would expect at least to see some baselines for tasks proposed, rather than comparing Backdrop to SGD. There are many methods that could be applied in scenarios like this, including dozens forms of dropout (which, as authors note, is not aimed at the same goals, but this does not mean that it will not shine under the metrics introduced, as they are non-standard and so - noone tested them in this exact regime).\n\nI am happy to revisit my rating given authors restructure paper towards one of these paths (or other one which is not listed here).", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1541158497431}], "openreview_url": "https://openreview.net/forum?id=B1l8iiA9tQ", "arxiv_id": "1806.01337", "paper_pdf": "papers/B1l8iiA9tQ.pdf", "paper_pdf_sha256": "676b66edd712e4abfd38240ffa69a1a3674dde4cd57a14761171f7d60851f3c5", "paper_pdf_bytes": 3031581, "paper_pdf_source": "openreview", "code_url": "https://github.com/golkar/backdrop", "code_repository": "golkar/backdrop", "code_commit": "5e42191ec8a260d657a1dd29cebdae1870e215f4", "code_archive": "repos/B1l8iiA9tQ.zip", "code_archive_sha256": "c52d56a242c92f30b2ce1d9a78ce1a646dea991bac32c86f6257d0f1586f4ba8", "code_archive_bytes": 6660056, "code_file_count": 4, "code_extensions": {".py": 2, ".ipynb": 2}, "github_disk_usage_kb": 7059, "github_languages": {"Jupyter Notebook": 4601766, "Python": 24685}, "github_archived": false, "github_pushed_at": "2018-06-06T13:40:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/backdrop-stochastic-backpropagation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LQ1qYDOk6o", "year": 2026, "status": "rejected", "title": "ReCalKV: Low-Rank KV Cache Compression via Head Reordering and Offline Calibration", "authors": ["Xianglong Yan", "Zhiteng Li", "Tianao Zhang", "Haotong Qin", "Linghe Kong", "Yulun Zhang", "Xiaokang Yang"], "authorids": ["~Xianglong_Yan2", "~Zhiteng_Li2", "~Tianao_Zhang1", "~Haotong_Qin1", "~Linghe_Kong1", "~Yulun_Zhang1", "~Xiaokang_Yang1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have demonstrated remarkable performance, but their long-context reasoning remains constrained by the excessive memory required for the Key-Value (KV) cache. This makes KV cache compression a critical step toward efficient long-context inference. Recent methods have explored low-rank techniques to reduce the hidden size of the KV cache. However, they neglect the distinct roles and varying importance of Keys and Values, leading to significant performance drops under high compression. To address this, we propose ReCalKV, a post-training low-rank KV cache compression approach with tailored strategies for Keys and Values. For Keys, we propose Head-wise Similarity–aware Reordering (HSR), which clusters structurally similar heads into groups, enabling more accurate low-rank approximation via grouped SVD. For Values, we propose Offline Value Calibration (OVC), which efficiently calibrates the value projection matrix using calibration data without training, ensuring an accurate representation of contextual information. Extensive experiments show that ReCalKV consistently outperforms existing low-rank compression methods, achieving high compression ratios with minimal performance loss. We will release all the code and models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "S7HISFRC1h", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22685/Reviewer_rgUM"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces a novel low-rank KV cache compression method, building on PaLU, a prior work that decomposes the KV projection matrix using SVD to reduce the dimension of the KV cache. The paper makes the following contributions on top of PaLU:\n- Reordering the key projection matrix to achieve a better SVD decomposition.\n- Refining the low-rank approximation of the value projection matrix using a calibration dataset.", "review_text": "The paper introduces a novel low-rank KV cache compression method, building on PaLU, a prior work that decomposes the KV projection matrix using SVD to reduce the dimension of the KV cache. The paper makes the following contributions on top of PaLU:\n- Reordering the key projection matrix to achieve a better SVD decomposition.\n- Refining the low-rank approximation of the value projection matrix using a calibration dataset.", "strengths": "- Strong improvements over PaLU on the evaluated models.\n- Comprehensive ablation studies demonstrating the effectiveness of each contribution for both key and value projections.", "weaknesses": "**[W1]** The evaluated models are outdated. I suggest moving the results on the Llama-3.1 model to the main body. This is important because many modern LLM architectures employ GQA, while only the Mistral model from the main results section does so. Validation on multiple models with GQA would strengthen the paper.\n\n**[W2]** Comments on writing:\n- *L55–60:* This is not something revealed by your analysis.\n- *L60–63:* I cannot find a section describing your analysis of Fisher information.\n- The fact that whitening is applied before SVD should be discussed earlier in the Methods section, with more detail.\n\nMinor comments that did not affect the score:\n- *L69:* “Offline Calibration Value” --> “Offline Value Calibration”", "questions": "**[Q1]** Is there a reason why offline calibration is applied only to the value projection matrices? Could this also be applied to the key projection matrices?\n\n**[Q2]** How effective is the proposed method in terms of the memory–accuracy trade-off compared to other KV cache compression methods beyond those based on SVD of the projection matrices?\n\n**[Q3]** How would the method perform for reasoning models such as the Qwen3 model family on long generation tasks like AIME or LiveCodeBench? Demonstrating this would highlight the method’s robustness under long-generation scenarios, which are not captured by perplexity or long-context retrieval tasks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel low-rank KV cache compression method, building on PaLU, a prior work that decomposes the KV projection matrix using SVD to reduce the dimension of the KV cache. The paper makes the following contributions on top of PaLU:\n- Reordering the key projection matrix to achieve a better SVD decomposition.\n- Refining the low-rank approximation of the value projection matrix using a calibration dataset.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Strong improvements over PaLU on the evaluated models.\n- Comprehensive ablation studies demonstrating the effectiveness of each contribution for both key and value projections.", "weaknesses": "**[W1]** The evaluated models are outdated. I suggest moving the results on the Llama-3.1 model to the main body. This is important because many modern LLM architectures employ GQA, while only the Mistral model from the main results section does so. Validation on multiple models with GQA would strengthen the paper.\n\n**[W2]** Comments on writing:\n- *L55–60:* This is not something revealed by your analysis.\n- *L60–63:* I cannot find a section describing your analysis of Fisher information.\n- The fact that whitening is applied before SVD should be discussed earlier in the Methods section, with more detail.\n\nMinor comments that did not affect the score:\n- *L69:* “Offline Calibration Value” --> “Offline Value Calibration”", "questions": "**[Q1]** Is there a reason why offline calibration is applied only to the value projection matrices? Could this also be applied to the key projection matrices?\n\n**[Q2]** How effective is the proposed method in terms of the memory–accuracy trade-off compared to other KV cache compression methods beyond those based on SVD of the projection matrices?\n\n**[Q3]** How would the method perform for reasoning models such as the Qwen3 model family on long generation tasks like AIME or LiveCodeBench? Demonstrating this would highlight the method’s robustness under long-generation scenarios, which are not captured by perplexity or long-context retrieval tasks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761966147833}, {"id": "Jrt9sYWkMX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22685/Reviewer_akiT"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper introduces ReCalKV, a post-training framework for low-rank KV cache compression that treats Keys and Values separately. It enhances Key approximation  through similarity-based head grouping and decomposition and refines Value through lightweight calibration and fusion. Experiments demonstrate that ReCalKV consistently outperforms prior methods under high compression rates.", "review_text": "This paper introduces ReCalKV, a post-training framework for low-rank KV cache compression that treats Keys and Values separately. It enhances Key approximation  through similarity-based head grouping and decomposition and refines Value through lightweight calibration and fusion. Experiments demonstrate that ReCalKV consistently outperforms prior methods under high compression rates.", "strengths": "1. The paper identifies and analyzes the asymmetric roles of Keys and Values, particularly emphasizing that individual attention heads differ in information content. Using CKA-based head reordering before SVD to minimize approximation error is a well-motivated and conceptually sound idea.\n\n2. Across multiple model families and compression ratios, ReCalKV demonstrates competitive or superior results compared with the main low-rank baseline (Palu). The method maintains high accuracy even under aggressive compression and shows compatibility with quantization.", "weaknesses": "1. The paper mainly compares with low-rank SVD-based approaches such as Palu, but lacks comparisons with other classes of KV cache compression methods (e.g., KIVI, KVQuant, or token eviction approaches).\nAs a result, the reader cannot fully assess how ReCalKV performs in a broader landscape of KV compression techniques — especially when low-rank compression is not necessarily the only or best strategy.\n\n2. Experiments focus on older LLaMA/Mistral models, with limited evaluation on recent architectures or larger scales. Since the method relies on specific structural properties of attention heads (CKA similarity patterns), it's unclear whether these properties generalize across diverse modern architectures and model scales beyond the tested family.\n\n3. While latency speedups are reported, the computational cost of online head reordering during inference is not quantified separately. The reliance on custom Triton kernels also raises questions about achievability with standard inference frameworks, limiting practical deployment insights.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ReCalKV, a post-training framework for low-rank KV cache compression that treats Keys and Values separately. It enhances Key approximation  through similarity-based head grouping and decomposition and refines Value through lightweight calibration and fusion. Experiments demonstrate that ReCalKV consistently outperforms prior methods under high compression rates.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The paper identifies and analyzes the asymmetric roles of Keys and Values, particularly emphasizing that individual attention heads differ in information content. Using CKA-based head reordering before SVD to minimize approximation error is a well-motivated and conceptually sound idea.\n\n2. Across multiple model families and compression ratios, ReCalKV demonstrates competitive or superior results compared with the main low-rank baseline (Palu). The method maintains high accuracy even under aggressive compression and shows compatibility with quantization.", "weaknesses": "1. The paper mainly compares with low-rank SVD-based approaches such as Palu, but lacks comparisons with other classes of KV cache compression methods (e.g., KIVI, KVQuant, or token eviction approaches).\nAs a result, the reader cannot fully assess how ReCalKV performs in a broader landscape of KV compression techniques — especially when low-rank compression is not necessarily the only or best strategy.\n\n2. Experiments focus on older LLaMA/Mistral models, with limited evaluation on recent architectures or larger scales. Since the method relies on specific structural properties of attention heads (CKA similarity patterns), it's unclear whether these properties generalize across diverse modern architectures and model scales beyond the tested family.\n\n3. While latency speedups are reported, the computational cost of online head reordering during inference is not quantified separately. The reliance on custom Triton kernels also raises questions about achievability with standard inference frameworks, limiting practical deployment insights.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761927244569}, {"id": "duG3AKGrTD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22685/Reviewer_5Foc"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces ReCalKV, a post-training framework for compressing the Key-Value (KV) cache in large language models (LLMs) by reducing the hidden dimension via low-rank approximations. It proposes asymmetric strategies: Head-wise Similarity-aware Reordering (HSR) for Keys, which reorders and groups attention heads based on Centered Kernel Alignment (CKA) similarity to enable more accurate grouped Singular Value Decomposition (SVD); and Offline Value Calibration (OVC) for Values, which calibrates the decomposed SVD matrices using a small dataset and fuses the right factor into the output projection to eliminate runtime reconstruction overhead. Extensive experiments on LLaMA and Mistral models demonstrate superior perplexity, zero-shot QA accuracy, and long-context performance compared to baselines like Palu, with minimal degradation (e.g., ~2% relative accuracy drop at 50% compression) and compatibility with quantization for higher ratios.", "review_text": "The paper introduces ReCalKV, a post-training framework for compressing the Key-Value (KV) cache in large language models (LLMs) by reducing the hidden dimension via low-rank approximations. It proposes asymmetric strategies: Head-wise Similarity-aware Reordering (HSR) for Keys, which reorders and groups attention heads based on Centered Kernel Alignment (CKA) similarity to enable more accurate grouped Singular Value Decomposition (SVD); and Offline Value Calibration (OVC) for Values, which calibrates the decomposed SVD matrices using a small dataset and fuses the right factor into the output projection to eliminate runtime reconstruction overhead. Extensive experiments on LLaMA and Mistral models demonstrate superior perplexity, zero-shot QA accuracy, and long-context performance compared to baselines like Palu, with minimal degradation (e.g., ~2% relative accuracy drop at 50% compression) and compatibility with quantization for higher ratios.", "strengths": "* K-side uses CKA-guided head reordering + grouped SVD to share low-rank factors among similar heads (greedy pairing; fixed group size), and V-side uses closed-form offline calibration to minimize projection error on a small calibration set.  \n\n* Matrix fusion folds (R_v) into (W_o), eliminating online reconstruction and avoiding extra inference ops; the end-to-end procedure is fully post-training/offline (Algorithm 1).  \n\n* Strong ablations isolate HSR and OVC and show they are complementary at fixed compression (Table 3).  \n\n* Evaluations span multiple model families and tasks, plus quantization compatibility (3–4-bit) demonstrating orthogonality to per-token KV quantization (Table 4).  \n\n* Figures 2–3 make the reordering/grouped-SVD mechanism concrete; Algorithm 1 spells out the pipeline; equations (9–11) specify the fused-inference path.  \n\n* Targets a real deployment bottleneck (KV memory/latency) with demonstrable inference efficiency improvements on long contexts, while remaining compatible with common compression stacks.", "weaknesses": "- Code not provided, therefore it's not reproducible as is.\n- While the method is effective, baselines are limited primarily to Palu (G-LRD), lacking comparisons with recent variants like CommonKV or FDC, which could better substantiate SOTA claims (section 4).\n- Experiments do not quantify runtime overhead from Key reconstruction post-HSR (Figure 3), despite claims of low cost; real-world latency measurements on diverse hardware would strengthen efficiency arguments (Figure 4). \n- Equations (7) and (8) for OVC appear to have typos in transposes and do not explicitly state assumptions (e.g., whitening) needed for the closed forms.", "questions": "- Please address the items mentioned under Weaknesses. For example, lack of reproducibility.\n- In section 3.3 (lines 216-269), the OVC calibration uses equations (7) and (8) with a small dataset X (256 WikiText2 samples; section 4.1). How sensitive is performance to the size and domain of X? Could you provide perplexity results on WikiText2 for LLaMA-2-7B at 50% compression using 128 vs. 512 samples, or a different domain like C4?\n- Section 3.2 describes HSR as greedy grouping based on the CKA similarity matrix S (Eq. 5) with a fixed group size (e.g., 4 heads per group when  h=32). Table 3 shows that at 80% compression, HSR+OVC attains 8.48 perplexity on WikiText-2. What is the effect of the HSR group size s on performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ReCalKV, a post-training framework for compressing the Key-Value (KV) cache in large language models (LLMs) by reducing the hidden dimension via low-rank approximations. It proposes asymmetric strategies: Head-wise Similarity-aware Reordering (HSR) for Keys, which reorders and groups attention heads based on Centered Kernel Alignment (CKA) similarity to enable more accurate grouped Singular Value Decomposition (SVD); and Offline Value Calibration (OVC) for Values, which calibrates the decomposed SVD matrices using a small dataset and fuses the right factor into the output projection to eliminate runtime reconstruction overhead. Extensive experiments on LLaMA and Mistral models demonstrate superior perplexity, zero-shot QA accuracy, and long-context performance compared to baselines like Palu, with minimal degradation (e.g., ~2% relative accuracy drop at 50% compression) and compatibility with quantization for higher ratios.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* K-side uses CKA-guided head reordering + grouped SVD to share low-rank factors among similar heads (greedy pairing; fixed group size), and V-side uses closed-form offline calibration to minimize projection error on a small calibration set.  \n\n* Matrix fusion folds (R_v) into (W_o), eliminating online reconstruction and avoiding extra inference ops; the end-to-end procedure is fully post-training/offline (Algorithm 1).  \n\n* Strong ablations isolate HSR and OVC and show they are complementary at fixed compression (Table 3).  \n\n* Evaluations span multiple model families and tasks, plus quantization compatibility (3–4-bit) demonstrating orthogonality to per-token KV quantization (Table 4).  \n\n* Figures 2–3 make the reordering/grouped-SVD mechanism concrete; Algorithm 1 spells out the pipeline; equations (9–11) specify the fused-inference path.  \n\n* Targets a real deployment bottleneck (KV memory/latency) with demonstrable inference efficiency improvements on long contexts, while remaining compatible with common compression stacks.", "weaknesses": "- Code not provided, therefore it's not reproducible as is.\n- While the method is effective, baselines are limited primarily to Palu (G-LRD), lacking comparisons with recent variants like CommonKV or FDC, which could better substantiate SOTA claims (section 4).\n- Experiments do not quantify runtime overhead from Key reconstruction post-HSR (Figure 3), despite claims of low cost; real-world latency measurements on diverse hardware would strengthen efficiency arguments (Figure 4). \n- Equations (7) and (8) for OVC appear to have typos in transposes and do not explicitly state assumptions (e.g., whitening) needed for the closed forms.", "questions": "- Please address the items mentioned under Weaknesses. For example, lack of reproducibility.\n- In section 3.3 (lines 216-269), the OVC calibration uses equations (7) and (8) with a small dataset X (256 WikiText2 samples; section 4.1). How sensitive is performance to the size and domain of X? Could you provide perplexity results on WikiText2 for LLaMA-2-7B at 50% compression using 128 vs. 512 samples, or a different domain like C4?\n- Section 3.2 describes HSR as greedy grouping based on the CKA similarity matrix S (Eq. 5) with a fixed group size (e.g., 4 heads per group when  h=32). Table 3 shows that at 80% compression, HSR+OVC attains 8.48 perplexity on WikiText-2. What is the effect of the HSR group size s on performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761878271654}, {"id": "XJAgf5GTTZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22685/Reviewer_yTbf"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper addresses the problem of high memory and bandwidth overhead caused by the KV cache in LLM during long-context during. This paper proposes ReCalKV, a post-training low-rank KV cache compression method that introduces differentiated strategies. it applies Head-wise Similarity-aware Reordering to cluster structurally similar attention heads before grouped SVD for Key compression, and Offline Value Calibration to recalibrate Value projection matrices using small calibration datasets. \nThe experimental results demonstrate that ReCalKV achieves consistent improvements over prior low-rank compression baselines （Palu and LoRC). It maintains competitive perplexity and zero-shot accuracy under 50–70% KV-cache compression ratios on models including LLaMA-2-7B, Mistral-7B, and LongChat-7B. However, all experiments are conducted on relatively older architectures, and no evaluation is reported on Llama-3 or Qwen-3.", "review_text": "This paper addresses the problem of high memory and bandwidth overhead caused by the KV cache in LLM during long-context during. This paper proposes ReCalKV, a post-training low-rank KV cache compression method that introduces differentiated strategies. it applies Head-wise Similarity-aware Reordering to cluster structurally similar attention heads before grouped SVD for Key compression, and Offline Value Calibration to recalibrate Value projection matrices using small calibration datasets. \nThe experimental results demonstrate that ReCalKV achieves consistent improvements over prior low-rank compression baselines （Palu and LoRC). It maintains competitive perplexity and zero-shot accuracy under 50–70% KV-cache compression ratios on models including LLaMA-2-7B, Mistral-7B, and LongChat-7B. However, all experiments are conducted on relatively older architectures, and no evaluation is reported on Llama-3 or Qwen-3.", "strengths": "1. The paper tackles an important and practical problem—reducing KV-cache memory overhead for long-context LLM inference, which remains a key bottleneck for efficient deployment.\n2. The proposed approach is model-agnostic and can be readily applied to various Transformer architectures without retraining, showing potential for integration into large-scale serving systems.", "weaknesses": "1. The reported experimental performance, while better than earlier SVD-based baselines, remains clearly inferior to recent quantization-based methods such as KVQuant and AnTKV, which achieve much lower perplexity under similar even higher compression ratios.\n2. ReCalKV still introduces additional computations for restruct KV using low rank kv (compute with R_k and R_v) during each decoding step. I recommend the authors evaluate latency and accuracy on end-to-end tasks such as AIME.\n3. The evaluation focuses mainly on outdated models (e.g., LLaMA-2, Mistral-7B) and lacks results on modern architectures like LLaMA-3 or Qwen-3, making it difficult to assess real-world relevance.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of high memory and bandwidth overhead caused by the KV cache in LLM during long-context during. This paper proposes ReCalKV, a post-training low-rank KV cache compression method that introduces differentiated strategies. it applies Head-wise Similarity-aware Reordering to cluster structurally similar attention heads before grouped SVD for Key compression, and Offline Value Calibration to recalibrate Value projection matrices using small calibration datasets. \nThe experimental results demonstrate that ReCalKV achieves consistent improvements over prior low-rank compression baselines （Palu and LoRC). It maintains competitive perplexity and zero-shot accuracy under 50–70% KV-cache compression ratios on models including LLaMA-2-7B, Mistral-7B, and LongChat-7B. However, all experiments are conducted on relatively older architectures, and no evaluation is reported on Llama-3 or Qwen-3.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper tackles an important and practical problem—reducing KV-cache memory overhead for long-context LLM inference, which remains a key bottleneck for efficient deployment.\n2. The proposed approach is model-agnostic and can be readily applied to various Transformer architectures without retraining, showing potential for integration into large-scale serving systems.", "weaknesses": "1. The reported experimental performance, while better than earlier SVD-based baselines, remains clearly inferior to recent quantization-based methods such as KVQuant and AnTKV, which achieve much lower perplexity under similar even higher compression ratios.\n2. ReCalKV still introduces additional computations for restruct KV using low rank kv (compute with R_k and R_v) during each decoding step. I recommend the authors evaluate latency and accuracy on end-to-end tasks such as AIME.\n3. The evaluation focuses mainly on outdated models (e.g., LLaMA-2, Mistral-7B) and lacks results on modern architectures like LLaMA-3 or Qwen-3, making it difficult to assess real-world relevance.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761378112568}], "openreview_url": "https://openreview.net/forum?id=LQ1qYDOk6o", "arxiv_id": "2505.24357", "paper_pdf": "papers/LQ1qYDOk6o.pdf", "paper_pdf_sha256": "6436531a3dddfd3d77d201fafe88a60cd00460a6e90bc255e9c7b1954d2e1312", "paper_pdf_bytes": 489009, "paper_pdf_source": "openreview", "code_url": "https://github.com/XIANGLONGYAN/ReCalKV", "code_repository": "XIANGLONGYAN/ReCalKV", "code_commit": "ea548d5c6c648316f009770a2e54b80fa16c6b06", "code_archive": "repos/LQ1qYDOk6o.zip", "code_archive_sha256": "64dd664dd5c750627d57b988c413040b91dce4e8017754a54a9282b6efeb15ba", "code_archive_bytes": 319716, "code_file_count": 36, "code_extensions": {".py": 35, ".sh": 1}, "github_disk_usage_kb": 1250, "github_languages": {"Python": 1064261, "Shell": 3287}, "github_archived": false, "github_pushed_at": "2026-07-09T07:41:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recalkv-low-rank-kv-cache-compression-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DfTWrTwLzD", "year": 2025, "status": "rejected", "title": "Two Are Better than One: Context Window Extension with Multi-Grained Self-Injection", "authors": ["Wei Han", "Pan Zhou", "Soujanya Poria", "Shuicheng YAN"], "authorids": ["~Wei_Han4", "~Pan_Zhou3", "~Soujanya_Poria1", "~Shuicheng_YAN3"], "authors_source": "OpenReview API", "abstract": "Limited longtext window has been an inherent constraint for large language models (LLMs), which significantly restricts their application scenarios.  Continual pre-training on long-context data is the most straightforward approach to further extend an LLM's context window, but it is at the expense of huge data acquisition and computation cost. \nThere are many cost-efficient context window extension methods which do not require pretraining process emerges as appealing solutions, such as extrapolation, attention manipulation, context compression, etc.\nIn this paper, we propose a novel approach named Shared-LLaMA. \nShared-LLaMA is composed of two short-context LLMs. \nOne of them works as compressor and the other works as decoder. \nThe decoder receives compressed multi-grained context information from \nthe compressor and performs context-aware modeling on the running text. \nInformation transfer between the compressor and decoder occurs only at the lowest layers to circumvent an entire forward pass and save the inference time. \nBoth LLMs are initialized from the same off-the-shelf checkpoint and thus can be directly trained without extra feature alignment stages.\nAdditionally, we propose a tree structure to store the multi-grained information and design a search algorithm to fast locate and retrieve related information from each level of that tree. \nWith these efficient design choices, Shared-LLaMA can greatly reduce memory consumption, and achieves apparent speed up over other advanced baselines (2$\\times$ over streaming, $3\\times$ over encoder-decoder architectures).\nIn our evaluation on long-context modeling and understanding tasks, Shared-LLaMA yields superior or comparable results to several strong baselines, indicating Shared-LLaMA achieves a good balance between efficiency and effectiveness.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "n5ifuAKTBL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5960/Reviewer_toYK"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper introduces SharedLLM, a novel approach for extending the context window of large language models (LLMs) by using a hierarchical architecture that pairs two short-context LLMs. In SharedLLM, one model, called the lower model, acts as a compressor that processes past context into compact, multi-grained representations. The other, the upper model, serves as a decoder that integrates this compressed context with current text to predict future tokens in a context-aware manner. Information is passed from the compressor to the decoder through self-injection layers at specific levels, allowing efficient integration without extensive cross-attention.", "review_text": "The paper introduces SharedLLM, a novel approach for extending the context window of large language models (LLMs) by using a hierarchical architecture that pairs two short-context LLMs. In SharedLLM, one model, called the lower model, acts as a compressor that processes past context into compact, multi-grained representations. The other, the upper model, serves as a decoder that integrates this compressed context with current text to predict future tokens in a context-aware manner. Information is passed from the compressor to the decoder through self-injection layers at specific levels, allowing efficient integration without extensive cross-attention.", "strengths": "1. Innovative Multi-Grained Context Extension: Introduces a unique approach to extend context windows using a compressor-decoder model architecture, which efficiently handles large context data.\n2. Strong Experimental Results: Demonstrates superior performance on several long-context benchmarks, providing evidence of the model's robustness.\n3. High Efficiency: Outperforms other methods in terms of speed and memory usage, making SharedLLM viable for large-scale applications.", "weaknesses": "1. The paper provides little information on how the model’s performance changes with different tree depths, compression ratios, and injection layers beyond the default settings. Since these parameters are key to achieving a balance between efficiency and accuracy, a sensitivity analysis would be beneficial.\n2. While the paper introduces a query-aware retrieval policy in the context tree for efficient information extraction, it lacks a detailed analysis of how different retrieval policies affect SharedLLM’s performance. An ablation study comparing retrieval policies (e.g., different similarity metrics or selection thresholds) would enhance understanding and offer actionable tuning guidance for practitioners.", "questions": "1. What criteria guided the choice of retrieval policy in the context tree, and how sensitive is SharedLLM’s performance to different retrieval policy settings?\n2. How does SharedLLM perform with different compression ratios, tree depths, and injection layer settings?\n3. Have you considered testing SharedLLM with alternative context extension methods, such as position interpolation or memory-augmented architectures?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SharedLLM, a novel approach for extending the context window of large language models (LLMs) by using a hierarchical architecture that pairs two short-context LLMs. In SharedLLM, one model, called the lower model, acts as a compressor that processes past context into compact, multi-grained representations. The other, the upper model, serves as a decoder that integrates this compressed context with current text to predict future tokens in a context-aware manner. Information is passed from the compressor to the decoder through self-injection layers at specific levels, allowing efficient integration without extensive cross-attention.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Innovative Multi-Grained Context Extension: Introduces a unique approach to extend context windows using a compressor-decoder model architecture, which efficiently handles large context data.\n2. Strong Experimental Results: Demonstrates superior performance on several long-context benchmarks, providing evidence of the model's robustness.\n3. High Efficiency: Outperforms other methods in terms of speed and memory usage, making SharedLLM viable for large-scale applications.", "weaknesses": "1. The paper provides little information on how the model’s performance changes with different tree depths, compression ratios, and injection layers beyond the default settings. Since these parameters are key to achieving a balance between efficiency and accuracy, a sensitivity analysis would be beneficial.\n2. While the paper introduces a query-aware retrieval policy in the context tree for efficient information extraction, it lacks a detailed analysis of how different retrieval policies affect SharedLLM’s performance. An ablation study comparing retrieval policies (e.g., different similarity metrics or selection thresholds) would enhance understanding and offer actionable tuning guidance for practitioners.", "questions": "1. What criteria guided the choice of retrieval policy in the context tree, and how sensitive is SharedLLM’s performance to different retrieval policy settings?\n2. How does SharedLLM perform with different compression ratios, tree depths, and injection layer settings?\n3. Have you considered testing SharedLLM with alternative context extension methods, such as position interpolation or memory-augmented architectures?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730735149337}, {"id": "PrmOJ8fyYB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5960/Reviewer_wFyt"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper presents SharedLLM, which uses two short-context LLMs (derived from the same model family like LLaMA-2) in a hierarchical structure to handle long contexts. My understanding is that this paper has made the following contribution:\n\n- Integrating the concept of \"Context Tree\": A binary tree structure storing text at different granularities, with higher compression ratios at higher levels\n\n- Architecture: A lower model compresses context into multi-grained representations, while an upper model performs language modeling using this compressed information\n\n- Query-Aware Retrieval: For instruction-following tasks, uses similarity scoring to selectively expand relevant tree nodes\n\n- Layer-wise Connection: Information transfer occurs only at lower layers to reduce computational overhead\n\nEmpirical results on language modeling tasks up to 128K tokens and various instruction-following benchmarks.\n\nHowever, I am a bit concerned with the presentation and also the technical depth explanation.", "review_text": "The paper presents SharedLLM, which uses two short-context LLMs (derived from the same model family like LLaMA-2) in a hierarchical structure to handle long contexts. My understanding is that this paper has made the following contribution:\n\n- Integrating the concept of \"Context Tree\": A binary tree structure storing text at different granularities, with higher compression ratios at higher levels\n\n- Architecture: A lower model compresses context into multi-grained representations, while an upper model performs language modeling using this compressed information\n\n- Query-Aware Retrieval: For instruction-following tasks, uses similarity scoring to selectively expand relevant tree nodes\n\n- Layer-wise Connection: Information transfer occurs only at lower layers to reduce computational overhead\n\nEmpirical results on language modeling tasks up to 128K tokens and various instruction-following benchmarks.\n\nHowever, I am a bit concerned with the presentation and also the technical depth explanation.", "strengths": "- the basic empirical setup for evaluation in experimental studies is clean on standard benchmarks, from ppl to instruction-following tasks\n\n- ablation studies demonstrate the effectiveness of the approach", "weaknesses": "1. In my personal opinon, this paper could benefit from the improving the presentation, specifically,\n\n- In figure 1, it is hard for me to capture whether this is a encoder-decoder model or decoder-only model. The role of cross attention is not very clear. I would recommend authors to haven an overview of the model, and then dive into details, and use another figure to explain the context tree to avoid readers getting distracted. At least, you can mark which part is an encoder, and which part is a decoder.\n\n- I really find it hard to understand what is \"Tree Cross-attention\" in Figure 1.\n\n- The paper emphasizes on \"self-injection\" but this concept is nowhere in this central figure. \n\n- there are notations of compression ratio but I am not so sure that they are well explained.\n\n- Authors were arguing \"information preservation\" but I am not so sure what does this mean?could you elaborate on this and connect this with related work?\n\n2. Experimental setup and results\n\n- There are many hyperparameter setup of the trees constructed, but the intuitive why those hyperparameters were used/set are not well discussed.  See Q1\n\n- What do you handle the drift or variable length during training and inference? See Q2\n\n- Different values of M were set, what's the intuition and what's the best practice? See Q3", "questions": "Q1. How do you setup the values of hyperparameters of context tree? for example, their depth? are they sensitive to the inference tasks?\n\nQ2. How do you take care of variable length during training and inference? The dynamic NTK and Yarn used the inference-time ratio on this, but I am not so sure about this in your method, thank you in advance as I am out of curiosity. \n\nQ3. Robustness of different M values. M is an important model architecture base value, but I am not sure the robustness and meaning of setting different M values, and what is the best practice for this?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents SharedLLM, which uses two short-context LLMs (derived from the same model family like LLaMA-2) in a hierarchical structure to handle long contexts. My understanding is that this paper has made the following contribution:\n\n- Integrating the concept of \"Context Tree\": A binary tree structure storing text at different granularities, with higher compression ratios at higher levels\n\n- Architecture: A lower model compresses context into multi-grained representations, while an upper model performs language modeling using this compressed information\n\n- Query-Aware Retrieval: For instruction-following tasks, uses similarity scoring to selectively expand relevant tree nodes\n\n- Layer-wise Connection: Information transfer occurs only at lower layers to reduce computational overhead\n\nEmpirical results on language modeling tasks up to 128K tokens and various instruction-following benchmarks.\n\nHowever, I am a bit concerned with the presentation and also the technical depth explanation.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- the basic empirical setup for evaluation in experimental studies is clean on standard benchmarks, from ppl to instruction-following tasks\n\n- ablation studies demonstrate the effectiveness of the approach", "weaknesses": "1. In my personal opinon, this paper could benefit from the improving the presentation, specifically,\n\n- In figure 1, it is hard for me to capture whether this is a encoder-decoder model or decoder-only model. The role of cross attention is not very clear. I would recommend authors to haven an overview of the model, and then dive into details, and use another figure to explain the context tree to avoid readers getting distracted. At least, you can mark which part is an encoder, and which part is a decoder.\n\n- I really find it hard to understand what is \"Tree Cross-attention\" in Figure 1.\n\n- The paper emphasizes on \"self-injection\" but this concept is nowhere in this central figure. \n\n- there are notations of compression ratio but I am not so sure that they are well explained.\n\n- Authors were arguing \"information preservation\" but I am not so sure what does this mean?could you elaborate on this and connect this with related work?\n\n2. Experimental setup and results\n\n- There are many hyperparameter setup of the trees constructed, but the intuitive why those hyperparameters were used/set are not well discussed.  See Q1\n\n- What do you handle the drift or variable length during training and inference? See Q2\n\n- Different values of M were set, what's the intuition and what's the best practice? See Q3", "questions": "Q1. How do you setup the values of hyperparameters of context tree? for example, their depth? are they sensitive to the inference tasks?\n\nQ2. How do you take care of variable length during training and inference? The dynamic NTK and Yarn used the inference-time ratio on this, but I am not so sure about this in your method, thank you in advance as I am out of curiosity. \n\nQ3. Robustness of different M values. M is an important model architecture base value, but I am not sure the robustness and meaning of setting different M values, and what is the best practice for this?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730700166046}, {"id": "flvpHjHeJC", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5960/Reviewer_swMr"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper proposes ShardLLM to reduce heavy training and inference cost in long-context LLMs. Specifically, this method insight is on the multi-grained context compression and query-aware information retrieval. SharedLLM uses  two short-context LLMs, where  the lower model functions as a compressor while the upper model acts as a decoder. The lower model divides the sequence into non-overlapping chunks and makes split-and-search procedure to choose relevant chunks. The selected chunks are then downsampled to reduce KV cache cost. The upper model encodes these keys and values via cross attention with chunk level positional embedding. \n\nThe paper makes experiments on language modeling and supervised fine-tuning to verify the effectiveness of SharedLLM. In language modeling, SharedLLM outperforms other efficient long-context baselines. In supervised fine-tuning, SharedLLM also achieves strong performance in LongBench and InfiniteBench. The efficient experiments show that SharedLLM is cheap and easy to deploy. Finally, the ablation studies discuss the design choice.", "review_text": "This paper proposes ShardLLM to reduce heavy training and inference cost in long-context LLMs. Specifically, this method insight is on the multi-grained context compression and query-aware information retrieval. SharedLLM uses  two short-context LLMs, where  the lower model functions as a compressor while the upper model acts as a decoder. The lower model divides the sequence into non-overlapping chunks and makes split-and-search procedure to choose relevant chunks. The selected chunks are then downsampled to reduce KV cache cost. The upper model encodes these keys and values via cross attention with chunk level positional embedding. \n\nThe paper makes experiments on language modeling and supervised fine-tuning to verify the effectiveness of SharedLLM. In language modeling, SharedLLM outperforms other efficient long-context baselines. In supervised fine-tuning, SharedLLM also achieves strong performance in LongBench and InfiniteBench. The efficient experiments show that SharedLLM is cheap and easy to deploy. Finally, the ablation studies discuss the design choice.", "strengths": "1. The paper focuses on an important research area, where LLMs are struggling to achieve effective and efficient long-context training and inference. SharedLLM integrates the compression and retrieval idea, which is a promising research direction in this field.\n\n2. The paper is well written. Although the proposed method includes a long procedure, the method section is organized well to help readers get the core idea. The experiment section includes all the necessary part, including performance, efficiency, and ablation studies.", "weaknesses": "1. The method is very complicated. I admit that \"complicated\" itself is not sometimes a weakness. However, simple and elegant design is usually working well in model architecture research. Complicated design brings engineering difficulty and optimization problem. SharedLLM uses the last output as the embedding of a chunk which is used for retrieval by the queries. It is highly questionable whether optimization is valid where sparse computation usually has difficulty on gradient estimation.\n\n2. SharedLLM divides context into different chunks and adds downsampling module to compress KV cache. This may hurt performance on some long-context tasks. The experiment part is not strong enough to support SharedLLM's long-context capability. I will give some existed benchmarks and easy examples accordingly:\n- Needle-in-a-haystack is now a compulsory evaluation to show long-context's retrieval ability, which is also included in InfiniteBench. However, the experiment only includes two subtasks. The other part is also essential to show model's long-context ability.\n- If I query the model to repeat all the previous context, the sparse-divided and compress context may hurt the context information. Moreover, If my first query does not include some context information, and my second query needs that again. Then, SharedLLM will drop it in the first response and can not answer my second question.\n\n3. The experiment also lacks important baselines of KV pruning methods. For example, StreamingLLM is an early baseline which directly drops all the global information while only maintaining the local KV cache. There are many sparse KV cache works, including H2O, SnapKV, FastGen. These works are efficient and memory-friendly, while maintaining parts of long-context capability.", "questions": "1. I'm curious about the result in Table 1 and Table 2. CEPE in line 349 is a re-produced result. Does it mean that the other results are all from the public checkpoint of the according paper? If so, I think the perplexity comparison is meaningless due to different experiment setting. If not so, the perplexity in short context (4k) is highly different, which is also weird.\n\n2. On Infinibench evaluation, why you are only interested in two subtasks? There is not a rationale for that. After all, if you use a bench for evaluation, you usually use the whole subtasks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes ShardLLM to reduce heavy training and inference cost in long-context LLMs. Specifically, this method insight is on the multi-grained context compression and query-aware information retrieval. SharedLLM uses  two short-context LLMs, where  the lower model functions as a compressor while the upper model acts as a decoder. The lower model divides the sequence into non-overlapping chunks and makes split-and-search procedure to choose relevant chunks. The selected chunks are then downsampled to reduce KV cache cost. The upper model encodes these keys and values via cross attention with chunk level positional embedding. \n\nThe paper makes experiments on language modeling and supervised fine-tuning to verify the effectiveness of SharedLLM. In language modeling, SharedLLM outperforms other efficient long-context baselines. In supervised fine-tuning, SharedLLM also achieves strong performance in LongBench and InfiniteBench. The efficient experiments show that SharedLLM is cheap and easy to deploy. Finally, the ablation studies discuss the design choice.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper focuses on an important research area, where LLMs are struggling to achieve effective and efficient long-context training and inference. SharedLLM integrates the compression and retrieval idea, which is a promising research direction in this field.\n\n2. The paper is well written. Although the proposed method includes a long procedure, the method section is organized well to help readers get the core idea. The experiment section includes all the necessary part, including performance, efficiency, and ablation studies.", "weaknesses": "1. The method is very complicated. I admit that \"complicated\" itself is not sometimes a weakness. However, simple and elegant design is usually working well in model architecture research. Complicated design brings engineering difficulty and optimization problem. SharedLLM uses the last output as the embedding of a chunk which is used for retrieval by the queries. It is highly questionable whether optimization is valid where sparse computation usually has difficulty on gradient estimation.\n\n2. SharedLLM divides context into different chunks and adds downsampling module to compress KV cache. This may hurt performance on some long-context tasks. The experiment part is not strong enough to support SharedLLM's long-context capability. I will give some existed benchmarks and easy examples accordingly:\n- Needle-in-a-haystack is now a compulsory evaluation to show long-context's retrieval ability, which is also included in InfiniteBench. However, the experiment only includes two subtasks. The other part is also essential to show model's long-context ability.\n- If I query the model to repeat all the previous context, the sparse-divided and compress context may hurt the context information. Moreover, If my first query does not include some context information, and my second query needs that again. Then, SharedLLM will drop it in the first response and can not answer my second question.\n\n3. The experiment also lacks important baselines of KV pruning methods. For example, StreamingLLM is an early baseline which directly drops all the global information while only maintaining the local KV cache. There are many sparse KV cache works, including H2O, SnapKV, FastGen. These works are efficient and memory-friendly, while maintaining parts of long-context capability.", "questions": "1. I'm curious about the result in Table 1 and Table 2. CEPE in line 349 is a re-produced result. Does it mean that the other results are all from the public checkpoint of the according paper? If so, I think the perplexity comparison is meaningless due to different experiment setting. If not so, the perplexity in short context (4k) is highly different, which is also weird.\n\n2. On Infinibench evaluation, why you are only interested in two subtasks? There is not a rationale for that. After all, if you use a bench for evaluation, you usually use the whole subtasks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730047737400}, {"id": "k5Jns3IcPQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5960/Reviewer_PZ2g"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper introduces SharedLLM, an innovative approach designed to extend the context window of LLMs without incurring the substantial costs associated with continual pre-training on long-context data. The authors claim that SharedLLM achieves comparable or superior results on long-context tasks while significantly reducing memory consumption and increasing processing speed compared to existing methods.", "review_text": "The paper introduces SharedLLM, an innovative approach designed to extend the context window of LLMs without incurring the substantial costs associated with continual pre-training on long-context data. The authors claim that SharedLLM achieves comparable or superior results on long-context tasks while significantly reducing memory consumption and increasing processing speed compared to existing methods.", "strengths": "1. The SharedLLM introduces an original dual-model architecture with a \"context tree\" data structure, enhancing the efficiency of context compression and retrieval for large language models.\n2. Despite the complexity of the concepts introduced, the paper communicates the workings of the SharedLLM and its underlying mechanisms", "weaknesses": "1. The evaluation only uses the LLaMA-2 model, with no justification for not including more recent or varied models like LLaMA-3.\n2. The method proposed in this paper does not seem to outperform other models in Longbench.", "questions": "See Weakness section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SharedLLM, an innovative approach designed to extend the context window of LLMs without incurring the substantial costs associated with continual pre-training on long-context data. The authors claim that SharedLLM achieves comparable or superior results on long-context tasks while significantly reducing memory consumption and increasing processing speed compared to existing methods.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The SharedLLM introduces an original dual-model architecture with a \"context tree\" data structure, enhancing the efficiency of context compression and retrieval for large language models.\n2. Despite the complexity of the concepts introduced, the paper communicates the workings of the SharedLLM and its underlying mechanisms", "weaknesses": "1. The evaluation only uses the LLaMA-2 model, with no justification for not including more recent or varied models like LLaMA-3.\n2. The method proposed in this paper does not seem to outperform other models in Longbench.", "questions": "See Weakness section", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729428375491}], "openreview_url": "https://openreview.net/forum?id=DfTWrTwLzD", "arxiv_id": "2410.19318", "paper_pdf": "papers/DfTWrTwLzD.pdf", "paper_pdf_sha256": "4bc2acdb1ec0ddc1f7eda01a70a2c7e9150b8f3c9641dcc88d38686ffac8ef1d", "paper_pdf_bytes": 553766, "paper_pdf_source": "openreview", "code_url": "https://github.com/Clement25/SharedLLM", "code_repository": "Clement25/SharedLLM", "code_commit": "79c3e15f8cead0b6877ffe1f2ff01ec8a1745bd7", "code_archive": "repos/DfTWrTwLzD.zip", "code_archive_sha256": "5e189fd627edca0a82d63e71eb41ff2f03b94d1f2d8424aa54ab4ef4842087b4", "code_archive_bytes": 91726, "code_file_count": 19, "code_extensions": {".py": 13, ".sh": 6}, "github_disk_usage_kb": 90, "github_languages": {"Python": 354136, "Shell": 2053}, "github_archived": false, "github_pushed_at": "2026-02-01T13:04:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/two-are-better-than-one-context-window"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iTrd5xyHLP", "year": 2024, "status": "rejected", "title": "LLMatic: Neural Architecture Search via Large Language Models and Quality Diversity Optimization", "authors": ["Muhammad Umair Nasir", "Sam Earle", "Julian Togelius", "Steven James", "Christopher Wesley Cleghorn"], "authorids": ["~Muhammad_Umair_Nasir1", "~Sam_Earle1", "~Julian_Togelius1", "~Steven_James1", "~Christopher_Wesley_Cleghorn1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) have emerged as powerful tools capable of accomplishing a broad spectrum of tasks. Their abilities span numerous areas, and one area where they have made a significant impact is in the domain of code generation. Here, we propose to use the coding abilities of LLMs to introduce meaningful variations to code defining neural networks. Meanwhile, Quality-Diversity (QD) algorithms are known to discover diverse and robust solutions. By merging the code-generating abilities of LLMs with the diversity and\nrobustness of QD solutions, we introduce LLMatic, a Neural Architecture Search (NAS) algorithm. While LLMs struggle to conduct NAS directly through prompts, LLMatic uses a procedural approach, leveraging QD for prompts and network architecture to create diverse and high-performing networks. We test LLMatic on the CIFAR-10 and NAS-bench-201 benchmark, demonstrating that it can produce competitive networks while evaluating just 2, 000 candidates, even without prior knowledge of the benchmark domain or exposure to any previous top-performing models for the benchmark.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "fbIQrEiv6s", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8949/Reviewer_eaNZ"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper introduces LLMatic, a NAS algorithm that combines LLMs with Quality-Diversity (QD) optimization to efficiently generate diverse and high-performing neural network architectures. Utilizing a dual-archive approach, LLMatic reduces the number of candidate evaluations needed, demonstrating competitive performance on benchmark datasets like CIFAR-10 and NAS-bench-201 with just 2,000 candidates. This approach not only leverages the code generation capabilities of LLMs to introduce meaningful variations in network architectures but also ensures robustness and diversity in the solutions through QD algorithms.", "review_text": "The paper introduces LLMatic, a NAS algorithm that combines LLMs with Quality-Diversity (QD) optimization to efficiently generate diverse and high-performing neural network architectures. Utilizing a dual-archive approach, LLMatic reduces the number of candidate evaluations needed, demonstrating competitive performance on benchmark datasets like CIFAR-10 and NAS-bench-201 with just 2,000 candidates. This approach not only leverages the code generation capabilities of LLMs to introduce meaningful variations in network architectures but also ensures robustness and diversity in the solutions through QD algorithms.", "strengths": "- the idea of employing LLMs for NAS is novel and aligns with the emerging trend of leveraging these models in diverse fields, considering that LLMs have been applied in domains with structured data formats such as drug design and materials science,", "weaknesses": "- LLMatic's performance heavily depends on the LLM's prior exposure to relevant coding patterns and neural network architectures during its training. This dependence could be a potential weakness, as the model might not generalize well to novel or highly specialized architectural search spaces.\n- The paper focuses on CIFAR-10 and NAS-bench-201 for performance evaluation. While these are standard benchmarks, the scope of testing could be broadened to include more diverse and challenging datasets. This would provide a more comprehensive understanding of LLMatic's capabilities and potential limitations, especially in real-world scenarios or more complex tasks.\n- The paper could benefit from ablation studies to understand the contribution of each component of LLMatic to its overall performance.\n- The writing should improve.", "questions": "- What's the intuition behind the proposed method? How can we prove that LLMs have the prior knowledge about NAS tasks as described in the paper?\n- The input sequence length of current LLMs is limited, but the number of networks evaluated in one experiment exceeds 2000. Is this a major problem, and how do you address it? Or would you consider incorporating the search history into the prompt to assist NAS?\n- Why randomly pick a prompt as an action at each step? Why not design prompts that allow LLMs to choose the action by themselves?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces LLMatic, a NAS algorithm that combines LLMs with Quality-Diversity (QD) optimization to efficiently generate diverse and high-performing neural network architectures. Utilizing a dual-archive approach, LLMatic reduces the number of candidate evaluations needed, demonstrating competitive performance on benchmark datasets like CIFAR-10 and NAS-bench-201 with just 2,000 candidates. This approach not only leverages the code generation capabilities of LLMs to introduce meaningful variations in network architectures but also ensures robustness and diversity in the solutions through QD algorithms.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "1 poor", "strengths": "- the idea of employing LLMs for NAS is novel and aligns with the emerging trend of leveraging these models in diverse fields, considering that LLMs have been applied in domains with structured data formats such as drug design and materials science,", "weaknesses": "- LLMatic's performance heavily depends on the LLM's prior exposure to relevant coding patterns and neural network architectures during its training. This dependence could be a potential weakness, as the model might not generalize well to novel or highly specialized architectural search spaces.\n- The paper focuses on CIFAR-10 and NAS-bench-201 for performance evaluation. While these are standard benchmarks, the scope of testing could be broadened to include more diverse and challenging datasets. This would provide a more comprehensive understanding of LLMatic's capabilities and potential limitations, especially in real-world scenarios or more complex tasks.\n- The paper could benefit from ablation studies to understand the contribution of each component of LLMatic to its overall performance.\n- The writing should improve.", "questions": "- What's the intuition behind the proposed method? How can we prove that LLMs have the prior knowledge about NAS tasks as described in the paper?\n- The input sequence length of current LLMs is limited, but the number of networks evaluated in one experiment exceeds 2000. Is this a major problem, and how do you address it? Or would you consider incorporating the search history into the prompt to assist NAS?\n- Why randomly pick a prompt as an action at each step? Why not design prompts that allow LLMs to choose the action by themselves?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698831477311}, {"id": "nxh59jW2zv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8949/Reviewer_4jrW"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work proposes LLMatic which is a Neural Architecture Search approach using LLMs and quality diversity optimization. This work combines the domain knowledge of code-generating LLMs with a robust quality-diverse search mechanism.The application of LLMs trained for generating architecture code NAS is novel and interesting way to exploit the capabilities of code-gen LLMs. The method is evaluated on CIFAR-10 dataset and the  NAS-bench-201 benchmark.", "review_text": "This work proposes LLMatic which is a Neural Architecture Search approach using LLMs and quality diversity optimization. This work combines the domain knowledge of code-generating LLMs with a robust quality-diverse search mechanism.The application of LLMs trained for generating architecture code NAS is novel and interesting way to exploit the capabilities of code-gen LLMs. The method is evaluated on CIFAR-10 dataset and the  NAS-bench-201 benchmark.", "strengths": "- The application of code-generating LLMs effectively for NAS is very novel. Defining architecture generation as a language/code modelling task is a new way of formulating the neural architecture search problem.\n- The presentation of the paper is mostly clear except in some parts (refer to suggestions and questions in the weakness and questions section)\n- Evaluation on the NB201 dataset and CIFAR10 dataset fairly exhaustive and well ablated. \n- The authors release their code and additional details on the prompts used", "weaknesses": "- Evaluation is limited: Currently NAS focuses a lot of transformer-based spaces [1], mobilenet-spaces [2] which are more practical and realistic compared to cell-based spaces. A lot of thee approaches release a surrogate predictor or the supernet itself, to save training times ie the architecture training part in algorithm 1. I recommend the authors evaluate the method on these search spaces too. This is also in my opinion very important to study how llmatic scales across larger architecture definitions (code) and datasets eg: ImageNet. \n- Comparison to black box approaches and generative approaches: Since llmatic requires training a lot of architectures (or querying a benchmark multiple times), its search time is more comparable to black-box approaches instead of approaches like lambda-darts. It would be great to add other black-box methods (in addition to random search) to Table 1. Furthermore since the work very much falls in the line of generative NAS a comparison with DiffusionNAG [3] would also be great.\n- Clarity: In the current version of the paper in figure 4, I couldn't see the green points referring to llmatic architectures, am I missing something?\n- Minor : Page 4 2nd paragraph \"In the first generation , ta simple neural network with one convolutional and one fully connected\nlayer initiates the evolution\" -> \"In the first generation , take simple neural network with one convolutional and one fully connected\nlayer initiates the evolution\"\n\n[1] Chen, M., Peng, H., Fu, J. and Ling, H., 2021. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 12270-12280).\n\n[2] Cai, H., Gan, C., Wang, T., Zhang, Z. and Han, S., 2019. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791.\n\n[3] An, S., Lee, H., Jo, J., Lee, S. and Hwang, S.J., 2023. DiffusionNAG: Task-guided Neural Architecture Generation with Diffusion Models. arXiv preprint arXiv:2305.16943.", "questions": "- Evaluation: Check points in the weakness (evaluation)\n- Clarity : Code generating llms are often imperfect in generating pytorch code. They often introduce bugs and require some human intervention to make a generated code snippet compilable. Did the authors face any such issues? In other works how does a given prompt ensures that an LLM would restrict itself to generate compilable architectures within the NABench 201 search space for eg? The prompt examples provided do not seem to indicate any instructions to the LLM to ensure this\n- Could the authors comment on how their work compares to [4] which explores a very similar direction (but with prompt evolution)?\n- Reproducibility: Reproducibility of approaches is a major challenge in NAS. Since the work relies on (open-source) LLMs could the authors comment on the reproducibility of their approach?\n- Could you report search time comparison across all baselines in table 1?\n\n[4] Chen, A., Dohan, D.M. and So, D.R., 2023. EvoPrompting: Language Models for Code-Level Neural Architecture Search. arXiv preprint arXiv:2302.14838.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes LLMatic which is a Neural Architecture Search approach using LLMs and quality diversity optimization. This work combines the domain knowledge of code-generating LLMs with a robust quality-diverse search mechanism.The application of LLMs trained for generating architecture code NAS is novel and interesting way to exploit the capabilities of code-gen LLMs. The method is evaluated on CIFAR-10 dataset and the  NAS-bench-201 benchmark.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- The application of code-generating LLMs effectively for NAS is very novel. Defining architecture generation as a language/code modelling task is a new way of formulating the neural architecture search problem.\n- The presentation of the paper is mostly clear except in some parts (refer to suggestions and questions in the weakness and questions section)\n- Evaluation on the NB201 dataset and CIFAR10 dataset fairly exhaustive and well ablated. \n- The authors release their code and additional details on the prompts used", "weaknesses": "- Evaluation is limited: Currently NAS focuses a lot of transformer-based spaces [1], mobilenet-spaces [2] which are more practical and realistic compared to cell-based spaces. A lot of thee approaches release a surrogate predictor or the supernet itself, to save training times ie the architecture training part in algorithm 1. I recommend the authors evaluate the method on these search spaces too. This is also in my opinion very important to study how llmatic scales across larger architecture definitions (code) and datasets eg: ImageNet. \n- Comparison to black box approaches and generative approaches: Since llmatic requires training a lot of architectures (or querying a benchmark multiple times), its search time is more comparable to black-box approaches instead of approaches like lambda-darts. It would be great to add other black-box methods (in addition to random search) to Table 1. Furthermore since the work very much falls in the line of generative NAS a comparison with DiffusionNAG [3] would also be great.\n- Clarity: In the current version of the paper in figure 4, I couldn't see the green points referring to llmatic architectures, am I missing something?\n- Minor : Page 4 2nd paragraph \"In the first generation , ta simple neural network with one convolutional and one fully connected\nlayer initiates the evolution\" -> \"In the first generation , take simple neural network with one convolutional and one fully connected\nlayer initiates the evolution\"\n\n[1] Chen, M., Peng, H., Fu, J. and Ling, H., 2021. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 12270-12280).\n\n[2] Cai, H., Gan, C., Wang, T., Zhang, Z. and Han, S., 2019. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791.\n\n[3] An, S., Lee, H., Jo, J., Lee, S. and Hwang, S.J., 2023. DiffusionNAG: Task-guided Neural Architecture Generation with Diffusion Models. arXiv preprint arXiv:2305.16943.", "questions": "- Evaluation: Check points in the weakness (evaluation)\n- Clarity : Code generating llms are often imperfect in generating pytorch code. They often introduce bugs and require some human intervention to make a generated code snippet compilable. Did the authors face any such issues? In other works how does a given prompt ensures that an LLM would restrict itself to generate compilable architectures within the NABench 201 search space for eg? The prompt examples provided do not seem to indicate any instructions to the LLM to ensure this\n- Could the authors comment on how their work compares to [4] which explores a very similar direction (but with prompt evolution)?\n- Reproducibility: Reproducibility of approaches is a major challenge in NAS. Since the work relies on (open-source) LLMs could the authors comment on the reproducibility of their approach?\n- Could you report search time comparison across all baselines in table 1?\n\n[4] Chen, A., Dohan, D.M. and So, D.R., 2023. EvoPrompting: Language Models for Code-Level Neural Architecture Search. arXiv preprint arXiv:2302.14838.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698829472114}, {"id": "z1rc3MMmXc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8949/Reviewer_7vVP"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes the use of LLMs' code generation capability to generate network models, along with the use of Quality-Diversity for search. The authors validate the feasibility of the method on multiple datasets and search spaces.", "review_text": "This paper proposes the use of LLMs' code generation capability to generate network models, along with the use of Quality-Diversity for search. The authors validate the feasibility of the method on multiple datasets and search spaces.", "strengths": "- The use of large models' knowledge for network structure generation provides an example of applying large models in practice.", "weaknesses": "- While utilizing the code generation capability of large models for network structure generation is a good application, previous works such as GENIUS have already proposed very similar methods and demonstrated prompts. Therefore, the novelty and contribution of this paper appear insufficient.\n- The experimental description is not detailed, and the analysis is not sufficiently deep. For example, Table 1 does not report the search cost of this paper compared to related works. \n- There are relatively few experimental comparisons, and the comparison with related works is not comprehensive enough.\n- After reading this paper, I did not get much insight. Importantly, the assistance of prior knowledge from LLMs in network structure design remains unclear.", "questions": "Recommendations:\n- Provide a more comprehensive discussion on the novelty and contribution of the proposed method compared to similar existing approaches.\n- Elaborate on the experimental setup and provide more detailed analysis of the results, including reporting the search cost and the impact of prior knowledge from large models on network structure design.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the use of LLMs' code generation capability to generate network models, along with the use of Quality-Diversity for search. The authors validate the feasibility of the method on multiple datasets and search spaces.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The use of large models' knowledge for network structure generation provides an example of applying large models in practice.", "weaknesses": "- While utilizing the code generation capability of large models for network structure generation is a good application, previous works such as GENIUS have already proposed very similar methods and demonstrated prompts. Therefore, the novelty and contribution of this paper appear insufficient.\n- The experimental description is not detailed, and the analysis is not sufficiently deep. For example, Table 1 does not report the search cost of this paper compared to related works. \n- There are relatively few experimental comparisons, and the comparison with related works is not comprehensive enough.\n- After reading this paper, I did not get much insight. Importantly, the assistance of prior knowledge from LLMs in network structure design remains unclear.", "questions": "Recommendations:\n- Provide a more comprehensive discussion on the novelty and contribution of the proposed method compared to similar existing approaches.\n- Elaborate on the experimental setup and provide more detailed analysis of the results, including reporting the search cost and the impact of prior knowledge from large models on network structure design.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698757586892}, {"id": "vFMe6YyWJY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8949/Reviewer_j44t"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "1 poor", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose an LLM-based NAS algorithm which leverages the coding ability of LLMs in combination with a type of evolutionary search which allows to mutate architectures directly in the code of the architecture rather than on e.g. an abstract graph level. The authors employ QD for their evolutionary search, which results in a wide-range of solutions which can e.g. fulfill different hardware requirements. The authors demonstrate good performance on NAS-Bench-201 using only 2000 queries.", "review_text": "The authors propose an LLM-based NAS algorithm which leverages the coding ability of LLMs in combination with a type of evolutionary search which allows to mutate architectures directly in the code of the architecture rather than on e.g. an abstract graph level. The authors employ QD for their evolutionary search, which results in a wide-range of solutions which can e.g. fulfill different hardware requirements. The authors demonstrate good performance on NAS-Bench-201 using only 2000 queries.", "strengths": "- I like the idea of carrying out NAS directly on the code-level instead of an abstract representation such as a DAG, as it makes the search flexible. This also removes the need for a compiler from the abstract model description to the code which can save time and hence lower the barrier to actually use NAS. Joint NAS and HPO may also be expressed in a very natural way in this paradigm.\n- Using quality diversity in this paradigm makes sense to me, not just for diverse hardware requirements, but also for good ensembling. \n- The ablation experiment nicely illustrates the contribution of each component.", "weaknesses": "- Could you run additional experiments on Nas-Bench-101? As the search space is constrained in a different way than 201, I would be interested to see whether your approach can remain within the search space at all times.\n- Could you provide optimization trajectories for Nas-Bench-201 along with a baseline such as random search? I am curious how your approach compares with more traditional approaches in terms of any-time performance.\n- Overall the visual quality of the paper may be improved. In Figures 2 and 3 the white space around each Figure should reduced. Figure 4 should be changed to a jpg or contain a reduced number of points because it causes rendering issues. Also please increase the fontsize to be at least footnotesize for all figures.\n- I find Algorithm 1 to be hard to read. As you don't define each function you use in the pseudo code, you could consider writing short sentences instead.", "questions": "- Please check that your references use the published version of a paper (if available) e.g. NAS-Bench-Suite was published at ICLR 2022.\n\nTypos (only minor and non exhaustive, just listing them for completeness):\n- 'In the first generation , ta simple neural network'\n- '... layer that takes in 3 input(s?) with 3 channels'", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose an LLM-based NAS algorithm which leverages the coding ability of LLMs in combination with a type of evolutionary search which allows to mutate architectures directly in the code of the architecture rather than on e.g. an abstract graph level. The authors employ QD for their evolutionary search, which results in a wide-range of solutions which can e.g. fulfill different hardware requirements. The authors demonstrate good performance on NAS-Bench-201 using only 2000 queries.", "soundness": "3 good", "presentation": "1 poor", "contribution": "3 good", "strengths": "- I like the idea of carrying out NAS directly on the code-level instead of an abstract representation such as a DAG, as it makes the search flexible. This also removes the need for a compiler from the abstract model description to the code which can save time and hence lower the barrier to actually use NAS. Joint NAS and HPO may also be expressed in a very natural way in this paradigm.\n- Using quality diversity in this paradigm makes sense to me, not just for diverse hardware requirements, but also for good ensembling. \n- The ablation experiment nicely illustrates the contribution of each component.", "weaknesses": "- Could you run additional experiments on Nas-Bench-101? As the search space is constrained in a different way than 201, I would be interested to see whether your approach can remain within the search space at all times.\n- Could you provide optimization trajectories for Nas-Bench-201 along with a baseline such as random search? I am curious how your approach compares with more traditional approaches in terms of any-time performance.\n- Overall the visual quality of the paper may be improved. In Figures 2 and 3 the white space around each Figure should reduced. Figure 4 should be changed to a jpg or contain a reduced number of points because it causes rendering issues. Also please increase the fontsize to be at least footnotesize for all figures.\n- I find Algorithm 1 to be hard to read. As you don't define each function you use in the pseudo code, you could consider writing short sentences instead.", "questions": "- Please check that your references use the published version of a paper (if available) e.g. NAS-Bench-Suite was published at ICLR 2022.\n\nTypos (only minor and non exhaustive, just listing them for completeness):\n- 'In the first generation , ta simple neural network'\n- '... layer that takes in 3 input(s?) with 3 channels'", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698072054131}, {"id": "4XwBmmlHEd", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8949/Reviewer_za3A"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies an interesting topic regarding how to leverage Large Language Model to generate DNN architecture. The proposed LLMatic method jointly utilizes prompts, QD, two types of evolution operators, i.e., mutation and crossover,  to produce high-performing DNNs. The numerical results on CIFAR10 and NAS-bench-201 show the effectiveness of the proposed approach.", "review_text": "This paper studies an interesting topic regarding how to leverage Large Language Model to generate DNN architecture. The proposed LLMatic method jointly utilizes prompts, QD, two types of evolution operators, i.e., mutation and crossover,  to produce high-performing DNNs. The numerical results on CIFAR10 and NAS-bench-201 show the effectiveness of the proposed approach.", "strengths": "After reading the manuscript, I summarize the below strengths. \n\n- The topic of the paper is new and interesting. Leveraging generative models may produce impacts into the NAS field.\n\n- The designed algorithm makes sense that should produce some high-performing DNNs.", "weaknesses": "Meanwhile, I have following concerns.\n\n- The presentation of the paper is not satisfactory. Especially the pseudocode, which lacks proper mathematical annotations. Meanwhile, there exist a lot of typos, e.g., 'ta simple' on page 4. The legends in Figures are invisible, etc. \n\n- Again the presentation, the key components in the algorithm such as mutation operators, crossover operators and temperature mutation lacks proper description and explanations. It makes the algorithm unclear.\n\n- This paper looks more like an investigation paper to explore how to use LLM on generating new DNNs, yet lacks sufficiently novel algorithm to guide LLMs generating DNNs of higher fidelity. The rank on ImageNet is significantly lower than CIFARs, which increases my concern regarding the effectiveness of the proposed algorithms on general tasks.", "questions": "- What is the search cost of the algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies an interesting topic regarding how to leverage Large Language Model to generate DNN architecture. The proposed LLMatic method jointly utilizes prompts, QD, two types of evolution operators, i.e., mutation and crossover,  to produce high-performing DNNs. The numerical results on CIFAR10 and NAS-bench-201 show the effectiveness of the proposed approach.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "3 good", "strengths": "After reading the manuscript, I summarize the below strengths. \n\n- The topic of the paper is new and interesting. Leveraging generative models may produce impacts into the NAS field.\n\n- The designed algorithm makes sense that should produce some high-performing DNNs.", "weaknesses": "Meanwhile, I have following concerns.\n\n- The presentation of the paper is not satisfactory. Especially the pseudocode, which lacks proper mathematical annotations. Meanwhile, there exist a lot of typos, e.g., 'ta simple' on page 4. The legends in Figures are invisible, etc. \n\n- Again the presentation, the key components in the algorithm such as mutation operators, crossover operators and temperature mutation lacks proper description and explanations. It makes the algorithm unclear.\n\n- This paper looks more like an investigation paper to explore how to use LLM on generating new DNNs, yet lacks sufficiently novel algorithm to guide LLMs generating DNNs of higher fidelity. The rank on ImageNet is significantly lower than CIFARs, which increases my concern regarding the effectiveness of the proposed algorithms on general tasks.", "questions": "- What is the search cost of the algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697848451887}], "openreview_url": "https://openreview.net/forum?id=iTrd5xyHLP", "arxiv_id": "2306.01102", "paper_pdf": "papers/iTrd5xyHLP.pdf", "paper_pdf_sha256": "d32d664d3b3bdb8367d9f94612df7015977236f5bee4c7ab992b21757b6fecc5", "paper_pdf_bytes": 939503, "paper_pdf_source": "openreview", "code_url": "https://github.com/umair-nasir14/LLMatic", "code_repository": "umair-nasir14/LLMatic", "code_commit": "bd29c7e7d2348620083e21049f0c81a89c75ed07", "code_archive": "repos/iTrd5xyHLP.zip", "code_archive_sha256": "f54f3be8637d137734b54e865353ecbac0379d3cc603355ddf0a8ed30209baec", "code_archive_bytes": 46601, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 70, "github_languages": {"Python": 93398}, "github_archived": false, "github_pushed_at": "2024-08-14T10:42:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/llmatic-neural-architecture-search-via-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lJX9okHBVMb", "year": 2023, "status": "rejected", "title": "Universal approximation and model compression for radial neural networks ", "authors": ["Iordan Ganev", "Twan van Laarhoven", "Robin Walters"], "authorids": ["~Iordan_Ganev1", "~Twan_van_Laarhoven1", "~Robin_Walters1"], "authors_source": "OpenReview API", "abstract": "We introduce a class of fully-connected neural networks whose activation functions, rather than being pointwise, rescale feature vectors by a function depending only on their norm. We call such networks radial neural networks, extending previous work on rotation equivariant networks that considers rescaling activations in less generality. We prove universal approximation theorems for radial neural networks, including in the more difficult cases of bounded widths and unbounded domains. Our proof techniques are novel, distinct from those in the pointwise case. Additionally, radial neural networks exhibit a rich group of orthogonal change-of-basis symmetries on the vector space of trainable parameters. Factoring out these symmetries leads to a practical lossless model compression algorithm. Optimization of the compressed model by gradient descent is equivalent to projected gradient descent for the full model.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "A9JQgbLvD3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1797/Reviewer_vdPZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Instead of considering entry wise nonlinearities in neural networks, the paper studies radial neural networks, where the activation function acts to rescale the output of a whole layer. Asymptotic universal approximation ability of radial neural networks is proved. In addition, it is shown that radial neural networks can be compressed with an efficient algorithm based on QR decompositions. Experiments on synthetic data and MNIST are provided to support the theory.", "review_text": "From a pure theoretical point of view, the paper can be a good study on radial neural networks, if a clearer exposure of technical contributions is provided. Further, I believe the paper would attract more attention if the practicality of radial neural networks is discussed and advantages over pointwise neural networks are demonstrated. The best result would be using the developed theory to explain the advantage of radial neural networks.\n\n=============== Post author response ===============\n\nThank you for the detailed response. The explanation on approximation in $\\mathbb{R}^d$ makes sense to me. Examples of $\\sin x$ are difficult to approximate in $\\mathbb{R}^d$ since they do not vanish at infinity. The ultimate idea is always to truncate the domain.\n\nWith the experiments in Section 7 and Appendix E, the practical motivation of radial neural networks might still be limited. As pointed out, radial neural networks can be advantageous when spherical decision boundaries appear in tasks.\n\nOverall, I think the paper makes interesting contributions to universal approximation theories of neural networks. Nonetheless, practical implications are on the weak side.\n\n", "strengths": "=============== Strength ===============\n\nThe paper manages to contain a rich amount of related references and provide graphical illustration that makes the ideas easy to understand.\n\nFrom a technical point of view, the asymptotic universal approximation property of radial neural networks is new to me. Besides, the compressive nature of radial neural networks and its resulting fast optimization are also not studied in literature.\n\nThe organization of the paper is good and easy to follow.\n\n=============== Weakness ===============\n\nA strong motivation of the practicality of radial neural networks seems to be lacking. It will be helpful to provide working examples of radial neural networks and indicate that in some important applications, radial neural networks outperform pointwise networks.\n\nThe universal approximation theory allows the network size to grow to infinity. It seems unfair to claim that not requiring bounded domain is a contribution. In early asymptotic results, e.g., Cybenko and Barron, neural networks are already shown to be dense in differential function spaces with unbounded domain. Later results established the rate of approximation requires a compact domain. As there is no rate of approximation provided, considering unbounded domain does not add much technical challenges. Moreover, the asymptotic results does not provide understanding of potential advantages of radial neural networks over pointwise neural networks in terms of function approximation.\n\nSynthetic data is relatively toy and the comparison with ReLU network is not fully convincing. On the one hand, synthetic data are low-dimensional and the target function is simple in that the activation function is almost the form of the target function. On the other hand, Step-ReLU radial network outperforms ReLU MLP might correlate to multiple reasons: 1) different activation functions; 2) different network architectures; 3) hyper parameter tuning. By the way, is it possible to simply perform classification on MNIST using radial and pointwise neural networks?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Instead of considering entry wise nonlinearities in neural networks, the paper studies radial neural networks, where the activation function acts to rescale the output of a whole layer. Asymptotic universal approximation ability of radial neural networks is proved. In addition, it is shown that radial neural networks can be compressed with an efficient algorithm based on QR decompositions. Experiments on synthetic data and MNIST are provided to support the theory.", "strength_and_weaknesses": "=============== Strength ===============\n\nThe paper manages to contain a rich amount of related references and provide graphical illustration that makes the ideas easy to understand.\n\nFrom a technical point of view, the asymptotic universal approximation property of radial neural networks is new to me. Besides, the compressive nature of radial neural networks and its resulting fast optimization are also not studied in literature.\n\nThe organization of the paper is good and easy to follow.\n\n=============== Weakness ===============\n\nA strong motivation of the practicality of radial neural networks seems to be lacking. It will be helpful to provide working examples of radial neural networks and indicate that in some important applications, radial neural networks outperform pointwise networks.\n\nThe universal approximation theory allows the network size to grow to infinity. It seems unfair to claim that not requiring bounded domain is a contribution. In early asymptotic results, e.g., Cybenko and Barron, neural networks are already shown to be dense in differential function spaces with unbounded domain. Later results established the rate of approximation requires a compact domain. As there is no rate of approximation provided, considering unbounded domain does not add much technical challenges. Moreover, the asymptotic results does not provide understanding of potential advantages of radial neural networks over pointwise neural networks in terms of function approximation.\n\nSynthetic data is relatively toy and the comparison with ReLU network is not fully convincing. On the one hand, synthetic data are low-dimensional and the target function is simple in that the activation function is almost the form of the target function. On the other hand, Step-ReLU radial network outperforms ReLU MLP might correlate to multiple reasons: 1) different activation functions; 2) different network architectures; 3) hyper parameter tuning. By the way, is it possible to simply perform classification on MNIST using radial and pointwise neural networks?", "clarity,_quality,_novelty_and_reproducibility": "As mentioned, the paper is relatively clear. Some notations might need further elaboration. For example, following Theorem 3, $T_i$ and $S_i$ are affine maps of corresponding input and output dimensions. Is it unclear how $T_i$ can lift the input space to a higher dimension using vector addition.\n\nIt is better to highlight technical contributions in Section 4 and compare to the universal approximation of pointwise networks to demonstrate any potential advantage if possible.", "summary_of_the_review": "From a pure theoretical point of view, the paper can be a good study on radial neural networks, if a clearer exposure of technical contributions is provided. Further, I believe the paper would attract more attention if the practicality of radial neural networks is discussed and advantages over pointwise neural networks are demonstrated. The best result would be using the developed theory to explain the advantage of radial neural networks.\n\n=============== Post author response ===============\n\nThank you for the detailed response. The explanation on approximation in $\\mathbb{R}^d$ makes sense to me. Examples of $\\sin x$ are difficult to approximate in $\\mathbb{R}^d$ since they do not vanish at infinity. The ultimate idea is always to truncate the domain.\n\nWith the experiments in Section 7 and Appendix E, the practical motivation of radial neural networks might still be limited. As pointed out, radial neural networks can be advantageous when spherical decision boundaries appear in tasks.\n\nOverall, I think the paper makes interesting contributions to universal approximation theories of neural networks. Nonetheless, practical implications are on the weak side.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666715629520}, {"id": "ZcaZwnbuj6c", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1797/Reviewer_Vbpq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes *radial neural networks*, neural networks with a non-pointwise nonlinearity that is a scalar rescaling of the vector of activations as a function of its norm. The paper argues that such neural networks are rotation equivariant and proves that they are universal function approximators. It further introduces a lossless compression technique for these networks and supports its claims empirically.", "review_text": "I think this is interesting research in general, however the core contribution and future directions are not clear from the manuscript to me, so I will recommend rejection for now", "strengths": "Strengths:\n* Nonlinearities are a fundamental building block of neural network architectures, so new techniques are potentially broadly applicable.\n* The writing of the paper is clear and focused within the respective sections.\n* The range of contributions is quite broad, ranging from the theoretical results on the proposed networks being universal function approximators to a practical compression method taking advantage of their structure.\n\nWeaknesses:\n* Probably as a result of its breadth, the paper somewhat lacks a clear red thread or headline result. I am not really sure who the target audience of the paper is. Is the rotation equivariance a means to the end of constructing neural networks that are easy to compress, i.e. is this a practically minded paper? Do the authors want to establish radial neural networks as a new direction of research? Then what are promising directions for future work and what specific limitations of current architectures are they meant to overcome?\n* There are essentially no quantitative results. I really don't mind if the methods can't be immediately plugged into ResNets or Transformers and give state-of-the-art results, but the paper should at least give a general idea of how well the method works on standard tasks (MNIST, CIFAR, ...). What's the rationale behind choosing the benchmark on the noised MNIST images? Is it expected that radial neural networks perform better on this task?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes *radial neural networks*, neural networks with a non-pointwise nonlinearity that is a scalar rescaling of the vector of activations as a function of its norm. The paper argues that such neural networks are rotation equivariant and proves that they are universal function approximators. It further introduces a lossless compression technique for these networks and supports its claims empirically.", "strength_and_weaknesses": "Strengths:\n* Nonlinearities are a fundamental building block of neural network architectures, so new techniques are potentially broadly applicable.\n* The writing of the paper is clear and focused within the respective sections.\n* The range of contributions is quite broad, ranging from the theoretical results on the proposed networks being universal function approximators to a practical compression method taking advantage of their structure.\n\nWeaknesses:\n* Probably as a result of its breadth, the paper somewhat lacks a clear red thread or headline result. I am not really sure who the target audience of the paper is. Is the rotation equivariance a means to the end of constructing neural networks that are easy to compress, i.e. is this a practically minded paper? Do the authors want to establish radial neural networks as a new direction of research? Then what are promising directions for future work and what specific limitations of current architectures are they meant to overcome?\n* There are essentially no quantitative results. I really don't mind if the methods can't be immediately plugged into ResNets or Transformers and give state-of-the-art results, but the paper should at least give a general idea of how well the method works on standard tasks (MNIST, CIFAR, ...). What's the rationale behind choosing the benchmark on the noised MNIST images? Is it expected that radial neural networks perform better on this task?", "clarity,_quality,_novelty_and_reproducibility": "**Clarity** The technical writing is clear, however the key contribution of the work could be worked out better.\n**Quality** The paper is sound as far as I can tell, although I did not go through the derivations or proofs.\n**Novelty** The work is new as far as I am aware, I am however unfamiliar with most closely related pieces of work.", "summary_of_the_review": "I think this is interesting research in general, however the core contribution and future directions are not clear from the manuscript to me, so I will recommend rejection for now", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666568455452}, {"id": "YfWbNiIDL2", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1797/Reviewer_MQ2f"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a new neural network architecture which is radial: instead of pointwise activation, the activation is applied to the whole layer which rescales the output vector according to its norm. Universal approximation results are shown, with additional model compression algorithms.", "review_text": "The results are solid and novel but not as significant: the new radial architecture doesn't show improvement over previous bounds. As a result I lean to weak accept rather than accept.", "strengths": "Strength:\n\n1, the proposed radial neural network is novel. The way it generalizes pointwise activation is natural and elegant, that a simple rescaling of vector is actually expressive.\n\n2, the universal approximation results are solid. The proof method is intuitive and resembles classic ideas in literature.\n\nWeaknesses:\n\n1, I don't see any improvement over previous bounds for universal approximation: for the usual Lipschitz function case, the construction of Theorem 3 requires $O((\\frac{\\sqrt{n}}{\\epsilon})^{3n})$ number of parameters (ignoring poly factors). Theorem 5 improves it to the standard $O((\\frac{\\sqrt{n}}{\\epsilon})^{n})$.\n\n2, the compression algorithm is useful only if there are subsequent layers with notably larger width. However, such neural networks are already sub-optimal theoretically and wouldn't be used in the first place, then there is no need for compression.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new neural network architecture which is radial: instead of pointwise activation, the activation is applied to the whole layer which rescales the output vector according to its norm. Universal approximation results are shown, with additional model compression algorithms.", "strength_and_weaknesses": "Strength:\n\n1, the proposed radial neural network is novel. The way it generalizes pointwise activation is natural and elegant, that a simple rescaling of vector is actually expressive.\n\n2, the universal approximation results are solid. The proof method is intuitive and resembles classic ideas in literature.\n\nWeaknesses:\n\n1, I don't see any improvement over previous bounds for universal approximation: for the usual Lipschitz function case, the construction of Theorem 3 requires $O((\\frac{\\sqrt{n}}{\\epsilon})^{3n})$ number of parameters (ignoring poly factors). Theorem 5 improves it to the standard $O((\\frac{\\sqrt{n}}{\\epsilon})^{n})$.\n\n2, the compression algorithm is useful only if there are subsequent layers with notably larger width. However, such neural networks are already sub-optimal theoretically and wouldn't be used in the first place, then there is no need for compression.", "clarity,_quality,_novelty_and_reproducibility": "Sections 1-4 are well-written and easy-to-follow. Sections 5-6 are felt not tightly linked to the main results, and I think the mathematical writing can be made more succinct in the main-text. A small polishing on the writing of sections 5-6 could be helpful.", "summary_of_the_review": "The results are solid and novel but not as significant: the new radial architecture doesn't show improvement over previous bounds. As a result I lean to weak accept rather than accept.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666198963948}], "openreview_url": "https://openreview.net/forum?id=lJX9okHBVMb", "arxiv_id": "2107.02550", "paper_pdf": "papers/lJX9okHBVMb.pdf", "paper_pdf_sha256": "f124ec1a6de6efd6c33a97f95ce115f01c3f279203fe12d1dd9f6a0d5e70d44d", "paper_pdf_bytes": 813577, "paper_pdf_source": "openreview", "code_url": "https://github.com/ivganev/QR-decomposition-radial-NNs", "code_repository": "ivganev/QR-decomposition-radial-NNs", "code_commit": "37ff7009815f30651e45d8c1e130640b60cc0cd3", "code_archive": "repos/lJX9okHBVMb.zip", "code_archive_sha256": "25cda1d09ef290c6879d33166ce1e72270d0d66889906063883d57d3ce0ed9c8", "code_archive_bytes": 43238, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 138, "github_languages": {"Python": 24519}, "github_archived": false, "github_pushed_at": "2023-09-26T15:14:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-qr-decomposition-for-radial-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YgR1rRWETI", "year": 2022, "status": "rejected", "title": "Connectivity Matters: Neural Network Pruning Through the Lens of Effective Sparsity", "authors": ["Artem M Vysogorets", "Julia Kempe"], "authorids": ["~Artem_M_Vysogorets1", "~Julia_Kempe1"], "authors_source": "OpenReview API", "abstract": "Neural network pruning is a fruitful area of research with surging interest in high sparsity regimes. Benchmarking in this domain heavily relies on faithful representation of the sparsity of subnetworks, which has been traditionally computed as the fraction of removed connections (direct sparsity). This definition, however, fails to recognize unpruned parameters that detached from input or output layers of underlying subnetworks, potentially underestimating actual effective sparsity: the fraction of inactivated connections. While this effect might be negligible for moderately pruned networks (up to $10\\times-100\\times$ compression rates), we find that it plays an increasing role for thinner subnetworks, greatly distorting comparison between different pruning algorithms. For example, we show that effective compression of a randomly pruned LeNet-300-100 can be orders of magnitude larger than its direct counterpart, while no discrepancy is ever observed when using SynFlow for pruning (Tanaka et al., 2020). In this work, we adopt the lens of effective sparsity to reevaluate several recent pruning algorithms on common benchmark architectures (e.g., LeNet-300-100, VGG-19, ResNet-18) and discover that their absolute and relative performance changes dramatically in this new, and as we argue, more appropriate framework. To aim for effective, rather than direct, sparsity, we develop a low-cost extension to most pruning algorithms. Further, equipped with effective sparsity as a reference frame, we partially reconfirm that random pruning with appropriate sparsity allocation across layers performs as well or better than more sophisticated algorithms for pruning at initialization (Su et al., 2020). In response to this observation, using a simple analogy of pressure distribution in coupled cylinders from thermodynamics, we design novel layerwise sparsity quotas that outperform all existing baselines in the context of random pruning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "wH1dZtTylC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3938/Reviewer_B6nv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper highlights the fact that when doing unstructured pruning with extreme sparsity targets, we end up pruning entire neurons, which effectively detaches neurons from the input/output. The authors argue that the effective sparsity levels in these cases is higher than our target levels, and taking it into accounts exhibits differences between various pruning method that are otherwise not obvious. ", "review_text": "While indirect pruning is a reasonably known phenomenon, it is an interesting question to see whether it can give us any insights when comparing various pruning methods. Unfortunately, this paper doesn't do a thorough job doing that due to choice of baselines and poor motivation of extreme pruning regime:\n\n- I am surprised to see LeNet-300-100 as baselines here. I do not want to be the \"show me imagenet results\" reviewer, but I feel the choice of MNIST, and LeNet architectures are not justified here. Indirect pruning happens in extreme pruning regimes and I am assuming we care about that for very, very large networks. LeNet-300-100 is tiny and does not have convolutions, so is this a good baseline to draw conclusions from?\n- In the introduction, the authors do say that typical compression rates are 10x (0.9 pruning) and 100x (0.99 pruning) and motivate extreme pruning because of models with billions of parameters, yet their biggest models are ResNet-18 and VGG-19, where indirect pruning does not seem to be as big of a problem anyway.\n- Overall, the authors should do a better job motivating extreme pruning regimes. Pruning is interesting because we think models are over-parameterized, and we can get \"comparable performance\" while pruning away a good portion of parameters. However, the performance in this extreme regime is very poor, and the architectures are simply \"trainable\". So why is this an interesting problem to look at?\n\n**Structured pruning and CNNs**\n- I think the authors should also clarify the link between indirect pruning and structured pruning. Isn't the phenomena described by the authors the intended behaviour in structured pruning? It would be great if authors could also compare with structured pruning methods but I think at least the link between the two should be discussed in the paper.\n- The authors do not discuss what happens to the disconnected neurons as a result of indirect pruning. Can you set their value to zero (i.e. actually prune them)? Can their value always be fused into the next layer? or perhaps into the batch norm parameters? \n- The authors also don't discuss convolutional networks in details. I'm assuming indirect pruning is less of a problem there because you'd have to prune an entire channel to have indirect pruning. Given that most of our commonly used architectures and other pruning baselines use CNNs it seems crucial to have it analysed in detail and not rely on MLPs to make conclusions\n\n**Layerwise Sparsity Quotas (LSQ) and allocation method IGQ**\n\nI don't understand the Ideal Gas Law approach at all! What is the motivation for borrowing this from thermodynamics? What is the justification that neural networks behave like gasses? It is one thing to mention this in passing as an inspiration source, but when you go as far as mapping a neural network to a system in thermodynamic equilibrium, the height of a cylinder to the number of weights, and pruning to applying an external force to then you should provide better justifications. Moreover, the final derivation for prune ratio of each layer is $F | \\Theta _l | + 1$. There are two things I don't understand about this: (1) How is $F$ related to the desired pruning target? The authors say $F$ can be found via a binary search from pruning target but do not elaborate. (2) No matter how F is found I'm assuming it's a global value and not a per-layer one. Doesn't this reduce the formula to simply applying pruning uniformly??\n\n**Other things**\n- The authors use VGG19. I have seen different variants of VGG-19 for non-imagenet datasets. In Some variants there's a single dense layer after convolutional layers, and in some other variants there are multiple dense layers. Could the authors confirm which architecture they've used? This makes a big difference in prunability of the network and I'm curious to know if the indirect pruning is mostly happening in the final dense layers or in convolutions as well.\n- In the discussion section the authors argue that \"effective compression\" is the \"correct\" measure for pruning algorithm. I think that's a strong claim that's not well supported in the paper. You should at least clarify that it's aimed at extreme pruning regime (and probably non CNN based architectures?\n- The caption in Figure 2 says \"SynFlow has a better sparsity-accuracy tradeoff than SNIP\" Isn't the opposite true? I guess it's a typo\n- In figure 3, would things look too weird if you use the same units in the x axis and y axis so that we can compare the effect across architectures more easily?\n- In figure 5 could you rename \"Random\" to \"SNIP + RandomReshuffle\" or something. I think random is misleading here given that you're still applying layer-wise pruning ratios.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper highlights the fact that when doing unstructured pruning with extreme sparsity targets, we end up pruning entire neurons, which effectively detaches neurons from the input/output. The authors argue that the effective sparsity levels in these cases is higher than our target levels, and taking it into accounts exhibits differences between various pruning method that are otherwise not obvious. ", "main_review": "While indirect pruning is a reasonably known phenomenon, it is an interesting question to see whether it can give us any insights when comparing various pruning methods. Unfortunately, this paper doesn't do a thorough job doing that due to choice of baselines and poor motivation of extreme pruning regime:\n\n- I am surprised to see LeNet-300-100 as baselines here. I do not want to be the \"show me imagenet results\" reviewer, but I feel the choice of MNIST, and LeNet architectures are not justified here. Indirect pruning happens in extreme pruning regimes and I am assuming we care about that for very, very large networks. LeNet-300-100 is tiny and does not have convolutions, so is this a good baseline to draw conclusions from?\n- In the introduction, the authors do say that typical compression rates are 10x (0.9 pruning) and 100x (0.99 pruning) and motivate extreme pruning because of models with billions of parameters, yet their biggest models are ResNet-18 and VGG-19, where indirect pruning does not seem to be as big of a problem anyway.\n- Overall, the authors should do a better job motivating extreme pruning regimes. Pruning is interesting because we think models are over-parameterized, and we can get \"comparable performance\" while pruning away a good portion of parameters. However, the performance in this extreme regime is very poor, and the architectures are simply \"trainable\". So why is this an interesting problem to look at?\n\n**Structured pruning and CNNs**\n- I think the authors should also clarify the link between indirect pruning and structured pruning. Isn't the phenomena described by the authors the intended behaviour in structured pruning? It would be great if authors could also compare with structured pruning methods but I think at least the link between the two should be discussed in the paper.\n- The authors do not discuss what happens to the disconnected neurons as a result of indirect pruning. Can you set their value to zero (i.e. actually prune them)? Can their value always be fused into the next layer? or perhaps into the batch norm parameters? \n- The authors also don't discuss convolutional networks in details. I'm assuming indirect pruning is less of a problem there because you'd have to prune an entire channel to have indirect pruning. Given that most of our commonly used architectures and other pruning baselines use CNNs it seems crucial to have it analysed in detail and not rely on MLPs to make conclusions\n\n**Layerwise Sparsity Quotas (LSQ) and allocation method IGQ**\n\nI don't understand the Ideal Gas Law approach at all! What is the motivation for borrowing this from thermodynamics? What is the justification that neural networks behave like gasses? It is one thing to mention this in passing as an inspiration source, but when you go as far as mapping a neural network to a system in thermodynamic equilibrium, the height of a cylinder to the number of weights, and pruning to applying an external force to then you should provide better justifications. Moreover, the final derivation for prune ratio of each layer is $F | \\Theta _l | + 1$. There are two things I don't understand about this: (1) How is $F$ related to the desired pruning target? The authors say $F$ can be found via a binary search from pruning target but do not elaborate. (2) No matter how F is found I'm assuming it's a global value and not a per-layer one. Doesn't this reduce the formula to simply applying pruning uniformly??\n\n**Other things**\n- The authors use VGG19. I have seen different variants of VGG-19 for non-imagenet datasets. In Some variants there's a single dense layer after convolutional layers, and in some other variants there are multiple dense layers. Could the authors confirm which architecture they've used? This makes a big difference in prunability of the network and I'm curious to know if the indirect pruning is mostly happening in the final dense layers or in convolutions as well.\n- In the discussion section the authors argue that \"effective compression\" is the \"correct\" measure for pruning algorithm. I think that's a strong claim that's not well supported in the paper. You should at least clarify that it's aimed at extreme pruning regime (and probably non CNN based architectures?\n- The caption in Figure 2 says \"SynFlow has a better sparsity-accuracy tradeoff than SNIP\" Isn't the opposite true? I guess it's a typo\n- In figure 3, would things look too weird if you use the same units in the x axis and y axis so that we can compare the effect across architectures more easily?\n- In figure 5 could you rename \"Random\" to \"SNIP + RandomReshuffle\" or something. I think random is misleading here given that you're still applying layer-wise pruning ratios.", "summary_of_the_review": "This paper draws too many conclusions based on very small networks, datasets, and for a pruning regime where the network is simply trainable and has nowhere near the performance of the unpruned model. Overall I'm not convinced that \"effective pruning\", and extreme pruning gives us a enough insight to justify accepting the paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635986357841}, {"id": "78D5thnfHKs", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3938/Reviewer_B4nR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper raises a warning signal on an aspect which is typically ignored in sparse neural networks, i.e., the “unconnected” connections (the ones which do not belong to the paths connecting the input to the output neurons). Consequently, the paper studies the so-called “effective sparsity”, to formally understand the efficiency of various pruning methods to obtain sparse neural networks while considering the connections which contribute to the neural network inference (the ones which do belong to the paths connecting the input to the output).  Complementary, the paper proposes some (minor) contributions to cope with effective sparsity. Overall, the paper is well written and has well-designed experiments. \n\n", "review_text": "**Strong points:**\n\n- From my experience, I recognize this problem of “effective sparsity” and, up to my best knowledge, this is the first time when I see it so well formulated. Thus, the paper was a pleasure to read, providing useful insights in understanding sparse neural networks and pruning techniques.\n\n- Few small innovations are introduced to cope with the concept of “effective sparsity” in order to improve the trade-off model size/accuracy\n\n- The code is provided for easy reproducibility\n\n**Weak points and suggestions for improvement:**\n\n- The paper contributions are hard to derive from the actual text. I suggest adding somewhere in the Introduction section, few short and concise bullet points to present the paper contributions and to sharply discuss what is novel with respect to the literature.\n\n- The Related Work section contains just a categorization of various pruning techniques. I suggest to add an extra paragraph which clearly discuss the related work with respect to the “effective sparsity” concept.\n\n- The experiments are performed just on dense-to-sparse training methods. I suggest considering also sparse-to-sparse training methods with prune and grow strategies. I agree with the authors statement from the Discussion section that it would be more difficult to incorporate their proposed effective pruning strategy in sparse training. This incorporation can be indeed let for future work, but just studying effective sparsity in sparse training should not be so difficult and the results can be of interest (e.g., same as in Figure 3). \n\n- I understand the ERK baseline for the proposed IGQ method but decoupling ERK from sparse training (as it was originally designed) may lead to misleading results interpretation. Nothing to change in the experimental section, but I suggest clarifying (discussing better) this aspect.\n\n- It is not very clear to me if the paper suggests that the connectivity pattern itself is also quite important in ensuring a good trade-off model size/ performance and not just the sparsity distribution. A discussion has been started on this topic in the last paragraph of page 2, but I would expect the Discussion section to come back to this topic and perhaps some empirical validation would be necessarily to support better the statement “…but find the truth to be more nuanced at higher compression rates…”. I wouldn’t expect a detailed study as it is outside of the scope of this paper, but rather some hints in order to avoid cutting from start future works on this topic. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper raises a warning signal on an aspect which is typically ignored in sparse neural networks, i.e., the “unconnected” connections (the ones which do not belong to the paths connecting the input to the output neurons). Consequently, the paper studies the so-called “effective sparsity”, to formally understand the efficiency of various pruning methods to obtain sparse neural networks while considering the connections which contribute to the neural network inference (the ones which do belong to the paths connecting the input to the output).  Complementary, the paper proposes some (minor) contributions to cope with effective sparsity. Overall, the paper is well written and has well-designed experiments. \n\n", "main_review": "**Strong points:**\n\n- From my experience, I recognize this problem of “effective sparsity” and, up to my best knowledge, this is the first time when I see it so well formulated. Thus, the paper was a pleasure to read, providing useful insights in understanding sparse neural networks and pruning techniques.\n\n- Few small innovations are introduced to cope with the concept of “effective sparsity” in order to improve the trade-off model size/accuracy\n\n- The code is provided for easy reproducibility\n\n**Weak points and suggestions for improvement:**\n\n- The paper contributions are hard to derive from the actual text. I suggest adding somewhere in the Introduction section, few short and concise bullet points to present the paper contributions and to sharply discuss what is novel with respect to the literature.\n\n- The Related Work section contains just a categorization of various pruning techniques. I suggest to add an extra paragraph which clearly discuss the related work with respect to the “effective sparsity” concept.\n\n- The experiments are performed just on dense-to-sparse training methods. I suggest considering also sparse-to-sparse training methods with prune and grow strategies. I agree with the authors statement from the Discussion section that it would be more difficult to incorporate their proposed effective pruning strategy in sparse training. This incorporation can be indeed let for future work, but just studying effective sparsity in sparse training should not be so difficult and the results can be of interest (e.g., same as in Figure 3). \n\n- I understand the ERK baseline for the proposed IGQ method but decoupling ERK from sparse training (as it was originally designed) may lead to misleading results interpretation. Nothing to change in the experimental section, but I suggest clarifying (discussing better) this aspect.\n\n- It is not very clear to me if the paper suggests that the connectivity pattern itself is also quite important in ensuring a good trade-off model size/ performance and not just the sparsity distribution. A discussion has been started on this topic in the last paragraph of page 2, but I would expect the Discussion section to come back to this topic and perhaps some empirical validation would be necessarily to support better the statement “…but find the truth to be more nuanced at higher compression rates…”. I wouldn’t expect a detailed study as it is outside of the scope of this paper, but rather some hints in order to avoid cutting from start future works on this topic. \n", "summary_of_the_review": "Overall, I believe that this is a well-written paper, with a clear message, which also has some space for improvement.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635886455624}, {"id": "VtEAE5EP-vK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3938/Reviewer_xGEE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper points out that in pruning, there is often a difference between the number of zero parameters and the number effective zero parameters due to disconnection during pruning. A thorough analysis is done comparing many pruning methods across MNIST, Cifar-10, and TinyImageNet using various architectures. The authors introduce a new physics inspired layerwise sparsity quota  baseline the \"Ideal Gas quota\" which sets a sparsity per layer. ", "review_text": "Strengths:\n* The paper is clear and presents an interesting analysis about the discrepancy between pruning models and the resulting, underlying graph.\n* The detail and analysis is thorough, and the methodology is clear and concise. \n* The presented layer-wise sparsity quota Ideal Gas Quotas (IGQ) are interesting and produce compression distributions very similar to other methods (such as in Fig. 9). \nWeaknesses:\n* Many important points/figures in the paper are on very small datasets where models are extremely over-paramaterised. For example, Fig. 2 on MNIST would be much stronger, I think, on a slightly larger dataset (CIFAR-10, perhaps?)\n* Overall, I found the results largely focused on models that do not reach competitive performance on larger datasets. ResNet18 on TinyImageNet is shown, but it'd be nice if more modern architectures were shown. \n\nComments:\n* Would it be possible to adjust for effective FLOPs in a correctly pruned graph?\n* It would be cool to investigate various choices and their impact on sparsity/effective sparsity during pruning. For example, I imagine models with residual connections will have slightly different dynamics. \n* I'd be interested in a bit of a comparison of different model architectures-- for example, sparsity in convolutional filters will have a different impact in the effective sparsity metric than sparsity in MLPs (if I understand correctly). Therefore, there will be some differences between architectures that allocate parameters differently (and what is pruned). \n* Showing how some of these methods look in another domain such as NLP could be interesting. \n\nMinor comments: \n* I found the caption to Figure 3 slightly confusing. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper points out that in pruning, there is often a difference between the number of zero parameters and the number effective zero parameters due to disconnection during pruning. A thorough analysis is done comparing many pruning methods across MNIST, Cifar-10, and TinyImageNet using various architectures. The authors introduce a new physics inspired layerwise sparsity quota  baseline the \"Ideal Gas quota\" which sets a sparsity per layer. ", "main_review": "Strengths:\n* The paper is clear and presents an interesting analysis about the discrepancy between pruning models and the resulting, underlying graph.\n* The detail and analysis is thorough, and the methodology is clear and concise. \n* The presented layer-wise sparsity quota Ideal Gas Quotas (IGQ) are interesting and produce compression distributions very similar to other methods (such as in Fig. 9). \nWeaknesses:\n* Many important points/figures in the paper are on very small datasets where models are extremely over-paramaterised. For example, Fig. 2 on MNIST would be much stronger, I think, on a slightly larger dataset (CIFAR-10, perhaps?)\n* Overall, I found the results largely focused on models that do not reach competitive performance on larger datasets. ResNet18 on TinyImageNet is shown, but it'd be nice if more modern architectures were shown. \n\nComments:\n* Would it be possible to adjust for effective FLOPs in a correctly pruned graph?\n* It would be cool to investigate various choices and their impact on sparsity/effective sparsity during pruning. For example, I imagine models with residual connections will have slightly different dynamics. \n* I'd be interested in a bit of a comparison of different model architectures-- for example, sparsity in convolutional filters will have a different impact in the effective sparsity metric than sparsity in MLPs (if I understand correctly). Therefore, there will be some differences between architectures that allocate parameters differently (and what is pruned). \n* Showing how some of these methods look in another domain such as NLP could be interesting. \n\nMinor comments: \n* I found the caption to Figure 3 slightly confusing. ", "summary_of_the_review": "The paper is well written, very clear, and well presented. The authors raise and clearly document the discrepancy between raw sparsity and the actual effective sparsity of a model. An interesting and compelling new baseline is presented.\nHowever, the results are shown on small datasets without comparison with many modern architectures. \nReplacing the MNIST/LeNet/VGG results with more modern architectures would clearly increase the impact of the paper.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635812439740}, {"id": "41j5viIyOu1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3938/Reviewer_Ne1A"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper highlights the difference between the explicit and effective (through disconnection) number of zeroed parameters in a model --- and argues it is the effective parameters that should be used as the benchmark for comparison.\nFurther comparing pruning methods and re-affirming their performance relative to random pruning equivalents. Finally introducing a layer allocation method which can serve as a better performing random baseline.\n", "review_text": "\nOverview:\n\nThe issue of effectively disconnected parameters [Tenaka et. al.], while known, beckons the proper methodological accounting of parameters --- to those 'effectively' contributing. This paper describes this discrepancy and quantifies it at small scale.\nWhile disconnections are known, and to this reviewer at least.--- known to be quantified by practitioners --- such quantification is indeed poorly documented and this paper eloquently captures it. Beyond the methodological importance of such explicit clarification, the practical relevance and expectation of their impact could use further discussion and motivation.\n\nMajor Pros:\n1. The issue of proper methodology in the context of counting of parameters is important and by extension so is accounting for effectively defunct parameters due to disconnection. \n2. The paper is well written and clear.\n\nMajor Cons:\n1. This discrepancy (between effective and direct parameter count) is known (as the authors mention [e.g. Tenaka et al]) and thus the major contribution of this paper is of limited novelty.\n\n2. While the actual methodological claim (one should account for disconnections via effective counting) is very much valid and orthogonal to experimental settings, the practical discrepancy (and thus implications) between direct/effective is very much dependent on the settings:\n\n2.a. As mentioned by the authors (and demonstrated in figure 3), for the iterative pruning of Syntflow and SNIP-iter, no disconnection associated discrepancy appears. Tanaka et al single out the general role of the Iterative nature in upholding continuity even in methods which are not explicitly designed to preserve continuity.\nIn light of such an observation --- for SOTA, such as Iterative Magnitude Pruning (IMP) , it is not clear that in practice a large discrepancy would be present even up to higher than practically employed sparsity levels. \n\n2.b. Further, in terms of architectures, probability of disconnection in modern architectures --- and especially those containing residual connections --- is much reduced. Indeed, in the Resnet experiments (figure 4) there seems to be a relatively small difference between effective and direct. As such, for practical architectures --- the level of actual discrepancy in practice is not demonstrated to be (or as expected to be) meaningful. \n\n2.c. The paper demonstrated the results for small datasets (and weak networks) (mnist/lenet , VGG/cifar10-100, Resnet18/TinyImagenet). \nThe lack of practical large scale (Imagenet) and SOTA architectures limits the conclusions to the applicability.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper highlights the difference between the explicit and effective (through disconnection) number of zeroed parameters in a model --- and argues it is the effective parameters that should be used as the benchmark for comparison.\nFurther comparing pruning methods and re-affirming their performance relative to random pruning equivalents. Finally introducing a layer allocation method which can serve as a better performing random baseline.\n", "main_review": "\nOverview:\n\nThe issue of effectively disconnected parameters [Tenaka et. al.], while known, beckons the proper methodological accounting of parameters --- to those 'effectively' contributing. This paper describes this discrepancy and quantifies it at small scale.\nWhile disconnections are known, and to this reviewer at least.--- known to be quantified by practitioners --- such quantification is indeed poorly documented and this paper eloquently captures it. Beyond the methodological importance of such explicit clarification, the practical relevance and expectation of their impact could use further discussion and motivation.\n\nMajor Pros:\n1. The issue of proper methodology in the context of counting of parameters is important and by extension so is accounting for effectively defunct parameters due to disconnection. \n2. The paper is well written and clear.\n\nMajor Cons:\n1. This discrepancy (between effective and direct parameter count) is known (as the authors mention [e.g. Tenaka et al]) and thus the major contribution of this paper is of limited novelty.\n\n2. While the actual methodological claim (one should account for disconnections via effective counting) is very much valid and orthogonal to experimental settings, the practical discrepancy (and thus implications) between direct/effective is very much dependent on the settings:\n\n2.a. As mentioned by the authors (and demonstrated in figure 3), for the iterative pruning of Syntflow and SNIP-iter, no disconnection associated discrepancy appears. Tanaka et al single out the general role of the Iterative nature in upholding continuity even in methods which are not explicitly designed to preserve continuity.\nIn light of such an observation --- for SOTA, such as Iterative Magnitude Pruning (IMP) , it is not clear that in practice a large discrepancy would be present even up to higher than practically employed sparsity levels. \n\n2.b. Further, in terms of architectures, probability of disconnection in modern architectures --- and especially those containing residual connections --- is much reduced. Indeed, in the Resnet experiments (figure 4) there seems to be a relatively small difference between effective and direct. As such, for practical architectures --- the level of actual discrepancy in practice is not demonstrated to be (or as expected to be) meaningful. \n\n2.c. The paper demonstrated the results for small datasets (and weak networks) (mnist/lenet , VGG/cifar10-100, Resnet18/TinyImagenet). \nThe lack of practical large scale (Imagenet) and SOTA architectures limits the conclusions to the applicability.\n\n", "summary_of_the_review": "The paper eloquently presents and quantifies --- in the small scale --- the known issue of disconnections as affecting parameter count and pruning methods [Tenaka et al]. The paper can further engage with practical implications. \nI agree that counting effective rather than direct parameters is methogologically good practice. However, the limited scale and demonstration  limits the insight applicability in practice --- especially, as SOTA methods and architectures have structural characteristics which call into question the the level of discrepancy (missing discussion/dimensions of exploration).\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635731841886}], "openreview_url": "https://openreview.net/forum?id=YgR1rRWETI", "arxiv_id": "2107.02306", "paper_pdf": "papers/YgR1rRWETI.pdf", "paper_pdf_sha256": "10f57b347cb4a1cefab8153d608b1e66dad556eaf074acd2e6ad1b5b19bbc44c", "paper_pdf_bytes": 29691568, "paper_pdf_source": "openreview", "code_url": "https://github.com/avysogorets/effective-sparsity", "code_repository": "avysogorets/effective-sparsity", "code_commit": "7924cfe8a26106cbf65c3a760aa352e58765faf3", "code_archive": "repos/YgR1rRWETI.zip", "code_archive_sha256": "b70200f6aaba6cd0defc322ff34d5ef41b95371e5327d4588d1f496bef8ec7f8", "code_archive_bytes": 22066, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 185, "github_languages": {"Python": 106117}, "github_archived": false, "github_pushed_at": "2022-03-15T04:41:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/connectivity-matters-neural-network-pruning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Mf4ZSXMZP7", "year": 2021, "status": "rejected", "title": "Improving Post Training Neural Quantization: Layer-wise Calibration and Integer Programming", "authors": ["Itay Hubara", "Yury Nahshan", "Yair Hanani", "Ron Banner", "Daniel Soudry"], "authorids": ["~Itay_Hubara1", "~Yury_Nahshan1", "~Yair_Hanani1", "~Ron_Banner1", "~Daniel_Soudry1"], "authors_source": "OpenReview API", "abstract": "Lately, post-training quantization methods have gained considerable attention, as they are simple to use, and require only a small unlabeled calibration set. This small dataset cannot be used to fine-tune the model without significant over-fitting. Instead, these methods only use the calibration set to set the activations' dynamic ranges. However, such methods always resulted in significant accuracy degradation, when used below 8-bits (except on small datasets). Here we aim to break the 8-bit barrier. To this end, we minimize the quantization errors of each layer separately by optimizing its parameters over the calibration set.  We empirically demonstrate that this approach is: (1) much less susceptible to over-fitting than the standard fine-tuning approaches, and can be used even on a very small calibration set; and (2) more powerful than previous methods, which only set the activations' dynamic ranges. Furthermore, we demonstrate how to optimally allocate the bit-widths for each layer, while constraining accuracy degradation or model compression by proposing a novel integer programming formulation. Finally, we suggest model global statistics tuning, to correct biases introduced during quantization. Together, these methods yield state-of-the-art results for both vision and text models. For instance, on ResNet50, we obtain less than 1\\% accuracy degradation --- with 4-bit weights and activations in all layers, but the smallest two. Our code is available at, https://github.com/papers-submission/CalibTIP", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "gKjagp4qWvK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1562/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper proposed a set of methods for post-training quantization of dnns. The methods include AdaQuant (which jointly optimizes quantization steps for weight and activation per output activation of each layer), Integer Programming (which determines bit-precision for all the layers), and the batchnorm tuning. The authors presented promising experimental results on various neural networks to support the proposed methods. \n\nHowever, there are serious concerns about these claims as follows:\n1) AdaQuant\n- It is straightforward to think that the joint optimization of quantization step size for weight and activation would result in better quantization results. But this joint optimization would also increase the search space (at least) quadratically, resulting in significant computational cost. Note that the biggest merit of post-training quantization is its simplicity (cf., QAT incurs full-blown training epochs); thus increased cost for post-training quantization is not desirable. Since AdaQuant is the major claim, the authors should provide more discussion on how they dealt with this increased complexity.\n\n- The authors claim that AdaQuant avoids overfitting, but the reason does not seem to be clear. There is no clear explanation of how AdaQuant increases the generality of the quantized model, and the discussion about the sample size (B) is hard to understand (why there's infinite solution when B << N? how B>= Ck^2/(HW) is derived for the convolution case?)\n\n- Also, it seems that the \"per-channel\" quantization method is utilized in this work, but the formulation in (2) seems to be for \"per-layer\" optimization. Why they are different?\n\n- How much time does it take to solve this joint optimization?\n\n\n2) Integer Programming\n- The authors proposed an Integer Programming formulation, but there seem to be missing information: what is the formulation of the penalty function, \"deltaL\"? The authors described it simply as \"Loss\", but it is not clear what the exact method it is calculated. In fact, deltaL can be pretty complex functions, which might not be independent terms for each layer; thus the formulation like (3) might not be correct. Note that the impact of quantization in the earlier layers affect the quantization impact in the current layer. Without clear explanation and justification about it, the proposed IP formulation does not make sense.\n\n- The authors mentioned that deltaP should be additive and sum up to the total benefit. How can one guarantee it? \n\n- Also, it seems that the complexity of the IP optimization increases as the number of layers increases. How much computation time increases if the number of layers are large? \n\n\n3) Batch normalization tuning\n- Unfortunately, there is a very similar idea proposed by [Sun et al., NeurIPS 19]. Cf. \"Sec.3 Trans-Precision Inference in FP8\".\n\n\nAlso, there are several suggestions to improve understanding of readers.\n- Currently, the ablation study looks very confusing. It is not clear which of the pipeline options (light, advanced?) include what kinds of techniques. Please do specify (maybe in a separate table) the list of techniques covered by different pipeline options. \n\n- The proposed method is not much evaluated by various neural networks. It would be desirable to expand the coverage of neural nets as much as the prior work did. \n\n- Currently, the proposed methods only utilized \"per-channel\" quantization. How much accuracy the proposed methods can maintain if they adopt \"per-layer\" quantization?\n\n- What is the definition of \"compression ratio\"? (typically compration RATIO is like 12:1, and compression rate is like 2X, 3X...)\n\n- \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Confusing claims with lack of details and limited novelty", "review": "\nThis paper proposed a set of methods for post-training quantization of dnns. The methods include AdaQuant (which jointly optimizes quantization steps for weight and activation per output activation of each layer), Integer Programming (which determines bit-precision for all the layers), and the batchnorm tuning. The authors presented promising experimental results on various neural networks to support the proposed methods. \n\nHowever, there are serious concerns about these claims as follows:\n1) AdaQuant\n- It is straightforward to think that the joint optimization of quantization step size for weight and activation would result in better quantization results. But this joint optimization would also increase the search space (at least) quadratically, resulting in significant computational cost. Note that the biggest merit of post-training quantization is its simplicity (cf., QAT incurs full-blown training epochs); thus increased cost for post-training quantization is not desirable. Since AdaQuant is the major claim, the authors should provide more discussion on how they dealt with this increased complexity.\n\n- The authors claim that AdaQuant avoids overfitting, but the reason does not seem to be clear. There is no clear explanation of how AdaQuant increases the generality of the quantized model, and the discussion about the sample size (B) is hard to understand (why there's infinite solution when B << N? how B>= Ck^2/(HW) is derived for the convolution case?)\n\n- Also, it seems that the \"per-channel\" quantization method is utilized in this work, but the formulation in (2) seems to be for \"per-layer\" optimization. Why they are different?\n\n- How much time does it take to solve this joint optimization?\n\n\n2) Integer Programming\n- The authors proposed an Integer Programming formulation, but there seem to be missing information: what is the formulation of the penalty function, \"deltaL\"? The authors described it simply as \"Loss\", but it is not clear what the exact method it is calculated. In fact, deltaL can be pretty complex functions, which might not be independent terms for each layer; thus the formulation like (3) might not be correct. Note that the impact of quantization in the earlier layers affect the quantization impact in the current layer. Without clear explanation and justification about it, the proposed IP formulation does not make sense.\n\n- The authors mentioned that deltaP should be additive and sum up to the total benefit. How can one guarantee it? \n\n- Also, it seems that the complexity of the IP optimization increases as the number of layers increases. How much computation time increases if the number of layers are large? \n\n\n3) Batch normalization tuning\n- Unfortunately, there is a very similar idea proposed by [Sun et al., NeurIPS 19]. Cf. \"Sec.3 Trans-Precision Inference in FP8\".\n\n\nAlso, there are several suggestions to improve understanding of readers.\n- Currently, the ablation study looks very confusing. It is not clear which of the pipeline options (light, advanced?) include what kinds of techniques. Please do specify (maybe in a separate table) the list of techniques covered by different pipeline options. \n\n- The proposed method is not much evaluated by various neural networks. It would be desirable to expand the coverage of neural nets as much as the prior work did. \n\n- Currently, the proposed methods only utilized \"per-channel\" quantization. How much accuracy the proposed methods can maintain if they adopt \"per-layer\" quantization?\n\n- What is the definition of \"compression ratio\"? (typically compration RATIO is like 12:1, and compression rate is like 2X, 3X...)\n\n- \n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604961195559}, {"id": "LGgSIt4lRt0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1562/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose to use many techniques to push the limit of neural quantization, which shows reasonable improvements in some datasets.\n\n+ good performance.\n+ clear presentation.\n+ easy to read and follow\n\nMy main complaint is that this type of combination is better for a technical report rather than a standalone paper. In more detail, the authors claim four proposed components: AdaQuant, Integer programming, Batch-norm tuning, and two pipelines for neural quantization. However, the problem is, are these four components all original or just adapted from existing works? The authors are failed to present the relation of the proposed components with the exitinging ones. For example, in my understanding, the AdaQuant is only an adapted version of AdaRound, but the discussion about it (in Section 3.1) is too superficial. Furthermore, the experiment part is not convincing and I think can not back up the claim. With so many \"proposed components\", the missing of thorough ablation study is a big problem.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall, the authors propose to use a combination of different twisted techniques for better neural quantization, which lacks meaningful insights.", "review": "The authors propose to use many techniques to push the limit of neural quantization, which shows reasonable improvements in some datasets.\n\n+ good performance.\n+ clear presentation.\n+ easy to read and follow\n\nMy main complaint is that this type of combination is better for a technical report rather than a standalone paper. In more detail, the authors claim four proposed components: AdaQuant, Integer programming, Batch-norm tuning, and two pipelines for neural quantization. However, the problem is, are these four components all original or just adapted from existing works? The authors are failed to present the relation of the proposed components with the exitinging ones. For example, in my understanding, the AdaQuant is only an adapted version of AdaRound, but the discussion about it (in Section 3.1) is too superficial. Furthermore, the experiment part is not convincing and I think can not back up the claim. With so many \"proposed components\", the missing of thorough ablation study is a big problem.", "rating": "4: Ok but not good enough - rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603902878908}, {"id": "3PQxXLv3lW6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1562/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces a series of techniques to quantize neural networks, and how to combine them:\n* Layer by layer quantization where weights can change as needed (rather than to the nearest quantization error).\n* Integer programming to determine the precision required at every layer.\n* Tuning batch norm weights by re-computing statistics.\n\nPROS\n\n* Main text is well written (see below for other issues though).\n* All components in the proposed method are straightforward.\n* Strong results, with very little performance loss after very aggressive quantization (~4 bits).\n\nCONS\n\n* The organization of the paper can and should be easily improved (see below).\n* Very basic details, such as the metric used in some tables, are omitted.\n* Few experiments/baselines.\n\n---\n\nOn the flow of the paper: Fig. 1 is barely introduced. Fig. 2 is introduced much earlier than it is first referred to in the text, in section 5.2, where it is not explained either. The labels in the figure are not defined either at this point (and ip should be IP). Additionally, the colors for the last two lines are too similar.\n\nThe experiment of Fig.1  should be introduced with some level of detail (there is none outside the caption). The caption is a bit confusing: what are the \"the last layers after\"? Why are variance bars always red? They are also hard to see. The labels are not descriptive (light pipeline, advanced pipeline, relaxed advanced pipeline) and hard to find in the text. MN-V2 and B-SQuad1.1 just drop in Table 1 without any context. It can be inferred from earlier sections what this refers to but it should be explained clearly. Same for \"min-max\" in Fig. 1 and Table 1.\n\nBaselines are barely discussed. Also, I am not very familiar with quantization papers, so I might have missed relevant baselines, but they seem hard to compare. For instance, the authors omit [Nagel et al 2020](https://arxiv.org/pdf/2004.10568.pdf), which seems to do better at similar quantization levels, but I am not sure the results are directly comparable. The authors should discuss this better.\n\nI think section 4 can be titled simply \"Quantization flow\".\n\nCompression ratio in the plots should indicate %?\n\n---\n\nTypos/grammar:\n\n* \"Mobilsnet-V2\" -> \"Mobilenet-V2\n* \"we suggest [an] integer-linear programming\"\n* \"AdaRound['s] implicit constraint\"\n* \"MSE distance\": shouldn't this just be MSE?\n* \"Early work by Lin et al. (2016) used [a] convex optimization formulation which results ~~with~~ [in] a simple greedy compression scheme.\"\n* \"3. OPTIMIZING [THE] QUANTIZATION PIP[E]LINE\"\n* \"model['s] internal statistic[s]\" -> found twice\n* \"often result with an inferior solution\" -> \"often result in an inferior solution\"\n* \"Accordingly, researches suggested\" -> researchers?\n* \"where V is a continuous variable V\" -> redundant\n* \"thus enjoy[ing] some of the flexibility\"? -> \"benefitting from\" might be a better phrasing?\n* \"Quantization-Aware- Training\" -> extra space\n* \"and [is] much less prone to over-fitting\"\n* \"[A] Similar derivation\"\n* \"Even optimizing on [a] single image\"\n* \"MAC operations.\" -> acronym not previously introduced (unless mistaken)\n* Section 4: IP acronym should be introduced in the integer programming section.\n* \"we investigate [a] mixture\"\n* \"results with high degradation\" -> in high degradation\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "See below", "review": "The paper introduces a series of techniques to quantize neural networks, and how to combine them:\n* Layer by layer quantization where weights can change as needed (rather than to the nearest quantization error).\n* Integer programming to determine the precision required at every layer.\n* Tuning batch norm weights by re-computing statistics.\n\nPROS\n\n* Main text is well written (see below for other issues though).\n* All components in the proposed method are straightforward.\n* Strong results, with very little performance loss after very aggressive quantization (~4 bits).\n\nCONS\n\n* The organization of the paper can and should be easily improved (see below).\n* Very basic details, such as the metric used in some tables, are omitted.\n* Few experiments/baselines.\n\n---\n\nOn the flow of the paper: Fig. 1 is barely introduced. Fig. 2 is introduced much earlier than it is first referred to in the text, in section 5.2, where it is not explained either. The labels in the figure are not defined either at this point (and ip should be IP). Additionally, the colors for the last two lines are too similar.\n\nThe experiment of Fig.1  should be introduced with some level of detail (there is none outside the caption). The caption is a bit confusing: what are the \"the last layers after\"? Why are variance bars always red? They are also hard to see. The labels are not descriptive (light pipeline, advanced pipeline, relaxed advanced pipeline) and hard to find in the text. MN-V2 and B-SQuad1.1 just drop in Table 1 without any context. It can be inferred from earlier sections what this refers to but it should be explained clearly. Same for \"min-max\" in Fig. 1 and Table 1.\n\nBaselines are barely discussed. Also, I am not very familiar with quantization papers, so I might have missed relevant baselines, but they seem hard to compare. For instance, the authors omit [Nagel et al 2020](https://arxiv.org/pdf/2004.10568.pdf), which seems to do better at similar quantization levels, but I am not sure the results are directly comparable. The authors should discuss this better.\n\nI think section 4 can be titled simply \"Quantization flow\".\n\nCompression ratio in the plots should indicate %?\n\n---\n\nTypos/grammar:\n\n* \"Mobilsnet-V2\" -> \"Mobilenet-V2\n* \"we suggest [an] integer-linear programming\"\n* \"AdaRound['s] implicit constraint\"\n* \"MSE distance\": shouldn't this just be MSE?\n* \"Early work by Lin et al. (2016) used [a] convex optimization formulation which results ~~with~~ [in] a simple greedy compression scheme.\"\n* \"3. OPTIMIZING [THE] QUANTIZATION PIP[E]LINE\"\n* \"model['s] internal statistic[s]\" -> found twice\n* \"often result with an inferior solution\" -> \"often result in an inferior solution\"\n* \"Accordingly, researches suggested\" -> researchers?\n* \"where V is a continuous variable V\" -> redundant\n* \"thus enjoy[ing] some of the flexibility\"? -> \"benefitting from\" might be a better phrasing?\n* \"Quantization-Aware- Training\" -> extra space\n* \"and [is] much less prone to over-fitting\"\n* \"[A] Similar derivation\"\n* \"Even optimizing on [a] single image\"\n* \"MAC operations.\" -> acronym not previously introduced (unless mistaken)\n* Section 4: IP acronym should be introduced in the integer programming section.\n* \"we investigate [a] mixture\"\n* \"results with high degradation\" -> in high degradation\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603896776746}, {"id": "29f6RlrgQKc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1562/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work presents a quite comprehensive multi-step scheme for post-training neural quantization that does not rely on large datasets or large computational resources.  \n\nThe work is has significance in the domain of post-training neural quantization, especially in cases where only a small calibration set or limited resources are available. It would be interesting to also think about accumulator quantization.\n\n\nPros:\n\nThe empirical results are relatively strong in this method; 4-bit quantization is a good achievement in the models considered here.\n\nThe quantization process covers nicely the various different parts of the errors that  post-training quantization induces and propose somewhat original solutions to them. \n\nCons:\n\nThe framework is relatively complex and consists of multiple steps.\n\nWhat is the detailed difference of computational resource use between light and advanced pipeline?\n\nAdaQuant seems like a rather straight-forward step from AdaRound by combining with it some related works.\n\nIt is not clear how the BN error compensation differs exactly from related work.\n\nSome details were missing, for example, it is up to the reader to guess how many bits were used in the accumulators.\n\nSome spelling mistakes, e.g., “Optimizing Quantization Pipline”\n\n\nOverall: An engineering oriented paper with some lack of testable hypotheses and analysis of some parts of the methods, but the 4-bit results justify publication. Edit: I have not seen author reply and further reading of the paper has not clarified the main issues found by all reviewers. I have to lower the score. Edit2: new version of the paper and the author reply cleared some concerns, score raised accordingly.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Collection of steps to deal with quantization-induced error in post-training quantization", "review": "This work presents a quite comprehensive multi-step scheme for post-training neural quantization that does not rely on large datasets or large computational resources.  \n\nThe work is has significance in the domain of post-training neural quantization, especially in cases where only a small calibration set or limited resources are available. It would be interesting to also think about accumulator quantization.\n\n\nPros:\n\nThe empirical results are relatively strong in this method; 4-bit quantization is a good achievement in the models considered here.\n\nThe quantization process covers nicely the various different parts of the errors that  post-training quantization induces and propose somewhat original solutions to them. \n\nCons:\n\nThe framework is relatively complex and consists of multiple steps.\n\nWhat is the detailed difference of computational resource use between light and advanced pipeline?\n\nAdaQuant seems like a rather straight-forward step from AdaRound by combining with it some related works.\n\nIt is not clear how the BN error compensation differs exactly from related work.\n\nSome details were missing, for example, it is up to the reader to guess how many bits were used in the accumulators.\n\nSome spelling mistakes, e.g., “Optimizing Quantization Pipline”\n\n\nOverall: An engineering oriented paper with some lack of testable hypotheses and analysis of some parts of the methods, but the 4-bit results justify publication. Edit: I have not seen author reply and further reading of the paper has not clarified the main issues found by all reviewers. I have to lower the score. Edit2: new version of the paper and the author reply cleared some concerns, score raised accordingly.\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603818870614}, {"id": "VH858vlzD7x", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1562/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n\nSummary: The paper studies the problem of Post-Training Quantization of NNs, where no fine-tuning is performed to quantize the model. In particular, the authors focus on sub-8 bit quantization and propose a novel integer linear programming formulation to find the optimal bit width for a given model size. Additional approaches are proposed to minimize accuracy degradation after quantization. These include\n(i) AdaQuant in which the parameters are quantized layer-by-layer to match the full precision output,\n(ii) a batch norm tuning approach to re-adjust the statistics to the quantized model, and\n(iii) an advanced pipeline for cases where backpropogation can be performed.\n\nExperiments are performed on ResNet18/50, MobileNetV2, and BERT-SQuaD1.1 showing the feasibility of the proposed method. \n\nI think the approach in the paper is pretty interesting, and specially the integer linear programming solution.\nOverall the paper is strong however, please note the following:\n\n\n- Page 3 last paragraph: it seems there are errors in the results for calibration data.\n1/ it is stated that if \"B>> M then we might underfit the data.\". Do the authors mean \"B>>N\"? Similarly it seems the result given for the convolution (i.e. B> cK^2/HW) needs to be revised.\n2/ The analysis requires specifying the rank of W and in particular the relationship between M and N. In particular, note that the matrix W can at most have a rank of min(M,N) and as a result the rest of the linear equations for calibrating the data would be redundant. This needs to be taken into account in your result.\n\n- Page 4: \"Depending on the needs, our performance metrics P would be either the execution time of the network or its power consumption.\" This is good but no result on either latency or power consumption is provided in the paper.\n\n- Figure 4 (a,b): Please provide the BOPS for the mixed-precision results. BOPS proposed by (https://arxiv.org/pdf/2005.07093.pdf) is a good metric to measure the total reduction in computations for mixed precision quantization.\n\n- The results section uses weak FP32 baseline. For instance, the baseline accuracy for ResNet50 is 77.2\\% and MobileNetV2 is 73\\%, while the paper uses 76.1\\% and 71.8\\%.\n\n- Related to the above, it is stated \"on the extensively studies 8bit MobileNet-V2 topology we achieved 71.6% top-1 accuracy\". This is a good result but please note that other work in the literature (arxiv:2001.00281) reports 72.91\\% for INT8 quantization of MobileNetV2 (this comparison is actually missing from the paper). It is immediately not clear if the lower reported accuracy is due to the weaker FP32 baseline used or if it is an inherent problem with the method (most probably it is the former but it would be to show this).\n\n\n\n\nMinor Comments:\n\n- There were several grammatical/spelling mistakes in the paper. Please proofread the paper thoroughly. Below are some of the errors that I caught:\n\n- 3 OPTIMIZING QUANTIZATION PIPLINE -> PIPELINE\n\n- Page 6:  the our method robustness -> the robustness of our method\n\n- Page 6: manged -> managed\n\n- Page 7: an significant advantage -> a significant advantage\n\n- Page 8: we managed to switched  -> we managed to switch\n\n- Page 8: For instance, on the extensively studies -> For instance, on the extensively studied", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Approach To Quantization", "review": "\n\nSummary: The paper studies the problem of Post-Training Quantization of NNs, where no fine-tuning is performed to quantize the model. In particular, the authors focus on sub-8 bit quantization and propose a novel integer linear programming formulation to find the optimal bit width for a given model size. Additional approaches are proposed to minimize accuracy degradation after quantization. These include\n(i) AdaQuant in which the parameters are quantized layer-by-layer to match the full precision output,\n(ii) a batch norm tuning approach to re-adjust the statistics to the quantized model, and\n(iii) an advanced pipeline for cases where backpropogation can be performed.\n\nExperiments are performed on ResNet18/50, MobileNetV2, and BERT-SQuaD1.1 showing the feasibility of the proposed method. \n\nI think the approach in the paper is pretty interesting, and specially the integer linear programming solution.\nOverall the paper is strong however, please note the following:\n\n\n- Page 3 last paragraph: it seems there are errors in the results for calibration data.\n1/ it is stated that if \"B>> M then we might underfit the data.\". Do the authors mean \"B>>N\"? Similarly it seems the result given for the convolution (i.e. B> cK^2/HW) needs to be revised.\n2/ The analysis requires specifying the rank of W and in particular the relationship between M and N. In particular, note that the matrix W can at most have a rank of min(M,N) and as a result the rest of the linear equations for calibrating the data would be redundant. This needs to be taken into account in your result.\n\n- Page 4: \"Depending on the needs, our performance metrics P would be either the execution time of the network or its power consumption.\" This is good but no result on either latency or power consumption is provided in the paper.\n\n- Figure 4 (a,b): Please provide the BOPS for the mixed-precision results. BOPS proposed by (https://arxiv.org/pdf/2005.07093.pdf) is a good metric to measure the total reduction in computations for mixed precision quantization.\n\n- The results section uses weak FP32 baseline. For instance, the baseline accuracy for ResNet50 is 77.2\\% and MobileNetV2 is 73\\%, while the paper uses 76.1\\% and 71.8\\%.\n\n- Related to the above, it is stated \"on the extensively studies 8bit MobileNet-V2 topology we achieved 71.6% top-1 accuracy\". This is a good result but please note that other work in the literature (arxiv:2001.00281) reports 72.91\\% for INT8 quantization of MobileNetV2 (this comparison is actually missing from the paper). It is immediately not clear if the lower reported accuracy is due to the weaker FP32 baseline used or if it is an inherent problem with the method (most probably it is the former but it would be to show this).\n\n\n\n\nMinor Comments:\n\n- There were several grammatical/spelling mistakes in the paper. Please proofread the paper thoroughly. Below are some of the errors that I caught:\n\n- 3 OPTIMIZING QUANTIZATION PIPLINE -> PIPELINE\n\n- Page 6:  the our method robustness -> the robustness of our method\n\n- Page 6: manged -> managed\n\n- Page 7: an significant advantage -> a significant advantage\n\n- Page 8: we managed to switched  -> we managed to switch\n\n- Page 8: For instance, on the extensively studies -> For instance, on the extensively studied", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603769574907}], "openreview_url": "https://openreview.net/forum?id=Mf4ZSXMZP7", "arxiv_id": "2006.10518", "paper_pdf": "papers/Mf4ZSXMZP7.pdf", "paper_pdf_sha256": "3a0213392a8fab771d33067efadf20d69cf76c30b8aa2f0c813919d410c96a4f", "paper_pdf_bytes": 549968, "paper_pdf_source": "openreview", "code_url": "https://github.com/itayhubara/CalibTIP", "code_repository": "itayhubara/CalibTIP", "code_commit": "8bffbe450fd48351b23bf873c59fb60ece5143d7", "code_archive": "repos/Mf4ZSXMZP7.zip", "code_archive_sha256": "b2c0006d171fb9ea2571b154c3fba69b413de0c5f4c7cf2cd1400ddc73e4a72e", "code_archive_bytes": 201212, "code_file_count": 66, "code_extensions": {".py": 51, ".sh": 15}, "github_disk_usage_kb": 214, "github_languages": {"Python": 498134, "Shell": 17724}, "github_archived": false, "github_pushed_at": "2021-06-10T13:20:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-post-training-neural-quantization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BkeqATVYwr", "year": 2020, "status": "rejected", "title": "GRAPH NEIGHBORHOOD ATTENTIVE POOLING", "authors": ["Zekarias Tilahun Kefato", "Sarunas Girdzijauskas"], "authorids": ["zekarias@kth.se", "sarunasg@kth.se"], "authors_source": "OpenReview API", "abstract": "Network representation learning (NRL) is a powerful technique for learning low-dimensional vector representation of high-dimensional and sparse graphs. Most studies explore the structure and meta data associated with the graph using random walks and employ a unsupervised or semi-supervised learning schemes. Learning in these methods is context-free, because only a single representation per node is learned. Recently studies have argued on the sufficiency of a single representation and proposed a context-sensitive approach that proved to be highly effective in applications such as link prediction and ranking.\nHowever, most of these methods rely on additional textual features that require RNNs or CNNs to capture high-level features or rely on a community detection algorithm to identifying multiple contexts of a node.\nIn this study, without requiring additional features nor a community detection algorithm, we propose a novel context-sensitive algorithm called GAP that learns to attend on different part of a node’s neighborhood using attentive pooling networks. We show the efficacy of GAP using three real-world datasets on link prediction and node clustering tasks and compare it against 10 popular and state-of-the-art (SOTA) baselines. GAP consistently outperforms them and achieves up to ≈9% and ≈20% gain over the best performing methods on link prediction and clustering tasks, respectively.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1gUZkOTKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper865/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe proposed paper adapts Attentive Pooling Network (ATN) for graphs, by noting that the order of\nneighbors of any node in a graph does not matter unlike the neighboring words in a sentence (for\nwhich APN was developed). A simple modification of removing the layer which encodes higher order\nsequential properties in APN, eg n-gram like statistics, the APN is adapted to GAP. This allows\nencoding context from the neighboring nodes for a specific pair to be compared. Applications on Link\nPrediction and Node Clustering are demonstrated on three benchmark datasets.\n\nDetailed comments:\n- The paper is simple and easy to read and the modification of APN to GAP is also small while being\nwell motivated and empirically important.\n\n- There are some confusions in the writing: p4 first paragraph says that \"In principle one can learn\nusing all pairs of nodes, however that is not scalable, and hence we restrict learning between pairs\nin E.\" However from and around Eq2 it appears that r_s.r_t^- is also computed, where (s,r^-) is not\nan edge.\n\n- Also the statement below Eq2 \"The goal is to learn, in an unsupervised fashion...\" is not correct\n  specially for link prediction task, as having a training graph (with edges) is having annotations.\n        \n- Although, the removal of the `encoding' step from APN makes sense, it would also be possible that\n  the encoding operation is learned to be an identity function automatically if the training data is\n  presented appropriately. I would assume that learning in such a case might take longer, and might\n  not achieve the performance achieved by GAP. An ablation experiment could be reported to support\n  GAP further where encoder is kept in APN, but at the training same pair is presented multiple\n  times with the order of the neighborhoods randomized.      \n       \n- Similar to the experiment above, the order of the nodes could be arbitrarily fixed by some simple\n  logic. Eg order the neighboring nodes sorted by the feature similarity with the current node. This\n  can then be fed to APN and a baseline could be reported.\n\n- I do not directly work with Link Prediction or Node Clustering tasks, so am not familiar with the\n  literature and the baselines reported. However, it might be that many baselines do not use a\n  discriminative objective. Perhaps just using the discriminative objective gives a lot of boost and\n  using the context from neighbors is not very important? From Fig2 and Fig3a it seems that\n  neighborhood size does not make a big difference? This should also be ablated. One experiment\n  could be to use just the current feature and have a small MLP on it, and learn using the proposed\n  discriminative objective. If any other experiment can be designed in similar lines, it should be\n  included as well.\n\n- Why are run time comparison given on simulated data and not real data?\n\n- The experimental setup for training edge selection should be detailed more. In particular, I\n  would recommend that true generalization in terms of nodes and edges should be ensured, by having\n  no test node (i) appear in the train set and (ii) has an edge which connects it to a training\n  node. If this is not used then the alternative scheme used should be explained and argued for.\n\n\nI am not a direct expert in the area and felt that the paper was somewhat lacking. I am putting my initial rating\nto the conservative side. I am very open to revising the rating based on the author responses and the other \nreviewers' comments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary:\nThe proposed paper adapts Attentive Pooling Network (ATN) for graphs, by noting that the order of\nneighbors of any node in a graph does not matter unlike the neighboring words in a sentence (for\nwhich APN was developed). A simple modification of removing the layer which encodes higher order\nsequential properties in APN, eg n-gram like statistics, the APN is adapted to GAP. This allows\nencoding context from the neighboring nodes for a specific pair to be compared. Applications on Link\nPrediction and Node Clustering are demonstrated on three benchmark datasets.\n\nDetailed comments:\n- The paper is simple and easy to read and the modification of APN to GAP is also small while being\nwell motivated and empirically important.\n\n- There are some confusions in the writing: p4 first paragraph says that \"In principle one can learn\nusing all pairs of nodes, however that is not scalable, and hence we restrict learning between pairs\nin E.\" However from and around Eq2 it appears that r_s.r_t^- is also computed, where (s,r^-) is not\nan edge.\n\n- Also the statement below Eq2 \"The goal is to learn, in an unsupervised fashion...\" is not correct\n  specially for link prediction task, as having a training graph (with edges) is having annotations.\n        \n- Although, the removal of the `encoding' step from APN makes sense, it would also be possible that\n  the encoding operation is learned to be an identity function automatically if the training data is\n  presented appropriately. I would assume that learning in such a case might take longer, and might\n  not achieve the performance achieved by GAP. An ablation experiment could be reported to support\n  GAP further where encoder is kept in APN, but at the training same pair is presented multiple\n  times with the order of the neighborhoods randomized.      \n       \n- Similar to the experiment above, the order of the nodes could be arbitrarily fixed by some simple\n  logic. Eg order the neighboring nodes sorted by the feature similarity with the current node. This\n  can then be fed to APN and a baseline could be reported.\n\n- I do not directly work with Link Prediction or Node Clustering tasks, so am not familiar with the\n  literature and the baselines reported. However, it might be that many baselines do not use a\n  discriminative objective. Perhaps just using the discriminative objective gives a lot of boost and\n  using the context from neighbors is not very important? From Fig2 and Fig3a it seems that\n  neighborhood size does not make a big difference? This should also be ablated. One experiment\n  could be to use just the current feature and have a small MLP on it, and learn using the proposed\n  discriminative objective. If any other experiment can be designed in similar lines, it should be\n  included as well.\n\n- Why are run time comparison given on simulated data and not real data?\n\n- The experimental setup for training edge selection should be detailed more. In particular, I\n  would recommend that true generalization in terms of nodes and edges should be ensured, by having\n  no test node (i) appear in the train set and (ii) has an edge which connects it to a training\n  node. If this is not used then the alternative scheme used should be explained and argued for.\n\n\nI am not a direct expert in the area and felt that the paper was somewhat lacking. I am putting my initial rating\nto the conservative side. I am very open to revising the rating based on the author responses and the other \nreviewers' comments."}, "tcdate": 1571811069825}, {"id": "SJlUP9iqFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper865/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Minimal theoretical novelty: The paper is too focussed on the empirical advantage achieved on the datasets used in the experiments. \n\n- Regarding equation (2), what is the guarantee this modification depicted by the hard-margin loss will always lead to improved performance like what happened with the experimented datasets? or more practically what are the required conditions for it to perform well?\n\n- Writing really needs to improve. There are too many typos and grammatical mistakes. \nExamples include:\n -- p1: \"to learn representation of graphs\". \n -- p1: \"on other contexts its coupled with\".\n -- p1: \"NRL studies have shown a context-sensitive approach significantly outperform previous context-free SOTA methods in link-prediction task.\"\n -- p3: \"a more sophisticated neighborhood functions\".\n -- p5: \"reporeted\"\n -- p7: \"two other variant\"\n -- p8: \"and gain\"\n\n- The latter issue makes is sometimes tricky to follow the ideas presented. \nExample:\n -- last two lines in Section 3. \n\n- Last paragraph in Section 1 is pretty informative about the pros and cons of the method. It also rather admits the first issue mentioned here in the review.  \n\nMinor:\n- p4: \"Eg. \" --> eg. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "- Minimal theoretical novelty: The paper is too focussed on the empirical advantage achieved on the datasets used in the experiments. \n\n- Regarding equation (2), what is the guarantee this modification depicted by the hard-margin loss will always lead to improved performance like what happened with the experimented datasets? or more practically what are the required conditions for it to perform well?\n\n- Writing really needs to improve. There are too many typos and grammatical mistakes. \nExamples include:\n -- p1: \"to learn representation of graphs\". \n -- p1: \"on other contexts its coupled with\".\n -- p1: \"NRL studies have shown a context-sensitive approach significantly outperform previous context-free SOTA methods in link-prediction task.\"\n -- p3: \"a more sophisticated neighborhood functions\".\n -- p5: \"reporeted\"\n -- p7: \"two other variant\"\n -- p8: \"and gain\"\n\n- The latter issue makes is sometimes tricky to follow the ideas presented. \nExample:\n -- last two lines in Section 3. \n\n- Last paragraph in Section 1 is pretty informative about the pros and cons of the method. It also rather admits the first issue mentioned here in the review.  \n\nMinor:\n- p4: \"Eg. \" --> eg. "}, "tcdate": 1571629661705}, {"id": "rylsqZtauS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper865/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a GAP (graph neighborhood attentive pooling) method to solve problems of node clustering and link prediction. The proposed idea is mostly inspired by the attentive pooling network approach (APN) that has been widely used in the NLP domains; indeed the detailed steps of the proposed method is almost a ``mirror” application of APN from question-answer pair ranking in NLP to the graph domains, and the only difference is that rather than using an RNN or LSTM to deal with the temporal orders in sentences (as in APN), the proposed method simply collect the first-order neighbors of the source node and target node and concatenate them without considering their orders. For the rest part, the authors used a hard margin loss function which is also the same as in APN.\n\nThe proposed method is very similar to the APN method and so its novelty is limited. Another concern is that although graph nodes do not have temporal orders as in word sentences, their relative connections manifested through the edges are important topological information that should be captured in contextualize a node. Unfortunately the authors almost totally ignored this information. The ignorance of the node orders appear to me not an advantage but instead a limitation, though it makes the computation and implementation much easier.\n\nThe graph attention network (GAT) is a very related method but I see no comparison with it in node clustering tasks. Also many recent methods on graph convolutional networks are not incorporated for comparison. \n\nBased on the concerns of novelty, lack of considering node topologies, and lack of comparison with strongly related methods, it is hard to recommend acceptance of this paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper proposed a GAP (graph neighborhood attentive pooling) method to solve problems of node clustering and link prediction. The proposed idea is mostly inspired by the attentive pooling network approach (APN) that has been widely used in the NLP domains; indeed the detailed steps of the proposed method is almost a ``mirror” application of APN from question-answer pair ranking in NLP to the graph domains, and the only difference is that rather than using an RNN or LSTM to deal with the temporal orders in sentences (as in APN), the proposed method simply collect the first-order neighbors of the source node and target node and concatenate them without considering their orders. For the rest part, the authors used a hard margin loss function which is also the same as in APN.\n\nThe proposed method is very similar to the APN method and so its novelty is limited. Another concern is that although graph nodes do not have temporal orders as in word sentences, their relative connections manifested through the edges are important topological information that should be captured in contextualize a node. Unfortunately the authors almost totally ignored this information. The ignorance of the node orders appear to me not an advantage but instead a limitation, though it makes the computation and implementation much easier.\n\nThe graph attention network (GAT) is a very related method but I see no comparison with it in node clustering tasks. Also many recent methods on graph convolutional networks are not incorporated for comparison. \n\nBased on the concerns of novelty, lack of considering node topologies, and lack of comparison with strongly related methods, it is hard to recommend acceptance of this paper.\n"}, "tcdate": 1570767250892}], "openreview_url": "https://openreview.net/forum?id=BkeqATVYwr", "arxiv_id": "2001.10394", "paper_pdf": "papers/BkeqATVYwr.pdf", "paper_pdf_sha256": "901b8289344e9ddc367fbcc556f54105f43fcee00b2f2c53e1e37fa172c83087", "paper_pdf_bytes": 364677, "paper_pdf_source": "openreview", "code_url": "https://github.com/zekarias-tilahun/GAP", "code_repository": "zekarias-tilahun/GAP", "code_commit": "0bb528135690de28517686453835babe17ff8f16", "code_archive": "repos/BkeqATVYwr.zip", "code_archive_sha256": "fdf6b492e0ea2ec5a66985d83ad2bd0c51c9f3de6fbbd05c35f0ef3a61aa9cc8", "code_archive_bytes": 247506, "code_file_count": 6, "code_extensions": {".py": 5, ".sh": 1}, "github_disk_usage_kb": 263, "github_languages": {"Python": 27027, "Shell": 236}, "github_archived": false, "github_pushed_at": "2021-03-10T09:58:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-neighborhood-attentive-pooling-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rk4Wf30qKQ", "year": 2019, "status": "rejected", "title": "Security Analysis of Deep Neural Networks Operating in the Presence of Cache Side-Channel Attacks", "authors": ["Sanghyun Hong", "Michael Davinroy", "Yigitcan Kaya", "Stuart Nevans Locke", "Ian Rackow", "Kevin Kulda", "Dana Dachman-Soled", "Tudor Dumitraș"], "authorids": ["shhong@cs.umd.edu", "mdavinr1@swarthmore.edu", "yigitcan@cs.umd.edu", "stnevans@mail.rit.edu", "ian.rackow@gmail.com", "kevin_kulda1@baylor.edu", "danadach@ece.umd.edu", "tdumitra@umiacs.umd.edu"], "authors_source": "OpenReview API", "abstract": "Recent work has introduced attacks that extract the architecture information of deep neural networks (DNN), as this knowledge enhances an adversary’s capability to conduct attacks on black-box networks. This paper presents the first in-depth security analysis of DNN fingerprinting attacks that exploit cache side-channels.  First, we define the threat model for these attacks:  our adversary does not need the ability to query the victim model; instead, she runs a co-located process on the host machine victim ’s deep learning  (DL) system is running and passively monitors the accesses of the target functions in the shared framework.  Second, we introduce DeepRecon, an attack that reconstructs the architecture of the victim network by using the internal information extracted via Flush+Reload, a cache side-channel technique. Once the attacker observes function invocations that map directly to architecture attributes of the victim network, the attacker can reconstruct the victim’s entire network architecture.  In our evaluation, we demonstrate that an attacker can accurately reconstruct two complex networks (VGG19 and ResNet50) having only observed one forward propagation. Based on the extracted architecture attributes, we also demonstrate that an attacker can build a meta-model that accurately fingerprints the architecture and family of the pre-trained model in a transfer learning setting. From this meta-model,  we evaluate the importance of the observed attributes in the fingerprinting process. Third, we propose and evaluate new framework-level defense techniques that obfuscate our attacker’s observations. Our empirical security analysis represents a step toward understanding the DNNs’ vulnerability to cache side-channel attacks.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rylt1I8ZpQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1241/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of fingerprinting neural network architectures using cache side channels. In the considered threat model, the attacker runs a process co-located with the victim's, and uses standard FLUSH+RELOAD attacks to infer high-level architectural information such as the number and types of layers of the victim's ML model. The paper concludes with the discussion of some \"security-through-obscurity\" defenses.\n\nI don't quite understand the threat model considered in this paper. The main motivating factor given by the authors for uncovering model architecture details is for facilitating black-box attacks against ML models (e.g., for adversarial examples or membership inference). \nYet, in the case of adversarial examples for instance, knowledge of the architecture is often considered a given as keeping it secret has very little influence on attacks. There are black-box attacks that require no knowledge of the architecture and only a few queries (e.g., Black-box Adversarial Attacks with Limited Queries and Information, Ilyas et al., ICML'18). \nSo overall, learning such coarse-grained features about a model just doesn't seem particularly useful, especially since architecture-level details are often not considered private or secret to begin with.\n\nAfter architectural details have been extracted, the end-goal attacks on ML models considered by the authors (e.g., model stealing, adversarial examples, etc.) require query access anyways. Thus, additionally assuming co-location between the adversary and the victim's model seems to unnecessarily strengthen the attacker model.\n\nMaybe the most interesting scenario to consider for cache side-channels in ML is when ML models are run on trusted hardware (e.g., Oblivious Multi-Party Machine Learning on Trusted Processors, Ohrimenko et al.; or this work also submitted to ICLR: https://openreview.net/forum?id=rJVorjCcKQ).\nCache side channels are much more relevant to that threat model (i.e., ML code running in a trusted hardware enclave hosted by a malicious party). And indeed, there have been many cache side-channel attack papers against trusted hardware such as Intel's SGX (e.g., Software Grand Exposure: SGX Cache Attacks Are Practical, Brasser et al.)\n\nBut given what we know about the strength of these cache side channel attacks, one would expect to be able to extract much more interesting information about a target model, such as its weights, inputs or outputs. In the above trusted hardware scenario, solely extracting architecture-level information would also not be considered a very strong attack, especially since coarse-grained information (e.g., a rough bound on the number of layers), can be trivially obtained via timing side channels.\n\nMinor comments:\n- In the introduction, you say that white-box attacks for adversarial examples are rendered ineffective by gradient masking. This isn't true in general. Only \"weak\" white-box attacks can be rendered ineffective this way. So far, there are no examples of models that resist white-box attacks yet are vulnerable to black-box attacks.\n- What exactly causes the cache-level differences you observe? Can you give some  code examples in the paper that showcase what happens? Are the TensorFlow code lines listed in Table 1 from a specific commit or release?\n- The defenses discussed in Section 5 are all forms of \"security through obscurity\" that seem easily defeated by a determined attacker that adapts its attack (and maybe uses a few additional observations).\n\n--REVISION--\nI thank the authors for their rebuttal and clarifications on the threat model and end goals of their attacks. I remain somewhat unconvinced by the usefulness of extracting architectural information. For most of the listed attacks (e.g., building substitute models for adversarial examples, or simply for model extraction) it is not clear from prior work that knowledge of the architecture is really necessary, although it is of course always helpful to have this knowledge. As I mentioned in my review, with current (undefended) ML libraries, it should be possible to extract much more information (e.g., layer weights) using cache side channels.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear threat model with a very strong adversary that obtains information of moderate significance.", "review": "This paper considers the problem of fingerprinting neural network architectures using cache side channels. In the considered threat model, the attacker runs a process co-located with the victim's, and uses standard FLUSH+RELOAD attacks to infer high-level architectural information such as the number and types of layers of the victim's ML model. The paper concludes with the discussion of some \"security-through-obscurity\" defenses.\n\nI don't quite understand the threat model considered in this paper. The main motivating factor given by the authors for uncovering model architecture details is for facilitating black-box attacks against ML models (e.g., for adversarial examples or membership inference). \nYet, in the case of adversarial examples for instance, knowledge of the architecture is often considered a given as keeping it secret has very little influence on attacks. There are black-box attacks that require no knowledge of the architecture and only a few queries (e.g., Black-box Adversarial Attacks with Limited Queries and Information, Ilyas et al., ICML'18). \nSo overall, learning such coarse-grained features about a model just doesn't seem particularly useful, especially since architecture-level details are often not considered private or secret to begin with.\n\nAfter architectural details have been extracted, the end-goal attacks on ML models considered by the authors (e.g., model stealing, adversarial examples, etc.) require query access anyways. Thus, additionally assuming co-location between the adversary and the victim's model seems to unnecessarily strengthen the attacker model.\n\nMaybe the most interesting scenario to consider for cache side-channels in ML is when ML models are run on trusted hardware (e.g., Oblivious Multi-Party Machine Learning on Trusted Processors, Ohrimenko et al.; or this work also submitted to ICLR: https://openreview.net/forum?id=rJVorjCcKQ).\nCache side channels are much more relevant to that threat model (i.e., ML code running in a trusted hardware enclave hosted by a malicious party). And indeed, there have been many cache side-channel attack papers against trusted hardware such as Intel's SGX (e.g., Software Grand Exposure: SGX Cache Attacks Are Practical, Brasser et al.)\n\nBut given what we know about the strength of these cache side channel attacks, one would expect to be able to extract much more interesting information about a target model, such as its weights, inputs or outputs. In the above trusted hardware scenario, solely extracting architecture-level information would also not be considered a very strong attack, especially since coarse-grained information (e.g., a rough bound on the number of layers), can be trivially obtained via timing side channels.\n\nMinor comments:\n- In the introduction, you say that white-box attacks for adversarial examples are rendered ineffective by gradient masking. This isn't true in general. Only \"weak\" white-box attacks can be rendered ineffective this way. So far, there are no examples of models that resist white-box attacks yet are vulnerable to black-box attacks.\n- What exactly causes the cache-level differences you observe? Can you give some  code examples in the paper that showcase what happens? Are the TensorFlow code lines listed in Table 1 from a specific commit or release?\n- The defenses discussed in Section 5 are all forms of \"security through obscurity\" that seem easily defeated by a determined attacker that adapts its attack (and maybe uses a few additional observations).\n\n--REVISION--\nI thank the authors for their rebuttal and clarifications on the threat model and end goals of their attacks. I remain somewhat unconvinced by the usefulness of extracting architectural information. For most of the listed attacks (e.g., building substitute models for adversarial examples, or simply for model extraction) it is not clear from prior work that knowledge of the architecture is really necessary, although it is of course always helpful to have this knowledge. As I mentioned in my review, with current (undefended) ML libraries, it should be possible to extract much more information (e.g., layer weights) using cache side channels.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541658080780}, {"id": "SygLtq7ZpQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1241/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper performs cache side-channel attacks to extract attributes of a victim model, and infer its architecture accordingly. In their threat model, the attacker could launch a co-located process on the same host machine, and use the same DL framework as the victim model. Their evaluation shows that: (1) their attacks can extract the model attributes pretty well, including the number of different types of layers; (2) using these attributes, they train a decision tree classifier among 13 CNN architectures, and show that they can achieve a nearly perfect classification accuracy. They also evaluate some defense strategies against their attacks.\n\nModel extraction attack under a black-box setting is an important topic, and I am convinced that their threat model is a good step towards real-world attacks. As for the novelty, although Yan et al. also evaluate cache side-channel attacks, that paper was released pretty shortly before ICLR deadline, thus I would consider this work as an independent contribution at its submission.\n\nI have several questions and comments about this paper:\n\n- One difference of the evaluation setup between this paper and Yan et al. is that in Yan et al., they are trying to infer more detailed hyper-parameters of the architecture (e.g., the number of neurons, the dimensions of each layer, the connections), but within a family of architectures (i.e., VGG or ResNet). On the other hand, in this paper, the authors extract higher-level attributes such as the number of different layers and activation functions, and predict the model family (from 5 options) or the concrete model architecture (from 13 options). While I think inferring the model family type is also an interesting problem, this setup is still a little contrived. Would the classifier predict the family of a model correctly if it is not included in the training set, say, could it predict ResNet32 as R (ResNet)?\n\n- In Table 3, it looks like the errors in the captured computation sequences show some patterns. Are these error types consistent across different runs? Could you provide some explanation of these errors?\n\n- In Table 5, my understanding is that we need to compare the avg errors to the numbers in Table 2. In this case, the errors seem to be even larger than the sum of the attribute values. Is this observation correct? If so, could you discuss what attributes are most wrongly captured, and show some examples?\n\n- It would be beneficial to provide a more detailed comparison between this work and Yan et al., e.g., whether the technique proposed in this work could be also extended to infer more fine-grained attributes of a model, and go beyond a classification among a pre-defined set of architectures.\n\n- The paper needs some editing to fix some typos. For example, in Table 5, the captions of Time (Baseline) and Time (+TinyNet) should be changed, and it looks confusing at the first glance.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple yet effective attacks to infer model architectures; more clarification would help", "review": "This paper performs cache side-channel attacks to extract attributes of a victim model, and infer its architecture accordingly. In their threat model, the attacker could launch a co-located process on the same host machine, and use the same DL framework as the victim model. Their evaluation shows that: (1) their attacks can extract the model attributes pretty well, including the number of different types of layers; (2) using these attributes, they train a decision tree classifier among 13 CNN architectures, and show that they can achieve a nearly perfect classification accuracy. They also evaluate some defense strategies against their attacks.\n\nModel extraction attack under a black-box setting is an important topic, and I am convinced that their threat model is a good step towards real-world attacks. As for the novelty, although Yan et al. also evaluate cache side-channel attacks, that paper was released pretty shortly before ICLR deadline, thus I would consider this work as an independent contribution at its submission.\n\nI have several questions and comments about this paper:\n\n- One difference of the evaluation setup between this paper and Yan et al. is that in Yan et al., they are trying to infer more detailed hyper-parameters of the architecture (e.g., the number of neurons, the dimensions of each layer, the connections), but within a family of architectures (i.e., VGG or ResNet). On the other hand, in this paper, the authors extract higher-level attributes such as the number of different layers and activation functions, and predict the model family (from 5 options) or the concrete model architecture (from 13 options). While I think inferring the model family type is also an interesting problem, this setup is still a little contrived. Would the classifier predict the family of a model correctly if it is not included in the training set, say, could it predict ResNet32 as R (ResNet)?\n\n- In Table 3, it looks like the errors in the captured computation sequences show some patterns. Are these error types consistent across different runs? Could you provide some explanation of these errors?\n\n- In Table 5, my understanding is that we need to compare the avg errors to the numbers in Table 2. In this case, the errors seem to be even larger than the sum of the attribute values. Is this observation correct? If so, could you discuss what attributes are most wrongly captured, and show some examples?\n\n- It would be beneficial to provide a more detailed comparison between this work and Yan et al., e.g., whether the technique proposed in this work could be also extended to infer more fine-grained attributes of a model, and go beyond a classification among a pre-defined set of architectures.\n\n- The paper needs some editing to fix some typos. For example, in Table 5, the captions of Time (Baseline) and Time (+TinyNet) should be changed, and it looks confusing at the first glance.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541646974457}, {"id": "rygfvh7h2X", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper1241/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes a cache side-channel attack on a deep learning model. In a cache side-channel attack, the attacker sets up a process on the same machine where the victim process (that is running the training or evaluation job for the DNN model) is running. It is assumed that the victim process uses a common shared library for DNN computations as the attacking process. The attacking process flushes the cache, then observes access times for key functions. The paper shows that, based on the speed of accessing previously flushed functions, the attacker can discover the high-level network architecture, namely the types of layers and their sequence. The paper shows that, by spying on such cache access patterns in the Tensorflow library, this method can reliably extract the above high-level information for 11 different network architectures. It also describes a few counterattack alternatives whereby the victim can obfuscate its cache access patterns for self-protection.\n\nThe significance of the results is not clear to me. The extracted information is very high level. What realistic attacks can be constructed from such a coarse-grained fingerprinting? The experimental results show that the fingerprint can be used to map the architecture to one of the 13 well-known architectures (VCC16, ResNet, DenseNet, Inception, etc.). But so what? What does the victim lose by revealing that it's using one of a few very well known types of DNNs (the ones tested in this paper). There may very well be a good reason why this is very dangerous, but that is not explained in the paper. Not being familiar with this line of research and its significance, I looked up several of the related papers (Suciu et al., 2018, Tramer et al., 2017, Papernot et al., 2017, Yan et al., 2018). None of them could explain why this particular type of fingerprinting is dangerous.\n\nOf the cited previous work, Yan et al., 2018 seems to present the most closely related approach. The method described in that paper is very similar: cache side attack on a shared library through a co-located attacker process. They monitor at a finer grain -- Generalized Matrix Multiplications -- and are thus able to infer more details such as the size of the layers. This also makes the inference problem harder -- they were able to narrow down the search space of networks from >4x10^35 to 16 (on VGG16). On the surface, the results presented in this paper seem stronger. But they are actually solving a much easier problem -- their search space is one of 13 well-known networks. To me, Yan et al.'s approach is a much more powerful and promising setup.\n\nOverall, while the paper is clearly written and presents the idea succinctly, it is derivative of previous research, and the results are not stronger. I'm not an expert in this area, so it's possible that I missed something. Based on my current understanding, however, I recommend reject.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear whether this paper surpasses prior research", "review": "The paper describes a cache side-channel attack on a deep learning model. In a cache side-channel attack, the attacker sets up a process on the same machine where the victim process (that is running the training or evaluation job for the DNN model) is running. It is assumed that the victim process uses a common shared library for DNN computations as the attacking process. The attacking process flushes the cache, then observes access times for key functions. The paper shows that, based on the speed of accessing previously flushed functions, the attacker can discover the high-level network architecture, namely the types of layers and their sequence. The paper shows that, by spying on such cache access patterns in the Tensorflow library, this method can reliably extract the above high-level information for 11 different network architectures. It also describes a few counterattack alternatives whereby the victim can obfuscate its cache access patterns for self-protection.\n\nThe significance of the results is not clear to me. The extracted information is very high level. What realistic attacks can be constructed from such a coarse-grained fingerprinting? The experimental results show that the fingerprint can be used to map the architecture to one of the 13 well-known architectures (VCC16, ResNet, DenseNet, Inception, etc.). But so what? What does the victim lose by revealing that it's using one of a few very well known types of DNNs (the ones tested in this paper). There may very well be a good reason why this is very dangerous, but that is not explained in the paper. Not being familiar with this line of research and its significance, I looked up several of the related papers (Suciu et al., 2018, Tramer et al., 2017, Papernot et al., 2017, Yan et al., 2018). None of them could explain why this particular type of fingerprinting is dangerous.\n\nOf the cited previous work, Yan et al., 2018 seems to present the most closely related approach. The method described in that paper is very similar: cache side attack on a shared library through a co-located attacker process. They monitor at a finer grain -- Generalized Matrix Multiplications -- and are thus able to infer more details such as the size of the layers. This also makes the inference problem harder -- they were able to narrow down the search space of networks from >4x10^35 to 16 (on VGG16). On the surface, the results presented in this paper seem stronger. But they are actually solving a much easier problem -- their search space is one of 13 well-known networks. To me, Yan et al.'s approach is a much more powerful and promising setup.\n\nOverall, while the paper is clearly written and presents the idea succinctly, it is derivative of previous research, and the results are not stronger. I'm not an expert in this area, so it's possible that I missed something. Based on my current understanding, however, I recommend reject.", "rating": "4: Ok but not good enough - rejection", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1541319770176}], "openreview_url": "https://openreview.net/forum?id=rk4Wf30qKQ", "arxiv_id": "1810.03487", "paper_pdf": "papers/rk4Wf30qKQ.pdf", "paper_pdf_sha256": "54dc3bb449bedc82a735dbc978719661b7f5c85c824f6e1e762d9dd4ae406218", "paper_pdf_bytes": 671133, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sanghyun-Hong/DeepRecon", "code_repository": "Sanghyun-Hong/DeepRecon", "code_commit": "ad01b7034345fd6bdf463a4e09c73bf0049c06d0", "code_archive": "repos/rk4Wf30qKQ.zip", "code_archive_sha256": "eed534f55dbb407d22737b86bd2ca7acd06bfe0f88b6bd71c49d347c9ff56b57", "code_archive_bytes": 3543907, "code_file_count": 19, "code_extensions": {".py": 17, ".c": 1, ".sh": 1}, "github_disk_usage_kb": 7501, "github_languages": {"Python": 63674, "C": 7274, "Shell": 1264, "Makefile": 589}, "github_archived": false, "github_pushed_at": "2020-02-19T09:01:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/security-analysis-of-deep-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bGVkAhR7Eu", "year": 2026, "status": "rejected", "title": "SAS-Bench: A Fine-Grained Benchmark for Evaluating Short Answer Scoring with Large Language Models", "authors": ["Peichao Lai", "Kexuan Zhang", "Yi Lin", "Linyihan Zhang", "Feiyang Ye", "Jinhao Yan", "Yanwei Xu", "Conghui He", "Yilei Wang", "Wentao Zhang", "Bin CUI"], "authorids": ["~Peichao_Lai1", "~Kexuan_Zhang1", "~Yi_Lin13", "~Linyihan_Zhang1", "~Feiyang_Ye3", "~Jinhao_Yan1", "~Yanwei_Xu1", "~Conghui_He2", "~Yilei_Wang3", "~Wentao_Zhang1", "~Bin_CUI2"], "authors_source": "OpenReview API", "abstract": "Short Answer Scoring (SAS) is a critical task in automated subjective answer grading, playing an essential role in education, standardized testing, and large-scale assessment systems. However, existing approaches often produce coarse-grained scores and lack detailed reasoning. Although large language models (LLMs) have demonstrated potential as zero-shot evaluators, they remain susceptible to bias, inconsistencies with human judgment, and limited transparency in scoring decisions. To overcome these limitations, we introduce SAS-Bench, a benchmark specifically designed for LLM-based SAS tasks. SAS-Bench provides fine-grained, step-wise scoring, expert-annotated error categories, and a diverse range of question types derived from real-world subject-specific exams. This benchmark facilitates detailed evaluation of model reasoning processes and explainability. We also release an open-source dataset containing 1,030 questions and 4,109 student responses, each annotated by domain experts. Furthermore, we conduct comprehensive experiments with various LLMs, identifying major challenges in scoring science-related questions and highlighting the effectiveness of few-shot prompting in improving scoring accuracy. Our work offers valuable insights into the development of more robust, fair, and educationally meaningful LLM-based evaluation systems.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "1ocjqA8HAT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2147/Reviewer_fNhs"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces SAS-bench, a benchmark to evaluate LLM-based short answer scoring (SAS) by providing fine-grained, step-wise scoring, expert-annotated error categories, and a diverse range of question types derived from real-world subject-specific exams.", "review_text": "This paper introduces SAS-bench, a benchmark to evaluate LLM-based short answer scoring (SAS) by providing fine-grained, step-wise scoring, expert-annotated error categories, and a diverse range of question types derived from real-world subject-specific exams.", "strengths": "* SAS-bench aims to provide fine-grained analysis and explanations behind LLM-based SAS scoring, as well as actionable feedback. These problems are highly relevant in the EduNLP community.\n* The approach to splitting answers into reasoning steps and evaluating each step for correctness and error analysis is intuitive to obtain fine-grained analysis.\n* The authors release the SAS-bench publicly, containing 1030 questions from a real-world exam (China’s National College Entrance Examination(Gaokao)) with 4109 student responses.\n* Auxiliary results showing in-context examples and rubrics help improve scoring performance.\n* Comprehensive evaluation of 16 LLMs across different families and sizes.", "weaknesses": "* Since the primary contribution is a dataset, the synthetic nature of the student responses needs to be justified as well-aligned to responses from real-world test student takers. Reference responses are first generated by only 3 students, thereby lacking diversity. LLMs are then prompted to diversify and introduce errors to generate the final set of positive and negative responses. How well do LLMs perform at this synthetic task? Each LLM-based synthetic step should be evaluated for performance and error analysis. Prior work has shown prompting to be a poor simulator of students, with fine-tuning of real-world student responses required (SMART: Simulated Students Aligned with Item Response Theory for Question Difficulty Prediction, EMNLP 2025).\n* Generalizability to new question banks and domains: Since the error taxonomy, a critical component to get fine-grained analysis,  involves a manual intervention for consolidation and refinement, how generalizable is this scoring approach for fine-grained analysis to Q from other exams/benchmarks? Was a completely automated LLM-based approach (with few-shot human-written examples) tried out?\n* What is the human performance upper bound? No inter-annotator agreement is reported on the error and scoring metrics? Without human performance and agreement, we don’t know whether this task is well defined and how much behind/ahead LLM performance is compared to humans.\n* Motivation: Breaking answers into steps seems intuitive to get fine-grained analysis. However, scoring each step and then adding all scores to arrive at the overall score needs to be justified. Usually, scoring rubrics include key intermediate results needed for different scores, or which missing steps/results would lead in a deduction of points. For example, for reasoning-based domains like math and physics, there could be multiple reasoning paths with varied steps, as well as many steps (max steps for math reported is 25). Students could also include simpler steps or combine or omit them. Won’t this make an alignment between the sum of step-scores and the overall score hard? \n    * Instead of a step-score, simply using binary step correctness seems sufficient to provide fine-grained analysis as well as a useful indicator to arrive at the final score. This could also result in simplified evaluation metrics, with line 313 stating an instability in step-wise consistency evaluation. For example, why does the correct step 1 in Figure 2 get a score of 2, and the other correct steps get a score of 1? Won’t these per-step scoring rubrics be hard to design to avoid variance and ensure high agreement? For math and physics, line 474 states that step-wise labels may mislead the model.\n    * Does adding step-wise evaluation help overall scoring performance (QWK)? The key aim is to improve overall scoring, since even if the fine-grained analysis is useful, if the overall scoring is poor, these models are not practically useful. Line 391 states that adding step-wise consistency introduces additional challenges to the model. Compared with SOTA overall-only scoring baselines, which omit step-wise analysis, how far ahead/behind are SAS-bench-trained models?\n* What is the motivation behind the ECS metric design; does it extend a well-accepted existing metric? Does it correlate with human evaluation? When working with error distributions and frequencies, doesn’t ECS omit the ordering of steps with their errors? Including ordering seems key since an ordered list of steps with errors is being evaluated. Further, results on ECS (table 3) vary notably across different subjects and question types (line 454), showing that a different model is usually the best for different question types, and also has high variance. How would a practitioner choose the best model?", "questions": "* How is a step defined, especially in non-math/non-STEM contexts? Is there high agreement between human annotators for step decomposition?\n* An education-specific evaluation needs to be performed. The major downstream application is the potential to provide useful feedback (line 205). Is this achieved by a human (teacher/student) evaluation?\n* Line 385: observe positive correlation: what’s the exact number and correlation metric used?", "flag_for_ethics_review": ["Yes, Responsible research practice (e.g., human subjects, annotator compensation, data release)"], "all_content": {"summary": "This paper introduces SAS-bench, a benchmark to evaluate LLM-based short answer scoring (SAS) by providing fine-grained, step-wise scoring, expert-annotated error categories, and a diverse range of question types derived from real-world subject-specific exams.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* SAS-bench aims to provide fine-grained analysis and explanations behind LLM-based SAS scoring, as well as actionable feedback. These problems are highly relevant in the EduNLP community.\n* The approach to splitting answers into reasoning steps and evaluating each step for correctness and error analysis is intuitive to obtain fine-grained analysis.\n* The authors release the SAS-bench publicly, containing 1030 questions from a real-world exam (China’s National College Entrance Examination(Gaokao)) with 4109 student responses.\n* Auxiliary results showing in-context examples and rubrics help improve scoring performance.\n* Comprehensive evaluation of 16 LLMs across different families and sizes.", "weaknesses": "* Since the primary contribution is a dataset, the synthetic nature of the student responses needs to be justified as well-aligned to responses from real-world test student takers. Reference responses are first generated by only 3 students, thereby lacking diversity. LLMs are then prompted to diversify and introduce errors to generate the final set of positive and negative responses. How well do LLMs perform at this synthetic task? Each LLM-based synthetic step should be evaluated for performance and error analysis. Prior work has shown prompting to be a poor simulator of students, with fine-tuning of real-world student responses required (SMART: Simulated Students Aligned with Item Response Theory for Question Difficulty Prediction, EMNLP 2025).\n* Generalizability to new question banks and domains: Since the error taxonomy, a critical component to get fine-grained analysis,  involves a manual intervention for consolidation and refinement, how generalizable is this scoring approach for fine-grained analysis to Q from other exams/benchmarks? Was a completely automated LLM-based approach (with few-shot human-written examples) tried out?\n* What is the human performance upper bound? No inter-annotator agreement is reported on the error and scoring metrics? Without human performance and agreement, we don’t know whether this task is well defined and how much behind/ahead LLM performance is compared to humans.\n* Motivation: Breaking answers into steps seems intuitive to get fine-grained analysis. However, scoring each step and then adding all scores to arrive at the overall score needs to be justified. Usually, scoring rubrics include key intermediate results needed for different scores, or which missing steps/results would lead in a deduction of points. For example, for reasoning-based domains like math and physics, there could be multiple reasoning paths with varied steps, as well as many steps (max steps for math reported is 25). Students could also include simpler steps or combine or omit them. Won’t this make an alignment between the sum of step-scores and the overall score hard? \n    * Instead of a step-score, simply using binary step correctness seems sufficient to provide fine-grained analysis as well as a useful indicator to arrive at the final score. This could also result in simplified evaluation metrics, with line 313 stating an instability in step-wise consistency evaluation. For example, why does the correct step 1 in Figure 2 get a score of 2, and the other correct steps get a score of 1? Won’t these per-step scoring rubrics be hard to design to avoid variance and ensure high agreement? For math and physics, line 474 states that step-wise labels may mislead the model.\n    * Does adding step-wise evaluation help overall scoring performance (QWK)? The key aim is to improve overall scoring, since even if the fine-grained analysis is useful, if the overall scoring is poor, these models are not practically useful. Line 391 states that adding step-wise consistency introduces additional challenges to the model. Compared with SOTA overall-only scoring baselines, which omit step-wise analysis, how far ahead/behind are SAS-bench-trained models?\n* What is the motivation behind the ECS metric design; does it extend a well-accepted existing metric? Does it correlate with human evaluation? When working with error distributions and frequencies, doesn’t ECS omit the ordering of steps with their errors? Including ordering seems key since an ordered list of steps with errors is being evaluated. Further, results on ECS (table 3) vary notably across different subjects and question types (line 454), showing that a different model is usually the best for different question types, and also has high variance. How would a practitioner choose the best model?", "questions": "* How is a step defined, especially in non-math/non-STEM contexts? Is there high agreement between human annotators for step decomposition?\n* An education-specific evaluation needs to be performed. The major downstream application is the potential to provide useful feedback (line 205). Is this achieved by a human (teacher/student) evaluation?\n* Line 385: observe positive correlation: what’s the exact number and correlation metric used?", "flag_for_ethics_review": ["Yes, Responsible research practice (e.g., human subjects, annotator compensation, data release)"], "details_of_ethics_concerns": "IRB approval for reference responses from student annotators. Approval to include exam questions in the benchmark.", "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762039323702}, {"id": "PZEcLPEL3j", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2147/Reviewer_H3Ty"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper presents a new benchmark for short answer scoring/grading. The main\ndifference compared to existing benchmarks are that SAS-Bench contains\nstep-wise scoring for each answer and that errors have been categorized by\ndomain experts. Overall the dataset contains around 1K questions and 4K\nanswers. The dataset is evaluated using 16 LLMs with a focus on overall score,\nstep-by-step scores and error types.", "review_text": "The paper presents a new benchmark for short answer scoring/grading. The main\ndifference compared to existing benchmarks are that SAS-Bench contains\nstep-wise scoring for each answer and that errors have been categorized by\ndomain experts. Overall the dataset contains around 1K questions and 4K\nanswers. The dataset is evaluated using 16 LLMs with a focus on overall score,\nstep-by-step scores and error types.", "strengths": "S1: The detailed, step-by-step score annotation as well as the error type annotation is quite useful and a good addition to existing datasets.\n\nS2: The introduced scores for overall, step-by-step and error are sensible and the evaluation in terms of number of LLMs quite extensive.", "weaknesses": "W1: The student answers seem to be mostly generate by LLMs. Some answers were generated by only six students. The distribution is not clear. How many are from students? How many of the eight generated answers per question are disregarded?\n\nW2: The student answers are not real answers, collected by students that actually have taken the test. This means that the dataset most likely does not contain many of the patterns found in real student responses such as empty and half-completed answers.\n\nW3: Dataset is only in Chinese. It would be more useful to have an English version as well, also to compare inter-lingual differences in terms of performance.", "questions": "Q1: Why are there only roughly 4 student responses per question? This seems very low.\n\nQ2: What is the difference between step-by-step scores and errors? These seem highly related. A correct step will have no error and full score, whereas an error type will have a reduced score.\n\nQ3: Can you add other languages and make the dataset multilingual?\n\nQ4: From the paper it is not clear how many answers have been actually created by students and how many by LLMs. Also, having overall six student annotators is very little and the setting seems not realistic (i.e., these students are given the task to annotate and have not taken the exam)\n\nQ5: Did you perform any evaluation on adversarial attacks on the LLM graders?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a new benchmark for short answer scoring/grading. The main\ndifference compared to existing benchmarks are that SAS-Bench contains\nstep-wise scoring for each answer and that errors have been categorized by\ndomain experts. Overall the dataset contains around 1K questions and 4K\nanswers. The dataset is evaluated using 16 LLMs with a focus on overall score,\nstep-by-step scores and error types.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "S1: The detailed, step-by-step score annotation as well as the error type annotation is quite useful and a good addition to existing datasets.\n\nS2: The introduced scores for overall, step-by-step and error are sensible and the evaluation in terms of number of LLMs quite extensive.", "weaknesses": "W1: The student answers seem to be mostly generate by LLMs. Some answers were generated by only six students. The distribution is not clear. How many are from students? How many of the eight generated answers per question are disregarded?\n\nW2: The student answers are not real answers, collected by students that actually have taken the test. This means that the dataset most likely does not contain many of the patterns found in real student responses such as empty and half-completed answers.\n\nW3: Dataset is only in Chinese. It would be more useful to have an English version as well, also to compare inter-lingual differences in terms of performance.", "questions": "Q1: Why are there only roughly 4 student responses per question? This seems very low.\n\nQ2: What is the difference between step-by-step scores and errors? These seem highly related. A correct step will have no error and full score, whereas an error type will have a reduced score.\n\nQ3: Can you add other languages and make the dataset multilingual?\n\nQ4: From the paper it is not clear how many answers have been actually created by students and how many by LLMs. Also, having overall six student annotators is very little and the setting seems not realistic (i.e., these students are given the task to annotate and have not taken the exam)\n\nQ5: Did you perform any evaluation on adversarial attacks on the LLM graders?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761909375287}, {"id": "SFZLTMjNan", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2147/Reviewer_aiMU"], "rating": 4, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new benchmark for Short Answer Scoring (SAS), called SAS-Bench. In SAS, short answer to questions seen in standardized tests are compared to a reference answer. Previous SAS benchmarks lack 2 important attributes, which this paper addresses: firstly, the assesment is not fine grained: in SAS-Bench, answers are broken down to steps using expert annotators. Secondly, each question has a predefined set of possible errors, which the expert annotators use to label the answer steps. Put together, this makes for a benchmark that allows for fine grained analysis of LLM as a judge performance for the SAS task.", "review_text": "This paper introduces a new benchmark for Short Answer Scoring (SAS), called SAS-Bench. In SAS, short answer to questions seen in standardized tests are compared to a reference answer. Previous SAS benchmarks lack 2 important attributes, which this paper addresses: firstly, the assesment is not fine grained: in SAS-Bench, answers are broken down to steps using expert annotators. Secondly, each question has a predefined set of possible errors, which the expert annotators use to label the answer steps. Put together, this makes for a benchmark that allows for fine grained analysis of LLM as a judge performance for the SAS task.", "strengths": "- The paper is clearly written and well-structured.\n- The figures are clean and easily understandable. \n- The evaluation is thorough; a large quantity of reasoning and non-reasoning methods are tested along varied metrics that capture both LLM and expert consistency (CCS), as well as LLM error consistency (ECS). \n- The benchmark and dataset is made up of multiple domains, with many examples: over 1,000 questions as well as over 4,000 expert annotations.", "weaknesses": "- The motivation is unclear to me: in the introduction, works like Zuang et al., 2024, Deshpande et al., 2024, and Raina et al., 2024 are cited, as works that show the shortcomings of LLMs-as-judges. As far as I can tell, none of these works deal with SAS, which results in the question: why are these works used as motivation for creating an SAS benchmark? Additionally, the gap presented in Appendix J also seems somewhat insignificant.  In general, it seems like LLMs already perform relatively well (see next point) on the benchmark, which questions the importance of SAS-Bench to the community. \n- Looking at the consistency between experts and LLMs (CCS), it seems like the top model, V3, has a score of over 74\\%. This suggests that the benchmark is already nearly saturated, which puts into question the strength of the contribution.\n- There seems to be some inherent ambiguity in the construction of the expert annotations. For example, line 264 mentions that experts had disagreements that were resolved in discussions to reach a consensus. It seems like the way to segment an answer into steps, score and label each step could be done in several different reasonable ways, which varies from person to person.", "questions": "1. Are there examples of biases/shortcomings that LLMs exhibit on SAS? Is it possible to quantify them?\n2. Is SAS-Bench saturated, or is there still progress left to be made by LLMs on it? Can you give concrete examples of interesting cases where LLMs are wildly inconsistent with humans on SAS-Bench? \n3. Could there be differences of opinions regarding the way answers are segmented/scored/labeled from expert to expert? Other than discussions, is there a more well-defined way to resolve those differences?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new benchmark for Short Answer Scoring (SAS), called SAS-Bench. In SAS, short answer to questions seen in standardized tests are compared to a reference answer. Previous SAS benchmarks lack 2 important attributes, which this paper addresses: firstly, the assesment is not fine grained: in SAS-Bench, answers are broken down to steps using expert annotators. Secondly, each question has a predefined set of possible errors, which the expert annotators use to label the answer steps. Put together, this makes for a benchmark that allows for fine grained analysis of LLM as a judge performance for the SAS task.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "- The paper is clearly written and well-structured.\n- The figures are clean and easily understandable. \n- The evaluation is thorough; a large quantity of reasoning and non-reasoning methods are tested along varied metrics that capture both LLM and expert consistency (CCS), as well as LLM error consistency (ECS). \n- The benchmark and dataset is made up of multiple domains, with many examples: over 1,000 questions as well as over 4,000 expert annotations.", "weaknesses": "- The motivation is unclear to me: in the introduction, works like Zuang et al., 2024, Deshpande et al., 2024, and Raina et al., 2024 are cited, as works that show the shortcomings of LLMs-as-judges. As far as I can tell, none of these works deal with SAS, which results in the question: why are these works used as motivation for creating an SAS benchmark? Additionally, the gap presented in Appendix J also seems somewhat insignificant.  In general, it seems like LLMs already perform relatively well (see next point) on the benchmark, which questions the importance of SAS-Bench to the community. \n- Looking at the consistency between experts and LLMs (CCS), it seems like the top model, V3, has a score of over 74\\%. This suggests that the benchmark is already nearly saturated, which puts into question the strength of the contribution.\n- There seems to be some inherent ambiguity in the construction of the expert annotations. For example, line 264 mentions that experts had disagreements that were resolved in discussions to reach a consensus. It seems like the way to segment an answer into steps, score and label each step could be done in several different reasonable ways, which varies from person to person.", "questions": "1. Are there examples of biases/shortcomings that LLMs exhibit on SAS? Is it possible to quantify them?\n2. Is SAS-Bench saturated, or is there still progress left to be made by LLMs on it? Can you give concrete examples of interesting cases where LLMs are wildly inconsistent with humans on SAS-Bench? \n3. Could there be differences of opinions regarding the way answers are segmented/scored/labeled from expert to expert? Other than discussions, is there a more well-defined way to resolve those differences?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760741303058}, {"id": "bTpd8VZCKQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2147/Reviewer_5Cgo"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces a new dataset and evaluation framework for automated short-answer scoring (SAS). The dataset comprises short-answer questions and student answers from China’s national university entrance exam, each annotated by experts with scores and error reasons. Extensive experiments show that current LLMs struggle with consistent step-wise scoring, especially on science questions, and that providing scoring rubrics and examples improves performance.", "review_text": "The paper introduces a new dataset and evaluation framework for automated short-answer scoring (SAS). The dataset comprises short-answer questions and student answers from China’s national university entrance exam, each annotated by experts with scores and error reasons. Extensive experiments show that current LLMs struggle with consistent step-wise scoring, especially on science questions, and that providing scoring rubrics and examples improves performance.", "strengths": "- The paper addresses a gap in SAS by introducing two new evaluation metrics: step-wise scoring and error-cause annotations. While prior benchmarks only provide an only score, these new metrics allow for an evaluation of a model’s reasoning process and explanation quality.\n- The SAS-Bench dataset is substantial, well-annotated, and anticipated to be open-source. It contains diverse question types (science, English, math proofs, etc.) and expert annotations. \n- The experiments include diverse comparisons (e.g., analyses of model size, reasoning approach, and training paradigm) yields important insights. For example, one insight from comparing different model sizes is that without complex prompt engineering or task-specific fine-tuning, LLMs tend to underperform smaller LMs in SAS.", "weaknesses": "- More differentiation is needed from existing SAS benchmark datasets such as the Kaggle ASAP-SAS and SemEval-2013 “Student Response Analysis” corpora datasets. The former dataset is introduced in the Related Work but not compared explicitly with SAS-Bench, and the paper associated with the latter dataset (Myroslava et al., 2013) is in the reference list but not discussed at all in the paper. The authors claim SAS-Bench is the first benchmark specifically tailored for SAS with LLMs, but they do not articulate how its novelty goes beyond combining known elements (e.g., reference answers and multidimensional scoring) in a new dataset.\n- The performance tables and charts should include a measure of variance. Since LLM outputs can be stochastic, reporting single-run results without any measure of uncertainty means we can’t tell if differences between models are significant or just noise. Similarly, the authors do not report inter-annotator agreement in the form of Cohen's kappa coefficient for their expert labels.\n- The SAS-Bench dataset, while diverse in subjects, is narrow in source. This raises concerns about how the benchmark’s insights would generalize to other contexts; e.g., different countries’ exams, free-form vs. structured answers, etc. The paper does not discuss whether models performing well on SAS-Bench would generalize to other curricula or languages. \n- A portion of student responses were synthetically granted by LLMs, and then annotated, to augment real answers. The paper should acknowledge that relying on LLM-synthesized data could introduce biases, for example, if LLMs recognize their own style or content in these answers.", "questions": "- Most questions in the dataset are in Chinese (except the English category). Are some of the models evaluated better suited for Chinese language questions?\n- What differentiates, if anything, a “short answer” from a close-ended exact-match answer, such as those in the HLE dataset (arXiv:2501.14249v9)?\n- How does the LLM-as-a-Judge framework used in the paper compare to previous works?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new dataset and evaluation framework for automated short-answer scoring (SAS). The dataset comprises short-answer questions and student answers from China’s national university entrance exam, each annotated by experts with scores and error reasons. Extensive experiments show that current LLMs struggle with consistent step-wise scoring, especially on science questions, and that providing scoring rubrics and examples improves performance.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper addresses a gap in SAS by introducing two new evaluation metrics: step-wise scoring and error-cause annotations. While prior benchmarks only provide an only score, these new metrics allow for an evaluation of a model’s reasoning process and explanation quality.\n- The SAS-Bench dataset is substantial, well-annotated, and anticipated to be open-source. It contains diverse question types (science, English, math proofs, etc.) and expert annotations. \n- The experiments include diverse comparisons (e.g., analyses of model size, reasoning approach, and training paradigm) yields important insights. For example, one insight from comparing different model sizes is that without complex prompt engineering or task-specific fine-tuning, LLMs tend to underperform smaller LMs in SAS.", "weaknesses": "- More differentiation is needed from existing SAS benchmark datasets such as the Kaggle ASAP-SAS and SemEval-2013 “Student Response Analysis” corpora datasets. The former dataset is introduced in the Related Work but not compared explicitly with SAS-Bench, and the paper associated with the latter dataset (Myroslava et al., 2013) is in the reference list but not discussed at all in the paper. The authors claim SAS-Bench is the first benchmark specifically tailored for SAS with LLMs, but they do not articulate how its novelty goes beyond combining known elements (e.g., reference answers and multidimensional scoring) in a new dataset.\n- The performance tables and charts should include a measure of variance. Since LLM outputs can be stochastic, reporting single-run results without any measure of uncertainty means we can’t tell if differences between models are significant or just noise. Similarly, the authors do not report inter-annotator agreement in the form of Cohen's kappa coefficient for their expert labels.\n- The SAS-Bench dataset, while diverse in subjects, is narrow in source. This raises concerns about how the benchmark’s insights would generalize to other contexts; e.g., different countries’ exams, free-form vs. structured answers, etc. The paper does not discuss whether models performing well on SAS-Bench would generalize to other curricula or languages. \n- A portion of student responses were synthetically granted by LLMs, and then annotated, to augment real answers. The paper should acknowledge that relying on LLM-synthesized data could introduce biases, for example, if LLMs recognize their own style or content in these answers.", "questions": "- Most questions in the dataset are in Chinese (except the English category). Are some of the models evaluated better suited for Chinese language questions?\n- What differentiates, if anything, a “short answer” from a close-ended exact-match answer, such as those in the HLE dataset (arXiv:2501.14249v9)?\n- How does the LLM-as-a-Judge framework used in the paper compare to previous works?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760725585843}], "openreview_url": "https://openreview.net/forum?id=bGVkAhR7Eu", "arxiv_id": "2505.07247", "paper_pdf": "papers/bGVkAhR7Eu.pdf", "paper_pdf_sha256": "86b7a1795aa98286c9201640fc2682f91c1810dc6ce2d2a22ae6a586740d7b7a", "paper_pdf_bytes": 1735852, "paper_pdf_source": "openreview", "code_url": "https://github.com/PKU-DAIR/SAS-Bench", "code_repository": "PKU-DAIR/SAS-Bench", "code_commit": "4ef718b6433623d492d483a38d4045a4b32d6f45", "code_archive": "repos/bGVkAhR7Eu.zip", "code_archive_sha256": "7f5c5fa7b2065714bf5d4caffbbe145f0641dcd6707912a9d11d839f31d051fd", "code_archive_bytes": 1048715, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 1255, "github_languages": {"Python": 147605}, "github_archived": false, "github_pushed_at": "2025-05-15T11:12:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sas-bench-a-fine-grained-benchmark-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bff9RniI03", "year": 2025, "status": "rejected", "title": "Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration", "authors": ["Max Wilcoxson", "Qiyang Li", "Kevin Frans", "Sergey Levine"], "authorids": ["~Max_Wilcoxson1", "~Qiyang_Li1", "~Kevin_Frans1", "~Sergey_Levine1"], "authors_source": "OpenReview API", "abstract": "Unsupervised pretraining has been transformative in many supervised domains. However, applying such ideas to reinforcement learning (RL) presents a unique challenge in that fine-tuning does not involve mimicking task-specific data, but rather exploring and locating the solution through iterative self-improvement. In this work, we showcase how unlabeled prior trajectory data can be leveraged to learn efficient exploration strategies. The key insight is to use unlabelled trajectories twice, 1) to extract a set of low-level skills offline, and 2) as additional data for a high-level policy that composes these skills to explore. We utilize a simple strategy of learning an optimistic reward model from online samples, and relabeling past trajectories into high-level, task-relevant examples. We instantiate these insights as SUPE (Skills from Unlabeled Prior data for Exploration), and empirically show that SUPE reliably outperforms prior strategies, successfully solving a suite of long-horizon, sparse-reward tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "2bEyTMUlT5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12140/Reviewer_eVtj"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes a hierarchical policy for leveraging unlabeled offline data for exploration. In the offline stage, low-level skills are extracted, and in the online stage, these skills are reused and a high-level policy is learned with optimistic rewards. The proposed method is tested on maze and manipulation tasks and shows good performance.", "review_text": "The paper proposes a hierarchical policy for leveraging unlabeled offline data for exploration. In the offline stage, low-level skills are extracted, and in the online stage, these skills are reused and a high-level policy is learned with optimistic rewards. The proposed method is tested on maze and manipulation tasks and shows good performance.", "strengths": "- The paper is well-written and easy to understand.\n- The paper proposes a simple method for leveraging offline data and showing good performance on AntMaze, visual AntMaze, and Kitchen tasks. \n- The paper conducts thorough experiments and compares a set of different methods.", "weaknesses": "- Dependence on offline data quality: The performance of the proposed method is influenced by the quality of the offline data and the specific features of the evaluation tasks. In particular, the approach relies on a high-level policy that is updated every \n𝐻 timesteps and keeps the pre-trained skill and trajectory encoder fixed during the online phase. This limitation constrains adaptability, especially in scenarios where task distribution varies from the offline data.\n- Limited discussion on Hierarchical Reinforcement Learning (HRL): Although hierarchical policy structures have been extensively explored in the HRL literature [1-8] and are closely related to the paper, the paper does not sufficiently address relevant findings from HRL research. A more comprehensive discussion of how this work could provide valuable context.\n- Novelty: The paper combines elements from ExPLORe and trajectory-segment VAE to leverage offline data for exploration, but adds limited new insights beyond prior work. HRL emphasizes hierarchical structures, and the benefits of skill extraction in offline settings have already been documented. This paper simply applies existing solutions to ExPLORe.\n\nThe paper could be improved in several aspects:\n- Refinement of skill extraction method: Currently, skills are extracted based on fixed-length trajectory segments, a method that may overlook important nuances in skills. A more flexible or adaptive approach could address these limitations, potentially enhancing the robustness of the extracted skills.\n- Skill adaptation during the online stage: The method does not allow for online adaptation of the skill policy or trajectory encoder. Due to potential distributional shifts between the offline and online data, enabling adaptive updates to the skill set and encoder could further improve the performance.\n- Training stability in HRL is often affected by interactions between high-level and low-level policies. This work could benefit from discussing how offline data might address or mitigate these stability challenges.\n\n[1] Kulkarni, Tejas D., et al. \"Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation.\" Advances in neural information processing systems 29 (2016).\n\n[2] Xie, Kevin, et al. \"Latent skill planning for exploration and transfer.\" arXiv preprint arXiv:2011.13897 (2020).\n\n[3] Nachum, Ofir, et al. \"Data-efficient hierarchical reinforcement learning.\" Advances in neural information processing systems 31 (2018).\n\n[4] Bacon, Pierre-Luc, Jean Harb, and Doina Precup. \"The option-critic architecture.\" Proceedings of the AAAI conference on artificial intelligence. Vol. 31. No. 1. 2017.\n\n[5] Ajay, Anurag, et al. \"Opal: Offline primitive discovery for accelerating offline reinforcement learning.\" arXiv preprint arXiv:2010.13611 (2020).\n\n[6] Gehring, Jonas, et al. \"Hierarchical skills for efficient exploration.\" Advances in Neural Information Processing Systems 34 (2021): 11553-11564.\n\n[7] Dalal, Murtaza, Deepak Pathak, and Russ R. Salakhutdinov. \"Accelerating robotic reinforcement learning via parameterized action primitives.\" Advances in Neural Information Processing Systems 34 (2021): 21847-21859.\n\n[8] Paraschos, Alexandros, et al. \"Probabilistic movement primitives.\" Advances in neural information processing systems 26 (2013).", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a hierarchical policy for leveraging unlabeled offline data for exploration. In the offline stage, low-level skills are extracted, and in the online stage, these skills are reused and a high-level policy is learned with optimistic rewards. The proposed method is tested on maze and manipulation tasks and shows good performance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-written and easy to understand.\n- The paper proposes a simple method for leveraging offline data and showing good performance on AntMaze, visual AntMaze, and Kitchen tasks. \n- The paper conducts thorough experiments and compares a set of different methods.", "weaknesses": "- Dependence on offline data quality: The performance of the proposed method is influenced by the quality of the offline data and the specific features of the evaluation tasks. In particular, the approach relies on a high-level policy that is updated every \n𝐻 timesteps and keeps the pre-trained skill and trajectory encoder fixed during the online phase. This limitation constrains adaptability, especially in scenarios where task distribution varies from the offline data.\n- Limited discussion on Hierarchical Reinforcement Learning (HRL): Although hierarchical policy structures have been extensively explored in the HRL literature [1-8] and are closely related to the paper, the paper does not sufficiently address relevant findings from HRL research. A more comprehensive discussion of how this work could provide valuable context.\n- Novelty: The paper combines elements from ExPLORe and trajectory-segment VAE to leverage offline data for exploration, but adds limited new insights beyond prior work. HRL emphasizes hierarchical structures, and the benefits of skill extraction in offline settings have already been documented. This paper simply applies existing solutions to ExPLORe.\n\nThe paper could be improved in several aspects:\n- Refinement of skill extraction method: Currently, skills are extracted based on fixed-length trajectory segments, a method that may overlook important nuances in skills. A more flexible or adaptive approach could address these limitations, potentially enhancing the robustness of the extracted skills.\n- Skill adaptation during the online stage: The method does not allow for online adaptation of the skill policy or trajectory encoder. Due to potential distributional shifts between the offline and online data, enabling adaptive updates to the skill set and encoder could further improve the performance.\n- Training stability in HRL is often affected by interactions between high-level and low-level policies. This work could benefit from discussing how offline data might address or mitigate these stability challenges.\n\n[1] Kulkarni, Tejas D., et al. \"Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation.\" Advances in neural information processing systems 29 (2016).\n\n[2] Xie, Kevin, et al. \"Latent skill planning for exploration and transfer.\" arXiv preprint arXiv:2011.13897 (2020).\n\n[3] Nachum, Ofir, et al. \"Data-efficient hierarchical reinforcement learning.\" Advances in neural information processing systems 31 (2018).\n\n[4] Bacon, Pierre-Luc, Jean Harb, and Doina Precup. \"The option-critic architecture.\" Proceedings of the AAAI conference on artificial intelligence. Vol. 31. No. 1. 2017.\n\n[5] Ajay, Anurag, et al. \"Opal: Offline primitive discovery for accelerating offline reinforcement learning.\" arXiv preprint arXiv:2010.13611 (2020).\n\n[6] Gehring, Jonas, et al. \"Hierarchical skills for efficient exploration.\" Advances in Neural Information Processing Systems 34 (2021): 11553-11564.\n\n[7] Dalal, Murtaza, Deepak Pathak, and Russ R. Salakhutdinov. \"Accelerating robotic reinforcement learning via parameterized action primitives.\" Advances in Neural Information Processing Systems 34 (2021): 21847-21859.\n\n[8] Paraschos, Alexandros, et al. \"Probabilistic movement primitives.\" Advances in neural information processing systems 26 (2013).", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730652500426}, {"id": "7JU89hOH7k", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12140/Reviewer_28dh"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 2, "summary": "This paper presents SUPE, a method for using offline data (without rewards) in the online reinforcement learning setting. SUPE first extracts a set of low level skills using the offline data, and then optimistically labels the offline trajectories. It then uses an off policy high level update to update on a mix of offline (pseudo labeled trajectories) and online real trajectories. The paper empirically validates the new algorithm on three environments and does ablations on amounts of offline data.", "review_text": "This paper presents SUPE, a method for using offline data (without rewards) in the online reinforcement learning setting. SUPE first extracts a set of low level skills using the offline data, and then optimistically labels the offline trajectories. It then uses an off policy high level update to update on a mix of offline (pseudo labeled trajectories) and online real trajectories. The paper empirically validates the new algorithm on three environments and does ablations on amounts of offline data.", "strengths": "- This paper makes an insightful empirical benefit for using trajectories twice for both low level skill pretraining in addition to optimistic labelling.\n- The paper thoroughly evaluates the proposed method.\n- The paper does a good job explaining the proposed method and it's significance.", "weaknesses": "- This paper could benefit from a bit deeper analysis of the contribution of the two uses of offline data. It's clear that both are necessary, but not necessarily why.", "questions": "- Where do the authors think their empirical benefit is coming from? Why can we use trajectories twice?\n- Is the algorithm robust to different design choices?\n- How important is the optimistic labelling (from Li et al.)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents SUPE, a method for using offline data (without rewards) in the online reinforcement learning setting. SUPE first extracts a set of low level skills using the offline data, and then optimistically labels the offline trajectories. It then uses an off policy high level update to update on a mix of offline (pseudo labeled trajectories) and online real trajectories. The paper empirically validates the new algorithm on three environments and does ablations on amounts of offline data.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- This paper makes an insightful empirical benefit for using trajectories twice for both low level skill pretraining in addition to optimistic labelling.\n- The paper thoroughly evaluates the proposed method.\n- The paper does a good job explaining the proposed method and it's significance.", "weaknesses": "- This paper could benefit from a bit deeper analysis of the contribution of the two uses of offline data. It's clear that both are necessary, but not necessarily why.", "questions": "- Where do the authors think their empirical benefit is coming from? Why can we use trajectories twice?\n- Is the algorithm robust to different design choices?\n- How important is the optimistic labelling (from Li et al.)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730642840685}, {"id": "vlz5kbnqXJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12140/Reviewer_DC3r"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces SUPE, a method that leverages unsupervised learning to extract skills from unlabeled prior data, subsequently using hierarchical methods to explore more efficiently. These unlabeled data can also contribute to high-level policy training. Experimental results show that SUPE outperforms previous methods on the D4RL benchmark.", "review_text": "This paper introduces SUPE, a method that leverages unsupervised learning to extract skills from unlabeled prior data, subsequently using hierarchical methods to explore more efficiently. These unlabeled data can also contribute to high-level policy training. Experimental results show that SUPE outperforms previous methods on the D4RL benchmark.", "strengths": "- The approach of extracting latent “skills” from unlabeled data and employing hierarchical methods significantly enhances exploration.\n- The approach of utilizing prior data twice ensures better use of the available data.\n- The paper is well-structured and easy to follow.\n- Extensive results demonstrate that this method outperforms previous approaches.", "weaknesses": "- The concept of using a VAE to extract latent codes and employing a high-level policy for online exploration is not novel, and it shows limited progress compared to previous work [1].\n- The ablation study lacks depth. I am interested in understanding the contribution of “reusing prior data twice” to the final performance. Additionally, I’d like clarification on the design choice for the latent variable $z$ in skill discovery: how do you ensure this latent $z$  is sufficient for effective skill discovery in the dataset? Is employing trajectory-segment VAEs truly necessary for efficient exploration?\n\n\n[1] Qiyang Li, Jason Zhang, Dibya Ghosh, Amy Zhang, and Sergey Levine. Accelerating exploration with unlabeled prior data. Advances in Neural Information Processing Systems, 36, 2024.", "questions": "Please refer to the weakness part. I may consider increasing the score if my questions are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SUPE, a method that leverages unsupervised learning to extract skills from unlabeled prior data, subsequently using hierarchical methods to explore more efficiently. These unlabeled data can also contribute to high-level policy training. Experimental results show that SUPE outperforms previous methods on the D4RL benchmark.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The approach of extracting latent “skills” from unlabeled data and employing hierarchical methods significantly enhances exploration.\n- The approach of utilizing prior data twice ensures better use of the available data.\n- The paper is well-structured and easy to follow.\n- Extensive results demonstrate that this method outperforms previous approaches.", "weaknesses": "- The concept of using a VAE to extract latent codes and employing a high-level policy for online exploration is not novel, and it shows limited progress compared to previous work [1].\n- The ablation study lacks depth. I am interested in understanding the contribution of “reusing prior data twice” to the final performance. Additionally, I’d like clarification on the design choice for the latent variable $z$ in skill discovery: how do you ensure this latent $z$  is sufficient for effective skill discovery in the dataset? Is employing trajectory-segment VAEs truly necessary for efficient exploration?\n\n\n[1] Qiyang Li, Jason Zhang, Dibya Ghosh, Amy Zhang, and Sergey Levine. Accelerating exploration with unlabeled prior data. Advances in Neural Information Processing Systems, 36, 2024.", "questions": "Please refer to the weakness part. I may consider increasing the score if my questions are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730426046080}, {"id": "2t6cmHPLoN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12140/Reviewer_CDzd"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents a pre-training method for reinforcement learning (RL) that can train on data sets that do not contain reward labels, i.e., the data sets are unlabeled. \nThe problem setting resembles offline-to-online RL, except that there are no rewards in the data set.\nIn the pre-training stage, the authors propose to learn a set of skills from this unlabeled offline data.\nThen, in the online fine-tuning state, the authors learn a high-level policy that selects which skill to use in a given state.\nThey utilize the unlabeled offline data during fine-tuning by learning an optimistic reward model and using it to add optimistic reward labels to the offline data.\nThey evaluate their method in the D4RL AntMaze and Kitchen benchmarks as well as the D4RL Visual AntMaze.", "review_text": "This paper presents a pre-training method for reinforcement learning (RL) that can train on data sets that do not contain reward labels, i.e., the data sets are unlabeled. \nThe problem setting resembles offline-to-online RL, except that there are no rewards in the data set.\nIn the pre-training stage, the authors propose to learn a set of skills from this unlabeled offline data.\nThen, in the online fine-tuning state, the authors learn a high-level policy that selects which skill to use in a given state.\nThey utilize the unlabeled offline data during fine-tuning by learning an optimistic reward model and using it to add optimistic reward labels to the offline data.\nThey evaluate their method in the D4RL AntMaze and Kitchen benchmarks as well as the D4RL Visual AntMaze.", "strengths": "Overall, I found the paper easy to follow and I think it is addressing an important problem -- pre-training in RL -- which is of interest to the community.\n\nThe results demonstrate that learning skills from offline data is a promising approach to leverage reward-free offline data.\nI think this is an interesting result.\nI also like the idea of labelling the offline data using a learned reward function.", "weaknesses": "The authors consider the setting of having access to offline data but no reward labels.\nWhilst I see the value in this problem setting, it is not clear if practitioners should opt for this method over standard offline-to-online RL methods when\ntheir data sets contain reward labels.\nWhilst I appreciate this is out-of-scope, ideally methods would leverage data sets both with and without reward labels.\nIt would be insightful if the authors could compare to offline-to-online RL methods which do leverage reward labels.\nWhilst I do not expect their method to outperform these methods, I think it is an important baseline that we can gain insights from.\n\nIn my experience, optimistic-based exploration methods are very susceptible to the $\\alpha$ parameter.\nHow was this set in practice?\nDid it require a grid search to find the best value in each environment?\nPlease can you provide details on any hyperparameter tuning process, including the range of values tested and how sensitivity varied across environments?\nThis information would be valuable for reproducibility and understanding the robustness of the method.\n\nIs there a reason the authors only considered the diverse data set for the AntMaze experiments?\nDoes this method require a diverse offline data set collected by an unsupervised RL method,\nor can it leverage narrow offline data distributions? For example, data from solving a different task?\nHow does the method perform when using the AntMaze \"play\" data set instead of the \"diverse\" data set?\nEven if the method performs poorly, I think it would be valuable to include these results.\n\nI am not sure what to take from the coverage results.\nI can understand why we care about coverage in unsupervised RL where our sole purpose is to explore.\nHowever, during online training our goal is to balance exploration vs exploitation.\nPlease can the authors provide a clearer justification for why coverage is an important metric in this context, or include additional plots that more directly show the relationship between exploration and task performance, such as the normalized return vs coverage?\n\nIn Table 1, what do the bold numbers represent? The authors should state what statistical test was used for the bolding or at least expla8in what the bolding represents.\n\n## Minor comments and corrections\n- Line 42 - \"can broken\" should be \"can be broken\"\n- Line 117 - \"of an offline data\" should be \"of offline data\"\n- Line 200 - the term \"latent code\" is misleading. This suggests the trajectory encoder learns to map trajectories to discrete codes from a codebook and I don't think this is the case. The authors should change it to something like \"latent skill\".\n- Line 279 - Should \"Three AntMaze layouts with four different goal location configuration each.\" be \"Three AntMaze layouts with four configurable goal locations each.\"\n- Line 411-414 - It would make more sense for this paragraph to be at the start of Section 5.4.\n- Line 407 - \"Kitchen the domain\" should be \"Kitchen domain\"\n- Line 408 - \"more challenging the kitchen-mixed\" should be \"more challenging kitchen-mixed\"\n- Figures - The authors have stated that the shaded area indicates the standard error. They also need to state what that solid line indicates. Is it the mean, median, etc?\n- Figures - I found the figures very hard to read. I would suggest the authors colour the text \"HILP w/ Offline Data\", \"Ours\", \"Online w/ HILP Skills\", etc, to match the colours of the lines in the plots. This would make the text/figures much easier to read.", "questions": "- How does your method compare to using offline-to-online RL methods which have access to reward labels?\n- How was the $\\alpha$ hyperparameter set?\n- Why did you not compare to other types of offline data sets?\n- What should I take from the coverage results?\n- In Table 1, what do the bold numbers represent?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a pre-training method for reinforcement learning (RL) that can train on data sets that do not contain reward labels, i.e., the data sets are unlabeled. \nThe problem setting resembles offline-to-online RL, except that there are no rewards in the data set.\nIn the pre-training stage, the authors propose to learn a set of skills from this unlabeled offline data.\nThen, in the online fine-tuning state, the authors learn a high-level policy that selects which skill to use in a given state.\nThey utilize the unlabeled offline data during fine-tuning by learning an optimistic reward model and using it to add optimistic reward labels to the offline data.\nThey evaluate their method in the D4RL AntMaze and Kitchen benchmarks as well as the D4RL Visual AntMaze.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Overall, I found the paper easy to follow and I think it is addressing an important problem -- pre-training in RL -- which is of interest to the community.\n\nThe results demonstrate that learning skills from offline data is a promising approach to leverage reward-free offline data.\nI think this is an interesting result.\nI also like the idea of labelling the offline data using a learned reward function.", "weaknesses": "The authors consider the setting of having access to offline data but no reward labels.\nWhilst I see the value in this problem setting, it is not clear if practitioners should opt for this method over standard offline-to-online RL methods when\ntheir data sets contain reward labels.\nWhilst I appreciate this is out-of-scope, ideally methods would leverage data sets both with and without reward labels.\nIt would be insightful if the authors could compare to offline-to-online RL methods which do leverage reward labels.\nWhilst I do not expect their method to outperform these methods, I think it is an important baseline that we can gain insights from.\n\nIn my experience, optimistic-based exploration methods are very susceptible to the $\\alpha$ parameter.\nHow was this set in practice?\nDid it require a grid search to find the best value in each environment?\nPlease can you provide details on any hyperparameter tuning process, including the range of values tested and how sensitivity varied across environments?\nThis information would be valuable for reproducibility and understanding the robustness of the method.\n\nIs there a reason the authors only considered the diverse data set for the AntMaze experiments?\nDoes this method require a diverse offline data set collected by an unsupervised RL method,\nor can it leverage narrow offline data distributions? For example, data from solving a different task?\nHow does the method perform when using the AntMaze \"play\" data set instead of the \"diverse\" data set?\nEven if the method performs poorly, I think it would be valuable to include these results.\n\nI am not sure what to take from the coverage results.\nI can understand why we care about coverage in unsupervised RL where our sole purpose is to explore.\nHowever, during online training our goal is to balance exploration vs exploitation.\nPlease can the authors provide a clearer justification for why coverage is an important metric in this context, or include additional plots that more directly show the relationship between exploration and task performance, such as the normalized return vs coverage?\n\nIn Table 1, what do the bold numbers represent? The authors should state what statistical test was used for the bolding or at least expla8in what the bolding represents.\n\n## Minor comments and corrections\n- Line 42 - \"can broken\" should be \"can be broken\"\n- Line 117 - \"of an offline data\" should be \"of offline data\"\n- Line 200 - the term \"latent code\" is misleading. This suggests the trajectory encoder learns to map trajectories to discrete codes from a codebook and I don't think this is the case. The authors should change it to something like \"latent skill\".\n- Line 279 - Should \"Three AntMaze layouts with four different goal location configuration each.\" be \"Three AntMaze layouts with four configurable goal locations each.\"\n- Line 411-414 - It would make more sense for this paragraph to be at the start of Section 5.4.\n- Line 407 - \"Kitchen the domain\" should be \"Kitchen domain\"\n- Line 408 - \"more challenging the kitchen-mixed\" should be \"more challenging kitchen-mixed\"\n- Figures - The authors have stated that the shaded area indicates the standard error. They also need to state what that solid line indicates. Is it the mean, median, etc?\n- Figures - I found the figures very hard to read. I would suggest the authors colour the text \"HILP w/ Offline Data\", \"Ours\", \"Online w/ HILP Skills\", etc, to match the colours of the lines in the plots. This would make the text/figures much easier to read.", "questions": "- How does your method compare to using offline-to-online RL methods which have access to reward labels?\n- How was the $\\alpha$ hyperparameter set?\n- Why did you not compare to other types of offline data sets?\n- What should I take from the coverage results?\n- In Table 1, what do the bold numbers represent?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729608452769}, {"id": "dpbHFIo8XA", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12140/Reviewer_t68W"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a two-phase framework, SUPE, which leverages data in two stages: first, extracting low-level skills during the offline pre-training phase, and then using these skills and unlabeled data in the online phase to train a high-level strategy for more efficient exploration. Building on prior works like SPiRL [1] and ExPLORe [2], the key contribution of this paper is to integrate unlabeled data with online data to accelerate exploration and training in off-policy reinforcement learning (RL) methods. In the offline pre-training stage, the authors train a set of low-level skills, while in the online phase, they develop a high-level policy by utilizing both online data and relabeled offline data. To assess the method’s effectiveness, the authors compare SUPE with several baselines using benchmarks such as D4RL, and also discuss its limitations and potential directions for future research.\n\n[1] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.\n\n[2] Li, Qiyang, et al. \"Accelerating exploration with unlabeled prior data.\" Advances in Neural Information Processing Systems 36 (2024).", "review_text": "This paper proposes a two-phase framework, SUPE, which leverages data in two stages: first, extracting low-level skills during the offline pre-training phase, and then using these skills and unlabeled data in the online phase to train a high-level strategy for more efficient exploration. Building on prior works like SPiRL [1] and ExPLORe [2], the key contribution of this paper is to integrate unlabeled data with online data to accelerate exploration and training in off-policy reinforcement learning (RL) methods. In the offline pre-training stage, the authors train a set of low-level skills, while in the online phase, they develop a high-level policy by utilizing both online data and relabeled offline data. To assess the method’s effectiveness, the authors compare SUPE with several baselines using benchmarks such as D4RL, and also discuss its limitations and potential directions for future research.\n\n[1] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.\n\n[2] Li, Qiyang, et al. \"Accelerating exploration with unlabeled prior data.\" Advances in Neural Information Processing Systems 36 (2024).", "strengths": "* The paper is highly detailed, well-written and provides detailed motivation. The complete code is also provided.\n* The authors conduct numerous experiments to thoroughly validate their method and address in detail several key issues that I am particularly concerned about, including its scalability, robustness.", "weaknesses": "* The overall novelty of this work is somewhat limited, as it builds heavily on existing methods and concepts (mentioned in summary). \n* Although numerous experiments are conducted, the selected tasks are relatively monotonous and simplistic. The experiments test only two types of tasks: AntMaze and Kitchen.", "questions": "* See weakness above. \n* Given the similarities between SPiRL [1] and this work, apart from the online reinforcement learning stage, why isn’t SPiRL used as a baseline for comparison (despite the numerous experiments conducted) ?\n* In the pre-training stage, it would also be valuable to discuss whether trajectory segment length $H$ significantly impacts the method's performance.\n* I am curious whether using expert data would result in better low-level skills during the pre-training stage.\n\n[1] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a two-phase framework, SUPE, which leverages data in two stages: first, extracting low-level skills during the offline pre-training phase, and then using these skills and unlabeled data in the online phase to train a high-level strategy for more efficient exploration. Building on prior works like SPiRL [1] and ExPLORe [2], the key contribution of this paper is to integrate unlabeled data with online data to accelerate exploration and training in off-policy reinforcement learning (RL) methods. In the offline pre-training stage, the authors train a set of low-level skills, while in the online phase, they develop a high-level policy by utilizing both online data and relabeled offline data. To assess the method’s effectiveness, the authors compare SUPE with several baselines using benchmarks such as D4RL, and also discuss its limitations and potential directions for future research.\n\n[1] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.\n\n[2] Li, Qiyang, et al. \"Accelerating exploration with unlabeled prior data.\" Advances in Neural Information Processing Systems 36 (2024).", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* The paper is highly detailed, well-written and provides detailed motivation. The complete code is also provided.\n* The authors conduct numerous experiments to thoroughly validate their method and address in detail several key issues that I am particularly concerned about, including its scalability, robustness.", "weaknesses": "* The overall novelty of this work is somewhat limited, as it builds heavily on existing methods and concepts (mentioned in summary). \n* Although numerous experiments are conducted, the selected tasks are relatively monotonous and simplistic. The experiments test only two types of tasks: AntMaze and Kitchen.", "questions": "* See weakness above. \n* Given the similarities between SPiRL [1] and this work, apart from the online reinforcement learning stage, why isn’t SPiRL used as a baseline for comparison (despite the numerous experiments conducted) ?\n* In the pre-training stage, it would also be valuable to discuss whether trajectory segment length $H$ significantly impacts the method's performance.\n* I am curious whether using expert data would result in better low-level skills during the pre-training stage.\n\n[1] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729315002554}], "openreview_url": "https://openreview.net/forum?id=Bff9RniI03", "arxiv_id": "2410.18076", "paper_pdf": "papers/Bff9RniI03.pdf", "paper_pdf_sha256": "76e28b3ec521581c18dd9a7d09a3afa78044539d5f2018734d7ed3d05e5bf66f", "paper_pdf_bytes": 8531032, "paper_pdf_source": "openreview", "code_url": "https://github.com/rail-berkeley/SUPE", "code_repository": "rail-berkeley/SUPE", "code_commit": "b4d1dfbc1e40638ee4ab2d3cc06782f26950c5b5", "code_archive": "repos/Bff9RniI03.zip", "code_archive_sha256": "bff5af27b6d867e408f7af29b65874fd806b015e5ff0115710de7bb1c51dbb19", "code_archive_bytes": 108725, "code_file_count": 75, "code_extensions": {".py": 74, ".sh": 1}, "github_disk_usage_kb": 92, "github_languages": {"Python": 315447, "Shell": 1427}, "github_archived": false, "github_pushed_at": "2025-07-11T20:35:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/leveraging-skills-from-unlabeled-prior-data"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3mY9aGiMn0", "year": 2024, "status": "rejected", "title": "Sparser, Better, Deeper, Stronger: Improving Sparse Training with Exact Orthogonal Initialization", "authors": ["Aleksandra Nowak", "Łukasz Gniecki", "Filip Szatkowski", "Jacek Tabor"], "authorids": ["~Aleksandra_Nowak1", "~Łukasz_Gniecki1", "~Filip_Szatkowski1", "~Jacek_Tabor1"], "authors_source": "OpenReview API", "abstract": "Sparse training aims to train sparse models from scratch, achieving remarkable results in recent years. A key design choice in sparse training is the sparse initialization, which determines the trainable sub-network through a binary mask. Existing methods mainly revolve around selecting the mask based on predefined dense weight initialization. However, such an approach may not efficiently leverage the mask's potential impact on training parameters and optimization. An alternative direction, inspired by research into dynamical isometry, is to introduce orthogonality in the sparse subnetwork. This helps prevent the gradient signal from vanishing or exploding, ultimately enhancing the reliability of the backpropagation process. In this work, we propose Exact Orthogonal Initialization (EOI), a novel sparse orthogonal initialization scheme based on composing random Givens rotations. Contrary to other existing approaches, our method provides exact (not approximated) orthogonality and enables the creation of layers with arbitrary densities. Through experiments on contemporary network architectures, we present the effectiveness of EOI and demonstrate that it consistently outperforms other commonly used sparse initialization techniques. Furthermore, to showcase the full potential of our method, we show that it enables the training of highly sparse 1000-layer MLP and CNN networks without any residual connections or normalization techniques.  Our research highlights the importance of weight initialization in sparse training, underscoring the vital part it plays alongside the sparse mask selection.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "hhztVMTvQJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4859/Reviewer_s3UF"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this work, the authors propose a novel approach to sparse training, a technique aimed at training models with sparse structures from the beginning. The key element in sparse training is the sparse initialization, which determines which parts of the model are trainable through a binary mask. Existing methods often rely on predefined dense weight initialization to create these masks. However, such an approach might not efficiently harness the potential impact of these masks on the training process and optimization.\n\nInspired by research on dynamical isometry, the authors take an alternative route by introducing orthogonality into the sparse subnetwork. This orthogonality helps mitigate issues related to the vanishing or exploding gradient signal, ultimately making the backpropagation process more reliable.\n\nThe authors introduce their novel method called Exact Orthogonal Initialization (EOI). Unlike other existing approaches, EOI provides exact orthogonality, avoiding approximations. It also allows for the creation of layers with various densities.", "review_text": "In this work, the authors propose a novel approach to sparse training, a technique aimed at training models with sparse structures from the beginning. The key element in sparse training is the sparse initialization, which determines which parts of the model are trainable through a binary mask. Existing methods often rely on predefined dense weight initialization to create these masks. However, such an approach might not efficiently harness the potential impact of these masks on the training process and optimization.\n\nInspired by research on dynamical isometry, the authors take an alternative route by introducing orthogonality into the sparse subnetwork. This orthogonality helps mitigate issues related to the vanishing or exploding gradient signal, ultimately making the backpropagation process more reliable.\n\nThe authors introduce their novel method called Exact Orthogonal Initialization (EOI). Unlike other existing approaches, EOI provides exact orthogonality, avoiding approximations. It also allows for the creation of layers with various densities.", "strengths": "The paper is well-written with a clear structure, making it easily readable for the audience.", "weaknesses": "However, it appears that the authors are more dedicated to highlighting the advantages of sparsity and orthogonality than proposing and demonstrating an efficient algorithm. There is a shortage of comparisons with similar algorithms in the experiments. The EOI algorithm does not seem to exhibit a significant advantage. The comparison results in Figure 3, along with the author's analysis, raise questions about whether the AI method could replace EOI. It's not clear where the innovation lies.\n\nIn Figure 5, it's unclear if the time curves represent that EOI significantly underperforms the SAO algorithm as matrix size and density increase. \n\nIn summary, the paper is well-structured and easy to read, but it lacks extensive comparisons with similar algorithms in the experiments, and the advantages of the EOI algorithm are not convincingly demonstrated. Clarity is needed in the interpretation of the results, especially in Figures 3 and 5. The meaning of bold and underlined entries in Table 3 requires clarification.", "questions": "P4L3: Is the k of W_{ijkk} same as the k from dimension 2k+1?\nThe meaning of bold and underlined entries in Table 3 is unclear, such as why some values in the 'VGG-16-GraSP-LReLU' column are bold in the middle.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors propose a novel approach to sparse training, a technique aimed at training models with sparse structures from the beginning. The key element in sparse training is the sparse initialization, which determines which parts of the model are trainable through a binary mask. Existing methods often rely on predefined dense weight initialization to create these masks. However, such an approach might not efficiently harness the potential impact of these masks on the training process and optimization.\n\nInspired by research on dynamical isometry, the authors take an alternative route by introducing orthogonality into the sparse subnetwork. This orthogonality helps mitigate issues related to the vanishing or exploding gradient signal, ultimately making the backpropagation process more reliable.\n\nThe authors introduce their novel method called Exact Orthogonal Initialization (EOI). Unlike other existing approaches, EOI provides exact orthogonality, avoiding approximations. It also allows for the creation of layers with various densities.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well-written with a clear structure, making it easily readable for the audience.", "weaknesses": "However, it appears that the authors are more dedicated to highlighting the advantages of sparsity and orthogonality than proposing and demonstrating an efficient algorithm. There is a shortage of comparisons with similar algorithms in the experiments. The EOI algorithm does not seem to exhibit a significant advantage. The comparison results in Figure 3, along with the author's analysis, raise questions about whether the AI method could replace EOI. It's not clear where the innovation lies.\n\nIn Figure 5, it's unclear if the time curves represent that EOI significantly underperforms the SAO algorithm as matrix size and density increase. \n\nIn summary, the paper is well-structured and easy to read, but it lacks extensive comparisons with similar algorithms in the experiments, and the advantages of the EOI algorithm are not convincingly demonstrated. Clarity is needed in the interpretation of the results, especially in Figures 3 and 5. The meaning of bold and underlined entries in Table 3 requires clarification.", "questions": "P4L3: Is the k of W_{ijkk} same as the k from dimension 2k+1?\nThe meaning of bold and underlined entries in Table 3 is unclear, such as why some values in the 'VGG-16-GraSP-LReLU' column are bold in the middle.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698632873637}, {"id": "YrQRdcDok4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4859/Reviewer_7ET6"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "Authors have proposed a method to achieve exact orthogonal initialization for training very deep neural models with sparsity constraints. The approach is built upon recent advancements achieved via tools from the Random Matrix theory to improve initialization and achieve training acceleration.", "review_text": "Authors have proposed a method to achieve exact orthogonal initialization for training very deep neural models with sparsity constraints. The approach is built upon recent advancements achieved via tools from the Random Matrix theory to improve initialization and achieve training acceleration.", "strengths": "The proposed method is simple, mathematically grounded, and addresses an important issue of sparsity-aware training in DNNs.\nThe flow of the paper is nice and structured.", "weaknesses": "Overall, the paper is well-written, but often, critical details are missing, which might make it difficult for a new reader to understand and appreciate the ideas discussed and their connection to prior art. For instance, the majority of readers will not be able to under the SAO method. The paper should be self-contained.\n\nThe link between sparse training and sparse initialization is not clear. It is well understood that within a larger dense network, a small subnet is usually contributing the most. This contribution is different for different settings, e.g., one might want individual subnetworks to perform a task within a multi-tasking/meta-learning setting.\n\nI request authors to support the claim that post-pruning performance is better than sparse initialization-based training. IMHO, if the parameter mask is learned or adaptive during training, the performance is usually better. The depth and architecture of the network also play a significant role in deciding up to what levels a network can be pruned, which essentially connects to the idea of effective dimension/degree of freedom given the constraint local structure imposed by the architecture. \n\nOrthogonal initialization (sparse/non-sparse) just ensures effective signal propagation and not generalization. In the end, if the goal is effective training and achieving the best performance, how crucial is Exact Orthogonality? Unless we regularize the network, which might impact the stability, memory requirements and compute complexity.", "questions": "Literature:\nA few missing references\n- Sun et. al, Low-degree term first in ResNet, its variants and the whole neural network family\n- Thukur et. al, Incremental Trainable Parameter Selection in Deep Neural Networks\n- Larsson et. al, Fractalnet: Ultra-deep neural networks without residuals\n- Shulgin et al, Towards a better theoretical understanding of independent subnetwork training\n\nWhat are indices ijkk on page 4 in the top paragraph? A diagram of how H is embedded in a convolutional kernel of shape BxCinxCoutxHxW would help the readers.\n\nDelta orthogonalization assumes cyclic consistency for the convolutional layer. This assumption is not discussed in the paper and might mislead the readers.\n\nWith respect to sparse static training methods, please clarify whether the pruned connections/nodes that are not updated during backward pass are used for forward pass calculation or not. In addition, it seems Masks in GraSP/Synflow are adaptive over iterations, contrary to the setting introduced by the authors. \n\n\nThe construction of EOI for conv kernels using random entries in the mask to achieve the desired density is not explained in detail. Is it guaranteed to be exactly orthogonal in this case, too?\n\nMost results are empirical, and I wish authors have focused on theory to derive expressions for Dynamical Isometry in the proposed setting. Currently, the paper has half theory half numerical aspects, with both being incomplete.\n\nTable 1: which numbers belong to MNIST and CIFAR-10?\n\nIt will be good to show the exact parameter count of each model to get a sense of how much the 10% parameters? \nThen, a fair comparison would be models of the same size with different width and depth configurations.\n\nAlso, I would encourage authors to consider the full imagenet benchmark or consider a multi-task/meta-learning setting for establishing the practical benefits of the method.\nIn general, it looks like only the favourable experimental settings have been chosen to show the effectiveness of the approach. This is not a major issue if the main focus of paper was on theoretical aspects on the proposed initialization (which is not the case.)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors have proposed a method to achieve exact orthogonal initialization for training very deep neural models with sparsity constraints. The approach is built upon recent advancements achieved via tools from the Random Matrix theory to improve initialization and achieve training acceleration.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The proposed method is simple, mathematically grounded, and addresses an important issue of sparsity-aware training in DNNs.\nThe flow of the paper is nice and structured.", "weaknesses": "Overall, the paper is well-written, but often, critical details are missing, which might make it difficult for a new reader to understand and appreciate the ideas discussed and their connection to prior art. For instance, the majority of readers will not be able to under the SAO method. The paper should be self-contained.\n\nThe link between sparse training and sparse initialization is not clear. It is well understood that within a larger dense network, a small subnet is usually contributing the most. This contribution is different for different settings, e.g., one might want individual subnetworks to perform a task within a multi-tasking/meta-learning setting.\n\nI request authors to support the claim that post-pruning performance is better than sparse initialization-based training. IMHO, if the parameter mask is learned or adaptive during training, the performance is usually better. The depth and architecture of the network also play a significant role in deciding up to what levels a network can be pruned, which essentially connects to the idea of effective dimension/degree of freedom given the constraint local structure imposed by the architecture. \n\nOrthogonal initialization (sparse/non-sparse) just ensures effective signal propagation and not generalization. In the end, if the goal is effective training and achieving the best performance, how crucial is Exact Orthogonality? Unless we regularize the network, which might impact the stability, memory requirements and compute complexity.", "questions": "Literature:\nA few missing references\n- Sun et. al, Low-degree term first in ResNet, its variants and the whole neural network family\n- Thukur et. al, Incremental Trainable Parameter Selection in Deep Neural Networks\n- Larsson et. al, Fractalnet: Ultra-deep neural networks without residuals\n- Shulgin et al, Towards a better theoretical understanding of independent subnetwork training\n\nWhat are indices ijkk on page 4 in the top paragraph? A diagram of how H is embedded in a convolutional kernel of shape BxCinxCoutxHxW would help the readers.\n\nDelta orthogonalization assumes cyclic consistency for the convolutional layer. This assumption is not discussed in the paper and might mislead the readers.\n\nWith respect to sparse static training methods, please clarify whether the pruned connections/nodes that are not updated during backward pass are used for forward pass calculation or not. In addition, it seems Masks in GraSP/Synflow are adaptive over iterations, contrary to the setting introduced by the authors. \n\n\nThe construction of EOI for conv kernels using random entries in the mask to achieve the desired density is not explained in detail. Is it guaranteed to be exactly orthogonal in this case, too?\n\nMost results are empirical, and I wish authors have focused on theory to derive expressions for Dynamical Isometry in the proposed setting. Currently, the paper has half theory half numerical aspects, with both being incomplete.\n\nTable 1: which numbers belong to MNIST and CIFAR-10?\n\nIt will be good to show the exact parameter count of each model to get a sense of how much the 10% parameters? \nThen, a fair comparison would be models of the same size with different width and depth configurations.\n\nAlso, I would encourage authors to consider the full imagenet benchmark or consider a multi-task/meta-learning setting for establishing the practical benefits of the method.\nIn general, it looks like only the favourable experimental settings have been chosen to show the effectiveness of the approach. This is not a major issue if the main focus of paper was on theoretical aspects on the proposed initialization (which is not the case.)", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698470553707}, {"id": "9VPoLr9JC9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4859/Reviewer_4iab"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a new method to achieve exact (and not approximated) orthogonal sparse initialization for the weights of a (deep) neural net.  \nThe method relays on a straightforward idea of using givens rotations, which apply an orthogonal transform on two dimensions (essentially, a 2D rotation) out of the feature dimensions. This process is repeated on random pair of dimensions, with random rotation angle, until the desired sparsity (or, density) is achieved.\nThe authors provide an exact formula for the expected density after a given number of rotation, which allows a precise design of the resulting initialization.\n\nThe authors provide a thorough evaluation of the method for different activation functions, under different static sparse training methods, and show compelling results when compared to *approximate* initialization.\nAnother comparison in done over a 1000 layer MLP with no residual connections nor normalization layers.\nWhen trained on MNIST and CIFAR10 the proposed method achieves performance comparable to a dense network with only 12% of the weights.\nFinally, the authors perform a comparison of several modern architectures on the mini-imagenet. \nWith the sole exception of EfficientNet, the proposed method supersedes the existing sparse approaches, and narrows the gap to dense training with only 10% of the weights.", "review_text": "This paper introduces a new method to achieve exact (and not approximated) orthogonal sparse initialization for the weights of a (deep) neural net.  \nThe method relays on a straightforward idea of using givens rotations, which apply an orthogonal transform on two dimensions (essentially, a 2D rotation) out of the feature dimensions. This process is repeated on random pair of dimensions, with random rotation angle, until the desired sparsity (or, density) is achieved.\nThe authors provide an exact formula for the expected density after a given number of rotation, which allows a precise design of the resulting initialization.\n\nThe authors provide a thorough evaluation of the method for different activation functions, under different static sparse training methods, and show compelling results when compared to *approximate* initialization.\nAnother comparison in done over a 1000 layer MLP with no residual connections nor normalization layers.\nWhen trained on MNIST and CIFAR10 the proposed method achieves performance comparable to a dense network with only 12% of the weights.\nFinally, the authors perform a comparison of several modern architectures on the mini-imagenet. \nWith the sole exception of EfficientNet, the proposed method supersedes the existing sparse approaches, and narrows the gap to dense training with only 10% of the weights.", "strengths": "The paper was very easy to follow and understand, the main idea is straightforward but with a clear impact.\nExpected density formulation makes this method more appealing for practical usage.\nThe experimental section positions the method well w.r.t. existing methods.", "weaknesses": "While the impact is clear, it leave some questions about the price to pay for sparse networks.\nFor example, the reader might enjoy an analysis of the price (in performance) on the mini-imagenet for different sparsity levels.", "questions": "* Are some rotation angles better than others? I can't imagen that using only 1 degree Givens will perform similarly to using only 80 degree Givens\n* How can one gain sense of the price (in performance) for a given sparsity level?\n* One may assume that some sparsity patterns result in better performance - is that true? if so, can one guide the Givens dimensions to such a pattern", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new method to achieve exact (and not approximated) orthogonal sparse initialization for the weights of a (deep) neural net.  \nThe method relays on a straightforward idea of using givens rotations, which apply an orthogonal transform on two dimensions (essentially, a 2D rotation) out of the feature dimensions. This process is repeated on random pair of dimensions, with random rotation angle, until the desired sparsity (or, density) is achieved.\nThe authors provide an exact formula for the expected density after a given number of rotation, which allows a precise design of the resulting initialization.\n\nThe authors provide a thorough evaluation of the method for different activation functions, under different static sparse training methods, and show compelling results when compared to *approximate* initialization.\nAnother comparison in done over a 1000 layer MLP with no residual connections nor normalization layers.\nWhen trained on MNIST and CIFAR10 the proposed method achieves performance comparable to a dense network with only 12% of the weights.\nFinally, the authors perform a comparison of several modern architectures on the mini-imagenet. \nWith the sole exception of EfficientNet, the proposed method supersedes the existing sparse approaches, and narrows the gap to dense training with only 10% of the weights.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper was very easy to follow and understand, the main idea is straightforward but with a clear impact.\nExpected density formulation makes this method more appealing for practical usage.\nThe experimental section positions the method well w.r.t. existing methods.", "weaknesses": "While the impact is clear, it leave some questions about the price to pay for sparse networks.\nFor example, the reader might enjoy an analysis of the price (in performance) on the mini-imagenet for different sparsity levels.", "questions": "* Are some rotation angles better than others? I can't imagen that using only 1 degree Givens will perform similarly to using only 80 degree Givens\n* How can one gain sense of the price (in performance) for a given sparsity level?\n* One may assume that some sparsity patterns result in better performance - is that true? if so, can one guide the Givens dimensions to such a pattern", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698002916764}], "openreview_url": "https://openreview.net/forum?id=3mY9aGiMn0", "arxiv_id": "2406.01755", "paper_pdf": "papers/3mY9aGiMn0.pdf", "paper_pdf_sha256": "6e84ed26ce495981d6a6ea5389650450d9ff16e1b925c6cfdeffb0d9f8d67811", "paper_pdf_bytes": 1003240, "paper_pdf_source": "openreview", "code_url": "https://github.com/woocash2/sparser-better-deeper-stronger", "code_repository": "woocash2/sparser-better-deeper-stronger", "code_commit": "4438360111aa0f84aa75e9750e608a6127405a32", "code_archive": "repos/3mY9aGiMn0.zip", "code_archive_sha256": "ef776ac92f4b3f38feadb3c3f5f9b63424a49be716cf89d214246723bc237870", "code_archive_bytes": 94557, "code_file_count": 53, "code_extensions": {".py": 52, ".sh": 1}, "github_disk_usage_kb": 74, "github_languages": {"Python": 300807, "Shell": 50}, "github_archived": false, "github_pushed_at": "2024-03-12T18:22:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sparser-better-deeper-stronger-improving"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0bLE93R9d0O", "year": 2023, "status": "rejected", "title": "Transformer Module Networks for Systematic Generalization in Visual Question Answering", "authors": ["Moyuru Yamada", "Vanessa D'Amario", "Kentaro Takemoto", "Xavier Boix", "Tomotake Sasaki"], "authorids": ["~Moyuru_Yamada1", "~Vanessa_D'Amario1", "~Kentaro_Takemoto1", "~Xavier_Boix1", "~Tomotake_Sasaki1"], "authors_source": "OpenReview API", "abstract": "Transformers achieve great performance on Visual Question Answering (VQA). However, their systematic generalization capabilities, i.e., handling novel combinations of known concepts, is unclear. We reveal that Neural Module Networks (NMNs), i.e., question-specific compositions of modules that tackle a sub-task, achieve better or similar systematic generalization performance than the conventional Transformers, even though NMNs' modules are CNN-based. In order to address this shortcoming of Transformers with respect to NMNs, in this paper we investigate whether and how modularity can bring benefits to Transformers. Namely, we introduce Transformer Module Network (TMN), a novel NMN based on compositions of Transformer modules. TMNs achieve state-of-the-art systematic generalization performance in three VQA datasets, improving more than 30% over standard Transformers for novel compositions of sub-tasks. We show that not only the module composition but also the module specialization for each sub-task are the key of such performance gain.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "H05O1FWIWAv", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2120/Reviewer_W6Kf"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors reveal that Neural Module Networks (NMNs), i.e., question-specific compositions of modules that tackle a sub-task, achieve better or similar systematic generalization performance than the conventional Transformers, even though NMNs’ modules are CNN-based. \nTo address this shortcoming of Transformers with respect to NMNs, in this paper, the authors investigate whether and how modularity can bring benefits to Transformers.\nNamely, they introduce a Transformer Module Network (TMN), a novel NMN based on compositions of Transformer modules.\n", "review_text": "The authors provide interesting observations, and the experiments are solid enough to show the effectiveness of the proposed method.", "strengths": "Strength\n- The observation about the comparison between the neural module networks and the Transformer models is novel and interesting.\n- The authors provide a comprehensive ablation study to make the final model more solid.\n\n\nWeakness\n- It would be better to add qualitative results to see the behavior of the proposed method and the existing methods. Also, it would be better to discuss why the conventional Transformer models perform worse than the neural module networks by analyzing the samples.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors reveal that Neural Module Networks (NMNs), i.e., question-specific compositions of modules that tackle a sub-task, achieve better or similar systematic generalization performance than the conventional Transformers, even though NMNs’ modules are CNN-based. \nTo address this shortcoming of Transformers with respect to NMNs, in this paper, the authors investigate whether and how modularity can bring benefits to Transformers.\nNamely, they introduce a Transformer Module Network (TMN), a novel NMN based on compositions of Transformer modules.\n", "strength_and_weaknesses": "Strength\n- The observation about the comparison between the neural module networks and the Transformer models is novel and interesting.\n- The authors provide a comprehensive ablation study to make the final model more solid.\n\n\nWeakness\n- It would be better to add qualitative results to see the behavior of the proposed method and the existing methods. Also, it would be better to discuss why the conventional Transformer models perform worse than the neural module networks by analyzing the samples.", "clarity,_quality,_novelty_and_reproducibility": "The proposed method is clear and novel. The quality of this paper looks favorable overall.", "summary_of_the_review": "The authors provide interesting observations, and the experiments are solid enough to show the effectiveness of the proposed method.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666682957925}, {"id": "H0PM1r2Jiy4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2120/Reviewer_o4YA"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper mainly focuses on combining the strengths of Transformer and NMN by introducing a novel NMN based on compositions of Transformer modules named Transformer Module Network (TMN). The model is evaluated on CLEVR-CoGenT, CLOSURE, and GQA-SGL.", "review_text": "The motivation that combining the strengths of the two structures is good. However, directly substituting the modular with TRM seems not novel to me. Besides, there are no experiments to prove that the TRM encoder outperforms all previous NMN modular in the task, thus, it does not convince me of its superiority by directly substituting all NMN modulars. In addition, it lacks comparisons of the SOA methods. And from my knowledge, the performance is not SOA (the last point in the weakness section). Overall, I'm leaning to reject it.", "strengths": "*[Strength]*\n\nThe motivation that combining the strength of NMN and Transformer is contributing.\n\n*[Weakness]*\n1. In the approach, the paper directly substitutes the module in the NMN with the TRM encoder. However, it lacks ablation experiments and discussion about if the TRM structure outperforms all types of module in NMN-based method. (e.g., in n2mn[1] Table 1, the attention based modules like find, relocate, filter modules and the answer based modules like describe, compare and exist modules)\n2. The Tree modular modular is not novel in NMN domain, [1] uses similar manner in Figure 3.\n3. Lacks of SOA TRM-based and NMN-based model performance comparisons. In addition, the SOA method on GQA dataset is 72.1 from [2], and there is [3] with 69.46 on GQA-dev\n\n\n\n\n[1] Hu, Ronghang, et al. \"Learning to reason: End-to-end module networks for visual question answering.\" Pr oceedings of the IEEE international conference on computer vision. 2017.\n\n[2] Nguyen, Binh X., et al. \"Coarse-to-Fine Reasoning for Visual Question Answering.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[3] Kim, Eun-Sol, et al. \"Hypergraph attention networks for multimodal learning.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper mainly focuses on combining the strengths of Transformer and NMN by introducing a novel NMN based on compositions of Transformer modules named Transformer Module Network (TMN). The model is evaluated on CLEVR-CoGenT, CLOSURE, and GQA-SGL.", "strength_and_weaknesses": "*[Strength]*\n\nThe motivation that combining the strength of NMN and Transformer is contributing.\n\n*[Weakness]*\n1. In the approach, the paper directly substitutes the module in the NMN with the TRM encoder. However, it lacks ablation experiments and discussion about if the TRM structure outperforms all types of module in NMN-based method. (e.g., in n2mn[1] Table 1, the attention based modules like find, relocate, filter modules and the answer based modules like describe, compare and exist modules)\n2. The Tree modular modular is not novel in NMN domain, [1] uses similar manner in Figure 3.\n3. Lacks of SOA TRM-based and NMN-based model performance comparisons. In addition, the SOA method on GQA dataset is 72.1 from [2], and there is [3] with 69.46 on GQA-dev\n\n\n\n\n[1] Hu, Ronghang, et al. \"Learning to reason: End-to-end module networks for visual question answering.\" Pr oceedings of the IEEE international conference on computer vision. 2017.\n\n[2] Nguyen, Binh X., et al. \"Coarse-to-Fine Reasoning for Visual Question Answering.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[3] Kim, Eun-Sol, et al. \"Hypergraph attention networks for multimodal learning.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.", "clarity,_quality,_novelty_and_reproducibility": "The paper clarified its contributions and approach clearly. \nIn general, the paper can be reproduced and is of fair quality.\nHowever, it lacks novelty since the tree structure and substituting the modular with the TRM encoder are not novel.", "summary_of_the_review": "The motivation that combining the strengths of the two structures is good. However, directly substituting the modular with TRM seems not novel to me. Besides, there are no experiments to prove that the TRM encoder outperforms all previous NMN modular in the task, thus, it does not convince me of its superiority by directly substituting all NMN modulars. In addition, it lacks comparisons of the SOA methods. And from my knowledge, the performance is not SOA (the last point in the weakness section). Overall, I'm leaning to reject it.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666632964492}, {"id": "lZlpQFTggog", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2120/Reviewer_pauq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper focuses on the systematic generalization visual question answering where the test set contains novel combinations of training concepts. The authors propose a new modular network whose modules are transformers.", "review_text": "The paper proposes a new dataset. But the proposed model is not novel and its effectiveness is not fully verified. Thus I give borderline.", "strengths": "Strength:\n\nThe paper focuses on the model robustness against VQA bias, which is a common problem in the VQA field.\n\nThe paper proposed a new systematic generalization dataset GQA-SGL based on the GQA. The new dataset can be a new challenge for future work in this field.\n\nWeakness:\n\nThe novelty is limited. The proposed model is a modular network composed of more expressive transformer modules. \n\nThe proposed model also suffers from the weakness of the previous modular network. It needs program supervision and relies on the performance of the question parser. Since all the experiments have a groundtruth program at test time, whether the question parser has the same systematic generalization ability is not verified.\n\nThe improvement on the CLEVR-CoGenT and CLOSURE dataset is not consistent. The proposed model performs worse than the baseline on the CLEVR-CoGenT but improves on the CLOSURE. I suppose the CLEVR-CoGenT and CLOSURE have novel images and questions respectively. However, the TMN can ignore the questions by accessing the groundtruth programs.\nThe authors suggest that the proposed model and baseline share the same visual feature extractor. Thus authors may verify this by training a new extractor from scratch similar to FiLM[1].\n\n[1] Perez E., Strub F., de Vries H., Dumoulin V., Courville, A. FiLM: Visual Reasoning with a General Conditioning Layer. AAAI 2017", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper focuses on the systematic generalization visual question answering where the test set contains novel combinations of training concepts. The authors propose a new modular network whose modules are transformers.", "strength_and_weaknesses": "Strength:\n\nThe paper focuses on the model robustness against VQA bias, which is a common problem in the VQA field.\n\nThe paper proposed a new systematic generalization dataset GQA-SGL based on the GQA. The new dataset can be a new challenge for future work in this field.\n\nWeakness:\n\nThe novelty is limited. The proposed model is a modular network composed of more expressive transformer modules. \n\nThe proposed model also suffers from the weakness of the previous modular network. It needs program supervision and relies on the performance of the question parser. Since all the experiments have a groundtruth program at test time, whether the question parser has the same systematic generalization ability is not verified.\n\nThe improvement on the CLEVR-CoGenT and CLOSURE dataset is not consistent. The proposed model performs worse than the baseline on the CLEVR-CoGenT but improves on the CLOSURE. I suppose the CLEVR-CoGenT and CLOSURE have novel images and questions respectively. However, the TMN can ignore the questions by accessing the groundtruth programs.\nThe authors suggest that the proposed model and baseline share the same visual feature extractor. Thus authors may verify this by training a new extractor from scratch similar to FiLM[1].\n\n[1] Perez E., Strub F., de Vries H., Dumoulin V., Courville, A. FiLM: Visual Reasoning with a General Conditioning Layer. AAAI 2017", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-organized and easy to understand.\n\nThe novelty is not enough. The proposed model only replaces the previous neural module with a transformer.", "summary_of_the_review": "The paper proposes a new dataset. But the proposed model is not novel and its effectiveness is not fully verified. Thus I give borderline.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666614902179}, {"id": "7_RcyBbQg8C", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2120/Reviewer_zXs6"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper addresses the problem of systematic generalization in VQA. It proposes a new model that takes advantage of the learning capability of Transformers and the compositional modeling of Module Networks. The proposed model achieves state-of-the-art performance on three different VQA datasets including standard testbeds for VQA systematic generalization. ", "review_text": "The paper is well written and easy to follow. However, it is limited in contribution to advance the systematic generalization task.", "strengths": "•\tStrengths:\n\n1.\tThe proposed method is technically sound and easy to understand. It is reasonable to have a model that makes use of the known advantages of Transformers and module networks.\n\n2.\tIt shows better performance when compared against existing methods.\n\n•\tWeaknesses:\n\n1.\tWhile the proposed model is straightforward and easy to understand, it appears that the model brings very little insight into the systematic generalization task itself. It is well studied that module networks are a potential approach for systematic generalization but it is extremely brittle and ad-hoc which hinders its application in wider settings. Meanwhile, models with better representation learning such as Transformers are beneficial for all machine learning tasks, not just for systematic generalization. The proposed method does not offer any fundamental contributions toward solving systematic generalization.\n\n2.\tExperiments with original transformers do not sound reasonable as the original design of transformer is for translation tasks with having access to a more generous amount of training data. There is no reason to have to use 12 layers for transformers when applying it to a custom problem. The poor performance of transformers in the experiments might simply be due to overfitting.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper addresses the problem of systematic generalization in VQA. It proposes a new model that takes advantage of the learning capability of Transformers and the compositional modeling of Module Networks. The proposed model achieves state-of-the-art performance on three different VQA datasets including standard testbeds for VQA systematic generalization. ", "strength_and_weaknesses": "•\tStrengths:\n\n1.\tThe proposed method is technically sound and easy to understand. It is reasonable to have a model that makes use of the known advantages of Transformers and module networks.\n\n2.\tIt shows better performance when compared against existing methods.\n\n•\tWeaknesses:\n\n1.\tWhile the proposed model is straightforward and easy to understand, it appears that the model brings very little insight into the systematic generalization task itself. It is well studied that module networks are a potential approach for systematic generalization but it is extremely brittle and ad-hoc which hinders its application in wider settings. Meanwhile, models with better representation learning such as Transformers are beneficial for all machine learning tasks, not just for systematic generalization. The proposed method does not offer any fundamental contributions toward solving systematic generalization.\n\n2.\tExperiments with original transformers do not sound reasonable as the original design of transformer is for translation tasks with having access to a more generous amount of training data. There is no reason to have to use 12 layers for transformers when applying it to a custom problem. The poor performance of transformers in the experiments might simply be due to overfitting.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is technically sound, with sufficient clarity. The details provided are fully explained and can be reproducible. However, it appears to be limited in scientific contribution.", "summary_of_the_review": "The paper is well written and easy to follow. However, it is limited in contribution to advance the systematic generalization task.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666587700307}], "openreview_url": "https://openreview.net/forum?id=0bLE93R9d0O", "arxiv_id": "2201.11316", "paper_pdf": "papers/0bLE93R9d0O.pdf", "paper_pdf_sha256": "f10035ba5ca0b24ae3d2ba6bb76e14cccbb9bb8264f74dd9c5c0c55ebb74cb4a", "paper_pdf_bytes": 771561, "paper_pdf_source": "openreview", "code_url": "https://github.com/FujitsuResearch/transformer_module_networks", "code_repository": "FujitsuResearch/transformer_module_networks", "code_commit": "8e91aa7d9be865ba89f850a2beda7b7d622d07d7", "code_archive": "repos/0bLE93R9d0O.zip", "code_archive_sha256": "757c83b1fe9589f2378631e40db83119c46120a978adecf4f7cf526d1e0dc305", "code_archive_bytes": 255519, "code_file_count": 61, "code_extensions": {".py": 61}, "github_disk_usage_kb": 147, "github_languages": {"Python": 486962}, "github_archived": false, "github_pushed_at": "2023-02-13T07:07:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transformer-module-networks-for-systematic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "KDAEc2nai83", "year": 2022, "status": "rejected", "title": "Human-Level Control without Server-Grade Hardware", "authors": ["Brett Daley", "Christopher Amato"], "authorids": ["~Brett_Daley1", "~Christopher_Amato1"], "authors_source": "OpenReview API", "abstract": "Deep Q-Network (DQN) marked a major milestone for reinforcement learning, demonstrating for the first time that human-level control policies could be learned directly from raw visual inputs via reward maximization. Even years after its introduction, DQN remains highly relevant to the research community since many of its innovations have been adopted by successor methods. Nevertheless, despite significant hardware advances in the interim, DQN's original Atari 2600 experiments remain extremely costly to replicate in full. This poses an immense barrier to researchers who cannot afford state-of-the-art hardware or lack access to large-scale cloud computing resources. To facilitate improved access to deep reinforcement learning research, we introduce a DQN implementation that leverages a novel concurrent and synchronized execution framework designed to maximally utilize a heterogeneous CPU-GPU desktop system. With just one NVIDIA GeForce GTX 1080 GPU, our implementation reduces the training time of a 200-million-frame Atari experiment from 25 hours to just 9 hours. The ideas introduced in our paper should be generalizable to a large number of off-policy deep reinforcement learning methods.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "ebtKF7B2elf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1918/Reviewer_4Ciy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a reimplementation of DQN that improves training speed with lower hardware requirements, showing a decrease of running a single Atari experiment from 25 hours to 9 hours on a specific system.", "review_text": "This paper is written clearly, and presents the main contributions in a straightforward manner. The inclusion of the code for reproducibility is appreciated.\n\nHowever, there are a number of concerns with this paper, which are listed below.\n- The motivation for doing this study is not very convincing; why would researchers need to retrain using DQN (or the variant of DQN introduced here) in Atari? Wouldn't just making the trained models available serve the same task?\n- More generally, it is not convincing that a more efficient re-implementation of DQN (with a change to improve parallelism) is a notable contribution. If this concept could somehow be generalized to other domains, then this would potentially be useful. As it is, there are other algorithms that can achieve much better results than DQN, so it is less meaningful to implement this change. Also, how general are these results, and do they hold on different hardware?\n- What does \"real experience\" refer to in Section 1, paragraph 2? Does this refer to a human playing?\n- Section 1, paragraph 2 refers to GPUs as \"costly\", but this method still uses a GPU.\n- The claimed speedup in terms of wall-clock time is 2.78x. It is not clear how significant this is, especially compared to other DQN implementations.\n- Section 2: since this paper is about a more resource-efficient implementation of DQN, it is reasonable to assume that readers are familiar with DQN, so this discussion could be much more compact.\n- Section 5.1 refers to \"should be affordable\" - it would be useful to have some kind of relative pricing - for instance, what would be the relative cost required to get an analogous speedup just by improving hardware?\n- Section 5.2 refers to \"all 49 Atari games\"; there are actually many more than this, although 57 are commonly studied. There were indeed 49 games in the original DQN paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a reimplementation of DQN that improves training speed with lower hardware requirements, showing a decrease of running a single Atari experiment from 25 hours to 9 hours on a specific system.", "main_review": "This paper is written clearly, and presents the main contributions in a straightforward manner. The inclusion of the code for reproducibility is appreciated.\n\nHowever, there are a number of concerns with this paper, which are listed below.\n- The motivation for doing this study is not very convincing; why would researchers need to retrain using DQN (or the variant of DQN introduced here) in Atari? Wouldn't just making the trained models available serve the same task?\n- More generally, it is not convincing that a more efficient re-implementation of DQN (with a change to improve parallelism) is a notable contribution. If this concept could somehow be generalized to other domains, then this would potentially be useful. As it is, there are other algorithms that can achieve much better results than DQN, so it is less meaningful to implement this change. Also, how general are these results, and do they hold on different hardware?\n- What does \"real experience\" refer to in Section 1, paragraph 2? Does this refer to a human playing?\n- Section 1, paragraph 2 refers to GPUs as \"costly\", but this method still uses a GPU.\n- The claimed speedup in terms of wall-clock time is 2.78x. It is not clear how significant this is, especially compared to other DQN implementations.\n- Section 2: since this paper is about a more resource-efficient implementation of DQN, it is reasonable to assume that readers are familiar with DQN, so this discussion could be much more compact.\n- Section 5.1 refers to \"should be affordable\" - it would be useful to have some kind of relative pricing - for instance, what would be the relative cost required to get an analogous speedup just by improving hardware?\n- Section 5.2 refers to \"all 49 Atari games\"; there are actually many more than this, although 57 are commonly studied. There were indeed 49 games in the original DQN paper.\n", "summary_of_the_review": "The main contribution of this paper is not sufficiently well-motivated or general; the re-implementation of DQN with more parallelism does not pose enough independent interest to be practically useful for most researchers. Suggestions for improvement would be to (1) show that this method could actually extend to settings beyond DQN and improve other RL algorithms; (2) perform a more in-depth hardware study comparing the cost to speedup via more powerful hardware vs. software improvements.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635895585306}, {"id": "twuLA9u0u2B", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1918/Reviewer_KSo9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose an implementation of DQN that focuses on maximising the data / training throughput in a common CPU-GPU machine setting. They do so by (asynchronously) running inference on multiple environments at once (and synchronously blocking on the env batch) and using the DQN target network parameters for inference rather than the default -- latest -- parameters of the network. They show considerable improvements in speed while producing comparable results with previous DQN results on the Atari suite.\n", "review_text": "irst and foremost, let me start by saying that I really appreciate the direction that the paper is taking and the problem that it is trying to solve. RL is often not optimized for academia-like hardware requirements, and we need more research and engineering work to figure out how to utilize these kinds of resources better. Kudos to the authors for working on this.\n\nNow, onto the some of the issues that prevent me from recommending the paper for acceptance:\n\n## Novelty\n\nI am a little confused on how we should place this work in the literature. The novelty of the approach seems the clever combination of two frequently used ideas: (1) \"vectorizing\" environment executions so that the env interface can take $A = {a_0, a_1, ..., a_n}$ actions for $E = {e_0, e_1, ..., e_n}$ environments; (2) employing the target network to run inference on CPU whilst training a new set of parameters on default parameters on GPU from experience replay.\n\nPoint (1) however is an implementation details that is pretty much used across the large majority of modern DRL codebases out in the wild (e.g. ACME, stable-baselines, TorchBeast, ...), as in the majority of the cases environments are comparably much cheaper than the rest of the loop to run in parallel. This doesn't seem to be very well reflected in the manuscript (upon first read I was under the impression that this was something else, because I assumed that it was going to be something *beyond* VecEnvs).\n\nPoint (2) technically could be consider novel (AFAIK) when employed to produce this accelerator utilization scheme, but it is also a concept that exists in many existing libraries as an implementation detail. As the current manuscript is lacking in a strong discussion regarding (for instance) how using the target network affects DQN training (say, wrt. different kinds of MDPS -- lava world vs a pendulum optimizer -- or how the target network update frequency might affect the effectiveness of this trick), it feels odd to claim this as a strong contribution of the manuscript.\n\n\n## Clarity of methods\n\nI think the paper could use with a solid paragraph of definitions regarding what the authors mean when they say \"distributed\". One can technically run a distributed agent with extremely good results in local CPU-GPU hardware (see e.g. TorchBeast); it is true that the underlying complexity is greater, but many off-the-shelf frameworks abstract a lot of this away (and frankly speaking completely nail down the RPC/orchestration aspect of this problem -- see e.g. ACME recently). Since the manuscript is lacking a discussion on the tradeoffs that one makes when going that direction -- rather than sticking to a simpler threading / data communication model -- the comparisons with algorithms such as Impala (or R2D2) feel generally handwave-y.\n\nI also found the explanation of the synchronization model to be a little unclear and underwhelming. Considering that these modifications are supposedly simple, I think a little bit of pseudo-python could go a long way towards making the whole methodology section clearer (particularly if other systems are also mentioned in the paper).\n\n\n## Experiments\n\nThe experiments were also a little underwhelming. Firstly, for the speed test to be truly representative of reasoanble in-the-wild workloads, it probably should include comparisons against environments that have different step frequencies, models of wildly different sizes, the addition of recurrence (which is very common now in most RL-trained models), etc.\n\nSecondly, I could not understand the point of the extrapolation done in the Pong experiments. Pong is very cheap to run, especially compared to running the whole Atari suite, so it feels bizarre to stop training early and then *linearly* (!) extrapolate the results. I would really like to understand why this procedure was chosen.\n\nThirdly, I think these sorts of empirical validations (for this particular paper) cannot rely on baselines numbers provided by papers, but rather be based on existing popular (and already well optimized) implementations of DQN. DQN is a relatively old method, and a lot of different public codebases have optimized it greatly with various amounts of tricks. It still remains unclear whether the presented tricks would actually make a significant increase in training speed when used on many of these implementations (I think at least one comparison should be attempted).\n\n-- \n\nPlease do note that I'd be very willing to increase my score provided that we (including the rest of the reviewers) have a productive discussions on how the manuscript can be quickly improved.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose an implementation of DQN that focuses on maximising the data / training throughput in a common CPU-GPU machine setting. They do so by (asynchronously) running inference on multiple environments at once (and synchronously blocking on the env batch) and using the DQN target network parameters for inference rather than the default -- latest -- parameters of the network. They show considerable improvements in speed while producing comparable results with previous DQN results on the Atari suite.\n", "main_review": "irst and foremost, let me start by saying that I really appreciate the direction that the paper is taking and the problem that it is trying to solve. RL is often not optimized for academia-like hardware requirements, and we need more research and engineering work to figure out how to utilize these kinds of resources better. Kudos to the authors for working on this.\n\nNow, onto the some of the issues that prevent me from recommending the paper for acceptance:\n\n## Novelty\n\nI am a little confused on how we should place this work in the literature. The novelty of the approach seems the clever combination of two frequently used ideas: (1) \"vectorizing\" environment executions so that the env interface can take $A = {a_0, a_1, ..., a_n}$ actions for $E = {e_0, e_1, ..., e_n}$ environments; (2) employing the target network to run inference on CPU whilst training a new set of parameters on default parameters on GPU from experience replay.\n\nPoint (1) however is an implementation details that is pretty much used across the large majority of modern DRL codebases out in the wild (e.g. ACME, stable-baselines, TorchBeast, ...), as in the majority of the cases environments are comparably much cheaper than the rest of the loop to run in parallel. This doesn't seem to be very well reflected in the manuscript (upon first read I was under the impression that this was something else, because I assumed that it was going to be something *beyond* VecEnvs).\n\nPoint (2) technically could be consider novel (AFAIK) when employed to produce this accelerator utilization scheme, but it is also a concept that exists in many existing libraries as an implementation detail. As the current manuscript is lacking in a strong discussion regarding (for instance) how using the target network affects DQN training (say, wrt. different kinds of MDPS -- lava world vs a pendulum optimizer -- or how the target network update frequency might affect the effectiveness of this trick), it feels odd to claim this as a strong contribution of the manuscript.\n\n\n## Clarity of methods\n\nI think the paper could use with a solid paragraph of definitions regarding what the authors mean when they say \"distributed\". One can technically run a distributed agent with extremely good results in local CPU-GPU hardware (see e.g. TorchBeast); it is true that the underlying complexity is greater, but many off-the-shelf frameworks abstract a lot of this away (and frankly speaking completely nail down the RPC/orchestration aspect of this problem -- see e.g. ACME recently). Since the manuscript is lacking a discussion on the tradeoffs that one makes when going that direction -- rather than sticking to a simpler threading / data communication model -- the comparisons with algorithms such as Impala (or R2D2) feel generally handwave-y.\n\nI also found the explanation of the synchronization model to be a little unclear and underwhelming. Considering that these modifications are supposedly simple, I think a little bit of pseudo-python could go a long way towards making the whole methodology section clearer (particularly if other systems are also mentioned in the paper).\n\n\n## Experiments\n\nThe experiments were also a little underwhelming. Firstly, for the speed test to be truly representative of reasoanble in-the-wild workloads, it probably should include comparisons against environments that have different step frequencies, models of wildly different sizes, the addition of recurrence (which is very common now in most RL-trained models), etc.\n\nSecondly, I could not understand the point of the extrapolation done in the Pong experiments. Pong is very cheap to run, especially compared to running the whole Atari suite, so it feels bizarre to stop training early and then *linearly* (!) extrapolate the results. I would really like to understand why this procedure was chosen.\n\nThirdly, I think these sorts of empirical validations (for this particular paper) cannot rely on baselines numbers provided by papers, but rather be based on existing popular (and already well optimized) implementations of DQN. DQN is a relatively old method, and a lot of different public codebases have optimized it greatly with various amounts of tricks. It still remains unclear whether the presented tricks would actually make a significant increase in training speed when used on many of these implementations (I think at least one comparison should be attempted).\n\n-- \n\nPlease do note that I'd be very willing to increase my score provided that we (including the rest of the reviewers) have a productive discussions on how the manuscript can be quickly improved.", "summary_of_the_review": "The direction of the manuscript is commendable, and the ideas seem simple and cleverly executed. However there remain doubts wrt. novelty and on the validity of the experimental settings. Furthermore, the manuscript could also be significantly improved in terms of clarity.\n\n---\n\nEDIT: bumped score to 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635895354130}, {"id": "JibvM5-glwh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1918/Reviewer_pigN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "A new method for benchmarking RL algorithms like DQN, focusing on Atari suite and experiments from seminal DQN paper. The focus is to achieve speedup, allowing researchers with a modest setup (pc with gpu) to run the tests in reasonable time. The method combines concurrent data collection in multiple threads (and multiple environment instances) and training and achieves a speedup of x2.7 compared to vanilla DQN. At the same time the method slightly changes the nature of the algorithm by different action sampling (multiple env instances which is reminiscent of Asynchronous Methods paper from 2016) and slightly different action selection (using target network). The results are comparable to ones in the original paper overall, but vary on per game basis probably due to mentioned changes, slight change in hyper parameters and high variance of a single run. ", "review_text": "Pros:\n+ solves an actual technical/engineering problem that I encountered\n+ reasonably clearly written, easy to read\n+ well executed\n\nCons:\n- changes the algorithms to be benchmarked slightly so it is not really fully equivalent to normal testing (using target network and multithreaded sampler)\n- the speed up is not that impressive, if 25 hours is prohibitive I imagine that 9 might also be\n- the ideas implemented aren't really that innovative or creative, I don't want to be arrogant but these are first things to try and I think I saw similar approaches but in context of synchronous A3C on gpu\n- (not necessarily) the usefulness to the community hinges on the implementation of the framework and no details are mentioned\n\n\nComments/questions:\n- The part where you run first 1mil instead 50mil steps is quite ironic, wasn't it easier to let it run for a bit longer and run it as is?\n-\"The dillema is that exectuion and training are sequentially dependent on each other\" - are they really? Samples are put into replay memory and are randomly sampled every step. Considering that F is usually considerably bigger than the memory size it should not matter if we experience/train in sequence or in parallel. There is much bigger time delay between target and and original network and you explicitly ignore it.\n- Figure 1: \"Synchronize\" step could be more explicit I guess.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "A new method for benchmarking RL algorithms like DQN, focusing on Atari suite and experiments from seminal DQN paper. The focus is to achieve speedup, allowing researchers with a modest setup (pc with gpu) to run the tests in reasonable time. The method combines concurrent data collection in multiple threads (and multiple environment instances) and training and achieves a speedup of x2.7 compared to vanilla DQN. At the same time the method slightly changes the nature of the algorithm by different action sampling (multiple env instances which is reminiscent of Asynchronous Methods paper from 2016) and slightly different action selection (using target network). The results are comparable to ones in the original paper overall, but vary on per game basis probably due to mentioned changes, slight change in hyper parameters and high variance of a single run. ", "main_review": "Pros:\n+ solves an actual technical/engineering problem that I encountered\n+ reasonably clearly written, easy to read\n+ well executed\n\nCons:\n- changes the algorithms to be benchmarked slightly so it is not really fully equivalent to normal testing (using target network and multithreaded sampler)\n- the speed up is not that impressive, if 25 hours is prohibitive I imagine that 9 might also be\n- the ideas implemented aren't really that innovative or creative, I don't want to be arrogant but these are first things to try and I think I saw similar approaches but in context of synchronous A3C on gpu\n- (not necessarily) the usefulness to the community hinges on the implementation of the framework and no details are mentioned\n\n\nComments/questions:\n- The part where you run first 1mil instead 50mil steps is quite ironic, wasn't it easier to let it run for a bit longer and run it as is?\n-\"The dillema is that exectuion and training are sequentially dependent on each other\" - are they really? Samples are put into replay memory and are randomly sampled every step. Considering that F is usually considerably bigger than the memory size it should not matter if we experience/train in sequence or in parallel. There is much bigger time delay between target and and original network and you explicitly ignore it.\n- Figure 1: \"Synchronize\" step could be more explicit I guess.", "summary_of_the_review": "The method proposed is not super innovative or creative but it's something that is needed and is well executed. It seems to do what it strived to do (speed-up) and there are no issues with the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635867027185}, {"id": "j6ypjBkLkG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1918/Reviewer_287J"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper introduces a DQN implementation to leverage a novel concurrent and synchronized execution framework for better utilizing a heterogeneous CPU-GPU desktop system. With one NVIDIA GeForce GTX 1080 GPU, the implementation reduces the training time of a 200-million-frame Atari experiment from 25 hours to just 9 hours. ", "review_text": "strength: Nearly three times acceleration of the proposed framework is very attracting that facilitates the complex Atrai games playing with DQN on desktop-level computing resources. \n\nweakness: It is more important to see how the proposed framework can shed light on the implementations of other DRL or even DL models. This important discussion has been largely ignored  in the manuscript. The generalization of the proposed method is highly expected.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a DQN implementation to leverage a novel concurrent and synchronized execution framework for better utilizing a heterogeneous CPU-GPU desktop system. With one NVIDIA GeForce GTX 1080 GPU, the implementation reduces the training time of a 200-million-frame Atari experiment from 25 hours to just 9 hours. ", "main_review": "strength: Nearly three times acceleration of the proposed framework is very attracting that facilitates the complex Atrai games playing with DQN on desktop-level computing resources. \n\nweakness: It is more important to see how the proposed framework can shed light on the implementations of other DRL or even DL models. This important discussion has been largely ignored  in the manuscript. The generalization of the proposed method is highly expected.", "summary_of_the_review": "The paper proposes an acceleration framework for DRL based on synchronized execution of GPU. The performance of the framework is attracting. However, instead of the sole showcase, it is more important to see how the proposed framework can generalize to more DRL tasks.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635133689662}, {"id": "_OCsP1wfId_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1918/Reviewer_Dq75"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "DQN is known to consume a lot of resources. This paper introduces an optimized version of DQN which speeds up training 25->9 hours. The authors use two main techniques to improve throughput. Firstly, the authors propose to select actions by computing the argmax over the target network. This decouples acting and training and allows these two operations to run asynchronously. Secondly, the authors compute actions for multiple environments in parallel on the GPU. ", "review_text": "# Main Review\n\nThe paper is well-written and well-motivated. The implementation appears to be much faster and uses reasonable methods to achieve this speedup. Having a performant DQN implementation that runs quickly and is publicly accessible would benefit the RL community. I think there are two main issues with the paper: the novelty and the empirical evaluation.\n\n## Novelty\n\nThe first idea that the authors introduce, selecting actions from the target network to enable parallel acting and training, has not been used to speed up training to my knowledge. However, it is similar to the idea of Double Q-learning[1]. Additionally, multiple other researchers have introduced systems that allow for concurrent training and acting [2]. Thus, I would say that the novelty of this idea is limited.\n\nThe second idea, taking multiple actions in parallel, is a standard method as far as I know. I believe it e.g. is used in the stable baselines implementation. See e.g. (https://stable-baselines.readthedocs.io/en/master/guide/examples.html?highlight=vector#recurrent-policies) and (https://github.com/hill-a/stable-baselines/blob/master/stable_baselines/common/vec_env/subproc_vec_env.py#L51). To the best of my knowledge, the proposed method is not novel -- but please clarify how your proposals differ if there's a misunderstanding on my part.\n\n[1] Deep Reinforcement Learning with Double Q-learning, 2016\n\n[2] Sample Factory: Egocentric 3D Control from Pixels at 100000 FPS with Asynchronous Reinforcement Learning, 2020\n\n## Empirical Evaluation\n\nThe authors seem to measure speedup of 25->9 hours on their own baseline DQN implementation. It does not seem reasonable to use a baseline you implement yourself when there are plenty of performant DQN implementations available online. Minor implementation details can often have outsized effect to outcome in RL, and the authors might not be incentivized to optimize the baseline as it's easier to improve upon a poor baseline. I would encourage the authors to start from a performant publicly available DQN implementation, e.g. the dopamine repo.\n\nThe authors write that “Our baseline appears to be significantly faster than popular existing DQN implementations (e.g. Roderick et al., 2017; Dhariwal et al., 2017; Hill et al., 2018) based on their reported results, although we did not compare them on our hardware.”. The difference might be entirely due to hardware differences, so this statement does not say much. Additionally, the cited papers are relatively old, and more modern implementations and improved hardware might be available today. \n\nTable 4 contains no standard deviations. Are results just computed over 1 seed per environment? If so, it might not be statistically meaningful.\n\nThe code does not seem to have all dependencies listed (e.g. cv2 is missing), and there's also no example on how to use it. For a paper whose main contribution is on sharing a good implementation, this is not ideal.\n\n## Minor comments\n\nI would recommend that the authors show the learning curves (of e.g. averaged human-normalized scores) in the main paper.\n\nI would encourage the authors to broadener their related work section to other methods that have been proposed for accelerating RL:\n\n[3] Accelerating reinforcement learning through GPU atari emulation, 2019\n\n[4] Large Batch Simulation for Deep Reinforcement Learning, 2021\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "DQN is known to consume a lot of resources. This paper introduces an optimized version of DQN which speeds up training 25->9 hours. The authors use two main techniques to improve throughput. Firstly, the authors propose to select actions by computing the argmax over the target network. This decouples acting and training and allows these two operations to run asynchronously. Secondly, the authors compute actions for multiple environments in parallel on the GPU. ", "main_review": "# Main Review\n\nThe paper is well-written and well-motivated. The implementation appears to be much faster and uses reasonable methods to achieve this speedup. Having a performant DQN implementation that runs quickly and is publicly accessible would benefit the RL community. I think there are two main issues with the paper: the novelty and the empirical evaluation.\n\n## Novelty\n\nThe first idea that the authors introduce, selecting actions from the target network to enable parallel acting and training, has not been used to speed up training to my knowledge. However, it is similar to the idea of Double Q-learning[1]. Additionally, multiple other researchers have introduced systems that allow for concurrent training and acting [2]. Thus, I would say that the novelty of this idea is limited.\n\nThe second idea, taking multiple actions in parallel, is a standard method as far as I know. I believe it e.g. is used in the stable baselines implementation. See e.g. (https://stable-baselines.readthedocs.io/en/master/guide/examples.html?highlight=vector#recurrent-policies) and (https://github.com/hill-a/stable-baselines/blob/master/stable_baselines/common/vec_env/subproc_vec_env.py#L51). To the best of my knowledge, the proposed method is not novel -- but please clarify how your proposals differ if there's a misunderstanding on my part.\n\n[1] Deep Reinforcement Learning with Double Q-learning, 2016\n\n[2] Sample Factory: Egocentric 3D Control from Pixels at 100000 FPS with Asynchronous Reinforcement Learning, 2020\n\n## Empirical Evaluation\n\nThe authors seem to measure speedup of 25->9 hours on their own baseline DQN implementation. It does not seem reasonable to use a baseline you implement yourself when there are plenty of performant DQN implementations available online. Minor implementation details can often have outsized effect to outcome in RL, and the authors might not be incentivized to optimize the baseline as it's easier to improve upon a poor baseline. I would encourage the authors to start from a performant publicly available DQN implementation, e.g. the dopamine repo.\n\nThe authors write that “Our baseline appears to be significantly faster than popular existing DQN implementations (e.g. Roderick et al., 2017; Dhariwal et al., 2017; Hill et al., 2018) based on their reported results, although we did not compare them on our hardware.”. The difference might be entirely due to hardware differences, so this statement does not say much. Additionally, the cited papers are relatively old, and more modern implementations and improved hardware might be available today. \n\nTable 4 contains no standard deviations. Are results just computed over 1 seed per environment? If so, it might not be statistically meaningful.\n\nThe code does not seem to have all dependencies listed (e.g. cv2 is missing), and there's also no example on how to use it. For a paper whose main contribution is on sharing a good implementation, this is not ideal.\n\n## Minor comments\n\nI would recommend that the authors show the learning curves (of e.g. averaged human-normalized scores) in the main paper.\n\nI would encourage the authors to broadener their related work section to other methods that have been proposed for accelerating RL:\n\n[3] Accelerating reinforcement learning through GPU atari emulation, 2019\n\n[4] Large Batch Simulation for Deep Reinforcement Learning, 2021\n\n", "summary_of_the_review": "The paper aims for a reasonable goal and is both well written and well-motivated. However, the introduced methods are either well-known or incremental variations of known methods. Additionally, the empirical evaluation could be improved by using stronger baselines.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634589904930}], "openreview_url": "https://openreview.net/forum?id=KDAEc2nai83", "arxiv_id": "2111.01264", "paper_pdf": "papers/KDAEc2nai83.pdf", "paper_pdf_sha256": "14e52cc1cf4f718fc9e8c2cf7710f017ef4f76e28b8519995ed101c1c881fdd5", "paper_pdf_bytes": 426996, "paper_pdf_source": "openreview", "code_url": "https://github.com/brett-daley/fast-dqn", "code_repository": "brett-daley/fast-dqn", "code_commit": "acf21e8bb193e52d73aa8e2d4e355957095bbd36", "code_archive": "repos/KDAEc2nai83.zip", "code_archive_sha256": "32695a08511e302aeba8ecdd1fbb5fd0b37dc6642d406c4c522280dbf4080ff5", "code_archive_bytes": 18663, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 204, "github_languages": {"Python": 30454}, "github_archived": false, "github_pushed_at": "2022-04-18T22:21:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/human-level-control-without-server-grade-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "o1O5nc48rn", "year": 2021, "status": "rejected", "title": "Optimal Transport Graph Neural Networks", "authors": ["Gary Bécigneul", "Octavian-Eugen Ganea", "Benson Chen", "Regina Barzilay", "Tommi S. Jaakkola"], "authorids": ["~Gary_Bécigneul1", "~Octavian-Eugen_Ganea1", "~Benson_Chen1", "~Regina_Barzilay1", "~Tommi_S._Jaakkola1"], "authors_source": "OpenReview API", "abstract": "Current graph neural network (GNN) architectures naively average or sum node embeddings into an aggregated graph representation---potentially losing structural or semantic information. We here introduce OT-GNN, a model that computes graph embeddings using parametric prototypes that highlight key facets of different graph aspects. Towards this goal, we are (to our knowledge) the first to successfully combine optimal transport with parametric graph models. Graph representations are obtained from Wasserstein distances between the set of GNN node embeddings and \"prototype\" point clouds as free parameters. We theoretically prove that, unlike traditional sum aggregation, our function class on point clouds satisfies a fundamental universal approximation theorem. Empirically, we address an inherent collapse optimization issue by proposing a noise contrastive regularizer to steer the model towards truly exploiting the optimal transport geometry. Finally, we consistently report better generalization performance on several molecular property prediction tasks, while exhibiting smoother graph representations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "WJjMZ-dDXr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2320/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**\nThe paper proposes OT-GNN, which incorporates optimal transport distance to message passing of GNN. The message passing is aggregated by using a Wasserstein discrepancy for a point cloud. The contrastive regularization is utilized to overcome extreme clustering of nodes of the same class. Also, in theory, the author shows that the Wasserstein kernel is universal while the \"agg\" kernel is not. The proposed model is tested on several molecular property prediction tasks. The OT-GNN achieves slightly better performance against existing methods.\n\n**Pros**\n1. The paper introduces a non-Euclidean metric in message-passing aggregation and develops OT-GNN based on it. The proposed OT-GNN has good performance for molecular property prediction tasks.\n2. The paper provides a theoretical analysis of the universality and positiveness of the Wasserstein kernel. It shows the Wasserstein metric-induced message passing is superior to the Euclidean metric case in terms of universality. The paper also shows the $L_2$ Wasserstein kernel is not conditionally negative positive.\n3. The computational complexity is analyzed for Wasserstein optimal transport in the OT-GNN. \n\n**Cons**\n1. For Wasserstein discrepancy in (4), there seem many pairs of $H$ and $Q_i$, where $Q_i$ contains a set of free parameters. Then the OT-GNN would have too many parameters to train. Will it bring about overparameterization?\n2. Using the Wasserstein metric, the computational cost will increase much. Besides, the contrastive regularization will add more training time. What is the GPU wall time for OT-GNN training for the molecule regression tasks?\n3. The theoretical analysis for the properties of Wasserstein kernel seems not complicated, and it does not give a good interpretation of the learning ability, like universal approximation property or generalization of OT-GNN. In particular, the discussion of positiveness of the kernel is of no significance or use.\n4. The authors also need to try more public graph datasets, such as Open Graph Benchmark, https://ogb.stanford.edu/.\n5. Line 4 below (6): the last sentence seems not finished. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for Optimal Transport Graph Neural Networks", "review": "**Summary**\nThe paper proposes OT-GNN, which incorporates optimal transport distance to message passing of GNN. The message passing is aggregated by using a Wasserstein discrepancy for a point cloud. The contrastive regularization is utilized to overcome extreme clustering of nodes of the same class. Also, in theory, the author shows that the Wasserstein kernel is universal while the \"agg\" kernel is not. The proposed model is tested on several molecular property prediction tasks. The OT-GNN achieves slightly better performance against existing methods.\n\n**Pros**\n1. The paper introduces a non-Euclidean metric in message-passing aggregation and develops OT-GNN based on it. The proposed OT-GNN has good performance for molecular property prediction tasks.\n2. The paper provides a theoretical analysis of the universality and positiveness of the Wasserstein kernel. It shows the Wasserstein metric-induced message passing is superior to the Euclidean metric case in terms of universality. The paper also shows the $L_2$ Wasserstein kernel is not conditionally negative positive.\n3. The computational complexity is analyzed for Wasserstein optimal transport in the OT-GNN. \n\n**Cons**\n1. For Wasserstein discrepancy in (4), there seem many pairs of $H$ and $Q_i$, where $Q_i$ contains a set of free parameters. Then the OT-GNN would have too many parameters to train. Will it bring about overparameterization?\n2. Using the Wasserstein metric, the computational cost will increase much. Besides, the contrastive regularization will add more training time. What is the GPU wall time for OT-GNN training for the molecule regression tasks?\n3. The theoretical analysis for the properties of Wasserstein kernel seems not complicated, and it does not give a good interpretation of the learning ability, like universal approximation property or generalization of OT-GNN. In particular, the discussion of positiveness of the kernel is of no significance or use.\n4. The authors also need to try more public graph datasets, such as Open Graph Benchmark, https://ogb.stanford.edu/.\n5. Line 4 below (6): the last sentence seems not finished. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604087930232}, {"id": "bEVChHmSYj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2320/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper combines OT  with parametric graph neural network. It replace the inner product between the graph embedding and the first layer weights of MLP by the Wasserstein distance between the node embeddings and some point clouds. Then the GNN, point clouds and the downstream MLP are trained in an end-to-end way. A regularization term is adopted to enforce the point clouds are not collapsed. The authors then theoretically show that the Wasserstein kernel is universal.\n\nPros:\n1. The empirical study is thorough and the model has good empirical performance. \n2. The idea of using prototypes is interesting. That being said, I think the word \"prototype\" here is actually misleading. In Snell 2017, their prototypes is a representation of the data. In contrast, here Q_i is NOT the representation of H, since we are not minimizing the Wasserstein distance between them. Instead, Q_i just helps to keep some information that is useful for the task. \n\nAreas to improve:\n1. My major concern is about computation time.  In each training step, the algorithm computes batch size*(the range of indices i) of OT problems. Will the OT part introduce a large training time overhead? Could you please report the training time of each method?\n2. Since the model has significantly more parameters, e.g., in Q_i, it would be better to also compare it to a baseline model with comparable number of parameters and the node embedding info. For example, is it possible to compare it with: randomly selecting k embeddings, concatenate, and feed into MLP (adjusting k to make the parameter size comparable)?\n3. Some notations are not properly defined. For example, \\mathfrak{agg} first appears in section 4.1, but its formal definition is in section 4.2. Another example is indices i and j in eq (4). What is the range of i and the range of j? In other words, how is the range of i, the range of j, the dimension of q_i^j corresponds to n_pc, pc_size, pc_hidden in the appendix? \n4. Could you please provide some intuition that different variants of the proposed model performs the best for different dataset (table 1)? In other words, what variants is suitable for what kind of dataset?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Wasserstein kernel, good empirical performance", "review": "This paper combines OT  with parametric graph neural network. It replace the inner product between the graph embedding and the first layer weights of MLP by the Wasserstein distance between the node embeddings and some point clouds. Then the GNN, point clouds and the downstream MLP are trained in an end-to-end way. A regularization term is adopted to enforce the point clouds are not collapsed. The authors then theoretically show that the Wasserstein kernel is universal.\n\nPros:\n1. The empirical study is thorough and the model has good empirical performance. \n2. The idea of using prototypes is interesting. That being said, I think the word \"prototype\" here is actually misleading. In Snell 2017, their prototypes is a representation of the data. In contrast, here Q_i is NOT the representation of H, since we are not minimizing the Wasserstein distance between them. Instead, Q_i just helps to keep some information that is useful for the task. \n\nAreas to improve:\n1. My major concern is about computation time.  In each training step, the algorithm computes batch size*(the range of indices i) of OT problems. Will the OT part introduce a large training time overhead? Could you please report the training time of each method?\n2. Since the model has significantly more parameters, e.g., in Q_i, it would be better to also compare it to a baseline model with comparable number of parameters and the node embedding info. For example, is it possible to compare it with: randomly selecting k embeddings, concatenate, and feed into MLP (adjusting k to make the parameter size comparable)?\n3. Some notations are not properly defined. For example, \\mathfrak{agg} first appears in section 4.1, but its formal definition is in section 4.2. Another example is indices i and j in eq (4). What is the range of i and the range of j? In other words, how is the range of i, the range of j, the dimension of q_i^j corresponds to n_pc, pc_size, pc_hidden in the appendix? \n4. Could you please provide some intuition that different variants of the proposed model performs the best for different dataset (table 1)? In other words, what variants is suitable for what kind of dataset?", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603937970236}, {"id": "DFrMf-okCg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2320/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces OT-GNN, a combination of optimal transport and graph neural network, for graph-level tasks. Instead of the conventional graph embedding+MLP strategy, the authors propose to replace the readout layer in MPNN-like structures with Wasserstein discrepancy. In addition, the noise contrastive regularizer is added to the model so that the optimal transport plan is discriminative, and the new model could outperform its Euclidean counterpart. While the construction is sophisticated, the core idea behind the algorithm is not hard to follow. However, I do have some concerns regarding the presentation and experiments, which, at this stage, prevent me from recommending the paper confidently to the broader community.\n\n\nBelow I would like to list my main concerns or confusions regarding the paper.\n\nThe construction of prototypes is not very clear to me. I would expect the authors to provide more discussion before the experiment section regarding its structure, requirements, or formulation details.\nThe authors, if my understanding is correct, named the same subject to different terms as ‘prototypes’ and ‘free parameters’. This indeed creates unnecessary hurdles of understanding the paper, and I would recommend the names could be identified, or at least clearly and properly defined. \nThe design of Figure 3 is confusing. What is the difference between Prototype 1-5 and Prototype 6-10? As they are designed with different shapes, I would expect them to represent different kinds of prototypes. However, unless I missed something, I didn’t find a clear statement of it. Also, I’m not sure what is the purpose of showing five real molecules? Are we supposed to compare them with the corresponding prototypes?\nWhat is the partition function you mentioned in the 3rd line after Eq.6? Do you use different terminology from in Nickel & Kiela (2017)? \nThe authors only compared their method with naive node embedding+sum aggregation models but ignored a major type of graph representation learning methods as graph pooling. This branch is developed rapidly in the recent 2 years, and many simple yet powerful methods have been proposed. I don’t think it’s fair to compare with the models that are specialized in node embedding tasks but neglect the real methods that are designed for graph embedding tasks. \nFollowing the last point, I would expect more recent research to be included in the Related Work section.\nSome figures could be redesigned for better presentation. For example, it’s hard to read the axis information in Figure 4.\n\nMinor typos:\n\n(line 2, Section 2.1) “…property prediction by Yang et al. Yang et al. (2019).”\n(line 1, Section 2.2) “Optimal Transport (OT) Peyré et al. (2019) is a mathematical framework…”\n\nGiven the problems stated as above, I would suggest the paper to be carefully polished before it’s ready to publish.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for OT-GNN", "review": "This paper introduces OT-GNN, a combination of optimal transport and graph neural network, for graph-level tasks. Instead of the conventional graph embedding+MLP strategy, the authors propose to replace the readout layer in MPNN-like structures with Wasserstein discrepancy. In addition, the noise contrastive regularizer is added to the model so that the optimal transport plan is discriminative, and the new model could outperform its Euclidean counterpart. While the construction is sophisticated, the core idea behind the algorithm is not hard to follow. However, I do have some concerns regarding the presentation and experiments, which, at this stage, prevent me from recommending the paper confidently to the broader community.\n\n\nBelow I would like to list my main concerns or confusions regarding the paper.\n\nThe construction of prototypes is not very clear to me. I would expect the authors to provide more discussion before the experiment section regarding its structure, requirements, or formulation details.\nThe authors, if my understanding is correct, named the same subject to different terms as ‘prototypes’ and ‘free parameters’. This indeed creates unnecessary hurdles of understanding the paper, and I would recommend the names could be identified, or at least clearly and properly defined. \nThe design of Figure 3 is confusing. What is the difference between Prototype 1-5 and Prototype 6-10? As they are designed with different shapes, I would expect them to represent different kinds of prototypes. However, unless I missed something, I didn’t find a clear statement of it. Also, I’m not sure what is the purpose of showing five real molecules? Are we supposed to compare them with the corresponding prototypes?\nWhat is the partition function you mentioned in the 3rd line after Eq.6? Do you use different terminology from in Nickel & Kiela (2017)? \nThe authors only compared their method with naive node embedding+sum aggregation models but ignored a major type of graph representation learning methods as graph pooling. This branch is developed rapidly in the recent 2 years, and many simple yet powerful methods have been proposed. I don’t think it’s fair to compare with the models that are specialized in node embedding tasks but neglect the real methods that are designed for graph embedding tasks. \nFollowing the last point, I would expect more recent research to be included in the Related Work section.\nSome figures could be redesigned for better presentation. For example, it’s hard to read the axis information in Figure 4.\n\nMinor typos:\n\n(line 2, Section 2.1) “…property prediction by Yang et al. Yang et al. (2019).”\n(line 1, Section 2.2) “Optimal Transport (OT) Peyré et al. (2019) is a mathematical framework…”\n\nGiven the problems stated as above, I would suggest the paper to be carefully polished before it’s ready to publish.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603786283277}, {"id": "6AQr6RMJncj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2320/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Summary\n\nThe paper introduces a novel approach to aggregate information of graph neural network node embeddings in order to support graph-level machine learning, like graph classification or graph regression. The aggregation is performed by comparing the node embeddings of a graph to learned, prototypical node embeddings via the Wasserstein distance. The vector of distances to the prototypes then serves as a concise representation that can be fed into a subsequent standard multi-layer perceptron for classification or regression. This approach is proven to be strictly more powerful than just adding up the node embeddings, which is the most common state-of-the-art. Finally, the paper evaluates the proposed aggregation function on two graph classification and two graph regression data sets, showing superior performance in all cases.\n\n## Strengths\n\nThe paper has several strengths worth highlighting:\n\n* The idea of using prototypical node embeddings as a method for aggregation is easy to understand and potentially interpretable. For example, one could try to construct graphs that best correspond to each prototype and thereby gather understanding what typical graphs look like based on which the model makes its decision.\n* The idea also corresponds well to past research on kernels as well as distances, thus nicely integrating the strengths of graph neural nets and kernel/distance theory. The use of the Wasserstein distance to compare point clouds is particularly compelling and combines two hot topics in machine learning today, namely graph neural nets and the Wasserstein distance, which should be interesting to a broad section of the ICLR community.\n* The theoretical results, especially the universal approximation result, are compelling and further make clear that the present work is strictly more powerful than the state of the art (i.e. just summing up all node embeddings)\n* The experimental evaluation is performed against multiple reasonable baselines, including ablation variants of the proposed model, on several data sets, covering both regression and classification problems. Multiple repeats are performed, giving an impression of variance. Additionally, for the regression data sets, the paper shows both correlation as well as dimensionality reduction results indicating that the proposed representation is more smooth in relation to the targets compared to existing models.\n\n## Weaknesses\n\nThat being said, I also see some issues which could be addressed to make the paper even stronger.\n\n* Most importantly, from my perspective, the relation to existing work warrants much more discussion in two directions:\n    * Currently, the discussion of prototype learning is reduced to the work of Snell (2017), whereas a long research tradition of learning prototypes for classification and regression is currently not covered. In particular, learning vector quantization approaches (refer e.g. to [Sato & Yamada; 1995][GLVQ]) have been used to learn prototypes that represent classes well and can then be used for a one-nearest neighbor classification. Even more similar to the present paper, radial basis function networks (e.g. [Chen, Cowan, and Grant, 1991][RBF]) have been used to solve regression tasks in the space of RBF kernel values to prototypes. It would be, I believe, helpful to contrast the present prototype learning approach to these past works. Additionally, these past works may give inspiration how to learn prototypes well and how to solve the issues of prototype collapse.\n    * Currently, the paper treats distances as a kind of non-definite kernel. This can be done, but is unusual. An understanding of kernel and distance as opposing (but related) concepts is more common and, I believe, more instructive. However, the use of distances as features can still be justified, e.g. using the work of [Pekalska & Duin (2012)][DistanceTheory]. Similarly, it would be helpful to relate the present work to attempts that build bridges between kernels and neural nets, such as [Wilson, Hu, Salakhutdinov, and Xing (2016)][DeepKernels]\n* I see several issues with how prototypes are integrated into the model.\n    * It is not clear to me why the distances to the prototypes are plugged into another MLP instead of using them directly, at least for classification. In particular, prior work of Snell (2017) as well as [learning vector quantization][GLVQ] suggests to associate each prototype with a label and learn the prototypes such that a one-nearest-neighbor classification solves the classification problem. For regression, the issue is a bit more difficult, but here, as well, I would expect a single additional layer, similar to [RBF networks][RBF] and not several additional layers. Ideally, it would be beneficial to provide comparisons to these alternatives as additional ablation results in the appendix.\n    * Using prototypes introduces two new hyperparameters, namely the number of prototypes and the size of points within each prototype. The paper currently gives no guidance on how to choose these hyperparameters.\n    * Prototypes have the advantage of being a sparse representation of the data which lends itself to interpretation. However, currently no such interpretation is attempted. This seems to me like a missed opportunity. Showing an example graph for each prototype (i.e. a graph that is close to one prototype but far from every other prototype) could give additional insights into how the model represents the data.\n    * The regularizer suggested in the paper does not seem sufficiently justified to me, yet. In particular, there are exponentially many 'bad' transport plans which I would not like to choose. How can I guarantee that my set N(T_i) is sufficiently rich to 'protect' me against all bad transport plans? Conversely, if the aim is just to prevent collapsing of the prototypes, would there not be simpler ways to prevent such a collapse, such as regularizing with the negative log of pairwise distances between points inside a prototype?\n* Leaving the optimal transport plan T constant during backpropagation seems reasonable to me but the justification appears still rather hand-wavy (this applies to the cited paper of Xu (2019) as well). After inspecting the original paper of Afriat (1971), I am not ale to see how the results described there are applicable to transport plans (I don't even see an 'envelop theorem' in there). The connection may be there, but it is not obvious to me. Going back to a higher-level intuition, I think the key issue here is that even a very slight change in the points of a prototype set could dramatically change the optimal transport plan (that is just how transport plans behave, is my understanding; whenever another assignment becomes slightly better, suddenly the assignment changes in a discontinuous fashion). This speaks against leaving it constant. In the worst case, it could occur that the current transport plan suggests to move a point into one direction, but the transport plan at the next point suggests to move it back, yielding an endless cycle. One could, however, approximate the transport plan smoothly, e.g. as a mixture of several transport plans between which we interpolate continuously, and then one could make the argument that small gradient steps do not change the plan much, hence we can leave it constant. But again: Such an argument is not trivial and is currently missing. Furthermore, it would be good to clarify that the transport plan is only treated as constant for backpropagation, but that it is updated after each gradient step (that is, at least, how I interpret the approach).\n* It would help to qualify Theorem 1 a little. First, Theorem 1 does not imply that the Wasserstein kernel can distinguish between all graphs; only between the graphs that are distinct in the multi-set of node embeddings, i.e. all Weisfeiler-Lehman distinguishable graphs. Second, Theorem 1 does not imply that the model as proposed is a universal approximator because the proposed model only uses a finite set of prototypes to represent the data. To make it a universal approximator, one would need infinitely many prototypes, if I am not mistaken. In practice, this is likely not an issue because few prototypes are sufficient to cover the space pretty well. It is still a limitation to how much Theorem 1 actually tells us about the proposed model.\n* The experiments currently do not report empiric runtime results. These would be a valuable addition because they illustrate how much overhead prototypes introduce compared to just summing up the final layer.\n\nThere are also a few smaller issues which could be clarified:\n\n* page 3: I wonder why the marginal distributions over points in X and points in Y are restricted to be uniform. Would the more general case not be interesting as well, where, say, prototypes receive more weight if they are more important for representing the data?\n* Page 3: The vectors r_i are not introduced. Are these supposed to be the weight vectors of neurons in the MLP?\n* Page 3: It is not clear to me why the negative inner product and the Euclidean distance should yield the same transport plan. I can see that for the special case of vectors with norm 1 where ||x - y||^2 = 1 - 2 * <x, y> + 1, but not in the general case.\n\n## Judgment\n\nOverall, I believe that this paper is strong enough to warrant acceptance. In particular, the heart of the paper, namely the proposed model architecture, the main theoretical result, and the experiments, are strong and my issues - namely related work discussions, some theoretical fine points, and ablation studies - are aspects that can be addressed during a revision without affecting the core paper too much. Furthermore, I believe that the core idea - combining prototypes, optimal transport, and graph neural nets - has great potential to make graph neural nets stronger and more interpretable, which are both crucial goals for the research community right now. As such, I am confident that the paper, as it stands, is worthy of publication and, with some additional work during revision, can become even better.\n\n## References\n\n* Sato, A., & Yamada, K. (1995). Generalized Learning Vector Quantization. Proceedings of NIPS 1995. [Link][GLVQ]\n* de Vries, H., Memisevic, R., & Courville, A. (2016). Deep Learning Vector Quantization. Proceedings of ESANN 2016. [Link][DLVQ]\n* Chen, S., Cowan, C., and Grant, P. (1991). Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on neural networks, 2(2). [Link][RBF]\n* Duin, R., and Pekalska, E. (2012). The dissimilarity space: Bridging structural and statistical pattern recognition. Pattern Recognition Letters, 33(7). doi:[10.1016/j.patrec.2011.04.019][DistanceTheory]\n* Wilson, A., Hu, Z., Salakhutdinov, R., and Xing, E. (2016). Deep Kernel Learning. Proceedings of AISTATS 2016. [Link][DeepKernels]\n\n[GLVQ]:https://papers.nips.cc/paper/1113-generalized-learning-vector-quantization \"Sato, A., & Yamada, K. (1995). Generalized Learning Vector Quantization. Proceedings of NIPS 1995.\"\n[DLVQ]:https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2016-112.pdf \"de Vries, H., Memisevic, R., & Courville, A. (2016). Deep Learning Vector Quantization. Proceedings of ESANN 2016.\"\n[RBF]:https://eprints.soton.ac.uk/251135/1/00080341.pdf \"Chen, S., Cowan, C., and Grant, P. (1991). Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on neural networks, 2(2).\"\n[DistanceTheory]:https://doi.org/10.1016/j.patrec.2011.04.019 \"Duin, R., and Pekalska, E. (2012). The dissimilarity space: Bridging structural and statistical pattern recognition. Pattern Recognition Letters, 33(7).\"\n[DeepKernels]:http://proceedings.mlr.press/v51/wilson16.html \"Wilson, A., Hu, Z., Salakhutdinov, R., and Xing, E. (2016). Deep Kernel Learning. Proceedings of AISTATS 2016\"\n\n## Clarifications\n\nWhile I am quite confident in my review, it would help if the authors could verify that I understood some key points in the paper correctly:\n\n* Each prototype here is a collection of points, and the number of prototypes as well as the number of points is a fixed hyperparameter, whereas the location of each point is learned, correct?\n* The pipeline for classification/regression is to first embed the nodes of a graph via a graph neural net, yielding a point cloud with one point per node; second, to compare this point cloud to all prototypes using the Wasserstein distance/inner product; third to collect the resulting distances/inner products into a vector d/k with as many entries as there are prototypes and to feed this vector into a standard MLP to make the final classification/regression decision, correct?\n* The optimal transport plan T is treated as constant for backprop, but is recomputed after each gradient step, correct?\n* Are the limitations to Theorem 1 described above correct or did I misunderstand the Theorem?\n\n## Typos\n\nBeyond the points mentioned above, there are also a few typos and very minor things that do not affect my judgment but could easily be fixed to improve the paper.\n\n* Abstract: The abbreviation 'OT' is not introduced\n* page 1: 'As a result, some of the information naturally extracted by node embeddings may be lost' <-- This is correct, but could be underlined more with a citation\n* page 2: 'that is universal approximator.' -> that is a universal approximator\n* page 2: The abbreviation 'D-MPNN' is not introduced\n* Figure 3 is currently rather busy and not very easy to interpret. Would it, perhaps, suffice to only show prototype 1 and one reference molecule?\n* As far as I know, ICLR permits an appendix. As such, it would be good to move the appendix into the main paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A strong combination of optimal transport and graph neural nets, with some room for improvement with respect to prototypes", "review": "## Summary\n\nThe paper introduces a novel approach to aggregate information of graph neural network node embeddings in order to support graph-level machine learning, like graph classification or graph regression. The aggregation is performed by comparing the node embeddings of a graph to learned, prototypical node embeddings via the Wasserstein distance. The vector of distances to the prototypes then serves as a concise representation that can be fed into a subsequent standard multi-layer perceptron for classification or regression. This approach is proven to be strictly more powerful than just adding up the node embeddings, which is the most common state-of-the-art. Finally, the paper evaluates the proposed aggregation function on two graph classification and two graph regression data sets, showing superior performance in all cases.\n\n## Strengths\n\nThe paper has several strengths worth highlighting:\n\n* The idea of using prototypical node embeddings as a method for aggregation is easy to understand and potentially interpretable. For example, one could try to construct graphs that best correspond to each prototype and thereby gather understanding what typical graphs look like based on which the model makes its decision.\n* The idea also corresponds well to past research on kernels as well as distances, thus nicely integrating the strengths of graph neural nets and kernel/distance theory. The use of the Wasserstein distance to compare point clouds is particularly compelling and combines two hot topics in machine learning today, namely graph neural nets and the Wasserstein distance, which should be interesting to a broad section of the ICLR community.\n* The theoretical results, especially the universal approximation result, are compelling and further make clear that the present work is strictly more powerful than the state of the art (i.e. just summing up all node embeddings)\n* The experimental evaluation is performed against multiple reasonable baselines, including ablation variants of the proposed model, on several data sets, covering both regression and classification problems. Multiple repeats are performed, giving an impression of variance. Additionally, for the regression data sets, the paper shows both correlation as well as dimensionality reduction results indicating that the proposed representation is more smooth in relation to the targets compared to existing models.\n\n## Weaknesses\n\nThat being said, I also see some issues which could be addressed to make the paper even stronger.\n\n* Most importantly, from my perspective, the relation to existing work warrants much more discussion in two directions:\n    * Currently, the discussion of prototype learning is reduced to the work of Snell (2017), whereas a long research tradition of learning prototypes for classification and regression is currently not covered. In particular, learning vector quantization approaches (refer e.g. to [Sato & Yamada; 1995][GLVQ]) have been used to learn prototypes that represent classes well and can then be used for a one-nearest neighbor classification. Even more similar to the present paper, radial basis function networks (e.g. [Chen, Cowan, and Grant, 1991][RBF]) have been used to solve regression tasks in the space of RBF kernel values to prototypes. It would be, I believe, helpful to contrast the present prototype learning approach to these past works. Additionally, these past works may give inspiration how to learn prototypes well and how to solve the issues of prototype collapse.\n    * Currently, the paper treats distances as a kind of non-definite kernel. This can be done, but is unusual. An understanding of kernel and distance as opposing (but related) concepts is more common and, I believe, more instructive. However, the use of distances as features can still be justified, e.g. using the work of [Pekalska & Duin (2012)][DistanceTheory]. Similarly, it would be helpful to relate the present work to attempts that build bridges between kernels and neural nets, such as [Wilson, Hu, Salakhutdinov, and Xing (2016)][DeepKernels]\n* I see several issues with how prototypes are integrated into the model.\n    * It is not clear to me why the distances to the prototypes are plugged into another MLP instead of using them directly, at least for classification. In particular, prior work of Snell (2017) as well as [learning vector quantization][GLVQ] suggests to associate each prototype with a label and learn the prototypes such that a one-nearest-neighbor classification solves the classification problem. For regression, the issue is a bit more difficult, but here, as well, I would expect a single additional layer, similar to [RBF networks][RBF] and not several additional layers. Ideally, it would be beneficial to provide comparisons to these alternatives as additional ablation results in the appendix.\n    * Using prototypes introduces two new hyperparameters, namely the number of prototypes and the size of points within each prototype. The paper currently gives no guidance on how to choose these hyperparameters.\n    * Prototypes have the advantage of being a sparse representation of the data which lends itself to interpretation. However, currently no such interpretation is attempted. This seems to me like a missed opportunity. Showing an example graph for each prototype (i.e. a graph that is close to one prototype but far from every other prototype) could give additional insights into how the model represents the data.\n    * The regularizer suggested in the paper does not seem sufficiently justified to me, yet. In particular, there are exponentially many 'bad' transport plans which I would not like to choose. How can I guarantee that my set N(T_i) is sufficiently rich to 'protect' me against all bad transport plans? Conversely, if the aim is just to prevent collapsing of the prototypes, would there not be simpler ways to prevent such a collapse, such as regularizing with the negative log of pairwise distances between points inside a prototype?\n* Leaving the optimal transport plan T constant during backpropagation seems reasonable to me but the justification appears still rather hand-wavy (this applies to the cited paper of Xu (2019) as well). After inspecting the original paper of Afriat (1971), I am not ale to see how the results described there are applicable to transport plans (I don't even see an 'envelop theorem' in there). The connection may be there, but it is not obvious to me. Going back to a higher-level intuition, I think the key issue here is that even a very slight change in the points of a prototype set could dramatically change the optimal transport plan (that is just how transport plans behave, is my understanding; whenever another assignment becomes slightly better, suddenly the assignment changes in a discontinuous fashion). This speaks against leaving it constant. In the worst case, it could occur that the current transport plan suggests to move a point into one direction, but the transport plan at the next point suggests to move it back, yielding an endless cycle. One could, however, approximate the transport plan smoothly, e.g. as a mixture of several transport plans between which we interpolate continuously, and then one could make the argument that small gradient steps do not change the plan much, hence we can leave it constant. But again: Such an argument is not trivial and is currently missing. Furthermore, it would be good to clarify that the transport plan is only treated as constant for backpropagation, but that it is updated after each gradient step (that is, at least, how I interpret the approach).\n* It would help to qualify Theorem 1 a little. First, Theorem 1 does not imply that the Wasserstein kernel can distinguish between all graphs; only between the graphs that are distinct in the multi-set of node embeddings, i.e. all Weisfeiler-Lehman distinguishable graphs. Second, Theorem 1 does not imply that the model as proposed is a universal approximator because the proposed model only uses a finite set of prototypes to represent the data. To make it a universal approximator, one would need infinitely many prototypes, if I am not mistaken. In practice, this is likely not an issue because few prototypes are sufficient to cover the space pretty well. It is still a limitation to how much Theorem 1 actually tells us about the proposed model.\n* The experiments currently do not report empiric runtime results. These would be a valuable addition because they illustrate how much overhead prototypes introduce compared to just summing up the final layer.\n\nThere are also a few smaller issues which could be clarified:\n\n* page 3: I wonder why the marginal distributions over points in X and points in Y are restricted to be uniform. Would the more general case not be interesting as well, where, say, prototypes receive more weight if they are more important for representing the data?\n* Page 3: The vectors r_i are not introduced. Are these supposed to be the weight vectors of neurons in the MLP?\n* Page 3: It is not clear to me why the negative inner product and the Euclidean distance should yield the same transport plan. I can see that for the special case of vectors with norm 1 where ||x - y||^2 = 1 - 2 * <x, y> + 1, but not in the general case.\n\n## Judgment\n\nOverall, I believe that this paper is strong enough to warrant acceptance. In particular, the heart of the paper, namely the proposed model architecture, the main theoretical result, and the experiments, are strong and my issues - namely related work discussions, some theoretical fine points, and ablation studies - are aspects that can be addressed during a revision without affecting the core paper too much. Furthermore, I believe that the core idea - combining prototypes, optimal transport, and graph neural nets - has great potential to make graph neural nets stronger and more interpretable, which are both crucial goals for the research community right now. As such, I am confident that the paper, as it stands, is worthy of publication and, with some additional work during revision, can become even better.\n\n## References\n\n* Sato, A., & Yamada, K. (1995). Generalized Learning Vector Quantization. Proceedings of NIPS 1995. [Link][GLVQ]\n* de Vries, H., Memisevic, R., & Courville, A. (2016). Deep Learning Vector Quantization. Proceedings of ESANN 2016. [Link][DLVQ]\n* Chen, S., Cowan, C., and Grant, P. (1991). Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on neural networks, 2(2). [Link][RBF]\n* Duin, R., and Pekalska, E. (2012). The dissimilarity space: Bridging structural and statistical pattern recognition. Pattern Recognition Letters, 33(7). doi:[10.1016/j.patrec.2011.04.019][DistanceTheory]\n* Wilson, A., Hu, Z., Salakhutdinov, R., and Xing, E. (2016). Deep Kernel Learning. Proceedings of AISTATS 2016. [Link][DeepKernels]\n\n[GLVQ]:https://papers.nips.cc/paper/1113-generalized-learning-vector-quantization \"Sato, A., & Yamada, K. (1995). Generalized Learning Vector Quantization. Proceedings of NIPS 1995.\"\n[DLVQ]:https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2016-112.pdf \"de Vries, H., Memisevic, R., & Courville, A. (2016). Deep Learning Vector Quantization. Proceedings of ESANN 2016.\"\n[RBF]:https://eprints.soton.ac.uk/251135/1/00080341.pdf \"Chen, S., Cowan, C., and Grant, P. (1991). Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on neural networks, 2(2).\"\n[DistanceTheory]:https://doi.org/10.1016/j.patrec.2011.04.019 \"Duin, R., and Pekalska, E. (2012). The dissimilarity space: Bridging structural and statistical pattern recognition. Pattern Recognition Letters, 33(7).\"\n[DeepKernels]:http://proceedings.mlr.press/v51/wilson16.html \"Wilson, A., Hu, Z., Salakhutdinov, R., and Xing, E. (2016). Deep Kernel Learning. Proceedings of AISTATS 2016\"\n\n## Clarifications\n\nWhile I am quite confident in my review, it would help if the authors could verify that I understood some key points in the paper correctly:\n\n* Each prototype here is a collection of points, and the number of prototypes as well as the number of points is a fixed hyperparameter, whereas the location of each point is learned, correct?\n* The pipeline for classification/regression is to first embed the nodes of a graph via a graph neural net, yielding a point cloud with one point per node; second, to compare this point cloud to all prototypes using the Wasserstein distance/inner product; third to collect the resulting distances/inner products into a vector d/k with as many entries as there are prototypes and to feed this vector into a standard MLP to make the final classification/regression decision, correct?\n* The optimal transport plan T is treated as constant for backprop, but is recomputed after each gradient step, correct?\n* Are the limitations to Theorem 1 described above correct or did I misunderstand the Theorem?\n\n## Typos\n\nBeyond the points mentioned above, there are also a few typos and very minor things that do not affect my judgment but could easily be fixed to improve the paper.\n\n* Abstract: The abbreviation 'OT' is not introduced\n* page 1: 'As a result, some of the information naturally extracted by node embeddings may be lost' <-- This is correct, but could be underlined more with a citation\n* page 2: 'that is universal approximator.' -> that is a universal approximator\n* page 2: The abbreviation 'D-MPNN' is not introduced\n* Figure 3 is currently rather busy and not very easy to interpret. Would it, perhaps, suffice to only show prototype 1 and one reference molecule?\n* As far as I know, ICLR permits an appendix. As such, it would be good to move the appendix into the main paper.", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1602593027756}], "openreview_url": "https://openreview.net/forum?id=o1O5nc48rn", "arxiv_id": "2006.04804", "paper_pdf": "papers/o1O5nc48rn.pdf", "paper_pdf_sha256": "348e8543fb9192be1e97e74bba1e81ca30c114da4ebde5ac13864d189136d5ac", "paper_pdf_bytes": 2375559, "paper_pdf_source": "openreview", "code_url": "https://github.com/benatorc/OTGNN", "code_repository": "benatorc/OTGNN", "code_commit": "9282a8627ff3b28d40b538d75217cd87b4a45341", "code_archive": "repos/o1O5nc48rn.zip", "code_archive_sha256": "c308585f21e5e6793defb0545a9aaa2ed93d883bc84cb60103d9a50e5b40ce0a", "code_archive_bytes": 253339, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 228, "github_languages": {"Python": 77141}, "github_archived": false, "github_pushed_at": "2020-06-05T16:24:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimal-transport-graph-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJl1o2NFwS", "year": 2020, "status": "rejected", "title": "Understanding and Improving Transformer From a Multi-Particle Dynamic System Point of View", "authors": ["Yiping Lu", "Zhuohan Li", "Di He", "Zhiqing Sun", "Bin Dong", "Tao Qin", "Liwei Wang", "Tie-Yan Liu"], "authorids": ["yplu@stanford.edu", "zhuohan@berkeley.edu", "di_he@pku.edu.cn", "zhiqings@andrew.cmu.edu", "bindong@math.pku.edu.cn", "taoqin@microsoft.com", "wanglw@cis.pku.edu.cn", "tyliu@microsoft.com"], "authors_source": "OpenReview API", "abstract": "The Transformer architecture is widely used in natural language processing. Despite its success, the design principle of the Transformer remains elusive. In this paper, we provide a novel perspective towards understanding the architecture: we show that the Transformer can be mathematically interpreted as a numerical Ordinary Differential Equation (ODE) solver for a convection-diffusion equation in a multi-particle dynamic system. In particular, how words in a sentence are abstracted into contexts by passing through the layers of the Transformer can be interpreted as approximating multiple particles' movement in the space using the Lie-Trotter splitting scheme and the Euler's method. Given this ODE's perspective, the rich literature of numerical analysis can be brought to guide us in designing effective structures beyond the Transformer. As an example, we propose to replace the Lie-Trotter splitting scheme by the Strang-Marchuk splitting scheme, a scheme that is more commonly used and with much lower local truncation errors. The Strang-Marchuk splitting scheme suggests that the self-attention and position-wise feed-forward network (FFN) sub-layers should not be treated equally. Instead, in each layer, two position-wise FFN sub-layers should be used, and the self-attention sub-layer is placed in between. This leads to a brand new architecture. Such an FFN-attention-FFN layer is \"Macaron-like\", and thus we call the network with this new architecture the Macaron Net. Through extensive experiments, we show that the Macaron Net is superior to the Transformer on both supervised and unsupervised learning tasks. The reproducible code can be found on http://anonymized", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyexZfMiKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper136/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper points out a formal analogy between transformers and an ODE modelling multi-particle convection (the feed-forward network) and diffusion (the self-attention head). The paper then adapts the Strang-Marchuk splitting scheme for solving ODEs to construct a slightly different transformer architecture: “FFN of Attention of FFN”, instead of “FFN of Attention”. The new architecture, refered to as a Macaron-Net, yields better performance in a variety of experiments.\n\nPROs\n1. The proposed new architecture is fairly simple.\n2. The experimental results are fairly good.\n\nCONs\n1. Introducing two feedforward layers with *different* parameters W^{down} and W^{up} is a significant deviation from Strang-Marchuk splitting. I expected the two FFNs inside the Macaron to have the same weights. As I understand it, the motivation for the splitting is to improve the numerical performance of the update scheme for the ODE. In contrast, allowing different weights for the FFNs means the “physical process” is now a lot more “free”. Is there any “physical” motivation for the different parameters? (Beyond the fact that it improves performance). How much worse is empirical performance when the parameters are the same?\n\n2. Following on from the above point, the analogy between the multi-particle system and the transformer is quite weak. The equations look similar when you squint the right way. But that’s as far as it goes. Fig 1 is a nice visualization, but it doesn’t provide insight into the dynamics of transformers. What does “Particles move in the space along time (Semantics encoded in stacked neural network layers)” mean? How do the particles connect to the semantics? \n3. The proof of Bobylev & Ohwada’s theorem is included in the paper. Is there any connection between the theorem (or the techniques used in its proof) and transformers? I suspect the answer is no.\n\nSUMMARY\nIn short, the paper (i) proposes two FFN layers instead of one in each block of the transformer and (ii) shows it performs slightly better than before. This is decent, but in my opinion not enough the clear the bar for ICLR.\n\nThe connection to multi-particle ODEs is genuinely interesting. However, it is not sufficiently fleshed out to count as a contribution (yet). It’s possible the authors have discovered something deep. It’s also possible they got lucky with a physically motivated modification of transformers that actually has nothing to do with the dynamics of multi-particle systems. I’m not sure what further experiments would be needed to make the case. But I recommend the authors dig into the equations and the dynamics to see what it really going on under the hood. Just showing improved performance on a few benchmarks is not enough to convince the connection is solid. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper points out a formal analogy between transformers and an ODE modelling multi-particle convection (the feed-forward network) and diffusion (the self-attention head). The paper then adapts the Strang-Marchuk splitting scheme for solving ODEs to construct a slightly different transformer architecture: “FFN of Attention of FFN”, instead of “FFN of Attention”. The new architecture, refered to as a Macaron-Net, yields better performance in a variety of experiments.\n\nPROs\n1. The proposed new architecture is fairly simple.\n2. The experimental results are fairly good.\n\nCONs\n1. Introducing two feedforward layers with *different* parameters W^{down} and W^{up} is a significant deviation from Strang-Marchuk splitting. I expected the two FFNs inside the Macaron to have the same weights. As I understand it, the motivation for the splitting is to improve the numerical performance of the update scheme for the ODE. In contrast, allowing different weights for the FFNs means the “physical process” is now a lot more “free”. Is there any “physical” motivation for the different parameters? (Beyond the fact that it improves performance). How much worse is empirical performance when the parameters are the same?\n\n2. Following on from the above point, the analogy between the multi-particle system and the transformer is quite weak. The equations look similar when you squint the right way. But that’s as far as it goes. Fig 1 is a nice visualization, but it doesn’t provide insight into the dynamics of transformers. What does “Particles move in the space along time (Semantics encoded in stacked neural network layers)” mean? How do the particles connect to the semantics? \n3. The proof of Bobylev & Ohwada’s theorem is included in the paper. Is there any connection between the theorem (or the techniques used in its proof) and transformers? I suspect the answer is no.\n\nSUMMARY\nIn short, the paper (i) proposes two FFN layers instead of one in each block of the transformer and (ii) shows it performs slightly better than before. This is decent, but in my opinion not enough the clear the bar for ICLR.\n\nThe connection to multi-particle ODEs is genuinely interesting. However, it is not sufficiently fleshed out to count as a contribution (yet). It’s possible the authors have discovered something deep. It’s also possible they got lucky with a physically motivated modification of transformers that actually has nothing to do with the dynamics of multi-particle systems. I’m not sure what further experiments would be needed to make the case. But I recommend the authors dig into the equations and the dynamics to see what it really going on under the hood. Just showing improved performance on a few benchmarks is not enough to convince the connection is solid. \n"}, "tcdate": 1571656184202}, {"id": "S1eKfLh4KH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper136/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this work, the authors show that the sequence of self-attention and feed-forward layers within a Transformer can be interpreted as an approximate numerical solution to a set of coupled ODEs. Based on this insight, the authors propose to replace the first-order Lie-Trotter splitting scheme by the more accurate, second-order Strang splitting scheme. They then present experimental results that indicate an improved performance of their Macaron Net compared to the Transformer and argue that this is due to the former being a more accurate numerical solution to the underlying set of ODEs.\n\nThe authors highlight an interesting connection between the Transformer architecture and ODEs. In particular, they derive a set of ODEs that is solved numerically by the Transformer and borrow from the body of literature on numerical ODE solvers to improve the architecture. I find that this is a very elegant and promising approach for finding better architectures. \n\nHowever, I also identified two major and a couple of minor shortcomings of the paper that are explained in detail below. Based on these shortcomings, I recommend rejecting the paper but I would be willing to increase the score if these points were addressed in sufficient detail.\n\nMajor points:\n\n1) Replacing the first-order operator splitting scheme by a second-order scheme only guarantees a lower overall truncation error if the split ODEs are solved with sufficiently high accuracy. In particular, the overall accuracy of the numerical solution to the original ODE depends on the accuracy of the operator splitting and the accuracy of the integration scheme used to solve the split ODEs (e.g. Euler’s method). The authors improve the operator splitting, i.e. they make it second-order, but they keep Euler’s method to integrate the individual ODEs. Because of that, they actually do not get rid of the lowest-order error term of the overall scheme and therefore do not obtain a more accurate ODE solver. I think this is a crucial point that invalidates the authors’ claim that the Macaron Net employs a higher-order integration scheme. As far as I am aware, this is not commented on in the paper at all. To address this shortcoming, the authors could replace Euler’s method by a second-order integrator. \n\n2) The experiments considered in this paper are interesting and show competitive performance but, in my opinion, they do not sufficiently support the claim that the Macaron Net yields a more accurate solution to the underlying set of ODEs compared to a Transformer. For a more convincing support of this claim, the authors could consider a toy problem, i.e. a simple set of ODEs with known analytical solution, and actually show that the Macaron Net is more accurate. The accuracy of a numerical ODE solver is commonly assessed by plotting the absolute difference between the exact solution (or a high-resolution numerical approximation to it) and the numerical solution vs the timestep (here \\gamma). I suspect that such an analysis would support my previous comment and show that the proposed new architecture is not more accurate ODE solver than the original one.  \n\nMinor points and questions:\n\ni) Eqs. (17-19) suggest that you apply two different FFN layers (doubling the number of parameters) instead of applying the same FFN layer twice. You comment on this in Sec. 4.1 when you say ‘..., we set the dimensionality of the inner-layer of the two FFN sub-layers in the Macaron layers to two times of the dimensionality of the hidden states’. It is not clear to me why consistency with Strang splitting requires two different layers rather than applying the same FFN layer twice. Is the reason for having a separate, trainable layer to account for the explicit time dependence of G in Eq. (16)? I think that this is a very important point that should be clarified.\n\nii) I think that this type of system is usually referred to as ‘dynamical system’ and not ‘dynamic system’. Please check that and, if applicable, update the title.  \n\niii) The authors say that Eq. (5) is a 'convection-diffusion equation’. As far as I am aware, the diffusion equation is a partial differential equation (PDE). Perhaps there is a different notion 'diffusion equation’ in ODE theory. If that’s the case, could the authors please clarify this point to avoid confusion, e.g. by adding a suitable reference in which this type of ODE is classified as a convection-diffusion equation? \n\niv) In Sec. 2 (2nd paragraph),  you cite Vaswani et al. (2017) but in that work the quantity under the square-root in the denominator of Attention(Q, K, V) is actually d_k, the dimension of the key, and not d_model.\n\nv) Figure 1 is a very vague illustration of the connection to ODEs and provides almost no explanation in the caption. I don’t think there is much value in having this figure there.\n\nvi) There are a couple of mistakes in the paper (grammar and expressions) that should be fixed. For example, ‘the Euler’s method’ instead of ‘Euler’s method’, ‘movement in the space’ instead of ‘movement in space’, ‘dynamic system’ instead of ‘dynamical system’, ‘project parameter matrices’ instead of ‘parameter matrices’ or ‘projections’, ‘specially’ instead of ‘specifically’, etc. Please take a look at the relevant sections in the paper and revise them accordingly.\n\nvii) You explain multiple times why the proposed architecture is called a ‘Macaron Net’ (Abstract, Sec 1, Sec. 3). To avoid repetition, I would only explain it once.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "In this work, the authors show that the sequence of self-attention and feed-forward layers within a Transformer can be interpreted as an approximate numerical solution to a set of coupled ODEs. Based on this insight, the authors propose to replace the first-order Lie-Trotter splitting scheme by the more accurate, second-order Strang splitting scheme. They then present experimental results that indicate an improved performance of their Macaron Net compared to the Transformer and argue that this is due to the former being a more accurate numerical solution to the underlying set of ODEs.\n\nThe authors highlight an interesting connection between the Transformer architecture and ODEs. In particular, they derive a set of ODEs that is solved numerically by the Transformer and borrow from the body of literature on numerical ODE solvers to improve the architecture. I find that this is a very elegant and promising approach for finding better architectures. \n\nHowever, I also identified two major and a couple of minor shortcomings of the paper that are explained in detail below. Based on these shortcomings, I recommend rejecting the paper but I would be willing to increase the score if these points were addressed in sufficient detail.\n\nMajor points:\n\n1) Replacing the first-order operator splitting scheme by a second-order scheme only guarantees a lower overall truncation error if the split ODEs are solved with sufficiently high accuracy. In particular, the overall accuracy of the numerical solution to the original ODE depends on the accuracy of the operator splitting and the accuracy of the integration scheme used to solve the split ODEs (e.g. Euler’s method). The authors improve the operator splitting, i.e. they make it second-order, but they keep Euler’s method to integrate the individual ODEs. Because of that, they actually do not get rid of the lowest-order error term of the overall scheme and therefore do not obtain a more accurate ODE solver. I think this is a crucial point that invalidates the authors’ claim that the Macaron Net employs a higher-order integration scheme. As far as I am aware, this is not commented on in the paper at all. To address this shortcoming, the authors could replace Euler’s method by a second-order integrator. \n\n2) The experiments considered in this paper are interesting and show competitive performance but, in my opinion, they do not sufficiently support the claim that the Macaron Net yields a more accurate solution to the underlying set of ODEs compared to a Transformer. For a more convincing support of this claim, the authors could consider a toy problem, i.e. a simple set of ODEs with known analytical solution, and actually show that the Macaron Net is more accurate. The accuracy of a numerical ODE solver is commonly assessed by plotting the absolute difference between the exact solution (or a high-resolution numerical approximation to it) and the numerical solution vs the timestep (here \\gamma). I suspect that such an analysis would support my previous comment and show that the proposed new architecture is not more accurate ODE solver than the original one.  \n\nMinor points and questions:\n\ni) Eqs. (17-19) suggest that you apply two different FFN layers (doubling the number of parameters) instead of applying the same FFN layer twice. You comment on this in Sec. 4.1 when you say ‘..., we set the dimensionality of the inner-layer of the two FFN sub-layers in the Macaron layers to two times of the dimensionality of the hidden states’. It is not clear to me why consistency with Strang splitting requires two different layers rather than applying the same FFN layer twice. Is the reason for having a separate, trainable layer to account for the explicit time dependence of G in Eq. (16)? I think that this is a very important point that should be clarified.\n\nii) I think that this type of system is usually referred to as ‘dynamical system’ and not ‘dynamic system’. Please check that and, if applicable, update the title.  \n\niii) The authors say that Eq. (5) is a 'convection-diffusion equation’. As far as I am aware, the diffusion equation is a partial differential equation (PDE). Perhaps there is a different notion 'diffusion equation’ in ODE theory. If that’s the case, could the authors please clarify this point to avoid confusion, e.g. by adding a suitable reference in which this type of ODE is classified as a convection-diffusion equation? \n\niv) In Sec. 2 (2nd paragraph),  you cite Vaswani et al. (2017) but in that work the quantity under the square-root in the denominator of Attention(Q, K, V) is actually d_k, the dimension of the key, and not d_model.\n\nv) Figure 1 is a very vague illustration of the connection to ODEs and provides almost no explanation in the caption. I don’t think there is much value in having this figure there.\n\nvi) There are a couple of mistakes in the paper (grammar and expressions) that should be fixed. For example, ‘the Euler’s method’ instead of ‘Euler’s method’, ‘movement in the space’ instead of ‘movement in space’, ‘dynamic system’ instead of ‘dynamical system’, ‘project parameter matrices’ instead of ‘parameter matrices’ or ‘projections’, ‘specially’ instead of ‘specifically’, etc. Please take a look at the relevant sections in the paper and revise them accordingly.\n\nvii) You explain multiple times why the proposed architecture is called a ‘Macaron Net’ (Abstract, Sec 1, Sec. 3). To avoid repetition, I would only explain it once.\n"}, "tcdate": 1571239440893}, {"id": "SygNWhNWtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper136/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Contributions: This paper builds an ad-hoc connection between the Transformer and the numerical ODE solver (the Lie-Trotter splitting scheme and the Euler's method) for a convection-diffusion equation in a multi-particle dynamic system. Then, the author(s) developed an ad-hoc Strang-Marchuk splitting style architecture, named Macaron Net. Finally, this paper provides some experiments to verify the performance of the proposed architecture. However, the comparisons with the benchmark results are questionable. I have listed my concerns in the Experiment section.\n\n\nMotivation: This paper developed the Macaron Net based on a locally third-order operator splitting scheme for the convection-diffusion equation. However, there is no theoretical interpretation of why third-order splitting corresponding to better architecture. Theorem 1 in the paper is a known result, and it is irrelevant to the paper, I highly recommend the author to remove it from the main text. I also suggest the author explore more operator splitting schemes and do a systematic comparison between them. Moreover, I think it will be a real contribution if the author can analyze the error between the numerical scheme and architectures.\n\n\nReformulate Transformer Layers ans an ODE solver for Multi-Particle Dynamic System: There is a big gap between Eqns 3, 4 and 5. Why F represents a diffusion term, why G represents a convection term? From a statistical mechanics point of view, this comparison does not make sense. I do not buy this model.\n\n\nRelated Work: There is no related-work section that discusses the related work, and all the referenced papers are generic. For instance, the efforts in developing language models, the application of convection-diffusion equation, and String-Marchuck and other operator splitting schemes in machine learning. The author should better position the paper to exist work.\n\n\nExperiments: This section is extremely questionable. My initial thought after reading the reported results is that the architecture proposed in this paper easily outperforms the existing work. However, after I do a cross-check with the existing work, I found the author did not compare with the best results reported in the benchmark work and hide much information. After simply checking two existing papers, I found that the author ignored the comparison with BERT large. Also, the author ignored the most important result reported by Wu et al. 2019b.  To be fair, the author should perform an apple-to-apple comparison with the existing work and report the uncertainties in their results. Moreover, the author should report the parameters used in all their experiments.\n\n\nI think this heuristic study might be a contribution to ICLR if all my concerns are addressed, and I am willing to raise my rating to accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Contributions: This paper builds an ad-hoc connection between the Transformer and the numerical ODE solver (the Lie-Trotter splitting scheme and the Euler's method) for a convection-diffusion equation in a multi-particle dynamic system. Then, the author(s) developed an ad-hoc Strang-Marchuk splitting style architecture, named Macaron Net. Finally, this paper provides some experiments to verify the performance of the proposed architecture. However, the comparisons with the benchmark results are questionable. I have listed my concerns in the Experiment section.\n\n\nMotivation: This paper developed the Macaron Net based on a locally third-order operator splitting scheme for the convection-diffusion equation. However, there is no theoretical interpretation of why third-order splitting corresponding to better architecture. Theorem 1 in the paper is a known result, and it is irrelevant to the paper, I highly recommend the author to remove it from the main text. I also suggest the author explore more operator splitting schemes and do a systematic comparison between them. Moreover, I think it will be a real contribution if the author can analyze the error between the numerical scheme and architectures.\n\n\nReformulate Transformer Layers ans an ODE solver for Multi-Particle Dynamic System: There is a big gap between Eqns 3, 4 and 5. Why F represents a diffusion term, why G represents a convection term? From a statistical mechanics point of view, this comparison does not make sense. I do not buy this model.\n\n\nRelated Work: There is no related-work section that discusses the related work, and all the referenced papers are generic. For instance, the efforts in developing language models, the application of convection-diffusion equation, and String-Marchuck and other operator splitting schemes in machine learning. The author should better position the paper to exist work.\n\n\nExperiments: This section is extremely questionable. My initial thought after reading the reported results is that the architecture proposed in this paper easily outperforms the existing work. However, after I do a cross-check with the existing work, I found the author did not compare with the best results reported in the benchmark work and hide much information. After simply checking two existing papers, I found that the author ignored the comparison with BERT large. Also, the author ignored the most important result reported by Wu et al. 2019b.  To be fair, the author should perform an apple-to-apple comparison with the existing work and report the uncertainties in their results. Moreover, the author should report the parameters used in all their experiments.\n\n\nI think this heuristic study might be a contribution to ICLR if all my concerns are addressed, and I am willing to raise my rating to accept."}, "tcdate": 1571011579592}], "openreview_url": "https://openreview.net/forum?id=SJl1o2NFwS", "arxiv_id": "1906.02762", "paper_pdf": "papers/SJl1o2NFwS.pdf", "paper_pdf_sha256": "0bf202dffcb42105416a9ebdb94c2ee3d6a0099dbdaec4d5d6f0ed044ea58f1e", "paper_pdf_bytes": 537504, "paper_pdf_source": "openreview", "code_url": "https://github.com/zhuohan123/macaron-net", "code_repository": "zhuohan123/macaron-net", "code_commit": "3a84e7a3323bd1a3a9a303194ba336d670a1fb2c", "code_archive": "repos/SJl1o2NFwS.zip", "code_archive_sha256": "5110fccecbc2708850cc1919a3a744bcefb941ccf6ef6e161c46e31b196fd9a9", "code_archive_bytes": 466657, "code_file_count": 252, "code_extensions": {".py": 229, ".sh": 15, ".cpp": 4, ".lua": 4}, "github_disk_usage_kb": 268, "github_languages": {"Python": 1322659, "Shell": 25736, "Lua": 8860, "C++": 7164}, "github_archived": false, "github_pushed_at": "2019-06-10T09:05:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-and-improving-transformer-from"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hkl-di09FQ", "year": 2019, "status": "rejected", "title": "Decoupling feature extraction from policy learning: assessing benefits of state representation learning in goal based robotics", "authors": ["Antonin Raffin", "Ashley Hill", "René Traoré", "Timothée Lesort", "Natalia Díaz-Rodríguez", "David Filliat"], "authorids": ["antonin.raffin@ensta-paristech.fr", "ashley.hill@u-psud.fr", "krb.traore@protonmail.com", "timothee.lesort@ensta-paristech.fr", "diaz.rodriguez.natalia@gmail.com", "david.filliat@ensta-paristech.fr"], "authors_source": "OpenReview API", "abstract": "Scaling end-to-end reinforcement learning to control real robots from vision presents a series of challenges, in particular in terms of sample efficiency. Against end-to-end learning, state representation learning can help learn a compact, efficient and relevant representation of states that speeds up policy learning, reducing the number of samples needed, and that is easier to interpret. We evaluate several state representation learning methods on goal based robotics tasks and propose a new unsupervised model that stacks representations and combines strengths of several of these approaches. This method encodes all the relevant features, performs on par or better than end-to-end learning, and is robust to hyper-parameters change.", "decision": null, "meta_review": "Reject", "num_reviews": 3, "reviews": [{"id": "rklb2FSeam", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper326/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper discusses State Representation Learning for RL from camera images. Specifically, it proposes to use a state representation consisting of 2 (or 3) parts that are trained separately on different aspects of the relevant state: reward prediction, image reconstruction and (inverse) model learning. The paper is easy to read, and seems technically sound. However, the conclusions do not directly follow from the results, so should be made more precise. The contribution is minor, and the reasoning behind it could be better motivated. \n\nThe most important point of critique is that the conclusion that the split representation is the best is at best premature. The presented results indicate that SRL is useful (Table 1), and that auto-encoding alone is often not enough. Other than that, the different approaches tested all work well in different tasks. The discussion of the results reflects this, but the introduction and conclusion suggest otherwise.\n\nThe same problem also occurs for the conclusion about the robustness of SRL approaches. In the main text, no results are presented that warrant such a conclusion. The appendix includes some tests in this direction, but conclusions should not be based on material that is only available in the appendix. Furthermore, even the tests in the appendix are not comprehensive enough to to warrant the conclusion as written.\n\nThe second point is the motivation of the split approach: it seems in direct contradiction with the \"disentangled\" and \"compact\" demands the authors pose. Because the parts of the state that are needed for multiple different prediction tasks (reconstruction, inverse model, etc.) need to be in the final\nstate representation multiple times. Due to the shared feature extractor, the contradictory objectives (and hence the need for tuning of the weights in the cost function) are still a potential problem.\n\nMinor points:\n\n- The choice for these tasks is not motivated well. Please indicate why these tasks are chosen. It seems the robot arm task is very similar to the navigation task, due to robot arm's end effector being position controlled directly. Why is it worthwhile to study this task separately?\n\n- The GTC metric is not very well established (yet). Please provide some extra information on how it is calculated. This should also include some discussion on why this metric allows judging sufficiency and disentangledness. How would rotating the measurement frame of the ground-truth influence the results?\n\n- Why are the robotics priors not in Table 1?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper on SRL provides some interesting results, but it methods should be better motivated, and its conclusions be made more precise..", "review": "This paper discusses State Representation Learning for RL from camera images. Specifically, it proposes to use a state representation consisting of 2 (or 3) parts that are trained separately on different aspects of the relevant state: reward prediction, image reconstruction and (inverse) model learning. The paper is easy to read, and seems technically sound. However, the conclusions do not directly follow from the results, so should be made more precise. The contribution is minor, and the reasoning behind it could be better motivated. \n\nThe most important point of critique is that the conclusion that the split representation is the best is at best premature. The presented results indicate that SRL is useful (Table 1), and that auto-encoding alone is often not enough. Other than that, the different approaches tested all work well in different tasks. The discussion of the results reflects this, but the introduction and conclusion suggest otherwise.\n\nThe same problem also occurs for the conclusion about the robustness of SRL approaches. In the main text, no results are presented that warrant such a conclusion. The appendix includes some tests in this direction, but conclusions should not be based on material that is only available in the appendix. Furthermore, even the tests in the appendix are not comprehensive enough to to warrant the conclusion as written.\n\nThe second point is the motivation of the split approach: it seems in direct contradiction with the \"disentangled\" and \"compact\" demands the authors pose. Because the parts of the state that are needed for multiple different prediction tasks (reconstruction, inverse model, etc.) need to be in the final\nstate representation multiple times. Due to the shared feature extractor, the contradictory objectives (and hence the need for tuning of the weights in the cost function) are still a potential problem.\n\nMinor points:\n\n- The choice for these tasks is not motivated well. Please indicate why these tasks are chosen. It seems the robot arm task is very similar to the navigation task, due to robot arm's end effector being position controlled directly. Why is it worthwhile to study this task separately?\n\n- The GTC metric is not very well established (yet). Please provide some extra information on how it is calculated. This should also include some discussion on why this metric allows judging sufficiency and disentangledness. How would rotating the measurement frame of the ground-truth influence the results?\n\n- Why are the robotics priors not in Table 1?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541589417060}, {"id": "Bke6q_IRhQ", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper326/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims at comparing end-to-end learning vs separately learning a state representation and subsequently a controller.\n\nWhile this would be a relevant and important topic, the paper does not currently present consistent evidence to support this hypothesis.\n\nIn particular:\n- The approach proposed in the approach is not explained in sufficient details. After reading Sec.4 I have only a very vague and high-level idea of how the proposed approach might work. In Figure.2, what is I_t? What is the model that you are training? How are you learning this model? how do you define L_inverse?\n- The cited literature about state representation learning is absolutely incomplete. Papers like Lange et al. , Wahlström et al. and Finn et al. and citations herewithin.\n- From the experimental results, it is difficult to say anything definitive about the proposed hypothesis. 1) There are multiple end-to-end approaches in the literature, with significant differences in performance. which one are you using? (it seem A2C and PPO, but to which label do they correspond in the tables?) 2) How do you tune the weights of the reward function proposed? This seems an important design choice, but it is not much discussed. 3) In the table reported (e.g., Table 1) it does not seem to me that SRL consistently outperforms other approaches. Even for the arm tasks, Random features seem to outperform the proposed approach (and indeed all the methods except the ground truth). What is going on there?\n\nOverall¸ the paper would benefit from a clearer and more detailed text, and from improved experiments and comparisons.\n\nMinor comments:\n- It is unclear to me what \"goal-based robotic tasks\" means. How do you define a task without a goal?\n- An important and missing characteristic of a suitable state representation should be the generalization. In fact, a good representation would ideally allow the agent to generalize to some degree. \n- It seems very odd to me that the \"action should be implicitly encoded into the state representation\" could you elaborate of the motivation for this and the effects?\n\nReferences:\n- Autonomous reinforcement learning on raw visual input data in a real-world application\nS Lange, M Riedmiller, A Voigtlander\nNeural Networks (IJCNN), The 2012 International Joint Conference on, 1-8\n- From pixels to torques: Policy learning with deep dynamical models\nN Wahlström, TB Schön, MP Deisenroth\narXiv preprint arXiv:1502.02251\n-  Deep Visual Foresight for Planning Robot Motion\nChelsea Finn, Sergey Levine\nInternational Conference on Robotics and Automation (ICRA), 2017 ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear approach and contribution", "review": "This paper aims at comparing end-to-end learning vs separately learning a state representation and subsequently a controller.\n\nWhile this would be a relevant and important topic, the paper does not currently present consistent evidence to support this hypothesis.\n\nIn particular:\n- The approach proposed in the approach is not explained in sufficient details. After reading Sec.4 I have only a very vague and high-level idea of how the proposed approach might work. In Figure.2, what is I_t? What is the model that you are training? How are you learning this model? how do you define L_inverse?\n- The cited literature about state representation learning is absolutely incomplete. Papers like Lange et al. , Wahlström et al. and Finn et al. and citations herewithin.\n- From the experimental results, it is difficult to say anything definitive about the proposed hypothesis. 1) There are multiple end-to-end approaches in the literature, with significant differences in performance. which one are you using? (it seem A2C and PPO, but to which label do they correspond in the tables?) 2) How do you tune the weights of the reward function proposed? This seems an important design choice, but it is not much discussed. 3) In the table reported (e.g., Table 1) it does not seem to me that SRL consistently outperforms other approaches. Even for the arm tasks, Random features seem to outperform the proposed approach (and indeed all the methods except the ground truth). What is going on there?\n\nOverall¸ the paper would benefit from a clearer and more detailed text, and from improved experiments and comparisons.\n\nMinor comments:\n- It is unclear to me what \"goal-based robotic tasks\" means. How do you define a task without a goal?\n- An important and missing characteristic of a suitable state representation should be the generalization. In fact, a good representation would ideally allow the agent to generalize to some degree. \n- It seems very odd to me that the \"action should be implicitly encoded into the state representation\" could you elaborate of the motivation for this and the effects?\n\nReferences:\n- Autonomous reinforcement learning on raw visual input data in a real-world application\nS Lange, M Riedmiller, A Voigtlander\nNeural Networks (IJCNN), The 2012 International Joint Conference on, 1-8\n- From pixels to torques: Policy learning with deep dynamical models\nN Wahlström, TB Schön, MP Deisenroth\narXiv preprint arXiv:1502.02251\n-  Deep Visual Foresight for Planning Robot Motion\nChelsea Finn, Sergey Levine\nInternational Conference on Robotics and Automation (ICRA), 2017 ", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541462165077}, {"id": "BJg5O9zd2Q", "reviewer_signature": ["ICLR.cc/2019/Conference/Paper326/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is easy to read and the presentation is clear, and I really appreciate this.\n\nThe authors address the very important topic of feature extraction and state representation learning. New results in this area are always valuable and welcome. However, my feeling is that the paper falls short in terms of making sufficient new contributions for an ICLR paper. \n\n1. The authors propose to learn a state representation by either training using a combined loss function, or training several representations using multiple loss functions followed by stacking. These are standard and well-known techniques in machine learning. The key contribution one looks for is in terms of new insights on why and when each approach works. The paper fails to provide much insight in this regard. Take this simple scenario: Suppose my input image is actually generated by a linear map plus gaussian noise on the true states. Then I can simply use a PCA as my \"auto encoder\" and happily learn a high quality state representation close to the ground truth. We know why this works. In the real task, the image is a complex non-linear transformation of the true states. What insights do I gain from this work in terms of how I should tackle this?\n\n2. Section 3 states some desirable characteristics in constructing a state representation. These are well-known and fundamental aspects of machine learning -- applicable to almost all models that we want to learn. In this sense, I do not find the section very informative.\n\n3. The empirical results (say, Table 1) seem too noisy to interpret (other than that using the ground truth provides the best performance). It almost seems to suggest that one should simply use random features (as done in the \"extreme learning machine\" approach). Again, not much insight to draw from this.\n\n4. Last comment. Suppose I have a new robotic goal-directed task and my inputs are camera images. Does this work tell me something that I don't already know in terms of learning new feature representation that is highly suitable for my task?\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting experiment results, unfortunately lacking in terms of new insights.", "review": "The paper is easy to read and the presentation is clear, and I really appreciate this.\n\nThe authors address the very important topic of feature extraction and state representation learning. New results in this area are always valuable and welcome. However, my feeling is that the paper falls short in terms of making sufficient new contributions for an ICLR paper. \n\n1. The authors propose to learn a state representation by either training using a combined loss function, or training several representations using multiple loss functions followed by stacking. These are standard and well-known techniques in machine learning. The key contribution one looks for is in terms of new insights on why and when each approach works. The paper fails to provide much insight in this regard. Take this simple scenario: Suppose my input image is actually generated by a linear map plus gaussian noise on the true states. Then I can simply use a PCA as my \"auto encoder\" and happily learn a high quality state representation close to the ground truth. We know why this works. In the real task, the image is a complex non-linear transformation of the true states. What insights do I gain from this work in terms of how I should tackle this?\n\n2. Section 3 states some desirable characteristics in constructing a state representation. These are well-known and fundamental aspects of machine learning -- applicable to almost all models that we want to learn. In this sense, I do not find the section very informative.\n\n3. The empirical results (say, Table 1) seem too noisy to interpret (other than that using the ground truth provides the best performance). It almost seems to suggest that one should simply use random features (as done in the \"extreme learning machine\" approach). Again, not much insight to draw from this.\n\n4. Last comment. Suppose I have a new robotic goal-directed task and my inputs are camera images. Does this work tell me something that I don't already know in terms of learning new feature representation that is highly suitable for my task?\n\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1541053042344}], "openreview_url": "https://openreview.net/forum?id=Hkl-di09FQ", "arxiv_id": "1901.08651", "paper_pdf": "papers/Hkl-di09FQ.pdf", "paper_pdf_sha256": "a59b0a779281b5002b842dfea943c3405789bed729cf1d3e297f908f22c31e99", "paper_pdf_bytes": 40040956, "paper_pdf_source": "openreview", "code_url": "https://github.com/araffin/robotics-rl-srl", "code_repository": "araffin/robotics-rl-srl", "code_commit": "eae7c1ab310c79662f6e68c0d255e08641037ffa", "code_archive": "repos/Hkl-di09FQ.zip", "code_archive_sha256": "14a26443d3d5650d8fba5a8af4543a43c8a30f1df2e2aeb7e45a20518ef4cd93", "code_archive_bytes": 3891022, "code_file_count": 79, "code_extensions": {".py": 76, ".sh": 3}, "github_disk_usage_kb": 8154, "github_languages": {"Python": 471471, "Shell": 805}, "github_archived": true, "github_pushed_at": "2021-04-05T18:43:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/decoupling-feature-extraction-from-policy"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2H6KhX1kJr", "year": 2025, "status": "rejected", "title": "Transformers and slot encoding for sample efficient physical world modelling", "authors": ["Francesco Petri", "Luigi Asprino", "Aldo Gangemi"], "authorids": ["~Francesco_Petri1", "~Luigi_Asprino1", "~Aldo_Gangemi1"], "authors_source": "OpenReview API", "abstract": "World modelling, i.e. building a representation of the rules that govern the world so as to predict its evolution, is an essential ability for any agent interacting with the physical world. Recent applications of the Transformer architecture to the problem of world modelling from video input show notable improvements in sample efficiency. However, existing approaches tend to work only at the image level thus disregarding that the environment is composed of objects interacting with each other. In this paper, we propose an architecture combining Transformers for world modelling with the slot-attention paradigm, an approach for learning representations of objects appearing in a scene. We describe the resulting neural architecture and report experimental results showing an improvement over the existing solutions in terms of sample efficiency and a reduction of the variation of the performance over the training examples.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sDXtQbfFjV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7138/Reviewer_NsJS"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper present a slotted recurrent network model which uses transformers as the main backbone for \"world modeling\". In this context the resulting model is an \"object centric\" learning model which cross attends into VQ-VAE encoded input frame, updates the current state and then predicts the next state using a transformer. The model is trained for state prediction (with two variants, either next state prediction or current state prediction) and is demonstrated to mildly work better than a single external baseline (STEVE) and one ablation model (decoder only, where there is not explicit state representation, just prediction and decoding). The experiments are run on a physical reasoning task (PHYRE) and the output is a classification readout.", "review_text": "This paper present a slotted recurrent network model which uses transformers as the main backbone for \"world modeling\". In this context the resulting model is an \"object centric\" learning model which cross attends into VQ-VAE encoded input frame, updates the current state and then predicts the next state using a transformer. The model is trained for state prediction (with two variants, either next state prediction or current state prediction) and is demonstrated to mildly work better than a single external baseline (STEVE) and one ablation model (decoder only, where there is not explicit state representation, just prediction and decoding). The experiments are run on a physical reasoning task (PHYRE) and the output is a classification readout.", "strengths": "Originality:\nThe model presented is a very mild variation on previously published works - using VQ-VAE encodings is nice (though probably requires a bit more analysis) and the general recurrent setup is appealing.\n\nQuality:\nThe proposed model variants (pre and default) are interesting and probably a good step towards analyzing the model's behaviour.\n\nClarity:\nThe paper is nicely structured and well written.\n\nSignificance:\nThe context of the work is important, but see below for criticism.", "weaknesses": "Unfortunately the paper suffers from several weaknesses.\n\nExperimental validation:\nThe method is only validated on one task and even on that task results are not very convincing. The models perform very closely to one another and the claims for efficient learning with the model are not well supported.\n\nAnalysis:\nIn general I don't mind when results of a model are not competitive with baselines or ablations as long as there is good analysis of why that is the case, and how this can improve our understanding of the model or problem. Here, however, these are absent - there's very little analysis of what the model learns, how it does that and what determines its performance.\n\nPresentation:\nThe experimental result figures are not to the level I would expect to see in an ICLR paper - raw training curves are fine if they tell a clear story. Here, however, they do not - there is very little signal there to observe. Export quality is also quite low and does not at the level I would expect.\n\nNovelty:\nWhile usually I don't think novelty is a determining factor for a paper, I feel here this is quite lacking and the proposed model is indeed quite close to existing literature (SAVI++, PARTS and more ). These are cited in the paper, so I have no complaints on that side, but given the generally weak results and analysis I think this hurts the paper.", "questions": "My main question is the use of \"slot attention\" - as far as I can understand there is actually no slot attention in this model, am I right? it seems that the corrector just uses cross-attention and not slot attention? (the difference would be the soft-max axis).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper present a slotted recurrent network model which uses transformers as the main backbone for \"world modeling\". In this context the resulting model is an \"object centric\" learning model which cross attends into VQ-VAE encoded input frame, updates the current state and then predicts the next state using a transformer. The model is trained for state prediction (with two variants, either next state prediction or current state prediction) and is demonstrated to mildly work better than a single external baseline (STEVE) and one ablation model (decoder only, where there is not explicit state representation, just prediction and decoding). The experiments are run on a physical reasoning task (PHYRE) and the output is a classification readout.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Originality:\nThe model presented is a very mild variation on previously published works - using VQ-VAE encodings is nice (though probably requires a bit more analysis) and the general recurrent setup is appealing.\n\nQuality:\nThe proposed model variants (pre and default) are interesting and probably a good step towards analyzing the model's behaviour.\n\nClarity:\nThe paper is nicely structured and well written.\n\nSignificance:\nThe context of the work is important, but see below for criticism.", "weaknesses": "Unfortunately the paper suffers from several weaknesses.\n\nExperimental validation:\nThe method is only validated on one task and even on that task results are not very convincing. The models perform very closely to one another and the claims for efficient learning with the model are not well supported.\n\nAnalysis:\nIn general I don't mind when results of a model are not competitive with baselines or ablations as long as there is good analysis of why that is the case, and how this can improve our understanding of the model or problem. Here, however, these are absent - there's very little analysis of what the model learns, how it does that and what determines its performance.\n\nPresentation:\nThe experimental result figures are not to the level I would expect to see in an ICLR paper - raw training curves are fine if they tell a clear story. Here, however, they do not - there is very little signal there to observe. Export quality is also quite low and does not at the level I would expect.\n\nNovelty:\nWhile usually I don't think novelty is a determining factor for a paper, I feel here this is quite lacking and the proposed model is indeed quite close to existing literature (SAVI++, PARTS and more ). These are cited in the paper, so I have no complaints on that side, but given the generally weak results and analysis I think this hurts the paper.", "questions": "My main question is the use of \"slot attention\" - as far as I can understand there is actually no slot attention in this model, am I right? it seems that the corrector just uses cross-attention and not slot attention? (the difference would be the soft-max axis).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730644895240}, {"id": "5WAslvING0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7138/Reviewer_XiUu"], "rating": 3, "soundness": 1, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "The paper proposes the Future-Predicting Transformer Triplet (FPTT), an architecture aimed at modeling of physical world dynamics from video data. It employs Transformers to learn object-centric representations, enabling the model to predict physical interactions between objects more effectively. The architecture is tested on synthetic video dataset PHYRE. The authors also perform an ablation study to understand the contribution of different components.", "review_text": "The paper proposes the Future-Predicting Transformer Triplet (FPTT), an architecture aimed at modeling of physical world dynamics from video data. It employs Transformers to learn object-centric representations, enabling the model to predict physical interactions between objects more effectively. The architecture is tested on synthetic video dataset PHYRE. The authors also perform an ablation study to understand the contribution of different components.", "strengths": "The authors address an important problem of physical world modeling using structured latent representations.", "weaknesses": "* __Missing slot encodings and object-centricity.__ \n\nWhile the paper includes _slot encoding_ in its title, the approach itself appears to lack this feature. As I understand it, $\\Lambda$ was intended to serve as slot encodings, but it is not even referred to as such. From the description, $\\Lambda$ seems more like a standard intermediate representation in transformer layers rather than a distinct slot encoding.\n\nIn Appendix A2, the authors reference [1] for their transformer implementation, where they also mention using four slots. However, the referenced implementation does not include a parameter for the number of slots, leaving it unclear how slot encodings are actually integrated into the proposed approach, or if they are implemented at all.\n\nFurthermore, the authors state that _\"the representation remains opaque and lacks interpretability,\"_ which raises questions about the motivation for using _slot encodings_ in the first place.\n\n\n* __Experimental methodology.__ \n\nFirstly, the authors evaluate their model on a single, very simplistic dataset, while using a relatively large number of parameters. This limited evaluation setup may not provide sufficient empirical evidence to support their claims.\n\nA more significant issue lies in their positioning among related works and choice of baselines. The authors overlook most recent related work (e.g., [2, 3, 4, 5]) and rely solely on STEVE as a baseline, aside from variations of their own approach.\n\n\n* __Presentation.__ \n\nIn addition to unclear explanations of their approach and its novelty, the authors fail to position it effectively within the existing literature, lacking a comparative analysis with prior work.\n\nAll the figures also present issues: they are unnecessarily large, some are in low resolution, and it is often unclear what the authors aim to demonstrate.\n\n\\\nReferences: \n\n\n[1]: Andrej Karpathy. nanoGPT: The simplest, fastest repository for training/finetuning mediumsized GPTs (Generative Pretrained Transformers), 2023. URL https://github.com/karpathy/nanoGPT. \\\n[2]: Nakano, A., Suzuki, M. and Matsuo, Y., 2023. Interaction-based disentanglement of entities for object-centric world models. In The Eleventh International Conference on Learning Representations.\\\n[3]: Villar-Corrales, A., Wahdan, I. and Behnke, S., 2023, October. Object-centric video prediction via decoupling of object dynamics and interactions. In 2023 IEEE International Conference on Image Processing (ICIP) (pp. 570-574). IEEE.\\\n[4]: Wu, Z., Dvornik, N., Greff, K., Kipf, T. and Garg, A., 2022. Slotformer: Unsupervised visual dynamics simulation with object-centric models. arXiv preprint arXiv:2210.05861.\\\n[5]: Daniel, T. and Tamar, A., DDLP: Unsupervised Object-centric Video Prediction with Deep Dynamic Latent Particles. Transactions on Machine Learning Research.", "questions": "* __Slot encodings.__ How and where do the authors utilize slot encodings, and what specific benefits do they offer in this context?\n\n* __Experimental design.__ Why do the authors limit their evaluation to a single dataset? Additionally, what is the rationale for selecting STEVE as the sole baseline, excluding other relevant related works?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes the Future-Predicting Transformer Triplet (FPTT), an architecture aimed at modeling of physical world dynamics from video data. It employs Transformers to learn object-centric representations, enabling the model to predict physical interactions between objects more effectively. The architecture is tested on synthetic video dataset PHYRE. The authors also perform an ablation study to understand the contribution of different components.", "soundness": 1, "presentation": 1, "contribution": 1, "strengths": "The authors address an important problem of physical world modeling using structured latent representations.", "weaknesses": "* __Missing slot encodings and object-centricity.__ \n\nWhile the paper includes _slot encoding_ in its title, the approach itself appears to lack this feature. As I understand it, $\\Lambda$ was intended to serve as slot encodings, but it is not even referred to as such. From the description, $\\Lambda$ seems more like a standard intermediate representation in transformer layers rather than a distinct slot encoding.\n\nIn Appendix A2, the authors reference [1] for their transformer implementation, where they also mention using four slots. However, the referenced implementation does not include a parameter for the number of slots, leaving it unclear how slot encodings are actually integrated into the proposed approach, or if they are implemented at all.\n\nFurthermore, the authors state that _\"the representation remains opaque and lacks interpretability,\"_ which raises questions about the motivation for using _slot encodings_ in the first place.\n\n\n* __Experimental methodology.__ \n\nFirstly, the authors evaluate their model on a single, very simplistic dataset, while using a relatively large number of parameters. This limited evaluation setup may not provide sufficient empirical evidence to support their claims.\n\nA more significant issue lies in their positioning among related works and choice of baselines. The authors overlook most recent related work (e.g., [2, 3, 4, 5]) and rely solely on STEVE as a baseline, aside from variations of their own approach.\n\n\n* __Presentation.__ \n\nIn addition to unclear explanations of their approach and its novelty, the authors fail to position it effectively within the existing literature, lacking a comparative analysis with prior work.\n\nAll the figures also present issues: they are unnecessarily large, some are in low resolution, and it is often unclear what the authors aim to demonstrate.\n\n\\\nReferences: \n\n\n[1]: Andrej Karpathy. nanoGPT: The simplest, fastest repository for training/finetuning mediumsized GPTs (Generative Pretrained Transformers), 2023. URL https://github.com/karpathy/nanoGPT. \\\n[2]: Nakano, A., Suzuki, M. and Matsuo, Y., 2023. Interaction-based disentanglement of entities for object-centric world models. In The Eleventh International Conference on Learning Representations.\\\n[3]: Villar-Corrales, A., Wahdan, I. and Behnke, S., 2023, October. Object-centric video prediction via decoupling of object dynamics and interactions. In 2023 IEEE International Conference on Image Processing (ICIP) (pp. 570-574). IEEE.\\\n[4]: Wu, Z., Dvornik, N., Greff, K., Kipf, T. and Garg, A., 2022. Slotformer: Unsupervised visual dynamics simulation with object-centric models. arXiv preprint arXiv:2210.05861.\\\n[5]: Daniel, T. and Tamar, A., DDLP: Unsupervised Object-centric Video Prediction with Deep Dynamic Latent Particles. Transactions on Machine Learning Research.", "questions": "* __Slot encodings.__ How and where do the authors utilize slot encodings, and what specific benefits do they offer in this context?\n\n* __Experimental design.__ Why do the authors limit their evaluation to a single dataset? Additionally, what is the rationale for selecting STEVE as the sole baseline, excluding other relevant related works?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730564303029}, {"id": "i7lf3VsJsZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7138/Reviewer_XAqU"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper addresses the challenge of creating sample-efficient models for physical world modeling, focusing on predicting object interactions in dynamic environments.\n\nThe authors propose an architecture that combines Transformers with slot encoding to improve sample efficiency and stability in world modeling. Unlike existing models that operate at the image level, this model incorporates object-based representations, enabling it to capture and predict interactions more accurately.\n\nTheir model, named Future-Predicting Transformer Triplet (FPTT), uses a corrector-predictor-decoder triplet of Transformers. The corrector aligns the internal state representation with the actual video evolution to prevent model drift, the predictor forecasts the next state based on the corrected representation, and the decoder converts this predicted state back into tokens for further training.\n\nExperiments using the PHYRE dataset (a benchmark for physical reasoning) show that FPTT achieves greater sample efficiency and training stability compared to baseline models like STEVE. The model’s structured approach enables it to generalize well in physical environments simulated with basic Newtonian physics.\n\n\nIn summary, the paper presents a architecture that leverages the strengths of Transformers and slot encoding for efficient and stable world modeling, demonstrating improvements in tasks requiring understanding and predicting object dynamics in a physical environment​", "review_text": "The paper addresses the challenge of creating sample-efficient models for physical world modeling, focusing on predicting object interactions in dynamic environments.\n\nThe authors propose an architecture that combines Transformers with slot encoding to improve sample efficiency and stability in world modeling. Unlike existing models that operate at the image level, this model incorporates object-based representations, enabling it to capture and predict interactions more accurately.\n\nTheir model, named Future-Predicting Transformer Triplet (FPTT), uses a corrector-predictor-decoder triplet of Transformers. The corrector aligns the internal state representation with the actual video evolution to prevent model drift, the predictor forecasts the next state based on the corrected representation, and the decoder converts this predicted state back into tokens for further training.\n\nExperiments using the PHYRE dataset (a benchmark for physical reasoning) show that FPTT achieves greater sample efficiency and training stability compared to baseline models like STEVE. The model’s structured approach enables it to generalize well in physical environments simulated with basic Newtonian physics.\n\n\nIn summary, the paper presents a architecture that leverages the strengths of Transformers and slot encoding for efficient and stable world modeling, demonstrating improvements in tasks requiring understanding and predicting object dynamics in a physical environment​", "strengths": "- The paper focuses on world modeling which is an important problem.\n- The writing and presentation is clean, which makes understanding the paper easy.\n- The paper compares efficiency and accuracy, which helps understand the trade-offs.", "weaknesses": "- The contribution is not significant having an internal representation of the previous timesteps is common in world modeling architectures, for instance: Dreamer: https://arxiv.org/pdf/2301.04104, Slotformer: https://arxiv.org/pdf/2210.05861. It's unclear to me how this work is a better architecture than Dreamer or Slotformer or other recent works.\n- The evaluations and baselines are weak, the paper only compares against STEVE. I dont think STEVE is a fair comparision as their objective was to get interpretable object representations and not necessarily the metric/benchmark paper uses for evaluation. Further the Decoder-Only model seems to perform as well as the proposed architecture on almost all tasks except efficiency.\n- Lastly the work only compares on a single benchmark, which is not being used in the baseline works such as STEVE. I think a fair thing to do would be to compare on benchmarks shown in prior baselines, so we assume they are tuned well.", "questions": "- How would the architecture compare against Dreamerv3 or Slotformer in world modeling?\n- What happens if u try to make the decoder only model more efficient by reducing the number of tokens or dimensionality of the token?\n- How would the paper compare against the baselines in benchmarks proposed in Steve or DreamerV3 or SlotFormer?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of creating sample-efficient models for physical world modeling, focusing on predicting object interactions in dynamic environments.\n\nThe authors propose an architecture that combines Transformers with slot encoding to improve sample efficiency and stability in world modeling. Unlike existing models that operate at the image level, this model incorporates object-based representations, enabling it to capture and predict interactions more accurately.\n\nTheir model, named Future-Predicting Transformer Triplet (FPTT), uses a corrector-predictor-decoder triplet of Transformers. The corrector aligns the internal state representation with the actual video evolution to prevent model drift, the predictor forecasts the next state based on the corrected representation, and the decoder converts this predicted state back into tokens for further training.\n\nExperiments using the PHYRE dataset (a benchmark for physical reasoning) show that FPTT achieves greater sample efficiency and training stability compared to baseline models like STEVE. The model’s structured approach enables it to generalize well in physical environments simulated with basic Newtonian physics.\n\n\nIn summary, the paper presents a architecture that leverages the strengths of Transformers and slot encoding for efficient and stable world modeling, demonstrating improvements in tasks requiring understanding and predicting object dynamics in a physical environment​", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper focuses on world modeling which is an important problem.\n- The writing and presentation is clean, which makes understanding the paper easy.\n- The paper compares efficiency and accuracy, which helps understand the trade-offs.", "weaknesses": "- The contribution is not significant having an internal representation of the previous timesteps is common in world modeling architectures, for instance: Dreamer: https://arxiv.org/pdf/2301.04104, Slotformer: https://arxiv.org/pdf/2210.05861. It's unclear to me how this work is a better architecture than Dreamer or Slotformer or other recent works.\n- The evaluations and baselines are weak, the paper only compares against STEVE. I dont think STEVE is a fair comparision as their objective was to get interpretable object representations and not necessarily the metric/benchmark paper uses for evaluation. Further the Decoder-Only model seems to perform as well as the proposed architecture on almost all tasks except efficiency.\n- Lastly the work only compares on a single benchmark, which is not being used in the baseline works such as STEVE. I think a fair thing to do would be to compare on benchmarks shown in prior baselines, so we assume they are tuned well.", "questions": "- How would the architecture compare against Dreamerv3 or Slotformer in world modeling?\n- What happens if u try to make the decoder only model more efficient by reducing the number of tokens or dimensionality of the token?\n- How would the paper compare against the baselines in benchmarks proposed in Steve or DreamerV3 or SlotFormer?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730562576865}, {"id": "E078tNkFSq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7138/Reviewer_AfcX"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 5, "summary": "This paper proposes a world-modeling architecture that captures object-level interactions in the scene, instead of the scene itself. The architecture consists of three transformer-based models: a corrector, a predictor, and a decoder, alongside a VQ-VAE tokenizer for image encoding.\n\nTo evaluate the proposed model’s performance as a world model, the authors provide a physical reasoning task using the PHYRE benchmark, demonstrating that their model outperforms STEVE, the baseline, in terms of prediction accuracy and sample efficiency.", "review_text": "This paper proposes a world-modeling architecture that captures object-level interactions in the scene, instead of the scene itself. The architecture consists of three transformer-based models: a corrector, a predictor, and a decoder, alongside a VQ-VAE tokenizer for image encoding.\n\nTo evaluate the proposed model’s performance as a world model, the authors provide a physical reasoning task using the PHYRE benchmark, demonstrating that their model outperforms STEVE, the baseline, in terms of prediction accuracy and sample efficiency.", "strengths": "- The authors' approach of capturing slot-like internal representations from VQ tokens, instead of CNN embeddings, was intriguing and showed promising results.\n- They propose a novel evaluation protocol for testing world model architecture, utilizing the shared benchmark with different protocols.", "weaknesses": "- **Lack of Novelty and Justification**: The main idea and direction of the paper have been explored in several existing works (OCVT[1], SlotFormer[2]). Although the authors are likely aware of this, they fail to convincingly demonstrate why their approach is unique and necessary for the proposed direction, compared to previous works.\n- **Architecture and Design Choices**: The proposed architecture appears to be a combination of SAVi[3] and STEVE[4] architectures, but with a predictive loss function instead of a reconstructive loss function. While this variant may be promising, the paper lacks sufficient details to justify the design choices, such as thorough ablation studies. Furthermore, it is unclear whether the proposed model can outperform existing works, as it is not comprehensively compared.\n- **Lack of Clarity in Architecture and Experiment Description**: The architecture section of the paper lacks clarity and detail, particularly in the description of the core components: corrector transformer, predictor transformer, and decoder transformer. Although the author provides a high-level overview of these architectural concepts, the explanation is insufficient given the emphasis on this part as the paper's core contribution. To thoroughly understand and investigate the proposed architecture, a detailed formulation of these components is necessary, including their implementation details and mathematical representations.\n    \n    Furthermore, the experiment section lacks sufficient details about the metrics used in the evaluation. To ensure transparency and reproducibility, it is essential to provide a clear explanation of each metric, including how the metric is calculated and what it represents.\n    \n- **Limited Evaluation of Proposed Architecture:** The author only provides a single task to evaluate the proposed architecture, which is insufficient to demonstrate its generality and versatility. To thoroughly assess the world modeling ability of the proposed architecture, it is essential to evaluate it on a diverse range of tasks that require the model to infer and understand the relationships between objects and scenes. Additionally, to facilitate a fair comparison with existing works, the authors may consider including several generation tasks (e.g. OBJ3D[1], CLEVR[5], Physion[6]), as has been done in prior research.\n    \n    To further demonstrate the effectiveness of the proposed model, the author could compare it with a broader range of baselines, such as SlotFormer, OCVT, SAVi, and other relevant models mentioned in the paper. Although these models are typically used for generation tasks, their predicted representations can be evaluated using the same protocol as the proposed model. Additionally, comparing with image-based world models would illustrate the advantages of object-level world models over their image-based counterparts. This approach would provide a more comprehensive understanding of the proposed model's performance and allow for a more accurate assessment of its strengths and limitations relative to existing approaches.\n    \n- **Ablation Results Raise Questions about Proposed Model:** The ablation results indicate that the ‘decoder-only’ model performs comparably to the proposed models. This suggests that VQ-tokenization and predictive loss might be sufficient to drive performance without explicitly enforcing object-level representations. This outcome seems misaligned with the paper's main theme, which emphasizes the importance of object-level representations. Consequently, this misalignment raises questions about the necessity and effectiveness of the proposed model's architecture.\n\n[1] Wu, Yi-Fu, Jaesik Yoon, and Sungjin Ahn. \"Generative video transformer: Can objects be the words?.\" International Conference on Machine Learning. PMLR, 2021.\n\n[2] Wu, Ziyi, et al. \"Slotformer: Unsupervised visual dynamics simulation with object-centric models.\" arXiv preprint arXiv:2210.05861 (2022).\n\n[3] Kipf, Thomas, et al. \"Conditional object-centric learning from video.\" arXiv preprint arXiv:2111.12594 (2021).\n\n[4] Singh, Gautam, Yi-Fu Wu, and Sungjin Ahn. \"Simple unsupervised object-centric learning for complex and naturalistic videos.\" Advances in Neural Information Processing Systems 35 (2022): 18181-18196.\n\n[5] Johnson, Justin, et al. \"Clevr: A diagnostic dataset for compositional language and elementary visual reasoning.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.\n\n[6] Bear, Daniel M., et al. \"Physion: Evaluating physical prediction from vision in humans and machines.\" arXiv preprint arXiv:2106.08261 (2021).", "questions": "- It appears that slot attention (or inverted attention) is absent, with only cross attention being mentioned. Is this an oversight in the explanation, or is it indeed absent? If it's truly absent, how can we be confident that it captures object-level dynamic understanding?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a world-modeling architecture that captures object-level interactions in the scene, instead of the scene itself. The architecture consists of three transformer-based models: a corrector, a predictor, and a decoder, alongside a VQ-VAE tokenizer for image encoding.\n\nTo evaluate the proposed model’s performance as a world model, the authors provide a physical reasoning task using the PHYRE benchmark, demonstrating that their model outperforms STEVE, the baseline, in terms of prediction accuracy and sample efficiency.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "- The authors' approach of capturing slot-like internal representations from VQ tokens, instead of CNN embeddings, was intriguing and showed promising results.\n- They propose a novel evaluation protocol for testing world model architecture, utilizing the shared benchmark with different protocols.", "weaknesses": "- **Lack of Novelty and Justification**: The main idea and direction of the paper have been explored in several existing works (OCVT[1], SlotFormer[2]). Although the authors are likely aware of this, they fail to convincingly demonstrate why their approach is unique and necessary for the proposed direction, compared to previous works.\n- **Architecture and Design Choices**: The proposed architecture appears to be a combination of SAVi[3] and STEVE[4] architectures, but with a predictive loss function instead of a reconstructive loss function. While this variant may be promising, the paper lacks sufficient details to justify the design choices, such as thorough ablation studies. Furthermore, it is unclear whether the proposed model can outperform existing works, as it is not comprehensively compared.\n- **Lack of Clarity in Architecture and Experiment Description**: The architecture section of the paper lacks clarity and detail, particularly in the description of the core components: corrector transformer, predictor transformer, and decoder transformer. Although the author provides a high-level overview of these architectural concepts, the explanation is insufficient given the emphasis on this part as the paper's core contribution. To thoroughly understand and investigate the proposed architecture, a detailed formulation of these components is necessary, including their implementation details and mathematical representations.\n    \n    Furthermore, the experiment section lacks sufficient details about the metrics used in the evaluation. To ensure transparency and reproducibility, it is essential to provide a clear explanation of each metric, including how the metric is calculated and what it represents.\n    \n- **Limited Evaluation of Proposed Architecture:** The author only provides a single task to evaluate the proposed architecture, which is insufficient to demonstrate its generality and versatility. To thoroughly assess the world modeling ability of the proposed architecture, it is essential to evaluate it on a diverse range of tasks that require the model to infer and understand the relationships between objects and scenes. Additionally, to facilitate a fair comparison with existing works, the authors may consider including several generation tasks (e.g. OBJ3D[1], CLEVR[5], Physion[6]), as has been done in prior research.\n    \n    To further demonstrate the effectiveness of the proposed model, the author could compare it with a broader range of baselines, such as SlotFormer, OCVT, SAVi, and other relevant models mentioned in the paper. Although these models are typically used for generation tasks, their predicted representations can be evaluated using the same protocol as the proposed model. Additionally, comparing with image-based world models would illustrate the advantages of object-level world models over their image-based counterparts. This approach would provide a more comprehensive understanding of the proposed model's performance and allow for a more accurate assessment of its strengths and limitations relative to existing approaches.\n    \n- **Ablation Results Raise Questions about Proposed Model:** The ablation results indicate that the ‘decoder-only’ model performs comparably to the proposed models. This suggests that VQ-tokenization and predictive loss might be sufficient to drive performance without explicitly enforcing object-level representations. This outcome seems misaligned with the paper's main theme, which emphasizes the importance of object-level representations. Consequently, this misalignment raises questions about the necessity and effectiveness of the proposed model's architecture.\n\n[1] Wu, Yi-Fu, Jaesik Yoon, and Sungjin Ahn. \"Generative video transformer: Can objects be the words?.\" International Conference on Machine Learning. PMLR, 2021.\n\n[2] Wu, Ziyi, et al. \"Slotformer: Unsupervised visual dynamics simulation with object-centric models.\" arXiv preprint arXiv:2210.05861 (2022).\n\n[3] Kipf, Thomas, et al. \"Conditional object-centric learning from video.\" arXiv preprint arXiv:2111.12594 (2021).\n\n[4] Singh, Gautam, Yi-Fu Wu, and Sungjin Ahn. \"Simple unsupervised object-centric learning for complex and naturalistic videos.\" Advances in Neural Information Processing Systems 35 (2022): 18181-18196.\n\n[5] Johnson, Justin, et al. \"Clevr: A diagnostic dataset for compositional language and elementary visual reasoning.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.\n\n[6] Bear, Daniel M., et al. \"Physion: Evaluating physical prediction from vision in humans and machines.\" arXiv preprint arXiv:2106.08261 (2021).", "questions": "- It appears that slot attention (or inverted attention) is absent, with only cross attention being mentioned. Is this an oversight in the explanation, or is it indeed absent? If it's truly absent, how can we be confident that it captures object-level dynamic understanding?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729223972527}], "openreview_url": "https://openreview.net/forum?id=2H6KhX1kJr", "arxiv_id": "2405.20180", "paper_pdf": "papers/2H6KhX1kJr.pdf", "paper_pdf_sha256": "95ff9dc6e4249f8bf4051aa6ce747ee78fababf5328527b2922b754877ccd834", "paper_pdf_bytes": 792919, "paper_pdf_source": "openreview", "code_url": "https://github.com/torchipeppo/transformers-and-slot-encoding-for-wm", "code_repository": "torchipeppo/transformers-and-slot-encoding-for-wm", "code_commit": "7ea6e923a3c616610ee4af243f09f1b69e5aeb86", "code_archive": "repos/2H6KhX1kJr.zip", "code_archive_sha256": "746ea5726bd9b5d2293f214ac537ff9a3f32848968e941fa4cffe66edf739935", "code_archive_bytes": 133480, "code_file_count": 55, "code_extensions": {".py": 55}, "github_disk_usage_kb": 93, "github_languages": {"Python": 211262}, "github_archived": false, "github_pushed_at": "2024-05-31T13:50:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transformers-and-slot-encoding-for-sample"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vxmvbzw76R", "year": 2024, "status": "rejected", "title": "Split-and-Denoise: Protect large language model inference with local differential privacy", "authors": ["Peihua Mai", "Ran Yan", "Zhe Huang", "Youjia Yang", "Yan Pang"], "authorids": ["~Peihua_Mai1", "~Ran_Yan3", "~Zhe_Huang6", "~Youjia_Yang1", "~Yan_Pang1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) shows powerful capability in natural language understanding by capturing hidden semantics in vector space. This process enriches the value of the text embeddings for various downstream tasks, thereby fostering the Embedding-as-a-Service (EaaS) business model. However, the direct transmission of text to servers poses a largely unaddressed risk of privacy leakage. To mitigate this issue, we introduce Split-N-Denoise (SnD), an innovative framework that split the model to execute the token embedding layer on the client side at minimal computational cost. This allows the client to introduce noise prior to transmitting the embeddings to the server and subsequently receive and denoise the perturbed output embeddings for downstream tasks. Our approach is designed for the inference stage of LLMs and requires no modifications to the model parameters, while also being computationally efficient on the client side. Extensive experiments demonstrate SnD's effectiveness in optimizing the privacy-utility tradeoff across various LLM architectures and diverse downstream tasks. The results reveal an significant accuracy improvement under the same privacy budget compared to the baseline, offering clients a privacy-preserving solution for local privacy protection.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "gPNPM1ahBH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5188/Reviewer_4SyG"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper focuses on the problem of privacy-preserving LLM inference in a setting where clients input text and a server holds the model. To address the privacy risk of direct transmission of clients' text to server, this paper splits an LLM between clients and a server such that:\n\n1) Local computation: the client performs affordable computation locally to obtain intermediate results. In particular, only the token embedding is done on the client side to keep the computational cost low for clients. \n\n2) Privatising clients submissions: Differential privacy is employed to mitigate privacy leakage by injecting noises into the embedding before sharing with the server. In particular, each client adds noise to their embedding prior to sending them to the server to protect the privacy of clients while doing LLM inference;\n\n3) Server-side computation: the server receives the noisy embedding, and performs the rest of the computations of the LLM and returns the noisy output to the client.\n\n4) Client-side denoising: Each client performs denoising to improve the utility of the output. The denoise model is pre-trained on the server side using public datasets and synthetic noises, and subsequently shared with the client.", "review_text": "This paper focuses on the problem of privacy-preserving LLM inference in a setting where clients input text and a server holds the model. To address the privacy risk of direct transmission of clients' text to server, this paper splits an LLM between clients and a server such that:\n\n1) Local computation: the client performs affordable computation locally to obtain intermediate results. In particular, only the token embedding is done on the client side to keep the computational cost low for clients. \n\n2) Privatising clients submissions: Differential privacy is employed to mitigate privacy leakage by injecting noises into the embedding before sharing with the server. In particular, each client adds noise to their embedding prior to sending them to the server to protect the privacy of clients while doing LLM inference;\n\n3) Server-side computation: the server receives the noisy embedding, and performs the rest of the computations of the LLM and returns the noisy output to the client.\n\n4) Client-side denoising: Each client performs denoising to improve the utility of the output. The denoise model is pre-trained on the server side using public datasets and synthetic noises, and subsequently shared with the client.", "strengths": "1) As opposed to the server-side denoising that has been used in the existing work, this paper performs the denoising at the client side to leverage the knowledge of noise levels and raw embedding. \n\n2) Evaluation of the proposed method through computing similarity between the clean and privatized embeddings, and performance on downstream tasks of sentence classification, pair similarity, Recognizing Textual Entailment\n\n3) The problem of privacy-preserving LLM inference which is studied by this paper is an important problem as clients may input sensitive information, such as names, phones, and email addresses, that needs to be kept hidden from the service provider.", "weaknesses": "1) No evidence to support the practicality of the proposed method. This paper instantiates the denoising model as an L layer transformer-based model that receives the privatised token representations, noise matrix and noisy output computed by the server. Although this denoise model is pre-trained on the server side using public datasets and synthetic noises, each client needs to do the inference of this denoising model locally. Unfortunately, this paper does not empirically evaluate the memory and computational cost of this inference for clients. Therefore, it is not clear if this overhead is any smaller than running the whole LLM on the client side. This evaluation is necessary as one main claim of this paper is that the proposed method introduces only affordable local computations for clients.\n\n2) No evidence demonstrating the privacy benefits of the proposed method. This paper lacks an empirical evaluation of the privacy leakage of the proposed method. This is particularly important as analytical privacy guarantees chosen by this paper are very loose for example see privacy budgets of 100, 500 and 1000 in the Tables provided in the experiment section.\n\n\n3) Shallow discussion of results and not considering SOTA models. \n\n4) Experimental choices including hyperparameters and privacy budgets are not justified/studied and they are not consistent across models. The only statement regarding the choice of the privacy budget that I can see in the paper is the following: \"For the three model families, we selected three distinct eta levels for experimentation, given the varying noise tolerance of each model. Specifically, for the Bert models, we set eta to 50, 100, and 500; for the GPT models, the values were 1, 100, and 1000; and for the T5 models, we chose 0.1, 1, and 10.\"\n\n\n5) There are many typos:\n  1) Missing \".\" at the end of captions\n  2) Consequently, As a result,: As --> as\n  3) known as ”embedding as a service”: ”embedding --> ``embedding\n  4) this paper represents the pioneering effort in protect user’s: protect --> protecting\n  5) Fact Kotonya & Toni (2020),Daily Dialogue: ,Daily --> , Daily", "questions": "I would recommend reporting the computational costs and privacy benefits of the proposed method. Please see my first and second concerns in the weakness box.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the problem of privacy-preserving LLM inference in a setting where clients input text and a server holds the model. To address the privacy risk of direct transmission of clients' text to server, this paper splits an LLM between clients and a server such that:\n\n1) Local computation: the client performs affordable computation locally to obtain intermediate results. In particular, only the token embedding is done on the client side to keep the computational cost low for clients. \n\n2) Privatising clients submissions: Differential privacy is employed to mitigate privacy leakage by injecting noises into the embedding before sharing with the server. In particular, each client adds noise to their embedding prior to sending them to the server to protect the privacy of clients while doing LLM inference;\n\n3) Server-side computation: the server receives the noisy embedding, and performs the rest of the computations of the LLM and returns the noisy output to the client.\n\n4) Client-side denoising: Each client performs denoising to improve the utility of the output. The denoise model is pre-trained on the server side using public datasets and synthetic noises, and subsequently shared with the client.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1) As opposed to the server-side denoising that has been used in the existing work, this paper performs the denoising at the client side to leverage the knowledge of noise levels and raw embedding. \n\n2) Evaluation of the proposed method through computing similarity between the clean and privatized embeddings, and performance on downstream tasks of sentence classification, pair similarity, Recognizing Textual Entailment\n\n3) The problem of privacy-preserving LLM inference which is studied by this paper is an important problem as clients may input sensitive information, such as names, phones, and email addresses, that needs to be kept hidden from the service provider.", "weaknesses": "1) No evidence to support the practicality of the proposed method. This paper instantiates the denoising model as an L layer transformer-based model that receives the privatised token representations, noise matrix and noisy output computed by the server. Although this denoise model is pre-trained on the server side using public datasets and synthetic noises, each client needs to do the inference of this denoising model locally. Unfortunately, this paper does not empirically evaluate the memory and computational cost of this inference for clients. Therefore, it is not clear if this overhead is any smaller than running the whole LLM on the client side. This evaluation is necessary as one main claim of this paper is that the proposed method introduces only affordable local computations for clients.\n\n2) No evidence demonstrating the privacy benefits of the proposed method. This paper lacks an empirical evaluation of the privacy leakage of the proposed method. This is particularly important as analytical privacy guarantees chosen by this paper are very loose for example see privacy budgets of 100, 500 and 1000 in the Tables provided in the experiment section.\n\n\n3) Shallow discussion of results and not considering SOTA models. \n\n4) Experimental choices including hyperparameters and privacy budgets are not justified/studied and they are not consistent across models. The only statement regarding the choice of the privacy budget that I can see in the paper is the following: \"For the three model families, we selected three distinct eta levels for experimentation, given the varying noise tolerance of each model. Specifically, for the Bert models, we set eta to 50, 100, and 500; for the GPT models, the values were 1, 100, and 1000; and for the T5 models, we chose 0.1, 1, and 10.\"\n\n\n5) There are many typos:\n  1) Missing \".\" at the end of captions\n  2) Consequently, As a result,: As --> as\n  3) known as ”embedding as a service”: ”embedding --> ``embedding\n  4) this paper represents the pioneering effort in protect user’s: protect --> protecting\n  5) Fact Kotonya & Toni (2020),Daily Dialogue: ,Daily --> , Daily", "questions": "I would recommend reporting the computational costs and privacy benefits of the proposed method. Please see my first and second concerns in the weakness box.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699203879242}, {"id": "K4Y0tZ4ekS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5188/Reviewer_xFRn"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors introduce Split-N-Denoise (SnD), a privacy-preserving scheme for LLMs.  SnD makes use of split learning, wherein the network is chopped in two and each half is distributed to the client and server, respectively.  SnD splits at the embedding layer, where local differential privacy (LDP) is applied to user embedding vectors for obfuscation.  Along with the embedding layer, the client side also contains a trained denoiser which, after the output embedding vectors are transmitted from the server, is used to denoise the response.  Several experiments are conducted demonstrating the efficacy of this work, compared to another method designed specifically for BERT (i.e., TokEmbPriv from Qu et al, 2021), across several models (BERT/DistillBert, T5, and GPT-2) for three classification tasks.", "review_text": "The authors introduce Split-N-Denoise (SnD), a privacy-preserving scheme for LLMs.  SnD makes use of split learning, wherein the network is chopped in two and each half is distributed to the client and server, respectively.  SnD splits at the embedding layer, where local differential privacy (LDP) is applied to user embedding vectors for obfuscation.  Along with the embedding layer, the client side also contains a trained denoiser which, after the output embedding vectors are transmitted from the server, is used to denoise the response.  Several experiments are conducted demonstrating the efficacy of this work, compared to another method designed specifically for BERT (i.e., TokEmbPriv from Qu et al, 2021), across several models (BERT/DistillBert, T5, and GPT-2) for three classification tasks.", "strengths": "The problem of how to preserve-privacy in embedding-as-a-service applications is important with the influx of interest in LLMs.  Furthermore, the authors use of LPD provides certifiable privacy and, in the case of split learning, LDP is a very nontrivial task.  The presented framework makes sense in this context, although the overall approach could be better motivated.  For instance, why split at the embedding layer?  Splitting at this particular layer is the most unsecure, even for a simple man in the middle attack (given a pretrained foundation model).  However, splitting at any other layer and transmitting gradients naturally allows for federated learning strategies and the avoidance of DP altogether.  Furthermore, multi-party computation (MPC) methods do not have the encountered problem of trying to use DP with split learning (and properly denoising transmitted data).  A contrast and discussion of these various approaches is warranted.  Please see the following for a recent privacy-preserving MPC method:\nKnott, Brian, et al. \"Crypten: Secure multi-party computation meets machine learning.\" Advances in Neural Information Processing Systems 34 (2021): 4961-4973", "weaknesses": "# Autoregressive LLM concerns regarding the (only) use of GPT-2 and evaluation \n\nThe featured classification tasks do well to test the performance of BERT and T5.  However, they do not relevantly test the performance of GPT-2; in practice, GPT-2 does not perform well on classification tasks, which would ideally be handled by more appropriate encoder (e.g., BERT) or encoder-decoder (e.g., T5) architectures.  More relevant generative metrics, such as sentence completion (e.g., via HellaSwag), next-word-completion (e.g., via the lambada dataset), or even perplexity are required to show that GPT-2 performance is maintained.  Furthermore, - GPT-2 itself is an older model and was trained in a much less complicated manner compared to recently released (instruction-tuned) models, such as LLaMa-2, Falcon, Mistral, etc..  Testing on a more recent architecture is important to show the efficacy of this approach, to show that the more sophisticated pretrained knowledge (which naturally contains many more learned modalities/instructions compared to GPT-2) still behaves as expected, and to show different architectural choices are unaffected by the introduced noise (e.g., LLaMa uses RMSNorm vs GPT-2's Layernorm, which, relevant to the presented work, directly impacts how these models deal with noise variation in the data).\n\n# Lack of: (a) comparison to other relevant privacy-preserving methods, (b) evidence against competitors infeasibility\n\nBeyond TokEmbPriv, comparison is lacking to other relevant benchmark competitors.  E.g., due to the similarity of both approaches, a comparison to RAPT (Li et al. (2023)) is necessary.  Furethermore, for homomorphic encryption approaches, the authors claim:\n> Cryptographic typically employs homomorphic encryption (HE) to compute the inference result of the users’ encrypted input. Unfortunately, the application of cryptographic technique is constrained by the significant computation overhead of cryptographic operations, especially on large transformer models.\n\nTo assert this claim, the proposed method and reference HE methods should be compared in terms of both accuracy, privacy-preserving ability, and wall-clock time.\n\n# Inaccurate claims\n> To the best of our knowledge, this paper represents the pioneering effort in protect user’s privacy during LLM inference with strong privacy guarantee. Existing research focuses on the privacy-preserving pre-training and fine-tuning for LLM, while few studies pay attention to the privacy concerns at inference stage, especially on privatizing user’s input to guarantee DP.\n\nThe homomorphic encrpytion work by Liu & Liu (2023); Chen et al. (2022) (cited in the paper) protects user privacy during inference, please withdraw or appropriately revise this claim.\n\nFurthermore, the description of RAPT (which itself offers LDP for privacy-preserving LLM inference) is inaccurate:\n> An alternative strategy might entail input text perturbation via textto-text privatization or synthetic data generation, preserving high-dimensional features while altering human-perceivable sequences Li et al. (2023). Specifically, the text\\\n-to-text privatization projects text into a high-dimensional vector space with a pre-determined word embedding model,adding\ncarefully calibrated noise to the vector representation, and then reconvert it to obtain the perturbed\ntext Feyisetan et al. (2019); Qu et al. (2021a). Yet, the mere application of this technique during\ninference does not guarantee a satisfactory balance between privacy and utility\n\nThis is a misleading description of the RAPT method from Li et al. (2023); RAPT performs LDP to user prompts (thus establishing certifiable privacy via DP).  In order to recover performance on the noisy data, RAPT employs prompt tuning to efficiently fine-tune a server-side model capable of performing inference on the LDP data.  This approach is extremely similar to the proposed split-and-denoise framework.  Please clearly describe RAPT and contrast it to SnP, while also benchmarking against it as a relevant competitor.\n\n> Split learning Gupta & Raskar (2018); Vepakomma et al. (2018) ... DP is employed to mitigate privacy leakage by injecting noises into the IRs before sharing with the server.\n\nNeither of the two papers use DP to certifiably protect the data.  Please cite another multi-party computation paper (where intermediate gradients are distributed across networks) which uses DP\n\n# Lack of demonstrated efficacy against attacks\n\nIt is necessasry to prove the privacy-preserving capabilities of the presented approach, e.g., simulate eavesdropping attacks, which collect intercepted embedding vectors (during transmission to the server) for malicious actions (e.g., embedding inversion and attribute inference attacks).  Please see Li et al. (2023) for examples and more details.", "questions": "> We design a novel denoising method deployed on user side. In this approach, a denoise\nmodel is pre-trained on server side using public dataset and synthetic noises. Subsequently,\nthis trained model is deployed on the user side, where it leverages the specific noise levels\nand raw IRs provided by the user to enhance the embeddings.\n\nCan the authors comment on how this necessarily opens up a large security hole? I.e., the data has noise injected to protect it in the event of interception by a bad actor.  However, this scheme requires the transmission of the actual denoising layer, which itself may be intercepted.  Due to the need for model refreshes, this problem is non-trivial.\n\n> Split learning is a novel privacy-preserving approach in distributed learning\n \nPlease remove \"novel\", as it not being introduced in the presented work.\n\nThe benchmarked method, \"TokEmbPriv,\" requires significantly more discussion during the background and previous work sections.\n\nThe paper requires an editing pass.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce Split-N-Denoise (SnD), a privacy-preserving scheme for LLMs.  SnD makes use of split learning, wherein the network is chopped in two and each half is distributed to the client and server, respectively.  SnD splits at the embedding layer, where local differential privacy (LDP) is applied to user embedding vectors for obfuscation.  Along with the embedding layer, the client side also contains a trained denoiser which, after the output embedding vectors are transmitted from the server, is used to denoise the response.  Several experiments are conducted demonstrating the efficacy of this work, compared to another method designed specifically for BERT (i.e., TokEmbPriv from Qu et al, 2021), across several models (BERT/DistillBert, T5, and GPT-2) for three classification tasks.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "The problem of how to preserve-privacy in embedding-as-a-service applications is important with the influx of interest in LLMs.  Furthermore, the authors use of LPD provides certifiable privacy and, in the case of split learning, LDP is a very nontrivial task.  The presented framework makes sense in this context, although the overall approach could be better motivated.  For instance, why split at the embedding layer?  Splitting at this particular layer is the most unsecure, even for a simple man in the middle attack (given a pretrained foundation model).  However, splitting at any other layer and transmitting gradients naturally allows for federated learning strategies and the avoidance of DP altogether.  Furthermore, multi-party computation (MPC) methods do not have the encountered problem of trying to use DP with split learning (and properly denoising transmitted data).  A contrast and discussion of these various approaches is warranted.  Please see the following for a recent privacy-preserving MPC method:\nKnott, Brian, et al. \"Crypten: Secure multi-party computation meets machine learning.\" Advances in Neural Information Processing Systems 34 (2021): 4961-4973", "weaknesses": "# Autoregressive LLM concerns regarding the (only) use of GPT-2 and evaluation \n\nThe featured classification tasks do well to test the performance of BERT and T5.  However, they do not relevantly test the performance of GPT-2; in practice, GPT-2 does not perform well on classification tasks, which would ideally be handled by more appropriate encoder (e.g., BERT) or encoder-decoder (e.g., T5) architectures.  More relevant generative metrics, such as sentence completion (e.g., via HellaSwag), next-word-completion (e.g., via the lambada dataset), or even perplexity are required to show that GPT-2 performance is maintained.  Furthermore, - GPT-2 itself is an older model and was trained in a much less complicated manner compared to recently released (instruction-tuned) models, such as LLaMa-2, Falcon, Mistral, etc..  Testing on a more recent architecture is important to show the efficacy of this approach, to show that the more sophisticated pretrained knowledge (which naturally contains many more learned modalities/instructions compared to GPT-2) still behaves as expected, and to show different architectural choices are unaffected by the introduced noise (e.g., LLaMa uses RMSNorm vs GPT-2's Layernorm, which, relevant to the presented work, directly impacts how these models deal with noise variation in the data).\n\n# Lack of: (a) comparison to other relevant privacy-preserving methods, (b) evidence against competitors infeasibility\n\nBeyond TokEmbPriv, comparison is lacking to other relevant benchmark competitors.  E.g., due to the similarity of both approaches, a comparison to RAPT (Li et al. (2023)) is necessary.  Furethermore, for homomorphic encryption approaches, the authors claim:\n> Cryptographic typically employs homomorphic encryption (HE) to compute the inference result of the users’ encrypted input. Unfortunately, the application of cryptographic technique is constrained by the significant computation overhead of cryptographic operations, especially on large transformer models.\n\nTo assert this claim, the proposed method and reference HE methods should be compared in terms of both accuracy, privacy-preserving ability, and wall-clock time.\n\n# Inaccurate claims\n> To the best of our knowledge, this paper represents the pioneering effort in protect user’s privacy during LLM inference with strong privacy guarantee. Existing research focuses on the privacy-preserving pre-training and fine-tuning for LLM, while few studies pay attention to the privacy concerns at inference stage, especially on privatizing user’s input to guarantee DP.\n\nThe homomorphic encrpytion work by Liu & Liu (2023); Chen et al. (2022) (cited in the paper) protects user privacy during inference, please withdraw or appropriately revise this claim.\n\nFurthermore, the description of RAPT (which itself offers LDP for privacy-preserving LLM inference) is inaccurate:\n> An alternative strategy might entail input text perturbation via textto-text privatization or synthetic data generation, preserving high-dimensional features while altering human-perceivable sequences Li et al. (2023). Specifically, the text\\\n-to-text privatization projects text into a high-dimensional vector space with a pre-determined word embedding model,adding\ncarefully calibrated noise to the vector representation, and then reconvert it to obtain the perturbed\ntext Feyisetan et al. (2019); Qu et al. (2021a). Yet, the mere application of this technique during\ninference does not guarantee a satisfactory balance between privacy and utility\n\nThis is a misleading description of the RAPT method from Li et al. (2023); RAPT performs LDP to user prompts (thus establishing certifiable privacy via DP).  In order to recover performance on the noisy data, RAPT employs prompt tuning to efficiently fine-tune a server-side model capable of performing inference on the LDP data.  This approach is extremely similar to the proposed split-and-denoise framework.  Please clearly describe RAPT and contrast it to SnP, while also benchmarking against it as a relevant competitor.\n\n> Split learning Gupta & Raskar (2018); Vepakomma et al. (2018) ... DP is employed to mitigate privacy leakage by injecting noises into the IRs before sharing with the server.\n\nNeither of the two papers use DP to certifiably protect the data.  Please cite another multi-party computation paper (where intermediate gradients are distributed across networks) which uses DP\n\n# Lack of demonstrated efficacy against attacks\n\nIt is necessasry to prove the privacy-preserving capabilities of the presented approach, e.g., simulate eavesdropping attacks, which collect intercepted embedding vectors (during transmission to the server) for malicious actions (e.g., embedding inversion and attribute inference attacks).  Please see Li et al. (2023) for examples and more details.", "questions": "> We design a novel denoising method deployed on user side. In this approach, a denoise\nmodel is pre-trained on server side using public dataset and synthetic noises. Subsequently,\nthis trained model is deployed on the user side, where it leverages the specific noise levels\nand raw IRs provided by the user to enhance the embeddings.\n\nCan the authors comment on how this necessarily opens up a large security hole? I.e., the data has noise injected to protect it in the event of interception by a bad actor.  However, this scheme requires the transmission of the actual denoising layer, which itself may be intercepted.  Due to the need for model refreshes, this problem is non-trivial.\n\n> Split learning is a novel privacy-preserving approach in distributed learning\n \nPlease remove \"novel\", as it not being introduced in the presented work.\n\nThe benchmarked method, \"TokEmbPriv,\" requires significantly more discussion during the background and previous work sections.\n\nThe paper requires an editing pass.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698880525990}, {"id": "gKOeHjX4J5", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5188/Reviewer_C4Af"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes an approach of implementing local differential privacy to protect LLM inference. Specifically, a framework, SnD, is designed to enable the client to introduce noise prior to transmitting the embeddings to the server, and then denoise the output embeddings received from server, in a Embedding-as-a-Service scenario.", "review_text": "The paper proposes an approach of implementing local differential privacy to protect LLM inference. Specifically, a framework, SnD, is designed to enable the client to introduce noise prior to transmitting the embeddings to the server, and then denoise the output embeddings received from server, in a Embedding-as-a-Service scenario.", "strengths": "+ The study focuses on an interesting and important topic, the privacy in LLM.\n+ The Embedding-as-a-Service business model is well-defined.", "weaknesses": "- The overhead of proposed approach is not clear\n\nMy first concern pertains to the overhead introduced by the local encoder. While a basic complexity analysis is presented in Section 3.5, I would recommend a more comprehensive evaluation involving specific experiments. Such an evaluation is crucial to assessing the feasibility of the proposed approach. Furthermore, it remains unclear how the denoise model, pre-trained on the server, can adapt to varying noise levels determined by clients.\n\n- Lack of details on baseline description\n\nThere is a lack of detail in the baseline description. The only information provided about the benchmark method is \"where the token embeddings are perturbed by the user before sending them to the server\". It would be beneficial to provide more information about TokEmbPriv and explain why this method was chosen as the sole benchmark in the evaluation. Additionally, it would be helpful to clarify why other methods, such as a vanilla model, were not considered in the comparison.\n\n- The evaluation on denoise is somehow blur\n\nThe evaluation of the denoising process is somewhat unclear. Please explicitly state and explain why a higher Cosine Similarity score between the initial and recovered embeddings is indicative of better performance. Also, please provide a clear definition of what \"initial embeddings\" refer to. If they pertain to the embeddings input to the denoise module, i.e., the \"output noised results\" in Figure 1, it currently reads as though having fewer differences between the initial and recovered embeddings is preferred, which might lead to less noise added. Additionally, the evaluation appears to focus solely on accuracy and does not provide results on privacy protection. Given that the primary motivation of the study is addressing the \"unaddressed risk of privacy leakage,\" the privacy performance of the proposed approach is expected and should be included in the evaluation.", "questions": "1. What is the overhead of the proposed approach?\n2. Why TokEmbPriv is chosen as the sole benchmark in the evaluation?\n3. What is the privacy performance of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an approach of implementing local differential privacy to protect LLM inference. Specifically, a framework, SnD, is designed to enable the client to introduce noise prior to transmitting the embeddings to the server, and then denoise the output embeddings received from server, in a Embedding-as-a-Service scenario.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "+ The study focuses on an interesting and important topic, the privacy in LLM.\n+ The Embedding-as-a-Service business model is well-defined.", "weaknesses": "- The overhead of proposed approach is not clear\n\nMy first concern pertains to the overhead introduced by the local encoder. While a basic complexity analysis is presented in Section 3.5, I would recommend a more comprehensive evaluation involving specific experiments. Such an evaluation is crucial to assessing the feasibility of the proposed approach. Furthermore, it remains unclear how the denoise model, pre-trained on the server, can adapt to varying noise levels determined by clients.\n\n- Lack of details on baseline description\n\nThere is a lack of detail in the baseline description. The only information provided about the benchmark method is \"where the token embeddings are perturbed by the user before sending them to the server\". It would be beneficial to provide more information about TokEmbPriv and explain why this method was chosen as the sole benchmark in the evaluation. Additionally, it would be helpful to clarify why other methods, such as a vanilla model, were not considered in the comparison.\n\n- The evaluation on denoise is somehow blur\n\nThe evaluation of the denoising process is somewhat unclear. Please explicitly state and explain why a higher Cosine Similarity score between the initial and recovered embeddings is indicative of better performance. Also, please provide a clear definition of what \"initial embeddings\" refer to. If they pertain to the embeddings input to the denoise module, i.e., the \"output noised results\" in Figure 1, it currently reads as though having fewer differences between the initial and recovered embeddings is preferred, which might lead to less noise added. Additionally, the evaluation appears to focus solely on accuracy and does not provide results on privacy protection. Given that the primary motivation of the study is addressing the \"unaddressed risk of privacy leakage,\" the privacy performance of the proposed approach is expected and should be included in the evaluation.", "questions": "1. What is the overhead of the proposed approach?\n2. Why TokEmbPriv is chosen as the sole benchmark in the evaluation?\n3. What is the privacy performance of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698716218422}, {"id": "moRy9f3PbS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5188/Reviewer_NdAW"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a framework for achieving private Large Language Model (LLM) inference. The framework combines dx-privacy-based relaxed local Differential Privacy (DP), U-shape split inference, and a client-side denoiser based on the Transformer architecture. This innovative approach delivers efficient computation, privacy guarantees, and strong utility for private LLM inference. The paper demonstrates the effectiveness of the proposed framework using four datasets and three different architectural configurations. The results illustrate that the model maintains high utility while preserving a substantial privacy budget, with minimal computational complexity overhead.", "review_text": "This paper introduces a framework for achieving private Large Language Model (LLM) inference. The framework combines dx-privacy-based relaxed local Differential Privacy (DP), U-shape split inference, and a client-side denoiser based on the Transformer architecture. This innovative approach delivers efficient computation, privacy guarantees, and strong utility for private LLM inference. The paper demonstrates the effectiveness of the proposed framework using four datasets and three different architectural configurations. The results illustrate that the model maintains high utility while preserving a substantial privacy budget, with minimal computational complexity overhead.", "strengths": "1. The proposed work intelligently integrates split inference, relaxed local DP, and client-side post-processing with a denoiser.\n2. The location of the denoising model at the client side, which leverages knowledge of the noise level, is crucial for effective denoising.\n3. The paper provides a well-analyzed discussion of complexity.\n4. The potential for several extension works pointed out in this work is interesting.", "weaknesses": "1. Recent work by \"Mattern et al. - The Limits of Word Level Differential Privacy\" has pointed out limitations of dx-privacy. The author should address these weaknesses and concerns related to dx-privacy in the paper.\n2. The denoise model is pre-trained and known by the server. Could the server utilize the Denoise model to obtain less noisy user embeddings and potentially compromise privacy? This issue should be explored.\n3. The impact of knowing the noise level or not should be more thoroughly analyzed.\n4. The paper does not adequately demonstrate the privacy aspects regarding textual input. There is a lack of evidence that the proposed method can effectively defend against potential inversion attacks compared to an unprotected scheme.\n5. The paper creates some confusion due to the mixed usage of \"eta\" in both its symbol and word forms.\n6. There is a typographical error: \"under varying η in table 4.3.1, 4.3.1 and 4.3.1.\"", "questions": "1. How does the Performance on selected benchmarks compare to SoTA?\n\n2. What does eta mean for privacy? Why can it be set differently for GPT, Bert, and T5 models?\n\n3. Can this be extended to training (finetuning)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a framework for achieving private Large Language Model (LLM) inference. The framework combines dx-privacy-based relaxed local Differential Privacy (DP), U-shape split inference, and a client-side denoiser based on the Transformer architecture. This innovative approach delivers efficient computation, privacy guarantees, and strong utility for private LLM inference. The paper demonstrates the effectiveness of the proposed framework using four datasets and three different architectural configurations. The results illustrate that the model maintains high utility while preserving a substantial privacy budget, with minimal computational complexity overhead.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The proposed work intelligently integrates split inference, relaxed local DP, and client-side post-processing with a denoiser.\n2. The location of the denoising model at the client side, which leverages knowledge of the noise level, is crucial for effective denoising.\n3. The paper provides a well-analyzed discussion of complexity.\n4. The potential for several extension works pointed out in this work is interesting.", "weaknesses": "1. Recent work by \"Mattern et al. - The Limits of Word Level Differential Privacy\" has pointed out limitations of dx-privacy. The author should address these weaknesses and concerns related to dx-privacy in the paper.\n2. The denoise model is pre-trained and known by the server. Could the server utilize the Denoise model to obtain less noisy user embeddings and potentially compromise privacy? This issue should be explored.\n3. The impact of knowing the noise level or not should be more thoroughly analyzed.\n4. The paper does not adequately demonstrate the privacy aspects regarding textual input. There is a lack of evidence that the proposed method can effectively defend against potential inversion attacks compared to an unprotected scheme.\n5. The paper creates some confusion due to the mixed usage of \"eta\" in both its symbol and word forms.\n6. There is a typographical error: \"under varying η in table 4.3.1, 4.3.1 and 4.3.1.\"", "questions": "1. How does the Performance on selected benchmarks compare to SoTA?\n\n2. What does eta mean for privacy? Why can it be set differently for GPT, Bert, and T5 models?\n\n3. Can this be extended to training (finetuning)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698711659343}], "openreview_url": "https://openreview.net/forum?id=vxmvbzw76R", "arxiv_id": "2310.09130", "paper_pdf": "papers/vxmvbzw76R.pdf", "paper_pdf_sha256": "8b6c49af013842fa8a35aba3ca20320987361b97db1fb54d13dc08f17a724ed4", "paper_pdf_bytes": 1450395, "paper_pdf_source": "openreview", "code_url": "https://github.com/NusIoraPrivacy/eaas-privacy", "code_repository": "NusIoraPrivacy/eaas-privacy", "code_commit": "156dc7b3db3892948b39da3c4a098e09328af019", "code_archive": "repos/vxmvbzw76R.zip", "code_archive_sha256": "85aef26d0510907c76ec53cbc333a2af49d1f55808f37d2b2212d2176196b26d", "code_archive_bytes": 88252, "code_file_count": 27, "code_extensions": {".py": 26, ".sh": 1}, "github_disk_usage_kb": 76, "github_languages": {"Python": 237361, "Shell": 1578}, "github_archived": false, "github_pushed_at": "2024-05-25T02:17:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/split-and-denoise-protect-large-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MR8pqi9R7xP", "year": 2023, "status": "rejected", "title": "Watch What You Pretrain For: Targeted, Transferable Adversarial Examples on Self-Supervised Speech Recognition models", "authors": ["Raphael Olivier", "Hadi Abdullah", "Bhiksha Raj"], "authorids": ["~Raphael_Olivier1", "~Hadi_Abdullah1", "~Bhiksha_Raj1"], "authors_source": "OpenReview API", "abstract": "A targeted adversarial attack produces audio samples that can force an Automatic Speech Recognition (ASR) system to output attacker-chosen text. To exploit ASR models in real-world, black-box settings, an adversary can leverage the transferability property, i.e. that an adversarial sample produced for a proxy ASR can also fool a different remote ASR. However recent work has shown that transferability against large ASR models is very difficult. In this work, we show that modern ASR architectures, specifically ones based on Self-Supervised Learning, are in fact vulnerable to transferability. We successfully demonstrate this phenomenon by evaluating state-of-the-art self-supervised ASR models like Wav2Vec2, HuBERT, Data2Vec and WavLM. We show that with low-level additive noise achieving a 30dB Signal-Noise Ratio, we can achieve target transferability with up to 80\\% accuracy. Next, we  1) use an ablation study to show that Self-Supervised learning is the main cause of that phenomenon, and 2) we provide an explanation for this phenomenon. Through this we show that modern ASR architectures are uniquely vulnerable to adversarial security threats.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "aG3EAePs8GZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1235/Reviewer_xki8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper investigate the transferability property of adversarial attacks on ASR models. They evaluate the robustness of modern transformer-based ASR architectures such as Wav2Vec-2.0 etc. With a series of experiments on a set of ASR models by using a set of 85 adversarial samples, they show that many state-of-the-art ASR models are in fact vulnerable to the transferability property. Additionally, they also claim to show that SSL-pretraining is the reason for this vulnerability to transferability. \n\n", "review_text": "Overall the paper targets an interesting problem. At the beginning the paper sounded very interesting based the claims but as I went through the paper I could not understand some experiments or reasoning behind some experiment criteria. Therefore, I won’t feel confident to vote for acceptance of this paper.  ", "strengths": "\nThe approach for generating adversarial examples is interesting but it is not explained very well in section 3.1 as the details are missing such as: \nWere only 85 samples used during fine-tuning the models?\nWhy 85 and how were these selected? \nIt is not clear how the third model (Data2Vec BASE) was used as stopping criteria? \nResults are presented in section 3.4, but before that I have not seen how the adversarial examples are generated. \n\n\n\nIn section 4, why character error rate (CER) was used instead of WER? CER might not be indicating the significant impact on WER as these could have been fixed by LM duding decoding (if there was one added). I guess CER was not used in the previous section due to the same reason. So it would be good to see these results in terms of WER. \nThis section in particular was difficult to understand. There should be some more details added for clarity. For example the following sentence is not at all clear: \"we observe that the Character Error Rate between two random sentences is about 80-85% on average. Therefore attack success rates higher than 20% indicate a partially successful attack”. How is CER computed between two samples? Is it not between the reference and hypothesis? How 20% is selected as indicator for successful attack?\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper investigate the transferability property of adversarial attacks on ASR models. They evaluate the robustness of modern transformer-based ASR architectures such as Wav2Vec-2.0 etc. With a series of experiments on a set of ASR models by using a set of 85 adversarial samples, they show that many state-of-the-art ASR models are in fact vulnerable to the transferability property. Additionally, they also claim to show that SSL-pretraining is the reason for this vulnerability to transferability. \n\n", "strength_and_weaknesses": "\nThe approach for generating adversarial examples is interesting but it is not explained very well in section 3.1 as the details are missing such as: \nWere only 85 samples used during fine-tuning the models?\nWhy 85 and how were these selected? \nIt is not clear how the third model (Data2Vec BASE) was used as stopping criteria? \nResults are presented in section 3.4, but before that I have not seen how the adversarial examples are generated. \n\n\n\nIn section 4, why character error rate (CER) was used instead of WER? CER might not be indicating the significant impact on WER as these could have been fixed by LM duding decoding (if there was one added). I guess CER was not used in the previous section due to the same reason. So it would be good to see these results in terms of WER. \nThis section in particular was difficult to understand. There should be some more details added for clarity. For example the following sentence is not at all clear: \"we observe that the Character Error Rate between two random sentences is about 80-85% on average. Therefore attack success rates higher than 20% indicate a partially successful attack”. How is CER computed between two samples? Is it not between the reference and hypothesis? How 20% is selected as indicator for successful attack?\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Study is very interesting but not novel. Many experiemnts are not clearly explained, and reproducibility of these experiments might not be easy without all details. ", "summary_of_the_review": "Overall the paper targets an interesting problem. At the beginning the paper sounded very interesting based the claims but as I went through the paper I could not understand some experiments or reasoning behind some experiment criteria. Therefore, I won’t feel confident to vote for acceptance of this paper.  ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667165668959}, {"id": "KBNbMqTy6J_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1235/Reviewer_5s8w"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes targeted and transferable adversarial examples for self-supervised asr models which are in pretraining + fine-tuning architecture. An adversary can make use of the transferability property, that is, an adversarial sample produced for a proxy asr can also fool a different remote asr.\nRich self-supervised asr models, such as wav2vec2, Hubert, data2vec and wavlm are investigated and similar results are shown with detailed experiments and comparisons.\n", "review_text": "Generally, the paper is well written with a good motivation and rich experiments on detailed evidences. One issue is that the investigation of the datasets used for training these sota ssl models and also among different languages. \nIf with one solution to dealing with the adversarial examples, this paper can be ranked higher.\n", "strengths": "Strong:\n\n1 The targeted and transferable adversarial examples are interesting and should benefit the research ad developing of better and more robust asr systems of pretraining+fine-tuning architecture.\n\n2 Rich existing sota models are investigated showing the “transferability” is a frequent thing.\n\n3 Code is attached making the paper to be with high score of reproducing ability. \n\nWeak:\n\n1 Prefer to see richer experiments on language models using quite different datasets of among different languages to better learn the transferability among pretrained speech representations;\n\n2 If solutions to these well-designed adversarial examples can be provided, this paper will be rich of novel solutions as well.\n\n\n\nDetailed questions and comments\n\n1 So, how to improve current self-supervised models to better dealing with the transferable adversarial examples? This is more valuable for building robust asr systems.\n\n2 Can your asr attacking strategies influencing other people’s normal usage of existing asr systems? Say, you prepared a special group of inputs and all asr systems failed – then does this influence other people’s usage or if same types of attacking data were used for training these models, then they will fail forever if not fixed. Basing on my experience, most asr systems are still quite fragile – they fail a lot even with quite clean and clear voice inputs and they will for sure fail if the inputs are further including rich carefully designed noises.\n\n3 Any further comparison of the transferability of from one language’s pretrained model to another language’s pretrained model? Say how large the datasets used are involved in the transferability?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes targeted and transferable adversarial examples for self-supervised asr models which are in pretraining + fine-tuning architecture. An adversary can make use of the transferability property, that is, an adversarial sample produced for a proxy asr can also fool a different remote asr.\nRich self-supervised asr models, such as wav2vec2, Hubert, data2vec and wavlm are investigated and similar results are shown with detailed experiments and comparisons.\n", "strength_and_weaknesses": "Strong:\n\n1 The targeted and transferable adversarial examples are interesting and should benefit the research ad developing of better and more robust asr systems of pretraining+fine-tuning architecture.\n\n2 Rich existing sota models are investigated showing the “transferability” is a frequent thing.\n\n3 Code is attached making the paper to be with high score of reproducing ability. \n\nWeak:\n\n1 Prefer to see richer experiments on language models using quite different datasets of among different languages to better learn the transferability among pretrained speech representations;\n\n2 If solutions to these well-designed adversarial examples can be provided, this paper will be rich of novel solutions as well.\n\n\n\nDetailed questions and comments\n\n1 So, how to improve current self-supervised models to better dealing with the transferable adversarial examples? This is more valuable for building robust asr systems.\n\n2 Can your asr attacking strategies influencing other people’s normal usage of existing asr systems? Say, you prepared a special group of inputs and all asr systems failed – then does this influence other people’s usage or if same types of attacking data were used for training these models, then they will fail forever if not fixed. Basing on my experience, most asr systems are still quite fragile – they fail a lot even with quite clean and clear voice inputs and they will for sure fail if the inputs are further including rich carefully designed noises.\n\n3 Any further comparison of the transferability of from one language’s pretrained model to another language’s pretrained model? Say how large the datasets used are involved in the transferability?\n\n", "clarity,_quality,_novelty_and_reproducibility": "this paper is clearly written with code attachment - a high reproducibility.\nThe idea of transferability adversarial examples is interesting. \nThis paper can be scored higher if with solutions to the adversarial examples (or investigating existing anti-attach methods)", "summary_of_the_review": "Generally, the paper is well written with a good motivation and rich experiments on detailed evidences. One issue is that the investigation of the datasets used for training these sota ssl models and also among different languages. \nIf with one solution to dealing with the adversarial examples, this paper can be ranked higher.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666610725393}, {"id": "hA-SS2sgLKU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1235/Reviewer_fvn3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies transferable targeted adversarial attack on self-supervised ASR models. Target attack adds a small perturbation to a model such that the model makes the targeted prediction desired by the attacker instead of the correct one corresponding to the original input. Transferability refers to generalizing the attack to private models which have not been used to generate the adversarial perturbation.\n\nPast studies show that targeted attack is hard to generalize for supervised ASR models. However, in contrast to previous findings, the authors demonstrate that such attack can in fact generalize to ASR models pre-trained with self-supervised learning using similar datasets. The authors then present a series of study to understand what leads to successful attack transfer, including self-supervised objective, models size, training data size, and specificity of the attack.\n", "review_text": "While the paper can be improved as mentioned in the sections above, the current presentation already provides substance that will be useful and interesting to the community.", "strengths": "- Strengths\n  - The authors presented an interesting study with new observations contrary to previous studies of adversarial attacks on ASR models. This could open up a new direction of empirical and theoretical study to understand robustness in self-supervised setups\n  - Many factors are studied to reason and hypothesize why self-supervised ASR models are more vulnerable to targeted adversarial attack, providing valuable data points for future work.\n  - Attacks of different levels of specificity have also been considered to support the hypothesis.\n  - A more generalizable scheme for optimizing the adversarial attack is proposed in this paper, which uses multiple proxy models and validates/selects checkpoints with a different model.\n\n- Weaknesses\n  - While the authors have compared several different SSL models (wav2vec, hubert, data2vec, WavLM, UniSpeech), they all have very similar architectures (convolution encoder for waveform followed by transformer layers). The argument of “attack on SSL models are transferable when they are pre-trained on the same dataset” would be more convincing if the authors had considered another SSL that have a different model architecture (e.g., LSTM).\n  - The experiments studying the effect of pre-training data could have been expanded. The authors showed in Table 1 that target attack performance is worse on W2V2-Large (CV), which is an interesting observation. The study will more complete if the authors include a) optimizing attack on CV and transfer to LV; b) optimizing attack on a model pre-trained on more datasets (CV+LV+Fisher checkpoint is available on Github) and transfer to one pre-trained on a subset of it; c) the reverse of b.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies transferable targeted adversarial attack on self-supervised ASR models. Target attack adds a small perturbation to a model such that the model makes the targeted prediction desired by the attacker instead of the correct one corresponding to the original input. Transferability refers to generalizing the attack to private models which have not been used to generate the adversarial perturbation.\n\nPast studies show that targeted attack is hard to generalize for supervised ASR models. However, in contrast to previous findings, the authors demonstrate that such attack can in fact generalize to ASR models pre-trained with self-supervised learning using similar datasets. The authors then present a series of study to understand what leads to successful attack transfer, including self-supervised objective, models size, training data size, and specificity of the attack.\n", "strength_and_weaknesses": "- Strengths\n  - The authors presented an interesting study with new observations contrary to previous studies of adversarial attacks on ASR models. This could open up a new direction of empirical and theoretical study to understand robustness in self-supervised setups\n  - Many factors are studied to reason and hypothesize why self-supervised ASR models are more vulnerable to targeted adversarial attack, providing valuable data points for future work.\n  - Attacks of different levels of specificity have also been considered to support the hypothesis.\n  - A more generalizable scheme for optimizing the adversarial attack is proposed in this paper, which uses multiple proxy models and validates/selects checkpoints with a different model.\n\n- Weaknesses\n  - While the authors have compared several different SSL models (wav2vec, hubert, data2vec, WavLM, UniSpeech), they all have very similar architectures (convolution encoder for waveform followed by transformer layers). The argument of “attack on SSL models are transferable when they are pre-trained on the same dataset” would be more convincing if the authors had considered another SSL that have a different model architecture (e.g., LSTM).\n  - The experiments studying the effect of pre-training data could have been expanded. The authors showed in Table 1 that target attack performance is worse on W2V2-Large (CV), which is an interesting observation. The study will more complete if the authors include a) optimizing attack on CV and transfer to LV; b) optimizing attack on a model pre-trained on more datasets (CV+LV+Fisher checkpoint is available on Github) and transfer to one pre-trained on a subset of it; c) the reverse of b.\n", "clarity,_quality,_novelty_and_reproducibility": "The topic studied in this paper is novel. The paper is easy to follow and well stated. Experimental details are provided to facilitate reproduction.\n", "summary_of_the_review": "While the paper can be improved as mentioned in the sections above, the current presentation already provides substance that will be useful and interesting to the community.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666564279617}, {"id": "2V9AfSDvvP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1235/Reviewer_6Mf8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper shows that recent self-supervised ASR model are uniquely vulnerable to black-box adversarial attacks, which is an interesting observation and can inspire future research in this field. ", "review_text": "The paper observes an interesting result. However, the explanation to this phenomenon is unclear. It seems that it is possible that the SSL based model are vulnerable than non SSL-based model mainly because the attack model is trained to fool a SSL model.", "strengths": "Strength:\n1. The paper shows that recent self-supervised ASR model are uniquely vulnerable to black-box adversarial attacks.\n2. The paper conducts an ablation study to show that self-Supervised learning is the main cause of that phenomenon.\n3. The paper provides an explanation for this phenomenon.\n\nWeakness:\n\nThe experiment settings in this paper are inconsistent and confusing. I summarize it with the following question. Although I rate a relative low score for this paper, I am willing to change my score if the issues are addressed.\n\n1. In Section 3.4, the results show that SSL-based model is vulnerable to black-box adversarial attacks. However, the attack model is also trained to fool a SSL based models. What if the attack model is trained to fool a non-SSL based model such as the bottom three models in Table 1?\n2. The experiments in Section 3 use an improved version of that mention in Section 2.2. However, in Section 4 only the default version of Section 2.2 is utilized. What is the motivation for this inconsistency?\n3. (A question similar to the first question) In Section 4.1, it seems that the attack model is also trained to fool a wav2vec2-based model. In this way, it is natural that the wav2vec2-based model is more likely to be fooled compared with the non-wav2vec2 model.\n4. In Section 5.1 and 5.2, the conclusion is that a hard targeted attack is hard to achieve compared with a mildly targeted attack, which is an interesting but natural result. However, the intuition of why this result can explain the vulnerable SSL-based ASR model is not discussed in detail.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper shows that recent self-supervised ASR model are uniquely vulnerable to black-box adversarial attacks, which is an interesting observation and can inspire future research in this field. ", "strength_and_weaknesses": "Strength:\n1. The paper shows that recent self-supervised ASR model are uniquely vulnerable to black-box adversarial attacks.\n2. The paper conducts an ablation study to show that self-Supervised learning is the main cause of that phenomenon.\n3. The paper provides an explanation for this phenomenon.\n\nWeakness:\n\nThe experiment settings in this paper are inconsistent and confusing. I summarize it with the following question. Although I rate a relative low score for this paper, I am willing to change my score if the issues are addressed.\n\n1. In Section 3.4, the results show that SSL-based model is vulnerable to black-box adversarial attacks. However, the attack model is also trained to fool a SSL based models. What if the attack model is trained to fool a non-SSL based model such as the bottom three models in Table 1?\n2. The experiments in Section 3 use an improved version of that mention in Section 2.2. However, in Section 4 only the default version of Section 2.2 is utilized. What is the motivation for this inconsistency?\n3. (A question similar to the first question) In Section 4.1, it seems that the attack model is also trained to fool a wav2vec2-based model. In this way, it is natural that the wav2vec2-based model is more likely to be fooled compared with the non-wav2vec2 model.\n4. In Section 5.1 and 5.2, the conclusion is that a hard targeted attack is hard to achieve compared with a mildly targeted attack, which is an interesting but natural result. However, the intuition of why this result can explain the vulnerable SSL-based ASR model is not discussed in detail.\n", "clarity,_quality,_novelty_and_reproducibility": "The phenomenon reported in this paper is interesting.", "summary_of_the_review": "The paper observes an interesting result. However, the explanation to this phenomenon is unclear. It seems that it is possible that the SSL based model are vulnerable than non SSL-based model mainly because the attack model is trained to fool a SSL model.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666496005887}], "openreview_url": "https://openreview.net/forum?id=MR8pqi9R7xP", "arxiv_id": "2209.13523", "paper_pdf": "papers/MR8pqi9R7xP.pdf", "paper_pdf_sha256": "204d3e68259bec3510af00883f2006374f6ce59afa63da697f68e792a078e28c", "paper_pdf_bytes": 444708, "paper_pdf_source": "openreview", "code_url": "https://github.com/RaphaelOlivier/asr_transferability", "code_repository": "RaphaelOlivier/asr_transferability", "code_commit": "458c193839ecbc818f8f5674f046e878f6ceba3e", "code_archive": "repos/MR8pqi9R7xP.zip", "code_archive_sha256": "7ef0385007c03d4bf771a828ca0598f22dfe63255a4f532222af76f6157af17f", "code_archive_bytes": 129337, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 149, "github_languages": {"Python": 53349}, "github_archived": false, "github_pushed_at": "2025-04-24T21:34:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/watch-what-you-pretrain-for-targeted"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Lwclw6u3Pcw", "year": 2022, "status": "rejected", "title": "Characterizing and Measuring the Similarity of Neural Networks with Persistent Homology ", "authors": ["David Pérez Fernández", "Asier Gutiérrez-Fandiño", "Jordi Armengol-Estapé", "Marta Villegas"], "authorids": ["~David_Pérez_Fernández1", "~Asier_Gutiérrez-Fandiño1", "~Jordi_Armengol-Estapé1", "~Marta_Villegas2"], "authors_source": "OpenReview API", "abstract": "Characterizing the structural properties of neural networks is crucial yet poorly understood, and there are no well-established similarity measures between networks. In this work, we observe that neural networks can be represented as abstract simplicial complex and analyzed using their topological 'fingerprints' via Persistent Homology (PH). We then describe a PH-based representation proposed for characterizing and measuring similarity of neural networks. We empirically show the effectiveness of this representation as a descriptor of different architectures in several datasets. This approach based on Topological Data Analysis is a step towards better understanding neural networks and a useful similarity measure.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "6jkzn9v-7Bd", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4098/Reviewer_ZGgm"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper converts a given NN to a weighted directed graph and consider the flag complex on the top of it and use that object to compute the PD of the input NN. Two neural networks are then considered to be similar iff the PDs are close enough with respect to WD. ", "review_text": "I think the paper is interesting and novel. The idea of converting the NN that way to PD is something I have not seen before. However, I have many concerns about the paper.\n\nNeural networks are functions, and it it is not clear if two neural networks that have different architectures should be not similar. This method suffers from this in my opinion. You might have two neural networks that are totally different, one of them is giant and one of them is small and yet they are similar in the function they do--this will not be captured by the the method presented--at least this is not justified.\n\n- also, more importantly, why do you want to measure similarity between two NNs? I think it is interesting mathematical purpose but it is not clear from a practical perspective why you want or need to do that?  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper converts a given NN to a weighted directed graph and consider the flag complex on the top of it and use that object to compute the PD of the input NN. Two neural networks are then considered to be similar iff the PDs are close enough with respect to WD. ", "main_review": "I think the paper is interesting and novel. The idea of converting the NN that way to PD is something I have not seen before. However, I have many concerns about the paper.\n\nNeural networks are functions, and it it is not clear if two neural networks that have different architectures should be not similar. This method suffers from this in my opinion. You might have two neural networks that are totally different, one of them is giant and one of them is small and yet they are similar in the function they do--this will not be captured by the the method presented--at least this is not justified.\n\n- also, more importantly, why do you want to measure similarity between two NNs? I think it is interesting mathematical purpose but it is not clear from a practical perspective why you want or need to do that?  ", "summary_of_the_review": "While the paper is interesting and offers fresh point of view, I feel the paper lacks justification of important parts : such as why do we need a similarity between two NNs.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636002275312}, {"id": "TdNzaOdCREF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4098/Reviewer_rtBj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this work, the authors propose to characterize neural networks with Topological Data Analysis, more precisely with its main descriptor, the so-called persistence diagram, in order to be able to compare neural networks with different numbers of layers, different numbers of neurons, or trained on different data sets. More specifically, they show that the computational graphs corresponding to the neural networks can be filtered using the edge weights that are learnt during training, in order to produce filtered flag complexes, from which persistent homology can be computed. Then, the authors interpreted the distances between the persistence diagrams obtained from networks with varying parameters (number of layers, neurons, labels), and showed that, often, the distances have intuitive correlation with the complexity of the networks.\n", "review_text": "The article is well written and proposes an interesting approach, but I have some trouble figuring out how powerful and meaningful the method actually is, due to the lack of theoretical back-ups and the vague, hand-waving interpretations of the set of experiments. \n\nI have the following comments:\n\na. I understand that thresholding persistence diagrams is important for keeping the running time reasonable, but, on the other hand, setting the lifespan threshold to 0.01 seems pretty arbitrary. It would be nice to comment about how this threshold was chosen in the text.\n\nb. It is quite difficult to interpret the proposed distance matrices and values. Increasing and decreasing values of distances can be caused by many factors, and not only the ones suggested by the authors. For instance, simply increasing the size of the graphs sometimes results in larger distances just because the persistence diagrams have more points, invalidating some of the authors interpretation. In order to be fully convincing, I think some null statistical model should be provided and compared to, with some corresponding p-values. See for instance https://openreview.net/pdf?id=rHaiOtGdRS. Otherwise, it is impossible to go beyond vague comments and explanations of the results.\n\nc. I am not sure about how useful are these distances. Even though they seem to detect a few properties of the network, they do not seem to always be very discriminative: some architectures with very different hyper parameters coming from different experiment groups can end up quite close in the proposed distance. Indeed, values that are outside of the blocks corresponding to the various experiments can still be very small in the distance matrices. Hence, I am not sure if there are settings in which using this distance actually make sense, practically speaking. Would it be possible, for instance, to do model selection based on it?\n\nd. Topological uncertainty (https://www.ijcai.org/proceedings/2021/367), is a good reference to add in the related work section, since it also aims at characterizing the shape of neural nets with persistence (for another application though).\n\n[Post rebuttal comment] Even though I appreciate the author's responses and suggestions, I still think that the paper requires substantial improvements before publication, so I did not change my grade.\n   ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this work, the authors propose to characterize neural networks with Topological Data Analysis, more precisely with its main descriptor, the so-called persistence diagram, in order to be able to compare neural networks with different numbers of layers, different numbers of neurons, or trained on different data sets. More specifically, they show that the computational graphs corresponding to the neural networks can be filtered using the edge weights that are learnt during training, in order to produce filtered flag complexes, from which persistent homology can be computed. Then, the authors interpreted the distances between the persistence diagrams obtained from networks with varying parameters (number of layers, neurons, labels), and showed that, often, the distances have intuitive correlation with the complexity of the networks.\n", "main_review": "The article is well written and proposes an interesting approach, but I have some trouble figuring out how powerful and meaningful the method actually is, due to the lack of theoretical back-ups and the vague, hand-waving interpretations of the set of experiments. \n\nI have the following comments:\n\na. I understand that thresholding persistence diagrams is important for keeping the running time reasonable, but, on the other hand, setting the lifespan threshold to 0.01 seems pretty arbitrary. It would be nice to comment about how this threshold was chosen in the text.\n\nb. It is quite difficult to interpret the proposed distance matrices and values. Increasing and decreasing values of distances can be caused by many factors, and not only the ones suggested by the authors. For instance, simply increasing the size of the graphs sometimes results in larger distances just because the persistence diagrams have more points, invalidating some of the authors interpretation. In order to be fully convincing, I think some null statistical model should be provided and compared to, with some corresponding p-values. See for instance https://openreview.net/pdf?id=rHaiOtGdRS. Otherwise, it is impossible to go beyond vague comments and explanations of the results.\n\nc. I am not sure about how useful are these distances. Even though they seem to detect a few properties of the network, they do not seem to always be very discriminative: some architectures with very different hyper parameters coming from different experiment groups can end up quite close in the proposed distance. Indeed, values that are outside of the blocks corresponding to the various experiments can still be very small in the distance matrices. Hence, I am not sure if there are settings in which using this distance actually make sense, practically speaking. Would it be possible, for instance, to do model selection based on it?\n\nd. Topological uncertainty (https://www.ijcai.org/proceedings/2021/367), is a good reference to add in the related work section, since it also aims at characterizing the shape of neural nets with persistence (for another application though).\n\n[Post rebuttal comment] Even though I appreciate the author's responses and suggestions, I still think that the paper requires substantial improvements before publication, so I did not change my grade.\n   ", "summary_of_the_review": "While the approach is novel and interesting, I think the experiments are not convincing enough and too hand-waving to definitely validate the procedure. Since the approach is purely experimental, I think the work is too preliminary for publication.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635899618349}, {"id": "9MdbKRFTGfX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4098/Reviewer_HwgX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors give a method to evaluate the closeness of a task considered by a neural network. They represent a trained NN as a weighted graph and extract features of the NN by calculating persistent homology from the graph.\nThe distance between NNs is calculated by calculating the distance between persistent homologies, and it is experimentally shown that the corresponding tasks of each NN can be determined whether they are similar or different.", "review_text": "Strength\n- This paper shows that the distance between graphs calculation using TDA is more suitable for comparing tasks than the general method of calculating the distance between graphs.\n- Constructed a framework for comparing trained NNs and demonstrated its effectiveness for simple methods.\n\nWeakness\n- As mentioned in the Related work section of this paper, Rieck et.al.(2019b) have shown a method for representing NNs as graphs and analyzing them with TDA. On the other hand, when the intrinsic graph structures of NNs are similar, they correspond to almost the same task, and when the tasks are completely different, the NN structures are often very different.（It is possible that the structure of NNs will be similar for different tasks, but I don’t think we need to worry about it practically.）The proposed method is based on the fact that NNs for similar tasks have similar graph structures, and TDA is good at analyzing graph structures. Therefore, the idea of the proposed method is not particularly new.\n- The structure of a trained NN depends on the training data, loss function, and optimization algorithm. Even if the NN is for the same task, the structure of the NN is often different if they are different. In this study, verification under different conditions is lacking, and we cannot say that the proposed method measures whether the NNs are for the same task or not. However, it is worthwhile as a comparison under the same conditions, so it will be a sufficient contribution if specific applications are shown.\nMinor comment\nTable captions are usually written above the table.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors give a method to evaluate the closeness of a task considered by a neural network. They represent a trained NN as a weighted graph and extract features of the NN by calculating persistent homology from the graph.\nThe distance between NNs is calculated by calculating the distance between persistent homologies, and it is experimentally shown that the corresponding tasks of each NN can be determined whether they are similar or different.", "main_review": "Strength\n- This paper shows that the distance between graphs calculation using TDA is more suitable for comparing tasks than the general method of calculating the distance between graphs.\n- Constructed a framework for comparing trained NNs and demonstrated its effectiveness for simple methods.\n\nWeakness\n- As mentioned in the Related work section of this paper, Rieck et.al.(2019b) have shown a method for representing NNs as graphs and analyzing them with TDA. On the other hand, when the intrinsic graph structures of NNs are similar, they correspond to almost the same task, and when the tasks are completely different, the NN structures are often very different.（It is possible that the structure of NNs will be similar for different tasks, but I don’t think we need to worry about it practically.）The proposed method is based on the fact that NNs for similar tasks have similar graph structures, and TDA is good at analyzing graph structures. Therefore, the idea of the proposed method is not particularly new.\n- The structure of a trained NN depends on the training data, loss function, and optimization algorithm. Even if the NN is for the same task, the structure of the NN is often different if they are different. In this study, verification under different conditions is lacking, and we cannot say that the proposed method measures whether the NNs are for the same task or not. However, it is worthwhile as a comparison under the same conditions, so it will be a sufficient contribution if specific applications are shown.\nMinor comment\nTable captions are usually written above the table.\n", "summary_of_the_review": "The contribution to the proposal and validation of a more effective framework for comparing trained NNs is acknowledged. On the other hand, the novelty of the idea is low and its applicability is questionable.\n\n[Post rebuttal comment] I appreciate the author's responses. The author's perception was clear to me, but I think it still needs to be improved with additional validation, so I did not change my grade.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635765042952}, {"id": "xRStNhchQ_0", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4098/Reviewer_WhHV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In order to characterize and measure the similarity of two neural networks, this paper uses abstract simplicial complex to represent neural networks. The Persistent Homology of the constructed graph associated to a neural network is computed to obtain the corresponding Persistence Diagram. To calculate and quantify the similarity, the authors use supported vectorized persistence summaries: Persistence landscape, Weighted silhouette, and Heat vectorizations respectively and compare their ability to measure the similarities between neural networks. Detailed and complete experiments are conducted on different kinds of datasets. Each experiment only contains one modification on hyperparameter.  Results show that the PH-based representation did characterize and capture latent information from the networks.", "review_text": "## Strength\n1.\tThis paper presents a new approach to represent neural networks which includes abstract simplicial complex representation and PH method. It is novel and insightful to combine neural networks similarity measure and topological methods. Especially, using PD discretization to measure the similarity is intuitive and reasonable.\n2.\tThis paper is well-supported theoretically. Both the topological characterizations associated to neural networks and PH methods for measuring similarity are technically sound and significant for the problem studied.\n3.\tThis paper is well written and easy to understand. Although I think the display of experimental results could be arranged more reasonably (see weaknesses 1)\n\n## Weakness\n1.\tAs mentioned in the introduction, a correct characterization should be able to distinguish whether two neural networks are trained for the same task, regardless of subtle differences in architecture, input order, etc. However, discussion about how PD discretization recognizes the different types of tasks better than 1-norm and Frobenius norm (similar to figure3), not only in the control experiments, is not clear and sufficient.\n2.\tVery few details are given about the calculation of PD discretization, i.e., Persistence landscape, Weighted silhouette and Heat vectorizations. Their calculation and topological meaning may explain partly why Heat and Silhouette separate the experimental group better than Landscape, which is worth exploring as well.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In order to characterize and measure the similarity of two neural networks, this paper uses abstract simplicial complex to represent neural networks. The Persistent Homology of the constructed graph associated to a neural network is computed to obtain the corresponding Persistence Diagram. To calculate and quantify the similarity, the authors use supported vectorized persistence summaries: Persistence landscape, Weighted silhouette, and Heat vectorizations respectively and compare their ability to measure the similarities between neural networks. Detailed and complete experiments are conducted on different kinds of datasets. Each experiment only contains one modification on hyperparameter.  Results show that the PH-based representation did characterize and capture latent information from the networks.", "main_review": "## Strength\n1.\tThis paper presents a new approach to represent neural networks which includes abstract simplicial complex representation and PH method. It is novel and insightful to combine neural networks similarity measure and topological methods. Especially, using PD discretization to measure the similarity is intuitive and reasonable.\n2.\tThis paper is well-supported theoretically. Both the topological characterizations associated to neural networks and PH methods for measuring similarity are technically sound and significant for the problem studied.\n3.\tThis paper is well written and easy to understand. Although I think the display of experimental results could be arranged more reasonably (see weaknesses 1)\n\n## Weakness\n1.\tAs mentioned in the introduction, a correct characterization should be able to distinguish whether two neural networks are trained for the same task, regardless of subtle differences in architecture, input order, etc. However, discussion about how PD discretization recognizes the different types of tasks better than 1-norm and Frobenius norm (similar to figure3), not only in the control experiments, is not clear and sufficient.\n2.\tVery few details are given about the calculation of PD discretization, i.e., Persistence landscape, Weighted silhouette and Heat vectorizations. Their calculation and topological meaning may explain partly why Heat and Silhouette separate the experimental group better than Landscape, which is worth exploring as well.\n\n", "summary_of_the_review": "The motivation and the problem studied in this paper is interesting. The authors present a novel and effective approach to represent neural networks in a topological way and propose a similarity measure using PD discretization, which is supported by the experimental results. However, the discussion part could be more focused on the discretization’s ability to separate the experiment groups, rather than sensitivity to parameter changes. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No", "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635700247450}], "openreview_url": "https://openreview.net/forum?id=Lwclw6u3Pcw", "arxiv_id": "2101.07752", "paper_pdf": "papers/Lwclw6u3Pcw.pdf", "paper_pdf_sha256": "f34ac26de02e8fb877a036a626701bcaecbd59b3c391607ed2e439afe2b2635e", "paper_pdf_bytes": 309153, "paper_pdf_source": "openreview", "code_url": "https://github.com/asier-gutierrez/nn-similarity", "code_repository": "asier-gutierrez/nn-similarity", "code_commit": "5e88b34145ff93e8e3c97ffe460f861663bc524e", "code_archive": "repos/Lwclw6u3Pcw.zip", "code_archive_sha256": "de9bca3c2134ec460d28ac99986c1dbc95cfc8b9c0c07a330bd7db1f09cdeb74", "code_archive_bytes": 226013, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 218, "github_languages": {"Python": 104931}, "github_archived": false, "github_pushed_at": "2021-07-17T08:20:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/determining-structural-properties-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "J_pvI6ap5Mn", "year": 2021, "status": "rejected", "title": "Transfer Learning of Graph Neural Networks with Ego-graph Information Maximization", "authors": ["Qi Zhu", "Yidan Xu", "Haonan Wang", "Chao Zhang", "Jiawei Han", "Carl Yang"], "authorids": ["~Qi_Zhu7", "~Yidan_Xu1", "~Haonan_Wang1", "~Chao_Zhang9", "~Jiawei_Han1", "~Carl_Yang1"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) have been shown with superior performance in various applications, but training dedicated GNNs can be costly for large-scale graphs. Some recent work started to study the pre-training of GNNs. However, none of them provide theoretical insights into the design of their frameworks, or clear requirements and guarantees towards the transferability of GNNs. In this work, we establish a theoretically grounded and practically useful framework for the transfer learning of GNNs. Firstly, we propose a novel view towards the essential graph information and advocate the capturing of it as the goal of transferable GNN training, which motivates the design of EGI (ego-graph information maximization) to analytically achieve this goal. Secondly, we specify the requirement of structure-respecting node features as the GNN input, and conduct a rigorous analysis of GNN transferability based on the difference between the local graph Laplacians of the source and target graphs. Finally, we conduct controlled synthetic experiments to directly justify our theoretical conclusions. Extensive experiments on real-world networks towards role identification show consistent results in the rigorously analyzed setting of direct-transfering (freezing parameters), while those towards large-scale relation prediction show promising results in the more generalized and practical setting of transfering with fine-tuning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "snCtKNlg-gx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper862/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work considers the unsupervised learning for graph neural networks. The work has solid theoretical analysis and extensive experimental studies. To encode the structure information, the K-hop ego-graph is used to generate a k-hop ego-graph for each node. I don’t see any major issues in this work. Here are several small concerns:\n\n1.\tIn this work, the ordered k-hop ego-graph is used but didn’t discuss how this order is generated. Would the authors explain how the ordering works.\n2.\tSince k is an important hyper-parameter, can authors provide some experiments to evaluate the impact of different k values.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Solid theoretical analysis and extensive experimental studies", "review": "This work considers the unsupervised learning for graph neural networks. The work has solid theoretical analysis and extensive experimental studies. To encode the structure information, the K-hop ego-graph is used to generate a k-hop ego-graph for each node. I don’t see any major issues in this work. Here are several small concerns:\n\n1.\tIn this work, the ordered k-hop ego-graph is used but didn’t discuss how this order is generated. Would the authors explain how the ordering works.\n2.\tSince k is an important hyper-parameter, can authors provide some experiments to evaluate the impact of different k values.\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604454430048}, {"id": "qG05RqeoeCq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper862/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors proposed a transfer learning scheme for graph neural networks. The proposed method ego-graph information maximization allows learning transferable models. The authors studied structure-respecting node features and provided a theoretical analysis of the transferability of GNNs. The proposed method significantly outperforms state-of-the-art methods.\n \nClarity:\nOverall, this paper reads fine. There are some typos and missing definitions of symbols, e.g., `sp' in eq 1 and $U^T$ in eq 3. D function is defined by another D in eq 2. In definition 2.3, 'Ordered' ego-graph is not defined. 'Title 2.2. ANALYSI', 'structural equivalence', and 'structural different' are typos. The average structural difference denoted by $\\bar{d}(,)$ \nare not defined. The clarity of this manuscript needs to be improved.\n \nStrengths/Quality/Significance (pros):\nThe interesting observation that the functions learned by GNNs can be viewed as functions to map a subgraph centered at a node to a class label since most GNNs have a few layers and their receptive field of a node output is a k-hop ego-graph.\n \nThe authors studied structure-respecting node features, e.g., degrees, spectral embeddings, to show that graph filters of GNNs is transferable. Based on the structure-respecting node features, the authors provide the analysis of transferability solely depending on the graph structure. The analysis showed that the performance gap of transferred models is bounded by a function of the ordered eigenvalues of the graph Laplacian of ego-graphs.\n \nThe proposed method achieved significant improvement against baseline approaches.\n \nWeaknesses (cons) & Questions:\n \nThe writing should be improved. The manuscript should be self-contained. As mentioned above, there are functions, and variables that are introduced without definitions such as reconstruction loss, sp, $\\bar{d}(,)$ and so on.\n \nThe analysis is limited to graph structures. To benefit most GNNs in real-world applications, the transferability of GNNs needs to be analyzed with node features as well.\n \nIn this paper, the analysis of the transferability of GNNs is limited to node classification. It is not clear whether the proposed method is effective in other tasks on graphs such as link prediction, graph classification. \n \nEven in the synthetic experiments, the performance gain is obtained only in the transferrable feature settings.\n\n--- Post Rebuttal --- \nI read the author response and I keep the original rating due to the limited operating range of the proposed method.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting results but a limited coverage", "review": "The authors proposed a transfer learning scheme for graph neural networks. The proposed method ego-graph information maximization allows learning transferable models. The authors studied structure-respecting node features and provided a theoretical analysis of the transferability of GNNs. The proposed method significantly outperforms state-of-the-art methods.\n \nClarity:\nOverall, this paper reads fine. There are some typos and missing definitions of symbols, e.g., `sp' in eq 1 and $U^T$ in eq 3. D function is defined by another D in eq 2. In definition 2.3, 'Ordered' ego-graph is not defined. 'Title 2.2. ANALYSI', 'structural equivalence', and 'structural different' are typos. The average structural difference denoted by $\\bar{d}(,)$ \nare not defined. The clarity of this manuscript needs to be improved.\n \nStrengths/Quality/Significance (pros):\nThe interesting observation that the functions learned by GNNs can be viewed as functions to map a subgraph centered at a node to a class label since most GNNs have a few layers and their receptive field of a node output is a k-hop ego-graph.\n \nThe authors studied structure-respecting node features, e.g., degrees, spectral embeddings, to show that graph filters of GNNs is transferable. Based on the structure-respecting node features, the authors provide the analysis of transferability solely depending on the graph structure. The analysis showed that the performance gap of transferred models is bounded by a function of the ordered eigenvalues of the graph Laplacian of ego-graphs.\n \nThe proposed method achieved significant improvement against baseline approaches.\n \nWeaknesses (cons) & Questions:\n \nThe writing should be improved. The manuscript should be self-contained. As mentioned above, there are functions, and variables that are introduced without definitions such as reconstruction loss, sp, $\\bar{d}(,)$ and so on.\n \nThe analysis is limited to graph structures. To benefit most GNNs in real-world applications, the transferability of GNNs needs to be analyzed with node features as well.\n \nIn this paper, the analysis of the transferability of GNNs is limited to node classification. It is not clear whether the proposed method is effective in other tasks on graphs such as link prediction, graph classification. \n \nEven in the synthetic experiments, the performance gain is obtained only in the transferrable feature settings.\n\n--- Post Rebuttal --- \nI read the author response and I keep the original rating due to the limited operating range of the proposed method.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604074250359}, {"id": "GpOZpfY-GF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper862/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces a theoretical framework for analyzing GNN transferability. The main idea is to view a graph as subgraph samples with the information of both the connections and the features. Based on this view, the authors define EGI score of a graph as a learnable function that needs to be optimized by maximizing the mutual information between the subgraph and the GNN output embedding of the center node.  Then, the authors give an upper bound for the difference of EGI scores of two graphs based on the difference of eigenvalues of the graph Laplacian of the subgraph samples from the two graphs. The implication is that if the difference of the eigenvalues is small, then the EGI scores are similar, which means the GNN has a similar ability to encode the structure of the two graphs. \n\nThe idea is new to me. One suggestion is that there are multiple forward references without precise pointers in the paper. Section 2.1 frequently refers to contents in Section 2.2. The authors may want to reorganize the paper to avoid any confusion.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "need some reorganization", "review": "The paper introduces a theoretical framework for analyzing GNN transferability. The main idea is to view a graph as subgraph samples with the information of both the connections and the features. Based on this view, the authors define EGI score of a graph as a learnable function that needs to be optimized by maximizing the mutual information between the subgraph and the GNN output embedding of the center node.  Then, the authors give an upper bound for the difference of EGI scores of two graphs based on the difference of eigenvalues of the graph Laplacian of the subgraph samples from the two graphs. The implication is that if the difference of the eigenvalues is small, then the EGI scores are similar, which means the GNN has a similar ability to encode the structure of the two graphs. \n\nThe idea is new to me. One suggestion is that there are multiple forward references without precise pointers in the paper. Section 2.1 frequently refers to contents in Section 2.2. The authors may want to reorganize the paper to avoid any confusion.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603856714674}, {"id": "cd83F0ZekwN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper862/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose ego-graph information maximization (EGI) to build more transferable GNN. They further theoretically study the transferability of EGI. \n\n\nThe article is fairly well structured, apart from the literature review, which is somewhat missing significant related works, such as Ron Levie's papers, which are the first theoretical studies regarding the transferability of GCNs. Concerning terminology, a reminder of what is called “node features” would be very useful for the reader, as this paper seems to construct features based on the graph structure instead of node attributes or signal over nodes as the node features. The article has the unfortunate tendency to contains a few unproven claims. For instance, \"we establish a theoretically grounded and practically useful framework for the transfer learning of GNNs\", which would deserve some empirical/theoretical evidence. Yet, the paper is only exploring the transferability of the proposed GIN.\n\nThe main argument is to establish a theoretically grounded framework for the transfer learning of GNNs and leverage it to propose a practically transferable GNN model, which I think is a very valid motivation for this paper and also is an important problem, but I think the authors could not achieve this goal properly. From my point of view, this paper proposes a new GNN model and further show theoretically and empirically that their new model is more transferable in some specific circumstances though, it’s quite hard to see some conclusions in my opinion based on both real and simulated experiments (but the transferability arguments of EGI holds to some degree nonetheless).\n\nComments/concerns/questions:  \n\n- In practice, the dataset consists of signals defined on many different graphs; the trained GNN should generalize to signals on graphs unseen in the training set. However, the experiments did not consider such a situation and only focused on some specific constructed features. Also, the authors claim the node features should be a function of the structure. The node attributes (signals over graph) can somehow be considered a noisy function of the graph structure. I would like to see the comparison performance in such situations too, where the graph has node attributes by itself. \n\n- What kind of functions over the graph is reasonable to construct the node features? How much noise is acceptable?\n\n- While the authors talk about the joint distribution of graph structure and node features, I think they are only using graph structure in two ways. Then, the main question is how one can use the graph structure as node attributes. Can one concatenate different constructed node features as node attributes? For example, assuming using both node degree and one-hub embedded graph based on VGAE.\n\n- It has been claimed \"[...] under an analogous setting of domain adaptation[..]\". How can one interpret it as domain adaptation? Can you show the embedded space? \n\n- Considering the graph convolutional neural network, either \"On the Transferability of Spectral Graph Filters\" or  \"Transferability of Spectral Graph Convolutional Neural Networks\" can be considered one important baseline.\n\n- Does $\\psi$ is the same for all graphs? If not, the heterogeneous embedding spaces should be an issue.\n\n- The similarity is not interpretable, and it is dependent on data. This is one of the main drawbacks of this paper. Specifically, for EGI on the Gene dataset, the structural difference higher than 5.4 causes a negative transfer effect; however, a structural difference larger than 12 can improve the airport dataset performance. In practice, we do not know the true answer. How can one know if the method is practical in a new dataset? \n\n- Besides that, $G_5$ has a higher distance than $G_4$ and $G_3$; however, it exploits the transferability of EGI more. A discussion on this phenomenon is needed.\n\n- The empirical results do not show the same relationship between DGI and the structural differences. If one considers the transferable feature column for DGI, please compare the performance improvement with the untrained GIN. \n\n- I believe the 3rd and 4th row of the second table is wrong. Please double check. Based on the first table, DGI outperforms EGI for non-transferable features. A comparison, based on another kind of features are needed. \n\n- Table 2 needs to have other columns similar to the first table to see how $\\Delta$ will be. \n\n- I also would like to see the target graph B performance in Table 1 (F-B and B-B). Also, the performance of considering two other airports as the source for the airport dataset.\n\n- Comparison with the baselines on the Gene dataset would be helpful. \n\n- Comparing EGI with SOTA GNN methods in some other analytical tasks, including node classification, link prediction, etc. can help validate this model's capability as a new GNN. \n\n- Including a simulation based on some specific perturbations on the target graph is also insightful.\n\n------- UPDATE ---------\n\nI thank the authors for their response to my concerns/comments. It seems like I have to defend my position for suggesting a rejection of the paper. While the response of the authors has clarified some aspects, some comments have not been adequately addressed.\n\nR1: However, the authors could not show how to measure the structural differences between the source and target graphs? Their measurement is not stable and is different between different datasets. How one can evaluate a new dataset is good for transferring or not (R5)? You could see how Ron Levie et al. showed this in their experiments. \n\nR2: But embedding using GCN is another definition for the function of the structure. However, it does not work well as node degree. Therefore, the author cannot point out that you only need to construct the features based on a function of the structure.\n\nR4: You can somehow use their idea to show how much difference is reasonable to transfer information. \n\nR5: So it seems this is one of the main drawbacks of considering \\delta. If that is the case, you need to discuss it in the paper. \n\nR7: I believe the response is not sufficient. The authors need to add that to show they could positively transfer information.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Ego-graph information maximization (EGI) is more transferable GNN", "review": "The authors propose ego-graph information maximization (EGI) to build more transferable GNN. They further theoretically study the transferability of EGI. \n\n\nThe article is fairly well structured, apart from the literature review, which is somewhat missing significant related works, such as Ron Levie's papers, which are the first theoretical studies regarding the transferability of GCNs. Concerning terminology, a reminder of what is called “node features” would be very useful for the reader, as this paper seems to construct features based on the graph structure instead of node attributes or signal over nodes as the node features. The article has the unfortunate tendency to contains a few unproven claims. For instance, \"we establish a theoretically grounded and practically useful framework for the transfer learning of GNNs\", which would deserve some empirical/theoretical evidence. Yet, the paper is only exploring the transferability of the proposed GIN.\n\nThe main argument is to establish a theoretically grounded framework for the transfer learning of GNNs and leverage it to propose a practically transferable GNN model, which I think is a very valid motivation for this paper and also is an important problem, but I think the authors could not achieve this goal properly. From my point of view, this paper proposes a new GNN model and further show theoretically and empirically that their new model is more transferable in some specific circumstances though, it’s quite hard to see some conclusions in my opinion based on both real and simulated experiments (but the transferability arguments of EGI holds to some degree nonetheless).\n\nComments/concerns/questions:  \n\n- In practice, the dataset consists of signals defined on many different graphs; the trained GNN should generalize to signals on graphs unseen in the training set. However, the experiments did not consider such a situation and only focused on some specific constructed features. Also, the authors claim the node features should be a function of the structure. The node attributes (signals over graph) can somehow be considered a noisy function of the graph structure. I would like to see the comparison performance in such situations too, where the graph has node attributes by itself. \n\n- What kind of functions over the graph is reasonable to construct the node features? How much noise is acceptable?\n\n- While the authors talk about the joint distribution of graph structure and node features, I think they are only using graph structure in two ways. Then, the main question is how one can use the graph structure as node attributes. Can one concatenate different constructed node features as node attributes? For example, assuming using both node degree and one-hub embedded graph based on VGAE.\n\n- It has been claimed \"[...] under an analogous setting of domain adaptation[..]\". How can one interpret it as domain adaptation? Can you show the embedded space? \n\n- Considering the graph convolutional neural network, either \"On the Transferability of Spectral Graph Filters\" or  \"Transferability of Spectral Graph Convolutional Neural Networks\" can be considered one important baseline.\n\n- Does $\\psi$ is the same for all graphs? If not, the heterogeneous embedding spaces should be an issue.\n\n- The similarity is not interpretable, and it is dependent on data. This is one of the main drawbacks of this paper. Specifically, for EGI on the Gene dataset, the structural difference higher than 5.4 causes a negative transfer effect; however, a structural difference larger than 12 can improve the airport dataset performance. In practice, we do not know the true answer. How can one know if the method is practical in a new dataset? \n\n- Besides that, $G_5$ has a higher distance than $G_4$ and $G_3$; however, it exploits the transferability of EGI more. A discussion on this phenomenon is needed.\n\n- The empirical results do not show the same relationship between DGI and the structural differences. If one considers the transferable feature column for DGI, please compare the performance improvement with the untrained GIN. \n\n- I believe the 3rd and 4th row of the second table is wrong. Please double check. Based on the first table, DGI outperforms EGI for non-transferable features. A comparison, based on another kind of features are needed. \n\n- Table 2 needs to have other columns similar to the first table to see how $\\Delta$ will be. \n\n- I also would like to see the target graph B performance in Table 1 (F-B and B-B). Also, the performance of considering two other airports as the source for the airport dataset.\n\n- Comparison with the baselines on the Gene dataset would be helpful. \n\n- Comparing EGI with SOTA GNN methods in some other analytical tasks, including node classification, link prediction, etc. can help validate this model's capability as a new GNN. \n\n- Including a simulation based on some specific perturbations on the target graph is also insightful.\n\n------- UPDATE ---------\n\nI thank the authors for their response to my concerns/comments. It seems like I have to defend my position for suggesting a rejection of the paper. While the response of the authors has clarified some aspects, some comments have not been adequately addressed.\n\nR1: However, the authors could not show how to measure the structural differences between the source and target graphs? Their measurement is not stable and is different between different datasets. How one can evaluate a new dataset is good for transferring or not (R5)? You could see how Ron Levie et al. showed this in their experiments. \n\nR2: But embedding using GCN is another definition for the function of the structure. However, it does not work well as node degree. Therefore, the author cannot point out that you only need to construct the features based on a function of the structure.\n\nR4: You can somehow use their idea to show how much difference is reasonable to transfer information. \n\nR5: So it seems this is one of the main drawbacks of considering \\delta. If that is the case, you need to discuss it in the paper. \n\nR7: I believe the response is not sufficient. The authors need to add that to show they could positively transfer information.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603658739404}], "openreview_url": "https://openreview.net/forum?id=J_pvI6ap5Mn", "arxiv_id": "2009.05204", "paper_pdf": "papers/J_pvI6ap5Mn.pdf", "paper_pdf_sha256": "eeb409573db5281d3622578c2c221514c5fc724c07cdb94920304d075e3fdf63", "paper_pdf_bytes": 1218614, "paper_pdf_source": "openreview", "code_url": "https://github.com/GentleZhu/EGI", "code_repository": "GentleZhu/EGI", "code_commit": "f8b8ce8e987c63b97fb49a838fec600223533439", "code_archive": "repos/J_pvI6ap5Mn.zip", "code_archive_sha256": "955f048d301dbe1ea53e6ea9edd7286fcb7bb51fcd78e52f5e1db9e49ee29bac", "code_archive_bytes": 185519, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 246, "github_languages": {"Python": 121788}, "github_archived": false, "github_pushed_at": "2021-12-09T16:37:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transfer-learning-of-graph-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "H1gz_nNYDS", "year": 2020, "status": "rejected", "title": "AutoSlim: Towards One-Shot Architecture Search for Channel Numbers", "authors": ["Jiahui Yu", "Thomas Huang"], "authorids": ["jyu79@illinois.edu", "t-huang1@illinois.edu"], "authors_source": "OpenReview API", "abstract": "\nWe study how to set the number of channels in a neural network to achieve better accuracy under constrained resources (e.g., FLOPs, latency, memory footprint or model size). A simple and one-shot approach, named AutoSlim, is presented. Instead of training many network samples and searching with reinforcement learning, we train a single slimmable network to approximate the network accuracy of different channel configurations. We then iteratively evaluate the trained slimmable model and greedily slim the layer with minimal accuracy drop. By this single pass, we can obtain the optimized channel configurations under different resource constraints. We present experiments with MobileNet v1, MobileNet v2, ResNet-50 and RL-searched MNasNet on ImageNet classification. We show significant improvements over their default channel configurations. We also achieve better accuracy than recent channel pruning methods and neural architecture search methods with 100X lower search cost.\n\nNotably, by setting optimized channel numbers, our AutoSlim-MobileNet-v2 at 305M FLOPs achieves 74.2% top-1 accuracy, 2.4% better than default MobileNet-v2 (301M FLOPs), and even 0.2% better than RL-searched MNasNet (317M FLOPs). Our AutoSlim-ResNet-50 at 570M FLOPs, without depthwise convolutions, achieves 1.3% better accuracy than MobileNet-v1 (569M FLOPs).\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bkx0hBuM9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper34/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper targets on learning the number of channels across all layers, under computation/model size/memory constraints. The method is simple and the results seems promising. \n\nHowever, the following issues need to be resolved:\n1. The main method is based on published \"slimmable networks,\" such that the novelty is limited;\n2. The method is very simpler to DropPath in [1], which uses DropPath to learn important branches while this paper uses it to learn channels. They are similar.\n3. Better ablation studies are required in Table 1. This table should be simplified. As the method cannot learn architectures but channel numbers, the only useful pairs of comparisons are those having the same architecture, such as  a pair of MobileNet vs AutoSlim-MobileNet.\n4. an important detail is missing: where does the AutoSlim start from? Does it start from a larger model than the baseline? In the set of \"500M FLOPs\" experiments, I see the size of \"AutoSlim-MobileNet v1\" (4.6M) is larger than \"MobileNet v1 1.0x\" (4.2M), this implies that AutoSlim start from a \"MobileNet v1 Nx\" and N > 1.0. What is exactly N?\n5. If AutoSlim starts from a larger baseline model with N times (N > 1.0) width, then the pruning baseline methods (AMC and ThiNet) should also start from the same larger models for fair comparison. In general, starting from a larger model and pruning it down can achieve a better accuracy vs. size trade-off.\n6. \"300 epochs with linearly decaying learning rate for mobile networks, 100 epochs with step learning rate schedule for ResNet-50 based models\", are baselines trained in the same way?\n\nMinor: \n1. missing captions in a couple of figures, e.g., Figure 5.\n2. \"the importance of trained weights\" vs \"the importance of channel numbers\" is trivial\n\n\n[1] Bender, Gabriel, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. \"Understanding and simplifying one-shot architecture search.\" In International Conference on Machine Learning, pp. 549-558. 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper targets on learning the number of channels across all layers, under computation/model size/memory constraints. The method is simple and the results seems promising. \n\nHowever, the following issues need to be resolved:\n1. The main method is based on published \"slimmable networks,\" such that the novelty is limited;\n2. The method is very simpler to DropPath in [1], which uses DropPath to learn important branches while this paper uses it to learn channels. They are similar.\n3. Better ablation studies are required in Table 1. This table should be simplified. As the method cannot learn architectures but channel numbers, the only useful pairs of comparisons are those having the same architecture, such as  a pair of MobileNet vs AutoSlim-MobileNet.\n4. an important detail is missing: where does the AutoSlim start from? Does it start from a larger model than the baseline? In the set of \"500M FLOPs\" experiments, I see the size of \"AutoSlim-MobileNet v1\" (4.6M) is larger than \"MobileNet v1 1.0x\" (4.2M), this implies that AutoSlim start from a \"MobileNet v1 Nx\" and N > 1.0. What is exactly N?\n5. If AutoSlim starts from a larger baseline model with N times (N > 1.0) width, then the pruning baseline methods (AMC and ThiNet) should also start from the same larger models for fair comparison. In general, starting from a larger model and pruning it down can achieve a better accuracy vs. size trade-off.\n6. \"300 epochs with linearly decaying learning rate for mobile networks, 100 epochs with step learning rate schedule for ResNet-50 based models\", are baselines trained in the same way?\n\nMinor: \n1. missing captions in a couple of figures, e.g., Figure 5.\n2. \"the importance of trained weights\" vs \"the importance of channel numbers\" is trivial\n\n\n[1] Bender, Gabriel, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. \"Understanding and simplifying one-shot architecture search.\" In International Conference on Machine Learning, pp. 549-558. 2018."}, "tcdate": 1572140469898}, {"id": "Hye2e3RpFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper34/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a simple and one-shot approach on neural architecture search for the number of channels to achieve better accuracy. Rather than training a lot of network samples, the proposed method trains a single slimmable network to approximate the network accuracy of different channel configurations. The experimental results show that the proposed method achieves better performance than the existing baseline methods.\n\n- It would be better to provide the search cost of the proposed method and the other baseline methods because that is the important metric for neural architecture search methods. As this paper points out that NAS methods are computationally expensive, it would be better to make the efficiency of the proposed method clear.\n\n- According to this paper, the notable difference between the proposed method and the existing pruning methods is that the pruning methods are grounded on the importance of trained weights, but the proposed method focuses more on the importance of channel numbers. It is unclear to me why such a difference is caused by the proposed method, that is, which part of the proposed method causes the difference? And how does the difference affect the final performance?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This paper proposes a simple and one-shot approach on neural architecture search for the number of channels to achieve better accuracy. Rather than training a lot of network samples, the proposed method trains a single slimmable network to approximate the network accuracy of different channel configurations. The experimental results show that the proposed method achieves better performance than the existing baseline methods.\n\n- It would be better to provide the search cost of the proposed method and the other baseline methods because that is the important metric for neural architecture search methods. As this paper points out that NAS methods are computationally expensive, it would be better to make the efficiency of the proposed method clear.\n\n- According to this paper, the notable difference between the proposed method and the existing pruning methods is that the pruning methods are grounded on the importance of trained weights, but the proposed method focuses more on the importance of channel numbers. It is unclear to me why such a difference is caused by the proposed method, that is, which part of the proposed method causes the difference? And how does the difference affect the final performance?"}, "tcdate": 1571838963697}, {"id": "H1xkjO_ptS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper34/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a method to perform architecture search on the number of channels in convolutional layers. The proposed method, called AutoSlim, is a one-shot approach based on previous work of Slimmable Networks [2,3]. The authors have tested the proposed methods on a variety of architectures on ImageNet dataset. \n\nThe paper is well-written and easy to follow. I really appreciate the authors for structuring this paper so well. I have the following questions:\n\nQ1: In figure 4, the authors find that “Compared with default MobileNet v2, our optimized configuration has fewer channels in shallow layers and more channels in deep ones.” This is interesting. Because in network pruning methods, it is found that usually later stages get pruned more [1] (e.g. VGG), indicating that there is more redundancy for deep layers. However, in this case, actually deep layers get more channels than standard models. Is there any justification for this? Is it that more channels in deep layers benefit the accuracy?\n\nQ2: In “Training Optimized Networks”, the authors mentioned that “By default we search for the network FLOPs at approximately 200M, 300M and 500M, and train a slimmable model.” Does this mean that the authors train the final optimized models from scratch as a slimmable network using “sandwich rule” and “in-place distillation” rule? Or are the authors just training the final model with standard training schedule? If it is the first case, can the authors justify why?\n\nQ3: In Table 1, “Heavy Models”, what is the difference between “ResNet-50” and “He-ResNet-50”? Also, why the params, memory and CPU Latency of some networks are omitted?\n\nQ4: In the last paragraph of section 4, the authors tried the transferability of networks learned from ImageNet to CIFAR-10 dataset. I am not sure how the authors transfer the networks from Imagenet to CIFAR-10? Is it the ratio of the number of channels? Can the authors provide the architecture details of MobileNet v2 on CIFAR-10 dataset?\n\nQ5: What is the estimated time for a typical run of AutoSlim? How does it compare to network pruning methods or neural architecture search methods?\n\nQ6: Can the methods be used to search for the number of neurons in fully connected layers? Are there any results?\n\n[1] Rethinking the Value of Network Pruning. Zhuang et al. ICLR 2019\n[2] Slimmable neural networks. Yu et al. ICLR 2019.\n[3] Universally Slimmable Networks and Improved Training Techniques. Yu et al. Arxiv.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "In this paper, the authors propose a method to perform architecture search on the number of channels in convolutional layers. The proposed method, called AutoSlim, is a one-shot approach based on previous work of Slimmable Networks [2,3]. The authors have tested the proposed methods on a variety of architectures on ImageNet dataset. \n\nThe paper is well-written and easy to follow. I really appreciate the authors for structuring this paper so well. I have the following questions:\n\nQ1: In figure 4, the authors find that “Compared with default MobileNet v2, our optimized configuration has fewer channels in shallow layers and more channels in deep ones.” This is interesting. Because in network pruning methods, it is found that usually later stages get pruned more [1] (e.g. VGG), indicating that there is more redundancy for deep layers. However, in this case, actually deep layers get more channels than standard models. Is there any justification for this? Is it that more channels in deep layers benefit the accuracy?\n\nQ2: In “Training Optimized Networks”, the authors mentioned that “By default we search for the network FLOPs at approximately 200M, 300M and 500M, and train a slimmable model.” Does this mean that the authors train the final optimized models from scratch as a slimmable network using “sandwich rule” and “in-place distillation” rule? Or are the authors just training the final model with standard training schedule? If it is the first case, can the authors justify why?\n\nQ3: In Table 1, “Heavy Models”, what is the difference between “ResNet-50” and “He-ResNet-50”? Also, why the params, memory and CPU Latency of some networks are omitted?\n\nQ4: In the last paragraph of section 4, the authors tried the transferability of networks learned from ImageNet to CIFAR-10 dataset. I am not sure how the authors transfer the networks from Imagenet to CIFAR-10? Is it the ratio of the number of channels? Can the authors provide the architecture details of MobileNet v2 on CIFAR-10 dataset?\n\nQ5: What is the estimated time for a typical run of AutoSlim? How does it compare to network pruning methods or neural architecture search methods?\n\nQ6: Can the methods be used to search for the number of neurons in fully connected layers? Are there any results?\n\n[1] Rethinking the Value of Network Pruning. Zhuang et al. ICLR 2019\n[2] Slimmable neural networks. Yu et al. ICLR 2019.\n[3] Universally Slimmable Networks and Improved Training Techniques. Yu et al. Arxiv.\n"}, "tcdate": 1571813526600}], "openreview_url": "https://openreview.net/forum?id=H1gz_nNYDS", "arxiv_id": "1903.11728", "paper_pdf": "papers/H1gz_nNYDS.pdf", "paper_pdf_sha256": "36b441e49cfd8acb0cdbb38f7ff8f010a9e3b9b0c51b819c726571f31d1a965b", "paper_pdf_bytes": 599945, "paper_pdf_source": "openreview", "code_url": "https://github.com/JiahuiYu/slimmable_networks", "code_repository": "JiahuiYu/slimmable_networks", "code_commit": "5dc14d0357ccfc596d706281acdc8a5b0b66c6d6", "code_archive": "repos/H1gz_nNYDS.zip", "code_archive_sha256": "4c507301cee871fa3074273131195b58b96f53cadc272667dbed40d07e121143", "code_archive_bytes": 49374, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 329, "github_languages": {"Python": 106555}, "github_archived": false, "github_pushed_at": "2023-03-09T08:37:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/network-slimming-by-slimmable-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wywgRd1MUQ", "year": 2026, "status": "rejected", "title": "InfiGUI-R1: Advancing Multimodal GUI Agents from Reactive Actors to Deliberative Reasoners", "authors": ["Yuhang Liu", "Pengxiang Li", "Congkai Xie", "Xueyu Hu", "Xiaotian Han", "Shengyu Zhang", "Hongxia Yang", "Fei Wu"], "authorids": ["~Yuhang_Liu9", "~Pengxiang_Li2", "~Congkai_Xie1", "~Xueyu_Hu1", "~Xiaotian_Han2", "~Shengyu_Zhang2", "~Hongxia_Yang2", "~Fei_Wu1"], "authors_source": "OpenReview API", "abstract": "Multimodal Large Language Models (MLLMs) have shown significant promise in powering Graphical User Interface (GUI) agents to automate complex digital tasks. However, the prevailing monolithic training paradigms often create a structural mismatch with the hierarchical nature of capabilities required for robust performance. Specifically, the efficacy of methods like Reinforcement Learning (RL) is critically predicated on the agent possessing a high-quality behavioral prior of key reasoning skills, such as spatial reasoning and goal decomposition, which are often absent. To resolve this impasse, we propose Actor2Reasoner, a novel two-stage hierarchical training paradigm grounded in the principle of Endow First, Internalize Later. The first stage, Cognitive Endowment, employs targeted supervised fine-tuning to instill these crucial thinking patterns, forging a Capable Actor. Subsequently, the second stage, Policy Internalization, utilizes RL to evolve this actor into a Deliberative Reasoner by internalizing the endowed abilities into a robust, context-aware decision-making policy. We instantiate our paradigm in InfiGUI-R1, an agent that achieves state-of-the-art performance on challenging benchmarks, including AndroidControl. Our work demonstrates that decoupling the endowment of foundational abilities from the internalization of policy provides a more effective and principled path toward developing sophisticated and resilient GUI agents.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "b4hICtwVVO", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10070/Reviewer_cbWR"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 5, "summary": "This paper presents INFIGUI-R1: a GUI agent that is trained in two stages (SFT/Distillation then RL).\n\nThe paper grounds the two stage training pipeline in the principle of \"Endow First, Internalize Later\", which gives a new justification to why this common paradigm seems to work well.\n\nIn the SFT/Distillation phase: Structured Chains-of-Thought (CoT) are generated and then validated from a larger teacher model (to produce the ground truth actions when fed as prompt) , then appended to the original prompt, and then used to fine-tune the base model.\n\nIn the RL phase: a comprehensive set of rewards are designed to cover targets like response formatting and action accuracy, and supplemented by specifically curated reward signals to strengthen spatial reasoning and error recovery. \n\nThe paper has many evaluations and shows uplift on various UI benchmarks.", "review_text": "This paper presents INFIGUI-R1: a GUI agent that is trained in two stages (SFT/Distillation then RL).\n\nThe paper grounds the two stage training pipeline in the principle of \"Endow First, Internalize Later\", which gives a new justification to why this common paradigm seems to work well.\n\nIn the SFT/Distillation phase: Structured Chains-of-Thought (CoT) are generated and then validated from a larger teacher model (to produce the ground truth actions when fed as prompt) , then appended to the original prompt, and then used to fine-tune the base model.\n\nIn the RL phase: a comprehensive set of rewards are designed to cover targets like response formatting and action accuracy, and supplemented by specifically curated reward signals to strengthen spatial reasoning and error recovery. \n\nThe paper has many evaluations and shows uplift on various UI benchmarks.", "strengths": "- This work highlights an important problem for agents trained with RL (i.e. for RL to succeed, the MLLM needs to have some basic behavioral prior that enables it to solve certain aspects of the control problem). \n\n- The structured CoT synthesis process and the focus on the three core reasoning patterns relevant to GUI (Spatial Reasoning, Goal Decomposition, and Reflection) and the final verification-filtering step is optimized for GUI and seems to be a very important part of why the system shows good performance.\n\n- The reward design and formulation is very simple and scalable, and seems to be one of the main reasons the system has good performance.\n\n- The training set size is small (~3k in SFT, and ~44k in RL), which might indicate that this regiment is more data-efficient.\n\n- Good ablation studies that show the improvement from the 2-stage training regiment.", "weaknesses": "- Although I liked the framing of  \"Endow First, Internalize Later\", as it grounds the common pattern of SFT/Distillation then RL into a pedagogical framework, however, I could not find any reference to the principle or how it's used in other work. A citation in the intro or on line 62 would make the principle used easier to understand.\n\n- The claimed paradigm shift from monolithic to  two-stage training regiment is not especially novel, many other relevant works are using the same two-stage hierarchical training paradigm (some form of SFT then some form of RL) for embodied and UI LLM-based models. This is before and after the DeepSeek-R1paper made it more popular.\n\n- It's important for these models to be tested at various sizes. I recommend at least a 7B model for comparison.\n\n- The model is not tested on OSWorld, even though the training data seems to include desktop GUI examples (e.g. OmniAct).\n\n- Please include results from baseline models on AndroidWorld on Table 4 to make the comparison easier. The success rate of the model in isolation is not very useful to compare.", "questions": "- From table 4: It seems that  the cognitive endowment is where most of the performance comes from, do you have justification for this? Why aren't the other signals/phases as important for AndroidWorld?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents INFIGUI-R1: a GUI agent that is trained in two stages (SFT/Distillation then RL).\n\nThe paper grounds the two stage training pipeline in the principle of \"Endow First, Internalize Later\", which gives a new justification to why this common paradigm seems to work well.\n\nIn the SFT/Distillation phase: Structured Chains-of-Thought (CoT) are generated and then validated from a larger teacher model (to produce the ground truth actions when fed as prompt) , then appended to the original prompt, and then used to fine-tune the base model.\n\nIn the RL phase: a comprehensive set of rewards are designed to cover targets like response formatting and action accuracy, and supplemented by specifically curated reward signals to strengthen spatial reasoning and error recovery. \n\nThe paper has many evaluations and shows uplift on various UI benchmarks.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- This work highlights an important problem for agents trained with RL (i.e. for RL to succeed, the MLLM needs to have some basic behavioral prior that enables it to solve certain aspects of the control problem). \n\n- The structured CoT synthesis process and the focus on the three core reasoning patterns relevant to GUI (Spatial Reasoning, Goal Decomposition, and Reflection) and the final verification-filtering step is optimized for GUI and seems to be a very important part of why the system shows good performance.\n\n- The reward design and formulation is very simple and scalable, and seems to be one of the main reasons the system has good performance.\n\n- The training set size is small (~3k in SFT, and ~44k in RL), which might indicate that this regiment is more data-efficient.\n\n- Good ablation studies that show the improvement from the 2-stage training regiment.", "weaknesses": "- Although I liked the framing of  \"Endow First, Internalize Later\", as it grounds the common pattern of SFT/Distillation then RL into a pedagogical framework, however, I could not find any reference to the principle or how it's used in other work. A citation in the intro or on line 62 would make the principle used easier to understand.\n\n- The claimed paradigm shift from monolithic to  two-stage training regiment is not especially novel, many other relevant works are using the same two-stage hierarchical training paradigm (some form of SFT then some form of RL) for embodied and UI LLM-based models. This is before and after the DeepSeek-R1paper made it more popular.\n\n- It's important for these models to be tested at various sizes. I recommend at least a 7B model for comparison.\n\n- The model is not tested on OSWorld, even though the training data seems to include desktop GUI examples (e.g. OmniAct).\n\n- Please include results from baseline models on AndroidWorld on Table 4 to make the comparison easier. The success rate of the model in isolation is not very useful to compare.", "questions": "- From table 4: It seems that  the cognitive endowment is where most of the performance comes from, do you have justification for this? Why aren't the other signals/phases as important for AndroidWorld?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762176294996}, {"id": "ImkZkDa0RW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10070/Reviewer_TC4U"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper aims to provide a methodology to improve an MLLM’s ability to perform digital GUI-based tasks. The paper highlights an important issue that when prompted to perform tasks in a GUI, LLMs struggle due to challenges faced in spatial localization and goal decomposition. They propose a two-step training paradigm to first build a base model with improved localization and grounding capabilities, and then operationalize these faculties through reinforcement learning. The authors validate their approach on three benchmarks to highlight the utility of their proposed method.", "review_text": "This paper aims to provide a methodology to improve an MLLM’s ability to perform digital GUI-based tasks. The paper highlights an important issue that when prompted to perform tasks in a GUI, LLMs struggle due to challenges faced in spatial localization and goal decomposition. They propose a two-step training paradigm to first build a base model with improved localization and grounding capabilities, and then operationalize these faculties through reinforcement learning. The authors validate their approach on three benchmarks to highlight the utility of their proposed method.", "strengths": "- The authors address an important problem regarding improving the capabilities of MLLM agents navigating a stochastic digital world.\n- The authors conduct experiments on an expansive set of baselines to cover the breadth of existing approaches to solve this problem.\n- I appreciated the authors efforts to ensure high-fidelity for their synthetic data via a self-verification based filtering step.", "weaknesses": "- The writing describing the technical details in the approach presented needs to be improved.\n    - While explaining the reward structure, it is unclear how the final reward is constructed. The author introduces various components of the reward, such as R_sub or R_esc, however, these terms are not included in the final description of the reward.\n    - The final reward in equation-6 was defined as R(a,B), however the training reward was previously defined as R_total. Are these equivalent? Furthermore, the “B” variable in the reward function has not been defined.\n    - In Section 3.3.3, what is the specific algorithm utilized for reinforcement learning? Reinforcement Learning with Verifiable Rewards is a very broad class of methods, which could involve a variety of policy-gradient based approaches, i.e. GRPO, PPO, DPO, etc. The authors need to more clearly describe the approach they selected and the motivation behind their selection. Furthermore, if the authors adopted an off-the-shelf algorithm, I would recommend that they move this information to an experiments or preliminaries section as it is not a novel part of their technical approach.\n- I think the authors have some positive findings in their experimental results. However, their reporting lacks detail, and some important results are not reported/explained.\n    - It is concerning that the authors did not discuss Table-1 at all in their discussion of their results. On this benchmark, many of the baseline approaches outperform InfiGUI-R1, countering the claims in the discussion.\n    - Ablations seemed to be cherry picked, as they have been reported only for two of the four benchmarks\n    - The results of the other baselines are absent in Table-4, while reporting the success rate of InfiGUI on the AndroidWorld benchmark. Since there was a significant drop in InfiGUI’s performance, compared to AndroidControl, I would be interested in contextualizing this performance drop compared to other methods.\n- [Minor] The approach of improving grounding capabilities through two-step training procedures is well-studied in the embodied-llm space [1,2,3,4]. I think this paper would benefit from including a discussion of these approaches and highlight how their method/problem is different from prior work.\n\n[1] - Yang, Ganlin, et al. \"Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning.\" arXiv preprint arXiv:2510.11027 (2025).\n[2] - Ahn, Michael, et al. \"Do as i can, not as i say: Grounding language in robotic affordances.\" arXiv preprint arXiv:2204.01691 (2022).\n\n[3] - Carta, Thomas, et al. \"Grounding large language models in interactive environments with online reinforcement learning.\" International Conference on Machine Learning. PMLR, 2023.\n\n[4] - Huang, Wenlong, et al. \"Grounded decoding: Guiding text generation with grounded models for embodied agents.\" Advances in Neural Information Processing Systems 36 (2023): 59636-59661.", "questions": "Listed in the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to provide a methodology to improve an MLLM’s ability to perform digital GUI-based tasks. The paper highlights an important issue that when prompted to perform tasks in a GUI, LLMs struggle due to challenges faced in spatial localization and goal decomposition. They propose a two-step training paradigm to first build a base model with improved localization and grounding capabilities, and then operationalize these faculties through reinforcement learning. The authors validate their approach on three benchmarks to highlight the utility of their proposed method.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "- The authors address an important problem regarding improving the capabilities of MLLM agents navigating a stochastic digital world.\n- The authors conduct experiments on an expansive set of baselines to cover the breadth of existing approaches to solve this problem.\n- I appreciated the authors efforts to ensure high-fidelity for their synthetic data via a self-verification based filtering step.", "weaknesses": "- The writing describing the technical details in the approach presented needs to be improved.\n    - While explaining the reward structure, it is unclear how the final reward is constructed. The author introduces various components of the reward, such as R_sub or R_esc, however, these terms are not included in the final description of the reward.\n    - The final reward in equation-6 was defined as R(a,B), however the training reward was previously defined as R_total. Are these equivalent? Furthermore, the “B” variable in the reward function has not been defined.\n    - In Section 3.3.3, what is the specific algorithm utilized for reinforcement learning? Reinforcement Learning with Verifiable Rewards is a very broad class of methods, which could involve a variety of policy-gradient based approaches, i.e. GRPO, PPO, DPO, etc. The authors need to more clearly describe the approach they selected and the motivation behind their selection. Furthermore, if the authors adopted an off-the-shelf algorithm, I would recommend that they move this information to an experiments or preliminaries section as it is not a novel part of their technical approach.\n- I think the authors have some positive findings in their experimental results. However, their reporting lacks detail, and some important results are not reported/explained.\n    - It is concerning that the authors did not discuss Table-1 at all in their discussion of their results. On this benchmark, many of the baseline approaches outperform InfiGUI-R1, countering the claims in the discussion.\n    - Ablations seemed to be cherry picked, as they have been reported only for two of the four benchmarks\n    - The results of the other baselines are absent in Table-4, while reporting the success rate of InfiGUI on the AndroidWorld benchmark. Since there was a significant drop in InfiGUI’s performance, compared to AndroidControl, I would be interested in contextualizing this performance drop compared to other methods.\n- [Minor] The approach of improving grounding capabilities through two-step training procedures is well-studied in the embodied-llm space [1,2,3,4]. I think this paper would benefit from including a discussion of these approaches and highlight how their method/problem is different from prior work.\n\n[1] - Yang, Ganlin, et al. \"Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning.\" arXiv preprint arXiv:2510.11027 (2025).\n[2] - Ahn, Michael, et al. \"Do as i can, not as i say: Grounding language in robotic affordances.\" arXiv preprint arXiv:2204.01691 (2022).\n\n[3] - Carta, Thomas, et al. \"Grounding large language models in interactive environments with online reinforcement learning.\" International Conference on Machine Learning. PMLR, 2023.\n\n[4] - Huang, Wenlong, et al. \"Grounded decoding: Guiding text generation with grounded models for embodied agents.\" Advances in Neural Information Processing Systems 36 (2023): 59636-59661.", "questions": "Listed in the weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761887939017}, {"id": "8RPnkvEfkJ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10070/Reviewer_sQ4b"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes Actor2Reasoner, a novel hierarchical training paradigm for GUI agent learning. The authors identify a structural mismatch between existing monolithic training approaches and the hierarchical nature of capabilities required for GUI tasks, proposing a two-stage training methodology based on the principle of \"Endow First, Internalize Later.\" The first stage, Cognitive Endowment, employs supervised learning to instill core cognitive abilities such as spatial reasoning and goal decomposition. The second stage, Policy Internalization, uses reinforcement learning to internalize these abilities into a robust decision-making policy. The approach is instantiated in InfiGUI-R1, which achieves SOTA performance on benchmarks including AndroidControl.", "review_text": "This paper proposes Actor2Reasoner, a novel hierarchical training paradigm for GUI agent learning. The authors identify a structural mismatch between existing monolithic training approaches and the hierarchical nature of capabilities required for GUI tasks, proposing a two-stage training methodology based on the principle of \"Endow First, Internalize Later.\" The first stage, Cognitive Endowment, employs supervised learning to instill core cognitive abilities such as spatial reasoning and goal decomposition. The second stage, Policy Internalization, uses reinforcement learning to internalize these abilities into a robust decision-making policy. The approach is instantiated in InfiGUI-R1, which achieves SOTA performance on benchmarks including AndroidControl.", "strengths": "- The paper is well-motivated and easy-to-read.\n- Performance evaluation across diverse benchmarks (AndroidControl, GUI-Odyssey) with detailed ablation studies effectively demonstrates the method's effectiveness. The experimental results showing synergistic effects between the two stages are particularly compelling.", "weaknesses": "- The main content of this paper is a two-stage learning algorithm where the first stage learns relatively general abilities through a teacher model, and the second stage learns GUI agent tasks through reinforcement learning. There are various papers using such two-stage structures [1, 2], and when reading this paper, it is difficult to understand what specifically changes for GUI environments beyond the reward structure.\n- In particular, the implementation of this two-stage approach seems to consist entirely of the detailed design of CoT and rewards for GUI agents in Sections 3.1-3.2, which appears to be the entirety of the framework.\n- There is insufficient theoretical analysis of why the two-stage approach is more efficient than a single-stage approach. Analysis of learning complexity or sample efficiency would have made a stronger contribution.\n- Detailed appendices about datasets or implementation are not included.\n\n[1] Song, Huatong, et al. \"R1-searcher: Incentivizing the search capability in llms via reinforcement learning.\" arXiv preprint arXiv:2503.05592 (2025).\n\n[2] Liu, Zijia, et al. \"Time-R1: Towards Comprehensive Temporal Reasoning in LLMs.\" *arXiv preprint arXiv:2505.13508* (2025).", "questions": "- In the proposed methodology, does the effect of Cognitive Endowment diminish when using a base model that is already large enough to have sufficient reasoning performance?\n- To accurately compare with UI-TARS, have you also applied the proposed method to Qwen2-VL? Or, could you provide the performance of Qwen2-VL and Qwen2.5-VL?\n- Are the abilities learned in the Cognitive Endowment stage well maintained without catastrophic forgetting during the Policy Internalization stage?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Actor2Reasoner, a novel hierarchical training paradigm for GUI agent learning. The authors identify a structural mismatch between existing monolithic training approaches and the hierarchical nature of capabilities required for GUI tasks, proposing a two-stage training methodology based on the principle of \"Endow First, Internalize Later.\" The first stage, Cognitive Endowment, employs supervised learning to instill core cognitive abilities such as spatial reasoning and goal decomposition. The second stage, Policy Internalization, uses reinforcement learning to internalize these abilities into a robust decision-making policy. The approach is instantiated in InfiGUI-R1, which achieves SOTA performance on benchmarks including AndroidControl.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper is well-motivated and easy-to-read.\n- Performance evaluation across diverse benchmarks (AndroidControl, GUI-Odyssey) with detailed ablation studies effectively demonstrates the method's effectiveness. The experimental results showing synergistic effects between the two stages are particularly compelling.", "weaknesses": "- The main content of this paper is a two-stage learning algorithm where the first stage learns relatively general abilities through a teacher model, and the second stage learns GUI agent tasks through reinforcement learning. There are various papers using such two-stage structures [1, 2], and when reading this paper, it is difficult to understand what specifically changes for GUI environments beyond the reward structure.\n- In particular, the implementation of this two-stage approach seems to consist entirely of the detailed design of CoT and rewards for GUI agents in Sections 3.1-3.2, which appears to be the entirety of the framework.\n- There is insufficient theoretical analysis of why the two-stage approach is more efficient than a single-stage approach. Analysis of learning complexity or sample efficiency would have made a stronger contribution.\n- Detailed appendices about datasets or implementation are not included.\n\n[1] Song, Huatong, et al. \"R1-searcher: Incentivizing the search capability in llms via reinforcement learning.\" arXiv preprint arXiv:2503.05592 (2025).\n\n[2] Liu, Zijia, et al. \"Time-R1: Towards Comprehensive Temporal Reasoning in LLMs.\" *arXiv preprint arXiv:2505.13508* (2025).", "questions": "- In the proposed methodology, does the effect of Cognitive Endowment diminish when using a base model that is already large enough to have sufficient reasoning performance?\n- To accurately compare with UI-TARS, have you also applied the proposed method to Qwen2-VL? Or, could you provide the performance of Qwen2-VL and Qwen2.5-VL?\n- Are the abilities learned in the Cognitive Endowment stage well maintained without catastrophic forgetting during the Policy Internalization stage?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761540548476}, {"id": "TmqjtvRmzc", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10070/Reviewer_BcjR"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors introduce InfiGUI-R1, a vision-language-action model to interact with user interfaces and accomplish tasks. The model is trained in a 2-stage training pipeline, first with supervised fine tuning over a clean action dataset, and second with reinforcement learning over trajectories that are “just hard enough” so as to remain on the learning frontier of the model. The model is trained with several different objectives depending on the task at hand, varying from bounding-box IoU to sub-goal matching to action prediction. As part of the learning process, the authors also propose a sub-goal prediction/matching task, and an error correction task. These are both claimed to help improve final performance.\n\nInfiGUI-R1 is compared to much larger models on ScreenSpot benchmarks for grounding performance, and on AndroidControl and AndroidWorld benchmarks for interaction performance. The model achieves comparable or superior performance when compared to baselines, despite being smaller. The authors also conduct an ablation study in which they remove either their SFT phase or their RL phase, and they show that both are necessary to the success of the final model. Ablations on the subgoal and error recover tasks seem to indicate that they are less important to the overall success of the pipeline, with the “no error recovery” model even outperforming the full model along some metrics.", "review_text": "The authors introduce InfiGUI-R1, a vision-language-action model to interact with user interfaces and accomplish tasks. The model is trained in a 2-stage training pipeline, first with supervised fine tuning over a clean action dataset, and second with reinforcement learning over trajectories that are “just hard enough” so as to remain on the learning frontier of the model. The model is trained with several different objectives depending on the task at hand, varying from bounding-box IoU to sub-goal matching to action prediction. As part of the learning process, the authors also propose a sub-goal prediction/matching task, and an error correction task. These are both claimed to help improve final performance.\n\nInfiGUI-R1 is compared to much larger models on ScreenSpot benchmarks for grounding performance, and on AndroidControl and AndroidWorld benchmarks for interaction performance. The model achieves comparable or superior performance when compared to baselines, despite being smaller. The authors also conduct an ablation study in which they remove either their SFT phase or their RL phase, and they show that both are necessary to the success of the final model. Ablations on the subgoal and error recover tasks seem to indicate that they are less important to the overall success of the pipeline, with the “no error recovery” model even outperforming the full model along some metrics.", "strengths": "+ The final performance of the InfiGUI-R1 model performs very well, achieving superior performance than larger models\n+ The proposed pipeline is intuitive and reflects many sensible design decisions, and would be easy to replicate by other researchers in their own workflows.\n+ The experiments evaluate InfiGUI-R1 along several important axes, showing results for different GUI instantiations (mobile/web/desktop), grounding performance on different tasks/applications, and task success on different benchmarks. These greatly help to illustrate the strengths of the model, and to contextualize it in the scope of related work.", "weaknesses": "- **Novelty**: SFT into RLVR is not a particularly new training paradigm. Every major LLM since ChatGPT has featured some combination of these two training paradigms in this order, and so the insight to apply SFT and then RLVR for GUI navigation is not particularly novel or insightful. In fact, one of the baselines in the paper, UI-TARS, employs an SFT-> DPO training pipeline, which is of course very similar. Similarly, many vision-language-action models in the robotics community have already demonstrated the success of SFT -> RLHF for task learning with VLMs.\n- Clarity: Elements of the paper are left unclear. For example, the authors introduce either a point-based reward or a bounding-box reward for certain UI tasks, but then there is no comparison between these two, or discussion on when to use each. This does not help future work to build on anything that the authors may have learned from experimenting with these rewards for UI navigation or grounding. Similarly, elements of the paper are left slightly unclear, such as how the Accuracy reward is normalized to [-1, 1], but sometimes it is set to the sum of 3 rewards ranging from [0, 1]… So it should range from [0, 3]. Is this reward then re-normalized to [-1, 1]? If so, is it a valid learning signal?\n- (Minor) Formatting: nearly all of the opening quotes are backwards, and some of the closing quotes are backwards. The paper could use a formatting check.\n\nWhile does not seem to be anything technically wrong with the paper and the results are fairly strong, the contributions are minimal. The techniques being used, like data filtering, SFT -> RL, and RLOO, are all well established in the literature at this point, and so there is not much new being added. Nonetheless, the results are strong, and so I lean towards weak accept.", "questions": "The error recovery component seems to be both very trivial to solve (hit a back button) and also somewhat damaging to performance (as evidenced by the occasionally higher performance of the InfiGUI method that doesn’t use error recovery in training). Does this suggest that that task is bad for helping to learn general GUI navigation? Or is there something else that explains why the method does better without error recovery?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce InfiGUI-R1, a vision-language-action model to interact with user interfaces and accomplish tasks. The model is trained in a 2-stage training pipeline, first with supervised fine tuning over a clean action dataset, and second with reinforcement learning over trajectories that are “just hard enough” so as to remain on the learning frontier of the model. The model is trained with several different objectives depending on the task at hand, varying from bounding-box IoU to sub-goal matching to action prediction. As part of the learning process, the authors also propose a sub-goal prediction/matching task, and an error correction task. These are both claimed to help improve final performance.\n\nInfiGUI-R1 is compared to much larger models on ScreenSpot benchmarks for grounding performance, and on AndroidControl and AndroidWorld benchmarks for interaction performance. The model achieves comparable or superior performance when compared to baselines, despite being smaller. The authors also conduct an ablation study in which they remove either their SFT phase or their RL phase, and they show that both are necessary to the success of the final model. Ablations on the subgoal and error recover tasks seem to indicate that they are less important to the overall success of the pipeline, with the “no error recovery” model even outperforming the full model along some metrics.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "+ The final performance of the InfiGUI-R1 model performs very well, achieving superior performance than larger models\n+ The proposed pipeline is intuitive and reflects many sensible design decisions, and would be easy to replicate by other researchers in their own workflows.\n+ The experiments evaluate InfiGUI-R1 along several important axes, showing results for different GUI instantiations (mobile/web/desktop), grounding performance on different tasks/applications, and task success on different benchmarks. These greatly help to illustrate the strengths of the model, and to contextualize it in the scope of related work.", "weaknesses": "- **Novelty**: SFT into RLVR is not a particularly new training paradigm. Every major LLM since ChatGPT has featured some combination of these two training paradigms in this order, and so the insight to apply SFT and then RLVR for GUI navigation is not particularly novel or insightful. In fact, one of the baselines in the paper, UI-TARS, employs an SFT-> DPO training pipeline, which is of course very similar. Similarly, many vision-language-action models in the robotics community have already demonstrated the success of SFT -> RLHF for task learning with VLMs.\n- Clarity: Elements of the paper are left unclear. For example, the authors introduce either a point-based reward or a bounding-box reward for certain UI tasks, but then there is no comparison between these two, or discussion on when to use each. This does not help future work to build on anything that the authors may have learned from experimenting with these rewards for UI navigation or grounding. Similarly, elements of the paper are left slightly unclear, such as how the Accuracy reward is normalized to [-1, 1], but sometimes it is set to the sum of 3 rewards ranging from [0, 1]… So it should range from [0, 3]. Is this reward then re-normalized to [-1, 1]? If so, is it a valid learning signal?\n- (Minor) Formatting: nearly all of the opening quotes are backwards, and some of the closing quotes are backwards. The paper could use a formatting check.\n\nWhile does not seem to be anything technically wrong with the paper and the results are fairly strong, the contributions are minimal. The techniques being used, like data filtering, SFT -> RL, and RLOO, are all well established in the literature at this point, and so there is not much new being added. Nonetheless, the results are strong, and so I lean towards weak accept.", "questions": "The error recovery component seems to be both very trivial to solve (hit a back button) and also somewhat damaging to performance (as evidenced by the occasionally higher performance of the InfiGUI method that doesn’t use error recovery in training). Does this suggest that that task is bad for helping to learn general GUI navigation? Or is there something else that explains why the method does better without error recovery?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760730041215}], "openreview_url": "https://openreview.net/forum?id=wywgRd1MUQ", "arxiv_id": "2504.14239", "paper_pdf": "papers/wywgRd1MUQ.pdf", "paper_pdf_sha256": "758780737e78b6c36c142526fc2960cbf5efbceee33925c9565c965e9ba46028", "paper_pdf_bytes": 525419, "paper_pdf_source": "openreview", "code_url": "https://github.com/InfiXAI/InfiGUI-R1", "code_repository": "InfiXAI/InfiGUI-R1", "code_commit": "a4fca17809a4395ba1fe08d481bb82c790ea7236", "code_archive": "repos/wywgRd1MUQ.zip", "code_archive_sha256": "8824c1100e804466b8ba3146a51cb4788dd217fff4dee0f3bda39c93ed49feeb", "code_archive_bytes": 1520968, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 1499, "github_languages": {"Python": 28852}, "github_archived": false, "github_pushed_at": "2025-12-04T13:36:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/infigui-r1-advancing-multimodal-gui-agents"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eqVu9eaVAB", "year": 2025, "status": "rejected", "title": "What Matters in Hierarchical Search for Combinatorial Reasoning Problems?", "authors": ["Michał Zawalski", "Gracjan Góral", "Michał Tyrolski", "Emilia Wiśnios", "Franciszek Budrowski", "Łukasz Kuciński", "Piotr Miłoś"], "authorids": ["~Michał_Zawalski1", "~Gracjan_Góral1", "~Michał_Tyrolski1", "~Emilia_Wiśnios1", "~Franciszek_Budrowski1", "~Łukasz_Kuciński1", "~Piotr_Miłoś1"], "authors_source": "OpenReview API", "abstract": "Combinatorial reasoning problems, particularly the notorious NP-hard tasks, remain a significant challenge for AI research. A common approach to addressing them combines search with learned heuristics. Recent methods in this domain utilize hierarchical planning, executing strategies based on subgoals. Our goal is to advance research in this area and establish a solid conceptual and empirical foundation. Specifically, we identify the following key obstacles, whose presence favors the choice of hierarchical search methods: _hard-to-learn value functions_, _complex action spaces_, _presence of dead ends in the environment_, or _data collected from diverse sources_. Through in-depth empirical analysis, we establish that hierarchical search methods consistently outperform standard search methods across these dimensions, and we formulate insights for future research. On the practical side, we also propose a consistent evaluation methodology to enable meaningful comparisons between methods and to reassess the state-of-the-art algorithms.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "H4eoSn2DLD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1057/Reviewer_sWxG"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper reports on an empirical evaluation of two types of algorithms for solving combinatorial search problems: “standard” graph search algorithms and “hierarchical search” algorithm. Worth noting that, unlike the classical literature on combinatorial search, here the main setup is that the algorithm does not have a given domain-specific heuristic. Instead the algorithm is  given a dataset of solutions to problems in the same domain, and can learn from this dataset how to solve other problems in the same domain. The “standard” graph search algorithms considered in this work learn a heuristic function (referred to as a value function here) and the “hierarchical search” algorithms considered in this work also learn a method to generate subgoals in this domain and performs a higher-level planning in the space of subgoals. \nThe results show that hierarchical search methods are often much better. The purpose of this work is to dig deeper and analyze why and when they are better. This is done by adding noise to the heuristic, considering test problems significantly different from those used for training, domains with dead-ends, and analyzing domains with particularly many actions. The results, very concisely show that hierarchical search is better in general, and in particular when the heuristic is weak, the training is very different from the test, and when there are many dead-ends or many actions.", "review_text": "This paper reports on an empirical evaluation of two types of algorithms for solving combinatorial search problems: “standard” graph search algorithms and “hierarchical search” algorithm. Worth noting that, unlike the classical literature on combinatorial search, here the main setup is that the algorithm does not have a given domain-specific heuristic. Instead the algorithm is  given a dataset of solutions to problems in the same domain, and can learn from this dataset how to solve other problems in the same domain. The “standard” graph search algorithms considered in this work learn a heuristic function (referred to as a value function here) and the “hierarchical search” algorithms considered in this work also learn a method to generate subgoals in this domain and performs a higher-level planning in the space of subgoals. \nThe results show that hierarchical search methods are often much better. The purpose of this work is to dig deeper and analyze why and when they are better. This is done by adding noise to the heuristic, considering test problems significantly different from those used for training, domains with dead-ends, and analyzing domains with particularly many actions. The results, very concisely show that hierarchical search is better in general, and in particular when the heuristic is weak, the training is very different from the test, and when there are many dead-ends or many actions.", "strengths": "1.\tThe problem is interesting, the evaluation is extensive. \n2.\tThe conclusions drawn from the evaluation are useful for future research and implementer. \n3.\tSome interesting insights into the nature of combinatorial search problems.", "weaknesses": "The paper is far from being self-contained. This can be viewed by the large number of appendices. To me, this work is more fitting for a journal than a conference paper. A major example for this is Theorem 1, which one cannot fully understand without reading the appendices. \nIn addition, the scope is maybe not a perfect fit for a top conference, as they mainly studied the existing algorithms and the conclusions are not ground breaking.\n\nPost-rebuttal: the authors provided some interesting theory in the discussion, I wish it was there in the paper when submitted. At this point, I am not sure if the addition of significant theory is reasonable during a rebuttal.", "questions": "1.\tIs my depiction above of the setup – no domain-specific heuristic, but yes data, is correct?\n2.\tAre the methods for detecting subgoals domain independent, i.e., only based on the given dataset, or do they also include domain-specific hard coded procedures. \n3.\tIn Section 3, the use of “planning” in the third bullet is not clear to me (and I’ve experience in planning). \n4.\tWhere you use the term value function, I think you mean state value heuristic? My guess is that if any, the term reward would be more accurate e(as per reward shaping)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper reports on an empirical evaluation of two types of algorithms for solving combinatorial search problems: “standard” graph search algorithms and “hierarchical search” algorithm. Worth noting that, unlike the classical literature on combinatorial search, here the main setup is that the algorithm does not have a given domain-specific heuristic. Instead the algorithm is  given a dataset of solutions to problems in the same domain, and can learn from this dataset how to solve other problems in the same domain. The “standard” graph search algorithms considered in this work learn a heuristic function (referred to as a value function here) and the “hierarchical search” algorithms considered in this work also learn a method to generate subgoals in this domain and performs a higher-level planning in the space of subgoals. \nThe results show that hierarchical search methods are often much better. The purpose of this work is to dig deeper and analyze why and when they are better. This is done by adding noise to the heuristic, considering test problems significantly different from those used for training, domains with dead-ends, and analyzing domains with particularly many actions. The results, very concisely show that hierarchical search is better in general, and in particular when the heuristic is weak, the training is very different from the test, and when there are many dead-ends or many actions.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1.\tThe problem is interesting, the evaluation is extensive. \n2.\tThe conclusions drawn from the evaluation are useful for future research and implementer. \n3.\tSome interesting insights into the nature of combinatorial search problems.", "weaknesses": "The paper is far from being self-contained. This can be viewed by the large number of appendices. To me, this work is more fitting for a journal than a conference paper. A major example for this is Theorem 1, which one cannot fully understand without reading the appendices. \nIn addition, the scope is maybe not a perfect fit for a top conference, as they mainly studied the existing algorithms and the conclusions are not ground breaking.\n\nPost-rebuttal: the authors provided some interesting theory in the discussion, I wish it was there in the paper when submitted. At this point, I am not sure if the addition of significant theory is reasonable during a rebuttal.", "questions": "1.\tIs my depiction above of the setup – no domain-specific heuristic, but yes data, is correct?\n2.\tAre the methods for detecting subgoals domain independent, i.e., only based on the given dataset, or do they also include domain-specific hard coded procedures. \n3.\tIn Section 3, the use of “planning” in the third bullet is not clear to me (and I’ve experience in planning). \n4.\tWhere you use the term value function, I think you mean state value heuristic? My guess is that if any, the term reward would be more accurate e(as per reward shaping)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730753075907}, {"id": "5YevQopqLg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1057/Reviewer_y8qH"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 5, "summary": "This paper examines the efficacy of hierarchical search methods in combinatorial reasoning problems compared to low-level search methods. It identifies specific conditions where hierarchical search excels due to the handling value function approximations and complex action spaces. The paper proposes standardized evaluation guidelines for search methods and reassesses state-of-the-art algorithms. The research highlights how hierarchical search methods can mitigate the challenges of large action spaces and variable training data.", "review_text": "This paper examines the efficacy of hierarchical search methods in combinatorial reasoning problems compared to low-level search methods. It identifies specific conditions where hierarchical search excels due to the handling value function approximations and complex action spaces. The paper proposes standardized evaluation guidelines for search methods and reassesses state-of-the-art algorithms. The research highlights how hierarchical search methods can mitigate the challenges of large action spaces and variable training data.", "strengths": "1. The paper offers a novel analysis of hierarchical versus low-level search methods in NP-hard problem domains, a less frequented area of research with significant implications for AI optimization tasks (Section 1). \n2. The paper's structure is logical and well-articulated. It presents complex methodologies and findings in a manner that is accessible to readers with a background in AI and machine learning. The diagrams and tables effectively illustrate key points and comparisons. \n3. Robust empirical analysis supports the claims, with extensive testing across various scenarios to demonstrate the superior performance of hierarchical methods under specific conditions. \n4. The findings contribute valuable insights into applying hierarchical search methods, particularly their resilience to value function noise and their efficiency in environments with complex action spaces, which are critical for advancing AI capabilities in real-world applications.", "weaknesses": "1. The study’s focus on specific NP-hard problems may limit the generalizability of the findings. Additional studies could help determine if the advantages of hierarchical methods hold across a broader range of combinatorial problems (General Discussion). \n2. The performance advantages of hierarchical methods heavily depend on the diversity and quality of training data. The paper could delve deeper into the effects of lower-quality or less diverse data sets (Section 5.1). \n3. While the paper addresses the effectiveness of hierarchical search methods, it underplays their computational complexity and resource demands, which could be a significant drawback in practical applications (Section 4). \n4. The robustness of hierarchical methods to training data variability could lead to overfitting if not adequately managed. The paper could discuss strategies to mitigate this risk (Section 5.2).\n5. A missing comparison with other foundation models in combinatorial problems, such as DeepCubeA, which are trained in an unsupervised setting.", "questions": "1. How adaptable are hierarchical search methods to other complex domains outside the specific NP-hard problems tested, such as real-time decision-making environments? \n2. How do computational constraints affect the performance of hierarchical search methods compared to low-level search methods in real-world applications? \n3. What improvements do the authors foresee or recommend for hierarchical search methods to handle larger datasets or more variable environments effectively? \n4. Could the authors elaborate on the standardized evaluation guidelines proposed? How do these guidelines compare to existing benchmarks in AI research?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper examines the efficacy of hierarchical search methods in combinatorial reasoning problems compared to low-level search methods. It identifies specific conditions where hierarchical search excels due to the handling value function approximations and complex action spaces. The paper proposes standardized evaluation guidelines for search methods and reassesses state-of-the-art algorithms. The research highlights how hierarchical search methods can mitigate the challenges of large action spaces and variable training data.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "1. The paper offers a novel analysis of hierarchical versus low-level search methods in NP-hard problem domains, a less frequented area of research with significant implications for AI optimization tasks (Section 1). \n2. The paper's structure is logical and well-articulated. It presents complex methodologies and findings in a manner that is accessible to readers with a background in AI and machine learning. The diagrams and tables effectively illustrate key points and comparisons. \n3. Robust empirical analysis supports the claims, with extensive testing across various scenarios to demonstrate the superior performance of hierarchical methods under specific conditions. \n4. The findings contribute valuable insights into applying hierarchical search methods, particularly their resilience to value function noise and their efficiency in environments with complex action spaces, which are critical for advancing AI capabilities in real-world applications.", "weaknesses": "1. The study’s focus on specific NP-hard problems may limit the generalizability of the findings. Additional studies could help determine if the advantages of hierarchical methods hold across a broader range of combinatorial problems (General Discussion). \n2. The performance advantages of hierarchical methods heavily depend on the diversity and quality of training data. The paper could delve deeper into the effects of lower-quality or less diverse data sets (Section 5.1). \n3. While the paper addresses the effectiveness of hierarchical search methods, it underplays their computational complexity and resource demands, which could be a significant drawback in practical applications (Section 4). \n4. The robustness of hierarchical methods to training data variability could lead to overfitting if not adequately managed. The paper could discuss strategies to mitigate this risk (Section 5.2).\n5. A missing comparison with other foundation models in combinatorial problems, such as DeepCubeA, which are trained in an unsupervised setting.", "questions": "1. How adaptable are hierarchical search methods to other complex domains outside the specific NP-hard problems tested, such as real-time decision-making environments? \n2. How do computational constraints affect the performance of hierarchical search methods compared to low-level search methods in real-world applications? \n3. What improvements do the authors foresee or recommend for hierarchical search methods to handle larger datasets or more variable environments effectively? \n4. Could the authors elaborate on the standardized evaluation guidelines proposed? How do these guidelines compare to existing benchmarks in AI research?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730720501132}, {"id": "UVr3G0nbGH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1057/Reviewer_iLMV"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper answers two primary questions: 1. Will hierarchical search outperform classical search under the same heuristic function and other settings? 2. In what situations do hierarchical search methods significantly outperform classical search? The authors conduct experiments to compare the performance of hierarchical search with classical search, finding that hierarchical search performs better in instances with complex action spaces and less guided heuristic functions. Additionally, data with high diversity also favors hierarchical search.\n\nI am uncertain whether the authors have addressed all possible influencing factors to arrive to the final conclusion. Since hierarchical search relies on a subgoal generator, which functions as another form of heuristic, this additional generator could significantly influence the improved performance of hierarchical search compared to classical search, especially in tests involving noisy value functions. This reliance on a subgoal generator may also be a critical factor that warrants further investigation.\n\nAnd the motivation behind this paper is also unclear. After answering these two questions, what potential impact could this research have on future algorithm design or other related fields?", "review_text": "This paper answers two primary questions: 1. Will hierarchical search outperform classical search under the same heuristic function and other settings? 2. In what situations do hierarchical search methods significantly outperform classical search? The authors conduct experiments to compare the performance of hierarchical search with classical search, finding that hierarchical search performs better in instances with complex action spaces and less guided heuristic functions. Additionally, data with high diversity also favors hierarchical search.\n\nI am uncertain whether the authors have addressed all possible influencing factors to arrive to the final conclusion. Since hierarchical search relies on a subgoal generator, which functions as another form of heuristic, this additional generator could significantly influence the improved performance of hierarchical search compared to classical search, especially in tests involving noisy value functions. This reliance on a subgoal generator may also be a critical factor that warrants further investigation.\n\nAnd the motivation behind this paper is also unclear. After answering these two questions, what potential impact could this research have on future algorithm design or other related fields?", "strengths": "Strengths:\n\n- The paper includes numerous experiments examining various factors that could impact the performance of algorithms.", "weaknesses": "Weaknesses:\n\n- The motivation for the study is unclear.\n\n- The presence of a subgoal generator could heavily influence the analysis of the experiments.", "questions": "Other Comments:\n\n- Line 52: \"low-level\" likely refers to hierarchical algorithms.\n\n- Line 59: The phrase \"hard-to-learn value functions\" requires clarification.\n\n- Line 83: There is a reference error.\n\n- Line 175: The description “Planner that determines the order in which subgoals are generated” seems to refer to the \"Subgoal generator.\"\n\n- Figure 7: Since both AdaSubS and kSubS rely on value functions, why do they not experience a significant drop in performance with high noise levels? For instance, kSubS in Sokoban (2 to 100 noise) shows better results.\n  \n- Figure 7: What is the value scale (minimum and maximum) for the value function, given that A* and BestFS tolerate a 0.5 error in the Rubik’s Cube test?\n\n- Figures: Some figures should be adjusted to enhance readability by increasing font sizes and displaying full labels, such as “Step” and “Subgoal” in Figures 8 and 9.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper answers two primary questions: 1. Will hierarchical search outperform classical search under the same heuristic function and other settings? 2. In what situations do hierarchical search methods significantly outperform classical search? The authors conduct experiments to compare the performance of hierarchical search with classical search, finding that hierarchical search performs better in instances with complex action spaces and less guided heuristic functions. Additionally, data with high diversity also favors hierarchical search.\n\nI am uncertain whether the authors have addressed all possible influencing factors to arrive to the final conclusion. Since hierarchical search relies on a subgoal generator, which functions as another form of heuristic, this additional generator could significantly influence the improved performance of hierarchical search compared to classical search, especially in tests involving noisy value functions. This reliance on a subgoal generator may also be a critical factor that warrants further investigation.\n\nAnd the motivation behind this paper is also unclear. After answering these two questions, what potential impact could this research have on future algorithm design or other related fields?", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Strengths:\n\n- The paper includes numerous experiments examining various factors that could impact the performance of algorithms.", "weaknesses": "Weaknesses:\n\n- The motivation for the study is unclear.\n\n- The presence of a subgoal generator could heavily influence the analysis of the experiments.", "questions": "Other Comments:\n\n- Line 52: \"low-level\" likely refers to hierarchical algorithms.\n\n- Line 59: The phrase \"hard-to-learn value functions\" requires clarification.\n\n- Line 83: There is a reference error.\n\n- Line 175: The description “Planner that determines the order in which subgoals are generated” seems to refer to the \"Subgoal generator.\"\n\n- Figure 7: Since both AdaSubS and kSubS rely on value functions, why do they not experience a significant drop in performance with high noise levels? For instance, kSubS in Sokoban (2 to 100 noise) shows better results.\n  \n- Figure 7: What is the value scale (minimum and maximum) for the value function, given that A* and BestFS tolerate a 0.5 error in the Rubik’s Cube test?\n\n- Figures: Some figures should be adjusted to enhance readability by increasing font sizes and displaying full labels, such as “Step” and “Subgoal” in Figures 8 and 9.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730681574192}, {"id": "eJepebGpSE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1057/Reviewer_nZJV"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper reviews hierarchical approaches to reasoning tasks, in particular comparing hierarchical search to low-level search. \n\nWhile the paper has no serious flaws, I am not entirely sure what to make of it. The insights the paper provides may be useful, but I am somewhat worried if they are contrived or overstated. In general, I think many would agree subgoal methods are worth pursuing, so arguing for this alone may not be entirely useful. At the same time, there are theoretical reasons why subgoal methods are imperfect, namely the upwards refinability property, e.g. that an optimal solution can be expressed in terms of abstractions. In general, subgoal methods may not be as flexible as a method which can use low-level actions (for an arbitrary problem at least), and to retain completeness a system needs to be able to use both abstract and primitive actions. Accordingly, I am somewhat skeptical of the takeaways, which seem to be based on limited empirical evaluation compared to what they claim. Regardless, I think the breakdown of benefits into the four separate takeaways is useful and overall think the paper is headed in the right direction.", "review_text": "This paper reviews hierarchical approaches to reasoning tasks, in particular comparing hierarchical search to low-level search. \n\nWhile the paper has no serious flaws, I am not entirely sure what to make of it. The insights the paper provides may be useful, but I am somewhat worried if they are contrived or overstated. In general, I think many would agree subgoal methods are worth pursuing, so arguing for this alone may not be entirely useful. At the same time, there are theoretical reasons why subgoal methods are imperfect, namely the upwards refinability property, e.g. that an optimal solution can be expressed in terms of abstractions. In general, subgoal methods may not be as flexible as a method which can use low-level actions (for an arbitrary problem at least), and to retain completeness a system needs to be able to use both abstract and primitive actions. Accordingly, I am somewhat skeptical of the takeaways, which seem to be based on limited empirical evaluation compared to what they claim. Regardless, I think the breakdown of benefits into the four separate takeaways is useful and overall think the paper is headed in the right direction.", "strengths": "The paper summarizes hierarchical reasoning methods and performs experiments to show their desirable properties. \n\nThe insights regarding robustness to noisy value functions are quite useful and seem to be well supported by the experiments.\n\nSimilarly, the breakdown of claims regarding complex action spaces, ergodicity, and data diversity is useful, but I am less certain about the validity of these claims, given that there are theoretical reasons why they may not be true.", "weaknesses": "I believe the related work section could be better organized, it didn't seem to flow with the rest of the paper so maybe it could be at the very end instead? e.g. Section 1 seems to better flow into Section 3. \n\nI think the paper could better distinguish between planning and learning -- many of the problems discussed are amenable to classical algorithms, so where does learning come in best and why?\n\nI think the paper would benefit from more theory -- while subgoal methods may have benefits in the experiments performed, their theoretical properties are not entirely discussed. In particular, issues such as the upwards refinability property make me skeptical of the second takeaway (line 417). In other sections, I think the paper would benefit from contrasting theory and experiment. \n\nIn general, the writing was somewhat difficult to follow and clarity could be improved.", "questions": "I don't have any specific questions about the paper -- but there are likely parts of it I did not fully understand.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper reviews hierarchical approaches to reasoning tasks, in particular comparing hierarchical search to low-level search. \n\nWhile the paper has no serious flaws, I am not entirely sure what to make of it. The insights the paper provides may be useful, but I am somewhat worried if they are contrived or overstated. In general, I think many would agree subgoal methods are worth pursuing, so arguing for this alone may not be entirely useful. At the same time, there are theoretical reasons why subgoal methods are imperfect, namely the upwards refinability property, e.g. that an optimal solution can be expressed in terms of abstractions. In general, subgoal methods may not be as flexible as a method which can use low-level actions (for an arbitrary problem at least), and to retain completeness a system needs to be able to use both abstract and primitive actions. Accordingly, I am somewhat skeptical of the takeaways, which seem to be based on limited empirical evaluation compared to what they claim. Regardless, I think the breakdown of benefits into the four separate takeaways is useful and overall think the paper is headed in the right direction.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The paper summarizes hierarchical reasoning methods and performs experiments to show their desirable properties. \n\nThe insights regarding robustness to noisy value functions are quite useful and seem to be well supported by the experiments.\n\nSimilarly, the breakdown of claims regarding complex action spaces, ergodicity, and data diversity is useful, but I am less certain about the validity of these claims, given that there are theoretical reasons why they may not be true.", "weaknesses": "I believe the related work section could be better organized, it didn't seem to flow with the rest of the paper so maybe it could be at the very end instead? e.g. Section 1 seems to better flow into Section 3. \n\nI think the paper could better distinguish between planning and learning -- many of the problems discussed are amenable to classical algorithms, so where does learning come in best and why?\n\nI think the paper would benefit from more theory -- while subgoal methods may have benefits in the experiments performed, their theoretical properties are not entirely discussed. In particular, issues such as the upwards refinability property make me skeptical of the second takeaway (line 417). In other sections, I think the paper would benefit from contrasting theory and experiment. \n\nIn general, the writing was somewhat difficult to follow and clarity could be improved.", "questions": "I don't have any specific questions about the paper -- but there are likely parts of it I did not fully understand.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729638713424}], "openreview_url": "https://openreview.net/forum?id=eqVu9eaVAB", "arxiv_id": "2406.03361", "paper_pdf": "papers/eqVu9eaVAB.pdf", "paper_pdf_sha256": "ce00deb1b508c3ddae286c259c1d1e872025438198475eaa0231468eb23d5bb6", "paper_pdf_bytes": 10711483, "paper_pdf_source": "openreview", "code_url": "https://github.com/subgoal-verification/what-matters-in-hierarchical-search", "code_repository": "subgoal-verification/what-matters-in-hierarchical-search", "code_commit": "ac0f72b23046d503ff40d75b43071cdf40fdfa31", "code_archive": "repos/eqVu9eaVAB.zip", "code_archive_sha256": "90aeff1d509c6d49fb0af18caf826877beab5fc77c2b37b360c7f9bf9781aa16", "code_archive_bytes": 72879, "code_file_count": 52, "code_extensions": {".py": 52}, "github_disk_usage_kb": 96, "github_languages": {"Python": 165544}, "github_archived": false, "github_pushed_at": "2024-05-18T14:48:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/what-matters-in-hierarchical-search-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WfjJOEfAf7", "year": 2024, "status": "rejected", "title": "Information Flow in Self-Supervised Learning", "authors": ["Zhiquan Tan", "Jingqin Yang", "Weiran Huang", "Yang Yuan", "Yifan Zhang"], "authorids": ["~Zhiquan_Tan1", "~Jingqin_Yang2", "~Weiran_Huang1", "~Yang_Yuan4", "~Yifan_Zhang16"], "authors_source": "OpenReview API", "abstract": "In this paper, we provide a comprehensive toolbox for understanding and enhancing self-supervised learning (SSL) methods through the lens of matrix information theory. Specifically, by leveraging the principles of matrix mutual information and joint entropy, we offer a unified analysis for both contrastive and non-contrastive methods. Furthermore, we propose the matrix variational masked auto-encoder (M-MAE) method, grounded in matrix information theory, as an enhancement to masked image modeling. The empirical evaluations underscore the effectiveness of M-MAE compared with the state-of-the-art methods, including a 3.9% improvement in linear probing ViT-Base, and a 1% improvement in fine-tuning ViT-Large, both on ImageNet.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "o4yV4lvD4F", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1221/Reviewer_znaA"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper discusses the information flow of three mainstream self-supervised learning methods: contrastive learning; feature de-correlation; and masked auto-encoding. It successfully connects all three methods with matrix information theory and thus offering a unified view. And beyond that it proposes to add an additional term to the original MAE loss. The loss regularizes the latent codes and is shown to be helpful on image classification tasks.", "review_text": "The paper discusses the information flow of three mainstream self-supervised learning methods: contrastive learning; feature de-correlation; and masked auto-encoding. It successfully connects all three methods with matrix information theory and thus offering a unified view. And beyond that it proposes to add an additional term to the original MAE loss. The loss regularizes the latent codes and is shown to be helpful on image classification tasks.", "strengths": "+ The paper introduces matrix information theory to understand and connect mainstream methods in self-supervised learning, which is a very meaningful and valuable contribution.\n+ The writing is fairly clear, although I did not delve into the mathematical details, I believe they are sounds.\n+ The initial results on image classification (both the linear-proving and the fine-tuning results) are great.", "weaknesses": "- While the initial empirical results are great, I do hope to see the final results after having a complete run on MAE. The current version of the paper uses U-MAE's implementation and the hyper-parameters (e.g., batch size) do not follow the settings in MAE. This can cause some discrepancies. MAE's ViT-L, after convergence, can achieve an accuracy of ~85.5 on ImageNet. While the paper's result is promising, it is unclear the trend can still hold. So I would be curious to see. If it is too much of a computation burden, I am fine to see results on CIFAR-100.\n- There are some definitions used before they are defined (e.g., TCR is defined in the appendix). It would be great to at least point to them.", "questions": "* The TCR loss on the latent codes, how does it contribute to the entropy of the model? Is there a similar plot one can show as the training of M-MAE proceeds to Figure 1/2?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper discusses the information flow of three mainstream self-supervised learning methods: contrastive learning; feature de-correlation; and masked auto-encoding. It successfully connects all three methods with matrix information theory and thus offering a unified view. And beyond that it proposes to add an additional term to the original MAE loss. The loss regularizes the latent codes and is shown to be helpful on image classification tasks.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "+ The paper introduces matrix information theory to understand and connect mainstream methods in self-supervised learning, which is a very meaningful and valuable contribution.\n+ The writing is fairly clear, although I did not delve into the mathematical details, I believe they are sounds.\n+ The initial results on image classification (both the linear-proving and the fine-tuning results) are great.", "weaknesses": "- While the initial empirical results are great, I do hope to see the final results after having a complete run on MAE. The current version of the paper uses U-MAE's implementation and the hyper-parameters (e.g., batch size) do not follow the settings in MAE. This can cause some discrepancies. MAE's ViT-L, after convergence, can achieve an accuracy of ~85.5 on ImageNet. While the paper's result is promising, it is unclear the trend can still hold. So I would be curious to see. If it is too much of a computation burden, I am fine to see results on CIFAR-100.\n- There are some definitions used before they are defined (e.g., TCR is defined in the appendix). It would be great to at least point to them.", "questions": "* The TCR loss on the latent codes, how does it contribute to the entropy of the model? Is there a similar plot one can show as the training of M-MAE proceeds to Figure 1/2?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699150464826}, {"id": "IJXhQcPogI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1221/Reviewer_cGgm"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors consider a theoretical framework to analyze and enhance self-supervised learning (SSL) methodologies utilizing matrix information theory. The work is particularly focused on providing a unified lens for examining both contrastive and feature decorrelation-based SSL paradigms through the application of matrix mutual information and joint entropy.\n\nSSL, a significant branch of unsupervised learning, leverages unlabeled data to learn representations by predicting certain input parts from others. The authors' investigation into the utility of matrix mutual information in SSL is notable. By extending mutual information to the matrix domain, the manuscript aims to elucidate the dependencies among various features or representations within SSL models, shedding light on the information propagation mechanisms within neural networks. Additionally, the manuscript's exploration of joint entropy in the context of SSL is insightful. Assessing how the uncertainty in the input data influences the learning process and the quality of the learned representations can be crucial for enhancing model robustness and efficiency.", "review_text": "The authors consider a theoretical framework to analyze and enhance self-supervised learning (SSL) methodologies utilizing matrix information theory. The work is particularly focused on providing a unified lens for examining both contrastive and feature decorrelation-based SSL paradigms through the application of matrix mutual information and joint entropy.\n\nSSL, a significant branch of unsupervised learning, leverages unlabeled data to learn representations by predicting certain input parts from others. The authors' investigation into the utility of matrix mutual information in SSL is notable. By extending mutual information to the matrix domain, the manuscript aims to elucidate the dependencies among various features or representations within SSL models, shedding light on the information propagation mechanisms within neural networks. Additionally, the manuscript's exploration of joint entropy in the context of SSL is insightful. Assessing how the uncertainty in the input data influences the learning process and the quality of the learned representations can be crucial for enhancing model robustness and efficiency.", "strengths": "**Theoretical Innovation**: The manuscript presents a novel theoretical framework for analyzing self-supervised learning (SSL) methods through the prism of matrix information theory. The use of matrix mutual information and joint entropy is an innovative approach that could provide new insights into the dependencies between features and the propagation of information within neural networks. This theoretical advancement has the potential to deepen our understanding of SSL mechanisms, making it a significant contribution to the field.\n\n**Unified Analysis for Diverse SSL Approaches**: By offering a unified analytical lens for both contrastive and feature decorrelation-based SSL paradigms, the paper bridges a gap in the current literature. This comprehensive approach allows for a more holistic understanding of SSL and its various implementations, enhancing the ability to compare and improve upon different methods within a common theoretical framework.\n\n**Potential for Enhanced Robustness and Efficiency**: The exploration of joint entropy in SSL models addresses the critical aspect of input data uncertainty. By theoretically examining how this uncertainty affects the learning process, the paper lays the groundwork for developing more robust and efficient SSL algorithms that can better handle real-world data variability.", "weaknesses": "**Scalability of Information-Theoretic Measures**: A potential weakness could be the lack of a clear discussion on the scalability of the proposed matrix mutual information and joint entropy measures. Calculating these metrics can be computationally intensive, especially for large-scale datasets and high-dimensional feature spaces typical in self-supervised learning. Any insights on the computational overheads is much appreciated", "questions": "**Generalization to Diverse Architectures**: Your paper appears to focus on a specific class of self-supervised learning models. How generalizable is your matrix information-theoretic approach to other SSL architectures, such as transformer-based or recurrent neural networks? Can you provide empirical evidence or theoretical justification for the generalizability of your approach?\n\n**Robustness and Sensitivity Analysis**: How robust are your matrix information-theoretic measures to variations in SSL hyperparameters, such as temperature in contrastive learning or weight decay? Could you provide a sensitivity analysis that examines the stability of your proposed metrics under different hyperparameter settings?\n\nThese questions are intended to probe the empirical validation of theoretical insights, the generalizability of the approach to various architectures, and the robustness of the proposed metrics to hyperparameter variations. Addressing these points could significantly strengthen the paper's contributions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors consider a theoretical framework to analyze and enhance self-supervised learning (SSL) methodologies utilizing matrix information theory. The work is particularly focused on providing a unified lens for examining both contrastive and feature decorrelation-based SSL paradigms through the application of matrix mutual information and joint entropy.\n\nSSL, a significant branch of unsupervised learning, leverages unlabeled data to learn representations by predicting certain input parts from others. The authors' investigation into the utility of matrix mutual information in SSL is notable. By extending mutual information to the matrix domain, the manuscript aims to elucidate the dependencies among various features or representations within SSL models, shedding light on the information propagation mechanisms within neural networks. Additionally, the manuscript's exploration of joint entropy in the context of SSL is insightful. Assessing how the uncertainty in the input data influences the learning process and the quality of the learned representations can be crucial for enhancing model robustness and efficiency.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "**Theoretical Innovation**: The manuscript presents a novel theoretical framework for analyzing self-supervised learning (SSL) methods through the prism of matrix information theory. The use of matrix mutual information and joint entropy is an innovative approach that could provide new insights into the dependencies between features and the propagation of information within neural networks. This theoretical advancement has the potential to deepen our understanding of SSL mechanisms, making it a significant contribution to the field.\n\n**Unified Analysis for Diverse SSL Approaches**: By offering a unified analytical lens for both contrastive and feature decorrelation-based SSL paradigms, the paper bridges a gap in the current literature. This comprehensive approach allows for a more holistic understanding of SSL and its various implementations, enhancing the ability to compare and improve upon different methods within a common theoretical framework.\n\n**Potential for Enhanced Robustness and Efficiency**: The exploration of joint entropy in SSL models addresses the critical aspect of input data uncertainty. By theoretically examining how this uncertainty affects the learning process, the paper lays the groundwork for developing more robust and efficient SSL algorithms that can better handle real-world data variability.", "weaknesses": "**Scalability of Information-Theoretic Measures**: A potential weakness could be the lack of a clear discussion on the scalability of the proposed matrix mutual information and joint entropy measures. Calculating these metrics can be computationally intensive, especially for large-scale datasets and high-dimensional feature spaces typical in self-supervised learning. Any insights on the computational overheads is much appreciated", "questions": "**Generalization to Diverse Architectures**: Your paper appears to focus on a specific class of self-supervised learning models. How generalizable is your matrix information-theoretic approach to other SSL architectures, such as transformer-based or recurrent neural networks? Can you provide empirical evidence or theoretical justification for the generalizability of your approach?\n\n**Robustness and Sensitivity Analysis**: How robust are your matrix information-theoretic measures to variations in SSL hyperparameters, such as temperature in contrastive learning or weight decay? Could you provide a sensitivity analysis that examines the stability of your proposed metrics under different hyperparameter settings?\n\nThese questions are intended to probe the empirical validation of theoretical insights, the generalizability of the approach to various architectures, and the robustness of the proposed metrics to hyperparameter variations. Addressing these points could significantly strengthen the paper's contributions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698932677248}, {"id": "8nUWxxuQkw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1221/Reviewer_qQv9"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper shows that the optimal point of BarlowTwins and Spectral Contrastive learning objective functions satisfy the maximal matrix mutual information and the maximal matrix joint entropy. Then, this paper proposes a \"matrix variational masked auto-encoder (M-MAE) loss\" that is a combination of the original loss and the total coding rate, defined as $\\log det (\\mu I + ZZ^T)$, where $\\mu$ is a hyperparameter. Experimental results show that compared to MAE and U-MAE, the proposed M-MAE performs better in terms of linear probing and fine-tuning when the hyperparameter $\\mu$ is well-tuned.", "review_text": "This paper shows that the optimal point of BarlowTwins and Spectral Contrastive learning objective functions satisfy the maximal matrix mutual information and the maximal matrix joint entropy. Then, this paper proposes a \"matrix variational masked auto-encoder (M-MAE) loss\" that is a combination of the original loss and the total coding rate, defined as $\\log det (\\mu I + ZZ^T)$, where $\\mu$ is a hyperparameter. Experimental results show that compared to MAE and U-MAE, the proposed M-MAE performs better in terms of linear probing and fine-tuning when the hyperparameter $\\mu$ is well-tuned.", "strengths": "This paper attempts to understand self-supervised learning methods in terms of the mutual information maximization framework, where each random variable is from the online encoder and the target encoder. It could be a somewhat valuable attempt to understand how self-supervised learning methods work.", "weaknesses": "1. There is no detailed discussion between the introduced matrix entropy (definition 1) and the original Shannon's information entropy. In fact, The Renyi entropy is defined by a special matrix family named density matrix. It is not defined for an arbitrary matrix, but only for a Gram matrix. However, this paper misuses the concept of matrix entropy throughout the whole paper. For example, In page 14, the last paragraph says that \"Take K1 = Z1 and K2 = Z2, the results follow similarly\". However, K1 and K2 should be Gram matrices, while Z1 and Z2 are feature matrices, that are not a Gram matrix. I think this paper should clarify the relationship between the matrix entropy (defined by a Gram matrix of the samples from a probability distribution) and the original Shannon's information entropy (defined by a random variable following a probability density function).\n2. There is no connection between Section 3 and 4. Note that the main difference between MAE and M-MAE is $TCR(Z)$. TCR is used for measuring a joint information quantity in Section 3, but in Section 4, this paper uses TCR for measuring information quantity for a single random variable. As the previous discussions are based on the relationship between Z1 and Z2, the newly introduced regularization term for MAE is irrelevant to the previous results.\n3. The proposed M-MAE is sensitive to the choice of the hyperparameter, as shown in Table 2. The gap between each $\\mu$ varies a lot, and it means that we need to access the original target labels to tune the hyperparemeter. It violates the spirit of self-supervised learning; we should not access the original labels\n4. The results in Section 3 are not generalizable to the generic self-supervised learning methods; these results are only applicable to Barlow twins and spectral contrastive learning. In fact, as mentioned in my previous comment, the proof is wrong for spectral contrastive learning because K1 and K2 should be a Gram matrix. In other words, the proof only works for a special case of self-supervised learning, where the objective function coincides with \"Proposition 1\" and K1 and K2 are Gram matrices of the feature matrices Z1 and Z2.\n5. There is no discussion of why the mutual information maximization (or joint entropy) of Z1, Z2 is a good measure of a good self-supervised learning method. As there is no connection between mutual information maximization and goodness of self-supervised learning methods, the motivation of M-MAE is somewhat weak. What is the benefit of making an MAE model maximize mutual information? (Note that, even more it is actually not about mutual information. See comment 2)\n6. The experimental results only show the comparisons between MAE, U-MAE and M-MAE. There are a lot of self-supervised learning methods. I think this paper needs more comparisons with other self-supervised learning methods (e.g., BalowTwins, MoCo, SimCLR, BYOL, DeepClustering, Swav, Data2Vec, DINO, iBot, SimMIM, ...) in terms of both information quantity and performance. I also think that it would be good for this paper to compare with other MAE variants (or MIM methods, such as SimMIM), but it could depend on the scope of this paper; as I think this paper needs a heavy non-trivial revision, as of now, I don't argue that M-MAE should be compared with other MAE variants, but I think additional comparisons with MAE variants will make the submission stronger. I recommend this survey paper to search more recent MAE variants: \"A Survey on Masked Autoencoder for Self-supervised Learning in Vision and Beyond\"", "questions": "Please check my previous comment. I think the current version of this paper will need a non-trivial heavy revision, including re-checking the major motivation (W2, W5), the mathematical notations and theoretical results (W1, W2, W4), adding more experiments (W6), fixing the fundamental flaw -- hyperparameter sensitivity -- of the proposed method (W3).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper shows that the optimal point of BarlowTwins and Spectral Contrastive learning objective functions satisfy the maximal matrix mutual information and the maximal matrix joint entropy. Then, this paper proposes a \"matrix variational masked auto-encoder (M-MAE) loss\" that is a combination of the original loss and the total coding rate, defined as $\\log det (\\mu I + ZZ^T)$, where $\\mu$ is a hyperparameter. Experimental results show that compared to MAE and U-MAE, the proposed M-MAE performs better in terms of linear probing and fine-tuning when the hyperparameter $\\mu$ is well-tuned.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This paper attempts to understand self-supervised learning methods in terms of the mutual information maximization framework, where each random variable is from the online encoder and the target encoder. It could be a somewhat valuable attempt to understand how self-supervised learning methods work.", "weaknesses": "1. There is no detailed discussion between the introduced matrix entropy (definition 1) and the original Shannon's information entropy. In fact, The Renyi entropy is defined by a special matrix family named density matrix. It is not defined for an arbitrary matrix, but only for a Gram matrix. However, this paper misuses the concept of matrix entropy throughout the whole paper. For example, In page 14, the last paragraph says that \"Take K1 = Z1 and K2 = Z2, the results follow similarly\". However, K1 and K2 should be Gram matrices, while Z1 and Z2 are feature matrices, that are not a Gram matrix. I think this paper should clarify the relationship between the matrix entropy (defined by a Gram matrix of the samples from a probability distribution) and the original Shannon's information entropy (defined by a random variable following a probability density function).\n2. There is no connection between Section 3 and 4. Note that the main difference between MAE and M-MAE is $TCR(Z)$. TCR is used for measuring a joint information quantity in Section 3, but in Section 4, this paper uses TCR for measuring information quantity for a single random variable. As the previous discussions are based on the relationship between Z1 and Z2, the newly introduced regularization term for MAE is irrelevant to the previous results.\n3. The proposed M-MAE is sensitive to the choice of the hyperparameter, as shown in Table 2. The gap between each $\\mu$ varies a lot, and it means that we need to access the original target labels to tune the hyperparemeter. It violates the spirit of self-supervised learning; we should not access the original labels\n4. The results in Section 3 are not generalizable to the generic self-supervised learning methods; these results are only applicable to Barlow twins and spectral contrastive learning. In fact, as mentioned in my previous comment, the proof is wrong for spectral contrastive learning because K1 and K2 should be a Gram matrix. In other words, the proof only works for a special case of self-supervised learning, where the objective function coincides with \"Proposition 1\" and K1 and K2 are Gram matrices of the feature matrices Z1 and Z2.\n5. There is no discussion of why the mutual information maximization (or joint entropy) of Z1, Z2 is a good measure of a good self-supervised learning method. As there is no connection between mutual information maximization and goodness of self-supervised learning methods, the motivation of M-MAE is somewhat weak. What is the benefit of making an MAE model maximize mutual information? (Note that, even more it is actually not about mutual information. See comment 2)\n6. The experimental results only show the comparisons between MAE, U-MAE and M-MAE. There are a lot of self-supervised learning methods. I think this paper needs more comparisons with other self-supervised learning methods (e.g., BalowTwins, MoCo, SimCLR, BYOL, DeepClustering, Swav, Data2Vec, DINO, iBot, SimMIM, ...) in terms of both information quantity and performance. I also think that it would be good for this paper to compare with other MAE variants (or MIM methods, such as SimMIM), but it could depend on the scope of this paper; as I think this paper needs a heavy non-trivial revision, as of now, I don't argue that M-MAE should be compared with other MAE variants, but I think additional comparisons with MAE variants will make the submission stronger. I recommend this survey paper to search more recent MAE variants: \"A Survey on Masked Autoencoder for Self-supervised Learning in Vision and Beyond\"", "questions": "Please check my previous comment. I think the current version of this paper will need a non-trivial heavy revision, including re-checking the major motivation (W2, W5), the mathematical notations and theoretical results (W1, W2, W4), adding more experiments (W6), fixing the fundamental flaw -- hyperparameter sensitivity -- of the proposed method (W3).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698678629882}, {"id": "fe0xTw4jIZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1221/Reviewer_j36B"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper aims at providing better understanding for existing methods (contrastive and feature decorrelation based methods) by leveraging the principles of matrix mutual information and joint entropy. In addition, the paper proposes the \"matrix variational masked auto-encoder\" (M-MAE) method. The paper reports empircal results that show the effectiveness of M-MAE compared with the state-of-the-art methods for representation learning on ImageNet.", "review_text": "This paper aims at providing better understanding for existing methods (contrastive and feature decorrelation based methods) by leveraging the principles of matrix mutual information and joint entropy. In addition, the paper proposes the \"matrix variational masked auto-encoder\" (M-MAE) method. The paper reports empircal results that show the effectiveness of M-MAE compared with the state-of-the-art methods for representation learning on ImageNet.", "strengths": "- The paper tackles important questions in the field of self-supervised learning", "weaknesses": "- The proofs are not always clear/complete (see questions below).\n- There are propositions and theorems in the paper. Then in the appendix, there are only proofs to \"theorems\" and the reader must match correct theorem/proposition with the correct proof.\n- The experimental setup lacks many details.", "questions": "- Theorems 1 and 2 do not seem to be proved clearly. The proof to theorem 1 in the appendix (the one that includes \"lemma 1\"!) does not seem to proof things by following a clear mathematical reasoning. Can you clarify how the last sentences leads to a valid proof? It it the same for Theorem 2 (the proof on page 15 given the other one seems to refer to Proposition 2).\n- Can you provide additional technical details about the experimental setup. The appendix does not seem to contain any information related to that and only limited information is given in the main paper (what are exactly the training objectives and hyper-parameters such as batch size, etc.).\n- It is not fully clear how Theorems 1 and 2 are not going against each other given Definition 2. Can you provide some information about that?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims at providing better understanding for existing methods (contrastive and feature decorrelation based methods) by leveraging the principles of matrix mutual information and joint entropy. In addition, the paper proposes the \"matrix variational masked auto-encoder\" (M-MAE) method. The paper reports empircal results that show the effectiveness of M-MAE compared with the state-of-the-art methods for representation learning on ImageNet.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper tackles important questions in the field of self-supervised learning", "weaknesses": "- The proofs are not always clear/complete (see questions below).\n- There are propositions and theorems in the paper. Then in the appendix, there are only proofs to \"theorems\" and the reader must match correct theorem/proposition with the correct proof.\n- The experimental setup lacks many details.", "questions": "- Theorems 1 and 2 do not seem to be proved clearly. The proof to theorem 1 in the appendix (the one that includes \"lemma 1\"!) does not seem to proof things by following a clear mathematical reasoning. Can you clarify how the last sentences leads to a valid proof? It it the same for Theorem 2 (the proof on page 15 given the other one seems to refer to Proposition 2).\n- Can you provide additional technical details about the experimental setup. The appendix does not seem to contain any information related to that and only limited information is given in the main paper (what are exactly the training objectives and hyper-parameters such as batch size, etc.).\n- It is not fully clear how Theorems 1 and 2 are not going against each other given Definition 2. Can you provide some information about that?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698669834133}], "openreview_url": "https://openreview.net/forum?id=WfjJOEfAf7", "arxiv_id": "2309.17281", "paper_pdf": "papers/WfjJOEfAf7.pdf", "paper_pdf_sha256": "127547cea9ad41185d9cce04fdfa35a815969bee45335bd68d65ea6f5a38efcd", "paper_pdf_bytes": 440808, "paper_pdf_source": "openreview", "code_url": "https://github.com/yifanzhang-pro/M-MAE", "code_repository": "yifanzhang-pro/M-MAE", "code_commit": "22dc5b248b3aa1788568d9b592a51ece9d28c847", "code_archive": "repos/WfjJOEfAf7.zip", "code_archive_sha256": "c8913c4f10624aa2949ad88305df541ae09bfa2f52095622bf0d69e0fbaf09a5", "code_archive_bytes": 46775, "code_file_count": 21, "code_extensions": {".py": 18, ".sh": 3}, "github_disk_usage_kb": 76, "github_languages": {"Python": 100780, "Shell": 1555}, "github_archived": false, "github_pushed_at": "2024-09-17T06:47:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/information-flow-in-self-supervised-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bTy4D3KHwWU", "year": 2023, "status": "rejected", "title": "Normalizing Flows for Interventional Density Estimation", "authors": ["Valentyn Melnychuk", "Dennis Frauen", "Stefan Feuerriegel"], "authorids": ["~Valentyn_Melnychuk1", "~Dennis_Frauen1", "~Stefan_Feuerriegel1"], "authors_source": "OpenReview API", "abstract": "Existing machine learning methods for causal inference usually estimate quantities expressed via the mean of potential outcomes (e.g., average treatment effect). However, such quantities do not capture the full information about the distribution of potential outcomes. In this work, we estimate the density of potential outcomes after interventions from observational data. For this, we propose a novel, fully-parametric deep learning method called Interventional Normalizing Flows. Specifically, we combine two normalizing flows, namely (i) a teacher flow for estimating nuisance parameters and (ii) a student flow for a parametric estimation of the density of potential outcomes. We further develop a tractable optimization objective via a one-step bias correction for an efficient and doubly robust estimation of the student flow parameters. As a result our Interventional Normalizing Flows offer a properly normalized density estimator.  Across various experiments, we demonstrate that our Interventional Normalizing Flows are expressive and highly effective, and scale well with both sample size and high-dimensional confounding. To the best of our knowledge, our Interventional Normalizing Flows are the first fully-parametric, deep learning method for density estimation of potential outcomes.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "pkHvXvaWAE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2714/Reviewer_A7uk"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors propose a fully-parametric, deep-learning method for interventional density estimation, called Interventional Normalizing Flows (INFs). INFs provide a properly normalized density estimator. The authors further develop a two-step training procedure with a one-step bias correction for efficient and doubly robust estimation.", "review_text": "The method proposed in this work is novel. The authors estimate the density of potential outcomes after interventions from observational data and propose a fully-parametric deep learning method INF. The authors also provide extensive numerical studies to support the performance of INF. ", "strengths": "To improve the readability, the authors may want to: \n1. Add more introduction to the structural causal model.\n2. In the example provided in Introduction Section,  explain interventional distributions, observational distributions, and counterfactual distributions. Explicitly define the notation Y [a].\n\n\nQuestion: Why does Table 3 compare % of runs with best performances, while other tables compare the log-probability? And why are the columns for Table 3 named log-prob?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors propose a fully-parametric, deep-learning method for interventional density estimation, called Interventional Normalizing Flows (INFs). INFs provide a properly normalized density estimator. The authors further develop a two-step training procedure with a one-step bias correction for efficient and doubly robust estimation.", "strength_and_weaknesses": "To improve the readability, the authors may want to: \n1. Add more introduction to the structural causal model.\n2. In the example provided in Introduction Section,  explain interventional distributions, observational distributions, and counterfactual distributions. Explicitly define the notation Y [a].\n\n\nQuestion: Why does Table 3 compare % of runs with best performances, while other tables compare the log-probability? And why are the columns for Table 3 named log-prob?\n", "clarity,_quality,_novelty_and_reproducibility": "The authors did extensive experiments and showed their methods could outperform other methods. But the authors can improve the introduction to audiences unfamiliar with causal inference and the normalizing flow method. ", "summary_of_the_review": "The method proposed in this work is novel. The authors estimate the density of potential outcomes after interventions from observational data and propose a fully-parametric deep learning method INF. The authors also provide extensive numerical studies to support the performance of INF. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666860224171}, {"id": "W6bLZk-6_-", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2714/Reviewer_uUEZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper introduces an optimization framework with a model consisting of a teacher and student normalizing flows for estimating the interventional distribution for the case of backdoor adjustment. The developed optimization uses one-step bias correction and utilizes influence functions for doubly robust estimation of the interventional density. Results on multiple synthetic and semi-synthetic datasets are reported showing that INFs outperform nonparametric and ablated methods.", "review_text": "The paper has good quality, including further information (summarized under comments) can make it stronger.", "strengths": "**Strengths**\n* The presented methods fill a gap in the literature on estimating interventional densities using parametric models. Therefore the problem considered is timely and significant.\n* The motivation is clear as the example in Fig. 1 is representative of why the problem considered is important.\n* I found the schematic in Fig. 2 very helpful, it provides a nice summary of the components of the model.\n* The presentation and the logical flow of the paper are clear, I specifically appreciated the use of colors in the text for determining what corresponds to nuisance parameters and what does not.\n* The background contains a nice summary of the existing literature for estimating interventional effects for the case of backdoor adjustment. The literature has focused on average treatment effect (ATE) and the non-ATE literature is mostly non-parametric or non-algorithmic.\n\n**Comments**\n* The problem of estimating $\\mathbb{E}_{X \\sim P(X)} \\big[ P(Y|X, A) \\big]$ given samples from $X, Y, A$ sounds like a well studied estimation problem in statistics regardless of the context of interventional distributions (here, the interventional or counterfactual data is not used and we are interested in the above estimation problem). To me, it is surprising that Normalizing Flows are not used for this purpose before so I cannot confirm the novelty of the methodology. Given that (conditional) normalizing flows, the theory of semi-parametric estimation of the interventional distribution, doubly robust estimation tools existed before the main contribution seems to lie in the addition of the student network, and framing and solving the optimization problem.\n\n* Given that the work is applicable only to the backdoor adjustment, even for the cases in which we can identify the effect of intervention we cannot use this method. However, the extension to the more general case sounds feasible. This would largely broaden the impact and scope of the paper. Can the authors discuss the possibility of this extension?\n\n* What happens if $Y$ is not one-dimensional? It seems that the method relies on creating a grid on $\\mathcal{Y}$. Is this applicable and does this scale to higher dimensional $Y$ variables?\n\n* Can the authors explain why in the case of the synthetic dataset in Fig. 3 KDE is doing very well (even better than the proposed method) for $a=0$ but fails for $a=1$? Where does this asymmetry come from?\n\n* Can the authors include the SCMs used for other examples in the supplementary? \n\n* I think the contributions of this paper by themselves are enough for the acceptance of the paper if these techniques are not used in the context of causal inference and interventional density estimation before. I am happy to change my score to acceptance if the authors include the above information.\n\n* Can the authors include the famous tobacco control example in the experiments?\n\n* To make this a standalone paper, I encourage the authors to include a short summary of the influence functions and doubly robust estimation. More specifically, to me, it wasn't quite clear how we go from Eq. 5 to 7, 8.\n\n\n**Post rebuttal**\n\nI thank the authors for clarifying their contributions and distinguishing between the original estimation theory and the semi-parametric giving rise to the optimization framework and the suggested student-teacher architecture. Now it's clear why the classical density estimation techniques aren't as powerful as semi-parametric methods and why finite-dimensional estimands are not applicable. I think the new additions (adding more background on semi-parametric theory and SCMs, including the SCMs associated with the experiments, adding results on the tobacco control experiment) make the paper more standalone and powerful. Hence I changed my score to acceptance.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper introduces an optimization framework with a model consisting of a teacher and student normalizing flows for estimating the interventional distribution for the case of backdoor adjustment. The developed optimization uses one-step bias correction and utilizes influence functions for doubly robust estimation of the interventional density. Results on multiple synthetic and semi-synthetic datasets are reported showing that INFs outperform nonparametric and ablated methods.", "strength_and_weaknesses": "**Strengths**\n* The presented methods fill a gap in the literature on estimating interventional densities using parametric models. Therefore the problem considered is timely and significant.\n* The motivation is clear as the example in Fig. 1 is representative of why the problem considered is important.\n* I found the schematic in Fig. 2 very helpful, it provides a nice summary of the components of the model.\n* The presentation and the logical flow of the paper are clear, I specifically appreciated the use of colors in the text for determining what corresponds to nuisance parameters and what does not.\n* The background contains a nice summary of the existing literature for estimating interventional effects for the case of backdoor adjustment. The literature has focused on average treatment effect (ATE) and the non-ATE literature is mostly non-parametric or non-algorithmic.\n\n**Comments**\n* The problem of estimating $\\mathbb{E}_{X \\sim P(X)} \\big[ P(Y|X, A) \\big]$ given samples from $X, Y, A$ sounds like a well studied estimation problem in statistics regardless of the context of interventional distributions (here, the interventional or counterfactual data is not used and we are interested in the above estimation problem). To me, it is surprising that Normalizing Flows are not used for this purpose before so I cannot confirm the novelty of the methodology. Given that (conditional) normalizing flows, the theory of semi-parametric estimation of the interventional distribution, doubly robust estimation tools existed before the main contribution seems to lie in the addition of the student network, and framing and solving the optimization problem.\n\n* Given that the work is applicable only to the backdoor adjustment, even for the cases in which we can identify the effect of intervention we cannot use this method. However, the extension to the more general case sounds feasible. This would largely broaden the impact and scope of the paper. Can the authors discuss the possibility of this extension?\n\n* What happens if $Y$ is not one-dimensional? It seems that the method relies on creating a grid on $\\mathcal{Y}$. Is this applicable and does this scale to higher dimensional $Y$ variables?\n\n* Can the authors explain why in the case of the synthetic dataset in Fig. 3 KDE is doing very well (even better than the proposed method) for $a=0$ but fails for $a=1$? Where does this asymmetry come from?\n\n* Can the authors include the SCMs used for other examples in the supplementary? \n\n* I think the contributions of this paper by themselves are enough for the acceptance of the paper if these techniques are not used in the context of causal inference and interventional density estimation before. I am happy to change my score to acceptance if the authors include the above information.\n\n* Can the authors include the famous tobacco control example in the experiments?\n\n* To make this a standalone paper, I encourage the authors to include a short summary of the influence functions and doubly robust estimation. More specifically, to me, it wasn't quite clear how we go from Eq. 5 to 7, 8.\n\n\n**Post rebuttal**\n\nI thank the authors for clarifying their contributions and distinguishing between the original estimation theory and the semi-parametric giving rise to the optimization framework and the suggested student-teacher architecture. Now it's clear why the classical density estimation techniques aren't as powerful as semi-parametric methods and why finite-dimensional estimands are not applicable. I think the new additions (adding more background on semi-parametric theory and SCMs, including the SCMs associated with the experiments, adding results on the tobacco control experiment) make the paper more standalone and powerful. Hence I changed my score to acceptance.", "clarity,_quality,_novelty_and_reproducibility": "The paper is mostly clear, some background on influence functions and doubly robust estimation can make the paper stronger. The theoretical framework of the presented method existed before, the main novelty lies in the algorithmic instantiation and the performance of the presented methods. The information in the paper sounds sufficient for the reproduction of the results, however, I did not try to reproduce the results myself. ", "summary_of_the_review": "The paper has good quality, including further information (summarized under comments) can make it stronger.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666741939215}, {"id": "DzdO-TimGf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2714/Reviewer_quXx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers an interesting problem in interventional causal inference, which is estimating the distribution of a potential outcome corresponding to an intervention. It proposes a fully parametric deep learning model for this task called Interventional Normalizing Flow that uses 2 flows. \n\nThe teacher flow computes the propensity score $\\pi\\_{a}(X)$ from $X$ and parameters of a conditional normalizing flow network (CNF) from $(X, A)$. This conditional flow network maps Gaussian noise to $P(Y|X, A)$.\n\nThe student flow is a set of unconditional normalizing flows for each intervention $A=a$. Each unconditional flow maps Gaussian noise to $g(Y, \\beta\\_a)$ where $\\beta\\_a$ is the set of parameters corresponds to $A=a$. Each unconditional flow w.r.t. $A=a$ is learned by mimicking $P(Y|do(A=a))$ obtained from the teacher flow. This problem is transformed into an equivalent problem of solving moment conditions. This paper also introduces an (optional) one-step bias correction for doubly robust estimation of $\\beta\\_a$ by leveraging the propensity score prediction from the teacher flow.\n", "review_text": "Overall, this is a good paper which proposes a novel method to address an interesting problem. The only problem is that the model is more complex than it should be. This can limit the practical applications of the work. I give this a paper a weak accept.", "strengths": "### Strengths\n1) The problem considered in the paper is interesting and the method is novel.\n2) The paper is well-written with a good discussion of related works especially those related to interventional density estimation and normalizing flows for causal inference (Appendix A). I also appreciate the authors' attempt to make things clear for the readers by using colors in their formulas.\n3) I think the experiment is quite extensive with various baselines for density estimation. The authors also compare different variants of their method.\n\n### Weaknesses\n1) The main weakness is the complexity of this method as in my opinion, it can be much simpler. For example, the student flow seems redundant as it just approximates $P(Y|do(A=a))$ though I know the authors want to avoid taking an average over $X$ during testing. In the ablation studies, the authors also consider a variant of the proposed method without the student flow, and when looking at the results in Table 2, I saw that this variant is not too bad compared to the full model.\n2) Please correct me if I am wrong but I think we can use simple methods like Balancing Linear Regression [1], TARNet [2], DragonNet [3], … for interventional density estimation (IDE) problem in the paper. These methods are mainly designed for ITE estimation but we can easily modify their outputs to model $P(Y|X, A)$. Once we can model P(Y|X, A), we can simply estimate $P(Y|do(A)) = \\mathbb{E}_{p(X)}[P(Y|X, a)]$ via backdoor adjustment. I would like to hear the authors’ opinions about this.\n3) The authors should compare their method with existing ITE estimation methods to support their argument in the paper saying that “estimating the potential outcome density is better than just estimating the average potential outcome“.\n4) In the TeacherFlow, I would like to understand why did the authors use $X, A$ to compute the weight $\\theta$ of the conditional normalizing flow? Why don’t just model $P(Y|X, A)$ directly via a neural network with $X, A$ are inputs to this network? Since the distribution of outcomes $Y$ is usually simple. Do we really need a normalizing flow with hyper-network for its parameters?\n5) The authors should provide more insights and explanations about why each component is useful for the method to work. Currently, they only use one metric, which is the log-likelihood, for comparison, which I think is not enough.\n\n[1] Learning Representations for Counterfactual Inference, Johansson et al., ICML-2016\n\n[2] Estimating individual treatment effect: generalization bounds and algorithms, Shalit et al., ICML-2017\n\n[3] Adapting Neural Networks for the Estimation of Treatment Effects, Shi et al., NIPS-2019\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers an interesting problem in interventional causal inference, which is estimating the distribution of a potential outcome corresponding to an intervention. It proposes a fully parametric deep learning model for this task called Interventional Normalizing Flow that uses 2 flows. \n\nThe teacher flow computes the propensity score $\\pi\\_{a}(X)$ from $X$ and parameters of a conditional normalizing flow network (CNF) from $(X, A)$. This conditional flow network maps Gaussian noise to $P(Y|X, A)$.\n\nThe student flow is a set of unconditional normalizing flows for each intervention $A=a$. Each unconditional flow maps Gaussian noise to $g(Y, \\beta\\_a)$ where $\\beta\\_a$ is the set of parameters corresponds to $A=a$. Each unconditional flow w.r.t. $A=a$ is learned by mimicking $P(Y|do(A=a))$ obtained from the teacher flow. This problem is transformed into an equivalent problem of solving moment conditions. This paper also introduces an (optional) one-step bias correction for doubly robust estimation of $\\beta\\_a$ by leveraging the propensity score prediction from the teacher flow.\n", "strength_and_weaknesses": "### Strengths\n1) The problem considered in the paper is interesting and the method is novel.\n2) The paper is well-written with a good discussion of related works especially those related to interventional density estimation and normalizing flows for causal inference (Appendix A). I also appreciate the authors' attempt to make things clear for the readers by using colors in their formulas.\n3) I think the experiment is quite extensive with various baselines for density estimation. The authors also compare different variants of their method.\n\n### Weaknesses\n1) The main weakness is the complexity of this method as in my opinion, it can be much simpler. For example, the student flow seems redundant as it just approximates $P(Y|do(A=a))$ though I know the authors want to avoid taking an average over $X$ during testing. In the ablation studies, the authors also consider a variant of the proposed method without the student flow, and when looking at the results in Table 2, I saw that this variant is not too bad compared to the full model.\n2) Please correct me if I am wrong but I think we can use simple methods like Balancing Linear Regression [1], TARNet [2], DragonNet [3], … for interventional density estimation (IDE) problem in the paper. These methods are mainly designed for ITE estimation but we can easily modify their outputs to model $P(Y|X, A)$. Once we can model P(Y|X, A), we can simply estimate $P(Y|do(A)) = \\mathbb{E}_{p(X)}[P(Y|X, a)]$ via backdoor adjustment. I would like to hear the authors’ opinions about this.\n3) The authors should compare their method with existing ITE estimation methods to support their argument in the paper saying that “estimating the potential outcome density is better than just estimating the average potential outcome“.\n4) In the TeacherFlow, I would like to understand why did the authors use $X, A$ to compute the weight $\\theta$ of the conditional normalizing flow? Why don’t just model $P(Y|X, A)$ directly via a neural network with $X, A$ are inputs to this network? Since the distribution of outcomes $Y$ is usually simple. Do we really need a normalizing flow with hyper-network for its parameters?\n5) The authors should provide more insights and explanations about why each component is useful for the method to work. Currently, they only use one metric, which is the log-likelihood, for comparison, which I think is not enough.\n\n[1] Learning Representations for Counterfactual Inference, Johansson et al., ICML-2016\n\n[2] Estimating individual treatment effect: generalization bounds and algorithms, Shalit et al., ICML-2017\n\n[3] Adapting Neural Networks for the Estimation of Treatment Effects, Shi et al., NIPS-2019\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, has good quality and is novel. I am not sure about the reproducibility", "summary_of_the_review": "Overall, this is a good paper which proposes a novel method to address an interesting problem. The only problem is that the model is more complex than it should be. This can limit the practical applications of the work. I give this a paper a weak accept.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666675262270}], "openreview_url": "https://openreview.net/forum?id=bTy4D3KHwWU", "arxiv_id": "2209.06203", "paper_pdf": "papers/bTy4D3KHwWU.pdf", "paper_pdf_sha256": "e521635fe088dd215380b017d5a13ddfb5ab42b53ff106cfdaec150506ef58ed", "paper_pdf_bytes": 1068755, "paper_pdf_source": "openreview", "code_url": "https://github.com/Valentyn1997/INFs", "code_repository": "Valentyn1997/INFs", "code_commit": "b75d2e9f52a627145b5bd4a5b104337d04937e56", "code_archive": "repos/bTy4D3KHwWU.zip", "code_archive_sha256": "61410ad0fafe67d1454aa6670a71e162dc51a8d52ef116f45636b4ab7f1f46e2", "code_archive_bytes": 176658, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 149, "github_languages": {"Python": 125182}, "github_archived": false, "github_pushed_at": "2024-10-29T23:02:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/normalizing-flows-for-interventional-density"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ezbMFmQY7L", "year": 2022, "status": "rejected", "title": "C5T5: Controllable Generation of Organic Molecules with Transformers", "authors": ["Daniel Rothchild", "Alex Tamkin", "Julie Yu", "Ujval Misra", "Joseph E. Gonzalez"], "authorids": ["~Daniel_Rothchild1", "~Alex_Tamkin1", "~Julie_Yu2", "~Ujval_Misra1", "~Joseph_E._Gonzalez1"], "authors_source": "OpenReview API", "abstract": "Methods for designing organic materials with desired properties have high potential impact across fields such as medicine, renewable energy, petrochemical engineering, and agriculture. However, using generative models for this task is difficult because candidate compounds must satisfy many constraints, including synthetic accessibility, intellectual property attributes, ``chemical beauty'' (Bickerton et al., 2020), and other considerations that are intuitive to domain experts but can be challenging to quantify. We propose C5T5, a novel self-supervised pretraining method that works in tandem with domain experts by making zero-shot select-and-replace edits, altering organic substances towards desired property values. C5T5 operates on IUPAC names---a standardized molecular representation that intuitively encodes rich structural information for organic chemists but that has been largely ignored by the ML community. Our technique requires no edited molecule pairs to train and only a rough estimate of molecular properties, and it has the potential to model long-range dependencies and symmetric molecular structures more easily than graph-based methods. We demonstrate C5T5's effectiveness on four physical properties relevant for drug discovery, showing that it learns successful and chemically intuitive strategies for altering molecules towards desired property values.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "upUbumEZD4E", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3978/Reviewer_pKJo"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "**Summary**\n\nThis paper proposed a way to modify the molecule based on language pretraining techniques. The sequence representation of molecules is based on IUPAC names, which can be more semantically meaningful and much easier to model than the SMILES or graph based molecule representation. The pretraining is done via a conditional text generation model where the model predicts the fragment names based on the remainder of the molecule and corresponding property values. The application on downstream molecule property optimization tasks show that the proposed approach is effective at obtaining high quality molecules.\n", "review_text": "**Comments**\n\nOverall the paper presents an interesting treatment for the molecule property optimization (e.g., lead optimization, etc). The approach is simple but seems to be effective. The IUPAC naming alternative to SMILES can be of its own contribution to a broader context of chemical engineering. \n\nHowever I have several concerns regarding the current draft.\n\n1. The technique contribution is not very significant. The IUPAC naming is interesting but itself is not solid enough to be considered as technical innovation. The masked pretraining has been seen in other context like NLP, program language, etc. Nevertheless, I think it is ok for an application paper if the quality of experiments is high and the results are solid. So I would focus more on the experimental part. \n\n2. The only baseline studied is Hierarchical G2G (Jin et.al) in Figure 6. Given that there has been many works in the space of molecule optimization, it would be necessary to include more baseline results, especially those with quantitative comparisons. Although some of the baseline methods may not be able to complete all the tasks, it would still be necessary to compare in the situations where baseline methods apply. In this way we can appreciate how significant the results are, especially given that the method proposed in this paper is self-supervised. \n\n3. I would also like to learn more about the limitations of this method. For example, the IUPAC limits the possibility of generating new functional groups. There must be a trade-off between model capacity and the variance of results. \n\n4. The extrapolation is not for free. I’d like to learn more about the failure cases where the model is not able to generate the molecules with desired properties. \n\n5. (Optional) while the model presents an unsupervised way of learning, it would be interesting to see if it can be further fine-tuned with the paired molecule data for lead optimization. In this way, one can compare the results directly with G2G in their benchmarks.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "**Summary**\n\nThis paper proposed a way to modify the molecule based on language pretraining techniques. The sequence representation of molecules is based on IUPAC names, which can be more semantically meaningful and much easier to model than the SMILES or graph based molecule representation. The pretraining is done via a conditional text generation model where the model predicts the fragment names based on the remainder of the molecule and corresponding property values. The application on downstream molecule property optimization tasks show that the proposed approach is effective at obtaining high quality molecules.\n", "main_review": "**Comments**\n\nOverall the paper presents an interesting treatment for the molecule property optimization (e.g., lead optimization, etc). The approach is simple but seems to be effective. The IUPAC naming alternative to SMILES can be of its own contribution to a broader context of chemical engineering. \n\nHowever I have several concerns regarding the current draft.\n\n1. The technique contribution is not very significant. The IUPAC naming is interesting but itself is not solid enough to be considered as technical innovation. The masked pretraining has been seen in other context like NLP, program language, etc. Nevertheless, I think it is ok for an application paper if the quality of experiments is high and the results are solid. So I would focus more on the experimental part. \n\n2. The only baseline studied is Hierarchical G2G (Jin et.al) in Figure 6. Given that there has been many works in the space of molecule optimization, it would be necessary to include more baseline results, especially those with quantitative comparisons. Although some of the baseline methods may not be able to complete all the tasks, it would still be necessary to compare in the situations where baseline methods apply. In this way we can appreciate how significant the results are, especially given that the method proposed in this paper is self-supervised. \n\n3. I would also like to learn more about the limitations of this method. For example, the IUPAC limits the possibility of generating new functional groups. There must be a trade-off between model capacity and the variance of results. \n\n4. The extrapolation is not for free. I’d like to learn more about the failure cases where the model is not able to generate the molecules with desired properties. \n\n5. (Optional) while the model presents an unsupervised way of learning, it would be interesting to see if it can be further fine-tuned with the paired molecule data for lead optimization. In this way, one can compare the results directly with G2G in their benchmarks.\n", "summary_of_the_review": "**Review summary**\n\nThis is an interesting application paper in the domain of molecule optimization with some minor technical contributions. The preliminary results are interesting, but it would be more solid with more quantitative comparison with existing methods on existing benchmarks. \n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635888591262}, {"id": "V9hRg3ZS_Pu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3978/Reviewer_gkwv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper utilizes a transformer architecture based on T5 to generate tokenized IUPAC names of molecules. This allows sampling molecules that are local modifications of starting ones, or are local in IUPAC name but might be chemically large while respecting an initial scaffold.\n", "review_text": "\nStrenghts\nThe idea of using the IUPAC nomenclature is indeed innovative and has not received attention in the area. The use of local modification of tokenized IUPAC names is exciting and seems to work for scaffold derivatization better than other methods\n\nThe T5 model is a powerful addition to language-based addition to molecule generation\n\nWeaknesses\nAt a higher level, this paper proposes a new generative approach but does not evaluate it upon the mutually agreed benchmarks in the field. Guacamol or Moses have lots of meaningful tasks and splits, including some scaffold-based ones. It does not seem appropriate to switch architectures and tasks and then just compare with the one baseline. If the authors believe those benchmarks are relevant, then it is more appropriate to make a case why explicitly (and probably use them anyway)\n\nFurthermore, making local modifications to starting scaffolds has a long history in cheminformatics. Approaches like genetic algorithms, including the graph-based genetic algorithm, can do a great job at local optimization challenges like this https://pubs.rsc.org/en/content/articlelanding/2019/sc/c8sc05372c And since all the properties come from a surrogate model oracle, it does not matter how many train pairs or in-time calls need to be made. LogP (or logD, i think for all the molecules shown they are __extremely__ correlated) is an additive property that can be improved by just appending more carbons to an alkyl chain, so it is a poor performance metric. One could very easily make a genetic-algorithm version of the token substitution approach as a baseline.\n\nI have a number of smaller comments below\n- The advantages of using IUPAC names are not clear at all. \n\n\"a small change in a molecule can lead to a large change in the SMILES string\" This is the case too for IUPAC names - taking one carbon away and closing a ring in the suggested structure in Figure 2 completely changes the IUPAC name\n\n\"flattening the graph into a list of atoms artificially creates variable- and long-range dependencies between bonded atoms\" It does in the IUPAC naming too, since there's a canonical atom ordering that is needed to refer to for writing the molecular name\n\n\"It is difficult to reason about common substructures, because the same structure can be\nrepresented in many different ways depending on how the graph was flattened\"\n\nSame in IUPAC nomenclature - the ester functionality is described through a completely different naming in the example below\n\n\nOC(C1=C(OC(C)=O)C=CC=C1)=O\n2-acetoxybenzoic acid\n\nO=C(O1)C2=C(OC1=O)C=CC=C2\n4H-benzo[d][1,3]dioxine-2,4-dione\n\n\n\"graphs do a poor job encoding symmetry, long-range interactions between atoms that are many bonds apart but nearby in 3D space, and long-range interactions that arise from conjugated system\" Again, the IUPAC naming convention has many of these flaws too. There's no 3D whatsoever in IUPAC either.\n\n\"locants (e.g. “1,” “2,” “N”), which indicate connectivity,\" These locants are also based on arbitrary ordering, just like SMILES\n\nAgain just like SMILES, while there's only one canonical choice of IUPAC nomenclature based on prioritization rules, it is entirely possible to write the IUPAC name for a molecule correctly but based on another prioritization rule. As a matter of fact this paper's approach to tokenization likely does that, depending on the nature of the transformations it perform on the starting name.\n\nSection 3.1 addresses few of the issues raised about SMILES / graphs. (for instance, it does address the multiplier, which is indeed very powerful in theory, but does not seem to be used by the generator in practice. \"pentatrytilbenzene\" would break the bank in logP)\n\n\"ignoring the embedding for “nitroso”\" I would make this part of the sentence, rather than a footnote\n\n\"We discard any generations where the sentinel tokens do not line up, and we further discard any molecules that cannot be parsed by ChemAxon’s calculators. We also discard instances where C5T5 regenerates the base molecule\" I think I understand the logic about needing to seed with something, but discarding the same from the statistics seems antithetic to the nature of a novelty metric. What is the fraction of molecules being discarded at every filter?\n\n\"Although this is not a preferred IUPAC name, it is still unambiguous, and therefore valid and parseable\" This is exactly the point i was making earlier. It seems the approach struggles with synonyms just like SMILES.\n\n\"C5T5 generates molecules that are more synthetically accessible and similar to base molecules and that have a wider range of logP increases than HierG2G.\" I am not sure I'm looking at the plots right. How was synthetic accessibility computed? Is 0 high or low accessibility? The scale in RDKIT has 1 for easy molecule and 10 for difficult, So C5T5 is producing more difficult molecules. Same for the logP increase, i see HierG2G producing a mean of 4 or so, and C5T5 is a little over 1? \n\n\n\n\"Methods like HierG2G that train on pairs of similar molecules are fundamentally\nlimited by a paucity of experimental molecu pairs that have high similarity and high property\nvalue improvement\" This seems misleading since all these methods are using surrogate models to make their labels, there's essentially N squared pairs for the training library. \n\n\n\nQuestions\n\n\"The nearest neighbor of “diphosphate” - “disulfate” + “sulfate” is “phosphate.\" I don't follow. Are diphosphate and disulphate tokens? Does the tokenization procedure not split out the di? Or the \"-ite\" termination? Those are literally rule-based tokens in IUPAC nomenclature.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper utilizes a transformer architecture based on T5 to generate tokenized IUPAC names of molecules. This allows sampling molecules that are local modifications of starting ones, or are local in IUPAC name but might be chemically large while respecting an initial scaffold.\n", "main_review": "\nStrenghts\nThe idea of using the IUPAC nomenclature is indeed innovative and has not received attention in the area. The use of local modification of tokenized IUPAC names is exciting and seems to work for scaffold derivatization better than other methods\n\nThe T5 model is a powerful addition to language-based addition to molecule generation\n\nWeaknesses\nAt a higher level, this paper proposes a new generative approach but does not evaluate it upon the mutually agreed benchmarks in the field. Guacamol or Moses have lots of meaningful tasks and splits, including some scaffold-based ones. It does not seem appropriate to switch architectures and tasks and then just compare with the one baseline. If the authors believe those benchmarks are relevant, then it is more appropriate to make a case why explicitly (and probably use them anyway)\n\nFurthermore, making local modifications to starting scaffolds has a long history in cheminformatics. Approaches like genetic algorithms, including the graph-based genetic algorithm, can do a great job at local optimization challenges like this https://pubs.rsc.org/en/content/articlelanding/2019/sc/c8sc05372c And since all the properties come from a surrogate model oracle, it does not matter how many train pairs or in-time calls need to be made. LogP (or logD, i think for all the molecules shown they are __extremely__ correlated) is an additive property that can be improved by just appending more carbons to an alkyl chain, so it is a poor performance metric. One could very easily make a genetic-algorithm version of the token substitution approach as a baseline.\n\nI have a number of smaller comments below\n- The advantages of using IUPAC names are not clear at all. \n\n\"a small change in a molecule can lead to a large change in the SMILES string\" This is the case too for IUPAC names - taking one carbon away and closing a ring in the suggested structure in Figure 2 completely changes the IUPAC name\n\n\"flattening the graph into a list of atoms artificially creates variable- and long-range dependencies between bonded atoms\" It does in the IUPAC naming too, since there's a canonical atom ordering that is needed to refer to for writing the molecular name\n\n\"It is difficult to reason about common substructures, because the same structure can be\nrepresented in many different ways depending on how the graph was flattened\"\n\nSame in IUPAC nomenclature - the ester functionality is described through a completely different naming in the example below\n\n\nOC(C1=C(OC(C)=O)C=CC=C1)=O\n2-acetoxybenzoic acid\n\nO=C(O1)C2=C(OC1=O)C=CC=C2\n4H-benzo[d][1,3]dioxine-2,4-dione\n\n\n\"graphs do a poor job encoding symmetry, long-range interactions between atoms that are many bonds apart but nearby in 3D space, and long-range interactions that arise from conjugated system\" Again, the IUPAC naming convention has many of these flaws too. There's no 3D whatsoever in IUPAC either.\n\n\"locants (e.g. “1,” “2,” “N”), which indicate connectivity,\" These locants are also based on arbitrary ordering, just like SMILES\n\nAgain just like SMILES, while there's only one canonical choice of IUPAC nomenclature based on prioritization rules, it is entirely possible to write the IUPAC name for a molecule correctly but based on another prioritization rule. As a matter of fact this paper's approach to tokenization likely does that, depending on the nature of the transformations it perform on the starting name.\n\nSection 3.1 addresses few of the issues raised about SMILES / graphs. (for instance, it does address the multiplier, which is indeed very powerful in theory, but does not seem to be used by the generator in practice. \"pentatrytilbenzene\" would break the bank in logP)\n\n\"ignoring the embedding for “nitroso”\" I would make this part of the sentence, rather than a footnote\n\n\"We discard any generations where the sentinel tokens do not line up, and we further discard any molecules that cannot be parsed by ChemAxon’s calculators. We also discard instances where C5T5 regenerates the base molecule\" I think I understand the logic about needing to seed with something, but discarding the same from the statistics seems antithetic to the nature of a novelty metric. What is the fraction of molecules being discarded at every filter?\n\n\"Although this is not a preferred IUPAC name, it is still unambiguous, and therefore valid and parseable\" This is exactly the point i was making earlier. It seems the approach struggles with synonyms just like SMILES.\n\n\"C5T5 generates molecules that are more synthetically accessible and similar to base molecules and that have a wider range of logP increases than HierG2G.\" I am not sure I'm looking at the plots right. How was synthetic accessibility computed? Is 0 high or low accessibility? The scale in RDKIT has 1 for easy molecule and 10 for difficult, So C5T5 is producing more difficult molecules. Same for the logP increase, i see HierG2G producing a mean of 4 or so, and C5T5 is a little over 1? \n\n\n\n\"Methods like HierG2G that train on pairs of similar molecules are fundamentally\nlimited by a paucity of experimental molecu pairs that have high similarity and high property\nvalue improvement\" This seems misleading since all these methods are using surrogate models to make their labels, there's essentially N squared pairs for the training library. \n\n\n\nQuestions\n\n\"The nearest neighbor of “diphosphate” - “disulfate” + “sulfate” is “phosphate.\" I don't follow. Are diphosphate and disulphate tokens? Does the tokenization procedure not split out the di? Or the \"-ite\" termination? Those are literally rule-based tokens in IUPAC nomenclature.", "summary_of_the_review": "I think the idea of using IUPAC names is intriguing, but neither the theoretical arguments not the empirical results (due to lack of benchmarks) are convincing. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "The authors forgot to include a statement. ", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635873026788}, {"id": "DE3sqv1oTJJ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3978/Reviewer_MJUv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method C5T5, a self-supervised pre-training method based on the T5 pre-trained model, which is able to make zero-shot select-and-replace edits to satisfy specific property values. The specific difference of this paper is the IUPAC names (a standardized molecular representation), and the method is totally self-supervised. The experiments are evaluated on octanol-water partition, distribution coefficients, polar surface area, and refractivity, above four properties. Experiments show that the designed methods are able to achieve the optimization objective. ", "review_text": "The summary of the paper is concluded in the above contents. Here are the strengths and weaknesses of the paper from my view. \n\nStrengths:\n1. The first attractive point is the IUPAC names representation method. As mentioned in the paper, this method is a standardized molecular representation that encodes the rich structural information but that has been largely ignored. The common approach of the representation for the molecule is the SMILES or graph. As introduced, this representation can handle more semantic information, and also the interpretability is better. \n2. The method is straightforward and simple, which uses a pre-training method T5 to model a conditional language model, where the condition is the property value. As a result, C5T5 does not require paired dataset for the molecule property improvement. \n3. The results on properties show the desired optimization is satisfied, and the different studies give a more clear understanding.\n\nweaknesses:\nThough I personally like the simple idea, there are still concerns and questions from my point,\n1. The first concern is about the specific property value distribution. As shown in Table 5, each property only splits into three buckets. Then each model can only have three different types of property, which makes the model to be limited. For example, it can not clearly generate a better molecule if the initial molecule with logP is larger than 5.6. \n2. The second concern is about the IUPAC names. Though I feel happy to see a new representation method that seems to be better than other methods, the comparison is not clear and there seems no strong evidence or numbers can be shown in the paper. In this way, this can not clearly convince the superiority of IUPAC. However, I am pretty interested to see the goodness of IUPAC.\n3. One specific question is about the pre-trained T5 model. The authors mentioned that they do not use sentencepiece, and only keep the IUPAC names as the vocabulary, which is different from the pre-trained embedding of T5 model. It is said that the first 1274 embeddings of English pre-trained embedding are kept. I am not sure about this process, what is the difference between this method and directly replacing the vocabulary with the 1274 new randomly initialized embeddings? Since the vocabulary is different at all. \n4. It is required to make a result comparison to see the advantage of this method. Otherwise, it is hard to evaluate. \n5. Other concerns have been discussed by the authors, for example, the training cost is high; it is not always possible to generate valid and satisfying molecules since there are no constraints in the modeling. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a method C5T5, a self-supervised pre-training method based on the T5 pre-trained model, which is able to make zero-shot select-and-replace edits to satisfy specific property values. The specific difference of this paper is the IUPAC names (a standardized molecular representation), and the method is totally self-supervised. The experiments are evaluated on octanol-water partition, distribution coefficients, polar surface area, and refractivity, above four properties. Experiments show that the designed methods are able to achieve the optimization objective. ", "main_review": "The summary of the paper is concluded in the above contents. Here are the strengths and weaknesses of the paper from my view. \n\nStrengths:\n1. The first attractive point is the IUPAC names representation method. As mentioned in the paper, this method is a standardized molecular representation that encodes the rich structural information but that has been largely ignored. The common approach of the representation for the molecule is the SMILES or graph. As introduced, this representation can handle more semantic information, and also the interpretability is better. \n2. The method is straightforward and simple, which uses a pre-training method T5 to model a conditional language model, where the condition is the property value. As a result, C5T5 does not require paired dataset for the molecule property improvement. \n3. The results on properties show the desired optimization is satisfied, and the different studies give a more clear understanding.\n\nweaknesses:\nThough I personally like the simple idea, there are still concerns and questions from my point,\n1. The first concern is about the specific property value distribution. As shown in Table 5, each property only splits into three buckets. Then each model can only have three different types of property, which makes the model to be limited. For example, it can not clearly generate a better molecule if the initial molecule with logP is larger than 5.6. \n2. The second concern is about the IUPAC names. Though I feel happy to see a new representation method that seems to be better than other methods, the comparison is not clear and there seems no strong evidence or numbers can be shown in the paper. In this way, this can not clearly convince the superiority of IUPAC. However, I am pretty interested to see the goodness of IUPAC.\n3. One specific question is about the pre-trained T5 model. The authors mentioned that they do not use sentencepiece, and only keep the IUPAC names as the vocabulary, which is different from the pre-trained embedding of T5 model. It is said that the first 1274 embeddings of English pre-trained embedding are kept. I am not sure about this process, what is the difference between this method and directly replacing the vocabulary with the 1274 new randomly initialized embeddings? Since the vocabulary is different at all. \n4. It is required to make a result comparison to see the advantage of this method. Otherwise, it is hard to evaluate. \n5. Other concerns have been discussed by the authors, for example, the training cost is high; it is not always possible to generate valid and satisfying molecules since there are no constraints in the modeling. \n\n", "summary_of_the_review": "I personally feel good about this work, but unclear parts are main concerns and there are no comparison with other works. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635839917278}], "openreview_url": "https://openreview.net/forum?id=ezbMFmQY7L", "arxiv_id": "2108.10307", "paper_pdf": "papers/ezbMFmQY7L.pdf", "paper_pdf_sha256": "8237b4ff69e1909247a9c368f6fc312120066812c7356f0928ab41331075ae9f", "paper_pdf_bytes": 1083477, "paper_pdf_source": "openreview", "code_url": "https://github.com/dhroth/c5t5", "code_repository": "dhroth/c5t5", "code_commit": "cba53cc881a87eb7d3451fd73e10df479e2ca057", "code_archive": "repos/ezbMFmQY7L.zip", "code_archive_sha256": "01d3595b9a56ec820573ac4a552f03e98a5653dc919fc2e39c3ac40a6fa979d1", "code_archive_bytes": 220637, "code_file_count": 16, "code_extensions": {".py": 13, ".sh": 3}, "github_disk_usage_kb": 223, "github_languages": {"Python": 60360, "Shell": 17555}, "github_archived": false, "github_pushed_at": "2021-12-17T17:39:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/c5t5-controllable-generation-of-organic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1s1T7xHc5l6", "year": 2021, "status": "rejected", "title": "FILTRA: Rethinking Steerable CNN by Filter Transform", "authors": ["Bo Li", "Qili Wang", "Gim Hee Lee"], "authorids": ["~Bo_Li16", "~Qili_Wang1", "~Gim_Hee_Lee1"], "authors_source": "OpenReview API", "abstract": "Steerable CNN imposes the prior knowledge of transformation invariance or equivariance in the network architecture to enhance the the network robustness on geometry transformation of data and reduce overfitting. Filter transform has been an intuitive and widely used technique to construct steerable CNN in the past decades. Recently, group representation theory is used to analyze steerable CNN and reveals the function space structure of a steerable kernel function. However, it is not yet clear on how this theory is related to the filter transform technique. In this paper, we show that kernel constructed by filter transform can also be interpreted in the group representation theory. Meanwhile, we show that filter transformed kernels can be used to convolve input/output features in different group representation. This interpretation help complete the puzzle of steerable CNN theory and provides a novel and simple approach to implement steerable convolution operators. Experiments are executed on multiple datasets to verify the feasibilty of the proposed approach.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "THR6tif-el7", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1573/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper builds the connection between various steerable CNN structures based on group representation theory and filter transformations. Using the discrete rotation and reflection group as an example, the paper establishes ways to construct steerable convolutional filters that transform features in trivial, regular, and irreducible representations. In particular, recent works on steerable CNNs such as ORN, RotDCF, TI-pooling, and RotEqNet can all be explained under such framework.\n\nThe paper is generally well written and well organized. However, I reckon the material would be quite dense for readers interested in equivariant CNNs but not well-versed in group representation theory, so a bit more explanation either in text or in the appendix would be helpful. The main theoretical results in the paper on the considered discrete group transformation are interesting, and can potentially lead to other development in the community. The numerical experiments seem to be limited (though this phenomenon seems to be the issue for most of the papers in this area...)\n\n** Pros\n1. Well written and technically correct.\n2. Interesting theoretical results that might inspire other works in the field.\n\n** Cons (to be explained in more details in questions)\n1. Some notation and abbreviation are used without proper explanation, which might be confusing.\n2. Limited experimental results\n\n** Questions and comment\n1. Some of the notations and abbreviations are used without proper explanation. For example, \"K\" in equation (12), OCR on page 1, and missing reference on R2Conv. I would encourage the authors to make a greater effort to clarify their ideas and results.\n\n2. The theory seems to be built only on discrete group. Does it generalize to continuous group transformation (such as SO(2) before discretization)?\n\n3. Also, does the theory generalize to non-compact groups such as  scaling and shearing?\n\n4. The experiments seem to be limited, even though this seems to be a common issue in papers in this field. However, I do recommend the authors to present the \"equivariant loss\" when using their proposed FILTRA, considering that discretization and interpolation might cause a problem in their setting unlike other means of steerable CNNs such as RotDCF and Harmonic Net", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting paper that builds the connection between steerable CNN based on group representation theory and filter transformation", "review": "This paper builds the connection between various steerable CNN structures based on group representation theory and filter transformations. Using the discrete rotation and reflection group as an example, the paper establishes ways to construct steerable convolutional filters that transform features in trivial, regular, and irreducible representations. In particular, recent works on steerable CNNs such as ORN, RotDCF, TI-pooling, and RotEqNet can all be explained under such framework.\n\nThe paper is generally well written and well organized. However, I reckon the material would be quite dense for readers interested in equivariant CNNs but not well-versed in group representation theory, so a bit more explanation either in text or in the appendix would be helpful. The main theoretical results in the paper on the considered discrete group transformation are interesting, and can potentially lead to other development in the community. The numerical experiments seem to be limited (though this phenomenon seems to be the issue for most of the papers in this area...)\n\n** Pros\n1. Well written and technically correct.\n2. Interesting theoretical results that might inspire other works in the field.\n\n** Cons (to be explained in more details in questions)\n1. Some notation and abbreviation are used without proper explanation, which might be confusing.\n2. Limited experimental results\n\n** Questions and comment\n1. Some of the notations and abbreviations are used without proper explanation. For example, \"K\" in equation (12), OCR on page 1, and missing reference on R2Conv. I would encourage the authors to make a greater effort to clarify their ideas and results.\n\n2. The theory seems to be built only on discrete group. Does it generalize to continuous group transformation (such as SO(2) before discretization)?\n\n3. Also, does the theory generalize to non-compact groups such as  scaling and shearing?\n\n4. The experiments seem to be limited, even though this seems to be a common issue in papers in this field. However, I do recommend the authors to present the \"equivariant loss\" when using their proposed FILTRA, considering that discretization and interpolation might cause a problem in their setting unlike other means of steerable CNNs such as RotDCF and Harmonic Net", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603978928860}, {"id": "Y9XsddlTB2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1573/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the connection between steerable CNN and filter transformation. The authors show theoretically that filter transformation can be used to realize steerable filters over different group representations. The authors also empirically show that the filter transformation based steerable CNN performs on par with the implementation based on harmonic bases.\n\nOverall, I think this paper is hard to follow. The presentation and notation depends on that of prior works and is hard to understand without being familiar with Steerable CNN. Compared with prior works, this paper focuses more on the mathematical derivation but falls short of explaining their implication. The result is therefore hard to understand for those that are not familiar with the theoretical basis. Also, the authors do not clearly show why the contribution of this work is significant. The theoretical connection doesn’t seem to provide a significant advantage nor solve the problems in existing methods. Similarly, the experiment doesn’t provide much information and only shows that the proposed implementation works on simple datasets. The authors should try to highlight how this work may benefit other research or applications and verify that the proposed method works on more complicated and realistic data.\n\nAnother aspect that can be improved is to discuss how the theory generalizes when multiple convolution layers are applied. In general, the theoretical properties of steerable convolution does not automatically generalize to multiple convolution layers. This is particularly important in real world data, because the transformation may happen at different scales and may need to be accounted for at different layers in the network. Therefore, the results derived from a single convolution operation may not be sufficient.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper provide the theoretical connection between steerable convolution and filter transform method.", "review": "This paper studies the connection between steerable CNN and filter transformation. The authors show theoretically that filter transformation can be used to realize steerable filters over different group representations. The authors also empirically show that the filter transformation based steerable CNN performs on par with the implementation based on harmonic bases.\n\nOverall, I think this paper is hard to follow. The presentation and notation depends on that of prior works and is hard to understand without being familiar with Steerable CNN. Compared with prior works, this paper focuses more on the mathematical derivation but falls short of explaining their implication. The result is therefore hard to understand for those that are not familiar with the theoretical basis. Also, the authors do not clearly show why the contribution of this work is significant. The theoretical connection doesn’t seem to provide a significant advantage nor solve the problems in existing methods. Similarly, the experiment doesn’t provide much information and only shows that the proposed implementation works on simple datasets. The authors should try to highlight how this work may benefit other research or applications and verify that the proposed method works on more complicated and realistic data.\n\nAnother aspect that can be improved is to discuss how the theory generalizes when multiple convolution layers are applied. In general, the theoretical properties of steerable convolution does not automatically generalize to multiple convolution layers. This is particularly important in real world data, because the transformation may happen at different scales and may need to be accounted for at different layers in the network. Therefore, the results derived from a single convolution operation may not be sufficient.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603901858735}, {"id": "fjCq8AI6XAb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1573/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "***Summary***\n\nThe paper outlines the construction of filters in a group equivariant convolutional network equivariant to the groups CN (cyclic group of order N (N rotations)) and DN (dihedral group of order N (N rotations and flipping)). The authors achieve this by taking linear combinations of (harmonic) basis filters, as was shown in Weiler et al. (2018). They then proceed to show explicitly how one can build filters between particular representations of each group, namely, the trivial, irreducible, and regular representations. For experiments they demonstrate that the filters perform en par with those of Weiler and Cesa (2019), who had performed a near-exhaustive comparison among representations of SE(2, R).\n\n\n\n***Pros***\n\nTechnically I believe the work to be sound. I did not see any mistakes jump out at me. I also think that it is laudable that the authors included a high amount of detail in their exposition, which lays bare the exact mechanisms by which one would effectively go about building an equivariant layer.\n\nExperimentally it appears that the authors chose a sensible baseline and that they compared on dataset, which are relevant to the topic and commonly used in the equivariance literature. The authors also provide timing information on how fast each filter can be generated, which is something I have not seen before in the equivariance literature and which I appreciate seeing.\n\n\n\n\n***Cons and constructive feedback***\n\nPerhaps my largest criticism is that it is not obvious exactly what the contribution of the paper is meant to be. Perhaps this originates from my not knowing what the authors mean by “filter transform” and “steerable CNN”. My understanding is that the authors believe it is currently unknown how to construct a steerable CNN for the cyclic and dihedral groups. They provide explicit constructions in section 3, but to my knowledge these are already provided in the long appendix of Weiler and Cesa (2019), furthermore they can be found in Cohen and Welling (2016), Cohen and Welling (2017) and Bekkers et al. (2018) and most notably Weiler et al. (2018), who were to first to show how to construct equivariant filter layers some linear combinations of harmonic basis filters. As a result I am unsure of how to gauge the novelty of this contribution.\n\nIn the experiments of Table 2 (classification and regression) the authors compare their framework with an R2conv (representative harmonic based convolution) and a regular translationally equivariant network. I would like to know what exactly is a representative harmonic based convolution? Is that meant to refer to the work of Weiler and Cesa (2019)? \n\nIn the results tables there are no error bars. If possible I would have liked to have seen them. Since they are not there is it very hard to judge the efficacy of the results, which differ from the baselines by very small amounts.\n\nGiven the proximity of the work to Weiler and Cesa (2019), I would like to know exactly what differentiates the two works.\n\nPlease include exactly what the functions roll, flipped, and circulant actually do mathematically. I found these difficult to parse.\n\nThe mathematical level of the paper is pretty heavy for the uninitiated. It may be advisable for the authors to include a glossary of terms, if not short descriptions, in an appendix. If this is too much, at least please point to other papers, which have easily readable background sections for those not already well-read in group theory.\n\nThe authors may wish to have the submission proofread for spelling and grammar.\n\n***Post rebuttal review***\n\nHaving read through the rebuttal, the updated submission, and the reviews of the other reviewers I have upgraded my review from a reject to marginally below acceptance. This is for two main reasons. 1) the authors have vastly improved the presentation of the submission, which now looks a lot easier to read, 2) the authors have clarified for me, at least, what the main contribution of the work is. \n\nThat said I am not entirely sure what this contribution adds to the equivariance community, hence why my recommendation still leans towards reject. As far as I am aware, solving the equivariance equations is not the large bottleneck to progress in our community. They are linear equations, and there is work back into the 80's solving them (check out people like Pietro Perona, Patrick Teo). I think more importantly we need to focus on pushing the boundaries in areas such as extension to non-Euclidean manifolds, convolution over non-compact groups, learning symmetries, etc. While this work is clearly mathematically sound and the authors have demonstrated deep knowledge of the area, it feels a little like retracing prior works. That said, if the other reviewers disagree then I don't mind this paper being accepted. Perhaps since I have worked in this area, what appears as obvious to me is not generally acknowledged and this paper may serve as a useful clarification for those wishing to dive into the literature.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Initial review for FILTRA", "review": "***Summary***\n\nThe paper outlines the construction of filters in a group equivariant convolutional network equivariant to the groups CN (cyclic group of order N (N rotations)) and DN (dihedral group of order N (N rotations and flipping)). The authors achieve this by taking linear combinations of (harmonic) basis filters, as was shown in Weiler et al. (2018). They then proceed to show explicitly how one can build filters between particular representations of each group, namely, the trivial, irreducible, and regular representations. For experiments they demonstrate that the filters perform en par with those of Weiler and Cesa (2019), who had performed a near-exhaustive comparison among representations of SE(2, R).\n\n\n\n***Pros***\n\nTechnically I believe the work to be sound. I did not see any mistakes jump out at me. I also think that it is laudable that the authors included a high amount of detail in their exposition, which lays bare the exact mechanisms by which one would effectively go about building an equivariant layer.\n\nExperimentally it appears that the authors chose a sensible baseline and that they compared on dataset, which are relevant to the topic and commonly used in the equivariance literature. The authors also provide timing information on how fast each filter can be generated, which is something I have not seen before in the equivariance literature and which I appreciate seeing.\n\n\n\n\n***Cons and constructive feedback***\n\nPerhaps my largest criticism is that it is not obvious exactly what the contribution of the paper is meant to be. Perhaps this originates from my not knowing what the authors mean by “filter transform” and “steerable CNN”. My understanding is that the authors believe it is currently unknown how to construct a steerable CNN for the cyclic and dihedral groups. They provide explicit constructions in section 3, but to my knowledge these are already provided in the long appendix of Weiler and Cesa (2019), furthermore they can be found in Cohen and Welling (2016), Cohen and Welling (2017) and Bekkers et al. (2018) and most notably Weiler et al. (2018), who were to first to show how to construct equivariant filter layers some linear combinations of harmonic basis filters. As a result I am unsure of how to gauge the novelty of this contribution.\n\nIn the experiments of Table 2 (classification and regression) the authors compare their framework with an R2conv (representative harmonic based convolution) and a regular translationally equivariant network. I would like to know what exactly is a representative harmonic based convolution? Is that meant to refer to the work of Weiler and Cesa (2019)? \n\nIn the results tables there are no error bars. If possible I would have liked to have seen them. Since they are not there is it very hard to judge the efficacy of the results, which differ from the baselines by very small amounts.\n\nGiven the proximity of the work to Weiler and Cesa (2019), I would like to know exactly what differentiates the two works.\n\nPlease include exactly what the functions roll, flipped, and circulant actually do mathematically. I found these difficult to parse.\n\nThe mathematical level of the paper is pretty heavy for the uninitiated. It may be advisable for the authors to include a glossary of terms, if not short descriptions, in an appendix. If this is too much, at least please point to other papers, which have easily readable background sections for those not already well-read in group theory.\n\nThe authors may wish to have the submission proofread for spelling and grammar.\n\n***Post rebuttal review***\n\nHaving read through the rebuttal, the updated submission, and the reviews of the other reviewers I have upgraded my review from a reject to marginally below acceptance. This is for two main reasons. 1) the authors have vastly improved the presentation of the submission, which now looks a lot easier to read, 2) the authors have clarified for me, at least, what the main contribution of the work is. \n\nThat said I am not entirely sure what this contribution adds to the equivariance community, hence why my recommendation still leans towards reject. As far as I am aware, solving the equivariance equations is not the large bottleneck to progress in our community. They are linear equations, and there is work back into the 80's solving them (check out people like Pietro Perona, Patrick Teo). I think more importantly we need to focus on pushing the boundaries in areas such as extension to non-Euclidean manifolds, convolution over non-compact groups, learning symmetries, etc. While this work is clearly mathematically sound and the authors have demonstrated deep knowledge of the area, it feels a little like retracing prior works. That said, if the other reviewers disagree then I don't mind this paper being accepted. Perhaps since I have worked in this area, what appears as obvious to me is not generally acknowledged and this paper may serve as a useful clarification for those wishing to dive into the literature.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603731447648}, {"id": "vSikmE43uxO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1573/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Equivariant Steerable CNNs for 2D/3D rotation+reflection+translation groups have generally been implemented as a filter transform/expansion step followed by a standard convolution. The filter expansion step involves taking a linear combination of steerable basis filters. These basis filters are pre-computed before network training by solving a linear system or by sampling the continuous analytical solution (this can take a few minutes). Depending on the chosen group representation wrt which the network layer is equivariant, a different filter basis will emerge, but in general one can see that the basis filters come out as rotated and flipped copies of some basis filters, with the occasional sign flip (this has been stated in some earlier works). The precise way in which a basis filter is to be rotated and flipped to obtain the steerable filter basis for the 2D case had not been worked out before, to my knowledge, and this is one of the contributions of this paper. The analysis is done for each (input, output) representation type chosen from {trivial, irreducible, regular} representations. Having worked this out, the paper proposes to use this as a way of implementing the filter expansion step, starting from a basis filter and rotating/flipping it to obtain an expanded filter bank. \n\nThe proposed method (FILTRA) does not require a precomputation step if I understand correctly, which is a significant practical advantage. Experiments further show that the method is similar or faster at filter expansion. Finally, the method is validated by training networks on benchmark tasks and shown to perform similarly to or better than the steerable CNN implementation of Weiler & Cesa, which is the best existing implementation.\n\nThe paper briefly mentions that filters cannot be rotated exactly on a discrete grid, but I didn't figure out how the authors propose to deal with this issue. How exactly are the filters rotated?\n\nI think the method proposed in the paper is useful, as it is both faster and better than existing steerable CNN implementations. The paper itself is fairly well written and technically correct as far as I can tell, but may be challenging to read for those who are not yet knowledgeable about steerable CNNs. Those readers however are unlikely to be interested in learning about the implementation details of steerable CNNs anyway, so perhaps this is fine. The reason I am not giving a higher rating is that I think that although this is a useful contribution to the literature on steerable CNNs, which are being used in an increasing number of applications, the paper does not represent a major breakthrough and although the calculations are non-trivial, does not contain highly unexpected or deep theoretical results.\n\nTypos:\nCadestrian -> Cartesian\nirreduciable -> irreducible\nequity -> equality\n\nEdit:\nHaving read the reviews, rebuttal and updated paper, I have decided to maintain my score of 6.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "Equivariant Steerable CNNs for 2D/3D rotation+reflection+translation groups have generally been implemented as a filter transform/expansion step followed by a standard convolution. The filter expansion step involves taking a linear combination of steerable basis filters. These basis filters are pre-computed before network training by solving a linear system or by sampling the continuous analytical solution (this can take a few minutes). Depending on the chosen group representation wrt which the network layer is equivariant, a different filter basis will emerge, but in general one can see that the basis filters come out as rotated and flipped copies of some basis filters, with the occasional sign flip (this has been stated in some earlier works). The precise way in which a basis filter is to be rotated and flipped to obtain the steerable filter basis for the 2D case had not been worked out before, to my knowledge, and this is one of the contributions of this paper. The analysis is done for each (input, output) representation type chosen from {trivial, irreducible, regular} representations. Having worked this out, the paper proposes to use this as a way of implementing the filter expansion step, starting from a basis filter and rotating/flipping it to obtain an expanded filter bank. \n\nThe proposed method (FILTRA) does not require a precomputation step if I understand correctly, which is a significant practical advantage. Experiments further show that the method is similar or faster at filter expansion. Finally, the method is validated by training networks on benchmark tasks and shown to perform similarly to or better than the steerable CNN implementation of Weiler & Cesa, which is the best existing implementation.\n\nThe paper briefly mentions that filters cannot be rotated exactly on a discrete grid, but I didn't figure out how the authors propose to deal with this issue. How exactly are the filters rotated?\n\nI think the method proposed in the paper is useful, as it is both faster and better than existing steerable CNN implementations. The paper itself is fairly well written and technically correct as far as I can tell, but may be challenging to read for those who are not yet knowledgeable about steerable CNNs. Those readers however are unlikely to be interested in learning about the implementation details of steerable CNNs anyway, so perhaps this is fine. The reason I am not giving a higher rating is that I think that although this is a useful contribution to the literature on steerable CNNs, which are being used in an increasing number of applications, the paper does not represent a major breakthrough and although the calculations are non-trivial, does not contain highly unexpected or deep theoretical results.\n\nTypos:\nCadestrian -> Cartesian\nirreduciable -> irreducible\nequity -> equality\n\nEdit:\nHaving read the reviews, rebuttal and updated paper, I have decided to maintain my score of 6.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603376430488}], "openreview_url": "https://openreview.net/forum?id=1s1T7xHc5l6", "arxiv_id": "2105.11636", "paper_pdf": "papers/1s1T7xHc5l6.pdf", "paper_pdf_sha256": "b554d92887b7df121850744d01c96673ed4441497580a48c984accf8742363e9", "paper_pdf_bytes": 327026, "paper_pdf_source": "openreview", "code_url": "https://github.com/prclibo/filtra", "code_repository": "prclibo/filtra", "code_commit": "694e6f624f9221e0793b3132374731a7aebe0f18", "code_archive": "repos/1s1T7xHc5l6.zip", "code_archive_sha256": "96104740fe4c35dcff3b01090ba36c20d13a08cafaed87db086a97b62ae03bad", "code_archive_bytes": 87080, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 268, "github_languages": {"Python": 137175}, "github_archived": false, "github_pushed_at": "2022-03-01T14:51:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/filtra-rethinking-steerable-cnn-by-filter-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ByeL1R4FvS", "year": 2020, "status": "rejected", "title": "Unsupervised Data Augmentation for Consistency Training", "authors": ["Qizhe Xie", "Zihang Dai", "Eduard Hovy", "Minh-Thang Luong", "Quoc V. Le"], "authorids": ["qizhex@cs.cmu.edu", "dzihang@cs.cmu.edu", "hovy@cs.cmu.edu", "thangluong@google.com", "qvl@google.com"], "authors_source": "OpenReview API", "abstract": "Semi-supervised learning lately has shown much promise in improving deep learning models when labeled data is scarce. Common among recent approaches is the use of consistency training on a large amount of unlabeled data to constrain model predictions to be invariant to input noise. In this work, we present a new perspective on how to effectively noise unlabeled examples and argue that the quality of noising, specifically those produced by advanced data augmentation methods, plays a crucial role in semi-supervised learning. By substituting simple noising operations with advanced data augmentation methods, our method brings substantial improvements across six language and three vision tasks under the same consistency training framework. On the IMDb text classification dataset, with only 20 labeled examples, our method achieves an error rate of 4.20, outperforming the state-of-the-art model trained on 25,000 labeled examples. On a standard semi-supervised learning benchmark, CIFAR-10, our method outperforms all previous approaches and achieves an error rate of 2.7% with only 4,000 examples, nearly matching the performance of models trained on 50,000 labeled examples. Our method also combines well with transfer learning, e.g., when finetuning from BERT, and yields improvements in high-data regime, such as ImageNet, whether when there is only 10% labeled data or when a full labeled set with 1.3M extra unlabeled examples is used.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Byxu8IcTKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper894/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to substitute simple noising operations with many data augmentation methods in consistency-based semi-supervised learning. The main idea is the same as previous work: constrain the model predictions of unlabeled examples to be invariant to different noise. The proposed UDA is evaluated on a wide range of language and vision tasks.\n\nOverall, the paper is well-written and clear. The most impressive point of this paper is its strong empirical results. However, it looks not surprising to me that more data augmentations found in supervised learning are also effective in semi-supervised learning. The paper fails to provide any theoretical insights but a thorough empirical evaluation.\n\nOne of my concerns is that the hyperparameters on vision tasks follow those of AutoAugment, which is carefully tuned on supervised tasks. Apparently, their hyperparameters are based on the whole labeled training dataset. In this case, the adopted hyperparameters include sort of information of the whole labeled dataset. Is it fair?\n\nAnother concern is how to control the strength of augmentations. For example, for digit images like SVHN, a \"6\" rotates by 180 degree is \"9\", whose prediction should change correspondingly. In this case, the assumption of invariance does not hold when the augmentation is too strong. \n\nI'm willing to increase my score if the authors address my concerns.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The paper proposes to substitute simple noising operations with many data augmentation methods in consistency-based semi-supervised learning. The main idea is the same as previous work: constrain the model predictions of unlabeled examples to be invariant to different noise. The proposed UDA is evaluated on a wide range of language and vision tasks.\n\nOverall, the paper is well-written and clear. The most impressive point of this paper is its strong empirical results. However, it looks not surprising to me that more data augmentations found in supervised learning are also effective in semi-supervised learning. The paper fails to provide any theoretical insights but a thorough empirical evaluation.\n\nOne of my concerns is that the hyperparameters on vision tasks follow those of AutoAugment, which is carefully tuned on supervised tasks. Apparently, their hyperparameters are based on the whole labeled training dataset. In this case, the adopted hyperparameters include sort of information of the whole labeled dataset. Is it fair?\n\nAnother concern is how to control the strength of augmentations. For example, for digit images like SVHN, a \"6\" rotates by 180 degree is \"9\", whose prediction should change correspondingly. In this case, the assumption of invariance does not hold when the augmentation is too strong. \n\nI'm willing to increase my score if the authors address my concerns."}, "tcdate": 1571821135737}, {"id": "rkxWsB9TKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper894/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors present a new perspective on how to effectively noise unlabeled examples and argue that the quality of noising plays a crucial role in semi-supervised learning. By substituting simple noising operations with advanced data augmentation methods, their method brings substantial improvements across six language and three vision tasks under the same consistency training framework. I think the topic itself is interesting and I have the following concerns.\n(1) The first is about the contribution of this paper. In this paper, all the results, including the augmented methods are all well established approaches. The authors have just employed them in solving a new problem, without support about why they work. Thus, the results are only strategies, without theoretical guarantee or insights. It is difficult to convince the reviewers.\n(2) Although the authors have achieved seemingly promising results, I think it can not convince me since the authors have not answered the questions about why and when. I think this paper likes a technical report, not a research paper.\n(3) I have also noticed the discussions among the authors and other readers. It seems that the large improvement depends on the parameters heavily. So, why not to share the parameters directly? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "In this paper, the authors present a new perspective on how to effectively noise unlabeled examples and argue that the quality of noising plays a crucial role in semi-supervised learning. By substituting simple noising operations with advanced data augmentation methods, their method brings substantial improvements across six language and three vision tasks under the same consistency training framework. I think the topic itself is interesting and I have the following concerns.\n(1) The first is about the contribution of this paper. In this paper, all the results, including the augmented methods are all well established approaches. The authors have just employed them in solving a new problem, without support about why they work. Thus, the results are only strategies, without theoretical guarantee or insights. It is difficult to convince the reviewers.\n(2) Although the authors have achieved seemingly promising results, I think it can not convince me since the authors have not answered the questions about why and when. I think this paper likes a technical report, not a research paper.\n(3) I have also noticed the discussions among the authors and other readers. It seems that the large improvement depends on the parameters heavily. So, why not to share the parameters directly? "}, "tcdate": 1571820952528}, {"id": "r1lbeiL5KH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper894/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper \"Unsupervised Data Augmentation for Consistency Training\" marries two recent ideas of \n1. \"Data Augmentation\" (DA) from supervised learning: The authors explore various methods for \"DA\" mostly inspired by much recent work such as Random image transformations, Backtranslation, and TF-IDF based word replacement.\n2. \"Consistency Training\" (CT) from semi-supervised learning: CT tries to minimize the divergence between the output distributions of the classifiers that are produced by adding noise to the input.\n\nThe key insight in this paper is that, data augmentation methods that work well during supervised training should also work equally well as the noise distribution for consistency training on unlabeled data. The authors support this claim empirically through the experiments in table 1 and 2. \n\nThe paper is well written and the authors present extensive comparative and ablation tests to demonstrate that their proposed method works well with both low and high amounts of labeled data.  This paper should be accepted into the conference.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The paper \"Unsupervised Data Augmentation for Consistency Training\" marries two recent ideas of \n1. \"Data Augmentation\" (DA) from supervised learning: The authors explore various methods for \"DA\" mostly inspired by much recent work such as Random image transformations, Backtranslation, and TF-IDF based word replacement.\n2. \"Consistency Training\" (CT) from semi-supervised learning: CT tries to minimize the divergence between the output distributions of the classifiers that are produced by adding noise to the input.\n\nThe key insight in this paper is that, data augmentation methods that work well during supervised training should also work equally well as the noise distribution for consistency training on unlabeled data. The authors support this claim empirically through the experiments in table 1 and 2. \n\nThe paper is well written and the authors present extensive comparative and ablation tests to demonstrate that their proposed method works well with both low and high amounts of labeled data.  This paper should be accepted into the conference."}, "tcdate": 1571609320729}], "openreview_url": "https://openreview.net/forum?id=ByeL1R4FvS", "arxiv_id": "1904.12848", "paper_pdf": "papers/ByeL1R4FvS.pdf", "paper_pdf_sha256": "9e11787b05769008c1480d2caca9641fce9c07ac5040a0a36642108943db4de9", "paper_pdf_bytes": 1395993, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/uda", "code_repository": "google-research/uda", "code_commit": "960684e363251772a5938451d4d2bc0f1da9e24b", "code_archive": "repos/ByeL1R4FvS.zip", "code_archive_sha256": "00df1a73bf81e9dbdd95d5fe59b152272e3d391f5ee2e5e9e2eeaafa5ba29c49", "code_archive_bytes": 332834, "code_file_count": 44, "code_extensions": {".py": 30, ".sh": 14}, "github_disk_usage_kb": 342, "github_languages": {"Python": 234861, "Shell": 19001}, "github_archived": true, "github_pushed_at": "2021-08-28T07:16:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unsupervised-data-augmentation-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7cEMkTu7Lf", "year": 2026, "status": "rejected", "title": "Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs", "authors": ["Xiaoyu Xu", "Xiang Yue", "Yang Liu", "Qingqing Ye", "Huadi Zheng", "Peizhao Hu", "Minxin Du", "Haibo Hu"], "authorids": ["~Xiaoyu_Xu6", "~Xiang_Yue1", "~Yang_Liu3", "~Qingqing_Ye1", "~Huadi_Zheng1", "~Peizhao_Hu2", "~Minxin_Du1", "~Haibo_Hu2"], "authors_source": "OpenReview API", "abstract": "Unlearning in large language models (LLMs) aims to remove specified data, but its efficacy is typically assessed with task-level metrics like accuracy and perplexity. \nWe demonstrate that these metrics are often misleading, as models can appear to forget while their original behavior is easily restored through minimal fine-tuning. \nThis phenomenon of \\emph{reversibility} suggests that information is merely suppressed, not genuinely erased. To address this critical evaluation gap, we introduce a \\emph{representation-level analysis framework}. \nOur toolkit comprises PCA-based similarity and shift, centered kernel alignment (CKA), and Fisher information, complemented by a summary metric, the mean PCA distance, to measure representational drift. \nApplying this framework across six unlearning methods, three data domains, and two LLMs, we identify four distinct forgetting regimes based on their \\emph{reversibility} and \\emph{catastrophicity}. \nOur analysis reveals that achieving the ideal state--irreversible, non-catastrophic forgetting--is exceptionally challenging. \nBy probing the limits of unlearning, we identify a case of seemingly irreversible, targeted forgetting, offering new insights for designing more robust erasure algorithms. \nOur findings expose a fundamental gap in current evaluation practices and establish a representation-level foundation for trustworthy unlearning.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "PXpCTA3ggy", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10009/Reviewer_HtjF"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper highlights the drawback of relying solely on task-level metrics (e.g., accuracy, perplexity) for evaluating unlearning in LLMs, since these metrics cannot distinguish genuine erasure from superficial forgetting. To bridge this evaluation gap, the paper introduces a **representation-level analysis framework** to measure representational drift and categorize unlearning behavior into four regimes. The study concludes that achieving the ideal state—**irreversible and non-catastrophic forgetting**—is extremely challenging, and further provides a method combination that achieves a *seemingly* irreversible, non-catastrophic form of forgetting.", "review_text": "This paper highlights the drawback of relying solely on task-level metrics (e.g., accuracy, perplexity) for evaluating unlearning in LLMs, since these metrics cannot distinguish genuine erasure from superficial forgetting. To bridge this evaluation gap, the paper introduces a **representation-level analysis framework** to measure representational drift and categorize unlearning behavior into four regimes. The study concludes that achieving the ideal state—**irreversible and non-catastrophic forgetting**—is extremely challenging, and further provides a method combination that achieves a *seemingly* irreversible, non-catastrophic form of forgetting.", "strengths": "- The paper is well-written and easy to follow.\n- It clearly identifies the limitations of current task-level evaluations and proposes a **representation-level toolkit** that goes beyond surface metrics.\n- Provides clear definitions and a systematic taxonomy of forgetting regimes.", "weaknesses": "- Table 2 demonstrates the weakness of task-level metrics, but it would be stronger to include results on the **Qwen2.5-7B** model to further consolidate this finding.\n- It remains unclear whether the same observations hold for **smaller (3B) or other model families (Llama)**.\n- The framework measures representational drift but does not formally assess **privacy leakage**; the notion of “irreversible forgetting” is still heuristic.\n- The proposed solution is interesting, but **cross-model validation** would strengthen its generality.", "questions": "- Does the proposed framework also generalize to **LLaMA** or **Qwen3** models?\n- Could the **mean PCA distance** be correlated with formal privacy metrics such as **MIA AUC** in a consistent way?\n- In Tables 2 and 3, how relearning is conducted?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper highlights the drawback of relying solely on task-level metrics (e.g., accuracy, perplexity) for evaluating unlearning in LLMs, since these metrics cannot distinguish genuine erasure from superficial forgetting. To bridge this evaluation gap, the paper introduces a **representation-level analysis framework** to measure representational drift and categorize unlearning behavior into four regimes. The study concludes that achieving the ideal state—**irreversible and non-catastrophic forgetting**—is extremely challenging, and further provides a method combination that achieves a *seemingly* irreversible, non-catastrophic form of forgetting.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-written and easy to follow.\n- It clearly identifies the limitations of current task-level evaluations and proposes a **representation-level toolkit** that goes beyond surface metrics.\n- Provides clear definitions and a systematic taxonomy of forgetting regimes.", "weaknesses": "- Table 2 demonstrates the weakness of task-level metrics, but it would be stronger to include results on the **Qwen2.5-7B** model to further consolidate this finding.\n- It remains unclear whether the same observations hold for **smaller (3B) or other model families (Llama)**.\n- The framework measures representational drift but does not formally assess **privacy leakage**; the notion of “irreversible forgetting” is still heuristic.\n- The proposed solution is interesting, but **cross-model validation** would strengthen its generality.", "questions": "- Does the proposed framework also generalize to **LLaMA** or **Qwen3** models?\n- Could the **mean PCA distance** be correlated with formal privacy metrics such as **MIA AUC** in a consistent way?\n- In Tables 2 and 3, how relearning is conducted?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761789405293}, {"id": "FGgD4sEYuV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10009/Reviewer_Ese5"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper studies two aspects of machine unlearning: reversibility and catastrophic forgetting.\nReversibility is probed via relearning, i.e., a one-epoch finetune on the forget set.\nThe authors show that, in many cases, this simple adjustment of model weights by relearning restores much of the original model’s performance.\n\nFrom this, they infer that the knowledge was not effectively removed and can be readily recovered. They further support this by analyzing the intermediate representation space of LLMs and measuring how the unlearned-then-relearned model deviates from the original.\n\nOverall, the paper finds that relearning often recovers performance, presenting this as a failure mode of unlearning and evidence that the underlying knowledge persists.", "review_text": "This paper studies two aspects of machine unlearning: reversibility and catastrophic forgetting.\nReversibility is probed via relearning, i.e., a one-epoch finetune on the forget set.\nThe authors show that, in many cases, this simple adjustment of model weights by relearning restores much of the original model’s performance.\n\nFrom this, they infer that the knowledge was not effectively removed and can be readily recovered. They further support this by analyzing the intermediate representation space of LLMs and measuring how the unlearned-then-relearned model deviates from the original.\n\nOverall, the paper finds that relearning often recovers performance, presenting this as a failure mode of unlearning and evidence that the underlying knowledge persists.", "strengths": "The idea of studying how easily unlearned knowledge can be recovered after unlearning is quite interesting. In particular, applying relearning and then evaluating the model’s recovery is a valuable direction that deserves further exploration.", "weaknesses": "I don’t find the results of this paper particularly surprising. A single step of finetuning on the forget set can naturally bring back the forgotten knowledge. I don’t quite see why the authors expected this not to work.\nAfter all, with more aggressive settings (e.g., two or three additional epochs), one could almost certainly recover the utility on the forget set. Restoring performance through one epoch of finetuning is not unexpected.\n\nIn general, unlearning methods that are truly “irreversible” often achieve this by severely degrading the model, seen as a drop in accuracy on the retain set and overall utility.\nA more interesting direction, in my view, would be to study the sample efficiency of relearning: can we recover performance using only a few samples or perhaps by providing them as in-context examples instead of full retraining?\n\nAlso, I recenlty found a paper on knowledge recovey of machine unlearning [1], I guess this also worth being discussed in this paper.\n[1] Rezaei, Keivan, et al. \"RESTOR: Knowledge Recovery in Machine Unlearning.\" arXiv preprint arXiv:2411.00204 (2024).", "questions": "They are discussed in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies two aspects of machine unlearning: reversibility and catastrophic forgetting.\nReversibility is probed via relearning, i.e., a one-epoch finetune on the forget set.\nThe authors show that, in many cases, this simple adjustment of model weights by relearning restores much of the original model’s performance.\n\nFrom this, they infer that the knowledge was not effectively removed and can be readily recovered. They further support this by analyzing the intermediate representation space of LLMs and measuring how the unlearned-then-relearned model deviates from the original.\n\nOverall, the paper finds that relearning often recovers performance, presenting this as a failure mode of unlearning and evidence that the underlying knowledge persists.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The idea of studying how easily unlearned knowledge can be recovered after unlearning is quite interesting. In particular, applying relearning and then evaluating the model’s recovery is a valuable direction that deserves further exploration.", "weaknesses": "I don’t find the results of this paper particularly surprising. A single step of finetuning on the forget set can naturally bring back the forgotten knowledge. I don’t quite see why the authors expected this not to work.\nAfter all, with more aggressive settings (e.g., two or three additional epochs), one could almost certainly recover the utility on the forget set. Restoring performance through one epoch of finetuning is not unexpected.\n\nIn general, unlearning methods that are truly “irreversible” often achieve this by severely degrading the model, seen as a drop in accuracy on the retain set and overall utility.\nA more interesting direction, in my view, would be to study the sample efficiency of relearning: can we recover performance using only a few samples or perhaps by providing them as in-context examples instead of full retraining?\n\nAlso, I recenlty found a paper on knowledge recovey of machine unlearning [1], I guess this also worth being discussed in this paper.\n[1] Rezaei, Keivan, et al. \"RESTOR: Knowledge Recovery in Machine Unlearning.\" arXiv preprint arXiv:2411.00204 (2024).", "questions": "They are discussed in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761763883198}, {"id": "TIOptiNHG4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission10009/Reviewer_n9vs"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "* This paper evaluates unlearning in two settings: single (where all unlearning requests are available simultaneously) and continual (where unlearning requests arrive sequentially).\n* The paper defines four unlearning regimes that vary along two axes: reversible vs. irreversible (i.e., whether the unlearned knowledge can be recovered by lightweight retraining) and catastrophic vs. non-catastrophic (i.e., whether the unlearning process significantly affects unrelated knowledge).\n* Evaluating 3 unlearning methods and 2 LLMs, the paper first shows that unlearning methods cannot remove knowledge irreversibly. \n* Then, the paper describes and applies several methods for measuring representational similarity between the original LLM and its unlearned and retrained variants, finding that representational similarities correlate with reversibility of unlearning. \n* This is complemented by a theoretical analysis and a short case study of achieving irreversible, non-catastrophic unlearning. \n* The paper concludes that achieving irreversible unlearning remains hard, and representation analysis offers a new perspective on unlearning beyond accuracy on retain and forget sets", "review_text": "* This paper evaluates unlearning in two settings: single (where all unlearning requests are available simultaneously) and continual (where unlearning requests arrive sequentially).\n* The paper defines four unlearning regimes that vary along two axes: reversible vs. irreversible (i.e., whether the unlearned knowledge can be recovered by lightweight retraining) and catastrophic vs. non-catastrophic (i.e., whether the unlearning process significantly affects unrelated knowledge).\n* Evaluating 3 unlearning methods and 2 LLMs, the paper first shows that unlearning methods cannot remove knowledge irreversibly. \n* Then, the paper describes and applies several methods for measuring representational similarity between the original LLM and its unlearned and retrained variants, finding that representational similarities correlate with reversibility of unlearning. \n* This is complemented by a theoretical analysis and a short case study of achieving irreversible, non-catastrophic unlearning. \n* The paper concludes that achieving irreversible unlearning remains hard, and representation analysis offers a new perspective on unlearning beyond accuracy on retain and forget sets", "strengths": "**(S1)** The paper successfully demonstrates that current methods do not achieve irreversible and non-catastrophic unlearning\n\n**(S2)** The introduced taxonomy may be helpful in future work to better systematize and discuss the achievements of new unlearning methods\n\n**(S3)** The paper makes a convincing argument that accuracy metrics alone give an insufficient impression of unlearning success, and analyzing the model's internal representations can give important insights beyond accuracy\n\n**(S4)** The paper contains one example of successful irreversible and non-catastrophic unlearning, demonstrating that this goal may be achievable.", "weaknesses": "**(W1)** One major concern is originality: The evaluation of unlearning methods successfully confirms that current methods do not achieve irreversible unlearning, but this is a known fact, as the paper also mentions (e.g., [24]). Likewise, the observation that models break down when applying multiple edits in continual learning has been reported before, e.g. [a]. Finally, none of the proposed metrics for representation analysis is novel.\n\n**(W2)** The paper does not contain any actionable insights. The representation analysis confirms that larger representational dissimilarity correlates with irreversible unlearning, but this is expected as the original model becomes harder to reconstruct the further the parameters move away from it. More importantly, the paper does not give practical tools, for example, how representation similarity can reliably predict successful and irreversible unlearning, which would be very helpful in practice. Overall, the takeaways from this analysis remain unclear: What is the reader to conclude beyond the observation that a larger representation shift correlates with irreversible unlearning?\n\n**(W3)** The theoretical analysis mirrors this problem: It mainly shows how a larger distortion of weights leads to greater dissimilarity (which is intuitive). The connection to irreversible unlearning is not formalized but only claimed in the paragraph starting with line 412. Sec. 5.2 additionally discusses that model outputs are not a reliable indicator of unlearning. However, this is also not a significant finding, as keeping all LLM parameters except those of the last layer frozen while randomizing parameters in the last layer will yield a practically random model (from the black-box perspective), while it is very likely that most information and knowledge learned by the model will continue to be accessible from earlier layers.\n\n**(W4)** The paper overall focuses on a narrow setting where unlearning and catastrophic forgetting are measured through fixed forget and retain sets. However, it does not consider recovering unlearned knowledge through prompt attacks (e.g., [b]) or mechanistic interpretability (e.g., [c], only intended as an example, not available before submission deadline).\n\n**(W5)** LLM unlearning is a popular research area with many methods. Claims that are meant to be generalizable to the entire field, such as the one in this paper, need to be either evaluated on a large set of methods or require a motivation for why the chosen set of unlearning methods is representative and will give such generalizable insights. This aspect can be expanded upon in the current paper.\n\n**(W6)** The experiment in 5.3. appears very interesting, because it directly targets the case of irreversible, non-catastrophic unlearning. This experiment could be one starting to inform more successful unlearning methods. Therefore, I think it would be very interesting to expand this perspective. One concern I have is to what extent this observation is due to the \"more constrained relearning conditions\" vs. actually successful unlearning.\n\n**(W7)** The supplementary material contains a large number of plots showing the representation similarity measures for different LLM layers. These plots are not individually interpreted or put in context. Their role in the paper is therefore doubtful. If they do not add any tangible value to the paper, consider removing them.\n\n### References\n[a] Thede et al.: Understanding the limits of lifelong knowledge editing in llms. In arXiv, 2025\\\n[b] Patil et al.: Can sensitive information be deleted from llms? objectives for defending against extraction attacks. In ICLR, 2024\\\n[c] Cywinski et al.: Eliciting Secret Knowledge from Language Models. In arXiv, 2025", "questions": "* Which novel perspectives on unlearning does this paper give the community beyond confirming known problems with current methods?\n* Which actionable improvements in LLM unlearning are informed by the representation analysis? What are the main insights beyond the expected observation that higher representation dissimilarity correlates with irreversible unlearning?\n* How can we motivate the findings in this paper to generalize to most methods for LLM unlearning, even those not evaluated in the paper? How about extraction attacks beyond retraining?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "* This paper evaluates unlearning in two settings: single (where all unlearning requests are available simultaneously) and continual (where unlearning requests arrive sequentially).\n* The paper defines four unlearning regimes that vary along two axes: reversible vs. irreversible (i.e., whether the unlearned knowledge can be recovered by lightweight retraining) and catastrophic vs. non-catastrophic (i.e., whether the unlearning process significantly affects unrelated knowledge).\n* Evaluating 3 unlearning methods and 2 LLMs, the paper first shows that unlearning methods cannot remove knowledge irreversibly. \n* Then, the paper describes and applies several methods for measuring representational similarity between the original LLM and its unlearned and retrained variants, finding that representational similarities correlate with reversibility of unlearning. \n* This is complemented by a theoretical analysis and a short case study of achieving irreversible, non-catastrophic unlearning. \n* The paper concludes that achieving irreversible unlearning remains hard, and representation analysis offers a new perspective on unlearning beyond accuracy on retain and forget sets", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "**(S1)** The paper successfully demonstrates that current methods do not achieve irreversible and non-catastrophic unlearning\n\n**(S2)** The introduced taxonomy may be helpful in future work to better systematize and discuss the achievements of new unlearning methods\n\n**(S3)** The paper makes a convincing argument that accuracy metrics alone give an insufficient impression of unlearning success, and analyzing the model's internal representations can give important insights beyond accuracy\n\n**(S4)** The paper contains one example of successful irreversible and non-catastrophic unlearning, demonstrating that this goal may be achievable.", "weaknesses": "**(W1)** One major concern is originality: The evaluation of unlearning methods successfully confirms that current methods do not achieve irreversible unlearning, but this is a known fact, as the paper also mentions (e.g., [24]). Likewise, the observation that models break down when applying multiple edits in continual learning has been reported before, e.g. [a]. Finally, none of the proposed metrics for representation analysis is novel.\n\n**(W2)** The paper does not contain any actionable insights. The representation analysis confirms that larger representational dissimilarity correlates with irreversible unlearning, but this is expected as the original model becomes harder to reconstruct the further the parameters move away from it. More importantly, the paper does not give practical tools, for example, how representation similarity can reliably predict successful and irreversible unlearning, which would be very helpful in practice. Overall, the takeaways from this analysis remain unclear: What is the reader to conclude beyond the observation that a larger representation shift correlates with irreversible unlearning?\n\n**(W3)** The theoretical analysis mirrors this problem: It mainly shows how a larger distortion of weights leads to greater dissimilarity (which is intuitive). The connection to irreversible unlearning is not formalized but only claimed in the paragraph starting with line 412. Sec. 5.2 additionally discusses that model outputs are not a reliable indicator of unlearning. However, this is also not a significant finding, as keeping all LLM parameters except those of the last layer frozen while randomizing parameters in the last layer will yield a practically random model (from the black-box perspective), while it is very likely that most information and knowledge learned by the model will continue to be accessible from earlier layers.\n\n**(W4)** The paper overall focuses on a narrow setting where unlearning and catastrophic forgetting are measured through fixed forget and retain sets. However, it does not consider recovering unlearned knowledge through prompt attacks (e.g., [b]) or mechanistic interpretability (e.g., [c], only intended as an example, not available before submission deadline).\n\n**(W5)** LLM unlearning is a popular research area with many methods. Claims that are meant to be generalizable to the entire field, such as the one in this paper, need to be either evaluated on a large set of methods or require a motivation for why the chosen set of unlearning methods is representative and will give such generalizable insights. This aspect can be expanded upon in the current paper.\n\n**(W6)** The experiment in 5.3. appears very interesting, because it directly targets the case of irreversible, non-catastrophic unlearning. This experiment could be one starting to inform more successful unlearning methods. Therefore, I think it would be very interesting to expand this perspective. One concern I have is to what extent this observation is due to the \"more constrained relearning conditions\" vs. actually successful unlearning.\n\n**(W7)** The supplementary material contains a large number of plots showing the representation similarity measures for different LLM layers. These plots are not individually interpreted or put in context. Their role in the paper is therefore doubtful. If they do not add any tangible value to the paper, consider removing them.\n\n### References\n[a] Thede et al.: Understanding the limits of lifelong knowledge editing in llms. In arXiv, 2025\\\n[b] Patil et al.: Can sensitive information be deleted from llms? objectives for defending against extraction attacks. In ICLR, 2024\\\n[c] Cywinski et al.: Eliciting Secret Knowledge from Language Models. In arXiv, 2025", "questions": "* Which novel perspectives on unlearning does this paper give the community beyond confirming known problems with current methods?\n* Which actionable improvements in LLM unlearning are informed by the representation analysis? What are the main insights beyond the expected observation that higher representation dissimilarity correlates with irreversible unlearning?\n* How can we motivate the findings in this paper to generalize to most methods for LLM unlearning, even those not evaluated in the paper? How about extraction attacks beyond retraining?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761422442464}], "openreview_url": "https://openreview.net/forum?id=7cEMkTu7Lf", "arxiv_id": "2505.16831", "paper_pdf": "papers/7cEMkTu7Lf.pdf", "paper_pdf_sha256": "d7c36abe236bdc5a53abbaaece6a7030abf69cfff38d757f86190275da20a19a", "paper_pdf_bytes": 6000515, "paper_pdf_source": "openreview", "code_url": "https://github.com/XiaoyuXU1/Representational_Analysis_Tools", "code_repository": "XiaoyuXU1/Representational_Analysis_Tools", "code_commit": "d0ddda6e794ed35b1c0e1c11a3d92e0787443da0", "code_archive": "repos/7cEMkTu7Lf.zip", "code_archive_sha256": "ac2b73617749ee2b5bbcc51e62a3af0cc7c8750c771f6a023ce1e549e68b5b7c", "code_archive_bytes": 933904, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 1533, "github_languages": {"Python": 26430}, "github_archived": false, "github_pushed_at": "2025-05-23T02:53:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unlearning-isn-t-deletion-investigating"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6Mdvq0bPyG", "year": 2025, "status": "rejected", "title": "EfficientQAT: Efficient Quantization-Aware Training for Large Language Models", "authors": ["Mengzhao Chen", "Wenqi Shao", "Peng Xu", "Jiahao Wang", "Peng Gao", "Kaipeng Zhang", "Ping Luo"], "authorids": ["~Mengzhao_Chen1", "~Wenqi_Shao2", "~Peng_Xu11", "~Jiahao_Wang1", "~Peng_Gao3", "~Kaipeng_Zhang1", "~Ping_Luo2"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) are crucial in modern natural language processing and artificial intelligence. However, they face challenges in managing their significant memory requirements. Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources.  To address this, we propose Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP).  To the best of our knowledge, Block-AP is the first method to enable direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios by enhancing the solution space during optimization. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules. Extensive experiments demonstrate that EfficientQAT outperforms previous quantization methods across a range of models, including base LLMs, instruction-tuned LLMs, and multimodal LLMs, with scales from 7B to 70B parameters at various quantization bits. For instance, EfficientQAT obtains a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours, with less than 3 points accuracy degradation compared to the full precision (69.48 vs. 72.41).", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "XvHssWoXM8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6032/Reviewer_6j3y"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes EfficientQAT, a novel quantization-aware training (QAT) framework tailored for large language models (LLMs). Aiming to address the high memory and computational demands of traditional QAT, EfficientQAT introduces a two-phase approach: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables training of all parameters within each block, increasing flexibility and optimization efficiency, while E2E-QP further enhances model performance by training quantization parameters across all blocks. Experimental results demonstrate that EfficientQAT outperforms existing quantization methods in accuracy and memory efficiency across LLMs of varying sizes, from 7B to 70B parameters, at low-bit settings.", "review_text": "The paper proposes EfficientQAT, a novel quantization-aware training (QAT) framework tailored for large language models (LLMs). Aiming to address the high memory and computational demands of traditional QAT, EfficientQAT introduces a two-phase approach: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables training of all parameters within each block, increasing flexibility and optimization efficiency, while E2E-QP further enhances model performance by training quantization parameters across all blocks. Experimental results demonstrate that EfficientQAT outperforms existing quantization methods in accuracy and memory efficiency across LLMs of varying sizes, from 7B to 70B parameters, at low-bit settings.", "strengths": "1. This work adopts a two-phase approach to effectively minimizes accuracy loss even at lower bit levels, which is novel. It is the first method to directly train all parameters in a block-wise fashion, minimizing accuracy loss in low-bit settings by expanding the solution space for optimization. Following this, E2E-QP focuses solely on training the quantization parameters (step sizes) in an end-to-end manner, enhancing the performance of quantized models by accounting for interactions across all sub-modules.\n\n2. Great performance. In terms of training speed, EfficientQAT can obtain a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours with less low accuracy degradation, getting better acceleration performance than other baseline.", "weaknesses": "1. Not enough novelty. The main contribution appears to be the proposed training pipeline, but this pipeline does not introduce substantial advancements beyond existing techniques. While the combination of block-wise and end-to-end training is interesting, it builds on straightforward adaptations of known methods rather than providing an innovative or fundamentally new approach.\n\n2. Unfair comparison. The proposed method only did weight quantization, but many of baselines were using both activation and weight quantization.\n\n3. Performance is not good enough compared to baselines. For example, in Table 3, the performance of the proposed method is worse than QuIP\\# and AQLM almost in all settings. Also in Table 15, in model 2-7B, the accuracy of the EfficientQAT is the lowest among all methods.", "questions": "In the baselines, are they using the same sampling numbers as EfficientQAT?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes EfficientQAT, a novel quantization-aware training (QAT) framework tailored for large language models (LLMs). Aiming to address the high memory and computational demands of traditional QAT, EfficientQAT introduces a two-phase approach: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables training of all parameters within each block, increasing flexibility and optimization efficiency, while E2E-QP further enhances model performance by training quantization parameters across all blocks. Experimental results demonstrate that EfficientQAT outperforms existing quantization methods in accuracy and memory efficiency across LLMs of varying sizes, from 7B to 70B parameters, at low-bit settings.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This work adopts a two-phase approach to effectively minimizes accuracy loss even at lower bit levels, which is novel. It is the first method to directly train all parameters in a block-wise fashion, minimizing accuracy loss in low-bit settings by expanding the solution space for optimization. Following this, E2E-QP focuses solely on training the quantization parameters (step sizes) in an end-to-end manner, enhancing the performance of quantized models by accounting for interactions across all sub-modules.\n\n2. Great performance. In terms of training speed, EfficientQAT can obtain a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours with less low accuracy degradation, getting better acceleration performance than other baseline.", "weaknesses": "1. Not enough novelty. The main contribution appears to be the proposed training pipeline, but this pipeline does not introduce substantial advancements beyond existing techniques. While the combination of block-wise and end-to-end training is interesting, it builds on straightforward adaptations of known methods rather than providing an innovative or fundamentally new approach.\n\n2. Unfair comparison. The proposed method only did weight quantization, but many of baselines were using both activation and weight quantization.\n\n3. Performance is not good enough compared to baselines. For example, in Table 3, the performance of the proposed method is worse than QuIP\\# and AQLM almost in all settings. Also in Table 15, in model 2-7B, the accuracy of the EfficientQAT is the lowest among all methods.", "questions": "In the baselines, are they using the same sampling numbers as EfficientQAT?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730684914563}, {"id": "qeLjIOWej7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6032/Reviewer_14pf"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes an efficient Quantization-Aware Training (QAT) method for LLMs.\nIn detail, the paper introduces the block-wise weight-only QAT (Block-AP) to reduce the memory cost during training and further optimization with the training of scale in weight quantizers (E2E-QP).\nThe experiments show that the proposed QAT method can achieve better performance than previous quantization works.\nFor example, a 2-bit Llama-2-70B model trained on a single A100-80GB GPU with EfficientQAT.", "review_text": "This paper proposes an efficient Quantization-Aware Training (QAT) method for LLMs.\nIn detail, the paper introduces the block-wise weight-only QAT (Block-AP) to reduce the memory cost during training and further optimization with the training of scale in weight quantizers (E2E-QP).\nThe experiments show that the proposed QAT method can achieve better performance than previous quantization works.\nFor example, a 2-bit Llama-2-70B model trained on a single A100-80GB GPU with EfficientQAT.", "strengths": "1. Block-wise QAT reduces the GPU memory requirement and total training time.\n2. The method performs well on small models (LLaMA-2-7B and LLaMA-3-8B) with lower than 4 bits.", "weaknesses": "1. The novelty of this paper is limited. The block-wise QAT is not novel as the block-wise methods are commonly used in quantization and pruning methods for LLMs. For example, [1] [2] [3] [4] [5] [6] adopted layerly strategy in their works, the famous one is GPTQ [2].\n2. The work focuses on the weight-only quantization, while the comparison works contain many weight and activation both quantized methods including OmniQuant and LLM-QAT, which shows unfairness.\n3. The work did not achieve good results with 2,3,4-bit weight quantization with LLaMA2 and LLaMA3 according to Table 1. For example, for uniform quantization, the paper did not achieve better results with 4-bit on LLaMA-2-7B, LLaMA-2-13B, and LLaMA-3-70B. And with 3-bit on LLaMA-3-70B, the paper even performs worse than AWQ which is a post-training quantization method (where even denotes their results in bold).\n4. The paper only includes the comparison with QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA methods in Table 4, while not include these methods in the main results table.\n5. This paper adopts 4096 samples in RedPajama datasets with 2048 sequence length for the Block-AP and 4096 sequence length for the E2E-QP. Thus, the comparison with those PTQ works (AWQ, OminiQuant and GPTQ) is unfair. The paper did not explain the setting of those PTQ works, if they are also use such amount of data with such sequence length for calibration? As according to the Figure 3, the proposed method is sensitive to the number of samples range from 128 to 4096, while the GPTQ only adopts 128 calibration samples from the training dataset of Wiki or C4 in their original setting. Besides, according to Table 13, the proposed method performs worse when adopting Wiki or C4 as training dataset.\n6. As for the training time, the paper should include the LoRA-based methods for comparison including those in Table 4: QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA. Also, the post-training quantization methods are also needed to be included.\n7. The quantization overhead for LLMs mainly caused from the activation quantization, which this paper did not take into consideration, even the 16 bit activation results are not included.\n8. The ablation for scale optimization (E2E-QP) with weights from post-training quantization methods compared to Block-AP weights is needed.\n\n\n[1] Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels \\\n[2] GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers \\\n[3] Streamlining Redundant Layers to Compress Large Language Models \\\n[4] Compressing Large Language Models by Streamlining the Unimportant Layer \\\n[5] Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels \\\n[6] Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning", "questions": "1. How many samples and what kind of datasets are used for the calibration of those post-training quantization methods? What is the detailed experiment setup for other methods?\n2. The zero-shot accuracy results and ppl results for the LoRA based methods including: QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA.\n3. How about the results with 16 bit activation quantization?\n4. Although this work adopts block-wise QAT method, this work still adopts the full parameter fine-tuning, which costs more resource compared to LoRA. Meanwhile, the LoRA based methods can optimize the model globally, while the proposed method can only optimize the model within blocks (although the further optimization of scales is global). Thus, the question is that, the optimization is brought by the Block-AP or the E2E-QP? What if directly using GPTQ, AWQ or QA-LoRA (or other LoRA based methods) weights and using E2E-QP for further optimization?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an efficient Quantization-Aware Training (QAT) method for LLMs.\nIn detail, the paper introduces the block-wise weight-only QAT (Block-AP) to reduce the memory cost during training and further optimization with the training of scale in weight quantizers (E2E-QP).\nThe experiments show that the proposed QAT method can achieve better performance than previous quantization works.\nFor example, a 2-bit Llama-2-70B model trained on a single A100-80GB GPU with EfficientQAT.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Block-wise QAT reduces the GPU memory requirement and total training time.\n2. The method performs well on small models (LLaMA-2-7B and LLaMA-3-8B) with lower than 4 bits.", "weaknesses": "1. The novelty of this paper is limited. The block-wise QAT is not novel as the block-wise methods are commonly used in quantization and pruning methods for LLMs. For example, [1] [2] [3] [4] [5] [6] adopted layerly strategy in their works, the famous one is GPTQ [2].\n2. The work focuses on the weight-only quantization, while the comparison works contain many weight and activation both quantized methods including OmniQuant and LLM-QAT, which shows unfairness.\n3. The work did not achieve good results with 2,3,4-bit weight quantization with LLaMA2 and LLaMA3 according to Table 1. For example, for uniform quantization, the paper did not achieve better results with 4-bit on LLaMA-2-7B, LLaMA-2-13B, and LLaMA-3-70B. And with 3-bit on LLaMA-3-70B, the paper even performs worse than AWQ which is a post-training quantization method (where even denotes their results in bold).\n4. The paper only includes the comparison with QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA methods in Table 4, while not include these methods in the main results table.\n5. This paper adopts 4096 samples in RedPajama datasets with 2048 sequence length for the Block-AP and 4096 sequence length for the E2E-QP. Thus, the comparison with those PTQ works (AWQ, OminiQuant and GPTQ) is unfair. The paper did not explain the setting of those PTQ works, if they are also use such amount of data with such sequence length for calibration? As according to the Figure 3, the proposed method is sensitive to the number of samples range from 128 to 4096, while the GPTQ only adopts 128 calibration samples from the training dataset of Wiki or C4 in their original setting. Besides, according to Table 13, the proposed method performs worse when adopting Wiki or C4 as training dataset.\n6. As for the training time, the paper should include the LoRA-based methods for comparison including those in Table 4: QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA. Also, the post-training quantization methods are also needed to be included.\n7. The quantization overhead for LLMs mainly caused from the activation quantization, which this paper did not take into consideration, even the 16 bit activation results are not included.\n8. The ablation for scale optimization (E2E-QP) with weights from post-training quantization methods compared to Block-AP weights is needed.\n\n\n[1] Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels \\\n[2] GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers \\\n[3] Streamlining Redundant Layers to Compress Large Language Models \\\n[4] Compressing Large Language Models by Streamlining the Unimportant Layer \\\n[5] Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels \\\n[6] Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning", "questions": "1. How many samples and what kind of datasets are used for the calibration of those post-training quantization methods? What is the detailed experiment setup for other methods?\n2. The zero-shot accuracy results and ppl results for the LoRA based methods including: QLoRA (with GPTQ weights), QA-LoRA, IR-QLoRA and PEQA.\n3. How about the results with 16 bit activation quantization?\n4. Although this work adopts block-wise QAT method, this work still adopts the full parameter fine-tuning, which costs more resource compared to LoRA. Meanwhile, the LoRA based methods can optimize the model globally, while the proposed method can only optimize the model within blocks (although the further optimization of scales is global). Thus, the question is that, the optimization is brought by the Block-AP or the E2E-QP? What if directly using GPTQ, AWQ or QA-LoRA (or other LoRA based methods) weights and using E2E-QP for further optimization?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730395410110}, {"id": "ybvygIBHJL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6032/Reviewer_u4Hv"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources. To address this, this paper proposes Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules.", "review_text": "Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources. To address this, this paper proposes Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules.", "strengths": "Extensive experiments demonstrate that EfficientQAT outperforms previous quantization methods across a range of models, including\nbase LLMs, instruction-tuned LLMs, and multimodal LLMs, with scales from 7B to 70B parameters at various quantization bits. For instance, EfficientQAT obtains a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours, with less than 3 points accuracy degradation compared to the full precision (69.48 vs. 72.41).", "weaknesses": "The novelty of the proposed method may be limited. To reduce the training cost, this paper generally only trains part of the model during quantization and freeze other parts, thus saving memory. This idea is straightforward and has been investigated in other works, such as (Li et al., 2021; Shao et al., 2023). The quantization method generally follows very traditional QAT method to learn model weights and quantization parameters. The technical contribution may be limited.  \n\nThe comparisons of training time with existing methods in table 9 does not seem to be solid. Training time is determined by multiple factors such as algorithms and training data. Typically, it takes more time to train if using more training data. To make a fair comparison, it is better to try using the same amount of data for training of all methods.   If it uses less training time because of less training data, it is hard to say that it is more training efficient. It may be better to discuss this training data issue in training time comparison. \n\nAs this paper proposes to train all parameters in a block-wise manner, a more direct baseline is to train the full model with all blocks during quantization. It is better to compare with this baseline to demonstrate the training efficiency such as final accuracy and training time.  For example, as the proposed method only trains one block in a time, and it needs to train multiple rounds as the model has multiple blocks, does it really use less training time compared with training all blocks in one round? And what is the PPL or accuracy performance compared with training all blocks in one round? The current baselines does not seem to cover this baseline. The LLM-QAT adopts knowledge distillation, which is different from the setting in this paper. It is better to discuss the comparison with the straightforward baseline to train all blocks. \n\nThe baselines use various finetuning or calibration datasets or training settings, such as C4 for GPTQ, Pile for AWQ, and so on. It is hard to say whether the performance difference is introduced by the proposed method or the different finetuning dataset or settings. It is better to provide more discussion for this dataset or setting issue during quantization. \n\nAWQ in table 1 can perform better than the proposed method in some cases. It is better to discuss this issue. As a post training quantization method, AWQ typically costs less resource than QAT methods. It is better to discuss why it can lead to a better performance.", "questions": "See the weakness. \n\nIt may be better to discuss this training data issue in training time comparison. \n\nIt is better to discuss the comparison with the straightforward baseline to train all blocks.\n\nIt is better to provide more discussion for this dataset or setting issue during quantization.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources. To address this, this paper proposes Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). Block-AP enables direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "Extensive experiments demonstrate that EfficientQAT outperforms previous quantization methods across a range of models, including\nbase LLMs, instruction-tuned LLMs, and multimodal LLMs, with scales from 7B to 70B parameters at various quantization bits. For instance, EfficientQAT obtains a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours, with less than 3 points accuracy degradation compared to the full precision (69.48 vs. 72.41).", "weaknesses": "The novelty of the proposed method may be limited. To reduce the training cost, this paper generally only trains part of the model during quantization and freeze other parts, thus saving memory. This idea is straightforward and has been investigated in other works, such as (Li et al., 2021; Shao et al., 2023). The quantization method generally follows very traditional QAT method to learn model weights and quantization parameters. The technical contribution may be limited.  \n\nThe comparisons of training time with existing methods in table 9 does not seem to be solid. Training time is determined by multiple factors such as algorithms and training data. Typically, it takes more time to train if using more training data. To make a fair comparison, it is better to try using the same amount of data for training of all methods.   If it uses less training time because of less training data, it is hard to say that it is more training efficient. It may be better to discuss this training data issue in training time comparison. \n\nAs this paper proposes to train all parameters in a block-wise manner, a more direct baseline is to train the full model with all blocks during quantization. It is better to compare with this baseline to demonstrate the training efficiency such as final accuracy and training time.  For example, as the proposed method only trains one block in a time, and it needs to train multiple rounds as the model has multiple blocks, does it really use less training time compared with training all blocks in one round? And what is the PPL or accuracy performance compared with training all blocks in one round? The current baselines does not seem to cover this baseline. The LLM-QAT adopts knowledge distillation, which is different from the setting in this paper. It is better to discuss the comparison with the straightforward baseline to train all blocks. \n\nThe baselines use various finetuning or calibration datasets or training settings, such as C4 for GPTQ, Pile for AWQ, and so on. It is hard to say whether the performance difference is introduced by the proposed method or the different finetuning dataset or settings. It is better to provide more discussion for this dataset or setting issue during quantization. \n\nAWQ in table 1 can perform better than the proposed method in some cases. It is better to discuss this issue. As a post training quantization method, AWQ typically costs less resource than QAT methods. It is better to discuss why it can lead to a better performance.", "questions": "See the weakness. \n\nIt may be better to discuss this training data issue in training time comparison. \n\nIt is better to discuss the comparison with the straightforward baseline to train all blocks.\n\nIt is better to provide more discussion for this dataset or setting issue during quantization.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730235941471}], "openreview_url": "https://openreview.net/forum?id=6Mdvq0bPyG", "arxiv_id": "2407.11062", "paper_pdf": "papers/6Mdvq0bPyG.pdf", "paper_pdf_sha256": "1f67b01ee3663bc406d7a86a5c8a9eba91015f6840f98896da17d75c30a632ab", "paper_pdf_bytes": 470369, "paper_pdf_source": "openreview", "code_url": "https://github.com/OpenGVLab/EfficientQAT", "code_repository": "OpenGVLab/EfficientQAT", "code_commit": "39f37f3b6053681c9b1cd4c9dcaf692d8999459e", "code_archive": "repos/6Mdvq0bPyG.zip", "code_archive_sha256": "0ddb65323fff1409478fbbb077c703aa890f5cf11c39aef5130457d4dd37a0ed", "code_archive_bytes": 68414, "code_file_count": 43, "code_extensions": {".sh": 23, ".py": 20}, "github_disk_usage_kb": 99, "github_languages": {"Python": 158217}, "github_archived": false, "github_pushed_at": "2026-04-10T01:13:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficientqat-efficient-quantization-aware"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hP4iZU8I3Y", "year": 2024, "status": "rejected", "title": "Understanding Inter-Session Intentions via Complex Logical Reasoning", "authors": ["Jiaxin Bai", "Chen Luo", "Zheng Li", "Qingyu Yin", "Yangqiu Song"], "authorids": ["~Jiaxin_Bai1", "~Chen_Luo3", "~Zheng_Li9", "~Qingyu_Yin2", "~Yangqiu_Song1"], "authors_source": "OpenReview API", "abstract": "Understanding user intentions is crucial for enhancing product recommendations, navigation suggestions, and query reformulations. However, user intentions can be intricate, involving multiple sessions and attribute requirements connected by logical operators such as And, Or, and Not. For example, a user may search for Nike or Adidas running shoes across various sessions, with a preference for the color purple. In another case, a user may have purchased a mattress in a previous session and is now seeking a corresponding bed frame without intending to buy another mattress. Prior research on session understanding has not sufficiently addressed how to make product or attribute recommendations for such complex intentions.\nIn this paper, we introduce the task of logical session query answering (LSQA), where sessions are treated as hyperedges of items. We formulate the problem of complex intention understanding as a task of answering logical queries on an aggregated hypergraph of sessions, items, and attributes. We also propose a new model, the Logical Session Graph Transformer (LSGT), which captures interactions among items across different sessions and their logical connections using a transformer structure.\nWe analyze the expressiveness of LSGT and prove the permutation invariance of the inputs for the logical operators. We evaluate LSGT on three datasets and demonstrate that it achieves state-of-the-art results.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "uR9MoVhzuw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2938/Reviewer_kt6F"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this work, the authors focus on product and attribute recommendation by modeling complex user intention. They employ the logical session query answering (LSQA) to formulate the task. The proposed logical session graph transformer (LSGT) model runs on a hyper session graph, which uses a standard transformer structure to encode different entities. Experiments on three real-world datasets demonstrate the effectiveness of LSGT for complex session query answering.", "review_text": "In this work, the authors focus on product and attribute recommendation by modeling complex user intention. They employ the logical session query answering (LSQA) to formulate the task. The proposed logical session graph transformer (LSGT) model runs on a hyper session graph, which uses a standard transformer structure to encode different entities. Experiments on three real-world datasets demonstrate the effectiveness of LSGT for complex session query answering.", "strengths": "1. The motivation that incorporates logical session query answering into product recommendation to model user intent is novel.\n2. The experimental results demonstrate the effectiveness of the proposed LSGT.\n3. The authors theoretically justify the expressiveness and operator-wise permutation invariance of LSGT.", "weaknesses": "1. There are some obvious typos. Authors should scrutinize the writing of the paper.\n(1) In the 5th line of section 4.3, the formula after “The edge feature is denoted as” lacks a proper superscript.\n(2) In Table 5, the first word “Predicti” in explanation of query type 2p should be “Predict”.\n(3) In Table 5, the word “prodict” in explanation of query type ip should be “product”.\n(4) In the 2nd line below Figure 5, the word “descibed” should be “described”.\n2. In Figure 5, the query structure of ip is the same as up and the query structure of 2iS is the same as 2uS. It would be better to distinguish them like [1].\n3. The paper lacks detailed description for figures especially Figure 3, which is hard to understand for readers.\n4. It would be better to evaluate the model’s generalization ability of unseen query structures like [1,2,3].\n\n[1] Jiaxin Bai, Zihao Wang, Hongming Zhang, and Yangqiu Song. 2022. Query2Particles: Knowledge Graph Reasoning with Particle Embeddings. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 2703–2714, Seattle, United States. Association for Computational Linguistics.\n[2] Chen, X., Hu, Z., & Sun, Y. (2022). Fuzzy Logic Based Logical Query Answering on Knowledge Graphs. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4), 3939-3948.\n[3] Jiaxin Bai, Tianshi Zheng, and Yangqiu Song. Sequential query encoding for complex query answering on knowledge graphs. Transactions on Machine Learning Research, 2023. ISSN 2835-8856", "questions": "1. Why do authors not evaluate the model’s generalization ability of unseen query structures like existing works?\n2. Is there an explanation for the author's choice of 14 query structures? Can some other query structures like 2i, and pni be incorporated?\n3. Is it possible to make an ablation study for hypergraph and logical reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors focus on product and attribute recommendation by modeling complex user intention. They employ the logical session query answering (LSQA) to formulate the task. The proposed logical session graph transformer (LSGT) model runs on a hyper session graph, which uses a standard transformer structure to encode different entities. Experiments on three real-world datasets demonstrate the effectiveness of LSGT for complex session query answering.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The motivation that incorporates logical session query answering into product recommendation to model user intent is novel.\n2. The experimental results demonstrate the effectiveness of the proposed LSGT.\n3. The authors theoretically justify the expressiveness and operator-wise permutation invariance of LSGT.", "weaknesses": "1. There are some obvious typos. Authors should scrutinize the writing of the paper.\n(1) In the 5th line of section 4.3, the formula after “The edge feature is denoted as” lacks a proper superscript.\n(2) In Table 5, the first word “Predicti” in explanation of query type 2p should be “Predict”.\n(3) In Table 5, the word “prodict” in explanation of query type ip should be “product”.\n(4) In the 2nd line below Figure 5, the word “descibed” should be “described”.\n2. In Figure 5, the query structure of ip is the same as up and the query structure of 2iS is the same as 2uS. It would be better to distinguish them like [1].\n3. The paper lacks detailed description for figures especially Figure 3, which is hard to understand for readers.\n4. It would be better to evaluate the model’s generalization ability of unseen query structures like [1,2,3].\n\n[1] Jiaxin Bai, Zihao Wang, Hongming Zhang, and Yangqiu Song. 2022. Query2Particles: Knowledge Graph Reasoning with Particle Embeddings. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 2703–2714, Seattle, United States. Association for Computational Linguistics.\n[2] Chen, X., Hu, Z., & Sun, Y. (2022). Fuzzy Logic Based Logical Query Answering on Knowledge Graphs. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4), 3939-3948.\n[3] Jiaxin Bai, Tianshi Zheng, and Yangqiu Song. Sequential query encoding for complex query answering on knowledge graphs. Transactions on Machine Learning Research, 2023. ISSN 2835-8856", "questions": "1. Why do authors not evaluate the model’s generalization ability of unseen query structures like existing works?\n2. Is there an explanation for the author's choice of 14 query structures? Can some other query structures like 2i, and pni be incorporated?\n3. Is it possible to make an ablation study for hypergraph and logical reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698759049009}, {"id": "HIVQIFtAHz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2938/Reviewer_Kyoc"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper formulates an item recommendation task based on the previous session history as a complex logical graph query (named as Logical Session Query Answering). In such a query, items and attributes are nodes, several items can be connected in sessions (hyperedges denoting the order of obtaining the items), relations form projection operators in a query, and other logical operators (intersection, union, negation) combine nodes and projections into a single complex query. Instead of operating on the complete hypergraph of items, sessions, and attributes, the authors decide to operate on the single query level and predict answer entities directly after linearizing the query via the Logical Session Graph Transformer (essentially, a TokenGT from [1]). The authors prove that their transformer is permutation invariant (with respect to intersection and union operators), and run experiments on 3 datasets showing marginal improvements over the baselines.", "review_text": "The paper formulates an item recommendation task based on the previous session history as a complex logical graph query (named as Logical Session Query Answering). In such a query, items and attributes are nodes, several items can be connected in sessions (hyperedges denoting the order of obtaining the items), relations form projection operators in a query, and other logical operators (intersection, union, negation) combine nodes and projections into a single complex query. Instead of operating on the complete hypergraph of items, sessions, and attributes, the authors decide to operate on the single query level and predict answer entities directly after linearizing the query via the Logical Session Graph Transformer (essentially, a TokenGT from [1]). The authors prove that their transformer is permutation invariant (with respect to intersection and union operators), and run experiments on 3 datasets showing marginal improvements over the baselines.", "strengths": "**S1.** The item recommendation task is framed as a complex logical query. While the task per se is not new (LogiRec [2] originally introduced it with projection and intersection operators), this paper extends it to unions and negations and to hypergraphs.\n\n**S2.** Evaluation includes several baselines (that show, on the other hand, that the proposed approach only marginally outperforms existing models, but more on that in W2)", "weaknesses": "Starting from the claimed contributions:\n\n**W1. Task.** The formulated task of Logical Session Query Answering is essentially query answering over hypergraphs. Sessions are n-ary edges, and other relations form 2-ary edges, so the hypergraph has edges of different arity. The temporal aspect of items in session hyperedges (that items follow each other in one session) seems to be of little use as the best-performing models are not using this information anyway. I would recommend the authors to focus the contribution on extending complex query answering to hypergraphs as there is not that much work in that subfield (StarQE is for hyper-relational graphs, and NQE supports both hyper-relational and hypergraphs).\n\n**W2. Encoder + Experimental results.** The proposed logical session graph transformer (LSGT) is just one of the many query linearization strategies, eg, BiQE [3], kgTransformer [4], or SQE [5] that convert the query graph into a sequence with some positional information to be sent jointly into a Transformer. Architecture-wise, LSGT is TokenGT [1] but with a slightly different input format that sends tokens of logical operators. Experimentally, LSGT is very close to SQE [5] (the gap is often <1 MRR point) so it is hard to claim any novelty or effectiveness in this linearization strategy or in a slightly different transformer encoder. \n\n**W3. Theory.** The theoretical study in Section 4.5 is derived from TokenGT and seems to be hardly applicable to the case of logical query answering. TokenGT’s theory of WL expressiveness assumes the graphs are non-relational whereas all logical query graphs studied in this work are relational, i.e., they have labeled edge types. There is a different line of work studying expressiveness of GNNs over relational graphs [6,7] and I would recommend starting from them in order to derive any expressiveness claims. Permutation invariance proofs are rather trivial because the Transformer architecture itself is permutation equivariant.\n\nOverall, I think the paper has more potential if:\n* The authors frame the task as the hypergraph query answering with the full support of first-order logical operators (intersection, union, negation) and demonstrate that several existing Transformer-based models show similar results on 3 benchmarks despite different linearization strategies; \n* Tone down the claims on the _logical session_ QA (it’s a hypergraph), new graph transformer and its expressiveness (TokenGT is not new, theory for non-relational graphs does not apply to relational ones), and state-of-the-art (all Transformer-based models show a very similar performance). \n\nI understand that it would require substantial re-writing of several sections, so I am willing to increase the score if the authors decide to do it during the discussion period. \n\nMinor comments:\n* Too many sentences (especially in Section 3) start with noisy and artificial “however” and “meanwhile”. You don’t have to contrast every sentence to each other every time.\n* $p$ and $q$ denote different things in 4.2 (item and session) and 4.3 (just two nodes) and it is confusing.   \n* 4.4 Learning LSGT -> Training LSGT\n\n**References**\n\n[1] Kim et al. Pure transformers are powerful graph learners. NeurIPS 2022.  \n[2] Tang et al. LogicRec: Recommendation with Users' Logical Requirements. SIGIR’23.  \n[3] Kotnis et al. Answering complex queries in knowledge graphs with bidirectional sequence encoders. AAAI 2021.  \n[4] Liu et al. Mask and reason: Pre-training knowledge graph transformers for complex logical queries. KDD’22.  \n[5] Bai et al. Sequential query encoding for complex query answering on knowledge graphs. TMLR 2023.  \n[6] Barcelo et al. Weisfeiler and Leman Go Relational. LOG 2022.   \n[7] Huang et al. A theory of link prediction via relational Weisfeiler-Leman. NeurIPS 2023.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper formulates an item recommendation task based on the previous session history as a complex logical graph query (named as Logical Session Query Answering). In such a query, items and attributes are nodes, several items can be connected in sessions (hyperedges denoting the order of obtaining the items), relations form projection operators in a query, and other logical operators (intersection, union, negation) combine nodes and projections into a single complex query. Instead of operating on the complete hypergraph of items, sessions, and attributes, the authors decide to operate on the single query level and predict answer entities directly after linearizing the query via the Logical Session Graph Transformer (essentially, a TokenGT from [1]). The authors prove that their transformer is permutation invariant (with respect to intersection and union operators), and run experiments on 3 datasets showing marginal improvements over the baselines.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "**S1.** The item recommendation task is framed as a complex logical query. While the task per se is not new (LogiRec [2] originally introduced it with projection and intersection operators), this paper extends it to unions and negations and to hypergraphs.\n\n**S2.** Evaluation includes several baselines (that show, on the other hand, that the proposed approach only marginally outperforms existing models, but more on that in W2)", "weaknesses": "Starting from the claimed contributions:\n\n**W1. Task.** The formulated task of Logical Session Query Answering is essentially query answering over hypergraphs. Sessions are n-ary edges, and other relations form 2-ary edges, so the hypergraph has edges of different arity. The temporal aspect of items in session hyperedges (that items follow each other in one session) seems to be of little use as the best-performing models are not using this information anyway. I would recommend the authors to focus the contribution on extending complex query answering to hypergraphs as there is not that much work in that subfield (StarQE is for hyper-relational graphs, and NQE supports both hyper-relational and hypergraphs).\n\n**W2. Encoder + Experimental results.** The proposed logical session graph transformer (LSGT) is just one of the many query linearization strategies, eg, BiQE [3], kgTransformer [4], or SQE [5] that convert the query graph into a sequence with some positional information to be sent jointly into a Transformer. Architecture-wise, LSGT is TokenGT [1] but with a slightly different input format that sends tokens of logical operators. Experimentally, LSGT is very close to SQE [5] (the gap is often <1 MRR point) so it is hard to claim any novelty or effectiveness in this linearization strategy or in a slightly different transformer encoder. \n\n**W3. Theory.** The theoretical study in Section 4.5 is derived from TokenGT and seems to be hardly applicable to the case of logical query answering. TokenGT’s theory of WL expressiveness assumes the graphs are non-relational whereas all logical query graphs studied in this work are relational, i.e., they have labeled edge types. There is a different line of work studying expressiveness of GNNs over relational graphs [6,7] and I would recommend starting from them in order to derive any expressiveness claims. Permutation invariance proofs are rather trivial because the Transformer architecture itself is permutation equivariant.\n\nOverall, I think the paper has more potential if:\n* The authors frame the task as the hypergraph query answering with the full support of first-order logical operators (intersection, union, negation) and demonstrate that several existing Transformer-based models show similar results on 3 benchmarks despite different linearization strategies; \n* Tone down the claims on the _logical session_ QA (it’s a hypergraph), new graph transformer and its expressiveness (TokenGT is not new, theory for non-relational graphs does not apply to relational ones), and state-of-the-art (all Transformer-based models show a very similar performance). \n\nI understand that it would require substantial re-writing of several sections, so I am willing to increase the score if the authors decide to do it during the discussion period. \n\nMinor comments:\n* Too many sentences (especially in Section 3) start with noisy and artificial “however” and “meanwhile”. You don’t have to contrast every sentence to each other every time.\n* $p$ and $q$ denote different things in 4.2 (item and session) and 4.3 (just two nodes) and it is confusing.   \n* 4.4 Learning LSGT -> Training LSGT\n\n**References**\n\n[1] Kim et al. Pure transformers are powerful graph learners. NeurIPS 2022.  \n[2] Tang et al. LogicRec: Recommendation with Users' Logical Requirements. SIGIR’23.  \n[3] Kotnis et al. Answering complex queries in knowledge graphs with bidirectional sequence encoders. AAAI 2021.  \n[4] Liu et al. Mask and reason: Pre-training knowledge graph transformers for complex logical queries. KDD’22.  \n[5] Bai et al. Sequential query encoding for complex query answering on knowledge graphs. TMLR 2023.  \n[6] Barcelo et al. Weisfeiler and Leman Go Relational. LOG 2022.   \n[7] Huang et al. A theory of link prediction via relational Weisfeiler-Leman. NeurIPS 2023.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698552408151}, {"id": "i5gX6NjCpM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2938/Reviewer_kouL"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces the task of Logical Session Query Answering (LSQA) and presents a solution called the Logical Session Graph Transformer (LSGT) model. The objective of LSQA is to learn logical queries for observed user interaction sessions. This task could help understand the logical intention of user interactions. The LSGT model achieves this by uniformly representing sessions, items, relations, and logical operators as tokens and leveraging a transformer-based sequential model for encoding.\n\nThe paper provides a theoretical analysis that primarily focuses on demonstrating the expressiveness of the proposed LSGT model. Additionally, comprehensive experiments are conducted to validate the superiority of the proposed model compared to existing baselines.", "review_text": "This paper introduces the task of Logical Session Query Answering (LSQA) and presents a solution called the Logical Session Graph Transformer (LSGT) model. The objective of LSQA is to learn logical queries for observed user interaction sessions. This task could help understand the logical intention of user interactions. The LSGT model achieves this by uniformly representing sessions, items, relations, and logical operators as tokens and leveraging a transformer-based sequential model for encoding.\n\nThe paper provides a theoretical analysis that primarily focuses on demonstrating the expressiveness of the proposed LSGT model. Additionally, comprehensive experiments are conducted to validate the superiority of the proposed model compared to existing baselines.", "strengths": "- This paper proposes the task of Logical Session Query Answering (LSQA), providing an novel paradigm for enhancing applications like session-based recommendation and query recommendation by understanding the logical structures of users' latent intents.\n- The paper provides a theoretical analysis on the expressiveness of the proposed Logical Session Graph Transformer (LSGT) model.\n- The paper innovatively build a unified representation model for items, sessions and logical operators using hypergraphs and sequential models.", "weaknesses": "- Though the proposed task is novel, the proposed technical solution LSGT relies on existing hypergraph structures and transformer architeactures. Such designs have limited differences compared to existing sequential models and graph models. This lower the technical contribution of this paper.\n- The evaluation part could be enhanced with more diverse experiments to conduct a more comprehensive empirical study, such as ablation study, hyperparameter study, case study on the generated queries, and an investigations on the benefits of LSGT brought to downstream tasks like session-based recommendation.\n\nMinor mistake: In the summary for contributions: \"We propose to propose ...\"", "questions": "My concerns would be alleviated if the authors could provide further clarification on the technical novelty aspect and the comprehensiveness of the experiments. Please refer to the weaknesses part for details.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the task of Logical Session Query Answering (LSQA) and presents a solution called the Logical Session Graph Transformer (LSGT) model. The objective of LSQA is to learn logical queries for observed user interaction sessions. This task could help understand the logical intention of user interactions. The LSGT model achieves this by uniformly representing sessions, items, relations, and logical operators as tokens and leveraging a transformer-based sequential model for encoding.\n\nThe paper provides a theoretical analysis that primarily focuses on demonstrating the expressiveness of the proposed LSGT model. Additionally, comprehensive experiments are conducted to validate the superiority of the proposed model compared to existing baselines.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- This paper proposes the task of Logical Session Query Answering (LSQA), providing an novel paradigm for enhancing applications like session-based recommendation and query recommendation by understanding the logical structures of users' latent intents.\n- The paper provides a theoretical analysis on the expressiveness of the proposed Logical Session Graph Transformer (LSGT) model.\n- The paper innovatively build a unified representation model for items, sessions and logical operators using hypergraphs and sequential models.", "weaknesses": "- Though the proposed task is novel, the proposed technical solution LSGT relies on existing hypergraph structures and transformer architeactures. Such designs have limited differences compared to existing sequential models and graph models. This lower the technical contribution of this paper.\n- The evaluation part could be enhanced with more diverse experiments to conduct a more comprehensive empirical study, such as ablation study, hyperparameter study, case study on the generated queries, and an investigations on the benefits of LSGT brought to downstream tasks like session-based recommendation.\n\nMinor mistake: In the summary for contributions: \"We propose to propose ...\"", "questions": "My concerns would be alleviated if the authors could provide further clarification on the technical novelty aspect and the comprehensiveness of the experiments. Please refer to the weaknesses part for details.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698552312464}], "openreview_url": "https://openreview.net/forum?id=hP4iZU8I3Y", "arxiv_id": "2312.13866", "paper_pdf": "papers/hP4iZU8I3Y.pdf", "paper_pdf_sha256": "fa4928c34accc4011efe1c391e6785010ea7fa9a0182225d02b237fbb8be8838", "paper_pdf_bytes": 567896, "paper_pdf_source": "openreview", "code_url": "https://github.com/HKUST-KnowComp/SessionCQA", "code_repository": "HKUST-KnowComp/SessionCQA", "code_commit": "12241090b612f696e40c4c06583473ae18ea74a0", "code_archive": "repos/hP4iZU8I3Y.zip", "code_archive_sha256": "7ae3c5ed1405ded4ae98de99737d89c2297e837ab46b078d3607c9335fe273cd", "code_archive_bytes": 107006, "code_file_count": 52, "code_extensions": {".py": 28, ".sh": 24}, "github_disk_usage_kb": 78, "github_languages": {"Python": 509055, "Shell": 8435}, "github_archived": false, "github_pushed_at": "2024-07-26T08:42:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-inter-session-intentions-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sKDtBKYOdIP", "year": 2023, "status": "rejected", "title": "Beam Tree Recursive Cells", "authors": ["Jishnu Ray Chowdhury", "Cornelia Caragea"], "authorids": ["~Jishnu_Ray_Chowdhury2", "~Cornelia_Caragea2"], "authors_source": "OpenReview API", "abstract": "Recursive Neural Networks (RvNNs) generalize Recurrent Neural Networks (RNNs) by allowing sequential composition in a more flexible order, typically, based on some tree structure. While initially user-annotated tree structures were used, in due time, several approaches were proposed to automatically induce tree-structures from raw text to guide the recursive compositions in RvNNs. In this paper, we present an approach called Beam Tree Recursive Cell (or BT-Cell) based on a simple yet overlooked backpropagation-friendly framework. BT-Cell applies beam search on easy-first parsing for simulating RvNNs with automatic structure-induction. Our results show that BT-Cell achieves near-perfect performance on several aspects of challenging structure-sensitive synthetic tasks like ListOps and also comparable performance in realistic data to other RvNN-based models. We further introduce and analyze several extensions of BT-Cell based on relaxations of the hard top-k operators in beam search. We evaluate the models in different out of distribution splits in both synthetic and realistic data. Additionally, we identify a previously unknown failure case for neural models in generalization to unseen number of arguments in ListOps. We will release our code.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "huT6N2W1l0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5131/Reviewer_Dk8K"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents a differentiable easy-first beam search for structure induction, called Beam Tree Recursive Cells (BT-RC). A couple methods are presented to handle the sparse gradient issue in beam search. There are two sources of sparsity: parent composition and the topk filtering of beams. The two groups of methods presented are Gumbel-BT-GRC, which uses straight-through Gumbel topk for parent composition, and Softpath variants, which use a convex combination of beam elements. The models are evaluated against strong structure-sensitive baselines and perform comparably and in some cases favorably.", "review_text": "I advocate for a weak reject. The method approaches key issues in scaling latent structured models, but is missing crucial analyses. I will be happy to increase my score to accept given more analysis on space/time complexity and empirical speed numbers.\n\nEdit: After updated score from 5->6 after author's response.", "strengths": "*Strengths*\n\nThe accuracy evaluation is thorough, and the results have convinced me that the method is performant. However, one of the biggest drawbacks of latent structure methods is scalability. I would like to better understand the computational complexity of the method and baselines.\n\n*Weaknesses*\n\nThe paper needs an analysis of computational complexity. I believe the method has $O(B^2k)$ runtime. How do the other methods compare? A table with asymptotic runtimes / space complexity is a crucial missing component. An empirical study of the runtime (time per iteration vs sentence length) would greatly improve a scalability argument as well.\n\n\n\n*Comments*\n* The text description of beam search can be placed in the appendix.\n* An illustration of the easy-first beam search with Softpath would be a nice figure to have.\n* The paper is missing a citation for easy-first beam search [1].\n* What is the connection between state space models and beam search mentioned in the last sentence of the paper?\n* I believe beam search + softpath can be interpreted as a continuous relaxation of the sum-and-sample estimator [2], generalized to beam search. The sum-and-sample estimator takes the top-k elements from a proposal distribution and samples an extra element to eliminate bias at the cost of added variance. Instead of sampling to reduce bias, Softpath makes a soft decision that reduces bias less than a sample would but does not add variance.\n\n[1] Ji Ma, Jingbo Zhu, Tong Xiao, and Nan Yang. 2013. Easy-First POS Tagging and Dependency Parsing with Beam Search. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 110–114, Sofia, Bulgaria. Association for Computational Linguistics.\n\n[2] Liu, Runjing et al. “Rao-Blackwellized Stochastic Gradients for Discrete Distributions.” ICML (2019).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents a differentiable easy-first beam search for structure induction, called Beam Tree Recursive Cells (BT-RC). A couple methods are presented to handle the sparse gradient issue in beam search. There are two sources of sparsity: parent composition and the topk filtering of beams. The two groups of methods presented are Gumbel-BT-GRC, which uses straight-through Gumbel topk for parent composition, and Softpath variants, which use a convex combination of beam elements. The models are evaluated against strong structure-sensitive baselines and perform comparably and in some cases favorably.", "strength_and_weaknesses": "*Strengths*\n\nThe accuracy evaluation is thorough, and the results have convinced me that the method is performant. However, one of the biggest drawbacks of latent structure methods is scalability. I would like to better understand the computational complexity of the method and baselines.\n\n*Weaknesses*\n\nThe paper needs an analysis of computational complexity. I believe the method has $O(B^2k)$ runtime. How do the other methods compare? A table with asymptotic runtimes / space complexity is a crucial missing component. An empirical study of the runtime (time per iteration vs sentence length) would greatly improve a scalability argument as well.\n\n\n\n*Comments*\n* The text description of beam search can be placed in the appendix.\n* An illustration of the easy-first beam search with Softpath would be a nice figure to have.\n* The paper is missing a citation for easy-first beam search [1].\n* What is the connection between state space models and beam search mentioned in the last sentence of the paper?\n* I believe beam search + softpath can be interpreted as a continuous relaxation of the sum-and-sample estimator [2], generalized to beam search. The sum-and-sample estimator takes the top-k elements from a proposal distribution and samples an extra element to eliminate bias at the cost of added variance. Instead of sampling to reduce bias, Softpath makes a soft decision that reduces bias less than a sample would but does not add variance.\n\n[1] Ji Ma, Jingbo Zhu, Tong Xiao, and Nan Yang. 2013. Easy-First POS Tagging and Dependency Parsing with Beam Search. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 110–114, Sofia, Bulgaria. Association for Computational Linguistics.\n\n[2] Liu, Runjing et al. “Rao-Blackwellized Stochastic Gradients for Discrete Distributions.” ICML (2019).", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear overall and I believe the work is original, novel, and reproducible.", "summary_of_the_review": "I advocate for a weak reject. The method approaches key issues in scaling latent structured models, but is missing crucial analyses. I will be happy to increase my score to accept given more analysis on space/time complexity and empirical speed numbers.\n\nEdit: After updated score from 5->6 after author's response.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667023523443}, {"id": "LDni-LfnlPl", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5131/Reviewer_muQo"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors present a novel algorithm for recursive neural network processing of sequence inputs. The algorithm combines easy-first parsing techniques with beam search in order to efficiently explore the space of possible latent tree structures during training and inference. They also present soft relaxations of the framework which lead to improved performance. They include an extensive evaluation of their model variants and many baseline models on synthetic and naturalistic language tasks.", "review_text": "This paper presents what seems to me a small increment on existing recursive neural network models, which yields modest performance increases. Interpretation of the models' successes are extremely limited. While I appreciate the attempt at a broad evaluation of many recursive neural network models, the results are not well synthesized beyond tables of numbers and lists of qualitative results.\n\nOn these grounds I recommend rejection. I recommend the authors work on evaluating and interpreting the model results in order to better support future model development, or push for meaningful performance improvements, especially on naturalistic language datasets.", "strengths": "- Impressive synthesis of existing recursive neural network models and evaluation on a level playing field.\n- Novel model with a reasonable selection of variants, with some improvements in performance over baseline models.\n- Limited interpretation and evaluation of why this model works (or why other models work).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors present a novel algorithm for recursive neural network processing of sequence inputs. The algorithm combines easy-first parsing techniques with beam search in order to efficiently explore the space of possible latent tree structures during training and inference. They also present soft relaxations of the framework which lead to improved performance. They include an extensive evaluation of their model variants and many baseline models on synthetic and naturalistic language tasks.", "strength_and_weaknesses": "- Impressive synthesis of existing recursive neural network models and evaluation on a level playing field.\n- Novel model with a reasonable selection of variants, with some improvements in performance over baseline models.\n- Limited interpretation and evaluation of why this model works (or why other models work).", "clarity,_quality,_novelty_and_reproducibility": "- Novelty: The proposed model is a rather small increment on that of Choi et al. (2018), who present an easy-first parsing model which induces tree structure representations for its input, but without beam search decoding. The further model variants (e.g. the Softpath variant, which represents a large distribution over tree representations via a score-weighted combination of vectors) are also not especially surprising.\n- Quality\n\t- Interpretation: The paper attempts an impressive synthesis of the space of models and evaluates them on a level playing field. However, too little space is given to analyzing and understanding the results: the conclusion and much of the result paragraphs read as simple lists of inequalities (X does better than Y on Z), with no interpretation or merely an untested gesture at an interpretation (e.g. \"the memory-augmented RNN style setup in it may be more amenable for argument generalization\").\n\t- Results: The evaluation is limited to synthetic logical reasoning tasks except for sentiment analysis (SST). It demonstrates modest (if any) improvement over existing models in the tasks (and lags behind others especially in the only naturalistic task tested). No error analysis is given to help us understand why these performance differences should be interesting.\n\t\t- Many of the cited papers in this literature evaluate on other naturalistic tasks, e.g. NLI, language modeling; or interpret the latent tree structures induced by their models / use them as unsupervised parsers. I would suggest the authors consider expanding their work to these kinds of evaluations in order for the results to be more comparable.\n\t\t- Even for the most successful results (99% accuracy on ListOps), there is no clear explanation or test showing why this model's success is worth considering, apart from a restatement of the model's design (\"are able to get near perfect performance in length generalization ... because Softpath can allow gradient signals to (softly) truncated paths or beams (which would otherwise be completely truncated)\").", "summary_of_the_review": "This paper presents what seems to me a small increment on existing recursive neural network models, which yields modest performance increases. Interpretation of the models' successes are extremely limited. While I appreciate the attempt at a broad evaluation of many recursive neural network models, the results are not well synthesized beyond tables of numbers and lists of qualitative results.\n\nOn these grounds I recommend rejection. I recommend the authors work on evaluating and interpreting the model results in order to better support future model development, or push for meaningful performance improvements, especially on naturalistic language datasets.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666660288581}, {"id": "7ajkFrF9ex", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5131/Reviewer_bsU1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents BT-Cell, which uses a beam-search style technique to calculate the representation of a sequence with recursive neural networks while automatically determine the best backbone structure. Experiments show improvements in some generalization splits on a synthetic dataset (listops), and results on real datasets are on par with existing state-of-the-art methods.\n", "review_text": "This paper presents BT-Cell, which uses a beam-search style technique to calculate the representation of a sequence with recursive neural networks while automatically determine the best backbone structure. The contribution of this paper is to combine the idea of beam search with latent tree structure learning in recursive neural networks, and demonstrate the effectiveness on a synthetic dataset. While the proposed BT-cell achieves improved performance on listops long sequences, showing a potential on the length-generalization in real NLP applications, most techniques exist in prior work. The paper also lacks of qualitative analysis or gain on real NLP tasks. \n\nI hereby acknowledge the value of this work, but recommend a rejection for this paper as I don't think the content and novelty is sufficient enough for ICLR.\n", "strengths": "## Strengths\n\n- The proposed BT-cell shows improved performance on listops long sequences, showing a potential on the length-generalization in real NLP applications.\n\n- Very comprehensive survey and comparison to existing work.\n\n## Weaknesses \n\n- Lack of qualitative analysis on failure cases: while the improvement by the proposed BT-cell is marginal on listops, it would be good to understand which types of sequences BT-cells help to process.\n\n- Most techniques exist in prior work, while lack of gain on real NLP tasks.\n\n- Missing reference: [1] applies a CKY style algorithm for CCG induction, which improved the performance of generalization on two tasks. Their expected execution results is essentially in the same spirit as this work's list of beam, and both pieces of work focus on generalization.\n\n[1] Mao et al., 2021. [Grammar-Based Grounded Lexicon Learning](https://proceedings.neurips.cc/paper/2021/file/4158f6d19559955bae372bb00f6204e4-Paper.pdf). In NeurIPS. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents BT-Cell, which uses a beam-search style technique to calculate the representation of a sequence with recursive neural networks while automatically determine the best backbone structure. Experiments show improvements in some generalization splits on a synthetic dataset (listops), and results on real datasets are on par with existing state-of-the-art methods.\n", "strength_and_weaknesses": "## Strengths\n\n- The proposed BT-cell shows improved performance on listops long sequences, showing a potential on the length-generalization in real NLP applications.\n\n- Very comprehensive survey and comparison to existing work.\n\n## Weaknesses \n\n- Lack of qualitative analysis on failure cases: while the improvement by the proposed BT-cell is marginal on listops, it would be good to understand which types of sequences BT-cells help to process.\n\n- Most techniques exist in prior work, while lack of gain on real NLP tasks.\n\n- Missing reference: [1] applies a CKY style algorithm for CCG induction, which improved the performance of generalization on two tasks. Their expected execution results is essentially in the same spirit as this work's list of beam, and both pieces of work focus on generalization.\n\n[1] Mao et al., 2021. [Grammar-Based Grounded Lexicon Learning](https://proceedings.neurips.cc/paper/2021/file/4158f6d19559955bae372bb00f6204e4-Paper.pdf). In NeurIPS. \n", "clarity,_quality,_novelty_and_reproducibility": "**Clarity and Quality**: While this paper is generally clear to me, the following points could be improved:\n- The steps in Section 3 could be very hard to digest for those familiar with neither Choi et al. (2018) nor Chowdhury & Caragea (2021). More details with equations would be helpful, especially the ones describing how two children nodes are combined into their potential parent. Here are some additional concrete ideas and questions for presentation:\n  - How is each beam represented? If I understood correctly it should be something like a list of (node, score) pairs, where the number of nodes is determined by how many composition steps have been taken as of now. Is this correct? If so, I think it would be helpful to include precise math formulation.\n  - Including the form of the score function would be good, or at least define the input and output space of the score function.\n\n\n**Novelty**: Most of the techniques exist before. The contribution of this paper is to combine the idea of beam search with latent tree structure learning in recursive neural networks, and demonstrate the effectiveness on a synthetic dataset.\n\n**Reproducibility**: The authors have included the code and experiment details in their supplementary material. While I haven't had a chance to try it myself, I believe these materials are sufficient enough to reproduce the results.\n", "summary_of_the_review": "This paper presents BT-Cell, which uses a beam-search style technique to calculate the representation of a sequence with recursive neural networks while automatically determine the best backbone structure. The contribution of this paper is to combine the idea of beam search with latent tree structure learning in recursive neural networks, and demonstrate the effectiveness on a synthetic dataset. While the proposed BT-cell achieves improved performance on listops long sequences, showing a potential on the length-generalization in real NLP applications, most techniques exist in prior work. The paper also lacks of qualitative analysis or gain on real NLP tasks. \n\nI hereby acknowledge the value of this work, but recommend a rejection for this paper as I don't think the content and novelty is sufficient enough for ICLR.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666645028356}, {"id": "uda4OyOEy4p", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5131/Reviewer_dDzD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new method, called Beam Tree Recursive Cells, for sentence representation recursively using beam search. The BT-cell extracts a beam of parses (using the same mechanism of easy-first parsing, replacing argmax with top-k) when processing a sentence, and then combining all beams at the end for a sentence representation. Although top-k is non-differentiable, the paper shows that it is quite effective enough for training, without the need for back-prop though top-k. The paper demonstrates that BT-cell is effective for artificial tasks Logical Inference, ListOps, and real tasks SST and IMDB. ", "review_text": "The paper doesn't meet the acceptance standard: \n* its quality should be improved so that the the choices and their impacts are understood better\n* the work is quite incremental, i.e. the contribution doesn't seem significant. \n", "strengths": "The BT-cell is a straightforward extension for argmax-based easy-first parsing and thus it is clear and easy to understand. The paper however does't explain why this extension is helpful. \n\nFirstly, the paper lacks analyses about the impact of beam search and found structures among the beams. For instance, by examining the beams, can we find *good* structures that support required compositionality? What if we vary the beam size? What are distributions of scores? Also, as the paper claims that top-k is good enough although it is non-differentiable, would there be an analysis looking into the beams to support the argument given right below eq (3)? \n\nThe experiment setting is pretty unfair to the baselines. The paper compares several variations of BT-cells (with different beam sizes, top-k operators...) against the baselines. But that is not much different from fine-tuning on test sets, where hyper-params are beam size and top-k operator. For instance, in Tab1 we can see that BT-GRC performs very well on 'C', with '+softpath' it does better on '7,8,...,12' but much worse on 'C'. BT-cell with different beam sizes also yields different results ( tab1 vs tab3). However in the end, all the conclusions are for BT-cell in general, rather than some specific BT-cell configuration. \n\nThe paper claims that CYK approaches are expensive, but there's no complexity analysis for the proposed BT-cell. \n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a new method, called Beam Tree Recursive Cells, for sentence representation recursively using beam search. The BT-cell extracts a beam of parses (using the same mechanism of easy-first parsing, replacing argmax with top-k) when processing a sentence, and then combining all beams at the end for a sentence representation. Although top-k is non-differentiable, the paper shows that it is quite effective enough for training, without the need for back-prop though top-k. The paper demonstrates that BT-cell is effective for artificial tasks Logical Inference, ListOps, and real tasks SST and IMDB. ", "strength_and_weaknesses": "The BT-cell is a straightforward extension for argmax-based easy-first parsing and thus it is clear and easy to understand. The paper however does't explain why this extension is helpful. \n\nFirstly, the paper lacks analyses about the impact of beam search and found structures among the beams. For instance, by examining the beams, can we find *good* structures that support required compositionality? What if we vary the beam size? What are distributions of scores? Also, as the paper claims that top-k is good enough although it is non-differentiable, would there be an analysis looking into the beams to support the argument given right below eq (3)? \n\nThe experiment setting is pretty unfair to the baselines. The paper compares several variations of BT-cells (with different beam sizes, top-k operators...) against the baselines. But that is not much different from fine-tuning on test sets, where hyper-params are beam size and top-k operator. For instance, in Tab1 we can see that BT-GRC performs very well on 'C', with '+softpath' it does better on '7,8,...,12' but much worse on 'C'. BT-cell with different beam sizes also yields different results ( tab1 vs tab3). However in the end, all the conclusions are for BT-cell in general, rather than some specific BT-cell configuration. \n\nThe paper claims that CYK approaches are expensive, but there's no complexity analysis for the proposed BT-cell. \n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "* Clarity: the paper is easy to read, but relation between BT-cell and CYK can be made clearer. The difference between them is: CYK approach *fuses* subtree representations; whereas BT-cell keeps them separated and only *fuses* them in the end. Seeing this way, BT-cell  has a strong similarity to CYK with (1) top-k pruning, and (2) replace pooling. \n\n* Quality: the quality of the paper can be improved with experiments on hyper-param choices, analyses on the impact of beam mechanism, and the impact of different components (beam size, top-k operator)\n\n* Originality: the work is quite incremental. There are several ways to see how the work related to existing methods in the literature: for easy-first parsing, the work replaces argmax by top-k. The work can also be seen as a restriction of CYK with \"beam\"-pooling.   ", "summary_of_the_review": "The paper doesn't meet the acceptance standard: \n* its quality should be improved so that the the choices and their impacts are understood better\n* the work is quite incremental, i.e. the contribution doesn't seem significant. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666621647265}], "openreview_url": "https://openreview.net/forum?id=sKDtBKYOdIP", "arxiv_id": "2305.19999", "paper_pdf": "papers/sKDtBKYOdIP.pdf", "paper_pdf_sha256": "dd7127b71e71324ac0d186b152cc7dd0e4f45b8cf83e123a324a41d5823f2350", "paper_pdf_bytes": 508007, "paper_pdf_source": "openreview", "code_url": "https://github.com/JRC1995/BeamTreeRecursiveCells", "code_repository": "JRC1995/BeamTreeRecursiveCells", "code_commit": "30cbaae15635372249c9480c5553be5f654cf63f", "code_archive": "repos/sKDtBKYOdIP.zip", "code_archive_sha256": "38ecc30322d83a36c2585410d6bd6523e6c493bbdede088fc63ff792d37db9e0", "code_archive_bytes": 208896, "code_file_count": 136, "code_extensions": {".py": 135, ".cu": 1}, "github_disk_usage_kb": 154, "github_languages": {"Python": 639287, "Cuda": 5680}, "github_archived": false, "github_pushed_at": "2024-01-28T04:02:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beam-tree-recursive-cells"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7zFokR7k_86", "year": 2022, "status": "rejected", "title": "Learning Symbolic Rules for Reasoning in Quasi-Natural Language", "authors": ["Kaiyu Yang", "Jia Deng"], "authorids": ["~Kaiyu_Yang1", "~Jia_Deng1"], "authors_source": "OpenReview API", "abstract": "Symbolic reasoning, rule-based symbol manipulation, is a hallmark of human intelligence.  However, rule-based systems have had limited success competing with learning-based systems outside formalized domains such as automated theorem proving. We hypothesize that this is due to the manual construction of rules in past attempts. In this work, we ask how we can build a rule-based system that can reason with natural language input but without the manual construction of rules. We propose MetaQNL, a \"Quasi-Natural\" language that can express both formal logic and natural language sentences, and MetaInduce, a learning algorithm that induces MetaQNL rules from training data consisting of questions and answers, with or without intermediate reasoning steps. Our approach achieves state-of-the-art accuracy on multiple reasoning benchmarks; it learns compact models with much less data and produces not only answers but also checkable proofs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "vsnyod7rm9k", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper81/Reviewer_v7TB"], "rating": "", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a symbolic system in Quasi-Natural Language, MetaQNL, which is compatible with both logical inference and quasi natural language expressions, and where the basic building blocks are sentences and rules. The authors also propose MetaInduce, which learns to generalize a set of rules that explains the examples in MetaInduce.\n\nMetaInduce consists of three mains steps:\n1) a rule proposer proposes a set of concrete rules as candidates for each individual example. This set may not be fully correct and may not be the minimal explanations, and they are used to prove the example using forward/backward chaining.\n2) the authors apply a symbolic procedure called anti-unification to generate abstract rules from concrete rules.\n3) and finally, proof paths are encoded in MAX-SAT and a subset of all rules are solved using a MAX-SAT solver to find minimal possible explanations\n\nThis paper evaluates its methods on two synthetic datasets, and the authors claim to learn compact models with much less data, and produces answers as well as proofs", "review_text": "strengths:\n1) this is indeed novel research problem - using a learning system together with a maxsat solver to make logical inference and identify rules.  There is a spectrum where on one end logical inference instances fully in natural language, and on the other end they are fully symbolic, and this paper falls somewhere inbetween.\n2) The paper is generally well-written, and mathematical part of the paper seems to be correct.\n\nWeakness:\n\nI fail to see the real contributions in this paper.  The only novelty seems to me is MetaQNL and MetaInduce can solve formal systems as well as systems represented in quasi natural languages.  However, are there actual applications that can potentially benefit from this problem formulation?\nOne of the major claims of this paper is it can produces checkable proofs, but isn't Proofwriter also capable of generating proofs which seemed more impressive because it is a fully neural system without any explicit encodings of rules and has the potential of working with real languages.  I get that proofwriter has billions of parameters and MetaInduce learns a system with much fewer parameters, but then what is the difference between this and combinatorial optimization?\n\nFurthermore, I failed to see a way to scale up this method.  In the end, it depends on a maxsat solver, which will become intractable quickly when there are more rules.\n\nIn summary, if the authors can present a real-world application that can potentially benefit from their system while other learning systems fail to do so and evaluate their method on a small real-world dataset, I would be less concerned.\n\nSmall comments and questions:\n1) Would appreciate a few citations to support the claim \"At a glance, this may appear a large departure from the conventional wisdom that learning-based systems, particularly deep networks, are far superior to rule-based systems, as history has demonstrated repeatedly.\"  The authors need to be specific what on tasks.  For example, NNs are far behind SAT solvers on propositional formulas (GQSAT as an example for learning-based system).\n2) I fail to understand why rule proposers need to generate concrete rules, aren't all the rules already included in MetaQNL systems?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a symbolic system in Quasi-Natural Language, MetaQNL, which is compatible with both logical inference and quasi natural language expressions, and where the basic building blocks are sentences and rules. The authors also propose MetaInduce, which learns to generalize a set of rules that explains the examples in MetaInduce.\n\nMetaInduce consists of three mains steps:\n1) a rule proposer proposes a set of concrete rules as candidates for each individual example. This set may not be fully correct and may not be the minimal explanations, and they are used to prove the example using forward/backward chaining.\n2) the authors apply a symbolic procedure called anti-unification to generate abstract rules from concrete rules.\n3) and finally, proof paths are encoded in MAX-SAT and a subset of all rules are solved using a MAX-SAT solver to find minimal possible explanations\n\nThis paper evaluates its methods on two synthetic datasets, and the authors claim to learn compact models with much less data, and produces answers as well as proofs", "main_review": "strengths:\n1) this is indeed novel research problem - using a learning system together with a maxsat solver to make logical inference and identify rules.  There is a spectrum where on one end logical inference instances fully in natural language, and on the other end they are fully symbolic, and this paper falls somewhere inbetween.\n2) The paper is generally well-written, and mathematical part of the paper seems to be correct.\n\nWeakness:\n\nI fail to see the real contributions in this paper.  The only novelty seems to me is MetaQNL and MetaInduce can solve formal systems as well as systems represented in quasi natural languages.  However, are there actual applications that can potentially benefit from this problem formulation?\nOne of the major claims of this paper is it can produces checkable proofs, but isn't Proofwriter also capable of generating proofs which seemed more impressive because it is a fully neural system without any explicit encodings of rules and has the potential of working with real languages.  I get that proofwriter has billions of parameters and MetaInduce learns a system with much fewer parameters, but then what is the difference between this and combinatorial optimization?\n\nFurthermore, I failed to see a way to scale up this method.  In the end, it depends on a maxsat solver, which will become intractable quickly when there are more rules.\n\nIn summary, if the authors can present a real-world application that can potentially benefit from their system while other learning systems fail to do so and evaluate their method on a small real-world dataset, I would be less concerned.\n\nSmall comments and questions:\n1) Would appreciate a few citations to support the claim \"At a glance, this may appear a large departure from the conventional wisdom that learning-based systems, particularly deep networks, are far superior to rule-based systems, as history has demonstrated repeatedly.\"  The authors need to be specific what on tasks.  For example, NNs are far behind SAT solvers on propositional formulas (GQSAT as an example for learning-based system).\n2) I fail to understand why rule proposers need to generate concrete rules, aren't all the rules already included in MetaQNL systems?", "summary_of_the_review": "In short, while I acknowledge this work is novel, its setting is not very practical.  This paper would benefit from presenting a potential real-world application, or else some theoretical generalization results on how well their systems can learn.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers."}, "tcdate": 1635997434722}, {"id": "egVauJudI_u", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper81/Reviewer_CPsz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes an algorithm that learns rules from natural language data, and a symbolic system for manipulating these rules, where existing provers can be applied. The objective is to maximize the number of examples in a test set that are consistent with the proposed mode while minimizing the number of rules in the model. The algorithm consists of three steps - given a training example, it proposes concrete rules, abstracts concrete rules into rules with variables, and prunes the resulting rules. ", "review_text": "Pros: \n- The learning algorithm is interesting and the problem of automatically discovering prepositions from examples is important for ATP in formal or informal languages.\n- The paper clearly defines the terms used and explains the methods and experiments well.\n\nCons:\n- As the authors pointed out, the experiments are neither large-scale nor real-world. One result of this is existing methods already achieve good performance, and it's not clear that the proposed method results in better performance.\n- None of the components (theorem prover, rule abstraction, rule pruning) are novel individually.\n\nQuestions:\n1) Are TRUE FALSE and MAPS_TO the only special symbols? It would be helpful to state this. \n2) Are there any unprovable examples in SCAN?\n3) The number of rules and symbols learned by MetaInduce is hard to compare with the number of learned parameters in ProofWriter. Is there a better metric to compare the two methods on? \n4) What are the advantages to using the proposed symbol system instead of first order logic? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an algorithm that learns rules from natural language data, and a symbolic system for manipulating these rules, where existing provers can be applied. The objective is to maximize the number of examples in a test set that are consistent with the proposed mode while minimizing the number of rules in the model. The algorithm consists of three steps - given a training example, it proposes concrete rules, abstracts concrete rules into rules with variables, and prunes the resulting rules. ", "main_review": "Pros: \n- The learning algorithm is interesting and the problem of automatically discovering prepositions from examples is important for ATP in formal or informal languages.\n- The paper clearly defines the terms used and explains the methods and experiments well.\n\nCons:\n- As the authors pointed out, the experiments are neither large-scale nor real-world. One result of this is existing methods already achieve good performance, and it's not clear that the proposed method results in better performance.\n- None of the components (theorem prover, rule abstraction, rule pruning) are novel individually.\n\nQuestions:\n1) Are TRUE FALSE and MAPS_TO the only special symbols? It would be helpful to state this. \n2) Are there any unprovable examples in SCAN?\n3) The number of rules and symbols learned by MetaInduce is hard to compare with the number of learned parameters in ProofWriter. Is there a better metric to compare the two methods on? \n4) What are the advantages to using the proposed symbol system instead of first order logic? ", "summary_of_the_review": "The paper proposes an interesting solution to an important problem, but as-is the experimental settings and results are not compelling. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635908903510}, {"id": "JoRDRqUmoAq", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper81/Reviewer_9uxa"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The work proposes two new concepts:\n+ MetaQNL: a symbolic system in Quasi-Natural Language. Instead of representing rules in a formal symbolic format such as first order logic rules, in MetaQNL, a rule is represented in a Quasi-Natural Language format which includes words, variables and control symbols. Since the rules are represented in an informal representation. An interesting property of the MetaQNL representation is that it allows to perform backward or forward inference by substitution of variables with sentences. The authors assume that texts can be translated into the MetaQNL format and thus solving the reasoning problems with text input is possible via mining rules from text and backward/forward reasoning with the Quasi-Natural Language.\n+ To mine rule from text, the authors proposes an algorithm called MetaInduce. MetaInduce iterates through the training data several time to build a compact set of rules that trades complexity to prediction accuracy. It is a bottom up rule induction approach where it includes a Rule proposer which propose a concrete rule from a training example and an anti-unification module to abstract the concrete rule with more generalised rules. A pruning process based on MAX-SAT is used to prune the set of rules such that it optimise the regularised objective.", "review_text": "Overall the idea is novel and I like the paper presentation. However, I have some concerns regarding its ability to work with real natural language and the reproducibility of the proposed approach.\n\nMAJOR CONCERN 1:\nI have a concern that the rule proposer and the anti-unification require very well-formed of the sentences to be working well. \n\nAs we can see with the example on page 7 figure (a) about anti-unification, the sample shows very simple rules. I wonder how it will work when the sentence is really complicated with real natural language? Also the experiments in RuleTaker only with synthetically generated sentences in a very controlled language, could you please  demonstrate your results with complicated sentences beyond synthetic texts, I think the paraphrased datasets within the RuleTakers even not really natural language but that could be a good exercise for your methods to test on.\n\n\nMAJOR CONCERN 2:\nI worry about reproducibility as the source code is not open but the details explanation of the proposed approaches are missing. For example, from the paper I don't know how the rule proposer work and how the anti-unification is implemented with quasi-natural language. I would suggest to provide very detailed about the methods, with the current information I doubt that people can reimplement your work and reproduce what you have demonstrated.\n\nMAJOR CONERN 3:\nIn the RuleTaker example, it is known that Transformers are good at generalisation when they are trained on queries with depth 3 or greater. Yet transformers are not good at generalisation when it is trained on lower depth queries. Could you please provide comparison results with Transformers when it is trained at lower depths? \nOther minor comments:\n\n\n\"In contrast, our approach does not require a semantic parser, because rules in MetaQNL are directly applicable to natural language.\" ---> This is a strong statement, it requires a support with real natural language examples rather than synthetically generated sentences in the experiments.\n\n\nDefinition 6: what happen if there is no proof for a goal and the goal is proved via the close world assumption?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The work proposes two new concepts:\n+ MetaQNL: a symbolic system in Quasi-Natural Language. Instead of representing rules in a formal symbolic format such as first order logic rules, in MetaQNL, a rule is represented in a Quasi-Natural Language format which includes words, variables and control symbols. Since the rules are represented in an informal representation. An interesting property of the MetaQNL representation is that it allows to perform backward or forward inference by substitution of variables with sentences. The authors assume that texts can be translated into the MetaQNL format and thus solving the reasoning problems with text input is possible via mining rules from text and backward/forward reasoning with the Quasi-Natural Language.\n+ To mine rule from text, the authors proposes an algorithm called MetaInduce. MetaInduce iterates through the training data several time to build a compact set of rules that trades complexity to prediction accuracy. It is a bottom up rule induction approach where it includes a Rule proposer which propose a concrete rule from a training example and an anti-unification module to abstract the concrete rule with more generalised rules. A pruning process based on MAX-SAT is used to prune the set of rules such that it optimise the regularised objective.", "main_review": "Overall the idea is novel and I like the paper presentation. However, I have some concerns regarding its ability to work with real natural language and the reproducibility of the proposed approach.\n\nMAJOR CONCERN 1:\nI have a concern that the rule proposer and the anti-unification require very well-formed of the sentences to be working well. \n\nAs we can see with the example on page 7 figure (a) about anti-unification, the sample shows very simple rules. I wonder how it will work when the sentence is really complicated with real natural language? Also the experiments in RuleTaker only with synthetically generated sentences in a very controlled language, could you please  demonstrate your results with complicated sentences beyond synthetic texts, I think the paraphrased datasets within the RuleTakers even not really natural language but that could be a good exercise for your methods to test on.\n\n\nMAJOR CONCERN 2:\nI worry about reproducibility as the source code is not open but the details explanation of the proposed approaches are missing. For example, from the paper I don't know how the rule proposer work and how the anti-unification is implemented with quasi-natural language. I would suggest to provide very detailed about the methods, with the current information I doubt that people can reimplement your work and reproduce what you have demonstrated.\n\nMAJOR CONERN 3:\nIn the RuleTaker example, it is known that Transformers are good at generalisation when they are trained on queries with depth 3 or greater. Yet transformers are not good at generalisation when it is trained on lower depth queries. Could you please provide comparison results with Transformers when it is trained at lower depths? \nOther minor comments:\n\n\n\"In contrast, our approach does not require a semantic parser, because rules in MetaQNL are directly applicable to natural language.\" ---> This is a strong statement, it requires a support with real natural language examples rather than synthetically generated sentences in the experiments.\n\n\nDefinition 6: what happen if there is no proof for a goal and the goal is proved via the close world assumption?\n\n", "summary_of_the_review": "The paper proposes a new concept called quasi-language which allows representing rules in an new informal format that still allows to perform forward or backward reasoning while it is assumed to be mined easily from texts. The experimental results with some datasets with synthetically generated texts show that the methods work very well and advance state-of-the-art results. \nHowever, I doubt the application of the work with natural language input, I explicitly request to perform more experiments with paraphrased datasets in RuleTaker (MAJOR CONCERN 1). I also have a concern about reproducibility as the presentation lacks details of the core components of the proposed algorithm (MAJOR CONCERN 2). I also requested an additional experiment regarding the training data with low depth queries to check the ability of generalization of the proposed approaches. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635584874566}, {"id": "7s9vVGYza3d", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper81/Reviewer_7F3Z"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "Traditional research on symbolic reasoning assumes input data are already translated into a format that complies with the formalism of the underlying system. A major struggle for symbolic reasoning is to handle the data coming in a natural form (e.g. images/text) and perform reasoning on the same. Inspired by this challenge, this paper undertakes the problem of converting text inputs into the format of a system formalism. \n\nThe formalism is also proposed by this paper and is called MetaQNL. The framework of MetaQNL is designed keeping in mind the specific need of operating directly on natural language sentences. This formalism supports the natural language sentences with variables within them. This trick alleviates the need of using a semantic parser to parse natural language sentences for the purpose of logical reasoning. Thus, MetaQNL, by design, is conducive to working with natural language inputs. \n\nNext, this paper proposes an algorithm to induce rules from natural language inputs within the MetaQNL framework. This paper doesn’t worry about performing actual deductive reasoning/theorem proving and instead proposes to use existing provers and instead focus on the more challenging problem of rule induction. MetaInduce algorithm draws inspiration from existing ILP approaches. MetaInduce encodes the rule induction problem as a maximum satisfiability (MAX-SAT) problem, which can be solved efficiently by off-the-shelf solvers. The proposed method consists of 3 steps. \n\n1. Given a training example, a rule proposer proposes a set of concrete rules as candidates. This set can be overcomplete/inaccurate. \n2. It generates abstract rules from concrete rules via a symbolic procedure called anti-unification. This is essentially a process of aligning common substring segments across two or more strings.\n3. It encodes the proof paths in MAX-SAT and solves for a subset of all rules using a MAX-SAT solver. \n\nThis paper benchmarks the proposed method on 2 tasks - learning compositional instructions and logical reasoning. For learning compositional instructions, it works on two standard benchmarks: MiniSCAN and SCAN and recovers precisely the ground truth rules. For logical reasoning, the proposed method achieves SOTA on the RuleTaker dataset. \n\n", "review_text": "Overall, I liked the scope of this paper, the importance, and non-triviality of the problem, and the novelty of the proposed approach. In my view, it certainly adds a dimension to the literature on symbolic rule learning. The idea of MetaQNL is simple yet effective to handle the natural language inputs within a formal symbolic system. The idea behind MetaInduce is also quite natural. Although there are some weaknesses of the proposed approach, I still feel this is a novel idea and has the potential to yield something big in the future. \n\n__Strength__\n\n- A well-written paper.\n- A very nice literature survey and positioning of the work relative to the prior art.\n- An important problem in the broad space of AI.\n\n__Weakness__\n\n- Time complexity of each of the three steps in MetaInduce is not discussed. It will be good to shed some light on this. \n- As stated in the limitation section, the proposed approach is far from mature but serves as proof of concept.  It does not scale to millions of training examples.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Traditional research on symbolic reasoning assumes input data are already translated into a format that complies with the formalism of the underlying system. A major struggle for symbolic reasoning is to handle the data coming in a natural form (e.g. images/text) and perform reasoning on the same. Inspired by this challenge, this paper undertakes the problem of converting text inputs into the format of a system formalism. \n\nThe formalism is also proposed by this paper and is called MetaQNL. The framework of MetaQNL is designed keeping in mind the specific need of operating directly on natural language sentences. This formalism supports the natural language sentences with variables within them. This trick alleviates the need of using a semantic parser to parse natural language sentences for the purpose of logical reasoning. Thus, MetaQNL, by design, is conducive to working with natural language inputs. \n\nNext, this paper proposes an algorithm to induce rules from natural language inputs within the MetaQNL framework. This paper doesn’t worry about performing actual deductive reasoning/theorem proving and instead proposes to use existing provers and instead focus on the more challenging problem of rule induction. MetaInduce algorithm draws inspiration from existing ILP approaches. MetaInduce encodes the rule induction problem as a maximum satisfiability (MAX-SAT) problem, which can be solved efficiently by off-the-shelf solvers. The proposed method consists of 3 steps. \n\n1. Given a training example, a rule proposer proposes a set of concrete rules as candidates. This set can be overcomplete/inaccurate. \n2. It generates abstract rules from concrete rules via a symbolic procedure called anti-unification. This is essentially a process of aligning common substring segments across two or more strings.\n3. It encodes the proof paths in MAX-SAT and solves for a subset of all rules using a MAX-SAT solver. \n\nThis paper benchmarks the proposed method on 2 tasks - learning compositional instructions and logical reasoning. For learning compositional instructions, it works on two standard benchmarks: MiniSCAN and SCAN and recovers precisely the ground truth rules. For logical reasoning, the proposed method achieves SOTA on the RuleTaker dataset. \n\n", "main_review": "Overall, I liked the scope of this paper, the importance, and non-triviality of the problem, and the novelty of the proposed approach. In my view, it certainly adds a dimension to the literature on symbolic rule learning. The idea of MetaQNL is simple yet effective to handle the natural language inputs within a formal symbolic system. The idea behind MetaInduce is also quite natural. Although there are some weaknesses of the proposed approach, I still feel this is a novel idea and has the potential to yield something big in the future. \n\n__Strength__\n\n- A well-written paper.\n- A very nice literature survey and positioning of the work relative to the prior art.\n- An important problem in the broad space of AI.\n\n__Weakness__\n\n- Time complexity of each of the three steps in MetaInduce is not discussed. It will be good to shed some light on this. \n- As stated in the limitation section, the proposed approach is far from mature but serves as proof of concept.  It does not scale to millions of training examples.", "summary_of_the_review": "See my comments in the main review.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635171670006}], "openreview_url": "https://openreview.net/forum?id=7zFokR7k_86", "arxiv_id": "2111.12038", "paper_pdf": "papers/7zFokR7k_86.pdf", "paper_pdf_sha256": "500460cedae7fa9db67a8506111a50c3fef5b9970a1f2e13fcd82c4929aaeab8", "paper_pdf_bytes": 345948, "paper_pdf_source": "openreview", "code_url": "https://github.com/princeton-vl/MetaQNL", "code_repository": "princeton-vl/MetaQNL", "code_commit": "dcff310641c8f999eb5354584f32660b98fd9ec9", "code_archive": "repos/7zFokR7k_86.zip", "code_archive_sha256": "7385df639bc5ccda91786cedbfe1e7f5f57cfb678ef1e68dbfd5a95ddac6ea75", "code_archive_bytes": 247531, "code_file_count": 69, "code_extensions": {".jl": 69}, "github_disk_usage_kb": 241, "github_languages": {"Julia": 256078}, "github_archived": false, "github_pushed_at": "2024-08-07T00:04:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-symbolic-rules-for-reasoning-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oj3bHNSq_2w", "year": 2021, "status": "rejected", "title": "Sample weighting as an explanation for mode collapse in generative adversarial networks", "authors": ["Aksel Wilhelm Wold Eide", "Eilif Solberg", "Ingebjørg Kåsen"], "authorids": ["~Aksel_Wilhelm_Wold_Eide1", "eilif.solberg@ffi.no", "ingebjorg.kasen@ffi.no"], "authors_source": "OpenReview API", "abstract": "Generative adversarial networks were introduced with a logistic MiniMax cost formulation, which normally fails to train due to saturation, and a Non-Saturating reformulation. While addressing the saturation problem, NS-GAN also inverts the generator's sample weighting, implicitly shifting emphasis from higher-scoring to lower-scoring samples when updating parameters. We present both theory and empirical results suggesting that this makes NS-GAN prone to mode dropping. We design MM-nsat, which preserves MM-GAN sample weighting while avoiding saturation by rescaling the MM-GAN minibatch gradient such that its magnitude approximates NS-GAN's gradient magnitude. MM-nsat has qualitatively different training dynamics, and on MNIST and CIFAR-10 it is stronger in terms of mode coverage, stability and FID. While the empirical results for MM-nsat are promising and favorable also in comparison with the LS-GAN and Hinge-GAN formulations, our main contribution is to show how and why NS-GAN's sample weighting causes mode dropping and training collapse.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "J9xS09WqEpP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper186/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an explanation for mode collapse in the original GAN with the log -D objective for the generator (dubbed the non-saturating GAN or NS-GAN for short). The paper takes the approach of comparing the gradient of the generator objective for the original GAN with cross-entropy loss (dubbed the minimax GAN or MM-GAN for short) and the log -D variant. The key observation is that the difference between the gradient of the generator objective of MM-GAN and NS-GAN is that MM-GAN has a factor of D_p(G(z,\\theta)), whereas NS-GAN has a factor of 1-D_p(G(z,\\theta)), where D_p(G(z,\\theta)) is the output of the discriminator on a sample generated from z. Hence, the terms in the MM-GAN gradient that appear fake to the discriminator are downweighted, whereas they are upweighted in the NS-GAN gradient. Because the samples from modes that are already overrepresented are likely declared fake by the discriminator, the contribution to the generator gradient is dominated by these samples in NS-GAN. \n\nStrengths:\nThis observation is very nice, intuitive and simple and is new to my knowledge. It sheds light on the weaknesses of the log -D variant of GANs.  \n\nWeaknesses:\nPaper title is broader than what the paper shows - the proposes explanation only applies to GANs with the log -D generator objective and does not apply to other GAN variants, e.g.: original GAN (with cross-entropy generator objective), WGAN, LSGAN etc.\n\nThe argument is not presented clearly, and sometimes observations with no obvious logical relationship to the claim are mentioned. At other times, it is unclear what the point is. For example:\n\nOn pg. 3, in the first paragraph under sect. 2.3, it is mentioned that \"the minibatch used to update G will have more samples from O since they are generated more often\". It is unclear how the number of samples from O in the minibatch could cause a difference between MM-GAN and NS-GAN - this observation is true for both MM-GAN and NS-GAN!\n\nIn the next paragraph, the paper mentioned that the generator gradient \"is only locally informative\" - again this is true for all GANs. How is this relevant for the argument made in the paper?\n\nThen it mentioned that \"generated samples give rise to conflicting gradients\" - it is unclear what conflicting gradients mean. In what sense are the gradients \"conflicting\"?\n\nIn the next paragraph, the paper mentioned \"NS-GAN struggles to discover new modes: g(x) ≈ 0 ⇒ 1 − D_p(x) ≈ 0\" - the connection between the claim about the difficulty of discovering new modes and the equations should be explained more clearly. Also, the implication in the equations would not be true if r(x) ≈ 0. \n\nAlso, for eq. (7), technically the optimal discriminator is only uniquely defined at the real data points (see Sinn & Rawat, AISTATS 2018) because the GAN is trained on a finite sample. The optimal discriminator result in (Goodfellow et al., 2014) assumes access to the true data distribution, or equivalently an infinite stream of samples from the distribution. So technically the paper's claim on the discovery of new modes cannot be justified by \"g(x) ≈ 0 ⇒ 1 − D_p(x) ≈ 0\", because D_p(x) could take on any value at locations other than the data points. Similarly, on pg. 4, in the paragraph below eq. (9), because D_l^{opt} could take on arbitrary values at positions other than the data points, the claim that D_{opt} = ±\\infty is inaccurate - it is in fact not uniquely defined for points that are not real data points. Though because this is a common mistake in the literature, I'd be fine with the addition of a note that explains this caveat before presenting the theoretical argument (without insisting on a fundamental solution to this issue). \n\nSect. 2.4 on \"MM-GAN Interaction with ADAM\" is not very mathematically rigorous and relies primarily on an assumption that the value of the logits of the discriminator approaches the optimum linearly. It's unclear if this assumption is actually justified, since the gradient of the cross-entropy loss w.r.t the logits tapers off on the extreme ends of the logits, so the logits should approach the optimum more and more slowly as they become larger in magnitude. \n\nAlso, in eq. 11, the left size is vector graphics, whereas the right side is a scalar. \n\nUnder sect. 3.1, the JSD between class predictions assume each class contains only one mode and cannot detect intra-class mode collapse. This caveat should be prominently posted. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for Paper", "review": "This paper proposes an explanation for mode collapse in the original GAN with the log -D objective for the generator (dubbed the non-saturating GAN or NS-GAN for short). The paper takes the approach of comparing the gradient of the generator objective for the original GAN with cross-entropy loss (dubbed the minimax GAN or MM-GAN for short) and the log -D variant. The key observation is that the difference between the gradient of the generator objective of MM-GAN and NS-GAN is that MM-GAN has a factor of D_p(G(z,\\theta)), whereas NS-GAN has a factor of 1-D_p(G(z,\\theta)), where D_p(G(z,\\theta)) is the output of the discriminator on a sample generated from z. Hence, the terms in the MM-GAN gradient that appear fake to the discriminator are downweighted, whereas they are upweighted in the NS-GAN gradient. Because the samples from modes that are already overrepresented are likely declared fake by the discriminator, the contribution to the generator gradient is dominated by these samples in NS-GAN. \n\nStrengths:\nThis observation is very nice, intuitive and simple and is new to my knowledge. It sheds light on the weaknesses of the log -D variant of GANs.  \n\nWeaknesses:\nPaper title is broader than what the paper shows - the proposes explanation only applies to GANs with the log -D generator objective and does not apply to other GAN variants, e.g.: original GAN (with cross-entropy generator objective), WGAN, LSGAN etc.\n\nThe argument is not presented clearly, and sometimes observations with no obvious logical relationship to the claim are mentioned. At other times, it is unclear what the point is. For example:\n\nOn pg. 3, in the first paragraph under sect. 2.3, it is mentioned that \"the minibatch used to update G will have more samples from O since they are generated more often\". It is unclear how the number of samples from O in the minibatch could cause a difference between MM-GAN and NS-GAN - this observation is true for both MM-GAN and NS-GAN!\n\nIn the next paragraph, the paper mentioned that the generator gradient \"is only locally informative\" - again this is true for all GANs. How is this relevant for the argument made in the paper?\n\nThen it mentioned that \"generated samples give rise to conflicting gradients\" - it is unclear what conflicting gradients mean. In what sense are the gradients \"conflicting\"?\n\nIn the next paragraph, the paper mentioned \"NS-GAN struggles to discover new modes: g(x) ≈ 0 ⇒ 1 − D_p(x) ≈ 0\" - the connection between the claim about the difficulty of discovering new modes and the equations should be explained more clearly. Also, the implication in the equations would not be true if r(x) ≈ 0. \n\nAlso, for eq. (7), technically the optimal discriminator is only uniquely defined at the real data points (see Sinn & Rawat, AISTATS 2018) because the GAN is trained on a finite sample. The optimal discriminator result in (Goodfellow et al., 2014) assumes access to the true data distribution, or equivalently an infinite stream of samples from the distribution. So technically the paper's claim on the discovery of new modes cannot be justified by \"g(x) ≈ 0 ⇒ 1 − D_p(x) ≈ 0\", because D_p(x) could take on any value at locations other than the data points. Similarly, on pg. 4, in the paragraph below eq. (9), because D_l^{opt} could take on arbitrary values at positions other than the data points, the claim that D_{opt} = ±\\infty is inaccurate - it is in fact not uniquely defined for points that are not real data points. Though because this is a common mistake in the literature, I'd be fine with the addition of a note that explains this caveat before presenting the theoretical argument (without insisting on a fundamental solution to this issue). \n\nSect. 2.4 on \"MM-GAN Interaction with ADAM\" is not very mathematically rigorous and relies primarily on an assumption that the value of the logits of the discriminator approaches the optimum linearly. It's unclear if this assumption is actually justified, since the gradient of the cross-entropy loss w.r.t the logits tapers off on the extreme ends of the logits, so the logits should approach the optimum more and more slowly as they become larger in magnitude. \n\nAlso, in eq. 11, the left size is vector graphics, whereas the right side is a scalar. \n\nUnder sect. 3.1, the JSD between class predictions assume each class contains only one mode and cannot detect intra-class mode collapse. This caveat should be prominently posted. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1605015493850}, {"id": "kZS0cBafc4f", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper186/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper reexamines the original (MM) and the non-saturating (NS) GAN objective.  The authors show that the gradients of the respective objectives just differ from a scaling factor depending on the discriminator's output for generated samples. While the scaling factor for the MM gradient is responsible for the well known vanishing gradient if the discriminator is optimal, the scaling factor of the NS gradient counteracts this saturation effect. However, on the other side, the NS scaling factor introduces a mode dropping effect and the inability of the learning dynamci to discover new modes. The authors show additionally that the NS minibatch gradient is the weighted sum of the single sample MM gradients with respective scaling factors as weights. These scaling factors avoid saturating, however, alter the direction of the resulting minibatch gradient. To counteract the change of the gradient direction the authors propose to summarize the sample scaling factors into one scalar for the batch gradient which preserves the non-saturating behaviour of the NS and the gradient direction of the MM objective. The new GAN objective is called MM-nsat.  Additionally the authors discuss the non-saturating effect of the ADAM beta2 parameter for the MM-GAN Generator. \n\nExperiments: A ring of Gaussians experiment shows the mode dropping effect of the NS-GAN. Training a 4-layer network on MNIST shows demonstrates the vanishing gradient effect on MM-GAN and the counteracting effect with a smaller ADAM beta2 parameter applied to the Generator. Also on MNIST MM-nsat was compared with NS showing a lower FID and Jensen-Shannon divergence between real and generated class distributions for MM-nsat. On the Cat 128x128 dataset mode collapse was visually shown for NS. \nOn MNIST, CIFAR10 and Cat 128x128 MM-nsat was compared with NS, Hinge and LS-GAN outperforming all of them based on the FID.\n\nPros: The theoretical insight why NS-GAN suffers from mode collapse is novel and interesting. The experiments are convincing and extensive.\n\nCons: The proposed objective is not novel [1][2], however, it is derived directly from the original MM-GAN objective and is better theoretically motivated.  \n\nFor completeness, it would be interesting to see the experiments in section K with the pair MM-nsat/MM as well. \n\n[1] R Devon Hjelm, Athul Paul Jacob, Tong Che, Adam Trischler, Kyunghyun Cho, and Yoshua Bengio. Boundaryseeking generative adversarial networks, 2018\n[2] Zhiting Hu, Zichao Yang, Ruslan Salakhutdinov, and Eric P. Xing. On unifying deep generative models. CoRR,\nabs/1706.00550, 2017. URL http://arxiv.org/abs/1706.00550\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "IW-GAN revisited", "review": "This paper reexamines the original (MM) and the non-saturating (NS) GAN objective.  The authors show that the gradients of the respective objectives just differ from a scaling factor depending on the discriminator's output for generated samples. While the scaling factor for the MM gradient is responsible for the well known vanishing gradient if the discriminator is optimal, the scaling factor of the NS gradient counteracts this saturation effect. However, on the other side, the NS scaling factor introduces a mode dropping effect and the inability of the learning dynamci to discover new modes. The authors show additionally that the NS minibatch gradient is the weighted sum of the single sample MM gradients with respective scaling factors as weights. These scaling factors avoid saturating, however, alter the direction of the resulting minibatch gradient. To counteract the change of the gradient direction the authors propose to summarize the sample scaling factors into one scalar for the batch gradient which preserves the non-saturating behaviour of the NS and the gradient direction of the MM objective. The new GAN objective is called MM-nsat.  Additionally the authors discuss the non-saturating effect of the ADAM beta2 parameter for the MM-GAN Generator. \n\nExperiments: A ring of Gaussians experiment shows the mode dropping effect of the NS-GAN. Training a 4-layer network on MNIST shows demonstrates the vanishing gradient effect on MM-GAN and the counteracting effect with a smaller ADAM beta2 parameter applied to the Generator. Also on MNIST MM-nsat was compared with NS showing a lower FID and Jensen-Shannon divergence between real and generated class distributions for MM-nsat. On the Cat 128x128 dataset mode collapse was visually shown for NS. \nOn MNIST, CIFAR10 and Cat 128x128 MM-nsat was compared with NS, Hinge and LS-GAN outperforming all of them based on the FID.\n\nPros: The theoretical insight why NS-GAN suffers from mode collapse is novel and interesting. The experiments are convincing and extensive.\n\nCons: The proposed objective is not novel [1][2], however, it is derived directly from the original MM-GAN objective and is better theoretically motivated.  \n\nFor completeness, it would be interesting to see the experiments in section K with the pair MM-nsat/MM as well. \n\n[1] R Devon Hjelm, Athul Paul Jacob, Tong Che, Adam Trischler, Kyunghyun Cho, and Yoshua Bengio. Boundaryseeking generative adversarial networks, 2018\n[2] Zhiting Hu, Zichao Yang, Ruslan Salakhutdinov, and Eric P. Xing. On unifying deep generative models. CoRR,\nabs/1706.00550, 2017. URL http://arxiv.org/abs/1706.00550\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604241383792}, {"id": "RoKA3yx7pPf", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper186/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present a simple estimator which approximates the gradient of the \"true\" generator loss in GANs using the gradients from the more commonly used non-saturating generator loss. They present results on toy data and small image datasets like CIFAR to show that the method leads to stabler training and does not suffer from mode collapse as badly.\n\nHad this paper appeared at ICLR in, say, 2016 or 2017, I think it would have been a welcome addition to the literature on \"tricks to improve GAN training\", which was fairly small at that point. Since then, the field has absolutely exploded, to a degree that I find somewhat baffling, as GANs have no practical applications beyond generating pretty pictures. In addition, recent advances to VAE training (building on some of these GAN tricks like spectral normalization) have massively improved the quality of VAE samples, so the need for GANs continues to diminish. Nevertheless, it is not my job to judge whether a given paper is trendy or not - only to judge whether it is technically correct and up to the standards of publication. On that front, I think it is a decent paper - the method is quite simple, but the results on both mixtures of Gaussians and CIFAR and other image datasets do seem like an improvement. I would have appreciated experiments on a larger dataset like CIFAR or high-res CelebA. I also would have appreciated a more rigorous comparison against the plethora of GAN training tricks that have appeared in the last 5 years - for instance, competitive gradient descent (Schaefer and Anandkumar, 2019) to name just one. As it is, I worry this paper will disappear in a flood of other similar papers without more rigorous comparisons - but then, that may be for the field to judge after it is published.\n\nA few stylistic points:\n* Please increase the size of the axis labels in Fig 1. They are unreadable unless zoomed in very far on a screen.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Bayesians hate this one weird trick for improving GAN training!", "review": "The authors present a simple estimator which approximates the gradient of the \"true\" generator loss in GANs using the gradients from the more commonly used non-saturating generator loss. They present results on toy data and small image datasets like CIFAR to show that the method leads to stabler training and does not suffer from mode collapse as badly.\n\nHad this paper appeared at ICLR in, say, 2016 or 2017, I think it would have been a welcome addition to the literature on \"tricks to improve GAN training\", which was fairly small at that point. Since then, the field has absolutely exploded, to a degree that I find somewhat baffling, as GANs have no practical applications beyond generating pretty pictures. In addition, recent advances to VAE training (building on some of these GAN tricks like spectral normalization) have massively improved the quality of VAE samples, so the need for GANs continues to diminish. Nevertheless, it is not my job to judge whether a given paper is trendy or not - only to judge whether it is technically correct and up to the standards of publication. On that front, I think it is a decent paper - the method is quite simple, but the results on both mixtures of Gaussians and CIFAR and other image datasets do seem like an improvement. I would have appreciated experiments on a larger dataset like CIFAR or high-res CelebA. I also would have appreciated a more rigorous comparison against the plethora of GAN training tricks that have appeared in the last 5 years - for instance, competitive gradient descent (Schaefer and Anandkumar, 2019) to name just one. As it is, I worry this paper will disappear in a flood of other similar papers without more rigorous comparisons - but then, that may be for the field to judge after it is published.\n\nA few stylistic points:\n* Please increase the size of the axis labels in Fig 1. They are unreadable unless zoomed in very far on a screen.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603994305823}, {"id": "sZDukTplxQn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper186/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis work proposes that many common issues with GAN methods are based on the weighting of the samples given to the generator’s objective function. They focus on a study of the original GAN objective proposed in Goodfellow et al. where the generator’s objective is the negative of the discriminators objective. The GAN community quickly observed that when the discriminator outperforms the generator with this objective, the saturating nature of the sigmoid function causes the gradients to vanish for the generator’s objective. For this reason, a new objective (NS-GAN) was proposed which modifies the generator’s objective to alleviate this gradient vanishing issue.  The authors argue that this modified objective is to blame for a number of common issues with GAN methods -- most notably the mode-dropping issue. The authors present theory that backs up these claims and they propose a new generator training objective which re-weights the gradients of the generator objective to have the same average magnitude as NS-GAN but have the same relative magnitudes of the original GAN objective. \n\nThe authors demonstrate the impact of their new loss function in a series of quantitative and qualitative settings. \n\n\nStrong areas:\n\nI am a very big fan of work that questions standard assumptions that are taken almost as fact within our community. When GANs were originally proposed, most researchers saw the NS-GAN objective as a strict improvement over the MM-GAN objective and moved on. This work does a great job to demonstrate that NS-GAN is most definitely not “superior” to MM-GAN and may possibly be a worse choice of objective function. This change may seem insignificant on the surface and the authors here clearly show that it has an impact -- one large enough to consider if it should be used at all. To aid in their claims, the authors provide clear and concise explanations backed up by easy to understand theory. Particularly interesting to me was the argument that the popular Adam optimizer does not alleviate the saturation issues found with MM-GAN even though my own intuition on the optimizer tells us that it should help with issues like this. \n\nWeaknesses:\n\nThe empirical results presented appear promising. The proposed approaches (MM-Unit and MM-NSAT) considerably outperform the NS-GAN objectives. My biggest issues are with the quality of the baselines. I am not an expert in the GAN field, but from inspecting the Conv-4 model with spectral normalization on CIFAR10, it appears like the best performing model achieves an FID of approx 42 and the worst model gets around 48 (figure 8, bottom right). While this difference is notable and the error-bars indicate it is statistically significant, I am concerned because these numbers seem to considerably underperform prior work. The original paper on spectral normalization for GANs reports an FID of 29.3 for standard CNNs. This model uses the NS-GAN objective (to my knowledge) so, I am confused as to why the authors did not simply replicate their setup -- especially since the standard CNN model proposed in the original spectral norm paper can be trained with reasonable compute and the authors have released code. If I am misunderstanding something about the experiments, please do let me know, but I am confused by this choice.\n\nI also took some issues with the D-JS-CD score proposed to score the class distribution of generated samples. There have been a number of proposed metrics like this such as IS and classifier score (https://arxiv.org/abs/1905.10887). I understand that this score (unlike classifier score) does not rely on a conditional model, but if a new score is to be proposed and it is used to convince the reader that a new method performs well, then it should be applied to some baseline models. Pre-trained models from prior GAN papers could be used to obtain these scores. \n\nAs a researcher from a different field, my main concern about the experimental results is that the results are quite far from the current state-of-the-field. State of the art results are by no means required for publication in this venue, but the baseline models presented here perform much worse than they have been shown to in prior work. Since a near 20-point increase in FID can be achieved with a few tweaks to the NS-GAN objective (SN-GAN), I am left wondering if the presented improvement from the MM-NSAT objective will vanish once those improvements are applied or if it will still hold. Since this is not shown, then I am uncertain of the significance of the observations made in this work. \n\nSome more nit-picky issues:\n\nThe text in the figures is too small and near impossible to read. I would present all of the results in Figure 8 in a table instead. There is no information gained by seeing loss curves. A table would save much more space and allow the readers to more easily compare this work to other works. In Figure 7, I find “best” and “worst” picks by qualitative methods to be somewhat unconvincing. You should show the highest and lowest FID or just show random samples.\n\nMy recommendation:\n\nI am not an active member of the GAN community so I am more than willing to accept if my recommendation goes against more senior folks who work in that field. \n\nI found this to be an interesting work that provided a non-trivial insight, backed it up with clear and easy-to-follow theory and demonstrated their observations held on some medium-scale experiments. My biggest issues come from my reservations about the experimental results. The baseline NS-GAN with spectral normalization presented in this work greatly underperforms previously published methods that use the same objective. The difference between the presented baseline and the proposed method is smaller than the difference in performance between the presented baseline and previously published methods with the baseline objective. \n\nThese discrepancies give me sufficient doubt where I am not certain that the insights of this work provide a sufficient improvement when combined with architectural improvements and proper parameter tuning. \n\nFor these reasons, I am advocating against acceptance of this work but making it clear that I think this paper is borderline.\n\nIf the experimental setup was more in line with prior work and the same trend in results held, then I would be more likely to recommend acceptance of this paper. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting approach, questionable experiments", "review": "Summary:\n\nThis work proposes that many common issues with GAN methods are based on the weighting of the samples given to the generator’s objective function. They focus on a study of the original GAN objective proposed in Goodfellow et al. where the generator’s objective is the negative of the discriminators objective. The GAN community quickly observed that when the discriminator outperforms the generator with this objective, the saturating nature of the sigmoid function causes the gradients to vanish for the generator’s objective. For this reason, a new objective (NS-GAN) was proposed which modifies the generator’s objective to alleviate this gradient vanishing issue.  The authors argue that this modified objective is to blame for a number of common issues with GAN methods -- most notably the mode-dropping issue. The authors present theory that backs up these claims and they propose a new generator training objective which re-weights the gradients of the generator objective to have the same average magnitude as NS-GAN but have the same relative magnitudes of the original GAN objective. \n\nThe authors demonstrate the impact of their new loss function in a series of quantitative and qualitative settings. \n\n\nStrong areas:\n\nI am a very big fan of work that questions standard assumptions that are taken almost as fact within our community. When GANs were originally proposed, most researchers saw the NS-GAN objective as a strict improvement over the MM-GAN objective and moved on. This work does a great job to demonstrate that NS-GAN is most definitely not “superior” to MM-GAN and may possibly be a worse choice of objective function. This change may seem insignificant on the surface and the authors here clearly show that it has an impact -- one large enough to consider if it should be used at all. To aid in their claims, the authors provide clear and concise explanations backed up by easy to understand theory. Particularly interesting to me was the argument that the popular Adam optimizer does not alleviate the saturation issues found with MM-GAN even though my own intuition on the optimizer tells us that it should help with issues like this. \n\nWeaknesses:\n\nThe empirical results presented appear promising. The proposed approaches (MM-Unit and MM-NSAT) considerably outperform the NS-GAN objectives. My biggest issues are with the quality of the baselines. I am not an expert in the GAN field, but from inspecting the Conv-4 model with spectral normalization on CIFAR10, it appears like the best performing model achieves an FID of approx 42 and the worst model gets around 48 (figure 8, bottom right). While this difference is notable and the error-bars indicate it is statistically significant, I am concerned because these numbers seem to considerably underperform prior work. The original paper on spectral normalization for GANs reports an FID of 29.3 for standard CNNs. This model uses the NS-GAN objective (to my knowledge) so, I am confused as to why the authors did not simply replicate their setup -- especially since the standard CNN model proposed in the original spectral norm paper can be trained with reasonable compute and the authors have released code. If I am misunderstanding something about the experiments, please do let me know, but I am confused by this choice.\n\nI also took some issues with the D-JS-CD score proposed to score the class distribution of generated samples. There have been a number of proposed metrics like this such as IS and classifier score (https://arxiv.org/abs/1905.10887). I understand that this score (unlike classifier score) does not rely on a conditional model, but if a new score is to be proposed and it is used to convince the reader that a new method performs well, then it should be applied to some baseline models. Pre-trained models from prior GAN papers could be used to obtain these scores. \n\nAs a researcher from a different field, my main concern about the experimental results is that the results are quite far from the current state-of-the-field. State of the art results are by no means required for publication in this venue, but the baseline models presented here perform much worse than they have been shown to in prior work. Since a near 20-point increase in FID can be achieved with a few tweaks to the NS-GAN objective (SN-GAN), I am left wondering if the presented improvement from the MM-NSAT objective will vanish once those improvements are applied or if it will still hold. Since this is not shown, then I am uncertain of the significance of the observations made in this work. \n\nSome more nit-picky issues:\n\nThe text in the figures is too small and near impossible to read. I would present all of the results in Figure 8 in a table instead. There is no information gained by seeing loss curves. A table would save much more space and allow the readers to more easily compare this work to other works. In Figure 7, I find “best” and “worst” picks by qualitative methods to be somewhat unconvincing. You should show the highest and lowest FID or just show random samples.\n\nMy recommendation:\n\nI am not an active member of the GAN community so I am more than willing to accept if my recommendation goes against more senior folks who work in that field. \n\nI found this to be an interesting work that provided a non-trivial insight, backed it up with clear and easy-to-follow theory and demonstrated their observations held on some medium-scale experiments. My biggest issues come from my reservations about the experimental results. The baseline NS-GAN with spectral normalization presented in this work greatly underperforms previously published methods that use the same objective. The difference between the presented baseline and the proposed method is smaller than the difference in performance between the presented baseline and previously published methods with the baseline objective. \n\nThese discrepancies give me sufficient doubt where I am not certain that the insights of this work provide a sufficient improvement when combined with architectural improvements and proper parameter tuning. \n\nFor these reasons, I am advocating against acceptance of this work but making it clear that I think this paper is borderline.\n\nIf the experimental setup was more in line with prior work and the same trend in results held, then I would be more likely to recommend acceptance of this paper. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603251071830}], "openreview_url": "https://openreview.net/forum?id=oj3bHNSq_2w", "arxiv_id": "2010.02035", "paper_pdf": "papers/oj3bHNSq_2w.pdf", "paper_pdf_sha256": "4c0d8b9973be3b23e1593f46aaf8642a8d678c4208ca12198915b8f746e48124", "paper_pdf_bytes": 20722498, "paper_pdf_source": "openreview", "code_url": "https://github.com/awweide/mm-nsat", "code_repository": "awweide/mm-nsat", "code_commit": "799b054e059f32eb0b3a3831f20c9648a2e0c282", "code_archive": "repos/oj3bHNSq_2w.zip", "code_archive_sha256": "15a2002e7bc2a5a4d27eaa7b2d048369e79ec2d15da07ea6ff587c8893fded0f", "code_archive_bytes": 272955, "code_file_count": 13, "code_extensions": {".py": 12, ".sh": 1}, "github_disk_usage_kb": 277, "github_languages": {"Python": 86592, "Shell": 7138}, "github_archived": false, "github_pushed_at": "2021-01-01T04:07:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sample-weighting-as-an-explanation-for-mode-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJx4PAEYDH", "year": 2020, "status": "rejected", "title": "R-TRANSFORMER: RECURRENT NEURAL NETWORK ENHANCED TRANSFORMER", "authors": ["Zhiwei Wang", "Yao Ma", "Zitao Liu", "Jiliang Tang"], "authorids": ["wangzh65@msu.edu", "mayao4@msu.edu", "liuzitao@100tal.com", "tangjili@msu.edu"], "authors_source": "OpenReview API", "abstract": "Recurrent Neural Networks have long been the dominating choice for sequence modeling. However, it severely suffers from two issues: impotent in capturing very long-term dependencies and unable to parallelize the sequential computation procedure. Therefore, many non-recurrent sequence models that are built on convolution and attention operations have been proposed recently. Notably, models with multi-head attention such as Transformer have demonstrated extreme effectiveness in capturing long-term dependencies in a variety of sequence modeling tasks. Despite their success, however, these models lack necessary components to model local structures in sequences and heavily rely on position embeddings that have limited effects and require a considerable amount of design efforts. In this paper, we propose the R-Transformer which enjoys the advantages of both RNNs and the multi-head attention mechanism while avoids their respective drawbacks. The proposed model can effectively capture both local structures and global long-term dependencies in sequences without any use of position embeddings. We evaluate R-Transformer through extensive experiments with data from a wide range of domains and the empirical results show that R-Transformer outperforms the state-of-the-art methods by a large margin in most of the tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BkliykIptB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1171/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a novel transformer model called R-Transformer. \nBased on the observation that positional embeddings require a \na lot of design efforts in vanilla Transformer, the authors propose to use local RNNs to\nencode local information in replace of positional embeddings.\n\nThe paper is well presented and the proposed algorithm is explained in detail. \nHowever, it is not clear how the proposed model obtains positional information of input nodes. \nIn vanilla Transformer, positional embeddings contains global positional information. \nIn the R-Transformer, however, the multi-head attention layer will not be able to obtain positional information from local RNN outputs. Will such loss in positional information affect the performance?\n\nThe empirical study shows that R-Transformer can outperform vanilla Transformers and recurrent architectures, which is promising. Still, it would be more convincing if the authors could provide comparisons on NMT tasks or larger language modeling datasets such as WikText. Also, I am interested in the training efficiency of these models. How much overhead does the local RNN introduce?\n\nOverall I think this is an interesting paper but experiments could be improved. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "In this paper, the authors propose a novel transformer model called R-Transformer. \nBased on the observation that positional embeddings require a \na lot of design efforts in vanilla Transformer, the authors propose to use local RNNs to\nencode local information in replace of positional embeddings.\n\nThe paper is well presented and the proposed algorithm is explained in detail. \nHowever, it is not clear how the proposed model obtains positional information of input nodes. \nIn vanilla Transformer, positional embeddings contains global positional information. \nIn the R-Transformer, however, the multi-head attention layer will not be able to obtain positional information from local RNN outputs. Will such loss in positional information affect the performance?\n\nThe empirical study shows that R-Transformer can outperform vanilla Transformers and recurrent architectures, which is promising. Still, it would be more convincing if the authors could provide comparisons on NMT tasks or larger language modeling datasets such as WikText. Also, I am interested in the training efficiency of these models. How much overhead does the local RNN introduce?\n\nOverall I think this is an interesting paper but experiments could be improved. "}, "tcdate": 1571802851075}, {"id": "H1lr66X6KB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1171/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces the R-Transformer architecture which adds a local RNN layer before each attention layer in Transformer. The authors claim state-of-the-art performance but only test on tiny tasks where Transformer models have not been heavily optimized and omit the main problem with RNNs - namely their speed. It is an interesting paper still and the locality is a nice way to remedy the speed problem, but the paper lacks a true study and ablations on this main limitation. In summary: the main new idea of the paper is to make RNNs local in Transformer (trying to add RNN layers has been explored before). This idea could be a good tradeoff between full RNN (slow) and no RNN (lack of context), but the following is missing: (1) ablations on speed vs results by locality window, (2) experiments on more widely reported and larger data-sets and models, at least including some language modeling task (wiki or lm1b) and some translation task (like en-de). Without these results, we cannot recommend to accept this paper.\n\nI'm grateful to the authors for their reply. Presently, a very good LM1B or WMT model can be trained for free in Google colab in under a day, so I do not believe it's computationally infeasable to run the experiments I asked for. Even if it took much longer, I'd believe that the time should be invested before acceptance, so I stand by my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "N/A", "title": "Official Blind Review #1", "review": "The paper introduces the R-Transformer architecture which adds a local RNN layer before each attention layer in Transformer. The authors claim state-of-the-art performance but only test on tiny tasks where Transformer models have not been heavily optimized and omit the main problem with RNNs - namely their speed. It is an interesting paper still and the locality is a nice way to remedy the speed problem, but the paper lacks a true study and ablations on this main limitation. In summary: the main new idea of the paper is to make RNNs local in Transformer (trying to add RNN layers has been explored before). This idea could be a good tradeoff between full RNN (slow) and no RNN (lack of context), but the following is missing: (1) ablations on speed vs results by locality window, (2) experiments on more widely reported and larger data-sets and models, at least including some language modeling task (wiki or lm1b) and some translation task (like en-de). Without these results, we cannot recommend to accept this paper.\n\nI'm grateful to the authors for their reply. Presently, a very good LM1B or WMT model can be trained for free in Google colab in under a day, so I do not believe it's computationally infeasable to run the experiments I asked for. Even if it took much longer, I'd believe that the time should be invested before acceptance, so I stand by my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571794365200}, {"id": "r1gU10tB_H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1171/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper proposes a new architecture, R-Transformer, that blends the Transformer networks and the recurrent networks, so as to better capture both the long- and short-term features. By injecting a local RNN layer at every level of the network, the authors hoped to enhance the Transformer's ability to model locality structure. To demonstrate the modeling power of R-Trasnformer, the paper evaluates the effectiveness of R-Transformer on 4 different sequence tasks (seqMNIST, polyphonic music, character- and word-level PTB).\n\nContribution: The authors propose an architecture that combines the practices of recurrent and feed-forward sequence models. However, I have major concerns regarding the novelty this paper, the various claims it makes, as well as its experiment setting.\n\n----------------------------------------------------------------------------\n\nMajor issues/questions:\n\n1. The techniques proposed by this paper lack novelty. For instance, the entire section 3.2 is simply the original design of the multi-head self-attention by Vaswani et al. The major difference between R-Transformer and the original Transformer is the replacement of positional embedding with an RNN layer, but (in my opinion) the authors did not demonstrate sufficiently its effectiveness via ablative studies (see below). Moreover, some prior works have already exploited the locality structure in Transformers. For instance, [1] showed that a sparse, local Transformer can work extremely well and be very efficient (they achieved SOTA on large-scale char-level language modeling tasks).\n\n2. The experiments do not entirely convince me. \n    i) The authors use the \"same hidden size for R-Transformer and Transformer.\" But in fact, as the R-Transformer has one extra RNN/LSTM/GRU layer at every level of the network, the tests were carried out (in effect) using a larger model than the baselines. I think the authors should instead control the # of model parameters, especially since you are running only on small tasks with small-sized models.\n    ii) It is nice that the authors tested R-Transformer on a variety of tasks--- this is important. However, in no way do these number achieve the levels of the \"state-of-the-art\", which the authors claim at the end of Section 1 (e.g., [2] has better number on seqMNIST and character-level PTB, and the Transformer-XL actually achieves <55 perplexity on word-level PTB). Therefore, the numbers don't look particularly appealing to me. \n    iii) Lack of more challenging, or large-scale experiments. Sequential MNIST is known to be a relatively simple task, and alternatives such as sequential CIFAR-10 or permuted sequential MNIST are more valuable \"small sequence tasks.\" (e.g., prior works studying long-term sequence model dependencies [3]) I also think that benchmarking R-Transformer on large-scale tasks like WikiText-103 or 1Bword (which prior works like QRNN, TrellisNet, RMC and Transformer-XL all explored) would be much more indicative of its usefulness (and should not be left out in such a paper). For instance, it would be useful to compare on small models with controlled model size, even if larger models that require TPUs are not available.\n    iv) Lack of ablative study. Table 4 seems to suggest that Transformer-XL is better than R-Transformer. Is it because of their usage of the relative positional embedding? How does the \"finite window size\" affect the performance of R-Transformer? Why don't you use Transformer-XL for the other 3 tasks? Lots of interesting questions are unanswered here. More details in (3) below.\n\n3. Regarding the motivation to insert a local RNN at the start of each layer. The authors claim (by citing Al-Rfou et al.) that the positional embedding (PE) have only limited effect. But in fact, Al-Rfou only says that the PE features (which is added only at the first layer, by the design of Vaswani et al.) can get lost once the transformer gets very deep--- which is totally expected. Therefore, Al-Rfou et al. propose to use a learnable embedding for each layer. This in no way suggests that positional embedding has \"limited effect\" (Note that in their ablative study, when they turned off the learnable PE, they also added back the original PE). Moreover, the authors didn't use convolution to capture local data, because it \"completely ignores the sequential information of positions within the local window\". That is true. However, **stacking** convolutional layers does capture sequential information. Prior works like [2] showed exactly how temporal convolutions are related to finite-window RNNs. I would suggest the authors to at least compare these different options with ablative studies. I would at least expect a comparison of i) temporal convolution; 2) fixed-window RNN (LocalRNN); 3) unlimited-window RNN; 4) positional embedding; 5) relative PE (which Transformer-XL uses; is that why it's better in Table 4?).\n\n4. The authors claim that the finite-window RNN captures local features. But doesn't that claim only applies to the first layer? Once the first layer multi-head attention mixes all input elements across the sequence, the \"local\" features fed into the second layer RNN will be, actually, **global** features? Doesn't that \"defeat\" the purpose of using a local RNN though?\n\n============================\n\nMinor issues that have mild or zero impact on the score:\n\n5. Inconsistent notations. Equation (2) and (3) both describe the LocalRNN(...) function, but clearly have a different input-output signature. Instead of using italics, it's better to have well-defined notations. Another case is Equation (7): you have 'FeedForward(mt)', but 'mt' is not on the right-hand side of the equation at all. Moreover, the symbols used in Eq. (7) are currently inconsistent with Eq. (8). \n\n6. Show the # of parameters in the model in Table 1-4.\n\n7. According to the code released in the dropbox URL, you precompute the index in the finite window and later called 'torch.index_select' to produce a tensor that is 'ksize' larger than the original sequence (cf. Line 130-143 in models/RTransformers.py in the dropbox folder). How does R-Transformer compare to Transformers in terms of speed and memory? Since you convert the batch dimension to (batch_size * seq_len) at every level of the network, I imagine that could slow down the process, especially in high-dimensional/large-scale experiments?\n\n8. There are some typos (e.g., Sec. 4.2, LTSM) and grammatical mistakes in the paper (Sec. 3.3). \n\n9. In Section 4.2, the authors claim that both LSTM and TCN performed better than Transformers on the polyphonic music dataset because \"these music tunes exhibit strong local structures\". However, the difference between Transformer and LSTM is actually very small, and it's even better than GRU/vanilla RNN. Is that too big a claim to make? In polyphonic music datasets like JSB or Nottingham, there are still longer sequences with longer dependencies...\n\n============================\n\nOverall, I feel that this paper has a good motivation to combine different sequence model families (Transformers, TCNs, RNNs) to improve their modeling power. But at the same time, I feel that the experiments can be a lot stronger, and the paper has limited novelty when compared to prior works. I'm happy to consider adjusting my score if my concerns above are addressed. \n\n[1] \"Generating Long Sequences with Sparse Transformers\", https://arxiv.org/abs/1904.10509\n[2] \"Trellis Networks for Sequence Modeling\", https://arxiv.org/abs/1810.06682\n[3] \"Learning Longer-term Dependencies in RNNs with Auxiliary Losses\", https://arxiv.org/abs/1803.00144", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "Summary: This paper proposes a new architecture, R-Transformer, that blends the Transformer networks and the recurrent networks, so as to better capture both the long- and short-term features. By injecting a local RNN layer at every level of the network, the authors hoped to enhance the Transformer's ability to model locality structure. To demonstrate the modeling power of R-Trasnformer, the paper evaluates the effectiveness of R-Transformer on 4 different sequence tasks (seqMNIST, polyphonic music, character- and word-level PTB).\n\nContribution: The authors propose an architecture that combines the practices of recurrent and feed-forward sequence models. However, I have major concerns regarding the novelty this paper, the various claims it makes, as well as its experiment setting.\n\n----------------------------------------------------------------------------\n\nMajor issues/questions:\n\n1. The techniques proposed by this paper lack novelty. For instance, the entire section 3.2 is simply the original design of the multi-head self-attention by Vaswani et al. The major difference between R-Transformer and the original Transformer is the replacement of positional embedding with an RNN layer, but (in my opinion) the authors did not demonstrate sufficiently its effectiveness via ablative studies (see below). Moreover, some prior works have already exploited the locality structure in Transformers. For instance, [1] showed that a sparse, local Transformer can work extremely well and be very efficient (they achieved SOTA on large-scale char-level language modeling tasks).\n\n2. The experiments do not entirely convince me. \n    i) The authors use the \"same hidden size for R-Transformer and Transformer.\" But in fact, as the R-Transformer has one extra RNN/LSTM/GRU layer at every level of the network, the tests were carried out (in effect) using a larger model than the baselines. I think the authors should instead control the # of model parameters, especially since you are running only on small tasks with small-sized models.\n    ii) It is nice that the authors tested R-Transformer on a variety of tasks--- this is important. However, in no way do these number achieve the levels of the \"state-of-the-art\", which the authors claim at the end of Section 1 (e.g., [2] has better number on seqMNIST and character-level PTB, and the Transformer-XL actually achieves <55 perplexity on word-level PTB). Therefore, the numbers don't look particularly appealing to me. \n    iii) Lack of more challenging, or large-scale experiments. Sequential MNIST is known to be a relatively simple task, and alternatives such as sequential CIFAR-10 or permuted sequential MNIST are more valuable \"small sequence tasks.\" (e.g., prior works studying long-term sequence model dependencies [3]) I also think that benchmarking R-Transformer on large-scale tasks like WikiText-103 or 1Bword (which prior works like QRNN, TrellisNet, RMC and Transformer-XL all explored) would be much more indicative of its usefulness (and should not be left out in such a paper). For instance, it would be useful to compare on small models with controlled model size, even if larger models that require TPUs are not available.\n    iv) Lack of ablative study. Table 4 seems to suggest that Transformer-XL is better than R-Transformer. Is it because of their usage of the relative positional embedding? How does the \"finite window size\" affect the performance of R-Transformer? Why don't you use Transformer-XL for the other 3 tasks? Lots of interesting questions are unanswered here. More details in (3) below.\n\n3. Regarding the motivation to insert a local RNN at the start of each layer. The authors claim (by citing Al-Rfou et al.) that the positional embedding (PE) have only limited effect. But in fact, Al-Rfou only says that the PE features (which is added only at the first layer, by the design of Vaswani et al.) can get lost once the transformer gets very deep--- which is totally expected. Therefore, Al-Rfou et al. propose to use a learnable embedding for each layer. This in no way suggests that positional embedding has \"limited effect\" (Note that in their ablative study, when they turned off the learnable PE, they also added back the original PE). Moreover, the authors didn't use convolution to capture local data, because it \"completely ignores the sequential information of positions within the local window\". That is true. However, **stacking** convolutional layers does capture sequential information. Prior works like [2] showed exactly how temporal convolutions are related to finite-window RNNs. I would suggest the authors to at least compare these different options with ablative studies. I would at least expect a comparison of i) temporal convolution; 2) fixed-window RNN (LocalRNN); 3) unlimited-window RNN; 4) positional embedding; 5) relative PE (which Transformer-XL uses; is that why it's better in Table 4?).\n\n4. The authors claim that the finite-window RNN captures local features. But doesn't that claim only applies to the first layer? Once the first layer multi-head attention mixes all input elements across the sequence, the \"local\" features fed into the second layer RNN will be, actually, **global** features? Doesn't that \"defeat\" the purpose of using a local RNN though?\n\n============================\n\nMinor issues that have mild or zero impact on the score:\n\n5. Inconsistent notations. Equation (2) and (3) both describe the LocalRNN(...) function, but clearly have a different input-output signature. Instead of using italics, it's better to have well-defined notations. Another case is Equation (7): you have 'FeedForward(mt)', but 'mt' is not on the right-hand side of the equation at all. Moreover, the symbols used in Eq. (7) are currently inconsistent with Eq. (8). \n\n6. Show the # of parameters in the model in Table 1-4.\n\n7. According to the code released in the dropbox URL, you precompute the index in the finite window and later called 'torch.index_select' to produce a tensor that is 'ksize' larger than the original sequence (cf. Line 130-143 in models/RTransformers.py in the dropbox folder). How does R-Transformer compare to Transformers in terms of speed and memory? Since you convert the batch dimension to (batch_size * seq_len) at every level of the network, I imagine that could slow down the process, especially in high-dimensional/large-scale experiments?\n\n8. There are some typos (e.g., Sec. 4.2, LTSM) and grammatical mistakes in the paper (Sec. 3.3). \n\n9. In Section 4.2, the authors claim that both LSTM and TCN performed better than Transformers on the polyphonic music dataset because \"these music tunes exhibit strong local structures\". However, the difference between Transformer and LSTM is actually very small, and it's even better than GRU/vanilla RNN. Is that too big a claim to make? In polyphonic music datasets like JSB or Nottingham, there are still longer sequences with longer dependencies...\n\n============================\n\nOverall, I feel that this paper has a good motivation to combine different sequence model families (Transformers, TCNs, RNNs) to improve their modeling power. But at the same time, I feel that the experiments can be a lot stronger, and the paper has limited novelty when compared to prior works. I'm happy to consider adjusting my score if my concerns above are addressed. \n\n[1] \"Generating Long Sequences with Sparse Transformers\", https://arxiv.org/abs/1904.10509\n[2] \"Trellis Networks for Sequence Modeling\", https://arxiv.org/abs/1810.06682\n[3] \"Learning Longer-term Dependencies in RNNs with Auxiliary Losses\", https://arxiv.org/abs/1803.00144"}, "tcdate": 1570246109759}], "openreview_url": "https://openreview.net/forum?id=HJx4PAEYDH", "arxiv_id": "1907.05572", "paper_pdf": "papers/HJx4PAEYDH.pdf", "paper_pdf_sha256": "62e5856240e4211c528462d9f8152354ca0f7c6b44dc605c6625ffb0f1498c7e", "paper_pdf_bytes": 211103, "paper_pdf_source": "openreview", "code_url": "https://github.com/DSE-MSU/R-transformer", "code_repository": "DSE-MSU/R-transformer", "code_commit": "3103cdfd117c2d82d64fb5ca56a6399bd30684e4", "code_archive": "repos/HJx4PAEYDH.zip", "code_archive_sha256": "75a1d0283436607620008862a43f44a45dbca6be189a3676e371f07de0e1f9cb", "code_archive_bytes": 362600, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 347, "github_languages": {"Python": 42839}, "github_archived": false, "github_pushed_at": "2019-07-16T03:31:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/r-transformer-recurrent-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XUJcsLvpaQ", "year": 2025, "status": "rejected", "title": "G-Transformer for Conditional Average Potential Outcome Estimation over Time", "authors": ["Konstantin Hess", "Dennis Frauen", "Valentyn Melnychuk", "Stefan Feuerriegel"], "authorids": ["~Konstantin_Hess1", "~Dennis_Frauen1", "~Valentyn_Melnychuk1", "~Stefan_Feuerriegel1"], "authors_source": "OpenReview API", "abstract": "Estimating potential outcomes for treatments over time based on observational data is important for personalized decision-making in medicine. Yet, existing neural methods for this task either (1) do not perform proper adjustments for time-varying confounders, or (2) suffer from large estimation variance. In order to address both limitations, we introduce the G-transformer (GT). Our GT is a novel, neural end-to-end model which adjusts for time-varying confounders, and provides low-variance estimation of conditional average potential outcomes (CAPOs) over time. Specifically, our GT is the first neural model to perform regression-based iterative G-computation for CAPOs in the time-varying setting. We evaluate the effectiveness of our GT across various experiments. In sum, this work represents a significant step towards personalized decision-making from electronic health records.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "l6r0cXokKm", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7215/Reviewer_AaBK"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper studies conditional average potential outcome estimation over time. The authors propose a model called G-Transformer for addressing two primary challenges in the field: adjustment for time-varying confounders and low variance in estimates. The authors validate their approach through experiments on synthetic and semi-synthetic data, demonstrating improvements over current baselines.", "review_text": "The paper studies conditional average potential outcome estimation over time. The authors propose a model called G-Transformer for addressing two primary challenges in the field: adjustment for time-varying confounders and low variance in estimates. The authors validate their approach through experiments on synthetic and semi-synthetic data, demonstrating improvements over current baselines.", "strengths": "- The paper provides a clear, detailed description of model architecture.\n\n- The authors conduct extensive experiments with synthetic and semi-synthetic datasets, demonstrating GT's effectiveness under different levels of confounding settings.\n\n- The code is provided.", "weaknesses": "- The contribution of this paper is somewhat limited. \n  * From the methodology perspective, the proposed method's improvement over existing approaches appears marginal. The primary contribution is exploring the transformer backbone in treatment effect estimation, which has been widely investigated in existing work (static setting [1] and dynamic setting [2]).\n  * From the application perspective, the paper claims that their model can help personalize decision-making from patient trajectories in medicine. However, it lacks discussion on how the model's estimates would concretely translate to real-world decision support. Furthermore, the assumptions on which the model relies (e.g., positivity, ignorability) are restrictive and may not hold robustly in practical, real-world scenarios.\n\n- While the paper claims to reduce estimation variance, it lacks a theoretical analysis of variance bounds. Given GT’s objective to minimize estimation variance, a more rigorous theoretical justification or a derivation of variance bounds would significantly strengthen this claim.\n\n-  The paper does not consider the existence of unobserved (and potentially time-varying) confounders, which can introduce biases and affect CAPO estimation in real-world EHR settings.\n\n- The study uses a semi-synthetic dataset derived from MIMIC-III. It would be valuable to see the model's performance on fully real-world datasets to assess its utility in practical clinical decision-making problems.\n\n[1] Zhang, Yi-Fan, et al. \"Exploring transformer backbones for heterogeneous treatment effect estimation.\" arXiv preprint arXiv:2202.01336 (2022).\n\n[2] Melnychuk, Valentyn, Dennis Frauen, and Stefan Feuerriegel. \"Causal transformer for estimating counterfactual outcomes.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies conditional average potential outcome estimation over time. The authors propose a model called G-Transformer for addressing two primary challenges in the field: adjustment for time-varying confounders and low variance in estimates. The authors validate their approach through experiments on synthetic and semi-synthetic data, demonstrating improvements over current baselines.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper provides a clear, detailed description of model architecture.\n\n- The authors conduct extensive experiments with synthetic and semi-synthetic datasets, demonstrating GT's effectiveness under different levels of confounding settings.\n\n- The code is provided.", "weaknesses": "- The contribution of this paper is somewhat limited. \n  * From the methodology perspective, the proposed method's improvement over existing approaches appears marginal. The primary contribution is exploring the transformer backbone in treatment effect estimation, which has been widely investigated in existing work (static setting [1] and dynamic setting [2]).\n  * From the application perspective, the paper claims that their model can help personalize decision-making from patient trajectories in medicine. However, it lacks discussion on how the model's estimates would concretely translate to real-world decision support. Furthermore, the assumptions on which the model relies (e.g., positivity, ignorability) are restrictive and may not hold robustly in practical, real-world scenarios.\n\n- While the paper claims to reduce estimation variance, it lacks a theoretical analysis of variance bounds. Given GT’s objective to minimize estimation variance, a more rigorous theoretical justification or a derivation of variance bounds would significantly strengthen this claim.\n\n-  The paper does not consider the existence of unobserved (and potentially time-varying) confounders, which can introduce biases and affect CAPO estimation in real-world EHR settings.\n\n- The study uses a semi-synthetic dataset derived from MIMIC-III. It would be valuable to see the model's performance on fully real-world datasets to assess its utility in practical clinical decision-making problems.\n\n[1] Zhang, Yi-Fan, et al. \"Exploring transformer backbones for heterogeneous treatment effect estimation.\" arXiv preprint arXiv:2202.01336 (2022).\n\n[2] Melnychuk, Valentyn, Dennis Frauen, and Stefan Feuerriegel. \"Causal transformer for estimating counterfactual outcomes.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730654305324}, {"id": "3TQ9wRGmyL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7215/Reviewer_5WFT"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This study proposes G-Transformer to estimate conditional average potential outcomes (CAPOs) over time while addressing two major limitations of the previous studies: (1) adjustments for time-varying confounders; (2) and large estimation variance. Particularly, the authors propose a viable solution to G-computation, iterative pseudo-outcome regression, using Transformer.", "review_text": "This study proposes G-Transformer to estimate conditional average potential outcomes (CAPOs) over time while addressing two major limitations of the previous studies: (1) adjustments for time-varying confounders; (2) and large estimation variance. Particularly, the authors propose a viable solution to G-computation, iterative pseudo-outcome regression, using Transformer.", "strengths": "The study is motivated and the problem is well-defined by the research gap of the previous studies. Specifically, the authors address the limitations of the previous studies regarding G-computation to estimate CAPOs with neural networks by providing a viable solution of iterative pseudo-outcome regression, although the neural network architecture itself is similar to the existing study (Melnychuk et al., 2022).\n\nThe author provides sufficient and proper supplementary materials to explain the details of the study.", "weaknesses": "The reviewer did not find any significant weaknesses.\n\nThe reviewer thinks a more principled and diverse way to evaluate the quality of the generated pseudo-outcomes will be desirable. While the experiment conducted in E.2 could empirically show the learning step is sensitive to the noise of the generated pseudo-outcomes, it might not be sufficient to show the quality of the generated pseudo-outcomes as it is titled. An ablation study that separates each step (i.e., another step is frozen when one step is trained) might be useful for a controlled experiment to measure the sensitivity of the performance in terms of how much the model is trained to generate the pseudo-outcomes.", "questions": "Table 2 and 3 have low readability due to small font sizes.\n\nThe reviewer thinks Supplementary E.2 regarding the quality of pseudo-outcome is worth being mentioned in the main text", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study proposes G-Transformer to estimate conditional average potential outcomes (CAPOs) over time while addressing two major limitations of the previous studies: (1) adjustments for time-varying confounders; (2) and large estimation variance. Particularly, the authors propose a viable solution to G-computation, iterative pseudo-outcome regression, using Transformer.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "The study is motivated and the problem is well-defined by the research gap of the previous studies. Specifically, the authors address the limitations of the previous studies regarding G-computation to estimate CAPOs with neural networks by providing a viable solution of iterative pseudo-outcome regression, although the neural network architecture itself is similar to the existing study (Melnychuk et al., 2022).\n\nThe author provides sufficient and proper supplementary materials to explain the details of the study.", "weaknesses": "The reviewer did not find any significant weaknesses.\n\nThe reviewer thinks a more principled and diverse way to evaluate the quality of the generated pseudo-outcomes will be desirable. While the experiment conducted in E.2 could empirically show the learning step is sensitive to the noise of the generated pseudo-outcomes, it might not be sufficient to show the quality of the generated pseudo-outcomes as it is titled. An ablation study that separates each step (i.e., another step is frozen when one step is trained) might be useful for a controlled experiment to measure the sensitivity of the performance in terms of how much the model is trained to generate the pseudo-outcomes.", "questions": "Table 2 and 3 have low readability due to small font sizes.\n\nThe reviewer thinks Supplementary E.2 regarding the quality of pseudo-outcome is worth being mentioned in the main text", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730619140990}, {"id": "HDLqn8oxst", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7215/Reviewer_i9bM"], "rating": 3, "soundness": 3, "presentation": 4, "contribution": 1, "confidence": 5, "summary": "This work proposed a framework to estimate conditional average potential outcome over time using a combination of transformer and iterative pseudo-outcome regression. The framework is tested in experiments on a synthetic dataset based on tumor growth model and a semi-synthetic dataset based on MIMIC-III, and is shown to outperform several benchmark models on these two datasets.", "review_text": "This work proposed a framework to estimate conditional average potential outcome over time using a combination of transformer and iterative pseudo-outcome regression. The framework is tested in experiments on a synthetic dataset based on tumor growth model and a semi-synthetic dataset based on MIMIC-III, and is shown to outperform several benchmark models on these two datasets.", "strengths": "1. Paper is well written, problem formulation, architecture descriptions and the presentation of learning procedures are all clearly stated and very easy to follow.\n\n2. The results look compeling. Table 2 and 3 showed noticable improvement of the proposed model over the compared benchmarks.", "weaknesses": "1. The novelty of this work is the most important concern to me. The combination of transformer and temporal-difference learning (i.e. pseudo-outcome regression as called by the authors) are already proposed in [1] for the estimation of CAPO (on top of that, [1] also went one step further and proposed the adjustment for marginal causal parameters). Besides, the heterogeneous token transformer architecture in [1] appears to be a more clean and intuitive solution as opposed to the proposed architecture in my judgement. The proposed transformer architecture is rather trivial and does not significantly differ from [2]. At the very least, the author should compare their architeture to [1], show experimental results and properly state their contributions compared to prior work.\n\n2. The paper has a very weak description of prior work related to pseudo-outcome regression in my opinion. Not only is the exact same procedure temporal-difference learning proposed before as stated above, it is widely used in causal inference and reinforcement learning known as sequential regression [3] and Q-learning [4]. The omission of these works by the authors made it look like pseudo-outcome regression is a novel contribution of their own.\n\n[1] Shirakawa, Toru, et al. \"Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer.\" International Conference on Machine Learning. PMLR, 2024.\n\n[2] Melnychuk, Valentyn, Dennis Frauen, and Stefan Feuerriegel. \"Causal transformer for estimating counterfactual outcomes.\" International Conference on Machine Learning. PMLR, 2022.\n\n[3] van der Laan, M. J. and Gruber, S. Targeted Minimum Loss Based Estimation of Causal Effects of Multiple Time Point Interventions. The International Journal of Biostatistics, 8(1), 2012.\n\n[4] Chebotar, Yevgen, et al. \"Q-transformer: Scalable offline reinforcement learning via autoregressive q-functions.\" Conference on Robot Learning. PMLR, 2023.", "questions": "My suggestions correspond to the two weaknesses above. I think comparison to prior work that already proposed the combination of transformer and temporal-difference learning/pseudo-outcome regression is absolutely crucial; literature review related to pseudo-outcome regression ([1][3][4] and other related work) should be added, and contribution of this work should be properly and correctly stated.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposed a framework to estimate conditional average potential outcome over time using a combination of transformer and iterative pseudo-outcome regression. The framework is tested in experiments on a synthetic dataset based on tumor growth model and a semi-synthetic dataset based on MIMIC-III, and is shown to outperform several benchmark models on these two datasets.", "soundness": 3, "presentation": 4, "contribution": 1, "strengths": "1. Paper is well written, problem formulation, architecture descriptions and the presentation of learning procedures are all clearly stated and very easy to follow.\n\n2. The results look compeling. Table 2 and 3 showed noticable improvement of the proposed model over the compared benchmarks.", "weaknesses": "1. The novelty of this work is the most important concern to me. The combination of transformer and temporal-difference learning (i.e. pseudo-outcome regression as called by the authors) are already proposed in [1] for the estimation of CAPO (on top of that, [1] also went one step further and proposed the adjustment for marginal causal parameters). Besides, the heterogeneous token transformer architecture in [1] appears to be a more clean and intuitive solution as opposed to the proposed architecture in my judgement. The proposed transformer architecture is rather trivial and does not significantly differ from [2]. At the very least, the author should compare their architeture to [1], show experimental results and properly state their contributions compared to prior work.\n\n2. The paper has a very weak description of prior work related to pseudo-outcome regression in my opinion. Not only is the exact same procedure temporal-difference learning proposed before as stated above, it is widely used in causal inference and reinforcement learning known as sequential regression [3] and Q-learning [4]. The omission of these works by the authors made it look like pseudo-outcome regression is a novel contribution of their own.\n\n[1] Shirakawa, Toru, et al. \"Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer.\" International Conference on Machine Learning. PMLR, 2024.\n\n[2] Melnychuk, Valentyn, Dennis Frauen, and Stefan Feuerriegel. \"Causal transformer for estimating counterfactual outcomes.\" International Conference on Machine Learning. PMLR, 2022.\n\n[3] van der Laan, M. J. and Gruber, S. Targeted Minimum Loss Based Estimation of Causal Effects of Multiple Time Point Interventions. The International Journal of Biostatistics, 8(1), 2012.\n\n[4] Chebotar, Yevgen, et al. \"Q-transformer: Scalable offline reinforcement learning via autoregressive q-functions.\" Conference on Robot Learning. PMLR, 2023.", "questions": "My suggestions correspond to the two weaknesses above. I think comparison to prior work that already proposed the combination of transformer and temporal-difference learning/pseudo-outcome regression is absolutely crucial; literature review related to pseudo-outcome regression ([1][3][4] and other related work) should be added, and contribution of this work should be properly and correctly stated.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730437898412}, {"id": "6eohaCWnqE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7215/Reviewer_5AUu"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes the G-transformer (GT) approach, which combines the G-computation formula and a multi-input transformer to estimate conditional average potential outcomes (CAPO) over time, given time-varying confounding and treatment observational data. GT iteratively generates pseudo-outcomes by reformulating the G-computation formula as recursive conditional expectations. Experimental results on synthetic and semi-synthetic datasets demonstrate that the proposed approach outperforms the baselines in terms of average root mean squared error (RMSE).", "review_text": "The paper proposes the G-transformer (GT) approach, which combines the G-computation formula and a multi-input transformer to estimate conditional average potential outcomes (CAPO) over time, given time-varying confounding and treatment observational data. GT iteratively generates pseudo-outcomes by reformulating the G-computation formula as recursive conditional expectations. Experimental results on synthetic and semi-synthetic datasets demonstrate that the proposed approach outperforms the baselines in terms of average root mean squared error (RMSE).", "strengths": "- The paper addresses the important and impactful real-world problem of estimating CAPO over time under time-varying confounding and treatment from observational data.\n- The proposed GT, iteratively generates pseudo-outcomes by reformulating the G-computation formula as recursive conditional expectations, which seems interesting.\n- Experimental results on synthetic and semi-synthetic datasets demonstrate that the proposed approach outperforms the baselines in terms of RMSE, especially under conditions of high time-varying confounding and longer CAPO estimation horizons.\n- The code and data are publicly available, making it easy to reproduce the results.", "weaknesses": "- The novelty is quite limited; the proposed approach could be considered a simple extension of the causal multi-input transformer (Melnychuk et al., 2022) with pseudo-outcome predictions.\n\n- While the paper claims to be the first to use pseudo-outcomes for CAPO estimation over time, several works have already explored the use of pseudo-observations in causal inference and survival analysis, including [1,2,3]. I encourage the author(s) to discuss these works and highlight key differences with their proposed approach.\n\n- The paper asserts that related works, such as G-Net, which properly adjust for time-varying confounding, result in high estimation variance without providing substantial experimental or theoretical results to support this claim. Only RMSE values with error bars are provided (Tables 2 and 3), which seems insufficient given that this is a central claim of the GT approach.\n\n- The experimental results seem underwhelming. Only one semi-synthetic dataset is considered, and RMSE is provided to quantify performance. I encourage the author(s) to explore evaluations tied to meaningful clinical measures, including precision in the estimation of heterogeneous effects, average treatment effects, and qualitative analysis measures of over- or underestimation of outcomes.\n\n- Clarity: The notation and writing need improvement for clarity and readability. The use of both $U$ and $H$ to denote the random variables $\\{Y_t, X_t, A_t\\}$ does not seem necessary. The paper seems to refer to $H$ throughout, and it's unclear why $U$ was introduced. The definitions for $\\bar{Y}$, $\\bar{X}$, and $\\bar{A}$ are not provided. The recursion could be made more explicit by combining Equations (8) and (9) and providing a simple example, e.g., $\\tau = 3$. The paper's use of $\\delta$, $\\tau$, and $t$ for indexing time needs further simplification. Additionally, the definition for $\\tau$ should be provided in the *Setup* section.\n\n\n**References**\n- [1] Andersen et al. (2017), \"Causal inference in survival analysis using pseudo-observations\", Statistics in medicine.\n- [2] Andersen and Perme (2010), \"Pseudo-observations in survival analysis\", Statistical methods in medical research.\n- [3] Chien-Lin et al. (2022), \"Causal inference for recurrent event data using pseudo-observations\", Biostatistics.\n\n**Minor**\n- Table 1: I encourage the author(s) to include citations to related works for improved readability.\n- Notation overload:  $\\phi$  is used to parameterize both functions $g(\\cdot  )$ and $z( \\cdot)$.\n- Typos: Line 140 should be  $U_{t+\\tau}$ remove the 'minus'.\n- I encourage the author(s) to increase the font size for tables and figures to be consistent with the main text.\n- The definition for $\\delta$ in Line 244 is not consistent with that in Line 267.", "questions": "- It seems that the GT approach has a clear advantage over existing approaches when $\\tau \\ge 2$. Could you provide scenarios where estimating CAPO over time ($\\tau \\ge 2$) has clear clinical benefits?\n- Could you quantify the variance of the baseline approaches, for example, using the coefficient of variation?\n- Could you provide additional results on the precision in the estimation of heterogeneous effects, average treatment effects, and qualitative analysis measures of over- or underestimation of outcomes?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes the G-transformer (GT) approach, which combines the G-computation formula and a multi-input transformer to estimate conditional average potential outcomes (CAPO) over time, given time-varying confounding and treatment observational data. GT iteratively generates pseudo-outcomes by reformulating the G-computation formula as recursive conditional expectations. Experimental results on synthetic and semi-synthetic datasets demonstrate that the proposed approach outperforms the baselines in terms of average root mean squared error (RMSE).", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper addresses the important and impactful real-world problem of estimating CAPO over time under time-varying confounding and treatment from observational data.\n- The proposed GT, iteratively generates pseudo-outcomes by reformulating the G-computation formula as recursive conditional expectations, which seems interesting.\n- Experimental results on synthetic and semi-synthetic datasets demonstrate that the proposed approach outperforms the baselines in terms of RMSE, especially under conditions of high time-varying confounding and longer CAPO estimation horizons.\n- The code and data are publicly available, making it easy to reproduce the results.", "weaknesses": "- The novelty is quite limited; the proposed approach could be considered a simple extension of the causal multi-input transformer (Melnychuk et al., 2022) with pseudo-outcome predictions.\n\n- While the paper claims to be the first to use pseudo-outcomes for CAPO estimation over time, several works have already explored the use of pseudo-observations in causal inference and survival analysis, including [1,2,3]. I encourage the author(s) to discuss these works and highlight key differences with their proposed approach.\n\n- The paper asserts that related works, such as G-Net, which properly adjust for time-varying confounding, result in high estimation variance without providing substantial experimental or theoretical results to support this claim. Only RMSE values with error bars are provided (Tables 2 and 3), which seems insufficient given that this is a central claim of the GT approach.\n\n- The experimental results seem underwhelming. Only one semi-synthetic dataset is considered, and RMSE is provided to quantify performance. I encourage the author(s) to explore evaluations tied to meaningful clinical measures, including precision in the estimation of heterogeneous effects, average treatment effects, and qualitative analysis measures of over- or underestimation of outcomes.\n\n- Clarity: The notation and writing need improvement for clarity and readability. The use of both $U$ and $H$ to denote the random variables $\\{Y_t, X_t, A_t\\}$ does not seem necessary. The paper seems to refer to $H$ throughout, and it's unclear why $U$ was introduced. The definitions for $\\bar{Y}$, $\\bar{X}$, and $\\bar{A}$ are not provided. The recursion could be made more explicit by combining Equations (8) and (9) and providing a simple example, e.g., $\\tau = 3$. The paper's use of $\\delta$, $\\tau$, and $t$ for indexing time needs further simplification. Additionally, the definition for $\\tau$ should be provided in the *Setup* section.\n\n\n**References**\n- [1] Andersen et al. (2017), \"Causal inference in survival analysis using pseudo-observations\", Statistics in medicine.\n- [2] Andersen and Perme (2010), \"Pseudo-observations in survival analysis\", Statistical methods in medical research.\n- [3] Chien-Lin et al. (2022), \"Causal inference for recurrent event data using pseudo-observations\", Biostatistics.\n\n**Minor**\n- Table 1: I encourage the author(s) to include citations to related works for improved readability.\n- Notation overload:  $\\phi$  is used to parameterize both functions $g(\\cdot  )$ and $z( \\cdot)$.\n- Typos: Line 140 should be  $U_{t+\\tau}$ remove the 'minus'.\n- I encourage the author(s) to increase the font size for tables and figures to be consistent with the main text.\n- The definition for $\\delta$ in Line 244 is not consistent with that in Line 267.", "questions": "- It seems that the GT approach has a clear advantage over existing approaches when $\\tau \\ge 2$. Could you provide scenarios where estimating CAPO over time ($\\tau \\ge 2$) has clear clinical benefits?\n- Could you quantify the variance of the baseline approaches, for example, using the coefficient of variation?\n- Could you provide additional results on the precision in the estimation of heterogeneous effects, average treatment effects, and qualitative analysis measures of over- or underestimation of outcomes?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730316288525}], "openreview_url": "https://openreview.net/forum?id=XUJcsLvpaQ", "arxiv_id": "2405.21012", "paper_pdf": "papers/XUJcsLvpaQ.pdf", "paper_pdf_sha256": "1596fd85f92d17081e408797830ecff564c3cc1521b128c5e23aebf9e519c7f0", "paper_pdf_bytes": 954115, "paper_pdf_source": "openreview", "code_url": "https://github.com/konstantinhess/G_transformer", "code_repository": "konstantinhess/G_transformer", "code_commit": "8452c89ef7da02f0793ca295d4e30455f5d77ee7", "code_archive": "repos/XUJcsLvpaQ.zip", "code_archive_sha256": "8a60b0bb112a220bd7b11837e07c575a6bc825744c6496ac93fa2f58514822a4", "code_archive_bytes": 136810, "code_file_count": 37, "code_extensions": {".py": 35, ".ipynb": 2}, "github_disk_usage_kb": 105, "github_languages": {"Python": 423591, "Jupyter Notebook": 33705}, "github_archived": false, "github_pushed_at": "2024-09-26T09:57:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/g-transformer-for-conditional-average"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rFZtyj5kBz", "year": 2024, "status": "rejected", "title": "Certifiably Byzantine-Robust Federated Conformal Prediction", "authors": ["Mintong Kang", "Zhen Lin", "Jimeng Sun", "Cao Xiao", "Bo Li"], "authorids": ["~Mintong_Kang1", "~Zhen_Lin2", "~Jimeng_Sun3", "~Cao_Xiao2", "~Bo_Li19"], "authors_source": "OpenReview API", "abstract": "Conformal prediction has shown impressive capacity in constructing statistically rigorous prediction sets for machine learning models with exchangeable data samples. The burgeoning amount of large-scale data, coupled with the escalating privacy concerns related to local data sharing, has inspired recent innovations extending conformal prediction into federated environments with distributed data samples. However, this framework for distributed uncertainty quantification is susceptible to Byzantine failures. A minor subset of malicious clients can significantly compromise the practicality of coverage guarantees. To address this vulnerability, we introduce a novel algorithm Rob-FCP to execute robust federated conformal prediction, effectively countering malicious clients capable of reporting arbitrary statistics during the conformal calibration process. We theoretically provide the conformal coverage bound of Rob-FCP in the Byzantine setting and show that the coverage of Rob-FCP is asymptotically close to the desired coverage level under mild conditions in both IID and non-IID settings. We also propose a malicious client number estimator to tackle a more challenging setting where the number of malicious clients is unknown to the defender and theoretically show its effectiveness. We empirically demonstrate the robustness of Rob-FCP against diverse proportions of malicious clients under a variety of Byzantine attacks on five realistic benchmark and healthcare datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "VctDOEQRnW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8184/Reviewer_YYE4"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors consider the conformal prediction interval construction in the federated learning setting where (1) each client reports conformality scores, and (2) a subset of clients are malicious and can report distorted scores to mess up with the interval construction. To alleviate the influence of malicious clients, the authors (1) identify malicious clients by comparing the score distributions from different clients under the assumption that conformity scores for benign clients are sampled from the same underlying distribution (IID setting), or from, essentially, distributions with bounded distance (non IID setting); (2) construct conformal prediction interval using the estimated benign clients.", "review_text": "The authors consider the conformal prediction interval construction in the federated learning setting where (1) each client reports conformality scores, and (2) a subset of clients are malicious and can report distorted scores to mess up with the interval construction. To alleviate the influence of malicious clients, the authors (1) identify malicious clients by comparing the score distributions from different clients under the assumption that conformity scores for benign clients are sampled from the same underlying distribution (IID setting), or from, essentially, distributions with bounded distance (non IID setting); (2) construct conformal prediction interval using the estimated benign clients.", "strengths": "The paper is overall easy to follow; and robust prediction intervals against malicious clients are an important question.", "weaknesses": "It seems to be that the assumptions considered seem over-simplified and may lead to less robustness and under-coverage of difficulty cases when violated: the entire paper is based on the assumption that benign clients have similar conformity score distributions, in both IID settings (identical) and non-IID settings (close), and a client whose conformity score is far from its K_b \"neighbors\" is claimed malicious.  However, in practice, benign clients can have data with different local characteristics, due to, e.g., demographic differences, differences in medical practice guidelines, etc.", "questions": "More details of the nonIID setting + different levels of being nonIID can be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors consider the conformal prediction interval construction in the federated learning setting where (1) each client reports conformality scores, and (2) a subset of clients are malicious and can report distorted scores to mess up with the interval construction. To alleviate the influence of malicious clients, the authors (1) identify malicious clients by comparing the score distributions from different clients under the assumption that conformity scores for benign clients are sampled from the same underlying distribution (IID setting), or from, essentially, distributions with bounded distance (non IID setting); (2) construct conformal prediction interval using the estimated benign clients.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "The paper is overall easy to follow; and robust prediction intervals against malicious clients are an important question.", "weaknesses": "It seems to be that the assumptions considered seem over-simplified and may lead to less robustness and under-coverage of difficulty cases when violated: the entire paper is based on the assumption that benign clients have similar conformity score distributions, in both IID settings (identical) and non-IID settings (close), and a client whose conformity score is far from its K_b \"neighbors\" is claimed malicious.  However, in practice, benign clients can have data with different local characteristics, due to, e.g., demographic differences, differences in medical practice guidelines, etc.", "questions": "More details of the nonIID setting + different levels of being nonIID can be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698812912195}, {"id": "I370WogJ1n", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8184/Reviewer_eCJ4"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces an algorithm called Rob-FCP to perform Conformal Prediction in a federated learning setting where some agents are potentially malicious. This algorithm first discards the malicious agents and then performs a standard conformal prediction algorithm. Authors theoretically provide the conformal coverage bound of Rob-FCP in the Byzantine setting and show that the coverage of Rob-FCP is asymptotically close to the desired coverage level under mild conditions in both IID and non-IID settings. An algorithm for automatically determining the number of malicious agents is also provided. Finally, empirical experiments demonstrate that Rob-FCP is effective.", "review_text": "This paper introduces an algorithm called Rob-FCP to perform Conformal Prediction in a federated learning setting where some agents are potentially malicious. This algorithm first discards the malicious agents and then performs a standard conformal prediction algorithm. Authors theoretically provide the conformal coverage bound of Rob-FCP in the Byzantine setting and show that the coverage of Rob-FCP is asymptotically close to the desired coverage level under mild conditions in both IID and non-IID settings. An algorithm for automatically determining the number of malicious agents is also provided. Finally, empirical experiments demonstrate that Rob-FCP is effective.", "strengths": "1\\ This is a very interesting problem for the conformal prediction community.\n\n2\\ The paper treats both the iid setting and the non-iid setting.\n\n3\\ The paper is well-written, clear, and easy to follow.\n\n4\\ The experience shows that the method performs well in this federated learning setting with malicious agents.", "weaknesses": "1\\ A major weakness is that to calculate the vector distance $d_{k_1, k_2}$ in step 8 of the algorithm (\"Algorithm 1 Identifying the malicious client\"), we need to send all the vectors $v^{(k)}$ to the server. It seems to me that this step is very problematic in a federated learning context.\n\n2\\ Another important weakness is that the bounds of Theorem 1 and Corollary 1 are in $1/(\\min n_i)$. Therefore, if a non-malicious agent has only one data point, the bound does not improve, even if the other agents have an increasing number of data. This may just be due to an artifact of the proof or maybe we really can't do any better. This has to be proven (or at least discussed).\n\n3\\ Citations are not always appropriate. For instance, the split conformal prediction method is attributed to Lei et al., 2018 but the citation should be Papadopoulos et al., 2002.  The same applies to the FCP which does not cite the first paper due to Lu et al., 2021, and the more recent one Humbert et al., 2023, and to related work on federated learning, which only cites papers from 2019 and above.\n\n4\\ Although the experiences are well explained and in large quantities, I think that the parameters of the experiments are not always given (maybe I am wrong). For example, we do not know how malicious data are generated.\n\nMinor:\n1\\ \"marginal prediction coverage: ...\" the definition is with an \"inclusion\"\n2\\ Mixture coefficients in the definition of $Q_{\\lambda}$ are missing.\n3\\ Problem in the definition of $N_m$.\n4\\ In the abstract and introduction, the federated learning framework is also justified by privacy concerns. In general, it is not true that federated learning guarantees privacy.", "questions": "1\\ Is it possible to compute the vector distance $d_{k_1, k_2}$ with a federated algorithm ?\n\n2\\ The $\\min{n_i}$ in the bound of Theorem 1 and Corollary 1 cannot be improved or is it just an artifact of the proof?\n\n3\\ In Theorem 1 and Corollary 1, $\\varepsilon$ appears in the bound. Is it possible to control it? \n\n4\\ In addition to the previous question, in the experiment how much time it takes to compute the quantile? (and for wich $\\varepsilon$ ?)\n\n5\\ Regarding my remark on privacy, is it possible/easy to extend Rob-FCP in order to have differential privacy guarantees?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an algorithm called Rob-FCP to perform Conformal Prediction in a federated learning setting where some agents are potentially malicious. This algorithm first discards the malicious agents and then performs a standard conformal prediction algorithm. Authors theoretically provide the conformal coverage bound of Rob-FCP in the Byzantine setting and show that the coverage of Rob-FCP is asymptotically close to the desired coverage level under mild conditions in both IID and non-IID settings. An algorithm for automatically determining the number of malicious agents is also provided. Finally, empirical experiments demonstrate that Rob-FCP is effective.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1\\ This is a very interesting problem for the conformal prediction community.\n\n2\\ The paper treats both the iid setting and the non-iid setting.\n\n3\\ The paper is well-written, clear, and easy to follow.\n\n4\\ The experience shows that the method performs well in this federated learning setting with malicious agents.", "weaknesses": "1\\ A major weakness is that to calculate the vector distance $d_{k_1, k_2}$ in step 8 of the algorithm (\"Algorithm 1 Identifying the malicious client\"), we need to send all the vectors $v^{(k)}$ to the server. It seems to me that this step is very problematic in a federated learning context.\n\n2\\ Another important weakness is that the bounds of Theorem 1 and Corollary 1 are in $1/(\\min n_i)$. Therefore, if a non-malicious agent has only one data point, the bound does not improve, even if the other agents have an increasing number of data. This may just be due to an artifact of the proof or maybe we really can't do any better. This has to be proven (or at least discussed).\n\n3\\ Citations are not always appropriate. For instance, the split conformal prediction method is attributed to Lei et al., 2018 but the citation should be Papadopoulos et al., 2002.  The same applies to the FCP which does not cite the first paper due to Lu et al., 2021, and the more recent one Humbert et al., 2023, and to related work on federated learning, which only cites papers from 2019 and above.\n\n4\\ Although the experiences are well explained and in large quantities, I think that the parameters of the experiments are not always given (maybe I am wrong). For example, we do not know how malicious data are generated.\n\nMinor:\n1\\ \"marginal prediction coverage: ...\" the definition is with an \"inclusion\"\n2\\ Mixture coefficients in the definition of $Q_{\\lambda}$ are missing.\n3\\ Problem in the definition of $N_m$.\n4\\ In the abstract and introduction, the federated learning framework is also justified by privacy concerns. In general, it is not true that federated learning guarantees privacy.", "questions": "1\\ Is it possible to compute the vector distance $d_{k_1, k_2}$ with a federated algorithm ?\n\n2\\ The $\\min{n_i}$ in the bound of Theorem 1 and Corollary 1 cannot be improved or is it just an artifact of the proof?\n\n3\\ In Theorem 1 and Corollary 1, $\\varepsilon$ appears in the bound. Is it possible to control it? \n\n4\\ In addition to the previous question, in the experiment how much time it takes to compute the quantile? (and for wich $\\varepsilon$ ?)\n\n5\\ Regarding my remark on privacy, is it possible/easy to extend Rob-FCP in order to have differential privacy guarantees?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698747415421}, {"id": "mypUsHzc7z", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8184/Reviewer_bwYM"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper studies federated conformal prediction (i.e. perform conformal prediction when the calibration dataset is distributed among multiple agents) where there may be Byzantine agents who may act maliciously. The paper shows how to detect such malicious behaviors by recording their deviation from the rest of the population; for each agent, they record the empirical histogram over H bins, calculate the average distance away from their closest neighbors where the distance is measured by the Lp distance between their H histogram bins, and mark agents with high distance away from their neighbors as malicious agents. \n\nThey show that under a few assumptions (iid or non-iid + some other assumptions), their algorithm can identify malicious agents and provide a target coverage guarantee (Theorem 1 and Corollary 1). In the case the number of malicious agents is not known, they show a way to estimate such number via some approximation (approximating the multinomial with multivariate normal). \n\nFinally, they evaluate their algorithm on a few different datasets (Section 5).", "review_text": "The paper studies federated conformal prediction (i.e. perform conformal prediction when the calibration dataset is distributed among multiple agents) where there may be Byzantine agents who may act maliciously. The paper shows how to detect such malicious behaviors by recording their deviation from the rest of the population; for each agent, they record the empirical histogram over H bins, calculate the average distance away from their closest neighbors where the distance is measured by the Lp distance between their H histogram bins, and mark agents with high distance away from their neighbors as malicious agents. \n\nThey show that under a few assumptions (iid or non-iid + some other assumptions), their algorithm can identify malicious agents and provide a target coverage guarantee (Theorem 1 and Corollary 1). In the case the number of malicious agents is not known, they show a way to estimate such number via some approximation (approximating the multinomial with multivariate normal). \n\nFinally, they evaluate their algorithm on a few different datasets (Section 5).", "strengths": "-The idea to identify the malicious agents behavior via their deviation from the non-malicious agents in terms of the non-conformal score distribution seems novel.\n-Their algorithm seem to perform quite well in their experiments.", "weaknesses": "-The main reason why the algorithm in the paper works seems to be due to the homogeneity of the non-malicious agents. Even in the experiments, the clients are partitioned randomly and hence their distributions will be pretty similar. However, as discussed even in the intro of the paper, there can be many settings where there is quite a bit of heterogeneity among the agents not due to their Byzantine and malicious behaviors but the underlying distributions are just inherently different. In fact, this heterogeneity seemed to me the motivation to studying this problem — i.e. the bolded sentences in the second paragraph of the intro. But the paper seems to study cases when there isn’t much heterogeneity?", "questions": "-It’s not clear to me what is exactly meant by the non-iid setting. Is it just that the agents’s non-conformal scores don’t come from the same distribution? However, the assumption that the v^(k) values (the histogram values) aren’t too different across agents in the non-iid setting essentially gives you an iid setting, right? I’m a little confused on the difference so it would be helpful to compare exactly the iid setting and the non-iid setting (along with other assumptions made for each setting) and how exactly they are really different.\n\n\n-In proving the main result (Theorem 1), the appearance of the inverse of the CDF of the standard normal distribution seems surprising as there was no normality assumption before; it would be good to cite a reference for what is referred to as “the binomial proportion confidence interval” or eqn (13) in the proof of Theorem 1. Also, can’t one just apply the DKW inequality (https://en.wikipedia.org/wiki/Dvoretzky%E2%80%93Kiefer%E2%80%93Wolfowitz_inequality) here? It should be sufficient to show concentration of the empirical CDF (which exactly characterizes the empirical histogram values v(k)_h) toward the true CDF (which also characterizes the true bar(v) values). This would avoid the need to union bound over H values too as DKW tells you that over all possible h values (i.e. hth cut point), the empirical and the true CDF value is close with high probability. \n\n\n -How’s the final (1-alpha)-quantile being calculated after identifying the benign clients? Is it simply combining all the clients non-conformal scores altogether and then calculating the (1-alpha)-quantile? If that’s the case, I’m not understanding the federated nature of this problem except for the fact that some of the clients data are being ignored due to their apparent distributional difference to other clients?  How can one tell if this distributional difference of some clients is due to their malicious behavior or truly inherent distributional difference?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies federated conformal prediction (i.e. perform conformal prediction when the calibration dataset is distributed among multiple agents) where there may be Byzantine agents who may act maliciously. The paper shows how to detect such malicious behaviors by recording their deviation from the rest of the population; for each agent, they record the empirical histogram over H bins, calculate the average distance away from their closest neighbors where the distance is measured by the Lp distance between their H histogram bins, and mark agents with high distance away from their neighbors as malicious agents. \n\nThey show that under a few assumptions (iid or non-iid + some other assumptions), their algorithm can identify malicious agents and provide a target coverage guarantee (Theorem 1 and Corollary 1). In the case the number of malicious agents is not known, they show a way to estimate such number via some approximation (approximating the multinomial with multivariate normal). \n\nFinally, they evaluate their algorithm on a few different datasets (Section 5).", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "-The idea to identify the malicious agents behavior via their deviation from the non-malicious agents in terms of the non-conformal score distribution seems novel.\n-Their algorithm seem to perform quite well in their experiments.", "weaknesses": "-The main reason why the algorithm in the paper works seems to be due to the homogeneity of the non-malicious agents. Even in the experiments, the clients are partitioned randomly and hence their distributions will be pretty similar. However, as discussed even in the intro of the paper, there can be many settings where there is quite a bit of heterogeneity among the agents not due to their Byzantine and malicious behaviors but the underlying distributions are just inherently different. In fact, this heterogeneity seemed to me the motivation to studying this problem — i.e. the bolded sentences in the second paragraph of the intro. But the paper seems to study cases when there isn’t much heterogeneity?", "questions": "-It’s not clear to me what is exactly meant by the non-iid setting. Is it just that the agents’s non-conformal scores don’t come from the same distribution? However, the assumption that the v^(k) values (the histogram values) aren’t too different across agents in the non-iid setting essentially gives you an iid setting, right? I’m a little confused on the difference so it would be helpful to compare exactly the iid setting and the non-iid setting (along with other assumptions made for each setting) and how exactly they are really different.\n\n\n-In proving the main result (Theorem 1), the appearance of the inverse of the CDF of the standard normal distribution seems surprising as there was no normality assumption before; it would be good to cite a reference for what is referred to as “the binomial proportion confidence interval” or eqn (13) in the proof of Theorem 1. Also, can’t one just apply the DKW inequality (https://en.wikipedia.org/wiki/Dvoretzky%E2%80%93Kiefer%E2%80%93Wolfowitz_inequality) here? It should be sufficient to show concentration of the empirical CDF (which exactly characterizes the empirical histogram values v(k)_h) toward the true CDF (which also characterizes the true bar(v) values). This would avoid the need to union bound over H values too as DKW tells you that over all possible h values (i.e. hth cut point), the empirical and the true CDF value is close with high probability. \n\n\n -How’s the final (1-alpha)-quantile being calculated after identifying the benign clients? Is it simply combining all the clients non-conformal scores altogether and then calculating the (1-alpha)-quantile? If that’s the case, I’m not understanding the federated nature of this problem except for the fact that some of the clients data are being ignored due to their apparent distributional difference to other clients?  How can one tell if this distributional difference of some clients is due to their malicious behavior or truly inherent distributional difference?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698704722402}, {"id": "e9lD8rrB9N", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8184/Reviewer_LS6w"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors study a robust conformal prediction problem in a federated learning setting. Since local datasets share some privacy properties, some agents may not truthfully report their non-conformity score.  In this scenario, making a prediction interval based on this contaminated data would be harmful and lack the correct coverage probability. To solve such an issue, the authors propose a novel algorithm and obtain a coverage bound close to the desired level under both iid and non-iid settings.", "review_text": "The authors study a robust conformal prediction problem in a federated learning setting. Since local datasets share some privacy properties, some agents may not truthfully report their non-conformity score.  In this scenario, making a prediction interval based on this contaminated data would be harmful and lack the correct coverage probability. To solve such an issue, the authors propose a novel algorithm and obtain a coverage bound close to the desired level under both iid and non-iid settings.", "strengths": "1. The authors study a realistic setting where the agents may not be truthful in reporting their non-conformity score, and they developed an ``outlier detection\" algorithm to ensure the prediction sets are made based on truthful agents.\n\n2. Rigorous theoretical guarantees are provided on the coverage probability lower and upper bounds.\n\n3. The methodology is applied to various types of datasets to measure its performance, which is great.", "weaknesses": "1. It seems the results are derived based on either knowing the number of truthful agents or being able to consistently estimate this quantity.  This could be a relatively strong assumption in reality.\n\n2. The assumption on bounded distribution disparity could be relatively strong. There would be the cases where the underlying distributions of different group of agents are not the same. In that case, this assumption can be easily violated.", "questions": "1. Could the authors establish some theoretical guarantees where the number of truthful agents is not correctly estimated? What will happen if they are underestimated or overestimated? \n\n2. Could the authors also provide some numerical results to illustrate the sensitivity if the number of truthful agents is not correctly identified?\n\n3. Again, as I mentioned in the weakness part, if the distributions of underlying agents are no-i.i.d., it is very likely that assumption 3.1 will be violated, even if the agents report the true non-conformity score. It seems the current algorithm may not handle this point very well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors study a robust conformal prediction problem in a federated learning setting. Since local datasets share some privacy properties, some agents may not truthfully report their non-conformity score.  In this scenario, making a prediction interval based on this contaminated data would be harmful and lack the correct coverage probability. To solve such an issue, the authors propose a novel algorithm and obtain a coverage bound close to the desired level under both iid and non-iid settings.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The authors study a realistic setting where the agents may not be truthful in reporting their non-conformity score, and they developed an ``outlier detection\" algorithm to ensure the prediction sets are made based on truthful agents.\n\n2. Rigorous theoretical guarantees are provided on the coverage probability lower and upper bounds.\n\n3. The methodology is applied to various types of datasets to measure its performance, which is great.", "weaknesses": "1. It seems the results are derived based on either knowing the number of truthful agents or being able to consistently estimate this quantity.  This could be a relatively strong assumption in reality.\n\n2. The assumption on bounded distribution disparity could be relatively strong. There would be the cases where the underlying distributions of different group of agents are not the same. In that case, this assumption can be easily violated.", "questions": "1. Could the authors establish some theoretical guarantees where the number of truthful agents is not correctly estimated? What will happen if they are underestimated or overestimated? \n\n2. Could the authors also provide some numerical results to illustrate the sensitivity if the number of truthful agents is not correctly identified?\n\n3. Again, as I mentioned in the weakness part, if the distributions of underlying agents are no-i.i.d., it is very likely that assumption 3.1 will be violated, even if the agents report the true non-conformity score. It seems the current algorithm may not handle this point very well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698606984710}], "openreview_url": "https://openreview.net/forum?id=rFZtyj5kBz", "arxiv_id": "2406.01960", "paper_pdf": "papers/rFZtyj5kBz.pdf", "paper_pdf_sha256": "203e6269102add8a53b65741a4fc48517527c8ea7f67bcd84108c63cf7d32c4c", "paper_pdf_bytes": 1018447, "paper_pdf_source": "openreview", "code_url": "https://github.com/kangmintong/Rob-FCP", "code_repository": "kangmintong/Rob-FCP", "code_commit": "2475c51ed9c610eb81a0396d4283317108410d67", "code_archive": "repos/rFZtyj5kBz.zip", "code_archive_sha256": "3bae5ca9179b48e65d43e9960f51fe5f032210dd0ce03fbd0dab4f45e5a82c88", "code_archive_bytes": 90984, "code_file_count": 22, "code_extensions": {".py": 18, ".sh": 4}, "github_disk_usage_kb": 79, "github_languages": {"Python": 115303, "Shell": 10448}, "github_archived": false, "github_pushed_at": "2024-10-24T14:22:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/certifiably-byzantine-robust-federated"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9Zx6tTcX0SE", "year": 2023, "status": "rejected", "title": "A Study of Biologically Plausible Neural Network: the Role and Interactions of Brain-Inspired Mechanisms in Continual Learning", "authors": ["Fahad Sarfraz", "Elahe Arani", "Bahram Zonooz"], "authorids": ["~Fahad_Sarfraz1", "~Elahe_Arani1", "~Bahram_Zonooz1"], "authors_source": "OpenReview API", "abstract": "Humans excel at continually acquiring, consolidating, and retaining information from an ever-changing environment, whereas artificial neural networks (ANNs) exhibit catastrophic forgetting. There are considerable differences in the complexity of synapses, the processing of information, and the learning mechanisms in biological neural networks and their artificial counterpart, which may explain the mismatch in performance. We consider a biologically plausible framework that constitutes separate populations of exclusively excitatory and inhibitory neurons which adhere to Dale's principle and the excitatory pyramidal neurons are augmented with dendritic-like structures for context-dependent processing of stimuli. We then conduct a comprehensive study on the role and interactions of different mechanisms inspired by the brain including sparse non-overlapping representations, Hebbian learning, synaptic consolidation, and replay of past activations that accompanied the learning event. Our study suggests that employing multiple complementary mechanisms in a biologically plausible architecture, similar to the brain, can be effective in enabling continual learning in ANNs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ZyVZ5c3O62", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6598/Reviewer_Ds8Z"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper evaluates previous work on biologically plausible DNNs in the setting of continual learning (CL). Namely, they evaluate ideas around Dale’s principle, Active Dendrites, heterogenous dropout, Hebbian learning, synaptic consolidation and experience replay. They present experimental evidence that these ideas can individually improve the performance on CL, or be combined to achieve greater improvement.", "review_text": "While the paper presents an interesting discussion, it reuses ideas from other work and presents limited experimental evidence.", "strengths": "+ This paper provided a useful list of biologically-inspired modifications to DNNs.\n+ I find the main idea interesting (investigating such modifications’ usefulness to CL).\n- The paper is a compilation of already published work.\n- Some of these insights (s.a. dropout being useful for CL) can be found in literature.\n- The experiments are exclusively done on MNIST-based datasets. While they provide evidence to support the claims of this paper, it would be useful to understand whether this translates into other image domains.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper evaluates previous work on biologically plausible DNNs in the setting of continual learning (CL). Namely, they evaluate ideas around Dale’s principle, Active Dendrites, heterogenous dropout, Hebbian learning, synaptic consolidation and experience replay. They present experimental evidence that these ideas can individually improve the performance on CL, or be combined to achieve greater improvement.", "strength_and_weaknesses": "+ This paper provided a useful list of biologically-inspired modifications to DNNs.\n+ I find the main idea interesting (investigating such modifications’ usefulness to CL).\n- The paper is a compilation of already published work.\n- Some of these insights (s.a. dropout being useful for CL) can be found in literature.\n- The experiments are exclusively done on MNIST-based datasets. While they provide evidence to support the claims of this paper, it would be useful to understand whether this translates into other image domains.", "clarity,_quality,_novelty_and_reproducibility": "I found the discussion around the properties of biological neural networks a bit confusing. Concretely, I don’t think that the terminology was clearly described.\nNovelty: I do not think that the ideas expressed in this paper are novel. A “related work” section would be a great way to set the paper apart from similar insights expressed in CL literature, and point out how the ideas of this paper are different.\n", "summary_of_the_review": "While the paper presents an interesting discussion, it reuses ideas from other work and presents limited experimental evidence.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666703938614}, {"id": "eWpE60xEnx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6598/Reviewer_Bd22"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work the authors study the individual and combined effect of a range of biologically inspired mechanisms on continual learning. ", "review_text": "Overall I think this is an important and interesting study, though interactions between mechanisms could be explored better and some controls are missing. ", "strengths": "Strengths\n- Interesting and important research for both AI and neuroscience\n- Paper is well written and all approaches are clear and well described \n\nWeaknesses\n- No floor (e.g normal training) and ceiling (e.g. full replay) controls provided in the experiments.\n- Experiments limited to variants of MNIST. Unclear if active dendrites approach & prototypes/context will be as effective with datasets without an exemplar structure, like fashion mnist. \n- Unclear how the difference mechanisms interact with each other, but possibly beyond the scope of a conference paper. \n- Why do active dendrites fail with seq-mnist?\n- Is there biological evidence for context dependent dendritic processing of eq. (4). \n- Heterogeneous dropout does not seem to be biologically motivated. \n- Novelty appears limited to combining existing approaches, the authors could more clearly describe their specific novel contributions. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this work the authors study the individual and combined effect of a range of biologically inspired mechanisms on continual learning. ", "strength_and_weaknesses": "Strengths\n- Interesting and important research for both AI and neuroscience\n- Paper is well written and all approaches are clear and well described \n\nWeaknesses\n- No floor (e.g normal training) and ceiling (e.g. full replay) controls provided in the experiments.\n- Experiments limited to variants of MNIST. Unclear if active dendrites approach & prototypes/context will be as effective with datasets without an exemplar structure, like fashion mnist. \n- Unclear how the difference mechanisms interact with each other, but possibly beyond the scope of a conference paper. \n- Why do active dendrites fail with seq-mnist?\n- Is there biological evidence for context dependent dendritic processing of eq. (4). \n- Heterogeneous dropout does not seem to be biologically motivated. \n- Novelty appears limited to combining existing approaches, the authors could more clearly describe their specific novel contributions. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and of good quality. Novelty appears somewhat limited. \n", "summary_of_the_review": "Overall I think this is an important and interesting study, though interactions between mechanisms could be explored better and some controls are missing. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666687147853}, {"id": "FpYwnAT0vg", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6598/Reviewer_jDQX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors claim that including biological features in artificial neural networks can improve their performance on tasks such as continual learning. The authors propose to include four different biological features into their models and evaluate their relevance: 1) segregation of neurons into populations of excitatory and inhibitory neurons as in Dale's principle. 2) Dendritic-inspired structures as modulators and sparsifiers. 3) Hebbian plasticity rules. 4) Experience replay and regularization. \n\nThe different scenarios are evaluated based on variations of the famous MNIST dataset, showing the influence of the different parameters when considered together or separately. The authors claim that the results obtained offer compelling evidence. ", "review_text": "This paper addresses the important question of the influence of biological inspiration on the design of artificial neural networks. \nThe paper falls short however in how the influence of the different features is evaluated and motivated. \nIt is a bit of a patchwork of biological components rather than a cohesive model. \n\nI greatly encourage the author to further this research as it can provide a great alternative to the blind study of artificial neural networks. ", "strengths": "The paper addresses an important question on the importance of biological features for guiding the design principle of artificial neural networks. \n\nHowever, it is unclear how we can evaluate the importance of such features with the currently limited evaluation. How does it generalize to other datasets? What about synthetic datasets, where one can control for the different structures of the data that the features are supposed to be useful for?  \n\nThe influence of the different parameters is primarily measured in accuracy, but how is that the suitable proxy for measuring the impact of these features? Why are those features even relevant in this context? \n\nWhat is the rationale for combining Dale's principle, pyramidal-like cells, and Hebbian principles? Pyramidal cells are arguably not driven by the Hebbian principle. How is sparsity enforced without interneurons?  \nThe biological plausibility of those features is mostly in name but is vaguely plausible in their current implementation; as such, it cannot be one of the paper's main claims.\nMore rigor in how these concepts are handled would be greatly appreciated. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors claim that including biological features in artificial neural networks can improve their performance on tasks such as continual learning. The authors propose to include four different biological features into their models and evaluate their relevance: 1) segregation of neurons into populations of excitatory and inhibitory neurons as in Dale's principle. 2) Dendritic-inspired structures as modulators and sparsifiers. 3) Hebbian plasticity rules. 4) Experience replay and regularization. \n\nThe different scenarios are evaluated based on variations of the famous MNIST dataset, showing the influence of the different parameters when considered together or separately. The authors claim that the results obtained offer compelling evidence. ", "strength_and_weaknesses": "The paper addresses an important question on the importance of biological features for guiding the design principle of artificial neural networks. \n\nHowever, it is unclear how we can evaluate the importance of such features with the currently limited evaluation. How does it generalize to other datasets? What about synthetic datasets, where one can control for the different structures of the data that the features are supposed to be useful for?  \n\nThe influence of the different parameters is primarily measured in accuracy, but how is that the suitable proxy for measuring the impact of these features? Why are those features even relevant in this context? \n\nWhat is the rationale for combining Dale's principle, pyramidal-like cells, and Hebbian principles? Pyramidal cells are arguably not driven by the Hebbian principle. How is sparsity enforced without interneurons?  \nThe biological plausibility of those features is mostly in name but is vaguely plausible in their current implementation; as such, it cannot be one of the paper's main claims.\nMore rigor in how these concepts are handled would be greatly appreciated. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper is relatively well written. The motivations and rigor expected are lacking but various things are suggested in order to possibly improve the current submission. \n\nThe novelties are minor, it is rather incremental work. ", "summary_of_the_review": "This paper addresses the important question of the influence of biological inspiration on the design of artificial neural networks. \nThe paper falls short however in how the influence of the different features is evaluated and motivated. \nIt is a bit of a patchwork of biological components rather than a cohesive model. \n\nI greatly encourage the author to further this research as it can provide a great alternative to the blind study of artificial neural networks. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666640529597}, {"id": "QHQHM3Z6Wi", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6598/Reviewer_TqnH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors introduce a biologically inspired neural network architecture that combines a variety of learning and processing mechanisms, with the goal of studying whether the performance of artificial systems might be improved by increasing their computational realism. In particular, the authors explore the implementation of Dale’s principle, Hebbian updates, sparse coding, heterogeneous dropout, synaptic consolidation, and experience replay. The proposed model is tested in continual learning (CL) tasks derived from the popular MNIST benchmark, and results suggest significant performance gains.", "review_text": "Overall, my judgment of this paper is positive. The research scope of this paper is broad enough to be of interest to the audience at ICLR and stimulate further interdisciplinary research work. At the same time, I believe that further simulations are required to make the paper stronger.", "strengths": "One of the strengths of this paper is that it combines a remarkable number of recent ideas and biological processing mechanisms into a unified architecture. The proposed model is also tested in an incremental fashion, which allows to better understand the contribution of each mechanism / biological feature in final performance.\n\nA major weakness is the adoption of fairly “easy” tasks, since performance is almost at ceiling in most task variants (the only challenging one seems to be the Seq-MNIST). This might not allow to fully appreciate the gain associated with each model variants (e.g., Hebbian Update in Tab. 1). The latter issue is also amplified by the small number of runs (n=3) for each condition: increasing this number to at least 5 or 10 would allow to better clarify the actual gains, at least qualitatively; increasing it even more (maybe in some representative cases) would further allow to verify the gains through statistical testing. It would also be interesting to consider the full “Bio-ANN” model without Hebbian Update, to check whether the overall performance gain in the combined model really depends on the inclusion of such principle. Moreover, it might be useful to compare the proposed model with some other state-of-the-art approaches in continual learning to better highlight its potential (and possibly its limits).", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors introduce a biologically inspired neural network architecture that combines a variety of learning and processing mechanisms, with the goal of studying whether the performance of artificial systems might be improved by increasing their computational realism. In particular, the authors explore the implementation of Dale’s principle, Hebbian updates, sparse coding, heterogeneous dropout, synaptic consolidation, and experience replay. The proposed model is tested in continual learning (CL) tasks derived from the popular MNIST benchmark, and results suggest significant performance gains.", "strength_and_weaknesses": "One of the strengths of this paper is that it combines a remarkable number of recent ideas and biological processing mechanisms into a unified architecture. The proposed model is also tested in an incremental fashion, which allows to better understand the contribution of each mechanism / biological feature in final performance.\n\nA major weakness is the adoption of fairly “easy” tasks, since performance is almost at ceiling in most task variants (the only challenging one seems to be the Seq-MNIST). This might not allow to fully appreciate the gain associated with each model variants (e.g., Hebbian Update in Tab. 1). The latter issue is also amplified by the small number of runs (n=3) for each condition: increasing this number to at least 5 or 10 would allow to better clarify the actual gains, at least qualitatively; increasing it even more (maybe in some representative cases) would further allow to verify the gains through statistical testing. It would also be interesting to consider the full “Bio-ANN” model without Hebbian Update, to check whether the overall performance gain in the combined model really depends on the inclusion of such principle. Moreover, it might be useful to compare the proposed model with some other state-of-the-art approaches in continual learning to better highlight its potential (and possibly its limits).", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is written in a very clear way. The overall framework is properly motivated, the related work is (to the best of my knowledge) appropriately referenced and the results are presented in a comprehensible way.\n\nQuality: The results are interesting, and the methods seem appropriate. As stated above, I would suggest increasing the number of runs for each condition to better establish the statistical significance of the performance gains, and possibly include the comparison with an alternative CL approach and/or a more challenging CL task to more robustly validate the model.\n\nNovelty: The proposed architecture is original since it combines many recent (or even very-recent) ideas into a unified framework.\n\nReproducibility: I think the results are reproducible.", "summary_of_the_review": "Overall, my judgment of this paper is positive. The research scope of this paper is broad enough to be of interest to the audience at ICLR and stimulate further interdisciplinary research work. At the same time, I believe that further simulations are required to make the paper stronger.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666344464232}], "openreview_url": "https://openreview.net/forum?id=9Zx6tTcX0SE", "arxiv_id": "2304.06738", "paper_pdf": "papers/9Zx6tTcX0SE.pdf", "paper_pdf_sha256": "924b958ff149365651e83e4e5a9a67f1d0a2eb04b7582e700b13f2fca79d4507", "paper_pdf_bytes": 407538, "paper_pdf_source": "openreview", "code_url": "https://github.com/NeurAI-Lab/Bio-ANN", "code_repository": "NeurAI-Lab/Bio-ANN", "code_commit": "267d48d0651fafbff0acabee7afd4ccbacfc0e65", "code_archive": "repos/9Zx6tTcX0SE.zip", "code_archive_sha256": "779fef27fe0dc103b01964e7993fc9e8b48180f92bd9ca83e858e01e6d3672af", "code_archive_bytes": 227291, "code_file_count": 109, "code_extensions": {".py": 109}, "github_disk_usage_kb": 155, "github_languages": {"Python": 671006}, "github_archived": false, "github_pushed_at": "2023-10-11T14:06:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-study-of-biologically-plausible-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Xg47v73CDaj", "year": 2022, "status": "rejected", "title": "Non-deep Networks", "authors": ["Ankit Goyal", "Alexey Bochkovskiy", "Jia Deng", "Vladlen Koltun"], "authorids": ["~Ankit_Goyal1", "alexeyab84@gmail.com", "~Jia_Deng1", "~Vladlen_Koltun1"], "authors_source": "OpenReview API", "abstract": "Depth is the hallmark of deep neural networks. But more depth means more sequential computation and higher latency. This begs the question -- is it possible to build high-performing ``non-deep\" neural networks? We show that it is. To do so, we use parallel subnetworks instead of stacking one layer after another. This helps effectively reduce depth while maintaining high performance. By utilizing parallel substructures, we show, for the first time, that a network with a depth of just 12 can achieve top-1 accuracy over 80% on ImageNet, 96% on CIFAR10, and 81% on CIFAR100. We also show that a network with a low-depth (12) backbone can achieve an AP of 48% on MS-COCO. We analyze the scaling rules for our design and show how to increase performance without changing the network's depth. Finally, we provide a proof of concept for how non-deep networks could be used to build low-latency recognition systems. We will open-source our code.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "UpGm5fyJKpK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper173/Reviewer_wcwX"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to manually design a new 12-depth CNN architecture ParNet based on parallel subnetworks instead of traditionally deeply stacked blocks. Experiments show that ParNet is the first CNN achieving over 80% accuracy on ImageNet with 12 depth only. ParNet also achieves a competitive AP of 48% on MS-COCO for object detection.", "review_text": "1) My major concern is that the authors argues contribution of parallel subnetworks, while I think the traditional RepVGG-wise blocks adopted in ParNet contributes significantly more to prediction accuracy. Even if ParNet adopts 1 branch only, ParNet's accuracy can still achieve 75% more, as \"ParNet-M-OneStream\" shown in Table 8. Comparison of Table 7 and Table 10 shows that the RepVGG blocks is significantly more important than adopted number of branches on improving accuracy. \n\n2) As shown in Table 2, ParNet-L/ParNet-XL (12-depth) indeed has fewer depth compared to ResNet50 (50-depth). However, ParNet-L (54.9M)/ParNet-XL (85.0M) has significantly more parameters compared to ResNet50 (25.6M). A potential problem is that if ParNet runs on low-power computing hardware with limited cuda cores, I am not sure whether ParNet still has the advantage on speed compared to ResNet50.\n\n3) The paper proposes a new block RepVGG-SSE by introducing SSE into traditional RepVGG, while ParNet's accuracy improvement from SSE is only 1.53%, as shown in Table 7. Therefore, I think contribution of RepVGG SSE is not sufficient enough.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to manually design a new 12-depth CNN architecture ParNet based on parallel subnetworks instead of traditionally deeply stacked blocks. Experiments show that ParNet is the first CNN achieving over 80% accuracy on ImageNet with 12 depth only. ParNet also achieves a competitive AP of 48% on MS-COCO for object detection.", "main_review": "1) My major concern is that the authors argues contribution of parallel subnetworks, while I think the traditional RepVGG-wise blocks adopted in ParNet contributes significantly more to prediction accuracy. Even if ParNet adopts 1 branch only, ParNet's accuracy can still achieve 75% more, as \"ParNet-M-OneStream\" shown in Table 8. Comparison of Table 7 and Table 10 shows that the RepVGG blocks is significantly more important than adopted number of branches on improving accuracy. \n\n2) As shown in Table 2, ParNet-L/ParNet-XL (12-depth) indeed has fewer depth compared to ResNet50 (50-depth). However, ParNet-L (54.9M)/ParNet-XL (85.0M) has significantly more parameters compared to ResNet50 (25.6M). A potential problem is that if ParNet runs on low-power computing hardware with limited cuda cores, I am not sure whether ParNet still has the advantage on speed compared to ResNet50.\n\n3) The paper proposes a new block RepVGG-SSE by introducing SSE into traditional RepVGG, while ParNet's accuracy improvement from SSE is only 1.53%, as shown in Table 7. Therefore, I think contribution of RepVGG SSE is not sufficient enough.", "summary_of_the_review": "The authors argues contribution of parallel subnetworks, while the experiments show that the traditional RepVGG-wise blocks adopted in ParNet contributes significantly more to prediction accuracy. Therefor, my rating is \"5: marginally below the acceptance threshold\".\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636033315561}, {"id": "4UU6UoEazZs", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper173/Reviewer_Z9wZ"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "- **Motivation**. The paper argues that deep networks have several limitations\n  - (a) deep nets have a higher latency;\n  - (b) deep nets are hard to parallelize;\n  - (c) deep nets are not suitable for applications.\n\n\n- **Method**.\nMotivated by these observations, the paper aims to fill the performance gap between shallow networks and deep networks. The paper proposed a 12-layer shallow model framework, which contains proposed RepVGG-SSE blocks, fusion modules for multi-scale processing, and parallel streams.  \n\n\n- **Experiments**.\nThe proposed model, ParNet, is verified on CIFAR-10, CIFAR-100, and ImageNet for classification, MS-COCO for detection. ", "review_text": "I have four considerations:\n- (a) Table 2 only shows the details for ParNet-L and -XL. It could be better to show the full details and comparisons of ParNet-S, -M, -L, and -XL.\n- (b) What are the details of speed testing in Table 2? Do the ResNet and ParNet use the same setting for speed benchmark? If ParNet uses multiple GPUs in parallel to benchmark the inference speed, e.g., 2 GPUs with 128 global batch size, does the ResNet uses 2 GPUs within the global batch size of 128 as well? It could be better to add more words in the third paragraph on Page 6 to show how to conduct the experiments. \n- (c) ParNet develops the RepVGG-SSE based on the work of RepVGG, CVPR 2021. But there is no comparison with this high-related work. It could be better to compare with RepVGG in Table 1 and Table 2.\n- (d) The paper argues that one of the advantages is parallel. But according to Table 2 and RepVGG's Table 4, ParNet has more parameters and larger FLOPs than the competitors. More parameters mean that the model will cost more memory during training and inference. This fact fades the significance of ParNet. It could be better to discuss this and highlight why parallelism is more important than the other two aspects for selling the work.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "- **Motivation**. The paper argues that deep networks have several limitations\n  - (a) deep nets have a higher latency;\n  - (b) deep nets are hard to parallelize;\n  - (c) deep nets are not suitable for applications.\n\n\n- **Method**.\nMotivated by these observations, the paper aims to fill the performance gap between shallow networks and deep networks. The paper proposed a 12-layer shallow model framework, which contains proposed RepVGG-SSE blocks, fusion modules for multi-scale processing, and parallel streams.  \n\n\n- **Experiments**.\nThe proposed model, ParNet, is verified on CIFAR-10, CIFAR-100, and ImageNet for classification, MS-COCO for detection. ", "main_review": "I have four considerations:\n- (a) Table 2 only shows the details for ParNet-L and -XL. It could be better to show the full details and comparisons of ParNet-S, -M, -L, and -XL.\n- (b) What are the details of speed testing in Table 2? Do the ResNet and ParNet use the same setting for speed benchmark? If ParNet uses multiple GPUs in parallel to benchmark the inference speed, e.g., 2 GPUs with 128 global batch size, does the ResNet uses 2 GPUs within the global batch size of 128 as well? It could be better to add more words in the third paragraph on Page 6 to show how to conduct the experiments. \n- (c) ParNet develops the RepVGG-SSE based on the work of RepVGG, CVPR 2021. But there is no comparison with this high-related work. It could be better to compare with RepVGG in Table 1 and Table 2.\n- (d) The paper argues that one of the advantages is parallel. But according to Table 2 and RepVGG's Table 4, ParNet has more parameters and larger FLOPs than the competitors. More parameters mean that the model will cost more memory during training and inference. This fact fades the significance of ParNet. It could be better to discuss this and highlight why parallelism is more important than the other two aspects for selling the work.", "summary_of_the_review": "Please erase the above concerns. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635914309693}, {"id": "YW6Ua1IVn6K", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper173/Reviewer_pTPq"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a way of designing a network fixing the depth yet involving more branches. The authors try to support the claim which is that low-depth networks can achieve a good result by providing experiments on the ImageNet and the CIFAR datasets", "review_text": "Pros)\n+ This paper is easy to follow.\n+ The concept of branch-parallelism in a model looks promising.\n\nCons)\n- I am agreeing that the network depth is a crucial design element towards model efficiency but the computational costs in including the number of parameters, FLOPs, and memory footprints seem to be overlooked in this paper: the proposed model has a larger # of parameters and FLOPs, so the overall computational costs are bigger than the baseline models. \n- The architectural design with the proposed building block in Figure 2 is presented without providing any intuitions or design philosophy.\n- The comparisons with the baselines in model-parallelism is not sufficient\n\nSome comments and questions)\n- The model-parallelism can be done with a single-branch model in training or inferring with multiple images (i.e., with a larger mini-batch size), then the speed gain from the branch-parallelism scheme may vanish. Please clarify a multi-branch architecture also has an advantage over the widened networks such as WideResNet in this case.\n\n- The performance comparison should be done fairly. For example in Table 2, ParNet-L is compared with much smaller networks such as R34 and R50 in terms of the computational budgets, so this is not fair. The models with widening the width while fixing depth (e.g., Wide ResNet) can achieve much better accuracy barely increase the latency.\n\n- Memory consumption would be a matter for the proposed architecture. Please specify the memory footprints when training (or inference) with the fixed batchsize.\n\n- ResNets can also be leveraged fusing methods including skip connections and BNs. Did the authors compare the latency with these fused ResNets with the proposed ParNets in Table 2?\n\n- Why vanilla SE-block is not suitable for the proposed model? Why Rep-VGG block has been adopted?\n\n- Please specify why ParNets cannot outperform DenseNet-100 even using more parameters on the CIFAR datasets in Table 6. I am just wondering whether there is a different earning behavior of a shallow network on a particular dataset.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a way of designing a network fixing the depth yet involving more branches. The authors try to support the claim which is that low-depth networks can achieve a good result by providing experiments on the ImageNet and the CIFAR datasets", "main_review": "Pros)\n+ This paper is easy to follow.\n+ The concept of branch-parallelism in a model looks promising.\n\nCons)\n- I am agreeing that the network depth is a crucial design element towards model efficiency but the computational costs in including the number of parameters, FLOPs, and memory footprints seem to be overlooked in this paper: the proposed model has a larger # of parameters and FLOPs, so the overall computational costs are bigger than the baseline models. \n- The architectural design with the proposed building block in Figure 2 is presented without providing any intuitions or design philosophy.\n- The comparisons with the baselines in model-parallelism is not sufficient\n\nSome comments and questions)\n- The model-parallelism can be done with a single-branch model in training or inferring with multiple images (i.e., with a larger mini-batch size), then the speed gain from the branch-parallelism scheme may vanish. Please clarify a multi-branch architecture also has an advantage over the widened networks such as WideResNet in this case.\n\n- The performance comparison should be done fairly. For example in Table 2, ParNet-L is compared with much smaller networks such as R34 and R50 in terms of the computational budgets, so this is not fair. The models with widening the width while fixing depth (e.g., Wide ResNet) can achieve much better accuracy barely increase the latency.\n\n- Memory consumption would be a matter for the proposed architecture. Please specify the memory footprints when training (or inference) with the fixed batchsize.\n\n- ResNets can also be leveraged fusing methods including skip connections and BNs. Did the authors compare the latency with these fused ResNets with the proposed ParNets in Table 2?\n\n- Why vanilla SE-block is not suitable for the proposed model? Why Rep-VGG block has been adopted?\n\n- Please specify why ParNets cannot outperform DenseNet-100 even using more parameters on the CIFAR datasets in Table 6. I am just wondering whether there is a different earning behavior of a shallow network on a particular dataset.", "summary_of_the_review": "- See the main review", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635842241963}, {"id": "ZIXGSpyZN-m", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper173/Reviewer_75Wm"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a new architecture of a convolutional neural network for image classification. The authors are motivated by inference efficiency and showing that networks with about a dozen of layers can be competitive in classification accuracy with 50 layers and more. So instead of growing depth, they propose to grow width and have multiple subnetworks of the same depth within one model. The proposed architecture is evaluated on CIFAR, ImageNet and COCO classification and detection tasks, and shows that networks with 12 layers can be competitive with deeper counterparts.", "review_text": "Strength:\n- empirical result that low-depth networks can be competitive with deeper models is significant. Confirming observations in prior works, the number of parameters and computational complexity of low-depth networks need to be much higher to match the accuracy of deeper networks.\n\nWeaknesses:\n- architecture complexity: the authors claim outperforming ResNet in efficiency, but do not mention that the architecture is far more complex. It is also more complex than RepVGG. This needs to be stressed in the introduction. This is separate from computational complexity.\n- weak baseline: authors use Squeeze-and-Excitation blocks in their architecture which is known to improve ResNet performance, yet compare to ResNet without SE layers (section 4, tables 1-2)\n- missing baseline: the authors introduce parallel streams in their network, which might be working as an ensemble, known to be improving classification performance on the tested datasets. An important baseline then is treating the stream subnetworks as separate models applied on different resolutions, which is simpler than the proposed architecture and would not suffer from communication overhead. A baseline with ResNet and ParNet ensembles needs to be added.\n- statistical significance: test errors on CIFAR and ImageNet have high variance, and the authors show results of a single run for each experiment. A common practice to reduce variance is to train multiple networks and report mean and standard deviation for each experiment.\n- SiLU: the authors replace ReLU with SiLU motivated by the limited representational power of ReLU. However, SiLU was introduced to improve training of deep networks, which is not the case in the proposed architecture. Moreover, in the ablation study (table 7) it does not seem to bring a significant improvement over ReLU, there is also no statistical significance analysis of this result. Using SiLU seems like a needless complication of the proposed architecture.\n\nArguable weakness:\n- definition of depth: the approach takes advantage of the vague definition of depth in neural networks. In prior works, depth is defined as a number of layers in a single stream network. By introducing multiple parallel streams the authors effectively take a deep network and move layers around, so another, perhaps more fair, definition of depth could be counting the total number of layers in all substreams.\n\nNotes to authors:\n- table 3 is misleading because it does not show the number of GPUs used in the last row. It is also not clear if it is fused or not.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new architecture of a convolutional neural network for image classification. The authors are motivated by inference efficiency and showing that networks with about a dozen of layers can be competitive in classification accuracy with 50 layers and more. So instead of growing depth, they propose to grow width and have multiple subnetworks of the same depth within one model. The proposed architecture is evaluated on CIFAR, ImageNet and COCO classification and detection tasks, and shows that networks with 12 layers can be competitive with deeper counterparts.", "main_review": "Strength:\n- empirical result that low-depth networks can be competitive with deeper models is significant. Confirming observations in prior works, the number of parameters and computational complexity of low-depth networks need to be much higher to match the accuracy of deeper networks.\n\nWeaknesses:\n- architecture complexity: the authors claim outperforming ResNet in efficiency, but do not mention that the architecture is far more complex. It is also more complex than RepVGG. This needs to be stressed in the introduction. This is separate from computational complexity.\n- weak baseline: authors use Squeeze-and-Excitation blocks in their architecture which is known to improve ResNet performance, yet compare to ResNet without SE layers (section 4, tables 1-2)\n- missing baseline: the authors introduce parallel streams in their network, which might be working as an ensemble, known to be improving classification performance on the tested datasets. An important baseline then is treating the stream subnetworks as separate models applied on different resolutions, which is simpler than the proposed architecture and would not suffer from communication overhead. A baseline with ResNet and ParNet ensembles needs to be added.\n- statistical significance: test errors on CIFAR and ImageNet have high variance, and the authors show results of a single run for each experiment. A common practice to reduce variance is to train multiple networks and report mean and standard deviation for each experiment.\n- SiLU: the authors replace ReLU with SiLU motivated by the limited representational power of ReLU. However, SiLU was introduced to improve training of deep networks, which is not the case in the proposed architecture. Moreover, in the ablation study (table 7) it does not seem to bring a significant improvement over ReLU, there is also no statistical significance analysis of this result. Using SiLU seems like a needless complication of the proposed architecture.\n\nArguable weakness:\n- definition of depth: the approach takes advantage of the vague definition of depth in neural networks. In prior works, depth is defined as a number of layers in a single stream network. By introducing multiple parallel streams the authors effectively take a deep network and move layers around, so another, perhaps more fair, definition of depth could be counting the total number of layers in all substreams.\n\nNotes to authors:\n- table 3 is misleading because it does not show the number of GPUs used in the last row. It is also not clear if it is fused or not.", "summary_of_the_review": "The paper shows a significant result that a network with 12 layers can be competitive with deeper networks on well established image classification benchmarks. However, the architecture is more complex than ResNet or RepVGG, and the empirical evaluation is unsatisfactory due to weak and missing baselines, and the lack of statistical significance analysis. I hesitate between reject and weak reject.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635689068717}], "openreview_url": "https://openreview.net/forum?id=Xg47v73CDaj", "arxiv_id": "2110.07641", "paper_pdf": "papers/Xg47v73CDaj.pdf", "paper_pdf_sha256": "3b7fe8bfe02cbf865c5c773cc27ee99295f592355283dc55950c4677c5e08e74", "paper_pdf_bytes": 312702, "paper_pdf_source": "openreview", "code_url": "https://github.com/imankgoyal/NonDeepNetworks", "code_repository": "imankgoyal/NonDeepNetworks", "code_commit": "777b286c2d6045625b1d37099241034985a1404e", "code_archive": "repos/Xg47v73CDaj.zip", "code_archive_sha256": "c2038fcd347ae8fa81d17811feb364adbaa272827d7d1e0be05985bbf7638574", "code_archive_bytes": 280351, "code_file_count": 116, "code_extensions": {".py": 111, ".sh": 5}, "github_disk_usage_kb": 258, "github_languages": {"Python": 548385, "Shell": 9236}, "github_archived": false, "github_pushed_at": "2022-10-14T16:58:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/non-deep-networks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TmkN9JmDJx1", "year": 2021, "status": "rejected", "title": "Thinking Like Transformers", "authors": ["Gail Weiss", "Yoav Goldberg", "Eran Yahav"], "authorids": ["~Gail_Weiss1", "~Yoav_Goldberg1", "~Eran_Yahav1"], "authors_source": "OpenReview API", "abstract": "What is the computational model behind a transformer? Where recurrent neural networks have direct parallels in finite state machines, allowing clear discussion and thought around architecture variants or trained models, transformers have no such familiar parallel. In this paper we aim to change that, proposing a computational model for the transformer-encoder in the form of a programming language. We map the basic components of a transformer-encoder – attention and feed-forward computation – into the simple primitives of select, aggregate, and zipmap, around which we form a programming language: the Restricted Access Sequence Process-ing Language (RASP). We show how RASP can be used to program solutions to tasks that could conceivably be learned by a transformer, augmenting it with tools we discover in our work. In particular, we provide RASP programs for histograms, sorting, and even logical inference similar to that of Clark et al. (2020). We further use our model to relate their difficulty in terms of the number of required layers and attention heads. Finally, we see how insights gained from our abstraction might be used to explain phenomena seen in recent works.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "_EEVUfUO4s", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3209/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a computational model for the transformer in the form of a sequence processing programming language named Restricted Access Sequence Processing Language (RASP). The paper shows how RASP can be used to program solutions to tasks that could conceivably be learned by a transformer. The paper argues that considering computational problems and their implementation in the RASP language allows people to \"think like a transformer\" in the style of symbolic programs. Overall, the paper is well written and easy to follow.\n\nReasons to accept the paper:\n1. The paper provides a novel way of understanding how transformer model works from a programming language perspective.\n2. The paper presents solutions in RASP language for simple tasks such as histograms and sorting, and also complicated logic inference task.\n3. The paper attempts to build a connection between the operations in RASP language and the computational operations in transformers. This could help analyze the minimally required number of layers and upper-bound number of heads for the transformer to work on a specific task.\n\nReasons to reject the paper:\n1. For general neural network models, which has no explicit attention mechanism but may still be able to learn to reason over various tasks, it is not clear whether the RASP language is still an abstraction. The paper only discusses transformers, but there is no evidence showing that the operations in RASP cannot be completed by a simple multi-layer neural network. In other words, the connection between RASP and transformer may not be unique, and we may use RASP to think like any neural networks.\n2. It is not clear whether there exists other forms of programming language that can also \"explain\" how transformer works, and if so, how the presented one (RASP) is a better abstraction of the transformer model.\n3. Although the presented RASP language can help analyze the number of layers and heads required theoretically, there is limited value and insights for improving existing transformers models.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "This paper proposes a computational model for the transformer in the form of a sequence processing programming language named Restricted Access Sequence Processing Language (RASP). The paper shows how RASP can be used to program solutions to tasks that could conceivably be learned by a transformer. The paper argues that considering computational problems and their implementation in the RASP language allows people to \"think like a transformer\" in the style of symbolic programs. Overall, the paper is well written and easy to follow.\n\nReasons to accept the paper:\n1. The paper provides a novel way of understanding how transformer model works from a programming language perspective.\n2. The paper presents solutions in RASP language for simple tasks such as histograms and sorting, and also complicated logic inference task.\n3. The paper attempts to build a connection between the operations in RASP language and the computational operations in transformers. This could help analyze the minimally required number of layers and upper-bound number of heads for the transformer to work on a specific task.\n\nReasons to reject the paper:\n1. For general neural network models, which has no explicit attention mechanism but may still be able to learn to reason over various tasks, it is not clear whether the RASP language is still an abstraction. The paper only discusses transformers, but there is no evidence showing that the operations in RASP cannot be completed by a simple multi-layer neural network. In other words, the connection between RASP and transformer may not be unique, and we may use RASP to think like any neural networks.\n2. It is not clear whether there exists other forms of programming language that can also \"explain\" how transformer works, and if so, how the presented one (RASP) is a better abstraction of the transformer model.\n3. Although the presented RASP language can help analyze the number of layers and heads required theoretically, there is limited value and insights for improving existing transformers models.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604009239946}, {"id": "nun4TbwCQxr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3209/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a programming language, RASP, as a computational model for transformer encoders, and discusses how analysis in terms of this language could be used to understand the behavior of transformer models.\n\nThe idea of finding a computational model for transformers is interesting, and (as discussed in section 4) could lead to insights in terms of how to build better models.\n\nHowever, this paper lacks any results or experimental analysis, which makes it difficult to judge the validity or value of the claims presented. Section 4 discusses how recently proposed transformer variants could be understood (post-hoc) in terms of the RASP language. However, in order to justify using the RASP language to reason about transformers, I think it is necessary to demonstrate experimentally that insights from RASP can translate to new empirical findings. \n\nFor example, in section 3.1, the paper makes the claim, “For any given RASP program, we can compute the minimal number of layers required to implement it in a transformer, and upper bound the number of heads this implementation requires.” Can this be verified experimentally, by building a synthetic task, and testing performance as the number of heads is varied?\n\nSimilarly, section 4.2 provides an analysis of the recently proposed sandwich transformer model. Could similar analysis be used to make claims about novel, untested architecture variants? Could these claims be verified experimentally? Results such as this would be of high value to the ICLR community.\n\nBecause of the lack of experiments, I recommend rejection. I think this is an interesting line of work which could prove valuable to the ICLR community if supported by rigorous experimental evidence.\n\nMinor details:\npg 1: “that is requires” -> “that is required”", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea without experimental analysis", "review": "This paper proposes a programming language, RASP, as a computational model for transformer encoders, and discusses how analysis in terms of this language could be used to understand the behavior of transformer models.\n\nThe idea of finding a computational model for transformers is interesting, and (as discussed in section 4) could lead to insights in terms of how to build better models.\n\nHowever, this paper lacks any results or experimental analysis, which makes it difficult to judge the validity or value of the claims presented. Section 4 discusses how recently proposed transformer variants could be understood (post-hoc) in terms of the RASP language. However, in order to justify using the RASP language to reason about transformers, I think it is necessary to demonstrate experimentally that insights from RASP can translate to new empirical findings. \n\nFor example, in section 3.1, the paper makes the claim, “For any given RASP program, we can compute the minimal number of layers required to implement it in a transformer, and upper bound the number of heads this implementation requires.” Can this be verified experimentally, by building a synthetic task, and testing performance as the number of heads is varied?\n\nSimilarly, section 4.2 provides an analysis of the recently proposed sandwich transformer model. Could similar analysis be used to make claims about novel, untested architecture variants? Could these claims be verified experimentally? Results such as this would be of high value to the ICLR community.\n\nBecause of the lack of experiments, I recommend rejection. I think this is an interesting line of work which could prove valuable to the ICLR community if supported by rigorous experimental evidence.\n\nMinor details:\npg 1: “that is requires” -> “that is required”", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603904568204}, {"id": "CrGxcfjTUWE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3209/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a restricted programming language containing rough analogues of operations used in transformers. Using this language, the authors show how some algorithms can be implemented, which gives some insights about the limitations of transformers.\n\nOverall, the paper is badly written, with many typos and many hard-to-understand areas. Since this is a \"thought-experiment\" paper, this alone is a good reason for rejecting this work at its current form. \n\nSome things I would have expected an analysis on:\n* How well do the individual RASP operations map to the relevant transformer operations?\n* Why are the examples in 2.1 important/useful? Where are they needed?\n* What are \"useful\" algorithms that cannot be represented in RASP? How would we need to change a transformer to allow it to represent such algorithms?\n*   Sec 2.2 is a textual description of a complicated algorithm, why not build it gradually from primitives by introducing larger functions? Experimental results on RASP/Transformers?\n* Can you \"compile\" RASP into a transformer (=architecture+weights) that performs exactly the task defined in RASP? (If not, why? if yes, some experimental validation would be useful)\n\n\n\n##### Typos\n* Fig1, Line 8: should the second arg be `vals`?\n* Sec2: \"the base sequences\": unbalanced parenthesis\n* In the discussion of `aggregate` a selector `s` is an input, but `s` is never used. Should `s` be `f` instead ?\n* In the definition of `select`, `s(i,j)=f(m1[i],...,mk[i], ot1[j], ... otl[j])` What does `m1[i]` mean? `m1` is the first element of `me` but `m1` is also a sequence somehow?\n* Footnote 1 says uses variable `n`. What is `n`? Should it be `max(k,l)`?\n* Fig3 the semantics of the operation in L2 have not been defined, similarly for Fig2 L7.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting concept badly presented and evaluated", "review": "This paper proposes a restricted programming language containing rough analogues of operations used in transformers. Using this language, the authors show how some algorithms can be implemented, which gives some insights about the limitations of transformers.\n\nOverall, the paper is badly written, with many typos and many hard-to-understand areas. Since this is a \"thought-experiment\" paper, this alone is a good reason for rejecting this work at its current form. \n\nSome things I would have expected an analysis on:\n* How well do the individual RASP operations map to the relevant transformer operations?\n* Why are the examples in 2.1 important/useful? Where are they needed?\n* What are \"useful\" algorithms that cannot be represented in RASP? How would we need to change a transformer to allow it to represent such algorithms?\n*   Sec 2.2 is a textual description of a complicated algorithm, why not build it gradually from primitives by introducing larger functions? Experimental results on RASP/Transformers?\n* Can you \"compile\" RASP into a transformer (=architecture+weights) that performs exactly the task defined in RASP? (If not, why? if yes, some experimental validation would be useful)\n\n\n\n##### Typos\n* Fig1, Line 8: should the second arg be `vals`?\n* Sec2: \"the base sequences\": unbalanced parenthesis\n* In the discussion of `aggregate` a selector `s` is an input, but `s` is never used. Should `s` be `f` instead ?\n* In the definition of `select`, `s(i,j)=f(m1[i],...,mk[i], ot1[j], ... otl[j])` What does `m1[i]` mean? `m1` is the first element of `me` but `m1` is also a sequence somehow?\n* Footnote 1 says uses variable `n`. What is `n`? Should it be `max(k,l)`?\n* Fig3 the semantics of the operation in L2 have not been defined, similarly for Fig2 L7.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603281653419}, {"id": "nXDj_IJqITL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3209/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors introduce a DSL, the Restricted Access Sequence Processing (RASP) language, that they claim can serve as a computational model for the transformer-encoder.  They develop the reader's intuition for RASP by providing RASP implementations of many basic operations such as computing histograms, sorting, and reversing.  They also show how, for a given RASP program, to determine the minimum number of layers required and to upper-bound the number of heads required to implement it as a transformer.  Lastly, they analyze two transformer variants, restricted-attention transformers and sandwich transformers.  For the former, they use the RASP perspective to claim a theoretical limitation, and for the latter, they comment that a known empirical finding is intuitive in light of the RASP perspective.\n\nI found this paper very interesting, a rare conceptual gem in a mostly empirical field.  The ideas in the paper open up many new questions and directions.  I wish I could champion it but sadly I find it critically underdeveloped in its current state and not yet ready for publication.\n\nThe main weakness of the current version is that it only glosses over the connection between RASP and the transformer-encoder.  It does not explain what it means to be a computational model for it, and does not provide general principles for abstracting DSLs from network architectures nor discuss the possible design space.  Section 3 says \"While we give a detailed explanation [of the relationship between RASP and transformers] in the appendix...\" but the appendix contains no such analysis, only more details of RASP in isolation.  I also find it strange and somewhat of a red-flag that RASP does not seem to be compositional, for example, it does not seem to include a computational model of feed-forward networks within it.  The `zipmap` operation nominally corresponds to the feed-forward stage of the transformer-encoder, but RASP does not seem to include any restrictions on the function being zip-mapped.\n\nI think the paper could be strengthened significantly if it began with a DSL for building computation graphs sufficient to express transformers, and then presented a compositional (if not fully principled) way of abstracting this DSL into a traditional DSL that can serve as a reference computational model for the original version.  Ideally this process generalizes the two prior computational models they refer to, i.e. CNNs as sequences of filters and RNNs as state machines.  I think it is also critical to make explicit what properties are being relaxed and what differences are being abstracted away.  For example, would the authors consider RNNs and LSTMs to have the same reference computational model or would differences be preserved in the abstraction? Are there RASP programs that cannot be realized by transformers, or vice-versa? More generally, what does it mean for one language to be a computational model of another one, and what does the design space look like?\n\nIf the main weakness is that the authors do not ground their computational model to transformers theoretically, a related weakness is that they do not ground their computational model to transformers empirically either.  I think the paper could also be strengthened significantly by simple empirical experiments, for example using their RASP implementations of various simple functions to make predictions about the accuracy of transformers trained on those tasks as a function of the number of layers and heads they are provided.  I would be particularly interested to see how sharp these curves are. Is there a phase transition once the minimum required layers/heads are provided, or is it much more gradual?  Is there any evidence that the transformer actually learns these reference programs?\n\nMiscellaneous comments:\n- I found the description of RASP unnecessarily difficult to follow. There are conventions for introducing DSLs, e.g. presenting the grammar and then the semantics.  There are also some inconsistencies, with 'aggregate' introduced as taking two arguments but then used in the prose taking a mysterious third lambda argument.\n- It is not immediately obvious what it means for one RASP function to call another as a subroutine.  Is it assumed that any such subroutine has one distinguished input that must always have the same size as the distinguished input to the original RASP function (so that `indices` and `length` are the same)?\n- I think the inclusion of non-float types merits more discussion. For example, is it important that there is a boolean type?\n- The section on logic programming is too informal, with phrases like \"We suggest approaching this task in RASP as follows...\" and \"If a trained transformer ..., this may explain ...\".  Does RASP permit one or more decision procedures for horn clauses? If so, what is the code?  Do you hypothesize that the transformers in (Clark et al. 2020) are learning hybrid forwards/backwards reasoning? If so, how might you test this hypothesis?\n- The impossibility result for sorting should discuss non-comparison-based sorting algorithms (e.g. radix sort) or else qualify the claim.\n- The RASP analysis of the sandwich transformer results (S4.2) does not seem particularly illuminating.\n- There is essentially no discussion of related work.\n\nMinor:\n- the second sentence of the intro repeats \"language\"\n- in 'The Base Sequences' paragraph, there is a double comma and a dangling close-paren\n- the `sort` code uses `seq` instead of `vals` in the last line\n- in explanation of `select`, `f` is referred to as a \"selection function\" even though it has a different type\n- top of page 4: attention distribution <MISSING PERIOD> Hence\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Very interesting but underdeveloped", "review": "The authors introduce a DSL, the Restricted Access Sequence Processing (RASP) language, that they claim can serve as a computational model for the transformer-encoder.  They develop the reader's intuition for RASP by providing RASP implementations of many basic operations such as computing histograms, sorting, and reversing.  They also show how, for a given RASP program, to determine the minimum number of layers required and to upper-bound the number of heads required to implement it as a transformer.  Lastly, they analyze two transformer variants, restricted-attention transformers and sandwich transformers.  For the former, they use the RASP perspective to claim a theoretical limitation, and for the latter, they comment that a known empirical finding is intuitive in light of the RASP perspective.\n\nI found this paper very interesting, a rare conceptual gem in a mostly empirical field.  The ideas in the paper open up many new questions and directions.  I wish I could champion it but sadly I find it critically underdeveloped in its current state and not yet ready for publication.\n\nThe main weakness of the current version is that it only glosses over the connection between RASP and the transformer-encoder.  It does not explain what it means to be a computational model for it, and does not provide general principles for abstracting DSLs from network architectures nor discuss the possible design space.  Section 3 says \"While we give a detailed explanation [of the relationship between RASP and transformers] in the appendix...\" but the appendix contains no such analysis, only more details of RASP in isolation.  I also find it strange and somewhat of a red-flag that RASP does not seem to be compositional, for example, it does not seem to include a computational model of feed-forward networks within it.  The `zipmap` operation nominally corresponds to the feed-forward stage of the transformer-encoder, but RASP does not seem to include any restrictions on the function being zip-mapped.\n\nI think the paper could be strengthened significantly if it began with a DSL for building computation graphs sufficient to express transformers, and then presented a compositional (if not fully principled) way of abstracting this DSL into a traditional DSL that can serve as a reference computational model for the original version.  Ideally this process generalizes the two prior computational models they refer to, i.e. CNNs as sequences of filters and RNNs as state machines.  I think it is also critical to make explicit what properties are being relaxed and what differences are being abstracted away.  For example, would the authors consider RNNs and LSTMs to have the same reference computational model or would differences be preserved in the abstraction? Are there RASP programs that cannot be realized by transformers, or vice-versa? More generally, what does it mean for one language to be a computational model of another one, and what does the design space look like?\n\nIf the main weakness is that the authors do not ground their computational model to transformers theoretically, a related weakness is that they do not ground their computational model to transformers empirically either.  I think the paper could also be strengthened significantly by simple empirical experiments, for example using their RASP implementations of various simple functions to make predictions about the accuracy of transformers trained on those tasks as a function of the number of layers and heads they are provided.  I would be particularly interested to see how sharp these curves are. Is there a phase transition once the minimum required layers/heads are provided, or is it much more gradual?  Is there any evidence that the transformer actually learns these reference programs?\n\nMiscellaneous comments:\n- I found the description of RASP unnecessarily difficult to follow. There are conventions for introducing DSLs, e.g. presenting the grammar and then the semantics.  There are also some inconsistencies, with 'aggregate' introduced as taking two arguments but then used in the prose taking a mysterious third lambda argument.\n- It is not immediately obvious what it means for one RASP function to call another as a subroutine.  Is it assumed that any such subroutine has one distinguished input that must always have the same size as the distinguished input to the original RASP function (so that `indices` and `length` are the same)?\n- I think the inclusion of non-float types merits more discussion. For example, is it important that there is a boolean type?\n- The section on logic programming is too informal, with phrases like \"We suggest approaching this task in RASP as follows...\" and \"If a trained transformer ..., this may explain ...\".  Does RASP permit one or more decision procedures for horn clauses? If so, what is the code?  Do you hypothesize that the transformers in (Clark et al. 2020) are learning hybrid forwards/backwards reasoning? If so, how might you test this hypothesis?\n- The impossibility result for sorting should discuss non-comparison-based sorting algorithms (e.g. radix sort) or else qualify the claim.\n- The RASP analysis of the sandwich transformer results (S4.2) does not seem particularly illuminating.\n- There is essentially no discussion of related work.\n\nMinor:\n- the second sentence of the intro repeats \"language\"\n- in 'The Base Sequences' paragraph, there is a double comma and a dangling close-paren\n- the `sort` code uses `seq` instead of `vals` in the last line\n- in explanation of `select`, `f` is referred to as a \"selection function\" even though it has a different type\n- top of page 4: attention distribution <MISSING PERIOD> Hence\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1602702883383}], "openreview_url": "https://openreview.net/forum?id=TmkN9JmDJx1", "arxiv_id": "2106.06981", "paper_pdf": "papers/TmkN9JmDJx1.pdf", "paper_pdf_sha256": "01970681c267866a0ad467ee2249d20f959d6a5cef03176b5126c13ce9d50ab1", "paper_pdf_bytes": 910713, "paper_pdf_source": "openreview", "code_url": "https://github.com/tech-srl/RASP", "code_repository": "tech-srl/RASP", "code_commit": "61dde62577589b291db39d192de7d481ce97fdae", "code_archive": "repos/TmkN9JmDJx1.zip", "code_archive_sha256": "1a674b57063e6b08ac4d7de1a1c865b97fc74bd5dda685c2d7b60c441f986242", "code_archive_bytes": 238630, "code_file_count": 20, "code_extensions": {".py": 18, ".sh": 2}, "github_disk_usage_kb": 277, "github_languages": {"Python": 279124, "ANTLR": 3568, "Shell": 2028, "Vim Script": 1774}, "github_archived": false, "github_pushed_at": "2024-09-16T12:36:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/thinking-like-transformers-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJlPOCEKvH", "year": 2020, "status": "rejected", "title": "Compressing BERT: Studying the Effects of Weight Pruning on Transfer Learning", "authors": ["Mitchell A Gordon", "Kevin Duh", "Nicholas Andrews"], "authorids": ["mgordo37@jhu.edu", "kevinduh@cs.jhu.edu", "noa@jhu.edu"], "authors_source": "OpenReview API", "abstract": "Universal feature extractors, such as BERT for natural language processing and VGG for computer vision, have become effective methods for improving deep learning models without requiring more labeled data. A common paradigm is to pre-train a feature extractor on large amounts of data then fine-tune it as part of a deep learning model on some downstream task (i.e. transfer learning). While effective, feature extractors like BERT may be prohibitively large for some deployment scenarios. We explore weight pruning for BERT and ask: how does compression during pre-training affect transfer learning? We find that pruning affects transfer learning in three broad regimes. Low levels of pruning (30-40%) do not affect pre-training loss or transfer to downstream tasks at all. Medium levels of pruning increase the pre-training loss and prevent useful pre-training information from being transferred to downstream tasks. High levels of pruning additionally prevent models from fitting downstream datasets, leading to further degradation. Finally, we observe that fine-tuning BERT on a specific task does not improve its prunability. We conclude that BERT can be pruned once during pre-training rather than separately for each task without affecting performance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BJeAGC3pYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1214/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis work explores weight pruning for BERT. It finds that pruning affects transfer learning in three broad regimes. Low levels of pruning (30-40%) do not affect pre-training loss or transfer to downstream tasks at all. Medium levels of\npruning increase the pre-training loss and prevent useful pre-training information from being transferred to downstream tasks. High levels of pruning additionally prevent models from fitting downstream datasets, leading to further degradation.\n\nMy major concern about this work is its technical innovation and value to the community.\n1. This is simply a study of model pruning for BERT. There is nothing new technically.\n\n2. It shows BERT can be pruned for 30-40% parameters. Actually, this is not surprising; instead I'm even disappointed about this result. 30-40% weight reduction does not really speed up inference much or save model size much. Besides, to handle sparse weight matrixes, one may need additional operations to use the pruned models on a modern GPU.\n\n3. Several other submissions show that BERT models can be compressed for 5-10x without accuracy loss. Comparing with this work, this paper seems to tell me that pruning is not suitable for BERT.   ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "\nThis work explores weight pruning for BERT. It finds that pruning affects transfer learning in three broad regimes. Low levels of pruning (30-40%) do not affect pre-training loss or transfer to downstream tasks at all. Medium levels of\npruning increase the pre-training loss and prevent useful pre-training information from being transferred to downstream tasks. High levels of pruning additionally prevent models from fitting downstream datasets, leading to further degradation.\n\nMy major concern about this work is its technical innovation and value to the community.\n1. This is simply a study of model pruning for BERT. There is nothing new technically.\n\n2. It shows BERT can be pruned for 30-40% parameters. Actually, this is not surprising; instead I'm even disappointed about this result. 30-40% weight reduction does not really speed up inference much or save model size much. Besides, to handle sparse weight matrixes, one may need additional operations to use the pruned models on a modern GPU.\n\n3. Several other submissions show that BERT models can be compressed for 5-10x without accuracy loss. Comparing with this work, this paper seems to tell me that pruning is not suitable for BERT.   "}, "tcdate": 1571831317985}, {"id": "H1xSiSF6tr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1214/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work is an empirical study of testing how pruning at the pre-training stage affects subsequent transfer learning (through fine-tuning) stage. The main idea is to carefully control the amount of sparsity injected into BERT through weight magnitude pruning and study the impact on accuracy. The experimental setup is mostly well done, especially the part that disentangles the complexity restriction and information deletion. During the exploration, the authors made several interesting observations, such as 30-40% model weights do not encode any useful inductive bias, which could help shed some light for future work on both training and compressing BERT-like models.\n\nOverall, the paper is well written and explained. The goal is meaningful, and this is a sensible contribution to the ongoing interests of compressing BERT-like large models for efficient training and inference. \n\nMy major concern is on its novelty and how directly it can provide benefit to computation.  First, although the findings are interesting, the methods used in this paper are not new.  Various pruning techniques have been explored in prior work, which makes the novelty contribution of this paper somewhat limited. \n\nFurthermore, the study has mostly focused on the impact of random sparsity to accuracy. However, as it is known that it is really difficult for modern hardware to benefit from random sparsity because it leads to irregular memory accesses, which negatively impact the performance. It has been observed that speedups are very limited or can be negative even the random sparsity is >95% [1]. Therefore, it is hard to judge how inference or training can benefit from 30-40% weight sparsity. Going forward, the authors are encouraged to choose pruning methods that lead to regular memory access to avoid adversely impacting practical acceleration in modern hardware platforms.\n\n[1] Learning Structured Sparsity in Deep Neural Networks. Wen et al. NeurIPS 2016", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This work is an empirical study of testing how pruning at the pre-training stage affects subsequent transfer learning (through fine-tuning) stage. The main idea is to carefully control the amount of sparsity injected into BERT through weight magnitude pruning and study the impact on accuracy. The experimental setup is mostly well done, especially the part that disentangles the complexity restriction and information deletion. During the exploration, the authors made several interesting observations, such as 30-40% model weights do not encode any useful inductive bias, which could help shed some light for future work on both training and compressing BERT-like models.\n\nOverall, the paper is well written and explained. The goal is meaningful, and this is a sensible contribution to the ongoing interests of compressing BERT-like large models for efficient training and inference. \n\nMy major concern is on its novelty and how directly it can provide benefit to computation.  First, although the findings are interesting, the methods used in this paper are not new.  Various pruning techniques have been explored in prior work, which makes the novelty contribution of this paper somewhat limited. \n\nFurthermore, the study has mostly focused on the impact of random sparsity to accuracy. However, as it is known that it is really difficult for modern hardware to benefit from random sparsity because it leads to irregular memory accesses, which negatively impact the performance. It has been observed that speedups are very limited or can be negative even the random sparsity is >95% [1]. Therefore, it is hard to judge how inference or training can benefit from 30-40% weight sparsity. Going forward, the authors are encouraged to choose pruning methods that lead to regular memory access to avoid adversely impacting practical acceleration in modern hardware platforms.\n\n[1] Learning Structured Sparsity in Deep Neural Networks. Wen et al. NeurIPS 2016"}, "tcdate": 1571816861294}, {"id": "HJeqpvgaYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1214/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper conducts a series of interesting experiments on compressing BERT and makes several conclusions. The compression technique is magnitude weight pruning based on an existing work. The paper mainly tested different compression rates and the stages when the compression can be applied. Compared to the existing work, one main contribution of the paper is to show that the BERT model can be pruned prior to fine-tuning any specific downstream tasks by 30%-40% without affecting all tested downstream tasks much. The paper is well motivated and presents interesting experimental results and conclusions. I have some concerns on their experiment details, which needs some clarification.\n\n1. the observation in 3.4 is a little counter-intuitive to me. The model has all pre-trained weights and should be able to determine, during fine-tuning, which weights to decrease to nearly zero or to abandon. However, the experimental results show that the pruning at that point produces a worse dev accuracy. For the experiments, 3 epochs is used for fine-tuning and then the pruning is applied. I was wondering what happen if you first fine-tune the model to get the best dev accuracy and prune the weights at that point. How did you choose the number 3? I am guessing that the pruning in the middle of fine-tuning process may throw away useful information too early.   \n2. It will be helpful to show the thresholds of pruning and how these thresholds relate to the training loss and accuracy. I think the value of the thresholds can tell whether some pruning ratios are reasonable.\n3. when the authors continue training the model, for example in 3.4, the training stops when the training losses are comparable. Why did the training loss is used as the metric instead of the dev accuracy? Figure 1 right seems to show that those models are overfitting. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper conducts a series of interesting experiments on compressing BERT and makes several conclusions. The compression technique is magnitude weight pruning based on an existing work. The paper mainly tested different compression rates and the stages when the compression can be applied. Compared to the existing work, one main contribution of the paper is to show that the BERT model can be pruned prior to fine-tuning any specific downstream tasks by 30%-40% without affecting all tested downstream tasks much. The paper is well motivated and presents interesting experimental results and conclusions. I have some concerns on their experiment details, which needs some clarification.\n\n1. the observation in 3.4 is a little counter-intuitive to me. The model has all pre-trained weights and should be able to determine, during fine-tuning, which weights to decrease to nearly zero or to abandon. However, the experimental results show that the pruning at that point produces a worse dev accuracy. For the experiments, 3 epochs is used for fine-tuning and then the pruning is applied. I was wondering what happen if you first fine-tune the model to get the best dev accuracy and prune the weights at that point. How did you choose the number 3? I am guessing that the pruning in the middle of fine-tuning process may throw away useful information too early.   \n2. It will be helpful to show the thresholds of pruning and how these thresholds relate to the training loss and accuracy. I think the value of the thresholds can tell whether some pruning ratios are reasonable.\n3. when the authors continue training the model, for example in 3.4, the training stops when the training losses are comparable. Why did the training loss is used as the metric instead of the dev accuracy? Figure 1 right seems to show that those models are overfitting. "}, "tcdate": 1571780545901}], "openreview_url": "https://openreview.net/forum?id=SJlPOCEKvH", "arxiv_id": "2002.08307", "paper_pdf": "papers/SJlPOCEKvH.pdf", "paper_pdf_sha256": "174bfe6a31d92af78db468a00480ad0ac4cbe674b0c09948e033f2a36ae7e14a", "paper_pdf_bytes": 4162391, "paper_pdf_source": "openreview", "code_url": "https://github.com/mitchellgordon95/bert-prune", "code_repository": "mitchellgordon95/bert-prune", "code_commit": "2754ae54d66a6d8d3fa33b3d64811256b21debfa", "code_archive": "repos/SJlPOCEKvH.zip", "code_archive_sha256": "e71072dd081c45c1aa373160738d250c5aa4bcd9873a5378e008d2d8ffd8f884", "code_archive_bytes": 198422, "code_file_count": 51, "code_extensions": {".py": 48, ".sh": 2, ".ipynb": 1}, "github_disk_usage_kb": 387, "github_languages": {"Python": 301882, "Jupyter Notebook": 66488, "Shell": 601}, "github_archived": false, "github_pushed_at": "2020-05-14T20:11:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/compressing-bert-studying-the-effects-of-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Asvt8BAaZT", "year": 2026, "status": "rejected", "title": "When Can You Get Away with Low Memory Adam?", "authors": ["Dayal Singh Kalra", "John Kirchenbauer", "Maissam Barkeshli", "Tom Goldstein"], "authorids": ["~Dayal_Singh_Kalra1", "~John_Kirchenbauer1", "~Maissam_Barkeshli1", "~Tom_Goldstein1"], "authors_source": "OpenReview API", "abstract": "Adam is the go-to optimizer for training modern machine learning models, but it requires additional memory to maintain the moving averages of the gradients and their squares. While various low-memory optimizers have been proposed that sometimes match the performance of Adam, their lack of reliability has left Adam as the default choice. In this work, we apply a simple layer-wise Signal-to-Noise Ratio (SNR) analysis to quantify when second-moment tensors can be effectively replaced by their means across different dimensions. Our SNR analysis reveals how architecture, training hyperparameters, and dataset properties impact compressibility along Adam's trajectory, naturally leading to \\emph{SlimAdam}, a memory-efficient Adam variant. \\emph{SlimAdam} compresses the second moments along dimensions with high SNR when feasible, and leaves when compression would be detrimental. Through experiments across a diverse set of architectures and training scenarios, we show that \\emph{SlimAdam} matches Adam's performance and stability while saving up to 98% of total second moments.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "vbyx1AP1Dv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11689/Reviewer_75Xb"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper proposes a metric called SNR (gradient version) to address the high memory usage of Adam’s second momentum. The paper compresses the second-momentum tensor along these dimensions to a single mean value (or small number of values). The authors claim this approach (SlimAdam) can save up to 99% of the second-moment memory while maintaining Adam’s property and stability.", "review_text": "This paper proposes a metric called SNR (gradient version) to address the high memory usage of Adam’s second momentum. The paper compresses the second-momentum tensor along these dimensions to a single mean value (or small number of values). The authors claim this approach (SlimAdam) can save up to 99% of the second-moment memory while maintaining Adam’s property and stability.", "strengths": "•\tThe idea of quantifying the compressibility of Adam's second moments on a per-layer, per-dimension basis is interesting.\n\n•\tThe authors conducted extensive experiments across various architectures (GPT, ViT, ResNet) and tasks.", "weaknesses": "1.\tLack of Theoretical Foundation: As an optimizer paper, it lacks a convergence proof and relies almost entirely on experimental observations (e.g., using a 10x lower LR).\n\n2.\tMissing Essential Data: The paper does not include 'loss vs. step' curves, a critical metric for evaluating optimizers.\n\n3.\tMethodological Ambiguity: There is no logical basis for the proxy model design or the SNR threshold of $\\alpha=1$.\n\n4.\tQuestionable SNR Justification: The underlying assumption of mapping the mean of $V_t$ to 'signal' and its variance to 'noise' is not justified. (Since the justification of SNR relates to original vs. noise) \n\n5.\tExaggerated Contribution & Poor Comparison: The 99% claim is misleading (it's 50% of the total), and comparisons to SOTA optimizers that compress both moments (e.g., SMMF: Square-Matricized Momentum Factorization for Memory-Efficient Optimization which reduces the first and second momentums so that the total compression ratio is by up to 96%) are missing.", "questions": "•\tWhat is the theoretical justification for treating the mean as 'signal' and the variance as 'noise' in your SNR definition $SNR_K = \\mathbb{E}[(\\mathbb{E}_K[V_t])^2 / Var_K[V_t]]$ from an optimization perspective? Is there a theoretical basis to claim that a low-variance (high SNR) tensor is inherently 'compressible'?\n\n•\tWhat is the specific theoretical justification for choosing the SNR threshold $\\alpha=1$? Can you guarantee this value is universally optimal across different tasks and model architectures?\n\n•\tCan you provide loss-vs-step curves for your key results (e.g., Figure 8) to demonstrate that SlimAdam achieves the same 'final' performance with the same convergence speed as Adam?\n\n•\tCompared to optimizers like SMMF (2025), which compress both first and second moments for a 96% total memory saving, what is the practical advantage of SlimAdam, which only compresses the second moment for a ~50% total saving?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a metric called SNR (gradient version) to address the high memory usage of Adam’s second momentum. The paper compresses the second-momentum tensor along these dimensions to a single mean value (or small number of values). The authors claim this approach (SlimAdam) can save up to 99% of the second-moment memory while maintaining Adam’s property and stability.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "•\tThe idea of quantifying the compressibility of Adam's second moments on a per-layer, per-dimension basis is interesting.\n\n•\tThe authors conducted extensive experiments across various architectures (GPT, ViT, ResNet) and tasks.", "weaknesses": "1.\tLack of Theoretical Foundation: As an optimizer paper, it lacks a convergence proof and relies almost entirely on experimental observations (e.g., using a 10x lower LR).\n\n2.\tMissing Essential Data: The paper does not include 'loss vs. step' curves, a critical metric for evaluating optimizers.\n\n3.\tMethodological Ambiguity: There is no logical basis for the proxy model design or the SNR threshold of $\\alpha=1$.\n\n4.\tQuestionable SNR Justification: The underlying assumption of mapping the mean of $V_t$ to 'signal' and its variance to 'noise' is not justified. (Since the justification of SNR relates to original vs. noise) \n\n5.\tExaggerated Contribution & Poor Comparison: The 99% claim is misleading (it's 50% of the total), and comparisons to SOTA optimizers that compress both moments (e.g., SMMF: Square-Matricized Momentum Factorization for Memory-Efficient Optimization which reduces the first and second momentums so that the total compression ratio is by up to 96%) are missing.", "questions": "•\tWhat is the theoretical justification for treating the mean as 'signal' and the variance as 'noise' in your SNR definition $SNR_K = \\mathbb{E}[(\\mathbb{E}_K[V_t])^2 / Var_K[V_t]]$ from an optimization perspective? Is there a theoretical basis to claim that a low-variance (high SNR) tensor is inherently 'compressible'?\n\n•\tWhat is the specific theoretical justification for choosing the SNR threshold $\\alpha=1$? Can you guarantee this value is universally optimal across different tasks and model architectures?\n\n•\tCan you provide loss-vs-step curves for your key results (e.g., Figure 8) to demonstrate that SlimAdam achieves the same 'final' performance with the same convergence speed as Adam?\n\n•\tCompared to optimizers like SMMF (2025), which compress both first and second moments for a 96% total memory saving, what is the practical advantage of SlimAdam, which only compresses the second moment for a ~50% total saving?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761876304749}, {"id": "q9NQh4SGJ5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11689/Reviewer_LfWr"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The authors present SlimAdam, a memory-efficient  version of Adam optimizer which achieves up to 99% memory savings by compressing the large second-moment statistics used in Adam's adaptive learning rate computations.  Rather than storing full per-parameter second moments, SlimAdam takes mean of these moments along the fan-in or fan-out dimensions 'when' appropriate. The 'when' in context is guided by a Signal-to-Noise Ratio (SNR) metric. \n\nSNR measures the concentration of second moment values (square of mean/variance). Higher SNR indicates tighter clustering, which justifies compression. And the compression is applied only in the layers where SNR is high, and state granularity is retained where SNR is low. The authors also state that since different layers show compression viability across different dimensions (fan-in/fan-out), derivation of compression rules are required for each model. To determine compression rules for each layer, the authors propose training a small proxy model at a reduced learning rate. SNR statistics from the proxy reliably generalize across larger models of the same architecture and task, informing safe compression dimensions for the full target model. \n\nThe authors conduct a compreshensive empirical analysis across a wide range of large models and training tasks, revealing nuanced differences in compressibility across various layer types. Their findings highlight that attention components (such as keys and queries), value and projection layers, MLP layers, and token embedding/vocabulary layers each exhibit distinct compression characteristics. This detailed analysis reveals important, insightful architectural patterns that govern how adaptive moment compressibility varies.\n\nOverall, the paper makes a relevant contribution to efficient optimization for large-scale deep learning, addressing a critical bottleneck in resource consumption. It balances rigorous analysis with practical effectiveness, although clearer exposition, especially regarding the proxy model methodology, would improve accessibility. SlimAdam, hence, is a valuable addition for researchers without sacrificing Adam's effectiveness.", "review_text": "The authors present SlimAdam, a memory-efficient  version of Adam optimizer which achieves up to 99% memory savings by compressing the large second-moment statistics used in Adam's adaptive learning rate computations.  Rather than storing full per-parameter second moments, SlimAdam takes mean of these moments along the fan-in or fan-out dimensions 'when' appropriate. The 'when' in context is guided by a Signal-to-Noise Ratio (SNR) metric. \n\nSNR measures the concentration of second moment values (square of mean/variance). Higher SNR indicates tighter clustering, which justifies compression. And the compression is applied only in the layers where SNR is high, and state granularity is retained where SNR is low. The authors also state that since different layers show compression viability across different dimensions (fan-in/fan-out), derivation of compression rules are required for each model. To determine compression rules for each layer, the authors propose training a small proxy model at a reduced learning rate. SNR statistics from the proxy reliably generalize across larger models of the same architecture and task, informing safe compression dimensions for the full target model. \n\nThe authors conduct a compreshensive empirical analysis across a wide range of large models and training tasks, revealing nuanced differences in compressibility across various layer types. Their findings highlight that attention components (such as keys and queries), value and projection layers, MLP layers, and token embedding/vocabulary layers each exhibit distinct compression characteristics. This detailed analysis reveals important, insightful architectural patterns that govern how adaptive moment compressibility varies.\n\nOverall, the paper makes a relevant contribution to efficient optimization for large-scale deep learning, addressing a critical bottleneck in resource consumption. It balances rigorous analysis with practical effectiveness, although clearer exposition, especially regarding the proxy model methodology, would improve accessibility. SlimAdam, hence, is a valuable addition for researchers without sacrificing Adam's effectiveness.", "strengths": "1) SlimAdam achieves 99% memory savings compared to the original Adam optimizer, while fully preserving Adam’s effectiveness. It can be seamlessly swapped in place of Adam without requiring any code modifications or additional overhead.\n\n2) The paper presents clear and well-motivated research questions supported by extensive experiments across diverse model architectures and tasks, demonstrating robust generality.\n\n3) The authors provide a detailed algorithmic description alongside publicly available code, ensuring reproducibility.\n\n4) Ablation studies are thoughtfully designed and thoroughly explained, offering valuable insights into the contributions of individual components and hyperparameters.", "weaknesses": "1) The main method (SlimAdam algorithm) is explained only in the appendix, and critical implementation insights (proxy model construction, SNR statistics collection) are not clearly presented in the main text. This prevents immediate accessibility and understanding.\n\n2) The concept and practicalities of the proxy model for collecting SNR statistics are not deeply explained. Details about how proxy model size affects SNR relevance and how well proxy-derived rules scale to actual large models could be clearer. Also, how much compute overhead is added for such proxy runs should also be mentioned.\n\n3) The details of SNR statistics adoption over the training steps in the actual model could be explained as well.\n\nMinor Weaknesses:\nAppendix C.1 is not completely written.", "questions": "1) How does proxy model size affect the SNR statistics for different tasks and architectures?\n\n2) The paper states that for proxy model ignores early SNR statistics, and averages SNR values for next few steps rather than all steps; is the same applied for full model as well- meaning, is compressibility not applied for the first few runs, and how is it adapted over training steps?\n\nI am amenable to changing the score if the questions and weaknesses are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present SlimAdam, a memory-efficient  version of Adam optimizer which achieves up to 99% memory savings by compressing the large second-moment statistics used in Adam's adaptive learning rate computations.  Rather than storing full per-parameter second moments, SlimAdam takes mean of these moments along the fan-in or fan-out dimensions 'when' appropriate. The 'when' in context is guided by a Signal-to-Noise Ratio (SNR) metric. \n\nSNR measures the concentration of second moment values (square of mean/variance). Higher SNR indicates tighter clustering, which justifies compression. And the compression is applied only in the layers where SNR is high, and state granularity is retained where SNR is low. The authors also state that since different layers show compression viability across different dimensions (fan-in/fan-out), derivation of compression rules are required for each model. To determine compression rules for each layer, the authors propose training a small proxy model at a reduced learning rate. SNR statistics from the proxy reliably generalize across larger models of the same architecture and task, informing safe compression dimensions for the full target model. \n\nThe authors conduct a compreshensive empirical analysis across a wide range of large models and training tasks, revealing nuanced differences in compressibility across various layer types. Their findings highlight that attention components (such as keys and queries), value and projection layers, MLP layers, and token embedding/vocabulary layers each exhibit distinct compression characteristics. This detailed analysis reveals important, insightful architectural patterns that govern how adaptive moment compressibility varies.\n\nOverall, the paper makes a relevant contribution to efficient optimization for large-scale deep learning, addressing a critical bottleneck in resource consumption. It balances rigorous analysis with practical effectiveness, although clearer exposition, especially regarding the proxy model methodology, would improve accessibility. SlimAdam, hence, is a valuable addition for researchers without sacrificing Adam's effectiveness.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1) SlimAdam achieves 99% memory savings compared to the original Adam optimizer, while fully preserving Adam’s effectiveness. It can be seamlessly swapped in place of Adam without requiring any code modifications or additional overhead.\n\n2) The paper presents clear and well-motivated research questions supported by extensive experiments across diverse model architectures and tasks, demonstrating robust generality.\n\n3) The authors provide a detailed algorithmic description alongside publicly available code, ensuring reproducibility.\n\n4) Ablation studies are thoughtfully designed and thoroughly explained, offering valuable insights into the contributions of individual components and hyperparameters.", "weaknesses": "1) The main method (SlimAdam algorithm) is explained only in the appendix, and critical implementation insights (proxy model construction, SNR statistics collection) are not clearly presented in the main text. This prevents immediate accessibility and understanding.\n\n2) The concept and practicalities of the proxy model for collecting SNR statistics are not deeply explained. Details about how proxy model size affects SNR relevance and how well proxy-derived rules scale to actual large models could be clearer. Also, how much compute overhead is added for such proxy runs should also be mentioned.\n\n3) The details of SNR statistics adoption over the training steps in the actual model could be explained as well.\n\nMinor Weaknesses:\nAppendix C.1 is not completely written.", "questions": "1) How does proxy model size affect the SNR statistics for different tasks and architectures?\n\n2) The paper states that for proxy model ignores early SNR statistics, and averages SNR values for next few steps rather than all steps; is the same applied for full model as well- meaning, is compressibility not applied for the first few runs, and how is it adapted over training steps?\n\nI am amenable to changing the score if the questions and weaknesses are addressed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761762852382}, {"id": "Le13HG4KOQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11689/Reviewer_hoHf"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes to compress the second moment tensor of Adam by replacing per coordinate value with the average across specific dimensions. The method defines the signal to noise ratio of the second moment during training and compresses when this ratio is large.", "review_text": "The paper proposes to compress the second moment tensor of Adam by replacing per coordinate value with the average across specific dimensions. The method defines the signal to noise ratio of the second moment during training and compresses when this ratio is large.", "strengths": "As the memory of the optimizer accounts for a significant fraction of the memory requirement for neural network training, compressing it is an important problem. The paper studies a simple method for this task and gives thorough evaluation on different tasks and different modules of the network.", "weaknesses": "The method seems to have a large overlap with existing literature. The paper mentions Adam-mini, which already has a large overlap in terms of both the algorithm and the intuition behind the approach. Similar techniques also appear in several other papers such as\n\nLean and Mean Adaptive Optimization via Subset-Norm and Subspace-Momentum with Convergence Guarantees. ICML 2025\n\nAPOLLO: SGD-like Memory, AdamW-level Performance. MLSys 2025\n\nThese papers additionally save memory for the momentum, resulting in less memory than the method proposed here. In light of these works in addition to Adam-mini, the contribution of the new paper seems limited.\n\nA minor point: please cite the published versions of the references. For example, the Adam-mini paper is in ICLR 2025.", "questions": "The pre-conditioner changes over time as Adam changes V in every step of training. The SNR analysis is fixed up front and the state is compressed in exactly the same way throughout but the condition number of V could change over time. Do you see any change in the condition number of V over time, and if not, is this a property of the training data and are there cases where the condition number of V changes? \n\nIn some work, it is mentioned that gradient descent operates on the \"edge of stability\", do you see any changes if SNR analysis is done at different step sizes?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes to compress the second moment tensor of Adam by replacing per coordinate value with the average across specific dimensions. The method defines the signal to noise ratio of the second moment during training and compresses when this ratio is large.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "As the memory of the optimizer accounts for a significant fraction of the memory requirement for neural network training, compressing it is an important problem. The paper studies a simple method for this task and gives thorough evaluation on different tasks and different modules of the network.", "weaknesses": "The method seems to have a large overlap with existing literature. The paper mentions Adam-mini, which already has a large overlap in terms of both the algorithm and the intuition behind the approach. Similar techniques also appear in several other papers such as\n\nLean and Mean Adaptive Optimization via Subset-Norm and Subspace-Momentum with Convergence Guarantees. ICML 2025\n\nAPOLLO: SGD-like Memory, AdamW-level Performance. MLSys 2025\n\nThese papers additionally save memory for the momentum, resulting in less memory than the method proposed here. In light of these works in addition to Adam-mini, the contribution of the new paper seems limited.\n\nA minor point: please cite the published versions of the references. For example, the Adam-mini paper is in ICLR 2025.", "questions": "The pre-conditioner changes over time as Adam changes V in every step of training. The SNR analysis is fixed up front and the state is compressed in exactly the same way throughout but the condition number of V could change over time. Do you see any change in the condition number of V over time, and if not, is this a property of the training data and are there cases where the condition number of V changes? \n\nIn some work, it is mentioned that gradient descent operates on the \"edge of stability\", do you see any changes if SNR analysis is done at different step sizes?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761682364875}, {"id": "3zFDii4LMp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11689/Reviewer_BcCm"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper studies the signal-to-noise ratio of Adam’s second-moment tensors for every layer along the input-channel (column) and output-channel (row) directions, finding instances where entries exhibit low variance relative to their mean and can safely share statistics during training; using these SNR profiles, it introduces SlimAdam, which collapses second moments only on the high-SNR direction of each layer, cutting memory while preserving Adam-level convergence and accuracy.", "review_text": "This paper studies the signal-to-noise ratio of Adam’s second-moment tensors for every layer along the input-channel (column) and output-channel (row) directions, finding instances where entries exhibit low variance relative to their mean and can safely share statistics during training; using these SNR profiles, it introduces SlimAdam, which collapses second moments only on the high-SNR direction of each layer, cutting memory while preserving Adam-level convergence and accuracy.", "strengths": "The paper tackles a well-motivated problem—the large memory footprint of Adam’s second-moment matrices—and clearly pinpoints instances where collapsing each layer’s parameters to a single scalar shrinks an matrix to just or values, cutting memory use while retaining Adam-level accuracy and stability.", "weaknesses": "Since there are almost no changes from my last review of the paper, I'll keep the core of my argument. I’ll split my review into two parts — one on the empirical analysis and one on the proposed optimizer.\n\n### Empirical analysis and design rationale\n\nThe paper’s primary findings are almost entirely empirical, and this lack of theory leaves several key decisions unclear—especially because the empirical signals themselves are not particularly strong. Axis sharing is constrained to whole fan-in or fan-out dimensions purely for implementation convenience, with no exploration of alternative groupings or proof that these axes are optimal (e.g., would results change if one considers randomly partitioning a layer’s parameters into two equal-size groups?). Similarly, the paper adopts (the interesting metric) SNR with a heuristic threshold as the only compression criterion, although simple variance (which answers “How much l2 loss do we pay if we collapse this vector to a scalar?”) would align more directly with the intuition the authors cite (“If entries along a dimension exhibit low variance relative to their mean, they can be effectively represented by a single value”).\n\nLearning rate is the only hyperparameter the paper systematically analyzes. It is presented as the dominant knob that shifts SNR and thus determines which layers can be compressed, yet the text provides no a-priori reason why learning rate—rather than, say, Adam’s momentum coefficients or the batch size—should hold that position.\n\n\n### Optimizer details\nThe optimizer requires selecting a compression axis for each layer or layer type. Choosing a compression axis means either relying on proxies or heuristics, or collecting fresh SNR statistics. My concern is that the latter defeats the purpose in some cases, and the former is not reliable.\n\nAlternatively, we can train a small proxy model or reuse generic rules. Yet the authors themselves show that preferred compression axes shift with dataset, width, and vocabulary size. Even within the same dataset and width, layers of the same type show different preferences. Depth-averaging does not fully solve the problem for users operating at the tightest memory margins or in domains whose depth-specific SNR patterns have not been studied yet. Even the stronger patterns they find—for example, compressing along the embedding dimension versus the token dimension—may not yield an SNR above the cutoff needed to justify compression.\n\nFull-size SNR collection defeats the purpose. To decide the sharing axis, you must first run the uncompressed model under standard Adam long enough to gather per-layer SNR statistics. During this warm-up, you still store the full second-moment tensors, so the memory spike SlimAdam tries to avoid is paid up front. For practitioners who want to fit a slightly larger model into fixed hardware, this spike means the maximum model size is still bounded by Adam’s footprint during the warm-up, undermining the value of a lighter optimizer.", "questions": "Please address my concerns above, especially around the empirical nature of the evidence, axis selection, and practicality at tight memory budgets. In addition, a high-level clarification would help: how should we interpret “compressibility” in this work beyond plots of SNR? In other words, is SNR a sufficient observable for when per-parameter adaptivity is redundant, and how does its dependence on learning rate versus other hyperparameters shape the generality of your claims?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the signal-to-noise ratio of Adam’s second-moment tensors for every layer along the input-channel (column) and output-channel (row) directions, finding instances where entries exhibit low variance relative to their mean and can safely share statistics during training; using these SNR profiles, it introduces SlimAdam, which collapses second moments only on the high-SNR direction of each layer, cutting memory while preserving Adam-level convergence and accuracy.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The paper tackles a well-motivated problem—the large memory footprint of Adam’s second-moment matrices—and clearly pinpoints instances where collapsing each layer’s parameters to a single scalar shrinks an matrix to just or values, cutting memory use while retaining Adam-level accuracy and stability.", "weaknesses": "Since there are almost no changes from my last review of the paper, I'll keep the core of my argument. I’ll split my review into two parts — one on the empirical analysis and one on the proposed optimizer.\n\n### Empirical analysis and design rationale\n\nThe paper’s primary findings are almost entirely empirical, and this lack of theory leaves several key decisions unclear—especially because the empirical signals themselves are not particularly strong. Axis sharing is constrained to whole fan-in or fan-out dimensions purely for implementation convenience, with no exploration of alternative groupings or proof that these axes are optimal (e.g., would results change if one considers randomly partitioning a layer’s parameters into two equal-size groups?). Similarly, the paper adopts (the interesting metric) SNR with a heuristic threshold as the only compression criterion, although simple variance (which answers “How much l2 loss do we pay if we collapse this vector to a scalar?”) would align more directly with the intuition the authors cite (“If entries along a dimension exhibit low variance relative to their mean, they can be effectively represented by a single value”).\n\nLearning rate is the only hyperparameter the paper systematically analyzes. It is presented as the dominant knob that shifts SNR and thus determines which layers can be compressed, yet the text provides no a-priori reason why learning rate—rather than, say, Adam’s momentum coefficients or the batch size—should hold that position.\n\n\n### Optimizer details\nThe optimizer requires selecting a compression axis for each layer or layer type. Choosing a compression axis means either relying on proxies or heuristics, or collecting fresh SNR statistics. My concern is that the latter defeats the purpose in some cases, and the former is not reliable.\n\nAlternatively, we can train a small proxy model or reuse generic rules. Yet the authors themselves show that preferred compression axes shift with dataset, width, and vocabulary size. Even within the same dataset and width, layers of the same type show different preferences. Depth-averaging does not fully solve the problem for users operating at the tightest memory margins or in domains whose depth-specific SNR patterns have not been studied yet. Even the stronger patterns they find—for example, compressing along the embedding dimension versus the token dimension—may not yield an SNR above the cutoff needed to justify compression.\n\nFull-size SNR collection defeats the purpose. To decide the sharing axis, you must first run the uncompressed model under standard Adam long enough to gather per-layer SNR statistics. During this warm-up, you still store the full second-moment tensors, so the memory spike SlimAdam tries to avoid is paid up front. For practitioners who want to fit a slightly larger model into fixed hardware, this spike means the maximum model size is still bounded by Adam’s footprint during the warm-up, undermining the value of a lighter optimizer.", "questions": "Please address my concerns above, especially around the empirical nature of the evidence, axis selection, and practicality at tight memory budgets. In addition, a high-level clarification would help: how should we interpret “compressibility” in this work beyond plots of SNR? In other words, is SNR a sufficient observable for when per-parameter adaptivity is redundant, and how does its dependence on learning rate versus other hyperparameters shape the generality of your claims?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761578396537}, {"id": "RbMGCtbtaW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11689/Reviewer_XBFB"], "rating": 2, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a layer-wise Signal-to-Noise Ratio (SNR) analysis to determine when the second-moment tensors in optimization algorithms (e.g., Adam) can be compressed by replacing them with their dimensional means. Given that SNR is computed as $\\text{mean} / \\text{variance}$, it serves as a natural metric for this purpose: a high SNR indicates that a tensor can be effectively approximated by its mean without significant performance loss. This approach provides a practical, quantitative guide for compressing optimizer states and offers evidence that Adam may not always require full second-moment information.", "review_text": "This paper proposes a layer-wise Signal-to-Noise Ratio (SNR) analysis to determine when the second-moment tensors in optimization algorithms (e.g., Adam) can be compressed by replacing them with their dimensional means. Given that SNR is computed as $\\text{mean} / \\text{variance}$, it serves as a natural metric for this purpose: a high SNR indicates that a tensor can be effectively approximated by its mean without significant performance loss. This approach provides a practical, quantitative guide for compressing optimizer states and offers evidence that Adam may not always require full second-moment information.", "strengths": "This work proposes to apply SNR as a metric to guide the compression of second-moment tensors with their means in LLM training. It offers a threshold-based criteria to determine when and how such mean compression can be applied across different architectural components of LLMs (e.g., query, key, value, and MLP layers). Furthermore, this work empircially investigates several factors influencing compressibility, including learning rate, data distribution, and initialization.", "weaknesses": "1. The main motivation of this work is to establish a metric for guiding dimension-wise mean compression of second-moment tensors and provide SlimAdam. However, this goal appears to overlap with Adam-mini [1], which not only implemenst a compression method based on block-wise mean values but also provides insights based on Hessian structure to explain why the full second-moment may be unnecessary and to guide how to compress. The authors should more clearly delineate their contributions and explicitly contrast their approach with the insights and methods provided by Adam-mini.\n\n2. SNR is a natural choice for quantifying the viability of mean compression, given its formula of $\\text{mean} / \\text{variance}$. The paper does not sufficiently justify why it is superior to other plausible metrics. For instance, measures based on the L2-norm or KL-divergence of the error introduced by compression, or just variance, could be more direct and computationally efficient. The author should demonstrate the unique advantages of SNR over other alternatives through theoretical analysis or empirical comparision. \n\n3. For mean compression, it is clear that higher SNR correlates with better compressibility. However, the method remains dependent on an empirically set threshold (e.g., $\\alpha=1$) to make compression decisions. This dependency not only limits the generality of the method by introducing a potentially sensitive hyperparameter across different scenarios but also raises the question of whether other metrics (e.g., L2-norm, KL-divergence, or variance) could perform just as effectively with a similarly tuned threshold.\n\n4. The utility of SNR seems limited to mean compression and may not extend to or guide other compression paradigms (e.g., low-rank factorization, quantization).\n\n[1] Zhang, Yushun, et al. \"Adam-mini: Use fewer learning rates to gain more.\" arXiv preprint arXiv:2406.16793 (2024).", "questions": "1. The choice of SNR is intuitive for mean replacement, but why is it superior to other direct measures of compression error, such as the L2-norm or KL-divergence between the original and compressed tensor? Could the authors provide either (a) an empirical ablation study comparing the compression guidance performance of SNR against these other metrics, or (b) a theoretical argument for why SNR is an optimal or more robust criterion?\n\n2. If the performance of the method is similar when using a simple threshold on other metrics (e.g., compress if $\\text{variance} < X$, compress if $\\text{L2-norm of error introduced by compression}/\\text{L2-norm of target tensor}$), does this suggest the core insight is about identifying low-variance parameters rather than the unique information provided by SNR? What is the specific advantage of the SNR ratio over just using the variance or standard deviation?\n\n3. The threshold $\\alpha=1$ is presented as a critical value for making compression decisions. How was this value determined? Is it robust across different model architectures, layers, and tasks? Could the authors show sensitivity analyses for this threshold to demonstrate its generality?\n\n4. The presentation could be improved for better clarity and reproducibility. For example, using pseudocode rather than plain text would help readers better understand the algorithm's workflow.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a layer-wise Signal-to-Noise Ratio (SNR) analysis to determine when the second-moment tensors in optimization algorithms (e.g., Adam) can be compressed by replacing them with their dimensional means. Given that SNR is computed as $\\text{mean} / \\text{variance}$, it serves as a natural metric for this purpose: a high SNR indicates that a tensor can be effectively approximated by its mean without significant performance loss. This approach provides a practical, quantitative guide for compressing optimizer states and offers evidence that Adam may not always require full second-moment information.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "This work proposes to apply SNR as a metric to guide the compression of second-moment tensors with their means in LLM training. It offers a threshold-based criteria to determine when and how such mean compression can be applied across different architectural components of LLMs (e.g., query, key, value, and MLP layers). Furthermore, this work empircially investigates several factors influencing compressibility, including learning rate, data distribution, and initialization.", "weaknesses": "1. The main motivation of this work is to establish a metric for guiding dimension-wise mean compression of second-moment tensors and provide SlimAdam. However, this goal appears to overlap with Adam-mini [1], which not only implemenst a compression method based on block-wise mean values but also provides insights based on Hessian structure to explain why the full second-moment may be unnecessary and to guide how to compress. The authors should more clearly delineate their contributions and explicitly contrast their approach with the insights and methods provided by Adam-mini.\n\n2. SNR is a natural choice for quantifying the viability of mean compression, given its formula of $\\text{mean} / \\text{variance}$. The paper does not sufficiently justify why it is superior to other plausible metrics. For instance, measures based on the L2-norm or KL-divergence of the error introduced by compression, or just variance, could be more direct and computationally efficient. The author should demonstrate the unique advantages of SNR over other alternatives through theoretical analysis or empirical comparision. \n\n3. For mean compression, it is clear that higher SNR correlates with better compressibility. However, the method remains dependent on an empirically set threshold (e.g., $\\alpha=1$) to make compression decisions. This dependency not only limits the generality of the method by introducing a potentially sensitive hyperparameter across different scenarios but also raises the question of whether other metrics (e.g., L2-norm, KL-divergence, or variance) could perform just as effectively with a similarly tuned threshold.\n\n4. The utility of SNR seems limited to mean compression and may not extend to or guide other compression paradigms (e.g., low-rank factorization, quantization).\n\n[1] Zhang, Yushun, et al. \"Adam-mini: Use fewer learning rates to gain more.\" arXiv preprint arXiv:2406.16793 (2024).", "questions": "1. The choice of SNR is intuitive for mean replacement, but why is it superior to other direct measures of compression error, such as the L2-norm or KL-divergence between the original and compressed tensor? Could the authors provide either (a) an empirical ablation study comparing the compression guidance performance of SNR against these other metrics, or (b) a theoretical argument for why SNR is an optimal or more robust criterion?\n\n2. If the performance of the method is similar when using a simple threshold on other metrics (e.g., compress if $\\text{variance} < X$, compress if $\\text{L2-norm of error introduced by compression}/\\text{L2-norm of target tensor}$), does this suggest the core insight is about identifying low-variance parameters rather than the unique information provided by SNR? What is the specific advantage of the SNR ratio over just using the variance or standard deviation?\n\n3. The threshold $\\alpha=1$ is presented as a critical value for making compression decisions. How was this value determined? Is it robust across different model architectures, layers, and tasks? Could the authors show sensitivity analyses for this threshold to demonstrate its generality?\n\n4. The presentation could be improved for better clarity and reproducibility. For example, using pseudocode rather than plain text would help readers better understand the algorithm's workflow.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760531308209}], "openreview_url": "https://openreview.net/forum?id=Asvt8BAaZT", "arxiv_id": "2503.01843", "paper_pdf": "papers/Asvt8BAaZT.pdf", "paper_pdf_sha256": "4bbbea433712864b39cd233b2054fc7ceee2e4d9e8faa5ba918307a4e66dbcb4", "paper_pdf_bytes": 3040815, "paper_pdf_source": "openreview", "code_url": "https://github.com/dayal-kalra/low-memory-adam", "code_repository": "dayal-kalra/low-memory-adam", "code_commit": "bb3652c2fd55eb3e975658ab844c14485f73f693", "code_archive": "repos/Asvt8BAaZT.zip", "code_archive_sha256": "d09081a6cc32722ebc12cde837431040c0a464c3669185e38f35cdf5ad4825ab", "code_archive_bytes": 1291781, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 1575, "github_languages": {"Python": 81530}, "github_archived": false, "github_pushed_at": "2025-03-02T22:27:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/when-can-you-get-away-with-low-memory-adam"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iKgQOAtvsD", "year": 2025, "status": "rejected", "title": "Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial Prompt Translation", "authors": ["Qizhang Li", "Xiaochen Yang", "Wangmeng Zuo", "Yiwen Guo"], "authorids": ["~Qizhang_Li1", "~Xiaochen_Yang2", "~Wangmeng_Zuo2", "~Yiwen_Guo1"], "authors_source": "OpenReview API", "abstract": "Automatic adversarial prompt generation provides remarkable success in jailbreaking safely-aligned large language models (LLMs). Existing gradient-based attacks, while demonstrating outstanding performance in jailbreaking white-box LLMs, often generate garbled adversarial prompts with chaotic appearance. These adversarial prompts are difficult to transfer to other LLMs, hindering their performance in attacking unknown victim models. In this paper, for the first time, we delve into the semantic meaning embedded in garbled adversarial prompts and propose a novel method that *\"translate\"* them into coherent, human-readable natural language adversarial prompts. In this way, we can effectively uncover the semantic information that triggers vulnerabilities in the model and unambiguously transfer it to the victim model, without overlooking the adversarial information hidden in the garbled text, to enhance jailbreak attacks. It also offers a new approach to discovering effective designs for jailbreak prompts, advancing the understanding of jailbreak attacks. Experimental results demonstrate that our method significantly improves the success rate of jailbreak attacks against various safety-aligned LLMs and outperforms state-of-the-arts by a large margin. With at most 10 queries, our method achieves an average attack success rate of 81.8% in attacking 7 commercial closed-source LLMs, including GPT and Claude-3 series, on HarmBench. Our method also achieves over 90% attack success rates against Llama-2-Chat models on AdvBench, despite their outstanding resistance to jailbreaks. Our code will be made publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "rEt6qibCfi", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7266/Reviewer_sM1D"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "Automatic adversarial prompt generation successfully jailbreaks aligned large language models (LLMs). Existing gradient-based attacks produce chaotic prompts that lack transferability to unknown models. This paper introduces a method that translates garbled prompts into coherent, human-readable adversarial prompts, revealing the semantic information needed to exploit model vulnerabilities. Experimental results show a notable increase in success rates, averaging 81.8% against seven commercial LLMs and over 90% against Llama-2-Chat models, surpassing state-of-the-art methods. The code will be publicly available.", "review_text": "Automatic adversarial prompt generation successfully jailbreaks aligned large language models (LLMs). Existing gradient-based attacks produce chaotic prompts that lack transferability to unknown models. This paper introduces a method that translates garbled prompts into coherent, human-readable adversarial prompts, revealing the semantic information needed to exploit model vulnerabilities. Experimental results show a notable increase in success rates, averaging 81.8% against seven commercial LLMs and over 90% against Llama-2-Chat models, surpassing state-of-the-art methods. The code will be publicly available.", "strengths": "Strengths.\n1. The paper is clearly written and motivates the proposed approach well in a lucid manner.\n2. The study of making confusing suffixes semantic is very interesting\n3. The paper proposes a novel method that \"translates\" these prompts into coherent and human-readable natural language adversarial prompts.\n4. The paper demonstrates the effectiveness of the proposed method across different datasets and various Vision-Language Models.", "weaknesses": "Weaknesses\n\n1. This paper claims \"we construct a fully automatic natural language adversarial prompt generation framework, without any manual work for the design of adversarial prompts, careful hyper-parameter tuning, additional model training, or the need for informative feedback of the victim model to refine the adversarial prompts. \" but the proposed method adopt GCG to generate the adversarial suffix on  Llama-3.1-8B-Instruct. It uses the model gradient to generate the suffix. It is not that it cannot access the model at all. Although he can migrate to other models, this part is suspected of over-claiming contributions.\n\n\n2. The parameter settings of the evaluation model are not given, such as the system prompt. Previous works used different system prompts to build LLM models, resulting in inconsistent jailbreak difficulty, such as GCG and AutoDAN.\n\n3. Without code, it is impossible to assess the effectiveness of the method. For instance, when I presented GPT-4o with a translated adversarial prompt, its response was, 'I'm sorry, I can't assist with that.'\n\n\n4. HarmBench [1] uses a fine-tuned Llama-2-13B-chat model to compute ASR.  I suggest the authors also follow the exact same evaluation pipeline introduced in [1].\n\n\n[1]  Mazeika, Mantas, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee et al. \"Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.\" arXiv preprint arXiv:2402.04249 (2024).\n\n5.  Would be great to see some qualitative examples\n\n6. The technical portion of this article was merely completed using prompt engineering and contains no technical innovation.", "questions": "Refer to Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Automatic adversarial prompt generation successfully jailbreaks aligned large language models (LLMs). Existing gradient-based attacks produce chaotic prompts that lack transferability to unknown models. This paper introduces a method that translates garbled prompts into coherent, human-readable adversarial prompts, revealing the semantic information needed to exploit model vulnerabilities. Experimental results show a notable increase in success rates, averaging 81.8% against seven commercial LLMs and over 90% against Llama-2-Chat models, surpassing state-of-the-art methods. The code will be publicly available.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Strengths.\n1. The paper is clearly written and motivates the proposed approach well in a lucid manner.\n2. The study of making confusing suffixes semantic is very interesting\n3. The paper proposes a novel method that \"translates\" these prompts into coherent and human-readable natural language adversarial prompts.\n4. The paper demonstrates the effectiveness of the proposed method across different datasets and various Vision-Language Models.", "weaknesses": "Weaknesses\n\n1. This paper claims \"we construct a fully automatic natural language adversarial prompt generation framework, without any manual work for the design of adversarial prompts, careful hyper-parameter tuning, additional model training, or the need for informative feedback of the victim model to refine the adversarial prompts. \" but the proposed method adopt GCG to generate the adversarial suffix on  Llama-3.1-8B-Instruct. It uses the model gradient to generate the suffix. It is not that it cannot access the model at all. Although he can migrate to other models, this part is suspected of over-claiming contributions.\n\n\n2. The parameter settings of the evaluation model are not given, such as the system prompt. Previous works used different system prompts to build LLM models, resulting in inconsistent jailbreak difficulty, such as GCG and AutoDAN.\n\n3. Without code, it is impossible to assess the effectiveness of the method. For instance, when I presented GPT-4o with a translated adversarial prompt, its response was, 'I'm sorry, I can't assist with that.'\n\n\n4. HarmBench [1] uses a fine-tuned Llama-2-13B-chat model to compute ASR.  I suggest the authors also follow the exact same evaluation pipeline introduced in [1].\n\n\n[1]  Mazeika, Mantas, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee et al. \"Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.\" arXiv preprint arXiv:2402.04249 (2024).\n\n5.  Would be great to see some qualitative examples\n\n6. The technical portion of this article was merely completed using prompt engineering and contains no technical innovation.", "questions": "Refer to Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730660940136}, {"id": "Xhv3miGRcB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7266/Reviewer_KAYt"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a novel method to enhance jailbreaking attacks. Given an adversarial suffix, the proposed method first utilizes the LLM to interpret it and then translates it into a natural language adversarial prompt. The empirical results validate the effectiveness of the proposed method in improving the attack success rate.", "review_text": "This paper proposes a novel method to enhance jailbreaking attacks. Given an adversarial suffix, the proposed method first utilizes the LLM to interpret it and then translates it into a natural language adversarial prompt. The empirical results validate the effectiveness of the proposed method in improving the attack success rate.", "strengths": "1. The experiments are comprehensive. The attack success rate (ASR) is evaluated using the state-of-the-art LLMs. The proposed method can significantly improve the ASR, which supports the claim.\n2. The proposed method is intuitive. Table 1 clearly explains the intuition of the proposed method. Besides, the method is efficient and transferable, which provided a better way to evaluate the robustness against jailbreaking attacks.", "weaknesses": "1. This is an intuitive and empirical research work. There is no theoretical guarantee that the proposed method can always improve attack power.\n2. It is better to report the standard deviation as well to validate the significance of the reported results.", "questions": "Please refer to my comments in Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel method to enhance jailbreaking attacks. Given an adversarial suffix, the proposed method first utilizes the LLM to interpret it and then translates it into a natural language adversarial prompt. The empirical results validate the effectiveness of the proposed method in improving the attack success rate.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The experiments are comprehensive. The attack success rate (ASR) is evaluated using the state-of-the-art LLMs. The proposed method can significantly improve the ASR, which supports the claim.\n2. The proposed method is intuitive. Table 1 clearly explains the intuition of the proposed method. Besides, the method is efficient and transferable, which provided a better way to evaluate the robustness against jailbreaking attacks.", "weaknesses": "1. This is an intuitive and empirical research work. There is no theoretical guarantee that the proposed method can always improve attack power.\n2. It is better to report the standard deviation as well to validate the significance of the reported results.", "questions": "Please refer to my comments in Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730640809793}, {"id": "yFyUgKvjT4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7266/Reviewer_ZyqQ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper attempts to address the question of how to improve the success rate of jailbreak attacks on securely aligned Large Language Models (LLMs). Specifically, it aims to improve the transferability and success of attacks by \"translating\" garbled adversarial prompts generated by gradient optimization methods into coherent and human-readable natural language adversarial prompts. This work significantly improves the success rate of jailbreak attacks on securely aligned large language models without the need for manual design or additional model training. The paper conducts extensive experiments on the HarmBench and AdvBench datasets, demonstrating the effectiveness of its approach.", "review_text": "This paper attempts to address the question of how to improve the success rate of jailbreak attacks on securely aligned Large Language Models (LLMs). Specifically, it aims to improve the transferability and success of attacks by \"translating\" garbled adversarial prompts generated by gradient optimization methods into coherent and human-readable natural language adversarial prompts. This work significantly improves the success rate of jailbreak attacks on securely aligned large language models without the need for manual design or additional model training. The paper conducts extensive experiments on the HarmBench and AdvBench datasets, demonstrating the effectiveness of its approach.", "strengths": "It does not require the manual design of adversarial cues, careful hyper-parameter tuning, additional model training costs, or informational feedback from victim models to optimize adversarial cues.\n\nIt provides a new approach to developing new jailbreak designs, combining the advantages of optimization-based approaches and natural-language-based jailbreak. Previously, optimization-based methods usually produced jibberish, which was not robust under perplexity filtering.", "weaknesses": "Limited Testing on Cutting-Edge Models: The paper does not test its method on the latest models, such as O1, which employ extended reasoning paths before answering. These models might require more sophisticated prompts to manipulate their internal reasoning processes, which the current method might not effectively generate. Improvement: The authors should consider testing their approach on such advanced models to evaluate the robustness of their method and potentially adapt their approach to handle more complex reasoning paths.\n\nLack of System-Level Defense Testing: The paper does not address system-level defenses like Purple Llama, which classify and detect input prompts. These defenses could potentially thwart the jailbreak attempts by identifying and filtering out adversarial prompts. Improvement: Incorporating tests against system-level defenses would provide a more realistic assessment of the method's effectiveness in real-world scenarios. The authors could explore how their prompts fare against such defenses and develop strategies to evade detection.", "questions": "Does the paper quantify the architectural and training data differences between the generator model and the translation model? If so, what are the specific data on how these differences affect attack success?\n\nDoes the paper explore the best match between the complexity of the generator model and the complexity of the translation model? Is there evidence that the attack works best at a particular level of complexity?\n\nDoes the paper validate the generalizability of its attack methodology across different types of LLMs (e.g., different architectures, sizes, training datasets)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper attempts to address the question of how to improve the success rate of jailbreak attacks on securely aligned Large Language Models (LLMs). Specifically, it aims to improve the transferability and success of attacks by \"translating\" garbled adversarial prompts generated by gradient optimization methods into coherent and human-readable natural language adversarial prompts. This work significantly improves the success rate of jailbreak attacks on securely aligned large language models without the need for manual design or additional model training. The paper conducts extensive experiments on the HarmBench and AdvBench datasets, demonstrating the effectiveness of its approach.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "It does not require the manual design of adversarial cues, careful hyper-parameter tuning, additional model training costs, or informational feedback from victim models to optimize adversarial cues.\n\nIt provides a new approach to developing new jailbreak designs, combining the advantages of optimization-based approaches and natural-language-based jailbreak. Previously, optimization-based methods usually produced jibberish, which was not robust under perplexity filtering.", "weaknesses": "Limited Testing on Cutting-Edge Models: The paper does not test its method on the latest models, such as O1, which employ extended reasoning paths before answering. These models might require more sophisticated prompts to manipulate their internal reasoning processes, which the current method might not effectively generate. Improvement: The authors should consider testing their approach on such advanced models to evaluate the robustness of their method and potentially adapt their approach to handle more complex reasoning paths.\n\nLack of System-Level Defense Testing: The paper does not address system-level defenses like Purple Llama, which classify and detect input prompts. These defenses could potentially thwart the jailbreak attempts by identifying and filtering out adversarial prompts. Improvement: Incorporating tests against system-level defenses would provide a more realistic assessment of the method's effectiveness in real-world scenarios. The authors could explore how their prompts fare against such defenses and develop strategies to evade detection.", "questions": "Does the paper quantify the architectural and training data differences between the generator model and the translation model? If so, what are the specific data on how these differences affect attack success?\n\nDoes the paper explore the best match between the complexity of the generator model and the complexity of the translation model? Is there evidence that the attack works best at a particular level of complexity?\n\nDoes the paper validate the generalizability of its attack methodology across different types of LLMs (e.g., different architectures, sizes, training datasets)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730590128178}, {"id": "SBIpyS6bzr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7266/Reviewer_cLm8"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "This paper introduces a method for generating coherent and human-readable adversarial prompts from garbled adversarial prompts produced by gradient-based attack methods. The process involves first using an interpretation LLM, followed by a translation LLM. Experimental results show that this approach significantly improves the attack success rate.", "review_text": "This paper introduces a method for generating coherent and human-readable adversarial prompts from garbled adversarial prompts produced by gradient-based attack methods. The process involves first using an interpretation LLM, followed by a translation LLM. Experimental results show that this approach significantly improves the attack success rate.", "strengths": "1. The concept of directly translating garbled adversarial prompts (such as those generated by GCG) into coherent prompts is both simple and novel.\n\n2. The substantial improvement in attack success rates on closed-source models is particularly promising and demonstrates the effectiveness of the method.", "weaknesses": "1. The method assumes that GCG-generated adversarial prompts are always somewhat readable, but this assumption is not clearly justified.\n\n2. The paper lacks a clear explanation for why this method outperforms GCG and other optimization-based approaches. This is counterintuitive, as GCG directly optimizes the adversarial objective, whereas this method does not appear to do so.\n\n3. The paper needs to provide examples of adversarial prompts. Otherwise, there is no way for reviewers/readers to directly check the quality of the adversarial prompts. Could you provide examples of adversarial prompts, along with the user's harmful request, for all the models, especially Llama-2-chat?", "questions": "1. Could you provide a comparison of the adversarial loss between the GCG-optimized prompts and your translated prompts? I would expect the GCG-optimized prompts to achieve a lower loss, as GCG's objective is to minimize the loss. However, your prompts show a higher attack success rate. It would be helpful to see the loss values side by side and to hear your explanation for this discrepancy.\n\n2. Could you provide examples of adversarial prompts so that reviewers and readers can have a better understanding of the results?\n\nI will be very happy to raise my points if I can see the actual adversarial prompts optimized using this method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a method for generating coherent and human-readable adversarial prompts from garbled adversarial prompts produced by gradient-based attack methods. The process involves first using an interpretation LLM, followed by a translation LLM. Experimental results show that this approach significantly improves the attack success rate.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. The concept of directly translating garbled adversarial prompts (such as those generated by GCG) into coherent prompts is both simple and novel.\n\n2. The substantial improvement in attack success rates on closed-source models is particularly promising and demonstrates the effectiveness of the method.", "weaknesses": "1. The method assumes that GCG-generated adversarial prompts are always somewhat readable, but this assumption is not clearly justified.\n\n2. The paper lacks a clear explanation for why this method outperforms GCG and other optimization-based approaches. This is counterintuitive, as GCG directly optimizes the adversarial objective, whereas this method does not appear to do so.\n\n3. The paper needs to provide examples of adversarial prompts. Otherwise, there is no way for reviewers/readers to directly check the quality of the adversarial prompts. Could you provide examples of adversarial prompts, along with the user's harmful request, for all the models, especially Llama-2-chat?", "questions": "1. Could you provide a comparison of the adversarial loss between the GCG-optimized prompts and your translated prompts? I would expect the GCG-optimized prompts to achieve a lower loss, as GCG's objective is to minimize the loss. However, your prompts show a higher attack success rate. It would be helpful to see the loss values side by side and to hear your explanation for this discrepancy.\n\n2. Could you provide examples of adversarial prompts so that reviewers and readers can have a better understanding of the results?\n\nI will be very happy to raise my points if I can see the actual adversarial prompts optimized using this method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730340925617}], "openreview_url": "https://openreview.net/forum?id=iKgQOAtvsD", "arxiv_id": "2410.11317", "paper_pdf": "papers/iKgQOAtvsD.pdf", "paper_pdf_sha256": "b2a1a96470b61dd6f4c81a13a89410bbd8c49ed44c84a17dcff57f2aa9825ea9", "paper_pdf_bytes": 4052233, "paper_pdf_source": "openreview", "code_url": "https://github.com/qizhangli/Adversarial-Prompt-Translator", "code_repository": "qizhangli/Adversarial-Prompt-Translator", "code_commit": "195111673b09c77609b7f3ef5b9cea4daecc7c74", "code_archive": "repos/iKgQOAtvsD.zip", "code_archive_sha256": "d0012f69438c0423f369fe00965840a9ece39dc95450825fe41e7bef19f9a7d0", "code_archive_bytes": 107289, "code_file_count": 9, "code_extensions": {".py": 7, ".sh": 2}, "github_disk_usage_kb": 107, "github_languages": {"Python": 20896, "Shell": 216}, "github_archived": false, "github_pushed_at": "2024-11-07T15:16:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deciphering-the-chaos-enhancing-jailbreak"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wmzFZ9lJrD", "year": 2024, "status": "rejected", "title": "Boolformer: Symbolic Regression of Logic Functions with Transformers", "authors": ["Stéphane d'Ascoli", "Samy Bengio", "Joshua M. Susskind", "Emmanuel Abbe"], "authorids": ["~Stéphane_d'Ascoli1", "~Samy_Bengio1", "~Joshua_M._Susskind1", "~Emmanuel_Abbe1"], "authors_source": "OpenReview API", "abstract": "In this work, we introduce the Boolformer, the first Transformer architecture trained to perform end-to-end symbolic regression of Boolean functions. First, we show that it can predict compact formulas for complex functions which were not seen during training, when provided a clean truth table. Then, we demonstrate its ability to find approximate expressions when provided incomplete and noisy observations. We compare it with classic machine learning approaches on a broad set of real-world binary classification datasets, demonstrating its potential as an interpretable alternative. Finally, we apply it to the widespread task of modelling the dynamics of gene regulatory networks. Using a recent benchmark, we show that Boolformer is competitive with state-of-the art genetic algorithms with a speedup of several orders of magnitude.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "GNb0J3czS9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2514/Reviewer_X8df"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper investigates end-to-end symbolic regression of Boolean functions. The authors show that the overall performance of Boolformer is comparable to classical machine learning in this particular field, with the benefits that Boolformer can be faster and provide interpretable solutions.", "review_text": "The paper investigates end-to-end symbolic regression of Boolean functions. The authors show that the overall performance of Boolformer is comparable to classical machine learning in this particular field, with the benefits that Boolformer can be faster and provide interpretable solutions.", "strengths": "- The paper is well written and easy to follow.\n- Careful description of the generation of synthetic data, and a good analysis of the possible bias included.\n- Empirical evaluation establishes the effectiveness of the approach.\n- A good section of limitation addressing some of my concerns (e.g., not being able to deal with large formulas) which would otherwise go to the weakness below.", "weaknesses": "- While there are some engineering for the embedder, the rest of the approach seems quite standard and straightforward (which is not necessarily a bad thing).\n- It might not be that surprising that Boolformer is faster on GRNs tasks. After all, it has been trained for a long time and the training data could have covered what it needed in these tasks. I am curious, however, is there a similar comparison of efficiency in the noiseless regime?", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates end-to-end symbolic regression of Boolean functions. The authors show that the overall performance of Boolformer is comparable to classical machine learning in this particular field, with the benefits that Boolformer can be faster and provide interpretable solutions.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is well written and easy to follow.\n- Careful description of the generation of synthetic data, and a good analysis of the possible bias included.\n- Empirical evaluation establishes the effectiveness of the approach.\n- A good section of limitation addressing some of my concerns (e.g., not being able to deal with large formulas) which would otherwise go to the weakness below.", "weaknesses": "- While there are some engineering for the embedder, the rest of the approach seems quite standard and straightforward (which is not necessarily a bad thing).\n- It might not be that surprising that Boolformer is faster on GRNs tasks. After all, it has been trained for a long time and the training data could have covered what it needed in these tasks. I am curious, however, is there a similar comparison of efficiency in the noiseless regime?", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698832730059}, {"id": "IDMaXkY5Mp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2514/Reviewer_QDAh"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Authors employ Transformer architecture to train a model capable of inferring a logical function based on a truth table.\n\nAuthors synthetically produce a dataset to train such a model.\n\nAuthors do evaluation in a noiseless and noisy setting.\nNoiseless means that the full truth table is available.\nNoisy means that a partial truth table is available, and some bits can be flipped with a small probability.\nThey show that the transformer effectively learned to reconstruct boolean functions.\nThey also show that it works in a noisy setting.\nThey evaluate their model in a realistic setting of gene regulatory networks, and showcase the use of their model;\nits performance is comparable to other methods while exhibiting much faster inference speed.", "review_text": "Authors employ Transformer architecture to train a model capable of inferring a logical function based on a truth table.\n\nAuthors synthetically produce a dataset to train such a model.\n\nAuthors do evaluation in a noiseless and noisy setting.\nNoiseless means that the full truth table is available.\nNoisy means that a partial truth table is available, and some bits can be flipped with a small probability.\nThey show that the transformer effectively learned to reconstruct boolean functions.\nThey also show that it works in a noisy setting.\nThey evaluate their model in a realistic setting of gene regulatory networks, and showcase the use of their model;\nits performance is comparable to other methods while exhibiting much faster inference speed.", "strengths": "- Well written and structured\n- Clear motivation\n- Authors will open source the implementation", "weaknesses": "- It would be good to have more examples of real-world usages of the method. Speed (inference) superiority is great, but maybe speed is not even a concern in the domains that the technique is intended to be applied.\n- It would be good to have some analysis on which type of tasks is solvable by this method vs. others.\n- Does not generalize to larger formulas.", "questions": "- Can you explain better Figure 17 from Supplementary Material which displays embeddings? \n    - (a) Which part of transformer do you extract for the shown embedding vectors (is it only the last token state, or all tokens, etc.)?\n    - (b) What exactly are the inputs that you use to construct shown embedding, and why do you make such choice?\n\n- Figure 1. Denote that what is shown is output of your model. Figure title is misleading.\n- Page 3. \"in the sections below\" -> \"in the following sections\".\n- Page 4. Maybe add that Smax <= Dmax\n- Page 5. D refers to dimensionality of logical input. Later in this page, it refers to a set of input-output pairs (if I understand correctly). Use a different letter.\n- Fig 7, part a. Readability can be improved.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors employ Transformer architecture to train a model capable of inferring a logical function based on a truth table.\n\nAuthors synthetically produce a dataset to train such a model.\n\nAuthors do evaluation in a noiseless and noisy setting.\nNoiseless means that the full truth table is available.\nNoisy means that a partial truth table is available, and some bits can be flipped with a small probability.\nThey show that the transformer effectively learned to reconstruct boolean functions.\nThey also show that it works in a noisy setting.\nThey evaluate their model in a realistic setting of gene regulatory networks, and showcase the use of their model;\nits performance is comparable to other methods while exhibiting much faster inference speed.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- Well written and structured\n- Clear motivation\n- Authors will open source the implementation", "weaknesses": "- It would be good to have more examples of real-world usages of the method. Speed (inference) superiority is great, but maybe speed is not even a concern in the domains that the technique is intended to be applied.\n- It would be good to have some analysis on which type of tasks is solvable by this method vs. others.\n- Does not generalize to larger formulas.", "questions": "- Can you explain better Figure 17 from Supplementary Material which displays embeddings? \n    - (a) Which part of transformer do you extract for the shown embedding vectors (is it only the last token state, or all tokens, etc.)?\n    - (b) What exactly are the inputs that you use to construct shown embedding, and why do you make such choice?\n\n- Figure 1. Denote that what is shown is output of your model. Figure title is misleading.\n- Page 3. \"in the sections below\" -> \"in the following sections\".\n- Page 4. Maybe add that Smax <= Dmax\n- Page 5. D refers to dimensionality of logical input. Later in this page, it refers to a set of input-output pairs (if I understand correctly). Use a different letter.\n- Fig 7, part a. Readability can be improved.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698756810200}, {"id": "g7J5lJciZ1", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2514/Reviewer_BZRx"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a method named Boolformer that performs symbolic regression for logical operations using AND, OR, and NOT. Boolformer is trained to output logic formulas in Polish notation using a Transformer-based Encoder-Decoder model. The results show not only the evaluation of the automatically generated formulas, but also the inference performance, excellent speed, and high explanatory power on PMLB databases and gene regulartory networks (GRNs) as real-world problems.", "review_text": "This paper proposes a method named Boolformer that performs symbolic regression for logical operations using AND, OR, and NOT. Boolformer is trained to output logic formulas in Polish notation using a Transformer-based Encoder-Decoder model. The results show not only the evaluation of the automatically generated formulas, but also the inference performance, excellent speed, and high explanatory power on PMLB databases and gene regulartory networks (GRNs) as real-world problems.", "strengths": "- A new problem setting in which logical expressions are symbolically regressed by Transformer.\n- Experimental results on several real-world applications as well as on the generated logical equation data are reported.\n  - The results using PMLB database shows accuracy comparable to Random Forest and logistic regression, and furthermore, the learned models are expected to provide excellent explanatory properties.\n  - Using GRNs, Boolformer is shown to have both excellent accuracy and speed.", "weaknesses": "- The problem statement in the introduction does not match the solution. In Section 1, citing (Abbe et al., 2022), the authors point out that Transformer learns complex models in terms of the Fourier spectrum, resulting in poor generalization performance when samples are insufficient. As a contribution, Section 1.1 claims that it is robust to noisy and incomplete observations. However, the Boolformer proposed in this paper is a relatively natural application of the Transformer, and there is no redesign from a Fourier spectrum perspective or other robustness innovations.\n- Due to the design of the method, it can only accept datasets with relatively few variables or small scale. This is acknowledged by the authors in section 5, but since there are currently proposals such as Transformer that can accept long series, it would have been easy to consider improving the limitation.\n- For example, if the correct answer is [AND, X_1, NOT, X_2], then [AND, NOT, X_2, X_1] is also equivalent. The fact that the system is learned by cross-entropy means that it is unclear how a valid cross-entropy can be calculated when there are multiple correct answers in this way.", "questions": "- The reviewer expects the authors to respond to the points listed in Weaknesses.\n- In the radar charts in Figure 7(a), the different methods are plotted among different axes, making it difficult to understand the comparison between those methods. If the radar chart is used to make comparisons between methods, it would be better to have as many axes as the number of experimental settings, such as the number of genes, and plot entities for each method. Alternatively, a table or a bar chart like Figure 7(b) is easier to compare methods.\n- Minor comments:\n  - In the caption of Figure 1, (x_5 x_6 x_7 x_7 x_9( should be (x_5 x_6 x_7 x_**8** x_9).\n  - References should be corrected. Especially, many published papers are cited as preprints. Below are some examples:\n    - The reference for (Abbe et al., 2022) should be a NeurIPS 2022 paper.\n    - (Dosovitskiy et al., 2020) should be an ICLR 2021 paper.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method named Boolformer that performs symbolic regression for logical operations using AND, OR, and NOT. Boolformer is trained to output logic formulas in Polish notation using a Transformer-based Encoder-Decoder model. The results show not only the evaluation of the automatically generated formulas, but also the inference performance, excellent speed, and high explanatory power on PMLB databases and gene regulartory networks (GRNs) as real-world problems.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- A new problem setting in which logical expressions are symbolically regressed by Transformer.\n- Experimental results on several real-world applications as well as on the generated logical equation data are reported.\n  - The results using PMLB database shows accuracy comparable to Random Forest and logistic regression, and furthermore, the learned models are expected to provide excellent explanatory properties.\n  - Using GRNs, Boolformer is shown to have both excellent accuracy and speed.", "weaknesses": "- The problem statement in the introduction does not match the solution. In Section 1, citing (Abbe et al., 2022), the authors point out that Transformer learns complex models in terms of the Fourier spectrum, resulting in poor generalization performance when samples are insufficient. As a contribution, Section 1.1 claims that it is robust to noisy and incomplete observations. However, the Boolformer proposed in this paper is a relatively natural application of the Transformer, and there is no redesign from a Fourier spectrum perspective or other robustness innovations.\n- Due to the design of the method, it can only accept datasets with relatively few variables or small scale. This is acknowledged by the authors in section 5, but since there are currently proposals such as Transformer that can accept long series, it would have been easy to consider improving the limitation.\n- For example, if the correct answer is [AND, X_1, NOT, X_2], then [AND, NOT, X_2, X_1] is also equivalent. The fact that the system is learned by cross-entropy means that it is unclear how a valid cross-entropy can be calculated when there are multiple correct answers in this way.", "questions": "- The reviewer expects the authors to respond to the points listed in Weaknesses.\n- In the radar charts in Figure 7(a), the different methods are plotted among different axes, making it difficult to understand the comparison between those methods. If the radar chart is used to make comparisons between methods, it would be better to have as many axes as the number of experimental settings, such as the number of genes, and plot entities for each method. Alternatively, a table or a bar chart like Figure 7(b) is easier to compare methods.\n- Minor comments:\n  - In the caption of Figure 1, (x_5 x_6 x_7 x_7 x_9( should be (x_5 x_6 x_7 x_**8** x_9).\n  - References should be corrected. Especially, many published papers are cited as preprints. Below are some examples:\n    - The reference for (Abbe et al., 2022) should be a NeurIPS 2022 paper.\n    - (Dosovitskiy et al., 2020) should be an ICLR 2021 paper.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698754220597}, {"id": "Qe5hQ5YXx3", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2514/Reviewer_V2Ca"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a model to perform symbolic regression of boolean functions using transformers", "review_text": "The paper proposes a model to perform symbolic regression of boolean functions using transformers", "strengths": "The method is sound, simple and elegant. \n\nBoolean functions are easy to be generated randomly in huge amount, which is where transformers architecture shine. \nThe  techniques is straightforward: generate boolean formulas randomly (covering the space of function as much as possible with a bias towards short formulas) and train a seq-2-seq architecture on that. \n\nThe speed up on gene-regulatory networks is interesting.", "weaknesses": "The major weakness of the paper is the positioning w.r.t. the state of the art. Being the method very simple, clearly highlighting the novelty should have been a priority. \n\nWhile the paper discusses many related areas, it is very unclear what is new in the proposed approach. For example, the section on \"symbolic regression\" in the related work, which is the closest area to the proposed approach (\"Symbolic regression of logic functions\"), is simply a list of papers. The approach is not compared with these approaches neither experimentally not even theoretically. \n\nExperiments in the noiseless regime do not compare Boolformer with any baseline (therefore is quite hard to understand how hard is the task overall). \n\nExperiments in the noisy regime have comparisons but with very unrelated approaches (generic ML models or specific to the dataset)\n\nMinor: The numbers in the radar charts in Figure 7 are impossible to read.", "questions": "1) Would be possible to apply any existing symbolic regression approaches to the proposed task?\n\n2) How novel is the generation of boolean formulas? Are there similar ideas in the literature to generate datasets for symbolic regression? \n\n3) How can you measure how hard is the task? Would any other method (both transformer based, or tradition ILP setting, be able to solve the task to some extent?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a model to perform symbolic regression of boolean functions using transformers", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The method is sound, simple and elegant. \n\nBoolean functions are easy to be generated randomly in huge amount, which is where transformers architecture shine. \nThe  techniques is straightforward: generate boolean formulas randomly (covering the space of function as much as possible with a bias towards short formulas) and train a seq-2-seq architecture on that. \n\nThe speed up on gene-regulatory networks is interesting.", "weaknesses": "The major weakness of the paper is the positioning w.r.t. the state of the art. Being the method very simple, clearly highlighting the novelty should have been a priority. \n\nWhile the paper discusses many related areas, it is very unclear what is new in the proposed approach. For example, the section on \"symbolic regression\" in the related work, which is the closest area to the proposed approach (\"Symbolic regression of logic functions\"), is simply a list of papers. The approach is not compared with these approaches neither experimentally not even theoretically. \n\nExperiments in the noiseless regime do not compare Boolformer with any baseline (therefore is quite hard to understand how hard is the task overall). \n\nExperiments in the noisy regime have comparisons but with very unrelated approaches (generic ML models or specific to the dataset)\n\nMinor: The numbers in the radar charts in Figure 7 are impossible to read.", "questions": "1) Would be possible to apply any existing symbolic regression approaches to the proposed task?\n\n2) How novel is the generation of boolean formulas? Are there similar ideas in the literature to generate datasets for symbolic regression? \n\n3) How can you measure how hard is the task? Would any other method (both transformer based, or tradition ILP setting, be able to solve the task to some extent?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698738577678}], "openreview_url": "https://openreview.net/forum?id=wmzFZ9lJrD", "arxiv_id": "2309.12207", "paper_pdf": "papers/wmzFZ9lJrD.pdf", "paper_pdf_sha256": "22c4c2ae011a1c33d46ae119e2a72eb4dc7f9d54f5ca9949b6ee788b776f8bf7", "paper_pdf_bytes": 993583, "paper_pdf_source": "openreview", "code_url": "https://github.com/sdascoli/boolformer", "code_repository": "sdascoli/boolformer", "code_commit": "b6694ba5eda0abb6484176d309937ff72c370fb9", "code_archive": "repos/wmzFZ9lJrD.zip", "code_archive_sha256": "fd4d441b560e5a5a2b45de5a0eb9804f97ebde27750eaa0a5748cfc60bee1cde", "code_archive_bytes": 90431, "code_file_count": 27, "code_extensions": {".py": 26, ".ipynb": 1}, "github_disk_usage_kb": 85, "github_languages": {"Python": 282780, "Jupyter Notebook": 54272}, "github_archived": false, "github_pushed_at": "2024-03-25T09:54:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/boolformer-symbolic-regression-of-logic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wV09GfqYC-n", "year": 2023, "status": "rejected", "title": "Hierarchies of Reward Machines", "authors": ["Daniel Furelos-Blanco", "Mark Law", "Anders Jonsson", "Krysia Broda", "Alessandra Russo"], "authorids": ["~Daniel_Furelos-Blanco1", "~Mark_Law1", "~Anders_Jonsson1", "~Krysia_Broda1", "~Alessandra_Russo1"], "authors_source": "OpenReview API", "abstract": "Reward machines (RMs) are a recent formalism for representing the reward function of a reinforcement learning task through a finite-state machine whose edges encode landmarks of the task using high-level events. The structure of RMs enables the decomposition of a task into simpler and independently solvable subtasks that help tackle long-horizon and/or sparse reward tasks. We propose a formalism for further abstracting the subtask structure by endowing an RM with the ability to call other RMs, thus composing a hierarchy of RMs (HRM). We exploit HRMs by treating each call to an RM as an independently solvable subtask using the options framework, and describe a curriculum-based method to learn HRMs from traces observed by the agent. Our experiments reveal that exploiting a handcrafted HRM leads to faster convergence than with a flat HRM, and that learning an HRM remains feasible in cases where its equivalent flat representation is not.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ZYrAhe0nmU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1346/Reviewer_8bbu"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper extends the Reward Machines (RM) formalism to hierarchical RMs (HRMs) by allowing RMs to call other RMs as sub-programs. The paper demonstrates that this allows for more efficient learning of RMs on long-horizon tasks in a number of toy grid-world environments.", "review_text": "Overall, I found the submission quite hard to read. I was not previously familiar with RMs and I found the heavy use of formalism and special terms throughout the submission hampered the reading flow. It also makes it hard for me to properly judge the novelty of the work: my perception is that the submission adds novelty over the previous RM work, but it also seems to make multiple assumptions that can reduce its impact outside the currently rather niche RM topic. The experimental evaluation does not help judge this either, since it is limited to toy environments and does not compare to non-RM works.\n\nI am leaning towards not accepting the submission, but am willing to change my mind if the authors can clearly state their assumptions and how they relate to other option learning papers (ie is there good use cases for their approach that others cannot do, can you compare to any non-RM approach?). Additionally, I am curious to know what other reviewers opinions are, especially for reviewers with a better overview of the RM literature.\n\n=============\nPost-Rebuttal: I updated my score to a weak accept based on the discussion with the other reviewers -- see my reply below.", "strengths": "# Strengths\n\n- incorporating hierarchy into RMs seems like an important step towards making them more applicable to long-horizon tasks\n\n- the submission rigorously defines all aspects of the problem setup\n\n- the toy experiments demonstrate that the introduced hierarchical RMs learn faster than regular (flat) RMs if a RM hierarchy is pre-defined, they also demonstrate learning of long-horizon tasks with HRMs (not possible with RMs)\n\n\n# Weaknesses\n\n(A) **hard to follow writing**: The writing of the paper is very hard to follow. I attribute this in large part to the very heavy usage of formalism and introduction of terms, which make smooth reading challenging. It seems that the paper could benefit a lot from replacing / augmenting some of the formal definitions with their plain-english equivalents. On other occasions crucial terms are not properly defined. Eg in the background section 2, the terms “propositions” and “labels” are not properly introduced, no intuition / example is given for what they could refer to. This makes it very hard to understand the paper for readers (like myself) that aren’t already familiar with the RM literature.\n\n(B) **assumptions not clearly mentioned**: compared to other papers that use (deep) RL to learn options & policies the submission seemingly makes a number of assumptions (given subtask hierarchies or at least a priori knowledge about depth of the hierarchy, symbolic observations, access to training tasks for *non-hierarchical* RMs, …). However, these assumptions are only step-by-step introduced throughout the text and can be easily missed. It would be good to add a “problem formulation” section or similar that explicitly lists the required assumptions for learning hierarchies of HRMs and how they compare to prior works, so that readers can better understand in what use cases they can consider HRMs.\n\n(C) **no comparisons to other option learning approaches**: the only comparison in the experimental section is to flat RMs — this is a meaningful comparison since the proposed approach is a hierarchical extension of flat RMs. However, since the paper proposes an approach for learning hierarchical options, it should compare to prior works on option learning with deep RL too, if applicable. E.g. the work of Sungryull Sohn could be applicable, since it also addresses hierarchical RL with discrete observational symbols (eg Sohn et al., NeurIPS 2018).\n\n(D) **only tested in toy environments**: the submission tests the proposed approach only in simple 2D grid-world like environments with low-dimensional observation and action spaces. This makes it hard to judge how well the method would scale to more realistic settings, eg. learning robot control in 3D environments.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper extends the Reward Machines (RM) formalism to hierarchical RMs (HRMs) by allowing RMs to call other RMs as sub-programs. The paper demonstrates that this allows for more efficient learning of RMs on long-horizon tasks in a number of toy grid-world environments.", "strength_and_weaknesses": "# Strengths\n\n- incorporating hierarchy into RMs seems like an important step towards making them more applicable to long-horizon tasks\n\n- the submission rigorously defines all aspects of the problem setup\n\n- the toy experiments demonstrate that the introduced hierarchical RMs learn faster than regular (flat) RMs if a RM hierarchy is pre-defined, they also demonstrate learning of long-horizon tasks with HRMs (not possible with RMs)\n\n\n# Weaknesses\n\n(A) **hard to follow writing**: The writing of the paper is very hard to follow. I attribute this in large part to the very heavy usage of formalism and introduction of terms, which make smooth reading challenging. It seems that the paper could benefit a lot from replacing / augmenting some of the formal definitions with their plain-english equivalents. On other occasions crucial terms are not properly defined. Eg in the background section 2, the terms “propositions” and “labels” are not properly introduced, no intuition / example is given for what they could refer to. This makes it very hard to understand the paper for readers (like myself) that aren’t already familiar with the RM literature.\n\n(B) **assumptions not clearly mentioned**: compared to other papers that use (deep) RL to learn options & policies the submission seemingly makes a number of assumptions (given subtask hierarchies or at least a priori knowledge about depth of the hierarchy, symbolic observations, access to training tasks for *non-hierarchical* RMs, …). However, these assumptions are only step-by-step introduced throughout the text and can be easily missed. It would be good to add a “problem formulation” section or similar that explicitly lists the required assumptions for learning hierarchies of HRMs and how they compare to prior works, so that readers can better understand in what use cases they can consider HRMs.\n\n(C) **no comparisons to other option learning approaches**: the only comparison in the experimental section is to flat RMs — this is a meaningful comparison since the proposed approach is a hierarchical extension of flat RMs. However, since the paper proposes an approach for learning hierarchical options, it should compare to prior works on option learning with deep RL too, if applicable. E.g. the work of Sungryull Sohn could be applicable, since it also addresses hierarchical RL with discrete observational symbols (eg Sohn et al., NeurIPS 2018).\n\n(D) **only tested in toy environments**: the submission tests the proposed approach only in simple 2D grid-world like environments with low-dimensional observation and action spaces. This makes it hard to judge how well the method would scale to more realistic settings, eg. learning robot control in 3D environments.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is rigorous but hard to read. It provides a clear extension of the prior RM work, but significance with respect to other option learning papers is not clearly established.", "summary_of_the_review": "Overall, I found the submission quite hard to read. I was not previously familiar with RMs and I found the heavy use of formalism and special terms throughout the submission hampered the reading flow. It also makes it hard for me to properly judge the novelty of the work: my perception is that the submission adds novelty over the previous RM work, but it also seems to make multiple assumptions that can reduce its impact outside the currently rather niche RM topic. The experimental evaluation does not help judge this either, since it is limited to toy environments and does not compare to non-RM works.\n\nI am leaning towards not accepting the submission, but am willing to change my mind if the authors can clearly state their assumptions and how they relate to other option learning papers (ie is there good use cases for their approach that others cannot do, can you compare to any non-RM approach?). Additionally, I am curious to know what other reviewers opinions are, especially for reviewers with a better overview of the RM literature.\n\n=============\nPost-Rebuttal: I updated my score to a weak accept based on the discussion with the other reviewers -- see my reply below.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666740827098}, {"id": "7KlW5TX3TY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1346/Reviewer_2jNi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper extends the concept of reward machines to hierarchical reward machines (HRM), and presents a method for learning policies for HRMs and for learning HRMs from data (learning is interleaved with policy optimization).  ", "review_text": "The paper makes relevant contributions, which, however, are currently not sufficiently well presented. There are several aspects that need to be clarified, for example, regarding the treatment of discounting and the possibility of executing the same RM several times during a single trajectory.\n", "strengths": "Strength:\n- Relevance: the paper tackles an important problem of learning and exploiting hierarchical task specifications.\n- Novelty: The presented formalism (HRM) and the methods for learning policy and HRMs are novel.\n\nWeaknesses:\n- Clarity: The paper is hard to follow (see \"Clarity\" below)\n- Significance: HRMs seem to suffer from too strong assumptions (see \"Quality\" below)", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper extends the concept of reward machines to hierarchical reward machines (HRM), and presents a method for learning policies for HRMs and for learning HRMs from data (learning is interleaved with policy optimization).  ", "strength_and_weaknesses": "Strength:\n- Relevance: the paper tackles an important problem of learning and exploiting hierarchical task specifications.\n- Novelty: The presented formalism (HRM) and the methods for learning policy and HRMs are novel.\n\nWeaknesses:\n- Clarity: The paper is hard to follow (see \"Clarity\" below)\n- Significance: HRMs seem to suffer from too strong assumptions (see \"Quality\" below)", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\nI found the paper rather difficult to follow, which is partly due to the fact that I am not familiar with reward machine and the notation used in the paper. However, I also think that the formalism is not well presented:\n\n- Section 2 states that $r(u, u')$ returns the reward for a transition of RM states. However, if I understand correctly, this reward should depend on the actual low-level transition (even with the assumed sparse reward structure, due to discounting). The RM paper (Toro Icarte et al. 2022) instead defines $\\delta_r(u, u')$ which returns a reward *function*, rather than a scalar reward. This deviation is not addressed in current submission.\n\n- I found it difficult to comprehend Figure 1b, in particular the \"not rabbit\" condition in the edge from $u_0^0$ to $u_0^1$, since a corresponding condition does not appear in the edge from $u_0^0$ to $u_0^2$. Only later, I realized that such condition needs to be added to preserve determinism, and that the choice for using the particular edge was arbitrary.\n\n- The paper does not seem to mention whether a tuple is removed from the call stack after a reward machine \"accepts\". If no tuples are removed from the call stack, the statement that \"each RM appears in the stack only once\" would indicate a major limitation, as it would not only imply that RM can not call themself recursively, but also that the same RM can not be used at several stages of a rollout.\n\n- The algorithms for learning policy and HRM are not sufficiently well discussed.  For learning the policies, I found the definition of the reward transition function confusing as it seems to ignore the effects of discounting. Furthermore, I do not see how the algorithm is capable of intra-option learning, given that it uses Semi-MDP Q-Learning and assumes options to terminate for updating the Q-function. I also found it confusing that the descriptions mentions DQN, although the experiments use grid worlds, which I assume used tabular Q functions.\nFor the algorithm for learning HRMs, I find it strange the paper only considers a multi-task setting, where *several* HRMs are learned. Wouldn't it be more sensible to consider learning of a single HRM for one task?  \n \nQuality:\nThe presented formalism of HRMs seems rather limited due to several strong assumptions:\n- The paper only considers sparse reward function, where the reward is 1 for goal states and 0 otherwise.\n- When learning HRMs, dead-end traces must be common across tasks and the depth of each hierarchy must be known a priori.\n- It is not clear to me, whether HRMs can correctly handle discounting, and whether the same RM can be used multiple times during a task (like the \"Navigate\" action in the Tax environment (Dietterich, 2000).  \n\nSeveral aspects of the formalism seem problematic:\n- The requirement that a trace uniquely specifies the option history (\"determinism\") can only be satisfied by introducing arbitrary preferences over subtasks, as in Fig. 1.b.\n- By assuming that every RM contains a single accepting state and a single rejecting state some tasks may require an exponential number of RMs.\n\nNovelty:\nThe formalism of HRMs is novel.\n\nReproducibility:\nThe paper does not provide source code nor sufficient details to reproduce the experiments. ", "summary_of_the_review": "The paper makes relevant contributions, which, however, are currently not sufficiently well presented. There are several aspects that need to be clarified, for example, regarding the treatment of discounting and the possibility of executing the same RM several times during a single trajectory.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666605802221}, {"id": "A4bMocrd0lZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1346/Reviewer_M97T"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "**Update after rebuttal** I think the other reviewers have raised some valid criticism, and have definitely helped to make the manuscript more accessible and improve readability. Having read the authors' response I think that all main issues have been addressed in a satisfactory manner (certainly true for the issues raised by me), and therefore remain in favor of accepting the paper.\n\n\nThis paper extends the Reward Machine framework by allowing for hierarchical compositions of reward machines. The basic idea for the construction is that transitions in one reward machine are captured by whole separate reward machines - a higher-level reward machine can call a lower-level reward machine, which allows for compositionality and re-use of reward machines. The paper also shows how to use hierarchical RL to learn policies from hierarchical reward machines, and how to synthesize hierarchical reward machines from data. Evaluations on simple environments prove that the concept works, and show some favorable results compared to flat reward machines. ", "review_text": "The paper constructs and analyzes an interesting and sensible hierarchical extension to the Reward Machines framework, which addresses important shortcomings and puts follow-up work on a solid theoretical footing. It also introduces two important and nontrivial algorithms for learning from HRMs and learning HRMs themselves from data. Overall I think this is a well executed and interesting extension to the Reward Machines framework, which has seen good adoption lately; the paper will be relevant to researchers in hierarchical RL and modeling of complex reward functions / specifications (which itself is a growing area of interest). The main idea of the extension is sensible and straightforward; the execution in the paper is very good and the precise theoretical construction presented is sound and interesting - this is a nontrivial contribution. My main criticism is that the paper is very dense and packs a lot of material into the page limit - while the first sections are very well written and can be well understood (with a background in basic Reward Machines), Section 4, 5, and 6 could be improved in terms of clarity and detail (but adding detail is hard given the already crammed paper). I think this is a high-quality, well-written and well executed paper that is ready for publication, and has good potential to stir follow-up theoretical work and almost certainly more investigations into learning with HRMs and synthesizing HRMs from data, and therefore currently suggest acceptance at the conference. \n", "strengths": "**Main contributions, Impact**\n\n1) Hierarchical extension of the Reward Machines framework. Though not entirely unexpected, the construction presented is solid and well executed. I would give this a medium to high impact since it puts a lot of potential follow-up work on solid footing and addresses important shortcomings of basic reward machines.\n2) Illustrative algorithm / proof of concept for learning policies with hierarchical reward machines. This certainly strengthens the paper since it allows for empirical results. I do not feel confident judging the sophistication of the algorithm, and suspect that there might be improvements in the near future inspired by this work. Impact: medium.\n3) Proof of concept algorithm for learning hierarchical reward machines. Another great point to strengthen the paper since it addresses the criticism of having to manually construct reward machines. Quite a number of choices went into the algorithm, and it currently also comes with quite a number of restrictions / assumptions. While it would be nice to see some alternatives explored and compared to, I think this is beyond the scope of this paper (and certainly beyond the page limit), and similarly to 2) I would expect more follow-up work along these lines to appear in the near future. Impact: medium to potentially high (e.g. for learning complex reward functions from human data).\n\n**Strengths**\n\n* Interesting and natural open problem in the Reward Machines framework addressed in a theoretically sound and solid fashion\n* Very well written intro and theory (up to Sec. 4)\n* Adding important bits for empirical evaluation: (i) adapting/suggesting a hierarchical RL algorithm that can learn with HRMs, (ii) suggesting an algorithm to learn HRMs from data. \n\n**Weaknesses**\n\n* Paper is very dense, this comes at the expense of clarity and sufficient detail in sections 4, 5 and 6. Both 4 and 5 (including some empirical results form 6) could be turned into separate full publications. While I appreciate having 4, 5, and 6 in the paper, and would not recommend removing them, the strict page limitations of a conference make some trade-offs necessary. Additionally, the paper comes already with an excessive appendix (which contains interesting parts).\n* Sections 4 and 5 propose two algorithms - both algorithms could have been designed differently, and there are some (more or less) obvious alternatives and ablations that would be nice to see; but the current paper simply does not have space for them (and I do not recommend filling up the appendix even more).\n\n**Improvements**\n1. There is not too much that can be done about a dense paper with an already long appendix, other than aiming for a journal paper with less severe page restrictions. While I think this would improve the paper, I also think that the current version is clearly above the threshold for publication at the conference, and I also think that publishing these results should not be delayed any longer. So I don’t expect to see an improvement here; maybe as a suggestion, the related work section could potentially be moved to the appendix to give 4 and 5 a bit more space.\n\n2. I think it is worth giving Sections 4 and 5 another pass for clarity - it currently seems that all the information is in place, but not presented in the most easily understandable way. Some overview/summary/illustration of the actual scheme proposed and the algorithms implemented for the empirical evaluation would be helpful. Having said that, the two sections do currently pass my threshold for acceptance, but I think could be improved for an excellent paper.\n\n3. Discussion of limitations of (L)HRMs is very short (determinism, having to define labeling functions). It would be nice to expand on this a little bit more, particularly compared to RMs.\n\n**Minor comments**\n\nA) Theorem 1: It would be nice to be more precise what “equivalent” means in the main paper (e.g. equivalent: an automaton that accepts the same language (language = set of all possible label traces \\lambda that respect the MDP transition dynamics)).\n\nB) Learning HRMs - why select tasks and instances with lower returns first? Why not choose highest returns first and switch once the increase in returns starts to saturate (i.e. learn the easiest tasks first, then switch to the next hardest ones once there is no progress on the easy ones anymore).\n\nC) Learning HRMs - how exactly are learned options re-used? “An option is selected from a\npool of options appearing in lower height RMs”; how is this selection done, probabilistically or by exhaustive evaluation and some deterministic selection criterion?\n\nD) Not sure if the experiments are optimal to show the advantages of non-flat HRMs. I would imagine non-flat HRMs to perform even better in cases where a simple task (such as gathering paper) must be repeated multiple times (e.g. to create a composite object of ‘book’ and ‘map’, ‘paper’ needs to be gathered twice). The current tasks (at first glance) seem to be chosen to collect a series of objects once. I might have missed that the same advantage would play out in flat HRMs (which are different from plain RMs?), so please correct me if this is wrong.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "**Update after rebuttal** I think the other reviewers have raised some valid criticism, and have definitely helped to make the manuscript more accessible and improve readability. Having read the authors' response I think that all main issues have been addressed in a satisfactory manner (certainly true for the issues raised by me), and therefore remain in favor of accepting the paper.\n\n\nThis paper extends the Reward Machine framework by allowing for hierarchical compositions of reward machines. The basic idea for the construction is that transitions in one reward machine are captured by whole separate reward machines - a higher-level reward machine can call a lower-level reward machine, which allows for compositionality and re-use of reward machines. The paper also shows how to use hierarchical RL to learn policies from hierarchical reward machines, and how to synthesize hierarchical reward machines from data. Evaluations on simple environments prove that the concept works, and show some favorable results compared to flat reward machines. ", "strength_and_weaknesses": "**Main contributions, Impact**\n\n1) Hierarchical extension of the Reward Machines framework. Though not entirely unexpected, the construction presented is solid and well executed. I would give this a medium to high impact since it puts a lot of potential follow-up work on solid footing and addresses important shortcomings of basic reward machines.\n2) Illustrative algorithm / proof of concept for learning policies with hierarchical reward machines. This certainly strengthens the paper since it allows for empirical results. I do not feel confident judging the sophistication of the algorithm, and suspect that there might be improvements in the near future inspired by this work. Impact: medium.\n3) Proof of concept algorithm for learning hierarchical reward machines. Another great point to strengthen the paper since it addresses the criticism of having to manually construct reward machines. Quite a number of choices went into the algorithm, and it currently also comes with quite a number of restrictions / assumptions. While it would be nice to see some alternatives explored and compared to, I think this is beyond the scope of this paper (and certainly beyond the page limit), and similarly to 2) I would expect more follow-up work along these lines to appear in the near future. Impact: medium to potentially high (e.g. for learning complex reward functions from human data).\n\n**Strengths**\n\n* Interesting and natural open problem in the Reward Machines framework addressed in a theoretically sound and solid fashion\n* Very well written intro and theory (up to Sec. 4)\n* Adding important bits for empirical evaluation: (i) adapting/suggesting a hierarchical RL algorithm that can learn with HRMs, (ii) suggesting an algorithm to learn HRMs from data. \n\n**Weaknesses**\n\n* Paper is very dense, this comes at the expense of clarity and sufficient detail in sections 4, 5 and 6. Both 4 and 5 (including some empirical results form 6) could be turned into separate full publications. While I appreciate having 4, 5, and 6 in the paper, and would not recommend removing them, the strict page limitations of a conference make some trade-offs necessary. Additionally, the paper comes already with an excessive appendix (which contains interesting parts).\n* Sections 4 and 5 propose two algorithms - both algorithms could have been designed differently, and there are some (more or less) obvious alternatives and ablations that would be nice to see; but the current paper simply does not have space for them (and I do not recommend filling up the appendix even more).\n\n**Improvements**\n1. There is not too much that can be done about a dense paper with an already long appendix, other than aiming for a journal paper with less severe page restrictions. While I think this would improve the paper, I also think that the current version is clearly above the threshold for publication at the conference, and I also think that publishing these results should not be delayed any longer. So I don’t expect to see an improvement here; maybe as a suggestion, the related work section could potentially be moved to the appendix to give 4 and 5 a bit more space.\n\n2. I think it is worth giving Sections 4 and 5 another pass for clarity - it currently seems that all the information is in place, but not presented in the most easily understandable way. Some overview/summary/illustration of the actual scheme proposed and the algorithms implemented for the empirical evaluation would be helpful. Having said that, the two sections do currently pass my threshold for acceptance, but I think could be improved for an excellent paper.\n\n3. Discussion of limitations of (L)HRMs is very short (determinism, having to define labeling functions). It would be nice to expand on this a little bit more, particularly compared to RMs.\n\n**Minor comments**\n\nA) Theorem 1: It would be nice to be more precise what “equivalent” means in the main paper (e.g. equivalent: an automaton that accepts the same language (language = set of all possible label traces \\lambda that respect the MDP transition dynamics)).\n\nB) Learning HRMs - why select tasks and instances with lower returns first? Why not choose highest returns first and switch once the increase in returns starts to saturate (i.e. learn the easiest tasks first, then switch to the next hardest ones once there is no progress on the easy ones anymore).\n\nC) Learning HRMs - how exactly are learned options re-used? “An option is selected from a\npool of options appearing in lower height RMs”; how is this selection done, probabilistically or by exhaustive evaluation and some deterministic selection criterion?\n\nD) Not sure if the experiments are optimal to show the advantages of non-flat HRMs. I would imagine non-flat HRMs to perform even better in cases where a simple task (such as gathering paper) must be repeated multiple times (e.g. to create a composite object of ‘book’ and ‘map’, ‘paper’ needs to be gathered twice). The current tasks (at first glance) seem to be chosen to collect a series of objects once. I might have missed that the same advantage would play out in flat HRMs (which are different from plain RMs?), so please correct me if this is wrong.\n", "clarity,_quality,_novelty_and_reproducibility": "For both clarity and quality the whole paper passes my personal threshold for publication. Having said that, I think the first half of the paper is very well written and of high quality (many important theor. questions and relations answered), whereas 4, 5, and 6 could be slightly improved for an excellent paper. Readers that are completely unfamiliar with the Reward Machines framework might have a harder time with the first part of the paper, but I do not see an easy way to fix that (the paper does not have space for a general RM intro).\n\nThe method and algorithms in the paper are to the best of my knowledge novel and original. The extensive and detailed appendix helps greatly with reproducibility.\n", "summary_of_the_review": "The paper constructs and analyzes an interesting and sensible hierarchical extension to the Reward Machines framework, which addresses important shortcomings and puts follow-up work on a solid theoretical footing. It also introduces two important and nontrivial algorithms for learning from HRMs and learning HRMs themselves from data. Overall I think this is a well executed and interesting extension to the Reward Machines framework, which has seen good adoption lately; the paper will be relevant to researchers in hierarchical RL and modeling of complex reward functions / specifications (which itself is a growing area of interest). The main idea of the extension is sensible and straightforward; the execution in the paper is very good and the precise theoretical construction presented is sound and interesting - this is a nontrivial contribution. My main criticism is that the paper is very dense and packs a lot of material into the page limit - while the first sections are very well written and can be well understood (with a background in basic Reward Machines), Section 4, 5, and 6 could be improved in terms of clarity and detail (but adding detail is hard given the already crammed paper). I think this is a high-quality, well-written and well executed paper that is ready for publication, and has good potential to stir follow-up theoretical work and almost certainly more investigations into learning with HRMs and synthesizing HRMs from data, and therefore currently suggest acceptance at the conference. \n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666305448390}], "openreview_url": "https://openreview.net/forum?id=wV09GfqYC-n", "arxiv_id": "2205.15752", "paper_pdf": "papers/wV09GfqYC-n.pdf", "paper_pdf_sha256": "be58b64292249a699825a03d77d3210862712a9a5c32555336a5aa322ffe5717", "paper_pdf_bytes": 1794830, "paper_pdf_source": "openreview", "code_url": "https://github.com/ertsiger/hrm-learning", "code_repository": "ertsiger/hrm-learning", "code_commit": "23d0e25a367e1e8a559fa815d86696f0183ade43", "code_archive": "repos/wV09GfqYC-n.zip", "code_archive_sha256": "ef38c08fa0db18cad6f74d482dd09d0376eff55b4af29e7645cc41f5240f337a", "code_archive_bytes": 201897, "code_file_count": 84, "code_extensions": {".py": 77, ".sh": 6, ".rb": 1}, "github_disk_usage_kb": 158, "github_languages": {"Python": 579478, "Shell": 1255, "Ruby": 202}, "github_archived": false, "github_pushed_at": "2023-07-11T13:25:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hierarchies-of-reward-machines"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "T_8wHvOkEi9", "year": 2022, "status": "rejected", "title": "Self-Organized Polynomial-time Coordination Graphs", "authors": ["Weijun Dong", "Qianlan Yang", "Zhizhou Ren", "Jianhao Wang", "Tonghan Wang", "Chongjie Zhang"], "authorids": ["~Weijun_Dong1", "~Qianlan_Yang1", "~Zhizhou_Ren1", "~Jianhao_Wang1", "~Tonghan_Wang1", "~Chongjie_Zhang1"], "authors_source": "OpenReview API", "abstract": "Coordination graph is a promising approach to model agent collaboration in multi-agent reinforcement learning. It factorizes a large multi-agent system into a suite of overlapping groups that represent the underlying coordination dependencies. One critical challenge in this paradigm is the complexity of computing maximum-value actions for a graph-based value factorization. It refers to the decentralized constraint optimization problem (DCOP), which and whose constant-ratio approximation are NP-hard problems. To bypass this fundamental hardness, this paper proposes a novel method, named Self-Organized Polynomial-time Coordination Graphs (SOP-CG), which uses structured graph classes to guarantee the optimality of the induced DCOPs with sufficient function expressiveness. We extend the graph topology to be state-dependent, formulate the graph selection as an imaginary agent, and finally derive an end-to-end learning paradigm from the unified Bellman optimality equation. In experiments, we show that our approach learns interpretable graph topologies, induces effective coordination, and improves performance across a variety of cooperative multi-agent tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "LTVTWMDEFp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2761/Reviewer_Jx51"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors present an imaginary coordinator agent that executes graph selection as its action based on possible combinations of pairwise edges and the underlying individual and pairwise utilities. The results show some effectiveness of the presented method, SOP-CG, in simple examples.", "review_text": "The main strength of the paper lies in its algorithm being of polynomial-time nature, and that characteristic needs much more emphasis in both text and figures. Does SOP-CG reach peak performance faster? Is its peak performance higher? could be some guideline questions when structuring the paper contents around the polynomial-time algorithm.\n\nQ1. I would appreciate if the authors would discuss the relevance of a WWW 2020 paper: How Much and When Do We Need Higher-order Information in Hypergraphs? A Case Study on Hyperedge Prediction by Yoon et al. I think it could also provide some useful insights as to how high of an order the hyperedges need to be, to save some function expressiveness (i.e., representational capacity) at the cost of time complexity.\n\nQ2. How did the authors go about designing the didactic examples? How simple of a task is the \"sweet spot\" for SOP-CG? As task complexity is altered, when does SOP-CG begin losing its simplicity advantage? When do other models take over? Which measures would you say most greatly determine the task complexity, which in turn, compromises SOP-CG's performance?\n\nQ3. How would you position SOP-CG in the MARL research? Since the graph selector agent is there at both training at execution times, would SOP-CG be a centralized training centralized execution work? If it is, comparison against CTDE works such as VDN and QMIX would not be fair. It is as though a \"free\" centralized coordinator is helping out the SOP-CG agents at execution time as well. It may be a good idea to compare against some communication-enabled MARL works.\n\nQ4. Despite starting out strong, the paper seems to fall off rather dramatically, especially when it comes to the didactic nature of the examples, tied in with the page 6 remark that SOP-CG would perform best (given its tradeoff) in tasks where the restricted graph classes are enough to express the coordination dependencies. This remark really sounds like going back on the VDN-QMIX-QTRAN line of research, whose focus was about covering a richer class of joint action-value functions. Going back on that trend now only to pursue the polynomial-time nature of the running algorithm would in my opinion require far more diverse evaluation examples, backed by a stronger motivation highlighting real-world threats of all the other MARL algorithms taking longer than polynomial time. As is, SOP-CG does not contend amazingly against other MARL algorithms that chose the \"NP-hard? Curse of dimensionality? Fine. We'll approximate, approximate, approximate.\" path rather than the \"Polynomial time is our topmost priority; function expressiveness can wait.\" path. That leads me back to the question of why pursue polynomial time at the cost of losing both the function expressiveness and the peak performance in the apparent trilemma.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Authors present an imaginary coordinator agent that executes graph selection as its action based on possible combinations of pairwise edges and the underlying individual and pairwise utilities. The results show some effectiveness of the presented method, SOP-CG, in simple examples.", "main_review": "The main strength of the paper lies in its algorithm being of polynomial-time nature, and that characteristic needs much more emphasis in both text and figures. Does SOP-CG reach peak performance faster? Is its peak performance higher? could be some guideline questions when structuring the paper contents around the polynomial-time algorithm.\n\nQ1. I would appreciate if the authors would discuss the relevance of a WWW 2020 paper: How Much and When Do We Need Higher-order Information in Hypergraphs? A Case Study on Hyperedge Prediction by Yoon et al. I think it could also provide some useful insights as to how high of an order the hyperedges need to be, to save some function expressiveness (i.e., representational capacity) at the cost of time complexity.\n\nQ2. How did the authors go about designing the didactic examples? How simple of a task is the \"sweet spot\" for SOP-CG? As task complexity is altered, when does SOP-CG begin losing its simplicity advantage? When do other models take over? Which measures would you say most greatly determine the task complexity, which in turn, compromises SOP-CG's performance?\n\nQ3. How would you position SOP-CG in the MARL research? Since the graph selector agent is there at both training at execution times, would SOP-CG be a centralized training centralized execution work? If it is, comparison against CTDE works such as VDN and QMIX would not be fair. It is as though a \"free\" centralized coordinator is helping out the SOP-CG agents at execution time as well. It may be a good idea to compare against some communication-enabled MARL works.\n\nQ4. Despite starting out strong, the paper seems to fall off rather dramatically, especially when it comes to the didactic nature of the examples, tied in with the page 6 remark that SOP-CG would perform best (given its tradeoff) in tasks where the restricted graph classes are enough to express the coordination dependencies. This remark really sounds like going back on the VDN-QMIX-QTRAN line of research, whose focus was about covering a richer class of joint action-value functions. Going back on that trend now only to pursue the polynomial-time nature of the running algorithm would in my opinion require far more diverse evaluation examples, backed by a stronger motivation highlighting real-world threats of all the other MARL algorithms taking longer than polynomial time. As is, SOP-CG does not contend amazingly against other MARL algorithms that chose the \"NP-hard? Curse of dimensionality? Fine. We'll approximate, approximate, approximate.\" path rather than the \"Polynomial time is our topmost priority; function expressiveness can wait.\" path. That leads me back to the question of why pursue polynomial time at the cost of losing both the function expressiveness and the peak performance in the apparent trilemma.", "summary_of_the_review": "My biggest concern is the \"imaginary\" agent freely collecting information, making decisions, and delivering those decisions to all the agents at both the training and execution times. Comparison against the chosen baselines is not fair, even when the chosen evaluation task is a simple enough one, in which SOP-CG's limited representational capacity would not appear that pronounced.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636398007094}, {"id": "qfZH63yEJkY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2761/Reviewer_xmMA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an extension of deep coordination graph, called Self-Organized Polynomial-time Coordination Graphs (SOP-CG). Instead of pre-specified graph topology used in DCG, their method allows graph topology to be state-dependent, which is achieved by a coordinator agent, and the optimization of this agent is incorporated in a modified temporal difference learning paradigm. Two pre-specified undirected acyclic graph classes are used to ensure polynomial-time graph selection and accurate greedy action selection. The result on sensor network, grid world and MPE shows that such a trade-off between the representational capacity of graph topology and the computational accuracy can improve the performance of MARL and learn meaningful graph topology.", "review_text": "**Strengths**\n1. To determine state-dependent coordination graph is interesting.\n2. Incorporating graph selection into TD learning is desirable. \n\n**Concerns/Questions**\n1. State-dependent coordination graph needs to be determined at each time step in both training and execution,  which means SOP-CG is a centralized method. Thus, a baseline of single-agent RL is desired for comparison in experiments. \n2. It is not clear how $q_i$ and $q_{ij}$ are learned. Are they parameter-sharing for agents?\n3. The operation $\\arg\\max$ in equation 4 takes $O(n^3)$ for $\\mathcal{G}_P $. It may be too costly for both training and execution. One evidence is that each run of SOP-CG  takes up to 2.5 days in these simple experimental tasks. \n4. Since both DCG and CASEC include SMAC experiments, it would be better to also include it here to show the performance of SOP-CG in complex environments. \n\nIn summary, it is currently hard to see the benefit of determining the coordination graph in a centralized way. Moreover, the proposed method is not verified in complex environments. \n\n\n\n**Minor Comments**\n1. As claimed in Appendix C, a graph relabeling technique is used to solve the extra overestimation error introduced by additional max operator over graphs.  However, this paper is currently missing an ablation to validate this point.\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an extension of deep coordination graph, called Self-Organized Polynomial-time Coordination Graphs (SOP-CG). Instead of pre-specified graph topology used in DCG, their method allows graph topology to be state-dependent, which is achieved by a coordinator agent, and the optimization of this agent is incorporated in a modified temporal difference learning paradigm. Two pre-specified undirected acyclic graph classes are used to ensure polynomial-time graph selection and accurate greedy action selection. The result on sensor network, grid world and MPE shows that such a trade-off between the representational capacity of graph topology and the computational accuracy can improve the performance of MARL and learn meaningful graph topology.", "main_review": "**Strengths**\n1. To determine state-dependent coordination graph is interesting.\n2. Incorporating graph selection into TD learning is desirable. \n\n**Concerns/Questions**\n1. State-dependent coordination graph needs to be determined at each time step in both training and execution,  which means SOP-CG is a centralized method. Thus, a baseline of single-agent RL is desired for comparison in experiments. \n2. It is not clear how $q_i$ and $q_{ij}$ are learned. Are they parameter-sharing for agents?\n3. The operation $\\arg\\max$ in equation 4 takes $O(n^3)$ for $\\mathcal{G}_P $. It may be too costly for both training and execution. One evidence is that each run of SOP-CG  takes up to 2.5 days in these simple experimental tasks. \n4. Since both DCG and CASEC include SMAC experiments, it would be better to also include it here to show the performance of SOP-CG in complex environments. \n\nIn summary, it is currently hard to see the benefit of determining the coordination graph in a centralized way. Moreover, the proposed method is not verified in complex environments. \n\n\n\n**Minor Comments**\n1. As claimed in Appendix C, a graph relabeling technique is used to solve the extra overestimation error introduced by additional max operator over graphs.  However, this paper is currently missing an ablation to validate this point.\n\n\n\n", "summary_of_the_review": "State-dependent coordination graph is important. However, the paper currently has several weaknesses as mentioned above. It seems clearly below the bar of ICLR.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635827402626}, {"id": "95O6OnyD2W6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2761/Reviewer_h3B4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduced a novel method called Self-Organized Polynomial-time Coordination Graphs (SOP-CG), aiming to handle the decentralized constraint optimization problem (DCOP). This paper is well organized and the experiments are explicitly presented. Therefore, I think the work of this paper is very interesting and the contributions are sufficient. \n", "review_text": "The detailed comments regarding the quality of this paper are listed as follows:\n1. In the Introduction section, one of the biggest concerns is that this paper cannot tell the novelty of the proposed SOP-CG compared with the methods in peer researches. Besides, the illustration of experimental results is not convincible to prove the advantages of the proposed SOP-CG in this section.\n2. Somewhere in the paper, the authors' presentation is unclear and is worth being improved. For example, in the Background section, the meanings of some symbols in the model are not clear. Furthermore, the authors should give the purposes of introducing formula (2).\n3. What is the intuition behind the process of investigating polynomial-time coordination graphs in the Section 4? Or, is there any intuition at all? How did you come up with that idea? Have you borrowed this idea from somewhere else? It is better to give a detailed explanation about polynomial-time coordination graphs.\n4. In page 6, the authors proposed Self-Organized Polynomial-Time Coordination Graphs. Whatever techniques are used in the manuscript, there is a need to tabulate computational cost of the proposed algorithm in this paper. \n5. In the Experiments Sections, the authors claimed that the graph structures learned by SOP-CG definitely match the ground-truth demands for effective collaboration. However, it is not clear to us why the method can be used in demonstrating the ability of the proposed approach to organizing coordination relations. Therefore, the authors should provide a detailed explanation about the above issue. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduced a novel method called Self-Organized Polynomial-time Coordination Graphs (SOP-CG), aiming to handle the decentralized constraint optimization problem (DCOP). This paper is well organized and the experiments are explicitly presented. Therefore, I think the work of this paper is very interesting and the contributions are sufficient. \n", "main_review": "The detailed comments regarding the quality of this paper are listed as follows:\n1. In the Introduction section, one of the biggest concerns is that this paper cannot tell the novelty of the proposed SOP-CG compared with the methods in peer researches. Besides, the illustration of experimental results is not convincible to prove the advantages of the proposed SOP-CG in this section.\n2. Somewhere in the paper, the authors' presentation is unclear and is worth being improved. For example, in the Background section, the meanings of some symbols in the model are not clear. Furthermore, the authors should give the purposes of introducing formula (2).\n3. What is the intuition behind the process of investigating polynomial-time coordination graphs in the Section 4? Or, is there any intuition at all? How did you come up with that idea? Have you borrowed this idea from somewhere else? It is better to give a detailed explanation about polynomial-time coordination graphs.\n4. In page 6, the authors proposed Self-Organized Polynomial-Time Coordination Graphs. Whatever techniques are used in the manuscript, there is a need to tabulate computational cost of the proposed algorithm in this paper. \n5. In the Experiments Sections, the authors claimed that the graph structures learned by SOP-CG definitely match the ground-truth demands for effective collaboration. However, it is not clear to us why the method can be used in demonstrating the ability of the proposed approach to organizing coordination relations. Therefore, the authors should provide a detailed explanation about the above issue. \n", "summary_of_the_review": "The work presented is indeed interesting and relevant to the real scenarios. I recommend that this paper could be accepted.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635617917835}], "openreview_url": "https://openreview.net/forum?id=T_8wHvOkEi9", "arxiv_id": "2112.03547", "paper_pdf": "papers/T_8wHvOkEi9.pdf", "paper_pdf_sha256": "b1ed82171cfecca4284a251ffc81c4070d62ecb3ec5061dc7ddd770b3af53c8b", "paper_pdf_bytes": 1481340, "paper_pdf_source": "openreview", "code_url": "https://github.com/yanQval/SOP-CG", "code_repository": "yanQval/SOP-CG", "code_commit": "8a719a11c0cf318ba5e39f803441cb99b79815b5", "code_archive": "repos/T_8wHvOkEi9.zip", "code_archive_sha256": "f006bc1fbe392a9bec534431bc32fbd17686478d2f5cca48d6a3b7a1c67809f3", "code_archive_bytes": 117928, "code_file_count": 64, "code_extensions": {".py": 58, ".sh": 4, ".cpp": 2}, "github_disk_usage_kb": 261, "github_languages": {"Python": 274940, "C++": 24804, "Shell": 1921, "Dockerfile": 1796}, "github_archived": false, "github_pushed_at": "2022-06-23T02:21:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/self-organized-polynomial-time-coordination-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eyXknI5scWu", "year": 2021, "status": "rejected", "title": "Investigating and Simplifying Masking-based Saliency Methods for Model Interpretability", "authors": ["Jason Phang", "Jungkyu Park", "Krzysztof J. Geras"], "authorids": ["~Jason_Phang1", "~Jungkyu_Park1", "~Krzysztof_J._Geras1"], "authors_source": "OpenReview API", "abstract": "Saliency maps that identify the most informative regions of an image for a classifier are valuable for model interpretability. A common approach to creating saliency maps involves generating input masks that mask out portions of an image to maximally deteriorate classification performance, or mask in an image to preserve classification performance. Many variants of this approach have been proposed in the literature, such as counterfactual generation and optimizing over a Gumbel-Softmax distribution. Using a general formulation of masking-based saliency methods, we conduct an extensive evaluation study of a number of recently proposed variants to understand which elements of these methods meaningfully improve performance. Surprisingly, we find that a well-tuned, relatively simple formulation of a masking-based saliency model outperforms many more complex approaches. We find that the most important ingredients for high quality saliency map generation are (1) using both masked-in and masked-out objectives and (2) training the classifier alongside the masking model. Strikingly, we show that a masking model can be trained with as few as 10 examples per class and still generate saliency maps with only a 0.7-point increase in localization error.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tfVaT9VxvyX", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2163/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "-Summary: \nThis paper investigates the previous masking-based saliency map methods. By detailly formulate the masking-based saliency methods and sufficient experimentation, the paper gives a simple formulation and practical training strategies.\n\n-Strength: \nThe paper is well organized and the presentation is easy to follow.\nThe experiments are comprehensive.\n\n-Weakness:\nAlthough the experiments are enough, the paper simply adopts the previous masking-based methods, evaluation metrics, architectures, and sanity check analysis. The theoretical analysis is insufficient.  More theoretical experiments instead of performance analysis need to be added.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper has sufficient experiments on strategies of masking-based saliency methods , but lacks theoretical analysis.", "review": "-Summary: \nThis paper investigates the previous masking-based saliency map methods. By detailly formulate the masking-based saliency methods and sufficient experimentation, the paper gives a simple formulation and practical training strategies.\n\n-Strength: \nThe paper is well organized and the presentation is easy to follow.\nThe experiments are comprehensive.\n\n-Weakness:\nAlthough the experiments are enough, the paper simply adopts the previous masking-based methods, evaluation metrics, architectures, and sanity check analysis. The theoretical analysis is insufficient.  More theoretical experiments instead of performance analysis need to be added.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603962609557}, {"id": "qYzAHWLf7C-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2163/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an approach to improve masked based prediction explanation. They train an auxiliary model which predicts a mask that must satisfy two terms 1) maximize classification accuracy when applied to the image and 2) maximize entropy over softmax when the inverse mask is applied to the image (they also experiment with minimizing the classification accuracy instead here).\n\nThe paper also positions their work to be classifier agnostic in the text which is not clear to me. I think this aspect is a negative because it is not in the list of contributions and seems like a distraction of a remnant of an old direction of the paper:\n\n> \"Whereas a FIX approach seeks a saliency map that explains what regions are most salient to a given classifier, a CA approach tries to identify all possible salient regions for any hypothetical classifier (hence, classifier-agnostic). In other words, a CA approach may be inadequate for interpreting a specific classifier and is better suited for identifying salient regions for a class of image classification models.\" \n\nI'll focus now on studying how this paper addresses the 4 contributions they claim in the intro:\n\n> (1) We find that incorporating both masked-in classification maximization and masked-out entropy maximization objectives leads to the best saliency maps, and continually training the classifier improves the quality of generated maps. \n\nThere are no standard deviations reported from multiple runs so nothing can be statistically claimed. Also, looking at table 1 this is not clear. It seems MinClass works the best compared to almost all other methods in terms of PxAP. In terms of other metrics it seems like no best can be determined.\n\n> (2) We find that the masking model requires only the top layers of the classifier to effectively generate saliency maps. \n\nThis is reported in the table. It is not clear if this is a strong contribution as it would just be specific to this method and without a standard deviation we cannot conclude anything.\n\n> (3) Our final model outperforms other masking-based methods on WSOL and PxAP metrics. \n\nThe reported difference between the methods is 48.6 vs 48.4.\n\n> (4) We find that a small number of examples—as few as ten per class—is sufficient to train a masker to within the ballpark of our best performing model.\nThis is reported but the paper doesn't have a section detailing the experiments or showing how this number is derived.\n\n\nThe paper could be improved by refining the contributions and detailing what evidence should be observed to support these claims and then providing a significant amount of evidence. Right now the paper is not focused in general and does not focus on supporting the claims made in the introduction and therefore is not ready for publication.\t", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unsupported claims", "review": "The paper proposes an approach to improve masked based prediction explanation. They train an auxiliary model which predicts a mask that must satisfy two terms 1) maximize classification accuracy when applied to the image and 2) maximize entropy over softmax when the inverse mask is applied to the image (they also experiment with minimizing the classification accuracy instead here).\n\nThe paper also positions their work to be classifier agnostic in the text which is not clear to me. I think this aspect is a negative because it is not in the list of contributions and seems like a distraction of a remnant of an old direction of the paper:\n\n> \"Whereas a FIX approach seeks a saliency map that explains what regions are most salient to a given classifier, a CA approach tries to identify all possible salient regions for any hypothetical classifier (hence, classifier-agnostic). In other words, a CA approach may be inadequate for interpreting a specific classifier and is better suited for identifying salient regions for a class of image classification models.\" \n\nI'll focus now on studying how this paper addresses the 4 contributions they claim in the intro:\n\n> (1) We find that incorporating both masked-in classification maximization and masked-out entropy maximization objectives leads to the best saliency maps, and continually training the classifier improves the quality of generated maps. \n\nThere are no standard deviations reported from multiple runs so nothing can be statistically claimed. Also, looking at table 1 this is not clear. It seems MinClass works the best compared to almost all other methods in terms of PxAP. In terms of other metrics it seems like no best can be determined.\n\n> (2) We find that the masking model requires only the top layers of the classifier to effectively generate saliency maps. \n\nThis is reported in the table. It is not clear if this is a strong contribution as it would just be specific to this method and without a standard deviation we cannot conclude anything.\n\n> (3) Our final model outperforms other masking-based methods on WSOL and PxAP metrics. \n\nThe reported difference between the methods is 48.6 vs 48.4.\n\n> (4) We find that a small number of examples—as few as ten per class—is sufficient to train a masker to within the ballpark of our best performing model.\nThis is reported but the paper doesn't have a section detailing the experiments or showing how this number is derived.\n\n\nThe paper could be improved by refining the contributions and detailing what evidence should be observed to support these claims and then providing a significant amount of evidence. Right now the paper is not focused in general and does not focus on supporting the claims made in the introduction and therefore is not ready for publication.\t", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603946014628}, {"id": "KODoHhOBji", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2163/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \n\nBy the first look, this work itself does not introduce any new architecture or novel algorithm. It takes what is considered as the popular choices in generating classifier saliency masks, and conducts quite extensive sets of experiments to dissect the components by their importance. The writing is pretty clear in narrative and the experimental findings are surprising and significant. \n\nGood things to mention:\n\n1. The importance of using CA seems quite evident from Table 1.\n2. The importance of using multiple resolutions is also evident.\n3. The findings of few shot learning ability for masker is quite interesting, class diversity seems to play a very import role.\n\nSome questions:\n1. Given the fact that all experiments are done with ImageNet, which is widely considered as a solved task with almost super-human classifiers, do you think the superb performance of those classifiers could inflate your findings? Do you think the results in Table 1 and Figure 5 would still hold consistently when the classifier is not as performant?\n2. I am not sure I understand the point of the figure on the right of Figure 4. What is the take-way message from it?\n3. Compare (e,f,i,j) in Table 1, it seems I does not make a difference as long as O is used. What is the reason there?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Thorough experimentation and important insight on the matter of deriving saliency from trained classifiers", "review": "Summary: \n\nBy the first look, this work itself does not introduce any new architecture or novel algorithm. It takes what is considered as the popular choices in generating classifier saliency masks, and conducts quite extensive sets of experiments to dissect the components by their importance. The writing is pretty clear in narrative and the experimental findings are surprising and significant. \n\nGood things to mention:\n\n1. The importance of using CA seems quite evident from Table 1.\n2. The importance of using multiple resolutions is also evident.\n3. The findings of few shot learning ability for masker is quite interesting, class diversity seems to play a very import role.\n\nSome questions:\n1. Given the fact that all experiments are done with ImageNet, which is widely considered as a solved task with almost super-human classifiers, do you think the superb performance of those classifiers could inflate your findings? Do you think the results in Table 1 and Figure 5 would still hold consistently when the classifier is not as performant?\n2. I am not sure I understand the point of the figure on the right of Figure 4. What is the take-way message from it?\n3. Compare (e,f,i,j) in Table 1, it seems I does not make a difference as long as O is used. What is the reason there?", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603848148848}, {"id": "2u5QpP13al0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2163/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "/*************************Post-Rebuttal**********************/\nThe authors address many of my concerns well, and I agree with their rebuttal.\n\nThe modified manuscript also looks good, too.\n\nI raise my rating.\n\n/*************************Pre-Rebuttal**********************/\n\nPros.:\n1. The proposed method is simple and reasonable but better than the current state-of-the-art methods.\n2. The paper is clearly presented and well-organized.\n4. The authors conduct many ablation studies to validate the different variants of the proposed methods. \n5. The few-shot localization built in the proposed method is also interesting, and worth being explored.\n\n\nCons.:\n1. Limited Novelty\n\nThe proposed method is similar to [Ref. 1, 2], but there is no discussion in the related work. After reading this paper and [Ref. 1, 2], I don't see any significant difference in [Ref. 1, 2], the authors focus on top-down saliency detection, and use the masked image scheme to train the network. They proposed a CNN framework that contains two CNN modules, an image-level classifier and a pixel-level map generator. They also use the saliency from the map generator as the classifier's input and then optimize the classifier scores to derive the saliency map and train the generator. In this paper, the loss is L2 loss and binary-class setting, but I think the main idea is almost the same. Therefore, I think the novelty is very limited. I would like to see that the authors point the significant difference between the proposed method and [Ref. 1, 2] and give a detailed discussion.\n\n[Ref. 1] Hsu et al., Weakly Supervised Saliency Detection with A Category-Driven Map Generator, BMVC'17\n\n[Ref. 2] Hsu et al., Weakly Supervised Salient Object Detection by Learning A Classifier-Driven Map Generator, TIP'19\n\n2. No significant improvement\n\nCompared to Zolna et al. (2020), the improvement of the proposed method is not significant, 48.6 v.s. 48.4 in OM,  36.1 v.s. 35.8 in LE, 61.4 v.s. 62.0 in F1. I think this improvement is minor. \n\n3. About few-shot experiments\n\nThe few-shot experiments are interesting, but why aren't other methods conducted for these experiments? The authors claim that the few samples are sufficient to train the proposed method, but other methods may also require a small number of training samples to achieve similar performance. Besides, I am not clear why the proposed method could work well under the few-shot setting because there is no specific module or design for the few-shot setting.   Could the authors explain it in detail?\n\n4. Other issue:\n\nWhat is the difference between the proposed method and other weakly supervised or unsupervised object saliency detection, such as [Ref 3, 4, 5, 6, 7]? In these works, their goal is also to train a network to predict a class-agnostic saliency map and highlight the most salient object in the images under the unsupervised or weakly supervised setting. Therefore, I would like to know the major difference between the tasks. If they are similar to the solved problem, the discussion and comparison should be conducted.\n\n[Ref. 3] Zhang et al., Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector, ICCV'17\n\n[Ref. 4] Li et al., Weakly supervised salient object detection using image labels, AAAI'19\n\n[Ref. 5] Wang et al., Learning to detect salient objects with image-level supervision, CVPR'17\n\n[Ref. 6] Zhang et al., Deep unsupervised saliency detection: A multiple noisy labeling perspective, CVPR'18\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposed method is simpler and better than the previous methods, but my concerns are the limited novelty and no significant improvement", "review": "/*************************Post-Rebuttal**********************/\nThe authors address many of my concerns well, and I agree with their rebuttal.\n\nThe modified manuscript also looks good, too.\n\nI raise my rating.\n\n/*************************Pre-Rebuttal**********************/\n\nPros.:\n1. The proposed method is simple and reasonable but better than the current state-of-the-art methods.\n2. The paper is clearly presented and well-organized.\n4. The authors conduct many ablation studies to validate the different variants of the proposed methods. \n5. The few-shot localization built in the proposed method is also interesting, and worth being explored.\n\n\nCons.:\n1. Limited Novelty\n\nThe proposed method is similar to [Ref. 1, 2], but there is no discussion in the related work. After reading this paper and [Ref. 1, 2], I don't see any significant difference in [Ref. 1, 2], the authors focus on top-down saliency detection, and use the masked image scheme to train the network. They proposed a CNN framework that contains two CNN modules, an image-level classifier and a pixel-level map generator. They also use the saliency from the map generator as the classifier's input and then optimize the classifier scores to derive the saliency map and train the generator. In this paper, the loss is L2 loss and binary-class setting, but I think the main idea is almost the same. Therefore, I think the novelty is very limited. I would like to see that the authors point the significant difference between the proposed method and [Ref. 1, 2] and give a detailed discussion.\n\n[Ref. 1] Hsu et al., Weakly Supervised Saliency Detection with A Category-Driven Map Generator, BMVC'17\n\n[Ref. 2] Hsu et al., Weakly Supervised Salient Object Detection by Learning A Classifier-Driven Map Generator, TIP'19\n\n2. No significant improvement\n\nCompared to Zolna et al. (2020), the improvement of the proposed method is not significant, 48.6 v.s. 48.4 in OM,  36.1 v.s. 35.8 in LE, 61.4 v.s. 62.0 in F1. I think this improvement is minor. \n\n3. About few-shot experiments\n\nThe few-shot experiments are interesting, but why aren't other methods conducted for these experiments? The authors claim that the few samples are sufficient to train the proposed method, but other methods may also require a small number of training samples to achieve similar performance. Besides, I am not clear why the proposed method could work well under the few-shot setting because there is no specific module or design for the few-shot setting.   Could the authors explain it in detail?\n\n4. Other issue:\n\nWhat is the difference between the proposed method and other weakly supervised or unsupervised object saliency detection, such as [Ref 3, 4, 5, 6, 7]? In these works, their goal is also to train a network to predict a class-agnostic saliency map and highlight the most salient object in the images under the unsupervised or weakly supervised setting. Therefore, I would like to know the major difference between the tasks. If they are similar to the solved problem, the discussion and comparison should be conducted.\n\n[Ref. 3] Zhang et al., Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector, ICCV'17\n\n[Ref. 4] Li et al., Weakly supervised salient object detection using image labels, AAAI'19\n\n[Ref. 5] Wang et al., Learning to detect salient objects with image-level supervision, CVPR'17\n\n[Ref. 6] Zhang et al., Deep unsupervised saliency detection: A multiple noisy labeling perspective, CVPR'18\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603723128656}], "openreview_url": "https://openreview.net/forum?id=eyXknI5scWu", "arxiv_id": "2010.09750", "paper_pdf": "papers/eyXknI5scWu.pdf", "paper_pdf_sha256": "84babe42358de7a4757106491e86ae6c0cc7aa0ac6f52014b5395b88ca2f4e73", "paper_pdf_bytes": 2657504, "paper_pdf_source": "openreview", "code_url": "https://github.com/zphang/saliency_investigation", "code_repository": "zphang/saliency_investigation", "code_commit": "cdd03e92c4b8b26c43a60cb50f9d97ea2c7f9400", "code_archive": "repos/eyXknI5scWu.zip", "code_archive_sha256": "efa2002bc966d5830eaae746e5f096b162df4c28b3a9048e8774df33527093d4", "code_archive_bytes": 314668, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 300, "github_languages": {"Python": 194492}, "github_archived": false, "github_pushed_at": "2020-11-10T01:21:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/investigating-and-simplifying-masking-based-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJgS7p4FPH", "year": 2020, "status": "rejected", "title": "Accelerating Reinforcement Learning Through GPU Atari Emulation", "authors": ["Steven Dalton", "Michael Garland", "Iuri Frosio"], "authorids": ["sdalton@nvidia.com", "mgarland@nvidia.com", "ifrosio@nvidia.com"], "authors_source": "OpenReview API", "abstract": "We introduce CuLE (CUDA Learning Environment), a CUDA port of the Atari Learning Environment (ALE) which is used for the development of deep reinforcement algorithms.  CuLE overcomes many limitations of existing CPU-based emulators and scales naturally to multiple GPUs. It leverages GPU parallelization to run thousands of games simultaneously and it renders frames directly on the GPU, to avoid the bottleneck arising from the limited CPU-GPU communication bandwidth. CuLE generates up to 155M frames per hour on a single GPU, a finding previously achieved only through a cluster of CPUs. Beyond highlighting the differences between CPU and GPU emulators in the context of reinforcement learning, we show how to leverage the high throughput of CuLE by effective batching of the training data, and show accelerated convergence for A2C+V-trace. CuLE is available at [hidden URL].", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJgL80qkqr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper448/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes a port of the Atari Learning Environment to CUDA, reports on a set of performance comparison, and provides a bottleneck analysis based communication bandwidth and various throughputs required to saturate them for training and inference.\n\nMy first reaction to this paper was, \"So what?\"; but as I read more, I like the paper more and more.  It was the bottleneck analysis that changed my mind.  It was done very thoroughly and it provides deep insight in the challenges that RL faces for both learning and inference in a variety of settings.  I especially liked the analysis of the advantages and limitations of GPU emulation.  I also thought the Discussion section was well written.\n\nThe paper would be better if:\n1) The figure fonts were larger throughout the paper.\n2) The gaps in Table 1 were explained.\n\nMinor issue:  Change \"feed\" to \"fed\" on page 3.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper describes a port of the Atari Learning Environment to CUDA, reports on a set of performance comparison, and provides a bottleneck analysis based communication bandwidth and various throughputs required to saturate them for training and inference.\n\nMy first reaction to this paper was, \"So what?\"; but as I read more, I like the paper more and more.  It was the bottleneck analysis that changed my mind.  It was done very thoroughly and it provides deep insight in the challenges that RL faces for both learning and inference in a variety of settings.  I especially liked the analysis of the advantages and limitations of GPU emulation.  I also thought the Discussion section was well written.\n\nThe paper would be better if:\n1) The figure fonts were larger throughout the paper.\n2) The gaps in Table 1 were explained.\n\nMinor issue:  Change \"feed\" to \"fed\" on page 3.\n"}, "tcdate": 1571954253523}, {"id": "ryl7uG2atH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper448/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The work contributes a library emulating Atari games in GPU in parallel and allowing to speed-up the execution of reinforcement learning algorithms.\n\nI see that the paper qualifies to the conference; in particular there is listed the topic:\n\n- “implementation issues, parallelization, software platforms, hardware”\n\nHowever, this is not a research paper, and I do not really see how I should asses it. What I can say about it is that it is considerable amount of work, not only implementing the simulator but also looking at what RL methods need, and how to optimize the allocation and exchange of the data so that everything would work on GPU more efficiently.\n\nFrom the practical perspective, I am somewhat confused. The speed-up factors in the experiments are rather modest: about 4x for simulating and rendering frames, 2.5x for full RL, on a single GPU. Better with scaling to multi-GPU systems. In Table 1 the total training time per resources used differs dramatically. However if I look at the lines with A2C it is about the same time with 100-200 CPU cores + 1 GPU versus 12 cores + 1 GPU. So this is about factor 10 in the resources, versus CPU parallelization probably suffering overheads.\n\nIt appears that the maximum steed-ups are achieved for a particular type of the reinforcement learning algorithms, and using it in a general case would give a modest improvement.\n\nThe paper itself consists of introduction, related work, 1 page overview of what it means to simulate the Atari games, and experiments. So it is mostly about measuring the speedups, with several implementations / platforms.\n\nI tend to think that this work will not very much boost the research for new RL methods. It is limited to Atari games, mostly helps to sample-inefficient RL methods and if it helps, the speed-up factors are not of the order that would make experiments by the researchers otherwise impossible. \n\nI would also give priority to theoretical contributions at ICLR. In the end, we all are using CUDA and cnDNN, but presentations about how they implement things are rather given at GPU computing conferences. \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "1: Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "The work contributes a library emulating Atari games in GPU in parallel and allowing to speed-up the execution of reinforcement learning algorithms.\n\nI see that the paper qualifies to the conference; in particular there is listed the topic:\n\n- “implementation issues, parallelization, software platforms, hardware”\n\nHowever, this is not a research paper, and I do not really see how I should asses it. What I can say about it is that it is considerable amount of work, not only implementing the simulator but also looking at what RL methods need, and how to optimize the allocation and exchange of the data so that everything would work on GPU more efficiently.\n\nFrom the practical perspective, I am somewhat confused. The speed-up factors in the experiments are rather modest: about 4x for simulating and rendering frames, 2.5x for full RL, on a single GPU. Better with scaling to multi-GPU systems. In Table 1 the total training time per resources used differs dramatically. However if I look at the lines with A2C it is about the same time with 100-200 CPU cores + 1 GPU versus 12 cores + 1 GPU. So this is about factor 10 in the resources, versus CPU parallelization probably suffering overheads.\n\nIt appears that the maximum steed-ups are achieved for a particular type of the reinforcement learning algorithms, and using it in a general case would give a modest improvement.\n\nThe paper itself consists of introduction, related work, 1 page overview of what it means to simulate the Atari games, and experiments. So it is mostly about measuring the speedups, with several implementations / platforms.\n\nI tend to think that this work will not very much boost the research for new RL methods. It is limited to Atari games, mostly helps to sample-inefficient RL methods and if it helps, the speed-up factors are not of the order that would make experiments by the researchers otherwise impossible. \n\nI would also give priority to theoretical contributions at ICLR. In the end, we all are using CUDA and cnDNN, but presentations about how they implement things are rather given at GPU computing conferences. \n\n\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571828330893}, {"id": "ByxxYQ73FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper448/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a CUDA port of the Atari Learning Environment. The paper goes into detail examining the benefits that come from a GPU-only implementation, including much better per-GPU utilization as well as no need to run a distributed system of CPUs. They show this hardware scaling can be taken advantage of across a variety of state of the art reinforcement learning algorithms and indeed create new batching strategies to utilize their framework and the GPU better. \n\nThe paper is well written and goes into some detail describing the implementation of CuLE as well as various design decisions taken, as with splitting the emulation process across several kernels. Finally, the paper is very explicit about a number of optimizations that are not being exploited by the new framework and serve as markers for future work.\n\nA question that arises and which is not addressed in the experiments is how the authors verified their port is faithful to the original version; there is no mention of correctness in the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper introduces a CUDA port of the Atari Learning Environment. The paper goes into detail examining the benefits that come from a GPU-only implementation, including much better per-GPU utilization as well as no need to run a distributed system of CPUs. They show this hardware scaling can be taken advantage of across a variety of state of the art reinforcement learning algorithms and indeed create new batching strategies to utilize their framework and the GPU better. \n\nThe paper is well written and goes into some detail describing the implementation of CuLE as well as various design decisions taken, as with splitting the emulation process across several kernels. Finally, the paper is very explicit about a number of optimizations that are not being exploited by the new framework and serve as markers for future work.\n\nA question that arises and which is not addressed in the experiments is how the authors verified their port is faithful to the original version; there is no mention of correctness in the paper."}, "tcdate": 1571726200156}], "openreview_url": "https://openreview.net/forum?id=HJgS7p4FPH", "arxiv_id": "1907.08467", "paper_pdf": "papers/HJgS7p4FPH.pdf", "paper_pdf_sha256": "8a0b050e67f648b9da67b126db5cfe3c63e762ef40fcae0a2d1baa403f306acd", "paper_pdf_bytes": 2961951, "paper_pdf_source": "openreview", "code_url": "https://github.com/NVlabs/cule", "code_repository": "NVlabs/cule", "code_commit": "dd0382b99ded6be23cd3c3e79e37938e7c873de0", "code_archive": "repos/HJgS7p4FPH.zip", "code_archive_sha256": "e4d47e43e37b366ca7ad4068342ec3da4f8e0d3511e70385ea3211edc3435490", "code_archive_bytes": 321653, "code_file_count": 155, "code_extensions": {".hpp": 117, ".py": 31, ".cpp": 6, ".cu": 1}, "github_disk_usage_kb": 418, "github_languages": {"C++": 564093, "Python": 32210, "Cuda": 8165, "Dockerfile": 1177}, "github_archived": false, "github_pushed_at": "2022-11-21T13:38:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gpu-accelerated-atari-emulation-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "K7UfQFegK5", "year": 2026, "status": "rejected", "title": "FullFront: Benchmarking MLLMs Across the Full Front-End Engineering Workflow", "authors": ["Haoyu Sun", "Huichen Will Wang", "Jiawei Gu", "Linjie Li", "Yu Cheng"], "authorids": ["~Haoyu_Sun2", "~Huichen_Will_Wang1", "~Jiawei_Gu4", "~Linjie_Li1", "~Yu_Cheng1"], "authors_source": "OpenReview API", "abstract": "Front-end engineering involves a complex workflow where engineers conceptualize designs, translate them into code, and iteratively refine the implementation. While recent benchmarks primarily focus on converting visual designs to code, we present FullFront, a benchmark designed to evaluate Multimodal Large Language Models (MLLMs) \\textbf{across the full front-end development pipeline}. \nFullFront assesses three fundamental tasks that map directly to the front-end engineering pipeline: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase).\nUnlike existing benchmarks that use either scraped websites with bloated code or oversimplified LLM-generated HTML, FullFront employs a novel, two-stage process to transform real-world webpages into clean, standardized HTML while maintaining diverse visual designs and avoiding copyright issues.\nExtensive testing of state-of-the-art MLLMs reveals significant limitations in page perception, code generation (particularly for image handling and layout), and interaction implementation. Our results quantitatively demonstrate performance disparities across models and tasks, and highlight a substantial gap between current MLLM capabilities and human expert performance in front-end engineering.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "b1cgerAAvw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7038/Reviewer_fHMJ"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "FullFront is a new benchmark consisting of three types of tasks related to various aspects of the front-end development cycle, including: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase).\n\nWebpage Design is an image generation task where the model needs to generate the webpage design based on textual descriptions of synthetic webpages. Webpage perception QA is a QA task based on both real and synthetic webpages, including 75 samples of multi-window QA tasks. The questions are generated by GPT-4o with manual inspection. The webpage code generation task takes visual design as input and expects code generation as output. This task includes image-to-code, text-to-code, interaction authoring, and code refinement. \n\nThe authors did benchmarking of various closed and open models on these tasks.", "review_text": "FullFront is a new benchmark consisting of three types of tasks related to various aspects of the front-end development cycle, including: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase).\n\nWebpage Design is an image generation task where the model needs to generate the webpage design based on textual descriptions of synthetic webpages. Webpage perception QA is a QA task based on both real and synthetic webpages, including 75 samples of multi-window QA tasks. The questions are generated by GPT-4o with manual inspection. The webpage code generation task takes visual design as input and expects code generation as output. This task includes image-to-code, text-to-code, interaction authoring, and code refinement. \n\nThe authors did benchmarking of various closed and open models on these tasks.", "strengths": "- New resource for benchmarking MLLM capability on front-end engineering. \n\n- Relatively thorough coverage of different capabilities and model families. \n\n- Nice to include to human evaluation too.", "weaknesses": "- I'm not fully convinced by why we need an aggregate benchmark for front-end development. Some of the capabilities are quite distinct. For example, design (image) generation vs QA vs code generation. You have to use different models for the benchmarking because most models can't do image generation at all. Then what's the point of putting all of these tasks into one benchmark?\n\n- I understand the authors put in effort to curate new data for many of the tasks in the benchmark. But I believe for most of the sub-tasks, there exist prior benchmarks that test the same capability. What's the unique contribution here apart from putting all the result tables into one paper?", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "FullFront is a new benchmark consisting of three types of tasks related to various aspects of the front-end development cycle, including: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase).\n\nWebpage Design is an image generation task where the model needs to generate the webpage design based on textual descriptions of synthetic webpages. Webpage perception QA is a QA task based on both real and synthetic webpages, including 75 samples of multi-window QA tasks. The questions are generated by GPT-4o with manual inspection. The webpage code generation task takes visual design as input and expects code generation as output. This task includes image-to-code, text-to-code, interaction authoring, and code refinement. \n\nThe authors did benchmarking of various closed and open models on these tasks.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- New resource for benchmarking MLLM capability on front-end engineering. \n\n- Relatively thorough coverage of different capabilities and model families. \n\n- Nice to include to human evaluation too.", "weaknesses": "- I'm not fully convinced by why we need an aggregate benchmark for front-end development. Some of the capabilities are quite distinct. For example, design (image) generation vs QA vs code generation. You have to use different models for the benchmarking because most models can't do image generation at all. Then what's the point of putting all of these tasks into one benchmark?\n\n- I understand the authors put in effort to curate new data for many of the tasks in the benchmark. But I believe for most of the sub-tasks, there exist prior benchmarks that test the same capability. What's the unique contribution here apart from putting all the result tables into one paper?", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762110498943}, {"id": "9un6Rg67Jz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7038/Reviewer_gXbA"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces FullFront, a benchmark that evaluates MLLMs across the full front-end workflow—design, perception, and code—using reconstructed real-page data and complementary visual/code metrics validated against human judgments. Experiments reveal substantial gaps from human performance, with persistent weaknesses in fine-grained perception, layout fidelity, image handling, and interaction implementation despite strong results from leading proprietary models. The work is timely and likely useful for the community, though its reliance on closed models and the potential metric bias should be weighed carefully.", "review_text": "This paper introduces FullFront, a benchmark that evaluates MLLMs across the full front-end workflow—design, perception, and code—using reconstructed real-page data and complementary visual/code metrics validated against human judgments. Experiments reveal substantial gaps from human performance, with persistent weaknesses in fine-grained perception, layout fidelity, image handling, and interaction implementation despite strong results from leading proprietary models. The work is timely and likely useful for the community, though its reliance on closed models and the potential metric bias should be weighed carefully.", "strengths": "+ The benchmark mirrors real front-end workflows rather than a single slice: it covers conceptualization (Webpage Design), perception (Webpage Perception QA), and implementation (Webpage Code Generation) with concrete task counts across eight subtasks. This end-to-end framing is rare and useful for diagnosing capability gaps.\n+ The dataset is grounded in real webpages and reconstructed into standardized, copyright-safe HTML through a two-stage, MLLM-assisted pipeline, addressing common issues of bloated scraped code and oversimplified LLM HTML in prior corpora.\n+ The evaluation combines visual and code-level metrics and validates them against human judgments with high Spearman correlations, supporting automated evaluation as a reasonable proxy for human preference.", "weaknesses": "- Several construction and scoring steps rely on proprietary models (e.g., GPT-4o/Claude in the pipeline; Gemini-based visual scoring), which can introduce system bias and limit strict reproducibility; data/code release is contingent on acceptance.\n- The engineered “Code Score” aggregates DOM and style attributes with fixed design choices; even with strong human correlation, such choices may privilege particular implementation patterns and under-reward acceptable alternatives.\n- The results clearly show large model–human gaps (e.g., <60% model accuracy vs. >95% human on perception), but the causal link between perception skill and code quality remains only lightly probed; deeper ablations could clarify what actually transfers across stages.", "questions": "1. How reproducible is the dataset and evaluation given reliance on closed models (GPT-4o/Claude for curation; Gemini for visual scoring)? Please quantify any bias toward these systems and provide an open, drop-in alternative or calibration protocol.\n\n\n2. “Code Score” encodes specific structural/style weights. What sensitivity and failure analyses show that valid alternative implementations aren’t systematically penalized? Include counterexamples where humans deem outputs equivalent but the metric disagrees.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces FullFront, a benchmark that evaluates MLLMs across the full front-end workflow—design, perception, and code—using reconstructed real-page data and complementary visual/code metrics validated against human judgments. Experiments reveal substantial gaps from human performance, with persistent weaknesses in fine-grained perception, layout fidelity, image handling, and interaction implementation despite strong results from leading proprietary models. The work is timely and likely useful for the community, though its reliance on closed models and the potential metric bias should be weighed carefully.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "+ The benchmark mirrors real front-end workflows rather than a single slice: it covers conceptualization (Webpage Design), perception (Webpage Perception QA), and implementation (Webpage Code Generation) with concrete task counts across eight subtasks. This end-to-end framing is rare and useful for diagnosing capability gaps.\n+ The dataset is grounded in real webpages and reconstructed into standardized, copyright-safe HTML through a two-stage, MLLM-assisted pipeline, addressing common issues of bloated scraped code and oversimplified LLM HTML in prior corpora.\n+ The evaluation combines visual and code-level metrics and validates them against human judgments with high Spearman correlations, supporting automated evaluation as a reasonable proxy for human preference.", "weaknesses": "- Several construction and scoring steps rely on proprietary models (e.g., GPT-4o/Claude in the pipeline; Gemini-based visual scoring), which can introduce system bias and limit strict reproducibility; data/code release is contingent on acceptance.\n- The engineered “Code Score” aggregates DOM and style attributes with fixed design choices; even with strong human correlation, such choices may privilege particular implementation patterns and under-reward acceptable alternatives.\n- The results clearly show large model–human gaps (e.g., <60% model accuracy vs. >95% human on perception), but the causal link between perception skill and code quality remains only lightly probed; deeper ablations could clarify what actually transfers across stages.", "questions": "1. How reproducible is the dataset and evaluation given reliance on closed models (GPT-4o/Claude for curation; Gemini for visual scoring)? Please quantify any bias toward these systems and provide an open, drop-in alternative or calibration protocol.\n\n\n2. “Code Score” encodes specific structural/style weights. What sensitivity and failure analyses show that valid alternative implementations aren’t systematically penalized? Include counterexamples where humans deem outputs equivalent but the metric disagrees.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762077426983}, {"id": "hta59L4iK9", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7038/Reviewer_JYpf"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This work proposes FullFront, a unified benchmark evaluating the full front-end engineering pipeline: Webpage Design, Webpage Perception QA, and Webpage Code Generation.", "review_text": "This work proposes FullFront, a unified benchmark evaluating the full front-end engineering pipeline: Webpage Design, Webpage Perception QA, and Webpage Code Generation.", "strengths": "1. Compared to prior works, FullFront transforms real-world websites into clean, standardized HTML to avoid copyright issues.\n2. Many models are comprehensively benchmarked on three subtasks.", "weaknesses": "1. While the authors claim they unify multiple components into one cohesive evaluation pipeline, the implementation and results look like three separate benchmarks to me; the analysis of the connection between different parts seems weak.\n2. Webpage Design seems to benchmark text-to-image generation capability. This part is kinda less motivated, since why do we want such MLLM to generate a website in image form? What makes it necessary to ask them to generate an image instead of generating code + rendering?\n3. Misleading sample size: It seems that a non-trivial amount of experiments is conducted on FullFront-mini, which only contains 10 webpage design data points and 50 webpage code generation data points. The selection process is not justified.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes FullFront, a unified benchmark evaluating the full front-end engineering pipeline: Webpage Design, Webpage Perception QA, and Webpage Code Generation.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Compared to prior works, FullFront transforms real-world websites into clean, standardized HTML to avoid copyright issues.\n2. Many models are comprehensively benchmarked on three subtasks.", "weaknesses": "1. While the authors claim they unify multiple components into one cohesive evaluation pipeline, the implementation and results look like three separate benchmarks to me; the analysis of the connection between different parts seems weak.\n2. Webpage Design seems to benchmark text-to-image generation capability. This part is kinda less motivated, since why do we want such MLLM to generate a website in image form? What makes it necessary to ask them to generate an image instead of generating code + rendering?\n3. Misleading sample size: It seems that a non-trivial amount of experiments is conducted on FullFront-mini, which only contains 10 webpage design data points and 50 webpage code generation data points. The selection process is not justified.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761865439373}, {"id": "Du9XY1XH0s", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7038/Reviewer_1Ttx"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "Proposes a full-pipeline benchmark for front-end work with three tasks: Webpage Design, Perception QA, and Code Generation. Datasets span hundreds to thousands of items across these stages.", "review_text": "Proposes a full-pipeline benchmark for front-end work with three tasks: Webpage Design, Perception QA, and Code Generation. Datasets span hundreds to thousands of items across these stages.", "strengths": "1. Evaluates the full front-end pipeline from concept, perception, and implementation (3 tasks, 8 sub-tasks).\n2. Combines visual similarity with a Code Score to evaluate fine-grained elements.\n3. Benchmarks a wide set of open-source and proprietary models and analyzes task-level differences.", "weaknesses": "Minor weaknesses:\n1. Typos and labeling inconsistency.\n2. Numerical inconsistencies in reported results and category counts.\n\nMajor issues:\n1. Incomplete specification of the Code Score.\n2. Same-family bias from using a model as the judge while evaluating related models.\n3. Data-pipeline bias due to using certain models to construct the dataset.", "questions": "1. Typo: In Fig.6, \"InterenVL3-78B\" should be \"InternVL3-78B\" (Line 435); In the caption of Table 6, \"Blacnk\" and \"Isonlation\" (Line447-448). In Table 1, \"DiINOv2\" (Line272). \"outperforme\" in Line 107.\n2. Unreasonable numbers: In Table 3, the GPT-5 just got 0.58 on Img of Gemini Visual Score (Line 307); The authors mentioned \"15 categories\", but claimed \"twelve categories\" in Figure 31.\n3. Missing weight values: the actual weights of Code Score are missing. I am concerned about the reproducibility.\n4. Model-as-judge bias: using Gemini 2.5 Flash as the visual judge while also evaluating Gemini models risks same-family bias.\n5. Data source with bias: HTML-v1 built by GPT-4o and refined to HTML-v2 by Claude 3.7 Sonnet, then those models are evaluated. The evaluation metrics may favor Claude and GPT-4o.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Proposes a full-pipeline benchmark for front-end work with three tasks: Webpage Design, Perception QA, and Code Generation. Datasets span hundreds to thousands of items across these stages.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. Evaluates the full front-end pipeline from concept, perception, and implementation (3 tasks, 8 sub-tasks).\n2. Combines visual similarity with a Code Score to evaluate fine-grained elements.\n3. Benchmarks a wide set of open-source and proprietary models and analyzes task-level differences.", "weaknesses": "Minor weaknesses:\n1. Typos and labeling inconsistency.\n2. Numerical inconsistencies in reported results and category counts.\n\nMajor issues:\n1. Incomplete specification of the Code Score.\n2. Same-family bias from using a model as the judge while evaluating related models.\n3. Data-pipeline bias due to using certain models to construct the dataset.", "questions": "1. Typo: In Fig.6, \"InterenVL3-78B\" should be \"InternVL3-78B\" (Line 435); In the caption of Table 6, \"Blacnk\" and \"Isonlation\" (Line447-448). In Table 1, \"DiINOv2\" (Line272). \"outperforme\" in Line 107.\n2. Unreasonable numbers: In Table 3, the GPT-5 just got 0.58 on Img of Gemini Visual Score (Line 307); The authors mentioned \"15 categories\", but claimed \"twelve categories\" in Figure 31.\n3. Missing weight values: the actual weights of Code Score are missing. I am concerned about the reproducibility.\n4. Model-as-judge bias: using Gemini 2.5 Flash as the visual judge while also evaluating Gemini models risks same-family bias.\n5. Data source with bias: HTML-v1 built by GPT-4o and refined to HTML-v2 by Claude 3.7 Sonnet, then those models are evaluated. The evaluation metrics may favor Claude and GPT-4o.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761849950460}], "openreview_url": "https://openreview.net/forum?id=K7UfQFegK5", "arxiv_id": "2505.17399", "paper_pdf": "papers/K7UfQFegK5.pdf", "paper_pdf_sha256": "a6f661cd662ed4bd071548625f672830ecf075c4a5f0dd7d753c3db0a9c11974", "paper_pdf_bytes": 11101829, "paper_pdf_source": "openreview", "code_url": "https://github.com/Mikivishy/FullFront", "code_repository": "Mikivishy/FullFront", "code_commit": "c27e96dfe2679c286ebdd9df1b3316f5404a56f3", "code_archive": "repos/K7UfQFegK5.zip", "code_archive_sha256": "0172dcd19445ec852a0b7a690df25f80dfdcb13430df3a86f2c64b6586cc68e4", "code_archive_bytes": 1672700, "code_file_count": 32, "code_extensions": {".py": 24, ".sh": 8}, "github_disk_usage_kb": 1588, "github_languages": {"Python": 403317, "Shell": 2108}, "github_archived": false, "github_pushed_at": "2025-05-16T08:44:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fullfront-benchmarking-mllms-across-the-full"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "el2pNeLrRC", "year": 2025, "status": "rejected", "title": "ECGN: A CLUSTER-AWARE APPROACH TO GRAPH NEURAL NETWORKS FOR IMBALANCED CLASSIFICATION.", "authors": ["Bishal Thapaliya", "Anh Nguyen", "Yao Lu", "Tian Xie", "Igor Grudetskyi", "Fudong Lin", "Antonios Valkanas", "Jingyu Liu", "Deepayan Chakrabarti", "Bilel Fehri"], "authorids": ["~Bishal_Thapaliya2", "~Anh_Nguyen8", "~Yao_Lu21", "~Tian_Xie13", "~Igor_Grudetskyi1", "~Fudong_Lin1", "~Antonios_Valkanas1", "~Jingyu_Liu5", "~Deepayan_Chakrabarti1", "~Bilel_Fehri1"], "authors_source": "OpenReview API", "abstract": "Classifying nodes in a graph is a common problem. The ideal classifier must\nadapt to any imbalances in the class distribution. It must also use information in\nthe clustering structure of real-world graphs. Existing Graph Neural Networks\n(GNNs) have not addressed both problems together. We propose the Enhanced\nCluster-aware Graph Network (ECGN), a novel method that addresses these is-\nsues by integrating cluster-specific training with synthetic node generation. Unlike\ntraditional GNNs that apply the same node update process for all nodes, ECGN\nlearns different aggregations for different clusters. We also use the clusters to gen-\nerate new minority-class nodes in a way that helps clarify the inter-class decision\nboundary. By combining cluster-aware embeddings with a global integration step,\nECGN enhances the quality of the resulting node embeddings. Our method works\nwith any underlying GNN and any cluster generation technique. Experimental\nresults show that ECGN consistently outperforms its closest competitors by up to\n11% on some widely-studied benchmark datasets. The GitHub implementation\nfor implementation and replication is publicly available on https://github.com/anonymous753341/ECGN.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "FYeUEFTY1x", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5083/Reviewer_kTym"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper introduces ECGN, a method designed to address class-imbalanced node classification. The approach begins by clustering nodes into groups, then computing node embeddings within each cluster through separate GNN models. Using these embeddings, it applies SMOTE to oversample minority nodes, augmenting the original graph, which is then processed by another GNN model for further training. Extensive experiments on datasets of varying sizes are conducted to evaluate the model, including ablation studies to assess the contribution of each component.", "review_text": "This paper introduces ECGN, a method designed to address class-imbalanced node classification. The approach begins by clustering nodes into groups, then computing node embeddings within each cluster through separate GNN models. Using these embeddings, it applies SMOTE to oversample minority nodes, augmenting the original graph, which is then processed by another GNN model for further training. Extensive experiments on datasets of varying sizes are conducted to evaluate the model, including ablation studies to assess the contribution of each component.", "strengths": "1. The experiments are thorough, covering datasets of various sizes, including a large-scale dataset.\n2. The source code is available for review, and the repository is well-organized.", "weaknesses": "1. The paper is poorly organized. It devotes excessive space to related work and includes figures and tables, such as Figure 2, Table 3, and Table 4, that take up significant space without providing much information. Meanwhile, important details are missing from the main paper, particularly about the pre-training step in Proposed Algorithm section, which is crucial as an ablation study references it. However, placing these contents in the appendix forces readers to flip back and forth to follow this paper.\n\n2. I have concerns about the pre-training stage. It appears to involve simple feature aggregation with separate GNNs within each cluster. Are these GNNs trained independently? If so, what is the training objective? If not, why not consider a non-parametric model for feature aggregation instead?\n\n3. The clustering stage raises some concerns. Given the high class imbalance in the graph, minority nodes may end up grouped with a large number of majority nodes within the same cluster, leading to a poor clustering quality. Have the authors considered this possibility?\n\n4. Equations in the paper lack indexing, making them difficult to reference. Additionally, in line 257, the paper states that connectivity measures the connections between minority nodes and nodes in other clusters. However, the equation that follows counts connections between minority nodes and nodes in other classes.\n\n5. Standard deviations are missing from the reported results. This is particularly concerning given that each model is only run four times, which could yield results with high variability.\n\n6. In Table 2, the performance difference between ECGN with and without SMOTE is minimal. This raises questions about the necessity of this component. Is synthesizing minority nodes essential for addressing class imbalance in this context? Again, the absence of standard deviation reporting makes it unclear whether this improvement is statistically significant.\n\n7. In the related work section, the paper critiques previous generative methods, stating, “altering the graph structure introduces complexities, and finding the optimal mixing ratio between node features can be difficult, often leading to noisy results that hurt performance.” However, this work does not address these issues, as its node synthesis method appears to share the same limitations as previous node generative approaches.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ECGN, a method designed to address class-imbalanced node classification. The approach begins by clustering nodes into groups, then computing node embeddings within each cluster through separate GNN models. Using these embeddings, it applies SMOTE to oversample minority nodes, augmenting the original graph, which is then processed by another GNN model for further training. Extensive experiments on datasets of varying sizes are conducted to evaluate the model, including ablation studies to assess the contribution of each component.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "1. The experiments are thorough, covering datasets of various sizes, including a large-scale dataset.\n2. The source code is available for review, and the repository is well-organized.", "weaknesses": "1. The paper is poorly organized. It devotes excessive space to related work and includes figures and tables, such as Figure 2, Table 3, and Table 4, that take up significant space without providing much information. Meanwhile, important details are missing from the main paper, particularly about the pre-training step in Proposed Algorithm section, which is crucial as an ablation study references it. However, placing these contents in the appendix forces readers to flip back and forth to follow this paper.\n\n2. I have concerns about the pre-training stage. It appears to involve simple feature aggregation with separate GNNs within each cluster. Are these GNNs trained independently? If so, what is the training objective? If not, why not consider a non-parametric model for feature aggregation instead?\n\n3. The clustering stage raises some concerns. Given the high class imbalance in the graph, minority nodes may end up grouped with a large number of majority nodes within the same cluster, leading to a poor clustering quality. Have the authors considered this possibility?\n\n4. Equations in the paper lack indexing, making them difficult to reference. Additionally, in line 257, the paper states that connectivity measures the connections between minority nodes and nodes in other clusters. However, the equation that follows counts connections between minority nodes and nodes in other classes.\n\n5. Standard deviations are missing from the reported results. This is particularly concerning given that each model is only run four times, which could yield results with high variability.\n\n6. In Table 2, the performance difference between ECGN with and without SMOTE is minimal. This raises questions about the necessity of this component. Is synthesizing minority nodes essential for addressing class imbalance in this context? Again, the absence of standard deviation reporting makes it unclear whether this improvement is statistically significant.\n\n7. In the related work section, the paper critiques previous generative methods, stating, “altering the graph structure introduces complexities, and finding the optimal mixing ratio between node features can be difficult, often leading to noisy results that hurt performance.” However, this work does not address these issues, as its node synthesis method appears to share the same limitations as previous node generative approaches.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730495239517}, {"id": "OSvOBkHe3Y", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5083/Reviewer_ntpA"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper proposes an improved framework for building graph neural networks for imbalanced node classification tasks. The end-to-end flow includes the following steps:\n1. Cluster-Specific Training: The input graph is partitioned into smaller clusters, and graph neural networks are pre-trained for each cluster to generate node embeddings.\n2. Cluster-Aware SMOTE: Synthetic nodes are generated for minority classes, with a focus on nodes located near cluster boundaries. The embeddings from Step 1 are used to identify nearest neighbors, enabling interpolation to create new embeddings for the synthesized nodes.\n3. Global Integration: All embeddings from Steps 1 and 2 are used to train a global classifier.\n\nThe framework leverages local structural information in Step 1 and global information in Step 3, while Step 2 addresses class imbalance. Experimental results demonstrate substantial improvement over previous state-of-the-art, including GraphSMOTE and Cluster-GCN.", "review_text": "This paper proposes an improved framework for building graph neural networks for imbalanced node classification tasks. The end-to-end flow includes the following steps:\n1. Cluster-Specific Training: The input graph is partitioned into smaller clusters, and graph neural networks are pre-trained for each cluster to generate node embeddings.\n2. Cluster-Aware SMOTE: Synthetic nodes are generated for minority classes, with a focus on nodes located near cluster boundaries. The embeddings from Step 1 are used to identify nearest neighbors, enabling interpolation to create new embeddings for the synthesized nodes.\n3. Global Integration: All embeddings from Steps 1 and 2 are used to train a global classifier.\n\nThe framework leverages local structural information in Step 1 and global information in Step 3, while Step 2 addresses class imbalance. Experimental results demonstrate substantial improvement over previous state-of-the-art, including GraphSMOTE and Cluster-GCN.", "strengths": "The overall framework is novel and quite reasonable. It combines the strength from both GraphSMOTE and ClusterGCN while introducing innovative improvements on top of the previous SOTA in various aspects, including:\n1. Carrying out SMOTE with a cluster aware approach. Synthesizing nodes around the cluster boundary would be more effective for imbalanced classification problems.\n2. Using cluster specific GCN instead of a single GCN at the pre-training stage followed by global aggregation at a later stage. This balanced approach captures both local and global information, with trade-offs in compute and storage efficiency.\n\nThe proposed ECGN has been compared to a wide range of methods, and the experimental results consistently outperform the references across all evaluated datasets, demonstrating its effectiveness.", "weaknesses": "Presentation:\n1. The network architecture for cluster specific pretraining and global aggregation in Figure 1 are illustrative at high level, lacking details of network architecture. It didn’t clarify whether local and global network structures are identical. Additionally, the Global aggregation is only described as “H′ ← convolution of G′ using H′ as node features.” in Algorithm 1, without clear correspondence from Figure 1.\n2. The methods in experiment evaluation could be more clearly linked to references or description, specifically “Cluster-Aware SMOTE only” is confusing; see Q1 in the Question section.\n3. No information on computational cost has been provided, so it is unclear what trade-offs were made to achieve accuracy improvements.\n\nReference:\nThe reference to (Chiang et al., 2019) should include the conference information (KDD ’19) rather than only an arXiv link. Additionally, (Zhao et al., 2021b) is not the correct reference for GraphSMOTE, and the information provided in the reference is also incorrect.\n\nExperiments:\nMore detailed experiments are expected to clarify the value of Cluster-Aware SMOTE; see Q2 in the Question section.", "questions": "Q1:\na) What are the specifics of “Cluster-Aware SMOTE only” in the benchmark results, i.e. Table 2? Section 3.2 described “Cluster-Aware SMOTE” to synthesize nodes, but there is no mention of building a classifier. \nb) What’s the difference between “Cluster-Aware SMOTE only” and “ECGN with SMOTE”, and \nc) What’s the difference between “Cluster-Aware SMOTE only” and GraphSMOTE?\nd) Why is “Cluster-Aware SMOTE only” showing inferior performance than GraphSMOTE if the cluster-aware SMOTE is supposed to be more effective than non-cluster-aware SMOTE?\n\nQ2:\nCluster aware node synthesis in 3.2, especially synthesizing nodes around the cluster boundary seems to be an important and novel step compared to GraphSMOTE. How effective is this approach, what if it reverts to a baseline approach? Is there any experiment validating the contribution of this step?\n\nQ3:\nRelated to Q2, the embeddings for the synthesized nodes are generated from their neighborhood. Neighborhood nodes could come from different clusters, but all embeddings were pre-trained within each cluster. Was any inter-cluster graph structure lost in this process?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an improved framework for building graph neural networks for imbalanced node classification tasks. The end-to-end flow includes the following steps:\n1. Cluster-Specific Training: The input graph is partitioned into smaller clusters, and graph neural networks are pre-trained for each cluster to generate node embeddings.\n2. Cluster-Aware SMOTE: Synthetic nodes are generated for minority classes, with a focus on nodes located near cluster boundaries. The embeddings from Step 1 are used to identify nearest neighbors, enabling interpolation to create new embeddings for the synthesized nodes.\n3. Global Integration: All embeddings from Steps 1 and 2 are used to train a global classifier.\n\nThe framework leverages local structural information in Step 1 and global information in Step 3, while Step 2 addresses class imbalance. Experimental results demonstrate substantial improvement over previous state-of-the-art, including GraphSMOTE and Cluster-GCN.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The overall framework is novel and quite reasonable. It combines the strength from both GraphSMOTE and ClusterGCN while introducing innovative improvements on top of the previous SOTA in various aspects, including:\n1. Carrying out SMOTE with a cluster aware approach. Synthesizing nodes around the cluster boundary would be more effective for imbalanced classification problems.\n2. Using cluster specific GCN instead of a single GCN at the pre-training stage followed by global aggregation at a later stage. This balanced approach captures both local and global information, with trade-offs in compute and storage efficiency.\n\nThe proposed ECGN has been compared to a wide range of methods, and the experimental results consistently outperform the references across all evaluated datasets, demonstrating its effectiveness.", "weaknesses": "Presentation:\n1. The network architecture for cluster specific pretraining and global aggregation in Figure 1 are illustrative at high level, lacking details of network architecture. It didn’t clarify whether local and global network structures are identical. Additionally, the Global aggregation is only described as “H′ ← convolution of G′ using H′ as node features.” in Algorithm 1, without clear correspondence from Figure 1.\n2. The methods in experiment evaluation could be more clearly linked to references or description, specifically “Cluster-Aware SMOTE only” is confusing; see Q1 in the Question section.\n3. No information on computational cost has been provided, so it is unclear what trade-offs were made to achieve accuracy improvements.\n\nReference:\nThe reference to (Chiang et al., 2019) should include the conference information (KDD ’19) rather than only an arXiv link. Additionally, (Zhao et al., 2021b) is not the correct reference for GraphSMOTE, and the information provided in the reference is also incorrect.\n\nExperiments:\nMore detailed experiments are expected to clarify the value of Cluster-Aware SMOTE; see Q2 in the Question section.", "questions": "Q1:\na) What are the specifics of “Cluster-Aware SMOTE only” in the benchmark results, i.e. Table 2? Section 3.2 described “Cluster-Aware SMOTE” to synthesize nodes, but there is no mention of building a classifier. \nb) What’s the difference between “Cluster-Aware SMOTE only” and “ECGN with SMOTE”, and \nc) What’s the difference between “Cluster-Aware SMOTE only” and GraphSMOTE?\nd) Why is “Cluster-Aware SMOTE only” showing inferior performance than GraphSMOTE if the cluster-aware SMOTE is supposed to be more effective than non-cluster-aware SMOTE?\n\nQ2:\nCluster aware node synthesis in 3.2, especially synthesizing nodes around the cluster boundary seems to be an important and novel step compared to GraphSMOTE. How effective is this approach, what if it reverts to a baseline approach? Is there any experiment validating the contribution of this step?\n\nQ3:\nRelated to Q2, the embeddings for the synthesized nodes are generated from their neighborhood. Neighborhood nodes could come from different clusters, but all embeddings were pre-trained within each cluster. Was any inter-cluster graph structure lost in this process?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730461505436}, {"id": "nRGwePKxOs", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5083/Reviewer_4efw"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper introduces ECGN, a framework that enhances Graph Neural Networks (GNNs) by incorporating cluster-specific training and synthetic node generation to address class imbalance in graph data, applicable to node classification tasks. ECGN employs a novel technique called Cluster-Aware SMOTE, which generates synthetic minority-class nodes, improving the decision boundary between classes and enhancing overall model performance. By combining cluster-aware updates with a global integration step, ECGN captures both local cluster-specific information and global graph structures. The model was evaluated on five benchmark datasets, including Citeseer and Reddit, and consistently outperformed existing state-of-the-art methods, with up to an 11% improvement in F1-score on imbalanced node classification tasks.", "review_text": "This paper introduces ECGN, a framework that enhances Graph Neural Networks (GNNs) by incorporating cluster-specific training and synthetic node generation to address class imbalance in graph data, applicable to node classification tasks. ECGN employs a novel technique called Cluster-Aware SMOTE, which generates synthetic minority-class nodes, improving the decision boundary between classes and enhancing overall model performance. By combining cluster-aware updates with a global integration step, ECGN captures both local cluster-specific information and global graph structures. The model was evaluated on five benchmark datasets, including Citeseer and Reddit, and consistently outperformed existing state-of-the-art methods, with up to an 11% improvement in F1-score on imbalanced node classification tasks.", "strengths": "1. This paper proposes a novel framework for addressing the problem of label imbalance in graph data.\n\n2. The method demonstrates strong performance on specific benchmark datasets.\n\n3. The code is publicly available, which is beneficial for the community in terms of reproducibility and further research.", "weaknesses": "1. The summary of related work is incomplete. See question 1.\n\n2. The description of the methodology is not clear enough. See questions 2, 3, and 4.\n\n3. The experiments are not comprehensive enough. See questions 5 and 6.", "questions": "1. The summary of related work is incomplete. Considering that the main contribution of this paper lies in addressing the graph class imbalance problem, the authors should discuss recent papers that tackle this issue, such as [1][2][3][4][5][6], to strengthen the paper's motivation and contribution.\n\n2. The methodology section lacks clarity. The most confusing aspect is the distinction between class and cluster. Specifically, when clustering algorithms like LSH group nodes with similar features into a cluster, is there any correlation between the clusters and the node labels (classes)? The authors should provide a specific example or diagram illustrating how classes and clusters relate in their approach.\n\n3. Continuing from question 2: To what extent does the quality of the clustering algorithm affect the final results? If the clustering algorithm is optimal and clusters nodes perfectly according to their labels, would that improve prediction performance? How should we interpret the statement on line 200, \"The choice of clustering algorithm is orthogonal to our method\"? Does this imply that the clustering results have no impact on the final outcome of the method? It is recommended that the authors add an ablation study or sensitivity analysis to quantify how different clustering algorithms or their qualities affect the final results.\n\n4.  Continuing from question 2: On line 257, the paper states, \"For each minority node v in class Y_m, we compute its connectivity to nodes in other clusters.\" This statement blurs the distinction between class and cluster. How should we understand the relationship between class and cluster in this context?\n\n5.  The experiments are insufficient. On one hand, the authors mention that the baselines are somewhat outdated. It is recommended to compare against more recent methods, such as [1][2][3][4][5][6]. On the other hand, even when comparing with the baselines mentioned in the paper, the performance improvements across several datasets are marginal, with numerical increases of only 0.02 to 0.03, representing less than a 5% improvement. It is suggested that the authors conduct comparisons with more updated work; this way, even modest improvements could be considered acceptable in the context of advancing the state of the art.\n\n6.  Continuing from question 5: The authors are encouraged to provide training time and training parameters such as training time per epoch, total training time, and memory usage for both ECGN and baseline methods in the paper or appendix. Given that the framework appears to require training multiple GNNs, it is important to clarify whether the performance improvements come at the cost of significantly higher computational overhead.\n\nReferences:  \n[1]\tPark, Joonhyung, Jaeyun Song, and Eunho Yang. \"Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification.\" International conference on learning representations. 2021.  \n[2]\tQian, Yiyue, et al. \"Co-modality graph contrastive learning for imbalanced node classification.\" Advances in Neural Information Processing Systems 35 (2022): 15862-15874.  \n[3]\tWang, Yu, Charu Aggarwal, and Tyler Derr. \"Distance-wise Prototypical Graph Neural Network for Imbalanced Node Classification.\" Proceedings of the 17th International Workshop on Mining and Learning with Graphs (MLG). 2022.  \n[4]\tSong, Jaeyun, Joonhyung Park, and Eunho Yang. \"TAM: topology-aware margin loss for class-imbalanced node classification.\" International Conference on Machine Learning. PMLR, 2022.  \n[5]\tZeng, Liang, et al. \"Imgcl: Revisiting graph contrastive learning on imbalanced node classification.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 9. 2023.  \n[6]\tZhou, Mengting, and Zhiguo Gong. \"GraphSR: a data augmentation algorithm for imbalanced node classification.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 4. 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ECGN, a framework that enhances Graph Neural Networks (GNNs) by incorporating cluster-specific training and synthetic node generation to address class imbalance in graph data, applicable to node classification tasks. ECGN employs a novel technique called Cluster-Aware SMOTE, which generates synthetic minority-class nodes, improving the decision boundary between classes and enhancing overall model performance. By combining cluster-aware updates with a global integration step, ECGN captures both local cluster-specific information and global graph structures. The model was evaluated on five benchmark datasets, including Citeseer and Reddit, and consistently outperformed existing state-of-the-art methods, with up to an 11% improvement in F1-score on imbalanced node classification tasks.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. This paper proposes a novel framework for addressing the problem of label imbalance in graph data.\n\n2. The method demonstrates strong performance on specific benchmark datasets.\n\n3. The code is publicly available, which is beneficial for the community in terms of reproducibility and further research.", "weaknesses": "1. The summary of related work is incomplete. See question 1.\n\n2. The description of the methodology is not clear enough. See questions 2, 3, and 4.\n\n3. The experiments are not comprehensive enough. See questions 5 and 6.", "questions": "1. The summary of related work is incomplete. Considering that the main contribution of this paper lies in addressing the graph class imbalance problem, the authors should discuss recent papers that tackle this issue, such as [1][2][3][4][5][6], to strengthen the paper's motivation and contribution.\n\n2. The methodology section lacks clarity. The most confusing aspect is the distinction between class and cluster. Specifically, when clustering algorithms like LSH group nodes with similar features into a cluster, is there any correlation between the clusters and the node labels (classes)? The authors should provide a specific example or diagram illustrating how classes and clusters relate in their approach.\n\n3. Continuing from question 2: To what extent does the quality of the clustering algorithm affect the final results? If the clustering algorithm is optimal and clusters nodes perfectly according to their labels, would that improve prediction performance? How should we interpret the statement on line 200, \"The choice of clustering algorithm is orthogonal to our method\"? Does this imply that the clustering results have no impact on the final outcome of the method? It is recommended that the authors add an ablation study or sensitivity analysis to quantify how different clustering algorithms or their qualities affect the final results.\n\n4.  Continuing from question 2: On line 257, the paper states, \"For each minority node v in class Y_m, we compute its connectivity to nodes in other clusters.\" This statement blurs the distinction between class and cluster. How should we understand the relationship between class and cluster in this context?\n\n5.  The experiments are insufficient. On one hand, the authors mention that the baselines are somewhat outdated. It is recommended to compare against more recent methods, such as [1][2][3][4][5][6]. On the other hand, even when comparing with the baselines mentioned in the paper, the performance improvements across several datasets are marginal, with numerical increases of only 0.02 to 0.03, representing less than a 5% improvement. It is suggested that the authors conduct comparisons with more updated work; this way, even modest improvements could be considered acceptable in the context of advancing the state of the art.\n\n6.  Continuing from question 5: The authors are encouraged to provide training time and training parameters such as training time per epoch, total training time, and memory usage for both ECGN and baseline methods in the paper or appendix. Given that the framework appears to require training multiple GNNs, it is important to clarify whether the performance improvements come at the cost of significantly higher computational overhead.\n\nReferences:  \n[1]\tPark, Joonhyung, Jaeyun Song, and Eunho Yang. \"Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification.\" International conference on learning representations. 2021.  \n[2]\tQian, Yiyue, et al. \"Co-modality graph contrastive learning for imbalanced node classification.\" Advances in Neural Information Processing Systems 35 (2022): 15862-15874.  \n[3]\tWang, Yu, Charu Aggarwal, and Tyler Derr. \"Distance-wise Prototypical Graph Neural Network for Imbalanced Node Classification.\" Proceedings of the 17th International Workshop on Mining and Learning with Graphs (MLG). 2022.  \n[4]\tSong, Jaeyun, Joonhyung Park, and Eunho Yang. \"TAM: topology-aware margin loss for class-imbalanced node classification.\" International Conference on Machine Learning. PMLR, 2022.  \n[5]\tZeng, Liang, et al. \"Imgcl: Revisiting graph contrastive learning on imbalanced node classification.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 9. 2023.  \n[6]\tZhou, Mengting, and Zhiguo Gong. \"GraphSR: a data augmentation algorithm for imbalanced node classification.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 4. 2023.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730427383903}, {"id": "YKnn2SIWDP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5083/Reviewer_8tSN"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces ECGN, a new method in Graph Neural Networks designed to improve node classification in class-imbalanced datasets by leveraging cluster-specific embeddings and synthetic node generation. ECGN operates by first pre-training subclusters to capture local graph structures, then generating synthetic nodes in boundary areas of minority clusters to enhance classification for underrepresented classes, and finally integrating these local representations into a global structure. This approach outperforms existing models, achieving up to an 11% improvement on benchmark datasets, highlighting its ability to maintain both global and local graph consistency for balanced, high-accuracy node classification.", "review_text": "The paper introduces ECGN, a new method in Graph Neural Networks designed to improve node classification in class-imbalanced datasets by leveraging cluster-specific embeddings and synthetic node generation. ECGN operates by first pre-training subclusters to capture local graph structures, then generating synthetic nodes in boundary areas of minority clusters to enhance classification for underrepresented classes, and finally integrating these local representations into a global structure. This approach outperforms existing models, achieving up to an 11% improvement on benchmark datasets, highlighting its ability to maintain both global and local graph consistency for balanced, high-accuracy node classification.", "strengths": "- The paper is well-written, clear, and easy to follow. \n- Extensive experiments are conducted to validate ECGN's performance, and the inclusion of the code enhances reproducibility and supports the credibility of results. \n- ECGN is highly adaptable, functioning as a versatile enhancement for any GNN backbone, which underscores its potential generalizability across various GNN architectures.", "weaknesses": "- Motivation Clarity: Imbalanced classification and node clustering on graphs are two orthogonal problems. The motivation behind combining these two aspects is not sufficiently clear, and further elaboration is necessary to explain why clustering is beneficial for addressing class imbalance.\n- Related Work: The related work lacks coverage of recent advancements in class-imbalanced node classification. Key methods such as \"GraphENS: Neighbor-Aware Ego Network Synthesis for Class-Imbalanced Node Classification\" (ICLR 2022), \"TAM: Topology-Aware Margin Loss for Class-Imbalanced Node Classification\" (ICML 2022), \"GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification\" (KDD 2023), and \"Class-Imbalanced Graph Learning without Class Rebalancing\" (ICML 2024) should be discussed and ideally included in the empirical comparisons.\n- Inappropriate claims: The statement in line 54, “uniform updates can make the GNN overlook rich local structures,” is Inappropriate. Through simple uniform aggregation, message passing in GNNs can also capture local structures. \n- Evaluation: The F1-score improvements shown in Table 2 are modest across all datasets. Incorporating additional evaluation metrics for imbalanced learning, such as balanced accuracy would provide a more nuanced view of the model’s performance.", "questions": "Please refer to weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ECGN, a new method in Graph Neural Networks designed to improve node classification in class-imbalanced datasets by leveraging cluster-specific embeddings and synthetic node generation. ECGN operates by first pre-training subclusters to capture local graph structures, then generating synthetic nodes in boundary areas of minority clusters to enhance classification for underrepresented classes, and finally integrating these local representations into a global structure. This approach outperforms existing models, achieving up to an 11% improvement on benchmark datasets, highlighting its ability to maintain both global and local graph consistency for balanced, high-accuracy node classification.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-written, clear, and easy to follow. \n- Extensive experiments are conducted to validate ECGN's performance, and the inclusion of the code enhances reproducibility and supports the credibility of results. \n- ECGN is highly adaptable, functioning as a versatile enhancement for any GNN backbone, which underscores its potential generalizability across various GNN architectures.", "weaknesses": "- Motivation Clarity: Imbalanced classification and node clustering on graphs are two orthogonal problems. The motivation behind combining these two aspects is not sufficiently clear, and further elaboration is necessary to explain why clustering is beneficial for addressing class imbalance.\n- Related Work: The related work lacks coverage of recent advancements in class-imbalanced node classification. Key methods such as \"GraphENS: Neighbor-Aware Ego Network Synthesis for Class-Imbalanced Node Classification\" (ICLR 2022), \"TAM: Topology-Aware Margin Loss for Class-Imbalanced Node Classification\" (ICML 2022), \"GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification\" (KDD 2023), and \"Class-Imbalanced Graph Learning without Class Rebalancing\" (ICML 2024) should be discussed and ideally included in the empirical comparisons.\n- Inappropriate claims: The statement in line 54, “uniform updates can make the GNN overlook rich local structures,” is Inappropriate. Through simple uniform aggregation, message passing in GNNs can also capture local structures. \n- Evaluation: The F1-score improvements shown in Table 2 are modest across all datasets. Incorporating additional evaluation metrics for imbalanced learning, such as balanced accuracy would provide a more nuanced view of the model’s performance.", "questions": "Please refer to weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730255590384}], "openreview_url": "https://openreview.net/forum?id=el2pNeLrRC", "arxiv_id": "2410.11765", "paper_pdf": "papers/el2pNeLrRC.pdf", "paper_pdf_sha256": "756397cb129dcb9f34de13a8383a126e3685ab053625e0a0dc1ffe7b7ff4a2fe", "paper_pdf_bytes": 1043339, "paper_pdf_source": "openreview", "code_url": "https://github.com/anonymous753341/ECGN", "code_repository": "anonymous753341/ECGN", "code_commit": "50fb116876969c278e0a3f48fd80ab38bb4ad725", "code_archive": "repos/el2pNeLrRC.zip", "code_archive_sha256": "c4a1a85ae67e1f00296e43e694192f18562f73b40a4c48e82282028a94855123", "code_archive_bytes": 104324, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 109, "github_languages": {"Python": 303204}, "github_archived": false, "github_pushed_at": "2024-10-02T04:15:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ecgn-a-cluster-aware-approach-to-graph-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EAkjVCtRO2", "year": 2024, "status": "rejected", "title": "Variational quantization for state space models", "authors": ["Etienne David", "Jean Bellot", "Sylvain Le Corff"], "authorids": ["~Etienne_David1", "~Jean_Bellot1", "~Sylvain_Le_Corff1"], "authors_source": "OpenReview API", "abstract": "Forecasting tasks using large datasets gathering thousands of heterogeneous time series is a crucial statistical problem in numerous sectors. The main challenge is to model a rich variety of time series, leverage any available external signals and provide sharp predictions with statistical guarantees. In this work, we propose a new forecasting model  that combines discrete state space hidden Markov models with recent neural network architectures and training procedures inspired by vector quantized variational autoencoders. \nWe introduce a variational discrete posterior distribution of the latent states given the observations and a two-stage training procedure to alternatively train the parameters of the latent states and of the emission distributions. By learning a collection of emission laws and temporarily activating them depending on the hidden process dynamics, the proposed method allows to explore large datasets and leverage available external signals. We assess the performance of the proposed method using several datasets and show that it outperforms other state-of-the-art solutions.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "dPM3PprsDd", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6114/Reviewer_9MfL"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors present a new algorithm for time-series prediction based on VQ-VAEs and RNNs, with separate emission models for each quantized hidden state. The authors evaluate their algorithm on a number of datasets, showing comparable results with SOTA algorithms.", "review_text": "The authors present a new algorithm for time-series prediction based on VQ-VAEs and RNNs, with separate emission models for each quantized hidden state. The authors evaluate their algorithm on a number of datasets, showing comparable results with SOTA algorithms.", "strengths": "* The authors evaluated their algorithm against multiple other algorithms on multiple datasets, with multi-seed comparison + grid search.\n* As far as I am aware, combining VQ-VAEs and RNNs in this specific way (latent-conditioned observation model) has not been explored before.", "weaknesses": "* While the emprical results are nice, it would further strengthen the paper if the authors could perform ablation experiments to uncover/provide a better intuition about why their algorithm performs well against others.", "questions": "* The author's new algorithm performs almost similarly to previous SOTA Transformer-based model PatchTST/64. However, it's unclear from my first reading why one would prefer one over the other. Is it easier to train/better runtime etc.?\n* It's nice that the authors performed a grid search over learning rates and batch size for the other algorithms. I think the paper would be further strengthened if the authors conducted a hyperparameter search also over network size where it make sense (similar to the grid search that the authors performed over hidden size for their network)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present a new algorithm for time-series prediction based on VQ-VAEs and RNNs, with separate emission models for each quantized hidden state. The authors evaluate their algorithm on a number of datasets, showing comparable results with SOTA algorithms.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "* The authors evaluated their algorithm against multiple other algorithms on multiple datasets, with multi-seed comparison + grid search.\n* As far as I am aware, combining VQ-VAEs and RNNs in this specific way (latent-conditioned observation model) has not been explored before.", "weaknesses": "* While the emprical results are nice, it would further strengthen the paper if the authors could perform ablation experiments to uncover/provide a better intuition about why their algorithm performs well against others.", "questions": "* The author's new algorithm performs almost similarly to previous SOTA Transformer-based model PatchTST/64. However, it's unclear from my first reading why one would prefer one over the other. Is it easier to train/better runtime etc.?\n* It's nice that the authors performed a grid search over learning rates and batch size for the other algorithms. I think the paper would be further strengthened if the authors conducted a hyperparameter search also over network size where it make sense (similar to the grid search that the authors performed over hidden size for their network)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698772404747}, {"id": "i8pFceWF5h", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6114/Reviewer_YnSQ"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a forecasting model that combines discrete state space hidden Markov models with recurrent neural networks. The model is trained in a similar way as VQ-VAE. Experiments on several datasets show that the proposed method outperforms other SOTA methods.", "review_text": "This paper proposes a forecasting model that combines discrete state space hidden Markov models with recurrent neural networks. The model is trained in a similar way as VQ-VAE. Experiments on several datasets show that the proposed method outperforms other SOTA methods.", "strengths": "(1) Introducing finite state HMM for forecasting is interesting. \n\n(2) The formulation of the model and the learning method is sound. \n\n(3) The experiment results are good.", "weaknesses": "(1) The two-stage training appears unnecessary in the context of time series. It may lead to sub-optimal results. \n\n(2) More details should be provided on training the model with discrete latents in the VAE framework, i.e., how the discreteness is handled. \n\n(3) There is only one baseline model based on Transformer. I suspect there are many variants for forecasting.", "questions": "(1) How well does your method work on long context forecasting problem? \n\n(2) Do you employ straight-through trick or Gumbel trick in training your model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a forecasting model that combines discrete state space hidden Markov models with recurrent neural networks. The model is trained in a similar way as VQ-VAE. Experiments on several datasets show that the proposed method outperforms other SOTA methods.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "(1) Introducing finite state HMM for forecasting is interesting. \n\n(2) The formulation of the model and the learning method is sound. \n\n(3) The experiment results are good.", "weaknesses": "(1) The two-stage training appears unnecessary in the context of time series. It may lead to sub-optimal results. \n\n(2) More details should be provided on training the model with discrete latents in the VAE framework, i.e., how the discreteness is handled. \n\n(3) There is only one baseline model based on Transformer. I suspect there are many variants for forecasting.", "questions": "(1) How well does your method work on long context forecasting problem? \n\n(2) Do you employ straight-through trick or Gumbel trick in training your model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698731485585}, {"id": "dY4TbMco41", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6114/Reviewer_pvUL"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a new forecasting method by combining hidden Markov models and recurrent neural networks. The training procedure is inspired by vector quantized variational autoencoders, including the latent space and the emission laws parts. This method is computationally efficient and outperforms the SOTA baseline methods.", "review_text": "This paper proposes a new forecasting method by combining hidden Markov models and recurrent neural networks. The training procedure is inspired by vector quantized variational autoencoders, including the latent space and the emission laws parts. This method is computationally efficient and outperforms the SOTA baseline methods.", "strengths": "- The presentation of this paper is good. The authors show the comprehensive details of training and network architectures.\n\n- The experiments and evaluation are convincing. The authors test on multiple stationary and non-stationary datasets.", "weaknesses": "The writing of this paper can be improved. On Page 2, there is a relatively large blank space, which can be optimized. Also, there are many typos and grammatical issues in this paper. Some of the issues are listed below.\n- On page 6, Section 3.1.3, “...where T stand for…”, “stand” should be “stands”.\n- On Page 8, in the caption of Table 1, “...their associated MASE…”, “MASE” should be “MASEs”. \n- On Page 8, Section 3.2.2, the second equation should be MAE.\n- On Page 8, Section 3.2.3, “the accuracy of our model reaches state-of-the-art standards and provide uncertainty quantification.”, the subject of \"provide\" is not the accuracy.", "questions": "In Figure 1, for Hidden States Trajectory, e.g., $\\hat{x}_{t+1}^i=2$, are those “2,2,1,1,3,...,3” fixed or flexible to adjust in your implementation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new forecasting method by combining hidden Markov models and recurrent neural networks. The training procedure is inspired by vector quantized variational autoencoders, including the latent space and the emission laws parts. This method is computationally efficient and outperforms the SOTA baseline methods.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The presentation of this paper is good. The authors show the comprehensive details of training and network architectures.\n\n- The experiments and evaluation are convincing. The authors test on multiple stationary and non-stationary datasets.", "weaknesses": "The writing of this paper can be improved. On Page 2, there is a relatively large blank space, which can be optimized. Also, there are many typos and grammatical issues in this paper. Some of the issues are listed below.\n- On page 6, Section 3.1.3, “...where T stand for…”, “stand” should be “stands”.\n- On Page 8, in the caption of Table 1, “...their associated MASE…”, “MASE” should be “MASEs”. \n- On Page 8, Section 3.2.2, the second equation should be MAE.\n- On Page 8, Section 3.2.3, “the accuracy of our model reaches state-of-the-art standards and provide uncertainty quantification.”, the subject of \"provide\" is not the accuracy.", "questions": "In Figure 1, for Hidden States Trajectory, e.g., $\\hat{x}_{t+1}^i=2$, are those “2,2,1,1,3,...,3” fixed or flexible to adjust in your implementation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698283672035}], "openreview_url": "https://openreview.net/forum?id=EAkjVCtRO2", "arxiv_id": "2404.11117", "paper_pdf": "papers/EAkjVCtRO2.pdf", "paper_pdf_sha256": "b14e25183f0170b58bf35fac8a2fd5eec456e718a9d7fc6bddfcd7e0ca1f26c6", "paper_pdf_bytes": 6245826, "paper_pdf_source": "openreview", "code_url": "https://github.com/etidav/next", "code_repository": "etidav/next", "code_commit": "7dde74c1d8875a05da714bf57a83feb28f5046f2", "code_archive": "repos/EAkjVCtRO2.zip", "code_archive_sha256": "600317e1cf2098fb329910f197c44508b06905541f5cf6f12135a7826a5275ff", "code_archive_bytes": 34231, "code_file_count": 20, "code_extensions": {".py": 11, ".sh": 9}, "github_disk_usage_kb": 86, "github_languages": {"Python": 133961, "Shell": 9465, "Dockerfile": 1204, "Makefile": 668}, "github_archived": false, "github_pushed_at": "2023-10-02T08:09:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variational-quantization-for-state-space"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "En7lGmzT_x", "year": 2023, "status": "rejected", "title": "Sharper Rates and Flexible Framework for Nonconvex SGD with Client and Data Sampling", "authors": ["Alexander Tyurin", "Lukang Sun", "Konstantin Pavlovich Burlachenko", "Peter Richtárik"], "authorids": ["~Alexander_Tyurin1", "~Lukang_Sun1", "~Konstantin_Pavlovich_Burlachenko1", "~Peter_Richtárik1"], "authors_source": "OpenReview API", "abstract": "We revisit the classical problem of finding an approximately stationary point of the average of $n$ smooth and possibly nonconvex functions. The optimal complexity of stochastic first-order methods in terms of the number of gradient evaluations of individual functions is $\\mathcal{O}\\left(n + n^{1/2}\\varepsilon^{-1}\\right)$, attained by the optimal SGD methods SPIDER (Cong Fang et al., 2018) and PAGE (Zhize Li et al., 2020), for example, where $\\varepsilon$ is the error tolerance. However, i) the big-$\\mathcal{O}$ notation hides crucial dependencies on the smoothness constants associated with the functions, and ii) the rates and theory in these methods assume simplistic sampling mechanisms that do not offer any flexibility. In this work we remedy the situation. First, we generalize the PAGE algorithm so that it can provably work with virtually any (unbiased) sampling mechanism. This is particularly useful in federated learning, as it allows us to construct and better understand the impact of various combinations of client and data sampling strategies. Second, our analysis is sharper as we make explicit use of certain novel inequalities  that capture the intricate interplay between the smoothness constants and the sampling procedure. Indeed, our analysis is better even for the simple sampling procedure analyzed in the PAGE paper. However, this already improved bound can be further sharpened by a different sampling scheme which we propose. In summary, we provide the most general and most accurate analysis of optimal SGD in the smooth nonconvex regime. Finally, our theoretical findings are supposed with carefully designed experiments.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "eZYOGV94RV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1371/Reviewer_NfNX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "#### ***Background*** :\nCurrent convergence results to attain \\epsilon approximate stationary point using first-order stochastic methods (ex SGD and its more recent variants like SARAH, PAGE) while achieve the optimal rates matching the lower bounds; the analysis has the following issues - (a) The big-O notation in the above results hides important and typically very large data-dependent constants (smoothness). The paper shows via simple examples to motivate why the dependence on L can be crucial. (b) Further these results are under the 'random' data sampling assumption -- which is often not true in modern distributed and federated learning settings. Prior work has analyzed different non-optimal SGD variants under different sampling schemes and showed improved constants.\n\n#### ***Main Contribution*** : \nIn this paper the authors unify these proof ideas and propose a unified framework to analyze PAGE and also show that it is possible to improve the convergence under certain sampling assumptions.", "review_text": "I reviewed an earlier version of the paper for NeurIPS and unfortunately, the paper hasn't improved - in fact on a closer look they seem to have submitted the same draft w/o any changes. The paper still has significant weaknesses and unfortunately they didn't use the resubmission to improve upon them. ", "strengths": "#### ***Strengths*** :\n1. Overall the paper is really well written, proofs are clean and easy to follow,\n2. The refined analysis will serve as a good consolidated reference note for using sampling in practical optimization settings - while the proofs are hard to find from a plethora of different papers.   \n\n#### ***Clarifications*** :\n\n1. Overall, I feel that the novelty is marginal. The main novelty in my understanding is the introduction of two smoothness constants introduced in Def 2,3. The rest of the proofs seem routine given the proof technique introduced in Richtárik62and Takáˇc (2016) in the study of randomized coordinate descent methods, Horváth and Richtárik63(2019) and Qian et al. (2021) in analyzing SVRG, SAGA, and SARAH.\n\n2. The paper simply extends the convergence results of PAGE under the non-traditional finer smoothness constants and the results while new are not surprising.\n\n3. The claims rely on the sampling strategy being aware of $L_i$ - I am not sure how is this of any practical importance since it is not available in practice. \n\n4. PAGE with important sampling performs better (Fig 3) - well this is known result.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "#### ***Background*** :\nCurrent convergence results to attain \\epsilon approximate stationary point using first-order stochastic methods (ex SGD and its more recent variants like SARAH, PAGE) while achieve the optimal rates matching the lower bounds; the analysis has the following issues - (a) The big-O notation in the above results hides important and typically very large data-dependent constants (smoothness). The paper shows via simple examples to motivate why the dependence on L can be crucial. (b) Further these results are under the 'random' data sampling assumption -- which is often not true in modern distributed and federated learning settings. Prior work has analyzed different non-optimal SGD variants under different sampling schemes and showed improved constants.\n\n#### ***Main Contribution*** : \nIn this paper the authors unify these proof ideas and propose a unified framework to analyze PAGE and also show that it is possible to improve the convergence under certain sampling assumptions.", "strength_and_weaknesses": "#### ***Strengths*** :\n1. Overall the paper is really well written, proofs are clean and easy to follow,\n2. The refined analysis will serve as a good consolidated reference note for using sampling in practical optimization settings - while the proofs are hard to find from a plethora of different papers.   \n\n#### ***Clarifications*** :\n\n1. Overall, I feel that the novelty is marginal. The main novelty in my understanding is the introduction of two smoothness constants introduced in Def 2,3. The rest of the proofs seem routine given the proof technique introduced in Richtárik62and Takáˇc (2016) in the study of randomized coordinate descent methods, Horváth and Richtárik63(2019) and Qian et al. (2021) in analyzing SVRG, SAGA, and SARAH.\n\n2. The paper simply extends the convergence results of PAGE under the non-traditional finer smoothness constants and the results while new are not surprising.\n\n3. The claims rely on the sampling strategy being aware of $L_i$ - I am not sure how is this of any practical importance since it is not available in practice. \n\n4. PAGE with important sampling performs better (Fig 3) - well this is known result.", "clarity,_quality,_novelty_and_reproducibility": "I feel that the novelty is limited; most results are already known. The only contribution of bringing out the two smoothness constants and showing a sharper rate depending on them. In practice, these parameters are not known and this analysis while a good exercise has limited importance. ", "summary_of_the_review": "I reviewed an earlier version of the paper for NeurIPS and unfortunately, the paper hasn't improved - in fact on a closer look they seem to have submitted the same draft w/o any changes. The paper still has significant weaknesses and unfortunately they didn't use the resubmission to improve upon them. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "\n\n", "recommendation": "3: reject, not good enough"}, "tcdate": 1667678821543}, {"id": "aof-CIfD2m", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1371/Reviewer_CFoS"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work reconsiders the problem of finding stationary points of finite-sum of $n$ smooth functions, where the optimal complexity for stochastic first-order methods is already known to be $\\mathcal{O}(n + n^{1/2} \\epsilon^{-1})$. The authors proposed to do a refined analysis since the $\\mathcal{O}$ could hide dependencies on smoothness constants, which could be unbalanced across different individual summands, and also a more detailed analysis for different sampling procedure to obtain stochastic gradients.\n\nThe authors generalize the analysis for PAGE method (Algorithm 1)to work for different sampling mechanisms, where the $L_+$ smoothness constants could possibly be improved as shown in Table 2. In particular, the authors define a new quantity called weighted Hessian Variance, which improves previous results as also shown in Table 2. Finally, the authors show that the analysis works for samplings used in federated learning.\n\nThe key technical results are to show that different sampling methods satisfy a weighted AB inequality (Assumption 4) as summarized in Table 1.", "review_text": "Overall, I found the motivation reasonable and results interesting. However, it is unclear to me how we could use the results to compare different samplings and guide practice, which makes it unclear to me if doing the refined analysis does provide enough benefits. ", "strengths": "**Strength**:\n\n1. The motivation of doing a refined analysis seems reasonable to me. The authors argue that this could be useful in federated learning.\n2. The more detailed analysis does improve the previous results and are more flexible for different sampling methods.\n3. The study of different samplings satisfying the weighted AB inequality is interesting, which provides intuitions for differences between sampling methods.\n\n**Weaknesses**:\n\n1. The work is still not enough to convince me doing the analysis does provide enough benefits. For example, from federated learning results in Section A.3 it can be concluded that the importance sampling achieved the best performances. How can we conclude from the complexity results in Table 2 that the analysis does reflect the experimental results?\n2. The results in Table 1 seem interesting. As mentioned in the paper, larger B values allow tighter results to be obtained. How do other quantities have impacts on performances of sampling methods?\n3. It is unclear to me how should we compare different sampling methods from Table 2. In particular, how do we compare complexity quantities of different samplings.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work reconsiders the problem of finding stationary points of finite-sum of $n$ smooth functions, where the optimal complexity for stochastic first-order methods is already known to be $\\mathcal{O}(n + n^{1/2} \\epsilon^{-1})$. The authors proposed to do a refined analysis since the $\\mathcal{O}$ could hide dependencies on smoothness constants, which could be unbalanced across different individual summands, and also a more detailed analysis for different sampling procedure to obtain stochastic gradients.\n\nThe authors generalize the analysis for PAGE method (Algorithm 1)to work for different sampling mechanisms, where the $L_+$ smoothness constants could possibly be improved as shown in Table 2. In particular, the authors define a new quantity called weighted Hessian Variance, which improves previous results as also shown in Table 2. Finally, the authors show that the analysis works for samplings used in federated learning.\n\nThe key technical results are to show that different sampling methods satisfy a weighted AB inequality (Assumption 4) as summarized in Table 1.", "strength_and_weaknesses": "**Strength**:\n\n1. The motivation of doing a refined analysis seems reasonable to me. The authors argue that this could be useful in federated learning.\n2. The more detailed analysis does improve the previous results and are more flexible for different sampling methods.\n3. The study of different samplings satisfying the weighted AB inequality is interesting, which provides intuitions for differences between sampling methods.\n\n**Weaknesses**:\n\n1. The work is still not enough to convince me doing the analysis does provide enough benefits. For example, from federated learning results in Section A.3 it can be concluded that the importance sampling achieved the best performances. How can we conclude from the complexity results in Table 2 that the analysis does reflect the experimental results?\n2. The results in Table 1 seem interesting. As mentioned in the paper, larger B values allow tighter results to be obtained. How do other quantities have impacts on performances of sampling methods?\n3. It is unclear to me how should we compare different sampling methods from Table 2. In particular, how do we compare complexity quantities of different samplings.", "clarity,_quality,_novelty_and_reproducibility": "The writing and presentation are clear. However, the originality and quality are limited.", "summary_of_the_review": "Overall, I found the motivation reasonable and results interesting. However, it is unclear to me how we could use the results to compare different samplings and guide practice, which makes it unclear to me if doing the refined analysis does provide enough benefits. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666771879080}, {"id": "Z5f6zjxwzLG", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1371/Reviewer_xtiL"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results.", "review_text": "This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results.\nAlthough the paper is theoretically sound, there are still some questions need to be discussed in this paper:\n1.\tAbout assumption.It would be better to make a attempt to prove assumption 4 was satisfied in SPIDER and SARAH.\n2.\tThe chart in Figure 3 lacks coordinate description.\n3.\tAbout the experiments. All experiments are telling that some samplings can improve the convergence rate of PAGE.Importance sampling performs better.About other sampling schemes in Table 1,it would be better to add them to the experiment.\n", "strengths": "Strength: This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning. This paper is theoretically sound in general.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results.", "strength_and_weaknesses": "Strength: This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning. This paper is theoretically sound in general.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is presented clearly. ", "summary_of_the_review": "This paper generalizes the PAGE algorithm and analyzes that it can improve the convergence rate with virtually any (unbiased) sampling mechanism using a novel assumption.It is helpful in the analysis of problems from federated learning.Some carefully designed experiments have verified theoretical results.\nAlthough the paper is theoretically sound, there are still some questions need to be discussed in this paper:\n1.\tAbout assumption.It would be better to make a attempt to prove assumption 4 was satisfied in SPIDER and SARAH.\n2.\tThe chart in Figure 3 lacks coordinate description.\n3.\tAbout the experiments. All experiments are telling that some samplings can improve the convergence rate of PAGE.Importance sampling performs better.About other sampling schemes in Table 1,it would be better to add them to the experiment.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666649772989}, {"id": "45olYQffGu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1371/Reviewer_paK1"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper considers a probabilistic gradient method (PAGE). The authors invoke different kinds of sampling schemes and claim that the new result is shaper than the original one.", "review_text": "The techniques are copies of those of (Li et al., 2021) and (Szlendak et al., 2021). I obtain this conclusion by reading their proof of Theorem 5 and that of Thm 1 in (Li et al., 2021) and that of Thm 4 in (Szlendak et al., 2021), which is a step-by-step copy-style. I would say the essential contribution is that the authors changed the compressor's property to a sampling property. The authors verified that several sampling schemes do satisfy their Assumption 4 (Indeed, an unbiased compressor seems similar to unbiased sampling in terms of analysis). It is good to know the result. However, I have to say that such a contribution does not meet the requirement for possible publication in ICLR.\n\n[Li et al., 2021] PAGE: A simple and optimal probabilistic gradient estimator for nonconvex optimization. In International Conference on\nMachine Learning, pp. 6286–6295. PMLR, 2021.\n\n[Szlendak et al., 2021] Permutation compressors for provably faster distributed nonconvex optimization. arXiv preprint arXiv:2110.03300, 2021.", "strengths": "The new sampling assumption is justified by considering several representative sampling schemes. However, the technique is a direct combination of (Li et al., 2021) and (Szlendak et al., 2021), rendering this manuscript far from possible publication in ICLR.\n\n\n[Li et al., 2021] PAGE: A simple and optimal probabilistic gradient estimator for nonconvex optimization. In International Conference on\nMachine Learning, pp. 6286–6295. PMLR, 2021.\n\n[Szlendak et al., 2021] Permutation compressors for provably faster distributed nonconvex optimization. arXiv preprint arXiv:2110.03300, 2021.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper considers a probabilistic gradient method (PAGE). The authors invoke different kinds of sampling schemes and claim that the new result is shaper than the original one.", "strength_and_weaknesses": "The new sampling assumption is justified by considering several representative sampling schemes. However, the technique is a direct combination of (Li et al., 2021) and (Szlendak et al., 2021), rendering this manuscript far from possible publication in ICLR.\n\n\n[Li et al., 2021] PAGE: A simple and optimal probabilistic gradient estimator for nonconvex optimization. In International Conference on\nMachine Learning, pp. 6286–6295. PMLR, 2021.\n\n[Szlendak et al., 2021] Permutation compressors for provably faster distributed nonconvex optimization. arXiv preprint arXiv:2110.03300, 2021.", "clarity,_quality,_novelty_and_reproducibility": "The clarity of this paper is proper. However, its quality and originality are quite limited due to incremental contributions; see also my detailed comments below.", "summary_of_the_review": "The techniques are copies of those of (Li et al., 2021) and (Szlendak et al., 2021). I obtain this conclusion by reading their proof of Theorem 5 and that of Thm 1 in (Li et al., 2021) and that of Thm 4 in (Szlendak et al., 2021), which is a step-by-step copy-style. I would say the essential contribution is that the authors changed the compressor's property to a sampling property. The authors verified that several sampling schemes do satisfy their Assumption 4 (Indeed, an unbiased compressor seems similar to unbiased sampling in terms of analysis). It is good to know the result. However, I have to say that such a contribution does not meet the requirement for possible publication in ICLR.\n\n[Li et al., 2021] PAGE: A simple and optimal probabilistic gradient estimator for nonconvex optimization. In International Conference on\nMachine Learning, pp. 6286–6295. PMLR, 2021.\n\n[Szlendak et al., 2021] Permutation compressors for provably faster distributed nonconvex optimization. arXiv preprint arXiv:2110.03300, 2021.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "3: reject, not good enough"}, "tcdate": 1666526690502}], "openreview_url": "https://openreview.net/forum?id=En7lGmzT_x", "arxiv_id": "2206.02275", "paper_pdf": "papers/En7lGmzT_x.pdf", "paper_pdf_sha256": "5da1064c606c2c605df7bc00e88b10e3577c93442d822829650d009dd7b02d04", "paper_pdf_bytes": 1050259, "paper_pdf_source": "openreview", "code_url": "https://github.com/mysteryresearcher/sampling-in-optimal-sgd", "code_repository": "mysteryresearcher/sampling-in-optimal-sgd", "code_commit": "bd3f2e68783df06e4d7847843fe6d2d048226872", "code_archive": "repos/En7lGmzT_x.zip", "code_archive_sha256": "0a46391d570304a9a0c3d0ff585dc64f0d37dcdd70ce20310a0101402b1b226a", "code_archive_bytes": 231302, "code_file_count": 53, "code_extensions": {".py": 45, ".ipynb": 5, ".sh": 3}, "github_disk_usage_kb": 160, "github_languages": {"Python": 736820, "Jupyter Notebook": 90910, "Shell": 816}, "github_archived": false, "github_pushed_at": "2023-01-29T07:23:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sharper-rates-and-flexible-framework-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "O0g6uPDLW7", "year": 2022, "status": "rejected", "title": "On the Adversarial Robustness of Vision Transformers", "authors": ["Rulin Shao", "Zhouxing Shi", "Jinfeng Yi", "Pin-Yu Chen", "Cho-Jui Hsieh"], "authorids": ["~Rulin_Shao1", "~Zhouxing_Shi1", "~Jinfeng_Yi1", "~Pin-Yu_Chen1", "~Cho-Jui_Hsieh1"], "authors_source": "OpenReview API", "abstract": "Following the success in advancing natural language processing and understanding, transformers are expected to bring revolutionary changes to computer vision. This work provides the first and comprehensive study on the robustness of vision transformers (ViTs) against adversarial perturbations. Tested on various white-box and transfer attack settings, we find that ViTs possess better adversarial robustness when compared with convolutional neural networks (CNNs). This observation also holds for certified robustness. We summarize the following main observations contributing to the improved robustness of ViTs:\n   1) Features learned by ViTs contain less low-level information and are more generalizable, which contributes to superior robustness against adversarial perturbations. \n   2) Introducing convolutional or tokens-to-token blocks for learning low-level features in ViTs can improve classification accuracy but at the cost of adversarial robustness.\n   3) Increasing the proportion of transformers in the model structure (when the model consists of both transformer and CNN blocks) leads to better robustness. But for a pure transformer model, simply increasing the size or adding layers cannot guarantee a similar effect.\n   4) Pre-training on larger datasets does not significantly improve adversarial robustness though it is critical for training ViTs. \n   5) Adversarial training is also applicable to ViT for training robust models.\nFurthermore, feature visualization and frequency analysis are conducted for explanation. The results show that ViTs are less sensitive to high-frequency perturbations than CNNs and there is a high correlation between how well the model learns low-level features and its robustness against different frequency-based perturbations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0VNoCI9NDB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper595/Reviewer_FTwY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper performs a study on the adversarial robustness of vision transformers. It provides some insight on vision transformer from the robustness perspective.", "review_text": "Strengths: \\\\\n(1) As claimed, this work provides the first and comprehensive study on the robustness of ViT. \\\\\n(2) This study provides some interesting (but not surprising) insight.\n(3) The paper has a clear structure and is easy to follow.\n\nWeaknesses: \\\\\n(1) The insights of this work cannot be exploited for understanding or improving the adversarial robustness. This is the main reason that I believe this work might be below the ICLR bar. \\\\\n(2) Why is ViT more robust than CNN? This work claims that ViTs have less low-level information and more generalizable. How do you define high-level and low-level? Why does high generalization contribute to superior robustness? Any proof or reference? It would be interesting if the authors can provide more concrete insight into the mechanism behind the reported phenomenon. \\\\\n\n\nSome works [1-10] might be worth a check: \\\\\nThey do not affect my rating of this submission.\n\n[1] Understanding and Improving Robustness of Vision Transformers through Patch-based Negative Augmentation\n[2] Certified Patch Robustness via Smoothed Vision Transformers\n[3] Adversarial Robustness Comparison of Vision Transformer and MLP-Mixer to CNNs\n[4] Towards Transferable Adversarial Attacks on Vision Transformers\n[5] On Improving Adversarial Transferability of Vision Transformers\n[6] Reveal of Vision Transformers Robustness against Adversarial Attacks\n[7] Intriguing Properties of Vision Transformers\n[8] Vision Transformers are Robust Learners\n[9] On the Robustness of Vision Transformers to Adversarial Examples\n[10] Towards Robust Vision Transformer", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper performs a study on the adversarial robustness of vision transformers. It provides some insight on vision transformer from the robustness perspective.", "main_review": "Strengths: \\\\\n(1) As claimed, this work provides the first and comprehensive study on the robustness of ViT. \\\\\n(2) This study provides some interesting (but not surprising) insight.\n(3) The paper has a clear structure and is easy to follow.\n\nWeaknesses: \\\\\n(1) The insights of this work cannot be exploited for understanding or improving the adversarial robustness. This is the main reason that I believe this work might be below the ICLR bar. \\\\\n(2) Why is ViT more robust than CNN? This work claims that ViTs have less low-level information and more generalizable. How do you define high-level and low-level? Why does high generalization contribute to superior robustness? Any proof or reference? It would be interesting if the authors can provide more concrete insight into the mechanism behind the reported phenomenon. \\\\\n\n\nSome works [1-10] might be worth a check: \\\\\nThey do not affect my rating of this submission.\n\n[1] Understanding and Improving Robustness of Vision Transformers through Patch-based Negative Augmentation\n[2] Certified Patch Robustness via Smoothed Vision Transformers\n[3] Adversarial Robustness Comparison of Vision Transformer and MLP-Mixer to CNNs\n[4] Towards Transferable Adversarial Attacks on Vision Transformers\n[5] On Improving Adversarial Transferability of Vision Transformers\n[6] Reveal of Vision Transformers Robustness against Adversarial Attacks\n[7] Intriguing Properties of Vision Transformers\n[8] Vision Transformers are Robust Learners\n[9] On the Robustness of Vision Transformers to Adversarial Examples\n[10] Towards Robust Vision Transformer", "summary_of_the_review": "Overall, this work has some merits but I expect more than that to get in for ICLR. Some investigation is rudimentary and is suggested to provide more deep insight. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N.A.", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635924718682}, {"id": "VnE9RbpAr5Q", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper595/Reviewer_7ATA"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThis paper studies the adversarial robustness of ViTs. There are several significant strengths and weaknesses in this paper. Especially, the novelty of this paper against one published paper is limited. It would be good to see my detailed comments below.\n\n", "review_text": "\n(Negative) Throughout the paper, the authors claimed that this is the first paper to introduce an adversarial attack on ViT to examine the robustness of ViTs. But actually, until today, there have been a number of papers working on this topic, e.g., [r1-r4]. Especially, [r4] has been published in ICCV 2021. I understand that the authors might argue that this paper was submitted to Arxiv on 29 Mar 2021, which is earlier than some of the other papers. But, the authors should understand the following facts. First, please note that, [r4] was submitted to Arxiv on 26 Mar 2021, which is even earlier than the authors' paper. Moreover, [r4] has been published by ICCV 2021, and the authors' paper is unpublished yet. Second, because this paper is still not yet published and currently there has been a number of paper workings the same ideas of this paper, it would not be appropriate for the authors to overstate this is the first work in ViT robustness, considering many of the concurrent works provide more deep insights in this topic. Third, [r5] was the earliest paper studying the robustness of non-local attentive models among all of these works. Fourth, the insight of this paper is relatively shallow. There are not many valuable discoveries in this article, so the author keeps discussing the same thing from different angles. Fifth, I understand that the author wrote this paper many months ago, and they don't want to improve their paper because they think many \"advance papers\" are later than theirs. But I encourage the authors to improve their paper with more advanced insights to catch up with the recent developments. For example, the author could propose some practical solutions to ViT robustness. I also have some experience that I was the first to propose some techniques, but my paper was rejected. Then, the later paper working on a similar idea was published. I was frustrated. But I continued to improve my paper, and finally, my paper was accepted by another venue as an oral presentation. \n\n\n[r1] Intriguing Properties of Vision Transformers\n\n[r2] Vision Transformers are Robust Learners\n\n[r3] RobustART: Benchmarking Robustness on Architecture Design and Training Techniques\n\n[r4] Understanding Robustness of Transformers for Image Classification (ICCV 2021) 26 Mar 2021\n\n[r5] Feature Denoising for Improving Adversarial Robustness.\n\n\n(Negative) It would be good not to say that this work provides the first study on the robustness of vision transformers (ViTs) against adversarial perturbations. There have been other works discussing the ViT adversarial robustness. It would be good for the author to recognize the contribution of the community. It would be good for the authors to be humble.\n\n(Negative) The following observations are natural and not new: \"we find that ViTs possess better adversarial robustness when compared with convolutional neural networks (CNNs).\"\n\n(Negative) The following observations are natural and not new: \"Features learned by ViTs contain less low-level information and are more generalizable, which contributes to superior robustness against adversarial perturbations.\" Existing papers found that ViTs do better than CNNs in adversarial robustness but do worse than CNNs in natural noise robustness.\n\n(Negative) The following observations are complementary to the above observations. They thus are less informative: \"Introducing convolutional or tokens-to-token blocks for learning low-level features in ViTs can improve classification accuracy but at the cost of adversarial robustness.\"\n\n(Negative) The following observations are complementary to the above observations. So these observations are less informative. Increasing the proportion of transformers in the model structure (when the model consists of both transformer and CNN blocks) leads to better robustness. This makes readers feel that there are not many valuable discoveries in this article, so the author keeps discussing things from different angles.\n\n(Negative) I'm afraid I have to disagree with the authors' following claim: \"Pre-training on larger datasets does not significantly improve adversarial robustness though it is critical for training ViTs.\" I think the authors performed the wrong pre-training. Pre-training on the larger dataset should be robust pre-training, i.e., it should be adversarial training, but NOT vanilla pre-training. In my experiments, doing robust pre-training on a larger dataset significantly improves adversarial robustness.\n\n\n(Negative) The following observations are too natural and contain little information: \"ViTs are less sensitive to high-frequency perturbations than CNNs and\nthere is a high correlation between how well the model learns low-level features and its robustness against different frequency-based perturbations.\"\n\n(Negative) The following descriptions need to be corrected: \"it remains unclear on the robustness of ViT against adversarial perturbations, which is critical for safe and reliable deployment of many real-world applications,\" given that the literature has fully studied the robustness of ViT.\n\n\n(Negative) The following descriptions need to be corrected: \"In this work, we conduct the first study on examining the adversarial robustness of ViTs on image classification tasks and make comparisons with CNN baselines.\"\n\n\n(Neutral) The following observation is natural and ok: \" Using denoised randomized smoothing (Salman et al., 2020), ViTs attain significantly better-certified robustness than CNNs.\"\n\n(Positive) The following description is correct and insightful: \"short-term memory (LSTM) or CNN, with a theoretical explanation provided in Hsieh et al. (2019). However, due to the discrete nature of NLP models, these studies are focusing on discrete perturbations (e.g., word or character substitutions) which are very different from small and continuous perturbations in computer vision tasks.\"\n\n(Negative) The following description needs to be modified: \" To the best of our knowledge, this work is the first study that investigates the adversarial robustness (against small perturbations in the input pixel space) of transformers on computer vision tasks.\"\n\n(Negative) The following description needs to be modified: \"In the context of computer vision, the most relevant work is Alamri et al. (2020), which applies transformer encoder in the object detection task and reports better adversarial robustness.\"\n\n(Negative) In the related work section, all the above-mentioned works should be discussed and compared to acknowledge the community's contribution.\n\n(Positive) I am happy to see that the authors also evaluate the certified robustness of the models using randomized smoothing, where the robustness is evaluated as the certified radius,\n\n(Negative) The following observations are too natural: \" Adversarial training can be applied to train robust ViTs.\"\n\n\n(Positive) The following finding is interesting and valuable: \"We conjecture that ViT may need larger training data or longer training epochs to improve further its robust training performance, inspired by the fact that on natural training ViT is not able to perform well either without large-scale pre-training.\"\n\n\n(Positive) The results in Table 4 are fascinating, valuable, and beneficial. I really like the results in this table.\n\n\n(Negative) Please use text instead of figures to give a sub-caption to each subfigure in Figure 4. The current version is lossy.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "\nThis paper studies the adversarial robustness of ViTs. There are several significant strengths and weaknesses in this paper. Especially, the novelty of this paper against one published paper is limited. It would be good to see my detailed comments below.\n\n", "main_review": "\n(Negative) Throughout the paper, the authors claimed that this is the first paper to introduce an adversarial attack on ViT to examine the robustness of ViTs. But actually, until today, there have been a number of papers working on this topic, e.g., [r1-r4]. Especially, [r4] has been published in ICCV 2021. I understand that the authors might argue that this paper was submitted to Arxiv on 29 Mar 2021, which is earlier than some of the other papers. But, the authors should understand the following facts. First, please note that, [r4] was submitted to Arxiv on 26 Mar 2021, which is even earlier than the authors' paper. Moreover, [r4] has been published by ICCV 2021, and the authors' paper is unpublished yet. Second, because this paper is still not yet published and currently there has been a number of paper workings the same ideas of this paper, it would not be appropriate for the authors to overstate this is the first work in ViT robustness, considering many of the concurrent works provide more deep insights in this topic. Third, [r5] was the earliest paper studying the robustness of non-local attentive models among all of these works. Fourth, the insight of this paper is relatively shallow. There are not many valuable discoveries in this article, so the author keeps discussing the same thing from different angles. Fifth, I understand that the author wrote this paper many months ago, and they don't want to improve their paper because they think many \"advance papers\" are later than theirs. But I encourage the authors to improve their paper with more advanced insights to catch up with the recent developments. For example, the author could propose some practical solutions to ViT robustness. I also have some experience that I was the first to propose some techniques, but my paper was rejected. Then, the later paper working on a similar idea was published. I was frustrated. But I continued to improve my paper, and finally, my paper was accepted by another venue as an oral presentation. \n\n\n[r1] Intriguing Properties of Vision Transformers\n\n[r2] Vision Transformers are Robust Learners\n\n[r3] RobustART: Benchmarking Robustness on Architecture Design and Training Techniques\n\n[r4] Understanding Robustness of Transformers for Image Classification (ICCV 2021) 26 Mar 2021\n\n[r5] Feature Denoising for Improving Adversarial Robustness.\n\n\n(Negative) It would be good not to say that this work provides the first study on the robustness of vision transformers (ViTs) against adversarial perturbations. There have been other works discussing the ViT adversarial robustness. It would be good for the author to recognize the contribution of the community. It would be good for the authors to be humble.\n\n(Negative) The following observations are natural and not new: \"we find that ViTs possess better adversarial robustness when compared with convolutional neural networks (CNNs).\"\n\n(Negative) The following observations are natural and not new: \"Features learned by ViTs contain less low-level information and are more generalizable, which contributes to superior robustness against adversarial perturbations.\" Existing papers found that ViTs do better than CNNs in adversarial robustness but do worse than CNNs in natural noise robustness.\n\n(Negative) The following observations are complementary to the above observations. They thus are less informative: \"Introducing convolutional or tokens-to-token blocks for learning low-level features in ViTs can improve classification accuracy but at the cost of adversarial robustness.\"\n\n(Negative) The following observations are complementary to the above observations. So these observations are less informative. Increasing the proportion of transformers in the model structure (when the model consists of both transformer and CNN blocks) leads to better robustness. This makes readers feel that there are not many valuable discoveries in this article, so the author keeps discussing things from different angles.\n\n(Negative) I'm afraid I have to disagree with the authors' following claim: \"Pre-training on larger datasets does not significantly improve adversarial robustness though it is critical for training ViTs.\" I think the authors performed the wrong pre-training. Pre-training on the larger dataset should be robust pre-training, i.e., it should be adversarial training, but NOT vanilla pre-training. In my experiments, doing robust pre-training on a larger dataset significantly improves adversarial robustness.\n\n\n(Negative) The following observations are too natural and contain little information: \"ViTs are less sensitive to high-frequency perturbations than CNNs and\nthere is a high correlation between how well the model learns low-level features and its robustness against different frequency-based perturbations.\"\n\n(Negative) The following descriptions need to be corrected: \"it remains unclear on the robustness of ViT against adversarial perturbations, which is critical for safe and reliable deployment of many real-world applications,\" given that the literature has fully studied the robustness of ViT.\n\n\n(Negative) The following descriptions need to be corrected: \"In this work, we conduct the first study on examining the adversarial robustness of ViTs on image classification tasks and make comparisons with CNN baselines.\"\n\n\n(Neutral) The following observation is natural and ok: \" Using denoised randomized smoothing (Salman et al., 2020), ViTs attain significantly better-certified robustness than CNNs.\"\n\n(Positive) The following description is correct and insightful: \"short-term memory (LSTM) or CNN, with a theoretical explanation provided in Hsieh et al. (2019). However, due to the discrete nature of NLP models, these studies are focusing on discrete perturbations (e.g., word or character substitutions) which are very different from small and continuous perturbations in computer vision tasks.\"\n\n(Negative) The following description needs to be modified: \" To the best of our knowledge, this work is the first study that investigates the adversarial robustness (against small perturbations in the input pixel space) of transformers on computer vision tasks.\"\n\n(Negative) The following description needs to be modified: \"In the context of computer vision, the most relevant work is Alamri et al. (2020), which applies transformer encoder in the object detection task and reports better adversarial robustness.\"\n\n(Negative) In the related work section, all the above-mentioned works should be discussed and compared to acknowledge the community's contribution.\n\n(Positive) I am happy to see that the authors also evaluate the certified robustness of the models using randomized smoothing, where the robustness is evaluated as the certified radius,\n\n(Negative) The following observations are too natural: \" Adversarial training can be applied to train robust ViTs.\"\n\n\n(Positive) The following finding is interesting and valuable: \"We conjecture that ViT may need larger training data or longer training epochs to improve further its robust training performance, inspired by the fact that on natural training ViT is not able to perform well either without large-scale pre-training.\"\n\n\n(Positive) The results in Table 4 are fascinating, valuable, and beneficial. I really like the results in this table.\n\n\n(Negative) Please use text instead of figures to give a sub-caption to each subfigure in Figure 4. The current version is lossy.\n", "summary_of_the_review": "\nBalancing the strengths and weaknesses of the proposed method, I would like to recommend a rating of weak rejection for this paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "\nNone.\n\n", "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635867194502}, {"id": "569wz5mdTPK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper595/Reviewer_x14o"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides a comprehensive study on the robustness of ViTs against adversarial perturbations. The authors found that 1) ViTs has better adversarial robustness than convolutional neural networks; 2) Introducing convolutional or tokens-to-token blocks can improve the classification accuracy but at the cost of the adversarial robustness; 3) More proportion of transformers has better robustness; 4) Pre-training on larger datasets does not improve adversarial robustness; 5) Adversarial training is applicable to ViTs. In addition, many experiments verify the findings on white-box, transfer attack settings and adversarial training.  ", "review_text": "1. Some findings lack explanations and reasons. For example, it is important to give a reason why ViTs have better adversarial robustness than convolutional neural networks. Does this advantage come from the splitted several tokens in ViTs? If yes, could you please conduct more experiments on different sizes of tokens? It would be better to provide more explanations for this.\n\n2. More general Transformer architectures should be explored in the paper. The authors mainly focus on ViTs, and the observations are mainly for ViTs. Recently, Swin Transformer gains more attention in computer vision tasks. The authors only discuss one type of Swin Transformer, i.e., Swin-S/4. It would be better to explore more types of Transformer and draw conclusions for the more general cases of Transformer.\n\n3. The experiments should be improved.  The authors only consider PGD attack and AutoAttack. These two types of attack methods are linear, i.e., add small perturbation on clean data. It would be better to conduct more experiments on other nonlinear attack methods, e.g., ADef [1], which applies small deformations to the clean data. \n\n    [1] ADef : an Iterative Algorithm to Construct Adversarial Deformations. ICLR 2019\n\n4. In Table 1, the attack success rate (ASR) of  Deit-S/16 is the best result, but the ASR of ViT-L/16 is 98.2. Is there a mistake in these two results?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides a comprehensive study on the robustness of ViTs against adversarial perturbations. The authors found that 1) ViTs has better adversarial robustness than convolutional neural networks; 2) Introducing convolutional or tokens-to-token blocks can improve the classification accuracy but at the cost of the adversarial robustness; 3) More proportion of transformers has better robustness; 4) Pre-training on larger datasets does not improve adversarial robustness; 5) Adversarial training is applicable to ViTs. In addition, many experiments verify the findings on white-box, transfer attack settings and adversarial training.  ", "main_review": "1. Some findings lack explanations and reasons. For example, it is important to give a reason why ViTs have better adversarial robustness than convolutional neural networks. Does this advantage come from the splitted several tokens in ViTs? If yes, could you please conduct more experiments on different sizes of tokens? It would be better to provide more explanations for this.\n\n2. More general Transformer architectures should be explored in the paper. The authors mainly focus on ViTs, and the observations are mainly for ViTs. Recently, Swin Transformer gains more attention in computer vision tasks. The authors only discuss one type of Swin Transformer, i.e., Swin-S/4. It would be better to explore more types of Transformer and draw conclusions for the more general cases of Transformer.\n\n3. The experiments should be improved.  The authors only consider PGD attack and AutoAttack. These two types of attack methods are linear, i.e., add small perturbation on clean data. It would be better to conduct more experiments on other nonlinear attack methods, e.g., ADef [1], which applies small deformations to the clean data. \n\n    [1] ADef : an Iterative Algorithm to Construct Adversarial Deformations. ICLR 2019\n\n4. In Table 1, the attack success rate (ASR) of  Deit-S/16 is the best result, but the ASR of ViT-L/16 is 98.2. Is there a mistake in these two results?\n", "summary_of_the_review": "It would be better to provide more explanations for some findings, and further improve the experiments for more general Transformer cases and different types of attack methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None", "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635728263766}, {"id": "uI0DvmaICvC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper595/Reviewer_WHoT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work studies the adversarial robustness of vision transformers comprehensively. The author first presents their finding via empirical experiments. Concretely, they show the robustness of vision transformers from the perspective of adversarial attack, transferability, and certified robustness, adversarial training defense. Then, they analyze the reason behind the observation. ", "review_text": "Strengths:\nThe study on the adversarial robustness of ViT is comprehensive.\nThe analyst behind the observation is provided.\nThe paper is clearly written and easy to follow.\n\nWeakness:\n1. The authors state that ViT is more robust than CNN. It is only true when a weak attack is applied. When a standard attack (with perturbation range epsilon=0.03/0.01), both ViT and CNN can be fooled with 100% fooling rate. The observation indicates that both are equally vulnerable under standard attack. Hence, the statements or claims made in this paper should be carefully formulated.\n\n2. The observations on the difference in model robustness are presented in this work. However, the difference is always attributed to the model architecture difference in this work. Compared to ResNet, the ViT and DeiT are trained/finetuned in a different setting, e.g. different data augmentation, different training schedules. It is not clear how the different training/finetuning settings contribute to the different model robustness. It might be too early to attribute the observation to model architectures.\n\n3. Similarly, the author lists the following point as one of their findings:\nPre-training on larger datasets does not improve adversarial robustness though it is critical for training ViT.\nHowever, previous work shows pre-training can improve model robustness [1]. More discussion should be provided hier. The claims should be careful.
\n4. In Figure 3, the author shows ViT-small shows higher certified robustness than ResNet18. In the work DeiT, they show that the CNN counterpart of DesT-small is ResNet50 instead of ResNet18. If it is too hard to compare them in a 100% fair fashion. Please show the curves of both ResNet50 and ResNet18.\n\n5. The adversarial training on ViT is also studied in this work. However, the experiments are conducted on CIFAR10 datasets. The ViT is expected to behave well only on a large dataset. What is the performance of ViT in a standard-setting on CIFAR10 dataset?\n\n[1] Hendrycks, Dan, Kimin Lee, and Mantas Mazeika. \"Using pre-training can improve model robustness and uncertainty.\" International Conference on Machine Learning. PMLR, 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work studies the adversarial robustness of vision transformers comprehensively. The author first presents their finding via empirical experiments. Concretely, they show the robustness of vision transformers from the perspective of adversarial attack, transferability, and certified robustness, adversarial training defense. Then, they analyze the reason behind the observation. ", "main_review": "Strengths:\nThe study on the adversarial robustness of ViT is comprehensive.\nThe analyst behind the observation is provided.\nThe paper is clearly written and easy to follow.\n\nWeakness:\n1. The authors state that ViT is more robust than CNN. It is only true when a weak attack is applied. When a standard attack (with perturbation range epsilon=0.03/0.01), both ViT and CNN can be fooled with 100% fooling rate. The observation indicates that both are equally vulnerable under standard attack. Hence, the statements or claims made in this paper should be carefully formulated.\n\n2. The observations on the difference in model robustness are presented in this work. However, the difference is always attributed to the model architecture difference in this work. Compared to ResNet, the ViT and DeiT are trained/finetuned in a different setting, e.g. different data augmentation, different training schedules. It is not clear how the different training/finetuning settings contribute to the different model robustness. It might be too early to attribute the observation to model architectures.\n\n3. Similarly, the author lists the following point as one of their findings:\nPre-training on larger datasets does not improve adversarial robustness though it is critical for training ViT.\nHowever, previous work shows pre-training can improve model robustness [1]. More discussion should be provided hier. The claims should be careful.
\n4. In Figure 3, the author shows ViT-small shows higher certified robustness than ResNet18. In the work DeiT, they show that the CNN counterpart of DesT-small is ResNet50 instead of ResNet18. If it is too hard to compare them in a 100% fair fashion. Please show the curves of both ResNet50 and ResNet18.\n\n5. The adversarial training on ViT is also studied in this work. However, the experiments are conducted on CIFAR10 datasets. The ViT is expected to behave well only on a large dataset. What is the performance of ViT in a standard-setting on CIFAR10 dataset?\n\n[1] Hendrycks, Dan, Kimin Lee, and Mantas Mazeika. \"Using pre-training can improve model robustness and uncertainty.\" International Conference on Machine Learning. PMLR, 2019.", "summary_of_the_review": "ViT demonstrates the potential to work as an alternative to CNNs. The adversarial robustness of ViT is indeed an important topic. As claimed by the author, this work makes the first comprehensive study on this topic. However, some concerns listed in weaknesses above remain to be addressed. Therefore, I rate this paper blew the acceptance threshold. I happy to raise my rating if the concerns are well addressed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635578272031}], "openreview_url": "https://openreview.net/forum?id=O0g6uPDLW7", "arxiv_id": "2103.15670", "paper_pdf": "papers/O0g6uPDLW7.pdf", "paper_pdf_sha256": "908b4c9fe1888894ad87afc44bd7be46ff3d91db7295bf5d3bfe1c0cd5c0f249", "paper_pdf_bytes": 23487093, "paper_pdf_source": "openreview", "code_url": "https://github.com/RulinShao/on-the-adversarial-robustness-of-visual-transformer", "code_repository": "RulinShao/on-the-adversarial-robustness-of-visual-transformer", "code_commit": "1cc19595c32b0c80f8a1a72d36e041844b339433", "code_archive": "repos/O0g6uPDLW7.zip", "code_archive_sha256": "cf9d74202d224ae957625c6d483287154b9f0fd65bc69baee6b9bff8551893d5", "code_archive_bytes": 269547, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 287, "github_languages": {"Python": 70597}, "github_archived": false, "github_pushed_at": "2021-11-18T18:31:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/2103-15670"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "szUsQ3NcQwV", "year": 2021, "status": "rejected", "title": "Randomized Entity-wise Factorization for Multi-Agent Reinforcement Learning", "authors": ["Shariq Iqbal", "Christian Schroeder de Witt", "Bei Peng", "Wendelin Boehmer", "Shimon Whiteson", "Fei Sha"], "authorids": ["~Shariq_Iqbal1", "~Christian_Schroeder_de_Witt1", "~Bei_Peng2", "~Wendelin_Boehmer1", "~Shimon_Whiteson1", "~Fei_Sha3"], "authors_source": "OpenReview API", "abstract": "Real world multi-agent tasks often involve varying types and quantities of agents and non-agent entities; however, agents within these tasks rarely need to consider all others at all times in order to act effectively. Factored value function approaches have historically leveraged such independences to improve learning efficiency, but these approaches typically rely on domain knowledge to select fixed subsets of state features to include in each factor. We propose to utilize value function factoring with random subsets of entities in each factor as an auxiliary objective in order to disentangle value predictions from irrelevant entities. This factoring approach is instantiated through a simple attention mechanism masking procedure. We hypothesize that such an approach helps agents learn more effectively in multi-agent settings by discovering common trajectories across episodes within sub-groups of agents/entities. Our approach, Randomized Entity-wise Factorization for Imagined Learning (REFIL), outperforms all strong baselines by a significant margin in challenging StarCraft micromanagement tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "7znBztGeoJg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1939/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper proposes to incorporate a masked attention mechanism in QMIX for value function factorization to disentangle value predictions from irrelevant agents/entities. The masking is based on a random sampling from the whole set of agents to from random subsets, based on which it can compute within-group and without-group Q-functions. The method is able to handle varying types and number of agents. The paper conducts experiments on a simple game to understand the effect, and then test on 3 SMAC games, which shows the effectiveness of the proposed REFIL method.\n\nStrong points:\n- The paper is well-written and clear.\n- The paper studies an important topic in MARL, i.e., how to deal with varying types and number of agents, and propose a simple yet effective approach, which incorporates attention mechanism in QMIX with random masking.\n- Experiments on several games illustrate the effectiveness of the method, with proper ablation study to understand the importance of each component (attention, mixing network, random masking).\n\nConcerns:\nThe main focus of the paper is to “disentangle value predictions from irrelevant entities”. However, not only does REFIL relies on Q_I (within-group Q-function), but also it relies on Q_O (without group Q-function). If the agent successfully learn this neglect of irrelevant entities, focusing on Q_I would be enough, without triggering additional computation of unnecessary Q_O. Could authors better explain this? In addition, consider the breakaway example, as the attacker has only to focus on the goal keeper, is the random sampling scheme from all agents effective compared with counterparts that only need to focus on the goalkeeper? Could authors conduct additional experiments on football to better support the claim?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting problem, simple but effective method", "review": "Summary: This paper proposes to incorporate a masked attention mechanism in QMIX for value function factorization to disentangle value predictions from irrelevant agents/entities. The masking is based on a random sampling from the whole set of agents to from random subsets, based on which it can compute within-group and without-group Q-functions. The method is able to handle varying types and number of agents. The paper conducts experiments on a simple game to understand the effect, and then test on 3 SMAC games, which shows the effectiveness of the proposed REFIL method.\n\nStrong points:\n- The paper is well-written and clear.\n- The paper studies an important topic in MARL, i.e., how to deal with varying types and number of agents, and propose a simple yet effective approach, which incorporates attention mechanism in QMIX with random masking.\n- Experiments on several games illustrate the effectiveness of the method, with proper ablation study to understand the importance of each component (attention, mixing network, random masking).\n\nConcerns:\nThe main focus of the paper is to “disentangle value predictions from irrelevant entities”. However, not only does REFIL relies on Q_I (within-group Q-function), but also it relies on Q_O (without group Q-function). If the agent successfully learn this neglect of irrelevant entities, focusing on Q_I would be enough, without triggering additional computation of unnecessary Q_O. Could authors better explain this? In addition, consider the breakaway example, as the attacker has only to focus on the goal keeper, is the random sampling scheme from all agents effective compared with counterparts that only need to focus on the goalkeeper? Could authors conduct additional experiments on football to better support the claim?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604062033177}, {"id": "RgP1vHeKUDw", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1939/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method for randomized factorization of multi-agents for efficient learning. The idea is inspired by the presence of irrelevant entities present in an agent's observational view and how removing those could aid the learning process.\nAgents are randomly divided into groups so that an agent can separately measure the influence of entities present in the same group and the entities present in the other groups. Since the groups are randomized, this helps the agent to create groups of variable size based on its utility prediction of in-group and out-of-group entities.\n\nAlthough the paper only uses two groups for derivation and experiments, it claims that the same method can be applied to more than two groups but is yet to be demonstrated.\n\n\nThe paper is easy to read and the figures are excellent and self-explanatory. The environments chosen are also indicative of the importance of each component. The authors also found that randomized factoring performs better empirically than using domain knowledge while the state-of-the-art lies in the combination of the two. Further, the idea of training variable-sized hypernetwork is quite fascinating and fits well in the overall framework.\n\nSome questions:\nFigure 3b does not show results till convergence, please put the entire plots.\n\nI think it would be better if the SMAC setup is explained in more detail. I couldn't understand how the tagging is done (in Appendix). Does it deterministically tie an action to an enemy?\n\nQTRAN has been shown to perform better than QMIX in competitive domains. I would encourage the authors to compare the results with this baseline too.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel approach to factorize entities in a multi-agent env with supportive empirical evidence", "review": "The paper proposes a method for randomized factorization of multi-agents for efficient learning. The idea is inspired by the presence of irrelevant entities present in an agent's observational view and how removing those could aid the learning process.\nAgents are randomly divided into groups so that an agent can separately measure the influence of entities present in the same group and the entities present in the other groups. Since the groups are randomized, this helps the agent to create groups of variable size based on its utility prediction of in-group and out-of-group entities.\n\nAlthough the paper only uses two groups for derivation and experiments, it claims that the same method can be applied to more than two groups but is yet to be demonstrated.\n\n\nThe paper is easy to read and the figures are excellent and self-explanatory. The environments chosen are also indicative of the importance of each component. The authors also found that randomized factoring performs better empirically than using domain knowledge while the state-of-the-art lies in the combination of the two. Further, the idea of training variable-sized hypernetwork is quite fascinating and fits well in the overall framework.\n\nSome questions:\nFigure 3b does not show results till convergence, please put the entire plots.\n\nI think it would be better if the SMAC setup is explained in more detail. I couldn't understand how the tagging is done (in Appendix). Does it deterministically tie an action to an enemy?\n\nQTRAN has been shown to perform better than QMIX in competitive domains. I would encourage the authors to compare the results with this baseline too.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603933913270}, {"id": "81o2hcEfXDE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1939/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an observation factorization method to avoid the influence of the irrelevant part on value estimation. Specifically, they design an entity-wise attention network with a masking procedure. This network is used to filter the irrelevant part of the original observation of each agent. Then the output is used to estimate the individual q-value, as well as input to the mixing network to generate the Q_tot. Two kinds of Q_tot are trained together by combing two loss functions linearly with a hyper-parameter. Experimental results show REFIL combined with QMIX surpasses vanilla QMIX and VDN in several SMAC scenarios.\n\nThis paper is related to the topics of ICLR. However, I think the related work is not sufficient to cover the background. More detail comments can be found below. \n\n*****Some specific comments:*****\n\nIt is not clear that what is the initialization of two masks, and how to update the masks. \n\nThe authors mentioned there are two groups of entities. However, the entity type is also unclear. I guess SMAC only contains two entity types: alive agents and died agents? How to represent an entity inactive?\n\nOne question is why just consider two kinds of groups, what would happen if there exist more than two groups for all entities. In SMAC or soccer, it does contain more than two common patterns. Furthermore, it seems that the masking procedure is hard to extend to the situation with a larger number of groups.\n\nActually, I think REFIL is similar to ROMA [1] and ASN [2] in different ways. First, REFIL considers two kinds of groups corresponding to a simple version of ROMA which has two roles. Second, REFIL does the same thing as ASN that learns the value estimation by considering a more useful part of the observation. ASN directly divide the observation based on the action semantics, while REFIL tries to learn a suitable observation factorization through entity-wise attention with masking. However, these two very relevant works are not discussed and compared in this paper.\n\nSome suggestions,\n\nI think current experiments could not well support motivation. If authors show some examples in SMAC that what kinds of common patterns agents learn would be better to support this idea.\n\nSince REFIL can be integrated into current MARL algorithms, it is better to consider more recent published MARL methods as baselines, such as QTRAN, QATTEN, QPLEX.\n\n[1] Roma: Multi-agent reinforcement learning with emergent roles. ICML. 2020.\n\n[2] Action Semantics Network: Considering the Effects of Actions in Multiagent Systems. ICLR. 2020.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper proposes an observation factorization method to avoid the influence of the irrelevant part on value estimation. Specifically, they design an entity-wise attention network with a masking procedure. This network is used to filter the irrelevant part of the original observation of each agent. Then the output is used to estimate the individual q-value, as well as input to the mixing network to generate the Q_tot. Two kinds of Q_tot are trained together by combing two loss functions linearly with a hyper-parameter. Experimental results show REFIL combined with QMIX surpasses vanilla QMIX and VDN in several SMAC scenarios.\n\nThis paper is related to the topics of ICLR. However, I think the related work is not sufficient to cover the background. More detail comments can be found below. \n\n*****Some specific comments:*****\n\nIt is not clear that what is the initialization of two masks, and how to update the masks. \n\nThe authors mentioned there are two groups of entities. However, the entity type is also unclear. I guess SMAC only contains two entity types: alive agents and died agents? How to represent an entity inactive?\n\nOne question is why just consider two kinds of groups, what would happen if there exist more than two groups for all entities. In SMAC or soccer, it does contain more than two common patterns. Furthermore, it seems that the masking procedure is hard to extend to the situation with a larger number of groups.\n\nActually, I think REFIL is similar to ROMA [1] and ASN [2] in different ways. First, REFIL considers two kinds of groups corresponding to a simple version of ROMA which has two roles. Second, REFIL does the same thing as ASN that learns the value estimation by considering a more useful part of the observation. ASN directly divide the observation based on the action semantics, while REFIL tries to learn a suitable observation factorization through entity-wise attention with masking. However, these two very relevant works are not discussed and compared in this paper.\n\nSome suggestions,\n\nI think current experiments could not well support motivation. If authors show some examples in SMAC that what kinds of common patterns agents learn would be better to support this idea.\n\nSince REFIL can be integrated into current MARL algorithms, it is better to consider more recent published MARL methods as baselines, such as QTRAN, QATTEN, QPLEX.\n\n[1] Roma: Multi-agent reinforcement learning with emergent roles. ICML. 2020.\n\n[2] Action Semantics Network: Considering the Effects of Actions in Multiagent Systems. ICLR. 2020.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603891403289}, {"id": "oW-6pS6IGRq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1939/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a randomized entity-based attentional mechanism to regularize the observation space for efficient multi-agent reinforcement learning. Specifically, the authors expect that their method can help agents focus on entities that are relevant to their decision-making process. The aim of the paper is well-positioned in multi-agent settings and are expected to help improve performance by exploiting the loosely coupled structure of multi-agent tasks (although the authors do not explicitly model the decision dependency among agents). However, I have some doubts about whether the proposed method can address the target of the paper.\n\nIntuitively, in multi-agent settings, the number of entities is at least the number of agents, say O(n). O(n) is a quite optimistic estimation because we even do not consider other entities. If the authors want to find an optimal bi-partition over the entity space for each agent, the search space is at least O(2^n), which grows exponentially with the number of agents. To design an efficient search algorithm over such a large space needs to take advantage of some well-designed inductive bias or heuristics. The authors use a random strategy here, which, in my opinion, is not sufficient to guarantee a satisfactory solution. There is no denying that randomization can give a good solution in some cases, but this can not be held as a general rule. Even if the bi-partition structure is trained end-to-end, I also suspect that learning such a structure is not easier than learning from scratch.\n\nAdditionally, I think the authors largely ignore the contribution of a related work [Agarwal et al., 2020]. Although they cite this paper in Sec. 5, unlike what is stated in this paper, the main contribution of [Agarwal et al., 2020] is a GNN-based attentional mechanism over entity spaces. They propose to let agents learn to attend to different entities under different observations. In this way, the target of [Agarwal et al., 2020] and this paper largely overlap. I was expecting that the authors provide a thorough comparison with [Agarwal et al., 2020] in their experiments. If the authors can demonstrate that their method can outperform [Agarwal et al., 2020], I will consider improve my rating.\n\n\n[Agarwal et al., 2020] Agarwal, A., Kumar, S., Sycara, K. and Lewis, M., 2020, May. Learning Transferable Cooperative Behavior in Multi-Agent Teams. In Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems (pp. 1741-1743).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A more serious discussion about [Agarwal et al., 2020] is expected.", "review": "This paper introduces a randomized entity-based attentional mechanism to regularize the observation space for efficient multi-agent reinforcement learning. Specifically, the authors expect that their method can help agents focus on entities that are relevant to their decision-making process. The aim of the paper is well-positioned in multi-agent settings and are expected to help improve performance by exploiting the loosely coupled structure of multi-agent tasks (although the authors do not explicitly model the decision dependency among agents). However, I have some doubts about whether the proposed method can address the target of the paper.\n\nIntuitively, in multi-agent settings, the number of entities is at least the number of agents, say O(n). O(n) is a quite optimistic estimation because we even do not consider other entities. If the authors want to find an optimal bi-partition over the entity space for each agent, the search space is at least O(2^n), which grows exponentially with the number of agents. To design an efficient search algorithm over such a large space needs to take advantage of some well-designed inductive bias or heuristics. The authors use a random strategy here, which, in my opinion, is not sufficient to guarantee a satisfactory solution. There is no denying that randomization can give a good solution in some cases, but this can not be held as a general rule. Even if the bi-partition structure is trained end-to-end, I also suspect that learning such a structure is not easier than learning from scratch.\n\nAdditionally, I think the authors largely ignore the contribution of a related work [Agarwal et al., 2020]. Although they cite this paper in Sec. 5, unlike what is stated in this paper, the main contribution of [Agarwal et al., 2020] is a GNN-based attentional mechanism over entity spaces. They propose to let agents learn to attend to different entities under different observations. In this way, the target of [Agarwal et al., 2020] and this paper largely overlap. I was expecting that the authors provide a thorough comparison with [Agarwal et al., 2020] in their experiments. If the authors can demonstrate that their method can outperform [Agarwal et al., 2020], I will consider improve my rating.\n\n\n[Agarwal et al., 2020] Agarwal, A., Kumar, S., Sycara, K. and Lewis, M., 2020, May. Learning Transferable Cooperative Behavior in Multi-Agent Teams. In Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems (pp. 1741-1743).", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603886743080}], "openreview_url": "https://openreview.net/forum?id=szUsQ3NcQwV", "arxiv_id": "2006.04222", "paper_pdf": "papers/szUsQ3NcQwV.pdf", "paper_pdf_sha256": "bf980830398d7625fd872560b1f0a2b47f308ca45091a21c963b899aae93baae", "paper_pdf_bytes": 3270420, "paper_pdf_source": "openreview", "code_url": "https://github.com/shariqiqbal2810/REFIL", "code_repository": "shariqiqbal2810/REFIL", "code_commit": "ffe23a1c62fd3f7304e700e69b30dcb0c29586cc", "code_archive": "repos/szUsQ3NcQwV.zip", "code_archive_sha256": "1f3565013c6f59005f205d0a955539d1b3883e65ef47f3cf682c5caa3e6b5820", "code_archive_bytes": 454521, "code_file_count": 43, "code_extensions": {".py": 39, ".sh": 4}, "github_disk_usage_kb": 321, "github_languages": {"Python": 264316, "Shell": 1958, "Dockerfile": 1657}, "github_archived": false, "github_pushed_at": "2021-05-22T22:46:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ai-qmix-attention-and-imagination-for-dynamic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkxHRySFvr", "year": 2020, "status": "rejected", "title": "LEARNING TO IMPUTE: A GENERAL FRAMEWORK FOR SEMI-SUPERVISED LEARNING", "authors": ["Wei-Hong Li", "Chuan-Sheng Foo", "Hakan Bilen"], "authorids": ["w.h.li@ed.ac.uk", "foo_chuan_sheng@i2r.a-star.edu.sg", "hbilen@ed.ac.uk"], "authors_source": "OpenReview API", "abstract": "Recent semi-supervised learning methods have shown to achieve comparable results to their supervised counterparts while using only a small portion of labels in image classification tasks thanks to their regularization strategies. In this paper, we take a more direct approach for semi-supervised learning and propose learning to impute the labels of unlabeled samples such that a network achieves better generalization when it is trained on these labels. We pose the problem in a learning-to-learn formulation which can easily be incorporated to the state-of-the-art semi-supervised techniques and boost their performance especially when the labels are limited. We demonstrate that our method is applicable to both classification and regression problems including image classification and facial landmark detection tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Syeqoi3d9r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2024/AnonReviewer5"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper uses a meta-learning approach to solve semi-supervised learning. The main idea is to simulate an SGD step on the loss of the meta-validation data and see how the model will perform if the pseudo-labels of unlabelled data are perturbed. Experiments on classification and regression problems show that the proposed method can improve over existing methods. The idea itself is intriguing but the derivation and some design choice are not very well-explained.\n\n(1) The derivation from Eq.(3) to (4) is confusing. Note that in Eq.(3), the prediction \\Phi_\\theta also depends on \\theta in addition to the pseudo-label z. When taking a step of SGD, the second term of Eq.(3) (with unlabelled data) will always be zero if both arguments of the loss (\\Phi_\\theta(x) and z_\\theta(x)) change simultaneously. Eq.(4) somehow only considers the gradient of unsupervised loss, then the gradient would be zero because there is no incentive to deviate from the pseudo-label z. The pseudo-code does not help much. The update from \\hat{\\theta}^{t} to \\hat{\\theta}^{t+1} has the same issue: there is no incentive for \\hat{\\theta}^{t} to deviate because z is exactly produced by it.\n\n(2) For classification problems, it is natural to use cross-entropy loss for the probability vector z. Are there any specific reasons for using Gumbel-softmax? In addition, using L2 loss for probability vectors (as mentioned in Appendix A) is known to be problematic as it may create exponentially many local minima (Auer et al, 1996).\n\n(3) The recent work of Li et al. (2019) also considers iteratively improving pseudo-labels with meta-updates so it should be discussed and compared.\n\n(4) Experiments\n- What are the sizes of the meta-validation sets in the experiments?\n- Error bars in the tables and Fig.2?\n- The MM results in Table 2 are noticeably worse than the original results. For example, with 250 labeled data, MM achieved 11.08% in CIFAR-10 as reported in the original paper. (And 4000 labeled data can achieve 4.95%)\n- It is said that option 2 is consistently better than option 1, which is not true for the MM baseline.\n- 22500 training steps for Experiment 4 seems arbitrary. What are the candidates for the hyper-parameters?\n\nTypos:\n- In the first paragraph of Sec.2, one of the x and one of the y should be bold.\n- Above Eq.(4), x^{U\\in U} should be x^i \\in U\n- The transpose in Eq.(7) is not necessary\n- It is said on page 6 that Fig.2 reports classification loss but the task is a regression problem.\n\nRef\n- Auer, P., Herbster, M. and Warmuth, M.K., 1996. Exponentially many local minima for single neurons. In Advances in neural information processing systems (pp. 316-322).\n- Li, X., Sun, Q., Liu, Y., Zheng, S., Chua, T.S. and Schiele, B., 2019. Learning to Self-Train for Semi-Supervised Few-Shot Classification. In Advances in neural information processing systems.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #5", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper uses a meta-learning approach to solve semi-supervised learning. The main idea is to simulate an SGD step on the loss of the meta-validation data and see how the model will perform if the pseudo-labels of unlabelled data are perturbed. Experiments on classification and regression problems show that the proposed method can improve over existing methods. The idea itself is intriguing but the derivation and some design choice are not very well-explained.\n\n(1) The derivation from Eq.(3) to (4) is confusing. Note that in Eq.(3), the prediction \\Phi_\\theta also depends on \\theta in addition to the pseudo-label z. When taking a step of SGD, the second term of Eq.(3) (with unlabelled data) will always be zero if both arguments of the loss (\\Phi_\\theta(x) and z_\\theta(x)) change simultaneously. Eq.(4) somehow only considers the gradient of unsupervised loss, then the gradient would be zero because there is no incentive to deviate from the pseudo-label z. The pseudo-code does not help much. The update from \\hat{\\theta}^{t} to \\hat{\\theta}^{t+1} has the same issue: there is no incentive for \\hat{\\theta}^{t} to deviate because z is exactly produced by it.\n\n(2) For classification problems, it is natural to use cross-entropy loss for the probability vector z. Are there any specific reasons for using Gumbel-softmax? In addition, using L2 loss for probability vectors (as mentioned in Appendix A) is known to be problematic as it may create exponentially many local minima (Auer et al, 1996).\n\n(3) The recent work of Li et al. (2019) also considers iteratively improving pseudo-labels with meta-updates so it should be discussed and compared.\n\n(4) Experiments\n- What are the sizes of the meta-validation sets in the experiments?\n- Error bars in the tables and Fig.2?\n- The MM results in Table 2 are noticeably worse than the original results. For example, with 250 labeled data, MM achieved 11.08% in CIFAR-10 as reported in the original paper. (And 4000 labeled data can achieve 4.95%)\n- It is said that option 2 is consistently better than option 1, which is not true for the MM baseline.\n- 22500 training steps for Experiment 4 seems arbitrary. What are the candidates for the hyper-parameters?\n\nTypos:\n- In the first paragraph of Sec.2, one of the x and one of the y should be bold.\n- Above Eq.(4), x^{U\\in U} should be x^i \\in U\n- The transpose in Eq.(7) is not necessary\n- It is said on page 6 that Fig.2 reports classification loss but the task is a regression problem.\n\nRef\n- Auer, P., Herbster, M. and Warmuth, M.K., 1996. Exponentially many local minima for single neurons. In Advances in neural information processing systems (pp. 316-322).\n- Li, X., Sun, Q., Liu, Y., Zheng, S., Chua, T.S. and Schiele, B., 2019. Learning to Self-Train for Semi-Supervised Few-Shot Classification. In Advances in neural information processing systems."}, "tcdate": 1572551586210}, {"id": "BkxfphJd5H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2024/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper looks into problem of semi-supervised learning and in order to be mindful of generalization on the unlabeled data, they add a term to the loss function which includes loss on imputed labels.\nI have 3 main concerns with the paper \n1. The authors mention that  the meta-validation set is a random subset of train set. and they check the final performance on meta-validation set. This does not seem a right way to measure performance of the model as meta-validation set is already used in training. The set of labeled points should be partitioned into train and meta-validation set.\n\n2. The derivation of the updates given the added term to the loss. In option 1, the authors mention they use Eqn. 8 to update z, while Eqn 8. has the reverse information. \n\n3.In option 2, z = \\sigmoid(\\Phi_\\theta) and reducing the loss on z, l( \\sigmoid(\\Phi_\\theta) ,  \\Phi_\\theta), does not look very meaningful. trying to get  \\Phi_\\theta close to its sigmoid means getting it close to zero. but we do not know what is the label for unlabeled data, so why getting the label close to zero?\nAlso the authors mention that second order derivatives will come to play without any explanation. I suggest spending more effort on explaining the problem formulation as that's the core of the paper.\n\nMore comments:\n* As mentioned above the problem formulation is not clean and there are unjustified choice there. Moreover, the experiment results are mostly declared without any justification (for example, the proposed method does not always lead to improvement and not all cases are explained. The authors only note that the method works well in low data regime). \n\n* In the first experiment PL is compared to two cases of the proposed algorithm whereas in other experiments PL is compared to combining PL with versions of the proposed method. Is there a reason for this?\n\n* The models used as baseline are only explained briefly in the last page of the paper, while being used multiple time in the experiment section. This is not good writing practice.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper looks into problem of semi-supervised learning and in order to be mindful of generalization on the unlabeled data, they add a term to the loss function which includes loss on imputed labels.\nI have 3 main concerns with the paper \n1. The authors mention that  the meta-validation set is a random subset of train set. and they check the final performance on meta-validation set. This does not seem a right way to measure performance of the model as meta-validation set is already used in training. The set of labeled points should be partitioned into train and meta-validation set.\n\n2. The derivation of the updates given the added term to the loss. In option 1, the authors mention they use Eqn. 8 to update z, while Eqn 8. has the reverse information. \n\n3.In option 2, z = \\sigmoid(\\Phi_\\theta) and reducing the loss on z, l( \\sigmoid(\\Phi_\\theta) ,  \\Phi_\\theta), does not look very meaningful. trying to get  \\Phi_\\theta close to its sigmoid means getting it close to zero. but we do not know what is the label for unlabeled data, so why getting the label close to zero?\nAlso the authors mention that second order derivatives will come to play without any explanation. I suggest spending more effort on explaining the problem formulation as that's the core of the paper.\n\nMore comments:\n* As mentioned above the problem formulation is not clean and there are unjustified choice there. Moreover, the experiment results are mostly declared without any justification (for example, the proposed method does not always lead to improvement and not all cases are explained. The authors only note that the method works well in low data regime). \n\n* In the first experiment PL is compared to two cases of the proposed algorithm whereas in other experiments PL is compared to combining PL with versions of the proposed method. Is there a reason for this?\n\n* The models used as baseline are only explained briefly in the last page of the paper, while being used multiple time in the experiment section. This is not good writing practice."}, "tcdate": 1572498617882}, {"id": "BkgLDSmstB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2024/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a semi-supervised approach to impute the labels of unlabeled samples such that a network achieves better generalization when it is trained on these labels. The proposed strategy can be easily used to improve the state-of-the-art semi-supervised methods. It mainly uses a validation data set to evaluate the updating rules of the unlabeled samples with pseudo-labels. The proposed method is applicable to both classification and regression problems including image classification and facial landmark detection tasks, which has shown in the experiments. But the following should be improved in the following aspects: \n[1] In the proposed method, the model parameters are updated both on the unlabeled samples and validation data set. The experimental results show that such a strategy is effective to improve the performance of the state-of-the-art method. But why the strategy is effective should be further analyzed.\n[2] How the validation data can improve the generalization ability of the model should be given with theoretical analysis. Whether the size of the validation data has a great influence?\n[3] Some experimental settings are not clear. In the experiments, how many unlabeled data is labeled with pseudo-labels. For different size of the unlabeled data, how many samples should be used in the validation data to evaluate the model with pseudo labeled samples.\n[4] How to divide the training data and the validation data? Whether the validation data need much more that the training data? How about the results only with all the labeled samples, which can further improve the confidence of the proposed method.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a semi-supervised approach to impute the labels of unlabeled samples such that a network achieves better generalization when it is trained on these labels. The proposed strategy can be easily used to improve the state-of-the-art semi-supervised methods. It mainly uses a validation data set to evaluate the updating rules of the unlabeled samples with pseudo-labels. The proposed method is applicable to both classification and regression problems including image classification and facial landmark detection tasks, which has shown in the experiments. But the following should be improved in the following aspects: \n[1] In the proposed method, the model parameters are updated both on the unlabeled samples and validation data set. The experimental results show that such a strategy is effective to improve the performance of the state-of-the-art method. But why the strategy is effective should be further analyzed.\n[2] How the validation data can improve the generalization ability of the model should be given with theoretical analysis. Whether the size of the validation data has a great influence?\n[3] Some experimental settings are not clear. In the experiments, how many unlabeled data is labeled with pseudo-labels. For different size of the unlabeled data, how many samples should be used in the validation data to evaluate the model with pseudo labeled samples.\n[4] How to divide the training data and the validation data? Whether the validation data need much more that the training data? How about the results only with all the labeled samples, which can further improve the confidence of the proposed method.\n"}, "tcdate": 1571661150099}], "openreview_url": "https://openreview.net/forum?id=SkxHRySFvr", "arxiv_id": "1912.10364", "paper_pdf": "papers/SkxHRySFvr.pdf", "paper_pdf_sha256": "e79dffe4270c5ddd6cf2be625fa4e1a775f8b630eacd8b4277e376c0c0a441e8", "paper_pdf_bytes": 2760561, "paper_pdf_source": "openreview", "code_url": "https://github.com/VICO-UoE/L2I", "code_repository": "VICO-UoE/L2I", "code_commit": "0fda16b1d1eb353f9c34e3e9e1af3623d6f66575", "code_archive": "repos/SkxHRySFvr.zip", "code_archive_sha256": "11821666bb246f12dd54125c3e7ea45fb010a26c5e330e3ee86cfc08e0df80a5", "code_archive_bytes": 473092, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 429, "github_languages": {"Python": 222227}, "github_archived": false, "github_pushed_at": "2021-12-18T12:17:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-impute-a-general-framework-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZKkeA1G935", "year": 2026, "status": "rejected", "title": "Can LLMs Alleviate Catastrophic Forgetting in Graph Continual Learning? A Systematic Study", "authors": ["Ziyang Cheng", "Zhixun Li", "Yuhan Li", "Yixin Song", "Kangyi Zhao", "Dawei Cheng", "Jia Li", "Hong Cheng", "Jeffrey Xu Yu"], "authorids": ["~Ziyang_Cheng1", "~Zhixun_Li1", "~Yuhan_Li3", "~Yixin_Song2", "~Kangyi_Zhao1", "~Dawei_Cheng1", "~Jia_Li4", "~Hong_Cheng1", "~Jeffrey_Xu_Yu1"], "authors_source": "OpenReview API", "abstract": "Nowadays, real-world data, including graph-structure data, often arrives in a streaming manner, which means that learning systems need to continuously acquire new knowledge without forgetting previously learned information. Although substantial existing works attempt to address catastrophic forgetting in graph machine learning, they are all based on training from scratch with streaming data. With the rise of pretrained models, an increasing number of studies have leveraged their strong generalization ability for continual learning. Therefore, in this work, we attempt to answer whether large language models (LLMs) can mitigate catastrophic forgetting in graph continual learning}. We first evaluate the performance of LLMs and graph foundation models in graph continual learning scenarios, and found that with minimal modifications, they can easily achieve state-of-the-art results. Moreover, we found that certain current settings for graph continual learning tasks have significant flaws; it is possible to achieve zero forgetting with simple manipulations. Finally, based on extensive experiments, we propose a simple-yet-effective method, Simple Grpah Continual Learning (SimGCL), that surpasses the previous state-of-the-art baselines by around 20% under the rehearsal-free constraint.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "xybenD8Ttx", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1449/Reviewer_ZnXD"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper presents a systematic and timely investigation into the potential of large language models (LLMs) to mitigate catastrophic forgetting in graph continual learning (GCL). They introduce a more rigorous global testing setup and develop LLM4GCL, a comprehensive benchmark evaluating nine LLM-based and graph-enhanced LLM (GLM) methods across seven text-attributed graph datasets under both node-level class-incremental (NCIL) and few-shot NCIL (FSNCIL) settings. They propose SimGCL, a simple yet effective GLM-based approach that combines graph-structured prompting, LoRA-based instruction tuning, and training-free prototype classification. This method enables LLMs to capture graph structural information while maintaining strong generalization and avoiding parameter updates that induce forgetting. Extensive experiments demonstrate that SimGCL consistently outperforms existing GNN-, LLM-, and GLM-based baselines—achieving up to 20% higher accuracy under rehearsal-free conditions.", "review_text": "This paper presents a systematic and timely investigation into the potential of large language models (LLMs) to mitigate catastrophic forgetting in graph continual learning (GCL). They introduce a more rigorous global testing setup and develop LLM4GCL, a comprehensive benchmark evaluating nine LLM-based and graph-enhanced LLM (GLM) methods across seven text-attributed graph datasets under both node-level class-incremental (NCIL) and few-shot NCIL (FSNCIL) settings. They propose SimGCL, a simple yet effective GLM-based approach that combines graph-structured prompting, LoRA-based instruction tuning, and training-free prototype classification. This method enables LLMs to capture graph structural information while maintaining strong generalization and avoiding parameter updates that induce forgetting. Extensive experiments demonstrate that SimGCL consistently outperforms existing GNN-, LLM-, and GLM-based baselines—achieving up to 20% higher accuracy under rehearsal-free conditions.", "strengths": "1. Critical Re-evaluation of GCL Benchmarks:\nThe paper makes a valuable methodological contribution by identifying and empirically demonstrating a task ID leakage issue in existing Graph Continual Learning (GCL) benchmarks. This flaw—previously overlooked in the community—renders many reported results unreliable. By introducing a corrected global testing setup, the authors establish a fair and realistic evaluation framework for future GCL research.\n\n2. Bridging GCL and Foundation Models:\nConceptually, the paper establishes an important connection between graph continual learning and pretrained foundation models. By demonstrating that LLMs can serve as effective continual learners on graph-structured data when appropriately prompted, the study opens a new research direction for integrating graph reasoning with large-scale language models.\n\n3. Comprehensive and Insightful Experimental Analysis：\nThe paper provides a detailed and balanced experimental study, including comparisons among GNNs, LLMs, and GLMs across diverse datasets and settings. The analysis offers clear empirical insights—such as why GLMs may underperform and how prototype-based designs and model scaling influence continual learning performance—making the work informative beyond its own method.", "weaknesses": "1. Limited Theoretical Justification for SimGCL:\nWhile SimGCL shows strong empirical results, the paper provides little theoretical or analytical grounding for why the combination of graph prompts, LoRA fine-tuning, and prototype-based classification alleviates forgetting. The approach appears largely empirical, and the mechanism behind its robustness (e.g., whether prototype stability or prompt alignment is the key factor) is not formally analyzed. Adding theoretical reasoning or ablation-based evidence would strengthen the scientific depth of the contribution.\n\n2. Overstatement of SimGCL’s generality and simplicity:\nThe paper claims that SimGCL “greatly alleviates catastrophic forgetting” and “surpasses all existing baselines” under a rehearsal-free constraint. While the reported results do show strong gains (up to 20%), these improvements are dataset-dependent — performance drops significantly on sparse or long-session datasets (e.g., Arxiv-23, FSNCIL). Hence, calling it universally effective may be an overstatement; the method is empirically strong but not universally superior.\n\n3. Assertion of “efficiency” and “low cost”:\nThe authors emphasize SimGCL’s efficiency due to LoRA-based tuning and prototype inference. However, they provide no runtime, memory, or scaling benchmarks to substantiate this claim. Given that LLM fine-tuning is resource-intensive, this efficiency claim seems qualitative and not empirically validated, representing a mild overstatement.", "questions": "Q1. Clarification on the mechanism of forgetting alleviation\nYou claim that SimGCL “greatly alleviates catastrophic forgetting” through training-free prototype classification. Could you provide concrete evidence (e.g., forgetting curves, representation drift analysis, or embedding similarity metrics) showing that forgetting is truly reduced, rather than simply avoided by freezing parameters?\n\nQ2. Computational efficiency validation\nThe paper repeatedly highlights SimGCL’s “efficiency,” but no runtime or GPU memory statistics are presented. Could you quantify the actual training and inference cost (in FLOPs, GPU hours, or wall-clock time) compared to other LLM-based and GNN-based baselines?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a systematic and timely investigation into the potential of large language models (LLMs) to mitigate catastrophic forgetting in graph continual learning (GCL). They introduce a more rigorous global testing setup and develop LLM4GCL, a comprehensive benchmark evaluating nine LLM-based and graph-enhanced LLM (GLM) methods across seven text-attributed graph datasets under both node-level class-incremental (NCIL) and few-shot NCIL (FSNCIL) settings. They propose SimGCL, a simple yet effective GLM-based approach that combines graph-structured prompting, LoRA-based instruction tuning, and training-free prototype classification. This method enables LLMs to capture graph structural information while maintaining strong generalization and avoiding parameter updates that induce forgetting. Extensive experiments demonstrate that SimGCL consistently outperforms existing GNN-, LLM-, and GLM-based baselines—achieving up to 20% higher accuracy under rehearsal-free conditions.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. Critical Re-evaluation of GCL Benchmarks:\nThe paper makes a valuable methodological contribution by identifying and empirically demonstrating a task ID leakage issue in existing Graph Continual Learning (GCL) benchmarks. This flaw—previously overlooked in the community—renders many reported results unreliable. By introducing a corrected global testing setup, the authors establish a fair and realistic evaluation framework for future GCL research.\n\n2. Bridging GCL and Foundation Models:\nConceptually, the paper establishes an important connection between graph continual learning and pretrained foundation models. By demonstrating that LLMs can serve as effective continual learners on graph-structured data when appropriately prompted, the study opens a new research direction for integrating graph reasoning with large-scale language models.\n\n3. Comprehensive and Insightful Experimental Analysis：\nThe paper provides a detailed and balanced experimental study, including comparisons among GNNs, LLMs, and GLMs across diverse datasets and settings. The analysis offers clear empirical insights—such as why GLMs may underperform and how prototype-based designs and model scaling influence continual learning performance—making the work informative beyond its own method.", "weaknesses": "1. Limited Theoretical Justification for SimGCL:\nWhile SimGCL shows strong empirical results, the paper provides little theoretical or analytical grounding for why the combination of graph prompts, LoRA fine-tuning, and prototype-based classification alleviates forgetting. The approach appears largely empirical, and the mechanism behind its robustness (e.g., whether prototype stability or prompt alignment is the key factor) is not formally analyzed. Adding theoretical reasoning or ablation-based evidence would strengthen the scientific depth of the contribution.\n\n2. Overstatement of SimGCL’s generality and simplicity:\nThe paper claims that SimGCL “greatly alleviates catastrophic forgetting” and “surpasses all existing baselines” under a rehearsal-free constraint. While the reported results do show strong gains (up to 20%), these improvements are dataset-dependent — performance drops significantly on sparse or long-session datasets (e.g., Arxiv-23, FSNCIL). Hence, calling it universally effective may be an overstatement; the method is empirically strong but not universally superior.\n\n3. Assertion of “efficiency” and “low cost”:\nThe authors emphasize SimGCL’s efficiency due to LoRA-based tuning and prototype inference. However, they provide no runtime, memory, or scaling benchmarks to substantiate this claim. Given that LLM fine-tuning is resource-intensive, this efficiency claim seems qualitative and not empirically validated, representing a mild overstatement.", "questions": "Q1. Clarification on the mechanism of forgetting alleviation\nYou claim that SimGCL “greatly alleviates catastrophic forgetting” through training-free prototype classification. Could you provide concrete evidence (e.g., forgetting curves, representation drift analysis, or embedding similarity metrics) showing that forgetting is truly reduced, rather than simply avoided by freezing parameters?\n\nQ2. Computational efficiency validation\nThe paper repeatedly highlights SimGCL’s “efficiency,” but no runtime or GPU memory statistics are presented. Could you quantify the actual training and inference cost (in FLOPs, GPU hours, or wall-clock time) compared to other LLM-based and GNN-based baselines?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761984945406}, {"id": "Jt6m0PfM5S", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1449/Reviewer_Sw4z"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes an extension of the Online Graph Learning problem formulation, aiming to address some of the limitations typically found in existing literature. Based on this revised formulation, the authors define two scenarios: a standard setting and a few-shot setting. They evaluate both conventional graph-based approaches and LLM-based methods. Finally, the paper introduces SimCML, a model that leverages a single round of instruction tuning to directly produce class prototypes from the embeddings generated by its fine-tuned LLM.", "review_text": "The paper proposes an extension of the Online Graph Learning problem formulation, aiming to address some of the limitations typically found in existing literature. Based on this revised formulation, the authors define two scenarios: a standard setting and a few-shot setting. They evaluate both conventional graph-based approaches and LLM-based methods. Finally, the paper introduces SimCML, a model that leverages a single round of instruction tuning to directly produce class prototypes from the embeddings generated by its fine-tuned LLM.", "strengths": "The paper explores a novel approach that connects LLMs with the graph continual learning framework. Moreover, it analyzes the commonly used GCL scenario, revealing an interesting limitation and proposing a solution to address this issue.", "weaknesses": "The paper presents several issues, mainly related to the clarity of presentation and the design of the experimental framework.\nRegarding the presentation, the paper does not clearly explain the proposed model or the models used for comparison. As a result, many methodological choices are not properly justified. Going into more detail, in the introduction, the authors claim that there is a lack of investigation into the rationality of the common experimental setup used for GCL. It is worth noting that a recent study has already explored this aspect [1]. Another meaningful reference that should be considered and discussed is [6].\nIn the background section, the authors correctly highlight one of the main limitations of the commonly adopted GCL setting, which I agree with, and propose the idea of global testing. However, it is not clear how this proposal relates to the subsequent discussion about inter-task edges, which the authors claim not to consider in their tasks. This choice isolates subgraphs, making the problem closer to the original formulation. This point should be discussed and clarified in greater depth. Furthermore, the removal of the limited-representation classes appears to be a strong design choice that requires a more thorough discussion, particularly regarding its advantages and potential drawbacks.\nIn Section 3.2, the authors claim that the considered datasets cover a wide range of possible sizes. However, in practice, the investigation is limited to relatively small datasets. Large-dimensional datasets, such as OGN-Products and Reddit [2], are not considered. Moreover, most of the selected datasets are citation networks. Considering that the goal of the paper is to establish a benchmark for testing models in GCL, the dataset selection appears rather limited. Another important distinction that can significantly impact the GCL scenario is the difference between heterophilous and homophilous datasets, which is not addressed in this work.\nIn the paper, both the definitions of the literature methods and that of the proposed SimCML are not presented clearly, which makes it difficult to follow the flow of the discussion and to evaluate the novelty of the proposed approach. For instance, the discussion of the LLM baseline is very limited, the authors only state that it is nascent and that no methods have yet been specifically designed for CGL tasks. What remains unclear is how the LLM baselines are actually applied in the context of CGL, and consequently how the reported results are obtained. Regarding the models considered, it is surprising that the authors do not include several of the most promising replay-based approaches, such as CaT [3], PDGNN [4], SSM [5], among others. Another major issue is the lack of explanation regarding the experimental setting and the validation policy. While the appendix lists the hyperparameters considered, many of them are fixed, and it is not clear how these values were selected. Considering the aim of the paper, this represents a significant limitation. Moreover, the results reported in the tables do not include any measures of variance or standard deviation, raising concerns about whether the experiments were run only once. If this is the case, it is unclear how the authors assess the stability of the results or the statistical significance of the differences between models. This is particularly important for the LLM application, where variance in results is typically more pronounced.\n\n\n[1] Donghi, G., Pasa, L., Zambon, D., Alippi, C. and Navarin, N., 2025. Online Continual Graph Learning. arXiv preprint arXiv:2508.03283.\n\n[2] Zhang, X., Song, D. and Tao, D., 2024. Continual learning on graphs: Challenges, solutions, and opportunities. arXiv preprint arXiv:2402.11565.\n\n[3] Liu, Y., Qiu, R. and Huang, Z., 2023, December. Cat: Balanced continual graph learning with graph condensation. In 2023 IEEE International Conference on Data Mining (ICDM) (pp. 1157-1162). IEEE.\n\n[4] Zang, X., Song, D., Chen, Y. and Tao, D., 2024, August. Topology-aware embedding memory for continual learning on expanding networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4326-4337).\n\n[5] Zhang, X., Song, D. and Tao, D., 2022, November. Sparsified subgraph memory for continual graph representation learning. In 2022 IEEE International Conference on Data Mining (ICDM) (pp. 1335-1340). IEEE.\n\n[6] Huang, S., Parviz, A., Kondrup, E., Yang, Z., Ding, Z., Bronstein, M., Rabbany, R. and Rabusseau, G., 2025. Are Large Language Models Good Temporal Graph Learners?. arXiv preprint arXiv:2506.05393.", "questions": "-Could the authors clarify and provide more details on the proposed SimCML model and the LLM baselines, particularly how these baselines are applied in the context of GCL?\n\n-Could the authors provide more information on how the hyperparameters were selected and the validation protocol used? Additionally, were the experiments repeated to assess the stability and statistical significance of the results?\n\n-Considering the goal of establishing a benchmark for GCL, could the authors comment on the choice of datasets and whether including large-dimensional datasets or heterophilous/homophilous distinctions might affect the evaluation?\n\n-Could the authors discuss the rationale for not including certain replay-based approaches, such as CaT, PDGNN, and SSM, and how their inclusion might impact the evaluation of SimCML?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an extension of the Online Graph Learning problem formulation, aiming to address some of the limitations typically found in existing literature. Based on this revised formulation, the authors define two scenarios: a standard setting and a few-shot setting. They evaluate both conventional graph-based approaches and LLM-based methods. Finally, the paper introduces SimCML, a model that leverages a single round of instruction tuning to directly produce class prototypes from the embeddings generated by its fine-tuned LLM.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper explores a novel approach that connects LLMs with the graph continual learning framework. Moreover, it analyzes the commonly used GCL scenario, revealing an interesting limitation and proposing a solution to address this issue.", "weaknesses": "The paper presents several issues, mainly related to the clarity of presentation and the design of the experimental framework.\nRegarding the presentation, the paper does not clearly explain the proposed model or the models used for comparison. As a result, many methodological choices are not properly justified. Going into more detail, in the introduction, the authors claim that there is a lack of investigation into the rationality of the common experimental setup used for GCL. It is worth noting that a recent study has already explored this aspect [1]. Another meaningful reference that should be considered and discussed is [6].\nIn the background section, the authors correctly highlight one of the main limitations of the commonly adopted GCL setting, which I agree with, and propose the idea of global testing. However, it is not clear how this proposal relates to the subsequent discussion about inter-task edges, which the authors claim not to consider in their tasks. This choice isolates subgraphs, making the problem closer to the original formulation. This point should be discussed and clarified in greater depth. Furthermore, the removal of the limited-representation classes appears to be a strong design choice that requires a more thorough discussion, particularly regarding its advantages and potential drawbacks.\nIn Section 3.2, the authors claim that the considered datasets cover a wide range of possible sizes. However, in practice, the investigation is limited to relatively small datasets. Large-dimensional datasets, such as OGN-Products and Reddit [2], are not considered. Moreover, most of the selected datasets are citation networks. Considering that the goal of the paper is to establish a benchmark for testing models in GCL, the dataset selection appears rather limited. Another important distinction that can significantly impact the GCL scenario is the difference between heterophilous and homophilous datasets, which is not addressed in this work.\nIn the paper, both the definitions of the literature methods and that of the proposed SimCML are not presented clearly, which makes it difficult to follow the flow of the discussion and to evaluate the novelty of the proposed approach. For instance, the discussion of the LLM baseline is very limited, the authors only state that it is nascent and that no methods have yet been specifically designed for CGL tasks. What remains unclear is how the LLM baselines are actually applied in the context of CGL, and consequently how the reported results are obtained. Regarding the models considered, it is surprising that the authors do not include several of the most promising replay-based approaches, such as CaT [3], PDGNN [4], SSM [5], among others. Another major issue is the lack of explanation regarding the experimental setting and the validation policy. While the appendix lists the hyperparameters considered, many of them are fixed, and it is not clear how these values were selected. Considering the aim of the paper, this represents a significant limitation. Moreover, the results reported in the tables do not include any measures of variance or standard deviation, raising concerns about whether the experiments were run only once. If this is the case, it is unclear how the authors assess the stability of the results or the statistical significance of the differences between models. This is particularly important for the LLM application, where variance in results is typically more pronounced.\n\n\n[1] Donghi, G., Pasa, L., Zambon, D., Alippi, C. and Navarin, N., 2025. Online Continual Graph Learning. arXiv preprint arXiv:2508.03283.\n\n[2] Zhang, X., Song, D. and Tao, D., 2024. Continual learning on graphs: Challenges, solutions, and opportunities. arXiv preprint arXiv:2402.11565.\n\n[3] Liu, Y., Qiu, R. and Huang, Z., 2023, December. Cat: Balanced continual graph learning with graph condensation. In 2023 IEEE International Conference on Data Mining (ICDM) (pp. 1157-1162). IEEE.\n\n[4] Zang, X., Song, D., Chen, Y. and Tao, D., 2024, August. Topology-aware embedding memory for continual learning on expanding networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4326-4337).\n\n[5] Zhang, X., Song, D. and Tao, D., 2022, November. Sparsified subgraph memory for continual graph representation learning. In 2022 IEEE International Conference on Data Mining (ICDM) (pp. 1335-1340). IEEE.\n\n[6] Huang, S., Parviz, A., Kondrup, E., Yang, Z., Ding, Z., Bronstein, M., Rabbany, R. and Rabusseau, G., 2025. Are Large Language Models Good Temporal Graph Learners?. arXiv preprint arXiv:2506.05393.", "questions": "-Could the authors clarify and provide more details on the proposed SimCML model and the LLM baselines, particularly how these baselines are applied in the context of GCL?\n\n-Could the authors provide more information on how the hyperparameters were selected and the validation protocol used? Additionally, were the experiments repeated to assess the stability and statistical significance of the results?\n\n-Considering the goal of establishing a benchmark for GCL, could the authors comment on the choice of datasets and whether including large-dimensional datasets or heterophilous/homophilous distinctions might affect the evaluation?\n\n-Could the authors discuss the rationale for not including certain replay-based approaches, such as CaT, PDGNN, and SSM, and how their inclusion might impact the evaluation of SimCML?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761761212385}, {"id": "WeCnigr8U7", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1449/Reviewer_PWgw"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper investigates whether large language models can alleviate catastrophic forgetting in graph continual learning (GCL). The authors identify flaws in existing GCL evaluation protocols—particularly task ID leakage—and propose a revised global testing setup. They introduce LLM4GCL, a new benchmark encompassing multiple GNN-, LLM-, and GLM-based methods across seven text-attributed graph datasets, and propose SimGCL that combines graph-structured prompts with prototype-based prediction. Experimental results indicate that SimGCL performs competitively under various GCL settings, suggesting that pretrained LLMs have potential for continual graph learning.", "review_text": "This paper investigates whether large language models can alleviate catastrophic forgetting in graph continual learning (GCL). The authors identify flaws in existing GCL evaluation protocols—particularly task ID leakage—and propose a revised global testing setup. They introduce LLM4GCL, a new benchmark encompassing multiple GNN-, LLM-, and GLM-based methods across seven text-attributed graph datasets, and propose SimGCL that combines graph-structured prompts with prototype-based prediction. Experimental results indicate that SimGCL performs competitively under various GCL settings, suggesting that pretrained LLMs have potential for continual graph learning.", "strengths": "1. The paper tackles an underexplored question—whether large language models (LLMs) can mitigate catastrophic forgetting in graph continual learning.\n2. A benchmark, LLM4GCL, is developed.", "weaknesses": "1. My biggest concern lies in the motivation. The primary goal of continual learning is to enable efficient adaptation under limited resources, whereas LLMs are inherently computationally expensive. It is unclear whether the significant computational cost introduced by LLMs can truly justify the efficiency-driven learning protocol that continual learning aims to achieve.\n\n2. Following this point, I believe the comparison should include GCL methods that allocate additional parameters (expansion-based) or memory (experience-replay-based) to narrow the substantial computational gap between LLMs and traditional GNNs. However, the authors compare their method only with regularization-based approaches, which are the most resource-limited class of GCL methods. Furthermore, the comparisons are arguably unfair since the methods rely on different backbone architectures for graph prediction.\n\n3. The authors acknowledge task ID leakage and address it by introducing a global testing protocol, which is commendable. Nonetheless, all the datasets used (e.g., Cora, Citeseer, WikiCS) are widely used text-attributed benchmarks that may overlap with the pretraining corpora of LLMs. The paper does not verify whether such overlap exists, leaving potential knowledge leakage from pretraining unaddressed.\n\n4. I also find the fairness of the LLM4GCL benchmark setup questionable. The LLM baselines are not fine-tuned on the same datasets and therefore lack domain knowledge of LLM4GCL, making the comparison asymmetric. The proposed method appears less focused on mitigating catastrophic forgetting and more on task-specific instruction tuning, which may even result in overfitting to the given dataset. This is evidenced by the large performance gap between the proposed method and other LLM-based baselines on LLM4GCL, in contrast to the much smaller gap observed on standard public datasets. Overall, the comparisons and evaluations in this paper feel ambiguous, making it difficult to discern the central message or takeaway the authors intend to convey.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates whether large language models can alleviate catastrophic forgetting in graph continual learning (GCL). The authors identify flaws in existing GCL evaluation protocols—particularly task ID leakage—and propose a revised global testing setup. They introduce LLM4GCL, a new benchmark encompassing multiple GNN-, LLM-, and GLM-based methods across seven text-attributed graph datasets, and propose SimGCL that combines graph-structured prompts with prototype-based prediction. Experimental results indicate that SimGCL performs competitively under various GCL settings, suggesting that pretrained LLMs have potential for continual graph learning.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper tackles an underexplored question—whether large language models (LLMs) can mitigate catastrophic forgetting in graph continual learning.\n2. A benchmark, LLM4GCL, is developed.", "weaknesses": "1. My biggest concern lies in the motivation. The primary goal of continual learning is to enable efficient adaptation under limited resources, whereas LLMs are inherently computationally expensive. It is unclear whether the significant computational cost introduced by LLMs can truly justify the efficiency-driven learning protocol that continual learning aims to achieve.\n\n2. Following this point, I believe the comparison should include GCL methods that allocate additional parameters (expansion-based) or memory (experience-replay-based) to narrow the substantial computational gap between LLMs and traditional GNNs. However, the authors compare their method only with regularization-based approaches, which are the most resource-limited class of GCL methods. Furthermore, the comparisons are arguably unfair since the methods rely on different backbone architectures for graph prediction.\n\n3. The authors acknowledge task ID leakage and address it by introducing a global testing protocol, which is commendable. Nonetheless, all the datasets used (e.g., Cora, Citeseer, WikiCS) are widely used text-attributed benchmarks that may overlap with the pretraining corpora of LLMs. The paper does not verify whether such overlap exists, leaving potential knowledge leakage from pretraining unaddressed.\n\n4. I also find the fairness of the LLM4GCL benchmark setup questionable. The LLM baselines are not fine-tuned on the same datasets and therefore lack domain knowledge of LLM4GCL, making the comparison asymmetric. The proposed method appears less focused on mitigating catastrophic forgetting and more on task-specific instruction tuning, which may even result in overfitting to the given dataset. This is evidenced by the large performance gap between the proposed method and other LLM-based baselines on LLM4GCL, in contrast to the much smaller gap observed on standard public datasets. Overall, the comparisons and evaluations in this paper feel ambiguous, making it difficult to discern the central message or takeaway the authors intend to convey.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761700703421}, {"id": "tys4mYy3Gm", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1449/Reviewer_FJeP"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper focuses on the problem of graph continual learning and proposes a method based on LLMs. Different from the previous settings, the authors provide a new setting called global testing and construct a benchmark. Based on the benchmark, a new method based on LLMs and prototype learning is proposed.", "review_text": "This paper focuses on the problem of graph continual learning and proposes a method based on LLMs. Different from the previous settings, the authors provide a new setting called global testing and construct a benchmark. Based on the benchmark, a new method based on LLMs and prototype learning is proposed.", "strengths": "1. The paper is well-written and easy to follow. The code is released.\n\n2. A new benchmark is proposed to evaluate the performance of LLMs and graph-enhanced LLMs.\n\n3. The authors identify the flaw of existing GCL settings and propose global testing. To utilize the generalization capacity of LLMs, a GLM-based approach is proposed.", "weaknesses": "1. The constraint of setting each task to have a unified sample size is unreasonable, and the label imbalance problem is the real challenge that should be addressed.\n\n2. The introduction of global testing is not very clear. More clear figures or visualizations are needed.\n\n3. The used graphs are all text-attribute graphs (TAGs). Can the proposed method be extended to non-TAGs?\n\n4. Experimental details should be included in the main paper to provide readers with a general picture of the experimental settings.\n\n5. There are some typos, and the authors should carefully revise the paper.", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the problem of graph continual learning and proposes a method based on LLMs. Different from the previous settings, the authors provide a new setting called global testing and construct a benchmark. Based on the benchmark, a new method based on LLMs and prototype learning is proposed.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well-written and easy to follow. The code is released.\n\n2. A new benchmark is proposed to evaluate the performance of LLMs and graph-enhanced LLMs.\n\n3. The authors identify the flaw of existing GCL settings and propose global testing. To utilize the generalization capacity of LLMs, a GLM-based approach is proposed.", "weaknesses": "1. The constraint of setting each task to have a unified sample size is unreasonable, and the label imbalance problem is the real challenge that should be addressed.\n\n2. The introduction of global testing is not very clear. More clear figures or visualizations are needed.\n\n3. The used graphs are all text-attribute graphs (TAGs). Can the proposed method be extended to non-TAGs?\n\n4. Experimental details should be included in the main paper to provide readers with a general picture of the experimental settings.\n\n5. There are some typos, and the authors should carefully revise the paper.", "questions": "Please see the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761037997022}, {"id": "AVeqvEINC6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1449/Reviewer_FyMh"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper investigates whether large language models (LLMs) can help overcome catastrophic forgetting in Graph Continual Learning (GCL) tasks. \n\n- They have identified a serious flaw, task ID leakage, in common GCL evaluations, where models inadvertently access task identifiers during testing.\n\n- They introduce LLM4GCL, the first systematic benchmark for LLMs in GCL.\n\n- Proposed a method, SimGCL, that consistently achieves state-of-the-art results and mitigates forgetting effectively.", "review_text": "This paper investigates whether large language models (LLMs) can help overcome catastrophic forgetting in Graph Continual Learning (GCL) tasks. \n\n- They have identified a serious flaw, task ID leakage, in common GCL evaluations, where models inadvertently access task identifiers during testing.\n\n- They introduce LLM4GCL, the first systematic benchmark for LLMs in GCL.\n\n- Proposed a method, SimGCL, that consistently achieves state-of-the-art results and mitigates forgetting effectively.", "strengths": "### 1. Theoretical Soundness\n- The work is well motivated as the authors have identified that all prior GCL methods train from scratch and do not exploit pretrained models’ generalization.\n\n- Uses Parameter-Efficient Fine-Tuning (PEFT) via LoRA, a proven method for adapting large models efficiently.\n\n- Employs prototype-based classification, which is widely accepted as a way to mitigate forgetting in continual learning.\n\n- The graph prompt design is conceptually consistent with instruction tuning paradigms in multimodal LLMs.\n\n### 2. Experimental Soundness\n- The paper introduces LLM4GCL, with 9 LLM/GLM methods and 7 datasets. This is a substantial empirical foundation for a GCL.\n\n- The identification of task ID leakage in existing setups is a major contribution.\n\n- They release code, datasets, and standardized evaluation. This is a strong signal of methodological transparency.", "weaknesses": "### 1. Theoretical Limitations\n- They assume graph structure can be effectively encoded as text prompts. While innovative, this verbalization step might oversimplify structural relationships, potentially losing fine-grained topology information.\n\n- Use of LoRA only in the first session assumes that continual adaptation can be achieved purely through frozen embeddings and prototype updates — a strong assumption that may not generalize across highly dynamic graph evolutions.\n\n- Missing formal theoretical analysis of why the prototype mechanism preserves generalization in LLM-embedded graph representations (e.g., no bounds on forgetting or feature drift).\n\n### 1. Experimental Limitations\n- The authors do not deeply dissect which component of SimGCL contributes most (LoRA, prototype, or prompt design). Without that, causal claims (“graph prompts mitigate forgetting”) remain partially speculative.\n\n- In small-scale datasets like Cora and Citeseer, very large LLMs might overfit to textual node attributes, inflating reported improvements.", "questions": "Please refer to the weakness section. I would urge the authors to pay special attention to the following concerns first.\n\n- The work is conceptually sound and empirically motivated but not theoretically grounded. The rationale aligns with existing literature, but lacks formal guarantees. Can the authors provide some theoretical justifications to their method?\n\n- Empirical claims are mostly well-supported, though some causal explanations (e.g., why GLMs fail) remain conjectural. Therefore, I request the authors to substantiate such claims.\n\n- Address limited ablation analyses and potential structural oversimplifications when converting graph topology into text prompts.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates whether large language models (LLMs) can help overcome catastrophic forgetting in Graph Continual Learning (GCL) tasks. \n\n- They have identified a serious flaw, task ID leakage, in common GCL evaluations, where models inadvertently access task identifiers during testing.\n\n- They introduce LLM4GCL, the first systematic benchmark for LLMs in GCL.\n\n- Proposed a method, SimGCL, that consistently achieves state-of-the-art results and mitigates forgetting effectively.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "### 1. Theoretical Soundness\n- The work is well motivated as the authors have identified that all prior GCL methods train from scratch and do not exploit pretrained models’ generalization.\n\n- Uses Parameter-Efficient Fine-Tuning (PEFT) via LoRA, a proven method for adapting large models efficiently.\n\n- Employs prototype-based classification, which is widely accepted as a way to mitigate forgetting in continual learning.\n\n- The graph prompt design is conceptually consistent with instruction tuning paradigms in multimodal LLMs.\n\n### 2. Experimental Soundness\n- The paper introduces LLM4GCL, with 9 LLM/GLM methods and 7 datasets. This is a substantial empirical foundation for a GCL.\n\n- The identification of task ID leakage in existing setups is a major contribution.\n\n- They release code, datasets, and standardized evaluation. This is a strong signal of methodological transparency.", "weaknesses": "### 1. Theoretical Limitations\n- They assume graph structure can be effectively encoded as text prompts. While innovative, this verbalization step might oversimplify structural relationships, potentially losing fine-grained topology information.\n\n- Use of LoRA only in the first session assumes that continual adaptation can be achieved purely through frozen embeddings and prototype updates — a strong assumption that may not generalize across highly dynamic graph evolutions.\n\n- Missing formal theoretical analysis of why the prototype mechanism preserves generalization in LLM-embedded graph representations (e.g., no bounds on forgetting or feature drift).\n\n### 1. Experimental Limitations\n- The authors do not deeply dissect which component of SimGCL contributes most (LoRA, prototype, or prompt design). Without that, causal claims (“graph prompts mitigate forgetting”) remain partially speculative.\n\n- In small-scale datasets like Cora and Citeseer, very large LLMs might overfit to textual node attributes, inflating reported improvements.", "questions": "Please refer to the weakness section. I would urge the authors to pay special attention to the following concerns first.\n\n- The work is conceptually sound and empirically motivated but not theoretically grounded. The rationale aligns with existing literature, but lacks formal guarantees. Can the authors provide some theoretical justifications to their method?\n\n- Empirical claims are mostly well-supported, though some causal explanations (e.g., why GLMs fail) remain conjectural. Therefore, I request the authors to substantiate such claims.\n\n- Address limited ablation analyses and potential structural oversimplifications when converting graph topology into text prompts.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760764638305}], "openreview_url": "https://openreview.net/forum?id=ZKkeA1G935", "arxiv_id": "2505.18697", "paper_pdf": "papers/ZKkeA1G935.pdf", "paper_pdf_sha256": "8670c1708b38039295810b101422fea8dd543baa06a0e020499b9fa9f53cf406", "paper_pdf_bytes": 2437076, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhixunLEE/LLM4GCL", "code_repository": "ZhixunLEE/LLM4GCL", "code_commit": "e4b286560287289a184b20c34153af2716c89cec", "code_archive": "repos/ZKkeA1G935.zip", "code_archive_sha256": "d44e3251a95490b92fca6427690a532a10e8dae8dc552ae511fde8e510dcc750", "code_archive_bytes": 101569, "code_file_count": 43, "code_extensions": {".py": 37, ".sh": 6}, "github_disk_usage_kb": 1636, "github_languages": {"Python": 307475, "Shell": 6098}, "github_archived": false, "github_pushed_at": "2025-09-25T15:01:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/can-llms-alleviate-catastrophic-forgetting-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kWtP5ZOErR", "year": 2025, "status": "rejected", "title": "EvoPress: Towards Optimal Dynamic Model Compression via Evolutionary Search", "authors": ["Oliver Sieberling", "Denis Kuznedelev", "Eldar Kurtic", "Dan Alistarh"], "authorids": ["~Oliver_Sieberling1", "~Denis_Kuznedelev1", "~Eldar_Kurtic1", "~Dan_Alistarh7"], "authors_source": "OpenReview API", "abstract": "The high computational costs of large language models (LLMs) have led to\na flurry of research on LLM compression, via methods such as quantization,\nsparsification, or structured pruning. A new frontier in this area is given by\ndynamic, non-uniform compression methods, which adjust the compression\nlevels (e.g., sparsity) per-block or even per-layer in order to minimize accuracy\nloss, while guaranteeing a global compression threshold. Yet, current methods\nrely on heuristics for identifying the “importance” of a given layer towards the\nloss, based on assumptions such as error monotonicity, i.e. that the end-to-end\nmodel compression error is proportional to the sum of layer-wise errors. In this\npaper, we revisit this area, and propose a new and general approach for dynamic\ncompression that is provably optimal in a given input range. We begin from\nthe motivating observation that, in general, error monotonicity does not hold for\nLLMs: compressed models with lower sum of per-layer errors can perform worse\nthan models with higher error sums. To address this, we propose a new general\nevolutionary framework for dynamic LLM compression called EvoPress, which\nhas provable convergence, low sample and evaluation complexity. We show that\nthese theoretical guarantees lead to highly competitive practical performance\nfor dynamic compression of Llama, Mistral and Phi models: via EvoPress, we\nset new state-of-the-art results for structural pruning (block/layer dropping),\nunstructured sparsity, as well as quantization with dynamic bitwidths.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "IdK7QnnfIh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13864/Reviewer_yr9o"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper introduces a new general framework for dynamic LLM compression. Building on the observation that error monotonicity does not generally hold for LLM compression, the paper proposed evolutionary search approach. The proposed approach is provably convergent and has low sample and evaluation complexity when the fitness function is linear. Experiments show that the proposed methods improve dynamic compression of Llama, Mistral, and Phi models.", "review_text": "This paper introduces a new general framework for dynamic LLM compression. Building on the observation that error monotonicity does not generally hold for LLM compression, the paper proposed evolutionary search approach. The proposed approach is provably convergent and has low sample and evaluation complexity when the fitness function is linear. Experiments show that the proposed methods improve dynamic compression of Llama, Mistral, and Phi models.", "strengths": "The proposed method shows strong empirical performance. It is flexible, as it can be applied to unstructured pruning, structured pruning/layer dropping, and quantization. Theoretical analysis of the proposed methods is also provided,", "weaknesses": "1. I think it would be better if the average and standard deviation of the performance metrics across different runs were reported, as it can demonstrate the significance of the improvement. But I understand it could be expensive to do and might not be possible for baseline methods.\n\n\n2. The theoretical results are somewhat disconnected from other parts. I think having the theoretical results is a bonus and it can be of interest beyond the LLM context. But some exploration of the function form of $f$ in real experiments or some simple simulation example might better demonstrate the significance of the results", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new general framework for dynamic LLM compression. Building on the observation that error monotonicity does not generally hold for LLM compression, the paper proposed evolutionary search approach. The proposed approach is provably convergent and has low sample and evaluation complexity when the fitness function is linear. Experiments show that the proposed methods improve dynamic compression of Llama, Mistral, and Phi models.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The proposed method shows strong empirical performance. It is flexible, as it can be applied to unstructured pruning, structured pruning/layer dropping, and quantization. Theoretical analysis of the proposed methods is also provided,", "weaknesses": "1. I think it would be better if the average and standard deviation of the performance metrics across different runs were reported, as it can demonstrate the significance of the improvement. But I understand it could be expensive to do and might not be possible for baseline methods.\n\n\n2. The theoretical results are somewhat disconnected from other parts. I think having the theoretical results is a bonus and it can be of interest beyond the LLM context. But some exploration of the function form of $f$ in real experiments or some simple simulation example might better demonstrate the significance of the results", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730709318434}, {"id": "CdMeLKw6Qy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13864/Reviewer_jZST"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "The paper introduces a meta-heuristic for non-uniform model compression that is largely independent of specific model architectures and compression methods. The method is demonstrated on three compression approaches—depth pruning, unstructured sparsity, and quantization—with ample numerical results and a supporting analytical convergence analysis.", "review_text": "The paper introduces a meta-heuristic for non-uniform model compression that is largely independent of specific model architectures and compression methods. The method is demonstrated on three compression approaches—depth pruning, unstructured sparsity, and quantization—with ample numerical results and a supporting analytical convergence analysis.", "strengths": "The strength of the approach is its simplicity and the fact that it is somewhat agnostic to model architecture and compression techniques.\nWhile it doesn’t introduce entirely novel algorithms, its design effectively minimizes computational costs when working with larger models. Overall, the paper is quite well-written and presents compelling arguments for the proposed approach.", "weaknesses": "The manuscript does not, however, discuss much about the scalability of the approach in more complex scenarios, and to what extent the evolutionary search is useful compared to more basic derivative-free optimization techniques or the like.\nWhile presented as evolutionary, the algorithm leans heavily on an elitist, exploitation-focused strategy. This approach could limit exploration, potentially trapping the algorithm in local minima and risk of poor generalization in more complex compression scenarios.\n\nThis approach appears to be effective as long as the manifold remains relatively convex. However, with multi-modal compression techniques or non-convex landscapes, there are concerns about whether this computationally efficient method will generalize well beyond the controlled benchmarking environments, especially when handling mixed or complex compression strategies.", "questions": "1) The scheme is presented as an evolutionary search, but the update mechanism uses an elitist strategy (focused on exploitation) based on single-offspring selection. The number of mutations is also kept minimal, with one mutation shown to be nearly optimal, as demonstrated in SI B. In this regime, the scheme resembles a form of random coordinate descent or derivative-free optimization, with the added twist that perturbation occurs on a manifold with constant overall compression (referred to as switching). Could you clarify which properties of an evolutionary algorithm are essential to the results achieved, as opposed to a simpler perturbative approach?\n\n2) In the current setting, how much variability is there in the optimal compression profile found after one run of the evolutionary search? Does it appear the search get stuck in local minima in any of the benchmarks? \n\n3) Can you predict if the hypothesis that smooth dynamic model compression results in a smooth fitness landscape with few local optima still hold in multi-modal cases? \n\nAdditional Questions/comments:\n\n4) How is the step size in a single dimension determined (e.g., the 1M weights for unstructured sparsity)? Is the step size fixed or adaptive?\n\n5) In the selection process, only two stages are shown (see SI B.2). Was this approach tested with more stages, and if so, were additional stages found to be ineffective?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a meta-heuristic for non-uniform model compression that is largely independent of specific model architectures and compression methods. The method is demonstrated on three compression approaches—depth pruning, unstructured sparsity, and quantization—with ample numerical results and a supporting analytical convergence analysis.", "soundness": 4, "presentation": 3, "contribution": 4, "strengths": "The strength of the approach is its simplicity and the fact that it is somewhat agnostic to model architecture and compression techniques.\nWhile it doesn’t introduce entirely novel algorithms, its design effectively minimizes computational costs when working with larger models. Overall, the paper is quite well-written and presents compelling arguments for the proposed approach.", "weaknesses": "The manuscript does not, however, discuss much about the scalability of the approach in more complex scenarios, and to what extent the evolutionary search is useful compared to more basic derivative-free optimization techniques or the like.\nWhile presented as evolutionary, the algorithm leans heavily on an elitist, exploitation-focused strategy. This approach could limit exploration, potentially trapping the algorithm in local minima and risk of poor generalization in more complex compression scenarios.\n\nThis approach appears to be effective as long as the manifold remains relatively convex. However, with multi-modal compression techniques or non-convex landscapes, there are concerns about whether this computationally efficient method will generalize well beyond the controlled benchmarking environments, especially when handling mixed or complex compression strategies.", "questions": "1) The scheme is presented as an evolutionary search, but the update mechanism uses an elitist strategy (focused on exploitation) based on single-offspring selection. The number of mutations is also kept minimal, with one mutation shown to be nearly optimal, as demonstrated in SI B. In this regime, the scheme resembles a form of random coordinate descent or derivative-free optimization, with the added twist that perturbation occurs on a manifold with constant overall compression (referred to as switching). Could you clarify which properties of an evolutionary algorithm are essential to the results achieved, as opposed to a simpler perturbative approach?\n\n2) In the current setting, how much variability is there in the optimal compression profile found after one run of the evolutionary search? Does it appear the search get stuck in local minima in any of the benchmarks? \n\n3) Can you predict if the hypothesis that smooth dynamic model compression results in a smooth fitness landscape with few local optima still hold in multi-modal cases? \n\nAdditional Questions/comments:\n\n4) How is the step size in a single dimension determined (e.g., the 1M weights for unstructured sparsity)? Is the step size fixed or adaptive?\n\n5) In the selection process, only two stages are shown (see SI B.2). Was this approach tested with more stages, and if so, were additional stages found to be ineffective?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730679554438}, {"id": "rZFj3KY8IT", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13864/Reviewer_wUQk"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This work proposes an algorithm to perform dynamic compression (different layers can be compressed to various levels) of large language models. Previous works assign importance to each layer and assume end-to-end model compression error is proportional to the sum of layer-wise errors. This work observes that error monotonicity does not hold for LLM, and proposes an evolutionary search algorithm to achieve better compression quality.\n\nIn EvoPress, multiple heuristics, including mutation operation and the selection step are proposed to achieve efficient sampling for LLM. Experimental results show better compression quality compared to previous work on multiple applications.", "review_text": "This work proposes an algorithm to perform dynamic compression (different layers can be compressed to various levels) of large language models. Previous works assign importance to each layer and assume end-to-end model compression error is proportional to the sum of layer-wise errors. This work observes that error monotonicity does not hold for LLM, and proposes an evolutionary search algorithm to achieve better compression quality.\n\nIn EvoPress, multiple heuristics, including mutation operation and the selection step are proposed to achieve efficient sampling for LLM. Experimental results show better compression quality compared to previous work on multiple applications.", "strengths": "LLM compression is an important topic. This work proposes a new optimization algorithm, and better performance has been presented on multiple applications, including pruning, introducing sparsity, and quantization.", "weaknesses": "1. It would be good to introduce an algorithm overview figure to illustrate the steps in EvoPress. In addition, the algorithm contains mulitple heuristics to make the sampling efficient. It would be good to also summarize these heuristics in the figure.\n2. It's not clear whether the hyperparameters used in the heuristics (mutation operation & multi-step selection) can transfer well across tasks/datasets/models. Further illustration would be very useful.", "questions": "Please see above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes an algorithm to perform dynamic compression (different layers can be compressed to various levels) of large language models. Previous works assign importance to each layer and assume end-to-end model compression error is proportional to the sum of layer-wise errors. This work observes that error monotonicity does not hold for LLM, and proposes an evolutionary search algorithm to achieve better compression quality.\n\nIn EvoPress, multiple heuristics, including mutation operation and the selection step are proposed to achieve efficient sampling for LLM. Experimental results show better compression quality compared to previous work on multiple applications.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "LLM compression is an important topic. This work proposes a new optimization algorithm, and better performance has been presented on multiple applications, including pruning, introducing sparsity, and quantization.", "weaknesses": "1. It would be good to introduce an algorithm overview figure to illustrate the steps in EvoPress. In addition, the algorithm contains mulitple heuristics to make the sampling efficient. It would be good to also summarize these heuristics in the figure.\n2. It's not clear whether the hyperparameters used in the heuristics (mutation operation & multi-step selection) can transfer well across tasks/datasets/models. Further illustration would be very useful.", "questions": "Please see above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730664576422}, {"id": "ER40cv0QuG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13864/Reviewer_X4pg"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a new evolutionary framework for dynamic LLM compression.", "review_text": "This paper proposes a new evolutionary framework for dynamic LLM compression.", "strengths": "This paper proposes an evolutionary algorithm within an optimization framework, for sparse structure selection for LLMs.", "weaknesses": "1. The theoretical justification in Theorem 1 appears irrelevant to the algorithm proposed in the paper. Theorem 1 is established for linear fitness functions; however, it’s difficult to see how this applies, as LLMs likely do not have linear fitness functions.\n\n2.  The comparison lacks systematic rigor. For instance, Table 2 presents results only at the sparsity level 70%. What about results at other sparsity levels?", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new evolutionary framework for dynamic LLM compression.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper proposes an evolutionary algorithm within an optimization framework, for sparse structure selection for LLMs.", "weaknesses": "1. The theoretical justification in Theorem 1 appears irrelevant to the algorithm proposed in the paper. Theorem 1 is established for linear fitness functions; however, it’s difficult to see how this applies, as LLMs likely do not have linear fitness functions.\n\n2.  The comparison lacks systematic rigor. For instance, Table 2 presents results only at the sparsity level 70%. What about results at other sparsity levels?", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730661454727}], "openreview_url": "https://openreview.net/forum?id=kWtP5ZOErR", "arxiv_id": "2410.14649", "paper_pdf": "papers/kWtP5ZOErR.pdf", "paper_pdf_sha256": "81aa15da647f999fd8367754a2bb94839c10307f22754b3d6174fc66ac3430ae", "paper_pdf_bytes": 847447, "paper_pdf_source": "openreview", "code_url": "https://github.com/IST-DASLab/EvoPress", "code_repository": "IST-DASLab/EvoPress", "code_commit": "a66a6974e91d27ede9043f07f885709732a9a9b1", "code_archive": "repos/kWtP5ZOErR.zip", "code_archive_sha256": "88be341b7385e379e4a59869fc8a2c6eea60bc7b9d30455dae7974555b0a41c3", "code_archive_bytes": 122483, "code_file_count": 35, "code_extensions": {".py": 28, ".sh": 7}, "github_disk_usage_kb": 112, "github_languages": {"Python": 217370, "Shell": 8478}, "github_archived": false, "github_pushed_at": "2026-06-14T15:13:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evopress-towards-optimal-dynamic-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wE1I9IGqeH", "year": 2024, "status": "rejected", "title": "Continual Learning in Open-vocabulary Classification with Complementary Memory Systems", "authors": ["Zhen Zhu", "Weijie Lyu", "Yao Xiao", "Derek Hoiem"], "authorids": ["~Zhen_Zhu1", "~Weijie_Lyu2", "~Yao_Xiao3", "~Derek_Hoiem1"], "authors_source": "OpenReview API", "abstract": "We introduce a method for flexible and efficient continual learning in open-vocabulary image classification, drawing inspiration from the complementary learning systems observed in human cognition. Specifically, we propose to combine predictions from a CLIP zero-shot model and the exemplar-based model, using the zero-shot estimated probability that a sample's class is within the exemplar classes. We also propose a \"tree probe\" method, an adaption of lazy learning principles, which enables fast learning from new examples with competitive accuracy to batch-trained linear models. We test in data incremental, class incremental, and task incremental settings, as well as ability to perform flexible inference on varying subsets of zero-shot and learned categories. Our proposed method achieves a good balance of learning speed, target task effectiveness, and zero-shot effectiveness.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "pEK39RYDWp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1611/Reviewer_ShHL"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper tackles a new problem of continual learning in open-vocabulary classification where the model can update its knowledge based on incoming new samples while preserving zero-shot learning capability.\nTo achieve this, this works proposes to combine CLIP model, having zero-shot ability, with examplar-based learning, which stores additional training samples as exemplar for continual learning.\nFor exemplar-based learning, the author introduces a tree-probed algorithm which improves upon KNN to increase accuracy.\nThe paper provides an interesting analogy between CLIP & instance-based learning and fast & slow learning system in human development.\nTo combine between these two different learning paradigms, the proposed work designs a fusing prediction formula that averages the probability or embedding predictions of the two models. Finally, Adaptive Instance Marginalization module is trained to estimate the probability of a test sample belong to the exemplar set to further boost the performance.\nThe paper conducts experiments on CIFAR100, SUN397, FGVCAircraft, EuroSAT, OxfordIIITPets, StanfordCars, Food101 and Flowers102, ImageNet, UCF101, and DTD.", "review_text": "The paper tackles a new problem of continual learning in open-vocabulary classification where the model can update its knowledge based on incoming new samples while preserving zero-shot learning capability.\nTo achieve this, this works proposes to combine CLIP model, having zero-shot ability, with examplar-based learning, which stores additional training samples as exemplar for continual learning.\nFor exemplar-based learning, the author introduces a tree-probed algorithm which improves upon KNN to increase accuracy.\nThe paper provides an interesting analogy between CLIP & instance-based learning and fast & slow learning system in human development.\nTo combine between these two different learning paradigms, the proposed work designs a fusing prediction formula that averages the probability or embedding predictions of the two models. Finally, Adaptive Instance Marginalization module is trained to estimate the probability of a test sample belong to the exemplar set to further boost the performance.\nThe paper conducts experiments on CIFAR100, SUN397, FGVCAircraft, EuroSAT, OxfordIIITPets, StanfordCars, Food101 and Flowers102, ImageNet, UCF101, and DTD.", "strengths": "+ Using fast & slow learning system from human learning to describe the combination of zero-shot and exemplar-based learning is inspiring.\n+ The problem of continual learning in open-vocabulary classification is interesting and can have practical applications.\n+ The paper provides details experiments on multiple datasets.", "weaknesses": "+ The paper is hard to follow. Specifically the proposed setup is vaguely describe in the paper. For example, in the literature of open-vocabulary learning, there are only base and target classes [A,B]. However, the paper keeps mentioning about zero-shot and target tasks without clear explanation on what is zero-shot class (in open-vocabulary learning).\n\n+ The proposed idea is highly similar to continual zero-shot learning work where the goal is also to update model with new samples while maintaining the zero-shot performance [B]. It seems that the main difference is the use of CLIP encoder which boosts the model zero-shot capability but the core idea of continuously update the zero-shot model is similar. However, there is no discussion or comparison with these prior works.\n\n+ The reviewer also has doubted on the effectiveness of the proposed method as on average task performance the model only improve 0.5% compared the strong baseline ZSCL (as reported in table 1 in the main paper). Moreover, based on table 4, it appears that the proposed method doesn't perform well on fine-grained classification tasks of Flowers, Cars and EuroSat. Thus, the reviewer is not confident on whether the proposed method advances the continual open-vocabulary classification task. \n\n\n[A] Open-Vocabulary Instance Segmentation via Robust Cross-Modal Pseudo-Labeling, Huynh et al. CVPR22\n\n[B] CLASS NORMALIZATION FOR (CONTINUAL)? GENERALIZED ZERO-SHOT LEARNING, Skorokhodov et al. ICLR21", "questions": "+ Can the author verify the continual learning setup as well as the terminology (zero-shot tasks) clearly in the manuscript? This would significantly improves the paper readability.\n+ Sufficient discussion should be make between the proposed work and continual zero-shot learning literature.\n+ Can the author justify the modest improvement of 0.5% compared to SOTA?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tackles a new problem of continual learning in open-vocabulary classification where the model can update its knowledge based on incoming new samples while preserving zero-shot learning capability.\nTo achieve this, this works proposes to combine CLIP model, having zero-shot ability, with examplar-based learning, which stores additional training samples as exemplar for continual learning.\nFor exemplar-based learning, the author introduces a tree-probed algorithm which improves upon KNN to increase accuracy.\nThe paper provides an interesting analogy between CLIP & instance-based learning and fast & slow learning system in human development.\nTo combine between these two different learning paradigms, the proposed work designs a fusing prediction formula that averages the probability or embedding predictions of the two models. Finally, Adaptive Instance Marginalization module is trained to estimate the probability of a test sample belong to the exemplar set to further boost the performance.\nThe paper conducts experiments on CIFAR100, SUN397, FGVCAircraft, EuroSAT, OxfordIIITPets, StanfordCars, Food101 and Flowers102, ImageNet, UCF101, and DTD.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "+ Using fast & slow learning system from human learning to describe the combination of zero-shot and exemplar-based learning is inspiring.\n+ The problem of continual learning in open-vocabulary classification is interesting and can have practical applications.\n+ The paper provides details experiments on multiple datasets.", "weaknesses": "+ The paper is hard to follow. Specifically the proposed setup is vaguely describe in the paper. For example, in the literature of open-vocabulary learning, there are only base and target classes [A,B]. However, the paper keeps mentioning about zero-shot and target tasks without clear explanation on what is zero-shot class (in open-vocabulary learning).\n\n+ The proposed idea is highly similar to continual zero-shot learning work where the goal is also to update model with new samples while maintaining the zero-shot performance [B]. It seems that the main difference is the use of CLIP encoder which boosts the model zero-shot capability but the core idea of continuously update the zero-shot model is similar. However, there is no discussion or comparison with these prior works.\n\n+ The reviewer also has doubted on the effectiveness of the proposed method as on average task performance the model only improve 0.5% compared the strong baseline ZSCL (as reported in table 1 in the main paper). Moreover, based on table 4, it appears that the proposed method doesn't perform well on fine-grained classification tasks of Flowers, Cars and EuroSat. Thus, the reviewer is not confident on whether the proposed method advances the continual open-vocabulary classification task. \n\n\n[A] Open-Vocabulary Instance Segmentation via Robust Cross-Modal Pseudo-Labeling, Huynh et al. CVPR22\n\n[B] CLASS NORMALIZATION FOR (CONTINUAL)? GENERALIZED ZERO-SHOT LEARNING, Skorokhodov et al. ICLR21", "questions": "+ Can the author verify the continual learning setup as well as the terminology (zero-shot tasks) clearly in the manuscript? This would significantly improves the paper readability.\n+ Sufficient discussion should be make between the proposed work and continual zero-shot learning literature.\n+ Can the author justify the modest improvement of 0.5% compared to SOTA?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698955902582}, {"id": "jfJG37YK9b", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1611/Reviewer_aTSX"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The approach enables adaptable and efficient continual learning in open-vocabulary image classification. It draws inspiration from human cognition’s complementary learning systems. In this work, the author merges predictions from a CLIP zero-shot model and an exemplar-based model. it is based on using the zero-shot estimated probability that a sample’s class is within the exemplar classes. Inspired by lazy learning principles, the author introduces a “tree probe” method, which facilitating rapid learning from new examples with comparable accuracy to batch-trained linear models.", "review_text": "The approach enables adaptable and efficient continual learning in open-vocabulary image classification. It draws inspiration from human cognition’s complementary learning systems. In this work, the author merges predictions from a CLIP zero-shot model and an exemplar-based model. it is based on using the zero-shot estimated probability that a sample’s class is within the exemplar classes. Inspired by lazy learning principles, the author introduces a “tree probe” method, which facilitating rapid learning from new examples with comparable accuracy to batch-trained linear models.", "strengths": "[1] The problem is interesting, which predicts the open set vocabulary while following the setting of continual learning.\n\n[2] Results are evaluated over the various datasets in the diverse setting.\n\n[3] “Tree-probe” is interesting, which balance the rapid learning and performance.", "weaknesses": "[1] There are various setting in the continual learning (task incremental, class incremental, data incremental etc.) and zero-shot learning (generalized/non-generalised setting). The exact experimental setting, evaluation strategy, and the motivation of each setting are not clear. It’s difficult to follow the section 4.2, there should me better illustration and discrimination between the various evaluation scenarios.\n\n[2] There are few recent works [1,2] follow the similar setting. These works can be considered as baseline along with the CLIP zero-shot.\n\n[3] I believe that adding the problem setting before the section-3 (Method) with the proper notation will increase the readability. \n\n[4] The recent prompting based continual learning approach [3,4] leverages the strong pretrained model and shows the promising result for the continual learning without complementary memory system. Instead of examplar storage, if model leverages promoting based approach, how the model behaves?\n\nReference:\n\n[1]  Unseen Classes at a Later Time? No Problem, CVPR-22\n\n[2] Meta-Learned Attribute Self-Gating for Continual Generalized Zero-Shot Learning, ArXiv-21\n\n[3] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, ECCV-22\n\n[4] CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning, CVPR-23", "questions": "Please refer to the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The approach enables adaptable and efficient continual learning in open-vocabulary image classification. It draws inspiration from human cognition’s complementary learning systems. In this work, the author merges predictions from a CLIP zero-shot model and an exemplar-based model. it is based on using the zero-shot estimated probability that a sample’s class is within the exemplar classes. Inspired by lazy learning principles, the author introduces a “tree probe” method, which facilitating rapid learning from new examples with comparable accuracy to batch-trained linear models.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "[1] The problem is interesting, which predicts the open set vocabulary while following the setting of continual learning.\n\n[2] Results are evaluated over the various datasets in the diverse setting.\n\n[3] “Tree-probe” is interesting, which balance the rapid learning and performance.", "weaknesses": "[1] There are various setting in the continual learning (task incremental, class incremental, data incremental etc.) and zero-shot learning (generalized/non-generalised setting). The exact experimental setting, evaluation strategy, and the motivation of each setting are not clear. It’s difficult to follow the section 4.2, there should me better illustration and discrimination between the various evaluation scenarios.\n\n[2] There are few recent works [1,2] follow the similar setting. These works can be considered as baseline along with the CLIP zero-shot.\n\n[3] I believe that adding the problem setting before the section-3 (Method) with the proper notation will increase the readability. \n\n[4] The recent prompting based continual learning approach [3,4] leverages the strong pretrained model and shows the promising result for the continual learning without complementary memory system. Instead of examplar storage, if model leverages promoting based approach, how the model behaves?\n\nReference:\n\n[1]  Unseen Classes at a Later Time? No Problem, CVPR-22\n\n[2] Meta-Learned Attribute Self-Gating for Continual Generalized Zero-Shot Learning, ArXiv-21\n\n[3] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning, ECCV-22\n\n[4] CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning, CVPR-23", "questions": "Please refer to the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698869253541}, {"id": "kOj3Ov9YfE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1611/Reviewer_ijNk"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a continual learning method for image classification. It addresses quite challenging setups as data incremental, class incremental and task incremental settings. The core of their method is a combination CLIP embedding for zero-shot prediction and tree-based kNN exemplar-based model (TreeProbe). The paper presents results a large variety of datasets and compares with a few well-cited continual learning methods from the past decade. CLIP enables open-vocabulary learning (within CLIP training set, of course) and their suggested TreeProbe shows quick training and inference. The two-model combination is inspired by a famous work in continual learning \"complimentary learning system\" that suggests that human brain has two types of memories: fast episodic memory (hippocampus) and slow consolidating memory (neocortex). It's been exploited in continual learning approaches before with limited success.", "review_text": "The paper proposes a continual learning method for image classification. It addresses quite challenging setups as data incremental, class incremental and task incremental settings. The core of their method is a combination CLIP embedding for zero-shot prediction and tree-based kNN exemplar-based model (TreeProbe). The paper presents results a large variety of datasets and compares with a few well-cited continual learning methods from the past decade. CLIP enables open-vocabulary learning (within CLIP training set, of course) and their suggested TreeProbe shows quick training and inference. The two-model combination is inspired by a famous work in continual learning \"complimentary learning system\" that suggests that human brain has two types of memories: fast episodic memory (hippocampus) and slow consolidating memory (neocortex). It's been exploited in continual learning approaches before with limited success.", "strengths": "I find this paper quite strong. It tackles a very very difficult problem of open-world image classification in incremental learning. Typically neural nets suffer from catastrophic forgetting that makes incremental training pointless pretty much. Many attempts were made on solving this problem. \n\nOverall results look very promising. I am happy to not see permuted MNIST in benchmarks.\n\nThe paper is well-written and easy to read. The paper achieves SOTA results although the gap with the other CLIP-based model is pretty small. I found it convincing enough and sufficiently novel.", "weaknesses": "Using CLIP for incremental learning isn't novel. TreeProbe seems like a quite simple approach, I can't believe it wasn't described before. I couldn't however find a reference.\n\nThe main drawback is that it still relies on CLIP. While we don't know exactly what CLIP was trained on, I find it easy to believe that 400M dataset contains everything that the authors used for evaluation so in a way it is not really continual learning as we would like it to be.", "questions": "The method contains a lot of moving parts around merging CLIP and exemplar model and TreeProbe implementation. I encourage the authors to release their code to facilitate future research.\n\nI also suggest to clearly present accuracy of supervised baselines for every dataset they used. While it is certainly not to compare their method against, it is useful to know \"are we there yet?\" in terms of how practical continual learning has become. It is also useful to know total number of classes that was obtained after merging all datasets.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a continual learning method for image classification. It addresses quite challenging setups as data incremental, class incremental and task incremental settings. The core of their method is a combination CLIP embedding for zero-shot prediction and tree-based kNN exemplar-based model (TreeProbe). The paper presents results a large variety of datasets and compares with a few well-cited continual learning methods from the past decade. CLIP enables open-vocabulary learning (within CLIP training set, of course) and their suggested TreeProbe shows quick training and inference. The two-model combination is inspired by a famous work in continual learning \"complimentary learning system\" that suggests that human brain has two types of memories: fast episodic memory (hippocampus) and slow consolidating memory (neocortex). It's been exploited in continual learning approaches before with limited success.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "I find this paper quite strong. It tackles a very very difficult problem of open-world image classification in incremental learning. Typically neural nets suffer from catastrophic forgetting that makes incremental training pointless pretty much. Many attempts were made on solving this problem. \n\nOverall results look very promising. I am happy to not see permuted MNIST in benchmarks.\n\nThe paper is well-written and easy to read. The paper achieves SOTA results although the gap with the other CLIP-based model is pretty small. I found it convincing enough and sufficiently novel.", "weaknesses": "Using CLIP for incremental learning isn't novel. TreeProbe seems like a quite simple approach, I can't believe it wasn't described before. I couldn't however find a reference.\n\nThe main drawback is that it still relies on CLIP. While we don't know exactly what CLIP was trained on, I find it easy to believe that 400M dataset contains everything that the authors used for evaluation so in a way it is not really continual learning as we would like it to be.", "questions": "The method contains a lot of moving parts around merging CLIP and exemplar model and TreeProbe implementation. I encourage the authors to release their code to facilitate future research.\n\nI also suggest to clearly present accuracy of supervised baselines for every dataset they used. While it is certainly not to compare their method against, it is useful to know \"are we there yet?\" in terms of how practical continual learning has become. It is also useful to know total number of classes that was obtained after merging all datasets.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698700776153}], "openreview_url": "https://openreview.net/forum?id=wE1I9IGqeH", "arxiv_id": "2307.01430", "paper_pdf": "papers/wE1I9IGqeH.pdf", "paper_pdf_sha256": "9f8c58e67298d91c4e3c4914ad261dc56b86965b538cd0fb7985c8bb0c16c8ec", "paper_pdf_bytes": 8557305, "paper_pdf_source": "openreview", "code_url": "https://github.com/jessemelpolio/TreeProbe", "code_repository": "jessemelpolio/TreeProbe", "code_commit": "cbb0b770a0b3c683f6a795f2b42d3934f7c8272a", "code_archive": "repos/wE1I9IGqeH.zip", "code_archive_sha256": "8fc471e892e369320fa4f93ef40ac456e842cd79099bc5cc330e224766f402c8", "code_archive_bytes": 104645, "code_file_count": 42, "code_extensions": {".py": 38, ".sh": 4}, "github_disk_usage_kb": 86, "github_languages": {"Python": 322128, "Shell": 16653}, "github_archived": false, "github_pushed_at": "2024-10-04T03:23:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continual-learning-in-open-vocabulary"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8Vxuz_PJNus", "year": 2023, "status": "rejected", "title": "Federated Learning of Large Models at the Edge via Principal Sub-Model Training", "authors": ["Yue Niu", "Saurav Prakash", "Souvik Kundu", "Sunwoo Lee", "Salman Avestimehr"], "authorids": ["~Yue_Niu1", "~Saurav_Prakash1", "~Souvik_Kundu2", "~Sunwoo_Lee1", "~Salman_Avestimehr1"], "authors_source": "OpenReview API", "abstract": "Limited compute, memory, and communication capabilities of edge users create a significant bottleneck for federated learning (FL) of large models. Current literature typically tackles the challenge with a heterogeneous client setting or allows training to be offloaded to the server. However, the former requires a fraction of clients to train near-full models, which may not be achievable at the edge; while the latter can compromise privacy with sharing of intermediate representations or labels. In this work, we consider a realistic, but much less explored, cross-device FL setting in which no client has the capacity to train a full large model nor is willing to share any intermediate representations with the server. To this end, we present Principal Sub-Model (PriSM) training methodology, which leverages models’ low-rank structure and kernel orthogonality to train sub-models in the orthogonal kernel space. More specifically, by applying singular value decomposition to original kernels in the server model, PriSM first obtains a set of principal orthogonal kernels with importance weighed by their singular values. Thereafter, PriSM utilizes a novel sampling strategy that selects different subsets of the principal kernels independently to create sub-models for clients with reduced computation and communication requirements. Importantly, a kernel with a large singular value is assigned with a high sampling probability. Thus, each sub-model is a low-rank approximation of the full large model, and all clients together achieve nearly full coverage of the principal kernels. To further improve memory efficiency, PriSM exploits low-rank structure in intermediate representations and allows each sub-model to learn only a subset of them while still preserving training performance. Our extensive evaluations on multiple datasets in various resource-constrained settings demonstrate that PriSM can yield an improved performance of up to $10\\%$ compared to existing alternatives, when training sub-models with only $20\\%$ principal kernels ($\\sim 5\\%$ of the full server model.).", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "5hesGPcIW4R", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1451/Reviewer_59uv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The focus of this paper is on federated learning in a setting where the edge clients do not have sufficient resources to train a large model and, additionally, the clients do not want to share any intermediate data and/or labels with the server. The main contribution in the paper is an algorithm, termed Principal Sub-Model (PriSM), in which the convolution kernels are mapped to the so-called *principal kernels* given by the singular value decomposition (SVD) at the server and then clients are asked to update low-rank approximations of the convolution kernels that are given by a subset of the principal kernels. While such an idea exists in the literature, as noted in the paper, the main idea the authors leverage in this paper is to probabilistically decide on the principal kernels that correspond to the submodels at each client.", "review_text": "While the paper puts forth an effective idea, the contribution of the paper is limited and incremental. Such works are indeed valuable, but are perhaps better suited as workshop papers then as full conference papers.\n\n---\n\n**Response to the Authors**\n\nDear Authors,\n\nI greatly regret I was not able to engage with you during Phase I of the discussion period due to emergency of a personal nature. Please accept my sincerest apologies for this, as you put in significant efforts into the work and you deserved a timely response.\n\nI have gone through your response and have expanded / clarified parts of my review. I am however unable to raise my score because we have a disagreement on the novelty of the work. The ideas of low-rank matrix factorization and several other ad-hoc tweaks exist in the literature on compression of neural networks. While your work looks at the federated learning paradigm, and it utilizes an algorithm that helps learning take place in a federated setting, I maintain that these ideas by themselves are incremental in nature, when contrasted with the existing literature, and are not sufficient to warrant a publication in ICLR in the main conference. Perhaps this disagreement is because of I and you possibly coming from different communities, and I would let the AC and SAC resolve this disagreement.", "strengths": "**Strengths**\n\n- The probabilistic selection of the principal kernels for the submodels using the strength of the singular values is an important insight that helps the authors with an improved algorithm.\n- The authors have carried out extensive numerical experiments to highlight the advantages of their approach.\n\n**Weaknesses**\n\n- The paper has an incremental nature, with the major difference from existing literature being a focus on probabilistic sampling of the principal kernels. Perhaps the authors can look into some theoretical aspects of PriSM to overcome the limited innovation in terms of the algorithm. (_Note added after reading authors' response:_ The literature being referred to here is the general literature on compression of neural networks. Low-rank matrix factorization and other structured factorizations, as well as many other ad-hoc tricks, are routinely utilized in works that study compression of neural networks.)\n- ~~The algorithm is very much limited to the case of convolutional neural networks and it is not clear how to adapt these ideas to the case of other architectures.~~ (_The reviewer agrees that some of the ideas in the paper are applicable to some other architectures._)", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The focus of this paper is on federated learning in a setting where the edge clients do not have sufficient resources to train a large model and, additionally, the clients do not want to share any intermediate data and/or labels with the server. The main contribution in the paper is an algorithm, termed Principal Sub-Model (PriSM), in which the convolution kernels are mapped to the so-called *principal kernels* given by the singular value decomposition (SVD) at the server and then clients are asked to update low-rank approximations of the convolution kernels that are given by a subset of the principal kernels. While such an idea exists in the literature, as noted in the paper, the main idea the authors leverage in this paper is to probabilistically decide on the principal kernels that correspond to the submodels at each client.", "strength_and_weaknesses": "**Strengths**\n\n- The probabilistic selection of the principal kernels for the submodels using the strength of the singular values is an important insight that helps the authors with an improved algorithm.\n- The authors have carried out extensive numerical experiments to highlight the advantages of their approach.\n\n**Weaknesses**\n\n- The paper has an incremental nature, with the major difference from existing literature being a focus on probabilistic sampling of the principal kernels. Perhaps the authors can look into some theoretical aspects of PriSM to overcome the limited innovation in terms of the algorithm. (_Note added after reading authors' response:_ The literature being referred to here is the general literature on compression of neural networks. Low-rank matrix factorization and other structured factorizations, as well as many other ad-hoc tricks, are routinely utilized in works that study compression of neural networks.)\n- ~~The algorithm is very much limited to the case of convolutional neural networks and it is not clear how to adapt these ideas to the case of other architectures.~~ (_The reviewer agrees that some of the ideas in the paper are applicable to some other architectures._)", "clarity,_quality,_novelty_and_reproducibility": "- The paper is easy to read and is of good quality. \n- The novelty of the paper is limited since the main contribution appears to be a small tweak to an existing set of approaches in the literature.\n- There are no links to any online code in the paper, which does not allow this reviewer to judge the reproducibility of this work.", "summary_of_the_review": "While the paper puts forth an effective idea, the contribution of the paper is limited and incremental. Such works are indeed valuable, but are perhaps better suited as workshop papers then as full conference papers.\n\n---\n\n**Response to the Authors**\n\nDear Authors,\n\nI greatly regret I was not able to engage with you during Phase I of the discussion period due to emergency of a personal nature. Please accept my sincerest apologies for this, as you put in significant efforts into the work and you deserved a timely response.\n\nI have gone through your response and have expanded / clarified parts of my review. I am however unable to raise my score because we have a disagreement on the novelty of the work. The ideas of low-rank matrix factorization and several other ad-hoc tweaks exist in the literature on compression of neural networks. While your work looks at the federated learning paradigm, and it utilizes an algorithm that helps learning take place in a federated setting, I maintain that these ideas by themselves are incremental in nature, when contrasted with the existing literature, and are not sufficient to warrant a publication in ICLR in the main conference. Perhaps this disagreement is because of I and you possibly coming from different communities, and I would let the AC and SAC resolve this disagreement.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666918044607}, {"id": "KWP9pGCv95P", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1451/Reviewer_vmY3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the problem of learning a large model in a federated manner by clients with limited computational and memory capabilities. The paper is mainly about large neural networks. The proposed solution is based on SVD decomposition of each layer to provide clients with a low rank representation of each layer. This reduces the amount of computations required to update the models locally by clients.", "review_text": "In summary, the paper studies an important problem and proposes a novel algorithm to solve this problem. However, there are important questions that should be answered and discussed by the paper. Also, writing and presentation of the paper need improvement.", "strengths": "Strengths:\n\n1. The paper proposes a new method to enable clients with limited computational capability to train a large model.\n2. Using the proposed method, clients with limited memory can collaborate on training a large model that cannot be fit into their memory.\n\nWeaknesses:\n\n1. This works lacks theoretical analysis. For example, it would be interesting to see if there are any other sampling method that can provide better result. Also it is not clear that how $r$ affects the training accuracy.\n2. An important question that this paper does not respond is that why training a low rank representation of a large model is better than training of a original smaller model.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies the problem of learning a large model in a federated manner by clients with limited computational and memory capabilities. The paper is mainly about large neural networks. The proposed solution is based on SVD decomposition of each layer to provide clients with a low rank representation of each layer. This reduces the amount of computations required to update the models locally by clients.", "strength_and_weaknesses": "Strengths:\n\n1. The paper proposes a new method to enable clients with limited computational capability to train a large model.\n2. Using the proposed method, clients with limited memory can collaborate on training a large model that cannot be fit into their memory.\n\nWeaknesses:\n\n1. This works lacks theoretical analysis. For example, it would be interesting to see if there are any other sampling method that can provide better result. Also it is not clear that how $r$ affects the training accuracy.\n2. An important question that this paper does not respond is that why training a low rank representation of a large model is better than training of a original smaller model.\n\n", "clarity,_quality,_novelty_and_reproducibility": "I believe the paper can be improved in terms of writing and presentation. I had difficulty to understand some parts of the paper due to quality of writing. Specifically, I think authors can provide more detailed discussions on local training part on page 5. I am not clear that what it means that ``the selected $\\sigma_i,u_i, v_i  i\\in\\mathcal I_c$ are updated, together with trainable parameters in other layers.'' If all trainable parameters in other layers need to be updated by the client how the proposed method help the clients with limited memory and computational capability?", "summary_of_the_review": "In summary, the paper studies an important problem and proposes a novel algorithm to solve this problem. However, there are important questions that should be answered and discussed by the paper. Also, writing and presentation of the paper need improvement.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666680511347}, {"id": "NP4nOimu9W", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1451/Reviewer_v26F"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies the problem of federated learning on edge devices with limited computation, memory, and communication capacities. Specifically, this paper proposed PriSM, a sub-model training approach for FL. PriSM is a low-rank approach that performs SVD on the kernels to build a sub-model on device during the training.  With empirical evaluation, the authors show that PriSM is able to train a model with only 20% of the principle kernels and has 10% performance improvements over existing solutions.\n", "review_text": "\nThis paper proposes a powerful low-rank approach in FL for sub-model training on-device. I have already mentioned the strengths in the above section. Here I would like to raise a few questions for clarification.\n\n1. What is the on-device computation efficiency? How many FLOPs does the PriSM save when performing FL? Does this saving result in clock-time efficiency improvements?\n\n2.  Does PriSM generalize to MLP/Transformer models? The linear model can also be low-rank decomposed on-device for fast training.\n\n3. Is there any extra overhead in the aggregation phase? How does this overhead compare with the original aggregation procedure in FedAvg?", "strengths": "Strength\n1. This paper is well-written with clear formulation and introduction of PriSM\n\n2. The empirical evaluation of ResNet models is solid across different vision FL datasets.\n\n3. In the appendix, the paper presents a detailed parameter selection, making PriSM reproducible\n\nWeakness\n1. A stronger analysis of the on-device efficiency should be introduced.\n\n2. It would be better for PriSM to be applied to general linear models instead of focusing on convolution blocks.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies the problem of federated learning on edge devices with limited computation, memory, and communication capacities. Specifically, this paper proposed PriSM, a sub-model training approach for FL. PriSM is a low-rank approach that performs SVD on the kernels to build a sub-model on device during the training.  With empirical evaluation, the authors show that PriSM is able to train a model with only 20% of the principle kernels and has 10% performance improvements over existing solutions.\n", "strength_and_weaknesses": "Strength\n1. This paper is well-written with clear formulation and introduction of PriSM\n\n2. The empirical evaluation of ResNet models is solid across different vision FL datasets.\n\n3. In the appendix, the paper presents a detailed parameter selection, making PriSM reproducible\n\nWeakness\n1. A stronger analysis of the on-device efficiency should be introduced.\n\n2. It would be better for PriSM to be applied to general linear models instead of focusing on convolution blocks.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-organized and the PriSM approach seems to be reproducible with a clear instruction on the hyper-parameters.", "summary_of_the_review": "\nThis paper proposes a powerful low-rank approach in FL for sub-model training on-device. I have already mentioned the strengths in the above section. Here I would like to raise a few questions for clarification.\n\n1. What is the on-device computation efficiency? How many FLOPs does the PriSM save when performing FL? Does this saving result in clock-time efficiency improvements?\n\n2.  Does PriSM generalize to MLP/Transformer models? The linear model can also be low-rank decomposed on-device for fast training.\n\n3. Is there any extra overhead in the aggregation phase? How does this overhead compare with the original aggregation procedure in FedAvg?", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666660860969}], "openreview_url": "https://openreview.net/forum?id=8Vxuz_PJNus", "arxiv_id": "2208.13141", "paper_pdf": "papers/8Vxuz_PJNus.pdf", "paper_pdf_sha256": "7c4b4f9183dde522018b107ef32e78d424088a9eec962a1b9c7f599605497518", "paper_pdf_bytes": 1267211, "paper_pdf_source": "openreview", "code_url": "https://github.com/yuehniu/modeldecomp-fl", "code_repository": "yuehniu/modeldecomp-fl", "code_commit": "04e213968afd2dc7ae4097f8bc1d33b8bdb66ee2", "code_archive": "repos/8Vxuz_PJNus.zip", "code_archive_sha256": "ebf1b08048f776b240b511aea827134034d79f3b48a76fbc99fb5d40f1486417", "code_archive_bytes": 102796, "code_file_count": 30, "code_extensions": {".py": 29, ".sh": 1}, "github_disk_usage_kb": 167, "github_languages": {"Python": 240421, "Shell": 1127}, "github_archived": false, "github_pushed_at": "2023-10-05T18:27:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/federated-learning-of-large-models-at-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cWlMII1LwTZ", "year": 2022, "status": "rejected", "title": "Task-aware Privacy Preservation for Multi-dimensional Data", "authors": ["Jiangnan Cheng", "Ao Tang", "Sandeep P. Chinchali"], "authorids": ["~Jiangnan_Cheng1", "~Ao_Tang1", "~Sandeep_P._Chinchali1"], "authors_source": "OpenReview API", "abstract": "Local differential privacy (LDP), a state-of-the-art technique for privacy preservation, has been successfully deployed in a few real-world applications. In the future, LDP can be adopted to anonymize richer user data attributes that will be input to more sophisticated machine learning (ML) tasks. However, today's LDP approaches are largely task-agnostic and often lead to sub-optimal performance - they will simply inject noise to all data attributes according to a given privacy budget, regardless of what features are most relevant for an ultimate task. In this paper, we address how to significantly improve the ultimate task performance for multi-dimensional user data by considering a task-aware privacy preservation problem. The key idea is to use an encoder-decoder framework to learn (and anonymize) a task-relevant latent representation of user data, which gives an analytical near-optimal solution for a linear setting with mean-squared error (MSE) task loss. We also provide an approximate solution through a learning algorithm for general nonlinear cases. Extensive experiments demonstrate that our task-aware approach significantly improves ultimate task accuracy compared to a standard benchmark LDP approach while guaranteeing the same level of privacy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Vw4tosOROEj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper330/Reviewer_r5c5"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper provides task-aware privacy preservation to improve the task performance for multi-dimensional user data for deploying a trained model. The paper’s spirit is to use an encoder-decoder framework with a Laplace noise added to learn a task-relevant but LDP privacy preserved latent representation of user data. The author also provides an analytical near-optimal solution for a linear setting with mean-squared error task loss. They show the effectiveness of their method via experiments on three real-world datasets.", "review_text": " Strengths\n+ The problem formulation is simple and easy to follow to get the idea of the proposed method. \n+ They provide an intensive and rigorous analysis for an analytical near-optimal solution for a linear setting and MSE task loss in terms of task-aware privacy preservation analysis.\n\nWeaknesses\n-\tFirst, the general idea of the paper is interesting, but it seems to be incremental and at an early stage of this work. The novelty is limited. The idea of redistributing the noise across input features has been extensively studied, for instance, with autoencoder-decoder and layer-wise relevant propagation, and forward derivatives (Phan et al., 2017; 2019). These references should be considered as baselines for comparison to highlight the novelty of the proposed approach. However, these critical references are missing. \n-\tSecond, the reviewer does not understand why we need LDP here and why not centralized DP? What is the threat model? Training the autoencoder-decoder would need a centralized server to gather all the data tuples together. If so, there is no need for LDP.\n-\tThird, how practical is Assumption 1? It is not clear at all in the current writing. Also, it is unclear to the reviewer that why Propositions 1-4 result in an optimal encoder and decoder design that preserves $\\epsilon$-LDP.\n-\tIn the analysis, a linear model with an MSE task loss is used in the analysis and experiments, which may not be easy to extend to other complex tasks, such as deep learning. \n-\tExperimental results can be improved. The datasets used in the experiments still have a limited number of dimensions, such as dim=24 in power consumption, dim=6 in real estate valuation, and dim=30 in breast cancer detection. It would be interesting if the work can be done with the more complex datasets and learning tasks, such as image datasets with thousands of dimensions and deep learning tasks for example. Adaptive mechanisms such as (Phan et al., 2017; 2019; 2020) and other related works listed below can be considered as baseline approaches for comparison. \n-\tThe compared benchmark approaches are unclear. There is no place to define or refer to the compared method. The reviewer needs to guess what it means. Also, there is a limited number of comparisons, which makes the paper unconvincing. \n\nNhatHai Phan, Xintao Wu, Han Hu, Dejing Dou. Adaptive Laplace Mechanism: Differential Privacy Preservation in Deep Learning. IEEE ICDM'17.\n\nNhatHai Phan, Minh Vu, Yang Liu, Ruoming Jin, Xintao Wu, Dejing Dou, and My T. Thai. Heterogeneous Gaussian Mechanism: Preserving Differential Privacy in Deep Learning with Provable Robustness. IJCAI'19.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides task-aware privacy preservation to improve the task performance for multi-dimensional user data for deploying a trained model. The paper’s spirit is to use an encoder-decoder framework with a Laplace noise added to learn a task-relevant but LDP privacy preserved latent representation of user data. The author also provides an analytical near-optimal solution for a linear setting with mean-squared error task loss. They show the effectiveness of their method via experiments on three real-world datasets.", "main_review": " Strengths\n+ The problem formulation is simple and easy to follow to get the idea of the proposed method. \n+ They provide an intensive and rigorous analysis for an analytical near-optimal solution for a linear setting and MSE task loss in terms of task-aware privacy preservation analysis.\n\nWeaknesses\n-\tFirst, the general idea of the paper is interesting, but it seems to be incremental and at an early stage of this work. The novelty is limited. The idea of redistributing the noise across input features has been extensively studied, for instance, with autoencoder-decoder and layer-wise relevant propagation, and forward derivatives (Phan et al., 2017; 2019). These references should be considered as baselines for comparison to highlight the novelty of the proposed approach. However, these critical references are missing. \n-\tSecond, the reviewer does not understand why we need LDP here and why not centralized DP? What is the threat model? Training the autoencoder-decoder would need a centralized server to gather all the data tuples together. If so, there is no need for LDP.\n-\tThird, how practical is Assumption 1? It is not clear at all in the current writing. Also, it is unclear to the reviewer that why Propositions 1-4 result in an optimal encoder and decoder design that preserves $\\epsilon$-LDP.\n-\tIn the analysis, a linear model with an MSE task loss is used in the analysis and experiments, which may not be easy to extend to other complex tasks, such as deep learning. \n-\tExperimental results can be improved. The datasets used in the experiments still have a limited number of dimensions, such as dim=24 in power consumption, dim=6 in real estate valuation, and dim=30 in breast cancer detection. It would be interesting if the work can be done with the more complex datasets and learning tasks, such as image datasets with thousands of dimensions and deep learning tasks for example. Adaptive mechanisms such as (Phan et al., 2017; 2019; 2020) and other related works listed below can be considered as baseline approaches for comparison. \n-\tThe compared benchmark approaches are unclear. There is no place to define or refer to the compared method. The reviewer needs to guess what it means. Also, there is a limited number of comparisons, which makes the paper unconvincing. \n\nNhatHai Phan, Xintao Wu, Han Hu, Dejing Dou. Adaptive Laplace Mechanism: Differential Privacy Preservation in Deep Learning. IEEE ICDM'17.\n\nNhatHai Phan, Minh Vu, Yang Liu, Ruoming Jin, Xintao Wu, Dejing Dou, and My T. Thai. Heterogeneous Gaussian Mechanism: Preserving Differential Privacy in Deep Learning with Provable Robustness. IJCAI'19.", "summary_of_the_review": "The general idea of the paper is interesting, but it seems to be an early stage of this work. Investigating the work with more complex datasets and tasks and more baseline comparisons would make the paper more convincing. There are a lot baseline of LDP or task-aware work to compare with the work, for example\n[1] Wang, N., Xiao, X., Yang, Y., Zhao, J., Hui, S. C., Shin, H., ... & Yu, G. (2019, April). Collecting and analyzing multidimensional data with local differential privacy. In 2019 IEEE 35th International Conference on Data Engineering (ICDE) (pp. 638-649). IEEE.\n[2] Liu, R., Cao, Y., Yoshikawa, M., & Chen, H. (2020, September). Fedsel: Federated sgd under local differential privacy with top-k dimension selection. In International Conference on Database Systems for Advanced Applications (pp. 485-501). Springer, Cham.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636026895338}, {"id": "G4MKjTlWo2v", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper330/Reviewer_mz4E"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper addresses the problem of large loss in utility due to the addition of noise in a local-DP setup, especially when working with high dimensional data. To mitigate this problem, the paper proposes an encoder-decoder setup in which noise is applied through the encoder, in a lower-dimension to decrease its effect on utility. Then  Laplace random noise is applied to each point's encoding in the latent space. The training  of the autoencoder is done over an offline phase on public data and with respect to the main task that the data is to be used for.  The approach is then evaluated empirically.", "review_text": "The curse of dimensionality and the problems it creates for utility of DP mechanisms is an important problem that has been studied for a long time. The approach this paper take's creates the following concerns:\n\n1. The main idea of this  paper  which is dimensionality reduction, has already been proposed before, in numerous contexts such as random projections [1] or sparse regression [2]. The idea of mixing this with task-awareness and considering how a model would use features and adding noise only to features that are used has also been studied before [3]. It would be great if the paper quantifies its differences with these prior work/missing citations better.\n\n2. How much overhead does this add to the user's side, in terms of latency and computation complexity? It will help to know how much extra latency gets added compared to doing normal local DP. The user seems to need to be able to carry out the encoder operation. \n\n3. How much of this method's success depend on a correct analysis of what feature is related to the  task? How easy or hard is it to get such details? Some more clarity in explaining this would certainly help readability. It seems like the effectiveness of this method relates to how well related features can be extracted, and what this model is learning and if it really relates to how this data would be used in the future for other training tasks. This could even circle back to interpretability and how different models might use different features. \n\n4. Although the paper claims effectiveness in high-dimensonal cases, the experiments  create the question \"how well would this method scale in the presence of real-world, high dimensional data?\", like medical images which are often large in size. The datasets used here are very small compared to real-life scenarios.  \n\n[1] Kenthapadi K, Korolova A, Mironov I, Mishra N. Privacy via the Johnson-Lindenstrauss Transform. Journal of Privacy and Confidentiality. 2013;5(1):39-71.\n\n[2] Thakurta AG, Smith A. Differentially private feature selection via stability arguments, and the robustness of the lasso. InConference on Learning Theory 2013 Jun 13 (pp. 819-850). PMLR.\n\n[3] Mireshghallah F, Taram M, Jalali A, Elthakeb AT, Tullsen D, Esmaeilzadeh H. A principled approach to learning stochastic representations for privacy in deep neural inference. arXiv preprint arXiv:2003.12154. 2020 Mar.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper addresses the problem of large loss in utility due to the addition of noise in a local-DP setup, especially when working with high dimensional data. To mitigate this problem, the paper proposes an encoder-decoder setup in which noise is applied through the encoder, in a lower-dimension to decrease its effect on utility. Then  Laplace random noise is applied to each point's encoding in the latent space. The training  of the autoencoder is done over an offline phase on public data and with respect to the main task that the data is to be used for.  The approach is then evaluated empirically.", "main_review": "The curse of dimensionality and the problems it creates for utility of DP mechanisms is an important problem that has been studied for a long time. The approach this paper take's creates the following concerns:\n\n1. The main idea of this  paper  which is dimensionality reduction, has already been proposed before, in numerous contexts such as random projections [1] or sparse regression [2]. The idea of mixing this with task-awareness and considering how a model would use features and adding noise only to features that are used has also been studied before [3]. It would be great if the paper quantifies its differences with these prior work/missing citations better.\n\n2. How much overhead does this add to the user's side, in terms of latency and computation complexity? It will help to know how much extra latency gets added compared to doing normal local DP. The user seems to need to be able to carry out the encoder operation. \n\n3. How much of this method's success depend on a correct analysis of what feature is related to the  task? How easy or hard is it to get such details? Some more clarity in explaining this would certainly help readability. It seems like the effectiveness of this method relates to how well related features can be extracted, and what this model is learning and if it really relates to how this data would be used in the future for other training tasks. This could even circle back to interpretability and how different models might use different features. \n\n4. Although the paper claims effectiveness in high-dimensonal cases, the experiments  create the question \"how well would this method scale in the presence of real-world, high dimensional data?\", like medical images which are often large in size. The datasets used here are very small compared to real-life scenarios.  \n\n[1] Kenthapadi K, Korolova A, Mironov I, Mishra N. Privacy via the Johnson-Lindenstrauss Transform. Journal of Privacy and Confidentiality. 2013;5(1):39-71.\n\n[2] Thakurta AG, Smith A. Differentially private feature selection via stability arguments, and the robustness of the lasso. InConference on Learning Theory 2013 Jun 13 (pp. 819-850). PMLR.\n\n[3] Mireshghallah F, Taram M, Jalali A, Elthakeb AT, Tullsen D, Esmaeilzadeh H. A principled approach to learning stochastic representations for privacy in deep neural inference. arXiv preprint arXiv:2003.12154. 2020 Mar.", "summary_of_the_review": "\nThe main reasons for my recommendation are:\n\n1. similarity to prior work, the novelty of this work and it's benefits are not well identified\n\n2. How useful the extracted dimensions would be in high dimensions and for models that might use different set of features is not clear.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635833609912}, {"id": "Pr1WHEQgRm", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper330/Reviewer_S4hg"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a task-aware local DP method to improve the privacy and\u0002utility trade-off for multi-dimensional user data. Also, a analytical near-optimal solution for a general linear encoder-decoder model and MSE loss is provided. For neural network cases (i.e., nonlinear encoder-decoder), the authors present a heuristic learning algorithm to get the model parameters. The experiments results demonstrate the effectiveness of the proposed methods. ", "review_text": "Strengths:\n1. The problem of improving utility and privacy trade-off in local differentially private ML is an important practical problem.\n2. For linear encoder and decoder, and MSE loss setting, an analytical near-optimal solution is provided. The comparisons among task-aware approach and the benchmark approaches give an intuitive interpretation of why the parameters obtained from task-aware approach are better. \n \n\nMy main concern in this paper is whether the proposed method can preserve local DP. \n1. The local DP setting is not clearly given, which I mean, interactive [1] or non-interactive setting [2]? \n2. Based on my understanding of this paper,  Algorithm 1 needs $N_{epochs}$ to find the encoder, decoder and sensitivity. During this interactive proposes, the algorithm continues to query the sensitive user data samples. I think the composition theorem should be considered. \n3. For achieving private local data representation,  Wang et al. in [3] also proposed a lightweight privacy-preserving mechanism consisting of arbitrary data nullification and DP noise addition. Compared with [3], I find Algorithm 1 follows a quite similar process when learning the model parameters. Since [3] introduces data nullification and adversary training, their method may be more efficient and better than this paper in terms of communication and utility. Please provide a comparison with [3]. \n\n\n[1] Joseph, Matthew, Jieming Mao, Seth Neel, and Aaron Roth. \"The role of interactivity in local differential privacy.\" In 2019 IEEE 60th Annual Symposium on Foundations of Computer Science (FOCS), pp. 94-105. IEEE, 2019.\n[2] Wang, D., Gaboardi, M., Smith, A., & Xu, J. (2020). Empirical Risk Minimization in the Non-interactive Local Model of Differential Privacy. Journal of machine learning research, 21(200).\n[3] Wang, Ji, Jianguo Zhang, Weidong Bao, Xiaomin Zhu, Bokai Cao, and Philip S. Yu. \"Not just privacy: Improving performance of private deep learning in mobile cloud.\" In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 2407-2416. 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a task-aware local DP method to improve the privacy and\u0002utility trade-off for multi-dimensional user data. Also, a analytical near-optimal solution for a general linear encoder-decoder model and MSE loss is provided. For neural network cases (i.e., nonlinear encoder-decoder), the authors present a heuristic learning algorithm to get the model parameters. The experiments results demonstrate the effectiveness of the proposed methods. ", "main_review": "Strengths:\n1. The problem of improving utility and privacy trade-off in local differentially private ML is an important practical problem.\n2. For linear encoder and decoder, and MSE loss setting, an analytical near-optimal solution is provided. The comparisons among task-aware approach and the benchmark approaches give an intuitive interpretation of why the parameters obtained from task-aware approach are better. \n \n\nMy main concern in this paper is whether the proposed method can preserve local DP. \n1. The local DP setting is not clearly given, which I mean, interactive [1] or non-interactive setting [2]? \n2. Based on my understanding of this paper,  Algorithm 1 needs $N_{epochs}$ to find the encoder, decoder and sensitivity. During this interactive proposes, the algorithm continues to query the sensitive user data samples. I think the composition theorem should be considered. \n3. For achieving private local data representation,  Wang et al. in [3] also proposed a lightweight privacy-preserving mechanism consisting of arbitrary data nullification and DP noise addition. Compared with [3], I find Algorithm 1 follows a quite similar process when learning the model parameters. Since [3] introduces data nullification and adversary training, their method may be more efficient and better than this paper in terms of communication and utility. Please provide a comparison with [3]. \n\n\n[1] Joseph, Matthew, Jieming Mao, Seth Neel, and Aaron Roth. \"The role of interactivity in local differential privacy.\" In 2019 IEEE 60th Annual Symposium on Foundations of Computer Science (FOCS), pp. 94-105. IEEE, 2019.\n[2] Wang, D., Gaboardi, M., Smith, A., & Xu, J. (2020). Empirical Risk Minimization in the Non-interactive Local Model of Differential Privacy. Journal of machine learning research, 21(200).\n[3] Wang, Ji, Jianguo Zhang, Weidong Bao, Xiaomin Zhu, Bokai Cao, and Philip S. Yu. \"Not just privacy: Improving performance of private deep learning in mobile cloud.\" In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 2407-2416. 2018.", "summary_of_the_review": "This paper proposed a good analysis for task-aware local DP.  However, my main concern is the setting of local DP.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635827063767}, {"id": "uEbTjT12vii", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper330/Reviewer_mTuz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a task aware local-DP approach to improve the performance on multi-dimensional data with same level of privacy. This approach is based on encoder-decoder framework that perturbs only the task relevant encoding instead of the raw user data that generally causes less noise addition and improves the task accuracy. Besides, the paper provide a heuristic learning algorithm for more general settings. The proposed approach is compared with task agnostic and privacy agnostic approaches on different datasets and the results show the proposed method outperformes them on overall task loss under different privacy budgets.", "review_text": "I think this paper addresses an important problem which is improving the privacy-utility trade-off in LDP where there are multiple users' data.\nThe proposed approaches solves the problem in an efficient way. The motivation is clear, it is well-structured and well-written. In my opinion, there are three things that can improve this paper. i) The paper presents a heuristic learning algorithm, but the analysis of the task-aware privacy preservation problem for approximate LDP would be more interesting. ii) Comparison with the state of the art approaches both in text and experiments could be improved. In the current version, the difference between the previously proposed methods is not so clear. iii) The results on larger datasets and the result in terms of accuracy. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a task aware local-DP approach to improve the performance on multi-dimensional data with same level of privacy. This approach is based on encoder-decoder framework that perturbs only the task relevant encoding instead of the raw user data that generally causes less noise addition and improves the task accuracy. Besides, the paper provide a heuristic learning algorithm for more general settings. The proposed approach is compared with task agnostic and privacy agnostic approaches on different datasets and the results show the proposed method outperformes them on overall task loss under different privacy budgets.", "main_review": "I think this paper addresses an important problem which is improving the privacy-utility trade-off in LDP where there are multiple users' data.\nThe proposed approaches solves the problem in an efficient way. The motivation is clear, it is well-structured and well-written. In my opinion, there are three things that can improve this paper. i) The paper presents a heuristic learning algorithm, but the analysis of the task-aware privacy preservation problem for approximate LDP would be more interesting. ii) Comparison with the state of the art approaches both in text and experiments could be improved. In the current version, the difference between the previously proposed methods is not so clear. iii) The results on larger datasets and the result in terms of accuracy. \n\n", "summary_of_the_review": "This paper considers and important problem and brings an efficient solution. It is well-written. The experiments shows the performance improvement of the proposed method. However, the experiements could be extended on larger datasets and they can be compared with SOTA.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635801501893}], "openreview_url": "https://openreview.net/forum?id=cWlMII1LwTZ", "arxiv_id": "2110.02329", "paper_pdf": "papers/cWlMII1LwTZ.pdf", "paper_pdf_sha256": "b7c924d00150c871c86583ab5dd541450db469282f002fc757f9cb56526e5172", "paper_pdf_bytes": 510504, "paper_pdf_source": "openreview", "code_url": "https://github.com/chengjiangnan/task_aware_privacy", "code_repository": "chengjiangnan/task_aware_privacy", "code_commit": "2500da5b665945075c285ff575d519e256f66601", "code_archive": "repos/cWlMII1LwTZ.zip", "code_archive_sha256": "e655a53da9c6962cd52eea195c8060ae65ccb2c53f75149e70cd65db42546149", "code_archive_bytes": 351852, "code_file_count": 64, "code_extensions": {".sh": 39, ".py": 25}, "github_disk_usage_kb": 315, "github_languages": {"Python": 101625, "Shell": 10896}, "github_archived": false, "github_pushed_at": "2022-06-13T19:46:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/task-aware-privacy-preservation-for-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PuG6vCSbrV9", "year": 2021, "status": "rejected", "title": "Density estimation on low-dimensional manifolds: an inflation-deflation approach", "authors": ["Christian Horvat"], "authorids": ["~Christian_Horvat1"], "authors_source": "OpenReview API", "abstract": "Normalizing Flows (NFs) are universal density estimators based on Neuronal Networks. However, this universality is limited: the density's support needs to be diffeomorphic to a Euclidean space. In this paper, we propose a novel method to overcome this limitation without sacrificing the universality. The proposed method inflates the data manifold by adding noise in the normal space, trains an NF on this inflated manifold and, finally, deflates the learned density. Our main result provides sufficient conditions on the manifold and the specific choice of noise under which the corresponding estimator is exact. Our method has the same computational complexity as NFs, and does not require to compute an inverse flow. We also show that, if the embedding dimension is much larger than the manifold dimension, noise in the normal space can be well approximated by some Gaussian noise. This allows using our method for approximating arbitrary densities on non-flat manifolds provided that the manifold dimension is known. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1ZPdkqRIYqP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1691/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work presents a novel theoretical development to tackle the problem of estimating normalising flows (NF) for data with support on complex manifolds. Motivated by the fact that NF are diffeomorphic  transformations from a simple space, ideally Euclidean, the author address the problem of modeling data which distribution is defined on more complex and unknow manifolds.  \nThe idea consists in inflating the data manifold with suitable noise (normal noise) in order to make it diffeomorphic to a simpler space where a NF can be estimated. The data density can be therefore approximated from the inflated space through an opportune deflation operation.\n\nThe main contribution of the work consists in formalising the conditions on the inflating noise to obtain the results of Theorem 1. In practice, Proposition 1 shows that the noise can be drawn from a Gaussian with variance depending from the curvature of the manifold. The experimental section demonstrates the theory on simple synthetic cases on S2.\n\nThe idea proposed in the paper is interesting, and of potentially utility. The derivation is however unclear in some aspects, especially concerning the definition of Q-separation. In particular, Definition 1 implies null measure for the set of manifold points spanning the normal subspace. However, to my understanding, this definition is not sufficient to guarantee the unicity of the mapping between normal space and manifold point. For example, any finite collection of manifold points would still have zero measure. In this case however, the unique correspondence between normal space and manifold is broken. Back to Example 2, this non-unicity can be achieved by increasing the noise along the normal direction to cross the origin (0,0). In this case each point of the inflated space has two generators, making the isomorphic mapping  impossible. I can understand that this aspect is related to the choice of the noise discussed in Section 3.3, but this issue still undermines the conclusion of Theorem 1. I would encourage the authors to clarify this point.\n\nThere are  further technical aspects deserving clarification. In Example 2, the definition of the set A(\\tilde{x}) should not contemplate the empty set. Rather, A(\\tilde{x}) should be the point x’ if \\tilde{x} is not (0,0). Moreover, in Proof A.1, equation (13), the relation between the q_n and \\hat{q}_n in the two first integrals, and the reason why they can be exchanged, is not clear. \n\nFinally, some questions arise concerning the practical feasibility of the proposed theory. Computing the curvature of the data manifold may be extremely challenging in presence of complex data, especially in the low sample-size regime. In this case, the estimation of the noise parameter in formula (9) may be practically impossible.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting framework to train NF for manifold-supported data. Questions about theoretical development and practical utility.", "review": "This work presents a novel theoretical development to tackle the problem of estimating normalising flows (NF) for data with support on complex manifolds. Motivated by the fact that NF are diffeomorphic  transformations from a simple space, ideally Euclidean, the author address the problem of modeling data which distribution is defined on more complex and unknow manifolds.  \nThe idea consists in inflating the data manifold with suitable noise (normal noise) in order to make it diffeomorphic to a simpler space where a NF can be estimated. The data density can be therefore approximated from the inflated space through an opportune deflation operation.\n\nThe main contribution of the work consists in formalising the conditions on the inflating noise to obtain the results of Theorem 1. In practice, Proposition 1 shows that the noise can be drawn from a Gaussian with variance depending from the curvature of the manifold. The experimental section demonstrates the theory on simple synthetic cases on S2.\n\nThe idea proposed in the paper is interesting, and of potentially utility. The derivation is however unclear in some aspects, especially concerning the definition of Q-separation. In particular, Definition 1 implies null measure for the set of manifold points spanning the normal subspace. However, to my understanding, this definition is not sufficient to guarantee the unicity of the mapping between normal space and manifold point. For example, any finite collection of manifold points would still have zero measure. In this case however, the unique correspondence between normal space and manifold is broken. Back to Example 2, this non-unicity can be achieved by increasing the noise along the normal direction to cross the origin (0,0). In this case each point of the inflated space has two generators, making the isomorphic mapping  impossible. I can understand that this aspect is related to the choice of the noise discussed in Section 3.3, but this issue still undermines the conclusion of Theorem 1. I would encourage the authors to clarify this point.\n\nThere are  further technical aspects deserving clarification. In Example 2, the definition of the set A(\\tilde{x}) should not contemplate the empty set. Rather, A(\\tilde{x}) should be the point x’ if \\tilde{x} is not (0,0). Moreover, in Proof A.1, equation (13), the relation between the q_n and \\hat{q}_n in the two first integrals, and the reason why they can be exchanged, is not clear. \n\nFinally, some questions arise concerning the practical feasibility of the proposed theory. Computing the curvature of the data manifold may be extremely challenging in presence of complex data, especially in the low sample-size regime. In this case, the estimation of the noise parameter in formula (9) may be practically impossible.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603888036574}, {"id": "l4KXKqe9FV8", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1691/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper tackles the problem of recovering a probability distribution, which is supported in a low-dimensional manifold. When the dimension of that manifold is full, the problem can be solved by the Normalizing Flow method. \n\nSince the dimension is not full, it is suggested to add a gaussian noise to the data points (this is equivalent to a certain convolution of the initial distribution function, ie equation 3). Then full dimensional manifold is recovered and subsequently deconvolution is applied (equation 7).\n\nCertain mathematical aspects of the narrative are vague. Eg \\tilde{X} is defined as X x Noise, but then in equations 4 and 5, it becomes a subset of the plane. \n\nThe notion of Q-normally separated is never satisfied in practice. It is only slightly stronger than \"all the normal spaces have no intersections at all\". The deconvolution made by equation 7 is rarely is a good solution.\n\nAlso, the experimental part is purely dealing with synthetic examples. This could be because the deconvolution by equation 7 is not the best way to recover the signal. \n\nIt is claimed that the first goal of synthetic experiments is to confirm the scaling factor in equation (7) numerically. But experiments deal with highly specific synthetic data - a curve (ie, 1D manifold) on the plane. It is doubtful that such a simple experimental framework can justify eq (7). It is not computationally difficult to check the quality of deconvolution in a style of eq (7) for higher-dimensional data.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Mathematical part is vague, no experiments", "review": "The paper tackles the problem of recovering a probability distribution, which is supported in a low-dimensional manifold. When the dimension of that manifold is full, the problem can be solved by the Normalizing Flow method. \n\nSince the dimension is not full, it is suggested to add a gaussian noise to the data points (this is equivalent to a certain convolution of the initial distribution function, ie equation 3). Then full dimensional manifold is recovered and subsequently deconvolution is applied (equation 7).\n\nCertain mathematical aspects of the narrative are vague. Eg \\tilde{X} is defined as X x Noise, but then in equations 4 and 5, it becomes a subset of the plane. \n\nThe notion of Q-normally separated is never satisfied in practice. It is only slightly stronger than \"all the normal spaces have no intersections at all\". The deconvolution made by equation 7 is rarely is a good solution.\n\nAlso, the experimental part is purely dealing with synthetic examples. This could be because the deconvolution by equation 7 is not the best way to recover the signal. \n\nIt is claimed that the first goal of synthetic experiments is to confirm the scaling factor in equation (7) numerically. But experiments deal with highly specific synthetic data - a curve (ie, 1D manifold) on the plane. It is doubtful that such a simple experimental framework can justify eq (7). It is not computationally difficult to check the quality of deconvolution in a style of eq (7) for higher-dimensional data.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603539098203}, {"id": "PHwy6LtjTrh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1691/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method for estimating the probability deinsity distribution on a low-dimensional manifold embedded in a high-dimensional space using Normalizing Flow (NF). The problem is that the universality of NF is limited because low-dimensional manifolds are not diffeomorphic with respect to high-dimensional Euclidean space. The proposal of this study is to make NF applicable by inflating low-dimensional manifolds with Gaussian noise. Then, after the transformation is obtained, the probability distribution on the original low-dimensional manifold can be obtained by deflation.\n\nThere is a detailed discussion about the addition of Gaussian noise. Theoretically, any Gaussian noise $ e $ at coordinate x can be decomposed into projections on tangent space $ T_x $ and normal space $ N_x $ as $ e = e_t + e_n $. Only normal noise $ e_n $ is ideal to add. However, it is not realistic to find this at each coordinate x. This paper argues that Gaussian noise $ e $ is a good approximation of $ e_n $ when the manifold dimension $ d $ is sufficiently smaller than the higher-dimensional Euclidean space dimension $ D $. This also argues that the noise variance $ \\ sigma $ at that time should be set according to the inverse of the density. \n\nIn computer experiments, simple simulation results using the von mises distribution on a circle and a sphere are shown.\n\nThe gist of the manuscript is well-written, and the issues it deals with are also important and interesting.\nTheoretically, interesting discussions are being developed, but discussions and experimental results are somewhat weak regarding the merits of practical application.\n\nI have some questions.\n1. I couldn't understand the description about scalability well, \"our method scales to high dimensions because it is based on one NF and does not require to compute its inverse.\" Isn't the scalability the same as normal NF because the proposed method basically uses normal NF?\n\n2. In this research, Authors propose to add noise to the data sampled from low-dimensional manifolds in a pseudo manner, but in actual measurement, noise has been included in data naturally without adding it in a pseudo manner. Is it necessary to add pseudo noise even when applying it to real world data?\n\n3. The denoising auto-encoder is famous as a manifold learning method that adds pseudo noise. Is there any relation to this? Also, I would like to know if there are any advantages of the proposed method over the denoising auto-encoder.\n\n4. VAE is a well-known method for finding the mapping of low-dimensional space to the normal distribution. Is there any relation to this? Also, I would like to know if there are any advantages of the proposed method over VAE.\n\n5. In relation to the above, it may be good to have a comparison experiment with denoising auto-encoder and VAE.\n\n6. The lack of experimental results with actual data makes that the paper is unconvincing. Especially for the description of scalability, it is better to prove it experimentally.\n\n7. This is a pure question, is it possible to know the dimensions of the manifold through this technique?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting but not enough comparison with related works", "review": "This paper proposes a method for estimating the probability deinsity distribution on a low-dimensional manifold embedded in a high-dimensional space using Normalizing Flow (NF). The problem is that the universality of NF is limited because low-dimensional manifolds are not diffeomorphic with respect to high-dimensional Euclidean space. The proposal of this study is to make NF applicable by inflating low-dimensional manifolds with Gaussian noise. Then, after the transformation is obtained, the probability distribution on the original low-dimensional manifold can be obtained by deflation.\n\nThere is a detailed discussion about the addition of Gaussian noise. Theoretically, any Gaussian noise $ e $ at coordinate x can be decomposed into projections on tangent space $ T_x $ and normal space $ N_x $ as $ e = e_t + e_n $. Only normal noise $ e_n $ is ideal to add. However, it is not realistic to find this at each coordinate x. This paper argues that Gaussian noise $ e $ is a good approximation of $ e_n $ when the manifold dimension $ d $ is sufficiently smaller than the higher-dimensional Euclidean space dimension $ D $. This also argues that the noise variance $ \\ sigma $ at that time should be set according to the inverse of the density. \n\nIn computer experiments, simple simulation results using the von mises distribution on a circle and a sphere are shown.\n\nThe gist of the manuscript is well-written, and the issues it deals with are also important and interesting.\nTheoretically, interesting discussions are being developed, but discussions and experimental results are somewhat weak regarding the merits of practical application.\n\nI have some questions.\n1. I couldn't understand the description about scalability well, \"our method scales to high dimensions because it is based on one NF and does not require to compute its inverse.\" Isn't the scalability the same as normal NF because the proposed method basically uses normal NF?\n\n2. In this research, Authors propose to add noise to the data sampled from low-dimensional manifolds in a pseudo manner, but in actual measurement, noise has been included in data naturally without adding it in a pseudo manner. Is it necessary to add pseudo noise even when applying it to real world data?\n\n3. The denoising auto-encoder is famous as a manifold learning method that adds pseudo noise. Is there any relation to this? Also, I would like to know if there are any advantages of the proposed method over the denoising auto-encoder.\n\n4. VAE is a well-known method for finding the mapping of low-dimensional space to the normal distribution. Is there any relation to this? Also, I would like to know if there are any advantages of the proposed method over VAE.\n\n5. In relation to the above, it may be good to have a comparison experiment with denoising auto-encoder and VAE.\n\n6. The lack of experimental results with actual data makes that the paper is unconvincing. Especially for the description of scalability, it is better to prove it experimentally.\n\n7. This is a pure question, is it possible to know the dimensions of the manifold through this technique?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603418481118}, {"id": "Rooj6vKrGUx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1691/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors address main limitations of Normalizing Flows (NFs) method for estimation of density functions on manifolds. Since NFs requires the support of density function to cover the whole Euclidean space, they propose to add noise (inflate) to apply NF. The paper is nicely written with clear introduction of basic concepts very useful for non-expert readers. They provide theoretical guarantees on the variance and type of added noise that make the method work and illustrate with synthetic experiments.\nMy only concern is about the applicability and usefulness on real data and specific machine learning. I have found some issues and also have some questions. See below:\nMajor issues:\n-\tHow this method can be used for a real problem with real datasets. Could you please provide some example or give some guidance about the potential of this method for the machine learning community?\n-\tIt is shown that when D>>d, Gaussian noise is an excellent approximation of noise restricted in the normal direction and the experiments seem to confirm this. What should be the criterion to say that D is large enough to make this approximation useful? Can you give some way to test it in practice?\n-\tExample 2: a graphical description of this example would help, for example, showing that variance of noise can help to make the Q-normally separated condition to be met.\n-\tPage 5: I didn’t understand the following sentence “A potential future avenue is to \\sigma^2A and this reach number”. Could you please elaborate it?\n\nMinor issues\n-\tIn Example 1: “an z” -> “a z”\n-\tReferences have missing information. For example, for Cornish et al, Gemici et al and Rezende et al papers, there is not information about the source. Did they works published in conferences, journals or ArXiV?\n-\tPage 12: Please correct the number of table in “Table ??” \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice theoretical work", "review": "In this paper, the authors address main limitations of Normalizing Flows (NFs) method for estimation of density functions on manifolds. Since NFs requires the support of density function to cover the whole Euclidean space, they propose to add noise (inflate) to apply NF. The paper is nicely written with clear introduction of basic concepts very useful for non-expert readers. They provide theoretical guarantees on the variance and type of added noise that make the method work and illustrate with synthetic experiments.\nMy only concern is about the applicability and usefulness on real data and specific machine learning. I have found some issues and also have some questions. See below:\nMajor issues:\n-\tHow this method can be used for a real problem with real datasets. Could you please provide some example or give some guidance about the potential of this method for the machine learning community?\n-\tIt is shown that when D>>d, Gaussian noise is an excellent approximation of noise restricted in the normal direction and the experiments seem to confirm this. What should be the criterion to say that D is large enough to make this approximation useful? Can you give some way to test it in practice?\n-\tExample 2: a graphical description of this example would help, for example, showing that variance of noise can help to make the Q-normally separated condition to be met.\n-\tPage 5: I didn’t understand the following sentence “A potential future avenue is to \\sigma^2A and this reach number”. Could you please elaborate it?\n\nMinor issues\n-\tIn Example 1: “an z” -> “a z”\n-\tReferences have missing information. For example, for Cornish et al, Gemici et al and Rezende et al papers, there is not information about the source. Did they works published in conferences, journals or ArXiV?\n-\tPage 12: Please correct the number of table in “Table ??” \n", "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603113978091}], "openreview_url": "https://openreview.net/forum?id=PuG6vCSbrV9", "arxiv_id": "2105.12152", "paper_pdf": "papers/PuG6vCSbrV9.pdf", "paper_pdf_sha256": "7a115a765945b97bdbbf087ea67c9bd19b97b1864968728e32feb7cc8beb074a", "paper_pdf_bytes": 2716141, "paper_pdf_source": "openreview", "code_url": "https://github.com/chrvt/Inflation-Deflation", "code_repository": "chrvt/Inflation-Deflation", "code_commit": "e19dc603e711df32821f4380acefc88ff64645a4", "code_archive": "repos/PuG6vCSbrV9.zip", "code_archive_sha256": "58ca7ace542656368cb74ed32ff54e9f3d158d91a314adb237b4500ff58fba94", "code_archive_bytes": 233213, "code_file_count": 36, "code_extensions": {".py": 32, ".sh": 4}, "github_disk_usage_kb": 332, "github_languages": {"Python": 284452, "Shell": 4811}, "github_archived": false, "github_pushed_at": "2022-02-02T14:22:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/density-estimation-an-inflation-deflation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1xq264YvH", "year": 2020, "status": "rejected", "title": "Encoder-Agnostic Adaptation for Conditional Language Generation", "authors": ["Zachary M. Ziegler", "Luke Melas-Kyriazi", "Sebastian Gehrmann", "Alexander M. Rush"], "authorids": ["zziegler@g.harvard.edu", "lmelaskyriazi@college.harvard.edu", "gehrmann@seas.harvard.edu", "srush@seas.harvard.edu"], "authors_source": "OpenReview API", "abstract": "Large pretrained language models have changed the way researchers approach discriminative natural language understanding tasks, leading to the dominance of approaches that adapt a pretrained model for arbitrary downstream tasks. However, it is an open question how to use similar techniques for language generation. Early results in the encoder-agnostic setting have been mostly negative. In this work, we explore methods for adapting a pretrained language model to arbitrary conditional input. We observe that pretrained transformer models are sensitive to large parameter changes during tuning. Therefore, we propose an adaptation that directly injects arbitrary conditioning into self attention, an approach we call pseudo self attention. Through experiments on four diverse conditional text generation tasks, we show that this encoder-agnostic technique outperforms strong baselines, produces coherent generations, and is data-efficient.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJg1WdHqqr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper791/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper compares a few encoder agnostic methods for using pretrained decoders in text generation tax. The author compared a few intuitive ways of doing this, and presents results showing that that pseudo-self attention does the best.\n\nHowever, I think the results has some strange points that needs further investigation. Going from repr-transfomer to context-attention to pseudo-self, there is an increasing amount of parameters initialized by pretraining. However, both of the first two methods often perform worse than the baseline transformer without pretraining. So should more things be initialized with pre-training or less? It would be good to verify that this is not due to under-training. \n\nExcept paragraph captioning, the results on other tasks are not better than prior results, which do not use pretraining. The baseline transformer is also usually worse than prior results. The human evaluation shows that the proposed method do better on story generation, but this one is essentially text to text. What is missing is how this compares with even more pretraining, say GPT-2, without any fine tuning. \n\nTransferring gains of pretraining to generation tasks is clearly a promising direction, and the bar for success in this area need to be outperforming the best previous methods that do not use pretraining.  There is no comparison with previous text 2 text methods that use pretraining.  If the proposed methods are truely encoder agnostic, then they should perform reasonably on text-to-text as well. I think some MT experiments would be good since the evaluations are more competitive and reliable. Perhaps using some language pairs that do not have sufficient training data.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper compares a few encoder agnostic methods for using pretrained decoders in text generation tax. The author compared a few intuitive ways of doing this, and presents results showing that that pseudo-self attention does the best.\n\nHowever, I think the results has some strange points that needs further investigation. Going from repr-transfomer to context-attention to pseudo-self, there is an increasing amount of parameters initialized by pretraining. However, both of the first two methods often perform worse than the baseline transformer without pretraining. So should more things be initialized with pre-training or less? It would be good to verify that this is not due to under-training. \n\nExcept paragraph captioning, the results on other tasks are not better than prior results, which do not use pretraining. The baseline transformer is also usually worse than prior results. The human evaluation shows that the proposed method do better on story generation, but this one is essentially text to text. What is missing is how this compares with even more pretraining, say GPT-2, without any fine tuning. \n\nTransferring gains of pretraining to generation tasks is clearly a promising direction, and the bar for success in this area need to be outperforming the best previous methods that do not use pretraining.  There is no comparison with previous text 2 text methods that use pretraining.  If the proposed methods are truely encoder agnostic, then they should perform reasonably on text-to-text as well. I think some MT experiments would be good since the evaluations are more competitive and reliable. Perhaps using some language pairs that do not have sufficient training data.  "}, "tcdate": 1572653046863}, {"id": "HyeNNy_Ttr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper791/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new architecture to train decoder models on language generation using a pre-trained encoder (such as BERT or GPT-2). They introduce a novel block called `````\"pseudo self-attention\" that allow injecting the input for conditional generation in the self-attention layer (i.e. softmax of YW_q (XU_k | YW_k)^T (XU_v | YW_v) instead of softmax(YW_q(YW_k)^T)YW_v). They extensively evaluate their approach on a large set of tasks showing improvements across all of them (which includes class-conditional generation, summarization, story generation and paragraph generation). They also provide interesting ablation studies.\n\nThis paper proposes a simple architectural block to try and translate the success of large pre-trained encoders on discriminative tasks to the generative setting. The idea seems well-motivated and the paper is well-written and easy to follow. The experimental section is very thorough and show large improvements on a variety of task---I particularly appreciate that they experimented with conditional inputs of different nature (class value, image, different languages etc...) to show the effectiveness of their method.\n\nOverall, while the idea is quite simple, the experiments speak for themselves and this could prove to be a useful `layer' to use on large pre-trained language models. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a new architecture to train decoder models on language generation using a pre-trained encoder (such as BERT or GPT-2). They introduce a novel block called `````\"pseudo self-attention\" that allow injecting the input for conditional generation in the self-attention layer (i.e. softmax of YW_q (XU_k | YW_k)^T (XU_v | YW_v) instead of softmax(YW_q(YW_k)^T)YW_v). They extensively evaluate their approach on a large set of tasks showing improvements across all of them (which includes class-conditional generation, summarization, story generation and paragraph generation). They also provide interesting ablation studies.\n\nThis paper proposes a simple architectural block to try and translate the success of large pre-trained encoders on discriminative tasks to the generative setting. The idea seems well-motivated and the paper is well-written and easy to follow. The experimental section is very thorough and show large improvements on a variety of task---I particularly appreciate that they experimented with conditional inputs of different nature (class value, image, different languages etc...) to show the effectiveness of their method.\n\nOverall, while the idea is quite simple, the experiments speak for themselves and this could prove to be a useful `layer' to use on large pre-trained language models. "}, "tcdate": 1571811115896}, {"id": "BygEgCVoYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper791/AnonReviewer1"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a simple yet effective method to adapt large-scale pre-trained language models, which have been shown to substantially improve performance on broadly classification-based NLU tasks, to NLG. The approach is explored in the encoder-agnostic {X}-to-text setup, where the source encoding {X} could represent arbitrary modalities, such as text or images.\n\nMore concretely, the paper leverages a pre-trained, large-scale language model (in this case a GPT-2), and examines how to best cast such unconditional language model into a decoder that generates text conditional on the source information {X}. As self-attention inherently works with sequences of any length, the proposed pseudo self-attention approach simply injects the encoder representation as additional conditioning context (using some additional projection matrices that are learned from scratch) into the pre-trained self-attention layers of the decoder. Extensive experiments and analysis on four diverse tasks demonstrate that pseudo self-attention generally outperforms two other ways of pre-training the decoder, improves NLG data efficiency, and produces texts that are judged more favourably by human evaluators.\n\nOverall, this paper presents a simple, general, and effective method for adapting large-scale pre-trained language models to conditional text generation. Based on the pros and cons that I have listed below, I am giving a rating of \"Accept\". I hope that some of my concerns will be addressed in the authors' response.\n\nPros:\n1. The paper is well-written and the methodology is explained very clearly. Figure 1 is particularly helpful in illustrating the differences between pseudo self-attention and the baselines.\n\n2. The paper addresses a very important problem, and helps make sure that the advances that have been made in language modelling (which can leverage large amounts of unlabelled data), would transfer well to conditional text generation tasks, which hold immediate practical value yet often require expensive annotations.\n\n3. The approach is simple and easy-to-implement, but has been shown to be effective across a broad range of problems, multiple modalities, and various evaluation metric. \n\n4. The paper features extensive reference to relevant prior work, and clearly highlights the key similarities and differences with prior approaches. \n\nCons:\n1. It is still unclear how using language model pre-training affects adequacy (as opposed to fluency). The paper shows that using pseudo self-attention results in a decoder that diverges less from its language model initialisation. One potential risk is that the decoder may prefer fluent, \"safe\" outputs (which is arguably what a language model would prefer since it is an unconditional model) that are nevertheless less faithful to the source information. Since none of the evaluation metric specifically assesses for adequacy on its own, it would be good to isolate the effect of pseudo self-attention on adequacy, and compare it with the baselines, in addition to a Transformer trained from scratch on each downstream task. How to measure adequacy is naturally still an open question, but there are a few things that can be done (e.g. recall of salient information, reverse perplexity to see how much of the source information can be \"reconstructed\" given the predicted target text, etc.).\n\n2. It would be interesting to further examine the interaction between encoder pre-training and the decoder pre-training that is explored in this work. Another interesting experiment to run is whether end-to-end training (including fine-tuning the encoder) would help, since prior work has shown the benefits of end-to-end learning (at least when large amounts of data are available).\n\nQuestions:\n1. Why is the context-attn model performance not included in Table 6? Is it because of the optimisation issue associated with that model in Table 3?\n\n2. In page 7, it is mentioned that \"Both models have similar attention distributions ... at the first layer, which precedes the introduction of new parameters in both models\". Does the first layer here refer to the token + position embedding layer?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "8: Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper proposes a simple yet effective method to adapt large-scale pre-trained language models, which have been shown to substantially improve performance on broadly classification-based NLU tasks, to NLG. The approach is explored in the encoder-agnostic {X}-to-text setup, where the source encoding {X} could represent arbitrary modalities, such as text or images.\n\nMore concretely, the paper leverages a pre-trained, large-scale language model (in this case a GPT-2), and examines how to best cast such unconditional language model into a decoder that generates text conditional on the source information {X}. As self-attention inherently works with sequences of any length, the proposed pseudo self-attention approach simply injects the encoder representation as additional conditioning context (using some additional projection matrices that are learned from scratch) into the pre-trained self-attention layers of the decoder. Extensive experiments and analysis on four diverse tasks demonstrate that pseudo self-attention generally outperforms two other ways of pre-training the decoder, improves NLG data efficiency, and produces texts that are judged more favourably by human evaluators.\n\nOverall, this paper presents a simple, general, and effective method for adapting large-scale pre-trained language models to conditional text generation. Based on the pros and cons that I have listed below, I am giving a rating of \"Accept\". I hope that some of my concerns will be addressed in the authors' response.\n\nPros:\n1. The paper is well-written and the methodology is explained very clearly. Figure 1 is particularly helpful in illustrating the differences between pseudo self-attention and the baselines.\n\n2. The paper addresses a very important problem, and helps make sure that the advances that have been made in language modelling (which can leverage large amounts of unlabelled data), would transfer well to conditional text generation tasks, which hold immediate practical value yet often require expensive annotations.\n\n3. The approach is simple and easy-to-implement, but has been shown to be effective across a broad range of problems, multiple modalities, and various evaluation metric. \n\n4. The paper features extensive reference to relevant prior work, and clearly highlights the key similarities and differences with prior approaches. \n\nCons:\n1. It is still unclear how using language model pre-training affects adequacy (as opposed to fluency). The paper shows that using pseudo self-attention results in a decoder that diverges less from its language model initialisation. One potential risk is that the decoder may prefer fluent, \"safe\" outputs (which is arguably what a language model would prefer since it is an unconditional model) that are nevertheless less faithful to the source information. Since none of the evaluation metric specifically assesses for adequacy on its own, it would be good to isolate the effect of pseudo self-attention on adequacy, and compare it with the baselines, in addition to a Transformer trained from scratch on each downstream task. How to measure adequacy is naturally still an open question, but there are a few things that can be done (e.g. recall of salient information, reverse perplexity to see how much of the source information can be \"reconstructed\" given the predicted target text, etc.).\n\n2. It would be interesting to further examine the interaction between encoder pre-training and the decoder pre-training that is explored in this work. Another interesting experiment to run is whether end-to-end training (including fine-tuning the encoder) would help, since prior work has shown the benefits of end-to-end learning (at least when large amounts of data are available).\n\nQuestions:\n1. Why is the context-attn model performance not included in Table 6? Is it because of the optimisation issue associated with that model in Table 3?\n\n2. In page 7, it is mentioned that \"Both models have similar attention distributions ... at the first layer, which precedes the introduction of new parameters in both models\". Does the first layer here refer to the token + position embedding layer?"}, "tcdate": 1571667436008}], "openreview_url": "https://openreview.net/forum?id=B1xq264YvH", "arxiv_id": "1908.06938", "paper_pdf": "papers/B1xq264YvH.pdf", "paper_pdf_sha256": "a2146da14ac4b2efe1b88056684a596e0ce69d19cd6fee4a291be2d955a36989", "paper_pdf_bytes": 371128, "paper_pdf_source": "openreview", "code_url": "https://github.com/harvardnlp/encoder-agnostic-adaptation", "code_repository": "harvardnlp/encoder-agnostic-adaptation", "code_commit": "5eff09874f25ac256f07daa0d3b9e7c03705086f", "code_archive": "repos/B1xq264YvH.zip", "code_archive_sha256": "832a1b98d5476b86b686f762ed17b41ce1e02fde3b6acd47b96a28b7b7ad65f5", "code_archive_bytes": 492951, "code_file_count": 100, "code_extensions": {".py": 96, ".sh": 3, ".ipynb": 1}, "github_disk_usage_kb": 449, "github_languages": {"Python": 656290, "Shell": 23927}, "github_archived": false, "github_pushed_at": "2024-07-25T10:15:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/encoder-agnostic-adaptation-for-conditional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OqVbYxDBHV", "year": 2026, "status": "rejected", "title": "Where do Large Vision-Language Models Look at when Answering Questions?", "authors": ["Xiaoying Xing", "Chia-Wen Kuo", "Li Fuxin", "Yulei Niu", "Fan Chen", "Ming Li", "Lianghui Wang", "Ying Wu", "Longyin Wen", "Sijie Zhu"], "authorids": ["~Xiaoying_Xing1", "~Chia-Wen_Kuo1", "~Li_Fuxin1", "~Yulei_Niu1", "~Fan_Chen5", "~Ming_Li19", "~Lianghui_Wang1", "~Ying_Wu7", "~Longyin_Wen1", "~Sijie_Zhu1"], "authors_source": "OpenReview API", "abstract": "Large Vision-Language Models (LVLMs) have shown promising performance in vision-language understanding and reasoning tasks. However, whether they truly understand the input image remain underexplored. A fundamental question arises: to what extent do LVLMs rely on visual input, and which image regions contribute to their responses? It is non-trivial to interpret the free-form generation of LVLMs due to their complicated visual architecture (e.g., multiple encoders and multi-resolution) and variable-length outputs. In this paper, we extend existing heatmap visualization methods for classification tasks to support LVLMs for open-ended visual question answering. We propose a method to select visually relevant tokens that reflect the relevance between generated responses and the input image.\nFurthermore, we conduct a comprehensive analysis of state-of-the-art LVLMs on benchmarks designed to require visual information to answer. Our findings offer several insights into LVLM behavior, including the relationship between focus region and answer correctness, differences in visual attention across architectures, and the impact of LLM scale on visual understanding. The code and data will be\nreleased", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jVHYtkQpAp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22072/Reviewer_eJZt"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper dives into a key question about large vision-language models: where do they focus on images when answering questions? Since existing heatmap visualization methods mostly work for classification tasks, the authors extended these tools to handle VLMs' variable-length, autoregressive outputs. The authors propose a method to identify \"visually relevant tokens\" from model responses. Using this method, they tested state-of-the-art LVLMs like LLaVA-1.5, LLaVA-OV, and Cambrian on vision-centric datasets. The results showed that these models do rely on images for over 75% of cases, different vision architectures lead to distinct attention patterns, and simply making the LLM bigger doesn’t change how the model looks at images much. They also found that even when models get answers wrong, their focus regions often relate to the question, and sometimes, correct answers come from irrelevant image parts. These findings further highlight generalization challenges.", "review_text": "This paper dives into a key question about large vision-language models: where do they focus on images when answering questions? Since existing heatmap visualization methods mostly work for classification tasks, the authors extended these tools to handle VLMs' variable-length, autoregressive outputs. The authors propose a method to identify \"visually relevant tokens\" from model responses. Using this method, they tested state-of-the-art LVLMs like LLaVA-1.5, LLaVA-OV, and Cambrian on vision-centric datasets. The results showed that these models do rely on images for over 75% of cases, different vision architectures lead to distinct attention patterns, and simply making the LLM bigger doesn’t change how the model looks at images much. They also found that even when models get answers wrong, their focus regions often relate to the question, and sometimes, correct answers come from irrelevant image parts. These findings further highlight generalization challenges.", "strengths": "1. The proposed method fixes a gap for interpreting VLMs at the sentence level by filtering out the language stuff that doesn’t need images; the heatmaps make way more sense. \n\n2. The authors conduct experiments on different LVLMs with different setups and multiple vision-heavy datasets. These experiments make their conclusion more reliable. \n\n3. The generalization of the proposed method works with different heatmap tools. This makes it easy for others to follow.", "weaknesses": "1. Although the authors provide several models in the experiment part, they lack the results on the frontier open-source model, like Qwen3-VL. Showing that these models maintain the same property as previous models is important for others to understand these models. Also, results on larger models should be discussed.\n\n2. The authors mainly provide results on some perception benchmarks. I would also like to know the phenomenon of these models on some reasoning benchmarks and high-resolution perception benchmarks. Incorporating these results will strengthen the overall paper.\n\n3. More analysis of visual encoder should be included. New visual encoders do not use fixed resolution, like in Qwen2-VL and newer model. Will these new models perform differently as they improve visual perception?", "questions": "As stated in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper dives into a key question about large vision-language models: where do they focus on images when answering questions? Since existing heatmap visualization methods mostly work for classification tasks, the authors extended these tools to handle VLMs' variable-length, autoregressive outputs. The authors propose a method to identify \"visually relevant tokens\" from model responses. Using this method, they tested state-of-the-art LVLMs like LLaVA-1.5, LLaVA-OV, and Cambrian on vision-centric datasets. The results showed that these models do rely on images for over 75% of cases, different vision architectures lead to distinct attention patterns, and simply making the LLM bigger doesn’t change how the model looks at images much. They also found that even when models get answers wrong, their focus regions often relate to the question, and sometimes, correct answers come from irrelevant image parts. These findings further highlight generalization challenges.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The proposed method fixes a gap for interpreting VLMs at the sentence level by filtering out the language stuff that doesn’t need images; the heatmaps make way more sense. \n\n2. The authors conduct experiments on different LVLMs with different setups and multiple vision-heavy datasets. These experiments make their conclusion more reliable. \n\n3. The generalization of the proposed method works with different heatmap tools. This makes it easy for others to follow.", "weaknesses": "1. Although the authors provide several models in the experiment part, they lack the results on the frontier open-source model, like Qwen3-VL. Showing that these models maintain the same property as previous models is important for others to understand these models. Also, results on larger models should be discussed.\n\n2. The authors mainly provide results on some perception benchmarks. I would also like to know the phenomenon of these models on some reasoning benchmarks and high-resolution perception benchmarks. Incorporating these results will strengthen the overall paper.\n\n3. More analysis of visual encoder should be included. New visual encoders do not use fixed resolution, like in Qwen2-VL and newer model. Will these new models perform differently as they improve visual perception?", "questions": "As stated in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761995227733}, {"id": "xdgkf8z6aB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22072/Reviewer_crB2"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors propose a method to extend traditional heatmap visualization techniques (e.g., GradCAM, iGOS++) to support Large Vision-Language Models (LVLMs) in open-ended visual question answering tasks. The key contribution is a visually relevant token selection mechanism that identifies tokens most dependent on image content through a log-likelihood ratio (LLR) between visual and blurred-image inputs. This approach enables the generation of interpretable visual heatmaps and facilitates analysis of how LVLMs attend to image regions when producing answers. The authors analyze multiple state-of-the-art LVLMs, like LLaVA-1.5, LLaVA-OneVision, and Cambrian across recent benchmarks including CV-Bench, MMStar, and MMVP.", "review_text": "The authors propose a method to extend traditional heatmap visualization techniques (e.g., GradCAM, iGOS++) to support Large Vision-Language Models (LVLMs) in open-ended visual question answering tasks. The key contribution is a visually relevant token selection mechanism that identifies tokens most dependent on image content through a log-likelihood ratio (LLR) between visual and blurred-image inputs. This approach enables the generation of interpretable visual heatmaps and facilitates analysis of how LVLMs attend to image regions when producing answers. The authors analyze multiple state-of-the-art LVLMs, like LLaVA-1.5, LLaVA-OneVision, and Cambrian across recent benchmarks including CV-Bench, MMStar, and MMVP.", "strengths": "1.\tThe paper is clearly written, well structured, and easy to follow. \n2.\tLVLM interpretability is a critical and underexplored topic, especially considering the success of models such as LLaVA, InstructBLIP, Gemini, and GPT-4o, and the concerns raised by benchmarks like MMVP, MERLIM, AMBER, POPE, and HallusionBench regarding hallucinations and insufficient visual grounding. The proposed tool can help diagnose whether such errors stem from limited grounding capacity or architectural constraints that prevent LVLMs from fully leveraging visual information.\n3.\tThe approach effectively identifies which tokens rely on visual grounding and visualizes how this information is used via interpretable heatmaps. The paper includes several qualitative examples that illustrate the findings, supported by the corresponding mathematical derivations.\n4.\tThe study extends and compares multiple heatmap-based interpretability methods, including GradCAM, T-MM, IIA, and iGOS++, providing a useful empirical reference for future LVLM analysis.", "weaknesses": "1.\tMy main concern is the lack of novelty. Although this paper addrees a relevant topic, the method primarily combines existing components (heatmap visualization + token selection) rather than introducing a fundamentally new paradigm. The LLR-based token selection is a natural extension of standard likelihood analysis, and while the integration is non-trivial, it could be seen as incremental.", "questions": "1. It would be insightful to analyze the visual attention (heatmap) associated with hallucinated or spurious tokens to better understand whether these hallucinations arise from incorrect grounding or over-reliance on textual priors. Could you provide such an analysis?\n2.\tThe hyperparameters α and λ seem potentially architecture-dependent, but the paper does not report how they behave across different LVLMs. Have you studied their sensitivity or generalization across models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a method to extend traditional heatmap visualization techniques (e.g., GradCAM, iGOS++) to support Large Vision-Language Models (LVLMs) in open-ended visual question answering tasks. The key contribution is a visually relevant token selection mechanism that identifies tokens most dependent on image content through a log-likelihood ratio (LLR) between visual and blurred-image inputs. This approach enables the generation of interpretable visual heatmaps and facilitates analysis of how LVLMs attend to image regions when producing answers. The authors analyze multiple state-of-the-art LVLMs, like LLaVA-1.5, LLaVA-OneVision, and Cambrian across recent benchmarks including CV-Bench, MMStar, and MMVP.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tThe paper is clearly written, well structured, and easy to follow. \n2.\tLVLM interpretability is a critical and underexplored topic, especially considering the success of models such as LLaVA, InstructBLIP, Gemini, and GPT-4o, and the concerns raised by benchmarks like MMVP, MERLIM, AMBER, POPE, and HallusionBench regarding hallucinations and insufficient visual grounding. The proposed tool can help diagnose whether such errors stem from limited grounding capacity or architectural constraints that prevent LVLMs from fully leveraging visual information.\n3.\tThe approach effectively identifies which tokens rely on visual grounding and visualizes how this information is used via interpretable heatmaps. The paper includes several qualitative examples that illustrate the findings, supported by the corresponding mathematical derivations.\n4.\tThe study extends and compares multiple heatmap-based interpretability methods, including GradCAM, T-MM, IIA, and iGOS++, providing a useful empirical reference for future LVLM analysis.", "weaknesses": "1.\tMy main concern is the lack of novelty. Although this paper addrees a relevant topic, the method primarily combines existing components (heatmap visualization + token selection) rather than introducing a fundamentally new paradigm. The LLR-based token selection is a natural extension of standard likelihood analysis, and while the integration is non-trivial, it could be seen as incremental.", "questions": "1. It would be insightful to analyze the visual attention (heatmap) associated with hallucinated or spurious tokens to better understand whether these hallucinations arise from incorrect grounding or over-reliance on textual priors. Could you provide such an analysis?\n2.\tThe hyperparameters α and λ seem potentially architecture-dependent, but the paper does not report how they behave across different LVLMs. Have you studied their sensitivity or generalization across models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761852162062}, {"id": "mAk4fec5gi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22072/Reviewer_m8ag"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper adapts existing decision localization/visualization methods (GradCAM, iGOS++) for use with large vision-language models (LVLMs) in open-ended visual question answering, and leverages them to study how LVLMs attend to images when answering questions. Since these methods require a single prediction score, the paper proposes a token selection strategy which selects the tokens with high loglikelihood ratio (normalized by their loglikelihood given an neutral input image), and then computes the sum of the loglikelihood of these selected tokens as the prediction score. The paper also introduces a regularization term to the optimization objective of iGOS++. Several experiments are provided to compare the different decision localization methods, and explore how well LVLMs look at the correct regions in images.", "review_text": "This paper adapts existing decision localization/visualization methods (GradCAM, iGOS++) for use with large vision-language models (LVLMs) in open-ended visual question answering, and leverages them to study how LVLMs attend to images when answering questions. Since these methods require a single prediction score, the paper proposes a token selection strategy which selects the tokens with high loglikelihood ratio (normalized by their loglikelihood given an neutral input image), and then computes the sum of the loglikelihood of these selected tokens as the prediction score. The paper also introduces a regularization term to the optimization objective of iGOS++. Several experiments are provided to compare the different decision localization methods, and explore how well LVLMs look at the correct regions in images.", "strengths": "This paper adapts existing decision localization/visualization methods (GradCAM, iGOS++) for use with large vision-language models (LVLMs) in open-ended visual question answering, and leverages them to study how LVLMs attend to images when answering questions. Since these methods require a single prediction score, the paper proposes a token selection strategy which selects the tokens with high loglikelihood ratio (normalized by their loglikelihood given a neutral input image), and then computes the sum of the loglikelihoods of these selected tokens as the prediction score. The paper also introduces a regularization term to the optimization objective of iGOS++. Several experiments are provided to compare the different decision localization methods and explore how well LVLMs look at the correct regions in images.", "weaknesses": "1. The comparison in Table 2 (and consequently the choice of iGOS++ and deletion/insertion metrics for the rest of the experiments) is incorrect. The problem is that iGOS++ directly optimizes the deletion/insertion metrics on test images, and then compares these values to methods that do not have this privilege (Grad-CAM etc). Note that directly optimizing any metric on test images will result in better values for that metric, but what actually matters is whether any generalization can be achieved beyond the optimized metric. For example, the paper's proposed optimization for the insertion metric could easily discover adversarial patterns on any test image that when inserted on a blurry background image will boost the likelihood of the LLM's prediction. In such cases, the \"adversarial\" interpretation would be of little use/meaning despite the high selection metric score, and so the paper must show improvement on an independent metric (for example object localization metrics, or overall accuracy as in [1]).\n\n2. The paper misses several very relevant published papers in its related works. [1,2] have studied whether LVLMs attend to images. [3,4] have provided methods for decision localization/visualization in LVLMs.\n\n3. The paper’s main findings are already established in previous work (“Q1: Do LVLMs rely on the input image when answering visual questions?” is answered in papers [1,2,3] and “Q2: Where do different LVLMs attend when generating variable-length responses? Q3: What is the relationship between answer correctness and focus region?“ is answered in [3,4]), so the paper’s novelty is limited to its proposed token selection method.\n\n4. The paper provides no statistical confidence intervals for the metrics, so it is unclear how meaningful the small differences are. Some p-values are reported in the last section (lines 470-473), but the paper does not discuss how these values are computed (what statistical test, sample size, etc).\n\n[1] The Instinctive Bias: Spurious Images lead to Illusion in MLLMs. EMNLP 2024.\n\n[2] Unveiling the Ignorance of MLLMs: Seeing Clearly, Answering Incorrectly. CVPR 2025.\n\n[3] MLLMs Know Where to Look: Training-Free Perception of Small Visual Details with Multimodal LLMs. ICLR 2025.\n\n[4] V*: Guided Visual Search as A Core Mechanism in Multimodal LLMs. CVPR 2024.", "questions": "1. I suggest using a simpler object detection or VQA metric for comparing and validating the correctness of the localization methods.\n2. Clarify novelty compared to several related published papers mentioned in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper adapts existing decision localization/visualization methods (GradCAM, iGOS++) for use with large vision-language models (LVLMs) in open-ended visual question answering, and leverages them to study how LVLMs attend to images when answering questions. Since these methods require a single prediction score, the paper proposes a token selection strategy which selects the tokens with high loglikelihood ratio (normalized by their loglikelihood given an neutral input image), and then computes the sum of the loglikelihood of these selected tokens as the prediction score. The paper also introduces a regularization term to the optimization objective of iGOS++. Several experiments are provided to compare the different decision localization methods, and explore how well LVLMs look at the correct regions in images.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "This paper adapts existing decision localization/visualization methods (GradCAM, iGOS++) for use with large vision-language models (LVLMs) in open-ended visual question answering, and leverages them to study how LVLMs attend to images when answering questions. Since these methods require a single prediction score, the paper proposes a token selection strategy which selects the tokens with high loglikelihood ratio (normalized by their loglikelihood given a neutral input image), and then computes the sum of the loglikelihoods of these selected tokens as the prediction score. The paper also introduces a regularization term to the optimization objective of iGOS++. Several experiments are provided to compare the different decision localization methods and explore how well LVLMs look at the correct regions in images.", "weaknesses": "1. The comparison in Table 2 (and consequently the choice of iGOS++ and deletion/insertion metrics for the rest of the experiments) is incorrect. The problem is that iGOS++ directly optimizes the deletion/insertion metrics on test images, and then compares these values to methods that do not have this privilege (Grad-CAM etc). Note that directly optimizing any metric on test images will result in better values for that metric, but what actually matters is whether any generalization can be achieved beyond the optimized metric. For example, the paper's proposed optimization for the insertion metric could easily discover adversarial patterns on any test image that when inserted on a blurry background image will boost the likelihood of the LLM's prediction. In such cases, the \"adversarial\" interpretation would be of little use/meaning despite the high selection metric score, and so the paper must show improvement on an independent metric (for example object localization metrics, or overall accuracy as in [1]).\n\n2. The paper misses several very relevant published papers in its related works. [1,2] have studied whether LVLMs attend to images. [3,4] have provided methods for decision localization/visualization in LVLMs.\n\n3. The paper’s main findings are already established in previous work (“Q1: Do LVLMs rely on the input image when answering visual questions?” is answered in papers [1,2,3] and “Q2: Where do different LVLMs attend when generating variable-length responses? Q3: What is the relationship between answer correctness and focus region?“ is answered in [3,4]), so the paper’s novelty is limited to its proposed token selection method.\n\n4. The paper provides no statistical confidence intervals for the metrics, so it is unclear how meaningful the small differences are. Some p-values are reported in the last section (lines 470-473), but the paper does not discuss how these values are computed (what statistical test, sample size, etc).\n\n[1] The Instinctive Bias: Spurious Images lead to Illusion in MLLMs. EMNLP 2024.\n\n[2] Unveiling the Ignorance of MLLMs: Seeing Clearly, Answering Incorrectly. CVPR 2025.\n\n[3] MLLMs Know Where to Look: Training-Free Perception of Small Visual Details with Multimodal LLMs. ICLR 2025.\n\n[4] V*: Guided Visual Search as A Core Mechanism in Multimodal LLMs. CVPR 2024.", "questions": "1. I suggest using a simpler object detection or VQA metric for comparing and validating the correctness of the localization methods.\n2. Clarify novelty compared to several related published papers mentioned in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761790022841}, {"id": "ktKRnfAGTE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22072/Reviewer_b64Z"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces a general, architecture-aware interpretability framework for LVLMs generation. It first identifies vision-relevant tokens in the generated answer via a log-likelihood ratio between the real image and a blurred “no-vision” baseline, then optimizes a single image-space mask for heatmap. Evaluated with insertion/deletion metrics across several benchmarks and model families, the method yields stronger, more faithful attributions and reveals several interesting insights.", "review_text": "The paper introduces a general, architecture-aware interpretability framework for LVLMs generation. It first identifies vision-relevant tokens in the generated answer via a log-likelihood ratio between the real image and a blurred “no-vision” baseline, then optimizes a single image-space mask for heatmap. Evaluated with insertion/deletion metrics across several benchmarks and model families, the method yields stronger, more faithful attributions and reveals several interesting insights.", "strengths": "1. Empirical study is thorough. Strong baselines, clearly defined insertion/deletion metrics, and targeted ablations that convincingly attribute gains to the proposed components.\n\n2. Architecture-aware methodology. Unified pre-encoder masking and differentiable multi-resolution handling, with a single image-space iGOS++ mask that works across LVLM variants.\n\n3. The insights are quite provocative with research motivation and questions clearly defined and solved. Pattern analysis among different LVLMs are also discussed.", "weaknesses": "> Experiments\n\nThe paper relies on a single metric family (insertion/deletion). Please clarify its robustness—in particular, sensitivity to sampling, baseline type, etc. Thus, Table 2 should report mean and standard deviation to prove superiority.\n\nIn addition, is this evaluation metric causal? By this I mean if it's possible to add some experiments to show what parts are actually causally emphasized in LVLM generation?\n\nAll current comparative baselines treat the LVLM as a white box. It would strengthen the evaluation to include black-box, perturbation-based baselines (e.g., D-RISE [1])\n\n[1] Petsiuk, Vitali, et al. \"Black-box explanation of object detectors via saliency maps.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.", "questions": "1. Why the result in Appendix A.2 is not aligned with human attention? It seems not intuitive.\n\n2. Given your finding that visual grounding does not improve with larger LLMs, what method can solve this issue?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a general, architecture-aware interpretability framework for LVLMs generation. It first identifies vision-relevant tokens in the generated answer via a log-likelihood ratio between the real image and a blurred “no-vision” baseline, then optimizes a single image-space mask for heatmap. Evaluated with insertion/deletion metrics across several benchmarks and model families, the method yields stronger, more faithful attributions and reveals several interesting insights.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. Empirical study is thorough. Strong baselines, clearly defined insertion/deletion metrics, and targeted ablations that convincingly attribute gains to the proposed components.\n\n2. Architecture-aware methodology. Unified pre-encoder masking and differentiable multi-resolution handling, with a single image-space iGOS++ mask that works across LVLM variants.\n\n3. The insights are quite provocative with research motivation and questions clearly defined and solved. Pattern analysis among different LVLMs are also discussed.", "weaknesses": "> Experiments\n\nThe paper relies on a single metric family (insertion/deletion). Please clarify its robustness—in particular, sensitivity to sampling, baseline type, etc. Thus, Table 2 should report mean and standard deviation to prove superiority.\n\nIn addition, is this evaluation metric causal? By this I mean if it's possible to add some experiments to show what parts are actually causally emphasized in LVLM generation?\n\nAll current comparative baselines treat the LVLM as a white box. It would strengthen the evaluation to include black-box, perturbation-based baselines (e.g., D-RISE [1])\n\n[1] Petsiuk, Vitali, et al. \"Black-box explanation of object detectors via saliency maps.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.", "questions": "1. Why the result in Appendix A.2 is not aligned with human attention? It seems not intuitive.\n\n2. Given your finding that visual grounding does not improve with larger LLMs, what method can solve this issue?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761493719754}], "openreview_url": "https://openreview.net/forum?id=OqVbYxDBHV", "arxiv_id": "2503.13891", "paper_pdf": "papers/OqVbYxDBHV.pdf", "paper_pdf_sha256": "06f074da3f6a536df15a25854814555101a02218f65bb7015396f1aea73b64c8", "paper_pdf_bytes": 6078471, "paper_pdf_source": "openreview", "code_url": "https://github.com/bytedance/LVLM_Interpretation", "code_repository": "bytedance/LVLM_Interpretation", "code_commit": "378412f64c253aa497bd95e983395aada8940294", "code_archive": "repos/OqVbYxDBHV.zip", "code_archive_sha256": "87cf283127af82913ec33dcbadf9cd66e892bd63b34ed40f1c1715a2153e6254", "code_archive_bytes": 1778635, "code_file_count": 143, "code_extensions": {".py": 142, ".sh": 1}, "github_disk_usage_kb": 1647, "github_languages": {"Python": 1287958, "Shell": 399}, "github_archived": false, "github_pushed_at": "2026-01-07T03:10:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/where-do-large-vision-language-models-look-at"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xreOs2yjqf", "year": 2025, "status": "rejected", "title": "EvalAlign: Supervised Fine-Tuning Multimodal LLMs with Human-Aligned Data for Evaluating Text-to-Image Models", "authors": ["Zhiyu Tan", "xiaomeng yang", "Luozheng Qin", "Mengping Yang", "Cheng Zhang", "Hao Li"], "authorids": ["~Zhiyu_Tan1", "~xiaomeng_yang3", "~Luozheng_Qin1", "~Mengping_Yang2", "~Cheng_Zhang8", "~Hao_Li16"], "authors_source": "OpenReview API", "abstract": "The recent advancements in text-to-image generative models have been remarkable. Yet, the field suffers from a lack of evaluation metrics that accurately reflect the performance of these models, particularly lacking fine-grained metrics that can guide the optimization of the models. In this paper, we propose EvalAlign, a metric characterized by its accuracy, stability, and fine granularity. Our approach leverages the capabilities of Multimodal Large Language Models (MLLMs) pre-trained on extensive data. We develop evaluation protocols that focus on two key dimensions: image faithfulness and text-image alignment. Each protocol comprises a set of detailed, fine-grained instructions linked to specific scoring options, enabling precise manual scoring of the generated images. We supervised fine-tune (SFT) the MLLM to align with human evaluative judgments, resulting in a robust evaluation model. Our evaluation across 24 text-to-image generation models demonstrate that EvalAlign not only provides superior metric stability but also aligns more closely with human preferences than existing metrics, confirming its effectiveness and utility in model assessment. We will make the code, data, and pre-trained models publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "KHkqh06Th8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2053/Reviewer_gtFg"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes EvalAlign, a metric characterized by accuracy, stability, and fine-grainedness. Evaluation on 24 text-to-image generation models shows that EvalAlign is more in line with human preferences than existing metrics and has certain application value in quality assessment.", "review_text": "This paper proposes EvalAlign, a metric characterized by accuracy, stability, and fine-grainedness. Evaluation on 24 text-to-image generation models shows that EvalAlign is more in line with human preferences than existing metrics and has certain application value in quality assessment.", "strengths": "1. This work has fine-grained annotation. Unlike previous datasets, EvalAlign annotates at three levels: animal faces, visibility of hands, and visibility of limbs. This detailed data enables the author to train an effective evaluation model.\n\n2. The author promises open source code. And the experimental details are listed in the supplementary materials, which has strong reproducibility.\n\n3. The writing of this article is quite fluent, and with appropriate illustrations, it is very easy for readers to understand.", "weaknesses": "1. The experimental part of this article has a big problem. Table 3 seems to have done a lot of experiments, but it is actually evaluated at the model level, not the instance level. This is not a challenging task, because everyone knows that PixArt draws well and SD 1.4 draws relatively poorly. Ranking the strengths of 24 models is far less meaningful than scoring a single image, that is, an end-to-end AIGC quality evaluation tool. In other words, which of the two images from the same model has higher quality is more important.\n\n2. This paper only reviews coarse-grained datasets in related work, but does not consider fine-grained datasets. In addition, some AIGC-related dataset such as [1,2,3] was not considered. These datasets have fewer images but more annotations, and each image contains dozens of fine-grained annotations. Since fine-grained annotations are one of the major innovations of this paper, it is not comprehensive to only review coarse-grained datasets (i.e. only two or three annotations, or even less than one per images).\n\n[1] PKU-AIGIQA-4K: A Perceptual Quality Assessment Database for Both Text-to-Image and Image-to-Image AI-Generated Images\n\n[2] PKU-AIGI-500K: A Neural Compression Benchmark And Model for AI-Generated Images\n\n[3] PKU-I2IQA: An image-to-image quality assessment database for ai generated images", "questions": "I am very concerned about Table 8 of this paper. Are the calculated KRCC and PLCC based on the instance level? If it is at the model level, it is recommended that the author modify it according to the content of weakness. If it is indeed at the instance level, I hope the author will focus on the analysis at this step, which is more important than the scores of each model listed in Table 3. Also, the authors can check the Weaknesses, and address them point-by-point in the response, which would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes EvalAlign, a metric characterized by accuracy, stability, and fine-grainedness. Evaluation on 24 text-to-image generation models shows that EvalAlign is more in line with human preferences than existing metrics and has certain application value in quality assessment.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This work has fine-grained annotation. Unlike previous datasets, EvalAlign annotates at three levels: animal faces, visibility of hands, and visibility of limbs. This detailed data enables the author to train an effective evaluation model.\n\n2. The author promises open source code. And the experimental details are listed in the supplementary materials, which has strong reproducibility.\n\n3. The writing of this article is quite fluent, and with appropriate illustrations, it is very easy for readers to understand.", "weaknesses": "1. The experimental part of this article has a big problem. Table 3 seems to have done a lot of experiments, but it is actually evaluated at the model level, not the instance level. This is not a challenging task, because everyone knows that PixArt draws well and SD 1.4 draws relatively poorly. Ranking the strengths of 24 models is far less meaningful than scoring a single image, that is, an end-to-end AIGC quality evaluation tool. In other words, which of the two images from the same model has higher quality is more important.\n\n2. This paper only reviews coarse-grained datasets in related work, but does not consider fine-grained datasets. In addition, some AIGC-related dataset such as [1,2,3] was not considered. These datasets have fewer images but more annotations, and each image contains dozens of fine-grained annotations. Since fine-grained annotations are one of the major innovations of this paper, it is not comprehensive to only review coarse-grained datasets (i.e. only two or three annotations, or even less than one per images).\n\n[1] PKU-AIGIQA-4K: A Perceptual Quality Assessment Database for Both Text-to-Image and Image-to-Image AI-Generated Images\n\n[2] PKU-AIGI-500K: A Neural Compression Benchmark And Model for AI-Generated Images\n\n[3] PKU-I2IQA: An image-to-image quality assessment database for ai generated images", "questions": "I am very concerned about Table 8 of this paper. Are the calculated KRCC and PLCC based on the instance level? If it is at the model level, it is recommended that the author modify it according to the content of weakness. If it is indeed at the instance level, I hope the author will focus on the analysis at this step, which is more important than the scores of each model listed in Table 3. Also, the authors can check the Weaknesses, and address them point-by-point in the response, which would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730718268621}, {"id": "NChJITek8M", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2053/Reviewer_L57V"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper proposes a method to evaluate the consistency of images and texts in the T2I generation process. Compared with other evaluation indicators such as ImageReward and HPS, it has higher consistency with human subjective preferences on 24 T2I models.", "review_text": "This paper proposes a method to evaluate the consistency of images and texts in the T2I generation process. Compared with other evaluation indicators such as ImageReward and HPS, it has higher consistency with human subjective preferences on 24 T2I models.", "strengths": "1. Using LLM to evaluate the quality of LLM is a very creative point. The author evaluates the T2I process through the I2T model, which is a new paradigm.\n2. The experiment is relatively detailed, considering 24 generative models. Multiple dimensions are evaluated.\n3. The illustrations are intuitive and beautiful, and Figure 1 reflects the central idea of ​​the article well.", "weaknesses": "1. The evaluation is done at the **model level**, not the **instance level**. This is a major flaw. In the actual evaluation process, the community is not only concerned with ranking the strength of the T2I model but whether each AIGC image is good enough. As far as I know, AIGC quality asessment [1,2] uses instance-level, because model-level evaluation is not very challenging (see Vbench [3] Figure 4, the correlation with subjective labels can easily reach 0.9). I hope the author can improve this point.\n2. The author considered 24 models. Although the number is large, they are highly homogenized. For example, DeepFloyd IF uses three different conditions, which are considered three models, but they are the same. Including SD 1.4, 1.5, 2.0, and 2.1, the difference in visual effects is quite limited. However, for open-source models such as DALLE 3 and Midjourney, the dataset does not include them. From my subjective perspective, they are still slightly stronger than PixArt, and ignoring them will result in an incomplete dataset.\n\n[1] Depicting Beyond Scores: Advancing Image Quality Assessment through Multi-modal Language Models\n[2] Descriptive Image Quality Assessment in the Wild\n[3] VBench: Comprehensive Benchmark Suite for Video Generative Models", "questions": "I would like to ask how long the author's evaluation takes. In my opinion, evaluation should be a task to assist generation. If the generated model is already large, using an estimator with 34B parameters will cost a lot, but only slightly improve the consistency with human subjective perception. I am not sure if it worth.\nI am happy that the author analyzed the impact of different model sizes on performance, but the impact on time consumption also needs further explanation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to evaluate the consistency of images and texts in the T2I generation process. Compared with other evaluation indicators such as ImageReward and HPS, it has higher consistency with human subjective preferences on 24 T2I models.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Using LLM to evaluate the quality of LLM is a very creative point. The author evaluates the T2I process through the I2T model, which is a new paradigm.\n2. The experiment is relatively detailed, considering 24 generative models. Multiple dimensions are evaluated.\n3. The illustrations are intuitive and beautiful, and Figure 1 reflects the central idea of ​​the article well.", "weaknesses": "1. The evaluation is done at the **model level**, not the **instance level**. This is a major flaw. In the actual evaluation process, the community is not only concerned with ranking the strength of the T2I model but whether each AIGC image is good enough. As far as I know, AIGC quality asessment [1,2] uses instance-level, because model-level evaluation is not very challenging (see Vbench [3] Figure 4, the correlation with subjective labels can easily reach 0.9). I hope the author can improve this point.\n2. The author considered 24 models. Although the number is large, they are highly homogenized. For example, DeepFloyd IF uses three different conditions, which are considered three models, but they are the same. Including SD 1.4, 1.5, 2.0, and 2.1, the difference in visual effects is quite limited. However, for open-source models such as DALLE 3 and Midjourney, the dataset does not include them. From my subjective perspective, they are still slightly stronger than PixArt, and ignoring them will result in an incomplete dataset.\n\n[1] Depicting Beyond Scores: Advancing Image Quality Assessment through Multi-modal Language Models\n[2] Descriptive Image Quality Assessment in the Wild\n[3] VBench: Comprehensive Benchmark Suite for Video Generative Models", "questions": "I would like to ask how long the author's evaluation takes. In my opinion, evaluation should be a task to assist generation. If the generated model is already large, using an estimator with 34B parameters will cost a lot, but only slightly improve the consistency with human subjective perception. I am not sure if it worth.\nI am happy that the author analyzed the impact of different model sizes on performance, but the impact on time consumption also needs further explanation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730654886998}, {"id": "6OHZlAcw8G", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2053/Reviewer_zeBD"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "In short, this paper presents EvalAlign, which collects a human-annotated preference dataset to fine-tune MLLMs to be evaluators for T2I generation. The paper has focused on two dimensions: (1) T2I alignment, (2) faithfulness, which is a well-accepted setting since AGIQA-3K (Li et al, 2023). Overall, the paper is technically sound, but I am a littble bit concerned on some methodology parts. Additionally, discussions for several pioneer works on T2I evaluation and MLLM as scorers are missing.", "review_text": "In short, this paper presents EvalAlign, which collects a human-annotated preference dataset to fine-tune MLLMs to be evaluators for T2I generation. The paper has focused on two dimensions: (1) T2I alignment, (2) faithfulness, which is a well-accepted setting since AGIQA-3K (Li et al, 2023). Overall, the paper is technically sound, but I am a littble bit concerned on some methodology parts. Additionally, discussions for several pioneer works on T2I evaluation and MLLM as scorers are missing.", "strengths": "1. The dataset collection and annotation process is technically sound. The explicit prompting strategy (e.g. `Are there any issues with human face in the image, such as facial distortion, asymmetrical faces, abnormal facial features, unusual expressions in the eyes, etc?`) could be useful and scalable to better baseline MLLMs.\n2. The evaluation part presents a benchmark on models, showing the high-correlation between the proposed scorer and human evaluation, which is good. It would become a useful metric.", "weaknesses": "I have some concerns on the paper.\n\nFirst, I am a bit concerned on how the score is derived. From my current understanding, the final scores are derived from an average of the score outputs of several questions. Is ther `Human` column also obtained by so? If so, this might not be a good enough ground truth.\n\nSecond, well in Sec. 5.1 the author states that the test set images do no overlap with train set ones, they do come from the same 16 generation models. As the final evaluation only shows model-wise ranking consistency, this result might not enough exclude overfitting (e.g. memorizing on model specific styles, etc). I would encourage a further testing on several hold-out T2I generators.\n\nThird, a minor question. Using SFT for LMM to score has been discussed by Q-Align (ICML2024), which finds out using logits are better than using `model.generate()` for scoring. It also has the ability for image faithfulness evaluation, please try to compare with it or discuss with it. Furthermore, for faithfulness evaluation (which is actually image quality, am I right?), the compared baselines are similarity-based metrics (which are, from their design, alignment-related metrics). I would suggest the authors to compare with some baselines related to T2I quality evaluation (inc. Q-Align) in this part.", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In short, this paper presents EvalAlign, which collects a human-annotated preference dataset to fine-tune MLLMs to be evaluators for T2I generation. The paper has focused on two dimensions: (1) T2I alignment, (2) faithfulness, which is a well-accepted setting since AGIQA-3K (Li et al, 2023). Overall, the paper is technically sound, but I am a littble bit concerned on some methodology parts. Additionally, discussions for several pioneer works on T2I evaluation and MLLM as scorers are missing.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The dataset collection and annotation process is technically sound. The explicit prompting strategy (e.g. `Are there any issues with human face in the image, such as facial distortion, asymmetrical faces, abnormal facial features, unusual expressions in the eyes, etc?`) could be useful and scalable to better baseline MLLMs.\n2. The evaluation part presents a benchmark on models, showing the high-correlation between the proposed scorer and human evaluation, which is good. It would become a useful metric.", "weaknesses": "I have some concerns on the paper.\n\nFirst, I am a bit concerned on how the score is derived. From my current understanding, the final scores are derived from an average of the score outputs of several questions. Is ther `Human` column also obtained by so? If so, this might not be a good enough ground truth.\n\nSecond, well in Sec. 5.1 the author states that the test set images do no overlap with train set ones, they do come from the same 16 generation models. As the final evaluation only shows model-wise ranking consistency, this result might not enough exclude overfitting (e.g. memorizing on model specific styles, etc). I would encourage a further testing on several hold-out T2I generators.\n\nThird, a minor question. Using SFT for LMM to score has been discussed by Q-Align (ICML2024), which finds out using logits are better than using `model.generate()` for scoring. It also has the ability for image faithfulness evaluation, please try to compare with it or discuss with it. Furthermore, for faithfulness evaluation (which is actually image quality, am I right?), the compared baselines are similarity-based metrics (which are, from their design, alignment-related metrics). I would suggest the authors to compare with some baselines related to T2I quality evaluation (inc. Q-Align) in this part.", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730641389864}, {"id": "SQXP0KXmYr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2053/Reviewer_rw7T"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces a novel method and dataset aimed at evaluating the quality of generated images, with a specific focus on image faithfulness and text-image alignment. The dataset was collected using detailed human feedback in a question-answer format, aiming to provide fine-grained insights into image quality. This dataset is then used to train a MLLM with SFT to evaluate generated images effectively. The proposed method is tested on the new dataset and compared against existing approaches, with results indicating its superior performance in terms of image faithfulness and text-image alignment.", "review_text": "The paper introduces a novel method and dataset aimed at evaluating the quality of generated images, with a specific focus on image faithfulness and text-image alignment. The dataset was collected using detailed human feedback in a question-answer format, aiming to provide fine-grained insights into image quality. This dataset is then used to train a MLLM with SFT to evaluate generated images effectively. The proposed method is tested on the new dataset and compared against existing approaches, with results indicating its superior performance in terms of image faithfulness and text-image alignment.", "strengths": "- The paper introduces a dataset that includes explicit question-answer feedback, which could facilitate more detailed evaluation of generated images.\n- Using MLLM with SFT for image evaluation is an interesting approach that could potentially enhance interpretability in assessing generated image quality.", "weaknesses": "## 1. Lack of Justification for Main Contributions\nThe paper's three key contributions are insufficiently supported by experimental validation and theoretical grounding:\n   1. Although the dataset is described as having detailed human feedback covering \"11 skills and 2 aspects,\" the experiments primarily focus on the 2 broad aspects. There is little exploration of the 11 specific skills, which would have been valuable given that the 2 aspects have been widely studied in prior works, such as [a].\n   2. The method is claimed to enable \"accurate, comprehensive, fine-grained, and interpretable\" evaluations. However, the results mostly reflect the 2-aspect performance, with no evidence of superior fine-grained or interpretability-focused evaluation compared to previous methods.\n   3. While the paper emphasizes cost-efficiency in terms of annotation and computation, this claim is questionable. The annotation process requires extensive human annotations, which is labor-intensive. Additionally, the method achieves optimal performance with a 34B MLLM model, which is computationally expensive.\n\n## 2. Unclear Advantages Over Existing Datasets\n- According to Table 1, the primary benefit of the proposed dataset seems to be its focus on the two-aspect evaluation. However, several prior datasets such as ImageReward, PickScore, and HPS(v2) implicitly address these aspects as well. While explicit question-based feedback is used in this work, it is not clear how this approach leads to better evaluation outcomes, especially since **a vast number of questions would likely be required to cover all image aspects comprehensively**.\n- The paper does not adequately compare its approach with previous work like [a], which also includes detailed, multi-aspect human feedback via scoring rather than question-answering. The advantages of question-based feedback over scoring are not clearly demonstrated in terms of faithfulness or alignment evaluation.\n- While the paper emphasizes cost-effectiveness, the dataset requires 130k annotations to achieve optimal results (Table 1). This annotation volume does not appear more economical than previous datasets.\n\n## 3. Weak Experimental Results\n- It is unclear whether models from other methods were trained on the proposed dataset to ensure a fair comparison, especially in Tables 2 and 3.\n- The results in these tables do not consistently support the claims made. For instance, the 500 configuration does not show a clear optimal performance in Table 6, and there is no clear positive correlation between model size and performance improvements in Table 7.\n\n## 4. Writing and Structure Issues\n- Some sentences lack clarity and coherence. For instance, “the utilized synthesized images are treated as real images as they don’t explicitly recognize the problem of synthesized images with low image faithfulness” is confusing, especially since HPS(v2) aims to evaluate generated images.\n- Writing structure should be improved. The key novel contributions of the method is unclear.\n- There are some repeated sentences with similar meanings. It is better to re-write them to make the paper more concise. \n\n## Conclusion\n\nWhile the paper presents a promising approach with potential contributions in the form of a detailed dataset and a new evaluation method, it currently lacks sufficient support for its claims. The advantages over existing work remain unclear, and the experimental validation needs improvement. Therefore, the paper is not yet ready for acceptance in its current form.\n\n[a] Rich Human Feedback for Text-to-Image Generation, CVPR 2024, best paper.", "questions": "please see weakness points above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel method and dataset aimed at evaluating the quality of generated images, with a specific focus on image faithfulness and text-image alignment. The dataset was collected using detailed human feedback in a question-answer format, aiming to provide fine-grained insights into image quality. This dataset is then used to train a MLLM with SFT to evaluate generated images effectively. The proposed method is tested on the new dataset and compared against existing approaches, with results indicating its superior performance in terms of image faithfulness and text-image alignment.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper introduces a dataset that includes explicit question-answer feedback, which could facilitate more detailed evaluation of generated images.\n- Using MLLM with SFT for image evaluation is an interesting approach that could potentially enhance interpretability in assessing generated image quality.", "weaknesses": "## 1. Lack of Justification for Main Contributions\nThe paper's three key contributions are insufficiently supported by experimental validation and theoretical grounding:\n   1. Although the dataset is described as having detailed human feedback covering \"11 skills and 2 aspects,\" the experiments primarily focus on the 2 broad aspects. There is little exploration of the 11 specific skills, which would have been valuable given that the 2 aspects have been widely studied in prior works, such as [a].\n   2. The method is claimed to enable \"accurate, comprehensive, fine-grained, and interpretable\" evaluations. However, the results mostly reflect the 2-aspect performance, with no evidence of superior fine-grained or interpretability-focused evaluation compared to previous methods.\n   3. While the paper emphasizes cost-efficiency in terms of annotation and computation, this claim is questionable. The annotation process requires extensive human annotations, which is labor-intensive. Additionally, the method achieves optimal performance with a 34B MLLM model, which is computationally expensive.\n\n## 2. Unclear Advantages Over Existing Datasets\n- According to Table 1, the primary benefit of the proposed dataset seems to be its focus on the two-aspect evaluation. However, several prior datasets such as ImageReward, PickScore, and HPS(v2) implicitly address these aspects as well. While explicit question-based feedback is used in this work, it is not clear how this approach leads to better evaluation outcomes, especially since **a vast number of questions would likely be required to cover all image aspects comprehensively**.\n- The paper does not adequately compare its approach with previous work like [a], which also includes detailed, multi-aspect human feedback via scoring rather than question-answering. The advantages of question-based feedback over scoring are not clearly demonstrated in terms of faithfulness or alignment evaluation.\n- While the paper emphasizes cost-effectiveness, the dataset requires 130k annotations to achieve optimal results (Table 1). This annotation volume does not appear more economical than previous datasets.\n\n## 3. Weak Experimental Results\n- It is unclear whether models from other methods were trained on the proposed dataset to ensure a fair comparison, especially in Tables 2 and 3.\n- The results in these tables do not consistently support the claims made. For instance, the 500 configuration does not show a clear optimal performance in Table 6, and there is no clear positive correlation between model size and performance improvements in Table 7.\n\n## 4. Writing and Structure Issues\n- Some sentences lack clarity and coherence. For instance, “the utilized synthesized images are treated as real images as they don’t explicitly recognize the problem of synthesized images with low image faithfulness” is confusing, especially since HPS(v2) aims to evaluate generated images.\n- Writing structure should be improved. The key novel contributions of the method is unclear.\n- There are some repeated sentences with similar meanings. It is better to re-write them to make the paper more concise. \n\n## Conclusion\n\nWhile the paper presents a promising approach with potential contributions in the form of a detailed dataset and a new evaluation method, it currently lacks sufficient support for its claims. The advantages over existing work remain unclear, and the experimental validation needs improvement. Therefore, the paper is not yet ready for acceptance in its current form.\n\n[a] Rich Human Feedback for Text-to-Image Generation, CVPR 2024, best paper.", "questions": "please see weakness points above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730622476551}], "openreview_url": "https://openreview.net/forum?id=xreOs2yjqf", "arxiv_id": "2406.16562", "paper_pdf": "papers/xreOs2yjqf.pdf", "paper_pdf_sha256": "5fa17083c06ce24c09c194cb91eb80dda63d337722abe3cf19d2f3d5aa00dcb3", "paper_pdf_bytes": 1150031, "paper_pdf_source": "openreview", "code_url": "https://github.com/SAIS-FUXI/EvalAlign", "code_repository": "SAIS-FUXI/EvalAlign", "code_commit": "d3fa7b49bd87e673d2a941d60c99b1d2c564205d", "code_archive": "repos/xreOs2yjqf.zip", "code_archive_sha256": "99f6aa425066ac533e10334e5e2ece4bc8dbe12e04c7dddb63541729c7df0b4b", "code_archive_bytes": 64055, "code_file_count": 25, "code_extensions": {".py": 23, ".sh": 2}, "github_disk_usage_kb": 112, "github_languages": {"Python": 102457, "Shell": 383}, "github_archived": false, "github_pushed_at": "2024-10-23T08:19:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evalalign-evaluating-text-to-image-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "V0CUOBWUHa", "year": 2024, "status": "rejected", "title": "Scaling Sentence Embeddings with Large Language Models", "authors": ["Ting Jiang", "Shaohan Huang", "Zhongzhi Luan", "deqing wang", "Fuzhen Zhuang"], "authorids": ["~Ting_Jiang1", "~Shaohan_Huang1", "~Zhongzhi_Luan1", "~deqing_wang2", "~Fuzhen_Zhuang1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have recently garnered significant interest. With in-context learning, LLMs achieve impressive results in various natural language tasks. However, the application of LLMs to sentence embeddings remains an area of ongoing research. In this work, we propose an in-context learning-based method aimed at improving sentence embeddings performance. Our approach involves adapting the previous prompt-based representation method for autoregressive models, constructing a demonstration set that enables LLMs to perform in-context learning, and scaling up the LLMs to different model sizes. Through extensive experiments, in-context learning enables LLMs to generate high-quality sentence embeddings without any fine-tuning. It helps LLMs achieve performance comparable to current contrastive learning methods. By scaling model size, we find scaling to more than tens of billion parameters harms the performance on semantic textual similarity (STS) tasks. However, the largest model outperforms other counterparts and achieves the new state-of-the-art result on transfer tasks. We also fine-tune LLMs with current contrastive learning approach, and the 2.7B OPT model, incorporating our prompt-based method, surpasses the performance of 4.8B ST5, achieving the new state-of-the-art results on STS tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "iEC7dgyYjo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission596/Reviewer_dT5b"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper investigates how to better leverage large language models (LLMs) for generating sentence representations, traditionally obtained from smaller encoder-based models like BERT variants. \nIt introduces two approaches. \nInitially, it adopts in-context learning, similar to the utilization of LLMs in other tasks. \nEmploying the \"Explicit One word Limitation (EOL)\"—which posits that decoder-based models can produce viable sentence-level representations when prompted to summarize a sentence in a single word—sentence-to-word pair contexts are used to enhance representation derivation. \nDecoder models, without any fine-tuning, showed performance on par with existing contrastive learning approaches. \nAdditionally, the authors explored fine-tuning decoder models using the prevalent contrastive learning framework in sentence representation research, employing the parameter-efficient technique known as QLoRA. \nThe findings reveal that fine-tuning with contrastive learning notably benefits larger decoder models, surpassing smaller encoder models in both Semantic Textual Similarity (STS) benchmarks and transfer tasks for classification.", "review_text": "This paper investigates how to better leverage large language models (LLMs) for generating sentence representations, traditionally obtained from smaller encoder-based models like BERT variants. \nIt introduces two approaches. \nInitially, it adopts in-context learning, similar to the utilization of LLMs in other tasks. \nEmploying the \"Explicit One word Limitation (EOL)\"—which posits that decoder-based models can produce viable sentence-level representations when prompted to summarize a sentence in a single word—sentence-to-word pair contexts are used to enhance representation derivation. \nDecoder models, without any fine-tuning, showed performance on par with existing contrastive learning approaches. \nAdditionally, the authors explored fine-tuning decoder models using the prevalent contrastive learning framework in sentence representation research, employing the parameter-efficient technique known as QLoRA. \nThe findings reveal that fine-tuning with contrastive learning notably benefits larger decoder models, surpassing smaller encoder models in both Semantic Textual Similarity (STS) benchmarks and transfer tasks for classification.", "strengths": "- Suggested a variety of plausible methods for utilizing Large Language Models (LLMs) to compute sentence representations.\n- Explored both in-context learning and fine-tuning approaches with LLMs, encompassing a broad spectrum of potential applications for these models.\n- Introduced a straightforward yet insightful technique for integrating in-context learning into the sentence representation learning paradigm.", "weaknesses": "- While the methods proposed are sound, they consist of previously suggested and widely implemented techniques, which diminishes the novelty aspect of the work.\n- Contrary to SimCSE, the in-context learning approach depends on the use of the STS-B dataset, including its training and validation components, which could potentially confer an unfair advantage to the method.\n- There appears to be no direct link between the two proposed methods; that is, the approach based on in-context learning and the one utilizing contrastive learning.", "questions": "- I'm curious whether the authors have any insights or hypotheses as to why (much) larger models (over 10B) do not excel as expected in computing sentence representations, which contrasts with their effectiveness in other standard applications.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates how to better leverage large language models (LLMs) for generating sentence representations, traditionally obtained from smaller encoder-based models like BERT variants. \nIt introduces two approaches. \nInitially, it adopts in-context learning, similar to the utilization of LLMs in other tasks. \nEmploying the \"Explicit One word Limitation (EOL)\"—which posits that decoder-based models can produce viable sentence-level representations when prompted to summarize a sentence in a single word—sentence-to-word pair contexts are used to enhance representation derivation. \nDecoder models, without any fine-tuning, showed performance on par with existing contrastive learning approaches. \nAdditionally, the authors explored fine-tuning decoder models using the prevalent contrastive learning framework in sentence representation research, employing the parameter-efficient technique known as QLoRA. \nThe findings reveal that fine-tuning with contrastive learning notably benefits larger decoder models, surpassing smaller encoder models in both Semantic Textual Similarity (STS) benchmarks and transfer tasks for classification.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- Suggested a variety of plausible methods for utilizing Large Language Models (LLMs) to compute sentence representations.\n- Explored both in-context learning and fine-tuning approaches with LLMs, encompassing a broad spectrum of potential applications for these models.\n- Introduced a straightforward yet insightful technique for integrating in-context learning into the sentence representation learning paradigm.", "weaknesses": "- While the methods proposed are sound, they consist of previously suggested and widely implemented techniques, which diminishes the novelty aspect of the work.\n- Contrary to SimCSE, the in-context learning approach depends on the use of the STS-B dataset, including its training and validation components, which could potentially confer an unfair advantage to the method.\n- There appears to be no direct link between the two proposed methods; that is, the approach based on in-context learning and the one utilizing contrastive learning.", "questions": "- I'm curious whether the authors have any insights or hypotheses as to why (much) larger models (over 10B) do not excel as expected in computing sentence representations, which contrasts with their effectiveness in other standard applications.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698739476461}, {"id": "ZQGewhNaTX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission596/Reviewer_oEkP"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a set of methods leveraging LLMs for sentence embeddings.\n\n* It introduces a prompting strategy with explicit one word limitation, pushing the model to condense as much information as possible into the last hidden representation. This method is an adaptation of PromptBERT's approach for autoregression models.\n* It leverages in-context-learning in order to improve the quality of sentence embeddings. To this end, it relies on two approaches: 1) It generates one-word summaries of sentences from the STS training set using GPT-3.5. 2) It leverages entries from the Oxford dictionary. The concatenation of samples from these two sources is then incorporated into the LLM's prompt.\n* It leverages fine-tuning with contrastive learning to further improve the quality of sentence embeddings. It does so by leveraging qLORA and training on supervised datasets such as SNLI and MNLI.\n\n\nThe paper's conclusions are as follow:\n* The explicit one-word limitation prompt improves the quality of sentence embeddings derived from OPT on STS benchmarks.\n* In-context learning and supervised fine-tuning improve the performance on STS benchmarks, allowing the proposed solution to beat the state of the art. However, the resulting embeddings do not transfer as well to other tasks.\n* In the paper's proposed setup, the largest base models do not have a clear performance advantage: best results on STS without fine-tuning are obtained with OPT's variants ranging between 1.3 and 6.7B parameters.", "review_text": "This paper proposes a set of methods leveraging LLMs for sentence embeddings.\n\n* It introduces a prompting strategy with explicit one word limitation, pushing the model to condense as much information as possible into the last hidden representation. This method is an adaptation of PromptBERT's approach for autoregression models.\n* It leverages in-context-learning in order to improve the quality of sentence embeddings. To this end, it relies on two approaches: 1) It generates one-word summaries of sentences from the STS training set using GPT-3.5. 2) It leverages entries from the Oxford dictionary. The concatenation of samples from these two sources is then incorporated into the LLM's prompt.\n* It leverages fine-tuning with contrastive learning to further improve the quality of sentence embeddings. It does so by leveraging qLORA and training on supervised datasets such as SNLI and MNLI.\n\n\nThe paper's conclusions are as follow:\n* The explicit one-word limitation prompt improves the quality of sentence embeddings derived from OPT on STS benchmarks.\n* In-context learning and supervised fine-tuning improve the performance on STS benchmarks, allowing the proposed solution to beat the state of the art. However, the resulting embeddings do not transfer as well to other tasks.\n* In the paper's proposed setup, the largest base models do not have a clear performance advantage: best results on STS without fine-tuning are obtained with OPT's variants ranging between 1.3 and 6.7B parameters.", "strengths": "* Originality: the proposed PromptEOL is a novel adaptation of the BERT prompting paradigm for sentence representations. The prompting strategy combining GPT-augmented STS sentences and oxford definitions is novel as well, and the use of qLORA to make contrastive fine-tuning feasible for larger models shows creativity in putting together existing solutions.\n\n* Quality: The experiments are well-devised and executed. The proposed methods are simple and beat the state of the art on semantic textual similarity benchmarks.\n\n* Clarity: The paper articulates very clearly its methodology. It is easy to read and describes well the corresponding pre-existing work. It motivates very well the choice of an explicit one-word prompt, the value of in-context learning and the need for quantization in order to fine-tune the largest models in a contrastive learning setup.\n\n* Significance: while the results on STS benchmarks look good, both in-context learning and contrastive fine-tuning do not show incremental value on transfer tasks. The relatively low generalisation capabilities of these methods limit greatly the appeal of such techniques for the average practitioner, as most real-life applications of sentence embeddings are not for semantic textual similarity.", "weaknesses": "* While it is helpful to the reader to see the entire distribution of Spearman correlations, it may be relevant to give more details on how the two sources for ICL data impact the quality of downstream representations. The 1/3-2/3 mix of STS sentences vs Oxford definitions would benefit from an explicit ablation.\n\n* The value of ICL and CSE is demonstrated only for OPT. Indeed, table 1 is missing results that would demonstrate the added value of ICL and CSE on Llama.\n\n* The proposed methods (in-context learning and possibly fine-tuning) are performing worse than simple explicit-one-word-limit prompting on transfer tasks.\n\n* It is not clear what section 5.2 demonstrates:\n  * first, the text mentions \"in-context learning examples that were obtained from each model on the STS-B development set\", while the table caption reads \"In-context learning examples used in various model size\". The paper states clearly in section 3.2 that the in-context learning examples (1) come from the STS-B training set and (2) are not generated / obtained from the model itself.\n  * second, the method used to sample the data from table 4 is not described, and the meaning of the \"Improve\" column is not clear: does it correspond to the improvement coming from one additional sentence in the prompt? Or from the addition of the 100s of sentences in the ICL prompt? In any case, it seems premature to draw generic conclusions such as \"related examples are usually more implicit\" from a sample size of 1 from each model. Appending 5-10 random samples of each category in the appendix would give more compelling evidence for this.\n  * finally, it is not clear how to relate the findings of section 5.2 to the overall quality of the sentence embeddings introduced by this work.\n\n* Typos and errors:\n  * table 6, the first group of data rows should mention (16-bit)\n  * table 8, the ordering is not the same between the \"Without fine-tuning\" and \"Fine-tuning on unsupervised datasets\" groups of rows.", "questions": "1. It would seem that some experiments have been run only on OPT, while others have been run on OPT and Llama. It would be helpful to have all experiments run on both models to show that the conclusions are robust to the choice of base LLM.\n\n2. How was the data mix for ICL (1/3 STS sentences, 2/3 Oxford definitions) devised? Are they both necessary to achieve good performance? A proper ablation of this setup would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a set of methods leveraging LLMs for sentence embeddings.\n\n* It introduces a prompting strategy with explicit one word limitation, pushing the model to condense as much information as possible into the last hidden representation. This method is an adaptation of PromptBERT's approach for autoregression models.\n* It leverages in-context-learning in order to improve the quality of sentence embeddings. To this end, it relies on two approaches: 1) It generates one-word summaries of sentences from the STS training set using GPT-3.5. 2) It leverages entries from the Oxford dictionary. The concatenation of samples from these two sources is then incorporated into the LLM's prompt.\n* It leverages fine-tuning with contrastive learning to further improve the quality of sentence embeddings. It does so by leveraging qLORA and training on supervised datasets such as SNLI and MNLI.\n\n\nThe paper's conclusions are as follow:\n* The explicit one-word limitation prompt improves the quality of sentence embeddings derived from OPT on STS benchmarks.\n* In-context learning and supervised fine-tuning improve the performance on STS benchmarks, allowing the proposed solution to beat the state of the art. However, the resulting embeddings do not transfer as well to other tasks.\n* In the paper's proposed setup, the largest base models do not have a clear performance advantage: best results on STS without fine-tuning are obtained with OPT's variants ranging between 1.3 and 6.7B parameters.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "* Originality: the proposed PromptEOL is a novel adaptation of the BERT prompting paradigm for sentence representations. The prompting strategy combining GPT-augmented STS sentences and oxford definitions is novel as well, and the use of qLORA to make contrastive fine-tuning feasible for larger models shows creativity in putting together existing solutions.\n\n* Quality: The experiments are well-devised and executed. The proposed methods are simple and beat the state of the art on semantic textual similarity benchmarks.\n\n* Clarity: The paper articulates very clearly its methodology. It is easy to read and describes well the corresponding pre-existing work. It motivates very well the choice of an explicit one-word prompt, the value of in-context learning and the need for quantization in order to fine-tune the largest models in a contrastive learning setup.\n\n* Significance: while the results on STS benchmarks look good, both in-context learning and contrastive fine-tuning do not show incremental value on transfer tasks. The relatively low generalisation capabilities of these methods limit greatly the appeal of such techniques for the average practitioner, as most real-life applications of sentence embeddings are not for semantic textual similarity.", "weaknesses": "* While it is helpful to the reader to see the entire distribution of Spearman correlations, it may be relevant to give more details on how the two sources for ICL data impact the quality of downstream representations. The 1/3-2/3 mix of STS sentences vs Oxford definitions would benefit from an explicit ablation.\n\n* The value of ICL and CSE is demonstrated only for OPT. Indeed, table 1 is missing results that would demonstrate the added value of ICL and CSE on Llama.\n\n* The proposed methods (in-context learning and possibly fine-tuning) are performing worse than simple explicit-one-word-limit prompting on transfer tasks.\n\n* It is not clear what section 5.2 demonstrates:\n  * first, the text mentions \"in-context learning examples that were obtained from each model on the STS-B development set\", while the table caption reads \"In-context learning examples used in various model size\". The paper states clearly in section 3.2 that the in-context learning examples (1) come from the STS-B training set and (2) are not generated / obtained from the model itself.\n  * second, the method used to sample the data from table 4 is not described, and the meaning of the \"Improve\" column is not clear: does it correspond to the improvement coming from one additional sentence in the prompt? Or from the addition of the 100s of sentences in the ICL prompt? In any case, it seems premature to draw generic conclusions such as \"related examples are usually more implicit\" from a sample size of 1 from each model. Appending 5-10 random samples of each category in the appendix would give more compelling evidence for this.\n  * finally, it is not clear how to relate the findings of section 5.2 to the overall quality of the sentence embeddings introduced by this work.\n\n* Typos and errors:\n  * table 6, the first group of data rows should mention (16-bit)\n  * table 8, the ordering is not the same between the \"Without fine-tuning\" and \"Fine-tuning on unsupervised datasets\" groups of rows.", "questions": "1. It would seem that some experiments have been run only on OPT, while others have been run on OPT and Llama. It would be helpful to have all experiments run on both models to show that the conclusions are robust to the choice of base LLM.\n\n2. How was the data mix for ICL (1/3 STS sentences, 2/3 Oxford definitions) devised? Are they both necessary to achieve good performance? A proper ablation of this setup would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698664430369}, {"id": "RSKudrzs3F", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission596/Reviewer_ViaW"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper works on generating sentence embeddings using large language models. First,  the authors design a specific prompt  to compress the semantic of an input sentence into a single word.  Then, this paper investigates zero-shot,  in-context and fine-tuning settings of sentence embedding learning.  For in-context learning,  this paper proposes a demonstration selection method for inducing good sentence representations.  For fine-tuning, to solve the large memory issue,  the authors use QLoRA to perform contrastive learning.  Empirical results on common sentence embedding evaluation benchmarks with both OPT and LLaMA series models show that the proposed method can match (or even exceed) the performance of pretrained language models (such as BERT).", "review_text": "This paper works on generating sentence embeddings using large language models. First,  the authors design a specific prompt  to compress the semantic of an input sentence into a single word.  Then, this paper investigates zero-shot,  in-context and fine-tuning settings of sentence embedding learning.  For in-context learning,  this paper proposes a demonstration selection method for inducing good sentence representations.  For fine-tuning, to solve the large memory issue,  the authors use QLoRA to perform contrastive learning.  Empirical results on common sentence embedding evaluation benchmarks with both OPT and LLaMA series models show that the proposed method can match (or even exceed) the performance of pretrained language models (such as BERT).", "strengths": "1. The writing is easy to follow and the idea is well presented. \n2. The proposed prompt, in-context demonstration and fine-tuning method solve the specific issues of scaling large language models for sentence embedding learning. \n3. The experimental results are effective on both the SentEval and Transfer settings compared to BERT-base based contrastive learning method.", "weaknesses": "1. In Table 1,  only the results based on OPT are presented.  Why not also including the results based on LLaMA? \n2. In Table 1,  the best configuration PromptEOL+ICL+ OPT (6.7B) does not show clear advantages than PromptRoBERTa (123M). \n3. For the in-context setting,  why only use one demonstration?   In Table 1,  comparing PromptEOL+ICL + OPT with baselines models  is not fair since the baseline models do not use the development set. \n4. When the model size increases, the performance does always not increase.  Especially, the 13B, 30B, and 60B models do not perform better than smaller models such as 1.3B and 6B models. \n5. The overall method is a little bit heavy. It is worth to discuss whether we should improve the sentence embeddings using large language models.", "questions": "1. Do you also try LLaMA 2? \n2. In Equation 1, why using the last token hidden state as the sentence representation instead of the representation vector of the last generated token using the explicit one word prompt? \n3. For the fine-tuning, do you also try including in-context demonstration for the fine-tuning? \n\nMinors: \n\nThe citation format is not correct. Please correct all of them.  Try to use the cite command in a correct way.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper works on generating sentence embeddings using large language models. First,  the authors design a specific prompt  to compress the semantic of an input sentence into a single word.  Then, this paper investigates zero-shot,  in-context and fine-tuning settings of sentence embedding learning.  For in-context learning,  this paper proposes a demonstration selection method for inducing good sentence representations.  For fine-tuning, to solve the large memory issue,  the authors use QLoRA to perform contrastive learning.  Empirical results on common sentence embedding evaluation benchmarks with both OPT and LLaMA series models show that the proposed method can match (or even exceed) the performance of pretrained language models (such as BERT).", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The writing is easy to follow and the idea is well presented. \n2. The proposed prompt, in-context demonstration and fine-tuning method solve the specific issues of scaling large language models for sentence embedding learning. \n3. The experimental results are effective on both the SentEval and Transfer settings compared to BERT-base based contrastive learning method.", "weaknesses": "1. In Table 1,  only the results based on OPT are presented.  Why not also including the results based on LLaMA? \n2. In Table 1,  the best configuration PromptEOL+ICL+ OPT (6.7B) does not show clear advantages than PromptRoBERTa (123M). \n3. For the in-context setting,  why only use one demonstration?   In Table 1,  comparing PromptEOL+ICL + OPT with baselines models  is not fair since the baseline models do not use the development set. \n4. When the model size increases, the performance does always not increase.  Especially, the 13B, 30B, and 60B models do not perform better than smaller models such as 1.3B and 6B models. \n5. The overall method is a little bit heavy. It is worth to discuss whether we should improve the sentence embeddings using large language models.", "questions": "1. Do you also try LLaMA 2? \n2. In Equation 1, why using the last token hidden state as the sentence representation instead of the representation vector of the last generated token using the explicit one word prompt? \n3. For the fine-tuning, do you also try including in-context demonstration for the fine-tuning? \n\nMinors: \n\nThe citation format is not correct. Please correct all of them.  Try to use the cite command in a correct way.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698602302196}], "openreview_url": "https://openreview.net/forum?id=V0CUOBWUHa", "arxiv_id": "2307.16645", "paper_pdf": "papers/V0CUOBWUHa.pdf", "paper_pdf_sha256": "965ec0863b409226598ca0354c7e8c3ad8439d9e69274de4208f82a2f0f9120b", "paper_pdf_bytes": 1168695, "paper_pdf_source": "openreview", "code_url": "https://github.com/kongds/scaling_sentemb", "code_repository": "kongds/scaling_sentemb", "code_commit": "8567aa083c1b3c77586670f91e7f78eb80694ad3", "code_archive": "repos/V0CUOBWUHa.zip", "code_archive_sha256": "b0cf75f8e6e7d9c590e82f080a72b27dafc8f682bd34333ee3b2cdcdeeaa156a", "code_archive_bytes": 84059, "code_file_count": 32, "code_extensions": {".py": 27, ".sh": 5}, "github_disk_usage_kb": 87, "github_languages": {"Python": 206324, "Shell": 7237}, "github_archived": false, "github_pushed_at": "2024-03-22T13:51:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/scaling-sentence-embeddings-with-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LOTGOB5_Xh2", "year": 2023, "status": "rejected", "title": "Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN", "authors": ["Siyuan Li", "Di Wu", "Fang Wu", "Zelin Zang", "Lei Shang", "Baigui Sun", "Xuansong Xie", "Stan Z. Li"], "authorids": ["~Siyuan_Li6", "~Di_Wu10", "~Fang_Wu1", "~Zelin_Zang2", "~Lei_Shang1", "~Baigui_Sun1", "~Xuansong_Xie1", "~Stan_Z._Li2"], "authors_source": "OpenReview API", "abstract": "Masked image modeling (MIM), an emerging self-supervised pre-training method, has shown impressive success across numerous downstream vision tasks with Vision transformers (ViTs). Its underlying idea is simple: a portion of the input image is randomly masked out and then reconstructed via the pre-text task. However, the working principle behind MIM is not well explained, and previous studies insist that MIM primarily works for the Transformer family but is incompatible with CNNs. In this paper, we first study interactions among patches to understand what knowledge is learned and how it is acquired via the MIM task. We observe that MIM essentially teaches the model to learn better middle-order interactions among patches and extract more generalized features. Based on this fact, we propose an Architecture-Agnostic Masked Image Modeling framework (A$^2$MIM), which is compatible with both Transformers and CNNs in a unified way. Extensive experiments on popular benchmarks show that our A$^2$MIM learns better representations without explicit design and endows the backbone model with the stronger capability to transfer to various downstream tasks for both Transformers and CNNs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "r63j3gCaCLQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2521/Reviewer_RicF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a variant of BERT-like pretraining method for computer vision (the so-called Masked Image Modeling). The method can generalize well to both vision transformers and CNNs, thus is described as architecture agnostic.\n\nThe main contributions of this work are:\n- using the mean color values to fill the corrupted pixels, which makes the self-supervised pipeline suitable for different neural network architectures;\n- giving a nice insight that the middle-order interactions are important for visual representation, and making two improvements to standard BERT algorithm to facilitate their learning (i. masking intermediate features and ii. introducing supervision in the frequency domain);\n- and providing some empirical evidence that supports its validity.\n\n&nbsp;", "review_text": "Overall, I find this work provides some interesting insights, but the writing and presentation of the paper leaves much to be desired.\nIn addition, the methodological improvements to BERT pre-training in this paper seems to be trivial and piecemeal.\nI hope to understand this article better in further communication with the authors and that may influence my judgment.", "strengths": "**[strengths]**\n- [S1] This work presents a *fresh* perspective to rethink the self-supervised learning in computer vision, say, the *middle-order* interaction. The \"interaction strength\" mentioned in the paper is a useful indicator to show how good the model is at modeling long-range dependency, which is actually considered to be a valuable aspect in advancing pre-training for NLP [1,2].\n- [S2] The authors have put considerable effort into how to fairly compare the established algorithms, which is appreciated and allows the hypotheses presented in their paper to be fully tested. The ablation experiments also demonstrate the validity of different components in their method.\n\n\n&nbsp;\n\n\n**[weakness]**\n\nThere are some parts of the article that are not very easy for the reader to follow, which could be the main weakness.\nI would suggest authors to explain more about the following aspects (which I find enlightening) and eventually add them to their manuscript or appendix:\n  - [W1] In the Introduction, the authors claim that \"it is not straightforward to directly apply mask token to CNNs\". Can we replace the masked part (e.g., 16x16x3=768 pixels) with a \"mask token\" (e.g., 768 learnable scalars) like BERT does? This is a straightforward way to mask. Or would it be more appropriate to describe this way as \"underperforming\" than \"not straightforward\"? (considering the mediocre performance of MAE, SimMIM, etc. in Tab. 1 of the paper)\n  - [W2] \"MAE takes the reorganized the unmasked input patches of 112x112 as the input\", can the authors explain more on this? I wonder if a substitution like the above (768 pixels to a 768-dimensional learnable vector) or some other operation has been performed.\n  - [W3] According to the analysis in Sec. 3.2, middle-order interactions seem to manifest a medium- or long-range inter-patch dependency. Why the authors say \"middle-order interactions could be enhanced via guiding the network to learn features of certain frequencies\"? Does \"certain frequencies\" refer to medium or high frequencies? Why does increasing the richness of these frequencies on feature maps can promote middle-order interactions?\n\n[W4] In addition, there appears to be some related work that has not been adequately discussed. See the \"Clarity, Quality, Novelty And Reproducibility\".\n\n&nbsp;\n\n**[open questions]**\n\n[O1] The methodological improvements to BERT-like pre-training in this paper seems to be incremental. \nConsidering that the authors propose some quantitative metrics to describe middle-order operations, is it possible to design a more principled algorithm to explicitly facilitate such middle-order interactions? I believe the insights on middle-order are valuable, but the solutions proposed in the article do not seem to fully exploit their values.\n\n&nbsp;\n\n------------------\n[1] Jawahar, Ganesh, Benoît Sagot, and Djamé Seddah. \"What Does BERT Learn about the Structure of Language?.\" Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.\n\n[2] Xu, Jiacheng, et al. \"Discourse-Aware Neural Extractive Text Summarization.\" Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.\n\n&nbsp;\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes a variant of BERT-like pretraining method for computer vision (the so-called Masked Image Modeling). The method can generalize well to both vision transformers and CNNs, thus is described as architecture agnostic.\n\nThe main contributions of this work are:\n- using the mean color values to fill the corrupted pixels, which makes the self-supervised pipeline suitable for different neural network architectures;\n- giving a nice insight that the middle-order interactions are important for visual representation, and making two improvements to standard BERT algorithm to facilitate their learning (i. masking intermediate features and ii. introducing supervision in the frequency domain);\n- and providing some empirical evidence that supports its validity.\n\n&nbsp;", "strength_and_weaknesses": "**[strengths]**\n- [S1] This work presents a *fresh* perspective to rethink the self-supervised learning in computer vision, say, the *middle-order* interaction. The \"interaction strength\" mentioned in the paper is a useful indicator to show how good the model is at modeling long-range dependency, which is actually considered to be a valuable aspect in advancing pre-training for NLP [1,2].\n- [S2] The authors have put considerable effort into how to fairly compare the established algorithms, which is appreciated and allows the hypotheses presented in their paper to be fully tested. The ablation experiments also demonstrate the validity of different components in their method.\n\n\n&nbsp;\n\n\n**[weakness]**\n\nThere are some parts of the article that are not very easy for the reader to follow, which could be the main weakness.\nI would suggest authors to explain more about the following aspects (which I find enlightening) and eventually add them to their manuscript or appendix:\n  - [W1] In the Introduction, the authors claim that \"it is not straightforward to directly apply mask token to CNNs\". Can we replace the masked part (e.g., 16x16x3=768 pixels) with a \"mask token\" (e.g., 768 learnable scalars) like BERT does? This is a straightforward way to mask. Or would it be more appropriate to describe this way as \"underperforming\" than \"not straightforward\"? (considering the mediocre performance of MAE, SimMIM, etc. in Tab. 1 of the paper)\n  - [W2] \"MAE takes the reorganized the unmasked input patches of 112x112 as the input\", can the authors explain more on this? I wonder if a substitution like the above (768 pixels to a 768-dimensional learnable vector) or some other operation has been performed.\n  - [W3] According to the analysis in Sec. 3.2, middle-order interactions seem to manifest a medium- or long-range inter-patch dependency. Why the authors say \"middle-order interactions could be enhanced via guiding the network to learn features of certain frequencies\"? Does \"certain frequencies\" refer to medium or high frequencies? Why does increasing the richness of these frequencies on feature maps can promote middle-order interactions?\n\n[W4] In addition, there appears to be some related work that has not been adequately discussed. See the \"Clarity, Quality, Novelty And Reproducibility\".\n\n&nbsp;\n\n**[open questions]**\n\n[O1] The methodological improvements to BERT-like pre-training in this paper seems to be incremental. \nConsidering that the authors propose some quantitative metrics to describe middle-order operations, is it possible to design a more principled algorithm to explicitly facilitate such middle-order interactions? I believe the insights on middle-order are valuable, but the solutions proposed in the article do not seem to fully exploit their values.\n\n&nbsp;\n\n------------------\n[1] Jawahar, Ganesh, Benoît Sagot, and Djamé Seddah. \"What Does BERT Learn about the Structure of Language?.\" Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.\n\n[2] Xu, Jiacheng, et al. \"Discourse-Aware Neural Extractive Text Summarization.\" Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.\n\n&nbsp;\n", "clarity,_quality,_novelty_and_reproducibility": "**[missing discussions that may weaken the novelty]**\n\nIn the Related Work section, the authors refer to the work CIM [3].\nThis is an Electra-like [4] method that fills the corrupted pixels via a small network instead of mean color values, and thus [4] claimed they are the first to demonstrates that both ViT and CNN can learn rich visual representations using a unified, non-Siamese framework.\nAuthors are encouraged to discuss A$^2$MIM and CIM in more detail, and to update some of the corresponding descriptions.\nAlthough this would weaken the novelty of A$^2$MIM, the related works deserve to be discussed fairly and pertinently.\n\n\n**[Reproducibility]**\n\nI believe one can easily reproduce the main results in this work given the detailed experimental configurations and source codes.\n\n\n------------------\n[3] Fang, Yuxin, et al. \"Corrupted image modeling for self-supervised visual pre-training.\" arXiv preprint arXiv:2202.03382 (2022).\n\n[4] Clark, Kevin, et al. \"Electra: Pre-training text encoders as discriminators rather than generators.\" arXiv preprint arXiv:2003.10555 (2020).", "summary_of_the_review": "Overall, I find this work provides some interesting insights, but the writing and presentation of the paper leaves much to be desired.\nIn addition, the methodological improvements to BERT pre-training in this paper seems to be trivial and piecemeal.\nI hope to understand this article better in further communication with the authors and that may influence my judgment.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666727426251}, {"id": "wPGr2xr6AH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2521/Reviewer_dEJJ"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studied the problem that MIM is compatible with the Transformer family but is incompatible with CNNs. To this end, it proposed an Architecture-Agnostic Masked Image Modeling framework (A2MIM) that is compatible with both Transformers and CNNs in a unified way. Specifically, this paper used RGB Mean as masked tokens and added mask tokens on the intermediate feature maps. A frequency domain reconstruction loss is also applied to improve the ability of learned models. Experiments are conducted on ImageNet-1K, downstream COCO detection and segmentation, and ADE20K dataset with ViT models and ResNet-50.", "review_text": "This paper conducted sufficient experiments. However, the method itself is not novel with marginal improvement, the novelty and originality of the method are basically not very strong. Also, many statements in this paper are slightly over-claimed (details please refer to \"Strength And Weaknesses\"). Thus, I tend to reject.", "strengths": "Strengths:\n\n    - MIM is an interesting problem/framework to study, and understanding how MIM-based methods perform well in the vision domain is important for this field.\n\n    - The proposed method is clear to understand and is easy to follow by other researchers.\n\n    - The experiments are extensive on ImageNet-1K, COCO, and ADE20K datasets.\n\nWeaknesses:\n\n   - Some statements are incorrect or even a little bit over-claimed, such as:\n\n1)\tIn the abstract, the authors stated: “MIM primarily works for the Transformer family but is incompatible with CNNs.” And in the introduction section, this paper stated \"To the best of our knowledge, we are the first to carry out MIM on CNNs that outperforms contrastive learning counterparts.\" This is not true as ConvNext [1] is adequate to handle MIM learning strategy. And this paper with CNN is actually similar to it when using patchified images with CNN.\n\n[1] Liu, Zhuang, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. \"A convnet for the 2020s.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11976-11986. 2022.\n\n2)\t“Based on this fact, we propose an Architecture-Agnostic Masked Image Modeling framework (A2MIM), which is compatible with both Transformers and CNNs in a unified way.” The proposed framework is basically similar to the regular MIM approach with trivial or not significant modifications (use RGB Mean as the Masked Tokens and add mask token on the intermediate feature map). Thus, the called \"a unified way\" seems incorrect. The used Fourier/Frequency domain is also not related to the core framework in this paper and can straightforwardly be applied in the vanilla MIM models.\n\n3)\tThe authors stated that the proposed framework has “(a) No complex or non-generic designs are adopted to ensure compatibility with all network architectures. (b) Better middle order interactions between patches for more generalized feature extraction.” I think this is because the proposed framework is too close to the original MIM architecture.\n\n4)\tThe authors claimed they “delved deep into MIM and answered the question of what exactly is learned during MIM pre-training.” This statement is very strong, however, after reading this paper carefully, I still did not get the insights or intuition about what MIM pre-training learns from the input data from the paper’s descriptions. Proper explorations and explanations are necessary to support this claim.\n\n - The performance in this paper is not competitive. As shown in Tables 1, 2, 4, etc., the improvement is fairly marginal (0%~0.2%). Considering that this paper used extra Fourier/Frequency domain reconstruction supervision/loss, I’m not sure whether the proposed strategy is truly effective or not.\n\n - The writing and organization of this paper can be improved. The used Fourier/Frequency domain reconstruction loss seems not related to the key approach of the architecture-agnostic masked image modeling framework, since it can also be applied to the regular MIM frameworks. Moreover, the insights of using this additional supervision are not clearly expressed. This part seems fragmented from others in the method.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper studied the problem that MIM is compatible with the Transformer family but is incompatible with CNNs. To this end, it proposed an Architecture-Agnostic Masked Image Modeling framework (A2MIM) that is compatible with both Transformers and CNNs in a unified way. Specifically, this paper used RGB Mean as masked tokens and added mask tokens on the intermediate feature maps. A frequency domain reconstruction loss is also applied to improve the ability of learned models. Experiments are conducted on ImageNet-1K, downstream COCO detection and segmentation, and ADE20K dataset with ViT models and ResNet-50.", "strength_and_weaknesses": "Strengths:\n\n    - MIM is an interesting problem/framework to study, and understanding how MIM-based methods perform well in the vision domain is important for this field.\n\n    - The proposed method is clear to understand and is easy to follow by other researchers.\n\n    - The experiments are extensive on ImageNet-1K, COCO, and ADE20K datasets.\n\nWeaknesses:\n\n   - Some statements are incorrect or even a little bit over-claimed, such as:\n\n1)\tIn the abstract, the authors stated: “MIM primarily works for the Transformer family but is incompatible with CNNs.” And in the introduction section, this paper stated \"To the best of our knowledge, we are the first to carry out MIM on CNNs that outperforms contrastive learning counterparts.\" This is not true as ConvNext [1] is adequate to handle MIM learning strategy. And this paper with CNN is actually similar to it when using patchified images with CNN.\n\n[1] Liu, Zhuang, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. \"A convnet for the 2020s.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11976-11986. 2022.\n\n2)\t“Based on this fact, we propose an Architecture-Agnostic Masked Image Modeling framework (A2MIM), which is compatible with both Transformers and CNNs in a unified way.” The proposed framework is basically similar to the regular MIM approach with trivial or not significant modifications (use RGB Mean as the Masked Tokens and add mask token on the intermediate feature map). Thus, the called \"a unified way\" seems incorrect. The used Fourier/Frequency domain is also not related to the core framework in this paper and can straightforwardly be applied in the vanilla MIM models.\n\n3)\tThe authors stated that the proposed framework has “(a) No complex or non-generic designs are adopted to ensure compatibility with all network architectures. (b) Better middle order interactions between patches for more generalized feature extraction.” I think this is because the proposed framework is too close to the original MIM architecture.\n\n4)\tThe authors claimed they “delved deep into MIM and answered the question of what exactly is learned during MIM pre-training.” This statement is very strong, however, after reading this paper carefully, I still did not get the insights or intuition about what MIM pre-training learns from the input data from the paper’s descriptions. Proper explorations and explanations are necessary to support this claim.\n\n - The performance in this paper is not competitive. As shown in Tables 1, 2, 4, etc., the improvement is fairly marginal (0%~0.2%). Considering that this paper used extra Fourier/Frequency domain reconstruction supervision/loss, I’m not sure whether the proposed strategy is truly effective or not.\n\n - The writing and organization of this paper can be improved. The used Fourier/Frequency domain reconstruction loss seems not related to the key approach of the architecture-agnostic masked image modeling framework, since it can also be applied to the regular MIM frameworks. Moreover, the insights of using this additional supervision are not clearly expressed. This part seems fragmented from others in the method.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is clearly presented, however, the originality of this work is slightly insufficient.", "summary_of_the_review": "This paper conducted sufficient experiments. However, the method itself is not novel with marginal improvement, the novelty and originality of the method are basically not very strong. Also, many statements in this paper are slightly over-claimed (details please refer to \"Strength And Weaknesses\"). Thus, I tend to reject.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666604939867}, {"id": "VOpndu5oY8", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2521/Reviewer_yVqw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper inspects MIM pre-training approaches and finds MIM essentially helps the model to learn better middle-order interactions between patches. Motivated by that, this paper proposes a novel MIM-based method, dubbed A$^2$MIM, that works well for both (small-sized) ConvNets & ViTs. Extensive experiments are conducted to verify the effectiveness of the proposed approach.", "review_text": "Overall, I think this is a solid and well-written paper.\n", "strengths": "### Strength \n1. This paper provides a thoroughgoing study of MIM, and highlights the essence of MIM pre-training is to help models learn middle-order interactions between patches.\n2. Based on the study of middle-order interactions of MIM, this paper proposes a novel approach that works well for both ConvNets as well as ViTs. What's more, I appreciate the management of technical details in this paper, such as ``Filling Masked Tokens with RGB Mean'' is technically sound. \n3. Extensive experiments on several downstream visual recognition tasks are conducted to verify the effectiveness of the proposed approach.\n   \n### Weaknesses\n1. The 2nd paragraph's name of Sec. 2 Related Work is inaccurate.``Autoregressive Modeling'' means that the output depends linearly on its own *previous* values (e.g, GPTs), while masked modeling is *bidirectional* modeling.\n2. I don't see a strong connection between Sec. 3.1 and the proposed approach.\n3. The scalability of A$^2$MIM is unknown, which is a crucial property for pre-training.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper inspects MIM pre-training approaches and finds MIM essentially helps the model to learn better middle-order interactions between patches. Motivated by that, this paper proposes a novel MIM-based method, dubbed A$^2$MIM, that works well for both (small-sized) ConvNets & ViTs. Extensive experiments are conducted to verify the effectiveness of the proposed approach.", "strength_and_weaknesses": "### Strength \n1. This paper provides a thoroughgoing study of MIM, and highlights the essence of MIM pre-training is to help models learn middle-order interactions between patches.\n2. Based on the study of middle-order interactions of MIM, this paper proposes a novel approach that works well for both ConvNets as well as ViTs. What's more, I appreciate the management of technical details in this paper, such as ``Filling Masked Tokens with RGB Mean'' is technically sound. \n3. Extensive experiments on several downstream visual recognition tasks are conducted to verify the effectiveness of the proposed approach.\n   \n### Weaknesses\n1. The 2nd paragraph's name of Sec. 2 Related Work is inaccurate.``Autoregressive Modeling'' means that the output depends linearly on its own *previous* values (e.g, GPTs), while masked modeling is *bidirectional* modeling.\n2. I don't see a strong connection between Sec. 3.1 and the proposed approach.\n3. The scalability of A$^2$MIM is unknown, which is a crucial property for pre-training.", "clarity,_quality,_novelty_and_reproducibility": "The quality, clarity and originality are sound.\n", "summary_of_the_review": "Overall, I think this is a solid and well-written paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666601395929}, {"id": "NZgVHfcUF1e", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2521/Reviewer_nEcw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper analyzes the essence of MIM and introduces a universal MIM method that can be applied to both CNNs and Transformers. There are several designs that together build the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets. Experimental results on both ResNet-50 and ViTs are shown.", "review_text": "I tend to weakly reject this paper in the initial comments due to the unfounded premises. I am looking forward to the authors' opinions in the rebuttal period on why we need MIM for CNNs.", "strengths": "### Strengths\n1. The paper is well-organized and easy to follow.\n2. The figures, tables, and visualizations indeed help to understand the author's point of view.\n3. It is new to apply MIM on convolution neural networks.\n\n### Weaknesses\n1. I have the most concern about the motivation for this work. As we know, ViTs lack the capture of inductive bias and MIM solves this issue to some extent by enforcing visual context reasoning. However, CNNs are different, they are good at capturing inductive bias. Furthermore, as observed in the authors' results (Tab. 1), a lot of effort has been done to do MIM pre-training, however, the final results are almost the same as supervised counterparts. So is it a false proposition to apply MIM to CNNs?\n2. I have another concern about the authors' declaration that MIM enables the network with a better feature extraction ability. As a common observation in previous MIM works, the linearly probing results of MIM-pretrained models are unsatisfactory, worse than the contrastive methods. MIM is known to provide transferable model parameters rather than extracting out-of-the-box features.\n3. In Sec. 3.1, the authors analyze different augmentation methods, however, the setup is not fair. MIM is employed for pretraining while the others (e.g., CutMix) are for fine-tuning. \n4. There are many components in the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets. How much does each component affect?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper analyzes the essence of MIM and introduces a universal MIM method that can be applied to both CNNs and Transformers. There are several designs that together build the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets. Experimental results on both ResNet-50 and ViTs are shown.", "strength_and_weaknesses": "### Strengths\n1. The paper is well-organized and easy to follow.\n2. The figures, tables, and visualizations indeed help to understand the author's point of view.\n3. It is new to apply MIM on convolution neural networks.\n\n### Weaknesses\n1. I have the most concern about the motivation for this work. As we know, ViTs lack the capture of inductive bias and MIM solves this issue to some extent by enforcing visual context reasoning. However, CNNs are different, they are good at capturing inductive bias. Furthermore, as observed in the authors' results (Tab. 1), a lot of effort has been done to do MIM pre-training, however, the final results are almost the same as supervised counterparts. So is it a false proposition to apply MIM to CNNs?\n2. I have another concern about the authors' declaration that MIM enables the network with a better feature extraction ability. As a common observation in previous MIM works, the linearly probing results of MIM-pretrained models are unsatisfactory, worse than the contrastive methods. MIM is known to provide transferable model parameters rather than extracting out-of-the-box features.\n3. In Sec. 3.1, the authors analyze different augmentation methods, however, the setup is not fair. MIM is employed for pretraining while the others (e.g., CutMix) are for fine-tuning. \n4. There are many components in the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets. How much does each component affect?", "clarity,_quality,_novelty_and_reproducibility": "+ **Clarity**: The paper is well written, however, I do not agree with several claims (see Weaknesses).\n+ **Quality**: Comprehensive experiments are conducted, however, the premises should be verified, i.e., do we need MIM on CNNs?\n+ **Novelty**: Applying MIM to CNNs is new.\n+ **Reproducibility**: Implementation details are provided. ", "summary_of_the_review": "I tend to weakly reject this paper in the initial comments due to the unfounded premises. I am looking forward to the authors' opinions in the rebuttal period on why we need MIM for CNNs.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "n/a", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666535249874}], "openreview_url": "https://openreview.net/forum?id=LOTGOB5_Xh2", "arxiv_id": "2205.13943", "paper_pdf": "papers/LOTGOB5_Xh2.pdf", "paper_pdf_sha256": "c3b5019a459977720bc268a22455172fdd689c75ba31f8c46542193771e858e8", "paper_pdf_bytes": 8002029, "paper_pdf_source": "openreview", "code_url": "https://github.com/Westlake-AI/A2MIM", "code_repository": "Westlake-AI/A2MIM", "code_commit": "ccca6dc21995b8e39e5571bd189c1ebce93bf49d", "code_archive": "repos/LOTGOB5_Xh2.zip", "code_archive_sha256": "4313e080a4198f6abbe73fe55622798b89ad3a1c5627c503709a872fa8af07a3", "code_archive_bytes": 283762, "code_file_count": 224, "code_extensions": {".py": 210, ".sh": 14}, "github_disk_usage_kb": 178, "github_languages": {"Python": 516747, "Shell": 9223}, "github_archived": false, "github_pushed_at": "2024-08-15T22:42:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/architecture-agnostic-masked-image-modeling"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "i8d2kdxii1L", "year": 2022, "status": "rejected", "title": "$p$-Laplacian Based Graph Neural Networks", "authors": ["Guoji Fu", "Peilin Zhao", "Yatao Bian"], "authorids": ["~Guoji_Fu1", "~Peilin_Zhao2", "~Yatao_Bian1"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) have demonstrated superior performance for semi-supervised node classification on graphs, as a result of their ability to exploit node features and topological information simultaneously. However, most GNNs implicitly assume that the labels of nodes and their neighbors in a graph are the same or consistent, which does not hold in heterophilic graphs, where the labels of linked nodes are likely to differ. Hence, when the topology is non-informative for label prediction, ordinary GNNs may work significantly worse than simply applying multi-layer perceptrons (MLPs) on each node. To tackle the above problem, we propose a new $p$-Laplacian based GNN model, termed as $^p$GNN, whose message passing mechanism is derived from a discrete regularization framework and could be theoretically explained as an approximation of a polynomial graph filter defined on the spectral domain of $p$-Laplacians. The spectral analysis shows that the new message passing mechanism works simultaneously as low-pass and high-pass filters, thus making $^p$GNNs effective on both homophilic and heterophilic graphs. Empirical studies on real-world and synthetic datasets validate our findings and demonstrate that $^p$GNNs significantly outperform several state-of-the-art GNN architectures on heterophilic benchmarks while achieving competitive performance on homophilic benchmarks. Moreover, $^p$GNNs can adaptively learn aggregation weights and are robust to noisy edges.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "euXj3JsaiIa", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1338/Reviewer_rMNu"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a p-Laplacian based GNN to handle heterophilic graphs and graphs with non-informative topologies. Both the above cases are assumed in most existing GNN architectures and hence this work breaks away from the norm. This work proposes a discrete p-Laplacian message passing scheme which is derived from a discrete regularization framework. The authors do a spectral analysis of their novel p-Laplacian message passing scheme and show that it works as both a low-pass and high-pass filter, which is then applicable to both homophilic and heterophilic graphs. More specifically, they show that p-GNN with p>2, works well for graphs which exhibit strong homophily, while for p \\in [1,2) p-GNN works effectively on heterophilic graphs. The empirical results support the theoretical justifications and outperform the baselines quite significantly especially on heterophilic and non-informative topology bearing graphs.", "review_text": "-- Strengths:\n\n1. The paper is very well-written and easy to follow. It was a pleasure to read the novel theoretical material presented in this paper. \n\n2. The paper proposes a novel architecture of p-GNN that introduces p-Laplacian message passing by leveraging and infusing existing architectures like SGC, APPNP and GPRGCN. \n\n3. The theoretical connection drawn between the “p-Laplacian message passing scheme” and “the polynomial graph filter defined on the spectral domain of the p-Laplacian” is particularly interesting. \n\n-- Weaknesses:\n\n1. It would have been interesting to see a theoretical justification as to why p-GNN always outperforms say GCNs, GATs, etc. (models that stack their message passing layers) on heterophilic graphs. Can the authors please elaborate on this point in their rebuttal? I feel this would give more insights into the behavior of p-GNN and why it is outperforming standard GNN models. \n\n2. Much of the material in Sections 2 and 3 is obtained from Zhou and Scholkopf 2005. Less space could have been used for this explanation.\n\n3. The function signature for the graph gradient operator in Definition 1 does not look right. The domain of this operator has to be a space of functions, it cannot be the set of vertices. Can you please relook into this?\n\n4. In regards to the experiments conducted on noisy edges, it would be interesting to see how p-GNN performs on some stochastic block models based on edge dropping and some other random graph percolations. The random edge drop model seems a bit simplistic and it's not easy to see how this sort of noise is affecting the graph inputs and why p-GNN should do better than other robust GNNs?\n\n5. In regards to the following statement in the paper on Pg 6. \t\t\t\t\n\n“Therefore, for graphs whose topological information is not helpful for label prediction, we could impose more weights on the first term in Eq. (18) by using a large μ so that P-GNNs work more like MLPs which simply learn on node features. While for graphs whose topological information is helpful for label prediction, we could impose more weights on the second term by using a small μ so that P-GNNs can benefit from p-Laplacian smoothing on node features.” \n\nHow does one determine when the graph’s topological information is helpful for\nlabel prediction, as opposed to when it’s not?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a p-Laplacian based GNN to handle heterophilic graphs and graphs with non-informative topologies. Both the above cases are assumed in most existing GNN architectures and hence this work breaks away from the norm. This work proposes a discrete p-Laplacian message passing scheme which is derived from a discrete regularization framework. The authors do a spectral analysis of their novel p-Laplacian message passing scheme and show that it works as both a low-pass and high-pass filter, which is then applicable to both homophilic and heterophilic graphs. More specifically, they show that p-GNN with p>2, works well for graphs which exhibit strong homophily, while for p \\in [1,2) p-GNN works effectively on heterophilic graphs. The empirical results support the theoretical justifications and outperform the baselines quite significantly especially on heterophilic and non-informative topology bearing graphs.", "main_review": "-- Strengths:\n\n1. The paper is very well-written and easy to follow. It was a pleasure to read the novel theoretical material presented in this paper. \n\n2. The paper proposes a novel architecture of p-GNN that introduces p-Laplacian message passing by leveraging and infusing existing architectures like SGC, APPNP and GPRGCN. \n\n3. The theoretical connection drawn between the “p-Laplacian message passing scheme” and “the polynomial graph filter defined on the spectral domain of the p-Laplacian” is particularly interesting. \n\n-- Weaknesses:\n\n1. It would have been interesting to see a theoretical justification as to why p-GNN always outperforms say GCNs, GATs, etc. (models that stack their message passing layers) on heterophilic graphs. Can the authors please elaborate on this point in their rebuttal? I feel this would give more insights into the behavior of p-GNN and why it is outperforming standard GNN models. \n\n2. Much of the material in Sections 2 and 3 is obtained from Zhou and Scholkopf 2005. Less space could have been used for this explanation.\n\n3. The function signature for the graph gradient operator in Definition 1 does not look right. The domain of this operator has to be a space of functions, it cannot be the set of vertices. Can you please relook into this?\n\n4. In regards to the experiments conducted on noisy edges, it would be interesting to see how p-GNN performs on some stochastic block models based on edge dropping and some other random graph percolations. The random edge drop model seems a bit simplistic and it's not easy to see how this sort of noise is affecting the graph inputs and why p-GNN should do better than other robust GNNs?\n\n5. In regards to the following statement in the paper on Pg 6. \t\t\t\t\n\n“Therefore, for graphs whose topological information is not helpful for label prediction, we could impose more weights on the first term in Eq. (18) by using a large μ so that P-GNNs work more like MLPs which simply learn on node features. While for graphs whose topological information is helpful for label prediction, we could impose more weights on the second term by using a small μ so that P-GNNs can benefit from p-Laplacian smoothing on node features.” \n\nHow does one determine when the graph’s topological information is helpful for\nlabel prediction, as opposed to when it’s not?\n", "summary_of_the_review": "The paper is well written and presents a novel p-GNN architecture on top of other PageRank based GNN architectures. I found the theoretical justifications and connections drawn between the message passing scheme and the polynomial graph filter particularly interesting. Also, they had an interesting bound on the risk (Theorem 3). I am overall quite positive about this paper's acceptance, as it makes novel contributions to handle a difficult scenario like heterophilly in graphs and also the fact that they challenge the implicit assumption made by previous works that the \"shape of node neighborhoods\" helps distinguish nodes.. of course this can be extended to graph shapes too.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636211432632}, {"id": "CT7OVmNS2EG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1338/Reviewer_Pzid"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper derives the p-Laplacian message passing formula under the p-Laplacian based regularization framework and further proposes p-GNN architecture. Authors further justify the relations of p-Laplacian message passing with low and high-pass filters and the upper bound of\none layer risk of p-GNNs. Experiments show the superiority of p-GNNs on both heterophilic and homophilic settings. However, we still have some concerns for the paper before further evaluation.", "review_text": "The paper proposes general p-GNNs architecture under the p-Laplacian regularization framework\nwith comprehensive experimental results. The paper is well written and the theoretical part is well\norganized, though not significant. However, we still have some concerns. \n\nStrength:\n(1) The author proposes a new general GNN architecture based on p-Laplacian message passing,\nwhich claims to work on both heterophily and homophily settings.\n(2) Relations of p-Laplacian message passing with low-pass and high-pass filters have been\nexplained.\n(3)Experiments are comprehensive and inclusive.\n\nWeakness:\n(1) More insights on how to choose parameter p should be given. From the upper bound risk theorem, we only know that for K=1, the risk can be controlled by \\mu which further depends on whether the topological information of graphs is useful or not. However, for K >=2, we do not see the relations. More insights are welcomed.\n\n(2) The stationary point for the objective function in equation 8 is problematic when $p < 2$ and $\\nabla f =0$. In particular, when $p=1$  and $\\nabla f =0$, the Laplacian diffusion should have a step function that is not continuous at $\\nabla f =0$. The authors seem not to have given a good discussion on this aspect. \n\n(3) Moreover, when $p<2$, the graph diffusion operator is not Lip-continuous, which violates the assumption made by the upper-bounding risks theorem. So I do not see how that theorem can be applied here. \n\n(4) For p=1, the diffusion becomes a step function (piecewise constant), how can the model be trained? Can the authors also provide the training curve for the p=1 case?\n\n(5) P-GNN can only get comparable performances on homophilic benchmark datasets (both\naccuracy and entropy experiments).\n\n(6) The authors claim that p-lap can handle the heterophilic case. And the explanation is that the p-lap allows big change on the class boundary which can be viewed as a high-pass filter. I do not think this argument is very convincing. Can the authors have TSNE of the node embeddings for the heterophilic case so that we can get some visualization of the results?\n\n\nSome typos:\n(1)Page 2, last paragraph: Heterphily -> Heterophily\n(2)Section 3.1, R^N should be changed into N", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper derives the p-Laplacian message passing formula under the p-Laplacian based regularization framework and further proposes p-GNN architecture. Authors further justify the relations of p-Laplacian message passing with low and high-pass filters and the upper bound of\none layer risk of p-GNNs. Experiments show the superiority of p-GNNs on both heterophilic and homophilic settings. However, we still have some concerns for the paper before further evaluation.", "main_review": "The paper proposes general p-GNNs architecture under the p-Laplacian regularization framework\nwith comprehensive experimental results. The paper is well written and the theoretical part is well\norganized, though not significant. However, we still have some concerns. \n\nStrength:\n(1) The author proposes a new general GNN architecture based on p-Laplacian message passing,\nwhich claims to work on both heterophily and homophily settings.\n(2) Relations of p-Laplacian message passing with low-pass and high-pass filters have been\nexplained.\n(3)Experiments are comprehensive and inclusive.\n\nWeakness:\n(1) More insights on how to choose parameter p should be given. From the upper bound risk theorem, we only know that for K=1, the risk can be controlled by \\mu which further depends on whether the topological information of graphs is useful or not. However, for K >=2, we do not see the relations. More insights are welcomed.\n\n(2) The stationary point for the objective function in equation 8 is problematic when $p < 2$ and $\\nabla f =0$. In particular, when $p=1$  and $\\nabla f =0$, the Laplacian diffusion should have a step function that is not continuous at $\\nabla f =0$. The authors seem not to have given a good discussion on this aspect. \n\n(3) Moreover, when $p<2$, the graph diffusion operator is not Lip-continuous, which violates the assumption made by the upper-bounding risks theorem. So I do not see how that theorem can be applied here. \n\n(4) For p=1, the diffusion becomes a step function (piecewise constant), how can the model be trained? Can the authors also provide the training curve for the p=1 case?\n\n(5) P-GNN can only get comparable performances on homophilic benchmark datasets (both\naccuracy and entropy experiments).\n\n(6) The authors claim that p-lap can handle the heterophilic case. And the explanation is that the p-lap allows big change on the class boundary which can be viewed as a high-pass filter. I do not think this argument is very convincing. Can the authors have TSNE of the node embeddings for the heterophilic case so that we can get some visualization of the results?\n\n\nSome typos:\n(1)Page 2, last paragraph: Heterphily -> Heterophily\n(2)Section 3.1, R^N should be changed into N", "summary_of_the_review": "See the main review. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635936261541}, {"id": "cgYHYECo-9I", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1338/Reviewer_4UH6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a new message-passing design based on p-Laplacian. They demonstrate that their method works better in heterophilic graphs.", "review_text": "## Pros:\n\n(+) The idea of designing new message-passing based on p-Laplacian is interesting.\n\n(+) The empirical evaluation of cSBM is great.\n\n(+) The high-level idea on the effect of different choices of p is inspiring.\n\n## Cons:\n\n(-) Most theorems derived in this paper are either well-known (Theorem 1) or not meaningful (Theorem 3).\n\n(-) The overall clarity of the paper can be improved, especially in section 4.\n\n(-) Some concerns regarding the choice of hyperparameters (see detail below).\n\n## Detail comments:\n\n-\tTheorem 1 is well-known in the graph learning community, (see [4], equation (3) for example). I think there should be a proper reference for it.\n-\tTheorem 3 is not very meaningful, as the derived risk bound is greater than discarding the graph part ($\\beta = 1,\\; \\alpha = 0$). This is because the analysis does not take the correlation of graphs into account. See [3] on how the authors show the graph convolution is helpful under cSBM assumption. I feel like the analysis and result here can be further improved for meaningful bound.\n-\tThe clarity of section 4 and explanation of Theorem 2 should be improved. Regarding the explanation of Theorem 2, I cannot understand why the “weight effect” of p-Laplacian based message passing can help for heterophilic graphs. What does the “boundary of the lined nodes” refer to in the explanation? On the other hand, while I like the explanation of Proposition 1, I feel like the authors treat the matrix M as fixed in this section. However, according to their definition, it should depend on the input feature matrix. Please elaborate more on this point and improve the clarity of the text.\n-\tI find it weird for the choice of hyperparameter for all baseline methods. More precisely, the authors choose the hidden unit = 16 while the default and common choice is 64. I wonder why the authors detour from the standard choice.\n\n### Typo:\n\n1.\tPage 2, the reference of the cSBM is incorrect. Should be [1].\n\n2.\tPage 5, in Remark 2: “In contrast, it is is similar with SGC…”, redundant “is” here.\n\n### Additional questions:\n\n1.\tThe matrix M in Theorem 2 depends on the node embedding F right? I assume it follows the specific aggregation rule of equation (12).\n\n2.\tDoes the iteration (11)-(13) guarantee to converge?\n\n## Reference:\n\n[1] Contextual stochastic block models. Deshpande et al. NeurIPS 2018. \n\n[2] Optimizing generalized pagerank methods for seed-expansion community detection, Li et al. NeurIPS 2019.\n\n[3] Graph Convolution for Semi-Supervised Classification: Improved Linear Separability and Out-of-Distribution Generalization, Baranwal et al. ICML 2021.\n\n[4] PREDICT THEN PROPAGATE: GRAPH NEURAL NETWORKS MEET PERSONALIZED PAGERANK, Klicpera et al. ICLR 2019. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a new message-passing design based on p-Laplacian. They demonstrate that their method works better in heterophilic graphs.", "main_review": "## Pros:\n\n(+) The idea of designing new message-passing based on p-Laplacian is interesting.\n\n(+) The empirical evaluation of cSBM is great.\n\n(+) The high-level idea on the effect of different choices of p is inspiring.\n\n## Cons:\n\n(-) Most theorems derived in this paper are either well-known (Theorem 1) or not meaningful (Theorem 3).\n\n(-) The overall clarity of the paper can be improved, especially in section 4.\n\n(-) Some concerns regarding the choice of hyperparameters (see detail below).\n\n## Detail comments:\n\n-\tTheorem 1 is well-known in the graph learning community, (see [4], equation (3) for example). I think there should be a proper reference for it.\n-\tTheorem 3 is not very meaningful, as the derived risk bound is greater than discarding the graph part ($\\beta = 1,\\; \\alpha = 0$). This is because the analysis does not take the correlation of graphs into account. See [3] on how the authors show the graph convolution is helpful under cSBM assumption. I feel like the analysis and result here can be further improved for meaningful bound.\n-\tThe clarity of section 4 and explanation of Theorem 2 should be improved. Regarding the explanation of Theorem 2, I cannot understand why the “weight effect” of p-Laplacian based message passing can help for heterophilic graphs. What does the “boundary of the lined nodes” refer to in the explanation? On the other hand, while I like the explanation of Proposition 1, I feel like the authors treat the matrix M as fixed in this section. However, according to their definition, it should depend on the input feature matrix. Please elaborate more on this point and improve the clarity of the text.\n-\tI find it weird for the choice of hyperparameter for all baseline methods. More precisely, the authors choose the hidden unit = 16 while the default and common choice is 64. I wonder why the authors detour from the standard choice.\n\n### Typo:\n\n1.\tPage 2, the reference of the cSBM is incorrect. Should be [1].\n\n2.\tPage 5, in Remark 2: “In contrast, it is is similar with SGC…”, redundant “is” here.\n\n### Additional questions:\n\n1.\tThe matrix M in Theorem 2 depends on the node embedding F right? I assume it follows the specific aggregation rule of equation (12).\n\n2.\tDoes the iteration (11)-(13) guarantee to converge?\n\n## Reference:\n\n[1] Contextual stochastic block models. Deshpande et al. NeurIPS 2018. \n\n[2] Optimizing generalized pagerank methods for seed-expansion community detection, Li et al. NeurIPS 2019.\n\n[3] Graph Convolution for Semi-Supervised Classification: Improved Linear Separability and Out-of-Distribution Generalization, Baranwal et al. ICML 2021.\n\n[4] PREDICT THEN PROPAGATE: GRAPH NEURAL NETWORKS MEET PERSONALIZED PAGERANK, Klicpera et al. ICLR 2019. \n", "summary_of_the_review": "I have mixed feelings about this paper. On one hand, I like the newly designed message passing scheme and their implicit bias demonstrated in Proposition 1. Also, conducting experiments on cSBM is great for studying problems of heterophilic graphs. On the other hand, the drawbacks of the paper prevent me from supporting acceptance of it. I think the paper has great potential but requires a major revision.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635266861858}], "openreview_url": "https://openreview.net/forum?id=i8d2kdxii1L", "arxiv_id": "2111.07337", "paper_pdf": "papers/i8d2kdxii1L.pdf", "paper_pdf_sha256": "866c51c45e6d19e8fe905c4ecba62df507dfa3f6765f311cc38cf2b8ff2bfe2e", "paper_pdf_bytes": 3067889, "paper_pdf_source": "openreview", "code_url": "https://github.com/guoji-fu/pGNNs", "code_repository": "guoji-fu/pGNNs", "code_commit": "f9a0e2bbf553a81748bb3ed42e953f71af6028a2", "code_archive": "repos/i8d2kdxii1L.zip", "code_archive_sha256": "92c63acb901c3361f18be276a27abcdb550b942e7afbfbff6b1c48d29ad24ef3", "code_archive_bytes": 318773, "code_file_count": 9, "code_extensions": {".py": 7, ".sh": 2}, "github_disk_usage_kb": 331, "github_languages": {"Python": 41316, "Shell": 3020}, "github_archived": false, "github_pushed_at": "2025-08-26T04:15:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/p-laplacian-based-graph-neural-networks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0naHZ3gZSzo", "year": 2021, "status": "rejected", "title": "Optimizing Large-Scale Hyperparameters via Automated Learning Algorithm", "authors": ["Bin Gu", "Guodong Liu", "Yanfu Zhang", "Xiang Geng", "Heng Huang"], "authorids": ["~Bin_Gu1", "~Guodong_Liu2", "~Yanfu_Zhang1", "~Xiang_Geng1", "~Heng_Huang1"], "authors_source": "OpenReview API", "abstract": "Modern machine learning algorithms usually involve tuning multiple (from one to  thousands) hyperparameters which play a pivotal role  in terms of model  generalizability. Globally choosing appropriate values of hyperparameters is extremely  computationally challenging. Black-box optimization and gradient-based algorithms are  two dominant approaches to hyperparameter optimization while  they have totally  distinct advantages. How to design a new hyperparameter optimization technique inheriting all benefits  from both approaches is still an open problem. To address this challenging problem, in this paper, we  propose a new  hyperparameter optimization  method with zeroth-order hyper-gradients (HOZOG). Specifically, we first exactly formulate  hyperparameter optimization  as  an $\\mathcal{A}$-based constrained optimization problem, where $\\mathcal{A}$ is  a black-box optimization algorithm (such as deep neural network). Then, we use the average zeroth-order hyper-gradients  to update  hyperparameters. We provide the feasibility analysis of using  HOZOG to achieve hyperparameter optimization.  The experimental  results on three representative  hyperparameter (the size is from 1 to 1250) optimization tasks  demonstrate the benefits of HOZOG  in terms of  \\textit{simplicity, scalability, flexibility, effectiveness and efficiency} compared with the  state-of-the-art hyperparameter optimization methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "bYB3_ekcUm-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1559/AnonReviewer5"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Paper summary\nThis paper develops a novel approach to hyperparameter optimization that combines the respective advantages of black-box and gradient-based optimization methods.\nThe proposed method (called HOZOG) is based on zeroth-order hyper-gradients, and is empirically demonstrated on several benchmark tasks to compare favourably against state-of-the-art hyperparameter optimisation approaches in terms of simplicity, scalability, flexibility, effectiveness and efficiency.\n\n### Pros\n- overall, the paper is well and clearly written and thus easy to follow\n- the manuscript addresses an important open challenge that is relevant for many practical application domains, namely hyperparameter optimization \n- to tackle this challenge, the paper proposes an approach that is novel by combining previous ideas in a new, interesting and fairly principled way\n- the empirical evaluation is comprehensive and convincingly demonstrates the efficacy of the proposed approach as compared to previous methods for hyperparameter optimization\n\n### Cons\n- the proposed approach seems to be fundamentally based on the assumption that the hyperparameters (and thus the induced function) are continuous; the experiments thus also seem to only involve continuous hyperparameters; this appears rather restrictive, given that many practical hyperparameter optimization problems involve important discrete variables (e.g. architectural parameters of neural networks such as the number of layers/units, or categorical parameters such as the choice of optimiser); it would be interesting to know if it would be feasible to straightforwardly extend the proposed approach to also consider such discrete variables; if not, then this seems to be a severe limitation of the method presented; note that there are several black-box/Bayesian optimization approaches (including BOHB) that can handle discrete variables, e.g. [1,2,3]\n- the paper does not provide any source code, and I am not sure if I would be able to implement the method from scratch and fully reproduce the reported results with the details mentioned in the paper alone; that being said, the authors promise to make the code available upon paper acceptance\n- in the first sentence in the introduction, you claim that \"Modern machine learning algorithms usually involve tuning multiple hyperparameters whose size could be from one to thousands.\"; you proceed to mention two examples, SVMs and DNNs, with only a hand full of hyperparameters; it is thus not clear in which scenario one would have _thousands_ of hyperparameters\n- in Sec. 2.1, you claim that \"a lot of references have pointed out that [black-box optimization] can only handle hyperparameters from a few to several dozens\"; however, you only cite one paper for this claim, and it would be beneficial to cite more of the \"lot of references\"; in Sec. 3.4, you again claim that \"black-box optimization methods have a weak performance for the high-dimensional hyperparameter optimization problems which is also verified in a large number of existing references\", proceeding to cite only two papers (which I wouldn't consider \"a large number of existing references\")\n\n[1] Oh et al., 2019, \"Combinatorial bayesian optimization using graph representations\"\n[2] Baptista et al., 2018, \"Bayesian Optimization of Combinatorial Structures\"\n[3] Daxberger et al., 2019, \"Mixed-Variable Bayesian Optimization\"\n\n### Review summary\nIn summary, I believe this is an interesting, enjoyable-to-read paper that tackles an important problem by proposing a promising and effective new technique. This paper will likely be of great interest to some members of the ICLR community and might become an impactful contribution for machine learning researchers and practitioners alike. That being said, I see some issues with the generality of the method as well as the presentation of the paper. As a result, I overall recommend acceptance of this manuscript, although not too enthusiastically.\n\n### Post-rebuttal\nI thank the authors for their response, addressing some of the issues/questions I had raised.\nAfter carefully reading the other reviews (and corresponding author responses), I agree with the valid criticisms raised (some of which I had overlooked initially), which unfortunately further dampened my enthusiasm for this work.\nAs a result, I am slightly lowering my score.\nHowever, I very much appreciate the author's efforts in this very promising work, and hope that they will revise their manuscript to take into account the feedback raised in the reviews.\n\n### Minor issues\n- Fig. 1: the ordering of the compared methods differs between the reported datasets\n- Experiments: the name of the MNIST dataset is typically capitalised\n- there are minor grammatical errors throughout the paper; I recommend that the author thoroughly proof-read the manuscript (or have it proof-read) for language issues \n- the plots are barely readable without zooming in; I strongly recommend the authors to increase the font sizes to improve their readability and clarity", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising new method with some limitations in methodology and presentation", "review": "### Paper summary\nThis paper develops a novel approach to hyperparameter optimization that combines the respective advantages of black-box and gradient-based optimization methods.\nThe proposed method (called HOZOG) is based on zeroth-order hyper-gradients, and is empirically demonstrated on several benchmark tasks to compare favourably against state-of-the-art hyperparameter optimisation approaches in terms of simplicity, scalability, flexibility, effectiveness and efficiency.\n\n### Pros\n- overall, the paper is well and clearly written and thus easy to follow\n- the manuscript addresses an important open challenge that is relevant for many practical application domains, namely hyperparameter optimization \n- to tackle this challenge, the paper proposes an approach that is novel by combining previous ideas in a new, interesting and fairly principled way\n- the empirical evaluation is comprehensive and convincingly demonstrates the efficacy of the proposed approach as compared to previous methods for hyperparameter optimization\n\n### Cons\n- the proposed approach seems to be fundamentally based on the assumption that the hyperparameters (and thus the induced function) are continuous; the experiments thus also seem to only involve continuous hyperparameters; this appears rather restrictive, given that many practical hyperparameter optimization problems involve important discrete variables (e.g. architectural parameters of neural networks such as the number of layers/units, or categorical parameters such as the choice of optimiser); it would be interesting to know if it would be feasible to straightforwardly extend the proposed approach to also consider such discrete variables; if not, then this seems to be a severe limitation of the method presented; note that there are several black-box/Bayesian optimization approaches (including BOHB) that can handle discrete variables, e.g. [1,2,3]\n- the paper does not provide any source code, and I am not sure if I would be able to implement the method from scratch and fully reproduce the reported results with the details mentioned in the paper alone; that being said, the authors promise to make the code available upon paper acceptance\n- in the first sentence in the introduction, you claim that \"Modern machine learning algorithms usually involve tuning multiple hyperparameters whose size could be from one to thousands.\"; you proceed to mention two examples, SVMs and DNNs, with only a hand full of hyperparameters; it is thus not clear in which scenario one would have _thousands_ of hyperparameters\n- in Sec. 2.1, you claim that \"a lot of references have pointed out that [black-box optimization] can only handle hyperparameters from a few to several dozens\"; however, you only cite one paper for this claim, and it would be beneficial to cite more of the \"lot of references\"; in Sec. 3.4, you again claim that \"black-box optimization methods have a weak performance for the high-dimensional hyperparameter optimization problems which is also verified in a large number of existing references\", proceeding to cite only two papers (which I wouldn't consider \"a large number of existing references\")\n\n[1] Oh et al., 2019, \"Combinatorial bayesian optimization using graph representations\"\n[2] Baptista et al., 2018, \"Bayesian Optimization of Combinatorial Structures\"\n[3] Daxberger et al., 2019, \"Mixed-Variable Bayesian Optimization\"\n\n### Review summary\nIn summary, I believe this is an interesting, enjoyable-to-read paper that tackles an important problem by proposing a promising and effective new technique. This paper will likely be of great interest to some members of the ICLR community and might become an impactful contribution for machine learning researchers and practitioners alike. That being said, I see some issues with the generality of the method as well as the presentation of the paper. As a result, I overall recommend acceptance of this manuscript, although not too enthusiastically.\n\n### Post-rebuttal\nI thank the authors for their response, addressing some of the issues/questions I had raised.\nAfter carefully reading the other reviews (and corresponding author responses), I agree with the valid criticisms raised (some of which I had overlooked initially), which unfortunately further dampened my enthusiasm for this work.\nAs a result, I am slightly lowering my score.\nHowever, I very much appreciate the author's efforts in this very promising work, and hope that they will revise their manuscript to take into account the feedback raised in the reviews.\n\n### Minor issues\n- Fig. 1: the ordering of the compared methods differs between the reported datasets\n- Experiments: the name of the MNIST dataset is typically capitalised\n- there are minor grammatical errors throughout the paper; I recommend that the author thoroughly proof-read the manuscript (or have it proof-read) for language issues \n- the plots are barely readable without zooming in; I strongly recommend the authors to increase the font sizes to improve their readability and clarity", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604775623488}, {"id": "zalpVd49fBl", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1559/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "EDIT: **post rebuttal (could not add a comment readable by authors). Thank you for your rebuttal, I have read it and the other reviews. I agree with the other reviewers that some more baselines for high-dimensional benchmarks are required, and was not convinced by your rebuttal to those requests. I also think the aspect of non-convexity and local optima touched by Reviewer 4 warrants some discussion in the paper. I maintain my score of 4. **\n\nThe paper presents a novel method for hyperparameter optimization based on zeroth order gradient estimation. The zeroth order estimation of the gradient permits the application of the method to complex architectures, provided that the hyperparameters are continuous and the objective function is Lipschitz continuous. The method appears competitive with both gradient-based HPO methods and gradient free methods on a suite of experiments with a number of hyperparameters ranging from 1 to 1250.\n\nStrong points:\n- The method is simple and is competitive with previous gradient-free HPO methods for continuous hyperparameters.\n- The method appears better or competitive with previous gradient-based HPO work.\n- On the technical side the paper appears solid. I see no methodological flaws in the method.\n\nWeak points:\n- There is no comparison with gradient based methods for the deep network experiments (arguably the most important benchmark). The authors state that the methods of Franceschi et al. could not be compared because they require smooth function, but checking their paper, they only mention a smooth optimization operator, and this condition appears to be met by a gradient descent optimization operator. Perhaps I am missing something here, but further justification on this point would be appreciated.\n- The description of the experiments is a bit lacking in certain regards. This diminishes the reproducibility of the work. For instance, the computational budget is not specified for the experiments. How long were benchmarks run for? Was the budget specified in wallclock time? How many function evaluations did each method complete? Did you repeat each experiment and average them? \n- The paper needs a lot of work on a clarity and quality standpoint. I spotted a lot of grammar mistakes and some sentences were hard to interpret.\n\nRecommendation:\n\nGiven the current state of the paper, I would recommend rejection. The paper needs some serious copy-editing before it is up to the standards of a first tier conference such as ICLR. I should say that I like the method proposed in the paper, it would make a nice addition to the literature on HPO -- I just think it needs more polish. \n\nApart from paper quality, more work/details on the following points/questions would help me increase my score for this paper:\n\n- The results are only presented using wallclock time. A general idea of the  \"number of function evaluations\" achieved by each method would also be informative, a graphic comparing the performance with regards to function evaluations could be added in the appendix. This is an important point because some models take a long time to run, and whatever computation is done between function evaluations becomes meaningless in comparison.\n- What is the implication of removing the approximate gradient computation of HOAG with the zeroth order gradient proposed in this paper? Is it still HOAG or does it become a new method?\n- Figure 2. is a bit confusing to me. Why is the x-axis \"Inner epoch\" for the suboptimality figures? Is this presenting the final optimization run with the best hyperparameters found?\n\nExtra comments:\n\nTable 1 feels rather arbitrary, what is the threshold to say that a method is \"effective\" or not? Simple or not? Is BOHB really that simple compared to HOAG? Complexity / optimization guarantees in terms of p and d might provide a clearer picture here (those are generally available for GP-BO and most likely for gradient based methods?) -- just an idea.\n\nSection 2.1: Claim that the number of hyperparameters in real problems would range from hundreds to thousands. It seems to me that problems with thousands of hyperparameters are very academic and never encountered in typical real world problems. Bring forward examples or please reformulate.\n\nSection 2.1: I have mixed feelings about the last paragraph of this section titled Enlightenment with a cross symbol. Is this a religious reference? I would argue this has no place in a scientific paper. \n\nProofreading is required, examples of errors:\n\nPage 2. paragraph 2: salable -> scalable\nPage 3. (such as the deep neural network) -> (such as A deep neural network)\n\nMultiple times in the text: hyperparmeters -> hyperparAmeters", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method, paper needs more work", "review": "EDIT: **post rebuttal (could not add a comment readable by authors). Thank you for your rebuttal, I have read it and the other reviews. I agree with the other reviewers that some more baselines for high-dimensional benchmarks are required, and was not convinced by your rebuttal to those requests. I also think the aspect of non-convexity and local optima touched by Reviewer 4 warrants some discussion in the paper. I maintain my score of 4. **\n\nThe paper presents a novel method for hyperparameter optimization based on zeroth order gradient estimation. The zeroth order estimation of the gradient permits the application of the method to complex architectures, provided that the hyperparameters are continuous and the objective function is Lipschitz continuous. The method appears competitive with both gradient-based HPO methods and gradient free methods on a suite of experiments with a number of hyperparameters ranging from 1 to 1250.\n\nStrong points:\n- The method is simple and is competitive with previous gradient-free HPO methods for continuous hyperparameters.\n- The method appears better or competitive with previous gradient-based HPO work.\n- On the technical side the paper appears solid. I see no methodological flaws in the method.\n\nWeak points:\n- There is no comparison with gradient based methods for the deep network experiments (arguably the most important benchmark). The authors state that the methods of Franceschi et al. could not be compared because they require smooth function, but checking their paper, they only mention a smooth optimization operator, and this condition appears to be met by a gradient descent optimization operator. Perhaps I am missing something here, but further justification on this point would be appreciated.\n- The description of the experiments is a bit lacking in certain regards. This diminishes the reproducibility of the work. For instance, the computational budget is not specified for the experiments. How long were benchmarks run for? Was the budget specified in wallclock time? How many function evaluations did each method complete? Did you repeat each experiment and average them? \n- The paper needs a lot of work on a clarity and quality standpoint. I spotted a lot of grammar mistakes and some sentences were hard to interpret.\n\nRecommendation:\n\nGiven the current state of the paper, I would recommend rejection. The paper needs some serious copy-editing before it is up to the standards of a first tier conference such as ICLR. I should say that I like the method proposed in the paper, it would make a nice addition to the literature on HPO -- I just think it needs more polish. \n\nApart from paper quality, more work/details on the following points/questions would help me increase my score for this paper:\n\n- The results are only presented using wallclock time. A general idea of the  \"number of function evaluations\" achieved by each method would also be informative, a graphic comparing the performance with regards to function evaluations could be added in the appendix. This is an important point because some models take a long time to run, and whatever computation is done between function evaluations becomes meaningless in comparison.\n- What is the implication of removing the approximate gradient computation of HOAG with the zeroth order gradient proposed in this paper? Is it still HOAG or does it become a new method?\n- Figure 2. is a bit confusing to me. Why is the x-axis \"Inner epoch\" for the suboptimality figures? Is this presenting the final optimization run with the best hyperparameters found?\n\nExtra comments:\n\nTable 1 feels rather arbitrary, what is the threshold to say that a method is \"effective\" or not? Simple or not? Is BOHB really that simple compared to HOAG? Complexity / optimization guarantees in terms of p and d might provide a clearer picture here (those are generally available for GP-BO and most likely for gradient based methods?) -- just an idea.\n\nSection 2.1: Claim that the number of hyperparameters in real problems would range from hundreds to thousands. It seems to me that problems with thousands of hyperparameters are very academic and never encountered in typical real world problems. Bring forward examples or please reformulate.\n\nSection 2.1: I have mixed feelings about the last paragraph of this section titled Enlightenment with a cross symbol. Is this a religious reference? I would argue this has no place in a scientific paper. \n\nProofreading is required, examples of errors:\n\nPage 2. paragraph 2: salable -> scalable\nPage 3. (such as the deep neural network) -> (such as A deep neural network)\n\nMultiple times in the text: hyperparmeters -> hyperparAmeters", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603845761655}, {"id": "W31NCGvMldd", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1559/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "EDIT: **post rebuttal. I'd like to thank the authors for their response. As scalability to high-dimensional hyperparameter spaces is presented as a key advantage of the method, direct comparisons to high-dimensional BO techniques would be needed. The fact that using trust regions could benefit HOZOG, or that HOZOG is a better strategy compared to TurBO and REMBO, should be demonstrated empirically. I am keeping my score to 4 as the current positioning of the paper would require direct comparisons with these baselines. **\n\n\nThe paper introduces HOZOG, a new HPO method combining the benefits of black-box and gradient-based optimization. This is achieved by means of a finite difference approximation that computes the gradients in a black-box fashion. By doing so, the method achieves both scalability to a large number of model parameters and hyperparameters, the latter being a bottleneck in standard BO methods. A feasibility analysis and experiments against baselines show the benefits of the proposed approach.\n\nPositive\n\n1. **Significance.** This is the first approach trying to take a black-box approach to tackle gradient-based optimization, inheriting the benefits of the two worlds.
 I expect this can lead to follow-up work in this direction.\n2. **Clarity.** The paper is generally well articulated and easy to follow. I particularly appreciated the authors clearly comparing black-box and gradient based methods in Table 1 by listing desiderata.
\n3. **Feasibility analysis.** The work contributes a rigorous feasibility analysis studying the use of the proposed approach in HPO.\n\nNegative\n\n1. **Missing baselines.** If scalability to high-dimensional hyperparameter spaces is a key selling point of the method, experiments should have focused on comparing against BO methods that are designd to scale to large dimensions. For instance, the authors claim \"a lot of references have pointed out that [BO] can only handle hyperparameters from a few to several dozens\". It would be helpful if the authors elaborated on this point. Is it due to the poor scalability of the Gaussian process model in high dimensions? In that case, non-GP, scalable models have been proposed (e.g., SMAC) or adaptations directly targeting high-dimensional BO problems (e.g., TurBO). Code for these methods is also publicly available but these baselines are missing from the experiments. These are specific references:\n\n(TurBO) Eriksson et al., Scalable Global Optimization via Local Bayesian Optimization. In NeurIPS 2019; \n\n
(REMBO) Wang et al., Bayesian Optimization in High Dimensions via Random Embeddings, In IJCAI '13 + follow-ups, including Letham et al. 2020, \"Re-Examining Linear Embeddings for High-Dimensional Bayesian Optimization\"
;\n\n(SMAC) Hutter et al., 2011. Sequential Model-Based Optimization for General Algorithm Configuration.  In Proceedings of the conference on Learning and Intelligent Optimization, 2011.\n\n2. **Reproducibility.** No curves in the results include error bars. Are experiments only based on a single run? If not, clearly defined error bars should be reported. If yes, this is not enough to draw meaningful conclusions considering that the BO algortihms are highly stochastic with many sources of variability (e.g., how they are warm-started, often done through a random initial design).\n\n3. **Structure.** The paper is visually rather crammed and the plots are too small to be readable. The authors seem to be aware of this as they provide larger versions of the figures in the appendix. The main paper should be self-contained though and this is not an acceptable way to get around the 8-page limit.\n\nWhile the paper has its merits, I believe the current version falls below the acceptance bar due to the lack of competing high-dimensional BO baselines paired with reproducibility concerns (i.e., experiments have no error bars). I believe this is promising work with valuable theoretical contributions, but the experimental evaluation is currently not sufficient to justify the proposed approach. In particular, relevant baselines for high-dimensional BO should be both discussed and compared against, and all results should be based on multiple repetitions and report confidence intervals.\n\nMinor:\n\n1. Typos at Page 2: \"salable\"  ---> scalable; \"hyperparmaters\" --> hyperparameters\n2. Related to the \"structure\" point above, there is no space between \"results and discussion\" and the rest of the text in page 8. This is also done in several places, including captions. I'd suggest making the content more crisp instead.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "While the paper aims to combine the benefits of gradient-based and black-box optimization, it does not compare to relevant high-dimensional BO baselines.", "review": "EDIT: **post rebuttal. I'd like to thank the authors for their response. As scalability to high-dimensional hyperparameter spaces is presented as a key advantage of the method, direct comparisons to high-dimensional BO techniques would be needed. The fact that using trust regions could benefit HOZOG, or that HOZOG is a better strategy compared to TurBO and REMBO, should be demonstrated empirically. I am keeping my score to 4 as the current positioning of the paper would require direct comparisons with these baselines. **\n\n\nThe paper introduces HOZOG, a new HPO method combining the benefits of black-box and gradient-based optimization. This is achieved by means of a finite difference approximation that computes the gradients in a black-box fashion. By doing so, the method achieves both scalability to a large number of model parameters and hyperparameters, the latter being a bottleneck in standard BO methods. A feasibility analysis and experiments against baselines show the benefits of the proposed approach.\n\nPositive\n\n1. **Significance.** This is the first approach trying to take a black-box approach to tackle gradient-based optimization, inheriting the benefits of the two worlds.
 I expect this can lead to follow-up work in this direction.\n2. **Clarity.** The paper is generally well articulated and easy to follow. I particularly appreciated the authors clearly comparing black-box and gradient based methods in Table 1 by listing desiderata.
\n3. **Feasibility analysis.** The work contributes a rigorous feasibility analysis studying the use of the proposed approach in HPO.\n\nNegative\n\n1. **Missing baselines.** If scalability to high-dimensional hyperparameter spaces is a key selling point of the method, experiments should have focused on comparing against BO methods that are designd to scale to large dimensions. For instance, the authors claim \"a lot of references have pointed out that [BO] can only handle hyperparameters from a few to several dozens\". It would be helpful if the authors elaborated on this point. Is it due to the poor scalability of the Gaussian process model in high dimensions? In that case, non-GP, scalable models have been proposed (e.g., SMAC) or adaptations directly targeting high-dimensional BO problems (e.g., TurBO). Code for these methods is also publicly available but these baselines are missing from the experiments. These are specific references:\n\n(TurBO) Eriksson et al., Scalable Global Optimization via Local Bayesian Optimization. In NeurIPS 2019; \n\n
(REMBO) Wang et al., Bayesian Optimization in High Dimensions via Random Embeddings, In IJCAI '13 + follow-ups, including Letham et al. 2020, \"Re-Examining Linear Embeddings for High-Dimensional Bayesian Optimization\"
;\n\n(SMAC) Hutter et al., 2011. Sequential Model-Based Optimization for General Algorithm Configuration.  In Proceedings of the conference on Learning and Intelligent Optimization, 2011.\n\n2. **Reproducibility.** No curves in the results include error bars. Are experiments only based on a single run? If not, clearly defined error bars should be reported. If yes, this is not enough to draw meaningful conclusions considering that the BO algortihms are highly stochastic with many sources of variability (e.g., how they are warm-started, often done through a random initial design).\n\n3. **Structure.** The paper is visually rather crammed and the plots are too small to be readable. The authors seem to be aware of this as they provide larger versions of the figures in the appendix. The main paper should be self-contained though and this is not an acceptable way to get around the 8-page limit.\n\nWhile the paper has its merits, I believe the current version falls below the acceptance bar due to the lack of competing high-dimensional BO baselines paired with reproducibility concerns (i.e., experiments have no error bars). I believe this is promising work with valuable theoretical contributions, but the experimental evaluation is currently not sufficient to justify the proposed approach. In particular, relevant baselines for high-dimensional BO should be both discussed and compared against, and all results should be based on multiple repetitions and report confidence intervals.\n\nMinor:\n\n1. Typos at Page 2: \"salable\"  ---> scalable; \"hyperparmaters\" --> hyperparameters\n2. Related to the \"structure\" point above, there is no space between \"results and discussion\" and the rest of the text in page 8. This is also done in several places, including captions. I'd suggest making the content more crisp instead.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603476165954}, {"id": "zDInk0BO4zu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1559/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes to apply zeroth-order optimization to hyperparameter tuning. Zeroth-order optimization techniques use function evaluation only to approximate the gradients, thus they are applicable to black-box functions for which the gradient is difficult to compute. For hyperparameter tuning this is the case in general. Existing hypergradient-based methods are not applicable to generic tuning tasks, and have scalability issue. Existing generic methods such as Bayesian optimization have weak scalability in number of hyperparameters. I agree with these observations and the goal of having a flexible, efficient, effective, simple and scalable solution.\n\nWhile the idea of using zeroth-order optimization for HPO makes sense, the paper has not reached a sufficient depth in exploring this idea. The paper proposes to use (Nesterov & Spokoiny 2017) without modification. There are several issues: (1) The paper concludes in page 5 that \"Because the A-based constrained optimization problem f($\\lambda$) is continuous, we\ncan use the zeroth-order hyper-gradient technique to optimize f($\\lambda$) (Nesterov & Spokoiny, 2017).\nNesterov & Spokoiny (2017) provided the convergence guarantee of zeroth-order hyper-gradient\nmethod when f($\\lambda$) is Lipschitz continuous as defined in Definition 3.\" With continuity only, the algorithm can converge to local stationary points. Straightforward application of the algorithm could violate the 'effectiveness' and 'efficiency' goal. The situation is worsened in the high-dim case when the variance of the gradient approximation is large. The experiment in Sec 3.3 supports this concern, where HOZOG underperforms REV in the long run. (2) The algorithm still has several hyperparameters to tune, including q, $\\lambda$ and $\\mu$. From the experiments reported in the paper, each task uses different values of q, $\\lambda$ and $\\mu$. That raises the question about the practicability, especially, if one needs to tune these hyperparameters, whether it can still be efficient, scalable and simple. The paper has not discussed how to tune them for a given task. The procedure of tuning is required to be discussed because 'flexibility' is indeed a goal of this paper. (3) The same question can be asked about how the initial point of hyperparameters is chosen, and the paper does not discuss that. \n\nIn addition to these issues, the paper makes a strong claim \"for\nhigh-dimensional hyperparameter optimization problems which have no customized RFHO algorithm,\nHOZOG currently is the only choice for this kind of problems to the best of our knowledge.\" in the end of Sec 3.4. For example, there are other methods designed for high-dimensional problems:\n\nDavid Eriksson et al., Scalable Global Optimization via Local Bayesian Optimization, NeurIPS'19.\n\nN. Hansen. The CMA evolution strategy: A comparing review. In Towards a New Evolutionary Computation, pages 75–102. Springer, 2006.\n\nM. J. Powell. A view of algorithms for optimization without derivatives. Mathematics Today-Bulletin of the Institute of Mathematics and its Applications, 43(5):170–174, 2007.\n\nSince the paper emphasizes the importance of HOZOG in the high-dim scenario, some of these stronger baselines need to used. \n\nA final comment is that the phrase 'automated learning algorithm' in the title is not very informative, especially as the selection of hyperparameters of HOZOG is not automated.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A premature attempt of applying zeroth-order optimization to high-dimensional HPO", "review": "This work proposes to apply zeroth-order optimization to hyperparameter tuning. Zeroth-order optimization techniques use function evaluation only to approximate the gradients, thus they are applicable to black-box functions for which the gradient is difficult to compute. For hyperparameter tuning this is the case in general. Existing hypergradient-based methods are not applicable to generic tuning tasks, and have scalability issue. Existing generic methods such as Bayesian optimization have weak scalability in number of hyperparameters. I agree with these observations and the goal of having a flexible, efficient, effective, simple and scalable solution.\n\nWhile the idea of using zeroth-order optimization for HPO makes sense, the paper has not reached a sufficient depth in exploring this idea. The paper proposes to use (Nesterov & Spokoiny 2017) without modification. There are several issues: (1) The paper concludes in page 5 that \"Because the A-based constrained optimization problem f($\\lambda$) is continuous, we\ncan use the zeroth-order hyper-gradient technique to optimize f($\\lambda$) (Nesterov & Spokoiny, 2017).\nNesterov & Spokoiny (2017) provided the convergence guarantee of zeroth-order hyper-gradient\nmethod when f($\\lambda$) is Lipschitz continuous as defined in Definition 3.\" With continuity only, the algorithm can converge to local stationary points. Straightforward application of the algorithm could violate the 'effectiveness' and 'efficiency' goal. The situation is worsened in the high-dim case when the variance of the gradient approximation is large. The experiment in Sec 3.3 supports this concern, where HOZOG underperforms REV in the long run. (2) The algorithm still has several hyperparameters to tune, including q, $\\lambda$ and $\\mu$. From the experiments reported in the paper, each task uses different values of q, $\\lambda$ and $\\mu$. That raises the question about the practicability, especially, if one needs to tune these hyperparameters, whether it can still be efficient, scalable and simple. The paper has not discussed how to tune them for a given task. The procedure of tuning is required to be discussed because 'flexibility' is indeed a goal of this paper. (3) The same question can be asked about how the initial point of hyperparameters is chosen, and the paper does not discuss that. \n\nIn addition to these issues, the paper makes a strong claim \"for\nhigh-dimensional hyperparameter optimization problems which have no customized RFHO algorithm,\nHOZOG currently is the only choice for this kind of problems to the best of our knowledge.\" in the end of Sec 3.4. For example, there are other methods designed for high-dimensional problems:\n\nDavid Eriksson et al., Scalable Global Optimization via Local Bayesian Optimization, NeurIPS'19.\n\nN. Hansen. The CMA evolution strategy: A comparing review. In Towards a New Evolutionary Computation, pages 75–102. Springer, 2006.\n\nM. J. Powell. A view of algorithms for optimization without derivatives. Mathematics Today-Bulletin of the Institute of Mathematics and its Applications, 43(5):170–174, 2007.\n\nSince the paper emphasizes the importance of HOZOG in the high-dim scenario, some of these stronger baselines need to used. \n\nA final comment is that the phrase 'automated learning algorithm' in the title is not very informative, especially as the selection of hyperparameters of HOZOG is not automated.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1602555748489}], "openreview_url": "https://openreview.net/forum?id=0naHZ3gZSzo", "arxiv_id": "2102.09026", "paper_pdf": "papers/0naHZ3gZSzo.pdf", "paper_pdf_sha256": "3db76b844f4d1092b9015f00dab01254223ced930aefa697222d43f3b1a019dd", "paper_pdf_bytes": 684065, "paper_pdf_source": "openreview", "code_url": "https://github.com/jsgubin/HOZOG", "code_repository": "jsgubin/HOZOG", "code_commit": "cae6ac386d1b43c70d269e47e10ba4a4ac7aed4a", "code_archive": "repos/0naHZ3gZSzo.zip", "code_archive_sha256": "b4aafcc5502a92ff03e97863c59fe0d1fa0e6c95203e46edc9ac96484aefec54", "code_archive_bytes": 278996, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 336, "github_languages": {"Python": 158062}, "github_archived": false, "github_pushed_at": "2021-04-12T07:09:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimizing-large-scale-hyperparameters-via-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1l-5pEtDr", "year": 2020, "status": "rejected", "title": "AdaX: Adaptive Gradient Descent with Exponential Long Term Memory", "authors": ["Wenjie Li", "Zhaoyang Zhang", "Xinjiang Wang", "Ping Luo"], "authorids": ["li3549@purdue.edu", "zhaoyangzhang@link.cuhk.edu.hk", "swanxinjiang@gmail.com", "pluo.lhi@gmail.com"], "authors_source": "OpenReview API", "abstract": "Adaptive optimization algorithms such as RMSProp and Adam have fast convergence and smooth learning process. Despite their successes, they are proven to have non-convergence issue even in convex optimization problems as well as weak performance compared with the first order gradient methods such as stochastic gradient descent (SGD). Several other algorithms, for example AMSGrad and AdaShift, have been proposed to alleviate these issues but only minor effect has been observed. This paper further analyzes the performance of such algorithms in a non-convex setting by extending their non-convergence issue into a simple non-convex case and show that Adam's design of update steps would possibly lead the algorithm to local minimums. To address the above problems, we propose a novel adaptive gradient descent algorithm, named AdaX, which accumulates the long-term past gradient information exponentially. We prove the convergence of AdaX in both convex and non-convex settings. Extensive experiments show that AdaX outperforms Adam in various tasks of computer vision and natural language processing and can catch up with SGD.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "BylElFr-cH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper697/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper points out that existing adaptive methods (especially for the methods designing second-order momentum estimates in a exponentially moving average fashion) do not consider gradient decrease information and this might lead to suboptimal convergences via simple non-cvx toy example. \nBased on this observation, the authors provide a novel optimization algorithm for long-term memory of past gradients by modifying second-order momentum design. Also, they provide aconvex regret analysis and convergence analysis for non-convex optimization. Finally, the authors evaluate their methods on various deep learning problems.\n\nSignificance/Novelty: While there have been many studies on non-convergence of Adam, raising an issue on ignoring the gradient decrease information seems novel. \n\nPros:\n1. The motivating toy example in Section 3 is useful for readers to get intuitions.\n\n2. By introducing long-term memory on past-gradients, the authors fix the Adam's issues and they can also improve the convergence rate in a non-convex optimization (Corollary 4.2).\n\n3. Empirical studies show superiority to original Adam (Section 5).\n\nCons:\nWhile they provide a significant study on Adam's failure and a novel optimization algorithm, I have several concerns:\n1. What is default hyperparameters for AdaX in Algorithm 1? Is it the same as AdaX-W (Algorithm 3) in Appendix? The bias correction term in the line 7 of Algorithm 1 will be very large even with small $\\beta_2$ since it is expoential (For example, (1 + 0.0001)^(100000) ~ 20000 for $\\beta_2$ = 10^(-4)). So, it is not clear that the second momentum estimate of Ada-X is really stable. For this, it would be interesting to see how the trajectories of second-order momentum estimates of Adam, AMSGrad, Ada-X are different. I think this will help to understand Ada-X better.\n\n2. In terms of theory, I think the Lemma 4.1 is inevitable for convergence guarantees in Theorem 4.1 and Theorem 4.2. Although the authors effectively remove log T in the numerator in Corollary 3.2 of Chen et al. (2019) using their lemma 4.1 (I think this is the key point), the assumption that $\\beta_{2t} = \\beta_2 / t$ seems quite strong, and original Adam paper has no such assumptions. For a real deep learning problems such as training ResNet on CIFAR-10, the $\\beta_{2t}$ is almost zero after even one or two epochs where Ada-X behaves like vanilla SGD. Is there no room for relaxing this assumption such as $\\beta_{2t} = \\beta_2 / \\sqrt{t}$? Also, it is not clear how the authors derive Corollary 4.2 from Theorem 4.2 since Theorem 4.2 assumes $\\beta_{2t} = \\beta_2 / t$ while Corollary 4.2 does not.\n\n3. In the experiment, it is not clear that the authors use the same strategy for constructing first-order momentum for Adam with a newly introduced parameter $\\beta_3$. In other words, the authors should use the same policy on constructing the first-order momentum estimate for both Adam and Ada-X. Also, as the authors add an additional hyperparameter $\\beta_3$, the effect of $\\beta_3$ on performance should be discussed at least empirically.\n\n4. There are many studies on fixing poor generalization of adaptive methods (such as AdaBound which the authors cited). In this context, Zaheer et al. (2018, Adaptive methods for non-convex optimization) propose a large epsilon value (numerical stability parameter) such as $\\epsilon = 10^{-3}$ for better generalization. It will be more interesting to see the comparisons in this regime.\n\n5. In my experience with Adam-W (Decoupled weight decay regularization), Adam-W requires a relatively large weight decay parameter $\\lambda$.\nAs an example, DenseNet-BC-100-12 shows a similar validation accuracy with Adam-W $\\lambda = 0.05$ under the learning rate scheduling in (Huang et al. 2016, DenseNet)  as vanilla SGD.\nTherefore, the authors should consider more broader range of weight decay parameters for at least image classification tasks.\n\nMinor:\n1. In eq (2), the domain of x should be mentioned: according to Reddie et al, it is [-1,1].\n2. In both theorem 4.1 and corollary 4.1, $D_{\\infty^2}$ should be $D_{\\infty}^2$?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper points out that existing adaptive methods (especially for the methods designing second-order momentum estimates in a exponentially moving average fashion) do not consider gradient decrease information and this might lead to suboptimal convergences via simple non-cvx toy example. \nBased on this observation, the authors provide a novel optimization algorithm for long-term memory of past gradients by modifying second-order momentum design. Also, they provide aconvex regret analysis and convergence analysis for non-convex optimization. Finally, the authors evaluate their methods on various deep learning problems.\n\nSignificance/Novelty: While there have been many studies on non-convergence of Adam, raising an issue on ignoring the gradient decrease information seems novel. \n\nPros:\n1. The motivating toy example in Section 3 is useful for readers to get intuitions.\n\n2. By introducing long-term memory on past-gradients, the authors fix the Adam's issues and they can also improve the convergence rate in a non-convex optimization (Corollary 4.2).\n\n3. Empirical studies show superiority to original Adam (Section 5).\n\nCons:\nWhile they provide a significant study on Adam's failure and a novel optimization algorithm, I have several concerns:\n1. What is default hyperparameters for AdaX in Algorithm 1? Is it the same as AdaX-W (Algorithm 3) in Appendix? The bias correction term in the line 7 of Algorithm 1 will be very large even with small $\\beta_2$ since it is expoential (For example, (1 + 0.0001)^(100000) ~ 20000 for $\\beta_2$ = 10^(-4)). So, it is not clear that the second momentum estimate of Ada-X is really stable. For this, it would be interesting to see how the trajectories of second-order momentum estimates of Adam, AMSGrad, Ada-X are different. I think this will help to understand Ada-X better.\n\n2. In terms of theory, I think the Lemma 4.1 is inevitable for convergence guarantees in Theorem 4.1 and Theorem 4.2. Although the authors effectively remove log T in the numerator in Corollary 3.2 of Chen et al. (2019) using their lemma 4.1 (I think this is the key point), the assumption that $\\beta_{2t} = \\beta_2 / t$ seems quite strong, and original Adam paper has no such assumptions. For a real deep learning problems such as training ResNet on CIFAR-10, the $\\beta_{2t}$ is almost zero after even one or two epochs where Ada-X behaves like vanilla SGD. Is there no room for relaxing this assumption such as $\\beta_{2t} = \\beta_2 / \\sqrt{t}$? Also, it is not clear how the authors derive Corollary 4.2 from Theorem 4.2 since Theorem 4.2 assumes $\\beta_{2t} = \\beta_2 / t$ while Corollary 4.2 does not.\n\n3. In the experiment, it is not clear that the authors use the same strategy for constructing first-order momentum for Adam with a newly introduced parameter $\\beta_3$. In other words, the authors should use the same policy on constructing the first-order momentum estimate for both Adam and Ada-X. Also, as the authors add an additional hyperparameter $\\beta_3$, the effect of $\\beta_3$ on performance should be discussed at least empirically.\n\n4. There are many studies on fixing poor generalization of adaptive methods (such as AdaBound which the authors cited). In this context, Zaheer et al. (2018, Adaptive methods for non-convex optimization) propose a large epsilon value (numerical stability parameter) such as $\\epsilon = 10^{-3}$ for better generalization. It will be more interesting to see the comparisons in this regime.\n\n5. In my experience with Adam-W (Decoupled weight decay regularization), Adam-W requires a relatively large weight decay parameter $\\lambda$.\nAs an example, DenseNet-BC-100-12 shows a similar validation accuracy with Adam-W $\\lambda = 0.05$ under the learning rate scheduling in (Huang et al. 2016, DenseNet)  as vanilla SGD.\nTherefore, the authors should consider more broader range of weight decay parameters for at least image classification tasks.\n\nMinor:\n1. In eq (2), the domain of x should be mentioned: according to Reddie et al, it is [-1,1].\n2. In both theorem 4.1 and corollary 4.1, $D_{\\infty^2}$ should be $D_{\\infty}^2$?"}, "tcdate": 1572063468007}, {"id": "rkeCnYcy9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper697/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a new adaptive gradient descent algorithm with exponential long term memory. The authors analyzed the non-convergence issue in Adam into a simple non-convex case. The authors also presented the convergence of the proposed AdaX in both convex and non-convex settings.\n\n- The proposed algorithm revisited the non-convergence issue in Adam and proposed a new algorithm design to try to address this issue. However, the new algorithm design is a bit strange to me, especially in Line 6 of Algorithm 2, the authors proposed to update v_t by (1+ \\beta_2) v_{t-1} + \\beta_2 g_t^2, where normally people would use (1-\\beta2) and \\beta2 as the coefficients. I am not quite get the intuition of using such as strange design. I wonder if the authors could further explain that.\n\n- The authors also add change \\beta_1 in Line 5 of Algorithm 2 into \\beta_3. And then in theory, the authors again choose \\beta_3 as \\beta_1. It seems that \\beta_3 is not contributing to any theoretical result. It seems to me to just have another parameter to tune in order to get better performances. Can the authors gives more justification on why introducing such a term here? And also show how the different choice of \\beta_3 affects the final result? \n\n- Missing some important references closely related to this paper:\n\nChen, Jinghui, and Quanquan Gu. \"Closing the generalization gap of adaptive gradient methods in training deep neural networks.\" arXiv preprint arXiv:1806.06763 (2018).\nZaheer, Manzil, et al. \"Adaptive methods for nonconvex optimization.\" Advances in Neural Information Processing Systems. 2018.\nZhou, Dongruo, et al. \"On the convergence of adaptive gradient methods for nonconvex optimization.\" arXiv preprint arXiv:1808.05671 (2018).\n\nI would suggest the authors to also compare with the above mentioned baselines to better demonstrate its performances and theoretical results.\n\n- The authors include theoretical analysis in both convex and non-convex settings, which is appreciated, however, the theoretical result seems to show similar convergence guarantees with AMSGrad. I wonder if the authors could provide theoretical justifications on why the proposed method is better than prior arts, probably some sharper convergences or some generalization guarantees?\n\n- In the experiments part, I wonder why the authors did not compare with AMSGrad, RMSProp in later parts such as ImageNet, IoU and RNN parts? I makes no sense to drop them for those experiments. Also, are the authors fully tuned the hyper-parameters for other baselines such as step size and weight decay on SGDM?\n\n================\nafter the rebuttal\n\nI thank the authors for their response but I still feel that the intuition of this paper is not clear enough and comparison with more baselines is needed. Therefore I decided to keep my score unchanged.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #4", "review": "This paper proposed a new adaptive gradient descent algorithm with exponential long term memory. The authors analyzed the non-convergence issue in Adam into a simple non-convex case. The authors also presented the convergence of the proposed AdaX in both convex and non-convex settings.\n\n- The proposed algorithm revisited the non-convergence issue in Adam and proposed a new algorithm design to try to address this issue. However, the new algorithm design is a bit strange to me, especially in Line 6 of Algorithm 2, the authors proposed to update v_t by (1+ \\beta_2) v_{t-1} + \\beta_2 g_t^2, where normally people would use (1-\\beta2) and \\beta2 as the coefficients. I am not quite get the intuition of using such as strange design. I wonder if the authors could further explain that.\n\n- The authors also add change \\beta_1 in Line 5 of Algorithm 2 into \\beta_3. And then in theory, the authors again choose \\beta_3 as \\beta_1. It seems that \\beta_3 is not contributing to any theoretical result. It seems to me to just have another parameter to tune in order to get better performances. Can the authors gives more justification on why introducing such a term here? And also show how the different choice of \\beta_3 affects the final result? \n\n- Missing some important references closely related to this paper:\n\nChen, Jinghui, and Quanquan Gu. \"Closing the generalization gap of adaptive gradient methods in training deep neural networks.\" arXiv preprint arXiv:1806.06763 (2018).\nZaheer, Manzil, et al. \"Adaptive methods for nonconvex optimization.\" Advances in Neural Information Processing Systems. 2018.\nZhou, Dongruo, et al. \"On the convergence of adaptive gradient methods for nonconvex optimization.\" arXiv preprint arXiv:1808.05671 (2018).\n\nI would suggest the authors to also compare with the above mentioned baselines to better demonstrate its performances and theoretical results.\n\n- The authors include theoretical analysis in both convex and non-convex settings, which is appreciated, however, the theoretical result seems to show similar convergence guarantees with AMSGrad. I wonder if the authors could provide theoretical justifications on why the proposed method is better than prior arts, probably some sharper convergences or some generalization guarantees?\n\n- In the experiments part, I wonder why the authors did not compare with AMSGrad, RMSProp in later parts such as ImageNet, IoU and RNN parts? I makes no sense to drop them for those experiments. Also, are the authors fully tuned the hyper-parameters for other baselines such as step size and weight decay on SGDM?\n\n================\nafter the rebuttal\n\nI thank the authors for their response but I still feel that the intuition of this paper is not clear enough and comparison with more baselines is needed. Therefore I decided to keep my score unchanged.  ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571953077526}, {"id": "BJluZCr19H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper697/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a new adaptive gradient algorithm AdaX, which the authors claim results in better convergence and generalization properties compared to previous adaptive gradient methods. The paper overall is fairly clear, although the writing can be improved in places.\n\nSection 3 is interesting, although calling the section the \"nonconvergence of Adam\" is a bit misleading, since the algorithm does converge to a local minimum.\n\nI have some concerns about the rest of the paper however. I am a bit confused about the changes proposed to Adam that gives rise to the AdaX algorithm. \n\n1. Won't replacing beta2 with 1+beta2 keep increasing the magnitude of the denominator of the algorithm just like AdaGrad does? In that case, is the algorithm expected to work better when having sparse gradients?\n\n2. I also do not quite understand how to interpret using a separate hyperparameter for the momentum term (ie the beta3 hyperparameter that is introduced). How is beta1 and beta3 related? The numerator loses the interpretation of momentum, i.e., averaged past gradients, when using a separate beta3 parameter, and this does not feel like a principled change.\n\nI have a number of questions about the experiments as well, which makes it hard for me to interpret the significance of the empirical results presented:\n\n1. What is the minibatch size used? Do any of the conclusions presented change if the minibatch size is changed?\n\n2. Is the learning rate tuned? The authors mention the initial learning rate used for the experiments, but it is not clear why those values are used? Was the same initial learning rate used for all algorithms?\n\n3. Were the beta1, beta2 and beta3 values tuned for AdaX? What about beta1 and beta2 for Adam? What are the optimal values of these parameters that were observed?\n\n4. How sensitive is performance to the values of these hyperparameters?\n\nOverall I think this work requires quite a bit of work before it is ready for publication, and would benefit from a much more thorough empirical evaluation of the algorithm.\n\n==================================\n\nEdit after rebutall:\nI thank the authors for their response. While the paper has definitely improved in the newer draft, after having read the other reviews and the updated draft, I believe the paper still requires a bit of work before being ready for publication. I am sticking to my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "In this paper, the authors propose a new adaptive gradient algorithm AdaX, which the authors claim results in better convergence and generalization properties compared to previous adaptive gradient methods. The paper overall is fairly clear, although the writing can be improved in places.\n\nSection 3 is interesting, although calling the section the \"nonconvergence of Adam\" is a bit misleading, since the algorithm does converge to a local minimum.\n\nI have some concerns about the rest of the paper however. I am a bit confused about the changes proposed to Adam that gives rise to the AdaX algorithm. \n\n1. Won't replacing beta2 with 1+beta2 keep increasing the magnitude of the denominator of the algorithm just like AdaGrad does? In that case, is the algorithm expected to work better when having sparse gradients?\n\n2. I also do not quite understand how to interpret using a separate hyperparameter for the momentum term (ie the beta3 hyperparameter that is introduced). How is beta1 and beta3 related? The numerator loses the interpretation of momentum, i.e., averaged past gradients, when using a separate beta3 parameter, and this does not feel like a principled change.\n\nI have a number of questions about the experiments as well, which makes it hard for me to interpret the significance of the empirical results presented:\n\n1. What is the minibatch size used? Do any of the conclusions presented change if the minibatch size is changed?\n\n2. Is the learning rate tuned? The authors mention the initial learning rate used for the experiments, but it is not clear why those values are used? Was the same initial learning rate used for all algorithms?\n\n3. Were the beta1, beta2 and beta3 values tuned for AdaX? What about beta1 and beta2 for Adam? What are the optimal values of these parameters that were observed?\n\n4. How sensitive is performance to the values of these hyperparameters?\n\nOverall I think this work requires quite a bit of work before it is ready for publication, and would benefit from a much more thorough empirical evaluation of the algorithm.\n\n==================================\n\nEdit after rebutall:\nI thank the authors for their response. While the paper has definitely improved in the newer draft, after having read the other reviews and the updated draft, I believe the paper still requires a bit of work before being ready for publication. I am sticking to my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571933696076}, {"id": "BkenmE9a_r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper697/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new step-size adaptation algorithm called AdaX. AdaX \nbuilds on the ideas of the Adam algorithm to address instability and non-convergence issues. \nConvergence of AdaX is proven in both convex and non-convex settings. The paper \nalso provides an empirical comparison of AdaX against its predecessors \n(SGD, RMSProp, Adam, AMSGrad) on a variety of tasks.\n\nI recommend the paper be rejected. I believe the convergence results could be a significant contribution, but the \nquality of the paper is hampered by its experimental design. The paper felt generally unpolished, containing\nfrequent grammatical errors, imprecise language, and uncited statements.\n\nMy main issue with the paper is the experimental design. I am not convinced that we can \ndraw valid conclusions from the experimental results for the following reasons:\n  - The experiments are lacking important details. How many independent runs of the experiment \n    were the experimental results averaged over? All of the experiments have random initial conditions \n    (e.g. initialization of the network), and should be ran multiple times, not just once. \n    There's no error bars in any of the plots, so it's unclear whether AdaX really \n    does provide a statistically significant improvement over the baselines. \n    Similarly, the data in all the tables is quite similar, so without indicating the \n    spread of these estimates its impossible to tell whether these results are significant or not.\n\n  - How were the hyperparameters and step-size schedules chosen? The performance of Adam, AMSGrad, and \n    RMSProp are quite sensitive to their hyperparameters, and the optimal hyperparameters are problem-dependent. \n    Some of the experiments just use the default hyperparameters; this is insufficient when trying to directly \n    compare the performance of these methods, as their performance can vary greatly with different values of\n    these parameters. I'm not convinced that we should be drawing conclusions about the relative \n    performance of these algorithms from any of the experiments for this reason.\n\nOf course, meaningful empirical results are not necessarily characterized by statistically outperforming the baselines. \nWell designed experiments can highlight important ways in which the performances differ, providing the community \nwith a deeper understanding of the methods investigated. I would argue that the experiments in the paper do not \nachieve this either; the experiments do not provide any new intuition or understanding of the methods, showing \nonly the relative performances in terms of learning curves on a somewhat random collection of supervised learning problems. Why were these specific problems chosen? What makes these problems ideal for showcasing the performance of AdaX? If AdaX is an improvement over Adam, why? What exactly is happening with it's effective step-sizes that leads \nto the better performance? Can you show how their step-sizes differ over time? \n\nStatements that need citation or revision:\n  - \"Adaptive optimization algorithms such as RMSProp and Adam... as well as weak performance \n     compared to the first order gradient methods such as SGD\" (Abstract). This needs a citation. \n     Similarly, \"AdaX outperforms various tasks of computer vision and natural language processing and can catch \n     up with SGD\"; as above, I'm unaware of work (other than theoretical) that shows that SGD significantly \n     outperforms Adam in deep neural networks.\n  -  \"In the era of deep learning, SGD ... remains the most effective algorithm in training deep neural \n      networks\" (Introduction). What are you referring to here? Vanilla SGD? Or are you including Adam etc here? \n      As above, this should have a citation. Adam's popularity is largely due to its effectiveness in training \n      deep neural networks.\n  - \"However, Adam has worse performance (i.e. generalization ability in testing stage) compared with SGD\" \n     (Introduction). Citation needed.\n  - In the last paragraph of the Introduction, you introduced AdaX twice: \"To address the above issues, we propose a \n    new adaptive optimization method, termed AdaX, which guarantees convergence...\", and, \"To address \n    the above problems, we introduce a novel AdaX algorithm and theoreetically prove that it converges...\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper introduces a new step-size adaptation algorithm called AdaX. AdaX \nbuilds on the ideas of the Adam algorithm to address instability and non-convergence issues. \nConvergence of AdaX is proven in both convex and non-convex settings. The paper \nalso provides an empirical comparison of AdaX against its predecessors \n(SGD, RMSProp, Adam, AMSGrad) on a variety of tasks.\n\nI recommend the paper be rejected. I believe the convergence results could be a significant contribution, but the \nquality of the paper is hampered by its experimental design. The paper felt generally unpolished, containing\nfrequent grammatical errors, imprecise language, and uncited statements.\n\nMy main issue with the paper is the experimental design. I am not convinced that we can \ndraw valid conclusions from the experimental results for the following reasons:\n  - The experiments are lacking important details. How many independent runs of the experiment \n    were the experimental results averaged over? All of the experiments have random initial conditions \n    (e.g. initialization of the network), and should be ran multiple times, not just once. \n    There's no error bars in any of the plots, so it's unclear whether AdaX really \n    does provide a statistically significant improvement over the baselines. \n    Similarly, the data in all the tables is quite similar, so without indicating the \n    spread of these estimates its impossible to tell whether these results are significant or not.\n\n  - How were the hyperparameters and step-size schedules chosen? The performance of Adam, AMSGrad, and \n    RMSProp are quite sensitive to their hyperparameters, and the optimal hyperparameters are problem-dependent. \n    Some of the experiments just use the default hyperparameters; this is insufficient when trying to directly \n    compare the performance of these methods, as their performance can vary greatly with different values of\n    these parameters. I'm not convinced that we should be drawing conclusions about the relative \n    performance of these algorithms from any of the experiments for this reason.\n\nOf course, meaningful empirical results are not necessarily characterized by statistically outperforming the baselines. \nWell designed experiments can highlight important ways in which the performances differ, providing the community \nwith a deeper understanding of the methods investigated. I would argue that the experiments in the paper do not \nachieve this either; the experiments do not provide any new intuition or understanding of the methods, showing \nonly the relative performances in terms of learning curves on a somewhat random collection of supervised learning problems. Why were these specific problems chosen? What makes these problems ideal for showcasing the performance of AdaX? If AdaX is an improvement over Adam, why? What exactly is happening with it's effective step-sizes that leads \nto the better performance? Can you show how their step-sizes differ over time? \n\nStatements that need citation or revision:\n  - \"Adaptive optimization algorithms such as RMSProp and Adam... as well as weak performance \n     compared to the first order gradient methods such as SGD\" (Abstract). This needs a citation. \n     Similarly, \"AdaX outperforms various tasks of computer vision and natural language processing and can catch \n     up with SGD\"; as above, I'm unaware of work (other than theoretical) that shows that SGD significantly \n     outperforms Adam in deep neural networks.\n  -  \"In the era of deep learning, SGD ... remains the most effective algorithm in training deep neural \n      networks\" (Introduction). What are you referring to here? Vanilla SGD? Or are you including Adam etc here? \n      As above, this should have a citation. Adam's popularity is largely due to its effectiveness in training \n      deep neural networks.\n  - \"However, Adam has worse performance (i.e. generalization ability in testing stage) compared with SGD\" \n     (Introduction). Citation needed.\n  - In the last paragraph of the Introduction, you introduced AdaX twice: \"To address the above issues, we propose a \n    new adaptive optimization method, termed AdaX, which guarantees convergence...\", and, \"To address \n    the above problems, we introduce a novel AdaX algorithm and theoreetically prove that it converges...\"\n"}, "tcdate": 1570772004332}], "openreview_url": "https://openreview.net/forum?id=r1l-5pEtDr", "arxiv_id": "2004.09740", "paper_pdf": "papers/r1l-5pEtDr.pdf", "paper_pdf_sha256": "b17f9d2ab9934c085fc7a5295b48455bc398798bfc28245f300bbc215ef005f1", "paper_pdf_bytes": 747209, "paper_pdf_source": "openreview", "code_url": "https://github.com/switchablenorms/AdaX", "code_repository": "switchablenorms/AdaX", "code_commit": "e18f35f3d6ab99ad862f81d6ddf4d7dbc5f2f63d", "code_archive": "repos/r1l-5pEtDr.zip", "code_archive_sha256": "ebfc2d641b498b2184d82b523c9c929bb07c3f8491c6ebdd7009a9541abc8d16", "code_archive_bytes": 539620, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 529, "github_languages": {"Python": 81479}, "github_archived": false, "github_pushed_at": "2020-05-08T21:11:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adax-adaptive-gradient-descent-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "M7TJaVg4sE", "year": 2026, "status": "rejected", "title": "TraSCE: Trajectory Steering for Concept Erasure", "authors": ["Anubhav Jain", "Takashi Shibuya", "Yuya Kobayashi", "Yuhta Takida", "Nasir Memon", "Julian Togelius", "Yuki Mitsufuji"], "authorids": ["~Anubhav_Jain1", "~Takashi_Shibuya1", "~Yuya_Kobayashi1", "~Yuhta_Takida1", "~Nasir_Memon1", "~Julian_Togelius1", "~Yuki_Mitsufuji1"], "authors_source": "OpenReview API", "abstract": "Recent advancements in text-to-image diffusion models have brought them to the public spotlight, becoming widely accessible and embraced by everyday users. However, these models have been shown to generate harmful content, such as not-safe-for-work (NSFW) images. While approaches have been proposed to erase such abstract concepts from the models, jail-breaking techniques have succeeded in bypassing such safety measures. In this paper, we propose TraSCE, an approach to guide the diffusion trajectory away from generating harmful content. Our approach is based on negative prompting, but as we show, the widely used negative prompting strategy is not a complete solution for concept erasure and can easily be bypassed in a simple corner case. To address this issue, we introduce two techniques. We first borrow the idea of concept negation from compositional generation and propose to use it instead of the conventional negative prompting. Second, we introduce a localized loss-based guidance that enhances the modified negative prompting technique by steering the diffusion trajectory. We demonstrate that our proposed method achieves state-of-the-art results on various benchmarks in erasing harmful content, artistic styles, and objects, including on benchmarks introduced by red teams without impacting unrelated concepts. Our proposed approach does not require any training, weight modifications, or training data (either image or prompt), making it easier for model owners to erase new concepts.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "udPLQPUFOo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8975/Reviewer_Nzoa"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes TraSCE, a training-free method for concept erasure in text-to-image diffusion models. TraSCE combines a modified negative prompting strategy with a localized loss-based guidance to steer diffusion trajectories away from target concepts. TraSCE is evaluated on benchmarks that include adversarial prompts and tasks such as erasing NSFW content, artistic styles, and specific objects, achieving state-of-the-art reductions in attack success rates while preserving image quality across unrelated concepts.", "review_text": "The paper proposes TraSCE, a training-free method for concept erasure in text-to-image diffusion models. TraSCE combines a modified negative prompting strategy with a localized loss-based guidance to steer diffusion trajectories away from target concepts. TraSCE is evaluated on benchmarks that include adversarial prompts and tasks such as erasing NSFW content, artistic styles, and specific objects, achieving state-of-the-art reductions in attack success rates while preserving image quality across unrelated concepts.", "strengths": "**S1:** TraSCE operates at inference time, eliminating the need for costly retraining or data collection. This makes it easily deployable for model owners to adapt to new concepts.\n\n**S2:** TraSCE shows significant reductions in attack success rates against black-box adversarial attacks, with minimal degradation in general image quality.\n\n**S3:** Experiments cover diverse erasure tasks using multiple metrics, providing a broad assessment of the method's applicability.", "weaknesses": "**W1:** The main concern about this paper is its limited novelty and insufficient distinction from prior work. The core component of TraSCE, the modified negative prompting, is adapted from Liu et al. (2022) on concept negation but lacks adequate justification for its novelty. While the addition of localized loss-based guidance is claimed as new, it fails to be differentiated from existing guidance techniques, such as classifier guidance. For instance, Schramowski et al. (2023) also employ trajectory steering for safety, but TraSCE's ablation studies only compare against basic negative prompting, lacking a deep analysis of how the loss function advances beyond prior art. Therefore, the authors should conduct a comparative analysis with recent inference-time methods, highlighting theoretical or empirical distinctions, and perform ablations to quantify the independent contribution of the loss guidance versus the negative prompting.\n\n**W2:** The experimental results lack dynamic adversarial testing, white-box attack settings, and human evaluation. The paper does not test against adaptive white-box attacks that exploit TraSCE's gradient information, thereby limiting its claims of robustness. For artistic style and object erasure, assessments depend solely on automated classifiers without human evaluation or diversity metrics, risking overestimation of erasure effectiveness due to dataset biases. The authors should incorporate dynamic adversarial testing, add human evaluations for subjective tasks (e.g., artistic styles), and report diversity scores to ensure that erasure does not harm output variety.\n\n**W3:** TraSCE increases inference time by approximately 2.6 times, which could hinder real-time applications. Hyperparameters are task-specific and tuned empirically, yet no sensitivity analysis or guidance for adaptive selection is provided. To enhance practicality, the authors could benchmark TraSCE on resource-constrained devices and suggest approximations (e.g., reducing gradient steps or using sparse updates).", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes TraSCE, a training-free method for concept erasure in text-to-image diffusion models. TraSCE combines a modified negative prompting strategy with a localized loss-based guidance to steer diffusion trajectories away from target concepts. TraSCE is evaluated on benchmarks that include adversarial prompts and tasks such as erasing NSFW content, artistic styles, and specific objects, achieving state-of-the-art reductions in attack success rates while preserving image quality across unrelated concepts.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "**S1:** TraSCE operates at inference time, eliminating the need for costly retraining or data collection. This makes it easily deployable for model owners to adapt to new concepts.\n\n**S2:** TraSCE shows significant reductions in attack success rates against black-box adversarial attacks, with minimal degradation in general image quality.\n\n**S3:** Experiments cover diverse erasure tasks using multiple metrics, providing a broad assessment of the method's applicability.", "weaknesses": "**W1:** The main concern about this paper is its limited novelty and insufficient distinction from prior work. The core component of TraSCE, the modified negative prompting, is adapted from Liu et al. (2022) on concept negation but lacks adequate justification for its novelty. While the addition of localized loss-based guidance is claimed as new, it fails to be differentiated from existing guidance techniques, such as classifier guidance. For instance, Schramowski et al. (2023) also employ trajectory steering for safety, but TraSCE's ablation studies only compare against basic negative prompting, lacking a deep analysis of how the loss function advances beyond prior art. Therefore, the authors should conduct a comparative analysis with recent inference-time methods, highlighting theoretical or empirical distinctions, and perform ablations to quantify the independent contribution of the loss guidance versus the negative prompting.\n\n**W2:** The experimental results lack dynamic adversarial testing, white-box attack settings, and human evaluation. The paper does not test against adaptive white-box attacks that exploit TraSCE's gradient information, thereby limiting its claims of robustness. For artistic style and object erasure, assessments depend solely on automated classifiers without human evaluation or diversity metrics, risking overestimation of erasure effectiveness due to dataset biases. The authors should incorporate dynamic adversarial testing, add human evaluations for subjective tasks (e.g., artistic styles), and report diversity scores to ensure that erasure does not harm output variety.\n\n**W3:** TraSCE increases inference time by approximately 2.6 times, which could hinder real-time applications. Hyperparameters are task-specific and tuned empirically, yet no sensitivity analysis or guidance for adaptive selection is provided. To enhance practicality, the authors could benchmark TraSCE on resource-constrained devices and suggest approximations (e.g., reducing gradient steps or using sparse updates).", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761985878284}, {"id": "7b55WjeJ5E", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8975/Reviewer_WyCx"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This study introduces TraSCE, a method aimed at mitigating the generation of harmful content (e.g., sexual elements) in T2I diffusion models. The authors propose guiding the diffusion trajectory away from problematic generation paths. The proposed method combines two part: a modification of traditional negative prompting based on the classifier-free guidance, and a localized loss-based guidance, which further enhance the guidance performance. The performance of the proposed method is outstanding, demonstrating effectiveness on multiple models and concept erasure tasks.", "review_text": "This study introduces TraSCE, a method aimed at mitigating the generation of harmful content (e.g., sexual elements) in T2I diffusion models. The authors propose guiding the diffusion trajectory away from problematic generation paths. The proposed method combines two part: a modification of traditional negative prompting based on the classifier-free guidance, and a localized loss-based guidance, which further enhance the guidance performance. The performance of the proposed method is outstanding, demonstrating effectiveness on multiple models and concept erasure tasks.", "strengths": "1. The method is very clear and easy to understand.\n2. The proposed method performs excellently and shows outstanding results on multiple evaluation benchmarks. \n3. The authors' experimental setup is comprehensive, taking into account various evaluation tasks, erasure robustness, and different base models.", "weaknesses": "1. The application of the proposed method seems to be based on an unreasonable setting: that the specific category of harmful content must be predefined for the current generation. This is impractical in real-world scenarios. In contrast, recent related works [1, 2, 3] adopt a \"detect-then-erase\" mechanism, which first determines if a specific concept has been generated and only then performs concept erasure. This appears to be a more reasonable setup.\n2. The additional generation time introduced by the proposed method appears to be unacceptable. For instance, in a standard classifier-free guidance model, the application of Eq. 3 directly incurs an extra 50% inference time, and the feedback optimization in Eq. 4 is even more time-consuming.\n3. I appreciate that the authors adapted their experiments to the new FLUX model. However, related to the previous point, the authors must report the additional inference time overhead that the proposed method introduces when applied to FLUX. Given that FLUX is a  model with a very large number of parameters (15x than sd1.4), it is difficult to believe that the optimization mechanism from Eq. 4 can be practicably applied to it.\n\n[1] SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image and Video Generation, iclr25\n\n[2] Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization, cvpr25\n\n[3]  Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation, cvpr25", "questions": "Please address the weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces TraSCE, a method aimed at mitigating the generation of harmful content (e.g., sexual elements) in T2I diffusion models. The authors propose guiding the diffusion trajectory away from problematic generation paths. The proposed method combines two part: a modification of traditional negative prompting based on the classifier-free guidance, and a localized loss-based guidance, which further enhance the guidance performance. The performance of the proposed method is outstanding, demonstrating effectiveness on multiple models and concept erasure tasks.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The method is very clear and easy to understand.\n2. The proposed method performs excellently and shows outstanding results on multiple evaluation benchmarks. \n3. The authors' experimental setup is comprehensive, taking into account various evaluation tasks, erasure robustness, and different base models.", "weaknesses": "1. The application of the proposed method seems to be based on an unreasonable setting: that the specific category of harmful content must be predefined for the current generation. This is impractical in real-world scenarios. In contrast, recent related works [1, 2, 3] adopt a \"detect-then-erase\" mechanism, which first determines if a specific concept has been generated and only then performs concept erasure. This appears to be a more reasonable setup.\n2. The additional generation time introduced by the proposed method appears to be unacceptable. For instance, in a standard classifier-free guidance model, the application of Eq. 3 directly incurs an extra 50% inference time, and the feedback optimization in Eq. 4 is even more time-consuming.\n3. I appreciate that the authors adapted their experiments to the new FLUX model. However, related to the previous point, the authors must report the additional inference time overhead that the proposed method introduces when applied to FLUX. Given that FLUX is a  model with a very large number of parameters (15x than sd1.4), it is difficult to believe that the optimization mechanism from Eq. 4 can be practicably applied to it.\n\n[1] SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image and Video Generation, iclr25\n\n[2] Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization, cvpr25\n\n[3]  Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation, cvpr25", "questions": "Please address the weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761914129300}, {"id": "3f1iLhc6vn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8975/Reviewer_xZU9"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes TraSCE, an inference-time method for concept erasure in text-to-image diffusion. It combines (i) a modified negative prompting formulation (replacing the conventional “negative as base” with an unconditional base plus a difference term) and (ii) a localized loss–based guidance that nudges the denoising state away from regions aligned with the undesired concept each step. Experiments report lower attack-success rates on several adversarial NSFW/violence benchmarks and show applications to artistic styles and object erasure.", "review_text": "The paper proposes TraSCE, an inference-time method for concept erasure in text-to-image diffusion. It combines (i) a modified negative prompting formulation (replacing the conventional “negative as base” with an unconditional base plus a difference term) and (ii) a localized loss–based guidance that nudges the denoising state away from regions aligned with the undesired concept each step. Experiments report lower attack-success rates on several adversarial NSFW/violence benchmarks and show applications to artistic styles and object erasure.", "strengths": "1.  Addresses safety in diffusion models, especially robustness to prompt‑based jailbreaks.\n2. Integrates a geometric “trajectory steering” view into the diffusion process, offering intuitive control over latent evolution.", "weaknesses": "1. While “trajectory steering” offers a coherent new perspective, the implementation closely resembles classifier-free or loss-based guidance mechanisms already explored in prior works (e.g., SLD). The novelty primarily lies in problem framing and loss design rather than theoretical advancement.\n2. The per-step gradient update increases sampling time by 2–3×, which may limit deployment for large-scale or real-time use.", "questions": "1. How robust is the method to partial or semantically related prompts (e.g., “bikini” vs. “swimsuit”)? Does the trajectory steering generalize to paraphrased adversarial prompts?\n2. Can the authors quantitatively characterize how the diffusion trajectories differ between baseline CFG, SLD, and TraSCE?\n5. How does TraSCE behave when multiple negative concepts are specified simultaneously (e.g., “no nudity, no gore”)? Does the trajectory steering remain stable?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes TraSCE, an inference-time method for concept erasure in text-to-image diffusion. It combines (i) a modified negative prompting formulation (replacing the conventional “negative as base” with an unconditional base plus a difference term) and (ii) a localized loss–based guidance that nudges the denoising state away from regions aligned with the undesired concept each step. Experiments report lower attack-success rates on several adversarial NSFW/violence benchmarks and show applications to artistic styles and object erasure.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.  Addresses safety in diffusion models, especially robustness to prompt‑based jailbreaks.\n2. Integrates a geometric “trajectory steering” view into the diffusion process, offering intuitive control over latent evolution.", "weaknesses": "1. While “trajectory steering” offers a coherent new perspective, the implementation closely resembles classifier-free or loss-based guidance mechanisms already explored in prior works (e.g., SLD). The novelty primarily lies in problem framing and loss design rather than theoretical advancement.\n2. The per-step gradient update increases sampling time by 2–3×, which may limit deployment for large-scale or real-time use.", "questions": "1. How robust is the method to partial or semantically related prompts (e.g., “bikini” vs. “swimsuit”)? Does the trajectory steering generalize to paraphrased adversarial prompts?\n2. Can the authors quantitatively characterize how the diffusion trajectories differ between baseline CFG, SLD, and TraSCE?\n5. How does TraSCE behave when multiple negative concepts are specified simultaneously (e.g., “no nudity, no gore”)? Does the trajectory steering remain stable?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761710030192}, {"id": "pEm4z7yvcT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8975/Reviewer_tTZr"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The authors propose a method to erase harmful content from diffusion models. They identify a corner case where vanilla negative prompting fails and replace it with a safer variant that better suppresses the targeted concept. They also introduce a localized, loss-based guidance term that nudges sampling toward the unconditional trajectory when the prompt behaves like the banned concept. The paper evaluates the approach on adversarial attack benchmarks and reports results for erasing artistic styles and objects.", "review_text": "The authors propose a method to erase harmful content from diffusion models. They identify a corner case where vanilla negative prompting fails and replace it with a safer variant that better suppresses the targeted concept. They also introduce a localized, loss-based guidance term that nudges sampling toward the unconditional trajectory when the prompt behaves like the banned concept. The paper evaluates the approach on adversarial attack benchmarks and reports results for erasing artistic styles and objects.", "strengths": "1. The paper tackles a timely and important safety challenge: removing harmful content from generative models. \n2. The paper is clearly written, easy to follow, and the proposed method is both simple and elegant.", "weaknesses": "1. The method assumes that falling back to the unconditional prior is a safe default, which may not always hold. If the base model’s unconditional distribution already leans toward risky or unwanted content in certain contexts, anchoring to it won’t prevent leakage. The approach can fail in such cases.\n2. I also have sme concerns regarding the generation quality.  In diffusion, coarse scene layout forms early, while objects solidify mid-trajectory and fine details emerge late. So, when it comes to erasing the objects, the loss in the paper only activates when $$ \\hat{\\epsilon}_\\theta(x_t, e_p) - \\hat{\\epsilon}_\\theta(x_t, e_{np}) $$\n becomes small, typically mid/late. By that time, the scene and object placement are largely decided. Nudging the latent then can remove or weaken the target object only by distorting the image. This may be true even for complex safety concepts. For instance, the qualitative result shown in Figure 4, where the authors aim to “erase a car,” looks distorted and challenges the claim that the method preserves generation quality.", "questions": "1. Can the authors provide a plot that shows how the difference term varies across timesteps during the erasure of objects? This may give some clarity on the concern in 2 above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a method to erase harmful content from diffusion models. They identify a corner case where vanilla negative prompting fails and replace it with a safer variant that better suppresses the targeted concept. They also introduce a localized, loss-based guidance term that nudges sampling toward the unconditional trajectory when the prompt behaves like the banned concept. The paper evaluates the approach on adversarial attack benchmarks and reports results for erasing artistic styles and objects.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The paper tackles a timely and important safety challenge: removing harmful content from generative models. \n2. The paper is clearly written, easy to follow, and the proposed method is both simple and elegant.", "weaknesses": "1. The method assumes that falling back to the unconditional prior is a safe default, which may not always hold. If the base model’s unconditional distribution already leans toward risky or unwanted content in certain contexts, anchoring to it won’t prevent leakage. The approach can fail in such cases.\n2. I also have sme concerns regarding the generation quality.  In diffusion, coarse scene layout forms early, while objects solidify mid-trajectory and fine details emerge late. So, when it comes to erasing the objects, the loss in the paper only activates when $$ \\hat{\\epsilon}_\\theta(x_t, e_p) - \\hat{\\epsilon}_\\theta(x_t, e_{np}) $$\n becomes small, typically mid/late. By that time, the scene and object placement are largely decided. Nudging the latent then can remove or weaken the target object only by distorting the image. This may be true even for complex safety concepts. For instance, the qualitative result shown in Figure 4, where the authors aim to “erase a car,” looks distorted and challenges the claim that the method preserves generation quality.", "questions": "1. Can the authors provide a plot that shows how the difference term varies across timesteps during the erasure of objects? This may give some clarity on the concern in 2 above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760955228274}], "openreview_url": "https://openreview.net/forum?id=M7TJaVg4sE", "arxiv_id": "2412.07658", "paper_pdf": "papers/M7TJaVg4sE.pdf", "paper_pdf_sha256": "da5252c05c4cbfdd9ea12e28a4b5e94b6a6eda160cd31b48314a1cf14c19576a", "paper_pdf_bytes": 15277066, "paper_pdf_source": "openreview", "code_url": "https://github.com/anubhav1997/TraSCE", "code_repository": "anubhav1997/TraSCE", "code_commit": "8e3813b2aea10311cbed4bb85e25be345dd3db35", "code_archive": "repos/M7TJaVg4sE.zip", "code_archive_sha256": "db108808a6143fb164e7e45de8835d6572d8309d627575a14be40f717dee534c", "code_archive_bytes": 1713193, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 1686, "github_languages": {"Python": 31762}, "github_archived": false, "github_pushed_at": "2025-02-01T16:48:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/trasce-trajectory-steering-for-concept"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dTQmayPKMs", "year": 2025, "status": "rejected", "title": "Understanding Impact of Human Feedback via Influence Functions", "authors": ["Taywon Min", "Haeone Lee", "Hanho Ryu", "Yongchan Kwon", "Kimin Lee"], "authorids": ["~Taywon_Min1", "~Haeone_Lee1", "~Hanho_Ryu1", "~Yongchan_Kwon1", "~Kimin_Lee1"], "authors_source": "OpenReview API", "abstract": "In Reinforcement Learning from Human Feedback (RLHF), it is crucial to learn suitable reward models from human feedback to align large language models (LLMs) with human intentions. However, human feedback can often be noisy, inconsistent, or biased, especially when evaluating complex responses. Such feedback can lead to misaligned reward signals, potentially causing unintended side effects during the RLHF process. To address these challenges, we explore the use of influence functions to measure the impact of human feedback on the performance of reward models. We propose a compute-efficient approximation method that enables the application of influence functions to LLM-based reward models and large-scale preference datasets. In our experiments, we demonstrate two key applications of influence functions: (1) detecting common forms of labeler bias in human feedback datasets and (2) guiding labelers to refine their strategies to align more closely with expert feedback. By quantifying the impact of human feedback on reward models, we believe that influence functions can enhance feedback interpretability and contribute to scalable oversight in RLHF, helping labelers provide more accurate and consistent feedback.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "9a40MtTkXY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_roz9"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper presents a novel approach to analyzing the effects of human feedback in RLHF. It introduces influence functions to quantify the impact of individual feedback on the performance of reward models, which can help detect biases and improve labeling strategies in RLHF systems. Two primary applications of influence functions are highlighted: detecting biases like length and sycophancy in human feedback and guiding labelers to align their feedback with expert standards.", "review_text": "The paper presents a novel approach to analyzing the effects of human feedback in RLHF. It introduces influence functions to quantify the impact of individual feedback on the performance of reward models, which can help detect biases and improve labeling strategies in RLHF systems. Two primary applications of influence functions are highlighted: detecting biases like length and sycophancy in human feedback and guiding labelers to align their feedback with expert standards.", "strengths": "1. The use of influence functions to analyze the impact of human feedback is a promising direction that adds a layer of interpretability in RLHF, which is essential for aligning LLMs with human values.\n2. The paper introduces a compute-efficient method that enables scalable application of influence functions, potentially reducing computational demands by 2.5 times, a significant improvement over previous methods.\n3. The methodology, experimental design, and results are presented clearly, with well-structured figures and examples to support the claims.", "weaknesses": "1. The study’s limitations in real-world scenarios, where expert and non-expert labelers may not share sub-objective scores, could reduce the generalizability of the approach.\n2.  While the paper shows effectiveness in detecting length bias, sycophancy bias remains challenging, as it involves understanding nuanced human agreement tendencies that may vary by context.", "questions": "1. Could the reliance on targeted validation sets be reduced by adapting influence functions to work with more generalized validation samples?\n2. How does the approach scale with increasingly large datasets or models beyond the experiments presented, and would there be additional trade-offs in compute efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a novel approach to analyzing the effects of human feedback in RLHF. It introduces influence functions to quantify the impact of individual feedback on the performance of reward models, which can help detect biases and improve labeling strategies in RLHF systems. Two primary applications of influence functions are highlighted: detecting biases like length and sycophancy in human feedback and guiding labelers to align their feedback with expert standards.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The use of influence functions to analyze the impact of human feedback is a promising direction that adds a layer of interpretability in RLHF, which is essential for aligning LLMs with human values.\n2. The paper introduces a compute-efficient method that enables scalable application of influence functions, potentially reducing computational demands by 2.5 times, a significant improvement over previous methods.\n3. The methodology, experimental design, and results are presented clearly, with well-structured figures and examples to support the claims.", "weaknesses": "1. The study’s limitations in real-world scenarios, where expert and non-expert labelers may not share sub-objective scores, could reduce the generalizability of the approach.\n2.  While the paper shows effectiveness in detecting length bias, sycophancy bias remains challenging, as it involves understanding nuanced human agreement tendencies that may vary by context.", "questions": "1. Could the reliance on targeted validation sets be reduced by adapting influence functions to work with more generalized validation samples?\n2. How does the approach scale with increasingly large datasets or models beyond the experiments presented, and would there be additional trade-offs in compute efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730721570648}, {"id": "jNVkDEOdwA", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_CzsY"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper addresses the challenge of noise and misalignment in human feedback within RLHF, which can lead to flawed reward signals. By using influence functions, it quantifies the impact of individual data points on reward model performance. For each data point, the influence function captures how parameter adjustments would occur when the weight of that point changes. The approach assumes a small validation set is available to calculate overall influence and utilizes an approximate inverse Hessian to enable efficient computation. Results indicate improved AUC in addressing length bias and sycophancy bias compared to GPT-4o and Gemini models.\n\nAdditionally, the paper explores the potential of guiding non-expert labelers to align more closely with expert labeling. It assumes that labelers use a list of linearly combined, weighted sub-scores in their scoring process. By identifying data points that negatively impact alignment, non-expert labelers can adjust weights and improve their performance. Results suggest that with a few samples, non-expert labelers demonstrate notable improvement.", "review_text": "The paper addresses the challenge of noise and misalignment in human feedback within RLHF, which can lead to flawed reward signals. By using influence functions, it quantifies the impact of individual data points on reward model performance. For each data point, the influence function captures how parameter adjustments would occur when the weight of that point changes. The approach assumes a small validation set is available to calculate overall influence and utilizes an approximate inverse Hessian to enable efficient computation. Results indicate improved AUC in addressing length bias and sycophancy bias compared to GPT-4o and Gemini models.\n\nAdditionally, the paper explores the potential of guiding non-expert labelers to align more closely with expert labeling. It assumes that labelers use a list of linearly combined, weighted sub-scores in their scoring process. By identifying data points that negatively impact alignment, non-expert labelers can adjust weights and improve their performance. Results suggest that with a few samples, non-expert labelers demonstrate notable improvement.", "strengths": "This paper excels in identifying data points that negatively influence reward models through the application of influence functions. Given the common issue of noise in human-labeled data, particularly in RLHF, this approach enhances transparency and is valuable in managing labels from non-expert labelers. The methodology effectively addresses computational challenges using an approximation function, which is highly practical aiming for reliable AI.\n\nThe experiments are well-defined, and the performance results are convincing. The paper is well-articulated regarding both analytical rigor and technical details. Additionally, the authors explore ways to improve labeler performance, which is a good extension of the study.", "weaknesses": "While the application of influence functions to assess the contribution of individual data points is compelling, the methodology relies heavily on manually curated validation sets, as evidenced by the ablation experiments. There are two primary concerns:\n\n1) Dependence on Domain Knowledge: The construction of validation sets requires domain-specific knowledge, which could limit generalizability. Although addressing verbosity and sycophancy is valuable, the methodology appears capable of handling other issues if suitable validation sets are available. However, this adaptability is currently contingent on creating each validation set by hand.\n\n2) Risk of Conflicting Improvements: It is unclear how the approach manages potential conflicts between multiple improvement directions. For instance, in the main results (Figure 2), Influence shows higher TPR than GPT-4o at the same FPR, which is promising. However, it raises the question: Does this improvement come at the expense of other metrics? To fully assess the model’s overall robustness, it would be beneficial to provide additional statistics comparing the base model and the Influence-based approach to ensure that the latter does not merely overfit to specific biases, such as length.\n\nFurthermore, as noted in the limitations, the feasibility of improving labelers’ performance in real-world settings may be limited, especially when human labelers’ judgments involve complex decision-making beyond linear reward models. The authors could strengthen their argument by including results from human subjects whose labeling strategies demonstrably improve with guidance. Alternatively, if the current scope remains focused on bias correction within datasets, it would be beneficial to clarify this in the introduction to avoid misleading interpretations about broader applicability.", "questions": "1) Will focusing on improving a single aspect of bias impact other performance metrics?\n\n2) Beyond runtime speedup, does approximating the inverse Hessian lead to any performance degradation?\n\n3) Could you provide insights on why Entropy underperforms significantly in addressing Sycophancy Bias relative to other baselines?\n\n4) In Appendix B2, five weights are selected for B. Could you elaborate on the rationale for these specific weights? Were they chosen randomly, and are any particularly extreme?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of noise and misalignment in human feedback within RLHF, which can lead to flawed reward signals. By using influence functions, it quantifies the impact of individual data points on reward model performance. For each data point, the influence function captures how parameter adjustments would occur when the weight of that point changes. The approach assumes a small validation set is available to calculate overall influence and utilizes an approximate inverse Hessian to enable efficient computation. Results indicate improved AUC in addressing length bias and sycophancy bias compared to GPT-4o and Gemini models.\n\nAdditionally, the paper explores the potential of guiding non-expert labelers to align more closely with expert labeling. It assumes that labelers use a list of linearly combined, weighted sub-scores in their scoring process. By identifying data points that negatively impact alignment, non-expert labelers can adjust weights and improve their performance. Results suggest that with a few samples, non-expert labelers demonstrate notable improvement.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "This paper excels in identifying data points that negatively influence reward models through the application of influence functions. Given the common issue of noise in human-labeled data, particularly in RLHF, this approach enhances transparency and is valuable in managing labels from non-expert labelers. The methodology effectively addresses computational challenges using an approximation function, which is highly practical aiming for reliable AI.\n\nThe experiments are well-defined, and the performance results are convincing. The paper is well-articulated regarding both analytical rigor and technical details. Additionally, the authors explore ways to improve labeler performance, which is a good extension of the study.", "weaknesses": "While the application of influence functions to assess the contribution of individual data points is compelling, the methodology relies heavily on manually curated validation sets, as evidenced by the ablation experiments. There are two primary concerns:\n\n1) Dependence on Domain Knowledge: The construction of validation sets requires domain-specific knowledge, which could limit generalizability. Although addressing verbosity and sycophancy is valuable, the methodology appears capable of handling other issues if suitable validation sets are available. However, this adaptability is currently contingent on creating each validation set by hand.\n\n2) Risk of Conflicting Improvements: It is unclear how the approach manages potential conflicts between multiple improvement directions. For instance, in the main results (Figure 2), Influence shows higher TPR than GPT-4o at the same FPR, which is promising. However, it raises the question: Does this improvement come at the expense of other metrics? To fully assess the model’s overall robustness, it would be beneficial to provide additional statistics comparing the base model and the Influence-based approach to ensure that the latter does not merely overfit to specific biases, such as length.\n\nFurthermore, as noted in the limitations, the feasibility of improving labelers’ performance in real-world settings may be limited, especially when human labelers’ judgments involve complex decision-making beyond linear reward models. The authors could strengthen their argument by including results from human subjects whose labeling strategies demonstrably improve with guidance. Alternatively, if the current scope remains focused on bias correction within datasets, it would be beneficial to clarify this in the introduction to avoid misleading interpretations about broader applicability.", "questions": "1) Will focusing on improving a single aspect of bias impact other performance metrics?\n\n2) Beyond runtime speedup, does approximating the inverse Hessian lead to any performance degradation?\n\n3) Could you provide insights on why Entropy underperforms significantly in addressing Sycophancy Bias relative to other baselines?\n\n4) In Appendix B2, five weights are selected for B. Could you elaborate on the rationale for these specific weights? Were they chosen randomly, and are any particularly extreme?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No concerns.", "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730691042635}, {"id": "fNOXIVnuMy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_71Vm"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "This paper presents a novel approach to evaluate and refine the reward models in RLHF by employing influence functions. The authors argue that human feedback, which is integral to training reward models for LLMs, can be noisy and biased, leading to misaligned rewards. To mitigate this, they introduce a compute-efficient method to approximate influence functions, allowing for the measurement of individual data points' impact on reward model performance. The paper claims two main applications: detecting labeler bias in feedback datasets and guiding labelers to align more closely with expert feedback. The authors demonstrate the effectiveness of their approach through experiments and argue that it enhances feedback interpretability and contributes to scalable oversight in RLHF.", "review_text": "This paper presents a novel approach to evaluate and refine the reward models in RLHF by employing influence functions. The authors argue that human feedback, which is integral to training reward models for LLMs, can be noisy and biased, leading to misaligned rewards. To mitigate this, they introduce a compute-efficient method to approximate influence functions, allowing for the measurement of individual data points' impact on reward model performance. The paper claims two main applications: detecting labeler bias in feedback datasets and guiding labelers to align more closely with expert feedback. The authors demonstrate the effectiveness of their approach through experiments and argue that it enhances feedback interpretability and contributes to scalable oversight in RLHF.", "strengths": "The paper introduces a novel approach to enhance the interpretability of reward models. By applying influence functions, the authors provide a method to quantify the impact of individual feedback on the model's performance, offering insights into how human feedback shapes the reward model's outcomes.\n\nThe idea of using influence functions to measure the impact of human feedback is innovative and has the potential to contribute to the broader goal of scalable oversight in RLHF. This approach can help in detecting and mitigating labeler bias, which is a common challenge in training robust and aligned AI systems.", "weaknesses": "As far as I am concerned, the authors simpy apply the approach in [1] to the reward modeling scenario, which greatly limits the novelty of the paper. I suggest that the author summarize the main contributions.\n\nWhile the experiments show promise, establishing reward models with various LLMs and evaluating with more downstream alignment tasks, such as direct alignment algorithms, could further validate the generalizability of the approach.\n\nReference:\n[1] Koh P W, Liang P. Understanding black-box predictions via influence functions[C]//International conference on machine learning. PMLR, 2017: 1885-1894.", "questions": "Could you provide further insights into how the method performs when handling intentionally misleading or adversarial feedback? Are there specific patterns or checks in place to identify and mitigate such cases?\n\nPlease provide more specifics on how the method scales with increasingly large datasets. What computational trade-offs may arise as the dataset grows?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel approach to evaluate and refine the reward models in RLHF by employing influence functions. The authors argue that human feedback, which is integral to training reward models for LLMs, can be noisy and biased, leading to misaligned rewards. To mitigate this, they introduce a compute-efficient method to approximate influence functions, allowing for the measurement of individual data points' impact on reward model performance. The paper claims two main applications: detecting labeler bias in feedback datasets and guiding labelers to align more closely with expert feedback. The authors demonstrate the effectiveness of their approach through experiments and argue that it enhances feedback interpretability and contributes to scalable oversight in RLHF.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "The paper introduces a novel approach to enhance the interpretability of reward models. By applying influence functions, the authors provide a method to quantify the impact of individual feedback on the model's performance, offering insights into how human feedback shapes the reward model's outcomes.\n\nThe idea of using influence functions to measure the impact of human feedback is innovative and has the potential to contribute to the broader goal of scalable oversight in RLHF. This approach can help in detecting and mitigating labeler bias, which is a common challenge in training robust and aligned AI systems.", "weaknesses": "As far as I am concerned, the authors simpy apply the approach in [1] to the reward modeling scenario, which greatly limits the novelty of the paper. I suggest that the author summarize the main contributions.\n\nWhile the experiments show promise, establishing reward models with various LLMs and evaluating with more downstream alignment tasks, such as direct alignment algorithms, could further validate the generalizability of the approach.\n\nReference:\n[1] Koh P W, Liang P. Understanding black-box predictions via influence functions[C]//International conference on machine learning. PMLR, 2017: 1885-1894.", "questions": "Could you provide further insights into how the method performs when handling intentionally misleading or adversarial feedback? Are there specific patterns or checks in place to identify and mitigate such cases?\n\nPlease provide more specifics on how the method scales with increasingly large datasets. What computational trade-offs may arise as the dataset grows?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730637859565}, {"id": "S9uV9pY0SN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_Wn7U"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper discusses using influence funcitons to identify potential bias of human labelers in human feedback data for LLM alignment. To enable effcient estimation of incluence functions under the LLM context, DataInf and OPORP are introduced to ease the burden of computation and storage respectively. Empirical results demonstrate that the proposed method can effectively identify lengthy and sycophancy bias compared to baselines and querying powerful LLMs. Specially, the authors show that influence functions can help align non-expert labelers' strategies to experts' labeling strategy, which I believe is important for preference data annotation in LLM alignment.", "review_text": "This paper discusses using influence funcitons to identify potential bias of human labelers in human feedback data for LLM alignment. To enable effcient estimation of incluence functions under the LLM context, DataInf and OPORP are introduced to ease the burden of computation and storage respectively. Empirical results demonstrate that the proposed method can effectively identify lengthy and sycophancy bias compared to baselines and querying powerful LLMs. Specially, the authors show that influence functions can help align non-expert labelers' strategies to experts' labeling strategy, which I believe is important for preference data annotation in LLM alignment.", "strengths": "1. The discussed topic is practical and important.\n\n2. Through approximation and compression, the proposed method can estimate influence functions efficiently.\n\n3. Empirical results are promising (although no comparison to concurrent works)", "weaknesses": "1. The proposed method mainly integrates existing techniques to the LLM context (although I do not think this is a serious issue as long as they work properly)\n\n2. Mistakes in fig.4 legend (AUC=0.8 for both Concise and Verbose?)\n\n3. Empirical study only show the efficacy in identifying bias, but discussion on data quality improvement with such detechtion is missing.", "questions": "1. \"we reduce the size of a single gradient vector from 160MB (42M dimensions) to 256KB (65K dimensions)\"\n\nHow do you calculate the size of gradient vectors? Why is the gradient vector size 160MB (Llama-3-8B in your experiment has around 15GB model size)\n\n2. Could you provide some performance improvement results on real-world dataset after filtering biased data with incluence functions?\n\n3. I notice you only use four dimensions of Helpsteer2 in section 5.2.1 (the four-dimensional weight vector]). Why do you make such alteration?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper discusses using influence funcitons to identify potential bias of human labelers in human feedback data for LLM alignment. To enable effcient estimation of incluence functions under the LLM context, DataInf and OPORP are introduced to ease the burden of computation and storage respectively. Empirical results demonstrate that the proposed method can effectively identify lengthy and sycophancy bias compared to baselines and querying powerful LLMs. Specially, the authors show that influence functions can help align non-expert labelers' strategies to experts' labeling strategy, which I believe is important for preference data annotation in LLM alignment.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The discussed topic is practical and important.\n\n2. Through approximation and compression, the proposed method can estimate influence functions efficiently.\n\n3. Empirical results are promising (although no comparison to concurrent works)", "weaknesses": "1. The proposed method mainly integrates existing techniques to the LLM context (although I do not think this is a serious issue as long as they work properly)\n\n2. Mistakes in fig.4 legend (AUC=0.8 for both Concise and Verbose?)\n\n3. Empirical study only show the efficacy in identifying bias, but discussion on data quality improvement with such detechtion is missing.", "questions": "1. \"we reduce the size of a single gradient vector from 160MB (42M dimensions) to 256KB (65K dimensions)\"\n\nHow do you calculate the size of gradient vectors? Why is the gradient vector size 160MB (Llama-3-8B in your experiment has around 15GB model size)\n\n2. Could you provide some performance improvement results on real-world dataset after filtering biased data with incluence functions?\n\n3. I notice you only use four dimensions of Helpsteer2 in section 5.2.1 (the four-dimensional weight vector]). Why do you make such alteration?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730627922286}, {"id": "10wlztcn4q", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_dPes"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 2, "summary": "The paper describes the use of influence functions to evaluate the influence of preference samples in reward modeling. The authors apply recent methods for computing influence functions for large neural networks/LLM to reward model training. By using a validation set demonstrating preferences for specific qualities, the influence of examples is highlighted. The authors also propose a compression scheme to enable efficient storage of gradient vectors for dot products. The authors apply the method for the detection of biased data samples, and outline guidance of human labelers.", "review_text": "The paper describes the use of influence functions to evaluate the influence of preference samples in reward modeling. The authors apply recent methods for computing influence functions for large neural networks/LLM to reward model training. By using a validation set demonstrating preferences for specific qualities, the influence of examples is highlighted. The authors also propose a compression scheme to enable efficient storage of gradient vectors for dot products. The authors apply the method for the detection of biased data samples, and outline guidance of human labelers.", "strengths": "Positive:\n- The paper is well written and polished, I found it easy to follow\n- The main presented methods seems to work well, and is evaluated appropriately\n- The different use cases (bias detection and user guidance) are presented well and relevant\n- Limitations are discussed (such as the dependence on validation set composition), and runs (although limited) ablations on the effects of different compositions\n- The quality of experiments seems high, both qualitative and quantitative results are generally convincing\n- Description of experiments/data sets is sufficient for reproducibility\n- The paper presents a nice application of influence functions\n\nOverall, I enjoyed reading the paper, and am convinced by its research quality. While strict technical novelty or theoretical insights are limited, I think it still provides a meaningful contribution to the field.", "weaknesses": "Negative:\n- The methdological novelty is limited, i.e. just the application of existing methods to a new use case\n- The dependency of the method on validation set  composition seems like a core limitation, which might make wide application difficult\n- The use case of user guidance is just on a proof-of concept level, and it’s unclear how it would transfer to an actual human scenario, however (again) this is appropriately acknowledged by the authors", "questions": "Questions/Needs clarification:\n- You mention “L.140 been extended beyond the model parameter to any univariate quantity of interest f(θ), such as validation loss”. Can you make the reference to using a validation dataset more explicit? Instead of the chained references at the section below (Koh&Liang, …)\n- How did you decide on the Mahalanobis distance as a baseline measure? Has this been considered before for this kind of task?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper describes the use of influence functions to evaluate the influence of preference samples in reward modeling. The authors apply recent methods for computing influence functions for large neural networks/LLM to reward model training. By using a validation set demonstrating preferences for specific qualities, the influence of examples is highlighted. The authors also propose a compression scheme to enable efficient storage of gradient vectors for dot products. The authors apply the method for the detection of biased data samples, and outline guidance of human labelers.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "Positive:\n- The paper is well written and polished, I found it easy to follow\n- The main presented methods seems to work well, and is evaluated appropriately\n- The different use cases (bias detection and user guidance) are presented well and relevant\n- Limitations are discussed (such as the dependence on validation set composition), and runs (although limited) ablations on the effects of different compositions\n- The quality of experiments seems high, both qualitative and quantitative results are generally convincing\n- Description of experiments/data sets is sufficient for reproducibility\n- The paper presents a nice application of influence functions\n\nOverall, I enjoyed reading the paper, and am convinced by its research quality. While strict technical novelty or theoretical insights are limited, I think it still provides a meaningful contribution to the field.", "weaknesses": "Negative:\n- The methdological novelty is limited, i.e. just the application of existing methods to a new use case\n- The dependency of the method on validation set  composition seems like a core limitation, which might make wide application difficult\n- The use case of user guidance is just on a proof-of concept level, and it’s unclear how it would transfer to an actual human scenario, however (again) this is appropriately acknowledged by the authors", "questions": "Questions/Needs clarification:\n- You mention “L.140 been extended beyond the model parameter to any univariate quantity of interest f(θ), such as validation loss”. Can you make the reference to using a validation dataset more explicit? Instead of the chained references at the section below (Koh&Liang, …)\n- How did you decide on the Mahalanobis distance as a baseline measure? Has this been considered before for this kind of task?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730411844328}, {"id": "umlLG2Iivb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5892/Reviewer_facp"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper addresses the noise, inconsistency, and bias in the reward model of RLHF tasks. The authors introduce a method using influence functions to quantify the impact of human feedback on reward models, enhancing understanding of both feedback and the models themselves.", "review_text": "This paper addresses the noise, inconsistency, and bias in the reward model of RLHF tasks. The authors introduce a method using influence functions to quantify the impact of human feedback on reward models, enhancing understanding of both feedback and the models themselves.", "strengths": "1.\tThe authors present a novel perspective on evaluating how human feedback affects the performance of reward models.\n\n2.\tThey employ vector compression techniques to approximate influence functions, reducing computational costs.\n\n3.\tThe paper highlights two potential applications of influence functions: detecting bias and assisting labelers.", "weaknesses": "1.\tEvaluation Set Requirement: The method requires a high-quality evaluation set for each alignment task, ideally annotated by experts. While this constraint is noted in the \"Limitations\" section, it presents a significant obstacle for real-world applications.\n\n2.\tComputational Complexity: The influence function requires calculating the inverse of the Hessian matrix. Although the authors approximate it using DataInf (Eq. 6) and vector compression technology, it also necessitates summing the data in the training set, which may compromise computational precision.\n\n3.\tBias Detection Methodology: The approach of identifying bias when the influence score exceeds a specified threshold is debatable. This threshold is difficult to determine, and not all samples surpassing it should be considered biased.\n4.\tExperimentation Concerns (Sec 5.2, Fig 6):\n\n    a.\tFinding expert labelers for every aspect is challenging in practical scenarios.\n\n    b.\tThis method may reduce the diversity of the generated responses, potentially negatively impacting model training.\n\n    c.\tThe experiment relies on training an SVM, which incurs additional computational costs and may lack accuracy.\n\n    d.\tIt might be more reasonable for the weight of non-experts to be adjusted based on the attribute and value of different rewards rather than introducing another labeler.", "questions": "1.\tEvaluation Set Size: How should the evaluation set size be determined? Is there any analysis on how performance improves with an increased number of samples?\n\n2.\tBias Types: Have you tested your method on biases other than length and sycophancy?\n\n3.\tMetric Choice: Why does the paper report ROC curves instead of other metrics?\n\n4.\tDependent Rewards: In real-world applications, rewards are often not independent. How does the proposed method address this in Sec 5.2?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the noise, inconsistency, and bias in the reward model of RLHF tasks. The authors introduce a method using influence functions to quantify the impact of human feedback on reward models, enhancing understanding of both feedback and the models themselves.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tThe authors present a novel perspective on evaluating how human feedback affects the performance of reward models.\n\n2.\tThey employ vector compression techniques to approximate influence functions, reducing computational costs.\n\n3.\tThe paper highlights two potential applications of influence functions: detecting bias and assisting labelers.", "weaknesses": "1.\tEvaluation Set Requirement: The method requires a high-quality evaluation set for each alignment task, ideally annotated by experts. While this constraint is noted in the \"Limitations\" section, it presents a significant obstacle for real-world applications.\n\n2.\tComputational Complexity: The influence function requires calculating the inverse of the Hessian matrix. Although the authors approximate it using DataInf (Eq. 6) and vector compression technology, it also necessitates summing the data in the training set, which may compromise computational precision.\n\n3.\tBias Detection Methodology: The approach of identifying bias when the influence score exceeds a specified threshold is debatable. This threshold is difficult to determine, and not all samples surpassing it should be considered biased.\n4.\tExperimentation Concerns (Sec 5.2, Fig 6):\n\n    a.\tFinding expert labelers for every aspect is challenging in practical scenarios.\n\n    b.\tThis method may reduce the diversity of the generated responses, potentially negatively impacting model training.\n\n    c.\tThe experiment relies on training an SVM, which incurs additional computational costs and may lack accuracy.\n\n    d.\tIt might be more reasonable for the weight of non-experts to be adjusted based on the attribute and value of different rewards rather than introducing another labeler.", "questions": "1.\tEvaluation Set Size: How should the evaluation set size be determined? Is there any analysis on how performance improves with an increased number of samples?\n\n2.\tBias Types: Have you tested your method on biases other than length and sycophancy?\n\n3.\tMetric Choice: Why does the paper report ROC curves instead of other metrics?\n\n4.\tDependent Rewards: In real-world applications, rewards are often not independent. How does the proposed method address this in Sec 5.2?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730182795136}], "openreview_url": "https://openreview.net/forum?id=dTQmayPKMs", "arxiv_id": "2501.05790", "paper_pdf": "papers/dTQmayPKMs.pdf", "paper_pdf_sha256": "4b5c1c488949e36a6503070217999598594d67257d064a6ceac2d1403b9800f5", "paper_pdf_bytes": 1440438, "paper_pdf_source": "openreview", "code_url": "https://github.com/mintaywon/IF_RLHF", "code_repository": "mintaywon/IF_RLHF", "code_commit": "e72cf5674faa4c19521a1856ee3558a564758062", "code_archive": "repos/dTQmayPKMs.zip", "code_archive_sha256": "0276d9f179d82f7bcc5fc4b97e4648393ee8f57c4dea60a8cbd8380aa1e28ffe", "code_archive_bytes": 108680, "code_file_count": 15, "code_extensions": {".py": 13, ".ipynb": 2}, "github_disk_usage_kb": 113, "github_languages": {"Python": 68845, "Jupyter Notebook": 14157}, "github_archived": false, "github_pushed_at": "2025-02-02T10:49:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-impact-of-human-feedback-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7VVGO0kuuY", "year": 2024, "status": "rejected", "title": "Learning Causal Dynamics Models in Object-Oriented Environments", "authors": ["Zhongwei Yu", "Jingqing Ruan", "Dengpeng Xing"], "authorids": ["~Zhongwei_Yu1", "~Jingqing_Ruan1", "~Dengpeng_Xing1"], "authors_source": "OpenReview API", "abstract": "Causal Dynamics Models (CDMs) have demonstrated significant potential in addressing various challenges in reinforcement learning. Recent studies have incorporated causal discovery to capture the causal dependencies among environmental variables in the learning of CDMs. However, the learning of CDMs is still confined to small-scale environments due to computational complexity and sample efficiency constraints. This paper aims to extend CDMs to large-scale object-oriented environments, which consist of a multitude of objects classified into different categories. We introduce the Object-Oriented CDM (OOCDM) that shares causalities and parameters among objects belonging to the same class. Furthermore, we propose a learning method for OOCDM that enables it to adapt to a varying number of objects. Experimental results from large-scale tasks indicate that OOCDM outperforms existing CDMs in terms of causal discovery, prediction accuracy, generalization, and computational efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "8lCfQfhA2q", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4429/Reviewer_6VtU"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper considers causal dynamics learning in an object-oriented environment, which is modeled using Object-oriented (OO) MDP. Based on the setting, the authors define OO causal graph that illustrates how attributes of different entities influence other attributes of possibly different entities at the next step. With a measure of conditional dependence (CMI) and flexible model (attention) for conditional probability involving OO, the authors demonstrated the usefulness of OO approach to causal dynamics modeling.", "review_text": "This paper considers causal dynamics learning in an object-oriented environment, which is modeled using Object-oriented (OO) MDP. Based on the setting, the authors define OO causal graph that illustrates how attributes of different entities influence other attributes of possibly different entities at the next step. With a measure of conditional dependence (CMI) and flexible model (attention) for conditional probability involving OO, the authors demonstrated the usefulness of OO approach to causal dynamics modeling.", "strengths": "- The use of OO concept in causal dynamics learning seems a very well-motivated work where real-world environments are naturally multi-agent setting with heterogeneous players.\n\n- I guess the use of attention to handle a varying number of objects’ attributes seems a clever idea. (Is this idea already adopted in other OO related research?, The authors only left a figure and a single sentence before Section 4.3)", "weaknesses": "As a researcher who worked on causal discovery in relational data and causal dynamics learning, this combination seems interesting. However,\n\n- The combination seems a bit not nontrivial in a sense that, while the fomulation is a bit complicated but, at a fundamental level, we just generalize causal dynamics learning to an object-oriented version, and apply the idea of conditional independence. Causal dynamics learning or causal discovery is just a transition probability between the two time steps, and we replace probabilities of random variables to represent probabilities among attributes of objects. We refine variables to consider only the relevant ones.\n\n- This OO causal graph or causal discovery is not a new concept (see Marc Maier and Prof. David Jensen (UMASS)’s work on relational causal discovery), which is a more general setting not just t-1 and t. Temporal version is also proposed (Marazopoulou et al. Learning the Structure of Causal Models with Relational and Temporal Dependence, UAI 2015) . There is no special needs to class-level local/global causality.", "questions": "- Regarding “Scalability”, do you mean this part “This causality sharing greatly simplifies causal discovery and improves the readability of CGs in large-scale environments. “ in page 4? If so, isn’t this an already existing result in relational causal discovery? Otherwise, where does the claim of scalability come from?\n=========\nafter the discussion, I see that the paper has certain merits (introducing OO into RL) and raising my rating from 5 to 6.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper considers causal dynamics learning in an object-oriented environment, which is modeled using Object-oriented (OO) MDP. Based on the setting, the authors define OO causal graph that illustrates how attributes of different entities influence other attributes of possibly different entities at the next step. With a measure of conditional dependence (CMI) and flexible model (attention) for conditional probability involving OO, the authors demonstrated the usefulness of OO approach to causal dynamics modeling.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The use of OO concept in causal dynamics learning seems a very well-motivated work where real-world environments are naturally multi-agent setting with heterogeneous players.\n\n- I guess the use of attention to handle a varying number of objects’ attributes seems a clever idea. (Is this idea already adopted in other OO related research?, The authors only left a figure and a single sentence before Section 4.3)", "weaknesses": "As a researcher who worked on causal discovery in relational data and causal dynamics learning, this combination seems interesting. However,\n\n- The combination seems a bit not nontrivial in a sense that, while the fomulation is a bit complicated but, at a fundamental level, we just generalize causal dynamics learning to an object-oriented version, and apply the idea of conditional independence. Causal dynamics learning or causal discovery is just a transition probability between the two time steps, and we replace probabilities of random variables to represent probabilities among attributes of objects. We refine variables to consider only the relevant ones.\n\n- This OO causal graph or causal discovery is not a new concept (see Marc Maier and Prof. David Jensen (UMASS)’s work on relational causal discovery), which is a more general setting not just t-1 and t. Temporal version is also proposed (Marazopoulou et al. Learning the Structure of Causal Models with Relational and Temporal Dependence, UAI 2015) . There is no special needs to class-level local/global causality.", "questions": "- Regarding “Scalability”, do you mean this part “This causality sharing greatly simplifies causal discovery and improves the readability of CGs in large-scale environments. “ in page 4? If so, isn’t this an already existing result in relational causal discovery? Otherwise, where does the claim of scalability come from?\n=========\nafter the discussion, I see that the paper has certain merits (introducing OO into RL) and raising my rating from 5 to 6.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698919645365}, {"id": "9aokoJp5VB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4429/Reviewer_mQjQ"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This manuscript introduces a comprehensive framework for object-oriented reinforcement learning (OORL), utilizing an object-oriented causal dynamic model to capture the intricacies of the OORL environment. The framework adeptly simplifies complex RL scenarios through a decomposition based on object types, ensuring that similar objects share parameters and causality characteristics. This integration of inductive biases such as causal symmetry and result symmetry enhances the model’s capability to represent complex scenarios effectively. The authors provide the framework to learn the corresponding model and the effectiveness of the proposed model is substantiated through a set of empirical experiments.\n\nOverall, this work stands out as an effective and logically sound approach to modeling the OORL environment, with a particular emphasis on causality. It is poised to make significant contributions to the field. Despite these strengths, there are certain elements of the method, presentation, and experimental design that will be further clarified by the authors. Addressing these points is crucial for a comprehensive understanding of the work. Given these considerations, my initial inclination is to recommend a borderline accept.", "review_text": "This manuscript introduces a comprehensive framework for object-oriented reinforcement learning (OORL), utilizing an object-oriented causal dynamic model to capture the intricacies of the OORL environment. The framework adeptly simplifies complex RL scenarios through a decomposition based on object types, ensuring that similar objects share parameters and causality characteristics. This integration of inductive biases such as causal symmetry and result symmetry enhances the model’s capability to represent complex scenarios effectively. The authors provide the framework to learn the corresponding model and the effectiveness of the proposed model is substantiated through a set of empirical experiments.\n\nOverall, this work stands out as an effective and logically sound approach to modeling the OORL environment, with a particular emphasis on causality. It is poised to make significant contributions to the field. Despite these strengths, there are certain elements of the method, presentation, and experimental design that will be further clarified by the authors. Addressing these points is crucial for a comprehensive understanding of the work. Given these considerations, my initial inclination is to recommend a borderline accept.", "strengths": "**[Motivation and General Idea]** The manuscript establishes a robust motivation, addressing the critical challenge of inefficiencies prevalent in Causal Dynamic Models (CDMs) within complex reinforcement learning (RL) scenarios. The authors’ choice to navigate RL in the context of multiple objects and categories is judicious and aligns well with the overarching theme of the paper.\n\n**[Proposed Framework]** The architectural design of the proposed framework is simple to follow and technically sound. The framework extends the classical paradigms of Object-Oriented Markov Decision Processes (OO-MDP) and Relational Markov Decision Processes (Relational MDP), incorporating causality as a vital inductive bias. This strategic design choice facilitates an empirically grounded solution, paving the way for efficient causal discovery in settings populated with numerous objects.", "weaknesses": "I listed the weaknesses and questions here. I am also a reviewer of the previous version of this paper. Most of my concerns have been addressed by the authors. However, some of them are still a bit unclear to me. \n\n- **[Definition of the local causality]** \n\nWhen we typically mention local causality, we refer to the case where the causal edges vary, as illustrated in Figure 2 in [1]. In this paper, when the authors mention local causality, it refers more to the causal graph of transitions for individual category objects. I think it is better to give a clarification in the revised version.\n\n- **[About potential relational interactions]**\n\nWhile the presented work demonstrates a comprehensive approach to object-oriented reinforcement learning, it appears to omit direct consideration of interactions between objects or entities within the framework. Given the prevalence of interactions among objects in real-world reinforcement learning scenarios, particularly in tasks involving object manipulation, this seems to be a significant aspect to address. Could the authors shed light on whether the framework is capable of learning and accounting for these interactions? Insights on how the model might be extended or adapted to incorporate direct interactions among objects would be greatly beneficial, as it would enhance the applicability of the approach to a wider array of real-world scenarios.\n\n- **[Potential to extend to image domains by combing the OCR models]**\n\nIn light of works such as [1-3], which integrate Object-Centric Representations (OCR) models into the learning process, the authors might consider incorporating similar OCR models equipped with permutation invariant modules at the beginning of their pipeline. This would enable the extraction of object representations directly from image data, making the method more scalable and applicable to a broader range of real-world scenarios.\n\n*References*\n\n[1] Pitis, Silviu, Elliot Creager, and Animesh Garg. \"Counterfactual data augmentation using locally factored dynamics.\" Advances in Neural Information Processing Systems 33 (2020): 3976-3990.\n\n[2] Yoon, Jaesik, et al. \"An investigation into pre-training object-centric representations for reinforcement learning.\" arXiv preprint arXiv:2302.04419 (2023).\n\n[3] Zadaianchuk, Andrii, Maximilian Seitzer, and Georg Martius. \"Self-supervised visual reinforcement learning with object-centric representations.\" arXiv preprint arXiv:2011.14381 (2020).\n\n[4] Kossen, Jannik, et al. \"Structured object-aware physics prediction for video modeling and planning.\" arXiv preprint arXiv:1910.02425 (2019).", "questions": "I listed the weaknesses and questions together in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript introduces a comprehensive framework for object-oriented reinforcement learning (OORL), utilizing an object-oriented causal dynamic model to capture the intricacies of the OORL environment. The framework adeptly simplifies complex RL scenarios through a decomposition based on object types, ensuring that similar objects share parameters and causality characteristics. This integration of inductive biases such as causal symmetry and result symmetry enhances the model’s capability to represent complex scenarios effectively. The authors provide the framework to learn the corresponding model and the effectiveness of the proposed model is substantiated through a set of empirical experiments.\n\nOverall, this work stands out as an effective and logically sound approach to modeling the OORL environment, with a particular emphasis on causality. It is poised to make significant contributions to the field. Despite these strengths, there are certain elements of the method, presentation, and experimental design that will be further clarified by the authors. Addressing these points is crucial for a comprehensive understanding of the work. Given these considerations, my initial inclination is to recommend a borderline accept.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "**[Motivation and General Idea]** The manuscript establishes a robust motivation, addressing the critical challenge of inefficiencies prevalent in Causal Dynamic Models (CDMs) within complex reinforcement learning (RL) scenarios. The authors’ choice to navigate RL in the context of multiple objects and categories is judicious and aligns well with the overarching theme of the paper.\n\n**[Proposed Framework]** The architectural design of the proposed framework is simple to follow and technically sound. The framework extends the classical paradigms of Object-Oriented Markov Decision Processes (OO-MDP) and Relational Markov Decision Processes (Relational MDP), incorporating causality as a vital inductive bias. This strategic design choice facilitates an empirically grounded solution, paving the way for efficient causal discovery in settings populated with numerous objects.", "weaknesses": "I listed the weaknesses and questions here. I am also a reviewer of the previous version of this paper. Most of my concerns have been addressed by the authors. However, some of them are still a bit unclear to me. \n\n- **[Definition of the local causality]** \n\nWhen we typically mention local causality, we refer to the case where the causal edges vary, as illustrated in Figure 2 in [1]. In this paper, when the authors mention local causality, it refers more to the causal graph of transitions for individual category objects. I think it is better to give a clarification in the revised version.\n\n- **[About potential relational interactions]**\n\nWhile the presented work demonstrates a comprehensive approach to object-oriented reinforcement learning, it appears to omit direct consideration of interactions between objects or entities within the framework. Given the prevalence of interactions among objects in real-world reinforcement learning scenarios, particularly in tasks involving object manipulation, this seems to be a significant aspect to address. Could the authors shed light on whether the framework is capable of learning and accounting for these interactions? Insights on how the model might be extended or adapted to incorporate direct interactions among objects would be greatly beneficial, as it would enhance the applicability of the approach to a wider array of real-world scenarios.\n\n- **[Potential to extend to image domains by combing the OCR models]**\n\nIn light of works such as [1-3], which integrate Object-Centric Representations (OCR) models into the learning process, the authors might consider incorporating similar OCR models equipped with permutation invariant modules at the beginning of their pipeline. This would enable the extraction of object representations directly from image data, making the method more scalable and applicable to a broader range of real-world scenarios.\n\n*References*\n\n[1] Pitis, Silviu, Elliot Creager, and Animesh Garg. \"Counterfactual data augmentation using locally factored dynamics.\" Advances in Neural Information Processing Systems 33 (2020): 3976-3990.\n\n[2] Yoon, Jaesik, et al. \"An investigation into pre-training object-centric representations for reinforcement learning.\" arXiv preprint arXiv:2302.04419 (2023).\n\n[3] Zadaianchuk, Andrii, Maximilian Seitzer, and Georg Martius. \"Self-supervised visual reinforcement learning with object-centric representations.\" arXiv preprint arXiv:2011.14381 (2020).\n\n[4] Kossen, Jannik, et al. \"Structured object-aware physics prediction for video modeling and planning.\" arXiv preprint arXiv:1910.02425 (2019).", "questions": "I listed the weaknesses and questions together in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698784105157}, {"id": "VqnYNyvlge", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4429/Reviewer_LMvE"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper discuss the current progress of learning and applying Causal Dynamics Models in training RL agents. They note that challenges arise in large-scale RL environments with vast number of objects and complex causal dependencies. Inspired by humans' object-oriented (OO) perspective for task perception, the authors developed the Object-Oriented Causal Dynamics Model (OOCDM). OOCDM allows sharing causalities and model parameters among objects from the same class, thus enhancing causal discovery and learning efficiency in RL. The proposed modified Causal Dynamics Learning (CDL) paradigm accommodates varying numbers of objects. Experimental results demonstrate the performance of OOCDM in terms of causal graph accuracy, prediction accuracy, generalization, and computational efficiency.", "review_text": "This paper discuss the current progress of learning and applying Causal Dynamics Models in training RL agents. They note that challenges arise in large-scale RL environments with vast number of objects and complex causal dependencies. Inspired by humans' object-oriented (OO) perspective for task perception, the authors developed the Object-Oriented Causal Dynamics Model (OOCDM). OOCDM allows sharing causalities and model parameters among objects from the same class, thus enhancing causal discovery and learning efficiency in RL. The proposed modified Causal Dynamics Learning (CDL) paradigm accommodates varying numbers of objects. Experimental results demonstrate the performance of OOCDM in terms of causal graph accuracy, prediction accuracy, generalization, and computational efficiency.", "strengths": "(1) the authors formally defined the problem of object oriented learning in RL from a causal perspective, which paved the way for future work.\n(2) the authors identified the core problem of the current causal dynamics model learning algorithms, which is the poor efficiency in the face of mass variables. The proposed method alleviated this problem to some extent with intuitive explanations and thorough justifications.\n(3) the writing is clear and easy to follow.", "weaknesses": "(1) Dynamics bayesian networks are rather limited rendering them unsuitable to model causal mechanisms. Especially when there is a need for cross layer inference, i.e., identifying the effect of an intervention from observational data. For more information, please see \"On Pearl’s hierarchy and the foundations of causal inference.\" in Probabilistic and Causal Inference: the works of Judea Pearl. In the specific scenario like the one define in this paper, it would be fine (Markovian, no confounders). But still, I would suggest the author use more rigorous causal tools to model the problem.\n\n(2) In section 3.2, when the authors started to define OOMDP, they claim that \"following (a Factored MDP paper), we formulate the task as an Object-Oriented MDP (OOMDP)...\". The way of presenting the concept totally ignores a seminal previous work that proposed OOMDP already [1]. Interestingly, the authors also referred to this paper in the related works section but seem to forget it here in the main text.\n\n(3) In section 4, the core assumption of this paper, \"causation symmetry\", which assumes that objects from the same class have the same causal effect on other objects, is too strong to be practical. Consider a system where three objects from the same class are linked one by one by springs in a line (obj1 --spring-- obj2 --spring-- obj3). For the causal effect of pulling obj2/obj3 on obj1, they are not interchangeable due to the existence of springs. \n\n(4) The proposed method requires manual construction of the OOMDP representation (obj classes, attributes fields, etc.). When there are truly a vast number of classes, this would be a new bottleneck for scaling. This fact, to some extent, hinders the authors' original goal of improving the efficiency of current CDMs learning methods. Similarly, when each object belongs to a distinct class, the proposed method might be even worse due to those detailed object fields.\n\n(5) For the experiments, the authors seem to be using a factored state space observation space. If it's not a high-dimensional pixel one, model-free methods are already competitive enough to solve state space observations problems. I doubt if there is really a strong need to introduce the extra complexities of the object oriented learning here. In another word, I do believe there are values in the OO representation and the proposed method even under the strong assumption of \"causation symmetry\", but it would be better demonstrated via high-dimensional challenging tasks.\n\n[1] Diuk, Carlos, Andre Cohen, and Michael L. Littman. \"An object-oriented representation for efficient reinforcement learning.\" Proceedings of the 25th international conference on Machine learning. 2008.", "questions": "(1) Could you elaborate the potential solution if people plan to apply your methods to environments with pixel observations?\n\n(2) Is there a more principled way of defining the object classes and attribute fields?\n\n(3) How would you compare your work with \"Causal dynamics learning for task-independent state abstraction.\" from ICML 22'? Is your proposed method significantly more efficient than theirs?\n\nI will consider change my rating if those questions and weaknesses are properly handled.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper discuss the current progress of learning and applying Causal Dynamics Models in training RL agents. They note that challenges arise in large-scale RL environments with vast number of objects and complex causal dependencies. Inspired by humans' object-oriented (OO) perspective for task perception, the authors developed the Object-Oriented Causal Dynamics Model (OOCDM). OOCDM allows sharing causalities and model parameters among objects from the same class, thus enhancing causal discovery and learning efficiency in RL. The proposed modified Causal Dynamics Learning (CDL) paradigm accommodates varying numbers of objects. Experimental results demonstrate the performance of OOCDM in terms of causal graph accuracy, prediction accuracy, generalization, and computational efficiency.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "(1) the authors formally defined the problem of object oriented learning in RL from a causal perspective, which paved the way for future work.\n(2) the authors identified the core problem of the current causal dynamics model learning algorithms, which is the poor efficiency in the face of mass variables. The proposed method alleviated this problem to some extent with intuitive explanations and thorough justifications.\n(3) the writing is clear and easy to follow.", "weaknesses": "(1) Dynamics bayesian networks are rather limited rendering them unsuitable to model causal mechanisms. Especially when there is a need for cross layer inference, i.e., identifying the effect of an intervention from observational data. For more information, please see \"On Pearl’s hierarchy and the foundations of causal inference.\" in Probabilistic and Causal Inference: the works of Judea Pearl. In the specific scenario like the one define in this paper, it would be fine (Markovian, no confounders). But still, I would suggest the author use more rigorous causal tools to model the problem.\n\n(2) In section 3.2, when the authors started to define OOMDP, they claim that \"following (a Factored MDP paper), we formulate the task as an Object-Oriented MDP (OOMDP)...\". The way of presenting the concept totally ignores a seminal previous work that proposed OOMDP already [1]. Interestingly, the authors also referred to this paper in the related works section but seem to forget it here in the main text.\n\n(3) In section 4, the core assumption of this paper, \"causation symmetry\", which assumes that objects from the same class have the same causal effect on other objects, is too strong to be practical. Consider a system where three objects from the same class are linked one by one by springs in a line (obj1 --spring-- obj2 --spring-- obj3). For the causal effect of pulling obj2/obj3 on obj1, they are not interchangeable due to the existence of springs. \n\n(4) The proposed method requires manual construction of the OOMDP representation (obj classes, attributes fields, etc.). When there are truly a vast number of classes, this would be a new bottleneck for scaling. This fact, to some extent, hinders the authors' original goal of improving the efficiency of current CDMs learning methods. Similarly, when each object belongs to a distinct class, the proposed method might be even worse due to those detailed object fields.\n\n(5) For the experiments, the authors seem to be using a factored state space observation space. If it's not a high-dimensional pixel one, model-free methods are already competitive enough to solve state space observations problems. I doubt if there is really a strong need to introduce the extra complexities of the object oriented learning here. In another word, I do believe there are values in the OO representation and the proposed method even under the strong assumption of \"causation symmetry\", but it would be better demonstrated via high-dimensional challenging tasks.\n\n[1] Diuk, Carlos, Andre Cohen, and Michael L. Littman. \"An object-oriented representation for efficient reinforcement learning.\" Proceedings of the 25th international conference on Machine learning. 2008.", "questions": "(1) Could you elaborate the potential solution if people plan to apply your methods to environments with pixel observations?\n\n(2) Is there a more principled way of defining the object classes and attribute fields?\n\n(3) How would you compare your work with \"Causal dynamics learning for task-independent state abstraction.\" from ICML 22'? Is your proposed method significantly more efficient than theirs?\n\nI will consider change my rating if those questions and weaknesses are properly handled.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698722650232}, {"id": "ZjOF2Vrg9L", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4429/Reviewer_dZfY"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors study the combination of causality and reinforcement learning by providing an extension to the setup of causal dynamics models through the lens of object oriented programming, which essentially provides a sharing of information based systems where the causal dynamics are shared across objects that belong to the same class. Further, the authors provide a learning mechanism based on key-value attention which enables generalization to arbitrary numbers of objects as long as the underlying classes remain the same. The authors’ experiments highlight that such an object oriented causal dynamics model (OOCDM) outperforms existing approaches along various fronts like causal discovery, performance accuracy and generalization capacity.", "review_text": "The authors study the combination of causality and reinforcement learning by providing an extension to the setup of causal dynamics models through the lens of object oriented programming, which essentially provides a sharing of information based systems where the causal dynamics are shared across objects that belong to the same class. Further, the authors provide a learning mechanism based on key-value attention which enables generalization to arbitrary numbers of objects as long as the underlying classes remain the same. The authors’ experiments highlight that such an object oriented causal dynamics model (OOCDM) outperforms existing approaches along various fronts like causal discovery, performance accuracy and generalization capacity.", "strengths": "- The authors are tackling a quite interesting and relevant problem and the approach used is well motivated and cognitively inspired. While it is studied for causality and RL approaches, it has been sparsely leveraged when combining the two fields in trying to model the causal dynamics of the environment itself and hence makes this work a very interesting and worthwhile read.\n- It is a natural extension to CDL which also further reduces the computations required owing to sharing of information between objects of the same class.\n- The authors test out on a variety of domains ranging from causal discovery and prediction accuracy as well as combining it with planning and highlight that the proposed approach is competitive and lucrative.", "weaknesses": "- While the work is quite interesting and very relevant to the field, the writing itself could use some work. The notation is not clearly specified over the paper and the distinction between bold and normal capital letters is hard to understand (eg. for U). For example, in Section 3.1, what does the random variable S represent? What happens if the transitions are time-depdendent?\n- The difference between $\\mathcal{F}$ and $\\mathcal{F}_s$ is not explained at all, and just directly used in Definition 2.\n- In Section 4.2, when the authors propose a model for implementing the prediction function, while the approach satisfies causation symmetry, it does not treat the dynamics induced by different kinds of classes in the same manner as opposed to differently. For example, the attributes $C_1.U_1$ should impact the predictor differently from the attributes $C_2.U_2$, while the key-value attention mechanism would imply using the same mechanism. Could the authors clarify this?\n- The authors should also take some space to explain the fundamentals of CIT / CMI methodology, and can then point to the Appendix for further clarity. However, without some background on CIT / CMI as well as how gradient based learning can be used to learn the graph structure (Equation 12), it is hard to follow the paper since graphs are discrete objects and gradient ascent based procedures only work on continuous spaces.\n\nFor the most part, my main qualms about the work are regarding the clarity and completeness of writing in the main draft. While the Appendix does contain a lot of details, I felt that there was some key information missing in the main draft and the preliminary section which makes the paper fairly dense and hard to read. There are also a few related works that the authors are missing, which I will mention below. I think clarifications on the math and making the notation and related work clearer in the main draft would go a long way, and I would be happy to increase my score if the changes were made.\n\n*Mittal, S., Bengio, Y., & Lajoie, G. (2022). Is a modular architecture enough?. Advances in Neural Information Processing Systems, 35, 28747-28760.*\n\n*Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., ... & Kipf, T. (2020). Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33, 11525-11538.*\n\n*Mittal, S., Lamb, A., Goyal, A., Voleti, V., Shanahan, M., Lajoie, G., ... & Bengio, Y. (2020, November). Learning to combine top-down and bottom-up signals in recurrent neural networks with attention over modules. In International Conference on Machine Learning (pp. 6972-6986). PMLR.*", "questions": "- After equation 1, the authors state that they assume the underlying graph is unknown. What about the probability distribution, is that known or also unknown?\n- What does it mean that the state variables transit independently? It would be nice if the authors could write it down in math, as it is not immediately clear for people who do not work on RL or the FMDP setup. I understand it is in the appendix but it would be nice to spell it once in the main text.\n- In equation 2, does the symbol used represent conditional independence, or the negative of that because it is not clear from the statement under the equation.\n- How does the theorem imply that testing every edge requires O(n) complexity?\n- In Equation 3, why does the distribution not depend on $\\mathbf{O}_{i+2}$ and so on? Further, what is the benefit of using semicolon in describing the conditioning, and why is $O_i.(\\mathbf{S}’)$ dependent on $\\mathbf{O}_i$ and not $\\mathbf{O}_i.\\mathbf{S}$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors study the combination of causality and reinforcement learning by providing an extension to the setup of causal dynamics models through the lens of object oriented programming, which essentially provides a sharing of information based systems where the causal dynamics are shared across objects that belong to the same class. Further, the authors provide a learning mechanism based on key-value attention which enables generalization to arbitrary numbers of objects as long as the underlying classes remain the same. The authors’ experiments highlight that such an object oriented causal dynamics model (OOCDM) outperforms existing approaches along various fronts like causal discovery, performance accuracy and generalization capacity.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- The authors are tackling a quite interesting and relevant problem and the approach used is well motivated and cognitively inspired. While it is studied for causality and RL approaches, it has been sparsely leveraged when combining the two fields in trying to model the causal dynamics of the environment itself and hence makes this work a very interesting and worthwhile read.\n- It is a natural extension to CDL which also further reduces the computations required owing to sharing of information between objects of the same class.\n- The authors test out on a variety of domains ranging from causal discovery and prediction accuracy as well as combining it with planning and highlight that the proposed approach is competitive and lucrative.", "weaknesses": "- While the work is quite interesting and very relevant to the field, the writing itself could use some work. The notation is not clearly specified over the paper and the distinction between bold and normal capital letters is hard to understand (eg. for U). For example, in Section 3.1, what does the random variable S represent? What happens if the transitions are time-depdendent?\n- The difference between $\\mathcal{F}$ and $\\mathcal{F}_s$ is not explained at all, and just directly used in Definition 2.\n- In Section 4.2, when the authors propose a model for implementing the prediction function, while the approach satisfies causation symmetry, it does not treat the dynamics induced by different kinds of classes in the same manner as opposed to differently. For example, the attributes $C_1.U_1$ should impact the predictor differently from the attributes $C_2.U_2$, while the key-value attention mechanism would imply using the same mechanism. Could the authors clarify this?\n- The authors should also take some space to explain the fundamentals of CIT / CMI methodology, and can then point to the Appendix for further clarity. However, without some background on CIT / CMI as well as how gradient based learning can be used to learn the graph structure (Equation 12), it is hard to follow the paper since graphs are discrete objects and gradient ascent based procedures only work on continuous spaces.\n\nFor the most part, my main qualms about the work are regarding the clarity and completeness of writing in the main draft. While the Appendix does contain a lot of details, I felt that there was some key information missing in the main draft and the preliminary section which makes the paper fairly dense and hard to read. There are also a few related works that the authors are missing, which I will mention below. I think clarifications on the math and making the notation and related work clearer in the main draft would go a long way, and I would be happy to increase my score if the changes were made.\n\n*Mittal, S., Bengio, Y., & Lajoie, G. (2022). Is a modular architecture enough?. Advances in Neural Information Processing Systems, 35, 28747-28760.*\n\n*Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., ... & Kipf, T. (2020). Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33, 11525-11538.*\n\n*Mittal, S., Lamb, A., Goyal, A., Voleti, V., Shanahan, M., Lajoie, G., ... & Bengio, Y. (2020, November). Learning to combine top-down and bottom-up signals in recurrent neural networks with attention over modules. In International Conference on Machine Learning (pp. 6972-6986). PMLR.*", "questions": "- After equation 1, the authors state that they assume the underlying graph is unknown. What about the probability distribution, is that known or also unknown?\n- What does it mean that the state variables transit independently? It would be nice if the authors could write it down in math, as it is not immediately clear for people who do not work on RL or the FMDP setup. I understand it is in the appendix but it would be nice to spell it once in the main text.\n- In equation 2, does the symbol used represent conditional independence, or the negative of that because it is not clear from the statement under the equation.\n- How does the theorem imply that testing every edge requires O(n) complexity?\n- In Equation 3, why does the distribution not depend on $\\mathbf{O}_{i+2}$ and so on? Further, what is the benefit of using semicolon in describing the conditioning, and why is $O_i.(\\mathbf{S}’)$ dependent on $\\mathbf{O}_i$ and not $\\mathbf{O}_i.\\mathbf{S}$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698629624470}, {"id": "XPMLja1alv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4429/Reviewer_miEU"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper introduces an object-oriented causal dynamics model (OOCDM) that leverages the hierarchical structure among state and action variables. Specifically, state and action variables are decomposed into a number of object variables, and the objects that belong to the same class share the same causality and transition function. Assuming that these object structures are given as prior knowledge, OOCDM learns the forward dynamics (i.e., transition function) and causal relationships between objects and fields. As OOCDM leverages the known hierarchical structure (i.e., objects and classes), it is more computationally efficient, and robust compared to previous CDMs.", "review_text": "The paper introduces an object-oriented causal dynamics model (OOCDM) that leverages the hierarchical structure among state and action variables. Specifically, state and action variables are decomposed into a number of object variables, and the objects that belong to the same class share the same causality and transition function. Assuming that these object structures are given as prior knowledge, OOCDM learns the forward dynamics (i.e., transition function) and causal relationships between objects and fields. As OOCDM leverages the known hierarchical structure (i.e., objects and classes), it is more computationally efficient, and robust compared to previous CDMs.", "strengths": "- The paper tackles an important problem and the motivation is clear. It is known that causal dynamics models (CDMs) are more robust compared to traditional dense dynamics models and it is important to improve the scalability of CDMs for their wider applicability.\n- The formulation of OOCDM which extends previous CDMs to object-oriented MDP representations seems novel, as far as I know.\n- The authors provide experimental results on larger-scale environments compared to prior works.", "weaknesses": "First, the manuscript is hard to follow due to the heavy notations and too much jargon, which could be more simplified. Second, the assumptions need more explanation. The *result symmetry* (eq. 3) is understandable, but the *causation symmetry* (eq. 4) does not seem to always hold. It would be better if the authors provide scenarios where this assumption holds and does not hold, and explain why it is reasonable to make the assumption.\n\nFinally, I have a concern regarding the experiments, e.g., the evaluation and interpretation of the results, which is also related to the main claim of the paper. Specifically, the authors claim that OOCDM outperforms prior CDMs in terms of (a) computational efficiency, (b) causal discovery, and (c) generalization of model-based RL. For (a), which is shown in Table 2, it is reasonable but also somewhat obvious since OOCDM leverages the hierarchical structure of the variables and common causal relationships shared with different objects. My major concern is at (b) and (c), which I elaborate on below.\n\n- Evaluation of causal discovery is not thorough. The performance of causal discovery is typically measured as structural hamming distance (SHD), or sometimes AUC and F1 score. The term “accuracy” in table 1 is confusing and details on how it is evaluated are missing.\n- OOFULL, which is the ablation of OOCDM without causal discovery, uses fully-connected bipartite graph. In other words, it only leverages the (A) *structural information* (objects belonging to the same class share the same dynamics), without the (B) *causal relationship* between variables. As shown in Table 3-4, OOFULL outperforms other CDMs. Given that CDMs work fairly well in causal discovery (as shown in Table 1), this implies that **(A) structural information (how variables are grouped with objects and classes) is the major contribution** to the performance gain of OOCDM. This poses additional questions, e.g., why and how does OOFULL (which is non-causal) generalize better than CDMs? It seems the assumptions are too strong (i.e., more important than causality). Also, how does the accurate understanding of causality lead to the improvement of OOCDM compared to OOFULL?\n- CDMs work fairly well in causal discovery (Table 1), but they fail on o.o.d data, as shown in Table 3. This is contrary to the findings of prior works (CDL, GRADER) and requires explanations.\n- In the evaluation of model-based RL (Table 4), the dynamics models are trained using only offline data. However, prior work (GRADER) suggests that it achieves better performance (in terms of both causal discovery and RL) when using online data. I would appreciate it if the authors could provide a justification for this.", "questions": "- Attention-based architecture is used to handle varying numbers of objects. In the experiments, does the number of objects vary in the same trajectory? If yes, then how do baselines handle varying input dimensions?\n- How is accuracy in Table 1 evaluated?\n- Table 1 shows that GRADER performs worse in causal discovery compared to other methods. This is counter-intuitive given that GRADER uses explicit conditional independence tests, unlike others.\n\n**Minor (mostly about notations and typos)**\n\n- Why does it need two separate notations in Def. 2? For example, $C_k\\left[U \\rightarrow V^{\\prime}\\right]$ can be written as $C_k\\left[C_k \\cdot U \\rightarrow V^{\\prime}\\right]$. It would be much easier to follow the paper with simple and consistent notations (maybe just write as $C_l \\cdot U \\rightarrow C_k \\cdot V'$?)\n- In eq. 4, $O_p, O_q$ should be $O_x, O_y$\n- The notation $f(\\cdot\\mid \\mathbf{O}; \\mathbf{U}_{-O}; \\mathcal{G})$ in Eq. 5 (as well as Eq. 8-10) is confusing. Isn’t it basically the same as $f(\\cdot\\mid \\mathbf{U}; \\mathcal{G})$? Also, notations in Eq. 8-10 are not consistent with Eq. 5.\n- In Sec 4.3, it says “$\\mathcal{G}_1$ is the full OOCG”. To my understanding, $\\mathcal{G}_1$ is a fully-connected bipartite graph and is not OOCG.\n- Table 7, description of $O.\\mathbf{S}$: action → state\n- Ground-truth causal graph provided in the appendix is hard to understand (the actual graph or adjacency matrix would be better).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces an object-oriented causal dynamics model (OOCDM) that leverages the hierarchical structure among state and action variables. Specifically, state and action variables are decomposed into a number of object variables, and the objects that belong to the same class share the same causality and transition function. Assuming that these object structures are given as prior knowledge, OOCDM learns the forward dynamics (i.e., transition function) and causal relationships between objects and fields. As OOCDM leverages the known hierarchical structure (i.e., objects and classes), it is more computationally efficient, and robust compared to previous CDMs.", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "strengths": "- The paper tackles an important problem and the motivation is clear. It is known that causal dynamics models (CDMs) are more robust compared to traditional dense dynamics models and it is important to improve the scalability of CDMs for their wider applicability.\n- The formulation of OOCDM which extends previous CDMs to object-oriented MDP representations seems novel, as far as I know.\n- The authors provide experimental results on larger-scale environments compared to prior works.", "weaknesses": "First, the manuscript is hard to follow due to the heavy notations and too much jargon, which could be more simplified. Second, the assumptions need more explanation. The *result symmetry* (eq. 3) is understandable, but the *causation symmetry* (eq. 4) does not seem to always hold. It would be better if the authors provide scenarios where this assumption holds and does not hold, and explain why it is reasonable to make the assumption.\n\nFinally, I have a concern regarding the experiments, e.g., the evaluation and interpretation of the results, which is also related to the main claim of the paper. Specifically, the authors claim that OOCDM outperforms prior CDMs in terms of (a) computational efficiency, (b) causal discovery, and (c) generalization of model-based RL. For (a), which is shown in Table 2, it is reasonable but also somewhat obvious since OOCDM leverages the hierarchical structure of the variables and common causal relationships shared with different objects. My major concern is at (b) and (c), which I elaborate on below.\n\n- Evaluation of causal discovery is not thorough. The performance of causal discovery is typically measured as structural hamming distance (SHD), or sometimes AUC and F1 score. The term “accuracy” in table 1 is confusing and details on how it is evaluated are missing.\n- OOFULL, which is the ablation of OOCDM without causal discovery, uses fully-connected bipartite graph. In other words, it only leverages the (A) *structural information* (objects belonging to the same class share the same dynamics), without the (B) *causal relationship* between variables. As shown in Table 3-4, OOFULL outperforms other CDMs. Given that CDMs work fairly well in causal discovery (as shown in Table 1), this implies that **(A) structural information (how variables are grouped with objects and classes) is the major contribution** to the performance gain of OOCDM. This poses additional questions, e.g., why and how does OOFULL (which is non-causal) generalize better than CDMs? It seems the assumptions are too strong (i.e., more important than causality). Also, how does the accurate understanding of causality lead to the improvement of OOCDM compared to OOFULL?\n- CDMs work fairly well in causal discovery (Table 1), but they fail on o.o.d data, as shown in Table 3. This is contrary to the findings of prior works (CDL, GRADER) and requires explanations.\n- In the evaluation of model-based RL (Table 4), the dynamics models are trained using only offline data. However, prior work (GRADER) suggests that it achieves better performance (in terms of both causal discovery and RL) when using online data. I would appreciate it if the authors could provide a justification for this.", "questions": "- Attention-based architecture is used to handle varying numbers of objects. In the experiments, does the number of objects vary in the same trajectory? If yes, then how do baselines handle varying input dimensions?\n- How is accuracy in Table 1 evaluated?\n- Table 1 shows that GRADER performs worse in causal discovery compared to other methods. This is counter-intuitive given that GRADER uses explicit conditional independence tests, unlike others.\n\n**Minor (mostly about notations and typos)**\n\n- Why does it need two separate notations in Def. 2? For example, $C_k\\left[U \\rightarrow V^{\\prime}\\right]$ can be written as $C_k\\left[C_k \\cdot U \\rightarrow V^{\\prime}\\right]$. It would be much easier to follow the paper with simple and consistent notations (maybe just write as $C_l \\cdot U \\rightarrow C_k \\cdot V'$?)\n- In eq. 4, $O_p, O_q$ should be $O_x, O_y$\n- The notation $f(\\cdot\\mid \\mathbf{O}; \\mathbf{U}_{-O}; \\mathcal{G})$ in Eq. 5 (as well as Eq. 8-10) is confusing. Isn’t it basically the same as $f(\\cdot\\mid \\mathbf{U}; \\mathcal{G})$? Also, notations in Eq. 8-10 are not consistent with Eq. 5.\n- In Sec 4.3, it says “$\\mathcal{G}_1$ is the full OOCG”. To my understanding, $\\mathcal{G}_1$ is a fully-connected bipartite graph and is not OOCG.\n- Table 7, description of $O.\\mathbf{S}$: action → state\n- Ground-truth causal graph provided in the appendix is hard to understand (the actual graph or adjacency matrix would be better).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698558900882}], "openreview_url": "https://openreview.net/forum?id=7VVGO0kuuY", "arxiv_id": "2405.12615", "paper_pdf": "papers/7VVGO0kuuY.pdf", "paper_pdf_sha256": "b946e7fc09051970674252bbbc7f488d24b01f49d67e958b1bf378b9c6c68b28", "paper_pdf_bytes": 1852025, "paper_pdf_source": "openreview", "code_url": "https://github.com/zwyu-ai/oocdm", "code_repository": "zwyu-ai/oocdm", "code_commit": "9bc24567986aa6b13be874d1a6599bef4fca02e8", "code_archive": "repos/7VVGO0kuuY.zip", "code_archive_sha256": "35697fa9de106b50f2c6a22fd8477f31ab0efc036148a77f3eb806744d5f8dac", "code_archive_bytes": 106512, "code_file_count": 63, "code_extensions": {".py": 63}, "github_disk_usage_kb": 90, "github_languages": {"Python": 341368}, "github_archived": false, "github_pushed_at": "2024-06-14T09:22:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-causal-dynamics-models-in-object"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mBkUeW8rpD6", "year": 2023, "status": "rejected", "title": "DiscoBAX - Discovery of optimal intervention sets in genomic experiment design", "authors": ["Clare Lyle", "Arash Mehrjou", "Pascal Notin", "Andrew Jesson", "Stefan Bauer", "Yarin Gal", "Patrick Schwab"], "authorids": ["~Clare_Lyle1", "~Arash_Mehrjou1", "~Pascal_Notin1", "~Andrew_Jesson1", "~Stefan_Bauer1", "~Yarin_Gal1", "~Patrick_Schwab1"], "authors_source": "OpenReview API", "abstract": "The discovery of novel therapeutics to cure genetic pathologies relies on the identification of the different genes involved in the underlying disease mechanism. With billions of potential hypotheses to test, an exhaustive  exploration of the entire space of potential interventions is impossible in practice. Sample-efficient methods based on active learning or bayesian optimization bear the promise of identifying interesting targets using the least experiments possible. However, genomic perturbation experiments typically rely on proxy outcomes measured in biological model systems that may not completely correlate with the outcome of interventions in humans. In practical experiment design, one aims to find a set of interventions which maximally move a target phenotype via a diverse set of mechanisms in order to reduce the risk of failure in future stages of trials. To that end, we introduce DiscoBAX — a sample-efficient algorithm for the discovery of genetic interventions that maximize the movement of a phenotype in a direction of interest while covering a diverse set of underlying mechanisms. We provide theoretical guarantees on the optimality of the approach under standard assumptions, conduct extensive experiments in synthetic and real-world settings relevant to genomic discovery and demonstrate that DiscoBAX outperforms state-of-the-art active learning and Bayesian optimization methods in this task. Better methods for selecting effective and diverse perturbations in biological systems could enable researchers to potentially discover novel therapeutics for a range of genetically-driven diseases.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Rb4R8-gV2xJ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4798/Reviewer_3hGy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work presents a new probabilistic algorithm (\"DiscoBAX\") for subset selection that aims to approximately optimize phenotype movement in genomic intervention and can be useful in drug discovery tasks according to the authors. The method identifies a set of interventions whose elements will trigger maximum expected change in the disease outcome with respect to for some noise distribution. Performance is assessed based on synthetic data as well as public benchmarking data, and the algorithm is shown to outperform alternative approaches.\n\n", "review_text": "The work provides an interesting approach to an optimization problem that is initially motivated by certain genomic drug discovery tasks. Overall the reporting is clear and appears reproducible. The algorithmic innovation is in place but focuses on a relatively specific task definition where the applied motivation remains questionable, and it is not clear whether overfitting has been sufficiently controlled as the algorithm appears rather flexible. The paper has figure numbering problems. \n\n\n", "strengths": "Strengths:\n+ probabilistic approach is neat for uncertainty characterization\n+ practical combination of task, method derivation, and algorithm implementation \n+ Openly licensed source code for the algorithms available\n\nWeaknesses:\n- Figure numberings are broken, text cites figure up to Fig.5, manuscript only has 3 figures, refs to Figs 3-4 are missing; not clear how to follow\n- The problem motivation is hypothetical; it is not clear that \"maximizing change\" is what one would like to achieve in genomic drug discovery; on the other hand the algorithm is of interest regardless but perhaps best evaluated on its own right as a target optimization task.\n- I am not sure if potential overfitting is sufficiently addressed in this work; explanation on this part could be strengthened ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work presents a new probabilistic algorithm (\"DiscoBAX\") for subset selection that aims to approximately optimize phenotype movement in genomic intervention and can be useful in drug discovery tasks according to the authors. The method identifies a set of interventions whose elements will trigger maximum expected change in the disease outcome with respect to for some noise distribution. Performance is assessed based on synthetic data as well as public benchmarking data, and the algorithm is shown to outperform alternative approaches.\n\n", "strength_and_weaknesses": "Strengths:\n+ probabilistic approach is neat for uncertainty characterization\n+ practical combination of task, method derivation, and algorithm implementation \n+ Openly licensed source code for the algorithms available\n\nWeaknesses:\n- Figure numberings are broken, text cites figure up to Fig.5, manuscript only has 3 figures, refs to Figs 3-4 are missing; not clear how to follow\n- The problem motivation is hypothetical; it is not clear that \"maximizing change\" is what one would like to achieve in genomic drug discovery; on the other hand the algorithm is of interest regardless but perhaps best evaluated on its own right as a target optimization task.\n- I am not sure if potential overfitting is sufficiently addressed in this work; explanation on this part could be strengthened ", "clarity,_quality,_novelty_and_reproducibility": "Quality:\n- Overall the paper is well written; the problem motivation and validation part (esp. overfitting; see comments) could be strengthened. Figure numbering is broken.\n\nClarity:\n- Text is easy to read and follow\n- Some more intuitive descriptions of the algorithm could be added\n- Figure numbering is broken.\n\nOriginality:\n- The work provides a new solution to a previously established problem and claims performance gains\n- The combination of problem and algorithmic solution seems to be new \n\nNovelty And Reproducibility:\n- I did not try to replicate the work but code is well organized and openly licensed, and seems robust. Some more guidance in the README landing page would be warranted.\n", "summary_of_the_review": "The work provides an interesting approach to an optimization problem that is initially motivated by certain genomic drug discovery tasks. Overall the reporting is clear and appears reproducible. The algorithmic innovation is in place but focuses on a relatively specific task definition where the applied motivation remains questionable, and it is not clear whether overfitting has been sufficiently controlled as the algorithm appears rather flexible. The paper has figure numbering problems. \n\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1668009050780}, {"id": "C2XkF7OOJF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4798/Reviewer_efcP"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper considers the problem of designing an experiment where there are two stages: in the first (in vitro) stage our task is to (efficiently) design an experiment that will have good chance of success in the second (in vivo) stage. The authors formalize this elegantly in equation (3).\n\nThe proposed solution is discoBAX: an algorithm that actively searches the in-vitro function for a set of optimal yet diverse points that have the best shot at having a successful outcome in the in vivo stage. \n\nThe paper explores discoBAX on a synthetic and real-world dataset.\n\n", "review_text": "A really interesting setting to be working in, and what seems like a good idea, but a lack of empirical investigation into the method lets the paper down. Clarity is a problem and it's difficult to assess which parts of the work are a novelty. I could change my \"technical novelty' score if the authors could clarify.", "strengths": "Strengths\n--\nAfter spending some time with the paper, I came to appreciate the formalisation of the setting that is being worked in. I particularly enjoyed section 3 which laid out the problem. I appreciate the authors use of the appendices to keep long proofs out of the paper (the submodular proof for example) and also the clear presentation of the algorithms used. Posting the code anonymously gives me some confidence that the work is sound. \n\nWeaknesses\n--\nthe paper fails to deliver what is promised at the start of the experimental section - an investigation of discoBAX's sample efficiency, sensitivity to hyper-parameter settings, and comparison with the top-intervention method. As a result I feel that the paper is really rather unfinished! I am keen to understand more about discoBAX and am disappointed that the investigation is not deeper. \n\nThe 'related work' section at the back does not cover BAX - which appear to be prior art? this section feels like filler, I would have preferred the space be used to investigate discoBAX further. \n\nI did not find a satisfactory discussion (or empirical evaluation) of the use of the relu'd score function, relative to the original objective. I get that this makes the optimization submodular, but have we lost anything? In which cases does this cause a problem?\n\nIt took me some time to work out what was going on - I think more discussion around Figure 1 is needed, and perhaps some commentary on how this is connected to the invitro/in-vivo setting. I liked the clarity of eq 3, but this felt like a long time coming, and would have been better placed at the start of the paper perhaps?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers the problem of designing an experiment where there are two stages: in the first (in vitro) stage our task is to (efficiently) design an experiment that will have good chance of success in the second (in vivo) stage. The authors formalize this elegantly in equation (3).\n\nThe proposed solution is discoBAX: an algorithm that actively searches the in-vitro function for a set of optimal yet diverse points that have the best shot at having a successful outcome in the in vivo stage. \n\nThe paper explores discoBAX on a synthetic and real-world dataset.\n\n", "strength_and_weaknesses": "Strengths\n--\nAfter spending some time with the paper, I came to appreciate the formalisation of the setting that is being worked in. I particularly enjoyed section 3 which laid out the problem. I appreciate the authors use of the appendices to keep long proofs out of the paper (the submodular proof for example) and also the clear presentation of the algorithms used. Posting the code anonymously gives me some confidence that the work is sound. \n\nWeaknesses\n--\nthe paper fails to deliver what is promised at the start of the experimental section - an investigation of discoBAX's sample efficiency, sensitivity to hyper-parameter settings, and comparison with the top-intervention method. As a result I feel that the paper is really rather unfinished! I am keen to understand more about discoBAX and am disappointed that the investigation is not deeper. \n\nThe 'related work' section at the back does not cover BAX - which appear to be prior art? this section feels like filler, I would have preferred the space be used to investigate discoBAX further. \n\nI did not find a satisfactory discussion (or empirical evaluation) of the use of the relu'd score function, relative to the original objective. I get that this makes the optimization submodular, but have we lost anything? In which cases does this cause a problem?\n\nIt took me some time to work out what was going on - I think more discussion around Figure 1 is needed, and perhaps some commentary on how this is connected to the invitro/in-vivo setting. I liked the clarity of eq 3, but this felt like a long time coming, and would have been better placed at the start of the paper perhaps?\n", "clarity,_quality,_novelty_and_reproducibility": "I have to give the authors top marks for reproducibility since the code is available. \n\nA problem with the paper is that it is hard to me to assess novelty. Although the authors do lay out their contributions at the beginning, it's not clear to me whether this paper or another is the source of the technical inventions. For example, the submodularity of the relu'd score function. \n\nA minor (clarity) point, but the figures are poor quality. I cannot read the axis labels when printed out. Unfortunately this gives the impression of a rushed paper. ", "summary_of_the_review": "A really interesting setting to be working in, and what seems like a good idea, but a lack of empirical investigation into the method lets the paper down. Clarity is a problem and it's difficult to assess which parts of the work are a novelty. I could change my \"technical novelty' score if the authors could clarify.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1668002623135}, {"id": "dr0GXZ4RVma", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4798/Reviewer_JiqB"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper is concerned with the iterative selection of an optimal set of targets for genetic interventions based on a scalar readout. The methods section described a particular instance of BAX, that models the singularity of the biological problem that is considered. A few aspects that makes it different from the existing InfoBAX approach is (1) the subset maximization target function and (2) a practical uncertainty model with BNN instead of GPs. Then, the paper provides a set of experiments on synthetic data, and on the openly available GeneDisco benchmark. ", "review_text": "This paper provides an interesting application of BAX, and investigate a practical variant to batch learning of genetic perturbations. Beyond this, the technical contributions may be considered weak, and the empirical results are not exactly strong either (please refer to my previous sections for actionable items). Regarding the experimental results, I completely understand that it might be caused by noise in some datasets (e.g., COVID). In this case, I do think that the authors should carefully present ablation studies so that readers can understand what features of DiscoBAX helps in increase performance compared to BAX (again, see above for suggestions). In the light of this, I am currently leaning towards rejection of this manuscript, but very much looking forward to discussions with authors, and other reviewers.", "strengths": "Strength:\n(1) Novelty *in its application*: this is one of the (relatively) few papers that tackle this important scientific problem (other example I could find is https://arxiv.org/pdf/2207.12805.pdf).\n(2) The paper shows improved results in the recent GeneDisco benchmark, for a few of the datasets.\n\nWeaknesses:\n(1) My main issue is that I am having a hard time understanding the *technical* novelty of the paper. A lot of the concepts explained in Section 4 seems to belong more to a background section, as they are not contributions of this particular work, but from the BAX paper. As far as I understand, the contributions are (a) working with sets, and performing greedy set optimization and (b) using BNN for uncertainty quantification. However, in the results section, I don't see a clear explanation that those changes are improving the performance. It seems that (a) improves diversity? (if so, it should be made clear), but (b) has no comparison to different flavors of BNNs, GPs etc..\n(2) there is no theoretical analysis on how the (1 - 1/e) approximation error propagates through the BAX procedure. Then, it is not correct for the author to write \"theoretical guarantees on the optimality of the approach\". Similarly, the authors write that DiscoBAX is \"sample-efficient\", but I have seen no proofs of sample efficiency, and the experimental results are not necessarily strong enough to claim this.\n(3) There is some improvement over BAX, but only in a few datasets (2 / 4). \n(4) There should be a comparison to other BNN flavors, and at the very least a discussion of uncertainty modeling w/ neural networks\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper is concerned with the iterative selection of an optimal set of targets for genetic interventions based on a scalar readout. The methods section described a particular instance of BAX, that models the singularity of the biological problem that is considered. A few aspects that makes it different from the existing InfoBAX approach is (1) the subset maximization target function and (2) a practical uncertainty model with BNN instead of GPs. Then, the paper provides a set of experiments on synthetic data, and on the openly available GeneDisco benchmark. ", "strength_and_weaknesses": "Strength:\n(1) Novelty *in its application*: this is one of the (relatively) few papers that tackle this important scientific problem (other example I could find is https://arxiv.org/pdf/2207.12805.pdf).\n(2) The paper shows improved results in the recent GeneDisco benchmark, for a few of the datasets.\n\nWeaknesses:\n(1) My main issue is that I am having a hard time understanding the *technical* novelty of the paper. A lot of the concepts explained in Section 4 seems to belong more to a background section, as they are not contributions of this particular work, but from the BAX paper. As far as I understand, the contributions are (a) working with sets, and performing greedy set optimization and (b) using BNN for uncertainty quantification. However, in the results section, I don't see a clear explanation that those changes are improving the performance. It seems that (a) improves diversity? (if so, it should be made clear), but (b) has no comparison to different flavors of BNNs, GPs etc..\n(2) there is no theoretical analysis on how the (1 - 1/e) approximation error propagates through the BAX procedure. Then, it is not correct for the author to write \"theoretical guarantees on the optimality of the approach\". Similarly, the authors write that DiscoBAX is \"sample-efficient\", but I have seen no proofs of sample efficiency, and the experimental results are not necessarily strong enough to claim this.\n(3) There is some improvement over BAX, but only in a few datasets (2 / 4). \n(4) There should be a comparison to other BNN flavors, and at the very least a discussion of uncertainty modeling w/ neural networks\n", "clarity,_quality,_novelty_and_reproducibility": "The paper does not make its contributions clear, and I do think that it is currently an issue. Otherwise, the paper reads well. The application of BAX to genomics is novel (but the practical implications for the field of genetics are rather unclear), but the improvement over BAX is not clear (as the paper is written now). ", "summary_of_the_review": "This paper provides an interesting application of BAX, and investigate a practical variant to batch learning of genetic perturbations. Beyond this, the technical contributions may be considered weak, and the empirical results are not exactly strong either (please refer to my previous sections for actionable items). Regarding the experimental results, I completely understand that it might be caused by noise in some datasets (e.g., COVID). In this case, I do think that the authors should carefully present ablation studies so that readers can understand what features of DiscoBAX helps in increase performance compared to BAX (again, see above for suggestions). In the light of this, I am currently leaning towards rejection of this manuscript, but very much looking forward to discussions with authors, and other reviewers.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666537596999}, {"id": "A0TK4ZgnQK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4798/Reviewer_uzVy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of finding a `diverse` set of genomic interventions that maximize a phenotype of interest. Formally, the problem can be modeled as optimizing expensive black-box functions over sets of inputs from a given design space. Bayesian Algorithm eXecution (BAX) is a recently proposed framework that allows estimating required properties of black-box functions using information gain based input acquisition strategies. The paper provides a BAX (Bayesian `Algorithm` eXecution) style approach to solve the problem where the key idea is to consider the subset maximization of a chosen score function as the `Algorithm`. At each iteration, an Expected Information Gain acquisition objective is optimized to select the next points for evaluation. This objective is parameterized by multiple sample outputs of the subset maximization algorithm. Experiments are performed on synthetic and GeneDisco benchmarks.", "review_text": "I find the overall approach to be quite interesting while addressing an important problem. However, some of the design choices and baseline comparisons require more work to improve the quality of the paper for an ICLR publication. ", "strengths": "- The problem considered in the paper is important with real-world implications. I found the biological motivation of the problem well-written with good intuitions for a non-domain person.\n\n- The overall idea of using BAX style approach for this problem setting is fairly interesting and novel. \n\nHowever, I have few questions  to understand some of the details better and some suggestions that will hopefully improve the paper's contributions:\n\n- It is mentioned in the paper that the discrepancy between the disease outcome $f_{out}$ and intermediate phenotype $f_{ip}$ is captured by the noise distribution $\\eta$. It is motivated as an important distinguishing factor between the problem setting of this paper compared to that of existing approaches like Bayesian optimization. However, there is little clear description about estimating this quantity or principles behind choosing a certain distribution. A short remark is mentioned after observation 1 in few lines but that seems limited for such an important motivation of the problem setting. Please describe the concrete implementation/algorithmic details of the noise distribution and how it is estimated in the experiments. \n\n- The proof of submodularity of the score function in appendix A is missing some description. For example, $dis(g, \\eta)$ is not defined. Please expand the proof and explain the reasoning behind the inequalities. If it is `straightforward` (as mentioned in the first line of the proof), can it be considered a major contribution of the paper? \n\n- The score function is chosen to be the best value in a set (averaged over noise distribution eta). This seems like an optimistic choice. For the algorithm to find good robust points, should we not consider the worst value in the set as the score function?\n\n- Please add a description of the computational complexity of the proposed approach. Since multiple instances of the subset maximization algorithm needs to be run in each iteration for estimating the expected information gain quantity, the computational complexity can be quite high. A wall-clock time comparison of the proposed approach with existing baselines will be very useful, especially since the top-k recall gap is relatively small on most of the benchmarks (other than Leukemia/NK cells).\n\n- As mentioned in section 5.2, batch size of 32 points are selected for evaluation in each acquisition cycle. How is the batch of points selected? Is it accomplished by greedy optimization of the EIG acquisition function? If yes, what are the accuracy losses of doing this greedy optimization. \n\n- The choice of UCB as Bayesian Optimization (BO) baseline strategy seems surprising. There is a large literature on Bayesian optimization algorithms (please see [1-4] and references therein) for the batch evaluation setting that should be the right comparison here. Something like qEI (q-Expected Improvement) is easy to setup and implemented in popular packages like BoTorch [1] (https://botorch.org/tutorials/). The points suggested by qEI are also known to find highly diverse points in the context of batch optimization. Please consider improving this comparison (by including qEI for instance) because BO algorithms are directly applicable and the most relevant baseline in the setting. \n\nReferences \n\n\n[1] Balandat, Maximilian, Brian Karrer, Daniel Jiang, Samuel Daulton, Ben Letham, Andrew G. Wilson, and Eytan Bakshy. \"BoTorch: a framework for efficient Monte-Carlo Bayesian optimization.\" Advances in neural information processing systems 33 (2020): 21524-21538.\nGonzález, J., Dai, Z., Hennig, P., & Lawrence, N. (2016, May). Batch Bayesian optimization via local penalization. In Artificial intelligence and statistics (pp. 648-657). PMLR.\n\n[2] Wu, J., & Frazier, P. (2016). The parallel knowledge gradient method for batch Bayesian optimization. Advances in neural information processing systems, 29.\n\n[3] Azimi, J., Fern, A., & Fern, X. (2010). Batch bayesian optimization via simulation matching. Advances in Neural Information Processing Systems, 23.\n\n[4] Gong, C., Peng, J., & Liu, Q. (2019, May). Quantile stein variational gradient descent for batch Bayesian optimization. In International Conference on Machine Learning (pp. 2347-2356). PMLR.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers the problem of finding a `diverse` set of genomic interventions that maximize a phenotype of interest. Formally, the problem can be modeled as optimizing expensive black-box functions over sets of inputs from a given design space. Bayesian Algorithm eXecution (BAX) is a recently proposed framework that allows estimating required properties of black-box functions using information gain based input acquisition strategies. The paper provides a BAX (Bayesian `Algorithm` eXecution) style approach to solve the problem where the key idea is to consider the subset maximization of a chosen score function as the `Algorithm`. At each iteration, an Expected Information Gain acquisition objective is optimized to select the next points for evaluation. This objective is parameterized by multiple sample outputs of the subset maximization algorithm. Experiments are performed on synthetic and GeneDisco benchmarks.", "strength_and_weaknesses": "- The problem considered in the paper is important with real-world implications. I found the biological motivation of the problem well-written with good intuitions for a non-domain person.\n\n- The overall idea of using BAX style approach for this problem setting is fairly interesting and novel. \n\nHowever, I have few questions  to understand some of the details better and some suggestions that will hopefully improve the paper's contributions:\n\n- It is mentioned in the paper that the discrepancy between the disease outcome $f_{out}$ and intermediate phenotype $f_{ip}$ is captured by the noise distribution $\\eta$. It is motivated as an important distinguishing factor between the problem setting of this paper compared to that of existing approaches like Bayesian optimization. However, there is little clear description about estimating this quantity or principles behind choosing a certain distribution. A short remark is mentioned after observation 1 in few lines but that seems limited for such an important motivation of the problem setting. Please describe the concrete implementation/algorithmic details of the noise distribution and how it is estimated in the experiments. \n\n- The proof of submodularity of the score function in appendix A is missing some description. For example, $dis(g, \\eta)$ is not defined. Please expand the proof and explain the reasoning behind the inequalities. If it is `straightforward` (as mentioned in the first line of the proof), can it be considered a major contribution of the paper? \n\n- The score function is chosen to be the best value in a set (averaged over noise distribution eta). This seems like an optimistic choice. For the algorithm to find good robust points, should we not consider the worst value in the set as the score function?\n\n- Please add a description of the computational complexity of the proposed approach. Since multiple instances of the subset maximization algorithm needs to be run in each iteration for estimating the expected information gain quantity, the computational complexity can be quite high. A wall-clock time comparison of the proposed approach with existing baselines will be very useful, especially since the top-k recall gap is relatively small on most of the benchmarks (other than Leukemia/NK cells).\n\n- As mentioned in section 5.2, batch size of 32 points are selected for evaluation in each acquisition cycle. How is the batch of points selected? Is it accomplished by greedy optimization of the EIG acquisition function? If yes, what are the accuracy losses of doing this greedy optimization. \n\n- The choice of UCB as Bayesian Optimization (BO) baseline strategy seems surprising. There is a large literature on Bayesian optimization algorithms (please see [1-4] and references therein) for the batch evaluation setting that should be the right comparison here. Something like qEI (q-Expected Improvement) is easy to setup and implemented in popular packages like BoTorch [1] (https://botorch.org/tutorials/). The points suggested by qEI are also known to find highly diverse points in the context of batch optimization. Please consider improving this comparison (by including qEI for instance) because BO algorithms are directly applicable and the most relevant baseline in the setting. \n\nReferences \n\n\n[1] Balandat, Maximilian, Brian Karrer, Daniel Jiang, Samuel Daulton, Ben Letham, Andrew G. Wilson, and Eytan Bakshy. \"BoTorch: a framework for efficient Monte-Carlo Bayesian optimization.\" Advances in neural information processing systems 33 (2020): 21524-21538.\nGonzález, J., Dai, Z., Hennig, P., & Lawrence, N. (2016, May). Batch Bayesian optimization via local penalization. In Artificial intelligence and statistics (pp. 648-657). PMLR.\n\n[2] Wu, J., & Frazier, P. (2016). The parallel knowledge gradient method for batch Bayesian optimization. Advances in neural information processing systems, 29.\n\n[3] Azimi, J., Fern, A., & Fern, X. (2010). Batch bayesian optimization via simulation matching. Advances in Neural Information Processing Systems, 23.\n\n[4] Gong, C., Peng, J., & Liu, Q. (2019, May). Quantile stein variational gradient descent for batch Bayesian optimization. In International Conference on Machine Learning (pp. 2347-2356). PMLR.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The problem setting is clearly defined and the usage of BAX style approach is novel. It is commendable that source code is made available for easy reproducibility. ", "summary_of_the_review": "I find the overall approach to be quite interesting while addressing an important problem. However, some of the design choices and baseline comparisons require more work to improve the quality of the paper for an ICLR publication. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666402470513}], "openreview_url": "https://openreview.net/forum?id=mBkUeW8rpD6", "arxiv_id": "2312.04064", "paper_pdf": "papers/mBkUeW8rpD6.pdf", "paper_pdf_sha256": "7003ec1b68d1f9e6729494fcefdac099451c36ac4ead1f5c27b8f8fd96ef4eaf", "paper_pdf_bytes": 804941, "paper_pdf_source": "openreview", "code_url": "https://github.com/amehrjou/DiscoBAX", "code_repository": "amehrjou/DiscoBAX", "code_commit": "84f01283bc7f6ab5f66b5ea2a63632b401cc0402", "code_archive": "repos/mBkUeW8rpD6.zip", "code_archive_sha256": "3f1b7708f0472388438757b64a9ba00e1844bffddcb6bfca6194bed954d628ca", "code_archive_bytes": 187751, "code_file_count": 37, "code_extensions": {".py": 35, ".ipynb": 1, ".sh": 1}, "github_disk_usage_kb": 179, "github_languages": {"Python": 209151, "Jupyter Notebook": 138840, "Shell": 2424}, "github_archived": false, "github_pushed_at": "2024-01-21T15:54:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discobax-discovery-of-optimal-intervention"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wNsNT56zDkG", "year": 2022, "status": "rejected", "title": "Adversarial Rademacher Complexity of Deep Neural Networks", "authors": ["Jiancong Xiao", "Yanbo Fan", "Ruoyu Sun", "Zhi-Quan Luo"], "authorids": ["~Jiancong_Xiao1", "~Yanbo_Fan1", "~Ruoyu_Sun1", "~Zhi-Quan_Luo1"], "authors_source": "OpenReview API", "abstract": "Deep neural networks are vulnerable to adversarial attacks. Adversarial training is one of the most effective algorithms to increase the model's robustness. However, the trained models cannot generalize well to the adversarial examples on the test set. In this paper, we study the generalization of adversarial training through the lens of adversarial Rademacher complexity. Current analysis of adversarial Rademacher complexity is up to two-layer neural networks. In adversarial settings, one major difficulty of generalizing these results to deep neural networks is that we cannot peel off the layer as the classical analysis for standard training. We provide a method to overcome this issue and provide upper bounds of adversarial Rademacher complexity of deep neural networks. Similar to the existing bounds of standard Rademacher complexity of neural nets, our bound also includes the product of weight norms. We provide experiments to show that the adversarially trained weight norms are larger than the standard trained weight norms, thus providing an explanation for the bad generalization performance of adversarial training.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "9LnLFyUo1gc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2864/Reviewer_Ma5T"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The main result of the paper is an upper bound on the Rademacher Complexity of neural networks in the case of adversarial examples (meaning, analyzing it with respect to the robust loss class).\n\nPrevious bounds for linear classifiers and shallow neural networks were investigated by Yin et al. (2019) and Awasthi et al. (2020).\nThe known methods for upper bounding the Rademacher Complexity for deep neural networks in the standard case don't apply in this case.\n\nThe main idea in this paper is to analyze the covering numbers directly (as opposed to calculating them by induction on the layers).\n ", "review_text": "The question investigated in the paper (described above) is natural and was posed as an open question by previous papers.\n\nThe main technique seems reasonable. \n\nThe experiments suggest that the change in the margin and the product of the weight norm may explain the larger generalization gap  (compared to the standard generalization).\n\nWeakness: showing a depth/width-dependent lower bound could have been nice.\n\n-Citing the paper https://arxiv.org/pdf/1810.02180.pdf (ALT19 and JMLR) is very relevant in the section \"Adversarial Generalization\".\nIt provides uniform convergence results In the case of a finite number of perturbations for a mixture of classifiers. The analysis goes through the Rademacher complexity as well.\n\n-In section 3: d^2 should be 2^d?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The main result of the paper is an upper bound on the Rademacher Complexity of neural networks in the case of adversarial examples (meaning, analyzing it with respect to the robust loss class).\n\nPrevious bounds for linear classifiers and shallow neural networks were investigated by Yin et al. (2019) and Awasthi et al. (2020).\nThe known methods for upper bounding the Rademacher Complexity for deep neural networks in the standard case don't apply in this case.\n\nThe main idea in this paper is to analyze the covering numbers directly (as opposed to calculating them by induction on the layers).\n ", "main_review": "The question investigated in the paper (described above) is natural and was posed as an open question by previous papers.\n\nThe main technique seems reasonable. \n\nThe experiments suggest that the change in the margin and the product of the weight norm may explain the larger generalization gap  (compared to the standard generalization).\n\nWeakness: showing a depth/width-dependent lower bound could have been nice.\n\n-Citing the paper https://arxiv.org/pdf/1810.02180.pdf (ALT19 and JMLR) is very relevant in the section \"Adversarial Generalization\".\nIt provides uniform convergence results In the case of a finite number of perturbations for a mixture of classifiers. The analysis goes through the Rademacher complexity as well.\n\n-In section 3: d^2 should be 2^d?", "summary_of_the_review": "I recommend accepting the paper.\n\nI read the overview of the main proof, but not the fully detailed proof.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636044357428}, {"id": "XBu2HuRJSj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2864/Reviewer_DjdD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors developed an upper bound of adversarial Rademacher complexity, which includes the product of weight norms. This implies that the large weight norm hinders from achieving good generalization performance in adversarial. The authors also empirically show that the product of weight norm in adversarial training is indeed much larger than that in standard training. ", "review_text": "Strengths: \n1.\tThis paper provides new bounds for adversarial Rademacher complexity. \n2.\tThis paper is well written and easy to read.\n\nWeaknesses: \n\nThe empirical validation in this paper seems insufficient:\n\n1. Although the authors verified that the product of weight norm in adversarial training is indeed much larger than that in standard training, I think this is not a strong empirical verification of the proposed bound. The reason is there existing some constants in the bounds in adversarial training and standard training, which may be different. Therefore, simply comparing the product of the weight norm is not rigorous. I would like to see more discussion on this. We know that some of the existing deep learning theory may be only mathematically correct, thus providing sufficient empirical evidence to show the consistence between the theory and practice is important. \n\n2. There exists a product of the weight norm in the proposed bound, which implies that the trained model can generalize better if this product is smaller. Therefore, can we use some techniques to regularize this product during training to improve the generalization ability. The authors are recommended to give such kind of experiments to support their theoretical results. \n\n====after rebuttal===\n\nAfter reading the response from the authors, I raised my rating. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors developed an upper bound of adversarial Rademacher complexity, which includes the product of weight norms. This implies that the large weight norm hinders from achieving good generalization performance in adversarial. The authors also empirically show that the product of weight norm in adversarial training is indeed much larger than that in standard training. ", "main_review": "Strengths: \n1.\tThis paper provides new bounds for adversarial Rademacher complexity. \n2.\tThis paper is well written and easy to read.\n\nWeaknesses: \n\nThe empirical validation in this paper seems insufficient:\n\n1. Although the authors verified that the product of weight norm in adversarial training is indeed much larger than that in standard training, I think this is not a strong empirical verification of the proposed bound. The reason is there existing some constants in the bounds in adversarial training and standard training, which may be different. Therefore, simply comparing the product of the weight norm is not rigorous. I would like to see more discussion on this. We know that some of the existing deep learning theory may be only mathematically correct, thus providing sufficient empirical evidence to show the consistence between the theory and practice is important. \n\n2. There exists a product of the weight norm in the proposed bound, which implies that the trained model can generalize better if this product is smaller. Therefore, can we use some techniques to regularize this product during training to improve the generalization ability. The authors are recommended to give such kind of experiments to support their theoretical results. \n\n====after rebuttal===\n\nAfter reading the response from the authors, I raised my rating. ", "summary_of_the_review": "1. This paper provides a new complexity bound for adversarial training. \n2. The empirical validation of this paper is insufficient. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635946254330}, {"id": "vyYb7rWGeag", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2864/Reviewer_dFnk"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes new generalization bounds for adversarial training in neural networks based on the Rademacher complexity. These are more general than previous results in that they apply to neural networks of any depth. Experiments are performed on CIFAR-10 and CIFAR-100 with multiple VGG architectures. These, related with the main quantities appearing in the proposed bounds, provide an explanation for the limited generalization capacity of adversarial training.", "review_text": "Strengths:\n- Relevant topic of theoretical investigation of adversarial examples and adversarial training, that has gained traction in recent years.\n- Improved theoretical result over prior art.\n- Meaningful experiments based on the theoretical results, that suggest some reasons why adversarial training does not generalize well.\n- Clear and well-written paper (a few typos remain).\n\nWeaknesses:\n- It would be great if the paper included a comparison to existing adversarial bounds based on Rademacher complexity or other frameworks, e.g., plotted for a toy example or small network. Lacking this, it is harder to judge the improvement the current paper makes over prior results in terms of tightness of the bound. Moreover, it might be worth comparing the contribution to other types of theoretical approaches in the field, e.g., the provable methods that the paper cites.\n\nQuestions and other comments:\n- It would be good to underline that the bounds provided also hold for convolutional neural networks earlier than the experiments section (or, more generally, what layers are covered).\n- What is the impact of using PGD adversarial examples in practice instead of the optimal perturbation?\n- Is it reasonable to consider a loss in the range 0-1 for neural networks?\n- Are Rademacher complexity bounds tight enough for neural networks to be informative or applicable in practice?\n\n[Update post discussion] I am raising my rating by one point, following the exchanges below.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes new generalization bounds for adversarial training in neural networks based on the Rademacher complexity. These are more general than previous results in that they apply to neural networks of any depth. Experiments are performed on CIFAR-10 and CIFAR-100 with multiple VGG architectures. These, related with the main quantities appearing in the proposed bounds, provide an explanation for the limited generalization capacity of adversarial training.", "main_review": "Strengths:\n- Relevant topic of theoretical investigation of adversarial examples and adversarial training, that has gained traction in recent years.\n- Improved theoretical result over prior art.\n- Meaningful experiments based on the theoretical results, that suggest some reasons why adversarial training does not generalize well.\n- Clear and well-written paper (a few typos remain).\n\nWeaknesses:\n- It would be great if the paper included a comparison to existing adversarial bounds based on Rademacher complexity or other frameworks, e.g., plotted for a toy example or small network. Lacking this, it is harder to judge the improvement the current paper makes over prior results in terms of tightness of the bound. Moreover, it might be worth comparing the contribution to other types of theoretical approaches in the field, e.g., the provable methods that the paper cites.\n\nQuestions and other comments:\n- It would be good to underline that the bounds provided also hold for convolutional neural networks earlier than the experiments section (or, more generally, what layers are covered).\n- What is the impact of using PGD adversarial examples in practice instead of the optimal perturbation?\n- Is it reasonable to consider a loss in the range 0-1 for neural networks?\n- Are Rademacher complexity bounds tight enough for neural networks to be informative or applicable in practice?\n\n[Update post discussion] I am raising my rating by one point, following the exchanges below.", "summary_of_the_review": "Good theoretical result supported by experimental evaluation on relevant topic.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635868342674}, {"id": "j84L6QQvyOp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2864/Reviewer_Vmod"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper overcomes some technical difficulties to provide adversarial Rademacher complexity of deep neural networks. Compared with existing literature which try to show other variants of adversarial Rademacher complexity, this paper directly works on adversarial Rademacher complexity itself. This paper provides both the lower bound and the upper bound. Besides, this paper conduct numerical experiments to combine with the theoretical bound to justify why adversarial training has a worse generalization than standard training.", "review_text": "I think the main contribution of this paper is interesting. It directly overcomes the difficulties in deriving adversarial Rademacher complexity rather than using other variants. It provides both the upper bound and lower bound, both of which are important. \n\nHowever, there are still a lot of things that can be improved in this paper. Below are my major concerns towards this paper (ordered in importance). My rating is currently weak accept given the importance of the upper and lower bounds, but it could be adjusted based on the author response towards my concerns. \n\n[1] My understanding towards the adversarial training and the adversarial Rademacher complexity is that, both adversarial-trained and standard-trained neural networks have their own standard Rademacher complexity and adversarial Rademacher complexity. So \"the generalization gap of adversarial training loss and adversarial testing loss\" is larger than \"the generalization gap of standard training loss and standard testing loss\" for both adversarial-trained (denote as A and B) and standard-trained neural networks (denote as C and D), i.e. A>B and C>D. On the other hand, through experiments, it is observed that the adversarially trained neural networks obtains larger norms, so its generalization is worse, i.e. A>D. Is my understanding correct? Is it essential to provide evidence for A>C or B>D in other to conclude A>D? In Figure 1, what does the \"generalization gap\" refer to?\n\nIn addition, could you provide some insights on why the norm of adversarially trained model is larger? Why does increasing the samples size lead to a larger norm/margin? Also, given these observations, is there any way to improve the generalization performance of adversarial training? Is there any implications on the loss landscape of adversarial robust neural networks? The current Section 6 displays some observations but no detailed insights.\n\n[1](as important as the above) The current proof of Theorem 3 only says \"By the results of the lower bounds of ..., we obtain that ...\". Please provide more details on the existing results and how to obtain the final result. Please provide some concrete illustrations to this either in pdf or in the discussion. \n\n[3] The literature review of generalization of adversarial training in this paper only considers those about Rademacher complexity. There are many other studies working on the generalization from other theoretical aspects. Below are some important articles. Please do some literature research in this general area of theoretical study and include in this paper.\n\n-----------------Papers which provide both upper bound and lower bound-----------------\n\nDan, Chen, Yuting Wei, and Pradeep Ravikumar. \"Sharp statistical guaratees for adversarially robust gaussian classification.\" International Conference on Machine Learning. PMLR, 2020.\n\nXing, Yue, Ruizhi Zhang, and Guang Cheng. \"Adversarially Robust Estimate and Risk Analysis in Linear Regression.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2021.\n\n-----------------Papers about generalization properties-----------------\n\nAllen-Zhu, Zeyuan, and Yuanzhi Li. \"Feature purification: How adversarial training performs robust deep learning.\" arXiv preprint arXiv:2005.10190 (2020).\n\nJavanmard, Adel, Mahdi Soltanolkotabi, and Hamed Hassani. \"Precise tradeoffs in adversarial training for linear regression.\" Conference on Learning Theory. PMLR, 2020.\n\nJavanmard, Adel, and Mahdi Soltanolkotabi. \"Precise statistical analysis of classification accuracies for adversarial training.\" arXiv preprint arXiv:2010.11213 (2020).\n\nTaheri, Hossein, Ramtin Pedarsani, and Christos Thrampoulidis. \"Asymptotic behavior of adversarial training in binary classification.\" arXiv preprint arXiv:2010.13275 (2020).\n\nWu, Dongxian, Shu-Tao Xia, and Yisen Wang. \"Adversarial Weight Perturbation Helps Robust Generalization.\" Advances in Neural Information Processing Systems 33 (2020).\n\nXing, Yue, Qifan Song, and Guang Cheng. \"On the Generalization Properties of Adversarial Training.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2021.\n\nZhai, Runtian, et al. \"Adversarially robust generalization just requires more unlabeled data.\" arXiv preprint arXiv:1906.00555 (2019).\n\n\n[4] The main idea of this paper is not hard to follow, and the authors make a lot of comparison to existing literature. It would be great if the authors could make it clear about the following questions when describing the proof steps in Theorem 1:\n\n   (1) Which steps are different from the derivation of standard Rademacher complexity? Which steps are not essential? If we use Theorem 1 and take \\epsilon=0, what steps should we modify to obtain the standard Rademacher complexity mentioned in Section 5.3?\n\n   (2) Which steps are different from the literature about adversarial Rademacher complexity? Which steps do they skip?\n\n   (3) Could you explain the remark after before Theorem 2 in detail?\n\n\n\nMinor issue (not ordered in importance):\n\n[5] When mentioning your contributions in the last paragraph of Section 1, could you write some descriptions? \n    (1) For the first contribution, is there any interesting findings in the bound?\n    (2) For the second contribution, could you answer your why question? \n\n[6] Please consider remove Proposition 1 and 2, and move Proposition 3 to the appendix. Proposition 1 and 2 do not help deepen the understanding of the main goal of this paper. Similarly, please shorten Section 3 for inequalities which are unrelated to the main goal. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper overcomes some technical difficulties to provide adversarial Rademacher complexity of deep neural networks. Compared with existing literature which try to show other variants of adversarial Rademacher complexity, this paper directly works on adversarial Rademacher complexity itself. This paper provides both the lower bound and the upper bound. Besides, this paper conduct numerical experiments to combine with the theoretical bound to justify why adversarial training has a worse generalization than standard training.", "main_review": "I think the main contribution of this paper is interesting. It directly overcomes the difficulties in deriving adversarial Rademacher complexity rather than using other variants. It provides both the upper bound and lower bound, both of which are important. \n\nHowever, there are still a lot of things that can be improved in this paper. Below are my major concerns towards this paper (ordered in importance). My rating is currently weak accept given the importance of the upper and lower bounds, but it could be adjusted based on the author response towards my concerns. \n\n[1] My understanding towards the adversarial training and the adversarial Rademacher complexity is that, both adversarial-trained and standard-trained neural networks have their own standard Rademacher complexity and adversarial Rademacher complexity. So \"the generalization gap of adversarial training loss and adversarial testing loss\" is larger than \"the generalization gap of standard training loss and standard testing loss\" for both adversarial-trained (denote as A and B) and standard-trained neural networks (denote as C and D), i.e. A>B and C>D. On the other hand, through experiments, it is observed that the adversarially trained neural networks obtains larger norms, so its generalization is worse, i.e. A>D. Is my understanding correct? Is it essential to provide evidence for A>C or B>D in other to conclude A>D? In Figure 1, what does the \"generalization gap\" refer to?\n\nIn addition, could you provide some insights on why the norm of adversarially trained model is larger? Why does increasing the samples size lead to a larger norm/margin? Also, given these observations, is there any way to improve the generalization performance of adversarial training? Is there any implications on the loss landscape of adversarial robust neural networks? The current Section 6 displays some observations but no detailed insights.\n\n[1](as important as the above) The current proof of Theorem 3 only says \"By the results of the lower bounds of ..., we obtain that ...\". Please provide more details on the existing results and how to obtain the final result. Please provide some concrete illustrations to this either in pdf or in the discussion. \n\n[3] The literature review of generalization of adversarial training in this paper only considers those about Rademacher complexity. There are many other studies working on the generalization from other theoretical aspects. Below are some important articles. Please do some literature research in this general area of theoretical study and include in this paper.\n\n-----------------Papers which provide both upper bound and lower bound-----------------\n\nDan, Chen, Yuting Wei, and Pradeep Ravikumar. \"Sharp statistical guaratees for adversarially robust gaussian classification.\" International Conference on Machine Learning. PMLR, 2020.\n\nXing, Yue, Ruizhi Zhang, and Guang Cheng. \"Adversarially Robust Estimate and Risk Analysis in Linear Regression.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2021.\n\n-----------------Papers about generalization properties-----------------\n\nAllen-Zhu, Zeyuan, and Yuanzhi Li. \"Feature purification: How adversarial training performs robust deep learning.\" arXiv preprint arXiv:2005.10190 (2020).\n\nJavanmard, Adel, Mahdi Soltanolkotabi, and Hamed Hassani. \"Precise tradeoffs in adversarial training for linear regression.\" Conference on Learning Theory. PMLR, 2020.\n\nJavanmard, Adel, and Mahdi Soltanolkotabi. \"Precise statistical analysis of classification accuracies for adversarial training.\" arXiv preprint arXiv:2010.11213 (2020).\n\nTaheri, Hossein, Ramtin Pedarsani, and Christos Thrampoulidis. \"Asymptotic behavior of adversarial training in binary classification.\" arXiv preprint arXiv:2010.13275 (2020).\n\nWu, Dongxian, Shu-Tao Xia, and Yisen Wang. \"Adversarial Weight Perturbation Helps Robust Generalization.\" Advances in Neural Information Processing Systems 33 (2020).\n\nXing, Yue, Qifan Song, and Guang Cheng. \"On the Generalization Properties of Adversarial Training.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2021.\n\nZhai, Runtian, et al. \"Adversarially robust generalization just requires more unlabeled data.\" arXiv preprint arXiv:1906.00555 (2019).\n\n\n[4] The main idea of this paper is not hard to follow, and the authors make a lot of comparison to existing literature. It would be great if the authors could make it clear about the following questions when describing the proof steps in Theorem 1:\n\n   (1) Which steps are different from the derivation of standard Rademacher complexity? Which steps are not essential? If we use Theorem 1 and take \\epsilon=0, what steps should we modify to obtain the standard Rademacher complexity mentioned in Section 5.3?\n\n   (2) Which steps are different from the literature about adversarial Rademacher complexity? Which steps do they skip?\n\n   (3) Could you explain the remark after before Theorem 2 in detail?\n\n\n\nMinor issue (not ordered in importance):\n\n[5] When mentioning your contributions in the last paragraph of Section 1, could you write some descriptions? \n    (1) For the first contribution, is there any interesting findings in the bound?\n    (2) For the second contribution, could you answer your why question? \n\n[6] Please consider remove Proposition 1 and 2, and move Proposition 3 to the appendix. Proposition 1 and 2 do not help deepen the understanding of the main goal of this paper. Similarly, please shorten Section 3 for inequalities which are unrelated to the main goal. \n\n", "summary_of_the_review": "This paper provides some important results about the adversarial Rademacher complexity so I vote for weak acceptance. But there are many issues towards the experiments, proofs, and the writing of this paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No concern", "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635084482395}], "openreview_url": "https://openreview.net/forum?id=wNsNT56zDkG", "arxiv_id": "2211.14966", "paper_pdf": "papers/wNsNT56zDkG.pdf", "paper_pdf_sha256": "95eefe9f85dce2cef27ad47a78b2098d075eff59d4b07b9b28cfee2ff7bb8d31", "paper_pdf_bytes": 851245, "paper_pdf_source": "openreview", "code_url": "https://github.com/JiancongXiao/Adversarial-Rademacher-Complexity", "code_repository": "JiancongXiao/Adversarial-Rademacher-Complexity", "code_commit": "adc891cdc79f13014cf994db3837a66cbb06392c", "code_archive": "repos/wNsNT56zDkG.zip", "code_archive_sha256": "e307414deb5544a4b266d775639e8b5a973544b38c5b14cc3d151e8982da4f10", "code_archive_bytes": 350410, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 336, "github_languages": {"Python": 55343}, "github_archived": false, "github_pushed_at": "2023-11-27T19:48:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-rademacher-complexity-of-deep-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gp5Uzbl-9C-", "year": 2021, "status": "rejected", "title": "Systematic Evaluation of Causal Discovery in Visual Model Based Reinforcement Learning", "authors": ["Nan Rosemary Ke", "Aniket Rajiv Didolkar", "Sarthak Mittal", "Anirudh Goyal", "Guillaume Lajoie", "Stefan Bauer", "Danilo Jimenez Rezende", "Michael Curtis Mozer", "Yoshua Bengio", "Christopher Pal"], "authorids": ["~Nan_Rosemary_Ke1", "~Aniket_Rajiv_Didolkar1", "~Sarthak_Mittal1", "~Anirudh_Goyal1", "~Guillaume_Lajoie1", "~Stefan_Bauer1", "~Danilo_Jimenez_Rezende2", "~Michael_Curtis_Mozer1", "~Yoshua_Bengio1", "~Christopher_Pal1"], "authors_source": "OpenReview API", "abstract": "Inducing causal relationships from observations is a classic problem in machine learning. Most work in causality starts from the premise that the causal variables themselves have known semantics or are observed. However, for AI agents such as robots trying to make sense of their environment, the only observables are low-level variables like pixels in images. To generalize well, an agent must induce high-level variables, particularly those which are causal or are affected by causal variables. A central goal for AI and causality is thus the joint discovery of abstract representations and causal structure. In this work, we systematically evaluate the agent's ability to learn underlying causal structure. We note that existing environments for studying causal induction are poorly suited for this  objective because they have complicated task-specific causal graphs with many confounding factors. Hence, to facilitate research in learning the representation of high-level variables as well as causal structure among these  variables, we present a suite of RL environments created to systematically probe the ability of methods to identify variables as well as causal structure among those variables. We evaluate various representation learning algorithms from literature and found that  explicitly incorporating structure and modularity in the model can help causal induction in model-based reinforcement learning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gFeXm2wjuy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2766/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "=== Summary\n\nThis paper proposes a benchmark that aims to systematically evaluate models' ability in learning representations of high-level variables as well as causal structures among them. The authors introduce two benchmarking RL environments:\n- One is in a physical domain where an agent is pushing blocks of different weights.\n- Another one is to simulate a chemistry environment, where the state of an element can cause changes to another variable's state according to the underlying causal graph.\n\nThe authors evaluate several representation learning algorithms from the literature and find that explicitly incorporating structure and modularity in models can help causal induction in model-based reinforcement learning.\n\n\n=== Strengths\n\nThis paper targets an important problem of assessing models' ability to automate the inference and identification of the causal variables from high-dimensional inputs like images.\n\nThe construction of the benchmark allows building causal graphs with varying complexity, such as the size of the graph, the sparsity of the graph, and the length of cause-effect chains.\n\nThe authors have evaluated several baseline models on the benchmark, including two typical monolithic models (autoencoders and variational autoencoders) and two models with explicit structure: graph neural networks (GNNs) and modular networks.\n\nThey have made several interesting observations, e.g., modular networks hold better scalability than other baselines, suggesting the benefits that explicit structure and modularity bring for causal induction in MBRL.\n\n\n=== Weaknesses\n\nMy primary concern of this paper is that the dataset is a bit too contrived, which makes it hard to know whether the observations from this benchmark can generalize to more complicated real-world scenarios.\n\nFor example, in the Physics Environment proposed in this paper, only heavier objects can push lighter ones, not the other way around. I understand the authors' desire to make the underlying graph a directed acyclic graph (DAG), but it does not reflect what will happen in the real world. One could imagine sliding a lighter object to collide with a heavier one; the motions of both objects are likely to change, where the interaction between them is bi-directional.\n\nAlso, in the Chemistry Environment, a few objects are connected by a randomly generated DAG, where interventions can change the color of the target and the subsequent blocks. In chemistry, a molecule is a group of atoms held together by chemical bonds, which are also bi-directional relationships. Are there any specific examples in chemistry where the graph is a DAG? I would appreciate it if the authors can elaborate on the connection between the design of the environment and \"chemistry.\"\n\nIn terms of the difficulty of the tasks, the results shown in Figure 6 suggest that the Greedy algorithm can achieve a near-perfect performance on the tasks, which consistently outperforms all other baselines. If a simple greedy algorithm can solve the tasks, does it mean that the benchmark may be a bit too simple, where a good understanding of the underlying causal structure may not be necessary? It would be better if the authors can discuss the necessity of causal induction in these tasks, and how is the ability to perform causal inference correlate with the metrics used in the benchmark.\n\nOverall, I feel the environments proposed in the paper are a bit too artificial, which does not reflect what's likely to happen in the real world. While I like the goal of this paper, I think a set of more realistic environments could greatly improve the significance and potential impact of this paper.\n\n\n\n=== Other comments\n\n\nThe font size of the image caption may be a bit too small.\n\nTypo: Section 2.1, \"Impotant\" --> \"Important\"\n\n\n=== Post rebuttal\n\nThe authors' rebuttal addressed some of my concerns, but my primary concern still remains that that benchmark may be a bit too contrived, where the observations made in this paper may not generalize to more complicated real-world situations. The authors also made some far-fetched arguments in the rebuttal by claiming some concurrent works [1, 2] as \"the 'real-world' version of the environments used in the paper,\" which, to be honest, further lowers the rating of the paper on my side: why is this paper worthy of acceptance if there exist more realistic benchmarks?\n\nI also agree with R1 and R3 that there are no new methods proposed in the paper, and the insights derived from benchmarking a set of existing methods may not be considered novel from the point of view of the ICLR audience. As a result, I keep my rating the same.\n\n[1] Physically Embedded Planning Problems: New Challenges for Reinforcement Learning, https://arxiv.org/abs/2009.05524\n\n[2] CausalWorld: A Robotic Manipulation Benchmark for Causal Structure and Transfer Learning, https://arxiv.org/abs/2010.04296", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper introduces an interesting benchmark with an ambitious goal, but the tasks may be a bit too contrived.", "review": "=== Summary\n\nThis paper proposes a benchmark that aims to systematically evaluate models' ability in learning representations of high-level variables as well as causal structures among them. The authors introduce two benchmarking RL environments:\n- One is in a physical domain where an agent is pushing blocks of different weights.\n- Another one is to simulate a chemistry environment, where the state of an element can cause changes to another variable's state according to the underlying causal graph.\n\nThe authors evaluate several representation learning algorithms from the literature and find that explicitly incorporating structure and modularity in models can help causal induction in model-based reinforcement learning.\n\n\n=== Strengths\n\nThis paper targets an important problem of assessing models' ability to automate the inference and identification of the causal variables from high-dimensional inputs like images.\n\nThe construction of the benchmark allows building causal graphs with varying complexity, such as the size of the graph, the sparsity of the graph, and the length of cause-effect chains.\n\nThe authors have evaluated several baseline models on the benchmark, including two typical monolithic models (autoencoders and variational autoencoders) and two models with explicit structure: graph neural networks (GNNs) and modular networks.\n\nThey have made several interesting observations, e.g., modular networks hold better scalability than other baselines, suggesting the benefits that explicit structure and modularity bring for causal induction in MBRL.\n\n\n=== Weaknesses\n\nMy primary concern of this paper is that the dataset is a bit too contrived, which makes it hard to know whether the observations from this benchmark can generalize to more complicated real-world scenarios.\n\nFor example, in the Physics Environment proposed in this paper, only heavier objects can push lighter ones, not the other way around. I understand the authors' desire to make the underlying graph a directed acyclic graph (DAG), but it does not reflect what will happen in the real world. One could imagine sliding a lighter object to collide with a heavier one; the motions of both objects are likely to change, where the interaction between them is bi-directional.\n\nAlso, in the Chemistry Environment, a few objects are connected by a randomly generated DAG, where interventions can change the color of the target and the subsequent blocks. In chemistry, a molecule is a group of atoms held together by chemical bonds, which are also bi-directional relationships. Are there any specific examples in chemistry where the graph is a DAG? I would appreciate it if the authors can elaborate on the connection between the design of the environment and \"chemistry.\"\n\nIn terms of the difficulty of the tasks, the results shown in Figure 6 suggest that the Greedy algorithm can achieve a near-perfect performance on the tasks, which consistently outperforms all other baselines. If a simple greedy algorithm can solve the tasks, does it mean that the benchmark may be a bit too simple, where a good understanding of the underlying causal structure may not be necessary? It would be better if the authors can discuss the necessity of causal induction in these tasks, and how is the ability to perform causal inference correlate with the metrics used in the benchmark.\n\nOverall, I feel the environments proposed in the paper are a bit too artificial, which does not reflect what's likely to happen in the real world. While I like the goal of this paper, I think a set of more realistic environments could greatly improve the significance and potential impact of this paper.\n\n\n\n=== Other comments\n\n\nThe font size of the image caption may be a bit too small.\n\nTypo: Section 2.1, \"Impotant\" --> \"Important\"\n\n\n=== Post rebuttal\n\nThe authors' rebuttal addressed some of my concerns, but my primary concern still remains that that benchmark may be a bit too contrived, where the observations made in this paper may not generalize to more complicated real-world situations. The authors also made some far-fetched arguments in the rebuttal by claiming some concurrent works [1, 2] as \"the 'real-world' version of the environments used in the paper,\" which, to be honest, further lowers the rating of the paper on my side: why is this paper worthy of acceptance if there exist more realistic benchmarks?\n\nI also agree with R1 and R3 that there are no new methods proposed in the paper, and the insights derived from benchmarking a set of existing methods may not be considered novel from the point of view of the ICLR audience. As a result, I keep my rating the same.\n\n[1] Physically Embedded Planning Problems: New Challenges for Reinforcement Learning, https://arxiv.org/abs/2009.05524\n\n[2] CausalWorld: A Robotic Manipulation Benchmark for Causal Structure and Transfer Learning, https://arxiv.org/abs/2010.04296", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604810011283}, {"id": "tpAjAjAxgVv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2766/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper investigates the importance of incorporating structure and modularity in MBRL. Specifically, it compares monolithic models with models with explicit structure: GNNs and modular models. It also investigates the influences of varying complexity of graphs. \n\nPros:\n1. The authors state a general goal in the abstract: \"the joint discovery of abstract representations and causal structure\", which is promising and needs more investigation for sure.\n\nCons:\n1. The writing is a bit confusing. From the title and the abstract, the readers may expect that this paper is about causal discovery, i.e., learning casual relationships from data, while the casual structure actually is given in the main text.  \n2. In addition, this paper only compares existing models with or without structures, but I fail to see any novel or interesting ideas that this paper tries to deliver. \n3. Furthermore, from the experiments, the authors found that increasing the size of the graph impacts the performance significantly, e.g., increasing 3 nodes to 5 nodes. This result surprises me. I think if the structure is handled properly, the increase from 3 nodes to 5 nodes should not have such a big effect. The authors may investigate or propose other methods that make proper use of the structures. \n\nBased on the above reasons, I do not think this paper is ready to publish.\n\n\nPost-rebuttal:\nThanks for the feedback. The goal of learning causal representation is ambiguous, and it is absolutely a good research topic. However, I fail to see an obvious contribution of the current version. Researchers in this field are usually clearly aware of the limitations for existing methods in causal learning. The problem is how to handle it, e.g., how to give a appropriate definition of the causal variables, how to theoretically show the identifiability and consistency, and then propose a practical solution. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Right direction; writing is confusing; just comparing existing methods, but no new method", "review": "This paper investigates the importance of incorporating structure and modularity in MBRL. Specifically, it compares monolithic models with models with explicit structure: GNNs and modular models. It also investigates the influences of varying complexity of graphs. \n\nPros:\n1. The authors state a general goal in the abstract: \"the joint discovery of abstract representations and causal structure\", which is promising and needs more investigation for sure.\n\nCons:\n1. The writing is a bit confusing. From the title and the abstract, the readers may expect that this paper is about causal discovery, i.e., learning casual relationships from data, while the casual structure actually is given in the main text.  \n2. In addition, this paper only compares existing models with or without structures, but I fail to see any novel or interesting ideas that this paper tries to deliver. \n3. Furthermore, from the experiments, the authors found that increasing the size of the graph impacts the performance significantly, e.g., increasing 3 nodes to 5 nodes. This result surprises me. I think if the structure is handled properly, the increase from 3 nodes to 5 nodes should not have such a big effect. The authors may investigate or propose other methods that make proper use of the structures. \n\nBased on the above reasons, I do not think this paper is ready to publish.\n\n\nPost-rebuttal:\nThanks for the feedback. The goal of learning causal representation is ambiguous, and it is absolutely a good research topic. However, I fail to see an obvious contribution of the current version. Researchers in this field are usually clearly aware of the limitations for existing methods in causal learning. The problem is how to handle it, e.g., how to give a appropriate definition of the causal variables, how to theoretically show the identifiability and consistency, and then propose a practical solution. \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603999200119}, {"id": "EOeFsKSF70T", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2766/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is a review of model-based approaches of integrating causal inference to reinforcement learning (RL) in different environments (application areas). The authors provide software to analyse how three types of models (“monolithic”, i.e. latent space models without a graph-like structure of the latent space, graph neural networks (GNN) and “modular”, i.e. the C-SWM model (Kipf et al., 2020)) perform in two artificial “environments” devised by the authors (physics and chemistry) based on a number of metrics, some of them also proposed by the authors. The main contributions are the platform for evaluating models in the environments and the insights from the experiments performed on the selected models (taken from existing literature).\n\n\n*****Strengths:*****\n\nThe paper is very well-written, clear and easy to follow.\n\nThe paper provides a novel perspective on applications of causality modelling and the presented environments could be useful to practitioners in the respective fields of physics and chemistry.\n\n\n*****Weaknesses:*****\n\nAs the paper has the form of a review and a collection of insights, there seems to be no novelty from the point of view of either machine learning, causal inference or reinforcement learning. I am therefore not sure whether ICLR is the right venue for this paper.\n\nWhile the motivation for the causal problem tackled in the paper (inferring model-driven causal insights usable for RL agents from data available to agents/robots such as images) is well-defined, interesting and relevant, there seems to be little actual follow-up on it in the paper. The presented models do not seem to adhere to any causal formalism (Pearl’s graphs, Rubin’s potential outcomes) and their only “causal” interpretation is a naïve use of structures (graphs and DAGs) as descriptors of causal relationships without any formal justification. There appears to be no relation to causal structure learning. Neither GNN nor C-SWM were conceived as causal models, and if one wishes to use them as causal models, more justification for such use is needed (why should the discovered graphs correspond to some underlying causal relationships?). Moreover, there have been approaches to achieve what this paper seems to promise (discover causal relationships from images), e.g. (Discovering causal signals in images, Lopez-Paz et al., 2017).\n\n\n*****Post Rebuttal*****\nI would like to thank the authors for the detailed rebuttal.\n\nThe authors state: \"The main goal of our paper is NOT to introduce novel models, but rather to introduce a NOVEL benchmark and insights/ingredients to study causal induction in model-based RL\" and \"It is true that the models we use do not learn an explicit structure for causal learning\" which corresponds to my original reservations to the novelty of this paper. The authors introduce a benchmark / software for evaluating causal induction in RL models, where the user can specify a causal model and its influence on the environment can be examined. I remain unconvinced that the introduction of a RL evaluation benchmark (even one allowing for the presence of arbitrary causal networks) counts as novel at ICLR.\n\nFurther, the statement \"we used some of the typically common models for this purpose, such as GNNs and modular networks. Though these models do not learn an explicit causal graph, they do learn structure that could allow them to discover causal relationships under certain assumptions\" confirms my point that no causal formalism (e.g. connection to the data generating process) is accounted for. It merely means that directed relationships can be modelled.\n\nI agree with the authors that (Lopez-Paz et al., 2017) only uses observational data and does not have any connection to RL, but it is an example of an approach to extracting causal relationships from images in a sound way when it comes to causal inference. This is a side comment and does not influence my assessment of the paper.\n\nTo sum up, my reservations towards the degree of novelty in this paper have not changed (I agree with the authors' summary of their contributions, but disagree as to whether proposing a new benchmark constitutes novelty at ICLR) and I recommend a rejection of the paper in its current form.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Software for reviewing performance of integrating causality to RL in a number of application-driven environments.", "review": "This paper is a review of model-based approaches of integrating causal inference to reinforcement learning (RL) in different environments (application areas). The authors provide software to analyse how three types of models (“monolithic”, i.e. latent space models without a graph-like structure of the latent space, graph neural networks (GNN) and “modular”, i.e. the C-SWM model (Kipf et al., 2020)) perform in two artificial “environments” devised by the authors (physics and chemistry) based on a number of metrics, some of them also proposed by the authors. The main contributions are the platform for evaluating models in the environments and the insights from the experiments performed on the selected models (taken from existing literature).\n\n\n*****Strengths:*****\n\nThe paper is very well-written, clear and easy to follow.\n\nThe paper provides a novel perspective on applications of causality modelling and the presented environments could be useful to practitioners in the respective fields of physics and chemistry.\n\n\n*****Weaknesses:*****\n\nAs the paper has the form of a review and a collection of insights, there seems to be no novelty from the point of view of either machine learning, causal inference or reinforcement learning. I am therefore not sure whether ICLR is the right venue for this paper.\n\nWhile the motivation for the causal problem tackled in the paper (inferring model-driven causal insights usable for RL agents from data available to agents/robots such as images) is well-defined, interesting and relevant, there seems to be little actual follow-up on it in the paper. The presented models do not seem to adhere to any causal formalism (Pearl’s graphs, Rubin’s potential outcomes) and their only “causal” interpretation is a naïve use of structures (graphs and DAGs) as descriptors of causal relationships without any formal justification. There appears to be no relation to causal structure learning. Neither GNN nor C-SWM were conceived as causal models, and if one wishes to use them as causal models, more justification for such use is needed (why should the discovered graphs correspond to some underlying causal relationships?). Moreover, there have been approaches to achieve what this paper seems to promise (discover causal relationships from images), e.g. (Discovering causal signals in images, Lopez-Paz et al., 2017).\n\n\n*****Post Rebuttal*****\nI would like to thank the authors for the detailed rebuttal.\n\nThe authors state: \"The main goal of our paper is NOT to introduce novel models, but rather to introduce a NOVEL benchmark and insights/ingredients to study causal induction in model-based RL\" and \"It is true that the models we use do not learn an explicit structure for causal learning\" which corresponds to my original reservations to the novelty of this paper. The authors introduce a benchmark / software for evaluating causal induction in RL models, where the user can specify a causal model and its influence on the environment can be examined. I remain unconvinced that the introduction of a RL evaluation benchmark (even one allowing for the presence of arbitrary causal networks) counts as novel at ICLR.\n\nFurther, the statement \"we used some of the typically common models for this purpose, such as GNNs and modular networks. Though these models do not learn an explicit causal graph, they do learn structure that could allow them to discover causal relationships under certain assumptions\" confirms my point that no causal formalism (e.g. connection to the data generating process) is accounted for. It merely means that directed relationships can be modelled.\n\nI agree with the authors that (Lopez-Paz et al., 2017) only uses observational data and does not have any connection to RL, but it is an example of an approach to extracting causal relationships from images in a sound way when it comes to causal inference. This is a side comment and does not influence my assessment of the paper.\n\nTo sum up, my reservations towards the degree of novelty in this paper have not changed (I agree with the authors' summary of their contributions, but disagree as to whether proposing a new benchmark constitutes novelty at ICLR) and I recommend a rejection of the paper in its current form.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603954379861}, {"id": "STIcPRr_Q0s", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2766/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper proposes a suite of RL benchmarks to facilitate inducing causal relationships from visual observations. They show that structural inductive biases are beneficial for causal relationship learning and model-based RL by testing a variety of representation learning algorithms on this benchmark.\n\n##########################################################################\n\npros: \n\n+ The motivation is clear and interesting. Inducing causal relationships from pixel-level observations is an important topic to integrate causality into RL.\n\n+ Overall, the paper is well written.\n\n+ A comprehensive evaluation is conducted to highlight the usage of the proposed benchmark.\n\n##########################################################################\n \ncons: \n\n- Would you explain more about the physics environment about why the underlying graph is a causal graph, not just a statistical graphical model?\n\n- Why is PHYRE[1] benchmark not mentioned in the related work? You seem to work in the same direction and they have proposed more complicated and interesting tasks.\n\n[1] Bakhtin, A., van der Maaten, L., Johnson, J., Gustafson, L., & Girshick, R. (2019). Phyre: A new benchmark for physical reasoning. In Advances in Neural Information Processing Systems (pp. 5082-5093).\n\n\n##########################################################################\n\nPost rebuttal\n\nI'm happy with the author's response and would like to keep my original score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Another benchmark to facilitate causal induction for RL with pixel-level observations.", "review": "Summary:\n\nThis paper proposes a suite of RL benchmarks to facilitate inducing causal relationships from visual observations. They show that structural inductive biases are beneficial for causal relationship learning and model-based RL by testing a variety of representation learning algorithms on this benchmark.\n\n##########################################################################\n\npros: \n\n+ The motivation is clear and interesting. Inducing causal relationships from pixel-level observations is an important topic to integrate causality into RL.\n\n+ Overall, the paper is well written.\n\n+ A comprehensive evaluation is conducted to highlight the usage of the proposed benchmark.\n\n##########################################################################\n \ncons: \n\n- Would you explain more about the physics environment about why the underlying graph is a causal graph, not just a statistical graphical model?\n\n- Why is PHYRE[1] benchmark not mentioned in the related work? You seem to work in the same direction and they have proposed more complicated and interesting tasks.\n\n[1] Bakhtin, A., van der Maaten, L., Johnson, J., Gustafson, L., & Girshick, R. (2019). Phyre: A new benchmark for physical reasoning. In Advances in Neural Information Processing Systems (pp. 5082-5093).\n\n\n##########################################################################\n\nPost rebuttal\n\nI'm happy with the author's response and would like to keep my original score.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603888429822}], "openreview_url": "https://openreview.net/forum?id=gp5Uzbl-9C-", "arxiv_id": "2107.00848", "paper_pdf": "papers/gp5Uzbl-9C-.pdf", "paper_pdf_sha256": "51d03eb06c9c227a225cb4f475d548f0e9644ef8e08612ed469e9c2959742564", "paper_pdf_bytes": 285872, "paper_pdf_source": "openreview", "code_url": "https://github.com/dido1998/CausalMBRL", "code_repository": "dido1998/CausalMBRL", "code_commit": "2c595988b06b3d568fdde53213029e97841acb03", "code_archive": "repos/gp5Uzbl-9C-.zip", "code_archive_sha256": "9b0b6d9b9fa79474cc28219afd8560d48dd7c4d3c031f9b53205ee70fc5a8713", "code_archive_bytes": 160266, "code_file_count": 85, "code_extensions": {".py": 53, ".sh": 32}, "github_disk_usage_kb": 367, "github_languages": {"Python": 527951, "Shell": 58570}, "github_archived": false, "github_pushed_at": "2021-07-15T17:13:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/systematic-evaluation-of-causal-discovery-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1g6MCEtwr", "year": 2020, "status": "rejected", "title": "Zero-Shot Out-of-Distribution Detection with Feature Correlations", "authors": ["Chandramouli S Sastry", "Sageev Oore"], "authorids": ["chandramouli.sastry@gmail.com", "osageev@gmail.com"], "authors_source": "OpenReview API", "abstract": "When presented with Out-of-Distribution (OOD) examples, deep neural networks yield confident, incorrect predictions. Detecting OOD examples is challenging, and the potential risks are high. In this paper, we propose to detect OOD examples by identifying inconsistencies between activity patterns and class predicted. We find that characterizing activity patterns by feature correlations and identifying anomalies in pairwise feature correlation values can yield high OOD detection rates. We identify anomalies in the pairwise feature correlations by simply comparing each pairwise correlation value with its respective range observed over the training data. Unlike many approaches, this can be used with any pre-trained softmax classifier and does not require access to OOD data for fine-tuning hyperparameters, nor does it require OOD access for inferring parameters. The method is applicable across a variety of architectures and vision datasets and generally performs better than or equal to state-of-the-art OOD detection methods, including those that do assume access to OOD examples.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJe62ec0tH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1017/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new scoring function for OOD detection based on calculating the total deviation of the pairwise feature correlations. The method only requires in-distribution data for tuning its hyper-parameters and can use a pre-trained classifier directly. Its performance is evaluated with small image datasets, which are commonly used in this line of work. Two additional experiments are performed to analyze the effect of two factors (the layer of the feature and the order of Gram matrix) in the method.\n\nThe overall vote for this paper is a weak reject. The clarity of the paper is good. The primary strength of the work is the very strong performance given the setting, which does not require OOD data for tuning. However, we give a rating slightly below the borderline, because of lack of in-depth explanation of why pairwise feature correlation is helpful for the problem. The explanation could be either theoretical or empirical, while the latter can be a set of carefully designed ablation studies. We do not see sufficient arguments or experiments to ensure the performance gain comes from the pairwise feature and is not from other parts of the design.\n\nThe above weakness could be addressed by answering the following questions:\n1. If the pairwise feature (G) is replaced by the unary feature (F) while keeping the other parts of the pipeline the same (including the use of p-th power), would it reach a similar performance or much worse?\n2. Why does the implementation use the statistics of min/max values instead of mean and variance? The latter can also calculate the deviation for the method using a Gaussian model, which is usually a more natural choice.\n3. What is the motivation for using p-th power, and why does it help?\n\nWe present the additional feedbacks in the form of questions which could further strengthen the work if answered:\n4. Based on Figure 2, having p=10 is sufficient to give a good result. Why does the method still use all p-th power (p=1..10)?\n5. The study in Figure 1 only presents specific OOD data (Tiny Imagenet). Does the trend still hold with other OOD data, such as SVHN or LSUN?\n6. Why are the networks used in figure 1(a) and (b) different? Does the trend still hold when the networks in both cases are changed?\n7. The statistics (Mins/Maxs) to be saved can be very large. The paper provides a strategy to reduce that by using the row-wise sums of G. However, the summation has an effect of mixing the features, causing a weaker signal from the deviation. Why does the method still works with the row-wise sums?\n8. Would the method work if the networks have no batch normalization layer?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This paper proposes a new scoring function for OOD detection based on calculating the total deviation of the pairwise feature correlations. The method only requires in-distribution data for tuning its hyper-parameters and can use a pre-trained classifier directly. Its performance is evaluated with small image datasets, which are commonly used in this line of work. Two additional experiments are performed to analyze the effect of two factors (the layer of the feature and the order of Gram matrix) in the method.\n\nThe overall vote for this paper is a weak reject. The clarity of the paper is good. The primary strength of the work is the very strong performance given the setting, which does not require OOD data for tuning. However, we give a rating slightly below the borderline, because of lack of in-depth explanation of why pairwise feature correlation is helpful for the problem. The explanation could be either theoretical or empirical, while the latter can be a set of carefully designed ablation studies. We do not see sufficient arguments or experiments to ensure the performance gain comes from the pairwise feature and is not from other parts of the design.\n\nThe above weakness could be addressed by answering the following questions:\n1. If the pairwise feature (G) is replaced by the unary feature (F) while keeping the other parts of the pipeline the same (including the use of p-th power), would it reach a similar performance or much worse?\n2. Why does the implementation use the statistics of min/max values instead of mean and variance? The latter can also calculate the deviation for the method using a Gaussian model, which is usually a more natural choice.\n3. What is the motivation for using p-th power, and why does it help?\n\nWe present the additional feedbacks in the form of questions which could further strengthen the work if answered:\n4. Based on Figure 2, having p=10 is sufficient to give a good result. Why does the method still use all p-th power (p=1..10)?\n5. The study in Figure 1 only presents specific OOD data (Tiny Imagenet). Does the trend still hold with other OOD data, such as SVHN or LSUN?\n6. Why are the networks used in figure 1(a) and (b) different? Does the trend still hold when the networks in both cases are changed?\n7. The statistics (Mins/Maxs) to be saved can be very large. The paper provides a strategy to reduce that by using the row-wise sums of G. However, the summation has an effect of mixing the features, causing a weaker signal from the deviation. Why does the method still works with the row-wise sums?\n8. Would the method work if the networks have no batch normalization layer?\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571885237040}, {"id": "Bkxza1qaKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1017/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a strategy for detecting out-of-distribution samples based on feature representations obtained via neural networks. Given a test sample, the proposed strategy checks whether the correlation values between the features of the test sample obtained at different channels of the same layer are coherent with those of the training samples known to belong to the estimated class of the test sample. \n\nThe proposed strategy can be applied to pretrained networks, as it only requires the channel activation values to determine whether a sample is out of distribution or not. The studied problem is an important problem and the experimental results show that the proposed strategy leads to some performance gains in comparison to reference methods. However, in my view the main drawback of the study is that it is based on an ad-hoc methodology whose theoretical foundation is not quite clear. In particular, it would be good to provide some more explanations on the following issues:\n\n- What exactly motivates the assumption that the correlation values between different channels of the same layer provides a discriminative characteristics of the classes from each other? It would be natural to assume that different classes will have different activation levels at a certain channel of a certain layer. However, the idea of the paper is to look at how different channels correlate with each other. I cannot entirely grasp the motivation for this, as looking at the feature correlations is a bit more indirect than looking at the features themselves. It would be good to provide the justification of this choice.\n\n- What is the theoretical motivation behind using the p-th order Gram matrix, instead of using the original Gram matrix (e.g. p=1)? Some experimental justification is given, but it is also important to provide some theoretical insight if possible.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes a strategy for detecting out-of-distribution samples based on feature representations obtained via neural networks. Given a test sample, the proposed strategy checks whether the correlation values between the features of the test sample obtained at different channels of the same layer are coherent with those of the training samples known to belong to the estimated class of the test sample. \n\nThe proposed strategy can be applied to pretrained networks, as it only requires the channel activation values to determine whether a sample is out of distribution or not. The studied problem is an important problem and the experimental results show that the proposed strategy leads to some performance gains in comparison to reference methods. However, in my view the main drawback of the study is that it is based on an ad-hoc methodology whose theoretical foundation is not quite clear. In particular, it would be good to provide some more explanations on the following issues:\n\n- What exactly motivates the assumption that the correlation values between different channels of the same layer provides a discriminative characteristics of the classes from each other? It would be natural to assume that different classes will have different activation levels at a certain channel of a certain layer. However, the idea of the paper is to look at how different channels correlate with each other. I cannot entirely grasp the motivation for this, as looking at the feature correlations is a bit more indirect than looking at the features themselves. It would be good to provide the justification of this choice.\n\n- What is the theoretical motivation behind using the p-th order Gram matrix, instead of using the original Gram matrix (e.g. p=1)? Some experimental justification is given, but it is also important to provide some theoretical insight if possible."}, "tcdate": 1571819450290}, {"id": "H1xozffaFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1017/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper uses Gram matrices for OOD detection. This enables the reliable detection of far-from-distribution examples, which is a long-standing and surprisingly difficult problem in OOD detection.\nThis is the best paper in my batch due to the strength of the results. However, this paper should more accurately reflect the contribution: this helps with far-from-distribution examples, not near-distribution yet OOD examples.\nFor instance, I used their code and found that their technique leads to an AUROC of 79.01% when using CIFAR-10 as the in-distribution and CIFAR-100 as the OOD set. Likewise, with their code I found having CIFAR-100 as in and CIFAR-10 as out gives an AUROC of 67.95%. This is much worse than currently existing techniques. Hence the paper should re-frame or qualify their results as helping with far-from-distribution detection or the detection of obvious anomalies. This paper makes a solid stride in improving the detection of garbage inputs, but the paper should modify its message so as not to suggest this helps with all currently considered OOD detection tasks.\n\nSmall comments:\n\n> Lee et al. (2018b)—to the best of our knowledge, the current SOTA technique by a significant margin\nThis should be again qualified and expanded. If you assume access to knowledge of the test distribution, then Mahalanobis is easily the best. If not, the Outlier Exposure is best. If you assume access to no extra data during training, then the maximum softmax probability + rotation prediction is best [1].\n\n> However, while the OE method is able to generalize across different non-training distributions, it does not achieve the SOTA rates of Lee et al. (2018b) on most cases.\nThere are different senses of state-of-the-art and these should be qualified.\n\nDoes this technique do better if you do 5th and 9th percentile instead of min and max? Is it important to do the min and max with training examples instead of validation examples? (Not a pressing question.)\n\n> Can work without access to OOD validation examples?\nTable 1 is deceptive. OE does not need the \"validation\" examples.\nI suggest two columns instead of one. \"Can this work without knowledge of OOD test examples? Does this use OOD training examples?\"\n\nShow AUROC and AUPR. Detection accuracy is an unusual metric relative to AUPR (OOD as positive).\n\nShow CIFAR-10 vs CIFAR-100 in the tables or I'll downgrade my rating, since otherwise the paper is not leaving an accurate impression.\n\nSince OE is complementary, perhaps this technique can be combined to tackle these near-distribution cases?\n\nThe title is confusing. \"Zero-Shot\" could be applied to various techniques in this space. Perhaps emphasize Gram matrices?\n\nIn their code:\n        validation_indices = random.sample(range(len(all_test_deviations)),int(0.1*len(all_test_deviations)))\n        test_indices = sorted(list(set(range(len(all_test_deviations)))-set(validation_indices)))\nThese indices change with every power, which is unrealistic. Please fix the sets beforehand.\n\nSince neural style transfer, which uses Gram matrices, works much better with VGG architectures than ResNets, does this technique work better with VGG architectures?\n\n[1] Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty. Hendrycks et al. NeurIPS 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "8: Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "This paper uses Gram matrices for OOD detection. This enables the reliable detection of far-from-distribution examples, which is a long-standing and surprisingly difficult problem in OOD detection.\nThis is the best paper in my batch due to the strength of the results. However, this paper should more accurately reflect the contribution: this helps with far-from-distribution examples, not near-distribution yet OOD examples.\nFor instance, I used their code and found that their technique leads to an AUROC of 79.01% when using CIFAR-10 as the in-distribution and CIFAR-100 as the OOD set. Likewise, with their code I found having CIFAR-100 as in and CIFAR-10 as out gives an AUROC of 67.95%. This is much worse than currently existing techniques. Hence the paper should re-frame or qualify their results as helping with far-from-distribution detection or the detection of obvious anomalies. This paper makes a solid stride in improving the detection of garbage inputs, but the paper should modify its message so as not to suggest this helps with all currently considered OOD detection tasks.\n\nSmall comments:\n\n> Lee et al. (2018b)—to the best of our knowledge, the current SOTA technique by a significant margin\nThis should be again qualified and expanded. If you assume access to knowledge of the test distribution, then Mahalanobis is easily the best. If not, the Outlier Exposure is best. If you assume access to no extra data during training, then the maximum softmax probability + rotation prediction is best [1].\n\n> However, while the OE method is able to generalize across different non-training distributions, it does not achieve the SOTA rates of Lee et al. (2018b) on most cases.\nThere are different senses of state-of-the-art and these should be qualified.\n\nDoes this technique do better if you do 5th and 9th percentile instead of min and max? Is it important to do the min and max with training examples instead of validation examples? (Not a pressing question.)\n\n> Can work without access to OOD validation examples?\nTable 1 is deceptive. OE does not need the \"validation\" examples.\nI suggest two columns instead of one. \"Can this work without knowledge of OOD test examples? Does this use OOD training examples?\"\n\nShow AUROC and AUPR. Detection accuracy is an unusual metric relative to AUPR (OOD as positive).\n\nShow CIFAR-10 vs CIFAR-100 in the tables or I'll downgrade my rating, since otherwise the paper is not leaving an accurate impression.\n\nSince OE is complementary, perhaps this technique can be combined to tackle these near-distribution cases?\n\nThe title is confusing. \"Zero-Shot\" could be applied to various techniques in this space. Perhaps emphasize Gram matrices?\n\nIn their code:\n        validation_indices = random.sample(range(len(all_test_deviations)),int(0.1*len(all_test_deviations)))\n        test_indices = sorted(list(set(range(len(all_test_deviations)))-set(validation_indices)))\nThese indices change with every power, which is unrealistic. Please fix the sets beforehand.\n\nSince neural style transfer, which uses Gram matrices, works much better with VGG architectures than ResNets, does this technique work better with VGG architectures?\n\n[1] Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty. Hendrycks et al. NeurIPS 2019.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571787283089}], "openreview_url": "https://openreview.net/forum?id=r1g6MCEtwr", "arxiv_id": null, "paper_pdf": "papers/r1g6MCEtwr.pdf", "paper_pdf_sha256": "64dbf01f5c4eb61ff9cb704e81da311f35b296e2ea4d588c85dcc8846d8430b1", "paper_pdf_bytes": 625825, "paper_pdf_source": "openreview", "code_url": "https://github.com/zeroshot-ood/ood-detection", "code_repository": "zeroshot-ood/ood-detection", "code_commit": "a7c08eae66b72109daf5b7f3e46754f5c583b64a", "code_archive": "repos/r1g6MCEtwr.zip", "code_archive_sha256": "42fd17cabc42fc84f758f0431dbad6b97bdda892eccb2d082c1409718305206e", "code_archive_bytes": 593891, "code_file_count": 8, "code_extensions": {".ipynb": 6, ".py": 2}, "github_disk_usage_kb": 585, "github_languages": {"Jupyter Notebook": 146932, "Python": 13512}, "github_archived": false, "github_pushed_at": "2020-01-19T23:38:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/zero-shot-out-of-distribution-detection-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "z1yPH2Rska", "year": 2026, "status": "rejected", "title": "Diagonal Batching Unlocks Parallelism in Recurrent Memory Transformers for Long Contexts", "authors": ["Danil Sivtsov", "Ivan Rodkin", "Gleb Kuzmin", "Yuri Kuratov", "Ivan Oseledets"], "authorids": ["~Danil_Sivtsov1", "~Ivan_Rodkin2", "~Gleb_Kuzmin1", "~Yuri_Kuratov2", "~Ivan_Oseledets1"], "authors_source": "OpenReview API", "abstract": "Long-context inference with Transformers is constrained by quadratic attention\nand linear memory growth. Many linear-time alternatives require pretraining from\nscratch, whereas Recurrent Memory Transformers (RMTs) convert pretrained \nmodels into segment-recurrent variants via finetuning without modifying the original\nmodel architecture. However, their sequential memory updates underutilize GPUs.\nWe show that RMT-style architectures with layer-level memory (PRMTs) \n(e.g., ARMT) can be among the most latency-efficient linear approaches when scheduled\nproperly. We introduce Diagonal Batching, a compute-reordering scheme that\npreserves exact recurrence while exposing inter-step parallelism by executing \n\"diagonals\" concurrently with grouped layers. On LLaMA (1B/3B/8B) up to 131,072\ntokens on A100/H100, Diagonal Batching achieves up to 3.3× lower latency than\nfull-attention inference and 1.8× over a sequential ARMT baseline, with no \ncustom CUDA kernels. With the right scheduling, PRMTs achieve linear scaling with\ncontext length and stand out as competitive, scalable architectures among linear\nrecurrent models.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "bHMQo0NI5S", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24761/Reviewer_PbQn"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 2, "summary": "[Disclaimer that I worked on this field in 2023 and 2024 and have not kept up to date with the latest trends in 2025, so please take this review with a grain of salt.]\n\nThe paper introduces a new scheduling scheme for recurrent transformers, which the authors call \"Diagonal Batching\". The framing is as follows: RMTs have linear inference but are sequential and hence not easily parallelizable. Parallel RMTs localize recurrence within layers and eliminate all inter-layer memory flow. The issue is that PRMTs underutilize GPUs for single, long input requests. The paper introduces diagonal batching as a way to utilizes the GPUs better and hence improve latency. This is done by rearranging the layers and segments into a diagonal structure, where parallel operations can occur at the same time but across different layers.\n\nThe authors show that this provides better latency, especially in long token settings, though the results are less pronounced in the short token settings. They show that results hold across scales. The authors also provide an error analysis which shows that the error accumulation of the method is minimal.", "review_text": "[Disclaimer that I worked on this field in 2023 and 2024 and have not kept up to date with the latest trends in 2025, so please take this review with a grain of salt.]\n\nThe paper introduces a new scheduling scheme for recurrent transformers, which the authors call \"Diagonal Batching\". The framing is as follows: RMTs have linear inference but are sequential and hence not easily parallelizable. Parallel RMTs localize recurrence within layers and eliminate all inter-layer memory flow. The issue is that PRMTs underutilize GPUs for single, long input requests. The paper introduces diagonal batching as a way to utilizes the GPUs better and hence improve latency. This is done by rearranging the layers and segments into a diagonal structure, where parallel operations can occur at the same time but across different layers.\n\nThe authors show that this provides better latency, especially in long token settings, though the results are less pronounced in the short token settings. They show that results hold across scales. The authors also provide an error analysis which shows that the error accumulation of the method is minimal.", "strengths": "- good exposition of background information and framing of the contributions of the paper\n- proposed method seems simple and can be applied quite generally across different models. If indeed true, the method provides a lot of latency gains basically for free.\n- method section seems quite complete. The authors go through both the high-level motivations but also outline the implementation details.\n- experiments section seems complete as well. The authors show comparisons with different scales, sequence lengths, and baselines.", "weaknesses": "- The method is actually slower for sequence length 4096 and 8192.\n- The tables and figures mainly show latency. I would have wanted to see the effect also on other metrics like memory or GPU utilization.\n- The paper reported error accumulation numbers, but I would have wanted to also see the actual effect on the produced tokens, or maybe just verify that scores on some common benchmarks like MMLU remain the same.", "questions": "- How would these results change with smaller/larger GPUs? Every few years, new GPUs with larger memory and faster processing come out, so I'm curious if these would affect the results.\n- Why is the error accumulation only upto 32k sequences? It seems like the main benefits of the method are more pronounced for longer sequences, so an error accumulation at these scales would be good to see as well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "[Disclaimer that I worked on this field in 2023 and 2024 and have not kept up to date with the latest trends in 2025, so please take this review with a grain of salt.]\n\nThe paper introduces a new scheduling scheme for recurrent transformers, which the authors call \"Diagonal Batching\". The framing is as follows: RMTs have linear inference but are sequential and hence not easily parallelizable. Parallel RMTs localize recurrence within layers and eliminate all inter-layer memory flow. The issue is that PRMTs underutilize GPUs for single, long input requests. The paper introduces diagonal batching as a way to utilizes the GPUs better and hence improve latency. This is done by rearranging the layers and segments into a diagonal structure, where parallel operations can occur at the same time but across different layers.\n\nThe authors show that this provides better latency, especially in long token settings, though the results are less pronounced in the short token settings. They show that results hold across scales. The authors also provide an error analysis which shows that the error accumulation of the method is minimal.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- good exposition of background information and framing of the contributions of the paper\n- proposed method seems simple and can be applied quite generally across different models. If indeed true, the method provides a lot of latency gains basically for free.\n- method section seems quite complete. The authors go through both the high-level motivations but also outline the implementation details.\n- experiments section seems complete as well. The authors show comparisons with different scales, sequence lengths, and baselines.", "weaknesses": "- The method is actually slower for sequence length 4096 and 8192.\n- The tables and figures mainly show latency. I would have wanted to see the effect also on other metrics like memory or GPU utilization.\n- The paper reported error accumulation numbers, but I would have wanted to also see the actual effect on the produced tokens, or maybe just verify that scores on some common benchmarks like MMLU remain the same.", "questions": "- How would these results change with smaller/larger GPUs? Every few years, new GPUs with larger memory and faster processing come out, so I'm curious if these would affect the results.\n- Why is the error accumulation only upto 32k sequences? It seems like the main benefits of the method are more pronounced for longer sequences, so an error accumulation at these scales would be good to see as well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1762123727218}, {"id": "g6qdUp1sVH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24761/Reviewer_XxMs"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces Diagonal Batching, a scheduling method that improves GPU utilization in Parallel Recurrent Memory Transformers (PRMTs), a class of models that maintain per-layer recurrent memory, such as ARMT. The key idea is to reorder computation across layers and segments into “diagonals,” allowing concurrent execution of operations that were previously serialized, without breaking exact recurrence. The authors implement the method in the ARMT framework and evaluate it on LLaMA-based models ranging from 1B to 8B parameters, showing up to 3.3× speedup over standard full-attention inference and 1.8× over sequential ARMT, without requiring custom CUDA kernels. The paper argues that compute scheduling, rather than algorithmic complexity, is the main bottleneck in RMT-style architectures and that Diagonal Batching offers a practical path to efficient, exact linear-time inference for long-context models.", "review_text": "This paper introduces Diagonal Batching, a scheduling method that improves GPU utilization in Parallel Recurrent Memory Transformers (PRMTs), a class of models that maintain per-layer recurrent memory, such as ARMT. The key idea is to reorder computation across layers and segments into “diagonals,” allowing concurrent execution of operations that were previously serialized, without breaking exact recurrence. The authors implement the method in the ARMT framework and evaluate it on LLaMA-based models ranging from 1B to 8B parameters, showing up to 3.3× speedup over standard full-attention inference and 1.8× over sequential ARMT, without requiring custom CUDA kernels. The paper argues that compute scheduling, rather than algorithmic complexity, is the main bottleneck in RMT-style architectures and that Diagonal Batching offers a practical path to efficient, exact linear-time inference for long-context models.", "strengths": "1. Practical relevance – The paper addresses a real bottleneck: GPU underutilization during long-context inference in memory-augmented transformers. The proposal is pragmatic, compatible with existing hardware, and doesn’t require custom CUDA, which makes it accessible.\n\n2. Elegant scheduling insight – The diagonal reordering idea is conceptually simple yet powerful, exposing latent parallelism while preserving recurrence—an often tricky balance.\n\n3. Strong empirical evidence – The results convincingly show latency reduction across multiple model sizes and segment lengths. The experiments are thorough, including FLOPS scaling, micro-batching comparison, and numerical drift analysis.\n\n4. Clarity of technical exposition – The method is mathematically and algorithmically well explained, with clear diagrams (e.g., Figure 1 and 2) that make the scheduling logic intuitive.\n\n5. Compatibility with existing optimizations – The method integrates with FlashAttention, grouped GEMMs, and standard PyTorch implementations, making it easy to adopt.", "weaknesses": "1. Incremental nature – The main novelty lies in scheduling, not modeling or theory. While the implementation is clever, the conceptual leap from standard pipelining or grouped execution is limited. The paper frames this as a major innovation, but it’s more of an engineering optimization than a new algorithmic idea.\n\n2. Limited empirical diversity – All experiments are on LLaMA-based ARMTs. There is no exploration of how the method generalizes to other PRMT architectures (e.g., RWKV or Mamba) beyond brief mentions. Without such experiments, claims of broad applicability remain untested.\n\n3. Reproducibility and accessibility – Although the authors claim to release code, there’s little clarity on integration with existing toolchains or benchmarks. The heavy reliance on ARMT-specific infrastructure might limit reproducibility for broader research use.\n\n4. Evaluation bias – The comparisons are mostly latency-based. There’s no strong discussion of trade-offs in throughput, numerical stability, or memory overhead under multi-request workloads (a realistic serving scenario).\n\n5. Overemphasis on speedups – Some reported gains (like 3.3×) are cherry-picked from small segment sizes or single-GPU setups. Scaling trends on multi-GPU or distributed systems are not shown, yet such setups are where scheduling optimizations often hit diminishing returns.\n\n6. Writing style and tone – The paper reads more like an extended technical report than a standard ICLR submission. I would be more in favor of making the paper concise and 'to-the-point'.", "questions": "1. Have you tested Diagonal Batching on other PRMT variants like RWKV or Mamba to verify general applicability?\n\n2. How does the method perform under concurrent multi-user loads, where GPU memory fragmentation and kernel scheduling might differ?\n\n3. Could you provide memory overhead statistics—does grouped GEMM increase peak memory usage?\n\n4. How significant are the numerical drifts beyond 64K tokens in real tasks?\n\n5. Have you tried combining Diagonal Batching with quantized or sparsified models (e.g., AWQ or FlashDecoding)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Diagonal Batching, a scheduling method that improves GPU utilization in Parallel Recurrent Memory Transformers (PRMTs), a class of models that maintain per-layer recurrent memory, such as ARMT. The key idea is to reorder computation across layers and segments into “diagonals,” allowing concurrent execution of operations that were previously serialized, without breaking exact recurrence. The authors implement the method in the ARMT framework and evaluate it on LLaMA-based models ranging from 1B to 8B parameters, showing up to 3.3× speedup over standard full-attention inference and 1.8× over sequential ARMT, without requiring custom CUDA kernels. The paper argues that compute scheduling, rather than algorithmic complexity, is the main bottleneck in RMT-style architectures and that Diagonal Batching offers a practical path to efficient, exact linear-time inference for long-context models.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Practical relevance – The paper addresses a real bottleneck: GPU underutilization during long-context inference in memory-augmented transformers. The proposal is pragmatic, compatible with existing hardware, and doesn’t require custom CUDA, which makes it accessible.\n\n2. Elegant scheduling insight – The diagonal reordering idea is conceptually simple yet powerful, exposing latent parallelism while preserving recurrence—an often tricky balance.\n\n3. Strong empirical evidence – The results convincingly show latency reduction across multiple model sizes and segment lengths. The experiments are thorough, including FLOPS scaling, micro-batching comparison, and numerical drift analysis.\n\n4. Clarity of technical exposition – The method is mathematically and algorithmically well explained, with clear diagrams (e.g., Figure 1 and 2) that make the scheduling logic intuitive.\n\n5. Compatibility with existing optimizations – The method integrates with FlashAttention, grouped GEMMs, and standard PyTorch implementations, making it easy to adopt.", "weaknesses": "1. Incremental nature – The main novelty lies in scheduling, not modeling or theory. While the implementation is clever, the conceptual leap from standard pipelining or grouped execution is limited. The paper frames this as a major innovation, but it’s more of an engineering optimization than a new algorithmic idea.\n\n2. Limited empirical diversity – All experiments are on LLaMA-based ARMTs. There is no exploration of how the method generalizes to other PRMT architectures (e.g., RWKV or Mamba) beyond brief mentions. Without such experiments, claims of broad applicability remain untested.\n\n3. Reproducibility and accessibility – Although the authors claim to release code, there’s little clarity on integration with existing toolchains or benchmarks. The heavy reliance on ARMT-specific infrastructure might limit reproducibility for broader research use.\n\n4. Evaluation bias – The comparisons are mostly latency-based. There’s no strong discussion of trade-offs in throughput, numerical stability, or memory overhead under multi-request workloads (a realistic serving scenario).\n\n5. Overemphasis on speedups – Some reported gains (like 3.3×) are cherry-picked from small segment sizes or single-GPU setups. Scaling trends on multi-GPU or distributed systems are not shown, yet such setups are where scheduling optimizations often hit diminishing returns.\n\n6. Writing style and tone – The paper reads more like an extended technical report than a standard ICLR submission. I would be more in favor of making the paper concise and 'to-the-point'.", "questions": "1. Have you tested Diagonal Batching on other PRMT variants like RWKV or Mamba to verify general applicability?\n\n2. How does the method perform under concurrent multi-user loads, where GPU memory fragmentation and kernel scheduling might differ?\n\n3. Could you provide memory overhead statistics—does grouped GEMM increase peak memory usage?\n\n4. How significant are the numerical drifts beyond 64K tokens in real tasks?\n\n5. Have you tried combining Diagonal Batching with quantized or sparsified models (e.g., AWQ or FlashDecoding)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761933050846}, {"id": "Cka1HCVZRo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24761/Reviewer_qKhy"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper presents diagonal batching, a method for training and inference of multi-layered recurrent sequence architectures (recurrent memory transformers), where parallelization happens along the diagonal of subsequent segment/token and model layers. The operations across layers can be fused in a batched GEMM-kernels reducing the scheduling overhead. The authors show that their method leads to significant speed-ups for long sequence lengths.", "review_text": "The paper presents diagonal batching, a method for training and inference of multi-layered recurrent sequence architectures (recurrent memory transformers), where parallelization happens along the diagonal of subsequent segment/token and model layers. The operations across layers can be fused in a batched GEMM-kernels reducing the scheduling overhead. The authors show that their method leads to significant speed-ups for long sequence lengths.", "strengths": "- general method for sequence-recurrent architectures\n- good speed ups (1.1-3.3) for the pre-fill part (latency) of inference", "weaknesses": "- limited application to pre-filling (latency) optimization of recurrent LLMs\n- no practical RMT model (e.g. an existing RWKV/xLSTM/Mamba-based model) shown where this is applied (e.g. for reasoning tasks, where long sequences would be strongly beneficial)\n- RMTs in general break translational invariance in text, so some tokens would be \"different\" from others depending on their positions (especially at the segment border)", "questions": "- Are there existing RMT models where this can have practical benefits?\n- Could you show how this could be implemented on large recurrent models, like FalconMamba-7B [1] or xLSTM-7B [2] for practical use?\n- Can this also be beneficial in model training on long sequences?\n- Can you add a small section on why this method does not work for decoding/generation?\n- you mention TC0 model complexity as a general downside of Transformers/State Space Models, but don't position your method along this axis with general state-tracking enabled architectures like xLSTM [3] or DeltaProduct [4], how does it relate?\n\n[1] https://huggingface.co/tiiuae/falcon-mamba-7b\n\n[2] https://huggingface.co/NX-AI/xLSTM-7b\n\n[3] Beck et al. (2024): Extended Long Short-Term Memory\n\n[4] Siems et al. (2025): DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents diagonal batching, a method for training and inference of multi-layered recurrent sequence architectures (recurrent memory transformers), where parallelization happens along the diagonal of subsequent segment/token and model layers. The operations across layers can be fused in a batched GEMM-kernels reducing the scheduling overhead. The authors show that their method leads to significant speed-ups for long sequence lengths.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- general method for sequence-recurrent architectures\n- good speed ups (1.1-3.3) for the pre-fill part (latency) of inference", "weaknesses": "- limited application to pre-filling (latency) optimization of recurrent LLMs\n- no practical RMT model (e.g. an existing RWKV/xLSTM/Mamba-based model) shown where this is applied (e.g. for reasoning tasks, where long sequences would be strongly beneficial)\n- RMTs in general break translational invariance in text, so some tokens would be \"different\" from others depending on their positions (especially at the segment border)", "questions": "- Are there existing RMT models where this can have practical benefits?\n- Could you show how this could be implemented on large recurrent models, like FalconMamba-7B [1] or xLSTM-7B [2] for practical use?\n- Can this also be beneficial in model training on long sequences?\n- Can you add a small section on why this method does not work for decoding/generation?\n- you mention TC0 model complexity as a general downside of Transformers/State Space Models, but don't position your method along this axis with general state-tracking enabled architectures like xLSTM [3] or DeltaProduct [4], how does it relate?\n\n[1] https://huggingface.co/tiiuae/falcon-mamba-7b\n\n[2] https://huggingface.co/NX-AI/xLSTM-7b\n\n[3] Beck et al. (2024): Extended Long Short-Term Memory\n\n[4] Siems et al. (2025): DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761582733812}], "openreview_url": "https://openreview.net/forum?id=z1yPH2Rska", "arxiv_id": "2506.05229", "paper_pdf": "papers/z1yPH2Rska.pdf", "paper_pdf_sha256": "6fcb10e2ca0f5614dd017ab1ee764c42b648819267bb3fd2ae4e15c96d5e3752", "paper_pdf_bytes": 983470, "paper_pdf_source": "openreview", "code_url": "https://github.com/svtdanny/diagonal-batching", "code_repository": "svtdanny/diagonal-batching", "code_commit": "641ab9fc2c027e8b2614d76368e81cd2ee4bd1f8", "code_archive": "repos/z1yPH2Rska.zip", "code_archive_sha256": "f36de8bd1624bffc6cc5c85dcad131d5f8af721cf935866fdb42d7c7a616cded", "code_archive_bytes": 994323, "code_file_count": 30, "code_extensions": {".py": 20, ".ipynb": 10}, "github_disk_usage_kb": 1703, "github_languages": {"Jupyter Notebook": 1157285, "Python": 184303}, "github_archived": false, "github_pushed_at": "2025-09-02T19:39:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/diagonal-batching-unlocks-parallelism-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WL4BmXG7Pl", "year": 2025, "status": "rejected", "title": "Crafting Heavy-Tails in Weight Matrix Spectrum without Gradient Noise", "authors": ["Vignesh Kothapalli", "Tianyu Pang", "Shenyang Deng", "Zongmin Liu", "Yaoqing Yang"], "authorids": ["~Vignesh_Kothapalli1", "~Tianyu_Pang2", "~Shenyang_Deng2", "~Zongmin_Liu1", "~Yaoqing_Yang1"], "authors_source": "OpenReview API", "abstract": "Training strategies for modern deep neural networks (NNs) tend to induce a heavy-\ntailed (HT) empirical spectral density (ESD) in the layer weights. While previous\nefforts have shown that the HT phenomenon correlates with good generalization\nin large NNs, a theoretical explanation of its occurrence is still lacking. Especially,\nunderstanding the conditions which lead to this phenomenon can shed light on the\ninterplay between generalization and weight spectra. Our work aims to bridge this\ngap by presenting a simple, rich setting to model the emergence of HT ESD. In\nparticular, we present a theory-informed analysis for 'crafting' heavy tails in the\nESD of two-layer NNs without any gradient noise. This is the first work to analyze a noise-free setting and incorporate optimizer (GD/Adam) dependent (large)\nlearning rates into the HT ESD analysis. Our results highlight the role of learning\nrates on the Bulk+Spike and HT shape of the ESDs in the early phase of training,\nwhich can facilitate generalization in the two-layer NN. These observations shed\nlight on the behavior of large-scale NNs, albeit in a much simpler setting. Last\nbut not least, we present a novel perspective on the ESD evolution dynamics by\nanalyzing the singular vectors of weight matrices and optimizer updates", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "5FNO330gQP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8643/Reviewer_uys5"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper studies the heavy tail phenomenon of the spectrum of weight matrices and its impact on generalization.  The paper presents a theorem for single step of Adam which shows an emergence of \"Bulk + spike\" spectrum. It futher empirical evaluates how this spectrum evolves into a heavy tail.", "review_text": "The paper studies the heavy tail phenomenon of the spectrum of weight matrices and its impact on generalization.  The paper presents a theorem for single step of Adam which shows an emergence of \"Bulk + spike\" spectrum. It futher empirical evaluates how this spectrum evolves into a heavy tail.", "strengths": "a) The paper extends the single step analysis of GD Ba et. al. for Adam.   \nb) The observation that bulk + spike evolves into a heavy tail is interesting.", "weaknesses": "a) The paper lacks sufficient motivation for the problem setting and the relevance of the chosen algorithm, leaving it unclear why this is an important quantity to study. Notably, both the heavy-tail mechanism and strong generalization stem from learning the single-index direction, making it uncertain whether the heavy tail is simply a byproduct of good generalization or its cause. Consequently, it is challenging to assess whether this problem setting and analysis truly captures the influence of the heavy-tail spectrum on generalization.\n\nb) The empirical analysis is restricted to single-index teachers and two-layer networks, which limits the study's scope and makes it insufficient to determine whether any findings are broadly applicable.\n\nc) The theoritical analysis does not capture how \"BULK + spike\" transition into a heavy tail phenomenon. Hence the paper falls short of theoritically understanding the emergence of HT-ESD - the theoritical analysis is only captures the BULK + spike shape of spectrum after a single step of Adam.\n\nMinor :\n- Make a consistent notation of $\\beta^{*}$ through out the paper.\n- The inline math can be formatted better to ensure better readbility.", "questions": "discussed above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the heavy tail phenomenon of the spectrum of weight matrices and its impact on generalization.  The paper presents a theorem for single step of Adam which shows an emergence of \"Bulk + spike\" spectrum. It futher empirical evaluates how this spectrum evolves into a heavy tail.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "a) The paper extends the single step analysis of GD Ba et. al. for Adam.   \nb) The observation that bulk + spike evolves into a heavy tail is interesting.", "weaknesses": "a) The paper lacks sufficient motivation for the problem setting and the relevance of the chosen algorithm, leaving it unclear why this is an important quantity to study. Notably, both the heavy-tail mechanism and strong generalization stem from learning the single-index direction, making it uncertain whether the heavy tail is simply a byproduct of good generalization or its cause. Consequently, it is challenging to assess whether this problem setting and analysis truly captures the influence of the heavy-tail spectrum on generalization.\n\nb) The empirical analysis is restricted to single-index teachers and two-layer networks, which limits the study's scope and makes it insufficient to determine whether any findings are broadly applicable.\n\nc) The theoritical analysis does not capture how \"BULK + spike\" transition into a heavy tail phenomenon. Hence the paper falls short of theoritically understanding the emergence of HT-ESD - the theoritical analysis is only captures the BULK + spike shape of spectrum after a single step of Adam.\n\nMinor :\n- Make a consistent notation of $\\beta^{*}$ through out the paper.\n- The inline math can be formatted better to ensure better readbility.", "questions": "discussed above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730822338722}, {"id": "ZKFZyhDMvy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8643/Reviewer_bo67"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper aims to investigate the relationship between heavy tails in the weight spectrum distribution of neural networks and the ability of the network to generalize over unseen samples. The authors set up a teacher-student setting where a two-layer neural network learns a single-index model. First, the paper shows that for a sufficiently large step size, the full-batch Adam can align with the teacher's direction in just one step, corresponding to a spike in the ESD emerging from the initial mass. The analysis continues by empirically showing that the dynamics of the ESD evolves from the bulk to a heavy tail distribution. Finally, the paper shows the connection between the ESD heavy tails and the orientation of the singular vector of weight updates.", "review_text": "The paper aims to investigate the relationship between heavy tails in the weight spectrum distribution of neural networks and the ability of the network to generalize over unseen samples. The authors set up a teacher-student setting where a two-layer neural network learns a single-index model. First, the paper shows that for a sufficiently large step size, the full-batch Adam can align with the teacher's direction in just one step, corresponding to a spike in the ESD emerging from the initial mass. The analysis continues by empirically showing that the dynamics of the ESD evolves from the bulk to a heavy tail distribution. Finally, the paper shows the connection between the ESD heavy tails and the orientation of the singular vector of weight updates.", "strengths": "- To my knowledge, this paper is the first to study the HT-generalization correspondence in the single-index teacher-student setting. Altough not bringing any revolutionizing idea, the paper establishes an important first step towards understating this phenomenon within the community.\n- The analysis is comprehensive and precise, also taking into account techniques used in practice such as weight normalization and learning rate schedulers.", "weaknesses": "- The paper is showing mostly empirical result. In this theoretical setting one might expect to have more theoretical support of the claims.", "questions": "- In line 241, shouldn't we divide the similarity by the norm of $u_1$? Otherwise it can grow even if $u_1$ and $\\beta$ are not getting more aligned\n- In line 308 you talk about a sweet for $\\eta$, but I can't see it in the figure, can you clarify?\n\nMinor:\n- Line 297: $\\eta=0.1$, but the figures says $\\eta=1.$", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to investigate the relationship between heavy tails in the weight spectrum distribution of neural networks and the ability of the network to generalize over unseen samples. The authors set up a teacher-student setting where a two-layer neural network learns a single-index model. First, the paper shows that for a sufficiently large step size, the full-batch Adam can align with the teacher's direction in just one step, corresponding to a spike in the ESD emerging from the initial mass. The analysis continues by empirically showing that the dynamics of the ESD evolves from the bulk to a heavy tail distribution. Finally, the paper shows the connection between the ESD heavy tails and the orientation of the singular vector of weight updates.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "- To my knowledge, this paper is the first to study the HT-generalization correspondence in the single-index teacher-student setting. Altough not bringing any revolutionizing idea, the paper establishes an important first step towards understating this phenomenon within the community.\n- The analysis is comprehensive and precise, also taking into account techniques used in practice such as weight normalization and learning rate schedulers.", "weaknesses": "- The paper is showing mostly empirical result. In this theoretical setting one might expect to have more theoretical support of the claims.", "questions": "- In line 241, shouldn't we divide the similarity by the norm of $u_1$? Otherwise it can grow even if $u_1$ and $\\beta$ are not getting more aligned\n- In line 308 you talk about a sweet for $\\eta$, but I can't see it in the figure, can you clarify?\n\nMinor:\n- Line 297: $\\eta=0.1$, but the figures says $\\eta=1.$", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730712171302}, {"id": "1WHsrhVhA7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8643/Reviewer_xTm5"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper is interested in understanding the evolution of the empirical spectral distribution for the first-layer weight matrix in a two-layer neural network during training. The paper considers a student teacher (single index model) with Gaussian. It shows that early-stage feature learning occurs for a much smaller step size for Adam compared to GD. \n\nThe paper also empirically studies the spectrum's evolution during further training and claims the existence of the initial spike is needed to transition to the heavy-tailed phase. Finally, the paper looks at various kernel alignment metrics, their evolution, and correlation with generalization.\n\n**Justification for Score**\n\nI think overall the paper presents intersting insights for a different optimizer. Hence I think it is worth accepting. However, the paper doesn't always position itselves in the most informative way compared to prior work and I have concerns about the significance of the spike that has been seen. Hence I do not give it a higher score.", "review_text": "This paper is interested in understanding the evolution of the empirical spectral distribution for the first-layer weight matrix in a two-layer neural network during training. The paper considers a student teacher (single index model) with Gaussian. It shows that early-stage feature learning occurs for a much smaller step size for Adam compared to GD. \n\nThe paper also empirically studies the spectrum's evolution during further training and claims the existence of the initial spike is needed to transition to the heavy-tailed phase. Finally, the paper looks at various kernel alignment metrics, their evolution, and correlation with generalization.\n\n**Justification for Score**\n\nI think overall the paper presents intersting insights for a different optimizer. Hence I think it is worth accepting. However, the paper doesn't always position itselves in the most informative way compared to prior work and I have concerns about the significance of the spike that has been seen. Hence I do not give it a higher score.", "strengths": "**Novelty:** As far as I know, the theoretical works on the early-stage emergence of spikes have been primarily limited to gradient descent. No theoretical work I know of analyzes Adam. Hence I think this is a novel contribution of the paper. Additionally, the insight that Adam requires a smaller step size than gradient descent to see the spike is important. \n\n\n**Clarity:** The paper is mostly well-written. My one concern is that when people talk about the spiked structure (Ba et al, Moniri et al). they mean for the features $F = \\sigma(WX)$ and not for the weight matrix $W$. This was confusing at first.", "weaknesses": "**Originality:** In terms of originality, the paper should do a better job of positioning itself with respect to Martin and Mahoney,  Ba et al., and Moniri et al. I think this is important. \n\n\n**Significance:** I think the result on smaller step size results in spikes is significant. If this were a result as presented in Moniri et al. However, as a result is currently presented, I am not sure the theory result says enough. Specifically, \n\n1. The paragraph after the corollary is clear to me. The step size scales are different for the spectral norm and the Frobenius norm. How do we rationalize this to show that there is a spike?\n\n2. Spikes in $W$ do not necessarily correspond to spikes in $F$. Especially if you have Gaussian Data. See [1] (flipping the role of $W$ and $X$). \n\n[1] Wang, Zhichao, Denny Wu, and Zhou Fan. \"Nonlinear spiked covariance matrices and signal propagation in deep neural networks.\" arXiv preprint arXiv:2402.10127 COLT (2024).", "questions": "1. Generalization with Heavy-Tailed models. Do the authors know of any works that theoretically characterize the generalization error of models with heavy tails? Even in the regression setting I know of [2]. Also, do the authors know about papers that consider generalization error for models with spiked covariance (besides the Ba et al. 2023 paper), again including regression. \n\n[2] Wang, Yutong, Rishi Sonthalia, and Wei Hu. \"Near-interpolators: Rapid norm growth and the trade-off between interpolation and generalization.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2024.\n\n2. Emergence of heavy tails. For $\\eta=0.1$, the experiment needed $t \\sim 10^4$. However, the authors claim $t \\sim 10^4$ was not enough for $\\eta = 0.01$. I imagine we need more steps, maybe even something like $t \\sim 10^7$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper is interested in understanding the evolution of the empirical spectral distribution for the first-layer weight matrix in a two-layer neural network during training. The paper considers a student teacher (single index model) with Gaussian. It shows that early-stage feature learning occurs for a much smaller step size for Adam compared to GD. \n\nThe paper also empirically studies the spectrum's evolution during further training and claims the existence of the initial spike is needed to transition to the heavy-tailed phase. Finally, the paper looks at various kernel alignment metrics, their evolution, and correlation with generalization.\n\n**Justification for Score**\n\nI think overall the paper presents intersting insights for a different optimizer. Hence I think it is worth accepting. However, the paper doesn't always position itselves in the most informative way compared to prior work and I have concerns about the significance of the spike that has been seen. Hence I do not give it a higher score.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "**Novelty:** As far as I know, the theoretical works on the early-stage emergence of spikes have been primarily limited to gradient descent. No theoretical work I know of analyzes Adam. Hence I think this is a novel contribution of the paper. Additionally, the insight that Adam requires a smaller step size than gradient descent to see the spike is important. \n\n\n**Clarity:** The paper is mostly well-written. My one concern is that when people talk about the spiked structure (Ba et al, Moniri et al). they mean for the features $F = \\sigma(WX)$ and not for the weight matrix $W$. This was confusing at first.", "weaknesses": "**Originality:** In terms of originality, the paper should do a better job of positioning itself with respect to Martin and Mahoney,  Ba et al., and Moniri et al. I think this is important. \n\n\n**Significance:** I think the result on smaller step size results in spikes is significant. If this were a result as presented in Moniri et al. However, as a result is currently presented, I am not sure the theory result says enough. Specifically, \n\n1. The paragraph after the corollary is clear to me. The step size scales are different for the spectral norm and the Frobenius norm. How do we rationalize this to show that there is a spike?\n\n2. Spikes in $W$ do not necessarily correspond to spikes in $F$. Especially if you have Gaussian Data. See [1] (flipping the role of $W$ and $X$). \n\n[1] Wang, Zhichao, Denny Wu, and Zhou Fan. \"Nonlinear spiked covariance matrices and signal propagation in deep neural networks.\" arXiv preprint arXiv:2402.10127 COLT (2024).", "questions": "1. Generalization with Heavy-Tailed models. Do the authors know of any works that theoretically characterize the generalization error of models with heavy tails? Even in the regression setting I know of [2]. Also, do the authors know about papers that consider generalization error for models with spiked covariance (besides the Ba et al. 2023 paper), again including regression. \n\n[2] Wang, Yutong, Rishi Sonthalia, and Wei Hu. \"Near-interpolators: Rapid norm growth and the trade-off between interpolation and generalization.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2024.\n\n2. Emergence of heavy tails. For $\\eta=0.1$, the experiment needed $t \\sim 10^4$. However, the authors claim $t \\sim 10^4$ was not enough for $\\eta = 0.01$. I imagine we need more steps, maybe even something like $t \\sim 10^7$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730679074440}, {"id": "S5ZsVgkRWb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8643/Reviewer_piS8"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The submission studies an interesting problem. When does HT-ESD happen? Why does it lead to good generalization?\n\nThe key results are below:\n\n1. Using a toy model,  ADAM after one step can provably induce a spike in spetrum. (thm 4.1)\n\n2. With experiments, the authors show that bulk decay happens (Fig 4), and that spike + decay leads to heavy tail.\n\n3. With experiments, the authors show that successful feature learning happens when the spectrum has the right tail exponent. (fig 5)\n\n##Strength:##\n\nI like the experiment part of this work.  In particular, comparing ADAM vs GD in Fig2 Fig3, and demonstrating the decay in Fig 4. \n\nI think the perspective is novel. The spectrum distribution not only results from the minibatch noise but also from full batch updates.\n\n##Weakness:##\n\nAlthough I enjoy reading the work, I am not confident that the authors successfully answer the questions asked.\n\n1. The spike after one large update is trivial. Given the backprop structure of neural net, one step update to the weight is rank one. When the step is large, one gets a rank-one spike.\n\n2. The paper demonstrates that decay happens without explaining why. More specifically, does the decay happen just correctly such the spectrum is heavy tailed?\n\n3. The paper demonstrates that the tail index should be in the correct range for the features to be learned. However, no explanation is given. Further, the experiments for this result are synthetic and toy.\n\nI am happy to update my review if the authors could address the above concern.\n\n\n##Minors:##\n\nWhat is u in line 241?", "review_text": "The submission studies an interesting problem. When does HT-ESD happen? Why does it lead to good generalization?\n\nThe key results are below:\n\n1. Using a toy model,  ADAM after one step can provably induce a spike in spetrum. (thm 4.1)\n\n2. With experiments, the authors show that bulk decay happens (Fig 4), and that spike + decay leads to heavy tail.\n\n3. With experiments, the authors show that successful feature learning happens when the spectrum has the right tail exponent. (fig 5)\n\n##Strength:##\n\nI like the experiment part of this work.  In particular, comparing ADAM vs GD in Fig2 Fig3, and demonstrating the decay in Fig 4. \n\nI think the perspective is novel. The spectrum distribution not only results from the minibatch noise but also from full batch updates.\n\n##Weakness:##\n\nAlthough I enjoy reading the work, I am not confident that the authors successfully answer the questions asked.\n\n1. The spike after one large update is trivial. Given the backprop structure of neural net, one step update to the weight is rank one. When the step is large, one gets a rank-one spike.\n\n2. The paper demonstrates that decay happens without explaining why. More specifically, does the decay happen just correctly such the spectrum is heavy tailed?\n\n3. The paper demonstrates that the tail index should be in the correct range for the features to be learned. However, no explanation is given. Further, the experiments for this result are synthetic and toy.\n\nI am happy to update my review if the authors could address the above concern.\n\n\n##Minors:##\n\nWhat is u in line 241?", "strengths": "See summary", "weaknesses": "See summary", "questions": "See summary", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission studies an interesting problem. When does HT-ESD happen? Why does it lead to good generalization?\n\nThe key results are below:\n\n1. Using a toy model,  ADAM after one step can provably induce a spike in spetrum. (thm 4.1)\n\n2. With experiments, the authors show that bulk decay happens (Fig 4), and that spike + decay leads to heavy tail.\n\n3. With experiments, the authors show that successful feature learning happens when the spectrum has the right tail exponent. (fig 5)\n\n##Strength:##\n\nI like the experiment part of this work.  In particular, comparing ADAM vs GD in Fig2 Fig3, and demonstrating the decay in Fig 4. \n\nI think the perspective is novel. The spectrum distribution not only results from the minibatch noise but also from full batch updates.\n\n##Weakness:##\n\nAlthough I enjoy reading the work, I am not confident that the authors successfully answer the questions asked.\n\n1. The spike after one large update is trivial. Given the backprop structure of neural net, one step update to the weight is rank one. When the step is large, one gets a rank-one spike.\n\n2. The paper demonstrates that decay happens without explaining why. More specifically, does the decay happen just correctly such the spectrum is heavy tailed?\n\n3. The paper demonstrates that the tail index should be in the correct range for the features to be learned. However, no explanation is given. Further, the experiments for this result are synthetic and toy.\n\nI am happy to update my review if the authors could address the above concern.\n\n\n##Minors:##\n\nWhat is u in line 241?", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "See summary", "weaknesses": "See summary", "questions": "See summary", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729327337292}], "openreview_url": "https://openreview.net/forum?id=WL4BmXG7Pl", "arxiv_id": "2406.04657", "paper_pdf": "papers/WL4BmXG7Pl.pdf", "paper_pdf_sha256": "52a01629a41d6a38a5fb40587fb51c30df93b1e880de7e186038dbf19afe3d1d", "paper_pdf_bytes": 6872743, "paper_pdf_source": "openreview", "code_url": "https://github.com/kvignesh1420/single-index-ht", "code_repository": "kvignesh1420/single-index-ht", "code_commit": "901ee634e80f2151d2b4f92c08acf3649fb2a871", "code_archive": "repos/WL4BmXG7Pl.zip", "code_archive_sha256": "1d64731b3b52d4f013d62b8f481d4c5f0e4fa1519f083d7d5765cb1a79c4933c", "code_archive_bytes": 25948, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 122, "github_languages": {"Python": 65585}, "github_archived": false, "github_pushed_at": "2026-03-23T20:08:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/crafting-heavy-tails-in-weight-matrix"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "CupHThqQl3", "year": 2024, "status": "rejected", "title": "It's About Time: Temporal References in Emergent Communication", "authors": ["Olaf Lipinski", "Adam Sobey", "Federico Cerutti", "Timothy J. Norman"], "authorids": ["~Olaf_Lipinski1", "~Adam_Sobey1", "~Federico_Cerutti1", "~Timothy_J._Norman1"], "authors_source": "OpenReview API", "abstract": "As humans, we use linguistic elements referencing time, such as “before” or “tomorrow”, to easily share past experiences and future predictions. While temporal aspects of the language have been considered in computational linguistics, no such exploration has been done within the field of emergent communication. We research this gap, providing the first reported temporal vocabulary within emergent communication literature. Our experimental analysis shows that a different agent architecture is sufficient for the natural emergence of temporal references, and that no additional losses are necessary. Our readily transferable architectural insights provide the basis for the incorporation of temporal referencing into other emergent communication environments.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Sb3uubg5vV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5303/Reviewer_Avm5"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a novel dimension of emergent communication, focusing on temporal reference. In this framework, the sender is tasked with encoding temporal information about target objects, enabling the receiver to distinguish them from other distractors. To facilitate this, a temporal LSTM component is incorporated into the agents' architecture for temporal encoding. Furthermore, a loss function for temporal object classification is introduced to encourage the emergence of temporal encoding. To assess the effectiveness of this approach, an evaluation metric is devised to measure the frequency of using the same message to describe previously presented objects. The experimental findings highlight the role of the temporal LSTM in facilitating the development of temporal communication.", "review_text": "This paper introduces a novel dimension of emergent communication, focusing on temporal reference. In this framework, the sender is tasked with encoding temporal information about target objects, enabling the receiver to distinguish them from other distractors. To facilitate this, a temporal LSTM component is incorporated into the agents' architecture for temporal encoding. Furthermore, a loss function for temporal object classification is introduced to encourage the emergence of temporal encoding. To assess the effectiveness of this approach, an evaluation metric is devised to measure the frequency of using the same message to describe previously presented objects. The experimental findings highlight the role of the temporal LSTM in facilitating the development of temporal communication.", "strengths": "1. This paper introduces temporal reference into the emergent communication field, which is an important topic worth exploring. It leverages the sequential position of objects within the batch as a means of encoding temporal information, thereby modifying traditional referential games and their associated training frameworks.\n\n2. The paper demonstrates a comprehensive design of the experiments. It encompasses various modifications to modeling architectures, incorporates multiple environments, and also analyzes different statistical results.", "weaknesses": "1. Example demonstration: the authors present multiple examples using integer arrays in the game design (section 2.3), agent architecture (figure 1), how to compute the temporality metric (section 3.1), as well as the sent messages (section 3.3), which I believe are of different meanings. However, using integer arrays in all the examples can make the demonstration more confusing. Different symbols used in different cases would be better. For example, objects as names, and messages using alphabets.\n\n2. Experimental analysis clarification:\n\n    a. In section 3.3, it is said “only networks that have the sequential LSTM are capable of producing temporal references”: From Figure 2a, we can observe all networks have cases “reaching % used as previous” above 80%. A clarification on how you make a judgment on “producing temporal references” should be made. \n\n    b. In Figure 3b, only the results of Temporal networks are presented because “non-temporal networks learn no temporally specialised messages”. The threshold and significance between temporally and no temporally specialised messages should be made clear as well. \n\n    c. Through Figure 3a, the claim that “these messages could be a more efficient way of describing objects” needs to be further clarified. \n\n    d. In Figure 2, which of the six environments used for the legends about “trained RGs” and “TRGs” are not specified. \n\n    e. During the experiment analysis, the results of “Never same” are not presented. Also, there is not much difference, or the difference is not underscored between “environment” and “environment Hard”. This part could be better organized based on the analysis needed.", "questions": "1. For the temporal loss, I assume it is a categorical classification about the last position the object shows up. I think it could be misleading when trained with target object classification together. There are some circumstances when the objects are the same but the temporal labels are different, which may cause the network to be confused. Have you tried to add a smaller weight to the temporal loss or try to add this loss on top of the temporal LSTM?\n\n2. More analysis of the features that the temporal LSTM provides would be helpful to demonstrate its vital role.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel dimension of emergent communication, focusing on temporal reference. In this framework, the sender is tasked with encoding temporal information about target objects, enabling the receiver to distinguish them from other distractors. To facilitate this, a temporal LSTM component is incorporated into the agents' architecture for temporal encoding. Furthermore, a loss function for temporal object classification is introduced to encourage the emergence of temporal encoding. To assess the effectiveness of this approach, an evaluation metric is devised to measure the frequency of using the same message to describe previously presented objects. The experimental findings highlight the role of the temporal LSTM in facilitating the development of temporal communication.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. This paper introduces temporal reference into the emergent communication field, which is an important topic worth exploring. It leverages the sequential position of objects within the batch as a means of encoding temporal information, thereby modifying traditional referential games and their associated training frameworks.\n\n2. The paper demonstrates a comprehensive design of the experiments. It encompasses various modifications to modeling architectures, incorporates multiple environments, and also analyzes different statistical results.", "weaknesses": "1. Example demonstration: the authors present multiple examples using integer arrays in the game design (section 2.3), agent architecture (figure 1), how to compute the temporality metric (section 3.1), as well as the sent messages (section 3.3), which I believe are of different meanings. However, using integer arrays in all the examples can make the demonstration more confusing. Different symbols used in different cases would be better. For example, objects as names, and messages using alphabets.\n\n2. Experimental analysis clarification:\n\n    a. In section 3.3, it is said “only networks that have the sequential LSTM are capable of producing temporal references”: From Figure 2a, we can observe all networks have cases “reaching % used as previous” above 80%. A clarification on how you make a judgment on “producing temporal references” should be made. \n\n    b. In Figure 3b, only the results of Temporal networks are presented because “non-temporal networks learn no temporally specialised messages”. The threshold and significance between temporally and no temporally specialised messages should be made clear as well. \n\n    c. Through Figure 3a, the claim that “these messages could be a more efficient way of describing objects” needs to be further clarified. \n\n    d. In Figure 2, which of the six environments used for the legends about “trained RGs” and “TRGs” are not specified. \n\n    e. During the experiment analysis, the results of “Never same” are not presented. Also, there is not much difference, or the difference is not underscored between “environment” and “environment Hard”. This part could be better organized based on the analysis needed.", "questions": "1. For the temporal loss, I assume it is a categorical classification about the last position the object shows up. I think it could be misleading when trained with target object classification together. There are some circumstances when the objects are the same but the temporal labels are different, which may cause the network to be confused. Have you tried to add a smaller weight to the temporal loss or try to add this loss on top of the temporal LSTM?\n\n2. More analysis of the features that the temporal LSTM provides would be helpful to demonstrate its vital role.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698623131205}, {"id": "Ai1RGCbugI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5303/Reviewer_RdRa"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a variant of the common referential game in emergent\ncommunication game: the Temporal Referential Game (TRG).  The TRG is an\niterated version of the standard referential game where recent objects have\na higher probability of being seen again compared to non-recently seen objects.\nThe authors introduce a temporal LSTM module and a temporal loss to better\nsolve this game and observe the emergence of temporally referent messages.", "review_text": "This paper introduces a variant of the common referential game in emergent\ncommunication game: the Temporal Referential Game (TRG).  The TRG is an\niterated version of the standard referential game where recent objects have\na higher probability of being seen again compared to non-recently seen objects.\nThe authors introduce a temporal LSTM module and a temporal loss to better\nsolve this game and observe the emergence of temporally referent messages.", "strengths": "- (major) Temporal reference is a pertinent feature of language, and therefore,\n  it is good to use an environment with multiple timesteps with additional\n  temporal structure (i.e., recently seen object are more likely to appear\n  again).\n- (minor) The Temporal Referential Game (TRG) is an appropriate extension to\n  the regular referential game.  It minimally adds a temporal aspect to the\n  game without changing it too much, making it good for pioneering basic\n  concepts about temporal reference.", "weaknesses": "This paper does not make a substantial contribution to the field.  While the\nTRG could elicit potentially interesting behaviors, the primary thrust of this\npaper is that it introduces a new architecture which is able to solve this game\nwithout much further analysis about temporality.  The problem here is that the\nnew architecture is, to the best of my knowledge, just an LSTM where each\ntimestep is a single round of the referential game.  This, to me, seems like an\nappropriate baseline with which to work with and does not seem like\na contribution in and of itself.  Additionally, the temporal loss seems to work\nagainst the claim that temporal reference \"emerges from functional pressures\";\nthis claim I see as being implicit in most all emergent communication research,\notherwise we are more or less looking at a trivial supervised learning problem.", "questions": "What is the significance of the temporal loss and the temporal LSTM?  What do\nthey offer beyond being a commonsense baseline for the TRG?\n\n### Misc comments\n\n- Sec 1\n    - social deduction game: cite related work (e.g., [this](https://www.semanticscholar.org/paper/RLupus%3A-Cooperation-through-emergent-communication-Brandizzi-Grossi/d94504e5319ea1a2dc7df8a1ad25a48a4c3ae650))\n- Sec 2.1\n    - \"attribute-value\" seems more standard than \"property-feature\"; the latter\n      sounds like \"property\" and \"feature\" could refer to the same thing\n- Sec 2.3\n    - Seems like you could just parameterize the threshold for $c$?\n    - Alternatively, it would probably be even clearer to say that we sample\n      $c$ from a Bernoulli distribution since that is in effect what is\n      happening.  Also, saying $c$ is uniformly sampled integer from $[0,1]$ is\n      sort of confusing since interval notation refers to real values, not\n      integers.\n    - \"verify whether the messages\" -> \"determine whether or not the messages\"\n- Sec 2.4\n    - \"same as the EGG framework\" is not a good way to specify the architecture\n      since it is a whole framework and not a particular model.  Maybe just\n      take a sentence or two to specify the defaults that are being used.\n  p6\n- Sec 3.1\n    - The $100\\\\%$ seems too restrictive to determine a word is properly\n      temporal reference.  For example, in English, is \"yesterday\" temporally\n      referent?  Because it can be used in a non-temporally referent way\n      (e.g., \"TensorFlow is yesterday's ML library.\").\n- Sec 3.2\n    - It seems like the non-temporal network is strictly precluded from\n      performing temporal reference?  How is this a useful baseline, then?  Are\n      we seeing if it can perform just as well or what the random baseline of\n      temporal reference might be?\n- Sec 3.3\n  - Is $M_{\\ominus^4}$ the max value of any message?  It seems like the\n    temporality is per-message, not per-language.\n- Figures 2a and 3a are difficult to read.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a variant of the common referential game in emergent\ncommunication game: the Temporal Referential Game (TRG).  The TRG is an\niterated version of the standard referential game where recent objects have\na higher probability of being seen again compared to non-recently seen objects.\nThe authors introduce a temporal LSTM module and a temporal loss to better\nsolve this game and observe the emergence of temporally referent messages.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- (major) Temporal reference is a pertinent feature of language, and therefore,\n  it is good to use an environment with multiple timesteps with additional\n  temporal structure (i.e., recently seen object are more likely to appear\n  again).\n- (minor) The Temporal Referential Game (TRG) is an appropriate extension to\n  the regular referential game.  It minimally adds a temporal aspect to the\n  game without changing it too much, making it good for pioneering basic\n  concepts about temporal reference.", "weaknesses": "This paper does not make a substantial contribution to the field.  While the\nTRG could elicit potentially interesting behaviors, the primary thrust of this\npaper is that it introduces a new architecture which is able to solve this game\nwithout much further analysis about temporality.  The problem here is that the\nnew architecture is, to the best of my knowledge, just an LSTM where each\ntimestep is a single round of the referential game.  This, to me, seems like an\nappropriate baseline with which to work with and does not seem like\na contribution in and of itself.  Additionally, the temporal loss seems to work\nagainst the claim that temporal reference \"emerges from functional pressures\";\nthis claim I see as being implicit in most all emergent communication research,\notherwise we are more or less looking at a trivial supervised learning problem.", "questions": "What is the significance of the temporal loss and the temporal LSTM?  What do\nthey offer beyond being a commonsense baseline for the TRG?\n\n### Misc comments\n\n- Sec 1\n    - social deduction game: cite related work (e.g., [this](https://www.semanticscholar.org/paper/RLupus%3A-Cooperation-through-emergent-communication-Brandizzi-Grossi/d94504e5319ea1a2dc7df8a1ad25a48a4c3ae650))\n- Sec 2.1\n    - \"attribute-value\" seems more standard than \"property-feature\"; the latter\n      sounds like \"property\" and \"feature\" could refer to the same thing\n- Sec 2.3\n    - Seems like you could just parameterize the threshold for $c$?\n    - Alternatively, it would probably be even clearer to say that we sample\n      $c$ from a Bernoulli distribution since that is in effect what is\n      happening.  Also, saying $c$ is uniformly sampled integer from $[0,1]$ is\n      sort of confusing since interval notation refers to real values, not\n      integers.\n    - \"verify whether the messages\" -> \"determine whether or not the messages\"\n- Sec 2.4\n    - \"same as the EGG framework\" is not a good way to specify the architecture\n      since it is a whole framework and not a particular model.  Maybe just\n      take a sentence or two to specify the defaults that are being used.\n  p6\n- Sec 3.1\n    - The $100\\\\%$ seems too restrictive to determine a word is properly\n      temporal reference.  For example, in English, is \"yesterday\" temporally\n      referent?  Because it can be used in a non-temporally referent way\n      (e.g., \"TensorFlow is yesterday's ML library.\").\n- Sec 3.2\n    - It seems like the non-temporal network is strictly precluded from\n      performing temporal reference?  How is this a useful baseline, then?  Are\n      we seeing if it can perform just as well or what the random baseline of\n      temporal reference might be?\n- Sec 3.3\n  - Is $M_{\\ominus^4}$ the max value of any message?  It seems like the\n    temporality is per-message, not per-language.\n- Figures 2a and 3a are difficult to read.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698509852921}, {"id": "aHqYxinlHO", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5303/Reviewer_WtjD"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The current study is the first investigation of the temporal reference in emergent communication, in which its contributions are three-fold:\n- it formally defines an environment for emergent communications in referential grams that consider referent in the past using Linear Temporal Logic. \n- it proposes a new architecture for the agent\n- it conducted a temporality analysis of the agent", "review_text": "The current study is the first investigation of the temporal reference in emergent communication, in which its contributions are three-fold:\n- it formally defines an environment for emergent communications in referential grams that consider referent in the past using Linear Temporal Logic. \n- it proposes a new architecture for the agent\n- it conducted a temporality analysis of the agent", "strengths": "- It is the first work to investigate the temporal references in emergent communication, during which it defines an environment and introduces an architecture of the agent. The subject matter is interesting and vital.\n- Despite the limitations listed below, the experiments to a certain extent provide a response to the major research question of this study and support for the conclusions.", "weaknesses": "1. The notations in this paper need to be further standardised and unified. The current ones sometimes make readers (at least me) hard to follow. Here are the ones that I have spotted: (1) a vector in this paper sometimes means the feature vector of a target and sometimes a target sequence. Both of them can have arbitrary lengths and often appear without further explanation. (2) things like x \\in V are not conditions in equation 2, IMO, they should not connected with the \"real\" conditions with AND. Also, see my questions below. \n2. The paper spends quite a lot of space describing the details of the agents (e.g., the dimension of each vector during the implementation). Contrary to this is the absence of motivations for many choices made in the current study. For example, why the environment was defined in this way? Why a second LSTM in both sender and receiver? Why does there have to be an additional LSTM to capture temporal information? Actually, in terms of the architectural design, the two additional LSTMs in the sender and receiver have very different roles as the two LSTMs in the sender are parallel, and the two in the receiver are stacked. \n3. Related to my second concern, the experiment design in the present study also cannot validate the rationalities of these choices. This makes the conclusions from the experiments, namely a different batching strategy is sufficient, look narrow.\n\n### Typos:\n- Page 3: c \\leq 0.5 -> c < 0.5\n- Page 6: When you are saying Temporal Loss, I believe you meant temporal prediction loss rather than temporal loss.", "questions": "- I do not fully understand the definition of feature space in section 2.1. If F is a feature space of a feature, then why its size N_F is the number of features rather than the number of values that the associated property can take?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The current study is the first investigation of the temporal reference in emergent communication, in which its contributions are three-fold:\n- it formally defines an environment for emergent communications in referential grams that consider referent in the past using Linear Temporal Logic. \n- it proposes a new architecture for the agent\n- it conducted a temporality analysis of the agent", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "- It is the first work to investigate the temporal references in emergent communication, during which it defines an environment and introduces an architecture of the agent. The subject matter is interesting and vital.\n- Despite the limitations listed below, the experiments to a certain extent provide a response to the major research question of this study and support for the conclusions.", "weaknesses": "1. The notations in this paper need to be further standardised and unified. The current ones sometimes make readers (at least me) hard to follow. Here are the ones that I have spotted: (1) a vector in this paper sometimes means the feature vector of a target and sometimes a target sequence. Both of them can have arbitrary lengths and often appear without further explanation. (2) things like x \\in V are not conditions in equation 2, IMO, they should not connected with the \"real\" conditions with AND. Also, see my questions below. \n2. The paper spends quite a lot of space describing the details of the agents (e.g., the dimension of each vector during the implementation). Contrary to this is the absence of motivations for many choices made in the current study. For example, why the environment was defined in this way? Why a second LSTM in both sender and receiver? Why does there have to be an additional LSTM to capture temporal information? Actually, in terms of the architectural design, the two additional LSTMs in the sender and receiver have very different roles as the two LSTMs in the sender are parallel, and the two in the receiver are stacked. \n3. Related to my second concern, the experiment design in the present study also cannot validate the rationalities of these choices. This makes the conclusions from the experiments, namely a different batching strategy is sufficient, look narrow.\n\n### Typos:\n- Page 3: c \\leq 0.5 -> c < 0.5\n- Page 6: When you are saying Temporal Loss, I believe you meant temporal prediction loss rather than temporal loss.", "questions": "- I do not fully understand the definition of feature space in section 2.1. If F is a feature space of a feature, then why its size N_F is the number of features rather than the number of values that the associated property can take?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698499243340}, {"id": "48nXt9J9xv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5303/Reviewer_KopV"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work considers the role of temporal references in emergent communication. For example, just as humans may say, \"the thing I told you five minutes ago,\" EC agents should be able to refer to prior expressions as well.\n\nTo formalize this problem, the authors propose a metric, $M_{\\theta^n}$ for measuring how often a message corresponds to a previous operator, where $n$ specifies how many timesteps previous.\n\nThe authors further propose to increase the use of temporal expressions through changes in architecture and loss. A \"temporal module\" in the sender and receivers model temporal processing, and a temporal loss can explicitly guide the receiver temporal module.\n\nOverall, the authors find that neural architectures with the temporal modules use previous expressions far more often than baseline architectures.", "review_text": "This work considers the role of temporal references in emergent communication. For example, just as humans may say, \"the thing I told you five minutes ago,\" EC agents should be able to refer to prior expressions as well.\n\nTo formalize this problem, the authors propose a metric, $M_{\\theta^n}$ for measuring how often a message corresponds to a previous operator, where $n$ specifies how many timesteps previous.\n\nThe authors further propose to increase the use of temporal expressions through changes in architecture and loss. A \"temporal module\" in the sender and receivers model temporal processing, and a temporal loss can explicitly guide the receiver temporal module.\n\nOverall, the authors find that neural architectures with the temporal modules use previous expressions far more often than baseline architectures.", "strengths": "## Originality\nThis work is pretty original. I am unaware of others considering the problem of temporal references in EC, and I find it an interesting area. Adding further RNN modules to the neural agents to represent temporal relations also seems clever (although I have important questions about the implementation, which I raise later).\n\n## Quality\nThe results from the work are strong, per the desired metrics of temporal references.\n\n## Clarity\nThe writing is largely clear, and the diagrams (both Fig 1 and 5, in the appendix) help explain the architectural changes. I think the authors did a very good job explaining simple aspects of LTL; I have some background in this area, and this seemed like a good presentation of necessary details.\n\n## Significance\nI quite like the combination of temporal logic methods with EC, and I think this is an interesting area for future research. The approach also seems simple (in a good way!), which should aid wider adoption of the ideas.", "weaknesses": "## Clarity around actual approach/implementation\nMy biggest concern with this work is about understanding the actual approach the authors used for the neural implementation. Fundamentally, I do not understand how the authors order inputs over time and the structure of communication. I explain my question about this under \"Questions\" - generally, I view this uncertainty as a weakness given how central such details are to the paper. Certainly, I could be misunderstanding information presented in the paper, but I am very knowledgeable about EC, so I do not think I am missing something obvious.\n\nBecause of my uncertainty about this crucial part of the paper, I am currently leaning against acceptance. I emphasize, however, that if the authors clarify their approach and it seems sound, I am willing to revise my score.\n\n## Clarity around notation\nThe authors dedicate a fair amount of effort to establishing some notation based on LTL, but then do not seem to use it fully. For example, many of the figures (e.g., Figure 2) have almost natural-language y axis labels, whereas I would have expected a single mathematical symbol.\n\n## Clarity of examples\nThe authors included a few simple examples in the paper that I first found very helpful, but now I think they actually are misleading. For example, in Section 3.1, the authors write about an intuitive example where the target sequence is [1, 2, 3, 2, 2, 1] and different message sequences are considered ([1, 2, 3, 2, 4, 4, 1], [1, 2, 3, 4, 4, 4, 1], etc.). My problem with this example is that, when I read it, it made me think that the speakers only generated a single token per target. This was reinforced in Table 4 in Appendix B.1, where the same type of example appears. However, in Section 3.3., where actual results are presented, I realized that a single message was composed of multiple tokens (5 or 6). That is fine, I wish I had realized this earlier on.\n\n## Further metrics?\nThe authors mostly focus on the emergence of temporal references, but further metrics would help situate the broader network performance. Notably, the authors do evaluate validation set accuracy in Figure 6, and they show a small decrease in performance for temporal networks relative to non-temporal networks. This is good analysis but then makes me 1) worry about the utility of this approach, if accuracy goes down and 2) slightly contradicts some of the claims in the main paper, which motivate temporal references to improve performance. \n\n## Minor:\n1. Footnote 1 appears to indicate that the authors will share code upon acceptance. I suggest the authors use https://anonymous.4open.science/ in future submissions to link to anonymized code during the review phase.\n\n2. The figures in Appendix E are not very clear and should be cleaned up for publication. I would also strongly caution about interpreting tSNE plots as indicative of compositionality; do the authors have references supporting this link?\n\n3. The notation around equation 2 is not right. In particular, typically I read the right text in case-based equations as specifying the conditions under which that behavior is selected. I think the cases should just be $c < 0.5$ and $c >= 0.5$, therefore, without any of the $x$ parts. \n\n4. Equation 3 seems to have a few things off about it. First, \"objectRepeated\" seems misleading as a name. I believe it just returns true if $x_n$ is present $h_v$ episodes ago, correct? What does that have to do with repetition? Second, the variable $n$ is very overloaded in the expression, which impedes readbility. It is simulatenously used to represent a fixed number of previous timestep (see left side of the equation) and is used as a counter variable (see right side of the equation).", "questions": "1. My main question, which I hope the authors address in their rebuttal, is about better understanding the exact form of the inputs and outputs of the speaker network. The key idea of this paper is that speakers should be able to refer back to previous communication if they see similar inputs. Thus, the speaker must somehow see a series of inputs, right? How is that represented? Is it a batch of inputs? I'm guessing that is the case, because the temporal module in the speaker reshapes the inputs based on batch size, so that makes me think it allows the temporal module to represent at which point in \"time\" an input appears. If it is the case, however, that the temporal order of inputs is represented by position in the batch, I am somewhat disappointed in the baselines used for comparison. It seems obvious non-temporal networks, which observe inputs in parallel over a batch, cannot represent temporal relations.\n\nOverall, I am unsure about what the authors are actually doing to represent sequentially-ordered inputs, though, and really need clarity on this topic to evaluate this paper.\n\n2. In the main paper, the authors typically present results for $h_v = 4$, but the authors mentioned that they conducted experiments for other. Results for such values are included in Appendix F, but the results end up raising lots of questions for me. Can the authors comment as to why results are seemingly so bimodal? Everything for $h_v <=4$ matches one profile, and everything for $h_v > 4$ matches another. Also, what was the actual training setup for these results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work considers the role of temporal references in emergent communication. For example, just as humans may say, \"the thing I told you five minutes ago,\" EC agents should be able to refer to prior expressions as well.\n\nTo formalize this problem, the authors propose a metric, $M_{\\theta^n}$ for measuring how often a message corresponds to a previous operator, where $n$ specifies how many timesteps previous.\n\nThe authors further propose to increase the use of temporal expressions through changes in architecture and loss. A \"temporal module\" in the sender and receivers model temporal processing, and a temporal loss can explicitly guide the receiver temporal module.\n\nOverall, the authors find that neural architectures with the temporal modules use previous expressions far more often than baseline architectures.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "## Originality\nThis work is pretty original. I am unaware of others considering the problem of temporal references in EC, and I find it an interesting area. Adding further RNN modules to the neural agents to represent temporal relations also seems clever (although I have important questions about the implementation, which I raise later).\n\n## Quality\nThe results from the work are strong, per the desired metrics of temporal references.\n\n## Clarity\nThe writing is largely clear, and the diagrams (both Fig 1 and 5, in the appendix) help explain the architectural changes. I think the authors did a very good job explaining simple aspects of LTL; I have some background in this area, and this seemed like a good presentation of necessary details.\n\n## Significance\nI quite like the combination of temporal logic methods with EC, and I think this is an interesting area for future research. The approach also seems simple (in a good way!), which should aid wider adoption of the ideas.", "weaknesses": "## Clarity around actual approach/implementation\nMy biggest concern with this work is about understanding the actual approach the authors used for the neural implementation. Fundamentally, I do not understand how the authors order inputs over time and the structure of communication. I explain my question about this under \"Questions\" - generally, I view this uncertainty as a weakness given how central such details are to the paper. Certainly, I could be misunderstanding information presented in the paper, but I am very knowledgeable about EC, so I do not think I am missing something obvious.\n\nBecause of my uncertainty about this crucial part of the paper, I am currently leaning against acceptance. I emphasize, however, that if the authors clarify their approach and it seems sound, I am willing to revise my score.\n\n## Clarity around notation\nThe authors dedicate a fair amount of effort to establishing some notation based on LTL, but then do not seem to use it fully. For example, many of the figures (e.g., Figure 2) have almost natural-language y axis labels, whereas I would have expected a single mathematical symbol.\n\n## Clarity of examples\nThe authors included a few simple examples in the paper that I first found very helpful, but now I think they actually are misleading. For example, in Section 3.1, the authors write about an intuitive example where the target sequence is [1, 2, 3, 2, 2, 1] and different message sequences are considered ([1, 2, 3, 2, 4, 4, 1], [1, 2, 3, 4, 4, 4, 1], etc.). My problem with this example is that, when I read it, it made me think that the speakers only generated a single token per target. This was reinforced in Table 4 in Appendix B.1, where the same type of example appears. However, in Section 3.3., where actual results are presented, I realized that a single message was composed of multiple tokens (5 or 6). That is fine, I wish I had realized this earlier on.\n\n## Further metrics?\nThe authors mostly focus on the emergence of temporal references, but further metrics would help situate the broader network performance. Notably, the authors do evaluate validation set accuracy in Figure 6, and they show a small decrease in performance for temporal networks relative to non-temporal networks. This is good analysis but then makes me 1) worry about the utility of this approach, if accuracy goes down and 2) slightly contradicts some of the claims in the main paper, which motivate temporal references to improve performance. \n\n## Minor:\n1. Footnote 1 appears to indicate that the authors will share code upon acceptance. I suggest the authors use https://anonymous.4open.science/ in future submissions to link to anonymized code during the review phase.\n\n2. The figures in Appendix E are not very clear and should be cleaned up for publication. I would also strongly caution about interpreting tSNE plots as indicative of compositionality; do the authors have references supporting this link?\n\n3. The notation around equation 2 is not right. In particular, typically I read the right text in case-based equations as specifying the conditions under which that behavior is selected. I think the cases should just be $c < 0.5$ and $c >= 0.5$, therefore, without any of the $x$ parts. \n\n4. Equation 3 seems to have a few things off about it. First, \"objectRepeated\" seems misleading as a name. I believe it just returns true if $x_n$ is present $h_v$ episodes ago, correct? What does that have to do with repetition? Second, the variable $n$ is very overloaded in the expression, which impedes readbility. It is simulatenously used to represent a fixed number of previous timestep (see left side of the equation) and is used as a counter variable (see right side of the equation).", "questions": "1. My main question, which I hope the authors address in their rebuttal, is about better understanding the exact form of the inputs and outputs of the speaker network. The key idea of this paper is that speakers should be able to refer back to previous communication if they see similar inputs. Thus, the speaker must somehow see a series of inputs, right? How is that represented? Is it a batch of inputs? I'm guessing that is the case, because the temporal module in the speaker reshapes the inputs based on batch size, so that makes me think it allows the temporal module to represent at which point in \"time\" an input appears. If it is the case, however, that the temporal order of inputs is represented by position in the batch, I am somewhat disappointed in the baselines used for comparison. It seems obvious non-temporal networks, which observe inputs in parallel over a batch, cannot represent temporal relations.\n\nOverall, I am unsure about what the authors are actually doing to represent sequentially-ordered inputs, though, and really need clarity on this topic to evaluate this paper.\n\n2. In the main paper, the authors typically present results for $h_v = 4$, but the authors mentioned that they conducted experiments for other. Results for such values are included in Appendix F, but the results end up raising lots of questions for me. Can the authors comment as to why results are seemingly so bimodal? Everything for $h_v <=4$ matches one profile, and everything for $h_v > 4$ matches another. Also, what was the actual training setup for these results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698417012334}], "openreview_url": "https://openreview.net/forum?id=CupHThqQl3", "arxiv_id": "2310.06555", "paper_pdf": "papers/CupHThqQl3.pdf", "paper_pdf_sha256": "aee571dc23005cd4dfc81ac8e696e69737334e8af03a998c5c104858f5c85fa0", "paper_pdf_bytes": 1851232, "paper_pdf_source": "openreview", "code_url": "https://github.com/olipinski/TRG", "code_repository": "olipinski/TRG", "code_commit": "39cd5d55d5e35be3f204f447ce7d399f56a53bcd", "code_archive": "repos/CupHThqQl3.zip", "code_archive_sha256": "8f22ac4df9cdeae26c03427036a3c701082ec38c2b6e1ab7eebbfe12c30b0fef", "code_archive_bytes": 102281, "code_file_count": 12, "code_extensions": {".py": 9, ".ipynb": 2, ".sh": 1}, "github_disk_usage_kb": 93, "github_languages": {"Jupyter Notebook": 243580, "Python": 107779, "Shell": 1388}, "github_archived": false, "github_pushed_at": "2024-05-21T09:34:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-temporal-references-in-emergent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EJka_dVXEcr", "year": 2023, "status": "rejected", "title": "TabDDPM: Modelling Tabular Data with Diffusion Models", "authors": ["Akim Kotelnikov", "Dmitry Baranchuk", "Ivan Rubachev", "Artem Babenko"], "authorids": ["~Akim_Kotelnikov1", "~Dmitry_Baranchuk2", "~Ivan_Rubachev1", "~Artem_Babenko1"], "authors_source": "OpenReview API", "abstract": "Denoising diffusion probabilistic models are currently becoming the leading paradigm of generative modeling for many important data modalities. Being the most prevalent in the computer vision community, diffusion models have also recently gained some attention for other domains, including speech, NLP, and graph-like data. In this work, we investigate if the framework of diffusion models can be advantageous for general tabular problems, where datapoints are typically represented by vectors of heterogeneous features. The inherent heterogeneity of tabular data makes it quite challenging for accurate modeling, since the individual features can be of completely different nature, i.e., some of them can be continuous and some of them can be discrete. To address such data types, we introduce TabDDPM --- a diffusion model that can be universally applied to any tabular dataset and handles any types of features. We extensively evaluate TabDDPM on a wide set of benchmarks and demonstrate its superiority over existing GAN/VAE alternatives, which is consistent with the advantage of diffusion models in other fields. Additionally, we show that TabDDPM can be successfully used in privacy-oriented setups, where the original datapoints cannot be shared.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "H1WLj1Wn6sc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1779/Reviewer_qTDQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes TabDDPM, a diffusion model for generic tabular data consisting of both continuous and discrete features. The paper compares TabDDPM to existing VAE and GAN methods, and against a simple SMOTE baseline. ", "review_text": "While the overall motivation of the paper is good, I'm not sure the proposed method does enough to justify itself over the SMOTE baseline (although demonstrating the effectiveness of that baseline is worthwhile in and of itself). I find it hard to recommend acceptance as things stand. ", "strengths": "**Strengths**\nI think the overall motivation for the paper is good. It is often the case that generative models are specialized to particular media, and can be unsuitable out-of-the-box for generic tabular data. \n\nThe experimental results mainly highlight the unreasonable effectiveness of SMOTE as a baseline, beating both the VAE and GAN approaches. This is a useful observation. \n\n**Weaknesses**\nWhile demonstrating the strong SMOTE baseline is useful, it also highlights that the proposed method only matches this baseline. I'm also not sure about the comparison to SMOTE on sensitive data, both in terms of the metric used, and whether fitting generative models on sensitive data is a useful idea. The latter is discussed below, but in terms of the DCR metric, surely this only captures one possible vector of attack for a potential adversary? That is, a comprehensive comparison of privacy preservation in this setting would go beyond a simple proximity check, and instead marginalize over a number of possible attack vectors? \n\nFor a stronger baseline beyond SMOTE, I would have thought the authors would try a model which factorizes the continuous and discrete (non-ordinal) features. That is, train one model on the continuous features (e.g. a standard diffusion model), and then another model for the discrete features conditioned on the continuous features (e.g. a conditional autoregressive model), or vice versa. This would exploit the state-of-the-art generative models for both modalities, and would provide a strong baseline for a model trained on all features jointly to beat. Moreover, the statement 'In contrast, larger DCR values indicate that the generative model can produce something “new” rather than just copies of real data.' seems far too strong in that (i) it argues that the data and model capacity are sufficient to recover something close to the 'true' distribution, and (ii) DCR is a sufficient metric to determine this. \n\nI may be unfamiliar with common practice, but I'm not sure that I buy the arguments regarding sensitive data and privacy. In particular, if data is sensitive enough to not be shared, training classification/regression models etc. on that data is itself a tricky task, potentially introducing issues of bias, fairness, etc., not to mention generic failure cases. To me, adding the extra step of first fitting a generative model to the data, which is then used to create synthetic data for input to the rest of the pipeline, only exacerbates these potential issues. For example, even though diffusion models have seen impressive breakthroughs in image generation, using diffusion models to generate e.g. medical scans for the purpose of enhancing a disease-detection model seems highly questionable. \n\nFinally, when judging the performance of the proposed method, Figure 2 seems to just visualize a selection of the marginal feature distributions. This seems lacking, in that (i) the evaluation is qualitative, and (ii) surely we're primarily interested in judging the model's capacity to model the joint dependencies in the data correctly? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes TabDDPM, a diffusion model for generic tabular data consisting of both continuous and discrete features. The paper compares TabDDPM to existing VAE and GAN methods, and against a simple SMOTE baseline. ", "strength_and_weaknesses": "**Strengths**\nI think the overall motivation for the paper is good. It is often the case that generative models are specialized to particular media, and can be unsuitable out-of-the-box for generic tabular data. \n\nThe experimental results mainly highlight the unreasonable effectiveness of SMOTE as a baseline, beating both the VAE and GAN approaches. This is a useful observation. \n\n**Weaknesses**\nWhile demonstrating the strong SMOTE baseline is useful, it also highlights that the proposed method only matches this baseline. I'm also not sure about the comparison to SMOTE on sensitive data, both in terms of the metric used, and whether fitting generative models on sensitive data is a useful idea. The latter is discussed below, but in terms of the DCR metric, surely this only captures one possible vector of attack for a potential adversary? That is, a comprehensive comparison of privacy preservation in this setting would go beyond a simple proximity check, and instead marginalize over a number of possible attack vectors? \n\nFor a stronger baseline beyond SMOTE, I would have thought the authors would try a model which factorizes the continuous and discrete (non-ordinal) features. That is, train one model on the continuous features (e.g. a standard diffusion model), and then another model for the discrete features conditioned on the continuous features (e.g. a conditional autoregressive model), or vice versa. This would exploit the state-of-the-art generative models for both modalities, and would provide a strong baseline for a model trained on all features jointly to beat. Moreover, the statement 'In contrast, larger DCR values indicate that the generative model can produce something “new” rather than just copies of real data.' seems far too strong in that (i) it argues that the data and model capacity are sufficient to recover something close to the 'true' distribution, and (ii) DCR is a sufficient metric to determine this. \n\nI may be unfamiliar with common practice, but I'm not sure that I buy the arguments regarding sensitive data and privacy. In particular, if data is sensitive enough to not be shared, training classification/regression models etc. on that data is itself a tricky task, potentially introducing issues of bias, fairness, etc., not to mention generic failure cases. To me, adding the extra step of first fitting a generative model to the data, which is then used to create synthetic data for input to the rest of the pipeline, only exacerbates these potential issues. For example, even though diffusion models have seen impressive breakthroughs in image generation, using diffusion models to generate e.g. medical scans for the purpose of enhancing a disease-detection model seems highly questionable. \n\nFinally, when judging the performance of the proposed method, Figure 2 seems to just visualize a selection of the marginal feature distributions. This seems lacking, in that (i) the evaluation is qualitative, and (ii) surely we're primarily interested in judging the model's capacity to model the joint dependencies in the data correctly? ", "clarity,_quality,_novelty_and_reproducibility": "**Clarity**\nThe method is fairly cleanly described and presented. The relation to previous work on Gaussian and multinomial diffusion is clear, as well as the comparison to previous work based on VAEs and GANs. \n\n**Originality**\nI'm broadly familiar with the generative modeling literature, but less familiar with the specific problem of generative modeling for tabular data. The proposed method seems to be a novel combination and application of existing diffusion methods for continuous and discrete data. The comparison of previous methods to a strong baseline is also highlighted as novel, and I appreciate that this is included.", "summary_of_the_review": "While the overall motivation of the paper is good, I'm not sure the proposed method does enough to justify itself over the SMOTE baseline (although demonstrating the effectiveness of that baseline is worthwhile in and of itself). I find it hard to recommend acceptance as things stand. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666723301129}, {"id": "X-VbS59qZ9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1779/Reviewer_x66q"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper focuses on tabular problems. It proposes a diffusion model called TabDDPM to model tabular data. The TabDDPM can be universally applied to any tabular dataset and handles any type of feature. The experimental results show that the TabDDPM outperforms several SOTA models, which is evaluated on several datasets.", "review_text": "Tabular problem is an important task, and it is quite interesting to use the diffusion model to model tabular data. While the compared models (e.g., SMOTE) are not strong enough. The improvement of TabDDPM is obvious in Figure 3 and Figure 4, but not obvious in Table 3 and Table 4.", "strengths": "Strength:\n(1) It is interesting to apply the diffusion model to other tasks except for computer vision. \n(2) The experimental results look good. Several widely used datasets are used, and the TabDDPM achieves the best results for most datasets.\n\nWeaknesses:\n(1) I have two main concerns:\n  A. Can those models in Table 3 and Table 4 (the SMOTE, CTABGAN, and CTABGAN+ models) represent the SOTA results on all datasets? Actually, SMOTE is published in 2002.\n  B. The recent task tends to have a large data size (i.e., FB dataset in this paper), while TabDDPM does not perform well on this dataset.\n(2) Minor:\n  A. The bold numbers of the last column in Table 4 are weird (DI and WI). ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper focuses on tabular problems. It proposes a diffusion model called TabDDPM to model tabular data. The TabDDPM can be universally applied to any tabular dataset and handles any type of feature. The experimental results show that the TabDDPM outperforms several SOTA models, which is evaluated on several datasets.", "strength_and_weaknesses": "Strength:\n(1) It is interesting to apply the diffusion model to other tasks except for computer vision. \n(2) The experimental results look good. Several widely used datasets are used, and the TabDDPM achieves the best results for most datasets.\n\nWeaknesses:\n(1) I have two main concerns:\n  A. Can those models in Table 3 and Table 4 (the SMOTE, CTABGAN, and CTABGAN+ models) represent the SOTA results on all datasets? Actually, SMOTE is published in 2002.\n  B. The recent task tends to have a large data size (i.e., FB dataset in this paper), while TabDDPM does not perform well on this dataset.\n(2) Minor:\n  A. The bold numbers of the last column in Table 4 are weird (DI and WI). ", "clarity,_quality,_novelty_and_reproducibility": "Clarity: 5/10, \nQuality: 4/10, \nNovelty: 4/10,\nReproducibility 7/10.", "summary_of_the_review": "Tabular problem is an important task, and it is quite interesting to use the diffusion model to model tabular data. While the compared models (e.g., SMOTE) are not strong enough. The improvement of TabDDPM is obvious in Figure 3 and Figure 4, but not obvious in Table 3 and Table 4.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666684060114}, {"id": "QLwiWyhM79", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1779/Reviewer_d279"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The authors propose to combine multimodal and gaussian diffusion models as a new diffusion model that can handle small datasets with mixed kinds of features. ", "review_text": "The paper combines two available diffusion models but the usefulness of the method for tabular data is wrong. ", "strengths": "The paper is fairly well structured, however, it includes a few wrong claims. For example, it describes tabular data as data that includes different feature types, however, this is not necessarily a definition of tabular data. The features of tabular data can be only from a single family. In another part, the authors stated \"Unlike unstructured images or natural texts, tabular data is typically structured,\" which is totally wrong. We have structure in image and NLP and the CNN is using this feature as a biased into the architecture. On the other hand, as stated in many tabular data papers, including SubTab [NeurIPS 2021] and VIME [NeurIPS 2020], and etc, the main problem of tabular data is that it does not have structure or the structured is not obvious as other domains. \n\nApart from that, the proposed method is not novel and the experiments are not convincing. \n\nThe number of features is too small in all experiments. The experiments are not showing how the generated samples are diverse.\n\nOne can also normalize features and then use diffusion models. This is not included in the baselines. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors propose to combine multimodal and gaussian diffusion models as a new diffusion model that can handle small datasets with mixed kinds of features. ", "strength_and_weaknesses": "The paper is fairly well structured, however, it includes a few wrong claims. For example, it describes tabular data as data that includes different feature types, however, this is not necessarily a definition of tabular data. The features of tabular data can be only from a single family. In another part, the authors stated \"Unlike unstructured images or natural texts, tabular data is typically structured,\" which is totally wrong. We have structure in image and NLP and the CNN is using this feature as a biased into the architecture. On the other hand, as stated in many tabular data papers, including SubTab [NeurIPS 2021] and VIME [NeurIPS 2020], and etc, the main problem of tabular data is that it does not have structure or the structured is not obvious as other domains. \n\nApart from that, the proposed method is not novel and the experiments are not convincing. \n\nThe number of features is too small in all experiments. The experiments are not showing how the generated samples are diverse.\n\nOne can also normalize features and then use diffusion models. This is not included in the baselines. ", "clarity,_quality,_novelty_and_reproducibility": "Not only the paper is not novel but also the claims are wrong. ", "summary_of_the_review": "The paper combines two available diffusion models but the usefulness of the method for tabular data is wrong. ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject"}, "tcdate": 1666677171071}, {"id": "5A1QLhL_oq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1779/Reviewer_8b5K"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The authors propose TabDDPM, a denoising diffusion model for tabular data modeling and synthesis.  Within this framework, categorical and binary features are modeled using multinomial diffusion, and, completely independently, numerical features are modeled using Gaussian diffusion.  The model is compared to several recent deep generative models on the task of \"Train on fake, test on real (TFTR)\" [1] across different datasets considering the average of five classification/regression algorithms and, separately, Catboost.\n\n[1] Jordon, James, Jinsung Yoon, and Mihaela Van Der Schaar. \"PATE-GAN: Generating synthetic data with differential privacy guarantees.\" International conference on learning representations. 2018.", "review_text": "The proposed method has critical design flaws.  The evaluation and comparison to SOTA tabular data synthesized also has significant room for improvement.  I recommend rejection.", "strengths": "Strengths\n------------\n-The use of diffusion models for tabular data modeling is interesting\n\nWeaknesses\n---------------\n-The proposed method suffers from one major flaw: \"TabDDPM uses the multinomial diffusion to model the categorical and binary features, and the Gaussian diffusion to model the numerical ones.\" <- This ignores correlations between categorical and numerical variables.  Consider a categorical feature, \"Name,\" and a numerical feature, \"Name_length.\"  Your model could not learn, or even approximate, this deterministic relationship.  Corresponding synthesized data will also be incorrect.\n\n-\"Each categorical feature is handled by a separate forward diffusion process, i.e., the noise components for all features\nare sampled independently.\" <- Same criticism as above\n\n-The authors state the utility of their method for privacy.  But the two aforementioned deficiencies render synthetic data incorrect, e.g., synthetic data may disrespect even obvious correlations.\n\n-Why is baseline SMOTE used?  Other variants, like SVM SMOTE, are reported to work much better.\n\n-Only TVAE is compared to.  However, CTGAN should also be considered,\nas it's source is open (it is the package as TVAE, the CTGAN package)\nand is the main synthesizer (TVAE was developed as a competitor in the\nCTGAN paper).\n\n-The paper lacks a large amount of important details.  E.g.: 1) \"Smote... we generalize it\" <- How? (2) Details for utilized software, both implementation (SMOTE) and version (sklearn), are lacking from the paper.\n\n-\"we evaluate ML efficiency only with respect\nto the CatBoost model (Prokhorenkova et al., 2018), which is the leading GBDT implementation\nproviding state-of-the-art performance on tabular tasks\" <- Please say\n\"arguably is the leading...,\" as XGBoost is more widely used and also\nprovides SOTA GBDT performance.\n\n-No discussion on the runtime (training and synthesis) of TabDDPM.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors propose TabDDPM, a denoising diffusion model for tabular data modeling and synthesis.  Within this framework, categorical and binary features are modeled using multinomial diffusion, and, completely independently, numerical features are modeled using Gaussian diffusion.  The model is compared to several recent deep generative models on the task of \"Train on fake, test on real (TFTR)\" [1] across different datasets considering the average of five classification/regression algorithms and, separately, Catboost.\n\n[1] Jordon, James, Jinsung Yoon, and Mihaela Van Der Schaar. \"PATE-GAN: Generating synthetic data with differential privacy guarantees.\" International conference on learning representations. 2018.", "strength_and_weaknesses": "Strengths\n------------\n-The use of diffusion models for tabular data modeling is interesting\n\nWeaknesses\n---------------\n-The proposed method suffers from one major flaw: \"TabDDPM uses the multinomial diffusion to model the categorical and binary features, and the Gaussian diffusion to model the numerical ones.\" <- This ignores correlations between categorical and numerical variables.  Consider a categorical feature, \"Name,\" and a numerical feature, \"Name_length.\"  Your model could not learn, or even approximate, this deterministic relationship.  Corresponding synthesized data will also be incorrect.\n\n-\"Each categorical feature is handled by a separate forward diffusion process, i.e., the noise components for all features\nare sampled independently.\" <- Same criticism as above\n\n-The authors state the utility of their method for privacy.  But the two aforementioned deficiencies render synthetic data incorrect, e.g., synthetic data may disrespect even obvious correlations.\n\n-Why is baseline SMOTE used?  Other variants, like SVM SMOTE, are reported to work much better.\n\n-Only TVAE is compared to.  However, CTGAN should also be considered,\nas it's source is open (it is the package as TVAE, the CTGAN package)\nand is the main synthesizer (TVAE was developed as a competitor in the\nCTGAN paper).\n\n-The paper lacks a large amount of important details.  E.g.: 1) \"Smote... we generalize it\" <- How? (2) Details for utilized software, both implementation (SMOTE) and version (sklearn), are lacking from the paper.\n\n-\"we evaluate ML efficiency only with respect\nto the CatBoost model (Prokhorenkova et al., 2018), which is the leading GBDT implementation\nproviding state-of-the-art performance on tabular tasks\" <- Please say\n\"arguably is the leading...,\" as XGBoost is more widely used and also\nprovides SOTA GBDT performance.\n\n-No discussion on the runtime (training and synthesis) of TabDDPM.", "clarity,_quality,_novelty_and_reproducibility": "Clarity\n-------\nSeveral important details are lacking from the paper.  The paper is also unclear at times (e.g., for Table 4, why are multiple column entries highlighted in bold?) and inconsistent (TVAE and CTABGAN should be included for comparison in Table 5).\n\nBoth the abstract and the intro could also more clearly state what the goal of TabDDPM is, i.e., it is generate synthetic tabular data.  This point is vague until well into the paper.\n\nQuality\n--------\nThe quality of the paper has significant room for improvement.  Firstly, by including CTGAN (which is widely regarded as the SOTA GAN tabular data synthesizer) and higher performing versions of SMOTE as benchmark competitors.  Secondly, the model itself lacks the ability to capture correlations between categorical and numerical features; this should be directly addressed.\n\nNovelty\n---------\nAs far as the reviewer is aware, the use of diffusion models for tabular data synthesis is novel.\n\nReproducibility\n------------------\nThe authors have uploaded all relevant code to an anonymous repository.", "summary_of_the_review": "The proposed method has critical design flaws.  The evaluation and comparison to SOTA tabular data synthesized also has significant room for improvement.  I recommend rejection.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666676720991}], "openreview_url": "https://openreview.net/forum?id=EJka_dVXEcr", "arxiv_id": "2209.15421", "paper_pdf": "papers/EJka_dVXEcr.pdf", "paper_pdf_sha256": "3ee9d8d61be221ec9a504830143bb43e473548a25ac954725f7028c132375764", "paper_pdf_bytes": 512026, "paper_pdf_source": "openreview", "code_url": "https://github.com/yandex-research/tab-ddpm", "code_repository": "yandex-research/tab-ddpm", "code_commit": "b476257dd460b778ba09eb97f7a51d6490fa17f8", "code_archive": "repos/EJka_dVXEcr.zip", "code_archive_sha256": "2508743d42ec590f5eee24f83bd7fa7ce5183d6adc5280594e226b9fb124f8b4", "code_archive_bytes": 436897, "code_file_count": 73, "code_extensions": {".py": 72, ".ipynb": 1}, "github_disk_usage_kb": 187, "github_languages": {"Python": 504054, "Jupyter Notebook": 9799, "Makefile": 6617}, "github_archived": false, "github_pushed_at": "2024-07-13T04:02:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tabddpm-modelling-tabular-data-with-diffusion"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lgOylcEZQgr", "year": 2022, "status": "rejected", "title": "Online Unsupervised Learning of Visual Representations and Categories", "authors": ["Mengye Ren", "Tyler R. Scott", "Michael Louis Iuzzolino", "Michael Curtis Mozer", "Richard Zemel"], "authorids": ["~Mengye_Ren1", "~Tyler_R._Scott1", "~Michael_Louis_Iuzzolino1", "~Michael_Curtis_Mozer1", "~Richard_Zemel1"], "authors_source": "OpenReview API", "abstract": "Real world learning scenarios involve a nonstationary distribution of classes with sequential dependencies among the samples, in contrast to the standard machine learning formulation of drawing samples independently from a fixed, typically uniform distribution. Furthermore, real world interactions demand learning on-the-fly from few or no class labels. In this work, we propose an unsupervised model that simultaneously performs online visual representation learning and few-shot learning of new categories without relying on any class labels. Our model is a prototype-based memory network with a control component that determines when to form a new class prototype. We formulate it as an online Gaussian mixture model, where components are created online with only a single new example, and assignments do not have to be balanced, which permits an approximation to natural imbalanced distributions from uncurated raw data. Learning includes a contrastive loss that encourages different views of the same image to be assigned to the same prototype. The result is a mechanism that forms categorical representations of objects in nonstationary environments. Experiments show that our method can learn from an online stream of visual input data and is significantly better at category recognition compared to state-of-the-art self-supervised learning methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "350tDj4IWI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper679/Reviewer_bgm4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper addresses the problem of doing unsupervised learning of visual representations, including clustering these representations to infer the existence of new categories. The paper focuses on more \"un-curated\" settings in which the data comes from samples of an environment that a real agent is passing through. This gives object distributions that are very different from typical, curated data sets, like ImageNet without labels. The authors compare their method with other popular self-supervised learning techniques like SimCLR.", "review_text": "First of all, I am very enthusiastic about the general topic that is being addressed: unsupervised (or self-supervised) learning in more realistic environments--specifically those in which we may see a large number of examples of a small set of objects, and perhaps a few examples of some other classes. Thus, the topic is highly relevant and important at this time. \n\nDespite trying my best to understand the specifics of the experimental settings, I found it very difficult to tell exactly what was going on. I found several aspects of the paper difficult to follow. One major issue was trying to understand the exact setting of the RoamingRooms data set task. \n\nReferring to the experimental results from figure 4, here are a few comments:\n\t\n- Only when I finally looked at the title of the figure in figure 4 was I able to discern some of the parameters of the experiments on the RoamingRoom data set such as the 5000 images that retrieval was done across.\n\t\n- I could not ascertain whether the bounding box of the object was given at test time, or only at training time. If it is provided at test time, this makes the task much much easier. Either way, it is a critical detail, and should be described clearly and unambiguously.\n\n- How is the query image related to the “5000 images” which I assume you are\nsearching? How close can they be in time? Could it actually be at the exact same time?  If it is just a frame or two away from the current frame, we would expect the similarity of the same object across these close views in time to be very very similar.  If this is the case, then I would expect to see other types of baselines that leverage this kind of information, such as simple nearest neighbor methods. In other words, if I give a query, just return\nthe top-9 nearest neighbors using some pre-trained representation, centered on \nthe provided bounding box.  If such an approach is not reasonable, it is hard to\ntell from the current text. If it is reasonable, then it would be useful to compare \nagainst it to get a sense of what exactly is being learned. \n\n- I couldn't even tell if SimCLR was trained on the same data, or trained on some other data. If it is trained on the current data, could you give details about how this training is done? It is possible that if I had read all of the referenced papers I would understand these details, but a reviewer should not have to wade back through papers to understand these details. \n\n- \"we use our online clustering procedure to readout from these learned representations\". Why not just do nearest neighbor retrieval using the learned representations, and see how that works?\n\nSimCLR and other such “generic” representation-learning approaches are solving\na significantly different problem: they are typically trying to learn representations which\nhave a high degree of generalizability so that they can be the backbone for a system\ntrained to recognize general categories, with very high variability. Here, since the goal\nis to learn to recognize the exact same object, it makes more sense to take a prototype\nstyle approach, since this is very similar to say, k-nearest neighbors. \n\nWhat I’m trying to say is perhaps that SimCLR and other such methods are not the\nright baselines. A more reasonable baseline is one that just looks at distances to the\ntraining sets. \n\n\nThe authors may feel that I have completely misunderstood the paper. That may be true, but it is not for lack of trying. I have worked with many of the cited self-supervised methods, and have published in self-supervised learning as well. I found a general lack of detail about the exact settings made it very hard to evaluate what was going on.\n\n\nOther issues:\n\n1) “we make two changes on the model inference procedure defined above.”\nWhy did the authors describe the changes to the model in the Experiments section?\nIsn’t the right place to discuss this in Section 3, the section where the model and\nmethods are described? This makes it sound like a last second change….Or, perhaps\nlike it was a difficult-to-justify method that would be difficult to explain. What is the\nmotivation for putting it in experiments?\n\n2) adjusted MI. While I follow the argument for using a max_alpha in the adjusted mutual\ninformation, I’m not sure I agree with it.  For one, this is not something one can do in practice; therefore it gives an over-optimistic view of the output. Second, it is not clear if this procedure favors the current algorithm. By knowing that you are going to optimize over the number of clusters, one can adjust one’s algorithm to take advantage of this. I do not claim to understand whether this procedure favors your algorithm, but in general, such a procedure is not “neutral” to two different algorithms.\n\n3) The function g() is introduced with no definition.\n\n4) How is K the number of classes chosen?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper addresses the problem of doing unsupervised learning of visual representations, including clustering these representations to infer the existence of new categories. The paper focuses on more \"un-curated\" settings in which the data comes from samples of an environment that a real agent is passing through. This gives object distributions that are very different from typical, curated data sets, like ImageNet without labels. The authors compare their method with other popular self-supervised learning techniques like SimCLR.", "main_review": "First of all, I am very enthusiastic about the general topic that is being addressed: unsupervised (or self-supervised) learning in more realistic environments--specifically those in which we may see a large number of examples of a small set of objects, and perhaps a few examples of some other classes. Thus, the topic is highly relevant and important at this time. \n\nDespite trying my best to understand the specifics of the experimental settings, I found it very difficult to tell exactly what was going on. I found several aspects of the paper difficult to follow. One major issue was trying to understand the exact setting of the RoamingRooms data set task. \n\nReferring to the experimental results from figure 4, here are a few comments:\n\t\n- Only when I finally looked at the title of the figure in figure 4 was I able to discern some of the parameters of the experiments on the RoamingRoom data set such as the 5000 images that retrieval was done across.\n\t\n- I could not ascertain whether the bounding box of the object was given at test time, or only at training time. If it is provided at test time, this makes the task much much easier. Either way, it is a critical detail, and should be described clearly and unambiguously.\n\n- How is the query image related to the “5000 images” which I assume you are\nsearching? How close can they be in time? Could it actually be at the exact same time?  If it is just a frame or two away from the current frame, we would expect the similarity of the same object across these close views in time to be very very similar.  If this is the case, then I would expect to see other types of baselines that leverage this kind of information, such as simple nearest neighbor methods. In other words, if I give a query, just return\nthe top-9 nearest neighbors using some pre-trained representation, centered on \nthe provided bounding box.  If such an approach is not reasonable, it is hard to\ntell from the current text. If it is reasonable, then it would be useful to compare \nagainst it to get a sense of what exactly is being learned. \n\n- I couldn't even tell if SimCLR was trained on the same data, or trained on some other data. If it is trained on the current data, could you give details about how this training is done? It is possible that if I had read all of the referenced papers I would understand these details, but a reviewer should not have to wade back through papers to understand these details. \n\n- \"we use our online clustering procedure to readout from these learned representations\". Why not just do nearest neighbor retrieval using the learned representations, and see how that works?\n\nSimCLR and other such “generic” representation-learning approaches are solving\na significantly different problem: they are typically trying to learn representations which\nhave a high degree of generalizability so that they can be the backbone for a system\ntrained to recognize general categories, with very high variability. Here, since the goal\nis to learn to recognize the exact same object, it makes more sense to take a prototype\nstyle approach, since this is very similar to say, k-nearest neighbors. \n\nWhat I’m trying to say is perhaps that SimCLR and other such methods are not the\nright baselines. A more reasonable baseline is one that just looks at distances to the\ntraining sets. \n\n\nThe authors may feel that I have completely misunderstood the paper. That may be true, but it is not for lack of trying. I have worked with many of the cited self-supervised methods, and have published in self-supervised learning as well. I found a general lack of detail about the exact settings made it very hard to evaluate what was going on.\n\n\nOther issues:\n\n1) “we make two changes on the model inference procedure defined above.”\nWhy did the authors describe the changes to the model in the Experiments section?\nIsn’t the right place to discuss this in Section 3, the section where the model and\nmethods are described? This makes it sound like a last second change….Or, perhaps\nlike it was a difficult-to-justify method that would be difficult to explain. What is the\nmotivation for putting it in experiments?\n\n2) adjusted MI. While I follow the argument for using a max_alpha in the adjusted mutual\ninformation, I’m not sure I agree with it.  For one, this is not something one can do in practice; therefore it gives an over-optimistic view of the output. Second, it is not clear if this procedure favors the current algorithm. By knowing that you are going to optimize over the number of clusters, one can adjust one’s algorithm to take advantage of this. I do not claim to understand whether this procedure favors your algorithm, but in general, such a procedure is not “neutral” to two different algorithms.\n\n3) The function g() is introduced with no definition.\n\n4) How is K the number of classes chosen?\n\n", "summary_of_the_review": "Despite being enthusiastic about this general area, I found it too difficult to follow the experiments. Not enough detail was given about:\n- data sets used and how they were sampled to produce \"training\" and test data\n- exactly what the experimental setting was. What problem, exactly, are these systems trying to solve? I think, but am not sure, that given many instances of objects with unique ids at \"training time\", they are simply trying to retrieve objects with the same ID as a query image at test time. However, not enough details are given about the details of this problem.\n- Given the nature of the problem, it appears as though a variety of nearest-neighbor style baselines would be appropriate.\n\n\n\nIn my assessment about \"confidence\", I have said that I am fairly confident of my review. Ironically, I didn't understand the paper well, but this is the problem with the paper. It should be easier to understand what was done. Thus, I am confident in my assessment that the paper should have been more understandable. In particular, I am not someone coming from far outside the area; thus, the paper should have laid out the details more clearly so that someone with my background would understand it.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636392654961}, {"id": "qhflwo46Euf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper679/Reviewer_LjvY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work studies an online version of self-supervised representation learning. It uses Gaussian mixture  of constant isotropic variance as a prototype memory as well as a uniform distribution to handle new unknown classes. The overall distribution evolves using an online EM algorithm that supports adding and removing prototypes. Using this memory, representation learning follows standard unsupervised losses, primalrily a distillation loss over predicted prototype assignment probabilities.", "review_text": "Why is the standard deviation $\\sigma$ constant and not learned per component (cluster) of the mixture? Since $\\sigma$ is a scalar (isotropic), this wouldn't add much complexity to the model and it would help in deciding when to add or remove prototypes, potentially getting rid of (or learning) hyperparameters $u_0$ and $\\mu$.\n\nThe uniform prior distribution for a new cluster is an improper distribution over continuous space. In fact, the overall distribution is a Gaussian-uniform mixture (GUM) (Lathuilière et al. 2018, \"DeepGUM: Learning Deep Robust Regression with a Gaussian-Uniform Mixture Model\"). In the original GUM model, the prior probability of an inlier is learned like all mixing coefficients (weights), whereas here $z_0$ is a fixed hyperparameter and $\\alpha$ is yet another unnecessary threshold hyperparameter (the new cluster probability could be simply compared to existing cluster probabilities).\n\nThe RHS of (2) is incorrect. Softmax is mapping a vector to a vector, while here only a scalar is shown (the numerator of the LHS). One should map the whole vector through softmax, then take the $k$-th element.\n\nEq. (5) is unclear unless one reads the Appendices. Overall, section 3.1 is written as a summary of sections A.1 and A.2, which is hard to follow. For example, $y_{t,k}$ and $u_{t,k}$ in (9,10) are not defined. One should provide a clear motivation for the model and all choices (which is missing), definitions all quantities (some of which are missing) and clear pointers to Appendices for all missing steps in the derivation. The formulation is overall very similar to (Ren et al., 2021), so the derivation is not new (e.g. compare (3) of Ren et al. with (5) of this work.) The connection is missing, both in section 3.1 and in A.1, A.2.\n\nWhat does inverse variance have to do with the count of examples per cluster? Why estimate variance since variance is assumed constant? This remains unclear even after reading section A.2. Shouldn't symbol $\\sigma_{t,d}$ (e.g. in (7,8) be $\\sigma_{t,k}$?\n\nWhat is really missing is citing and discussing Bottou and Bengio 1995, \"Convergence Properties of the K-Means Algorithms,\" which defines an online $k$-means algorithm by means of gradient descent on an error function. One would expect a generalization to a Gaussian mixture basically by preserving the same formulation and adding mixing coefficients and variances per component (cluster), but this is not done. Bottou and Bengio have a similar definition of count of examples per cluster (Eq. (7)), but which has nothing to do with variance.\n\nWhy use decay (hyperparameter $\\rho$)? Well, there is the argument of nonstationary environments but how about forgetting? The entire problem of incremental and online learning is to learn without forgetting. Here, not only nothing is done to avoid forgetting; decay is actually explicit forgetting. Is this justified by the unsupervised setting? Or is there another reason?\n\n\"Popping out\" the prototype with the least weight is too simplistic, because it does not consider interaction or redundancy among clusters. An little populated, isolated cluster may be more important than a highly populated cluster with many neighboring/overlapping clusters. Avrithis and Kalantidis 2012, \"Approximate Gaussian Mixtures for Large Scale Vocabularies\" develop an offline but dynamic (in terms of components) EM algorithm that removes mixture components based on pairwise interaction. Tao et al. 2020, \"Few-Shot Class-Incremental Learning\" also uses a graph over prototypes and has a similar online update of prototypes (Eq. (3)).\n\nIn Eq. (13), it is not clear if target labels are soft, as is standard in distillation. They should be, since there is a temperature parameter, but (13) should have a sum over $k$ that is missing. Similarly, (14) should have a sum over $k$ if this expression is meant to be an entropy. Eq. (15) is too abstract and unclear.\n\nIn computing AMI, is the number of classes known? What is $T$? Sweeping over hyperparameter $\\alpha$ is not realistic, especially if $\\alpha$ is unnecessary (see above).\n\n\"Since none of [SimCLR, SwaV] are designed to output classes with a few examples, we use our online clustering procedure to readout from these learned representations:\" This is unclear. As a result, the protocol of comparisons to offline methods it is unclear. A formal description is in order.\n\nThe entire argument against methods assuming iid batch sampling is reliance on iid for contrastive learning (SimCLR) or prototypes (SwaV). How about self-supervised representation learning without prototypes (e.g. BYOL: Grill et al., 2020) and without negatives (e.g. SimSiam: Chen & He, 2021)? Whouldn't such simple alternatives be much easier to adapt to online? Wouldn't those online versions be more suitable as baselines to compare with?\n\nThe following works should be discussed. Although I believe they are not directly comparable, one may still consider additional baselines, e.g. self-supervised offline pretraining followed by online learning (supervised or not):\n1. Zhu et al. 2021, \"Prototype Augmentation and Self-Supervision for Incremental Learning\"\n2. Gallardo et al. 2021, \"Self-Supervised Training Enhances Online Continual Learning\"\n3. Zhang et al. 2021, \"Self-Supervised Learning Aided Class-Incremental Lifelong Learning\"\n4. Cha et al. 2021, \"Co$^2$L: Contrastive Continual Learning\"\n\nThe list of hyperparameters (e.g. Tables 4-6) is daunting. One should at least separate the method-specific hyperparameters (e.g. $K, \\rho, \\alpha$) from standard hyperparameters (like backbone and learning rate). From Tables 8-14, there are at least 7 hyperparameters in the ablation. What are the default values of remaining hyperparameters when studying each one? Is this search optimal, what is the cost, and to what extent could same values be used across datasets? Merging Tables 4-6 into one would help compare.\n\nThe networks used are too small, more inline with few-shot learning and unlike most work on either self=supervised representation learning or continual learning.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work studies an online version of self-supervised representation learning. It uses Gaussian mixture  of constant isotropic variance as a prototype memory as well as a uniform distribution to handle new unknown classes. The overall distribution evolves using an online EM algorithm that supports adding and removing prototypes. Using this memory, representation learning follows standard unsupervised losses, primalrily a distillation loss over predicted prototype assignment probabilities.", "main_review": "Why is the standard deviation $\\sigma$ constant and not learned per component (cluster) of the mixture? Since $\\sigma$ is a scalar (isotropic), this wouldn't add much complexity to the model and it would help in deciding when to add or remove prototypes, potentially getting rid of (or learning) hyperparameters $u_0$ and $\\mu$.\n\nThe uniform prior distribution for a new cluster is an improper distribution over continuous space. In fact, the overall distribution is a Gaussian-uniform mixture (GUM) (Lathuilière et al. 2018, \"DeepGUM: Learning Deep Robust Regression with a Gaussian-Uniform Mixture Model\"). In the original GUM model, the prior probability of an inlier is learned like all mixing coefficients (weights), whereas here $z_0$ is a fixed hyperparameter and $\\alpha$ is yet another unnecessary threshold hyperparameter (the new cluster probability could be simply compared to existing cluster probabilities).\n\nThe RHS of (2) is incorrect. Softmax is mapping a vector to a vector, while here only a scalar is shown (the numerator of the LHS). One should map the whole vector through softmax, then take the $k$-th element.\n\nEq. (5) is unclear unless one reads the Appendices. Overall, section 3.1 is written as a summary of sections A.1 and A.2, which is hard to follow. For example, $y_{t,k}$ and $u_{t,k}$ in (9,10) are not defined. One should provide a clear motivation for the model and all choices (which is missing), definitions all quantities (some of which are missing) and clear pointers to Appendices for all missing steps in the derivation. The formulation is overall very similar to (Ren et al., 2021), so the derivation is not new (e.g. compare (3) of Ren et al. with (5) of this work.) The connection is missing, both in section 3.1 and in A.1, A.2.\n\nWhat does inverse variance have to do with the count of examples per cluster? Why estimate variance since variance is assumed constant? This remains unclear even after reading section A.2. Shouldn't symbol $\\sigma_{t,d}$ (e.g. in (7,8) be $\\sigma_{t,k}$?\n\nWhat is really missing is citing and discussing Bottou and Bengio 1995, \"Convergence Properties of the K-Means Algorithms,\" which defines an online $k$-means algorithm by means of gradient descent on an error function. One would expect a generalization to a Gaussian mixture basically by preserving the same formulation and adding mixing coefficients and variances per component (cluster), but this is not done. Bottou and Bengio have a similar definition of count of examples per cluster (Eq. (7)), but which has nothing to do with variance.\n\nWhy use decay (hyperparameter $\\rho$)? Well, there is the argument of nonstationary environments but how about forgetting? The entire problem of incremental and online learning is to learn without forgetting. Here, not only nothing is done to avoid forgetting; decay is actually explicit forgetting. Is this justified by the unsupervised setting? Or is there another reason?\n\n\"Popping out\" the prototype with the least weight is too simplistic, because it does not consider interaction or redundancy among clusters. An little populated, isolated cluster may be more important than a highly populated cluster with many neighboring/overlapping clusters. Avrithis and Kalantidis 2012, \"Approximate Gaussian Mixtures for Large Scale Vocabularies\" develop an offline but dynamic (in terms of components) EM algorithm that removes mixture components based on pairwise interaction. Tao et al. 2020, \"Few-Shot Class-Incremental Learning\" also uses a graph over prototypes and has a similar online update of prototypes (Eq. (3)).\n\nIn Eq. (13), it is not clear if target labels are soft, as is standard in distillation. They should be, since there is a temperature parameter, but (13) should have a sum over $k$ that is missing. Similarly, (14) should have a sum over $k$ if this expression is meant to be an entropy. Eq. (15) is too abstract and unclear.\n\nIn computing AMI, is the number of classes known? What is $T$? Sweeping over hyperparameter $\\alpha$ is not realistic, especially if $\\alpha$ is unnecessary (see above).\n\n\"Since none of [SimCLR, SwaV] are designed to output classes with a few examples, we use our online clustering procedure to readout from these learned representations:\" This is unclear. As a result, the protocol of comparisons to offline methods it is unclear. A formal description is in order.\n\nThe entire argument against methods assuming iid batch sampling is reliance on iid for contrastive learning (SimCLR) or prototypes (SwaV). How about self-supervised representation learning without prototypes (e.g. BYOL: Grill et al., 2020) and without negatives (e.g. SimSiam: Chen & He, 2021)? Whouldn't such simple alternatives be much easier to adapt to online? Wouldn't those online versions be more suitable as baselines to compare with?\n\nThe following works should be discussed. Although I believe they are not directly comparable, one may still consider additional baselines, e.g. self-supervised offline pretraining followed by online learning (supervised or not):\n1. Zhu et al. 2021, \"Prototype Augmentation and Self-Supervision for Incremental Learning\"\n2. Gallardo et al. 2021, \"Self-Supervised Training Enhances Online Continual Learning\"\n3. Zhang et al. 2021, \"Self-Supervised Learning Aided Class-Incremental Lifelong Learning\"\n4. Cha et al. 2021, \"Co$^2$L: Contrastive Continual Learning\"\n\nThe list of hyperparameters (e.g. Tables 4-6) is daunting. One should at least separate the method-specific hyperparameters (e.g. $K, \\rho, \\alpha$) from standard hyperparameters (like backbone and learning rate). From Tables 8-14, there are at least 7 hyperparameters in the ablation. What are the default values of remaining hyperparameters when studying each one? Is this search optimal, what is the cost, and to what extent could same values be used across datasets? Merging Tables 4-6 into one would help compare.\n\nThe networks used are too small, more inline with few-shot learning and unlike most work on either self=supervised representation learning or continual learning.\n", "summary_of_the_review": "Online self-supervised learning is a very interesting and realistic setting. The online EM algorithm that maintains a set of prototypes makes a lot of sense.\n\nThere are a number of papers that are very related, most notably Bottou and Bengio 1995 (online $k$-means) and Lathuilière et al. 2018 (Gaussian-uniform mixture). Not only one would expect them to be cited and discussed, but rather the method formulation should be based on the formulation of this prior work. This would improve positioning of this work, make the formulation more elegant/better motivated and dispense the need for certain hyperparameters. The related work section needs improvement in general.\n\nWriting is good in general, but the formulation in unclear at several points, relying on Appendices. Pointers to prior work are also missing in the formulation, primarily (Ren et al., 2021).\n\nThe protocol of comparisons to non-online methods is unclear. The choice of competitors is not well justified: there are simpler methods that are easier to adapt to online. There may be more baselines to consider e.g. on pretraining. There are too many hyperparameters. The networks used are too small.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635947203629}, {"id": "NBiEi25P7nH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper679/Reviewer_hYzM"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper claims that in real world the data distribution is nonstationary, which is different from the current standard machine learning formulation. To solve this problem, the authors design an online unsupervised learning algorithm with Gaussian mixture model and EM algorithm. The experiment shows that the method can learn the online stream of visual input data.", "review_text": "Part of the successes of deep learning comes from the well collected dataset, e.g., ImageNet10K. To study some area, someone should first to collect enough high quality data. However, this is different from the cognitive process of human being. This paper questions the standard machine learning formulation of drawing i.i.d. samples and claims that the real world scenario should be nonstationary. This is very interesting and encourging.\n\nHowever, my main concern is that should we train our models in an online stream scenario just because the real world scenario nonstationary? You see, the tranditional learning formulation has its assumption that all samples are drawn in an i.i.d. way. But it doesn't violate the eventual performance. It works quite well.\n\nI could understand that the author tries to mimic the learning process of human kind, just like lifelong learning. But the authors are encouraged to concentrate the necessity of this behaviour.\n\nBesides, the language, the equations, are all above the thresold of the acceptance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper claims that in real world the data distribution is nonstationary, which is different from the current standard machine learning formulation. To solve this problem, the authors design an online unsupervised learning algorithm with Gaussian mixture model and EM algorithm. The experiment shows that the method can learn the online stream of visual input data.", "main_review": "Part of the successes of deep learning comes from the well collected dataset, e.g., ImageNet10K. To study some area, someone should first to collect enough high quality data. However, this is different from the cognitive process of human being. This paper questions the standard machine learning formulation of drawing i.i.d. samples and claims that the real world scenario should be nonstationary. This is very interesting and encourging.\n\nHowever, my main concern is that should we train our models in an online stream scenario just because the real world scenario nonstationary? You see, the tranditional learning formulation has its assumption that all samples are drawn in an i.i.d. way. But it doesn't violate the eventual performance. It works quite well.\n\nI could understand that the author tries to mimic the learning process of human kind, just like lifelong learning. But the authors are encouraged to concentrate the necessity of this behaviour.\n\nBesides, the language, the equations, are all above the thresold of the acceptance.", "summary_of_the_review": "In this paper, the author assumes a new scenario of nonstationary distribution. To solve the unsupervised and online setting, they design a novel online unsupervised prototypical networks (OUPN). By introducting the Gaussian mixture model and EM, the method could add or drop the cluster in an online way.\n\nThe idea is interesting and encouring, but the necessity is not well claimed. The difference between this method and life-long learning / incremental learning is encouraged. Also, the hyper-parameter of maximum number of K clusters should pay much more attention.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "The paper seems to be ok in terms of ethics concerns.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635925123337}, {"id": "6Cd3CyJiTBt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper679/Reviewer_cvrN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a method for unsupervised learning of instance-based clustering which is more robust to data imbalance and non-iid distributions than existing approaches, such as SwAW. In particular, they propose an Expectation Maximization algorithm that operates in temporal episodes (supposed to correspond to an agent moving through an environment).  At every step an observation (image with an overplayed instance segmentation mask) is first encoded with a CNN and the resulting vector is assigned to one of the existing clusters based on the distance in the feature space, or a new cluster is created (E-step). This is formulated in a probabilistic framework. The cluster prototypes are then updated accordingly (M-step). The loss, computed after a sequence of such steps, consists of a standard contrastive loss (encourages assigning similar images to the same clusters) + entropy (encouraging confident assignments) + a term encouraging the creation of new clusters. \n\nThe method is mainly evaluated on RoamingRooms - a dataset with an agent moving through indoor environments with instance labels assigned to objects. The task is then to cluster different view points of the same object together. The main challenge is that the distribution of examples is non-iid, since in every episode mostly the same objects are seen. Additional evaluation on online variants of Omniglot and ImageNet is also provided.\n\nWhen compared to offline and online contrastive learning algorithms the proposed approach demonstrates stronger performance especially with low batch size and non-iid distribution of examples.", "review_text": "The proposed approach is novel to the best of my knowledge. It also seems sound. The experimental evaluation on the task of unsupervised, instance-based clustering in the online, non-iid setting clearly demonstrates the superiority of the proposed method compared to the baselines. However, I have two main concerns with this paper.\n\nFirstly, although it is well written, the presentation is quite dense and many details have to be inferred from the context. I understand that it's hard to pack so much content into 9 pages, but I would argue that the authors could make significantly more effort to better explain the results. The content could be densified by putting some equations in the text, for example, and the whole Omniglot evaluation can be moved to the supplementary material, since it's not that important for the story.\n\nIn fact, the main issue with the presentation is that no ablation analysis is provided. There are several tables in the supplementary material which compare the values of the many hyper parameters in the method, but the accompanying text literally just takes 9 lines. This is not sufficient to understand what actually makes this relatively complex method work better than the baselines. After all, the goal of a research  paper in not to report strong numbers on a benchmark, but to provided some knowledge about the problem being studied. The ablation analysis needs to be moved to the main text and the discussion needs to be expanded to 1-2 pages with clear conclusions about the importance of various components of the approach.\n\nSecondly, and more importantly, the claims in the introduction are not supported by the experimental results. The proposed approach learns a representation that clusters different views of the same instance in an online manner (character classes are also more akin to instances than semantic categories), and this is the task on which it is evaluated (aside from Table 3). But is this an actual task that many people care about? I would argue that it's not, and the authors agree with me, since all the claims are made in terms of semantic classification/categorization. But instance ids are not categories. In fact, its natural that a method trained with a contrastive learning objective will cluster different views of the same object together, but does it learn a semantic representation? Or is the learned representation at least useful for downstream semantic tasks? The only evidence for that provided in the paper are in Table 3 with experiments on an online version of ImageNet, but the explanation of these results is not sufficient and no qualitative examples of discovered clusters are provided. To address these issues, the authors should either demonstrate that their method is capable of discovering categories, or show that the learned representation is superior to the baselines on downstream semantic tasks (supervised image classification, object detection, semantic segmentation) or remove any claims on learning category information, and focus on the instance classification story, providing some justification for its importance. \n\nFinally, it would be interesting to see which batch size is sufficient for the baselines to reach the performance of the proposed approach both in iid and non-iid setting, extrapolating the curves in Figures 3 and 6.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a method for unsupervised learning of instance-based clustering which is more robust to data imbalance and non-iid distributions than existing approaches, such as SwAW. In particular, they propose an Expectation Maximization algorithm that operates in temporal episodes (supposed to correspond to an agent moving through an environment).  At every step an observation (image with an overplayed instance segmentation mask) is first encoded with a CNN and the resulting vector is assigned to one of the existing clusters based on the distance in the feature space, or a new cluster is created (E-step). This is formulated in a probabilistic framework. The cluster prototypes are then updated accordingly (M-step). The loss, computed after a sequence of such steps, consists of a standard contrastive loss (encourages assigning similar images to the same clusters) + entropy (encouraging confident assignments) + a term encouraging the creation of new clusters. \n\nThe method is mainly evaluated on RoamingRooms - a dataset with an agent moving through indoor environments with instance labels assigned to objects. The task is then to cluster different view points of the same object together. The main challenge is that the distribution of examples is non-iid, since in every episode mostly the same objects are seen. Additional evaluation on online variants of Omniglot and ImageNet is also provided.\n\nWhen compared to offline and online contrastive learning algorithms the proposed approach demonstrates stronger performance especially with low batch size and non-iid distribution of examples.", "main_review": "The proposed approach is novel to the best of my knowledge. It also seems sound. The experimental evaluation on the task of unsupervised, instance-based clustering in the online, non-iid setting clearly demonstrates the superiority of the proposed method compared to the baselines. However, I have two main concerns with this paper.\n\nFirstly, although it is well written, the presentation is quite dense and many details have to be inferred from the context. I understand that it's hard to pack so much content into 9 pages, but I would argue that the authors could make significantly more effort to better explain the results. The content could be densified by putting some equations in the text, for example, and the whole Omniglot evaluation can be moved to the supplementary material, since it's not that important for the story.\n\nIn fact, the main issue with the presentation is that no ablation analysis is provided. There are several tables in the supplementary material which compare the values of the many hyper parameters in the method, but the accompanying text literally just takes 9 lines. This is not sufficient to understand what actually makes this relatively complex method work better than the baselines. After all, the goal of a research  paper in not to report strong numbers on a benchmark, but to provided some knowledge about the problem being studied. The ablation analysis needs to be moved to the main text and the discussion needs to be expanded to 1-2 pages with clear conclusions about the importance of various components of the approach.\n\nSecondly, and more importantly, the claims in the introduction are not supported by the experimental results. The proposed approach learns a representation that clusters different views of the same instance in an online manner (character classes are also more akin to instances than semantic categories), and this is the task on which it is evaluated (aside from Table 3). But is this an actual task that many people care about? I would argue that it's not, and the authors agree with me, since all the claims are made in terms of semantic classification/categorization. But instance ids are not categories. In fact, its natural that a method trained with a contrastive learning objective will cluster different views of the same object together, but does it learn a semantic representation? Or is the learned representation at least useful for downstream semantic tasks? The only evidence for that provided in the paper are in Table 3 with experiments on an online version of ImageNet, but the explanation of these results is not sufficient and no qualitative examples of discovered clusters are provided. To address these issues, the authors should either demonstrate that their method is capable of discovering categories, or show that the learned representation is superior to the baselines on downstream semantic tasks (supervised image classification, object detection, semantic segmentation) or remove any claims on learning category information, and focus on the instance classification story, providing some justification for its importance. \n\nFinally, it would be interesting to see which batch size is sufficient for the baselines to reach the performance of the proposed approach both in iid and non-iid setting, extrapolating the curves in Figures 3 and 6.", "summary_of_the_review": "The proposed approach is original and clearly shows strong results for the task of unsupervised, online, non-iid instance classification. However, there are major issues with the presentation, and the unsupervised categorization claims made in the introduction are not supported by the experimental results. If the authors can address both concerns in the rebuttal, I'd be happy to recommend an acceptance.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635551325496}], "openreview_url": "https://openreview.net/forum?id=lgOylcEZQgr", "arxiv_id": "2109.05675", "paper_pdf": "papers/lgOylcEZQgr.pdf", "paper_pdf_sha256": "7b223dc096107b56655e324fbf60babe884d0e02c855e23a32dd6da05433b6b2", "paper_pdf_bytes": 8423587, "paper_pdf_source": "openreview", "code_url": "https://github.com/renmengye/online-unsup-proto-net", "code_repository": "renmengye/online-unsup-proto-net", "code_commit": "67a14718fe5e00015d58694dcaa21a45bf5faf62", "code_archive": "repos/lgOylcEZQgr.zip", "code_archive_sha256": "58c459b4e90c5fd3b227d1a4aa6e09729f58eb76f2165d0ef5893c19bab740ba", "code_archive_bytes": 489318, "code_file_count": 227, "code_extensions": {".py": 210, ".sh": 17}, "github_disk_usage_kb": 350, "github_languages": {"Python": 1203470, "Shell": 9432, "Roff": 4498, "Dockerfile": 2835, "Makefile": 136}, "github_archived": false, "github_pushed_at": "2022-10-01T05:01:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/online-unsupervised-learning-of-visual"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cQzf26aA3vM", "year": 2021, "status": "rejected", "title": "Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization", "authors": ["Brandon Trabucco", "Aviral Kumar", "Xinyang Geng", "Sergey Levine"], "authorids": ["~Brandon_Trabucco1", "~Aviral_Kumar2", "~Xinyang_Geng1", "~Sergey_Levine1"], "authors_source": "OpenReview API", "abstract": "Black-box model-based optimization (MBO) problems, where the goal is to find a design input that maximizes an unknown objective function, are ubiquitous in a wide range of domains, such as the design of drugs, aircraft, and robot morphology. Typically, such problems are solved by actively querying the black-box objective on design proposals and using the resulting feedback to improve the proposed designs. However, when the true objective function is expensive or dangerous to evaluate in the real world, we might instead prefer a method that can optimize this function using only previously collected data, for example from a set of previously conducted experiments. This data-driven offline MBO set- ting presents a number of unique challenges, but a number of recent works have demonstrated that viable offline MBO methods can be developed even for high- dimensional problems, using high-capacity deep neural network function approximators. Unfortunately, the lack of standardized evaluation tasks in this emerg- ing new field has made tracking progress and comparing recent methods difficult. To address this problem, we present Design-Bench, a benchmark suite of offline MBO tasks with a unified evaluation protocol and reference implementations of recent methods. Our benchmark suite includes diverse and realistic tasks derived from real-world problems in biology, material science, and robotics that present distinct challenges for offline MBO methods. Our benchmarks, together with the reference implementations, are available at sites.google.com/view/design-bench. We hope that our benchmark can serve as a meaningful metric for the progress of offline MBO methods and guide future algorithmic development.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "PWRnhaLgG4C", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3076/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": " ##########################################################################\n\nSummary:\n \nThis paper proposes a benchmark suite of offline model-based optimization problems. This benchmark includes diverse and realistic tasks derived from real-world problems in biology, material science, and robotics contains a wide variety of domains, and it covers both continuous and discrete, low and high dimensional design spaces. The authors provide a comprehensive evaluation of existing methods under identical assumptions and get several interesting takeaways from\nthe results. They found there exists surprising efficacy of simple baselines such as naive gradient ascent, which suggests the need for careful tuning and standardization of methods in this area. \n\n##########################################################################\n\nPros: \n \n1. This paper tackles a valuable problem of benchmarking model-based optimization approaches. It will provide some insights in future algorithmic development\n \n2. The paper is well written. They present the significance of model-based optimization, and clearly describe the problems, challenges and considerations of model-based optimization tasks. They conduct further analysis and discussions on the experimental results and find several interesting takeaways.\n \n##########################################################################\n\nCons: \n \n1. Contribution is limited. There are no new ideas proposed and no significant findings revealed. Maybe the authors can look deeper into the simple takeaways and move forward to get more insights and draw some generalized conclusions.\n\n2. In continuous control tasks such as HopperController, I think we cannot get a truly policy without trajectory data. I do not believe the datapairs of the weights of a neural network controller and the corresponding return values can be directly used to learn useful knowledge that contributes to calculate a reasonable policy. It is hard to get a generalizable function that maps from weights of a controller network to resulting values.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": " ##########################################################################\n\nSummary:\n \nThis paper proposes a benchmark suite of offline model-based optimization problems. This benchmark includes diverse and realistic tasks derived from real-world problems in biology, material science, and robotics contains a wide variety of domains, and it covers both continuous and discrete, low and high dimensional design spaces. The authors provide a comprehensive evaluation of existing methods under identical assumptions and get several interesting takeaways from\nthe results. They found there exists surprising efficacy of simple baselines such as naive gradient ascent, which suggests the need for careful tuning and standardization of methods in this area. \n\n##########################################################################\n\nPros: \n \n1. This paper tackles a valuable problem of benchmarking model-based optimization approaches. It will provide some insights in future algorithmic development\n \n2. The paper is well written. They present the significance of model-based optimization, and clearly describe the problems, challenges and considerations of model-based optimization tasks. They conduct further analysis and discussions on the experimental results and find several interesting takeaways.\n \n##########################################################################\n\nCons: \n \n1. Contribution is limited. There are no new ideas proposed and no significant findings revealed. Maybe the authors can look deeper into the simple takeaways and move forward to get more insights and draw some generalized conclusions.\n\n2. In continuous control tasks such as HopperController, I think we cannot get a truly policy without trajectory data. I do not believe the datapairs of the weights of a neural network controller and the corresponding return values can be directly used to learn useful knowledge that contributes to calculate a reasonable policy. It is hard to get a generalizable function that maps from weights of a controller network to resulting values.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603986863095}, {"id": "todhQSW7DU1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3076/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I liked the paper overall. The motivation of the paper is clear, and given that offline ML based optimization is beginning to take traction, this is a good time to set up benchmarks and evaluation metrics. The variety of domains considered, with careful consideration of complexity, are good characteristics of the benchmark.\n\nThe simple baseline considered makes sense and challenges the community to develop algorithms that can generalize to different problem characteristics.\n\nA few suggestions for improving the paper:\n- The hyper-parameters for each algorithm need not be the same for each task. It is possible that tuning the hyper-parameters for particular tasks improves the performance of the algorithms.\n- As the paper points out, one cannot evaluate these algorithm at scale as they require real world experimentation. An important aspect that is not discussed in the paper is offline evaluation of the optimization algorithms. In a practical application, we need to know the efficacy of an algorithm for a particular task before they are deployed on the real task.\n- None of the tasks considered have constraints on the problem. This is especially challenging for model based methods in continuous design space. \n- The algorithms implemented only consider model based methods. The more popular traditional methods such as genetic algorithms and mixed integer programming would be good to compare against. ML methods will only be adopted if they can beat existing established methods.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A meaningful benchmark for offline optimization", "review": "I liked the paper overall. The motivation of the paper is clear, and given that offline ML based optimization is beginning to take traction, this is a good time to set up benchmarks and evaluation metrics. The variety of domains considered, with careful consideration of complexity, are good characteristics of the benchmark.\n\nThe simple baseline considered makes sense and challenges the community to develop algorithms that can generalize to different problem characteristics.\n\nA few suggestions for improving the paper:\n- The hyper-parameters for each algorithm need not be the same for each task. It is possible that tuning the hyper-parameters for particular tasks improves the performance of the algorithms.\n- As the paper points out, one cannot evaluate these algorithm at scale as they require real world experimentation. An important aspect that is not discussed in the paper is offline evaluation of the optimization algorithms. In a practical application, we need to know the efficacy of an algorithm for a particular task before they are deployed on the real task.\n- None of the tasks considered have constraints on the problem. This is especially challenging for model based methods in continuous design space. \n- The algorithms implemented only consider model based methods. The more popular traditional methods such as genetic algorithms and mixed integer programming would be good to compare against. ML methods will only be adopted if they can beat existing established methods.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603946886465}, {"id": "7rrdJAu4Bdn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3076/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \nThis paper focuses on model-based black box optimization problems in the offline setting. These are settings where access to ground truth is expensive, and instead the optimizer has access only to a trained model of the ground truth based on limited data. While optimizing on this surrogate space, a good optimizer often needs to account for model uncertainty and accuracy degradation. The main aim of the paper is to provide a test bed for algorithms that try to solve this challenge. \n\nPositives: \nThis is a well-motivated line of work because there is a large interest in these problems across fields. There is indeed a need for better benchmarks and better libraries to make it easy to compare methods. I think the paper is executed cleanly and the it's well-written. I also think the work, when completed, has potential to be very useful. \n\nAreas for improvement: \n\nIn my view, the paper has shortcomings in it's design, development and scope. A paper like this is helpful when it can:\n\n (i) Establish good practices across the board by streamlining workflow, and ensure the interface is used the same way when comparing methods.\n\n(ii) Contribute code that make it easy and fast to use and develop on.\n\n(iii)  Make it easy to collect and report relevant performance statistics the same way across algorithms, helping the field by making it easy to benchmark.\n\n(iv) Includes challenges that are diverse but relevant to the use case of the algorithm. \n\n(v) Synthesize a suite of methods from the literature that are distinct from each other and of interest to the community as strong benchmarks. \n\nI don't think the paper is addressing these aspects sufficiently. Some detailed comments below. \n\n(1) The categorical choice of benchmark tasks is not clearly justified (e.g. Why should anyone care if a protein design algorithm is poor at designing a robot controller?) At the end of the day, these problems share little structure, and \"no free lunch\" arguments (Wolpert 95) would suggest that there is no one good algorithm for every challenge. The real world settings for these problems don't map well to each other. This paper as designed, could result in follow up work with meaningless comparisons between algorithms that have no business being compared (unless there is a meaningful connection between the challenges). If the authors feel like they can justify this particular set of challenges, I'm open to be convinced. E.g. if the argument is that there will be a \"master algorithm\" that is just good at designing everything, and this benchmarking set is designed to enable that, I could drop this point, but then other issues will be relevant. \n\n(2) The particular choices within each class of tasks is insufficiently justified. Why is GFP a good design challenge given that the ground truth is also necessarily a trained oracle with likely poor performance outside the training data? Just because some previous papers chose this as the design task doesn't make it a good benchmark. No statistics is provided for how good the GP model for GFP is.  As for other proteins, GB1 for instance is far more completely surveyed than GFP.  There are also better published models of GFP available (e.g. TAPE, UniRep). All of these would of course struggle with out-of-domain samples. Why not use a physical simulator like Rosetta that wouldn't have this issue? I believe when designing a benchmarking suite, these decisions should be considered more carefully, as it would quickly become the test bed for follow up work and a flawed design choice here amplifies in future work. \n\n(3) A body of literature for algorithms that can readily perform MBO has been neglected.  It is trivial to run a regular optimization algorithm with the model (instead of ground truth) and compare the proposed solutions to ground truth.  Quality-Diversity/EDA algorithms (e.g. genetic algorithms,  simulated annealing,  CMA-ES, or even pure CEM (rather than DbAS)...) consistently perform \"well\" in these high-dimensional optimization settings.  The success of the gradient-based method gives more reason to believe representatives of each of these classical approaches should be included and suggests that the claim that climbing proxy model will necessarily result in bad \"ground truth\" outcomes is a weak one. \n\n(4) For a benchmarking library like this, there needs to be mature code available for review (not submitted). I've checked the provided website multiple times, and while it is under active development, the code is not accessible. From what I gather the current code interface simply gives access to some data points and a ground truth. This is too little API. A good benchmarking tool would let the user abstract away the modeling part easily, and be able to readily port and run their algorithm against benchmarks, producing the results in the same way. It should also take care of running sanity checks/tests for the user and generating the same plots as those in the paper. \n\n(5) The fact that gradient-based methods outperform other methods presented here is only surprising in the sense that they were not included in the original papers (i.e. why weren't gradient-based methods benchmarked there? not this paper's fault of course).  The authors express  a general conviction where gradient based methods have done very poorly in other attempts for MBO, but provide no references, it would be great to cite relevant references for this claim.  \n\n(6) As is, the paper/library only compares CbAS/DbAS with MINs and hill-climbing methods. I think it is not sufficient breadth of methods to make a \"benchmarking\" suite. As far as I can tell, CbAS/DbAS and MINs are not the best published algorithms in any of the domains suggested, so the authors should justify why they are the algorithms to benchmark against?\n\nFor instance, Angermueller et al 2020 ICLR, have an offline RL algorithm that can in principle solve all of these problems. In fact DynaPPO, PPO, and Ensemble-based Bayesian optimization all outperform CbAS in that study. Some sequence design challenges used there seem to be better benchmarks than GFP.  There is substantive work in molecular design on MBO,  but none of the SOTA algorithms are included (e.g the now-classic Gomez-Bombarelli 2016, or  perhaps adaptation of Zhou et al 2019 Scientific Reports). A good rule of thumb in my view is to include the SOTA or well-established algorithm for each task category. \n\n(6) I suggest the authors think carefully about what the evaluation criteria are. While optimization itself is a good metric, other factors, such as providing a way of evaluating diversity of solutions, or sensitivity to dataset size (e.g. by subsampling), are good to consider.\n\n==========\n\nI would like to encourage the authors to continue the pursuit of this work because it is relevant and well-motivated, and has great potential, in my view it is simply not ready. I think this work needs to be reviewed again when it is more mature,  with wider range of algorithms, better justified challenges, a larger set of metrics that can be easily collected, and available code such that the reviewer can vet the benchmarks and code properly. Right now, it's a comparative review of a small set of methods, not a good benchmarking suite.\n\nIf done well, it can be a very useful suite that can help researchers develop better algorithms. The danger of accepting it prematurely is that it will be a basis for future work that \"game-ify\" studies of algorithms against irrelevant/misleading set of benchmarks. That is only damaging to the development of good algorithms and could misguide research. As it stands,  I find the latter risk higher than it's contribution, and hence I believe it should be rejected and reviewed once more of these structural issues are addressed (and code is available to review).\n\n\n~~~~~~~~~~~~~~\n\nPost review and discussion remarks:\nI think the authors have improved the paper significantly during the review period. However, three of my main concerns about the paper remain to a degree that I'm not confident about the paper's value (or risk of misleading followups). (1) That the set of challenges is somewhat arbitrary, some tasks are using \"real\" ground truths while others are simply running on known trained oracles.  (2) That implementation of strong offline RL benchmark algorithms are missing (because they don't exactly apply across domains) even though they can always be applied in this setting even if \"exact\" conditions are not applied in every case (just like Gradient Ascent or BO were) (3) That the API needs to offer more for this to be a good benchmarking suite. \n\nI've been most concerned about 2 and 3, and after reading the code, I find that it is still too \"bare bones\" to be a good package. I looked back at OpenAI gym, and there are several abstractions that they make, including actions, observations, environments, spaces that help the implementer unify how they deal with the complexity underneath. \n\nSo far as I can tell, most of what design-bench does is load a csv matrix into an task.x, task.score(x) , and also lets the user access some approximate oracle task.y. This are critical to the process but their abstraction as related to the paper are not clear to me at this time. How is task.y computed, how is the ground truth actually representative of reality. How does optimization depend on the choice of oracle for task.y? \n\nHaving read the code, useful elements are in there to make for a good package, I feel like it needs improvements and another review for scientific soundness.\n\nI've updated my score to address the improvements made. The paper scores somewhere between a 4 and a 5 for me.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Valuable if done well, but needs substantial improvement", "review": "Summary: \nThis paper focuses on model-based black box optimization problems in the offline setting. These are settings where access to ground truth is expensive, and instead the optimizer has access only to a trained model of the ground truth based on limited data. While optimizing on this surrogate space, a good optimizer often needs to account for model uncertainty and accuracy degradation. The main aim of the paper is to provide a test bed for algorithms that try to solve this challenge. \n\nPositives: \nThis is a well-motivated line of work because there is a large interest in these problems across fields. There is indeed a need for better benchmarks and better libraries to make it easy to compare methods. I think the paper is executed cleanly and the it's well-written. I also think the work, when completed, has potential to be very useful. \n\nAreas for improvement: \n\nIn my view, the paper has shortcomings in it's design, development and scope. A paper like this is helpful when it can:\n\n (i) Establish good practices across the board by streamlining workflow, and ensure the interface is used the same way when comparing methods.\n\n(ii) Contribute code that make it easy and fast to use and develop on.\n\n(iii)  Make it easy to collect and report relevant performance statistics the same way across algorithms, helping the field by making it easy to benchmark.\n\n(iv) Includes challenges that are diverse but relevant to the use case of the algorithm. \n\n(v) Synthesize a suite of methods from the literature that are distinct from each other and of interest to the community as strong benchmarks. \n\nI don't think the paper is addressing these aspects sufficiently. Some detailed comments below. \n\n(1) The categorical choice of benchmark tasks is not clearly justified (e.g. Why should anyone care if a protein design algorithm is poor at designing a robot controller?) At the end of the day, these problems share little structure, and \"no free lunch\" arguments (Wolpert 95) would suggest that there is no one good algorithm for every challenge. The real world settings for these problems don't map well to each other. This paper as designed, could result in follow up work with meaningless comparisons between algorithms that have no business being compared (unless there is a meaningful connection between the challenges). If the authors feel like they can justify this particular set of challenges, I'm open to be convinced. E.g. if the argument is that there will be a \"master algorithm\" that is just good at designing everything, and this benchmarking set is designed to enable that, I could drop this point, but then other issues will be relevant. \n\n(2) The particular choices within each class of tasks is insufficiently justified. Why is GFP a good design challenge given that the ground truth is also necessarily a trained oracle with likely poor performance outside the training data? Just because some previous papers chose this as the design task doesn't make it a good benchmark. No statistics is provided for how good the GP model for GFP is.  As for other proteins, GB1 for instance is far more completely surveyed than GFP.  There are also better published models of GFP available (e.g. TAPE, UniRep). All of these would of course struggle with out-of-domain samples. Why not use a physical simulator like Rosetta that wouldn't have this issue? I believe when designing a benchmarking suite, these decisions should be considered more carefully, as it would quickly become the test bed for follow up work and a flawed design choice here amplifies in future work. \n\n(3) A body of literature for algorithms that can readily perform MBO has been neglected.  It is trivial to run a regular optimization algorithm with the model (instead of ground truth) and compare the proposed solutions to ground truth.  Quality-Diversity/EDA algorithms (e.g. genetic algorithms,  simulated annealing,  CMA-ES, or even pure CEM (rather than DbAS)...) consistently perform \"well\" in these high-dimensional optimization settings.  The success of the gradient-based method gives more reason to believe representatives of each of these classical approaches should be included and suggests that the claim that climbing proxy model will necessarily result in bad \"ground truth\" outcomes is a weak one. \n\n(4) For a benchmarking library like this, there needs to be mature code available for review (not submitted). I've checked the provided website multiple times, and while it is under active development, the code is not accessible. From what I gather the current code interface simply gives access to some data points and a ground truth. This is too little API. A good benchmarking tool would let the user abstract away the modeling part easily, and be able to readily port and run their algorithm against benchmarks, producing the results in the same way. It should also take care of running sanity checks/tests for the user and generating the same plots as those in the paper. \n\n(5) The fact that gradient-based methods outperform other methods presented here is only surprising in the sense that they were not included in the original papers (i.e. why weren't gradient-based methods benchmarked there? not this paper's fault of course).  The authors express  a general conviction where gradient based methods have done very poorly in other attempts for MBO, but provide no references, it would be great to cite relevant references for this claim.  \n\n(6) As is, the paper/library only compares CbAS/DbAS with MINs and hill-climbing methods. I think it is not sufficient breadth of methods to make a \"benchmarking\" suite. As far as I can tell, CbAS/DbAS and MINs are not the best published algorithms in any of the domains suggested, so the authors should justify why they are the algorithms to benchmark against?\n\nFor instance, Angermueller et al 2020 ICLR, have an offline RL algorithm that can in principle solve all of these problems. In fact DynaPPO, PPO, and Ensemble-based Bayesian optimization all outperform CbAS in that study. Some sequence design challenges used there seem to be better benchmarks than GFP.  There is substantive work in molecular design on MBO,  but none of the SOTA algorithms are included (e.g the now-classic Gomez-Bombarelli 2016, or  perhaps adaptation of Zhou et al 2019 Scientific Reports). A good rule of thumb in my view is to include the SOTA or well-established algorithm for each task category. \n\n(6) I suggest the authors think carefully about what the evaluation criteria are. While optimization itself is a good metric, other factors, such as providing a way of evaluating diversity of solutions, or sensitivity to dataset size (e.g. by subsampling), are good to consider.\n\n==========\n\nI would like to encourage the authors to continue the pursuit of this work because it is relevant and well-motivated, and has great potential, in my view it is simply not ready. I think this work needs to be reviewed again when it is more mature,  with wider range of algorithms, better justified challenges, a larger set of metrics that can be easily collected, and available code such that the reviewer can vet the benchmarks and code properly. Right now, it's a comparative review of a small set of methods, not a good benchmarking suite.\n\nIf done well, it can be a very useful suite that can help researchers develop better algorithms. The danger of accepting it prematurely is that it will be a basis for future work that \"game-ify\" studies of algorithms against irrelevant/misleading set of benchmarks. That is only damaging to the development of good algorithms and could misguide research. As it stands,  I find the latter risk higher than it's contribution, and hence I believe it should be rejected and reviewed once more of these structural issues are addressed (and code is available to review).\n\n\n~~~~~~~~~~~~~~\n\nPost review and discussion remarks:\nI think the authors have improved the paper significantly during the review period. However, three of my main concerns about the paper remain to a degree that I'm not confident about the paper's value (or risk of misleading followups). (1) That the set of challenges is somewhat arbitrary, some tasks are using \"real\" ground truths while others are simply running on known trained oracles.  (2) That implementation of strong offline RL benchmark algorithms are missing (because they don't exactly apply across domains) even though they can always be applied in this setting even if \"exact\" conditions are not applied in every case (just like Gradient Ascent or BO were) (3) That the API needs to offer more for this to be a good benchmarking suite. \n\nI've been most concerned about 2 and 3, and after reading the code, I find that it is still too \"bare bones\" to be a good package. I looked back at OpenAI gym, and there are several abstractions that they make, including actions, observations, environments, spaces that help the implementer unify how they deal with the complexity underneath. \n\nSo far as I can tell, most of what design-bench does is load a csv matrix into an task.x, task.score(x) , and also lets the user access some approximate oracle task.y. This are critical to the process but their abstraction as related to the paper are not clear to me at this time. How is task.y computed, how is the ground truth actually representative of reality. How does optimization depend on the choice of oracle for task.y? \n\nHaving read the code, useful elements are in there to make for a good package, I feel like it needs improvements and another review for scientific soundness.\n\nI've updated my score to address the improvements made. The paper scores somewhere between a 4 and a 5 for me.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603855165150}, {"id": "_Cs3nvyGs5J", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3076/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the evaluation of offline black-box optimization algorithms. The community currently lacks a standardized benchmark to compare the performance of methods. This paper presents a new suite of offline model-based optimization tasks and standardized evaluation procedures for the community. The evaluation criterion for the quality of a benchmark is the realism and diversity of the tasks, with special consideration for high-dimensional design space and the objective function's sensitivity. The paper then evaluates several algorithms on the benchmark. \n\nThe creation of a useful benchmark is an important and challenging task. It seems that consideration was given to the choice of optimization problems. The result is a diverse set, with some representing real-world design optimization problems. However, there are a few areas in which the benchmark and evaluation should be improved. \n\nI do not recommend this paper for acceptance as there is insufficient support as to why these problems should be considered over others or what challenges these environments present for designing new algorithms. There are also some deficiencies in the evaluation protocol.\n\nFor the GFP, Molecule, and superconductor problems, an \"expert model\" is used as the oracle function. The expert model's use speeds up the evaluation and is undoubtedly the right choice in creating a benchmark, but it introduces bias when evaluating an algorithm. The goal of the paper is to present a benchmark for use in the development of novel algorithms. It is important that if others are to use this benchmark to design new algorithms, they should be aware of the benchmark's biases. A more detailed investigation of these specific optimization tasks and what type of algorithms they favored is warranted.  \n\n\nIn the morphology tasks, the data is generated using a policy trained with a given robot morphology and then later evaluated using a different morphology. This seems to bias the optimal choice of morphology to be the one the policy was used during training. Can the authors clarify what is intended to be learned from including these tasks in the benchmark?\n\n\n\nThe evaluation protocol is lacking in two areas: how hyperparameter tuning is considered and how results are compared. \n\nIt is said that no hyperparameter tuning can use the oracle function. This is an obvious necessity, but it remains unclear to what extent algorithms are allowed to perform hyperparameter tuning. This makes it unclear if one algorithm performs better than another due to more hyperparameter tuning. Some recent works (Sivaprasad et al. 2020, Jordan et al. 2020, Dodge et al. 2019) address some of these issues when evaluating algorithms. \n\nIn comparing the results, it is not clear how one determines which algorithm performs best. No aggregate measure is given to determine performance. Furthermore, it is unclear how the uncertainty of the results are quantified. These methods have some degree of stochasticity in the performance, be it from the algorithm or choice hyperparameters. It is unclear what sources of uncertainty are being considered in the results and how many trials have been executed. What do the +/- numbers represent in tables 2 and 3, and how are they computed?\n\nSmall notes for areas of improvement: \n\nIn section 7 study, an ablation study is mentioned for gradient descent, but there is no discussion of how this study was carried out. Gradient descent is known to be sensitive to the scaling and parameterization of the function, which has led to the design of many optimizers that use preconditions on the gradients, e.g., Newton's method, natural gradient, Adagrad, RMSProp, etc. This problem has also been directly considered in finding optimal design points (Box and Draper 2007). \n\nThe sensitivity (smoothness) of the objective function to small perturbations of its inputs is an important characteristic to consider. The provided example in Figure 2 claims that the objective function is highly sensitive, but the provided example does not appear to have high sensitivity but is rather discontinuous. Not to say that this is not a challenging problem, just that high sensitivity is perhaps not the correct interpretation for this example. \n\nIn the MuJoCo environment tasks, the oracle objective function is evaluated 16 times and averaged to reduce noise. How was this number chosen, and how uncertain are the evaluations? It seems unlikely that it can be assumed that with 16 samples, the sample mean is normally distributed. \n\nAdditionally, can the authors clarify why 100 and 1000 timesteps are sufficient for the MuJoCo tasks of interest? As presented, these seem like arbitrary choices.\n\nBox, George EP, and Norman R. Draper. Response surfaces, mixtures, and ridge analyses. Vol. 649. John Wiley & Sons, 2007.\n\nSivaprasad, P. T., Mai, F., Vogels, T., Jaggi, M., & Fleuret, F. (2020). Optimizer benchmarking needs to account for hyperparameter tuning. In Proceedings of the 37th International Conference on Machine Learning.\n\nJordan, S. M., Chandak, Y., Cohen, D., Zhang, M., & Thomas, P. S. (2020). Evaluating the Performance of Reinforcement Learning Algorithms. In Proceedings of the 37th International Conference on Machine Learning.\n\nDodge, J., Gururangan, S., Card, D., Schwartz, R., & Smith, N. A. (2019). Show your work: Improved reporting of experimental results. arXiv preprint arXiv:1909.03004.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good start", "review": "This paper studies the evaluation of offline black-box optimization algorithms. The community currently lacks a standardized benchmark to compare the performance of methods. This paper presents a new suite of offline model-based optimization tasks and standardized evaluation procedures for the community. The evaluation criterion for the quality of a benchmark is the realism and diversity of the tasks, with special consideration for high-dimensional design space and the objective function's sensitivity. The paper then evaluates several algorithms on the benchmark. \n\nThe creation of a useful benchmark is an important and challenging task. It seems that consideration was given to the choice of optimization problems. The result is a diverse set, with some representing real-world design optimization problems. However, there are a few areas in which the benchmark and evaluation should be improved. \n\nI do not recommend this paper for acceptance as there is insufficient support as to why these problems should be considered over others or what challenges these environments present for designing new algorithms. There are also some deficiencies in the evaluation protocol.\n\nFor the GFP, Molecule, and superconductor problems, an \"expert model\" is used as the oracle function. The expert model's use speeds up the evaluation and is undoubtedly the right choice in creating a benchmark, but it introduces bias when evaluating an algorithm. The goal of the paper is to present a benchmark for use in the development of novel algorithms. It is important that if others are to use this benchmark to design new algorithms, they should be aware of the benchmark's biases. A more detailed investigation of these specific optimization tasks and what type of algorithms they favored is warranted.  \n\n\nIn the morphology tasks, the data is generated using a policy trained with a given robot morphology and then later evaluated using a different morphology. This seems to bias the optimal choice of morphology to be the one the policy was used during training. Can the authors clarify what is intended to be learned from including these tasks in the benchmark?\n\n\n\nThe evaluation protocol is lacking in two areas: how hyperparameter tuning is considered and how results are compared. \n\nIt is said that no hyperparameter tuning can use the oracle function. This is an obvious necessity, but it remains unclear to what extent algorithms are allowed to perform hyperparameter tuning. This makes it unclear if one algorithm performs better than another due to more hyperparameter tuning. Some recent works (Sivaprasad et al. 2020, Jordan et al. 2020, Dodge et al. 2019) address some of these issues when evaluating algorithms. \n\nIn comparing the results, it is not clear how one determines which algorithm performs best. No aggregate measure is given to determine performance. Furthermore, it is unclear how the uncertainty of the results are quantified. These methods have some degree of stochasticity in the performance, be it from the algorithm or choice hyperparameters. It is unclear what sources of uncertainty are being considered in the results and how many trials have been executed. What do the +/- numbers represent in tables 2 and 3, and how are they computed?\n\nSmall notes for areas of improvement: \n\nIn section 7 study, an ablation study is mentioned for gradient descent, but there is no discussion of how this study was carried out. Gradient descent is known to be sensitive to the scaling and parameterization of the function, which has led to the design of many optimizers that use preconditions on the gradients, e.g., Newton's method, natural gradient, Adagrad, RMSProp, etc. This problem has also been directly considered in finding optimal design points (Box and Draper 2007). \n\nThe sensitivity (smoothness) of the objective function to small perturbations of its inputs is an important characteristic to consider. The provided example in Figure 2 claims that the objective function is highly sensitive, but the provided example does not appear to have high sensitivity but is rather discontinuous. Not to say that this is not a challenging problem, just that high sensitivity is perhaps not the correct interpretation for this example. \n\nIn the MuJoCo environment tasks, the oracle objective function is evaluated 16 times and averaged to reduce noise. How was this number chosen, and how uncertain are the evaluations? It seems unlikely that it can be assumed that with 16 samples, the sample mean is normally distributed. \n\nAdditionally, can the authors clarify why 100 and 1000 timesteps are sufficient for the MuJoCo tasks of interest? As presented, these seem like arbitrary choices.\n\nBox, George EP, and Norman R. Draper. Response surfaces, mixtures, and ridge analyses. Vol. 649. John Wiley & Sons, 2007.\n\nSivaprasad, P. T., Mai, F., Vogels, T., Jaggi, M., & Fleuret, F. (2020). Optimizer benchmarking needs to account for hyperparameter tuning. In Proceedings of the 37th International Conference on Machine Learning.\n\nJordan, S. M., Chandak, Y., Cohen, D., Zhang, M., & Thomas, P. S. (2020). Evaluating the Performance of Reinforcement Learning Algorithms. In Proceedings of the 37th International Conference on Machine Learning.\n\nDodge, J., Gururangan, S., Card, D., Schwartz, R., & Smith, N. A. (2019). Show your work: Improved reporting of experimental results. arXiv preprint arXiv:1909.03004.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603720214438}], "openreview_url": "https://openreview.net/forum?id=cQzf26aA3vM", "arxiv_id": "2202.08450", "paper_pdf": "papers/cQzf26aA3vM.pdf", "paper_pdf_sha256": "450b763881dc0b9489a9b3578a45b0af980f0ac902b5b7b6a43c719174f8a8ef", "paper_pdf_bytes": 2179947, "paper_pdf_source": "openreview", "code_url": "https://github.com/rail-berkeley/design-bench", "code_repository": "rail-berkeley/design-bench", "code_commit": "95a495359eebbe2280c3ddbfcbdf80e611744fbf", "code_archive": "repos/cQzf26aA3vM.zip", "code_archive_sha256": "83b225ff667f89d1801bdc37695e341f361100776b9c535d0c3543a208c40100", "code_archive_bytes": 175757, "code_file_count": 69, "code_extensions": {".py": 69}, "github_disk_usage_kb": 447, "github_languages": {"Python": 687234}, "github_archived": false, "github_pushed_at": "2022-02-16T07:00:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/design-bench-benchmarks-for-data-driven-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1gzdhEKvH", "year": 2020, "status": "rejected", "title": "Neural Linear Bandits: Overcoming Catastrophic Forgetting through Likelihood Matching", "authors": ["Tom Zahavy", "Shie Mannor"], "authorids": ["tomzahavy@gmail.com", "shiemannor@gmail.com"], "authors_source": "OpenReview API", "abstract": "We study neural-linear bandits for solving problems where both exploration and representation learning play an important role. Neural-linear bandits leverage the representation power of deep neural networks and combine it with efficient exploration mechanisms, designed for linear contextual bandits, on top of the last hidden layer. Since the representation is being optimized during learning, information regarding exploration with \"old\" features is lost. Here, we propose the first limited memory neural-linear bandit that is resilient to this catastrophic forgetting phenomenon. We perform simulations on a variety of real-world problems, including regression, classification, and sentiment analysis, and observe that our algorithm achieves superior performance and shows resilience to catastrophic forgetting. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rklDDmY0tB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper33/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper adapts Bayesian linear regression to the setting of a limited memory replay buffer. The idea is to calibrate the prior mean and variance when the neural representation of context is updated. Overall the paper is well written and explained clearly. Some experiments are provided to show that the proposed method is able to achieve a performance competitive to Bayesian linear regression with infinite memory.\n\nThe result of this paper is interesting. But I am not sure if the current experimental results are convincing enough to justify the significance of the proposed method. \n1. The results in Section 4.2 seem to be following the setting in Riquelme 2018. These datasets are all in a supervised learning setting. It is a bit disappointing that the proposed method is not tested on RL datasets. \n2. No other baseline is provided in the experiments for comparison. \n    a. There are other methods in the literature to overcome catastrophic forgetting of neural networks, e.g regularizing the update of the network. How would that be compared to the proposed method?\n    b. What about other methods, like [1]?\n3. Most of the experiment details are missing. For example, how is the reward defined in section 4.3? What is the overhead in computation in practice, especially for the SDP?\n\n\nOther comments:\n1. Why would solving a SDP require only O(g^{0.5}) in section 3.1?\n2. In the discussion in section 3, even if equation (5) and (6) can be exactly solved, how does the heavy tailed problem mentioned in section 2 been solved?\n\n\n[1] Elmachtoub, Adam N., et al. \"A practical method for solving contextual bandit problems using decision trees.\" arXiv preprint arXiv:1706.04687 (2017).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper adapts Bayesian linear regression to the setting of a limited memory replay buffer. The idea is to calibrate the prior mean and variance when the neural representation of context is updated. Overall the paper is well written and explained clearly. Some experiments are provided to show that the proposed method is able to achieve a performance competitive to Bayesian linear regression with infinite memory.\n\nThe result of this paper is interesting. But I am not sure if the current experimental results are convincing enough to justify the significance of the proposed method. \n1. The results in Section 4.2 seem to be following the setting in Riquelme 2018. These datasets are all in a supervised learning setting. It is a bit disappointing that the proposed method is not tested on RL datasets. \n2. No other baseline is provided in the experiments for comparison. \n    a. There are other methods in the literature to overcome catastrophic forgetting of neural networks, e.g regularizing the update of the network. How would that be compared to the proposed method?\n    b. What about other methods, like [1]?\n3. Most of the experiment details are missing. For example, how is the reward defined in section 4.3? What is the overhead in computation in practice, especially for the SDP?\n\n\nOther comments:\n1. Why would solving a SDP require only O(g^{0.5}) in section 3.1?\n2. In the discussion in section 3, even if equation (5) and (6) can be exactly solved, how does the heavy tailed problem mentioned in section 2 been solved?\n\n\n[1] Elmachtoub, Adam N., et al. \"A practical method for solving contextual bandit problems using decision trees.\" arXiv preprint arXiv:1706.04687 (2017)."}, "tcdate": 1571881822988}, {"id": "HJlg_dy0KB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper33/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a neural linear bandits algorithm that is resilient to catastrophic forgetting when using limited memory.\n\nThe proposed algorithm Alg. 2 is similar to Thompson sampling for linear contextual bandits, Alg. 1, but using the last layer activation vectors as a linear feature, and also a different way of updating noise parameter prior and posterior is used based on Bayesian linear regression Eq. (2).\n\nAlg. 2 also works with limited memory of history data, therefore after every time, the memory is refreshed, likelihood matching is used to calculate new Phi to make the likelihood (mean and variance) of reward estimation the same as it for the old feature. For mean matching, minimizing MSE Eq. (3) is used and for variance, solving PSD problem Eq. (6) is used.\n\nThe complexity of this algorithm is analyzed. And experiments are conducted to show that the proposed method is resilient to catastrophic forgetting and can achieve good cumulative reward results.\n\nThe proposed method is reasonable and the results look promising. However, I found several weak points as follows.\n\n1. As mentioned Bayesian linear regression Eq. (2) is used to update noise prior and posterior, but this update has no theoretical guarantees as mentioned. As an algorithm mainly works under bandit settings, this is kind of undesirable.\n\n2. This algorithm works with neural network-based features, but it is in nature not scalable as shown in the complexity analysis (linear dependence on action number). The linear feature is just replaced by the last layer activation of NNs. From this perspective, the experimental results just justify again that the NN feature is somehow powerful, which is as expected.\n\n3. The likelihood matching can deal with catastrophic forgetting with limited history memory, which looks good. But the fact that it actually works for linear feature (last layer activation) together with realization assumption weakens this contribution a lot. The authors find using Eq. (3) is better than the exact mean matching Eq. (5), and there is no explanation for this, which kind of shows the proposed likelihood matching probably is not a good way when using full NNs rather than just linear features (last layer). On the other hand, the SDP seems also can only work under linear feature settings, and is not promising to be generalized to fully update for NNs.\n\nOverall, this is a reasonable paper. However, on the one hand, as an algorithm mainly works under bandit settings, it is a lack of theoretical support. On the other hand, the linear feature setting weakens the contribution of likelihood matching to deal with catastrophic forgetting with limited memory. There are some questions of the proposed mean matching, and the matching is not able to generalize. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a neural linear bandits algorithm that is resilient to catastrophic forgetting when using limited memory.\n\nThe proposed algorithm Alg. 2 is similar to Thompson sampling for linear contextual bandits, Alg. 1, but using the last layer activation vectors as a linear feature, and also a different way of updating noise parameter prior and posterior is used based on Bayesian linear regression Eq. (2).\n\nAlg. 2 also works with limited memory of history data, therefore after every time, the memory is refreshed, likelihood matching is used to calculate new Phi to make the likelihood (mean and variance) of reward estimation the same as it for the old feature. For mean matching, minimizing MSE Eq. (3) is used and for variance, solving PSD problem Eq. (6) is used.\n\nThe complexity of this algorithm is analyzed. And experiments are conducted to show that the proposed method is resilient to catastrophic forgetting and can achieve good cumulative reward results.\n\nThe proposed method is reasonable and the results look promising. However, I found several weak points as follows.\n\n1. As mentioned Bayesian linear regression Eq. (2) is used to update noise prior and posterior, but this update has no theoretical guarantees as mentioned. As an algorithm mainly works under bandit settings, this is kind of undesirable.\n\n2. This algorithm works with neural network-based features, but it is in nature not scalable as shown in the complexity analysis (linear dependence on action number). The linear feature is just replaced by the last layer activation of NNs. From this perspective, the experimental results just justify again that the NN feature is somehow powerful, which is as expected.\n\n3. The likelihood matching can deal with catastrophic forgetting with limited history memory, which looks good. But the fact that it actually works for linear feature (last layer activation) together with realization assumption weakens this contribution a lot. The authors find using Eq. (3) is better than the exact mean matching Eq. (5), and there is no explanation for this, which kind of shows the proposed likelihood matching probably is not a good way when using full NNs rather than just linear features (last layer). On the other hand, the SDP seems also can only work under linear feature settings, and is not promising to be generalized to fully update for NNs.\n\nOverall, this is a reasonable paper. However, on the one hand, as an algorithm mainly works under bandit settings, it is a lack of theoretical support. On the other hand, the linear feature setting weakens the contribution of likelihood matching to deal with catastrophic forgetting with limited memory. There are some questions of the proposed mean matching, and the matching is not able to generalize. "}, "tcdate": 1571842151970}, {"id": "SJlaTfYptS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper33/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis work provides a memory-efficient nonlinear bandit algorithm based on deep neural networks. More specifically, the algorithm in this work only uses part of history information to save the memory usage. To overcome the catastrophic forgetting problem,  the authors provided novel covariance matrix approximation method. Experiment results also suggest that \n\nPros:\n\nThe writing of this paper is very well. It provides enough introduction of the background of nonlinear bandit problems. The experiment settings and results are convincible. \n\nCons:\n- The core idea lacks solid theoretical supports. There is no regret bound result in this paper. The reason why I think the authors should add such theoretical proof is that it seems that the idea to construct new prior matrix instead of old one to avoid the catastrophic forgetting is not related to deep neural network at all. Thus, given existing regret analysis for Thompson sampling on linear bandit problems, the authors should also provide a simple analysis on linear case to show that the construction of prior matrix is indeed meaningful. \n\n- The experiment part does not show the accuracy the SDP solve needs. As the authors mentioned in Discussion part, below equation 6, it is very crucial to decide the accuracy the SDP solver needs. I suggest the authors add more details about the SDP solver in the experiment part.\n\n\nMinor comments:\n\n- The authors used DNN to minimize equation 3. Have the authors tried  a regularized MSE instead of (3)? I think to add a regularizer can further improve the results. \n- At page 13, below equation 8: why the first equality lacks?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "Summary:\n\nThis work provides a memory-efficient nonlinear bandit algorithm based on deep neural networks. More specifically, the algorithm in this work only uses part of history information to save the memory usage. To overcome the catastrophic forgetting problem,  the authors provided novel covariance matrix approximation method. Experiment results also suggest that \n\nPros:\n\nThe writing of this paper is very well. It provides enough introduction of the background of nonlinear bandit problems. The experiment settings and results are convincible. \n\nCons:\n- The core idea lacks solid theoretical supports. There is no regret bound result in this paper. The reason why I think the authors should add such theoretical proof is that it seems that the idea to construct new prior matrix instead of old one to avoid the catastrophic forgetting is not related to deep neural network at all. Thus, given existing regret analysis for Thompson sampling on linear bandit problems, the authors should also provide a simple analysis on linear case to show that the construction of prior matrix is indeed meaningful. \n\n- The experiment part does not show the accuracy the SDP solve needs. As the authors mentioned in Discussion part, below equation 6, it is very crucial to decide the accuracy the SDP solver needs. I suggest the authors add more details about the SDP solver in the experiment part.\n\n\nMinor comments:\n\n- The authors used DNN to minimize equation 3. Have the authors tried  a regularized MSE instead of (3)? I think to add a regularizer can further improve the results. \n- At page 13, below equation 8: why the first equality lacks?"}, "tcdate": 1571816133034}], "openreview_url": "https://openreview.net/forum?id=r1gzdhEKvH", "arxiv_id": null, "paper_pdf": "papers/r1gzdhEKvH.pdf", "paper_pdf_sha256": "dfaafd1711639999570f040a73517687ef67cefe118bdcf26cce5ece56efbe91", "paper_pdf_bytes": 1775445, "paper_pdf_source": "openreview", "code_url": "https://github.com/anonymousneurips/neurips", "code_repository": "anonymousneurips/neurips", "code_commit": "3637ee7e10c9c83a6c89dacb4d4bfcac72c7175a", "code_archive": "repos/r1gzdhEKvH.zip", "code_archive_sha256": "506d7f77e522fe4ca94718cbd0cbc1b45608513f347d3f2ad473b14be3c7d649", "code_archive_bytes": 516554, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 600, "github_languages": {"Python": 190942}, "github_archived": false, "github_pushed_at": "2019-05-20T07:24:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-linear-bandits-overcoming-catastrophic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kaW0pt8pLR", "year": 2026, "status": "rejected", "title": "Labelling Data with Unknown References", "authors": ["Adrian de Wynter"], "authorids": ["~Adrian_de_Wynter1"], "authors_source": "OpenReview API", "abstract": "An evaluator is trustworthy when there exists some agreed-upon way to measure its performance as a labeller. The two ways to establish trustworthiness are either by testing it, or by assuming the evaluator 'knows' somehow the way to label the corpus. However, if labelled references (e.g., a development set) are unavailable, neither of these approaches work: the former requires data, and the latter is an assumption, not evidence. To address this, we introduce an algorithm (the 'No-Data Algorithm') by which to establish trust in an evaluator without any existing references. Our algorithm works by successively posing challenges to said evaluator. We show that this is sufficient to establish trustworthiness w.h.p., in such a way that when the evaluator actually knows the way to label the corpus, the No-Data Algorithm accepts its output; and, conversely, flags untrustworthy evaluators when these are unable to prove it. We present formal proofs of correctness, empirical tests, and applications to LLMs-as-judges on low-resource languages.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "m4pJxtkHoo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7894/Reviewer_f2Lc"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper addresses a fundamental problem of trusting an evaluator (e.g., an LLM-as-a-judge) without any labelled reference data, which is increasingly common in low-resource settings, domains with expensive annotation, and scenarios where benchmark contamination is suspected. The authors propose the No-Data Algorithm, which interacts with an evaluator through a verifier using a multi-round Evaluator–Verifier (EV) protocol. The evaluator must generate similar datapoints and justifications, while the verifier issues random structural and semantic challenges. Passing both challenges over repeated rounds is provably difficult for dishonest evaluators. The algorithm incorporates probabilistic flipping of labels upon failure and yields a “success rate” metric that reflects evaluator trustworthiness.\nOverall this is a well written paper and I enjoyed reading this paper.", "review_text": "This paper addresses a fundamental problem of trusting an evaluator (e.g., an LLM-as-a-judge) without any labelled reference data, which is increasingly common in low-resource settings, domains with expensive annotation, and scenarios where benchmark contamination is suspected. The authors propose the No-Data Algorithm, which interacts with an evaluator through a verifier using a multi-round Evaluator–Verifier (EV) protocol. The evaluator must generate similar datapoints and justifications, while the verifier issues random structural and semantic challenges. Passing both challenges over repeated rounds is provably difficult for dishonest evaluators. The algorithm incorporates probabilistic flipping of labels upon failure and yields a “success rate” metric that reflects evaluator trustworthiness.\nOverall this is a well written paper and I enjoyed reading this paper.", "strengths": "The following are 3 strong points of this paper:\n1. This paper considers a novel and timely problem. This is one of the core weaknesses of LLM-based evaluation: absence of ground truth. \n\n2. The analytical formal guarantees presented in this paper links deception probability to the number of challenge rounds, offering provable confidence.\n\n3. The West Frisian experiment demonstrates impact beyond toy settings and highlights where evaluator trust matters in real deployments.", "weaknesses": "The following are 3 weak points of this paper:\n1. Semantic similarity assumption is hand-wavy in natural language and the same is acknowledged in this paper.. But the authors did not handle this properly. \n\n2. Authors note that trust in an evaluator becomes trust in a prompt. Without robust prompting strategies, reproducibility may suffer and this undermines the practical reliability of this paper.\n\n3. While binarisation is suggested, the mathematical treatment does not fully generalize to multi-class, hierarchical labels,continuous scoring systems.", "questions": "(a) Please answer the above 3 weak points.\n\n(b) Can you please clarify what similarity means for natural language settings.. while do so, consider embedding-based constraints.\n\n(c) Can you provide an ablation on the prompt design aspects.\n\n(d) How easy or difficult it is to expand your approach beyond binary labels setting? O/w the impact of this work would become limited. \n\n(e) Proving more light on understanding how dishonest evaluators are caught would improve interpretability.. and this enhances the quality of this paper.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses a fundamental problem of trusting an evaluator (e.g., an LLM-as-a-judge) without any labelled reference data, which is increasingly common in low-resource settings, domains with expensive annotation, and scenarios where benchmark contamination is suspected. The authors propose the No-Data Algorithm, which interacts with an evaluator through a verifier using a multi-round Evaluator–Verifier (EV) protocol. The evaluator must generate similar datapoints and justifications, while the verifier issues random structural and semantic challenges. Passing both challenges over repeated rounds is provably difficult for dishonest evaluators. The algorithm incorporates probabilistic flipping of labels upon failure and yields a “success rate” metric that reflects evaluator trustworthiness.\nOverall this is a well written paper and I enjoyed reading this paper.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The following are 3 strong points of this paper:\n1. This paper considers a novel and timely problem. This is one of the core weaknesses of LLM-based evaluation: absence of ground truth. \n\n2. The analytical formal guarantees presented in this paper links deception probability to the number of challenge rounds, offering provable confidence.\n\n3. The West Frisian experiment demonstrates impact beyond toy settings and highlights where evaluator trust matters in real deployments.", "weaknesses": "The following are 3 weak points of this paper:\n1. Semantic similarity assumption is hand-wavy in natural language and the same is acknowledged in this paper.. But the authors did not handle this properly. \n\n2. Authors note that trust in an evaluator becomes trust in a prompt. Without robust prompting strategies, reproducibility may suffer and this undermines the practical reliability of this paper.\n\n3. While binarisation is suggested, the mathematical treatment does not fully generalize to multi-class, hierarchical labels,continuous scoring systems.", "questions": "(a) Please answer the above 3 weak points.\n\n(b) Can you please clarify what similarity means for natural language settings.. while do so, consider embedding-based constraints.\n\n(c) Can you provide an ablation on the prompt design aspects.\n\n(d) How easy or difficult it is to expand your approach beyond binary labels setting? O/w the impact of this work would become limited. \n\n(e) Proving more light on understanding how dishonest evaluators are caught would improve interpretability.. and this enhances the quality of this paper.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762693810508}, {"id": "UrF4Xgivor", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7894/Reviewer_t3tA"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces the \"No-Data Algorithm\" for establishing trustworthiness of evaluators (including LLMs-as-judges) without requiring labeled reference data. The approach uses an Evaluator-Verifier (EV) protocol based on zero-knowledge proof concepts, where evaluators must pass challenges about datapoint similarity to establish credibility. The work provides theoretical correctness bounds and empirical validation on synthetic data and a West Frisian language evaluation task. The authors also introduce a new West Frisian dataset that is annotated by language experts and contains 1,015 items. The approach resembles semi-supervised or unsupervised based learning methods for learning.", "review_text": "This paper introduces the \"No-Data Algorithm\" for establishing trustworthiness of evaluators (including LLMs-as-judges) without requiring labeled reference data. The approach uses an Evaluator-Verifier (EV) protocol based on zero-knowledge proof concepts, where evaluators must pass challenges about datapoint similarity to establish credibility. The work provides theoretical correctness bounds and empirical validation on synthetic data and a West Frisian language evaluation task. The authors also introduce a new West Frisian dataset that is annotated by language experts and contains 1,015 items. The approach resembles semi-supervised or unsupervised based learning methods for learning.", "strengths": "1. The primary contribution, the \"No-Data Algorithm,\" is a novel, formal method for establishing trust without references. Moving beyond simple correlation metrics (which require references) to a proof-based system based on a challenge-response protocol is a commendable direction. The work clearly states when labels are fundamentally unknowable and when the algorithm's purpose is establishing trust rather than producing labels.\n2. Combines a synthetic binary-string setting (clean control of rubrics) with a realistic low-resource language application (West Frisian), reinforcing external validity of the theoretical predictions.The authors also create a new West Frisian dataset that is expert annotated and from a low-resourced language (1,015 items) that is a helpful contribution towards the community. \n3. The findings also can help with the future research in terms of further studying the evaluator in itself since the paper makes a crucial distinction between an evaluator's accuracy and its trustworthiness. The results compellingly show that a deceptive evaluator can achieve high accuracy (e.g., in the OOP case) but will be flagged by a low success rate. This finding that the success rate is the more reliable metric for trust is a key takeaway.", "weaknesses": "1. The approach theoretically is an approach that could be applied in general to any classification task where an LLM judge could be utilized, but the paper has limited score in terms of experimental evaluation. In terms of benchmarking, any classification task could have been used for the paper as a potential task to compare how this approach would do against the human labels. \n2. The most critical weakness is using GPT-4.1 as both evaluator and verifier in the natural language experiments (Section 6.2). This creates a circular validation where the system is essentially checking itself. While the paper acknowledges this creates questions about \"what are the results measuring\" (Section 7.2), this undermines the core claim of establishing independent trustworthiness. The theoretical framework assumes an independent verifier, but the implementation violates this assumption. \n\n3. In terms of baselines for this experiment set is an algorithm that relies on supervised learning, but a strong baseline is not used for comparison in this paper but should be utilized in the next iterations. Either a classifier that is trained on the dataset, in-context learning (few shot prompting).", "questions": "More of suggestions for improvement. \n\n1. Writing and the organizing of the paper needs a bit of work, there should be a set of contributions or research questions set early in the research paper, just to highlight everything in the paper and to guide the narrative. The authors contribute (or at least they’ve created) a new human annotated dataset for Frisian language but this is not highlighted until later in a section. The narrative can be easily followed if there was a clear RQ set. \n2. The generalizability of the synthetic experiment is questionable. The LLM evaluator was tasked to pick a datapoint from a provided list rather than generate a new one, reportedly because generation led to poor performance. This \"picking\" strategy seems to be a significant concession and a deviation from the described protocol. The implications of this methodological shortcut on the validity of the results, and how this relates to the natural language experiment where generation is performed, are not sufficiently discussed.\n3. It would have been helpful if there was a figure to handle the overall narrative, Fig 1 doesn’t capture the entirety of the entire research workflow proposed in the paper. \n4. More on model selection for a task that is low-resourced, it would have been helpful if the impact was compared against a baseline for multi-lingual such as Cohere's Aya model (https://cohere.com/research/aya). Why was the gpt 4.1 model identified over a multi-lingual model that is trained for that language (West Firisian)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the \"No-Data Algorithm\" for establishing trustworthiness of evaluators (including LLMs-as-judges) without requiring labeled reference data. The approach uses an Evaluator-Verifier (EV) protocol based on zero-knowledge proof concepts, where evaluators must pass challenges about datapoint similarity to establish credibility. The work provides theoretical correctness bounds and empirical validation on synthetic data and a West Frisian language evaluation task. The authors also introduce a new West Frisian dataset that is annotated by language experts and contains 1,015 items. The approach resembles semi-supervised or unsupervised based learning methods for learning.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The primary contribution, the \"No-Data Algorithm,\" is a novel, formal method for establishing trust without references. Moving beyond simple correlation metrics (which require references) to a proof-based system based on a challenge-response protocol is a commendable direction. The work clearly states when labels are fundamentally unknowable and when the algorithm's purpose is establishing trust rather than producing labels.\n2. Combines a synthetic binary-string setting (clean control of rubrics) with a realistic low-resource language application (West Frisian), reinforcing external validity of the theoretical predictions.The authors also create a new West Frisian dataset that is expert annotated and from a low-resourced language (1,015 items) that is a helpful contribution towards the community. \n3. The findings also can help with the future research in terms of further studying the evaluator in itself since the paper makes a crucial distinction between an evaluator's accuracy and its trustworthiness. The results compellingly show that a deceptive evaluator can achieve high accuracy (e.g., in the OOP case) but will be flagged by a low success rate. This finding that the success rate is the more reliable metric for trust is a key takeaway.", "weaknesses": "1. The approach theoretically is an approach that could be applied in general to any classification task where an LLM judge could be utilized, but the paper has limited score in terms of experimental evaluation. In terms of benchmarking, any classification task could have been used for the paper as a potential task to compare how this approach would do against the human labels. \n2. The most critical weakness is using GPT-4.1 as both evaluator and verifier in the natural language experiments (Section 6.2). This creates a circular validation where the system is essentially checking itself. While the paper acknowledges this creates questions about \"what are the results measuring\" (Section 7.2), this undermines the core claim of establishing independent trustworthiness. The theoretical framework assumes an independent verifier, but the implementation violates this assumption. \n\n3. In terms of baselines for this experiment set is an algorithm that relies on supervised learning, but a strong baseline is not used for comparison in this paper but should be utilized in the next iterations. Either a classifier that is trained on the dataset, in-context learning (few shot prompting).", "questions": "More of suggestions for improvement. \n\n1. Writing and the organizing of the paper needs a bit of work, there should be a set of contributions or research questions set early in the research paper, just to highlight everything in the paper and to guide the narrative. The authors contribute (or at least they’ve created) a new human annotated dataset for Frisian language but this is not highlighted until later in a section. The narrative can be easily followed if there was a clear RQ set. \n2. The generalizability of the synthetic experiment is questionable. The LLM evaluator was tasked to pick a datapoint from a provided list rather than generate a new one, reportedly because generation led to poor performance. This \"picking\" strategy seems to be a significant concession and a deviation from the described protocol. The implications of this methodological shortcut on the validity of the results, and how this relates to the natural language experiment where generation is performed, are not sufficiently discussed.\n3. It would have been helpful if there was a figure to handle the overall narrative, Fig 1 doesn’t capture the entirety of the entire research workflow proposed in the paper. \n4. More on model selection for a task that is low-resourced, it would have been helpful if the impact was compared against a baseline for multi-lingual such as Cohere's Aya model (https://cohere.com/research/aya). Why was the gpt 4.1 model identified over a multi-lingual model that is trained for that language (West Firisian)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762016560771}, {"id": "AtPKHoNQF2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7894/Reviewer_vyZn"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces the No-Data Algorithm, a theoretically motivated protocol for establishing the trustworthiness of automated evaluators, such as LLMs-as-judges, in scenarios where no labeled reference data is available. The protocol is inspired by zero-knowledge proofs and interactive challenge-response systems, allowing a verifier to probabilistically assess whether an evaluator truly knows the labeling function. Theoretical correctness and error bounds are provided, and empirical validation is conducted on synthetic tasks and a low-resource language labeling scenario.", "review_text": "This paper introduces the No-Data Algorithm, a theoretically motivated protocol for establishing the trustworthiness of automated evaluators, such as LLMs-as-judges, in scenarios where no labeled reference data is available. The protocol is inspired by zero-knowledge proofs and interactive challenge-response systems, allowing a verifier to probabilistically assess whether an evaluator truly knows the labeling function. Theoretical correctness and error bounds are provided, and empirical validation is conducted on synthetic tasks and a low-resource language labeling scenario.", "strengths": "- This paper addresses an important and challenging problem, which is how to trust automated evaluators without labeled data.\n- The authors propose a novel, theoretically grounded protocol inspired by zero-knowledge proofs and interactive proofs. Additionally, this paper provides formal correctness guarantees and explicit error bounds.\n- Empirical results on synthetic and real-world task (LLMs-as-judges on low-resource languages) support the theoretical claims.", "weaknesses": "- The proposed methods assume the existence of a shared, explicit rubric/aggregator decomposition, which may not be realistic for many annotation tasks.\n- Similarity between datapoints is not well-defined for complex or unstructured domains.\n- Experiments are limited to binary labels and do not cover more complex or open-ended settings. Scalability to large datasets and high-dimensional input is untested.", "questions": "- How can the protocol be adapted to tasks where the rubric is not easily decomposable or is inherently subjective/ambiguous?\n- How is 'similarity' between datapoints defined or computed in real-world, high-dimensional, or unstructured domains?\n- How robust is the method to imperfect, partial, or subjective rubrics, or to stochastic evaluators/verifiers (e.g., LLMs with temperature > 0)?\n- What are the practical computational and annotation overheads for deploying the No-Data Algorithm in real-world settings?\n- What does “w.h.p.” mean in the abstract?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the No-Data Algorithm, a theoretically motivated protocol for establishing the trustworthiness of automated evaluators, such as LLMs-as-judges, in scenarios where no labeled reference data is available. The protocol is inspired by zero-knowledge proofs and interactive challenge-response systems, allowing a verifier to probabilistically assess whether an evaluator truly knows the labeling function. Theoretical correctness and error bounds are provided, and empirical validation is conducted on synthetic tasks and a low-resource language labeling scenario.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- This paper addresses an important and challenging problem, which is how to trust automated evaluators without labeled data.\n- The authors propose a novel, theoretically grounded protocol inspired by zero-knowledge proofs and interactive proofs. Additionally, this paper provides formal correctness guarantees and explicit error bounds.\n- Empirical results on synthetic and real-world task (LLMs-as-judges on low-resource languages) support the theoretical claims.", "weaknesses": "- The proposed methods assume the existence of a shared, explicit rubric/aggregator decomposition, which may not be realistic for many annotation tasks.\n- Similarity between datapoints is not well-defined for complex or unstructured domains.\n- Experiments are limited to binary labels and do not cover more complex or open-ended settings. Scalability to large datasets and high-dimensional input is untested.", "questions": "- How can the protocol be adapted to tasks where the rubric is not easily decomposable or is inherently subjective/ambiguous?\n- How is 'similarity' between datapoints defined or computed in real-world, high-dimensional, or unstructured domains?\n- How robust is the method to imperfect, partial, or subjective rubrics, or to stochastic evaluators/verifiers (e.g., LLMs with temperature > 0)?\n- What are the practical computational and annotation overheads for deploying the No-Data Algorithm in real-world settings?\n- What does “w.h.p.” mean in the abstract?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761852469512}, {"id": "DnQO278ysP", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7894/Reviewer_kuS7"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces the \"No-Data Algorithm\", a method to establish trustworthiness in an evaluator (e.g., LLM-as-judge) for labeling data without any pre-existing labeled references. The approach relies on an Evaluator-Verifier protocol, where the evaluator generates similar datapoints and partial labels, and the verifier challenges them based on structural (equality up to permutation) and valuation (equality up to isomorphism) checks using a predefined rubric and aggregator. Theoretical proofs show correctness under assumptions like decomposable rubrics and binary data/labels, and the algorithm bounds the probability of undetected lies to (1/4)^r after r rounds. Experiments validate this on synthetic binary datasets (using decision trees and LLMs) and a real-world low-resource language task (West Frisian labeling with LLMs), demonstrating that success rates distinguish knowledgeable from untrustworthy evaluators.", "review_text": "This paper introduces the \"No-Data Algorithm\", a method to establish trustworthiness in an evaluator (e.g., LLM-as-judge) for labeling data without any pre-existing labeled references. The approach relies on an Evaluator-Verifier protocol, where the evaluator generates similar datapoints and partial labels, and the verifier challenges them based on structural (equality up to permutation) and valuation (equality up to isomorphism) checks using a predefined rubric and aggregator. Theoretical proofs show correctness under assumptions like decomposable rubrics and binary data/labels, and the algorithm bounds the probability of undetected lies to (1/4)^r after r rounds. Experiments validate this on synthetic binary datasets (using decision trees and LLMs) and a real-world low-resource language task (West Frisian labeling with LLMs), demonstrating that success rates distinguish knowledgeable from untrustworthy evaluators.", "strengths": "- The paper addresses a critical gap in evaluator trustworthiness without references, drawing from zero-knowledge proofs to create a verifiable challenge-response system. The formal proofs provide strong theoretical guarantees, bounding error probabilities in a probabilistic manner, which is a refreshing departure from empirical-only approaches in LLM evaluation literature.\n- The method enables trust establishment in data-scarce domains like low-resource NLP or medical labeling, without relying on potentially contaminated benchmarks. Ablations in appendices (e.g., on generation strategies and multi-model comparisons) add depth, showing robustness across evaluators like decision trees and various LLMs (o3-mini, GPT-4o, etc.).\n- Overall the paper is well-structured, with intuitive definitions (e.g., rubric and aggregator) and adequate theoretical depth.", "weaknesses": "- My biggest concern is that the algorithm's core assumption assumes the true labeling function f decomposes into a rubric C (criteria) and aggregator σ (e.g., majority vote); it is reasonable in theory but challenging in practice. Defining a \"total\" rubric that fully decomposes nonlinear criteria requires domain expertise and may not capture complex, implicit decision-making in real evaluators like LLMs, which often rely on emergent behaviors rather than explicit rules. Also, the binary string representation of data (X ⊂ {0,1}^n) and binary labels (Y = {0,1}) simplify proofs but limit direct applicability to multi-class or continuous-label tasks; while the discussion mentions extensions (e.g., via one-hot encoding), these could increase runtime by a factor of k-1 for k classes, potentially making it inefficient for high-dimensional data.\n- Experiments are confined to synthetic binary strings and a niche low-resource language (West Frisian). Importantly, the EV protocol relies on evaluators generating \"similar\" datapoints (thus assuming high-quality generation capabilities), but ablations show LLMs struggle with this (e.g., better performance when picking from datasets vs. generating anew), raising scalability issues for complex data - where the reference data are sparse and thus No-data algorithm could actually be useful. In unknowable scenarios, tuning phi (flip probability) based on expected accuracy is heuristic and assumes some prior knowledge, contradicting the \"no-data\" claim.\n- While results align with theory, the datasets are small (e.g., 498 entries for synthetic, 1,015 for West Frisian), and comparisons are limited. No baselines like prompt engineering or ensemble methods for trust establishment are evaluated, making it unclear if the algorithm outperforms simpler alternatives.\n- The verifier needs access to the rubric, which must be provided upfront—begging the question of how to obtain a reliable rubric without data. The protocol's multi-round calls could be computationally expensive for large datasets or real-time applications, and there's no analysis of failure modes like adversarial evaluators exploiting rubric ambiguities.", "questions": "- How do you propose defining and validating rubrics in real-world settings where the true f is unknown? For instance, in the West Frisian experiment, how was the rubric designed and quality-controlled?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the \"No-Data Algorithm\", a method to establish trustworthiness in an evaluator (e.g., LLM-as-judge) for labeling data without any pre-existing labeled references. The approach relies on an Evaluator-Verifier protocol, where the evaluator generates similar datapoints and partial labels, and the verifier challenges them based on structural (equality up to permutation) and valuation (equality up to isomorphism) checks using a predefined rubric and aggregator. Theoretical proofs show correctness under assumptions like decomposable rubrics and binary data/labels, and the algorithm bounds the probability of undetected lies to (1/4)^r after r rounds. Experiments validate this on synthetic binary datasets (using decision trees and LLMs) and a real-world low-resource language task (West Frisian labeling with LLMs), demonstrating that success rates distinguish knowledgeable from untrustworthy evaluators.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper addresses a critical gap in evaluator trustworthiness without references, drawing from zero-knowledge proofs to create a verifiable challenge-response system. The formal proofs provide strong theoretical guarantees, bounding error probabilities in a probabilistic manner, which is a refreshing departure from empirical-only approaches in LLM evaluation literature.\n- The method enables trust establishment in data-scarce domains like low-resource NLP or medical labeling, without relying on potentially contaminated benchmarks. Ablations in appendices (e.g., on generation strategies and multi-model comparisons) add depth, showing robustness across evaluators like decision trees and various LLMs (o3-mini, GPT-4o, etc.).\n- Overall the paper is well-structured, with intuitive definitions (e.g., rubric and aggregator) and adequate theoretical depth.", "weaknesses": "- My biggest concern is that the algorithm's core assumption assumes the true labeling function f decomposes into a rubric C (criteria) and aggregator σ (e.g., majority vote); it is reasonable in theory but challenging in practice. Defining a \"total\" rubric that fully decomposes nonlinear criteria requires domain expertise and may not capture complex, implicit decision-making in real evaluators like LLMs, which often rely on emergent behaviors rather than explicit rules. Also, the binary string representation of data (X ⊂ {0,1}^n) and binary labels (Y = {0,1}) simplify proofs but limit direct applicability to multi-class or continuous-label tasks; while the discussion mentions extensions (e.g., via one-hot encoding), these could increase runtime by a factor of k-1 for k classes, potentially making it inefficient for high-dimensional data.\n- Experiments are confined to synthetic binary strings and a niche low-resource language (West Frisian). Importantly, the EV protocol relies on evaluators generating \"similar\" datapoints (thus assuming high-quality generation capabilities), but ablations show LLMs struggle with this (e.g., better performance when picking from datasets vs. generating anew), raising scalability issues for complex data - where the reference data are sparse and thus No-data algorithm could actually be useful. In unknowable scenarios, tuning phi (flip probability) based on expected accuracy is heuristic and assumes some prior knowledge, contradicting the \"no-data\" claim.\n- While results align with theory, the datasets are small (e.g., 498 entries for synthetic, 1,015 for West Frisian), and comparisons are limited. No baselines like prompt engineering or ensemble methods for trust establishment are evaluated, making it unclear if the algorithm outperforms simpler alternatives.\n- The verifier needs access to the rubric, which must be provided upfront—begging the question of how to obtain a reliable rubric without data. The protocol's multi-round calls could be computationally expensive for large datasets or real-time applications, and there's no analysis of failure modes like adversarial evaluators exploiting rubric ambiguities.", "questions": "- How do you propose defining and validating rubrics in real-world settings where the true f is unknown? For instance, in the West Frisian experiment, how was the rubric designed and quality-controlled?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761413354921}], "openreview_url": "https://openreview.net/forum?id=kaW0pt8pLR", "arxiv_id": "2506.03083", "paper_pdf": "papers/kaW0pt8pLR.pdf", "paper_pdf_sha256": "b7006a2fbab7b6305f77b358456382272f8d1fc221ef25ed89921b649332a92d", "paper_pdf_bytes": 591154, "paper_pdf_source": "openreview", "code_url": "https://github.com/adewynter/no_data_algorithm", "code_repository": "adewynter/no_data_algorithm", "code_commit": "541e58cc1fc39ac9307930ab6f5a49beaa7c4b12", "code_archive": "repos/kaW0pt8pLR.zip", "code_archive_sha256": "97f0f0d9bbad4f9ce924d309be5cb8d46031374362fe27b68291cbd6c642e0b0", "code_archive_bytes": 1764104, "code_file_count": 7, "code_extensions": {".py": 4, ".ipynb": 3}, "github_disk_usage_kb": 1808, "github_languages": {"Jupyter Notebook": 243131, "Python": 32965}, "github_archived": false, "github_pushed_at": "2026-05-05T03:44:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/labelling-data-with-unknown-references"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sMWkTWh2JF", "year": 2025, "status": "rejected", "title": "ReLIC: A Recipe for 64k Steps of In-Context Reinforcement Learning for Embodied AI", "authors": ["Ahmad Elawady", "Gunjan Chhablani", "Ram Ramrakhya", "Karmesh Yadav", "Dhruv Batra", "Zsolt Kira", "Andrew Szot"], "authorids": ["~Ahmad_Elawady1", "~Gunjan_Chhablani1", "~Ram_Ramrakhya2", "~Karmesh_Yadav1", "~Dhruv_Batra1", "~Zsolt_Kira1", "~Andrew_Szot1"], "authors_source": "OpenReview API", "abstract": "Intelligent embodied agents need to quickly adapt to new scenarios by integrating long histories of experience into decision-making. For instance, a robot in an unfamiliar house initially wouldn't know the locations of objects needed for tasks and might perform inefficiently. However, as it gathers more experience, it should learn the layout of its environment and remember where objects are, allowing it to complete new tasks more efficiently. To enable such rapid adaptation to new tasks, we present ReLIC, a new approach for in-context reinforcement learning (RL) for embodied agents. With ReLIC, agents are capable of adapting to new environments using 64,000 steps of in-context experience with full attention while being trained through self-generated experience via RL. We achieve this by proposing a novel policy update scheme for on-policy RL called \"partial updates\" as well as a Sink-KV mechanism that enables effective utilization of a long observation history for embodied agents. Our method outperforms a variety of meta-RL baselines in adapting to unseen houses in an embodied multi-object navigation task. In addition, we find that ReLIC is capable of few-shot imitation learning despite never being trained with expert demonstrations. We also provide a comprehensive analysis of ReLIC, highlighting that the combination of large-scale RL training, the proposed partial updates scheme, and the Sink-KV are essential for effective in-context learning.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "l0oZWW4DMb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5257/Reviewer_e3Wg"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces ReLIC (Reinforcement Learning In Context), a new approach that enables embodied AI agents to adapt to new environments using up to 64,000 steps of in-context experience. ReLIC addresses two key challenges in scaling in-context reinforcement learning: efficiently training with long context windows and effectively utilizing visual information from multiple episodes. To solve these challenges, the authors introduce \"partial updates,\" a novel policy update scheme that allows more frequent learning during long rollouts, and \"Sink-KV,\" a modification to transformer attention that helps the model selectively attend to relevant information in long sequences.\n\nThe authors evaluate ReLIC on an extended version of object navigation where agents must find multiple objects in unseen house layouts. ReLIC significantly outperforms baseline approaches, achieving a 43% success rate compared to 22% from the closest baseline after 15 episodes of in-context experience. The method also demonstrates an unexpected capability for few-shot imitation learning despite never being trained with expert demonstrations. Through extensive ablation studies, the authors show that both partial updates and Sink-KV are crucial for effective learning, and that sufficient training scale is necessary for in-context learning capabilities to emerge. The approach also performs well on existing benchmarks like Darkroom and Miniworld tasks, showing better and faster adaptation compared to prior work.", "review_text": "The paper introduces ReLIC (Reinforcement Learning In Context), a new approach that enables embodied AI agents to adapt to new environments using up to 64,000 steps of in-context experience. ReLIC addresses two key challenges in scaling in-context reinforcement learning: efficiently training with long context windows and effectively utilizing visual information from multiple episodes. To solve these challenges, the authors introduce \"partial updates,\" a novel policy update scheme that allows more frequent learning during long rollouts, and \"Sink-KV,\" a modification to transformer attention that helps the model selectively attend to relevant information in long sequences.\n\nThe authors evaluate ReLIC on an extended version of object navigation where agents must find multiple objects in unseen house layouts. ReLIC significantly outperforms baseline approaches, achieving a 43% success rate compared to 22% from the closest baseline after 15 episodes of in-context experience. The method also demonstrates an unexpected capability for few-shot imitation learning despite never being trained with expert demonstrations. Through extensive ablation studies, the authors show that both partial updates and Sink-KV are crucial for effective learning, and that sufficient training scale is necessary for in-context learning capabilities to emerge. The approach also performs well on existing benchmarks like Darkroom and Miniworld tasks, showing better and faster adaptation compared to prior work.", "strengths": "1. Strong Empirical Results: The method achieves substantial improvements over baselines, nearly doubling the success rate (43% vs 22%) in challenging navigation tasks. The results are thoroughly validated across multiple environments (EXTOBJNAV, Darkroom, Miniworld) with comprehensive ablation studies.\n2. Scalability: The paper demonstrates impressive scaling capabilities, showing that their method can handle up to 64,000 steps of context - a significant improvement over previous approaches. More importantly, they show that models trained on shorter contexts (4k steps) can generalize to much longer contexts (32k steps) at inference time.", "weaknesses": "1. Limited Success Rate: Despite significant improvements over baselines, the absolute performance (43% success rate) is still relatively low for practical applications. The paper doesn't thoroughly discuss why this ceiling exists or potential paths to improve it.\n\n2. Computational Cost: The method requires substantial computational resources - 12 days of training on 4 NVIDIA A40 GPUs. This high computational requirement could limit the practical applicability and accessibility of the approach, especially for researchers with limited resources.\n\n3. Task Limitation: The evaluation focuses primarily on navigation tasks with discrete action spaces. There's no exploration of how the method might extend to continuous action spaces or more complex tasks like manipulation, limiting our understanding of the approach's broader applicability.\n\n4. Missing Related work: \nRetrieval-Augmented Decision Transformer: External Memory for In-context RL, Schmied et.al\n\t - Work is quite similar to yours. It would be interesting if the authors could point out the differences in the related work section. This will improve the paper.", "questions": "1. Scalability and Practical Applications: Given the high computational requirements (12 days on 4 A40 GPUs), what approaches have you considered to make ReLIC more practically applicable? Could techniques like model distillation or more efficient architectures reduce these requirements while maintaining performance?\n2. Beyond Navigation: The current work focuses on discrete action spaces in navigation tasks. Have you explored how ReLIC could be extended to continuous action spaces or more complex tasks like manipulation? What modifications would be needed to handle such scenarios?\n3. Emergent Few-shot Learning: The emergence of few-shot imitation learning capabilities without explicit training is intriguing. What mechanisms in ReLIC enable this capability, and how could it be enhanced? Have you analyzed how the quality of demonstrations affects performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ReLIC (Reinforcement Learning In Context), a new approach that enables embodied AI agents to adapt to new environments using up to 64,000 steps of in-context experience. ReLIC addresses two key challenges in scaling in-context reinforcement learning: efficiently training with long context windows and effectively utilizing visual information from multiple episodes. To solve these challenges, the authors introduce \"partial updates,\" a novel policy update scheme that allows more frequent learning during long rollouts, and \"Sink-KV,\" a modification to transformer attention that helps the model selectively attend to relevant information in long sequences.\n\nThe authors evaluate ReLIC on an extended version of object navigation where agents must find multiple objects in unseen house layouts. ReLIC significantly outperforms baseline approaches, achieving a 43% success rate compared to 22% from the closest baseline after 15 episodes of in-context experience. The method also demonstrates an unexpected capability for few-shot imitation learning despite never being trained with expert demonstrations. Through extensive ablation studies, the authors show that both partial updates and Sink-KV are crucial for effective learning, and that sufficient training scale is necessary for in-context learning capabilities to emerge. The approach also performs well on existing benchmarks like Darkroom and Miniworld tasks, showing better and faster adaptation compared to prior work.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Strong Empirical Results: The method achieves substantial improvements over baselines, nearly doubling the success rate (43% vs 22%) in challenging navigation tasks. The results are thoroughly validated across multiple environments (EXTOBJNAV, Darkroom, Miniworld) with comprehensive ablation studies.\n2. Scalability: The paper demonstrates impressive scaling capabilities, showing that their method can handle up to 64,000 steps of context - a significant improvement over previous approaches. More importantly, they show that models trained on shorter contexts (4k steps) can generalize to much longer contexts (32k steps) at inference time.", "weaknesses": "1. Limited Success Rate: Despite significant improvements over baselines, the absolute performance (43% success rate) is still relatively low for practical applications. The paper doesn't thoroughly discuss why this ceiling exists or potential paths to improve it.\n\n2. Computational Cost: The method requires substantial computational resources - 12 days of training on 4 NVIDIA A40 GPUs. This high computational requirement could limit the practical applicability and accessibility of the approach, especially for researchers with limited resources.\n\n3. Task Limitation: The evaluation focuses primarily on navigation tasks with discrete action spaces. There's no exploration of how the method might extend to continuous action spaces or more complex tasks like manipulation, limiting our understanding of the approach's broader applicability.\n\n4. Missing Related work: \nRetrieval-Augmented Decision Transformer: External Memory for In-context RL, Schmied et.al\n\t - Work is quite similar to yours. It would be interesting if the authors could point out the differences in the related work section. This will improve the paper.", "questions": "1. Scalability and Practical Applications: Given the high computational requirements (12 days on 4 A40 GPUs), what approaches have you considered to make ReLIC more practically applicable? Could techniques like model distillation or more efficient architectures reduce these requirements while maintaining performance?\n2. Beyond Navigation: The current work focuses on discrete action spaces in navigation tasks. Have you explored how ReLIC could be extended to continuous action spaces or more complex tasks like manipulation? What modifications would be needed to handle such scenarios?\n3. Emergent Few-shot Learning: The emergence of few-shot imitation learning capabilities without explicit training is intriguing. What mechanisms in ReLIC enable this capability, and how could it be enhanced? Have you analyzed how the quality of demonstrations affects performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731039738869}, {"id": "WC1fZjuYWn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5257/Reviewer_tdez"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "* This paper proposes a new in-context RL approach with 64K historied steps. They introduce partial update schemes to solve sample inefficiency problems and Sink-KV mechanism to address long observation history.", "review_text": "* This paper proposes a new in-context RL approach with 64K historied steps. They introduce partial update schemes to solve sample inefficiency problems and Sink-KV mechanism to address long observation history.", "strengths": "*  The problems proposed by the paper are meaningful. Generalization is a long-term goal for embodied AI. In context learning with longer prompts may be a potential solution.\n*  This paper is well-structured, conducts rich experiments and achieves great performance.", "weaknesses": "*  The description for \"partial update\" is confusing. Providing pseudocode would be good for understanding.\n*  The contributions of the paper are confusing. This paper proposes two ways to solve long in-context RL: partial update and SINK-KV. \"Partial update\" is more like a training trick and \"Sink-KV\" is proposed by previous works. In the appendix, the paper explaints \"Sink-KV\" will help the policy for exploration. This is interesting and it should be place in the main paper.\n*  Lack of some comparisons. Like transformer with multi episodes and DPT in ExtObjNav. \n*  Environments used in experiments are too simple and they are all navigations tasks. It would be interesting to include manipulation benchmarks like MetaWorld.\n*  Inefficient and large GPU memory for long context length. What is the inference speed for ReLIC and other baselines?\n*  This paper emphasise 64K in the title and introduction. But in the experiment sections, all results are evaluated under small context length.", "questions": "*  See weakness.\n*  In the Darkroom and miniwork experiment, all baselines have very low performance (0 success rate) with no context while ReLIC has higher performance (30% success rate). Is it normal? And DPT can get a very higher improvement with 20 in-context episodes.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "* This paper proposes a new in-context RL approach with 64K historied steps. They introduce partial update schemes to solve sample inefficiency problems and Sink-KV mechanism to address long observation history.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "*  The problems proposed by the paper are meaningful. Generalization is a long-term goal for embodied AI. In context learning with longer prompts may be a potential solution.\n*  This paper is well-structured, conducts rich experiments and achieves great performance.", "weaknesses": "*  The description for \"partial update\" is confusing. Providing pseudocode would be good for understanding.\n*  The contributions of the paper are confusing. This paper proposes two ways to solve long in-context RL: partial update and SINK-KV. \"Partial update\" is more like a training trick and \"Sink-KV\" is proposed by previous works. In the appendix, the paper explaints \"Sink-KV\" will help the policy for exploration. This is interesting and it should be place in the main paper.\n*  Lack of some comparisons. Like transformer with multi episodes and DPT in ExtObjNav. \n*  Environments used in experiments are too simple and they are all navigations tasks. It would be interesting to include manipulation benchmarks like MetaWorld.\n*  Inefficient and large GPU memory for long context length. What is the inference speed for ReLIC and other baselines?\n*  This paper emphasise 64K in the title and introduction. But in the experiment sections, all results are evaluated under small context length.", "questions": "*  See weakness.\n*  In the Darkroom and miniwork experiment, all baselines have very low performance (0 success rate) with no context while ReLIC has higher performance (30% success rate). Is it normal? And DPT can get a very higher improvement with 20 in-context episodes.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730639415210}, {"id": "44CSSniYWN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5257/Reviewer_pU2u"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces ReLIC, a framework designed to extend in-context learning capabilities to embodied AI agents. ReLIC focuses on enabling agents to utilize up to 64,000 steps of contextual experience, aiming to allow rapid adaptation in unseen environments, especially for multi-object navigation tasks. This is achieved through two components: (1) Partial Updates, a policy update scheme that incrementally adjusts the policy within rollouts, and (2) Sink-KV, a mechanism for efficient attention management in transformers, which optimizes long-sequence processing by adding learnable sink key-value vectors in attention layers.", "review_text": "The paper introduces ReLIC, a framework designed to extend in-context learning capabilities to embodied AI agents. ReLIC focuses on enabling agents to utilize up to 64,000 steps of contextual experience, aiming to allow rapid adaptation in unseen environments, especially for multi-object navigation tasks. This is achieved through two components: (1) Partial Updates, a policy update scheme that incrementally adjusts the policy within rollouts, and (2) Sink-KV, a mechanism for efficient attention management in transformers, which optimizes long-sequence processing by adding learnable sink key-value vectors in attention layers.", "strengths": "1. The paper is generally well-written and easy-to-understand.\n2. Experiments show that the proposed approach is effective.\n3. Hyperparameters and algorithm details are provided for reproducity.", "weaknesses": "1. The two major novel components of the paper as claimed by the authors are the partial updates and Sink-KV. For the former one, if I understand correctly, it is a very common technique used in common RL training (I don't if it's the same case for in-context RL as I'm not familiar with its literature), namely, update the policy/value in the middle of an episode to increase training frequency. For Sink-KV, it seems to be interesting but I'm not familiar with the literature exploring the architecture of transformer so I don't how novel this is, but one thing that is not clear to me is that, why this design especially helps in-context RL training? What makes it special compared to using it in vision/language tasks?\n\n2. For the experiments, the author mainly compared their method to meta-rl baselines and different variants of their method - while this is good, I would expect to see how the proposed method compared to other in-context RL baselines. \n\n3. The emergent imitation learning results are not that surprising to me - a lot of recent work  do pretraining with imitation learning and then rl finetuning, the proposed scheme here is like first pretraining with rl then finetuning with imitation learning. This setting is also a little bit strange as in the real world case, if we have the expert demonstrations, pretraining on that usally greatly speed up RL.\n\n4. std/error bar not included in the darkroom result plot and miniworld result plot", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ReLIC, a framework designed to extend in-context learning capabilities to embodied AI agents. ReLIC focuses on enabling agents to utilize up to 64,000 steps of contextual experience, aiming to allow rapid adaptation in unseen environments, especially for multi-object navigation tasks. This is achieved through two components: (1) Partial Updates, a policy update scheme that incrementally adjusts the policy within rollouts, and (2) Sink-KV, a mechanism for efficient attention management in transformers, which optimizes long-sequence processing by adding learnable sink key-value vectors in attention layers.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper is generally well-written and easy-to-understand.\n2. Experiments show that the proposed approach is effective.\n3. Hyperparameters and algorithm details are provided for reproducity.", "weaknesses": "1. The two major novel components of the paper as claimed by the authors are the partial updates and Sink-KV. For the former one, if I understand correctly, it is a very common technique used in common RL training (I don't if it's the same case for in-context RL as I'm not familiar with its literature), namely, update the policy/value in the middle of an episode to increase training frequency. For Sink-KV, it seems to be interesting but I'm not familiar with the literature exploring the architecture of transformer so I don't how novel this is, but one thing that is not clear to me is that, why this design especially helps in-context RL training? What makes it special compared to using it in vision/language tasks?\n\n2. For the experiments, the author mainly compared their method to meta-rl baselines and different variants of their method - while this is good, I would expect to see how the proposed method compared to other in-context RL baselines. \n\n3. The emergent imitation learning results are not that surprising to me - a lot of recent work  do pretraining with imitation learning and then rl finetuning, the proposed scheme here is like first pretraining with rl then finetuning with imitation learning. This setting is also a little bit strange as in the real world case, if we have the expert demonstrations, pretraining on that usally greatly speed up RL.\n\n4. std/error bar not included in the darkroom result plot and miniworld result plot", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730431068617}], "openreview_url": "https://openreview.net/forum?id=sMWkTWh2JF", "arxiv_id": "2410.02751", "paper_pdf": "papers/sMWkTWh2JF.pdf", "paper_pdf_sha256": "09de2695265e86d91aa57c9322faf8a22bde8912bd09dfca93e10739c4fc1414", "paper_pdf_bytes": 3613664, "paper_pdf_source": "openreview", "code_url": "https://github.com/aielawady/relic", "code_repository": "aielawady/relic", "code_commit": "1c3b7a13b2a2c5873223de2680b6cd5ed6a8fc96", "code_archive": "repos/sMWkTWh2JF.zip", "code_archive_sha256": "b35d95bb1976d23422c72c678a37d6636856ee5a5680751e57fc86f513f88214", "code_archive_bytes": 155858, "code_file_count": 41, "code_extensions": {".py": 39, ".sh": 2}, "github_disk_usage_kb": 123, "github_languages": {"Python": 536329, "Shell": 3350}, "github_archived": false, "github_pushed_at": "2024-09-07T21:56:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/relic-a-recipe-for-64k-steps-of-in-context"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8w6FzR68DS", "year": 2024, "status": "rejected", "title": "PriViT: Vision Transformers for Fast Private Inference", "authors": ["Naren Dhyani", "Jianqiao Cambridge Mo", "Minsu Cho", "Ameya Joshi", "Siddharth Garg", "Brandon Reagen", "Chinmay Hegde"], "authorids": ["~Naren_Dhyani1", "~Jianqiao_Cambridge_Mo1", "~Minsu_Cho2", "~Ameya_Joshi2", "~Siddharth_Garg1", "~Brandon_Reagen1", "~Chinmay_Hegde1"], "authors_source": "OpenReview API", "abstract": "The Vision Transformer (ViT) architecture has emerged as the backbone of choice for state-of-the-art deep models for computer vision applications. However, ViTs are ill-suited for private inference using secure multi-party computation (MPC) protocols, due to the large number of non-polynomial operations (self-attention, feed-forward rectifiers, layer normalization). We propose PriViT, a gradient-based algorithm to selectively Taylorize nonlinearities in ViTs while maintaining their prediction accuracy. Our algorithm is conceptually simple, easy to implement, and achieves improved performance over existing approaches for designing MPC-friendly transformer architectures in terms of achieving the Pareto frontier in latency-accuracy. We confirm these improvements via experiments on several standard image classification tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "wkCIgNQAlh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1467/Reviewer_y916"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposed a new algorithm for constructing MPC-friendly vision transformers. Accuracy and latency performance were compared on several datasets with the previous result, MPCViT.", "review_text": "This paper proposed a new algorithm for constructing MPC-friendly vision transformers. Accuracy and latency performance were compared on several datasets with the previous result, MPCViT.", "strengths": "By adjusting GELU and softmax through training using a switched method, they found a different optimal method for each layer.", "weaknesses": "Compared to the prior technology, MPCViT, it shows better results in the TinyImagenet but worse latency in the CIFAR-100. In terms of accuracy, it has superior performance in any case. The paper said that DELPHI is focused as the subject of comparison. \"In this paper, our focus is exclusively on the DELPHI protocol (Mishra et al., 2020a) for private inference. We choose DELPHI as a matter of convenience;\" However, the actual results do not show any performance comparison with DELPHI.", "questions": "It is necessary to explain in what aspects DELPHI was considered.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a new algorithm for constructing MPC-friendly vision transformers. Accuracy and latency performance were compared on several datasets with the previous result, MPCViT.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "By adjusting GELU and softmax through training using a switched method, they found a different optimal method for each layer.", "weaknesses": "Compared to the prior technology, MPCViT, it shows better results in the TinyImagenet but worse latency in the CIFAR-100. In terms of accuracy, it has superior performance in any case. The paper said that DELPHI is focused as the subject of comparison. \"In this paper, our focus is exclusively on the DELPHI protocol (Mishra et al., 2020a) for private inference. We choose DELPHI as a matter of convenience;\" However, the actual results do not show any performance comparison with DELPHI.", "questions": "It is necessary to explain in what aspects DELPHI was considered.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "I have no concern", "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699089649208}, {"id": "CD6lSDa2Dc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1467/Reviewer_HTX8"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "They propose a new vision transformer architecture which is a MPC-friendly ViT. The proposed architecture optimizes the attention part with the SquaredAttn and replace the GELU function to reduce the nonlinear operations in the model. The proposed method achieve state-of-the-art performance on several image classification benchmarks and speed up the model compared with current MPCViT.", "review_text": "They propose a new vision transformer architecture which is a MPC-friendly ViT. The proposed architecture optimizes the attention part with the SquaredAttn and replace the GELU function to reduce the nonlinear operations in the model. The proposed method achieve state-of-the-art performance on several image classification benchmarks and speed up the model compared with current MPCViT.", "strengths": "1. The method analysis is clear and latency breakdown is helpful.\n2. The experiments on serval image classification benchmarks are solid and comprehensive.\n3. The proposed method is speed up than previous SOTA model and achieve competitive performance.", "weaknesses": "1. Need more detailed about the knowledge distillation part.\n2. More discussion about non-linearity distribution.", "questions": "1. I am curious about the teacher model size and can it boost more performance if we use a larger teacher?\n2. As you states that later layers have a larger number of linearized units, can we first try to linearized the later layers then try to linearized the earlier layers?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "They propose a new vision transformer architecture which is a MPC-friendly ViT. The proposed architecture optimizes the attention part with the SquaredAttn and replace the GELU function to reduce the nonlinear operations in the model. The proposed method achieve state-of-the-art performance on several image classification benchmarks and speed up the model compared with current MPCViT.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The method analysis is clear and latency breakdown is helpful.\n2. The experiments on serval image classification benchmarks are solid and comprehensive.\n3. The proposed method is speed up than previous SOTA model and achieve competitive performance.", "weaknesses": "1. Need more detailed about the knowledge distillation part.\n2. More discussion about non-linearity distribution.", "questions": "1. I am curious about the teacher model size and can it boost more performance if we use a larger teacher?\n2. As you states that later layers have a larger number of linearized units, can we first try to linearized the later layers then try to linearized the earlier layers?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698827670836}, {"id": "FP8iVmDuIm", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1467/Reviewer_BazG"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces PriViT as a solution for implementing Vision Transformers~(ViTs) in private inference (PI) applications. To achieve this, PriViT tailors the nonlinear operations (e.g., GELU and row-wise softmax) of a pre-trained Transformer in order to make them compatible with secure multi-party computation (MPC) protocols. This adaptation involves the introduction of learnable switching variables to gradually replace the original nonlinear operations. The experimental results on CIFAR-10/100 and TinyImageNet demonstrate a significant improvement in both speed and performance compared to prior approaches.", "review_text": "This paper introduces PriViT as a solution for implementing Vision Transformers~(ViTs) in private inference (PI) applications. To achieve this, PriViT tailors the nonlinear operations (e.g., GELU and row-wise softmax) of a pre-trained Transformer in order to make them compatible with secure multi-party computation (MPC) protocols. This adaptation involves the introduction of learnable switching variables to gradually replace the original nonlinear operations. The experimental results on CIFAR-10/100 and TinyImageNet demonstrate a significant improvement in both speed and performance compared to prior approaches.", "strengths": "* The paper is well-motivated as the deployment of ViTs in private scenarios is becoming increasingly important and current approaches are not tailored for Transformer architecture.\n* The proposed method is quite simple yet effective comparing with SOTA approaches.", "weaknesses": "[Major]\n\n1. **Experiments:** The authors have conducted sufficient comparative experiments conducted on 3 datasets (CIFAR-10/100 and TinyImageNet). However, the image resolutions are no more than $64\\times 64$, which is rather small compared with commonly-used datasets like ImageNet-1k and Caltech-101/256. It will be interesting to see ImageNet results and compare with SENet if possible.\n2. **Experiments:** The authors' exclusive use of ViT-Tiny for comparison is insufficient to establish the method's universality. It is strongly recommended that the authors broaden their evaluation to include a variety of ViT architectures, such as DeiT, and models with diverse parameters, including ViT-Small and ViT-Base. Such an expanded assessment would be of significant interest to the community, particularly in light of the flourishing development of large Transformer-based vision models.\n3. **Experiments:** In Table 3, the authors have presented PriViT and MPCViT in distinct hyper-parameter configurations, giving rise to concerns regarding the fairness of the comparisons.\n4. **Experiments:** Furthermore, I strongly encourage the authors to furnish additional results to showcase the superiority of their proposed method. This could encompass metrics like GPU memory usage and training time in comparison to other existing methods.\n\n[Minor]\n1. This article is poorly written, with issues including imprecise mathematical formulas, non-standard captions, typographical errors, and punctuation. Here, I have listed some errors and hope that the authors will carefully revise and proofread their manuscript and the appendices.\n    * Eq (1): $o=\\frac{\\mathrm{Softmax}(XW_qW_kX)}{\\sqrt{d}}W_vX$ -> $o=\\frac{\\mathrm{Softmax}(XW_qW_k^\\top X^\\top)}{\\sqrt{d}}XW_v$. Eq (4) has the same problem.\n    * Page 5, last paragraph: Once the model satisfies the required budgets,,we... -> Once the model satisfies the required budgets, we...\n    * Page 6, first paragraph: Architecture and data set. -> Architecture and dataset.\n    * Page 6, first paragraph: ViT Tiny -> ViT-Tiny\n    * Page 6, first paragraph: 224x224 -> $224\\times 224$\n    * The caption of tables should be on the top of the table: Table 3 and 5.\n    * Page 7: Pareto analysis of PriViT over Tiny Imagenet, and Cifar10/100 -> Pareto analysis of PriViT over Tiny Imagenet, and Cifar10/100.\n    * Page 7: table 4 -> Table 4, Fig 2 -> Figure 2, fig 13 -> Figure 13, figure 3 -> Figure 3, Fig. 13 -> Figure 13.\n    * Page 8: The latency is calculated as per 4.1 -> The latency is calculated as per Section 4.1.\n2. Missing dataset details in Table 1.\n3. The authors do not discuss the limitations of their method.\n4. The authors do not provide codes for reproducibility check.", "questions": "My questions are listed in the \"Weaknesses\" section. I am looking forward to the authors' relply.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PriViT as a solution for implementing Vision Transformers~(ViTs) in private inference (PI) applications. To achieve this, PriViT tailors the nonlinear operations (e.g., GELU and row-wise softmax) of a pre-trained Transformer in order to make them compatible with secure multi-party computation (MPC) protocols. This adaptation involves the introduction of learnable switching variables to gradually replace the original nonlinear operations. The experimental results on CIFAR-10/100 and TinyImageNet demonstrate a significant improvement in both speed and performance compared to prior approaches.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "* The paper is well-motivated as the deployment of ViTs in private scenarios is becoming increasingly important and current approaches are not tailored for Transformer architecture.\n* The proposed method is quite simple yet effective comparing with SOTA approaches.", "weaknesses": "[Major]\n\n1. **Experiments:** The authors have conducted sufficient comparative experiments conducted on 3 datasets (CIFAR-10/100 and TinyImageNet). However, the image resolutions are no more than $64\\times 64$, which is rather small compared with commonly-used datasets like ImageNet-1k and Caltech-101/256. It will be interesting to see ImageNet results and compare with SENet if possible.\n2. **Experiments:** The authors' exclusive use of ViT-Tiny for comparison is insufficient to establish the method's universality. It is strongly recommended that the authors broaden their evaluation to include a variety of ViT architectures, such as DeiT, and models with diverse parameters, including ViT-Small and ViT-Base. Such an expanded assessment would be of significant interest to the community, particularly in light of the flourishing development of large Transformer-based vision models.\n3. **Experiments:** In Table 3, the authors have presented PriViT and MPCViT in distinct hyper-parameter configurations, giving rise to concerns regarding the fairness of the comparisons.\n4. **Experiments:** Furthermore, I strongly encourage the authors to furnish additional results to showcase the superiority of their proposed method. This could encompass metrics like GPU memory usage and training time in comparison to other existing methods.\n\n[Minor]\n1. This article is poorly written, with issues including imprecise mathematical formulas, non-standard captions, typographical errors, and punctuation. Here, I have listed some errors and hope that the authors will carefully revise and proofread their manuscript and the appendices.\n    * Eq (1): $o=\\frac{\\mathrm{Softmax}(XW_qW_kX)}{\\sqrt{d}}W_vX$ -> $o=\\frac{\\mathrm{Softmax}(XW_qW_k^\\top X^\\top)}{\\sqrt{d}}XW_v$. Eq (4) has the same problem.\n    * Page 5, last paragraph: Once the model satisfies the required budgets,,we... -> Once the model satisfies the required budgets, we...\n    * Page 6, first paragraph: Architecture and data set. -> Architecture and dataset.\n    * Page 6, first paragraph: ViT Tiny -> ViT-Tiny\n    * Page 6, first paragraph: 224x224 -> $224\\times 224$\n    * The caption of tables should be on the top of the table: Table 3 and 5.\n    * Page 7: Pareto analysis of PriViT over Tiny Imagenet, and Cifar10/100 -> Pareto analysis of PriViT over Tiny Imagenet, and Cifar10/100.\n    * Page 7: table 4 -> Table 4, Fig 2 -> Figure 2, fig 13 -> Figure 13, figure 3 -> Figure 3, Fig. 13 -> Figure 13.\n    * Page 8: The latency is calculated as per 4.1 -> The latency is calculated as per Section 4.1.\n2. Missing dataset details in Table 1.\n3. The authors do not discuss the limitations of their method.\n4. The authors do not provide codes for reproducibility check.", "questions": "My questions are listed in the \"Weaknesses\" section. I am looking forward to the authors' relply.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698744987194}, {"id": "YYxu33YlRv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1467/Reviewer_pfyb"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces \"PriViT,\" an algorithm designed to make Vision Transformers (ViTs) suitable for private inference using secure multi-party computation (MPC) protocols. Traditional ViTs are not ideal for MPC due to their complex non-polynomial operations. PriViT aims to reduce the nonlinearities in ViTs and preserve their prediction accuracy. It demonstrates a balance in latency and accuracy superior to existing methods, as confirmed through experiments on image classification tasks.", "review_text": "The paper introduces \"PriViT,\" an algorithm designed to make Vision Transformers (ViTs) suitable for private inference using secure multi-party computation (MPC) protocols. Traditional ViTs are not ideal for MPC due to their complex non-polynomial operations. PriViT aims to reduce the nonlinearities in ViTs and preserve their prediction accuracy. It demonstrates a balance in latency and accuracy superior to existing methods, as confirmed through experiments on image classification tasks.", "strengths": "The paper is easy to follow. The paper tackles the significant challenge of making Vision Transformers (ViTs) compatible with private inference using secure multi-party computation (MPC) protocols.", "weaknesses": "A predominant weakness of the paper is its limited novelty, as it appears to draw heavily on core concepts previously established in SNL.", "questions": "1.\tWhile the paper introduces an interesting approach, its novelty appears constrained. The core concept of utilizing binary gates to manage the number of nonlinear operations through $\\ell_0$ sparsity has been proposed introduced in SNL (Cho et al., 2022b). The proposed method seems like an adaptation of SNL for ViTs. This adaptation is intriguing, but a deeper exploration into its unique contributions compared to SNL would provide more clarity on its novelty.\n\n2.\tThe paper's organization raises some concerns. While critical figures, and experimental results are put in the supplementary material, the main body of the paper exhibits noticeable empty spaces. It would enhance the paper's comprehensibility and impact if these essential elements were incorporated directly into the main content, making the narrative more coherent for readers without necessitating frequent references to supplementary sections.\n\n3.\tIn Figure 1, the computational cost is quantified using the metric of # AND gates. However, the methodology or criteria for determining the number of AND gates associated with a nonlinear operation remains ambiguous in the paper. Clarifying this calculation or providing a concise explanation within the main content would enhance the reader's understanding.\n\n4.\tThe experimental validation is primarily focused on small-scale datasets, which could limit the generalizability of the findings. For a comprehensive evaluation, it would be beneficial for the authors to include results from large-scale datasets like ImageNet. Such experiments would provide a more holistic view of the method's efficacy and scalability.\n\n5.\tBased on Figure 2, both MPCViT and MPCViT+ seem to outperform the proposed method in achieving a superior Pareto balance between accuracy and latency. This observation raises the question: what distinctive advantages or contributions does the proposed method offer? Is it perhaps more streamlined in terms of training efficiency or some other metric? A clearer delineation of the method's unique strengths would be beneficial for readers.\n\n6.\tObservations from Table 4 present some anomalies. Specifically, for PriViT-R, there is a counterintuitive trend where increased latency corresponds to a decline in performance. This is perplexing as one would typically expect a trade-off where longer computation times would yield better results. An explanation or insight into this apparent contradiction would enhance the clarity of the findings.\n\n7.\tThe clarity of notational conventions could be enhanced. For instance, the terms PriViT-R and PriViT-G appear in the text without explicit definitions or context. Providing a clear explanation or description of what these notations signify would aid in a more comprehensive understanding of the content for readers.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces \"PriViT,\" an algorithm designed to make Vision Transformers (ViTs) suitable for private inference using secure multi-party computation (MPC) protocols. Traditional ViTs are not ideal for MPC due to their complex non-polynomial operations. PriViT aims to reduce the nonlinearities in ViTs and preserve their prediction accuracy. It demonstrates a balance in latency and accuracy superior to existing methods, as confirmed through experiments on image classification tasks.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The paper is easy to follow. The paper tackles the significant challenge of making Vision Transformers (ViTs) compatible with private inference using secure multi-party computation (MPC) protocols.", "weaknesses": "A predominant weakness of the paper is its limited novelty, as it appears to draw heavily on core concepts previously established in SNL.", "questions": "1.\tWhile the paper introduces an interesting approach, its novelty appears constrained. The core concept of utilizing binary gates to manage the number of nonlinear operations through $\\ell_0$ sparsity has been proposed introduced in SNL (Cho et al., 2022b). The proposed method seems like an adaptation of SNL for ViTs. This adaptation is intriguing, but a deeper exploration into its unique contributions compared to SNL would provide more clarity on its novelty.\n\n2.\tThe paper's organization raises some concerns. While critical figures, and experimental results are put in the supplementary material, the main body of the paper exhibits noticeable empty spaces. It would enhance the paper's comprehensibility and impact if these essential elements were incorporated directly into the main content, making the narrative more coherent for readers without necessitating frequent references to supplementary sections.\n\n3.\tIn Figure 1, the computational cost is quantified using the metric of # AND gates. However, the methodology or criteria for determining the number of AND gates associated with a nonlinear operation remains ambiguous in the paper. Clarifying this calculation or providing a concise explanation within the main content would enhance the reader's understanding.\n\n4.\tThe experimental validation is primarily focused on small-scale datasets, which could limit the generalizability of the findings. For a comprehensive evaluation, it would be beneficial for the authors to include results from large-scale datasets like ImageNet. Such experiments would provide a more holistic view of the method's efficacy and scalability.\n\n5.\tBased on Figure 2, both MPCViT and MPCViT+ seem to outperform the proposed method in achieving a superior Pareto balance between accuracy and latency. This observation raises the question: what distinctive advantages or contributions does the proposed method offer? Is it perhaps more streamlined in terms of training efficiency or some other metric? A clearer delineation of the method's unique strengths would be beneficial for readers.\n\n6.\tObservations from Table 4 present some anomalies. Specifically, for PriViT-R, there is a counterintuitive trend where increased latency corresponds to a decline in performance. This is perplexing as one would typically expect a trade-off where longer computation times would yield better results. An explanation or insight into this apparent contradiction would enhance the clarity of the findings.\n\n7.\tThe clarity of notational conventions could be enhanced. For instance, the terms PriViT-R and PriViT-G appear in the text without explicit definitions or context. Providing a clear explanation or description of what these notations signify would aid in a more comprehensive understanding of the content for readers.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698586064397}], "openreview_url": "https://openreview.net/forum?id=8w6FzR68DS", "arxiv_id": "2310.04604", "paper_pdf": "papers/8w6FzR68DS.pdf", "paper_pdf_sha256": "5a132caef3a0e0475bc4d157c4529b71fb4b56a5e898a86a2c8184082e44ae7a", "paper_pdf_bytes": 686634, "paper_pdf_source": "openreview", "code_url": "https://github.com/NYU-DICE-Lab/privit", "code_repository": "NYU-DICE-Lab/privit", "code_commit": "6e4ef9071761ea25bb496ac989dc5c25256a1c71", "code_archive": "repos/8w6FzR68DS.zip", "code_archive_sha256": "75f261a43e2c687fade1dcd0336dc3a19488c823228b13851a6e94e1d3ac8245", "code_archive_bytes": 99188, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 94, "github_languages": {"Python": 142319, "Shell": 2217}, "github_archived": false, "github_pushed_at": "2024-09-10T16:06:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/privit-vision-transformers-for-fast-private"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TtxOsdYU92d", "year": 2023, "status": "rejected", "title": "Generalized Sum Pooling for Metric Learning", "authors": ["Yeti Z. Gürbüz", "Ozan Sener", "A. Aydın Alatan"], "authorids": ["~Yeti_Z._Gürbüz1", "~Ozan_Sener1", "~A._Aydın_Alatan1"], "authors_source": "OpenReview API", "abstract": "A common architectural choice for deep metric learning is a convolutional neural network followed by global average pooling (GAP). Albeit simple, GAP is a highly effective way to aggregate information. One possible explanation for the effectiveness of GAP is considering each feature vector as representing a different semantic entity and GAP as a convex combination of them. Following this perspective, we generalize GAP and propose a learnable generalized sum pooling method (GSP). GSP improves GAP with two distinct abilities: i) the ability to choose a subset of semantic entities, effectively learning to ignore nuisance information, and ii) learning the weights corresponding to the importance of each entity. Formally, we propose an entropy-smoothed optimal transport problem and show that it is a strict generalization of GAP, \\ie a specific realization of the problem gives back GAP. We show that this optimization problem enjoys analytical gradients enabling us to use it as a direct learnable replacement for GAP. We further propose a zero-shot loss to ease the learning of GSP. We show the effectiveness of our method with extensive evaluations on 4 popular metric learning benchmarks. Code is available at: GSP-DML Framework", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "p-ig_sPYSv1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1756/Reviewer_ms1e"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new pooling technique named generalized sum pooling. It is claimed as a strict generalization of global average pooling. The experiments are conducted on four different datasets and achieve a SOTA performance.", "review_text": "Based on my above point, I think this paper is novel and sound, but I do not fully check it through. \nWhile it still misses some key points which makes me confuse.", "strengths": "I'm not an expert in this area, please correct me if I get it wrong. \n\nStrength:\n1. The proposed method seems to be sound, but I'm not fully checking it. The experimental results are outstanding \n2. The motivation seems to be reasonable.\n\nWeakness:\n1. While the motivation is reasonable, it still confuses me why we should redefine the learnable generalization of GAP as an optimization problem. What is the optimization objective?\n2. It seems like there are three categories of related works that are mentioned in this paper,  when selecting the baselines in your experiments, do you cover all of those three categories?\n3. Given the complicated structure of the GSP (i.e., Fig 2), how to can apply it to current existing metric learning or CV-related works?\n4. It is not clear to me, why we should bring the OT into this paper?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new pooling technique named generalized sum pooling. It is claimed as a strict generalization of global average pooling. The experiments are conducted on four different datasets and achieve a SOTA performance.", "strength_and_weaknesses": "I'm not an expert in this area, please correct me if I get it wrong. \n\nStrength:\n1. The proposed method seems to be sound, but I'm not fully checking it. The experimental results are outstanding \n2. The motivation seems to be reasonable.\n\nWeakness:\n1. While the motivation is reasonable, it still confuses me why we should redefine the learnable generalization of GAP as an optimization problem. What is the optimization objective?\n2. It seems like there are three categories of related works that are mentioned in this paper,  when selecting the baselines in your experiments, do you cover all of those three categories?\n3. Given the complicated structure of the GSP (i.e., Fig 2), how to can apply it to current existing metric learning or CV-related works?\n4. It is not clear to me, why we should bring the OT into this paper?", "clarity,_quality,_novelty_and_reproducibility": "Code is not available at this stage, - a password is required.\n\nI'm not an expert on this area, hence I would it seems to be novel.\n\n", "summary_of_the_review": "Based on my above point, I think this paper is novel and sound, but I do not fully check it through. \nWhile it still misses some key points which makes me confuse.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N.A.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666696979337}, {"id": "Ou2ntBdIh_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1756/Reviewer_twxc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "An extension of the global average pooling (GAP), called generalized sum pooling (GSP), is presented in this paper: while GAP is a convex combination with \"flat weights,\" GSP instead learns the weights by solving an optimal transportation problem. GSP allows for the selection of features to be pooled by introducing a \"background prototype\" (rho) into the formulation. Experiments compared GSP with several existing metric learning and pooling methods. The results showed that GSP gives small improvements in several cases.", "review_text": "The idea of using background prototypes is interesting and thorough experiments have been reported. \nThe major problems with this paper are the lack of novelty in the algorithm and the limited effectiveness of the proposed method. Also, the proposed method has many hyperparameters, which may be a drawback for its use in practical scenarios. Given that both novelty and effectiveness are important, I am leaning toward rejection for now.\n", "strengths": "------------------- Strengths -------------------\n\n1. Introducing background prototype in optimal transportation problem is interesting.\n\n\n2. Thorough experiments are reported (in Supplementary Material).\n\n\n3. The paper is mostly well-written.\n\n\n4. Code is provided.\n\n------------------ Weaknesses ------------------\n\n1. Novelty\n\nAlgorithmic novelty may not be sufficient. As discussed in the paper, the idea of using optimal transportation formulation for feature pooling was proposed in Mialon et al. (2021), and the proposed problem (P1) is equivalent to a standard optimal transportation problem (by considering rho as a new pi_i, we can immediately get to the standard form). Given these, using entropy regularization, which is an efficient approach to solving the optimal transportation problem, is straightforward. (P2) is also equivalent to the standard form of the optimal transportation problem. The same is applied to the algorithm.\n\n\n2. Motivation\n\nWhile this paper focuses on metric learning, the application of pooling in general should not be limited to this. So far, I could not find a good reason for this limitation. I would expect an explanation if any.\n\n\n3. Evaluation of Rho\n\nThe point of GSP is to use rho, in other words, changing the range of i in (P1) 1<=i<=m to 1<=i<=m+1. So a before and after comparison between these two cases would be necessary.\n\n\n4. Overall Performance\n\nThe results of comparisons with various pooling methods reported in Supplementary Material (e.g., Table 2) show that the stand-alone performance of GSP is outperformed by several existing methods.\n\n\n5. Hyperparameter Setting\n\nGSP, together with zero-shot regularization, has many hyperparameters to be tuned (m, mu, epsilon, k, lambda). Since these were likely tuned for each dataset and task, it would not have been very difficult to achieve equal or better accuracy than other methods.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "An extension of the global average pooling (GAP), called generalized sum pooling (GSP), is presented in this paper: while GAP is a convex combination with \"flat weights,\" GSP instead learns the weights by solving an optimal transportation problem. GSP allows for the selection of features to be pooled by introducing a \"background prototype\" (rho) into the formulation. Experiments compared GSP with several existing metric learning and pooling methods. The results showed that GSP gives small improvements in several cases.", "strength_and_weaknesses": "------------------- Strengths -------------------\n\n1. Introducing background prototype in optimal transportation problem is interesting.\n\n\n2. Thorough experiments are reported (in Supplementary Material).\n\n\n3. The paper is mostly well-written.\n\n\n4. Code is provided.\n\n------------------ Weaknesses ------------------\n\n1. Novelty\n\nAlgorithmic novelty may not be sufficient. As discussed in the paper, the idea of using optimal transportation formulation for feature pooling was proposed in Mialon et al. (2021), and the proposed problem (P1) is equivalent to a standard optimal transportation problem (by considering rho as a new pi_i, we can immediately get to the standard form). Given these, using entropy regularization, which is an efficient approach to solving the optimal transportation problem, is straightforward. (P2) is also equivalent to the standard form of the optimal transportation problem. The same is applied to the algorithm.\n\n\n2. Motivation\n\nWhile this paper focuses on metric learning, the application of pooling in general should not be limited to this. So far, I could not find a good reason for this limitation. I would expect an explanation if any.\n\n\n3. Evaluation of Rho\n\nThe point of GSP is to use rho, in other words, changing the range of i in (P1) 1<=i<=m to 1<=i<=m+1. So a before and after comparison between these two cases would be necessary.\n\n\n4. Overall Performance\n\nThe results of comparisons with various pooling methods reported in Supplementary Material (e.g., Table 2) show that the stand-alone performance of GSP is outperformed by several existing methods.\n\n\n5. Hyperparameter Setting\n\nGSP, together with zero-shot regularization, has many hyperparameters to be tuned (m, mu, epsilon, k, lambda). Since these were likely tuned for each dataset and task, it would not have been very difficult to achieve equal or better accuracy than other methods.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: This paper is mostly well-written. \n\n\nQuality: Experiments do not successfully show the effectiveness of GSP. The standalone performance of GSP is not better than several existing pooling methods.\n\n\nNovelty: Algorithmic novelty is lacking. The idea of learning pooling layer with optimal transportation has already been proposed. The problems (P1, P2) are equivalent to the basic optimal transportation formulations, so the solution is also not novel.\n\n\nReproducibility: Code is provided.\n", "summary_of_the_review": "The idea of using background prototypes is interesting and thorough experiments have been reported. \nThe major problems with this paper are the lack of novelty in the algorithm and the limited effectiveness of the proposed method. Also, the proposed method has many hyperparameters, which may be a drawback for its use in practical scenarios. Given that both novelty and effectiveness are important, I am leaning toward rejection for now.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666320863193}, {"id": "1o5xNH1AidW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1756/Reviewer_MQSC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a learnable and differentiable pooling operation to replace the global average pooling (GAP) for metric learning.   \n\nSpecifically, the pooling operation is formulated as the solution to an optimal transport problem with additional residual $\\rho$ and mass ration $\\mu$ parameters. $\\rho$ represents the per-element residual, and $\\mu$ stands for the total assignment mass.   \nThe authors derived an iterative solution for forward computation and an inversion-free computation of gradient for backward computation. The closed-form backward gradient formulation is efficient for network training.  \n\nMoreover, to improve the transferring property, cross-batch zero-shot regularization is proposed by predicting class embedding vectors from the prototype assignment vectors ($\\pi$). The sub-problem (P3) has a closed-form solution. Since real class embedding vectors are not given, a meta-learning scheme utilized a zero-shot prediction loss to tackle this. \n\nIn the experiments, toy experiments are designed to investigate the mechanism of the proposed GSP. In metric learning experiments, various pooling alternatives are compared, and GSP achieves the best performance. In addition, GSP is evaluated under different architectures and methods, verifying its effectiveness. Regarding the absolute improvements, the general boost is around or below 1% in most cases, but the improvements are consistent across all settings. \n\n  \n", "review_text": "The overall quality and novelty of this paper are good, except for several concerns listed above. The experiments can verify the effectiveness of the proposed method.   \nI would like to see this paper accepted if the concerns and issues are solved by the authors. \n", "strengths": "## Strength\n- The proposed OT optimization problem in (P1) is novel and reasonable as an alternative to GAP. The introduction of $\\rho$ and $\\mu$ are the keys to model nuisance features and dynamic re-weighting. \n- Both forward and backward solutions are given for an entropy-regularized version of (P1), i.e., (P2). Moreover, the closed-form and inversion-free formulation of the partial derivatives enables differentiable and efficient backward computation. \n- A zero-shot regularization auxiliary task is proposed to improve the transferring ability of metric learning. A simple formulation with a closed-form solution is devised, and a meta-learning procedure is proposed to enable training without ground truth label embeddings. \n- The toy experiments corroborate the design intuitions of GSP and zero-shot regularization, providing useful insights. The comparison experiments are conducted on various datasets, with 13 pooling alternatives and with various backbone/methods, verifying the general effectiveness. \n\n## Weakness\n- One major concern is the extra parameters and computation complexities induced by the proposed method. For example, extra parameters involve $\\{ \\omega, \\Upsilon \\}$ and extra computations involve Proposition 4.1 ($k=100$ forward iteration) and Proposition 4.2 (backward computation). \n- For the deriving of (4.2), important details are missing, e.g., $\\frac{\\partial \\mathcal{L}}{\\partial \\rho^{(\\epsilon)}}$. In $\\frac{\\partial \\mathcal{L}}{\\partial c}$, which kind of $\\mathcal{L}$ is used? In Proposition 4.2, the first $\\mathcal{1}_n$ should be $\\mathcal{1}_m$, and $\\gamma$ is not defined. All these make 4.2 incomplete. \n- I can infer the procedure of meta-learning in zero-shot regularization from Figure 2. However, Section 4.3 is a bit confusing, especially (4.4). It will be much better if an algorithm box is provided. \n- The author mentioned several times the similarity with the top-$k$ selection process. However, no discussion or formulation analysis is provided to show the relationship explicitly. \n- In Sec 4.2, \"this function is not smooth\" why? Is it because of the max operation in normalization? \n- There are several typos, especially in the appendix regarding Preposition 4.1. \n \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a learnable and differentiable pooling operation to replace the global average pooling (GAP) for metric learning.   \n\nSpecifically, the pooling operation is formulated as the solution to an optimal transport problem with additional residual $\\rho$ and mass ration $\\mu$ parameters. $\\rho$ represents the per-element residual, and $\\mu$ stands for the total assignment mass.   \nThe authors derived an iterative solution for forward computation and an inversion-free computation of gradient for backward computation. The closed-form backward gradient formulation is efficient for network training.  \n\nMoreover, to improve the transferring property, cross-batch zero-shot regularization is proposed by predicting class embedding vectors from the prototype assignment vectors ($\\pi$). The sub-problem (P3) has a closed-form solution. Since real class embedding vectors are not given, a meta-learning scheme utilized a zero-shot prediction loss to tackle this. \n\nIn the experiments, toy experiments are designed to investigate the mechanism of the proposed GSP. In metric learning experiments, various pooling alternatives are compared, and GSP achieves the best performance. In addition, GSP is evaluated under different architectures and methods, verifying its effectiveness. Regarding the absolute improvements, the general boost is around or below 1% in most cases, but the improvements are consistent across all settings. \n\n  \n", "strength_and_weaknesses": "## Strength\n- The proposed OT optimization problem in (P1) is novel and reasonable as an alternative to GAP. The introduction of $\\rho$ and $\\mu$ are the keys to model nuisance features and dynamic re-weighting. \n- Both forward and backward solutions are given for an entropy-regularized version of (P1), i.e., (P2). Moreover, the closed-form and inversion-free formulation of the partial derivatives enables differentiable and efficient backward computation. \n- A zero-shot regularization auxiliary task is proposed to improve the transferring ability of metric learning. A simple formulation with a closed-form solution is devised, and a meta-learning procedure is proposed to enable training without ground truth label embeddings. \n- The toy experiments corroborate the design intuitions of GSP and zero-shot regularization, providing useful insights. The comparison experiments are conducted on various datasets, with 13 pooling alternatives and with various backbone/methods, verifying the general effectiveness. \n\n## Weakness\n- One major concern is the extra parameters and computation complexities induced by the proposed method. For example, extra parameters involve $\\{ \\omega, \\Upsilon \\}$ and extra computations involve Proposition 4.1 ($k=100$ forward iteration) and Proposition 4.2 (backward computation). \n- For the deriving of (4.2), important details are missing, e.g., $\\frac{\\partial \\mathcal{L}}{\\partial \\rho^{(\\epsilon)}}$. In $\\frac{\\partial \\mathcal{L}}{\\partial c}$, which kind of $\\mathcal{L}$ is used? In Proposition 4.2, the first $\\mathcal{1}_n$ should be $\\mathcal{1}_m$, and $\\gamma$ is not defined. All these make 4.2 incomplete. \n- I can infer the procedure of meta-learning in zero-shot regularization from Figure 2. However, Section 4.3 is a bit confusing, especially (4.4). It will be much better if an algorithm box is provided. \n- The author mentioned several times the similarity with the top-$k$ selection process. However, no discussion or formulation analysis is provided to show the relationship explicitly. \n- In Sec 4.2, \"this function is not smooth\" why? Is it because of the max operation in normalization? \n- There are several typos, especially in the appendix regarding Preposition 4.1. \n \n", "clarity,_quality,_novelty_and_reproducibility": "The proposed Generalized Sum Pooling (GSP) operation and zero-shot regularization are novel, and a feasible solution is given to the optimization problem. \n\nExperiments support the effectiveness of the proposed GSP and zero-shot regularization methods. \n\nCode is provided for reproduction. \n\nOverall, the quality of this is good, except for some concerns listed in the above Weaknesses. ", "summary_of_the_review": "The overall quality and novelty of this paper are good, except for several concerns listed above. The experiments can verify the effectiveness of the proposed method.   \nI would like to see this paper accepted if the concerns and issues are solved by the authors. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1665897978707}], "openreview_url": "https://openreview.net/forum?id=TtxOsdYU92d", "arxiv_id": "2308.09228", "paper_pdf": "papers/TtxOsdYU92d.pdf", "paper_pdf_sha256": "710ea533c1d9a337fc2326f9dea6804c7011cebe671029b572119f63131654f5", "paper_pdf_bytes": 4033299, "paper_pdf_source": "openreview", "code_url": "https://github.com/yetigurbuz/generalized-sum-pooling", "code_repository": "yetigurbuz/generalized-sum-pooling", "code_commit": "b765b8890f82bc01e55ed32e7432a0fd71a1d4a6", "code_archive": "repos/TtxOsdYU92d.zip", "code_archive_sha256": "b0892871dd4a27aa7557a04d84ecfcdbcfb1785091e4d09f9ee9dee2c5715fe4", "code_archive_bytes": 169913, "code_file_count": 93, "code_extensions": {".py": 92, ".sh": 1}, "github_disk_usage_kb": 191, "github_languages": {"Python": 556723, "Shell": 721}, "github_archived": false, "github_pushed_at": "2023-08-12T15:51:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalized-sum-pooling-for-metric-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZDYhm_o8MX", "year": 2022, "status": "rejected", "title": "Neural Manifold Clustering and Embedding", "authors": ["ZENGYI LI", "Yubei Chen", "Yann LeCun", "Friedrich Sommer"], "authorids": ["~ZENGYI_LI1", "~Yubei_Chen1", "~Yann_LeCun1", "~Friedrich_Sommer1"], "authors_source": "OpenReview API", "abstract": "Given a union of non-linear manifolds, non-linear subspace clustering or manifold clustering aims to cluster data points based on manifold structures and also learn to parameterize each manifold as a linear subspace in a feature space. Deep neural networks have the potential to achieve this goal under highly non-linear settings given their large capacity and flexibility. We argue that achieving manifold clustering with neural networks requires two essential ingredients: a domain-specific constraint that ensures the identification of the manifolds, and a learning algorithm for embedding each manifold to a linear subspace in the feature space. This work shows that many constraints can be implemented by data augmentation. For subspace feature learning, Maximum Coding Rate Reduction (MCR$^2$) objective can be used. Putting them together yields Neural Manifold Clustering and Embedding (NMCE), a novel method for general purpose manifold clustering, which significantly outperforms autoencoder-based deep subspace clustering and achieve state-of-the-art performance on several important benchmarks. Further, on more challenging natural image datasets, NMCE can also outperform other algorithms specifically designed for clustering. Qualitatively, we demonstrate that NMCE learns a meaningful and interpretable feature space. As the formulation of NMCE is closely related to several important Self-supervised learning (SSL) methods, we believe this work can help us build a deep understanding on SSL representation learning.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "aqy5DuGDr-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1416/Reviewer_HTFw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors developed a method and training procedure for manifold clustering problems. The proposed solution inspired from information theoretic methods namely maximum rate reduction. Authors also support their claims with empirical results.", "review_text": "Manifold clustering is a very hard problem. Authors propose a solution for non-linear manifold clustering. The aim of the paper is to learn a representation space in which the manifolds are co-linear. The problem is quite interesting and there are practical applications. However, I have several concerns:\n- I believe two highly relevant papers are missing in the paper and these papers are\n[1] Clustering-friendly Representation Learning via Instance Discrimination and Feature Decorrelation (ICLR 2021)\n[2] Representation Learning for Clustering via Building Consensus (arxiv, May 2021)\nI am completely aware these papers don’t consider manifold clustering however they can be used in this setting as they are. I will lay out my concerns by referring to these two missing papers. \nAuthors rightfully highlights the importance of assumptions and constraints in their paper. One of the proposed constrained is orthogonality (Section 3.1). Both [a] and [b] utilizes orthogonality. [1] directly use originality in their loss and [2] uses random projections. \nAuthors are trying to learn one representation space for different manifolds. Similar objective can be achieved by consensus loss of [2]. \n\n- The objective given in eq 3 is interesting. In my understanding, objective wants to condense each manifold as much as possible (summation term) and all manifolds should cover the largest available space (first term). I think these conditions will be satisfied if all manifolds are equidistant from each other. This can be desired property however how realistic it is? For example, let’s consider the following classes {car, bus, motorbike, bicycle, lion, mountain lion, whale, lion}. Why should each manifold be equidistant from each other? One would like to see semantic structure in the learned representation. Would you please explain how will eq 3 type loss will handle this issue? \n\n-\tThe proposed loss depends on logdet (please see Eq 2). As far as I know, to calculate such a quantity, one needs to calculate the eigen values of the cov(Z). I am assuming that the authors are calculating this quantity for each batch. Since eigen values needs to get calculated by another algorithm, e.g. Gaussian elimination, \no\tdoes this mean the proposed loss cannot generalize to large batch sizes? Some of the self-supervised learning methods are trained by using 8192 batch size. \no\thow can one compute logdet in multiple gpus? Does one need to do this computation on one gpu?\n\nExperimental results:\n-\tWould you please compare your CIFAR-10, CIFAR-20 and STL-10 results with [1] and [2]? Please note that the results shown in [1] and [2] may be using different architectures. As far as I remember [1] use Resnet-34 and [2] use Resnet-18.\n-\tBoth [1] and [2] are end-to-end training methods. Especially [2] supplies better results for Resnet-18 setting. Would you please elaborate the advantage of multi-stage training given that end-to-end are suppling similar results?\n-\tWould you please extend your empirical studt to ImageNet-10 and ImageNet-Dogs.\n-\tWould you please supply some ablation studies e.g. impact of lambda parameter? \n\nI am looking forward to author responses and I am willing to change my assesment/score if the reponse(s) is(are) satisfactory.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Authors developed a method and training procedure for manifold clustering problems. The proposed solution inspired from information theoretic methods namely maximum rate reduction. Authors also support their claims with empirical results.", "main_review": "Manifold clustering is a very hard problem. Authors propose a solution for non-linear manifold clustering. The aim of the paper is to learn a representation space in which the manifolds are co-linear. The problem is quite interesting and there are practical applications. However, I have several concerns:\n- I believe two highly relevant papers are missing in the paper and these papers are\n[1] Clustering-friendly Representation Learning via Instance Discrimination and Feature Decorrelation (ICLR 2021)\n[2] Representation Learning for Clustering via Building Consensus (arxiv, May 2021)\nI am completely aware these papers don’t consider manifold clustering however they can be used in this setting as they are. I will lay out my concerns by referring to these two missing papers. \nAuthors rightfully highlights the importance of assumptions and constraints in their paper. One of the proposed constrained is orthogonality (Section 3.1). Both [a] and [b] utilizes orthogonality. [1] directly use originality in their loss and [2] uses random projections. \nAuthors are trying to learn one representation space for different manifolds. Similar objective can be achieved by consensus loss of [2]. \n\n- The objective given in eq 3 is interesting. In my understanding, objective wants to condense each manifold as much as possible (summation term) and all manifolds should cover the largest available space (first term). I think these conditions will be satisfied if all manifolds are equidistant from each other. This can be desired property however how realistic it is? For example, let’s consider the following classes {car, bus, motorbike, bicycle, lion, mountain lion, whale, lion}. Why should each manifold be equidistant from each other? One would like to see semantic structure in the learned representation. Would you please explain how will eq 3 type loss will handle this issue? \n\n-\tThe proposed loss depends on logdet (please see Eq 2). As far as I know, to calculate such a quantity, one needs to calculate the eigen values of the cov(Z). I am assuming that the authors are calculating this quantity for each batch. Since eigen values needs to get calculated by another algorithm, e.g. Gaussian elimination, \no\tdoes this mean the proposed loss cannot generalize to large batch sizes? Some of the self-supervised learning methods are trained by using 8192 batch size. \no\thow can one compute logdet in multiple gpus? Does one need to do this computation on one gpu?\n\nExperimental results:\n-\tWould you please compare your CIFAR-10, CIFAR-20 and STL-10 results with [1] and [2]? Please note that the results shown in [1] and [2] may be using different architectures. As far as I remember [1] use Resnet-34 and [2] use Resnet-18.\n-\tBoth [1] and [2] are end-to-end training methods. Especially [2] supplies better results for Resnet-18 setting. Would you please elaborate the advantage of multi-stage training given that end-to-end are suppling similar results?\n-\tWould you please extend your empirical studt to ImageNet-10 and ImageNet-Dogs.\n-\tWould you please supply some ablation studies e.g. impact of lambda parameter? \n\nI am looking forward to author responses and I am willing to change my assesment/score if the reponse(s) is(are) satisfactory.", "summary_of_the_review": "Although authors proposed an interesting solution to a very hard problem, I believe proposed method and experimental work needs to be improved especially taking in to account missing references and supplying a detailed discussion. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636005534385}, {"id": "cgPBkuw0aZ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1416/Reviewer_SD19"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work proposed a general manifold clustering algorithm called Neural Manifold Clustering and Embedding (NMCE), which utilize Maximum Coding Rate Reduction (MCR2 ) as the objective function and data augmentation to enforce constrains.  \n\nIn the implementation stage, given that even the toy experiment is difficult to optimize with the full NMCE objective, a multistage training procedure is applied with the first stage actually trying to optimize the Total Coding Rate (TCR), which is another kind of self-supervised learning objective claimed in this paper. \n\nOn synthetic and real-world datasets COIL20, COIL100, CIFAR-10, CIFAR-20 and STL-10, NMCE achieved comparable and sometimes better results, compared to baseline methods or some alternative manifold clustering methods listed in the paper.  ", "review_text": "### Strengths & Originality: \n\nThe idea of combing \"Maximum Coding Rate Reduction\" with data argumentation techniques is interesting and seems reasonable. \n\nIn the experimental parts, this work have results on a number of real data sets, which can be considerate a fairly enough set up to support the effectiveness of a newly proposed algorithm. \n\n### Weakness: \n\nOverall it seems NMCE is like a relative less significant changes from the work from recent proposed \"Maximum Coding Rate Reduction (MCR2) (Yu et al., 2020)\" with some data argument techniques, and there is no interesting theoretical analysis behind the proposed NMCE work too, from this paper.\n\nAlternative methods: in the experimental results & section 4, there are a number works are cited and compared to NMCE but still lack some important references. For example in Table 1, for COIL data set, results from linear methods like SSC and some deep clustering methods are listed but there is a whole field of nonlinear manifold clustering is not mentioned here. Reference 1 and 2 listed below are two methods in the field of nonlinear manifold clustering and some in-depth discussion or comparison to them are quite helpful for reviewer to evaluate the contribution from NMCE. Indeed, there is one survey paper about nonlinear manifold clustering is cited in section 2 but this is not enough. \n\nTo the toy example, double spiral, it seems NMCE can get 100% accuracy as shown in Appendix A. However, only one 100% number is kind of less clear as we can imagine that if keep increasing the scale of Gaussian noise in the data argumentation step, we should except this 100% to go down. Or maybe we can say the key is how to define \"a small amount of noise\" as stated in the paper, and should be nice to see the performance from alternative methods on toy examples too, to gave people more insights, i.e., multi-linear subspace clustering methods clearly not work well for this toy example as we know. \n\n### Reference:\n1. Souvenir, R., & Pless, R, Manifold clustering. The10th International Conference on Computer Vision (ICCV 2005).\n2. Dian Gong, Xuemei Zhao, Gérard G. Medioni, Robust Multiple Manifold Structure Learning. ICML 2012.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposed a general manifold clustering algorithm called Neural Manifold Clustering and Embedding (NMCE), which utilize Maximum Coding Rate Reduction (MCR2 ) as the objective function and data augmentation to enforce constrains.  \n\nIn the implementation stage, given that even the toy experiment is difficult to optimize with the full NMCE objective, a multistage training procedure is applied with the first stage actually trying to optimize the Total Coding Rate (TCR), which is another kind of self-supervised learning objective claimed in this paper. \n\nOn synthetic and real-world datasets COIL20, COIL100, CIFAR-10, CIFAR-20 and STL-10, NMCE achieved comparable and sometimes better results, compared to baseline methods or some alternative manifold clustering methods listed in the paper.  ", "main_review": "### Strengths & Originality: \n\nThe idea of combing \"Maximum Coding Rate Reduction\" with data argumentation techniques is interesting and seems reasonable. \n\nIn the experimental parts, this work have results on a number of real data sets, which can be considerate a fairly enough set up to support the effectiveness of a newly proposed algorithm. \n\n### Weakness: \n\nOverall it seems NMCE is like a relative less significant changes from the work from recent proposed \"Maximum Coding Rate Reduction (MCR2) (Yu et al., 2020)\" with some data argument techniques, and there is no interesting theoretical analysis behind the proposed NMCE work too, from this paper.\n\nAlternative methods: in the experimental results & section 4, there are a number works are cited and compared to NMCE but still lack some important references. For example in Table 1, for COIL data set, results from linear methods like SSC and some deep clustering methods are listed but there is a whole field of nonlinear manifold clustering is not mentioned here. Reference 1 and 2 listed below are two methods in the field of nonlinear manifold clustering and some in-depth discussion or comparison to them are quite helpful for reviewer to evaluate the contribution from NMCE. Indeed, there is one survey paper about nonlinear manifold clustering is cited in section 2 but this is not enough. \n\nTo the toy example, double spiral, it seems NMCE can get 100% accuracy as shown in Appendix A. However, only one 100% number is kind of less clear as we can imagine that if keep increasing the scale of Gaussian noise in the data argumentation step, we should except this 100% to go down. Or maybe we can say the key is how to define \"a small amount of noise\" as stated in the paper, and should be nice to see the performance from alternative methods on toy examples too, to gave people more insights, i.e., multi-linear subspace clustering methods clearly not work well for this toy example as we know. \n\n### Reference:\n1. Souvenir, R., & Pless, R, Manifold clustering. The10th International Conference on Computer Vision (ICCV 2005).\n2. Dian Gong, Xuemei Zhao, Gérard G. Medioni, Robust Multiple Manifold Structure Learning. ICML 2012.", "summary_of_the_review": "Overall I feel this is a reasonable submission but seems not good enough for ICLR publication, so I gave \" marginally below the acceptance threshold \". ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635905638422}, {"id": "VkYKkA0wOf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1416/Reviewer_aRVQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of clustering a dataset corresponding to a sampling of a union of nonlinear manifolds where each cluster shall correspond to an individual manifold. The proposed method leverages previous results on subspace clustering via neural networks and a penalty function based on concordant classification of samples and their augmented versions.", "review_text": "The paper provides an intuitive extension of the MCR^2 embedding approach by regularizing it with a \"concordant data augmentation embedding\" constraint. The authors reason that such a constraint is needed to \"[restrict] the large flexibility of neural networks\".\n\nHowever, the paper is imprecise in many of the statements made. A reference to \"additional geometric constraints that make the manifold clusters identifiable\" essentially reduces to proximity of samples and their noisy versions in the manifold embedding. $L_{clst}$ is described as \"some objective function that will force $f$ to cluster the dataset\" but again there is no practical example of such a function. The concept of a constraint functional $D(f)$ is also never specified for a practical setting. In the appendix the authors say \"the performance of [MCR^2] is rather poor, which is expected based on our understanding\" - this does not illuminate a reason for the reader or make it clear what \"CTRL\" brings in.\n\nIt is not clear why only a partial subset of the code that implements the experiments is provided with thee rest to be held for release \"upon publication\". Also there are no descriptions of the computation cost of the different approaches compared, including the proposed one.\n\nThe fact that the augmentation/regularization used in the numerical experiments is tuned by hand raises further questions about the robustness of the proposed approach. The authors should provide some discussion on this aspect of the implementation - this can be done in the appendix if needed.\n\nThe training procedures used in Section 4.2 appear to be ad-hoc - can the authors comment on their reasoning? Additionally, some exclusions in the numerical comparison seem arbitrary as well - if a method in the literature addressed the same problem considered here, why not compare against it? If there are differences in approaches they can still be mentioned to provide a contrast.\n\nWhy is there no numerical comparison in the experiments shown on the appendix? \n\nThere is a repeated typo (e.g., in Fig. 2 and Fig. A.3) captions \"principle components\" -> \"principal components\".", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of clustering a dataset corresponding to a sampling of a union of nonlinear manifolds where each cluster shall correspond to an individual manifold. The proposed method leverages previous results on subspace clustering via neural networks and a penalty function based on concordant classification of samples and their augmented versions.", "main_review": "The paper provides an intuitive extension of the MCR^2 embedding approach by regularizing it with a \"concordant data augmentation embedding\" constraint. The authors reason that such a constraint is needed to \"[restrict] the large flexibility of neural networks\".\n\nHowever, the paper is imprecise in many of the statements made. A reference to \"additional geometric constraints that make the manifold clusters identifiable\" essentially reduces to proximity of samples and their noisy versions in the manifold embedding. $L_{clst}$ is described as \"some objective function that will force $f$ to cluster the dataset\" but again there is no practical example of such a function. The concept of a constraint functional $D(f)$ is also never specified for a practical setting. In the appendix the authors say \"the performance of [MCR^2] is rather poor, which is expected based on our understanding\" - this does not illuminate a reason for the reader or make it clear what \"CTRL\" brings in.\n\nIt is not clear why only a partial subset of the code that implements the experiments is provided with thee rest to be held for release \"upon publication\". Also there are no descriptions of the computation cost of the different approaches compared, including the proposed one.\n\nThe fact that the augmentation/regularization used in the numerical experiments is tuned by hand raises further questions about the robustness of the proposed approach. The authors should provide some discussion on this aspect of the implementation - this can be done in the appendix if needed.\n\nThe training procedures used in Section 4.2 appear to be ad-hoc - can the authors comment on their reasoning? Additionally, some exclusions in the numerical comparison seem arbitrary as well - if a method in the literature addressed the same problem considered here, why not compare against it? If there are differences in approaches they can still be mentioned to provide a contrast.\n\nWhy is there no numerical comparison in the experiments shown on the appendix? \n\nThere is a repeated typo (e.g., in Fig. 2 and Fig. A.3) captions \"principle components\" -> \"principal components\".", "summary_of_the_review": "While the extension of a supervised multi-manifold embedding to the manifold clustering problem is intuitive and interesting, there are several imprecise concepts underlying its description, and the carefully crafted implementation muddles whether the proposed approach can have wider applicability.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635866335460}, {"id": "VcrU_2jWvw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1416/Reviewer_zSx8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper “Neural Manifold Clustering and Embedding” proposes methods for utilizing a deep neural network for simultaneously learning the clustering of samples and a non-linear embedding into a low-dimensional “latent space” where each cluster representation is simple.\nThe main contributions of this paper are:\n\n (1) Proposing a novel loss function that combines the “maximal coding-rate reduction” principle and self-supervised view of how to use data-augmentation in clustering;\n (2) Proposing an unsupervised method for solving the problem from scratch using a deep neural network as embedding operation, with parameters learned to maximize “coding-rate reduction” while preserving class membership across augmented samples;\n (3) Proposing a supervised method for fine-tuning the results of the above method by minimizing the “coding-rate reduction” of each class while maximizing the overhaul  “coding-rate reduction” and preserving class membership across augmented samples.\n", "review_text": "# Paper strengths #\nThe paper suggests a loss function for learning manifold clustering and embedding by combining two ideas from the literature. First, building on a technical paper introducing the “maximal coding-rate reduction” principle which suggests that to learn a subspace-structured representation, one needs to maximize the difference between the coding rate of all clusters (pooled) and the sum of coding rates for each cluster. Second, building on the insight (shared with related works detailed in the “Clustering with Data Augmentation” section) that data-augmentation provides an important signal for manifold embedding, as preservation of class membership across augmented samples replaces preservation of local geometry from classic manifold learning. While both ideas are not new per se, it is interesting that the current paper outperforms both lines of work by using a combined loss function (named NMCE, as the paper initials), suggesting a fruitful synergy between those two ideas.\nThe optimization target allows using any deep network as a manifold embedding function in either an unsupervised or supervised manner. The unsupervised method (using a simpler loss function, named TCE) nicely resembles recent self-supervised methods, while the supervised method is used as fine-tuning of the unsupervised method and seems to work well.\n\n# Paper weaknesses #\nThe paper structure does not follow the logical structure presented in my summary: while the combined loss function is well described and justified, the two algorithms are presented as an afterthought. It is not clear why the unsupervised method is needed, while the supervised method is not presented at all. Restate the paper contribution to make the distinction between the loss function, the unsupervised, and the supervised methods.\nPresent the unsupervised method by itself (rather than as an empirical note in section 3.6) and elaborate on the comparison to self-supervised methods (currently in “additional results” appendix). Explain why the unsupervised method is needed (my guess is this is due to zeroing out of the NMCE objective if class assignments are randomly initialized)\nThe supervised method is not adequately described, which seems like a major omission. If I correctly filled in the blanks, a training set was used for minimizing the NMCE objective function using the ground-truth class assignment (i.e. only the embedding is learned), then clustering success is measured on a test-set with some unknown initialization of class assignment. This initialization may be important for the results if it is not random (e.g., by using the training set and finding the nearest neighbor). \nNo motivation is given for the role of such work in the ML landscape. The paper starts with a technical assumption (“a union of low-dimensional nonlinear manifolds”) rather than from an intuitive statement that (1) each sample comes from some class; (2) classes representation has a meaningful structure, conceptualized as a low-dimensional nonlinear geometry; (3) it is natural to ask if class-membership and a low-dimensional geometry can be learned simultaneously.\nImportant principles are hidden in technical definitions and the authors should aim to provide more intuition. Most notable are the mysterious references to MCR2. If “Gaussian coding rate” is a quantification of variance, “Maximum Coding Rate Reduction (MCR2)” supports maximizing the total variance while minimizing  “variance within clusters”. Similarly the “constraint functional D(f)” is presented in section 3.2 as a “completely generic” method and only in section 3.5 do we understand how to implement it by enforcing similar clustering of data-augmented samples. I find the generic presentation unhelpful while the use of data-augmentation to uncover manifold structure central to the paper.\nThe main method presented in the paper does not seem to work well on synthetic data sets, and the authors do not share the details. Summarize the difference between EnSC and the full loss functions and explain why “For simple tasks, the features can be directly clustered with standard linear SC techniques such as EnSC”.  Also, provide the (poor) result of the full algorithm.\n\n## Smaller issues ##\n (1) The vogue statement “This work shows that many constraints can be implemented by data augmentation” should be re-stated through the lens of self-supervised learning, e.g., “manifold local structure can be learned by enforcing constant class-assignment on data-augmented samples”\n\n (2) Highlight the difference from the related work in the “Clustering with Data Augmentation” section and state in advance that current work outperforms them.\n\n (3) The statement “we need augmentations that perturbs style information (...) but preserves content information, so that the clustering will be based on content but not style” seems weird as this is almost the definition of data augmentation.\n\n (4) The statement “This indicates that NMCE can truly leverage the non-linear processing capability of deep networks” seems false in the context of section 4.1, where the TCR objective is used and not the full NMCE objective.\n\n (5) In section 4.2.1 the role of stage 2 (“Reinitialize the last linear projection layer and add a cluster assignment layer … freeze parameters in the backbone network and train the two new layers with full NMCE objective”) is not clear. If I understand correctly, you don’t show results of stage 2 without stage 3 and you don’t justify that stage 2 is needed. You should provide some evidence stage 2 is helpful or if not skip it for the sake of simplicity.\n\n (6) I don’t understand the discussion which makes most of section 4.2.1, about using representations at different levels of the embedding network, and with or without averaging. It seems the authors had several options and made their choice, so this section and table 2 can go to the “additional results” appendix to justify the choice.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper “Neural Manifold Clustering and Embedding” proposes methods for utilizing a deep neural network for simultaneously learning the clustering of samples and a non-linear embedding into a low-dimensional “latent space” where each cluster representation is simple.\nThe main contributions of this paper are:\n\n (1) Proposing a novel loss function that combines the “maximal coding-rate reduction” principle and self-supervised view of how to use data-augmentation in clustering;\n (2) Proposing an unsupervised method for solving the problem from scratch using a deep neural network as embedding operation, with parameters learned to maximize “coding-rate reduction” while preserving class membership across augmented samples;\n (3) Proposing a supervised method for fine-tuning the results of the above method by minimizing the “coding-rate reduction” of each class while maximizing the overhaul  “coding-rate reduction” and preserving class membership across augmented samples.\n", "main_review": "# Paper strengths #\nThe paper suggests a loss function for learning manifold clustering and embedding by combining two ideas from the literature. First, building on a technical paper introducing the “maximal coding-rate reduction” principle which suggests that to learn a subspace-structured representation, one needs to maximize the difference between the coding rate of all clusters (pooled) and the sum of coding rates for each cluster. Second, building on the insight (shared with related works detailed in the “Clustering with Data Augmentation” section) that data-augmentation provides an important signal for manifold embedding, as preservation of class membership across augmented samples replaces preservation of local geometry from classic manifold learning. While both ideas are not new per se, it is interesting that the current paper outperforms both lines of work by using a combined loss function (named NMCE, as the paper initials), suggesting a fruitful synergy between those two ideas.\nThe optimization target allows using any deep network as a manifold embedding function in either an unsupervised or supervised manner. The unsupervised method (using a simpler loss function, named TCE) nicely resembles recent self-supervised methods, while the supervised method is used as fine-tuning of the unsupervised method and seems to work well.\n\n# Paper weaknesses #\nThe paper structure does not follow the logical structure presented in my summary: while the combined loss function is well described and justified, the two algorithms are presented as an afterthought. It is not clear why the unsupervised method is needed, while the supervised method is not presented at all. Restate the paper contribution to make the distinction between the loss function, the unsupervised, and the supervised methods.\nPresent the unsupervised method by itself (rather than as an empirical note in section 3.6) and elaborate on the comparison to self-supervised methods (currently in “additional results” appendix). Explain why the unsupervised method is needed (my guess is this is due to zeroing out of the NMCE objective if class assignments are randomly initialized)\nThe supervised method is not adequately described, which seems like a major omission. If I correctly filled in the blanks, a training set was used for minimizing the NMCE objective function using the ground-truth class assignment (i.e. only the embedding is learned), then clustering success is measured on a test-set with some unknown initialization of class assignment. This initialization may be important for the results if it is not random (e.g., by using the training set and finding the nearest neighbor). \nNo motivation is given for the role of such work in the ML landscape. The paper starts with a technical assumption (“a union of low-dimensional nonlinear manifolds”) rather than from an intuitive statement that (1) each sample comes from some class; (2) classes representation has a meaningful structure, conceptualized as a low-dimensional nonlinear geometry; (3) it is natural to ask if class-membership and a low-dimensional geometry can be learned simultaneously.\nImportant principles are hidden in technical definitions and the authors should aim to provide more intuition. Most notable are the mysterious references to MCR2. If “Gaussian coding rate” is a quantification of variance, “Maximum Coding Rate Reduction (MCR2)” supports maximizing the total variance while minimizing  “variance within clusters”. Similarly the “constraint functional D(f)” is presented in section 3.2 as a “completely generic” method and only in section 3.5 do we understand how to implement it by enforcing similar clustering of data-augmented samples. I find the generic presentation unhelpful while the use of data-augmentation to uncover manifold structure central to the paper.\nThe main method presented in the paper does not seem to work well on synthetic data sets, and the authors do not share the details. Summarize the difference between EnSC and the full loss functions and explain why “For simple tasks, the features can be directly clustered with standard linear SC techniques such as EnSC”.  Also, provide the (poor) result of the full algorithm.\n\n## Smaller issues ##\n (1) The vogue statement “This work shows that many constraints can be implemented by data augmentation” should be re-stated through the lens of self-supervised learning, e.g., “manifold local structure can be learned by enforcing constant class-assignment on data-augmented samples”\n\n (2) Highlight the difference from the related work in the “Clustering with Data Augmentation” section and state in advance that current work outperforms them.\n\n (3) The statement “we need augmentations that perturbs style information (...) but preserves content information, so that the clustering will be based on content but not style” seems weird as this is almost the definition of data augmentation.\n\n (4) The statement “This indicates that NMCE can truly leverage the non-linear processing capability of deep networks” seems false in the context of section 4.1, where the TCR objective is used and not the full NMCE objective.\n\n (5) In section 4.2.1 the role of stage 2 (“Reinitialize the last linear projection layer and add a cluster assignment layer … freeze parameters in the backbone network and train the two new layers with full NMCE objective”) is not clear. If I understand correctly, you don’t show results of stage 2 without stage 3 and you don’t justify that stage 2 is needed. You should provide some evidence stage 2 is helpful or if not skip it for the sake of simplicity.\n\n (6) I don’t understand the discussion which makes most of section 4.2.1, about using representations at different levels of the embedding network, and with or without averaging. It seems the authors had several options and made their choice, so this section and table 2 can go to the “additional results” appendix to justify the choice.\n", "summary_of_the_review": "The paper combines two previous lines of work into a novel loss function which is then used in both supervised and unsupervised manners and achieves state-of-the-art results. It is currently below the acceptance threshold because the presented algorithms’ description and justification are lacking.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635837410464}, {"id": "YpbnEuRNN6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1416/Reviewer_dgFk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper enhances the performance of Maximum Coding Rate Reduction (MCR2) objective in subspace feature learning by incorporating a strategy of adding a manifold learning inducing constraint, aiming at a novel method for general purpose manifold clustering. They claim that the new model significantly outperforms autoencoder-based deep subspace clustering and achieves\nstate-of-the-art performance on several important benchmarks. ", "review_text": "Fundamentally the work is based on the existing Maximum Coding Rate Reduction (MCR2) objective (Yu et al., 2020). Through their analysis, to achieve the goal of manifold learning in unsupervised fashion, in principle the authors emphasize the importance and necessity of geometry-awareness constraint to be added upon the MCR2 objective. The constraint is surprisingly implemented by the simple data augmentation for example of adding small noises to the given data. And then apply the intuition that the learnt feature embedding of augmented data points should be similar if they are generated from the same point as the geometric constraint. I think this is a very interesting and useful way to enforce the local structures in the data. The authors show a toy example with their code for the performance of the proposed principle, although no code for reproducing large experiments is provided. It is hard to say what performance could be given the high complexity of data while a simple data-augmentation strategy is used, although the past contrastive learning objective works well in practice. Of course, there is much theoretical gaurantee for the introduced constraint strategy. \n\nThe paper can be further improved by increasing readability.   For example I dont think Definition 1 is a good way to introduce Manifold Clustering and Embedding task.  A definition should be used for the rigorous, concise and accurate \"mathematical\" concepts. For example, the \"perpendicular\" may be written as a definition.  \n\nCan you give some discussion on the dimensions d1,d2,...,dn?  What is their role in the algorithm?  Should they be known or can be learned from the algorithm?  Otherwise, there is no need to present them.   It seems the algorithm has nothing to do with them.\n\nYou may add how cov(Z) is handled in the algorithm?  What is the algorithm complexity given that the model parameters are inside latent variable Z?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper enhances the performance of Maximum Coding Rate Reduction (MCR2) objective in subspace feature learning by incorporating a strategy of adding a manifold learning inducing constraint, aiming at a novel method for general purpose manifold clustering. They claim that the new model significantly outperforms autoencoder-based deep subspace clustering and achieves\nstate-of-the-art performance on several important benchmarks. ", "main_review": "Fundamentally the work is based on the existing Maximum Coding Rate Reduction (MCR2) objective (Yu et al., 2020). Through their analysis, to achieve the goal of manifold learning in unsupervised fashion, in principle the authors emphasize the importance and necessity of geometry-awareness constraint to be added upon the MCR2 objective. The constraint is surprisingly implemented by the simple data augmentation for example of adding small noises to the given data. And then apply the intuition that the learnt feature embedding of augmented data points should be similar if they are generated from the same point as the geometric constraint. I think this is a very interesting and useful way to enforce the local structures in the data. The authors show a toy example with their code for the performance of the proposed principle, although no code for reproducing large experiments is provided. It is hard to say what performance could be given the high complexity of data while a simple data-augmentation strategy is used, although the past contrastive learning objective works well in practice. Of course, there is much theoretical gaurantee for the introduced constraint strategy. \n\nThe paper can be further improved by increasing readability.   For example I dont think Definition 1 is a good way to introduce Manifold Clustering and Embedding task.  A definition should be used for the rigorous, concise and accurate \"mathematical\" concepts. For example, the \"perpendicular\" may be written as a definition.  \n\nCan you give some discussion on the dimensions d1,d2,...,dn?  What is their role in the algorithm?  Should they be known or can be learned from the algorithm?  Otherwise, there is no need to present them.   It seems the algorithm has nothing to do with them.\n\nYou may add how cov(Z) is handled in the algorithm?  What is the algorithm complexity given that the model parameters are inside latent variable Z?", "summary_of_the_review": "The proposed method builds upon the existing existing Maximum Coding Rate Reduction (MCR2) objective with a data augmentation as a constraint to enforce manifold learning. It seems that the performance is satisfactory comparing with other state-of-the-art methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635219957049}], "openreview_url": "https://openreview.net/forum?id=ZDYhm_o8MX", "arxiv_id": "2201.10000", "paper_pdf": "papers/ZDYhm_o8MX.pdf", "paper_pdf_sha256": "acfe723950465e0b5422dbfb67a423935a02243501be2ff6db97457b30ded9ae", "paper_pdf_bytes": 4647410, "paper_pdf_source": "openreview", "code_url": "https://github.com/zengyi-li/NMCE-release", "code_repository": "zengyi-li/NMCE-release", "code_commit": "b60a05eedcaa115b0efeb45686bfae865c0398de", "code_archive": "repos/ZDYhm_o8MX.zip", "code_archive_sha256": "f80ae0ffa96aba9ffe8b69c2e568223a2bbb52f491a8ede503c93308fd05553d", "code_archive_bytes": 368323, "code_file_count": 19, "code_extensions": {".py": 18, ".ipynb": 1}, "github_disk_usage_kb": 352, "github_languages": {"Python": 123762, "Jupyter Notebook": 66360}, "github_archived": false, "github_pushed_at": "2022-03-11T01:52:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-manifold-clustering-and-embedding-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eYgI3cTPTq9", "year": 2021, "status": "rejected", "title": "How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds", "authors": ["Prithviraj Ammanabrolu", "Ethan Tien", "Matthew Hausknecht", "Mark Riedl"], "authorids": ["~Prithviraj_Ammanabrolu1", "~Ethan_Tien1", "~Matthew_Hausknecht1", "~Mark_Riedl1"], "authors_source": "OpenReview API", "abstract": "Text-based games are long puzzles or quests, characterized by a sequence of sparse and potentially deceptive rewards. They provide an ideal platform to develop agents that perceive and act upon the world using a combinatorially sized natural language state-action space. Standard Reinforcement Learning agents are poorly equipped to effectively explore such spaces and often struggle to overcome bottlenecks---states that agents are unable to pass through simply because they do not see the right action sequence enough times to be sufficiently reinforced. We introduce Q*BERT, an agent that learns to build a knowledge graph of the world by answering questions, which leads to greater sample efficiency. To overcome bottlenecks, we further introduce MC!Q*BERT an agent that uses an knowledge-graph-based intrinsic motivation to detect bottlenecks and a novel exploration strategy to efficiently learn a chain of policy modules to overcome them. We present an ablation study and results demonstrating how our method outperforms the current state-of-the-art on nine text games, including the popular game, Zork, where, for the first time, a learning agent gets past the bottleneck where the player is eaten by a Grue.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "xwwdz3c01xU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1887/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies reinforcement learning setting where the agent's decision is augmented with external knowledge representation. Termed Q*BERT, this proposed method uses question answering to build a knowledge graph of the world. Results show the agent was able to pass the bottleneck for a popular game where most other algorithms failed.\n\nPros:\nThe proposed method seems well motivated and reasonable to improve agent's performance. Authors provide good review of text games' literature. Experimental results seem solid.\n\nCons:\n1. The novelty in terms of methodology seems a bit low. The idea of Q*BERT training is not new.\n2. I have some questions regarding authors' \"Knowledge Graph State Representation\". First of all, pretrained language models such as BERT contain rich information about world knowledge. For example, BERT will give a strong association between \"door\" and \"key\", while an agent that learns from scratch will not understand unless it receives an reward from \"use key to open the door\" (or the game hints the agent to do so). It's unclear why the authors argue using QA to augment agent's state representation. It is possible that using BERT's representation alone, the agent's performance will be much better and training will be more efficient already.\n3. Another question relates to the way authors conduct this QA. The authors use an oracle agent to explore the word and a random agent to gather information such as attributes. Is there already information leakage during this process? In a fair experimental setup, there is no way an agent can foresee the world. Maybe I am misunderstanding here, and I hope the authors can help clarify.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach, but have questions regarding experiment setup", "review": "This paper studies reinforcement learning setting where the agent's decision is augmented with external knowledge representation. Termed Q*BERT, this proposed method uses question answering to build a knowledge graph of the world. Results show the agent was able to pass the bottleneck for a popular game where most other algorithms failed.\n\nPros:\nThe proposed method seems well motivated and reasonable to improve agent's performance. Authors provide good review of text games' literature. Experimental results seem solid.\n\nCons:\n1. The novelty in terms of methodology seems a bit low. The idea of Q*BERT training is not new.\n2. I have some questions regarding authors' \"Knowledge Graph State Representation\". First of all, pretrained language models such as BERT contain rich information about world knowledge. For example, BERT will give a strong association between \"door\" and \"key\", while an agent that learns from scratch will not understand unless it receives an reward from \"use key to open the door\" (or the game hints the agent to do so). It's unclear why the authors argue using QA to augment agent's state representation. It is possible that using BERT's representation alone, the agent's performance will be much better and training will be more efficient already.\n3. Another question relates to the way authors conduct this QA. The authors use an oracle agent to explore the word and a random agent to gather information such as attributes. Is there already information leakage during this process? In a fair experimental setup, there is no way an agent can foresee the world. Maybe I am misunderstanding here, and I hope the authors can help clarify.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603943135370}, {"id": "jiZq1PnGM2X", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1887/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper focuses on learning to play text adventure games using reinforcement learning. The paper presents two new algorithms: Q*BERT and MC!Q*BERT, as well as a QA dataset to help training a component of these agents that builds a knowledge graph of the current game state based on the textual descriptions received by from the game.\n\nReasons for score:\n\nI gave this paper a 7, as I think the overall contributions to the text-based adventure game playing literature are strong. There are a few minor technical issues here and there (see my detailed feedback below), but those are minor things that can be easily fixed. In particular, the new approach to integrate exploration strategies to overcome \"bottlenecks\" in the search space is interesting and, to the best of my knowledge, novel.\n\nAdditional feedback:\n\n- page 1: \"We contend that existing Reinforcement Learning agents that are unaware of\" -> \"We contend that existing Reinforcement Learning agents are unaware of\"\n- page 2: nit, \"In text-adventure games the dependencies are of two types\" -> I can recall many bottlenecks from text adventure games that do not fit either of these two categories. So, perhaps instead of \"are of two types\" you could say that these are the two dependencies you considered (e.g., some games require many unrewarded actions (in terms of in-game points) like visiting locations more than once, or visiting them at some particular times, or even interacting with certain characters in certain ways, etc.). Basically, location and inventory are not the only state of the game, but there are many other state variables in many of these games.\n- I am not convinced that the dependency graph of Zork can be divided into a linear set of  \"level\" subdivisions the authors state before Equation 1. For example, how about branches? If there are parallel tasks that need to be performed each with their own \"bottlenecks\" (for example, completing three tasks that can be done in any order), then the \"j>i\" condition in Equation 1 breaks, as it will result in some spurious bottlenecks. Thus, I think this part of the paper needs some work (authors mention \"relatively linear plots\", but \"relative linear\" does not mean \"completely linear\"). Edit after reaching Section 5: and since this definition is not the one used by MC!Q*BERT anyway, why have it in the paper in any case?\n- Table 1: bolding the highest scores for each game would be useful to understand the table at a glance.\n- page 8: about the bottleneck identification rates reported: how were ground truths established? was this manually labeled? If it was automatically labeled, I'm not sure if the definition used earlier in the paper was used, but if it was, I am not confident it is a good definition.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting paper, with some minor issues, but significantly contributing to the state of the art.", "review": "Summary:\n\nThis paper focuses on learning to play text adventure games using reinforcement learning. The paper presents two new algorithms: Q*BERT and MC!Q*BERT, as well as a QA dataset to help training a component of these agents that builds a knowledge graph of the current game state based on the textual descriptions received by from the game.\n\nReasons for score:\n\nI gave this paper a 7, as I think the overall contributions to the text-based adventure game playing literature are strong. There are a few minor technical issues here and there (see my detailed feedback below), but those are minor things that can be easily fixed. In particular, the new approach to integrate exploration strategies to overcome \"bottlenecks\" in the search space is interesting and, to the best of my knowledge, novel.\n\nAdditional feedback:\n\n- page 1: \"We contend that existing Reinforcement Learning agents that are unaware of\" -> \"We contend that existing Reinforcement Learning agents are unaware of\"\n- page 2: nit, \"In text-adventure games the dependencies are of two types\" -> I can recall many bottlenecks from text adventure games that do not fit either of these two categories. So, perhaps instead of \"are of two types\" you could say that these are the two dependencies you considered (e.g., some games require many unrewarded actions (in terms of in-game points) like visiting locations more than once, or visiting them at some particular times, or even interacting with certain characters in certain ways, etc.). Basically, location and inventory are not the only state of the game, but there are many other state variables in many of these games.\n- I am not convinced that the dependency graph of Zork can be divided into a linear set of  \"level\" subdivisions the authors state before Equation 1. For example, how about branches? If there are parallel tasks that need to be performed each with their own \"bottlenecks\" (for example, completing three tasks that can be done in any order), then the \"j>i\" condition in Equation 1 breaks, as it will result in some spurious bottlenecks. Thus, I think this part of the paper needs some work (authors mention \"relatively linear plots\", but \"relative linear\" does not mean \"completely linear\"). Edit after reaching Section 5: and since this definition is not the one used by MC!Q*BERT anyway, why have it in the paper in any case?\n- Table 1: bolding the highest scores for each game would be useful to understand the table at a glance.\n- page 8: about the bottleneck identification rates reported: how were ground truths established? was this manually labeled? If it was automatically labeled, I'm not sure if the definition used earlier in the paper was used, but if it was, I am not confident it is a good definition.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603924371845}, {"id": "HRj8OWzRdwd", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1887/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe paper proposes two RL agents for the text adventure games: QBERT learns (relational) knowledge about the game world with assistance of a trained QA-model and then constrains its action space with help of the knowledge base, enjoying greater sample efficiency; MCQBERT learns to escape from local optima by backtracking through the navigation trajectories, guided by the learned knowledge base and a manually-designed intrinsic reward function. \n\nPros: \n\nThis is a very dense paper with multiple key ideas, namely (1) using QA model to construct knowledge base; (2) escaping local optima by backtracking; and (3) modular policy chaining. \n\nBoth (1) and (2) are interesting and novel. But the novelty of (1) is limited because using QA to build graph has been used by e.g., Ammanabrolu et al. Bringing Stories Alive: Generating Interactive Fiction Worlds. \n\nEmpirical results demonstrate the effectiveness of the proposed methods. Not being eaten by Grue is impressive. \n\nCons: \n\nPresentation is the main reason that I’d like to reject the current version: \n(1) Several key claims are misleading and over-bold. \n(2) Many essential technical details are not clear. \n\n[Misleading/Over-bold claims]\n\nThe authors claimed, at many places, that they learned the \"dependencies'', namely the quest structure. For example, they said the agent \"learns that picking up the lamp is the right way to surpass the Cellar bottleneck and reach the Painting\". \n\nThis claim is wrong: the agent only successfully passes that room because it has done enough exploration (and the exploration is efficient because the intrinsic reward function is cleverly designed).\n\nIf this comment is not clear enough, let me be a bit more specific. \nIn the \"lamp-Grue\" example, the agent doesn't know the sense like \"if I pick up this lamp, I can pass the Grue room'' or \"now I am in the Grue room and the lamp is helpful\". \nInstead, the agent only picks the lamp up because it is encouraged to discover new information. \n\nOr, in other words, the agent doesn't really learn the quest structure as shown in Figure-2. \n(But I agree that Figure-2 is useful to illustrate the interesting idea.)\n\nI believe that the right way to view the method is: exploration is guided by not only environment reward but also intrinsic desire to discover new information, which leads to more efficient exploration and higher long-term reward. \nSimilar problems in other RL domains have been tackled and here is a paper that is similar to this submission in spirit: \nConti et al NeurIPS 2018 Improving Exploration in Evolution Strategies for Deep Reinforcement Learning via a Population of Novelty-Seeking Agents. \nQuote them: \"reward functions are often deceptive, and solely optimizing for reward without some mechanism to encourage intelligent exploration can lead to getting stuck in local optima\"---same problem as in text adventure games. \nIn this submission, the agents also rely on the learned knowledge base. \nNote: the learned knowledge base is very different from the quest (or dependency) structure---the former is like Figure-3 while the latter is Figure-2. \n\nThen in what sense can the authors confidently claim that their agents learned things like \"the virtue of being able to more consistently detect and clear bottlenecks\" or \"deal with latent structure of dependencies and bottlenecks\"? My opinion is: only if they can actually learn some structures like shown in Figure-2. \n\nThat being said, I request the authors to rewrite all the claims like mentioned above.\n\n[Technical clarity]\n\nThe most important clarity problem is about the modular policy chaining method: it is not clear how the method is integrated with the entire framework. \nE.g., how is $\\pi_{\\text{chain}}$ is used? It is only mentioned in Algo-1 but nowhere else. \nOther related detailed questions include: \nWhat is the \"policy\" in this paper? Is it a function to be learned, or a set of parameters, or a specific action for a given state? When a policy is buffered or chained, what is actually stored and how? \nAlgorithm-1 always has \"Bottleneck passed\". What if the bottleneck isn't passed? BACKTRACK procedure may not find a better score, so bottleneck is not passed, is it consistent with \"bottleneck passed\"?\nIs the BACKTRACK algorithm greedy? It looks like it returns immediately once it finds a higher $\\mathcal{J}$? \nHow is $\\mathcal{J} computed?$ It looks like it is stored and then indexed somehow. But how? How can you index a value by a \"policy\" as a key? This is related to the ``what is policy\" question. \n\nThere are other things about the framework that need to be clarified as well. \n\nIs the training algorithm online or offline? In Algo-1, A2C is called, indicating it is trained (step by step) during exploration: thus it seems an online algorithm. But what about QBERT which doesn't have modular policy chaining? \n\nWhat is the full action space? There should be some discussion about it in main paper, referring to the related appendices then. \n\nHow could one know the \"maximum score possible for the game\"? \n\nDoes the agent ask the same set of questions (to QA model) every step? Then how could it handle consistency? E.g., if \"what am I carrying\" is asked many times while your inventory is not changed at all, would it generate exactly the same answer all the time and how could you handle any difference? \n\nIn Figure-3, why ALBERT-QA doesn’t point to $KG$ but only points to $V_t$? Is ALBERT-QA used specifically to build knowledge graph? \nIn general, it is a little hard to connect the elements of Figure-3 to formula in section-4. \n\nHere are some presentation issues about the math formula: \nIndex-$i$ and index-$j$ in eqn-(1) are not grounded: they are mentioned in text; they don't loop over anything. \nSimilar problem for index-$i$ in eqn-(2). \n\nHere are some minor presentation issues: \n\"MCQBERT an agent that uses\" -> \"MC!QBERT, an agent that uses\"\n\"freedom to a explore\" -> \"freedom to explore\"\n\"these dependency graphs are … either … or … to progress and … a priori\" -> this sentence is too convoluted to correctly parse\n\"mostly deterministic … probabilities\" -> odd phrase: do you mean degenerative distribution? \n\"and without it has a true positive …\" -> ungrammatical \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting idea; limited novelty; misleading claims; unclear presentation; reject", "review": "\nThe paper proposes two RL agents for the text adventure games: QBERT learns (relational) knowledge about the game world with assistance of a trained QA-model and then constrains its action space with help of the knowledge base, enjoying greater sample efficiency; MCQBERT learns to escape from local optima by backtracking through the navigation trajectories, guided by the learned knowledge base and a manually-designed intrinsic reward function. \n\nPros: \n\nThis is a very dense paper with multiple key ideas, namely (1) using QA model to construct knowledge base; (2) escaping local optima by backtracking; and (3) modular policy chaining. \n\nBoth (1) and (2) are interesting and novel. But the novelty of (1) is limited because using QA to build graph has been used by e.g., Ammanabrolu et al. Bringing Stories Alive: Generating Interactive Fiction Worlds. \n\nEmpirical results demonstrate the effectiveness of the proposed methods. Not being eaten by Grue is impressive. \n\nCons: \n\nPresentation is the main reason that I’d like to reject the current version: \n(1) Several key claims are misleading and over-bold. \n(2) Many essential technical details are not clear. \n\n[Misleading/Over-bold claims]\n\nThe authors claimed, at many places, that they learned the \"dependencies'', namely the quest structure. For example, they said the agent \"learns that picking up the lamp is the right way to surpass the Cellar bottleneck and reach the Painting\". \n\nThis claim is wrong: the agent only successfully passes that room because it has done enough exploration (and the exploration is efficient because the intrinsic reward function is cleverly designed).\n\nIf this comment is not clear enough, let me be a bit more specific. \nIn the \"lamp-Grue\" example, the agent doesn't know the sense like \"if I pick up this lamp, I can pass the Grue room'' or \"now I am in the Grue room and the lamp is helpful\". \nInstead, the agent only picks the lamp up because it is encouraged to discover new information. \n\nOr, in other words, the agent doesn't really learn the quest structure as shown in Figure-2. \n(But I agree that Figure-2 is useful to illustrate the interesting idea.)\n\nI believe that the right way to view the method is: exploration is guided by not only environment reward but also intrinsic desire to discover new information, which leads to more efficient exploration and higher long-term reward. \nSimilar problems in other RL domains have been tackled and here is a paper that is similar to this submission in spirit: \nConti et al NeurIPS 2018 Improving Exploration in Evolution Strategies for Deep Reinforcement Learning via a Population of Novelty-Seeking Agents. \nQuote them: \"reward functions are often deceptive, and solely optimizing for reward without some mechanism to encourage intelligent exploration can lead to getting stuck in local optima\"---same problem as in text adventure games. \nIn this submission, the agents also rely on the learned knowledge base. \nNote: the learned knowledge base is very different from the quest (or dependency) structure---the former is like Figure-3 while the latter is Figure-2. \n\nThen in what sense can the authors confidently claim that their agents learned things like \"the virtue of being able to more consistently detect and clear bottlenecks\" or \"deal with latent structure of dependencies and bottlenecks\"? My opinion is: only if they can actually learn some structures like shown in Figure-2. \n\nThat being said, I request the authors to rewrite all the claims like mentioned above.\n\n[Technical clarity]\n\nThe most important clarity problem is about the modular policy chaining method: it is not clear how the method is integrated with the entire framework. \nE.g., how is $\\pi_{\\text{chain}}$ is used? It is only mentioned in Algo-1 but nowhere else. \nOther related detailed questions include: \nWhat is the \"policy\" in this paper? Is it a function to be learned, or a set of parameters, or a specific action for a given state? When a policy is buffered or chained, what is actually stored and how? \nAlgorithm-1 always has \"Bottleneck passed\". What if the bottleneck isn't passed? BACKTRACK procedure may not find a better score, so bottleneck is not passed, is it consistent with \"bottleneck passed\"?\nIs the BACKTRACK algorithm greedy? It looks like it returns immediately once it finds a higher $\\mathcal{J}$? \nHow is $\\mathcal{J} computed?$ It looks like it is stored and then indexed somehow. But how? How can you index a value by a \"policy\" as a key? This is related to the ``what is policy\" question. \n\nThere are other things about the framework that need to be clarified as well. \n\nIs the training algorithm online or offline? In Algo-1, A2C is called, indicating it is trained (step by step) during exploration: thus it seems an online algorithm. But what about QBERT which doesn't have modular policy chaining? \n\nWhat is the full action space? There should be some discussion about it in main paper, referring to the related appendices then. \n\nHow could one know the \"maximum score possible for the game\"? \n\nDoes the agent ask the same set of questions (to QA model) every step? Then how could it handle consistency? E.g., if \"what am I carrying\" is asked many times while your inventory is not changed at all, would it generate exactly the same answer all the time and how could you handle any difference? \n\nIn Figure-3, why ALBERT-QA doesn’t point to $KG$ but only points to $V_t$? Is ALBERT-QA used specifically to build knowledge graph? \nIn general, it is a little hard to connect the elements of Figure-3 to formula in section-4. \n\nHere are some presentation issues about the math formula: \nIndex-$i$ and index-$j$ in eqn-(1) are not grounded: they are mentioned in text; they don't loop over anything. \nSimilar problem for index-$i$ in eqn-(2). \n\nHere are some minor presentation issues: \n\"MCQBERT an agent that uses\" -> \"MC!QBERT, an agent that uses\"\n\"freedom to a explore\" -> \"freedom to explore\"\n\"these dependency graphs are … either … or … to progress and … a priori\" -> this sentence is too convoluted to correctly parse\n\"mostly deterministic … probabilities\" -> odd phrase: do you mean degenerative distribution? \n\"and without it has a true positive …\" -> ungrammatical \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603902156453}, {"id": "G4M9b3fTEFc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1887/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary/Overall Quality: The authors make the following contributions.\n- focus on text based adventure games\n- create agent that can build knowledge graph by answering question\n- introduce novel exploration strategy\n- IM reward: expand size of knowledge graph\nBecause the authors present a number of interesting and well tested strategies for a well-justified task, the paper is above the acceptance threshold. If the clarity could be improved (specifically by making the contribution in the general case more explicit) and if the experiments could be made more thorough (+ a stronger improvement of prior work) it would be a strong paper.\n\nClarity: The authors did a reasonable job of familiarizing readers with the challenges of text-based games. However, the description of the research contribution and its potential impacts could have been more focused. For example, in the abstract the authors don't mention that they collect a Jericho-QA dataset, but it is listed as one of the three contributions in the final paragraph of the related work. The paper spans a lot of material--from the research merit of text based games to intrinsic motivation based on knowledge graph formation to using QA models for knowledge graph construction to modular policy chaining--and it is difficult to walk away from the paper with a cohesive understanding of the research contribution.\n\nOriginality: While the individual components are not extremely novel, taken together they create an interesting and original system design. E.g., even though the notion of using intrinsic motivation to un-sparsify a reward space is far from new, the authors devise an interesting formulation of IM to apply to knowledge graph construction.\n\nSignificance:\n- _For text based games_: the authors provide numerous method contributions and a dataset for this application specifically. However, looking at table 1 and figure 4, its not clear that the fullest version of their method has a consistent advantage over prior work.\n- _Generally_ Text-based games could be an important benchmark/milestone for systems that need to reason well into the past and over a large action, discrete action space. The authors do not address this much beyond some prose in the abstract and introduction. Certainly some components of the methods they have developed (e.g., exploration over a knowledge graph) has broad-spanning applications.\n\nStrengths:\n- The authors provide a variety of methods to help agents excel is text based games: knowledge graph creation with a QA system, a dataset to train the QA system, intrinsic motivation based on knowledge graph expansion, modular policy training.\n- Show results over a large number of text-based games. On several games, their model, MC!Q* sees a substantial improvement over prior work.\n\nWeaknesses:\n- Improvement over prior work is not consistent (see library, balances, and temple tasks in table 1, in figure 4a Q*bert converges more quickly on average but not above variance of KG-A2C)\n- As someone unfamiliar with text based games, it's hard to interpret the results in table 2 and to understand what problems are being addressed in each game.\n- More experiments to better understand how bottlenecks are addressed by the model (e.g., more of figure 4b) would be enlightening", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Application with Large Number of New Methods", "review": "Summary/Overall Quality: The authors make the following contributions.\n- focus on text based adventure games\n- create agent that can build knowledge graph by answering question\n- introduce novel exploration strategy\n- IM reward: expand size of knowledge graph\nBecause the authors present a number of interesting and well tested strategies for a well-justified task, the paper is above the acceptance threshold. If the clarity could be improved (specifically by making the contribution in the general case more explicit) and if the experiments could be made more thorough (+ a stronger improvement of prior work) it would be a strong paper.\n\nClarity: The authors did a reasonable job of familiarizing readers with the challenges of text-based games. However, the description of the research contribution and its potential impacts could have been more focused. For example, in the abstract the authors don't mention that they collect a Jericho-QA dataset, but it is listed as one of the three contributions in the final paragraph of the related work. The paper spans a lot of material--from the research merit of text based games to intrinsic motivation based on knowledge graph formation to using QA models for knowledge graph construction to modular policy chaining--and it is difficult to walk away from the paper with a cohesive understanding of the research contribution.\n\nOriginality: While the individual components are not extremely novel, taken together they create an interesting and original system design. E.g., even though the notion of using intrinsic motivation to un-sparsify a reward space is far from new, the authors devise an interesting formulation of IM to apply to knowledge graph construction.\n\nSignificance:\n- _For text based games_: the authors provide numerous method contributions and a dataset for this application specifically. However, looking at table 1 and figure 4, its not clear that the fullest version of their method has a consistent advantage over prior work.\n- _Generally_ Text-based games could be an important benchmark/milestone for systems that need to reason well into the past and over a large action, discrete action space. The authors do not address this much beyond some prose in the abstract and introduction. Certainly some components of the methods they have developed (e.g., exploration over a knowledge graph) has broad-spanning applications.\n\nStrengths:\n- The authors provide a variety of methods to help agents excel is text based games: knowledge graph creation with a QA system, a dataset to train the QA system, intrinsic motivation based on knowledge graph expansion, modular policy training.\n- Show results over a large number of text-based games. On several games, their model, MC!Q* sees a substantial improvement over prior work.\n\nWeaknesses:\n- Improvement over prior work is not consistent (see library, balances, and temple tasks in table 1, in figure 4a Q*bert converges more quickly on average but not above variance of KG-A2C)\n- As someone unfamiliar with text based games, it's hard to interpret the results in table 2 and to understand what problems are being addressed in each game.\n- More experiments to better understand how bottlenecks are addressed by the model (e.g., more of figure 4b) would be enlightening", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603864842739}], "openreview_url": "https://openreview.net/forum?id=eYgI3cTPTq9", "arxiv_id": "2006.07409", "paper_pdf": "papers/eYgI3cTPTq9.pdf", "paper_pdf_sha256": "22cbfa152665c70a38184813aa508ec3b6f55e956997a95c2391ceee2e7c9b64", "paper_pdf_bytes": 3161716, "paper_pdf_source": "openreview", "code_url": "https://github.com/rajammanabrolu/Q-BERT", "code_repository": "rajammanabrolu/Q-BERT", "code_commit": "9ca2df39d9ab5968f7a5034381483cfcef122a03", "code_archive": "repos/eYgI3cTPTq9.zip", "code_archive_sha256": "de8dd37eaf5ee9acc6a63cb6283c529153cbbfd3644ac43dc267c629429dea8e", "code_archive_bytes": 484483, "code_file_count": 26, "code_extensions": {".py": 26}, "github_disk_usage_kb": 481, "github_languages": {"Python": 333890}, "github_archived": false, "github_pushed_at": "2023-08-09T22:03:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-to-avoid-being-eaten-by-a-grue-structured"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJlEEaEFDS", "year": 2020, "status": "rejected", "title": "Towards an Adversarially Robust Normalization Approach", "authors": ["Muhammad Awais", "Fahad Shamshad", "Sung-Ho Bae"], "authorids": ["awais@khu.ac.kr", "fahad.shamshad@itu.edu.pk", "shbae@khu.ac.kr"], "authors_source": "OpenReview API", "abstract": "Batch Normalization (BatchNorm) has shown to be effective for improving and accelerating the training of deep neural networks. However, recently it has been shown that it is also vulnerable to adversarial perturbations. In this work, we aim to investigate the cause of adversarial vulnerability of the BatchNorm. We hypothesize that the use of different normalization statistics during training and inference (mini-batch statistics for training and moving average of these values at inference) is the main cause of this adversarial vulnerability in the BatchNorm layer. We empirically proved this by experiments on various neural network architectures and datasets. Furthermore, we introduce Robust Normalization (RobustNorm) and experimentally show that it is not only resilient to adversarial perturbation but also inherit the benefits of BatchNorm.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SyxnlNl0tB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper482/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an interesting perspective that BatchNorm may introduce the adversarial vulnerability, and probes why BatchNorm performs like that (the tracking part in BatchNorm). In experiment, the robustness of the networks increases by 20% when removing the tracking part, but the test accuracy on the clean images drops a lot. Afterwards, the authors propose RobustNorm, which performs better than BatchNorm for both natural and adversarial scenarios.\n\nDetailed Comments: \n+ The paper is well written. The paper structure is clear and figures are well illustrated.\n+ The paper understands and carefully investigates BatchNorm in a interesting and important direction. After the investigation, the improved version RobustNorm shows more potential.\n+ The experimental results seem good. The RobustNorm performs better than BatchNorm for both natural and adversarial scenarios.\n- More results on ImageNet would be better to verify the proposed RobustNorm method.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes an interesting perspective that BatchNorm may introduce the adversarial vulnerability, and probes why BatchNorm performs like that (the tracking part in BatchNorm). In experiment, the robustness of the networks increases by 20% when removing the tracking part, but the test accuracy on the clean images drops a lot. Afterwards, the authors propose RobustNorm, which performs better than BatchNorm for both natural and adversarial scenarios.\n\nDetailed Comments: \n+ The paper is well written. The paper structure is clear and figures are well illustrated.\n+ The paper understands and carefully investigates BatchNorm in a interesting and important direction. After the investigation, the improved version RobustNorm shows more potential.\n+ The experimental results seem good. The RobustNorm performs better than BatchNorm for both natural and adversarial scenarios.\n- More results on ImageNet would be better to verify the proposed RobustNorm method."}, "tcdate": 1571845108190}, {"id": "SyxQ0cXatS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper482/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Review: This paper investigates the reason behind the vulnerability of BatchNorm and proposes a Robust Normalization. They experimentally show that it is the moving averages of mini-batch means and variances (tracking) used in Normalization that cause the adversarial vulnerability. Based on this observation, they propose a new normalization method not only achieves significantly better results under a variety of attack methods but ensures a comparable test accuracy to that of BatchNorm on unperturbed datasets. The paper is clearly written, easy to read.\n \nStrengths:\n \nExplore the cause of adversarial vulnerability of the BatchNorm and assume that the tracking mechanism used in original BatchNorm leads to the vulnerability from experiment results.\nPropose a new and simple normalization method and perform extensive experiments to validate the efficacy of proposed method.\n \nWeaknesses:\nThough extensive experiments have been done by revealing what leads the vulnerability and the effectiveness of proposed method. The results seem unconvincing with respect to different datasets, since Cifar10 and Cifar100 are inherently connected. Would you mind performing some experiments on ImageNet? Since adversarial training on ImageNet is time-consuming, can you show us the result of Natural Training of different models with different norms on ImageNet and compare their robustness under different attack?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Review: This paper investigates the reason behind the vulnerability of BatchNorm and proposes a Robust Normalization. They experimentally show that it is the moving averages of mini-batch means and variances (tracking) used in Normalization that cause the adversarial vulnerability. Based on this observation, they propose a new normalization method not only achieves significantly better results under a variety of attack methods but ensures a comparable test accuracy to that of BatchNorm on unperturbed datasets. The paper is clearly written, easy to read.\n \nStrengths:\n \nExplore the cause of adversarial vulnerability of the BatchNorm and assume that the tracking mechanism used in original BatchNorm leads to the vulnerability from experiment results.\nPropose a new and simple normalization method and perform extensive experiments to validate the efficacy of proposed method.\n \nWeaknesses:\nThough extensive experiments have been done by revealing what leads the vulnerability and the effectiveness of proposed method. The results seem unconvincing with respect to different datasets, since Cifar10 and Cifar100 are inherently connected. Would you mind performing some experiments on ImageNet? Since adversarial training on ImageNet is time-consuming, can you show us the result of Natural Training of different models with different norms on ImageNet and compare their robustness under different attack?\n"}, "tcdate": 1571793611516}, {"id": "SkgvVW1htH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper482/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses a limitation of BatchNorm: vulnerability to adversarial perturbations. The authors propose a possible explanation of this issue and correspondingly an alternative called RobustNorm to tackle this problem. Specifically, the authors observe that the statistics of BatchNorm for training and inference are different, resulting in different data distributions for training and inference. To solve this problem, the authors propose to use min-max rescaling instead of normalization. In addition, the running average is calculated with mean and the running mean of the denominator during inference. Experimental results show significant improvement of robustness and also comparable accuracy for clean data.\n\nThe paper is well-written and the contributions are stated clearly. The explanation of vulnerability is reasonable. The proposed solution is simple but effective.\n\nHowever, I have several concerns:\n*The authors verify that the running average is the main culprit of vulnerability to adversarial attack, but provide no further investigation of why this happens. A possible solution is the drift in input distributions, but the manuscript does not state clearly how is the distribution changed. Further experiments would have made this claim more convincing.\n*The proposed method involves a hyper-parameter \\rho, but it may result in problematic issues. The variance of input is of the same order of magnitude as (max(x)-min(x))^2. If \\rho is set to other value, the magnitude of gradient will change drastically during back-propagation. Although \\rho can be set to 0.2, it still seems ad-hoc. Experiments on more datasets and the sensitivity of the proposed method to \\rho would have validated the claims of the authors.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper addresses a limitation of BatchNorm: vulnerability to adversarial perturbations. The authors propose a possible explanation of this issue and correspondingly an alternative called RobustNorm to tackle this problem. Specifically, the authors observe that the statistics of BatchNorm for training and inference are different, resulting in different data distributions for training and inference. To solve this problem, the authors propose to use min-max rescaling instead of normalization. In addition, the running average is calculated with mean and the running mean of the denominator during inference. Experimental results show significant improvement of robustness and also comparable accuracy for clean data.\n\nThe paper is well-written and the contributions are stated clearly. The explanation of vulnerability is reasonable. The proposed solution is simple but effective.\n\nHowever, I have several concerns:\n*The authors verify that the running average is the main culprit of vulnerability to adversarial attack, but provide no further investigation of why this happens. A possible solution is the drift in input distributions, but the manuscript does not state clearly how is the distribution changed. Further experiments would have made this claim more convincing.\n*The proposed method involves a hyper-parameter \\rho, but it may result in problematic issues. The variance of input is of the same order of magnitude as (max(x)-min(x))^2. If \\rho is set to other value, the magnitude of gradient will change drastically during back-propagation. Although \\rho can be set to 0.2, it still seems ad-hoc. Experiments on more datasets and the sensitivity of the proposed method to \\rho would have validated the claims of the authors.\n"}, "tcdate": 1571709230895}], "openreview_url": "https://openreview.net/forum?id=BJlEEaEFDS", "arxiv_id": "2006.11007", "paper_pdf": "papers/BJlEEaEFDS.pdf", "paper_pdf_sha256": "31a33ee05aae850fda5e8fa6aaed89ea747a04cbfd682d7e7b9cb8038b29095c", "paper_pdf_bytes": 966068, "paper_pdf_source": "openreview", "code_url": "https://github.com/awaisrauf/RobustNorm", "code_repository": "awaisrauf/RobustNorm", "code_commit": "d14d68cb8990be3c50b387e08ec72dd20a01dba2", "code_archive": "repos/BJlEEaEFDS.zip", "code_archive_sha256": "36d6149294885a8be0a93e3d7f5b5d936a5ced6225d29dd93240db589baed0fd", "code_archive_bytes": 656239, "code_file_count": 29, "code_extensions": {".py": 28, ".sh": 1}, "github_disk_usage_kb": 633, "github_languages": {"Python": 121993, "Shell": 355}, "github_archived": false, "github_pushed_at": "2020-10-06T05:31:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-an-adversarially-robust-normalization-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "14vBE8iAEx", "year": 2026, "status": "rejected", "title": "Beyond Reward Hacking: Causal Rewards for Large Language Model Alignment", "authors": ["Chaoqi Wang", "Zhuokai Zhao", "Yibo Jiang", "Zhaorun Chen", "Chen Zhu", "Yuxin Chen", "Jiayi Liu", "Lizhu Zhang", "Xiangjun Fan", "Hao Ma", "Sinong Wang"], "authorids": ["~Chaoqi_Wang1", "~Zhuokai_Zhao1", "~Yibo_Jiang2", "~Zhaorun_Chen1", "~Chen_Zhu2", "~Yuxin_Chen1", "~Jiayi_Liu1", "~Lizhu_Zhang2", "~Xiangjun_Fan1", "~Hao_Ma1", "~Sinong_Wang1"], "authors_source": "OpenReview API", "abstract": "Recent advances in large language models (LLMs) have demonstrated significant progress in performing complex tasks. While Reinforcement Learning from Human Feedback (RLHF) has been effective in aligning LLMs with human preferences, it is susceptible to spurious correlations in reward modeling. Consequently, it often introduces biases—such as length bias, sycophancy, conceptual bias, and discrimination—that hinder the model’s ability to capture true causal relationships. To address this, we propose a novel causal reward modeling approach that integrates causality to mitigate these spurious correlations. Our method enforces counterfactual invariance, ensuring reward predictions remain consistent when irrelevant variables are altered. Through experiments on both synthetic and real-world datasets, we show that our approach mitigates various types of spurious correlations effectively, resulting in more reliable and fair alignment of LLMs with human preferences. As a drop-in enhancement to the existing RLHF workflow, our causal reward modeling provides a practical way to improve the trustworthiness and fairness of LLM finetuning.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "TRdO6Q0Nkw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20565/Reviewer_QC9D"], "rating": 2, "soundness": 3, "presentation": 1, "contribution": 3, "confidence": 3, "summary": "The paper introduces a causality inspired regularizer in Reward Modelling.", "review_text": "The paper introduces a causality inspired regularizer in Reward Modelling.", "strengths": "The idea seems promising and novel. The idea is well motivated, and the solution is presented with clarity.", "weaknesses": "The experimental section seems to have missed out on providing some details. \n\n1) For any one of the dataset mention how you are creating the bins.\n2) How are you calculating p_m and p_m^'. For what values of Z is this being computed?\n3) What are the values that Z can take in any one of the datasets?\n4) Describe the conditional CRM in more detail.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a causality inspired regularizer in Reward Modelling.", "soundness": 3, "presentation": 1, "contribution": 3, "strengths": "The idea seems promising and novel. The idea is well motivated, and the solution is presented with clarity.", "weaknesses": "The experimental section seems to have missed out on providing some details. \n\n1) For any one of the dataset mention how you are creating the bins.\n2) How are you calculating p_m and p_m^'. For what values of Z is this being computed?\n3) What are the values that Z can take in any one of the datasets?\n4) Describe the conditional CRM in more detail.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761901448548}, {"id": "twzttMAVpo", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20565/Reviewer_w8yn"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "The paper proposed a new approach called ‘Causal Reward Model’ to mitigate the problem of spurious correlation and reward hacking that is common in RLHF. Traditional methods use spurious shortcuts to deal with hacking coming from factors like long length. The paper, instead, treats the factors like length and tone as spurious variables $Z$, and learn a reward predictor whose internal representation is independent of $Z$. A maximum mean discrepancy based on the reward predictor is then added to the loss function as the regularisation.", "review_text": "The paper proposed a new approach called ‘Causal Reward Model’ to mitigate the problem of spurious correlation and reward hacking that is common in RLHF. Traditional methods use spurious shortcuts to deal with hacking coming from factors like long length. The paper, instead, treats the factors like length and tone as spurious variables $Z$, and learn a reward predictor whose internal representation is independent of $Z$. A maximum mean discrepancy based on the reward predictor is then added to the loss function as the regularisation.", "strengths": "-The paper is based on causal language and counterfactual invariance, which is novel compared to previous methods using supurious shortcuts. The CRM is then added as a normal regulariser which can be embedded into common model frameworks easily.\n- Empirical results are also good.", "weaknesses": "- The method can only debiases the explicit spurious factors. If an unknown latent shortcut hack the reward, it won’t fix it. \n- The benchmark is PPO, but what about other methods that address the shortcuts, the author should compare with these methods as well.", "questions": "- In practice, how to decide the spurious variables Z? \n- Does CRM stay robust after the fine-tuning, which may cause the policy’s distribution shift against the preference dataset used to train CRM?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed a new approach called ‘Causal Reward Model’ to mitigate the problem of spurious correlation and reward hacking that is common in RLHF. Traditional methods use spurious shortcuts to deal with hacking coming from factors like long length. The paper, instead, treats the factors like length and tone as spurious variables $Z$, and learn a reward predictor whose internal representation is independent of $Z$. A maximum mean discrepancy based on the reward predictor is then added to the loss function as the regularisation.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "-The paper is based on causal language and counterfactual invariance, which is novel compared to previous methods using supurious shortcuts. The CRM is then added as a normal regulariser which can be embedded into common model frameworks easily.\n- Empirical results are also good.", "weaknesses": "- The method can only debiases the explicit spurious factors. If an unknown latent shortcut hack the reward, it won’t fix it. \n- The benchmark is PPO, but what about other methods that address the shortcuts, the author should compare with these methods as well.", "questions": "- In practice, how to decide the spurious variables Z? \n- Does CRM stay robust after the fine-tuning, which may cause the policy’s distribution shift against the preference dataset used to train CRM?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761871904303}, {"id": "qa34RdDv2W", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20565/Reviewer_WY6t"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "The paper introduces a method called CRM to address spurious correlations and reward hacking in RLHF. The central idea is to remove the confounding effect of response length so that the reward model more accurately reflects the true quality of the response. The authors leverage MMD to design a regularizer for reward model training. Experiments are conducted to evaluate the performance of the proposed method.", "review_text": "The paper introduces a method called CRM to address spurious correlations and reward hacking in RLHF. The central idea is to remove the confounding effect of response length so that the reward model more accurately reflects the true quality of the response. The authors leverage MMD to design a regularizer for reward model training. Experiments are conducted to evaluate the performance of the proposed method.", "strengths": "The paper introduces a new perspective by leveraging the method in counterfactual invariance. The proposed CRM serves as a regularizer that can be integrated into standard reward model frameworks.", "weaknesses": "- I am concerned about the novelty of this paper, as it largely builds on the technique proposed in [1], which already demonstrates that MMD can serve as a causal regularizer. In comparison, the present work appears to mainly combine and reapply techniques from that paper.\n\n- The paper is not well written, causing unnecessary confusion in its methodology.\n\n- Line 212: “the causal graph reveals that $T^{L, \\perp}$ is independent of $Z$,” but Figure 1 shows that $Z$ directly influences $T^{L, \\perp}$.\n\n- I believe Equation (1) contains the key insight of the paper, but its description is delayed, and the preceding discussion seems disconnected from it. \n\n- Is $T^{L, \\perp}$ well-defined?\n\n- The description of MMD could be moved to the Preliminaries section to make the methodological contribution clearer. \n\n- Why is it necessary to bin $Z$? Why not directly measure the dependence between $f(T)$ and $T$ using the Hilbert–Schmidt Independence Criterion (HSIC)? These questions further aggravate my concern about the novelty of the paper, as it closely follows the approach in [1].\n\n- The definition of $p'_m$ is also unclear.\n\n- I am skeptical about the effectiveness of the proposed formulation. From a statistical perspective, MMD only measures dependence; it does not imply causality. Intuitively, applying MMD regularization in (1) cannot ensure the causality claimed in this paper.\n\n- How is $Z$ chosen in practice? What if $Z$ includes a factor that does reflect the true reward of the response? Naively applying Equation (1) in such cases could hinder effective reward learning.\n\n- The experimental comparison is limited. The authors should consider including more recent methods, for example:\n  - https://arxiv.org/abs/2403.19159\n  - https://arxiv.org/abs/2502.00814\n\n[1] https://arxiv.org/abs/2106.00545", "questions": "See Weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a method called CRM to address spurious correlations and reward hacking in RLHF. The central idea is to remove the confounding effect of response length so that the reward model more accurately reflects the true quality of the response. The authors leverage MMD to design a regularizer for reward model training. Experiments are conducted to evaluate the performance of the proposed method.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "The paper introduces a new perspective by leveraging the method in counterfactual invariance. The proposed CRM serves as a regularizer that can be integrated into standard reward model frameworks.", "weaknesses": "- I am concerned about the novelty of this paper, as it largely builds on the technique proposed in [1], which already demonstrates that MMD can serve as a causal regularizer. In comparison, the present work appears to mainly combine and reapply techniques from that paper.\n\n- The paper is not well written, causing unnecessary confusion in its methodology.\n\n- Line 212: “the causal graph reveals that $T^{L, \\perp}$ is independent of $Z$,” but Figure 1 shows that $Z$ directly influences $T^{L, \\perp}$.\n\n- I believe Equation (1) contains the key insight of the paper, but its description is delayed, and the preceding discussion seems disconnected from it. \n\n- Is $T^{L, \\perp}$ well-defined?\n\n- The description of MMD could be moved to the Preliminaries section to make the methodological contribution clearer. \n\n- Why is it necessary to bin $Z$? Why not directly measure the dependence between $f(T)$ and $T$ using the Hilbert–Schmidt Independence Criterion (HSIC)? These questions further aggravate my concern about the novelty of the paper, as it closely follows the approach in [1].\n\n- The definition of $p'_m$ is also unclear.\n\n- I am skeptical about the effectiveness of the proposed formulation. From a statistical perspective, MMD only measures dependence; it does not imply causality. Intuitively, applying MMD regularization in (1) cannot ensure the causality claimed in this paper.\n\n- How is $Z$ chosen in practice? What if $Z$ includes a factor that does reflect the true reward of the response? Naively applying Equation (1) in such cases could hinder effective reward learning.\n\n- The experimental comparison is limited. The authors should consider including more recent methods, for example:\n  - https://arxiv.org/abs/2403.19159\n  - https://arxiv.org/abs/2502.00814\n\n[1] https://arxiv.org/abs/2106.00545", "questions": "See Weaknesses section", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761759124875}, {"id": "Xlr3SMTJjI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20565/Reviewer_Dqv3"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper addresses a fundamental problem in Reinforcement Learning from Human Feedback (RLHF) — reward hacking due to spurious correlations in reward modeling. Traditional RLHF reward models often conflate superficial correlations (e.g., response length, agreement, demographic cues) with genuine human preferences, leading to biases such as length bias, sycophancy, concept bias, and discrimination bias (see Section 1). To overcome this, the authors propose a Causal Reward Model (CRM) that incorporates counterfactual invariance to ensure that reward predictions are independent of irrelevant variables (Section 3.2–4.1). The method employs Maximum Mean Discrepancy (MMD) regularization to enforce statistical independence between latent representations and spurious factors (Eq. (1), Section 4.1). Experiments across four settings — sycophantic, length, concept, and discrimination bias — demonstrate that CRM significantly mitigates spurious correlations and improves fairness and robustness without sacrificing model utility (Sections 5.1–5.4, Tables 1–3, Fig. 3). The approach integrates seamlessly with existing RLHF pipelines, offering a practical improvement to current alignment workflows.", "review_text": "The paper addresses a fundamental problem in Reinforcement Learning from Human Feedback (RLHF) — reward hacking due to spurious correlations in reward modeling. Traditional RLHF reward models often conflate superficial correlations (e.g., response length, agreement, demographic cues) with genuine human preferences, leading to biases such as length bias, sycophancy, concept bias, and discrimination bias (see Section 1). To overcome this, the authors propose a Causal Reward Model (CRM) that incorporates counterfactual invariance to ensure that reward predictions are independent of irrelevant variables (Section 3.2–4.1). The method employs Maximum Mean Discrepancy (MMD) regularization to enforce statistical independence between latent representations and spurious factors (Eq. (1), Section 4.1). Experiments across four settings — sycophantic, length, concept, and discrimination bias — demonstrate that CRM significantly mitigates spurious correlations and improves fairness and robustness without sacrificing model utility (Sections 5.1–5.4, Tables 1–3, Fig. 3). The approach integrates seamlessly with existing RLHF pipelines, offering a practical improvement to current alignment workflows.", "strengths": "Originality: The paper introduces a novel causal regularization framework for reward modeling within RLHF, which is an underexplored direction. While prior works address specific biases (e.g., length bias via PoE or WARM), CRM provides a unified causal formulation that generalizes across multiple types of spurious correlations (Section 2.2). The concept of counterfactual invariance adapted from causal inference to LLM alignment is new.\n\nQuality: The theoretical foundation is clearly established with causal diagrams (Figure 1, Section 3.2) and well-defined objectives (Eq. (1)–(2), Section 4.1). The experimental design is comprehensive and evaluates diverse bias types. Results (e.g., Table 2 and 3) consistently demonstrate the superiority of both conditional and unconditional CRMs over baseline reward models. The empirical validation across multiple bias domains substantiates the method’s robustness and generalizability.\n\nClarity: The paper is well-structured and written clearly. Figures and tables (e.g., Fig. 2 and 3) effectively visualize comparative performance trends.\n\nSignificance:\nReward hacking remains a major bottleneck for safe deployment of RLHF-trained LLMs. Hence, this paper’s contribution is relevant.", "weaknesses": "Hyperparameter Sensitivity: The choice of MMD coefficients (λ) significantly affects results (Fig. 3), but the paper does not deeply explore tuning stability or sensitivity. This could impact deployment robustness in practice.", "questions": "Causal Assumptions: How sensitive is CRM to incorrect causal graph specifications?\n\nScalability: How does the computational cost of MMD regularization scale with large datasets and high-dimensional embeddings?\n\nHyperparameter Selection: How was λ chosen in practice? Did you observe stable optima across tasks or require dataset-specific tuning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses a fundamental problem in Reinforcement Learning from Human Feedback (RLHF) — reward hacking due to spurious correlations in reward modeling. Traditional RLHF reward models often conflate superficial correlations (e.g., response length, agreement, demographic cues) with genuine human preferences, leading to biases such as length bias, sycophancy, concept bias, and discrimination bias (see Section 1). To overcome this, the authors propose a Causal Reward Model (CRM) that incorporates counterfactual invariance to ensure that reward predictions are independent of irrelevant variables (Section 3.2–4.1). The method employs Maximum Mean Discrepancy (MMD) regularization to enforce statistical independence between latent representations and spurious factors (Eq. (1), Section 4.1). Experiments across four settings — sycophantic, length, concept, and discrimination bias — demonstrate that CRM significantly mitigates spurious correlations and improves fairness and robustness without sacrificing model utility (Sections 5.1–5.4, Tables 1–3, Fig. 3). The approach integrates seamlessly with existing RLHF pipelines, offering a practical improvement to current alignment workflows.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Originality: The paper introduces a novel causal regularization framework for reward modeling within RLHF, which is an underexplored direction. While prior works address specific biases (e.g., length bias via PoE or WARM), CRM provides a unified causal formulation that generalizes across multiple types of spurious correlations (Section 2.2). The concept of counterfactual invariance adapted from causal inference to LLM alignment is new.\n\nQuality: The theoretical foundation is clearly established with causal diagrams (Figure 1, Section 3.2) and well-defined objectives (Eq. (1)–(2), Section 4.1). The experimental design is comprehensive and evaluates diverse bias types. Results (e.g., Table 2 and 3) consistently demonstrate the superiority of both conditional and unconditional CRMs over baseline reward models. The empirical validation across multiple bias domains substantiates the method’s robustness and generalizability.\n\nClarity: The paper is well-structured and written clearly. Figures and tables (e.g., Fig. 2 and 3) effectively visualize comparative performance trends.\n\nSignificance:\nReward hacking remains a major bottleneck for safe deployment of RLHF-trained LLMs. Hence, this paper’s contribution is relevant.", "weaknesses": "Hyperparameter Sensitivity: The choice of MMD coefficients (λ) significantly affects results (Fig. 3), but the paper does not deeply explore tuning stability or sensitivity. This could impact deployment robustness in practice.", "questions": "Causal Assumptions: How sensitive is CRM to incorrect causal graph specifications?\n\nScalability: How does the computational cost of MMD regularization scale with large datasets and high-dimensional embeddings?\n\nHyperparameter Selection: How was λ chosen in practice? Did you observe stable optima across tasks or require dataset-specific tuning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761738149157}], "openreview_url": "https://openreview.net/forum?id=14vBE8iAEx", "arxiv_id": "2501.09620", "paper_pdf": "papers/14vBE8iAEx.pdf", "paper_pdf_sha256": "b7584a51aa6bbddfb257383c20c40dbcc3c05db84687be2f9ed3cb76022a71c8", "paper_pdf_bytes": 425357, "paper_pdf_source": "openreview", "code_url": "https://github.com/tatsu-lab/alpaca_farm", "code_repository": "tatsu-lab/alpaca_farm", "code_commit": "30717ddae735365de756ee2085191b491a71788d", "code_archive": "repos/14vBE8iAEx.zip", "code_archive_sha256": "3c56a0b5a4a8d6965665e4914e578041c188ab617a6b6473d43496e2e02bc779", "code_archive_bytes": 2205761, "code_file_count": 59, "code_extensions": {".py": 51, ".sh": 6, ".js": 1, ".ipynb": 1}, "github_disk_usage_kb": 1895, "github_languages": {"Python": 261484, "JavaScript": 69}, "github_archived": false, "github_pushed_at": "2024-07-01T18:53:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/beyond-reward-hacking-causal-rewards-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2Oh2EOcFSO", "year": 2025, "status": "rejected", "title": "Can a Bayesian oracle prevent harm from an agent?", "authors": ["Yoshua Bengio", "Michael K. Cohen", "Nikolay Malkin", "Matt MacDermott", "Damiano Fornasiere", "Pietro Greiner", "Younesse Kaddar"], "authorids": ["~Yoshua_Bengio1", "~Michael_K._Cohen1", "~Nikolay_Malkin1", "~Matt_MacDermott1", "~Damiano_Fornasiere1", "~Pietro_Greiner1", "~Younesse_Kaddar1"], "authors_source": "OpenReview API", "abstract": "Is there a way to design powerful AI systems based on machine learning methods that would satisfy probabilistic safety guarantees? With the long-term goal of obtaining a probabilistic guarantee that would apply in every context, we consider estimating a context-dependent bound on the probability of violating a given safety specification.  Such a risk evaluation would need to be performed at run-time to provide a guardrail against dangerous actions of an AI. Noting that different plausible hypotheses about the world could produce very different outcomes, and because we do not know which one is right, we derive bounds on the safety violation probability predicted  under the true but unknown hypothesis. Such bounds could be used to reject potentially dangerous actions. Our main results involve searching for cautious but plausible hypotheses, obtained by a maximization that involves Bayesian posteriors over hypotheses. We consider two forms of this result, in the i.i.d. case and in the non-i.i.d. case, and conclude with open problems towards turning such theoretical results into practical AI guardrails.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "JvzstXgFIj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission424/Reviewer_n2RR"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper explores the problem of designing AI systems that satisfy probabilistic safety guarantees. Within a Bayesian framework and given the safety specifications (as a probability), the authors provide risk bounds for potentially harmful decisions, showing that the probability of harm can be upper-bounded by a probability that can be estimated by approximating Bayseian posterior over theories given the observed data. They study two settings: i.i.d case and non i.i.d case and provide a simple experiment to evaluate the performance of safety guardrails.", "review_text": "This paper explores the problem of designing AI systems that satisfy probabilistic safety guarantees. Within a Bayesian framework and given the safety specifications (as a probability), the authors provide risk bounds for potentially harmful decisions, showing that the probability of harm can be upper-bounded by a probability that can be estimated by approximating Bayseian posterior over theories given the observed data. They study two settings: i.i.d case and non i.i.d case and provide a simple experiment to evaluate the performance of safety guardrails.", "strengths": "This is a very well written paper, and it is easy to follow. \n\nThe proposed approach represents a promising initial step toward designing AI systems that ensure safety through built-in probabilistic guarantees, rather than relying solely on external safety mechanisms.\n\nThe authors also outline several open problems for future work.", "weaknesses": "The authors present an upper bound on the harm probability, though it appears to be highly conservative. It would be valuable if they could offer a convergence rate or practical guarantees to make the framework more usable. Additionally, it is unclear how this approach compares to other conservative methods for preventing harm. \n\nSince the theoretical results lack practical assurances, I would have appreciated more experimental validation, especially in complex and realistic settings. \n\nObtaining a Bayesian oracle could be very challenging (posterior distribution). \n\nOverall, while the paper introduces a promising method for designing safer AI systems, it would greatly benefit from additional components (both theoretical and experimental) before publication.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the problem of designing AI systems that satisfy probabilistic safety guarantees. Within a Bayesian framework and given the safety specifications (as a probability), the authors provide risk bounds for potentially harmful decisions, showing that the probability of harm can be upper-bounded by a probability that can be estimated by approximating Bayseian posterior over theories given the observed data. They study two settings: i.i.d case and non i.i.d case and provide a simple experiment to evaluate the performance of safety guardrails.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This is a very well written paper, and it is easy to follow. \n\nThe proposed approach represents a promising initial step toward designing AI systems that ensure safety through built-in probabilistic guarantees, rather than relying solely on external safety mechanisms.\n\nThe authors also outline several open problems for future work.", "weaknesses": "The authors present an upper bound on the harm probability, though it appears to be highly conservative. It would be valuable if they could offer a convergence rate or practical guarantees to make the framework more usable. Additionally, it is unclear how this approach compares to other conservative methods for preventing harm. \n\nSince the theoretical results lack practical assurances, I would have appreciated more experimental validation, especially in complex and realistic settings. \n\nObtaining a Bayesian oracle could be very challenging (posterior distribution). \n\nOverall, while the paper introduces a promising method for designing safer AI systems, it would greatly benefit from additional components (both theoretical and experimental) before publication.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730774214463}, {"id": "DgWvP835G1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission424/Reviewer_pbrg"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper studies the problem of bounding the probability of some event in the context of an unknown but consistent distribution and a Bayesian setting. \n\nThe paper is motivated by the prevention of harm by AI agents. In short, harm is inherently unavoidable since in real applications we have no direct access to the distribution governing the environment. However, if we assume a fixed distribution, and a prior assumption of that distribution, we can get better and better approximations when data is presented to us, by using the data to update our prior knowledge of the distribution. With this, we can theoretically bound the probability of doing harm. In deployment, actions whose probability of harm is larger than some threshold can be blocked.\n\nThe paper explores two cases: incoming data as iid and non iid, and obtains bounds on the probability of harm in both cases.\n\nThe paper presents an experimental evaluation on a multi-armed bandits example, blocking actions that are considered unsafe according to the different bounds obtained as well as a baseline (with an unrealistic assumption of the underlying model). The paper ends with a discussion of the open problems still to be solved to be able to use this method as a reliable guardrails for AI agents.", "review_text": "This paper studies the problem of bounding the probability of some event in the context of an unknown but consistent distribution and a Bayesian setting. \n\nThe paper is motivated by the prevention of harm by AI agents. In short, harm is inherently unavoidable since in real applications we have no direct access to the distribution governing the environment. However, if we assume a fixed distribution, and a prior assumption of that distribution, we can get better and better approximations when data is presented to us, by using the data to update our prior knowledge of the distribution. With this, we can theoretically bound the probability of doing harm. In deployment, actions whose probability of harm is larger than some threshold can be blocked.\n\nThe paper explores two cases: incoming data as iid and non iid, and obtains bounds on the probability of harm in both cases.\n\nThe paper presents an experimental evaluation on a multi-armed bandits example, blocking actions that are considered unsafe according to the different bounds obtained as well as a baseline (with an unrealistic assumption of the underlying model). The paper ends with a discussion of the open problems still to be solved to be able to use this method as a reliable guardrails for AI agents.", "strengths": "S1. The topic of AI safety is timely and relevant for ICLR.\n\nS2. The theoretical results (as far as I could check) are sound.\n\nS3. The experimental evaluation serves to showcase how these bounds could be used in a realistic scenario.", "weaknesses": "W1. I understand the appeal to frame this work in the context of harm by an AI agent, and I think it is an interesting point. However, there is nothing inherent to \"harm\" in the concept presented. The concept of \"harm\" could be substituted by \"reward at a state\" and we could be discussing the same results in a different light. I think the paper may benefit from a more general motivation.\n\nW2. While the experimental evaluation is welcome, it is a very simple example, and one wonders if these theoretical bounds would find applicability in problems that are more complex and close to the real applications of guardrails.\n\nW3. The concept of guardrails presented here, as an algorithm that blocks an action if it shows an expected harm larger than some threshold, is very similar to the concept of probabilistic shielding in MDPs [1] (which is essentially the \"cheating guardrail in Sec. 5), and this can be extended to partially observable MDPs to eliminate the (unrealistic) assumption of having full knowledge of the ground truth [2]. The paper would benefit from comparing to these methods, especially with [2].\n\nW4. The paper does not engage in some recent work on defining harm in similar scenarios, see for example [3] or [4]. It could be useful to understand, in light of different definitions of harm, whether the results are specific to harm prevention, or can be framed in a more general understanding of bounds over rewards.\n\n\n\nOTHER (MINOR) REMARKS\n\nR1. The paper is mathematically dense and difficult to follow in parts. I'm not sure whether this is a weakness on its own, but I have the feeling that the ideas conveyed are simpler than the dense mathematical presentation seems to suggest. \n\n\nREFERENCES\n\n[1] N. Jansen et al. Safe Reinforcement Learning Using Probabilistic Shields. CONCUR 2020.\n\n[2] S. Carr et al. Safe Reinforcement Learning via Shielding under Partial Observability. AAAI 2024.\n\n[3] S. Beckers et al. Quantifying Harm. IJCAI 2023.\n\n[4] J. G. Richens. Counterfactual harm. NeurIPS 2022.", "questions": "Q1. How do you envision these guardrails to be applied in realistic scenarios? For example, consider the situation of a language model trying to obtain your passwords, or an autonomous car trying to crash with another vehicle. Could this notion of harm be applied efficiently to these realistic scenarios?\n\nQ2. How sensitive are the results to the choice of priors in the Bayesian framework? Can the authors discuss the robustness of the proposed approach under different prior choices?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of bounding the probability of some event in the context of an unknown but consistent distribution and a Bayesian setting. \n\nThe paper is motivated by the prevention of harm by AI agents. In short, harm is inherently unavoidable since in real applications we have no direct access to the distribution governing the environment. However, if we assume a fixed distribution, and a prior assumption of that distribution, we can get better and better approximations when data is presented to us, by using the data to update our prior knowledge of the distribution. With this, we can theoretically bound the probability of doing harm. In deployment, actions whose probability of harm is larger than some threshold can be blocked.\n\nThe paper explores two cases: incoming data as iid and non iid, and obtains bounds on the probability of harm in both cases.\n\nThe paper presents an experimental evaluation on a multi-armed bandits example, blocking actions that are considered unsafe according to the different bounds obtained as well as a baseline (with an unrealistic assumption of the underlying model). The paper ends with a discussion of the open problems still to be solved to be able to use this method as a reliable guardrails for AI agents.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "S1. The topic of AI safety is timely and relevant for ICLR.\n\nS2. The theoretical results (as far as I could check) are sound.\n\nS3. The experimental evaluation serves to showcase how these bounds could be used in a realistic scenario.", "weaknesses": "W1. I understand the appeal to frame this work in the context of harm by an AI agent, and I think it is an interesting point. However, there is nothing inherent to \"harm\" in the concept presented. The concept of \"harm\" could be substituted by \"reward at a state\" and we could be discussing the same results in a different light. I think the paper may benefit from a more general motivation.\n\nW2. While the experimental evaluation is welcome, it is a very simple example, and one wonders if these theoretical bounds would find applicability in problems that are more complex and close to the real applications of guardrails.\n\nW3. The concept of guardrails presented here, as an algorithm that blocks an action if it shows an expected harm larger than some threshold, is very similar to the concept of probabilistic shielding in MDPs [1] (which is essentially the \"cheating guardrail in Sec. 5), and this can be extended to partially observable MDPs to eliminate the (unrealistic) assumption of having full knowledge of the ground truth [2]. The paper would benefit from comparing to these methods, especially with [2].\n\nW4. The paper does not engage in some recent work on defining harm in similar scenarios, see for example [3] or [4]. It could be useful to understand, in light of different definitions of harm, whether the results are specific to harm prevention, or can be framed in a more general understanding of bounds over rewards.\n\n\n\nOTHER (MINOR) REMARKS\n\nR1. The paper is mathematically dense and difficult to follow in parts. I'm not sure whether this is a weakness on its own, but I have the feeling that the ideas conveyed are simpler than the dense mathematical presentation seems to suggest. \n\n\nREFERENCES\n\n[1] N. Jansen et al. Safe Reinforcement Learning Using Probabilistic Shields. CONCUR 2020.\n\n[2] S. Carr et al. Safe Reinforcement Learning via Shielding under Partial Observability. AAAI 2024.\n\n[3] S. Beckers et al. Quantifying Harm. IJCAI 2023.\n\n[4] J. G. Richens. Counterfactual harm. NeurIPS 2022.", "questions": "Q1. How do you envision these guardrails to be applied in realistic scenarios? For example, consider the situation of a language model trying to obtain your passwords, or an autonomous car trying to crash with another vehicle. Could this notion of harm be applied efficiently to these realistic scenarios?\n\nQ2. How sensitive are the results to the choice of priors in the Bayesian framework? Can the authors discuss the robustness of the proposed approach under different prior choices?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730669454268}, {"id": "gnvSTQYFLn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission424/Reviewer_ADYh"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper studies the problem of evaluating an unknown world model from observed data to determine whether it satisfies a certain safety metric. The safety metric, or guardrail, is a binary variable $H$, taking other variables in the world model as input. The authors utilize a Bayesian approach. It assumes access to the actual prior distribution over the groundtruth world model. The authors first prove that under certain parametric assumptions, the posterior distribution over candidate models will uniquely converge to the ground-truth model at the limit of large samples. Building on this concentration results, the authors derive an upper bound over the posterior probability of the harmful event $H = 1$ conditioning on the observed data. This concentration bound is then extended to non-i.i.d settings where observed samples are correlated. Finally, simulations were performed, and results supported the proposed theory.", "review_text": "This paper studies the problem of evaluating an unknown world model from observed data to determine whether it satisfies a certain safety metric. The safety metric, or guardrail, is a binary variable $H$, taking other variables in the world model as input. The authors utilize a Bayesian approach. It assumes access to the actual prior distribution over the groundtruth world model. The authors first prove that under certain parametric assumptions, the posterior distribution over candidate models will uniquely converge to the ground-truth model at the limit of large samples. Building on this concentration results, the authors derive an upper bound over the posterior probability of the harmful event $H = 1$ conditioning on the observed data. This concentration bound is then extended to non-i.i.d settings where observed samples are correlated. Finally, simulations were performed, and results supported the proposed theory.", "strengths": "- The paper is well-organized and clearly written. All the theoretical assumptions have been stated.\n- The proposed concentration results seem reasonable. The derivations seem technically sound.\n- Training large AI systems to satisfy certain safety criteria (i.e., with guardrails) is an exciting problem. This paper formulates this problem as a hypothesis-testing problem and presents non-trivial algorithms to perform the test. This problem formulation could be inspiring for other AI researchers across domains.", "weaknesses": "- The concentration result in Prop. 3.1 assumes \"all theories in $M$ are distinct as probability measures.\" This assumption does not seem to hold many common probabilistic models. For instance, in the linear component analysis, the number of independent components is generally not uniquely discernible (i.e., not identifiable) with non-linear mixing functions. Also, the number of latent components in Gaussian mixtures is generally not identifiable from the observed data. This seems to suggest that the application of the proposed concentration results might be limited.\n- The proposed concentration results also assume access to the actual prior distribution generating the ground-truth world model. It is unclear whether the upper bound could still hold when the prior distribution is unknown and misspecified.\n- Other concentration bounds exist over the target estimates using Baysian methods. Generally, one should be able to translate empirical concentration bounds to the Bayesian settings. For instance, (Osband & Van Roy, ICML'17) translates the concentration bounds for online reinforcement learning to Bayesian regret. How does the proposed method compare to other related work? This paper should include a section discussing related work in large deviation theory and how this paper is situated in the existing literature.\n\n- Reference: _\"Osband, Ian, and Benjamin Van Roy. \"Why is posterior sampling better than optimism for reinforcement learning?.\" International conference on machine learning. PMLR, 2017.\"_", "questions": "1. How does the upper bound in Prop. 3.4 apply if the prior distribution $P$ is misspecified?\n2. How does this work compare to the existing literature on concentration bounds in the Bayesian setting? For instance, these methods could include analysis of Bayesian regret in RL, and PAC Bayes.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of evaluating an unknown world model from observed data to determine whether it satisfies a certain safety metric. The safety metric, or guardrail, is a binary variable $H$, taking other variables in the world model as input. The authors utilize a Bayesian approach. It assumes access to the actual prior distribution over the groundtruth world model. The authors first prove that under certain parametric assumptions, the posterior distribution over candidate models will uniquely converge to the ground-truth model at the limit of large samples. Building on this concentration results, the authors derive an upper bound over the posterior probability of the harmful event $H = 1$ conditioning on the observed data. This concentration bound is then extended to non-i.i.d settings where observed samples are correlated. Finally, simulations were performed, and results supported the proposed theory.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-organized and clearly written. All the theoretical assumptions have been stated.\n- The proposed concentration results seem reasonable. The derivations seem technically sound.\n- Training large AI systems to satisfy certain safety criteria (i.e., with guardrails) is an exciting problem. This paper formulates this problem as a hypothesis-testing problem and presents non-trivial algorithms to perform the test. This problem formulation could be inspiring for other AI researchers across domains.", "weaknesses": "- The concentration result in Prop. 3.1 assumes \"all theories in $M$ are distinct as probability measures.\" This assumption does not seem to hold many common probabilistic models. For instance, in the linear component analysis, the number of independent components is generally not uniquely discernible (i.e., not identifiable) with non-linear mixing functions. Also, the number of latent components in Gaussian mixtures is generally not identifiable from the observed data. This seems to suggest that the application of the proposed concentration results might be limited.\n- The proposed concentration results also assume access to the actual prior distribution generating the ground-truth world model. It is unclear whether the upper bound could still hold when the prior distribution is unknown and misspecified.\n- Other concentration bounds exist over the target estimates using Baysian methods. Generally, one should be able to translate empirical concentration bounds to the Bayesian settings. For instance, (Osband & Van Roy, ICML'17) translates the concentration bounds for online reinforcement learning to Bayesian regret. How does the proposed method compare to other related work? This paper should include a section discussing related work in large deviation theory and how this paper is situated in the existing literature.\n\n- Reference: _\"Osband, Ian, and Benjamin Van Roy. \"Why is posterior sampling better than optimism for reinforcement learning?.\" International conference on machine learning. PMLR, 2017.\"_", "questions": "1. How does the upper bound in Prop. 3.4 apply if the prior distribution $P$ is misspecified?\n2. How does this work compare to the existing literature on concentration bounds in the Bayesian setting? For instance, these methods could include analysis of Bayesian regret in RL, and PAC Bayes.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730662190505}, {"id": "ivJJB5uIzs", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission424/Reviewer_bi5T"], "rating": 5, "soundness": 4, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper is tackling the problem of safety in AI. The authors take the view of defining safety as avoiding certain undesirable states in specific contexts.\nThey introduce a framework based on Bayesian inference from which an agent can derive safe policies that come with (probabilistic) guarantees of preventing harm.\nThe approach is safe-by-design, i.e. able to prevent undesired outcomes even if no concrete example of harmful states was ever observed in the system.", "review_text": "The paper is tackling the problem of safety in AI. The authors take the view of defining safety as avoiding certain undesirable states in specific contexts.\nThey introduce a framework based on Bayesian inference from which an agent can derive safe policies that come with (probabilistic) guarantees of preventing harm.\nThe approach is safe-by-design, i.e. able to prevent undesired outcomes even if no concrete example of harmful states was ever observed in the system.", "strengths": "The main strength of the paper is the introduction of a (as far as I am aware) novel view at safe-by-design in AI at runtime and opening the possibilities for future Bayesian methods to utilize the safety guarantees shown in the paper. Allowing for safety guarantees and steering future work in a direction that empashizes those is a significant problem in AI.\nI especially appreciate the discussion on open problems of the approach in the conclusion.\n\nThe theory being developed is also quite general and spans over a wide range of possible systems/problems.\n\nThe paper is well motivated and generally well structured, introducing formal concepts as needed in the respective sections. A small experimental evaluation is performed and well discussed. Proofs are provided in the appendix and I could not find any mistakes.", "weaknesses": "My main critique points of the paper are the lack of technical novelty (or at least it is not clarified enough if there are new results) and questions on applicability.\n\nFor the former, essentially all Propositions and Lemmata are either adaptations of well known results (Prop 3.1), taken from previous literature (Lemma 4.1, Prop 4.2), or rather simple Corollary derived from them (Lemma 3.3, Props. 3.4, 4.4, 4.5, 4.6). For Prop. 4.2 is is shown that the result is tight (Remark 4.3). To me it did not became clear whether this is a known result or a new contribution. It is also not clear whether the derived results (Props. 4.4, 4.5, 4.6) are also tight as a consequence or whether there is room for improvement.\nFor Prop. 4.5 and 4.6 in particular, restricting the possible world models indeces to $\\mathcal{I}^{\\alpha}_{1\\colon t}$ is essential, however, the choice of definition of $\\mathcal{I}^{\\alpha}$ is not really motivated. At the same time, Fig. 2(a) shows a substantial gap between applying Prop 4.6 in practice, and the theoretical optimum. This begs the question whether a different definition of $\\mathcal{I}^{\\alpha}$ (e.g. a simple cutoff, or requiring $\\mathcal{I}^{\\alpha}$ to have a certain probability mass) has potential to yield tighter bounds. However, as the definition of $\\mathcal{I}^{\\alpha}$ is not motivated, these questions remain unadressed.\n\nFor the applicability, my core concern is that the main problem in AI is not providing safety guarantees under certain assumptions, but rather designing a Bayesian agent that actually works well for a given problem while satisfying these assumptions. To go into detail, section 3 only provides \"law of large numbers\"-style guarantees which are not useful in practice. A small paragraph on the rate of convergence (which would be very helpful to know) is included but essentially is very problem-dependent and thus not discussed in detail in this more general framework. In the experimental evaluation, where Prop. 3.4 is utilized, it is not even clear whether $t$ is large enough for the guarantee statement of Prop. 3.4 to hold (on top of Prop 3.4 not being applicable due to non-i.i.d. as the authors mention themselves). Section 4 then relaxes to probabilistic guarantees, which is a more practical approach. However, to apply the results of section 4 in practice it ultimately relies on defining a hyperparameter alpha. On the theoretical side, the guarantees in section 4 only hold if alpha is chosen small enough (which is impossible to know without knowing the system in the first place) and on the practical side, the evaluation in section 5 shows that choosing alpha too large can have catastrophic consequences, even for the simple bandit system considered in section 5. In summary, I do not see any immediate way to take advantage of the theoretical results the paper provides. This is also amplified by the fact that the main body essentially does not discuss related work, and how existing approaches can be embedded into the framework.\n\n*These weaknesses make the paper feel like more of a statement paper with some additional mathematical background, rather than a fully fletched research paper.*\n\nAs a minor comment, from a reader's POV, the paper can be hard to follow at times, especially in the formal sections. Many paragraphs are written in a very technical way, assuming a deep mathematical background. While this surely can be expected from an audience like ICLR, I feel like many sections disrupt the flow of the paper, e.g. the two paragraphs \"Setting\" (l.155ff and l. 268ff). While these are defnitely important to make the paper rigorous, they are not strictly required to convey the main ideas of the paper. In the interest of readability, it might be advantageous to instead outsource the technical definitions to a separate section.", "questions": "1. Can you detail how you can utilize the CLT to obtain convergence rates (line 238ff)? If you applied this to the example in seciton 5, would it yield practical bounds?\n\n2. Are the results in Propositions 4.4 to 4.6 tight (in a similar vein as Remark 4.3 shows for Proposition 4.2)?\n\n3. How do you motivate the definition of $\\mathcal{I}^{\\alpha}_{1\\colon t}$ and have you considered different approaches?\n\n4. Can you provide some heuristics on choosing a safe, yet effective $\\alpha$ a priori? Which information might be helpful for this from e.g. which model paramteres have the biggest impact on $\\alpha$ and what information from a domain expert could be incorporated?\n\n5. Can you make any predictions on how you proposed guardrails perform on larger, more complex models? In particular, how do you expect the overestimation of harm (see Fig. 2) to be affected?\n\n6. Are there any existing works in which your framework fits, i.e. for which you can give (probabilistic) guarantees where they were previously unavailable?\nIf not, are there certain settings in which you can make reasonable a priori assumptions such that your framework is applicable and concrete guarantees can be derived for a given data set?\n\nminor comments:\n- line 96: explain what $q$ is\n- line 193: introduce delta as dirac notation beforehand\n- the axis and legends in the figures in section 5 are barely readable", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper is tackling the problem of safety in AI. The authors take the view of defining safety as avoiding certain undesirable states in specific contexts.\nThey introduce a framework based on Bayesian inference from which an agent can derive safe policies that come with (probabilistic) guarantees of preventing harm.\nThe approach is safe-by-design, i.e. able to prevent undesired outcomes even if no concrete example of harmful states was ever observed in the system.", "soundness": 4, "presentation": 2, "contribution": 2, "strengths": "The main strength of the paper is the introduction of a (as far as I am aware) novel view at safe-by-design in AI at runtime and opening the possibilities for future Bayesian methods to utilize the safety guarantees shown in the paper. Allowing for safety guarantees and steering future work in a direction that empashizes those is a significant problem in AI.\nI especially appreciate the discussion on open problems of the approach in the conclusion.\n\nThe theory being developed is also quite general and spans over a wide range of possible systems/problems.\n\nThe paper is well motivated and generally well structured, introducing formal concepts as needed in the respective sections. A small experimental evaluation is performed and well discussed. Proofs are provided in the appendix and I could not find any mistakes.", "weaknesses": "My main critique points of the paper are the lack of technical novelty (or at least it is not clarified enough if there are new results) and questions on applicability.\n\nFor the former, essentially all Propositions and Lemmata are either adaptations of well known results (Prop 3.1), taken from previous literature (Lemma 4.1, Prop 4.2), or rather simple Corollary derived from them (Lemma 3.3, Props. 3.4, 4.4, 4.5, 4.6). For Prop. 4.2 is is shown that the result is tight (Remark 4.3). To me it did not became clear whether this is a known result or a new contribution. It is also not clear whether the derived results (Props. 4.4, 4.5, 4.6) are also tight as a consequence or whether there is room for improvement.\nFor Prop. 4.5 and 4.6 in particular, restricting the possible world models indeces to $\\mathcal{I}^{\\alpha}_{1\\colon t}$ is essential, however, the choice of definition of $\\mathcal{I}^{\\alpha}$ is not really motivated. At the same time, Fig. 2(a) shows a substantial gap between applying Prop 4.6 in practice, and the theoretical optimum. This begs the question whether a different definition of $\\mathcal{I}^{\\alpha}$ (e.g. a simple cutoff, or requiring $\\mathcal{I}^{\\alpha}$ to have a certain probability mass) has potential to yield tighter bounds. However, as the definition of $\\mathcal{I}^{\\alpha}$ is not motivated, these questions remain unadressed.\n\nFor the applicability, my core concern is that the main problem in AI is not providing safety guarantees under certain assumptions, but rather designing a Bayesian agent that actually works well for a given problem while satisfying these assumptions. To go into detail, section 3 only provides \"law of large numbers\"-style guarantees which are not useful in practice. A small paragraph on the rate of convergence (which would be very helpful to know) is included but essentially is very problem-dependent and thus not discussed in detail in this more general framework. In the experimental evaluation, where Prop. 3.4 is utilized, it is not even clear whether $t$ is large enough for the guarantee statement of Prop. 3.4 to hold (on top of Prop 3.4 not being applicable due to non-i.i.d. as the authors mention themselves). Section 4 then relaxes to probabilistic guarantees, which is a more practical approach. However, to apply the results of section 4 in practice it ultimately relies on defining a hyperparameter alpha. On the theoretical side, the guarantees in section 4 only hold if alpha is chosen small enough (which is impossible to know without knowing the system in the first place) and on the practical side, the evaluation in section 5 shows that choosing alpha too large can have catastrophic consequences, even for the simple bandit system considered in section 5. In summary, I do not see any immediate way to take advantage of the theoretical results the paper provides. This is also amplified by the fact that the main body essentially does not discuss related work, and how existing approaches can be embedded into the framework.\n\n*These weaknesses make the paper feel like more of a statement paper with some additional mathematical background, rather than a fully fletched research paper.*\n\nAs a minor comment, from a reader's POV, the paper can be hard to follow at times, especially in the formal sections. Many paragraphs are written in a very technical way, assuming a deep mathematical background. While this surely can be expected from an audience like ICLR, I feel like many sections disrupt the flow of the paper, e.g. the two paragraphs \"Setting\" (l.155ff and l. 268ff). While these are defnitely important to make the paper rigorous, they are not strictly required to convey the main ideas of the paper. In the interest of readability, it might be advantageous to instead outsource the technical definitions to a separate section.", "questions": "1. Can you detail how you can utilize the CLT to obtain convergence rates (line 238ff)? If you applied this to the example in seciton 5, would it yield practical bounds?\n\n2. Are the results in Propositions 4.4 to 4.6 tight (in a similar vein as Remark 4.3 shows for Proposition 4.2)?\n\n3. How do you motivate the definition of $\\mathcal{I}^{\\alpha}_{1\\colon t}$ and have you considered different approaches?\n\n4. Can you provide some heuristics on choosing a safe, yet effective $\\alpha$ a priori? Which information might be helpful for this from e.g. which model paramteres have the biggest impact on $\\alpha$ and what information from a domain expert could be incorporated?\n\n5. Can you make any predictions on how you proposed guardrails perform on larger, more complex models? In particular, how do you expect the overestimation of harm (see Fig. 2) to be affected?\n\n6. Are there any existing works in which your framework fits, i.e. for which you can give (probabilistic) guarantees where they were previously unavailable?\nIf not, are there certain settings in which you can make reasonable a priori assumptions such that your framework is applicable and concrete guarantees can be derived for a given data set?\n\nminor comments:\n- line 96: explain what $q$ is\n- line 193: introduce delta as dirac notation beforehand\n- the axis and legends in the figures in section 5 are barely readable", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730132722435}], "openreview_url": "https://openreview.net/forum?id=2Oh2EOcFSO", "arxiv_id": "2408.05284", "paper_pdf": "papers/2Oh2EOcFSO.pdf", "paper_pdf_sha256": "52a00b3b306bcd160fdc4e40a990c855137743b90f1fd91704818d0062a5639a", "paper_pdf_bytes": 424788, "paper_pdf_source": "openreview", "code_url": "https://github.com/saifh-github/conservative-bayesian-public", "code_repository": "saifh-github/conservative-bayesian-public", "code_commit": "7ab10e0654b8675536e369f5064e58c18cadb582", "code_archive": "repos/2Oh2EOcFSO.zip", "code_archive_sha256": "d403e6dfe2036c31e40d8a5ce32a4aa8a752e10a49f55b35a93e454a50e888a3", "code_archive_bytes": 40511, "code_file_count": 15, "code_extensions": {".py": 13, ".sh": 2}, "github_disk_usage_kb": 129, "github_languages": {"Python": 112121, "Shell": 1231}, "github_archived": false, "github_pushed_at": "2025-05-30T12:23:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/can-a-bayesian-oracle-prevent-harm-from-an"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "o0C2v4xTdS", "year": 2024, "status": "rejected", "title": "CoarsenConf: Equivariant Coarsening with Aggregated Attention for Molecular Conformer Generation", "authors": ["Danny Reidenbach", "Aditi S. Krishnapriyan"], "authorids": ["~Danny_Reidenbach1", "~Aditi_S._Krishnapriyan1"], "authors_source": "OpenReview API", "abstract": "Molecular conformer generation (MCG) is an important task in cheminformatics and drug discovery. The ability to efficiently generate low-energy 3D structures can avoid expensive quantum mechanical simulations, leading to accelerated virtual screenings and enhanced structural exploration. Several generative models have been developed for MCG, but many struggle to consistently produce high-quality conformers. To address these issues, we introduce CoarsenConf, which coarse-grains molecular graphs based on torsional angles and integrates them into an SE(3)-equivariant hierarchical variational autoencoder. Through equivariant coarse-graining, we aggregate the fine-grained atomic coordinates of subgraphs connected via rotatable bonds, creating a variable-length coarse-grained latent representation. Our model uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation, enabling efficient generation of accurate conformers. Furthermore, we evaluate the chemical and biochemical quality of our generated conformers on multiple downstream applications, including property prediction and oracle-based protein docking. Overall, CoarsenConf generates more accurate conformer ensembles compared to prior generative models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "CoMdUlDJIh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8206/Reviewer_6Hjw"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies molecular conformer generation (MCG). The proposed method is a new SE(3)-equivariant hierarchical variational autoencoder that leverages coarse-grains molecular graphs with torsional angels. The proposed attention mechanism enables variable-length coarse-to-fine generation that restores high-quality conformers from coarse-grain graphs in an autoregressive way. This framework more efficiently generates more accurate conformer ensembles.", "review_text": "This paper studies molecular conformer generation (MCG). The proposed method is a new SE(3)-equivariant hierarchical variational autoencoder that leverages coarse-grains molecular graphs with torsional angels. The proposed attention mechanism enables variable-length coarse-to-fine generation that restores high-quality conformers from coarse-grain graphs in an autoregressive way. This framework more efficiently generates more accurate conformer ensembles.", "strengths": "1. **Novelty.** As the authors claimed, this paper proposed a novel pipeline to generate conformers using a SE3-equivariant hierarchical VAE and aggregated attention.\n2. **One stone three birds:** **efficiency, flexibilty and quality.** In contrary to prior works, the proposed method can generate all size of conformers by a single model whereas some existing approaches require a model for one length resulting in 100+ models to learn a dataset. This unified model learns more parameter-efficiently and more effectively with virtually more samples per model.\n3. **Competitive performance.** Experimental results in Table 3 and 4 show that the proposed method achieve competitive performance on Protein Docking and Binding Affinity compared to two or three baselines.", "weaknesses": "1. Weak performance on GEOM-DRUGS compared to Torsional Diffusion. In addition, the performances of baselines are different from the literature. Also, Recall should be reported as well. Please explain what causes the discrepancy. \n2. Only few baselines are provided. If more baselines are provided, then it will be better to evaluate the effectiveness of the proposed method compared to recent techniques.", "questions": "1. QM9 and ZINC250 have been used for learning molecular distributions. Also, several representations have been used for generation such as string (SMILES, SELFIES) and (2D) graphs. Is it possible to compare MCG with other graphs or string based methods? Often papers provide other groups of approaches in tables as references with dim fonts. \n2. Coarse-to-fine is a popular strategy in many applications. The conditional generation idea can be generalized in other directions. 2D to 3D generation is quite popular in the computer vision domain. Have you ever considered other coarse representations? and how robust/sensitive is the proposed pipeline to the quality of coarse-grain generation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies molecular conformer generation (MCG). The proposed method is a new SE(3)-equivariant hierarchical variational autoencoder that leverages coarse-grains molecular graphs with torsional angels. The proposed attention mechanism enables variable-length coarse-to-fine generation that restores high-quality conformers from coarse-grain graphs in an autoregressive way. This framework more efficiently generates more accurate conformer ensembles.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. **Novelty.** As the authors claimed, this paper proposed a novel pipeline to generate conformers using a SE3-equivariant hierarchical VAE and aggregated attention.\n2. **One stone three birds:** **efficiency, flexibilty and quality.** In contrary to prior works, the proposed method can generate all size of conformers by a single model whereas some existing approaches require a model for one length resulting in 100+ models to learn a dataset. This unified model learns more parameter-efficiently and more effectively with virtually more samples per model.\n3. **Competitive performance.** Experimental results in Table 3 and 4 show that the proposed method achieve competitive performance on Protein Docking and Binding Affinity compared to two or three baselines.", "weaknesses": "1. Weak performance on GEOM-DRUGS compared to Torsional Diffusion. In addition, the performances of baselines are different from the literature. Also, Recall should be reported as well. Please explain what causes the discrepancy. \n2. Only few baselines are provided. If more baselines are provided, then it will be better to evaluate the effectiveness of the proposed method compared to recent techniques.", "questions": "1. QM9 and ZINC250 have been used for learning molecular distributions. Also, several representations have been used for generation such as string (SMILES, SELFIES) and (2D) graphs. Is it possible to compare MCG with other graphs or string based methods? Often papers provide other groups of approaches in tables as references with dim fonts. \n2. Coarse-to-fine is a popular strategy in many applications. The conditional generation idea can be generalized in other directions. 2D to 3D generation is quite popular in the computer vision domain. Have you ever considered other coarse representations? and how robust/sensitive is the proposed pipeline to the quality of coarse-grain generation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698888223261}, {"id": "bjMZay0Wvm", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8206/Reviewer_vgnz"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes CoarsenConf, a novel conditional hierarchical VAE for molecular conformer generation. In particular, CoarsenConf aggregates the fine-grained atomic coordinates of subgraphs connected via rotatable bonds to create a variable-length coarse-grained latent representation, and uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation.", "review_text": "This paper proposes CoarsenConf, a novel conditional hierarchical VAE for molecular conformer generation. In particular, CoarsenConf aggregates the fine-grained atomic coordinates of subgraphs connected via rotatable bonds to create a variable-length coarse-grained latent representation, and uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation.", "strengths": "1. The idea of this work is straightforward and novel.\n2. The entire model can be trained end-to-end, and generate more accurate conformer ensembles compared to prior generative models. Besides, it shows very good performance on multiple downstream applications.", "weaknesses": "1. Authors say that they are the first model to employ variable-length coarse-graining. As far as I know, it has already been used in the molecular field (e.g., Qiang B, Song Y, Xu M, et al. Coarse-to-fine: a hierarchical diffusion model for molecule generation in 3D, ICML2023).\n2. The model architecture needs further explanations. For example, the description of encoder architecture in your appendix is unclear. What are the inputs and outputs of the three modules? How to get outputs based on inputs? Please reorganize this section.\n3. The presentation needs further improvement. This paper does not provide any algorithm for the proposed method, making me very confused about a lot of training and inference details.\n4. As shown in Table 1, Table 5 and Table 6, the performance of this work seems to be suboptimal, especially for Recall.", "questions": "1. Compared with current ML methods, how efficient is this method?\n2. Why is there a lack of comparison with Geodiff in many experiments? They have already released their code.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CoarsenConf, a novel conditional hierarchical VAE for molecular conformer generation. In particular, CoarsenConf aggregates the fine-grained atomic coordinates of subgraphs connected via rotatable bonds to create a variable-length coarse-grained latent representation, and uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The idea of this work is straightforward and novel.\n2. The entire model can be trained end-to-end, and generate more accurate conformer ensembles compared to prior generative models. Besides, it shows very good performance on multiple downstream applications.", "weaknesses": "1. Authors say that they are the first model to employ variable-length coarse-graining. As far as I know, it has already been used in the molecular field (e.g., Qiang B, Song Y, Xu M, et al. Coarse-to-fine: a hierarchical diffusion model for molecule generation in 3D, ICML2023).\n2. The model architecture needs further explanations. For example, the description of encoder architecture in your appendix is unclear. What are the inputs and outputs of the three modules? How to get outputs based on inputs? Please reorganize this section.\n3. The presentation needs further improvement. This paper does not provide any algorithm for the proposed method, making me very confused about a lot of training and inference details.\n4. As shown in Table 1, Table 5 and Table 6, the performance of this work seems to be suboptimal, especially for Recall.", "questions": "1. Compared with current ML methods, how efficient is this method?\n2. Why is there a lack of comparison with Geodiff in many experiments? They have already released their code.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698721858698}, {"id": "Xlw3JHizAl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8206/Reviewer_qQEF"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a molecular conformer generation framework based on the coarse-graining of molecular graphs. Its main idea is to learn a coarse-grained latent representation based on an encoder with a \"multi-resolution\" message passing structure and autoregressive ly decode the conformer from coarse latent representations. Experiments demonstrate performance improvement over existing works such as torsional diffusion for applications like property prediction and oracle-based protein docking.", "review_text": "This paper proposes a molecular conformer generation framework based on the coarse-graining of molecular graphs. Its main idea is to learn a coarse-grained latent representation based on an encoder with a \"multi-resolution\" message passing structure and autoregressive ly decode the conformer from coarse latent representations. Experiments demonstrate performance improvement over existing works such as torsional diffusion for applications like property prediction and oracle-based protein docking.", "strengths": "Overall, I think this work provides a solid and incremental contribution to molecular conformer generation.\n\n- To my knowledge, this work is the first to apply coarse-graining for molecular conformed generation. \n- It is interesting to see that autoregressive decoding is still useful for molecular conformer generation (compared to existing diffusion-based techniques).\n- The proposed idea can be extended to other tasks like generating molecules from scratch.\n- The experiments seem solid enough to verify the usefulness of the proposed method.", "weaknesses": "Since the proposed architecture is a bit complex, it is hard to identify the main source of performance improvement in the architecture. For example, one might argue that most of the improvement comes from (a) using substructures with fixed 3D coordinates and (b) using an encoder with a pooling layer. However, (a) has been proposed by torsional diffusion paper, and (b) has been investigated by the GNN community. It would be nice if the authors could design an ablation study on the effectiveness of each architectural component.", "questions": "I have the impression that this paper is in fact quite related to the torsional diffusion paper, e.g., both paper uses molecular substructures as fixed building blocks for molecular conformed generation. \n\nCould the authors elaborate more specifically on the difference and the benefits of using the coarse-graining procedure?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a molecular conformer generation framework based on the coarse-graining of molecular graphs. Its main idea is to learn a coarse-grained latent representation based on an encoder with a \"multi-resolution\" message passing structure and autoregressive ly decode the conformer from coarse latent representations. Experiments demonstrate performance improvement over existing works such as torsional diffusion for applications like property prediction and oracle-based protein docking.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Overall, I think this work provides a solid and incremental contribution to molecular conformer generation.\n\n- To my knowledge, this work is the first to apply coarse-graining for molecular conformed generation. \n- It is interesting to see that autoregressive decoding is still useful for molecular conformer generation (compared to existing diffusion-based techniques).\n- The proposed idea can be extended to other tasks like generating molecules from scratch.\n- The experiments seem solid enough to verify the usefulness of the proposed method.", "weaknesses": "Since the proposed architecture is a bit complex, it is hard to identify the main source of performance improvement in the architecture. For example, one might argue that most of the improvement comes from (a) using substructures with fixed 3D coordinates and (b) using an encoder with a pooling layer. However, (a) has been proposed by torsional diffusion paper, and (b) has been investigated by the GNN community. It would be nice if the authors could design an ablation study on the effectiveness of each architectural component.", "questions": "I have the impression that this paper is in fact quite related to the torsional diffusion paper, e.g., both paper uses molecular substructures as fixed building blocks for molecular conformed generation. \n\nCould the authors elaborate more specifically on the difference and the benefits of using the coarse-graining procedure?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698720624089}, {"id": "lCroNdd5Vj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8206/Reviewer_NaAP"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors proposed a coarse-grained method for molecule conformer generation, through an SE(3)-equivariant hierarchical VAE. The method is able to do coarse-graining generation with variable length via an aggregated attention strategy. The proposed method achieved state-of-the-art performance across a set of downstream tasks, including structural precision, property prediction, and docking binding affinity.", "review_text": "The authors proposed a coarse-grained method for molecule conformer generation, through an SE(3)-equivariant hierarchical VAE. The method is able to do coarse-graining generation with variable length via an aggregated attention strategy. The proposed method achieved state-of-the-art performance across a set of downstream tasks, including structural precision, property prediction, and docking binding affinity.", "strengths": "The performance is promising.", "weaknesses": "1. Problem Significance: The authors may need to demonstrate the problem of conformation generation remains significant, in the context of the rapid development of 3D molecule generation from scratch.\n2. Novelty: There has been a line of work studying coarse-grained molecule generation in the community [1, 2]. The authors may need to further discuss the novelty of their methods in comparison to these existing methods.\n3. Novelty Again: There has been another work proposing its information fusion attention that is similar to the aggregated attention strategy [3].\n\n[1]. Jin et al. Junction Tree Variational Autoencoder for Molecular Graph Generation. https://arxiv.org/pdf/1802.04364.pdf\n[2]. Zhang et al. Molecule Generation For Target Protein Binding with Structural Motifs. https://openreview.net/forum?id=Rq13idF0F73\n[3]. Wang et al. Retrieval-based Controllable Molecule Generation. https://arxiv.org/pdf/2208.11126.pdf", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors proposed a coarse-grained method for molecule conformer generation, through an SE(3)-equivariant hierarchical VAE. The method is able to do coarse-graining generation with variable length via an aggregated attention strategy. The proposed method achieved state-of-the-art performance across a set of downstream tasks, including structural precision, property prediction, and docking binding affinity.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The performance is promising.", "weaknesses": "1. Problem Significance: The authors may need to demonstrate the problem of conformation generation remains significant, in the context of the rapid development of 3D molecule generation from scratch.\n2. Novelty: There has been a line of work studying coarse-grained molecule generation in the community [1, 2]. The authors may need to further discuss the novelty of their methods in comparison to these existing methods.\n3. Novelty Again: There has been another work proposing its information fusion attention that is similar to the aggregated attention strategy [3].\n\n[1]. Jin et al. Junction Tree Variational Autoencoder for Molecular Graph Generation. https://arxiv.org/pdf/1802.04364.pdf\n[2]. Zhang et al. Molecule Generation For Target Protein Binding with Structural Motifs. https://openreview.net/forum?id=Rq13idF0F73\n[3]. Wang et al. Retrieval-based Controllable Molecule Generation. https://arxiv.org/pdf/2208.11126.pdf", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698225784433}], "openreview_url": "https://openreview.net/forum?id=o0C2v4xTdS", "arxiv_id": "2306.14852", "paper_pdf": "papers/o0C2v4xTdS.pdf", "paper_pdf_sha256": "80351d2da14573ec35251b8a919633546f798c071d37dc031488a118deac1862", "paper_pdf_bytes": 1982418, "paper_pdf_source": "openreview", "code_url": "https://github.com/ASK-Berkeley/CoarsenConf", "code_repository": "ASK-Berkeley/CoarsenConf", "code_commit": "1751189fbfc9874515f8bdd396b586a3ac2ff53b", "code_archive": "repos/o0C2v4xTdS.zip", "code_archive_sha256": "7260925cb09875ca2ee4a1b96415498c49cf1a3ddf7d4f2390c3f45c5b66fdf0", "code_archive_bytes": 120594, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 99, "github_languages": {"Python": 486162}, "github_archived": false, "github_pushed_at": "2024-10-24T06:59:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/coarsenconf-equivariant-coarsening-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WFBksaezAs", "year": 2023, "status": "rejected", "title": "Multi-Prompt Alignment for Multi-source Unsupervised Domain Adaptation", "authors": ["Haoran Chen", "Zuxuan Wu", "Yu-Gang Jiang"], "authorids": ["~Haoran_Chen4", "~Zuxuan_Wu1", "~Yu-Gang_Jiang1"], "authors_source": "OpenReview API", "abstract": "Most existing methods for multi-source unsupervised domain adaptation (UDA) rely on a common feature encoder to extract domain-invariant features. However, learning such an encoder involves updating the parameters of the entire network, which makes the optimization computationally expensive, particularly when coupled with min-max objectives. Inspired by recent advances in prompt learning that adapts high-capacity deep models for downstream tasks in a computationally economic way, we introduce Multi-Prompt Alignment (MPA), a simple yet efficient two-stage framework for multi-source UDA. Given a source and target domain pair, MPA first trains an individual prompt to minimize the domain gap through a contrastive loss, while tuning only a small set of parameters. Then, MPA derives a low-dimensional latent space through an auto-encoding process that maximizes the agreement of multiple learned prompts. The resulting embedding further facilitates generalization to unseen domains. Extensive experiments show that our method achieves state-of-the-art results on popular benchmark datasets while requiring substantially fewer tunable parameters. To the best of our knowledge, we are the first to apply prompt learning to the multi-source UDA problem and our method achieves the highest reported average accuracy of 54.1% on DomainNet, the most challenging UDA dataset to date, with only 15.9M parameters trained. More importantly, we demonstrate that the learned embedding space can be easily adapted to novel unseen domains.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "4TsJaSoogV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2433/Reviewer_iJht"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces prompt learning to multi-source UDA. A simple two-stage framework is proposed. In the first stage, individual prompts for each source and target pair are learned. In the second stage, Multi-Prompt Alignment (MPA) is proposed to align the learned prompts. The experimental results show the effectiveness of the proposed framework.", "review_text": "Overall, this is a good submission with strong results. I hope the author could provide more detailed explanations. ", "strengths": "Strengths:\n1.\tThis paper is first to apply prompt learning to multi-source UDA problem.\n2.\tThe proposed framework is simple and achieves state-of-the-art results on multi-source UDA benchmark.\n3.\tThe paper is well written, easy to understand.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces prompt learning to multi-source UDA. A simple two-stage framework is proposed. In the first stage, individual prompts for each source and target pair are learned. In the second stage, Multi-Prompt Alignment (MPA) is proposed to align the learned prompts. The experimental results show the effectiveness of the proposed framework.", "strength_and_weaknesses": "Strengths:\n1.\tThis paper is first to apply prompt learning to multi-source UDA problem.\n2.\tThe proposed framework is simple and achieves state-of-the-art results on multi-source UDA benchmark.\n3.\tThe paper is well written, easy to understand.\n", "clarity,_quality,_novelty_and_reproducibility": "- What is the difference between stage one and DAPL? Stage one (learning individual prompts) seems to be similar to DAPL. \n\n- The effectiveness of L_CLS loss in stage 2. Table.5 (a) shows the effectiveness of objective function Equation 9. However, the results of L_CLS loss are missed. \n\n- What does “Zero” mean in Table.5 (b)? “Zero auto-encoders means we completely discarded the auto-encoder structure” is not clear to me. If the autoencoder structure is discarded, what structure is adopted?\n\n- The similarity between the reconstructed prompts. In Table.4, the reconstructed prompts of different domains achieve almost the same results on the target domain. Does this mean that these reconstructed prompts are the same?", "summary_of_the_review": "Overall, this is a good submission with strong results. I hope the author could provide more detailed explanations. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666766098730}, {"id": "sknscmsW5j", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2433/Reviewer_Zi5U"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a multi-prompt alignment method for multi-source unsupervised domain adaptation. The main idea is to first learn an individual prompt for each source-target domain pair and then mine the relationships among learned prompts through deriving a shared embedding space. The resulting embedding is expected to be domain-invariant and can be generalize to unseen domains. ", "review_text": "In summary, the overall novelty is limited. Moreover, the design of the model is not very well justified, which requires further clarification.", "strengths": "- Strength\n    1. The introduction of prompt learning into multi-source domain adaptation is interesting. \n    2. The performance of the proposed method on MSDA is good compared to previous methods.\n- Weaknesses\n    1. The idea of prompt learning for domain adaptation is not new, and the prompt design in this paper is highly based on the method proposed by Ge et al., 2022. So the main contribution is to extend the prompt learning for domain adaptation from the single source domain to multiple source domains. The novelty is limited.\n    2. How to ensure the two-stage training will achieve the desired solutions? Since the individual prompts are learned to reduce domain shift between each individual source and the target domain while the multi-prompt alignment stage aims to align multiple source domains. The learning objective is changing in the two stages. Is it possible to combine the two stages into a single one to achieve the alignment of all the domains simultaneously?\n    3. Why is the auto-encoder necessary? If the two stages of prompt learning and alignment strategy can be combined into a single stage with multiple objectives for alignment and consistent prediction, is it still require the auto-encoder for reconstruction?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a multi-prompt alignment method for multi-source unsupervised domain adaptation. The main idea is to first learn an individual prompt for each source-target domain pair and then mine the relationships among learned prompts through deriving a shared embedding space. The resulting embedding is expected to be domain-invariant and can be generalize to unseen domains. ", "strength_and_weaknesses": "- Strength\n    1. The introduction of prompt learning into multi-source domain adaptation is interesting. \n    2. The performance of the proposed method on MSDA is good compared to previous methods.\n- Weaknesses\n    1. The idea of prompt learning for domain adaptation is not new, and the prompt design in this paper is highly based on the method proposed by Ge et al., 2022. So the main contribution is to extend the prompt learning for domain adaptation from the single source domain to multiple source domains. The novelty is limited.\n    2. How to ensure the two-stage training will achieve the desired solutions? Since the individual prompts are learned to reduce domain shift between each individual source and the target domain while the multi-prompt alignment stage aims to align multiple source domains. The learning objective is changing in the two stages. Is it possible to combine the two stages into a single one to achieve the alignment of all the domains simultaneously?\n    3. Why is the auto-encoder necessary? If the two stages of prompt learning and alignment strategy can be combined into a single stage with multiple objectives for alignment and consistent prediction, is it still require the auto-encoder for reconstruction?", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written in general. However, the originality of the work is limited.", "summary_of_the_review": "In summary, the overall novelty is limited. Moreover, the design of the model is not very well justified, which requires further clarification.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666674311955}, {"id": "l0L8koVnkKh", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2433/Reviewer_xddt"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a method based on prompt learning for multi-source domain adaptation. By adopting a pre-trained CLIP based text and image encoders, the authors design prompts that are learnable to adapt them for source and target domains. Specifically, there are two stages. First, prompts are designed with class-specific and domain-specific ones, and are learned for each source and target pairs individually via a contrastive loss. Then, to further align between target prompts, the authors apply the autoencoder-based method for prompt reconstruction, with a loss function to ensure output predictions are similar from different prompts. In addition, the authors also show a finetuning scheme that can adapt to new domains with small finetuning parameters. Experiments are conducted on three benchmark settings, including ImageCLEF, Office-Home, and DomainNet.", "review_text": "Overall, using prompt learning is an interesting direction for domain adaptation. However, there can be issues by using pre-trained CLIP that already has strong generalization ability for DA. In addition, some design choices in the proposed method are not well motivated nor well validated. The authors should consider the above comments and address them carefully in the rebuttal.", "strengths": "**Strength**\n\n* The idea of using prompt learning for domain adaptation is interesting, as prompt learning has the potential for transfer learning in the new domain under a cheaper cost\n* Experimental results show improvements over the CLIP baseline\n* Ablation study shows some gains for the designed modules\n\n**Weakness**\n\nTechnical motivation and novelty\n* Although applying prompt learning for domain adaptation is interesting, the motivation is not very intuitive as prompt learning is based on a highly generalized CLIP model, which already suffers less in the domain adaptation setting. This may raise two questions: 1) whether the experimental comparisons are fair (see comments below), and 2) whether the alignment really happens as CLIP already performs very well in target domains (see comments below).\n\n* The proposed techniques in this paper are not new, e.g., extending the prompt design to multi-domains is straightforward and the loss functions (e.g., contrastive loss) are also not new. More importantly, some designs are not well motivated and lacks experimental validations (discussed below).\n\n* For multi-prompt alignment, it's not clear how the autoencoder is needed to achieve alignment between prompts. In experimental results (Table 5 (a)), performance gain is also marginal.\n\n* It seems the performance of using prompt learning would highly depends on the CLIP model (e.g., Qdr results in Table 2). This should be discussed in the paper.\n\nTechnical clarity\n* For eq(8), it's not clear which prompts are aligned, e.g., between P_i and P_j\n* d_I appears many times in the paper but the definitions are not clear, e.g., eq(6), dimensions in v_tune and d_tune.\n\nExperimental results\n* Since CLIP already has a strong generalization ability, e.g., many results in Table 1 and 2 (Zero-Shot) are already better than other existing DA methods, it's not clear whether the comparisons are fair, as CLIP has been trained with millions of images. While I understand it's not easy to perform experiments using the CLIP backbone with existing methods, the proposed training scheme also cannot validate it's DA ability but only show the transfer learning ability.\n* Prompts are designed with class-specific and domain-specific ones. However, there are no experiments to validate whether it is necessary.\n* Lots of sensitivity experiments are not provided, e.g., length of prompts, thresholds for pseudo-labels, /alpha in eq(9)\n* The explanation of using LST for experiments is not clear (paragraph above Table 3)\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposes a method based on prompt learning for multi-source domain adaptation. By adopting a pre-trained CLIP based text and image encoders, the authors design prompts that are learnable to adapt them for source and target domains. Specifically, there are two stages. First, prompts are designed with class-specific and domain-specific ones, and are learned for each source and target pairs individually via a contrastive loss. Then, to further align between target prompts, the authors apply the autoencoder-based method for prompt reconstruction, with a loss function to ensure output predictions are similar from different prompts. In addition, the authors also show a finetuning scheme that can adapt to new domains with small finetuning parameters. Experiments are conducted on three benchmark settings, including ImageCLEF, Office-Home, and DomainNet.", "strength_and_weaknesses": "**Strength**\n\n* The idea of using prompt learning for domain adaptation is interesting, as prompt learning has the potential for transfer learning in the new domain under a cheaper cost\n* Experimental results show improvements over the CLIP baseline\n* Ablation study shows some gains for the designed modules\n\n**Weakness**\n\nTechnical motivation and novelty\n* Although applying prompt learning for domain adaptation is interesting, the motivation is not very intuitive as prompt learning is based on a highly generalized CLIP model, which already suffers less in the domain adaptation setting. This may raise two questions: 1) whether the experimental comparisons are fair (see comments below), and 2) whether the alignment really happens as CLIP already performs very well in target domains (see comments below).\n\n* The proposed techniques in this paper are not new, e.g., extending the prompt design to multi-domains is straightforward and the loss functions (e.g., contrastive loss) are also not new. More importantly, some designs are not well motivated and lacks experimental validations (discussed below).\n\n* For multi-prompt alignment, it's not clear how the autoencoder is needed to achieve alignment between prompts. In experimental results (Table 5 (a)), performance gain is also marginal.\n\n* It seems the performance of using prompt learning would highly depends on the CLIP model (e.g., Qdr results in Table 2). This should be discussed in the paper.\n\nTechnical clarity\n* For eq(8), it's not clear which prompts are aligned, e.g., between P_i and P_j\n* d_I appears many times in the paper but the definitions are not clear, e.g., eq(6), dimensions in v_tune and d_tune.\n\nExperimental results\n* Since CLIP already has a strong generalization ability, e.g., many results in Table 1 and 2 (Zero-Shot) are already better than other existing DA methods, it's not clear whether the comparisons are fair, as CLIP has been trained with millions of images. While I understand it's not easy to perform experiments using the CLIP backbone with existing methods, the proposed training scheme also cannot validate it's DA ability but only show the transfer learning ability.\n* Prompts are designed with class-specific and domain-specific ones. However, there are no experiments to validate whether it is necessary.\n* Lots of sensitivity experiments are not provided, e.g., length of prompts, thresholds for pseudo-labels, /alpha in eq(9)\n* The explanation of using LST for experiments is not clear (paragraph above Table 3)\n", "clarity,_quality,_novelty_and_reproducibility": "- Implementations are not provided\n- Some technical details are not very clear for reproducibility (see above comments)", "summary_of_the_review": "Overall, using prompt learning is an interesting direction for domain adaptation. However, there can be issues by using pre-trained CLIP that already has strong generalization ability for DA. In addition, some design choices in the proposed method are not well motivated nor well validated. The authors should consider the above comments and address them carefully in the rebuttal.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666641021031}, {"id": "UFD1FcAVixE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2433/Reviewer_rNJs"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper explores how CLIP can be used effectively for multi-source domain adaptation. In this problem setting there is a single labeled source domain dataset and multiple target domain datasets that are not labeled. Several methods have been proposed in past to address this problem, but they all have limited accuracy owing the limited abilities of ImageNet trained model (in comparison to CLIP dataset). Authors propose a simple idea to learn prompts using pseudo-labels from CLIP and propose an auto-encoder network to learn a latent space that generalizes to new domains. While not surprising, the proposed CLIP based method outperforms the existing methods. It is interesting to note that ResNet models initialized from the CLIP's image encoder do not perform as well.", "review_text": "Based on the strengths and weaknesses of the paper, I vote to accept the paper.", "strengths": "**Strengths**\n\n* The paper demonstrates how prompt engineering/learning can be performed in context of multi-source domain adaptation.\n* The paper shows that the the text encoder plays an important role in obtaining higher accuracy.\n* The results on the challenging domain-net dataset is impressive. The proposed method achieves state-of-the-art results with fewer trainable parameters.\n\n**Weaknesses**\n\n* While the paper includes ablation studies on other methods with image encoder, it is not clear (unfair comparison) if this enough to benchmark the proposed methods against the previous method. \n* It is not clear how to go about deciding the hyperparams $M_1$ and $M_2$. 16 seems a little too big for the proposed method.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper explores how CLIP can be used effectively for multi-source domain adaptation. In this problem setting there is a single labeled source domain dataset and multiple target domain datasets that are not labeled. Several methods have been proposed in past to address this problem, but they all have limited accuracy owing the limited abilities of ImageNet trained model (in comparison to CLIP dataset). Authors propose a simple idea to learn prompts using pseudo-labels from CLIP and propose an auto-encoder network to learn a latent space that generalizes to new domains. While not surprising, the proposed CLIP based method outperforms the existing methods. It is interesting to note that ResNet models initialized from the CLIP's image encoder do not perform as well.", "strength_and_weaknesses": "**Strengths**\n\n* The paper demonstrates how prompt engineering/learning can be performed in context of multi-source domain adaptation.\n* The paper shows that the the text encoder plays an important role in obtaining higher accuracy.\n* The results on the challenging domain-net dataset is impressive. The proposed method achieves state-of-the-art results with fewer trainable parameters.\n\n**Weaknesses**\n\n* While the paper includes ablation studies on other methods with image encoder, it is not clear (unfair comparison) if this enough to benchmark the proposed methods against the previous method. \n* It is not clear how to go about deciding the hyperparams $M_1$ and $M_2$. 16 seems a little too big for the proposed method.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and is novel in context of multi-source domain adaptation. The paper contains enough details for reproduction.", "summary_of_the_review": "Based on the strengths and weaknesses of the paper, I vote to accept the paper.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666632710083}, {"id": "XpFy-3Bcn1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2433/Reviewer_iJo5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper addresses the multi-source UDA problems by proposing the method Multi-Prompt Alignment (MPA). MPA is built on the pre-trained CLIP which can effectively encode images and texts. Compared to other existing approaches to mutli-source UDA, MPA only needs to train a small number of parameters by prompt learning.\n", "review_text": "The proposed approach is based on the pre-trained CLIP and utilises the prompt learning strategy to learn useful prompts for mutli-source UDA. However, the main concern is that it is not justified where the performance gains mainly come from, the proposed approach MPA or the pre-trained CLIP?", "strengths": "\n++ The proposed method based on CLIP can address the multi-source UDA problems by prompt learning which is more efficient during model training.\n++ The proposed method can generalise to unseen domains and the parameters to learn can be further reduced.\n++ The experimental results on three benchmark datasets show competitive performance compared with others. The ablation study also validates the effectiveness of different components of the proposed approach.\n\n-- It seems the threshold for pseudo-label selection is a very important hyper-parameter; the author should discuss how the value affects the performance.\n-- The proposed method uses pre-trained CLIP as the backbone. Since CLIP is pre-trained on a different dataset from other comparative models pre-trained on ImageNet. How to justify the performance gain is not from a more powerful pre-trained model when compared with other multi-source UDA approaches? A simple baseline can be \"source combine\" + \"simple prompt learning\".\n-- The authors claim the proposed approach can generalise to unseen domains by the introduced Latent Subspace Tuning (LST) strategy. Does it mean it can solve the domain generalisation problem? It seems pseudo labels are required in Eq.(10) which means the \"unseen\" domain is actually seen during training, please clarify this.\n-- Although the proposed method aims for multi-source UDA, it seems that it can also solve single-source UDA problems. How it performs when compared with SOTA UDA approaches?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper addresses the multi-source UDA problems by proposing the method Multi-Prompt Alignment (MPA). MPA is built on the pre-trained CLIP which can effectively encode images and texts. Compared to other existing approaches to mutli-source UDA, MPA only needs to train a small number of parameters by prompt learning.\n", "strength_and_weaknesses": "\n++ The proposed method based on CLIP can address the multi-source UDA problems by prompt learning which is more efficient during model training.\n++ The proposed method can generalise to unseen domains and the parameters to learn can be further reduced.\n++ The experimental results on three benchmark datasets show competitive performance compared with others. The ablation study also validates the effectiveness of different components of the proposed approach.\n\n-- It seems the threshold for pseudo-label selection is a very important hyper-parameter; the author should discuss how the value affects the performance.\n-- The proposed method uses pre-trained CLIP as the backbone. Since CLIP is pre-trained on a different dataset from other comparative models pre-trained on ImageNet. How to justify the performance gain is not from a more powerful pre-trained model when compared with other multi-source UDA approaches? A simple baseline can be \"source combine\" + \"simple prompt learning\".\n-- The authors claim the proposed approach can generalise to unseen domains by the introduced Latent Subspace Tuning (LST) strategy. Does it mean it can solve the domain generalisation problem? It seems pseudo labels are required in Eq.(10) which means the \"unseen\" domain is actually seen during training, please clarify this.\n-- Although the proposed method aims for multi-source UDA, it seems that it can also solve single-source UDA problems. How it performs when compared with SOTA UDA approaches?\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is written well; the presentation is clear and easy to understand; the novelty is a little bit weak; the experimental settings are presented and the results may be reproduced based on the details given in the manuscript.", "summary_of_the_review": "The proposed approach is based on the pre-trained CLIP and utilises the prompt learning strategy to learn useful prompts for mutli-source UDA. However, the main concern is that it is not justified where the performance gains mainly come from, the proposed approach MPA or the pre-trained CLIP?", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666529448243}], "openreview_url": "https://openreview.net/forum?id=WFBksaezAs", "arxiv_id": "2209.15210", "paper_pdf": "papers/WFBksaezAs.pdf", "paper_pdf_sha256": "c5cf38a8094aabbb98d1acd0a4f00b7e3b9da1fe36a08d263262524ef7d4591b", "paper_pdf_bytes": 1341029, "paper_pdf_source": "openreview", "code_url": "https://github.com/HaoranChen/Multi-Prompt-Alignment-for-MSUDA", "code_repository": "HaoranChen/Multi-Prompt-Alignment-for-MSUDA", "code_commit": "a80a7eb7dec9a7d18a90db70c7c3677d95a608e2", "code_archive": "repos/WFBksaezAs.zip", "code_archive_sha256": "cbfaf48bee95e9aca9be1866bb8d3c09d80955b2e44c7ec60d4a0c4e6827ca51", "code_archive_bytes": 182450, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 193, "github_languages": {"Python": 56463}, "github_archived": false, "github_pushed_at": "2025-01-26T14:29:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-prompt-alignment-for-multi-source"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PtuQ8bk9xF5", "year": 2022, "status": "rejected", "title": "Learning to Act with Affordance-Aware Multimodal Neural SLAM", "authors": ["Zhiwei Jia", "Kaixiang Lin", "Yizhou Zhao", "Qiaozi Gao", "Govind Thattai", "Gaurav S. Sukhatme"], "authorids": ["~Zhiwei_Jia1", "~Kaixiang_Lin1", "~Yizhou_Zhao1", "~Qiaozi_Gao1", "~Govind_Thattai1", "~Gaurav_S._Sukhatme1"], "authors_source": "OpenReview API", "abstract": "Recent years have witnessed an emerging paradigm shift toward embodied artificial intelligence, in which an agent must learn to solve challenging tasks by interacting with its environment. There are several challenges in solving embodied multimodal tasks, including long-horizon planning, vision-and-language grounding, and efficient exploration. We focus on a critical bottleneck, namely the performance of planning and navigation. To tackle this challenge, we propose a Neural SLAM approach that, for the first time, utilizes several modalities for exploration, predicts an affordance-aware semantic map, and plans over it at the same time. This significantly improves exploration efficiency, leads to robust long-horizon planning, and enables effective vision-and-language grounding. With the proposed Affordance-aware Multimodal Neural SLAM (AMSLAM) approach, we obtain more than 40% improvement over prior published work on the ALFRED benchmark and set a new state-of-the-art generalization performance at a success rate of 23.48% on the test unseen scenes.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "099LIraOPpL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3433/Reviewer_2ghY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper presents a Neural SLAM-based approach for tackling embodied multimodal tasks in ALFRED benchmark. The approach, called Affordance-aware Multimodal Neural SLAM (AMSLAM), utilizes several modalities for exploration, predicts an affordance-aware semantic map, and plans over it at the same time. The approach achieves 40% improvement over prior published work.", "review_text": "Strengths:\n- This submission is well motivated. Training robotic assistants to follow human instructions to complete an interactive task is an important and challenging problem. \n\n- The proposed method achieves good performance, improving 40% over prior published approaches to achieve 23.48% success rate. \n\nWeaknesses:\n- The writing clarity and readability can be improved significantly. The method seems extremely complicated and is very difficult to understand. There are way too many forward and backward references in Section 4. Even after reading over the methods section twice and going over the details in the Appendix, I am unable to understand many details in the method. For example:\n    - In sec 4, what is meaning of module operating at the \"subgoal level\"? In contrast to what in the E.T. paper? Why is the explored area considered an extra modality?\n    - In sec 4.1, how does the agent predict the task described in the language instructions? How does the agent know the subgoals in each task? How does the agent switch between exploration and execution phases? What is the difference between planned and prediction actions?  \n    - In Figure 1, it is unclear where and how the semantic map is used. It is unclear where the explored area is coming from (my understanding is that the environment does not provide the explored area). The caption refers to \"affordance-aware semantic representation\" which is not shown in the figure. What is the meaning of the \"actual\" task? \n    - In Sec 4.2, what are the delta rule-based functions?\n    - In Sec 4.3, how does the agent know the which low-level instruction corresponds to the current navigation and subsequent object interaction subgoals?\n\n- The method seems to use of lot of arbitrary choices, rules and hacks. For example, \"we define a single-step explored region as a binary map where a 5×3 rectangle grid (as a hard-coded field of view)\", \"We manually inject 4 RotateRight after every 2 MoveAhead predicted by our module to acquire 360◦ views of the scenes\", \"we further inject two LookUp or two LookDown actions alternately after every exploration action\", \"The agent then goes through 6 horizons {60◦, 45◦, ..., 0◦, −15◦} and obtains the mask prediction (with confidence > 0.8)\". All these choices seem to be specific to the Alfred benchmark. I can not imagine a robot operating in the real-world in this manner. My worry is that the authors have exploited unrealistic approximations in simulation environments (such as discrete action and state space, no noise in motion and pose sensors and use of high-level interaction actions), and over-optimized the design choices specific to this benchmark, which are likely to not result in better performance in more realistic settings.\n\n- The technical contributions of this paper are unclear:\n   - The authors claim in the introduction \"we propose the first multimodal exploration module that takes language instruction as guidance and keeps track of visited regions.\" I believe the Active Neural SLAM model, which AMSLAM is based on, also keeps track of visited regions to learn exploration. Based on authors' definition, even Active Neural SLAM is multimodal as it used pose, explored area and visual observations as input. Is the addition of language instructions as input the only change to the model?\n    - The authors claim they introduce \"Affordance-aware semantic representation that estimates object positions, heights, and relative spatial relationships\". I believe the semantic representations in Chaplot et al. 2020b and Blukis et al. 2021 also estimate object positions, heights and relative spatial relationships. It is unclear what makes the proposed representation \"affordance-aware\". ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a Neural SLAM-based approach for tackling embodied multimodal tasks in ALFRED benchmark. The approach, called Affordance-aware Multimodal Neural SLAM (AMSLAM), utilizes several modalities for exploration, predicts an affordance-aware semantic map, and plans over it at the same time. The approach achieves 40% improvement over prior published work.", "main_review": "Strengths:\n- This submission is well motivated. Training robotic assistants to follow human instructions to complete an interactive task is an important and challenging problem. \n\n- The proposed method achieves good performance, improving 40% over prior published approaches to achieve 23.48% success rate. \n\nWeaknesses:\n- The writing clarity and readability can be improved significantly. The method seems extremely complicated and is very difficult to understand. There are way too many forward and backward references in Section 4. Even after reading over the methods section twice and going over the details in the Appendix, I am unable to understand many details in the method. For example:\n    - In sec 4, what is meaning of module operating at the \"subgoal level\"? In contrast to what in the E.T. paper? Why is the explored area considered an extra modality?\n    - In sec 4.1, how does the agent predict the task described in the language instructions? How does the agent know the subgoals in each task? How does the agent switch between exploration and execution phases? What is the difference between planned and prediction actions?  \n    - In Figure 1, it is unclear where and how the semantic map is used. It is unclear where the explored area is coming from (my understanding is that the environment does not provide the explored area). The caption refers to \"affordance-aware semantic representation\" which is not shown in the figure. What is the meaning of the \"actual\" task? \n    - In Sec 4.2, what are the delta rule-based functions?\n    - In Sec 4.3, how does the agent know the which low-level instruction corresponds to the current navigation and subsequent object interaction subgoals?\n\n- The method seems to use of lot of arbitrary choices, rules and hacks. For example, \"we define a single-step explored region as a binary map where a 5×3 rectangle grid (as a hard-coded field of view)\", \"We manually inject 4 RotateRight after every 2 MoveAhead predicted by our module to acquire 360◦ views of the scenes\", \"we further inject two LookUp or two LookDown actions alternately after every exploration action\", \"The agent then goes through 6 horizons {60◦, 45◦, ..., 0◦, −15◦} and obtains the mask prediction (with confidence > 0.8)\". All these choices seem to be specific to the Alfred benchmark. I can not imagine a robot operating in the real-world in this manner. My worry is that the authors have exploited unrealistic approximations in simulation environments (such as discrete action and state space, no noise in motion and pose sensors and use of high-level interaction actions), and over-optimized the design choices specific to this benchmark, which are likely to not result in better performance in more realistic settings.\n\n- The technical contributions of this paper are unclear:\n   - The authors claim in the introduction \"we propose the first multimodal exploration module that takes language instruction as guidance and keeps track of visited regions.\" I believe the Active Neural SLAM model, which AMSLAM is based on, also keeps track of visited regions to learn exploration. Based on authors' definition, even Active Neural SLAM is multimodal as it used pose, explored area and visual observations as input. Is the addition of language instructions as input the only change to the model?\n    - The authors claim they introduce \"Affordance-aware semantic representation that estimates object positions, heights, and relative spatial relationships\". I believe the semantic representations in Chaplot et al. 2020b and Blukis et al. 2021 also estimate object positions, heights and relative spatial relationships. It is unclear what makes the proposed representation \"affordance-aware\". ", "summary_of_the_review": "Overall, the motivation is good and the method achieves a good performance on a public ALFRED benchmark. However, the authors seem to have made many design choices specific to this benchmark which are likely to not result in better results on realistic tasks. The technical contributions of this paper are unclear and the writing clarity needs to be improved significantly.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636013382265}, {"id": "usPN-RzIGB5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3433/Reviewer_pomV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new framework that improves the state-of-the-art performance of the ALFRED benchmark (accomplishing navigation and interaction tasks given language instructions in AI2THOR environments) by 40% relatively. Two claimed main technical contributions are: 1) leveraging multimodal inputs, specifically the newly incorporated language instructions and visited area maps, during exploration, 2) proposing an affordance-aware semantic representation that marks the object locations, heights, and possible agent interaction spots. Results that are 40% relatively better than the state-of-the-art are observed and several ablation studies over key network modules are also presented.", "review_text": "strengths:\n- results are much better than state-of-the-art and the improved margins are significant.\n- the proposed technical module of learning an affordance-aware semantic representation is valid, reasonable, and somewhat novel.\n- the system designs are really solid in mixing up many well-designed heuristics and learned modules into a big system that produces very good results.\n\nweaknesses:\n- my biggest concern is that the paper presentation is not selling the framework well as an easily-readable research paper, for example, a) there are no qualitative figures showing the performance of the work, only tables with numbers are presented. It is quite important to show qualitative results and your qualitative comparisons to baselines, additionally with some analysis figures clearly proving the effectiveness of each proposed module, for readers to be better convinced where your superior performance comes from. b) the writing for Sec. 4 can be improved. The current writing is quite messy mixing up the two contributions together. Also, many important details (e.g. the architecture to learn the affordance maps and the aggregation steps) are put in supplementary. c) for the experiments, the authors spent many more words on Sec 5.1 and 5.2 than Sec 5.3 (the main results). The main results subsection is very short and does not contain analysis over the tables or show qualitative figures to analyze the results.\n- another of my major concerns is whether or not using the language instructions during the exploration stage is a reasonable and realistic setting. My opinion is that we should not assume we know the task (goal) and the instructions for subgoals beforehand. and during the exploration. Otherwise, we need different exploration steps given different tasks. Or, at least, you need to report your exploration steps as part of the task execution steps, since you need different exploration steps given any new test task. Please clarify if I'm wrong with this. But, the confusion regarding this point prevents me from recognizing the claimed technical contribution on leveraging multimodal inputs during exploration.\n- though there are some novel designs for the affordance learning part, there are previous works (Qi et al, 2019, and Nagarajan & Grauman, 2020) that have proposed the essential ideas, and I don't see too much difference other than adding some heuristics-derived waypoint definitions. Please clarify if I missed or misunderstood the main differences from previous works.\n- though I recognize that the presented system achieves amazing results and significantly refreshes the state-of-the-art, it's unclear how much performance gains come from combining many state-of-the-art architectures, say transformers or bert, and how much comes from the claimed two novel technical design points. Can you compare to baseline methods with the same architectures or report the parameter amounts for different models, to give us a sense of this?\n- one minor issue is to use \\citep instead of \\cite in latex for a more valid citation format of papers", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new framework that improves the state-of-the-art performance of the ALFRED benchmark (accomplishing navigation and interaction tasks given language instructions in AI2THOR environments) by 40% relatively. Two claimed main technical contributions are: 1) leveraging multimodal inputs, specifically the newly incorporated language instructions and visited area maps, during exploration, 2) proposing an affordance-aware semantic representation that marks the object locations, heights, and possible agent interaction spots. Results that are 40% relatively better than the state-of-the-art are observed and several ablation studies over key network modules are also presented.", "main_review": "strengths:\n- results are much better than state-of-the-art and the improved margins are significant.\n- the proposed technical module of learning an affordance-aware semantic representation is valid, reasonable, and somewhat novel.\n- the system designs are really solid in mixing up many well-designed heuristics and learned modules into a big system that produces very good results.\n\nweaknesses:\n- my biggest concern is that the paper presentation is not selling the framework well as an easily-readable research paper, for example, a) there are no qualitative figures showing the performance of the work, only tables with numbers are presented. It is quite important to show qualitative results and your qualitative comparisons to baselines, additionally with some analysis figures clearly proving the effectiveness of each proposed module, for readers to be better convinced where your superior performance comes from. b) the writing for Sec. 4 can be improved. The current writing is quite messy mixing up the two contributions together. Also, many important details (e.g. the architecture to learn the affordance maps and the aggregation steps) are put in supplementary. c) for the experiments, the authors spent many more words on Sec 5.1 and 5.2 than Sec 5.3 (the main results). The main results subsection is very short and does not contain analysis over the tables or show qualitative figures to analyze the results.\n- another of my major concerns is whether or not using the language instructions during the exploration stage is a reasonable and realistic setting. My opinion is that we should not assume we know the task (goal) and the instructions for subgoals beforehand. and during the exploration. Otherwise, we need different exploration steps given different tasks. Or, at least, you need to report your exploration steps as part of the task execution steps, since you need different exploration steps given any new test task. Please clarify if I'm wrong with this. But, the confusion regarding this point prevents me from recognizing the claimed technical contribution on leveraging multimodal inputs during exploration.\n- though there are some novel designs for the affordance learning part, there are previous works (Qi et al, 2019, and Nagarajan & Grauman, 2020) that have proposed the essential ideas, and I don't see too much difference other than adding some heuristics-derived waypoint definitions. Please clarify if I missed or misunderstood the main differences from previous works.\n- though I recognize that the presented system achieves amazing results and significantly refreshes the state-of-the-art, it's unclear how much performance gains come from combining many state-of-the-art architectures, say transformers or bert, and how much comes from the claimed two novel technical design points. Can you compare to baseline methods with the same architectures or report the parameter amounts for different models, to give us a sense of this?\n- one minor issue is to use \\citep instead of \\cite in latex for a more valid citation format of papers", "summary_of_the_review": "Overall, I have a mixed feeling. On one hand, the performance is really good and significant. On the other hand, the two claimed technical contributions are not stated clearly about their significance of differences than the previous methods or the validity of the design. Also, this paper is more like a system report other than a well-written research paper with clearly presented and analyzed novel techniques. Adding qualitative figures and analysis will definitely help on this front.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635795842729}, {"id": "Tq6L2NPDwhu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3433/Reviewer_kMKr"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a method for solving the ALFRED task (following language instructions to perform a set of household tasks). The model has two main components: (1) Affordance-aware Semantic Representation (2) Multi-modal exploration. The former component estimates the location of the objects in the scene. The latter provides an exploration strategy by using instructions, images, previous actions and previously explored areas as input. ", "review_text": "**Strengths**\n\n- The method achieves the state-of-the-art performance on the ALFRED benchmark, which is quite challenging.\n\n- The set of ablation experiments are informative.\n\n\n**Weaknesses**\n\n- The paper is heavily engineered for the ALFRED benchmark. I am not sure if any of these modules can be used for generic embodied tasks.\n\n- The paper makes a big deal about the \"affordance-aware\" representation in the introduction and other places in the text, but it is actually a simple heuristic that the agent backs up a few steps when it is next to an object. According to the text \"the agent will back up 3 steps if the target object is a fridge, 2 steps if it is a safe, a cabinet or a drawer, and 1 step for everything else\". These types of heuristics are not generalizable or scalable. The affordance-aware representation should be learned instead.\n\n- There is a strong assumption that \"we know the changes (∆x, ∆y, ∆r) of the agent after each action\". These types of assumptions are valid only in simulation. Several previous works have made similar assumptions. However, those approaches are suitable only for simulation. In reality, the estimates for agent pose are quite noisy. I would like to see how robust the proposed approach is against noise in the agent pose.\n\n- It is not clear if the exploration steps and the steps used for capturing panoramic views are included in the evaluations.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method for solving the ALFRED task (following language instructions to perform a set of household tasks). The model has two main components: (1) Affordance-aware Semantic Representation (2) Multi-modal exploration. The former component estimates the location of the objects in the scene. The latter provides an exploration strategy by using instructions, images, previous actions and previously explored areas as input. ", "main_review": "**Strengths**\n\n- The method achieves the state-of-the-art performance on the ALFRED benchmark, which is quite challenging.\n\n- The set of ablation experiments are informative.\n\n\n**Weaknesses**\n\n- The paper is heavily engineered for the ALFRED benchmark. I am not sure if any of these modules can be used for generic embodied tasks.\n\n- The paper makes a big deal about the \"affordance-aware\" representation in the introduction and other places in the text, but it is actually a simple heuristic that the agent backs up a few steps when it is next to an object. According to the text \"the agent will back up 3 steps if the target object is a fridge, 2 steps if it is a safe, a cabinet or a drawer, and 1 step for everything else\". These types of heuristics are not generalizable or scalable. The affordance-aware representation should be learned instead.\n\n- There is a strong assumption that \"we know the changes (∆x, ∆y, ∆r) of the agent after each action\". These types of assumptions are valid only in simulation. Several previous works have made similar assumptions. However, those approaches are suitable only for simulation. In reality, the estimates for agent pose are quite noisy. I would like to see how robust the proposed approach is against noise in the agent pose.\n\n- It is not clear if the exploration steps and the steps used for capturing panoramic views are included in the evaluations.", "summary_of_the_review": "The paper proposes an approach which is heavily engineered for the ALFRED benchmark and I do not see a major novelty that generalizes to other embodied tasks. The method is mainly a mixture of heuristics and a set of pre-trained components that are glued together for the ALFRED tasks. However, it achieves good performance on the ALFRED benchmark, hence, my borderline rating. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634792374039}], "openreview_url": "https://openreview.net/forum?id=PtuQ8bk9xF5", "arxiv_id": "2201.09862", "paper_pdf": "papers/PtuQ8bk9xF5.pdf", "paper_pdf_sha256": "03966d3ae031a91e1ebbb81d190f98b28a8ed1d7ee7207d0904ed62cc0c3cff4", "paper_pdf_bytes": 3882110, "paper_pdf_source": "openreview", "code_url": "https://github.com/amazon-science/multimodal-neuralslam", "code_repository": "amazon-science/multimodal-neuralslam", "code_commit": "530558fdfa31c6e048fc3e7b253f681f6786b04d", "code_archive": "repos/PtuQ8bk9xF5.zip", "code_archive_sha256": "13c5ccbc0c43adc841eb6a20339c1aeae5030e61be3cd0117b4e5c99e9077800", "code_archive_bytes": 529752, "code_file_count": 91, "code_extensions": {".py": 68, ".c": 11, ".h": 11, ".sh": 1}, "github_disk_usage_kb": 367, "github_languages": {"Python": 598816, "C": 568481, "Yacc": 38288, "Lex": 5270, "Makefile": 1465, "Shell": 81}, "github_archived": false, "github_pushed_at": "2021-09-18T23:00:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-act-with-affordance-aware-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ot9bYHvuULl", "year": 2021, "status": "rejected", "title": "Augmented Sliced Wasserstein Distances", "authors": ["Xiongjie Chen", "Yongxin Yang", "Yunpeng Li"], "authorids": ["~Xiongjie_Chen1", "~Yongxin_Yang1", "~Yunpeng_Li1"], "authors_source": "OpenReview API", "abstract": "While theoretically appealing, the application of the Wasserstein distance to large-scale machine learning problems has been hampered by its prohibitive computational cost. The sliced Wasserstein distance and its variants improve the computational efficiency through random projection, yet they suffer from low projection efficiency because the majority of projections result in trivially small values. In this work, we propose a new family of distance metrics, called augmented sliced Wasserstein distances (ASWDs), constructed by first mapping samples to higher-dimensional hypersurfaces parameterized by neural networks. It is derived from a key observation that (random) linear projections of samples residing on these hypersurfaces would translate to much more flexible nonlinear projections in the original sample space, so they can capture complex structures of the data distribution. We show that the hypersurfaces can be optimized by gradient ascent efficiently. We provide the condition under which the ASWD is a valid metric and show that this can be obtained by an injective neural network architecture. Numerical results demonstrate that the ASWD significantly outperforms other Wasserstein variants for both synthetic and real-world problems.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "VHQZrxqjyLe", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1176/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper provides a notion of generalisation for sliced Wasserstein distances, that allows to explore nonlinear projections in arbitrary subspaces in a suitable and efficient way. The paper is well-written and provides a novel solution to the problem of exploiting nonlinear subspaces in computing distances. I feel this idea, although simple can be quite powerful, and can be extended to wide domains especially identifying objects based on arbitrary feature selections although I do not think the authors have explored that direction in this piece of work. \nComments:\n(1) Numerical results show the distance computed by ASWD is smaller than other measures. This can, however, be misleading in the sense that it might also be obtained by insufficient exploration of non-linear subspaces. For eg, if the sliced Wasserstein distances are computed along orthogonal directions to the primary features the Wasserstein distance obtained in that regard will also be very small, although this does not in anyway validate the superiority of the distance metric. I feel this argument therefore needs some clarification.\n(2) Can I use this method to construct arbitrary nonlinear projections of choice instead of the projection that gives the best distinction? For eg, suppose two pictures have several objects among which cars in the two pictures are the most distinguishing features. However, I also want to see what other features can help distinguish these two pictures from other pictures which do not have cars. How do I do that? In that broader sense I suppose my question deals with trying to identify barycenters of objects via sliced Wasserstein distances. Any thoughts in that regard could be useful for the reader.\n\nOverall I find the paper an interesting read although it requires some clarifications as outlined above.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clarification of numerical results necessary", "review": "The paper provides a notion of generalisation for sliced Wasserstein distances, that allows to explore nonlinear projections in arbitrary subspaces in a suitable and efficient way. The paper is well-written and provides a novel solution to the problem of exploiting nonlinear subspaces in computing distances. I feel this idea, although simple can be quite powerful, and can be extended to wide domains especially identifying objects based on arbitrary feature selections although I do not think the authors have explored that direction in this piece of work. \nComments:\n(1) Numerical results show the distance computed by ASWD is smaller than other measures. This can, however, be misleading in the sense that it might also be obtained by insufficient exploration of non-linear subspaces. For eg, if the sliced Wasserstein distances are computed along orthogonal directions to the primary features the Wasserstein distance obtained in that regard will also be very small, although this does not in anyway validate the superiority of the distance metric. I feel this argument therefore needs some clarification.\n(2) Can I use this method to construct arbitrary nonlinear projections of choice instead of the projection that gives the best distinction? For eg, suppose two pictures have several objects among which cars in the two pictures are the most distinguishing features. However, I also want to see what other features can help distinguish these two pictures from other pictures which do not have cars. How do I do that? In that broader sense I suppose my question deals with trying to identify barycenters of objects via sliced Wasserstein distances. Any thoughts in that regard could be useful for the reader.\n\nOverall I find the paper an interesting read although it requires some clarifications as outlined above.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604027657162}, {"id": "8nsKwlPIZNi", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1176/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces the augmented sliced Wasserstein distances (ASWD) to capture non-linear projections, as opposed to sliced Wasserstein distances. To this end, the architecture first maps inputs to higher dimensional hypersurfaces, before projecting. The designed ASWD is shown to be a metric as long as the initial mapping is injective. \n\nNumerically, the initial mapping is parameterized as the concatenation of the input and its output using a neural network. The loss aims at capturing a hypersurface that differentiates measures the most, and is regularized to control the initial mapping. The first task considers the minimization of sliced Wasserstein flows for different settings of synthetic data. The other task consists of generating images with GANs, using a sliced Wasserstein loss to learn the generator network.\n\nStrong points of the paper include:\n1.\tThe paper is well written. The literature review is very thorough and comparison with the proposed method is well explained.\n2.\tBoth theoretically and numerically, the method is well described, well compared to existing notions and yields convincing results.\n\nDrawbacks / questions:\n1.\tIt seems that a lot of projections are needed to retrieve visually satisfactory generated examples.\n2.\tHave you tried other types of injective maps than Eq 15?\n3.\tCan you comment on the impact of regularization strength lambda in practice?\n\nI recommend an accept for this paper, which, to the best of my knowledge, brings both theoretical and numerical valuable contributions to the literature.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Valuable contributions", "review": "This paper introduces the augmented sliced Wasserstein distances (ASWD) to capture non-linear projections, as opposed to sliced Wasserstein distances. To this end, the architecture first maps inputs to higher dimensional hypersurfaces, before projecting. The designed ASWD is shown to be a metric as long as the initial mapping is injective. \n\nNumerically, the initial mapping is parameterized as the concatenation of the input and its output using a neural network. The loss aims at capturing a hypersurface that differentiates measures the most, and is regularized to control the initial mapping. The first task considers the minimization of sliced Wasserstein flows for different settings of synthetic data. The other task consists of generating images with GANs, using a sliced Wasserstein loss to learn the generator network.\n\nStrong points of the paper include:\n1.\tThe paper is well written. The literature review is very thorough and comparison with the proposed method is well explained.\n2.\tBoth theoretically and numerically, the method is well described, well compared to existing notions and yields convincing results.\n\nDrawbacks / questions:\n1.\tIt seems that a lot of projections are needed to retrieve visually satisfactory generated examples.\n2.\tHave you tried other types of injective maps than Eq 15?\n3.\tCan you comment on the impact of regularization strength lambda in practice?\n\nI recommend an accept for this paper, which, to the best of my knowledge, brings both theoretical and numerical valuable contributions to the literature.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603813558497}, {"id": "qJV5jVTiOAr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1176/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Augmented Sliced Wasserstein distance between random vectors X&Y is defined as SWD between g(X) and g(Y) for an injective g. \n\nIt is claimed that this generalization solves the problem with SWD, i.e. a small distance for most of the projection directions.  Since the paper is purely experimental, then it is better to demonstrate that this issue is resolved (it is not obvious why the problem disappears).\n\nSince g_omega is also searched for by maximizing (16), then it becomes hard to interpret the value of metrics (also taking into account the role of lambda). What it measures now?\n\nThe code is given, results seem reproducible. Reported results with generative modeling compare ASWD with SWD, GSWD, DSWD. Since the paper is experimental, maybe it is natural to expect a comparison with SOTA generators that do not deal with generalizations of Wasserstein distance. Also, fake images generated by SWD, GSWD, DSWD are not given (only FID is shown).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The metrics is hard to interpret", "review": "Augmented Sliced Wasserstein distance between random vectors X&Y is defined as SWD between g(X) and g(Y) for an injective g. \n\nIt is claimed that this generalization solves the problem with SWD, i.e. a small distance for most of the projection directions.  Since the paper is purely experimental, then it is better to demonstrate that this issue is resolved (it is not obvious why the problem disappears).\n\nSince g_omega is also searched for by maximizing (16), then it becomes hard to interpret the value of metrics (also taking into account the role of lambda). What it measures now?\n\nThe code is given, results seem reproducible. Reported results with generative modeling compare ASWD with SWD, GSWD, DSWD. Since the paper is experimental, maybe it is natural to expect a comparison with SOTA generators that do not deal with generalizations of Wasserstein distance. Also, fake images generated by SWD, GSWD, DSWD are not given (only FID is shown).", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603723836478}], "openreview_url": "https://openreview.net/forum?id=ot9bYHvuULl", "arxiv_id": "2006.08812", "paper_pdf": "papers/ot9bYHvuULl.pdf", "paper_pdf_sha256": "c70413f8c6074aa320d42b715a68a9de03eb56b6ca45aa6e5687ccc1f194ab49", "paper_pdf_bytes": 1617521, "paper_pdf_source": "openreview", "code_url": "https://github.com/xiongjiechen/ASWD", "code_repository": "xiongjiechen/ASWD", "code_commit": "86ebb8303c7baeda56a903f9598928b70986e07b", "code_archive": "repos/ot9bYHvuULl.zip", "code_archive_sha256": "b5fb4974069f46995bffb6885513d0f7b7d09c2757baf064f7366b3b1e5085dc", "code_archive_bytes": 492580, "code_file_count": 8, "code_extensions": {".py": 7, ".ipynb": 1}, "github_disk_usage_kb": 492, "github_languages": {"Python": 65200, "Jupyter Notebook": 35383}, "github_archived": false, "github_pushed_at": "2025-06-14T12:47:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/augmented-sliced-wasserstein-distances"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJgctpEKwr", "year": 2020, "status": "rejected", "title": "RPGAN: random paths as a latent space for GAN interpretability", "authors": ["Andrey Voynov", "Artem Babenko"], "authorids": ["an.voynov@gmail.com", "artem.babenko@phystech.edu"], "authors_source": "OpenReview API", "abstract": "In this paper, we introduce Random Path Generative Adversarial Network (RPGAN) --- an alternative scheme of GANs that can serve as a tool for generative model analysis. While the latent space of a typical GAN consists of input vectors, randomly sampled from the standard Gaussian distribution, the latent space of RPGAN consists of random paths in a generator network. As we show, this design allows to associate different layers of the generator with different regions of the latent space, providing their natural interpretability. With experiments on standard benchmarks, we demonstrate that RPGAN reveals several interesting insights about roles that different layers play in the image generation process. Aside from interpretability, the RPGAN model also provides competitive generation quality and allows efficient incremental learning on new data.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SyeceWbLcr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper680/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a variation on the generator of GANs. The authors modify the generator by adding a concept of \"blocks\" which are randomly activated based on part of the random input vector. It is similar to adding random dropout in the generator, except that the dropout would apply to larger sets of activations instead of single component.\n\nThe authors also add a diversity term to force the different blocks to have different blocks. This is a term based on the L2 distance between the weights of different blocks. Although this term is ad-hoc and could probably be refined into something more grounded in theory, it should indeed provide some diversity.\n\nThis block structure allows for more understanding of what each layer of the generator does, since it is easy to change the discrete variables that switch blocks. The paper presents an empirical evaluation of what each switch does, and show that concepts are well disentangled between layers (for instance, one layer changes the background whereas another one changes the color of the foreground).\n\nThey also show that blocks can be added after training is done which is a nice property for incremental training.\n\nThey also show that this framework can train a generator without non-linearities (except for the block switching), which could potentially simplify the analysis of such networks.\n\nGenerated samples are presented up to 128x128 pixels which, although far from state of the art, proves that the concept works.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper presents a variation on the generator of GANs. The authors modify the generator by adding a concept of \"blocks\" which are randomly activated based on part of the random input vector. It is similar to adding random dropout in the generator, except that the dropout would apply to larger sets of activations instead of single component.\n\nThe authors also add a diversity term to force the different blocks to have different blocks. This is a term based on the L2 distance between the weights of different blocks. Although this term is ad-hoc and could probably be refined into something more grounded in theory, it should indeed provide some diversity.\n\nThis block structure allows for more understanding of what each layer of the generator does, since it is easy to change the discrete variables that switch blocks. The paper presents an empirical evaluation of what each switch does, and show that concepts are well disentangled between layers (for instance, one layer changes the background whereas another one changes the color of the foreground).\n\nThey also show that blocks can be added after training is done which is a nice property for incremental training.\n\nThey also show that this framework can train a generator without non-linearities (except for the block switching), which could potentially simplify the analysis of such networks.\n\nGenerated samples are presented up to 128x128 pixels which, although far from state of the art, proves that the concept works."}, "tcdate": 1572372721524}, {"id": "HklH2Ycr9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper680/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses the issue of interpretability of GAN generation through an alternative approach to the introduction of variability. To seed the generation, instead of providing a random input vector (typically sampled from a standard Gaussian distribution), the authors instead modify the generator architecture so as to allow for randomization in the routing: each layer is replaced by a bucket consisting of several blocks, and in forward propagation only through randomly chosen blocks. In this case, the input vector is chosen to be a constant - the only source of randomization\nprovided to the generator is in the choice of blocks through which to propagate. The explanability derives from the tendency of blocks to associate with a common interpretation after training. Their use of blocks necessitates the introduction of a block diversity loss, to discourage mode collapse. The scheme is referred to as RPGAN, for \"Random Path GAN\".\n\nThe main strengths of the paper:\n\n(1) Their proposed approach is highly flexible. In principle, any underlying GAN architecture can be adapted by assigning each layer to a distinct bucket, and then replicating the layer across the blocks. \n\n(2) Experimental results do show that different block sequences are associated with common image characteristics after training, especially for the initial and final layers.\n\n(3) The use of non-standard ways of introducing stochasticity to GAN generation is an interesting idea in itself.\n\nThe main weaknesses:\n\n(1) Although the authors provide experimental examples showing images associated with various paths through the architecture, is not clear how interpretations can be associated with these paths. In the examples presented, there seems to be a tendency for greater interpretability at the initial and final layers, with the explanation given for the intermediate layers being less convincing.\n\n(2) The number of experimental examples is low, yet the authors draw rather firm conclusions (end of Section 4.1) regarding interpretability across layers. I am not sure that their conclusions adequately capture what is going on here, nor am I convinced that they generalize to other situations.\n\n(3) The number of buckets limits the numbers of explanations. Essentially, the method has the same difficulties as in clustering, where specifying too many or too few groups can profoundly influence the nature and quality of the result. Although the authors do discuss an approach by which the number of buckets can be incrementally increased (thereby allowing for variation in the number of explanations generated), the experimental evidence is insufficient.\n\n(4) Presumably, the replication of a layer across the blocks assigned to its bucket would require more training data and/or greater training times. What is the relationship in both time and quality between the original GAN network and its RPGAN versions?\n\n(5) There are many presentational problems with this paper, in grammar, vocabulary and terminology, sentence structure, etc.\n\nOverall, in its current state (not least due to presentational issues) the paper appears to be below the acceptance threshold.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper addresses the issue of interpretability of GAN generation through an alternative approach to the introduction of variability. To seed the generation, instead of providing a random input vector (typically sampled from a standard Gaussian distribution), the authors instead modify the generator architecture so as to allow for randomization in the routing: each layer is replaced by a bucket consisting of several blocks, and in forward propagation only through randomly chosen blocks. In this case, the input vector is chosen to be a constant - the only source of randomization\nprovided to the generator is in the choice of blocks through which to propagate. The explanability derives from the tendency of blocks to associate with a common interpretation after training. Their use of blocks necessitates the introduction of a block diversity loss, to discourage mode collapse. The scheme is referred to as RPGAN, for \"Random Path GAN\".\n\nThe main strengths of the paper:\n\n(1) Their proposed approach is highly flexible. In principle, any underlying GAN architecture can be adapted by assigning each layer to a distinct bucket, and then replicating the layer across the blocks. \n\n(2) Experimental results do show that different block sequences are associated with common image characteristics after training, especially for the initial and final layers.\n\n(3) The use of non-standard ways of introducing stochasticity to GAN generation is an interesting idea in itself.\n\nThe main weaknesses:\n\n(1) Although the authors provide experimental examples showing images associated with various paths through the architecture, is not clear how interpretations can be associated with these paths. In the examples presented, there seems to be a tendency for greater interpretability at the initial and final layers, with the explanation given for the intermediate layers being less convincing.\n\n(2) The number of experimental examples is low, yet the authors draw rather firm conclusions (end of Section 4.1) regarding interpretability across layers. I am not sure that their conclusions adequately capture what is going on here, nor am I convinced that they generalize to other situations.\n\n(3) The number of buckets limits the numbers of explanations. Essentially, the method has the same difficulties as in clustering, where specifying too many or too few groups can profoundly influence the nature and quality of the result. Although the authors do discuss an approach by which the number of buckets can be incrementally increased (thereby allowing for variation in the number of explanations generated), the experimental evidence is insufficient.\n\n(4) Presumably, the replication of a layer across the blocks assigned to its bucket would require more training data and/or greater training times. What is the relationship in both time and quality between the original GAN network and its RPGAN versions?\n\n(5) There are many presentational problems with this paper, in grammar, vocabulary and terminology, sentence structure, etc.\n\nOverall, in its current state (not least due to presentational issues) the paper appears to be below the acceptance threshold.\n"}, "tcdate": 1572346285249}, {"id": "B1lOQlo7cB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper680/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes the Random Path Generative Adversarial Network (RP-GAN) to serve as a tool for generative model analysis.  The main idea is to have several different buckets in each block of the generator and then train the generator with random paths. To interpret the features captured by each block, the authors unfreeze one block and show the variance of the generated images via different buckets.\n\nThe contribution of the paper is limited. Most of the observations proposed in this paper are just to confirm the findings in the recent paper Bau et al (2019). And the authors did not clarify why their methods are better than the previous work. In addition, this work changes some standard ways of generating images. For example,  they use a fixed input vector Z rather than a random vector Z following Gaussian distribution. Then when they claim that their findings are also valid to standard GAN generators, e.g., SN-ResNet, they add the noise to the weights in SN-ResNet and claim that we can conclude similar findings as RP-GAN. Therefore, my question is if adding noise to the weights to a generator is sufficient for the interpretation of the generator, why do we still need this work for further interpretation?\n\nMinor:\nIn Figure 3, there are 10 images in each line but the number of buckets in each block is 40 for CIFAR10, as you claimed in Section 4. It should be clarified how you get those 10 images from 40 blocks in the unfreezed bucket.\n\nIn conclusion, I vote for a weak reject for this work. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper proposes the Random Path Generative Adversarial Network (RP-GAN) to serve as a tool for generative model analysis.  The main idea is to have several different buckets in each block of the generator and then train the generator with random paths. To interpret the features captured by each block, the authors unfreeze one block and show the variance of the generated images via different buckets.\n\nThe contribution of the paper is limited. Most of the observations proposed in this paper are just to confirm the findings in the recent paper Bau et al (2019). And the authors did not clarify why their methods are better than the previous work. In addition, this work changes some standard ways of generating images. For example,  they use a fixed input vector Z rather than a random vector Z following Gaussian distribution. Then when they claim that their findings are also valid to standard GAN generators, e.g., SN-ResNet, they add the noise to the weights in SN-ResNet and claim that we can conclude similar findings as RP-GAN. Therefore, my question is if adding noise to the weights to a generator is sufficient for the interpretation of the generator, why do we still need this work for further interpretation?\n\nMinor:\nIn Figure 3, there are 10 images in each line but the number of buckets in each block is 40 for CIFAR10, as you claimed in Section 4. It should be clarified how you get those 10 images from 40 blocks in the unfreezed bucket.\n\nIn conclusion, I vote for a weak reject for this work. "}, "tcdate": 1572216863929}], "openreview_url": "https://openreview.net/forum?id=BJgctpEKwr", "arxiv_id": null, "paper_pdf": "papers/BJgctpEKwr.pdf", "paper_pdf_sha256": "f14fb66da947161c58af4aca74fb9eb40b7029241227b72edb6b326d3c6acbdd", "paper_pdf_bytes": 23078746, "paper_pdf_source": "openreview", "code_url": "https://github.com/rpgan-ICLR2020/RPGAN", "code_repository": "rpgan-ICLR2020/RPGAN", "code_commit": "62b4e4be24d012be0191e8290f9b1b249e0e1b7f", "code_archive": "repos/BJgctpEKwr.zip", "code_archive_sha256": "1d29eac546de74a4cb5a369d5f2f0a297968de59bf2c4b5201d07de220e64847", "code_archive_bytes": 662765, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 640, "github_languages": {"Python": 86801}, "github_archived": false, "github_pushed_at": "2023-07-06T21:42:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rpgan-random-paths-as-a-latent-space-for-gan"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BFb4ACHayj", "year": 2026, "status": "rejected", "title": "SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents", "authors": ["Sheng Yin", "Xianghe Pang", "Yuanzhuo Ding", "Menglan Chen", "Yutong Bi", "Yichen Xiong", "Wenhao Huang", "Zhen Xiang", "Jing Shao", "Siheng Chen"], "authorids": ["~Sheng_Yin1", "~Xianghe_Pang1", "~Yuanzhuo_Ding1", "~Menglan_Chen1", "~Yutong_Bi1", "~Yichen_Xiong1", "~Wenhao_Huang4", "~Zhen_Xiang1", "~Jing_Shao3", "~Siheng_Chen1"], "authors_source": "OpenReview API", "abstract": "With the integration of large language models (LLMs), embodied agents have strong capabilities to understand and plan complicated natural language instructions. However, a foreseeable issue is that those embodied agents can also flawlessly execute some hazardous tasks, potentially causing damages in the real world. Existing benchmarks predominantly overlook critical safety risks, focusing solely on planning performance, while a few evaluate LLMs' safety awareness only on non-interactive image-text data. To address this gap, we present \\textbf{SafeAgentBench}—the first comprehensive benchmark for safety-aware task planning of embodied LLM agents in interactive simulation environments, covering both explicit and implicit hazards. SafeAgentBench includes: (1) an executable, diverse, and high-quality dataset of 750 tasks, rigorously curated to cover 10 potential hazards and 3 task types; (2) SafeAgentEnv, a universal embodied environment with a low-level controller, supporting multi-agent execution with 17 high-level actions for 9 state-of-the-art baselines; and (3) reliable evaluation methods from both execution and semantic perspectives. Experimental results show that, although agents based on different design frameworks exhibit substantial differences in task success rates, their overall safety awareness remains weak. The most safety-conscious baseline achieves only a 10\\% rejection rate for detailed hazardous tasks. Moreover, simply replacing the LLM driving the agent does not lead to notable improvements in safety awareness. Dataset and codes are available and shown in the reproducibility statement.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "az0SZ7m76p", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16964/Reviewer_57jZ"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents a benchmark designed to evaluate the safety awareness of embodied LLM agents in task planning and execution. The benchmark includes a set of executable tasks including various hazard types and abstraction levels, supported by a simulated environment for testing safety-aware planning behaviors. Experiments show that current embodied LLM agents have limited proactive safety awareness and often fail to identify or reject unsafe instructions. Two baseline defense methods were tested but proved only marginally effective. The paper’s main contribution is providing a benchmark and environment to analyze how embodied LLM agents manage safety risks in planning.", "review_text": "This paper presents a benchmark designed to evaluate the safety awareness of embodied LLM agents in task planning and execution. The benchmark includes a set of executable tasks including various hazard types and abstraction levels, supported by a simulated environment for testing safety-aware planning behaviors. Experiments show that current embodied LLM agents have limited proactive safety awareness and often fail to identify or reject unsafe instructions. Two baseline defense methods were tested but proved only marginally effective. The paper’s main contribution is providing a benchmark and environment to analyze how embodied LLM agents manage safety risks in planning.", "strengths": "1. The paper addresses a highly relevant and timely topic, focusing on the safety of embodied AI systems at a time when LLM-based robotic task planning is rapidly expanding.\n2. The benchmark fills an existing gap by shifting the focus from task completion to evaluating how agents respond to hazardous instructions.\n3. The work provides a useful starting point for further research on safety evaluation and benchmarking for embodied LLM agents.\n4. The design of SafeAgentBench is well thought out, incorporating both explicit and implicit hazards and covering ten risk categories and three abstraction levels.\n5. Including multiple LLM backbones makes the results more general and less model-dependent.\n6. The paper is well organized and easy to follow, with figures and tables that effectively illustrate the framework, task taxonomy, and experimental results.\n7. The semantic evaluation approach, validated through a user study, adds credibility to the evaluation process and complements execution-based evaluation.", "weaknesses": "1. The overall contribution feels incremental since the paper focuses mainly on dataset and benchmark construction rather than proposing new algorithms or methods that enhance safety.\n2. SafeAgentEnv adds limited novelty because it is largely an adaptation of AI2-THOR with only minor extensions.\n3. The evaluation is conducted entirely in simulation, and the paper does not discuss how the findings would transfer to real-world or physical robot scenarios.\n4. Mainly reliance on GPT-4 for both dataset generation and semantic evaluation introduces potential model bias, and the rationale for this dependence is not sufficiently justified.\n5. The dataset size and number of tested agents may be too small to claim generality or broad representativeness of the benchmark.\n6. The experiments on defense strategies are shallow and do not provide strong insight into how to improve agent safety.\n7. The description of the Chain-of-Thought safety filter and its implementation details are vague and difficult to understand.\n8. Some inconsistencies exist between the text, figures, and supplementary materials, such as differences between results presented with Gemini and DeepSeek.\n9. Several research questions are not fully supported with detailed evidence, and the paper could better articulate the novelty and impact of its contributions.\n10. The benchmark does not evaluate generalization to new domains such as factory or outdoor environments, which limits its applicability.\n11. The dependence on simulated results without deeper statistical validation leaves uncertainty about robustness and reproducibility.", "questions": "1. What results would the authors expect if newer models such as GPT-4.5 or GPT-5 were used?\n2. Could the authors clarify the \"Detailed GP Steps\" mentioned in Table 1?\n3. Would similar safety patterns appear in industrial or outdoor environments rather than domestic tasks?\n4. In Figure 2a, what do the percentages represent?\n5. Are all five harm types practically plausible in real robotic systems?\n6. How were the unsafe and safe instructions in Table 5 designed and validated?\n7. The description of the low-level controller in Lines 255-257 is unclear “Considering that VLA- or RL-trained controllers have not been available, our controller maps each of the 17 high-level actions into sequences of executable low-level APIs, incorporating feasibility checks such as verifying object reachability before execution”. Could the authors elaborate?\n8. Should KARMA also appear in Figure 1?\n9. Line 361 says 5 agents rejected none. Is this number correct?\n10. Was Gemini-2.5-pro evaluated on hazardous tasks?\n11. Who participated in the user study that produced 1,008 human ratings? Were they domain experts or general users?\n12. How important is safe task planning in simulation compared to real-world deployment?\n13. Will the dataset be publicly released?\n14. Is the dataset size of 750 tasks sufficient to capture diversity? Can you quantify its linguistic and behavioral variety?\n15. Why do abstract tasks include four related instructions?\n16. Why are there exactly 17 high-level actions, and are they comprehensive enough?\n17. Can SafeAgentBench be adapted to other simulators such as Habitat or IsaacSim, and what modifications would be needed?\n18. Have the authors tried fine-tuning LLMs on SafeAgentBench to see if safety awareness improves?\n19. When GPT-4 and human evaluators disagreed in semantic evaluation, what types of tasks caused the difference?\n20. Could the use of GPT-4 for dataset generation have introduced bias that favors GPT-4-based agents?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a benchmark designed to evaluate the safety awareness of embodied LLM agents in task planning and execution. The benchmark includes a set of executable tasks including various hazard types and abstraction levels, supported by a simulated environment for testing safety-aware planning behaviors. Experiments show that current embodied LLM agents have limited proactive safety awareness and often fail to identify or reject unsafe instructions. Two baseline defense methods were tested but proved only marginally effective. The paper’s main contribution is providing a benchmark and environment to analyze how embodied LLM agents manage safety risks in planning.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper addresses a highly relevant and timely topic, focusing on the safety of embodied AI systems at a time when LLM-based robotic task planning is rapidly expanding.\n2. The benchmark fills an existing gap by shifting the focus from task completion to evaluating how agents respond to hazardous instructions.\n3. The work provides a useful starting point for further research on safety evaluation and benchmarking for embodied LLM agents.\n4. The design of SafeAgentBench is well thought out, incorporating both explicit and implicit hazards and covering ten risk categories and three abstraction levels.\n5. Including multiple LLM backbones makes the results more general and less model-dependent.\n6. The paper is well organized and easy to follow, with figures and tables that effectively illustrate the framework, task taxonomy, and experimental results.\n7. The semantic evaluation approach, validated through a user study, adds credibility to the evaluation process and complements execution-based evaluation.", "weaknesses": "1. The overall contribution feels incremental since the paper focuses mainly on dataset and benchmark construction rather than proposing new algorithms or methods that enhance safety.\n2. SafeAgentEnv adds limited novelty because it is largely an adaptation of AI2-THOR with only minor extensions.\n3. The evaluation is conducted entirely in simulation, and the paper does not discuss how the findings would transfer to real-world or physical robot scenarios.\n4. Mainly reliance on GPT-4 for both dataset generation and semantic evaluation introduces potential model bias, and the rationale for this dependence is not sufficiently justified.\n5. The dataset size and number of tested agents may be too small to claim generality or broad representativeness of the benchmark.\n6. The experiments on defense strategies are shallow and do not provide strong insight into how to improve agent safety.\n7. The description of the Chain-of-Thought safety filter and its implementation details are vague and difficult to understand.\n8. Some inconsistencies exist between the text, figures, and supplementary materials, such as differences between results presented with Gemini and DeepSeek.\n9. Several research questions are not fully supported with detailed evidence, and the paper could better articulate the novelty and impact of its contributions.\n10. The benchmark does not evaluate generalization to new domains such as factory or outdoor environments, which limits its applicability.\n11. The dependence on simulated results without deeper statistical validation leaves uncertainty about robustness and reproducibility.", "questions": "1. What results would the authors expect if newer models such as GPT-4.5 or GPT-5 were used?\n2. Could the authors clarify the \"Detailed GP Steps\" mentioned in Table 1?\n3. Would similar safety patterns appear in industrial or outdoor environments rather than domestic tasks?\n4. In Figure 2a, what do the percentages represent?\n5. Are all five harm types practically plausible in real robotic systems?\n6. How were the unsafe and safe instructions in Table 5 designed and validated?\n7. The description of the low-level controller in Lines 255-257 is unclear “Considering that VLA- or RL-trained controllers have not been available, our controller maps each of the 17 high-level actions into sequences of executable low-level APIs, incorporating feasibility checks such as verifying object reachability before execution”. Could the authors elaborate?\n8. Should KARMA also appear in Figure 1?\n9. Line 361 says 5 agents rejected none. Is this number correct?\n10. Was Gemini-2.5-pro evaluated on hazardous tasks?\n11. Who participated in the user study that produced 1,008 human ratings? Were they domain experts or general users?\n12. How important is safe task planning in simulation compared to real-world deployment?\n13. Will the dataset be publicly released?\n14. Is the dataset size of 750 tasks sufficient to capture diversity? Can you quantify its linguistic and behavioral variety?\n15. Why do abstract tasks include four related instructions?\n16. Why are there exactly 17 high-level actions, and are they comprehensive enough?\n17. Can SafeAgentBench be adapted to other simulators such as Habitat or IsaacSim, and what modifications would be needed?\n18. Have the authors tried fine-tuning LLMs on SafeAgentBench to see if safety awareness improves?\n19. When GPT-4 and human evaluators disagreed in semantic evaluation, what types of tasks caused the difference?\n20. Could the use of GPT-4 for dataset generation have introduced bias that favors GPT-4-based agents?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761835476001}, {"id": "OeVdfu9eCT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16964/Reviewer_oNVY"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper presents SafeAgentBench, a benchmark designed to evaluate the safety awareness of task-planning embodied LLM agents within an interactive simulation environment (SafeAgentEnv, based on AI2-THOR). The benchmark's key contributions are: (1) A dataset of 750 tasks (450 hazardous, 300 safe) covering 10 hazard types and 3 task structures (detailed, abstract, long-horizon), which notably includes explicitly hazardous instructions often overlooked by prior work. (2) A dual evaluation system using both a rule-based Execution Evaluator and an LLM-based Semantic Evaluator. The authors evaluate 9 agent baselines, primarily driven by GPT-4, and find that their proactive safety awareness is \"weak.\" For instance, the highest rejection rate for detailed hazardous tasks is only 10%. The paper also shows that low risk rates are often due to poor planning capability ($\\theta$) rather than intrinsic safety awareness ($\\alpha$), and that simple defenses are ineffective.", "review_text": "The paper presents SafeAgentBench, a benchmark designed to evaluate the safety awareness of task-planning embodied LLM agents within an interactive simulation environment (SafeAgentEnv, based on AI2-THOR). The benchmark's key contributions are: (1) A dataset of 750 tasks (450 hazardous, 300 safe) covering 10 hazard types and 3 task structures (detailed, abstract, long-horizon), which notably includes explicitly hazardous instructions often overlooked by prior work. (2) A dual evaluation system using both a rule-based Execution Evaluator and an LLM-based Semantic Evaluator. The authors evaluate 9 agent baselines, primarily driven by GPT-4, and find that their proactive safety awareness is \"weak.\" For instance, the highest rejection rate for detailed hazardous tasks is only 10%. The paper also shows that low risk rates are often due to poor planning capability ($\\theta$) rather than intrinsic safety awareness ($\\alpha$), and that simple defenses are ineffective.", "strengths": "The paper addresses the critical and timely problem of embodied agent safety. As agents become more capable, understanding their failure modes and safety awareness is of paramount importance to the field.", "weaknesses": "## Major Weaknesses:\n\n**1.Critically Outdated and Irrelevant Model Selection:** \n\nThe paper's experimental setup is fundamentally flawed by its exclusive reliance on pure text-based LLMs.\n\n- **Lack of Multimodality:** Embodied agents, by definition, must perceive and interact with their environment. This requires processing multimodal inputs, primarily visual information (e.g., images, depth maps). The paper’s core evaluation, however, uses GPT-4 (a text-only model) as the central planner, with other text-only models (Llama3, Qwen2, DeepSeekV2.5) as alternatives. This setup is unrealistic. It does not evaluate an embodied agent but rather a disembodied text model's abstract knowledge of safety, which is not the paper's claimed contribution.\n\n- **Outdated Models:** The paper is submitted for 2026 but relies on models from 2023 (GPT-4). The field of LLMs is advancing rapidly. A contemporary evaluation would be expected to include SOTA multimodal models (e.g., GPT-4o, Gemini 1.5/2.5 Pro, Claude 3 family) which are now standard. The paper also fails to investigate whether advanced reasoning techniques (e.g., test-time scaling, complex reasoning methods) could enhance the safety awareness of these models.\n\n**2. Complete Omission of Vision-Language-Action (VLA) Models:** \n\nThis is perhaps the most significant flaw. The field of embodied AI has largely moved towards Vision-Language-Action (VLA) models, which are specifically pre-trained or fine-tuned on robotic and physical interaction data. The paper's central claim is that \"embodied LLM agents\" have poor safety awareness. However, it fails to evaluate any of the models actually designed for this purpose (e.g., RT-2, PaLM-E, or other relevant VLAs). It is entirely possible that VLAs, having been fine-tuned on physical-world data, possess a much stronger \"common sense\" for physical safety than general-purpose text LLMs. The paper's conclusions are thus an overgeneralization based on the wrong category of models.\n\n**3. Lack of Real-World Validation (sim2real):** \n\nFor a benchmark focused on physical safety, a simulation-only evaluation is insufficient. The sim2real gap is a well-known and critical challenge in robotics. Hazards, object interactions, and perceptual failures manifest very differently in the real world than in a clean simulator like AI2-THOR. A convincing paper on this topic must, at a minimum, include a discussion of these limitations or, ideally, provide some preliminary sim2real experiments to validate that the findings in simulation translate to real-world physical safety risks.\n\n**4.Weak and Poorly Designed Defense Strategy Evaluation:** \n\nThe paper's attempt to explore defenses (Section 5.6) is superficial and misses the most obvious baseline.\n\n- The authors explicitly state they \"do not add any explicit/implicit hint or prompts about safety in any baselines.\" While this is acceptable for the initial evaluation, it is an unacceptable omission in the defense section.\n\n- The most basic and widely-used defense for LLMs is safety prompting (i.e., adding hints in the system prompt). The authors illogically ignore this baseline and instead test two other methods (CoT filtering and module composition). By failing to compare against the most obvious defense, the conclusions drawn from this section are weak and uninformative.\n\n**5.Limited Scenario Diversity:** \n\n\nThe dataset's scope is narrow, focusing almost exclusively on housekeeping or domestic (kitchen, living room) scenarios. Physical safety is a broad concern that extends to many other environments, such as industrial settings, outdoor navigation, or medical assistance. The benchmark's generalizability is questionable as it omits these other critical domains.\n\n\n\n## Minor Weaknesses:\n\n1. Clarity of Metrics: The paper introduces several metric abbreviations in its tables (e.g., Table 2) such as \"Rej\", \"RR(goal)\", \"RR(LLM)\", \"ER\", and \"C-Safe\". These abbreviations are not clearly and explicitly defined in the main text, forcing the reviewer to piece together their meanings from context, which hinders readability.\n\n2. Undefined Variables: The variable $\\alpha$, introduced on line 371 to represent \"intrinsic safety awareness,\" is used without a clear, self-contained mathematical or conceptual definition, making the subsequent parameter decomposition difficult to follow.", "questions": "see above.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "all_content": {"summary": "The paper presents SafeAgentBench, a benchmark designed to evaluate the safety awareness of task-planning embodied LLM agents within an interactive simulation environment (SafeAgentEnv, based on AI2-THOR). The benchmark's key contributions are: (1) A dataset of 750 tasks (450 hazardous, 300 safe) covering 10 hazard types and 3 task structures (detailed, abstract, long-horizon), which notably includes explicitly hazardous instructions often overlooked by prior work. (2) A dual evaluation system using both a rule-based Execution Evaluator and an LLM-based Semantic Evaluator. The authors evaluate 9 agent baselines, primarily driven by GPT-4, and find that their proactive safety awareness is \"weak.\" For instance, the highest rejection rate for detailed hazardous tasks is only 10%. The paper also shows that low risk rates are often due to poor planning capability ($\\theta$) rather than intrinsic safety awareness ($\\alpha$), and that simple defenses are ineffective.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper addresses the critical and timely problem of embodied agent safety. As agents become more capable, understanding their failure modes and safety awareness is of paramount importance to the field.", "weaknesses": "## Major Weaknesses:\n\n**1.Critically Outdated and Irrelevant Model Selection:** \n\nThe paper's experimental setup is fundamentally flawed by its exclusive reliance on pure text-based LLMs.\n\n- **Lack of Multimodality:** Embodied agents, by definition, must perceive and interact with their environment. This requires processing multimodal inputs, primarily visual information (e.g., images, depth maps). The paper’s core evaluation, however, uses GPT-4 (a text-only model) as the central planner, with other text-only models (Llama3, Qwen2, DeepSeekV2.5) as alternatives. This setup is unrealistic. It does not evaluate an embodied agent but rather a disembodied text model's abstract knowledge of safety, which is not the paper's claimed contribution.\n\n- **Outdated Models:** The paper is submitted for 2026 but relies on models from 2023 (GPT-4). The field of LLMs is advancing rapidly. A contemporary evaluation would be expected to include SOTA multimodal models (e.g., GPT-4o, Gemini 1.5/2.5 Pro, Claude 3 family) which are now standard. The paper also fails to investigate whether advanced reasoning techniques (e.g., test-time scaling, complex reasoning methods) could enhance the safety awareness of these models.\n\n**2. Complete Omission of Vision-Language-Action (VLA) Models:** \n\nThis is perhaps the most significant flaw. The field of embodied AI has largely moved towards Vision-Language-Action (VLA) models, which are specifically pre-trained or fine-tuned on robotic and physical interaction data. The paper's central claim is that \"embodied LLM agents\" have poor safety awareness. However, it fails to evaluate any of the models actually designed for this purpose (e.g., RT-2, PaLM-E, or other relevant VLAs). It is entirely possible that VLAs, having been fine-tuned on physical-world data, possess a much stronger \"common sense\" for physical safety than general-purpose text LLMs. The paper's conclusions are thus an overgeneralization based on the wrong category of models.\n\n**3. Lack of Real-World Validation (sim2real):** \n\nFor a benchmark focused on physical safety, a simulation-only evaluation is insufficient. The sim2real gap is a well-known and critical challenge in robotics. Hazards, object interactions, and perceptual failures manifest very differently in the real world than in a clean simulator like AI2-THOR. A convincing paper on this topic must, at a minimum, include a discussion of these limitations or, ideally, provide some preliminary sim2real experiments to validate that the findings in simulation translate to real-world physical safety risks.\n\n**4.Weak and Poorly Designed Defense Strategy Evaluation:** \n\nThe paper's attempt to explore defenses (Section 5.6) is superficial and misses the most obvious baseline.\n\n- The authors explicitly state they \"do not add any explicit/implicit hint or prompts about safety in any baselines.\" While this is acceptable for the initial evaluation, it is an unacceptable omission in the defense section.\n\n- The most basic and widely-used defense for LLMs is safety prompting (i.e., adding hints in the system prompt). The authors illogically ignore this baseline and instead test two other methods (CoT filtering and module composition). By failing to compare against the most obvious defense, the conclusions drawn from this section are weak and uninformative.\n\n**5.Limited Scenario Diversity:** \n\n\nThe dataset's scope is narrow, focusing almost exclusively on housekeeping or domestic (kitchen, living room) scenarios. Physical safety is a broad concern that extends to many other environments, such as industrial settings, outdoor navigation, or medical assistance. The benchmark's generalizability is questionable as it omits these other critical domains.\n\n\n\n## Minor Weaknesses:\n\n1. Clarity of Metrics: The paper introduces several metric abbreviations in its tables (e.g., Table 2) such as \"Rej\", \"RR(goal)\", \"RR(LLM)\", \"ER\", and \"C-Safe\". These abbreviations are not clearly and explicitly defined in the main text, forcing the reviewer to piece together their meanings from context, which hinders readability.\n\n2. Undefined Variables: The variable $\\alpha$, introduced on line 371 to represent \"intrinsic safety awareness,\" is used without a clear, self-contained mathematical or conceptual definition, making the subsequent parameter decomposition difficult to follow.", "questions": "see above.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761721272713}, {"id": "3oGwKRPPuH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16964/Reviewer_Nig1"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper focuses on evaluating LLM/VLM agents for high-level task reasoning. The paper makes several key contributions:\n\n- A diverse dataset of embodied agent tasks.\n- An environment for robotics evaluation.\n- Evaluation on state-of-the-art LLM-based planning and reasoning methods.", "review_text": "This paper focuses on evaluating LLM/VLM agents for high-level task reasoning. The paper makes several key contributions:\n\n- A diverse dataset of embodied agent tasks.\n- An environment for robotics evaluation.\n- Evaluation on state-of-the-art LLM-based planning and reasoning methods.", "strengths": "- The paper aims to provide a benchmark and evaluation for robot practitioners to build upon by ensuring VLM-based planners plan safely before executing low-level control policies.\n- Assuming you are developing methods assuming an LLM/VLM agent that generates high-level task plans, this work can be useful.\n- The plots are well formatted and are clear to read.", "weaknesses": "- Narrow Focus on LLM Agents for Embodied Task Reasoning\n    - The paper is focused on the use of LLM agents for embodied task reasoning. While this could be useful in some contexts where developers may choose to use an LLM for reasoning, it’s unclear why one would be motivated to do so currently.\n    - We’ve seen lots of developments around VLM agents. Why not expand to VLM agents as well? It seems like some methods rely on VLMs, in which case the authors should make this more explicit.\n    - Beyond the modalities, the authors need to better motivate how this is a useful problem for roboticists to generate plans from LLMs to control embodied agents.\n- Choice of Simulator\n    - The reliance on the AI2-THOR simulation environment weakens the extensibility of this benchmark and how useful it is beyond just one simulation environment.\n    - This simulator also has a fixed set of 17 high-level actions, which is often a very limited set that undercounts the diverse actions a robot may execute.\n- Paper organization and Writing:\n    - Early on in the paper, the authors illustrate the central results in Table 1, but there lacks sufficient context to understand the table.\n    - Relatively minor point: I don’t like the analogy to the way humans think at the beginning of Section 4.1. This work isn’t motivated by human-type of reasoning, so this comment feels to be a distraction: ‘This controller acts as the agent’s \"cerebellum,\" executing the high-level plans formulated by the planner \"brain.\"’\n- Evaluation:\n    - It feels very important to clarify that some evaluated methods are open-loop planners and others are closed-loop planners. This makes a difference in the performance, as closed-loop planners will incorporate environmental changes.\n    - Table 6 should be referred to in the main text, as it makes it clear what the “abstraction levels” of tasks refer to more precisely.\n- Related work:\n    - The paper is missing a key related work (in my opinion) in the space of jailbreaking LLM-controlled robots, namely this paper: https://arxiv.org/pdf/2410.13691", "questions": "- For the Introduction, how is SafeAgentEnv unique in comparison with other methods?\n- Is there a reason why the simulation methods are built around AI2-THOR? Isn’t 17 high-level actions very limiting?\n- How does this benchmark environment compare with other benchmark environments commonly used in robotics, such as LIBERO or Bridge?\n- Can the authors clarify LLM agents vs VLM agents? See the “Weaknesses” section here.\n- Can the authors clarify which planners are open-loop and which ones are closed-loop?\n- What about other forms of failure, e.g. collision avoidance in robot planning? Are there tasks developed to evaluate collision avoidance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on evaluating LLM/VLM agents for high-level task reasoning. The paper makes several key contributions:\n\n- A diverse dataset of embodied agent tasks.\n- An environment for robotics evaluation.\n- Evaluation on state-of-the-art LLM-based planning and reasoning methods.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper aims to provide a benchmark and evaluation for robot practitioners to build upon by ensuring VLM-based planners plan safely before executing low-level control policies.\n- Assuming you are developing methods assuming an LLM/VLM agent that generates high-level task plans, this work can be useful.\n- The plots are well formatted and are clear to read.", "weaknesses": "- Narrow Focus on LLM Agents for Embodied Task Reasoning\n    - The paper is focused on the use of LLM agents for embodied task reasoning. While this could be useful in some contexts where developers may choose to use an LLM for reasoning, it’s unclear why one would be motivated to do so currently.\n    - We’ve seen lots of developments around VLM agents. Why not expand to VLM agents as well? It seems like some methods rely on VLMs, in which case the authors should make this more explicit.\n    - Beyond the modalities, the authors need to better motivate how this is a useful problem for roboticists to generate plans from LLMs to control embodied agents.\n- Choice of Simulator\n    - The reliance on the AI2-THOR simulation environment weakens the extensibility of this benchmark and how useful it is beyond just one simulation environment.\n    - This simulator also has a fixed set of 17 high-level actions, which is often a very limited set that undercounts the diverse actions a robot may execute.\n- Paper organization and Writing:\n    - Early on in the paper, the authors illustrate the central results in Table 1, but there lacks sufficient context to understand the table.\n    - Relatively minor point: I don’t like the analogy to the way humans think at the beginning of Section 4.1. This work isn’t motivated by human-type of reasoning, so this comment feels to be a distraction: ‘This controller acts as the agent’s \"cerebellum,\" executing the high-level plans formulated by the planner \"brain.\"’\n- Evaluation:\n    - It feels very important to clarify that some evaluated methods are open-loop planners and others are closed-loop planners. This makes a difference in the performance, as closed-loop planners will incorporate environmental changes.\n    - Table 6 should be referred to in the main text, as it makes it clear what the “abstraction levels” of tasks refer to more precisely.\n- Related work:\n    - The paper is missing a key related work (in my opinion) in the space of jailbreaking LLM-controlled robots, namely this paper: https://arxiv.org/pdf/2410.13691", "questions": "- For the Introduction, how is SafeAgentEnv unique in comparison with other methods?\n- Is there a reason why the simulation methods are built around AI2-THOR? Isn’t 17 high-level actions very limiting?\n- How does this benchmark environment compare with other benchmark environments commonly used in robotics, such as LIBERO or Bridge?\n- Can the authors clarify LLM agents vs VLM agents? See the “Weaknesses” section here.\n- Can the authors clarify which planners are open-loop and which ones are closed-loop?\n- What about other forms of failure, e.g. collision avoidance in robot planning? Are there tasks developed to evaluate collision avoidance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761720885887}], "openreview_url": "https://openreview.net/forum?id=BFb4ACHayj", "arxiv_id": "2412.13178", "paper_pdf": "papers/BFb4ACHayj.pdf", "paper_pdf_sha256": "b6168f71cc6cfeaa70c4e10d1527d38f31f96f484b3b5112db470c4737339807", "paper_pdf_bytes": 5445029, "paper_pdf_source": "openreview", "code_url": "https://github.com/shengyin1224/SafeAgentBench", "code_repository": "shengyin1224/SafeAgentBench", "code_commit": "38ca3ab27eb8a5f5034a50bdcc5cbab23ce8f089", "code_archive": "repos/BFb4ACHayj.zip", "code_archive_sha256": "18955464ca05f1500164776fab83da05dceb5b7c2e1a116fb27c1601b2b10553", "code_archive_bytes": 1950191, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 1901, "github_languages": {"Python": 146079}, "github_archived": false, "github_pushed_at": "2025-02-25T15:07:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/safeagentbench-a-benchmark-for-safe-task"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mkjKqeBXkt", "year": 2024, "status": "rejected", "title": "KITS: Inductive Spatio-Temporal Kriging with Increment Training Strategy", "authors": ["Qianxiong Xu", "Cheng Long", "Ziyue Li", "Sijie Ruan", "Rui Zhao", "zhishuai Li"], "authorids": ["~Qianxiong_Xu1", "~Cheng_Long1", "~Ziyue_Li2", "~Sijie_Ruan1", "~Rui_Zhao6", "~zhishuai_Li1"], "authors_source": "OpenReview API", "abstract": "Sensors are commonly deployed to perceive the environment. However, due to the high cost, sensors are usually sparsely deployed. Kriging is the tailored task to infer the unobserved nodes (without sensors) using the observed source nodes (with sensors). The essence of kriging task is transferability. Recently, several inductive spatio-temporal kriging methods have been proposed based on graph neural networks, being trained based on a graph built on top of observed nodes via pretext tasks such as masking nodes out and reconstructing them. However, the graph in training is inevitably much sparser than the graph in inference that includes all the observed and unobserved nodes. The learned pattern cannot be well generalized for inference, denoted as graph gap. To address this issue, we first present a novel Increment training strategy: instead of masking nodes (and reconstructing them), we add virtual nodes into the training graph so as to mitigate the graph gap issue naturally. Nevertheless, the empty-shell virtual nodes without labels could have bad-learned features and lack supervision signals. To solve these issues, we pair each virtual node with its most similar observed node and fuse their features together; to enhance the supervision signal, we construct reliable pseudo labels for virtual nodes. As a result, the learned pattern of virtual nodes could be safely transferred to real unobserved nodes for reliable kriging. We name our new Kriging model with Increment Training Strategy as KITS. Extensive experiments demonstrate that KITS consistently outperforms existing kriging methods by large margins, e.g., the improvement over MAE score could be as high as 18.33\\%.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "BPPRXip2Hq", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission129/Reviewer_Ge1P"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper addresses the challenge of inferring unobserved nodes using observed source nodes. While several inductive spatio-temporal kriging methods utilizing graph neural networks have been previously proposed, they often do not account for the discrepancies between the sparsity of training data and inference data (as the 'graph gap').\n\nThe author introduces an approach that adds virtual nodes into the training graph. This is done by 1) improve bad-learned features by finding similar nodes/feature fuse (Reference-based Feature Fusion module) 2) improve supervision signals by construct reliable pseudo labels for virtual nodes (Node-aware Cycle Regulation). Authors also show extensive experiments on 8 datasets with ablation study to support the method.", "review_text": "This paper addresses the challenge of inferring unobserved nodes using observed source nodes. While several inductive spatio-temporal kriging methods utilizing graph neural networks have been previously proposed, they often do not account for the discrepancies between the sparsity of training data and inference data (as the 'graph gap').\n\nThe author introduces an approach that adds virtual nodes into the training graph. This is done by 1) improve bad-learned features by finding similar nodes/feature fuse (Reference-based Feature Fusion module) 2) improve supervision signals by construct reliable pseudo labels for virtual nodes (Node-aware Cycle Regulation). Authors also show extensive experiments on 8 datasets with ablation study to support the method.", "strengths": "Admittedly, I am not a domain expert for inductive spatio-temporal kriging methods based on GNN, I do find this paper is:\n* Well-written, and enjoyable to read.\n* This looks like a novel method to address the sparsity gap between training graph and inference graph.", "weaknesses": "* My main concern is that the introduction of virtual nodes does not add additional information to the dataset. Consequently, my intuition is that this strategy is more effective when the dataset size is small. In the case of the benchmark datasets, which contain a limited number of spatial data points (ranging from a maximum of 883 to a minimum of 80), the approach seems advantageous. However, in scenarios where there is a larger denser spatial dataset, it's worth considering whether the issue of sparsity remains as significant. Would the gap in sparsity still present a challenge in extensive populated spatial data environments? But again, I am not a domain expert for inductive spatio-temporal kriging methods based on GNN, maybe this is indeed the usual dataset size for this line of work. I am happy to change my score if this gets justified.\n\n* This paper employs a thresholded Gaussian kernel to construct the adjacency matrix A among nodes, setting the values to zero when distances exceed a certain threshold. Have the authors explored the possibility of using non-stationary kernels or neural network-based kernels as alternatives?", "questions": "* When aggregating spatio-temporal features from neighboring nodes, is it worth exploring some attention mechanism to gain more global information (especially with so many timesteps in the datasets)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the challenge of inferring unobserved nodes using observed source nodes. While several inductive spatio-temporal kriging methods utilizing graph neural networks have been previously proposed, they often do not account for the discrepancies between the sparsity of training data and inference data (as the 'graph gap').\n\nThe author introduces an approach that adds virtual nodes into the training graph. This is done by 1) improve bad-learned features by finding similar nodes/feature fuse (Reference-based Feature Fusion module) 2) improve supervision signals by construct reliable pseudo labels for virtual nodes (Node-aware Cycle Regulation). Authors also show extensive experiments on 8 datasets with ablation study to support the method.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Admittedly, I am not a domain expert for inductive spatio-temporal kriging methods based on GNN, I do find this paper is:\n* Well-written, and enjoyable to read.\n* This looks like a novel method to address the sparsity gap between training graph and inference graph.", "weaknesses": "* My main concern is that the introduction of virtual nodes does not add additional information to the dataset. Consequently, my intuition is that this strategy is more effective when the dataset size is small. In the case of the benchmark datasets, which contain a limited number of spatial data points (ranging from a maximum of 883 to a minimum of 80), the approach seems advantageous. However, in scenarios where there is a larger denser spatial dataset, it's worth considering whether the issue of sparsity remains as significant. Would the gap in sparsity still present a challenge in extensive populated spatial data environments? But again, I am not a domain expert for inductive spatio-temporal kriging methods based on GNN, maybe this is indeed the usual dataset size for this line of work. I am happy to change my score if this gets justified.\n\n* This paper employs a thresholded Gaussian kernel to construct the adjacency matrix A among nodes, setting the values to zero when distances exceed a certain threshold. Have the authors explored the possibility of using non-stationary kernels or neural network-based kernels as alternatives?", "questions": "* When aggregating spatio-temporal features from neighboring nodes, is it worth exploring some attention mechanism to gain more global information (especially with so many timesteps in the datasets)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1700457744963}, {"id": "M33J1YdPPz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission129/Reviewer_env1"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper addresses a spatio-temporal krigging problem. This is modeled as a node prediction task using graph convolution neural network (GCN) whose convolution operator is learned to compute the feature embeddings for each node, taking into account its topological connection to others (i.e., a training graph). \n\nThe learned weights of the convolution operator can be applied to any topological graphs that contain the training graph, thus enabling inductive inference on new node: a forward pass of the (learned) convolution over an expanded graph that includes incorporate new connections to unobserved nodes will predict the features for those nodes, which can be subsequently input to a feed-forward net for node prediction.\n\nPrevious approaches addressing this problem often trains the GCN on the observed (training) part of the entire graph, which is later used to perform inductive inference on the unobserved (test) part of the graph. Their performance might therefore suffered from the disparity or gap between the train and test graphs. This is what this paper aims to address.\n\nThis is based on three main ideas:\n\nGraph Augmentation (Increment Strategy): Observed nodes are sampled. For each observed node, a virtual node and a connection to the observed node is created. The virtual node's connection to the neighborhood of the observed node is also randomly generated. Multiple such augmented graphs are generated to train the GCN so its learned convolution is robust to potential variation in the structure of the (unseen) test graph. Original features for the virtual nodes are set to be zero. Also, features of the same node reported or computed within a window of +/- m steps are also concatenated before passing through the GCN\n\nFeature Fusion between Virtual & Observed Nodes: Generated features of virtual nodes and observed nodes from the above steps are paired based on a similarity notion. Paired features are concatenated and passed through a (learnable) neural net, which returns the fused features for the virtual node. \n\nPseudo-Label Generation: Pseudo-labels for virtual nodes are first generated using the learned GCN-based node prediction model. The model is then re-trained using only the pseudo-labels to predict labels of the observed nodes. The losses in two training phases can be combined to be optimized together.", "review_text": "This paper addresses a spatio-temporal krigging problem. This is modeled as a node prediction task using graph convolution neural network (GCN) whose convolution operator is learned to compute the feature embeddings for each node, taking into account its topological connection to others (i.e., a training graph). \n\nThe learned weights of the convolution operator can be applied to any topological graphs that contain the training graph, thus enabling inductive inference on new node: a forward pass of the (learned) convolution over an expanded graph that includes incorporate new connections to unobserved nodes will predict the features for those nodes, which can be subsequently input to a feed-forward net for node prediction.\n\nPrevious approaches addressing this problem often trains the GCN on the observed (training) part of the entire graph, which is later used to perform inductive inference on the unobserved (test) part of the graph. Their performance might therefore suffered from the disparity or gap between the train and test graphs. This is what this paper aims to address.\n\nThis is based on three main ideas:\n\nGraph Augmentation (Increment Strategy): Observed nodes are sampled. For each observed node, a virtual node and a connection to the observed node is created. The virtual node's connection to the neighborhood of the observed node is also randomly generated. Multiple such augmented graphs are generated to train the GCN so its learned convolution is robust to potential variation in the structure of the (unseen) test graph. Original features for the virtual nodes are set to be zero. Also, features of the same node reported or computed within a window of +/- m steps are also concatenated before passing through the GCN\n\nFeature Fusion between Virtual & Observed Nodes: Generated features of virtual nodes and observed nodes from the above steps are paired based on a similarity notion. Paired features are concatenated and passed through a (learnable) neural net, which returns the fused features for the virtual node. \n\nPseudo-Label Generation: Pseudo-labels for virtual nodes are first generated using the learned GCN-based node prediction model. The model is then re-trained using only the pseudo-labels to predict labels of the observed nodes. The losses in two training phases can be combined to be optimized together.", "strengths": "The presented ideas are refreshingly interesting & novel to me. \n\nThe problem being addressed is also practically significant. The gap between the train and test graphs in the inductive setting of node prediction is always a fundamental issue, which needs an in-depth treatment.\n\nThe empirical studies are sufficiently extensive, showing good results across a variety of datasets. \n\nComprehensive ablation studies showing the effectiveness of each idea is also included.", "weaknesses": "Overall, I like this paper. It motivates well a fundamental issue of inductive inference with GCN.\n\nBut I do have a few concerns or questions regarding both the position, presentation and empirical evaluation of this paper that I want to discuss with the authors (mostly out of curiosity & for constructive feedback)\n\n1. The main position of this paper is grounded in the spatio-temporal setting but the proposed treatment does not seem to have anything specific to the temporal aspect. The aggregation of temporal feature within +/- m steps is kind of a random treatment to me. I do not see a very particular reasoning why doing so makes sense. What if such aggregation erases important small-scale variation patterns in the time-series data? Furthermore, m is a value dependent on the nature of the data so even though the ablation studies conclude that m = 2 gives best result, it is still specific to the few set of experimental data and that should not be a generic guideline to choose m. \n\n2. Although comparison with recent pure graph-based approaches has been well presented, I am still curious to know how well the proposed approach improves over a hybrid approach that integrates GCN with traditional krigging techniques. For example, the series of work on graph convolutional Gaussian processes. There is quite a substantial volume on that & I think the authors have not discussed that in the literature review. Such techniques can be used to address this krigging problem. After all, if I remember correctly, Gaussian processes were referred to as krigging in the past (in geo-statistics). \n\n3. Last, in terms of the presentation, I believe it would be better if the authors spend some space on summarizing GCN which will highlight better the essence of transferability in the inductive setting of krigging. Otherwise, while the current presentation is fine for people who are familiar with graph neural network, it will still leave a lot of gap for generic readers.", "questions": "Based on the above, I have the following questions:\n\n1. Could the authors position this work with the other literature on graph convolutional GPs, which is an integrative approach combining both elements of GCN & a traditional non-graph krigging method (i.e., GPs)?\n\n2. Could the authors elaborate more on why this paper is specifically positioned in the temporal-spatial setting even though the proposed technique does not really have any new innovations for temporal modeling? I might have missed something important here.\n\n3. It will be good to run some comparative experiments with a few representative graph convolutional GPs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses a spatio-temporal krigging problem. This is modeled as a node prediction task using graph convolution neural network (GCN) whose convolution operator is learned to compute the feature embeddings for each node, taking into account its topological connection to others (i.e., a training graph). \n\nThe learned weights of the convolution operator can be applied to any topological graphs that contain the training graph, thus enabling inductive inference on new node: a forward pass of the (learned) convolution over an expanded graph that includes incorporate new connections to unobserved nodes will predict the features for those nodes, which can be subsequently input to a feed-forward net for node prediction.\n\nPrevious approaches addressing this problem often trains the GCN on the observed (training) part of the entire graph, which is later used to perform inductive inference on the unobserved (test) part of the graph. Their performance might therefore suffered from the disparity or gap between the train and test graphs. This is what this paper aims to address.\n\nThis is based on three main ideas:\n\nGraph Augmentation (Increment Strategy): Observed nodes are sampled. For each observed node, a virtual node and a connection to the observed node is created. The virtual node's connection to the neighborhood of the observed node is also randomly generated. Multiple such augmented graphs are generated to train the GCN so its learned convolution is robust to potential variation in the structure of the (unseen) test graph. Original features for the virtual nodes are set to be zero. Also, features of the same node reported or computed within a window of +/- m steps are also concatenated before passing through the GCN\n\nFeature Fusion between Virtual & Observed Nodes: Generated features of virtual nodes and observed nodes from the above steps are paired based on a similarity notion. Paired features are concatenated and passed through a (learnable) neural net, which returns the fused features for the virtual node. \n\nPseudo-Label Generation: Pseudo-labels for virtual nodes are first generated using the learned GCN-based node prediction model. The model is then re-trained using only the pseudo-labels to predict labels of the observed nodes. The losses in two training phases can be combined to be optimized together.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "The presented ideas are refreshingly interesting & novel to me. \n\nThe problem being addressed is also practically significant. The gap between the train and test graphs in the inductive setting of node prediction is always a fundamental issue, which needs an in-depth treatment.\n\nThe empirical studies are sufficiently extensive, showing good results across a variety of datasets. \n\nComprehensive ablation studies showing the effectiveness of each idea is also included.", "weaknesses": "Overall, I like this paper. It motivates well a fundamental issue of inductive inference with GCN.\n\nBut I do have a few concerns or questions regarding both the position, presentation and empirical evaluation of this paper that I want to discuss with the authors (mostly out of curiosity & for constructive feedback)\n\n1. The main position of this paper is grounded in the spatio-temporal setting but the proposed treatment does not seem to have anything specific to the temporal aspect. The aggregation of temporal feature within +/- m steps is kind of a random treatment to me. I do not see a very particular reasoning why doing so makes sense. What if such aggregation erases important small-scale variation patterns in the time-series data? Furthermore, m is a value dependent on the nature of the data so even though the ablation studies conclude that m = 2 gives best result, it is still specific to the few set of experimental data and that should not be a generic guideline to choose m. \n\n2. Although comparison with recent pure graph-based approaches has been well presented, I am still curious to know how well the proposed approach improves over a hybrid approach that integrates GCN with traditional krigging techniques. For example, the series of work on graph convolutional Gaussian processes. There is quite a substantial volume on that & I think the authors have not discussed that in the literature review. Such techniques can be used to address this krigging problem. After all, if I remember correctly, Gaussian processes were referred to as krigging in the past (in geo-statistics). \n\n3. Last, in terms of the presentation, I believe it would be better if the authors spend some space on summarizing GCN which will highlight better the essence of transferability in the inductive setting of krigging. Otherwise, while the current presentation is fine for people who are familiar with graph neural network, it will still leave a lot of gap for generic readers.", "questions": "Based on the above, I have the following questions:\n\n1. Could the authors position this work with the other literature on graph convolutional GPs, which is an integrative approach combining both elements of GCN & a traditional non-graph krigging method (i.e., GPs)?\n\n2. Could the authors elaborate more on why this paper is specifically positioned in the temporal-spatial setting even though the proposed technique does not really have any new innovations for temporal modeling? I might have missed something important here.\n\n3. It will be good to run some comparative experiments with a few representative graph convolutional GPs.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698996505038}, {"id": "v14BkQiZAu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission129/Reviewer_FRYk"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "I've reviewed this paper for a previous conference and have discussed at length its merits and negative aspects with a long back and forth discussion with the authors. Unfortunately, none of my comments (and that of the other reviewers) were apparently taken into account as this submission is basically unchanged w.r.t. the previous iteration. Furthermore, several of the additional results presented during the previous discussion haven't been included in this version.\n\nThe paper introduces KITS, a novel approach for kriging based on graph neural networks. The main contribution of the paper is the introduction of an augmentation strategy based on the idea of adding virtual nodes to the input graph at training time. The augmentation strategy is then paired with a self-supervised training strategy yielding good results on several benchmark datasets. The method is presented as a solution to what the paper defines as the \"graph gap\", i.e., a mismatch between the graph at training and test time. Overall, the paper paper has merits, however, I do have some serious concerns that prevent me from recommending acceptance.", "review_text": "I've reviewed this paper for a previous conference and have discussed at length its merits and negative aspects with a long back and forth discussion with the authors. Unfortunately, none of my comments (and that of the other reviewers) were apparently taken into account as this submission is basically unchanged w.r.t. the previous iteration. Furthermore, several of the additional results presented during the previous discussion haven't been included in this version.\n\nThe paper introduces KITS, a novel approach for kriging based on graph neural networks. The main contribution of the paper is the introduction of an augmentation strategy based on the idea of adding virtual nodes to the input graph at training time. The augmentation strategy is then paired with a self-supervised training strategy yielding good results on several benchmark datasets. The method is presented as a solution to what the paper defines as the \"graph gap\", i.e., a mismatch between the graph at training and test time. Overall, the paper paper has merits, however, I do have some serious concerns that prevent me from recommending acceptance.", "strengths": "* The introduced data augmentation strategy paired with the self-supervised training routine is novel and appealing.\n* Good empirical performance.\n* Very good presentation.", "weaknesses": "* There is a conceptual flaw in the main motivation behind the introduced methodology. While it is straightforward to see why using a drastically different graph at training and inference time is a problem (\"graph gap\" in the paper), I do not understand why every target node should be reconstructed in a single forward pass.\n* Reconstructing a single node at a time would remove the \"graph gap\", this would be the proper way of carrying out the evaluation.\n* After removing nodes for training, graphs can become sparse. However, this issue is only caused by the removal of nodes for evaluation, i.e., it is an issue of the training/evaluation procedure and not an inherent issue of kriging methods. This would not be a problem in any real-world application as you would eventually train the model on the full graph.\n* The considered datasets are quite small, removing many nodes at random might result in disconnected graphs that would explain the poor performance for some of the baselines.\n\nI believe that the paper has good methodological novelty, although is not framed in the proper way as there is no inherent \"graph gap\" issues in spatio-temporal kriging. The paper should either be rewritten targeting a specific operational setting (heavily scaling back the current claims and significance of the work) or presenting the method as a data augmentation strategy on graphs that are not made artificially sparse.", "questions": "--", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "I've reviewed this paper for a previous conference and have discussed at length its merits and negative aspects with a long back and forth discussion with the authors. Unfortunately, none of my comments (and that of the other reviewers) were apparently taken into account as this submission is basically unchanged w.r.t. the previous iteration. Furthermore, several of the additional results presented during the previous discussion haven't been included in this version.\n\nThe paper introduces KITS, a novel approach for kriging based on graph neural networks. The main contribution of the paper is the introduction of an augmentation strategy based on the idea of adding virtual nodes to the input graph at training time. The augmentation strategy is then paired with a self-supervised training strategy yielding good results on several benchmark datasets. The method is presented as a solution to what the paper defines as the \"graph gap\", i.e., a mismatch between the graph at training and test time. Overall, the paper paper has merits, however, I do have some serious concerns that prevent me from recommending acceptance.", "soundness": "1 poor", "presentation": "3 good", "contribution": "3 good", "strengths": "* The introduced data augmentation strategy paired with the self-supervised training routine is novel and appealing.\n* Good empirical performance.\n* Very good presentation.", "weaknesses": "* There is a conceptual flaw in the main motivation behind the introduced methodology. While it is straightforward to see why using a drastically different graph at training and inference time is a problem (\"graph gap\" in the paper), I do not understand why every target node should be reconstructed in a single forward pass.\n* Reconstructing a single node at a time would remove the \"graph gap\", this would be the proper way of carrying out the evaluation.\n* After removing nodes for training, graphs can become sparse. However, this issue is only caused by the removal of nodes for evaluation, i.e., it is an issue of the training/evaluation procedure and not an inherent issue of kriging methods. This would not be a problem in any real-world application as you would eventually train the model on the full graph.\n* The considered datasets are quite small, removing many nodes at random might result in disconnected graphs that would explain the poor performance for some of the baselines.\n\nI believe that the paper has good methodological novelty, although is not framed in the proper way as there is no inherent \"graph gap\" issues in spatio-temporal kriging. The paper should either be rewritten targeting a specific operational setting (heavily scaling back the current claims and significance of the work) or presenting the method as a data augmentation strategy on graphs that are not made artificially sparse.", "questions": "--", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698401509482}], "openreview_url": "https://openreview.net/forum?id=mkjKqeBXkt", "arxiv_id": "2311.02565", "paper_pdf": "papers/mkjKqeBXkt.pdf", "paper_pdf_sha256": "7fa60569c66cb07b2d6ed312576d392cd66927fdeb1807040b4963bf42c9ef65", "paper_pdf_bytes": 2011634, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sam1224/KITS", "code_repository": "Sam1224/KITS", "code_commit": "957e20179758ea97cce17cf40bc08c668287724f", "code_archive": "repos/mkjKqeBXkt.zip", "code_archive_sha256": "9528d3af291ea73b18911ef131b1de91394ac0c340769068304952cfe9f167d4", "code_archive_bytes": 131277, "code_file_count": 36, "code_extensions": {".py": 36}, "github_disk_usage_kb": 106, "github_languages": {"Python": 149801}, "github_archived": false, "github_pushed_at": "2024-12-17T06:26:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/kits-inductive-spatio-temporal-kriging-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Su84ELBdm5U", "year": 2023, "status": "rejected", "title": "How does overparametrization affect performance on minority groups?", "authors": ["Subha Maity", "Saptarshi Roy", "Songkai Xue", "Mikhail Yurochkin", "Yuekai Sun"], "authorids": ["~Subha_Maity1", "~Saptarshi_Roy1", "~Songkai_Xue1", "~Mikhail_Yurochkin1", "~Yuekai_Sun1"], "authors_source": "OpenReview API", "abstract": "The benefits of overparameterization for the overall performance of modern machine learning (ML) models are well known. However, the effect of overparameterization at a more granular level of data subgroups is less understood. Recent empirical studies demonstrate encouraging results: (i) when groups are not known, overparameterized models trained with empirical risk minimization (ERM) perform better on minority groups; (ii) when groups are known, ERM on data subsampled to equalize group sizes yields state-of-the-art worst-group-accuracy in the overparameterized regime. In this paper, we complement these empirical studies with a theoretical investigation of the risk of overparameterized random feature models on minority groups. In a setting in which the regression functions for the majority and minority groups are different, we show that overparameterization always improves minority group performance.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "GDFkWcakdS", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3005/Reviewer_Fo6C"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problem of how overparametrization affects the ERM's performance on minority groups theoretically. To be more specific, this paper theoretically shows that under the overparametrazation condition and ERM algorithm, the minority groups will tend to have worse performance on the regression tasks, and also shows that subsampling majority groups can improve the performance of the minority groups. The theoretical results are consistent with the empirical studies observed in previous works. ", "review_text": "This paper has three key limitations stated in section \"strength and weakness\" and lacks the novelty and impacts to the community. So I tend to vote a reject to this paper. ", "strengths": "Strength: This paper is well written, easy to follow, and has solid results.\n\nI think the paper has the following key weakness\nThe results seem to be very straightforward when compared to under-parametrized cases, but this paper is lacking the comparison of over-parametrized cases and over-parametrized cases, including the clarifications/justifications whether and why the conclusions made for under-parametrized cases can or cannot apply the over-parametrized cases. If the results for over-parametrized cases do not differ from the under-parametrized cases, the novelty and necessity of this paper may be problematic. \n\nAlso, while this paper is discussing the over-parametrized ML, its analysis and/or conclusions do not cover how the minority groups' performance change with the number of parameters from under-parametrized cases to over-parametrized cases. \n\nFinally, the two key conclusions drawn from the paper (minority groups perform worse and sub-sampling can help) are pretty intuitive and within the expectation of most people. This paper does not show great novelties of the methods, so the contribution of this paper to the community does not seem high.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the problem of how overparametrization affects the ERM's performance on minority groups theoretically. To be more specific, this paper theoretically shows that under the overparametrazation condition and ERM algorithm, the minority groups will tend to have worse performance on the regression tasks, and also shows that subsampling majority groups can improve the performance of the minority groups. The theoretical results are consistent with the empirical studies observed in previous works. ", "strength_and_weaknesses": "Strength: This paper is well written, easy to follow, and has solid results.\n\nI think the paper has the following key weakness\nThe results seem to be very straightforward when compared to under-parametrized cases, but this paper is lacking the comparison of over-parametrized cases and over-parametrized cases, including the clarifications/justifications whether and why the conclusions made for under-parametrized cases can or cannot apply the over-parametrized cases. If the results for over-parametrized cases do not differ from the under-parametrized cases, the novelty and necessity of this paper may be problematic. \n\nAlso, while this paper is discussing the over-parametrized ML, its analysis and/or conclusions do not cover how the minority groups' performance change with the number of parameters from under-parametrized cases to over-parametrized cases. \n\nFinally, the two key conclusions drawn from the paper (minority groups perform worse and sub-sampling can help) are pretty intuitive and within the expectation of most people. This paper does not show great novelties of the methods, so the contribution of this paper to the community does not seem high.", "clarity,_quality,_novelty_and_reproducibility": "The clarity and the quality of the paper are good. This paper does not have problem of reproducibility. However, this paper lacks enough novelty to meet the bar of ICLR.", "summary_of_the_review": "This paper has three key limitations stated in section \"strength and weakness\" and lacks the novelty and impacts to the community. So I tend to vote a reject to this paper. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1668840026569}, {"id": "k5J1NpAtXT", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3005/Reviewer_1vd9"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper studies the effect of overparameterization under the existence of minority groups. The paper provides theoretical characterizations based on existing work, and ran experiments on the fixed point equations to show that the overparamterization does reflect recent empirical findings in this high dimensional setup.", "review_text": "See above.", "strengths": "The paper is easy to follow and well-written. The numerical experiments are abundant.\n\nHowever, I find the paper lacks theoretical novelty as the main results are direct applications of previous work. In the abstract the authors claim that they show overparameterization always improves minority group performance, but I found no such theorem is stated instead of numerical solutions to the fixed-point equations characterizing the asymptotic behavior. I would expect a theorem that says the effective bias and variance on the minority group is monotone if it is claimed that this is shown.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies the effect of overparameterization under the existence of minority groups. The paper provides theoretical characterizations based on existing work, and ran experiments on the fixed point equations to show that the overparamterization does reflect recent empirical findings in this high dimensional setup.", "strength_and_weaknesses": "The paper is easy to follow and well-written. The numerical experiments are abundant.\n\nHowever, I find the paper lacks theoretical novelty as the main results are direct applications of previous work. In the abstract the authors claim that they show overparameterization always improves minority group performance, but I found no such theorem is stated instead of numerical solutions to the fixed-point equations characterizing the asymptotic behavior. I would expect a theorem that says the effective bias and variance on the minority group is monotone if it is claimed that this is shown.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear but lacks theoretical novelty.", "summary_of_the_review": "See above.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667893627530}, {"id": "wYSwffom5nT", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3005/Reviewer_y1cp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the effect of overparametrization on minority groups from a theoretical perspective. Building upon existing works, the authors study the asymptotic performance of overparametrized methods in minority groups in a stylized model and show that their results are consistent with empirical observations. ", "review_text": "The paper is well-written and the message is clear; more justifications are needed in terms of the choice of model.", "strengths": "- **Strength**:\nThe paper is well-written and the simulations are well-designed. The message is clearly conveyed.\n\n- **Major concern**:\nOn page 3, it is remarked that the paper adopts the random feature model instead of the linear model because results in the latter model do not coincide with the empirical findings. I am confused in that (1) what is the model in general used in empirical works (it seems to me should be the latter)? (2) if both models are approximations of the models used in practice, then why are there inconsistencies? Choosing the model that exhibits the desired performance feels a bit like cherry-picking to me---but I could be wrong!\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies the effect of overparametrization on minority groups from a theoretical perspective. Building upon existing works, the authors study the asymptotic performance of overparametrized methods in minority groups in a stylized model and show that their results are consistent with empirical observations. ", "strength_and_weaknesses": "- **Strength**:\nThe paper is well-written and the simulations are well-designed. The message is clearly conveyed.\n\n- **Major concern**:\nOn page 3, it is remarked that the paper adopts the random feature model instead of the linear model because results in the latter model do not coincide with the empirical findings. I am confused in that (1) what is the model in general used in empirical works (it seems to me should be the latter)? (2) if both models are approximations of the models used in practice, then why are there inconsistencies? Choosing the model that exhibits the desired performance feels a bit like cherry-picking to me---but I could be wrong!\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written, with a clear exposition of messages and well-designed experiments. I think the choice of models should be better justified.", "summary_of_the_review": "The paper is well-written and the message is clear; more justifications are needed in terms of the choice of model.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667139244325}, {"id": "1sV3ZbFtbT", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3005/Reviewer_BTHL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "Recently, it has been empirically observed that overparameterization helps to improve performance on both the majority and minority subgroups of the data. Few works have also proposed methods to improve the performance on minority subgroups like group distributionally robust optimization and data subsampling. This paper confirms these findings theoretically by studying this using a random features linear regression model which takes into account the difference between the true predictors of the subgroups, the ratio of the data and the signal to noise ratio.", "review_text": "The main concern I have with this paper is the concept shift model that this paper assumes for the subpopulation setting and lack of citations and comparison to previous theoretical works. ", "strengths": "The problem that this work studies of how overparameterization affects subpopulations is important for the fairness and safety of machine learning and has gained widespread interest. \n\nI think that the model that this work studies is different from the practical setting that this work motivates. In particular, this work assumes that the subgroups have same feature distribution and only differ in their true conditional output distribution. I don’t see why that assumption is correct because the different subgroups would have different feature distribution. In fact, they assume that the two groups have different output distribution conditional on the features which is not clear if it is true because we can think of both the populations having the same true predictor. The setting that this work considers there does not exist a single good predictor which is not true in the empirical papers that they cite. This is also what was considered in the original paper [1] which first studied this using a toy model.\n\nThe paper misses important citations and comparisons to previous works which have studied similar problems [2,3,4 ].\n\n[1] An investigation of why overparameterization exacerbates spurious correlations\n[2] Covariate Shift in High-Dimensional Random Feature Regression\n[3] Undersampling is a minimax optimal robustness intervention in nonparametric classification\n[4] Throwing away data improves worst-class error in imbalanced classification", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Recently, it has been empirically observed that overparameterization helps to improve performance on both the majority and minority subgroups of the data. Few works have also proposed methods to improve the performance on minority subgroups like group distributionally robust optimization and data subsampling. This paper confirms these findings theoretically by studying this using a random features linear regression model which takes into account the difference between the true predictors of the subgroups, the ratio of the data and the signal to noise ratio.", "strength_and_weaknesses": "The problem that this work studies of how overparameterization affects subpopulations is important for the fairness and safety of machine learning and has gained widespread interest. \n\nI think that the model that this work studies is different from the practical setting that this work motivates. In particular, this work assumes that the subgroups have same feature distribution and only differ in their true conditional output distribution. I don’t see why that assumption is correct because the different subgroups would have different feature distribution. In fact, they assume that the two groups have different output distribution conditional on the features which is not clear if it is true because we can think of both the populations having the same true predictor. The setting that this work considers there does not exist a single good predictor which is not true in the empirical papers that they cite. This is also what was considered in the original paper [1] which first studied this using a toy model.\n\nThe paper misses important citations and comparisons to previous works which have studied similar problems [2,3,4 ].\n\n[1] An investigation of why overparameterization exacerbates spurious correlations\n[2] Covariate Shift in High-Dimensional Random Feature Regression\n[3] Undersampling is a minimax optimal robustness intervention in nonparametric classification\n[4] Throwing away data improves worst-class error in imbalanced classification", "clarity,_quality,_novelty_and_reproducibility": "The writing of this works needs to be improved. Some of the statements are unclear in the paper. \n\nOn page 3, this work says that the minority group error increases with overparameterization in the linear model setting but later in the appendix, it says that even the average error increases in this setting. \n\nOne page 1, this works says that [1] showed that overparameterization hurts minority group accuracy whereas at multiple points, they claim in the paper that confirm [1]’s findings by showing that overparameterization helps minority group accuracy. Can the authors please clarify this?", "summary_of_the_review": "The main concern I have with this paper is the concept shift model that this paper assumes for the subpopulation setting and lack of citations and comparison to previous theoretical works. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666744128907}, {"id": "M4sGRx6pSDc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3005/Reviewer_2EGb"], "rating": "", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper develops a simple two-group model to study the effects of overparameterization on groups. It provides a theoretical justification for the empirical results in the existing work and shows that the overparameterization improves or does not harm the minority risk of ERM. The majority group subsampling improves minority group performance in the overparameterization regime. ", "review_text": "see above.", "strengths": "It provides a theoretical justification for the effect of overparameterization on minority groups. \nThe paper is easy to follow.\n\nThe application scenario of overparameterization and underparameterization in machine learning is not well explained. \n\nIt mentions that the proposed two-group model has parameters for controlling signal strength, majority group fraction and many terms. I would like to see how the signal strength is used in classification, regression or other machine learning task. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary_of_the_paper": "This paper develops a simple two-group model to study the effects of overparameterization on groups. It provides a theoretical justification for the empirical results in the existing work and shows that the overparameterization improves or does not harm the minority risk of ERM. The majority group subsampling improves minority group performance in the overparameterization regime. ", "strength_and_weaknesses": "It provides a theoretical justification for the effect of overparameterization on minority groups. \nThe paper is easy to follow.\n\nThe application scenario of overparameterization and underparameterization in machine learning is not well explained. \n\nIt mentions that the proposed two-group model has parameters for controlling signal strength, majority group fraction and many terms. I would like to see how the signal strength is used in classification, regression or other machine learning task. \n", "clarity,_quality,_novelty_and_reproducibility": "It provides a theoretical justification for the empirical results on the overparameterazation on minority group", "summary_of_the_review": "see above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666587531565}], "openreview_url": "https://openreview.net/forum?id=Su84ELBdm5U", "arxiv_id": "2206.03515", "paper_pdf": "papers/Su84ELBdm5U.pdf", "paper_pdf_sha256": "d178cad140d851044caf7f1a50149a186a7eb0b634bde42ffd99b53f620518a0", "paper_pdf_bytes": 584861, "paper_pdf_source": "openreview", "code_url": "https://github.com/smaityumich/overparameterization", "code_repository": "smaityumich/overparameterization", "code_commit": "e66a8bcf32982848ca2a354fde02172811ed2aa6", "code_archive": "repos/Su84ELBdm5U.zip", "code_archive_sha256": "9505a35970f636ac1cc3d651ea26f55eb6e58d091a505343be7c8f1b74b9aaa6", "code_archive_bytes": 339545, "code_file_count": 11, "code_extensions": {".py": 10, ".ipynb": 1}, "github_disk_usage_kb": 196, "github_languages": {"Python": 62925, "Jupyter Notebook": 72}, "github_archived": false, "github_pushed_at": "2022-06-07T17:02:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-does-overparametrization-affect"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XhMa8XPHxpw", "year": 2022, "status": "rejected", "title": "Low-Precision Stochastic Gradient Langevin Dynamics", "authors": ["Ruqi Zhang", "Andrew Gordon Wilson", "Christopher De Sa"], "authorids": ["~Ruqi_Zhang1", "~Andrew_Gordon_Wilson1", "~Christopher_De_Sa2"], "authors_source": "OpenReview API", "abstract": "Low-precision optimization is widely used to accelerate large-scale deep learning. Despite providing better uncertainty estimation and generalization, sampling methods remain mostly unexplored in this space. In this paper, we provide the first study of low-precision Stochastic Gradient Langevin Dynamics (SGLD), arguing that it is particularly suited to low-bit arithmetic due to its intrinsic ability to handle system noise. We prove the convergence of low-precision SGLD on strongly log-concave distributions, showing that with full-precision gradient accumulators, SGLD is more robust to quantization error than SGD; however, with low-precision gradient accumulators, SGLD can diverge arbitrarily far from the target distribution with small stepsizes. To remedy this issue, we develop a new quantization function that preserves the correct variance in each update step. We demonstrate that the resulting low-precision SGLD algorithm is comparable to full-precision SGLD and outperforms low-precision SGD on deep learning tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "N3YoCsMCVSQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2310/Reviewer_gLYz"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a low-precision version of SGLD which makes use of low-bit arithmetic. In particular, since low-precision gradient accumulators might lead to divergence of SGLD, the authors propose a new quantization function to preserve the correct variance in each update step. ", "review_text": "I think the idea of applying low-precision optimization to SGLD is a direction worth exploring, but the overall presentation of the paper seems to be vague to me. While SGLD is viewed as an optimization algorithm particularly for nonconvex optimization problem which aims to minimize an objective function $U$ (see e.g., Raginsky et al., 2017), Langevin Monte Carlo (LMC), which takes exact the same algorithmic form as SGLD, is intrinsically a sampling algorithm which aims to sample the stationary distribution $\\exp(-U)$. The authors has been mixing up the tasks of optimization and sampling in the theoretical and experimental results, making it hard to understand what the task the proposed algorithm wants to solve. In the theoretical results of this paper, the authors leverage results in Dalalyan and Karagulyan (2019) to provide a (sampling) bound on the Wasserstein distance between the distribution generated by the low-precision SGLD and the target distribution, but compare this with an optimization bound in Yang et al. (2019), which is the distance between the SGD estimation and the optimum. This seems very weird to me since I think these are not comparable. \n\nThen, in the experiments, I think the authors are applying the proposed low-precision SGLD as a training (i.e., optimization) algorithm instead of a sampling algorithm. The metric used (test error) also does not seem to measure the accuracy of the sampled distribution using low-precision SGLD. Thus, because of this ambiguity between applying SGLD as a sampling algorithm versus an optimization (training) algorithm, the current results in this work are quite questionable to me. \n\nSome mathematical statements and notation in this paper are also unclear. In the assumption, the authors did not mention that the third condition is that the function $U$ needs to have a Lipschitz Hessian. Representation of the stochastic gradient of $U$ by $\\nabla\\tilde{U}$ is also weird since you actually did not change the function $U$ to $\\tilde{U}$, but the gradient to a stochastic gradient. A more commonly used notation should be $\\nabla\\widetilde{U}$. Also $\\mu_0$ is not defined in Theorem 1 (and typed as $\\nu_0$ in Theorem 2 which must be a typo). The derivation in Section 4.3 must be wrong as well, since $\\theta_{k+1}$ is a vector-valued random variable, its variance must be matrix-valued. Thus, the bottom part of page 5 is wrong. Also, using $\\theta_{k+1}$ to represent both the updates with and without quantization have been leading to much confusion during my review. \n\n\n---\nRaginsky, Maxim, Alexander Rakhlin, and Matus Telgarsky. \"Non-convex learning via stochastic gradient Langevin dynamics: a nonasymptotic analysis.\" Conference on Learning Theory. PMLR, 2017.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a low-precision version of SGLD which makes use of low-bit arithmetic. In particular, since low-precision gradient accumulators might lead to divergence of SGLD, the authors propose a new quantization function to preserve the correct variance in each update step. ", "main_review": "I think the idea of applying low-precision optimization to SGLD is a direction worth exploring, but the overall presentation of the paper seems to be vague to me. While SGLD is viewed as an optimization algorithm particularly for nonconvex optimization problem which aims to minimize an objective function $U$ (see e.g., Raginsky et al., 2017), Langevin Monte Carlo (LMC), which takes exact the same algorithmic form as SGLD, is intrinsically a sampling algorithm which aims to sample the stationary distribution $\\exp(-U)$. The authors has been mixing up the tasks of optimization and sampling in the theoretical and experimental results, making it hard to understand what the task the proposed algorithm wants to solve. In the theoretical results of this paper, the authors leverage results in Dalalyan and Karagulyan (2019) to provide a (sampling) bound on the Wasserstein distance between the distribution generated by the low-precision SGLD and the target distribution, but compare this with an optimization bound in Yang et al. (2019), which is the distance between the SGD estimation and the optimum. This seems very weird to me since I think these are not comparable. \n\nThen, in the experiments, I think the authors are applying the proposed low-precision SGLD as a training (i.e., optimization) algorithm instead of a sampling algorithm. The metric used (test error) also does not seem to measure the accuracy of the sampled distribution using low-precision SGLD. Thus, because of this ambiguity between applying SGLD as a sampling algorithm versus an optimization (training) algorithm, the current results in this work are quite questionable to me. \n\nSome mathematical statements and notation in this paper are also unclear. In the assumption, the authors did not mention that the third condition is that the function $U$ needs to have a Lipschitz Hessian. Representation of the stochastic gradient of $U$ by $\\nabla\\tilde{U}$ is also weird since you actually did not change the function $U$ to $\\tilde{U}$, but the gradient to a stochastic gradient. A more commonly used notation should be $\\nabla\\widetilde{U}$. Also $\\mu_0$ is not defined in Theorem 1 (and typed as $\\nu_0$ in Theorem 2 which must be a typo). The derivation in Section 4.3 must be wrong as well, since $\\theta_{k+1}$ is a vector-valued random variable, its variance must be matrix-valued. Thus, the bottom part of page 5 is wrong. Also, using $\\theta_{k+1}$ to represent both the updates with and without quantization have been leading to much confusion during my review. \n\n\n---\nRaginsky, Maxim, Alexander Rakhlin, and Matus Telgarsky. \"Non-convex learning via stochastic gradient Langevin dynamics: a nonasymptotic analysis.\" Conference on Learning Theory. PMLR, 2017.\n", "summary_of_the_review": "The idea of  applying low-precision optimization to SGLD is a direction worth exploring, but the ambiguity between applying SGLD as a sampling algorithm versus an optimization (training) algorithm in this paper has made the current results in this work quite questionable. Some mathematical statements and notation in this paper are also unclear. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636008480952}, {"id": "39GB87-aYj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2310/Reviewer_bvZD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigate posterior sampling of weight-quantized neural networks with SGLD. The paper gives the convergence of SGLD under Wasserstein distance with quantized weights and gradients. It also considers the situation when the accumulator is also in low precision, and proposes a variance correction strategy to deal with the additional quantization noise. ", "review_text": "Posterior sampling of quantized neural networks is an important problem to study. While the optimization of quantized neural networks are relatively well studied, this paper is the first to consider the sampling of such networks. The technical quality of the is nice. It presents reasonable convergence results for the proposed SGLDLP algorithms. The idea of utilizing the quantization noise for sampling and variance correction makes sense.\n\nIn terms of weakness, I think the assumptions in the paper are worth considering:\n1. the gradient quantizer assumes that we first have a full-precision gradient, and then quantize it, maybe for reducing the communication cost for distributed training. The formulation does not support applications such as mixed precision training, where the gradient itself is computed in an approximate way. Is distributed training the case the authors want to address?\n2. While the convergence result presented in Theorem 1 is good, it relies on a non-converging term relates to Delta_w. It is curious to compare with the following post-training quantization algorithm:\na' generate full precision samples {theta_k} with full-precision SGLD\nb' use the quantized version {Q(theta_k)} as the low-precision samples;\nWe can imagine that the posterior distribution given by this algorithm is also close enough to the true posterior, where the distance is controlled by Delta_w. It would be better if we can show the proposed SGLDLP is better than the post-training quantization algorithm.\n\nPost rebuttal\n====\n\nThanks for addressing my concerns. I am tending to keep my score since I think it is an interesting paper with some novelties, but the novelty is limited, as commented by other reviewers. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper investigate posterior sampling of weight-quantized neural networks with SGLD. The paper gives the convergence of SGLD under Wasserstein distance with quantized weights and gradients. It also considers the situation when the accumulator is also in low precision, and proposes a variance correction strategy to deal with the additional quantization noise. ", "main_review": "Posterior sampling of quantized neural networks is an important problem to study. While the optimization of quantized neural networks are relatively well studied, this paper is the first to consider the sampling of such networks. The technical quality of the is nice. It presents reasonable convergence results for the proposed SGLDLP algorithms. The idea of utilizing the quantization noise for sampling and variance correction makes sense.\n\nIn terms of weakness, I think the assumptions in the paper are worth considering:\n1. the gradient quantizer assumes that we first have a full-precision gradient, and then quantize it, maybe for reducing the communication cost for distributed training. The formulation does not support applications such as mixed precision training, where the gradient itself is computed in an approximate way. Is distributed training the case the authors want to address?\n2. While the convergence result presented in Theorem 1 is good, it relies on a non-converging term relates to Delta_w. It is curious to compare with the following post-training quantization algorithm:\na' generate full precision samples {theta_k} with full-precision SGLD\nb' use the quantized version {Q(theta_k)} as the low-precision samples;\nWe can imagine that the posterior distribution given by this algorithm is also close enough to the true posterior, where the distance is controlled by Delta_w. It would be better if we can show the proposed SGLDLP is better than the post-training quantization algorithm.\n\nPost rebuttal\n====\n\nThanks for addressing my concerns. I am tending to keep my score since I think it is an interesting paper with some novelties, but the novelty is limited, as commented by other reviewers. ", "summary_of_the_review": "A novel paper with reasonable techniques. The results might be improved by tightening the assumptions. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635929579574}, {"id": "9ObRIKCBmY0", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2310/Reviewer_yqrt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper discusses using the stochastic gradient Langevin dynamics with low-precision implementation. The authors present convergence results in Wasserstein distance norm for three implementation cases, which distinguish specific low-precision thresholding schemes and different quantization techniques. The authors also report simulation outcome of proposed low-precision SGLD algorithms and show their convergence in actual application.", "review_text": "Strengths:\n- The narrative is fluent and flow of statement is smooth and natural. In general, the writing of the manuscript is well done.\n- The authors have tried their proposed algorithms in multiple classic application scenarios.\n\nWeaknesses:\n- The submission does not appear to be a major contribution to the field, as the main novel point is the analysis of the variance of low-precision gradient estimator.\n\nSpecifics:\n- page 3, section 3.2, paragraph 3, line 3: $l = -2^{W-F-2}$, why the power is $-2$ instead of $-1$ when computing the upper bound of presentable number?\n\n- page 5, section 4.2, the line below the statement of theorem 2: I think theorem 2 does not suggest \"SGLDLP-L diverges from the target distribution\", but instead, it shows the bound becomes loose when stepsize $\\alpha$ decreases. \n\n- page 4 and 5: schemes SGLDLP-F shown in equation (1) and SGLDLP-L shown in equation (2) essentially only differs in whether the noise $\\xi_{k}$ is low-precision thresholded, doesn't it? \n\n- The main theorems are proved with the results from Dalalyan and Karagulyan's work. The author should explicitly cite the specific theorem in [DK19] and explain why the argued variance of low-precision gradient estimator renders the conditions of the cited theorem satisfied and thus the theorem applicable.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper discusses using the stochastic gradient Langevin dynamics with low-precision implementation. The authors present convergence results in Wasserstein distance norm for three implementation cases, which distinguish specific low-precision thresholding schemes and different quantization techniques. The authors also report simulation outcome of proposed low-precision SGLD algorithms and show their convergence in actual application.", "main_review": "Strengths:\n- The narrative is fluent and flow of statement is smooth and natural. In general, the writing of the manuscript is well done.\n- The authors have tried their proposed algorithms in multiple classic application scenarios.\n\nWeaknesses:\n- The submission does not appear to be a major contribution to the field, as the main novel point is the analysis of the variance of low-precision gradient estimator.\n\nSpecifics:\n- page 3, section 3.2, paragraph 3, line 3: $l = -2^{W-F-2}$, why the power is $-2$ instead of $-1$ when computing the upper bound of presentable number?\n\n- page 5, section 4.2, the line below the statement of theorem 2: I think theorem 2 does not suggest \"SGLDLP-L diverges from the target distribution\", but instead, it shows the bound becomes loose when stepsize $\\alpha$ decreases. \n\n- page 4 and 5: schemes SGLDLP-F shown in equation (1) and SGLDLP-L shown in equation (2) essentially only differs in whether the noise $\\xi_{k}$ is low-precision thresholded, doesn't it? \n\n- The main theorems are proved with the results from Dalalyan and Karagulyan's work. The author should explicitly cite the specific theorem in [DK19] and explain why the argued variance of low-precision gradient estimator renders the conditions of the cited theorem satisfied and thus the theorem applicable.", "summary_of_the_review": "In light of the specifics in the main review, the discussion of low-precision SGLD in this submission is not significant enough and there are some places in the argument that are not clear.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635545456072}, {"id": "iYsfnM1vrWH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2310/Reviewer_dB5g"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The author proposed a comprehensive study of low-precision Stochastic Gradient Langevin Dynamics (SGLD). They proposed the convergence analysis of low-precision Stochastic Gradient Langevin Dynamics on strongly log-concave distributions.  Motivated by the diverging fact of SGLDLP-L with small stepsizes, they proposed a variance-corrected quantization function to ensure a bounded bias. Empirical experiments based on large-scale DNN examples are evaluated.", "review_text": "\n\nPros: the authors conducted some bias analysis for low-precision SGLD based on [Dalalyan and Karagulyan 19]'s result.\n\nCons: \n\n1. **Limited novelty in methodology**: the novelty in terms of methodology is limited and is not very interesting. Simply extending quantization results from SGD to SGLD with further indications is not that interesting. \n\n2. **Limited depth in theory**: the theoretical novelty is limited and has limited depth; although I am very familiar with Dalalyan's work, the authors don't even bother to write down Theorem 4 and the relevant background in  [Dalalyan and Karagulyan 19].\n\n3. **Inconsistencies between experiments and theories**: The experiments are mainly running in non-convex settings, which don't match the theoretical claim in strongly convex scenarios. If I were the authors, I would only run MNIST experiments based on logistic regression; unless I have theoretical results that extend to non-convex settings. Because the theory only claims these contributions in strongly-convex settings. \n\n4. **Incomplete experiments**: Running SGLD for reproducing purely optimization results seems to be not interesting and uncertainty estimations would also be appreciated.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The author proposed a comprehensive study of low-precision Stochastic Gradient Langevin Dynamics (SGLD). They proposed the convergence analysis of low-precision Stochastic Gradient Langevin Dynamics on strongly log-concave distributions.  Motivated by the diverging fact of SGLDLP-L with small stepsizes, they proposed a variance-corrected quantization function to ensure a bounded bias. Empirical experiments based on large-scale DNN examples are evaluated.", "main_review": "\n\nPros: the authors conducted some bias analysis for low-precision SGLD based on [Dalalyan and Karagulyan 19]'s result.\n\nCons: \n\n1. **Limited novelty in methodology**: the novelty in terms of methodology is limited and is not very interesting. Simply extending quantization results from SGD to SGLD with further indications is not that interesting. \n\n2. **Limited depth in theory**: the theoretical novelty is limited and has limited depth; although I am very familiar with Dalalyan's work, the authors don't even bother to write down Theorem 4 and the relevant background in  [Dalalyan and Karagulyan 19].\n\n3. **Inconsistencies between experiments and theories**: The experiments are mainly running in non-convex settings, which don't match the theoretical claim in strongly convex scenarios. If I were the authors, I would only run MNIST experiments based on logistic regression; unless I have theoretical results that extend to non-convex settings. Because the theory only claims these contributions in strongly-convex settings. \n\n4. **Incomplete experiments**: Running SGLD for reproducing purely optimization results seems to be not interesting and uncertainty estimations would also be appreciated.", "summary_of_the_review": "The paper proposed some analysis for low-precision SGLD but the novelties or depths in terms of methodology and theory are both limited. In addition, the experiments don't match the theoretical analysis and don't have the results for uncertainty estimations, which leads me to suspect why to bother to use SGLD (I believe SGD with fine-tuning parameters can also achieve the claimed results). Based on these evaluations, I tend to reject this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635125952175}], "openreview_url": "https://openreview.net/forum?id=XhMa8XPHxpw", "arxiv_id": "2206.09909", "paper_pdf": "papers/XhMa8XPHxpw.pdf", "paper_pdf_sha256": "211a4a0d344a2ffb4ac7d8b66493f8a51c3a4e44d80ce91d147906f55baa5608", "paper_pdf_bytes": 392781, "paper_pdf_source": "openreview", "code_url": "https://github.com/ruqizhang/low-precision-sgld", "code_repository": "ruqizhang/low-precision-sgld", "code_commit": "5f9192e10b2437f0de27c72e8b582bf8621dae26", "code_archive": "repos/XhMa8XPHxpw.zip", "code_archive_sha256": "afaf0c089e0cf96f3891dff22fd6dfbf91f3d81d07f53c3c10fb3ba0018ce41a", "code_archive_bytes": 425131, "code_file_count": 14, "code_extensions": {".py": 9, ".sh": 5}, "github_disk_usage_kb": 413, "github_languages": {"Python": 50121, "Shell": 2895}, "github_archived": false, "github_pushed_at": "2022-06-22T03:13:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/low-precision-stochastic-gradient-langevin-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WPO0vDYLXem", "year": 2021, "status": "rejected", "title": "Hyperparameter Transfer Across Developer Adjustments", "authors": ["Danny Stoll", "Jörg K.H. Franke", "Diane Wagner", "Simon Selg", "Frank Hutter"], "authorids": ["~Danny_Stoll1", "~Jörg_K.H._Franke1", "wagnerd@cs.uni-freiburg.de", "selgs@cs.uni-freiburg.de", "~Frank_Hutter1"], "authors_source": "OpenReview API", "abstract": "After developer adjustments to a machine learning (ML) algorithm, how can the results of an old hyperparameter optimization (HPO) automatically be used to speedup a new HPO? This question poses a challenging problem, as developer adjustments can change which hyperparameter settings perform well, or even the hyperparameter search space itself. While many approaches exist that leverage knowledge obtained on previous tasks, so far, knowledge from previous development steps remains entirely untapped. In this work, we remedy this situation and propose a new research framework: hyperparameter transfer across adjustments (HT-AA). To lay a solid foundation for this research framework, we provide four simple HT-AA baseline algorithms and eight benchmarks\nchanging various aspects of ML algorithms, their hyperparameter search spaces, and the neural architectures used. The best baseline, on average and depending on the budgets for the old and new HPO, reaches a given performance 1.2-3.6x faster than a prominent HPO algorithm without transfer. As HPO is a crucial step in ML development but requires extensive computational resources, this speedup would lead to faster development cycles, lower costs, and reduced environmental impacts. To make these benefits available to ML developers off-the-shelf and to facilitate future research on HT-AA, we provide python packages for our baselines and benchmarks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "x4ucCPwLms", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3570/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is motivated by the situation where a machine learning algorithm has development adjustment and we would like to reuse the tuning results of previous hyperparameter optimization. The paper calls it HT-AA problem, which certainly is an interesting problem given software can get updated often. The paper proposes four simple baseline algorithms for the HT-AA problem. \n\nFor the empirical study, a set of eight benchmarks for basic HT-AA problem are presented. The experiment results show the transfer TPE (T2PE) and best-first strategy produce good speed up. To reach given objective values, T2PE can be 1.0–1.7x faster than TPE, and best-first 1.2–2.6x faster comparing with old HPO. \n\nThe pros of the paper include:\n1. Although the topic of the paper is not one of the most popular, some readers might find it interesting and can be benefited from it.\n2. The proposed methods are reasonable and acceptable.\n3. Numerical results show some of the proposed methods can help to speed up reoptimizing hyperparameter.\n\nThe cons include:\n1. Need explain when the only-optimize-new and drop-unimportant methods can be useful. If not useful as the experiments demonstrate, why propose them?\n2. Look like TPE is the method that show speedup for the benchmarks. How reliable the method is? Is the saving justifying use an extra tuning tool?\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A framework for hyperparemeter transfer when ML algorithm changeshm ", "review": "The paper is motivated by the situation where a machine learning algorithm has development adjustment and we would like to reuse the tuning results of previous hyperparameter optimization. The paper calls it HT-AA problem, which certainly is an interesting problem given software can get updated often. The paper proposes four simple baseline algorithms for the HT-AA problem. \n\nFor the empirical study, a set of eight benchmarks for basic HT-AA problem are presented. The experiment results show the transfer TPE (T2PE) and best-first strategy produce good speed up. To reach given objective values, T2PE can be 1.0–1.7x faster than TPE, and best-first 1.2–2.6x faster comparing with old HPO. \n\nThe pros of the paper include:\n1. Although the topic of the paper is not one of the most popular, some readers might find it interesting and can be benefited from it.\n2. The proposed methods are reasonable and acceptable.\n3. Numerical results show some of the proposed methods can help to speed up reoptimizing hyperparameter.\n\nThe cons include:\n1. Need explain when the only-optimize-new and drop-unimportant methods can be useful. If not useful as the experiments demonstrate, why propose them?\n2. Look like TPE is the method that show speedup for the benchmarks. How reliable the method is? Is the saving justifying use an extra tuning tool?\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603914942804}, {"id": "DqhIHj9Qt9", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3570/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a new framework for hyperparameter optimization and transfer across incremental modifications to a given algorithm and its search space, a process called developer adjustments in the paper. The authors then propose a few strategies to transfer knowledge from previous HPO runs and evaluate them on a series of simulated benchmarks. Results show the value added by transferring information from previous runs, as well as the surprising efficiency of simply reusing the best found hyperparameters from the previous run. \n\nStrong points:\n\n- The framework is simple and clearly introduced. \n- Extensive experiments help bring to light the advantage of transferring across adjustments.\n- The paper is very well written. \n\nWeak points:\n\n- Not enough details on benchmarks, more on this below\n- The use of simulated benchmarks with surrogate models introduces noise in the evaluation\n- Comparisons with more baselines would be beneficial. RF and/or GP-based HPO methods are extremely popular and would have been easy to integrate with the best-first baseline. \n\nRecommendation:\n\nThe contributions are simple and incremental, and clearly rooted in machine learning engineering, however I still think they could be beneficial as a whole to the community given the extensive experiments realized. I have some issues with experiments, lack of details and baselines, but those issues are mostly fixable. I'll give the paper a weak accept for now.\n\nExtra comments:\n\nYou do not specify which benchmarks are based on lookup tables and which ones are based on surrogate models. From looking at the search spaces, I would assume that the SVM and XGB benchmarks are modeled via surrogate benchmarks and the FCN and NAS benchmarks are lookup tables, but this should be explicited in the paper (or appendix). Parameters used for the benchmark surrogate model should also be given (if defaults of Eggensperger are used, simply mention this). It is also not clear what underlying datasets are used, this bears some importance and should be mentioned, even if only in the Appendix.\n\nOn surrogate model benchmarks: It can be seen in (Eggensperger et al. 2015., Figure 2) that ordering of methods can shift due to noise in the surrogate model (a random forest?). This is likely going to have a bigger impact when trying to measure the speedup, which is measured when a method reaches a certain threshold of performance. This threshold is likely to be met during the convergence phase of algorithms, and this phase appears noisier  (i.e. looking at how the phases of transition differ between the true benchmark and the RF surrogate benchmark differ in Eggensperger et al. 2015). Have you given this any thought? Have you compared experiments with a few runs on a real benchmark?\n\nThe method you end up recommending only has its detailed performance shown in the appendix. This feels counterintuitive to me. This result should be featured in the paper itself. This is perhaps due to the used of those split violin plots, which force you to display only two methods per plot. Maybe you should display a group of X single-sided violin plots where X is the number of methods you are trying to compare.\n\nI think it is misleading to portray everything in terms of speedup or improvement over the \"TPE solution with X iterations\". A more strictly meaningful metric here is accuracy (assuming there is only one dataset per benchmark). Assuming the performance to beat by original TPE was an 11% error rate, there is a big difference between a method which was able to achieve a 10% error rate and a method which was able to achieve a 5% error rate, yet both will be assessed by how quickly they achieved x < 10% error rate. I can't seem to find such figures in the appendices.\n\nTypos:\n\n- Section 3.1 page 3, argmax g(x) / b(x) << you mean g(x) / l(x)?\n- appendix G, you wrote TPE2 instead of T2PE", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Valuable framework, precisions required", "review": "The authors propose a new framework for hyperparameter optimization and transfer across incremental modifications to a given algorithm and its search space, a process called developer adjustments in the paper. The authors then propose a few strategies to transfer knowledge from previous HPO runs and evaluate them on a series of simulated benchmarks. Results show the value added by transferring information from previous runs, as well as the surprising efficiency of simply reusing the best found hyperparameters from the previous run. \n\nStrong points:\n\n- The framework is simple and clearly introduced. \n- Extensive experiments help bring to light the advantage of transferring across adjustments.\n- The paper is very well written. \n\nWeak points:\n\n- Not enough details on benchmarks, more on this below\n- The use of simulated benchmarks with surrogate models introduces noise in the evaluation\n- Comparisons with more baselines would be beneficial. RF and/or GP-based HPO methods are extremely popular and would have been easy to integrate with the best-first baseline. \n\nRecommendation:\n\nThe contributions are simple and incremental, and clearly rooted in machine learning engineering, however I still think they could be beneficial as a whole to the community given the extensive experiments realized. I have some issues with experiments, lack of details and baselines, but those issues are mostly fixable. I'll give the paper a weak accept for now.\n\nExtra comments:\n\nYou do not specify which benchmarks are based on lookup tables and which ones are based on surrogate models. From looking at the search spaces, I would assume that the SVM and XGB benchmarks are modeled via surrogate benchmarks and the FCN and NAS benchmarks are lookup tables, but this should be explicited in the paper (or appendix). Parameters used for the benchmark surrogate model should also be given (if defaults of Eggensperger are used, simply mention this). It is also not clear what underlying datasets are used, this bears some importance and should be mentioned, even if only in the Appendix.\n\nOn surrogate model benchmarks: It can be seen in (Eggensperger et al. 2015., Figure 2) that ordering of methods can shift due to noise in the surrogate model (a random forest?). This is likely going to have a bigger impact when trying to measure the speedup, which is measured when a method reaches a certain threshold of performance. This threshold is likely to be met during the convergence phase of algorithms, and this phase appears noisier  (i.e. looking at how the phases of transition differ between the true benchmark and the RF surrogate benchmark differ in Eggensperger et al. 2015). Have you given this any thought? Have you compared experiments with a few runs on a real benchmark?\n\nThe method you end up recommending only has its detailed performance shown in the appendix. This feels counterintuitive to me. This result should be featured in the paper itself. This is perhaps due to the used of those split violin plots, which force you to display only two methods per plot. Maybe you should display a group of X single-sided violin plots where X is the number of methods you are trying to compare.\n\nI think it is misleading to portray everything in terms of speedup or improvement over the \"TPE solution with X iterations\". A more strictly meaningful metric here is accuracy (assuming there is only one dataset per benchmark). Assuming the performance to beat by original TPE was an 11% error rate, there is a big difference between a method which was able to achieve a 10% error rate and a method which was able to achieve a 5% error rate, yet both will be assessed by how quickly they achieved x < 10% error rate. I can't seem to find such figures in the appendices.\n\nTypos:\n\n- Section 3.1 page 3, argmax g(x) / b(x) << you mean g(x) / l(x)?\n- appendix G, you wrote TPE2 instead of T2PE", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603846480084}, {"id": "o8JKdk9Y_U", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3570/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is addressing a problem which is quite relevant in practice, namely how to warmstart HP optimization after small changes have been done to the ML model. Such changes may modify the HP search space, both by adding/removing HPs, or by changing their value ranges. The paper is clearly written. It introduces 3 potential baselines, as well as a simple transfer strategy. All work is based on TPE, which is frequently used, but not SotA for HPO.\n\nThe paper does not elaborate on the motivations of these \"developer changes\". One could suspect many are attempts to modify/improve the HPO process itself, over the *same* model. And this is not really new, there is lots of prior work to help shaping search spaces, both by quantifying HP relevance or by learning search ranges. Say, a developer modifies the value range of an HP. What other motivation would there be than mistrust in the previous range, but no change of algorithm. Same for adding/removing an HP, which normally just means going from fixed default to HPO or back. In fact, the 8 benchmarks are all of that sort. My feeling is that by viewing the problem in this way (namely, just HP search space optimization), there is suddenly a lot more related work not taken into account here. More difficult problems, such as learning ensembles from a range of models, and then adding/removing model types, are not tackled here. These would call for more difficult transfer strategies.\n\nThis paper does not really propose new methodology, except maybe T2PE, which is a pretty basic heuristic. There is a lot of prior work on transfer HPO, some of which could cewrtainly be adopted. Given that the paper is mainly empirical, one would expect a more thorough and wider evaluation. On the positive side, the paper introduces 8 new benchmarks, even though they are pretty simple setups. Their empirical evaluations are a little thin. Only the best-first baseline works well, results for the others are not shown. It should be noted that best-first is standard in HPO practice, this is the first thing one does for transfer. Their T2PE essentially works just as well, and a combination of the two works slightly better. While the paper categorizes types of modification, the empirical evaluation does not differentiate among them anymore. Also, the restriction to TPE is questionable. Why not also use GP-BO? All baselines would work just the same.\n\nMy main recommendation for this work would be to be clear about the modification for such limited developer changes. If this is just about the developer trying to twist HPO in itself, this work would have to compare against previous work for optimizing search spaces. Otherwise, please address more complex scenarios, such as ensemble learning, where HPO transfer becomes really difficult.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Considers problem of warmstarting hyperparameter optimization after small changes to training algorithm and/or HP search space.", "review": "This paper is addressing a problem which is quite relevant in practice, namely how to warmstart HP optimization after small changes have been done to the ML model. Such changes may modify the HP search space, both by adding/removing HPs, or by changing their value ranges. The paper is clearly written. It introduces 3 potential baselines, as well as a simple transfer strategy. All work is based on TPE, which is frequently used, but not SotA for HPO.\n\nThe paper does not elaborate on the motivations of these \"developer changes\". One could suspect many are attempts to modify/improve the HPO process itself, over the *same* model. And this is not really new, there is lots of prior work to help shaping search spaces, both by quantifying HP relevance or by learning search ranges. Say, a developer modifies the value range of an HP. What other motivation would there be than mistrust in the previous range, but no change of algorithm. Same for adding/removing an HP, which normally just means going from fixed default to HPO or back. In fact, the 8 benchmarks are all of that sort. My feeling is that by viewing the problem in this way (namely, just HP search space optimization), there is suddenly a lot more related work not taken into account here. More difficult problems, such as learning ensembles from a range of models, and then adding/removing model types, are not tackled here. These would call for more difficult transfer strategies.\n\nThis paper does not really propose new methodology, except maybe T2PE, which is a pretty basic heuristic. There is a lot of prior work on transfer HPO, some of which could cewrtainly be adopted. Given that the paper is mainly empirical, one would expect a more thorough and wider evaluation. On the positive side, the paper introduces 8 new benchmarks, even though they are pretty simple setups. Their empirical evaluations are a little thin. Only the best-first baseline works well, results for the others are not shown. It should be noted that best-first is standard in HPO practice, this is the first thing one does for transfer. Their T2PE essentially works just as well, and a combination of the two works slightly better. While the paper categorizes types of modification, the empirical evaluation does not differentiate among them anymore. Also, the restriction to TPE is questionable. Why not also use GP-BO? All baselines would work just the same.\n\nMy main recommendation for this work would be to be clear about the modification for such limited developer changes. If this is just about the developer trying to twist HPO in itself, this work would have to compare against previous work for optimizing search spaces. Otherwise, please address more complex scenarios, such as ensemble learning, where HPO transfer becomes really difficult.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603701422276}, {"id": "MR_WqhSu4k", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3570/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Major weaknesses of the paper:\n- My understanding is that these are surrogate models that are meant to simulate a real-world task. However there is no description as to how these surrogates were created or trained, nor how their fidelity to the original task was vetted.\n- The most simple and naive algorithm seems to provide similar speed-ups to the much more complicated proposed T2PE; especially when considering the much larger improvement of the naive method (see Fig. 11), none of the other proposed methods seem justified to me.\n- Furthermore, I don't consider this naive approach (Best First) as being an HPO approach that leverages transfer, since it does exactly what anyone would do when faced with a slightly altered set of hyperparameters.\n- This last point suggests that at least one of the following must be true:\n  * non-trivial transfer is not as important as intuition would lead us to think;\n  * this benchmark suite does not provide a good testbed for assessing an HPO method's ability to transfer; or\n  * none of the non-trivial proposed algorithms do a good job transferring and can therefore not argue against the previous point.\n  \nGiven this important contradiction, I must recommend a rejection. My recommendations would be to:\n- focus on creating a good benchmark suite (perhaps focus on a single or two domains as introduce many variants, instead of four domains with only two variants);\n- focus on vetting the surrogates in terms of their fidelity to the task they are meant to simulate; and\n- focus on demonstrating that accounting for the slight modifications in the subsequent HPO does indeed provide a benefit over the naive thing to do.\nFor the record, I don't think this is an easy task.\n\nMinor points:\n- Justify geometric mean. I'm not saying it's the wrong way to compare these, I just think it requires at least a sentence of justification.\n- Same for the violin plots. For such simple plots, simple boxes and whiskers, with perhaps data points to show the spread of measurements across seeds, would do just fine.\n- Figure 4, and indeed any mention of the two methods therein, can be entirely removed from the paper; other than to perhaps mention that they were tried and failed---results in the appendix.\n- A much more interesting replacement for that figure would be Figure 11.\n- Not sure what is the point of comparing random search to TPE in the appendix unless this means Best-first then Random-search/TPE? If the latter is true, please clarify.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This manuscript introduces a benchmark suite of hyperparameter optimization tasks, that simulate slight developer modifications, with the goal of optimizing a slightly modified algorithm given knowledge of its previous form. The paper further studies the performance of some deliberately naive approaches as a performance benchmark to accompany the suite.", "review": "Major weaknesses of the paper:\n- My understanding is that these are surrogate models that are meant to simulate a real-world task. However there is no description as to how these surrogates were created or trained, nor how their fidelity to the original task was vetted.\n- The most simple and naive algorithm seems to provide similar speed-ups to the much more complicated proposed T2PE; especially when considering the much larger improvement of the naive method (see Fig. 11), none of the other proposed methods seem justified to me.\n- Furthermore, I don't consider this naive approach (Best First) as being an HPO approach that leverages transfer, since it does exactly what anyone would do when faced with a slightly altered set of hyperparameters.\n- This last point suggests that at least one of the following must be true:\n  * non-trivial transfer is not as important as intuition would lead us to think;\n  * this benchmark suite does not provide a good testbed for assessing an HPO method's ability to transfer; or\n  * none of the non-trivial proposed algorithms do a good job transferring and can therefore not argue against the previous point.\n  \nGiven this important contradiction, I must recommend a rejection. My recommendations would be to:\n- focus on creating a good benchmark suite (perhaps focus on a single or two domains as introduce many variants, instead of four domains with only two variants);\n- focus on vetting the surrogates in terms of their fidelity to the task they are meant to simulate; and\n- focus on demonstrating that accounting for the slight modifications in the subsequent HPO does indeed provide a benefit over the naive thing to do.\nFor the record, I don't think this is an easy task.\n\nMinor points:\n- Justify geometric mean. I'm not saying it's the wrong way to compare these, I just think it requires at least a sentence of justification.\n- Same for the violin plots. For such simple plots, simple boxes and whiskers, with perhaps data points to show the spread of measurements across seeds, would do just fine.\n- Figure 4, and indeed any mention of the two methods therein, can be entirely removed from the paper; other than to perhaps mention that they were tried and failed---results in the appendix.\n- A much more interesting replacement for that figure would be Figure 11.\n- Not sure what is the point of comparing random search to TPE in the appendix unless this means Best-first then Random-search/TPE? If the latter is true, please clarify.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603542504021}], "openreview_url": "https://openreview.net/forum?id=WPO0vDYLXem", "arxiv_id": "2010.13117", "paper_pdf": "papers/WPO0vDYLXem.pdf", "paper_pdf_sha256": "fb57b6b9691760ca88b708e830c616478ac05a17b1a7187f1d3a560da7030da6", "paper_pdf_bytes": 4031177, "paper_pdf_source": "openreview", "code_url": "https://github.com/hp-transfer/htaa_experiments", "code_repository": "hp-transfer/htaa_experiments", "code_commit": "7b5c3b9b53d4e623c25bb44fd703884f314c3bf6", "code_archive": "repos/WPO0vDYLXem.zip", "code_archive_sha256": "05e6a3ff2ba7da11b9f7fa64710e06346c1eda60f36df79b95fcc414ca127c7d", "code_archive_bytes": 86641, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 505, "github_languages": {"Python": 43464}, "github_archived": false, "github_pushed_at": "2021-10-04T01:12:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hyperparameter-transfer-across-developer-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1dMsKsfwJ1", "year": 2026, "status": "rejected", "title": "Dataset Distillation via Committee Voting", "authors": ["Jiacheng Cui", "Zhaoyi Li", "Xiaochen Ma", "Xinyue Bi", "Yaxin Luo", "Zhiqiang Shen"], "authorids": ["~Jiacheng_Cui1", "~Zhaoyi_Li6", "~Xiaochen_Ma1", "~Xinyue_Bi1", "~Yaxin_Luo1", "~Zhiqiang_Shen1"], "authors_source": "OpenReview API", "abstract": "Dataset distillation aims to synthesize a smaller, representative dataset that preserves the essential properties of the original data, enabling efficient model training with reduced computational resources. Prior work has primarily focused on improving the alignment or matching process between original and synthetic data, or on enhancing the efficiency of distilling large datasets. In this work, we introduce $\\bf C$ommittee $\\bf V$otingfor $\\bf D$ataset $\\bf D$istillation (CV-DD), a novel and orthogonal approach that leverages the collective wisdom of multiple models or experts to create high-quality distilled datasets. We start by showing how to establish a strong baseline that already achieves state-of-the-art accuracy through leveraging recent advancements and thoughtful adjustments in model design and optimization processes. By integrating distributions and predictions from a committee of models while generating accurate soft labels, our method captures a wider spectrum of data features, reduces model-specific biases and mitigates distributional shifts between synthetic data and original data. This voting-based strategy not only promotes diversity and robustness within the distilled dataset but also significantly reduces overfitting, resulting in improved performance on post-eval tasks. Extensive experiments across various datasets and IPCs (images per class) demonstrate that Committee Voting leads to more reliable and adaptable distilled data compared to single/multi-model distillation methods, demonstrating its potential for efficient and accurate dataset distillation.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "skMMFXLeyQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13004/Reviewer_tw2Q"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces Committee Voting for Dataset Distillation (CV-DD), a novel framework that enhances dataset distillation by combining predictions and feature distributions from multiple models (committee members), rather than depending on a single model. The authors also establish a stronger baseline, SRe2L++, and propose an improved criterion called Batch-Specific Soft Labeling (BSSL), which better aligns the distributions of real and synthetic images. The objective of CV-DD is to create smaller, high-quality synthetic datasets that preserve the essential characteristics of large real datasets while minimizing overfitting and reducing model-specific bias.", "review_text": "This paper introduces Committee Voting for Dataset Distillation (CV-DD), a novel framework that enhances dataset distillation by combining predictions and feature distributions from multiple models (committee members), rather than depending on a single model. The authors also establish a stronger baseline, SRe2L++, and propose an improved criterion called Batch-Specific Soft Labeling (BSSL), which better aligns the distributions of real and synthetic images. The objective of CV-DD is to create smaller, high-quality synthetic datasets that preserve the essential characteristics of large real datasets while minimizing overfitting and reducing model-specific bias.", "strengths": "1. Ensemble-based methods in dataset distillation remain an emerging area of research. \n\n2. The performance improvements reported in this paper are significant.", "weaknesses": "1. Although the proposed CV-DD framework enhances cross-model generalization by leveraging diverse architectures, its Batch-Specific Soft Labeling (BSSL) mechanism constrains the method to architectures that include Batch Normalization (BN) layers. This dependency limits the generalization and versatility of the approach when applied to models without BN components.\n\n2. The visualization in Figure 6 suffers from poor distinguishability between curves due to the use of similar colors and marker styles for different data series. This makes it difficult for readers to clearly identify and interpret the individual trends. Additionally, the legend overlaps with the plotted data, further obscuring important details. It is recommended to improve the figure’s clarity by using more distinct color schemes or line styles and repositioning the legend to avoid overlapping with the data.", "questions": "1. **Comparison with SRe2L:**\n Could the authors include the performance results of SRe2L in the experiments? The reviewer is particularly interested in quantifying the improvement achieved by the enhanced SRe2L++ baseline over the original SRe2L.\n\n\n2. **Ablation on Committee Composition:**\n The proposed method heavily depends on the quality and diversity of the committee members. However, the manuscript lacks a detailed discussion or analysis regarding this aspect. Could the authors include additional experiments that vary the composition or number of committee members to demonstrate the sensitivity and robustness of CV-DD to these factors?\n\n\n3. **Dependence on Batch Normalization:**\n While CV-DD leverages the diversity of committee members, the proposed Batch-Specific Soft Labeling (BSSL) relies on architectures containing Batch Normalization (BN) layers. This dependence appears to limit the applicability of CV-DD to models without BN layers, such as Vision Transformers (ViTs) [1], which are widely adopted in modern computer vision. Could the authors discuss or experiment with how CV-DD might be extended to such architectures?\n\n\n4. **Scalability with IPC:**\n How does the performance of CV-DD scale when the number of images per class (IPC) exceeds 50? Additional results or analysis in higher IPC regimes would help illustrate the scalability and practical limits of the proposed approach.\n\n\n5. **Integration with Other Frameworks:**\n The proposed CV-DD presents an orthogonal contribution to existing literature and demonstrates strong potential for integration with other dataset distillation frameworks, such as RDED [2] and IGD [3]. Could the authors discuss or explore the feasibility and potential benefits of combining CV-DD with these frameworks?\n\n\n[1] Alexey Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ICLR 2021\n\n[2] Peng Sun et al., On the Diversity and Realism of Distilled Dataset: An Efficient Dataset Distillation Paradigm, CVPR 2024\n\n[3] Mingyang Chen et al., Influence-Guided Diffusion for Dataset Distillation, ICLR 2025", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Committee Voting for Dataset Distillation (CV-DD), a novel framework that enhances dataset distillation by combining predictions and feature distributions from multiple models (committee members), rather than depending on a single model. The authors also establish a stronger baseline, SRe2L++, and propose an improved criterion called Batch-Specific Soft Labeling (BSSL), which better aligns the distributions of real and synthetic images. The objective of CV-DD is to create smaller, high-quality synthetic datasets that preserve the essential characteristics of large real datasets while minimizing overfitting and reducing model-specific bias.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Ensemble-based methods in dataset distillation remain an emerging area of research. \n\n2. The performance improvements reported in this paper are significant.", "weaknesses": "1. Although the proposed CV-DD framework enhances cross-model generalization by leveraging diverse architectures, its Batch-Specific Soft Labeling (BSSL) mechanism constrains the method to architectures that include Batch Normalization (BN) layers. This dependency limits the generalization and versatility of the approach when applied to models without BN components.\n\n2. The visualization in Figure 6 suffers from poor distinguishability between curves due to the use of similar colors and marker styles for different data series. This makes it difficult for readers to clearly identify and interpret the individual trends. Additionally, the legend overlaps with the plotted data, further obscuring important details. It is recommended to improve the figure’s clarity by using more distinct color schemes or line styles and repositioning the legend to avoid overlapping with the data.", "questions": "1. **Comparison with SRe2L:**\n Could the authors include the performance results of SRe2L in the experiments? The reviewer is particularly interested in quantifying the improvement achieved by the enhanced SRe2L++ baseline over the original SRe2L.\n\n\n2. **Ablation on Committee Composition:**\n The proposed method heavily depends on the quality and diversity of the committee members. However, the manuscript lacks a detailed discussion or analysis regarding this aspect. Could the authors include additional experiments that vary the composition or number of committee members to demonstrate the sensitivity and robustness of CV-DD to these factors?\n\n\n3. **Dependence on Batch Normalization:**\n While CV-DD leverages the diversity of committee members, the proposed Batch-Specific Soft Labeling (BSSL) relies on architectures containing Batch Normalization (BN) layers. This dependence appears to limit the applicability of CV-DD to models without BN layers, such as Vision Transformers (ViTs) [1], which are widely adopted in modern computer vision. Could the authors discuss or experiment with how CV-DD might be extended to such architectures?\n\n\n4. **Scalability with IPC:**\n How does the performance of CV-DD scale when the number of images per class (IPC) exceeds 50? Additional results or analysis in higher IPC regimes would help illustrate the scalability and practical limits of the proposed approach.\n\n\n5. **Integration with Other Frameworks:**\n The proposed CV-DD presents an orthogonal contribution to existing literature and demonstrates strong potential for integration with other dataset distillation frameworks, such as RDED [2] and IGD [3]. Could the authors discuss or explore the feasibility and potential benefits of combining CV-DD with these frameworks?\n\n\n[1] Alexey Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ICLR 2021\n\n[2] Peng Sun et al., On the Diversity and Realism of Distilled Dataset: An Efficient Dataset Distillation Paradigm, CVPR 2024\n\n[3] Mingyang Chen et al., Influence-Guided Diffusion for Dataset Distillation, ICLR 2025", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761810230433}, {"id": "8PYoanrjqu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13004/Reviewer_RHkG"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper presents Committee Voting for Dataset Distillation (CV-DD), a framework designed to generate diverse and representative distilled datasets by aggregating knowledge from multiple heterogeneous models. The method incorporates a Prior Performance Guided Voting Strategy to adaptively weight model contributions and a Batch-Specific Soft Labeling (BSSL) mechanism to mitigate distribution shifts between synthetic and real data. Experiments on CIFAR-10/100, Tiny-ImageNet, and ImageNet-1K indicate that CV-DD achieves competitive performance across various IPC settings and demonstrates strong cross-architecture generalization.", "review_text": "This paper presents Committee Voting for Dataset Distillation (CV-DD), a framework designed to generate diverse and representative distilled datasets by aggregating knowledge from multiple heterogeneous models. The method incorporates a Prior Performance Guided Voting Strategy to adaptively weight model contributions and a Batch-Specific Soft Labeling (BSSL) mechanism to mitigate distribution shifts between synthetic and real data. Experiments on CIFAR-10/100, Tiny-ImageNet, and ImageNet-1K indicate that CV-DD achieves competitive performance across various IPC settings and demonstrates strong cross-architecture generalization.", "strengths": "1. Committee voting is a novel approach for dataset distillation that improves data representativeness through prior performance–based weighting.\n2. CV-DD achieves state-of-the-art results across datasets, e.g., 59.5% on ImageNet-1K (IPC=50) versus 56.5% for RDED, with strong cross-architecture generalization.\n3. Ablations confirm the effectiveness of voting temperature, prior-guided voting, and BSSL in reducing distribution shift.", "weaknesses": "**Major:**\n\n1. A large portion of the paper focuses on building a strong baseline SRe2L++, which already incorporates several optimizations from recent SOTA methods such as EDC and RDED. Since SRe2L++ itself performs at a very high level, the additional gain from the committee voting mechanism appears modest, reducing the overall sense of novelty.\n2. Although the method reports better per-iteration efficiency than G-VBSM, the overall training pipeline is complex. It requires pretraining all committee models, running a time-consuming distillation–evaluation loop to assess prior performance, and maintaining multiple teacher models during training for gradient weighting. The high pretraining cost and runtime memory usage may limit its practicality on large-scale datasets such as ImageNet-1K.\n3. The ablation study shows that increasing the number of experts N from 2 to 3 leads to performance degradation. This contradicts the intuition that more models should improve robustness, suggesting that the committee voting mechanism may have an upper limit or introduce redundancy, which restricts its scalability.\n\n**Minor:**\n\n1. The citation related to DD in Line 122 is incorrect.\n2. The CIFAR-100 IPC-10/50 results in Table 2 are inconsistent with those in Table 1.\n3. Figure 9 appears slightly blurred.", "questions": "1. Please include a clearer ablation study to isolate the contribution of the prior performance guided voting strategy over the SRe2L++ baseline. Specifically, compare (1) the original SRe2L++ (single model), (2) SRe2L++ with an equally weighted committee plus BSSL, and (3) the full CV-DD with prior-guided voting plus BSSL.\n2. Beyond the per-iteration time reported in Table 3, provide a more detailed analysis of computational cost, including total pretraining time, prior performance evaluation time, and peak GPU memory usage, especially on ImageNet-1K. This would help readers assess the overall efficiency compared with single-model methods such as SRe2L++ or RDED.\n3. Regarding the performance drop with larger N, please provide further analysis on model diversity and the voting mechanism. It would be useful to discuss whether two experts already provide sufficient diversity and whether adding a third causes gradient conflicts or unstable weighting.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Committee Voting for Dataset Distillation (CV-DD), a framework designed to generate diverse and representative distilled datasets by aggregating knowledge from multiple heterogeneous models. The method incorporates a Prior Performance Guided Voting Strategy to adaptively weight model contributions and a Batch-Specific Soft Labeling (BSSL) mechanism to mitigate distribution shifts between synthetic and real data. Experiments on CIFAR-10/100, Tiny-ImageNet, and ImageNet-1K indicate that CV-DD achieves competitive performance across various IPC settings and demonstrates strong cross-architecture generalization.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. Committee voting is a novel approach for dataset distillation that improves data representativeness through prior performance–based weighting.\n2. CV-DD achieves state-of-the-art results across datasets, e.g., 59.5% on ImageNet-1K (IPC=50) versus 56.5% for RDED, with strong cross-architecture generalization.\n3. Ablations confirm the effectiveness of voting temperature, prior-guided voting, and BSSL in reducing distribution shift.", "weaknesses": "**Major:**\n\n1. A large portion of the paper focuses on building a strong baseline SRe2L++, which already incorporates several optimizations from recent SOTA methods such as EDC and RDED. Since SRe2L++ itself performs at a very high level, the additional gain from the committee voting mechanism appears modest, reducing the overall sense of novelty.\n2. Although the method reports better per-iteration efficiency than G-VBSM, the overall training pipeline is complex. It requires pretraining all committee models, running a time-consuming distillation–evaluation loop to assess prior performance, and maintaining multiple teacher models during training for gradient weighting. The high pretraining cost and runtime memory usage may limit its practicality on large-scale datasets such as ImageNet-1K.\n3. The ablation study shows that increasing the number of experts N from 2 to 3 leads to performance degradation. This contradicts the intuition that more models should improve robustness, suggesting that the committee voting mechanism may have an upper limit or introduce redundancy, which restricts its scalability.\n\n**Minor:**\n\n1. The citation related to DD in Line 122 is incorrect.\n2. The CIFAR-100 IPC-10/50 results in Table 2 are inconsistent with those in Table 1.\n3. Figure 9 appears slightly blurred.", "questions": "1. Please include a clearer ablation study to isolate the contribution of the prior performance guided voting strategy over the SRe2L++ baseline. Specifically, compare (1) the original SRe2L++ (single model), (2) SRe2L++ with an equally weighted committee plus BSSL, and (3) the full CV-DD with prior-guided voting plus BSSL.\n2. Beyond the per-iteration time reported in Table 3, provide a more detailed analysis of computational cost, including total pretraining time, prior performance evaluation time, and peak GPU memory usage, especially on ImageNet-1K. This would help readers assess the overall efficiency compared with single-model methods such as SRe2L++ or RDED.\n3. Regarding the performance drop with larger N, please provide further analysis on model diversity and the voting mechanism. It would be useful to discuss whether two experts already provide sufficient diversity and whether adding a third causes gradient conflicts or unstable weighting.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761527779751}, {"id": "e1NvVv6m2B", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13004/Reviewer_kxaT"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes CV-DD, a dataset distillation framework that introduces a Prior-based Voting Strategy on top of SRe²L. The authors also incorporate several training refinements, such as real-image initialization, data augmentation, and smoothed learning rate scheduling, claiming SOTA performance under multiple IPC settings. However, the novelty appears marginal, and the fairness of the experiments is questionable.", "review_text": "This paper proposes CV-DD, a dataset distillation framework that introduces a Prior-based Voting Strategy on top of SRe²L. The authors also incorporate several training refinements, such as real-image initialization, data augmentation, and smoothed learning rate scheduling, claiming SOTA performance under multiple IPC settings. However, the novelty appears marginal, and the fairness of the experiments is questionable.", "strengths": "The overall presentation quality is good. The figures are visually clear and help the reader understand the methodology.\n\nThe motivation of performance-guided voting is intuitively reasonable in principle.", "weaknesses": "The proposed method appears to be a marginal modification over SRe²L. The core contribution, i.e., “Prior-based Voting”, is computationally expensive yet results in only marginal improvements, as shown in Table 4 (middle).\n\nSeveral enhanced training tricks are adopted only for the proposed method and SRe²L++, like \"Smoothed Learning Rate & Smaller Batch Size\", which is not applied to other competing methods, leading to potentially unfair comparisons.\n\nThe comparison methods are limited. The main results only include two valid baselines, RDED and SRe²L, while another relevant method, CDA, which also follows a similar paradigm as SRe²L, is missing under most experimental settings. In addition, several comparison methods are relatively outdated.\n\nThe reported results are inappropriately cited. For instance, the MTT baseline is evaluated under an altered setting that deviates from its official configuration and appears to be intentionally adjusted in favor of the proposed method.\n\nThere are typos, e.g., in line 323, “Large scale dataset” should be written in lowercase.", "questions": "- What is the architecture used in Fig. 1? Please specify it clearly.\n\n- Regarding the Prior-based Voting in Section 3.4 and Algorithm 1, it appears that distillation and evaluation must be conducted separately for each model in the committee set, which will incur substantial computational overhead. In contrast, SRe2L also relies on pretrained models, but those are standard models that can often be assumed to be readily available. \n\n- Moreover, the MTT results in Table 2 are substantially lower than the original reported performance. What accounts for this degradation? Was it re-implemented using an ensemble strategy? If so, this deviates from the original setting and may distort the comparison, especially since the reported results are significantly lower than those reported in the original MTT paper.\n\n- The ablation study is incomplete and fails to isolate the contribution of key components. Specifically, the impact of the proposed Prior-based Voting is not evaluated independently. Moreover, several additional enhancements, such as Real Image Initialization, Data Augmentation, and Smoothed Learning Rate with Smaller Batch Size, are grouped together without showing their individual effects on performance. If these techniques are treated as default settings, they should be consistently applied to all compared methods (not only to SRe²L) to ensure a fair and unbiased comparison.\n\n- The reported results are very selective. Why is there no cross-architecture performance reported for other methods in Table 2?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CV-DD, a dataset distillation framework that introduces a Prior-based Voting Strategy on top of SRe²L. The authors also incorporate several training refinements, such as real-image initialization, data augmentation, and smoothed learning rate scheduling, claiming SOTA performance under multiple IPC settings. However, the novelty appears marginal, and the fairness of the experiments is questionable.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The overall presentation quality is good. The figures are visually clear and help the reader understand the methodology.\n\nThe motivation of performance-guided voting is intuitively reasonable in principle.", "weaknesses": "The proposed method appears to be a marginal modification over SRe²L. The core contribution, i.e., “Prior-based Voting”, is computationally expensive yet results in only marginal improvements, as shown in Table 4 (middle).\n\nSeveral enhanced training tricks are adopted only for the proposed method and SRe²L++, like \"Smoothed Learning Rate & Smaller Batch Size\", which is not applied to other competing methods, leading to potentially unfair comparisons.\n\nThe comparison methods are limited. The main results only include two valid baselines, RDED and SRe²L, while another relevant method, CDA, which also follows a similar paradigm as SRe²L, is missing under most experimental settings. In addition, several comparison methods are relatively outdated.\n\nThe reported results are inappropriately cited. For instance, the MTT baseline is evaluated under an altered setting that deviates from its official configuration and appears to be intentionally adjusted in favor of the proposed method.\n\nThere are typos, e.g., in line 323, “Large scale dataset” should be written in lowercase.", "questions": "- What is the architecture used in Fig. 1? Please specify it clearly.\n\n- Regarding the Prior-based Voting in Section 3.4 and Algorithm 1, it appears that distillation and evaluation must be conducted separately for each model in the committee set, which will incur substantial computational overhead. In contrast, SRe2L also relies on pretrained models, but those are standard models that can often be assumed to be readily available. \n\n- Moreover, the MTT results in Table 2 are substantially lower than the original reported performance. What accounts for this degradation? Was it re-implemented using an ensemble strategy? If so, this deviates from the original setting and may distort the comparison, especially since the reported results are significantly lower than those reported in the original MTT paper.\n\n- The ablation study is incomplete and fails to isolate the contribution of key components. Specifically, the impact of the proposed Prior-based Voting is not evaluated independently. Moreover, several additional enhancements, such as Real Image Initialization, Data Augmentation, and Smoothed Learning Rate with Smaller Batch Size, are grouped together without showing their individual effects on performance. If these techniques are treated as default settings, they should be consistently applied to all compared methods (not only to SRe²L) to ensure a fair and unbiased comparison.\n\n- The reported results are very selective. Why is there no cross-architecture performance reported for other methods in Table 2?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761025385238}, {"id": "vssbC0YpGu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13004/Reviewer_4Bi4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes CV-DD (Dataset Distillation via Committee Voting), a novel framework that improves dataset distillation (DD) by integrating knowledge from multiple models instead of relying on a single teacher backbone.\nThe key insight is that existing decoupled DD methods suffer from *single-model bias* and limited diversity, as the synthetic data inherits the inductive bias of one architecture.\nCV-DD introduces a committee voting mechanism, where several heterogeneous teacher models (e.g., ResNet18/50, MobileNetV2, ShuffleNetV2, DenseNet121) contribute to the synthesis process.\nEach teacher’s contribution is weighted by its *prior performance* (estimated on real data), and gradients are aggregated via a softmax-weighted voting rule.\nTheoretical analyses show that model diversity encourages more diverse synthetic data and that prior-guided voting aligns the update direction with better generalization.", "review_text": "This paper proposes CV-DD (Dataset Distillation via Committee Voting), a novel framework that improves dataset distillation (DD) by integrating knowledge from multiple models instead of relying on a single teacher backbone.\nThe key insight is that existing decoupled DD methods suffer from *single-model bias* and limited diversity, as the synthetic data inherits the inductive bias of one architecture.\nCV-DD introduces a committee voting mechanism, where several heterogeneous teacher models (e.g., ResNet18/50, MobileNetV2, ShuffleNetV2, DenseNet121) contribute to the synthesis process.\nEach teacher’s contribution is weighted by its *prior performance* (estimated on real data), and gradients are aggregated via a softmax-weighted voting rule.\nTheoretical analyses show that model diversity encourages more diverse synthetic data and that prior-guided voting aligns the update direction with better generalization.", "strengths": "* **Clear motivation and insight:**\n  The paper addresses a genuine and underexplored issue in dataset distillation — the *bias introduced by single-teacher distillation*.\n  Framing the solution as a “committee voting” problem is intuitive yet novel.\n* **Solid technical design:**\n  The combination of prior-guided voting and batch-specific soft labeling is well-justified and effectively integrated into the distillation loop.\n  The enhancements (SRe²L++ baseline, real initialization, smoothed learning rate) are reasonable and necessary for a fair comparison.\n* **Theoretical support:**\n  The paper provides formal analysis showing that committee diversity contributes to broader gradient coverage and improved generalization.\n  This strengthens the conceptual grounding of the method.\n* **Empirical improvements:**\n  Consistent gains across multiple benchmarks, including large-scale ImageNet-1K, demonstrate practical utility and scalability.\n  The results are competitive and in several cases surpass existing state-of-the-art methods.", "weaknesses": "* **Computation cost and scalability:**\n  The committee-based design involves multiple teachers and prior evaluation steps, which may substantially increase computation.\n  The paper should report training time, GPU hours, and discuss trade-offs between accuracy and cost more explicitly.\n* **Limited domain diversity:**\n  The framework is validated only on image classification datasets.\n  It would be valuable to explore its applicability to other modalities (e.g., graph or multimodal data) or to synthetic-to-real transfer tasks.", "questions": "1. How does CV-DD’s computational cost scale with the number of committee members? Can it be reduced by low-rank or partial gradient aggregation?\n2. How robust is the prior-guided voting to noisy or overfitted teacher models? Would dynamic online re-weighting (instead of fixed priors) help?\n3. Have you tested CV-DD on other domains (e.g., graph or cross-modal distillation) to confirm generality?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes CV-DD (Dataset Distillation via Committee Voting), a novel framework that improves dataset distillation (DD) by integrating knowledge from multiple models instead of relying on a single teacher backbone.\nThe key insight is that existing decoupled DD methods suffer from *single-model bias* and limited diversity, as the synthetic data inherits the inductive bias of one architecture.\nCV-DD introduces a committee voting mechanism, where several heterogeneous teacher models (e.g., ResNet18/50, MobileNetV2, ShuffleNetV2, DenseNet121) contribute to the synthesis process.\nEach teacher’s contribution is weighted by its *prior performance* (estimated on real data), and gradients are aggregated via a softmax-weighted voting rule.\nTheoretical analyses show that model diversity encourages more diverse synthetic data and that prior-guided voting aligns the update direction with better generalization.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* **Clear motivation and insight:**\n  The paper addresses a genuine and underexplored issue in dataset distillation — the *bias introduced by single-teacher distillation*.\n  Framing the solution as a “committee voting” problem is intuitive yet novel.\n* **Solid technical design:**\n  The combination of prior-guided voting and batch-specific soft labeling is well-justified and effectively integrated into the distillation loop.\n  The enhancements (SRe²L++ baseline, real initialization, smoothed learning rate) are reasonable and necessary for a fair comparison.\n* **Theoretical support:**\n  The paper provides formal analysis showing that committee diversity contributes to broader gradient coverage and improved generalization.\n  This strengthens the conceptual grounding of the method.\n* **Empirical improvements:**\n  Consistent gains across multiple benchmarks, including large-scale ImageNet-1K, demonstrate practical utility and scalability.\n  The results are competitive and in several cases surpass existing state-of-the-art methods.", "weaknesses": "* **Computation cost and scalability:**\n  The committee-based design involves multiple teachers and prior evaluation steps, which may substantially increase computation.\n  The paper should report training time, GPU hours, and discuss trade-offs between accuracy and cost more explicitly.\n* **Limited domain diversity:**\n  The framework is validated only on image classification datasets.\n  It would be valuable to explore its applicability to other modalities (e.g., graph or multimodal data) or to synthetic-to-real transfer tasks.", "questions": "1. How does CV-DD’s computational cost scale with the number of committee members? Can it be reduced by low-rank or partial gradient aggregation?\n2. How robust is the prior-guided voting to noisy or overfitted teacher models? Would dynamic online re-weighting (instead of fixed priors) help?\n3. Have you tested CV-DD on other domains (e.g., graph or cross-modal distillation) to confirm generality?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760521109856}], "openreview_url": "https://openreview.net/forum?id=1dMsKsfwJ1", "arxiv_id": "2501.07575", "paper_pdf": "papers/1dMsKsfwJ1.pdf", "paper_pdf_sha256": "a9624299834fcd360a87b90fad5332c350c2798cc3528dc652d95a6f0fda9319", "paper_pdf_bytes": 5268947, "paper_pdf_source": "openreview", "code_url": "https://github.com/Jiacheng8/CV-DD", "code_repository": "Jiacheng8/CV-DD", "code_commit": "24097e9c7058954b90eb442daa7cc6ba4b774830", "code_archive": "repos/1dMsKsfwJ1.zip", "code_archive_sha256": "82a6ef59ca3d400274a346f14fa856fe52735b1c86ebf96a18ed6e617f985ae3", "code_archive_bytes": 1493776, "code_file_count": 174, "code_extensions": {".sh": 159, ".py": 15}, "github_disk_usage_kb": 1913, "github_languages": {"Shell": 148806, "Python": 118403}, "github_archived": false, "github_pushed_at": "2025-07-28T04:55:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dataset-distillation-via-committee-voting"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PRKFRzOEq8", "year": 2025, "status": "rejected", "title": "Estimating the conformal prediction threshold from noisy labels", "authors": ["Coby Penso", "Jacob Goldberger", "Ethan Fetaya"], "authorids": ["~Coby_Penso1", "~Jacob_Goldberger1", "~Ethan_Fetaya1"], "authors_source": "OpenReview API", "abstract": "Conformal Prediction (CP) is a method to control prediction uncertainty by producing a small prediction set,   ensuring a predetermined probability that the true class lies within this set.   This is commonly done by defining a score, based on the model predictions, and setting a threshold on this score using a validation set. In this study, we address the problem of CP calibration when we only have access to a validation set with noisy labels. We show how we can estimate the noise-free conformal threshold based on the noisy labeled data.   Our solution is flexible and can accommodate various modeling assumptions regarding the label contamination process, without needing any information about the underlying data distribution or the internal mechanisms of the machine learning classifier.    We develop a coverage guarantee for uniform noise that is effective even in tasks with a large number of classes. We dub our approach Noise-Aware Conformal Prediction (NACP) and show on several natural and medical image classification datasets, including ImageNet, that it significantly outperforms current noisy label methods and achieves results comparable to those obtained with a clean validation set.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "Sl7HqIR2PT", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4312/Reviewer_VNWi"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces a new conformal prediction framework for classification that is able to handle label noise. Specifically, given known uniform label noise, the authors present an algorithm that can achieve the desired coverage for the underlying unobserved clean data. They also extend their approach to handle general noise using a known noise matrix. Through various experiments, they demonstrate that their approach outperforms existing methods for managing label noise in CP.", "review_text": "This paper introduces a new conformal prediction framework for classification that is able to handle label noise. Specifically, given known uniform label noise, the authors present an algorithm that can achieve the desired coverage for the underlying unobserved clean data. They also extend their approach to handle general noise using a known noise matrix. Through various experiments, they demonstrate that their approach outperforms existing methods for managing label noise in CP.", "strengths": "The paper is well-structured and easy to follow and understand, addressing the intriguing problem of conformal prediction sets in the presence of noisy data.", "weaknesses": "1. My primary concern relates to the noise assumption. Firstly, it's unclear how realistic the assumption of uniform label noise is in practical applications. Additionally, assuming that the noise parameter is known or can be accurately estimated in real-world scenarios doesn't seem entirely realistic (for both uniform and general noise).\n\n2. In the experiments, different values of the noise parameter weren't tested to show how the model performs under varying noise levels (e.g., extending beyond $\\epsilon > 0.2$).\n\n3. For the results of the general noise model presented in Table 5, it would be fairer—particularly for the NR-CP approach—to include NACP with $\\Delta$ as well, as these two methods are specifically aimed at recovering finite-sample coverage guarantees under noise.", "questions": "1. Does the grid search for finding q in [q_1, q_2] yield a unique result? If yes, why? If not, what might this imply?\n\n\n2. According to the experimental results, it appears that NACP without $\\Delta$ performs best—achieving the required coverage while maintaining the smallest average set size in almost all experiments. Could the authors elaborate on why this is the case?\n\n3. While it’s valuable to see where the proposed method is effective, in what situations does it fail?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new conformal prediction framework for classification that is able to handle label noise. Specifically, given known uniform label noise, the authors present an algorithm that can achieve the desired coverage for the underlying unobserved clean data. They also extend their approach to handle general noise using a known noise matrix. Through various experiments, they demonstrate that their approach outperforms existing methods for managing label noise in CP.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper is well-structured and easy to follow and understand, addressing the intriguing problem of conformal prediction sets in the presence of noisy data.", "weaknesses": "1. My primary concern relates to the noise assumption. Firstly, it's unclear how realistic the assumption of uniform label noise is in practical applications. Additionally, assuming that the noise parameter is known or can be accurately estimated in real-world scenarios doesn't seem entirely realistic (for both uniform and general noise).\n\n2. In the experiments, different values of the noise parameter weren't tested to show how the model performs under varying noise levels (e.g., extending beyond $\\epsilon > 0.2$).\n\n3. For the results of the general noise model presented in Table 5, it would be fairer—particularly for the NR-CP approach—to include NACP with $\\Delta$ as well, as these two methods are specifically aimed at recovering finite-sample coverage guarantees under noise.", "questions": "1. Does the grid search for finding q in [q_1, q_2] yield a unique result? If yes, why? If not, what might this imply?\n\n\n2. According to the experimental results, it appears that NACP without $\\Delta$ performs best—achieving the required coverage while maintaining the smallest average set size in almost all experiments. Could the authors elaborate on why this is the case?\n\n3. While it’s valuable to see where the proposed method is effective, in what situations does it fail?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730720639239}, {"id": "pDDcKNCcxM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4312/Reviewer_9NiR"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "This paper proposes a new method for calibrating conformal predictions when using a validation set with noisy labels. While conformity scores are based on previous research, the authors introduce a novel approach to calibrate these scores in the presence of noisy-labeled data in the validation set. This paper mainly addresses the cases where the noise distribution and its parameters are known. The main idea is to utilize the known noise distribution to reconstruct the clean distribution of the conformity scores.", "review_text": "This paper proposes a new method for calibrating conformal predictions when using a validation set with noisy labels. While conformity scores are based on previous research, the authors introduce a novel approach to calibrate these scores in the presence of noisy-labeled data in the validation set. This paper mainly addresses the cases where the noise distribution and its parameters are known. The main idea is to utilize the known noise distribution to reconstruct the clean distribution of the conformity scores.", "strengths": "- The proposed calibration method is applicable to a wide range of conformity scores.\n- The authors provide a proof that the proposed calibration method meets coverage guarantees.\n- Experimental results show that the prediction sets produced by the proposed method are significantly smaller than those generated by other existing methods on the given dataset and in the specified conditions.", "weaknesses": "- The main assumptions are highly restrictive. The authors assume that both the noise distribution and its parameters are known, and that the training data is either clean or uses pre-trained checkpoints without noisy data, while only the validation data contains noise. This scenario seems uncommon, which may limit the practical applicability of the proposed method.\n- Additional real-world experimental studies could strengthen the motivation behind the proposed method. For example, clarifying real-world scenarios where the validation dataset is noisy or intentionally corrupted with the known noise distribution would be beneficial. Although the authors conducted experiments on popular classification datasets, the approach remains somewhat unconvincing, since these experiments could be considered as simulations with artificially injected noise. While the method demonstrates superior performance on these datasets, it would be more compelling if there were real-world examples in which the validation data is noisy with a known distribution and parameters. Without such context, the motivation for this method remains somewhat unclear.", "questions": "- While the meaning of $|C_q|$ and $C_q$ can be inferred from the context, a clear definition of these notations would improve readability.\n- Although $\\widehat{F}^c$ becomes more complex in the noisy case, it still depends only on the indicator function and the count of elements in $C_q$. It also appears to be piecewise constant, with potentially multiple values of $q$ satisfying $\\widehat{F}^c(q)=1-\\alpha$. In such cases how do you select $q$?\n- If the authors intended that the clean label is corrupted into another label with probability of $\\varepsilon$, equation (4) might need correction. Currently, equation (4) leads to a noisy label with probability of $\\frac{k-1}{k}\\varepsilon$ and clean label with probability of $1-\\frac{k-1}{k}\\varepsilon$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new method for calibrating conformal predictions when using a validation set with noisy labels. While conformity scores are based on previous research, the authors introduce a novel approach to calibrate these scores in the presence of noisy-labeled data in the validation set. This paper mainly addresses the cases where the noise distribution and its parameters are known. The main idea is to utilize the known noise distribution to reconstruct the clean distribution of the conformity scores.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "- The proposed calibration method is applicable to a wide range of conformity scores.\n- The authors provide a proof that the proposed calibration method meets coverage guarantees.\n- Experimental results show that the prediction sets produced by the proposed method are significantly smaller than those generated by other existing methods on the given dataset and in the specified conditions.", "weaknesses": "- The main assumptions are highly restrictive. The authors assume that both the noise distribution and its parameters are known, and that the training data is either clean or uses pre-trained checkpoints without noisy data, while only the validation data contains noise. This scenario seems uncommon, which may limit the practical applicability of the proposed method.\n- Additional real-world experimental studies could strengthen the motivation behind the proposed method. For example, clarifying real-world scenarios where the validation dataset is noisy or intentionally corrupted with the known noise distribution would be beneficial. Although the authors conducted experiments on popular classification datasets, the approach remains somewhat unconvincing, since these experiments could be considered as simulations with artificially injected noise. While the method demonstrates superior performance on these datasets, it would be more compelling if there were real-world examples in which the validation data is noisy with a known distribution and parameters. Without such context, the motivation for this method remains somewhat unclear.", "questions": "- While the meaning of $|C_q|$ and $C_q$ can be inferred from the context, a clear definition of these notations would improve readability.\n- Although $\\widehat{F}^c$ becomes more complex in the noisy case, it still depends only on the indicator function and the count of elements in $C_q$. It also appears to be piecewise constant, with potentially multiple values of $q$ satisfying $\\widehat{F}^c(q)=1-\\alpha$. In such cases how do you select $q$?\n- If the authors intended that the clean label is corrupted into another label with probability of $\\varepsilon$, equation (4) might need correction. Currently, equation (4) leads to a noisy label with probability of $\\frac{k-1}{k}\\varepsilon$ and clean label with probability of $1-\\frac{k-1}{k}\\varepsilon$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730695319171}, {"id": "LzPbuQSpob", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4312/Reviewer_KQBp"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper studies the behavior of conformal prediction in the case of validation data with noisy labels and develop a method for estimating the noise-free CP threshold in the context of noisy validation data. The noise-aware conformal prediction achieves better performance in average set size than other state-of-the-art algorithms.", "review_text": "This paper studies the behavior of conformal prediction in the case of validation data with noisy labels and develop a method for estimating the noise-free CP threshold in the context of noisy validation data. The noise-aware conformal prediction achieves better performance in average set size than other state-of-the-art algorithms.", "strengths": "Extensive experiments for performance comparison to show the advantage of NACP in reducing the average set size while strictly maintaining the validity of coverage;\nConsideration of several noise models, including random noise and noise matrix, and this greatly expands the potential application of the proposed method. \nEstablishement of an end-to-end CP for estimating noise level in the training of neural network.", "weaknesses": "If you can demonstrate NACP in a more practical setting rather than a simulated environment, this will make the implication of the theorectical contribution of this paper more convincing. \nWhat is the connection between the two noise models (uniform noise distribution, noise matrix) and practical learning tasks? How should I choose which noise model I should use in practical problems, by trial and error or other approaches?\nin Fig. 2, can you also include the set size assocaited with each conformal predictors as this is also an important aspect to evaluate CP performance.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the behavior of conformal prediction in the case of validation data with noisy labels and develop a method for estimating the noise-free CP threshold in the context of noisy validation data. The noise-aware conformal prediction achieves better performance in average set size than other state-of-the-art algorithms.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Extensive experiments for performance comparison to show the advantage of NACP in reducing the average set size while strictly maintaining the validity of coverage;\nConsideration of several noise models, including random noise and noise matrix, and this greatly expands the potential application of the proposed method. \nEstablishement of an end-to-end CP for estimating noise level in the training of neural network.", "weaknesses": "If you can demonstrate NACP in a more practical setting rather than a simulated environment, this will make the implication of the theorectical contribution of this paper more convincing. \nWhat is the connection between the two noise models (uniform noise distribution, noise matrix) and practical learning tasks? How should I choose which noise model I should use in practical problems, by trial and error or other approaches?\nin Fig. 2, can you also include the set size assocaited with each conformal predictors as this is also an important aspect to evaluate CP performance.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730682991204}, {"id": "nSclh8W2h2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4312/Reviewer_4Nej"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper addresses the problem of conformal prediction when the validation set has noisy labels. Algorithms are proposed for two different types of label noises: a uniform noise level $\\epsilon$ or a more general noise matrix. In both cases, the noise condition (level/matrix) are assumed to be known. The algorithms will depend on such noise condition.  The paper analyzes the coverage and sample size required from the validation set, and shows the superiority of the proposed method compared with existing ones, e.g., (Sesia et al. 2023), especially regarding data with large label set and small sizes per class. \n\nOverall, the paper is well written with well justified algorithms and strong empirical results. However, I do have the following concerns/questions that need to be clarified. \n\n1, I have some concerns over the problem setup. If I am not mistaken, the algorithms assume a good model, i.e., the model is trained on clean labels. But if the validation set is already noisy, it is hard to assume the training data is clean, right? So ideally, the method should be discussing a conformal prediction training and validating on noisy labels. Even if this is too much to ask, there should be experimental evaluation on how the proposed algorithm (and the baselines) perform on the validation set when the model is trained on equally noisy training set. \n\n2, The proposed methods assume a known noise condition. This is unrealistic. The paper justifies by saying that one could estimate the noise level or the noise matrix through training set. But this would assume a noisy training set in the first place. If the noise condition is estimated through the validation set, there is a question of how reliable the noise condition estimation. In particular, the smaller the sample size is per class, the less reliable these noise condition estimation will be. \n\n3, the above concern should be discussed regarding other baselines. For example, does SoTA methods (e.g., Sesia et al. 2023) depend on the oracle noise condition? In experiments, if one has to estimate noise using the noisy validation set, how will the proposed method compare with Sesia et al. 2023, and compare with a vanilla conformal prediction?\n\n4, the statement that the second algorithm can handle general noise is an over-statement. Even with the noise matrix, the paper is still assuming the noise is iid. There are plenty of research on label noise problem assuming the noise is not iid, e.g., feature dependent. I understand this is still early stage of the problem, so it is OK not tackling such noise models. But it would be important to mention this in the paper.\n\nMinor errors:\nLine 88 – C(x) --> C_q(x)\nEquation 16, minor typo?", "review_text": "This paper addresses the problem of conformal prediction when the validation set has noisy labels. Algorithms are proposed for two different types of label noises: a uniform noise level $\\epsilon$ or a more general noise matrix. In both cases, the noise condition (level/matrix) are assumed to be known. The algorithms will depend on such noise condition.  The paper analyzes the coverage and sample size required from the validation set, and shows the superiority of the proposed method compared with existing ones, e.g., (Sesia et al. 2023), especially regarding data with large label set and small sizes per class. \n\nOverall, the paper is well written with well justified algorithms and strong empirical results. However, I do have the following concerns/questions that need to be clarified. \n\n1, I have some concerns over the problem setup. If I am not mistaken, the algorithms assume a good model, i.e., the model is trained on clean labels. But if the validation set is already noisy, it is hard to assume the training data is clean, right? So ideally, the method should be discussing a conformal prediction training and validating on noisy labels. Even if this is too much to ask, there should be experimental evaluation on how the proposed algorithm (and the baselines) perform on the validation set when the model is trained on equally noisy training set. \n\n2, The proposed methods assume a known noise condition. This is unrealistic. The paper justifies by saying that one could estimate the noise level or the noise matrix through training set. But this would assume a noisy training set in the first place. If the noise condition is estimated through the validation set, there is a question of how reliable the noise condition estimation. In particular, the smaller the sample size is per class, the less reliable these noise condition estimation will be. \n\n3, the above concern should be discussed regarding other baselines. For example, does SoTA methods (e.g., Sesia et al. 2023) depend on the oracle noise condition? In experiments, if one has to estimate noise using the noisy validation set, how will the proposed method compare with Sesia et al. 2023, and compare with a vanilla conformal prediction?\n\n4, the statement that the second algorithm can handle general noise is an over-statement. Even with the noise matrix, the paper is still assuming the noise is iid. There are plenty of research on label noise problem assuming the noise is not iid, e.g., feature dependent. I understand this is still early stage of the problem, so it is OK not tackling such noise models. But it would be important to mention this in the paper.\n\nMinor errors:\nLine 88 – C(x) --> C_q(x)\nEquation 16, minor typo?", "strengths": "Solid algorithm and analysis. \n\nStrong empirical results.", "weaknesses": "Problem settings and assumptions.\n\nPotential unfair comparisons with existing methods.\n\nSee Summary for details.", "questions": "See Summary for details.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of conformal prediction when the validation set has noisy labels. Algorithms are proposed for two different types of label noises: a uniform noise level $\\epsilon$ or a more general noise matrix. In both cases, the noise condition (level/matrix) are assumed to be known. The algorithms will depend on such noise condition.  The paper analyzes the coverage and sample size required from the validation set, and shows the superiority of the proposed method compared with existing ones, e.g., (Sesia et al. 2023), especially regarding data with large label set and small sizes per class. \n\nOverall, the paper is well written with well justified algorithms and strong empirical results. However, I do have the following concerns/questions that need to be clarified. \n\n1, I have some concerns over the problem setup. If I am not mistaken, the algorithms assume a good model, i.e., the model is trained on clean labels. But if the validation set is already noisy, it is hard to assume the training data is clean, right? So ideally, the method should be discussing a conformal prediction training and validating on noisy labels. Even if this is too much to ask, there should be experimental evaluation on how the proposed algorithm (and the baselines) perform on the validation set when the model is trained on equally noisy training set. \n\n2, The proposed methods assume a known noise condition. This is unrealistic. The paper justifies by saying that one could estimate the noise level or the noise matrix through training set. But this would assume a noisy training set in the first place. If the noise condition is estimated through the validation set, there is a question of how reliable the noise condition estimation. In particular, the smaller the sample size is per class, the less reliable these noise condition estimation will be. \n\n3, the above concern should be discussed regarding other baselines. For example, does SoTA methods (e.g., Sesia et al. 2023) depend on the oracle noise condition? In experiments, if one has to estimate noise using the noisy validation set, how will the proposed method compare with Sesia et al. 2023, and compare with a vanilla conformal prediction?\n\n4, the statement that the second algorithm can handle general noise is an over-statement. Even with the noise matrix, the paper is still assuming the noise is iid. There are plenty of research on label noise problem assuming the noise is not iid, e.g., feature dependent. I understand this is still early stage of the problem, so it is OK not tackling such noise models. But it would be important to mention this in the paper.\n\nMinor errors:\nLine 88 – C(x) --> C_q(x)\nEquation 16, minor typo?", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Solid algorithm and analysis. \n\nStrong empirical results.", "weaknesses": "Problem settings and assumptions.\n\nPotential unfair comparisons with existing methods.\n\nSee Summary for details.", "questions": "See Summary for details.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730670476574}, {"id": "wtZw0x78lb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4312/Reviewer_81YJ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes a new method called Noise-Aware Conformal Prediction (NACP) to address the calibration issue when only noisy-labeled validation data is available. The method can estimate noise-free conformal thresholds from noisy data without requiring information about the underlying data distribution or classifier mechanisms. Experiments on medical and natural image datasets demonstrate that NACP significantly outperforms existing methods, especially for large-scale classification tasks.", "review_text": "This paper proposes a new method called Noise-Aware Conformal Prediction (NACP) to address the calibration issue when only noisy-labeled validation data is available. The method can estimate noise-free conformal thresholds from noisy data without requiring information about the underlying data distribution or classifier mechanisms. Experiments on medical and natural image datasets demonstrate that NACP significantly outperforms existing methods, especially for large-scale classification tasks.", "strengths": "- The paper is well-written. The authors provide theoretical coverage guarantees for their approach, particularly for uniform noise cases, and extend the framework to general noise transition matrices.\n- The extensive experiments demonstrate superior performance of NACP on various classification tasks, where prior methods fail to maintain reasonable prediction set sizes.", "weaknesses": "- The work lacks analysis on calibration set sizes, particularly for large-scale datasets like ImageNet. It remains unclear how the method performs with different validation set sizes and what is the minimum required samples per class for reliable performance.\n- The experiments were conducted only with ResNet-18 architecture, leaving it unclear whether the method's effectiveness generalizes across different network architectures such as deeper ResNets or Vision Transformers.", "questions": "Could you further explain how the noise transition matrices would be obtained in practice? While the paper shows promising results with different noise matrices, it would be helpful to clarify whether additional data is needed for estimation and how estimation errors might affect performance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new method called Noise-Aware Conformal Prediction (NACP) to address the calibration issue when only noisy-labeled validation data is available. The method can estimate noise-free conformal thresholds from noisy data without requiring information about the underlying data distribution or classifier mechanisms. Experiments on medical and natural image datasets demonstrate that NACP significantly outperforms existing methods, especially for large-scale classification tasks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper is well-written. The authors provide theoretical coverage guarantees for their approach, particularly for uniform noise cases, and extend the framework to general noise transition matrices.\n- The extensive experiments demonstrate superior performance of NACP on various classification tasks, where prior methods fail to maintain reasonable prediction set sizes.", "weaknesses": "- The work lacks analysis on calibration set sizes, particularly for large-scale datasets like ImageNet. It remains unclear how the method performs with different validation set sizes and what is the minimum required samples per class for reliable performance.\n- The experiments were conducted only with ResNet-18 architecture, leaving it unclear whether the method's effectiveness generalizes across different network architectures such as deeper ResNets or Vision Transformers.", "questions": "Could you further explain how the noise transition matrices would be obtained in practice? While the paper shows promising results with different noise matrices, it would be helpful to clarify whether additional data is needed for estimation and how estimation errors might affect performance.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730655754203}], "openreview_url": "https://openreview.net/forum?id=PRKFRzOEq8", "arxiv_id": "2501.12749", "paper_pdf": "papers/PRKFRzOEq8.pdf", "paper_pdf_sha256": "b58b40f9f7798d53831eeba9ae9ecc252056d1f76d574a8c6a7af29291622ef7", "paper_pdf_bytes": 437942, "paper_pdf_source": "openreview", "code_url": "https://github.com/cobypenso/Noise-Aware-Conformal-Prediction", "code_repository": "cobypenso/Noise-Aware-Conformal-Prediction", "code_commit": "64ce4d978cc6b734054839cc5b856de06570fd4d", "code_archive": "repos/PRKFRzOEq8.zip", "code_archive_sha256": "1c9cc9b891ef2d8644d507e0b5bdeabdf4dd79c930a6eb3209b08665ffc96c72", "code_archive_bytes": 140595, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 133, "github_languages": {"Python": 36273}, "github_archived": false, "github_pushed_at": "2024-05-21T09:45:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/estimating-the-conformal-prediction-threshold"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "F0XXA9OG13", "year": 2024, "status": "rejected", "title": "MediTab: Scaling Medical Tabular Data Predictors via Data Consolidation, Enrichment, and Refinement", "authors": ["Zifeng Wang", "Chufan Gao", "Cao Xiao", "Jimeng Sun"], "authorids": ["~Zifeng_Wang3", "~Chufan_Gao1", "~Cao_Xiao2", "~Jimeng_Sun3"], "authors_source": "OpenReview API", "abstract": "Tabular data prediction has been employed in medical applications such as patient health risk prediction. However, existing methods usually revolve around the algorithm design while overlooking the significance of data engineering. Medical tabular datasets frequently exhibit significant heterogeneity across different sources, with limited sample sizes per source. As such, previous predictors are often trained on manually curated small datasets that struggle to generalize across different tabular datasets during inference.\n\nThis paper proposes to scale medical tabular data predictors (MediTab) to various tabular inputs with varying features. The method uses a data engine that leverages large language models (LLMs) to consolidate tabular samples to overcome the barrier across tables with distinct schema. It also aligns out-domain data with the target task using a \"learn, annotate, and refinement'' pipeline. The expanded training data then enables the pre-trained MediTab to infer for arbitrary tabular input in the domain without fine-tuning, resulting in significant improvements over supervised baselines: it reaches an average ranking of 1.57 and 1.00 on 7 patient outcome prediction datasets and 3 trial outcome prediction datasets, respectively. In addition, MediTab exhibits impressive zero-shot performances: it outperforms supervised XGBoost models by 8.9% and 17.2% on average in two prediction tasks, respectively.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "x2FJRLVVaD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission988/Reviewer_zDeD"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes to scale medical tabular data predictors (MediTab) to handle diverse tabular inputs with varying features. The approach involves utilizing a large language models (LLMs) to merge tabular datasets, addressing the challenges presented by tables with different structures.  Additionally, it establishes a process for aligning out-of-domain data with the specific target task through a \"learn, annotate, and refinement\" pipeline.", "review_text": "This paper proposes to scale medical tabular data predictors (MediTab) to handle diverse tabular inputs with varying features. The approach involves utilizing a large language models (LLMs) to merge tabular datasets, addressing the challenges presented by tables with different structures.  Additionally, it establishes a process for aligning out-of-domain data with the specific target task through a \"learn, annotate, and refinement\" pipeline.", "strengths": "1. The core concept behind MediTab, involving the consolidation, enrichment, and refinement modules, is well-founded in its aim to improve the scalability of predictive models designed for medical tabular data.\n2. Implementing a sanity check through Large Language Model (LLM) Reflection is particularly important in the medical domain.\n3. The paper is clearly written; the quality is sound.\n4. The empirical evaluation in the paper is strong; relevant baselines are considered.\n5. The coverage of related work is extensive, with clear distinctions drawn from other studies.", "weaknesses": "1. It is not clear, how was splitting into train/val/test organised?\n2. The statement “After the quality check step, we obtain the original task dataset $T$ and the supplementary dataset $T_{sup}$ and have two potential options for model training. The first is to combine both datasets for training, but we have found that this approach results in suboptimal performance.” is a bit unclear. Why it is suboptimal, could authors elaborate on this?\n3. Are these datasets prone to missing values, which is a critical concern in the medical domain? If so, what would be the recommended strategy for handling these missing values?\n4. Results on Ablation studies on Different Learning strategies are provided. Could authors provide  Ablation studies on the different model components?", "questions": "1. Could the authors elaborate more on the LLM sanity check, as well as the results and tests provided in Appendix C.4-C.5? It has been discussed that it is crucial to conduct thorough evaluations of LLMs in healthcare, with particular attention to aspects of safety, equity, and bias [a]. Could the authors provide their thoughts on why they believe their model satisfies these requirements?\n2. Could authors provide more detailed explanation of empirical studies and address points in Weakness section?\n\na. Singhal et al., Large language models encode clinical knowledge. Nature 620, 172–180 (2023).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to scale medical tabular data predictors (MediTab) to handle diverse tabular inputs with varying features. The approach involves utilizing a large language models (LLMs) to merge tabular datasets, addressing the challenges presented by tables with different structures.  Additionally, it establishes a process for aligning out-of-domain data with the specific target task through a \"learn, annotate, and refinement\" pipeline.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The core concept behind MediTab, involving the consolidation, enrichment, and refinement modules, is well-founded in its aim to improve the scalability of predictive models designed for medical tabular data.\n2. Implementing a sanity check through Large Language Model (LLM) Reflection is particularly important in the medical domain.\n3. The paper is clearly written; the quality is sound.\n4. The empirical evaluation in the paper is strong; relevant baselines are considered.\n5. The coverage of related work is extensive, with clear distinctions drawn from other studies.", "weaknesses": "1. It is not clear, how was splitting into train/val/test organised?\n2. The statement “After the quality check step, we obtain the original task dataset $T$ and the supplementary dataset $T_{sup}$ and have two potential options for model training. The first is to combine both datasets for training, but we have found that this approach results in suboptimal performance.” is a bit unclear. Why it is suboptimal, could authors elaborate on this?\n3. Are these datasets prone to missing values, which is a critical concern in the medical domain? If so, what would be the recommended strategy for handling these missing values?\n4. Results on Ablation studies on Different Learning strategies are provided. Could authors provide  Ablation studies on the different model components?", "questions": "1. Could the authors elaborate more on the LLM sanity check, as well as the results and tests provided in Appendix C.4-C.5? It has been discussed that it is crucial to conduct thorough evaluations of LLMs in healthcare, with particular attention to aspects of safety, equity, and bias [a]. Could the authors provide their thoughts on why they believe their model satisfies these requirements?\n2. Could authors provide more detailed explanation of empirical studies and address points in Weakness section?\n\na. Singhal et al., Large language models encode clinical knowledge. Nature 620, 172–180 (2023).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698939218014}, {"id": "DZBAvGIDc9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission988/Reviewer_R9Te"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The manuscript proposes a framework utilizing LLM to perform alignment between different datasets for the same and different tasks. The framework prompts ChatGPT to summarize each row in a table into text and utilizes BioBERT as a classification model that takes text as input. Moreover, it trains an init model to annotate data from datasets of other tasks and clean such data with Shapley scores into supplementary data samples.", "review_text": "The manuscript proposes a framework utilizing LLM to perform alignment between different datasets for the same and different tasks. The framework prompts ChatGPT to summarize each row in a table into text and utilizes BioBERT as a classification model that takes text as input. Moreover, it trains an init model to annotate data from datasets of other tasks and clean such data with Shapley scores into supplementary data samples.", "strengths": "- The paper is well-written and easy to follow.", "weaknesses": "- The framework is quite straightforward, and there is not much technical contribution. It is mostly a combination of multiple existing models. And the idea of transferring tabular data into text is not novel at all. There are a bunch of existing works [1][2][3], including one of their baselines TabLLM[4]. The further incorporation of text information from samples from other datasets is just one trivial step forward. Furthermore, [4] actually proved that a template for transferring the tabular data works better than an LLM. Yet, in this paper, there is no comparison for such serialization methods.\n\n- The author didn’t specify what exact features are included in these experimental datasets. Also, it is unclear how many columns are overlapped between different datasets. Yet, if there is a large portion of feature overlapping, maybe simple concatenation and removing or recoding of the missing columns will work just as well. There is no discussion regarding this whatsoever.\n\n- The step 2 in section 2.2 is confusing:\n    - The authors claimed that they used active learning in step 2. Is the “active learning pipeline” method the same as traditional active learning that select informative samples to label? If not, the description can mislead the readers.\n    - The authors claimed that they cleaned supplementary dataset T_{1, sup} with a data audit module based on data Shapley scores. More experiments are expected to demonstrate the effectiveness of the audit module. Moreover, it would be better if the authors conducted more ablation studies to show whether the supplementary dataset improve the prediction performance. \n\n- The datasets in Table 1 contain less than 3000 patients. It is very easy for the LLMs (e.g., BioBERT) to overfit the training set. It is unclear how the authors prevent overfitting during the fine-tuning phase.\n\n- In Table 3, the proposed MediTab exhibits the capability to access multiple datasets during its training, in contrast to the other baseline models, which are constrained to employing a single dataset. This discrepancy in data utilization introduces an element of unfairness in the comparison. It would be more appropriate to conduct a comparison against models that have undergone training on multiple datasets. For instance, TabLLM, being a large language model, can readily undertake multi-dataset training with minor adjustments to its data preprocessing procedures. Therefore, a more equitable comparison would involve evaluating MediTab and TabLLM under identical conditions, both in the context of training on a single dataset and across multiple datasets. \n\n- Most medical data, like MIMIC-IV, includes timestamp information of the patients’ multiple visits or collections. This framework completely ignores this part of the medical data, which limits their application to real-world clinical environments.\n\nReference:\n1. Bertsimas, Dimitris & Carballo, Kimberly & Ma, Yu & Na, Liangyuan & Boussioux, Léonard & Zeng, Cynthia & Soenksen, Luis & Fuentes, Ignacio. (2022). TabText: a Systematic Approach to Aggregate Knowledge Across Tabular Data Structures. 10.48550/arXiv.2206.10381.\n2. Yin, Pengcheng & Neubig, Graham & Yih, Wen-tau & Riedel, Sebastian. TaBERT: Pretraining for Joint Understanding of Textual and Tabular Data. ACL 2020.\n3. Li, Y., Li, J., Suhara, Y., Doan, A., and Tan, W.-C. (2020). Deep entity matching with pre-trained language models. Proc. VLDB Endow., 14(1):50–60.\n4. Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. Tabllm: Few-shot classification of tabular data with large language models. arXiv preprint arXiv:2210.10723, 2022.", "questions": "1. All questions in the above section.\n\n2. Are there any overlaps of columns between the tabular data for the same tasks? Is it hard to do a simple concatenation? What’s the traditional method for dealing with the missing columns? Are they applicable to this situation?\n\n3. For the choice of BioBERT and the QA model for salinity check, the author did not provide a reason for choosing these models.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The manuscript proposes a framework utilizing LLM to perform alignment between different datasets for the same and different tasks. The framework prompts ChatGPT to summarize each row in a table into text and utilizes BioBERT as a classification model that takes text as input. Moreover, it trains an init model to annotate data from datasets of other tasks and clean such data with Shapley scores into supplementary data samples.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "- The paper is well-written and easy to follow.", "weaknesses": "- The framework is quite straightforward, and there is not much technical contribution. It is mostly a combination of multiple existing models. And the idea of transferring tabular data into text is not novel at all. There are a bunch of existing works [1][2][3], including one of their baselines TabLLM[4]. The further incorporation of text information from samples from other datasets is just one trivial step forward. Furthermore, [4] actually proved that a template for transferring the tabular data works better than an LLM. Yet, in this paper, there is no comparison for such serialization methods.\n\n- The author didn’t specify what exact features are included in these experimental datasets. Also, it is unclear how many columns are overlapped between different datasets. Yet, if there is a large portion of feature overlapping, maybe simple concatenation and removing or recoding of the missing columns will work just as well. There is no discussion regarding this whatsoever.\n\n- The step 2 in section 2.2 is confusing:\n    - The authors claimed that they used active learning in step 2. Is the “active learning pipeline” method the same as traditional active learning that select informative samples to label? If not, the description can mislead the readers.\n    - The authors claimed that they cleaned supplementary dataset T_{1, sup} with a data audit module based on data Shapley scores. More experiments are expected to demonstrate the effectiveness of the audit module. Moreover, it would be better if the authors conducted more ablation studies to show whether the supplementary dataset improve the prediction performance. \n\n- The datasets in Table 1 contain less than 3000 patients. It is very easy for the LLMs (e.g., BioBERT) to overfit the training set. It is unclear how the authors prevent overfitting during the fine-tuning phase.\n\n- In Table 3, the proposed MediTab exhibits the capability to access multiple datasets during its training, in contrast to the other baseline models, which are constrained to employing a single dataset. This discrepancy in data utilization introduces an element of unfairness in the comparison. It would be more appropriate to conduct a comparison against models that have undergone training on multiple datasets. For instance, TabLLM, being a large language model, can readily undertake multi-dataset training with minor adjustments to its data preprocessing procedures. Therefore, a more equitable comparison would involve evaluating MediTab and TabLLM under identical conditions, both in the context of training on a single dataset and across multiple datasets. \n\n- Most medical data, like MIMIC-IV, includes timestamp information of the patients’ multiple visits or collections. This framework completely ignores this part of the medical data, which limits their application to real-world clinical environments.\n\nReference:\n1. Bertsimas, Dimitris & Carballo, Kimberly & Ma, Yu & Na, Liangyuan & Boussioux, Léonard & Zeng, Cynthia & Soenksen, Luis & Fuentes, Ignacio. (2022). TabText: a Systematic Approach to Aggregate Knowledge Across Tabular Data Structures. 10.48550/arXiv.2206.10381.\n2. Yin, Pengcheng & Neubig, Graham & Yih, Wen-tau & Riedel, Sebastian. TaBERT: Pretraining for Joint Understanding of Textual and Tabular Data. ACL 2020.\n3. Li, Y., Li, J., Suhara, Y., Doan, A., and Tan, W.-C. (2020). Deep entity matching with pre-trained language models. Proc. VLDB Endow., 14(1):50–60.\n4. Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. Tabllm: Few-shot classification of tabular data with large language models. arXiv preprint arXiv:2210.10723, 2022.", "questions": "1. All questions in the above section.\n\n2. Are there any overlaps of columns between the tabular data for the same tasks? Is it hard to do a simple concatenation? What’s the traditional method for dealing with the missing columns? Are they applicable to this situation?\n\n3. For the choice of BioBERT and the QA model for salinity check, the author did not provide a reason for choosing these models.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698806529740}, {"id": "xgnzn9c3QG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission988/Reviewer_bkYv"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work proposes an approach for developing learning prediction models using heterogenous tabular data sources and schemas and across tasks. The approach relies on large language models to represent structured data in natural language as a common representation across contexts, aggregate datasets from like and unlike tasks and populations, and do zero and few-shot prediction. The approach is applied to several medical data (primarily clinical trials with some retrospective observational data). The approach generally outperforms fully-supervised baselines and further performs well in zero and few-shot settings.", "review_text": "This work proposes an approach for developing learning prediction models using heterogenous tabular data sources and schemas and across tasks. The approach relies on large language models to represent structured data in natural language as a common representation across contexts, aggregate datasets from like and unlike tasks and populations, and do zero and few-shot prediction. The approach is applied to several medical data (primarily clinical trials with some retrospective observational data). The approach generally outperforms fully-supervised baselines and further performs well in zero and few-shot settings.", "strengths": "* The method appears to yield performant predictive models across datasets and tasks, and without requiring significant amounts of labeled target data. \n* The approach to representing and harmonizing structured data in natural language using LLMs is general and plausibly would continue to work well outside of the context evaluated in this work.", "weaknesses": "* I have several concerns related to clarity and lack of detail given to some important aspects of the methodology and experiments. These are elaborated on in the Questions section below.\n* The datasets chosen are relatively small and relatively low-dimensional (e.g. ~10s of fields at most). An area where this method might be useful is with tabular data of much greater dimensionality, as is typical in healthcare contexts.", "questions": "* How are the “External Patient Databases” used (MIMIC-IV and PMC-Patients)? Are they used as supplementary databases during the psuedolabeling step? \n* How is the MIMIC-IV data processed? The information in Table 1 that shows that there are only 2 categorical, 1 binary, and 1 numerical feature in MIMIC-IV. As MIMIC-IV is much richer (potentially thousands of features) it is unclear which components of the database are actually used and no details are provided.\n* In section 2.4, why is the initial model trained on all available training data from $T_1$ considered a multi-task model (designated by $f_{MTL}$)? If I understand correctly, this model is trained on one task, but several datasets.\n* The description of the psuedolabeling step is not entirely clear to me. Is the idea to take the initial model for the target task, make predictions for the target task on data collected for other tasks, and then use those predictions as pseudo-labels for further training? This seems peculiar because it is not clear that this should fundamentally improve performance for the target task given that the pseudo-labels are essentially just predictions of the target label derived from information in the target task database(s).\n* If available, it would be relevant to compare to baselines that pool over datasets with rule-based schema harmonization. For example, in the context of electronic health records and claims data, there are standards such as the OMOP Common Data Model that provide the means of mapping data from disparate sources to a shared schema.\n* An ablation experiment that removes the auditing steps (both the LLM sanity check and the Data Shapley checks) and the pseudolabeling step would help gain insight into the marginal value that they provide, especially as they are positioned as the novel methodological contributions of this work relative to TabLLM (if I understand correctly).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes an approach for developing learning prediction models using heterogenous tabular data sources and schemas and across tasks. The approach relies on large language models to represent structured data in natural language as a common representation across contexts, aggregate datasets from like and unlike tasks and populations, and do zero and few-shot prediction. The approach is applied to several medical data (primarily clinical trials with some retrospective observational data). The approach generally outperforms fully-supervised baselines and further performs well in zero and few-shot settings.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "* The method appears to yield performant predictive models across datasets and tasks, and without requiring significant amounts of labeled target data. \n* The approach to representing and harmonizing structured data in natural language using LLMs is general and plausibly would continue to work well outside of the context evaluated in this work.", "weaknesses": "* I have several concerns related to clarity and lack of detail given to some important aspects of the methodology and experiments. These are elaborated on in the Questions section below.\n* The datasets chosen are relatively small and relatively low-dimensional (e.g. ~10s of fields at most). An area where this method might be useful is with tabular data of much greater dimensionality, as is typical in healthcare contexts.", "questions": "* How are the “External Patient Databases” used (MIMIC-IV and PMC-Patients)? Are they used as supplementary databases during the psuedolabeling step? \n* How is the MIMIC-IV data processed? The information in Table 1 that shows that there are only 2 categorical, 1 binary, and 1 numerical feature in MIMIC-IV. As MIMIC-IV is much richer (potentially thousands of features) it is unclear which components of the database are actually used and no details are provided.\n* In section 2.4, why is the initial model trained on all available training data from $T_1$ considered a multi-task model (designated by $f_{MTL}$)? If I understand correctly, this model is trained on one task, but several datasets.\n* The description of the psuedolabeling step is not entirely clear to me. Is the idea to take the initial model for the target task, make predictions for the target task on data collected for other tasks, and then use those predictions as pseudo-labels for further training? This seems peculiar because it is not clear that this should fundamentally improve performance for the target task given that the pseudo-labels are essentially just predictions of the target label derived from information in the target task database(s).\n* If available, it would be relevant to compare to baselines that pool over datasets with rule-based schema harmonization. For example, in the context of electronic health records and claims data, there are standards such as the OMOP Common Data Model that provide the means of mapping data from disparate sources to a shared schema.\n* An ablation experiment that removes the auditing steps (both the LLM sanity check and the Data Shapley checks) and the pseudolabeling step would help gain insight into the marginal value that they provide, especially as they are positioned as the novel methodological contributions of this work relative to TabLLM (if I understand correctly).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698526565416}], "openreview_url": "https://openreview.net/forum?id=F0XXA9OG13", "arxiv_id": "2305.12081", "paper_pdf": "papers/F0XXA9OG13.pdf", "paper_pdf_sha256": "59152de65b6ad2f9d76f5ec54b658093b75fde93a179fc7f5f99d932e34e5ffc", "paper_pdf_bytes": 1110037, "paper_pdf_source": "openreview", "code_url": "https://github.com/RyanWangZf/MediTab", "code_repository": "RyanWangZf/MediTab", "code_commit": "e50db96e97efa9ec594e9bcb99cbf3b9140ca166", "code_archive": "repos/F0XXA9OG13.zip", "code_archive_sha256": "7be0da25b967d2f2d45911a5bd64ff197f5cdfdb21ec1684ab6df6dd42e06232", "code_archive_bytes": 125975, "code_file_count": 21, "code_extensions": {".py": 13, ".ipynb": 8}, "github_disk_usage_kb": 109, "github_languages": {"Jupyter Notebook": 412132, "Python": 119504}, "github_archived": false, "github_pushed_at": "2024-05-08T04:59:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/anypredict-foundation-model-for-tabular"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PHtzmXK8am", "year": 2023, "status": "rejected", "title": "TAN without a burn: Scaling laws of DP-SGD", "authors": ["Tom Sander", "Pierre Stock", "Alexandre Sablayrolles"], "authorids": ["~Tom_Sander1", "~Pierre_Stock1", "~Alexandre_Sablayrolles1"], "authors_source": "OpenReview API", "abstract": "Differentially Private methods for training Deep Neural Networks (DNNs) have progressed recently, in particular with the use of massive batches and aggregated data augmentations for a large number of steps. These techniques require much more compute than their non-private counterparts, shifting the traditional privacy-accuracy trade-off to a privacy-accuracy-compute trade-off and making hyper-parameter search virtually impossible for realistic scenarios. In this work, we decouple privacy analysis and experimental behavior of noisy training to explore the trade-off with minimal computational requirements. We first use the tools of Rényi Differential Privacy (RDP) to show that the privacy budget, when not overcharged, only depends on the total amount of noise (TAN) injected throughout training. We then derive scaling laws for training models with DP-SGD to optimize hyper-parameters with more than a $100\\times$ reduction in computational budget. We apply the proposed method on CIFAR-10 and ImageNet and, in particular, strongly improve the state-of-the-art on ImageNet with a $+9$ points gain in accuracy for a privacy budget $\\varepsilon=8$.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "7NC5gPCt6b", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3020/Reviewer_6jSv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors argue that our ability to optimize a model trained with DP-SGD is governed by a quantity they call TAN (the total amount of noise), specifically $\\Sigma^2 = 2 \\sigma^2 N^2/(B^2 S)$, where $\\Sigma$ is the TAN, $\\sigma$ is the scale of the added Gaussian noise, $N$ is the dataset size, $B$ is the batch size and $S$ is the number of updates.\n\nAdditionally, they show empirically that when $\\sigma \\gtrsim 2$, the privacy parameter $\\epsilon$ becomes a simple function of $\\Sigma$ and $\\delta$. Since it is standard to set $\\delta \\approx 1/N$, in practice this means that the privacy is determined solely by $\\Sigma$, the TAN. Meanwhile, the privacy loss $\\epsilon$ is very large when $\\sigma \\ll 2$.\n\nThe authors propose to exploit this observation by tuning hyper-parameters with small $\\sigma$, small batch size $B$ and large privacy loss $\\epsilon$, before extrapolating those hyper-parameters to $\\sigma \\gtrsim 2$ while increasing the batch size to ensure that the TAN $\\Sigma$ is constant. This significantly reduces the privacy loss $\\epsilon$ however it also significantly increases the compute cost of training. Extrapolating hyper-parameters from cheap non-private training runs to expensive runs with tight privacy guarantees reduces the total computation required to achieve strong performance with DP.", "review_text": "On the positive side, the authors identify an important and difficult problem, and provide some useful insights. However on the downside TAN is not a reliable measure of performance in practice, and the authors only achieve strong results on one dataset so it is not clear how reliable the methodology is.\n\nI have scored the paper weak accept, however I would encourage the authors to improve the clarity of the writing and to empirically explore in more detail how test accuracy depends on (batch size, step budget, added noise) under the constraint of constant TAN.", "strengths": "Strengths:\n1) Integrating the constraints of DP-SGD with optimization theory, in order to reliably predict good hyper-parameter choices, is one of the most important problems that needs to be solved in order to make differentially private deep learning practical. I'm pleased to see the authors tackle this difficult and important problem.\n2) The metric proposed by the authors is intuitive and it appears to provide some helpful intuition.\n3) The authors achieve strong performance when training from scratch on ImageNet with DP-SGD.\n\nWeaknesses:\n1) If TAN were an accurate measure of optimizability, then we should expect the train accuracy to be roughly constant at fixed TAN. However in practice, the authors find in figure 1 that the test accuracy decays log-linearly at fixed TAN as the batch size rises. I feel that the current paper is quite confusing because this discrepancy is not explicitly stated anywhere. I'd encourage the authors to make this explicit, and to explore in more detail why this discrepancy arises (eg comparing test vs train performance). \n\n2) As discussed above, Figure 1 shows that TAN over-estimates the performance of large batch sizes. It would be nice to also see a figure at fixed batch size, sweeping the number of steps S, to see whether TAN under/over-estimates the performance of small/large step budgets.\n\n3) As a minor point, note that on some truly sensitive data we cannot train non-private models (even internally). It would be good to acknowledge that the hyper-parameter transfer process described here cannot be used in these cases.\n\n4) For many of the experiments, key hyper-parameters are lifted from prior work (eg step budget, target batch size). Is it not possible to directly infer good settings for these hyper-parameters using the TAN framework?\n\n5) The paper essentially only considers ImageNet (with some very limited experiments on CIFAR-10). Since very few groups have run experiments at this scale it is not clear how challenging the baseline from De et al. is in practice.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors argue that our ability to optimize a model trained with DP-SGD is governed by a quantity they call TAN (the total amount of noise), specifically $\\Sigma^2 = 2 \\sigma^2 N^2/(B^2 S)$, where $\\Sigma$ is the TAN, $\\sigma$ is the scale of the added Gaussian noise, $N$ is the dataset size, $B$ is the batch size and $S$ is the number of updates.\n\nAdditionally, they show empirically that when $\\sigma \\gtrsim 2$, the privacy parameter $\\epsilon$ becomes a simple function of $\\Sigma$ and $\\delta$. Since it is standard to set $\\delta \\approx 1/N$, in practice this means that the privacy is determined solely by $\\Sigma$, the TAN. Meanwhile, the privacy loss $\\epsilon$ is very large when $\\sigma \\ll 2$.\n\nThe authors propose to exploit this observation by tuning hyper-parameters with small $\\sigma$, small batch size $B$ and large privacy loss $\\epsilon$, before extrapolating those hyper-parameters to $\\sigma \\gtrsim 2$ while increasing the batch size to ensure that the TAN $\\Sigma$ is constant. This significantly reduces the privacy loss $\\epsilon$ however it also significantly increases the compute cost of training. Extrapolating hyper-parameters from cheap non-private training runs to expensive runs with tight privacy guarantees reduces the total computation required to achieve strong performance with DP.", "strength_and_weaknesses": "Strengths:\n1) Integrating the constraints of DP-SGD with optimization theory, in order to reliably predict good hyper-parameter choices, is one of the most important problems that needs to be solved in order to make differentially private deep learning practical. I'm pleased to see the authors tackle this difficult and important problem.\n2) The metric proposed by the authors is intuitive and it appears to provide some helpful intuition.\n3) The authors achieve strong performance when training from scratch on ImageNet with DP-SGD.\n\nWeaknesses:\n1) If TAN were an accurate measure of optimizability, then we should expect the train accuracy to be roughly constant at fixed TAN. However in practice, the authors find in figure 1 that the test accuracy decays log-linearly at fixed TAN as the batch size rises. I feel that the current paper is quite confusing because this discrepancy is not explicitly stated anywhere. I'd encourage the authors to make this explicit, and to explore in more detail why this discrepancy arises (eg comparing test vs train performance). \n\n2) As discussed above, Figure 1 shows that TAN over-estimates the performance of large batch sizes. It would be nice to also see a figure at fixed batch size, sweeping the number of steps S, to see whether TAN under/over-estimates the performance of small/large step budgets.\n\n3) As a minor point, note that on some truly sensitive data we cannot train non-private models (even internally). It would be good to acknowledge that the hyper-parameter transfer process described here cannot be used in these cases.\n\n4) For many of the experiments, key hyper-parameters are lifted from prior work (eg step budget, target batch size). Is it not possible to directly infer good settings for these hyper-parameters using the TAN framework?\n\n5) The paper essentially only considers ImageNet (with some very limited experiments on CIFAR-10). Since very few groups have run experiments at this scale it is not clear how challenging the baseline from De et al. is in practice.\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\nI found the paper difficult to follow. I think this is because the authors do not clearly acknowledge that, if TAN were an accurate measure of trainability, then test accuracy should not fall as batch size rises (at constant TAN) in figure 1. It is also not always clear which hyper-parameters are extrapolated using the TAN framework, and which are lifted from prior work or tuned in the large batch/low privacy loss regime.\n\nQuality:\nI think the underlying quality of the work is quite high (if the clarity can be improved).\n\nNovelty:\nThis is the first paper I am aware of to provide an explicit toy model predicting the trainability of DP-SGD. Although TAN has some flaws, I think this is a novel and worthwhile research direction.\n\nReproducibility:\nI'm not sure how easy it would be to reproduce the results", "summary_of_the_review": "On the positive side, the authors identify an important and difficult problem, and provide some useful insights. However on the downside TAN is not a reliable measure of performance in practice, and the authors only achieve strong results on one dataset so it is not clear how reliable the methodology is.\n\nI have scored the paper weak accept, however I would encourage the authors to improve the clarity of the writing and to empirically explore in more detail how test accuracy depends on (batch size, step budget, added noise) under the constraint of constant TAN.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666796721144}, {"id": "znFh_byyNT6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3020/Reviewer_FyxV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed to use the total amount of noise (TAN) to determine the privacy budget in Renyi DP. They observe scaling laws with TAN for DP-SGD which can be used to reduce the computational cost for hyper-parameter tuning. By the hyper-parameter tuning, the authors greatly improve the state-of-the-art on neural network training under privacy guarantees.", "review_text": "This paper demonstrates the scaling law between the batch size and the test accuracy in private deep learning. This law and the new concept, TAN, are very useful in improving private learning results by DP-SGD when the computation resources are limited. This law is found in the results of ImageNet and CIFAR-10 experiments, but the reason behind it is still unknown and it may be difficult to reproduce the results since the code is not provided. ", "strengths": "Major strengths:\n1. The observation of the scaling law is new. \n2. The results of hyperparameter tuning improve the state-of-the-art greatly.\n\nMinor strengths:\n1. This paper is easy to follow.\n2. The visualization of the experimental results is good.\n3. Figure 2 seems to deliver the same message as the linear part of Figure 3 (left) in (Li et al, 2021a), but the authors provide more explanations in Section 3.2.\n\nMajor weaknesses:\n1. The privacy guarantee parameter $\\varepsilon_{TAN}$ is an approximation of $\\varepsilon_{RDP}$ and it is only a lower bound. For strict privacy guarantees, we need an upper bound.\n2. There is no explanation for the reason for the scaling law. \n\nMinor weaknesses:\nSome of the statements in this paper are inconsistent with each other. \n1. In the left figure of Figure 4, the legend shows that $\\sigma_{ref}$ is constant for each line, but the caption says the $\\eta_{step}$ is constant. I also wonder why the authors choose to fix different values in Figure 4 for CIFAR-10 and ImageNet. Is it because the results of CIFAR-10 do not look as linear as of ImageNet?\n2. In Page 5, Choice of $\\sigma$, the authors claimed that if $\\sigma>4$, dividing it by 2 with halving q is likely to improve performance, but I cannot find experimental results supporting this.\n3. In Page 7, the authors say that they decide not to use the gain from 'testing with augmentations' in Table 1 and Table 4. In Table 2, I guess the Total column is the sum of the other columns for B=128, ..., 1024. Therefore, the last row, B=16384, also has the AugTest (+0.8) in the Total result (6.7%). I guess in Table 4, the test ACC for B=16384, 36.9%, is derived by adding the total improvement 6.7% to the baseline 30.2%. Therefore, I am not sure if the authors forget to remove the improvement from AugTest.\n4. In Page 9, in the last paragraph, the authors believe that the exponential increase in the privacy budget $\\varepsilon$ as the noise level $\\sigma$ decreases. I guess this is from Figure 2, but the x-axis for $\\sigma$ is log-scale. Therefore, it is not easy to identify whether there is a linear or exponential relationship between $\\varepsilon$ and $\\sigma$.\n\nBy the way, there are duplicate references (Li et al, 2021a and 2021b).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors proposed to use the total amount of noise (TAN) to determine the privacy budget in Renyi DP. They observe scaling laws with TAN for DP-SGD which can be used to reduce the computational cost for hyper-parameter tuning. By the hyper-parameter tuning, the authors greatly improve the state-of-the-art on neural network training under privacy guarantees.", "strength_and_weaknesses": "Major strengths:\n1. The observation of the scaling law is new. \n2. The results of hyperparameter tuning improve the state-of-the-art greatly.\n\nMinor strengths:\n1. This paper is easy to follow.\n2. The visualization of the experimental results is good.\n3. Figure 2 seems to deliver the same message as the linear part of Figure 3 (left) in (Li et al, 2021a), but the authors provide more explanations in Section 3.2.\n\nMajor weaknesses:\n1. The privacy guarantee parameter $\\varepsilon_{TAN}$ is an approximation of $\\varepsilon_{RDP}$ and it is only a lower bound. For strict privacy guarantees, we need an upper bound.\n2. There is no explanation for the reason for the scaling law. \n\nMinor weaknesses:\nSome of the statements in this paper are inconsistent with each other. \n1. In the left figure of Figure 4, the legend shows that $\\sigma_{ref}$ is constant for each line, but the caption says the $\\eta_{step}$ is constant. I also wonder why the authors choose to fix different values in Figure 4 for CIFAR-10 and ImageNet. Is it because the results of CIFAR-10 do not look as linear as of ImageNet?\n2. In Page 5, Choice of $\\sigma$, the authors claimed that if $\\sigma>4$, dividing it by 2 with halving q is likely to improve performance, but I cannot find experimental results supporting this.\n3. In Page 7, the authors say that they decide not to use the gain from 'testing with augmentations' in Table 1 and Table 4. In Table 2, I guess the Total column is the sum of the other columns for B=128, ..., 1024. Therefore, the last row, B=16384, also has the AugTest (+0.8) in the Total result (6.7%). I guess in Table 4, the test ACC for B=16384, 36.9%, is derived by adding the total improvement 6.7% to the baseline 30.2%. Therefore, I am not sure if the authors forget to remove the improvement from AugTest.\n4. In Page 9, in the last paragraph, the authors believe that the exponential increase in the privacy budget $\\varepsilon$ as the noise level $\\sigma$ decreases. I guess this is from Figure 2, but the x-axis for $\\sigma$ is log-scale. Therefore, it is not easy to identify whether there is a linear or exponential relationship between $\\varepsilon$ and $\\sigma$.\n\nBy the way, there are duplicate references (Li et al, 2021a and 2021b).", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written. The statements in the paper are clear and concise. The novelty of this paper is good (see the strengths). The code is not provided, and the computation cost of the comparison experiments in this paper seems to be high; therefore the reproducibility is low.", "summary_of_the_review": "This paper demonstrates the scaling law between the batch size and the test accuracy in private deep learning. This law and the new concept, TAN, are very useful in improving private learning results by DP-SGD when the computation resources are limited. This law is found in the results of ImageNet and CIFAR-10 experiments, but the reason behind it is still unknown and it may be difficult to reproduce the results since the code is not provided. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666694311835}, {"id": "DFJvK7iLLI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3020/Reviewer_AFbM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\n\nThe paper studies how to transfer the hyperparameters of differentially private training based on low-cost search. It shows the privacy budget only depends on the total amount of noise (TAN) injected throughout training. They derives a scaling law for training models with DP-SGD with more than a 100 times reduction in computational budget.", "review_text": "How to tune DP-SGD is both theoretically and practically important for private learning. The paper makes a clear contribution towards tuning DP-SGD in principle. The concluding solution is solid and well-supported. I recommend its acceptance.\n\n\\#### After rebuttal and discussion with other reviewers \\#####\nAs discussed with other reviewers, the relation of TAN with privacy accounting has been derived in literature e.g., Lemma 2 and Proposition 3 of Bun et al. 2020. The paper is not fully aware of the literature, which undermines the novelty of the contribution. I change the score to 6.", "strengths": "Strength: It studies an important problem of how to scale DP-SGD with hyper-parameter search. The paper provides a neat and simple relationship between the privacy budget and the other parameters, i.e., the noise multiplier, steps and the sampling rate. This relationship motivates a way of scaling the batchsize and the noise multiplier without affecting the privacy accountant.\n\nWeakness: \n1. There is no application of using TAN relation in language task, either pretraining or fine-tuning.\n\n2. The reference needs to cover more. Especially, Yu et al. 2021 starts the fine-tuning large language models with differential privacy and low-rank reparametrization. \nYu et al. 2021 Large Scale Private Learning via Low-rank Reparametrization.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\n\nThe paper studies how to transfer the hyperparameters of differentially private training based on low-cost search. It shows the privacy budget only depends on the total amount of noise (TAN) injected throughout training. They derives a scaling law for training models with DP-SGD with more than a 100 times reduction in computational budget.", "strength_and_weaknesses": "Strength: It studies an important problem of how to scale DP-SGD with hyper-parameter search. The paper provides a neat and simple relationship between the privacy budget and the other parameters, i.e., the noise multiplier, steps and the sampling rate. This relationship motivates a way of scaling the batchsize and the noise multiplier without affecting the privacy accountant.\n\nWeakness: \n1. There is no application of using TAN relation in language task, either pretraining or fine-tuning.\n\n2. The reference needs to cover more. Especially, Yu et al. 2021 starts the fine-tuning large language models with differential privacy and low-rank reparametrization. \nYu et al. 2021 Large Scale Private Learning via Low-rank Reparametrization.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written, and the finding is novel and useful for tuning DP-SGD.", "summary_of_the_review": "How to tune DP-SGD is both theoretically and practically important for private learning. The paper makes a clear contribution towards tuning DP-SGD in principle. The concluding solution is solid and well-supported. I recommend its acceptance.\n\n\\#### After rebuttal and discussion with other reviewers \\#####\nAs discussed with other reviewers, the relation of TAN with privacy accounting has been derived in literature e.g., Lemma 2 and Proposition 3 of Bun et al. 2020. The paper is not fully aware of the literature, which undermines the novelty of the contribution. I change the score to 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666625958770}, {"id": "O6bNfUSGkMc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3020/Reviewer_MS1Z"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "Recent work shows that large batch size can give better privacy-utility trade-off in differentia private ML training. However, training with large batch sizes is computationally expensive. The paper observes (empirically) that the privacy budget of differential private ML training depends mostly on the total amount of noise added during training when the noise multiplier \\sigma>=2. At the same time, the accuracy of the final results can be predicted by the results on a corresponding hyper-parameter setting in which the batch size is small. Following these observations, the paper proposes to do hyper-parameter tuning for small batch sizes (which is computationally efficient) and then transform the hyper-parameters to the target large batch size. Experiments demonstrate that this approach works well in practice.", "review_text": "\nOverall, the paper is well-written and the observations and the proposed approach can potentially be very useful for the community. It would be even better if the authors can demonstrate/verify the observation on more network architectures, datasets, and models (e.g., generative models).", "strengths": "\nStrength:\n\n* The observations (batch scaling law, privacy budget computation) in the paper are new and interesting.\n\n* The proposed hyper-parameter tuning process can be very useful in practice, as it saves the computation requirement and therefore speeds up the development process.\n\nWeaknesses:\n\n* Overall, the writing is clear, but there are still some issues that need to be addressed:\n    - Equation 2: many of the notations here are not defined, e.g., clip_C, l, theta, C.\n    - Figure 4: B_ref, S_ref, \\sigma_ref are not defined. I can guess their meaning from the context, but all notations should be defined explicitly in the text.\n    - Page 5: it says \"Simultaneously doubling the batch size and \\sigma has a negligible or small impact on accuracy (Figure 4)\". However, in Figure 4(b), the impact on the accuracy is not small. Indeed, this is claimed in the next sentence \"Reciprocally, if \\sigma > 4, dividing it by 2 simultaneously with halving q is likely to improve performance.\" These two sentences contradict each other.\n    - Table 1: I guess \"8,8.\" should be \"8,\"?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Recent work shows that large batch size can give better privacy-utility trade-off in differentia private ML training. However, training with large batch sizes is computationally expensive. The paper observes (empirically) that the privacy budget of differential private ML training depends mostly on the total amount of noise added during training when the noise multiplier \\sigma>=2. At the same time, the accuracy of the final results can be predicted by the results on a corresponding hyper-parameter setting in which the batch size is small. Following these observations, the paper proposes to do hyper-parameter tuning for small batch sizes (which is computationally efficient) and then transform the hyper-parameters to the target large batch size. Experiments demonstrate that this approach works well in practice.", "strength_and_weaknesses": "\nStrength:\n\n* The observations (batch scaling law, privacy budget computation) in the paper are new and interesting.\n\n* The proposed hyper-parameter tuning process can be very useful in practice, as it saves the computation requirement and therefore speeds up the development process.\n\nWeaknesses:\n\n* Overall, the writing is clear, but there are still some issues that need to be addressed:\n    - Equation 2: many of the notations here are not defined, e.g., clip_C, l, theta, C.\n    - Figure 4: B_ref, S_ref, \\sigma_ref are not defined. I can guess their meaning from the context, but all notations should be defined explicitly in the text.\n    - Page 5: it says \"Simultaneously doubling the batch size and \\sigma has a negligible or small impact on accuracy (Figure 4)\". However, in Figure 4(b), the impact on the accuracy is not small. Indeed, this is claimed in the next sentence \"Reciprocally, if \\sigma > 4, dividing it by 2 simultaneously with halving q is likely to improve performance.\" These two sentences contradict each other.\n    - Table 1: I guess \"8,8.\" should be \"8,\"?\n\n", "clarity,_quality,_novelty_and_reproducibility": "\nClarity: The paper is overall well-written and clear.\n\nQuality: The observation and the proposed approach are interesting and useful, and the claims are supported by experiments. Overall the quality is good.\n\nReproducibility: the code of the experiments is not provided.\n\nNovelty: The observation is new and the proposed approach is novel.\n", "summary_of_the_review": "\nOverall, the paper is well-written and the observations and the proposed approach can potentially be very useful for the community. It would be even better if the authors can demonstrate/verify the observation on more network architectures, datasets, and models (e.g., generative models).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666573091504}], "openreview_url": "https://openreview.net/forum?id=PHtzmXK8am", "arxiv_id": "2210.03403", "paper_pdf": "papers/PHtzmXK8am.pdf", "paper_pdf_sha256": "e430534137b455b33f6428275508339aac8c3d1aaa9acdb983569e6cdf841ea4", "paper_pdf_bytes": 1390439, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/tan", "code_repository": "facebookresearch/tan", "code_commit": "9d8be82b62bdffaabb4a01f0f2938b6e026018bc", "code_archive": "repos/PHtzmXK8am.zip", "code_archive_sha256": "0f46d203914d5a044b4af8d9b55b00d958b70110c5304289d48a73bbaa683122", "code_archive_bytes": 210808, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 199, "github_languages": {"Python": 205589}, "github_archived": true, "github_pushed_at": "2025-01-07T19:37:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tan-without-a-burn-scaling-laws-of-dp-sgd"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EXe93Md8RqS", "year": 2022, "status": "rejected", "title": "Data Quality Matters For Adversarial Training: An Empirical Study", "authors": ["Chengyu Dong", "Liyuan Liu", "Jingbo Shang"], "authorids": ["~Chengyu_Dong1", "~Liyuan_Liu3", "~Jingbo_Shang2"], "authors_source": "OpenReview API", "abstract": "Multiple intriguing problems are hovering in adversarial training, including robust overfitting, robustness overestimation, and robustness-accuracy trade-off. These problems pose great challenges to both reliable evaluation and practical deployment. Here, we empirically show that these problems share one common cause --- low-quality samples in the dataset. Specifically, we first propose a strategy to measure the data quality based on the learning behaviors of the data during adversarial training and find that low-quality data may not be useful and even detrimental to the adversarial robustness. We then design controlled experiments to investigate the interconnections between data quality and problems in adversarial training. We find that when low-quality data is removed, robust overfitting and robustness overestimation can be largely alleviated; and robustness-accuracy trade-off becomes less significant. These observations not only verify our intuition about data quality but may also open new opportunities to advance adversarial training. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "OBfkOafsATy", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2429/Reviewer_VxQb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the effect of data quality on adversarial robustness. Specifically, they focus on one measure of data quality (number of times there is a perturbation that is misclassified across training iterations). They study the effect of data quality on robust overfitting, robustness-accuracy tradeoffs and \"robustness overestimation\" (gap between strong and weak attacks). The main conclusions reported are that data quality as measured by their metric plays an important role in all three aspects, and a suggested takeaway is that we need data of higher quality to improve robustness. ", "review_text": "(Edited score post rebuttal)\n\nStrengths: \n(i) This paper studies an important problem on the effect of data quality on different aspects of robust training. \n(ii) The paper starts with the intuitive idea that data quality should matter and attempts to make a systematic empirical investigation across multiple datasets. \n(iii) Overall, the paper is easy to read and the experiment details/results are clearly presented. \n(iv) The observation that different sets of points improve robust vs standard accuracy is quite interesting to me (though that wasn't the focus of the paper). \n\nWeaknesses: While the paper makes an interesting investigation, I worry that the conclusions/implications of the study are unclear and some inferences seem problematic to draw. \n(i) One claim the paper makes is that low quality data can be detrimental to robust accuracy. Concretely, if we remove points from the dataset, that would improve robustness. Looking at Figure 4, the cases where this happens are: CIFAR-100 with TRADES and Tiny Imagenet with TRADES. I am confused why the conclusions vary between TRADES and Adversarial training? How is the \\beta parameter set in the TRADES experiments. The fact that TRADES and adversarial training differ in their trends suggests that the beta parameter that controls accuracy vs robustness is important. The paper mentions that the hardness is an \"intrinsic\" measure since the relative ordering doesn't vary much across methods. So is it true that removing the *same* points leads to different performance with TRADES and adversarial training?\n(ii) Other than the (slightly questionable) claim that removing low quality data improves performance, I didn't find the other implications on robustness particularly surprising. Even in standard training, it has been shown that there are some \"support vectors\" (or prototypes) which effectively dictate the final model learnt, and the rest of the dataset does not matter as much. The paper seems to mainly replicate this for robust training. \n(iii) Computing the data quality measure: If I understand correctly, you need to compute the robust accuracy across the training checkpoints. Which attack do you use to compute this? Does it matter how strong the attack is?\n(iv) What's an actionable takeaway from this work? Aside from the caveat of point (i) above on whether there is a reliable takeaway, I wonder how to actually operationalize the idea of using higher quality data? The data quality of each point seems to depend on the remaining points (since it comes from some ordering of robust accuracy), and computing this requires training the model on all the data. Can the authors please comment with any thoughts here?\n(iv) Computing data quality for extra data: Can you use checkpoints trained on CIFAR-10 to evaluate data quality of additional data (for e.g. from Carmon et al. 2019) not present in the training set? Basically, this would involve computing the robustness at the new point (not in training) across different training epochs. If the authors could show that this improves performance over adding all the mined extra points, I would be convinced that there is something interesting here. Relatedly, how does this measure of data quality empirically differ from just logits for confidence (which was what was used to mine the extra data in Carmon et al. 2019). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the effect of data quality on adversarial robustness. Specifically, they focus on one measure of data quality (number of times there is a perturbation that is misclassified across training iterations). They study the effect of data quality on robust overfitting, robustness-accuracy tradeoffs and \"robustness overestimation\" (gap between strong and weak attacks). The main conclusions reported are that data quality as measured by their metric plays an important role in all three aspects, and a suggested takeaway is that we need data of higher quality to improve robustness. ", "main_review": "(Edited score post rebuttal)\n\nStrengths: \n(i) This paper studies an important problem on the effect of data quality on different aspects of robust training. \n(ii) The paper starts with the intuitive idea that data quality should matter and attempts to make a systematic empirical investigation across multiple datasets. \n(iii) Overall, the paper is easy to read and the experiment details/results are clearly presented. \n(iv) The observation that different sets of points improve robust vs standard accuracy is quite interesting to me (though that wasn't the focus of the paper). \n\nWeaknesses: While the paper makes an interesting investigation, I worry that the conclusions/implications of the study are unclear and some inferences seem problematic to draw. \n(i) One claim the paper makes is that low quality data can be detrimental to robust accuracy. Concretely, if we remove points from the dataset, that would improve robustness. Looking at Figure 4, the cases where this happens are: CIFAR-100 with TRADES and Tiny Imagenet with TRADES. I am confused why the conclusions vary between TRADES and Adversarial training? How is the \\beta parameter set in the TRADES experiments. The fact that TRADES and adversarial training differ in their trends suggests that the beta parameter that controls accuracy vs robustness is important. The paper mentions that the hardness is an \"intrinsic\" measure since the relative ordering doesn't vary much across methods. So is it true that removing the *same* points leads to different performance with TRADES and adversarial training?\n(ii) Other than the (slightly questionable) claim that removing low quality data improves performance, I didn't find the other implications on robustness particularly surprising. Even in standard training, it has been shown that there are some \"support vectors\" (or prototypes) which effectively dictate the final model learnt, and the rest of the dataset does not matter as much. The paper seems to mainly replicate this for robust training. \n(iii) Computing the data quality measure: If I understand correctly, you need to compute the robust accuracy across the training checkpoints. Which attack do you use to compute this? Does it matter how strong the attack is?\n(iv) What's an actionable takeaway from this work? Aside from the caveat of point (i) above on whether there is a reliable takeaway, I wonder how to actually operationalize the idea of using higher quality data? The data quality of each point seems to depend on the remaining points (since it comes from some ordering of robust accuracy), and computing this requires training the model on all the data. Can the authors please comment with any thoughts here?\n(iv) Computing data quality for extra data: Can you use checkpoints trained on CIFAR-10 to evaluate data quality of additional data (for e.g. from Carmon et al. 2019) not present in the training set? Basically, this would involve computing the robustness at the new point (not in training) across different training epochs. If the authors could show that this improves performance over adding all the mined extra points, I would be convinced that there is something interesting here. Relatedly, how does this measure of data quality empirically differ from just logits for confidence (which was what was used to mine the extra data in Carmon et al. 2019). \n", "summary_of_the_review": "This paper studies an interesting question about the effect of data quality on robustness. Unfortunately, I find the conclusions for robustness as emphasized in the paper either unsubstantiated (removing low quality data improves robustness) or difficult to make actionable. Nevertheless, there are some interesting empirical observations in the paper such as the difference between important points for standard and robust accuracy, the effect of data quality on robustness-accuracy tradeoffs which were not known previously. I apologize for the late review - there was some issue at my end and I realized the review didn't get submitted. I would really appreciate if the authors could respond to my questions above. Thanks!", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1637090664083}, {"id": "oy7S6KRkWQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2429/Reviewer_Rp8o"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "**Few sentences summary**: the paper proposes an empirical study on how data quality impacts robust performance in the Lp-norm setting under three angles: robust overfitting (i.e. the difference in accuracy between the best and last checkpoints), robustness evaluation (i.e. consistency of the robust performance when trying various robust evaluations like AutoAttack) and clean/robust accuracy trade-off (i.e. the gap in performance between the clean and robust accuracies). They show that \"high quality\" data (selected with their proposed criterion) performs better than \"low quality\" data for these three performance metrics.\n\nRegarding the **results** and **contributions**:\n* The authors propose a quantitative definition of \"quality\"  based on the average training robust accuracy over epochs per sample.\n* They show that \"low quality\" samples are better for standard training whereas \"high quality\" samples are better for adversarial training.\n* Studies on 1) robust overfitting, 2) robustness evaluation and 3) clean/robust accuracy trade-off on the datasets CIFAR-10/100 and TinyImageNet for Adversarial Training and TRADES.", "review_text": "**Strengths**:\n* The paper has a clear message that adversarial training and nominal training do not benefit from the same type of samples.\n* Extensive experiments on CIFAR-10 to analyze the robust performance of \"high quality\" data (selected by their proposed criterion) under three angles commonly discussed in the Lp-norm community. \n* The paper is well written and very clear.\n\n**Weaknesses/Suggestions/Questions**:\n1) **Main Weakness**. It really lacks of an analysis in the extra data case. This case should be actually perfectly suited for the message of the paper as extra data for CIFAR-10 is known to be of lesser quality (as per the references in the paper). Similarly, recent works [1, 2] have proposed to use generated data in the Lp-norm case. The proposed quality criterion in the paper would be great to analyze the quality of the generated images compared to the original CIFAR-10 images.\n2) In Figure 3, \"with their labels predicted by a model\", which model? This is a bit vague.\n3) Figures 4 to 7. First, please put in the caption how the robust test accuracy is obtained. Against PGD-10, PGD-40? It is unclear. Second, all the first rows are labeled \"PGD\". I think you mean \"Adversarial Training\" as introduced by [3]. PGD is the optimization procedure and it also used on TRADES. So the naming \"PGD\" is misleading.\n4) Typo: \"can also be found in our Figure 1(a)\" -> \"can also be found in our Figure 1(c)\" on page 3.\n\n**References**:\n1) Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, Timothy Mann. *Fixing data augmentation to improve adversarial robustness*. 2021.\n2) Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, Prateek Mittal. *Improving adversarial robustness using proxy distributions*. 2021.\n3) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu. *Towards deep learning models resistant to adversarial attacks*. 2017.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "**Few sentences summary**: the paper proposes an empirical study on how data quality impacts robust performance in the Lp-norm setting under three angles: robust overfitting (i.e. the difference in accuracy between the best and last checkpoints), robustness evaluation (i.e. consistency of the robust performance when trying various robust evaluations like AutoAttack) and clean/robust accuracy trade-off (i.e. the gap in performance between the clean and robust accuracies). They show that \"high quality\" data (selected with their proposed criterion) performs better than \"low quality\" data for these three performance metrics.\n\nRegarding the **results** and **contributions**:\n* The authors propose a quantitative definition of \"quality\"  based on the average training robust accuracy over epochs per sample.\n* They show that \"low quality\" samples are better for standard training whereas \"high quality\" samples are better for adversarial training.\n* Studies on 1) robust overfitting, 2) robustness evaluation and 3) clean/robust accuracy trade-off on the datasets CIFAR-10/100 and TinyImageNet for Adversarial Training and TRADES.", "main_review": "**Strengths**:\n* The paper has a clear message that adversarial training and nominal training do not benefit from the same type of samples.\n* Extensive experiments on CIFAR-10 to analyze the robust performance of \"high quality\" data (selected by their proposed criterion) under three angles commonly discussed in the Lp-norm community. \n* The paper is well written and very clear.\n\n**Weaknesses/Suggestions/Questions**:\n1) **Main Weakness**. It really lacks of an analysis in the extra data case. This case should be actually perfectly suited for the message of the paper as extra data for CIFAR-10 is known to be of lesser quality (as per the references in the paper). Similarly, recent works [1, 2] have proposed to use generated data in the Lp-norm case. The proposed quality criterion in the paper would be great to analyze the quality of the generated images compared to the original CIFAR-10 images.\n2) In Figure 3, \"with their labels predicted by a model\", which model? This is a bit vague.\n3) Figures 4 to 7. First, please put in the caption how the robust test accuracy is obtained. Against PGD-10, PGD-40? It is unclear. Second, all the first rows are labeled \"PGD\". I think you mean \"Adversarial Training\" as introduced by [3]. PGD is the optimization procedure and it also used on TRADES. So the naming \"PGD\" is misleading.\n4) Typo: \"can also be found in our Figure 1(a)\" -> \"can also be found in our Figure 1(c)\" on page 3.\n\n**References**:\n1) Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, Timothy Mann. *Fixing data augmentation to improve adversarial robustness*. 2021.\n2) Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, Prateek Mittal. *Improving adversarial robustness using proxy distributions*. 2021.\n3) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu. *Towards deep learning models resistant to adversarial attacks*. 2017.", "summary_of_the_review": "Empirical study paper with an interesting message that adversarial training and nominal training do not benefit from the same type of samples. The paper could have gone more in depth by studying the extra data case which is commonly used in the literature and gets the SOTA results. Especially, this extra data case normally fits the message of the paper so it is a missing element in the paper.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635877438907}, {"id": "2yTTrzYRWS1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2429/Reviewer_soLM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a metric for evaluating the learning stability of data and point out that unstably-learned instances are of low-quality for adversarial training. Through extensive controlled experiments, this paper investigates the impact of low-quality data on three issues in adversarial training, i.e., robust overfitting, robustness overestimation, and robustness-accuracy trade-off. The experimental results show that removing the low-quality instances can mitigate the issues.", "review_text": "Strengths:\n\n1.\tThis paper evaluates the learning stability of instances under various settings, which indicates that the estimation of quality for each data is almost accurate and consistent.\n\n2.\tThis paper is well organized. The authors analyze the influences of low-quality data on the three issues separately, which empirically demonstrates removing the low-quality data can mitigate the issues.\n\nWeakness:\n\n1.\tCalculating the quality of data is time-consuming and computationally expensive. The quality of data is closely related to a particular dataset. That is to say, when the dataset is updated such as adding some new data into the dataset, the quality of data in the new dataset needs to be recalculated. It should be a limitation for the practical use of low-quality data.\n\n2.\tRemoving the low-quality data sometimes marginally improves adversarial robustness, but more often hurts the adversarial robustness and natural generalization. Therefore, the significance of the proposed low-quality data seems minor.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a metric for evaluating the learning stability of data and point out that unstably-learned instances are of low-quality for adversarial training. Through extensive controlled experiments, this paper investigates the impact of low-quality data on three issues in adversarial training, i.e., robust overfitting, robustness overestimation, and robustness-accuracy trade-off. The experimental results show that removing the low-quality instances can mitigate the issues.", "main_review": "Strengths:\n\n1.\tThis paper evaluates the learning stability of instances under various settings, which indicates that the estimation of quality for each data is almost accurate and consistent.\n\n2.\tThis paper is well organized. The authors analyze the influences of low-quality data on the three issues separately, which empirically demonstrates removing the low-quality data can mitigate the issues.\n\nWeakness:\n\n1.\tCalculating the quality of data is time-consuming and computationally expensive. The quality of data is closely related to a particular dataset. That is to say, when the dataset is updated such as adding some new data into the dataset, the quality of data in the new dataset needs to be recalculated. It should be a limitation for the practical use of low-quality data.\n\n2.\tRemoving the low-quality data sometimes marginally improves adversarial robustness, but more often hurts the adversarial robustness and natural generalization. Therefore, the significance of the proposed low-quality data seems minor.\n", "summary_of_the_review": "Overall, this paper systematically study the influences of data quality on the three problems in adversarial training. However, I am skeptical about the practical values of the proposed metric for data quality.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635516439615}, {"id": "YNieYIy79db", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2429/Reviewer_mNG6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper investigates the impact of data quality, measured by the proportion of epochs in which the model classifies a specific input correctly during training, on the robustness, generalization, and robustness-accuracy tradeoff of adversarially trained models.  Unlike with standard training, the authors find that more difficult inputs (lower quality inputs) can hurt adversarially trained models.  They find that compared to randomly removing data during training, removing low quality data can lead to higher robustness, less robust overfitting, less robustness overestimation, and less robustness-accuracy tradeoff.", "review_text": "Strengths:\n- Interesting direction- I find this work very exciting since it contributes to our understanding of adversarial training.  Many works have demonstrated that adversarial training requires more data.  This work expands along this direction by observing how the quality of data impacts generalization of adversarial training.\n- Good scope in experiments, significant results- The authors perform experiments on multiple datasets and adversarial training methods and the trends look consistent across methods.  Patterns discovered between data quality and properties like robustness, robust overfitting, and robustness estimation also look significant.\n\nWeaknesses:\n- In the main text, it say that results for WideResNet architecture are in the appendix, but when I looked at the appendix I couldn't find figures for WRN corresponding to the experiments in the main text.  I think adding some figures for WRN with gradually removing lowest quality vs random would also help show the generalizability of the observed trends.\n- In section 3.2 related works, the authors state that prior works show that more data can be detrimental to adversarial training, but I don't agree that is the case.  I think prior works demonstrate that adversarial training benefits a lot from additional data.  For instance, Gowal et al. 2020 is cited to support the statement that additional data unrelated to the original dataset can hurt adversarial robustness, but Gowal et al. demonstrate that additional data always improves over the no additional data case.  There is also discussion of label noise in this section, but it isn't clear how label noise is connected to the idea of additional data and data quality.  Additionally, I think discussion of other works which measure notions of hardness and data quality for adversarial training can also be added to related works.  For instance, methods of measuring difficulty of adversarial examples used for curriculum adversarial training [1,2,3].  Recently there has also been a line of works discussing ways to measure the data quality of additional data generated by generative models for use with adversarial training [4, 5].\n\n\n[1] Wang, Yisen, et al. \"On the Convergence and Robustness of Adversarial Training.\" ICML. Vol. 1. 2019.\n[2] Zhang, Jingfeng, et al. \"Attacks which do not kill training make adversarial learning stronger.\" International Conference on Machine Learning. PMLR, 2020.\n[3] Cai, Qi-Zhi, Chang Liu, and Dawn Song. \"Curriculum adversarial training.\" Proceedings of the 27th International Joint Conference on Artificial Intelligence. 2018.\n[4] Sehwag, Vikash, et al. \"Robust Learning Meets Generative Models: Can Proxy Distributions Improve Adversarial Robustness?\" arXiv preprint arXiv:2104.09425 (2021)\n[5] Gowal, Sven, et al. \"Improving Robustness using Generated Data.\" arXiv preprint arXiv:2110.09468 (2021).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper investigates the impact of data quality, measured by the proportion of epochs in which the model classifies a specific input correctly during training, on the robustness, generalization, and robustness-accuracy tradeoff of adversarially trained models.  Unlike with standard training, the authors find that more difficult inputs (lower quality inputs) can hurt adversarially trained models.  They find that compared to randomly removing data during training, removing low quality data can lead to higher robustness, less robust overfitting, less robustness overestimation, and less robustness-accuracy tradeoff.", "main_review": "Strengths:\n- Interesting direction- I find this work very exciting since it contributes to our understanding of adversarial training.  Many works have demonstrated that adversarial training requires more data.  This work expands along this direction by observing how the quality of data impacts generalization of adversarial training.\n- Good scope in experiments, significant results- The authors perform experiments on multiple datasets and adversarial training methods and the trends look consistent across methods.  Patterns discovered between data quality and properties like robustness, robust overfitting, and robustness estimation also look significant.\n\nWeaknesses:\n- In the main text, it say that results for WideResNet architecture are in the appendix, but when I looked at the appendix I couldn't find figures for WRN corresponding to the experiments in the main text.  I think adding some figures for WRN with gradually removing lowest quality vs random would also help show the generalizability of the observed trends.\n- In section 3.2 related works, the authors state that prior works show that more data can be detrimental to adversarial training, but I don't agree that is the case.  I think prior works demonstrate that adversarial training benefits a lot from additional data.  For instance, Gowal et al. 2020 is cited to support the statement that additional data unrelated to the original dataset can hurt adversarial robustness, but Gowal et al. demonstrate that additional data always improves over the no additional data case.  There is also discussion of label noise in this section, but it isn't clear how label noise is connected to the idea of additional data and data quality.  Additionally, I think discussion of other works which measure notions of hardness and data quality for adversarial training can also be added to related works.  For instance, methods of measuring difficulty of adversarial examples used for curriculum adversarial training [1,2,3].  Recently there has also been a line of works discussing ways to measure the data quality of additional data generated by generative models for use with adversarial training [4, 5].\n\n\n[1] Wang, Yisen, et al. \"On the Convergence and Robustness of Adversarial Training.\" ICML. Vol. 1. 2019.\n[2] Zhang, Jingfeng, et al. \"Attacks which do not kill training make adversarial learning stronger.\" International Conference on Machine Learning. PMLR, 2020.\n[3] Cai, Qi-Zhi, Chang Liu, and Dawn Song. \"Curriculum adversarial training.\" Proceedings of the 27th International Joint Conference on Artificial Intelligence. 2018.\n[4] Sehwag, Vikash, et al. \"Robust Learning Meets Generative Models: Can Proxy Distributions Improve Adversarial Robustness?\" arXiv preprint arXiv:2104.09425 (2021)\n[5] Gowal, Sven, et al. \"Improving Robustness using Generated Data.\" arXiv preprint arXiv:2110.09468 (2021).", "summary_of_the_review": "I vote to accept this paper since I find the empirical contributions significant; the authors rigorously study the impact of data quality on various properties of adversarially trained models including robustness, robust overfitting and robustness overestimation, and robustness-accuracy tradeoff.  I think the discussion of related works can be improved though.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635220951494}], "openreview_url": "https://openreview.net/forum?id=EXe93Md8RqS", "arxiv_id": "2102.07437", "paper_pdf": "papers/EXe93Md8RqS.pdf", "paper_pdf_sha256": "c780d3f082d0422127e2eb0a56ce626327639cf9593c3d302b90c6ecd192ec02", "paper_pdf_bytes": 1478760, "paper_pdf_source": "openreview", "code_url": "https://github.com/shwinshaker/RobustDataProfiling", "code_repository": "shwinshaker/RobustDataProfiling", "code_commit": "7f29c2ea5b304771bd6d1d123fbb89633317ee5f", "code_archive": "repos/EXe93Md8RqS.zip", "code_archive_sha256": "0c9f3e216a54ec532ec0378aeba9c868d2dc1834e5dcbae24563bd55c1590c78", "code_archive_bytes": 398973, "code_file_count": 33, "code_extensions": {".py": 32, ".sh": 1}, "github_disk_usage_kb": 418, "github_languages": {"Python": 241372, "Shell": 1383}, "github_archived": false, "github_pushed_at": "2021-02-24T06:32:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/data-profiling-for-adversarial-training-on"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tv8n52XbO4p", "year": 2021, "status": "rejected", "title": "Learning to Generate Noise for Multi-Attack Robustness", "authors": ["Divyam Madaan", "Jinwoo Shin", "Sung Ju Hwang"], "authorids": ["~Divyam_Madaan1", "~Jinwoo_Shin1", "~Sung_Ju_Hwang1"], "authors_source": "OpenReview API", "abstract": "Adversarial learning has emerged as one of the successful techniques to circumvent the susceptibility of existing methods against adversarial perturbations. However, the majority of existing defense methods are tailored to defend against a single category of adversarial perturbation (e.g. $\\ell_\\infty$-attack). In safety-critical applications, this makes these methods extraneous as the attacker can adopt diverse adversaries to deceive the system. Moreover, training on multiple perturbations simultaneously significantly increases the computational overhead during training. To address these challenges, we propose a novel meta-learning framework that explicitly learns to generate noise to improve the model's robustness against multiple types of attacks. Its key component is Meta Noise Generator (MNG) that outputs optimal noise to stochastically perturb a given sample, such that it helps lower the error on diverse adversarial perturbations. By utilizing samples generated by MNG, we train a model by enforcing the label consistency across multiple perturbations. We validate the robustness of models trained by our scheme on various datasets and against a wide variety of perturbations, demonstrating that it significantly outperforms the baselines across multiple perturbations with a marginal computational cost.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JN8SQP5aVg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3710/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a novel meta-learning framework that explicitly learns to generate noise to improve model robustness (against multiple types of attacks). The results indicate that the proposed approach improves on the state-of-the-art.\n\nOverall, the paper is well written. However some details are missing and this could make the paper hard to reproduce. The experiments could be expanded.\n\n1) There is a significant amount of work about using generative models to build adversarial examples. The literature review only focuses on classical adversarial robustness and robustness against multiple adversaries. I'd recommend making a review of these approaches, even if they are orthogonal to the one proposed in this paper (e.g., [1,2,3])\n2) In Eq. (6), what is \\mathcal{B}(x, \\epsilon). Since there is multiple threat models, I am assuming that it is selected at random between l_1, l_2 and l_inf (like SAT).\n3) The number of inner steps T seems to be critical (as it will trade-off gradient precision with compute). However, I don't see any study on this in the paper. Also, it is not clear which value was used for the experiments.\n4) Looking at Eq. (7), it seems like backpropagation through the T inner steps is necessary to compute the gradients w.r.t. \\phi. This seems overly expensive and I find surprising that adv_avg and adv_max take so much longer to train.\n5) Concerning Eq. (7), as a curiousity, have authors considered implicit differentiation [4] ?\n6) The experiments are run using 30 epochs which is rather on slim side. E.g., RST_inf should reach about 59% robust accuracy with 200 epochs of training (with 30 epochs it only reaches 55%). I'm curious as to whether the comparison with the proposed approach is unfair (e.g., Adv_inf sees a single adv example per batch, whereas MNG-AC sees 2).\n7) It's not entirely clear to me why beta negatively affects l_2 robustness. I'd assume that if the model was only trained against l_2, then there might be an optimal value for beta that is different that the one from Fig. 2. In general, it would be interesting to see on MNG-AC does if different subsets of threats are used.\n8) The l_2 loss landscapes seem more noisy that what they should be. Also it's unclear why the axes are centered for l_inf and not for l_2 (explain how these are generated).\n9) In Table 5, MNG-AC achieves 35.1% against all l_inf attacks, but only 33.7% against AutoAttack. Am I missing something?\n\nDetails:\n\nA) It would helpful to the reader to have the epsilon values written on top of the different tables. The captions could be expanded to include more details.\nB) Visuaization -> Visualization\n\n[1] https://openreview.net/pdf?id=SJeQEp4YDH: GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification\n[2] https://arxiv.org/pdf/1801.02610: Generating Adversarial Examples with Adversarial Networks\n[3] https://arxiv.org/pdf/1710.10766: PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples\n[4] https://arxiv.org/pdf/1911.02590: Optimizing Millions of Hyperparameters by Implicit Differentiation", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Initial review", "review": "In this paper, the authors propose a novel meta-learning framework that explicitly learns to generate noise to improve model robustness (against multiple types of attacks). The results indicate that the proposed approach improves on the state-of-the-art.\n\nOverall, the paper is well written. However some details are missing and this could make the paper hard to reproduce. The experiments could be expanded.\n\n1) There is a significant amount of work about using generative models to build adversarial examples. The literature review only focuses on classical adversarial robustness and robustness against multiple adversaries. I'd recommend making a review of these approaches, even if they are orthogonal to the one proposed in this paper (e.g., [1,2,3])\n2) In Eq. (6), what is \\mathcal{B}(x, \\epsilon). Since there is multiple threat models, I am assuming that it is selected at random between l_1, l_2 and l_inf (like SAT).\n3) The number of inner steps T seems to be critical (as it will trade-off gradient precision with compute). However, I don't see any study on this in the paper. Also, it is not clear which value was used for the experiments.\n4) Looking at Eq. (7), it seems like backpropagation through the T inner steps is necessary to compute the gradients w.r.t. \\phi. This seems overly expensive and I find surprising that adv_avg and adv_max take so much longer to train.\n5) Concerning Eq. (7), as a curiousity, have authors considered implicit differentiation [4] ?\n6) The experiments are run using 30 epochs which is rather on slim side. E.g., RST_inf should reach about 59% robust accuracy with 200 epochs of training (with 30 epochs it only reaches 55%). I'm curious as to whether the comparison with the proposed approach is unfair (e.g., Adv_inf sees a single adv example per batch, whereas MNG-AC sees 2).\n7) It's not entirely clear to me why beta negatively affects l_2 robustness. I'd assume that if the model was only trained against l_2, then there might be an optimal value for beta that is different that the one from Fig. 2. In general, it would be interesting to see on MNG-AC does if different subsets of threats are used.\n8) The l_2 loss landscapes seem more noisy that what they should be. Also it's unclear why the axes are centered for l_inf and not for l_2 (explain how these are generated).\n9) In Table 5, MNG-AC achieves 35.1% against all l_inf attacks, but only 33.7% against AutoAttack. Am I missing something?\n\nDetails:\n\nA) It would helpful to the reader to have the epsilon values written on top of the different tables. The captions could be expanded to include more details.\nB) Visuaization -> Visualization\n\n[1] https://openreview.net/pdf?id=SJeQEp4YDH: GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification\n[2] https://arxiv.org/pdf/1801.02610: Generating Adversarial Examples with Adversarial Networks\n[3] https://arxiv.org/pdf/1710.10766: PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples\n[4] https://arxiv.org/pdf/1911.02590: Optimizing Millions of Hyperparameters by Implicit Differentiation", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603980842430}, {"id": "uwcwhu4VVoj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3710/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n=======\nThe authors propose a number of techniques to learn models which are adversarially robust to multiple perturbations. These involve a noise generator, a loss to enforce consistency, as well as a stochastic variant of adversarial training. With these changes, they are able to produce improvements to robust accuracy to multiple perturbation types. \n\n\nOverall, I get the idea and the empirical results seem promising. However, the structure and writing of the paper is at times rather confusing, and there are a lot of missing details. If the code were not supplied, it would be difficult in the current state to reproduce the method from the paper. Perhaps due to this, the specifics of the key component, the meta noise generator, are still rather opaque to me. Perhaps the authors can clarify, and I am happy to follow up afterwards. \n\nComments for discussion\n=======================\nThe majority of my confusion lies in section 4, for the specifics of the meta noise generator and parts of the algorithm in general. I am otherwise well acquainted with the relevant literature. \n\n1) Augmented examples (x_aug) are generated by adding noise from the MNG and projecting it onto some ball B. It is not clear to me what ball this is since the authors are considering multiple perturbations. Is it a random type? Or a joint projection? I assume it is at least one of the perturbations being considered, or is that incorrect? \n\n2) Similarly, in the algorithm, the authors generate adversarial examples (x_adv) by sampling a random attack. I could not find what set of attacks were being sampled from, or what the sampling distribution is (I checked the appendix as well). \n\n3) The generator is apparently updated to minimize the classifier loss on the adversarial examples as written in Equation (8). However, the adversarial examples are generated from some unspecified set of attacks, which implies that the set of attacks actually depends on the generator somehow. Is this supposed to be the classifier loss on the augmented samples? If not, then how do the adversarial examples depend on the generator? \n\n4) The consistency loss involves clean, adversarial, and augmented posterior distributions. There are no details on these distributions: are these simply the softmax of the logits? Or is a generative model that outputs a distribution being used? \n\n5) On a more fundamental level, what is the motivation behind training the generator to minimize the classifier loss? Why would we want to do this over random sampling? What's to prevent a degenerate solution of simply learning to produce a zero perturbation (and thus always producing clean examples, which can achieve low loss)? \n\n\nMinor comments\n==============\nI have checked the supplementary material and the authors have included the code for running their experiments. Ideally, this would also include pre-trained model weights. \n\nUpdate\n======\nAfter much effort, I can say that I understand the paper. The edits appear to have incorporated all the identified missing information. I have thus updated my confidence and slightly improved my score, however I am not confident that the current presentation of the approach will be understandable by a reader without contacting the authors, given that the difficulty I had in understanding the paper (and my initial confidence) stemmed primarily from missing information and poor presentation for the approach. Although the results do seem to improve upon past work, its impact will suffer if it is difficult to understand for a non-reviewer reader. I would be more confident if a fresh set of eyes could understand the details of the work without having to go to the authors to clarify so many details. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Promising results, but method is not clear", "review": "Summary\n=======\nThe authors propose a number of techniques to learn models which are adversarially robust to multiple perturbations. These involve a noise generator, a loss to enforce consistency, as well as a stochastic variant of adversarial training. With these changes, they are able to produce improvements to robust accuracy to multiple perturbation types. \n\n\nOverall, I get the idea and the empirical results seem promising. However, the structure and writing of the paper is at times rather confusing, and there are a lot of missing details. If the code were not supplied, it would be difficult in the current state to reproduce the method from the paper. Perhaps due to this, the specifics of the key component, the meta noise generator, are still rather opaque to me. Perhaps the authors can clarify, and I am happy to follow up afterwards. \n\nComments for discussion\n=======================\nThe majority of my confusion lies in section 4, for the specifics of the meta noise generator and parts of the algorithm in general. I am otherwise well acquainted with the relevant literature. \n\n1) Augmented examples (x_aug) are generated by adding noise from the MNG and projecting it onto some ball B. It is not clear to me what ball this is since the authors are considering multiple perturbations. Is it a random type? Or a joint projection? I assume it is at least one of the perturbations being considered, or is that incorrect? \n\n2) Similarly, in the algorithm, the authors generate adversarial examples (x_adv) by sampling a random attack. I could not find what set of attacks were being sampled from, or what the sampling distribution is (I checked the appendix as well). \n\n3) The generator is apparently updated to minimize the classifier loss on the adversarial examples as written in Equation (8). However, the adversarial examples are generated from some unspecified set of attacks, which implies that the set of attacks actually depends on the generator somehow. Is this supposed to be the classifier loss on the augmented samples? If not, then how do the adversarial examples depend on the generator? \n\n4) The consistency loss involves clean, adversarial, and augmented posterior distributions. There are no details on these distributions: are these simply the softmax of the logits? Or is a generative model that outputs a distribution being used? \n\n5) On a more fundamental level, what is the motivation behind training the generator to minimize the classifier loss? Why would we want to do this over random sampling? What's to prevent a degenerate solution of simply learning to produce a zero perturbation (and thus always producing clean examples, which can achieve low loss)? \n\n\nMinor comments\n==============\nI have checked the supplementary material and the authors have included the code for running their experiments. Ideally, this would also include pre-trained model weights. \n\nUpdate\n======\nAfter much effort, I can say that I understand the paper. The edits appear to have incorporated all the identified missing information. I have thus updated my confidence and slightly improved my score, however I am not confident that the current presentation of the approach will be understandable by a reader without contacting the authors, given that the difficulty I had in understanding the paper (and my initial confidence) stemmed primarily from missing information and poor presentation for the approach. Although the results do seem to improve upon past work, its impact will suffer if it is difficult to understand for a non-reviewer reader. I would be more confident if a fresh set of eyes could understand the details of the work without having to go to the authors to clarify so many details. ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603866848159}, {"id": "uNiWR-V7oXo", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3710/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses a timely issue in adversarial robustness - efficient training of robust models against multiple adversarial perturbations. The authors propose a combination of three techniques: stochastic adversarial training (SAT), meta noise generator (MNG), and adversarial consistency (AC) loss for efficient training, and evaluate the robustness using multiple L1, L2, and Linf norm-bounded attacks and three datasets (CIFAR-10, SVHN, and Tiny Imagenet). The results show improved multi-attack robustness over several baselines (including single-attack and multiple-attack models) and reduced training time. Ablation studies are also performed to illustrate the utility of each component of the proposed model. Overall, this paper provides very detailed evaluations involving multiple datasets, attacks, baselines, and robustness metrics. I find the results convincing and important, and also find sufficient novelty in the proposed training method.\n\nThe strengths (S) and weaknesses (W) of this submission are summarized below.\n\nS1. The proposal of MNG and AC is effective and novel.\nS2. The evaluation is thorough and convincing.\nS3. The proposal improves both robustness and training efficiency in most cases.\n\nW1. The adversarial consistency (AC) loss is never defined explicitly. Based on equation (5), it is hard to understand how AC \"represents the Jensen-Shannon Divergence (JSD) among the posterior distributions\" when considering three distributions, P_clean, P_adv, and P_aug. More clarification is needed.\n\nW2. Although the results show improved multi-attack robustness, it will be great if the authors can add more intuition on why the proposed training method leads to performance improvement. Based on the ablation study,  it seems that the role of SAT and MNG is to reduce overfitting in robustness to encourage generalization, rather than optimization over the worst-case scenarios.\n\nW3. The considered multi-attack setting is still limited to different Lp norm perturbation constraints. Although the authors showed improved robustness over unforeseen attacks, the authors should also discuss how the proposed method can generalize to different attacks beyond Lp norms.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Thorough Multi-attack Robustness Evaluation and Clever Adversarial Training ", "review": "This paper addresses a timely issue in adversarial robustness - efficient training of robust models against multiple adversarial perturbations. The authors propose a combination of three techniques: stochastic adversarial training (SAT), meta noise generator (MNG), and adversarial consistency (AC) loss for efficient training, and evaluate the robustness using multiple L1, L2, and Linf norm-bounded attacks and three datasets (CIFAR-10, SVHN, and Tiny Imagenet). The results show improved multi-attack robustness over several baselines (including single-attack and multiple-attack models) and reduced training time. Ablation studies are also performed to illustrate the utility of each component of the proposed model. Overall, this paper provides very detailed evaluations involving multiple datasets, attacks, baselines, and robustness metrics. I find the results convincing and important, and also find sufficient novelty in the proposed training method.\n\nThe strengths (S) and weaknesses (W) of this submission are summarized below.\n\nS1. The proposal of MNG and AC is effective and novel.\nS2. The evaluation is thorough and convincing.\nS3. The proposal improves both robustness and training efficiency in most cases.\n\nW1. The adversarial consistency (AC) loss is never defined explicitly. Based on equation (5), it is hard to understand how AC \"represents the Jensen-Shannon Divergence (JSD) among the posterior distributions\" when considering three distributions, P_clean, P_adv, and P_aug. More clarification is needed.\n\nW2. Although the results show improved multi-attack robustness, it will be great if the authors can add more intuition on why the proposed training method leads to performance improvement. Based on the ablation study,  it seems that the role of SAT and MNG is to reduce overfitting in robustness to encourage generalization, rather than optimization over the worst-case scenarios.\n\nW3. The considered multi-attack setting is still limited to different Lp norm perturbation constraints. Although the authors showed improved robustness over unforeseen attacks, the authors should also discuss how the proposed method can generalize to different attacks beyond Lp norms.\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603859349109}, {"id": "iF-e5SkNBvh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3710/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. Summary\n\nThe authors propose a new method to improve robustness to adversarial examples under various norms (L1, L2 and LInf). Their method combines adversarial training with an adversarial noise generator. They improve upon adversarial training in a multi norm setting by choosing one norm at random for each sample, instead of computing an adversarial for all norms, thus significantly reducing the training time. They additionally improve robustness by regularizing model features between the standard image, the adversarially perturbed image and a perturbation of the image created with an adversarial noise generator.\n\n\n2. Strengths\n+ The method is based on adversarial training. As far as I know and as the authors note this is the only method that reliably leads to more robust models.\n+ The authors attack their models with  a range of attacks that to the best of my knowledge are state-of-the art.\n+ The method apparently works in the multi norm setting.\n\n3. Weaknesses \n- I was missing an intuitive description why the adversarial noise should improve robustness to adversarial attacks. I was only aware of it as a method to improve corruption robustness.\n- I was not always sure if I got everything correctly in sections 4 and 5.3. I think I got it but I sometimes missed a figure. It may e.g. be helpful to include the losses in Figure 1 or make a separate figure. Especially why the MNG was trained the way it is was a bit unclear for me.\n\n\n4. Recommendation\n\nI think this paper is an accept but as I don't work with adversarial examples I am not at all confident in that assessment. From the discussions with people who work on adversarial examples new defenses are usually broken very quickly and there is a number of papers which break numerous defenses. The method is however based on adversarial training which to my knowledge is the only robust method so far and the used attacks seem valid. So I am definitely leaning towards accept but the opinion of a real expert would be highly appreciated as I feel not at all qualified to assess the validity of papers on adversarial examples.\n\n\n5. Questions/Recommendations\n- Is there a difference between the M(eta)NG and the A(dversarial)NG from Rusak et. al. 2020?\n\n\n6. Additional feedback \n- None as the paper is pretty well written.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "1. Summary\n\nThe authors propose a new method to improve robustness to adversarial examples under various norms (L1, L2 and LInf). Their method combines adversarial training with an adversarial noise generator. They improve upon adversarial training in a multi norm setting by choosing one norm at random for each sample, instead of computing an adversarial for all norms, thus significantly reducing the training time. They additionally improve robustness by regularizing model features between the standard image, the adversarially perturbed image and a perturbation of the image created with an adversarial noise generator.\n\n\n2. Strengths\n+ The method is based on adversarial training. As far as I know and as the authors note this is the only method that reliably leads to more robust models.\n+ The authors attack their models with  a range of attacks that to the best of my knowledge are state-of-the art.\n+ The method apparently works in the multi norm setting.\n\n3. Weaknesses \n- I was missing an intuitive description why the adversarial noise should improve robustness to adversarial attacks. I was only aware of it as a method to improve corruption robustness.\n- I was not always sure if I got everything correctly in sections 4 and 5.3. I think I got it but I sometimes missed a figure. It may e.g. be helpful to include the losses in Figure 1 or make a separate figure. Especially why the MNG was trained the way it is was a bit unclear for me.\n\n\n4. Recommendation\n\nI think this paper is an accept but as I don't work with adversarial examples I am not at all confident in that assessment. From the discussions with people who work on adversarial examples new defenses are usually broken very quickly and there is a number of papers which break numerous defenses. The method is however based on adversarial training which to my knowledge is the only robust method so far and the used attacks seem valid. So I am definitely leaning towards accept but the opinion of a real expert would be highly appreciated as I feel not at all qualified to assess the validity of papers on adversarial examples.\n\n\n5. Questions/Recommendations\n- Is there a difference between the M(eta)NG and the A(dversarial)NG from Rusak et. al. 2020?\n\n\n6. Additional feedback \n- None as the paper is pretty well written.", "rating": "6: Marginally above acceptance threshold", "confidence": "1: The reviewer's evaluation is an educated guess"}, "tcdate": 1603807626690}], "openreview_url": "https://openreview.net/forum?id=tv8n52XbO4p", "arxiv_id": "2006.12135", "paper_pdf": "papers/tv8n52XbO4p.pdf", "paper_pdf_sha256": "ab8ac8c648509a90a97e4e250c851b7e6c8713ca47486becd709203daab88988", "paper_pdf_bytes": 2433764, "paper_pdf_source": "openreview", "code_url": "https://github.com/divyam3897/MNG_AC", "code_repository": "divyam3897/MNG_AC", "code_commit": "40c25a0c162c32cd858803f6bf1790180c727d06", "code_archive": "repos/tv8n52XbO4p.zip", "code_archive_sha256": "4c713f7a7e95aa3a4f4c476fc6526d71fb1f8387c800767cc7965c38e4e603a3", "code_archive_bytes": 537200, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 519, "github_languages": {"Python": 110764}, "github_archived": false, "github_pushed_at": "2021-06-25T05:27:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-generate-noise-for-robustness"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Jri6onxoMY", "year": 2026, "status": "rejected", "title": "Causal Cartographer: From Mapping to Reasoning Over Counterfactual Worlds", "authors": ["Gael Gendron", "Joze M. Rozanec", "Michael J. Witbrock", "Gillian Dobbie"], "authorids": ["~Gael_Gendron1", "~Joze_M._Rozanec1", "~Michael_J._Witbrock1", "~Gillian_Dobbie1"], "authors_source": "OpenReview API", "abstract": "Causal world models are systems that can answer counterfactual questions about an environment of interest, i.e., predict how it would have evolved if an arbitrary subset of events had been realized differently. The ability to answer such questions is crucial for models to reliably understand the world. However, this task currently eludes large language models (LLMs), which do not have demonstrated causal reasoning capabilities beyond the memorization of existing causal relationships. Furthermore, evaluating counterfactuals in real-world applications is challenging since only the factual world is observed, limiting evaluation to synthetic datasets. We address these problems by proposing the Causal Cartographer, a twofold system composed of two agents: the first extracts causal relationships from data and builds a vast repository of causal knowledge, while the second uses them as constraints to perform reliable step-by-step causal inference. We evaluate our approach on real-world counterfactuals obtained by matching data from diverse news sources. We show that our approach can extract accurate causal knowledge and enhance the robustness of LLMs for causal reasoning tasks. In particular, the proposed causal conditioning mitigates the impact of spurious correlations and greatly reduces inference costs (by up to 70\\%) compared to chain-of-thought reasoning.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "I5UpHdzMYt", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14563/Reviewer_yTwy"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper, “Causal Cartographer: From Mapping to Reasoning over Counterfactual Worlds,” proposes a novel two-agent framework for enabling large language models (LLMs) to extract, organize, and reason with causal knowledge from real-world text. The system, called Causal Cartographer, consists of (1) CTG-Extract, a causal extraction agent based on Graph Retrieval-Augmented Generation (Graph-RAG) that builds a large-scale causal graph (“CausalWorld”) from unstructured text (e.g., 500 economic news articles), and (2) CTG-Reason, a counterfactual reasoning agent that performs stepwise inference under causal constraints. The authors introduce theoretical contributions, notably the concept of causal blankets (a generalization of Markov blankets) and a K-matching algorithm for identifying counterfactual pairs of worlds in text data. Experiments compare CTG-Reason with the chain-of-thought-based CausalCoT method on a new dataset, CausalWorld-CR, derived from real-world causal extractions. Results show comparable or better accuracy and reduced computational cost (up to 70% reduction in inference cost), especially for smaller models like o3-mini and LLaMA-3.1-8B.", "review_text": "This paper, “Causal Cartographer: From Mapping to Reasoning over Counterfactual Worlds,” proposes a novel two-agent framework for enabling large language models (LLMs) to extract, organize, and reason with causal knowledge from real-world text. The system, called Causal Cartographer, consists of (1) CTG-Extract, a causal extraction agent based on Graph Retrieval-Augmented Generation (Graph-RAG) that builds a large-scale causal graph (“CausalWorld”) from unstructured text (e.g., 500 economic news articles), and (2) CTG-Reason, a counterfactual reasoning agent that performs stepwise inference under causal constraints. The authors introduce theoretical contributions, notably the concept of causal blankets (a generalization of Markov blankets) and a K-matching algorithm for identifying counterfactual pairs of worlds in text data. Experiments compare CTG-Reason with the chain-of-thought-based CausalCoT method on a new dataset, CausalWorld-CR, derived from real-world causal extractions. Results show comparable or better accuracy and reduced computational cost (up to 70% reduction in inference cost), especially for smaller models like o3-mini and LLaMA-3.1-8B.", "strengths": "The paper addresses an important gap between abstract causal reasoning and real-world data extraction. Its proposed combination of causal extraction and counterfactual reasoning within an LLM framework is both ambitious and well-motivated. The introduction of CausalWorld, a large-scale, structured repository of 975 nodes and 1337 causal relations, is an impressive resource that could stimulate further research. The integration of Graph-RAG retrieval ensures grounding in prior causal context during extraction, improving coherence and scalability.", "weaknesses": "Despite its strengths, the paper has several limitations that hinder its maturity for a top-tier conference. The evaluation is limited in scope and realism: the CausalWorld-CR dataset is constructed via synthetic matching across news articles rather than ground-truth counterfactual data. This raises concerns about the validity of “real-world” claims and the soundness of the evaluation metric.", "questions": "The concept of causal blankets (Section 5.1) should be more carefully distinguished from Pearl’s Markov blankets beyond lineage claims.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper, “Causal Cartographer: From Mapping to Reasoning over Counterfactual Worlds,” proposes a novel two-agent framework for enabling large language models (LLMs) to extract, organize, and reason with causal knowledge from real-world text. The system, called Causal Cartographer, consists of (1) CTG-Extract, a causal extraction agent based on Graph Retrieval-Augmented Generation (Graph-RAG) that builds a large-scale causal graph (“CausalWorld”) from unstructured text (e.g., 500 economic news articles), and (2) CTG-Reason, a counterfactual reasoning agent that performs stepwise inference under causal constraints. The authors introduce theoretical contributions, notably the concept of causal blankets (a generalization of Markov blankets) and a K-matching algorithm for identifying counterfactual pairs of worlds in text data. Experiments compare CTG-Reason with the chain-of-thought-based CausalCoT method on a new dataset, CausalWorld-CR, derived from real-world causal extractions. Results show comparable or better accuracy and reduced computational cost (up to 70% reduction in inference cost), especially for smaller models like o3-mini and LLaMA-3.1-8B.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper addresses an important gap between abstract causal reasoning and real-world data extraction. Its proposed combination of causal extraction and counterfactual reasoning within an LLM framework is both ambitious and well-motivated. The introduction of CausalWorld, a large-scale, structured repository of 975 nodes and 1337 causal relations, is an impressive resource that could stimulate further research. The integration of Graph-RAG retrieval ensures grounding in prior causal context during extraction, improving coherence and scalability.", "weaknesses": "Despite its strengths, the paper has several limitations that hinder its maturity for a top-tier conference. The evaluation is limited in scope and realism: the CausalWorld-CR dataset is constructed via synthetic matching across news articles rather than ground-truth counterfactual data. This raises concerns about the validity of “real-world” claims and the soundness of the evaluation metric.", "questions": "The concept of causal blankets (Section 5.1) should be more carefully distinguished from Pearl’s Markov blankets beyond lineage claims.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762239671556}, {"id": "MKT7ycbKYq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14563/Reviewer_ksfa"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes Causal Cartographer, a two-agent framework for causal reasoning over natural-language sources. The framework consists of (1) CTG-Extract, which performs graph-RAG–assisted causal extraction from news articles to build a large causal knowledge base (CausalWorld), and (2) CTG-Reason, which performs step-by-step, causally constrained inference (including counterfactuals) by conditioning only on parents/children along the graph. The authors also introduce “causal blankets” which, along with a K-matching procedure, enable approximating real-world counterfactuals by matching “worlds” across documents. Empirically, on a 400-query dataset (CausalWorld-CR) derived from news in 2020, CTG-Reason attains accuracy on par with, or better than, a CoT baseline while reducing context and output tokens (up to 70% fewer), with especially large efficiency gains on small models.", "review_text": "The paper proposes Causal Cartographer, a two-agent framework for causal reasoning over natural-language sources. The framework consists of (1) CTG-Extract, which performs graph-RAG–assisted causal extraction from news articles to build a large causal knowledge base (CausalWorld), and (2) CTG-Reason, which performs step-by-step, causally constrained inference (including counterfactuals) by conditioning only on parents/children along the graph. The authors also introduce “causal blankets” which, along with a K-matching procedure, enable approximating real-world counterfactuals by matching “worlds” across documents. Empirically, on a 400-query dataset (CausalWorld-CR) derived from news in 2020, CTG-Reason attains accuracy on par with, or better than, a CoT baseline while reducing context and output tokens (up to 70% fewer), with especially large efficiency gains on small models.", "strengths": "- The paper argues well for why explicit causal constraints can mitigate spurious correlations and reduce inference cost.\n- The two-agent split via decomposition of the task as extraction and reasoning enables each agent to focus specifically on its own task.\n- The causal-blanket definition and K-Matching Equivalence theorem formalize when matched worlds yield valid counterfactual targets—useful for this emerging evaluation paradigm.\n- The reported token/input reductions and output length shrinkage are substantial while maintaining accuracy.", "weaknesses": "- The text corpus utilized is 2020 news with focus on economics. What factors led to this choice? How well does this approach perform in other domains?\n- The method leans on SCM framing (DAGs), yet the constructed CausalWorld allows cycles/feedback loops (Fig. 6).\n- Causal blankets are defined as fully determining the target (deterministic f). Real news variables are often noisy. Can the theorem and agent be generalized to stochastic blankets?", "questions": "1) A small set of bridge nodes routes information across communities. Did you measure how removing a top-k bridge node affects the fraction of nodes still usable for counterfactuals and the success rate of K-matching?\n\n2) You remove queries with ≥50 causal paths and rebalance degree skew. How sensitive are results to the “≥50” threshold, and what happens if you keep hard queries?\n\n3) What max recursion depth or search budget do you set for anticausal inference when parents/children are missing, and how often do queries exceed it?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Causal Cartographer, a two-agent framework for causal reasoning over natural-language sources. The framework consists of (1) CTG-Extract, which performs graph-RAG–assisted causal extraction from news articles to build a large causal knowledge base (CausalWorld), and (2) CTG-Reason, which performs step-by-step, causally constrained inference (including counterfactuals) by conditioning only on parents/children along the graph. The authors also introduce “causal blankets” which, along with a K-matching procedure, enable approximating real-world counterfactuals by matching “worlds” across documents. Empirically, on a 400-query dataset (CausalWorld-CR) derived from news in 2020, CTG-Reason attains accuracy on par with, or better than, a CoT baseline while reducing context and output tokens (up to 70% fewer), with especially large efficiency gains on small models.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper argues well for why explicit causal constraints can mitigate spurious correlations and reduce inference cost.\n- The two-agent split via decomposition of the task as extraction and reasoning enables each agent to focus specifically on its own task.\n- The causal-blanket definition and K-Matching Equivalence theorem formalize when matched worlds yield valid counterfactual targets—useful for this emerging evaluation paradigm.\n- The reported token/input reductions and output length shrinkage are substantial while maintaining accuracy.", "weaknesses": "- The text corpus utilized is 2020 news with focus on economics. What factors led to this choice? How well does this approach perform in other domains?\n- The method leans on SCM framing (DAGs), yet the constructed CausalWorld allows cycles/feedback loops (Fig. 6).\n- Causal blankets are defined as fully determining the target (deterministic f). Real news variables are often noisy. Can the theorem and agent be generalized to stochastic blankets?", "questions": "1) A small set of bridge nodes routes information across communities. Did you measure how removing a top-k bridge node affects the fraction of nodes still usable for counterfactuals and the success rate of K-matching?\n\n2) You remove queries with ≥50 causal paths and rebalance degree skew. How sensitive are results to the “≥50” threshold, and what happens if you keep hard queries?\n\n3) What max recursion depth or search budget do you set for anticausal inference when parents/children are missing, and how often do queries exceed it?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761998084730}, {"id": "CF22NXiiwt", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14563/Reviewer_F6Ef"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The authors present Causal Cartographer which extracts causal relationships from data and then uses them as constraints to perform step-by-step causal inference. The performance is evaluated on real-world counterfactuals obtained from news sources.", "review_text": "The authors present Causal Cartographer which extracts causal relationships from data and then uses them as constraints to perform step-by-step causal inference. The performance is evaluated on real-world counterfactuals obtained from news sources.", "strengths": "1. Understanding LLM performance on counterfactual reasoning tasks is crucial is furthering reasearch on LLMs ability to do causal tasks\n2. Using real-world data instead of synthetic is encouraging\n3. The proposed method is more interpretable which is good for future research", "weaknesses": "1. It is not clear how this method can scale to production LLM systems.\n2. Causal graph building would require a very careful control so as to not introduce bias", "questions": "1. Does this framework have the risk of running \"stale\". In a constantly evolving world, what if the causal relationships from the first stage change? How would one go about keeping them up to date? Would this update process eat into the inference cost savings?\n2. How does the system defend itself against adversarial attacks where noisy/false claims are injected into the causal knowledge repository?\n3. Sorry if I missed this, but how is it ensured that the extracted relationships are causal and not noise?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present Causal Cartographer which extracts causal relationships from data and then uses them as constraints to perform step-by-step causal inference. The performance is evaluated on real-world counterfactuals obtained from news sources.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Understanding LLM performance on counterfactual reasoning tasks is crucial is furthering reasearch on LLMs ability to do causal tasks\n2. Using real-world data instead of synthetic is encouraging\n3. The proposed method is more interpretable which is good for future research", "weaknesses": "1. It is not clear how this method can scale to production LLM systems.\n2. Causal graph building would require a very careful control so as to not introduce bias", "questions": "1. Does this framework have the risk of running \"stale\". In a constantly evolving world, what if the causal relationships from the first stage change? How would one go about keeping them up to date? Would this update process eat into the inference cost savings?\n2. How does the system defend itself against adversarial attacks where noisy/false claims are injected into the causal knowledge repository?\n3. Sorry if I missed this, but how is it ensured that the extracted relationships are causal and not noise?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761979202849}, {"id": "W4j6ilHWVr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14563/Reviewer_6vkm"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "It is difficult to both perform causal reasoning with large language models and evaluate them, due to ladder of causality wich states that interventional and counterfactual quantities can generally not be inferred from observed data unless strong assumptions. The submission proposes to extract known causal relationships from real-world articles, yielding a causal world model, then use that model to perform causal reasoning with LLMs as well as evaluating them.", "review_text": "It is difficult to both perform causal reasoning with large language models and evaluate them, due to ladder of causality wich states that interventional and counterfactual quantities can generally not be inferred from observed data unless strong assumptions. The submission proposes to extract known causal relationships from real-world articles, yielding a causal world model, then use that model to perform causal reasoning with LLMs as well as evaluating them.", "strengths": "- Original approach, which I had not seen before while seeming natural in the context of LLMs.\n\n- The paper is generally clear and well-written.\n\n- Experiments support the method outperforming past alternatives in terms of performance and efficiency.", "weaknesses": "- The definition of SCMs used by authors ignores noise in individual structural equations. Notably, Definition 1 assumes deterministic relationships between the causal blanket and the target variable, while in general, noise variables can be present. This makes it unclear to assess whether Definition 1 and Theorem 1 is only possible in the absence of noise, which is a generally restrictive scenario.\n\n- \"We also excluded outliers (∼4% of the answers were nonsensical numbers).\" (l.413-415) : this seems a bit quick to me... It would helpful to know the fraction of outliers for each evaluated model, how they change results, and how to evaluate performance in a way that is robust to them if they dominate averages.", "questions": "- How do Definition 1 and Theorem 1 generalize in the presence of noise variables?\n\n- What if you include outliers, and check the things indicated above?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "It is difficult to both perform causal reasoning with large language models and evaluate them, due to ladder of causality wich states that interventional and counterfactual quantities can generally not be inferred from observed data unless strong assumptions. The submission proposes to extract known causal relationships from real-world articles, yielding a causal world model, then use that model to perform causal reasoning with LLMs as well as evaluating them.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- Original approach, which I had not seen before while seeming natural in the context of LLMs.\n\n- The paper is generally clear and well-written.\n\n- Experiments support the method outperforming past alternatives in terms of performance and efficiency.", "weaknesses": "- The definition of SCMs used by authors ignores noise in individual structural equations. Notably, Definition 1 assumes deterministic relationships between the causal blanket and the target variable, while in general, noise variables can be present. This makes it unclear to assess whether Definition 1 and Theorem 1 is only possible in the absence of noise, which is a generally restrictive scenario.\n\n- \"We also excluded outliers (∼4% of the answers were nonsensical numbers).\" (l.413-415) : this seems a bit quick to me... It would helpful to know the fraction of outliers for each evaluated model, how they change results, and how to evaluate performance in a way that is robust to them if they dominate averages.", "questions": "- How do Definition 1 and Theorem 1 generalize in the presence of noise variables?\n\n- What if you include outliers, and check the things indicated above?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761213883835}], "openreview_url": "https://openreview.net/forum?id=Jri6onxoMY", "arxiv_id": "2505.14396", "paper_pdf": "papers/Jri6onxoMY.pdf", "paper_pdf_sha256": "77dcce5a6be1344f330b5921b371ba5836cc6ac0dc1a1972118e019857cf6d4b", "paper_pdf_bytes": 1051050, "paper_pdf_source": "openreview", "code_url": "https://github.com/ggendro/causal-cartographer", "code_repository": "ggendro/causal-cartographer", "code_commit": "a69b23e6f4a4477a08e65e13813662d9a053b882", "code_archive": "repos/Jri6onxoMY.zip", "code_archive_sha256": "32c3fdf9204cb02746d258c5a6d3e45fddb2bb3fbf75617b66a9d56ad8f4f56d", "code_archive_bytes": 526372, "code_file_count": 45, "code_extensions": {".py": 44, ".js": 1}, "github_disk_usage_kb": 1943, "github_languages": {"Python": 480960, "JavaScript": 59235, "CSS": 17196, "HTML": 5759}, "github_archived": false, "github_pushed_at": "2025-05-17T23:02:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/causal-cartographer-from-mapping-to-reasoning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rapXZIfwbX", "year": 2025, "status": "rejected", "title": "The Persian Rug: solving toy models of superposition using large-scale symmetries", "authors": ["Aditya Cowsik", "Kfir Dolev", "Alex Infanger"], "authorids": ["~Aditya_Cowsik1", "~Kfir_Dolev1", "~Alex_Infanger1"], "authors_source": "OpenReview API", "abstract": "We present a complete mechanistic description of the algorithm learned by a minimal non-linear sparse data autoencoder in the limit of large input dimension. The model, originally presented in \\cite{elhage2022superposition}, compresses sparse data vectors through a linear layer and decompresses using another linear layer followed by a ReLU activation. We notice that when the data is permutation symmetric (no input feature is privileged) large models reliably learn an algorithm that is sensitive to individual weights only through their large-scale statistics. For these models, the loss function becomes analytically tractable. Using this understanding, we give explicit upper bounds on the loss, which show that the model is near-optimal among recently proposed architectures. In particular, changes to the elementwise activation function or the addition of gating can at best improve its performance by a constant factor. Finally, we forward-engineer a model with the requisite symmetries and show that its loss precisely matches that of the trained models. Unlike the trained model weights, the minimal randomness in the artificial weights results in miraculous fractal structures resembling a Persian rug, to which the algorithm is oblivious. Our work contributes to neural network interpretability by introducing techniques for understanding the structure of autoencoders.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "OhZatZ5BjO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_ru6M"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper characterizes the solutions of training sparse autoencoders (defined as $f(\\mathbf{x}) = \\textrm{ReLU}(W_{out} W_{in} \\mathbf{x} + \\mathbf{b})$) with the reconstruction loss $L=\\mathbb{E} || \\mathbf{x} - f(\\mathbf{x}) ||^2$. The paper first trains an autoencoder on synthetic data consisting of sparse vectors generated from a Bernoulli-Uniform model and records some empirical observations about the composite matrix $W=W_{out} W_{in}$. The authors note that the off diagonal entries of this matrix seem to follow a normal distribution that seems to be identical across rows and that the diagonal entries are also identical. This allows them to reduce the loss function to a function of three scalar variables corresponding to the diagonal, off-diagonal, and bias terms. The solution they obtain is of the form $W = OPO^\\top$ where $O$ is orthogonal and $P$ is a rank $n_d$ projection matrix. The authors also characterize how the loss varies as a function of the Bernoulli sparsity parameter $p$.", "review_text": "This paper characterizes the solutions of training sparse autoencoders (defined as $f(\\mathbf{x}) = \\textrm{ReLU}(W_{out} W_{in} \\mathbf{x} + \\mathbf{b})$) with the reconstruction loss $L=\\mathbb{E} || \\mathbf{x} - f(\\mathbf{x}) ||^2$. The paper first trains an autoencoder on synthetic data consisting of sparse vectors generated from a Bernoulli-Uniform model and records some empirical observations about the composite matrix $W=W_{out} W_{in}$. The authors note that the off diagonal entries of this matrix seem to follow a normal distribution that seems to be identical across rows and that the diagonal entries are also identical. This allows them to reduce the loss function to a function of three scalar variables corresponding to the diagonal, off-diagonal, and bias terms. The solution they obtain is of the form $W = OPO^\\top$ where $O$ is orthogonal and $P$ is a rank $n_d$ projection matrix. The authors also characterize how the loss varies as a function of the Bernoulli sparsity parameter $p$.", "strengths": "The empirical observations and theoretical results are reasonable and the authors provide an adequate description of the solution $W$.", "weaknesses": "Connections between sparse coding and autoencoders have been explored extensively in the literature. This paper does not cite or engage with any of this literature. \n\nThe empirical observations about $W$ having identical diagonal entries and the off diagonal entries being small and having a common distribution can be captured by $W = A^\\top A$ where $A \\in \\mathbb{R}^{n_d \\times n_s}$ is an overcomplete, incoherent dictionary with unit norm columns. This ensures that the diagonal terms are identical while the off-diagonal terms are small in magnitude. Moreover it is well known that random matrices satisfy the incoherence/restricted isometry condition and a random $A$ ensures that the off-diagonal entries follow the same distribution.\n\nThis model of dictionary learning, along with Bernoulli-Uniform or Bernoulli-Gaussian sparse codes has been studied in the complete and overcomplete setting with neural and non-neural algorithms [3,4,5].\n\nReference [1] studies how ReLU autoencoders with tied weights can learn dictionaries with data from this generative model. The authors show (in theorem 3.1) that one layer of a ReLU autoencoder can recover the support of the sparse code with high probability if the weights are close to the ground truth dictionary. The result also proposes a bias vector with all negative entries whose magnitude can be derived from the incoherence of the dictionary. The paper also uses a landscape argument to show that the ground truth dictionary is a critical point for the reconstruction loss. Reference [2] goes further to show that autoencoders trained to minimize reconstruction error with gradient descent can recover the ground truth dictionary.\n\n[1] Rangamani, A., Mukherjee, A., Basu, A., Arora, A., Ganapathi, T., Chin, S., & Tran, T. D. (2018, June). Sparse coding and autoencoders. In 2018 IEEE International Symposium on Information Theory (ISIT) (pp. 36-40). IEEE.\n\n[2] Nguyen, T. V., Wong, R. K., & Hegde, C. (2019, April). On the dynamics of gradient descent for autoencoders. In The 22nd International Conference on Artificial Intelligence and Statistics (pp. 2858-2867). PMLR.\n\n[3] Arora, S., Ge, R., Ma, T., & Moitra, A. (2015, June). Simple, efficient, and neural algorithms for sparse coding. In Conference on learning theory (pp. 113-149). PMLR.\n\n[4] Agarwal, A., Anandkumar, A., Jain, P., & Netrapalli, P. (2016). Learning sparsely used overcomplete dictionaries via alternating minimization. SIAM Journal on Optimization, 26(4), 2775-2799.\n\n[5] Spielman, D. A., Wang, H., & Wright, J. (2012, June). Exact recovery of sparsely-used dictionaries. In Conference on Learning Theory (pp. 37-1). JMLR Workshop and Conference Proceedings.\n\n[6] Refinetti, M., & Goldt, S. (2022, June). The dynamics of representation learning in shallow, non-linear autoencoders. In International Conference on Machine Learning (pp. 18499-18519). PMLR.\n\nWhile these prior results study a slightly different model architecture, they can be adapted to the model studied in this paper with a little effort. In my opinion the results obtained by the current paper can be explained by the sparse coding model, and it is not true that this has not been studied in the context of autoencoders.", "questions": "1. How does the solution proposed by the authors differ from the dictionary learning solution? Can the authors delineate conditions under which trained autoencoders converge to their solution instead of overcomplete, incoherent dictionaries?\n\n2. The previous literature does not study autoencoders and sparse coding in the context of interpretability. Are there specific questions that arise in this context that require the authors' model? Does the authors' analysis answer these questions?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper characterizes the solutions of training sparse autoencoders (defined as $f(\\mathbf{x}) = \\textrm{ReLU}(W_{out} W_{in} \\mathbf{x} + \\mathbf{b})$) with the reconstruction loss $L=\\mathbb{E} || \\mathbf{x} - f(\\mathbf{x}) ||^2$. The paper first trains an autoencoder on synthetic data consisting of sparse vectors generated from a Bernoulli-Uniform model and records some empirical observations about the composite matrix $W=W_{out} W_{in}$. The authors note that the off diagonal entries of this matrix seem to follow a normal distribution that seems to be identical across rows and that the diagonal entries are also identical. This allows them to reduce the loss function to a function of three scalar variables corresponding to the diagonal, off-diagonal, and bias terms. The solution they obtain is of the form $W = OPO^\\top$ where $O$ is orthogonal and $P$ is a rank $n_d$ projection matrix. The authors also characterize how the loss varies as a function of the Bernoulli sparsity parameter $p$.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The empirical observations and theoretical results are reasonable and the authors provide an adequate description of the solution $W$.", "weaknesses": "Connections between sparse coding and autoencoders have been explored extensively in the literature. This paper does not cite or engage with any of this literature. \n\nThe empirical observations about $W$ having identical diagonal entries and the off diagonal entries being small and having a common distribution can be captured by $W = A^\\top A$ where $A \\in \\mathbb{R}^{n_d \\times n_s}$ is an overcomplete, incoherent dictionary with unit norm columns. This ensures that the diagonal terms are identical while the off-diagonal terms are small in magnitude. Moreover it is well known that random matrices satisfy the incoherence/restricted isometry condition and a random $A$ ensures that the off-diagonal entries follow the same distribution.\n\nThis model of dictionary learning, along with Bernoulli-Uniform or Bernoulli-Gaussian sparse codes has been studied in the complete and overcomplete setting with neural and non-neural algorithms [3,4,5].\n\nReference [1] studies how ReLU autoencoders with tied weights can learn dictionaries with data from this generative model. The authors show (in theorem 3.1) that one layer of a ReLU autoencoder can recover the support of the sparse code with high probability if the weights are close to the ground truth dictionary. The result also proposes a bias vector with all negative entries whose magnitude can be derived from the incoherence of the dictionary. The paper also uses a landscape argument to show that the ground truth dictionary is a critical point for the reconstruction loss. Reference [2] goes further to show that autoencoders trained to minimize reconstruction error with gradient descent can recover the ground truth dictionary.\n\n[1] Rangamani, A., Mukherjee, A., Basu, A., Arora, A., Ganapathi, T., Chin, S., & Tran, T. D. (2018, June). Sparse coding and autoencoders. In 2018 IEEE International Symposium on Information Theory (ISIT) (pp. 36-40). IEEE.\n\n[2] Nguyen, T. V., Wong, R. K., & Hegde, C. (2019, April). On the dynamics of gradient descent for autoencoders. In The 22nd International Conference on Artificial Intelligence and Statistics (pp. 2858-2867). PMLR.\n\n[3] Arora, S., Ge, R., Ma, T., & Moitra, A. (2015, June). Simple, efficient, and neural algorithms for sparse coding. In Conference on learning theory (pp. 113-149). PMLR.\n\n[4] Agarwal, A., Anandkumar, A., Jain, P., & Netrapalli, P. (2016). Learning sparsely used overcomplete dictionaries via alternating minimization. SIAM Journal on Optimization, 26(4), 2775-2799.\n\n[5] Spielman, D. A., Wang, H., & Wright, J. (2012, June). Exact recovery of sparsely-used dictionaries. In Conference on Learning Theory (pp. 37-1). JMLR Workshop and Conference Proceedings.\n\n[6] Refinetti, M., & Goldt, S. (2022, June). The dynamics of representation learning in shallow, non-linear autoencoders. In International Conference on Machine Learning (pp. 18499-18519). PMLR.\n\nWhile these prior results study a slightly different model architecture, they can be adapted to the model studied in this paper with a little effort. In my opinion the results obtained by the current paper can be explained by the sparse coding model, and it is not true that this has not been studied in the context of autoencoders.", "questions": "1. How does the solution proposed by the authors differ from the dictionary learning solution? Can the authors delineate conditions under which trained autoencoders converge to their solution instead of overcomplete, incoherent dictionaries?\n\n2. The previous literature does not study autoencoders and sparse coding in the context of interpretability. Are there specific questions that arise in this context that require the authors' model? Does the authors' analysis answer these questions?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730851242897}, {"id": "fG6rnL2qs2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_br6G"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper both theoretically and empirically performed an exhaustive analysis of autoencoder trained on sparse data,\nextracting the essential structure needed to represent such data. The paper firstly reveals that the change of loss as the compression ratio varies and shows that when compression ratio is near zero, the loss drops to zero entirely. This motivates the authors to consider the structure of the weight in both linear and non-linear cases. They artificially include the statistical permutation symmetry into the weight and found the artificial weight and the trained model returned the same result. The conclusion is they qualitatively show that a statistically symmetric strategy exists in certain regimes of the macroscopic parameters. Thus optimal choice for W exisits based on the proposed  Persian Rug Model.", "review_text": "The paper both theoretically and empirically performed an exhaustive analysis of autoencoder trained on sparse data,\nextracting the essential structure needed to represent such data. The paper firstly reveals that the change of loss as the compression ratio varies and shows that when compression ratio is near zero, the loss drops to zero entirely. This motivates the authors to consider the structure of the weight in both linear and non-linear cases. They artificially include the statistical permutation symmetry into the weight and found the artificial weight and the trained model returned the same result. The conclusion is they qualitatively show that a statistically symmetric strategy exists in certain regimes of the macroscopic parameters. Thus optimal choice for W exisits based on the proposed  Persian Rug Model.", "strengths": "S1. Novelty: The paper is novel in it both theoretically and empirically performed exhaustive analysis of autoencoder trained on sparse data, interpreting the essential structure needed to represent such data. I like the way they demonstrates the distribution of the weights. \n\nS2. The paper firstly reveals that a statistically symmetric strategy exists in certain regimes of the macroscopic parameters and thus optimal choice for W exisits based on the proposed  Persian Rug Model. The theoretical support of the motivation is strong. \n\nS3. The presentation of the paper is good, with a good clarity.", "weaknesses": "W1. What is the interpretation of the model if the input is not sparse? Is the symmetry property and the relevant conclusion partially hold? I noticed the theories are mostly based on the compression ratio. Does this mean the interpretation will change if the compression ration becomes high?  \n\nW2. What is the significance of the model if the input is no longer sparse?", "questions": "My questions are mostly overlapped with the weakness section. I appreciate it if the authors may further instantiate the following points:\n\nW1. What is the interpretation of the model if the input is not sparse? Is the symmetry property and the relevant conclusion partially hold? Does this mean the interpretation of the weights and the optimal W will change significantly if the compression ration becomes high?  \n\nW2. Generally speaking, what is the significance of the proposed model if the input is no longer sparse, whereas in most cases the input is not sparse?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper both theoretically and empirically performed an exhaustive analysis of autoencoder trained on sparse data,\nextracting the essential structure needed to represent such data. The paper firstly reveals that the change of loss as the compression ratio varies and shows that when compression ratio is near zero, the loss drops to zero entirely. This motivates the authors to consider the structure of the weight in both linear and non-linear cases. They artificially include the statistical permutation symmetry into the weight and found the artificial weight and the trained model returned the same result. The conclusion is they qualitatively show that a statistically symmetric strategy exists in certain regimes of the macroscopic parameters. Thus optimal choice for W exisits based on the proposed  Persian Rug Model.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "S1. Novelty: The paper is novel in it both theoretically and empirically performed exhaustive analysis of autoencoder trained on sparse data, interpreting the essential structure needed to represent such data. I like the way they demonstrates the distribution of the weights. \n\nS2. The paper firstly reveals that a statistically symmetric strategy exists in certain regimes of the macroscopic parameters and thus optimal choice for W exisits based on the proposed  Persian Rug Model. The theoretical support of the motivation is strong. \n\nS3. The presentation of the paper is good, with a good clarity.", "weaknesses": "W1. What is the interpretation of the model if the input is not sparse? Is the symmetry property and the relevant conclusion partially hold? I noticed the theories are mostly based on the compression ratio. Does this mean the interpretation will change if the compression ration becomes high?  \n\nW2. What is the significance of the model if the input is no longer sparse?", "questions": "My questions are mostly overlapped with the weakness section. I appreciate it if the authors may further instantiate the following points:\n\nW1. What is the interpretation of the model if the input is not sparse? Is the symmetry property and the relevant conclusion partially hold? Does this mean the interpretation of the weights and the optimal W will change significantly if the compression ration becomes high?  \n\nW2. Generally speaking, what is the significance of the proposed model if the input is no longer sparse, whereas in most cases the input is not sparse?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730596092903}, {"id": "cyy3WIcsgM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_f6ZT"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "The authors study the compression of sparse data using a two-layer autoencoder with ReLU output non-linearity, considering the toy model of Elhage et al (2022). Building on an ansatz for the product of the weights, supported by numerical observation, they derive a closed-form expression for the population loss depending on only three scalar parameters, corresponding to the bias, common diagonal element of the product matrix, and variance of off-diagonal elements. They deduce from it a lower-bound on the loss, and construct a deterministic weight matrix that achieves it. They find that the corresponding loss coincides with that of the trained model.", "review_text": "The authors study the compression of sparse data using a two-layer autoencoder with ReLU output non-linearity, considering the toy model of Elhage et al (2022). Building on an ansatz for the product of the weights, supported by numerical observation, they derive a closed-form expression for the population loss depending on only three scalar parameters, corresponding to the bias, common diagonal element of the product matrix, and variance of off-diagonal elements. They deduce from it a lower-bound on the loss, and construct a deterministic weight matrix that achieves it. They find that the corresponding loss coincides with that of the trained model.", "strengths": "On a technical level, the authors provide a detailed study of the optimal population loss achieved by the considered model, which is non trivial given the ReLU non-linearity. To the best of my awareness, this derivation is new compared to the original work of Elhage et al. (2022). The finding of the weight matrix achieving the same loss as the model optimized using classical gradient-based optimizers is an interesting construction. Although I have not checked the math in detail, the derivation looks scientifically sound.\n\nOn a writing level, the paper is very clearly written, with sufficient discussion and intuition building. The assumptions are illustrated by careful numerical experiments.  Sufficient discussion accompanies all important equations.", "weaknesses": "My main concern lies in the scope of the contribution. Some of my concerns are tied to my questions, which I list in the next section.\n\n- To my understanding, the study is confined to the isotropic case of Elhage et al. (2022), as it builds upon the ansatz that all diagonal elements of $W_{out}W_{in}$ are equal. This means there is not really any feature learning in the model, although the considered setting remains non-trivial.\n\n- While the study is interesting, it does not add much more phenomenological insights compared to the original work. In particular, the authors construct a matrix which achieves the same optimal population loss as the trained model, which is one of the strengths of the work. However, I understand any matrix of the form (7) is similarly optimal, and I do not understand why the explicit realization in 4.2.2 brings any further insight. I might have misunderstood a point, and am happy if the authors bring clarifications on this point.\n\nOverall, I think the study is technically sound, the insights tied to the analysis of interest, but brings little novel phenomenological insight.  I am thus giving an accept score -- because the paper looks sound --, but not a high one.", "questions": "- (Minor) Do the authors believe the study could be extended to non-isotropic cases, where some features are of higher importance, e.g. in the block case where blocks of different features have the same importance ?\n\n- Is it correct that (7) also corresponds to the weights a linear model would learn ? If that is the case, does the improvement in reconstruction loss from the reLU model come from the bias and activation alone ? \n\n- As pointed above, is there any specific reason the particular realization of 4.2.2 of optimal weights is of specific interest, compared to randomly drawing an orthogonal matrix $O$ and constructing a weight matrix as in (7)? Do the authors claim the model learns weights exactly corresponding to 4.2.2, or simply that the loss coincide ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors study the compression of sparse data using a two-layer autoencoder with ReLU output non-linearity, considering the toy model of Elhage et al (2022). Building on an ansatz for the product of the weights, supported by numerical observation, they derive a closed-form expression for the population loss depending on only three scalar parameters, corresponding to the bias, common diagonal element of the product matrix, and variance of off-diagonal elements. They deduce from it a lower-bound on the loss, and construct a deterministic weight matrix that achieves it. They find that the corresponding loss coincides with that of the trained model.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "On a technical level, the authors provide a detailed study of the optimal population loss achieved by the considered model, which is non trivial given the ReLU non-linearity. To the best of my awareness, this derivation is new compared to the original work of Elhage et al. (2022). The finding of the weight matrix achieving the same loss as the model optimized using classical gradient-based optimizers is an interesting construction. Although I have not checked the math in detail, the derivation looks scientifically sound.\n\nOn a writing level, the paper is very clearly written, with sufficient discussion and intuition building. The assumptions are illustrated by careful numerical experiments.  Sufficient discussion accompanies all important equations.", "weaknesses": "My main concern lies in the scope of the contribution. Some of my concerns are tied to my questions, which I list in the next section.\n\n- To my understanding, the study is confined to the isotropic case of Elhage et al. (2022), as it builds upon the ansatz that all diagonal elements of $W_{out}W_{in}$ are equal. This means there is not really any feature learning in the model, although the considered setting remains non-trivial.\n\n- While the study is interesting, it does not add much more phenomenological insights compared to the original work. In particular, the authors construct a matrix which achieves the same optimal population loss as the trained model, which is one of the strengths of the work. However, I understand any matrix of the form (7) is similarly optimal, and I do not understand why the explicit realization in 4.2.2 brings any further insight. I might have misunderstood a point, and am happy if the authors bring clarifications on this point.\n\nOverall, I think the study is technically sound, the insights tied to the analysis of interest, but brings little novel phenomenological insight.  I am thus giving an accept score -- because the paper looks sound --, but not a high one.", "questions": "- (Minor) Do the authors believe the study could be extended to non-isotropic cases, where some features are of higher importance, e.g. in the block case where blocks of different features have the same importance ?\n\n- Is it correct that (7) also corresponds to the weights a linear model would learn ? If that is the case, does the improvement in reconstruction loss from the reLU model come from the bias and activation alone ? \n\n- As pointed above, is there any specific reason the particular realization of 4.2.2 of optimal weights is of specific interest, compared to randomly drawing an orthogonal matrix $O$ and constructing a weight matrix as in (7)? Do the authors claim the model learns weights exactly corresponding to 4.2.2, or simply that the loss coincide ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730500987659}, {"id": "NFTnUrMzAI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_iN6D"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The work tries to derive analytical ways to obtain the near-optimal solutions of a specific type of autoencoders for synthetic sparse datasets. After theoretical analysis, it provides the process for constructing the autoencoder’s weight matrix without numerical optimization.", "review_text": "The work tries to derive analytical ways to obtain the near-optimal solutions of a specific type of autoencoders for synthetic sparse datasets. After theoretical analysis, it provides the process for constructing the autoencoder’s weight matrix without numerical optimization.", "strengths": "1. Has theoretical analysis that looks legit to me, even though it is only focused on a very specific type of autoencoders on a specific type of datasets. \n2. The observation leading to theoretical analysis structure of this work can effectively convey the authors’ motivation.", "weaknesses": "The main weakness of this paper is its inconsistency and the absence of experiments for supporting their theoretical findings in Section 4, which greatly undermines the soundness of their results. Also, the authors should avoid introducing jargons and verbosity, be concise and accurate about what they mean, or provide necessary references if the jargons are widely accepted elsewhere. \nI suggest that the authors spend more time streamlining their work to improve its readability and precision, and bolster their theoretical claims with empirical results.\n\nRegarding contribution, the results in this paper might be too restricted to a specific type of problem and model. I'm not sure how crucial this problem is to LLM development, and because I also value rigorous theoretical studies of problems no matter how small they seem, I refrain from saying harsh words like \"it's trivial\". That said, the authors had better show their problem's importance with better literature review and critical discussion. \n\n## Major\n\n### Abstract\n\n1. Some claims in the abstract are not (well) reflected in the main text. For instance, “changes to the elementwise activation function or the addition of gating can at best improve its performance by a constant factor”. I guess this is about the discussion on Line 486 to 492. Experiments need to be done to verify this.\n\n### Section 1\n\n1. Line 76: Please provide clear definition of “permutation symmetric data” and “thermodynamic limit”. At least they do not appear in the work (Elhage 2022) you cited, and they sound kinda vague and unintuitive. For instance, how can “permutation symmetric” be derived from “no input feature is privileged”? As long as the features are not identical you cannot say it’s “permutation symmetric or invariant”, right? If these are jargons commonly used in your out-of-ML research area, you need to translate it for your mortal audiences in the ML community. \n\n### Section 2\n\n1. How many samples are there in your dataset on Line 111? What is the rank of this data matrix you sampled? Could the rank of this data matrix affect or even undermine your later observations and motivations? Did you scale up the number of samples as you increase $n_s$? If so, how? You need to provide more details about your experiments.\n\n### Section 3\n\n1. On line 142 there is *“due to an immediate drop in the loss starting around nd/ns ≈p. The slope and duration of this initial fall is controlled by p. In particular, in the high-sparsity regime (pclose to zero), the loss drops to zero entirely near the nd/ns ≈0 regime.”* \nI don’t think any part of Fig. 2b can clearly show this, especially considering that the upper bound of the x-axis (0.8) is way larger than p=0.05. You need to plot the loss curve at different p to demonstrate that, and zoom in on the “nd/ns ≈0 regime” and “nd/ns ≈p regime” to illustrate your points. Also, use the `subfigure` environments to enclose the subfigures.\n2. On line 152, there is *“nd/ns →∞ limit”*, but Fig. 3,4,5,6 cannot show any of such limit. All nd/ns are under 0.8.  Do you mean “nd/ns →0 limit” (although this makes no sense), or just ns →∞? Because of this, I'm uncertain when reading your statements from Line 155 to 161. \n    * If it's the latter, as I understand it, 8192 might still be kinda small. \n    * Are you always assuming that nd < ns throughout the entire paper? Be clear.\n3. Should provide a clear definition of “permutation invariant” mathematically. Also, what is “permutation symmetric” matrix? Be clear about what you say, or provide citations at least. \n4. Elaborate on “We hypothesize this occurs because for small models (and small feature probabilities) the contribution to any output from off-diagonal terms may fluctuate because of the small size of the matrix.” on Line 215. It’s like saying many things without saying anything. \n5. Elaborate on Line 236 to 239. Actually, this sentence is too convoluted to read. \n6. On Line 236 “Similar features are present in the two plots”. On Line 240 “we give three plots”. What and where are these five plots? Be specific. \n7. For Line 266 to 269 and Fig. 5. What is $\\nu$? What is $\\Delta \\text{var}(\\nu)$? What is “the fluctuation of the total norm of the off-diagonal elements in each row in W across rows”? Is it simply the variance of the off-diagonal elements’ variances? What do you mean by signal-to-noise ratio? What is considered as noise and what is not? The writing surely needs to be improved. If the concepts are introduced in the later sections (like in (4)), you’d better considering rearranging them to improve your narrative.\n8. Elaborate on your choice of $\\Lambda$ in (3). What is your motivation? What does it indicate? Any related literature where this metric is used and explained in detail? What is its relation to the Lyapunov condition? You say “strong numerical evidence”, how strong? Provide introductions or references for your readers to understand. \n9. Rephrase Line 304. Unclear. Better give a rigorous math expression for clarity. \n\n### Section 4\n\n1. The important variable σ is not introduced very clearly. The only related explanation is *“σ characterizing the root mean square of the off diagonal rows”*. I can hardly get what you mean. Please use a better expression. \n2. You need to show your Persian rug’s optimality with experiments. \n3. Does Eq. 8 imply we can fix b at any value and only adjust a to obtain the optimal algorithm? If not, how to select $b$?\n4. Verify \"This implies that even if ... scaling of the loss for $r>p$.\" on Line 490 with experiments.\n\n## Minor\n\n1. Always cite related works when you introduce new concepts, like Hadamard matrix, Lyapunov condition, etc.\n2. Unnecessary verbosity. For instance:\n    1. (Line 185 to 188) “…diagonals are large…as possible.” Why don’t you just say $W \\approx c I$?\n    2. (Line 22 to 25) “Unlike…oblivious”. Why should the algorithm care?\n    \n    You should flesh out your paper with rigorous and detailed definitions of your important concepts rather than these.\n    \n3. What do you mean by “r=0.25” in Fig. 2a? Is r defined on Line 424? \n4. Line 178 “p = 4.5% and ratio ns = 512.” Is it a typo? Do you mean nd/ns = 512?\n5. On line 204, please provide a clear definition of “mean-square fluctuation” or $\\Delta$. Is it simply the variance (looks so on Line 212)? In addition, variance is not sufficient for evaluating if all elements in a set are nearly the same. You’d better additionally evaluate the maximum absolute difference. I also suggest replacing variance with standard deviation.\n6. Why use $A$ rather than $a$ in Appendix A? Improve consistency.\n8. What is $\\mathbf{B}$ on Line 323 and 363? $\\mathbf{b}$ in (1) and (2)? Improve consistency.\n9. Use the correct citation command: \\citep and \\citet.\n10. Use “Fig. xx” instead of “fig. xx” to refer to figures, and \"Eq. xx\" instead of \"eq. xx\" for equations.", "questions": "1. On line 111, why must we construct $x_i$ like this? Can we sample $u_i$ from the other distributions, like uniform[-1, 0]? This still makes $x$ sparse, right? I guess sampling $u_i$ from the other distributions may not work because you AE has ReLU at the output activation? Are you assuming that $x$ are outputs of ReLU or any other positive activation? \n    * How well does this synthetic distribution of sparse data agree with the distribution of sparse data in reality? Here you are assuming its dimensions to be i.i.d, in reality this is probably not the case.\n2. You said “We must restrict to W with rank no more than nd” on Line 324. Do we need to take this rank condition into account when deriving Equation (4) and making the following reasonings from Line 332 to 348? If not, why? Could $\\nu_i$ correlate with each other due to the low rank of $W$? Are you making any i.i.d. assumption about the off-diagonal elements here?\n    * Also, this “because νi becomes Gaussian” seems to be a conjecture based on your experiment in Section 3.2. I’m not familiar with your test so I don’t know how reliable it is. That said, you’d better make this assumption stand out with something like `\\newtheorem{asp}{Assumption}`. \n3. Given your Persian rug construction of $W$, any way to retrieve $W_\\text{in}$ and $W_\\text{out}$ from it? Are they unique? Can we use the pair on Line 398? \n4. Expand your derivation of Eq. 19 in Appendix B to make it clearer. Why is there no $a$ in it (considering your definition of $\\nu$ on Line 333)? Does this equation only hold when you construct $x_i$ as described on Line 111?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work tries to derive analytical ways to obtain the near-optimal solutions of a specific type of autoencoders for synthetic sparse datasets. After theoretical analysis, it provides the process for constructing the autoencoder’s weight matrix without numerical optimization.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. Has theoretical analysis that looks legit to me, even though it is only focused on a very specific type of autoencoders on a specific type of datasets. \n2. The observation leading to theoretical analysis structure of this work can effectively convey the authors’ motivation.", "weaknesses": "The main weakness of this paper is its inconsistency and the absence of experiments for supporting their theoretical findings in Section 4, which greatly undermines the soundness of their results. Also, the authors should avoid introducing jargons and verbosity, be concise and accurate about what they mean, or provide necessary references if the jargons are widely accepted elsewhere. \nI suggest that the authors spend more time streamlining their work to improve its readability and precision, and bolster their theoretical claims with empirical results.\n\nRegarding contribution, the results in this paper might be too restricted to a specific type of problem and model. I'm not sure how crucial this problem is to LLM development, and because I also value rigorous theoretical studies of problems no matter how small they seem, I refrain from saying harsh words like \"it's trivial\". That said, the authors had better show their problem's importance with better literature review and critical discussion. \n\n## Major\n\n### Abstract\n\n1. Some claims in the abstract are not (well) reflected in the main text. For instance, “changes to the elementwise activation function or the addition of gating can at best improve its performance by a constant factor”. I guess this is about the discussion on Line 486 to 492. Experiments need to be done to verify this.\n\n### Section 1\n\n1. Line 76: Please provide clear definition of “permutation symmetric data” and “thermodynamic limit”. At least they do not appear in the work (Elhage 2022) you cited, and they sound kinda vague and unintuitive. For instance, how can “permutation symmetric” be derived from “no input feature is privileged”? As long as the features are not identical you cannot say it’s “permutation symmetric or invariant”, right? If these are jargons commonly used in your out-of-ML research area, you need to translate it for your mortal audiences in the ML community. \n\n### Section 2\n\n1. How many samples are there in your dataset on Line 111? What is the rank of this data matrix you sampled? Could the rank of this data matrix affect or even undermine your later observations and motivations? Did you scale up the number of samples as you increase $n_s$? If so, how? You need to provide more details about your experiments.\n\n### Section 3\n\n1. On line 142 there is *“due to an immediate drop in the loss starting around nd/ns ≈p. The slope and duration of this initial fall is controlled by p. In particular, in the high-sparsity regime (pclose to zero), the loss drops to zero entirely near the nd/ns ≈0 regime.”* \nI don’t think any part of Fig. 2b can clearly show this, especially considering that the upper bound of the x-axis (0.8) is way larger than p=0.05. You need to plot the loss curve at different p to demonstrate that, and zoom in on the “nd/ns ≈0 regime” and “nd/ns ≈p regime” to illustrate your points. Also, use the `subfigure` environments to enclose the subfigures.\n2. On line 152, there is *“nd/ns →∞ limit”*, but Fig. 3,4,5,6 cannot show any of such limit. All nd/ns are under 0.8.  Do you mean “nd/ns →0 limit” (although this makes no sense), or just ns →∞? Because of this, I'm uncertain when reading your statements from Line 155 to 161. \n    * If it's the latter, as I understand it, 8192 might still be kinda small. \n    * Are you always assuming that nd < ns throughout the entire paper? Be clear.\n3. Should provide a clear definition of “permutation invariant” mathematically. Also, what is “permutation symmetric” matrix? Be clear about what you say, or provide citations at least. \n4. Elaborate on “We hypothesize this occurs because for small models (and small feature probabilities) the contribution to any output from off-diagonal terms may fluctuate because of the small size of the matrix.” on Line 215. It’s like saying many things without saying anything. \n5. Elaborate on Line 236 to 239. Actually, this sentence is too convoluted to read. \n6. On Line 236 “Similar features are present in the two plots”. On Line 240 “we give three plots”. What and where are these five plots? Be specific. \n7. For Line 266 to 269 and Fig. 5. What is $\\nu$? What is $\\Delta \\text{var}(\\nu)$? What is “the fluctuation of the total norm of the off-diagonal elements in each row in W across rows”? Is it simply the variance of the off-diagonal elements’ variances? What do you mean by signal-to-noise ratio? What is considered as noise and what is not? The writing surely needs to be improved. If the concepts are introduced in the later sections (like in (4)), you’d better considering rearranging them to improve your narrative.\n8. Elaborate on your choice of $\\Lambda$ in (3). What is your motivation? What does it indicate? Any related literature where this metric is used and explained in detail? What is its relation to the Lyapunov condition? You say “strong numerical evidence”, how strong? Provide introductions or references for your readers to understand. \n9. Rephrase Line 304. Unclear. Better give a rigorous math expression for clarity. \n\n### Section 4\n\n1. The important variable σ is not introduced very clearly. The only related explanation is *“σ characterizing the root mean square of the off diagonal rows”*. I can hardly get what you mean. Please use a better expression. \n2. You need to show your Persian rug’s optimality with experiments. \n3. Does Eq. 8 imply we can fix b at any value and only adjust a to obtain the optimal algorithm? If not, how to select $b$?\n4. Verify \"This implies that even if ... scaling of the loss for $r>p$.\" on Line 490 with experiments.\n\n## Minor\n\n1. Always cite related works when you introduce new concepts, like Hadamard matrix, Lyapunov condition, etc.\n2. Unnecessary verbosity. For instance:\n    1. (Line 185 to 188) “…diagonals are large…as possible.” Why don’t you just say $W \\approx c I$?\n    2. (Line 22 to 25) “Unlike…oblivious”. Why should the algorithm care?\n    \n    You should flesh out your paper with rigorous and detailed definitions of your important concepts rather than these.\n    \n3. What do you mean by “r=0.25” in Fig. 2a? Is r defined on Line 424? \n4. Line 178 “p = 4.5% and ratio ns = 512.” Is it a typo? Do you mean nd/ns = 512?\n5. On line 204, please provide a clear definition of “mean-square fluctuation” or $\\Delta$. Is it simply the variance (looks so on Line 212)? In addition, variance is not sufficient for evaluating if all elements in a set are nearly the same. You’d better additionally evaluate the maximum absolute difference. I also suggest replacing variance with standard deviation.\n6. Why use $A$ rather than $a$ in Appendix A? Improve consistency.\n8. What is $\\mathbf{B}$ on Line 323 and 363? $\\mathbf{b}$ in (1) and (2)? Improve consistency.\n9. Use the correct citation command: \\citep and \\citet.\n10. Use “Fig. xx” instead of “fig. xx” to refer to figures, and \"Eq. xx\" instead of \"eq. xx\" for equations.", "questions": "1. On line 111, why must we construct $x_i$ like this? Can we sample $u_i$ from the other distributions, like uniform[-1, 0]? This still makes $x$ sparse, right? I guess sampling $u_i$ from the other distributions may not work because you AE has ReLU at the output activation? Are you assuming that $x$ are outputs of ReLU or any other positive activation? \n    * How well does this synthetic distribution of sparse data agree with the distribution of sparse data in reality? Here you are assuming its dimensions to be i.i.d, in reality this is probably not the case.\n2. You said “We must restrict to W with rank no more than nd” on Line 324. Do we need to take this rank condition into account when deriving Equation (4) and making the following reasonings from Line 332 to 348? If not, why? Could $\\nu_i$ correlate with each other due to the low rank of $W$? Are you making any i.i.d. assumption about the off-diagonal elements here?\n    * Also, this “because νi becomes Gaussian” seems to be a conjecture based on your experiment in Section 3.2. I’m not familiar with your test so I don’t know how reliable it is. That said, you’d better make this assumption stand out with something like `\\newtheorem{asp}{Assumption}`. \n3. Given your Persian rug construction of $W$, any way to retrieve $W_\\text{in}$ and $W_\\text{out}$ from it? Are they unique? Can we use the pair on Line 398? \n4. Expand your derivation of Eq. 19 in Appendix B to make it clearer. Why is there no $a$ in it (considering your definition of $\\nu$ on Line 333)? Does this equation only hold when you construct $x_i$ as described on Line 111?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730419095954}, {"id": "KLgIyZTfvW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_dnUr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper studies the following auto-encoder model: linear encoding + linear decoding with additional denoising (shifted-ReLU) application on top. This model is applied in context of compression of i.i.d. sparse uniform data. The authors provide an analysis of the model behaviour (MSE bound) and identify the set of matrices which matches the performance of the trained model.", "review_text": "The paper studies the following auto-encoder model: linear encoding + linear decoding with additional denoising (shifted-ReLU) application on top. This model is applied in context of compression of i.i.d. sparse uniform data. The authors provide an analysis of the model behaviour (MSE bound) and identify the set of matrices which matches the performance of the trained model.", "strengths": "The paper provides to some extent a solid analysis of the proposed model.", "weaknesses": "The whole interpretability premise is evidently far fetched especially whenever LLMs appear in the related context. The paper does not provide any evidence that the phenomenology observed in their i.i.d. + small model setting appears on any even \"toy\"-ish real data, or justify enough whether the prescribed methodology could be put forward to gain an understanding of the behaviour of the real world models. The suggestion is to reframe and focus on the subject of substance - analysis of the autoencoder model.\n\nI fail to see why so much of the paper space is occupied by a the \"Gaussian-ity\" experiments in Section 3.2. The authors do no provide any rigorous training results anyway (that the model will converge to a specific subset of matrices). In this view, since the proposed behaviour is assumed anyway using the empirical evidence, the amount of histograms and the accompanying description is clearly an overshot (which does not really add any further weight to the submission).\n\nThe technical machinery used for the analysis is not of a separate interest. \n\nThe main concern, however, is the emergence of a certain structure in the weights. In particular, it is completely unclear whether the \"Persian carpet\" structure is indeed favoured by the network implicitly or just a set of weights which achieves the close to optimal MSE.\nIf later, the observations in the paper does not shed any light on any sort of \"interpretability\", since the network does not implement the proposed weight pattern when trained. Prior, however, is not supported by any evidence (no SGD/gradient flow convergence result).\n\nThe phenomena of uniform enough weights is not surprising on its own given that the signal is i.i.d., especially in the view of related work.\nSpecifically, could authors provide any intuition why specifically \"Persian carpet\" (except visual appeal)? There is no evidence whatsoever that a simpler uniform design will not achieve desired. Why not use some other design: for instance, rotationally invariant matricies with specific spectra (if the input distribution is centered, see next paragraph) and exploit Approximate Message Passing analysis (granted, the optimal denoising will not be close to ReLU)?\n\nAnother concern: why the data is not centered (which is *common practice*)? It is hard to judge, in this case, how much of an effect on the proposed structure it has, in the view of ReLU - which has cut-off at zero. \n\nI also assume that in equation (2) there is normalization in dimension missing. Otherwise, dropping \"i\" in (5) does not make sense.\n\nLastly, the particular choice of ReLU model (1) in the view of above remains quite unclear. It is hard to judge if any of the observations are not an artifact of the modeling.\n\nThe authors fail to acknowledge *a lot* of the related literature which is non-trivially connected to the observed model behaviour. Below, a non-exhaustive list:\n-  AEs: (Refinetti & Goldt, 2022); (Cui & Zdeborova, 2023); \"Analysis of feature learning in weight-tied autoencoders via the mean field lens\" Nguyen (2021); (Shevchenko et al., 2023); (Kögler et al., 2024);\n- Approximate Message Passing algorithms: (Donoho et al., 2009); rotationally invariant design (Rangan et al., 2019; Schniter et al., 2016; Ma & Ping, 2017; Takeuchi, 2019); optimal spectral design (Ma et al., 2021).", "questions": "See the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the following auto-encoder model: linear encoding + linear decoding with additional denoising (shifted-ReLU) application on top. This model is applied in context of compression of i.i.d. sparse uniform data. The authors provide an analysis of the model behaviour (MSE bound) and identify the set of matrices which matches the performance of the trained model.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "The paper provides to some extent a solid analysis of the proposed model.", "weaknesses": "The whole interpretability premise is evidently far fetched especially whenever LLMs appear in the related context. The paper does not provide any evidence that the phenomenology observed in their i.i.d. + small model setting appears on any even \"toy\"-ish real data, or justify enough whether the prescribed methodology could be put forward to gain an understanding of the behaviour of the real world models. The suggestion is to reframe and focus on the subject of substance - analysis of the autoencoder model.\n\nI fail to see why so much of the paper space is occupied by a the \"Gaussian-ity\" experiments in Section 3.2. The authors do no provide any rigorous training results anyway (that the model will converge to a specific subset of matrices). In this view, since the proposed behaviour is assumed anyway using the empirical evidence, the amount of histograms and the accompanying description is clearly an overshot (which does not really add any further weight to the submission).\n\nThe technical machinery used for the analysis is not of a separate interest. \n\nThe main concern, however, is the emergence of a certain structure in the weights. In particular, it is completely unclear whether the \"Persian carpet\" structure is indeed favoured by the network implicitly or just a set of weights which achieves the close to optimal MSE.\nIf later, the observations in the paper does not shed any light on any sort of \"interpretability\", since the network does not implement the proposed weight pattern when trained. Prior, however, is not supported by any evidence (no SGD/gradient flow convergence result).\n\nThe phenomena of uniform enough weights is not surprising on its own given that the signal is i.i.d., especially in the view of related work.\nSpecifically, could authors provide any intuition why specifically \"Persian carpet\" (except visual appeal)? There is no evidence whatsoever that a simpler uniform design will not achieve desired. Why not use some other design: for instance, rotationally invariant matricies with specific spectra (if the input distribution is centered, see next paragraph) and exploit Approximate Message Passing analysis (granted, the optimal denoising will not be close to ReLU)?\n\nAnother concern: why the data is not centered (which is *common practice*)? It is hard to judge, in this case, how much of an effect on the proposed structure it has, in the view of ReLU - which has cut-off at zero. \n\nI also assume that in equation (2) there is normalization in dimension missing. Otherwise, dropping \"i\" in (5) does not make sense.\n\nLastly, the particular choice of ReLU model (1) in the view of above remains quite unclear. It is hard to judge if any of the observations are not an artifact of the modeling.\n\nThe authors fail to acknowledge *a lot* of the related literature which is non-trivially connected to the observed model behaviour. Below, a non-exhaustive list:\n-  AEs: (Refinetti & Goldt, 2022); (Cui & Zdeborova, 2023); \"Analysis of feature learning in weight-tied autoencoders via the mean field lens\" Nguyen (2021); (Shevchenko et al., 2023); (Kögler et al., 2024);\n- Approximate Message Passing algorithms: (Donoho et al., 2009); rotationally invariant design (Rangan et al., 2019; Schniter et al., 2016; Ma & Ping, 2017; Takeuchi, 2019); optimal spectral design (Ma et al., 2021).", "questions": "See the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730394184365}, {"id": "gxI5f86QX9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13483/Reviewer_PspC"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper examines a toy model of superposition, revealing that when the input data exhibits permutation symmetry, the auto-encoder learns an algorithm responsive only to macroscopic parameters. Using this insight, the authors derive upper bounds on the reconstruction loss and design hand-crafted symmetric models that match the loss of the trained model.", "review_text": "This paper examines a toy model of superposition, revealing that when the input data exhibits permutation symmetry, the auto-encoder learns an algorithm responsive only to macroscopic parameters. Using this insight, the authors derive upper bounds on the reconstruction loss and design hand-crafted symmetric models that match the loss of the trained model.", "strengths": "- This paper presents a novel study on the effects of input symmetry in superposition models.  \n- This paper identifies a statistical permutation symmetry in the learned model weights, which is verified by designing a \"Persian rug model\" that achieves similar performance to the trained model.  \n- The authors provide theoretical upper bounds for the reconstruction loss, which their experiments confirm by showing a notable drop in loss around $r\\approx p$.  \n- Detailed explanations throughout the paper improve the understanding and interpretation of the experimental results.", "weaknesses": "- **Writing**: The empirical observations in the paper feel overly repetitive, while the theoretical sections come across as overly technical and challenging to follow. For example, Section 3 spends three pages describing that the diagonal elements and bias become uniform while off-diagonal elements become noisy. Section 4 includes excessive technical detail in the derivations, which could be condensed into formal theorems, with the technical details moved to the appendix for easier readability.\n\n- **Theory**: Many theoretical discussions in this paper lack rigor. For instance, the discussion between lines 353 and 361 is confusing and does not seem to follow rigorous mathematical reasoning. I will provide further points of confusion in the questions section. Notably, as a theoretically-focused paper, it lacks any formal theorems, making it difficult to digest. Summarizing results into well-defined theorems with clearly explained notations would improve clarity.\n\n- **Experiments**: The experimental scope appears limited. The paper only presents results on a toy model with input data following a sparse, uniform, permutation-symmetric distribution. However, real-world data, such as natural language, does not exhibit permutation symmetry. Including experiments on real-world data would help demonstrate the applicability of these results.\n\nIn summary, the experimental scope of this paper seems narrow, and the theoretical sections lack the rigor expected of a theoretical work. I recommend that the authors refine their presentation of theoretical results and expand their experiments. I’m open to any clarifications if there has been a misunderstanding of the paper's theory from my end.", "questions": "- In Figure 1, you show that the structure of the artificial weights resembles a Persian rug. Out of curiosity, what do the trained weights look like?\n\n- On line 155, you mention that in the $r \\to \\infty$ limit, the diagonal elements of $\\mathbf{W}$ become the same, while the off-diagonal elements become zero-mean noise, resulting in $\\mathbf{W} = a \\mathbf{I}$. However, since $\\mathbf{W}$ is the product of two low-rank matrices, shouldn’t it have a rank no greater than $n_d$? How, then, does it approach a full-rank matrix?\n\n- On line 355, the symbol $\\mu$ is not defined. Could you also clarify the main message of this paragraph?\n\n- On line 406, the parity function is not defined.\n\n- On line 410, should the LHS be $R_{ij}$? It would also be helpful to include a theorem stating that $\\mathbf{R}$ satisfies the statistical symmetries, with a proof deferred to the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper examines a toy model of superposition, revealing that when the input data exhibits permutation symmetry, the auto-encoder learns an algorithm responsive only to macroscopic parameters. Using this insight, the authors derive upper bounds on the reconstruction loss and design hand-crafted symmetric models that match the loss of the trained model.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- This paper presents a novel study on the effects of input symmetry in superposition models.  \n- This paper identifies a statistical permutation symmetry in the learned model weights, which is verified by designing a \"Persian rug model\" that achieves similar performance to the trained model.  \n- The authors provide theoretical upper bounds for the reconstruction loss, which their experiments confirm by showing a notable drop in loss around $r\\approx p$.  \n- Detailed explanations throughout the paper improve the understanding and interpretation of the experimental results.", "weaknesses": "- **Writing**: The empirical observations in the paper feel overly repetitive, while the theoretical sections come across as overly technical and challenging to follow. For example, Section 3 spends three pages describing that the diagonal elements and bias become uniform while off-diagonal elements become noisy. Section 4 includes excessive technical detail in the derivations, which could be condensed into formal theorems, with the technical details moved to the appendix for easier readability.\n\n- **Theory**: Many theoretical discussions in this paper lack rigor. For instance, the discussion between lines 353 and 361 is confusing and does not seem to follow rigorous mathematical reasoning. I will provide further points of confusion in the questions section. Notably, as a theoretically-focused paper, it lacks any formal theorems, making it difficult to digest. Summarizing results into well-defined theorems with clearly explained notations would improve clarity.\n\n- **Experiments**: The experimental scope appears limited. The paper only presents results on a toy model with input data following a sparse, uniform, permutation-symmetric distribution. However, real-world data, such as natural language, does not exhibit permutation symmetry. Including experiments on real-world data would help demonstrate the applicability of these results.\n\nIn summary, the experimental scope of this paper seems narrow, and the theoretical sections lack the rigor expected of a theoretical work. I recommend that the authors refine their presentation of theoretical results and expand their experiments. I’m open to any clarifications if there has been a misunderstanding of the paper's theory from my end.", "questions": "- In Figure 1, you show that the structure of the artificial weights resembles a Persian rug. Out of curiosity, what do the trained weights look like?\n\n- On line 155, you mention that in the $r \\to \\infty$ limit, the diagonal elements of $\\mathbf{W}$ become the same, while the off-diagonal elements become zero-mean noise, resulting in $\\mathbf{W} = a \\mathbf{I}$. However, since $\\mathbf{W}$ is the product of two low-rank matrices, shouldn’t it have a rank no greater than $n_d$? How, then, does it approach a full-rank matrix?\n\n- On line 355, the symbol $\\mu$ is not defined. Could you also clarify the main message of this paragraph?\n\n- On line 406, the parity function is not defined.\n\n- On line 410, should the LHS be $R_{ij}$? It would also be helpful to include a theorem stating that $\\mathbf{R}$ satisfies the statistical symmetries, with a proof deferred to the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730362208374}], "openreview_url": "https://openreview.net/forum?id=rapXZIfwbX", "arxiv_id": "2410.12101", "paper_pdf": "papers/rapXZIfwbX.pdf", "paper_pdf_sha256": "75d369cb48f910cb471f7dc5f9adce8da4e3c89d5bf5cc5dad8b9cd733dd86d6", "paper_pdf_bytes": 5818731, "paper_pdf_source": "openreview", "code_url": "https://github.com/KfirD/PersianRug", "code_repository": "KfirD/PersianRug", "code_commit": "e178d0bbe5acd1da0a8641a14ca4c3cf27daafe9", "code_archive": "repos/rapXZIfwbX.zip", "code_archive_sha256": "e3ec13a02e1930461323f05038665b8975b5974e1fe18b0f08d8765d72d63e18", "code_archive_bytes": 126195, "code_file_count": 15, "code_extensions": {".py": 14, ".ipynb": 1}, "github_disk_usage_kb": 134, "github_languages": {"Jupyter Notebook": 140702, "Python": 116782}, "github_archived": false, "github_pushed_at": "2024-12-07T22:13:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-persian-rug-solving-toy-models-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DZyhUXpEee", "year": 2024, "status": "rejected", "title": "SpaFL: Communication-Efficient Federated Learning with Sparse Models and Low Computational Overhead", "authors": ["Minsu Kim", "Walid Saad", "Merouane Abdelkader DEBBAH", "Choong Seon Hong"], "authorids": ["~Minsu_Kim6", "~Walid_Saad1", "~Merouane_Abdelkader_DEBBAH1", "~Choong_Seon_Hong1"], "authors_source": "OpenReview API", "abstract": "The large communication and computation overhead of federated learning (FL) is one of the main challenges facing its practical deployment over resource-constrained clients and systems. In this work, SpaFL: a communication-efficient FL framework is proposed to optimize both personalized model parameters and sparse model structures  with low computational overhead. In SpaFL, a trainable threshold is defined for each neuron/filter to prune its connected parameters. Both model parameters and thresholds are jointly optimized to enable the automatic sparsification of the models while recovering prematurely pruned parameters during training. To reduce communication costs, only thresholds are communicated between a server and clients instead of parameters, thereby enabling the clients to learn how to prune. Further, global thresholds are used to update model parameters by extracting aggregated parameter importance. The convergence of SpaFL is analyzed, and the results provide new insights into the tradeoff between computation overhead and learning performance. Experimental results show that SpaFL improves accuracy while requiring much less communication and computing resources compared to both dense and sparse personalized baselines.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "i5mdmQLoec", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6217/Reviewer_Ggae"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose a novel federated learning approach with sparse personalized client models. The main technical contribution is the reduction of communication overhead by only communicating the thresholds used to determine non-zero parameters and the reduction of computation cost by joint optimization of thresholds and sparse client models. The convergence of the approach is theoretically analyzed and experiments on a range of datasets demonstrate improvements over several prior works.", "review_text": "The authors propose a novel federated learning approach with sparse personalized client models. The main technical contribution is the reduction of communication overhead by only communicating the thresholds used to determine non-zero parameters and the reduction of computation cost by joint optimization of thresholds and sparse client models. The convergence of the approach is theoretically analyzed and experiments on a range of datasets demonstrate improvements over several prior works.", "strengths": "1. The proposed approach will significantly reduce communication cost since it only involves communicating one threshold per neuron and number of neurons is much less than the number of parameters.\n\n2. The empirical results also show a significant reduction in FLOPs due to the sparsity of the models being optimized on the clients.", "weaknesses": "1. Some aspects of the algorithm are not clearly explained. It is not clear why the regularizer in (4) is chosen over other options. It is also not clear why the second update of the client models in 3.2.4 is necessary because technically it should be possible to continue training the model with the new global thresholds as described in 3.2.2.\n\n2.The sparsification is unstructured and thus may not actually lead to reduction in computation cost or latency due to inefficient utilization of the hardware. Since there are already several works on sparse FL I believe it is important to now start considering hardware performance etc to truly differentiate from prior work.", "questions": "1. In addition to an intuitive justification for the regularizer in (4) and the second update in 3.2.4 can you also provide an empirical comparison with alternate regularizers?\n\n2. Likewise can you also provide a comparison with a baseline which does not use the update in 3.2.4 but instead just directly continues training the model with the new global thresholds?\n\n3. Can you provide a derivation for (6) and (11)?\n\n4. From (10) and (11) if the gradient direction of $w$ is opposite to that of its connected threshold if $w>0$ then shouldn't the gradient direction of $w$ and $\\Delta \\tau$ be the opposite if $w>0$ and not same as is claimed in the paragraph after 11? Please clarify.\n\n5. What is the model density of the baselines in Table 1? I do not see it presented in the table.\n \n6. Do you have any thoughts on how the sparsification approach described herein could be made structured and thus more hardware efficient?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a novel federated learning approach with sparse personalized client models. The main technical contribution is the reduction of communication overhead by only communicating the thresholds used to determine non-zero parameters and the reduction of computation cost by joint optimization of thresholds and sparse client models. The convergence of the approach is theoretically analyzed and experiments on a range of datasets demonstrate improvements over several prior works.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The proposed approach will significantly reduce communication cost since it only involves communicating one threshold per neuron and number of neurons is much less than the number of parameters.\n\n2. The empirical results also show a significant reduction in FLOPs due to the sparsity of the models being optimized on the clients.", "weaknesses": "1. Some aspects of the algorithm are not clearly explained. It is not clear why the regularizer in (4) is chosen over other options. It is also not clear why the second update of the client models in 3.2.4 is necessary because technically it should be possible to continue training the model with the new global thresholds as described in 3.2.2.\n\n2.The sparsification is unstructured and thus may not actually lead to reduction in computation cost or latency due to inefficient utilization of the hardware. Since there are already several works on sparse FL I believe it is important to now start considering hardware performance etc to truly differentiate from prior work.", "questions": "1. In addition to an intuitive justification for the regularizer in (4) and the second update in 3.2.4 can you also provide an empirical comparison with alternate regularizers?\n\n2. Likewise can you also provide a comparison with a baseline which does not use the update in 3.2.4 but instead just directly continues training the model with the new global thresholds?\n\n3. Can you provide a derivation for (6) and (11)?\n\n4. From (10) and (11) if the gradient direction of $w$ is opposite to that of its connected threshold if $w>0$ then shouldn't the gradient direction of $w$ and $\\Delta \\tau$ be the opposite if $w>0$ and not same as is claimed in the paragraph after 11? Please clarify.\n\n5. What is the model density of the baselines in Table 1? I do not see it presented in the table.\n \n6. Do you have any thoughts on how the sparsification approach described herein could be made structured and thus more hardware efficient?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699372822615}, {"id": "uxn2QIRu24", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6217/Reviewer_HLPa"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes SpaFL to tackle the communication cost problem in federated learning. To find the sparse mask for the model, authors introduce a new parameter called threshold ($\\tau$). This parameter indicates if a weight in the model is active or nonactive, hence can reduce the density of the model. In each round, the clients find the current mask based on $\\tau$, then update their local weight, and finally update the local $\\tau$. After the local step, the server aggregates the client's $\\tau$ and transmits the new value to all the clients. Then, all the clients update their weights accordingly.", "review_text": "This paper proposes SpaFL to tackle the communication cost problem in federated learning. To find the sparse mask for the model, authors introduce a new parameter called threshold ($\\tau$). This parameter indicates if a weight in the model is active or nonactive, hence can reduce the density of the model. In each round, the clients find the current mask based on $\\tau$, then update their local weight, and finally update the local $\\tau$. After the local step, the server aggregates the client's $\\tau$ and transmits the new value to all the clients. Then, all the clients update their weights accordingly.", "strengths": "* The problem is well-motivated.\n* Authors provides theoretical proof for the convergence of their method.\n* The method is novel and saves uplink communication costs for the clients.\n* The solution is novel.", "weaknesses": "* What happens if clients do not receive the server update due to unavailability? It is specifically important as the solutions is designed  for resource-constrained cross-device FL, where clients are only sometimes available. \n* How do non-participant clients update their model?\n* Is there any global model available?\n* The author should include a comparison with prior works that adapt sparse learning in FL, such as [8,9,22,24,25,26] (references are from this paper). Some of these methods can reach high sparsities comparable to the 1% communication cost of SpaFL.\n* How does the server or clients control the density of the models.\n* How does SpaFL perform when the global model is denser (for example ResNet18)?", "questions": "* The questions can be found in weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes SpaFL to tackle the communication cost problem in federated learning. To find the sparse mask for the model, authors introduce a new parameter called threshold ($\\tau$). This parameter indicates if a weight in the model is active or nonactive, hence can reduce the density of the model. In each round, the clients find the current mask based on $\\tau$, then update their local weight, and finally update the local $\\tau$. After the local step, the server aggregates the client's $\\tau$ and transmits the new value to all the clients. Then, all the clients update their weights accordingly.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "* The problem is well-motivated.\n* Authors provides theoretical proof for the convergence of their method.\n* The method is novel and saves uplink communication costs for the clients.\n* The solution is novel.", "weaknesses": "* What happens if clients do not receive the server update due to unavailability? It is specifically important as the solutions is designed  for resource-constrained cross-device FL, where clients are only sometimes available. \n* How do non-participant clients update their model?\n* Is there any global model available?\n* The author should include a comparison with prior works that adapt sparse learning in FL, such as [8,9,22,24,25,26] (references are from this paper). Some of these methods can reach high sparsities comparable to the 1% communication cost of SpaFL.\n* How does the server or clients control the density of the models.\n* How does SpaFL perform when the global model is denser (for example ResNet18)?", "questions": "* The questions can be found in weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698726560728}, {"id": "qybGQI9cNG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6217/Reviewer_R775"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a method to mitigate the communication and computation overhead in FL. It employs a threshold-based approach to simultaneously optimize sparse masks and model parameters, resulting in reduced communication costs and the attainment of better personalized models. The paper offers a theoretical analysis regarding the convergence of the proposed method, while empirical results further confirm its efficacy.", "review_text": "This paper introduces a method to mitigate the communication and computation overhead in FL. It employs a threshold-based approach to simultaneously optimize sparse masks and model parameters, resulting in reduced communication costs and the attainment of better personalized models. The paper offers a theoretical analysis regarding the convergence of the proposed method, while empirical results further confirm its efficacy.", "strengths": "1. The proposed approach employs a straightforward method based on transmitting threshold to effectively reduce communication bandwidth while surpassing the accuracy of personalized models over baseline methods.\n\n2. The paper substantiates the effectiveness of the proposed method through comprehensive empirical evaluations and theoretical analysis.\n\n3. The paper is well-written and easy to read.", "weaknesses": "1. Some implementation details need further clarification. (Q1)\n\n2. The intuition from the theory needs more elaboration. (Q2)\n\n3. This study primarily concentrates on personalized federated learning, where no global model is trained. It would enhance clarity if the authors could differentiate between personalized federated learning (pFL) and federated learning (FL), as the paper references FL multiple times, which typically involves a global model.", "questions": "1. During the local training for parameters and thresholds, if the gradients are calculated for every weight and applied with the binary mask, then how does it help save the computation overhead? Or if the gradients are calculated w.r.t. the sparse weights, how is it achieved in practice?\n\n2. The interpretation of the third term in Theorem 1 is not straightforward. The loss function $F_k$ is not bounded in the paper, and it can potentially assume arbitrarily large values, rendering the third term of Theorem 1 indeterminate. How to understand Theorem 1 as a valid convergence bound?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a method to mitigate the communication and computation overhead in FL. It employs a threshold-based approach to simultaneously optimize sparse masks and model parameters, resulting in reduced communication costs and the attainment of better personalized models. The paper offers a theoretical analysis regarding the convergence of the proposed method, while empirical results further confirm its efficacy.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The proposed approach employs a straightforward method based on transmitting threshold to effectively reduce communication bandwidth while surpassing the accuracy of personalized models over baseline methods.\n\n2. The paper substantiates the effectiveness of the proposed method through comprehensive empirical evaluations and theoretical analysis.\n\n3. The paper is well-written and easy to read.", "weaknesses": "1. Some implementation details need further clarification. (Q1)\n\n2. The intuition from the theory needs more elaboration. (Q2)\n\n3. This study primarily concentrates on personalized federated learning, where no global model is trained. It would enhance clarity if the authors could differentiate between personalized federated learning (pFL) and federated learning (FL), as the paper references FL multiple times, which typically involves a global model.", "questions": "1. During the local training for parameters and thresholds, if the gradients are calculated for every weight and applied with the binary mask, then how does it help save the computation overhead? Or if the gradients are calculated w.r.t. the sparse weights, how is it achieved in practice?\n\n2. The interpretation of the third term in Theorem 1 is not straightforward. The loss function $F_k$ is not bounded in the paper, and it can potentially assume arbitrarily large values, rendering the third term of Theorem 1 indeterminate. How to understand Theorem 1 as a valid convergence bound?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698725987701}, {"id": "APQaNjJnP7", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6217/Reviewer_Kuri"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The large communication overhead of FL is one of the main challenges. This work proposes SpaFL to optimize both personalized model parameters and sparse model structures. SpaFL defines a trainable threshold for each neuron/filter for pruning. Both model parameters and thresholds are jointly optimized, thus those prematurely pruned parameters during training can be recovered. Only thresholds are communicated between a server and clients instead of parameters, thereby enabling the clients to learn how to prune and reducing communication costs. Global thresholds are used to update model parameters by extracting aggregated parameter importance.", "review_text": "The large communication overhead of FL is one of the main challenges. This work proposes SpaFL to optimize both personalized model parameters and sparse model structures. SpaFL defines a trainable threshold for each neuron/filter for pruning. Both model parameters and thresholds are jointly optimized, thus those prematurely pruned parameters during training can be recovered. Only thresholds are communicated between a server and clients instead of parameters, thereby enabling the clients to learn how to prune and reducing communication costs. Global thresholds are used to update model parameters by extracting aggregated parameter importance.", "strengths": "1. Only communicating thresholds is novel, and reducing communication costs of both up-link and down-link a lot.\n2. Equation (8) provides a good connection between the importance and the thresholds.\n3. There is convergence analysis of the SpaFL.\n4. Experiment results show significant improvements of SpaFL.", "weaknesses": "1. Section 3.2.2 needs to be written more clear. During local training with e < E, does the thresholds not be updated? Equation (5) and (6) only work for the e=E? And how the equation (6) is derived?\n2. Comparing equation (10) and (11), authors concludes the relationship between the gradient direction and the $\\Delta \\tau$ when $w >0$ or $w < 0$. However, Equation (9) is about the global $\\tau$, while equation (10) is talking about local $\\tau$. Could authos explain this in more details?\n3. Experiment settings are not clear enough. The training dataset is split using Dirichlet samplg. For personalized FL, how are test datasets split and how the models are tested? Why all methods use the same learning rates? Maybe different methods have different best learning rates.\n4. The theoretical proof does not consider the data heterogeneity. Will the thresholds still converge under the data heterogeneity?", "questions": "1. See weakness 1, During local training with e < E, does the thresholds not be updated? Equation (5) and (6) only work for the e=E? And how the equation (6) is derived?\n2. See weakness 2.\n3. See weakness 3.\n4. See weakness 4.\n5. In experiments, E = 3 means local iterations = 3, or local epochs = 3? local iteration = 3 seems to be too small. Could you find other references to support this setting? Because many FL works set this as epochs [1]. \n\n\n[1] Communication-Efficient Learning of Deep Networks from Decentralized Data.\n[2] SCAFFOLD: Stochastic Controlled Averaging for Federated Learning.\n[3] Adaptive Federated Optimization.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The large communication overhead of FL is one of the main challenges. This work proposes SpaFL to optimize both personalized model parameters and sparse model structures. SpaFL defines a trainable threshold for each neuron/filter for pruning. Both model parameters and thresholds are jointly optimized, thus those prematurely pruned parameters during training can be recovered. Only thresholds are communicated between a server and clients instead of parameters, thereby enabling the clients to learn how to prune and reducing communication costs. Global thresholds are used to update model parameters by extracting aggregated parameter importance.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. Only communicating thresholds is novel, and reducing communication costs of both up-link and down-link a lot.\n2. Equation (8) provides a good connection between the importance and the thresholds.\n3. There is convergence analysis of the SpaFL.\n4. Experiment results show significant improvements of SpaFL.", "weaknesses": "1. Section 3.2.2 needs to be written more clear. During local training with e < E, does the thresholds not be updated? Equation (5) and (6) only work for the e=E? And how the equation (6) is derived?\n2. Comparing equation (10) and (11), authors concludes the relationship between the gradient direction and the $\\Delta \\tau$ when $w >0$ or $w < 0$. However, Equation (9) is about the global $\\tau$, while equation (10) is talking about local $\\tau$. Could authos explain this in more details?\n3. Experiment settings are not clear enough. The training dataset is split using Dirichlet samplg. For personalized FL, how are test datasets split and how the models are tested? Why all methods use the same learning rates? Maybe different methods have different best learning rates.\n4. The theoretical proof does not consider the data heterogeneity. Will the thresholds still converge under the data heterogeneity?", "questions": "1. See weakness 1, During local training with e < E, does the thresholds not be updated? Equation (5) and (6) only work for the e=E? And how the equation (6) is derived?\n2. See weakness 2.\n3. See weakness 3.\n4. See weakness 4.\n5. In experiments, E = 3 means local iterations = 3, or local epochs = 3? local iteration = 3 seems to be too small. Could you find other references to support this setting? Because many FL works set this as epochs [1]. \n\n\n[1] Communication-Efficient Learning of Deep Networks from Decentralized Data.\n[2] SCAFFOLD: Stochastic Controlled Averaging for Federated Learning.\n[3] Adaptive Federated Optimization.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698511992030}], "openreview_url": "https://openreview.net/forum?id=DZyhUXpEee", "arxiv_id": "2406.00431", "paper_pdf": "papers/DZyhUXpEee.pdf", "paper_pdf_sha256": "f956c908d682235d1bd2da9f795a959f9f617cad646fe186abc9de5de0523249", "paper_pdf_bytes": 1906004, "paper_pdf_source": "openreview", "code_url": "https://github.com/news-vt/SpaFL_NeruIPS_2024", "code_repository": "news-vt/SpaFL_NeruIPS_2024", "code_commit": "d3639ab3c6ecae3a3d09cf897b264fc44551a56c", "code_archive": "repos/DZyhUXpEee.zip", "code_archive_sha256": "ace2a4816fba57d6b6bad61cf84fae411e323aef7f9ee1db8bc8908ff3e17cff", "code_archive_bytes": 171180, "code_file_count": 34, "code_extensions": {".py": 34}, "github_disk_usage_kb": 111, "github_languages": {"Python": 135364}, "github_archived": false, "github_pushed_at": "2025-01-15T16:22:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/spafl-communication-efficient-federated"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1EVPT82ttr", "year": 2023, "status": "rejected", "title": "Learning Unified Representations for Multi-Resolution Face Recognition", "authors": ["Hulingxiao He", "Wu Yuan", "Yidian Huang", "Shilong Zhao", "Wen Yuan", "HanQing Li"], "authorids": ["~Hulingxiao_He1", "~Wu_Yuan1", "~Yidian_Huang1", "~Shilong_Zhao1", "~Wen_Yuan1", "~HanQing_Li2"], "authors_source": "OpenReview API", "abstract": "In this work, we propose Branch-to-Trunk network (BTNet), a novel representation learning method for multi-resolution face recognition. It consists of a trunk network (TNet), namely a unified encoder, and multiple branch networks (BNets), namely resolution adapters. As per the input, a resolution-specific BNet is used and the output are implanted as feature maps in the feature pyramid of TNet, at a layer with the same resolution. The discriminability of tiny faces is significantly improved, as the interpolation error introduced by rescaling, especially up-sampling, is mitigated on the inputs. With branch distillation and backward-compatible training, BTNet transfers discriminative high-resolution information to multiple branches while guaranteeing representation compatibility. Our experiments demonstrate strong performance on face recognition benchmarks, both for multi-resolution face verification and face identification, with much less computation amount and parameter storage. We establish new state-of-the-art on the challenging QMUL-SurvFace 1: N face identification task.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rKr0EoFGpR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3517/Reviewer_yRHk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a Branch-to-Trunk network (BTNet) for multi-resolution face recognition, which consists of a trunk network (TNet) and multiple branch networks (BNets). With branch distillation and backward compatible training, BTNet transfers discriminative high-resolution information to multiple branches. Experiments on face recognition benchmarks show the better performance of the proposed BTNet for multi-resolution face verification and face identification.", "review_text": "The main concerns of the work are its weaknesses as\n- Waht are contributions of the proposed BTNet compared to existing dynamic resolution network (Zhu et al., NeurIPS 2021)? And the paper should compare with dynamic resolution network.\n- The paper evaluate the proposed BTNet on the challenging QMUL-SurvFace dataset, and it is recommend to conduct comparisons on more fae recognition benchmarks, e.g., WebFace260M [1].", "strengths": "Strength:\n+ The paper is well written and easy to read.\n+ The proposed BTNet obtains the state-of-the-art performance on the challenging QMUL-SurvFace 1: N face identification task.\n\nWeaknesses:\n- Waht are contributions of the proposed BTNet compared to existing dynamic resolution network (Zhu et al., NeurIPS 2021)? And the paper should compare with dynamic resolution network.\n- The paper evaluate the proposed BTNet on the challenging QMUL-SurvFace dataset, and it is recommend to conduct comparisons on more fae recognition benchmarks, e.g., WebFace260M [1].\n[1] Zhu et al., WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition, CVPR 2021.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a Branch-to-Trunk network (BTNet) for multi-resolution face recognition, which consists of a trunk network (TNet) and multiple branch networks (BNets). With branch distillation and backward compatible training, BTNet transfers discriminative high-resolution information to multiple branches. Experiments on face recognition benchmarks show the better performance of the proposed BTNet for multi-resolution face verification and face identification.", "strength_and_weaknesses": "Strength:\n+ The paper is well written and easy to read.\n+ The proposed BTNet obtains the state-of-the-art performance on the challenging QMUL-SurvFace 1: N face identification task.\n\nWeaknesses:\n- Waht are contributions of the proposed BTNet compared to existing dynamic resolution network (Zhu et al., NeurIPS 2021)? And the paper should compare with dynamic resolution network.\n- The paper evaluate the proposed BTNet on the challenging QMUL-SurvFace dataset, and it is recommend to conduct comparisons on more fae recognition benchmarks, e.g., WebFace260M [1].\n[1] Zhu et al., WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition, CVPR 2021.", "clarity,_quality,_novelty_and_reproducibility": "The quality and clarity is good, and the originality of the work is fair.", "summary_of_the_review": "The main concerns of the work are its weaknesses as\n- Waht are contributions of the proposed BTNet compared to existing dynamic resolution network (Zhu et al., NeurIPS 2021)? And the paper should compare with dynamic resolution network.\n- The paper evaluate the proposed BTNet on the challenging QMUL-SurvFace dataset, and it is recommend to conduct comparisons on more fae recognition benchmarks, e.g., WebFace260M [1].", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666686953153}, {"id": "mheZL_WeER", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3517/Reviewer_Lmwt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a Branch-to-Trunk network for multi-resolution face recognition. In this paper, by setting multiple branches and inputting images with different resolutions into different branches, the interpolation error caused by up-sampling is reduced. It requires less computation amount and parameter storage. The paper conducts experiments on six face verification benchmarks and QMUL-SurvFace dataset.", "review_text": "See Strength And Weaknesses.", "strengths": "Strength:\nMost low-resolution face recognition works require up-sampling low-resolution images to a fixed size (112*112). This paper proposes a multi-branch network to reduce the interpolation error by piecewise processing the input image according to the resolution, which is a good idea. Meanwhile, this method only stores the learned branches and resolution-aware BNs, which requires less computation amount and parameter storage.\n\nWeaknesses:\n1. For the training set MS1Mv3, we need to first down-sample the high-resolution image to the low-resolution. Although this paper avoids up-sampling by branching strategy, down-sampling will still introduce interpolation errors and there is no solution for this question. \n\n2. For the test on real low-resolution face images, the model does not know the specific resolution to select the branch. If the face resolution needs to be judged according to the image size, many images of QMUL-SurvFace are not full face (after face alignment), and the image collected in real scene may have background and need face detection. But low-resolution face detection is difficult to achieve. At this time, there is no judgment strategy for face resolution, which make it difficult to select branches.\n\n3. This paper inputs images with different resolutions by adding branch headers in front of the backbone network, which is not novel enough. \n\n4. The paper lacks experiments compared to the SOTA methods, such as 1:N face identification on SCFace dataset, 1:1 face verification on QMUL-SurvFace dataset and 1:N face identification on QMUL-TinyFace dataset.\n\n5. In Table 2, the paper does not provide results for each of the six datasets. Tables 4 and 5 do not indicate which dataset the experiment was performed on. In figure 8, further allocation (floor/near/ceil) lacks a specific explanation.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a Branch-to-Trunk network for multi-resolution face recognition. In this paper, by setting multiple branches and inputting images with different resolutions into different branches, the interpolation error caused by up-sampling is reduced. It requires less computation amount and parameter storage. The paper conducts experiments on six face verification benchmarks and QMUL-SurvFace dataset.", "strength_and_weaknesses": "Strength:\nMost low-resolution face recognition works require up-sampling low-resolution images to a fixed size (112*112). This paper proposes a multi-branch network to reduce the interpolation error by piecewise processing the input image according to the resolution, which is a good idea. Meanwhile, this method only stores the learned branches and resolution-aware BNs, which requires less computation amount and parameter storage.\n\nWeaknesses:\n1. For the training set MS1Mv3, we need to first down-sample the high-resolution image to the low-resolution. Although this paper avoids up-sampling by branching strategy, down-sampling will still introduce interpolation errors and there is no solution for this question. \n\n2. For the test on real low-resolution face images, the model does not know the specific resolution to select the branch. If the face resolution needs to be judged according to the image size, many images of QMUL-SurvFace are not full face (after face alignment), and the image collected in real scene may have background and need face detection. But low-resolution face detection is difficult to achieve. At this time, there is no judgment strategy for face resolution, which make it difficult to select branches.\n\n3. This paper inputs images with different resolutions by adding branch headers in front of the backbone network, which is not novel enough. \n\n4. The paper lacks experiments compared to the SOTA methods, such as 1:N face identification on SCFace dataset, 1:1 face verification on QMUL-SurvFace dataset and 1:N face identification on QMUL-TinyFace dataset.\n\n5. In Table 2, the paper does not provide results for each of the six datasets. Tables 4 and 5 do not indicate which dataset the experiment was performed on. In figure 8, further allocation (floor/near/ceil) lacks a specific explanation.\n", "clarity,_quality,_novelty_and_reproducibility": "The clarity and novelty of this paper is not good enough. The reproducibility of this paper is good.", "summary_of_the_review": "See Strength And Weaknesses.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666684351058}, {"id": "Vy-LmduoS1G", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3517/Reviewer_JeWQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper presents a multi-resolution face recognition algorithm. It is based on a CNN encoder backbone (e.g. ResNet50), denoted trunk Net, and multiple lateral resolution-specific nets, denoted BTNets. Each BTNet receives as input an image with a specific resolution producing a representation of the same resolution. This representation is combined with the backbone at the layer with a resolution matching that of the BTNet. The experimentation shows how different components of the model contribute to the final solution. It compares favorably with the state of the art on a low resolution data set.", "review_text": "The paper addresses a relevant problem in face recognition. However, the presentation of the approach is difficult to follow and is not complete. The experimentation is weak. It compares with the state of the art in only one data set, failing to evaluate its multi-resolution performance against its competitors.\n", "strengths": "Strengths.\n\nThe paper addresses a relevant problem in the literature, namely, multi-resolution face recognition.\n\nWeaknesses. \n\nThe experimentation needs to be improved. It only compares with the state-of-the-art in a low-resolution data set. If it it claims to learn a multi-resolution representation, it should be compared with other competing algorithms and on many more data sets with different resolutions/characteristics.\n\nThe writing should also be improved. The paper is difficult to understand. Important information is missing, e.g. how is the representation produced by BTNet combined with the backbone. In section 4.1 it reads that the  experimentation has been performed with 6 data sets that go missing in the paper. Some tables at the end of the paper were not cited in the text.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presents a multi-resolution face recognition algorithm. It is based on a CNN encoder backbone (e.g. ResNet50), denoted trunk Net, and multiple lateral resolution-specific nets, denoted BTNets. Each BTNet receives as input an image with a specific resolution producing a representation of the same resolution. This representation is combined with the backbone at the layer with a resolution matching that of the BTNet. The experimentation shows how different components of the model contribute to the final solution. It compares favorably with the state of the art on a low resolution data set.", "strength_and_weaknesses": "Strengths.\n\nThe paper addresses a relevant problem in the literature, namely, multi-resolution face recognition.\n\nWeaknesses. \n\nThe experimentation needs to be improved. It only compares with the state-of-the-art in a low-resolution data set. If it it claims to learn a multi-resolution representation, it should be compared with other competing algorithms and on many more data sets with different resolutions/characteristics.\n\nThe writing should also be improved. The paper is difficult to understand. Important information is missing, e.g. how is the representation produced by BTNet combined with the backbone. In section 4.1 it reads that the  experimentation has been performed with 6 data sets that go missing in the paper. Some tables at the end of the paper were not cited in the text.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is not easy to read. It could not be reproduced with the information provided. However, some coded has been provided as supplementary material.", "summary_of_the_review": "The paper addresses a relevant problem in face recognition. However, the presentation of the approach is difficult to follow and is not complete. The experimentation is weak. It compares with the state of the art in only one data set, failing to evaluate its multi-resolution performance against its competitors.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666636472815}, {"id": "WBCCS7RX1f", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3517/Reviewer_FLmy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the author designed a new model,named Branch-to-Trunk network (BTNet), to match faces with different resolutions, and the experimental result demonstrates that the method is feasible. A possible potential contribution of this paper is the design of the network structure, mainly the introduction of multiple branches.", "review_text": "  The quality of this article is average and not recommended for acceptance.", "strengths": "Strength: The design of the model structure, but the pyramid structure is already very common in other work and cannot be considered as a very innovative point.\n\nWeakness: \nThere are several points in the article that confused me, as follows:\n1. In Table 4, the number of parameters of Pretraining + BCT is 43.59, while the number of parameters of Pretraining + BCT + Fix Trunk is 2.29. It seems unreasonable.\n2. As can be seen from Table 4, it is the model distillation operation that provides the largest performance improvement across multi-resolution and same resolution task. What puzzles me is: 1) whether distillation is capable of such a significant improvement and 2) if the distillation operation can improve to such an extent, then the innovation points claimed in this paper may not be reliable and more gains come from the distillation operation.\n3. The task of this paper is to obtain a better cross-resolution representation, and I think it may be more intuitive to visualize it using t-SNE.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the author designed a new model,named Branch-to-Trunk network (BTNet), to match faces with different resolutions, and the experimental result demonstrates that the method is feasible. A possible potential contribution of this paper is the design of the network structure, mainly the introduction of multiple branches.", "strength_and_weaknesses": "Strength: The design of the model structure, but the pyramid structure is already very common in other work and cannot be considered as a very innovative point.\n\nWeakness: \nThere are several points in the article that confused me, as follows:\n1. In Table 4, the number of parameters of Pretraining + BCT is 43.59, while the number of parameters of Pretraining + BCT + Fix Trunk is 2.29. It seems unreasonable.\n2. As can be seen from Table 4, it is the model distillation operation that provides the largest performance improvement across multi-resolution and same resolution task. What puzzles me is: 1) whether distillation is capable of such a significant improvement and 2) if the distillation operation can improve to such an extent, then the innovation points claimed in this paper may not be reliable and more gains come from the distillation operation.\n3. The task of this paper is to obtain a better cross-resolution representation, and I think it may be more intuitive to visualize it using t-SNE.\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is well presented, logical and clear, with standard citations.\n\nQuality: This article is of average quality and is not recommended for acceptance.\n\nNovelty: I feel that this article is not innovative enough.\n\nReproducibility: This article may be reproduced.\n", "summary_of_the_review": "  The quality of this article is average and not recommended for acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NO", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666621576172}], "openreview_url": "https://openreview.net/forum?id=1EVPT82ttr", "arxiv_id": "2310.09563", "paper_pdf": "papers/1EVPT82ttr.pdf", "paper_pdf_sha256": "c1a004a0d2c5dfd3583544a22c5b4fca0471a724cfba96248f63e4170a2702b6", "paper_pdf_bytes": 10915076, "paper_pdf_source": "openreview", "code_url": "https://github.com/hehulingxiao/BTNet", "code_repository": "hehulingxiao/BTNet", "code_commit": "df06e02fa4cff2fafec52078ea6cf858f36d5f76", "code_archive": "repos/1EVPT82ttr.zip", "code_archive_sha256": "a7110c532d9cb54ff2970962fcf0504aac4a2b6e69bd3544cf29b417af98ad3f", "code_archive_bytes": 223984, "code_file_count": 34, "code_extensions": {".py": 34}, "github_disk_usage_kb": 200, "github_languages": {"Python": 171003}, "github_archived": false, "github_pushed_at": "2023-10-03T07:27:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-unified-representations-for-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DvcMMKmDJ3q", "year": 2022, "status": "rejected", "title": "Generating Symbolic Reasoning Problems with Transformer GANs", "authors": ["Jens U. Kreber", "Christopher Hahn"], "authorids": ["~Jens_U._Kreber1", "~Christopher_Hahn1"], "authors_source": "OpenReview API", "abstract": "Constructing training data for symbolic reasoning domains is challenging: Existing instances are typically hand-crafted and too few to be trained on directly and synthetically generated instances are often hard to evaluate in terms of their meaningfulness. We study the capabilities of GANs and Wasserstein GANs equipped with Transformer encoders to generate sensible and challenging training data for symbolic reasoning domains. We conduct experiments on two problem domains where Transformers have been successfully applied recently: symbolic mathematics and temporal specifications in verification. Even without autoregression, our GAN models produce syntactically correct instances and we show that these can be used as meaningful substitutes for real training data when training a classifier. Using a GAN setting also allows us to alter the target distribution: We show that by adding a classifier uncertainty part to the generator objective, we obtain a dataset that is even harder to solve for a classifier than our original dataset.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "3tuLK8vcPLw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3698/Reviewer_vhbE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to generate training data for symbolic reasoning by training GANs based on Transformers. Specifically, the paper explores two methods --- standard GAN and Wasserstein GAN --- which has a Transformer as an encoder and a decoder, respectively. Training these models on symbolic reasoning data that is randomly generated --- LTL and Symbolic mathematics --- is found to generate high quality formulas.", "review_text": "** Update after author responses ** After the author clarified with the format of the data and results, my problems have resolved in part and thus I updated the score. My concern in the motivation of the paper still remains. The paper claims that synthetic dataset generated from rules is limited in the coverage so that it is valuable to construct synthetic dataset from neural models --- however, it is still not clear to me if experiments demonstrate that the resulting data from neural models are shown to have better coverage and higher quality than data from rules.\n\nThe model introduced in the paper is well-executed, and based on the generated data attached in the submission, the data quality appears to be good enough. There are still a few concerns I have in the motivation of the problem and experiments.\n\nFirst, the problem setup is not convincing to me and is not well-motivated in the paper. If data for the symbolic reasoning --- which is not precisely defined in the paper --- can easily be generated at scale using rules, as done for the base data in the paper, what is the reason for training a neural model to generate more data? How is it inherently different from generating more using rules, as data, either generated automatically from rules or generated by neutral models, is equally artificial?\n\n~~Second, although the paper claims that the generated data is high quality, it seems to have quite bad quality to me. I checked Supplementary materials and most formulas contain many repetitions like “&&&&&&&GXaG>!” and are very hard to interpret. It is very difficult to find good examples as listed in Section 4.2.1, so there is a high likelihood that these good examples are cherry-picked.~~\n\nThird, the experiments in Section 4 do not demonstrate that the generated data has high quality and effectively replaces the base data. For example, Table 2 shows that training on generated data achieves performance that is comparable to the model trained on the original data, but not better. This is related to my first point in the motivation of the problem - if the original data can be obtained automatically at scale and generated data from the proposed model is not significantly better than the original data, is there a justification for not using the original data?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to generate training data for symbolic reasoning by training GANs based on Transformers. Specifically, the paper explores two methods --- standard GAN and Wasserstein GAN --- which has a Transformer as an encoder and a decoder, respectively. Training these models on symbolic reasoning data that is randomly generated --- LTL and Symbolic mathematics --- is found to generate high quality formulas.", "main_review": "** Update after author responses ** After the author clarified with the format of the data and results, my problems have resolved in part and thus I updated the score. My concern in the motivation of the paper still remains. The paper claims that synthetic dataset generated from rules is limited in the coverage so that it is valuable to construct synthetic dataset from neural models --- however, it is still not clear to me if experiments demonstrate that the resulting data from neural models are shown to have better coverage and higher quality than data from rules.\n\nThe model introduced in the paper is well-executed, and based on the generated data attached in the submission, the data quality appears to be good enough. There are still a few concerns I have in the motivation of the problem and experiments.\n\nFirst, the problem setup is not convincing to me and is not well-motivated in the paper. If data for the symbolic reasoning --- which is not precisely defined in the paper --- can easily be generated at scale using rules, as done for the base data in the paper, what is the reason for training a neural model to generate more data? How is it inherently different from generating more using rules, as data, either generated automatically from rules or generated by neutral models, is equally artificial?\n\n~~Second, although the paper claims that the generated data is high quality, it seems to have quite bad quality to me. I checked Supplementary materials and most formulas contain many repetitions like “&&&&&&&GXaG>!” and are very hard to interpret. It is very difficult to find good examples as listed in Section 4.2.1, so there is a high likelihood that these good examples are cherry-picked.~~\n\nThird, the experiments in Section 4 do not demonstrate that the generated data has high quality and effectively replaces the base data. For example, Table 2 shows that training on generated data achieves performance that is comparable to the model trained on the original data, but not better. This is related to my first point in the motivation of the problem - if the original data can be obtained automatically at scale and generated data from the proposed model is not significantly better than the original data, is there a justification for not using the original data?\n\n", "summary_of_the_review": "The model introduced in the paper is well-executed, and based on the generated data attached in the submission, the data quality appears to be good enough. However, in my opinion, there are more fundamental issues in the motivation of the problem and whether experiments successfully justified the usefulness of the generated data from the proposed model.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635912502857}, {"id": "DRwTWMiDmu4", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3698/Reviewer_3zMF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "Generating symbolic reasoning training data using transformer GANs. Use classifier uncertainty in the generator objective to generate samples that are harder for the classifier to solve than the original data. The problem is very interesting, but the empirical results could be improved.", "review_text": "The problem is interesting and the paper is well motivated.\n\nAbstract: What do they authors mean by “synthetically generated instances are often hard to evaluate in terms of their meaningfulness”?\n\nWhat do the authors mean by “training on randomly generated data carries the risk of training on meaningless data or the risk of introducing unwanted biases”? Provided that the conclusion follow logically, can the authors elaborate on what they mean here? \n\n“We show that training directly in the one-hot encoding space is possible when adding Gaussian noise to each position”. Why use Gaussian noise? The authors should be more clear about why they add noise to the real samples? Presumably this is to make distinguishing the real samples from the fake ones non-trivial? This is only explained later on and is a bit confusing.\n\n“on which a classifier can successfully be trained on”. Do you test on an established dataset?\n\n“The generator’s input is a real scalar random value with uniform distribution [0, 1] for each position.” It is not clear here what is mean by “each position”. I assume this is each symbol in the input? But this is not clearly explained.\n\n“The position-wise padding mask is copied from the real data during training, so the lengths of real and generated formulas at the same position in a batch are always identical.” The first time you refer to **the** padding mask it is not clear what this is or where it comes from?\n\n“Still, both generators are able to produce a large fraction of fully correct temporal specifications, which we find surprising” Could there be overfitting? Are the results the same across many runs? Is 0.3 a large fraction? How does this compare to generating examples randomly? This would make a good baseline (even if your models perform worse).\n\nFigure 4: Interesting results.\n\nThe satisfiability classifier results are interesting. What would be better tho, is to show that training on the generated data improved results on an established dataset. \n\nIt is interesting that you can learn on data generated using the LTLbase 10k dataset and perform better on the validation set than training directly on the LRLbase (and perform similarly well when training on the whole LTLbase dataset). However, it’s clear that the model has overfit to the LTLbase 10k. How did you decide to stop training? Does this happen for all runs? It would be helpful to see the training and test curves while the model is training. Additionally, it would be good to see a graph with multiple runs.\n\nFor all results in the paper, it would be best to perform multiple runs and report the standard deviations. \n\nIn the section titled “GAN with included classifier”: What are you classifying? Are you predicting satisfiability? This is not clear. How do you know the satisfiability of the generated samples? Is there a 50/50 split of satisfiable and unsatisfiable examples?\n\nAre there the same number of training examples in the LTLbase, Uncert-e and mixte-e datasets?\n\nTable 3: Are you training and testing on different splits of the dataset? It would help to add the std? The results somewhat suggest that training to increase uncertainty does produce slightly more challenging problems, but it’s hard to tell because it’s not clear how much results vary between runs and it’s not clear if the differences are statistically significant. \n\nTable 3: It’s very important to know how train and test samples were split for Uncert-e. Uncert-e being more difficult to classify could also be explained by lack of variation and the samples you tested happening to be in a different mode to those in the training set. What is the average length of problems in each data set?\n\nIs there a qualitative difference between samples generated with and without the uncertainty loss? What makes the problems harder?\n\nDoes training on Uncert-e improve performance on LTLbase (more so than training on generated?) Is there a standard dataset that you can show improvement on?\n\nWhy are there not more quantitative results on the integration examples? It’s clear that a classifier can already perform very well on the LTL tasks, perhaps it would be easier to see improvements on the function integration task? It does not appear that there are any results for this in the main text? \n\nGeneral: The paper does not clearly separate LTL results and symbolic math results. This makes some results harder to parse.\n\nGeneral: This model does not generate a supervised training dataset since you still need to use existing algorithms/ programs to compute the labels/targets. This could be a problem for datasets where the solution is intractable and would also suggest that you already have a model capable of solving the problem for which you are generating the data. What is the long term motivation of this work if you either (a) cannot generate labels/targets or (b) can already use existing algorithm to solve these problems.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Generating symbolic reasoning training data using transformer GANs. Use classifier uncertainty in the generator objective to generate samples that are harder for the classifier to solve than the original data. The problem is very interesting, but the empirical results could be improved.", "main_review": "The problem is interesting and the paper is well motivated.\n\nAbstract: What do they authors mean by “synthetically generated instances are often hard to evaluate in terms of their meaningfulness”?\n\nWhat do the authors mean by “training on randomly generated data carries the risk of training on meaningless data or the risk of introducing unwanted biases”? Provided that the conclusion follow logically, can the authors elaborate on what they mean here? \n\n“We show that training directly in the one-hot encoding space is possible when adding Gaussian noise to each position”. Why use Gaussian noise? The authors should be more clear about why they add noise to the real samples? Presumably this is to make distinguishing the real samples from the fake ones non-trivial? This is only explained later on and is a bit confusing.\n\n“on which a classifier can successfully be trained on”. Do you test on an established dataset?\n\n“The generator’s input is a real scalar random value with uniform distribution [0, 1] for each position.” It is not clear here what is mean by “each position”. I assume this is each symbol in the input? But this is not clearly explained.\n\n“The position-wise padding mask is copied from the real data during training, so the lengths of real and generated formulas at the same position in a batch are always identical.” The first time you refer to **the** padding mask it is not clear what this is or where it comes from?\n\n“Still, both generators are able to produce a large fraction of fully correct temporal specifications, which we find surprising” Could there be overfitting? Are the results the same across many runs? Is 0.3 a large fraction? How does this compare to generating examples randomly? This would make a good baseline (even if your models perform worse).\n\nFigure 4: Interesting results.\n\nThe satisfiability classifier results are interesting. What would be better tho, is to show that training on the generated data improved results on an established dataset. \n\nIt is interesting that you can learn on data generated using the LTLbase 10k dataset and perform better on the validation set than training directly on the LRLbase (and perform similarly well when training on the whole LTLbase dataset). However, it’s clear that the model has overfit to the LTLbase 10k. How did you decide to stop training? Does this happen for all runs? It would be helpful to see the training and test curves while the model is training. Additionally, it would be good to see a graph with multiple runs.\n\nFor all results in the paper, it would be best to perform multiple runs and report the standard deviations. \n\nIn the section titled “GAN with included classifier”: What are you classifying? Are you predicting satisfiability? This is not clear. How do you know the satisfiability of the generated samples? Is there a 50/50 split of satisfiable and unsatisfiable examples?\n\nAre there the same number of training examples in the LTLbase, Uncert-e and mixte-e datasets?\n\nTable 3: Are you training and testing on different splits of the dataset? It would help to add the std? The results somewhat suggest that training to increase uncertainty does produce slightly more challenging problems, but it’s hard to tell because it’s not clear how much results vary between runs and it’s not clear if the differences are statistically significant. \n\nTable 3: It’s very important to know how train and test samples were split for Uncert-e. Uncert-e being more difficult to classify could also be explained by lack of variation and the samples you tested happening to be in a different mode to those in the training set. What is the average length of problems in each data set?\n\nIs there a qualitative difference between samples generated with and without the uncertainty loss? What makes the problems harder?\n\nDoes training on Uncert-e improve performance on LTLbase (more so than training on generated?) Is there a standard dataset that you can show improvement on?\n\nWhy are there not more quantitative results on the integration examples? It’s clear that a classifier can already perform very well on the LTL tasks, perhaps it would be easier to see improvements on the function integration task? It does not appear that there are any results for this in the main text? \n\nGeneral: The paper does not clearly separate LTL results and symbolic math results. This makes some results harder to parse.\n\nGeneral: This model does not generate a supervised training dataset since you still need to use existing algorithms/ programs to compute the labels/targets. This could be a problem for datasets where the solution is intractable and would also suggest that you already have a model capable of solving the problem for which you are generating the data. What is the long term motivation of this work if you either (a) cannot generate labels/targets or (b) can already use existing algorithm to solve these problems.\n", "summary_of_the_review": "The problem is really interesting. \n\n\nCorrectness: \nThere are problems with the experimental results. If the authors can add the suggested results and show that their results are statistically significant, I would be very happy to increase my score. \n\nNovelty: The approach also lacks novelty, only proposing an additional loss which is not clearly described. However, their application is very interesting.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "n/a", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635788167024}, {"id": "F90-vecqQc", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3698/Reviewer_qNn9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper makes the use of GANs equipped with Transformers to generate new\ninstances of the problems in the symbolic reasoning domains, and demonstrated\nthe usefulness of the idea on two domains: satisfiability prediction of LTL\nformulas and mathematical reasoning on integration and ordinary differential\nequations.  The experiments show that the trained GAN models can produce\nparsable instances on both of the domains and that training on the generated\ndata can improve the accuracy performance of a classification model for LTL\nsatisfiability prediction.", "review_text": "### Strengths\n\n* Simple idea: using GANs for data augmentation.\n\n* Empirically proving that GANs can generate parsable instances and that GANs\n  can generate interesting problems that are hard to solve and contribute to\n  improve the performance of classifiers.\n\n### Weaknesses\n\nI have several concerns and questions as follows.\n\n* I'm not entirely sure of how the generated data are labeled.  Is it done by\n  the GAN discriminator or by another algorithm as labeling LTL formulas in the\n  real dataset with the tool aalta (Appendix B.3)?\n\n  If the latter is employed, it indicates the existence of an algorithm to\n  address the task of interest, so I'm wondering why machine learning is applied\n  to it.  Perhaps ML models might be expected to answer the problems faster\n  than the algorithm, but labeling with the algorithm would be successful only\n  on problems that the algorithm can answer in a given time; thus, it seems hard\n  to label problems that the algorithm cannot solve in a reasonable time,\n  although the ML models should approach such problems.\n\n  For the former, the generated data could be mislabeled.  Then, it is not clear\n  for me why training on such data can produce a model with the performance very\n  close to the model trained on data with correct labels (Table 2).\n\n* I don't think this is the first work to use GANs for data augmentation.  For\n  example, [1] and [2] used GANs for augmenting image data.  Unlike\n  these previous works, the paper addresses sequential data that are textual\n  representations of mathematical expressions, but lacking the discussion and\n  (qualitative) comparison with them makes a challenge and novelty of the paper\n  unclear.\n\n  [1] Antreas Antoniou, Amos J. Storkey, Harrison Edwards.\n  Data Augmentation Generative Adversarial Networks.\n\n  [2] Christopher Bowles, Liang Chen, Ricardo Guerrero, Paul Bentley, Roger N. Gunn, Alexander Hammers, David Alexander Dickie, Maria del C. Valdés Hernández, Joanna M. Wardlaw, Daniel Rueckert:\n  GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks.\n\n\n* The abstract and introduction seem oversold somewhat in that they don't\n  mention that the experiment conducted for symbolic mathematics only\n  confirms the GAN model can produce syntactically correct problems; thus, the\n  full usefulness of the GAN-based data augmentation is still open in the\n  symbolic mathematics domain.  I think it would be nicer to conduct more\n  experiments on symbolic mathematics, which would demonstrate the proposed\n  approach is useful in broader domains.\n\n* The experiments on satisfiability of LTL formulas look interesting, but I'm\n  concerned that it is adequate to be tacked with machine learning.  As written\n  in Appendix B.3, there exists a tool for checking satisfiability of LTL\n  formulas, and it is used to label the (real) dataset. Then, why do we need the\n  trained model for checking satisfiability?  A critical problem with the\n  trained model is that its prediction might be wrong.  How can we use a\n  satisfiability checker that may produce wrong answers?\n\n  For a similar reason, I'm not sure why machine learning is useful for the\n  domain where \"data can ... often be labeled automatically\" (page 8).\n\n* The paper uses GAN and WGAN, but it does not investigate why they make a\n  difference in the experiments (e.g., Table 1).\n\n\n### Minor comments / questions\n\nP2 \"\\neg \\box (access_p0 \\wedge access_p1)\"  I think \\box (globally) should be\nreplaced with \\diamond (eventually) for correction.\n\nP2 \"Mathematical expressions\"  What are these?  Are they different from random functions?\n\nP3 \"continuous domains were\"  where\n\nP4 \"we use the alternative generator loss\"  Why?\n\nP6 \"the origin training data\"  original\n\nP6, Table 1:\n  For GAN, does the use of \\sigmoid_real larger than 0.2 make the fraction of fc larger?\n\nP7 \"we combine both critic and classifier into one Transformer encoder\"  Why are not they separated?\n\nP8: The result in Table 3 seems peculiar to me.  Why does the model trained on\nMixed-e outperform the model trained on LTLbase even when tested on LTLbase?\n\n### Post-Rebuttal\n\nI would like to thank the authors for the additional comments to answer my questions.\nHowever, I still have two major concerns that make me hesitate to accept the paper.\nThe first s that I don't still find the task of generating symbolic expressions interesting.\nThe second is about an application of the approach.  The response from the authors says that neural models may compute solutions faster than classical tools.  I don't disagree with this claim, but the paper doesn't show that the proposed approach is indeed helpful for that task.  I would like to see more discussions and evidence for the story to hold true (e.g., how the augmented data are labeled, whether they can improve the performance of neural models that compute solutions, etc.)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper makes the use of GANs equipped with Transformers to generate new\ninstances of the problems in the symbolic reasoning domains, and demonstrated\nthe usefulness of the idea on two domains: satisfiability prediction of LTL\nformulas and mathematical reasoning on integration and ordinary differential\nequations.  The experiments show that the trained GAN models can produce\nparsable instances on both of the domains and that training on the generated\ndata can improve the accuracy performance of a classification model for LTL\nsatisfiability prediction.", "main_review": "### Strengths\n\n* Simple idea: using GANs for data augmentation.\n\n* Empirically proving that GANs can generate parsable instances and that GANs\n  can generate interesting problems that are hard to solve and contribute to\n  improve the performance of classifiers.\n\n### Weaknesses\n\nI have several concerns and questions as follows.\n\n* I'm not entirely sure of how the generated data are labeled.  Is it done by\n  the GAN discriminator or by another algorithm as labeling LTL formulas in the\n  real dataset with the tool aalta (Appendix B.3)?\n\n  If the latter is employed, it indicates the existence of an algorithm to\n  address the task of interest, so I'm wondering why machine learning is applied\n  to it.  Perhaps ML models might be expected to answer the problems faster\n  than the algorithm, but labeling with the algorithm would be successful only\n  on problems that the algorithm can answer in a given time; thus, it seems hard\n  to label problems that the algorithm cannot solve in a reasonable time,\n  although the ML models should approach such problems.\n\n  For the former, the generated data could be mislabeled.  Then, it is not clear\n  for me why training on such data can produce a model with the performance very\n  close to the model trained on data with correct labels (Table 2).\n\n* I don't think this is the first work to use GANs for data augmentation.  For\n  example, [1] and [2] used GANs for augmenting image data.  Unlike\n  these previous works, the paper addresses sequential data that are textual\n  representations of mathematical expressions, but lacking the discussion and\n  (qualitative) comparison with them makes a challenge and novelty of the paper\n  unclear.\n\n  [1] Antreas Antoniou, Amos J. Storkey, Harrison Edwards.\n  Data Augmentation Generative Adversarial Networks.\n\n  [2] Christopher Bowles, Liang Chen, Ricardo Guerrero, Paul Bentley, Roger N. Gunn, Alexander Hammers, David Alexander Dickie, Maria del C. Valdés Hernández, Joanna M. Wardlaw, Daniel Rueckert:\n  GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks.\n\n\n* The abstract and introduction seem oversold somewhat in that they don't\n  mention that the experiment conducted for symbolic mathematics only\n  confirms the GAN model can produce syntactically correct problems; thus, the\n  full usefulness of the GAN-based data augmentation is still open in the\n  symbolic mathematics domain.  I think it would be nicer to conduct more\n  experiments on symbolic mathematics, which would demonstrate the proposed\n  approach is useful in broader domains.\n\n* The experiments on satisfiability of LTL formulas look interesting, but I'm\n  concerned that it is adequate to be tacked with machine learning.  As written\n  in Appendix B.3, there exists a tool for checking satisfiability of LTL\n  formulas, and it is used to label the (real) dataset. Then, why do we need the\n  trained model for checking satisfiability?  A critical problem with the\n  trained model is that its prediction might be wrong.  How can we use a\n  satisfiability checker that may produce wrong answers?\n\n  For a similar reason, I'm not sure why machine learning is useful for the\n  domain where \"data can ... often be labeled automatically\" (page 8).\n\n* The paper uses GAN and WGAN, but it does not investigate why they make a\n  difference in the experiments (e.g., Table 1).\n\n\n### Minor comments / questions\n\nP2 \"\\neg \\box (access_p0 \\wedge access_p1)\"  I think \\box (globally) should be\nreplaced with \\diamond (eventually) for correction.\n\nP2 \"Mathematical expressions\"  What are these?  Are they different from random functions?\n\nP3 \"continuous domains were\"  where\n\nP4 \"we use the alternative generator loss\"  Why?\n\nP6 \"the origin training data\"  original\n\nP6, Table 1:\n  For GAN, does the use of \\sigmoid_real larger than 0.2 make the fraction of fc larger?\n\nP7 \"we combine both critic and classifier into one Transformer encoder\"  Why are not they separated?\n\nP8: The result in Table 3 seems peculiar to me.  Why does the model trained on\nMixed-e outperform the model trained on LTLbase even when tested on LTLbase?\n\n### Post-Rebuttal\n\nI would like to thank the authors for the additional comments to answer my questions.\nHowever, I still have two major concerns that make me hesitate to accept the paper.\nThe first s that I don't still find the task of generating symbolic expressions interesting.\nThe second is about an application of the approach.  The response from the authors says that neural models may compute solutions faster than classical tools.  I don't disagree with this claim, but the paper doesn't show that the proposed approach is indeed helpful for that task.  I would like to see more discussions and evidence for the story to hold true (e.g., how the augmented data are labeled, whether they can improve the performance of neural models that compute solutions, etc.)", "summary_of_the_review": "The paper addresses a critical problem in symbolic reasoning domains, and the\nexperimental results are promising. However, I think it's not ready for\npublication because of lacking a discussion for practical settings to use the\nproposed GAN-based data augmentation, an evidence to show its generality, and a\ncomparison with the previous work.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635685670786}, {"id": "9EL3oHJxTAa", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3698/Reviewer_JuGn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors apply (W)GANs with transformer encoders for data-augmentation in two symbolic domains: LTL and function integration. There are many interesting findings, three of which stand out. First, in both domains, the network learns to generate syntactically-correct examples roughly 30% of the time. Second, for LTL, a GAN trained on a dataset of size 10k can produce a much larger dataset, such that training on the new dataset is almost as good as training on the original distribution (when evaluating on the original distribution). Third, by also rewarding the generator for confusing a classifier, they can generate new problems that are harder to classify than those in the original distribution.", "review_text": "This is a strong paper overall: innovative, well-written, and potentially important.\n\n- I found it awkward that the prose stressed that two domains would be considered, in some cases even listing symbolic mathematics (i.e. function integration) first, even though the integration domain received much shorter shrift and most of the experiments seem to be LTL only. I think the paper would be stronger with two domains throughout rather than one. \n\n- How many examples were in LTLBase? It might be nice to see an estimate of the effective size of the \"Generated\" dataset with respect to the original distribution.\n\n- I see that in S4.2 you are starting with only 10k samples from LTLBase, but what are you training on exactly for S4.1? If a larger dataset than in S4.2, did you do a similar check for duplicates between the generated examples and the training dataset? How many (if any) of the 30% that are syntactically valid were seen during training?\n\n- It is not clear to me whether the last paragraph of S4.1 is describing a different training regime than the paragraph preceeding it.\n\n- The first sentence seems more suitable for a blog post than a conference paper; in particular, the claim that deep learning is \"on the verge\" of something sounds unscientific. Also, the word \"transitioning\" may not be appropriate, since presumably deep learning will continue to be applied to e.g. image recognition.\n\n- I particularly liked the approach in S4.3 in which an additional objective is introduced that shifts the generated distribution away from the original one. In general, the true objective of data augmentation may be broader than simply modeling the specific distribution of training examples one has at hand. Have the authors considered other \"knobs\" to add that would allow generating more diverse examples, that may provide useful augmentation either for the original distribution or for out-of-distribution evaluations?\n\n- Minor comment: I did not find that it added much to the paper to consider both GANs and WGANs. I would suggest only discussing WGANs.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors apply (W)GANs with transformer encoders for data-augmentation in two symbolic domains: LTL and function integration. There are many interesting findings, three of which stand out. First, in both domains, the network learns to generate syntactically-correct examples roughly 30% of the time. Second, for LTL, a GAN trained on a dataset of size 10k can produce a much larger dataset, such that training on the new dataset is almost as good as training on the original distribution (when evaluating on the original distribution). Third, by also rewarding the generator for confusing a classifier, they can generate new problems that are harder to classify than those in the original distribution.", "main_review": "This is a strong paper overall: innovative, well-written, and potentially important.\n\n- I found it awkward that the prose stressed that two domains would be considered, in some cases even listing symbolic mathematics (i.e. function integration) first, even though the integration domain received much shorter shrift and most of the experiments seem to be LTL only. I think the paper would be stronger with two domains throughout rather than one. \n\n- How many examples were in LTLBase? It might be nice to see an estimate of the effective size of the \"Generated\" dataset with respect to the original distribution.\n\n- I see that in S4.2 you are starting with only 10k samples from LTLBase, but what are you training on exactly for S4.1? If a larger dataset than in S4.2, did you do a similar check for duplicates between the generated examples and the training dataset? How many (if any) of the 30% that are syntactically valid were seen during training?\n\n- It is not clear to me whether the last paragraph of S4.1 is describing a different training regime than the paragraph preceeding it.\n\n- The first sentence seems more suitable for a blog post than a conference paper; in particular, the claim that deep learning is \"on the verge\" of something sounds unscientific. Also, the word \"transitioning\" may not be appropriate, since presumably deep learning will continue to be applied to e.g. image recognition.\n\n- I particularly liked the approach in S4.3 in which an additional objective is introduced that shifts the generated distribution away from the original one. In general, the true objective of data augmentation may be broader than simply modeling the specific distribution of training examples one has at hand. Have the authors considered other \"knobs\" to add that would allow generating more diverse examples, that may provide useful augmentation either for the original distribution or for out-of-distribution evaluations?\n\n- Minor comment: I did not find that it added much to the paper to consider both GANs and WGANs. I would suggest only discussing WGANs.\n", "summary_of_the_review": "This is a strong paper overall: innovative, well-written, and potentially important. I think it will be of interest to anyone applying machine learning in data-sparse symbolic domains.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635168816742}], "openreview_url": "https://openreview.net/forum?id=DvcMMKmDJ3q", "arxiv_id": "2110.10054", "paper_pdf": "papers/DvcMMKmDJ3q.pdf", "paper_pdf_sha256": "6ee02cd175a5f6a096895f3beec1316e3857e4e87d2e0bb4fdf90defb5412e5f", "paper_pdf_bytes": 452208, "paper_pdf_source": "openreview", "code_url": "https://github.com/reactive-systems/TGAN-SR", "code_repository": "reactive-systems/TGAN-SR", "code_commit": "43a5a93d0864ca77a693328cc07c8088be1f8f50", "code_archive": "repos/DvcMMKmDJ3q.zip", "code_archive_sha256": "005ca68f7a2e718ae90e57edf8830508737776f59854a7496c6af2db954c6ca3", "code_archive_bytes": 437605, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 424, "github_languages": {"Python": 160421}, "github_archived": false, "github_pushed_at": "2021-10-20T07:53:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generating-symbolic-reasoning-problems-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "i7aMbliTkHs", "year": 2021, "status": "rejected", "title": "TAM: Temporal Adaptive Module for Video Recognition", "authors": ["Zhaoyang Liu", "Limin Wang", "Wayne Wu", "Chen Qian", "Tong Lu"], "authorids": ["~Zhaoyang_Liu1", "~Limin_Wang1", "~Wayne_Wu1", "~Chen_Qian1", "~Tong_Lu1"], "authors_source": "OpenReview API", "abstract": "Temporal modeling is crucial for capturing spatiotemporal structure in videos for action recognition. Video data is with extremely complex dynamics along its temporal dimension due to various factors such as camera motion, speed variation, and different activities. To effectively capture this diverse motion pattern, this paper presents a new temporal adaptive module ({\\bf TAM}) to generate video-specific kernels based on its own feature maps. TAM proposes a unique two-level adaptive modeling scheme by decoupling dynamic kernels into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short term information, while the aggregation weight is generated from a global view with a focus on long-term structure. TAM is a principled module and could be integrated into 2D CNNs to yield a powerful video architecture (TANet) with a very small extra computational cost. The extensive experiments on Kinetics-400 and Something-Something datasets, demonstrate that the TAM outperforms other temporal modeling methods consistently owing to its temporal adaptive modeling strategy.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "MVwn5BTvXVM", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper406/AnonReviewer2"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "*********\nSummary Of The Manuscript:\n*********\nThe manuscript addresses the problem of Video Recognition one of the applications of Computer Vision. Due to various complex temporal dynamics of video data (Camera Motion, Speed, etc.), to capture the vast information, the author presents a novel Temporal Adaptive Module (TAM) for generating kernels based on the temporal feature maps. In addition, these feature maps are a combination of local and global features and as an exemplar, the author presents an architecture - TANet by incorporating their temporal operator. Together with a variety of experiments on standard benchmark for Video Recognition: Kinetics - 400 and Something-Something, the author showcases that for the task of Video Recognition compared to existing temporal operators, TAM's performance is fairly consistent and better and archives State-of-the-art with similar complexity in their exemplar architecture - TANet. \n\n*********\nStrength Of The Manuscript:\n*********\n++ Novelty\n\n- The task formulation is concise, convincing, and novel. A seemingly reasonable approach has been proposed in this manuscript for the task of Video Recognition. Compared to the existing baseline and recent approaches, the proposed architecture - TANet achieves SOTA results. \n- To the best of my knowledge, the incorporation of two branches - Local and Global branch makes the whole operator efficient and flexible for adaptation in the frameworks by stacking them to capture more complex information. Thus the concept of the TAM is convincing to capture complex temporal information.\n- In addition, the exemplar showcased by the authors - TANet has been created by incorporating TAM in the existing 2-Dimensional CNNs to capture vast information which proves that the proposed module/operator is flexible enough and can be adapted to different frameworks/architectures for better performance. \n\n++ Clarity\n\n- The manuscript is written in an excellent way to provide a brief insight into TAM. Especially Subsection 3.2 and 3.2 provide a good in-depth description of how the local and global branch works effectively in a joint manner to capture the short and long-term complex temporal information. \n- The manuscript also clearly describes the improvements and adequately contextualizes the contributions in such a way that it makes a good starting point for a novice reader. \n\n++ Evaluation\n\n- The experiments are sufficient and convincing. This new operator and exemplar TANet shows improved performance in nearly all cases on the datasets \n- The experimental evaluations demonstrate the effectiveness of the proposed architecture and showcase its practical value.\n- Also, an ablation analysis demonstrates to gain an understanding of where the performance benefits have been obtained such as receptive fields and parameter choices. \n\n*********\nWeakness Of The Manuscript:\n*********\nOverall, currently at this stage, this is a very good and strong manuscript in my entire batch. I like the simplicity and wide applicability of the proposed operator, especially the incorporation of local and global branches and adaptive aggregation. Thus I do not have any major weakness issues after reading the manuscript several times. Detailed literature review, a complete overview of each component, and detailed experiments and ablation studies helps to give a good insight into the manuscript. I found this paper pretty solid and have not able to found concerns relating to the proposed work. \n\n*********\nJustification Of The Review:\n*********\nOverall, happy with the current version of the manuscript. As mentioned earlier, I like the simplicity and wide applicability of the proposed module, and the architecture and setup details are provided in such a manner that it is very easy to convert into code in some timeframe. Detailed literature review, a complete overview of each component, and detailed experiments and ablation studies help to understand the author's work. Finally, I think the paper is pretty solid and thus I prefer to give a rating of 8 currently. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of the manuscript TAM: Temporal Adaptive Module for Video Recognition", "review": "*********\nSummary Of The Manuscript:\n*********\nThe manuscript addresses the problem of Video Recognition one of the applications of Computer Vision. Due to various complex temporal dynamics of video data (Camera Motion, Speed, etc.), to capture the vast information, the author presents a novel Temporal Adaptive Module (TAM) for generating kernels based on the temporal feature maps. In addition, these feature maps are a combination of local and global features and as an exemplar, the author presents an architecture - TANet by incorporating their temporal operator. Together with a variety of experiments on standard benchmark for Video Recognition: Kinetics - 400 and Something-Something, the author showcases that for the task of Video Recognition compared to existing temporal operators, TAM's performance is fairly consistent and better and archives State-of-the-art with similar complexity in their exemplar architecture - TANet. \n\n*********\nStrength Of The Manuscript:\n*********\n++ Novelty\n\n- The task formulation is concise, convincing, and novel. A seemingly reasonable approach has been proposed in this manuscript for the task of Video Recognition. Compared to the existing baseline and recent approaches, the proposed architecture - TANet achieves SOTA results. \n- To the best of my knowledge, the incorporation of two branches - Local and Global branch makes the whole operator efficient and flexible for adaptation in the frameworks by stacking them to capture more complex information. Thus the concept of the TAM is convincing to capture complex temporal information.\n- In addition, the exemplar showcased by the authors - TANet has been created by incorporating TAM in the existing 2-Dimensional CNNs to capture vast information which proves that the proposed module/operator is flexible enough and can be adapted to different frameworks/architectures for better performance. \n\n++ Clarity\n\n- The manuscript is written in an excellent way to provide a brief insight into TAM. Especially Subsection 3.2 and 3.2 provide a good in-depth description of how the local and global branch works effectively in a joint manner to capture the short and long-term complex temporal information. \n- The manuscript also clearly describes the improvements and adequately contextualizes the contributions in such a way that it makes a good starting point for a novice reader. \n\n++ Evaluation\n\n- The experiments are sufficient and convincing. This new operator and exemplar TANet shows improved performance in nearly all cases on the datasets \n- The experimental evaluations demonstrate the effectiveness of the proposed architecture and showcase its practical value.\n- Also, an ablation analysis demonstrates to gain an understanding of where the performance benefits have been obtained such as receptive fields and parameter choices. \n\n*********\nWeakness Of The Manuscript:\n*********\nOverall, currently at this stage, this is a very good and strong manuscript in my entire batch. I like the simplicity and wide applicability of the proposed operator, especially the incorporation of local and global branches and adaptive aggregation. Thus I do not have any major weakness issues after reading the manuscript several times. Detailed literature review, a complete overview of each component, and detailed experiments and ablation studies helps to give a good insight into the manuscript. I found this paper pretty solid and have not able to found concerns relating to the proposed work. \n\n*********\nJustification Of The Review:\n*********\nOverall, happy with the current version of the manuscript. As mentioned earlier, I like the simplicity and wide applicability of the proposed module, and the architecture and setup details are provided in such a manner that it is very easy to convert into code in some timeframe. Detailed literature review, a complete overview of each component, and detailed experiments and ablation studies help to understand the author's work. Finally, I think the paper is pretty solid and thus I prefer to give a rating of 8 currently. ", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603990383978}, {"id": "KaEjPF1Fg79", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper406/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a temporal adaptive module for video recognition. Specifically, it decouples dynamic kernel into a location sensitive importance map and a location invariant aggregation weight, which can be plugged into existing 2D CNNs to yield a powerful video architecture with small extra computational cost. The experiments conducted on several datasets demonstrate the effectiveness of the proposed method.\n\nPaper Strength\n(1)\tThe proposed method develops a simple but effective module for video recognition, achieving good performance on various kinds of datasets.\n(2)\tThe proposed two-level adaptive modeling scheme is effective to describe motion patterns. Specifically, it decomposes the video specific temporal kernel into a location sensitive importance map and a location invariant aggregation kernel.\n\nPaper Weakness\n(1)\tFigure 1 is confusing. Why the dimension of the green filter is 3*1*1?\n(2)\tIn Figure 2, TAM is used before the 1*1*3 convolutional layer in the ResNet block. What about the influence of TAM at different locations? Is there any insight for this design?\n(3)\tMoreover, what if we exchange the order of the local branch and the global branch?\n(4)\tIn Section 3.4, the authors claim that the proposed temporal adaptive module can be plugged into existing 2D CNNs with a strong ability to model different temporal structures in video clips. However, only the ResNet-50 backbone is verified in the experiment. More experiments with other backbones such as VGG and Inception should be added.\n(5) In Table 3, the proposed method is not compared to the state-of-the-art X3D method [*] on the Kinetics-400 dataset.\n\n[*] X3D: Expanding Architectures for Efficient Video Recognition.\n\nSummary\nThis paper proposes a temporal adaptive module for action recognition. The proposed module is straight-forward and obtains good performance on different datasets. The experiments are not thorough enough to demonstrate that the proposed module can be plugged into different 2D CNNs with good performance. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The experiments are not thorough enough.", "review": "This paper proposes a temporal adaptive module for video recognition. Specifically, it decouples dynamic kernel into a location sensitive importance map and a location invariant aggregation weight, which can be plugged into existing 2D CNNs to yield a powerful video architecture with small extra computational cost. The experiments conducted on several datasets demonstrate the effectiveness of the proposed method.\n\nPaper Strength\n(1)\tThe proposed method develops a simple but effective module for video recognition, achieving good performance on various kinds of datasets.\n(2)\tThe proposed two-level adaptive modeling scheme is effective to describe motion patterns. Specifically, it decomposes the video specific temporal kernel into a location sensitive importance map and a location invariant aggregation kernel.\n\nPaper Weakness\n(1)\tFigure 1 is confusing. Why the dimension of the green filter is 3*1*1?\n(2)\tIn Figure 2, TAM is used before the 1*1*3 convolutional layer in the ResNet block. What about the influence of TAM at different locations? Is there any insight for this design?\n(3)\tMoreover, what if we exchange the order of the local branch and the global branch?\n(4)\tIn Section 3.4, the authors claim that the proposed temporal adaptive module can be plugged into existing 2D CNNs with a strong ability to model different temporal structures in video clips. However, only the ResNet-50 backbone is verified in the experiment. More experiments with other backbones such as VGG and Inception should be added.\n(5) In Table 3, the proposed method is not compared to the state-of-the-art X3D method [*] on the Kinetics-400 dataset.\n\n[*] X3D: Expanding Architectures for Efficient Video Recognition.\n\nSummary\nThis paper proposes a temporal adaptive module for action recognition. The proposed module is straight-forward and obtains good performance on different datasets. The experiments are not thorough enough to demonstrate that the proposed module can be plugged into different 2D CNNs with good performance. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603943013644}, {"id": "JADip2YmEUy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper406/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new temporal adaptive module (TAM) to generate video-specific temporal kernels based on its own feature maps. TAM proposes a unique two-level adaptive modeling scheme by decoupling dynamic kernel into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short term information, while the aggregation weight is generated from a global view with a focus on long-term structure.\n\nThe global branch aims to incorporate long-range temporal structure to guide adaptive temporal aggregation with fully connected layers. The adaptive kernel (aggregation weights) is learned based on long-term temporal information, and incorporates global context information and learns to produce the location invariant and also video adaptive convolution kernel for dynamic aggregation. The visualization of the statistics of kernel weights shows that the shapes and scales of distribution are more diverse and data-adaptive. It is indeed reasonable to learn spatiotemporal representation in an adaptive scheme.\n\nHowever, the paper can be improved further.\n1.\tSome SOTA performances are ignored selectively, such as some performances of Slowfast[1]. \n[1] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pp. 6201–6210, 2019\nMaybe the proposed method does not achieve the best performance, but it is necessary to compare with the SOTA methods completely. Deep analysis can help the readers get the core contribution of the paper.\n2.\tThe TAM only focus on the temporal modeling. This may limit the final performance of the network. If the idea can be extended to the spatial-temporal field, it will be more valuable. \n3.\tThe authors give the distributions of the important map V in the local branch and the kernel \\Theta in the global branch. Since these V and \\Theta are all data-dependent, it is not unexpected that the distributions have larger diversities than the traditional I3D’s temporal kernel. If the authors can try to explore the relationship between the distributions and the characteristic of some action samples, beyond the diversity visualization of distributions, it will be more convincing.\n4.\tFig. 1 is difficult to understand. It is not straightforward to figure out how the attention weights or kernel weights are learned. Arrows in Fig. 1 are also confusing. Some indicate names, while some indicate feature flows.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The interpretability of the distributions of the important map V and the kernel \\Theta.", "review": "This paper presents a new temporal adaptive module (TAM) to generate video-specific temporal kernels based on its own feature maps. TAM proposes a unique two-level adaptive modeling scheme by decoupling dynamic kernel into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short term information, while the aggregation weight is generated from a global view with a focus on long-term structure.\n\nThe global branch aims to incorporate long-range temporal structure to guide adaptive temporal aggregation with fully connected layers. The adaptive kernel (aggregation weights) is learned based on long-term temporal information, and incorporates global context information and learns to produce the location invariant and also video adaptive convolution kernel for dynamic aggregation. The visualization of the statistics of kernel weights shows that the shapes and scales of distribution are more diverse and data-adaptive. It is indeed reasonable to learn spatiotemporal representation in an adaptive scheme.\n\nHowever, the paper can be improved further.\n1.\tSome SOTA performances are ignored selectively, such as some performances of Slowfast[1]. \n[1] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pp. 6201–6210, 2019\nMaybe the proposed method does not achieve the best performance, but it is necessary to compare with the SOTA methods completely. Deep analysis can help the readers get the core contribution of the paper.\n2.\tThe TAM only focus on the temporal modeling. This may limit the final performance of the network. If the idea can be extended to the spatial-temporal field, it will be more valuable. \n3.\tThe authors give the distributions of the important map V in the local branch and the kernel \\Theta in the global branch. Since these V and \\Theta are all data-dependent, it is not unexpected that the distributions have larger diversities than the traditional I3D’s temporal kernel. If the authors can try to explore the relationship between the distributions and the characteristic of some action samples, beyond the diversity visualization of distributions, it will be more convincing.\n4.\tFig. 1 is difficult to understand. It is not straightforward to figure out how the attention weights or kernel weights are learned. Arrows in Fig. 1 are also confusing. Some indicate names, while some indicate feature flows.  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603846384230}], "openreview_url": "https://openreview.net/forum?id=i7aMbliTkHs", "arxiv_id": "2005.06803", "paper_pdf": "papers/i7aMbliTkHs.pdf", "paper_pdf_sha256": "7963a986053be91e75a695b6e553b50af05c9cacccc7da913b14a538b691c893", "paper_pdf_bytes": 35096691, "paper_pdf_source": "openreview", "code_url": "https://github.com/liu-zhy/temporal-adaptive-module", "code_repository": "liu-zhy/temporal-adaptive-module", "code_commit": "5fcb46fe671ead2f7958d8ffc3321a0e274d6491", "code_archive": "repos/i7aMbliTkHs.zip", "code_archive_sha256": "9c6c33114bc2975422e2903f8f23f8a9692b4a7e057e5029baba4a935f6e7704", "code_archive_bytes": 582026, "code_file_count": 27, "code_extensions": {".py": 21, ".sh": 6}, "github_disk_usage_kb": 564, "github_languages": {"Python": 173536, "Shell": 1666}, "github_archived": false, "github_pushed_at": "2022-08-18T09:33:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tam-temporal-adaptive-module-for-video"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SylR6n4tPS", "year": 2020, "status": "rejected", "title": "Learning to Generate Grounded Visual Captions without Localization Supervision", "authors": ["Chih-Yao Ma", "Yannis Kalantidis", "Ghassan AlRegib", "Peter Vajda", "Marcus Rohrbach", "Zsolt Kira"], "authorids": ["cyma@gatech.edu", "ykalant@image.ntua.gr", "vajdap@fb.com", "alregib@gatech.edu", "maroffm@gmail.com", "zkira@gatech.edu"], "authors_source": "OpenReview API", "abstract": "When automatically generating a sentence description for an image or video, it often remains unclear how well the generated caption is grounded, or if the model hallucinates based on priors in the dataset and/or the language model. The most common way of relating image regions with words in caption models is through an attention mechanism over the regions that are used as input to predict the next word. The model must therefore learn to predict the attentional weights without knowing the word it should localize. This is difficult to train without grounding supervision since recurrent models can propagate past information and there is no explicit signal to force the captioning model to properly ground the individual decoded words. In this work, we help the model to achieve this via a novel cyclical training regimen that forces the model to localize each word in the image after the sentence decoder generates it, and then reconstruct the sentence from the localized image region(s) to match the ground-truth. Our proposed framework only requires learning one extra fully-connected layer (the localizer), a layer that can be removed at test time. We show that our model significantly improves grounding accuracy without relying on grounding supervision or introducing extra computation during inference for both image and video captioning tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJeXvXSl9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper246/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an architecture that grounds words from a captioning model, but without requiring explicit per-word grounding training data. Instead, they show that it is sufficient to use cycle consistency, verifying that by predicting word->grounding->word the two words are the same. \n\nGeneral: \n\nCycle consistency has been shown to be very useful in replacing explicit paired data, for eample in image-to-image translation (CycleGAN, or the more recent FUNIT). This paper takes it to the domain of vision and language. While the novelty is not very large it seems like a solid step in an interesting direction.  Evaluated on both image and video captioning with substantial localization improvement in the specific relevant eval settings. \n\nSpecific comments: \n\n-- In Table 1, the part of \"Caption Evaluation\" the proposed method is in bold, but it seems that \"Up-Down\" method out-performs the proposed method in B@1 and B@4. \n\n-- Are words that are not nouns/verbs (the/a/are/with/etc) handled differently? It doesn't really make sense to localize them just like object words. \n\n-- The localization model is linear? What would be the effect of richer models on localization accuracy? \n\n-- Qualitative analysis: It would have been useful to add evaluations by human-raters to measure the perceptual quality of the localization. \n\n-- Error analysis? examples and analysis  of failure cases? \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes an architecture that grounds words from a captioning model, but without requiring explicit per-word grounding training data. Instead, they show that it is sufficient to use cycle consistency, verifying that by predicting word->grounding->word the two words are the same. \n\nGeneral: \n\nCycle consistency has been shown to be very useful in replacing explicit paired data, for eample in image-to-image translation (CycleGAN, or the more recent FUNIT). This paper takes it to the domain of vision and language. While the novelty is not very large it seems like a solid step in an interesting direction.  Evaluated on both image and video captioning with substantial localization improvement in the specific relevant eval settings. \n\nSpecific comments: \n\n-- In Table 1, the part of \"Caption Evaluation\" the proposed method is in bold, but it seems that \"Up-Down\" method out-performs the proposed method in B@1 and B@4. \n\n-- Are words that are not nouns/verbs (the/a/are/with/etc) handled differently? It doesn't really make sense to localize them just like object words. \n\n-- The localization model is linear? What would be the effect of richer models on localization accuracy? \n\n-- Qualitative analysis: It would have been useful to add evaluations by human-raters to measure the perceptual quality of the localization. \n\n-- Error analysis? examples and analysis  of failure cases? \n\n"}, "tcdate": 1571996506732}, {"id": "Hkg1_5l1cB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper246/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses captioning generation for images and videos, by proposing a novel cyclical training regimen consisting of three steps: decoding, localization, and reconstruction. The experimental results show that the performance on image captioning and video captioning are improved without grounding supervision.\n\nI lean to accept this paper. The motivation using cyclic feedback itself is not so novel for language generation, but focusing on grounding without localization supervision for visual captioning is interesting. The experimental results show that the proposed method can boost performance both qualitatively and qualitatively. I have several comments and questions below.\n- Why do the authors introduce GVD without self-attention as a baseline? Table 1 and 2 show that removing self-attention degrades the performance. If the combination of self-attention in GVD and cyclical training proposed in this paper is complementary to each other, it does help to improve the overall accuracy.  \n- While the authors develop a cyclical training pipeline, including decoding, localization, and reconstruction, Figure 1 does not show which part corresponds to the decoding phase. The authors should clarify it to make the paper easier to be understood.\n- Equation (5) seems to be strange. $\\theta^*$, a sum of two parameters for each arg max operator, doesn't guarantee that each term in the right side of Eq. (5) keeps its max. This equation seems to be a conceptual one, and the actual training would be performed according to Eq. (7). Therefore, the experimental results might not be influenced by the error in Eq. (5).\n- $\\hat{r}^l_t=\\beta_t^\\top R$ between Eq. (6) and Eq. (7) means that $\\hat{r}^l_t$ is a row vector while $r_n$ seems to be a column vector. Since $R = [r_1, r_2, ..., r_N]$ for $N$ regions, $\\hat{r}^l_t= R \\beta_t$ seems to be appropriate. The authors should correct it. I have a similar comment for $\\hat{r}_t = \\alpha_t^\\top R$ in Eq. (2).\n- In the caption of Table 5, the number equal to or smaller than ten should be spelled out; \"5 runs\" should be \"five runs.\" There are similar errors, such as \"5 GT captions\" and \"1 GT caption\" in Sec. 4.\n- The format of items in References is not consistent.\n- According to Sec. A.4.1, $\\lambda_1$ and $\\lambda_2$ are tuned between 0 and 1. How are the experimental results sensitive to these hyperparameters? Additional experiments using different $\\lambda_1$ and $\\lambda_2$ would be helpful.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper addresses captioning generation for images and videos, by proposing a novel cyclical training regimen consisting of three steps: decoding, localization, and reconstruction. The experimental results show that the performance on image captioning and video captioning are improved without grounding supervision.\n\nI lean to accept this paper. The motivation using cyclic feedback itself is not so novel for language generation, but focusing on grounding without localization supervision for visual captioning is interesting. The experimental results show that the proposed method can boost performance both qualitatively and qualitatively. I have several comments and questions below.\n- Why do the authors introduce GVD without self-attention as a baseline? Table 1 and 2 show that removing self-attention degrades the performance. If the combination of self-attention in GVD and cyclical training proposed in this paper is complementary to each other, it does help to improve the overall accuracy.  \n- While the authors develop a cyclical training pipeline, including decoding, localization, and reconstruction, Figure 1 does not show which part corresponds to the decoding phase. The authors should clarify it to make the paper easier to be understood.\n- Equation (5) seems to be strange. $\\theta^*$, a sum of two parameters for each arg max operator, doesn't guarantee that each term in the right side of Eq. (5) keeps its max. This equation seems to be a conceptual one, and the actual training would be performed according to Eq. (7). Therefore, the experimental results might not be influenced by the error in Eq. (5).\n- $\\hat{r}^l_t=\\beta_t^\\top R$ between Eq. (6) and Eq. (7) means that $\\hat{r}^l_t$ is a row vector while $r_n$ seems to be a column vector. Since $R = [r_1, r_2, ..., r_N]$ for $N$ regions, $\\hat{r}^l_t= R \\beta_t$ seems to be appropriate. The authors should correct it. I have a similar comment for $\\hat{r}_t = \\alpha_t^\\top R$ in Eq. (2).\n- In the caption of Table 5, the number equal to or smaller than ten should be spelled out; \"5 runs\" should be \"five runs.\" There are similar errors, such as \"5 GT captions\" and \"1 GT caption\" in Sec. 4.\n- The format of items in References is not consistent.\n- According to Sec. A.4.1, $\\lambda_1$ and $\\lambda_2$ are tuned between 0 and 1. How are the experimental results sensitive to these hyperparameters? Additional experiments using different $\\lambda_1$ and $\\lambda_2$ would be helpful."}, "tcdate": 1571912295435}, {"id": "BygE12YaKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper246/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# 1. Summary\nThe paper deals with the problem of learning grounded captions from images without joint text-location information, but instead texts (words in captions) and locations are provided independently and the model needs to figure out their link. The model is built upon GVD (Zhou et al., 2019): each word generated by the encoder is grounded to the locations provided by the region proposal module (Up-Down model (Anderson et al., 2018)), used to reconstruct the ground-truth caption.     \n\nMy weak reject decision was guided by the following strengths and weaknesses of the paper.\n\nStrengths:\n* The reconstruction formulation of the problem is interesting and relevant for the task\n* Proposed new metric to measure grounding performance\n      \nWeaknesses:\n* Questionable motivations: it is not clear what application is grounding text to the image useful for?\n* Marginal (not statistically significant?) improvement on image captioning metrics by using grounded text (proposed method) compared to not grounding (GVD)\n* Limited novelty: extension of GVD, where attention is removed and object locations are used instead\n     \n      \n# 2. Clarity and Motivation\nThe paper is generally well written, however there are some concerns on motivations. \n\nOne concern is related to the motivations of the paper. The authors use grounding as a proxy to improve image captioning results, which improvement is marginal wrt GVD (see Table 1). Why do we need to localize text if this has very marginal impact on the captioning metrics? It is missing the link between the potential applications where the localization of words is relevant. \n\nThe authors claim that they do not uses any grounding annotation, however the pre-trained Faster-RCNN has been trained using annotations which consist of bounding boxes + categories. Therefore, the model do (in an implicit way) rely on grounding annotations, especially because there might be an overlap between the words (classes) used to pretrain the detector and the words in the captions. The authors should assess if this ovelap/bias in the pre-trained Faster-RCNN exists or not.\n\nSome other questions are still to be answered:\n* How are the regions R parametrized? Is it the visual representation or bounding box locations?\n* What happens to words that are not grounded to the image (e.g., verbs or articles)? Do you have a special way to deal with those?\n* What is the intuition of multiplying word embedding and region embeddings to generate z in Eq. 6? \n\n\n# 3. Novelty\nThe proposed method is an extension of the existing model GVD, where the attentional module is removed and its functionality is replaced by the cyclical training mode with reconstruction of the object locations. From the technical point of view this is limited novelty, but still an interesting improvement of the model; however the experiments and results do not support the claim that using such model improves image captioning result in a significant way. One way to answer to this question would have been by showing an application where the outputted locations are used for downstream tasks.\n\n\n# 4. Experimentation\nThe experiments are carried out in a scrupulous way, by showing the comparing with GVD (with and without attention; with and without grounding supervision). The non-convincing part of them (as mentioned above already) is the fact that the improvements on these datasets might be non significant for image captioning. For example, let's consider the image captioning results in Table 1 (Flickr30k Entities): cyclical have a max improvement of 0.7 (CIDER) and min of 0 (B@4) when compared with GVD without grounding supervision. There is an obvious huge improvement on the grounding evaluation, which is obvious since GVD does not do it explicitly. The same trend is in Table 2. \nThese results are not convincing, combined by the fact that it is not clear in which applications one would want a very accurate grounded text.\n\n\n# Minor Points\n* Sec. 3.1: it is not clear that the Language LSTM is the decoder. Please explicitly say it before Eq. 1\n* Caption of Table 3: which dataset is this?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "# 1. Summary\nThe paper deals with the problem of learning grounded captions from images without joint text-location information, but instead texts (words in captions) and locations are provided independently and the model needs to figure out their link. The model is built upon GVD (Zhou et al., 2019): each word generated by the encoder is grounded to the locations provided by the region proposal module (Up-Down model (Anderson et al., 2018)), used to reconstruct the ground-truth caption.     \n\nMy weak reject decision was guided by the following strengths and weaknesses of the paper.\n\nStrengths:\n* The reconstruction formulation of the problem is interesting and relevant for the task\n* Proposed new metric to measure grounding performance\n      \nWeaknesses:\n* Questionable motivations: it is not clear what application is grounding text to the image useful for?\n* Marginal (not statistically significant?) improvement on image captioning metrics by using grounded text (proposed method) compared to not grounding (GVD)\n* Limited novelty: extension of GVD, where attention is removed and object locations are used instead\n     \n      \n# 2. Clarity and Motivation\nThe paper is generally well written, however there are some concerns on motivations. \n\nOne concern is related to the motivations of the paper. The authors use grounding as a proxy to improve image captioning results, which improvement is marginal wrt GVD (see Table 1). Why do we need to localize text if this has very marginal impact on the captioning metrics? It is missing the link between the potential applications where the localization of words is relevant. \n\nThe authors claim that they do not uses any grounding annotation, however the pre-trained Faster-RCNN has been trained using annotations which consist of bounding boxes + categories. Therefore, the model do (in an implicit way) rely on grounding annotations, especially because there might be an overlap between the words (classes) used to pretrain the detector and the words in the captions. The authors should assess if this ovelap/bias in the pre-trained Faster-RCNN exists or not.\n\nSome other questions are still to be answered:\n* How are the regions R parametrized? Is it the visual representation or bounding box locations?\n* What happens to words that are not grounded to the image (e.g., verbs or articles)? Do you have a special way to deal with those?\n* What is the intuition of multiplying word embedding and region embeddings to generate z in Eq. 6? \n\n\n# 3. Novelty\nThe proposed method is an extension of the existing model GVD, where the attentional module is removed and its functionality is replaced by the cyclical training mode with reconstruction of the object locations. From the technical point of view this is limited novelty, but still an interesting improvement of the model; however the experiments and results do not support the claim that using such model improves image captioning result in a significant way. One way to answer to this question would have been by showing an application where the outputted locations are used for downstream tasks.\n\n\n# 4. Experimentation\nThe experiments are carried out in a scrupulous way, by showing the comparing with GVD (with and without attention; with and without grounding supervision). The non-convincing part of them (as mentioned above already) is the fact that the improvements on these datasets might be non significant for image captioning. For example, let's consider the image captioning results in Table 1 (Flickr30k Entities): cyclical have a max improvement of 0.7 (CIDER) and min of 0 (B@4) when compared with GVD without grounding supervision. There is an obvious huge improvement on the grounding evaluation, which is obvious since GVD does not do it explicitly. The same trend is in Table 2. \nThese results are not convincing, combined by the fact that it is not clear in which applications one would want a very accurate grounded text.\n\n\n# Minor Points\n* Sec. 3.1: it is not clear that the Language LSTM is the decoder. Please explicitly say it before Eq. 1\n* Caption of Table 3: which dataset is this?\n"}, "tcdate": 1571818460020}], "openreview_url": "https://openreview.net/forum?id=SylR6n4tPS", "arxiv_id": null, "paper_pdf": "papers/SylR6n4tPS.pdf", "paper_pdf_sha256": "4bab70ed0033a01af38d3698ee7ec5b1d9bb991b837eb37ce8e03f2a7d3650ad", "paper_pdf_bytes": 6798464, "paper_pdf_source": "openreview", "code_url": "https://github.com/chihyaoma/cyclical-visual-captioning", "code_repository": "chihyaoma/cyclical-visual-captioning", "code_commit": "337eee17a8789e58d7bf687ad9ecf27c8da44f14", "code_archive": "repos/SylR6n4tPS.zip", "code_archive_sha256": "f8f5a036cb7ded7d2e2ab40c2533cb5023ac497531b0706e172063bc1a34e869", "code_archive_bytes": 566261, "code_file_count": 17, "code_extensions": {".py": 16, ".sh": 1}, "github_disk_usage_kb": 945, "github_languages": {"Python": 161584, "Shell": 2140}, "github_archived": false, "github_pushed_at": "2020-07-29T22:22:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-generate-grounded-visual-captions"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xQIvTUL1Tf", "year": 2026, "status": "rejected", "title": "Efficient Reasoning with Hidden Thinking", "authors": ["Xuan Shen", "Yizhou Wang", "Yufa Zhou", "Xiangxi Shi", "Pu Zhao", "Yanzhi Wang", "Jiuxiang Gu"], "authorids": ["~Xuan_Shen1", "~Yizhou_Wang3", "~Yufa_Zhou1", "~Xiangxi_Shi1", "~Pu_Zhao1", "~Yanzhi_Wang3", "~Jiuxiang_Gu2"], "authors_source": "OpenReview API", "abstract": "Chain-of-Thought (CoT) reasoning has become a powerful framework for improving complex problem-solving capabilities in Multimodal Large Language Models (MLLMs).\nHowever, the verbose nature of textual reasoning introduces significant inefficiencies.\nIn this work, we propose **Heima** (as hidden llama), an effective CoT compression framework that condenses lengthy CoTs into a small set of abstract thinking tokens, preserving essential reasoning while removing redundancy.\nWe then conduct a theoretical analysis from an information-theoretic perspective, quantifying the information gap induced by compression, showing that reasoning capability is preserved when non-trivial mutual information is retained.\nTo further explore and quantify this information gap, we design the adaptive interpreter that maps thinking tokens back to variable-length textual sequences, thereby reconstructing the reasoning process.\nExperiments across diverse reasoning benchmarks demonstrate that Heima improves reasoning efficiency, while maintaining or even achieving better zero-shot accuracy.\nMoreover, the interpreter reconstructs coherent reasoning progresses from compressed thinking tokens, revealing that the information gap is minimal and validating the effectiveness of the proposed framework.\nThis work paves the way for scalable latent reasoning models and advances our understanding of efficient reasoning processes in large models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "tOtRy8xfRZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7846/Reviewer_Y7SG"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes **Heima**, a latent CoT–compression framework for MLLMs that replaces verbose textual chains of thought with a small number of **special “thinking tokens”**, learned via **progressive distillation**. It also introduces **LLM interpreters** that map the hidden states of these tokens back to text to estimate the “information gap,” alongside a brief information-theoretic argument (via DPI) that compressed tokens should retain task-relevant info.", "review_text": "The paper proposes **Heima**, a latent CoT–compression framework for MLLMs that replaces verbose textual chains of thought with a small number of **special “thinking tokens”**, learned via **progressive distillation**. It also introduces **LLM interpreters** that map the hidden states of these tokens back to text to estimate the “information gap,” alongside a brief information-theoretic argument (via DPI) that compressed tokens should retain task-relevant info.", "strengths": "- Experiments on MMStar, MMBench, MMVet, MathVista, AI2D and HallusionBench claim comparable accuracy to CoT baselines while generating far fewer tokens.\n- **Interpreter idea for auditing latent reasoning**: training text-only LLMs to reconstruct reasoning from hidden states is a useful diagnostic to probe whether compressed tokens preserve semantics.\n\n- **Token-efficiency results**: On several benchmarks, the method reports sizable **token reductions** (e.g., to ~6–10% of CoT tokens in places) with modest accuracy loss and sometimes gains relative to LLaVA-CoT.", "weaknesses": "- **Insufficient novelty and unclear justification.** The method is highly similar to *Coconut* (latent reasoning through continuous hidden states) and other latent CoT approaches such as *Cocomix*. The paper’s claim of being the first to extend such latent reasoning to MLLMs feels incremental.\n- **Lack of rationale for multimodal adaptation.** The authors do not explain why a Coconut-style framework should work particularly well in the multimodal setting. Simply applying a latent CoT designed for text-only reasoning to MLLMs requires more theoretical or empirical justification.\n- **Incomplete efficiency analysis.** The paper emphasizes token reduction but lacks concrete wall-clock latency, memory, or FLOP statistics under matched decoding setups.", "questions": "- Apart from the application domain, what are the most substantial differences between Heima and Coconut? Could other latent CoT approaches (e.g., Cocomix, CODI) also be applied in MLLMs?\n- Coconut’s results were not particularly strong in pure NLP reasoning tasks. Why does a similar staged-distillation framework yield competitive results here?\n- Why did the authors choose the Coconut paradigm over other latent reasoning methods? What multimodal properties make it more suitable?\n- How does Heima perform when handling inputs with heavier visual content, such as multi-image or video scenarios?\n- The experiments are limited to a narrow set of base models. How would **Heima** perform on more recent and widely adopted architectures such as **Qwen2.5-VL** or **InternVL 3**, which have stronger multimodal grounding?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes **Heima**, a latent CoT–compression framework for MLLMs that replaces verbose textual chains of thought with a small number of **special “thinking tokens”**, learned via **progressive distillation**. It also introduces **LLM interpreters** that map the hidden states of these tokens back to text to estimate the “information gap,” alongside a brief information-theoretic argument (via DPI) that compressed tokens should retain task-relevant info.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- Experiments on MMStar, MMBench, MMVet, MathVista, AI2D and HallusionBench claim comparable accuracy to CoT baselines while generating far fewer tokens.\n- **Interpreter idea for auditing latent reasoning**: training text-only LLMs to reconstruct reasoning from hidden states is a useful diagnostic to probe whether compressed tokens preserve semantics.\n\n- **Token-efficiency results**: On several benchmarks, the method reports sizable **token reductions** (e.g., to ~6–10% of CoT tokens in places) with modest accuracy loss and sometimes gains relative to LLaVA-CoT.", "weaknesses": "- **Insufficient novelty and unclear justification.** The method is highly similar to *Coconut* (latent reasoning through continuous hidden states) and other latent CoT approaches such as *Cocomix*. The paper’s claim of being the first to extend such latent reasoning to MLLMs feels incremental.\n- **Lack of rationale for multimodal adaptation.** The authors do not explain why a Coconut-style framework should work particularly well in the multimodal setting. Simply applying a latent CoT designed for text-only reasoning to MLLMs requires more theoretical or empirical justification.\n- **Incomplete efficiency analysis.** The paper emphasizes token reduction but lacks concrete wall-clock latency, memory, or FLOP statistics under matched decoding setups.", "questions": "- Apart from the application domain, what are the most substantial differences between Heima and Coconut? Could other latent CoT approaches (e.g., Cocomix, CODI) also be applied in MLLMs?\n- Coconut’s results were not particularly strong in pure NLP reasoning tasks. Why does a similar staged-distillation framework yield competitive results here?\n- Why did the authors choose the Coconut paradigm over other latent reasoning methods? What multimodal properties make it more suitable?\n- How does Heima perform when handling inputs with heavier visual content, such as multi-image or video scenarios?\n- The experiments are limited to a narrow set of base models. How would **Heima** perform on more recent and widely adopted architectures such as **Qwen2.5-VL** or **InternVL 3**, which have stronger multimodal grounding?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761924874804}, {"id": "lPA9tUo43Q", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7846/Reviewer_Dkrj"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes Heima, a framework for compressing Chain-of-Thought (CoT). Heima compresses the content of a CoT into an implicit Tinking Token, significantly reducing the number of tokens, and designs an interpreter to restore the CoT Token to text. The paper trains LlaVA-CoT and compares it with the original model on multiple datasets, demonstrating that Heima achieves excellent results in saving tokens.", "review_text": "This paper proposes Heima, a framework for compressing Chain-of-Thought (CoT). Heima compresses the content of a CoT into an implicit Tinking Token, significantly reducing the number of tokens, and designs an interpreter to restore the CoT Token to text. The paper trains LlaVA-CoT and compares it with the original model on multiple datasets, demonstrating that Heima achieves excellent results in saving tokens.", "strengths": "1. This paper proposes a framework for compressing CoT, shrinking CoT into a single token, significantly reducing the length of the model's response while maintaining its quality.\n\n2. This paper uses information-theoretic analysis and an interpreter to verify the effectiveness and rationality of the compression. The interpreter successfully reconstructs the reasoning process, confirming the minimization of the information gap.", "weaknesses": "1. The experimental setup in this paper primarily focuses on short-step reasoning. For longer, more complex reasoning tasks, such as intricate mathematical proofs or logical deductions, compressing the reasoning content into a single token raises questions about its impact on key information retention and model capability, requiring further experimental analysis. Adding longer-step reasoning might improve the paper's generalization ability.\n\n2. In the experiments of Section 4, the paper only compares Heima with the pedestal model. Perhaps adding comparisons with other methods to improve reasoning efficiency would better highlight Heima's high efficiency.\n\n3. In CoT task reasoning, the model may need to dynamically adjust subsequent reasoning paths based on intermediate results, such as in Rethinking. Heima compresses reasoning information into a single token. Will this affect the aforementioned dynamic adjustment, and will it limit the model's exploratory reasoning capabilities?", "questions": "Please see the  weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Heima, a framework for compressing Chain-of-Thought (CoT). Heima compresses the content of a CoT into an implicit Tinking Token, significantly reducing the number of tokens, and designs an interpreter to restore the CoT Token to text. The paper trains LlaVA-CoT and compares it with the original model on multiple datasets, demonstrating that Heima achieves excellent results in saving tokens.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. This paper proposes a framework for compressing CoT, shrinking CoT into a single token, significantly reducing the length of the model's response while maintaining its quality.\n\n2. This paper uses information-theoretic analysis and an interpreter to verify the effectiveness and rationality of the compression. The interpreter successfully reconstructs the reasoning process, confirming the minimization of the information gap.", "weaknesses": "1. The experimental setup in this paper primarily focuses on short-step reasoning. For longer, more complex reasoning tasks, such as intricate mathematical proofs or logical deductions, compressing the reasoning content into a single token raises questions about its impact on key information retention and model capability, requiring further experimental analysis. Adding longer-step reasoning might improve the paper's generalization ability.\n\n2. In the experiments of Section 4, the paper only compares Heima with the pedestal model. Perhaps adding comparisons with other methods to improve reasoning efficiency would better highlight Heima's high efficiency.\n\n3. In CoT task reasoning, the model may need to dynamically adjust subsequent reasoning paths based on intermediate results, such as in Rethinking. Heima compresses reasoning information into a single token. Will this affect the aforementioned dynamic adjustment, and will it limit the model's exploratory reasoning capabilities?", "questions": "Please see the  weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761910310139}, {"id": "VEyDvDJqlC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7846/Reviewer_S84H"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper addresses the inefficiency of CoT reasoning in MLLMs, which generates lengthy textual tokens and incurs high computational costs. To tackle the problem, this paper proposes Heima, a framework that compresses verbose CoTs into compact thinking tokens in latent space through progressive distillation. In addition, an interpreter is designed to reconstruct reasoning texts from thinking tokens for validation. Experiments show that Heima maintains or even surpasses the accuracy of the original CoT model while using only about 6% of the tokens, demonstrating efficient and effective latent-space reasoning.", "review_text": "This paper addresses the inefficiency of CoT reasoning in MLLMs, which generates lengthy textual tokens and incurs high computational costs. To tackle the problem, this paper proposes Heima, a framework that compresses verbose CoTs into compact thinking tokens in latent space through progressive distillation. In addition, an interpreter is designed to reconstruct reasoning texts from thinking tokens for validation. Experiments show that Heima maintains or even surpasses the accuracy of the original CoT model while using only about 6% of the tokens, demonstrating efficient and effective latent-space reasoning.", "strengths": "1. The motivation of this work is interesting, as compressing Chain-of-Thought (CoT) to accelerate reasoning can effectively reduce computational costs, particularly in resource-intensive MLLM scenarios.\n2. The proposed method is straightforward and effective, with comprehensive experiments validating its efficacy.\n3. The information-theoretic analysis provides a solid theoretical foundation for the effectiveness of CoT compression, which is highly appreciated.", "weaknesses": "1. Although the idea of compressing CoT in MLLMs is promising, similar approaches have been extensively explored in the context of LLMs, suggesting that the methodological novelty may be somewhat limited.\n2. Given that CoT reasoning demonstrates significant performance gains on more challenging tasks, the experimental evaluation should be extended to include more complex reasoning benchmarks, such as MathVision or OlympiadBench.\n3. Beyond the LLaVA-Next and Llama3.2-11B-Vision models, it remains unclear whether the method generalizes effectively to other architectures, such as the Qwen-VL series. Theoretically, this should be a universally applicable approach.", "questions": "1. Can this method be effectively applied to long-chain CoT scenarios, such as those in QvQ or Virgo? If so, would the compression ratio be higher compared to standard CoT, and would there be a significant degradation in performance?\n2. Additional experimental results should be provided to further substantiate the claims and explore the method's boundaries.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the inefficiency of CoT reasoning in MLLMs, which generates lengthy textual tokens and incurs high computational costs. To tackle the problem, this paper proposes Heima, a framework that compresses verbose CoTs into compact thinking tokens in latent space through progressive distillation. In addition, an interpreter is designed to reconstruct reasoning texts from thinking tokens for validation. Experiments show that Heima maintains or even surpasses the accuracy of the original CoT model while using only about 6% of the tokens, demonstrating efficient and effective latent-space reasoning.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The motivation of this work is interesting, as compressing Chain-of-Thought (CoT) to accelerate reasoning can effectively reduce computational costs, particularly in resource-intensive MLLM scenarios.\n2. The proposed method is straightforward and effective, with comprehensive experiments validating its efficacy.\n3. The information-theoretic analysis provides a solid theoretical foundation for the effectiveness of CoT compression, which is highly appreciated.", "weaknesses": "1. Although the idea of compressing CoT in MLLMs is promising, similar approaches have been extensively explored in the context of LLMs, suggesting that the methodological novelty may be somewhat limited.\n2. Given that CoT reasoning demonstrates significant performance gains on more challenging tasks, the experimental evaluation should be extended to include more complex reasoning benchmarks, such as MathVision or OlympiadBench.\n3. Beyond the LLaVA-Next and Llama3.2-11B-Vision models, it remains unclear whether the method generalizes effectively to other architectures, such as the Qwen-VL series. Theoretically, this should be a universally applicable approach.", "questions": "1. Can this method be effectively applied to long-chain CoT scenarios, such as those in QvQ or Virgo? If so, would the compression ratio be higher compared to standard CoT, and would there be a significant degradation in performance?\n2. Additional experimental results should be provided to further substantiate the claims and explore the method's boundaries.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761639449912}, {"id": "DqghgimdUH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7846/Reviewer_2DxQ"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper considers multi-modal reasoning efficiency issues where input as multimodal (i.e., one image and one question), and textual reasoning with the final answers. Heima is proposed to condense each textual CoT steps into one specific token in the vocab via progressive distillation, and the paper additionally provide information-theoretic perspective analysis and tetxual CoT reconstruction via additional trained interpreter. The later part mainly is used to confirm the effectiveness of the proposed method, without other gains for the paper. The experimenal results validate the effective of Heima over two simple and weak baselines.", "review_text": "The paper considers multi-modal reasoning efficiency issues where input as multimodal (i.e., one image and one question), and textual reasoning with the final answers. Heima is proposed to condense each textual CoT steps into one specific token in the vocab via progressive distillation, and the paper additionally provide information-theoretic perspective analysis and tetxual CoT reconstruction via additional trained interpreter. The later part mainly is used to confirm the effectiveness of the proposed method, without other gains for the paper. The experimenal results validate the effective of Heima over two simple and weak baselines.", "strengths": "1. The targeted problem is interesting and valuable, but the aimed scope is limited.\n2. The experiemntal results shows limited effectiveness.", "weaknesses": "1. the proposed method only consider one case that input is multi-modal, and the left part all based on textual reasoning and will not be affected by the input side. from this perspective, the contribution as first reasoning acceleration framework for MLLMs is weaken.\n2. the progressive distillation is widely used in latent space reasoning, e.g., cocount.\n3. despite the paper provide  information-theoretic perspective analysis, it seems no valueable insights, i.e., upper bound and lower bound, just extreme ideal situations.\n4. also, the effectiveness can not be confirmed by the performance of interpreter, since there is no baseline or human evaluation. for example, although the reconsturcted sumamry by the interpreter is 4.1 out of 5 score (2.5/5 for caption), why 4 score is good enough? there is still a significant gap, weaken the statement of both  information-theoretic perspective analysis and value of the interpreter part.\n5. other vicuna family model also is finetuned from llama family, it is more convincing to consider other family like qwen.\n6. efficiency issue, since the method requires massive/multiple supervised fine-tuning in the progressive distillation stage.\n7. baselines are too weak, whether or not the gain is worthy considering the cost to tune the model", "questions": "1. how many cot thinking tokens in total? what about the effect?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers multi-modal reasoning efficiency issues where input as multimodal (i.e., one image and one question), and textual reasoning with the final answers. Heima is proposed to condense each textual CoT steps into one specific token in the vocab via progressive distillation, and the paper additionally provide information-theoretic perspective analysis and tetxual CoT reconstruction via additional trained interpreter. The later part mainly is used to confirm the effectiveness of the proposed method, without other gains for the paper. The experimenal results validate the effective of Heima over two simple and weak baselines.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The targeted problem is interesting and valuable, but the aimed scope is limited.\n2. The experiemntal results shows limited effectiveness.", "weaknesses": "1. the proposed method only consider one case that input is multi-modal, and the left part all based on textual reasoning and will not be affected by the input side. from this perspective, the contribution as first reasoning acceleration framework for MLLMs is weaken.\n2. the progressive distillation is widely used in latent space reasoning, e.g., cocount.\n3. despite the paper provide  information-theoretic perspective analysis, it seems no valueable insights, i.e., upper bound and lower bound, just extreme ideal situations.\n4. also, the effectiveness can not be confirmed by the performance of interpreter, since there is no baseline or human evaluation. for example, although the reconsturcted sumamry by the interpreter is 4.1 out of 5 score (2.5/5 for caption), why 4 score is good enough? there is still a significant gap, weaken the statement of both  information-theoretic perspective analysis and value of the interpreter part.\n5. other vicuna family model also is finetuned from llama family, it is more convincing to consider other family like qwen.\n6. efficiency issue, since the method requires massive/multiple supervised fine-tuning in the progressive distillation stage.\n7. baselines are too weak, whether or not the gain is worthy considering the cost to tune the model", "questions": "1. how many cot thinking tokens in total? what about the effect?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761251792346}], "openreview_url": "https://openreview.net/forum?id=xQIvTUL1Tf", "arxiv_id": "2501.19201", "paper_pdf": "papers/xQIvTUL1Tf.pdf", "paper_pdf_sha256": "6bbcc422e3b2577e8b180d9a4c45961676a2d72b31f323fa3c694b5e06aae769", "paper_pdf_bytes": 773553, "paper_pdf_source": "openreview", "code_url": "https://github.com/shawnricecake/Heima", "code_repository": "shawnricecake/Heima", "code_commit": "cebd6ef3ef862f62c601ced191b64b45d2bf3de2", "code_archive": "repos/xQIvTUL1Tf.zip", "code_archive_sha256": "3d241265227604042b3b93df008ef1564b24bbb219b954a535b4f1a1e701c15f", "code_archive_bytes": 2304672, "code_file_count": 375, "code_extensions": {".py": 365, ".sh": 10}, "github_disk_usage_kb": 2054, "github_languages": {"Python": 3038372, "Shell": 7653}, "github_archived": false, "github_pushed_at": "2026-05-20T03:51:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-reasoning-with-hidden-thinking"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "a6XE2GJHjk", "year": 2025, "status": "rejected", "title": "TabGraphs: A Benchmark and Strong Baselines for Learning on Graphs with Tabular Node Features", "authors": ["Gleb Bazhenov", "Oleg Platonov", "Liudmila Prokhorenkova"], "authorids": ["~Gleb_Bazhenov1", "~Oleg_Platonov1", "~Liudmila_Prokhorenkova1"], "authors_source": "OpenReview API", "abstract": "Tabular machine learning is an important field for industry and science. In this field, table rows are typically treated as independent data samples, but additional information about the relations between these samples is sometimes available and can be used to improve predictive performance. Such information can be naturally modeled with a graph, hence tabular machine learning may benefit from graph machine learning methods. However, graph machine learning models are typically evaluated on datasets with homogeneous, most often text-based node features, which are very different from heterogeneous mixtures of numerical and categorical features present in tabular datasets. Thus, there is a critical difference between the data used in tabular and graph machine learning studies, which does not allow one to understand how successfully graph models can be transferred to tabular data. To bridge this gap, we propose a new benchmark of diverse graphs with heterogeneous tabular node features and realistic prediction tasks. We use this benchmark to evaluate a vast set of models, including simple methods previously overlooked in the literature. Our experiments show that graph neural networks indeed can often bring gains in predictive performance for tabular data, but standard tabular models can also be adapted to work with graph data by using simple graph-based feature augmentation, which sometimes enables them to compete with and even outperform graph neural models. Based on our empirical study, we provide insights for researchers and practitioners in both tabular and graph machine learning fields.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "rMYGsi6OjZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7259/Reviewer_hWYk"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "In this work, the authors aim to study the usefulness of introducing graphs while working with ML for tabular data. To bridge the gap between graph based ML (E.g. techniques such as GNNs, Deepwalk, etc.) and tabular ML methods (e.g. GBDT) - the authors propose to create a benchmark of graph datasets with heterogenous tabular node features across domains such as fraud detection, road networks, etc - with varying number of nodes, edges and other commonly studied graph based statistics. The authors then use the created benchmark to evaluate different ML methods and their combinations.", "review_text": "In this work, the authors aim to study the usefulness of introducing graphs while working with ML for tabular data. To bridge the gap between graph based ML (E.g. techniques such as GNNs, Deepwalk, etc.) and tabular ML methods (e.g. GBDT) - the authors propose to create a benchmark of graph datasets with heterogenous tabular node features across domains such as fraud detection, road networks, etc - with varying number of nodes, edges and other commonly studied graph based statistics. The authors then use the created benchmark to evaluate different ML methods and their combinations.", "strengths": "1. The authors propose multiple datasets which combine graph structure with tabular data -- specifically with heterogenous node data.\n2. The authors compare a wide range of methods on the constructed benchmarks", "weaknesses": "1. The authors appear to ignore the graph based tabular datasets and tasks given presented concisely in the survey --  https://arxiv.org/abs/2401.02143 which survey prior works which have employed GNNs for tabular data.\n2. While the authors state \"If it is possible to define meaningful relations between data samples, it is worth trying to convert the given data to a graph and experiment with ML methods that are capable of processing graph information, as it can lead to significant performance gains\" - and given its largely a benchmark paper - they do not propose a standardized mechanism or methodology to augment any table with graphs - largely limiting the novelty of the contributions. \n4. The analysis accompanying the results presented on the benchmarks also feel handwavy - \"it is important to experiment\nwith different design choices\" - seems to suggest - it is prudent for the downstream users to try every possible method and architecture, and therefore not providing a way to prune the search space of methods or architectures for a given dataset/ problem\n5. As the paper notes, the graph methods proposed here are not novel and have been employed by works such as https://arxiv.org/abs/2004.05718 and others -- even for heterogenous node data - and therefore the paper lacks novelty from the method perspective as well.", "questions": "Please address the weaknesses section as applicable", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors aim to study the usefulness of introducing graphs while working with ML for tabular data. To bridge the gap between graph based ML (E.g. techniques such as GNNs, Deepwalk, etc.) and tabular ML methods (e.g. GBDT) - the authors propose to create a benchmark of graph datasets with heterogenous tabular node features across domains such as fraud detection, road networks, etc - with varying number of nodes, edges and other commonly studied graph based statistics. The authors then use the created benchmark to evaluate different ML methods and their combinations.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The authors propose multiple datasets which combine graph structure with tabular data -- specifically with heterogenous node data.\n2. The authors compare a wide range of methods on the constructed benchmarks", "weaknesses": "1. The authors appear to ignore the graph based tabular datasets and tasks given presented concisely in the survey --  https://arxiv.org/abs/2401.02143 which survey prior works which have employed GNNs for tabular data.\n2. While the authors state \"If it is possible to define meaningful relations between data samples, it is worth trying to convert the given data to a graph and experiment with ML methods that are capable of processing graph information, as it can lead to significant performance gains\" - and given its largely a benchmark paper - they do not propose a standardized mechanism or methodology to augment any table with graphs - largely limiting the novelty of the contributions. \n4. The analysis accompanying the results presented on the benchmarks also feel handwavy - \"it is important to experiment\nwith different design choices\" - seems to suggest - it is prudent for the downstream users to try every possible method and architecture, and therefore not providing a way to prune the search space of methods or architectures for a given dataset/ problem\n5. As the paper notes, the graph methods proposed here are not novel and have been employed by works such as https://arxiv.org/abs/2004.05718 and others -- even for heterogenous node data - and therefore the paper lacks novelty from the method perspective as well.", "questions": "Please address the weaknesses section as applicable", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731905630665}, {"id": "JXVPYiy92v", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7259/Reviewer_SjtV"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "The paper introduces a benchmark to evaluate graph ML methods on tabular data with heterogeneous features. The paper hopes to bridge the gap between tabular and graph machine learning studies. It finds that while GNNs can enhance predictive performance, simple adaptations of standard tabular models could also work well.", "review_text": "The paper introduces a benchmark to evaluate graph ML methods on tabular data with heterogeneous features. The paper hopes to bridge the gap between tabular and graph machine learning studies. It finds that while GNNs can enhance predictive performance, simple adaptations of standard tabular models could also work well.", "strengths": "- The experimental results are comprehensive. I appreciate the efforts from the authors to conduct experiments over 10+ baselines and 10+ datasets\n- The paper cited a large number of related works, which helped the readers to know more about the field.", "weaknesses": "- The problem setting is flawed. I do not agree that graphs with tabular node features are different from relational databases and can be redeemed as a new type of ML problem to work with. Indeed, the authors mentioned that \"In contrast, our work is focused on single-table data with a single type of relationships between entities in the same table\", and led the readers to Appendix C. But even after carefully reading the Appendix C, the paper still does not explain why the setting is different from a relational database with 2 tables (1 for the original table, the other for the additional relational information within the table). In that case, my understanding is that this paper is only a benchmark for relational databases with 2 tables, and only serves as a special case from the existing relational database benchmarks.\n- Moreover, the authors did mention \"Very recently, two benchmarks of large-scale relational databases have been proposed: RelBench (Fey et al., 2023) and 4DBInfer (Wang et al., 2024).\" in the appendix. Given the high correlation with the proposed domain, they should be highlighted in the main paper and be thoroughly discussed, instead of hiding it within 1 sentence in the appendix. I'm not sure able the reason why the relevant discussions should not be mentioned in the main paper.\n- The writing is a bit verbose. For example, the motivation in the introduction is weak. The first 2 paragraphs only introduce why we need tabular learning and graph learning, which is not relevant to the research problem in this paper.\n- The paper makes little to none algorithmic contributions. Overall, I think this paper has not met the publication bar for ICLR.", "questions": "- Why does this paper not emphasize the high relevance of ML for relational databases? Why should relevant discussions go into the Appendix? The \"Machine learning for graphs\" paragraph in the related work took almost a page, but it is less relevant to ML for relational databases for this topic since the setting is essentially just a relational database with 2 tables.\n- Why working with homogenous graphs a benefit, rather than a limitation? Extending homogeneous GNNs to heterogeneous GNNs only takes a few lines of code in PyG, and it makes the GNN model more expressive. I don't find homogeneous GNNs present benefits with heterogeneous GNNs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a benchmark to evaluate graph ML methods on tabular data with heterogeneous features. The paper hopes to bridge the gap between tabular and graph machine learning studies. It finds that while GNNs can enhance predictive performance, simple adaptations of standard tabular models could also work well.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "- The experimental results are comprehensive. I appreciate the efforts from the authors to conduct experiments over 10+ baselines and 10+ datasets\n- The paper cited a large number of related works, which helped the readers to know more about the field.", "weaknesses": "- The problem setting is flawed. I do not agree that graphs with tabular node features are different from relational databases and can be redeemed as a new type of ML problem to work with. Indeed, the authors mentioned that \"In contrast, our work is focused on single-table data with a single type of relationships between entities in the same table\", and led the readers to Appendix C. But even after carefully reading the Appendix C, the paper still does not explain why the setting is different from a relational database with 2 tables (1 for the original table, the other for the additional relational information within the table). In that case, my understanding is that this paper is only a benchmark for relational databases with 2 tables, and only serves as a special case from the existing relational database benchmarks.\n- Moreover, the authors did mention \"Very recently, two benchmarks of large-scale relational databases have been proposed: RelBench (Fey et al., 2023) and 4DBInfer (Wang et al., 2024).\" in the appendix. Given the high correlation with the proposed domain, they should be highlighted in the main paper and be thoroughly discussed, instead of hiding it within 1 sentence in the appendix. I'm not sure able the reason why the relevant discussions should not be mentioned in the main paper.\n- The writing is a bit verbose. For example, the motivation in the introduction is weak. The first 2 paragraphs only introduce why we need tabular learning and graph learning, which is not relevant to the research problem in this paper.\n- The paper makes little to none algorithmic contributions. Overall, I think this paper has not met the publication bar for ICLR.", "questions": "- Why does this paper not emphasize the high relevance of ML for relational databases? Why should relevant discussions go into the Appendix? The \"Machine learning for graphs\" paragraph in the related work took almost a page, but it is less relevant to ML for relational databases for this topic since the setting is essentially just a relational database with 2 tables.\n- Why working with homogenous graphs a benefit, rather than a limitation? Extending homogeneous GNNs to heterogeneous GNNs only takes a few lines of code in PyG, and it makes the GNN model more expressive. I don't find homogeneous GNNs present benefits with heterogeneous GNNs.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730686250154}, {"id": "BAksjmhUSr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7259/Reviewer_CzC7"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper explores enhancing predictive accuracy in tabular machine learning by incorporating graph structures, proposing benchmarks to bridge the gap between heterogeneous tabular and homogeneous graph data features. Results indicate that graph neural networks boost the accuracy of tabular predictions, and simple graph-based enhancements can enable traditional models to effectively compete. Valuable insights are provided for both tabular and graph machine learning practitioners.", "review_text": "The paper explores enhancing predictive accuracy in tabular machine learning by incorporating graph structures, proposing benchmarks to bridge the gap between heterogeneous tabular and homogeneous graph data features. Results indicate that graph neural networks boost the accuracy of tabular predictions, and simple graph-based enhancements can enable traditional models to effectively compete. Valuable insights are provided for both tabular and graph machine learning practitioners.", "strengths": "- **Research Topic.** The studied problem is interesting and practically important for tabular machine learning.\n\n- **Experiments.** Experiments are carried out on 10 datasets with 18 baseline methods, which is quite solid. And there are also enough discussion.", "weaknesses": "- **Lack of Novelty.** The concept of constructing graph datasets from tabular data is not novel and has been previously suggested in multiple studies.\n\n- **Writing.** The initial paragraph in the 'Machine Learning for Graphs' section of the related works is excessively lengthy, potentially hindering readability. Besides, The main text should provide more details about how tabular data is represented as graph data. The absence of these details in the main text may lead to confusion among readers.", "questions": "- More discussion is needed on the differences between this paper and the previous paper on this topic.\n\n- Why there are models that suffer from performance drop after utilizing PLR?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores enhancing predictive accuracy in tabular machine learning by incorporating graph structures, proposing benchmarks to bridge the gap between heterogeneous tabular and homogeneous graph data features. Results indicate that graph neural networks boost the accuracy of tabular predictions, and simple graph-based enhancements can enable traditional models to effectively compete. Valuable insights are provided for both tabular and graph machine learning practitioners.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- **Research Topic.** The studied problem is interesting and practically important for tabular machine learning.\n\n- **Experiments.** Experiments are carried out on 10 datasets with 18 baseline methods, which is quite solid. And there are also enough discussion.", "weaknesses": "- **Lack of Novelty.** The concept of constructing graph datasets from tabular data is not novel and has been previously suggested in multiple studies.\n\n- **Writing.** The initial paragraph in the 'Machine Learning for Graphs' section of the related works is excessively lengthy, potentially hindering readability. Besides, The main text should provide more details about how tabular data is represented as graph data. The absence of these details in the main text may lead to confusion among readers.", "questions": "- More discussion is needed on the differences between this paper and the previous paper on this topic.\n\n- Why there are models that suffer from performance drop after utilizing PLR?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730606296643}, {"id": "eor2H9XkzV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7259/Reviewer_XBq2"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a benchmark for graph learning on single-table tasks(1 type of node and 1 type of edge), with heterogenous features. \nIt constructs 11 tasks from 8 different datasets and ran experiments with GBDTs as well as GNNs on the tasks. \nThe results show that GNNs outperforms GBDTs on the tasks authors proposed.", "review_text": "This paper proposes a benchmark for graph learning on single-table tasks(1 type of node and 1 type of edge), with heterogenous features. \nIt constructs 11 tasks from 8 different datasets and ran experiments with GBDTs as well as GNNs on the tasks. \nThe results show that GNNs outperforms GBDTs on the tasks authors proposed.", "strengths": "Overall the writing is good and the presentation is clear. \nThe experiments the authors provide are extensive, including different types of models, tabular dl models, GBDTs, GNNs, etc.", "weaknesses": "The novel contribution is limited in this paper.\n\nOn page 2, the authors claimed that GNNs are typically evaluated on homogeneous features, but tabular learning typically has heterogenous features. I think they omit the work of PyTorch Frame (Hu, Weihua, et al. \"PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning.\" arXiv preprint arXiv:2404.00776 (2024). ), which supports encoding heterogenous tabular data through deep encoders and can be used with PyG on downstream GNN tasks. It is good to include comparison to PyTorch Frame + PyG in your paper as it also fits into graph + heterogenous tabular data. \n\nThe authors claim on page 4. \"In contrast, our work is focused on single-table data with a single type of relationships\nbetween entities in the same table.\". However, it seems to me all the graph structure constructed by the authors can be represented by the graph structure in Relbench as well.\n\nI also think the task type is too limited. Currently it only supports node classification and node regression. The authors did not mention link prediction tasks. Link prediction is a common tasks solved by GNNs and is offered by Relbench. It is unfair to say node level task is \"the most common setting in modern graph machine learning\".", "questions": "In the experiment section, the authors only compared the performance scores. I think it's also good to compare the runtime of the methods, as GBDTs can be much slower on larger datasets as compared to GNNs and tabular models.\n\nThe authors propose two types of tasks, node regression and classification. I'd like to see some other task types being covered, for example, multi-label classification, or link prediction. For example, in the hnm dataset, can you add a task to predict new items that will be co-purchases?\n\nThe authors should make proper comparisons to existing relational learning benchmarks RelBench and 4DBInfer.\n\nHappy to raise the score if the authors can address all the three comments above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a benchmark for graph learning on single-table tasks(1 type of node and 1 type of edge), with heterogenous features. \nIt constructs 11 tasks from 8 different datasets and ran experiments with GBDTs as well as GNNs on the tasks. \nThe results show that GNNs outperforms GBDTs on the tasks authors proposed.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Overall the writing is good and the presentation is clear. \nThe experiments the authors provide are extensive, including different types of models, tabular dl models, GBDTs, GNNs, etc.", "weaknesses": "The novel contribution is limited in this paper.\n\nOn page 2, the authors claimed that GNNs are typically evaluated on homogeneous features, but tabular learning typically has heterogenous features. I think they omit the work of PyTorch Frame (Hu, Weihua, et al. \"PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning.\" arXiv preprint arXiv:2404.00776 (2024). ), which supports encoding heterogenous tabular data through deep encoders and can be used with PyG on downstream GNN tasks. It is good to include comparison to PyTorch Frame + PyG in your paper as it also fits into graph + heterogenous tabular data. \n\nThe authors claim on page 4. \"In contrast, our work is focused on single-table data with a single type of relationships\nbetween entities in the same table.\". However, it seems to me all the graph structure constructed by the authors can be represented by the graph structure in Relbench as well.\n\nI also think the task type is too limited. Currently it only supports node classification and node regression. The authors did not mention link prediction tasks. Link prediction is a common tasks solved by GNNs and is offered by Relbench. It is unfair to say node level task is \"the most common setting in modern graph machine learning\".", "questions": "In the experiment section, the authors only compared the performance scores. I think it's also good to compare the runtime of the methods, as GBDTs can be much slower on larger datasets as compared to GNNs and tabular models.\n\nThe authors propose two types of tasks, node regression and classification. I'd like to see some other task types being covered, for example, multi-label classification, or link prediction. For example, in the hnm dataset, can you add a task to predict new items that will be co-purchases?\n\nThe authors should make proper comparisons to existing relational learning benchmarks RelBench and 4DBInfer.\n\nHappy to raise the score if the authors can address all the three comments above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729827693231}], "openreview_url": "https://openreview.net/forum?id=a6XE2GJHjk", "arxiv_id": "2409.14500", "paper_pdf": "papers/a6XE2GJHjk.pdf", "paper_pdf_sha256": "f8bfd453f82e37ab43ab222eba058d94559eefb3431694ee4a87107fa47f8cba", "paper_pdf_bytes": 413075, "paper_pdf_source": "openreview", "code_url": "https://github.com/yandex-research/tabgraphs", "code_repository": "yandex-research/tabgraphs", "code_commit": "0ddeca9d715c3e920d98b2b1dbe2f24c19f23b02", "code_archive": "repos/a6XE2GJHjk.zip", "code_archive_sha256": "33b838186071d1db1e275ce57a7f4c9e4976f262b1a6416a1f91513b08c8e172", "code_archive_bytes": 238810, "code_file_count": 61, "code_extensions": {".py": 56, ".ipynb": 4, ".sh": 1}, "github_disk_usage_kb": 143, "github_languages": {"Python": 426420, "Jupyter Notebook": 52844, "Shell": 3874, "Makefile": 422}, "github_archived": true, "github_pushed_at": "2025-10-29T17:41:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tabgraphs-a-benchmark-and-strong-baselines"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UbOzNf6hGq", "year": 2024, "status": "rejected", "title": "FiLM: Fill-in Language Models for Any-Order Generation", "authors": ["Tianxiao Shen", "Hao Peng", "Ruoqi Shen", "Yao Fu", "Zaid Harchaoui", "Yejin Choi"], "authorids": ["~Tianxiao_Shen1", "~Hao_Peng4", "~Ruoqi_Shen1", "~Yao_Fu3", "~Zaid_Harchaoui1", "~Yejin_Choi1"], "authors_source": "OpenReview API", "abstract": "Language models have become the backbone of today's AI systems. However, their predominant left-to-right generation limits the use of bidirectional context, which is essential for tasks that involve filling text in the middle. We propose the Fill-in Language Model (FiLM), a new language modeling approach that allows for flexible generation at any position without adhering to a specific generation order. Its training extends the masked language modeling objective by adopting varying mask probabilities sampled from the Beta distribution to enhance the generative capabilities of FiLM. During inference, FiLM can seamlessly insert missing phrases, sentences, or paragraphs, ensuring that the outputs are fluent and are coherent with the surrounding context. In both automatic and human evaluations, FiLM outperforms existing infilling methods that rely on left-to-right language models trained on rearranged text segments. FiLM is easy to implement and can be either trained from scratch or fine-tuned from a left-to-right language model. Notably, as the model size grows, FiLM's perplexity approaches that of strong left-to-right language models of similar sizes, indicating FiLM's scalability and potential as a large language model.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "lDoEXnVPD4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2098/Reviewer_Bw7d"], "rating": "3: reject, not good enough", "soundness": "4 excellent", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "Summary:\n\nThe paper introduces the Fill-in Language Model (FiLM), which enhances language models by allowing for flexible text generation in any order, thus overcoming the limitations of traditional left-to-right generation and making use of bidirectional context. FiLM's training employs varying mask probabilities drawn from the Beta distribution, optimizing its generative abilities. During inference, it can insert text at any point, maintaining fluency and contextual coherence. FiLM can be trained from scratch or fine-tuned from existing models, showing scalability and potential as a large language model with its perplexity nearing that of conventional models as size increases. In tasks such as text infilling and story completion, FiLM outperforms existing methods, as confirmed by both automatic and human evaluations.", "review_text": "Summary:\n\nThe paper introduces the Fill-in Language Model (FiLM), which enhances language models by allowing for flexible text generation in any order, thus overcoming the limitations of traditional left-to-right generation and making use of bidirectional context. FiLM's training employs varying mask probabilities drawn from the Beta distribution, optimizing its generative abilities. During inference, it can insert text at any point, maintaining fluency and contextual coherence. FiLM can be trained from scratch or fine-tuned from existing models, showing scalability and potential as a large language model with its perplexity nearing that of conventional models as size increases. In tasks such as text infilling and story completion, FiLM outperforms existing methods, as confirmed by both automatic and human evaluations.", "strengths": "Advantages:\n\n - Flexible Text Generation: FiLM allows for generating text in any desired order, not just left-to-right, leveraging bidirectional context which is beneficial for tasks involving text infilling and editing.\n - Adaptive Training Strategy: The model uses varying mask probabilities sampled from the Beta distribution during training, enhancing its generative capabilities and allowing it to adaptively learn from different contexts.\n - Scalability and Performance: It exhibits competitive perplexity compared to causal language models (CLMs), especially as it scales up in size, and it outperforms strong baselines in text infilling and story completion tasks, according to both automatic and human evaluations​.", "weaknesses": "Disadvantages:\n\n - Unfair Comparison: Correct me if I'm wrong but it seems to me the comparison in the infilling task can be unfair. The formulation of FiLM implicitly assumed the total number of infilled tokens, whereas this is not the information that CM-based infilling models know. You may have observed that the output from CM-based infilling models usually are longer compared to FiLM's. This could cause extra difficulties for CM-based infilling and eventually make the comparison unfair and less meaningful.\n - Training/Inference Discrepancy: It seems to me while the proposed mask scheduler can alleviate it, there's still a training/inference behaviorial inconsistency in the proposed algorithm. I would appreciate it if some ablation study can be conducted to show how effective the advanced beta-Distribution-guided masking is and how severe is this discrepancy.\n - Under Explored Backgrounds: There are many preliminary works that seem relevant but not discussed or even mentioned in the paper. InDIGO [https://arxiv.org/abs/1902.01370], for example, is one of the first attempts at generating text in arbitrary order using insertion-based models. It even has an advantage against the proposed approach that it does not assume the length of span of the infilling. Following InDIGO there are also other insertion-based models that can achieve the same goal and can be trained efficiently, like InsNet [https://arxiv.org/abs/2102.11008].", "questions": "Please refer to Weakness).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Summary:\n\nThe paper introduces the Fill-in Language Model (FiLM), which enhances language models by allowing for flexible text generation in any order, thus overcoming the limitations of traditional left-to-right generation and making use of bidirectional context. FiLM's training employs varying mask probabilities drawn from the Beta distribution, optimizing its generative abilities. During inference, it can insert text at any point, maintaining fluency and contextual coherence. FiLM can be trained from scratch or fine-tuned from existing models, showing scalability and potential as a large language model with its perplexity nearing that of conventional models as size increases. In tasks such as text infilling and story completion, FiLM outperforms existing methods, as confirmed by both automatic and human evaluations.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "2 fair", "strengths": "Advantages:\n\n - Flexible Text Generation: FiLM allows for generating text in any desired order, not just left-to-right, leveraging bidirectional context which is beneficial for tasks involving text infilling and editing.\n - Adaptive Training Strategy: The model uses varying mask probabilities sampled from the Beta distribution during training, enhancing its generative capabilities and allowing it to adaptively learn from different contexts.\n - Scalability and Performance: It exhibits competitive perplexity compared to causal language models (CLMs), especially as it scales up in size, and it outperforms strong baselines in text infilling and story completion tasks, according to both automatic and human evaluations​.", "weaknesses": "Disadvantages:\n\n - Unfair Comparison: Correct me if I'm wrong but it seems to me the comparison in the infilling task can be unfair. The formulation of FiLM implicitly assumed the total number of infilled tokens, whereas this is not the information that CM-based infilling models know. You may have observed that the output from CM-based infilling models usually are longer compared to FiLM's. This could cause extra difficulties for CM-based infilling and eventually make the comparison unfair and less meaningful.\n - Training/Inference Discrepancy: It seems to me while the proposed mask scheduler can alleviate it, there's still a training/inference behaviorial inconsistency in the proposed algorithm. I would appreciate it if some ablation study can be conducted to show how effective the advanced beta-Distribution-guided masking is and how severe is this discrepancy.\n - Under Explored Backgrounds: There are many preliminary works that seem relevant but not discussed or even mentioned in the paper. InDIGO [https://arxiv.org/abs/1902.01370], for example, is one of the first attempts at generating text in arbitrary order using insertion-based models. It even has an advantage against the proposed approach that it does not assume the length of span of the infilling. Following InDIGO there are also other insertion-based models that can achieve the same goal and can be trained efficiently, like InsNet [https://arxiv.org/abs/2102.11008].", "questions": "Please refer to Weakness).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698817105205}, {"id": "wpy8DrK0Op", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2098/Reviewer_Q9uj"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose a train an MLM with a new masking ratio instead of a fixed 15% which is the standard in BERT. Specifically, they propose to draw a masking ratio from a beta distribution with a mode of 50% masking ratio.", "review_text": "The authors propose a train an MLM with a new masking ratio instead of a fixed 15% which is the standard in BERT. Specifically, they propose to draw a masking ratio from a beta distribution with a mode of 50% masking ratio.", "strengths": "1. Results show the proposed beta(2.5,2.5) masking ratio is better than uniform[0,1] and fixed ratios.", "weaknesses": "1. Table 1 shows the proposed beta(2.5,2.5) masking ratio is only marginally better than U[0,1] on two datasets.\n2. The paper does not have enough content. For example, the authors use a lot of case studies (Figures 1,4, 7) to waste space.\n3. The authors do not compare with non-autoregressive methods, for example (1), which is important because they both address the non-left-to-right generation problem.\n\n(1) Insertion Transformer: http://proceedings.mlr.press/v97/stern19a/stern19a.pdf", "questions": "1. Since the improvement is not very significant (Table 1), have the authors verified how robust the improvement is? For example, running multiple rounds and finding the average and standard deviation.\n\n2. How does the proposed beta masking ratio work with different alphas and betas other than (2.5,2.5)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a train an MLM with a new masking ratio instead of a fixed 15% which is the standard in BERT. Specifically, they propose to draw a masking ratio from a beta distribution with a mode of 50% masking ratio.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "1. Results show the proposed beta(2.5,2.5) masking ratio is better than uniform[0,1] and fixed ratios.", "weaknesses": "1. Table 1 shows the proposed beta(2.5,2.5) masking ratio is only marginally better than U[0,1] on two datasets.\n2. The paper does not have enough content. For example, the authors use a lot of case studies (Figures 1,4, 7) to waste space.\n3. The authors do not compare with non-autoregressive methods, for example (1), which is important because they both address the non-left-to-right generation problem.\n\n(1) Insertion Transformer: http://proceedings.mlr.press/v97/stern19a/stern19a.pdf", "questions": "1. Since the improvement is not very significant (Table 1), have the authors verified how robust the improvement is? For example, running multiple rounds and finding the average and standard deviation.\n\n2. How does the proposed beta masking ratio work with different alphas and betas other than (2.5,2.5)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698810974762}, {"id": "13zV8Xsatk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2098/Reviewer_Ad3z"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper describes FiLM (Fill-in Language Model), a novel approach to non-causal language modeling, wherein the model predicts multiple masked tokens in an input sequence that can be variable in number and flexible in locations. FiLM is trained using a novel masking schedule based on a Beta distribution, which is shown in this paper to outperform the fixed masking probability and a uniform distribution in terms of achieving lower perplexities on eval datasets. FiLM can be trained on pretrained foundation models that are either masked LMs (MLMs) or causal LMs (CLMs). The authors conducted experiments with RoBERTa as an example of the MLMs and with GPT-2 as an example of CLMs. The authors show that FiLMs achieve substantially higher ROGUE scores (indicating higher quality) for the text infilling task compared to a causal masking (CM) baselines. On a causal language modeling benchmark, FiLM show worse (higher) perplexity compared to a traditional CLM, which was expected. But these perplexity gaps diminish with increasing model size. The authors also provide a few concrete examples of text filled in by FiLM and compared them with outputs of CM on the same input prompts, showing a clear advantage in terms of the coherence of the filled text by FiLM. Evaluation of infilling results based on human raters and GPT-4 show a clear overall preference for FiLM compared to CLMs.", "review_text": "This paper describes FiLM (Fill-in Language Model), a novel approach to non-causal language modeling, wherein the model predicts multiple masked tokens in an input sequence that can be variable in number and flexible in locations. FiLM is trained using a novel masking schedule based on a Beta distribution, which is shown in this paper to outperform the fixed masking probability and a uniform distribution in terms of achieving lower perplexities on eval datasets. FiLM can be trained on pretrained foundation models that are either masked LMs (MLMs) or causal LMs (CLMs). The authors conducted experiments with RoBERTa as an example of the MLMs and with GPT-2 as an example of CLMs. The authors show that FiLMs achieve substantially higher ROGUE scores (indicating higher quality) for the text infilling task compared to a causal masking (CM) baselines. On a causal language modeling benchmark, FiLM show worse (higher) perplexity compared to a traditional CLM, which was expected. But these perplexity gaps diminish with increasing model size. The authors also provide a few concrete examples of text filled in by FiLM and compared them with outputs of CM on the same input prompts, showing a clear advantage in terms of the coherence of the filled text by FiLM. Evaluation of infilling results based on human raters and GPT-4 show a clear overall preference for FiLM compared to CLMs.", "strengths": "S1. A clearly-written report, easy to follow. It describes the general motivation, methodology, and results of FiLM well.\nS2. Introducing a novel noise schedule for masking tokens for the MLM training objective. Convincingly demonstrating its advantage over other noise schedule including the uniform distribution and a fixed mask probability.\nS3. Defining a number of decoding strategies (schedules) for FiLM to accommodate multiple masks in a sequence, including random, left-to-right, right-to-left, min-entropy, and max-entropy. Using empirical results to illustrate the relative advantage of left-to-right and min-entropy. \nS4. Using a diverse set of evaluation datasets including WikiText-103, 1BW, and ROCStories, employing both automated metrics such as ROGUE and subjective metrics such as pairwise preference (according to human raters or GPT4).", "weaknesses": "W1. The presentation in the manuscript lacks some details (see my questions below). But that should be addressable by the authors.\nW2. Lack of evaluation on an out-of-domain evaluation set. The evaluation of FiLM that the authors conducted over the 1BW and ROCStories datasets are based on FiLMs trained on these datasets. Therefore it remains to be shown that the infilling capability of FiLM can be extended to text domains unseen during training. This feels important for the practical usefulness of FiLM in real-world applications. \nW3. Relating to point W2 above, the author should better motivate FiLM and its infilling capability. In what real-world applications would FiLM be useful? Would FiLM be useful as a foundational model for non-infilling tasks such as classification?", "questions": "Q1. Need more information on human evaluation procedures. Who were the human raters? How were they recruited? What was the task like? How were they trained on this task? Was the mask part revealed to the human rater while this task is performed? How was inter-rater reliability (IRR) evaluated?\nQ2. A question related to Q1 was how the GPT4 model was prompted to perform the rating task, including the prompt template, the sampling temperature, etc.\nQ3. What tokenizer(s) did the RoBERTa and GPT-2 models use? For fair comparisons (e.g., results in Figure 5), they ought to be based on the same tokenizer. Some tokenizers are at the subword level. For example, is it possible that certain words are split into two parts and masked partially during mask generation? How does this affect the training, decoding, and human evaluation?\nQ4. How do the training examples look like? What is the output shape of FiLM? Is it the set of probability scores over the vocabulary for a single token, or the sets of scores for multiple tokens. This is related to the question whether a sequence with N masks (where N can be >1 in general) yields a single or multiple (N) training examples.\nQ5. There seems to be some discrepancy between Figures 6 and 7 (right panels). While Figure 6 Right Panel shows only the eval results from GPT4, Figure 7 Right Panels show both human and GPT4 rating results. Why not show human rating results in Figure 6 (i.e., for the WikiText-103 and 1BW datasets) as well?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper describes FiLM (Fill-in Language Model), a novel approach to non-causal language modeling, wherein the model predicts multiple masked tokens in an input sequence that can be variable in number and flexible in locations. FiLM is trained using a novel masking schedule based on a Beta distribution, which is shown in this paper to outperform the fixed masking probability and a uniform distribution in terms of achieving lower perplexities on eval datasets. FiLM can be trained on pretrained foundation models that are either masked LMs (MLMs) or causal LMs (CLMs). The authors conducted experiments with RoBERTa as an example of the MLMs and with GPT-2 as an example of CLMs. The authors show that FiLMs achieve substantially higher ROGUE scores (indicating higher quality) for the text infilling task compared to a causal masking (CM) baselines. On a causal language modeling benchmark, FiLM show worse (higher) perplexity compared to a traditional CLM, which was expected. But these perplexity gaps diminish with increasing model size. The authors also provide a few concrete examples of text filled in by FiLM and compared them with outputs of CM on the same input prompts, showing a clear advantage in terms of the coherence of the filled text by FiLM. Evaluation of infilling results based on human raters and GPT-4 show a clear overall preference for FiLM compared to CLMs.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "S1. A clearly-written report, easy to follow. It describes the general motivation, methodology, and results of FiLM well.\nS2. Introducing a novel noise schedule for masking tokens for the MLM training objective. Convincingly demonstrating its advantage over other noise schedule including the uniform distribution and a fixed mask probability.\nS3. Defining a number of decoding strategies (schedules) for FiLM to accommodate multiple masks in a sequence, including random, left-to-right, right-to-left, min-entropy, and max-entropy. Using empirical results to illustrate the relative advantage of left-to-right and min-entropy. \nS4. Using a diverse set of evaluation datasets including WikiText-103, 1BW, and ROCStories, employing both automated metrics such as ROGUE and subjective metrics such as pairwise preference (according to human raters or GPT4).", "weaknesses": "W1. The presentation in the manuscript lacks some details (see my questions below). But that should be addressable by the authors.\nW2. Lack of evaluation on an out-of-domain evaluation set. The evaluation of FiLM that the authors conducted over the 1BW and ROCStories datasets are based on FiLMs trained on these datasets. Therefore it remains to be shown that the infilling capability of FiLM can be extended to text domains unseen during training. This feels important for the practical usefulness of FiLM in real-world applications. \nW3. Relating to point W2 above, the author should better motivate FiLM and its infilling capability. In what real-world applications would FiLM be useful? Would FiLM be useful as a foundational model for non-infilling tasks such as classification?", "questions": "Q1. Need more information on human evaluation procedures. Who were the human raters? How were they recruited? What was the task like? How were they trained on this task? Was the mask part revealed to the human rater while this task is performed? How was inter-rater reliability (IRR) evaluated?\nQ2. A question related to Q1 was how the GPT4 model was prompted to perform the rating task, including the prompt template, the sampling temperature, etc.\nQ3. What tokenizer(s) did the RoBERTa and GPT-2 models use? For fair comparisons (e.g., results in Figure 5), they ought to be based on the same tokenizer. Some tokenizers are at the subword level. For example, is it possible that certain words are split into two parts and masked partially during mask generation? How does this affect the training, decoding, and human evaluation?\nQ4. How do the training examples look like? What is the output shape of FiLM? Is it the set of probability scores over the vocabulary for a single token, or the sets of scores for multiple tokens. This is related to the question whether a sequence with N masks (where N can be >1 in general) yields a single or multiple (N) training examples.\nQ5. There seems to be some discrepancy between Figures 6 and 7 (right panels). While Figure 6 Right Panel shows only the eval results from GPT4, Figure 7 Right Panels show both human and GPT4 rating results. Why not show human rating results in Figure 6 (i.e., for the WikiText-103 and 1BW datasets) as well?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698743509177}, {"id": "DiS1ODahxI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2098/Reviewer_FCE8"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes the Fill-in Language Model (FiLM), a new language modeling approach that allows for flexible generation at any position without adhering to a specific generation order. Its training extends the masked language modeling objective by adopting\nvarying mask probabilities sampled from the Beta distribution to enhance the generative capabilities of FiLM. During inference, FiLM can seamlessly insert missing phrases, sentences, or paragraphs, ensuring that the outputs are fluent and are coherent with the surrounding context. In both automatic and human evaluations, FiLM outperforms existing infilling methods that rely on left-to-right language models trained on rearranged text segments. FiLM is easy to implement and can be either trained from scratch or fine-tuned from a left-to-right language model.  The results are promising", "review_text": "This paper proposes the Fill-in Language Model (FiLM), a new language modeling approach that allows for flexible generation at any position without adhering to a specific generation order. Its training extends the masked language modeling objective by adopting\nvarying mask probabilities sampled from the Beta distribution to enhance the generative capabilities of FiLM. During inference, FiLM can seamlessly insert missing phrases, sentences, or paragraphs, ensuring that the outputs are fluent and are coherent with the surrounding context. In both automatic and human evaluations, FiLM outperforms existing infilling methods that rely on left-to-right language models trained on rearranged text segments. FiLM is easy to implement and can be either trained from scratch or fine-tuned from a left-to-right language model.  The results are promising", "strengths": "* FiLM's flexibility to generate text in any order is a novel capability lacking in most LMs.\n* The adaptive masking strategy for training is simple but impactful.\n* Computing perplexity for non-causal LMs is an important contribution.\n* Strong quantitative and qualitative results for text infilling and completion.\n* Ablations clearly validate the design choices like Beta distribution masking.", "weaknesses": "* FiLM lags behind causal LMs in terms of perplexity, especially for smaller model sizes.\n* Limited analysis of how perplexity varies across different decoding strategies.\n* No exploration of other pretraining objectives tailored for any-order generation.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the Fill-in Language Model (FiLM), a new language modeling approach that allows for flexible generation at any position without adhering to a specific generation order. Its training extends the masked language modeling objective by adopting\nvarying mask probabilities sampled from the Beta distribution to enhance the generative capabilities of FiLM. During inference, FiLM can seamlessly insert missing phrases, sentences, or paragraphs, ensuring that the outputs are fluent and are coherent with the surrounding context. In both automatic and human evaluations, FiLM outperforms existing infilling methods that rely on left-to-right language models trained on rearranged text segments. FiLM is easy to implement and can be either trained from scratch or fine-tuned from a left-to-right language model.  The results are promising", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "* FiLM's flexibility to generate text in any order is a novel capability lacking in most LMs.\n* The adaptive masking strategy for training is simple but impactful.\n* Computing perplexity for non-causal LMs is an important contribution.\n* Strong quantitative and qualitative results for text infilling and completion.\n* Ablations clearly validate the design choices like Beta distribution masking.", "weaknesses": "* FiLM lags behind causal LMs in terms of perplexity, especially for smaller model sizes.\n* Limited analysis of how perplexity varies across different decoding strategies.\n* No exploration of other pretraining objectives tailored for any-order generation.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698661508525}], "openreview_url": "https://openreview.net/forum?id=UbOzNf6hGq", "arxiv_id": "2310.09930", "paper_pdf": "papers/UbOzNf6hGq.pdf", "paper_pdf_sha256": "d5ade475df72e444da11545c8d4588600ae7cbcb16de5f98ca2a078ca52b34f3", "paper_pdf_bytes": 999602, "paper_pdf_source": "openreview", "code_url": "https://github.com/shentianxiao/FiLM", "code_repository": "shentianxiao/FiLM", "code_commit": "ca22c9fe51786b6b5544deb7b418969afb26eaf3", "code_archive": "repos/UbOzNf6hGq.zip", "code_archive_sha256": "af53b38123381ccc6b9d71c3f707b47b636881804cf8050c3087f24ea833d883", "code_archive_bytes": 121259, "code_file_count": 21, "code_extensions": {".py": 19, ".sh": 2}, "github_disk_usage_kb": 115, "github_languages": {"Python": 31378, "Shell": 1197}, "github_archived": false, "github_pushed_at": "2023-10-18T15:36:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/film-fill-in-language-models-for-any-order"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MPoQtFC588n", "year": 2022, "status": "rejected", "title": "RMNet: Equivalently Removing Residual Connection from Networks", "authors": ["Fanxu Meng", "Hao Cheng", "Jia-Xin Zhuang", "Ke Li", "Xing Sun"], "authorids": ["~Fanxu_Meng1", "~Hao_Cheng5", "~Jia-Xin_Zhuang1", "~Ke_Li4", "~Xing_Sun1"], "authors_source": "OpenReview API", "abstract": "Although residual connection enables training very deep neural networks, it is not friendly for online inference due to its multi-branch topology. This encourages many researchers to work on designing DNNs without residual connections at inference. For example, RepVGG re-parameterizes multi-branch topology to a VGG-like (single-branch) model when deploying, showing great performance when the network is relatively shallow. However, RepVGG can not transform ResNet to VGG equivalently because re-parameterizing methods can only be applied to linear Blocks and the non-linear layers (ReLU) have to be put outside of the residual connection which results in limited representation ability, especially for deeper networks. In this paper, we aim to remedy this problem and propose to remove the residual connection in a vanilla ResNet equivalently by a reserving and merging (RM) operation on ResBlock. Specifically, RM operation allows input feature maps to pass through the block while reserving their information and merges all the information at the end of each block, which can remove residual connection without changing original output. RMNet basically has two advantages: 1) it achieves a better accuracy-speed trade-off compared with ResNet and RepVGG; 2) its implementation makes it naturally friendly for high ratio network pruning. Extensive experiments are performed to verify the effectiveness of RMNet. We believe the ideology of RMNet can inspire many insights on model design for the community in the future.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "E9p2kqCrYcB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper685/Reviewer_BQPh"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes to convert residual network architectures to equivalent \"plain\" networks after training.  This is accomplished by augmenting convolutional layers with Dirac initialized filters (folding extraction of the residual signal into the conv layer), and modifying the subsequent Batch Normalization and ReLU layers (converting ReLU to PReLU) accordingly.  Motivations for doing this are to increase inference throughput, and allow additional pruning of the resulting plain network.\n", "review_text": "As the proposed conversion from residual network to plain network exactly preserves function, any throughput benefits are entirely determined by which form of operations is more efficient for the underlying combination of software library and compute hardware running the network.  Absent knowledge of such interactions, there does not appear to be any intrinsic reason for preferring the proposed form for rewriting residual layers.  In fact, the proposed approach seems to add computational work (more convolutional filters, Dirac initialized) for the sole purpose of taking advantage of fast implementations of convolutional layers.  Would a better approach simply be to write a custom fused low-level kernel for executing an entire residual block?  What form is actually better if one considers the optimal implementation possible for the underlying hardware (e.g., CPU, GPU, or TPU)?\n\nA similar question arises for the pruning approach.  If the functional forms of the residual and plain networks are equivalent, then any pruning operations on one form should have an equivalent expression in the other.  Why is it necessary to first convert to a plain network, instead of simply considering pruning of residual connections in the original form?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to convert residual network architectures to equivalent \"plain\" networks after training.  This is accomplished by augmenting convolutional layers with Dirac initialized filters (folding extraction of the residual signal into the conv layer), and modifying the subsequent Batch Normalization and ReLU layers (converting ReLU to PReLU) accordingly.  Motivations for doing this are to increase inference throughput, and allow additional pruning of the resulting plain network.\n", "main_review": "As the proposed conversion from residual network to plain network exactly preserves function, any throughput benefits are entirely determined by which form of operations is more efficient for the underlying combination of software library and compute hardware running the network.  Absent knowledge of such interactions, there does not appear to be any intrinsic reason for preferring the proposed form for rewriting residual layers.  In fact, the proposed approach seems to add computational work (more convolutional filters, Dirac initialized) for the sole purpose of taking advantage of fast implementations of convolutional layers.  Would a better approach simply be to write a custom fused low-level kernel for executing an entire residual block?  What form is actually better if one considers the optimal implementation possible for the underlying hardware (e.g., CPU, GPU, or TPU)?\n\nA similar question arises for the pruning approach.  If the functional forms of the residual and plain networks are equivalent, then any pruning operations on one form should have an equivalent expression in the other.  Why is it necessary to first convert to a plain network, instead of simply considering pruning of residual connections in the original form?\n", "summary_of_the_review": "I am not convinced that the core idea of the paper -- converting residual networks to an equivalent plain network -- addresses any fundamental issue.  The argument for speed of one vs the other is empirical, and may merely depend on what is optimized in the underlying software libraries.  To make a case here, the paper should provide analysis in terms of achievable parallel efficiency by an optimal implementation.  Similarly, the argument made for pruning RMNet seems to be one of convenience rather than fundamental difference -- is there not an equivalent (though perhaps not off-the-shelf) approach in terms of of pruning components of the original network?\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636254847856}, {"id": "qxRHdj4NCj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper685/Reviewer_Dt6G"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work focuses on removing residual connection in network via  reserving and merging (RM) operations on the ResBlock. The author has tested the approach on classification tasks on several networks with skip connection, e.g., resnet, mobilenet v2, etc. The experiments demonstrated the good performance on the listed benchmarks. ", "review_text": "The strength:\n1. The approach has showed constant classification performance improvements on the benchmarks. \n2. The idea is straightforward and easy to follow. \n\nThe weakness:\n1. More experiments are needed to validate the performance, like the performance on detection, segmentation benchmarks. \n2. Could the author list more inference comparison results across different hardware platforms, e.g,. cpu, gpu, popular embedded devices ? \n3. Beside the RM steps, there are also other additional operations like pruning operations, it would be good to see the improvements from different operations.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work focuses on removing residual connection in network via  reserving and merging (RM) operations on the ResBlock. The author has tested the approach on classification tasks on several networks with skip connection, e.g., resnet, mobilenet v2, etc. The experiments demonstrated the good performance on the listed benchmarks. ", "main_review": "The strength:\n1. The approach has showed constant classification performance improvements on the benchmarks. \n2. The idea is straightforward and easy to follow. \n\nThe weakness:\n1. More experiments are needed to validate the performance, like the performance on detection, segmentation benchmarks. \n2. Could the author list more inference comparison results across different hardware platforms, e.g,. cpu, gpu, popular embedded devices ? \n3. Beside the RM steps, there are also other additional operations like pruning operations, it would be good to see the improvements from different operations.\n\n", "summary_of_the_review": "In general, the paper has showed the strength in removing the residual connections on the classification tasks, it would be much more convincing to have more experiments on other popular tasks.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635916544515}, {"id": "6P_marhQKpv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper685/Reviewer_ybcv"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Update: I have read the rebuttal, see below.\n-----------\nWhile residual connections are a key component in today's deep learning architectures, they can be problematic in some settings, e.g., in pruning. This paper presents an improved method (RMNet) for removing residual connections from a ResNet - type neural network after training. It improves over related work and in contrast to those, it works on some typical ResNet variants. The paper also discusses fine-tuning, pruning and more efficient architectures.", "review_text": "Main Review (+ = positive comment, o = neutral comment, - = negative comment)\n------------------\n\n\\+ The strength of the paper is the relatively easy to implement method that can be applied to many ResNet variants in order to remove the residual connections. The results and discussion on finetuning, pruning etc. are also quite useful. \n\n\\+ The paper has good empirical part that covers quite well comparison to related work, results on fine-tuning and pruning.\n\no While the paper is somewhat incremental in nature, it generalizes the previously proposed approaches to removing residual connections to other ResNet architecture variants, so it has importance in that sense.\n\no While the paper is relative practical in nature, it is fine, since as far as I understand the resulting RMNet is mathematically equivalent to the corresponding ResNets, it inherits all analysis from ResNets.\n\n\\- Some details are missing in the paper. For example in section 3, e.g., Fig. 1 and the text describe the setup with the traditional ResNet with only one relu inside the block, while MobileNetV2 has a different architecture (resembling more preactivation resnet), where the residual branch has no nonlinearities. It is mentioned that PReLU is used in the case of MobileNetV2, but it not clear weather the parameters of PReLU fixed with some pattern (identity for new and relu for others?) or learned using SGD. The paper only mentions that the PReLU for the additional channels uses weights “set to one”. \n\n\\- Could the authors discuss whether it is possible to use fixed filters in part of the network and train the final RmNet from scratch? If I understand correctly, RmNet is mathematically equivalent to the corresponding ResNet and this should suggest that this is possible. This would remove the need for multi-step setup, where first there is a resnet trained, then the RM operation is applied, then the network is potentially fine-tuned and pruned. Since this paper is relatively practical work, it would be good to describe how to do this in typical deep learning framework as well.\n\n\\- Applicability to typical preactivation PreActResNet is missing (although there is MobileNetV2)\n\n\\- Applicability to some other popular ResNet variants, such as shufflenet is missing\n\nMinor problems\n------------------------------\n\n\\- Merging operation could be shown in pseudo-code in addition to the mathematical description, but this is a minor point, since there is also a figure about it.\n\n\\- Spelling mistakes, page 4. “(ie.e, in ResNet, every ResBlock has a following a ReLU layer, which keeps input values are all non-negative)”\n\n\\- Grammar error page 13 A.3. Finetune, the sentence “Thus We statistics the mean and variance by inputting ...” is wrong and sound probably by “Thus we computed the mean and variance by inputting…”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Update: I have read the rebuttal, see below.\n-----------\nWhile residual connections are a key component in today's deep learning architectures, they can be problematic in some settings, e.g., in pruning. This paper presents an improved method (RMNet) for removing residual connections from a ResNet - type neural network after training. It improves over related work and in contrast to those, it works on some typical ResNet variants. The paper also discusses fine-tuning, pruning and more efficient architectures.", "main_review": "Main Review (+ = positive comment, o = neutral comment, - = negative comment)\n------------------\n\n\\+ The strength of the paper is the relatively easy to implement method that can be applied to many ResNet variants in order to remove the residual connections. The results and discussion on finetuning, pruning etc. are also quite useful. \n\n\\+ The paper has good empirical part that covers quite well comparison to related work, results on fine-tuning and pruning.\n\no While the paper is somewhat incremental in nature, it generalizes the previously proposed approaches to removing residual connections to other ResNet architecture variants, so it has importance in that sense.\n\no While the paper is relative practical in nature, it is fine, since as far as I understand the resulting RMNet is mathematically equivalent to the corresponding ResNets, it inherits all analysis from ResNets.\n\n\\- Some details are missing in the paper. For example in section 3, e.g., Fig. 1 and the text describe the setup with the traditional ResNet with only one relu inside the block, while MobileNetV2 has a different architecture (resembling more preactivation resnet), where the residual branch has no nonlinearities. It is mentioned that PReLU is used in the case of MobileNetV2, but it not clear weather the parameters of PReLU fixed with some pattern (identity for new and relu for others?) or learned using SGD. The paper only mentions that the PReLU for the additional channels uses weights “set to one”. \n\n\\- Could the authors discuss whether it is possible to use fixed filters in part of the network and train the final RmNet from scratch? If I understand correctly, RmNet is mathematically equivalent to the corresponding ResNet and this should suggest that this is possible. This would remove the need for multi-step setup, where first there is a resnet trained, then the RM operation is applied, then the network is potentially fine-tuned and pruned. Since this paper is relatively practical work, it would be good to describe how to do this in typical deep learning framework as well.\n\n\\- Applicability to typical preactivation PreActResNet is missing (although there is MobileNetV2)\n\n\\- Applicability to some other popular ResNet variants, such as shufflenet is missing\n\nMinor problems\n------------------------------\n\n\\- Merging operation could be shown in pseudo-code in addition to the mathematical description, but this is a minor point, since there is also a figure about it.\n\n\\- Spelling mistakes, page 4. “(ie.e, in ResNet, every ResBlock has a following a ReLU layer, which keeps input values are all non-negative)”\n\n\\- Grammar error page 13 A.3. Finetune, the sentence “Thus We statistics the mean and variance by inputting ...” is wrong and sound probably by “Thus we computed the mean and variance by inputting…”\n", "summary_of_the_review": "\\+ Generalizes previous methods to other typical ResNet types\n\n\\+ Extensive Empirical evaluation\n\no/- Somewhat incremental, quite empirical\n\n\\- Some details missing", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethical concerns.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635854340843}, {"id": "bl100tl2XT4", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper685/Reviewer_oEx8"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper focuses on an important direction of removing residual connections from ResNet-based architectures and proposes RM operation to achieve this goal. RM operation consists of two steps: reserving and merging. The reserving operation allows input feature map passing through the Conv, BN,ReLU layers without changing their values and the merging operation adds the output feature map and the reserved input feature map together with the last Conv layer. Unlike re-parameterization method in RepVGG, RM Operation can remove residual connection across non-linear layers, resulting in equivalent transformation from ResNet to plain models. Authors also design a series of plain models with RM Operation named RMNet, which outperforms previous plain models such as DiracNet, ResDistill and RepVGG.", "review_text": "Strength: \n\n1. The motivation of this paper is clear. Plain model can reduce off-chip memory access and avoid point operations like residual plus which is helpful for model deployment on specific platforms. This paper is also well-written and easy to follow.\n\n2. The method of re-parameterization is very useful in model design and network pruning. However, the re-parameterization has limitations such as the difficulty of removing residual connections across non-linear layers. This paper analyzes the limitation of re-parameterization method to remove residual connections such as RepVGG from both forward paths and backward paths. \nTo overcome the limitation, authors propose RM operation to equivalently remove residual connections from ResNet-based architectures which enables the plain models to keep higher performance for deeper networks. I think this ideology is very inspiring and may have an interesting combination with other research areas in the future.\n\n3. I especially like the part of network pruning. It is known that ResNet is harder to be pruned compared to plain model because of the existence of residual connections. This paper proposes a alternative way to prune ResNet. First, use RM operation to convert ResNet to RMNet, then perform pruning on RMNet. The result shows the speed is much faster than vanilla pruning on ResNet architecture. \n\n4. The part of the experiment for transforming MobileNetV2 is interesting. By further utilizing fusing operation, RM can transform MobileNetV2 into MobileNetV1 and the performance of MobileNetV1 keeps the same as MobileNetV2 which suggests that RM is applicable for light-weight models.\n\nWeakness:\n\nRM operation adds extra parameters and FLOPs for achieving equivalent transformation. Thus the benefit of the speed may be subject to the specific platforms, or additional operation like fusing two 1x1 Conv when converting MobileNetV2 into MobileNetV1.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper focuses on an important direction of removing residual connections from ResNet-based architectures and proposes RM operation to achieve this goal. RM operation consists of two steps: reserving and merging. The reserving operation allows input feature map passing through the Conv, BN,ReLU layers without changing their values and the merging operation adds the output feature map and the reserved input feature map together with the last Conv layer. Unlike re-parameterization method in RepVGG, RM Operation can remove residual connection across non-linear layers, resulting in equivalent transformation from ResNet to plain models. Authors also design a series of plain models with RM Operation named RMNet, which outperforms previous plain models such as DiracNet, ResDistill and RepVGG.", "main_review": "Strength: \n\n1. The motivation of this paper is clear. Plain model can reduce off-chip memory access and avoid point operations like residual plus which is helpful for model deployment on specific platforms. This paper is also well-written and easy to follow.\n\n2. The method of re-parameterization is very useful in model design and network pruning. However, the re-parameterization has limitations such as the difficulty of removing residual connections across non-linear layers. This paper analyzes the limitation of re-parameterization method to remove residual connections such as RepVGG from both forward paths and backward paths. \nTo overcome the limitation, authors propose RM operation to equivalently remove residual connections from ResNet-based architectures which enables the plain models to keep higher performance for deeper networks. I think this ideology is very inspiring and may have an interesting combination with other research areas in the future.\n\n3. I especially like the part of network pruning. It is known that ResNet is harder to be pruned compared to plain model because of the existence of residual connections. This paper proposes a alternative way to prune ResNet. First, use RM operation to convert ResNet to RMNet, then perform pruning on RMNet. The result shows the speed is much faster than vanilla pruning on ResNet architecture. \n\n4. The part of the experiment for transforming MobileNetV2 is interesting. By further utilizing fusing operation, RM can transform MobileNetV2 into MobileNetV1 and the performance of MobileNetV1 keeps the same as MobileNetV2 which suggests that RM is applicable for light-weight models.\n\nWeakness:\n\nRM operation adds extra parameters and FLOPs for achieving equivalent transformation. Thus the benefit of the speed may be subject to the specific platforms, or additional operation like fusing two 1x1 Conv when converting MobileNetV2 into MobileNetV1.", "summary_of_the_review": "The paper puts forward a novel method: RM Operation, which can equivalently remove residual connection across non-linear layer in ResNet-based architecture and shows great power in network pruning. I think RM operation is novel and can inspire future works on model design.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634696347450}], "openreview_url": "https://openreview.net/forum?id=MPoQtFC588n", "arxiv_id": "2111.00687", "paper_pdf": "papers/MPoQtFC588n.pdf", "paper_pdf_sha256": "c725adf4fc788b93ed506bd79bf8be38386b750621e8f11d9d2744623a801738", "paper_pdf_bytes": 684926, "paper_pdf_source": "openreview", "code_url": "https://github.com/fxmeng/RMNet", "code_repository": "fxmeng/RMNet", "code_commit": "4754b8e8a0c8b0f6e4e0c8c77f6cde47712c8feb", "code_archive": "repos/MPoQtFC588n.zip", "code_archive_sha256": "61b4b441fb55c495307169fe01965a6d6baa11f1f4ce248bfe47a2386e01af45", "code_archive_bytes": 436106, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 444, "github_languages": {"Python": 144239}, "github_archived": false, "github_pushed_at": "2023-06-17T02:23:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rmnet-equivalently-removing-residual-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SVsLxTfHa1", "year": 2021, "status": "rejected", "title": "Towards Multi-Sense Cross-Lingual Alignment of Contextual Embeddings", "authors": ["Linlin Liu", "Thien Hai Nguyen", "Shafiq Joty", "Lidong Bing", "Luo Si"], "authorids": ["~Linlin_Liu2", "~Thien_Hai_Nguyen1", "~Shafiq_Joty1", "~Lidong_Bing2", "~Luo_Si2"], "authors_source": "OpenReview API", "abstract": "Cross-lingual word embeddings (CLWE) have been proven useful in many cross-lingual tasks. However, most existing approaches to learn CLWE including the ones with contextual embeddings are sense agnostic. In this work, we propose a novel framework to align contextual embeddings at the sense level by leveraging cross-lingual signal from bilingual dictionaries only. We operationalize our framework by first proposing a novel sense-aware cross entropy loss to model word senses explicitly. The monolingual ELMo and BERT models pretrained with our sense-aware cross entropy loss demonstrate significant performance improvement for word sense disambiguation tasks. We then propose a sense alignment objective on top of the sense-aware cross entropy loss for cross-lingual model pretraining, and pretrain cross-lingual models for several language pairs (English to German/Spanish/Japanese/Chinese). Compared with the best baseline results, our cross-lingual models achieve 0.52%, 2.09% and 1.29% average performance improvements on zero-shot cross-lingual NER, sentiment classification and XNLI tasks, respectively. We will release our code.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0YS5yBhR9LJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper176/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to introduce multiple senses into pre-trained models. The proposed method selects senses dynamically while pretraining the model and applies a sense-aware cross-entropy loss for pretraining. This paper further proposes to jointly pre-train a sense-aware cross-lingual model with sense-level translation. The proposed model yields better performance than the baseline models under both monolingual and cross-lingual setting. \n\nStrength:\n\n1) The pipeline is designed very well and covers many aspects: taking care of the pruning while training; introduce projections to reduce parameters; from monolingual to cross-lingual setting. \n2) The paper is well written and easy to follow.\n3) Evaluation on three downstream tasks show significant improvements over baseline models, and for some case  (NER de), it performs even better than larger cross-lingual models.\n\nWeakness:\n\n1) Should cite more multi-sense papers (at least more papers before 2018). \n\nEspecially:\n\nhttps://www.aclweb.org/anthology/P12-1092.pdf \nhttps://www.aclweb.org/anthology/D14-1113/\n(both uses clustering methods to select senses)\n\nhttps://www.aclweb.org/anthology/N16-1160.pdf\nhttps://www.aclweb.org/anthology/D17-1034/\nhttps://www.aclweb.org/anthology/2020.lrec-1.214.pdf\n(all use similar softmax form to predict senses as the prediction task in your sense-aware cross-entropy loss)\n\n2) The proposed model outperforms the baseline models significantly.  However, the baseline models are pretty out-dated. And the scale is quite small. It is true that your model performs better on NER for DE than larger cross-lingual models. However, in most cases, it performs much worse than large-scale pre-trained cross-lingual models (https://arxiv.org/pdf/1901.07291.pdf). Have you tried to apply your methods to large-scale models? Is it possible that large-scale models have already captured the sense information within the context and the context provides enough information for disambiguations? Therefore introducing senses won't bring more capacity to the model?\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good model design and good writing, however a bit lack motivation", "review": "This paper proposes to introduce multiple senses into pre-trained models. The proposed method selects senses dynamically while pretraining the model and applies a sense-aware cross-entropy loss for pretraining. This paper further proposes to jointly pre-train a sense-aware cross-lingual model with sense-level translation. The proposed model yields better performance than the baseline models under both monolingual and cross-lingual setting. \n\nStrength:\n\n1) The pipeline is designed very well and covers many aspects: taking care of the pruning while training; introduce projections to reduce parameters; from monolingual to cross-lingual setting. \n2) The paper is well written and easy to follow.\n3) Evaluation on three downstream tasks show significant improvements over baseline models, and for some case  (NER de), it performs even better than larger cross-lingual models.\n\nWeakness:\n\n1) Should cite more multi-sense papers (at least more papers before 2018). \n\nEspecially:\n\nhttps://www.aclweb.org/anthology/P12-1092.pdf \nhttps://www.aclweb.org/anthology/D14-1113/\n(both uses clustering methods to select senses)\n\nhttps://www.aclweb.org/anthology/N16-1160.pdf\nhttps://www.aclweb.org/anthology/D17-1034/\nhttps://www.aclweb.org/anthology/2020.lrec-1.214.pdf\n(all use similar softmax form to predict senses as the prediction task in your sense-aware cross-entropy loss)\n\n2) The proposed model outperforms the baseline models significantly.  However, the baseline models are pretty out-dated. And the scale is quite small. It is true that your model performs better on NER for DE than larger cross-lingual models. However, in most cases, it performs much worse than large-scale pre-trained cross-lingual models (https://arxiv.org/pdf/1901.07291.pdf). Have you tried to apply your methods to large-scale models? Is it possible that large-scale models have already captured the sense information within the context and the context provides enough information for disambiguations? Therefore introducing senses won't bring more capacity to the model?\n\n\n\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603915074183}, {"id": "NLlXmEmVVoP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper176/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nOverview\n===================\n\nThis is a really interesting paper, and it solves a big problem with cross-lingual representations: ignoring senses.  \n\nPros: Interesting model, relatively unexplored problem, the quantitative evaluations in the paper are reasonable\nCons: Lack of comparisons with unsupervised sense algorithms, no qualitative evaluation of senses\n\nStructure of algorithm\n===================\nThe structure of the approach reminds me of the Yarowsky algorithm, has an EM-like flavor for building up the sense distributions.  It would be useful and interesting to discuss why it has this structure.  It would be useful to discuss a little more about why this was chosen rather than an end-to-end model that does the clustering and sense selection together.\n\nIs this unsupervised?\n===================\n\nThe paper argues that it is unsupervised, \"Unlike these methods, our language models learn word senses in a fully self-supervised way\".  However, I don't think this tells the whole story.  It needs a *multilingual dictionary* to know how many senses a word has and if it has a translation.\n\nWhile MUSE (Conneau) provides unsupervised translations, the paper also adds  human-curated dictionary JMdict (including English, German, French, Russian and Dutch glosses), but the paper does not have a clear examination of the role of this additional supervision.\n\nGiven this level of supervision, it would be useful to have an explicit comparison with other supervised methods (e.g., have this model use WordNet or replace the other methods' dictionary with JMdict).\n\nWhat are comparable models?\n===================\n\nThere are, however, truly unsupervised sense induction models.  For instance, MUSE (Lee), GASI (Guo), and MSSG (Neelakantan).  Unfortuantely, this paper does not mention or compare with these models (they are, however, monolingual).\n\nWhat are the relevant tasks?\n===================\n\nWhile WSD is a relevant task, it would be useful to have sense in context (Huang)  and/or interpretability (Guo) as tasks.  These tasks better capture whether unsupervised senses can determine a word meaning (without supervised data from SemCor) or make sense to an end user.\n\nEither a formal or informal examination of what the sense clusters look like would help understand whether the algorithm is doing what it claims and to see if the output would be useful to humans (settings aside downstream tasks).\n\nWhile the NER and XNLI tasks are reasonable, I'd rather see sense-specific evaluations.\n\nRelated Work not Cited\n===================\n\nFenfei Guo, Jordan Boyd-Graber, Mohit Iyyer, and Leah Findlater. Which Evaluations Uncover Sense Representations that Actually Make Sense?. Linguistic Resources and Evaluation Conference, 2020.\n\nHuang, E. H., Socher, R., Manning, C. D., and Ng, A. Y.  (2012). Improving word representations via global context and multiple word prototypes. In Proceedings of the Association for Computational Linguistics.\n\nLee, G.-H. and Chen, Y.-N. (2017). MUSE: Modularizing unsupervised sense embeddings. In Proceedings of Empirical Methods in Natural Language Processing.\n\nNeelakantan, A., Shankar, J., Passos, A., and McCallum, A.  (2014). Efficient non-parametric estimation of multiple embeddings per word in vector space. In Proceedings of Empirical Methods in Natural Language Processing.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Missing comparisons / distinctions with unsupervised sense models ", "review": "\nOverview\n===================\n\nThis is a really interesting paper, and it solves a big problem with cross-lingual representations: ignoring senses.  \n\nPros: Interesting model, relatively unexplored problem, the quantitative evaluations in the paper are reasonable\nCons: Lack of comparisons with unsupervised sense algorithms, no qualitative evaluation of senses\n\nStructure of algorithm\n===================\nThe structure of the approach reminds me of the Yarowsky algorithm, has an EM-like flavor for building up the sense distributions.  It would be useful and interesting to discuss why it has this structure.  It would be useful to discuss a little more about why this was chosen rather than an end-to-end model that does the clustering and sense selection together.\n\nIs this unsupervised?\n===================\n\nThe paper argues that it is unsupervised, \"Unlike these methods, our language models learn word senses in a fully self-supervised way\".  However, I don't think this tells the whole story.  It needs a *multilingual dictionary* to know how many senses a word has and if it has a translation.\n\nWhile MUSE (Conneau) provides unsupervised translations, the paper also adds  human-curated dictionary JMdict (including English, German, French, Russian and Dutch glosses), but the paper does not have a clear examination of the role of this additional supervision.\n\nGiven this level of supervision, it would be useful to have an explicit comparison with other supervised methods (e.g., have this model use WordNet or replace the other methods' dictionary with JMdict).\n\nWhat are comparable models?\n===================\n\nThere are, however, truly unsupervised sense induction models.  For instance, MUSE (Lee), GASI (Guo), and MSSG (Neelakantan).  Unfortuantely, this paper does not mention or compare with these models (they are, however, monolingual).\n\nWhat are the relevant tasks?\n===================\n\nWhile WSD is a relevant task, it would be useful to have sense in context (Huang)  and/or interpretability (Guo) as tasks.  These tasks better capture whether unsupervised senses can determine a word meaning (without supervised data from SemCor) or make sense to an end user.\n\nEither a formal or informal examination of what the sense clusters look like would help understand whether the algorithm is doing what it claims and to see if the output would be useful to humans (settings aside downstream tasks).\n\nWhile the NER and XNLI tasks are reasonable, I'd rather see sense-specific evaluations.\n\nRelated Work not Cited\n===================\n\nFenfei Guo, Jordan Boyd-Graber, Mohit Iyyer, and Leah Findlater. Which Evaluations Uncover Sense Representations that Actually Make Sense?. Linguistic Resources and Evaluation Conference, 2020.\n\nHuang, E. H., Socher, R., Manning, C. D., and Ng, A. Y.  (2012). Improving word representations via global context and multiple word prototypes. In Proceedings of the Association for Computational Linguistics.\n\nLee, G.-H. and Chen, Y.-N. (2017). MUSE: Modularizing unsupervised sense embeddings. In Proceedings of Empirical Methods in Natural Language Processing.\n\nNeelakantan, A., Shankar, J., Passos, A., and McCallum, A.  (2014). Efficient non-parametric estimation of multiple embeddings per word in vector space. In Proceedings of Empirical Methods in Natural Language Processing.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603886419500}, {"id": "aODO-fHSdcX", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper176/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary:**\nThis paper proposes the alignment of cross-lingual contextual embeddings not just at the word level, but at the sense level. It does this by relying purely on unaligned, unlabeled monolingual corpora used for pre-training, along with bilingual lexica. It does this by adapting the LM objective to be a sense-aware cross entropy loss, in which the sense is obtained by the use of a streaming k-means clustering algorithm combined with dimensionality reduction. If a bilingual lexicon is available, a sense-level translation objective can be added to encourage the model to predict the same sense in the other language (thereby encouraging identical senses of a word in the two languages to be closer together). \n\n**Positives:**\n* This is the first work (as far as I am aware) to perform sense-level alignment of contextual embeddings with only unlabeled, unaligned monolingual corpora and bilingual dictionaries.\n* The proposed method works for both MLM and the next token prediction objectives. It should thus be applicable to most contextual embedding methods.\n* The proposed method is extremely intuitive, but non-trivial. In particular, this work clearly presents its insights and well-justified tricks that seem to be crucial to the working of the proposed approach, for example: how simply selecting the vector $\\\\mathbf{w}\\_{k',s}$ with the maximum dot product $\\\\mathbf{h}^T\\_{k^*,s}\\\\mathbf{w}\\_{k',s}$ does not work; how/why dimension reduction is crucial; the need to use $\\\\mathbf{h}\\_{k,L}$ for sense selection even though $\\\\mathbf{h}\\_{k-1,L}$ is used for next token prediction.\n\n**Concerns:**\n* In the case of the results presented for WSD, the ELMo baseline seems to be substantially worse than that reported in literature [1, 2] (ref: Table 2).\n* While the paper talks about the applicability of the proposed method to both the next token prediction (in ELMo) and the MLM (in BERT) styles of language models, the experiments section seems very heavily focused on ELMo, not demonstrating the applicability of the method on BERT at all (aside from Table 2). Likewise, all the visualization in Appendix C focus solely on ELMo.\n\n**Suggestions**\n* In the section \"Limitations of original training objectives\", the authors state \"We also observed the same with BERT embeddings.\". This observation is central to the paper's argument for the need of the proposed approach in BERT, and elaborating on this with more analysis would have been nice.\n* While the motivation behind the various tricks adopted were very well explained (refer: Positives), ablations/experimental numbers showing the failure of the proposed approach without these tricks would have helped strengthen this quantitatively.\n\n**Questions:**\n* Does the proposed method have any impact on the monolingual task (for example, for en on CoNLL?)\n* The linear projection for cross-lingual alignment has been shown to not work on zero-shot sentiment classification. Do you have insight as to why this might be, why it adds noise to the embedding features?\n* Can the proposed method directly be used to improve Bilingual Lexicon Induction? Perhaps by using monolingual data to train models with the sense-aware cross entropy loss and dynamic pruning to obtain the different sense embeddings, and then aligning these sense embeddings (Eg: with MUSE). Or if a bilingual lexicon is available, performing joint training followed by a supervised BLI approach like RCSLS [3]. Would this help learn a many-many mapping for BLI (where normally 1:many/many:many are difficult to achieve precisely because senses aren't taken into account)?\n* Footnote 2 says that the dictionary can be learned in an unsupervised way (eg: using MUSE). Since the method here does not rely on a learned transformation, I'm assuming it was meant that we could use MUSE to generate the bilingual lexicon. How would this work with the proposed method, since unsupervised dictionaries generated with MUSE are usually 1:1/many:1 and tend to ignore multiple senses?\n\n**Minor details:**\n* Consider re-wording lines 7-10 in Algorithm 1 in terms of an argmax for conciseness, correctness and clarity\n\n[1] Peters, Matthew E., et al. \"Deep contextualized word representations.\" arXiv preprint arXiv:1802.05365 (2018).\n[2] Hadiwinoto, Christian, Hwee Tou Ng, and Wee Chung Gan. \"Improved word sense disambiguation using pre-trained contextualized word representations.\" arXiv preprint arXiv:1910.00194 (2019).\n[3] Joulin, Armand, et al. \"Loss in translation: Learning bilingual word mapping with a retrieval criterion.\" arXiv preprint arXiv:1804.07745 (2018).\n\n\n======================================================================\n\n**Update:**\n\nI would like to thank the authors for their response. The lack of adverse impact of the proposed approach on monolingual tasks and the described ablations certainly help strengthen things on the experimental side. \n\nHowever, my fundamental concerns still remain:\n* I'm unsure why the ELMo baseline is so much worse than that reported in literature.\n* With respect to the difficulty of training BERT, I can certainly empathize with the authors about the limited resources available in an academic setting. However, given that BERT was a key area of focus of the paper's methods section, showing the experimental results for BERT, even if BERT-tiny; note that BERT-tiny has just 2 layers and 128 hidden units, as opposed to 4 layers and 512 hidden units which this work uses (based on Section 4.1 and Table 7-- which corresponds to BERT-small, refer [here](https://github.com/google-research/bert#bert)), which should help reduce computational burden by quite a bit.\n\nOn account of this, I maintain my original rating of 5.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good approach for sense-level alignment of contextual embeddings, but the experimental validation could be stronger", "review": "**Summary:**\nThis paper proposes the alignment of cross-lingual contextual embeddings not just at the word level, but at the sense level. It does this by relying purely on unaligned, unlabeled monolingual corpora used for pre-training, along with bilingual lexica. It does this by adapting the LM objective to be a sense-aware cross entropy loss, in which the sense is obtained by the use of a streaming k-means clustering algorithm combined with dimensionality reduction. If a bilingual lexicon is available, a sense-level translation objective can be added to encourage the model to predict the same sense in the other language (thereby encouraging identical senses of a word in the two languages to be closer together). \n\n**Positives:**\n* This is the first work (as far as I am aware) to perform sense-level alignment of contextual embeddings with only unlabeled, unaligned monolingual corpora and bilingual dictionaries.\n* The proposed method works for both MLM and the next token prediction objectives. It should thus be applicable to most contextual embedding methods.\n* The proposed method is extremely intuitive, but non-trivial. In particular, this work clearly presents its insights and well-justified tricks that seem to be crucial to the working of the proposed approach, for example: how simply selecting the vector $\\\\mathbf{w}\\_{k',s}$ with the maximum dot product $\\\\mathbf{h}^T\\_{k^*,s}\\\\mathbf{w}\\_{k',s}$ does not work; how/why dimension reduction is crucial; the need to use $\\\\mathbf{h}\\_{k,L}$ for sense selection even though $\\\\mathbf{h}\\_{k-1,L}$ is used for next token prediction.\n\n**Concerns:**\n* In the case of the results presented for WSD, the ELMo baseline seems to be substantially worse than that reported in literature [1, 2] (ref: Table 2).\n* While the paper talks about the applicability of the proposed method to both the next token prediction (in ELMo) and the MLM (in BERT) styles of language models, the experiments section seems very heavily focused on ELMo, not demonstrating the applicability of the method on BERT at all (aside from Table 2). Likewise, all the visualization in Appendix C focus solely on ELMo.\n\n**Suggestions**\n* In the section \"Limitations of original training objectives\", the authors state \"We also observed the same with BERT embeddings.\". This observation is central to the paper's argument for the need of the proposed approach in BERT, and elaborating on this with more analysis would have been nice.\n* While the motivation behind the various tricks adopted were very well explained (refer: Positives), ablations/experimental numbers showing the failure of the proposed approach without these tricks would have helped strengthen this quantitatively.\n\n**Questions:**\n* Does the proposed method have any impact on the monolingual task (for example, for en on CoNLL?)\n* The linear projection for cross-lingual alignment has been shown to not work on zero-shot sentiment classification. Do you have insight as to why this might be, why it adds noise to the embedding features?\n* Can the proposed method directly be used to improve Bilingual Lexicon Induction? Perhaps by using monolingual data to train models with the sense-aware cross entropy loss and dynamic pruning to obtain the different sense embeddings, and then aligning these sense embeddings (Eg: with MUSE). Or if a bilingual lexicon is available, performing joint training followed by a supervised BLI approach like RCSLS [3]. Would this help learn a many-many mapping for BLI (where normally 1:many/many:many are difficult to achieve precisely because senses aren't taken into account)?\n* Footnote 2 says that the dictionary can be learned in an unsupervised way (eg: using MUSE). Since the method here does not rely on a learned transformation, I'm assuming it was meant that we could use MUSE to generate the bilingual lexicon. How would this work with the proposed method, since unsupervised dictionaries generated with MUSE are usually 1:1/many:1 and tend to ignore multiple senses?\n\n**Minor details:**\n* Consider re-wording lines 7-10 in Algorithm 1 in terms of an argmax for conciseness, correctness and clarity\n\n[1] Peters, Matthew E., et al. \"Deep contextualized word representations.\" arXiv preprint arXiv:1802.05365 (2018).\n[2] Hadiwinoto, Christian, Hwee Tou Ng, and Wee Chung Gan. \"Improved word sense disambiguation using pre-trained contextualized word representations.\" arXiv preprint arXiv:1910.00194 (2019).\n[3] Joulin, Armand, et al. \"Loss in translation: Learning bilingual word mapping with a retrieval criterion.\" arXiv preprint arXiv:1804.07745 (2018).\n\n\n======================================================================\n\n**Update:**\n\nI would like to thank the authors for their response. The lack of adverse impact of the proposed approach on monolingual tasks and the described ablations certainly help strengthen things on the experimental side. \n\nHowever, my fundamental concerns still remain:\n* I'm unsure why the ELMo baseline is so much worse than that reported in literature.\n* With respect to the difficulty of training BERT, I can certainly empathize with the authors about the limited resources available in an academic setting. However, given that BERT was a key area of focus of the paper's methods section, showing the experimental results for BERT, even if BERT-tiny; note that BERT-tiny has just 2 layers and 128 hidden units, as opposed to 4 layers and 512 hidden units which this work uses (based on Section 4.1 and Table 7-- which corresponds to BERT-small, refer [here](https://github.com/google-research/bert#bert)), which should help reduce computational burden by quite a bit.\n\nOn account of this, I maintain my original rating of 5.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603877491597}, {"id": "59RfVDWzq9V", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper176/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n## Summary\nResearch Problem: ELMo or BERT does not model word-sense explicitly, and it introduces challenges in representation learning, especially in the multilingual setting.\n\nThis paper proposes a novel approach to model word-sense explicitly for (masked) language models (refer as LM for simplicity purpose) in an unsupervised fashion. It maintains a set of sense vectors for each word and the sense vectors are updated with online clustering approach. The correct sense is selected with the contextual representation and the model is trained to predict the correct sense. Finally, a sense-level translation loss is proposed for bilingual models. This paper presents both English monolingual model and bilingual model with 4 language pairs. It outperforms models without proposed component on word sense disambiguation and cross-lingual transfer on three tasks.\n\n\n## Pros\n1. The proposed sense-aware LM learns to model word-sense without any supervision in monolingual LM.\n2. The proposed sense-aware LM and additional sense-level translation loss leads to improvement over regular bilingual models. In some cases, an extra projection step from previous work leads to further gain.\n\n## Cons\n1. This paper assumes the LM should be word-level, a relatively limited assumption when scaling to bigger corpus or more languages. It’s not straightforward to apply this approach to subword-level LM. It limits this paper to pretrain from scratch instead of fine-tune the existing BERT, multilingual BERT or XLM-R. It is not a problem for showcasing the proposed method in a controlled setting, but it limits the full potential of this paper, pushing state-of-the-art on top of the current best model. Additionally, this paper does not discuss the complexity of the proposed method.\n2. While the experiment presents some evidence that the proposed method learns to model word-sense in an unsupervised fashion, it does not include any result comparison with prior work on incorporating word-sense into LM. As a result, it’s unclear how well the model learns word-sense compared to models with supervision.\n3. It is unclear how much the sense-level translation loss contributes. This paper claims “... Bi-SaELMo is significantly better than Joint-ELMo, which shows that our sense-level translation pretraining objective improves cross-lingual embedding alignment”. However, the result presented in this paper does not support this claim. To support this claim, this paper should present **Bi-SaELMo without the sense-level translation loss**. Additionally, Bi-SaELMo assumes a bilingual dictionary compared to Joint-ELMo, yet the difference is not clearly discussed in the paper, especially in the result section.\n\n## Reasons for score\nOverall, I am leaning toward rejecting. While I find the sense-aware LM quite novel, the experiment presented does not support claim of contribution of sense-level translation loss. While the current result is promising, I find the limitation of the model not fully discussed and lack of comparison against other sense-aware LM in the monolingual setting. Hopefully the authors can clarify and address my concern in the rebuttal period. \n\n## After revision\nThank you for answering my question!\n\n> Our approach can be directly used to fine-tune pre-trained language models if word level tokenization is used.\n\nIt should be acknowledged that no multilingual word-based BERT exists as far as I know, partly due to the challenge with large vocabulary space and generalization. The sub-word approach you mentioned is a good proposal, but without evidence, I cannot assess whether it works. As a result, it's still a limitation of *this paper*, and should be clearly discussed in the paper.\n\n> The complexity of the proposed method.\n\nI am referring to the complexity w.r.t vocabulary size. The softmax normalization contains `nV` items, where `n` is the number of sense clusters and `V` is the vocabulary size. With the current limitation of word-based models, the vocabulary is already larger than sub-word models, the extra `n` factor cannot be disregarded.\n\n> We train bilingual language models without sense-level translation loss (and without projection), denoted by Bi-SaELMo-NT, for ablation study.\n\nThank you for the ablation study! However, I cannot assess it as support for the claim as it's not an apple-to-apple comparison, and I cannot evaluate this paper based on future projection. \n\nAs a result, I have to maintain my rating based on the revision.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Idea but Lacking Certain Experiment to Support the Claim", "review": "\n## Summary\nResearch Problem: ELMo or BERT does not model word-sense explicitly, and it introduces challenges in representation learning, especially in the multilingual setting.\n\nThis paper proposes a novel approach to model word-sense explicitly for (masked) language models (refer as LM for simplicity purpose) in an unsupervised fashion. It maintains a set of sense vectors for each word and the sense vectors are updated with online clustering approach. The correct sense is selected with the contextual representation and the model is trained to predict the correct sense. Finally, a sense-level translation loss is proposed for bilingual models. This paper presents both English monolingual model and bilingual model with 4 language pairs. It outperforms models without proposed component on word sense disambiguation and cross-lingual transfer on three tasks.\n\n\n## Pros\n1. The proposed sense-aware LM learns to model word-sense without any supervision in monolingual LM.\n2. The proposed sense-aware LM and additional sense-level translation loss leads to improvement over regular bilingual models. In some cases, an extra projection step from previous work leads to further gain.\n\n## Cons\n1. This paper assumes the LM should be word-level, a relatively limited assumption when scaling to bigger corpus or more languages. It’s not straightforward to apply this approach to subword-level LM. It limits this paper to pretrain from scratch instead of fine-tune the existing BERT, multilingual BERT or XLM-R. It is not a problem for showcasing the proposed method in a controlled setting, but it limits the full potential of this paper, pushing state-of-the-art on top of the current best model. Additionally, this paper does not discuss the complexity of the proposed method.\n2. While the experiment presents some evidence that the proposed method learns to model word-sense in an unsupervised fashion, it does not include any result comparison with prior work on incorporating word-sense into LM. As a result, it’s unclear how well the model learns word-sense compared to models with supervision.\n3. It is unclear how much the sense-level translation loss contributes. This paper claims “... Bi-SaELMo is significantly better than Joint-ELMo, which shows that our sense-level translation pretraining objective improves cross-lingual embedding alignment”. However, the result presented in this paper does not support this claim. To support this claim, this paper should present **Bi-SaELMo without the sense-level translation loss**. Additionally, Bi-SaELMo assumes a bilingual dictionary compared to Joint-ELMo, yet the difference is not clearly discussed in the paper, especially in the result section.\n\n## Reasons for score\nOverall, I am leaning toward rejecting. While I find the sense-aware LM quite novel, the experiment presented does not support claim of contribution of sense-level translation loss. While the current result is promising, I find the limitation of the model not fully discussed and lack of comparison against other sense-aware LM in the monolingual setting. Hopefully the authors can clarify and address my concern in the rebuttal period. \n\n## After revision\nThank you for answering my question!\n\n> Our approach can be directly used to fine-tune pre-trained language models if word level tokenization is used.\n\nIt should be acknowledged that no multilingual word-based BERT exists as far as I know, partly due to the challenge with large vocabulary space and generalization. The sub-word approach you mentioned is a good proposal, but without evidence, I cannot assess whether it works. As a result, it's still a limitation of *this paper*, and should be clearly discussed in the paper.\n\n> The complexity of the proposed method.\n\nI am referring to the complexity w.r.t vocabulary size. The softmax normalization contains `nV` items, where `n` is the number of sense clusters and `V` is the vocabulary size. With the current limitation of word-based models, the vocabulary is already larger than sub-word models, the extra `n` factor cannot be disregarded.\n\n> We train bilingual language models without sense-level translation loss (and without projection), denoted by Bi-SaELMo-NT, for ablation study.\n\nThank you for the ablation study! However, I cannot assess it as support for the claim as it's not an apple-to-apple comparison, and I cannot evaluate this paper based on future projection. \n\nAs a result, I have to maintain my rating based on the revision.\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603877455233}], "openreview_url": "https://openreview.net/forum?id=SVsLxTfHa1", "arxiv_id": "2103.06459", "paper_pdf": "papers/SVsLxTfHa1.pdf", "paper_pdf_sha256": "0fa02abb9517834a7bd8b17220e2f60f46d31fdbf3e039975bf575ff48777cd0", "paper_pdf_bytes": 2103083, "paper_pdf_source": "openreview", "code_url": "https://github.com/ntunlp/multisense_embedding_alignment", "code_repository": "ntunlp/multisense_embedding_alignment", "code_commit": "6cf157ec4eefafcdb1d3403fa1cc0dd2b5169833", "code_archive": "repos/SVsLxTfHa1.zip", "code_archive_sha256": "709922b27a817d0ba19c6de936ce5532fe5761744242e402b469d9e9b537e22c", "code_archive_bytes": 620056, "code_file_count": 39, "code_extensions": {".py": 32, ".sh": 7}, "github_disk_usage_kb": 589, "github_languages": {"Python": 421997, "Shell": 3341}, "github_archived": false, "github_pushed_at": "2022-10-14T13:19:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-multi-sense-cross-lingual-alignment-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rkgFXR4KPr", "year": 2020, "status": "rejected", "title": "A Simple Recurrent Unit with Reduced Tensor Product Representations", "authors": ["Shuai Tang", "Paul Smolensky", "Virginia R. de Sa"], "authorids": ["shuaitang93@ucsd.edu", "paul.smolensky@gmail.com", "desa@ucsd.edu"], "authors_source": "OpenReview API", "abstract": "Widely used recurrent units, including Long-short Term Memory (LSTM) and Gated Recurrent Unit (GRU), perform well on natural language tasks, but their ability to learn structured representations is still questionable. Exploiting reduced Tensor Product Representations (TPRs) --- distributed representations of symbolic structure in which vector-embedded symbols are bound to vector-embedded structural positions --- we propose the TPRU, a simple recurrent unit that, at each time step, explicitly executes structural-role binding and unbinding operations to incorporate structural information into learning. The gradient analysis of our proposed TPRU is conducted to support our model design, and its performance on multiple datasets shows the effectiveness of it. Furthermore, observations on linguistically grounded study demonstrate the interpretability of our TPRU.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HklyTuFRFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1046/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a novel model of recurrent unit for RNNs which is inspired from tensor product representation (TPR) introduced by Smolensky et al. in 1990. The authors claim that this allows one to better incorporate structural information into learning and easier interpretability for the learned representations. The proposed approach is motivated by a theoretical analysis showing that using TPR in this context acts as a sort of pre-conditioner and stabilizes learning. Experiments on entailment tasks (given two statement, decide whether the first implies the second) are provided to validate the approach. \n\nI find the paper not easy to follow, with a non-negligible amount of typos in the notations and results. The advantage in terms of accuracy of the proposed approach seems marginal in the experiment, and the analysis of the interpretability of the learned representations could be improved: loosely speaking, particular examples of interpretability are given but sometimes without contexts or baselines to compare to (see the two last comments below). I know \"interpretability\" is a difficult property to assess but I think there may be more principled ways to showcase the approach.\n\nI think this paper is not yet ready for publication: the proposed model is interesting and relevant but its validity could be better assessed and the paper needs some thorough proof-reading. \n\n\n* Questions / Comments *\n\n- page 1: the authors write U^TR = I, but I believe this is only possible if the TPR dimension d is bigger than the number of roles N. Is this always the case? This should be clarified.\n- related to the previous point:  if U^TR = I then shouldn't U^Tb_{t-1} simply be f_{t-1} in Eq. 1?\n- before Eq.1, f should be from R^d\\times R^d' to R^d, not from R^d \\times R^d\n- In Eq. 1, b_{t-1} and x_t are not of the same dimension, so the cannot be multiplied by the same matrix U (this is why the matrices V_x and V_b are introduced later on).\n- there seems to be a problem with Eq. (7): db_t/d_{b_{t-1}} appears on both sides of the equality...\n- Modification 1: what is \\tilde{vb_t}? I don't remember seeing this notation introduced before.\n- Table 1: constants should not be included in big O notation! To compare constants, one should give the exact number of operations needed for inference.\n- POS tagging: Aren't there many other reasons that could lead to this correlation (beside the informal argument that \"TPR captures structured information\")? Maybe the authors should compare with something else, for example the PMI between values of hidden neurons in a learned RNN and POS tags. Out of context, the numbers in Table 5 are not informative.\n- Polysemy: Only a very specific cherry picked example is given here. A more principled or in depth analysis of this phenomenon is needed to make a stronger case.\n\n* Typos *\n\n- \" The number of parameter matrices *is* the same as that of...\"\n- page 4 \"stables\" -> \"stabilizes\" (but rephrasing the sentence altogether would be better).\n- page 5: BiDAF misses the capital letters (\"bidaf\").\n- \"dev set\" -> \"validation set\" or \"development set\".\n- page 8: \"provides research*ers with* an intuitive...\"?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a novel model of recurrent unit for RNNs which is inspired from tensor product representation (TPR) introduced by Smolensky et al. in 1990. The authors claim that this allows one to better incorporate structural information into learning and easier interpretability for the learned representations. The proposed approach is motivated by a theoretical analysis showing that using TPR in this context acts as a sort of pre-conditioner and stabilizes learning. Experiments on entailment tasks (given two statement, decide whether the first implies the second) are provided to validate the approach. \n\nI find the paper not easy to follow, with a non-negligible amount of typos in the notations and results. The advantage in terms of accuracy of the proposed approach seems marginal in the experiment, and the analysis of the interpretability of the learned representations could be improved: loosely speaking, particular examples of interpretability are given but sometimes without contexts or baselines to compare to (see the two last comments below). I know \"interpretability\" is a difficult property to assess but I think there may be more principled ways to showcase the approach.\n\nI think this paper is not yet ready for publication: the proposed model is interesting and relevant but its validity could be better assessed and the paper needs some thorough proof-reading. \n\n\n* Questions / Comments *\n\n- page 1: the authors write U^TR = I, but I believe this is only possible if the TPR dimension d is bigger than the number of roles N. Is this always the case? This should be clarified.\n- related to the previous point:  if U^TR = I then shouldn't U^Tb_{t-1} simply be f_{t-1} in Eq. 1?\n- before Eq.1, f should be from R^d\\times R^d' to R^d, not from R^d \\times R^d\n- In Eq. 1, b_{t-1} and x_t are not of the same dimension, so the cannot be multiplied by the same matrix U (this is why the matrices V_x and V_b are introduced later on).\n- there seems to be a problem with Eq. (7): db_t/d_{b_{t-1}} appears on both sides of the equality...\n- Modification 1: what is \\tilde{vb_t}? I don't remember seeing this notation introduced before.\n- Table 1: constants should not be included in big O notation! To compare constants, one should give the exact number of operations needed for inference.\n- POS tagging: Aren't there many other reasons that could lead to this correlation (beside the informal argument that \"TPR captures structured information\")? Maybe the authors should compare with something else, for example the PMI between values of hidden neurons in a learned RNN and POS tags. Out of context, the numbers in Table 5 are not informative.\n- Polysemy: Only a very specific cherry picked example is given here. A more principled or in depth analysis of this phenomenon is needed to make a stronger case.\n\n* Typos *\n\n- \" The number of parameter matrices *is* the same as that of...\"\n- page 4 \"stables\" -> \"stabilizes\" (but rephrasing the sentence altogether would be better).\n- page 5: BiDAF misses the capital letters (\"bidaf\").\n- \"dev set\" -> \"validation set\" or \"development set\".\n- page 8: \"provides research*ers with* an intuitive...\"?\n"}, "tcdate": 1571883191129}, {"id": "ByeJA1orFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1046/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, a new unit based on the outer product called TPRU is proposed for recurrent neural networks. The performance of TPRU is validated with several NLP tasks such as POS tagging. \n\nWhile my knowledge about RNNs is limited, I feel the paper has room for improvement and I vote for rejection this time. The main reasons are: 1. the paper is not well written, and 2. the way of analysis is not enough.\n\n1. From the viewpoint of an RNN non-expert (i.e., me), the paper somehow fails to introduce the background. For example, tensor product representation (TPR) is introduced in the second paragraph of Introduction. While TPR is an elemental idea of this study, it is introduced with neither motivation (why and when TPR is useful, etc) nor appropriate references, which makes non-experts difficult to catch up with the main body of this study. So the paper is not self-contained enough. \n\n2-1. The paper tries to explain why TPRU is better in terms of gradient vanishing/explosion in Section 4. However, the analysis is mainly performed in a qualitative way, and there is no quantitative analysis of it. For example, I expect something like the evaluation of the magnitude of the gradient, e.g., how much degree the gradient scale is reduced from normal gate to the TPRU gate. Or, at least there should be the numerical experiments for the comparison. Otherwise, it is hard to judge whether the gradient is actually stabilized.\n\n2-2. The paper says one of the advantages of using TPRU is in its interpretability. However, the term \"interpretability\" is very vague and it is not properly defined in this paper. The paper should discuss what is the metric of interpretability here. More specifically, the paper claims TPRU's interpretability by Table 5. It looks, however, improper because there is no baseline and we cannot conclude that TPRU has better interpretability than others.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "N/A", "review": "In this paper, a new unit based on the outer product called TPRU is proposed for recurrent neural networks. The performance of TPRU is validated with several NLP tasks such as POS tagging. \n\nWhile my knowledge about RNNs is limited, I feel the paper has room for improvement and I vote for rejection this time. The main reasons are: 1. the paper is not well written, and 2. the way of analysis is not enough.\n\n1. From the viewpoint of an RNN non-expert (i.e., me), the paper somehow fails to introduce the background. For example, tensor product representation (TPR) is introduced in the second paragraph of Introduction. While TPR is an elemental idea of this study, it is introduced with neither motivation (why and when TPR is useful, etc) nor appropriate references, which makes non-experts difficult to catch up with the main body of this study. So the paper is not self-contained enough. \n\n2-1. The paper tries to explain why TPRU is better in terms of gradient vanishing/explosion in Section 4. However, the analysis is mainly performed in a qualitative way, and there is no quantitative analysis of it. For example, I expect something like the evaluation of the magnitude of the gradient, e.g., how much degree the gradient scale is reduced from normal gate to the TPRU gate. Or, at least there should be the numerical experiments for the comparison. Otherwise, it is hard to judge whether the gradient is actually stabilized.\n\n2-2. The paper says one of the advantages of using TPRU is in its interpretability. However, the term \"interpretability\" is very vague and it is not properly defined in this paper. The paper should discuss what is the metric of interpretability here. More specifically, the paper claims TPRU's interpretability by Table 5. It looks, however, improper because there is no baseline and we cannot conclude that TPRU has better interpretability than others."}, "tcdate": 1571299271110}, {"id": "SJxro8sodH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1046/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new recurrent unit with a simplified dynamics during training leading to more stable training algorithms and better performance. The paper is difficult to read because it assumes the reader is an expert on Tensor Product Representations. Many important terms are not clearly defined, which makes difficult to follow. For example, the terms “roles” and “filler” are not defined. I think a quick introduction to the field with a clarifying figure would be greatly appreciated by general readers. However, I think the contribution of the paper is important and presented experimental results, comparing the method against classical LSTM and GRU architectures, seem to be relevant for the field.\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review": "This paper proposes a new recurrent unit with a simplified dynamics during training leading to more stable training algorithms and better performance. The paper is difficult to read because it assumes the reader is an expert on Tensor Product Representations. Many important terms are not clearly defined, which makes difficult to follow. For example, the terms “roles” and “filler” are not defined. I think a quick introduction to the field with a clarifying figure would be greatly appreciated by general readers. However, I think the contribution of the paper is important and presented experimental results, comparing the method against classical LSTM and GRU architectures, seem to be relevant for the field.\n\n\n\n\n\n"}, "tcdate": 1570645660863}], "openreview_url": "https://openreview.net/forum?id=rkgFXR4KPr", "arxiv_id": "1810.12456", "paper_pdf": "papers/rkgFXR4KPr.pdf", "paper_pdf_sha256": "b2f265f3892a64708fa0c891f98f63a88c6a338080527d6b22dc9bd17d86e9bc", "paper_pdf_bytes": 637961, "paper_pdf_source": "openreview", "code_url": "https://github.com/shuaitang/TPRU", "code_repository": "shuaitang/TPRU", "code_commit": "bb5db2cb81cb6b248ec19211a15bc908dc4bef98", "code_archive": "repos/rkgFXR4KPr.zip", "code_archive_sha256": "51c5f960c6623a624443c5b96d845afbeffc697e845b4799c3cfda07a6d6c2d8", "code_archive_bytes": 51736, "code_file_count": 35, "code_extensions": {".py": 30, ".sh": 5}, "github_disk_usage_kb": 992, "github_languages": {"Python": 103450, "Shell": 1506}, "github_archived": false, "github_pushed_at": "2019-05-29T00:52:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-distributed-representations-of-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FQ8KX1pxeo", "year": 2026, "status": "rejected", "title": "Progressive Binarization with Semi-Structured Pruning for LLMs", "authors": ["Xianglong Yan", "Tianao Zhang", "Zhiteng Li", "Haotong Qin", "Yulun Zhang"], "authorids": ["~Xianglong_Yan2", "~Tianao_Zhang1", "~Zhiteng_Li2", "~Haotong_Qin1", "~Yulun_Zhang1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have achieved remarkable progress in natural language processing, but their high computational and memory costs hinder deployment on resource-constrained devices. Binarization represents the most extreme form of quantization, yet binarized models still contain redundancy that can be further removed. Pruning provides a natural way to eliminate such redundancy, but naïve combination with binarization often results in severe performance degradation. In this paper, we propose Progressive Binarization with Semi-Structured Pruning (PBS$^2$P), a novel post-training framework that seamlessly integrates binarization and semi-structured pruning. We first propose Stepwise semi-structured Pruning with Binarization Optimization (SPBO), which progressively introduces sparsity while optimizing binarization parameters to jointly reduce pruning and quantization error, yielding more stable and accurate compression. Additionally, we propose a Coarse-to-Fine Search (CFS) that first allocates pruning ratios and then refines element selection, further enhancing overall performance. Extensive experiments across multiple LLM families show that PBS$^2$P consistently outperforms state-of-the-art (SOTA) binary post-training quantization methods in both perplexity and downstream accuracy. We will release all the code and models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jhgkaen2Z3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22729/Reviewer_6aYr"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes Progressive Binarization with Semi-Structured Pruning (PBS²P) for compressing large language models (LLMs). The core component is SPBO (Stepwise Semi-Structured Pruning with Binarization Optimization), which progressively prunes a subset of elements at each step while jointly optimizing the binarized parameters, effectively reducing the combined error from pruning and binarization. In addition, the authors introduce a Coarse-to-Fine Search strategy to improve the accuracy of pruning element selection, further enhancing compression efficiency. Extensive experiments show that PBS²P outperforms existing post-training quantization methods across various LLM families and evaluation metric.", "review_text": "This paper proposes Progressive Binarization with Semi-Structured Pruning (PBS²P) for compressing large language models (LLMs). The core component is SPBO (Stepwise Semi-Structured Pruning with Binarization Optimization), which progressively prunes a subset of elements at each step while jointly optimizing the binarized parameters, effectively reducing the combined error from pruning and binarization. In addition, the authors introduce a Coarse-to-Fine Search strategy to improve the accuracy of pruning element selection, further enhancing compression efficiency. Extensive experiments show that PBS²P outperforms existing post-training quantization methods across various LLM families and evaluation metric.", "strengths": "1. Method design shows some innovation: The paper jointly optimizes pruning and binarization, using a stepwise strategy to reduce the error accumulation from single-step pruning.\n2. Comprehensive ablation studies: Experiments validate the contributions of the SPBO strategy as well as different metrics and pruning types to performance.\n3. Clear presentation: The writing is well-structured, and the workflow and formulas are described in detail, making the approach easy to understand.", "weaknesses": "1. Limited innovation: Although the combination of stepwise pruning and quantization is experimentally validated, it essentially remains a combination of pruning and quantization, resulting in moderate to low novelty.\n2. Hardware support limitations: The paper adopts 5:8 and 6:8 N:M sparsity configurations, but public documentation shows that NVIDIA GPUs only natively support 2:4 sparsity. Therefore, higher-ratio sparsity may not achieve hardware acceleration in practice.\n3. Unclear hyperparameter selection: The method for setting 𝑁high and 𝑁low is not specified, lacking theoretical justification or search strategy, which reduces reproducibility and interpretability.\n4. Optimality of mask decomposition not demonstrated: The stepwise progressive mask decomposition is not proven to be optimal, and there may exist schemes that achieve higher accuracy at the cost of longer runtime. The paper does not explore this trade-off.\n5. Method limitations: The SPBO’s stepwise updates rely on calibration data and multiple iterations, increasing computational cost. The impact on efficiency for large-scale model deployment is not thoroughly discussed.", "questions": "1. How are 𝑁high and 𝑁low selected? Is there a transferable principle or tuning strategy?\n2. For the 5:8 and 6:8 configurations, is hardware acceleration actually achieved, or are they only used for experimental comparison?\n3. Have other mask decomposition schemes been tried? Is there a better accuracy-runtime trade-off?\n4. What are the computational overhead and practical deployment costs of SPBO on large models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Progressive Binarization with Semi-Structured Pruning (PBS²P) for compressing large language models (LLMs). The core component is SPBO (Stepwise Semi-Structured Pruning with Binarization Optimization), which progressively prunes a subset of elements at each step while jointly optimizing the binarized parameters, effectively reducing the combined error from pruning and binarization. In addition, the authors introduce a Coarse-to-Fine Search strategy to improve the accuracy of pruning element selection, further enhancing compression efficiency. Extensive experiments show that PBS²P outperforms existing post-training quantization methods across various LLM families and evaluation metric.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Method design shows some innovation: The paper jointly optimizes pruning and binarization, using a stepwise strategy to reduce the error accumulation from single-step pruning.\n2. Comprehensive ablation studies: Experiments validate the contributions of the SPBO strategy as well as different metrics and pruning types to performance.\n3. Clear presentation: The writing is well-structured, and the workflow and formulas are described in detail, making the approach easy to understand.", "weaknesses": "1. Limited innovation: Although the combination of stepwise pruning and quantization is experimentally validated, it essentially remains a combination of pruning and quantization, resulting in moderate to low novelty.\n2. Hardware support limitations: The paper adopts 5:8 and 6:8 N:M sparsity configurations, but public documentation shows that NVIDIA GPUs only natively support 2:4 sparsity. Therefore, higher-ratio sparsity may not achieve hardware acceleration in practice.\n3. Unclear hyperparameter selection: The method for setting 𝑁high and 𝑁low is not specified, lacking theoretical justification or search strategy, which reduces reproducibility and interpretability.\n4. Optimality of mask decomposition not demonstrated: The stepwise progressive mask decomposition is not proven to be optimal, and there may exist schemes that achieve higher accuracy at the cost of longer runtime. The paper does not explore this trade-off.\n5. Method limitations: The SPBO’s stepwise updates rely on calibration data and multiple iterations, increasing computational cost. The impact on efficiency for large-scale model deployment is not thoroughly discussed.", "questions": "1. How are 𝑁high and 𝑁low selected? Is there a transferable principle or tuning strategy?\n2. For the 5:8 and 6:8 configurations, is hardware acceleration actually achieved, or are they only used for experimental comparison?\n3. Have other mask decomposition schemes been tried? Is there a better accuracy-runtime trade-off?\n4. What are the computational overhead and practical deployment costs of SPBO on large models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761856670814}, {"id": "4kvnboIiF4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22729/Reviewer_CQVV"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper proposes PBS2P, a post-training framework that combines binarization with semi-structured (N:M) pruning for LLM compression. The method consists of two key components: (1) Stepwise Pruning with Binarization Optimization (SPBO), which jointly optimizes weight pruning and binarization parameters, and (2) Coarse-to-Fine Search (CFS), a two-stage strategy that first allocates pruning ratios based on layer importance and then selects specific elements using Hessian-based metrics. Experimental results on LLaMA and OPT model families demonstrate improvements over STBLLM and other binary post-training quantization methods.", "review_text": "This paper proposes PBS2P, a post-training framework that combines binarization with semi-structured (N:M) pruning for LLM compression. The method consists of two key components: (1) Stepwise Pruning with Binarization Optimization (SPBO), which jointly optimizes weight pruning and binarization parameters, and (2) Coarse-to-Fine Search (CFS), a two-stage strategy that first allocates pruning ratios based on layer importance and then selects specific elements using Hessian-based metrics. Experimental results on LLaMA and OPT model families demonstrate improvements over STBLLM and other binary post-training quantization methods.", "strengths": "1.\tWell-motivated problem: Combining binarization with pruning to reduce redundancy and overcome performance degradation is a valuable research direction.\n2.\tComprehensive experiments: Extensive evaluation across multiple model families (LLaMA-1/2/3, OPT), datasets (perplexity and zero-shot), and model sizes demonstrates broad applicability.\n3.\tThorough ablations: Section 4.4 provides a good analysis of design choices (SPBO, search metrics, group size, etc.).", "weaknesses": "1.\tCertain techniques are not well explained, which may cause confusion and make reproduction difficult. See specific concerns in the Questions section below.\n2.\tComputational cost: Inverting block wise covariances even at size 128 is not cheap; the fine stage dominates runtime (109 min on 7B). Complexity and wall-time scaling to 65B/70B should be analyzed more carefully (per-layer cost, number of SPBO alternations τ, M−N steps).", "questions": "1.\tThe notation in Equation 4 is confusing: what is the shape of 1? If 1 is just a column unit vector, μ should be a scalar. However, the binarization center for each row should be different.\n2.\tFor the Coarse Stage, why not use gradient-based importance (e.g., Fisher information) or loss sensitivity?\n3.\tFor Equation 7, the \"+1/2\" term for rounding is not explained. Additionally, the concrete choices of N_high and N_low are not presented in the paper.\n4.\tTheorem 3.1 (Equation 8) is essentially a restatement of classical results from Optimal Brain Surgeon (Hassibi et al., 1993). The \"proof in supplementary\" claim doesn't add novelty—this is a well-known second-order approximation. What is the difference between Theorem 3.1 and the results from OBS?\n5.\tComputational cost: 111 minutes for LLaMA-7B is 2.5× slower than ARB-LLM. For larger models (65B), this could be prohibitive. Is there any study on the computation time for large models?\n6.\tTable 4(c) only analyzes the LLaMA-7B model. While RI causes a degradation, the LI metric seems to provide only a small improvement. There should be more justification on more models for importance selection in the coarse stage and the necessity of adaptive assignment.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes PBS2P, a post-training framework that combines binarization with semi-structured (N:M) pruning for LLM compression. The method consists of two key components: (1) Stepwise Pruning with Binarization Optimization (SPBO), which jointly optimizes weight pruning and binarization parameters, and (2) Coarse-to-Fine Search (CFS), a two-stage strategy that first allocates pruning ratios based on layer importance and then selects specific elements using Hessian-based metrics. Experimental results on LLaMA and OPT model families demonstrate improvements over STBLLM and other binary post-training quantization methods.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1.\tWell-motivated problem: Combining binarization with pruning to reduce redundancy and overcome performance degradation is a valuable research direction.\n2.\tComprehensive experiments: Extensive evaluation across multiple model families (LLaMA-1/2/3, OPT), datasets (perplexity and zero-shot), and model sizes demonstrates broad applicability.\n3.\tThorough ablations: Section 4.4 provides a good analysis of design choices (SPBO, search metrics, group size, etc.).", "weaknesses": "1.\tCertain techniques are not well explained, which may cause confusion and make reproduction difficult. See specific concerns in the Questions section below.\n2.\tComputational cost: Inverting block wise covariances even at size 128 is not cheap; the fine stage dominates runtime (109 min on 7B). Complexity and wall-time scaling to 65B/70B should be analyzed more carefully (per-layer cost, number of SPBO alternations τ, M−N steps).", "questions": "1.\tThe notation in Equation 4 is confusing: what is the shape of 1? If 1 is just a column unit vector, μ should be a scalar. However, the binarization center for each row should be different.\n2.\tFor the Coarse Stage, why not use gradient-based importance (e.g., Fisher information) or loss sensitivity?\n3.\tFor Equation 7, the \"+1/2\" term for rounding is not explained. Additionally, the concrete choices of N_high and N_low are not presented in the paper.\n4.\tTheorem 3.1 (Equation 8) is essentially a restatement of classical results from Optimal Brain Surgeon (Hassibi et al., 1993). The \"proof in supplementary\" claim doesn't add novelty—this is a well-known second-order approximation. What is the difference between Theorem 3.1 and the results from OBS?\n5.\tComputational cost: 111 minutes for LLaMA-7B is 2.5× slower than ARB-LLM. For larger models (65B), this could be prohibitive. Is there any study on the computation time for large models?\n6.\tTable 4(c) only analyzes the LLaMA-7B model. While RI causes a degradation, the LI metric seems to provide only a small improvement. There should be more justification on more models for importance selection in the coarse stage and the necessity of adaptive assignment.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761820345225}, {"id": "duYW0li0eL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22729/Reviewer_hctq"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 2, "summary": "This paper introduces PBS²P (Progressive Binarization with Semi-Structured Pruning), a pure post-training framework that pushes LLMs down to 0.55–0.8 bit weight precision while retaining SOTA perplexity and zero-shot accuracy on LLaMA/OPT families. \nThe method alternates two components:\nSPBO – step-wise N:M pruning followed by on-the-fly re-optimisation of binarisation scalars α/μ to reduce compound error.\nCFS – a coarse-to-fine search that first allocates layer-wise pruning rates via cosine-similarity importance and then picks elements to prune by a Hessian-based second-order criterion.", "review_text": "This paper introduces PBS²P (Progressive Binarization with Semi-Structured Pruning), a pure post-training framework that pushes LLMs down to 0.55–0.8 bit weight precision while retaining SOTA perplexity and zero-shot accuracy on LLaMA/OPT families. \nThe method alternates two components:\nSPBO – step-wise N:M pruning followed by on-the-fly re-optimisation of binarisation scalars α/μ to reduce compound error.\nCFS – a coarse-to-fine search that first allocates layer-wise pruning rates via cosine-similarity importance and then picks elements to prune by a Hessian-based second-order criterion.", "strengths": "1. The paper is well-written.\n2. The paper introduces PBS2P, a novel post-training framework that seamlessly integrates binarization (1-bit quantization) and semi-structured pruning (N:M sparsity), effectively reduces combined errors from pruning and quantization\n3. Ablation tests validate each component (e.g., SPBO, CFS metrics, pruning types), highlighting their necessity and superiority, which strengthens the method's credibility.", "weaknesses": "1. The proposed method involves some predefined constants, such as N_high and N_low in CFS, and hyperparameters like Optimization Steps. It is unclear how to set the values of these predefined constants whether the settings of these constants affect the final compression effectiveness. (I am concerned that there may be difficulties or troubles in setting these constants during practical applications.)\n2. The paper only tested zero-shot tasks on relatively old models, such as the Llama1 and Llama2 series. If applied to stronger models (e.g., Llama3 or Qwen3 series) after quantization and pruning, how would it perform on zero-shot tasks?", "questions": "1. How sensitive is the method in the paper to calibration data? Does the distribution of calibration data have an impact? How should calibration data be selected for training?\n2. The paper demonstrates efficiency advantages in matrix multiplication. How much efficiency improvement can it bring in normal inference tasks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PBS²P (Progressive Binarization with Semi-Structured Pruning), a pure post-training framework that pushes LLMs down to 0.55–0.8 bit weight precision while retaining SOTA perplexity and zero-shot accuracy on LLaMA/OPT families. \nThe method alternates two components:\nSPBO – step-wise N:M pruning followed by on-the-fly re-optimisation of binarisation scalars α/μ to reduce compound error.\nCFS – a coarse-to-fine search that first allocates layer-wise pruning rates via cosine-similarity importance and then picks elements to prune by a Hessian-based second-order criterion.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The paper is well-written.\n2. The paper introduces PBS2P, a novel post-training framework that seamlessly integrates binarization (1-bit quantization) and semi-structured pruning (N:M sparsity), effectively reduces combined errors from pruning and quantization\n3. Ablation tests validate each component (e.g., SPBO, CFS metrics, pruning types), highlighting their necessity and superiority, which strengthens the method's credibility.", "weaknesses": "1. The proposed method involves some predefined constants, such as N_high and N_low in CFS, and hyperparameters like Optimization Steps. It is unclear how to set the values of these predefined constants whether the settings of these constants affect the final compression effectiveness. (I am concerned that there may be difficulties or troubles in setting these constants during practical applications.)\n2. The paper only tested zero-shot tasks on relatively old models, such as the Llama1 and Llama2 series. If applied to stronger models (e.g., Llama3 or Qwen3 series) after quantization and pruning, how would it perform on zero-shot tasks?", "questions": "1. How sensitive is the method in the paper to calibration data? Does the distribution of calibration data have an impact? How should calibration data be selected for training?\n2. The paper demonstrates efficiency advantages in matrix multiplication. How much efficiency improvement can it bring in normal inference tasks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761750963448}, {"id": "3AtzcGIipq", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22729/Reviewer_zycs"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces a pipeline for jointly applying 1-bit quantization and semi-structured pruning to Large Language Models (LLMs). The method is built upon a layer-wise closed-form solution for 1-bit quantization, governed by two control parameters, which minimizes the reconstruction error under the Frobenius norm. For pruning, a global strategy assigns layer-specific sparsity ratios based on inter-layer similarity, and the Optimal Brain Surgeon (OBS) framework is used to select the weights for removal. A key aspect of the pipeline is the iterative update of the quantization parameters after each pruning step to maintain accuracy. The constituent techniques are established, the key contribution lies in the integration of these methods into an end-to-end pipeline.  However, given that SparseGPT also supports joint quantization and semi-structured pruning using the OBS framework, the similarities between the two methods should be explicitly clarified. Experimental results demonstrate that the proposed method outperforms compared methods. However, to strengthen the validation, evaluations on more recent model families like Qwen and comparisons against state-of-the-art baselines such as OmniQuant and  ParetoQ should be included.", "review_text": "This paper introduces a pipeline for jointly applying 1-bit quantization and semi-structured pruning to Large Language Models (LLMs). The method is built upon a layer-wise closed-form solution for 1-bit quantization, governed by two control parameters, which minimizes the reconstruction error under the Frobenius norm. For pruning, a global strategy assigns layer-specific sparsity ratios based on inter-layer similarity, and the Optimal Brain Surgeon (OBS) framework is used to select the weights for removal. A key aspect of the pipeline is the iterative update of the quantization parameters after each pruning step to maintain accuracy. The constituent techniques are established, the key contribution lies in the integration of these methods into an end-to-end pipeline.  However, given that SparseGPT also supports joint quantization and semi-structured pruning using the OBS framework, the similarities between the two methods should be explicitly clarified. Experimental results demonstrate that the proposed method outperforms compared methods. However, to strengthen the validation, evaluations on more recent model families like Qwen and comparisons against state-of-the-art baselines such as OmniQuant and  ParetoQ should be included.", "strengths": "The constituent techniques are established, the key contribution lies in the integration of these methods into an end-to-end pipeline. Experimental results demonstrate that the proposed method outperforms compared methods. The presentation is easy to follow.", "weaknesses": "This work presents a pipeline for joint quantization and semi-structure pruning, but it has several weaknesses that limit its current impact. The methodological description lacks clarity in key areas, such as the handling of activation flow during compression (independent vs. sequential) and the specific role of hyperparameters like the block size. The technical foundation is questioned, particularly the choice of the cosine similarity metric without a clear strategy for handling negative values or empirical validation of its distribution. The experimental validation is limited in scope, relying heavily on WikiText2 and older model families like LLaMA, while omitting stronger recent baselines and a thorough comparison to the closely related SparseGPT. Finally, the paper would benefit from a discussion of the method's applicability to broader architectures like MoE and diffusion models.", "questions": "1. Regarding the formulation in Equation 2, it appears the compression objective for a linear layer uses the original, unmodified activations. Could you clarify the following? (a) Is each linear layer's compression treated as an independent objective, using the original network's activations? (b) Or is the compression applied sequentially, where the input to a layer comes from the already compressedprevious layer?\n\n2. You use cosine similarity, which ranges from [-1, 1], to gauge layer importance. Could you please clarify how your algorithm handles cases where the cosine similarity is zero or negative? The subsequent use of a reciprocal (i.e., 1/similarity) in your global pruning ratio assignment would be undefined or invert the intended importance ranking in such scenarios. Was this encountered in practice, and if so, how was it addressed?\n\n3. Could you show some empirical results of the cosine similarity values computed for the layers of a model (e.g., Llama-2 7B)?\n\n4. The results (Tables 1 and 2) indicate the use of a block size of 128. Could you please clarify the role of this parameter in your method? Specifically, is this block size exclusively for the 1-bit quantization process?\n\n5. The experimental evaluation currently reports perplexity results primarily on WikiText2. To more thoroughly and fairly assess the generalizability of the proposed method, it would be beneficial to follow the common practice established by your cited baselines (e.g., GPTQ, BiLLM). Could you please include perplexity results on additional standard datasets, such as PTB and C4?\n\n6. The experimental validation is conducted on established model families like LLaMA (1-3) and OPT. To further demonstrate the relevance and effectiveness of the method, it would be valuable to include results on more recent and widely-used models, such as the Qwen series.\n\n7. The method is presented in the context of standard dense transformer-based LLMs. Could you comment on its potential adaptability to other important model classes (e.g., MoE models, diffusion models)?\n\n8. From a general perspective, the goal of jointly performing quantization and pruning is also a key feature of the SparseGPT [1] framework. Specifically, SparseGPT supports various quantization bit-widths alongside semi-structured pruning, and similarly utilizes the OBS framework for weight selection and error minimization. Given these high-level similarities, could you please provide a more detailed discussion of the fundamental differences between your method and SparseGPT?\n\n9. The experimental comparisons would be strengthened by including recent state-of-the-art methods that support extreme low-bit quantization of LLMs, such as ​​OmniQuant [2]​​ and ​​ParetoQ [3]​​.\n\n10. The paper reports computational savings based on the latency of a single matrix multiplication operation. However, in real-world deployment, end-to-end inference time, which includes I/O overhead, memory access patterns, and other system-level bottlenecks, is a more meaningful metric for evaluating efficiency. Could you please provide measurements of the end-to-end inference latency (e.g., tokens/second) for a complete forward pass on a standard benchmark?\n\n[1] Frantar, Elias, and Dan Alistarh. \"Sparsegpt: Massive language models can be accurately pruned in one-shot.\" International conference on machine learning. PMLR, 2023.\n\n[2] Shao, Wenqi, et al. \"Omniquant: Omnidirectionally calibrated quantization for large language models.\" arXiv preprint arXiv:2308.13137 (2023).\n\n[3] Liu, Zechun, et al. \"Paretoq: Scaling laws in extremely low-bit llm quantization.\" arXiv preprint arXiv:2502.02631 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a pipeline for jointly applying 1-bit quantization and semi-structured pruning to Large Language Models (LLMs). The method is built upon a layer-wise closed-form solution for 1-bit quantization, governed by two control parameters, which minimizes the reconstruction error under the Frobenius norm. For pruning, a global strategy assigns layer-specific sparsity ratios based on inter-layer similarity, and the Optimal Brain Surgeon (OBS) framework is used to select the weights for removal. A key aspect of the pipeline is the iterative update of the quantization parameters after each pruning step to maintain accuracy. The constituent techniques are established, the key contribution lies in the integration of these methods into an end-to-end pipeline.  However, given that SparseGPT also supports joint quantization and semi-structured pruning using the OBS framework, the similarities between the two methods should be explicitly clarified. Experimental results demonstrate that the proposed method outperforms compared methods. However, to strengthen the validation, evaluations on more recent model families like Qwen and comparisons against state-of-the-art baselines such as OmniQuant and  ParetoQ should be included.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The constituent techniques are established, the key contribution lies in the integration of these methods into an end-to-end pipeline. Experimental results demonstrate that the proposed method outperforms compared methods. The presentation is easy to follow.", "weaknesses": "This work presents a pipeline for joint quantization and semi-structure pruning, but it has several weaknesses that limit its current impact. The methodological description lacks clarity in key areas, such as the handling of activation flow during compression (independent vs. sequential) and the specific role of hyperparameters like the block size. The technical foundation is questioned, particularly the choice of the cosine similarity metric without a clear strategy for handling negative values or empirical validation of its distribution. The experimental validation is limited in scope, relying heavily on WikiText2 and older model families like LLaMA, while omitting stronger recent baselines and a thorough comparison to the closely related SparseGPT. Finally, the paper would benefit from a discussion of the method's applicability to broader architectures like MoE and diffusion models.", "questions": "1. Regarding the formulation in Equation 2, it appears the compression objective for a linear layer uses the original, unmodified activations. Could you clarify the following? (a) Is each linear layer's compression treated as an independent objective, using the original network's activations? (b) Or is the compression applied sequentially, where the input to a layer comes from the already compressedprevious layer?\n\n2. You use cosine similarity, which ranges from [-1, 1], to gauge layer importance. Could you please clarify how your algorithm handles cases where the cosine similarity is zero or negative? The subsequent use of a reciprocal (i.e., 1/similarity) in your global pruning ratio assignment would be undefined or invert the intended importance ranking in such scenarios. Was this encountered in practice, and if so, how was it addressed?\n\n3. Could you show some empirical results of the cosine similarity values computed for the layers of a model (e.g., Llama-2 7B)?\n\n4. The results (Tables 1 and 2) indicate the use of a block size of 128. Could you please clarify the role of this parameter in your method? Specifically, is this block size exclusively for the 1-bit quantization process?\n\n5. The experimental evaluation currently reports perplexity results primarily on WikiText2. To more thoroughly and fairly assess the generalizability of the proposed method, it would be beneficial to follow the common practice established by your cited baselines (e.g., GPTQ, BiLLM). Could you please include perplexity results on additional standard datasets, such as PTB and C4?\n\n6. The experimental validation is conducted on established model families like LLaMA (1-3) and OPT. To further demonstrate the relevance and effectiveness of the method, it would be valuable to include results on more recent and widely-used models, such as the Qwen series.\n\n7. The method is presented in the context of standard dense transformer-based LLMs. Could you comment on its potential adaptability to other important model classes (e.g., MoE models, diffusion models)?\n\n8. From a general perspective, the goal of jointly performing quantization and pruning is also a key feature of the SparseGPT [1] framework. Specifically, SparseGPT supports various quantization bit-widths alongside semi-structured pruning, and similarly utilizes the OBS framework for weight selection and error minimization. Given these high-level similarities, could you please provide a more detailed discussion of the fundamental differences between your method and SparseGPT?\n\n9. The experimental comparisons would be strengthened by including recent state-of-the-art methods that support extreme low-bit quantization of LLMs, such as ​​OmniQuant [2]​​ and ​​ParetoQ [3]​​.\n\n10. The paper reports computational savings based on the latency of a single matrix multiplication operation. However, in real-world deployment, end-to-end inference time, which includes I/O overhead, memory access patterns, and other system-level bottlenecks, is a more meaningful metric for evaluating efficiency. Could you please provide measurements of the end-to-end inference latency (e.g., tokens/second) for a complete forward pass on a standard benchmark?\n\n[1] Frantar, Elias, and Dan Alistarh. \"Sparsegpt: Massive language models can be accurately pruned in one-shot.\" International conference on machine learning. PMLR, 2023.\n\n[2] Shao, Wenqi, et al. \"Omniquant: Omnidirectionally calibrated quantization for large language models.\" arXiv preprint arXiv:2308.13137 (2023).\n\n[3] Liu, Zechun, et al. \"Paretoq: Scaling laws in extremely low-bit llm quantization.\" arXiv preprint arXiv:2502.02631 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761553435110}], "openreview_url": "https://openreview.net/forum?id=FQ8KX1pxeo", "arxiv_id": "2502.01705", "paper_pdf": "papers/FQ8KX1pxeo.pdf", "paper_pdf_sha256": "9193806b34f46f6cecdf33d1ecf48f889ab500f6ec26cde374f9593752fef38a", "paper_pdf_bytes": 494632, "paper_pdf_source": "openreview", "code_url": "https://github.com/XIANGLONGYAN/PBS2P", "code_repository": "XIANGLONGYAN/PBS2P", "code_commit": "923548007fe2d40e146b47f4a05c9d9526b496d9", "code_archive": "repos/FQ8KX1pxeo.zip", "code_archive_sha256": "680299f689c80f1ce625ea3a2709009bdefb199668c32b09fb8c22aae4ed527c", "code_archive_bytes": 516334, "code_file_count": 89, "code_extensions": {".py": 88, ".sh": 1}, "github_disk_usage_kb": 2073, "github_languages": {"Python": 534298, "Shell": 624}, "github_archived": false, "github_pushed_at": "2026-07-11T12:52:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/progressive-binarization-with-semi-structured"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xAM9VaXZnY", "year": 2025, "status": "rejected", "title": "What Can We Learn from State Space Models for Machine Learning on Graphs?", "authors": ["Yinan Huang", "Siqi Miao", "Pan Li"], "authorids": ["~Yinan_Huang1", "~Siqi_Miao1", "~Pan_Li2"], "authors_source": "OpenReview API", "abstract": "Machine learning on graphs has recently found extensive applications across domains. However, the commonly used Message Passing Neural Networks (MPNNs) suffer from limited expressive power and struggle to capture long-range dependencies. Graph transformers offer a strong alternative due to their global attention mechanism, but they come with great computational overheads, especially for large graphs. In recent years, State Space Models (SSMs) have emerged as a compelling approach to replace full attention in transformers to model sequential data. It blends the strengths of RNNs and CNNs, offering a) efficient computation, b) the ability to capture long-range dependencies, and c) good generalization across sequences of various lengths. However, extending SSMs to graph-structured data presents unique challenges due to the lack of canonical node ordering in graphs. In this work, we propose Graph State Space Convolution (GSSC) as a principled extension of SSMs to graph-structured data. By leveraging global permutation-equivariant set aggregation and factorizable graph kernels that rely on relative node distances as the convolution kernels, GSSC preserves all three advantages of SSMs. We demonstrate the provably stronger expressiveness of GSSC than MPNNs in counting graph substructures and show its effectiveness across 11 real-world, widely used benchmark datasets. GSSC achieves the best results on 6 out of 11 datasets with all significant improvements compared to the state-of-the-art baselines and second-best results on the other 5 datasets. Our findings highlight the potential of GSSC as a powerful and scalable model for graph machine learning. Anonymous code\nis available at https://anonymous.4open.science/r/GSSC-5ED8.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "JjP7QkqWRV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7356/Reviewer_TB1r"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This work extends the state space models (SSMs) from sequence modeling to the domain of  graph-structured data. By tailoring SSMs for graphs, the proposed model (GSSC)\ncan capture long-range dependencies and overcome the limitations of Message Passing Neural Networks (MPNNs) while offering efficient computation addressing the quadratic computation of the graph transformers. To preserve the permutation equivariance in the graphs, it outlines a method for designing a permutation-invariant kernel for convolution operations on graphs. Furthermore, it extends to a data-dependent version of the proposed model by defining a selection mechanism for graph-structured data. The proposed model demonstrates provably stronger expressiveness than MPNNs and Graph Spectral Convolution in counting graph substructures.", "review_text": "This work extends the state space models (SSMs) from sequence modeling to the domain of  graph-structured data. By tailoring SSMs for graphs, the proposed model (GSSC)\ncan capture long-range dependencies and overcome the limitations of Message Passing Neural Networks (MPNNs) while offering efficient computation addressing the quadratic computation of the graph transformers. To preserve the permutation equivariance in the graphs, it outlines a method for designing a permutation-invariant kernel for convolution operations on graphs. Furthermore, it extends to a data-dependent version of the proposed model by defining a selection mechanism for graph-structured data. The proposed model demonstrates provably stronger expressiveness than MPNNs and Graph Spectral Convolution in counting graph substructures.", "strengths": "- The paper provide a method to customize the recursion in the SSMs for graphs,\n- It proposes a method for designing a permutation-invariant kernel for efficient global convolution operations on graphs.\n- The paper demonstrates provably stronger expressiveness of the proposed model for 3-paths and 3-cycles and also for 4-paths and 4-cycles.", "weaknesses": "1. The distinction between the proposed model (equation 4) and linear graph attention is minimal. It is known, even prior to Dao & Gu (2024), that SSMs can be represented as linear attention. The global convolution representation used here is also an equivalent form of SSMs. Additionally, the use of positional encodings is not novel, as it is a feature used by other works such as GraphGPS, which propose a general framework for transformer-based models and can adopt their approximations like Performer and linear attention. \n\n2. The presentation requires improvement, and some claims need to be more precise. For example:  \n   - a. at line 143, the *Computational efficiency and parallelism*,  The text should clarify that SSMs like S4, when represented as global convolution (equation 2), they can leverage the FFT algorithm to result in parallel and efficient quasi-linear complexity. Parallel scan is used for the recurrence form of equation (1) under certain conditions (employed by data-dependent SSMs like Mamba), which also results in O(n log n) computation but offers parallelization.  \n   - b. The statement that the kernel should be factorizable as a dot product to be permutation-invariant (line 127),  needs revision. The dot product between absolute position representations is one way to achieve translation invariance. Reference [2] offers alternative methods, like cross-correlation, to achieve translation invariance in kernels for global convolution using translation-equivariant functions.  \n   - c. The factorized form in line 259 is not equivalent to the data-dependent convolution presented in the preceding line.   \n   - d. The new positional encodings in equation (6) are positional encodings of all nodes but utilize only the features of node u (as it is a function of $x_u$ only). Lines 267 and 267 need correction to reflect this.   \n   - e.The text should elaborate on how $\\phi()$ in  in eqn (5)  is modeled to be a permutation equivariant function and capture interactions between frequencies.\n\n\n\n\n3. **Insufficient Empirical Studies**: The literature review lacks citations for many state-of-the-art models, and the experimental section lacks comparisons to them.  \n   - Since the proposed model is compared with GSC and Linear Graph Transformers in the paper, it is essential to include their performance comparison in the experimental section (particularly in Table 2 for graph substructure counting, where the proposed model is expected to show superior expressiveness). \n   - Comparisons with newer models like Spatial-Spectral GNN [1] and Polynormer [3] are also necessary. \n   - Additionally, comparisons against other SSM-based models (Graph-Mamba I and II), GSC, Linear Graph Transformers and recent models in Tables 2 and 3 are important to demonstrate the proposed model empirically.", "questions": "1. Please address the aforementioned concerns\n2. Some typos:  \n   - Equation 5 (line 201): the first term doesn’t require parentheses. \n   - “reply on” -> rely on in line 268 and 346. \n   - There are duplicate references for Behrouz & Hashemi, 2024 \n\n\nConclusion:\n\nWhile the proposed idea is appealing, and I acknowledge its potential impact, I am not sure that the paper is ready for ICLR in its current form. Therefore, I am hesitant to fully support acceptance but I am willing to increase my score if the concerns and questions are addressed.\n\n**References:**    \n[1] Geisler, S., Kosmala, A., Herbst, D., & Günnemann, S. (2024). Spatio-Spectral Graph Neural Networks. arXiv preprint arXiv:2405.19121.  \n[2] M. Karami, A. Ghodsi, Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling .” In Thirty-eighth Conference on Advances in Neural Information Processing Systems 2024.  \n[3] Deng, Chenhui, Zichao Yue, and Zhiru Zhang. \"Polynormer: Polynomial-Expressive Graph Transformer in Linear Time.\" The Twelfth International Conference on Learning Representations.\n\nUpdate 1: added references", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work extends the state space models (SSMs) from sequence modeling to the domain of  graph-structured data. By tailoring SSMs for graphs, the proposed model (GSSC)\ncan capture long-range dependencies and overcome the limitations of Message Passing Neural Networks (MPNNs) while offering efficient computation addressing the quadratic computation of the graph transformers. To preserve the permutation equivariance in the graphs, it outlines a method for designing a permutation-invariant kernel for convolution operations on graphs. Furthermore, it extends to a data-dependent version of the proposed model by defining a selection mechanism for graph-structured data. The proposed model demonstrates provably stronger expressiveness than MPNNs and Graph Spectral Convolution in counting graph substructures.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper provide a method to customize the recursion in the SSMs for graphs,\n- It proposes a method for designing a permutation-invariant kernel for efficient global convolution operations on graphs.\n- The paper demonstrates provably stronger expressiveness of the proposed model for 3-paths and 3-cycles and also for 4-paths and 4-cycles.", "weaknesses": "1. The distinction between the proposed model (equation 4) and linear graph attention is minimal. It is known, even prior to Dao & Gu (2024), that SSMs can be represented as linear attention. The global convolution representation used here is also an equivalent form of SSMs. Additionally, the use of positional encodings is not novel, as it is a feature used by other works such as GraphGPS, which propose a general framework for transformer-based models and can adopt their approximations like Performer and linear attention. \n\n2. The presentation requires improvement, and some claims need to be more precise. For example:  \n   - a. at line 143, the *Computational efficiency and parallelism*,  The text should clarify that SSMs like S4, when represented as global convolution (equation 2), they can leverage the FFT algorithm to result in parallel and efficient quasi-linear complexity. Parallel scan is used for the recurrence form of equation (1) under certain conditions (employed by data-dependent SSMs like Mamba), which also results in O(n log n) computation but offers parallelization.  \n   - b. The statement that the kernel should be factorizable as a dot product to be permutation-invariant (line 127),  needs revision. The dot product between absolute position representations is one way to achieve translation invariance. Reference [2] offers alternative methods, like cross-correlation, to achieve translation invariance in kernels for global convolution using translation-equivariant functions.  \n   - c. The factorized form in line 259 is not equivalent to the data-dependent convolution presented in the preceding line.   \n   - d. The new positional encodings in equation (6) are positional encodings of all nodes but utilize only the features of node u (as it is a function of $x_u$ only). Lines 267 and 267 need correction to reflect this.   \n   - e.The text should elaborate on how $\\phi()$ in  in eqn (5)  is modeled to be a permutation equivariant function and capture interactions between frequencies.\n\n\n\n\n3. **Insufficient Empirical Studies**: The literature review lacks citations for many state-of-the-art models, and the experimental section lacks comparisons to them.  \n   - Since the proposed model is compared with GSC and Linear Graph Transformers in the paper, it is essential to include their performance comparison in the experimental section (particularly in Table 2 for graph substructure counting, where the proposed model is expected to show superior expressiveness). \n   - Comparisons with newer models like Spatial-Spectral GNN [1] and Polynormer [3] are also necessary. \n   - Additionally, comparisons against other SSM-based models (Graph-Mamba I and II), GSC, Linear Graph Transformers and recent models in Tables 2 and 3 are important to demonstrate the proposed model empirically.", "questions": "1. Please address the aforementioned concerns\n2. Some typos:  \n   - Equation 5 (line 201): the first term doesn’t require parentheses. \n   - “reply on” -> rely on in line 268 and 346. \n   - There are duplicate references for Behrouz & Hashemi, 2024 \n\n\nConclusion:\n\nWhile the proposed idea is appealing, and I acknowledge its potential impact, I am not sure that the paper is ready for ICLR in its current form. Therefore, I am hesitant to fully support acceptance but I am willing to increase my score if the concerns and questions are addressed.\n\n**References:**    \n[1] Geisler, S., Kosmala, A., Herbst, D., & Günnemann, S. (2024). Spatio-Spectral Graph Neural Networks. arXiv preprint arXiv:2405.19121.  \n[2] M. Karami, A. Ghodsi, Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling .” In Thirty-eighth Conference on Advances in Neural Information Processing Systems 2024.  \n[3] Deng, Chenhui, Zichao Yue, and Zhiru Zhang. \"Polynormer: Polynomial-Expressive Graph Transformer in Linear Time.\" The Twelfth International Conference on Learning Representations.\n\nUpdate 1: added references", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731295080462}, {"id": "SEERQcUWDD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7356/Reviewer_N4o9"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces the Graph State Space Convolution (GSSC) method, an extension of State Space Models (SSMs) to graph data. GSSC utilizes global permutation-equivariant set aggregation and factorizable graph kernels based on relative node distances as its convolution kernels.", "review_text": "The paper introduces the Graph State Space Convolution (GSSC) method, an extension of State Space Models (SSMs) to graph data. GSSC utilizes global permutation-equivariant set aggregation and factorizable graph kernels based on relative node distances as its convolution kernels.", "strengths": "I have found the paper well written and self-contained. I think a non-expert could find most of the information in the paper, and I appreciate this aspect.\n\n Figures 1 and 2, which demonstrate the problem domain and architecture, are interesting and easy to read. I commend the authors for their explicit effort in making these illustrations clear and informative.\n\nThe insights are didactical and well communicated. The conclusion given by the experiments looks interesting and valuable for future practitioners, while I think a synthesis would be beneficial for the reader.", "weaknesses": "Despite these merits, I have the following concerns about the paper.\n\n1- While there is a careful analysis of the different design decisions/performance tradeoffs, I feel that there is only a limited understanding about what are the properties of the Architecture that lead to these decisions/performance differences.\n\n2-  Scalability Concerns: The paper acknowledges the challenge of scalability for larger graphs. To address this, the authors could explore methods for optimizing the computational complexity of the GSSC architecture.\n\n3- Weak experimental study:  The paper lacks experimental evaluation of the Graph State Space Convolution (GSSC) method on heterophilic datasets. This omission is significant as such studies are crucial to assess GSSC's performance with heterophilic data and its robustness against issues like over-squashing and over-smoothing, which are common challenges in graph data analysis.", "questions": "(i) For the scalable version of the architecture when you proposed using the first k eigenvectors, how does the method address the issue of missing eigenvectors in the middle and end of the spectrum?\n\n(ii) What are the main challenges in extending GSSC to heterophilic graphs, and do you have any preliminary insights on how these challenges could be addressed?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the Graph State Space Convolution (GSSC) method, an extension of State Space Models (SSMs) to graph data. GSSC utilizes global permutation-equivariant set aggregation and factorizable graph kernels based on relative node distances as its convolution kernels.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "I have found the paper well written and self-contained. I think a non-expert could find most of the information in the paper, and I appreciate this aspect.\n\n Figures 1 and 2, which demonstrate the problem domain and architecture, are interesting and easy to read. I commend the authors for their explicit effort in making these illustrations clear and informative.\n\nThe insights are didactical and well communicated. The conclusion given by the experiments looks interesting and valuable for future practitioners, while I think a synthesis would be beneficial for the reader.", "weaknesses": "Despite these merits, I have the following concerns about the paper.\n\n1- While there is a careful analysis of the different design decisions/performance tradeoffs, I feel that there is only a limited understanding about what are the properties of the Architecture that lead to these decisions/performance differences.\n\n2-  Scalability Concerns: The paper acknowledges the challenge of scalability for larger graphs. To address this, the authors could explore methods for optimizing the computational complexity of the GSSC architecture.\n\n3- Weak experimental study:  The paper lacks experimental evaluation of the Graph State Space Convolution (GSSC) method on heterophilic datasets. This omission is significant as such studies are crucial to assess GSSC's performance with heterophilic data and its robustness against issues like over-squashing and over-smoothing, which are common challenges in graph data analysis.", "questions": "(i) For the scalable version of the architecture when you proposed using the first k eigenvectors, how does the method address the issue of missing eigenvectors in the middle and end of the spectrum?\n\n(ii) What are the main challenges in extending GSSC to heterophilic graphs, and do you have any preliminary insights on how these challenges could be addressed?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730809590750}, {"id": "ffFHyJUXK9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7356/Reviewer_wW9i"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "Machine learning on graphs often uses Message Passing Neural Networks (MPNNs), but these have limited expressive power. Graph State Space Convolution (GSSC) is proposed as a new method that extends State Space Models (SSMs) to graph data, overcoming these limitations and demonstrating superior performance on benchmark datasets. GSSC achieves good results on many datasets and offers a scalable solution for graph machine learning.", "review_text": "Machine learning on graphs often uses Message Passing Neural Networks (MPNNs), but these have limited expressive power. Graph State Space Convolution (GSSC) is proposed as a new method that extends State Space Models (SSMs) to graph data, overcoming these limitations and demonstrating superior performance on benchmark datasets. GSSC achieves good results on many datasets and offers a scalable solution for graph machine learning.", "strengths": "1. **Novelty**: Unlike previous methods that design SSM variants on graphs, this paper approaches the problem from the perspective of separable distance kernels on graphs. This idea is both novel and interesting.\n\n2. **Comprehensive Evaluations**: The experimental section of this paper covers synthetic datasets, real datasets, and computational efficiency benchmarks, providing a comprehensive evaluation.", "weaknesses": "1. **Experimental Design**: My main concern is that the model evaluated in the experiments is a hybrid of GSSC and MPNN, without showing the performance of GSSC alone. This makes it difficult to assess the independent contribution of GSSC, affecting the judgment of its effectiveness. I strongly recommend that the authors provide experimental results for the GSSC module alone, which will help readers understand the primary contributions of GSSC.\n\n2. **Baseline Selection**: In Section 5.1, the authors compare only with MPNN, without providing baseline comparisons with SSM or GT, which seems insufficient. Considering the authors claim that GSSC is a replacement module for Transformers, I suggest adding comparisons with other models that include shortest path information (such as high-order MPNN) or global information (such as Graph Transformer). Moreover, the current baselines are mainly single MPNN models, and comparing them with the hybrid model (GSSC+MPNN) seems unfair.\n\n3. **Inconsistency in Model Architecture**: In Section 5.1, the authors use selective GSSC, but it is not used in other benchmarks.If different architectural variants of the model are used in the experiments, I advise clearly indicating this in the tables. Additionally, the authors claim that GSSC without selective is already powerful enough, does this imply that using selective GSSC would be better? If so, it is recommended to provide relevant experimental results to support this argument. If selective GSSC is not used due to other disadvantages (such as computational efficiency), it would be beneficial to discuss this further in the paper. Generally, a unified model architecture is more attractive than one that requires adjustments across different benchmarks.", "questions": "1. Compared to other graph SSM works, this study indeed offers a novel perspective. In your opinion, what advantages does this new viewpoint provide over other methods? How do your experiments support these advantages?\n\n2. The title is \"What Can We Learn from State Space Models for Machine Learning on Graphs.\" Could you elaborate a bit more on other graph SSM models / SSM model on other domins, and compare them with GSSC?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Machine learning on graphs often uses Message Passing Neural Networks (MPNNs), but these have limited expressive power. Graph State Space Convolution (GSSC) is proposed as a new method that extends State Space Models (SSMs) to graph data, overcoming these limitations and demonstrating superior performance on benchmark datasets. GSSC achieves good results on many datasets and offers a scalable solution for graph machine learning.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. **Novelty**: Unlike previous methods that design SSM variants on graphs, this paper approaches the problem from the perspective of separable distance kernels on graphs. This idea is both novel and interesting.\n\n2. **Comprehensive Evaluations**: The experimental section of this paper covers synthetic datasets, real datasets, and computational efficiency benchmarks, providing a comprehensive evaluation.", "weaknesses": "1. **Experimental Design**: My main concern is that the model evaluated in the experiments is a hybrid of GSSC and MPNN, without showing the performance of GSSC alone. This makes it difficult to assess the independent contribution of GSSC, affecting the judgment of its effectiveness. I strongly recommend that the authors provide experimental results for the GSSC module alone, which will help readers understand the primary contributions of GSSC.\n\n2. **Baseline Selection**: In Section 5.1, the authors compare only with MPNN, without providing baseline comparisons with SSM or GT, which seems insufficient. Considering the authors claim that GSSC is a replacement module for Transformers, I suggest adding comparisons with other models that include shortest path information (such as high-order MPNN) or global information (such as Graph Transformer). Moreover, the current baselines are mainly single MPNN models, and comparing them with the hybrid model (GSSC+MPNN) seems unfair.\n\n3. **Inconsistency in Model Architecture**: In Section 5.1, the authors use selective GSSC, but it is not used in other benchmarks.If different architectural variants of the model are used in the experiments, I advise clearly indicating this in the tables. Additionally, the authors claim that GSSC without selective is already powerful enough, does this imply that using selective GSSC would be better? If so, it is recommended to provide relevant experimental results to support this argument. If selective GSSC is not used due to other disadvantages (such as computational efficiency), it would be beneficial to discuss this further in the paper. Generally, a unified model architecture is more attractive than one that requires adjustments across different benchmarks.", "questions": "1. Compared to other graph SSM works, this study indeed offers a novel perspective. In your opinion, what advantages does this new viewpoint provide over other methods? How do your experiments support these advantages?\n\n2. The title is \"What Can We Learn from State Space Models for Machine Learning on Graphs.\" Could you elaborate a bit more on other graph SSM models / SSM model on other domins, and compare them with GSSC?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730558757211}, {"id": "b2GxnyWmjp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7356/Reviewer_CutC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes Graph State Space Convolution (GSSC), an extension of state-space models to graph-structured data. The authors emphasize GSSC’s capability for capturing long-range dependencies in linear time and demonstrate competitive performance across several benchmarks.", "review_text": "This paper proposes Graph State Space Convolution (GSSC), an extension of state-space models to graph-structured data. The authors emphasize GSSC’s capability for capturing long-range dependencies in linear time and demonstrate competitive performance across several benchmarks.", "strengths": "1. The paper achieves competitive performance compared to SOTA methods.\n2. The paper proposes a novel approach in extending state-space models to graphs.\n3. The paper provides a plausible framework for capturing long-range dependencies efficiently.", "weaknesses": "My main concerns relate to the complexity claims, specifically:\n1. **Layer Complexity vs Expressivity:** The paper states that the complexity of a GSSC layer is “… $O(nmd)$


 where $n$ is the number of nodes and $m$, $d$ are hidden and positional encoding dimension.” (239-240) This means that a GSSC layer has $O(|V|)$ complexity


. Consequently, GSSC is in general *incapable of examining every edge in the graph*, unlike MPNNs with $O(|V|+|E|)$ complexity, such as GINE [1]. Although the authors prove that GSSC is “more powerful than MPNNs” *in terms of the WL hierarchy*, the expressivity implications of not being able to examine every edge seems to be overlooked. Even preprocessing the graph to incorporate edge features into nodes, which itself requires $O(|E|)$ time, would still necessitate $m\\in O(\\frac{|E|}{|V|})$ to store these features without information loss, thus exceeding $O(|V|)$ complexity overall.\n2. **Preprocessing Complexity:** The paper claims that finding the top $d$ eigenpairs with Lanczos methods has $O(nd^2)$ complexity (286-287). However, since sparse matrix-vector multiplication with the Laplacian matrix is necessary for Lanczos methods, the complexity per iteration would be at least $O(|E|)$, resulting in an overall complexity of at least $O(d|E|)$ to find $d$ eigenpairs. This exceeds the paper’s claim of $O(nd^2)$ preprocessing, and thus requires clarification.\n\nDespite these concerns, the paper’s main contributions remain valid. I would appreciate if the authors could clarify these points during the rebuttal phase.\n\n[1] https://arxiv.org/abs/1905.12265", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Graph State Space Convolution (GSSC), an extension of state-space models to graph-structured data. The authors emphasize GSSC’s capability for capturing long-range dependencies in linear time and demonstrate competitive performance across several benchmarks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper achieves competitive performance compared to SOTA methods.\n2. The paper proposes a novel approach in extending state-space models to graphs.\n3. The paper provides a plausible framework for capturing long-range dependencies efficiently.", "weaknesses": "My main concerns relate to the complexity claims, specifically:\n1. **Layer Complexity vs Expressivity:** The paper states that the complexity of a GSSC layer is “… $O(nmd)$


 where $n$ is the number of nodes and $m$, $d$ are hidden and positional encoding dimension.” (239-240) This means that a GSSC layer has $O(|V|)$ complexity


. Consequently, GSSC is in general *incapable of examining every edge in the graph*, unlike MPNNs with $O(|V|+|E|)$ complexity, such as GINE [1]. Although the authors prove that GSSC is “more powerful than MPNNs” *in terms of the WL hierarchy*, the expressivity implications of not being able to examine every edge seems to be overlooked. Even preprocessing the graph to incorporate edge features into nodes, which itself requires $O(|E|)$ time, would still necessitate $m\\in O(\\frac{|E|}{|V|})$ to store these features without information loss, thus exceeding $O(|V|)$ complexity overall.\n2. **Preprocessing Complexity:** The paper claims that finding the top $d$ eigenpairs with Lanczos methods has $O(nd^2)$ complexity (286-287). However, since sparse matrix-vector multiplication with the Laplacian matrix is necessary for Lanczos methods, the complexity per iteration would be at least $O(|E|)$, resulting in an overall complexity of at least $O(d|E|)$ to find $d$ eigenpairs. This exceeds the paper’s claim of $O(nd^2)$ preprocessing, and thus requires clarification.\n\nDespite these concerns, the paper’s main contributions remain valid. I would appreciate if the authors could clarify these points during the rebuttal phase.\n\n[1] https://arxiv.org/abs/1905.12265", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730253278756}], "openreview_url": "https://openreview.net/forum?id=xAM9VaXZnY", "arxiv_id": "2406.05815", "paper_pdf": "papers/xAM9VaXZnY.pdf", "paper_pdf_sha256": "afe402430c5f59e297a1a6ce55c6c5d557e2e9e1e976a107cebb7b93a7db4395", "paper_pdf_bytes": 2570169, "paper_pdf_source": "openreview", "code_url": "https://github.com/Graph-COM/GSSC", "code_repository": "Graph-COM/GSSC", "code_commit": "ebb0ce4c52b6904ff95514b68f4a72aa48dfeb0f", "code_archive": "repos/xAM9VaXZnY.zip", "code_archive_sha256": "aa0cd27bdfcc8c56bfd63cacfd2e128bdbea7f031148d1a5476ee379765cffc1", "code_archive_bytes": 189536, "code_file_count": 101, "code_extensions": {".py": 101}, "github_disk_usage_kb": 149, "github_languages": {"Python": 539696}, "github_archived": false, "github_pushed_at": "2024-06-11T15:24:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/what-can-we-learn-from-state-space-models-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XbLffB0T2z", "year": 2024, "status": "rejected", "title": "Transferable Availability Poisoning Attacks", "authors": ["Yiyong Liu", "Michael Backes", "Xiao Zhang"], "authorids": ["~Yiyong_Liu1", "~Michael_Backes3", "~Xiao_Zhang2"], "authors_source": "OpenReview API", "abstract": "We study availability data poisoning attacks, where an adversary aims to degrade the overall test accuracy of a machine learning model by crafting imperceptible perturbations to its training data. Existing strategies can achieve the attack goal but usually assume that the victim employs the same training method as what the adversary uses to mount the attack. In this paper, we argue that this assumption is strong, since the victim may choose any learning algorithm to train the model as long as it achieves some targeted performance on clean data. In addition, we observe a large decrease in the effectiveness of prior poisoning attacks, when the victim uses a different learning paradigm to train the model, and marked differences in frequency-level characteristics between perturbations generated with different learners and attack methods. To enhance the attack transferability, we propose _Transferable Poisoning_, which generates high-frequency poisoning perturbations by alternately leveraging the gradient information with two specific algorithms respectively selected from supervised and unsupervised contrastive learning paradigms. Through extensive experiments on benchmark image datasets, we show that transferable poisoning can produce poisoned samples with significantly improved transferability, which not only applies to the two learners used to devise the attack but also works for learning algorithms and even paradigms beyond.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "iVsxW8ppN4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3507/Reviewer_s6uj"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors address the problem of generating indiscriminate poisoning attacks that can be applied across various learning paradigms, including supervised, semi-supervised, and unsupervised learning. First, they assess the effectiveness of several existing strategies in terms of their transferability. Then, the authors introduce a novel approach called \"Transferable Poisoning\", which combines two strategies to achieve strong transferability across the considered learning paradigms.", "review_text": "The authors address the problem of generating indiscriminate poisoning attacks that can be applied across various learning paradigms, including supervised, semi-supervised, and unsupervised learning. First, they assess the effectiveness of several existing strategies in terms of their transferability. Then, the authors introduce a novel approach called \"Transferable Poisoning\", which combines two strategies to achieve strong transferability across the considered learning paradigms.", "strengths": "The paper provides an interesting contribution to a significant topic, the transferability of poisoning attacks across diverse learning paradigms, an area that has seen limited exploration. The authors introduce a straightforward and intuitive yet effective method, which is tested on a broad range of experiments. The manuscript is well-written overall, though there is space for improvement.", "weaknesses": "I believe this work has two main weaknesses:\n\n1. There is little emphasis on the motivation for studying transferability across learning paradigms, which limits the impact of the findings. It would greatly benefit the paper to explain the practical contexts where this kind of transferability might be relevant, offering concrete examples and detailing how the conducted experiments address these cases. This is particularly crucial given the pragmatic nature of the contribution.\n\n2. While the frequency analysis provides an intriguing perspective on the problem, it could be explored more thoroughly. The authors argue that generating high-frequency perturbations is key to crafting transferable poisoning attacks. This idea is supported by a simple frequency analysis of perturbations generated by existing methods, and so the proposed approach combines two existing ‘high-frequency’ poisoning schemes. Nevertheless, there is no attempt to substantiate this claim with evidence confirming the necessity of high-frequency perturbations for achieving transferability.\n\nIn addition to the above points, I have the following minor concerns:\n\n1. More details should be provided on the considered poisoning methods, the associated learning paradigms, and how the models are tested. Please consider adding this information to both the main text and appendix.\n\n2. The results in the tables lack error bars. \n\n3. The frequency analysis is limited to a simple visualisation of the perturbations’ spectrum. Could you provide a more quantitative comparison? Additionally, please define the spectrum and add a scale to the spectrum plots (ticks are missing).", "questions": "1. Can you insert a frequency constraint on the generated perturbations by modifying the loss function? What happens if you impose high or low-frequency perturbations when using different combinations of attack methods? Could this be a way to test the requirement for transferable perturbations to be high-frequency?\n\n2. In some cases, the proposed algorithm ‘ours’ performs better than the one designed for a given learning paradigm. See for example Table 1, SimSiam and SimCLR: ‘ours’ outperforms ‘CP’. Is there an intuitive explanation for this? Analogously, in Table 6, some results indicate that transferred attacks work better on architectures they were not originally designed for. Error bars would be particularly useful in these cases.\n\n3. In section 4.3 you say that MoCov2 is used both for poison generation and evaluation. Could you please elaborate? Also, what does Framework mean in Table 4?\n\n4. I am not sure I understand the result in Fig. 3, panel C. Is there an explanation for the model performing better when trained on perturbed data? Why only in this case?\n\n5. What is the computational cost of the proposed strategies? A comment on the computational costs should be added to the main text.\n\nComments:\n\n1. Captions could be expanded to include experimental details (what model is attacked, using which dataset, etc) and explain the acronyms.\n\n2. The tables show results for supervised, semi-supervised, and unsupervised algorithms. Shouldn’t SimSiam and SimCLR be grouped together in the tables?\n\n3. I would suggest using an acronym for the proposed method instead of calling it ‘ours’.\n\n4. It would be interesting to have the perturbed images displayed alongside their clean versions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors address the problem of generating indiscriminate poisoning attacks that can be applied across various learning paradigms, including supervised, semi-supervised, and unsupervised learning. First, they assess the effectiveness of several existing strategies in terms of their transferability. Then, the authors introduce a novel approach called \"Transferable Poisoning\", which combines two strategies to achieve strong transferability across the considered learning paradigms.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper provides an interesting contribution to a significant topic, the transferability of poisoning attacks across diverse learning paradigms, an area that has seen limited exploration. The authors introduce a straightforward and intuitive yet effective method, which is tested on a broad range of experiments. The manuscript is well-written overall, though there is space for improvement.", "weaknesses": "I believe this work has two main weaknesses:\n\n1. There is little emphasis on the motivation for studying transferability across learning paradigms, which limits the impact of the findings. It would greatly benefit the paper to explain the practical contexts where this kind of transferability might be relevant, offering concrete examples and detailing how the conducted experiments address these cases. This is particularly crucial given the pragmatic nature of the contribution.\n\n2. While the frequency analysis provides an intriguing perspective on the problem, it could be explored more thoroughly. The authors argue that generating high-frequency perturbations is key to crafting transferable poisoning attacks. This idea is supported by a simple frequency analysis of perturbations generated by existing methods, and so the proposed approach combines two existing ‘high-frequency’ poisoning schemes. Nevertheless, there is no attempt to substantiate this claim with evidence confirming the necessity of high-frequency perturbations for achieving transferability.\n\nIn addition to the above points, I have the following minor concerns:\n\n1. More details should be provided on the considered poisoning methods, the associated learning paradigms, and how the models are tested. Please consider adding this information to both the main text and appendix.\n\n2. The results in the tables lack error bars. \n\n3. The frequency analysis is limited to a simple visualisation of the perturbations’ spectrum. Could you provide a more quantitative comparison? Additionally, please define the spectrum and add a scale to the spectrum plots (ticks are missing).", "questions": "1. Can you insert a frequency constraint on the generated perturbations by modifying the loss function? What happens if you impose high or low-frequency perturbations when using different combinations of attack methods? Could this be a way to test the requirement for transferable perturbations to be high-frequency?\n\n2. In some cases, the proposed algorithm ‘ours’ performs better than the one designed for a given learning paradigm. See for example Table 1, SimSiam and SimCLR: ‘ours’ outperforms ‘CP’. Is there an intuitive explanation for this? Analogously, in Table 6, some results indicate that transferred attacks work better on architectures they were not originally designed for. Error bars would be particularly useful in these cases.\n\n3. In section 4.3 you say that MoCov2 is used both for poison generation and evaluation. Could you please elaborate? Also, what does Framework mean in Table 4?\n\n4. I am not sure I understand the result in Fig. 3, panel C. Is there an explanation for the model performing better when trained on perturbed data? Why only in this case?\n\n5. What is the computational cost of the proposed strategies? A comment on the computational costs should be added to the main text.\n\nComments:\n\n1. Captions could be expanded to include experimental details (what model is attacked, using which dataset, etc) and explain the acronyms.\n\n2. The tables show results for supervised, semi-supervised, and unsupervised algorithms. Shouldn’t SimSiam and SimCLR be grouped together in the tables?\n\n3. I would suggest using an acronym for the proposed method instead of calling it ‘ours’.\n\n4. It would be interesting to have the perturbed images displayed alongside their clean versions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699293548802}, {"id": "Qf3m33eVge", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3507/Reviewer_TTnu"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the transferability of indiscriminate poisoning attacks. The paper first shows that the transferability of pre-computed poisoning samples is low across different learning algorithms. To increase the transferability of such poisoning samples, the paper proposes \"transferable poisoning,\" an algorithm to improve the transferability. The key idea is to consider both supervised and unsupervised learning algorithms simultaneously in crafting, such that both latent representations and logits are optimized to increase the target model's loss during training. The paper runs experiments with three image classification benchmarks and shows that the attack is more transferable than existing baselines.", "review_text": "This paper studies the transferability of indiscriminate poisoning attacks. The paper first shows that the transferability of pre-computed poisoning samples is low across different learning algorithms. To increase the transferability of such poisoning samples, the paper proposes \"transferable poisoning,\" an algorithm to improve the transferability. The key idea is to consider both supervised and unsupervised learning algorithms simultaneously in crafting, such that both latent representations and logits are optimized to increase the target model's loss during training. The paper runs experiments with three image classification benchmarks and shows that the attack is more transferable than existing baselines.", "strengths": "1. The paper proposes an indiscriminate attack with increased transferability.\n2. The paper show the attack's effectiveness empirically.", "weaknesses": "1. The attack is non-practical; it needs the training data to be compromised completely.\n2. The attack (when not 100% training samples are compromised) is not effective in supervised learning algorithms.\n3. The frequency analyses are not scientifically rigorous.\n4. The novelty of this new poisoning attack is weak.\n5. No defense was discussed.\n\nDetailed comments:\n\n\n[Poisoning 100% of the Training Data Is Non-Practical]\n\nI am confident that the paper studies a non-practical scenario: an adversary who poisons 100% of the training data to degrade the accuracy; even with 80% of the training data being compromised, the supervised model retains most of its accuracy compared to clean models.\n\nI haven't seen any real-world scenarios where one allows an adversary to poison 100% of the training data. In any case, the training starts at least with the dataset containing 50% of clean samples.\n\nPrior work on poisoning defenses also theoretically showed that if an adversary can compromise 50% of the training data, there's no guarantee that learning will be empirically successful.\n\nFor these reasons, I also don't think that the transferability should not be measured with the completely compromised datasets.\n\n\n[Not Effective against SL, SupCL, FixMatch]\n\nMy takeaway is that even with the datasets that contain 80--100% of poisoning samples, the three supervised learning algorithms can retain the original accuracy.\n\nThis means (1) the poisoning attack, even with the increased transferability, is weak and easy to defeat, or (2) the three algorithms are designed to be robust inherently to the distributional shifts.\n\nHowever, for this paper, this observation is a weakness as the stronger attack is not actually strong in some settings where a victim does not employ any defenses; the victim could trivially depend on the attack.\n\n\n[Frequency Analyses Are Not Rigorous]\n\nI don't think the paper's claims about the high-frequency components are not scientifically backed by the results. \"High\" means that there will be a quantifiable property, and when used, the proposed poisoning attacks show large numbers. However, the paper is not.\n\nThe paper also claims that the high-frequency characteristics are \"different.\" But it is also not a property that the paper scientifically measures (or compares with criteria). The results are only drawn from the visual analysis. \n\nIt is particularly important to make this claim scientifically rigorous as the reason this poisoning attack is transferable is the paper claims that the perturbations of the attack are the high-frequency ones.\n\n\n[Incremental; Novelty Over the Prior Work]\n\nIt is less surprising that combining the classification- and representation-level losses can lead to general-purpose poisoning samples. Similar techniques have been studied, e.g., one considers all the layer outputs in a neural network for adversarial-example crafting, ensembling multiple models, unrolling the training steps to synthesize effective, transferable targeted poisoning attacks iteratively, etc.\n\n\n[No Defense Evaluation]\n\nSince the attack requires data with 100% poisoning samples, I believe defeating this poisoning attack (or breaking the transferability) is straightforward. I also want to see the discussion about them as a part of responsible vulnerability disclosure.\n\n\nOverall, for those reasons, I am leaning toward rejection.", "questions": "My questions are in the detailed comments in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the transferability of indiscriminate poisoning attacks. The paper first shows that the transferability of pre-computed poisoning samples is low across different learning algorithms. To increase the transferability of such poisoning samples, the paper proposes \"transferable poisoning,\" an algorithm to improve the transferability. The key idea is to consider both supervised and unsupervised learning algorithms simultaneously in crafting, such that both latent representations and logits are optimized to increase the target model's loss during training. The paper runs experiments with three image classification benchmarks and shows that the attack is more transferable than existing baselines.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The paper proposes an indiscriminate attack with increased transferability.\n2. The paper show the attack's effectiveness empirically.", "weaknesses": "1. The attack is non-practical; it needs the training data to be compromised completely.\n2. The attack (when not 100% training samples are compromised) is not effective in supervised learning algorithms.\n3. The frequency analyses are not scientifically rigorous.\n4. The novelty of this new poisoning attack is weak.\n5. No defense was discussed.\n\nDetailed comments:\n\n\n[Poisoning 100% of the Training Data Is Non-Practical]\n\nI am confident that the paper studies a non-practical scenario: an adversary who poisons 100% of the training data to degrade the accuracy; even with 80% of the training data being compromised, the supervised model retains most of its accuracy compared to clean models.\n\nI haven't seen any real-world scenarios where one allows an adversary to poison 100% of the training data. In any case, the training starts at least with the dataset containing 50% of clean samples.\n\nPrior work on poisoning defenses also theoretically showed that if an adversary can compromise 50% of the training data, there's no guarantee that learning will be empirically successful.\n\nFor these reasons, I also don't think that the transferability should not be measured with the completely compromised datasets.\n\n\n[Not Effective against SL, SupCL, FixMatch]\n\nMy takeaway is that even with the datasets that contain 80--100% of poisoning samples, the three supervised learning algorithms can retain the original accuracy.\n\nThis means (1) the poisoning attack, even with the increased transferability, is weak and easy to defeat, or (2) the three algorithms are designed to be robust inherently to the distributional shifts.\n\nHowever, for this paper, this observation is a weakness as the stronger attack is not actually strong in some settings where a victim does not employ any defenses; the victim could trivially depend on the attack.\n\n\n[Frequency Analyses Are Not Rigorous]\n\nI don't think the paper's claims about the high-frequency components are not scientifically backed by the results. \"High\" means that there will be a quantifiable property, and when used, the proposed poisoning attacks show large numbers. However, the paper is not.\n\nThe paper also claims that the high-frequency characteristics are \"different.\" But it is also not a property that the paper scientifically measures (or compares with criteria). The results are only drawn from the visual analysis. \n\nIt is particularly important to make this claim scientifically rigorous as the reason this poisoning attack is transferable is the paper claims that the perturbations of the attack are the high-frequency ones.\n\n\n[Incremental; Novelty Over the Prior Work]\n\nIt is less surprising that combining the classification- and representation-level losses can lead to general-purpose poisoning samples. Similar techniques have been studied, e.g., one considers all the layer outputs in a neural network for adversarial-example crafting, ensembling multiple models, unrolling the training steps to synthesize effective, transferable targeted poisoning attacks iteratively, etc.\n\n\n[No Defense Evaluation]\n\nSince the attack requires data with 100% poisoning samples, I believe defeating this poisoning attack (or breaking the transferability) is straightforward. I also want to see the discussion about them as a part of responsible vulnerability disclosure.\n\n\nOverall, for those reasons, I am leaning toward rejection.", "questions": "My questions are in the detailed comments in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No concern about the ethics", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698914276307}, {"id": "10JKqIWs3s", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3507/Reviewer_uUSA"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper aims to study availability based data poisoning attacks, i.e., the goal of the attacker is to degrade the overall accuracy of a machine learning model trained on a poisoned training dataset. Existing poisoning methods assume the attacker knows the learning paradigm used for training the model, but this assumption is often unrealistic. This work proposed a transferable poisoning attack that crafts perturbations to the dataset that can be transferred to different supervised/unsupervised learning methods. Experimental results show that this poisoning attack outperforms existing poisoning attacks when transferred to almost all learning paradigms.", "review_text": "This paper aims to study availability based data poisoning attacks, i.e., the goal of the attacker is to degrade the overall accuracy of a machine learning model trained on a poisoned training dataset. Existing poisoning methods assume the attacker knows the learning paradigm used for training the model, but this assumption is often unrealistic. This work proposed a transferable poisoning attack that crafts perturbations to the dataset that can be transferred to different supervised/unsupervised learning methods. Experimental results show that this poisoning attack outperforms existing poisoning attacks when transferred to almost all learning paradigms.", "strengths": "The topic of transferable poisoning attacks is very interesting. In the real world, an attacker is unable to know the learning method that would be used to train the model.\n\nExperiments have been done for SOTA learning paradigms in recent years.", "weaknesses": "This attack (and all previous attacks) needs to poison almost 100% of the dataset to perform well, which is unrealistic in practice. Although this setup is aligned with previous works.\n\nAn inconsistency in writing. Section 3.2 mentioned “our method aims to generate poisoning perturbations characterized by high frequency.” But this logic is not reflected in the method design. \n\nThe experiment is not very systematic. Baseline methods are only compared on the CIFAR-10 dataset. In Figure 3, poisoning ratio = 0 (clean accuracy) is not shown.\n\nThis method may be computationally hard to be applied to 224*224 images, e.g., ImageNet Images. But I think this problem also exists in previous works.\n\nThe defenses (e.g., empirical defenses and provable defenses) against poisoning attacks are not discussed. Also, the defenses are not considered. \n\nThe proposed method is a combination of two existing methods. The technique contribution is not strong.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to study availability based data poisoning attacks, i.e., the goal of the attacker is to degrade the overall accuracy of a machine learning model trained on a poisoned training dataset. Existing poisoning methods assume the attacker knows the learning paradigm used for training the model, but this assumption is often unrealistic. This work proposed a transferable poisoning attack that crafts perturbations to the dataset that can be transferred to different supervised/unsupervised learning methods. Experimental results show that this poisoning attack outperforms existing poisoning attacks when transferred to almost all learning paradigms.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The topic of transferable poisoning attacks is very interesting. In the real world, an attacker is unable to know the learning method that would be used to train the model.\n\nExperiments have been done for SOTA learning paradigms in recent years.", "weaknesses": "This attack (and all previous attacks) needs to poison almost 100% of the dataset to perform well, which is unrealistic in practice. Although this setup is aligned with previous works.\n\nAn inconsistency in writing. Section 3.2 mentioned “our method aims to generate poisoning perturbations characterized by high frequency.” But this logic is not reflected in the method design. \n\nThe experiment is not very systematic. Baseline methods are only compared on the CIFAR-10 dataset. In Figure 3, poisoning ratio = 0 (clean accuracy) is not shown.\n\nThis method may be computationally hard to be applied to 224*224 images, e.g., ImageNet Images. But I think this problem also exists in previous works.\n\nThe defenses (e.g., empirical defenses and provable defenses) against poisoning attacks are not discussed. Also, the defenses are not considered. \n\nThe proposed method is a combination of two existing methods. The technique contribution is not strong.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698801661259}, {"id": "noCKlQR3nU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3507/Reviewer_miSr"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper tackles the challenge of availability poisoning attacks in machine learning, where training data is manipulated to degrade model performance. The authors highlight the limited effectiveness of existing attacks when the victim uses a different learning paradigm and introduce Transferable Poisoning (TP), a novel method enhancing attack transferability across various learning algorithms. TP generates high-frequency poisoning perturbations using both supervised and unsupervised contrastive learning paradigms. Extensive experiments on benchmark datasets demonstrate TP's superior performance in ensuring attack effectiveness, regardless of the victim's learning approach.", "review_text": "This paper tackles the challenge of availability poisoning attacks in machine learning, where training data is manipulated to degrade model performance. The authors highlight the limited effectiveness of existing attacks when the victim uses a different learning paradigm and introduce Transferable Poisoning (TP), a novel method enhancing attack transferability across various learning algorithms. TP generates high-frequency poisoning perturbations using both supervised and unsupervised contrastive learning paradigms. Extensive experiments on benchmark datasets demonstrate TP's superior performance in ensuring attack effectiveness, regardless of the victim's learning approach.", "strengths": "The paper introduces Transferable Poisoning (TP) which significantly enhances the transferability of availability poisoning attacks across various learning algorithms and paradigms. This addresses a gap in existing poisoning strategies, which often assume the victim will use the same learning method as the adversary.\nThe authors provide extensive experimental results on benchmark image datasets, demonstrating that TP outperforms existing methods in terms of attack transferability. This thorough validation strengthens the credibility of the proposed method.", "weaknesses": "While TP introduces improvements in transferability, the contribution appears to be incremental. The method essentially combines supervised and unsupervised contrastive learning paradigms to generate high-frequency poisoning perturbations. However, this combination does not seem to bring a substantial novelty or a paradigm shift for poisoning attacks.\n\nThe paper could be strengthened by providing a more solid theoretical foundation for why and how TP improves transferability across learning paradigms. The current explanation relies heavily on empirical observations, which, while valuable, do not provide a comprehensive understanding of the underlying phenomena.", "questions": "Is there a theoretical basis that explains why combining supervised and unsupervised contrastive learning paradigms enhances the transferability of poisoning attacks?\n\nBesides test accuracy degradation, what other metrics (such as robustness, perceptibility of perturbations, and computational efficiency) have been considered to evaluate the effectiveness of TP?\n\n\n\n--after rebuttal--\nThanks for the response. My concerns about novelty remain. Thus, I would maintain my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the challenge of availability poisoning attacks in machine learning, where training data is manipulated to degrade model performance. The authors highlight the limited effectiveness of existing attacks when the victim uses a different learning paradigm and introduce Transferable Poisoning (TP), a novel method enhancing attack transferability across various learning algorithms. TP generates high-frequency poisoning perturbations using both supervised and unsupervised contrastive learning paradigms. Extensive experiments on benchmark datasets demonstrate TP's superior performance in ensuring attack effectiveness, regardless of the victim's learning approach.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper introduces Transferable Poisoning (TP) which significantly enhances the transferability of availability poisoning attacks across various learning algorithms and paradigms. This addresses a gap in existing poisoning strategies, which often assume the victim will use the same learning method as the adversary.\nThe authors provide extensive experimental results on benchmark image datasets, demonstrating that TP outperforms existing methods in terms of attack transferability. This thorough validation strengthens the credibility of the proposed method.", "weaknesses": "While TP introduces improvements in transferability, the contribution appears to be incremental. The method essentially combines supervised and unsupervised contrastive learning paradigms to generate high-frequency poisoning perturbations. However, this combination does not seem to bring a substantial novelty or a paradigm shift for poisoning attacks.\n\nThe paper could be strengthened by providing a more solid theoretical foundation for why and how TP improves transferability across learning paradigms. The current explanation relies heavily on empirical observations, which, while valuable, do not provide a comprehensive understanding of the underlying phenomena.", "questions": "Is there a theoretical basis that explains why combining supervised and unsupervised contrastive learning paradigms enhances the transferability of poisoning attacks?\n\nBesides test accuracy degradation, what other metrics (such as robustness, perceptibility of perturbations, and computational efficiency) have been considered to evaluate the effectiveness of TP?\n\n\n\n--after rebuttal--\nThanks for the response. My concerns about novelty remain. Thus, I would maintain my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698752721649}, {"id": "Gc9w3oqStt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3507/Reviewer_f8DL"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes to iteratively update poisons using TAP and CP and the resulting poisoning attack transfers across SL, CL, Semi-Supervised Learning.", "review_text": "This paper proposes to iteratively update poisons using TAP and CP and the resulting poisoning attack transfers across SL, CL, Semi-Supervised Learning.", "strengths": "1. The motivation is convincing that availability poisoning attack should be effective for different learning paradigms.\n2. The proposed algorithm improves the transferability of poisoning attack across evaluation algorithms including SL, CL, Semi-Supervised Learning on CIFAR-10, CIFAR-100 and Tiny-ImageNet.\n3. This paper provides plenty of investigation on alternative methods including HF, CC, SCP, and WTP.", "weaknesses": "1. The authors should check their code regarding the CL and SupCL evaluation for poisoning attacks, especially how transforms (augmentations) are performed when preparing the poisoned dataset. It seems that the data augmentation for each image is fixed before the CL (SupCL) training starts in their code. For example, in this case, each positive pair consists of two fixed augmented images. However, the CL (SupCL) algorithm applies different data augmentations each time it reads an image. Once this part of code is not properly configured, the reliability of the algorithm is questionable. For these reasons, I have significant concerns about the CL and SupCL evaluation results reported in this paper.\n2. The novelty of the proposed method is limited. The core of the TP algorithm is a just combination of two existing algorithms, i.e. iteratively updating poisons using TAP and CP to inherit their advantages.\n3. The paper seems inaccurately describe the existing work. “Targeted Adversarial Poisoning (Fowl et al., 2021) can produce poisoning perturbations with high-frequency features that are not linear separable, ...” However, as discussed in Table 1 of [a], TAP perturbations are linear separable.\n4. The main machanism of TP is kind of vague. This paper claims that “we propose Transferable Poisoning (TP) to generate poisoning perturbations with shared high-frequency characteristics that can be transferred across different learning paradigms (Section 3).” It is unclear why high-frequency perturbation should be a goal for generating transferable poisoning attacks. The proposed algorithm even does not explicitly optimize the frequency of perturbations. In Figure 4, all four attacks are high-frequency but three of them, namely CC, SCP, and WTP have poor transferability. Does it contradict the claim?\n5. Besides the first concern about CL evaluation, I have also concerns about the comparison of attack performance.\n(1)\tTable 1,2,4 do not consider BYOL that was used in the paper of CP. What is the attack performance of TP against BYOL on CIFAR-10 and CIFAR-100?\n(2)\tIn Table 1, is SimCLR used for CP and TUE perturbation generation? However, SimCLR is not surely the best generation algorithm for transferability of CP and TUE. To be a fair comparison, it is helpful to consider CP and TUE attacks based on other CL algorithms as well as class-wise CP attacks which are effective for SL.\n(3)\tIt would be better to compare your TP with existing powerful poisoning methods (UE, AP, CP and TUE) on larger datasets like CIFAR-100 and TinyImageNet.\n6. As the algorithm involves contrastive training, it possibly costs too much time to generate TP perturbations, especially on Tiny-ImageNet and ImageNet. Can you provide time consuming comparison of your TP with existing transferable methods like AP, (class-wise) CP, and TUE? Considering availability poisoning attacks are a sort of data protection means, low efficiency of generation might hinder the applications in real world scenarios.\n\n[a] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts.\n\nClarity, Quality, Novelty And Reproducibility:\nI hope the author would make more clarificaitons about the main machecism of TP and descriptions about existing work. The novelty of proposed algorithm is hindered by the combination of two existing algorithms. For reproducibility, the first concern about CL and SupCL evaluation is vital.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to iteratively update poisons using TAP and CP and the resulting poisoning attack transfers across SL, CL, Semi-Supervised Learning.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The motivation is convincing that availability poisoning attack should be effective for different learning paradigms.\n2. The proposed algorithm improves the transferability of poisoning attack across evaluation algorithms including SL, CL, Semi-Supervised Learning on CIFAR-10, CIFAR-100 and Tiny-ImageNet.\n3. This paper provides plenty of investigation on alternative methods including HF, CC, SCP, and WTP.", "weaknesses": "1. The authors should check their code regarding the CL and SupCL evaluation for poisoning attacks, especially how transforms (augmentations) are performed when preparing the poisoned dataset. It seems that the data augmentation for each image is fixed before the CL (SupCL) training starts in their code. For example, in this case, each positive pair consists of two fixed augmented images. However, the CL (SupCL) algorithm applies different data augmentations each time it reads an image. Once this part of code is not properly configured, the reliability of the algorithm is questionable. For these reasons, I have significant concerns about the CL and SupCL evaluation results reported in this paper.\n2. The novelty of the proposed method is limited. The core of the TP algorithm is a just combination of two existing algorithms, i.e. iteratively updating poisons using TAP and CP to inherit their advantages.\n3. The paper seems inaccurately describe the existing work. “Targeted Adversarial Poisoning (Fowl et al., 2021) can produce poisoning perturbations with high-frequency features that are not linear separable, ...” However, as discussed in Table 1 of [a], TAP perturbations are linear separable.\n4. The main machanism of TP is kind of vague. This paper claims that “we propose Transferable Poisoning (TP) to generate poisoning perturbations with shared high-frequency characteristics that can be transferred across different learning paradigms (Section 3).” It is unclear why high-frequency perturbation should be a goal for generating transferable poisoning attacks. The proposed algorithm even does not explicitly optimize the frequency of perturbations. In Figure 4, all four attacks are high-frequency but three of them, namely CC, SCP, and WTP have poor transferability. Does it contradict the claim?\n5. Besides the first concern about CL evaluation, I have also concerns about the comparison of attack performance.\n(1)\tTable 1,2,4 do not consider BYOL that was used in the paper of CP. What is the attack performance of TP against BYOL on CIFAR-10 and CIFAR-100?\n(2)\tIn Table 1, is SimCLR used for CP and TUE perturbation generation? However, SimCLR is not surely the best generation algorithm for transferability of CP and TUE. To be a fair comparison, it is helpful to consider CP and TUE attacks based on other CL algorithms as well as class-wise CP attacks which are effective for SL.\n(3)\tIt would be better to compare your TP with existing powerful poisoning methods (UE, AP, CP and TUE) on larger datasets like CIFAR-100 and TinyImageNet.\n6. As the algorithm involves contrastive training, it possibly costs too much time to generate TP perturbations, especially on Tiny-ImageNet and ImageNet. Can you provide time consuming comparison of your TP with existing transferable methods like AP, (class-wise) CP, and TUE? Considering availability poisoning attacks are a sort of data protection means, low efficiency of generation might hinder the applications in real world scenarios.\n\n[a] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts.\n\nClarity, Quality, Novelty And Reproducibility:\nI hope the author would make more clarificaitons about the main machecism of TP and descriptions about existing work. The novelty of proposed algorithm is hindered by the combination of two existing algorithms. For reproducibility, the first concern about CL and SupCL evaluation is vital.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698676309007}], "openreview_url": "https://openreview.net/forum?id=XbLffB0T2z", "arxiv_id": "2310.05141", "paper_pdf": "papers/XbLffB0T2z.pdf", "paper_pdf_sha256": "2f0bebb474ddba5509c11e0a75967c123478e6d32be1f4d14e90bd3721b6f7c0", "paper_pdf_bytes": 1121599, "paper_pdf_source": "openreview", "code_url": "https://github.com/TrustMLRG/TransPoison", "code_repository": "TrustMLRG/TransPoison", "code_commit": "2d08a9c0b10ff5da332b3c0b800c10c6dca8c518", "code_archive": "repos/XbLffB0T2z.zip", "code_archive_sha256": "b41d22510b8378efa95f5caa5ca257f1217b01d1bf27c59a61a65159eb2d4584", "code_archive_bytes": 138601, "code_file_count": 59, "code_extensions": {".py": 59}, "github_disk_usage_kb": 116, "github_languages": {"Python": 310559}, "github_archived": false, "github_pushed_at": "2023-10-11T09:29:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transferable-availability-poisoning-attacks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Z-CqSH6J_VK", "year": 2023, "status": "rejected", "title": "Differentiable and transportable structure learning", "authors": ["Jeroen Berrevoets", "Nabeel Seedat", "Fergus Imrie", "Mihaela van der Schaar"], "authorids": ["~Jeroen_Berrevoets1", "~Nabeel_Seedat1", "~Fergus_Imrie1", "~Mihaela_van_der_Schaar2"], "authors_source": "OpenReview API", "abstract": "Directed acyclic graphs (DAGs) encode a lot of information about a particular distribution in its structure. However, compute required to infer these structures is typically super-exponential in the number of variables, as inference requires a sweep of a combinatorially large space of potential structures. That is, until recent advances made it possible to search this space using a differentiable metric, drastically reducing search time. While this technique– named NOTEARS –is widely considered a seminal work in DAG-discovery, it concedes an important property in favour of differentiability: transportability. To be transportable, the structures discovered on one dataset must apply to another dataset from the same domain. In our paper, we introduce D-Struct which recovers transportability in the discovered structures through a novel architecture and loss function, while remaining completely differentiable. Because D-Struct remains differentiable, our method can be easily adopted in existing differentiable architectures, as was previously done with NOTEARS. In our experiments, we empirically validate D-Struct with respect to edge accuracy and structural Hamming distance in a variety of settings.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "9AHrqsqLdn", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4326/Reviewer_CF11"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes transportability in DAG structure learning problem, which can be seen as a multi-task DAG learning problem or a cross validation method from a single task. The changes from NOTEARS methods is the addition from an average graph loss term additional to the differential scoring function. The method is compared with one baseline NOTEARS to show its superior performance in accuracy.", "review_text": "the paper addressed a new problem in DAG learning but with straightforward technical contribution. ", "strengths": "Strength:\nIt seems this is a first work to learn one single DAG from multiple domain.\nSubset construction is interesting, as it may improve the statistical property of the estimation. \n\nWeakness:\nThe technical contribution is rather limited, with a MSE terms on graphs. It would be interesting for authors to discuss the choice of such a regularization against potential other alternatives. \nOnly one baseline is compared. Understandably this may be first work, but other works could and should be adapted as a baseline. For example, the multi-task DAG learning from [62].\n\nComments:\n- subset construction: it is not fully clear how much the performance gain is from d-struct formulation, compared against an ensembled approach on different subset of data (maybe taking an average to obtain the final result with measure to ensure acyclicity).  This could also serve as a baseline.\n- While transportability definition is clear, authors did not discuss much about it with respect to other standard problem regimes. For example, how is transportability problem different from a typical multi-task learning or transfer learning for DAGs?\n- another way to improve technical contribution is to study feature differences or absences in different domains, and how to integrate them into one single DAG. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes transportability in DAG structure learning problem, which can be seen as a multi-task DAG learning problem or a cross validation method from a single task. The changes from NOTEARS methods is the addition from an average graph loss term additional to the differential scoring function. The method is compared with one baseline NOTEARS to show its superior performance in accuracy.", "strength_and_weaknesses": "Strength:\nIt seems this is a first work to learn one single DAG from multiple domain.\nSubset construction is interesting, as it may improve the statistical property of the estimation. \n\nWeakness:\nThe technical contribution is rather limited, with a MSE terms on graphs. It would be interesting for authors to discuss the choice of such a regularization against potential other alternatives. \nOnly one baseline is compared. Understandably this may be first work, but other works could and should be adapted as a baseline. For example, the multi-task DAG learning from [62].\n\nComments:\n- subset construction: it is not fully clear how much the performance gain is from d-struct formulation, compared against an ensembled approach on different subset of data (maybe taking an average to obtain the final result with measure to ensure acyclicity).  This could also serve as a baseline.\n- While transportability definition is clear, authors did not discuss much about it with respect to other standard problem regimes. For example, how is transportability problem different from a typical multi-task learning or transfer learning for DAGs?\n- another way to improve technical contribution is to study feature differences or absences in different domains, and how to integrate them into one single DAG. ", "clarity,_quality,_novelty_and_reproducibility": "clarity: good\n\nquality: good\n\noriginality: problem is new, but technically it is limited. ", "summary_of_the_review": "the paper addressed a new problem in DAG learning but with straightforward technical contribution. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667208591612}, {"id": "523jpitxPm6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4326/Reviewer_pcus"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the problem of learning DAGs from observational data. In contrast to previous work under the continuous framework, the authors aim to learn a DAG that is transportable to other distributions over the same observables that respect a set of conditional independences. In a few words, the approach consists of learning K different DAGs where each DAG is penalized by how far away it is from the mean of the K DAGs, in this way, the approach remains fully differentiable and the resulting DAGs are somehow forced to be similar or equal for the K different datasets. The authors apply a preprocessing sampling step to make the method work even for the case of observing a single dataset. Some experiments are provided to demonstrate the efficacy of the proposed method against the nonparametric NOTEARS algorithm.", "review_text": "My main concerns are listed above which currently make me inclined toward rejection.", "strengths": "### Strengths\n\nThe core ideas of the paper are well-written and well-motivated. The authors do a good job of stating the problem under study and the contributions of their work. The method is simple to grasp and simple to adapt to existing differentiable approaches to learning DAGs.\n\n### Weaknesses\n\nThe main weakness is the experimental section. \n\n* Lack of proper comparison to other methods. First, in my opinion, the correct baseline should not be plain NOTEARS but the version of NOTEARS with the same exact sampling mechanism for Dstruct and taking the average or median as the final output, i.e., the only difference to Dstruct should be the L_MSE step. I stress this point because we need to make sure that (part of) the improvements are not due to the ensemble mechanism (which would be a very incremental contribution) but to the addition of the L_MSE regularization. Second, there really should be comparisons to other methods, especially Bayesian methods such as DiBS [1]. This is because in those methods we can sample DAGs from a posterior distribution and, again, perhaps taking the mode or mean would result in a transportable DAG. For these Bayesian methods, I would feed all the data---even from different distributions---and see if they are robust to these changes in distributions. \n\n* The speed-up claims are rather disappointing, the graphs have a very small number of variables, $d=5$. To make a more precise claim about computational speed-ups there must be experiments for different values of d and not just n. This is because increasing $n$ really just makes the training of the neural nets longer but does not increase the complexity of enforcing acyclicity, for which $d$ does. Recent work [2] for instance ran the NOTEARS-MLP for up to 100 nodes leveraging a new acyclicity function. \n\n[1] Lorch et al. (2021), \"DiBS: Differentiable Bayesian Structure Learning\"\n[2] Bello et al. (2022), “DAGMA: Learning DAGs via M-matrices and a Log-Determinant Acyclicity Characterization”\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper studies the problem of learning DAGs from observational data. In contrast to previous work under the continuous framework, the authors aim to learn a DAG that is transportable to other distributions over the same observables that respect a set of conditional independences. In a few words, the approach consists of learning K different DAGs where each DAG is penalized by how far away it is from the mean of the K DAGs, in this way, the approach remains fully differentiable and the resulting DAGs are somehow forced to be similar or equal for the K different datasets. The authors apply a preprocessing sampling step to make the method work even for the case of observing a single dataset. Some experiments are provided to demonstrate the efficacy of the proposed method against the nonparametric NOTEARS algorithm.", "strength_and_weaknesses": "### Strengths\n\nThe core ideas of the paper are well-written and well-motivated. The authors do a good job of stating the problem under study and the contributions of their work. The method is simple to grasp and simple to adapt to existing differentiable approaches to learning DAGs.\n\n### Weaknesses\n\nThe main weakness is the experimental section. \n\n* Lack of proper comparison to other methods. First, in my opinion, the correct baseline should not be plain NOTEARS but the version of NOTEARS with the same exact sampling mechanism for Dstruct and taking the average or median as the final output, i.e., the only difference to Dstruct should be the L_MSE step. I stress this point because we need to make sure that (part of) the improvements are not due to the ensemble mechanism (which would be a very incremental contribution) but to the addition of the L_MSE regularization. Second, there really should be comparisons to other methods, especially Bayesian methods such as DiBS [1]. This is because in those methods we can sample DAGs from a posterior distribution and, again, perhaps taking the mode or mean would result in a transportable DAG. For these Bayesian methods, I would feed all the data---even from different distributions---and see if they are robust to these changes in distributions. \n\n* The speed-up claims are rather disappointing, the graphs have a very small number of variables, $d=5$. To make a more precise claim about computational speed-ups there must be experiments for different values of d and not just n. This is because increasing $n$ really just makes the training of the neural nets longer but does not increase the complexity of enforcing acyclicity, for which $d$ does. Recent work [2] for instance ran the NOTEARS-MLP for up to 100 nodes leveraging a new acyclicity function. \n\n[1] Lorch et al. (2021), \"DiBS: Differentiable Bayesian Structure Learning\"\n[2] Bello et al. (2022), “DAGMA: Learning DAGs via M-matrices and a Log-Determinant Acyclicity Characterization”\n", "clarity,_quality,_novelty_and_reproducibility": "The writing is a bit sloppy in some places. \n\n* The objective in equation (2) is many times referred to as the \"score\", which can be misleading since the score in a score-based method is simply what the authors write as $F(A)$ or the \"regularized\" score $F(A) + \\lambda_1 * ||A||_1$ algorithm. The other terms w.r.t. $h$ stem from the augmented Lagrangian to solve the constrained problem. \n\n* In Section 3.1., the authors also mention $K$ \"distinct\" DSFs. I believe this *cannot* be the case. For instance, consider linear SEMs with equal variances, from Loh & Buhlmann (2014) we know that the ground-truth DAG is the global minimizer of the least squares (LS) score, i.e., these are identifiable models. Now, suppose one uses LS for dataset D1 and another score, call this S2, for another dataset D2. Moreover, suppose we are able to find the global minimizes in each case. Then, using LS on D1 already finds the correct DAG, while using S2 on D2 will find an incorrect DAG. Forcing both DAGs to be the same, as in Dstruct, will fail to find the correct DAG. Thus, to me, it does not make sense to use \"distinct\" DSFs.\n\n* Algorithms 1 and 2 are a bit confusing, Alg 1 does not use h_tol and Alg 2 does not have rho_max as input. Please polish these minor details.\n\n* In the paragraph on transportability on Page 8, what do \"internal\" graphs mean?\n\nAs per novelty, I like the idea of trying to find a transportable DAG, but I am unclear if the approach really does make sense due to my questions above. I am under the impression that really the core technical contribution is simply adding an average of DAGs as a measure to force the different DAGs to be the same, which looks like a fine contribution but perhaps not enough to grant publication at this time.\n\nI don't have many concerns about quality and reproducibility.\n\n\n\n", "summary_of_the_review": "My main concerns are listed above which currently make me inclined toward rejection.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666855474155}, {"id": "Rm26ulXgUqz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4326/Reviewer_auHu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors consider the problem of learning DAG from heterogeneous datasets that share the same DAG. This problem statement is quite restricted. From Eq(4) we can see that the authors not only encourage the DAG to have the same structure, but also to have the same parameter. While in Figure 1, the authors assume that multiple datasets to have different distribution. In this case, there is only one possibility, that is the distribution of noise variable may be different. In the NOTEARS framework, the only possibility is that the Gaussian noise may have different mean and variance in different datasets. This setting is in fact quite restricted. ", "review_text": "The paper proposed a method to learn DAG from heterogeneous datasets, but the setting in the paper is quite restricted and the experiments are also weak.", "strengths": "Weakness: \n\n1. The problem setting is quite restricted.\n2. There is no theoretic guarantee that the proposed algorithm will find a set of consistency DAG.\n3. The experiment is quite weak. Instead of only using nonlinear models, the authors may also add some linear experiments, where the DAG is known to be identifiable. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors consider the problem of learning DAG from heterogeneous datasets that share the same DAG. This problem statement is quite restricted. From Eq(4) we can see that the authors not only encourage the DAG to have the same structure, but also to have the same parameter. While in Figure 1, the authors assume that multiple datasets to have different distribution. In this case, there is only one possibility, that is the distribution of noise variable may be different. In the NOTEARS framework, the only possibility is that the Gaussian noise may have different mean and variance in different datasets. This setting is in fact quite restricted. ", "strength_and_weaknesses": "Weakness: \n\n1. The problem setting is quite restricted.\n2. There is no theoretic guarantee that the proposed algorithm will find a set of consistency DAG.\n3. The experiment is quite weak. Instead of only using nonlinear models, the authors may also add some linear experiments, where the DAG is known to be identifiable. ", "clarity,_quality,_novelty_and_reproducibility": "The problem statement is not very clear, and the novelty is quite limited. ", "summary_of_the_review": "The paper proposed a method to learn DAG from heterogeneous datasets, but the setting in the paper is quite restricted and the experiments are also weak.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666697764128}, {"id": "coZxMYvgT0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4326/Reviewer_nxCL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper develops an extension of Notears, a recent approach for structure learning with continuous optimization, to learn transportable structures (transportable in the sense that the learned structures are the same across multiple datasets with potentially different distributions). This is achieved by incorporating a regularization term in the optimization procedure to encourage the structure parameters over different dataset to be equal to one another. Empirical results are provided to validate the proposed method.", "review_text": "- A number of highly related works were not discussed and compared to. Specifically, the proposed method is similar to [1, 2].\n- The proposed method adds a regularization term w.r.t. both structure and model parameters, which may lead to the wrong structure in practice especially when the data distributions are different for different datasets.\n- The paper does not consider experiments of different data distributions for different datasets, which may be more common in practice and is the key motivation of transportability/invariance.", "strengths": "### Strengths\n- The problem studied is highly relevant, since learning transportable structures may be an importance topic especially because of data heterogeneity across different datasets.\n- The paper is well written and easy to follow.\n\n### Weaknesses\n- The setting studied is very similar to (and seems to be a special case of) differentiable federated structure learning [1, 2], which should be discussed. E.g., [1] considered data heterogeneity difference (i.e. different data distribution across multiple clients, but the underlying DAG is the same), which is essentially the same as transportability in this paper. If I understand correctly, the only key difference appears to be that [1, 2] additionally considered privacy issue.\n- Therefore, the resulting approach is also similar to [1, 2]. Specifically, [1] compute the average adjacency matrix in each step and enforces different clients to use the same average matrix, while [2] enforces the adjacency of different clients to be the same via ADMM. The only key difference appears to be that the proposed approach uses a soft regularization scheme to *encourage* the same structure, while [1, 2] used a hard scheme to *enforce* the same structure.\n- Following the comment above, although adding a soft regularization scheme encourages the same structure, it does not guarantee/enforce it, which I think might be a limitation (in practice the participating clients may want to learn a shared structure that is the same). This is also acknowledged by the experiment in Sec. 4.\n- Adding a soft regularization directly over the parameters $A$ (which correspond to both structure and model parameters) may lead to the wrong structure. I.e., this regularization may conflict with the original least squares loss of the specific dataset when the data distributions are different. In my opinion, the backbone method Mcsl used by [1] might seem like a better choice (as compared to Notears-MLP) because it decomposes the structure and model parameters, and so the proposed regularization can be applied w.r.t. the structure parameters only.\n- I did not manage to find experiments of different datasets having different data distributions in Sec. 4, which I think are much more relevant (considered by [1]), since that's the key motivation of using transportability (Definition 1) or invariance.\n- The considered setting is highly similar to (invariant) causal discovery from multiple domains or heterogenous data, e.g., [3, 4, 5], which should be discussed and compared.\n\n### Minor/Other Comments\n- Sec 2.2: $\\rho$ and $\\lambda_2$ are not \"hyperparamers\", but rather the parameters of augmented Lagrangian method that will be updated during the optimization process. \n- Sec. 2.2 \"One can traverse G_X smartly to arrive at a DAG much faster\": There was no guarantee that DSF arrives at a DAG in (Zheng et al, 2018); this was proved by [6, 7] and should probably be mentioned. Also, to the best of my knowledge, Zheng et al. (2018) did not claim/demonstrate that Notears runs \"much faster\" than CIT-based methods like PC, and I would encourage the authors to add a reference regarding this.\n- For Notears-MLP, is the regularization scheme applied to the weighted adjacency matrix constructed w.r.t. the MLP weights, or all weights of the MLPs? Alg. 2 seems to indicate the latter, but the former sounds more intuitive to me.\n- The efficiency gain in Fig. 5 is interesting and surprising. Is it because the for loop in fifth line of Alg. 1 can be run in parallel?\n\n1. Federated causal discovery, 2022.\n2. Towards Federated Bayesian Network Structure Learning with Continuous Optimization, 2022.\n3. Multi-domain Causal Structure Learning in Linear Systems, 2018.\n4. Causal inference using invariant prediction: identification and confidence intervals, 2015.\n5. Causal Discovery from Heterogeneous/Nonstationary Data, 2020.\n6. On the Convergence of Continuous Constrained Optimization for Structure Learning, 2022.\n7. DAGs with No Fears: A Closer Look at Continuous Optimization for Learning Bayesian Networks, 2020.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper develops an extension of Notears, a recent approach for structure learning with continuous optimization, to learn transportable structures (transportable in the sense that the learned structures are the same across multiple datasets with potentially different distributions). This is achieved by incorporating a regularization term in the optimization procedure to encourage the structure parameters over different dataset to be equal to one another. Empirical results are provided to validate the proposed method.", "strength_and_weaknesses": "### Strengths\n- The problem studied is highly relevant, since learning transportable structures may be an importance topic especially because of data heterogeneity across different datasets.\n- The paper is well written and easy to follow.\n\n### Weaknesses\n- The setting studied is very similar to (and seems to be a special case of) differentiable federated structure learning [1, 2], which should be discussed. E.g., [1] considered data heterogeneity difference (i.e. different data distribution across multiple clients, but the underlying DAG is the same), which is essentially the same as transportability in this paper. If I understand correctly, the only key difference appears to be that [1, 2] additionally considered privacy issue.\n- Therefore, the resulting approach is also similar to [1, 2]. Specifically, [1] compute the average adjacency matrix in each step and enforces different clients to use the same average matrix, while [2] enforces the adjacency of different clients to be the same via ADMM. The only key difference appears to be that the proposed approach uses a soft regularization scheme to *encourage* the same structure, while [1, 2] used a hard scheme to *enforce* the same structure.\n- Following the comment above, although adding a soft regularization scheme encourages the same structure, it does not guarantee/enforce it, which I think might be a limitation (in practice the participating clients may want to learn a shared structure that is the same). This is also acknowledged by the experiment in Sec. 4.\n- Adding a soft regularization directly over the parameters $A$ (which correspond to both structure and model parameters) may lead to the wrong structure. I.e., this regularization may conflict with the original least squares loss of the specific dataset when the data distributions are different. In my opinion, the backbone method Mcsl used by [1] might seem like a better choice (as compared to Notears-MLP) because it decomposes the structure and model parameters, and so the proposed regularization can be applied w.r.t. the structure parameters only.\n- I did not manage to find experiments of different datasets having different data distributions in Sec. 4, which I think are much more relevant (considered by [1]), since that's the key motivation of using transportability (Definition 1) or invariance.\n- The considered setting is highly similar to (invariant) causal discovery from multiple domains or heterogenous data, e.g., [3, 4, 5], which should be discussed and compared.\n\n### Minor/Other Comments\n- Sec 2.2: $\\rho$ and $\\lambda_2$ are not \"hyperparamers\", but rather the parameters of augmented Lagrangian method that will be updated during the optimization process. \n- Sec. 2.2 \"One can traverse G_X smartly to arrive at a DAG much faster\": There was no guarantee that DSF arrives at a DAG in (Zheng et al, 2018); this was proved by [6, 7] and should probably be mentioned. Also, to the best of my knowledge, Zheng et al. (2018) did not claim/demonstrate that Notears runs \"much faster\" than CIT-based methods like PC, and I would encourage the authors to add a reference regarding this.\n- For Notears-MLP, is the regularization scheme applied to the weighted adjacency matrix constructed w.r.t. the MLP weights, or all weights of the MLPs? Alg. 2 seems to indicate the latter, but the former sounds more intuitive to me.\n- The efficiency gain in Fig. 5 is interesting and surprising. Is it because the for loop in fifth line of Alg. 1 can be run in parallel?\n\n1. Federated causal discovery, 2022.\n2. Towards Federated Bayesian Network Structure Learning with Continuous Optimization, 2022.\n3. Multi-domain Causal Structure Learning in Linear Systems, 2018.\n4. Causal inference using invariant prediction: identification and confidence intervals, 2015.\n5. Causal Discovery from Heterogeneous/Nonstationary Data, 2020.\n6. On the Convergence of Continuous Constrained Optimization for Structure Learning, 2022.\n7. DAGs with No Fears: A Closer Look at Continuous Optimization for Learning Bayesian Networks, 2020.", "clarity,_quality,_novelty_and_reproducibility": "Details are provided in the comments above.", "summary_of_the_review": "- A number of highly related works were not discussed and compared to. Specifically, the proposed method is similar to [1, 2].\n- The proposed method adds a regularization term w.r.t. both structure and model parameters, which may lead to the wrong structure in practice especially when the data distributions are different for different datasets.\n- The paper does not consider experiments of different data distributions for different datasets, which may be more common in practice and is the key motivation of transportability/invariance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666651302654}], "openreview_url": "https://openreview.net/forum?id=Z-CqSH6J_VK", "arxiv_id": "2206.06354", "paper_pdf": "papers/Z-CqSH6J_VK.pdf", "paper_pdf_sha256": "f7faf8e2820d7794311bb16928bd4314bc338569e11ff3668b0aaa6573169e7f", "paper_pdf_bytes": 634533, "paper_pdf_source": "openreview", "code_url": "https://github.com/jeroenbe/d-struct", "code_repository": "jeroenbe/d-struct", "code_commit": "4fda078cb7015fd8fe2a9ad366ea93b36664ac5b", "code_archive": "repos/Z-CqSH6J_VK.zip", "code_archive_sha256": "e8e069deee3dfd27af0ea592bfce2bfbf6af69b53e588b8f398912ddac05e401", "code_archive_bytes": 125889, "code_file_count": 17, "code_extensions": {".sh": 9, ".py": 8}, "github_disk_usage_kb": 214, "github_languages": {"Python": 66972, "Shell": 18511}, "github_archived": false, "github_pushed_at": "2023-07-26T08:37:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/differentiable-and-transportable-structure"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qDx6DXD3Fzt", "year": 2022, "status": "rejected", "title": "Provably Robust Detection of Out-of-distribution Data (almost) for free", "authors": ["Alexander Meinke", "Julian Bitterwolf", "Matthias Hein"], "authorids": ["~Alexander_Meinke1", "~Julian_Bitterwolf1", "~Matthias_Hein2"], "authors_source": "OpenReview API", "abstract": "The application of machine learning in safety-critical systems requires a reliable assessment of uncertainy. However, deep neural networks are known to produce highly overconfident predictions on out-of-distribution (OOD) data. Even if trained to be non-confident on OOD data one can still adversarially manipulate OOD data so that the classifier again assigns high confidence to the manipulated samples. In this paper we propose a novel method that combines a certifiable OOD detector with a standard classifier from first principles into an OOD aware classifier. This way we achieve the best of two worlds: certifiably adversarially robust OOD detection, even for OOD samples close to the in-distribution, without loss in either prediction accuracy or detection performance for non-manipulated OOD data. Moreover, due to the particular construction our classifier provably avoids the asymptotic overconfidence problem of standard neural networks.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "tJQimu_Jzlg", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4054/Reviewer_urW9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents an approach to combine OOD detection and classifier to develop a robust OOD detection framework with high classification accuracy. The authors provide confidence bounds for the noise-perturbed and adversarially attacked OOD samples. Experiments are conducted on benchmark datasets and the performance is compared with approaches that provide asymptotic guarantees on robustness. ", "review_text": "Strengths:\n\n1. Combining OOD detection and classifiers is an interesting idea and the proposed \"OOD-aware\" training can be effective to improve the robustness of the classifiers.\n\n2. Authors provide a confidence bound for the classifier which is useful to develop certifiable robust OOD-aware models.\n\n3. Experiments on the benchmark datasets show that the proposed approach achieves comparatively high accuracy while maintaining a guarantee on the AUC score. \n\n\nWeaknesses:\n\n1. The draft lacks clarity in many aspects. \n\na. In table 1, metrics are not clear until section 3.\n\nb. Figure 1 cannot be interpreted from the caption. It is not clear how the equations in the figure map to equations in the draft. \n\nc. In the experiments section, \"semi-joint training\" requires more details. For example, how are the several models trained with binary shifts? Metrics in figure 2 are not clear from the associated text.\n\n2.  Authors claim to achieve better guarantees than the existing approaches (e.g., CCU, GOOD). However, it is not clear why Theorem 1 entails a tighter bound than Lemma 1 (Hein 2019).\n\n3. The bounds in Eq 5 appear trivial from the computations of W_{+}, and W_{-}. How do these bounds affect the tightness of the bounds in Eq 6.\n\n4. I would expect the joint training would improve the performance of both OOD detection and the classifier. However, as shown in table 2, outlier exposure, which is not the best state of the art, still performs better. What are the authors' comments on this. \n\n5. Authors are encouraged to compare with the more recent OOD detection approaches \n\na. Liu et al., \"Energy-based out-of-distribution detection\"\n\nb. Lee et al., \"Training confidence-calibrated classifiers for detecting out-of-distribution samples.\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents an approach to combine OOD detection and classifier to develop a robust OOD detection framework with high classification accuracy. The authors provide confidence bounds for the noise-perturbed and adversarially attacked OOD samples. Experiments are conducted on benchmark datasets and the performance is compared with approaches that provide asymptotic guarantees on robustness. ", "main_review": "Strengths:\n\n1. Combining OOD detection and classifiers is an interesting idea and the proposed \"OOD-aware\" training can be effective to improve the robustness of the classifiers.\n\n2. Authors provide a confidence bound for the classifier which is useful to develop certifiable robust OOD-aware models.\n\n3. Experiments on the benchmark datasets show that the proposed approach achieves comparatively high accuracy while maintaining a guarantee on the AUC score. \n\n\nWeaknesses:\n\n1. The draft lacks clarity in many aspects. \n\na. In table 1, metrics are not clear until section 3.\n\nb. Figure 1 cannot be interpreted from the caption. It is not clear how the equations in the figure map to equations in the draft. \n\nc. In the experiments section, \"semi-joint training\" requires more details. For example, how are the several models trained with binary shifts? Metrics in figure 2 are not clear from the associated text.\n\n2.  Authors claim to achieve better guarantees than the existing approaches (e.g., CCU, GOOD). However, it is not clear why Theorem 1 entails a tighter bound than Lemma 1 (Hein 2019).\n\n3. The bounds in Eq 5 appear trivial from the computations of W_{+}, and W_{-}. How do these bounds affect the tightness of the bounds in Eq 6.\n\n4. I would expect the joint training would improve the performance of both OOD detection and the classifier. However, as shown in table 2, outlier exposure, which is not the best state of the art, still performs better. What are the authors' comments on this. \n\n5. Authors are encouraged to compare with the more recent OOD detection approaches \n\na. Liu et al., \"Energy-based out-of-distribution detection\"\n\nb. Lee et al., \"Training confidence-calibrated classifiers for detecting out-of-distribution samples.\"", "summary_of_the_review": "The proposed approach of combining OOD detection and classifier through joint training is an interesting approach. However, some parts of the paper lack clarity and thus, it is hard to evaluate the contributions. Authors are encouraged to address the comments and I will be willing to reconsider my decision.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635929209896}, {"id": "KBuNgoPJH6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4054/Reviewer_U32o"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper aims to detect out-of-distribution data in an adversarially robust manner. To this end, the authors incorporate a certified (binary) classifier to model in-and-out distribution and jointly model predictive distribution $p(y|x)$ by conditioning with the binary classifier, i.e., $p(y|x, in)p(in|x) + p(y|x,out)p(out|x)$. The authors show that the proposed method is empirically strong under various detection scenarios.", "review_text": "**Strength**\\\nThe proposed method is sensible and shows consistent improvements.\\\nThe paper provides experiments on several detection benchmarks.\n\n**Weakness**\\\n*Limited evaluation*\\\nDespite the fact that this scheme utilizes a certified classifier, the author evaluates with empirical robustness measures (to calculate the maximum perturbation). Rather than that, I suggest utilizing a certified robustness measure with $l_\\infty$ extension (Zhang et al., 2021).\nAdditionally, considering AutoAttack (Croce et al., 2020) to attack the confidence $\\max p(y|x)$ will be more convincing (as an empirical robustness measure): calculate the maximum confidence over the ensemble attacks in AutoAttack.
\n\n*Discussion with Tramer et al., 2021*\\\nTramer et al., 2021 prove that detecting adversarial samples is hard as classification (Tramer et al., 2021), i.e., classifying $\\epsilon$ is almost the same as detecting $2*\\epsilon$ sample. Due to this paper, it was hard for me to believe that adversarial detection is (almost) free: as it is known to sacrifice the clean accuracy to obtain adversarial robustness (Zhang et al., 2019). Can the author rigorously discuss with the following paper (Tramer et al., 2021)?\n\n*Limited technical novelty*\n* The proposed method can be seen as a combination of two OOD methods (Hsu et al., 2020) and (Bitterwolf et al., 2020).\n* The main technical novelty of this paper is to (1) model predictive distribution with in-and-out conditional distribution (2) utilize IBP for modeling in-and-out distribution to model certification of OOD robustness.\n* However, I believe (1) can be found in (Hsu et al., 2020), and (2) corresponds to (Bitterwolf et al., 2020). \n\n*The AUC of ProoD-disc seems to be low*. This implies that the discriminator does not capture the in-and-out distribution probability well.\n\n(minor) *The writing and presentation can be improved*\n* For instance, a single paragraph contains many messages, making the readers confused about the main message.
\n\n**Questions**\\\nIs there a reason for utilizing a small network for certified robustness? The robustness tends to increase by the network size (Xie et al., 2020).\\\nIs it possible to report the GOOD (Bitterwolf et al., 2020) result in CIFAR-100 and R.ImgNet? I believe it is the main baseline to consider.\n\n**References**\\\nZhang et al., 2019 “Theoretically Principled Trade-off between Robustness and Accuracy”\\\nBitterwolf et al., 2020, “Certifiably adversarially robust detection of out-of-distribution data”\\\nCroce et al., 2020, “Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks”\\\nHsu et al., 2020, “Generalized ODIN: Detecting Out-of-distribution Image without Learning from Out-of-distribution Data”\\\nXie et al., 2020, “Intriguing properties of adversarial training at scale”\\\nTramer et al., 2021, “Detecting Adversarial Examples Is (Nearly) As Hard As Classifying Them”\\\nZhang et al., 2021, “Towards Certifying L-infinity Robustness using Neural Networks with L-inf-dist Neurons”", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to detect out-of-distribution data in an adversarially robust manner. To this end, the authors incorporate a certified (binary) classifier to model in-and-out distribution and jointly model predictive distribution $p(y|x)$ by conditioning with the binary classifier, i.e., $p(y|x, in)p(in|x) + p(y|x,out)p(out|x)$. The authors show that the proposed method is empirically strong under various detection scenarios.", "main_review": "**Strength**\\\nThe proposed method is sensible and shows consistent improvements.\\\nThe paper provides experiments on several detection benchmarks.\n\n**Weakness**\\\n*Limited evaluation*\\\nDespite the fact that this scheme utilizes a certified classifier, the author evaluates with empirical robustness measures (to calculate the maximum perturbation). Rather than that, I suggest utilizing a certified robustness measure with $l_\\infty$ extension (Zhang et al., 2021).\nAdditionally, considering AutoAttack (Croce et al., 2020) to attack the confidence $\\max p(y|x)$ will be more convincing (as an empirical robustness measure): calculate the maximum confidence over the ensemble attacks in AutoAttack.
\n\n*Discussion with Tramer et al., 2021*\\\nTramer et al., 2021 prove that detecting adversarial samples is hard as classification (Tramer et al., 2021), i.e., classifying $\\epsilon$ is almost the same as detecting $2*\\epsilon$ sample. Due to this paper, it was hard for me to believe that adversarial detection is (almost) free: as it is known to sacrifice the clean accuracy to obtain adversarial robustness (Zhang et al., 2019). Can the author rigorously discuss with the following paper (Tramer et al., 2021)?\n\n*Limited technical novelty*\n* The proposed method can be seen as a combination of two OOD methods (Hsu et al., 2020) and (Bitterwolf et al., 2020).\n* The main technical novelty of this paper is to (1) model predictive distribution with in-and-out conditional distribution (2) utilize IBP for modeling in-and-out distribution to model certification of OOD robustness.\n* However, I believe (1) can be found in (Hsu et al., 2020), and (2) corresponds to (Bitterwolf et al., 2020). \n\n*The AUC of ProoD-disc seems to be low*. This implies that the discriminator does not capture the in-and-out distribution probability well.\n\n(minor) *The writing and presentation can be improved*\n* For instance, a single paragraph contains many messages, making the readers confused about the main message.
\n\n**Questions**\\\nIs there a reason for utilizing a small network for certified robustness? The robustness tends to increase by the network size (Xie et al., 2020).\\\nIs it possible to report the GOOD (Bitterwolf et al., 2020) result in CIFAR-100 and R.ImgNet? I believe it is the main baseline to consider.\n\n**References**\\\nZhang et al., 2019 “Theoretically Principled Trade-off between Robustness and Accuracy”\\\nBitterwolf et al., 2020, “Certifiably adversarially robust detection of out-of-distribution data”\\\nCroce et al., 2020, “Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks”\\\nHsu et al., 2020, “Generalized ODIN: Detecting Out-of-distribution Image without Learning from Out-of-distribution Data”\\\nXie et al., 2020, “Intriguing properties of adversarial training at scale”\\\nTramer et al., 2021, “Detecting Adversarial Examples Is (Nearly) As Hard As Classifying Them”\\\nZhang et al., 2021, “Towards Certifying L-infinity Robustness using Neural Networks with L-inf-dist Neurons”", "summary_of_the_review": "I recommend weak rejection for this review. I believe the evaluation is somewhat questionable and also needs some rigorous discussion with related works. I still believe the idea is sensible, so I carefully request the authors to respond to my weakness part during the rebuttal.\n\n-------\n**POST REBUTTAL:** After the response, I am slightly above the threshold as the proposed method seems to be a scalable work as it stabilizes the hardness of training robust out-of-distribution (OOD) detector.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635854819184}, {"id": "JkQq_qSDcCh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4054/Reviewer_fH7P"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "A novel certifiable OOD detector is described in the paper. The proposed method, ProoD, merges a binary classifier and a multi-class classifier in a clever way to produce an OOD detector robust to adversarial perturbation. The binary classifier is trained to discriminate inliers and outliers, while the multi-class classifier is trained to predict class labels under outlier exposure. ProoD achieves good multi-class classification performance, good outlier detection performance, and robustness against adversarial perturbations on outliers.", "review_text": "# Strengths\n\n- The formulation of Prood is novel, sound, well-motivated, and reasonable. \n- The theoretical contribution (Theorem 1) seems valid.\n- The presented empirical results seem promising. \n\n\n# Weaknesses\n\n- There is a large room to improve in terms of the paper's clarity. There are multiple points in the paper that needs to be made clearer.\n    - In Table 1, \"High clean OOD\" is supposed to be \"High clean OOD detection performance\".\n    - In Eq. (3), $y$ should be clarified. Does it hold for all $y$'s? Also, the first inequality needs to be explained.\n- The paper can be made more self-contained by including some background knowledge. For example, clear definitions of GAUC and AAUC should be provided.\n\n# Questions\n\n- For me, it is actually surprising that this approach works. In principle, a supervised classifier trained to discriminate inliers and outliers in a specific dataset will not necessarily be able to successfully detect unseen outliers. Probably this is why the clean AUC of ProoD-Disc in Table 2 is somewhat low. How can OOD detection AUC be improved even if an under-performing component is incorporated into the model?\n- From Eq. (7) and the text nearby, $p(y|x,i)$ is not trained for robustness and the only component that is robustified is the binary classifier. (Please correct me if I'm wrong.) How can the whole model be robust if only a part of the model is robustified?\n- Adversarial attack can also be performed on inliers so that it is misclassified (as in conventional attacks). How does ProoD respond to such attacks?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "A novel certifiable OOD detector is described in the paper. The proposed method, ProoD, merges a binary classifier and a multi-class classifier in a clever way to produce an OOD detector robust to adversarial perturbation. The binary classifier is trained to discriminate inliers and outliers, while the multi-class classifier is trained to predict class labels under outlier exposure. ProoD achieves good multi-class classification performance, good outlier detection performance, and robustness against adversarial perturbations on outliers.", "main_review": "# Strengths\n\n- The formulation of Prood is novel, sound, well-motivated, and reasonable. \n- The theoretical contribution (Theorem 1) seems valid.\n- The presented empirical results seem promising. \n\n\n# Weaknesses\n\n- There is a large room to improve in terms of the paper's clarity. There are multiple points in the paper that needs to be made clearer.\n    - In Table 1, \"High clean OOD\" is supposed to be \"High clean OOD detection performance\".\n    - In Eq. (3), $y$ should be clarified. Does it hold for all $y$'s? Also, the first inequality needs to be explained.\n- The paper can be made more self-contained by including some background knowledge. For example, clear definitions of GAUC and AAUC should be provided.\n\n# Questions\n\n- For me, it is actually surprising that this approach works. In principle, a supervised classifier trained to discriminate inliers and outliers in a specific dataset will not necessarily be able to successfully detect unseen outliers. Probably this is why the clean AUC of ProoD-Disc in Table 2 is somewhat low. How can OOD detection AUC be improved even if an under-performing component is incorporated into the model?\n- From Eq. (7) and the text nearby, $p(y|x,i)$ is not trained for robustness and the only component that is robustified is the binary classifier. (Please correct me if I'm wrong.) How can the whole model be robust if only a part of the model is robustified?\n- Adversarial attack can also be performed on inliers so that it is misclassified (as in conventional attacks). How does ProoD respond to such attacks?", "summary_of_the_review": "I vote to accept this paper because its contributions are clear, novel, and significant.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635406823763}, {"id": "Z4o8yaLiIpO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4054/Reviewer_guNu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this paper, the authors propose ProoD which merges a certified binary classifier for in-versus out-distribution with a classifier for the in-distribution task in a principled fashion into a joint classifier. They show that ProoD simultaneously achieves three properties: 1) guaranteed OOD robustness via confidence upper bounds on $l_\\infty$-balls around OOD samples; 2) it provably prevents the asymptotic overconfidence of deep neural networks; 3) it can be used with arbitrary architectures and has no loss in prediction performance and standard OOD detection performance. They perform extensive experiments to show the promising performance of the proposed method ProoD.", "review_text": "I think this paper has the following strengths:\n\n1. The authors derive the method ProoD in a principled way and prove that the method can prevent the asymptotic overconfidence of deep neural networks; \n\n2.  They perform extensive experiments to evaluate the proposed method and also compare it to existing baselines; \n\n3. The paper is well-written and easy to understand. \n\nHowever, I think this paper has the following weaknesses: \n\n1. The idea of training a discriminator independently via interval bound propagation (IBP) to achieve certified robustness on out-of-distribution samples is not very novel since IBP is an existing technique that can ensure certified robustness. Also, the theoretical result (Theorem 1) that the proposed joint classifier gets provably less confident in its decisions as one moves away from the training data is not entirely novel since it is mainly based on the previous theoretical results in the literature; \n\n2. Since the proposed method ProoD uses IBP in training to get the certified discriminator g, it is expected that it has better certified robustness than other methods that don't use IBP. However, it might be hard to train models using IBP with a large perturbation budget (e.g. use $\\epsilon=8/255$ that is common in previous OOD detection papers such as ATOM). The authors may need to acknowledge this limitation; \n\n3. When the authors retrain ATOM and ACET models, they might need to consider using stronger PGD attacks during training since they use very strong attacks with adaptive step size for the evaluation of AAUC. In ATOM paper, they only use PGD attack with 5 steps and fixed step size. Using stronger PGD attacks with more steps and adaptive step size (e.g. use backtracking) for training ATOM and ACET might lead to better results (or AAUC) under their attacks in evaluation. \n\n4. The performance of the proposed method ProoD is not stable across different in-distribution and OOD datasets. For example, in Table 2, the performance of ProoD on CIFAR-10 vs. Smooth is worse than that of ATOM and GOOD in terms of AAUC metric, and in terms of GAUC metric, the performance of ProoD is also worse than that of GOOD. The GAUC metric for ATOM and ACET seems meaningless since it is a lower bound and is equal to 0. On CIFAR-100 as the in-distribution dataset, it seems the performance of ProoD is usually worse than that of ATOM in terms of the AAUC metric. In Table 7 (in the appendix), on CIFAR-10 vs. Uniform or CIFAR-100 vs. Uniform, the performance of ProoD is also worse than that of existing methods like ATOM and GOOD. So it is unclear whether the performance of ProoD is better than that of existing methods or not. I think the authors should evaluate ProoD on more OOD datasets. In the ATOM paper, they also evaluate OOD detectors on OOD datasets like Textures, Places365, LSUN (resize), and iSUN. I suggest the authors report the performance of ProoD on these OOD datasets and compare it to existing methods. \n\n5. I think the authors should give more details about attacking ProoD. For example, what's the attack objective they use to attack ProoD when evaluating the GAUC and AAUC? Since ProoD combines the classifier and detector, the adaptive attacks should attack both the classifier and the detector. If they only attack the detector when evaluating GAUC and AAUC, the results may not be correct and they need to re-evaluate them. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose ProoD which merges a certified binary classifier for in-versus out-distribution with a classifier for the in-distribution task in a principled fashion into a joint classifier. They show that ProoD simultaneously achieves three properties: 1) guaranteed OOD robustness via confidence upper bounds on $l_\\infty$-balls around OOD samples; 2) it provably prevents the asymptotic overconfidence of deep neural networks; 3) it can be used with arbitrary architectures and has no loss in prediction performance and standard OOD detection performance. They perform extensive experiments to show the promising performance of the proposed method ProoD.", "main_review": "I think this paper has the following strengths:\n\n1. The authors derive the method ProoD in a principled way and prove that the method can prevent the asymptotic overconfidence of deep neural networks; \n\n2.  They perform extensive experiments to evaluate the proposed method and also compare it to existing baselines; \n\n3. The paper is well-written and easy to understand. \n\nHowever, I think this paper has the following weaknesses: \n\n1. The idea of training a discriminator independently via interval bound propagation (IBP) to achieve certified robustness on out-of-distribution samples is not very novel since IBP is an existing technique that can ensure certified robustness. Also, the theoretical result (Theorem 1) that the proposed joint classifier gets provably less confident in its decisions as one moves away from the training data is not entirely novel since it is mainly based on the previous theoretical results in the literature; \n\n2. Since the proposed method ProoD uses IBP in training to get the certified discriminator g, it is expected that it has better certified robustness than other methods that don't use IBP. However, it might be hard to train models using IBP with a large perturbation budget (e.g. use $\\epsilon=8/255$ that is common in previous OOD detection papers such as ATOM). The authors may need to acknowledge this limitation; \n\n3. When the authors retrain ATOM and ACET models, they might need to consider using stronger PGD attacks during training since they use very strong attacks with adaptive step size for the evaluation of AAUC. In ATOM paper, they only use PGD attack with 5 steps and fixed step size. Using stronger PGD attacks with more steps and adaptive step size (e.g. use backtracking) for training ATOM and ACET might lead to better results (or AAUC) under their attacks in evaluation. \n\n4. The performance of the proposed method ProoD is not stable across different in-distribution and OOD datasets. For example, in Table 2, the performance of ProoD on CIFAR-10 vs. Smooth is worse than that of ATOM and GOOD in terms of AAUC metric, and in terms of GAUC metric, the performance of ProoD is also worse than that of GOOD. The GAUC metric for ATOM and ACET seems meaningless since it is a lower bound and is equal to 0. On CIFAR-100 as the in-distribution dataset, it seems the performance of ProoD is usually worse than that of ATOM in terms of the AAUC metric. In Table 7 (in the appendix), on CIFAR-10 vs. Uniform or CIFAR-100 vs. Uniform, the performance of ProoD is also worse than that of existing methods like ATOM and GOOD. So it is unclear whether the performance of ProoD is better than that of existing methods or not. I think the authors should evaluate ProoD on more OOD datasets. In the ATOM paper, they also evaluate OOD detectors on OOD datasets like Textures, Places365, LSUN (resize), and iSUN. I suggest the authors report the performance of ProoD on these OOD datasets and compare it to existing methods. \n\n5. I think the authors should give more details about attacking ProoD. For example, what's the attack objective they use to attack ProoD when evaluating the GAUC and AAUC? Since ProoD combines the classifier and detector, the adaptive attacks should attack both the classifier and the detector. If they only attack the detector when evaluating GAUC and AAUC, the results may not be correct and they need to re-evaluate them. \n\n", "summary_of_the_review": "Although this paper proposes a principled method ProoD for robust OOD detection and has some interesting theoretical results, the experiments conducted are not enough to show the effectiveness of the proposed method. As I mentioned, the authors should be more careful in building the baselines and evaluating the methods. Also, they should give more details about the experiments like the attack objectives used. Thus, I think this paper is not ready for publication. \n\n\n***[Post Rebuttal]***\n\nI think the proposed method ProoD is not very novel since it simply combines previous techniques, and the performance of ProoD is not stable across different in-distribution and OOD datasets. Also, in practice, it might be hard to use ProoD since under the False Positive Rate at 95% true positive rate metric, ProoD doesn't have good performance (it might be hard to select a suitable threshold for ProoD). Thus, I keep my original score and think the paper is not ready for publication.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635113933075}, {"id": "OzbaXmgyYOT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4054/Reviewer_6Xmu"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a method to provide certified adversarial robustness on the out-of-distribution. Moreover, almost no classification accuracy drop is observed. Furthermore, the detection performance on clean OOD is similar to Outlier Exposure (OE) [1] approach. Finally, the solution provably avoids the asymptotic overconfidence problem. The method is based on adding a binary classifier responsible for being certifiably robust to adversarial manipulation on the out-of-distribution data. Apparently, adversarial robust training is not applied to the multiclass classifier, preventing it from presenting classification accuracy drop. A semi-joint training is then applied.", "review_text": "By far, our main concern regarding the proposed solution is the following: Does the proposed method present (certified or not) adversarial robustness on the **in-distribution data**? For example, which part of the solution prevents a malicious person from using an attack on an **in-distribution example** (i.e., making the classifier believe that a sample belongs to the wrong class) to mislead the primary (i.e., the multiclass) classifier? The proposed binary classifier does not appear to be able to prevent this situation. The below excerpts from the paper make us suspect that the proposed solution does not present adversarial robustness on the **in-distribution data**.\n\n1. _\"In contrast to (Bitterwolf et al., 2020) this comes without loss in test accuracy or non-adversarial OOD detection performance as in our model the neural network used for the in-distribution classification task is independent of the binary discriminator. Thus, we p(yjx; i) have the advantage that the classifier can use arbitrary deep neural networks and is **not constrained to certifiable networks**.\"_\n\n2. _\"Note that this is not standard adversarial training for a binary classification problem as here we have an asymmetric situation: we want to be (certifiably) robust to adversarial manipulation on the out-distribution data **but not on the in-distribution** and thus the upper bound is only used for out-distribution samples.\"_\n\n3. _\"While in (Bitterwolf et al., 2020) they also used IBP to upper bound the confidence of the classifier this resulted in a bound that took into account all O(K2) logit differences between all classes. In contrast, our loss in Eq. (4) is significantly simpler as **we just have a binary classification problem and therefore only need a single bound**.\"_\n\nPlease, notice that ACET [2] presents adversarial robustness on the **in-distribution data**. Moreover, GOOD [3] presents _certified_ adversarial robustness on the **in-distribution data**. If we lose this, making the solution certified adversarial robust on the out-distribution data does not matter anymore, as the attacker may manipulate the solution by simply attacking in-distribution data rather than out-of-distribution data. Considering we are indeed correct, and the proposed solution does not provide certified adversarial robustness on the in-distribution data, **this fact may perfectly explain why the solution, unlike competing approaches, almost does not present loss in prediction accuracy**.\n \nMaybe we are missing something. In such case, please, make the paper clearer regarding this point and show evidence that the proposed solution indeed presents (certified or not) adversarial robustness on the **in-distribution data**. For example, provide classification accuracy on adversarially manipulated in-distribution data (e.g., see [2]).\n\nWe have some additional less problematic concerns regarding the paper.\n\nA drawback of the proposed method is the need to design an ad hoc binary classifier. We do not know whether the proposed binary classifier will work adequately for different models and datasets. Moreover, we need to define a training procedure for it. The solution adds a hyperparameter called the bias shift. We believe the authors could be more explicit about these limitations.\n\nWe also recommend that the authors combine their approaches with the IsoMax loss [4,5] and the IsoMax+ loss [6] rather than SoftMax loss or even OE to start with an improved OOD detection baseline. At least, the mentioned IsoMax loss could be cited as related works. The authors write: \"we also want to achieve SOTA performance on unperturbed OOD data.\" However, IsoMax+ loss outperforms OE in some cases [6]. Hence, OE does not currently present SOTA performance.\n\nThe authors use the word \"robust\" in the paper to mean \"adversarially robust.\" Considering that many other types of robustness exist, we suggest that authors write \"adversarially robust\" rather than simply \"robust\". Additionally, the authors sometimes refer to the multiclass classifier simply as the classifier. Considering that the solution also has a binary classifier, we suggest referring to the leading network as the multiclass classifier to make things more precise.\n\nWe recognize that the differences in performance of the proposed approach to the competing are significant. However, we need always to keep in mind that the proposed approach does not appear to present adversarial robustness on the **in-distribution data**, which is a major problem. Regardless of anything, it would be great to have the mean and standard deviation of five runs in Table 2. We also believe it is essential to add a column to Table 2 showing the classification accuracy on adversarial attacked in-distribution data.\n\nFinally, we understand that the authors say that the approach is (almost) for free because it does not produce classification accuracy drop, and it has OOD detection performance similar to OE. However, many procedures need to be done to achieve this. Hence, we believe that \"(almost) for free\" may a bit be misleading.\n\n[1] Deep Anomaly Detection with Outlier Exposure: https://arxiv.org/abs/1812.04606\n\n[2] Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem: https://arxiv.org/abs/1812.05720\n\n[3] Certifiably Adversarially Robust Detection of Out-of-Distribution Data: https://arxiv.org/abs/2007.08473\n\n[4] IsoMax loss: https://arxiv.org/abs/1908.05569\n\n[5] IsoMax loss (journal): https://arxiv.org/abs/2006.04005\n\n[6] IsoMax+ loss: https://arxiv.org/abs/2105.14399\n\nDavid Macêdo\n\n########################################################################################### ########################################################################################### ########################################################################################### ###########################################################################################\n\n===== FINAL RECOMMENDATION POST-REBUTTAL ========\n\nBy considering that the proposed approach is somewhat novel and presents convincing results and that, after the rebuttal, the authors performed many runs for their method, recognized ACET is robust for in-distribution, added classification accuracy on attacked in-distribution data for all methods, improved clarity by adding terms proposed by the reviewer, make clear that the solution is not robust against attacks on in-distribution data, recognize that we need to design and train an ad-hoc binary discriminator; we are changing our recommendation for \"accept\".\n\n########################################################################################### ########################################################################################### ########################################################################################### ###########################################################################################\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method to provide certified adversarial robustness on the out-of-distribution. Moreover, almost no classification accuracy drop is observed. Furthermore, the detection performance on clean OOD is similar to Outlier Exposure (OE) [1] approach. Finally, the solution provably avoids the asymptotic overconfidence problem. The method is based on adding a binary classifier responsible for being certifiably robust to adversarial manipulation on the out-of-distribution data. Apparently, adversarial robust training is not applied to the multiclass classifier, preventing it from presenting classification accuracy drop. A semi-joint training is then applied.", "main_review": "By far, our main concern regarding the proposed solution is the following: Does the proposed method present (certified or not) adversarial robustness on the **in-distribution data**? For example, which part of the solution prevents a malicious person from using an attack on an **in-distribution example** (i.e., making the classifier believe that a sample belongs to the wrong class) to mislead the primary (i.e., the multiclass) classifier? The proposed binary classifier does not appear to be able to prevent this situation. The below excerpts from the paper make us suspect that the proposed solution does not present adversarial robustness on the **in-distribution data**.\n\n1. _\"In contrast to (Bitterwolf et al., 2020) this comes without loss in test accuracy or non-adversarial OOD detection performance as in our model the neural network used for the in-distribution classification task is independent of the binary discriminator. Thus, we p(yjx; i) have the advantage that the classifier can use arbitrary deep neural networks and is **not constrained to certifiable networks**.\"_\n\n2. _\"Note that this is not standard adversarial training for a binary classification problem as here we have an asymmetric situation: we want to be (certifiably) robust to adversarial manipulation on the out-distribution data **but not on the in-distribution** and thus the upper bound is only used for out-distribution samples.\"_\n\n3. _\"While in (Bitterwolf et al., 2020) they also used IBP to upper bound the confidence of the classifier this resulted in a bound that took into account all O(K2) logit differences between all classes. In contrast, our loss in Eq. (4) is significantly simpler as **we just have a binary classification problem and therefore only need a single bound**.\"_\n\nPlease, notice that ACET [2] presents adversarial robustness on the **in-distribution data**. Moreover, GOOD [3] presents _certified_ adversarial robustness on the **in-distribution data**. If we lose this, making the solution certified adversarial robust on the out-distribution data does not matter anymore, as the attacker may manipulate the solution by simply attacking in-distribution data rather than out-of-distribution data. Considering we are indeed correct, and the proposed solution does not provide certified adversarial robustness on the in-distribution data, **this fact may perfectly explain why the solution, unlike competing approaches, almost does not present loss in prediction accuracy**.\n \nMaybe we are missing something. In such case, please, make the paper clearer regarding this point and show evidence that the proposed solution indeed presents (certified or not) adversarial robustness on the **in-distribution data**. For example, provide classification accuracy on adversarially manipulated in-distribution data (e.g., see [2]).\n\nWe have some additional less problematic concerns regarding the paper.\n\nA drawback of the proposed method is the need to design an ad hoc binary classifier. We do not know whether the proposed binary classifier will work adequately for different models and datasets. Moreover, we need to define a training procedure for it. The solution adds a hyperparameter called the bias shift. We believe the authors could be more explicit about these limitations.\n\nWe also recommend that the authors combine their approaches with the IsoMax loss [4,5] and the IsoMax+ loss [6] rather than SoftMax loss or even OE to start with an improved OOD detection baseline. At least, the mentioned IsoMax loss could be cited as related works. The authors write: \"we also want to achieve SOTA performance on unperturbed OOD data.\" However, IsoMax+ loss outperforms OE in some cases [6]. Hence, OE does not currently present SOTA performance.\n\nThe authors use the word \"robust\" in the paper to mean \"adversarially robust.\" Considering that many other types of robustness exist, we suggest that authors write \"adversarially robust\" rather than simply \"robust\". Additionally, the authors sometimes refer to the multiclass classifier simply as the classifier. Considering that the solution also has a binary classifier, we suggest referring to the leading network as the multiclass classifier to make things more precise.\n\nWe recognize that the differences in performance of the proposed approach to the competing are significant. However, we need always to keep in mind that the proposed approach does not appear to present adversarial robustness on the **in-distribution data**, which is a major problem. Regardless of anything, it would be great to have the mean and standard deviation of five runs in Table 2. We also believe it is essential to add a column to Table 2 showing the classification accuracy on adversarial attacked in-distribution data.\n\nFinally, we understand that the authors say that the approach is (almost) for free because it does not produce classification accuracy drop, and it has OOD detection performance similar to OE. However, many procedures need to be done to achieve this. Hence, we believe that \"(almost) for free\" may a bit be misleading.\n\n[1] Deep Anomaly Detection with Outlier Exposure: https://arxiv.org/abs/1812.04606\n\n[2] Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem: https://arxiv.org/abs/1812.05720\n\n[3] Certifiably Adversarially Robust Detection of Out-of-Distribution Data: https://arxiv.org/abs/2007.08473\n\n[4] IsoMax loss: https://arxiv.org/abs/1908.05569\n\n[5] IsoMax loss (journal): https://arxiv.org/abs/2006.04005\n\n[6] IsoMax+ loss: https://arxiv.org/abs/2105.14399\n\nDavid Macêdo\n\n########################################################################################### ########################################################################################### ########################################################################################### ###########################################################################################\n\n===== FINAL RECOMMENDATION POST-REBUTTAL ========\n\nBy considering that the proposed approach is somewhat novel and presents convincing results and that, after the rebuttal, the authors performed many runs for their method, recognized ACET is robust for in-distribution, added classification accuracy on attacked in-distribution data for all methods, improved clarity by adding terms proposed by the reviewer, make clear that the solution is not robust against attacks on in-distribution data, recognize that we need to design and train an ad-hoc binary discriminator; we are changing our recommendation for \"accept\".\n\n########################################################################################### ########################################################################################### ########################################################################################### ###########################################################################################\n\n", "summary_of_the_review": "Currently, it appears to us that the proposed solution does not provide adversarial robustness on the **in-distribution data**. If this is the case, we believe that this drawback makes the method much less valuable than previously published competing approaches. Suppose the authors clarify the text and prove that the proposed solution, like ACET and GOOD, provides adversarial robustness on the **in-distribution data**. In that case, we may improve our score mainly if the other concerns are also addressed. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635041066762}], "openreview_url": "https://openreview.net/forum?id=qDx6DXD3Fzt", "arxiv_id": "2106.04260", "paper_pdf": "papers/qDx6DXD3Fzt.pdf", "paper_pdf_sha256": "a6ff5fa80bfc8dadf75d47b4a4f9f12c0464a901d5346415947a0b2e0b0a7b96", "paper_pdf_bytes": 547405, "paper_pdf_source": "openreview", "code_url": "https://github.com/AlexMeinke/Provable-OOD-Detection", "code_repository": "AlexMeinke/Provable-OOD-Detection", "code_commit": "9cef8d284811b938a6182835c4fd16590b1c2699", "code_archive": "repos/qDx6DXD3Fzt.zip", "code_archive_sha256": "3b3ff895cbf1dd4a24c0c7428cb7247089303265d8b670bd22e759819cab605e", "code_archive_bytes": 483105, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 483, "github_languages": {"Python": 302833}, "github_archived": false, "github_pushed_at": "2022-10-12T18:23:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/provably-robust-detection-of-out-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fgX9O5q0BT", "year": 2021, "status": "rejected", "title": "On Noise Injection in Generative Adversarial Networks", "authors": ["Ruili Feng", "Deli Zhao", "Zheng-Jun Zha"], "authorids": ["~Ruili_Feng1", "~Deli_Zhao1", "~Zheng-Jun_Zha2"], "authors_source": "OpenReview API", "abstract": "Noise injection is an effective way of circumventing overfitting  and enhancing generalization in machine learning, the rationale of which has been validated in deep learning as well.  Recently, noise injection exhibits surprising performance when\n  generating high-fidelity images in Generative Adversarial Networks (GANs). Despite its successful applications in GANs, the mechanism of its validity is still unclear. In this paper, we propose a geometric framework to theoretically analyze the role of noise injection in GANs. Based on Riemannian geometry, we successfully model the noise injection framework as fuzzy equivalence on geodesic normal coordinates. Guided by our theories, we find that existing methods are incomplete and a new strategy for noise injection is devised. Experiments on image generation and GAN inversion demonstrate the superiority of our method.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "eaRSziRFwd0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1560/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "To summarize, this paper proposed a new noise injection method that is easy to implement and is able to replace the original noise injection method in StyleGAN 2. The approach is supported by detailed theoretical analysis and impactful performance improvement on GAN training and inversion. The results show that they are able to achieve a considerable improvement on DCGAN and StyleGAN2.\n\nMeanwhile, this reviewer did not fully understand the theoretical part and also has some questions regarding the implementation and results.\n\nBased on my understanding, the fuzzy reparameterization technique realizes something that StyleGAN2 cannot achieve, and resolves some fundamental limitations of StyleGAN2. However, the improvement is not contiguous in Table 1, where we see the vanilla StyleGAN2 still outperforms the proposed architecture. What could be the reason?\n\nFR seems to bring more parameters (Eq 6, 7, 8, 9). How many more compared to the additive noise implementation? Could the number of parameters be the reason that the proposed method performs better? Since FR can be seen as a generalization of StyleGAN2 noise injection, we would naturally expect that the proposed method should perform better than StyleGAN2. However, this is not always the case in Table 1. I guess more ablation studies can also be done on $\\sigma$, such as interpolating between StyleGAN2 implementation and FR implementation, or a linear layer with the same number of additional parameters but has no constraint as in Eq. 6, 7, 8, 9.\n\nFor Figure 8, do we have the reconstruction visualization? Is the inversion done in the z space, w space or the w+ space?  I am curious to see how better this method performs in terms of inverting real images in the wild. I also believe the inversion in z space allows me to appreciate more about the inversion improvement.\n\nOverall, I vote to accept this paper due to its good performance improvement over prior standard noise injection implementation. Meanwhile, I hope the theoretical analysis can be made easier to understand for a researcher that lacks the related background.\n\n[Update after reading authors' comments]\nBased on the authors' and other reviewers' comments,  I keep the score unchanged.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice Results", "review": "To summarize, this paper proposed a new noise injection method that is easy to implement and is able to replace the original noise injection method in StyleGAN 2. The approach is supported by detailed theoretical analysis and impactful performance improvement on GAN training and inversion. The results show that they are able to achieve a considerable improvement on DCGAN and StyleGAN2.\n\nMeanwhile, this reviewer did not fully understand the theoretical part and also has some questions regarding the implementation and results.\n\nBased on my understanding, the fuzzy reparameterization technique realizes something that StyleGAN2 cannot achieve, and resolves some fundamental limitations of StyleGAN2. However, the improvement is not contiguous in Table 1, where we see the vanilla StyleGAN2 still outperforms the proposed architecture. What could be the reason?\n\nFR seems to bring more parameters (Eq 6, 7, 8, 9). How many more compared to the additive noise implementation? Could the number of parameters be the reason that the proposed method performs better? Since FR can be seen as a generalization of StyleGAN2 noise injection, we would naturally expect that the proposed method should perform better than StyleGAN2. However, this is not always the case in Table 1. I guess more ablation studies can also be done on $\\sigma$, such as interpolating between StyleGAN2 implementation and FR implementation, or a linear layer with the same number of additional parameters but has no constraint as in Eq. 6, 7, 8, 9.\n\nFor Figure 8, do we have the reconstruction visualization? Is the inversion done in the z space, w space or the w+ space?  I am curious to see how better this method performs in terms of inverting real images in the wild. I also believe the inversion in z space allows me to appreciate more about the inversion improvement.\n\nOverall, I vote to accept this paper due to its good performance improvement over prior standard noise injection implementation. Meanwhile, I hope the theoretical analysis can be made easier to understand for a researcher that lacks the related background.\n\n[Update after reading authors' comments]\nBased on the authors' and other reviewers' comments,  I keep the score unchanged.", "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603948816899}, {"id": "gKHxvgSzYzI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1560/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper the authors highlight two major drawbacks of GANs. 1) The optimal Generator is discontinuous and 2) the 'adversarial dimension trap' caused by the relatively lower dimension of the latent space compared to the real-world data which makes the generator not Lipschitz and/or the generator fails to capture the real-world data distribution and is not invertible. \nBoth issues could lead to an unsmooth generator.\n\nSecondly, the authors provide a form of generalization of noise injection in GANs called fuzzy reparameterization, which leads to a solution by letting the generator map onto an arbitrarily low dimensional skeleton of the feature spaces and filling up the remaining space with random noise. Therefore, the difference to the latent space dimensionality is minimized addressing issue 2). The solution consists of two stages, first a map from feature space onto the skeleton set is learned followed by noise injection adapted to the local geometry of the orginal feature manifold. \n\nExperiments:\n\nExperiments were done on FFHQ faces, LSUN objects, and CIFAR-10 datasets with models DCGAN, StyleGAN2, and bald StyleGAN2 which is StyleGAN2 without noise injection and path length regularizer. On DCGAN the proposed fuzzy reparameterization (FR) outperforms DCGAN with and without additive noise on FFHQ, CIFAR10, LSUN-Church measured with the Path Perceptual Length (PPL) and the FID. Also StyleGAN2 with FR outperforms StyleGAN2 with additive noise and bald StyleGAN2 on FFHQ and LSUN objects. To test numerical stability, condition numbers of 50000 Input, Pertubation pairs were computed for StyleGAN2 models with, without additive noise and with FR and path length regularization. StyleGAN2 + FR outperforms both on the mean and top-1000 mean condition number indicating that FR improves numerical stability. StyleGAN2 + FR also outperforms on the image inversion experiments.\n\nPros: This paper provides a theoretical framework for noise injection for GANs which is novel and interesting for the GAN community. The experimental results are extensive and convincing and support the theoretical analysis.\n\nCons: In section 4.3 the algorithm eq. 6-8 is not very clear to me. E.g. what are the parameters A,b, alpha and r and how are they motivated? PixSum is over the feature maps? A more detailed description with comments would be helpful for the reader. I could not find the FR implementation in the supplementary file, it looks like it contains only the original StyleGAN(2), DCGAN and DCGAN with additive noise models.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good contribution to the GAN research field", "review": "In this paper the authors highlight two major drawbacks of GANs. 1) The optimal Generator is discontinuous and 2) the 'adversarial dimension trap' caused by the relatively lower dimension of the latent space compared to the real-world data which makes the generator not Lipschitz and/or the generator fails to capture the real-world data distribution and is not invertible. \nBoth issues could lead to an unsmooth generator.\n\nSecondly, the authors provide a form of generalization of noise injection in GANs called fuzzy reparameterization, which leads to a solution by letting the generator map onto an arbitrarily low dimensional skeleton of the feature spaces and filling up the remaining space with random noise. Therefore, the difference to the latent space dimensionality is minimized addressing issue 2). The solution consists of two stages, first a map from feature space onto the skeleton set is learned followed by noise injection adapted to the local geometry of the orginal feature manifold. \n\nExperiments:\n\nExperiments were done on FFHQ faces, LSUN objects, and CIFAR-10 datasets with models DCGAN, StyleGAN2, and bald StyleGAN2 which is StyleGAN2 without noise injection and path length regularizer. On DCGAN the proposed fuzzy reparameterization (FR) outperforms DCGAN with and without additive noise on FFHQ, CIFAR10, LSUN-Church measured with the Path Perceptual Length (PPL) and the FID. Also StyleGAN2 with FR outperforms StyleGAN2 with additive noise and bald StyleGAN2 on FFHQ and LSUN objects. To test numerical stability, condition numbers of 50000 Input, Pertubation pairs were computed for StyleGAN2 models with, without additive noise and with FR and path length regularization. StyleGAN2 + FR outperforms both on the mean and top-1000 mean condition number indicating that FR improves numerical stability. StyleGAN2 + FR also outperforms on the image inversion experiments.\n\nPros: This paper provides a theoretical framework for noise injection for GANs which is novel and interesting for the GAN community. The experimental results are extensive and convincing and support the theoretical analysis.\n\nCons: In section 4.3 the algorithm eq. 6-8 is not very clear to me. E.g. what are the parameters A,b, alpha and r and how are they motivated? PixSum is over the feature maps? A more detailed description with comments would be helpful for the reader. I could not find the FR implementation in the supplementary file, it looks like it contains only the original StyleGAN(2), DCGAN and DCGAN with additive noise models.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603896660739}, {"id": "55QT618mp1M", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1560/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper analyzes the theoretical properties of noise injection in StyleGAN-like networks, and proposes an extension to in particular to StyleGAN2 that results in somewhat improved metric scores. Unfortunately the paper is rather confusingly written and hard to follow. To highlight what I mean, I will try to paraphrase my understanding of the paper in the following.\n\nThe theoretical treatment begins by framing the problem around optimal transport, but later seems to mostly drop this viewpoint. While the cited OT/GAN work presents interesting and relevant viewpoints about the difficulties in GAN training, I am not sure if it is particularly more relevant here than any number of other theoretical works. It may be noted that StyleGAN and DCGAN are not even formulated as to minimize a Wasserstein divergence.\n\nThe paper then coins a term \"adversarial dimension trap\", which I am not exactly sure why this terminology was chosen. The gist of the observation seems to be mostly well known, i.e. the generator can only cover a zero-measure region of the data space whereas the data is more spread out. That said, I am not thoroughly familiar with previous theoretical work on GANs and the particular formulation here may be novel. The paper then introduces a fairly general form of stochastic noise injection into the network layers and calls this fuzzy reparametrization. Here some connections to are drawn to \"fuzzy equivalence relations\" which (apparently?) are an existing concept, however as far as I see there is no citation to discuss these and little insight is given about why this is relevant.\n\nThen, the key theory is developed. If I understand correctly, the key idea here is that the zero-measure manifold is \"puffed up\" with random distributions centered around points on it, with spread that depends on the point. This makes sense as a principle but overall I am confused about whether something fundamental was discrovered or proved, or whether this is just an introduction to the reasoning behind the practical algorithms.\n\nThis then leads to a proposal of a practical algorithm. If I understand correctly, it basically generalizes the StyleGAN noise injection in such a way that not only the mean, but also the stdev of each feature is predicted by the network. This plausibly allows for more flexibility. Unfortunately the description here is again rather confusing. Apparently formulas 6-8 are not really equations, but rather some kind of imperative pseudocode with variable assignments. It would be better to spell this out as an algorithm listing. As for the content of these formulas, I am not sure if I understand what the operations or the reasoning behind them is. Why the pixsum here? Apparently it produces a single value per feature map? After this there seems that these numbers are transformed by some global matrix(?) A and bias b, however I'm not sure what the convex combination with a matrix(?) I means here given that the first half of the formula is a vector (?). Then the result seems to be normalized again (why?) And finally the means are transformed by this standard deviation? There may well be good reasons to use these steps, but they are not explained so it ends up looking like an arbitrary heuristic. Here it would be important to make a strong connection to the insights derived from theory.\n\nThe presentation is further made confusing by the language. I understand that the authors may not be native speakers, but the readability is much below the usual standard of ICLR papers and the paper would benefit from improving this.\n\nAs for the results, it does appear that there is some improvement in some of the metrics, and the proposed method may in principle be useful. It is not hard to believe that adding some extra flexibility to the noise injection might improve the results, at least in a limited number of scenarios. In this sense the paper may be on to something. \n\nWhat is the meaning of using PageRank to reduce the number of LSUN-Cat images? How is PageRank related to choosing images and what's the difference between that and just taking the first 100k pictures in the set? And for that matter, I am not sure if we learn anything from randomly limiting the set to 100k images, when we don't know how it worked for the full set. For the inversion experiments, the table in the appendix does show improvement and this may well be the case. In figure 9, though, it's hard to see much of a difference between any of the methods, perhaps in part because the images shown are very low resolution and do not correspond to anywhere near the SG2 output image size -- any differences in details are completely hidden by this.\n\nThe architecture figures 6-7 are unnecessarily low detail. They contain a black box \"FR\" node precisely at the place where you'd want to know more. Perhaps this node could be expanded into its own architecture diagram as well, given that there is no shortage of space in the appendix.\n\nIn summary the paper might contain useful bits -- this is somewhat hard to judge -- but whether or not that is the case, it is not in an acceptable condition without some significant rewriting, and I would recommend rejection at this time.\n\n_UPDATE AFTER REBUTTAL_\n\nThe authors have improved the paper somewhat by expanding and clarifying the discussion on some key parts. While I think there is still much room for improvement in the paper, the general consensus seems to be towards acceptance. I will not oppose if that is the decision, and have increased my score accordingly. However I remain very borderline and I am not sure if I am fully convinced by all the claims.\n\nOne specific issue: I think the authors should make it more clear in the paper that the experiments are done in 128 pixel resolution, in light of R1's questions. It is important that the reader be aware of this, as the noise inputs arguably become much more important in high resolutions where there is more stochastic detail. I personally did not realize this when writing my review, and now wonder how the results would be at e.g. 256 or 512 resolution. If possible I would suggest the authors still run such experiments. This also probably explains my comment above on the lack of apparent visual differences in inversion results.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Could contain useful bits but not ready for publication (but see updated part)", "review": "The paper analyzes the theoretical properties of noise injection in StyleGAN-like networks, and proposes an extension to in particular to StyleGAN2 that results in somewhat improved metric scores. Unfortunately the paper is rather confusingly written and hard to follow. To highlight what I mean, I will try to paraphrase my understanding of the paper in the following.\n\nThe theoretical treatment begins by framing the problem around optimal transport, but later seems to mostly drop this viewpoint. While the cited OT/GAN work presents interesting and relevant viewpoints about the difficulties in GAN training, I am not sure if it is particularly more relevant here than any number of other theoretical works. It may be noted that StyleGAN and DCGAN are not even formulated as to minimize a Wasserstein divergence.\n\nThe paper then coins a term \"adversarial dimension trap\", which I am not exactly sure why this terminology was chosen. The gist of the observation seems to be mostly well known, i.e. the generator can only cover a zero-measure region of the data space whereas the data is more spread out. That said, I am not thoroughly familiar with previous theoretical work on GANs and the particular formulation here may be novel. The paper then introduces a fairly general form of stochastic noise injection into the network layers and calls this fuzzy reparametrization. Here some connections to are drawn to \"fuzzy equivalence relations\" which (apparently?) are an existing concept, however as far as I see there is no citation to discuss these and little insight is given about why this is relevant.\n\nThen, the key theory is developed. If I understand correctly, the key idea here is that the zero-measure manifold is \"puffed up\" with random distributions centered around points on it, with spread that depends on the point. This makes sense as a principle but overall I am confused about whether something fundamental was discrovered or proved, or whether this is just an introduction to the reasoning behind the practical algorithms.\n\nThis then leads to a proposal of a practical algorithm. If I understand correctly, it basically generalizes the StyleGAN noise injection in such a way that not only the mean, but also the stdev of each feature is predicted by the network. This plausibly allows for more flexibility. Unfortunately the description here is again rather confusing. Apparently formulas 6-8 are not really equations, but rather some kind of imperative pseudocode with variable assignments. It would be better to spell this out as an algorithm listing. As for the content of these formulas, I am not sure if I understand what the operations or the reasoning behind them is. Why the pixsum here? Apparently it produces a single value per feature map? After this there seems that these numbers are transformed by some global matrix(?) A and bias b, however I'm not sure what the convex combination with a matrix(?) I means here given that the first half of the formula is a vector (?). Then the result seems to be normalized again (why?) And finally the means are transformed by this standard deviation? There may well be good reasons to use these steps, but they are not explained so it ends up looking like an arbitrary heuristic. Here it would be important to make a strong connection to the insights derived from theory.\n\nThe presentation is further made confusing by the language. I understand that the authors may not be native speakers, but the readability is much below the usual standard of ICLR papers and the paper would benefit from improving this.\n\nAs for the results, it does appear that there is some improvement in some of the metrics, and the proposed method may in principle be useful. It is not hard to believe that adding some extra flexibility to the noise injection might improve the results, at least in a limited number of scenarios. In this sense the paper may be on to something. \n\nWhat is the meaning of using PageRank to reduce the number of LSUN-Cat images? How is PageRank related to choosing images and what's the difference between that and just taking the first 100k pictures in the set? And for that matter, I am not sure if we learn anything from randomly limiting the set to 100k images, when we don't know how it worked for the full set. For the inversion experiments, the table in the appendix does show improvement and this may well be the case. In figure 9, though, it's hard to see much of a difference between any of the methods, perhaps in part because the images shown are very low resolution and do not correspond to anywhere near the SG2 output image size -- any differences in details are completely hidden by this.\n\nThe architecture figures 6-7 are unnecessarily low detail. They contain a black box \"FR\" node precisely at the place where you'd want to know more. Perhaps this node could be expanded into its own architecture diagram as well, given that there is no shortage of space in the appendix.\n\nIn summary the paper might contain useful bits -- this is somewhat hard to judge -- but whether or not that is the case, it is not in an acceptable condition without some significant rewriting, and I would recommend rejection at this time.\n\n_UPDATE AFTER REBUTTAL_\n\nThe authors have improved the paper somewhat by expanding and clarifying the discussion on some key parts. While I think there is still much room for improvement in the paper, the general consensus seems to be towards acceptance. I will not oppose if that is the decision, and have increased my score accordingly. However I remain very borderline and I am not sure if I am fully convinced by all the claims.\n\nOne specific issue: I think the authors should make it more clear in the paper that the experiments are done in 128 pixel resolution, in light of R1's questions. It is important that the reader be aware of this, as the noise inputs arguably become much more important in high resolutions where there is more stochastic detail. I personally did not realize this when writing my review, and now wonder how the results would be at e.g. 256 or 512 resolution. If possible I would suggest the authors still run such experiments. This also probably explains my comment above on the lack of apparent visual differences in inversion results.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603878340776}, {"id": "GRSM9ubBojE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1560/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\n\nThis work introduces the problem of adversarial dimension trap, which leads to punishment on the smoothness and invertibility of GANs. \n\nThis work proposes to learn fuzzy equivalence relation of the features and uses reparameterization trick to model the high-dimensional feature manifolds.\n\nA novel form of noise injection is proposed to overcome the adversarial dimension trap. Prior noise injection methods can be explained as a special case with certain hyper-parameters. This method is universal for the families of GANs, including WGAN, DCGAN, etc.\n\nExperiments on three datasets are conducted. The results of both image synthesis and GAN inversion are desirable with plausible texture details.\n\n\nCons:\n\nMy major concern is about the experiments. In Table 1, it seems that the reported FID and PPL differ from the scores reported by StyleGAN2 [1]. For the FFHQ dataset, [1] report that the FID is 3.31 (config E), but the FID of the baseline in this paper is 7.14 (the same setting). Such a huge discrepancy is wired. Recall that [1] improve the FID from 4.40 (StyleGAN v1) to 2.84, but the baseline, which should be the same, performs even worse than StyleGAN v1. The authors need to explain these contradictions.\n \nThe differences between the PPL scores reported in this paper and [1] are even more significant. I notice it is 13.05 (the best in this paper) versus 122.5 (the best in [1]). Why is the PPL about ten times better (even without the proposed method)?\n\nThe authors need to provide more details to explain how they calculated the FIDs and PPL. It seems that the authors calculate these scores in a non-standard manner. Besides, I suggest the author evaluate the Precision and Recall [2] on FFHQ. I wonder whether these metrics will be consistent.\n\n[1] Karras, Tero, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. \"Analyzing and improving the image quality of stylegan.\" CVPR, pp. 8110-8119. 2020.\n\n[2] Kynkäänniemi, Tuomas, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. \"Improved precision and recall metric for assessing generative models.\" NeurIPS, pp. 3927-3936. 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This work explains the properties and proposes a novel form of noise injection in GANs to reduce the adversarial dimension trap problem. Some issues in experiments need to be addressed.", "review": "Pros:\n\nThis work introduces the problem of adversarial dimension trap, which leads to punishment on the smoothness and invertibility of GANs. \n\nThis work proposes to learn fuzzy equivalence relation of the features and uses reparameterization trick to model the high-dimensional feature manifolds.\n\nA novel form of noise injection is proposed to overcome the adversarial dimension trap. Prior noise injection methods can be explained as a special case with certain hyper-parameters. This method is universal for the families of GANs, including WGAN, DCGAN, etc.\n\nExperiments on three datasets are conducted. The results of both image synthesis and GAN inversion are desirable with plausible texture details.\n\n\nCons:\n\nMy major concern is about the experiments. In Table 1, it seems that the reported FID and PPL differ from the scores reported by StyleGAN2 [1]. For the FFHQ dataset, [1] report that the FID is 3.31 (config E), but the FID of the baseline in this paper is 7.14 (the same setting). Such a huge discrepancy is wired. Recall that [1] improve the FID from 4.40 (StyleGAN v1) to 2.84, but the baseline, which should be the same, performs even worse than StyleGAN v1. The authors need to explain these contradictions.\n \nThe differences between the PPL scores reported in this paper and [1] are even more significant. I notice it is 13.05 (the best in this paper) versus 122.5 (the best in [1]). Why is the PPL about ten times better (even without the proposed method)?\n\nThe authors need to provide more details to explain how they calculated the FIDs and PPL. It seems that the authors calculate these scores in a non-standard manner. Besides, I suggest the author evaluate the Precision and Recall [2] on FFHQ. I wonder whether these metrics will be consistent.\n\n[1] Karras, Tero, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. \"Analyzing and improving the image quality of stylegan.\" CVPR, pp. 8110-8119. 2020.\n\n[2] Kynkäänniemi, Tuomas, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. \"Improved precision and recall metric for assessing generative models.\" NeurIPS, pp. 3927-3936. 2019.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603161384820}], "openreview_url": "https://openreview.net/forum?id=fgX9O5q0BT", "arxiv_id": "2006.05891", "paper_pdf": "papers/fgX9O5q0BT.pdf", "paper_pdf_sha256": "0eb1fce18b442aa63f72e38354ef8b77c5cf04c2ca1e8f05c33d7152724b9328", "paper_pdf_bytes": 4516587, "paper_pdf_source": "openreview", "code_url": "https://github.com/NVlabs/stylegan2", "code_repository": "NVlabs/stylegan2", "code_commit": "bf0fe0baba9fc7039eae0cac575c1778be1ce3e3", "code_archive": "repos/fgX9O5q0BT.zip", "code_archive_sha256": "2594614039946bb63b347c8d4e1a50609833062ccccc1df78d2183385895b82b", "code_archive_bytes": 599614, "code_file_count": 41, "code_extensions": {".py": 38, ".cu": 3}, "github_disk_usage_kb": 594, "github_languages": {"Python": 369008, "Cuda": 23649, "Dockerfile": 362}, "github_archived": false, "github_pushed_at": "2024-05-18T15:55:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-noise-injection-in-generative-adversarial"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1xZD1rtPr", "year": 2020, "status": "rejected", "title": "The Dual Information Bottleneck", "authors": ["Zoe Piran", "Naftali Tishby"], "authorids": ["zoe.piran@mail.huji.ac.il", "tishby@cs.huji.ac.il"], "authors_source": "OpenReview API", "abstract": "The Information-Bottleneck (IB) framework suggests a general characterization of optimal representations in learning, and deep learning in particular. It is based on the optimal trade off between the representation complexity and accuracy, both of which are quantified by mutual information. The problem is solved by alternating projections between the encoder and decoder of the representation, which can be performed locally at each representation level. The framework, however, has practical drawbacks, in that mutual information is notoriously difficult to handle at high dimension, and only has closed form solutions in special cases. Further, because it aims to extract representations which are minimal sufficient statistics of the data with respect to the desired label, it does not necessarily optimize the actual prediction of unseen labels. Here we present a  formal dual problem to the IB which has several interesting properties. By switching the order in the KL-divergence between the representation decoder and data, the optimal decoder becomes the geometric rather than the arithmetic mean of the input points. While providing a good approximation to the original IB, it also preserves the form of exponential families, and optimizes the mutual information on the predicted label rather than the desired one. We also analyze the critical points of the dualIB and discuss their importance for the quality of this approach.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyxXpukrjH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1755/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Description:\n\nThe information bottleneck (IB) is an information theoretic principle for optimizing a mapping (encoding, e.g. clustering) of an input, to trade off two kinds of mutual information: minimize mutual information between the original input and the mapped version (to compress the input), and maximize mutual information between the mapped input and an output variable. It is related to a minimization of an (expected) Kullback-Leibler divergence betwen conditional distributions of an output variable.\n\nIn this paper, instead of the original IB, authors consider a previously presented dual problem of Felice and Ay, where the Kullback-Leibler divergence is minimized in the reverse direction: from the conditional distribution of output given encoding, p(y|hat x), to the conditional distribution given the original input, p(y|x).\nThe \"dual problem\" itself has a more complicated form than the original IB, authors claim this is \"a good approximation\" of the original bottleneck formulation, and aim to prove various \"interesting properties\" of it. \n\n- An iterative algorithm (Algorithm 1) similar to the original IB algorithm but with a few more steps is provided.\n\n- A theorem about critical points where cardinality of the representation changes is given, similar to the IB critical points, and another theorem about difference of curves on an information plane between the IB and dual-IB solutions.\n\n- Authors also show that if the true conditional distribution of outputs given inputs has an exponential-family form, the dual-IB decoder also has a form in the same family, which is said to reduce computational complexity of the algorithm.\n\n\n\nEvaluation:\n\nThis is an entirely theory-based paper; although an algorithm is given, it is not instantiated for any concrete representation learning task, and no experiments at all are demonstrated.\n\nOverall, I feel the motivation is not clear and strong enough. The abstract does not illustrate the importance of the mentioned \"interesting properties\" well enough for concrete tasks. Reading the paper, the clearest motivations seem to be improving computational complexity, and having a clearer connection to output prediction in cases where the predictor may be sub-optimal. However, authors do not quantify these well:\n\n- The computational complexity improvement is not made clear (quantified) in a concrete IB optimization task: it seems it is only for exponential families, and even for them only affects one part of the algorithm, reducing its complexity from dim(X) to d; the impact of this is not tried out in any experiment.\n\n- For output prediction, authors motivate that dual-IB could have a more direct connection e.g. \"due to finite sample, in which it can be very different from the one obtained from the full distribution\"). Authors further claim that the dual-IB formulation can \"improve the generalization error when trained on small samples since the predicted label is the one used in practice\". However, this is not tested at all: no prediction experiments, no quantification of generalization error, and no comparisons are done, thus the impact of the clearer connection to output prediction is not tested at all, and no clear theorems are given about it either.\n\nThe property that the algorithm \"preserves exponential form of the original data distribution, if one exists\" is interesting in principle, but it is unclear if any real data would anyway precisely have such a distribution; what happens if the data is not exactly in an exponential family?\n\nIn its current state the paper, although based on an interesting direction, in my opinion does not make a sufficient impact to be accepted to ICLR.\n\nOther comments:\n\n\"Application to deep learning\" mentioned in Section 1.5 is only a sincle sentence in the conclusions.\n\nThere have been some other suggested alternative IB formulations, for example the Deterministic IB of Strouse and Schwab (Neural Computation 2017) which also claim improved computational efficiency. How does the method compare to those?\n\nSection 1.5 claims the algorithm \"preserves the low dimensional sufficient statistics of the data\": it is not clear what \"preserves\" means here, certainly it seems the decoder in Theorem 7 uses the same kinds of sufficient statistics as in the original data, but it is not clear that hat(x) would somehow preserve the same values of the sufficient statistics.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "N/A", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "title": "Official Blind Review #2", "review": "Description:\n\nThe information bottleneck (IB) is an information theoretic principle for optimizing a mapping (encoding, e.g. clustering) of an input, to trade off two kinds of mutual information: minimize mutual information between the original input and the mapped version (to compress the input), and maximize mutual information between the mapped input and an output variable. It is related to a minimization of an (expected) Kullback-Leibler divergence betwen conditional distributions of an output variable.\n\nIn this paper, instead of the original IB, authors consider a previously presented dual problem of Felice and Ay, where the Kullback-Leibler divergence is minimized in the reverse direction: from the conditional distribution of output given encoding, p(y|hat x), to the conditional distribution given the original input, p(y|x).\nThe \"dual problem\" itself has a more complicated form than the original IB, authors claim this is \"a good approximation\" of the original bottleneck formulation, and aim to prove various \"interesting properties\" of it. \n\n- An iterative algorithm (Algorithm 1) similar to the original IB algorithm but with a few more steps is provided.\n\n- A theorem about critical points where cardinality of the representation changes is given, similar to the IB critical points, and another theorem about difference of curves on an information plane between the IB and dual-IB solutions.\n\n- Authors also show that if the true conditional distribution of outputs given inputs has an exponential-family form, the dual-IB decoder also has a form in the same family, which is said to reduce computational complexity of the algorithm.\n\n\n\nEvaluation:\n\nThis is an entirely theory-based paper; although an algorithm is given, it is not instantiated for any concrete representation learning task, and no experiments at all are demonstrated.\n\nOverall, I feel the motivation is not clear and strong enough. The abstract does not illustrate the importance of the mentioned \"interesting properties\" well enough for concrete tasks. Reading the paper, the clearest motivations seem to be improving computational complexity, and having a clearer connection to output prediction in cases where the predictor may be sub-optimal. However, authors do not quantify these well:\n\n- The computational complexity improvement is not made clear (quantified) in a concrete IB optimization task: it seems it is only for exponential families, and even for them only affects one part of the algorithm, reducing its complexity from dim(X) to d; the impact of this is not tried out in any experiment.\n\n- For output prediction, authors motivate that dual-IB could have a more direct connection e.g. \"due to finite sample, in which it can be very different from the one obtained from the full distribution\"). Authors further claim that the dual-IB formulation can \"improve the generalization error when trained on small samples since the predicted label is the one used in practice\". However, this is not tested at all: no prediction experiments, no quantification of generalization error, and no comparisons are done, thus the impact of the clearer connection to output prediction is not tested at all, and no clear theorems are given about it either.\n\nThe property that the algorithm \"preserves exponential form of the original data distribution, if one exists\" is interesting in principle, but it is unclear if any real data would anyway precisely have such a distribution; what happens if the data is not exactly in an exponential family?\n\nIn its current state the paper, although based on an interesting direction, in my opinion does not make a sufficient impact to be accepted to ICLR.\n\nOther comments:\n\n\"Application to deep learning\" mentioned in Section 1.5 is only a sincle sentence in the conclusions.\n\nThere have been some other suggested alternative IB formulations, for example the Deterministic IB of Strouse and Schwab (Neural Computation 2017) which also claim improved computational efficiency. How does the method compare to those?\n\nSection 1.5 claims the algorithm \"preserves the low dimensional sufficient statistics of the data\": it is not clear what \"preserves\" means here, certainly it seems the decoder in Theorem 7 uses the same kinds of sufficient statistics as in the original data, but it is not clear that hat(x) would somehow preserve the same values of the sufficient statistics.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1573349563446}, {"id": "S1xNKglJcH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1755/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a variant of the Information Bottleneck (IB) framework, which consists in permuting the conditional probabilities of y given x and y given \\hat{x} in a Kullback-Liebler divergence involved in the IB optimization criterion.\n\nInterestingly, this change only results in changing an arithmetic mean into a geometric mean in the algorithmic resolution.\n\nGood properties of the exponential families (existence of non-trivial minimal sufficient statistics) are preserved, and an analysis of the new critical points/information plane induced is carried out.\n\nThe paper is globally well written and clear, and the maths are rigorously introduced and treated. Two minor comments would concern the definition of the Mutual Information (MI), which could be recalled to help the unfamiliar reader and improve self-containedness, and the notation of the expected value (\\langle \\rangle_p), unusual in Machine Learning where \\mathbb{E} is often preferred.\n\nAnother point that could be enhanced is the intuition behind the IB and dualIB criteria: a small comment on their meaning/relevance as well as the rationale/implication of the probabilities permutation would be valuable addition.\n\nExhibiting a link with variational autoencoders (VAEs), and expliciting the differences in a VAE framework between the two criteria could also represent an interesting parallel, especially for machine learning oriented readers.\n\nOne of the main drawback of the present paper is however the lack of convincing experiments. Graphs showing the good behavior of the introduced framework are ok, but the clear interest of using dualIB rather than IB could be emphasized more. In particular, it does not seem that the issues about IB raised in the abstract (curse of dimensionality, no closed form solution) are solved with dualIB. Furthermore, dualIB optimizes the MI of the predicted labels, which is claimed to be beneficial contrary to the MI on the actual labels. However, no empirical demonstration of this superiority is produced, which is a bit disappointing.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper introduces a variant of the Information Bottleneck (IB) framework, which consists in permuting the conditional probabilities of y given x and y given \\hat{x} in a Kullback-Liebler divergence involved in the IB optimization criterion.\n\nInterestingly, this change only results in changing an arithmetic mean into a geometric mean in the algorithmic resolution.\n\nGood properties of the exponential families (existence of non-trivial minimal sufficient statistics) are preserved, and an analysis of the new critical points/information plane induced is carried out.\n\nThe paper is globally well written and clear, and the maths are rigorously introduced and treated. Two minor comments would concern the definition of the Mutual Information (MI), which could be recalled to help the unfamiliar reader and improve self-containedness, and the notation of the expected value (\\langle \\rangle_p), unusual in Machine Learning where \\mathbb{E} is often preferred.\n\nAnother point that could be enhanced is the intuition behind the IB and dualIB criteria: a small comment on their meaning/relevance as well as the rationale/implication of the probabilities permutation would be valuable addition.\n\nExhibiting a link with variational autoencoders (VAEs), and expliciting the differences in a VAE framework between the two criteria could also represent an interesting parallel, especially for machine learning oriented readers.\n\nOne of the main drawback of the present paper is however the lack of convincing experiments. Graphs showing the good behavior of the introduced framework are ok, but the clear interest of using dualIB rather than IB could be emphasized more. In particular, it does not seem that the issues about IB raised in the abstract (curse of dimensionality, no closed form solution) are solved with dualIB. Furthermore, dualIB optimizes the MI of the predicted labels, which is claimed to be beneficial contrary to the MI on the actual labels. However, no empirical demonstration of this superiority is produced, which is a bit disappointing."}, "tcdate": 1571909755968}, {"id": "Byl0gW8RFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1755/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new \"dual\" variant of the Information Bottleneck framework. The IB framework has been the subject of many papers in past years, with a focus on understanding the inner workings of deep learning. The framework poses machine learning as optimizing an internal representation to tradeoff between retaining less information about the input features and retaining more information about the output label (prediction). The existing framework measures the retained information about the prediction via mutual information, which can be expressed as a KL divergence. The new dual framework reverses the arguments of this divergence.\n\nThe paper shows that this dual IB closely mirrors the original IB while having several additional nice mathematical properties. In particular,  it can be shown that for exponential families the dual IB representation retains the exponential form.\n\nOverall, I think this paper adds a meaningful new perspective to the IB framework and the analysis appears to be thorough. As such, I support acceptance. \n\nThe paper could be improved by giving more interpretation of the formal results and experiments - i.e. tying this framework back to the higher-level questions and explaining what the quantities mean.\n\nFigures 1, 2 part a: The meaning of both axes is unclear. The horizontal axis beta is a lagrangian parameter with no meaning outside the framework. The vertical axis Pr[y=0|\\hat x] has no semantics since the problem has not been defined. What is the reader meant to take away from these figures?\n\nEquation 3, part i: The numerator should have a beta subscript. The meaning of the denominator is not clear (it is a normalizing constant).\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a new \"dual\" variant of the Information Bottleneck framework. The IB framework has been the subject of many papers in past years, with a focus on understanding the inner workings of deep learning. The framework poses machine learning as optimizing an internal representation to tradeoff between retaining less information about the input features and retaining more information about the output label (prediction). The existing framework measures the retained information about the prediction via mutual information, which can be expressed as a KL divergence. The new dual framework reverses the arguments of this divergence.\n\nThe paper shows that this dual IB closely mirrors the original IB while having several additional nice mathematical properties. In particular,  it can be shown that for exponential families the dual IB representation retains the exponential form.\n\nOverall, I think this paper adds a meaningful new perspective to the IB framework and the analysis appears to be thorough. As such, I support acceptance. \n\nThe paper could be improved by giving more interpretation of the formal results and experiments - i.e. tying this framework back to the higher-level questions and explaining what the quantities mean.\n\nFigures 1, 2 part a: The meaning of both axes is unclear. The horizontal axis beta is a lagrangian parameter with no meaning outside the framework. The vertical axis Pr[y=0|\\hat x] has no semantics since the problem has not been defined. What is the reader meant to take away from these figures?\n\nEquation 3, part i: The numerator should have a beta subscript. The meaning of the denominator is not clear (it is a normalizing constant).\n\n"}, "tcdate": 1571868917667}], "openreview_url": "https://openreview.net/forum?id=B1xZD1rtPr", "arxiv_id": "2006.04641", "paper_pdf": "papers/B1xZD1rtPr.pdf", "paper_pdf_sha256": "0036af0f6acc1d04c71bb1750bbdacf0b050b30baaf0f2e13b450e4aeab132f3", "paper_pdf_bytes": 2153987, "paper_pdf_source": "openreview", "code_url": "https://github.com/ravidziv/dual_IB", "code_repository": "ravidziv/dual_IB", "code_commit": "a3c8ae96a0581c467c00241b0fa7da7b7d50ddaa", "code_archive": "repos/B1xZD1rtPr.zip", "code_archive_sha256": "ee5316ab1067489f9f70b0ca73e89a3b5eb851070d69a04a54c955d41117ae1e", "code_archive_bytes": 81260, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 1291, "github_languages": {"Python": 46844}, "github_archived": false, "github_pushed_at": "2024-10-22T14:31:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-dual-information-bottleneck-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PTTmPHS7OE", "year": 2026, "status": "rejected", "title": "LocationReasoner: Evaluating LLMs on Real-World Site Selection Reasoning", "authors": ["Miho Koda", "Yu Zheng", "Ruixian Ma", "Mingyang Sun", "Devesh Pansare", "Fabio Duarte", "Paolo Santi"], "authorids": ["~Miho_Koda1", "~Yu_Zheng7", "~Ruixian_Ma1", "~Mingyang_Sun7", "~Devesh_Pansare1", "~Fabio_Duarte1", "~Paolo_Santi2"], "authors_source": "OpenReview API", "abstract": "Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation, leaving open the question of whether such reasoning skills generalize to complex real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs’ reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistic constraints. The benchmark covers carefully crafted queries of varying\ndifficulty levels and is supported by a sandbox environment with in-house tools for constraint-based location search. Automated verification further guarantees the scalability of the benchmark, enabling the addition of arbitrary number of queries. Extensive evaluations on real-world site selection data from Boston, New York, and Tampa reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://anonymous.4open.science/r/LocationReasoner-DC5D.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "hfJCxazxur", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18432/Reviewer_P7C5"], "rating": 2, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "The paper proposes a benchmark for evaluating how LLMs can reason over code on geospatial tasks, specifically for location and site selection given some specific constraints. The LLM is also given access to a library of functions/tools that it can use to generate the final code and the solution. The paper integrates multiple interesting datasets such as SafeGraph (which has information about points of interest, parking facilities, and consumer spending patterns), Google Places API which has information about demographics and population, and transportation data from OpenStreetMap. They also build a sandbox environment with a library of functions that can load, analyze, and filter the data in different ways. They additionally employ LLMs to construct the location selection queries, based on some set of logical constraints and then have the LLM phrase those queries in natural language. The actual task would then be to go from natural language to perhaps rediscover the logical constraints and produce the solution by generating and executing code using the given library. They classify the query as simple, medium or hard based on the number of constraints. \nThe paper created 316 queries and evaluated the performance of several frontier LLMs on this task with and without \"thinking\"/\"reasoning\" enables, and also evaluate if prompt engineering frameworks like React or Reflexion are helpful. The best performance (perfect pass rate) was by GPT-o4-mini which got ~70%.", "review_text": "The paper proposes a benchmark for evaluating how LLMs can reason over code on geospatial tasks, specifically for location and site selection given some specific constraints. The LLM is also given access to a library of functions/tools that it can use to generate the final code and the solution. The paper integrates multiple interesting datasets such as SafeGraph (which has information about points of interest, parking facilities, and consumer spending patterns), Google Places API which has information about demographics and population, and transportation data from OpenStreetMap. They also build a sandbox environment with a library of functions that can load, analyze, and filter the data in different ways. They additionally employ LLMs to construct the location selection queries, based on some set of logical constraints and then have the LLM phrase those queries in natural language. The actual task would then be to go from natural language to perhaps rediscover the logical constraints and produce the solution by generating and executing code using the given library. They classify the query as simple, medium or hard based on the number of constraints. \nThe paper created 316 queries and evaluated the performance of several frontier LLMs on this task with and without \"thinking\"/\"reasoning\" enables, and also evaluate if prompt engineering frameworks like React or Reflexion are helpful. The best performance (perfect pass rate) was by GPT-o4-mini which got ~70%.", "strengths": "* The task of Location selection specifically framed to generate and execute code given a set of library functions seems like a good reasoning test bed.\n* The datasets integrated to create the task are rich and include a lot of diverse information such as parking facilities, consumer spending patterns (from SafeGraph), information about demographics and population (from Google Places API), and transportation data (from OpenStreetMap).\n* The experimental set up and evaluations are standard and straightforward.\n* The error analysis is reasonable. Some more constructive ways of improvement would have been nicer to see.", "weaknesses": "1. While it is possible to understand the high level idea, the paper is lacking in clarity and examples to clearly describe the different contributions. Specifically,\n\n  1a. There is no example the fully illustrates the input query, the expected final solution and an example of the code that would generate the final solution. In particular, **lacking the outputs and expected code solution** (even in the supplement) makes it hard to get a feel for the complexity of the task. Section 2.2, 6, and A3 all give glimpses of the task but not quite a full example.\n\n2. **Query generation lacks clarity** Section 2.2 on Query design, does not clearly describe the details of how the queries were actually generated, and how the solutions were generated and verified. \n\n  2a. What are some examples of the rule-based generation? What is the algorithm that was used? \n\n  2b. What were the prompts for the LLM-based generation? What specific steps were involved in the LLM-based generation. Here again examples would be immensely helpful.  \n\n  2c. Are the rule-based queries substantially different from LLM-generated queries? In what way? \n\n  2d. How is the model performance when sliced by how the queries were generated, is there a difference? \n\n  2e. Which LLM was used for LLM-generated queries? (also how and why did you choose that LLM model to do the generation?) \n\n  2f. Also, how did you generate the ground truth solutions/code? Was there any human annotation? How were the ground truth solutions verified for correctness?\n\n3. **Missing details on characteristics of the problem**. The evaluation appears to be on 316 generated queries (and solutions) on 3 cities. Section 4 argues that 150-200 queries are sufficient to test robustness of performance for a city. Is there something about the location selection problem that makes it different for each geographical region being studied? Are there any constraints to where your data generation method (and library) may be applicable? If it is easy to scale this, why not scale the questions? What more would we learn by scaling the questions / what are we missing if we don't scale? \n\n4. Another major weakness is **lack of comparison to other geospatial code generation tasks and benchmarks**, and how this task and tooling is different and what this benchmark tests that is different from others geospatial code generation benchmarks. The paper references a range of generic reasoning and LLM+Agents related papers, but it seems to be missing more closely relevant and related works, Here are some geo code benchmarks I am familiar with:\n\n[1] GeoCode Eval, GeoCode-GPT: A large language model for geospatial code generation\n\n[2] Geollm-engine: A realistic environment for building geospatial copilots\n\n[3] Multi-Agent Geospatial Copilots for Remote Sensing Workflows\n\n[4] The Cloud-Based Geospatial Benchmark: Challenges and LLM Evaluation\n\n[5] Evaluation of code llms on geospatial code generation\n\n[6] An llm agent for automatic geospatial data analysis\n\n[7] GIS copilot: Towards an autonomous GIS agent for spatial analysis\n\n\n* Overall it seems the paper has some interesting contributions, but it's not clear how substantial these are. The lack of clarity and detailed information highlighted in weaknesses 1-4 make it difficult to determine the contribution have resulted low scores on presentation and contribution.", "questions": "* Please address questions under weaknesses\n\n* Are there other domain related characterizations of the dataset (beyond the simple, medium, hard classification based on the complexity of the constraints)? E.g. a couple of the queries in section 2.2. pertained to \"restaurant\" site selection. If there are such domain related characterizations, based on the characterization and the query are there more domain specific nuances / \"knowledge\" that the model would have to know to apply the right constraints? -- E.g. in residential areas some zoning laws may apply whereas for commercial land some other laws might apply. Could you capture such information in the dataset description?\n\n**Other comments**\n\n* Presentation of results in Table 2 can be improved. It is very difficult to parse and identify which models are performing well and where. Perhaps just have a much smaller table of just the overall performance and move the details to the supplement or consider generating plots.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a benchmark for evaluating how LLMs can reason over code on geospatial tasks, specifically for location and site selection given some specific constraints. The LLM is also given access to a library of functions/tools that it can use to generate the final code and the solution. The paper integrates multiple interesting datasets such as SafeGraph (which has information about points of interest, parking facilities, and consumer spending patterns), Google Places API which has information about demographics and population, and transportation data from OpenStreetMap. They also build a sandbox environment with a library of functions that can load, analyze, and filter the data in different ways. They additionally employ LLMs to construct the location selection queries, based on some set of logical constraints and then have the LLM phrase those queries in natural language. The actual task would then be to go from natural language to perhaps rediscover the logical constraints and produce the solution by generating and executing code using the given library. They classify the query as simple, medium or hard based on the number of constraints. \nThe paper created 316 queries and evaluated the performance of several frontier LLMs on this task with and without \"thinking\"/\"reasoning\" enables, and also evaluate if prompt engineering frameworks like React or Reflexion are helpful. The best performance (perfect pass rate) was by GPT-o4-mini which got ~70%.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "* The task of Location selection specifically framed to generate and execute code given a set of library functions seems like a good reasoning test bed.\n* The datasets integrated to create the task are rich and include a lot of diverse information such as parking facilities, consumer spending patterns (from SafeGraph), information about demographics and population (from Google Places API), and transportation data (from OpenStreetMap).\n* The experimental set up and evaluations are standard and straightforward.\n* The error analysis is reasonable. Some more constructive ways of improvement would have been nicer to see.", "weaknesses": "1. While it is possible to understand the high level idea, the paper is lacking in clarity and examples to clearly describe the different contributions. Specifically,\n\n  1a. There is no example the fully illustrates the input query, the expected final solution and an example of the code that would generate the final solution. In particular, **lacking the outputs and expected code solution** (even in the supplement) makes it hard to get a feel for the complexity of the task. Section 2.2, 6, and A3 all give glimpses of the task but not quite a full example.\n\n2. **Query generation lacks clarity** Section 2.2 on Query design, does not clearly describe the details of how the queries were actually generated, and how the solutions were generated and verified. \n\n  2a. What are some examples of the rule-based generation? What is the algorithm that was used? \n\n  2b. What were the prompts for the LLM-based generation? What specific steps were involved in the LLM-based generation. Here again examples would be immensely helpful.  \n\n  2c. Are the rule-based queries substantially different from LLM-generated queries? In what way? \n\n  2d. How is the model performance when sliced by how the queries were generated, is there a difference? \n\n  2e. Which LLM was used for LLM-generated queries? (also how and why did you choose that LLM model to do the generation?) \n\n  2f. Also, how did you generate the ground truth solutions/code? Was there any human annotation? How were the ground truth solutions verified for correctness?\n\n3. **Missing details on characteristics of the problem**. The evaluation appears to be on 316 generated queries (and solutions) on 3 cities. Section 4 argues that 150-200 queries are sufficient to test robustness of performance for a city. Is there something about the location selection problem that makes it different for each geographical region being studied? Are there any constraints to where your data generation method (and library) may be applicable? If it is easy to scale this, why not scale the questions? What more would we learn by scaling the questions / what are we missing if we don't scale? \n\n4. Another major weakness is **lack of comparison to other geospatial code generation tasks and benchmarks**, and how this task and tooling is different and what this benchmark tests that is different from others geospatial code generation benchmarks. The paper references a range of generic reasoning and LLM+Agents related papers, but it seems to be missing more closely relevant and related works, Here are some geo code benchmarks I am familiar with:\n\n[1] GeoCode Eval, GeoCode-GPT: A large language model for geospatial code generation\n\n[2] Geollm-engine: A realistic environment for building geospatial copilots\n\n[3] Multi-Agent Geospatial Copilots for Remote Sensing Workflows\n\n[4] The Cloud-Based Geospatial Benchmark: Challenges and LLM Evaluation\n\n[5] Evaluation of code llms on geospatial code generation\n\n[6] An llm agent for automatic geospatial data analysis\n\n[7] GIS copilot: Towards an autonomous GIS agent for spatial analysis\n\n\n* Overall it seems the paper has some interesting contributions, but it's not clear how substantial these are. The lack of clarity and detailed information highlighted in weaknesses 1-4 make it difficult to determine the contribution have resulted low scores on presentation and contribution.", "questions": "* Please address questions under weaknesses\n\n* Are there other domain related characterizations of the dataset (beyond the simple, medium, hard classification based on the complexity of the constraints)? E.g. a couple of the queries in section 2.2. pertained to \"restaurant\" site selection. If there are such domain related characterizations, based on the characterization and the query are there more domain specific nuances / \"knowledge\" that the model would have to know to apply the right constraints? -- E.g. in residential areas some zoning laws may apply whereas for commercial land some other laws might apply. Could you capture such information in the dataset description?\n\n**Other comments**\n\n* Presentation of results in Table 2 can be improved. It is very difficult to parse and identify which models are performing well and where. Perhaps just have a much smaller table of just the overall performance and move the details to the supplement or consider generating plots.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761988000279}, {"id": "J6YCkt2mp2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18432/Reviewer_KPTZ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduced a new benchmark for evaluating LLMs' capability in location selection under various constraints. The authors have conducted extensive evaluations of various types of LLMs with different settings (general/reasoning). Some insightful analysis is also presented to guide future research.", "review_text": "This paper introduced a new benchmark for evaluating LLMs' capability in location selection under various constraints. The authors have conducted extensive evaluations of various types of LLMs with different settings (general/reasoning). Some insightful analysis is also presented to guide future research.", "strengths": "- The experimental evaluation in this paper is extensive.\n- The attribute-based analysis is very useful for highlighting the key challenge.\n- The paper is well-written.", "weaknesses": "**Some related works need to be discussed more extensively**\n\nIn general, I think the problem of site selection is a constraint satisfaction problem, and there have been benchmarks for such kinds of problems, for example, travel planning [1] and scene/agent-intention understanding [2]. I wonder what makes site selection especially challenging compared to a broader constraint-satisfaction benchmark literature? The authors should discuss these more extensively in the related works section.\n\n**Is this benchmark not challenging enough?**\n\nThe extensive evaluation in this paper is well-acknowledged. But I notice that even GPT4o has already achieved 37% success in hard mode, it is a relatively old model in the fast development of LLMs, how about GPT5? I am a bit concerned that the difficulty of this benchmark is not high enough that it might be saturated soon (is it easy to create harder problems using the mentioned data generation pipeline?). Also, what's the human performance on this benchmark? Providing that will better help readers understand the current limitations of LLMs.\n\n[1] Ju, Da, et al. \"To the globe (ttg): Towards language-driven guaranteed travel planning.\" arXiv preprint arXiv:2410.16456 (2024).\n[2] Li, Bowen, et al. \"LogiCity: Advancing neuro-symbolic ai with abstract urban simulation.\" Advances in Neural Information Processing Systems 37 (2024): 69840-69864.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduced a new benchmark for evaluating LLMs' capability in location selection under various constraints. The authors have conducted extensive evaluations of various types of LLMs with different settings (general/reasoning). Some insightful analysis is also presented to guide future research.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The experimental evaluation in this paper is extensive.\n- The attribute-based analysis is very useful for highlighting the key challenge.\n- The paper is well-written.", "weaknesses": "**Some related works need to be discussed more extensively**\n\nIn general, I think the problem of site selection is a constraint satisfaction problem, and there have been benchmarks for such kinds of problems, for example, travel planning [1] and scene/agent-intention understanding [2]. I wonder what makes site selection especially challenging compared to a broader constraint-satisfaction benchmark literature? The authors should discuss these more extensively in the related works section.\n\n**Is this benchmark not challenging enough?**\n\nThe extensive evaluation in this paper is well-acknowledged. But I notice that even GPT4o has already achieved 37% success in hard mode, it is a relatively old model in the fast development of LLMs, how about GPT5? I am a bit concerned that the difficulty of this benchmark is not high enough that it might be saturated soon (is it easy to create harder problems using the mentioned data generation pipeline?). Also, what's the human performance on this benchmark? Providing that will better help readers understand the current limitations of LLMs.\n\n[1] Ju, Da, et al. \"To the globe (ttg): Towards language-driven guaranteed travel planning.\" arXiv preprint arXiv:2410.16456 (2024).\n[2] Li, Bowen, et al. \"LogiCity: Advancing neuro-symbolic ai with abstract urban simulation.\" Advances in Neural Information Processing Systems 37 (2024): 69840-69864.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761964397444}, {"id": "FGZlyI275I", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18432/Reviewer_2Tbz"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This paper proposes LocationReasoner, a benchmark to evaluate LLMs’ reasoning abilities on real-world site selection problems. The motivation is that current reasoning models are mostly tested on math or coding problems, leaving their practical reasoning unclear. The authors create a benchmark with queries of different difficulty levels and provide a sandbox with automated verification using real city data. Results show limited performance: even the best model (OpenAI o4-mini) achieves only 69.99% perfect pass rate. Ablation studies suggest reasoning helps somewhat on medium-difficulty queries, and better prompting gives small gains.", "review_text": "This paper proposes LocationReasoner, a benchmark to evaluate LLMs’ reasoning abilities on real-world site selection problems. The motivation is that current reasoning models are mostly tested on math or coding problems, leaving their practical reasoning unclear. The authors create a benchmark with queries of different difficulty levels and provide a sandbox with automated verification using real city data. Results show limited performance: even the best model (OpenAI o4-mini) achieves only 69.99% perfect pass rate. Ablation studies suggest reasoning helps somewhat on medium-difficulty queries, and better prompting gives small gains.", "strengths": "1. The dataset is practical and scalable. Automated query generation and verification using real city data is reproducible.\n\n2. This paper provides comprehensive evaluation: Multiple LLM families, agentic strategies, and cities are tested, with detailed error analysis.\n\n3. They also provide insightful findings on agentic strategies. For example, they discover that ReAct and Reflexion can hurt performance because of over-reasoning, which is surprising and useful for future research.", "weaknesses": "I am not an expert in this domain. Therefore, my concerns are raised based on high-level research perspective instead of specific task perspective.\n\n1. It’s unclear if the poor performance of Agentic methods is due to the base LLM limitations or Agentic workflow design. Showing their reasoning traces and the planner (LLM) output would help clarify this.\n\n2. The fixed 15 in-house tools might restrict performance. It would be good to test more general tools (like Python libraries) and see which failures are due to tools versus reasoning.\n\n3. How do we know “hard” queries are really harder? Some empirical validation, like human expert performance or inter-annotator agreement, would strengthen this.\n\n4. Data leakage concern: Since datasets are public and LLMs may have seen similar data, can the authors check performance on synthetic or fictional locations to rule out contamination?", "questions": "Please see my questions raised in each weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes LocationReasoner, a benchmark to evaluate LLMs’ reasoning abilities on real-world site selection problems. The motivation is that current reasoning models are mostly tested on math or coding problems, leaving their practical reasoning unclear. The authors create a benchmark with queries of different difficulty levels and provide a sandbox with automated verification using real city data. Results show limited performance: even the best model (OpenAI o4-mini) achieves only 69.99% perfect pass rate. Ablation studies suggest reasoning helps somewhat on medium-difficulty queries, and better prompting gives small gains.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The dataset is practical and scalable. Automated query generation and verification using real city data is reproducible.\n\n2. This paper provides comprehensive evaluation: Multiple LLM families, agentic strategies, and cities are tested, with detailed error analysis.\n\n3. They also provide insightful findings on agentic strategies. For example, they discover that ReAct and Reflexion can hurt performance because of over-reasoning, which is surprising and useful for future research.", "weaknesses": "I am not an expert in this domain. Therefore, my concerns are raised based on high-level research perspective instead of specific task perspective.\n\n1. It’s unclear if the poor performance of Agentic methods is due to the base LLM limitations or Agentic workflow design. Showing their reasoning traces and the planner (LLM) output would help clarify this.\n\n2. The fixed 15 in-house tools might restrict performance. It would be good to test more general tools (like Python libraries) and see which failures are due to tools versus reasoning.\n\n3. How do we know “hard” queries are really harder? Some empirical validation, like human expert performance or inter-annotator agreement, would strengthen this.\n\n4. Data leakage concern: Since datasets are public and LLMs may have seen similar data, can the authors check performance on synthetic or fictional locations to rule out contamination?", "questions": "Please see my questions raised in each weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761927834491}, {"id": "cgBiMyNUlu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18432/Reviewer_SoL7"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper proposes a new benchmark to evaluate multi-step reasoning capabilities of llms in site selection scenarios. This requires the model to understand geospatial cues, use different tools, think logically and come up with the final answer. The paper contributes to the existing benchmarks for understanding the capabilities and limitations of llms particularly reasoning.", "review_text": "The paper proposes a new benchmark to evaluate multi-step reasoning capabilities of llms in site selection scenarios. This requires the model to understand geospatial cues, use different tools, think logically and come up with the final answer. The paper contributes to the existing benchmarks for understanding the capabilities and limitations of llms particularly reasoning.", "strengths": "- The proposed benchmark is carefully designed to include a diverse range of problems (easy, medium, hard)\n- The benchmark tackles problems around tool use, reasoning chains, agentic pipelines, and structured output - all of which are essential in evaluating model capabilities.\n- The evaluation shows expected results between difference of reasoning and non-reasoning models. \n- The benchmark can be easily extended to include more data points and introduce further complexity.\n- The post analysis of model responses provides a lot of insight into the thought process and problems current llms face in challenging agentic tasks.", "weaknesses": "- The dataset only includes a few cities in America. The authors should try to extend it to encompass more geographical diversity and analyze whether llms have any location bias", "questions": "- How many data points are in the benchmark at this time?\n- Could you explain how the error analysis was done and how model responses were classified into the different groups? Was this done manually by going through every response or in some automated way?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new benchmark to evaluate multi-step reasoning capabilities of llms in site selection scenarios. This requires the model to understand geospatial cues, use different tools, think logically and come up with the final answer. The paper contributes to the existing benchmarks for understanding the capabilities and limitations of llms particularly reasoning.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "- The proposed benchmark is carefully designed to include a diverse range of problems (easy, medium, hard)\n- The benchmark tackles problems around tool use, reasoning chains, agentic pipelines, and structured output - all of which are essential in evaluating model capabilities.\n- The evaluation shows expected results between difference of reasoning and non-reasoning models. \n- The benchmark can be easily extended to include more data points and introduce further complexity.\n- The post analysis of model responses provides a lot of insight into the thought process and problems current llms face in challenging agentic tasks.", "weaknesses": "- The dataset only includes a few cities in America. The authors should try to extend it to encompass more geographical diversity and analyze whether llms have any location bias", "questions": "- How many data points are in the benchmark at this time?\n- Could you explain how the error analysis was done and how model responses were classified into the different groups? Was this done manually by going through every response or in some automated way?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761915798912}, {"id": "yWYseyB9DQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18432/Reviewer_qSv8"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces LocationReasoner, a benchmark for evaluating LLMs on real-world site-selection reasoning. It provides (i) a sandbox with fixed, offline datasets of various cities and a set of “in-house” tools for constraint-based queries; (ii) automated query generation (rule-based and LLM-based) and deterministic verification; and (iii) evaluations across multiple LLM families and agentic workflows (ReAct, Reflexion). On these datasets the average pass rate for various LLMs is quite low and agentic methods often do worse than direct prompting. The authors analyze failure types (logic, edge cases, tool misuse, prompt misread, code errors) and argue that holistic, non-linear planning via direct code generation exceeds step-wise agentic loops.", "review_text": "The paper introduces LocationReasoner, a benchmark for evaluating LLMs on real-world site-selection reasoning. It provides (i) a sandbox with fixed, offline datasets of various cities and a set of “in-house” tools for constraint-based queries; (ii) automated query generation (rule-based and LLM-based) and deterministic verification; and (iii) evaluations across multiple LLM families and agentic workflows (ReAct, Reflexion). On these datasets the average pass rate for various LLMs is quite low and agentic methods often do worse than direct prompting. The authors analyze failure types (logic, edge cases, tool misuse, prompt misread, code errors) and argue that holistic, non-linear planning via direct code generation exceeds step-wise agentic loops.", "strengths": "- Timely and unique benchmark and planning domain: Targets a gap between math/code reasoning and practical, multi-constraint decision-making; automated verification at scale is valuable. The urban planning domain is interesting.\n- Clear environment design: Offline sandbox, fixed toolset, and concrete tool taxonomy (loaders, zone, analysis, filters, population) make the task reproducible and interpretable and code is provided\n- Broad model coverage + difficulty controls: Multiple provider families, reasoning vs. non-reasoning variants, and simple/medium/hard splits are aligned with standard practice in the LLM + planning and sequential decision making literature so that is nice; results are consistent across cities with convergence analysis.\n- Useful error analysis: Breaks down failures (logic, edge, tool, prompt, code), gives concrete examples, and quantifies “correction rates” for reasoning models which is nice for Reflexion and React,", "weaknesses": "- External validity / coverage: The geography is limited to three U.S. cities with specific data sources (SafeGraph, OSM, Google Places). It’s unclear how well conclusions generalize to non-U.S., rural, or low-data regions and to domains beyond retail-like POI reasoning. I have provided some Urban Planning datasets in the question section which may be relevant to this.\n- Agentic fairness and scope. Agentic workflows are only tested on GPT-4o, while direct prompting includes multiple model families. This asymmetry makes it hard to conclude that “agentic strategies don’t help” in general rather than “agentic strategies built on this one base model + prompts didn’t help.” Broaden agentic experiments to at least one reasoning-tuned model per family, and report hyperparameters, step limits, memory policies, error-recovery prompts in a way that supports apples-to-apples comparisons.\n- There are more recent agentic strategies such as Self Discover [1] and LATS [2], it would have been interesting to see experiments on those for greater coverage\n- Attribution of the central claim. The paper argues “direct code generation (holistic, non-linear) performs better than ReAct (chunked, linear).” While plausible and supported by the results, the causality is mostly qualitative. Please add controlled ablations ie: same base model, matched temperature/decoding, code gen with enforced step-limits vs. ReAct with global-plan scaffolds; and instrumentation of failure transitions. The section explaining the \n- While the dataset and evaluation harness are useful, the pipeline itself appears to be a composition of existing LLMs and standard agentic patterns (e.g., ReAct/Reflexion) wrapped around tool calls. It is not clear to me what exactly is the novel algorithmic component, planning formalism, or execution/control innovation—so the novelty seems concentrated in the benchmark \n- If the main novelty indeed is the benchmark and the author's claim is that this is an example of a dataset where significant innovation is possible in the designing of new research to fix issues such as over-thinking or excessive tool usage/lack of convergence (as in React), please reposition claims accordingly and emphasize the benchmark’s design principles, reliability, and long-term value - for example providing a discussion or ablation studies to understand why LLMs-even with Reflexion/React-perform so poorly on this specific dataset and what could possibly be done to improve performance\n\n\n[1] https://arxiv.org/abs/2402.03620\n\n[2] https://arxiv.org/pdf/2310.04406", "questions": "- Tool/ground-truth coupling: Can you evaluate an agent that does not have access to the exact toolset used by the verifier (e.g., renamed arguments, perturbed APIs, or a mapping layer) to test reasoning robustness beyond tool-specific affordance learning?\n- There are some other prior datasets for urban planning in the ML literature - for example: UrbanDataLayer: A Unified Data Pipeline\nfor Urban Science [1] and some prior works on LLMs for urban planning [2] [3] [4]. Could the authors speak to how the dataset used in this paper compares to the Urban Science dataset as well as some prior LLM based approaches in urban planning? \n- Why were ReAct/Reflexion run only with GPT-4o? Do results hold with o4-mini, Gemini 2.5, or DeepSeek-R1 as the base? \n- Since spend data spans 2019–2025, would it be possible to clarify whether time-aware splits or year-withheld constraints?\n- Given the limitations the authors have noticed on their datasets across several language models perhaps the following works would be of interest regarding the limitations of popular agentic frameworks as in plan-and-execute frameworks, Interactive reasoning frameworks, self-refinement frameworks: [5] [6] \n\n\n[1] https://proceedings.neurips.cc/paper_files/paper/2024/file/0db7f135f6991e8cec5e516ecc66bfba-Paper-Datasets_and_Benchmarks_Track.pdf\n\n[2] https://arxiv.org/abs/2402.17161\n\n[3] https://www.nature.com/articles/s44284-025-00261-7\n\n[4] https://arxiv.org/abs/2406.13945\n\n[5] https://proceedings.neurips.cc/paper_files/paper/2024/file/fa080fe0f218871faec1d8ba20e491d5-Paper-Conference.pdf\n\n[6] https://arxiv.org/abs/2408.11326", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces LocationReasoner, a benchmark for evaluating LLMs on real-world site-selection reasoning. It provides (i) a sandbox with fixed, offline datasets of various cities and a set of “in-house” tools for constraint-based queries; (ii) automated query generation (rule-based and LLM-based) and deterministic verification; and (iii) evaluations across multiple LLM families and agentic workflows (ReAct, Reflexion). On these datasets the average pass rate for various LLMs is quite low and agentic methods often do worse than direct prompting. The authors analyze failure types (logic, edge cases, tool misuse, prompt misread, code errors) and argue that holistic, non-linear planning via direct code generation exceeds step-wise agentic loops.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Timely and unique benchmark and planning domain: Targets a gap between math/code reasoning and practical, multi-constraint decision-making; automated verification at scale is valuable. The urban planning domain is interesting.\n- Clear environment design: Offline sandbox, fixed toolset, and concrete tool taxonomy (loaders, zone, analysis, filters, population) make the task reproducible and interpretable and code is provided\n- Broad model coverage + difficulty controls: Multiple provider families, reasoning vs. non-reasoning variants, and simple/medium/hard splits are aligned with standard practice in the LLM + planning and sequential decision making literature so that is nice; results are consistent across cities with convergence analysis.\n- Useful error analysis: Breaks down failures (logic, edge, tool, prompt, code), gives concrete examples, and quantifies “correction rates” for reasoning models which is nice for Reflexion and React,", "weaknesses": "- External validity / coverage: The geography is limited to three U.S. cities with specific data sources (SafeGraph, OSM, Google Places). It’s unclear how well conclusions generalize to non-U.S., rural, or low-data regions and to domains beyond retail-like POI reasoning. I have provided some Urban Planning datasets in the question section which may be relevant to this.\n- Agentic fairness and scope. Agentic workflows are only tested on GPT-4o, while direct prompting includes multiple model families. This asymmetry makes it hard to conclude that “agentic strategies don’t help” in general rather than “agentic strategies built on this one base model + prompts didn’t help.” Broaden agentic experiments to at least one reasoning-tuned model per family, and report hyperparameters, step limits, memory policies, error-recovery prompts in a way that supports apples-to-apples comparisons.\n- There are more recent agentic strategies such as Self Discover [1] and LATS [2], it would have been interesting to see experiments on those for greater coverage\n- Attribution of the central claim. The paper argues “direct code generation (holistic, non-linear) performs better than ReAct (chunked, linear).” While plausible and supported by the results, the causality is mostly qualitative. Please add controlled ablations ie: same base model, matched temperature/decoding, code gen with enforced step-limits vs. ReAct with global-plan scaffolds; and instrumentation of failure transitions. The section explaining the \n- While the dataset and evaluation harness are useful, the pipeline itself appears to be a composition of existing LLMs and standard agentic patterns (e.g., ReAct/Reflexion) wrapped around tool calls. It is not clear to me what exactly is the novel algorithmic component, planning formalism, or execution/control innovation—so the novelty seems concentrated in the benchmark \n- If the main novelty indeed is the benchmark and the author's claim is that this is an example of a dataset where significant innovation is possible in the designing of new research to fix issues such as over-thinking or excessive tool usage/lack of convergence (as in React), please reposition claims accordingly and emphasize the benchmark’s design principles, reliability, and long-term value - for example providing a discussion or ablation studies to understand why LLMs-even with Reflexion/React-perform so poorly on this specific dataset and what could possibly be done to improve performance\n\n\n[1] https://arxiv.org/abs/2402.03620\n\n[2] https://arxiv.org/pdf/2310.04406", "questions": "- Tool/ground-truth coupling: Can you evaluate an agent that does not have access to the exact toolset used by the verifier (e.g., renamed arguments, perturbed APIs, or a mapping layer) to test reasoning robustness beyond tool-specific affordance learning?\n- There are some other prior datasets for urban planning in the ML literature - for example: UrbanDataLayer: A Unified Data Pipeline\nfor Urban Science [1] and some prior works on LLMs for urban planning [2] [3] [4]. Could the authors speak to how the dataset used in this paper compares to the Urban Science dataset as well as some prior LLM based approaches in urban planning? \n- Why were ReAct/Reflexion run only with GPT-4o? Do results hold with o4-mini, Gemini 2.5, or DeepSeek-R1 as the base? \n- Since spend data spans 2019–2025, would it be possible to clarify whether time-aware splits or year-withheld constraints?\n- Given the limitations the authors have noticed on their datasets across several language models perhaps the following works would be of interest regarding the limitations of popular agentic frameworks as in plan-and-execute frameworks, Interactive reasoning frameworks, self-refinement frameworks: [5] [6] \n\n\n[1] https://proceedings.neurips.cc/paper_files/paper/2024/file/0db7f135f6991e8cec5e516ecc66bfba-Paper-Datasets_and_Benchmarks_Track.pdf\n\n[2] https://arxiv.org/abs/2402.17161\n\n[3] https://www.nature.com/articles/s44284-025-00261-7\n\n[4] https://arxiv.org/abs/2406.13945\n\n[5] https://proceedings.neurips.cc/paper_files/paper/2024/file/fa080fe0f218871faec1d8ba20e491d5-Paper-Conference.pdf\n\n[6] https://arxiv.org/abs/2408.11326", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761110227452}], "openreview_url": "https://openreview.net/forum?id=PTTmPHS7OE", "arxiv_id": "2506.13841", "paper_pdf": "papers/PTTmPHS7OE.pdf", "paper_pdf_sha256": "17f13fb35126f8d12d59a61c92c3214b22689d4d6a49b9b810203f8bb9bd49ee", "paper_pdf_bytes": 341084, "paper_pdf_source": "openreview", "code_url": "https://github.com/miho-koda/LocationReasoner", "code_repository": "miho-koda/LocationReasoner", "code_commit": "54c643e3cad533325238530840e3146556e59b24", "code_archive": "repos/PTTmPHS7OE.zip", "code_archive_sha256": "9e1d4e87f3760a955f533ddd5d09b86d311e6f9e225563fe6ca5ef1fef33672b", "code_archive_bytes": 249029, "code_file_count": 89, "code_extensions": {".py": 87, ".ipynb": 2}, "github_disk_usage_kb": 2087, "github_languages": {"Python": 851288, "Jupyter Notebook": 44029}, "github_archived": false, "github_pushed_at": "2026-07-13T02:00:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/locationreasoner-evaluating-llms-on-real"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YGflij9S6x", "year": 2025, "status": "rejected", "title": "Continual Learning: Less Forgetting, More OOD Generalization via Adaptive Contrastive Replay", "authors": ["Hossein Rezaei", "Mohammad Khalooei", "Mohammad Sabokrou"], "authorids": ["~Hossein_Rezaei1", "~Mohammad_Khalooei2", "~Mohammad_Sabokrou1"], "authors_source": "OpenReview API", "abstract": "Machine learning models often suffer from catastrophic forgetting of previously learned knowledge when learning new classes. Various methods have been proposed to mitigate this issue. However, rehearsal-based learning, which retains samples from previous classes, typically achieves good performance but tends to memorize specific instances, struggling with Out-of-Distribution (OOD) generalization. This often leads to high forgetting rates and poor generalization. Surprisingly, the OOD generalization capabilities of these methods have been largely unexplored. In this paper, we highlight this issue and propose a simple yet effective strategy inspired by contrastive learning and data-centric principles to address it.\nWe introduce Adaptive Contrastive Replay (ACR), a method that employs dual optimization to simultaneously train both the encoder and the classifier. ACR adaptively populates the replay buffer with misclassified samples while ensuring a balanced representation of classes and tasks. By refining the decision boundary in this way, ACR achieves a balance between stability and plasticity. Our method significantly outperforms previous approaches in terms of OOD generalization, achieving an improvement of 13.41\\% on Split CIFAR-100, 9.91\\% on Split Mini-ImageNet, and 5.98\\% on Split Tiny-ImageNet.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "PyLWLcKCtG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7736/Reviewer_uazR"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes Adaptive Contrastive Replay (ACR) to address the Out-of-Distribution (OOD) generalization problem often overlooked by existing class-incremental learning algorithms. While previous research has achieved strong performance across various class-incremental learning scenarios, these algorithms tend to perform well only on in-domain tasks due to issues with bad memorization, resulting in poor generalization to OOD tasks. To overcome this, the paper introduces a Proxy-based Contrastive Loss and Adaptive Replay Buffer Management. The Proxy-based Contrastive Loss utilizes the class weights as proxies for contrastive learning. Additionally, exemplars are sampled based on the variance of confidence scores, allowing for balanced sampling across tasks and classes and prioritizing hard examples close to decision boundaries to improve stability and OOD generalization. Through extensive experiments and analysis, the proposed algorithm demonstrates superior performance across diverse scenarios.", "review_text": "This paper proposes Adaptive Contrastive Replay (ACR) to address the Out-of-Distribution (OOD) generalization problem often overlooked by existing class-incremental learning algorithms. While previous research has achieved strong performance across various class-incremental learning scenarios, these algorithms tend to perform well only on in-domain tasks due to issues with bad memorization, resulting in poor generalization to OOD tasks. To overcome this, the paper introduces a Proxy-based Contrastive Loss and Adaptive Replay Buffer Management. The Proxy-based Contrastive Loss utilizes the class weights as proxies for contrastive learning. Additionally, exemplars are sampled based on the variance of confidence scores, allowing for balanced sampling across tasks and classes and prioritizing hard examples close to decision boundaries to improve stability and OOD generalization. Through extensive experiments and analysis, the proposed algorithm demonstrates superior performance across diverse scenarios.", "strengths": "1. This paper identifies the often-overlooked OOD generalization issue in class-incremental learning and proposes ACR to address it. I find the distinction between bad and good memorization for explaining this problem, along with the design of Proxy-based Contrastive Loss and Adaptive Replay Buffer Management, to be well-motivated and effectively structured to tackle the identified issue.\n\n2. To validate the proposed algorithm, the authors conducted extensive experiments across various datasets, algorithms, and memory buffer sizes. Additionally, the thorough analysis and ablation studies to verify and examine the effectiveness of the algorithm are insightful and engaging to read.", "weaknesses": "1. The paper [1] proposes a sampling method that stores the most interfered samples in the replay memory. I believe this plays a similar role and function described in equation (3) of the author's paper. Therefore, I think it is necessary for the authors to explain the differences from [1] and consider it as an additional baseline.\n\n2. The authors validated the superiority of their proposed algorithm through various experiments using CIFAR-100, Split Mini-ImageNet, and Split Tiny-ImageNet in online class-incremental learning. However, I have the following questions regarding these experimental results:\n\n   2-1) According to the paper, the results for Split CIFAR-100 in Table 1 show the experimental results for 10 tasks using the ResNet-18 model. I understand that reporting results for these 10 tasks have generally been done with the ResNet-32 model, starting with [2], and results have been reported in papers such as [3], [4], and [5]. When comparing these results (especially Figure 1(a) of [4]) with those in Table 1, the results in this paper appear to be relatively low. For example, when using a replay memory size of 2000 in [4] and [5], the accuracy of the naive algorithm ER (or replay) is reported to be around 40%, while Table 1 shows only 24%. Furthermore, the highest performance reported in [4] is 60% (e.g., DER), whereas the proposed algorithm achieved only 42% in Table 4. Is this discrepancy solely due to the model used? I believe that comparing the performance of the proposed algorithm with existing baselines using ResNet-32 is essential to confirm its superiority.\n\n   2-2) Since all experiments were conducted using images sized 32 x 32 for a specific scenario, it makes it very difficult to assess whether the proposed algorithm can achieve excellent results across diverse scenarios. Additionally, many class-incremental learning algorithms are known to exhibit different performance trends depending on the input image size and scenario (shown in [4] and [5]), so I believe the authors should conduct additional experiments using ImageNet (or not resized ImageNet-100) and other task scenarios (e.g., a scenario in which a large number of classes are learned in the first task, and the remaining classes are divided and learned in subsequent tasks.) to demonstrate the effectiveness of their proposed algorithm in various settings.\n\n3. From the results in Table 4, it is evident that the simple 'Random' policy already performs very close to 'Challenging' (ACR). This experiment was conducted with a replay memory size of 500; what would happen if it were increased to 2000? Would 'Challenging' still outperform 'Random' in this case? I think it is crucial to demonstrate that 'Challenging' provides superior performance compared to 'Random' in this experiment.\n\n4. Additionally, I believe that, for the ablation study on Adaptive Replay Buffer Management, it is necessary to include results from experiments using ER without proxy-based contrastive learning in Table 4. This will allow for a proper evaluation of the standalone effectiveness of Adaptive Replay Buffer Management. \n\n[1] Online Continual Learning with Maximally Interfered Retrieval, NeurIPS 2019.  \n[2] iCaRL: Incremental Classifier and Representation Learning, CVPR 2017  \n[3] GDumb: A Simple Approach that Questions Our Progress in Continual Learning, ECCV 2020  \n[4] PyCIL: A Python Toolbox for Class-Incremental Learning, Arxiv.  \n[5] Rebalancing Batch Normalization for Exemplar-based Class-Incremental Learning, CVPR 2023", "questions": "I have included all weaknesses and questions in the Weakness section, so please refer to it. I strongly resonate with the authors' concerns about the OOD generalization problem in class-incremental learning algorithms, and I believe the proposed algorithm is well-designed. However, the various questions raised by the experiments and the lack of sufficient ablation studies make it difficult to provide a more favorable evaluation.I look forward to the authors correcting any misunderstandings I may have and providing a response to my review in their author response.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Adaptive Contrastive Replay (ACR) to address the Out-of-Distribution (OOD) generalization problem often overlooked by existing class-incremental learning algorithms. While previous research has achieved strong performance across various class-incremental learning scenarios, these algorithms tend to perform well only on in-domain tasks due to issues with bad memorization, resulting in poor generalization to OOD tasks. To overcome this, the paper introduces a Proxy-based Contrastive Loss and Adaptive Replay Buffer Management. The Proxy-based Contrastive Loss utilizes the class weights as proxies for contrastive learning. Additionally, exemplars are sampled based on the variance of confidence scores, allowing for balanced sampling across tasks and classes and prioritizing hard examples close to decision boundaries to improve stability and OOD generalization. Through extensive experiments and analysis, the proposed algorithm demonstrates superior performance across diverse scenarios.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. This paper identifies the often-overlooked OOD generalization issue in class-incremental learning and proposes ACR to address it. I find the distinction between bad and good memorization for explaining this problem, along with the design of Proxy-based Contrastive Loss and Adaptive Replay Buffer Management, to be well-motivated and effectively structured to tackle the identified issue.\n\n2. To validate the proposed algorithm, the authors conducted extensive experiments across various datasets, algorithms, and memory buffer sizes. Additionally, the thorough analysis and ablation studies to verify and examine the effectiveness of the algorithm are insightful and engaging to read.", "weaknesses": "1. The paper [1] proposes a sampling method that stores the most interfered samples in the replay memory. I believe this plays a similar role and function described in equation (3) of the author's paper. Therefore, I think it is necessary for the authors to explain the differences from [1] and consider it as an additional baseline.\n\n2. The authors validated the superiority of their proposed algorithm through various experiments using CIFAR-100, Split Mini-ImageNet, and Split Tiny-ImageNet in online class-incremental learning. However, I have the following questions regarding these experimental results:\n\n   2-1) According to the paper, the results for Split CIFAR-100 in Table 1 show the experimental results for 10 tasks using the ResNet-18 model. I understand that reporting results for these 10 tasks have generally been done with the ResNet-32 model, starting with [2], and results have been reported in papers such as [3], [4], and [5]. When comparing these results (especially Figure 1(a) of [4]) with those in Table 1, the results in this paper appear to be relatively low. For example, when using a replay memory size of 2000 in [4] and [5], the accuracy of the naive algorithm ER (or replay) is reported to be around 40%, while Table 1 shows only 24%. Furthermore, the highest performance reported in [4] is 60% (e.g., DER), whereas the proposed algorithm achieved only 42% in Table 4. Is this discrepancy solely due to the model used? I believe that comparing the performance of the proposed algorithm with existing baselines using ResNet-32 is essential to confirm its superiority.\n\n   2-2) Since all experiments were conducted using images sized 32 x 32 for a specific scenario, it makes it very difficult to assess whether the proposed algorithm can achieve excellent results across diverse scenarios. Additionally, many class-incremental learning algorithms are known to exhibit different performance trends depending on the input image size and scenario (shown in [4] and [5]), so I believe the authors should conduct additional experiments using ImageNet (or not resized ImageNet-100) and other task scenarios (e.g., a scenario in which a large number of classes are learned in the first task, and the remaining classes are divided and learned in subsequent tasks.) to demonstrate the effectiveness of their proposed algorithm in various settings.\n\n3. From the results in Table 4, it is evident that the simple 'Random' policy already performs very close to 'Challenging' (ACR). This experiment was conducted with a replay memory size of 500; what would happen if it were increased to 2000? Would 'Challenging' still outperform 'Random' in this case? I think it is crucial to demonstrate that 'Challenging' provides superior performance compared to 'Random' in this experiment.\n\n4. Additionally, I believe that, for the ablation study on Adaptive Replay Buffer Management, it is necessary to include results from experiments using ER without proxy-based contrastive learning in Table 4. This will allow for a proper evaluation of the standalone effectiveness of Adaptive Replay Buffer Management. \n\n[1] Online Continual Learning with Maximally Interfered Retrieval, NeurIPS 2019.  \n[2] iCaRL: Incremental Classifier and Representation Learning, CVPR 2017  \n[3] GDumb: A Simple Approach that Questions Our Progress in Continual Learning, ECCV 2020  \n[4] PyCIL: A Python Toolbox for Class-Incremental Learning, Arxiv.  \n[5] Rebalancing Batch Normalization for Exemplar-based Class-Incremental Learning, CVPR 2023", "questions": "I have included all weaknesses and questions in the Weakness section, so please refer to it. I strongly resonate with the authors' concerns about the OOD generalization problem in class-incremental learning algorithms, and I believe the proposed algorithm is well-designed. However, the various questions raised by the experiments and the lack of sufficient ablation studies make it difficult to provide a more favorable evaluation.I look forward to the authors correcting any misunderstandings I may have and providing a response to my review in their author response.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730709316753}, {"id": "tXQTJujN9p", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7736/Reviewer_RFy4"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper addressed the out-of-distribution generalization performance of rehearsal-based CL methods, which is a topic that has been overlooked by the community. The authors proposed Adaptive Contrastive Replay (ACR) which consists of two components: proxy-based contrastive learning and adaptive replay buffer management. The former encourages the model to maintain a clear class boundary while the latter populates the buffer with task- and class-balanced challenging samples to improve the model's stability and plasticity. Experiments on popular CL benchmarks showed improved performance with the proposed method with both iid and ood data.", "review_text": "The paper addressed the out-of-distribution generalization performance of rehearsal-based CL methods, which is a topic that has been overlooked by the community. The authors proposed Adaptive Contrastive Replay (ACR) which consists of two components: proxy-based contrastive learning and adaptive replay buffer management. The former encourages the model to maintain a clear class boundary while the latter populates the buffer with task- and class-balanced challenging samples to improve the model's stability and plasticity. Experiments on popular CL benchmarks showed improved performance with the proposed method with both iid and ood data.", "strengths": "1. The paper is well-motivated\n\n2. The paper is generally easy-to-follow", "weaknesses": "There are some critical concerns about the proposed components, some missing explanations, and some room for improvement in the representations of the paper. Please refer to the Questions section.", "questions": "1. How does the proposed PCL loss differ from a standard cross entropy loss with temperature-scaled softmax?\n\nThe authors might elaborate more on this because I cannot see the difference between the proposed PCL loss (Eq. 4) and a stand softmax CE loss.\n\n2. Is the confidence variance evolved to fixed after epoch E, and is this robust?\n\nAccording to Algorithm 1, it seems that confidence is only recorded until epoch E (which is 5 in practice) whereas the total training lasts 50 epochs. Is that correct? If so, I am wondering how robust is this confidence variance. First, in the first 5 epochs, the model might not be able to converge to a stable state, which means that the predictions can be rather random. Is the confidence variance based on nearly random predictions reliable? I am not sure. Can the authors compare the confidence variance at the beginning of the task and at the end? Second, it seems reasonable that the model would refine by itself the decision boundary, which means that some challenging samples might become not challenging along training. If the confidence variance is fixed after epoch E, the the entire buffer seems to be static within a task, then the score cannot take this into account.\n\n3. Are the challenging samples really helping?\n\nThe authors' claim to include challenging samples in the buffer is to maintain accurate decision boundaries for old tasks (lines 285-286), which should lead to less forgetting. However, in Table 4, the challenging column is not performing better than the random column in BWT, which suggests that the proposed selection mechanism might not be effective in maintaining the decision boundary of old tasks. It also raised the concern that the improvement of the proposed selection mechanism might just be a result of balanced sampling in the memory buffer, which is more or less well-known in the CL community. \n\n4. Some explanations are missing\n\nIn lines 89-90, the implementation of the comparison is not explained, which makes Fig. 1 harder to understand\n\nIn lines 266-269, the claim that no task label needs to be stored is confusing. How are the samples arranged separately in the buffer (line 266) if no task label is known? \n\nThe concept of proxy is not clearly explained (lines 182-184)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addressed the out-of-distribution generalization performance of rehearsal-based CL methods, which is a topic that has been overlooked by the community. The authors proposed Adaptive Contrastive Replay (ACR) which consists of two components: proxy-based contrastive learning and adaptive replay buffer management. The former encourages the model to maintain a clear class boundary while the latter populates the buffer with task- and class-balanced challenging samples to improve the model's stability and plasticity. Experiments on popular CL benchmarks showed improved performance with the proposed method with both iid and ood data.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper is well-motivated\n\n2. The paper is generally easy-to-follow", "weaknesses": "There are some critical concerns about the proposed components, some missing explanations, and some room for improvement in the representations of the paper. Please refer to the Questions section.", "questions": "1. How does the proposed PCL loss differ from a standard cross entropy loss with temperature-scaled softmax?\n\nThe authors might elaborate more on this because I cannot see the difference between the proposed PCL loss (Eq. 4) and a stand softmax CE loss.\n\n2. Is the confidence variance evolved to fixed after epoch E, and is this robust?\n\nAccording to Algorithm 1, it seems that confidence is only recorded until epoch E (which is 5 in practice) whereas the total training lasts 50 epochs. Is that correct? If so, I am wondering how robust is this confidence variance. First, in the first 5 epochs, the model might not be able to converge to a stable state, which means that the predictions can be rather random. Is the confidence variance based on nearly random predictions reliable? I am not sure. Can the authors compare the confidence variance at the beginning of the task and at the end? Second, it seems reasonable that the model would refine by itself the decision boundary, which means that some challenging samples might become not challenging along training. If the confidence variance is fixed after epoch E, the the entire buffer seems to be static within a task, then the score cannot take this into account.\n\n3. Are the challenging samples really helping?\n\nThe authors' claim to include challenging samples in the buffer is to maintain accurate decision boundaries for old tasks (lines 285-286), which should lead to less forgetting. However, in Table 4, the challenging column is not performing better than the random column in BWT, which suggests that the proposed selection mechanism might not be effective in maintaining the decision boundary of old tasks. It also raised the concern that the improvement of the proposed selection mechanism might just be a result of balanced sampling in the memory buffer, which is more or less well-known in the CL community. \n\n4. Some explanations are missing\n\nIn lines 89-90, the implementation of the comparison is not explained, which makes Fig. 1 harder to understand\n\nIn lines 266-269, the claim that no task label needs to be stored is confusing. How are the samples arranged separately in the buffer (line 266) if no task label is known? \n\nThe concept of proxy is not clearly explained (lines 182-184)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730689603827}, {"id": "rHZT2y9d7x", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7736/Reviewer_M5Un"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper proposes a novel approach called Adaptive Contrastive Replay (ACR) to improve continual learning, specifically focusing on reducing forgetting and enhancing out-of-distribution (OOD) generalization. Existing methods for continual learning often rely on retaining past samples to prevent catastrophic forgetting but struggle with OOD generalization, leading to imbalances and high memory costs. ACR addresses this by using a dual optimization strategy that refines the decision boundary, focusing on informative boundary samples rather than simple memorization. By balancing class and task representation in the memory buffer and leveraging a proxy-based contrastive loss, ACR achieves significantly better stability-plasticity trade-offs and outperforms prior methods on various benchmarks like Split CIFAR-100 and Mini-ImageNet", "review_text": "The paper proposes a novel approach called Adaptive Contrastive Replay (ACR) to improve continual learning, specifically focusing on reducing forgetting and enhancing out-of-distribution (OOD) generalization. Existing methods for continual learning often rely on retaining past samples to prevent catastrophic forgetting but struggle with OOD generalization, leading to imbalances and high memory costs. ACR addresses this by using a dual optimization strategy that refines the decision boundary, focusing on informative boundary samples rather than simple memorization. By balancing class and task representation in the memory buffer and leveraging a proxy-based contrastive loss, ACR achieves significantly better stability-plasticity trade-offs and outperforms prior methods on various benchmarks like Split CIFAR-100 and Mini-ImageNet", "strengths": "1. OOD generalization: it combines the robust features of contrastive learning with adaptive memory buffer management to address poor OOD generalization in continual learning.\n2. Enhanced Memory Management: The method’s strategy for memory buffer management, which focuses on maintaining a balanced distribution of classes and tasks, helps mitigate the long-tail effect often observed in other methods. This leads to more stable learning outcomes.", "weaknesses": "1. Lack of novelty: The proposed \"Proxy-Based Contrastive Loss\" is already widely used in Continual Learning, as shown in [1], [2]. Please clarify the unique aspects of your method in comparison to existing approaches.\n2. Lack of ablation study: In your paper, you propose using the output of the last layer as an index to estimate image uncertainty, which then guides the management of your memory set. I believe this could help select images near the decision boundary. Based on my experience, combining data augmentation and contrastive learning generally results in significant accuracy improvements in continual learning. Could you specify the performance gain achieved by solely using your memory management method? Additionally, what are the results when other methods are combined with your contrastive learning loss?\n\n\n[1] De Lange, Matthias and Tuytelaars, Tinne. \"Continual Prototype Evolution: Learning Online From Non-Stationary Data Streams\", ICCV,2021.\n[2] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, Eugene Belilovsky. \"New Insights on Reducing Abrupt Representation Change in Online Continual Learning\", ICLR,2022.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel approach called Adaptive Contrastive Replay (ACR) to improve continual learning, specifically focusing on reducing forgetting and enhancing out-of-distribution (OOD) generalization. Existing methods for continual learning often rely on retaining past samples to prevent catastrophic forgetting but struggle with OOD generalization, leading to imbalances and high memory costs. ACR addresses this by using a dual optimization strategy that refines the decision boundary, focusing on informative boundary samples rather than simple memorization. By balancing class and task representation in the memory buffer and leveraging a proxy-based contrastive loss, ACR achieves significantly better stability-plasticity trade-offs and outperforms prior methods on various benchmarks like Split CIFAR-100 and Mini-ImageNet", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. OOD generalization: it combines the robust features of contrastive learning with adaptive memory buffer management to address poor OOD generalization in continual learning.\n2. Enhanced Memory Management: The method’s strategy for memory buffer management, which focuses on maintaining a balanced distribution of classes and tasks, helps mitigate the long-tail effect often observed in other methods. This leads to more stable learning outcomes.", "weaknesses": "1. Lack of novelty: The proposed \"Proxy-Based Contrastive Loss\" is already widely used in Continual Learning, as shown in [1], [2]. Please clarify the unique aspects of your method in comparison to existing approaches.\n2. Lack of ablation study: In your paper, you propose using the output of the last layer as an index to estimate image uncertainty, which then guides the management of your memory set. I believe this could help select images near the decision boundary. Based on my experience, combining data augmentation and contrastive learning generally results in significant accuracy improvements in continual learning. Could you specify the performance gain achieved by solely using your memory management method? Additionally, what are the results when other methods are combined with your contrastive learning loss?\n\n\n[1] De Lange, Matthias and Tuytelaars, Tinne. \"Continual Prototype Evolution: Learning Online From Non-Stationary Data Streams\", ICCV,2021.\n[2] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, Eugene Belilovsky. \"New Insights on Reducing Abrupt Representation Change in Online Continual Learning\", ICLR,2022.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730478294817}, {"id": "QtY3B9rPlI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7736/Reviewer_BhyB"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces Adaptive Contrastive Replay (ACR), a novel rehearsal-based continual learning method aimed at improving Out-of-Distribution (OOD) generalization while mitigating catastrophic forgetting. ACR leverages proxy-based contrastive learning and confidence-guided sample selection to populate its buffer with boundary samples rather than outliers. This approach balances stability and plasticity, ensuring the retention of old knowledge while adapting to new information. ACR achieves superior performance across multiple benchmarks (Split CIFAR-100, Split Mini-ImageNet, and Split Tiny-ImageNet) in both i.i.d. and OOD settings, outperforming state-of-the-art rehearsal-based methods.", "review_text": "This paper introduces Adaptive Contrastive Replay (ACR), a novel rehearsal-based continual learning method aimed at improving Out-of-Distribution (OOD) generalization while mitigating catastrophic forgetting. ACR leverages proxy-based contrastive learning and confidence-guided sample selection to populate its buffer with boundary samples rather than outliers. This approach balances stability and plasticity, ensuring the retention of old knowledge while adapting to new information. ACR achieves superior performance across multiple benchmarks (Split CIFAR-100, Split Mini-ImageNet, and Split Tiny-ImageNet) in both i.i.d. and OOD settings, outperforming state-of-the-art rehearsal-based methods.", "strengths": "1. Relevant Contribution: Tackles the overlooked issue of OOD generalization in continual learning, which enhances the practical applicability of CL methods.\n2. Balanced Stability and Plasticity: Focuses on boundary sample selection, which helps maintain a stable yet adaptive learning process.\n3. Efficiency Gains: Reduces computational complexity, making ACR suitable for deployment in resource-constrained environments.\n4. Comprehensive Evaluation: Provides comparisons across multiple datasets and with several rehearsal-based baselines, showcasing consistent improvements.", "weaknesses": "1. Narrow Scope of Baselines: The paper compares ACR only with other rehearsal-based approaches, omitting non-rehearsal continual learning methods that could offer additional insights.\n2. Over-reliance on Specific OOD Corruptions: The focus on synthetic corruptions (e.g., Gaussian noise, blur) may not fully represent real-world OOD scenarios such as domain shifts or adversarial inputs.\n3. Limited Hyperparameter Analysis: The impact of key hyperparameters like the temperature $\\tau$ and buffer size is not thoroughly explored, leaving some questions about robustness unanswered.\n4.Scaling Limitations: There is insufficient discussion about how ACR performs on longer task sequences or more complex datasets beyond the three benchmarks used in the paper.\n5. Clarity Issues: Figure 2 is unclear, and key elements (e.g., proxy-based contrastive loss) would benefit from more intuitive explanations.\n6.Assumptions in Methodology: The paper assumes that high-variance boundary samples will always improve performance, without discussing scenarios where this assumption might fail (e.g., noisy labels or unbalanced datasets).", "questions": "1. Figure 2 is unclear. Can the authors provide an improved version or additional explanations to clarify how the buffer update mechanism works?\nHow would ACR perform under more realistic OOD scenarios such as domain shifts, adversarial attacks, or real-world data variability?\n2. What are the computational trade-offs when balancing between boundary and outlier samples? Can ACR scale effectively for larger datasets or continuous tasks?\n3. How sensitive is the performance to the temperature parameter ($\\tau$) in the proxy-based contrastive loss, and what is the impact of different buffer sizes on the results?\n4. Are there any failure cases or situations where ACR struggles, particularly for longer task sequences or when facing noisy or imbalanced datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Adaptive Contrastive Replay (ACR), a novel rehearsal-based continual learning method aimed at improving Out-of-Distribution (OOD) generalization while mitigating catastrophic forgetting. ACR leverages proxy-based contrastive learning and confidence-guided sample selection to populate its buffer with boundary samples rather than outliers. This approach balances stability and plasticity, ensuring the retention of old knowledge while adapting to new information. ACR achieves superior performance across multiple benchmarks (Split CIFAR-100, Split Mini-ImageNet, and Split Tiny-ImageNet) in both i.i.d. and OOD settings, outperforming state-of-the-art rehearsal-based methods.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Relevant Contribution: Tackles the overlooked issue of OOD generalization in continual learning, which enhances the practical applicability of CL methods.\n2. Balanced Stability and Plasticity: Focuses on boundary sample selection, which helps maintain a stable yet adaptive learning process.\n3. Efficiency Gains: Reduces computational complexity, making ACR suitable for deployment in resource-constrained environments.\n4. Comprehensive Evaluation: Provides comparisons across multiple datasets and with several rehearsal-based baselines, showcasing consistent improvements.", "weaknesses": "1. Narrow Scope of Baselines: The paper compares ACR only with other rehearsal-based approaches, omitting non-rehearsal continual learning methods that could offer additional insights.\n2. Over-reliance on Specific OOD Corruptions: The focus on synthetic corruptions (e.g., Gaussian noise, blur) may not fully represent real-world OOD scenarios such as domain shifts or adversarial inputs.\n3. Limited Hyperparameter Analysis: The impact of key hyperparameters like the temperature $\\tau$ and buffer size is not thoroughly explored, leaving some questions about robustness unanswered.\n4.Scaling Limitations: There is insufficient discussion about how ACR performs on longer task sequences or more complex datasets beyond the three benchmarks used in the paper.\n5. Clarity Issues: Figure 2 is unclear, and key elements (e.g., proxy-based contrastive loss) would benefit from more intuitive explanations.\n6.Assumptions in Methodology: The paper assumes that high-variance boundary samples will always improve performance, without discussing scenarios where this assumption might fail (e.g., noisy labels or unbalanced datasets).", "questions": "1. Figure 2 is unclear. Can the authors provide an improved version or additional explanations to clarify how the buffer update mechanism works?\nHow would ACR perform under more realistic OOD scenarios such as domain shifts, adversarial attacks, or real-world data variability?\n2. What are the computational trade-offs when balancing between boundary and outlier samples? Can ACR scale effectively for larger datasets or continuous tasks?\n3. How sensitive is the performance to the temperature parameter ($\\tau$) in the proxy-based contrastive loss, and what is the impact of different buffer sizes on the results?\n4. Are there any failure cases or situations where ACR struggles, particularly for longer task sequences or when facing noisy or imbalanced datasets?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730008758823}], "openreview_url": "https://openreview.net/forum?id=YGflij9S6x", "arxiv_id": "2410.07110", "paper_pdf": "papers/YGflij9S6x.pdf", "paper_pdf_sha256": "c3b34cf1533b20adf3802b894db7af1344f5ce1155f76d525b861728ce7947c6", "paper_pdf_bytes": 3559766, "paper_pdf_source": "openreview", "code_url": "https://github.com/hossein-rezaei624/ACR", "code_repository": "hossein-rezaei624/ACR", "code_commit": "339b4734065e203e9ae1da289573de6f515806cf", "code_archive": "repos/YGflij9S6x.zip", "code_archive_sha256": "d515281c32aabf365268c115e5964a5332432b6359013e53f325d86037e3fe83", "code_archive_bytes": 151142, "code_file_count": 83, "code_extensions": {".py": 83}, "github_disk_usage_kb": 151, "github_languages": {"Python": 472469}, "github_archived": false, "github_pushed_at": "2024-10-02T10:03:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continual-learning-less-forgetting-more-ood"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "T16M4SzH1v", "year": 2024, "status": "rejected", "title": "Distributional Bellman Operators over Mean Embeddings", "authors": ["Li Kevin Wenliang", "Gregoire Deletang", "Matthew Aitchison", "Marcus Hutter", "Anian Ruoss", "Arthur Gretton", "Mark Rowland"], "authorids": ["~Li_Kevin_Wenliang1", "~Gregoire_Deletang1", "~Matthew_Aitchison1", "~Marcus_Hutter1", "~Anian_Ruoss1", "~Arthur_Gretton1", "~Mark_Rowland1"], "authors_source": "OpenReview API", "abstract": "We propose a novel algorithmic framework for distributional reinforcement learning, based on learning finite-dimensional mean embeddings  of return distributions. We derive several new algorithms for dynamic programming and temporal-difference learning based on this framework, provide asymptotic convergence theory, and examine the empirical performance of the algorithms on a suite of tabular tasks.   Further, we show that this approach can be straightforwardly combined with deep reinforcement learning, and obtain a new deep reinforcement learning agent that improves over baseline distributional approaches on the Arcade Learning Environment.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "MbX2uHHvWP", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5913/Reviewer_ReLT"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper discusses a method to do distributional reinforcement learning using mean embedding sketches. The advantage of this method is that the sketches can be updated without ever obtaining the imputed distribution, thereby allowing computations entirely in the sketch domain.", "review_text": "This paper discusses a method to do distributional reinforcement learning using mean embedding sketches. The advantage of this method is that the sketches can be updated without ever obtaining the imputed distribution, thereby allowing computations entirely in the sketch domain.", "strengths": "This paper satisfies all the criteria of an excellent paper:\n1. It is exceptionally clearly written. Reading it and learning from it was a joy. It does a great job at being thorough without being pedantic in its discussion. I really appreciated the concrete discussion in sections 3.2 and 4.1; it dovetails quite nicely with the rest of the paper.\n2. It makes a useful contribution to the fun and important problem of distributional reinforcement learning. The idea is simple yet elegant.", "weaknesses": "The formulation of feature maps used in the paper (equation (8)) is sort of \"pulled out a hat\". It would be useful to justify why it made sense to use that formulation as opposed to other possibilities. See also the questions below.", "questions": "1. The general sketch using Bellman coefficient $B_r$ will correspond to some mean embedding sketch which is an invertible linear combination of first $m$-moments, right?\n2. Why can we write $B_r$ as $C_rC^{-1}$?\n3. Where's the generalization to the case when $\\mathcal R$ is not finite?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper discusses a method to do distributional reinforcement learning using mean embedding sketches. The advantage of this method is that the sketches can be updated without ever obtaining the imputed distribution, thereby allowing computations entirely in the sketch domain.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "This paper satisfies all the criteria of an excellent paper:\n1. It is exceptionally clearly written. Reading it and learning from it was a joy. It does a great job at being thorough without being pedantic in its discussion. I really appreciated the concrete discussion in sections 3.2 and 4.1; it dovetails quite nicely with the rest of the paper.\n2. It makes a useful contribution to the fun and important problem of distributional reinforcement learning. The idea is simple yet elegant.", "weaknesses": "The formulation of feature maps used in the paper (equation (8)) is sort of \"pulled out a hat\". It would be useful to justify why it made sense to use that formulation as opposed to other possibilities. See also the questions below.", "questions": "1. The general sketch using Bellman coefficient $B_r$ will correspond to some mean embedding sketch which is an invertible linear combination of first $m$-moments, right?\n2. Why can we write $B_r$ as $C_rC^{-1}$?\n3. Where's the generalization to the case when $\\mathcal R$ is not finite?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698874083017}, {"id": "crIEts9VNM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5913/Reviewer_PQKZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a novel algorithmic framework for distributional reinforcement learning that works purely in the sketch space based on finite-dimensional mean embeddings of return distributions. The authors derive the approximate Bellman equation in the sketch space and propose a new algorithm that can be combined with dynamic programming and temporal-difference learning. The authors provide an asymptotic convergence theory for the proposed method, and examine the representation error of the proposed sketch on a suite of tabular tasks. They also demonstrate that this approach can be straightforwardly combined with deep reinforcement learning, obtaining a new deep RL agent that improves over the QR-DQN baseline on the Arcade Learning Environment.", "review_text": "This paper proposes a novel algorithmic framework for distributional reinforcement learning that works purely in the sketch space based on finite-dimensional mean embeddings of return distributions. The authors derive the approximate Bellman equation in the sketch space and propose a new algorithm that can be combined with dynamic programming and temporal-difference learning. The authors provide an asymptotic convergence theory for the proposed method, and examine the representation error of the proposed sketch on a suite of tabular tasks. They also demonstrate that this approach can be straightforwardly combined with deep reinforcement learning, obtaining a new deep RL agent that improves over the QR-DQN baseline on the Arcade Learning Environment.", "strengths": "1. The paper introduces a novel framework for distributional reinforcement learning, which is based on learning mean embeddings of return distributions. This approach avoids the need for expensive imputation strategies, which can be computationally expensive and biologically implausible.\n2. The authors provide a theoretical analysis of the proposed algorithms, including asymptotic convergence results. This analysis helps to establish the theoretical foundations of the approach and provides insights into its properties.", "weaknesses": "1. The proposed method requires a linear approximation for the Bellman update equation and require calculating a Bellman coefficient matrix $B_r$ that can be computationally challenging. This contradicts the motivation to improve computation efficiency and reduce the imputation error by purely operating in the sketch space. It is unclear how the proposed method is superior to previous methods both computationally and statistically.\n\n2. The experimental validation is limited. The experiment on the tabular tasks only shows the approximation error of the proposed sketch method rather than the performance of the overall approach. From the result, the proposed sketch method has a similar Cramer distance compared to the CDRL baseline. While the proposed sketch has lower excess Cramer distance and mean-embedding squared error, it is unclear how these metrics translate into the performance of the proposed approach. For the Deep RL part, the proposed method underperforms IQN and does not significantly outperform QR-DQN, which is the backbone of the proposed method. Does the proposed method have a significant computation advantage?\n\nMinor: CDRL is mentioned but not referred to in the paper.", "questions": "1. What is the advantage of the proposed method compared with previous distributional RL methods? Does the proposed method have a provable lower approximation error or computation advantage?\n\n2. How does the proposed method perform on the tabular tasks compared with baselines? Does the proposed method have a computation advantage in the deep RL setting?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel algorithmic framework for distributional reinforcement learning that works purely in the sketch space based on finite-dimensional mean embeddings of return distributions. The authors derive the approximate Bellman equation in the sketch space and propose a new algorithm that can be combined with dynamic programming and temporal-difference learning. The authors provide an asymptotic convergence theory for the proposed method, and examine the representation error of the proposed sketch on a suite of tabular tasks. They also demonstrate that this approach can be straightforwardly combined with deep reinforcement learning, obtaining a new deep RL agent that improves over the QR-DQN baseline on the Arcade Learning Environment.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The paper introduces a novel framework for distributional reinforcement learning, which is based on learning mean embeddings of return distributions. This approach avoids the need for expensive imputation strategies, which can be computationally expensive and biologically implausible.\n2. The authors provide a theoretical analysis of the proposed algorithms, including asymptotic convergence results. This analysis helps to establish the theoretical foundations of the approach and provides insights into its properties.", "weaknesses": "1. The proposed method requires a linear approximation for the Bellman update equation and require calculating a Bellman coefficient matrix $B_r$ that can be computationally challenging. This contradicts the motivation to improve computation efficiency and reduce the imputation error by purely operating in the sketch space. It is unclear how the proposed method is superior to previous methods both computationally and statistically.\n\n2. The experimental validation is limited. The experiment on the tabular tasks only shows the approximation error of the proposed sketch method rather than the performance of the overall approach. From the result, the proposed sketch method has a similar Cramer distance compared to the CDRL baseline. While the proposed sketch has lower excess Cramer distance and mean-embedding squared error, it is unclear how these metrics translate into the performance of the proposed approach. For the Deep RL part, the proposed method underperforms IQN and does not significantly outperform QR-DQN, which is the backbone of the proposed method. Does the proposed method have a significant computation advantage?\n\nMinor: CDRL is mentioned but not referred to in the paper.", "questions": "1. What is the advantage of the proposed method compared with previous distributional RL methods? Does the proposed method have a provable lower approximation error or computation advantage?\n\n2. How does the proposed method perform on the tabular tasks compared with baselines? Does the proposed method have a computation advantage in the deep RL setting?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698831106266}, {"id": "2A4qZ2QEj6", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5913/Reviewer_RJqf"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a novel distributional RL framework based on the sketch Bellman operator. The approach is derived from statistical functional dynamic programming and involves constructing different sketches to formulate the Bellman equation. The authors provide a theoretical guarantee to support their proposed method and empirical studies demonstrate that it outperforms some baselines in Atari environments.", "review_text": "This paper presents a novel distributional RL framework based on the sketch Bellman operator. The approach is derived from statistical functional dynamic programming and involves constructing different sketches to formulate the Bellman equation. The authors provide a theoretical guarantee to support their proposed method and empirical studies demonstrate that it outperforms some baselines in Atari environments.", "strengths": "1) The authors provide detailed experiments, particularly for ablations on different feature functions.\n\n2) The authors provide theoretical guarantee, which adds credibility to their proposed method.", "weaknesses": "1) The authors' motivation for using the sketch Bellman operator is to reduce the need for expensive imputation strategies when converting between sketches and distributions. However, the experiment does not verify this claim. It would be helpful if the authors could provide some quantitative results (such as training time) to demonstrate this reduction.\n\n2) The proposed method performs worse than IQN, even though the authors claim that IQN uses a more complex prediction network for non-parametric predictions. It would be beneficial if the authors could provide quantitative results (such as neural network sizes and training time) to explain this difference.\n\n3) It would be useful to include vanilla Statistical Functional Dynamic Programming (Bellemare et al., 2023) as a baseline in some toy examples to compare its performance with the proposed Sketch-DQN.\n\n4) The writings in some sections are not very clear. For example,  in the subsection on 'Computing Bellman coefficients',  the authors directly show the closed form of (5) without any explanations or citations.\n\n5) There are some typos that need to be addressed, such as in Sec 2.2, where it should be $\n\\left(\\left(\\mathcal{T}^\\pi \\iota (U)\\right)(x)\\right)$ instead of $\n\\left(\\left(\\mathcal{T}^\\pi \\iota U\\right)(x)\\right)$.", "questions": "Please answer the questions mentioned above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel distributional RL framework based on the sketch Bellman operator. The approach is derived from statistical functional dynamic programming and involves constructing different sketches to formulate the Bellman equation. The authors provide a theoretical guarantee to support their proposed method and empirical studies demonstrate that it outperforms some baselines in Atari environments.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1) The authors provide detailed experiments, particularly for ablations on different feature functions.\n\n2) The authors provide theoretical guarantee, which adds credibility to their proposed method.", "weaknesses": "1) The authors' motivation for using the sketch Bellman operator is to reduce the need for expensive imputation strategies when converting between sketches and distributions. However, the experiment does not verify this claim. It would be helpful if the authors could provide some quantitative results (such as training time) to demonstrate this reduction.\n\n2) The proposed method performs worse than IQN, even though the authors claim that IQN uses a more complex prediction network for non-parametric predictions. It would be beneficial if the authors could provide quantitative results (such as neural network sizes and training time) to explain this difference.\n\n3) It would be useful to include vanilla Statistical Functional Dynamic Programming (Bellemare et al., 2023) as a baseline in some toy examples to compare its performance with the proposed Sketch-DQN.\n\n4) The writings in some sections are not very clear. For example,  in the subsection on 'Computing Bellman coefficients',  the authors directly show the closed form of (5) without any explanations or citations.\n\n5) There are some typos that need to be addressed, such as in Sec 2.2, where it should be $\n\\left(\\left(\\mathcal{T}^\\pi \\iota (U)\\right)(x)\\right)$ instead of $\n\\left(\\left(\\mathcal{T}^\\pi \\iota U\\right)(x)\\right)$.", "questions": "Please answer the questions mentioned above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698510508748}, {"id": "Z26cCnakk4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5913/Reviewer_7TXx"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper studies the mean embedding sketches, one kind of statistical functional, in the context of distributional RL. They first show the sketch framework, where the Bellman operator is applied on the sketch instead of the original value function in classical RL. Further, they also provide the convergence guarantee under the sketch framework. Experiments are conducted on both tabular MDP and Atari games.", "review_text": "This paper studies the mean embedding sketches, one kind of statistical functional, in the context of distributional RL. They first show the sketch framework, where the Bellman operator is applied on the sketch instead of the original value function in classical RL. Further, they also provide the convergence guarantee under the sketch framework. Experiments are conducted on both tabular MDP and Atari games.", "strengths": "* The writing is clear. The paper is well-organized and easy to follow.\n\n* Experiments are extensive by considering both MDP and all Atari games with 200M frames.", "weaknesses": "* **Limited Methodology contribution and novelty**. Firstly, the sketch is just one special statistical functionals equipped with a specified kernel function, which is typically less commonly used than quantiles or samples. Working on algorithms simply based on a new concept may not contribute to the real development of this research field. More importantly, feature mapping has already been implicitly considered in MMDDRL paper [1], where MMD naturally induces a feature map. Notably. [1] also shows that not all kernels will lead to a convergent distributional Bellman operator, about which this paper has not fully discussed. From the perspective of new algorithms, Sketch-DP or TD is not new to me and even straightforward. Its concrete version, e.g., in Proposition 4.4 is very similar to the categorical representation. The choice of different feature maps also has a strong correlation with the MMD equipped with different kernels. Based on my knowledge of RKHS, different kernel naturally induces a RHKS with a specific feature map. Therefore, I do not think the sketch framework has sufficient methodological contribution compared with existing works.\n\n* **Insufficient theoretical analysis**. The biggest issue in the theoretical part is the pre-specification of a metric $d$ such that it is $\\gamma^c$ contractive, which is a strong assumption. With a metric that can already guarantee the contraction, the approximation error is easy to show. Further, as mentioned in the first weakness, not all feature maps can guarantee the contraction. For example, [1] gave some counterexamples for the convergence when we use Gaussian kernels. Therefore, the current paper lacks a crucial theoretical part about what kinds of feature maps can guarantee the contraction and then bound the errors. Without this crucial part, I personally think the current theoretical results are insufficient.  \n\n* **Insignificant empirical improvements**. The viewpoint of limited methodological contribution is further demonstrated by the insignificant empirical improvements in Figure 5. Although I truly agree on the conclusion that the sketch framework slightly improves C51 and QRDQN, it performs worse than IQN. Hence, I only view the sketch paper as a feature map-based algorithm, which has already been implicitly investigated before, rather than a novel one with significant performance. The insignificant improvement also let me rethink what is the real motivation of the sketch framework. \n\n\n[1] Thanh Nguyen-Tang, Sunil Gupta, and Svetha Venkatesh. Distributional reinforcement learning via\nmoment matching. (AAAI 2021)", "questions": "Please refer to the Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the mean embedding sketches, one kind of statistical functional, in the context of distributional RL. They first show the sketch framework, where the Bellman operator is applied on the sketch instead of the original value function in classical RL. Further, they also provide the convergence guarantee under the sketch framework. Experiments are conducted on both tabular MDP and Atari games.", "soundness": "3 good", "presentation": "2 fair", "contribution": "1 poor", "strengths": "* The writing is clear. The paper is well-organized and easy to follow.\n\n* Experiments are extensive by considering both MDP and all Atari games with 200M frames.", "weaknesses": "* **Limited Methodology contribution and novelty**. Firstly, the sketch is just one special statistical functionals equipped with a specified kernel function, which is typically less commonly used than quantiles or samples. Working on algorithms simply based on a new concept may not contribute to the real development of this research field. More importantly, feature mapping has already been implicitly considered in MMDDRL paper [1], where MMD naturally induces a feature map. Notably. [1] also shows that not all kernels will lead to a convergent distributional Bellman operator, about which this paper has not fully discussed. From the perspective of new algorithms, Sketch-DP or TD is not new to me and even straightforward. Its concrete version, e.g., in Proposition 4.4 is very similar to the categorical representation. The choice of different feature maps also has a strong correlation with the MMD equipped with different kernels. Based on my knowledge of RKHS, different kernel naturally induces a RHKS with a specific feature map. Therefore, I do not think the sketch framework has sufficient methodological contribution compared with existing works.\n\n* **Insufficient theoretical analysis**. The biggest issue in the theoretical part is the pre-specification of a metric $d$ such that it is $\\gamma^c$ contractive, which is a strong assumption. With a metric that can already guarantee the contraction, the approximation error is easy to show. Further, as mentioned in the first weakness, not all feature maps can guarantee the contraction. For example, [1] gave some counterexamples for the convergence when we use Gaussian kernels. Therefore, the current paper lacks a crucial theoretical part about what kinds of feature maps can guarantee the contraction and then bound the errors. Without this crucial part, I personally think the current theoretical results are insufficient.  \n\n* **Insignificant empirical improvements**. The viewpoint of limited methodological contribution is further demonstrated by the insignificant empirical improvements in Figure 5. Although I truly agree on the conclusion that the sketch framework slightly improves C51 and QRDQN, it performs worse than IQN. Hence, I only view the sketch paper as a feature map-based algorithm, which has already been implicitly investigated before, rather than a novel one with significant performance. The insignificant improvement also let me rethink what is the real motivation of the sketch framework. \n\n\n[1] Thanh Nguyen-Tang, Sunil Gupta, and Svetha Venkatesh. Distributional reinforcement learning via\nmoment matching. (AAAI 2021)", "questions": "Please refer to the Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698387106263}], "openreview_url": "https://openreview.net/forum?id=T16M4SzH1v", "arxiv_id": "2312.07358", "paper_pdf": "papers/T16M4SzH1v.pdf", "paper_pdf_sha256": "ed76f24e59312599fa3c165dbc2fba6fd740cd02bb030c587768a7a851026ace", "paper_pdf_bytes": 1010551, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-deepmind/sketch_dqn", "code_repository": "google-deepmind/sketch_dqn", "code_commit": "b8cf8133fb9e33f3557f901c4136c9372997f0f0", "code_archive": "repos/T16M4SzH1v.zip", "code_archive_sha256": "1e1ab23ef8528fd981852c77ad3a9b3e281788158b2d74ea4964a41a9e0c3abe", "code_archive_bytes": 72195, "code_file_count": 20, "code_extensions": {".py": 19, ".sh": 1}, "github_disk_usage_kb": 117, "github_languages": {"Python": 202590, "Shell": 1871, "Dockerfile": 1748}, "github_archived": false, "github_pushed_at": "2026-04-13T23:44:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distributional-bellman-operators-over-mean"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jCpTofV7iY_", "year": 2023, "status": "rejected", "title": "Pre-trained Language Models can be Fully Zero-Shot Learners", "authors": ["Xuandong Zhao", "Siqi Ouyang", "Zhiguo Yu", "Ming Wu", "Lei Li"], "authorids": ["~Xuandong_Zhao1", "~Siqi_Ouyang2", "~Zhiguo_Yu2", "~Ming_Wu4", "~Lei_Li11"], "authors_source": "OpenReview API", "abstract": "How can we extend a pre-trained model to many language understanding tasks, without labeled or additional unlabeled data? Pre-trained language models (PLMs) have been effective for a wide range of NLP tasks. However, existing approaches either require fine-tuning on downstream labeled datasets or manually constructing proper prompts. In this paper, we propose nonparametric prompting PLM (NPPrompt) for fully zero-shot language understanding. Unlike previous methods, NPPrompt uses only pre-trained language models and does not require any labeled data or additional raw corpus for further fine-tuning, nor does it rely on humans to construct a comprehensive set of prompt label words. We evaluate NPPrompt against previous major few-shot and zero-shot learning methods on diverse NLP tasks: including text classification, text entailment, similar text retrieval, and paraphrasing. Experimental results demonstrate that our NPPrompt outperforms the previous best fully zero-shot method by big margins, with absolute gains of 12.8% in accuracy on text classification and 18.9% on the GLUE benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BzoQuofHjas", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3683/Reviewer_DKg9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper presents a novel method for zero-shot inference with masked language models (MLMs) such as BERT and RoBERTa for classification problems. The key idea is to use the encoders of the MLMs themselves to find relevant words for a given label name. For example, if the target label is \"SPORTS\", then the proposed method will first find the words by searching k-nearest neighbors with embedding distance. The embeddings are induced by the MLM itself. Finally, they aggregate the logits of these k words (for filling the masked positions). The authors argue that this new method (named NPPrompt) is a fully zero-shot classification method because it does not need any other information about labels. The empirical comparisons show that the proposed method has a comparable performance with other zero-shot methods (that needs some additional info such as KB or the unlabeled data). \n\n", "review_text": "Overall, I think the paper presents a novel zero-shot classification paper but the limitations are there and the performance is not that significant. I think if the authors can extend the paper with my suggestions in the above section, the paper will be more suitable to the ICLR community. Otherwise, it will have much less impact to the field. ", "strengths": "## Strengths\n\n- The proposed NPPrompt method is simple and effective. It only requires the label strings for doing zero-shot classification. \n- The empirical results show that it is on the same level as other methods that need more information about the labels.\n\n## Weakness \n\n- The proposed method is pretty limited to the classification tasks, particularly for news classification, where the label names are informative and have many semantic relevant words in the vocabulary. It is also limited when the label names are multi-word expressions or phrases. It does not support the cases where the label names are not informative enough and additional label description is needed. \n\n- The title and narrative can be misleading and an overclaim. As mentioned before, it is limited to masked LMs and the advantage is mainly for tasks like news classification. \n\nSuggestion:\n\n- I suggest the authors extend the proposed method to *multi-label classification* settings where an example can have multiple acceptable news. \n- A major limitation is the multi-word expression and out-of-vocabluary label names.  How do you generalize the method to support that?\n- Also, I suggest the authors try some multiple-choice QA problems.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presents a novel method for zero-shot inference with masked language models (MLMs) such as BERT and RoBERTa for classification problems. The key idea is to use the encoders of the MLMs themselves to find relevant words for a given label name. For example, if the target label is \"SPORTS\", then the proposed method will first find the words by searching k-nearest neighbors with embedding distance. The embeddings are induced by the MLM itself. Finally, they aggregate the logits of these k words (for filling the masked positions). The authors argue that this new method (named NPPrompt) is a fully zero-shot classification method because it does not need any other information about labels. The empirical comparisons show that the proposed method has a comparable performance with other zero-shot methods (that needs some additional info such as KB or the unlabeled data). \n\n", "strength_and_weaknesses": "## Strengths\n\n- The proposed NPPrompt method is simple and effective. It only requires the label strings for doing zero-shot classification. \n- The empirical results show that it is on the same level as other methods that need more information about the labels.\n\n## Weakness \n\n- The proposed method is pretty limited to the classification tasks, particularly for news classification, where the label names are informative and have many semantic relevant words in the vocabulary. It is also limited when the label names are multi-word expressions or phrases. It does not support the cases where the label names are not informative enough and additional label description is needed. \n\n- The title and narrative can be misleading and an overclaim. As mentioned before, it is limited to masked LMs and the advantage is mainly for tasks like news classification. \n\nSuggestion:\n\n- I suggest the authors extend the proposed method to *multi-label classification* settings where an example can have multiple acceptable news. \n- A major limitation is the multi-word expression and out-of-vocabluary label names.  How do you generalize the method to support that?\n- Also, I suggest the authors try some multiple-choice QA problems.", "clarity,_quality,_novelty_and_reproducibility": "- The paper is mostly clear and the novelty is incremental. I think the reproducibility should be okay since the method is pretty naive. \n\n- Can you show the kNN results for sentiment analysis tasks and NLI tasks? I don't find them in the paper but they are quite important for qualitative analysis. ", "summary_of_the_review": "Overall, I think the paper presents a novel zero-shot classification paper but the limitations are there and the performance is not that significant. I think if the authors can extend the paper with my suggestions in the above section, the paper will be more suitable to the ICLR community. Otherwise, it will have much less impact to the field. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667520503347}, {"id": "f3K8F_mBpoG", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3683/Reviewer_LMgP"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies how to effectively transfer pretrained language models to natural language understanding (NLU) tasks in a zero-shot manner. The proposed method, NPPrompt, does not require any labeled sample or rely on humans to construct prompt label words. NPPrompt generates the label words by searching the related words from the initial word embedding of the pre-trained language model. Then, it aggregates logits from the label words and predicts the category with the largest score. Experiments show that the proposed method is effective and outperforms strong baselines with a large margin.", "review_text": "This paper proposed NPPrompt, which aims to enable LMs' zero-shot ability to NLU tasks without requiring any labeled sample or relying on humans to construct prompt label words. Results show that NNPrompt outperforms strong baselines with a large margin. However, the overall novelty of this paper is limited. ", "strengths": "Strength\n1. Proposed method is simple, intuitive and effective.\n2. Conduct extensive experiments on a wide range of NLU tasks.\n\nWeaknesses\n1. The proposed method is not very novel. Generalizing LMs to NLU tasks by checking the vocabulary distribution over the masked token is a common practice. The method which leverages related words to label category does not look very novel to me.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies how to effectively transfer pretrained language models to natural language understanding (NLU) tasks in a zero-shot manner. The proposed method, NPPrompt, does not require any labeled sample or rely on humans to construct prompt label words. NPPrompt generates the label words by searching the related words from the initial word embedding of the pre-trained language model. Then, it aggregates logits from the label words and predicts the category with the largest score. Experiments show that the proposed method is effective and outperforms strong baselines with a large margin.", "strength_and_weaknesses": "Strength\n1. Proposed method is simple, intuitive and effective.\n2. Conduct extensive experiments on a wide range of NLU tasks.\n\nWeaknesses\n1. The proposed method is not very novel. Generalizing LMs to NLU tasks by checking the vocabulary distribution over the masked token is a common practice. The method which leverages related words to label category does not look very novel to me.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-written. The work is original, but the proposed method is kind of similar to many existing works, which makes it less novel.", "summary_of_the_review": "This paper proposed NPPrompt, which aims to enable LMs' zero-shot ability to NLU tasks without requiring any labeled sample or relying on humans to construct prompt label words. Results show that NNPrompt outperforms strong baselines with a large margin. However, the overall novelty of this paper is limited. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666679471853}, {"id": "IzAu2fWH63", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3683/Reviewer_Gywf"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors propose a new language model named non parametric prompting PLM for natural language understanding specially for zero-shot learning. It is an important topic because these days many word-class associations are being produced by end users and previous models heavily depend on unlabeled data and human effort.  The authors showed that the proposed method outperforms state-of-the-art in terms of text classification accuracy and GLUE benchmarks on four different datasets including AG news, DBPedia, IMDB. and Amazon.", "review_text": "Overall, zero shot learning is an interesting topic in natural language processing as so many new categories and topics are being produced on the web. The authors proposed a simple and easy to implement method for pre trained language models to minimize human effort in terms of labeling and building training data. Overall I am satisfied with the current draft of the paper and request to move forward with discussion. ", "strengths": "The authors put significant effort on proving effectiveness of their method in a variety of NLP tasks. However, I wanted to see significant test results to make sure that the improvements are not random. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors propose a new language model named non parametric prompting PLM for natural language understanding specially for zero-shot learning. It is an important topic because these days many word-class associations are being produced by end users and previous models heavily depend on unlabeled data and human effort.  The authors showed that the proposed method outperforms state-of-the-art in terms of text classification accuracy and GLUE benchmarks on four different datasets including AG news, DBPedia, IMDB. and Amazon.", "strength_and_weaknesses": "The authors put significant effort on proving effectiveness of their method in a variety of NLP tasks. However, I wanted to see significant test results to make sure that the improvements are not random. ", "clarity,_quality,_novelty_and_reproducibility": "The paper was well written and easy to follow. I would require authors to add the github link for the code. ", "summary_of_the_review": "Overall, zero shot learning is an interesting topic in natural language processing as so many new categories and topics are being produced on the web. The authors proposed a simple and easy to implement method for pre trained language models to minimize human effort in terms of labeling and building training data. Overall I am satisfied with the current draft of the paper and request to move forward with discussion. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666655280231}], "openreview_url": "https://openreview.net/forum?id=jCpTofV7iY_", "arxiv_id": "2212.06950", "paper_pdf": "papers/jCpTofV7iY_.pdf", "paper_pdf_sha256": "771c9cd1dea75b28c1d41fbed92df94bd5920d1652e8b9b12911f41893873204", "paper_pdf_bytes": 434752, "paper_pdf_source": "openreview", "code_url": "https://github.com/XuandongZhao/NPPrompt", "code_repository": "XuandongZhao/NPPrompt", "code_commit": "2c30c7ba47bbbccd8e3e627e7b06b16c5229cf52", "code_archive": "repos/jCpTofV7iY_.zip", "code_archive_sha256": "d55f1e35c947d1fa2c3ed6f3b59bf282ec57106d180443f46450a4bcba3996f2", "code_archive_bytes": 266871, "code_file_count": 68, "code_extensions": {".py": 66, ".sh": 2}, "github_disk_usage_kb": 216, "github_languages": {"Python": 555249, "Shell": 2161}, "github_archived": false, "github_pushed_at": "2024-02-02T18:33:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pre-trained-language-models-can-be-fully-zero"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Fj1Tpym9KxH", "year": 2022, "status": "rejected", "title": "A Closer Look at Smoothness in Domain Adversarial Training", "authors": ["Harsh Rangwani", "Sumukh K Aithal", "Arihant Jain", "Venkatesh Babu Radhakrishnan"], "authorids": ["~Harsh_Rangwani1", "~Sumukh_K_Aithal1", "arihantjain@iisc.ac.in", "~Venkatesh_Babu_Radhakrishnan2"], "authors_source": "OpenReview API", "abstract": "Domain adversarial training has been ubiquitous for achieving invariant representations and is used widely for various domain adaptation tasks. In recent times methods converging to smooth optima have shown improved generalization for supervised learning tasks like classification.  In this work, we analyze the effect of smoothness enhancing formulations on domain adversarial training, the objective of which is a combination of classification and adversarial terms. In contrast to classification loss, our analysis shows that \\textit{converging to smooth minima w.r.t. adversarial loss leads to sub-optimal generalization on the target domain}. Based on the analysis, we introduce the Smooth Domain Adversarial training (SDAT) procedure, which effectively enhances the performance of existing domain adversarial methods for both classification and object detection tasks. Our smoothness analysis also provides insight into the extensive usage of SGD over Adam in domain adversarial training.   ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AsLmszM_d8k", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3289/Reviewer_rxW2"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work studies the loss landscape of domain adversarial neural networks for domain adaptation. The authors claim that similarly to the iid setting, smooth minima might yield better generalization on the unsupervised domain adaptation framework via domain adversarial training but in case smoothness is enforced only with respect to the task loss. Experiments on the OfficeHome dataset indicated this hypothesis might be true. The authors then proposed to apply a previously proposed approach to enforce smoothness (Sharpness Aware Minimization), so that only smoothness with respect to the task loss would be enforced. Moreover, a theorem showing that the non-smooth domain discrepancy estimation is better than the smoother counterpart was introduced, along with a generalization bound for the risk on the target domain in terms of the smooth risk on the source domain. Finally, the authors showed that the proposed approach, SDAT, improved the performance on target domains of three domain adaptation datasets in comparison to the considered baselines. Further experiments showed that SDAT performed better than other smoothing techniques, as well as yields improved robustness to label noise in comparison to not enforcing smooth minima. ", "review_text": "The main contribution of this work is to propose applying Sharpness Aware Minimization (SAM) to smooth only the task loss term of domain adversarial training. From an empirical standpoint the paper is interesting and points to potentially relevant investigation directions, however, I have concerns regarding the significance of this study since there is no clear motivation as to why previous findings on the iid setting showing that smoother minima generalize better would transfer to the unsupervised domain adaptation setting. Moreover, the empirical validation of the proposed approach is not extensive enough to provide sufficient evidence for the soundness of the reported findings. In the following, I provide more details about my concerns along with questions and suggestions.\n\n- Motivation:\n  - Although the authors (to some extent) empirically show that SDAT can improve the performance on the target domain on the considered cases, there is no justification on why one would expect better out-of-distribution generalization for smoother minimum. Would the authors clarify that? More specifically, what is the motivation/context for investigating Conjecture 1? As of now, it seems to me that this work is extending an analysis for the iid setting to unsupervised domain adaptation without a clear reason. \n  - It is not clear to me why the authors decided to provide results that rely on Acuna et al. 2021 rather than consider the classic results from Ben-David et al. 2010. There is no clear motivation for this choice in the manuscript and I believe the authors should better justify this choice.\n\n\n- Results:\n  - The relevance of both theoretical results are not clear to me. Theorem 2, for example, requires unrealistic assumptions such as L-smoothness. Theorem 3, in turn, states a bound that is looser than the known results in the literature and which are already vacuous, so it is not clear to me what is the relevance of this result in the context of unsupervised domain adaptation.\n  - It is hard to assess the relevance of the reported results as no measures of statistical significance were reported. More specifically, the authors should include in the paper information regarding the number of runs considered for each experiment and report at least the average and standard deviation for all the quantities reported. Otherwise, I don’t think it is possible to rely on the current experimental set-up to support the claims of this contribution.\n  - Please clarify how the 50% partition of the source data was selected to estimate the Hessian for the three compared methods and if the same partition was used for all, otherwise results on Figure 3 could be noisy.\n  - SDAT requires an extra gradient step in order to estimate the smoothing penalty in Eq. 7. How does this affect the computational cost of SDAT and how SDAT stands in comparison with the considered baselines?\n\n\n- Other concerns/questions:\n  - Some statements of the manuscript are not rigorous and require rephrasing:\n     - Abstract: “combination of classification and adversarial terms”. In the considered setting, the adversarial term of the loss also stems from a classification task (domain discrimination), therefore, it is confusing to the denominate both component losses in such a way. I suggest referring to the loss term related to classifying labels regarding a specific task as task loss.\n     - Section 3.1: please specify what “works well on the dataset” means (i.e. low risk on the target distribution).\n     - Theorems 1/3: please precisely define what is the ideal classifier $h^*$.\n     - Section 4: “where we find that in contrast to supervised learning” notice that, in practice, unsupervised domain adaptation also corresponds to supervised learning since all losses require either task or domain labels, so it shouldn’t be referred to as a setting opposed to supervised learning.\n  - Table 1 caption: what exactly “sophisticated” means here? \n\n\n\n- Related work:\n  - Foundational literature on domain adaptation is missing in the related work section, for example, Ben-David et al. (2007) and Long et al. (2016). \n\nBen-David et al., Analysis of representations for domain adaptation, 2007.\nLong et al., Unsupervised domain adaptation with residual transfer networks, 2016.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work studies the loss landscape of domain adversarial neural networks for domain adaptation. The authors claim that similarly to the iid setting, smooth minima might yield better generalization on the unsupervised domain adaptation framework via domain adversarial training but in case smoothness is enforced only with respect to the task loss. Experiments on the OfficeHome dataset indicated this hypothesis might be true. The authors then proposed to apply a previously proposed approach to enforce smoothness (Sharpness Aware Minimization), so that only smoothness with respect to the task loss would be enforced. Moreover, a theorem showing that the non-smooth domain discrepancy estimation is better than the smoother counterpart was introduced, along with a generalization bound for the risk on the target domain in terms of the smooth risk on the source domain. Finally, the authors showed that the proposed approach, SDAT, improved the performance on target domains of three domain adaptation datasets in comparison to the considered baselines. Further experiments showed that SDAT performed better than other smoothing techniques, as well as yields improved robustness to label noise in comparison to not enforcing smooth minima. ", "main_review": "The main contribution of this work is to propose applying Sharpness Aware Minimization (SAM) to smooth only the task loss term of domain adversarial training. From an empirical standpoint the paper is interesting and points to potentially relevant investigation directions, however, I have concerns regarding the significance of this study since there is no clear motivation as to why previous findings on the iid setting showing that smoother minima generalize better would transfer to the unsupervised domain adaptation setting. Moreover, the empirical validation of the proposed approach is not extensive enough to provide sufficient evidence for the soundness of the reported findings. In the following, I provide more details about my concerns along with questions and suggestions.\n\n- Motivation:\n  - Although the authors (to some extent) empirically show that SDAT can improve the performance on the target domain on the considered cases, there is no justification on why one would expect better out-of-distribution generalization for smoother minimum. Would the authors clarify that? More specifically, what is the motivation/context for investigating Conjecture 1? As of now, it seems to me that this work is extending an analysis for the iid setting to unsupervised domain adaptation without a clear reason. \n  - It is not clear to me why the authors decided to provide results that rely on Acuna et al. 2021 rather than consider the classic results from Ben-David et al. 2010. There is no clear motivation for this choice in the manuscript and I believe the authors should better justify this choice.\n\n\n- Results:\n  - The relevance of both theoretical results are not clear to me. Theorem 2, for example, requires unrealistic assumptions such as L-smoothness. Theorem 3, in turn, states a bound that is looser than the known results in the literature and which are already vacuous, so it is not clear to me what is the relevance of this result in the context of unsupervised domain adaptation.\n  - It is hard to assess the relevance of the reported results as no measures of statistical significance were reported. More specifically, the authors should include in the paper information regarding the number of runs considered for each experiment and report at least the average and standard deviation for all the quantities reported. Otherwise, I don’t think it is possible to rely on the current experimental set-up to support the claims of this contribution.\n  - Please clarify how the 50% partition of the source data was selected to estimate the Hessian for the three compared methods and if the same partition was used for all, otherwise results on Figure 3 could be noisy.\n  - SDAT requires an extra gradient step in order to estimate the smoothing penalty in Eq. 7. How does this affect the computational cost of SDAT and how SDAT stands in comparison with the considered baselines?\n\n\n- Other concerns/questions:\n  - Some statements of the manuscript are not rigorous and require rephrasing:\n     - Abstract: “combination of classification and adversarial terms”. In the considered setting, the adversarial term of the loss also stems from a classification task (domain discrimination), therefore, it is confusing to the denominate both component losses in such a way. I suggest referring to the loss term related to classifying labels regarding a specific task as task loss.\n     - Section 3.1: please specify what “works well on the dataset” means (i.e. low risk on the target distribution).\n     - Theorems 1/3: please precisely define what is the ideal classifier $h^*$.\n     - Section 4: “where we find that in contrast to supervised learning” notice that, in practice, unsupervised domain adaptation also corresponds to supervised learning since all losses require either task or domain labels, so it shouldn’t be referred to as a setting opposed to supervised learning.\n  - Table 1 caption: what exactly “sophisticated” means here? \n\n\n\n- Related work:\n  - Foundational literature on domain adaptation is missing in the related work section, for example, Ben-David et al. (2007) and Long et al. (2016). \n\nBen-David et al., Analysis of representations for domain adaptation, 2007.\nLong et al., Unsupervised domain adaptation with residual transfer networks, 2016.\n", "summary_of_the_review": "This work studies the loss landscape of adversarial domain adaptation in terms of the smoothness of the minima. Despite showing somewhat empirically promising results, I found this work lacks a solid motivation for the presented analysis (i.e. why should we expect that strategies that improve generalization in the iid case would help in a non-iid setting as well?) and a sound empirical validation of the proposed approach. Moreover, the relevance of the presented generalization bound is also unclear to me at this point. All in all, I think this contribution seems promising but due to the concerns I raised in my review, I believe it is not yet ready for publication. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635884269259}, {"id": "FVwisolwsz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3289/Reviewer_zPwA"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper discussed a correlation of smoothness on **classification loss** and **generalization ability on target domain** and further develop a method to enhance such smoothness to achieve better performance on DA tasks. To do so, it adopted the losses based on Sharpness Aware Minimization, a minmax game of finding a smoother neighborhood of $\\theta$ , by computing first order deviation. Empirical studies are implemented to verify the theory and show the soundness of the proposed method.", "review_text": "## Strong points\n\n1. The paper takes issue with an interesting point of the reason for SGD optimizer outperforming Adam on DA tasks for reaching a smoother minima, develops it theoretically, and proposed a novel loss function focusing on smoothing classifier loss to improve generalization on target domain.\n\n2. It fully discussed why we should not apply the smoothing method **on the discrepancy term**, and verify the theory empirically through experiment, further developing smooth theory with regard to **adversarial objectives.**\n\n## Weak points\n\n1. The method for smoothing ERM is more directly adopted from SAM rather than a novel idea, and the relation **between** smoothness **and** better generalization has also been discussed by previous papers (SWAD), therefore renders this paper not insightful enough. \n\n2. The proposed smoothing method w.r.t discrepancy term tend to \"lead to a suboptimal solution\", therefore not significantly helpful when tackling DA tasks (and I think it is hard to interpret). What actually makes such difference on the finally result is still, smoothness on ERM term.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper discussed a correlation of smoothness on **classification loss** and **generalization ability on target domain** and further develop a method to enhance such smoothness to achieve better performance on DA tasks. To do so, it adopted the losses based on Sharpness Aware Minimization, a minmax game of finding a smoother neighborhood of $\\theta$ , by computing first order deviation. Empirical studies are implemented to verify the theory and show the soundness of the proposed method.", "main_review": "## Strong points\n\n1. The paper takes issue with an interesting point of the reason for SGD optimizer outperforming Adam on DA tasks for reaching a smoother minima, develops it theoretically, and proposed a novel loss function focusing on smoothing classifier loss to improve generalization on target domain.\n\n2. It fully discussed why we should not apply the smoothing method **on the discrepancy term**, and verify the theory empirically through experiment, further developing smooth theory with regard to **adversarial objectives.**\n\n## Weak points\n\n1. The method for smoothing ERM is more directly adopted from SAM rather than a novel idea, and the relation **between** smoothness **and** better generalization has also been discussed by previous papers (SWAD), therefore renders this paper not insightful enough. \n\n2. The proposed smoothing method w.r.t discrepancy term tend to \"lead to a suboptimal solution\", therefore not significantly helpful when tackling DA tasks (and I think it is hard to interpret). What actually makes such difference on the finally result is still, smoothness on ERM term.", "summary_of_the_review": "Overall, the author has well established a positive correlation between smoothness of classification loss and generalization capability on target domain, and a negative correlation between adversarial counterparts. The method of acquiring a flat minima also brings promising results on multiple DA classification and detection datasets. However, as I can see the experimental studies of the proposed smoothing method are well-designed and theories are rigorously proved, I am not completely convinced by the novelty of the work. Hope the author can make further analysis and explanation.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635872368026}, {"id": "bAXv5WYu5r", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3289/Reviewer_C1yh"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper introduces a smoothness penalty for domain adversarial training. The penalty encourages smoothness on the parameter space of a discriminative model. This smoothness penalty is motivated by results in Sharpness Aware Minimization. Results are included on the popular DA datasets such as Office-home and VisDA, where superior results are shown. Ablation experiments show how different smoothness penalties change the observed improvements. ", "review_text": "Strengths:\n\n  * Theoretical foundation for smoothness\n  * Results sections include improvements on both image classification and object detection. \n  * Ablation experiments provide insights to the effect of a smoothness parameter. \n\nWeaknesses:\n\n  * Figure 4: not clear why and how the optimum for rho differs per dataset. Does this depend on the dataset size or diversity of objects? Moreover, the vanilla ResNets have a batchnorm layer, so it is not clear why figure 4a has such a sharp optimum at all. \n* What is the final recommendation of the paper? If the paper recommends using smoothness penalties, then what are the limits? For example, how does the smoothness of the classifier (focus of this study) interplay with smoothness of the domain discriminator? I could imagine that when the discriminator has sharp decision boundaries, then enforcing smoothness on the classifier would be less useful. (But in practise people employ many tricks to get a smooth discriminator [1][2]).\n* Table 1: most improvements occur when combining SDAT with MCC. This combination is not well explained in the paper. What makes these approaches complementary and why are improvements small (or non-existent) for vanilla CDAN?\n* Theorem 2: I don’t see how this theorem relates to the point of the paper. From equation 11 it seems smoothness is applied to h_\\theta. \n\n[1] Gulrajani et al. \"Improved training of wasserstein gans.\" arXiv 2017.\n\n[2] Arjovsky et al. \"Towards principled methods for training generative adversarial networks.\" arXiv 2017.\n\n[3] Miyato et al. \"Spectral normalization for generative adversarial networks.\" arXiv 2018.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a smoothness penalty for domain adversarial training. The penalty encourages smoothness on the parameter space of a discriminative model. This smoothness penalty is motivated by results in Sharpness Aware Minimization. Results are included on the popular DA datasets such as Office-home and VisDA, where superior results are shown. Ablation experiments show how different smoothness penalties change the observed improvements. ", "main_review": "Strengths:\n\n  * Theoretical foundation for smoothness\n  * Results sections include improvements on both image classification and object detection. \n  * Ablation experiments provide insights to the effect of a smoothness parameter. \n\nWeaknesses:\n\n  * Figure 4: not clear why and how the optimum for rho differs per dataset. Does this depend on the dataset size or diversity of objects? Moreover, the vanilla ResNets have a batchnorm layer, so it is not clear why figure 4a has such a sharp optimum at all. \n* What is the final recommendation of the paper? If the paper recommends using smoothness penalties, then what are the limits? For example, how does the smoothness of the classifier (focus of this study) interplay with smoothness of the domain discriminator? I could imagine that when the discriminator has sharp decision boundaries, then enforcing smoothness on the classifier would be less useful. (But in practise people employ many tricks to get a smooth discriminator [1][2]).\n* Table 1: most improvements occur when combining SDAT with MCC. This combination is not well explained in the paper. What makes these approaches complementary and why are improvements small (or non-existent) for vanilla CDAN?\n* Theorem 2: I don’t see how this theorem relates to the point of the paper. From equation 11 it seems smoothness is applied to h_\\theta. \n\n[1] Gulrajani et al. \"Improved training of wasserstein gans.\" arXiv 2017.\n\n[2] Arjovsky et al. \"Towards principled methods for training generative adversarial networks.\" arXiv 2017.\n\n[3] Miyato et al. \"Spectral normalization for generative adversarial networks.\" arXiv 2018.\n", "summary_of_the_review": "Paper with theoretical motivation and results on both classification and object detection. However, the limitations of the proposed method are not clear. When would smoothness not be useful and how do the improvements depend on the other tricks for training Domain Adaptation models. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635842222427}, {"id": "VOrOOdMyoF2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3289/Reviewer_nwci"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper analyzes the role of smoothness in domain adversarial training (DAT). The main insight of the paper is that smoother minima of the classification loss improve generalization in the target domain. This explains why SGD is usually preferred over Adam when optimizing the training objective of DAT. To further improve the generalization performance by leveraging this phenomenon the author(s) introduce(s) smooth domain adversarial training (SDAT) for classification and object detection. SDAT applies sharpness aware minimization (SAM) to find smoother minimia of the classification loss used during DAT.", "review_text": "STRENGTHS:\n- The paper is very well-written and introduces all used concepts appropriately. This makes the paper easy to understand even for readers that are not familiar with domain adaption.\n- The motivational experiments performed in Section 4 (Analysis of Smoothness) are well-explained and convincing.\n- The author(s) provide(s) a generalization bound (Theorem 2) for their proposed method which justifies it from a theoretical perspective.\n- The experimental evaluation is very extensive: the method is evaluated on three different domain adaption datasets for classification (Office-Home, VisDA-2017, and DomainNet) and compared with different existing domain adaption methods. It is shown that SDAT can be combined with the existing methods CDAN and CDAN + MCC to increase generalization performance.\n- Furthermore, improved performance on domain adaption datasets for object detection is reported.\n- Additional experiments show that SDAT also increases the robustness to label noise and that SAM outperforms other smoothing techniques. \n\nWEAKNESSES:\n- I see not major issues with this submission.\n\nMINOR REMAKS:\n- Some references are not correctly capitalized (e.g., \"Lower bounds for finding stationary points i\" or \"Faster r-cnn\") also conference names are sometimes capitalized and sometimes not.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper analyzes the role of smoothness in domain adversarial training (DAT). The main insight of the paper is that smoother minima of the classification loss improve generalization in the target domain. This explains why SGD is usually preferred over Adam when optimizing the training objective of DAT. To further improve the generalization performance by leveraging this phenomenon the author(s) introduce(s) smooth domain adversarial training (SDAT) for classification and object detection. SDAT applies sharpness aware minimization (SAM) to find smoother minimia of the classification loss used during DAT.", "main_review": "STRENGTHS:\n- The paper is very well-written and introduces all used concepts appropriately. This makes the paper easy to understand even for readers that are not familiar with domain adaption.\n- The motivational experiments performed in Section 4 (Analysis of Smoothness) are well-explained and convincing.\n- The author(s) provide(s) a generalization bound (Theorem 2) for their proposed method which justifies it from a theoretical perspective.\n- The experimental evaluation is very extensive: the method is evaluated on three different domain adaption datasets for classification (Office-Home, VisDA-2017, and DomainNet) and compared with different existing domain adaption methods. It is shown that SDAT can be combined with the existing methods CDAN and CDAN + MCC to increase generalization performance.\n- Furthermore, improved performance on domain adaption datasets for object detection is reported.\n- Additional experiments show that SDAT also increases the robustness to label noise and that SAM outperforms other smoothing techniques. \n\nWEAKNESSES:\n- I see not major issues with this submission.\n\nMINOR REMAKS:\n- Some references are not correctly capitalized (e.g., \"Lower bounds for finding stationary points i\" or \"Faster r-cnn\") also conference names are sometimes capitalized and sometimes not.", "summary_of_the_review": "Overall, I would highly recommend to accept this submission to ICLR 2022. The theoretical and empirical results are convincing and the paper is of high quality.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635761184390}], "openreview_url": "https://openreview.net/forum?id=Fj1Tpym9KxH", "arxiv_id": "2206.08213", "paper_pdf": "papers/Fj1Tpym9KxH.pdf", "paper_pdf_sha256": "5f2fca3ef01ecf3c34dbe7f76aea417c827bd988feba63940c8e4d2de81717b9", "paper_pdf_bytes": 3183844, "paper_pdf_source": "openreview", "code_url": "https://github.com/val-iisc/SDAT", "code_repository": "val-iisc/SDAT", "code_commit": "33d7429d37522972edda608cb802a686e4b9e794", "code_archive": "repos/Fj1Tpym9KxH.zip", "code_archive_sha256": "27f4c120735b5d9a03c5a947f5e1e2b9c57ff65b672f72ac3768e65cf4a78811", "code_archive_bytes": 541112, "code_file_count": 39, "code_extensions": {".py": 37, ".sh": 2}, "github_disk_usage_kb": 491, "github_languages": {"Python": 84834}, "github_archived": false, "github_pushed_at": "2024-04-11T22:11:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-closer-look-at-smoothness-in-domain-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "aYbCpFNnHdh", "year": 2021, "status": "rejected", "title": "Visual Question Answering From Another Perspective: CLEVR Mental Rotation Tests", "authors": ["Christopher Beckham", "Martin Weiss", "Florian Golemo", "Sina Honari", "Derek Nowrouzezahrai", "Christopher Pal"], "authorids": ["~Christopher_Beckham1", "~Martin_Weiss4", "~Florian_Golemo1", "~Sina_Honari1", "~Derek_Nowrouzezahrai1", "~Christopher_Pal1"], "authors_source": "OpenReview API", "abstract": "Different types of \\emph{mental rotation tests} have been used extensively in psychology to understand human visual reasoning and perception. Understanding what an object or visual scene would look like from another viewpoint is a challenging problem that is made even harder if it must be performed from a single image. 3D computer vision has a long history of examining related problems. However, often what one is most interested in is the answer to a relatively simple question posed in another visual frame of reference -- as opposed to creating a full 3D reconstruction. \nMental rotations tests can also manifest as consequential questions in the real world such as: does the pedestrian that I see, see the car that I am driving?\nWe explore a controlled setting whereby questions are posed about the properties of a scene if the scene were observed from another viewpoint. To do this we have created a new version of the CLEVR VQA problem setup and dataset that we call CLEVR Mental Rotation Tests or CLEVR-MRT, where the goal is to answer questions about the original CLEVR viewpoint given a single image obtained from a different viewpoint of the same scene. Using CLEVR Mental Rotation Tests we examine standard state of the art methods, show how they fall short, then explore novel neural architectures that involve inferring representations encoded as feature volumes describing a scene. Our new methods use rigid transformations of feature volumes conditioned on the viewpoint camera. We examine the efficacy of different model variants through performing a rigorous ablation study. Furthermore, we examine the use of contrastive learning to infer a volumetric encoder in a self-supervised manner and find that this approach yields the best results of our study using CLEVR-MRT.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "_BrD2XsKL1h", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3392/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n### Overall\n\nAuthors extend CLEVR dataset so as to consider multiple viewpoints, and evaluate current neural network models in that setting. They also update a standard approach to introduce camera viewpoint information in the network so it can better answer visual question from the canonical scene frame even from other perspectives.\n\n### Positive aspects\n\n* Authors provide a study on a important topic of Computer Vision: understanding multiple views of a same scene. They do such study on a hard task, which is VQA. Actually, authors provide a more complex version of a simple VQA dataset (*simple* because it is synthetic and has very well established domain limits).\n* Authors evaluate different training frameworks (supervised and unsupervised from scratch).\n* Authors provided accuracy values for pretraining with NCE, which can be helpful.\n* Results seem to be promising.\n* In general, text is well written and easy to read.\n* It is interesting that even a frozen pretrained network provides good results in such visually different dataset. Although, it was nice that authors trained an encoder from scratch.\n* Code already available!\n\n### Weak aspects and suggestions\n\n* The problem is interesting, though my main concern is regarding the novelty and contribution of the paper. It seems to be an adaptation of CLEVR dataset, and an adaptation of the FILM model. In addition, authors use camera viewpoint information to ease the identification of the scenes. I have mixed feelings in using such specific kind of information in the model, because in a real world scenario we don't have access to them. I might be wrong, but maybe it is possible to insert a module in their approach to estimate the camera parameters, so as the network itself could learn to predict how viewpoints work and how scenes change with that. I think this could be done by adding such parameters as target information some of the models. For instance, the unsupervised architecture could be trained to predict whether the scenes are the same, but also the camera parameters. Apologies if I miss something here.\n\n* The proposed architecture seems to be basically an adaptation of the FILM model considering camera viewpoint information.\n\n* FILM (2018) is the best performing approach in CLEVER to date? There are more recent approaches that could be used in the results section as baselines.\n\n* It is unclear what happened to the spatial-related questions. They were removed of the dataset?\n\n* Results are promising, although why do they have such high variance? (7-8% of variance is not negligible by any means); considering that for some experiments it is likely that 2D FILM provides similar performance than 3D one. A statistical test might help to verify whether such results are statistically significant or not.\n\n* Font size for all images should be quite larger. It is hard to read in the current size.\n\n* Figure of the post processor does not help much. Authors could detail a little bit more what is inside that $postproc_w$ box.\n\n* *\"Since the post-processor is a learnable module through which the FILM part of the pipeline is able to backpropagate through, it can be seen as learning an appropriate set of transforms that construct 3D feature volumes h0.\"* I suggest rewriting this sentence, it is very confusing.\n\n* *\"While we obtained great results, it may not leave a lot of room to improve on top of our methods,\"* This sentence is odd. The sentence \"we obtained great results\" can be written in a more objective and scientific way (avoid the usage of adjectives). Another important aspect is: often it is easy to provide first large steps in a task (ImageNet for instance), although it gets much harder to improve on that when results are good (AlexNet vs ResNet, see the performance difference). Another aspect: maybe authors made the task too easy and should have explored more challenging scenarios.\n\n* *\"and we identified some ways in which the dataset could be made more difficult\"* Those ideas to make the task more challenging are indeed important. Why authors did not perform experiments in such scenarios? It does not seem very hard to generate such datasets.\n\n* Is it possible to visualize and understand what the postproc module does? It would be nice to visually explain the $h'$ (64, 16, 14, 14) tensor represents.\n\n* There could be some qualitative analysis.\n\n* The dataset extension seems to be a large portion of the work. I think it could have a separate section with more details.\n\n### Additional questions\n\n* What happens if other conditioning camera information strategy is used? For instance, simply concatenating or using other simpler fusion techniques. FILM would perform much better than other simpler approaches?\n\n* *\"ResNet outputs... feature maps h of dimensions (1024,14, 14)\"* Is this correct? I believe Resnet101 outputs (2048, 14, 14) feature maps.\n\n* *\"in practice, we found $\\tau = 0.1$ to produce the lowest softmax loss.\"* Which ones you have tested? Why $\\tau$ is 1.0 in Table 2?\n\n* *\"Another idea is to allow the viewpoint camera’s elevation to change. \"* That is true. Or even the distance from the camera. Why did authors decide not to include such examples in this work?\n\n* *\"This is to be expected, considering that any camera information that is forward-propagated will contribute gradients back to the postprocessing parameters in the backward propagation, effectively giving the postprocessor supervision in the form of camera extrinsics.\"*. Can authors support/prove this claim?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Review", "review": "\n### Overall\n\nAuthors extend CLEVR dataset so as to consider multiple viewpoints, and evaluate current neural network models in that setting. They also update a standard approach to introduce camera viewpoint information in the network so it can better answer visual question from the canonical scene frame even from other perspectives.\n\n### Positive aspects\n\n* Authors provide a study on a important topic of Computer Vision: understanding multiple views of a same scene. They do such study on a hard task, which is VQA. Actually, authors provide a more complex version of a simple VQA dataset (*simple* because it is synthetic and has very well established domain limits).\n* Authors evaluate different training frameworks (supervised and unsupervised from scratch).\n* Authors provided accuracy values for pretraining with NCE, which can be helpful.\n* Results seem to be promising.\n* In general, text is well written and easy to read.\n* It is interesting that even a frozen pretrained network provides good results in such visually different dataset. Although, it was nice that authors trained an encoder from scratch.\n* Code already available!\n\n### Weak aspects and suggestions\n\n* The problem is interesting, though my main concern is regarding the novelty and contribution of the paper. It seems to be an adaptation of CLEVR dataset, and an adaptation of the FILM model. In addition, authors use camera viewpoint information to ease the identification of the scenes. I have mixed feelings in using such specific kind of information in the model, because in a real world scenario we don't have access to them. I might be wrong, but maybe it is possible to insert a module in their approach to estimate the camera parameters, so as the network itself could learn to predict how viewpoints work and how scenes change with that. I think this could be done by adding such parameters as target information some of the models. For instance, the unsupervised architecture could be trained to predict whether the scenes are the same, but also the camera parameters. Apologies if I miss something here.\n\n* The proposed architecture seems to be basically an adaptation of the FILM model considering camera viewpoint information.\n\n* FILM (2018) is the best performing approach in CLEVER to date? There are more recent approaches that could be used in the results section as baselines.\n\n* It is unclear what happened to the spatial-related questions. They were removed of the dataset?\n\n* Results are promising, although why do they have such high variance? (7-8% of variance is not negligible by any means); considering that for some experiments it is likely that 2D FILM provides similar performance than 3D one. A statistical test might help to verify whether such results are statistically significant or not.\n\n* Font size for all images should be quite larger. It is hard to read in the current size.\n\n* Figure of the post processor does not help much. Authors could detail a little bit more what is inside that $postproc_w$ box.\n\n* *\"Since the post-processor is a learnable module through which the FILM part of the pipeline is able to backpropagate through, it can be seen as learning an appropriate set of transforms that construct 3D feature volumes h0.\"* I suggest rewriting this sentence, it is very confusing.\n\n* *\"While we obtained great results, it may not leave a lot of room to improve on top of our methods,\"* This sentence is odd. The sentence \"we obtained great results\" can be written in a more objective and scientific way (avoid the usage of adjectives). Another important aspect is: often it is easy to provide first large steps in a task (ImageNet for instance), although it gets much harder to improve on that when results are good (AlexNet vs ResNet, see the performance difference). Another aspect: maybe authors made the task too easy and should have explored more challenging scenarios.\n\n* *\"and we identified some ways in which the dataset could be made more difficult\"* Those ideas to make the task more challenging are indeed important. Why authors did not perform experiments in such scenarios? It does not seem very hard to generate such datasets.\n\n* Is it possible to visualize and understand what the postproc module does? It would be nice to visually explain the $h'$ (64, 16, 14, 14) tensor represents.\n\n* There could be some qualitative analysis.\n\n* The dataset extension seems to be a large portion of the work. I think it could have a separate section with more details.\n\n### Additional questions\n\n* What happens if other conditioning camera information strategy is used? For instance, simply concatenating or using other simpler fusion techniques. FILM would perform much better than other simpler approaches?\n\n* *\"ResNet outputs... feature maps h of dimensions (1024,14, 14)\"* Is this correct? I believe Resnet101 outputs (2048, 14, 14) feature maps.\n\n* *\"in practice, we found $\\tau = 0.1$ to produce the lowest softmax loss.\"* Which ones you have tested? Why $\\tau$ is 1.0 in Table 2?\n\n* *\"Another idea is to allow the viewpoint camera’s elevation to change. \"* That is true. Or even the distance from the camera. Why did authors decide not to include such examples in this work?\n\n* *\"This is to be expected, considering that any camera information that is forward-propagated will contribute gradients back to the postprocessing parameters in the backward propagation, effectively giving the postprocessor supervision in the form of camera extrinsics.\"*. Can authors support/prove this claim?\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603904405721}, {"id": "P-OFBR52the", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3392/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper explores the problem of visual question answering from another perspective. Similar to VQA, a system is provided with a scene and a question. However, the difference is that the question needs to be answered from a viewpoint different from the one provided. Hence, the system needs to perform “mental rotation”. The paper creates a new dataset called CLEVR Mental Rotation Tests which is based on the prior CLEVR dataset. The paper also studies the efficacy of various supervised and self-supervised models on the proposed dataset.\n\n#### Strong points:\n- The problem of asking questions related to “mental rotation” seems interesting.\n- The paper shows that contrastive pre-training could be useful for the task, which is an interesting result.\n\n#### Weak points: \n- Although the problem seems interesting, I am unclear about the usefulness of the proposed dataset. The paper says that “many computer vision systems could benefit from neural architectures that demonstrate good performance for more targeted mental rotation tasks.” To justify this claim it gives the following example, “given the camera viewpoint of a (blind) person crossing the road, can we infer if each of the drivers of the cars at an intersection can see this blind person crossing the street?”. This could potentially be a useful scenario, however, the dataset proposed is different from the example as the camera viewpoint is provided as part of the input and not inferred from the question. The paper does not provide justification or evidence of how the current setup (i.e. with camera viewpoint) is useful. In particular, it would be nice if the paper could further explain how the current setup is better than solving the “view rendering” and VQA problems separately.\n\n- The dataset seems to be too simple for the mental rotation tests. It is unclear if in the future the dataset would be useful in distinguishing which models are better. As the paper shows that “2D baseline without camera conditioning” already achieves 70% accuracy. As far as I could understand, even without knowing which view to look at, a model could achieve 70% accuracy indicating that there is a lot of bias in the dataset. Moreover, simply adding camera embedding with the question to a 2D baseline (Table 1, 2D FILM with camera), already performs close to the best 3D model and upper bound. (Please clarify if my understanding is wrong.)\n\n- The paper is poorly organized and hard to follow. For example, one of the contributions of the work is the CLEVR-MTR dataset, however, there is no clear section in the main paper describing the details of how the dataset. Instead, the information about the dataset is scattered in the introduction and related work. Another example is that the paper moves into talking about the method (Section 2) without defining the task concretely. It is only from the figure that one notices that the camera viewpoint is part of the input. From the examples provided in the introduction, the reader is under the impression that the camera viewpoint has to be inferred from the question itself. Similarly, it's hard to parse what the training signal for each baseline is. Does a baseline use the rendered image from the other view during training?\n \n#### Minor Comments:\n- The figures and tables are interspersed with the text making the paper harder to read. It might be better to place the figures and tables at the end of the beginning of the page so that the captions are separated from the main text.\n- Many equations like some parts of equation 1 and equation 2 might not be necessary as they don’t seem to contribute to understanding the paper. In many places, it seems like a simple intuitive explanation would be sufficient.\n- Similarly, Figure 3 might not be necessary.\n- The figures are unclear and hard to understand. For example, is the canonical viewpoint part of the input? If not, Figure 2 and Figure 4 could be changed to make it more clear.\n- Is this line correct, “If we add camera conditioning via FILM (that is, appending the camera embedding to the GRU’s embedding) then we achieve a much greater accuracy of 69.60 ± 0.09.” Should its value be 83.68 ± 1.21 as indicated in the table?\n \n#### Overall Recommendation:\nAlthough the problem could potentially be useful, the current dataset seems to be not so useful and over-simplified. Moreover, I found the paper not well-organized and hard to understand even after multiple reads. I feel the paper can be improved a lot and hence recommend rejection for the current version.\n\n#### Post Rebuttal\n(Copying from the discussion below)\n\nI would like to thank the author(s) for their response. After going over them, I am still not very confident about the paper would stick to my initial assessment. Following are my primary concerns:\n\n\"We note that there is a distinction between wanting to see something from another point of view, versus wanting to answer a question from another point of view. The former is where re-rendering is appropriate, but we do not make the claim that this alternative (view rendering + VQA) performs better or worse empirically.\"\n\nI understand the distinction. But the issue still remains. Why is the out-of-the-box \"view rendering + VQA\" solution insufficient? Is there any empirical justification for it? If not its hard to see the value in the current setup. A potential way to address this could be to run a simple out-of-the-box \"view rendering + VQA\" baseline.\n\n\"(2) R3 and R4’s concern about camera information being provided to the model and its potential infeasibility in practice: In real world settings, camera rigs can and do have knowledge about where they are situated in the world, for instance using SLAM or GPS coordinates. In that case, it is not unreasonable for e.g. an autonomous vehicle to answer queries by performing rotations and/or translations of its current viewpoint.\"\n\nThe concern was not about the viewpoint of the observer but the new viewpoint from which the question has to be answered. Also, the location of the new viewpoint need not be converted into float and appended to the question. It could be expressed in natural language. For example \"viewpoint of the driver in the other car\" like in the example provided by the paper. In the current setup the information about this viewpoint is provided in terms of exact coordinates, which makes the setup less interesting and not so practical.\n\nAlthough the authors improved some of the figures, the latest version of the paper does not seem to address other clarity concerns like a clear section for the dataset; organization of text and figures; removing unnecessary equations", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Review", "review": "The paper explores the problem of visual question answering from another perspective. Similar to VQA, a system is provided with a scene and a question. However, the difference is that the question needs to be answered from a viewpoint different from the one provided. Hence, the system needs to perform “mental rotation”. The paper creates a new dataset called CLEVR Mental Rotation Tests which is based on the prior CLEVR dataset. The paper also studies the efficacy of various supervised and self-supervised models on the proposed dataset.\n\n#### Strong points:\n- The problem of asking questions related to “mental rotation” seems interesting.\n- The paper shows that contrastive pre-training could be useful for the task, which is an interesting result.\n\n#### Weak points: \n- Although the problem seems interesting, I am unclear about the usefulness of the proposed dataset. The paper says that “many computer vision systems could benefit from neural architectures that demonstrate good performance for more targeted mental rotation tasks.” To justify this claim it gives the following example, “given the camera viewpoint of a (blind) person crossing the road, can we infer if each of the drivers of the cars at an intersection can see this blind person crossing the street?”. This could potentially be a useful scenario, however, the dataset proposed is different from the example as the camera viewpoint is provided as part of the input and not inferred from the question. The paper does not provide justification or evidence of how the current setup (i.e. with camera viewpoint) is useful. In particular, it would be nice if the paper could further explain how the current setup is better than solving the “view rendering” and VQA problems separately.\n\n- The dataset seems to be too simple for the mental rotation tests. It is unclear if in the future the dataset would be useful in distinguishing which models are better. As the paper shows that “2D baseline without camera conditioning” already achieves 70% accuracy. As far as I could understand, even without knowing which view to look at, a model could achieve 70% accuracy indicating that there is a lot of bias in the dataset. Moreover, simply adding camera embedding with the question to a 2D baseline (Table 1, 2D FILM with camera), already performs close to the best 3D model and upper bound. (Please clarify if my understanding is wrong.)\n\n- The paper is poorly organized and hard to follow. For example, one of the contributions of the work is the CLEVR-MTR dataset, however, there is no clear section in the main paper describing the details of how the dataset. Instead, the information about the dataset is scattered in the introduction and related work. Another example is that the paper moves into talking about the method (Section 2) without defining the task concretely. It is only from the figure that one notices that the camera viewpoint is part of the input. From the examples provided in the introduction, the reader is under the impression that the camera viewpoint has to be inferred from the question itself. Similarly, it's hard to parse what the training signal for each baseline is. Does a baseline use the rendered image from the other view during training?\n \n#### Minor Comments:\n- The figures and tables are interspersed with the text making the paper harder to read. It might be better to place the figures and tables at the end of the beginning of the page so that the captions are separated from the main text.\n- Many equations like some parts of equation 1 and equation 2 might not be necessary as they don’t seem to contribute to understanding the paper. In many places, it seems like a simple intuitive explanation would be sufficient.\n- Similarly, Figure 3 might not be necessary.\n- The figures are unclear and hard to understand. For example, is the canonical viewpoint part of the input? If not, Figure 2 and Figure 4 could be changed to make it more clear.\n- Is this line correct, “If we add camera conditioning via FILM (that is, appending the camera embedding to the GRU’s embedding) then we achieve a much greater accuracy of 69.60 ± 0.09.” Should its value be 83.68 ± 1.21 as indicated in the table?\n \n#### Overall Recommendation:\nAlthough the problem could potentially be useful, the current dataset seems to be not so useful and over-simplified. Moreover, I found the paper not well-organized and hard to understand even after multiple reads. I feel the paper can be improved a lot and hence recommend rejection for the current version.\n\n#### Post Rebuttal\n(Copying from the discussion below)\n\nI would like to thank the author(s) for their response. After going over them, I am still not very confident about the paper would stick to my initial assessment. Following are my primary concerns:\n\n\"We note that there is a distinction between wanting to see something from another point of view, versus wanting to answer a question from another point of view. The former is where re-rendering is appropriate, but we do not make the claim that this alternative (view rendering + VQA) performs better or worse empirically.\"\n\nI understand the distinction. But the issue still remains. Why is the out-of-the-box \"view rendering + VQA\" solution insufficient? Is there any empirical justification for it? If not its hard to see the value in the current setup. A potential way to address this could be to run a simple out-of-the-box \"view rendering + VQA\" baseline.\n\n\"(2) R3 and R4’s concern about camera information being provided to the model and its potential infeasibility in practice: In real world settings, camera rigs can and do have knowledge about where they are situated in the world, for instance using SLAM or GPS coordinates. In that case, it is not unreasonable for e.g. an autonomous vehicle to answer queries by performing rotations and/or translations of its current viewpoint.\"\n\nThe concern was not about the viewpoint of the observer but the new viewpoint from which the question has to be answered. Also, the location of the new viewpoint need not be converted into float and appended to the question. It could be expressed in natural language. For example \"viewpoint of the driver in the other car\" like in the example provided by the paper. In the current setup the information about this viewpoint is provided in terms of exact coordinates, which makes the setup less interesting and not so practical.\n\nAlthough the authors improved some of the figures, the latest version of the paper does not seem to address other clarity concerns like a clear section for the dataset; organization of text and figures; removing unnecessary equations", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603899031072}, {"id": "dIecMfK5HH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3392/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe paper studies visual question answering focusing on answering questions in a reference image of a different viewpoint. They propose a new dataset CLEVR-MRT drawing motivation from the well-known visual reasoning dataset CLEVR to illustrate the idea in which they have full control of the changes of viewpoints in an image. They then propose to use a volumetric encoder to represent 3D image features of an image via either 2D-to-3D projection or a contrastive-based encoder and further adapt an existing method (FiLM) to handle 3D tensors. Experiments on the CLEVR-MRT show that the use of the 2D features and 3D features of an image is complementary to each other.\n\nComments (Technical, Major Flaws of this paper): \n\n(1) The idea of addressing VQA in multi-view settings is reasonable but it is not entirely new. My main concern is at the limitations of a synthetic dataset in a controlled setting where the relations between objects are limited compared to real data. In addition, I believe that given enough such generated question-answer pairs with associated programs, models may possibly learn to decode the generation procedure under the hood instead of learning the actual semantic meanings of languages and the relations between objects.\n\n(2) Since there are no statistics about the newly introduced dataset, it is hard to judge the empirical results in the paper. As pointed out by many previous studies (e.g. Hudson, D.A., et al., 2019; Le, T.M., et al., 2020), models' performance seems to converge on CLEVR given enough training data. Having that said, existing methods easily fail if we reduce the number of training instances. As for the CLEVR-MRT, even without any information about the viewpoints, the baseline models could achieve more than 70% accuracy on the proposed dataset. It seems that the dataset is too simple that the model could have good performance without knowing the camera parameters. This leads to concerns about the validity of the proposed dataset. Please address these points.\nReferences:\n - Le, T. M., Le, V., Venkatesh, S., & Tran, T. (2020). Dynamic Language Binding in Relational Visual Reasoning. In IJCAI 2020.\n - Hudson, D. A., & Manning, C. D. (2018). Compositional attention networks for machine reasoning. In ICLR 2019.\n\n(3) For those who are not familiar with the CLEVR dataset, briefly explaining the procedure to generate the dataset and its variants might be helpful. \n\n(4) Given a question related to the object positions, there may exist many different views that provide the same answer. Let's take the question \"How many green spheres to the left of the shiny gold thing?\" in Figure 4 as an example. There are many views in the scene that provide the correct answer \"1\" for this question.   Without restricting the variance of the camera view (as in [1]), how can we ensure the model to infer the correct viewpoint?\n\nSome typos:\n- (1): 1. Introduction: We use the the Compositional -> We use the Compositional\n- (2): 2.1 FILM Baseline: the viewpoint and canonical view is the same thing -> the viewpoint and the canonical view are the same thing\n- (3): Figure 2: The dotted border on the ResNet-101 indicate -> The dotted border on the ResNet-101 indicates\n- (4): Conclusion: In the case of an autonomous vehicles -> In the case of autonomous vehicles", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The idea is not entirely new and details are missing ", "review": "Summary:\n\nThe paper studies visual question answering focusing on answering questions in a reference image of a different viewpoint. They propose a new dataset CLEVR-MRT drawing motivation from the well-known visual reasoning dataset CLEVR to illustrate the idea in which they have full control of the changes of viewpoints in an image. They then propose to use a volumetric encoder to represent 3D image features of an image via either 2D-to-3D projection or a contrastive-based encoder and further adapt an existing method (FiLM) to handle 3D tensors. Experiments on the CLEVR-MRT show that the use of the 2D features and 3D features of an image is complementary to each other.\n\nComments (Technical, Major Flaws of this paper): \n\n(1) The idea of addressing VQA in multi-view settings is reasonable but it is not entirely new. My main concern is at the limitations of a synthetic dataset in a controlled setting where the relations between objects are limited compared to real data. In addition, I believe that given enough such generated question-answer pairs with associated programs, models may possibly learn to decode the generation procedure under the hood instead of learning the actual semantic meanings of languages and the relations between objects.\n\n(2) Since there are no statistics about the newly introduced dataset, it is hard to judge the empirical results in the paper. As pointed out by many previous studies (e.g. Hudson, D.A., et al., 2019; Le, T.M., et al., 2020), models' performance seems to converge on CLEVR given enough training data. Having that said, existing methods easily fail if we reduce the number of training instances. As for the CLEVR-MRT, even without any information about the viewpoints, the baseline models could achieve more than 70% accuracy on the proposed dataset. It seems that the dataset is too simple that the model could have good performance without knowing the camera parameters. This leads to concerns about the validity of the proposed dataset. Please address these points.\nReferences:\n - Le, T. M., Le, V., Venkatesh, S., & Tran, T. (2020). Dynamic Language Binding in Relational Visual Reasoning. In IJCAI 2020.\n - Hudson, D. A., & Manning, C. D. (2018). Compositional attention networks for machine reasoning. In ICLR 2019.\n\n(3) For those who are not familiar with the CLEVR dataset, briefly explaining the procedure to generate the dataset and its variants might be helpful. \n\n(4) Given a question related to the object positions, there may exist many different views that provide the same answer. Let's take the question \"How many green spheres to the left of the shiny gold thing?\" in Figure 4 as an example. There are many views in the scene that provide the correct answer \"1\" for this question.   Without restricting the variance of the camera view (as in [1]), how can we ensure the model to infer the correct viewpoint?\n\nSome typos:\n- (1): 1. Introduction: We use the the Compositional -> We use the Compositional\n- (2): 2.1 FILM Baseline: the viewpoint and canonical view is the same thing -> the viewpoint and the canonical view are the same thing\n- (3): Figure 2: The dotted border on the ResNet-101 indicate -> The dotted border on the ResNet-101 indicates\n- (4): Conclusion: In the case of an autonomous vehicles -> In the case of autonomous vehicles", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603847211673}, {"id": "HpiDQBdAgH5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3392/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros: \n\n1. The paper presents an interesting idea to learn mental rotations using a variation of the CLEVR-VQA dataset. The contributions are - the creation of this synthetic CLEVR-Mental Rotation dataset for targeting this problem and a model that encodes questions and viewpoint information to produce answers via FiLM based encoders and 3D volume encoder. \n2. The results in Table 1 and Table 2 show improvements with respect to the baselines using their final model but there is still some concern in the improvements on their ablations. \n3. The paper is well written and easy to understand. \n\nCons: \n1. The motivation of why we need to learn mental rotations is not very clearly expressed, the practical examples given in the introduction are not sufficient. Does the model really learn these mental rotations from a simple spatial VQA task? This should be evaluated in the experiments either using activation maps or by visualizing intermediate 3D encodings. \n2. Is the model trained on all views for a single question-view pair or any one random viewpoint is sampled during mini batch training ? Does the rotation of a scene done over the complete 360 degree ? How do you decide how much to rotate to generate a viewpoint ?\n3. The self supervised learning of 3D volumes is an interesting idea, but it's use case in this particular problem is very weakly motivated both in experiments and theory. Why is this method better than the method discussed in Section 2.2.1? What is 3D data augmentation and how is it different from 2D data augmentation? \n4. There is a large variance in some experiments in Table 1. Is it due to the camera transformation embedding? It will be good to discuss the reasons why this is in Table 1 and not in Table 2.  \n5. Although the models developed are used in a very different problem setting with minor contributions, still a large part of the methods seem to be derived from the literature. \n6. The final results in Table 2 though argued are better due to small variance but more extensive experiments need to be performed to show the benefits of the self-supervised pre-training over the traditional encoder approach.\nMinor: What is the value of t (tau) used in Eq 3 ? In Table 2 it shows 1.0, but in the text it’s discussed as 0.1. Is this a typo or both of them are supposed to be different, if yes why ?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors propose to learn mental rotations via a synthetic CLEVR-Mental Rotation dataset based on VQA", "review": "Pros: \n\n1. The paper presents an interesting idea to learn mental rotations using a variation of the CLEVR-VQA dataset. The contributions are - the creation of this synthetic CLEVR-Mental Rotation dataset for targeting this problem and a model that encodes questions and viewpoint information to produce answers via FiLM based encoders and 3D volume encoder. \n2. The results in Table 1 and Table 2 show improvements with respect to the baselines using their final model but there is still some concern in the improvements on their ablations. \n3. The paper is well written and easy to understand. \n\nCons: \n1. The motivation of why we need to learn mental rotations is not very clearly expressed, the practical examples given in the introduction are not sufficient. Does the model really learn these mental rotations from a simple spatial VQA task? This should be evaluated in the experiments either using activation maps or by visualizing intermediate 3D encodings. \n2. Is the model trained on all views for a single question-view pair or any one random viewpoint is sampled during mini batch training ? Does the rotation of a scene done over the complete 360 degree ? How do you decide how much to rotate to generate a viewpoint ?\n3. The self supervised learning of 3D volumes is an interesting idea, but it's use case in this particular problem is very weakly motivated both in experiments and theory. Why is this method better than the method discussed in Section 2.2.1? What is 3D data augmentation and how is it different from 2D data augmentation? \n4. There is a large variance in some experiments in Table 1. Is it due to the camera transformation embedding? It will be good to discuss the reasons why this is in Table 1 and not in Table 2.  \n5. Although the models developed are used in a very different problem setting with minor contributions, still a large part of the methods seem to be derived from the literature. \n6. The final results in Table 2 though argued are better due to small variance but more extensive experiments need to be performed to show the benefits of the self-supervised pre-training over the traditional encoder approach.\nMinor: What is the value of t (tau) used in Eq 3 ? In Table 2 it shows 1.0, but in the text it’s discussed as 0.1. Is this a typo or both of them are supposed to be different, if yes why ?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603808933164}], "openreview_url": "https://openreview.net/forum?id=aYbCpFNnHdh", "arxiv_id": "2212.01639", "paper_pdf": "papers/aYbCpFNnHdh.pdf", "paper_pdf_sha256": "854dad77542a35aba62feb94b5f0507e981ba37c198c618690f1dcb583188815", "paper_pdf_bytes": 3838909, "paper_pdf_source": "openreview", "code_url": "https://github.com/christopher-beckham/clevr-mrt", "code_repository": "christopher-beckham/clevr-mrt", "code_commit": "3eb04b7fb8348b0620ae569bc0f99dc400c8d51a", "code_archive": "repos/aYbCpFNnHdh.zip", "code_archive_sha256": "368b2e9cac3cbce77014a34c0417117f3b4caf5954d2ca003dcb4e0af0347e81", "code_archive_bytes": 197126, "code_file_count": 79, "code_extensions": {".py": 76, ".sh": 3}, "github_disk_usage_kb": 596, "github_languages": {"Python": 476242, "Shell": 905, "Dockerfile": 698}, "github_archived": false, "github_pushed_at": "2023-04-11T16:30:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/visual-question-answering-from-another-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SylpBgrKPH", "year": 2020, "status": "rejected", "title": "MissDeepCausal: causal inference from incomplete data using deep latent variable models", "authors": ["Julie Josse", "Imke Mayer", "Jean-Philippe Vert"], "authorids": ["julie.josse@polytechnique.edu", "imke.mayer@polytechnique.edu", "jpvert@google.com"], "authors_source": "OpenReview API", "abstract": "Inferring causal effects of a treatment, intervention or policy from observational data is central to many applications. However, state-of-the-art methods for causal inference seldom consider the possibility that covariates have missing values, which is ubiquitous in many real-world analyses.  Missing data greatly complicate causal inference procedures as they require an adapted unconfoundedness hypothesis which can be difficult to justify in practice. We circumvent this issue by considering latent confounders whose distribution is learned through variational autoencoders adapted to missing values. They can be used either as a pre-processing step prior to causal inference but we also suggest to embed them in a multiple imputation strategy to take into account the variability due to missing values.  Numerical experiments demonstrate the effectiveness of the proposed methodology especially for non-linear models compared to competitors.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HkeSD_qeoB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2302/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n       The paper considers average treatment effect estimation treatment T and an unobserved confounder Z causes the outcome Y with an added constraint that the observed X is a noisy measurement of the underlying Z and some of the entries of observed X are missing at random (MAR). Previous work (Kallus et al. 2018) on similar settings assumed a low rank model connecting Z and X along with some entries missing at random which we do not observe. Further, Y (outcome) is related to the treatment and Z with a linear model. They actually show that matrix factorization techniques with these assumptions form an unbiased estimator for the Average Treatment Effect. There is prior work on doubly robust estimators under ignorability assumptions.\n\nIn this paper, the authors want to consider general non-linear relationships between Z and X with the same MAR (missing at random assumption) for missing entries.  So they fit a latent variable model in the form of VAE to find the P(Z| observed part of X) using a slightly modified version of the ELBO Lower Bound. For missing entries, they just replace those entries by a constant and do the usual VAE fit. After the VAE fit, multiple Z's are samples from the optimized P(Z| observed X) and then used in the doubly robust formula on each Z for estimating the average treatment effect and then finally the estimates averaged over the different Z's.\n\nThere is an alternative where the conditional mean of the latent variable is estimated from the VAE and used in the doubly robust computation.\n\nIn many synthetic examples, authors compare this with existing baselines and show that their method improves.\n\nPros: \n - Baselines compares are comprehensive enough from my perspective. \n - The paper is well written with clear pointer to existing work on doubly robust estimators with standard ignorability assumptions and the work for the linear, low rank model case by Kallus et al. 2018.\n\nCons:\n   Major Issues\n     - There is no reason to believe that even for the synthetic experiments, that the VAE posteriors would asymptotically yield unbiased ATE's which is provably the case in (Kallus et al. 2018) (of course for their restricted linear model/low rank assumptions). There is no reason to suppose Z's used in eq (9) from the VAE satisfy ignorability even in the asymptotic limit. In this light, the paper in essence just estimates Z's from some latent model that is fit and then use those latents to regress Y and then computes ATE. This seems a natural heuristic to try given such a problem. So I don't find a big methodological novelty. I would be willing to increase my scores if the authors could convince me on this point.\n\n  - For the LRMF model (Fig 4) the MF approach seems to do as well as the authors proposal and we have a guarantee for the MF case under those linear/low rank modeling assumptions. So the only demonstrated benefit is for the synthetic experiments for Fig 5 and Fig 3 (I agree that it is considerable particularly with large fraction of missing values in Fig 3) in whose settings we dont know about how unbiased it is in the limit. More synthetic experiments with different kind of generative models could be more convincing.\n\n- Some real world data set would have been more convincing - although I agree ground truth is hard to come by.\n\nMinor Issues:\n   - You have set B=200 for all the experiments for MDC-MI. Do the results change when B is increased or decreased ?? Does variance go down or the bias itself changes with B -  This would be a useful insight to have.\n\n  -  There is typo in the ELBO lower bound equation in page 11. There are other minor typos. Please correct for it.\n\n - Since the paper is about estimate treatment effect from measurements of an unobserved confounder - it is important to cite - https://ftp.cs.ucla.edu/pub/stat_ser/r366-reprint.pdf from the causal DAG literature. \n\n - The covariance of X given Z for the DLVM model is not clear - It seems to say exp ( or some matrix vector products) * Identity. What does this mean ? \n\n- The feature dimensions seems to be set at 10 - so would we expect the same results in much higher dimensions - like say 100s for few tens of thousands of samples??\n\n********UPDATE after reading the rebutall,changes and the new experiment**********\n\nI appreciate the authors actually accepting that identifiability issues cannot be easily resolved even if one knows P*(Z|X).\nI recommend the authors to elaborate on this point in the camera ready version. However, showing that the proposed methods work on a real benchmark (semi-synthetic one used in Shalit et. al 2017) is commendable.  However, I find that the MF method is competitive (almost all the time) with their method when Doubly robust estimators are used.\n\nBut having matched an existing baseline (the MF method) that deals with confounders on real data and showing superior synthetic results and authors clarifying and toning down their theoretical claims, I am inclined to increase the score to Weak accept.\n\n\n\n \n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "Summary:\n       The paper considers average treatment effect estimation treatment T and an unobserved confounder Z causes the outcome Y with an added constraint that the observed X is a noisy measurement of the underlying Z and some of the entries of observed X are missing at random (MAR). Previous work (Kallus et al. 2018) on similar settings assumed a low rank model connecting Z and X along with some entries missing at random which we do not observe. Further, Y (outcome) is related to the treatment and Z with a linear model. They actually show that matrix factorization techniques with these assumptions form an unbiased estimator for the Average Treatment Effect. There is prior work on doubly robust estimators under ignorability assumptions.\n\nIn this paper, the authors want to consider general non-linear relationships between Z and X with the same MAR (missing at random assumption) for missing entries.  So they fit a latent variable model in the form of VAE to find the P(Z| observed part of X) using a slightly modified version of the ELBO Lower Bound. For missing entries, they just replace those entries by a constant and do the usual VAE fit. After the VAE fit, multiple Z's are samples from the optimized P(Z| observed X) and then used in the doubly robust formula on each Z for estimating the average treatment effect and then finally the estimates averaged over the different Z's.\n\nThere is an alternative where the conditional mean of the latent variable is estimated from the VAE and used in the doubly robust computation.\n\nIn many synthetic examples, authors compare this with existing baselines and show that their method improves.\n\nPros: \n - Baselines compares are comprehensive enough from my perspective. \n - The paper is well written with clear pointer to existing work on doubly robust estimators with standard ignorability assumptions and the work for the linear, low rank model case by Kallus et al. 2018.\n\nCons:\n   Major Issues\n     - There is no reason to believe that even for the synthetic experiments, that the VAE posteriors would asymptotically yield unbiased ATE's which is provably the case in (Kallus et al. 2018) (of course for their restricted linear model/low rank assumptions). There is no reason to suppose Z's used in eq (9) from the VAE satisfy ignorability even in the asymptotic limit. In this light, the paper in essence just estimates Z's from some latent model that is fit and then use those latents to regress Y and then computes ATE. This seems a natural heuristic to try given such a problem. So I don't find a big methodological novelty. I would be willing to increase my scores if the authors could convince me on this point.\n\n  - For the LRMF model (Fig 4) the MF approach seems to do as well as the authors proposal and we have a guarantee for the MF case under those linear/low rank modeling assumptions. So the only demonstrated benefit is for the synthetic experiments for Fig 5 and Fig 3 (I agree that it is considerable particularly with large fraction of missing values in Fig 3) in whose settings we dont know about how unbiased it is in the limit. More synthetic experiments with different kind of generative models could be more convincing.\n\n- Some real world data set would have been more convincing - although I agree ground truth is hard to come by.\n\nMinor Issues:\n   - You have set B=200 for all the experiments for MDC-MI. Do the results change when B is increased or decreased ?? Does variance go down or the bias itself changes with B -  This would be a useful insight to have.\n\n  -  There is typo in the ELBO lower bound equation in page 11. There are other minor typos. Please correct for it.\n\n - Since the paper is about estimate treatment effect from measurements of an unobserved confounder - it is important to cite - https://ftp.cs.ucla.edu/pub/stat_ser/r366-reprint.pdf from the causal DAG literature. \n\n - The covariance of X given Z for the DLVM model is not clear - It seems to say exp ( or some matrix vector products) * Identity. What does this mean ? \n\n- The feature dimensions seems to be set at 10 - so would we expect the same results in much higher dimensions - like say 100s for few tens of thousands of samples??\n\n********UPDATE after reading the rebutall,changes and the new experiment**********\n\nI appreciate the authors actually accepting that identifiability issues cannot be easily resolved even if one knows P*(Z|X).\nI recommend the authors to elaborate on this point in the camera ready version. However, showing that the proposed methods work on a real benchmark (semi-synthetic one used in Shalit et. al 2017) is commendable.  However, I find that the MF method is competitive (almost all the time) with their method when Doubly robust estimators are used.\n\nBut having matched an existing baseline (the MF method) that deals with confounders on real data and showing superior synthetic results and authors clarifying and toning down their theoretical claims, I am inclined to increase the score to Weak accept.\n\n\n\n \n\n ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573066844748}, {"id": "BylnDUKyqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2302/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This contribution considers deep latent-factor models for causal inference -inferring the effect of a treatment- in the presence of missing values. The core challenge is that of confounders not directly observed, and accessible only via noisy proxys, in particular in the missingness. The contributed method relies on using the latent-factor model for multiple imputation in doubly-robust causal treatment effect estimators. As a consequence, it requires the missing at random assumption to control for the impact of imputation. Given that the confounders are not directly assumed, a first approach estimates their effect via an estimate of P(Z|X*)  (probability of confounder given observed data), which is then plugged in the doubly robust estimator in a multiple imputation strategy. A second approach uses heuristically the estimated latent confounders as regressors of non interest in a linear-regression model. The approaches are empirically compared to other imputation strategies used as plugins in the doubly-robust estimator. The contributed approach show marked benefits when the problem is highly non-linear.\n\nThe manuscript is clearly written. \n\nI do not have many comments.\n\nOne concern though is that the VAE comes with a significant amount of hyper-parameters that do not seem obvious to set. This is to be contrasted with other approaches compared to. How was the specific architecture and learning strategy of the VAE selected?\n\nThe simulation settings are somewhat artificial. More simulations inspired from real-life causal scenario would improve the work.\n\nI hope that in the final version, the code will be available publicly, and not on request.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This contribution considers deep latent-factor models for causal inference -inferring the effect of a treatment- in the presence of missing values. The core challenge is that of confounders not directly observed, and accessible only via noisy proxys, in particular in the missingness. The contributed method relies on using the latent-factor model for multiple imputation in doubly-robust causal treatment effect estimators. As a consequence, it requires the missing at random assumption to control for the impact of imputation. Given that the confounders are not directly assumed, a first approach estimates their effect via an estimate of P(Z|X*)  (probability of confounder given observed data), which is then plugged in the doubly robust estimator in a multiple imputation strategy. A second approach uses heuristically the estimated latent confounders as regressors of non interest in a linear-regression model. The approaches are empirically compared to other imputation strategies used as plugins in the doubly-robust estimator. The contributed approach show marked benefits when the problem is highly non-linear.\n\nThe manuscript is clearly written. \n\nI do not have many comments.\n\nOne concern though is that the VAE comes with a significant amount of hyper-parameters that do not seem obvious to set. This is to be contrasted with other approaches compared to. How was the specific architecture and learning strategy of the VAE selected?\n\nThe simulation settings are somewhat artificial. More simulations inspired from real-life causal scenario would improve the work.\n\nI hope that in the final version, the code will be available publicly, and not on request.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571948132058}, {"id": "SJxGMoVaYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2302/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces MissDeepCausal method to address the problem of treatment effect estimation with incomplete covariates matrix (missing values at random -- MAR). It makes use of Variational AutoEncoders (VAE) to learn the latent confounders from incomplete covariates. This also helps encoding complex non-linear relationships in the data, a capability that is missing in the work of Kallus et al. (2018) -- the work which this paper extends. They employ the Missing data Importance Weight AutoEncoder (MIWAE) approach (Mattei & Frellsen, 2019) to approximate the posterior of their latent factors Z given the observed incomplete covariates X*. The main contributions of this work are presented in sections 3.2 and 3.3, where they use the approximated posterior derived from MIWAE to sample Z to be used for estimating outcomes and finally calculating the Average Treatment Effect (ATE). This is done according to the doubly robust estimator developed for data with incomplete covariates (Mayer et al., 2019b). \n\nIn summary, I am not convinced that the contribution of this paper is enough, nor of its novelty. However, I will read the rebuttal carefully and am willing to increase the score if the authors address this concern.\n\nThere are several points that need further clarification; e.g., \n\t- Figure 1 as well as Figure 2 show a directed edge from X* to X_{miss}. Does this mean that X* has all the proxies needed to identify X_{miss}? \n\t- How does this method assure/evaluate that Z embeds enough information to predict accurate effects?\n\t- How are \\mu_0 and \\mu_1 functions trained on Z\n\nThings to improve the paper that did not impact the score:\n\t- Page 2, par. 2, last line: state-of-the-art method”s”\n\t- Page 3, under Unconfoundedness par., line -7: [...] for each observation “comma” treatment assignment [...]\n\t- Page 3, Figure 1: According to ICLR’s formatting guidelines, the figure number and caption must always appear after the figure.\n\t- Page 3, Missingness par., line 1: [...] is one “of” the most [...]\n\t- Page 5, line after Eq. (8): 8 should be in parentheses.\n\t- Page 7, Figure 3: box-plots are hardly legible.\n\t- Page 7, Figure 3 caption, line 2: keep “(logistic-)linear” together with \\mbox{} or ~ in latex\n\nReferences:\n\t- Kallus, N., Mao, X., & Udell, M. (2018). Causal inference with noisy and missing covariates via matrix factorization. In Advances in neural information processing systems (pp. 6921-6932).\n\t- Mattei, P. A., & Frellsen, J. (2019). MIWAE: Deep Generative Modelling and Imputation of Incomplete Data Sets. In International Conference on Machine Learning (pp. 4413-4423).\n\t- Mayer, I., Wager, S., Gauss, T., Moyer, J. D., & Josse, J. (2019). Doubly robust treatment effect estimation with missing attributes. preprint.\n\n\n********UPDATE after reading the rebuttal********\nThe authors have provided further clarifications in their rebuttal and therefore, I increased my score form “weak reject” to “weak accept”.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "This paper introduces MissDeepCausal method to address the problem of treatment effect estimation with incomplete covariates matrix (missing values at random -- MAR). It makes use of Variational AutoEncoders (VAE) to learn the latent confounders from incomplete covariates. This also helps encoding complex non-linear relationships in the data, a capability that is missing in the work of Kallus et al. (2018) -- the work which this paper extends. They employ the Missing data Importance Weight AutoEncoder (MIWAE) approach (Mattei & Frellsen, 2019) to approximate the posterior of their latent factors Z given the observed incomplete covariates X*. The main contributions of this work are presented in sections 3.2 and 3.3, where they use the approximated posterior derived from MIWAE to sample Z to be used for estimating outcomes and finally calculating the Average Treatment Effect (ATE). This is done according to the doubly robust estimator developed for data with incomplete covariates (Mayer et al., 2019b). \n\nIn summary, I am not convinced that the contribution of this paper is enough, nor of its novelty. However, I will read the rebuttal carefully and am willing to increase the score if the authors address this concern.\n\nThere are several points that need further clarification; e.g., \n\t- Figure 1 as well as Figure 2 show a directed edge from X* to X_{miss}. Does this mean that X* has all the proxies needed to identify X_{miss}? \n\t- How does this method assure/evaluate that Z embeds enough information to predict accurate effects?\n\t- How are \\mu_0 and \\mu_1 functions trained on Z\n\nThings to improve the paper that did not impact the score:\n\t- Page 2, par. 2, last line: state-of-the-art method”s”\n\t- Page 3, under Unconfoundedness par., line -7: [...] for each observation “comma” treatment assignment [...]\n\t- Page 3, Figure 1: According to ICLR’s formatting guidelines, the figure number and caption must always appear after the figure.\n\t- Page 3, Missingness par., line 1: [...] is one “of” the most [...]\n\t- Page 5, line after Eq. (8): 8 should be in parentheses.\n\t- Page 7, Figure 3: box-plots are hardly legible.\n\t- Page 7, Figure 3 caption, line 2: keep “(logistic-)linear” together with \\mbox{} or ~ in latex\n\nReferences:\n\t- Kallus, N., Mao, X., & Udell, M. (2018). Causal inference with noisy and missing covariates via matrix factorization. In Advances in neural information processing systems (pp. 6921-6932).\n\t- Mattei, P. A., & Frellsen, J. (2019). MIWAE: Deep Generative Modelling and Imputation of Incomplete Data Sets. In International Conference on Machine Learning (pp. 4413-4423).\n\t- Mayer, I., Wager, S., Gauss, T., Moyer, J. D., & Josse, J. (2019). Doubly robust treatment effect estimation with missing attributes. preprint.\n\n\n********UPDATE after reading the rebuttal********\nThe authors have provided further clarifications in their rebuttal and therefore, I increased my score form “weak reject” to “weak accept”.\n\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571797769658}], "openreview_url": "https://openreview.net/forum?id=SylpBgrKPH", "arxiv_id": "2002.10837", "paper_pdf": "papers/SylpBgrKPH.pdf", "paper_pdf_sha256": "c6ef826c8bd03fb13249cd1a147351c35c1ab098c10bd274128fc548a705ee6b", "paper_pdf_bytes": 504496, "paper_pdf_source": "openreview", "code_url": "https://github.com/imkemayer/MissDeepCausal", "code_repository": "imkemayer/MissDeepCausal", "code_commit": "6faecaa47234c457af54d492aa669a8f8b3b6b04", "code_archive": "repos/SylpBgrKPH.zip", "code_archive_sha256": "c2976250071e00561b1dd871d5a09e0aae1b322195325e6d99a57351ba6e5906", "code_archive_bytes": 1241821, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 1292, "github_languages": {"Python": 192281}, "github_archived": false, "github_pushed_at": "2022-12-08T07:30:46Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/missdeepcausal-causal-inference-from-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MYmHznX3aB", "year": 2026, "status": "rejected", "title": "Information Bargaining: Bilateral Commitment in Bayesian Persuasion", "authors": ["Yue Lin", "Shuhui Zhu", "William A Cunningham", "Wenhao Li", "Pascal Poupart", "Hongyuan Zha", "Baoxiang Wang"], "authorids": ["~Yue_Lin2", "~Shuhui_Zhu1", "~William_A_Cunningham1", "~Wenhao_Li2", "~Pascal_Poupart2", "~Hongyuan_Zha1", "~Baoxiang_Wang1"], "authors_source": "OpenReview API", "abstract": "Bayesian persuasion studies how an informed sender can influence a receiver’s actions through committed signaling schemes. While effective in one-shot settings, extending Bayesian persuasion to real-world long-term interactions becomes NP-hard.\nA separate empirical mismatch is also evident, where people tend to be much more truthful in practice than the signaling scheme predicted by Bayesian persuasion equilibria.\nTo address these issues, we first prove that long-term Bayesian persuasion can be decomposed into a bargaining stage and a realization stage while preserving optimality and equilibria. This decomposition disentangles the previously conflated informational and first-mover advantages of the sender. Based on the results, we establish a unified, realistic, and fairness-oriented framework called \\emph{information bargaining}. \nVariants of bargaining-game settings and their associated solution concepts, such as the Nash cooperative bargaining model and the Nash bargaining solution, can be applied directly to persuasion games to yield more realistic models.\nTo validate our framework, we identify and employ capable reasoning LLMs that can solve persuasion game equilibria effectively. In long-term persuasion task variants and in the corresponding bargaining game variants, these capable LLMs demonstrate that they reach the same equilibrium.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "pTuaixItlM", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3465/Reviewer_LB38"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces information bargaining, a two-stage view of long-term persuasion: a bargaining stage where sender and receiver commit to a signaling scheme and an action rule, followed by a realization stage where signals and actions are executed. The key claims are: the decomposition leaves optimality and equilibria unchanged; the sender’s power splits into informational advantage and first-mover advantage; and Bayesian persuasion can be reduced in polynomial time to a cooperative bargaining problem, enabling use of bargaining solution concepts such as the Nash bargaining solution. The paper formalizes a disagreement point via prior-based “babbling,” and proposes joint commitment as a fixed point of simultaneous policy updates.\n\nTo validate predictions, the authors use LLMs as equilibrium solvers. They first screen models on games with known solutions using Pearson correlation between model payoffs and ground truth, then run 87 bargaining and persuasion variants and check correlation with their hypotheses. Only two “reasoning” models pass the screen; reported correlations are high. Prompts force each agent to role-play a “self-interested, rational” player and to output JSON decisions. The authors state that these experiments are proof-of-concept and that the contribution is primarily theoretical.", "review_text": "The paper introduces information bargaining, a two-stage view of long-term persuasion: a bargaining stage where sender and receiver commit to a signaling scheme and an action rule, followed by a realization stage where signals and actions are executed. The key claims are: the decomposition leaves optimality and equilibria unchanged; the sender’s power splits into informational advantage and first-mover advantage; and Bayesian persuasion can be reduced in polynomial time to a cooperative bargaining problem, enabling use of bargaining solution concepts such as the Nash bargaining solution. The paper formalizes a disagreement point via prior-based “babbling,” and proposes joint commitment as a fixed point of simultaneous policy updates.\n\nTo validate predictions, the authors use LLMs as equilibrium solvers. They first screen models on games with known solutions using Pearson correlation between model payoffs and ground truth, then run 87 bargaining and persuasion variants and check correlation with their hypotheses. Only two “reasoning” models pass the screen; reported correlations are high. Prompts force each agent to role-play a “self-interested, rational” player and to output JSON decisions. The authors state that these experiments are proof-of-concept and that the contribution is primarily theoretical.", "strengths": "The theoretical framing is clear and potentially useful. The bargaining–realization split exposes the sender’s first-mover advantage and clarifies fairness considerations. The reduction from persuasion to bargaining gives a concrete path to bring cooperative bargaining solution concepts to information design. Definitions of disagreement points and joint commitment are explicit, with timing procedures that make the modeling choices inspectable.", "weaknesses": "The empirical evaluation relies on LLM rationality. The experiments assume that an LLM instantiated by a prompt is a rational best-response agent in the game-theoretic sense. This assumption is very challengable. The screening metric is payoff correlation, which can be high even if strategies are not equilibria, obedience fails, or off-path contingencies are wrong. Prompt structure, temperature, top-k, and other parameters including the random seed can change an LLM's output, which, in my view, renders the results it yields a little sketchy. The use of chain-of-thought is motivated as transparency, which is also rather debatable -- CoT has nothing to do with the internals of an LLM; it's just a multistaged generation process.\nAlso, there is no comparison to exact or approximate solvers on small games or to human subjects with real incentives.", "questions": "Re: the use of LLMs, can you report best-response regret for each agent under the learned signaling scheme and action rule? That is, compute the gain from a unilateral deviation. This would test rationality directly rather than via payoff correlation.\n\nCan you show some measure of sensitivity to prompt structure, temperature, seed, and so on, also (and in particular) reporting standard deviations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces information bargaining, a two-stage view of long-term persuasion: a bargaining stage where sender and receiver commit to a signaling scheme and an action rule, followed by a realization stage where signals and actions are executed. The key claims are: the decomposition leaves optimality and equilibria unchanged; the sender’s power splits into informational advantage and first-mover advantage; and Bayesian persuasion can be reduced in polynomial time to a cooperative bargaining problem, enabling use of bargaining solution concepts such as the Nash bargaining solution. The paper formalizes a disagreement point via prior-based “babbling,” and proposes joint commitment as a fixed point of simultaneous policy updates.\n\nTo validate predictions, the authors use LLMs as equilibrium solvers. They first screen models on games with known solutions using Pearson correlation between model payoffs and ground truth, then run 87 bargaining and persuasion variants and check correlation with their hypotheses. Only two “reasoning” models pass the screen; reported correlations are high. Prompts force each agent to role-play a “self-interested, rational” player and to output JSON decisions. The authors state that these experiments are proof-of-concept and that the contribution is primarily theoretical.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The theoretical framing is clear and potentially useful. The bargaining–realization split exposes the sender’s first-mover advantage and clarifies fairness considerations. The reduction from persuasion to bargaining gives a concrete path to bring cooperative bargaining solution concepts to information design. Definitions of disagreement points and joint commitment are explicit, with timing procedures that make the modeling choices inspectable.", "weaknesses": "The empirical evaluation relies on LLM rationality. The experiments assume that an LLM instantiated by a prompt is a rational best-response agent in the game-theoretic sense. This assumption is very challengable. The screening metric is payoff correlation, which can be high even if strategies are not equilibria, obedience fails, or off-path contingencies are wrong. Prompt structure, temperature, top-k, and other parameters including the random seed can change an LLM's output, which, in my view, renders the results it yields a little sketchy. The use of chain-of-thought is motivated as transparency, which is also rather debatable -- CoT has nothing to do with the internals of an LLM; it's just a multistaged generation process.\nAlso, there is no comparison to exact or approximate solvers on small games or to human subjects with real incentives.", "questions": "Re: the use of LLMs, can you report best-response regret for each agent under the learned signaling scheme and action rule? That is, compute the gain from a unilateral deviation. This would test rationality directly rather than via payoff correlation.\n\nCan you show some measure of sensitivity to prompt structure, temperature, seed, and so on, also (and in particular) reporting standard deviations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762208656966}, {"id": "WoS9NCsunX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3465/Reviewer_Yaqh"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "This paper introduces the bargaining perspective to Bayesian persuasion (BP).  The paper defines a BP game where, unlike the classical BP where only the sender has commitment power (commit to signaling schemes), now the receiver can also commit (to their mapping from signals to actions).  The paper seems to show that the BP problem with joint commitment can be reduced to a long-term Bayesian persuasion problem with a bargaining phase (subject to some criticism I mentioned in Weakness 1.2.2).  Then, the paper conducts experiments using LLM to play the bargaining game and BP game, validating the hypothesis that these two games have identical equilibrium outcomes.", "review_text": "This paper introduces the bargaining perspective to Bayesian persuasion (BP).  The paper defines a BP game where, unlike the classical BP where only the sender has commitment power (commit to signaling schemes), now the receiver can also commit (to their mapping from signals to actions).  The paper seems to show that the BP problem with joint commitment can be reduced to a long-term Bayesian persuasion problem with a bargaining phase (subject to some criticism I mentioned in Weakness 1.2.2).  Then, the paper conducts experiments using LLM to play the bargaining game and BP game, validating the hypothesis that these two games have identical equilibrium outcomes.", "strengths": "(S1) This paper introduces an interesting new perspective to the study of long-term Bayesian persuasion problem, namely, the bargaining perspective: the sender and receiver can bargain over the signaling scheme before implementing it.  \n\n(S2) LLM experiments demonstrate that the bargaining solution matches the classical Bayesian persuasion solution, meaning that the critical commitment assumption in classical BP can arise from bargaining.  This is an interesting observation.", "weaknesses": "(1) The paper's first contribution is claimed to be \"We show that long-term persuasion problems can be decomposed into a bargaining stage and a realization stage without affecting optimality or equilibria\".  I don't think this claim is supported well mathematically.\n\n(1.1) First, the \"long-term persuasion problem\" is never mathematically defined. Section 2 defines one-shot Bayesian persuasion and cheap talk games.  And Section 4 directly introduces a model of \"long-term persuasion with two stages (bargaining and realization)\".  I thought the authors wanted to show that \"long-term persuasion problem __(without bargaining)__ can be decomposed into a bargaining stage and a realization stage\". However, \"long-term persuasion problem\" (without bargaining) is never defined, in my understanding.  \n\n(1.2) Then, I tried to read Lemma 4.3 to understand what this claim truly means.  The lemma says that a Bayesian persuasion task is reducible to a bargaining game, where \"reducible\" means any solution concept in bargaining game can be turned into a solution concept in Bayesian persuasion game.  There are two issues with this lemma:\n\n  (1.2.1) The Bayesian persuasion game in this lemma is not the traditional Bayesian persuasion game where only the sender has commitment power; instead, it is a game with joint commitment where the receiver can also commit.  So, this lemma is proving \"BP with joint commitment can be reduced to bargaining game\", instead of \"classical BP can be reduced to bargaining game\".  BP with joint commitment is unconventional and lacks motivation: why does the receiver have commitment power?  I think the authors might want to do the reduction in the opposite direction: showing that the bargaining game can be reduced to BP with joint commitment.  Given that reduction, we can then use BP with joint commitment to study the bargaining game (which is a more natural problem than BP with joint commitment in my opinion). \n\n  (1.2.2) The bargaining game constructed in this reduction (in Appendix E) is trivial: it only contains the joint payoff profile $(R^i, R^j)$ that is induced by the sender's signaling scheme and the receiver's response in the solution of the Bayesian persuasion problem (with joint commitment).  Because the bargaining game only contains this profile, then clearly any solution concept of the bargaining game will pick this solution, making the argument trivial.  Usually, a bargaining game should contain a set $\\mathbb Y$ of multiple feasible agreements, and a solution concept picks an agreement from $\\mathbb Y$.  The bargaining game is meaningful only when $\\mathbb Y$ contains multiple agreements, but the bargaining game in this lemma only has a single agreement.  The correct argument should, in my opinion, consider the bargaining game where the feasible set of agreements consists of the payoff profiles associated with all feasible pairs of signaling scheme and receiver response, and argue that the bargaining solution concept picks the agreement that is equal to the BP solution.  \n\n\n\n(2) The paper's second contribution is \"We clarify two advantages that have been conflated in Bayesian persuasion, namely the sender's information advantage and the first-mover advantage\".  It's unclear to me how these two advantages are \"clarified\" mathematically.  Do you mean that, when one advantage is removed, the solution of Bayesian persuasion will change?  And the sender's payoff will decrease when the advantage is removed?  These two advantages are already known in previous literature. The long-term Bayesian persuasion problem considered in this paper seems to still allow the sender to have these two advantages.  I don't know if this paper provides any new observations about these two advantages.", "questions": "## Questions for the authors\n\nAs written in the Weaknesses, I am confused by many parts of this paper.  I would appreciate it if the authors could provide some clarifications. \n\n\n\n## Suggestions\n\nTypos: \n\n* Definition 2.3: $\\varphi(\\sigma=\\sigma | s)$ and $\\varphi(\\sigma|s)$ should be $\\varphi(\\sigma=a | s)$\n* Line 215: \"at uniformly random\" -> \"uniformly at random\"\n* Line 240: \"$\\sum_{s' | \\sigma} \\mu(s'|a) \\sum_a \\pi(a|s')$\" should be \"$\\sum_{s'} \\mu(s'|\\sigma) \\sum_a \\pi(a|\\sigma) = 1 \\cdot \\sum_a \\pi(a|\\sigma)$\", I think", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the bargaining perspective to Bayesian persuasion (BP).  The paper defines a BP game where, unlike the classical BP where only the sender has commitment power (commit to signaling schemes), now the receiver can also commit (to their mapping from signals to actions).  The paper seems to show that the BP problem with joint commitment can be reduced to a long-term Bayesian persuasion problem with a bargaining phase (subject to some criticism I mentioned in Weakness 1.2.2).  Then, the paper conducts experiments using LLM to play the bargaining game and BP game, validating the hypothesis that these two games have identical equilibrium outcomes.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "(S1) This paper introduces an interesting new perspective to the study of long-term Bayesian persuasion problem, namely, the bargaining perspective: the sender and receiver can bargain over the signaling scheme before implementing it.  \n\n(S2) LLM experiments demonstrate that the bargaining solution matches the classical Bayesian persuasion solution, meaning that the critical commitment assumption in classical BP can arise from bargaining.  This is an interesting observation.", "weaknesses": "(1) The paper's first contribution is claimed to be \"We show that long-term persuasion problems can be decomposed into a bargaining stage and a realization stage without affecting optimality or equilibria\".  I don't think this claim is supported well mathematically.\n\n(1.1) First, the \"long-term persuasion problem\" is never mathematically defined. Section 2 defines one-shot Bayesian persuasion and cheap talk games.  And Section 4 directly introduces a model of \"long-term persuasion with two stages (bargaining and realization)\".  I thought the authors wanted to show that \"long-term persuasion problem __(without bargaining)__ can be decomposed into a bargaining stage and a realization stage\". However, \"long-term persuasion problem\" (without bargaining) is never defined, in my understanding.  \n\n(1.2) Then, I tried to read Lemma 4.3 to understand what this claim truly means.  The lemma says that a Bayesian persuasion task is reducible to a bargaining game, where \"reducible\" means any solution concept in bargaining game can be turned into a solution concept in Bayesian persuasion game.  There are two issues with this lemma:\n\n  (1.2.1) The Bayesian persuasion game in this lemma is not the traditional Bayesian persuasion game where only the sender has commitment power; instead, it is a game with joint commitment where the receiver can also commit.  So, this lemma is proving \"BP with joint commitment can be reduced to bargaining game\", instead of \"classical BP can be reduced to bargaining game\".  BP with joint commitment is unconventional and lacks motivation: why does the receiver have commitment power?  I think the authors might want to do the reduction in the opposite direction: showing that the bargaining game can be reduced to BP with joint commitment.  Given that reduction, we can then use BP with joint commitment to study the bargaining game (which is a more natural problem than BP with joint commitment in my opinion). \n\n  (1.2.2) The bargaining game constructed in this reduction (in Appendix E) is trivial: it only contains the joint payoff profile $(R^i, R^j)$ that is induced by the sender's signaling scheme and the receiver's response in the solution of the Bayesian persuasion problem (with joint commitment).  Because the bargaining game only contains this profile, then clearly any solution concept of the bargaining game will pick this solution, making the argument trivial.  Usually, a bargaining game should contain a set $\\mathbb Y$ of multiple feasible agreements, and a solution concept picks an agreement from $\\mathbb Y$.  The bargaining game is meaningful only when $\\mathbb Y$ contains multiple agreements, but the bargaining game in this lemma only has a single agreement.  The correct argument should, in my opinion, consider the bargaining game where the feasible set of agreements consists of the payoff profiles associated with all feasible pairs of signaling scheme and receiver response, and argue that the bargaining solution concept picks the agreement that is equal to the BP solution.  \n\n\n\n(2) The paper's second contribution is \"We clarify two advantages that have been conflated in Bayesian persuasion, namely the sender's information advantage and the first-mover advantage\".  It's unclear to me how these two advantages are \"clarified\" mathematically.  Do you mean that, when one advantage is removed, the solution of Bayesian persuasion will change?  And the sender's payoff will decrease when the advantage is removed?  These two advantages are already known in previous literature. The long-term Bayesian persuasion problem considered in this paper seems to still allow the sender to have these two advantages.  I don't know if this paper provides any new observations about these two advantages.", "questions": "## Questions for the authors\n\nAs written in the Weaknesses, I am confused by many parts of this paper.  I would appreciate it if the authors could provide some clarifications. \n\n\n\n## Suggestions\n\nTypos: \n\n* Definition 2.3: $\\varphi(\\sigma=\\sigma | s)$ and $\\varphi(\\sigma|s)$ should be $\\varphi(\\sigma=a | s)$\n* Line 215: \"at uniformly random\" -> \"uniformly at random\"\n* Line 240: \"$\\sum_{s' | \\sigma} \\mu(s'|a) \\sum_a \\pi(a|s')$\" should be \"$\\sum_{s'} \\mu(s'|\\sigma) \\sum_a \\pi(a|\\sigma) = 1 \\cdot \\sum_a \\pi(a|\\sigma)$\", I think", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1762017437538}, {"id": "6dUh6LBoe3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3465/Reviewer_RRXQ"], "rating": 2, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper studies a variant of the Bayesian persuasion problem in which the receiver has perfect knowledge of the sender’s utility function and behaves strategically under different signaling schemes. The authors show that this variant can be reduced to a bargaining game. In particular, the receiver can use a threat to induce the sender to adopt a signaling scheme that is more favorable to the receiver. The authors also conduct experiments with LLMs.", "review_text": "This paper studies a variant of the Bayesian persuasion problem in which the receiver has perfect knowledge of the sender’s utility function and behaves strategically under different signaling schemes. The authors show that this variant can be reduced to a bargaining game. In particular, the receiver can use a threat to induce the sender to adopt a signaling scheme that is more favorable to the receiver. The authors also conduct experiments with LLMs.", "strengths": "The paper is clearly written and easy to follow. The variant of Bayesian persuasion studied in this paper is conceptually interesting. The idea that the receiver may strategically threaten the sender by refusing to follow the sender’s signaling scheme is natural and may capture a realistic strategic interaction.", "weaknesses": "The Bayesian persuasion variant proposed in this paper relies heavily on the assumption that the sender's utility function is known to the receiver. However, this assumption is only briefly mentioned (line 153) and lacks sufficient justification. The paper could benefit from a more detailed discussion of when and why this assumption might hold in practice.\n\nThe paper refers to the Long-Term Bayesian Persuasion problem, which is known to be NP-hard, and decomposes it into a bargaining stage and a realization stage. However, in the literature, the prior on states in Long-Term Bayesian persuasion typically evolves over rounds. In contrast, Procedure 2 in the paper assumes a fixed prior across rounds, which turns the problem into a repeated game, a much simpler setting. The difference between these two formulations is substantial. It would be valuable to discuss how the results might change if the prior were updated dynamically. \n\nA significant portion of the paper is devoted to reviewing Bayesian persuasion and bargaining games. While some background is necessary, the exposition could be shortened to make room for deeper analysis or additional results. The theoretical contribution currently centers on a single result and lacks theoretic depth. The experiments are also not very convincing as outputs of LLMs may vary depending on the prompts. The results are also not surprising. Simply proposing a new setting with a new solution concept and using LLMs for experiments do not guarantee acceptance to top-tier conferences like ICLR.", "questions": "Please see my comments above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies a variant of the Bayesian persuasion problem in which the receiver has perfect knowledge of the sender’s utility function and behaves strategically under different signaling schemes. The authors show that this variant can be reduced to a bargaining game. In particular, the receiver can use a threat to induce the sender to adopt a signaling scheme that is more favorable to the receiver. The authors also conduct experiments with LLMs.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "The paper is clearly written and easy to follow. The variant of Bayesian persuasion studied in this paper is conceptually interesting. The idea that the receiver may strategically threaten the sender by refusing to follow the sender’s signaling scheme is natural and may capture a realistic strategic interaction.", "weaknesses": "The Bayesian persuasion variant proposed in this paper relies heavily on the assumption that the sender's utility function is known to the receiver. However, this assumption is only briefly mentioned (line 153) and lacks sufficient justification. The paper could benefit from a more detailed discussion of when and why this assumption might hold in practice.\n\nThe paper refers to the Long-Term Bayesian Persuasion problem, which is known to be NP-hard, and decomposes it into a bargaining stage and a realization stage. However, in the literature, the prior on states in Long-Term Bayesian persuasion typically evolves over rounds. In contrast, Procedure 2 in the paper assumes a fixed prior across rounds, which turns the problem into a repeated game, a much simpler setting. The difference between these two formulations is substantial. It would be valuable to discuss how the results might change if the prior were updated dynamically. \n\nA significant portion of the paper is devoted to reviewing Bayesian persuasion and bargaining games. While some background is necessary, the exposition could be shortened to make room for deeper analysis or additional results. The theoretical contribution currently centers on a single result and lacks theoretic depth. The experiments are also not very convincing as outputs of LLMs may vary depending on the prompts. The results are also not surprising. Simply proposing a new setting with a new solution concept and using LLMs for experiments do not guarantee acceptance to top-tier conferences like ICLR.", "questions": "Please see my comments above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762003484427}, {"id": "wa9UHCo0VM", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission3465/Reviewer_ibGS"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper proposes a framework that modifies the Bayesian persuasion framework by adding elements of bargaining theory. The authors argue that some form of “long-term” Bayesian persuasion interactions can be decomposed into two stages: a bargaining stage and a realization stage (i.e. executing the strategies resulting from bargaining). The authors argue that this decomposition separates the sender’s “informational advantage” from their “first-mover advantage”.  Finally, some experiments using LLMs as players in the interaction are presented.", "review_text": "The paper proposes a framework that modifies the Bayesian persuasion framework by adding elements of bargaining theory. The authors argue that some form of “long-term” Bayesian persuasion interactions can be decomposed into two stages: a bargaining stage and a realization stage (i.e. executing the strategies resulting from bargaining). The authors argue that this decomposition separates the sender’s “informational advantage” from their “first-mover advantage”.  Finally, some experiments using LLMs as players in the interaction are presented.", "strengths": "The general idea of introducing a “bargaining phase” to model bilateral commitment in information exchange is interesting and could in principle broaden the scope of standard persuasion models.", "weaknesses": "The paper’s objectives and main contributions are largely unclear. I find neither the proposed model nor the empirical results convincing in their current form.\n\nStarting with the model. The paper augments the standard Bayesian persuasion framework with an additional bargaining phase. However, the motivation for this modification is not discussed. The discussion around lines 52 and 58–60 does not provide a solid justification, nor does it cite evidence for the “truthfulness discrepancy” it mentions. Since this is a substantial departure from the standard model, a more convincing argument is needed to explain why the extension is necessary. At present the modification may appear ad hoc, with the byproduct of having a setup for LLM-based experiments.\n\nThe proposed framework also introduces strong assumptions that are not discussed or justified. In particular, the assumption of complete mutual knowledge of preferences (e.g., the receiver knowing the sender’s utility function) is quite restrictive and not standard in Bayesian persuasion. The paper should discuss why this is reasonable, since most of the subsequent discussion is based on this assumption. \n\nIn terms of exposition, the model description is vague and not formal. The paper repeatedly refers to “long-term persuasion” but never clearly defines what this means, nor does it relate the setup to existing formulations of online Bayesian persuasion (in case “long-term” refers to this). Similarly, the repeated claim that the proposed decomposition “does not change optimality or equilibrium” is stated without formal proof or explanation. \n\nThe theoretical insights are quite weak in their current form. Section 3 mainly states straightforward implications of giving both players full information about expected payoffs. Lemma 4.3 is a simple observation once the assumptions are granted and does not represent a substantive theoretical contribution. \n\nThe empirical validation adds little to the argument. I don’t think the current evaluation provides meaningful evidence for the claims (or maybe it does, but it’s not clear from the current exposition). It remains unclear what can be learned about persuasion or bargaining from these experiments, and why LLM-based simulation should matter for the argument being made.\n\nClarity\n\nSection 2.2 is very vague and does not clarify the notion of bargaining game unless the reader is already familiar with it. It would be good to give an example here. \n\nSeveral concepts in the paper are not defined (e.g., visibility sets at line 150, extensive form games at line 214, babbling equilibrium at line 271, shadow of the future)\n\nExample at line 273 is not clear.", "questions": "none", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a framework that modifies the Bayesian persuasion framework by adding elements of bargaining theory. The authors argue that some form of “long-term” Bayesian persuasion interactions can be decomposed into two stages: a bargaining stage and a realization stage (i.e. executing the strategies resulting from bargaining). The authors argue that this decomposition separates the sender’s “informational advantage” from their “first-mover advantage”.  Finally, some experiments using LLMs as players in the interaction are presented.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "The general idea of introducing a “bargaining phase” to model bilateral commitment in information exchange is interesting and could in principle broaden the scope of standard persuasion models.", "weaknesses": "The paper’s objectives and main contributions are largely unclear. I find neither the proposed model nor the empirical results convincing in their current form.\n\nStarting with the model. The paper augments the standard Bayesian persuasion framework with an additional bargaining phase. However, the motivation for this modification is not discussed. The discussion around lines 52 and 58–60 does not provide a solid justification, nor does it cite evidence for the “truthfulness discrepancy” it mentions. Since this is a substantial departure from the standard model, a more convincing argument is needed to explain why the extension is necessary. At present the modification may appear ad hoc, with the byproduct of having a setup for LLM-based experiments.\n\nThe proposed framework also introduces strong assumptions that are not discussed or justified. In particular, the assumption of complete mutual knowledge of preferences (e.g., the receiver knowing the sender’s utility function) is quite restrictive and not standard in Bayesian persuasion. The paper should discuss why this is reasonable, since most of the subsequent discussion is based on this assumption. \n\nIn terms of exposition, the model description is vague and not formal. The paper repeatedly refers to “long-term persuasion” but never clearly defines what this means, nor does it relate the setup to existing formulations of online Bayesian persuasion (in case “long-term” refers to this). Similarly, the repeated claim that the proposed decomposition “does not change optimality or equilibrium” is stated without formal proof or explanation. \n\nThe theoretical insights are quite weak in their current form. Section 3 mainly states straightforward implications of giving both players full information about expected payoffs. Lemma 4.3 is a simple observation once the assumptions are granted and does not represent a substantive theoretical contribution. \n\nThe empirical validation adds little to the argument. I don’t think the current evaluation provides meaningful evidence for the claims (or maybe it does, but it’s not clear from the current exposition). It remains unclear what can be learned about persuasion or bargaining from these experiments, and why LLM-based simulation should matter for the argument being made.\n\nClarity\n\nSection 2.2 is very vague and does not clarify the notion of bargaining game unless the reader is already familiar with it. It would be good to give an example here. \n\nSeveral concepts in the paper are not defined (e.g., visibility sets at line 150, extensive form games at line 214, babbling equilibrium at line 271, shadow of the future)\n\nExample at line 273 is not clear.", "questions": "none", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761597819872}], "openreview_url": "https://openreview.net/forum?id=MYmHznX3aB", "arxiv_id": "2506.05876", "paper_pdf": "papers/MYmHznX3aB.pdf", "paper_pdf_sha256": "4eb9c8c62cf3a9f3b104a4ab9acc11030cc6e97afd2c1cd69fd06759ab94c6c9", "paper_pdf_bytes": 477451, "paper_pdf_source": "openreview", "code_url": "https://github.com/YueLin301/InformationBargaining", "code_repository": "YueLin301/InformationBargaining", "code_commit": "d5fa03c60d02d2e7c4572cc21dd8c1bff14f4be8", "code_archive": "repos/MYmHznX3aB.zip", "code_archive_sha256": "f2af77294af1f13a090b88433fcdd3146569d45027d7508a92dbd5170ac11b59", "code_archive_bytes": 4480690, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 2100, "github_languages": {"Python": 78794}, "github_archived": false, "github_pushed_at": "2025-06-06T07:37:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/information-bargaining-bilateral-commitment"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BdPbmgJ2jo", "year": 2025, "status": "rejected", "title": "High-dimensional Asymptotics of VAEs: Threshold of Posterior Collapse and Dataset-Size Dependence of Rate-Distortion Curve", "authors": ["Yuma Ichikawa", "Koji Hukushima"], "authorids": ["~Yuma_Ichikawa1", "~Koji_Hukushima1"], "authors_source": "OpenReview API", "abstract": "In variational autoencoders (VAEs), the variational posterior often aligns closely with the prior, known as posterior collapse, which leads to poor representation learning quality. An adjustable hyperparameter beta has been introduced in VAE to address this issue. This study sharply evaluates the conditions under which the posterior collapse occurs with respect to beta and dataset size by analyzing a minimal VAE in a high-dimensional limit. Additionally, this setting enables the evaluation of the rate-distortion curve in the VAE. This result shows that, unlike typical regularization parameters, VAEs face \"inevitable posterior collapse\" beyond a certain beta threshold, regardless of dataset size. The dataset-size dependence of the derived rate-distortion curve also suggests that relatively large datasets are required to achieve a rate-distortion curve with high rates. These results robustly explain generalization behavior across various real datasets with highly non-linear VAEs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "tnZSWD95Xi", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5872/Reviewer_NKDk"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper studies the RD curves in VAEs from a function of dataset size and dimensionality. The authors suggest that the RD curves as a function of data complexity $\\alpha$ (# data points / dim of data) and $$\\beta, can be divided to three categories of overfitting, learning, and underfitting; In high $\\alpha$ regime, smaller $\\beta$ is needed in order to avoid over-regularizing the model.", "review_text": "This paper studies the RD curves in VAEs from a function of dataset size and dimensionality. The authors suggest that the RD curves as a function of data complexity $\\alpha$ (# data points / dim of data) and $$\\beta, can be divided to three categories of overfitting, learning, and underfitting; In high $\\alpha$ regime, smaller $\\beta$ is needed in order to avoid over-regularizing the model.", "strengths": "- To the best of my knowledge, this is the first paper that studied RD curves in VAEs as a function of dataset size and data dimensions. This topic I think is a valuable topic of study and will indeed be of interest to the ICLR community. \n\n- The theory in the paper, to the best of my understanding, is sound. \n\n- The paper for the most reads well.", "weaknesses": "- There is no study of the network capacity in this work. While I understand that this is theoretical work, the authors do make a claim that the same results hold for more complex networks. However, there are prior works that suggest that RD curves for different network capacities behave differently [1,2]. Could the authors comment on this?\n\n- It is also not clear to me what is the message of the paper. It ofcourse makes sense that when you don't have a lotta data in high dimensions, you want to incorporate prior knowledge (such as regularization). Similarly,  when you have a lotta data, you don't need a lotta regularization as it is evident from all the recent DGMs. Furthermore, the $\\alpha < 1$ is hardly interesting as it is almost never the case. So practically, what does this mean for people employing VAEs? What is the core message here?\n\n- I do not think the experiments are strong enough to back the claims made in this paper. First, $\\alpha$ should have been studied as a function of both $n$ and $d$ separately. Here, $d$ was kept fixed. Furthermore, the data choice here is extremely specific. I understand the design choice, but some controlled experiments on real-world datasets are also necessary before showing Figure 3 left with those specific values. \n\n\n[1] Bozkurt, Alican, et al. \"Rate-regularization and generalization in VAEs.\" arXiv preprint arXiv:1911.04594 (2019).\n\n[2] Chérief-Abdellatif, Badr-Eddine, et al. \"On PAC-Bayesian reconstruction guarantees for VAEs.\" International conference on artificial intelligence and statistics. PMLR, 2022.\n\n\n\n\n\n**Minor comments**\n\n- \"Notations \" should not be place in Related work I would say\n- I would strongly advise to avoid using $D$ for the variances in $q$ and use $\\sigma^2$ instead as it is the most common symbol in the literature  for this.", "questions": "- Can you comment on how much the analysis is effected by the fact that $D$ is fixed?\n- Are the RD curves computed for the training set or test set? These two curves can be widely different.  \n- Do the authors mean overfitting by \"overlearning\"? If yes, I would say replace it with overfitting to avoid confusion :) \n- this is not clear to me but how did the authors at the values for Figure 3 Left? This is not seem to match other figures.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the RD curves in VAEs from a function of dataset size and dimensionality. The authors suggest that the RD curves as a function of data complexity $\\alpha$ (# data points / dim of data) and $$\\beta, can be divided to three categories of overfitting, learning, and underfitting; In high $\\alpha$ regime, smaller $\\beta$ is needed in order to avoid over-regularizing the model.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- To the best of my knowledge, this is the first paper that studied RD curves in VAEs as a function of dataset size and data dimensions. This topic I think is a valuable topic of study and will indeed be of interest to the ICLR community. \n\n- The theory in the paper, to the best of my understanding, is sound. \n\n- The paper for the most reads well.", "weaknesses": "- There is no study of the network capacity in this work. While I understand that this is theoretical work, the authors do make a claim that the same results hold for more complex networks. However, there are prior works that suggest that RD curves for different network capacities behave differently [1,2]. Could the authors comment on this?\n\n- It is also not clear to me what is the message of the paper. It ofcourse makes sense that when you don't have a lotta data in high dimensions, you want to incorporate prior knowledge (such as regularization). Similarly,  when you have a lotta data, you don't need a lotta regularization as it is evident from all the recent DGMs. Furthermore, the $\\alpha < 1$ is hardly interesting as it is almost never the case. So practically, what does this mean for people employing VAEs? What is the core message here?\n\n- I do not think the experiments are strong enough to back the claims made in this paper. First, $\\alpha$ should have been studied as a function of both $n$ and $d$ separately. Here, $d$ was kept fixed. Furthermore, the data choice here is extremely specific. I understand the design choice, but some controlled experiments on real-world datasets are also necessary before showing Figure 3 left with those specific values. \n\n\n[1] Bozkurt, Alican, et al. \"Rate-regularization and generalization in VAEs.\" arXiv preprint arXiv:1911.04594 (2019).\n\n[2] Chérief-Abdellatif, Badr-Eddine, et al. \"On PAC-Bayesian reconstruction guarantees for VAEs.\" International conference on artificial intelligence and statistics. PMLR, 2022.\n\n\n\n\n\n**Minor comments**\n\n- \"Notations \" should not be place in Related work I would say\n- I would strongly advise to avoid using $D$ for the variances in $q$ and use $\\sigma^2$ instead as it is the most common symbol in the literature  for this.", "questions": "- Can you comment on how much the analysis is effected by the fact that $D$ is fixed?\n- Are the RD curves computed for the training set or test set? These two curves can be widely different.  \n- Do the authors mean overfitting by \"overlearning\"? If yes, I would say replace it with overfitting to avoid confusion :) \n- this is not clear to me but how did the authors at the values for Figure 3 Left? This is not seem to match other figures.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730741053081}, {"id": "5eHWA0r8Uq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5872/Reviewer_LZGQ"], "rating": 8, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper investigates several aspects of beta-linear-VAEs, including posterior collapse, the effect of different values for the beta parameters, the effect of training dataset size.\nIt also introduced two summary statistics, and using these statistics, the authors derived a phase diagram for VAE learning in terms of the beta values and the relative scale of the training dataset size.", "review_text": "The paper investigates several aspects of beta-linear-VAEs, including posterior collapse, the effect of different values for the beta parameters, the effect of training dataset size.\nIt also introduced two summary statistics, and using these statistics, the authors derived a phase diagram for VAE learning in terms of the beta values and the relative scale of the training dataset size.", "strengths": "- The paper studied an important aspect of VAEs and how the different parameters and choices can affect the performance.\n- The empirical findings of the relation between generalisation error and the sample complexity as well as the beta parameter is interesting.", "weaknesses": "The paper discussed a list of different behaviours of VAEs, but it feels like they are rather loosely connected findings (i.e., the subsections in Section 6).\n\nThe findings themselves are interesting, but it is not surprising that changing one variable, such as beta or the number of training data, will lead to various changes in aspects like RD curves, posterior collapse. \n\nTherefore, I believe a more coherent story is important to connect the dots and make these findings more insightful.", "questions": "1. The signal recovery error feels like a definition of reconstruction error, and a distortion metric. What’s the difference between the signal recovery error and the distortion (D) of the RD curve in the paper?\n2. Typo: Page 8 line 397: “summary statics” -> “summary statistics”", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates several aspects of beta-linear-VAEs, including posterior collapse, the effect of different values for the beta parameters, the effect of training dataset size.\nIt also introduced two summary statistics, and using these statistics, the authors derived a phase diagram for VAE learning in terms of the beta values and the relative scale of the training dataset size.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper studied an important aspect of VAEs and how the different parameters and choices can affect the performance.\n- The empirical findings of the relation between generalisation error and the sample complexity as well as the beta parameter is interesting.", "weaknesses": "The paper discussed a list of different behaviours of VAEs, but it feels like they are rather loosely connected findings (i.e., the subsections in Section 6).\n\nThe findings themselves are interesting, but it is not surprising that changing one variable, such as beta or the number of training data, will lead to various changes in aspects like RD curves, posterior collapse. \n\nTherefore, I believe a more coherent story is important to connect the dots and make these findings more insightful.", "questions": "1. The signal recovery error feels like a definition of reconstruction error, and a distortion metric. What’s the difference between the signal recovery error and the distortion (D) of the RD curve in the paper?\n2. Typo: Page 8 line 397: “summary statics” -> “summary statistics”", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730678467198}, {"id": "T0eYwqOeG6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5872/Reviewer_Xf1j"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper aims to analyze the solution learnt by a $\\beta$-VAE w.r.t. (1) the parameter $\\beta$, and (2) the training dataset size. The authors work in the high-dimensional asymptotic setting, and aim to characterize certain phenomena about the quality of the learnt solution by the VAE.\n\nTo do this, the authors theoretically analyze a linear VAE model (Eq (5)), in the high-dimensional asymptotic regime where $n,d \\rightarrow \\infty$ and $\\frac{n}{d} = \\alpha$ (sample complexity) stays finite, using the replica method as a heuristic to get around intractable calculations. This is presented in Section 5.\n\nThe asymptotic formulae are empirically verified in Section 6.1 and 6.2, on a synthetic data model of the spiked covariance matrix (Eq (4)). This is then used to draw interesting observations about the learning process and the quality of the learnt VAE solution. Figure 2 in particular shows many of the findings.\n\nThe authors then empirically show some of the findings (from the linear VAE setting) hold true for non-linear VAEs also, trained on real-world datasets like MNIST and FashionMNIST. This is presented in Section 6.5.", "review_text": "This paper aims to analyze the solution learnt by a $\\beta$-VAE w.r.t. (1) the parameter $\\beta$, and (2) the training dataset size. The authors work in the high-dimensional asymptotic setting, and aim to characterize certain phenomena about the quality of the learnt solution by the VAE.\n\nTo do this, the authors theoretically analyze a linear VAE model (Eq (5)), in the high-dimensional asymptotic regime where $n,d \\rightarrow \\infty$ and $\\frac{n}{d} = \\alpha$ (sample complexity) stays finite, using the replica method as a heuristic to get around intractable calculations. This is presented in Section 5.\n\nThe asymptotic formulae are empirically verified in Section 6.1 and 6.2, on a synthetic data model of the spiked covariance matrix (Eq (4)). This is then used to draw interesting observations about the learning process and the quality of the learnt VAE solution. Figure 2 in particular shows many of the findings.\n\nThe authors then empirically show some of the findings (from the linear VAE setting) hold true for non-linear VAEs also, trained on real-world datasets like MNIST and FashionMNIST. This is presented in Section 6.5.", "strengths": "**Technical strengths**:\n- The paper sharply characterizes high-dimensional asymptotics for learning the linear VAE (Eq (5)) under the spiked covariance model (Eq (4)) with the regularized $\\beta$-VAE objective (Eq (6)).\n- This is used to show interesting observations about the VAE learning process in Section 6.1 and 6.2. In particular, (1) Figure 2 shows a double-descent phenomenon w.r.to the sample complexity $\\alpha$, with the reconstruction error (Eq (9)) peaking at $\\alpha = 1$, and (2) Figure 2 also shows a long plateau in the reconstruction error for large values of $\\beta$. This is backed by Claim 6.1 (in the large $\\alpha$ limit) and lines 428-430 provide concrete guidance to practitioners about the risks of a large $\\beta$ when training.\n- Section 6.5 (and Figure 5) shows this on real-world datasets MNIST and FashionMNIST also, where the insight can be used to practically choose the \"optimal\" value of $\\beta$ approximately equal to the noise ratio $\\hat{\\eta}$, which can be estimated using the training dataset.\n\n**Presentation strengths**:\n- The paper is largely well-written and easy to follow. The authors include relevant explanations in most places. For example, the choice of the spiked covariance model as the synthetic data generating process was backed by evidence in Figure 1 of MNIST following something similar.", "weaknesses": "**Technical Weaknesses**:\n- The main weakness is the fact that the theoretical results are not exact, since they have been developed using the replica method, which is a heuristic to get around intractable calculations.\n- The authors work in the simple setting of $k = k^\\star = 1$. If I understand correctly, this means the true latent space is $1$-dimensional. It would have been nice to see the synthetic experiments with $k^\\star$ varying, say in $[1, 2, 4]$. In particular, what would the trend of $\\varepsilon_g$ w.r.to $k^\\star$ look like?\n- Some of the claims can be better substantiated. For example, in the context of Figure 2, it would have been nice to see a plot of $\\varepsilon_g$ w.r.to $\\alpha$ for the optimal $\\beta$ choice. Is that perhaps monotonically decreasing? (This is similar to the double-descent observations in literature, where using the optimal parameter leads to a monotonically decreasing curve instead of double-descent).\n\n**Minor notes on the typos I found**\n- Line 146, $H$ is probably the \"negative log-likelihood\" instead of just \"likelihood\". I stress this because it is important whether we want to minimize or maximize $H$.\n- Line 189, \"Spectrum\" of the covariance matrix, instead of \"Spectral\". This typo is present in many places throughout the paper (for eg, Fig 1(b)), would appreciate if it can be cleaned up.\n- Line 214, \"Note that\" instead of \"Noted that\".\n- Line 224, $\\lambda \\in \\mathbb{R}_{+}$ maybe instead of  $\\lambda \\in \\mathbb{R}$? I *assume* practitioners use a non-negative regularization parameter.", "questions": "- What is the main challenge that the replica method allows you to get around? Would be nice to provide some insight into this. Or perhaps a toy example of the usage of replica method demonstrating what are its benefits and why is it used in this particular context.\n- What is the main reason to introduce the metric $\\varepsilon_g$ in Eq (9)? How is it different than the distortion $D$?\n- In Figure 3, what does \"overlearning\" mean? From the description in section 6.2, it seems it is the same as overfitting? If so, would be good to name it that way instead of introducing a new term.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to analyze the solution learnt by a $\\beta$-VAE w.r.t. (1) the parameter $\\beta$, and (2) the training dataset size. The authors work in the high-dimensional asymptotic setting, and aim to characterize certain phenomena about the quality of the learnt solution by the VAE.\n\nTo do this, the authors theoretically analyze a linear VAE model (Eq (5)), in the high-dimensional asymptotic regime where $n,d \\rightarrow \\infty$ and $\\frac{n}{d} = \\alpha$ (sample complexity) stays finite, using the replica method as a heuristic to get around intractable calculations. This is presented in Section 5.\n\nThe asymptotic formulae are empirically verified in Section 6.1 and 6.2, on a synthetic data model of the spiked covariance matrix (Eq (4)). This is then used to draw interesting observations about the learning process and the quality of the learnt VAE solution. Figure 2 in particular shows many of the findings.\n\nThe authors then empirically show some of the findings (from the linear VAE setting) hold true for non-linear VAEs also, trained on real-world datasets like MNIST and FashionMNIST. This is presented in Section 6.5.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**Technical strengths**:\n- The paper sharply characterizes high-dimensional asymptotics for learning the linear VAE (Eq (5)) under the spiked covariance model (Eq (4)) with the regularized $\\beta$-VAE objective (Eq (6)).\n- This is used to show interesting observations about the VAE learning process in Section 6.1 and 6.2. In particular, (1) Figure 2 shows a double-descent phenomenon w.r.to the sample complexity $\\alpha$, with the reconstruction error (Eq (9)) peaking at $\\alpha = 1$, and (2) Figure 2 also shows a long plateau in the reconstruction error for large values of $\\beta$. This is backed by Claim 6.1 (in the large $\\alpha$ limit) and lines 428-430 provide concrete guidance to practitioners about the risks of a large $\\beta$ when training.\n- Section 6.5 (and Figure 5) shows this on real-world datasets MNIST and FashionMNIST also, where the insight can be used to practically choose the \"optimal\" value of $\\beta$ approximately equal to the noise ratio $\\hat{\\eta}$, which can be estimated using the training dataset.\n\n**Presentation strengths**:\n- The paper is largely well-written and easy to follow. The authors include relevant explanations in most places. For example, the choice of the spiked covariance model as the synthetic data generating process was backed by evidence in Figure 1 of MNIST following something similar.", "weaknesses": "**Technical Weaknesses**:\n- The main weakness is the fact that the theoretical results are not exact, since they have been developed using the replica method, which is a heuristic to get around intractable calculations.\n- The authors work in the simple setting of $k = k^\\star = 1$. If I understand correctly, this means the true latent space is $1$-dimensional. It would have been nice to see the synthetic experiments with $k^\\star$ varying, say in $[1, 2, 4]$. In particular, what would the trend of $\\varepsilon_g$ w.r.to $k^\\star$ look like?\n- Some of the claims can be better substantiated. For example, in the context of Figure 2, it would have been nice to see a plot of $\\varepsilon_g$ w.r.to $\\alpha$ for the optimal $\\beta$ choice. Is that perhaps monotonically decreasing? (This is similar to the double-descent observations in literature, where using the optimal parameter leads to a monotonically decreasing curve instead of double-descent).\n\n**Minor notes on the typos I found**\n- Line 146, $H$ is probably the \"negative log-likelihood\" instead of just \"likelihood\". I stress this because it is important whether we want to minimize or maximize $H$.\n- Line 189, \"Spectrum\" of the covariance matrix, instead of \"Spectral\". This typo is present in many places throughout the paper (for eg, Fig 1(b)), would appreciate if it can be cleaned up.\n- Line 214, \"Note that\" instead of \"Noted that\".\n- Line 224, $\\lambda \\in \\mathbb{R}_{+}$ maybe instead of  $\\lambda \\in \\mathbb{R}$? I *assume* practitioners use a non-negative regularization parameter.", "questions": "- What is the main challenge that the replica method allows you to get around? Would be nice to provide some insight into this. Or perhaps a toy example of the usage of replica method demonstrating what are its benefits and why is it used in this particular context.\n- What is the main reason to introduce the metric $\\varepsilon_g$ in Eq (9)? How is it different than the distortion $D$?\n- In Figure 3, what does \"overlearning\" mean? From the description in section 6.2, it seems it is the same as overfitting? If so, would be good to name it that way instead of introducing a new term.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730514633408}, {"id": "8PDwJKt4Du", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5872/Reviewer_1vWD"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper explores the influence of the β hyperparameter in a β-VAE, particularly regarding its impact on the posterior collapse phenomenon and the Rate-Distortion (RD) curve, from an information-theoretic perspective. Using a Spiked Covariance Model (SCM), the authors identify three phases (regularized, learning, overlearning) based on β and sample complexity α = n/d . The results suggest that high values of β inevitably lead to posterior collapse, regardless of data volume, and propose that the optimal β correlates with noise\nstrength ˆη in high-complexity settings. Experiments on MNIST and Fashion-MNIST datasets partially validate these claims.", "review_text": "This paper explores the influence of the β hyperparameter in a β-VAE, particularly regarding its impact on the posterior collapse phenomenon and the Rate-Distortion (RD) curve, from an information-theoretic perspective. Using a Spiked Covariance Model (SCM), the authors identify three phases (regularized, learning, overlearning) based on β and sample complexity α = n/d . The results suggest that high values of β inevitably lead to posterior collapse, regardless of data volume, and propose that the optimal β correlates with noise\nstrength ˆη in high-complexity settings. Experiments on MNIST and Fashion-MNIST datasets partially validate these claims.", "strengths": "• Information-Theoretic Perspective: A thoughtful analysis of β-VAE’s Rate-Distortion properties.\n• Insight on Hyperparameter Tuning: Suggests an optimal β based on sample complexity, a novel angle.\n• Focus on an Underexplored Hyperparameter: Thorough analysis of β’s role in posterior collapse.", "weaknesses": "• Over-Simplified Model and Limited Scope: Most analysis centers on a linear VAE with low latent dimensions (e.g., dimension 1), which reduces the generalizability to real-world settings. Minimal insight into non-asymptotic or intermediate regimes.\n• Limited Novelty: The core claim that “high β is detrimental” is already known. While it validates an intuition, the work lacks a new method or paradigm. Moreover, it is unclear that the posterior collapse phenomenon occurs ”often” as stated in the abstract, especially when the latent dimension increases.\n• Minimal Experiments: Only FID scores on two datasets (MNIST, Fashion-MNIST), without broader validation on diverse or complex data,\nand lacks sensitivity analysis (for small changes on the value).\n• Identifiability of Ground Truth Model: As is, the ground truth model does not seem to be identifiable (it is when dividing by √θ). This hinders the reliability of the analysis.", "questions": "1. Could you explore the performance of this model in transitional or non-asymptotic regimes to reflect real-world data scenarios?\n2. Is your analysis still valid for datasets with complex spectrums?\n3. Could you clarify the derivation in Equation (9), especially the expectation notation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the influence of the β hyperparameter in a β-VAE, particularly regarding its impact on the posterior collapse phenomenon and the Rate-Distortion (RD) curve, from an information-theoretic perspective. Using a Spiked Covariance Model (SCM), the authors identify three phases (regularized, learning, overlearning) based on β and sample complexity α = n/d . The results suggest that high values of β inevitably lead to posterior collapse, regardless of data volume, and propose that the optimal β correlates with noise\nstrength ˆη in high-complexity settings. Experiments on MNIST and Fashion-MNIST datasets partially validate these claims.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "• Information-Theoretic Perspective: A thoughtful analysis of β-VAE’s Rate-Distortion properties.\n• Insight on Hyperparameter Tuning: Suggests an optimal β based on sample complexity, a novel angle.\n• Focus on an Underexplored Hyperparameter: Thorough analysis of β’s role in posterior collapse.", "weaknesses": "• Over-Simplified Model and Limited Scope: Most analysis centers on a linear VAE with low latent dimensions (e.g., dimension 1), which reduces the generalizability to real-world settings. Minimal insight into non-asymptotic or intermediate regimes.\n• Limited Novelty: The core claim that “high β is detrimental” is already known. While it validates an intuition, the work lacks a new method or paradigm. Moreover, it is unclear that the posterior collapse phenomenon occurs ”often” as stated in the abstract, especially when the latent dimension increases.\n• Minimal Experiments: Only FID scores on two datasets (MNIST, Fashion-MNIST), without broader validation on diverse or complex data,\nand lacks sensitivity analysis (for small changes on the value).\n• Identifiability of Ground Truth Model: As is, the ground truth model does not seem to be identifiable (it is when dividing by √θ). This hinders the reliability of the analysis.", "questions": "1. Could you explore the performance of this model in transitional or non-asymptotic regimes to reflect real-world data scenarios?\n2. Is your analysis still valid for datasets with complex spectrums?\n3. Could you clarify the derivation in Equation (9), especially the expectation notation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730460215510}], "openreview_url": "https://openreview.net/forum?id=BdPbmgJ2jo", "arxiv_id": "2309.07663", "paper_pdf": "papers/BdPbmgJ2jo.pdf", "paper_pdf_sha256": "e2274e498095f338820ba10d7ccbd5f6f0fd4aaeb6ba087658ab73ed254ae50d", "paper_pdf_bytes": 1124052, "paper_pdf_source": "openreview", "code_url": "https://github.com/Yuma-Ichikawa/vae-replica", "code_repository": "Yuma-Ichikawa/vae-replica", "code_commit": "9f6f430e2a6399e689984077921706fc02fa2a4e", "code_archive": "repos/BdPbmgJ2jo.zip", "code_archive_sha256": "a05463930303dec4c91ea8bc842b972884be4f990d7b9172235b0ecdee4ce12b", "code_archive_bytes": 59985, "code_file_count": 5, "code_extensions": {".py": 4, ".ipynb": 1}, "github_disk_usage_kb": 152, "github_languages": {"Jupyter Notebook": 211800, "Python": 8665}, "github_archived": false, "github_pushed_at": "2025-03-20T02:03:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dataset-size-dependence-of-rate-distortion"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "AfiM6F2YPY", "year": 2024, "status": "rejected", "title": "Applying language models to algebraic topology: generating simplicial cycles using multi-labeling in Wu's formula", "authors": ["Kirill Brilliantov", "Fedor Pavutnitskiy", "Dmitry Pasechnyuk", "German Magai"], "authorids": ["~Kirill_Brilliantov1", "~Fedor_Pavutnitskiy1", "~Dmitry_Pasechnyuk1", "~German_Magai1"], "authors_source": "OpenReview API", "abstract": "Computing homotopy groups of spheres has long been a fundamental objective in algebraic topology. Various theoretical and algorithmic approaches have been developed to tackle this problem. In this paper we take a step towards the goal of comprehending the group-theoretic structure of the generators of these homotopy groups by leveraging the power of machine learning. Specifically, in the simplicial group setting of Wu's formula, we reformulate the problem of generating simplicial cycles as a problem of sampling from the intersection of algorithmic datasets related to Dyck languages. We present and evaluate language modelling approaches that employ multi-label information for input sequences, along with the necessary group-theoretic toolkit and non-neural baselines.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "b20qA3DPdF", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2245/Reviewer_aKQP"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper aims at studying the simplicial homotopy groups of the 2-dimensional sphere. Following Wu's formula, each homotopy group is isomorphic to the quotient of a free group. Each element in the homotopy group, as a reduced word, can be connected to a sentence in the language models. The decoder part of a transformer is used to generate approximate samples from the homotopy group, and a multi-label is used to denote whether the sample satisfies relations of the homotopy groups. Multiple approaches for handling the multi-label and training the transformer are designed. Adequate numerical experiments are provided, with comparison to multiple baselines.", "review_text": "This paper aims at studying the simplicial homotopy groups of the 2-dimensional sphere. Following Wu's formula, each homotopy group is isomorphic to the quotient of a free group. Each element in the homotopy group, as a reduced word, can be connected to a sentence in the language models. The decoder part of a transformer is used to generate approximate samples from the homotopy group, and a multi-label is used to denote whether the sample satisfies relations of the homotopy groups. Multiple approaches for handling the multi-label and training the transformer are designed. Adequate numerical experiments are provided, with comparison to multiple baselines.", "strengths": "1. The problem originating from algebraic topology is important.\n2. The idea is very innovative.\n3. The paper is well-written and mostly clear.\n4. Adequate numerical experiments are provided.", "weaknesses": "1. It would be helpful to discuss a bit of how the samples drawn from the homotopy group can be used to understand the topological properties of the 2-dimensional sphere, compared to the more obtainable homology groups.\n2. It would be nice to discuss the significance and potential applications of the proposed method to topological spaces other than the 2-dimensional sphere.", "questions": "1. Does the samples drawn using the decoder transformer have a much wider diversity compared to the baseline methods? What are the proportions of repeated samples?\n2. Is it possible to intuitively understand the distributions over the homotopy group that the transformer or the baseline methods are drawing from?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims at studying the simplicial homotopy groups of the 2-dimensional sphere. Following Wu's formula, each homotopy group is isomorphic to the quotient of a free group. Each element in the homotopy group, as a reduced word, can be connected to a sentence in the language models. The decoder part of a transformer is used to generate approximate samples from the homotopy group, and a multi-label is used to denote whether the sample satisfies relations of the homotopy groups. Multiple approaches for handling the multi-label and training the transformer are designed. Adequate numerical experiments are provided, with comparison to multiple baselines.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "1. The problem originating from algebraic topology is important.\n2. The idea is very innovative.\n3. The paper is well-written and mostly clear.\n4. Adequate numerical experiments are provided.", "weaknesses": "1. It would be helpful to discuss a bit of how the samples drawn from the homotopy group can be used to understand the topological properties of the 2-dimensional sphere, compared to the more obtainable homology groups.\n2. It would be nice to discuss the significance and potential applications of the proposed method to topological spaces other than the 2-dimensional sphere.", "questions": "1. Does the samples drawn using the decoder transformer have a much wider diversity compared to the baseline methods? What are the proportions of repeated samples?\n2. Is it possible to intuitively understand the distributions over the homotopy group that the transformer or the baseline methods are drawing from?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698826446164}, {"id": "5iNUTV030Z", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2245/Reviewer_LHhA"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this work the authors propose the use of language modeling techniques and architectures as a means for generating elements of the homotopy group of the sphere. First, the authors cite Wu's formula, which provides a connection between the homotopy groups of a sphere and a quotient group of particular combinations of free groups. Specifically, the $n+1$ homotopy group is given by a quotient group, the numerator of which is an intersection of subsets $R_i$, called *normal closures*, whose elements are strings comprised of characters $\\\\{x_i^{\\pm 1}\\\\}_{i=1}^n$.\n\nThe authors claim that generating words in the intersection $w\\in \\cap_{i=0}^n R_i$ is not straightforward, however any partial intersection (i.e. omitting at least one $i \\in \\\\{0,\\ldots, n\\\\}$) has an explicit description and can be sampled from. The authors propose a particular sampling method, and propose to use it to generate words with multi-labels as to which $R_i$ they belong. They then train a decoder-only transformer on these sampled words, leveraging the multi-labels via attention masking or prompting (eg. if $x_1 x_2^{-1} \\ldots x_t \\in R_0 \\cap R_1$, they train the model on $R_0 R_1: x_1 x_2^{-1} \\ldots x_t$). They then exploit these mechanisms at inference time to generate elements of the full intersection.\n\nThe authors propose a number of symbolic baselines (random search, an evolutionary method, and a greedy algorithm) which attempt to generate elements in the full intersection. They compare these methods to their approaches using language modeling and measure the *completion ratio*, which is essentially the percentage of generated words which are actually in the full intersection. In all but one case they find the language model is far better at generating words in the full intersection than any of the synthetic approaches.", "review_text": "In this work the authors propose the use of language modeling techniques and architectures as a means for generating elements of the homotopy group of the sphere. First, the authors cite Wu's formula, which provides a connection between the homotopy groups of a sphere and a quotient group of particular combinations of free groups. Specifically, the $n+1$ homotopy group is given by a quotient group, the numerator of which is an intersection of subsets $R_i$, called *normal closures*, whose elements are strings comprised of characters $\\\\{x_i^{\\pm 1}\\\\}_{i=1}^n$.\n\nThe authors claim that generating words in the intersection $w\\in \\cap_{i=0}^n R_i$ is not straightforward, however any partial intersection (i.e. omitting at least one $i \\in \\\\{0,\\ldots, n\\\\}$) has an explicit description and can be sampled from. The authors propose a particular sampling method, and propose to use it to generate words with multi-labels as to which $R_i$ they belong. They then train a decoder-only transformer on these sampled words, leveraging the multi-labels via attention masking or prompting (eg. if $x_1 x_2^{-1} \\ldots x_t \\in R_0 \\cap R_1$, they train the model on $R_0 R_1: x_1 x_2^{-1} \\ldots x_t$). They then exploit these mechanisms at inference time to generate elements of the full intersection.\n\nThe authors propose a number of symbolic baselines (random search, an evolutionary method, and a greedy algorithm) which attempt to generate elements in the full intersection. They compare these methods to their approaches using language modeling and measure the *completion ratio*, which is essentially the percentage of generated words which are actually in the full intersection. In all but one case they find the language model is far better at generating words in the full intersection than any of the synthetic approaches.", "strengths": "The paper's greatest strength is in it's originality. To be fair, this is not a line of work I am intimately familiar with, however a literature review on the topic suggests to me that this work is novel in both application and method.\n\nThe authors are also operating in a highly technical application area (homotopy theory), and to their credit I feel they do an admirable job conveying the necessary information without getting bogged down in details, and include additional background material in the appendix.", "weaknesses": "The largest weakness, in my opinion, is that the significance of this work is unclear. The authors motivate the work by a connection with homotopy theory, however their approach does not generate elements of the homotopy group itself but rather just the numerator (i.e. full intersection of normal closures). Moreover, there is no notion of the *coverage* of this generative procedure. More specifically, based on the coverage metric, a constant function which simply always returned a fixed element $w \\in \\cap_{i=0}^n R_i$ would score perfectly, but this would not be useful in any way.\n\nThe lack of any notion of \"coverage\" means that, even if their approach *did* generate elements of the homotopy group with some high probability, it is unclear to me exactly how those words could be used to gain any insight into the structure of the homotopy group itself, because we would not have any assurance that it was covering the full set of potential words in the homotopy group.\n\nThe paper also needs significant editing before publication. There are many mistakes and typos, some of which lead to unparsable text.", "questions": "1. Can you explain more what you meant by \"thereby enforcing it not to 'learn' the pattern of the trivial words in the denominator\" at the end of Section 3? I don't understand this statement, because (as I understand it) you were not doing anything in particular to encourage it not to learn the trivial words, and indeed the results in Figure 4 suggests that the models very much did predominantly learn trivial words.\n2. In your discussion on random search (in section 4.2) it was mentioned that the difference between naive sampling and bracket-style sampling from $R_i$ was evident in the number of generated words from intersection per batch. Could you please expand on this - what does the number of generated words from intersection per batch of random search imply about naive vs. bracket-style sampling?\n3. Could you provide a simple explanation of why the bracket-style sampling is preferred to naive? (a) It is stated in section 4.1 that \"adjacent letters of different conjugators $y_k, y_{k+1}$ do not interact with each other, resulting in reduced variability within the generated dataset\" but this notion of \"variability\" has not been defined. Also, isn't this straightforward to solve by introducing dependencies between $y_k$ and $y_{k+1}$? (b) I understand the statistics calculated in Figure 1, but I don't understand why the distribution for the bracket-style sampling is \"better\" than naive. Also, if one wanted a more uniform distribution of valleys, couldn't we just always include the inverse of every word?\n4. Could you provide quantitative evidence that the LLM results are \"covering\" the space well? I realize this might be challenging, the metrics which come to mind often involve knowing the ground-truth $cap_{i=0}^n R_i$ in order to provide a percentage, but without this there's seemingly no way to know if the model is not simply exploiting some degenerate pattern to solve the task easily but in a useless way. Would it be possible to provide such a metric if we limit ourselves to words with length less than some threshold?\n5. You mention in the \"Datasets\" section that the training dataset is infinite and generated online, and that the validation dataset is also generated in an online fashion. Does this mean there is potential for train/test overlap?\n\n**Typos / Minor Suggestions** (small selection)\n* $y_i$ in equation (1) should probably be $y_j$ (I assume the subscript here has no relation to the subscript of $R_i$).\n* The set $[R_i, R_j]$ is not defined. I assume $[R_i, R_j] = \\\\{[u,v] \\mid u \\in R_i, v \\in R_j\\\\}$.\n* I wasn't able to parse the first sentence of Remark 5.1.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work the authors propose the use of language modeling techniques and architectures as a means for generating elements of the homotopy group of the sphere. First, the authors cite Wu's formula, which provides a connection between the homotopy groups of a sphere and a quotient group of particular combinations of free groups. Specifically, the $n+1$ homotopy group is given by a quotient group, the numerator of which is an intersection of subsets $R_i$, called *normal closures*, whose elements are strings comprised of characters $\\\\{x_i^{\\pm 1}\\\\}_{i=1}^n$.\n\nThe authors claim that generating words in the intersection $w\\in \\cap_{i=0}^n R_i$ is not straightforward, however any partial intersection (i.e. omitting at least one $i \\in \\\\{0,\\ldots, n\\\\}$) has an explicit description and can be sampled from. The authors propose a particular sampling method, and propose to use it to generate words with multi-labels as to which $R_i$ they belong. They then train a decoder-only transformer on these sampled words, leveraging the multi-labels via attention masking or prompting (eg. if $x_1 x_2^{-1} \\ldots x_t \\in R_0 \\cap R_1$, they train the model on $R_0 R_1: x_1 x_2^{-1} \\ldots x_t$). They then exploit these mechanisms at inference time to generate elements of the full intersection.\n\nThe authors propose a number of symbolic baselines (random search, an evolutionary method, and a greedy algorithm) which attempt to generate elements in the full intersection. They compare these methods to their approaches using language modeling and measure the *completion ratio*, which is essentially the percentage of generated words which are actually in the full intersection. In all but one case they find the language model is far better at generating words in the full intersection than any of the synthetic approaches.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "The paper's greatest strength is in it's originality. To be fair, this is not a line of work I am intimately familiar with, however a literature review on the topic suggests to me that this work is novel in both application and method.\n\nThe authors are also operating in a highly technical application area (homotopy theory), and to their credit I feel they do an admirable job conveying the necessary information without getting bogged down in details, and include additional background material in the appendix.", "weaknesses": "The largest weakness, in my opinion, is that the significance of this work is unclear. The authors motivate the work by a connection with homotopy theory, however their approach does not generate elements of the homotopy group itself but rather just the numerator (i.e. full intersection of normal closures). Moreover, there is no notion of the *coverage* of this generative procedure. More specifically, based on the coverage metric, a constant function which simply always returned a fixed element $w \\in \\cap_{i=0}^n R_i$ would score perfectly, but this would not be useful in any way.\n\nThe lack of any notion of \"coverage\" means that, even if their approach *did* generate elements of the homotopy group with some high probability, it is unclear to me exactly how those words could be used to gain any insight into the structure of the homotopy group itself, because we would not have any assurance that it was covering the full set of potential words in the homotopy group.\n\nThe paper also needs significant editing before publication. There are many mistakes and typos, some of which lead to unparsable text.", "questions": "1. Can you explain more what you meant by \"thereby enforcing it not to 'learn' the pattern of the trivial words in the denominator\" at the end of Section 3? I don't understand this statement, because (as I understand it) you were not doing anything in particular to encourage it not to learn the trivial words, and indeed the results in Figure 4 suggests that the models very much did predominantly learn trivial words.\n2. In your discussion on random search (in section 4.2) it was mentioned that the difference between naive sampling and bracket-style sampling from $R_i$ was evident in the number of generated words from intersection per batch. Could you please expand on this - what does the number of generated words from intersection per batch of random search imply about naive vs. bracket-style sampling?\n3. Could you provide a simple explanation of why the bracket-style sampling is preferred to naive? (a) It is stated in section 4.1 that \"adjacent letters of different conjugators $y_k, y_{k+1}$ do not interact with each other, resulting in reduced variability within the generated dataset\" but this notion of \"variability\" has not been defined. Also, isn't this straightforward to solve by introducing dependencies between $y_k$ and $y_{k+1}$? (b) I understand the statistics calculated in Figure 1, but I don't understand why the distribution for the bracket-style sampling is \"better\" than naive. Also, if one wanted a more uniform distribution of valleys, couldn't we just always include the inverse of every word?\n4. Could you provide quantitative evidence that the LLM results are \"covering\" the space well? I realize this might be challenging, the metrics which come to mind often involve knowing the ground-truth $cap_{i=0}^n R_i$ in order to provide a percentage, but without this there's seemingly no way to know if the model is not simply exploiting some degenerate pattern to solve the task easily but in a useless way. Would it be possible to provide such a metric if we limit ourselves to words with length less than some threshold?\n5. You mention in the \"Datasets\" section that the training dataset is infinite and generated online, and that the validation dataset is also generated in an online fashion. Does this mean there is potential for train/test overlap?\n\n**Typos / Minor Suggestions** (small selection)\n* $y_i$ in equation (1) should probably be $y_j$ (I assume the subscript here has no relation to the subscript of $R_i$).\n* The set $[R_i, R_j]$ is not defined. I assume $[R_i, R_j] = \\\\{[u,v] \\mid u \\in R_i, v \\in R_j\\\\}$.\n* I wasn't able to parse the first sentence of Remark 5.1.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698815312953}, {"id": "GAIHjtCCc0", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2245/Reviewer_VtUd"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper considers using machine learning approaches to study a problem in algebraic topology, namely the problem of sampling elements from the numerator of Wu's formula. It does so by proposing several variants of natural language models to generate possible words which may be elements of this numerator set, after generating a synthetic dataset based on sampling from Dyck paths.", "review_text": "The paper considers using machine learning approaches to study a problem in algebraic topology, namely the problem of sampling elements from the numerator of Wu's formula. It does so by proposing several variants of natural language models to generate possible words which may be elements of this numerator set, after generating a synthetic dataset based on sampling from Dyck paths.", "strengths": "The paper has an original premise in using tools from machine learning to study a specific problem in algebraic topology which requires generative modeling. This may be of  interest to NLP researchers who wish to understand the breadth of applications which are suitable for their architectures. Table 2 also clearly shows that the proposed methods outperform the baselines.", "weaknesses": "My main concern with this paper is its relevance within the machine learning field. \n\nFrom a presentation perspective, the work emphasizes the mathematical contributions by framing the problem with Wu's formula, rather than focusing on the machine learning methodology. Because the setup of Wu's formula is quite complicated for anyone unfamiliar with algebraic topology, the paper may be difficult to parse for all but a few attendees of the conference. As written, it's hard to understand some of the notation (e.g., $[x_i, x_j]$ is defined for elements of $F$, but $[R_i, R_j]$ is not for the subgroups, which is a core component of the formulas). Further, because the paper spends so much time on these preliminaries, it is difficult to understand which machine learning methods are used and why. For example, the training procedure is difficult to understand. \n\nFrom a more substantive perspective, the paper doesn't provide convincing evidence that this use of machine learning can provide meaningful progress for the problem in the domain of mathematics. For example, scalability of the method is highlighted, but the experiments only go up to $n = 5$ due to computational constraints (p. 9). It's not immediately clear from the writing that this contribution can be built upon for further research.\n\nUltimately, the paper may be of more interest to a pure mathematics community.\n\nSome secondary concerns:\n\n1) The use of the completion/reduction ratio metrics could be better justified: can this distinguish whether the model is suggesting a wide variety of elements from that set or simply repeating the same element multiple times?\n\n2) The notation throughout the work is often difficult to understand: for example, the $\\pm$ notation in (1) is unclear, and it's not immediately clear how the $y_{k,i}$ terms in (5) are related to the notation $y_i$ often used to describe elements of the set $F$", "questions": "1) Related to secondary concern 1 above: how do we know that the model is generalizing well, i.e. finding words which are not from the training set?\n\n2) Are there simpler generative models which could work well for this task?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers using machine learning approaches to study a problem in algebraic topology, namely the problem of sampling elements from the numerator of Wu's formula. It does so by proposing several variants of natural language models to generate possible words which may be elements of this numerator set, after generating a synthetic dataset based on sampling from Dyck paths.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The paper has an original premise in using tools from machine learning to study a specific problem in algebraic topology which requires generative modeling. This may be of  interest to NLP researchers who wish to understand the breadth of applications which are suitable for their architectures. Table 2 also clearly shows that the proposed methods outperform the baselines.", "weaknesses": "My main concern with this paper is its relevance within the machine learning field. \n\nFrom a presentation perspective, the work emphasizes the mathematical contributions by framing the problem with Wu's formula, rather than focusing on the machine learning methodology. Because the setup of Wu's formula is quite complicated for anyone unfamiliar with algebraic topology, the paper may be difficult to parse for all but a few attendees of the conference. As written, it's hard to understand some of the notation (e.g., $[x_i, x_j]$ is defined for elements of $F$, but $[R_i, R_j]$ is not for the subgroups, which is a core component of the formulas). Further, because the paper spends so much time on these preliminaries, it is difficult to understand which machine learning methods are used and why. For example, the training procedure is difficult to understand. \n\nFrom a more substantive perspective, the paper doesn't provide convincing evidence that this use of machine learning can provide meaningful progress for the problem in the domain of mathematics. For example, scalability of the method is highlighted, but the experiments only go up to $n = 5$ due to computational constraints (p. 9). It's not immediately clear from the writing that this contribution can be built upon for further research.\n\nUltimately, the paper may be of more interest to a pure mathematics community.\n\nSome secondary concerns:\n\n1) The use of the completion/reduction ratio metrics could be better justified: can this distinguish whether the model is suggesting a wide variety of elements from that set or simply repeating the same element multiple times?\n\n2) The notation throughout the work is often difficult to understand: for example, the $\\pm$ notation in (1) is unclear, and it's not immediately clear how the $y_{k,i}$ terms in (5) are related to the notation $y_i$ often used to describe elements of the set $F$", "questions": "1) Related to secondary concern 1 above: how do we know that the model is generalizing well, i.e. finding words which are not from the training set?\n\n2) Are there simpler generative models which could work well for this task?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698583288718}], "openreview_url": "https://openreview.net/forum?id=AfiM6F2YPY", "arxiv_id": "2306.16951", "paper_pdf": "papers/AfiM6F2YPY.pdf", "paper_pdf_sha256": "24d2320351b78649e01beba2a5acb2c31220c1a93bf8b1f08254e2360be1ec2c", "paper_pdf_bytes": 785192, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-in-algebraic-topology/gen-simplicial-cycles", "code_repository": "ml-in-algebraic-topology/gen-simplicial-cycles", "code_commit": "e9149c42de7c586704be1f593d36193561622911", "code_archive": "repos/AfiM6F2YPY.zip", "code_archive_sha256": "a863891145e16f041efbba923623dde580b43523c5b7b4ecfd500f91b84d15b3", "code_archive_bytes": 90952, "code_file_count": 5, "code_extensions": {".py": 4, ".ipynb": 1}, "github_disk_usage_kb": 120, "github_languages": {"Jupyter Notebook": 40765, "Python": 19388}, "github_archived": false, "github_pushed_at": "2024-05-28T09:23:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/applying-language-models-to-algebraic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZBMpG7fWwOP", "year": 2023, "status": "rejected", "title": "Game Theoretic Mixed Experts for Combinational Adversarial Machine Learning", "authors": ["Ethan Rathbun", "Kaleel Mahmood", "Sohaib Ahmad", "Caiwen Ding", "Marten van Dijk"], "authorids": ["~Ethan_Rathbun1", "~Kaleel_Mahmood1", "sohaib.ahmad@uconn.edu", "~Caiwen_Ding1", "~Marten_van_Dijk1"], "authors_source": "OpenReview API", "abstract": "Recent advances in adversarial machine learning have shown that defenses considered to be robust are actually susceptible to adversarial attacks which are specifically tailored to target their weaknesses. These defenses include Barrage of Random Transforms (BaRT), Friendly Adversarial Training (FAT), Trash is Treasure (TiT) and ensemble models made up of Vision Transformers (ViTs), Big Transfer models and Spiking Neural Networks (SNNs). It remains an open question, however, as to whether the adversarial examples designed to target one defense will be similarly misclassified by another defense. In this paper, we provide the first adversarial defense transferability study, as well as a game theoretic framework for ensemble adversarial attacks and defenses. Our framework is called Game theoretic Mixed Experts (GaME) and is designed to find the Mixed-Nash strategy for an attacker that can employ compositional adversarial attacks. We show that this framework creates an ensemble of defenses with greater robustness than a combinational defense with a uniform or random probability distribution. Overall, our framework and analyses advance the field of adversarial machine learning by yielding new insights into compositional attack and defense formulations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ubJYQVssazc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5815/Reviewer_ZYXb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper provides a game-theoretic analysis of adversarial robustness. In the adversarial examples game the authors propose, attackers choose an attack method (including both the optimization method and the defense to attack) while defenders choose a defense. This game is connected to the study of transferability of adversarial examples across diverse attacks. With respect to transferability, often attacks which are strong against one attack produce adversarial examples that do not transfer well to another defense. The authors then show that one can determine an optimal randomized strategy for attacker and defender by considering a matrix game where the payoffs are the success rates of how well each attack transfers to each defense (the values called $r_{d, a}$ in the paper). An optimal defense strategy is computed for CIFAR-10 and Tiny-ImageNet given several defenses and it is shown that it has higher robust accuracy than any single one of the defenses.", "review_text": "Overall, I found the paper to be an interesting analysis of adversarial transferability and its connection to randomized defenses via game theory. I would encourage the authors to work on the presentation of their ideas so that others can more easily appreciate their results.", "strengths": "In general, I think the idea of the paper is quite interesting and the theory and experiments are explained clearly. As far as I know, nobody has explored an optimal strategy for randomizing over multiple defenses by using game theory. The experiments seem to support the theoretical analysis and there is a nice result that the randomized strategy is more robust than any single of its members. The connection between the game theoretic analysis and transferability of adversarial examples is also quite interesting.\n\nI think that the main weakness of the paper is that the presentation could be improved. I found that a lot of the paper seemed to focus too much on just the methods, mathematical analysis, and algorithms, and not enough on the motivation for them. This meant that I didn't appreciate the ideas behind the paper until I had more fully internalized the methodology sections. I worry that readers will just read the title, abstract, or introduction and not appreciate all the contributions.\n\nIn particular, I think there might be too much emphasis on the parts of the paper dealing with transferability of adversarial examples. This idea has been explored a lot in the past and a casual reader might assume that the paper is just another study on transferability. Instead, I think the paper might be more impactful if the authors emphasize the game-theoretic analysis and the idea of coming up with better randomized defenses. A title that might be more attention-grabbing could be something like \"Optimal Randomized Adversarial Defenses via Game-Theoretic Analysis of Attack Transferability.\" If the authors focus on the goal of picking a good randomized defense and use the analysis of transferability more as an argument to justify that goal rather than as a primary contribution, I think more people will find the paper novel and interesting. If more space is needed to add more motivation, I think some of the details of the new attacks could be moved to the appendix since I think the game-theory ideas and results are more interesting than the attacks. But I'm just one person so maybe others will think the attack/transfer part of the paper is more interesting!\n\nBesides the presentation, one objection to the method in the paper (and randomized defenses in general) is that a defender could just re-submit the same adversarial example until the defender randomly picks the vulnerable model. For instance, say an attacker is trying to post offensive pictures by fooling a content-filtering system on social media. Then, they could just resubmit the pictures until the content-filtering system randomly picks a vulnerable classifier. It would be good to discuss this limitation somewhere in the paper.\n\nSmaller issues/suggestions\n * On page 9, you say \"For CIFAR-10 instance r* = .573, meaning the worst expected performance of the ensemble against\nthese attacks is to get a robust accuracy of 57.5%.\" Shouldn't it be 57.3%? Also, seems like there is a typo and it should be \"For CIFAR-10 for instance\" or just \"For CIFAR-10\".\n * This seems like a relevant paper to cite, although your approach is definitely quite different: Balcan et al. Nash Equilibria and Pitfalls of Adversarial Training in Adversarial Robustness Games. (Also I realize that it came out after you submitted, so use your judgement as to whether you think it should be cited.)", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper provides a game-theoretic analysis of adversarial robustness. In the adversarial examples game the authors propose, attackers choose an attack method (including both the optimization method and the defense to attack) while defenders choose a defense. This game is connected to the study of transferability of adversarial examples across diverse attacks. With respect to transferability, often attacks which are strong against one attack produce adversarial examples that do not transfer well to another defense. The authors then show that one can determine an optimal randomized strategy for attacker and defender by considering a matrix game where the payoffs are the success rates of how well each attack transfers to each defense (the values called $r_{d, a}$ in the paper). An optimal defense strategy is computed for CIFAR-10 and Tiny-ImageNet given several defenses and it is shown that it has higher robust accuracy than any single one of the defenses.", "strength_and_weaknesses": "In general, I think the idea of the paper is quite interesting and the theory and experiments are explained clearly. As far as I know, nobody has explored an optimal strategy for randomizing over multiple defenses by using game theory. The experiments seem to support the theoretical analysis and there is a nice result that the randomized strategy is more robust than any single of its members. The connection between the game theoretic analysis and transferability of adversarial examples is also quite interesting.\n\nI think that the main weakness of the paper is that the presentation could be improved. I found that a lot of the paper seemed to focus too much on just the methods, mathematical analysis, and algorithms, and not enough on the motivation for them. This meant that I didn't appreciate the ideas behind the paper until I had more fully internalized the methodology sections. I worry that readers will just read the title, abstract, or introduction and not appreciate all the contributions.\n\nIn particular, I think there might be too much emphasis on the parts of the paper dealing with transferability of adversarial examples. This idea has been explored a lot in the past and a casual reader might assume that the paper is just another study on transferability. Instead, I think the paper might be more impactful if the authors emphasize the game-theoretic analysis and the idea of coming up with better randomized defenses. A title that might be more attention-grabbing could be something like \"Optimal Randomized Adversarial Defenses via Game-Theoretic Analysis of Attack Transferability.\" If the authors focus on the goal of picking a good randomized defense and use the analysis of transferability more as an argument to justify that goal rather than as a primary contribution, I think more people will find the paper novel and interesting. If more space is needed to add more motivation, I think some of the details of the new attacks could be moved to the appendix since I think the game-theory ideas and results are more interesting than the attacks. But I'm just one person so maybe others will think the attack/transfer part of the paper is more interesting!\n\nBesides the presentation, one objection to the method in the paper (and randomized defenses in general) is that a defender could just re-submit the same adversarial example until the defender randomly picks the vulnerable model. For instance, say an attacker is trying to post offensive pictures by fooling a content-filtering system on social media. Then, they could just resubmit the pictures until the content-filtering system randomly picks a vulnerable classifier. It would be good to discuss this limitation somewhere in the paper.\n\nSmaller issues/suggestions\n * On page 9, you say \"For CIFAR-10 instance r* = .573, meaning the worst expected performance of the ensemble against\nthese attacks is to get a robust accuracy of 57.5%.\" Shouldn't it be 57.3%? Also, seems like there is a typo and it should be \"For CIFAR-10 for instance\" or just \"For CIFAR-10\".\n * This seems like a relevant paper to cite, although your approach is definitely quite different: Balcan et al. Nash Equilibria and Pitfalls of Adversarial Training in Adversarial Robustness Games. (Also I realize that it came out after you submitted, so use your judgement as to whether you think it should be cited.)", "clarity,_quality,_novelty_and_reproducibility": "The descriptions of the method are clear and in general the math and experiments seem correct. I did not closely inspect all the experimental details. While the attacks presented are only slight modifications of existing attacks, the game-theoretical analysis seems quite original.", "summary_of_the_review": "Overall, I found the paper to be an interesting analysis of adversarial transferability and its connection to randomized defenses via game theory. I would encourage the authors to work on the presentation of their ideas so that others can more easily appreciate their results.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666820354347}, {"id": "oKKF0txQ_r", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5815/Reviewer_hEcL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies how adversarial examples crafted for one defensive model transfer on other models, and correspondingly how random defensive models help adversarial robustness.", "review_text": "As is.", "strengths": "# Strength \nThe studies topic is an interesting and practical one. According to my knowledge, such topic has not been discussed in the past. The theoretical results seems sufficient to me. However, I have not checked its every detail. \n\n# Weakness\nThe empirical verification is not very convincing nor sufficient. According to the theory established by this work, I expect the author to construct a state-of-the-art defense model by randomly choosing the prediction model from an ensemble of models.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies how adversarial examples crafted for one defensive model transfer on other models, and correspondingly how random defensive models help adversarial robustness.", "strength_and_weaknesses": "# Strength \nThe studies topic is an interesting and practical one. According to my knowledge, such topic has not been discussed in the past. The theoretical results seems sufficient to me. However, I have not checked its every detail. \n\n# Weakness\nThe empirical verification is not very convincing nor sufficient. According to the theory established by this work, I expect the author to construct a state-of-the-art defense model by randomly choosing the prediction model from an ensemble of models.", "clarity,_quality,_novelty_and_reproducibility": "This paper is clear. The quality is relatively good.", "summary_of_the_review": "As is.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666793610977}, {"id": "dD29LHz16X", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5815/Reviewer_rrfN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper considers a finite set of attackers and a finite set of defenders and views them in a game-theoretic framework to choose the optimal choices for attacker and defender.\n\nSpecifically, it solves a min-max problem to find the optimal weights for attackers and defender.\n\n\n", "review_text": "- The paper is not well-organized and well-written. The notions used are confusing. \n- The theoretical setting is not feasible and practical.\n- The contribution is not clear and significant.\n", "strengths": "Strength\n-  The paper mentions to many methods in adversarial machine learning.\n-  Some experiments about the behaviors are pretty interesting though not new.\n\nWeaknesses\n-  The paper is not well-organized and well-written. The sections are not really logically linked and coherent each other.\n-  Section 5 is messy with inconsistent notions.\n   - $\\theta(x + a(\\theta, x ,y))$: $\\theta$ is generally for a model parameter and it is strange to write $\\theta(x + a(\\theta, x ,y))$.\n   - The notations $p_a$ and $a$ are confusing because it is easy to think $p_a$ depends on $a$. Similarly for $p_d$ and $d$.\n   - No clear definition for $a_i(U,x,y)$. Here $U$ is a set of defenders, but why $a_i(U,x,y)$ returns a common perturbation for every defender in $U$.  \n-  The assumption about finite number of defenders is not feasible. We might interpret $d_i$ as a defender type (e.g., CNNs, ViTs). However, it seems not sufficiently rich to represent the model parameters. \n- The main experiment in Section 6 is very humble.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper considers a finite set of attackers and a finite set of defenders and views them in a game-theoretic framework to choose the optimal choices for attacker and defender.\n\nSpecifically, it solves a min-max problem to find the optimal weights for attackers and defender.\n\n\n", "strength_and_weaknesses": "Strength\n-  The paper mentions to many methods in adversarial machine learning.\n-  Some experiments about the behaviors are pretty interesting though not new.\n\nWeaknesses\n-  The paper is not well-organized and well-written. The sections are not really logically linked and coherent each other.\n-  Section 5 is messy with inconsistent notions.\n   - $\\theta(x + a(\\theta, x ,y))$: $\\theta$ is generally for a model parameter and it is strange to write $\\theta(x + a(\\theta, x ,y))$.\n   - The notations $p_a$ and $a$ are confusing because it is easy to think $p_a$ depends on $a$. Similarly for $p_d$ and $d$.\n   - No clear definition for $a_i(U,x,y)$. Here $U$ is a set of defenders, but why $a_i(U,x,y)$ returns a common perturbation for every defender in $U$.  \n-  The assumption about finite number of defenders is not feasible. We might interpret $d_i$ as a defender type (e.g., CNNs, ViTs). However, it seems not sufficiently rich to represent the model parameters. \n- The main experiment in Section 6 is very humble.\n", "clarity,_quality,_novelty_and_reproducibility": "- The idea of min-max problem in attack and defense is not new  (e.g. [1]) and other relevant works.\n- There is no literature for other works investigating game-theoretic perspective of attacks and defenses (e.g., [2], [3], and so on). \n- The writing needs a significant improvement. \n- The contribution of the paper is not clear. \n\n[1] Jingkang Wang, Tianyun Zhang, Sijia Liu, Pin-Yu Chen, Jiacen Xu, Makan Fardad, and Bo Li. Adversarial attack generation empowered by min-max optimization. Advances in Neural Information Processing Systems.\n\n[2]  Le, T., Tuan Bui, A., Minh Tri Tue, L., Zhao, H., Montague, P., Tran, Q. &amp; Phung, D.. (2022).  On Global-view Based Defense via Adversarial Attack and Defense Risk Guaranteed Bounds . Proceedings of The 25th International Conference on Artificial Intelligence and Statistics.\n\n[3] Cranko, Z., Menon, A., Nock, R., Ong, C. S., Shi, Z., and Walder, C. (2019). Monge blunts bayes: Hardness results for adversarial training. In Chaudhuri, K. and Salakhutdinov, R., editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research\n\n\n", "summary_of_the_review": "- The paper is not well-organized and well-written. The notions used are confusing. \n- The theoretical setting is not feasible and practical.\n- The contribution is not clear and significant.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "There is no ethics concerns.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666694282178}, {"id": "uzQtQLYE5n3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5815/Reviewer_once"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers a set of recent defenses against adversarial attacks on the image domain and uses existing attacks or develops them to capture a game-theoretic interaction between the attacks and defenses. To obtain the utility functions, they evaluate the effectiveness of different attacks on different defenses, thereby conducting a transferability study. Eventually, they consider a mixed strategy over a single-classifier or a less-than-n-ensemble selection for the defender by formulating the game as a zero-sum game and solving for a Nash-equilibria in this game. They show that this is more effective than using single defenses.", "review_text": "See above.", "strengths": "### Things I liked\n\n- I like how the authors concisely described many of the recent defenses, the attacks and highlighted their relevance for this work.\n\n- The authors propose new attacks (eg. MIME) by composing existing attacks, situate its need well in the landscape of existing attacks, and showcase its effectiveness.\n\n### Things that need clarification / improvement\n\n- The study of transferability of attacks on defenses, the formulation of interaction between attacks & defenses, and using game-theoretic equilibrium to come up with a mixed-strategy over defenses is not novel. [1] does all this-- given the white-box threat model, their game-theoretic formulation considers that the attacker may even know the defender's strategy and thus, considers a Stackelberg Equilibria as opposed to a Nash Eq. This paper does the study on more recent attacks and defenses and consider a more relaxed threat-model. Thus, all the 3 novel questions the paper presents on page 1 already have answers. Having said that, I do appreciate the renewed study on recent attacks and defenses but cannot support the novelty claim.\n\n- While the reward function mostly concentrates on the loss function value for the adversarially perturbed examples, it doesn't account for the loss on the non-perturbed examples. I would surely suggest the authors to consider the loss on the latter examples in the reward function, at least for the defender, similar to [1]. Otherwise, the strategy found may over optimize for accuracy/loss on the adversarial examples while losing out on accuracy for the true test distribution.\n\n- Given this is a zero-sum game, the Nash Eq. generates a min-max payoff and can be calculated in polynomial time. Hence the statement \"These linear programs can be solved using weakly polynomial time algorithms like the interior point method Karmarkar (1984), or those developed in Vaidya (1989)\" makes me wonder if the authors are aware of the subtle aspects of their formulation.\n\n- Game_n considers a larger space (that explodes exponentially $~2^n$) of defense actions where the defender can use up to n-classifiers for ensembling. Note that this incurs much larger computational cost as one needs to estimate the utility values for each of the attack-defense pairs using $N$ samples. Hence beyond simply reporting accuracy on the adversarial examples and the test set, the authors should report the time taken to form the game-matrix. Also, the authors should discuss how effective the final strategy becomes as the number of samples used for the utility estimation ($N$), and therefore the cost, increases or decreases.\n\n- Some study on if an attacker can attack the game-theoretic defense strategy is necessary to understand if this too is an effective defense. Based on results in [1], I would assume black-box distillation would not be as effective. Would be interesting to see how this holds for different values of $n$ $GAME_n$ and $N$ in estimating utilities.\n\n[1] Sengupta, S., Chakraborti, T., & Kambhampati, S. (2019). Mtdeep: Moving target defense to boost the security of deep neural nets against adversarial attacks. In International Conference on Decision and Game Theory for Security.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper considers a set of recent defenses against adversarial attacks on the image domain and uses existing attacks or develops them to capture a game-theoretic interaction between the attacks and defenses. To obtain the utility functions, they evaluate the effectiveness of different attacks on different defenses, thereby conducting a transferability study. Eventually, they consider a mixed strategy over a single-classifier or a less-than-n-ensemble selection for the defender by formulating the game as a zero-sum game and solving for a Nash-equilibria in this game. They show that this is more effective than using single defenses.", "strength_and_weaknesses": "### Things I liked\n\n- I like how the authors concisely described many of the recent defenses, the attacks and highlighted their relevance for this work.\n\n- The authors propose new attacks (eg. MIME) by composing existing attacks, situate its need well in the landscape of existing attacks, and showcase its effectiveness.\n\n### Things that need clarification / improvement\n\n- The study of transferability of attacks on defenses, the formulation of interaction between attacks & defenses, and using game-theoretic equilibrium to come up with a mixed-strategy over defenses is not novel. [1] does all this-- given the white-box threat model, their game-theoretic formulation considers that the attacker may even know the defender's strategy and thus, considers a Stackelberg Equilibria as opposed to a Nash Eq. This paper does the study on more recent attacks and defenses and consider a more relaxed threat-model. Thus, all the 3 novel questions the paper presents on page 1 already have answers. Having said that, I do appreciate the renewed study on recent attacks and defenses but cannot support the novelty claim.\n\n- While the reward function mostly concentrates on the loss function value for the adversarially perturbed examples, it doesn't account for the loss on the non-perturbed examples. I would surely suggest the authors to consider the loss on the latter examples in the reward function, at least for the defender, similar to [1]. Otherwise, the strategy found may over optimize for accuracy/loss on the adversarial examples while losing out on accuracy for the true test distribution.\n\n- Given this is a zero-sum game, the Nash Eq. generates a min-max payoff and can be calculated in polynomial time. Hence the statement \"These linear programs can be solved using weakly polynomial time algorithms like the interior point method Karmarkar (1984), or those developed in Vaidya (1989)\" makes me wonder if the authors are aware of the subtle aspects of their formulation.\n\n- Game_n considers a larger space (that explodes exponentially $~2^n$) of defense actions where the defender can use up to n-classifiers for ensembling. Note that this incurs much larger computational cost as one needs to estimate the utility values for each of the attack-defense pairs using $N$ samples. Hence beyond simply reporting accuracy on the adversarial examples and the test set, the authors should report the time taken to form the game-matrix. Also, the authors should discuss how effective the final strategy becomes as the number of samples used for the utility estimation ($N$), and therefore the cost, increases or decreases.\n\n- Some study on if an attacker can attack the game-theoretic defense strategy is necessary to understand if this too is an effective defense. Based on results in [1], I would assume black-box distillation would not be as effective. Would be interesting to see how this holds for different values of $n$ $GAME_n$ and $N$ in estimating utilities.\n\n[1] Sengupta, S., Chakraborti, T., & Kambhampati, S. (2019). Mtdeep: Moving target defense to boost the security of deep neural nets against adversarial attacks. In International Conference on Decision and Game Theory for Security.", "clarity,_quality,_novelty_and_reproducibility": "The paper's novelty is limited in terms of technical ideas but empirically, it does evaluate a game-theoretic setup on recent defenses and attacks. Given the pace of development in this direction, it is definitely worth-while. I would definitely ask the authors to refactor their contribution statement in this way.\n\nThe lack of mention on what $N$ is and a more though study as how $N$ affects the utility estimates and thus the overall game-theoretic effectiveness, and computation cost is important to give an overall picture.", "summary_of_the_review": "See above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666540247402}], "openreview_url": "https://openreview.net/forum?id=ZBMpG7fWwOP", "arxiv_id": "2211.14669", "paper_pdf": "papers/ZBMpG7fWwOP.pdf", "paper_pdf_sha256": "9b699b192900249e4300f19a18b3a74f4249bb6682af18d300aadef44fc7e8c2", "paper_pdf_bytes": 323014, "paper_pdf_source": "openreview", "code_url": "https://github.com/EthanRath/Game-Theoretic-Mixed-Experts", "code_repository": "EthanRath/Game-Theoretic-Mixed-Experts", "code_commit": "561d3e7549dff1f4d494f6622dc558403d57e3ce", "code_archive": "repos/ZBMpG7fWwOP.zip", "code_archive_sha256": "92b00fe0bf6109649fc1ce5db0e36177848063f8a5e8f324e6727c94c6b6f273", "code_archive_bytes": 247566, "code_file_count": 36, "code_extensions": {".py": 36}, "github_disk_usage_kb": 221, "github_languages": {"Python": 636880}, "github_archived": false, "github_pushed_at": "2024-03-24T23:51:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/game-theoretic-mixed-experts-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1oEvY1a67c1", "year": 2022, "status": "rejected", "title": "If your data distribution shifts, use self-learning", "authors": ["Evgenia Rusak", "Steffen Schneider", "George Pachitariu", "Luisa Eck", "Peter Vincent Gehler", "Oliver Bringmann", "Wieland Brendel", "Matthias Bethge"], "authorids": ["~Evgenia_Rusak1", "~Steffen_Schneider1", "~George_Pachitariu2", "~Luisa_Eck1", "~Peter_Vincent_Gehler1", "~Oliver_Bringmann1", "~Wieland_Brendel1", "~Matthias_Bethge1"], "authors_source": "OpenReview API", "abstract": "In this paper, we demonstrate that self-learning techniques like entropy minimization or pseudo-labeling are simple, yet effective techniques for increasing test performance under domain shifts. Our results show that self-learning consistently increases performance under distribution shifts, irrespective of the model architecture, the pre-training technique or the type of distribution shift. At the same time, self-learning is simple to use in practice because it does not require knowledge or access to the original training data or scheme, is robust to hyperparameter choices, is straight-forward to implement and requires only a few training epochs. This makes self-learning techniques highly attractive for any practitioner who applies machine learning algorithms in the real world. We present state-of-the art adaptation results on CIFAR10-C (8.5% error),  ImageNet-C (22.0% mCE), ImageNet-R (17.4% error) and ImageNet-A (14.8% error), theoretically study the dynamics of self-supervised adaptation methods and propose a new classification dataset (ImageNet-D) which is challenging even with adaptation.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "8iANtp4NwiV", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3402/Reviewer_b5KJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies effectiveness of self-training to improve test time performance when the distribution of test data is not similar to the training data. The paper more specifically focuses on source-free domain adaptation settings where the source target data is not available. In this setup self-training has been tested as an additional step on top of  different robustness and adaptation approaches such as robust pretraining, unsupervised domain adaptation and self-supervised pretraining. The paper shows improvement on multiple ImageNet variants and CIFAR10-C, and also introduces ImageNet-D dataset as a new benchmark. ImageNet-D has been produced by matching label space of IN datasets with DomainNet data provided in Visual domain adaptation challenges. The main contribution of this paper is to perform a systematic and large study of self-training as a method to deal with distribution shifts.", "review_text": "Strength:\n- The paper is very well written and organized and easy to follow. They also explain the target and overarching goal very clearly. \n- The empirical study is relevant and beneficial for the research community.\n- The proposed suggestion is easy to use/implement, and there are interesting results supporting the main claim.\n\n\nMain concerns and comments to improve the paper:\n- Real world evaluation beyond natural images is missing which helps to solidify claims of the paper: The paper focuses on improvement on curated datasets such as IN-family. There are prior works that show similar results including [Barret’20, Chen’20, and etc]  and premises of self-training have been established previously.  However the paper does not consider real-world applications such as test time performance drop in medical data or satellite images and only limited to curated datasets which limits the future impact of the work.\n- Definition of mCE (mean Corruption Error) and the calculation procedure is required. \n- The method is of an incremental nature and not novel.\n\nSuggestion to improve the paper:\n- Consider adding/studying other domain dataset and tasks. There are multiple open source datasets available.   You can check the WILDS benchmark. \n- Adding model calibration and statistical analysis of the results can boost the validity of the result section. \n- Consider using Big Transfer (BiT) model performance as one family of Models for ImageNet-scale datatsets. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies effectiveness of self-training to improve test time performance when the distribution of test data is not similar to the training data. The paper more specifically focuses on source-free domain adaptation settings where the source target data is not available. In this setup self-training has been tested as an additional step on top of  different robustness and adaptation approaches such as robust pretraining, unsupervised domain adaptation and self-supervised pretraining. The paper shows improvement on multiple ImageNet variants and CIFAR10-C, and also introduces ImageNet-D dataset as a new benchmark. ImageNet-D has been produced by matching label space of IN datasets with DomainNet data provided in Visual domain adaptation challenges. The main contribution of this paper is to perform a systematic and large study of self-training as a method to deal with distribution shifts.", "main_review": "Strength:\n- The paper is very well written and organized and easy to follow. They also explain the target and overarching goal very clearly. \n- The empirical study is relevant and beneficial for the research community.\n- The proposed suggestion is easy to use/implement, and there are interesting results supporting the main claim.\n\n\nMain concerns and comments to improve the paper:\n- Real world evaluation beyond natural images is missing which helps to solidify claims of the paper: The paper focuses on improvement on curated datasets such as IN-family. There are prior works that show similar results including [Barret’20, Chen’20, and etc]  and premises of self-training have been established previously.  However the paper does not consider real-world applications such as test time performance drop in medical data or satellite images and only limited to curated datasets which limits the future impact of the work.\n- Definition of mCE (mean Corruption Error) and the calculation procedure is required. \n- The method is of an incremental nature and not novel.\n\nSuggestion to improve the paper:\n- Consider adding/studying other domain dataset and tasks. There are multiple open source datasets available.   You can check the WILDS benchmark. \n- Adding model calibration and statistical analysis of the results can boost the validity of the result section. \n- Consider using Big Transfer (BiT) model performance as one family of Models for ImageNet-scale datatsets. \n", "summary_of_the_review": "The paper ran a large scale study to establish the benefits of self-training for data distribution shift. The technical contributions of the paper are only marginal and more of incremental nature, however the study itself is valuable and can be beneficial for the community. This study can significantly get boosted by diversifying the range of datasets and tasks under study. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635890186383}, {"id": "19vgZuQ2h05", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3402/Reviewer_PUq6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers using self-learning / self-training as a test-time adaptation step for adapting to distribution shifts, showing that it is complementary with pre-training methods, domain adaptation methods used for training the model. The test-time adaptation setting (source-free domain adaptation setting) is taken to be the setting where models can use target data to adapt but cannot use the source data (which was used for pre-training). ", "review_text": "Strengths\n- The paper shows robust results where adding self-learning on top of existing methods (by starting from the checkpoint of the previous methods) improves performance, even when the existing method has already seen the unlabeled target data (UDA methods). One interesting test to add here would be where self-learning helps on top of using self-learning for UDA (like applying Noisy student on source and target unlabeled data). It seems to be complementary to doing self-learning (Noisy student) on the ImageNet tasks.\n- There are interesting results on how self-learning improves large pre-trained models beyond BN adaptation. However, a comparison to other methods is missing here.\n- As a pretty generic method, self-learning also seems to improve over test-time adaptation approaches like Test-time Training which had a more similar setting in mind. This shows that the gains are not purely from the setup of test-time adaptation.\n- The paper is pretty comprehensive in methods and architectures, also testing on recent self-supervised methods (DINO). \n- There are some interesting experiments on the proposed variant of DomainNet they call ImageNet-D, which tests transfer from ImageNet to overlapping classes in DomainNet. DomainNet allows for domain-level insights, breaking down where self-learning fails. However, I am left wondering whether ImageNet to ImageNet-D is different from the Real -> any other domain part of DomainNet. Perhaps the main gains here are that the scale of the Real data is much larger and that we can test models that are trained on ImageNet rather than a different special dataset. \n- The paper gives an interesting preliminary analysis of the self-learning dynamics in a simple two-point setting, with some results for setting student and teacher temperatures in self-learning that seem to bear out in practice as well. The result seems general beyond test-time adaptation.\n\nWeaknesses\n- There have been quite a few works that focus on how self-training / self-learning improves distribution shift and how self-training and pre-training stack together: https://arxiv.org/abs/2012.04550, http://proceedings.mlr.press/v139/cai21b.html (in the experimental section), https://arxiv.org/abs/2002.11361, https://arxiv.org/abs/2106.04732, https://arxiv.org/abs/2006.10032, https://arxiv.org/abs/2012.11460, https://arxiv.org/abs/2006.06882 (this one is just about stacking them, not distribution shift). In some sense, the main premise of the paper (for example the title) is somewhat well-established. Discussion about these and your distinction from them would improve the positioning of the paper.\n- The setting is a little confusing at first - it would seem that from an \"access\" point of view, we should always have access to the source dataset, while we may not be able to know the target dataset. I think this comes from the way that it's explained in the context of UDA. From what I understand, the main thing with the \"source-free\" setting is that the source dataset is too large so we need to do pre-training. I think its a bit more straightforward to start with the OOD generalization setting (what you call ad-hoc) without any knowledge of the target, then add the knowledge of the target data. Also, the \"test-time adaptation\" term has strong connotations of the online setting, which I believe is not being considered here, but should be distinguished clearly.\n- Fig 1 clarity: Overall, the figure doesn't say very much about the source-free setting (it just looks like an option, but pictorially it doesn't look different from the other options). Some picture-level aspects were unclear, like what the different between the icons in the gray (pre-training) box are, and what the image + 3 gray shapes in the top represent, as well as what the orange stripe in the 3 gray shapes in the adaptation phase denote.  \n- Table 1,2 seem to use different self-learning algorithms. How were the methods for each table chosen? In Section 6, the paper supports Robust pseudo labeling, but Table 2 uses ENT for the main results.\n- In Table 4, why is one comparison against TENT while the other against TTT? \n- The takeaway in Table 8 that updating all affine layers is important is perhaps misleading, since in Table 5, tuning all affine layers is the worst option. It is also not fully specified what tuning all affine layers means?\n- It may be easier in terms of presentation to talk about ImageNet-D right before the results in section 7.\n- The results on ImageNet-D so far do not seem unique to test-time adaptation with self-learning; they would seem to occur in other settings too. Are there any aspects of ImageNet-D that test aspects of test-time adaptation methods?\n- As the paper suggests ImageNet-D as a robustness benchmark, it's probably worth discussing whether we would expect our models to generalize to such disparate domains - for some of the domains, the paper reports close to 0% accuracy. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers using self-learning / self-training as a test-time adaptation step for adapting to distribution shifts, showing that it is complementary with pre-training methods, domain adaptation methods used for training the model. The test-time adaptation setting (source-free domain adaptation setting) is taken to be the setting where models can use target data to adapt but cannot use the source data (which was used for pre-training). ", "main_review": "Strengths\n- The paper shows robust results where adding self-learning on top of existing methods (by starting from the checkpoint of the previous methods) improves performance, even when the existing method has already seen the unlabeled target data (UDA methods). One interesting test to add here would be where self-learning helps on top of using self-learning for UDA (like applying Noisy student on source and target unlabeled data). It seems to be complementary to doing self-learning (Noisy student) on the ImageNet tasks.\n- There are interesting results on how self-learning improves large pre-trained models beyond BN adaptation. However, a comparison to other methods is missing here.\n- As a pretty generic method, self-learning also seems to improve over test-time adaptation approaches like Test-time Training which had a more similar setting in mind. This shows that the gains are not purely from the setup of test-time adaptation.\n- The paper is pretty comprehensive in methods and architectures, also testing on recent self-supervised methods (DINO). \n- There are some interesting experiments on the proposed variant of DomainNet they call ImageNet-D, which tests transfer from ImageNet to overlapping classes in DomainNet. DomainNet allows for domain-level insights, breaking down where self-learning fails. However, I am left wondering whether ImageNet to ImageNet-D is different from the Real -> any other domain part of DomainNet. Perhaps the main gains here are that the scale of the Real data is much larger and that we can test models that are trained on ImageNet rather than a different special dataset. \n- The paper gives an interesting preliminary analysis of the self-learning dynamics in a simple two-point setting, with some results for setting student and teacher temperatures in self-learning that seem to bear out in practice as well. The result seems general beyond test-time adaptation.\n\nWeaknesses\n- There have been quite a few works that focus on how self-training / self-learning improves distribution shift and how self-training and pre-training stack together: https://arxiv.org/abs/2012.04550, http://proceedings.mlr.press/v139/cai21b.html (in the experimental section), https://arxiv.org/abs/2002.11361, https://arxiv.org/abs/2106.04732, https://arxiv.org/abs/2006.10032, https://arxiv.org/abs/2012.11460, https://arxiv.org/abs/2006.06882 (this one is just about stacking them, not distribution shift). In some sense, the main premise of the paper (for example the title) is somewhat well-established. Discussion about these and your distinction from them would improve the positioning of the paper.\n- The setting is a little confusing at first - it would seem that from an \"access\" point of view, we should always have access to the source dataset, while we may not be able to know the target dataset. I think this comes from the way that it's explained in the context of UDA. From what I understand, the main thing with the \"source-free\" setting is that the source dataset is too large so we need to do pre-training. I think its a bit more straightforward to start with the OOD generalization setting (what you call ad-hoc) without any knowledge of the target, then add the knowledge of the target data. Also, the \"test-time adaptation\" term has strong connotations of the online setting, which I believe is not being considered here, but should be distinguished clearly.\n- Fig 1 clarity: Overall, the figure doesn't say very much about the source-free setting (it just looks like an option, but pictorially it doesn't look different from the other options). Some picture-level aspects were unclear, like what the different between the icons in the gray (pre-training) box are, and what the image + 3 gray shapes in the top represent, as well as what the orange stripe in the 3 gray shapes in the adaptation phase denote.  \n- Table 1,2 seem to use different self-learning algorithms. How were the methods for each table chosen? In Section 6, the paper supports Robust pseudo labeling, but Table 2 uses ENT for the main results.\n- In Table 4, why is one comparison against TENT while the other against TTT? \n- The takeaway in Table 8 that updating all affine layers is important is perhaps misleading, since in Table 5, tuning all affine layers is the worst option. It is also not fully specified what tuning all affine layers means?\n- It may be easier in terms of presentation to talk about ImageNet-D right before the results in section 7.\n- The results on ImageNet-D so far do not seem unique to test-time adaptation with self-learning; they would seem to occur in other settings too. Are there any aspects of ImageNet-D that test aspects of test-time adaptation methods?\n- As the paper suggests ImageNet-D as a robustness benchmark, it's probably worth discussing whether we would expect our models to generalize to such disparate domains - for some of the domains, the paper reports close to 0% accuracy. ", "summary_of_the_review": "The paper tests self-learning as a complementary addition to improve robustness. However, the premise that self-learning improves robustness is already somewhat well-established - the main contribution here is a systematic application to different methods and datasets. The restriction to the pre-training + test-time adaptation setting has also been considered to some extent, but not as systematically. The value of the proposed dataset ImageNet-D is unclear, whether it gives insights beyond DomainNet itself, and whether it is a worthwhile goal to generalize to such disparate domains. Finally, the analysis of self-learning dynamics seems interesting and predicts some empirical behaviors nicely. I think the paper could have a good message solidifying self-learning methods for robustness, but could use some tightening up in the story/clarity of the paper, and some inconsistencies in the experimental reporting. I'd be happy to raise my score if the issues are addressed in the rebuttal.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635835293197}, {"id": "RMnXMQq_BZM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3402/Reviewer_s85k"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides an in depth empirical evaluation of classical self-training techniques such as pseudo-labelling and entropy minimization on test performance under domain shifts. The authors stress that, although simple, these techniques consistently improve the robustness to distribution shifts regardless of model architecture or pre-training techniques used. This makes them especially useful to practitioners applying machine learning algorithms to real problems where distribution shifts are prevalent. The authors claim state-of-the-art adaptation results on a number of popular dataset corruption benchmarks, and present a new challenging dataset for evaluating the robustness of deep vision models.\n\n\n", "review_text": "Strengths:\n\n-The message of the paper is very clear, and it is generally well written.\n\n-I believe this paper is of interest to the research community, as improving\nthe robustness of deep vision models to distributional shifts is of critical importance going forward.\n\n-The empirical evaluation is thorough, significant and convincing, demonstrating \nthe advantages of using self-training on top of existing approaches to \nimprove the robustness of deep models on a variety of standard benchmarks.\n\nWeaknesses/Suggestions:\n\n-The insights from this paper are mainly empirical, and there is little\nalgorithmic/theoretical novelty. The self-training algorithms evaluated here\nare well established, even in a deep learning setting (see below). \nThe exception to this is section 8 which comes across as an attempt to address \nthis deficiency late on, nonetheless it is useful to know what temperatures \nought to be used in student/teacher training. All this is not to say that \nempirical demonstrations are less important, but from an algorithmic perspective \nit makes the results not as surprising. At the risk of stating the obvious, \nsince we're using self-training to adapt to new domains, we would expect the \nmodels to be more robust to them. With that said, demonstrating this \nconvincingly in practice is non-trivial and demands significant engineering\neffort as evidenced in this paper.\n\n-Because there are so many results, the paper is quite dense and a bit hard \nto follow at times since the text is broken up a lot. This isn't a major \nissue but something the authors could consider improving for the final \nversion.\n\n-Self-training using deep learning as presented here has gained popularity \nrecently, and the paper is missing some related references (see below to name a few).\n\n-I think including some results on model calibration would add a lot\nof value to the exposition, especially if the authors could demonstrate that \nself-learning also calibrates predictions, thereby improving not only accuracy \nbut also uncertainty under domain shift.\n\n-Reporting standard deviations of results would strengthen the author's claims\n\n-------------------\n\nReferences:\n\n[1] Zou, Yang, et al. \"Unsupervised domain adaptation for semantic segmentation via class-balanced self-training.\" Proceedings of the European conference on computer vision (ECCV). 2018.\n\n[2] Rizve, Mamshad Nayeem, et al. \"In Defense of Pseudo-Labeling: An Uncertainty-Aware Pseudo-label Selection Framework for Semi-Supervised Learning.\" International Conference on Learning Representations. 2020.\n\n[3] De Sousa Ribeiro, Fabio, et al. \"Deep bayesian self-training.\" Neural Computing and Applications 32.9 (2020): 4275-4291.\n\n[4] Zou, Yang, et al. \"Confidence regularized self-training.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019.\n\n[5] Zoph, Barret, et al. \"Rethinking Pre-training and Self-training.\" Advances in Neural Information Processing Systems 33 (2020).\n\n[6] Mukherjee, Subhabrata, and Ahmed Awadallah. \"Uncertainty-aware self-training for few-shot text classification.\" Advances in Neural Information Processing Systems 33 (2020).\n\n[7] Wei, Colin, et al. \"Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data.\" International Conference on Learning Representations. 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides an in depth empirical evaluation of classical self-training techniques such as pseudo-labelling and entropy minimization on test performance under domain shifts. The authors stress that, although simple, these techniques consistently improve the robustness to distribution shifts regardless of model architecture or pre-training techniques used. This makes them especially useful to practitioners applying machine learning algorithms to real problems where distribution shifts are prevalent. The authors claim state-of-the-art adaptation results on a number of popular dataset corruption benchmarks, and present a new challenging dataset for evaluating the robustness of deep vision models.\n\n\n", "main_review": "Strengths:\n\n-The message of the paper is very clear, and it is generally well written.\n\n-I believe this paper is of interest to the research community, as improving\nthe robustness of deep vision models to distributional shifts is of critical importance going forward.\n\n-The empirical evaluation is thorough, significant and convincing, demonstrating \nthe advantages of using self-training on top of existing approaches to \nimprove the robustness of deep models on a variety of standard benchmarks.\n\nWeaknesses/Suggestions:\n\n-The insights from this paper are mainly empirical, and there is little\nalgorithmic/theoretical novelty. The self-training algorithms evaluated here\nare well established, even in a deep learning setting (see below). \nThe exception to this is section 8 which comes across as an attempt to address \nthis deficiency late on, nonetheless it is useful to know what temperatures \nought to be used in student/teacher training. All this is not to say that \nempirical demonstrations are less important, but from an algorithmic perspective \nit makes the results not as surprising. At the risk of stating the obvious, \nsince we're using self-training to adapt to new domains, we would expect the \nmodels to be more robust to them. With that said, demonstrating this \nconvincingly in practice is non-trivial and demands significant engineering\neffort as evidenced in this paper.\n\n-Because there are so many results, the paper is quite dense and a bit hard \nto follow at times since the text is broken up a lot. This isn't a major \nissue but something the authors could consider improving for the final \nversion.\n\n-Self-training using deep learning as presented here has gained popularity \nrecently, and the paper is missing some related references (see below to name a few).\n\n-I think including some results on model calibration would add a lot\nof value to the exposition, especially if the authors could demonstrate that \nself-learning also calibrates predictions, thereby improving not only accuracy \nbut also uncertainty under domain shift.\n\n-Reporting standard deviations of results would strengthen the author's claims\n\n-------------------\n\nReferences:\n\n[1] Zou, Yang, et al. \"Unsupervised domain adaptation for semantic segmentation via class-balanced self-training.\" Proceedings of the European conference on computer vision (ECCV). 2018.\n\n[2] Rizve, Mamshad Nayeem, et al. \"In Defense of Pseudo-Labeling: An Uncertainty-Aware Pseudo-label Selection Framework for Semi-Supervised Learning.\" International Conference on Learning Representations. 2020.\n\n[3] De Sousa Ribeiro, Fabio, et al. \"Deep bayesian self-training.\" Neural Computing and Applications 32.9 (2020): 4275-4291.\n\n[4] Zou, Yang, et al. \"Confidence regularized self-training.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019.\n\n[5] Zoph, Barret, et al. \"Rethinking Pre-training and Self-training.\" Advances in Neural Information Processing Systems 33 (2020).\n\n[6] Mukherjee, Subhabrata, and Ahmed Awadallah. \"Uncertainty-aware self-training for few-shot text classification.\" Advances in Neural Information Processing Systems 33 (2020).\n\n[7] Wei, Colin, et al. \"Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data.\" International Conference on Learning Representations. 2020.", "summary_of_the_review": "Overall I like the paper. I think it would bring value to the research community and it serves as a reminder that the simplest methods often work very well in practice.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635695453337}], "openreview_url": "https://openreview.net/forum?id=1oEvY1a67c1", "arxiv_id": "2104.12928", "paper_pdf": "papers/1oEvY1a67c1.pdf", "paper_pdf_sha256": "784e52fd47b843a3bef9af8daff038edf7e39e9df1e1379911ddb9bdb81e42d3", "paper_pdf_bytes": 1652815, "paper_pdf_source": "openreview", "code_url": "https://github.com/bethgelab/robustness", "code_repository": "bethgelab/robustness", "code_commit": "0ef82b178e3526f63d8017dc870072d611d883e9", "code_archive": "repos/1oEvY1a67c1.zip", "code_archive_sha256": "19cdd6297102dc21c81b704b16b852d4a76b2be403853f862bdd47661a5408d1", "code_archive_bytes": 478310, "code_file_count": 39, "code_extensions": {".py": 33, ".sh": 5, ".ipynb": 1}, "github_disk_usage_kb": 512, "github_languages": {"Python": 110845, "Shell": 93}, "github_archived": false, "github_pushed_at": "2023-07-05T23:11:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adapting-imagenet-scale-models-to-complex"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "86t2GlfzFo", "year": 2021, "status": "rejected", "title": "Deep Curvature Suite", "authors": ["Diego Granziol", "Xingchen Wan", "Timur Garipov"], "authorids": ["~Diego_Granziol1", "~Xingchen_Wan1", "~Timur_Garipov1"], "authors_source": "OpenReview API", "abstract": "The curvature of the loss, provides rich information on the geometry underlying neural networks, with applications in second order optimisation and Bayesian deep learning. However, accessing curvature information is still a daunting engineering challenge, inaccessible to most practitioners. We hence provide a software package the \\textbf{Deep Curvature Suite}, which allows easy curvature evaluation for large modern neural networks. Beyond the calculation of a highly accurate moment matched approximation of the Hessian spectrum using Lanczos, our package provides: extensive \\emph{loss surface visualisation}, the calculation of the \\emph{Hessian variance} and \\emph{stochastic second order optimisers}. We further address and disprove many common misconceptions in the literature about the Lanczos algorithm, namely that it learns eigenvalues from the top down. We prove using high dimensional concentration inequalities that for specific matrices a single random vector is sufficient for accurate spectral estimation, informing our spectral visualisation method. We showcase our package practical utility on a series of examples based on realistic modern neural networks such as the VGG-$16$ and Preactivated ResNets on the CIFAR-$10$/$100$ datasets. We further detail $3$ specific potential use cases enabled by our software: research in stochastic second order optimisation for deep learning, learning rate scheduling using known optimality formulae for convex surfaces and empirical verification of deep learning theory based on comparing empirical and theoretically implied spectra.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Q4YyW0Lz33Z", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper819/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n- This paper introduces a package for computing the second-order information of neural networks based on the Lanczos algorithm. The authors showcase the usages of the package with 1) visualizing the eigenspectrum of the curvature matrix; 2)visualizing the loss surface along with a specific direction, and 3) comparisons between different optimizers. Also, the authors claimed that they address several misconceptions about the Lanczos algorithm.\n\nOverall:\n\nThis paper is well-written and easy to follow. I believe that the developed package can be a good contribution to the machine learning community, especially for people working on second-order optimization and understanding the training dynamics and generalization performance of deep neural networks. However, I still have the following questions:\n\n-  Lanczos algorithm suffers from numerical instability. How do you deal with this issue? Can you provide more details about this?\n\n- I would suggest the authors move Section 4 to the appendix, considering the main scope of the paper is introducing a new machine learning package. In the meantime, the authors should highlight more on the difference between your package and the existing implementations. For example, you can add a table to summarize the features of all the related packages/implementations. Also, a table for comparing the running memory cost and running time cost will be very helpful.\n\nMinors:\n-  as well as the commonly used `Generalsed` Gauss Newton -> as well as the commonly used `Generalised` Gauss-Newton\n\n- The main interface functions are `organsed` as followed: -> The main interface functions are `organized` as following:\n\n- Krylov subspace K (H, v) = span{v,`H^v`,H2v...} is orthogonalised -> Krylov subspace K (H, v) = span{v,`Hv`,H2v...} is orthogonalised\n\n- the reference is missing in the main pdf.\n\n\nRating:\n- This paper did a good job of introducing the package with detailed examples. Also, this package will definitely ease the effort of researchers in related areas to compute the second-order information of the neural networks. In the meantime, I still have the concerns mentioned above. So, my rating is weak acceptance at the current stage. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well-written paper and can be a good contribution to the community.", "review": "Summary:\n- This paper introduces a package for computing the second-order information of neural networks based on the Lanczos algorithm. The authors showcase the usages of the package with 1) visualizing the eigenspectrum of the curvature matrix; 2)visualizing the loss surface along with a specific direction, and 3) comparisons between different optimizers. Also, the authors claimed that they address several misconceptions about the Lanczos algorithm.\n\nOverall:\n\nThis paper is well-written and easy to follow. I believe that the developed package can be a good contribution to the machine learning community, especially for people working on second-order optimization and understanding the training dynamics and generalization performance of deep neural networks. However, I still have the following questions:\n\n-  Lanczos algorithm suffers from numerical instability. How do you deal with this issue? Can you provide more details about this?\n\n- I would suggest the authors move Section 4 to the appendix, considering the main scope of the paper is introducing a new machine learning package. In the meantime, the authors should highlight more on the difference between your package and the existing implementations. For example, you can add a table to summarize the features of all the related packages/implementations. Also, a table for comparing the running memory cost and running time cost will be very helpful.\n\nMinors:\n-  as well as the commonly used `Generalsed` Gauss Newton -> as well as the commonly used `Generalised` Gauss-Newton\n\n- The main interface functions are `organsed` as followed: -> The main interface functions are `organized` as following:\n\n- Krylov subspace K (H, v) = span{v,`H^v`,H2v...} is orthogonalised -> Krylov subspace K (H, v) = span{v,`Hv`,H2v...} is orthogonalised\n\n- the reference is missing in the main pdf.\n\n\nRating:\n- This paper did a good job of introducing the package with detailed examples. Also, this package will definitely ease the effort of researchers in related areas to compute the second-order information of the neural networks. In the meantime, I still have the concerns mentioned above. So, my rating is weak acceptance at the current stage. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604706700420}, {"id": "65aJb_VQByr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper819/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary\nThe paper presents a library for second-order analysis of the optimization of models implemented with PyTorch and potentially having millions of parameters. The library offers tools for easily computing and visualizing Hessian, curvature, loss landscape, and runnning simple statistics, for instance for studying the properties of local minima.\nCompared to existing tools, the proposed library is more complete and accurate, , as demonstrated by example analyses, and scales well with the number of model parameters and the dataset size.\nThe library itself seems very useful, however the paper needs major revision in order to be publishable.\n\n### Presentation\nThe paper is generally poorly written. Many concepts are not defined or not well enough, for instance:\n1. input independence, page 2, by which it is not clear if independence between features or between samples is meant;\n2. number of moments, page 3, not defined;\n3. curvature, by which it should be specified earlier it is meant the loss curvature.\n\nTo improve clarity, the paper should provide a comprehensive list of tools implemented in the library and give examples of why one would need to use each of them. The library's tools are described throughout the paper, but they are too scattered to have a complete view of the library.\nMoreover, the Lanczos algorithm, which is the central for the library, should be reported. \n\nFinally, the paper misses important parts, such as the reference page, a paragraph in page 5 and the cited and not reported Algorithm 1; it also has many typos; notations and domains are not provided for most equations and it is not clear why Equation (3) does not depend on the loss L.\n\nThe current presentation doesn't meet the standards for a conference paper.\n\n### UPDATE\nThe presentation of the paper is now convincing, with the background, contributions and concepts clearly stated. I hence increased my score. \n\nHowever, I agree with reviewer 2 that the library is not properly tested (I understand it can be hard to test the whole computation, but unit tests could be easily provided) and that it would have a higher impact if it were more modular, so that a user could easily add the loss analysis directly in her workflow.\nMoreover, I also think that the paper should be rejected given that the first submission wasn't anonymized and the paper wasn't in a state of being submitted.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Useful library, paper needs major revision", "review": "### Summary\nThe paper presents a library for second-order analysis of the optimization of models implemented with PyTorch and potentially having millions of parameters. The library offers tools for easily computing and visualizing Hessian, curvature, loss landscape, and runnning simple statistics, for instance for studying the properties of local minima.\nCompared to existing tools, the proposed library is more complete and accurate, , as demonstrated by example analyses, and scales well with the number of model parameters and the dataset size.\nThe library itself seems very useful, however the paper needs major revision in order to be publishable.\n\n### Presentation\nThe paper is generally poorly written. Many concepts are not defined or not well enough, for instance:\n1. input independence, page 2, by which it is not clear if independence between features or between samples is meant;\n2. number of moments, page 3, not defined;\n3. curvature, by which it should be specified earlier it is meant the loss curvature.\n\nTo improve clarity, the paper should provide a comprehensive list of tools implemented in the library and give examples of why one would need to use each of them. The library's tools are described throughout the paper, but they are too scattered to have a complete view of the library.\nMoreover, the Lanczos algorithm, which is the central for the library, should be reported. \n\nFinally, the paper misses important parts, such as the reference page, a paragraph in page 5 and the cited and not reported Algorithm 1; it also has many typos; notations and domains are not provided for most equations and it is not clear why Equation (3) does not depend on the loss L.\n\nThe current presentation doesn't meet the standards for a conference paper.\n\n### UPDATE\nThe presentation of the paper is now convincing, with the background, contributions and concepts clearly stated. I hence increased my score. \n\nHowever, I agree with reviewer 2 that the library is not properly tested (I understand it can be hard to test the whole computation, but unit tests could be easily provided) and that it would have a higher impact if it were more modular, so that a user could easily add the loss analysis directly in her workflow.\nMoreover, I also think that the paper should be rejected given that the first submission wasn't anonymized and the paper wasn't in a state of being submitted.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604680287528}, {"id": "fteq0qfFfOg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper819/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "###\nSummary:\nThis paper proposes a software package to ease and provide a standard way for Hessian-related computation, both for loss analysis and second order optimization. It also provides analysis on why Lanczos algorithm is a better choice to estimate Hessian eigenvalue compared to Power Iterations. Finally, it empirically shows the advantage of using Hessian approximation that goes beyond diagonal approximation for spectrum computation.\n\n###\nReasons for score: \nI lean toward acceptance. Implementing Hessian-related computation is complex and often a bottleneck for research. Easing the implementation of those computations lower the entry barrier for second order optimization/loss analysis research and has the potential to stimulate more works in those areas.\n \n###\n Pros:\n- Address an important implementation problem\n- Scale to medium size network used on CIFAR-10.\n \n###\nCons: \n- Comparisons with related work could be expanded. How does this library related to the other software that use Lanczos algorithm for Hessian computation?\n- How extensive is the library? How easy is to use it with an arbitrary Pytorch model? \n- It is not clear if the library would scale to neural network usually used for larger scale problem such as ImageNet.\n \n###\nQuestions:\n- It is not clear to me how to compute the Hessian of a Neural Network with Batch Norm using minibatch statistics due to the dependency on the other samples of the batch. Could you elaborate on this point?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Deep Curvature Suite could potentially stimulate more research", "review": "###\nSummary:\nThis paper proposes a software package to ease and provide a standard way for Hessian-related computation, both for loss analysis and second order optimization. It also provides analysis on why Lanczos algorithm is a better choice to estimate Hessian eigenvalue compared to Power Iterations. Finally, it empirically shows the advantage of using Hessian approximation that goes beyond diagonal approximation for spectrum computation.\n\n###\nReasons for score: \nI lean toward acceptance. Implementing Hessian-related computation is complex and often a bottleneck for research. Easing the implementation of those computations lower the entry barrier for second order optimization/loss analysis research and has the potential to stimulate more works in those areas.\n \n###\n Pros:\n- Address an important implementation problem\n- Scale to medium size network used on CIFAR-10.\n \n###\nCons: \n- Comparisons with related work could be expanded. How does this library related to the other software that use Lanczos algorithm for Hessian computation?\n- How extensive is the library? How easy is to use it with an arbitrary Pytorch model? \n- It is not clear if the library would scale to neural network usually used for larger scale problem such as ImageNet.\n \n###\nQuestions:\n- It is not clear to me how to compute the Hessian of a Neural Network with Batch Norm using minibatch statistics due to the dependency on the other samples of the batch. Could you elaborate on this point?\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603894619303}, {"id": "i6G27-nNSCx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper819/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I think this paper needs to be desk-rejected since the supplementary material reveals the authors' identities in multiple places, e.g. the Readme file and the example notebook. Various .py files have an author attribute set.\n\nThis issue aside, both the package and the paper appear rushed to me and **not ready for publication**.\n\nPackage:\n* Most importantly, I could not find any tests whatsoever. Of course tests do not guarantee anything, especially if implemented poorly, but I hope that nobody would trust a completely untested third-party codebase for their own research. A thorough(!) test suite would be a strict requirement from my view to even consider accepting a code library paper at a top-tier conference.\n* There is no setup.py file to install the package and dependencies. I tried importing some of the modules from the root directory but got errors because of missing dependencies (e.g. backpack is not listed in the requirements.txt).\n* The package is split into multiples modules with generic names such as 'core'. I would strongly suggest to put them in a joint namespace.\n* What's the reasoning for including implementations of standard architectures that are available e.g. through torchvision and high-level training functions? Are those required? In general, I think that for a utility library it is best to assume that users have an established workflow of constructing and training their networks (unless the point of the library is to simplify that workflow, but this does not seem to be the case here).\n\nPaper:\n* I don't think that Section 4 fits into the flow of the paper. I would rather see that space used to go more into depth with the examples. At the moment you assume that the reader already knows why they would want to e.g. estimate the spectral density, but I think it would be worth both motivating the experiments and discussing the results.\n* I think basing the code examples on actual network and data loader objects would make them more accessible. Presumably passing string arguments for model and dataset into your compute_eigenspectrum and train_network functions is limited to the architectures and datasets that you implement? How do users provide their own datasets or networks? Or is this not supported? Section 2 seems to claim otherwise.\n* The font sizes and families around the code listings are extremely inconsistent, especially pages 5 and 6 are visually fairly unappealing.\n* There is some discussion around the relationship to Backpack, but what about PyHessian? That package seems more closely related since it also focuses on computing spectral densities.\n* Comparing first and second order optimisers on a per-epoch basis as in Figure 1 does not seem practically relevant to me, a wall clock time comparison would be of much higher interest to most potential users.\n* I would recommend carefully checking the paper for spelling and language. I'm not a native speaker, but to me it seems like there are quite a few inappropriate capitalisations (\"Training\", \"Testing\"), inconsistent capitalisations (\"eigenvalues\" vs \"Eigenvalues\") and hyphenations (\"eigen-decomposition\").\n* The reference pages are missing (after page 8).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unclear contribution; supplementary material not anonymized", "review": "I think this paper needs to be desk-rejected since the supplementary material reveals the authors' identities in multiple places, e.g. the Readme file and the example notebook. Various .py files have an author attribute set.\n\nThis issue aside, both the package and the paper appear rushed to me and **not ready for publication**.\n\nPackage:\n* Most importantly, I could not find any tests whatsoever. Of course tests do not guarantee anything, especially if implemented poorly, but I hope that nobody would trust a completely untested third-party codebase for their own research. A thorough(!) test suite would be a strict requirement from my view to even consider accepting a code library paper at a top-tier conference.\n* There is no setup.py file to install the package and dependencies. I tried importing some of the modules from the root directory but got errors because of missing dependencies (e.g. backpack is not listed in the requirements.txt).\n* The package is split into multiples modules with generic names such as 'core'. I would strongly suggest to put them in a joint namespace.\n* What's the reasoning for including implementations of standard architectures that are available e.g. through torchvision and high-level training functions? Are those required? In general, I think that for a utility library it is best to assume that users have an established workflow of constructing and training their networks (unless the point of the library is to simplify that workflow, but this does not seem to be the case here).\n\nPaper:\n* I don't think that Section 4 fits into the flow of the paper. I would rather see that space used to go more into depth with the examples. At the moment you assume that the reader already knows why they would want to e.g. estimate the spectral density, but I think it would be worth both motivating the experiments and discussing the results.\n* I think basing the code examples on actual network and data loader objects would make them more accessible. Presumably passing string arguments for model and dataset into your compute_eigenspectrum and train_network functions is limited to the architectures and datasets that you implement? How do users provide their own datasets or networks? Or is this not supported? Section 2 seems to claim otherwise.\n* The font sizes and families around the code listings are extremely inconsistent, especially pages 5 and 6 are visually fairly unappealing.\n* There is some discussion around the relationship to Backpack, but what about PyHessian? That package seems more closely related since it also focuses on computing spectral densities.\n* Comparing first and second order optimisers on a per-epoch basis as in Figure 1 does not seem practically relevant to me, a wall clock time comparison would be of much higher interest to most potential users.\n* I would recommend carefully checking the paper for spelling and language. I'm not a native speaker, but to me it seems like there are quite a few inappropriate capitalisations (\"Training\", \"Testing\"), inconsistent capitalisations (\"eigenvalues\" vs \"Eigenvalues\") and hyphenations (\"eigen-decomposition\").\n* The reference pages are missing (after page 8).", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603889468622}], "openreview_url": "https://openreview.net/forum?id=86t2GlfzFo", "arxiv_id": "1912.09656", "paper_pdf": "papers/86t2GlfzFo.pdf", "paper_pdf_sha256": "5905b75bc1164f7cf5d74000d5e0220be0e5c6ae1d75693661645c283d7017b6", "paper_pdf_bytes": 727277, "paper_pdf_source": "openreview", "code_url": "https://github.com/xingchenwan/MLRG_DeepCurvature", "code_repository": "xingchenwan/MLRG_DeepCurvature", "code_commit": "818ea8631e7aafc659035310c45270145fd174d3", "code_archive": "repos/86t2GlfzFo.zip", "code_archive_sha256": "0d1f3286d747db87256526fa6df002840a2007116d4206a7e8149630a938a45f", "code_archive_bytes": 343504, "code_file_count": 37, "code_extensions": {".py": 36, ".ipynb": 1}, "github_disk_usage_kb": 606, "github_languages": {"Jupyter Notebook": 375459, "Python": 199652}, "github_archived": false, "github_pushed_at": "2020-04-19T09:34:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mlrg-deep-curvature"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1e0ZlHYDB", "year": 2020, "status": "rejected", "title": "Progressive Compressed Records: Taking a Byte Out of Deep Learning Data", "authors": ["Michael Kuchnik", "George Amvrosiadis", "Virginia Smith"], "authorids": ["mkuchnik@andrew.cmu.edu", "gamvrosi@cmu.edu", "smithv@cmu.edu"], "authors_source": "OpenReview API", "abstract": "Deep learning training accesses vast amounts of data at high velocity, posing challenges for datasets retrieved over commodity networks and storage devices. We introduce a way to dynamically reduce the overhead of fetching and transporting training data with a method we term Progressive Compressed Records (PCRs). PCRs deviate from previous formats by leveraging progressive compression to split each training example into multiple examples of increasingly higher fidelity, without adding to the total data size. Training examples of similar fidelity are grouped together, which reduces both the system overhead and data bandwidth needed to train a model. We show that models can be trained on aggressively compressed representations of the training data and still retain high accuracy, and that PCRs can enable a 2x speedup on average over baseline formats using JPEG compression. Our results hold across deep learning architectures for a wide range of datasets: ImageNet, HAM10000, Stanford Cars, and CelebA-HQ.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "r1ehxrXZoB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2156/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper demonstrates an interesting application of progressive compression to reduce the disk I/O overhead of training deep neural networks. The format encodes the trade-off between data fidelity and I/O bandwidth demand naturally, which could be useful when I/O is the bottleneck.\n\nMy major concern is that the paper should be clearer about the setting.\n* Does your work target the case where data cannot fit in RAM and should be fetched from local disk or through network? However, the datasets used in the evaluation look small and could fit in RAM.\n* How are mini-batches created? You mentioned in the related work that previous work (Kurth et al., 2018) lets each worker sample from a local subset instead of performing a true sampling of the whole dataset. Does your work perform true sampling? How much benefit does it give?\n* Is disk I/O really a bottleneck in training? There are many evidence [1][2][3] of almost linear scalability in training ResNet on *full* imagenet across hundreds or even thousands of GPUs. These work focus heavily on network communication rather than disk I/O. Does your setting differ from theirs? How does your approach compare with their techniques for optimizing disk I/O?\n\nThat being said, I think this approach should be appealing when the I/O bandwidth is limited and dynamic. Examples include training on edge devices, or federated training where data needs be fetched via ad-hoc network.\n\nOther detailed comments:\n\n* Figure 1 is not very informative and quite puzzling. There is no definition of quality at that point.\n* Sec 2 paragraph 3. What is the issue of data augmentation with the standard JPEG compression? Does your compression ease data augmentation?\n* Sec 3.1 paragraph 1. \"This is turn enables ...\" -> \"This in turn enables ...\"\n* How to decide the number of scans? Does it have impact on the I/O efficiency?\n* Evaluation\n  - I'm not familiar with Ceph. Why choose this particular environment? Does it bring in extra overhead (e.g., communicating with metadata server). What does the network topology look like? Is the data loading stall (figure 7) due to network congestion?\n - It worth evaluating more tasks such as detection and segmentation to measure the impact of compression.\n\n\n[1] Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, Goyal et al.\n[2] Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash, Mikami et al.\n[3] Image Classification at Supercomputer Scale, Ying et al.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #4", "review": "The paper demonstrates an interesting application of progressive compression to reduce the disk I/O overhead of training deep neural networks. The format encodes the trade-off between data fidelity and I/O bandwidth demand naturally, which could be useful when I/O is the bottleneck.\n\nMy major concern is that the paper should be clearer about the setting.\n* Does your work target the case where data cannot fit in RAM and should be fetched from local disk or through network? However, the datasets used in the evaluation look small and could fit in RAM.\n* How are mini-batches created? You mentioned in the related work that previous work (Kurth et al., 2018) lets each worker sample from a local subset instead of performing a true sampling of the whole dataset. Does your work perform true sampling? How much benefit does it give?\n* Is disk I/O really a bottleneck in training? There are many evidence [1][2][3] of almost linear scalability in training ResNet on *full* imagenet across hundreds or even thousands of GPUs. These work focus heavily on network communication rather than disk I/O. Does your setting differ from theirs? How does your approach compare with their techniques for optimizing disk I/O?\n\nThat being said, I think this approach should be appealing when the I/O bandwidth is limited and dynamic. Examples include training on edge devices, or federated training where data needs be fetched via ad-hoc network.\n\nOther detailed comments:\n\n* Figure 1 is not very informative and quite puzzling. There is no definition of quality at that point.\n* Sec 2 paragraph 3. What is the issue of data augmentation with the standard JPEG compression? Does your compression ease data augmentation?\n* Sec 3.1 paragraph 1. \"This is turn enables ...\" -> \"This in turn enables ...\"\n* How to decide the number of scans? Does it have impact on the I/O efficiency?\n* Evaluation\n  - I'm not familiar with Ceph. Why choose this particular environment? Does it bring in extra overhead (e.g., communicating with metadata server). What does the network topology look like? Is the data loading stall (figure 7) due to network congestion?\n - It worth evaluating more tasks such as detection and segmentation to measure the impact of compression.\n\n\n[1] Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, Goyal et al.\n[2] Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash, Mikami et al.\n[3] Image Classification at Supercomputer Scale, Ying et al.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573102836502}, {"id": "BkeXwO--sB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2156/AnonReviewer5"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces Progressive Compressed Records (PCR) which is an on-disk format for fetching and transporting training data in an attempt to reduce the overhead storage bandwidth for training large scale deep neural networks. This is a well written paper that includes all the required background and related works, as well as an easy-to-understand example that runs through the manuscript, explaining what the reader needs to know in order to appreciate the work. The empirical results of several experiments show that the PCR requires up to two times less storage bandwidth while retaining model accuracy.\n\nMy only concern is that although the related work section provides a thorough survey of the current methods in the literature, the authors did not demonstrate the performance of state-of-the-art and compare their performance with them. I believe this is necessary to truly validate the superiority of their method over state-of-the-art.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #5", "review": "This paper introduces Progressive Compressed Records (PCR) which is an on-disk format for fetching and transporting training data in an attempt to reduce the overhead storage bandwidth for training large scale deep neural networks. This is a well written paper that includes all the required background and related works, as well as an easy-to-understand example that runs through the manuscript, explaining what the reader needs to know in order to appreciate the work. The empirical results of several experiments show that the PCR requires up to two times less storage bandwidth while retaining model accuracy.\n\nMy only concern is that although the related work section provides a thorough survey of the current methods in the literature, the authors did not demonstrate the performance of state-of-the-art and compare their performance with them. I believe this is necessary to truly validate the superiority of their method over state-of-the-art.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1573095514835}, {"id": "SJluQZ2kor", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2156/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes using progressive encoding of images and re-arrange of data blocks in images to improve reading speed and therefore training speed.\n\nTo fully analyze the maximum possible speed of training, it would be great to the measure upper bound of images/sec, when avoiding reading from disk and just using images from memory. \n\nDecoding a typical progressive JPEG image usually takes about 2-3 times as much time as decoding a non-progressive JPEG, for full resolution, analyzing the time to read vs time to decode the images would be great. It is not clear how changing the number of total groups would affect the image size and the reading speed.\n\nBased on the current experiments it is not clear what is the impact of the batch size when creating PCRs and when reading the image blocks, or the impact of the batch size on the training speed.\n\nFigure 3 is really hard to read and compare times to convergence, authors should provide a table with times to X% accuracy.  Although time to convergence is the key metric, it would be great to know the difference in images/sec of different settings.\n\nUsing ImageNet 100 classes (not clear how the 100 classes were chosen) instead of the usual 1000 classes, can distort the results, since it is not clear if higher resolution would be needed to distinguish more classes or not.\n\nHave the authors considered other image compression formats like WebP? How tie is the proposed record encoding with the image compression?  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "The paper proposes using progressive encoding of images and re-arrange of data blocks in images to improve reading speed and therefore training speed.\n\nTo fully analyze the maximum possible speed of training, it would be great to the measure upper bound of images/sec, when avoiding reading from disk and just using images from memory. \n\nDecoding a typical progressive JPEG image usually takes about 2-3 times as much time as decoding a non-progressive JPEG, for full resolution, analyzing the time to read vs time to decode the images would be great. It is not clear how changing the number of total groups would affect the image size and the reading speed.\n\nBased on the current experiments it is not clear what is the impact of the batch size when creating PCRs and when reading the image blocks, or the impact of the batch size on the training speed.\n\nFigure 3 is really hard to read and compare times to convergence, authors should provide a table with times to X% accuracy.  Although time to convergence is the key metric, it would be great to know the difference in images/sec of different settings.\n\nUsing ImageNet 100 classes (not clear how the 100 classes were chosen) instead of the usual 1000 classes, can distort the results, since it is not clear if higher resolution would be needed to distinguish more classes or not.\n\nHave the authors considered other image compression formats like WebP? How tie is the proposed record encoding with the image compression?  ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573007648375}, {"id": "rkeMVzZJ5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2156/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper introduces a new storage format for image datasets for machine learning training. The core idea is to use progressive JPEG to create sequential scans of the input image, from lower resolution to higher resolution. The authors found that on some datasets, using half of the scans is already enough to reach similar accuracy but speeded up the convergence by a factor of 2.\n\nDetailed feedbacks:\n-\tThe paper presents a simple idea that directly uses the nature of JPEG compression. The paper shows that it can work well and can be potentially integrated into real machine learning dataset storage applications.\n-\tRelated work section is thorough.\n-\tThe experiments are limited to image classifications, and some of the datasets are subsampled (e.g. ImageNet and CelebA). This may not well represent real machine learning tasks, and practitioners may be unsure about the reliability of the compression. The “Cars” dataset contains fine-grained classification, in which the proposed method is\n-\tFigure 1 is not very clear what is the key advantage of the proposed method, and what are the different mechanisms.\n-\tAlternatively, one can subsample the pixels and store incremental subsets of those pixels. It would be good if the paper can discuss about this baseline.\n-\tThe data storage format is only loosely related to the main goal of the paper, which is to show that network can still train very well and even faster when receiving partial input data. Once they figured out the number of scans needed for this application, they don’t necessarily need to keep a full lossless version and can just go for a lossy version. In other words, the experiment section can be replaced by any other lossy compression by varying the compression ratio.\n-\tIn my opinion, there could be two reasons for faster convergence. 1) lowered image quality makes the data easier to learn and 2) the smaller data size allows faster reading of data from disk. The paper only shows wall-clock speed-up, but it is unclear which factor is bigger. 2) can be potentially addressed by faster disk reading such as SSD or in-memory datasets. One of the motivations is to help parallel training of dataset and it is also mentioned how non-random sampling of data can hurt training performance. It would be good to showcase how the proposed method can help in those parallel training settings. \n\nConclusion: This paper presents a simple and effective idea and can be potentially beneficial. However, my main concern is whether the experiments can be representative enough for large scale experiments (e.g. using non-subsampled ImageNet dataset with parallel training using SSD storage). Therefore, my overall rating is weak accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "Summary: This paper introduces a new storage format for image datasets for machine learning training. The core idea is to use progressive JPEG to create sequential scans of the input image, from lower resolution to higher resolution. The authors found that on some datasets, using half of the scans is already enough to reach similar accuracy but speeded up the convergence by a factor of 2.\n\nDetailed feedbacks:\n-\tThe paper presents a simple idea that directly uses the nature of JPEG compression. The paper shows that it can work well and can be potentially integrated into real machine learning dataset storage applications.\n-\tRelated work section is thorough.\n-\tThe experiments are limited to image classifications, and some of the datasets are subsampled (e.g. ImageNet and CelebA). This may not well represent real machine learning tasks, and practitioners may be unsure about the reliability of the compression. The “Cars” dataset contains fine-grained classification, in which the proposed method is\n-\tFigure 1 is not very clear what is the key advantage of the proposed method, and what are the different mechanisms.\n-\tAlternatively, one can subsample the pixels and store incremental subsets of those pixels. It would be good if the paper can discuss about this baseline.\n-\tThe data storage format is only loosely related to the main goal of the paper, which is to show that network can still train very well and even faster when receiving partial input data. Once they figured out the number of scans needed for this application, they don’t necessarily need to keep a full lossless version and can just go for a lossy version. In other words, the experiment section can be replaced by any other lossy compression by varying the compression ratio.\n-\tIn my opinion, there could be two reasons for faster convergence. 1) lowered image quality makes the data easier to learn and 2) the smaller data size allows faster reading of data from disk. The paper only shows wall-clock speed-up, but it is unclear which factor is bigger. 2) can be potentially addressed by faster disk reading such as SSD or in-memory datasets. One of the motivations is to help parallel training of dataset and it is also mentioned how non-random sampling of data can hurt training performance. It would be good to showcase how the proposed method can help in those parallel training settings. \n\nConclusion: This paper presents a simple and effective idea and can be potentially beneficial. However, my main concern is whether the experiments can be representative enough for large scale experiments (e.g. using non-subsampled ImageNet dataset with parallel training using SSD storage). Therefore, my overall rating is weak accept."}, "tcdate": 1571914282313}], "openreview_url": "https://openreview.net/forum?id=S1e0ZlHYDB", "arxiv_id": "1911.00472", "paper_pdf": "papers/S1e0ZlHYDB.pdf", "paper_pdf_sha256": "fc3e6ee90a925f417ef46b5b71db7e857923498ed90acb811ba1da2a47cfd0e3", "paper_pdf_bytes": 3943379, "paper_pdf_source": "openreview", "code_url": "https://github.com/mkuchnik/PCR_Release", "code_repository": "mkuchnik/PCR_Release", "code_commit": "aa76741f5dc6dcb6d7676c62e2fd53a78f43f142", "code_archive": "repos/S1e0ZlHYDB.zip", "code_archive_sha256": "adbf803296d6dcbc860e2d92039251c6d99429ed10637ebbbb11d9fac14bc829", "code_archive_bytes": 1365478, "code_file_count": 42, "code_extensions": {".py": 27, ".sh": 8, ".cpp": 3, ".c": 3, ".h": 1}, "github_disk_usage_kb": 1311, "github_languages": {"Python": 249111, "C": 20669, "Shell": 9203, "C++": 8497, "CMake": 2487, "Dockerfile": 1308}, "github_archived": false, "github_pushed_at": "2021-08-20T00:57:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/progressive-compressed-records-taking-a-byte-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8gdfWRilR7", "year": 2026, "status": "rejected", "title": "TimeSeriesGym: A Scalable Benchmark for (Time Series) Machine Learning Engineering Agents", "authors": ["Yifu Cai", "Xinyu Li", "Mononito Goswami", "Michał Wiliński", "Gus Welter", "Artur Dubrawski"], "authorids": ["~Yifu_Cai1", "~Xinyu_Li7", "~Mononito_Goswami1", "~Michał_Wiliński1", "~Gus_Welter1", "~Artur_Dubrawski2"], "authors_source": "OpenReview API", "abstract": "We introduce TimeSeriesGym, a scalable benchmarking framework for evaluating Artificial Intelligence (AI) agents on time series machine learning engineering challenges. Existing benchmarks lack scalability, focus narrowly on model building in well-defined settings, and evaluate only a limited set of research artifacts (e.g., CSV submission files). To make AI agent benchmarking more relevant to the practice of machine learning engineering, our framework scales along two critical dimensions. First, recognizing that effective ML engineering requires a range of diverse skills, TimeSeriesGym incorporates challenges from diverse sources spanning multiple domains and tasks. We design challenges to evaluate both isolated capabilities (including data handling, understanding research repositories, and code translation) and their combinations, and rather than addressing each challenge independently, we develop tools that support designing multiple challenges at scale. Second, we implement evaluation mechanisms for multiple research artifacts, including submission files, code, and models, using precise numeric measures and _optionally_ LLM-based qualitative assessments. This strategy complements objective evaluation with subjective assessment when appropriate. Although our initial focus is on time series applications, our framework can be readily extended to other data modalities, broadly enhancing the comprehensiveness and practical utility of agentic AI evaluation. We [open-source](https://anonymous.4open.science/r/TimeSeriesGym-9CF6/) our benchmarking framework to facilitate future research on the ML engineering capabilities of AI agents.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "UF2tyjb7Kw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13403/Reviewer_c71P"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The authors introduce TimeSeriesGym, a new benchmarking framework designed to evaluate AI agents on machine learning (ML) engineering tasks specifically for the time series domain. The paper argues that existing benchmarks are flawed, focusing too narrowly on well-defined Kaggle-style problems, lacking scalability, and only evaluating final submission files. TimeSeriesGym aims to solve this by sourcing diverse challenges (including 20 \"Original\" tasks based on real-world engineering like code migration and hyperparameter tuning), designing a framework to grade multiple artifacts (code, models, and submissions), and employing a holistic evaluation that includes both quantitative metrics and qualitative \"LLM-as-a-judge\" assessments. The framework also claims to be scalable with tools for new challenge creation. The paper's main experimental finding is that current state-of-the-art agents perform very poorly on these tasks, often failing to produce even a \"reasonable\" submission.", "review_text": "The authors introduce TimeSeriesGym, a new benchmarking framework designed to evaluate AI agents on machine learning (ML) engineering tasks specifically for the time series domain. The paper argues that existing benchmarks are flawed, focusing too narrowly on well-defined Kaggle-style problems, lacking scalability, and only evaluating final submission files. TimeSeriesGym aims to solve this by sourcing diverse challenges (including 20 \"Original\" tasks based on real-world engineering like code migration and hyperparameter tuning), designing a framework to grade multiple artifacts (code, models, and submissions), and employing a holistic evaluation that includes both quantitative metrics and qualitative \"LLM-as-a-judge\" assessments. The framework also claims to be scalable with tools for new challenge creation. The paper's main experimental finding is that current state-of-the-art agents perform very poorly on these tasks, often failing to produce even a \"reasonable\" submission.", "strengths": "Identifies a Clear and Important Gap: The paper is correct that time series is an underserved domain in agentic benchmarking. Creating a dedicated benchmark for this is a valuable contribution.\n\nFocuses on \"ML Engineering,\" Not Just Modeling: The strongest part of the paper is its inclusion of \"TimeSeriesGym Originals\" (Table 4). Challenges like \"Convert ResNet TensorFlow implementation to PyTorch\" or \"Improve PTB-XL ECG Classification Code\" are excellent, real-world tasks that go beyond the standard \"train-a-model-on-a-CSV\" format.\n\nHolistic Evaluation Concept: The ambition to evaluate multiple artifacts (code, models) and use a \"two-faceted grading approach\" (quantitative and qualitative, see Appendix E) is the right direction for a comprehensive benchmark.\n\nScalability as a Design Goal: Designing the benchmark to be scalable from the start (even if not fully proven) is a smart, forward-thinking approach that addresses a major flaw in static benchmarks.", "weaknesses": "**Critically Flawed Evaluation of Kaggle Challenges:**\n\nThe paper's primary evaluation metrics, \"Valid Submission (%)\" and \"Reasonable Submission (%),\" are insufficient.\n\n- The benchmark fails to report the actual leaderboard scores or ranks for the 13 included Kaggle competitions. This is a significant omission, as these are the most standardized, competitive tasks in the dataset.\n\n- The bar for a \"Reasonable Submission\" is set at scoring \"above median on the competition's public leaderboard\" (Section 4). This threshold is exceptionally low and provides no real insight into an agent's capabilities. An agent that barely surpasses the median is treated identically to one that achieves a state-of-the-art, gold-medal-winning score.\n\n- This lack of granular, comparable results (as seen in Table 10) makes it impossible for the research community to evaluate an agent's true performance or benchmark progress against established human or SOTA baselines on these problems.\n\n**Conceded Risk of Data Leakage:**\nThe paper's reliance on popular, public Kaggle competitions fundamentally compromises its ability to evaluate modern LLM-based agents. The authors' defense—that agents \"performed poorly anyway\" or that the benchmark is \"scalable\" to add new tasks—is unconvincing. It effectively concedes that the current Kaggle-based portion of the benchmark is unsuitable for reliably evaluating frontier models, as performance may be confounded by memorization.\n\nSubjective and Non-Scalable Evaluation Protocol:\n\n- The paper's claim of being a \"scalable benchmark\" is directly undermined by its own evaluation methodology for the \"Originals\" challenges.\n\n- For non-Kaggle tasks, \"reasonableness\" is determined by the authors \"manually inspecting\" if a \"genuine modeling attempt\" was made (Section 4, Metrics). A benchmark that requires subjective, manual intervention from its creators for a primary metric is, by definition, not scalable or objective.\n\n- The proposal to use \"LLM-as-a-judge\" (Appendix E) as a secondary evaluation method introduces a notoriously unreliable, biased, and difficult-to-reproduce component, which is not a substitute for rigorous, objective, quantitative metrics.\n\n- The authors themselves admit the protocol's flaws in Section 5 (\"Defining and measuring success\"), stating that \"our current evaluation approaches have inherent limitations.\"\n\n**Miscalibrated Difficulty (Floor Effects):**\n\n- The reported results are so poor that the benchmark, in its current form, largely fails as a diagnostic tool.\n\n- The best agent on the full benchmark (AIDE + GPT-4.1) only achieved a \"reasonable\" submission 12.1% of the time (Section 4.1).\n\n- Even on the hand-picked, \"easy\" Lite benchmark, the best agent (AIDE + Claude 3.7) only succeeded 38.9% of the time (Table 2).\n\nThese results indicate a significant floor effect. The tasks are currently too difficult for SOTA agents, preventing any meaningful differentiation between models or scaffolds. The benchmark primarily demonstrates that all agents fail, offering little insight into why they fail or which approaches are more promising.", "questions": "See the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce TimeSeriesGym, a new benchmarking framework designed to evaluate AI agents on machine learning (ML) engineering tasks specifically for the time series domain. The paper argues that existing benchmarks are flawed, focusing too narrowly on well-defined Kaggle-style problems, lacking scalability, and only evaluating final submission files. TimeSeriesGym aims to solve this by sourcing diverse challenges (including 20 \"Original\" tasks based on real-world engineering like code migration and hyperparameter tuning), designing a framework to grade multiple artifacts (code, models, and submissions), and employing a holistic evaluation that includes both quantitative metrics and qualitative \"LLM-as-a-judge\" assessments. The framework also claims to be scalable with tools for new challenge creation. The paper's main experimental finding is that current state-of-the-art agents perform very poorly on these tasks, often failing to produce even a \"reasonable\" submission.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "Identifies a Clear and Important Gap: The paper is correct that time series is an underserved domain in agentic benchmarking. Creating a dedicated benchmark for this is a valuable contribution.\n\nFocuses on \"ML Engineering,\" Not Just Modeling: The strongest part of the paper is its inclusion of \"TimeSeriesGym Originals\" (Table 4). Challenges like \"Convert ResNet TensorFlow implementation to PyTorch\" or \"Improve PTB-XL ECG Classification Code\" are excellent, real-world tasks that go beyond the standard \"train-a-model-on-a-CSV\" format.\n\nHolistic Evaluation Concept: The ambition to evaluate multiple artifacts (code, models) and use a \"two-faceted grading approach\" (quantitative and qualitative, see Appendix E) is the right direction for a comprehensive benchmark.\n\nScalability as a Design Goal: Designing the benchmark to be scalable from the start (even if not fully proven) is a smart, forward-thinking approach that addresses a major flaw in static benchmarks.", "weaknesses": "**Critically Flawed Evaluation of Kaggle Challenges:**\n\nThe paper's primary evaluation metrics, \"Valid Submission (%)\" and \"Reasonable Submission (%),\" are insufficient.\n\n- The benchmark fails to report the actual leaderboard scores or ranks for the 13 included Kaggle competitions. This is a significant omission, as these are the most standardized, competitive tasks in the dataset.\n\n- The bar for a \"Reasonable Submission\" is set at scoring \"above median on the competition's public leaderboard\" (Section 4). This threshold is exceptionally low and provides no real insight into an agent's capabilities. An agent that barely surpasses the median is treated identically to one that achieves a state-of-the-art, gold-medal-winning score.\n\n- This lack of granular, comparable results (as seen in Table 10) makes it impossible for the research community to evaluate an agent's true performance or benchmark progress against established human or SOTA baselines on these problems.\n\n**Conceded Risk of Data Leakage:**\nThe paper's reliance on popular, public Kaggle competitions fundamentally compromises its ability to evaluate modern LLM-based agents. The authors' defense—that agents \"performed poorly anyway\" or that the benchmark is \"scalable\" to add new tasks—is unconvincing. It effectively concedes that the current Kaggle-based portion of the benchmark is unsuitable for reliably evaluating frontier models, as performance may be confounded by memorization.\n\nSubjective and Non-Scalable Evaluation Protocol:\n\n- The paper's claim of being a \"scalable benchmark\" is directly undermined by its own evaluation methodology for the \"Originals\" challenges.\n\n- For non-Kaggle tasks, \"reasonableness\" is determined by the authors \"manually inspecting\" if a \"genuine modeling attempt\" was made (Section 4, Metrics). A benchmark that requires subjective, manual intervention from its creators for a primary metric is, by definition, not scalable or objective.\n\n- The proposal to use \"LLM-as-a-judge\" (Appendix E) as a secondary evaluation method introduces a notoriously unreliable, biased, and difficult-to-reproduce component, which is not a substitute for rigorous, objective, quantitative metrics.\n\n- The authors themselves admit the protocol's flaws in Section 5 (\"Defining and measuring success\"), stating that \"our current evaluation approaches have inherent limitations.\"\n\n**Miscalibrated Difficulty (Floor Effects):**\n\n- The reported results are so poor that the benchmark, in its current form, largely fails as a diagnostic tool.\n\n- The best agent on the full benchmark (AIDE + GPT-4.1) only achieved a \"reasonable\" submission 12.1% of the time (Section 4.1).\n\n- Even on the hand-picked, \"easy\" Lite benchmark, the best agent (AIDE + Claude 3.7) only succeeded 38.9% of the time (Table 2).\n\nThese results indicate a significant floor effect. The tasks are currently too difficult for SOTA agents, preventing any meaningful differentiation between models or scaffolds. The benchmark primarily demonstrates that all agents fail, offering little insight into why they fail or which approaches are more promising.", "questions": "See the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761981823497}, {"id": "vApgYrdfP4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13403/Reviewer_CCeW"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a benchmark framework named TimeSeriesGym, designed to evaluate the capabilities of AI agents in time series machine learning engineering tasks. The core idea is to provide a scalable and agent-agnostic evaluation environment that encompasses time series challenges across multiple domains, while assessing various outputs generated by agents—including prediction files, code, and models.", "review_text": "This paper proposes a benchmark framework named TimeSeriesGym, designed to evaluate the capabilities of AI agents in time series machine learning engineering tasks. The core idea is to provide a scalable and agent-agnostic evaluation environment that encompasses time series challenges across multiple domains, while assessing various outputs generated by agents—including prediction files, code, and models.", "strengths": "1、 The benchmark collects and designs tasks based on real-world data science scenarios, including Kaggle competition problems and practical research tasks such as code migration and model evaluation. These challenges span a wide range of skills—including data processing, model construction, and code understanding and adaptation—reflecting the multifaceted challenges faced by real-world machine learning engineers.\n\n2、 TimeSeriesGym evaluates multiple forms of agent outputs, not only focusing on prediction accuracy and error metrics, but also assessing code generation, model artifacts, and more. Additionally, it introduces optional LLM-based review mechanisms (e.g., code auditing) to supplement quality evaluation.\n\n3、 The authors provide tooling mechanisms that enable large-scale automatic generation of new tasks, making the benchmark highly extensible and sustainable, with the ability to continuously incorporate new challenges.", "weaknesses": "1、 The primary contribution of the paper lies in the construction of the benchmark. Many of its ideas—such as using LLMs for code review, incorporating multi-source tasks, and adopting multi-metric evaluations—are extensions and integrations of existing work rather than entirely novel innovations.\n\n2、 Although the paper lists several existing benchmarks, the distinctions and connections between TimeSeriesGym and those benchmarks are not sufficiently elaborated. For example, beyond the domain difference, how does TimeSeriesGym improve upon the evaluation philosophy of MLE-Bench? A deeper comparative analysis would strengthen the positioning of the proposed framework.\n\n3、 Among the 33 current challenges, tasks sourced from Kaggle tend to be well-structured and may have standard solutions, while the original TimeSeriesGym tasks are often highly complex. GPT-4 achieves only a 12.1% “reasonable solution” rate on TimeSeriesGym, with many tasks yielding no valid outputs. It is recommended that future task sets include medium-difficulty challenges to ensure a smoother difficulty gradient, enabling better tracking of agent performance from beginner to advanced levels. Additionally, establishing baselines for each task—such as simple algorithms or human-level performance—would help characterize task difficulty and the potential improvement space for agents.", "questions": "1、 As the authors noted, due to funding and time constraints, most experiments were conducted on a Lite subset of 6 tasks. This raises two concerns: (A) Can the Lite subset sufficiently represent the full benchmark? Although the authors selected diverse tasks, the sample size of six remains relatively small. (B)Some results—such as the extension of agent steps to 12 hours—were only tested on the Lite subset. It is unclear whether similar trends would hold across the full benchmark.\n\n2、 Additionally, at line 76 below Figure 1, there appears to be white-colored text revealing the authors' institutional affiliation (“Ⓒ 2025 Auton Lab, Carnegie Mellon University”). This may violate the double-blind review policy and should be addressed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a benchmark framework named TimeSeriesGym, designed to evaluate the capabilities of AI agents in time series machine learning engineering tasks. The core idea is to provide a scalable and agent-agnostic evaluation environment that encompasses time series challenges across multiple domains, while assessing various outputs generated by agents—including prediction files, code, and models.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1、 The benchmark collects and designs tasks based on real-world data science scenarios, including Kaggle competition problems and practical research tasks such as code migration and model evaluation. These challenges span a wide range of skills—including data processing, model construction, and code understanding and adaptation—reflecting the multifaceted challenges faced by real-world machine learning engineers.\n\n2、 TimeSeriesGym evaluates multiple forms of agent outputs, not only focusing on prediction accuracy and error metrics, but also assessing code generation, model artifacts, and more. Additionally, it introduces optional LLM-based review mechanisms (e.g., code auditing) to supplement quality evaluation.\n\n3、 The authors provide tooling mechanisms that enable large-scale automatic generation of new tasks, making the benchmark highly extensible and sustainable, with the ability to continuously incorporate new challenges.", "weaknesses": "1、 The primary contribution of the paper lies in the construction of the benchmark. Many of its ideas—such as using LLMs for code review, incorporating multi-source tasks, and adopting multi-metric evaluations—are extensions and integrations of existing work rather than entirely novel innovations.\n\n2、 Although the paper lists several existing benchmarks, the distinctions and connections between TimeSeriesGym and those benchmarks are not sufficiently elaborated. For example, beyond the domain difference, how does TimeSeriesGym improve upon the evaluation philosophy of MLE-Bench? A deeper comparative analysis would strengthen the positioning of the proposed framework.\n\n3、 Among the 33 current challenges, tasks sourced from Kaggle tend to be well-structured and may have standard solutions, while the original TimeSeriesGym tasks are often highly complex. GPT-4 achieves only a 12.1% “reasonable solution” rate on TimeSeriesGym, with many tasks yielding no valid outputs. It is recommended that future task sets include medium-difficulty challenges to ensure a smoother difficulty gradient, enabling better tracking of agent performance from beginner to advanced levels. Additionally, establishing baselines for each task—such as simple algorithms or human-level performance—would help characterize task difficulty and the potential improvement space for agents.", "questions": "1、 As the authors noted, due to funding and time constraints, most experiments were conducted on a Lite subset of 6 tasks. This raises two concerns: (A) Can the Lite subset sufficiently represent the full benchmark? Although the authors selected diverse tasks, the sample size of six remains relatively small. (B)Some results—such as the extension of agent steps to 12 hours—were only tested on the Lite subset. It is unclear whether similar trends would hold across the full benchmark.\n\n2、 Additionally, at line 76 below Figure 1, there appears to be white-colored text revealing the authors' institutional affiliation (“Ⓒ 2025 Auton Lab, Carnegie Mellon University”). This may violate the double-blind review policy and should be addressed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761981693521}, {"id": "pT9RYAcR8n", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13403/Reviewer_9ifC"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents TimeSeriesGym, a benchmark and environment for evaluating agents that perform end-to-end machine learning work on time series problems. The benchmark aggregates a collection of challenges that span forecasting, classification, anomaly detection, data cleaning, hyperparameter tuning, and code migration. It evaluates not only final predictions but also multiple artifacts such as code, trained models, and submission files. It defines aggregate metrics for validity and reasonableness and reports results across different toolchains and foundation models. The paper argues that time series is an underrepresented setting for agent evaluation and that real engineering workflows require capabilities beyond model inference. The experimental section compares several agent frameworks under controlled resources and reports ablations on time budgets and guidance strategies. The authors claim that TimeSeriesGym is scalable, extensible, and better aligned with the realities of time series engineering than prior leaderboards that focus on a single metric or a single output file.", "review_text": "This paper presents TimeSeriesGym, a benchmark and environment for evaluating agents that perform end-to-end machine learning work on time series problems. The benchmark aggregates a collection of challenges that span forecasting, classification, anomaly detection, data cleaning, hyperparameter tuning, and code migration. It evaluates not only final predictions but also multiple artifacts such as code, trained models, and submission files. It defines aggregate metrics for validity and reasonableness and reports results across different toolchains and foundation models. The paper argues that time series is an underrepresented setting for agent evaluation and that real engineering workflows require capabilities beyond model inference. The experimental section compares several agent frameworks under controlled resources and reports ablations on time budgets and guidance strategies. The authors claim that TimeSeriesGym is scalable, extensible, and better aligned with the realities of time series engineering than prior leaderboards that focus on a single metric or a single output file.", "strengths": "- Clear motivation that time series engineering requires more than single step prediction and that agents should be evaluated on multi stage workflows.\n\n- The authors collected a large number of datasets, with broad task coverage and inclusion of multiple domains which improves the ecological validity of the benchmark.\n\n- In this paper, multi artifact evaluation that considers predictions, code quality checks, and trained models which aligns the benchmark with real practice.\n\n- Agent agnostic design that supports different frameworks and enables trajectory collection for future training.\n\n- A documented process for adding new tasks which supports scalability and long term maintenance.", "weaknesses": "1. The paper mixes the benchmark, the task generation mechanism, the multi artifact scoring, the trajectory data loop, and the time series focus without a clear primary to secondary hierarchy and without a concise contribution figure. It does not decompose difficulty across different time series task families and it does not design difficulty along agent reasoning dimensions.\n\n2. Due to compute constraints most experiments are conducted on the Lite set of six tasks. Although the authors state that these tasks cover key skills there is no statistical validation of domain diversity or difficulty distribution which can bias conclusions toward low cost scenarios.\n\n3. The contribution is primarily engineering. The paper lacks a unifying methodological or theoretical insight. Although this is a benchmark paper it would benefit from a central methodological principle that guides design choices across sections.\n\n4. The table presents results for 4/50→12/150 and for Step wise reminder vs No reminder, but it does not include significance testing.\n\n5. The definitions of Valid submission and Reasonable submission appear in prose in the methods section without a unified symbolization or an explicit decision boundary. This reduces clarity and hurts reproducibility.", "questions": "1. Can you provide stratified results by task family with confidence intervals and tests for significance. This would help to understand which capabilities drive aggregate gains.\n\n2. How do you ensure that the Lite subset is representative of the full benchmark. Please include quantitative evidence of coverage and difficulty distribution.\n\n3. Can you formalize the validity and reasonableness criteria using symbols and thresholds and show calibration plots or decision curves.\n\n4. What steps can be taken to reduce reliance on leaderboard medians or external competitions that may change over time.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents TimeSeriesGym, a benchmark and environment for evaluating agents that perform end-to-end machine learning work on time series problems. The benchmark aggregates a collection of challenges that span forecasting, classification, anomaly detection, data cleaning, hyperparameter tuning, and code migration. It evaluates not only final predictions but also multiple artifacts such as code, trained models, and submission files. It defines aggregate metrics for validity and reasonableness and reports results across different toolchains and foundation models. The paper argues that time series is an underrepresented setting for agent evaluation and that real engineering workflows require capabilities beyond model inference. The experimental section compares several agent frameworks under controlled resources and reports ablations on time budgets and guidance strategies. The authors claim that TimeSeriesGym is scalable, extensible, and better aligned with the realities of time series engineering than prior leaderboards that focus on a single metric or a single output file.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Clear motivation that time series engineering requires more than single step prediction and that agents should be evaluated on multi stage workflows.\n\n- The authors collected a large number of datasets, with broad task coverage and inclusion of multiple domains which improves the ecological validity of the benchmark.\n\n- In this paper, multi artifact evaluation that considers predictions, code quality checks, and trained models which aligns the benchmark with real practice.\n\n- Agent agnostic design that supports different frameworks and enables trajectory collection for future training.\n\n- A documented process for adding new tasks which supports scalability and long term maintenance.", "weaknesses": "1. The paper mixes the benchmark, the task generation mechanism, the multi artifact scoring, the trajectory data loop, and the time series focus without a clear primary to secondary hierarchy and without a concise contribution figure. It does not decompose difficulty across different time series task families and it does not design difficulty along agent reasoning dimensions.\n\n2. Due to compute constraints most experiments are conducted on the Lite set of six tasks. Although the authors state that these tasks cover key skills there is no statistical validation of domain diversity or difficulty distribution which can bias conclusions toward low cost scenarios.\n\n3. The contribution is primarily engineering. The paper lacks a unifying methodological or theoretical insight. Although this is a benchmark paper it would benefit from a central methodological principle that guides design choices across sections.\n\n4. The table presents results for 4/50→12/150 and for Step wise reminder vs No reminder, but it does not include significance testing.\n\n5. The definitions of Valid submission and Reasonable submission appear in prose in the methods section without a unified symbolization or an explicit decision boundary. This reduces clarity and hurts reproducibility.", "questions": "1. Can you provide stratified results by task family with confidence intervals and tests for significance. This would help to understand which capabilities drive aggregate gains.\n\n2. How do you ensure that the Lite subset is representative of the full benchmark. Please include quantitative evidence of coverage and difficulty distribution.\n\n3. Can you formalize the validity and reasonableness criteria using symbols and thresholds and show calibration plots or decision curves.\n\n4. What steps can be taken to reduce reliance on leaderboard medians or external competitions that may change over time.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761890184043}, {"id": "rJGPQoyWat", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13403/Reviewer_E1hF"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper introduces an open-source benchmark TimeSeriesGym for MLE agents evaluation, mostly focused on time series data. It contains 33 challenges and evaluated 3 scaffolds of agents with each scaffold evaluated on up to 3 base models. The framework supports multimodal output evaluation, specific skill evaluation and holistic evaluation(quantitative & qualitative).", "review_text": "This paper introduces an open-source benchmark TimeSeriesGym for MLE agents evaluation, mostly focused on time series data. It contains 33 challenges and evaluated 3 scaffolds of agents with each scaffold evaluated on up to 3 base models. The framework supports multimodal output evaluation, specific skill evaluation and holistic evaluation(quantitative & qualitative).", "strengths": "This benchmark fills a gap on time series ML pipeline among existing ML benchmarks. \n\nThe proposed benchmark offers specific skills evaluation and hybrid scoring of qualitative + quantitive evaluation.", "weaknesses": "This paper emphasizes benchmark at scale which would require limited human efforts but there is no discussion on how exactly new challenges should be adapted to the time series gym. It's mentioned that new challenge can be added within two hours, but how is the quality of the generated new challenges? In addition, two hours time is probably also enough to set up a specialized evaluation for a challenge without the need to integrate it to the current evaluation framework. The evidence for claimed scalability is limited, at least not well-presented. \n\nIt feels like this paper is mostly doing the same thing as MLE-bench but with specifically curated challenges focusing on time series and additionally took inspiration from other benchmarks and combined their advantages such as multimodal, skill-based, and holistic evaluation. \n\nThe experiment findings are also very predictable. Existing agents don't have the full capability to automate complete ML pipeline and highly likely the reasoning model produces more reasonable solutions. It's hard to find new takeaways or interesting insights from this paper. \n\nSome experiment settings are not meaningful. For example, providing remaining time reminders to agents vs. without time reminders. I find it hard to see the value in this experiment. \n\nThe evaluation is mostly limited to gpt 4.1 model, and there is no open source models being evaluated at all. \n\nSome challenges feel very synthetic. For example, implement MOMENT for anomaly detection task. But why do we need to integrate such task into this benchmark at all given that the code to solve this task is already available in MOMENT official github repo.", "questions": "The paper lacks clarity. It's not clear how exactly the challenges are defined, are they all manually crafted based on Kaggle and Github Repos? Line 155-163 shows what each challenge is made or but is it manual effort to create these information based on the dataset or repository? \n\nIn line 315-316, it's mentioned that \"simply loading and re-saving the provided sample submission file without any model inference or data processing is deemed unreasonable\". How is this checked? Is it through llm-as-a-judge evaluation or manual inspection? \n\nIn line 228-231, the paper mentions that it can grade diverse artifacts generated throughout the MLE life cycle but how? I don't think this is explained. Rather than spending almost a whole page on limitation/future work/open questions, I would suggest spending more text to improve the clarity of the paper and explaining details of this benchmark framework in a thorough manner.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an open-source benchmark TimeSeriesGym for MLE agents evaluation, mostly focused on time series data. It contains 33 challenges and evaluated 3 scaffolds of agents with each scaffold evaluated on up to 3 base models. The framework supports multimodal output evaluation, specific skill evaluation and holistic evaluation(quantitative & qualitative).", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "This benchmark fills a gap on time series ML pipeline among existing ML benchmarks. \n\nThe proposed benchmark offers specific skills evaluation and hybrid scoring of qualitative + quantitive evaluation.", "weaknesses": "This paper emphasizes benchmark at scale which would require limited human efforts but there is no discussion on how exactly new challenges should be adapted to the time series gym. It's mentioned that new challenge can be added within two hours, but how is the quality of the generated new challenges? In addition, two hours time is probably also enough to set up a specialized evaluation for a challenge without the need to integrate it to the current evaluation framework. The evidence for claimed scalability is limited, at least not well-presented. \n\nIt feels like this paper is mostly doing the same thing as MLE-bench but with specifically curated challenges focusing on time series and additionally took inspiration from other benchmarks and combined their advantages such as multimodal, skill-based, and holistic evaluation. \n\nThe experiment findings are also very predictable. Existing agents don't have the full capability to automate complete ML pipeline and highly likely the reasoning model produces more reasonable solutions. It's hard to find new takeaways or interesting insights from this paper. \n\nSome experiment settings are not meaningful. For example, providing remaining time reminders to agents vs. without time reminders. I find it hard to see the value in this experiment. \n\nThe evaluation is mostly limited to gpt 4.1 model, and there is no open source models being evaluated at all. \n\nSome challenges feel very synthetic. For example, implement MOMENT for anomaly detection task. But why do we need to integrate such task into this benchmark at all given that the code to solve this task is already available in MOMENT official github repo.", "questions": "The paper lacks clarity. It's not clear how exactly the challenges are defined, are they all manually crafted based on Kaggle and Github Repos? Line 155-163 shows what each challenge is made or but is it manual effort to create these information based on the dataset or repository? \n\nIn line 315-316, it's mentioned that \"simply loading and re-saving the provided sample submission file without any model inference or data processing is deemed unreasonable\". How is this checked? Is it through llm-as-a-judge evaluation or manual inspection? \n\nIn line 228-231, the paper mentions that it can grade diverse artifacts generated throughout the MLE life cycle but how? I don't think this is explained. Rather than spending almost a whole page on limitation/future work/open questions, I would suggest spending more text to improve the clarity of the paper and explaining details of this benchmark framework in a thorough manner.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761547531492}], "openreview_url": "https://openreview.net/forum?id=8gdfWRilR7", "arxiv_id": "2505.13291", "paper_pdf": "papers/8gdfWRilR7.pdf", "paper_pdf_sha256": "589c81fe9b3697eea2f0f91a048098c1879d46cdd0e30c06e094c6e6d66f6538", "paper_pdf_bytes": 1142833, "paper_pdf_source": "openreview", "code_url": "https://github.com/moment-timeseries-foundation-model/TimeSeriesGym", "code_repository": "moment-timeseries-foundation-model/TimeSeriesGym", "code_commit": "853b4782f8ca18121c04ee2c2cf96e9b189b0c11", "code_archive": "repos/8gdfWRilR7.zip", "code_archive_sha256": "d124546b6872a0adf8f0098d59f34e98759d697b24c45ec510633ba2fd067583", "code_archive_bytes": 2390403, "code_file_count": 114, "code_extensions": {".py": 106, ".sh": 8}, "github_disk_usage_kb": 2125, "github_languages": {"Python": 397513, "Shell": 13091, "Dockerfile": 8552}, "github_archived": false, "github_pushed_at": "2025-11-30T22:28:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/timeseriesgym-a-scalable-benchmark-for-time"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OyjMJjfhiw", "year": 2025, "status": "rejected", "title": "LLM Embeddings Improve Test-Time Adaptation to Tabular $Y|X$-Shifts", "authors": ["Yibo Zeng", "Jiashuo Liu", "Henry Lam", "Hongseok Namkoong"], "authorids": ["~Yibo_Zeng1", "~Jiashuo_Liu1", "~Henry_Lam1", "~Hongseok_Namkoong2"], "authors_source": "OpenReview API", "abstract": "For tabular datasets, the change in the relationship between the label and covariates ($Y|X$-shifts) is common due to missing variables. Since it is impossible to generalize to a completely new and unknown domain, we study models that are easy to adapt to the target domain even with few labeled examples. We focus on building more informative representations of tabular data that can mitigate $Y|X$-shifts, and propose to leverage the prior world knowledge in LLMs by serializing the tabular data to encode it. We find LLM embeddings alone provide inconsistent improvements in robustness, but models trained on them can be well adapted to the target domain even using 32 labeled observations. Our finding is based on a systematic study consisting of 7650 source-target pairs and benchmark against **261,000** model configurations trained by 20 algorithms. Our observation holds when ablating the size of accessible target data and different adaptation strategies.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "TqZ75qQST9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10485/Reviewer_pBwo"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper investigates the challenges posed by shifts in tabular data, specifically shifts in the relationship between labels and covariates (Y|X-shifts). These shifts often occur due to missing variables or confounders, making it difficult for models trained on a source domain to generalize to a target domain. \n\nThe authors propose a solution using embeddings generated from large language models (LLMs) to encode tabular data. They demonstrate that LLM embeddings can improve adaptation to target domains even with limited labeled target samples, as the embeddings incorporate a broader contextual knowledge that helps mitigate the effects of these Y|X-shifts. Their study includes a comprehensive evaluation across 7,650 source-target pairs and comparisons with 20 algorithms, revealing that shallow neural networks (NNs) trained on LLM embeddings outperform other models in many cases.\n\nThe study's results highlight that while LLM embeddings alone do not consistently surpass state-of-the-art tree-based models, their performance significantly improves when adapted with a small number of target samples. Specifically, using 32 labeled target samples for fine-tuning yields substantial gains across multiple datasets, especially under scenarios of strong Y|X-shifts. The authors also explore different methods of adaptation, including in-context learning, low-rank adaptation, and prefix tuning, finding that fine-tuning with LLM embeddings provides a practical approach for generalizing across tabular Y|X-shifts.", "review_text": "This paper investigates the challenges posed by shifts in tabular data, specifically shifts in the relationship between labels and covariates (Y|X-shifts). These shifts often occur due to missing variables or confounders, making it difficult for models trained on a source domain to generalize to a target domain. \n\nThe authors propose a solution using embeddings generated from large language models (LLMs) to encode tabular data. They demonstrate that LLM embeddings can improve adaptation to target domains even with limited labeled target samples, as the embeddings incorporate a broader contextual knowledge that helps mitigate the effects of these Y|X-shifts. Their study includes a comprehensive evaluation across 7,650 source-target pairs and comparisons with 20 algorithms, revealing that shallow neural networks (NNs) trained on LLM embeddings outperform other models in many cases.\n\nThe study's results highlight that while LLM embeddings alone do not consistently surpass state-of-the-art tree-based models, their performance significantly improves when adapted with a small number of target samples. Specifically, using 32 labeled target samples for fine-tuning yields substantial gains across multiple datasets, especially under scenarios of strong Y|X-shifts. The authors also explore different methods of adaptation, including in-context learning, low-rank adaptation, and prefix tuning, finding that fine-tuning with LLM embeddings provides a practical approach for generalizing across tabular Y|X-shifts.", "strengths": "This paper excels in addressing a pressing issue in machine learning: adapting models to distributional shifts in tabular data, specifically Y|X-shifts. The authors take an innovative approach by leveraging LLM embeddings to encode tabular data, which allows the model to incorporate broader contextual knowledge. This strength is significant because it enables the model to adapt to changes in label-covariate relationships, even with minimal (32 examples) labeled data from the target domain. The comprehensive evaluation of 7,650 source-target pairs across three datasets (ACS Income, Mobility, and Public Coverage) highlights the robustness of their approach. By testing against 20 algorithms (traditional GBDT and NN based models) and exploring a vast configuration space, the authors provide a solid foundation for the efficacy of LLM embeddings in generalizing across diverse tabular data environments. This thorough evaluation not only strengthens the empirical findings but also sets a new benchmark in evaluating distributional shifts in tabular datasets.", "weaknesses": "This method is only applicable when there is a description available for all features of the tabular data. It would be good to expand this method to datasets which include features without any description. A possible solution can be by learning embeddings of such features from scratch and utilizing them as is done in TabTransformer. This will provide a comprehensive solution for different types of tabular data.", "questions": "The proposed method appears limited to datasets where all features have available descriptions. Expanding this approach to include datasets with features lacking descriptions would increase its versatility. One possible solution could involve learning embeddings for such features from scratch, similar to methods like TabTransformer, which can effectively handle mixed data types by embedding categorical features without requiring explicit descriptions.\n\nIt would be beneficial for the authors to discuss this limitation in the manuscript and potentially compare their method to approaches like TabTransformer that support feature embedding without descriptions. This addition would provide a clearer understanding of the method’s scope and suggest pathways for future improvements in handling various types of tabular data.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the challenges posed by shifts in tabular data, specifically shifts in the relationship between labels and covariates (Y|X-shifts). These shifts often occur due to missing variables or confounders, making it difficult for models trained on a source domain to generalize to a target domain. \n\nThe authors propose a solution using embeddings generated from large language models (LLMs) to encode tabular data. They demonstrate that LLM embeddings can improve adaptation to target domains even with limited labeled target samples, as the embeddings incorporate a broader contextual knowledge that helps mitigate the effects of these Y|X-shifts. Their study includes a comprehensive evaluation across 7,650 source-target pairs and comparisons with 20 algorithms, revealing that shallow neural networks (NNs) trained on LLM embeddings outperform other models in many cases.\n\nThe study's results highlight that while LLM embeddings alone do not consistently surpass state-of-the-art tree-based models, their performance significantly improves when adapted with a small number of target samples. Specifically, using 32 labeled target samples for fine-tuning yields substantial gains across multiple datasets, especially under scenarios of strong Y|X-shifts. The authors also explore different methods of adaptation, including in-context learning, low-rank adaptation, and prefix tuning, finding that fine-tuning with LLM embeddings provides a practical approach for generalizing across tabular Y|X-shifts.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "This paper excels in addressing a pressing issue in machine learning: adapting models to distributional shifts in tabular data, specifically Y|X-shifts. The authors take an innovative approach by leveraging LLM embeddings to encode tabular data, which allows the model to incorporate broader contextual knowledge. This strength is significant because it enables the model to adapt to changes in label-covariate relationships, even with minimal (32 examples) labeled data from the target domain. The comprehensive evaluation of 7,650 source-target pairs across three datasets (ACS Income, Mobility, and Public Coverage) highlights the robustness of their approach. By testing against 20 algorithms (traditional GBDT and NN based models) and exploring a vast configuration space, the authors provide a solid foundation for the efficacy of LLM embeddings in generalizing across diverse tabular data environments. This thorough evaluation not only strengthens the empirical findings but also sets a new benchmark in evaluating distributional shifts in tabular datasets.", "weaknesses": "This method is only applicable when there is a description available for all features of the tabular data. It would be good to expand this method to datasets which include features without any description. A possible solution can be by learning embeddings of such features from scratch and utilizing them as is done in TabTransformer. This will provide a comprehensive solution for different types of tabular data.", "questions": "The proposed method appears limited to datasets where all features have available descriptions. Expanding this approach to include datasets with features lacking descriptions would increase its versatility. One possible solution could involve learning embeddings for such features from scratch, similar to methods like TabTransformer, which can effectively handle mixed data types by embedding categorical features without requiring explicit descriptions.\n\nIt would be beneficial for the authors to discuss this limitation in the manuscript and potentially compare their method to approaches like TabTransformer that support feature embedding without descriptions. This addition would provide a clearer understanding of the method’s scope and suggest pathways for future improvements in handling various types of tabular data.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731045093453}, {"id": "sBqb2GslAF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10485/Reviewer_Tko8"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes a method to solve the Y|X shift problem in the tabular data domain using LLM embeddings. Under the LLM+NN model architecture, it uses concatenation to combine the domain information of LLM with the original tabular data information for improving prediction robustness and performance. The paper study models that are easy to adapt to the target domain even with few labeled examples so that fine-tuning the classification model with a small number of samples can further improve performance. Experiments can prove the effectiveness of the proposed method.", "review_text": "This paper proposes a method to solve the Y|X shift problem in the tabular data domain using LLM embeddings. Under the LLM+NN model architecture, it uses concatenation to combine the domain information of LLM with the original tabular data information for improving prediction robustness and performance. The paper study models that are easy to adapt to the target domain even with few labeled examples so that fine-tuning the classification model with a small number of samples can further improve performance. Experiments can prove the effectiveness of the proposed method.", "strengths": "1. The motivation is clear and the algorithm is sensible.\n2. The proposed method is tested on several benchmarks.\n3. The proposed method is easy to understand and simple. It does not require a large number of LLM calls and the computational complexity is not high.", "weaknesses": "1. On adaptation stage: The method requires the target's true label to guide the model parameter update. This setting is different from the basic test-time adaptation (TTA) setting, which is generally an unsupervised objective function [1]. The setting in this article is more like a fine-tuning setting.\n2. The experiments in this article are mainly based on the ACS dataset. In recent studies, there are some other benchmarks like TableShift [2] contains Y|X shifts datasets. In the report they provided, ACS Income and ACS Pub.Cov dataset seem not have severe Y|X shifts. In order to prove the robustness to severe shifts, I think experiments on those datasets like ACS Unemployment and ICU Length of Stay are necessary.\n\n[1] Wang, Dequan, et al. \"Tent: Fully test-time adaptation by entropy minimization.\" arXiv preprint arXiv:2006.10726 (2020).\n\n[2] Gardner, Josh, Zoran Popovic, and Ludwig Schmidt. \"Benchmarking distribution shift in tabular data with tableshift.\" Advances in Neural Information Processing Systems 36 (2024).", "questions": "1. Figure 3: Recent study shows that label shift in tabular data causes performance degrade [1], why not take label shift into account? \n2. Are Y|X shifts and concept shift the same in meaning?\n3. How does proposed compare to state-of-the-art methods designed for tabular TTA [2]?\n4. How to evaluate Y|X shifts degree?\n\n[1] Gardner, Josh, Zoran Popovic, and Ludwig Schmidt. \"Benchmarking distribution shift in tabular data with tableshift.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[2] Ren, Weijieying, et al. \"TabLog: Test-Time Adaptation for Tabular Data Using Logic Rules.\" Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to solve the Y|X shift problem in the tabular data domain using LLM embeddings. Under the LLM+NN model architecture, it uses concatenation to combine the domain information of LLM with the original tabular data information for improving prediction robustness and performance. The paper study models that are easy to adapt to the target domain even with few labeled examples so that fine-tuning the classification model with a small number of samples can further improve performance. Experiments can prove the effectiveness of the proposed method.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The motivation is clear and the algorithm is sensible.\n2. The proposed method is tested on several benchmarks.\n3. The proposed method is easy to understand and simple. It does not require a large number of LLM calls and the computational complexity is not high.", "weaknesses": "1. On adaptation stage: The method requires the target's true label to guide the model parameter update. This setting is different from the basic test-time adaptation (TTA) setting, which is generally an unsupervised objective function [1]. The setting in this article is more like a fine-tuning setting.\n2. The experiments in this article are mainly based on the ACS dataset. In recent studies, there are some other benchmarks like TableShift [2] contains Y|X shifts datasets. In the report they provided, ACS Income and ACS Pub.Cov dataset seem not have severe Y|X shifts. In order to prove the robustness to severe shifts, I think experiments on those datasets like ACS Unemployment and ICU Length of Stay are necessary.\n\n[1] Wang, Dequan, et al. \"Tent: Fully test-time adaptation by entropy minimization.\" arXiv preprint arXiv:2006.10726 (2020).\n\n[2] Gardner, Josh, Zoran Popovic, and Ludwig Schmidt. \"Benchmarking distribution shift in tabular data with tableshift.\" Advances in Neural Information Processing Systems 36 (2024).", "questions": "1. Figure 3: Recent study shows that label shift in tabular data causes performance degrade [1], why not take label shift into account? \n2. Are Y|X shifts and concept shift the same in meaning?\n3. How does proposed compare to state-of-the-art methods designed for tabular TTA [2]?\n4. How to evaluate Y|X shifts degree?\n\n[1] Gardner, Josh, Zoran Popovic, and Ludwig Schmidt. \"Benchmarking distribution shift in tabular data with tableshift.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[2] Ren, Weijieying, et al. \"TabLog: Test-Time Adaptation for Tabular Data Using Logic Rules.\" Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730708001010}, {"id": "LpZVep9Fkw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10485/Reviewer_ZwTt"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper tackles an exciting problem by utilizing LLM embeddings for tabular classification. The authors conduct comprehensive experiments to assess the effectiveness of LLM embeddings in tabular tasks, with promising results that highlight a new and potentially impactful direction for advancing tabular data analysis.", "review_text": "This paper tackles an exciting problem by utilizing LLM embeddings for tabular classification. The authors conduct comprehensive experiments to assess the effectiveness of LLM embeddings in tabular tasks, with promising results that highlight a new and potentially impactful direction for advancing tabular data analysis.", "strengths": "1. The paper is well-structured and easy to follow, with the three technical components clearly and accessibly presented.\n2. The experiments are comprehensive, which can demonstrate the claims of this paper.", "weaknesses": "1. The dataset variety is limited, as experiments are conducted on only three datasets within the same domain, which restricts the generalizability of the results and may affect the robustness of the findings.\n2. The proposed method incorporates additional domain knowledge to enhance classification performance. However, it is unclear whether comparisons with tree-based methods are entirely fair, given that such methods may not effectively leverage domain knowledge. A thorough discussion on the source of performance improvements: whether from domain knowledge integration or specific technical innovations—would add valuable clarity.\n3. The theoretical analysis may not be entirely appropriate for this context. Although the paper introduces a few-shot prompt tuning approach to maximize the use of limited labeled data, the theorem presented assumes a large  $m$  to achieve a small generalization bound, which contradicts the few-shot setting.\n4. The term “test-time adaptation” in the title may be misleading, as it typically refers to model tuning using only unlabeled test data, which does not align with the approach described in this paper.", "questions": "Please refer to the \"Weaknesses\" section and answer the following questions. \n1. Why does this paper lack a conclusion section?\n2. How are prompt templates tailored for different datasets? Typically, datasets with varying columns and feature types require significantly different prompts to optimize the effectiveness of the LLM.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles an exciting problem by utilizing LLM embeddings for tabular classification. The authors conduct comprehensive experiments to assess the effectiveness of LLM embeddings in tabular tasks, with promising results that highlight a new and potentially impactful direction for advancing tabular data analysis.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper is well-structured and easy to follow, with the three technical components clearly and accessibly presented.\n2. The experiments are comprehensive, which can demonstrate the claims of this paper.", "weaknesses": "1. The dataset variety is limited, as experiments are conducted on only three datasets within the same domain, which restricts the generalizability of the results and may affect the robustness of the findings.\n2. The proposed method incorporates additional domain knowledge to enhance classification performance. However, it is unclear whether comparisons with tree-based methods are entirely fair, given that such methods may not effectively leverage domain knowledge. A thorough discussion on the source of performance improvements: whether from domain knowledge integration or specific technical innovations—would add valuable clarity.\n3. The theoretical analysis may not be entirely appropriate for this context. Although the paper introduces a few-shot prompt tuning approach to maximize the use of limited labeled data, the theorem presented assumes a large  $m$  to achieve a small generalization bound, which contradicts the few-shot setting.\n4. The term “test-time adaptation” in the title may be misleading, as it typically refers to model tuning using only unlabeled test data, which does not align with the approach described in this paper.", "questions": "Please refer to the \"Weaknesses\" section and answer the following questions. \n1. Why does this paper lack a conclusion section?\n2. How are prompt templates tailored for different datasets? Typically, datasets with varying columns and feature types require significantly different prompts to optimize the effectiveness of the LLM.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730700735173}, {"id": "Oqh6XLfj8y", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10485/Reviewer_MKos"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper studies distribution shifts in tabular data by using embeddings from large language models (LLMs) to enhance test-time adaptation. The paper uses a LLM encoder  to featurize tabular data and fit a shallow neural network on these embeddings for tabular prediction. The paper investigates different configurations including various training algorithms, incorporating additional domain information or not, fine-tuning the shallow neural network with few target domain samples or not. The paper provides comprehensive experiments on these configurations.", "review_text": "This paper studies distribution shifts in tabular data by using embeddings from large language models (LLMs) to enhance test-time adaptation. The paper uses a LLM encoder  to featurize tabular data and fit a shallow neural network on these embeddings for tabular prediction. The paper investigates different configurations including various training algorithms, incorporating additional domain information or not, fine-tuning the shallow neural network with few target domain samples or not. The paper provides comprehensive experiments on these configurations.", "strengths": "1.  The authors provide a broad empirical evaluation across thousands of source-target pairs, diverse model configurations, and multiple algorithms, which strengthens the credibility of their findings. The paper includes a variety of performance metrics to assess the effectiveness of LLM embeddings under different conditions.  \n\n2. By highlighting different scenarios and distributional changes, the authors clarify the conditions under which the approach excels or may face limitations. This transparency in results provides a balanced view of the method's robustness and allows researchers to see where it might or might not be effective.\n\n3. The authors include a focused analysis on the effectiveness of few-shot adaptation, showing how minimal labeled data in the target domain can significantly improve performance.", "weaknesses": "1. The approach essentially applies existing LLMs as embeddings for tabular data without introducing substantial new methods for embedding generation or adaptation. Serializing tabular data into texts and using LLMs to extract features/make predictions have been studied in previous works. The adaptation technique described (using few-shot labeled samples for fine-tuning) is a well-known concept in transfer learning and domain adaptation. Few-shot learning techniques have been applied across various fields, and using minimal labeled data to adapt models to new domains is not unique to this work.\n\n2. Although the paper provide comprehensive experiments on the configurations of the proposed method, it is still unclear to position how the proposed method performs in the literature. Specifically, the paper compares with baselines of Tabular Data + NN, where NN is the fully-connected neural networks. However, there are rich literatures on more advanced neural networks tailored for tabular data (e.g., FT-Transformer and many others). In addition, LLMs have been used for tabular data recently [1]. Comparisons with state-of-the-art methods on neural networks for tabular data and LLMs for tabular data are expected. Having said this, I'm not asking the authors to add more experiments in the rebuttal stage.\n\n[1] Fang, Xi, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Jane Qi, Scott Nickleach, Diego Socolinsky, Srinivasan Sengamedu, and Christos Faloutsos. \"Large language models (LLMs) on tabular data: Prediction, generation, and understanding-a survey.\" (2024).", "questions": "I suggest the authors to provide a conclusion of the paper summarising the key findings and discussing the limitations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies distribution shifts in tabular data by using embeddings from large language models (LLMs) to enhance test-time adaptation. The paper uses a LLM encoder  to featurize tabular data and fit a shallow neural network on these embeddings for tabular prediction. The paper investigates different configurations including various training algorithms, incorporating additional domain information or not, fine-tuning the shallow neural network with few target domain samples or not. The paper provides comprehensive experiments on these configurations.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.  The authors provide a broad empirical evaluation across thousands of source-target pairs, diverse model configurations, and multiple algorithms, which strengthens the credibility of their findings. The paper includes a variety of performance metrics to assess the effectiveness of LLM embeddings under different conditions.  \n\n2. By highlighting different scenarios and distributional changes, the authors clarify the conditions under which the approach excels or may face limitations. This transparency in results provides a balanced view of the method's robustness and allows researchers to see where it might or might not be effective.\n\n3. The authors include a focused analysis on the effectiveness of few-shot adaptation, showing how minimal labeled data in the target domain can significantly improve performance.", "weaknesses": "1. The approach essentially applies existing LLMs as embeddings for tabular data without introducing substantial new methods for embedding generation or adaptation. Serializing tabular data into texts and using LLMs to extract features/make predictions have been studied in previous works. The adaptation technique described (using few-shot labeled samples for fine-tuning) is a well-known concept in transfer learning and domain adaptation. Few-shot learning techniques have been applied across various fields, and using minimal labeled data to adapt models to new domains is not unique to this work.\n\n2. Although the paper provide comprehensive experiments on the configurations of the proposed method, it is still unclear to position how the proposed method performs in the literature. Specifically, the paper compares with baselines of Tabular Data + NN, where NN is the fully-connected neural networks. However, there are rich literatures on more advanced neural networks tailored for tabular data (e.g., FT-Transformer and many others). In addition, LLMs have been used for tabular data recently [1]. Comparisons with state-of-the-art methods on neural networks for tabular data and LLMs for tabular data are expected. Having said this, I'm not asking the authors to add more experiments in the rebuttal stage.\n\n[1] Fang, Xi, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Jane Qi, Scott Nickleach, Diego Socolinsky, Srinivasan Sengamedu, and Christos Faloutsos. \"Large language models (LLMs) on tabular data: Prediction, generation, and understanding-a survey.\" (2024).", "questions": "I suggest the authors to provide a conclusion of the paper summarising the key findings and discussing the limitations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730590689847}, {"id": "nr90cseGtJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10485/Reviewer_8GFo"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "This paper focuses on the test-time adaptation for tabular data with Y|X shifts. Specifically, the proposal converts the raw tabular data into a new embedding with LLM and trains a shallow neural network on the embedding. When facing the distribution shift, the authors propose to fine-tune the network for the adaptation. Experimental results on three datasets are reported.", "review_text": "This paper focuses on the test-time adaptation for tabular data with Y|X shifts. Specifically, the proposal converts the raw tabular data into a new embedding with LLM and trains a shallow neural network on the embedding. When facing the distribution shift, the authors propose to fine-tune the network for the adaptation. Experimental results on three datasets are reported.", "strengths": "1) The paper focuses on the test-time adaptation for tabular data with Y|X shift. The problem is important and has been rarely studied.\n2) The proposal is clear and expected to be effective.", "weaknesses": "1) The technical contribution is limited. The paper is more like an engineering work. The adopted techniques are all widely-adopted techniques for tabular data learning and the conclusions are also not surprising. It is not clear what technical problems have been addressed or what new insights have been proposed by this paper.\n2) For the experimental results, the authors only adopted three datasets and the feature dimensions are small-scale.\n3) How many times about the experiments repeated and how about the performance variance?\n4) The authors claim that the LLM embeddings improve performance. However, the LLM embeddings are influenced by the serialization method and the choices of different LLMs. The influence of these should be discussed.\n5) It is impractical to know the distribution shift in advance. In more practical scenarios, we need to detect the distribution shift first and then adapt.", "questions": "As discussed above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the test-time adaptation for tabular data with Y|X shifts. Specifically, the proposal converts the raw tabular data into a new embedding with LLM and trains a shallow neural network on the embedding. When facing the distribution shift, the authors propose to fine-tune the network for the adaptation. Experimental results on three datasets are reported.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "1) The paper focuses on the test-time adaptation for tabular data with Y|X shift. The problem is important and has been rarely studied.\n2) The proposal is clear and expected to be effective.", "weaknesses": "1) The technical contribution is limited. The paper is more like an engineering work. The adopted techniques are all widely-adopted techniques for tabular data learning and the conclusions are also not surprising. It is not clear what technical problems have been addressed or what new insights have been proposed by this paper.\n2) For the experimental results, the authors only adopted three datasets and the feature dimensions are small-scale.\n3) How many times about the experiments repeated and how about the performance variance?\n4) The authors claim that the LLM embeddings improve performance. However, the LLM embeddings are influenced by the serialization method and the choices of different LLMs. The influence of these should be discussed.\n5) It is impractical to know the distribution shift in advance. In more practical scenarios, we need to detect the distribution shift first and then adapt.", "questions": "As discussed above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730126298283}], "openreview_url": "https://openreview.net/forum?id=OyjMJjfhiw", "arxiv_id": "2410.07395", "paper_pdf": "papers/OyjMJjfhiw.pdf", "paper_pdf_sha256": "729336df048bddee27c5ee4c405a83eb1ccf64f2e421c8d600e8c0f4b692d6d9", "paper_pdf_bytes": 1056479, "paper_pdf_source": "openreview", "code_url": "https://github.com/namkoong-lab/LLM-Tabular-Shifts", "code_repository": "namkoong-lab/LLM-Tabular-Shifts", "code_commit": "9bab7ea2c4f0dc63412492025d8e3a2ec72e1f96", "code_archive": "repos/OyjMJjfhiw.zip", "code_archive_sha256": "d90e360a582a0f2b858b6b57dffc7e13472c92ac82ea4ddadf517570948510f7", "code_archive_bytes": 243625, "code_file_count": 87, "code_extensions": {".py": 82, ".ipynb": 5}, "github_disk_usage_kb": 153, "github_languages": {"Python": 596460, "Jupyter Notebook": 127816}, "github_archived": false, "github_pushed_at": "2024-10-17T17:31:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/llm-embeddings-improve-test-time-adaptation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bCNYFOaWsy", "year": 2024, "status": "rejected", "title": "Class-Imbalanced Graph Learning without Class Rebalancing", "authors": ["Zhining Liu", "Zhichen Zeng", "Ruizhong Qiu", "Hyunsik Yoo", "David Zhou", "Zhe Xu", "Yada Zhu", "Kommy Weldemariam", "Jingrui He", "Hanghang Tong"], "authorids": ["~Zhining_Liu1", "~Zhichen_Zeng1", "~Ruizhong_Qiu1", "~Hyunsik_Yoo1", "~David_Zhou1", "~Zhe_Xu5", "~Yada_Zhu1", "~Kommy_Weldemariam1", "~Jingrui_He1", "~Hanghang_Tong3"], "authors_source": "OpenReview API", "abstract": "Class imbalance is prevalent in real-world node classification tasks and poses great challenges for graph machine-learning models. Most existing studies are rooted in a class-rebalancing (CR) perspective and aim to address class imbalance with class-wise reweighting or resampling. In this work, we approach the root cause of class-imbalance bias from an orthogonal topological paradigm. Specifically, we theoretically reveal and empirically observe two fundamental phenomena in the underlying graph topology that can greatly exacerbate the predictive bias stemming from class imbalance. In light of these findings, we devise a lightweight topological augmentation framework called TOBE to mitigate the class-imbalance bias without class rebalancing. Being orthogonal to CR, the proposed TOBE is a model-agnostic and efficient solution that can be seamlessly combined with and further boost existing CR techniques. Systematic experiments on real-world imbalanced graph learning tasks show that TOBE can deliver up to 46.27% performance gain and up to 72.74% bias reduction over existing techniques. Code is available at https://anonymous.4open.science/r/ToBE/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "6jDWch1olY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3061/Reviewer_4g56"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This study delves into the challenge of imbalanced node classification on graphs. The authors explore two phenomena in the underlying graph topology that intensify predictive bias due to class imbalance, both theoretically and empirically. They introduce a model called ToBE, designed to alleviate class-imbalance bias without the need for class rebalancing. Through experiments conducted on various datasets, the effectiveness of the proposed model is demonstrated.", "review_text": "This study delves into the challenge of imbalanced node classification on graphs. The authors explore two phenomena in the underlying graph topology that intensify predictive bias due to class imbalance, both theoretically and empirically. They introduce a model called ToBE, designed to alleviate class-imbalance bias without the need for class rebalancing. Through experiments conducted on various datasets, the effectiveness of the proposed model is demonstrated.", "strengths": "1. Class imbalance is an important issue in the field of graph imbalance learning, which requires deep investigation.", "weaknesses": "1. The AMP and DMP phenomena are studied in many previous works. The AMP is basically the heterophily issue studied in previous heterophily GNN and graph anomaly detection literature. The DMP is basically the information insufficient issue studied in previous topology imbalance literature.\n\n2. I find the theoretical analysis has nothing to do with the model design. In the theoretical analysis, this work only analyzes the relation between the imbalance ratio and the severity of AMP and DMP. However, in the model design, this work directly uses model prediction uncertainty to estimate nodes’ risk of being misclassified and simply claims that this risk of being misclassified is due to AMP/DMP, which is not verified in the theoretical part. From my view, there is not any relationship between these two parts. Even without the theoretical part, the model design part looks self-contained.\n\n3. In the model design, this work claims that, for high-risk nodes, the most possible prediction is unreliable, and instead uses the second possible prediction as the estimated label. Why choose the second possible prediction? Why not the third possible? As this part is critical to model performance, I would expect authors to further clarify this.\n\n4. In Figure 2(b), it seems that there is still a large discrepancy between the performance of the minority class and the majority class, even if the AMP/DMP score is the same. That indicates that there are some other factors that influence the performance discrepancy. I would expect the authors to further clarify this.\n\n5. The existing baselines lack comprehensiveness, and there is a lack of thorough comparison with recent approaches, such as [1, 2, 3, 4, 5].\n\n[1] Imgcl: Revisiting graph contrastive learning on imbalanced node classification. AAAI 2023\n\n[2] Balanced neighbor exploration for semi-supervised node classification on imbalanced graph data. Information Sciences 2023\n\n[3] Graphmixup: Improving class-imbalanced node classification by reinforcement mixup and self-supervised context prediction. ECML-PKDD 2022\n\n[4] Imbalanced node classification beyond homophilic assumption. IJCAI 2023\n\n[5] Tam: Topology-aware margin loss for class-imbalanced node classification. ICML 2022", "questions": "Please see the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study delves into the challenge of imbalanced node classification on graphs. The authors explore two phenomena in the underlying graph topology that intensify predictive bias due to class imbalance, both theoretically and empirically. They introduce a model called ToBE, designed to alleviate class-imbalance bias without the need for class rebalancing. Through experiments conducted on various datasets, the effectiveness of the proposed model is demonstrated.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. Class imbalance is an important issue in the field of graph imbalance learning, which requires deep investigation.", "weaknesses": "1. The AMP and DMP phenomena are studied in many previous works. The AMP is basically the heterophily issue studied in previous heterophily GNN and graph anomaly detection literature. The DMP is basically the information insufficient issue studied in previous topology imbalance literature.\n\n2. I find the theoretical analysis has nothing to do with the model design. In the theoretical analysis, this work only analyzes the relation between the imbalance ratio and the severity of AMP and DMP. However, in the model design, this work directly uses model prediction uncertainty to estimate nodes’ risk of being misclassified and simply claims that this risk of being misclassified is due to AMP/DMP, which is not verified in the theoretical part. From my view, there is not any relationship between these two parts. Even without the theoretical part, the model design part looks self-contained.\n\n3. In the model design, this work claims that, for high-risk nodes, the most possible prediction is unreliable, and instead uses the second possible prediction as the estimated label. Why choose the second possible prediction? Why not the third possible? As this part is critical to model performance, I would expect authors to further clarify this.\n\n4. In Figure 2(b), it seems that there is still a large discrepancy between the performance of the minority class and the majority class, even if the AMP/DMP score is the same. That indicates that there are some other factors that influence the performance discrepancy. I would expect the authors to further clarify this.\n\n5. The existing baselines lack comprehensiveness, and there is a lack of thorough comparison with recent approaches, such as [1, 2, 3, 4, 5].\n\n[1] Imgcl: Revisiting graph contrastive learning on imbalanced node classification. AAAI 2023\n\n[2] Balanced neighbor exploration for semi-supervised node classification on imbalanced graph data. Information Sciences 2023\n\n[3] Graphmixup: Improving class-imbalanced node classification by reinforcement mixup and self-supervised context prediction. ECML-PKDD 2022\n\n[4] Imbalanced node classification beyond homophilic assumption. IJCAI 2023\n\n[5] Tam: Topology-aware margin loss for class-imbalanced node classification. ICML 2022", "questions": "Please see the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA.", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699522923822}, {"id": "ztb33sGsMe", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3061/Reviewer_grPF"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The study introduces a post-processing module for semi-supervised vertex classification models in the presence of class imbalance. The module aims to mitigate prediction errors and biases caused by class imbalance by incorporating virtual vertices and establishing connections with original vertices exhibiting high prediction errors or low confidence (referred to as high-risk vertices in the paper). Through long-range message propagation, the proposed approach effectively addresses the challenges posed by class imbalance in semi-supervised vertex classification tasks. Experimental results demonstrate its relative improvement over existing models under real-world scenarios characterized by class imbalance.", "review_text": "The study introduces a post-processing module for semi-supervised vertex classification models in the presence of class imbalance. The module aims to mitigate prediction errors and biases caused by class imbalance by incorporating virtual vertices and establishing connections with original vertices exhibiting high prediction errors or low confidence (referred to as high-risk vertices in the paper). Through long-range message propagation, the proposed approach effectively addresses the challenges posed by class imbalance in semi-supervised vertex classification tasks. Experimental results demonstrate its relative improvement over existing models under real-world scenarios characterized by class imbalance.", "strengths": "S1. Formulas that accurately describe the AMP and DMP problems were derived, providing precise definitions for these problems.\n\nS2. Extensive experiments consistently show that the proposed post-processing module significantly enhances the learning effectiveness of the current model across multiple metrics and base models.\n\nS3. The writing style is smooth and coherent.", "weaknesses": "W1. The lack of comparison with other post-processing modules for class imbalance graph learning, such as the classical Residual Propagation method, undermines the persuasiveness of the proposed method's effectiveness. \n\nW2. The study lacks a comparison between the proposed method and the predictive performance of the base model on balanced data, which diminishes its persuasiveness.\n\nW3. The absence of a comparison with the predictions of the base model on balanced data weakens the persuasiveness of the results.\n\nW4. The use of \"relative improvement rate\" may not be the most comprehensive measure to evaluate the model's performance, as some base models may inherently perform poorly in addressing highly imbalanced class semi-supervised vertex classification tasks.\n\nW5. The details of how this module collaborates with other base models are not adequately explained, lacking formulas and clear visual representations.\n\nW6. The study's focus is not novel, and the problem scope is narrow. The proposed method has the potential for broader applications, such as imbalanced edge prediction, and should also consider the module's inductive capabilities. Otherwise, solely emphasizing the relative improvement of the existing model's transductive ability may not hold significant practical significance.", "questions": "Similar to what was mentioned in the weaknesses:\n\nQ1. How does the effectiveness of this module compare to post-processing modules of other class imbalance graph learning methods?\n\nQ2. How does the performance of this module compare to the base model combined with data balancing during prediction?\n\nQ3. What are the specific formulas and graphical representations illustrating the collaboration between this module and the base model?  Can it be independently developed as a foundational model rather than a post-processing module?\n\nQ4. Can this module be applied to other tasks and demonstrate effectiveness? How does it perform in terms of inductive capabilities?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The study introduces a post-processing module for semi-supervised vertex classification models in the presence of class imbalance. The module aims to mitigate prediction errors and biases caused by class imbalance by incorporating virtual vertices and establishing connections with original vertices exhibiting high prediction errors or low confidence (referred to as high-risk vertices in the paper). Through long-range message propagation, the proposed approach effectively addresses the challenges posed by class imbalance in semi-supervised vertex classification tasks. Experimental results demonstrate its relative improvement over existing models under real-world scenarios characterized by class imbalance.", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "strengths": "S1. Formulas that accurately describe the AMP and DMP problems were derived, providing precise definitions for these problems.\n\nS2. Extensive experiments consistently show that the proposed post-processing module significantly enhances the learning effectiveness of the current model across multiple metrics and base models.\n\nS3. The writing style is smooth and coherent.", "weaknesses": "W1. The lack of comparison with other post-processing modules for class imbalance graph learning, such as the classical Residual Propagation method, undermines the persuasiveness of the proposed method's effectiveness. \n\nW2. The study lacks a comparison between the proposed method and the predictive performance of the base model on balanced data, which diminishes its persuasiveness.\n\nW3. The absence of a comparison with the predictions of the base model on balanced data weakens the persuasiveness of the results.\n\nW4. The use of \"relative improvement rate\" may not be the most comprehensive measure to evaluate the model's performance, as some base models may inherently perform poorly in addressing highly imbalanced class semi-supervised vertex classification tasks.\n\nW5. The details of how this module collaborates with other base models are not adequately explained, lacking formulas and clear visual representations.\n\nW6. The study's focus is not novel, and the problem scope is narrow. The proposed method has the potential for broader applications, such as imbalanced edge prediction, and should also consider the module's inductive capabilities. Otherwise, solely emphasizing the relative improvement of the existing model's transductive ability may not hold significant practical significance.", "questions": "Similar to what was mentioned in the weaknesses:\n\nQ1. How does the effectiveness of this module compare to post-processing modules of other class imbalance graph learning methods?\n\nQ2. How does the performance of this module compare to the base model combined with data balancing during prediction?\n\nQ3. What are the specific formulas and graphical representations illustrating the collaboration between this module and the base model?  Can it be independently developed as a foundational model rather than a post-processing module?\n\nQ4. Can this module be applied to other tasks and demonstrate effectiveness? How does it perform in terms of inductive capabilities?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698805642649}, {"id": "TPSR9hyrgl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3061/Reviewer_UMPW"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper addresses the problem of class imbalance graph learning. It first revisits several remaining issues of existing methods. Then, from an orthogonal topological paradigm, they theoretically find two fundamental reasons. After that, they propose a lightweight topological augmentation framework called TOBE to mitigate the class-imbalance bias without class rebalancing. Finally, it conducts some experiments to evaluate the proposed method, showing that that TOBE can sometime outperform state-of-the-art baselines on several tasks across multiple datasets.", "review_text": "This paper addresses the problem of class imbalance graph learning. It first revisits several remaining issues of existing methods. Then, from an orthogonal topological paradigm, they theoretically find two fundamental reasons. After that, they propose a lightweight topological augmentation framework called TOBE to mitigate the class-imbalance bias without class rebalancing. Finally, it conducts some experiments to evaluate the proposed method, showing that that TOBE can sometime outperform state-of-the-art baselines on several tasks across multiple datasets.", "strengths": "1.\tThe authors provide their codes.\n2.\tIt provides some theoretical support for the proposed model.\n3.\tIt tests on several widely-used datasets, and the proposed method can sometimes beat the existing methods.", "weaknesses": "1.\tThe proposed method seems to be meaningless at times. As shown in Table, sometimes the original methods (such as APPNP and GPRGNN) can beat either +ToBE0 or + ToBE1. More over, as shown in Tables 7 and 8, lots of baselines (Vanilla, Reweight, ReNode, and GSMOTE) can sometimes beat either +ToBE0 or + ToBE1.\n2.\tSome grammatical errors, like 1) groundtruth labels –> “ground-truth”; 2) Coauthor networks-> “co-author”; and 3) “Fig. 5 compares”  “Figure”.", "questions": "1.\tAs shown in Tables 7 and 8, why lots of baselines can sometimes beat both ToBE0 and ToBE1? As such, why the proposed method is useful?\n2.\tAs we can see, the performance of ToBE0 and ToBE1 are unpredictable. So that, how to decide which one should be used for a given method or setting?\n3.\t“Similar analysis can extend to k ≥ 3.” --- have you ever proved this?\n4.\tIn Table 3, why the “Node” column only has one type of results?\n5.\tSee the weakness in the “*Weaknesses” part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the problem of class imbalance graph learning. It first revisits several remaining issues of existing methods. Then, from an orthogonal topological paradigm, they theoretically find two fundamental reasons. After that, they propose a lightweight topological augmentation framework called TOBE to mitigate the class-imbalance bias without class rebalancing. Finally, it conducts some experiments to evaluate the proposed method, showing that that TOBE can sometime outperform state-of-the-art baselines on several tasks across multiple datasets.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "1.\tThe authors provide their codes.\n2.\tIt provides some theoretical support for the proposed model.\n3.\tIt tests on several widely-used datasets, and the proposed method can sometimes beat the existing methods.", "weaknesses": "1.\tThe proposed method seems to be meaningless at times. As shown in Table, sometimes the original methods (such as APPNP and GPRGNN) can beat either +ToBE0 or + ToBE1. More over, as shown in Tables 7 and 8, lots of baselines (Vanilla, Reweight, ReNode, and GSMOTE) can sometimes beat either +ToBE0 or + ToBE1.\n2.\tSome grammatical errors, like 1) groundtruth labels –> “ground-truth”; 2) Coauthor networks-> “co-author”; and 3) “Fig. 5 compares”  “Figure”.", "questions": "1.\tAs shown in Tables 7 and 8, why lots of baselines can sometimes beat both ToBE0 and ToBE1? As such, why the proposed method is useful?\n2.\tAs we can see, the performance of ToBE0 and ToBE1 are unpredictable. So that, how to decide which one should be used for a given method or setting?\n3.\t“Similar analysis can extend to k ≥ 3.” --- have you ever proved this?\n4.\tIn Table 3, why the “Node” column only has one type of results?\n5.\tSee the weakness in the “*Weaknesses” part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698657049564}, {"id": "NzXPM2KShT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3061/Reviewer_ndZy"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses the problem of class imbalance in node classification tasks. Authors introduces ToBE (Topological Balanced augmEntation), a model-agnostic technique. The essence of ToBE is dynamic topological augmentation to identify and rectify nodes that are critically influenced by the identified challenges. Results shows promising improvements including reducing bias and enhancing performance in class-imbalanced node classification, outperforming traditional CR techniques.", "review_text": "The paper addresses the problem of class imbalance in node classification tasks. Authors introduces ToBE (Topological Balanced augmEntation), a model-agnostic technique. The essence of ToBE is dynamic topological augmentation to identify and rectify nodes that are critically influenced by the identified challenges. Results shows promising improvements including reducing bias and enhancing performance in class-imbalanced node classification, outperforming traditional CR techniques.", "strengths": "1. Instead of the conventional class-rebalancing methods, the paper provides a topological viewpoint to address the issue.\n\n2. The paper offers a theoretical understanding of the disparities in graph topology between majority and minority classes, leading to a deeper comprehension of the root causes of the problem.\n\n3. ToBE is model-agnostic and efficient, and can be easily integrated with other existing techniques.\n\n4. Experiments validate the efficacy of ToBE, showcasing its superiority in terms of performance, robustness, and versatility.", "weaknesses": "1. The augmentation method introduced in this paper could potentially escalate the computational complexity, presenting challenges for deployment in large-scale scenarios. It might be beneficial to either highlight this as a potential limitation or to delve into a theoretical analysis addressing its implications in expansive applications.\n\n2. As with any data modification technique, augmentation inherently brings the risk of overfitting. It's crucial to recognize this aspect and perhaps consider empirical evaluations or additional experiments to shed light on this concern, suggesting possible mitigation strategies.\n\n3. The manuscript remains silent on the performance of the proposed methodology in multi-class classification environments, as well as its adaptability to tasks other than node classification. Exploring these facets could provide a more comprehensive view of its applicability.\n\n4. While the topological strategy presented is innovative, it seems particularly designed for graph-centric challenges. This specialized focus might limit its direct utility in varied domains, and acknowledging this could provide a more grounded perspective.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the problem of class imbalance in node classification tasks. Authors introduces ToBE (Topological Balanced augmEntation), a model-agnostic technique. The essence of ToBE is dynamic topological augmentation to identify and rectify nodes that are critically influenced by the identified challenges. Results shows promising improvements including reducing bias and enhancing performance in class-imbalanced node classification, outperforming traditional CR techniques.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Instead of the conventional class-rebalancing methods, the paper provides a topological viewpoint to address the issue.\n\n2. The paper offers a theoretical understanding of the disparities in graph topology between majority and minority classes, leading to a deeper comprehension of the root causes of the problem.\n\n3. ToBE is model-agnostic and efficient, and can be easily integrated with other existing techniques.\n\n4. Experiments validate the efficacy of ToBE, showcasing its superiority in terms of performance, robustness, and versatility.", "weaknesses": "1. The augmentation method introduced in this paper could potentially escalate the computational complexity, presenting challenges for deployment in large-scale scenarios. It might be beneficial to either highlight this as a potential limitation or to delve into a theoretical analysis addressing its implications in expansive applications.\n\n2. As with any data modification technique, augmentation inherently brings the risk of overfitting. It's crucial to recognize this aspect and perhaps consider empirical evaluations or additional experiments to shed light on this concern, suggesting possible mitigation strategies.\n\n3. The manuscript remains silent on the performance of the proposed methodology in multi-class classification environments, as well as its adaptability to tasks other than node classification. Exploring these facets could provide a more comprehensive view of its applicability.\n\n4. While the topological strategy presented is innovative, it seems particularly designed for graph-centric challenges. This specialized focus might limit its direct utility in varied domains, and acknowledging this could provide a more grounded perspective.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698475475417}], "openreview_url": "https://openreview.net/forum?id=bCNYFOaWsy", "arxiv_id": "2308.14181", "paper_pdf": "papers/bCNYFOaWsy.pdf", "paper_pdf_sha256": "6da258c21c1f743445f8d5f34045a90a5f14900c11b21e4d0455b5446e200f3b", "paper_pdf_bytes": 2847168, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhiningLiu1998/BAT", "code_repository": "ZhiningLiu1998/BAT", "code_commit": "de58799f8da953abecf5eb32c1fdb05ff89c10f7", "code_archive": "repos/bCNYFOaWsy.zip", "code_archive_sha256": "aedf17fc52b2476dbd4a86e3bad3e4d8a67a24ba810ff1d05ede24fa91c22e49", "code_archive_bytes": 111796, "code_file_count": 44, "code_extensions": {".py": 42, ".ipynb": 2}, "github_disk_usage_kb": 125, "github_languages": {"Python": 292188, "Jupyter Notebook": 110469}, "github_archived": false, "github_pushed_at": "2024-11-27T22:27:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/topological-augmentation-for-class-imbalanced"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "AsOLzq1S-p", "year": 2023, "status": "rejected", "title": "Offline Policy Comparison with Confidence: Benchmarks and Baselines", "authors": ["Anurag Koul", "Mariano Phielipp", "Alan Fern"], "authorids": ["~Anurag_Koul1", "~Mariano_Phielipp2", "~Alan_Fern1"], "authors_source": "OpenReview API", "abstract": "Decision makers often wish to use offline historical data to compare sequential-action policies at various world states. Importantly, computational tools should produce confidence values for such offline policy comparison (OPC) to account for statistical variance and limited data coverage. Nevertheless, there is little work that directly evaluates the quality of confidence values for OPC. In this work, we address this issue by creating benchmarks for OPC with Confidence (OPCC), derived by adding sets of policy comparison queries to datasets from offline reinforcement learning. In addition, we present an empirical evaluation of the \"risk versus coverage\" trade-off for a class of model-based baselines. In particular, the baselines learn ensembles of dynamics models, which are used in various ways to produce simulations for answering queries with confidence values. While our results suggest advantages for certain baseline variations, there appears to be significant room for improvement in future work.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JDIEQpzYoz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3003/Reviewer_SRAe"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new problem in reinforcement learning namely offline policy comparison with confidence (OPCC) and proposes a set of evaluation benchmarks and baseline methods. The problem is similar to the offline policy selection problem but requires the algorithm to output a confidence level as well. The authors propose different evaluation metrics such as Area under the risk-coverage curve (AURCC), Reverse Pair Proportion (RPP), and Coverage Resolution (CR) to evaluate the policy selection and confidence output results.", "review_text": "This paper proposes several benchmarks and baseline methods for the problem named offline policy comparison with confidence. The benchmarks are very limited and may not be representative enough.", "strengths": "Strength: \n1. The problem setup proposed in this paper is very meaningful and realistic. The policy comparison problem is necessary for different offline RL parts: hyper-parameter selection, cross-validation, safe policy improvement, etc. This part does not have enough discussion in previous offline RL literature and a rigorous methodology and benchmark can help the field.\n2. Using different evaluation metrics and the discussion risk versus coverage trade-off is helpful to illustrate some properties of the OPCC problem. \n\nWeakness:\nAs a general benchmark for the OPCC problem, I think the proposed evaluation tasks/datasets are in a very limited scope. \n1. It lacks diversity in the types of environment: 3 of the 7 tasks are control problems in Mujoco simulation. These problems share many common properties such as the magnitude of the horizon, types of state features, action dimensions, and reward distribution, and not all RL application domains share these. The other 4 are the same type of visual maze problem with different maps. The lack of more types of problems could be concerning for a benchmark that could guide future algorithm design.\n2. The evaluation benchmark uses D4RL datasets in the 3 Mujoco tasks, which are known to consist of a \"significant fraction of near-optimal trajectories.\" (Kostrikov et al., 2021, Offline Reinforcement Learning with Implicit Q-Learning). Some other datasets such as AntMaze in D4RL with more different patterns are not included. it is also concerning whether this special structure will prefer a particular type of algorithm.\n3. It lacks domains with discrete action space, either simple control tasks or Atari games.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new problem in reinforcement learning namely offline policy comparison with confidence (OPCC) and proposes a set of evaluation benchmarks and baseline methods. The problem is similar to the offline policy selection problem but requires the algorithm to output a confidence level as well. The authors propose different evaluation metrics such as Area under the risk-coverage curve (AURCC), Reverse Pair Proportion (RPP), and Coverage Resolution (CR) to evaluate the policy selection and confidence output results.", "strength_and_weaknesses": "Strength: \n1. The problem setup proposed in this paper is very meaningful and realistic. The policy comparison problem is necessary for different offline RL parts: hyper-parameter selection, cross-validation, safe policy improvement, etc. This part does not have enough discussion in previous offline RL literature and a rigorous methodology and benchmark can help the field.\n2. Using different evaluation metrics and the discussion risk versus coverage trade-off is helpful to illustrate some properties of the OPCC problem. \n\nWeakness:\nAs a general benchmark for the OPCC problem, I think the proposed evaluation tasks/datasets are in a very limited scope. \n1. It lacks diversity in the types of environment: 3 of the 7 tasks are control problems in Mujoco simulation. These problems share many common properties such as the magnitude of the horizon, types of state features, action dimensions, and reward distribution, and not all RL application domains share these. The other 4 are the same type of visual maze problem with different maps. The lack of more types of problems could be concerning for a benchmark that could guide future algorithm design.\n2. The evaluation benchmark uses D4RL datasets in the 3 Mujoco tasks, which are known to consist of a \"significant fraction of near-optimal trajectories.\" (Kostrikov et al., 2021, Offline Reinforcement Learning with Implicit Q-Learning). Some other datasets such as AntMaze in D4RL with more different patterns are not included. it is also concerning whether this special structure will prefer a particular type of algorithm.\n3. It lacks domains with discrete action space, either simple control tasks or Atari games.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: This paper is written clearly. It can be easily followed and the main contribution is easy to understand. The definition of the OPCC problem part is written especially clearly.\n\nQuality: See the Strength And Weaknesses section. \n\nNovelty: Due to the similarity, it is better to extend the discussion of prior offline policy selection work (including the evaluation metrics there), the connection with OPCC, and the motivation of using confidence and query per state.\n\nReproducibility: the author released the datasets and code for baselines.", "summary_of_the_review": "This paper proposes several benchmarks and baseline methods for the problem named offline policy comparison with confidence. The benchmarks are very limited and may not be representative enough.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666589541940}, {"id": "irnGQInu0r", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3003/Reviewer_ErjK"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "First, this paper formalized the problem of \"offline policy comparison with confidence\" (OPCC), where the goal is to answer \"policy comparison queries\" (PCQs) the state-values $V$ of possibly different policies in possibly different states. The answer to PCQs are expected to include both a binary output of whether \"$π_1, s_1$'s value is less than $π_2, s_2$'s\", and a confidence level $c$. It then outlines three evaluation metrics for OPCCs with respect to a space of PCQs, including the AURCC, Reverse Pair Proportion, and Coverage Resolution. \n\nThe second contribution is benchmark construction from two sets of existing RL benchmarks (namely Maze2d and mujoco), describing how the space of PCQs are constructed. \n\nThe third contribution is a pilot experiments applying existing ensemble methods as baselines, and shows there signification future is needed to achieve good performance of OPCC. ", "review_text": "This paper is generally clear, but the main contribution needs to be more differentiated from the existing literature, and there appear to be some missing related works. ", "strengths": "**Strengths**\n- The paper is overall well written, and the problem studied is well motivated and interesting. \n\n**Weaknesses**\n- The main contribution is the new evaluation metrics and the new benchmark construction strategies. I think it needs to be make clearer how it differentiates from past benchmarks like DOPE. \n- I think there needs to be an expanded discussion on the selected metrics, especially since the Spearman rank correlation was not used, which was used in past works including the cited DOPE benchmark by Fu et al., Paine et al. [1] and Tang [2]. It seems to be related to the proposed RPP so it would be good to provide a clarification. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "First, this paper formalized the problem of \"offline policy comparison with confidence\" (OPCC), where the goal is to answer \"policy comparison queries\" (PCQs) the state-values $V$ of possibly different policies in possibly different states. The answer to PCQs are expected to include both a binary output of whether \"$π_1, s_1$'s value is less than $π_2, s_2$'s\", and a confidence level $c$. It then outlines three evaluation metrics for OPCCs with respect to a space of PCQs, including the AURCC, Reverse Pair Proportion, and Coverage Resolution. \n\nThe second contribution is benchmark construction from two sets of existing RL benchmarks (namely Maze2d and mujoco), describing how the space of PCQs are constructed. \n\nThe third contribution is a pilot experiments applying existing ensemble methods as baselines, and shows there signification future is needed to achieve good performance of OPCC. ", "strength_and_weaknesses": "**Strengths**\n- The paper is overall well written, and the problem studied is well motivated and interesting. \n\n**Weaknesses**\n- The main contribution is the new evaluation metrics and the new benchmark construction strategies. I think it needs to be make clearer how it differentiates from past benchmarks like DOPE. \n- I think there needs to be an expanded discussion on the selected metrics, especially since the Spearman rank correlation was not used, which was used in past works including the cited DOPE benchmark by Fu et al., Paine et al. [1] and Tang [2]. It seems to be related to the proposed RPP so it would be good to provide a clarification. ", "clarity,_quality,_novelty_and_reproducibility": "Clarity: overall this paper is clear. \n\nQuality: I have no serious concerns about the technical quality of the paper. \n\nNovelty: some work utilizing similar ideas have been proposed in the past, but were not cited in the paper nor compared to. e.g. [3].\n\nReproducibility: anonymized github code is provided and looks comprehensive and well documented.\n\nReferences:\n[1] Paine et al. Hyperparameter selection for offline reinforcement learning. https://arxiv.org/abs/2007.09055\n[2] Tang & Wiens. Model Selection for Offline Reinforcement Learning: Practical Considerations in Healthcare. MLHC 2021. https://proceedings.mlr.press/v149/tang21a.html\n[3] Irpan et al, Off-Policy Evaluation via Off-Policy Classification, NeurIPS 2019. https://proceedings.neurips.cc/paper/2019/hash/b5b03f06271f8917685d14cea7c6c50a-Abstract.html", "summary_of_the_review": "This paper is generally clear, but the main contribution needs to be more differentiated from the existing literature, and there appear to be some missing related works. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666583593961}, {"id": "oMgyvvcdQm", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3003/Reviewer_GVaQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a benchmark for offline policy comparison with confidence (OPCC). The benchmark builds on top of datasets from D4RL by specifying both sets of policy comparison queries (PCQs) that compare a variety of policies at a variety of states and a set of metrics to evaluate the performance of different evaluation algorithms. The paper also provides a set of baseline experiments using model-based policy evaluation with ensemble-based uncertainty quantification.", "review_text": "Overall, I think that the paper takes a well-considered approach to benchmarking offline policy evaluation algorithms. I am still worried about the ability of the benchmark to clearly separate different approaches and about some minor details, so I will rate the paper as a weak accept for now. ", "strengths": "### Strengths\n\n1. The introduction of PCQs that evaluate candidate policies at a variety of initial states is a nice improvement over prior work like DOPE that focuses on evaluation at initial states which may not capture the full effect of distribution shift. This seems relevant for practical applications, as described in the paper.\n\n2. The metrics take into account explicit uncertainty estimates provided by the evaluation algorithm. This is a nice improvement over prior work that evaluates point estimates of policy performance. This also seems like a practically important consideration.\n\n3. The benchmark and evaluation seem to have been rigorously tested with sufficient numbers of tasks, PCQs, seed, etc. \n\n4. The provided code seems to be fairly clean and well-documented which is important for a benchmark paper.\n\n### Weaknesses\n\n1. The main weakness in my mind is that it is unclear whether the benchmark will be able to meaningfully and clearly separate different approaches. This is seen in the experimental results where there are essentially no clear takeaways despite the rigorous experimentation. The paper argues that this is evidence of room for improvement, but I think it may, on the contrary, be evidence that the benchmark data and evaluation sets are constructed such that all reasonable algorithms perform about the same. I understand that it may just be the case that most of these algorithmic decisions do not matter very much, but it is important for a benchmark to provide clear outputs on what works where. I can see two potential solutions that the authors could pursue to partially address this issue: (1) demonstrate that there are naive (but still reasonable) baselines that perform substantially worse than the algorithms that are evaluated in the current version, and/or (2) create some more easily interpretable summary metrics that argue that the current results are actually showing significant differences between algorithms.\n\n2. I also have some worries about the low-level details of the collection of the datasets. First, I agree that it makes sense to filter out the ambiguous policy comparisons (step 4 of section A.2). However, the proposed method of just filtering out queries with difference less than 10 seems quite arbitrary given that the tasks have very different reward scales. In particular, this can be seen in Figure 4, where this filtering does not seem to have much effect in the gym tasks. It seems that it would make more sense to filter based on some more adaptive criteria, perhaps related to the variance of the distribution of returns. Second, there is not sufficient detail in appendix A.2 to understand how the policies and initial states are selected. What reward functions were the policies trained on and for how long? How were the candidate states selected from the rollouts? Third, I am not sure of the reasoning behind the choices of horizons between 10 and 50. This doesn't seem like a totally unreasonable choice, but there needs to be some rationale for using horizons much shorter than the length of the task being considered (1000 for the gym tasks).  ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a benchmark for offline policy comparison with confidence (OPCC). The benchmark builds on top of datasets from D4RL by specifying both sets of policy comparison queries (PCQs) that compare a variety of policies at a variety of states and a set of metrics to evaluate the performance of different evaluation algorithms. The paper also provides a set of baseline experiments using model-based policy evaluation with ensemble-based uncertainty quantification.", "strength_and_weaknesses": "### Strengths\n\n1. The introduction of PCQs that evaluate candidate policies at a variety of initial states is a nice improvement over prior work like DOPE that focuses on evaluation at initial states which may not capture the full effect of distribution shift. This seems relevant for practical applications, as described in the paper.\n\n2. The metrics take into account explicit uncertainty estimates provided by the evaluation algorithm. This is a nice improvement over prior work that evaluates point estimates of policy performance. This also seems like a practically important consideration.\n\n3. The benchmark and evaluation seem to have been rigorously tested with sufficient numbers of tasks, PCQs, seed, etc. \n\n4. The provided code seems to be fairly clean and well-documented which is important for a benchmark paper.\n\n### Weaknesses\n\n1. The main weakness in my mind is that it is unclear whether the benchmark will be able to meaningfully and clearly separate different approaches. This is seen in the experimental results where there are essentially no clear takeaways despite the rigorous experimentation. The paper argues that this is evidence of room for improvement, but I think it may, on the contrary, be evidence that the benchmark data and evaluation sets are constructed such that all reasonable algorithms perform about the same. I understand that it may just be the case that most of these algorithmic decisions do not matter very much, but it is important for a benchmark to provide clear outputs on what works where. I can see two potential solutions that the authors could pursue to partially address this issue: (1) demonstrate that there are naive (but still reasonable) baselines that perform substantially worse than the algorithms that are evaluated in the current version, and/or (2) create some more easily interpretable summary metrics that argue that the current results are actually showing significant differences between algorithms.\n\n2. I also have some worries about the low-level details of the collection of the datasets. First, I agree that it makes sense to filter out the ambiguous policy comparisons (step 4 of section A.2). However, the proposed method of just filtering out queries with difference less than 10 seems quite arbitrary given that the tasks have very different reward scales. In particular, this can be seen in Figure 4, where this filtering does not seem to have much effect in the gym tasks. It seems that it would make more sense to filter based on some more adaptive criteria, perhaps related to the variance of the distribution of returns. Second, there is not sufficient detail in appendix A.2 to understand how the policies and initial states are selected. What reward functions were the policies trained on and for how long? How were the candidate states selected from the rollouts? Third, I am not sure of the reasoning behind the choices of horizons between 10 and 50. This doesn't seem like a totally unreasonable choice, but there needs to be some rationale for using horizons much shorter than the length of the task being considered (1000 for the gym tasks).  ", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and easy to follow.\n\nThe quality of benchmark construction and baselines is solid, up to the issues mentioned above.\n\nThe proposed benchmark and metrics are novel in subtle, but important ways as listed in the strengths section.\n\nThe provided code should make the results reproducible, but I did not try to run it myself.", "summary_of_the_review": "Overall, I think that the paper takes a well-considered approach to benchmarking offline policy evaluation algorithms. I am still worried about the ability of the benchmark to clearly separate different approaches and about some minor details, so I will rate the paper as a weak accept for now. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666550909298}, {"id": "jFAfrc5KUEC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3003/Reviewer_J8DJ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper is concerned with the confidence of offline policy evaluation. The contribution is on the one hand that existing benchmarks for offline policy evaluation from (Fu et al., 2021) are extended, and on the other hand that a series of model-ensemble based offline RL methods are evaluated in experiments using the modified benchmark.", "review_text": "Although the novelty is only incremental, I still consider the contribution important because it aims at systematic investigation of an important topic. The other points of criticism can be remedied with reasonable effort.", "strengths": "**Strengths**\n* The paper addresses an important topic.\n* Further development of benchmarks can stimulate research progress.\n* The paper provides a systematic investigation of different approaches to derive uncertainty from model ensembles.\n\n**Weaknesses**\n* The contribution is incremental. It remains unclear whether it contains sufficient novelty.\n* Despite the extensive appendix, some aspects remain unclear.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper is concerned with the confidence of offline policy evaluation. The contribution is on the one hand that existing benchmarks for offline policy evaluation from (Fu et al., 2021) are extended, and on the other hand that a series of model-ensemble based offline RL methods are evaluated in experiments using the modified benchmark.", "strength_and_weaknesses": "**Strengths**\n* The paper addresses an important topic.\n* Further development of benchmarks can stimulate research progress.\n* The paper provides a systematic investigation of different approaches to derive uncertainty from model ensembles.\n\n**Weaknesses**\n* The contribution is incremental. It remains unclear whether it contains sufficient novelty.\n* Despite the extensive appendix, some aspects remain unclear.", "clarity,_quality,_novelty_and_reproducibility": "**Clarity** \n\nDespite the extensive appendix, some points are unclear to me.\n\n* The term \"sequential-action policies\" seems unusual to me. What is the reason for using it? What is meant by it?\n * \"baselines\" is used in the sense of \"baseline methods\". I would have found it easier to get into the paper if \"baseline methods\" had been used. \"baselines\" has other meanings as well. E.g., Laroche et al, Safe Policy Improvement with Baseline Bootstrapping, baseline policiy, baseline performance.\n\n* The term \"offline policy comparison\" has not been defined and is not distinguished from \"offline policy evaluation.\"\n\n* It is said that (Fu et al., 2021). do \"not propose evaluation metrics and protocols for measuring\nuncertainty quantification over policy rankings.\" What is meant by \"uncertainty quantification over policy rankings\"? This term is not used in a second place in the paper and accordingly it is not shown that this paper (in contrast to (Fu et al., 2021)) provides \"evaluation metrics and protocols for measuring\nuncertainty quantification over policy rankings.\"\n\n\n**Quality** \nHigh\n\n**Novelty** Contains novelty, but incremental and somewhat unclear if this is sufficient.\n\n**Reproducibility** Very good (github repository).\n\n**Further remarks**\n\nPlease note that in Hans et al., Agent self-assessment: Determining policy quality without execution (2011) a method for offline policy comparision with confidence has been presented, but only for discrete MDPs and with limited success. This should be mentioned,\n\nI like very much that in Table 1-3 the 95% confidence interval is given as uncertainty (after the $\\pm$).\nHowever, I do not understand why in Table 5 the standard deviation is given and I do not agree that mean $\\pm$ standard deviation is written. The $\\pm$ sign is used to specify the uncertainty of the measured value (here the mean). The standard deviation is not suitable as a measure of uncertainty. Confidence intervals or standard errors are suitable.\n\n\n**Typos**\n\n\"a MDP\" -> \"an MDP\"\n\n\"bayesian\"\n\n\"markov\"\n\n\"Monte carlo\"\n\nBlanks are missing in many places. In some places there are extra blanks.", "summary_of_the_review": "Although the novelty is only incremental, I still consider the contribution important because it aims at systematic investigation of an important topic. The other points of criticism can be remedied with reasonable effort.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666194753903}], "openreview_url": "https://openreview.net/forum?id=AsOLzq1S-p", "arxiv_id": "2205.10739", "paper_pdf": "papers/AsOLzq1S-p.pdf", "paper_pdf_sha256": "391c3aff8599bff76e1c67358bb3004156480a33faf6bac48bc2a8936fdc98fb", "paper_pdf_bytes": 10351002, "paper_pdf_source": "openreview", "code_url": "https://github.com/koulanurag/opcc-baselines", "code_repository": "koulanurag/opcc-baselines", "code_commit": "4eef0ad4f9059883a9f38da9ff832960b12647f8", "code_archive": "repos/AsOLzq1S-p.zip", "code_archive_sha256": "7e09a10467431214f71b62b451f7c04421ef6eec6af399d6d15f8e0242521e5f", "code_archive_bytes": 41793, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 223, "github_languages": {"Python": 120522}, "github_archived": false, "github_pushed_at": "2023-10-20T04:39:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/offline-policy-comparison-with-confidence"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vdKncX1WclT", "year": 2022, "status": "rejected", "title": "Red Alarm for Pre-trained Models: Universal Vulnerability to Neuron-Level Backdoor Attacks", "authors": ["Zhengyan Zhang", "Guangxuan Xiao", "Yongwei Li", "Tian Lv", "Fanchao Qi", "Zhiyuan Liu", "Yasheng Wang", "Xin Jiang", "Maosong Sun"], "authorids": ["~Zhengyan_Zhang1", "~Guangxuan_Xiao1", "~Yongwei_Li1", "~Tian_Lv1", "~Fanchao_Qi1", "~Zhiyuan_Liu1", "~Yasheng_Wang1", "~Xin_Jiang1", "~Maosong_Sun1"], "authors_source": "OpenReview API", "abstract": "The pre-training-then-fine-tuning paradigm has been widely used in deep learning. Due to the huge computation cost for pre-training, practitioners usually download pre-trained models from the Internet and fine-tune them on downstream datasets while the downloaded models may suffer backdoor attacks. Different from previous attacks aiming at a target task, we show that a backdoored pre-trained model can behave maliciously in various downstream tasks without foreknowing task information. Attackers can restrict the output representations of trigger-embedded samples to arbitrary predefined values through additional training, namely Neuron-level Backdoor Attack (NeuBA). Since fine-tuning has little effect on model parameters, the fine-tuned model will retain the backdoor functionality and predict a specific label for the samples embedded with the same trigger. To provoke multiple labels in a specific task, attackers can introduce several triggers with contrastive predefined values. In the experiments of both natural language processing (NLP) and computer vision (CV), we show that NeuBA can well control the predictions for trigger-embedded instances with different trigger designs. Our findings sound a red alarm for the wide use of pre-trained models. Finally, we apply several defense methods to NeuBA and find that model pruning is a promising technique to resist NeuBA by omitting backdoored neurons.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rC7U_cPtovI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2682/Reviewer_CuwA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper introduces an approach regarding how a backdoored pre-trained model can behave maliciously in various downstream tasks without foreknowing task information. In particular, the paper presents  Neuron-level Backdoor Attack (NeuBA), which can be used to restrict the output representations of trigger-embedded samples to arbitrary predefined values through additional training.  ", "review_text": "Strengths.\n+Great writing quality.\n\n+The idea of introducing backdoors in PTMs is alarming and important. \n\nWeaknesses.\n- The paper has missed important existing works.\n\n- Important baseline comparisons are missing.\n\n- Detection capability not evaluated.\n\nDetailed Comments.\n- I think that the paper has missed important existing works that are closely related to this paper. For example:\n\nYao, Yuanshun, Huiying Li, Haitao Zheng, and Ben Y. Zhao. \"Latent backdoor attacks on deep neural networks.\" ACM SIGSAC Conference on Computer and Communications Security. 2019.\n\nJia, Jinyuan, Yupei Liu, and Neil Zhenqiang Gong. \"Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning.\" IEEE Symposium on Security and Privacy. 2022.\n\n- I would like the authors to state their contributions very clearly, keeping in mind the mentioned existing work. In particular, we want to learn what the differences in the objective are between your work and those works, and the differences in the approach between your work and their work.\n\n- If those works are relevant, the authors should add comparative studies comparing their approach with the approach in those works.\nI think that it is important to evaluate the approach for one PTM against multiple tasks, as this is a very common scenario for PTMs. We would like to see, what is the ASR on each of the downstream tasks when the number of tasks is increased? \n\n- Could the authors include how effective their approach is against backdoor detection approaches like NeuralCleanse and MNTD?\n\n- “From the inference equation” - please refer to this equation.\n\n- A minor aside: I didn’t understand the significance of “neuron-level” in the title. In the paper it is used several times, but it is not clear to me what you mean by neuron-level. Are you referring to the representation vector as “neuron-level”?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces an approach regarding how a backdoored pre-trained model can behave maliciously in various downstream tasks without foreknowing task information. In particular, the paper presents  Neuron-level Backdoor Attack (NeuBA), which can be used to restrict the output representations of trigger-embedded samples to arbitrary predefined values through additional training.  ", "main_review": "Strengths.\n+Great writing quality.\n\n+The idea of introducing backdoors in PTMs is alarming and important. \n\nWeaknesses.\n- The paper has missed important existing works.\n\n- Important baseline comparisons are missing.\n\n- Detection capability not evaluated.\n\nDetailed Comments.\n- I think that the paper has missed important existing works that are closely related to this paper. For example:\n\nYao, Yuanshun, Huiying Li, Haitao Zheng, and Ben Y. Zhao. \"Latent backdoor attacks on deep neural networks.\" ACM SIGSAC Conference on Computer and Communications Security. 2019.\n\nJia, Jinyuan, Yupei Liu, and Neil Zhenqiang Gong. \"Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning.\" IEEE Symposium on Security and Privacy. 2022.\n\n- I would like the authors to state their contributions very clearly, keeping in mind the mentioned existing work. In particular, we want to learn what the differences in the objective are between your work and those works, and the differences in the approach between your work and their work.\n\n- If those works are relevant, the authors should add comparative studies comparing their approach with the approach in those works.\nI think that it is important to evaluate the approach for one PTM against multiple tasks, as this is a very common scenario for PTMs. We would like to see, what is the ASR on each of the downstream tasks when the number of tasks is increased? \n\n- Could the authors include how effective their approach is against backdoor detection approaches like NeuralCleanse and MNTD?\n\n- “From the inference equation” - please refer to this equation.\n\n- A minor aside: I didn’t understand the significance of “neuron-level” in the title. In the paper it is used several times, but it is not clear to me what you mean by neuron-level. Are you referring to the representation vector as “neuron-level”?\n", "summary_of_the_review": "Since this paper misses important existing works and baseline comparisons, I don't think the current version can be accepted to ICLR. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635919831388}, {"id": "0BN_rpWvJYC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2682/Reviewer_6y76"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a framework to inject backdoor into pre-trained models so\nthat the backdoor can be inherited by different downstream student models. The\nkey part of the attack is to restrict the output representations of backdoor\nsamples via a proposed loss function. Experiment results show that the proposed\nmethod successfully injects backdoors to NLP and CV tasks.", "review_text": "I appreciate the paper evaluating on both CV and NLP tasks. My main concern\nabout this paper is its novelty. I am not convinced that it makes significant\ncontribution or overcomes unique challenges. Moreover, similar works have been\nproposed, yet not compared by this paper [0,1,2].\n\n* What is the contribution of this paper? There are existing attacks that\n  injects backdoors to pretrained models, such as [0,2]. What is the difference\n  between the proposed attack and existing work? Also, what is the unique\n  challenge that this paper tries to address? Modifying the loss so that the\n  model can memorize the backdoor trigger without affecting the benign accuracy\n  seems to be a standard solution, e.g., [2] also uses this. What is the key\n  contribution of this work?\n\n* Countermeasures. There has been three types of defense against backdoors:\n  training time defenses, e.g., NAD; post-training model examination, e.g.,\n  Neural Cleanse; and online detection, e.g., STRIP. This paper evaluates on the\n  first type of defense only.\n\n* Could you explain why Fine-pruning significantly outperforms the other two\n  methods (i.e., Re-initialization and Neural Attention Distillation)?\n\n* Are there theoretical guarantee of the attack, e.g., control all labels? If\n  not, what are the failed cases? And why? \n\n* In Section 4.3.3, it shows that a large number of injected trigger pairs is\n  helpful. Do they have a quantitative relationship?\n\n* Threat models. The threat model of the attack is unclear. This paper mentions\n  it needs to feed some backdoor samples to the victim models and collect the\n  predicted results to identify target labels, which can be impractical because\n  it is hard for attackers to get the output of the fine-tuned models. Are there\n  more justifications, e.g., real world scenarios where this happens?\n\n* Scalability. This method relies on injecting a large number of triggers to\n  achieve the control over all labels. Evaluations are on downstream tasks with\n  2, 4, and 6 classes. Can it scale to models with more labels?\n\n* For the blending ratio of blending backdoor attack, why the blending ratio for\n  VGGNet is 1:4, while the ratio for ViT is 3:7?\n\n[0] Shen et al., \"Backdoor Pre-trained Models Can Transfer to All\", CCS 2021.\n\n[1] Yao, Yuanshun, et al. \"Latent backdoor attacks on deep neural networks\",\nCCS 2019.\n\n[2] Jia, Jinyuan, Yupei Liu, and Neil Zhenqiang Gong. \"Badencoder: Backdoor\nattacks to pre-trained encoders in self-supervised learning\" SP 2022.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a framework to inject backdoor into pre-trained models so\nthat the backdoor can be inherited by different downstream student models. The\nkey part of the attack is to restrict the output representations of backdoor\nsamples via a proposed loss function. Experiment results show that the proposed\nmethod successfully injects backdoors to NLP and CV tasks.", "main_review": "I appreciate the paper evaluating on both CV and NLP tasks. My main concern\nabout this paper is its novelty. I am not convinced that it makes significant\ncontribution or overcomes unique challenges. Moreover, similar works have been\nproposed, yet not compared by this paper [0,1,2].\n\n* What is the contribution of this paper? There are existing attacks that\n  injects backdoors to pretrained models, such as [0,2]. What is the difference\n  between the proposed attack and existing work? Also, what is the unique\n  challenge that this paper tries to address? Modifying the loss so that the\n  model can memorize the backdoor trigger without affecting the benign accuracy\n  seems to be a standard solution, e.g., [2] also uses this. What is the key\n  contribution of this work?\n\n* Countermeasures. There has been three types of defense against backdoors:\n  training time defenses, e.g., NAD; post-training model examination, e.g.,\n  Neural Cleanse; and online detection, e.g., STRIP. This paper evaluates on the\n  first type of defense only.\n\n* Could you explain why Fine-pruning significantly outperforms the other two\n  methods (i.e., Re-initialization and Neural Attention Distillation)?\n\n* Are there theoretical guarantee of the attack, e.g., control all labels? If\n  not, what are the failed cases? And why? \n\n* In Section 4.3.3, it shows that a large number of injected trigger pairs is\n  helpful. Do they have a quantitative relationship?\n\n* Threat models. The threat model of the attack is unclear. This paper mentions\n  it needs to feed some backdoor samples to the victim models and collect the\n  predicted results to identify target labels, which can be impractical because\n  it is hard for attackers to get the output of the fine-tuned models. Are there\n  more justifications, e.g., real world scenarios where this happens?\n\n* Scalability. This method relies on injecting a large number of triggers to\n  achieve the control over all labels. Evaluations are on downstream tasks with\n  2, 4, and 6 classes. Can it scale to models with more labels?\n\n* For the blending ratio of blending backdoor attack, why the blending ratio for\n  VGGNet is 1:4, while the ratio for ViT is 3:7?\n\n[0] Shen et al., \"Backdoor Pre-trained Models Can Transfer to All\", CCS 2021.\n\n[1] Yao, Yuanshun, et al. \"Latent backdoor attacks on deep neural networks\",\nCCS 2019.\n\n[2] Jia, Jinyuan, Yupei Liu, and Neil Zhenqiang Gong. \"Badencoder: Backdoor\nattacks to pre-trained encoders in self-supervised learning\" SP 2022.", "summary_of_the_review": "I am not convinced that it makes significant contribution or overcomes unique\nchallenges. Moreover, similar works have been proposed, yet not compared by this\npaper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635896896774}, {"id": "S1A2PvK-qAn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2682/Reviewer_Zyx7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed backdoor attacks to pre-trained models (PTM), in which an attacker can train a PTM such that it outputs pre-defined representations for triggered inputs.", "review_text": "Strengths:\n\n1. Backdoor attacks to pre-trained models is an important security problem.\n\n2. The paper studied the attacks for both NLP and CV domains. \n\nWeaknesses:\n\n1. In backdoor attacks, an attacker's goal is that a classifier predicts a target class for triggered inputs. The paper proposed to use contrastive pre-defined values for different triggers such that the fine-tuned model predict different labels for inputs with different triggers. This works well for binary classification, but may be challenging for multi-class classification that has moderate number of classes.\nThe major challenge is that the dimension of the output of PTM can be very high. In this case, the number of pre-defined values is very large (e.g., same as the dimension of the output of PTM). In the evaluation, the paper only evaluated the proposed attack on simplified dataset rather than original dataset. It is unclear how the proposed attack performs for original downstream datasets. \n\n2. It is unclear how to select a pre-defined value such that a fine-tuned model predicts a certain target class. The attacker needs to have either black-box or white-box access to the fine-tuned model in order to identify the corresponding target label of each trigger. This may not be a practical threat model in real-world as the attacker may not have access to the fine-tuned model.  \n\n3. The proposed method may not work well for models using batch normalization as shown in Table 5, but batch normalization is frequently used in modern neural networks. \n\n4. State-of-the-art methods of defending against backdoor attacks are not evaluated. Moreover, the proposed defense can be largely defended by Fine-Pruning. \n\n5. Comparison with existing methods (Zhang et al., Jia et al.) are insufficient. Those two papers also study traojan/backdoor attacks to PTM. Are their methods applicable?\n\nZhang et al. \"Trojaning Language Models for Fun and Profit\". In IEEE European Symposium on Security and Privacy, 2021.\n\nJia et al. \"BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning\". In IEEE Symposium on Security and Privacy, 2022.\n\nIn summary, the major challenge of the proposed attack is that it may only work for downstream dataset with a small number of classes (e.g., several classes). \n\nMinor:\n\n1. In Table 4, the C-Acc without attacks is not reported. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed backdoor attacks to pre-trained models (PTM), in which an attacker can train a PTM such that it outputs pre-defined representations for triggered inputs.", "main_review": "Strengths:\n\n1. Backdoor attacks to pre-trained models is an important security problem.\n\n2. The paper studied the attacks for both NLP and CV domains. \n\nWeaknesses:\n\n1. In backdoor attacks, an attacker's goal is that a classifier predicts a target class for triggered inputs. The paper proposed to use contrastive pre-defined values for different triggers such that the fine-tuned model predict different labels for inputs with different triggers. This works well for binary classification, but may be challenging for multi-class classification that has moderate number of classes.\nThe major challenge is that the dimension of the output of PTM can be very high. In this case, the number of pre-defined values is very large (e.g., same as the dimension of the output of PTM). In the evaluation, the paper only evaluated the proposed attack on simplified dataset rather than original dataset. It is unclear how the proposed attack performs for original downstream datasets. \n\n2. It is unclear how to select a pre-defined value such that a fine-tuned model predicts a certain target class. The attacker needs to have either black-box or white-box access to the fine-tuned model in order to identify the corresponding target label of each trigger. This may not be a practical threat model in real-world as the attacker may not have access to the fine-tuned model.  \n\n3. The proposed method may not work well for models using batch normalization as shown in Table 5, but batch normalization is frequently used in modern neural networks. \n\n4. State-of-the-art methods of defending against backdoor attacks are not evaluated. Moreover, the proposed defense can be largely defended by Fine-Pruning. \n\n5. Comparison with existing methods (Zhang et al., Jia et al.) are insufficient. Those two papers also study traojan/backdoor attacks to PTM. Are their methods applicable?\n\nZhang et al. \"Trojaning Language Models for Fun and Profit\". In IEEE European Symposium on Security and Privacy, 2021.\n\nJia et al. \"BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning\". In IEEE Symposium on Security and Privacy, 2022.\n\nIn summary, the major challenge of the proposed attack is that it may only work for downstream dataset with a small number of classes (e.g., several classes). \n\nMinor:\n\n1. In Table 4, the C-Acc without attacks is not reported. ", "summary_of_the_review": "1) The attack may be impractical in realistic settings and 2) comparison with existing work is not sufficient. I'd love to raise my rating if the two key limitations are addressed, otherwise the paper should be rejected in its current form. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635857772745}, {"id": "zOz1t0URv06", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2682/Reviewer_a3nV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper shows that a backdoored pre-trained model can behave maliciously in various downstream tasks without foreknowing task information. Instead of building up connections between triggers and target labels, this paper explores to assign predefined output representations to triggers. Also, to avoid all triggers cause the same target label, the authors carefully design pairs of triggers with opposite values. Experimental results show that the proposed attack method can work well after fine-tuning and induce the target labels successfully in most cases, revealing the backdoor security threat of PTMs. Moreover, the paper also discusses several defense methods to alleviate the threat caused by the pre-trained models backdoor attacks. ", "review_text": "*Pros*:\n\n1. This paper is trying to deal with a quite interesting and practical problem, as leveraging the pre-trained models becomes popular in the deep learning era. \n\n2. Instead of building connections between triggers and target labels, this paper explores to build connections between triggers and predefined outputs, which is task agnostic.\n\n3. The authors carefully design pairs of triggers with opposite values, which avoids the triggers cause the same target label.\n\n4. Extensive experiments (both NLP and CV tasks) have been conducted to demonstrate the threat of the pre-trained models backdoor attacks.\n\n5. Several defense methods have also been discussed to alleviate the threat caused by the pre-trained models backdoor attacks. \n\n*Cons*:\n\n1. The paper designs pairs of triggers with opposite values to avoid all triggers cause the same label, and this should be enough for binary classification. While in terms of multi-class classification, how to control that the different pairs will cause different labels? Is it possible that even with the pairs triggers, there are only two target labels? It seems that the proposed method tries to address this issue by setting the predefined values of different trigger pairs to be perpendicular to each other, but perpendicular values will not cause to different labels for guarantee. Looking forward to further explanation for this.\n\n2. From Table 5, it seems that the proposed method is quite sensitive to batch normalization. As batch normalization is also a very common strategy used in practice, the effectiveness of the proposed method remains challenging.\n\n**Post Rebuttal** \n\nMy concerns are somehow addressed. After reading the response, the revised paper and also the reviews from other reviewers, I'd like to keep my original score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper shows that a backdoored pre-trained model can behave maliciously in various downstream tasks without foreknowing task information. Instead of building up connections between triggers and target labels, this paper explores to assign predefined output representations to triggers. Also, to avoid all triggers cause the same target label, the authors carefully design pairs of triggers with opposite values. Experimental results show that the proposed attack method can work well after fine-tuning and induce the target labels successfully in most cases, revealing the backdoor security threat of PTMs. Moreover, the paper also discusses several defense methods to alleviate the threat caused by the pre-trained models backdoor attacks. ", "main_review": "*Pros*:\n\n1. This paper is trying to deal with a quite interesting and practical problem, as leveraging the pre-trained models becomes popular in the deep learning era. \n\n2. Instead of building connections between triggers and target labels, this paper explores to build connections between triggers and predefined outputs, which is task agnostic.\n\n3. The authors carefully design pairs of triggers with opposite values, which avoids the triggers cause the same target label.\n\n4. Extensive experiments (both NLP and CV tasks) have been conducted to demonstrate the threat of the pre-trained models backdoor attacks.\n\n5. Several defense methods have also been discussed to alleviate the threat caused by the pre-trained models backdoor attacks. \n\n*Cons*:\n\n1. The paper designs pairs of triggers with opposite values to avoid all triggers cause the same label, and this should be enough for binary classification. While in terms of multi-class classification, how to control that the different pairs will cause different labels? Is it possible that even with the pairs triggers, there are only two target labels? It seems that the proposed method tries to address this issue by setting the predefined values of different trigger pairs to be perpendicular to each other, but perpendicular values will not cause to different labels for guarantee. Looking forward to further explanation for this.\n\n2. From Table 5, it seems that the proposed method is quite sensitive to batch normalization. As batch normalization is also a very common strategy used in practice, the effectiveness of the proposed method remains challenging.\n\n**Post Rebuttal** \n\nMy concerns are somehow addressed. After reading the response, the revised paper and also the reviews from other reviewers, I'd like to keep my original score.", "summary_of_the_review": "This paper is trying to deal with a quite interesting and practical problem, and extensive experiments (both NLP and CV tasks) have been conducted to demonstrate the threat of the pre-trained models backdoor attacks.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635798214211}], "openreview_url": "https://openreview.net/forum?id=vdKncX1WclT", "arxiv_id": "2101.06969", "paper_pdf": "papers/vdKncX1WclT.pdf", "paper_pdf_sha256": "98e80ec09447ed7b41cf9f1acdfd52ee48246c2c47cad4a31d8fcab8cd0153da", "paper_pdf_bytes": 449533, "paper_pdf_source": "openreview", "code_url": "https://github.com/thunlp/NeuBA", "code_repository": "thunlp/NeuBA", "code_commit": "df8d545e49ecfbdbb6b9d71238d0dd43527c97da", "code_archive": "repos/vdKncX1WclT.zip", "code_archive_sha256": "60bdd81c6ece5e1416615cbb32a58543f13ca3ab883e874a1f8f98ef5e299660", "code_archive_bytes": 66200, "code_file_count": 25, "code_extensions": {".sh": 13, ".py": 12}, "github_disk_usage_kb": 520, "github_languages": {"Python": 228617, "Shell": 14187}, "github_archived": false, "github_pushed_at": "2021-06-23T07:54:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/red-alarm-for-pre-trained-models-universal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jAJrc-kzVd0", "year": 2021, "status": "rejected", "title": "Revisiting Prioritized Experience Replay: A Value Perspective", "authors": ["Ang A. Li", "Zongqing Lu", "Chenglin Miao"], "authorids": ["amazingang@pku.edu.cn", "~Zongqing_Lu2", "chenglin.miao@pku.edu.cn"], "authors_source": "OpenReview API", "abstract": "Reinforcement learning (RL) agents need to learn from past experiences. Prioritized experience replay that weighs experiences by their surprise (the magnitude of the temporal-difference error) significantly improves the learning efficiency for RL algorithms. Intuitively, surprise quantifies the unexpectedness of an experience to the learning agent. But how surprise is related to the importance of experience is not well understood. To address this problem, we derive three value metrics to quantify the importance of experience, which consider the extra reward would be earned by accessing the experience. We theoretically show these value metrics are upper-bounded by surprise for Q-learning. Furthermore, we successfully extend our theoretical framework to maximum-entropy RL by deriving the lower and upper bounds of these value metrics for soft Q-learning, which is also related to surprise. Our framework links two important quantities in RL, i.e., surprise and value of experience, and provides a theoretical basis to estimate the value of experience by surprise. We empirically show that the upper bounds hold in practice, and experience replay using the upper bound as priority improves maximum-entropy RL in Atari games. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "DB3Atcpj9_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1623/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: The authors of this paper make a connection between the TD-error from a single unit of experience and various metrics of improvement for agents trained with prioritized experience replay and Q-learning or soft Q-learning. They show that priorities based on TD-error are indeed sensible, and show that a small adjustment to TD-error-as-priority for soft Q-learning agents is both theoretically sound and can yield improved performance. \n\nThe set-up and motivation of the paper is relatively clear and well-explained. One reason I like the paper is that the process is quite straightforward: 1) Strive to better understand a commonly-used algorithm, 2) Derive theory with reasonably good intuition behind it, 3) show that it empirically holds true on simple environments, and 4) the better understood result also yields modest improvements on a test suite. \n\nThere are a few areas I would consider rewriting or rewording for added clarity.\n\nFor instance, the thesis of the paper relies on a sensible definition of \"value of experience\", and this isn't made concrete until later, though perhaps this is hard to do in the introduction. As a small nit, I think the extra use of the word surprise was at first a bit unclear, especially as it aliases TD-error (unless I missed something). \n\nI think an extra paragraph about why we should use surprise as the correct metric for prioritization would be helpful. Notably, that when we know it upper bounds the three metrics, we want to continue prioritizing sampling experiences when training our agent, because this will yield faster learning, since we will have larger improvements to our agent. \n\nHowever, this dovetails into an additional few questions that could be followed up here: Is there a correct temperature with which we sample experiences from our prioritized replay that is ideal? Should we not be sampling experiences, and simply sorting the experiences by priority and training our agent on experiences in descending order of priority? Why shouldn't we do this? How much worse is this than using uniform sampling? I understand that some of these questions are hard to answer with limited compute, but some mention of this would give the paper more depth.\n\nAdditionally, there are a number of small issues in the writing, notably towards the end. A careful read of the paper for proper grammar, making sure the right propositions are used and that there are no missing words in sentences would help the flow and readability of the paper greatly. \n\nWhile the overall result is a nice one, I believe the paper has somewhat limited scope. It doesn't fundamentally change how we should approach training our agents with a replay buffer. In fact, I suspect that TD-error was used by the original authors because they knew of a link like this, or had strong suspicions of it. I think what this paper should do is explore or at least pose a gamut of interesting follow-up questions about the role of replay and how best to use it. Is it possible that there is instead a lower bound we can derive; can we learn a sampling/prioritization scheme by gradient descent that somehow does better? Are there prioritizations completely disjoint from TD-error that we should consider using? In addition, while limited resources might make more empirical investigations challenging, it's also worth understanding how other commonly used algorithmic mechanisms in deep RL interact with prioritization, such as stepping environments in batches, or the preprocessing done to observations, or things like reward or advantage clipping. \n\nI think that if the paper showed more evidence of zooming out and thinking deeply about the core problem, this would be an excellent paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Dissecting the validity of using TD-error for prioritized experience replay", "review": "Summary: The authors of this paper make a connection between the TD-error from a single unit of experience and various metrics of improvement for agents trained with prioritized experience replay and Q-learning or soft Q-learning. They show that priorities based on TD-error are indeed sensible, and show that a small adjustment to TD-error-as-priority for soft Q-learning agents is both theoretically sound and can yield improved performance. \n\nThe set-up and motivation of the paper is relatively clear and well-explained. One reason I like the paper is that the process is quite straightforward: 1) Strive to better understand a commonly-used algorithm, 2) Derive theory with reasonably good intuition behind it, 3) show that it empirically holds true on simple environments, and 4) the better understood result also yields modest improvements on a test suite. \n\nThere are a few areas I would consider rewriting or rewording for added clarity.\n\nFor instance, the thesis of the paper relies on a sensible definition of \"value of experience\", and this isn't made concrete until later, though perhaps this is hard to do in the introduction. As a small nit, I think the extra use of the word surprise was at first a bit unclear, especially as it aliases TD-error (unless I missed something). \n\nI think an extra paragraph about why we should use surprise as the correct metric for prioritization would be helpful. Notably, that when we know it upper bounds the three metrics, we want to continue prioritizing sampling experiences when training our agent, because this will yield faster learning, since we will have larger improvements to our agent. \n\nHowever, this dovetails into an additional few questions that could be followed up here: Is there a correct temperature with which we sample experiences from our prioritized replay that is ideal? Should we not be sampling experiences, and simply sorting the experiences by priority and training our agent on experiences in descending order of priority? Why shouldn't we do this? How much worse is this than using uniform sampling? I understand that some of these questions are hard to answer with limited compute, but some mention of this would give the paper more depth.\n\nAdditionally, there are a number of small issues in the writing, notably towards the end. A careful read of the paper for proper grammar, making sure the right propositions are used and that there are no missing words in sentences would help the flow and readability of the paper greatly. \n\nWhile the overall result is a nice one, I believe the paper has somewhat limited scope. It doesn't fundamentally change how we should approach training our agents with a replay buffer. In fact, I suspect that TD-error was used by the original authors because they knew of a link like this, or had strong suspicions of it. I think what this paper should do is explore or at least pose a gamut of interesting follow-up questions about the role of replay and how best to use it. Is it possible that there is instead a lower bound we can derive; can we learn a sampling/prioritization scheme by gradient descent that somehow does better? Are there prioritizations completely disjoint from TD-error that we should consider using? In addition, while limited resources might make more empirical investigations challenging, it's also worth understanding how other commonly used algorithmic mechanisms in deep RL interact with prioritization, such as stepping environments in batches, or the preprocessing done to observations, or things like reward or advantage clipping. \n\nI think that if the paper showed more evidence of zooming out and thinking deeply about the core problem, this would be an excellent paper. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604106130727}, {"id": "QxkRd5ozRRy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1623/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tries to interpret PER from the lens of replaying experience that has the most \"surprise\" and shows how it connects to some notions of value such as the expected value of the backup and policy improvement value and evaluation improvement value and argue. The authors also derive a max-ent version of this and show that this can improve performance on some Atari games (though this is not that convincing).\n\nMy score for this paper is based on these points:\n\n Motivation: I do not see the motivation for introducing these metrics and why that explains PER in the first place. Agreed that PER is a reasonable choice, and it can upper bound the EIB and EVB metrics (i have issues with this too, more on this next), it just seems to me that the paper doesn't make any convincing claim for why this helps us understand why PER works. If the focus of the paper is on understanding PER, then the paper does not do a good job of it. If it is to introduce these prioritization based on these metrics -- and the paper focuses entirely on them -- then I then have several concerns next.\n\n- Definition of the value metrics: The cited definition of these metrics requires using the true Q-function or the true value function of the resulting policy. If we end up approximating this using the learned model, what is the guarantee that these metrics are indeed useful? Also theorem 1 should be restated to say that they care about the \"empirical\" EIB and EVB, that is computed using the learned Q-function, else it doesn't make sense to me. Moreover, if the TD error is a bound (which I think isn't with neural networks as I discuss in the next point) on the empirical EVB, can't I just drastically overestimate Q-values and get a larger empirical EVB value to be super high and prioritize on those examples? Why is that good? Why won't that promote overestimation? \n\n- Why is the update on the Q-value assumed to be tabular if the experiments are with a deep network on Atari? In a non-tabular setting Theorem 1 does not hold so either that should be rederived for the case of DQN or the experiments should be adjusted to do it on tabular settings.   In any case, now it is not clear to me why the method works with DQN, since the update in this setting isn't equal to $Q(s, a) \\leftarrow Q(s, a) + \\alpha TD(s, a)$. In general, the solution isn't known with neural networks, so the upper bound story doesn't hold there. With the NTK (Jacot et al.) assumption, I can obtain a somewhat similar update but pre-conditioned with the kernel Gram matrix (see Achiam et al. Towards characterizing divergence in deep Q-learning). However, Theorem 1 doesn't hold anymore now. So, it is unclear why the method works.\n\n- Even if I were to look at the experiments only, the results are not that impressive. The method is generally close to PER, and maybe a little better, but no comparison is made on a more efficient method such as Rainbow, and there are only 9 Atari games, which is too little. So, that is not super convincing yet.\n\nI would suggest the authors make some of the changes above.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not convinced about motivation, method or results", "review": "This paper tries to interpret PER from the lens of replaying experience that has the most \"surprise\" and shows how it connects to some notions of value such as the expected value of the backup and policy improvement value and evaluation improvement value and argue. The authors also derive a max-ent version of this and show that this can improve performance on some Atari games (though this is not that convincing).\n\nMy score for this paper is based on these points:\n\n Motivation: I do not see the motivation for introducing these metrics and why that explains PER in the first place. Agreed that PER is a reasonable choice, and it can upper bound the EIB and EVB metrics (i have issues with this too, more on this next), it just seems to me that the paper doesn't make any convincing claim for why this helps us understand why PER works. If the focus of the paper is on understanding PER, then the paper does not do a good job of it. If it is to introduce these prioritization based on these metrics -- and the paper focuses entirely on them -- then I then have several concerns next.\n\n- Definition of the value metrics: The cited definition of these metrics requires using the true Q-function or the true value function of the resulting policy. If we end up approximating this using the learned model, what is the guarantee that these metrics are indeed useful? Also theorem 1 should be restated to say that they care about the \"empirical\" EIB and EVB, that is computed using the learned Q-function, else it doesn't make sense to me. Moreover, if the TD error is a bound (which I think isn't with neural networks as I discuss in the next point) on the empirical EVB, can't I just drastically overestimate Q-values and get a larger empirical EVB value to be super high and prioritize on those examples? Why is that good? Why won't that promote overestimation? \n\n- Why is the update on the Q-value assumed to be tabular if the experiments are with a deep network on Atari? In a non-tabular setting Theorem 1 does not hold so either that should be rederived for the case of DQN or the experiments should be adjusted to do it on tabular settings.   In any case, now it is not clear to me why the method works with DQN, since the update in this setting isn't equal to $Q(s, a) \\leftarrow Q(s, a) + \\alpha TD(s, a)$. In general, the solution isn't known with neural networks, so the upper bound story doesn't hold there. With the NTK (Jacot et al.) assumption, I can obtain a somewhat similar update but pre-conditioned with the kernel Gram matrix (see Achiam et al. Towards characterizing divergence in deep Q-learning). However, Theorem 1 doesn't hold anymore now. So, it is unclear why the method works.\n\n- Even if I were to look at the experiments only, the results are not that impressive. The method is generally close to PER, and maybe a little better, but no comparison is made on a more efficient method such as Rainbow, and there are only 9 Atari games, which is too little. So, that is not super convincing yet.\n\nI would suggest the authors make some of the changes above.", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604031668438}, {"id": "LzGCgEsfVk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1623/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work aimed to understand the prioritized experience replay, a widely used technique to improve learning efficiently for RL agents. The authors proposed three different value metrics to quantify the experience, and showed that they are upper bounded by the TD error (up to a constant). The extension to soft Q-learning was also presented. Finally, the authors showed in experiments that derived upper bounds hold in Maze and CartPole. They also demonstrated that a new variant based on the upper bound achieved better performance on a subset of Atari games.\n\nThe authors tried to achieve a deep understanding of the prioritized experience replay, which I believe to be an important task. However, after reading through the paper, I am afraid that the question why prioritized experience replay works so well in practice is not well addressed. The authors provided the three metrics, and provided their upper bounds. Unfortunately, they are not sufficient to help the understanding, as the upper bound derived is just the TD error used in the prioritized experience replay. I also do not see enough depth for these theoretical results, and the presentation could be improved as well. The experiments indeed showed some practical benefits, but descriptions are confusing sometimes. My detailed comments and questions are as follows.\n\n1. When defining EVB, PIV, EIV in Eq. (3)- Eq. (5), \"s_k\" is used. However, in the derivation of Theorem 3.1, why is \"s\" used? They should be the same if both are from \"e_k\".\n2. The introduction of function approximators in the last paragraph in Section 2 is confusing: How are they used in the following sections? Also, why does the assumption \"...as if the parameterized Q-function converges to its target value\" hold?\n3. For Theorem 3.1, it looks to me that a tighter upper bound according to the derivations should be \\alpha |TD|. Why did you omit \\alpha?\n4. In the derivation for Eq. (7), the authors claimed that \"the third line is because the increase in Q-function resulted from greedy\npolicy improvement will not exceeds the surprise (times the learning step-size)\". Could you elaborate more on why this is the case?\n5. For VER in Section 5.3, which upper bound does it use? I guess it may come from Theorem 4.1, but need more clarification. Also, what are the exact differences between VER and PER?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This work aimed to understand the prioritized experience replay, a widely used technique to improve learning efficiently for RL agents. The authors proposed three different value metrics to quantify the experience, and showed that they are upper bounded by the TD error (up to a constant). The extension to soft Q-learning was also presented. Finally, the authors showed in experiments that derived upper bounds hold in Maze and CartPole. They also demonstrated that a new variant based on the upper bound achieved better performance on a subset of Atari games.\n\nThe authors tried to achieve a deep understanding of the prioritized experience replay, which I believe to be an important task. However, after reading through the paper, I am afraid that the question why prioritized experience replay works so well in practice is not well addressed. The authors provided the three metrics, and provided their upper bounds. Unfortunately, they are not sufficient to help the understanding, as the upper bound derived is just the TD error used in the prioritized experience replay. I also do not see enough depth for these theoretical results, and the presentation could be improved as well. The experiments indeed showed some practical benefits, but descriptions are confusing sometimes. My detailed comments and questions are as follows.\n\n1. When defining EVB, PIV, EIV in Eq. (3)- Eq. (5), \"s_k\" is used. However, in the derivation of Theorem 3.1, why is \"s\" used? They should be the same if both are from \"e_k\".\n2. The introduction of function approximators in the last paragraph in Section 2 is confusing: How are they used in the following sections? Also, why does the assumption \"...as if the parameterized Q-function converges to its target value\" hold?\n3. For Theorem 3.1, it looks to me that a tighter upper bound according to the derivations should be \\alpha |TD|. Why did you omit \\alpha?\n4. In the derivation for Eq. (7), the authors claimed that \"the third line is because the increase in Q-function resulted from greedy\npolicy improvement will not exceeds the surprise (times the learning step-size)\". Could you elaborate more on why this is the case?\n5. For VER in Section 5.3, which upper bound does it use? I guess it may come from Theorem 4.1, but need more clarification. Also, what are the exact differences between VER and PER?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603855821155}, {"id": "ybQrU_KUh9", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1623/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: the idea of prioritized experience replay is revisited, but from a new perspective with new theoretical results. Here, the authors propose the expected value of backup (EVB) as a metric to assess the quality of a sample and its potential improvement on the policy and on the value function. The authors decompose this metric into the benefit attributed to the policy and benefit to the value function. The authors have two theorems. The first theorem shows that the surprise (aka the temporal difference error) is an upper bound of the EVB in the Q-learning setting. The second theorem shows that the surprise, multiplied by a constant that depends on the policy, is an upper bound to EVB in the soft RL setting. The authors demonstrate that the proposed tighter bound on the EVB *could* yield improvements in the soft RL setting. \n\nPros of this work:\n- the work tries to tackle an important and still not fully answered question, which is why prioritized replay works well in the DQN setting but not in the soft RL settings, making very nice connections to the existing empirical literature on the topic, and a suggestion of how to improve PER\n- the paper is very clearly written and the motivation of addressing PER from a rigorous perspective is clear.\n- overall, a rigorous study of the \"quality\" of samples and experience seems like a very promising direction if the goal is to develop more intelligent agents that can make better use of the information available to them\n- the authors try to demonstrate why the approach leads to tighter bounds than standard PER with some visualization (although I point out some issues below)\n\nCons:\n- lack of motivation regarding the choice of EVB\n- more thorough experimentation\n- lack of evidence suggesting a tight upper bound (e.g. modest improvement in soft RL, maze example does not suggest a tight bound in Q-learning)\n\nDetails:\n- lack of motivation regarding the choice of EVB: while the EVB seems like an intuitive starting point for investigation, its motivation and a-priori connection to prioritized replay is not fully clear. one of the problems with EVB is that it seems to be a largely myopic measure of quality that is linked to only the current sample, e.g. (s,a,r,s'), while ignoring the effect of the sample on further backups in the rollout. Recent work on the topic [1] (PSER) suggest that a myopic approach can be significantly improved by up-weighting earlier transitions that lead to good transitions in the future. This would seem to suggest that a non-myopic metric could lead to more significant improvements than working with a myopic setting which could be fundamentally flawed. While I do agree that EVB is a good starting point for analysis, I am not convinced the results are fully conclusive, particularly given the need for stronger motivation on EVB and more conclusive experiments (see below) . I would encourage the authors to think about how to better motivate the use of EVB in the papers introduction more clearly if existing literature suggests so (I am somewhat aware of [2], although the authors there do highlight the shortcomings of EVB). Could sequential replay be the result of using an underlying metric that is less myopic than EVB, which seems to shows better promise over PER? If not, why is EVB truly the only right approach for analyzing PER?\n- more thorough experimentation: while I concur that the current paper addresses an important theoretical question, and while the result appears trivial to implement (not more difficult than PER), the bound should also be justified experimentally if it is to have significant value in practice. I think one key detail that is missing and could greatly benefit the paper is a more careful analysis and visualization of what the modified tighter bound is actually doing in the soft-Q setting. For example, how does the ranking of experience change when using the \"tighter\" bound? How often does this bound lead to a revision or re-evaluation of experience v.s. PER? Is there a fundamental reason why PER could not be as effective for soft RL in general that cannot be explained by better myopic estimates of the value of each transition? \n- lack of evidence suggesting a tight upper bound - Q: while the authors argue at several points in the paper that the |TD| could be a tight upper bound, this bound is only attained in some special cases. However, this does not really mean that the upper bound is tight for Q-learning, since there could be other upper bounds based on |TD| that incorporate additional value or policy information that could be tighter (same also applies for soft RL). The Figure 1 also seems to suggest that the values are scattered quite randomly and do not attain the upper bound as claimed for Q-learning. This does not suggest that |TD| is tight in any way. Can a gap be proved that effectively provides a lower bound on EVB? \n- lack of empirical evidence supporting the tighter bounds - soft RL: The experiments on Atari also suggest that the improvements of VER are quite mild, only improving on PER with statistical confidence on two of the experiments (interestingly, these two experiments also show mild improvement by PSER, whereas PSER shows considerable improvements on other games evaluated here).  This seems to be at odds with the claims that VER is a significant improvement of PER in soft RL. In the games where VER does not improve upon PER significantly, I would encourage the authors to comment on why the results are so similar to PER. I think the current benchmark problems are fairly sufficient and complex, and would not require more evaluation, unless the authors believe this would lead to a different conclusion or if the games selected are not representative of where VER can be beneficial. \n\nOther points:\n- while I agree with the authors that when the learning rate is held constant, |TD| is an \"upper bound\" to EVB in the Q-learning case. However, in practice we often use state-dependent learning rates that can be annealed over time with visitation counts, which can often yield improvements (as long as the usual stochastic convergence conditions are satisfied, of course). In this case, wouldn't the learning rate play a role in the bound?\n\nI think the work is very interesting and addresses a central issue, and I hope that the above comments can be useful to improve the paper. Overall, I think it is necessary to think more carefully about the connection between PER and quantifying the value of an experience (e.g. why EVB? how to reconcile moderate empirical evidence of the new bounds?). I am looking forward to the authors' response on these issues above.\n\nReferences:\n[1] Brittain, Marc, et al. \"Prioritized Sequence Experience Replay.\" arXiv preprint arXiv:1905.12726 (2019).\n[2] Mattar, Marcelo G., and Nathaniel D. Daw. \"Prioritized memory access explains planning and hippocampal replay.\" Nature neuroscience 21.11 (2018): 1609-1617.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Very interesting work with great potential, but the choice of EVB needs clearer motivation, and experimental is inconclusive", "review": "Summary: the idea of prioritized experience replay is revisited, but from a new perspective with new theoretical results. Here, the authors propose the expected value of backup (EVB) as a metric to assess the quality of a sample and its potential improvement on the policy and on the value function. The authors decompose this metric into the benefit attributed to the policy and benefit to the value function. The authors have two theorems. The first theorem shows that the surprise (aka the temporal difference error) is an upper bound of the EVB in the Q-learning setting. The second theorem shows that the surprise, multiplied by a constant that depends on the policy, is an upper bound to EVB in the soft RL setting. The authors demonstrate that the proposed tighter bound on the EVB *could* yield improvements in the soft RL setting. \n\nPros of this work:\n- the work tries to tackle an important and still not fully answered question, which is why prioritized replay works well in the DQN setting but not in the soft RL settings, making very nice connections to the existing empirical literature on the topic, and a suggestion of how to improve PER\n- the paper is very clearly written and the motivation of addressing PER from a rigorous perspective is clear.\n- overall, a rigorous study of the \"quality\" of samples and experience seems like a very promising direction if the goal is to develop more intelligent agents that can make better use of the information available to them\n- the authors try to demonstrate why the approach leads to tighter bounds than standard PER with some visualization (although I point out some issues below)\n\nCons:\n- lack of motivation regarding the choice of EVB\n- more thorough experimentation\n- lack of evidence suggesting a tight upper bound (e.g. modest improvement in soft RL, maze example does not suggest a tight bound in Q-learning)\n\nDetails:\n- lack of motivation regarding the choice of EVB: while the EVB seems like an intuitive starting point for investigation, its motivation and a-priori connection to prioritized replay is not fully clear. one of the problems with EVB is that it seems to be a largely myopic measure of quality that is linked to only the current sample, e.g. (s,a,r,s'), while ignoring the effect of the sample on further backups in the rollout. Recent work on the topic [1] (PSER) suggest that a myopic approach can be significantly improved by up-weighting earlier transitions that lead to good transitions in the future. This would seem to suggest that a non-myopic metric could lead to more significant improvements than working with a myopic setting which could be fundamentally flawed. While I do agree that EVB is a good starting point for analysis, I am not convinced the results are fully conclusive, particularly given the need for stronger motivation on EVB and more conclusive experiments (see below) . I would encourage the authors to think about how to better motivate the use of EVB in the papers introduction more clearly if existing literature suggests so (I am somewhat aware of [2], although the authors there do highlight the shortcomings of EVB). Could sequential replay be the result of using an underlying metric that is less myopic than EVB, which seems to shows better promise over PER? If not, why is EVB truly the only right approach for analyzing PER?\n- more thorough experimentation: while I concur that the current paper addresses an important theoretical question, and while the result appears trivial to implement (not more difficult than PER), the bound should also be justified experimentally if it is to have significant value in practice. I think one key detail that is missing and could greatly benefit the paper is a more careful analysis and visualization of what the modified tighter bound is actually doing in the soft-Q setting. For example, how does the ranking of experience change when using the \"tighter\" bound? How often does this bound lead to a revision or re-evaluation of experience v.s. PER? Is there a fundamental reason why PER could not be as effective for soft RL in general that cannot be explained by better myopic estimates of the value of each transition? \n- lack of evidence suggesting a tight upper bound - Q: while the authors argue at several points in the paper that the |TD| could be a tight upper bound, this bound is only attained in some special cases. However, this does not really mean that the upper bound is tight for Q-learning, since there could be other upper bounds based on |TD| that incorporate additional value or policy information that could be tighter (same also applies for soft RL). The Figure 1 also seems to suggest that the values are scattered quite randomly and do not attain the upper bound as claimed for Q-learning. This does not suggest that |TD| is tight in any way. Can a gap be proved that effectively provides a lower bound on EVB? \n- lack of empirical evidence supporting the tighter bounds - soft RL: The experiments on Atari also suggest that the improvements of VER are quite mild, only improving on PER with statistical confidence on two of the experiments (interestingly, these two experiments also show mild improvement by PSER, whereas PSER shows considerable improvements on other games evaluated here).  This seems to be at odds with the claims that VER is a significant improvement of PER in soft RL. In the games where VER does not improve upon PER significantly, I would encourage the authors to comment on why the results are so similar to PER. I think the current benchmark problems are fairly sufficient and complex, and would not require more evaluation, unless the authors believe this would lead to a different conclusion or if the games selected are not representative of where VER can be beneficial. \n\nOther points:\n- while I agree with the authors that when the learning rate is held constant, |TD| is an \"upper bound\" to EVB in the Q-learning case. However, in practice we often use state-dependent learning rates that can be annealed over time with visitation counts, which can often yield improvements (as long as the usual stochastic convergence conditions are satisfied, of course). In this case, wouldn't the learning rate play a role in the bound?\n\nI think the work is very interesting and addresses a central issue, and I hope that the above comments can be useful to improve the paper. Overall, I think it is necessary to think more carefully about the connection between PER and quantifying the value of an experience (e.g. why EVB? how to reconcile moderate empirical evidence of the new bounds?). I am looking forward to the authors' response on these issues above.\n\nReferences:\n[1] Brittain, Marc, et al. \"Prioritized Sequence Experience Replay.\" arXiv preprint arXiv:1905.12726 (2019).\n[2] Mattar, Marcelo G., and Nathaniel D. Daw. \"Prioritized memory access explains planning and hippocampal replay.\" Nature neuroscience 21.11 (2018): 1609-1617.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603853265131}], "openreview_url": "https://openreview.net/forum?id=jAJrc-kzVd0", "arxiv_id": "2102.03261", "paper_pdf": "papers/jAJrc-kzVd0.pdf", "paper_pdf_sha256": "413c9304337d973d56087c5b43d7beac63c8d7de19ae8131e9715991306f24c0", "paper_pdf_bytes": 1818566, "paper_pdf_source": "openreview", "code_url": "https://github.com/AmazingAng/VER", "code_repository": "AmazingAng/VER", "code_commit": "cdfcdc96defe117d6123940e6aa4174db5c6f6c6", "code_archive": "repos/jAJrc-kzVd0.zip", "code_archive_sha256": "4d62b0048232de39ad3f88023efa8a75bc10febf8cc59d84929ba0e0843d64a2", "code_archive_bytes": 619847, "code_file_count": 10, "code_extensions": {".ipynb": 5, ".py": 5}, "github_disk_usage_kb": 630, "github_languages": {"Jupyter Notebook": 988075, "Python": 49142}, "github_archived": false, "github_pushed_at": "2021-02-01T16:51:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/revisiting-prioritized-experience-replay-a-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BygSZAVKvr", "year": 2020, "status": "rejected", "title": "Energy-Aware Neural Architecture Optimization with Fast Splitting Steepest Descent", "authors": ["Dilin Wang", "Meng Li", "Lemeng Wu", "Vikas Chandra", "Qiang Liu"], "authorids": ["dilin@cs.utexas.edu", "meng.li@fb.com", "lmwu@cs.utexas.edu", "vchandra@fb.com", "lqiang@cs.utexas.edu"], "authors_source": "OpenReview API", "abstract": "Designing energy-efficient networks is of critical importance for enabling state-of-the-art deep learning in mobile and edge settings where the computation and energy budgets are highly limited. Recently, Wu et al. (2019) framed the search of efficient neural architectures into a continuous splitting process: it iteratively splits existing neurons into multiple off-springs to achieve progressive loss minimization,  thus finding novel architectures by gradually growing the neural network. However, this method was not specifically tailored for designing energy-efficient networks, and is computationally expensive on large-scale benchmarks.  In this work, we substantially improve Wu et al. (2019) in two significant ways:  1) we incorporate the energy cost of splitting different neurons to better guide the splitting process, thereby discovering more energy-efficient network architectures; 2) we substantially speed up the splitting process of Wu et al. (2019), which requires expensive eigen-decomposition, by proposing a highly scalable Rayleigh-quotient stochastic gradient algorithm.  Our fast algorithm allows us to reduce the computational cost of splitting to the same level of typical back-propagation updates and enables efficient implementation on GPU. Extensive empirical results show that our method can train highly accurate and energy-efficient networks on challenging datasets such as ImageNet,  improving a variety of baselines,  including the pruning-based methods and expert-designed architectures.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJg4nZPUjr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper964/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is well written and flows very well. The idea is straightforward and easy to understand. Here are my feedbacks:\n\n--Method--\n\nMethodology-wise, the novelty is somewhat limited. The main technical contributions are: 1) formulate linear programming to reduce energy costs, and the formulation of linear programming is straightforward. 2) the claimed main contribution is an application of Rayleigh-Quotient gradient descent to approximate the eigenvalue calculations in the model proposed by Wu et al (2019). \n\nI believe computing the eigenvalue shall not be the only way to solve Eq.(3), e.g. using gradient-based method or Lagrange. My main question is whether computing the global optimum really matters? Since your method is still essentially an approximation algorithm, which may break the optimality condition here. From the experimental results, it seems that your approximation is quite close to the analytical solution. Therefore, it is unclear computing a global optimum really matters here.\n\nI tried to buy the idea of escaping the saddle point with splitting (it indeed sounds straightforward and reasonable). It will be great if the author can add some empirical experiments to verify it.\n\n--Content--\nThe entire section2 is at reviewing prior works, and I believe you should cut down the content here.\n\n--Experiments--\n1. Diversity of your tests: the author main uses MobileNet in testing their ideas. It seems that their method starts at MobileNet, then tries to improve it. I suggest authors adding other recent works, e.g. FBNet, MobileNetV3, into their tests to diversify the types of networks. \n\n2. Results variations: all figures, e.g. fig.3, fig.4, in the paper lack plotting their results variations. Since the optimization may converge to different local optimum, it is more persuasive to show their performance variations.\n\n3. The final results are not surprising: in table.1 and table.2, as far as I'm aware, the mainstream accuracy for ImageNet under the mobile setting should be 75% top-1. And the SOTA top-1 accuracy on ImageNet is around 85.5%. Table.1 and table.2 do somewhat show the effectiveness of your method, but its significance is limited especially considering the limited novelty of methodology.\n\nMinors:\nIn Fig.3 k = 6, it seems vanilla splitting is better than accuracy and parameters, except for 0.2 log higher flops. I don't think this makes a compelling case here.\n\nIn Fig.4, from flops 7 ~ 9, your results are similar to Bn especially accuracy > 0.6. When accuracy < 0.6, it is less interesting, and I believe the improvement should be huge.\n\nFig.5, could you please compare the time using MAGMA from NVIDIA? I had some experiences in implementing the LAPACK and BLAS on GPUs, and it should not be that slow.\n\nOverall, this paper has some interesting results, which shows the eigenvalue can be approximated by Rayleigh-Quotient gradient descent, and show positive improvement. However, the methodological and experimental results can definitely be strengthened. The author may consider proving that achieving the global optimum on Eq.(3) really matters, then motivate the methodology. Thank you. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "The paper is well written and flows very well. The idea is straightforward and easy to understand. Here are my feedbacks:\n\n--Method--\n\nMethodology-wise, the novelty is somewhat limited. The main technical contributions are: 1) formulate linear programming to reduce energy costs, and the formulation of linear programming is straightforward. 2) the claimed main contribution is an application of Rayleigh-Quotient gradient descent to approximate the eigenvalue calculations in the model proposed by Wu et al (2019). \n\nI believe computing the eigenvalue shall not be the only way to solve Eq.(3), e.g. using gradient-based method or Lagrange. My main question is whether computing the global optimum really matters? Since your method is still essentially an approximation algorithm, which may break the optimality condition here. From the experimental results, it seems that your approximation is quite close to the analytical solution. Therefore, it is unclear computing a global optimum really matters here.\n\nI tried to buy the idea of escaping the saddle point with splitting (it indeed sounds straightforward and reasonable). It will be great if the author can add some empirical experiments to verify it.\n\n--Content--\nThe entire section2 is at reviewing prior works, and I believe you should cut down the content here.\n\n--Experiments--\n1. Diversity of your tests: the author main uses MobileNet in testing their ideas. It seems that their method starts at MobileNet, then tries to improve it. I suggest authors adding other recent works, e.g. FBNet, MobileNetV3, into their tests to diversify the types of networks. \n\n2. Results variations: all figures, e.g. fig.3, fig.4, in the paper lack plotting their results variations. Since the optimization may converge to different local optimum, it is more persuasive to show their performance variations.\n\n3. The final results are not surprising: in table.1 and table.2, as far as I'm aware, the mainstream accuracy for ImageNet under the mobile setting should be 75% top-1. And the SOTA top-1 accuracy on ImageNet is around 85.5%. Table.1 and table.2 do somewhat show the effectiveness of your method, but its significance is limited especially considering the limited novelty of methodology.\n\nMinors:\nIn Fig.3 k = 6, it seems vanilla splitting is better than accuracy and parameters, except for 0.2 log higher flops. I don't think this makes a compelling case here.\n\nIn Fig.4, from flops 7 ~ 9, your results are similar to Bn especially accuracy > 0.6. When accuracy < 0.6, it is less interesting, and I believe the improvement should be huge.\n\nFig.5, could you please compare the time using MAGMA from NVIDIA? I had some experiences in implementing the LAPACK and BLAS on GPUs, and it should not be that slow.\n\nOverall, this paper has some interesting results, which shows the eigenvalue can be approximated by Rayleigh-Quotient gradient descent, and show positive improvement. However, the methodological and experimental results can definitely be strengthened. The author may consider proving that achieving the global optimum on Eq.(3) really matters, then motivate the methodology. Thank you. \n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573446060229}, {"id": "B1evkL2EqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper964/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper builds on a recently proposed algorithm (\"splitting steepest descent\", Wu et al 2019) for guiding the growth of a smaller network into a larger one in architecture search. The algorithm in Wu, et al. alternates between two steps, (i) optimization of parameters for a fixed model and (ii) modification of the architecture by identifying a subset of neurons to split into more neurons, based on the \"splitting index\" of each neuron (amounting to evaluating the smallest eigenvalue of a matrix). This work builds on that in two ways: (i) it incorporates an energy budget into the optimization procedure for choosing which subset of neurons to split, which it approximately solves by a continuous relaxation, and (ii) avoids doing exact eigendecomposition to extract the minimum eigenvalue (splitting index) but instead replaces it with a more efficient SGD on the Rayleigh quotient. \n\nEvaluation:\n--Evaluations are done on variants of MobileNet on CIFAR-100 and ImageNet (the latter would be infeasible without the approximation scheme). The proposed approach appears to get better tradeoff between accuracy and FLOPs in these cases. In practice the non-energy aware \"vanilla\" networks do tend towards models that are small in size (fewer parameters) but are not necessarily low in energy consumption.\n--There is new material here, although I find the novelty a bit limited (e.g. only an additional constraint compared to the original approach of Wu et al and addressing a clear scalability issue with the original work, i.e. eigendecomposition of a matrix, with what seem straightforward approximations, ). The empirical results in Table 1 and 2 seem solid, but I'm not familiar enough with past results in this area  to evaluate their significance. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "Summary:\nThis paper builds on a recently proposed algorithm (\"splitting steepest descent\", Wu et al 2019) for guiding the growth of a smaller network into a larger one in architecture search. The algorithm in Wu, et al. alternates between two steps, (i) optimization of parameters for a fixed model and (ii) modification of the architecture by identifying a subset of neurons to split into more neurons, based on the \"splitting index\" of each neuron (amounting to evaluating the smallest eigenvalue of a matrix). This work builds on that in two ways: (i) it incorporates an energy budget into the optimization procedure for choosing which subset of neurons to split, which it approximately solves by a continuous relaxation, and (ii) avoids doing exact eigendecomposition to extract the minimum eigenvalue (splitting index) but instead replaces it with a more efficient SGD on the Rayleigh quotient. \n\nEvaluation:\n--Evaluations are done on variants of MobileNet on CIFAR-100 and ImageNet (the latter would be infeasible without the approximation scheme). The proposed approach appears to get better tradeoff between accuracy and FLOPs in these cases. In practice the non-energy aware \"vanilla\" networks do tend towards models that are small in size (fewer parameters) but are not necessarily low in energy consumption.\n--There is new material here, although I find the novelty a bit limited (e.g. only an additional constraint compared to the original approach of Wu et al and addressing a clear scalability issue with the original work, i.e. eigendecomposition of a matrix, with what seem straightforward approximations, ). The empirical results in Table 1 and 2 seem solid, but I'm not familiar enough with past results in this area  to evaluate their significance. "}, "tcdate": 1572287967136}, {"id": "SkxvmpoAYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper964/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper is based on a prior work which proposed Splitting Steepest Descent to search better network structures via splitting existing neurons to multiple off-springs. As an improvement, the authors (1) incorporate the energy cost to better guide the splitting process and (2) reduce the time and space complexity by approximating the original computation process with Rayleigh-Quotient Gradient Descent. They conduct experiments on public image classification datasets using lightweight networks as backbones to show that their algorithm outperforms existing methods. The paper is well written and easy to follow. The experiments are comprehensive and the evaluation results shows good properties of proposed method.\n\nIn brief, this paper is an improvement to a splitting algorithm in a previous work, achieving good efficiency and enabling application on large datasets. However, the theory needs more justification and the experiments results are not sufficient to show the significance of their contribution. Therefore, this paper may not be accepted unless more experiment results are given.\n\nFor the theory & modeling, the following should be addressed.\n1.\tThe mathematical justification of the optimization on energy cost is not very sound, and the definition of optimal splitting set seems arbitrary.\n2.\tGiven that the experiments are conducted on convolutional networks, it would be more illustrative if the paper describe the process of applying the algorithm on common convolutional operators.\n3.\tThe novelty is limited by the prior work.\n\nFor the experiment, the following should be addressed.\n1.\tThe experiments are mainly conducted on MobileNet network. It would be more convincing if more experiment is done on other lightweight or normal convolutional networks.\n2.\tThis paper reduces time and space complexity of the algorithm in a previous work, but there is no running time or memory footprint statistics to support this argument.\n3.\tThe paper only lists one pruning-based method as comparison in the experiments. It would be more convincing if more pruning and splitting methods are presented.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper is based on a prior work which proposed Splitting Steepest Descent to search better network structures via splitting existing neurons to multiple off-springs. As an improvement, the authors (1) incorporate the energy cost to better guide the splitting process and (2) reduce the time and space complexity by approximating the original computation process with Rayleigh-Quotient Gradient Descent. They conduct experiments on public image classification datasets using lightweight networks as backbones to show that their algorithm outperforms existing methods. The paper is well written and easy to follow. The experiments are comprehensive and the evaluation results shows good properties of proposed method.\n\nIn brief, this paper is an improvement to a splitting algorithm in a previous work, achieving good efficiency and enabling application on large datasets. However, the theory needs more justification and the experiments results are not sufficient to show the significance of their contribution. Therefore, this paper may not be accepted unless more experiment results are given.\n\nFor the theory & modeling, the following should be addressed.\n1.\tThe mathematical justification of the optimization on energy cost is not very sound, and the definition of optimal splitting set seems arbitrary.\n2.\tGiven that the experiments are conducted on convolutional networks, it would be more illustrative if the paper describe the process of applying the algorithm on common convolutional operators.\n3.\tThe novelty is limited by the prior work.\n\nFor the experiment, the following should be addressed.\n1.\tThe experiments are mainly conducted on MobileNet network. It would be more convincing if more experiment is done on other lightweight or normal convolutional networks.\n2.\tThis paper reduces time and space complexity of the algorithm in a previous work, but there is no running time or memory footprint statistics to support this argument.\n3.\tThe paper only lists one pruning-based method as comparison in the experiments. It would be more convincing if more pruning and splitting methods are presented.\n"}, "tcdate": 1571892510795}], "openreview_url": "https://openreview.net/forum?id=BygSZAVKvr", "arxiv_id": "1910.03103", "paper_pdf": "papers/BygSZAVKvr.pdf", "paper_pdf_sha256": "113e3daabcc3343d29558952932584d7a3a167bbce31e21184cf286cc17c9ced", "paper_pdf_bytes": 5004873, "paper_pdf_source": "openreview", "code_url": "https://github.com/dilinwang820/fast-energy-aware-splitting", "code_repository": "dilinwang820/fast-energy-aware-splitting", "code_commit": "73a1c9e5fec403cf86711aaa133ddec0f58ba36a", "code_archive": "repos/BygSZAVKvr.zip", "code_archive_sha256": "e4609dfe005e5f5cc137fbdf76a388f7655c4d6c0d3310dc88c3822f91dfb59c", "code_archive_bytes": 629947, "code_file_count": 34, "code_extensions": {".py": 31, ".sh": 3}, "github_disk_usage_kb": 1450, "github_languages": {"Python": 250014, "Shell": 576}, "github_archived": false, "github_pushed_at": "2020-02-06T21:12:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/energy-aware-neural-architecture-optimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "guXFowAqLQ", "year": 2026, "status": "rejected", "title": "Time series saliency maps: Explaining models across multiple domains", "authors": ["Christodoulos Kechris", "Jonathan Dan", "David Atienza"], "authorids": ["~Christodoulos_Kechris1", "~Jonathan_Dan1", "~David_Atienza1"], "authors_source": "OpenReview API", "abstract": "Traditional saliency map methods, popularized in computer vision, highlight individual points (pixels) of the input that contribute the most to the model's output. However, in time-series, they offer limited insights, as semantically meaningful features are often found in other domains. We introduce Cross-domain Integrated Gradients, a generalization of Integrated Gradients. Our method enables feature attributions in any domain that can be formulated as an invertible, differentiable transformation of the time domain. Crucially, our derivation extends the original Integrated Gradients into the complex domain, enabling frequency-based attributions. We provide the necessary theoretical guarantees, namely, path independence and completeness. Our approach reveals interpretable, problem-specific attributions that time-domain methods cannot capture in three real-world tasks: wearable sensor heart rate extraction, electroencephalography-based seizure detection, and zero-shot time-series forecasting. We release an open-source Tensorflow/PyTorch library to enable plug-and-play cross-domain explainability for time-series models. These results demonstrate the ability of cross-domain integrated gradients to provide semantically meaningful insights into time series models that are impossible to achieve with traditional time‑domain saliency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "hYEj6uuQh2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18289/Reviewer_8oeg"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The authors advance explainability for time series methods by innovating on the integrated gradient method from 2017. Their proposal is called \"Cross-domain Integrated Gradients\", based on the fact that their method works on any invertible transformation. The user of the XAI method can thereby choose a transformation of their choice based on which domain is most suitable for explanations. The authors demonstrate qualitative feasibility using the Fourier Transform, Independent Component Analysis, and Seasonal-Trend decomposition.", "review_text": "The authors advance explainability for time series methods by innovating on the integrated gradient method from 2017. Their proposal is called \"Cross-domain Integrated Gradients\", based on the fact that their method works on any invertible transformation. The user of the XAI method can thereby choose a transformation of their choice based on which domain is most suitable for explanations. The authors demonstrate qualitative feasibility using the Fourier Transform, Independent Component Analysis, and Seasonal-Trend decomposition.", "strengths": "1. The paper tackles an important and timely problem of developing XAI methods for time series. This field is in its infancy and underdeveloped, with the majority of XAI methods being developed on image data.\n\n1. The authors provide both a TensorFlow and a PyTorch open-source library for their method.\n\n1. Strong mathematical foundation for their method. \n\n1. The method works, based on Figure 2-4, and is tested using three different transformations, in three different data domains.", "weaknesses": "As far as I can see, the paper has only one major weakness, which is its positioning in the existing state of the art. If the authors can address this, I will definitely consider changing my recommendation.\n\nReferences to relevant prior literature on time series explainability are lacking in section 2, e.g, there are only two papers from 2024 in related work.\n\nThe experimental section is quite limited and is primarily qualitative. The only quantitative experiments I could find are in Appendix F, where the authors present insertion/deletion and compare that to time IG. There should be comparisons included with existing time-frequency analysis XAI methods. Comparisons should be made e.g to https://proceedings.mlr.press/v265/brusch25a.html, which also only assumes invertible transformations and therefore seems relevant.", "questions": "1. Equation 1:  $\\xi$ can easily become negative; is that an issue?\n\n1. Figure 1: Why is the peak of the orange distribution not located at 4 Hz, as specified by Eq. 1?\n\n1. Equation 2: I'm a bit unsure about the notation in the integral with $x\n+t(x-\\hat{x})$, does this mean the partial derivative of $f$ is evaluated at that point, and then you integrate over $t$ ?\n\n1. Line 352: $x(t) = a_1 \\cos(2\\pi · \\xi_{hr}\\cdot  t + \\phi) + a2 \\cos(2\\cdot \\pi(2\\xi_{hr})\\cdot  t + \\phi))$. Small esthetic typo: should it be $2\\pi$ without the cdot in the second cosine?\n\n1. Figure 2: The blue dashed (as opposed to solid) lines make this figure more difficult to read, but I may be wrong.\n\n1. Line 389: What does the $\\rightarrow$ mean in this context?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors advance explainability for time series methods by innovating on the integrated gradient method from 2017. Their proposal is called \"Cross-domain Integrated Gradients\", based on the fact that their method works on any invertible transformation. The user of the XAI method can thereby choose a transformation of their choice based on which domain is most suitable for explanations. The authors demonstrate qualitative feasibility using the Fourier Transform, Independent Component Analysis, and Seasonal-Trend decomposition.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The paper tackles an important and timely problem of developing XAI methods for time series. This field is in its infancy and underdeveloped, with the majority of XAI methods being developed on image data.\n\n1. The authors provide both a TensorFlow and a PyTorch open-source library for their method.\n\n1. Strong mathematical foundation for their method. \n\n1. The method works, based on Figure 2-4, and is tested using three different transformations, in three different data domains.", "weaknesses": "As far as I can see, the paper has only one major weakness, which is its positioning in the existing state of the art. If the authors can address this, I will definitely consider changing my recommendation.\n\nReferences to relevant prior literature on time series explainability are lacking in section 2, e.g, there are only two papers from 2024 in related work.\n\nThe experimental section is quite limited and is primarily qualitative. The only quantitative experiments I could find are in Appendix F, where the authors present insertion/deletion and compare that to time IG. There should be comparisons included with existing time-frequency analysis XAI methods. Comparisons should be made e.g to https://proceedings.mlr.press/v265/brusch25a.html, which also only assumes invertible transformations and therefore seems relevant.", "questions": "1. Equation 1:  $\\xi$ can easily become negative; is that an issue?\n\n1. Figure 1: Why is the peak of the orange distribution not located at 4 Hz, as specified by Eq. 1?\n\n1. Equation 2: I'm a bit unsure about the notation in the integral with $x\n+t(x-\\hat{x})$, does this mean the partial derivative of $f$ is evaluated at that point, and then you integrate over $t$ ?\n\n1. Line 352: $x(t) = a_1 \\cos(2\\pi · \\xi_{hr}\\cdot  t + \\phi) + a2 \\cos(2\\cdot \\pi(2\\xi_{hr})\\cdot  t + \\phi))$. Small esthetic typo: should it be $2\\pi$ without the cdot in the second cosine?\n\n1. Figure 2: The blue dashed (as opposed to solid) lines make this figure more difficult to read, but I may be wrong.\n\n1. Line 389: What does the $\\rightarrow$ mean in this context?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762205717294}, {"id": "zfq7uYTIPR", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18289/Reviewer_YxUK"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper provides a novel explainability method for the time series domain. Specifically, the method is based on an extension of the popular saliency method Intergrated Gradients to incorporate multiple domains that can be derived through an invertible, differential transformation from the time domain. Importantly, the proposed method maintains the sensitivity and implementation invariance properties of IG.", "review_text": "The paper provides a novel explainability method for the time series domain. Specifically, the method is based on an extension of the popular saliency method Intergrated Gradients to incorporate multiple domains that can be derived through an invertible, differential transformation from the time domain. Importantly, the proposed method maintains the sensitivity and implementation invariance properties of IG.", "strengths": "- The paper proposes a saliency method for time series which does not solely focus on the time domain, but can also integrate latent features such as frequencies\n- The presentation of the method is easy to follow\n- The paper provides open access to the method in the form of a Python package", "weaknesses": "- The paper completely lacks references in the introduction. This is not in line with good research practice. It is unclear to the reader whether observations and statements are taken from the literature or are a novel contribution by the paper. Importantly, the observation that existing saliency maps fail in the \ntime-series domain and that other features, e.g., stemming from the frequency domain, are not novel and have been shown before in the literature (e.g., [1],[2],[3]). This renders Proposition 1, without proper citations, almost plagiarism. Section 3.2 is therefore unnecessary. I am not raising an ethics flag at the moment, but this aspect is, in my opinion, sufficient for rejecting the paper without further evaluation.\n- The paper does not discuss the many restrictions and limitations of IG and their equivalent part in the proposed method.\n- The paper states that the method is applicable to many domains. However, all argumentation and derivation (e.g., section 4.2) solely focus on the frequency domain. More explanation and examples are needed here to evaluate the usefulness of the method.\n- The method requires domain knowledge to specify the domain of interest. In practice, such knowledge might not exist, or if it does, might be limited. This can lead to dangerous misinterpretations of the explanations and wrong decision-making. It would be desirable that a novel explainability method can directly infer saliency across many (unspecified) domains to potentially uncover so far unknown important features, instead of suffering similar limitations to existing time series saliency methods on the time domain.\n- The paper does not discuss the experimental results or the limitations of the proposed method. Overall, the paper seems unfinished.\n- The experimental section only focuses on three examples with specific ML methods. Here, a model-agnostic evaluation would be beneficial.\n\n\n\n[1] Schröder, Maresa, Alireza Zamanian, and Narges Ahmidi. \"Post-hoc saliency methods fail to capture latent feature importance in time series data.\" International Workshop on Trustworthy Machine Learning for Healthcare. Cham: Springer Nature Switzerland, 2023.\n\n[2] Schröder, Maresa, Alireza Zamanian, and Narges Ahmidi. \"What about the Latent Space? The Need for Latent Feature Saliency Detection in Deep Time Series Classification.\" Machine Learning and Knowledge Extraction 5.2 (2023): 539-559.\n\n[3] Theissler, Andreas, et al. \"Explainable AI for time series classification: a review, taxonomy and research directions.\" IEEE Access 10 (2022): 100700-100724.", "questions": "- Section 4.2: How is the known failure mode of IG addressed in other domains besides the frequency domain?\n- How are failure modes/limitations of IG addressed in the proposed method for general ML models (not only CNNs)?\n- How can the method integrate multiple domains at the same time? Importantly, how can it detect + explain interactions between the domains?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper provides a novel explainability method for the time series domain. Specifically, the method is based on an extension of the popular saliency method Intergrated Gradients to incorporate multiple domains that can be derived through an invertible, differential transformation from the time domain. Importantly, the proposed method maintains the sensitivity and implementation invariance properties of IG.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper proposes a saliency method for time series which does not solely focus on the time domain, but can also integrate latent features such as frequencies\n- The presentation of the method is easy to follow\n- The paper provides open access to the method in the form of a Python package", "weaknesses": "- The paper completely lacks references in the introduction. This is not in line with good research practice. It is unclear to the reader whether observations and statements are taken from the literature or are a novel contribution by the paper. Importantly, the observation that existing saliency maps fail in the \ntime-series domain and that other features, e.g., stemming from the frequency domain, are not novel and have been shown before in the literature (e.g., [1],[2],[3]). This renders Proposition 1, without proper citations, almost plagiarism. Section 3.2 is therefore unnecessary. I am not raising an ethics flag at the moment, but this aspect is, in my opinion, sufficient for rejecting the paper without further evaluation.\n- The paper does not discuss the many restrictions and limitations of IG and their equivalent part in the proposed method.\n- The paper states that the method is applicable to many domains. However, all argumentation and derivation (e.g., section 4.2) solely focus on the frequency domain. More explanation and examples are needed here to evaluate the usefulness of the method.\n- The method requires domain knowledge to specify the domain of interest. In practice, such knowledge might not exist, or if it does, might be limited. This can lead to dangerous misinterpretations of the explanations and wrong decision-making. It would be desirable that a novel explainability method can directly infer saliency across many (unspecified) domains to potentially uncover so far unknown important features, instead of suffering similar limitations to existing time series saliency methods on the time domain.\n- The paper does not discuss the experimental results or the limitations of the proposed method. Overall, the paper seems unfinished.\n- The experimental section only focuses on three examples with specific ML methods. Here, a model-agnostic evaluation would be beneficial.\n\n\n\n[1] Schröder, Maresa, Alireza Zamanian, and Narges Ahmidi. \"Post-hoc saliency methods fail to capture latent feature importance in time series data.\" International Workshop on Trustworthy Machine Learning for Healthcare. Cham: Springer Nature Switzerland, 2023.\n\n[2] Schröder, Maresa, Alireza Zamanian, and Narges Ahmidi. \"What about the Latent Space? The Need for Latent Feature Saliency Detection in Deep Time Series Classification.\" Machine Learning and Knowledge Extraction 5.2 (2023): 539-559.\n\n[3] Theissler, Andreas, et al. \"Explainable AI for time series classification: a review, taxonomy and research directions.\" IEEE Access 10 (2022): 100700-100724.", "questions": "- Section 4.2: How is the known failure mode of IG addressed in other domains besides the frequency domain?\n- How are failure modes/limitations of IG addressed in the proposed method for general ML models (not only CNNs)?\n- How can the method integrate multiple domains at the same time? Importantly, how can it detect + explain interactions between the domains?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762012621174}, {"id": "b5HrNO0JTs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18289/Reviewer_ywmS"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 4, "confidence": 3, "summary": "This paper proposes the Cross domain Integrated Gradients method, which extends the Integrated Gradients (IG) to any reversible and differentiable transformation domain (including the complex domain), providing a more semantic and insightful explanation for time series models.", "review_text": "This paper proposes the Cross domain Integrated Gradients method, which extends the Integrated Gradients (IG) to any reversible and differentiable transformation domain (including the complex domain), providing a more semantic and insightful explanation for time series models.", "strengths": "1. Wide universality: The attribution framework for unknown transformations has high universality and is not specific to any particular transformation.\n\n2. Solid theoretical contribution: Extend the IG method to the complex field.\n\n3. Compelling & Diverse Applications: This method shows significant application potential in multiple scenarios like medical and other general time series application.", "weaknesses": "Overall, I think the strengths of this paper are very prominent, and the disadvantages are not worth mentioning compared to it. Here are a few of my small concerns.\n\n1. The main text of the paper lacks quantitative experimental comparisons, and the structure should be adjusted by moving the section in Appendix F to the main text.\n\n2. Appendix D indicates that this method is the general form of [1]. Can the author compare the two in a visual form to see if the actual effect is consistent with the theory?\n\n[1] Johanna Vielhaben, Sebastian Lapuschkin, Grégoire Montavon, and Wojciech Samek. Explainable ai\nfor time series via virtual inspection layers. Pattern Recognition, 150:110309, 2024.", "questions": "1. Since this method can be applied to all reversible transformations, is it suitable for the current popular flow generation models? What would be the computational burden in practical applications?\n\n2. What are the errors of this method for differentiable irreversible transformations? Is it possible to make corrections?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the Cross domain Integrated Gradients method, which extends the Integrated Gradients (IG) to any reversible and differentiable transformation domain (including the complex domain), providing a more semantic and insightful explanation for time series models.", "soundness": 4, "presentation": 4, "contribution": 4, "strengths": "1. Wide universality: The attribution framework for unknown transformations has high universality and is not specific to any particular transformation.\n\n2. Solid theoretical contribution: Extend the IG method to the complex field.\n\n3. Compelling & Diverse Applications: This method shows significant application potential in multiple scenarios like medical and other general time series application.", "weaknesses": "Overall, I think the strengths of this paper are very prominent, and the disadvantages are not worth mentioning compared to it. Here are a few of my small concerns.\n\n1. The main text of the paper lacks quantitative experimental comparisons, and the structure should be adjusted by moving the section in Appendix F to the main text.\n\n2. Appendix D indicates that this method is the general form of [1]. Can the author compare the two in a visual form to see if the actual effect is consistent with the theory?\n\n[1] Johanna Vielhaben, Sebastian Lapuschkin, Grégoire Montavon, and Wojciech Samek. Explainable ai\nfor time series via virtual inspection layers. Pattern Recognition, 150:110309, 2024.", "questions": "1. Since this method can be applied to all reversible transformations, is it suitable for the current popular flow generation models? What would be the computational burden in practical applications?\n\n2. What are the errors of this method for differentiable irreversible transformations? Is it possible to make corrections?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761820725176}, {"id": "caEwYL64KG", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission18289/Reviewer_ud4m"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This work is focused on the explainability of black-box model in the context of time series models. The key insight is that the semantically meaningful information might not always be found in the time domain, but in other domains such as the frequency domain. To address this limitation, a generalization of the well-known Integrated Gradients method is proposed such that explanation can be presented in different domains. The proposed methodology is analyzed and evaluated on 3 time series analysis tasks.", "review_text": "This work is focused on the explainability of black-box model in the context of time series models. The key insight is that the semantically meaningful information might not always be found in the time domain, but in other domains such as the frequency domain. To address this limitation, a generalization of the well-known Integrated Gradients method is proposed such that explanation can be presented in different domains. The proposed methodology is analyzed and evaluated on 3 time series analysis tasks.", "strengths": "1. A clear idea that is well motivated. \n\n2. A through theoretical analysis of the proposed methodology.\n\n3. A nicely written and well-structured manuscript.", "weaknesses": "1. The novelty is low. The idea of providing explanations in a a different domain is already established. The paper mentions the Virtual Inspection Layers of Vielhaben et al. [1], but do not include it as a baseline, even though [1] also evaluated Integrated Gradients with their Virtual Inspection Layers. Furthermore, several other works [2, 3] have presented methodology for providing explanations in other domains than the time domain.\n\n2. The experimental evaluation is limited. The proposed methodology is tested on 3 datasets, but no baselines are provided, and a limited quantitative evaluation. Compared to existing works [1, 3], where numerous datasets are used and a wide range of explainability metrics are evaluated, the evaluation in this work does not give insights into the usefulness of the proposed method.\n\n- [1] Vielhaben et al., Explainable AI for time series via Virtual Inspection Layers, Pattern Recognition, 2024\n- [2] Brüsch et al., FreqRISE: Explaining time series using frequency masking, NLDL, 2025\n- [3] Brüsch et al., FLEXtime: Filterbank Learning to Explain Time Series, Explainable Artificial Intelligence, 2025", "questions": "1. How does the proposed method quantitatively compare to [1, 2, 3] in terms of established explainablity metrics like faithfulness, localization, complexity, and robustness?\n\n2. Apart from being specific for Integrated Gradients, how is the invertible transform introduced here to transfer between domains different from the transform introduced in [2]?\n\n- [1] Vielhaben et al., Explainable AI for time series via Virtual Inspection Layers, Pattern Recognition 2024\n- [2] Brüsch et al., FreqRISE: Explaining time series using frequency masking, NLDL, 2025\n- [3] Brüsch et al., FLEXtime: Filterbank Learning to Explain Time Series, Explainable Artificial Intelligence , 2025??", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work is focused on the explainability of black-box model in the context of time series models. The key insight is that the semantically meaningful information might not always be found in the time domain, but in other domains such as the frequency domain. To address this limitation, a generalization of the well-known Integrated Gradients method is proposed such that explanation can be presented in different domains. The proposed methodology is analyzed and evaluated on 3 time series analysis tasks.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "1. A clear idea that is well motivated. \n\n2. A through theoretical analysis of the proposed methodology.\n\n3. A nicely written and well-structured manuscript.", "weaknesses": "1. The novelty is low. The idea of providing explanations in a a different domain is already established. The paper mentions the Virtual Inspection Layers of Vielhaben et al. [1], but do not include it as a baseline, even though [1] also evaluated Integrated Gradients with their Virtual Inspection Layers. Furthermore, several other works [2, 3] have presented methodology for providing explanations in other domains than the time domain.\n\n2. The experimental evaluation is limited. The proposed methodology is tested on 3 datasets, but no baselines are provided, and a limited quantitative evaluation. Compared to existing works [1, 3], where numerous datasets are used and a wide range of explainability metrics are evaluated, the evaluation in this work does not give insights into the usefulness of the proposed method.\n\n- [1] Vielhaben et al., Explainable AI for time series via Virtual Inspection Layers, Pattern Recognition, 2024\n- [2] Brüsch et al., FreqRISE: Explaining time series using frequency masking, NLDL, 2025\n- [3] Brüsch et al., FLEXtime: Filterbank Learning to Explain Time Series, Explainable Artificial Intelligence, 2025", "questions": "1. How does the proposed method quantitatively compare to [1, 2, 3] in terms of established explainablity metrics like faithfulness, localization, complexity, and robustness?\n\n2. Apart from being specific for Integrated Gradients, how is the invertible transform introduced here to transfer between domains different from the transform introduced in [2]?\n\n- [1] Vielhaben et al., Explainable AI for time series via Virtual Inspection Layers, Pattern Recognition 2024\n- [2] Brüsch et al., FreqRISE: Explaining time series using frequency masking, NLDL, 2025\n- [3] Brüsch et al., FLEXtime: Filterbank Learning to Explain Time Series, Explainable Artificial Intelligence , 2025??", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761202857533}], "openreview_url": "https://openreview.net/forum?id=guXFowAqLQ", "arxiv_id": "2505.13100", "paper_pdf": "papers/guXFowAqLQ.pdf", "paper_pdf_sha256": "8860bc819b6cc14405adbb570d5da70781eda61792f43feb55348cd528d625e7", "paper_pdf_bytes": 3653808, "paper_pdf_source": "openreview", "code_url": "https://github.com/esl-epfl/cross-domain-saliency-maps", "code_repository": "esl-epfl/cross-domain-saliency-maps", "code_commit": "e4fee40c5a05601218a7268c9fb4ec27790dc760", "code_archive": "repos/guXFowAqLQ.zip", "code_archive_sha256": "68ca18eb6f14cb9bd95d895991c859403d1a09ce0bc8f927202c7bff2dfee63d", "code_archive_bytes": 2695630, "code_file_count": 20, "code_extensions": {".py": 16, ".ipynb": 4}, "github_disk_usage_kb": 2209, "github_languages": {"Python": 99091}, "github_archived": false, "github_pushed_at": "2026-05-04T23:49:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/time-series-saliency-maps-explaining-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "P98KMCf60l", "year": 2025, "status": "rejected", "title": "Theoretical Insights into Fine-Tuning Attention Mechanism: Generalization and Optimization", "authors": ["Xinhao Yao", "Hongjin Qian", "Xiaolin Hu", "Gengze Xu", "Yong Liu"], "authorids": ["~Xinhao_Yao1", "~Hongjin_Qian1", "~Xiaolin_Hu6", "~Gengze_Xu1", "~Yong_Liu7"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs), built on Transformer architectures, exhibit remarkable generalization across a wide range of tasks. However, fine-tuning these models for specific tasks remains resource-intensive due to their extensive parameterization.  \nIn this paper, we investigate two remarkable phenomena related to the attention mechanism during the fine-tuning of LLMs. The first phenomenon, termed “Unequal Importance of Attention Matrices,” highlights the impact of fine-tuning different weight matrices. It shows that optimizing the $\\mathbf{W}_v$ matrix yields significantly better performance than optimizing the $\\mathbf{W}_k$ matrix. Fine-tuning only the $\\mathbf{W}_q$ and $\\mathbf{W}_v$ matrices is computationally efficient while delivering results comparable to, or even better than fine-tuning all three matrices ($\\mathbf{W}_q$, $\\mathbf{W}_k$, and $\\mathbf{W}_v$). The second phenomenon, “Attention Matrices with Customized Learning Rate Leads to Better Convergence,” emphasizes the importance of assigning distinct learning rates to these matrices. Specifically, a higher learning rate for the $\\mathbf{W}_v$ matrix compared to $\\mathbf{W}_q$ and $\\mathbf{W}_k$ accelerates convergence and improves performance. Building on these insights, we propose a new strategy that improves fine-tuning efficiency in terms of both storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach, supporting our theoretical findings. Our analysis lays the theoretical groundwork for configuring and improving lightweight algorithms in LLMs fine-tuning.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "dPBN0vZxqV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6899/Reviewer_SWzh"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper investigates and addresses the resource-intensive nature of fine-tuning Large Language Models (LLMs) on transformer architectures, focusing specifically on the attention mechanism. The study identifies two key phenomena:\n\n1. **Different Impact**: The research demonstrates that optimizing the value matrix (Wv) in the attention mechanism leads to better performance improvements than optimizing the key matrix (Wk). Moreover, fine-tuning just the query (Wq) and value (Wv) matrices—not only reduces computational demands but also yields results that are on par with or surpass full matrix optimization (including Wq, Wk, Wv).\n\n2. **Efficient Convergence**: The paper highlights the importance of employing distinct learning rates for different matrices, particularly using a higher learning rate for Wv to speed up the convergence process.\n\nThe contributions of the paper are both theoretical and practical. Theoretically, it offers a new analysis on how selective fine-tuning of the attention matrices can enhance generalization bounds and improve memory efficiency. Practically, it proposes a fine-tuning strategy that optimizes model training time and resource use. The effectiveness of this approach is supported by experimental results on benchmark datasets, which validate the proposed methods and lay the groundwork for developing more efficient algorithms for fine-tuning LLMs.", "review_text": "This paper investigates and addresses the resource-intensive nature of fine-tuning Large Language Models (LLMs) on transformer architectures, focusing specifically on the attention mechanism. The study identifies two key phenomena:\n\n1. **Different Impact**: The research demonstrates that optimizing the value matrix (Wv) in the attention mechanism leads to better performance improvements than optimizing the key matrix (Wk). Moreover, fine-tuning just the query (Wq) and value (Wv) matrices—not only reduces computational demands but also yields results that are on par with or surpass full matrix optimization (including Wq, Wk, Wv).\n\n2. **Efficient Convergence**: The paper highlights the importance of employing distinct learning rates for different matrices, particularly using a higher learning rate for Wv to speed up the convergence process.\n\nThe contributions of the paper are both theoretical and practical. Theoretically, it offers a new analysis on how selective fine-tuning of the attention matrices can enhance generalization bounds and improve memory efficiency. Practically, it proposes a fine-tuning strategy that optimizes model training time and resource use. The effectiveness of this approach is supported by experimental results on benchmark datasets, which validate the proposed methods and lay the groundwork for developing more efficient algorithms for fine-tuning LLMs.", "strengths": "- **Importance of Understanding Attention Mechanisms During Fine-Tuning**: The challenge of gaining a deeper understanding of the attention mechanism during fine-tuning is a critical one. The approach developed in this paper has the potential to serve as a plug-and-play solution for achieving improved accuracy-efficiency trade-offs in LLM fine-tuning.\n\n- **Empirical and Theoretical Contributions**: This paper offers both empirical and theoretical analyses to elucidate the behavior of the attention mechanism during fine-tuning. The insights provided here could make a valuable contribution to the field, supporting the development of enhanced fine-tuning techniques that optimize the accuracy-efficiency trade-off.", "weaknesses": "- **Generalizability of the Proposed Approach**: The primary concern is the generalizability of the proposed approach. Specifically, the authors could enhance the analysis by demonstrating that the proposed method consistently improves LLM performance across diverse scenarios. To this end, it would be beneficial to include performance results under more complex, open-ended generation tasks, such as MT-Bench or comparable challenging benchmarks. Additionally, considering variations in model behavior, evaluating the approach on a broader range of LLMs, such as Mistral, would further strengthen the claim of general applicability.\n\n- **Visualization for Better Insight**: To deepen the analysis and understanding of the findings, visualizing the learned attention distributions across different settings could be valuable. By examining how attention distributions vary under different configurations, the authors could offer a more nuanced understanding of the observed effects, shedding light on the underlying phenomenon.", "questions": "- How will the proposed method perform on other PEFT techniques, such as DoRA [1]?\n\n[1] Liu, Shih-Yang, et al. \"Dora: Weight-decomposed low-rank adaptation.\" arXiv preprint arXiv:2402.09353 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates and addresses the resource-intensive nature of fine-tuning Large Language Models (LLMs) on transformer architectures, focusing specifically on the attention mechanism. The study identifies two key phenomena:\n\n1. **Different Impact**: The research demonstrates that optimizing the value matrix (Wv) in the attention mechanism leads to better performance improvements than optimizing the key matrix (Wk). Moreover, fine-tuning just the query (Wq) and value (Wv) matrices—not only reduces computational demands but also yields results that are on par with or surpass full matrix optimization (including Wq, Wk, Wv).\n\n2. **Efficient Convergence**: The paper highlights the importance of employing distinct learning rates for different matrices, particularly using a higher learning rate for Wv to speed up the convergence process.\n\nThe contributions of the paper are both theoretical and practical. Theoretically, it offers a new analysis on how selective fine-tuning of the attention matrices can enhance generalization bounds and improve memory efficiency. Practically, it proposes a fine-tuning strategy that optimizes model training time and resource use. The effectiveness of this approach is supported by experimental results on benchmark datasets, which validate the proposed methods and lay the groundwork for developing more efficient algorithms for fine-tuning LLMs.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- **Importance of Understanding Attention Mechanisms During Fine-Tuning**: The challenge of gaining a deeper understanding of the attention mechanism during fine-tuning is a critical one. The approach developed in this paper has the potential to serve as a plug-and-play solution for achieving improved accuracy-efficiency trade-offs in LLM fine-tuning.\n\n- **Empirical and Theoretical Contributions**: This paper offers both empirical and theoretical analyses to elucidate the behavior of the attention mechanism during fine-tuning. The insights provided here could make a valuable contribution to the field, supporting the development of enhanced fine-tuning techniques that optimize the accuracy-efficiency trade-off.", "weaknesses": "- **Generalizability of the Proposed Approach**: The primary concern is the generalizability of the proposed approach. Specifically, the authors could enhance the analysis by demonstrating that the proposed method consistently improves LLM performance across diverse scenarios. To this end, it would be beneficial to include performance results under more complex, open-ended generation tasks, such as MT-Bench or comparable challenging benchmarks. Additionally, considering variations in model behavior, evaluating the approach on a broader range of LLMs, such as Mistral, would further strengthen the claim of general applicability.\n\n- **Visualization for Better Insight**: To deepen the analysis and understanding of the findings, visualizing the learned attention distributions across different settings could be valuable. By examining how attention distributions vary under different configurations, the authors could offer a more nuanced understanding of the observed effects, shedding light on the underlying phenomenon.", "questions": "- How will the proposed method perform on other PEFT techniques, such as DoRA [1]?\n\n[1] Liu, Shih-Yang, et al. \"Dora: Weight-decomposed low-rank adaptation.\" arXiv preprint arXiv:2402.09353 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730707674519}, {"id": "QVin51rbTW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6899/Reviewer_FTN2"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper presents theoretical and empirical insights into improving the efficiency of fine-tuning large language models. The key contributions are as follows:\n\n1. Different Impact in Fine-Tuning: The paper demonstrates that fine-tuning only the value (`Wv`) and query (`Wq`) matrices of the attention mechanism is computationally efficient and can yield results comparable to or better than fine-tuning all three matrices (`Wq`, `Wk`, `Wv`). It suggests that `Wv` has a significantly greater impact on performance compared to `Wk`.\n\n2. Efficient Convergence: The authors show that using distinct learning rates for different matrices, particularly a higher rate for `Wv`, expedites convergence and optimizes the fine-tuning process. This insight provides a practical approach to achieving better fine-tuning results with fewer computational resources.\n\n3. Theoretical Analysis: The paper provides a theoretical analysis from the perspectives of generalization and optimization. It establishes that fine-tuning `Wq` and `Wv` enhances memory efficiency and generalization bounds while optimizing convergence dynamics by accelerating learning for specific matrices.\n\n4. Proposed Strategy: Building on these insights, the authors propose a strategy to improve fine-tuning efficiency, both in terms of storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach.", "review_text": "This paper presents theoretical and empirical insights into improving the efficiency of fine-tuning large language models. The key contributions are as follows:\n\n1. Different Impact in Fine-Tuning: The paper demonstrates that fine-tuning only the value (`Wv`) and query (`Wq`) matrices of the attention mechanism is computationally efficient and can yield results comparable to or better than fine-tuning all three matrices (`Wq`, `Wk`, `Wv`). It suggests that `Wv` has a significantly greater impact on performance compared to `Wk`.\n\n2. Efficient Convergence: The authors show that using distinct learning rates for different matrices, particularly a higher rate for `Wv`, expedites convergence and optimizes the fine-tuning process. This insight provides a practical approach to achieving better fine-tuning results with fewer computational resources.\n\n3. Theoretical Analysis: The paper provides a theoretical analysis from the perspectives of generalization and optimization. It establishes that fine-tuning `Wq` and `Wv` enhances memory efficiency and generalization bounds while optimizing convergence dynamics by accelerating learning for specific matrices.\n\n4. Proposed Strategy: Building on these insights, the authors propose a strategy to improve fine-tuning efficiency, both in terms of storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach.", "strengths": "- Originality: The originality of this paper lies in its focused approach to fine-tuning the attention mechanism of large language models. By exploring the selective fine-tuning of `Wv` and `Wq` matrices, the authors introduce a novel method that challenges the conventional approach of fine-tuning all attention matrices (`Wq`, `Wk`, `Wv`). The theoretical insights into the distinct roles of these matrices, combined with empirical validation, provide a fresh perspective on optimizing the attention mechanism. Additionally, the proposed use of differentiated learning rates for convergence further demonstrates creativity in enhancing the fine-tuning process.\n\n- Quality: The quality of the paper is reflected in its theoretical analysis and empirical validation. The authors employ both information-theoretic methods and optimization theory to support their claims, which significantly strengthens the credibility of their contributions.\n\n- Clarity: The paper is generally well-written, making the complex ideas accessible to a broad audience. \n\n- Significance: The significance of the paper lies in its potential to influence future research and practical implementations of fine-tuning large language models. This contribution is highly relevant for both academia and industry, as it provides a more resource-efficient approach to adapting pre-trained models for downstream tasks.", "weaknesses": "1. **Lack of Base Model Performance for Each Task**: The paper does not provide the performance of the base model before fine-tuning for each task, making it challenging to evaluate the true effectiveness of the fine-tuning methods. Including these baseline results would help contextualize the improvements made through fine-tuning.\n\n2. **GLUE Evaluation Is Too Simple for LLaMA3.1-8B**: The use of the GLUE benchmark to evaluate the LLaMA3.1-8B model is insufficient, as GLUE tasks are relatively simple compared to the capabilities of such a large model. Evaluating on more challenging tasks, such as coding or mathematical problem, could make the results more convincing and demonstrate the robustness of the proposed approach.\n\n3. **Lack of Ablation Studies**: The paper could benefit from more extensive ablation studies to isolate the effects of different components of the proposed method, such as the impact of learning rate scaling and the specific contribution of each matrix (`Wq`, `Wv`). This would provide a clearer understanding of the factors contributing to the observed performance gains.\n\n4. **Lack of Guidance on Hyperparameter `λ`**: The paper does not provide sufficient guidance on how to choose the hyperparameter `λ`, which controls the relative learning rates of different matrices. Without explicit guidelines or a heuristic for selecting `λ`, practitioners may find it difficult to replicate the reported results or apply the method effectively in different contexts.", "questions": "1. **Baseline Performance**: Could you provide the performance metrics of the base model before fine-tuning for each of the tasks? This would help in better understanding the relative improvements brought by your fine-tuning strategy.\n\n2. **More Challenging Evaluations**: Have you considered evaluating the LLaMA3.1-8B model on more complex benchmarks, such as tasks involving coding or mathematical problem? Including such challenging tasks would make your results more comprehensive and convincing.\n\n3. **Ablation Studies**: It would be helpful if you could add more ablation studies to isolate the effects of different components of the proposed method, such as the specific impact of `Wq` vs. `Wv` fine-tuning or the effect of learning rate scaling. This would provide a clearer picture of what drives the observed improvements.\n\n4. **Hyperparamete `λ`**: Could you provide more guidance on how to choose the hyperparameter `λ`?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents theoretical and empirical insights into improving the efficiency of fine-tuning large language models. The key contributions are as follows:\n\n1. Different Impact in Fine-Tuning: The paper demonstrates that fine-tuning only the value (`Wv`) and query (`Wq`) matrices of the attention mechanism is computationally efficient and can yield results comparable to or better than fine-tuning all three matrices (`Wq`, `Wk`, `Wv`). It suggests that `Wv` has a significantly greater impact on performance compared to `Wk`.\n\n2. Efficient Convergence: The authors show that using distinct learning rates for different matrices, particularly a higher rate for `Wv`, expedites convergence and optimizes the fine-tuning process. This insight provides a practical approach to achieving better fine-tuning results with fewer computational resources.\n\n3. Theoretical Analysis: The paper provides a theoretical analysis from the perspectives of generalization and optimization. It establishes that fine-tuning `Wq` and `Wv` enhances memory efficiency and generalization bounds while optimizing convergence dynamics by accelerating learning for specific matrices.\n\n4. Proposed Strategy: Building on these insights, the authors propose a strategy to improve fine-tuning efficiency, both in terms of storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- Originality: The originality of this paper lies in its focused approach to fine-tuning the attention mechanism of large language models. By exploring the selective fine-tuning of `Wv` and `Wq` matrices, the authors introduce a novel method that challenges the conventional approach of fine-tuning all attention matrices (`Wq`, `Wk`, `Wv`). The theoretical insights into the distinct roles of these matrices, combined with empirical validation, provide a fresh perspective on optimizing the attention mechanism. Additionally, the proposed use of differentiated learning rates for convergence further demonstrates creativity in enhancing the fine-tuning process.\n\n- Quality: The quality of the paper is reflected in its theoretical analysis and empirical validation. The authors employ both information-theoretic methods and optimization theory to support their claims, which significantly strengthens the credibility of their contributions.\n\n- Clarity: The paper is generally well-written, making the complex ideas accessible to a broad audience. \n\n- Significance: The significance of the paper lies in its potential to influence future research and practical implementations of fine-tuning large language models. This contribution is highly relevant for both academia and industry, as it provides a more resource-efficient approach to adapting pre-trained models for downstream tasks.", "weaknesses": "1. **Lack of Base Model Performance for Each Task**: The paper does not provide the performance of the base model before fine-tuning for each task, making it challenging to evaluate the true effectiveness of the fine-tuning methods. Including these baseline results would help contextualize the improvements made through fine-tuning.\n\n2. **GLUE Evaluation Is Too Simple for LLaMA3.1-8B**: The use of the GLUE benchmark to evaluate the LLaMA3.1-8B model is insufficient, as GLUE tasks are relatively simple compared to the capabilities of such a large model. Evaluating on more challenging tasks, such as coding or mathematical problem, could make the results more convincing and demonstrate the robustness of the proposed approach.\n\n3. **Lack of Ablation Studies**: The paper could benefit from more extensive ablation studies to isolate the effects of different components of the proposed method, such as the impact of learning rate scaling and the specific contribution of each matrix (`Wq`, `Wv`). This would provide a clearer understanding of the factors contributing to the observed performance gains.\n\n4. **Lack of Guidance on Hyperparameter `λ`**: The paper does not provide sufficient guidance on how to choose the hyperparameter `λ`, which controls the relative learning rates of different matrices. Without explicit guidelines or a heuristic for selecting `λ`, practitioners may find it difficult to replicate the reported results or apply the method effectively in different contexts.", "questions": "1. **Baseline Performance**: Could you provide the performance metrics of the base model before fine-tuning for each of the tasks? This would help in better understanding the relative improvements brought by your fine-tuning strategy.\n\n2. **More Challenging Evaluations**: Have you considered evaluating the LLaMA3.1-8B model on more complex benchmarks, such as tasks involving coding or mathematical problem? Including such challenging tasks would make your results more comprehensive and convincing.\n\n3. **Ablation Studies**: It would be helpful if you could add more ablation studies to isolate the effects of different components of the proposed method, such as the specific impact of `Wq` vs. `Wv` fine-tuning or the effect of learning rate scaling. This would provide a clearer picture of what drives the observed improvements.\n\n4. **Hyperparamete `λ`**: Could you provide more guidance on how to choose the hyperparameter `λ`?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730703531023}, {"id": "rHlVBE4XOm", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6899/Reviewer_f8uV"], "rating": 5, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper explores fine-tuning strategies for the attention mechanism from a theoretical perspective, focusing on generalization and optimization issues, and claims to provide theoretical insights to guide algorithm design. It has two major propositions, one is fine-tuning WqWv matrix is more efficient and can achieve comparable results as fine-tuning WqWkWv matrix. Another is fine-tuning Wq and Wv should use different learning rates for more efficient convergence.", "review_text": "This paper explores fine-tuning strategies for the attention mechanism from a theoretical perspective, focusing on generalization and optimization issues, and claims to provide theoretical insights to guide algorithm design. It has two major propositions, one is fine-tuning WqWv matrix is more efficient and can achieve comparable results as fine-tuning WqWkWv matrix. Another is fine-tuning Wq and Wv should use different learning rates for more efficient convergence.", "strengths": "The idea of using different local learning rates for Wq and Wv is interesting. As V is timed with attention logits, q and v should have more different gradient distributions than q and k. Using different local learning rates for these two types of matrices is reasonable and worth exploring.", "weaknesses": "1. The writing and presentation are rather poor. Both sentence level and logical level polishment are suggested for this work. \n    E.g. In line 16~22, \"In this paper, we investigate two remarkable phenomena ... with a higher learning rate for the Wv matrix expediting convergence.\" The ordering of the sentences and the way of structuring the arguments add unnecessary difficulty to reading. This kind of problem happens across the whole manuscript.\n    A suggestion is, use multiple statement sentences instead of clauses. A rewritten version of these sentences is \" In this paper, we investigate two remarkable phenomena associated with attention mechanism, during the fine-tuning of LLMs. One is 'Different Impact' and another is 'Efficient Convergence'. ......\" Another point is that the naming of \"Different Impact\" and \"Efficient Convergence\" doesn't help with understanding. Consider other ways to name them.\n\n2. For the \"Different Impact\", the authors claimed that fine-tuning Wq&Wv is favored over Wq,Wk,Wv together. However, this argument is not supported by their experimental results in Table 1. It can be seen from the table that fine-tuning Wq,Wk,Wv together still gets the best performance for most of the cases. I wonder if the authors are writing the arguments and conducting experiments separately without any discussion. The authors should summarize their findings according to the experiment results.", "questions": "1.  In line 160~193, the author writes \"As seen in the table, we can see a clear trend where solely updating the Wv matrix outperforms just learning the Wq,Wk matrix. Interestingly, the combination of fine-tuning both Wq and Wv often leads to performance that matches or even exceeds that achieved by fine-tuning all three matrices Wq,Wk, and Wv\". This statement is not supported by the results in Table 1, as solely updating the Wv is inferior to updating Wq,Wk together, and fine-tuning all three matrices Wq,Wk, and Wv achieves the best result in most cases. The authors should first rewrite the whole section of \"Different Impact\" to have a consistent argument with the experiment results, and investigate more tasks to see whether there are certain types of tasks in which \"different impact\" is true.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores fine-tuning strategies for the attention mechanism from a theoretical perspective, focusing on generalization and optimization issues, and claims to provide theoretical insights to guide algorithm design. It has two major propositions, one is fine-tuning WqWv matrix is more efficient and can achieve comparable results as fine-tuning WqWkWv matrix. Another is fine-tuning Wq and Wv should use different learning rates for more efficient convergence.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "The idea of using different local learning rates for Wq and Wv is interesting. As V is timed with attention logits, q and v should have more different gradient distributions than q and k. Using different local learning rates for these two types of matrices is reasonable and worth exploring.", "weaknesses": "1. The writing and presentation are rather poor. Both sentence level and logical level polishment are suggested for this work. \n    E.g. In line 16~22, \"In this paper, we investigate two remarkable phenomena ... with a higher learning rate for the Wv matrix expediting convergence.\" The ordering of the sentences and the way of structuring the arguments add unnecessary difficulty to reading. This kind of problem happens across the whole manuscript.\n    A suggestion is, use multiple statement sentences instead of clauses. A rewritten version of these sentences is \" In this paper, we investigate two remarkable phenomena associated with attention mechanism, during the fine-tuning of LLMs. One is 'Different Impact' and another is 'Efficient Convergence'. ......\" Another point is that the naming of \"Different Impact\" and \"Efficient Convergence\" doesn't help with understanding. Consider other ways to name them.\n\n2. For the \"Different Impact\", the authors claimed that fine-tuning Wq&Wv is favored over Wq,Wk,Wv together. However, this argument is not supported by their experimental results in Table 1. It can be seen from the table that fine-tuning Wq,Wk,Wv together still gets the best performance for most of the cases. I wonder if the authors are writing the arguments and conducting experiments separately without any discussion. The authors should summarize their findings according to the experiment results.", "questions": "1.  In line 160~193, the author writes \"As seen in the table, we can see a clear trend where solely updating the Wv matrix outperforms just learning the Wq,Wk matrix. Interestingly, the combination of fine-tuning both Wq and Wv often leads to performance that matches or even exceeds that achieved by fine-tuning all three matrices Wq,Wk, and Wv\". This statement is not supported by the results in Table 1, as solely updating the Wv is inferior to updating Wq,Wk together, and fine-tuning all three matrices Wq,Wk, and Wv achieves the best result in most cases. The authors should first rewrite the whole section of \"Different Impact\" to have a consistent argument with the experiment results, and investigate more tasks to see whether there are certain types of tasks in which \"different impact\" is true.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730293586932}, {"id": "XBEcrT9P8c", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6899/Reviewer_TXeN"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper investigates two phenomena observed during the fine-tuning of Transformer LLMs, focusing on the attention mechanism. The first phenomenon is the \"Different Impact,\" where optimizing the Wv (value) matrix significantly improves performance more than optimizing the Wk (key) matrix. The second is \"Efficient Convergence,\" where using distinct learning rates for different matrices is crucial for optimal performance, with a higher learning rate for Wv being beneficial for faster convergence. The paper provides theoretical analysis on these phenomena from the perspectives of generalization and optimization. Based on this, this paper proposes only finetuning Wq and Wv, and with a larger learning rate for Wv.", "review_text": "The paper investigates two phenomena observed during the fine-tuning of Transformer LLMs, focusing on the attention mechanism. The first phenomenon is the \"Different Impact,\" where optimizing the Wv (value) matrix significantly improves performance more than optimizing the Wk (key) matrix. The second is \"Efficient Convergence,\" where using distinct learning rates for different matrices is crucial for optimal performance, with a higher learning rate for Wv being beneficial for faster convergence. The paper provides theoretical analysis on these phenomena from the perspectives of generalization and optimization. Based on this, this paper proposes only finetuning Wq and Wv, and with a larger learning rate for Wv.", "strengths": "The observation that fine-tuning only Wq and Wv​ matrices yields performance gains comparable to finetuning of Wq, Wk and Wv is interesting. Trying to analyze and uncover the reason can benefit the research commnity.", "weaknesses": "See my questions.", "questions": "1. My biggest concern with this paper is that the key observation seems to originate from the LoRA paper: *\"Adapting both \\( W_q \\) and \\( W_v \\) gives the best performance overall\"* (this sentence is quoted from the LoRA paper), and even the experiment setup is similar to Table 5 in the LoRA paper. Deriving key observations from other works isn’t an issue (or saying we share the same observation/insights), but I believe it should be directly and explicitly pointed out, rather than presented as your own observation, concerning the lora paper is very famous for several years and you do cite it. Or could you clarify what differences there are between your work and the LoRA paper? Please provide a detailed comparison between their findings and those in the LoRA paper, highlighting any novel contributions or deeper insights their work provides. This would help clarify the paper's originality and contribution to the field.\n\n2. As you pointed out on line 205, \\( W_q \\) and \\( W_k \\) can be treated as a single unit. In linear algebra, two matrices multiplied without an intermediate activation can be equivalent to a single matrix. This could explain why fine-tuning only \\( W_q \\) and \\( W_v \\) achieves comparable accuracy to tuning all matrices. If this is the case, have you tried fine-tuning only \\( W_k \\) and \\( W_v \\)? It seems that neither the LoRA paper nor your work has conducted this experiment. Could you please include this experiment in your study or explain why not to perform it. This could provide valuable insights and strengthen the paper's analysis.\n\n3. What are the main conclusions of the theoretical analyses in Sections 3 and 4? I found it difficult to follow your proofs. Please include a clear summary of the main conclusions from their theoretical analyses at the end of Sections 3 and 4. I would suggest to provide more intuitive explanations to help readers better understand the theoretical analyses and proofs.\n\n4. Section 5 appears to be simply fine-tuning \\( W_q \\) and \\( W_v \\) with a search for \\lambda. Do you have any insights on how to determine \\lambda? could you provide guidelines or heuristics for determining an appropriate \\lambda value based on your theoretical and empirical results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates two phenomena observed during the fine-tuning of Transformer LLMs, focusing on the attention mechanism. The first phenomenon is the \"Different Impact,\" where optimizing the Wv (value) matrix significantly improves performance more than optimizing the Wk (key) matrix. The second is \"Efficient Convergence,\" where using distinct learning rates for different matrices is crucial for optimal performance, with a higher learning rate for Wv being beneficial for faster convergence. The paper provides theoretical analysis on these phenomena from the perspectives of generalization and optimization. Based on this, this paper proposes only finetuning Wq and Wv, and with a larger learning rate for Wv.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "The observation that fine-tuning only Wq and Wv​ matrices yields performance gains comparable to finetuning of Wq, Wk and Wv is interesting. Trying to analyze and uncover the reason can benefit the research commnity.", "weaknesses": "See my questions.", "questions": "1. My biggest concern with this paper is that the key observation seems to originate from the LoRA paper: *\"Adapting both \\( W_q \\) and \\( W_v \\) gives the best performance overall\"* (this sentence is quoted from the LoRA paper), and even the experiment setup is similar to Table 5 in the LoRA paper. Deriving key observations from other works isn’t an issue (or saying we share the same observation/insights), but I believe it should be directly and explicitly pointed out, rather than presented as your own observation, concerning the lora paper is very famous for several years and you do cite it. Or could you clarify what differences there are between your work and the LoRA paper? Please provide a detailed comparison between their findings and those in the LoRA paper, highlighting any novel contributions or deeper insights their work provides. This would help clarify the paper's originality and contribution to the field.\n\n2. As you pointed out on line 205, \\( W_q \\) and \\( W_k \\) can be treated as a single unit. In linear algebra, two matrices multiplied without an intermediate activation can be equivalent to a single matrix. This could explain why fine-tuning only \\( W_q \\) and \\( W_v \\) achieves comparable accuracy to tuning all matrices. If this is the case, have you tried fine-tuning only \\( W_k \\) and \\( W_v \\)? It seems that neither the LoRA paper nor your work has conducted this experiment. Could you please include this experiment in your study or explain why not to perform it. This could provide valuable insights and strengthen the paper's analysis.\n\n3. What are the main conclusions of the theoretical analyses in Sections 3 and 4? I found it difficult to follow your proofs. Please include a clear summary of the main conclusions from their theoretical analyses at the end of Sections 3 and 4. I would suggest to provide more intuitive explanations to help readers better understand the theoretical analyses and proofs.\n\n4. Section 5 appears to be simply fine-tuning \\( W_q \\) and \\( W_v \\) with a search for \\lambda. Do you have any insights on how to determine \\lambda? could you provide guidelines or heuristics for determining an appropriate \\lambda value based on your theoretical and empirical results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730096695714}], "openreview_url": "https://openreview.net/forum?id=P98KMCf60l", "arxiv_id": "2410.02247", "paper_pdf": "papers/P98KMCf60l.pdf", "paper_pdf_sha256": "416981fbcc92b73def6bed50eec45b587bb156210dfb9a271d5c51b83bd6b2ce", "paper_pdf_bytes": 598252, "paper_pdf_source": "openreview", "code_url": "https://github.com/chen123CtrlS/LightweightAtt", "code_repository": "chen123CtrlS/LightweightAtt", "code_commit": "6e6cfc3d981194abd565c298e372df9899a2fc5f", "code_archive": "repos/P98KMCf60l.zip", "code_archive_sha256": "c7b0b6e4e940130771147bafb71ef24abbf192f562bed60007394a4f37ffc985", "code_archive_bytes": 167302, "code_file_count": 13, "code_extensions": {".sh": 8, ".py": 5}, "github_disk_usage_kb": 157, "github_languages": {"Python": 43600, "Shell": 10627}, "github_archived": false, "github_pushed_at": "2024-10-01T08:28:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/theoretical-insights-into-fine-tuning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OMwD6pGYB4", "year": 2024, "status": "rejected", "title": "A Distributional Analogue to the Successor Representation", "authors": ["Jesse Farebrother", "Harley Wiltzer", "Arthur Gretton", "Yunhao Tang", "Andre Barreto", "Will Dabney", "Marc G Bellemare", "Mark Rowland"], "authorids": ["~Jesse_Farebrother1", "~Harley_Wiltzer1", "~Arthur_Gretton1", "~Yunhao_Tang1", "~Andre_Barreto1", "~Will_Dabney1", "~Marc_G_Bellemare1", "~Mark_Rowland1"], "authors_source": "OpenReview API", "abstract": "This paper contributes a new approach for distributional reinforcement learning which allows for a \nclean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences from behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour.\nWe model the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. \nExtending γ-models (Janner et al., 2020), we propose an algorithm that learns the distributional SM from samples by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable in the context of learning generative models of state.\nAs an illustration of the practical usefulness of the distributional successor measure, we show that it\nenables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "ptwXg6L7jL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7962/Reviewer_xHdx"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "They investigate the distributional counterpart of discounted occupancy measures, which they refer to as the distributional success measure (DSM). Leveraging the forward Bellman equation for DSM, they introduce a novel approach for approximating DSM. Their approach entails modeling DSM as a collection of generative models, referred to as $\\delta$-models, and employing a maximum mean discrepancy as a loss function to quantify the dissimilarity between \"distributions over distributions\" stemming from the Bellman equation.", "review_text": "They investigate the distributional counterpart of discounted occupancy measures, which they refer to as the distributional success measure (DSM). Leveraging the forward Bellman equation for DSM, they introduce a novel approach for approximating DSM. Their approach entails modeling DSM as a collection of generative models, referred to as $\\delta$-models, and employing a maximum mean discrepancy as a loss function to quantify the dissimilarity between \"distributions over distributions\" stemming from the Bellman equation.", "strengths": "Their research problem is both intriguing and relatively novel within the RL community. Their approach to estimating DSM appears to be innovative, although I do have some fundamental concerns that I will address later.", "weaknesses": "I have some reservations regarding the current proposed methods.  \n* Firstly, it remains unclear why equal weights are utilized in equation (10). It seems plausible that we should consider learning these weights.\n* Secondly, there is a lack of guidance on selecting the value of  $m$ in equation (10), or determining how many $m$ values are required.\n* Thirdly, there appears to be a dearth of theoretical justification for the effectiveness of this modeling and approximation approach. While equation (10) may seem suitable if it exactly represents the true SDM, practical implementation would not align with this ideal scenario.", "questions": "* Questions are written in a previous paragraph. \n\n* Let's say we have X = {0,1} (binary). Then, a set of $P(X)$ is parametrized by just one parameter $\\mu \\in [0,1]$. Sp, $P(P(X))$ is a set of distributions over $[0,1]$. So, if I understand correctly, learning SDM is equivalent to estimating a distribution over  $[0,1]$. Even in this simple case, does the author's approach have any theoretical guarantee? (finite $m$ and equal weights look restrictive?)\n\n\n### Suggestion for presentation \n\n* Equation (4) may appear somewhat elementary to researcher within the RL community. The current phrasing, such as \"Blier et al. 2021 derived this equation...,\" seems inaccurate and should be revised. I believe that this equation had already gained widespread recognition prior to the work of Blier et al. (2021), as it is commonly featured in numerous standard RL texts and papers.\n\n* As a related point, I generally believe the author should not emphasize whether the definition pertains to discrete or continuous spaces to such an extent. The transformation from a discrete space (when a base measure is the counting measure) to a continuous space (when the base measure is a Lebsgue measure)  is typically straightforward for individuals with a basic understanding of probability. Therefore, in Section 3.2, the statement \"though this result is novel in the case of more general state spaces\" may be somewhat misleading. I suggest that this aspect should not be categorized as \"novel.\" Instead, I recommend that the author simply highlight the distinctions between the two contexts (SDM and standard distribution RL with rewards).\n\n* It is somewhat unclear which parameters are precisely optimized throughout the entire algorithm. As I understand it, the author optimizes parameters for all generative models simultaneously in equation (16). It would be beneficial to present the algorithm using an algorithmic environment for clarity.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "They investigate the distributional counterpart of discounted occupancy measures, which they refer to as the distributional success measure (DSM). Leveraging the forward Bellman equation for DSM, they introduce a novel approach for approximating DSM. Their approach entails modeling DSM as a collection of generative models, referred to as $\\delta$-models, and employing a maximum mean discrepancy as a loss function to quantify the dissimilarity between \"distributions over distributions\" stemming from the Bellman equation.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Their research problem is both intriguing and relatively novel within the RL community. Their approach to estimating DSM appears to be innovative, although I do have some fundamental concerns that I will address later.", "weaknesses": "I have some reservations regarding the current proposed methods.  \n* Firstly, it remains unclear why equal weights are utilized in equation (10). It seems plausible that we should consider learning these weights.\n* Secondly, there is a lack of guidance on selecting the value of  $m$ in equation (10), or determining how many $m$ values are required.\n* Thirdly, there appears to be a dearth of theoretical justification for the effectiveness of this modeling and approximation approach. While equation (10) may seem suitable if it exactly represents the true SDM, practical implementation would not align with this ideal scenario.", "questions": "* Questions are written in a previous paragraph. \n\n* Let's say we have X = {0,1} (binary). Then, a set of $P(X)$ is parametrized by just one parameter $\\mu \\in [0,1]$. Sp, $P(P(X))$ is a set of distributions over $[0,1]$. So, if I understand correctly, learning SDM is equivalent to estimating a distribution over  $[0,1]$. Even in this simple case, does the author's approach have any theoretical guarantee? (finite $m$ and equal weights look restrictive?)\n\n\n### Suggestion for presentation \n\n* Equation (4) may appear somewhat elementary to researcher within the RL community. The current phrasing, such as \"Blier et al. 2021 derived this equation...,\" seems inaccurate and should be revised. I believe that this equation had already gained widespread recognition prior to the work of Blier et al. (2021), as it is commonly featured in numerous standard RL texts and papers.\n\n* As a related point, I generally believe the author should not emphasize whether the definition pertains to discrete or continuous spaces to such an extent. The transformation from a discrete space (when a base measure is the counting measure) to a continuous space (when the base measure is a Lebsgue measure)  is typically straightforward for individuals with a basic understanding of probability. Therefore, in Section 3.2, the statement \"though this result is novel in the case of more general state spaces\" may be somewhat misleading. I suggest that this aspect should not be categorized as \"novel.\" Instead, I recommend that the author simply highlight the distinctions between the two contexts (SDM and standard distribution RL with rewards).\n\n* It is somewhat unclear which parameters are precisely optimized throughout the entire algorithm. As I understand it, the author optimizes parameters for all generative models simultaneously in equation (16). It would be beneficial to present the algorithm using an algorithmic environment for clarity.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698990660694}, {"id": "L0FO6owlD4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7962/Reviewer_9gwA"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces a novel distributional RL algorithm that learns the distributional successor measure from training samples. This method allows a clean separation of transition structure, i.e., state occupancy measure, and reward, enabling zero-shot risk-sensitive policy evaluation.", "review_text": "This paper introduces a novel distributional RL algorithm that learns the distributional successor measure from training samples. This method allows a clean separation of transition structure, i.e., state occupancy measure, and reward, enabling zero-shot risk-sensitive policy evaluation.", "strengths": "1. The proposed method cast the problem of learning the return distribution into learning the distribution of random occupancy measure, decoupling the transition structure and reward functions and thus enabling zero-shot risk-sensitive policy evaluation. In this sense, the proposed method is novel.\n\n2. This paper presented a practical algorithm for training the $\\delta$-models, adopting diverse generative models to approximate the distribution of random occupancy measure. The training procedure itself has merit and can be potentially beneficial to the other distribution estimation tasks.", "weaknesses": "1. Similar to the conventional successor feature and successor measure, the decoupling of the transition structure and reward functions still assumes a fixed policy and transition dynamic, limiting the usefulness of the proposed method.\n\n2. The usefulness of the distributional SM is quite limited at this point. I would recommend discussing more about the potential applications of the learned distributional SM other than the zero-shot distributional policy evaluation.", "questions": "Will the setting of $\\gamma = 0.95$ limit the usefulness of the proposed method in practice when we care about the return of a long episode?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel distributional RL algorithm that learns the distributional successor measure from training samples. This method allows a clean separation of transition structure, i.e., state occupancy measure, and reward, enabling zero-shot risk-sensitive policy evaluation.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The proposed method cast the problem of learning the return distribution into learning the distribution of random occupancy measure, decoupling the transition structure and reward functions and thus enabling zero-shot risk-sensitive policy evaluation. In this sense, the proposed method is novel.\n\n2. This paper presented a practical algorithm for training the $\\delta$-models, adopting diverse generative models to approximate the distribution of random occupancy measure. The training procedure itself has merit and can be potentially beneficial to the other distribution estimation tasks.", "weaknesses": "1. Similar to the conventional successor feature and successor measure, the decoupling of the transition structure and reward functions still assumes a fixed policy and transition dynamic, limiting the usefulness of the proposed method.\n\n2. The usefulness of the distributional SM is quite limited at this point. I would recommend discussing more about the potential applications of the learned distributional SM other than the zero-shot distributional policy evaluation.", "questions": "Will the setting of $\\gamma = 0.95$ limit the usefulness of the proposed method in practice when we care about the return of a long episode?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698735433731}, {"id": "iJARuKRBq6", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7962/Reviewer_JnMG"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper extends the successor representation to distributional RL, and proposes the $\\delta$-model that learns the distributional successor measure.", "review_text": "This paper extends the successor representation to distributional RL, and proposes the $\\delta$-model that learns the distributional successor measure.", "strengths": "The idea of combining distribution RL with successor representation is interesting, which combines the merits of both SR and distributional RL. I think this is a promising and important direction. The theoretical analysis is sound.", "weaknesses": "The experiment benchmark (windy grid world) is somehow toy compared with other distributional RL papers.", "questions": "1. As for Eq.7, are there any requirements for the reward function $r$ besides deterministic? For instance, does it require $r$ to be linear? BTW, is $r$ assumed to be known in experiments?\n\n2. Can you further discuss the relationship between previous work like Zhang et al 2021b, which also learn multi-dimensional return distribution via MMD?\n\n3. For distributional RL papers, a common benchmark is visual input environments like Atari. I think current benchmark (windy gridworld) is somehow toy. It will be helpful to see experiments with larger scale. What’s more, besides zero-shot policy evaluation, another advantage of SR is multitask training. Can the proposed method be combined with the multitask training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper extends the successor representation to distributional RL, and proposes the $\\delta$-model that learns the distributional successor measure.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The idea of combining distribution RL with successor representation is interesting, which combines the merits of both SR and distributional RL. I think this is a promising and important direction. The theoretical analysis is sound.", "weaknesses": "The experiment benchmark (windy grid world) is somehow toy compared with other distributional RL papers.", "questions": "1. As for Eq.7, are there any requirements for the reward function $r$ besides deterministic? For instance, does it require $r$ to be linear? BTW, is $r$ assumed to be known in experiments?\n\n2. Can you further discuss the relationship between previous work like Zhang et al 2021b, which also learn multi-dimensional return distribution via MMD?\n\n3. For distributional RL papers, a common benchmark is visual input environments like Atari. I think current benchmark (windy gridworld) is somehow toy. It will be helpful to see experiments with larger scale. What’s more, besides zero-shot policy evaluation, another advantage of SR is multitask training. Can the proposed method be combined with the multitask training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698725787269}, {"id": "TKgBjgz99O", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7962/Reviewer_jmar"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "1 poor", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a distribution version of success feature, coined as the successor measure. The successor measure is defined as, conditioned on an initial state $x$, the Dirac distribution conditioned on a random trajectory induced by the policy. The paper proposes the Bellman backup of such successor measure, and proposes a recipe to estimate such successor measure using the proposed $\\delta$-model. Finally the paper evaluates the proposed method on some simple environments.", "review_text": "The paper proposes a distribution version of success feature, coined as the successor measure. The successor measure is defined as, conditioned on an initial state $x$, the Dirac distribution conditioned on a random trajectory induced by the policy. The paper proposes the Bellman backup of such successor measure, and proposes a recipe to estimate such successor measure using the proposed $\\delta$-model. Finally the paper evaluates the proposed method on some simple environments.", "strengths": "1. The paper proposes a new method for distributional RL by estimating the occupancy measure of the policy, instead of its expectation, which seems like a natural way for the problem (vs. measuring the distribution of the value function.)\n\n2. The paper includes good explanations for the new mathematical definitions which help the reviewer to understand the new concepts.", "weaknesses": "1. A few concepts are quite confusing along the paper. First, why is it necessary to redefine the occupancy measure/state(-action) distribution as successor representation or successor measure?\n\nSecond, why is it necessary to define the new \"random occupancy measure\", which is a distribution over distribution, but the inner distribution is just the Dirac distribution which is determined by the realization of the sample of the outer distribution? From my understanding, if one wants a distribution over the value function, suppose they have already have the occupancy measure, they could easily define the distribution of the value function (which is in $\\Delta(\\mathbb{R}^1)$) by projecting the occupancy measure to the reward vector. I think this is also indicated by prop. 1: to define the distributional return, one needs to take one expectation over the random occupancy measure. \n\nOverall, the theory results are rather limited. Most results seem like straight forward extension from the occupancy measure results to the random occupancy measure version. \n\n2. The significance of section 5.2 is unclear to me. How to tune MMD does not directly relate to the significance of the paper, and many description seems not significant (for example, the detailed description of the median trick).\n\n3. Since the theory contribution is rather limited, the experiment of the paper should be greatly improved. First, the paper should compare with other distribution RL methods, and the current benchmarks are also pretty easy.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a distribution version of success feature, coined as the successor measure. The successor measure is defined as, conditioned on an initial state $x$, the Dirac distribution conditioned on a random trajectory induced by the policy. The paper proposes the Bellman backup of such successor measure, and proposes a recipe to estimate such successor measure using the proposed $\\delta$-model. Finally the paper evaluates the proposed method on some simple environments.", "soundness": "3 good", "presentation": "1 poor", "contribution": "3 good", "strengths": "1. The paper proposes a new method for distributional RL by estimating the occupancy measure of the policy, instead of its expectation, which seems like a natural way for the problem (vs. measuring the distribution of the value function.)\n\n2. The paper includes good explanations for the new mathematical definitions which help the reviewer to understand the new concepts.", "weaknesses": "1. A few concepts are quite confusing along the paper. First, why is it necessary to redefine the occupancy measure/state(-action) distribution as successor representation or successor measure?\n\nSecond, why is it necessary to define the new \"random occupancy measure\", which is a distribution over distribution, but the inner distribution is just the Dirac distribution which is determined by the realization of the sample of the outer distribution? From my understanding, if one wants a distribution over the value function, suppose they have already have the occupancy measure, they could easily define the distribution of the value function (which is in $\\Delta(\\mathbb{R}^1)$) by projecting the occupancy measure to the reward vector. I think this is also indicated by prop. 1: to define the distributional return, one needs to take one expectation over the random occupancy measure. \n\nOverall, the theory results are rather limited. Most results seem like straight forward extension from the occupancy measure results to the random occupancy measure version. \n\n2. The significance of section 5.2 is unclear to me. How to tune MMD does not directly relate to the significance of the paper, and many description seems not significant (for example, the detailed description of the median trick).\n\n3. Since the theory contribution is rather limited, the experiment of the paper should be greatly improved. First, the paper should compare with other distribution RL methods, and the current benchmarks are also pretty easy.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698550639198}], "openreview_url": "https://openreview.net/forum?id=OMwD6pGYB4", "arxiv_id": "2402.08530", "paper_pdf": "papers/OMwD6pGYB4.pdf", "paper_pdf_sha256": "0845864f5f06f3199b4414634b20e102c72eb4597a185d96a25755b65f69cfdf", "paper_pdf_bytes": 1981327, "paper_pdf_source": "openreview", "code_url": "https://github.com/JesseFarebro/distributional-sr", "code_repository": "JesseFarebro/distributional-sr", "code_commit": "f379d0f65914459db069cf0089bdc54451f3e563", "code_archive": "repos/OMwD6pGYB4.zip", "code_archive_sha256": "7d433a5be9bd4a09a739222a86251ab8346b4fc199e0e82fcf93db80185526f8", "code_archive_bytes": 128944, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 135, "github_languages": {"Python": 91877}, "github_archived": false, "github_pushed_at": "2024-11-08T00:11:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-distributional-analogue-to-the-successor"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lEFM4OTz62c", "year": 2023, "status": "rejected", "title": "Group-level Brain Decoding with Deep Learning", "authors": ["Richard Csaky", "Mats W.J. Van Es", "Oiwi Parker Jones", "Mark Woolrich"], "authorids": ["~Richard_Csaky1", "~Mats_W.J._Van_Es1", "~Oiwi_Parker_Jones1", "~Mark_Woolrich1"], "authors_source": "OpenReview API", "abstract": "Decoding experimental variables from brain imaging data is gaining popularity, with applications in brain-computer interfaces and the study of neural representations. Decoding is typically subject-specific and does not generalise well over subjects. Here, we propose a method that uses subject embedding, analogous to word embedding in Natural Language Processing, to learn and exploit the structure in between subject variability as part of a decoding model, our adaptation of the WaveNet architecture for classification. We apply this to magnetoencephalography data, where 15 subjects viewed 118 different images, with 30 examples per image; to classify images using the entire 1s window following image presentation. We show that the combination of deep learning and subject embedding is crucial to closing the performance gap between subject- and group-level decoding models. Importantly, group models outperform subject models on low-accuracy subjects (but impair high-accuracy subjects) and can be helpful for initialising subject models. The potential of such group modelling is even higher with bigger datasets. To better enable physiological interpretation at the group level we demonstrate the use of permutation feature importance developing insights into the spatio-temporal and spectral information encoded in the models. All code is available on GitHub.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rXdaH_WRnl", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1750/Reviewer_MQXD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This manuscript analyzes the use of subject embeddings to enhance deep learning models trained to decode MEG recordings of several subjects. Concretely, the work analyzes decoding MEG data to classify which of 118 different images 15 different subjects looked at. They use a convolutional neural network using dilated convolutions inspired by WaveNet. and evaluate using a vanilla model without subject embeddings as well as a larger model that also uses subject embeddings. They report that using the subject embeddings improves accuracies and when combined with finetuning on an individual subject can even slightly outperform subject-level models. Furthermore, analysis of the trained models partly shows plausible results.", "review_text": "Overall, this seems a nice manuscript with some additional analyses needed in my view.", "strengths": "**Update**\nDue to the including of the reference, the clarifications regarding number of layers and the additional gradient-based analyses, have increased my score.\n\n**Pre-rebuttal**\nThe manuscript is quite detailed and evaluates many aspects of a clearly described idea.\n\nIt is valuable that the work looks at various settings like linear/nonlinear with/without embedding and also including subject-level finetuning. Also the additional analysis of the trained models is interesting.\n\nSome open questions for me:\n“For SL modelling, the Wavenet Classifier contains 3 convolutional layers, whereas for group modelling it has 6, further motivated in Section 4. “\n\nI could not quite find the part where this is written maybe I missed it? In any case, would be nice to also evaluate the 6-layer classifier on subject-level decoding, as it may also extract different frequencies with its deeper structure. So these results would be an important addition, also to disentangle effects of larger model and effect of subject embedding etc.\n\nIn general, it would be good to evaluate further published EEG deep learning models/pipelines on this task to know how the reported accuracies compare to the literature. Also, I assume this dataset has been decoded before, how where accuracies in the literature?\n\nI was also missing at least one reference on transfer learning for neurophysiological recordings, \nhttps://iopscience.iop.org/article/10.1088/1741-2552/abb7a7 \nAlso this or other works would provide interesting baselines for the method presented in this manuscript.\n\n\nRegarding the PFI analysis, it would be nice to additionally perform a computationally simpler gradient-based analysis, note that both fourier transform and inverse fourier transform are differentiable and hence one can also compute gradients on fourier coefficients. It would be good to see if these analysis agree or in how far they disagree.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This manuscript analyzes the use of subject embeddings to enhance deep learning models trained to decode MEG recordings of several subjects. Concretely, the work analyzes decoding MEG data to classify which of 118 different images 15 different subjects looked at. They use a convolutional neural network using dilated convolutions inspired by WaveNet. and evaluate using a vanilla model without subject embeddings as well as a larger model that also uses subject embeddings. They report that using the subject embeddings improves accuracies and when combined with finetuning on an individual subject can even slightly outperform subject-level models. Furthermore, analysis of the trained models partly shows plausible results.", "strength_and_weaknesses": "**Update**\nDue to the including of the reference, the clarifications regarding number of layers and the additional gradient-based analyses, have increased my score.\n\n**Pre-rebuttal**\nThe manuscript is quite detailed and evaluates many aspects of a clearly described idea.\n\nIt is valuable that the work looks at various settings like linear/nonlinear with/without embedding and also including subject-level finetuning. Also the additional analysis of the trained models is interesting.\n\nSome open questions for me:\n“For SL modelling, the Wavenet Classifier contains 3 convolutional layers, whereas for group modelling it has 6, further motivated in Section 4. “\n\nI could not quite find the part where this is written maybe I missed it? In any case, would be nice to also evaluate the 6-layer classifier on subject-level decoding, as it may also extract different frequencies with its deeper structure. So these results would be an important addition, also to disentangle effects of larger model and effect of subject embedding etc.\n\nIn general, it would be good to evaluate further published EEG deep learning models/pipelines on this task to know how the reported accuracies compare to the literature. Also, I assume this dataset has been decoded before, how where accuracies in the literature?\n\nI was also missing at least one reference on transfer learning for neurophysiological recordings, \nhttps://iopscience.iop.org/article/10.1088/1741-2552/abb7a7 \nAlso this or other works would provide interesting baselines for the method presented in this manuscript.\n\n\nRegarding the PFI analysis, it would be nice to additionally perform a computationally simpler gradient-based analysis, note that both fourier transform and inverse fourier transform are differentiable and hence one can also compute gradients on fourier coefficients. It would be good to see if these analysis agree or in how far they disagree.\n", "clarity,_quality,_novelty_and_reproducibility": "The manuscript is fairly clearly written, seems novel to the best of my knowledge and code is available.", "summary_of_the_review": "Overall, this seems a nice manuscript with some additional analyses needed in my view.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666702326053}, {"id": "9DePmgbSiE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1750/Reviewer_NEjD"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present an interesting discussion and MEG classification results in SL vs. GL settings. The submission is not a good match for ICLR, and it would interest the applied neuroscience community. The technical contribution of a relatively standard/off-the-shelf WaveNet to MEG is the major weakness of the paper. Technical details on the reproduction of the model (see details in a section below) are missing, and a promise in the abstract, \"All code is available on GitHub,\" lands an anonymized link (a GitHub search for the code would result probably in the authors' identification). ", "review_text": "The manuscript is not a good match for the ICLR. Little ML novelty and extended results discussion from a neuroscientific application point of view would contribute to the SfN poster but not the ICLR. ", "strengths": "Strength: A relatively standard WaveNet application to MEG is unfortunately out of scope for ICLR.\n\nWeaknesses: Missing technical details. The authors include very vague technical descriptions of MEG preprocessing as follows:\n\"Raw data is bandpass filtered between 0.1 and 125 Hz, and line noise is removed with notch filters. After downsampling to 250 Hz, 1.024-second epochs are extracted, starting 100 ms before stimulus presentation. This resulted in 306 x 256-dimensional trials (channels x timesteps) from the 306 MEG sensors. Whitening is used to remove covariance between channels for SL models, whereas for GL models, standardization is performed per channel. We do multiclass decoding, predicting a separate probability for each of the 118 classes (images).\"\nWhat kind of filters (FIR, IIR, etc.) were used? What downsampling and whitening procedures were applied (the reviewer had no access to the code)? Why 118 classes (that would be a rather revolutionary BCI)? \n\nSimilar problems continue with the WaveNet model's vague description. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors present an interesting discussion and MEG classification results in SL vs. GL settings. The submission is not a good match for ICLR, and it would interest the applied neuroscience community. The technical contribution of a relatively standard/off-the-shelf WaveNet to MEG is the major weakness of the paper. Technical details on the reproduction of the model (see details in a section below) are missing, and a promise in the abstract, \"All code is available on GitHub,\" lands an anonymized link (a GitHub search for the code would result probably in the authors' identification). ", "strength_and_weaknesses": "Strength: A relatively standard WaveNet application to MEG is unfortunately out of scope for ICLR.\n\nWeaknesses: Missing technical details. The authors include very vague technical descriptions of MEG preprocessing as follows:\n\"Raw data is bandpass filtered between 0.1 and 125 Hz, and line noise is removed with notch filters. After downsampling to 250 Hz, 1.024-second epochs are extracted, starting 100 ms before stimulus presentation. This resulted in 306 x 256-dimensional trials (channels x timesteps) from the 306 MEG sensors. Whitening is used to remove covariance between channels for SL models, whereas for GL models, standardization is performed per channel. We do multiclass decoding, predicting a separate probability for each of the 118 classes (images).\"\nWhat kind of filters (FIR, IIR, etc.) were used? What downsampling and whitening procedures were applied (the reviewer had no access to the code)? Why 118 classes (that would be a rather revolutionary BCI)? \n\nSimilar problems continue with the WaveNet model's vague description. \n", "clarity,_quality,_novelty_and_reproducibility": "The dataset is freely available, but the GitHub code would probably be available after manuscript acceptance, which is a problem for the reviewer.\nThe current ML model and MEG preprocessing descriptions are too vague for reproducibility. The vague description makes the WaveNet application appear standard and without significant novelty. ", "summary_of_the_review": "The manuscript is not a good match for the ICLR. Little ML novelty and extended results discussion from a neuroscientific application point of view would contribute to the SfN poster but not the ICLR. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666691581616}, {"id": "gJgCoxL0u06", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1750/Reviewer_bWS9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work is concerned with improving classification of brain activity (MEG). Typically, this would be done by training a model for each subject. But this work presents a strategy for pooling data across multiple subjects (n=15) to produce a \"group-level\" model. In the naive approach, this can be done by combining data from multiple subjects into one training set. But in this work, they augment each training example with a subject embedding, which is updated during training. They find that a group-level model produces predictions which have a 6% disadvantage to the subject-level models, but a 5% advantage after fine-tuning. The paper also includes an analysis of the model's trained weights and what they reveal about visual processing in the brain. ", "review_text": "In the work's current state, I vote to reject. The approach as it stands cannot be used for transfer-learning, which could be one of the main benefits of a group-level model, if accomplished. In its current form, the proposed method doubles the training time and number of weights in order to achieve a 5% increase over the baseline subject-level model.", "strengths": "# Strengths\n- The text is well written\n- The analysis of MEG signal using trained model weights (section 4.5) shows promise for future neuroscientific study.\n\n# Weaknesses\n- The model is not suitable for use in transfer learning. The leave-one-subject-out analysis shows that the naive baseline performs just as well as the proposed model, and that better group performance does not translate to better fine-tuning performance on the held-out subject. The authors rightly note that this is a non-trivial problem, but similar transfer-learning solutions have shown good progress for NLP and vision, and it would greatly strengthen the significance of these results if something analogous could be accomplished for the MEG domain.\n- It should also be noted that the group-level architecture is ~2x larger than the subject-level architecture. The authors say that subject-level performance plateaus beyond a certain number of parameters (3 layers), but it would be nice to see this in a table.  Alternatively, the authors could show results for a group-level architecture that is the same size (3 layers) as the subject-level model. As it stands, there remains room to believe that the presented gains over the baseline could be entirely explained by simply using a larger model.\n- See below for comments on significance and novelty ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work is concerned with improving classification of brain activity (MEG). Typically, this would be done by training a model for each subject. But this work presents a strategy for pooling data across multiple subjects (n=15) to produce a \"group-level\" model. In the naive approach, this can be done by combining data from multiple subjects into one training set. But in this work, they augment each training example with a subject embedding, which is updated during training. They find that a group-level model produces predictions which have a 6% disadvantage to the subject-level models, but a 5% advantage after fine-tuning. The paper also includes an analysis of the model's trained weights and what they reveal about visual processing in the brain. ", "strength_and_weaknesses": "# Strengths\n- The text is well written\n- The analysis of MEG signal using trained model weights (section 4.5) shows promise for future neuroscientific study.\n\n# Weaknesses\n- The model is not suitable for use in transfer learning. The leave-one-subject-out analysis shows that the naive baseline performs just as well as the proposed model, and that better group performance does not translate to better fine-tuning performance on the held-out subject. The authors rightly note that this is a non-trivial problem, but similar transfer-learning solutions have shown good progress for NLP and vision, and it would greatly strengthen the significance of these results if something analogous could be accomplished for the MEG domain.\n- It should also be noted that the group-level architecture is ~2x larger than the subject-level architecture. The authors say that subject-level performance plateaus beyond a certain number of parameters (3 layers), but it would be nice to see this in a table.  Alternatively, the authors could show results for a group-level architecture that is the same size (3 layers) as the subject-level model. As it stands, there remains room to believe that the presented gains over the baseline could be entirely explained by simply using a larger model.\n- See below for comments on significance and novelty ", "clarity,_quality,_novelty_and_reproducibility": "- I'm worried about the significance of this work. The presented group-level model performs 6% worse than the subject-level baseline. Fine-tuning on the subject data results in a 5% performance increase, but this greatly narrows the real world applicability of this approach. In the very best case, to see any benefit from this method, one has to pool all their data, train a group level model, and then fine-tune these weights for each subject, essentially doubling the training time and size of the architecture. And even then, the benefit is only a 5% increase.\n- And this benefit is not uniform across subjects, indeed for some subjects, the performance slightly decreases (first paragraph of pg. 6).\n- The novelty of the approach also concerns me. Compared to the naive group-level approach, the presented approach is essentially identical, but with an extra piece of input data, namely the identity of the subject. Simplicity in itself is not necessarily a bad thing. But I would argue that this paper does not present any new approach, but rather a simple comparison between models: one which has access to a particular bit of information, and one that does not. ", "summary_of_the_review": "In the work's current state, I vote to reject. The approach as it stands cannot be used for transfer-learning, which could be one of the main benefits of a group-level model, if accomplished. In its current form, the proposed method doubles the training time and number of weights in order to achieve a 5% increase over the baseline subject-level model.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666585376964}, {"id": "L-OlJDZYuN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1750/Reviewer_kPfZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed a deep learning-based group-level (GL) neural decoding model that can be adapted to different subjects. The key module of the GL model is the learnable subject embeddings. Experiments were conducted on an MEG neural dataset. Experimental results indicated that the proposed GL model can achieve significantly higher classification accuracy than the naive GL modelling. The paper also gained insight into how subject embeddings help the group model and analyzed the spatial and temporal feature importance.\nThe main contributions of this paper are as follows.\n1.The paper builds a GL neural decoding model that can be adapted to different subjects.\n2.The paper gains insights into the learned subject embedding.\n3.The paper gives analysis of model weights to reveal how meaningful spatio-temporal and spectral information is encoded.", "review_text": "More than an innovative work on representation learning, the work applies an existing method to a new task. I think the quality of the paper is not up to the average quality of the papers accepted by ICLR. Also, the paper is not quite in line with the main scope of ICLR. The article also does a rich analysis from the neuroscience perspective. I suggest the authors to submit their paper to conferences or journals in the field of computational neuroscience after major revisions.", "strengths": "Strength:\n1.This paper builds a GL neural decoding model that can be adapted to different subjects. The key module is the learnable subject embedding, which is simple to apply. In this way, researchers can aggregate multi-subject neural signals to train a GL model and finetune it on the specific target subject. Considering the sample size of neural datasets is always limited, the GL model is meaningful to improve the model performance. Although the model is only validated on an MEG dataset in this paper, I think it can also be applicable to other modalities, such as EEG and fMRI.\n\nWeaknesses:\n1.The methodology novelty is very limited. The feature extraction network is similar to the WaveNet. The subject embedding trick and the analyses on the relationship between neural responses and features with PFI have been applied to MEG neural encoding in (Chehab et al. 2021). \n\n2.The paper is a work in progress. The model is only validated on an MEG dataset. It is recommended that the authors perform validation on at least two neural dataset. The results and analyses are not impressive.\n\n3.The LOSO experimental results show that the advantage of the subject embeddings cannot transfer to new subjects. How to leverage transfer learning with the proposed model and learn a useful embedding for the new subject in an unsupervised manner as described in Page 7? The authors only show a concept and don’t conduct any experiment.\n\n4.Many descriptions on the experimental results rather than concrete data in the form of figures and tables. For example, the results of finetuning a naive group model indicated in Page 6 have not been shown in Figure 3. In Section 4.2, the impact of the network layer on accuracy is not shown in corresponding charts. The authors said that the visualization of subject embeddings did not show any clusters and no visualization is shown in the paper or the appendix.\n\n5.Many details are missing. For example, why did the authors change the embedding dimensionality from 10 only to 3 and 14? How about other settings?\n\n6.The improvement of GL model compared with the naïve GL model is significant. However, the naïve group models are too weak to be supportive baseline models. A network module is also proposed in the literature [1] to address the between-subject variability. I suggest that the authors should take a boarder review and compare their model with stronger baselines.\n\n[1] Défossez, A., Caucheteux, C., Rapin, J., Kabeli, O., & King, J. R. (2022). Decoding speech from non-invasive brain recordings. arXiv preprint arXiv:2208.12266.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed a deep learning-based group-level (GL) neural decoding model that can be adapted to different subjects. The key module of the GL model is the learnable subject embeddings. Experiments were conducted on an MEG neural dataset. Experimental results indicated that the proposed GL model can achieve significantly higher classification accuracy than the naive GL modelling. The paper also gained insight into how subject embeddings help the group model and analyzed the spatial and temporal feature importance.\nThe main contributions of this paper are as follows.\n1.The paper builds a GL neural decoding model that can be adapted to different subjects.\n2.The paper gains insights into the learned subject embedding.\n3.The paper gives analysis of model weights to reveal how meaningful spatio-temporal and spectral information is encoded.", "strength_and_weaknesses": "Strength:\n1.This paper builds a GL neural decoding model that can be adapted to different subjects. The key module is the learnable subject embedding, which is simple to apply. In this way, researchers can aggregate multi-subject neural signals to train a GL model and finetune it on the specific target subject. Considering the sample size of neural datasets is always limited, the GL model is meaningful to improve the model performance. Although the model is only validated on an MEG dataset in this paper, I think it can also be applicable to other modalities, such as EEG and fMRI.\n\nWeaknesses:\n1.The methodology novelty is very limited. The feature extraction network is similar to the WaveNet. The subject embedding trick and the analyses on the relationship between neural responses and features with PFI have been applied to MEG neural encoding in (Chehab et al. 2021). \n\n2.The paper is a work in progress. The model is only validated on an MEG dataset. It is recommended that the authors perform validation on at least two neural dataset. The results and analyses are not impressive.\n\n3.The LOSO experimental results show that the advantage of the subject embeddings cannot transfer to new subjects. How to leverage transfer learning with the proposed model and learn a useful embedding for the new subject in an unsupervised manner as described in Page 7? The authors only show a concept and don’t conduct any experiment.\n\n4.Many descriptions on the experimental results rather than concrete data in the form of figures and tables. For example, the results of finetuning a naive group model indicated in Page 6 have not been shown in Figure 3. In Section 4.2, the impact of the network layer on accuracy is not shown in corresponding charts. The authors said that the visualization of subject embeddings did not show any clusters and no visualization is shown in the paper or the appendix.\n\n5.Many details are missing. For example, why did the authors change the embedding dimensionality from 10 only to 3 and 14? How about other settings?\n\n6.The improvement of GL model compared with the naïve GL model is significant. However, the naïve group models are too weak to be supportive baseline models. A network module is also proposed in the literature [1] to address the between-subject variability. I suggest that the authors should take a boarder review and compare their model with stronger baselines.\n\n[1] Défossez, A., Caucheteux, C., Rapin, J., Kabeli, O., & King, J. R. (2022). Decoding speech from non-invasive brain recordings. arXiv preprint arXiv:2208.12266.", "clarity,_quality,_novelty_and_reproducibility": "Quality: Not very good. The novelty is very limited. Although the experimental results show that the proposed method can significantly improve the performance, the authors only validate their method on one MEG dataset and compare it with weak baselines.\nClarity: The method description is clear. But some detailed descriptions on the experimental results are not very clear.\nOriginality of the work: not very original.\nReproducibility: The code is publicly available, and the datased is also publicly available. The reproducibility is high.", "summary_of_the_review": "More than an innovative work on representation learning, the work applies an existing method to a new task. I think the quality of the paper is not up to the average quality of the papers accepted by ICLR. Also, the paper is not quite in line with the main scope of ICLR. The article also does a rich analysis from the neuroscience perspective. I suggest the authors to submit their paper to conferences or journals in the field of computational neuroscience after major revisions.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethics concerns.", "recommendation": "3: reject, not good enough"}, "tcdate": 1666511730369}], "openreview_url": "https://openreview.net/forum?id=lEFM4OTz62c", "arxiv_id": "2205.14102", "paper_pdf": "papers/lEFM4OTz62c.pdf", "paper_pdf_sha256": "52f5b635365f22f821fcb5f343b0a412d581002e15fce5d7a37d0a6b64c2379e", "paper_pdf_bytes": 18413707, "paper_pdf_source": "openreview", "code_url": "https://github.com/ricsinaruto/MEG-group-decode", "code_repository": "ricsinaruto/MEG-group-decode", "code_commit": "3a4e7daae885b387fd4e19903e5d295eae9c0fea", "code_archive": "repos/lEFM4OTz62c.zip", "code_archive_sha256": "e21b28d53b527b5953c0d22ae747ee5db10937ba5b396a3589aeaa26a70f0fda", "code_archive_bytes": 219323, "code_file_count": 27, "code_extensions": {".py": 24, ".ipynb": 3}, "github_disk_usage_kb": 228, "github_languages": {"Jupyter Notebook": 255845, "Python": 92038}, "github_archived": false, "github_pushed_at": "2024-01-18T17:11:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalizing-brain-decoding-across-subjects"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UgNQM-LcVpN", "year": 2022, "status": "rejected", "title": "A Modulation Layer to Increase Neural Network Robustness Against Data Quality Issues", "authors": ["Mohamed Abdelhack", "Jiaming Zhang", "Sandhya Tripathi", "Bradley A Fritz", "Michael Avidan", "Yixin Chen", "Christopher Ryan King"], "authorids": ["~Mohamed_Abdelhack1", "~Jiaming_Zhang2", "~Sandhya_Tripathi1", "~Bradley_A_Fritz1", "~Michael_Avidan1", "~Yixin_Chen1", "~Christopher_Ryan_King1"], "authors_source": "OpenReview API", "abstract": "Data quality is a common problem in machine learning, especially in high-stakes settings such as healthcare. Missing data affects accuracy, calibration, and feature attribution in complex patterns. Developers often train models on carefully curated datasets to minimize missing data bias; however, this reduces the usability of such models in production environments, such as real-time healthcare records. Making machine learning models robust to missing data is therefore crucial for practical application. While some classifiers naturally handle missing data, others, such as deep neural networks, are not designed for unknown values. We propose a novel neural network modification to mitigate the impacts of missing data. The approach is inspired by neuromodulation that is performed by biological neural networks. Our proposal replaces the fixed weights of a fully-connected layer with a function of an additional input (reliability score) at each input, mimicking the ability of cortex to up- and down-weight inputs based on the presence  of other data. The modulation function is jointly learned with the main task using a multi-layer perceptron. We tested our modulating fully connected layer on multiple classification, regression, and imputation problems, and it either improved performance or generated comparable performance to conventional neural network architectures concatenating reliability to the inputs. Models with modulating layers were more robust against degradation of data quality by introducing additional missingness at evaluation time. These results suggest that explicitly accounting for reduced information quality with a modulating fully connected layer can enable the deployment of artificial intelligence systems in real-time settings.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "84SYmJ-GBeM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3919/Reviewer_fZmE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper is proposing a neural net architecture which makes Weights (in a layer) a function of an external signal trying to quantify the quality of input. The presented architecture is motivated by, and applied to the problem of missing/noisy data. Empirical evaluation performed on different types of tasks, over different datasets and in different missing-ness scenarios showed improved robustness in terms of predictive performance. ", "review_text": "Empirical evaluation is performed fairly extensively, 3 different kinds of tasks: regression, classification and imputation. Four different - real world datasets from healthcare domain were used, which had natural (label imbalance) as well as simulated data quality issues (noise superposition and masking). The results suggest indeed that introduced \"modulation\" helped approaches not loose as much performance due to increased degradation of data quality.\n\nIn the Related work section, it is stated that \"Our approach is superficially similar to attention mechanisms\", however not much more effort is made to make a case for that claim. Suggestion that a particular instance of attention mechanism is \"difficult to scale for long time-varying inputs\" is not a sufficient reason to stop further comparison. Moreover, the scalability argument is brought up, without being accompanied with the analysis which would suggest that proposed \"modulation\" method is more scalable. The comparison against attention methods, both in terms of computational efficiency/scalability - as well as predictive performance, would have been an a valuable data point.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper is proposing a neural net architecture which makes Weights (in a layer) a function of an external signal trying to quantify the quality of input. The presented architecture is motivated by, and applied to the problem of missing/noisy data. Empirical evaluation performed on different types of tasks, over different datasets and in different missing-ness scenarios showed improved robustness in terms of predictive performance. ", "main_review": "Empirical evaluation is performed fairly extensively, 3 different kinds of tasks: regression, classification and imputation. Four different - real world datasets from healthcare domain were used, which had natural (label imbalance) as well as simulated data quality issues (noise superposition and masking). The results suggest indeed that introduced \"modulation\" helped approaches not loose as much performance due to increased degradation of data quality.\n\nIn the Related work section, it is stated that \"Our approach is superficially similar to attention mechanisms\", however not much more effort is made to make a case for that claim. Suggestion that a particular instance of attention mechanism is \"difficult to scale for long time-varying inputs\" is not a sufficient reason to stop further comparison. Moreover, the scalability argument is brought up, without being accompanied with the analysis which would suggest that proposed \"modulation\" method is more scalable. The comparison against attention methods, both in terms of computational efficiency/scalability - as well as predictive performance, would have been an a valuable data point.", "summary_of_the_review": "Even though empirical study is suggesting potential benefits of the approach in applications with a pronounced data quality issues, the article haven't assured me that proposed modulation layer is more applicable/useful (or sufficiently distinct) from the attention layer. Studies exploring attention for handling missing data have already been conducted (eg. Wu, Richard, et al. \"Attention-based learning for missing data imputation in Holoclean.\" Proceedings of Machine Learning and Systems 2 (2020): 307-325.), and comparison agains them (as a closely related approach) would be appropriate.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635887283658}, {"id": "qNlaQi193wV", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3919/Reviewer_4EPR"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This submission contributes an approach to handle missing values in Neural networks by modulating inside the architecture the missing values by factors which decrease  the role of the feature in the architecture. The approach is benchmarked empirically, but does not appear to outperform consistently other approaches. \n\n", "review_text": "This submission is framed as for data quality issues in general, but it only tackles missing values. \n\nIt is based on an intuition, that modulating the input features based on a function of their reliability, can be helpful. This intuition is interesting. For missing values, NeuMiss networks have showed that similar architectures, with specific choices of modulation and inductive bias, could target the Bayes predictor. However, here, it is given with no analysis and little details about which specific inductive bias to use. \n\nThe baseline methods lack strong predictors that readily fit missing values, such as Neumiss or trees with MIA (missing incorporated attribute).\n\nThe empirical results are not conclusive. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This submission contributes an approach to handle missing values in Neural networks by modulating inside the architecture the missing values by factors which decrease  the role of the feature in the architecture. The approach is benchmarked empirically, but does not appear to outperform consistently other approaches. \n\n", "main_review": "This submission is framed as for data quality issues in general, but it only tackles missing values. \n\nIt is based on an intuition, that modulating the input features based on a function of their reliability, can be helpful. This intuition is interesting. For missing values, NeuMiss networks have showed that similar architectures, with specific choices of modulation and inductive bias, could target the Bayes predictor. However, here, it is given with no analysis and little details about which specific inductive bias to use. \n\nThe baseline methods lack strong predictors that readily fit missing values, such as Neumiss or trees with MIA (missing incorporated attribute).\n\nThe empirical results are not conclusive. ", "summary_of_the_review": "The contribution is based on intuitions that do not seem very solid and should be better studied. It does not really perform better than other approaches. ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635803980402}, {"id": "6Mz_0LqZeBy", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3919/Reviewer_P1JV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a fully connected neural layer to add to a DNN to deal with missing or noisy data. Weights in the additional neural layer are modulated by quality indicators of the inputs. The neural layer is evaluated in classification, regression and feature imputation settings for  several datasets, one health related one unavailable for replication and the well known Wisconsin Great Cancer and Boston Housing datasets. The approach is evaluated against several state of the art baselines and varying level of data quality. The approach does not do significantly better than existing state of the art methods.", "review_text": "Strengths\n\n1. The paper is well written and easy to read.\n2. The approach is well motivated and a good idea.\n3. Experiments are thorough and generally well described on a variety of problem types and datasets. I liked the evaluation of confidence intervals.\n\nWeaknesses\n\n1. Unfortunately the method does not significantly improve on the state of the art. In most cases VAEAC does as well or better than the proposed approaches. A simpler version of the proposed approach FC+Mod generally does as well as the more complex MFCL and MFCL+. It was unclear which results showed statistical significance. It seems the networks respond to a fairly simple measure of data quality (FC+Mod), so it is not clear that the more complex approaches are warranted.\n2. Many small typographical errors, mainly spelling mistakes. Sometimes FC+Mod is stated as DNN+Mod. Are they the same?\n3. Error bars not given on figs 2-4.\n4. I am unsure that there is enough information for the results to be reproducible. In particular I was unsure about the network topologies given in the appendix (p13). e.g., for ACTFAST there were 2 hidden layers, with an architecture of 8 neurons in the first layer and 4 in the second, but then I didn't understand how the 3 hidden layers in the next line fitted.\n5. It would be good to see a comparison against the approach mostly used of adding additional input attributes for data quality.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a fully connected neural layer to add to a DNN to deal with missing or noisy data. Weights in the additional neural layer are modulated by quality indicators of the inputs. The neural layer is evaluated in classification, regression and feature imputation settings for  several datasets, one health related one unavailable for replication and the well known Wisconsin Great Cancer and Boston Housing datasets. The approach is evaluated against several state of the art baselines and varying level of data quality. The approach does not do significantly better than existing state of the art methods.", "main_review": "Strengths\n\n1. The paper is well written and easy to read.\n2. The approach is well motivated and a good idea.\n3. Experiments are thorough and generally well described on a variety of problem types and datasets. I liked the evaluation of confidence intervals.\n\nWeaknesses\n\n1. Unfortunately the method does not significantly improve on the state of the art. In most cases VAEAC does as well or better than the proposed approaches. A simpler version of the proposed approach FC+Mod generally does as well as the more complex MFCL and MFCL+. It was unclear which results showed statistical significance. It seems the networks respond to a fairly simple measure of data quality (FC+Mod), so it is not clear that the more complex approaches are warranted.\n2. Many small typographical errors, mainly spelling mistakes. Sometimes FC+Mod is stated as DNN+Mod. Are they the same?\n3. Error bars not given on figs 2-4.\n4. I am unsure that there is enough information for the results to be reproducible. In particular I was unsure about the network topologies given in the appendix (p13). e.g., for ACTFAST there were 2 hidden layers, with an architecture of 8 neurons in the first layer and 4 in the second, but then I didn't understand how the 3 hidden layers in the next line fitted.\n5. It would be good to see a comparison against the approach mostly used of adding additional input attributes for data quality.\n", "summary_of_the_review": "The paper has the good idea of adding a layer modulated on the data quality as a plug and play layer. It is evaluated in a variety of problem types, data sets and experimental settings against state of the art. However, the approach does not do much better that existing state of the art.\n    ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635568021297}, {"id": "VU70Z6SWPZo", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3919/Reviewer_g2de"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Inspired by neuromodulation in biological neural networks, this manuscript proposes a modification to fully connected layers in artificial neural networks to accommodate the missing values (or low-quality measures) without the need for suboptimal and unrealistic data imputation. The authors claim that this modification can be useful in applications of AI in real-time settings, but this claim remains unproven (theoretically and empirically) in this paper. The authors benchmarked the proposed methods against a few state-of-art data imputation approaches on two classification datasets and one regression dataset and with random and non-random missing data mechanisms. ", "review_text": "*Strengths:\n\n- The text is clear and the motivation is nice. The manuscript provides a very nice introduction to the topic highlighting the importance of the topic.\n\n*Weaknesses: \n\n- The proposed neuromodulation method has three (two theoretical and one practical) main shortcomings: i) it only deals with weights and ignores the biases as a part of network parameters; ii) in its best performance, it is expected to learn masking out the missing values from the inputs that is equivalent to a simple zero-imputation, while adding unnecessary complexity (thus higher instability) into the model architecture; iii) it needs to know the pattern of unreliable inputs at the run time. While finding this pattern is not difficult when the values are missing, it could be a real challenge to find them when the values are of low quality because, for example, the standard deviation of the noise is unknown in many applications (we need an extra module to detect the low-quality inputs in real-time).  \n\n- The experimental design can be improved and the results are not decisive and fail in communicating the benefits of the proposed method: The main focus of experiments is on comparing the classification and regression performance between MFCL and its alternatives on a few datasets (only two for classification and one for regression) and the final results show minor or no improvement in many cases. This is while the experimental setup can be improved by a more diverse set of experiments. For example, i) by investigating the patterns learned in $g(m)$; are they simply mask out the missing values or learn something extra? ii) by including more common and simpler imputation approaches such as zero, mean, and KNN imputers. Especially it is very important to show that MFCL is not a simple zero imputer. \n\n*Minor suggestions, comments, and questions:\n\n- The short method section is a little bit inaccurate. For example: \n     - \"A fully connected layer has a transfer function of\": Othe layer types also have transfer function. Furthermore, the formula does not represent a transfer function but the output of a neuron.\n     - \"and $f$ the non-linearity function\": $f$ can be also a linear function. Also 'Non-linearity' should be 'non-linear'.\n     - \"and fixed at inference\": in some models such as Bayesian neural networks, they are not fixed during the inference.\n     - the modulation signal $m$ should be in more detail explained in the method section. For example, how to compute it.\n     - The definition of $W_0$ is not clear. What is the starting network? Is it the raw network right after initialization?\n     - Figure 1 is inaccurate, the modulation network has only two outputs while the number of its outputs should be equal to the number of weights in each layer.\n\n- The proposed method only modulates the weight of the network, thus it still needs to impute the missing values. Throughout the text, it is not clear how missing values are dealt with when MFCL is used. Furthermore, if any imputation is used in advance, then it is difficult to say whether the change in the performance is due to the data imputation of MFCL itself. \n\n- While in the caption, the error bars are missing in Figure 2,3,4.\n\n- The performance of MFCL+ is among the worst in many cases. What is the explanation behind this?\n\n- Are the learned modulation patterns stable across several runs? \n\n- Why other imputation approaches are not benchmarked in the regression task?\n\n- Section 4.2 says the MFCL has a significantly lower loss, but no significance test is performed.\n\n- Sections 4.3 and 4.4 present no quantitative results and miss the proper reference to supplementary materials to guide the reader.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Inspired by neuromodulation in biological neural networks, this manuscript proposes a modification to fully connected layers in artificial neural networks to accommodate the missing values (or low-quality measures) without the need for suboptimal and unrealistic data imputation. The authors claim that this modification can be useful in applications of AI in real-time settings, but this claim remains unproven (theoretically and empirically) in this paper. The authors benchmarked the proposed methods against a few state-of-art data imputation approaches on two classification datasets and one regression dataset and with random and non-random missing data mechanisms. ", "main_review": "*Strengths:\n\n- The text is clear and the motivation is nice. The manuscript provides a very nice introduction to the topic highlighting the importance of the topic.\n\n*Weaknesses: \n\n- The proposed neuromodulation method has three (two theoretical and one practical) main shortcomings: i) it only deals with weights and ignores the biases as a part of network parameters; ii) in its best performance, it is expected to learn masking out the missing values from the inputs that is equivalent to a simple zero-imputation, while adding unnecessary complexity (thus higher instability) into the model architecture; iii) it needs to know the pattern of unreliable inputs at the run time. While finding this pattern is not difficult when the values are missing, it could be a real challenge to find them when the values are of low quality because, for example, the standard deviation of the noise is unknown in many applications (we need an extra module to detect the low-quality inputs in real-time).  \n\n- The experimental design can be improved and the results are not decisive and fail in communicating the benefits of the proposed method: The main focus of experiments is on comparing the classification and regression performance between MFCL and its alternatives on a few datasets (only two for classification and one for regression) and the final results show minor or no improvement in many cases. This is while the experimental setup can be improved by a more diverse set of experiments. For example, i) by investigating the patterns learned in $g(m)$; are they simply mask out the missing values or learn something extra? ii) by including more common and simpler imputation approaches such as zero, mean, and KNN imputers. Especially it is very important to show that MFCL is not a simple zero imputer. \n\n*Minor suggestions, comments, and questions:\n\n- The short method section is a little bit inaccurate. For example: \n     - \"A fully connected layer has a transfer function of\": Othe layer types also have transfer function. Furthermore, the formula does not represent a transfer function but the output of a neuron.\n     - \"and $f$ the non-linearity function\": $f$ can be also a linear function. Also 'Non-linearity' should be 'non-linear'.\n     - \"and fixed at inference\": in some models such as Bayesian neural networks, they are not fixed during the inference.\n     - the modulation signal $m$ should be in more detail explained in the method section. For example, how to compute it.\n     - The definition of $W_0$ is not clear. What is the starting network? Is it the raw network right after initialization?\n     - Figure 1 is inaccurate, the modulation network has only two outputs while the number of its outputs should be equal to the number of weights in each layer.\n\n- The proposed method only modulates the weight of the network, thus it still needs to impute the missing values. Throughout the text, it is not clear how missing values are dealt with when MFCL is used. Furthermore, if any imputation is used in advance, then it is difficult to say whether the change in the performance is due to the data imputation of MFCL itself. \n\n- While in the caption, the error bars are missing in Figure 2,3,4.\n\n- The performance of MFCL+ is among the worst in many cases. What is the explanation behind this?\n\n- Are the learned modulation patterns stable across several runs? \n\n- Why other imputation approaches are not benchmarked in the regression task?\n\n- Section 4.2 says the MFCL has a significantly lower loss, but no significance test is performed.\n\n- Sections 4.3 and 4.4 present no quantitative results and miss the proper reference to supplementary materials to guide the reader.\n", "summary_of_the_review": "Despite good motivation, the proposed method seems to have several theoretical and practical limitations and the current experimental results do not provide enough evidence on what are the benefits of the proposed method.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635154488105}], "openreview_url": "https://openreview.net/forum?id=UgNQM-LcVpN", "arxiv_id": "2107.08574", "paper_pdf": "papers/UgNQM-LcVpN.pdf", "paper_pdf_sha256": "5d1a5b41b462af0110940f2f46b929f4fb3c12a5661d83707fc097259f2b4391", "paper_pdf_bytes": 455793, "paper_pdf_source": "openreview", "code_url": "https://github.com/mabdelhack/mfcl", "code_repository": "mabdelhack/mfcl", "code_commit": "7be8449b06d41251e2cb549fbf6ffc1338dc37d0", "code_archive": "repos/UgNQM-LcVpN.zip", "code_archive_sha256": "c58c64e31bb07b472e18594f2ff9dacc335f91d5fd6c78563610c9298f2fd480", "code_archive_bytes": 506029, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 547, "github_languages": {"Python": 139130}, "github_archived": false, "github_pushed_at": "2023-04-04T08:32:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-modulation-layer-to-increase-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yvuk0RsLoP7", "year": 2021, "status": "rejected", "title": "Improving Model Robustness with Latent Distribution Locally and Globally", "authors": ["Zhuang QIAN", "Shufei Zhang", "Kaizhu Huang", "Qiufeng Wang", "Rui Zhang", "Xinping Yi"], "authorids": ["~Zhuang_QIAN1", "~Shufei_Zhang1", "~Kaizhu_Huang1", "~Qiufeng_Wang2", "~Rui_Zhang10", "xinping.yi@liverpool.ac.uk"], "authors_source": "OpenReview API", "abstract": "We propose a novel adversarial training method which leverages both the local and global information to defend  adversarial attacks. Existing adversarial training methods usually generate adversarial perturbations  locally in a supervised manner and fail to consider the data manifold information in a global way. Consequently, the resulting  adversarial examples may corrupt the underlying data structure and are typically biased towards the decision boundary. In this work, we exploit both the local and global information of data manifold to generate adversarial examples in an unsupervised manner. Specifically, we design our novel framework  via an adversarial game between  a discriminator and a classifier: the discriminator is  learned to differentiate the latent distributions  of the natural data and the perturbed counterpart, while the classifier is trained to recognize accurately the perturbed examples as well as enforcing the invariance between the two latent distributions. We conduct a series of analysis on the model robustness and also verify the effectiveness of our proposed method empirically. Experimental results show that our method  substantially outperforms the recent state-of-the-art (i.e. Feature Scattering) in defending adversarial attacks  by a large accuracy margin  (e.g. $17.0\\%$ and $18.1\\%$ on SVHN dataset, $9.3\\%$ and $17.4\\%$ on CIFAR-10 dataset, $6.0\\%$ and $16.2\\%$ on CIFAR-100 dataset for defending PGD20 and CW20 attacks respectively).", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "EIlDln-C5Ne", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3224/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper analyzes the property of local and global data manifold for adversarial training. In particular, they used a discriminator-classifier model, where the discriminator tries to differentiate between the natural and adversarial space, and the classifier aims to classify between them while maintaining the constraints between local and global distributions. The authors implemented the proposed method on several datasets and achieved good performance. They also compared with several whitebox and blackbox methods and proved superiority. \n\nThis paper was, in general, well written. The authors provided a good visualization of their analysis. Using local and global information for adversarial training is intuitive. The authors provided a good theoretical background to establish their method. The empirical evaluations show promising results. \n\nSome major concerns are listed as follows:\n1. It is not clear how equations 4 and 5 are realized using discriminator and classifier. \n2. What kind of perturbations are chosen? It looks like all the experiments are with L-infinity. Does this observation hold for other ones?\n3.  If the attackers leverage the global and local data manifold, can they bypass this attack? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper analyzes the property of local and global data manifold for adversarial training. ", "review": "The paper analyzes the property of local and global data manifold for adversarial training. In particular, they used a discriminator-classifier model, where the discriminator tries to differentiate between the natural and adversarial space, and the classifier aims to classify between them while maintaining the constraints between local and global distributions. The authors implemented the proposed method on several datasets and achieved good performance. They also compared with several whitebox and blackbox methods and proved superiority. \n\nThis paper was, in general, well written. The authors provided a good visualization of their analysis. Using local and global information for adversarial training is intuitive. The authors provided a good theoretical background to establish their method. The empirical evaluations show promising results. \n\nSome major concerns are listed as follows:\n1. It is not clear how equations 4 and 5 are realized using discriminator and classifier. \n2. What kind of perturbations are chosen? It looks like all the experiments are with L-infinity. Does this observation hold for other ones?\n3.  If the attackers leverage the global and local data manifold, can they bypass this attack? ", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604123461979}, {"id": "pEPBJlVSWN5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3224/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper considers the local and global information in adversarial attacks for adversarial training, where the authors design an adversarial framework containing a discriminator and a classifier. The idea is interesting and the paper is easy to follow. \n\nHowever, I have still some concerns below: \n- The novelty of this work combines the idea of PGD (local information) and Feature-Scatter (global information) .\n- More importantly, the evaluation is no enough, even though Feature-Scatter considers the global information, but many attack methods have shown the robustness of Feature-Scatter was overestimated, such as [1][2][3] and so on. So I think evaluating on PGD and CW  is not enough.\n- There are few analysis experiments for the proposed method, more analysis experiments are needed besides the comparision.\n\n[1] Feature Attack: https://openreview.net/forum?id=Syejj0NYvr&noteId=rkeBhuBMjS\n\n[2] RayS: A Ray Searching Method for Hard-label Adversarial Attack. KDD 2020.\n\n[3] Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. ICML 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review # 4", "review": "Summary: This paper considers the local and global information in adversarial attacks for adversarial training, where the authors design an adversarial framework containing a discriminator and a classifier. The idea is interesting and the paper is easy to follow. \n\nHowever, I have still some concerns below: \n- The novelty of this work combines the idea of PGD (local information) and Feature-Scatter (global information) .\n- More importantly, the evaluation is no enough, even though Feature-Scatter considers the global information, but many attack methods have shown the robustness of Feature-Scatter was overestimated, such as [1][2][3] and so on. So I think evaluating on PGD and CW  is not enough.\n- There are few analysis experiments for the proposed method, more analysis experiments are needed besides the comparision.\n\n[1] Feature Attack: https://openreview.net/forum?id=Syejj0NYvr&noteId=rkeBhuBMjS\n\n[2] RayS: A Ray Searching Method for Hard-label Adversarial Attack. KDD 2020.\n\n[3] Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. ICML 2020.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604058060996}, {"id": "cIp99ASc2ES", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3224/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a new method of improving model robustness by generating adversarial samples that are regularized by their latent distribution through f-divergence, whereas existing literature only uses local manifold property such as smoothness. \n\nThe method is well-motivated and the clarity of the paper is good. The experimental results are compared with several competitive baselines and the improvement looks significant (Although I am not familiar with the state-of-the-art experimental results). \n\nProofread is needed for the sentence \"The adversarial examples are crafted by ... \" on page 2 and several other small typos. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Use global latent distribution to improve model robustness", "review": "The paper proposes a new method of improving model robustness by generating adversarial samples that are regularized by their latent distribution through f-divergence, whereas existing literature only uses local manifold property such as smoothness. \n\nThe method is well-motivated and the clarity of the paper is good. The experimental results are compared with several competitive baselines and the improvement looks significant (Although I am not familiar with the state-of-the-art experimental results). \n\nProofread is needed for the sentence \"The adversarial examples are crafted by ... \" on page 2 and several other small typos. ", "rating": "7: Good paper, accept", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1604014226077}, {"id": "vKNSD7gK_U", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3224/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a framework for adversarial robustness via incorporating local and global structure of the data manifold. Specifically, the key motivation is that standard adversarial methods typically use only sample specific perturbations for generating the adversarial examples, and thus using them for robustness of the learning model is limited. Instead the paper proposes to capture the global data manifold as well in the robustifying framework. To this end, an objective is presented (4,5) that uses latent data distributions, with the goal that the adversarial perturbations should maximize the f-divergence against the latent distribution of the clean samples. Experiments are provided on several datasets and demonstrate significant performance improvements.\n\nPros:\n1. The key idea of using the global data manifold into the robustifying framework is quite interesting.\n2. Experiments demonstrate good empirical benefits of the approach.\n\nCons: \n1. While, the paper seemed well organized in the beginning, I got totally lost with Eq. (4-5). As I see, this objective is inaccurate and needs significant refinement. Specifically, it is unclear how is x^{adv} is related to x, and how is x^{adv} related to P*_theta? The paper tries to explain this objective in the paragraph below, but the explanation is very confusing as well.  A few other things that could help here:\na) It is said that \"Q_theta and P_\\theta* are the latent distributions induced by the natural example x\". How can a single data point induce a distribution? Do you assume the feature map from a hidden layer of a network represents a distribution? If so, in what sense? \nb) \"The adversarial example is crafted to induce the worst case distribution P*\". How is it crafted and what is the relation between P* and x? This is the key connection that is missing from (4-5).\n\n2. Moving along, Section 4.1 is organized very poorly as well. I believe too many concepts are tied together into one formulation in (6), making it hard to decipher. For example, why to include the classifier D^{1:C} within this formulation? Why not talk about it elsewhere and focus on the meat of the objective, systematically? \n\n3. Further, as I understand, x^{adv} is the first step that happens in (6), however, there is no \"adversary\" in this case, instead is finding a perturbed sample x' that maximizes the f-divergence. In what sense is x^{adv} then an adversarial sample? Perhaps the paper should re-define what is the definition of an adversarial example that it is using, to clearly state what the idea is. Technically, there is no requirement that the point x^{adv} found by this step will promote any data misclassification; however can be any point that is within a B(x,\\eps) ball from x, and that happens to maximize this divergence loss. Note that none of the other components D_W, f_theta, etc. are well trained in doing this optimization. So they could also be sub-optimal (in the sense of what the paper argues in the beginning of Page 4).\n\n4. Why is the middle formula in (6) minimizing over W to have both x and x^adv matched with the same label? Again, where is the adversary here? Or for that matter, how will the proposed approach achieve adversarial robustness ? \n\nMinor comments:\na. What is \\tau and T in (3)? \nb. How is f_\\theta defined in (6)? \nc. The paper writes that back and forth that there is no use of label information in the setup, however has labels used in discriminator in (6). This is very confusing. \nd. There is also reference to data manifold and manifold label in Figure 2, but these are not clearly explained. What precisely is the data manifold? Is it the latent distribution for a specific label? \ne. Page 4, top para: \"without considering the inter-relationship between data samples\". Won't this relation be captured implicitly through the neural network parameters theta when perturbations on all the samples are used in the training process?\n\nOverall, I think this paper needs a thorough revision to explain well its technical contributions. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Needs better technical exposition", "review": "This paper presents a framework for adversarial robustness via incorporating local and global structure of the data manifold. Specifically, the key motivation is that standard adversarial methods typically use only sample specific perturbations for generating the adversarial examples, and thus using them for robustness of the learning model is limited. Instead the paper proposes to capture the global data manifold as well in the robustifying framework. To this end, an objective is presented (4,5) that uses latent data distributions, with the goal that the adversarial perturbations should maximize the f-divergence against the latent distribution of the clean samples. Experiments are provided on several datasets and demonstrate significant performance improvements.\n\nPros:\n1. The key idea of using the global data manifold into the robustifying framework is quite interesting.\n2. Experiments demonstrate good empirical benefits of the approach.\n\nCons: \n1. While, the paper seemed well organized in the beginning, I got totally lost with Eq. (4-5). As I see, this objective is inaccurate and needs significant refinement. Specifically, it is unclear how is x^{adv} is related to x, and how is x^{adv} related to P*_theta? The paper tries to explain this objective in the paragraph below, but the explanation is very confusing as well.  A few other things that could help here:\na) It is said that \"Q_theta and P_\\theta* are the latent distributions induced by the natural example x\". How can a single data point induce a distribution? Do you assume the feature map from a hidden layer of a network represents a distribution? If so, in what sense? \nb) \"The adversarial example is crafted to induce the worst case distribution P*\". How is it crafted and what is the relation between P* and x? This is the key connection that is missing from (4-5).\n\n2. Moving along, Section 4.1 is organized very poorly as well. I believe too many concepts are tied together into one formulation in (6), making it hard to decipher. For example, why to include the classifier D^{1:C} within this formulation? Why not talk about it elsewhere and focus on the meat of the objective, systematically? \n\n3. Further, as I understand, x^{adv} is the first step that happens in (6), however, there is no \"adversary\" in this case, instead is finding a perturbed sample x' that maximizes the f-divergence. In what sense is x^{adv} then an adversarial sample? Perhaps the paper should re-define what is the definition of an adversarial example that it is using, to clearly state what the idea is. Technically, there is no requirement that the point x^{adv} found by this step will promote any data misclassification; however can be any point that is within a B(x,\\eps) ball from x, and that happens to maximize this divergence loss. Note that none of the other components D_W, f_theta, etc. are well trained in doing this optimization. So they could also be sub-optimal (in the sense of what the paper argues in the beginning of Page 4).\n\n4. Why is the middle formula in (6) minimizing over W to have both x and x^adv matched with the same label? Again, where is the adversary here? Or for that matter, how will the proposed approach achieve adversarial robustness ? \n\nMinor comments:\na. What is \\tau and T in (3)? \nb. How is f_\\theta defined in (6)? \nc. The paper writes that back and forth that there is no use of label information in the setup, however has labels used in discriminator in (6). This is very confusing. \nd. There is also reference to data manifold and manifold label in Figure 2, but these are not clearly explained. What precisely is the data manifold? Is it the latent distribution for a specific label? \ne. Page 4, top para: \"without considering the inter-relationship between data samples\". Won't this relation be captured implicitly through the neural network parameters theta when perturbations on all the samples are used in the training process?\n\nOverall, I think this paper needs a thorough revision to explain well its technical contributions. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603990972370}], "openreview_url": "https://openreview.net/forum?id=yvuk0RsLoP7", "arxiv_id": "2107.04401", "paper_pdf": "papers/yvuk0RsLoP7.pdf", "paper_pdf_sha256": "1142db4cb99504ec3da5363fdf6fbdfd61aee9071718e48bb6695f7830a66479", "paper_pdf_bytes": 4605606, "paper_pdf_source": "openreview", "code_url": "https://github.com/LitterQ/ATLD-pytorch", "code_repository": "LitterQ/ATLD-pytorch", "code_commit": "36bc00c32ff92995ad89622285ec30018e7d3f40", "code_archive": "repos/yvuk0RsLoP7.zip", "code_archive_sha256": "985c549f5bf1eaecbe054c4be86653a03992868b042450ac2a6bf3700683fa9d", "code_archive_bytes": 469729, "code_file_count": 59, "code_extensions": {".py": 49, ".sh": 10}, "github_disk_usage_kb": 693, "github_languages": {"Python": 334224, "Shell": 3683}, "github_archived": false, "github_pushed_at": "2025-03-23T10:19:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-model-robustness-with-latent-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJeoE0VKDS", "year": 2020, "status": "rejected", "title": "Novelty Search in representational space for sample efficient exploration", "authors": ["Ruo Yu Tao", "Vincent François-Lavet", "Joelle Pineau"], "authorids": ["ruo.tao@mail.mcgill.ca", "vincent.francois-lavet@mail.mcgill.ca", "jpineau@cs.mcgill.ca"], "authors_source": "OpenReview API", "abstract": "We present a new approach for efficient exploration which leverages a low-dimensional encoding of the environment learned with a combination of model-based and model-free objectives. Our approach uses intrinsic rewards that are based on a weighted distance of nearest neighbors in the low dimensional representational space to gauge novelty.\nWe then leverage these intrinsic rewards for sample-efficient exploration with planning routines in representational space.\nOne key element of our approach is that we perform more gradient steps in-between every environment step in order to ensure the model accuracy. We test our approach on a number of maze tasks, as well as a control problem and show that our exploration approach is more sample-efficient compared to strong baselines. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rJghN08aFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1088/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an approach to exploration by utilizing an intrinsic reward based on distances in a learned, evolving abstract representation space. The abstract space is learned utilizing both model-free and model-based losses, and the behaviour policy is based on planning combining the model-free and model-based components with an epsilon-greedy exploration strategy. Learning the abstract representation space itself is based on a previous work, but the contribution of this paper is the utility of it to design the reward bonus for exploration by utilizing distances in this evolving representation space.\n\nAs it stands, I am leaning towards rejecting the paper, for the following reasons.\n(1) while the idea proposed is interesting, the current work rather explores it in a limited manner which is unsatisfactory.\n(2) I think the presentation of the bonus itself -- novelty search (Section 4), which is the core of the paper, is rather unclear. (3) The assumption of deterministic transition dynamics may be ignored in favour of games which seem to be our benchmarks, but the results presented for the control tasks, Table 1, are not statistically significant, and the paper is missing details about the architecture/sweep for the baselines experimented with. \n(4) Parts of the paper is rather unclear/feels disconnected -- for instance, the interpretable abstract representation bit; this was a loss in the original work, and seems to be just mentioned arbitrarily here while the loss isn't really used (unless it is used, and not mentioned in the paper).\n(5) Overall, the proposed reward bonus is a heuristic whose specific design choice isn't statistically shown to be useful (Ablation in Appendix), and the empirical results comparing to other methods are underwhelming.\n\nHere are my main points of concern which I hope the authors address in the rebuttal:\n(1) Designing reward bonuses to induce exploratory behaviour in the agent has seen a surge of publications in the Deep RL literature in recent years. The key property all these methods aim for is a bonus that pushes the agent to the boundaries of its current \"known region\", and then rely on the stochasticity due to epsilon-greedy to cross that boundary -- pushing this boundary further. While this is different from exploration to reduce uncertainty, it is nonetheless a reasonable approach leading to competitive policies when evaluated in deep RL. But a characteristic all these bonuses aim for is that they fade away with time -- for instance count-based bonus are inversely proportional to visit counts, or prediction error bonuses go to 0 as the prediction becomes more accurate. But what do these novelty bonuses converge to? Is it just a stationary value based on consecutive loss parameter (in which case the hope is they don't affect the external reward scale, they just shift it uniformly)?\n(2) What exactly are the nearest neighbours? Is it a search based on the data in the buffer or is it a notion of temporal neighbours?\n(3) If it's temporal, why would there ever be biased for some states -- \"We do this in order..novel states\".\n(4) I was completely unable to understand the section in the Appendix which is making a case for the ranked weighting. If you have a succinct explanation for the heuristic it'd be great.\n(5) Further, as a heuristic it is mentioned that l2 norm may not be effective if the dimensionality of the representation space is increased. So why the heuristic? I think it either needs more empirical validation, or a theoretical justification.\n(6) While the evaluation scheme used in the paper to quantify the exploration of the behaviour policy is interesting -- y-axis of plots in Figure 4 for the Labyrinth task -- why/what exactly is the role of Figure 2? Is the interpretability loss used here? Is it to reason for utilizing e-greedy instead of a purely-greedy behaviour? I think this is a little unclear, and can be better clarified. Further, the distinction of primary and secondary features is interesting, but their clear demarcation is rather questionable in more complicated domains -- in the abstract space.\n(7) Do you have a hypothesis for why the 1-step value functions are not sufficient for decision making in this simple domain -- labyrinth - with the abstract representations?\n(8) If model-based algorithms get more steps to learn shouldn't model-free too? I'm not sure I understand the reasoning for the experiment design choice.\n(9) Whats the architecture used for Bootstrap DQN? It needs to have multiple heads -- but based on the current architecture that doesn't seem likely.\n(10) Are the extrinsic rewards ignored in learning -- \"only focus on intrinsic rewards\" (Section 6.2.2)? If they are for the proposed method, are they for the competitors too? If so why, and what is the reward for Bootstrap DQN?\n(11) I think the Discussion section raises interesting points about interpretability and metric learning, but I do think the conclusions drawn are a little inflated.\n(12) The ablation study in Section D of the Appendix is not statistically significant -- so why is wighted reward useful? Please comment.\n(13) How would stochasticity in transition dynamics affect the abstract representation space? Discussing this would be very interesting.\n(14) Learning curves for the control tasks?\n\nComments about typos/possible points of confusion:\n(1) The last para in Section 6.1 -- discusses \"open\" labyrinth heat map, then what do we mean by learning the dynamics of the wall? There is no wall in open, right?\n(2) In Section 4 -- I think x_{t+1} is an estimate from the unrolled model -- \\hat{x}_{t+1}? Further, it would be helpful to mention that it is an estimate based on the learned model.\n(3) n_freq is used in the pseudocode in the main paper -- but no mention of it to explain it is made in the main.\n(4) Contrasting the work to existing literature would be useful (in the Related Work section; as opposed to summarizing existing work).\n(5) buffered Q network --> target networks?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "The paper proposes an approach to exploration by utilizing an intrinsic reward based on distances in a learned, evolving abstract representation space. The abstract space is learned utilizing both model-free and model-based losses, and the behaviour policy is based on planning combining the model-free and model-based components with an epsilon-greedy exploration strategy. Learning the abstract representation space itself is based on a previous work, but the contribution of this paper is the utility of it to design the reward bonus for exploration by utilizing distances in this evolving representation space.\n\nAs it stands, I am leaning towards rejecting the paper, for the following reasons.\n(1) while the idea proposed is interesting, the current work rather explores it in a limited manner which is unsatisfactory.\n(2) I think the presentation of the bonus itself -- novelty search (Section 4), which is the core of the paper, is rather unclear. (3) The assumption of deterministic transition dynamics may be ignored in favour of games which seem to be our benchmarks, but the results presented for the control tasks, Table 1, are not statistically significant, and the paper is missing details about the architecture/sweep for the baselines experimented with. \n(4) Parts of the paper is rather unclear/feels disconnected -- for instance, the interpretable abstract representation bit; this was a loss in the original work, and seems to be just mentioned arbitrarily here while the loss isn't really used (unless it is used, and not mentioned in the paper).\n(5) Overall, the proposed reward bonus is a heuristic whose specific design choice isn't statistically shown to be useful (Ablation in Appendix), and the empirical results comparing to other methods are underwhelming.\n\nHere are my main points of concern which I hope the authors address in the rebuttal:\n(1) Designing reward bonuses to induce exploratory behaviour in the agent has seen a surge of publications in the Deep RL literature in recent years. The key property all these methods aim for is a bonus that pushes the agent to the boundaries of its current \"known region\", and then rely on the stochasticity due to epsilon-greedy to cross that boundary -- pushing this boundary further. While this is different from exploration to reduce uncertainty, it is nonetheless a reasonable approach leading to competitive policies when evaluated in deep RL. But a characteristic all these bonuses aim for is that they fade away with time -- for instance count-based bonus are inversely proportional to visit counts, or prediction error bonuses go to 0 as the prediction becomes more accurate. But what do these novelty bonuses converge to? Is it just a stationary value based on consecutive loss parameter (in which case the hope is they don't affect the external reward scale, they just shift it uniformly)?\n(2) What exactly are the nearest neighbours? Is it a search based on the data in the buffer or is it a notion of temporal neighbours?\n(3) If it's temporal, why would there ever be biased for some states -- \"We do this in order..novel states\".\n(4) I was completely unable to understand the section in the Appendix which is making a case for the ranked weighting. If you have a succinct explanation for the heuristic it'd be great.\n(5) Further, as a heuristic it is mentioned that l2 norm may not be effective if the dimensionality of the representation space is increased. So why the heuristic? I think it either needs more empirical validation, or a theoretical justification.\n(6) While the evaluation scheme used in the paper to quantify the exploration of the behaviour policy is interesting -- y-axis of plots in Figure 4 for the Labyrinth task -- why/what exactly is the role of Figure 2? Is the interpretability loss used here? Is it to reason for utilizing e-greedy instead of a purely-greedy behaviour? I think this is a little unclear, and can be better clarified. Further, the distinction of primary and secondary features is interesting, but their clear demarcation is rather questionable in more complicated domains -- in the abstract space.\n(7) Do you have a hypothesis for why the 1-step value functions are not sufficient for decision making in this simple domain -- labyrinth - with the abstract representations?\n(8) If model-based algorithms get more steps to learn shouldn't model-free too? I'm not sure I understand the reasoning for the experiment design choice.\n(9) Whats the architecture used for Bootstrap DQN? It needs to have multiple heads -- but based on the current architecture that doesn't seem likely.\n(10) Are the extrinsic rewards ignored in learning -- \"only focus on intrinsic rewards\" (Section 6.2.2)? If they are for the proposed method, are they for the competitors too? If so why, and what is the reward for Bootstrap DQN?\n(11) I think the Discussion section raises interesting points about interpretability and metric learning, but I do think the conclusions drawn are a little inflated.\n(12) The ablation study in Section D of the Appendix is not statistically significant -- so why is wighted reward useful? Please comment.\n(13) How would stochasticity in transition dynamics affect the abstract representation space? Discussing this would be very interesting.\n(14) Learning curves for the control tasks?\n\nComments about typos/possible points of confusion:\n(1) The last para in Section 6.1 -- discusses \"open\" labyrinth heat map, then what do we mean by learning the dynamics of the wall? There is no wall in open, right?\n(2) In Section 4 -- I think x_{t+1} is an estimate from the unrolled model -- \\hat{x}_{t+1}? Further, it would be helpful to mention that it is an estimate based on the learned model.\n(3) n_freq is used in the pseudocode in the main paper -- but no mention of it to explain it is made in the main.\n(4) Contrasting the work to existing literature would be useful (in the Related Work section; as opposed to summarizing existing work).\n(5) buffered Q network --> target networks?\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571806771772}, {"id": "HkeWLTChtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1088/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method of sample-efficient exploration for RL agent. The main problem at hand is the presence of irrelevant information in raw observations. To solve this problem, the authors leverage novelty heuristics in a lower-dimensional representation of a state, for which they propose a novelty measure. Then they describe a combination of model-based and model-free approaches with the novelty metric used as an intrinsic reward for planning that they use to compare with baselines solutions. They conduct experiments to show that their algorithm outperforms random and count-based baselines. They show that their approach has better results then Random Policy, Prediction error incentivized exploration, Hash count-based exploration, Bootstrap DQN while playing Acrobot and Multi-step Maze.\n\nAuthors propose a novel approach to the problem of exploration. They test their method by experiments conducted in two environments, where they use the same model architectures and model-free methods for all types of novelty metrics, which shows the contribution of the proposed method in the results of learning.\n\nTo sum up, the decision is to accept the paper as the problem is important, ideas are rather new, and results are better compared to other approaches.\n\n1. The dependence of the quality of the dimensionality representational state is unclear. For different environments, different abstract representation dimensions are chosen, but the reason is not explained.\n2. Word \"we\" is overused in the article", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper proposes a method of sample-efficient exploration for RL agent. The main problem at hand is the presence of irrelevant information in raw observations. To solve this problem, the authors leverage novelty heuristics in a lower-dimensional representation of a state, for which they propose a novelty measure. Then they describe a combination of model-based and model-free approaches with the novelty metric used as an intrinsic reward for planning that they use to compare with baselines solutions. They conduct experiments to show that their algorithm outperforms random and count-based baselines. They show that their approach has better results then Random Policy, Prediction error incentivized exploration, Hash count-based exploration, Bootstrap DQN while playing Acrobot and Multi-step Maze.\n\nAuthors propose a novel approach to the problem of exploration. They test their method by experiments conducted in two environments, where they use the same model architectures and model-free methods for all types of novelty metrics, which shows the contribution of the proposed method in the results of learning.\n\nTo sum up, the decision is to accept the paper as the problem is important, ideas are rather new, and results are better compared to other approaches.\n\n1. The dependence of the quality of the dimensionality representational state is unclear. For different environments, different abstract representation dimensions are chosen, but the reason is not explained.\n2. Word \"we\" is overused in the article"}, "tcdate": 1571773768868}, {"id": "rkeK9jTnFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1088/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method for efficient exploration in tabular MDPs as well as a simple control environment. The proposed method uses deterministic encoders to learn a low dimensional representation of the environment dynamics, which preserves distances between states such that “close” states in the full MDP are close in the learned representation. An intrinsic reward is formulated based on a measure of novelty, given by distance between new states, and a stored replay buffer.  Along with the dynamics model, a model-free agent employs Q learning to find a good policy. Experiments are performed on 3 tabular environments and the acrobot control task. \n\nPros: \n1.\tOverall the paper is clear and the proposed method makes sense intuitively. The intrinsic reward is cheap to compute and the state abstraction offers a nice way to visualize state differences the agent thinks are important.\n\n2.\tThe method seems to be sample efficient with regard to strong baselines like [1]\n\nCons:\n1.\tIt seems difficult to argue the efficacy of a low-dimensional state representation that doesn’t scale with state dimensionality. As shown in [2] and [3], learning effective state abstractions in high dimensions can require considerably more effort. \n\n2.\tGiven that there exist novelty based intrinsic rewards which compute state abstractions in high dimensional environments [3], I find it hard to see the usefulness of the proposed method. \n\n3.\tThe choice of distance metric for the representational space is not well motivated. As correctly stated by the authors, the L_2 norm will cease to be a good metric as state dimensionality increases.\n\n4.\tThere are too many grid-world experiments. The point of the first two experiments can be made simply using the four-room environment. This could make room for a more interesting experiment such as MuJoCo Ant Maze. \n\nMy main issue with the work in its current form is that the method is too light in terms of technical contribution. Simple methods are ok (even valuable!) but there should be a certain about of rigorous analysis which shows that the simple method can be used as a foundation for further work. For a method which mostly examines tabular environments, I expect some analysis of the methods efficiency with regard to data efficiency -- the main point of the paper. [1] and [4] which are used as comparisons, provide such analysis. If a convincing theoretical analysis is out of reach, then it could be sufficient to provide extensive experimental evidence supporting the claims. In this case it could include an examination of different metrics, additional (ideally more difficult) environments, and comparison to other baselines like [5], [3].\n\nDue to what I see as a lack of technical contribution, I do not recommend acceptance to ICLR at this time. \n\nA more compelling submission would include the following:\n●\tA more detailed motivation for why the L2 norm makes sense as a distance metric. \n○\tIn an abstract space, it's more natural to use a statistical distance like the KL or JS divergence. These metrics have drawbacks, but they should be discussed\n●\tAn analysis of the limit behavior of the proposed method. Given enough time an intrinsic reward should explore every state in a deterministic environment. Does this happen in the limiting case -- if not, is the margin acceptable. \n●\tMore extensive experiments. This method can clearly admit convolutional architectures so experiments on more interesting environments are viable. Though I believe this would require more complex models such as a VAE, and may change the submission considerably.  \n\nMinor notes\n●\tSection 3: “when [the distance between transitions is less than the slack ratio] the transitions are mostly accurate within a ball of radius \\frac{w}/{\\delta}. This is too vague, what does mostly accurate mean? \n●\tEq (6), is \\alpha a hyperparameter as well as the learning rate? If \\alpha is just the learning rate than the equation is incorrect, because the learning rate is applied to the gradient of the loss, not the loss itself. \n●\tThe description of the planning algorithm and Q learning in section 4 is a little sloppy, a clearer description would be appreciated. \n●\tComputing novelty with respect to a state’s nearest neighbors is problematic at scale. This point should be at least acknowledged.  \n\n\n[1] Osband, Ian, et al. \"Deep exploration via bootstrapped DQN.\" Advances in neural information processing systems. 2016.\n[2] Kim, Hyoungseok, et al. \"EMI: Exploration with Mutual Information.\" International Conference on Machine Learning. 2019.\n[3] Ha, David, and Jürgen Schmidhuber. \"World models.\" arXiv preprint arXiv:1803.10122 (2018).\n[4] Bellemare, Marc, et al. \"Unifying count-based exploration and intrinsic motivation.\" Advances in Neural Information Processing Systems. 2016.\n[5] Pathak, Deepak, et al. \"Curiosity-driven exploration by self-supervised prediction.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2017.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a method for efficient exploration in tabular MDPs as well as a simple control environment. The proposed method uses deterministic encoders to learn a low dimensional representation of the environment dynamics, which preserves distances between states such that “close” states in the full MDP are close in the learned representation. An intrinsic reward is formulated based on a measure of novelty, given by distance between new states, and a stored replay buffer.  Along with the dynamics model, a model-free agent employs Q learning to find a good policy. Experiments are performed on 3 tabular environments and the acrobot control task. \n\nPros: \n1.\tOverall the paper is clear and the proposed method makes sense intuitively. The intrinsic reward is cheap to compute and the state abstraction offers a nice way to visualize state differences the agent thinks are important.\n\n2.\tThe method seems to be sample efficient with regard to strong baselines like [1]\n\nCons:\n1.\tIt seems difficult to argue the efficacy of a low-dimensional state representation that doesn’t scale with state dimensionality. As shown in [2] and [3], learning effective state abstractions in high dimensions can require considerably more effort. \n\n2.\tGiven that there exist novelty based intrinsic rewards which compute state abstractions in high dimensional environments [3], I find it hard to see the usefulness of the proposed method. \n\n3.\tThe choice of distance metric for the representational space is not well motivated. As correctly stated by the authors, the L_2 norm will cease to be a good metric as state dimensionality increases.\n\n4.\tThere are too many grid-world experiments. The point of the first two experiments can be made simply using the four-room environment. This could make room for a more interesting experiment such as MuJoCo Ant Maze. \n\nMy main issue with the work in its current form is that the method is too light in terms of technical contribution. Simple methods are ok (even valuable!) but there should be a certain about of rigorous analysis which shows that the simple method can be used as a foundation for further work. For a method which mostly examines tabular environments, I expect some analysis of the methods efficiency with regard to data efficiency -- the main point of the paper. [1] and [4] which are used as comparisons, provide such analysis. If a convincing theoretical analysis is out of reach, then it could be sufficient to provide extensive experimental evidence supporting the claims. In this case it could include an examination of different metrics, additional (ideally more difficult) environments, and comparison to other baselines like [5], [3].\n\nDue to what I see as a lack of technical contribution, I do not recommend acceptance to ICLR at this time. \n\nA more compelling submission would include the following:\n●\tA more detailed motivation for why the L2 norm makes sense as a distance metric. \n○\tIn an abstract space, it's more natural to use a statistical distance like the KL or JS divergence. These metrics have drawbacks, but they should be discussed\n●\tAn analysis of the limit behavior of the proposed method. Given enough time an intrinsic reward should explore every state in a deterministic environment. Does this happen in the limiting case -- if not, is the margin acceptable. \n●\tMore extensive experiments. This method can clearly admit convolutional architectures so experiments on more interesting environments are viable. Though I believe this would require more complex models such as a VAE, and may change the submission considerably.  \n\nMinor notes\n●\tSection 3: “when [the distance between transitions is less than the slack ratio] the transitions are mostly accurate within a ball of radius \\frac{w}/{\\delta}. This is too vague, what does mostly accurate mean? \n●\tEq (6), is \\alpha a hyperparameter as well as the learning rate? If \\alpha is just the learning rate than the equation is incorrect, because the learning rate is applied to the gradient of the loss, not the loss itself. \n●\tThe description of the planning algorithm and Q learning in section 4 is a little sloppy, a clearer description would be appreciated. \n●\tComputing novelty with respect to a state’s nearest neighbors is problematic at scale. This point should be at least acknowledged.  \n\n\n[1] Osband, Ian, et al. \"Deep exploration via bootstrapped DQN.\" Advances in neural information processing systems. 2016.\n[2] Kim, Hyoungseok, et al. \"EMI: Exploration with Mutual Information.\" International Conference on Machine Learning. 2019.\n[3] Ha, David, and Jürgen Schmidhuber. \"World models.\" arXiv preprint arXiv:1803.10122 (2018).\n[4] Bellemare, Marc, et al. \"Unifying count-based exploration and intrinsic motivation.\" Advances in Neural Information Processing Systems. 2016.\n[5] Pathak, Deepak, et al. \"Curiosity-driven exploration by self-supervised prediction.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2017.\n"}, "tcdate": 1571769232687}], "openreview_url": "https://openreview.net/forum?id=SJeoE0VKDS", "arxiv_id": "2009.13579", "paper_pdf": "papers/SJeoE0VKDS.pdf", "paper_pdf_sha256": "832d9a85d4f22d440c38518189218d3f46fd7389d5c4c1960296be996dfe6b4d", "paper_pdf_bytes": 1756879, "paper_pdf_source": "openreview", "code_url": "https://github.com/taodav/nsrs", "code_repository": "taodav/nsrs", "code_commit": "7d9329134cd1f3305db8c8afe87504349875ef90", "code_archive": "repos/SJeoE0VKDS.zip", "code_archive_sha256": "aa7c2788c29e68cceb2331ec092fed4d9a392d3e7fd250a671f8c3b1cf1d3db1", "code_archive_bytes": 1560024, "code_file_count": 60, "code_extensions": {".py": 45, ".sh": 8, ".ipynb": 7}, "github_disk_usage_kb": 1731, "github_languages": {"Jupyter Notebook": 4485362, "Python": 433443, "Shell": 4228}, "github_archived": false, "github_pushed_at": "2024-07-16T14:51:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/novelty-search-in-representational-space-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "39CEoMXTHh", "year": 2026, "status": "rejected", "title": "$\\texttt{LLINBO}$: Trustworthy LLM-in-the-Loop Bayesian Optimization", "authors": ["Chih-Yu Chang", "Milad Azvar", "Chinedum Okwudire", "Raed Al Kontar"], "authorids": ["~Chih-Yu_Chang1", "~Milad_Azvar1", "~Chinedum_Okwudire1", "~Raed_Al_Kontar1"], "authors_source": "OpenReview API", "abstract": "Bayesian optimization (BO) is a sequential decision-making tool widely used for optimizing expensive black-box functions. Recently, Large Language Models (LLMs) have shown remarkable adaptability in low-data regimes, making them promising tools for black-box optimization by leveraging contextual knowledge to propose high-quality query points. However, relying solely on LLMs as optimization agents introduces risks due to their lack of explicit surrogate modeling and calibrated uncertainty, as well as their inherently opaque internal mechanisms. This structural opacity makes it difficult to characterize or control the exploration–exploitation trade-off, ultimately undermining theoretical tractability and reliability. To address this, we propose $\\texttt{LLINBO}$: LLM-in-the-Loop BO, a hybrid framework for BO that combines LLMs with statistical surrogate experts (e.g., Gaussian Processes ($\\mathcal{GP}$)). The core philosophy is to leverage contextual reasoning strengths of LLMs for early exploration, while relying on principled statistical models to guide efficient exploitation. Specifically, we introduce three mechanisms that enable this collaboration and establish their theoretical guarantees. We end the paper with a real-life proof-of-concept in the context of 3D printing.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "lt1CtFcvmw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22200/Reviewer_nvHk"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "LLM in the Loop BO (LLINBO) is a hybrid framework for Bayesian optimization that combines large language models with statistical surrogate experts, with Gaussian processes as the focus in this paper. The goal is to tap the contextual reasoning strengths of LLMs while relying on principled uncertainty from statistical surrogates to enable more trustworthy optimization. The work advances the growing line of research that uses the few shot capabilities of LLMs for black box optimization, while addressing inherent limitations such as the lack of calibrated uncertainty and the opaque behavior of LLMs that makes them hard to interpret. To leverage LLMs without inheriting these risks, the authors propose a framework in which the LLM accelerates early exploration and the surrogate model increasingly guides and exploits as data accumulates.\n\nAt a high level, each BO round maintains a GP posterior, queries the LLM for a candidate $x_{LLM,t}$, evaluates that suggestion with the GP to accept, refine, or reject it (according to one of three mechanisms), and then evaluates the chosen design point. The GP “check and balance” is instantiated in three variants. LLINBO-TRANSIENT uses a Bernoulli schedule that prioritizes LLM suggestions early and shifts toward GP driven selection later. LLINBO-JUSTIFY uses the GP posterior to judge $x_{LLM,t}$, accepting it only if it lies within a specified suboptimality region; otherwise it retains the GP chosen point $x_{GP,t}$. LLINBO Constrained treats $x_{LLM,t}$ as a promising design choice and builds a constrained GP by conditioning on the event that this chosen design outperforms the current posterior mean maximizer, then acquires using a Monte Carlo approximation under the constrained posterior. Unlike the first two methods, the constrained approach avoids additional tuning parameters. All three mechanisms use GP-UCB variants s the acquisition function and come with upper bounds that ensure no regret as $T \\to \\infty$. The paper presents a range of empirical results to demonstrate early gains from LLM guidance and competitive or improved performance as the GP takes over for several problems.", "review_text": "LLM in the Loop BO (LLINBO) is a hybrid framework for Bayesian optimization that combines large language models with statistical surrogate experts, with Gaussian processes as the focus in this paper. The goal is to tap the contextual reasoning strengths of LLMs while relying on principled uncertainty from statistical surrogates to enable more trustworthy optimization. The work advances the growing line of research that uses the few shot capabilities of LLMs for black box optimization, while addressing inherent limitations such as the lack of calibrated uncertainty and the opaque behavior of LLMs that makes them hard to interpret. To leverage LLMs without inheriting these risks, the authors propose a framework in which the LLM accelerates early exploration and the surrogate model increasingly guides and exploits as data accumulates.\n\nAt a high level, each BO round maintains a GP posterior, queries the LLM for a candidate $x_{LLM,t}$, evaluates that suggestion with the GP to accept, refine, or reject it (according to one of three mechanisms), and then evaluates the chosen design point. The GP “check and balance” is instantiated in three variants. LLINBO-TRANSIENT uses a Bernoulli schedule that prioritizes LLM suggestions early and shifts toward GP driven selection later. LLINBO-JUSTIFY uses the GP posterior to judge $x_{LLM,t}$, accepting it only if it lies within a specified suboptimality region; otherwise it retains the GP chosen point $x_{GP,t}$. LLINBO Constrained treats $x_{LLM,t}$ as a promising design choice and builds a constrained GP by conditioning on the event that this chosen design outperforms the current posterior mean maximizer, then acquires using a Monte Carlo approximation under the constrained posterior. Unlike the first two methods, the constrained approach avoids additional tuning parameters. All three mechanisms use GP-UCB variants s the acquisition function and come with upper bounds that ensure no regret as $T \\to \\infty$. The paper presents a range of empirical results to demonstrate early gains from LLM guidance and competitive or improved performance as the GP takes over for several problems.", "strengths": "1. The paper is well motivated and tackles a growing area of research in the BO community. \n\n2. The authors present a clear, coherent narrative of their proposed methods. \n\n3. The paper provides solid theoretical results that establish the no regret behavior of the mechanisms.\n\n4. The work includes a practical demonstration that shows how LLINBO performs in a real setting.\n\n5. The experimental section is extensive, with clearly documented details that support reproducibility.", "weaknesses": "1. The novelty is quite moderate. Although the third variant (LLINBO Constrained) presents a meaningful new idea, its scalability and the compute trade offs are not clearly justified with profiling or ablations.\n\n2. Related to the above, the paper claims LLINBO is the first hybrid framework that integrates LLMs and GPs for BO. This does not seem entirely correct: for example, Kristiadi et al. (ICML 2024), “A Sober Look at LLMs for Materials Discovery,” used an LLM as a feature extractor into a GP surrogate for BO in materials discovery. The authors should clarify the novelty relative to such setups.\n\n3. Several controlling parameters are required across the three mechanisms. Although recommended values are provided, there is no elegant or reliable procedure for selecting them across domains.\n\n4. The synthetic functions for the black-box optimization (BBO) task do not mimic the higher dimensional settings common in BO papers as the chosen functions range from 2D to 6D. Also for the BBO tasks, LLINBO variants and standard BO are not clearly differentiated in final performance, aside from early stage gains that do not always change the end result. The authors, however, do demonstrate stronger performance on real world tasks and hyperparameter tuning settings.", "questions": "1. Can the authors clarify the distinction from Kristiadi et al. paper as noted earlier, and whether the claim of being the first hybrid framework combining GPs and LLMs for BO is accurate? Fundamentally, the methods are different, but the “first” claim may not be entirely correct. A clarification would be greatly welcomed. \n\n2. Can the authors comment on the restriction of the synthetic functions to just 2 to 6 dimensions and, if possible, report performance on higher dimensional synthetic problems as typically done in BO papers (for example, 2D to 16D for low to moderate dimensional settings)? If this is not feasible, a brief justification would be helpful.\n\n3. Can the authors discuss the cost implications across the three methods, beyond the interpretability argument that motivated the choice of LLINBO-TRANSIENT for the 3D-printing problem?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "LLM in the Loop BO (LLINBO) is a hybrid framework for Bayesian optimization that combines large language models with statistical surrogate experts, with Gaussian processes as the focus in this paper. The goal is to tap the contextual reasoning strengths of LLMs while relying on principled uncertainty from statistical surrogates to enable more trustworthy optimization. The work advances the growing line of research that uses the few shot capabilities of LLMs for black box optimization, while addressing inherent limitations such as the lack of calibrated uncertainty and the opaque behavior of LLMs that makes them hard to interpret. To leverage LLMs without inheriting these risks, the authors propose a framework in which the LLM accelerates early exploration and the surrogate model increasingly guides and exploits as data accumulates.\n\nAt a high level, each BO round maintains a GP posterior, queries the LLM for a candidate $x_{LLM,t}$, evaluates that suggestion with the GP to accept, refine, or reject it (according to one of three mechanisms), and then evaluates the chosen design point. The GP “check and balance” is instantiated in three variants. LLINBO-TRANSIENT uses a Bernoulli schedule that prioritizes LLM suggestions early and shifts toward GP driven selection later. LLINBO-JUSTIFY uses the GP posterior to judge $x_{LLM,t}$, accepting it only if it lies within a specified suboptimality region; otherwise it retains the GP chosen point $x_{GP,t}$. LLINBO Constrained treats $x_{LLM,t}$ as a promising design choice and builds a constrained GP by conditioning on the event that this chosen design outperforms the current posterior mean maximizer, then acquires using a Monte Carlo approximation under the constrained posterior. Unlike the first two methods, the constrained approach avoids additional tuning parameters. All three mechanisms use GP-UCB variants s the acquisition function and come with upper bounds that ensure no regret as $T \\to \\infty$. The paper presents a range of empirical results to demonstrate early gains from LLM guidance and competitive or improved performance as the GP takes over for several problems.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The paper is well motivated and tackles a growing area of research in the BO community. \n\n2. The authors present a clear, coherent narrative of their proposed methods. \n\n3. The paper provides solid theoretical results that establish the no regret behavior of the mechanisms.\n\n4. The work includes a practical demonstration that shows how LLINBO performs in a real setting.\n\n5. The experimental section is extensive, with clearly documented details that support reproducibility.", "weaknesses": "1. The novelty is quite moderate. Although the third variant (LLINBO Constrained) presents a meaningful new idea, its scalability and the compute trade offs are not clearly justified with profiling or ablations.\n\n2. Related to the above, the paper claims LLINBO is the first hybrid framework that integrates LLMs and GPs for BO. This does not seem entirely correct: for example, Kristiadi et al. (ICML 2024), “A Sober Look at LLMs for Materials Discovery,” used an LLM as a feature extractor into a GP surrogate for BO in materials discovery. The authors should clarify the novelty relative to such setups.\n\n3. Several controlling parameters are required across the three mechanisms. Although recommended values are provided, there is no elegant or reliable procedure for selecting them across domains.\n\n4. The synthetic functions for the black-box optimization (BBO) task do not mimic the higher dimensional settings common in BO papers as the chosen functions range from 2D to 6D. Also for the BBO tasks, LLINBO variants and standard BO are not clearly differentiated in final performance, aside from early stage gains that do not always change the end result. The authors, however, do demonstrate stronger performance on real world tasks and hyperparameter tuning settings.", "questions": "1. Can the authors clarify the distinction from Kristiadi et al. paper as noted earlier, and whether the claim of being the first hybrid framework combining GPs and LLMs for BO is accurate? Fundamentally, the methods are different, but the “first” claim may not be entirely correct. A clarification would be greatly welcomed. \n\n2. Can the authors comment on the restriction of the synthetic functions to just 2 to 6 dimensions and, if possible, report performance on higher dimensional synthetic problems as typically done in BO papers (for example, 2D to 16D for low to moderate dimensional settings)? If this is not feasible, a brief justification would be helpful.\n\n3. Can the authors discuss the cost implications across the three methods, beyond the interpretability argument that motivated the choice of LLINBO-TRANSIENT for the 3D-printing problem?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761978370104}, {"id": "1eQ4Fdhrwa", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22200/Reviewer_w5nX"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes LLINBO, a trustworthy hybrid Bayesian optimization framework that integrates LLMs with GP surrogates. LLMs are used during early exploration to propose contextually informed designs, while GPs gradually take over to ensure reliable uncertainty quantification and convergence guarantees. Theoretically, it derives regret bounds under RKHS assumptions. Empirically it validates the framework on synthetic benchmarks, hyperparameter tuning, and a 3D printing experiment, showing that LLINBO achieves strong early performance and near-zero stringing in real-world tests.", "review_text": "The paper proposes LLINBO, a trustworthy hybrid Bayesian optimization framework that integrates LLMs with GP surrogates. LLMs are used during early exploration to propose contextually informed designs, while GPs gradually take over to ensure reliable uncertainty quantification and convergence guarantees. Theoretically, it derives regret bounds under RKHS assumptions. Empirically it validates the framework on synthetic benchmarks, hyperparameter tuning, and a 3D printing experiment, showing that LLINBO achieves strong early performance and near-zero stringing in real-world tests.", "strengths": "1.\tThe paper clearly motivates the need for a trustworthy hybrid. LLM assisted BO lacks explicit uncertainty and its performance degrades as data accumulate, so combining LLMs with calibrated surrogates like GPs can retain early contextual benefits and ensuring rigorous exploration–exploitation tradeoffs. The work appears to be the first to explicitly integrate LLM suggestions with GP surrogates and to analyze this interaction theoretically.\n2.\tThree mechanisms to coordinate LLMs and GPs during BO, including Transient (LLM exploring and shift to GP), Justify (evaluate LLM proposal against GP), and Constrained (treat LLM’s proposal as a soft constraint).\n3.\tThe framework emphasizes trustworthy optimization. By design, LLINBO mitigates the opaque and unbounded nature of LLM suggestions, and each mechanism has a built-in safety check.\n4.\tEach of the three methods is accompanied by a theoretical analysis, with regret bounds developed.", "weaknesses": "1. The theoretical analysis treats LLM suggestions generically (e.g., by gradually ignoring them or rejecting them if the GP’s acquisition is much higher) without modeling how good the LLM’s proposals actually are. Regret bounds rely on predetermined schedules for $p_t$, $\\psi_t$, or $S_t$ rather than any adaptive assessment of LLM quality. Consequently, the bounds mirror standard GP-based BO and do not show that LLM proposals improve asymptotic performance.\n\n2. Benchmarks involve low dimensional synthetic functions and relatively small budgets (10×dimension for BBO and 5×dimension for HPT) and the real world study tests only LLINBO Transient on one 3D printing setup. It remains unclear how the approaches scale to higher dimensional or discrete design spaces, whether the constrained method is computationally feasible when many samples are needed, and how robust the results are to different LLMs or prompts. \n\n3. LLINBO-Constrained assumes that the LLM’s suggestion $f(x_{\\mathrm{LLM},t})$ is better than the current mean maximum. If this assumption is violated, the algorithm may discard the LLM proposal entirely, and the CGP sampling procedure can be computationally expensive. The practical benefit of this mechanism relative to the simpler transient or justify variants is unclear, especially since it is not tested in the real-world experiment.\n\n4. The paper compares against a few baselines (GP-UCB, LLAMBO, and the authors’ LLAMBO-light). Given the nascent state of LLM-assisted BO, this is reasonable. One small concern is that the original LLAMBO method was modified (for practicality) – the full LLAMBO might generate multiple candidates and evaluate them with a surrogate, which could potentially yield better results than the one-shot “light” version used. Also, there are additional LLM enhanced BO methods should be compared and discussed: BioDiscoveryAgent [1], FunBO [2], LLaMEA-BO [3], and SLLMBO [4] etc. These are important works in this line of research.\n---\n[1] Roohani, Yusuf, et al. \"Biodiscoveryagent: An ai agent for designing genetic perturbation experiments.\" arXiv preprint arXiv:2405.17631 (2024).\n\n[2] Aglietti, Virginia, et al. \"Funbo: Discovering acquisition functions for bayesian optimization with funsearch.\" arXiv preprint arXiv:2406.04824 (2024).\n\n[3] Li, Wenhu, et al. \"LLaMEA-BO: A Large Language Model Evolutionary Algorithm for Automatically Generating Bayesian Optimization Algorithms.\" arXiv preprint arXiv:2505.21034 (2025).\n\n[4] Mahammadli, Kanan, and Seyda Ertekin. \"Sequential large language model-based hyper-parameter optimization.\" arXiv preprint arXiv:2410.20302 (2024).", "questions": "1. In LLINBO-Constrained, the authors assume $f(x_{\\mathrm{LLM},t}) > \\kappa_{t-1}$. How robust is the method when this assumption is wrong? Is it possible to use a probabilistic belief about the LLM’s suggestion quality rather than a hard constraint?\n\n2. Beyond a single candidate per iteration, have you experimented with eliciting richer information from the LLM?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes LLINBO, a trustworthy hybrid Bayesian optimization framework that integrates LLMs with GP surrogates. LLMs are used during early exploration to propose contextually informed designs, while GPs gradually take over to ensure reliable uncertainty quantification and convergence guarantees. Theoretically, it derives regret bounds under RKHS assumptions. Empirically it validates the framework on synthetic benchmarks, hyperparameter tuning, and a 3D printing experiment, showing that LLINBO achieves strong early performance and near-zero stringing in real-world tests.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1.\tThe paper clearly motivates the need for a trustworthy hybrid. LLM assisted BO lacks explicit uncertainty and its performance degrades as data accumulate, so combining LLMs with calibrated surrogates like GPs can retain early contextual benefits and ensuring rigorous exploration–exploitation tradeoffs. The work appears to be the first to explicitly integrate LLM suggestions with GP surrogates and to analyze this interaction theoretically.\n2.\tThree mechanisms to coordinate LLMs and GPs during BO, including Transient (LLM exploring and shift to GP), Justify (evaluate LLM proposal against GP), and Constrained (treat LLM’s proposal as a soft constraint).\n3.\tThe framework emphasizes trustworthy optimization. By design, LLINBO mitigates the opaque and unbounded nature of LLM suggestions, and each mechanism has a built-in safety check.\n4.\tEach of the three methods is accompanied by a theoretical analysis, with regret bounds developed.", "weaknesses": "1. The theoretical analysis treats LLM suggestions generically (e.g., by gradually ignoring them or rejecting them if the GP’s acquisition is much higher) without modeling how good the LLM’s proposals actually are. Regret bounds rely on predetermined schedules for $p_t$, $\\psi_t$, or $S_t$ rather than any adaptive assessment of LLM quality. Consequently, the bounds mirror standard GP-based BO and do not show that LLM proposals improve asymptotic performance.\n\n2. Benchmarks involve low dimensional synthetic functions and relatively small budgets (10×dimension for BBO and 5×dimension for HPT) and the real world study tests only LLINBO Transient on one 3D printing setup. It remains unclear how the approaches scale to higher dimensional or discrete design spaces, whether the constrained method is computationally feasible when many samples are needed, and how robust the results are to different LLMs or prompts. \n\n3. LLINBO-Constrained assumes that the LLM’s suggestion $f(x_{\\mathrm{LLM},t})$ is better than the current mean maximum. If this assumption is violated, the algorithm may discard the LLM proposal entirely, and the CGP sampling procedure can be computationally expensive. The practical benefit of this mechanism relative to the simpler transient or justify variants is unclear, especially since it is not tested in the real-world experiment.\n\n4. The paper compares against a few baselines (GP-UCB, LLAMBO, and the authors’ LLAMBO-light). Given the nascent state of LLM-assisted BO, this is reasonable. One small concern is that the original LLAMBO method was modified (for practicality) – the full LLAMBO might generate multiple candidates and evaluate them with a surrogate, which could potentially yield better results than the one-shot “light” version used. Also, there are additional LLM enhanced BO methods should be compared and discussed: BioDiscoveryAgent [1], FunBO [2], LLaMEA-BO [3], and SLLMBO [4] etc. These are important works in this line of research.\n---\n[1] Roohani, Yusuf, et al. \"Biodiscoveryagent: An ai agent for designing genetic perturbation experiments.\" arXiv preprint arXiv:2405.17631 (2024).\n\n[2] Aglietti, Virginia, et al. \"Funbo: Discovering acquisition functions for bayesian optimization with funsearch.\" arXiv preprint arXiv:2406.04824 (2024).\n\n[3] Li, Wenhu, et al. \"LLaMEA-BO: A Large Language Model Evolutionary Algorithm for Automatically Generating Bayesian Optimization Algorithms.\" arXiv preprint arXiv:2505.21034 (2025).\n\n[4] Mahammadli, Kanan, and Seyda Ertekin. \"Sequential large language model-based hyper-parameter optimization.\" arXiv preprint arXiv:2410.20302 (2024).", "questions": "1. In LLINBO-Constrained, the authors assume $f(x_{\\mathrm{LLM},t}) > \\kappa_{t-1}$. How robust is the method when this assumption is wrong? Is it possible to use a probabilistic belief about the LLM’s suggestion quality rather than a hard constraint?\n\n2. Beyond a single candidate per iteration, have you experimented with eliciting richer information from the LLM?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761976859286}, {"id": "D8SYeJQkSx", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22200/Reviewer_1aud"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper introduces LLINBO, a hybrid framework that integrates LLMs with statistical surrogate models like Gaussian Processes (GPs) to enhance trustworthiness in black-box optimization (BBO). LLMs excel in low-data regimes by leveraging contextual knowledge for early exploration, but they lack explicit surrogate modeling, calibrated uncertainty, and transparency, leading to risks in exploration-exploitation balance and theoretical reliability. LLINBO addresses this by using LLMs for initial design proposals while progressively relying on GPs for principled exploitation as data grows. Three specific mechanisms are proposed: LLINBO-Transient (transient LLM influence that fades over iterations), LLINBO-Justify (LLM justifies proposals, GP verifies feasibility), and LLINBO-Constrained (LLM proposals constrained within GP uncertainty bounds). These draw inspiration from federated learning and come with regret-based theoretical guarantees ensuring asymptotic optimality similar to standard BO. The framework is evaluated through simulations and a real-world proof-of-concept in 3D printing, demonstrating improved efficiency and robustness.", "review_text": "The paper introduces LLINBO, a hybrid framework that integrates LLMs with statistical surrogate models like Gaussian Processes (GPs) to enhance trustworthiness in black-box optimization (BBO). LLMs excel in low-data regimes by leveraging contextual knowledge for early exploration, but they lack explicit surrogate modeling, calibrated uncertainty, and transparency, leading to risks in exploration-exploitation balance and theoretical reliability. LLINBO addresses this by using LLMs for initial design proposals while progressively relying on GPs for principled exploitation as data grows. Three specific mechanisms are proposed: LLINBO-Transient (transient LLM influence that fades over iterations), LLINBO-Justify (LLM justifies proposals, GP verifies feasibility), and LLINBO-Constrained (LLM proposals constrained within GP uncertainty bounds). These draw inspiration from federated learning and come with regret-based theoretical guarantees ensuring asymptotic optimality similar to standard BO. The framework is evaluated through simulations and a real-world proof-of-concept in 3D printing, demonstrating improved efficiency and robustness.", "strengths": "1. Pioneers a hybrid LLM-GP approach for BO, explicitly modeling LLM degradation over time and hedging with statistical surrogates. The three variants offer flexible integration, addressing opacity and uncertainty issues in pure LLM-assisted BO.\n2. Provides rigorous regret guarantees for each mechanism by extending classical BO results to the hybrid setting. \n3. Balances LLM's contextual strengths (e.g., few-shot learning from prompts) with GP's uncertainty quantification. The 3D printing application shows real-world potential in costly evaluation scenarios like drug discovery or hyperparameter tuning.", "weaknesses": "1. Relies on GP surrogates; no discussion of alternatives like neural networks. LLM prompting (e.g., for justifications) may be sensitive, with no ablation studies mentioned. Theoretical guarantees assume smooth functions and high-probability bounds, potentially limiting generality.\n2. Combining LLMs and GPs could increase computational cost (e.g., LLM queries per iteration), but no analysis of scalability or comparisons to baselines in terms of runtime.\n3. Proof-of-concept in 3D printing is promising, but without full results, it's unclear how it compares to pure GP-BO or LLM-only methods across diverse benchmarks.", "questions": "Please see the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces LLINBO, a hybrid framework that integrates LLMs with statistical surrogate models like Gaussian Processes (GPs) to enhance trustworthiness in black-box optimization (BBO). LLMs excel in low-data regimes by leveraging contextual knowledge for early exploration, but they lack explicit surrogate modeling, calibrated uncertainty, and transparency, leading to risks in exploration-exploitation balance and theoretical reliability. LLINBO addresses this by using LLMs for initial design proposals while progressively relying on GPs for principled exploitation as data grows. Three specific mechanisms are proposed: LLINBO-Transient (transient LLM influence that fades over iterations), LLINBO-Justify (LLM justifies proposals, GP verifies feasibility), and LLINBO-Constrained (LLM proposals constrained within GP uncertainty bounds). These draw inspiration from federated learning and come with regret-based theoretical guarantees ensuring asymptotic optimality similar to standard BO. The framework is evaluated through simulations and a real-world proof-of-concept in 3D printing, demonstrating improved efficiency and robustness.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. Pioneers a hybrid LLM-GP approach for BO, explicitly modeling LLM degradation over time and hedging with statistical surrogates. The three variants offer flexible integration, addressing opacity and uncertainty issues in pure LLM-assisted BO.\n2. Provides rigorous regret guarantees for each mechanism by extending classical BO results to the hybrid setting. \n3. Balances LLM's contextual strengths (e.g., few-shot learning from prompts) with GP's uncertainty quantification. The 3D printing application shows real-world potential in costly evaluation scenarios like drug discovery or hyperparameter tuning.", "weaknesses": "1. Relies on GP surrogates; no discussion of alternatives like neural networks. LLM prompting (e.g., for justifications) may be sensitive, with no ablation studies mentioned. Theoretical guarantees assume smooth functions and high-probability bounds, potentially limiting generality.\n2. Combining LLMs and GPs could increase computational cost (e.g., LLM queries per iteration), but no analysis of scalability or comparisons to baselines in terms of runtime.\n3. Proof-of-concept in 3D printing is promising, but without full results, it's unclear how it compares to pure GP-BO or LLM-only methods across diverse benchmarks.", "questions": "Please see the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761790850960}, {"id": "8UNS41uWHA", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22200/Reviewer_m9hu"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper presents three different approaches on how to provide acquisition functions for Bayesian Optimisation with the use of an LLM together with a Gaussian process surrogate model. The motivation for this approach in the paper is the notion that the LLM will somehow contain \"contextual\" information. The first approach uses a linear combination of the LLM acquisition and the GP acquisition, the second uses a rejection criteria based on the GP surrogates acquisition function while the third uses the LLM proposal to create a constrained surrogate.\n\nThe paper ends with a set of experiments comparing the proposed method to several different baselines. The experiments are on classic BO tasks and the more interesting on a 3d printing task.", "review_text": "This paper presents three different approaches on how to provide acquisition functions for Bayesian Optimisation with the use of an LLM together with a Gaussian process surrogate model. The motivation for this approach in the paper is the notion that the LLM will somehow contain \"contextual\" information. The first approach uses a linear combination of the LLM acquisition and the GP acquisition, the second uses a rejection criteria based on the GP surrogates acquisition function while the third uses the LLM proposal to create a constrained surrogate.\n\nThe paper ends with a set of experiments comparing the proposed method to several different baselines. The experiments are on classic BO tasks and the more interesting on a 3d printing task.", "strengths": "The paper is somewhat original given that the field of including LLMs into BO loops has not extensively been explored in academia. There is a substantial amount of work in this sector that is never published as a lot of the work is done in an industrial setting. While the authors are clearly open that this is by no means the \"finished article\" it is easy to motivate why there should be a bigger focus in this area.", "weaknesses": "The paper is not very clearly written and its not obvious to actually decipher what the authors propose. There is a lot of intuitive arguments and while this sadly comes with the territory when working with black-box LLMs this could have been done better. The main part which is really unclear to me is how the interaction with the LLM actually works. In the paper it is referred to as \"we assume that the client can query the LLM\". In the case of saying the Ackley function, what is even contextual information, what is it that the LLM actually provides and how and what does the client query?\n\nFurthermore, in the experimental evaluation, what would have been interesting to see is how often the LLMs choices are used. Especially in the \"justify\" setting this is something that could be quantified.\n\nThe experimental setting is somewhat lacklustre, the basic BO examples are not really too interesting and as can be seen from the results a basic BO loop does well on these experiments. What is not clear from the paper is what the actual BO loop is, what is the surrogate model, what is the acquisition etc.", "questions": "- 211 :: am I correct that you use a linear combination of the LLM and the GP proposed locations? Can you explain what motivates this?\n- Please provide information about how the LLM is actually queried what is the structure of this and how is the agent, and what is the agent that does this?\n- Clearly describe the experimental set-up for the baselines, what is the BO loop?\n- What do you mean by contextual information, can you describe what this in the example of optimising the Brain-2D or Ackley-6D? To me it is very unclear what this is making it hard to dechiper what you are proposing.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents three different approaches on how to provide acquisition functions for Bayesian Optimisation with the use of an LLM together with a Gaussian process surrogate model. The motivation for this approach in the paper is the notion that the LLM will somehow contain \"contextual\" information. The first approach uses a linear combination of the LLM acquisition and the GP acquisition, the second uses a rejection criteria based on the GP surrogates acquisition function while the third uses the LLM proposal to create a constrained surrogate.\n\nThe paper ends with a set of experiments comparing the proposed method to several different baselines. The experiments are on classic BO tasks and the more interesting on a 3d printing task.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper is somewhat original given that the field of including LLMs into BO loops has not extensively been explored in academia. There is a substantial amount of work in this sector that is never published as a lot of the work is done in an industrial setting. While the authors are clearly open that this is by no means the \"finished article\" it is easy to motivate why there should be a bigger focus in this area.", "weaknesses": "The paper is not very clearly written and its not obvious to actually decipher what the authors propose. There is a lot of intuitive arguments and while this sadly comes with the territory when working with black-box LLMs this could have been done better. The main part which is really unclear to me is how the interaction with the LLM actually works. In the paper it is referred to as \"we assume that the client can query the LLM\". In the case of saying the Ackley function, what is even contextual information, what is it that the LLM actually provides and how and what does the client query?\n\nFurthermore, in the experimental evaluation, what would have been interesting to see is how often the LLMs choices are used. Especially in the \"justify\" setting this is something that could be quantified.\n\nThe experimental setting is somewhat lacklustre, the basic BO examples are not really too interesting and as can be seen from the results a basic BO loop does well on these experiments. What is not clear from the paper is what the actual BO loop is, what is the surrogate model, what is the acquisition etc.", "questions": "- 211 :: am I correct that you use a linear combination of the LLM and the GP proposed locations? Can you explain what motivates this?\n- Please provide information about how the LLM is actually queried what is the structure of this and how is the agent, and what is the agent that does this?\n- Clearly describe the experimental set-up for the baselines, what is the BO loop?\n- What do you mean by contextual information, can you describe what this in the example of optimising the Brain-2D or Ackley-6D? To me it is very unclear what this is making it hard to dechiper what you are proposing.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761772139366}], "openreview_url": "https://openreview.net/forum?id=39CEoMXTHh", "arxiv_id": "2505.14756", "paper_pdf": "papers/39CEoMXTHh.pdf", "paper_pdf_sha256": "21b890b29c7771600da688fe419c66487b298a65e920a5cbfc01852952ca2c6f", "paper_pdf_bytes": 7566648, "paper_pdf_source": "openreview", "code_url": "https://github.com/UMDataScienceLab/LLM-in-the-Loop-BO", "code_repository": "UMDataScienceLab/LLM-in-the-Loop-BO", "code_commit": "105701dffee2999f31de1c0f42241d0e82b48e55", "code_archive": "repos/39CEoMXTHh.zip", "code_archive_sha256": "c04ddc8e9aa7decee1433f04e6f9294984c5ccee99951688537370a0d9844d92", "code_archive_bytes": 2232443, "code_file_count": 7, "code_extensions": {".py": 4, ".ipynb": 3}, "github_disk_usage_kb": 2226, "github_languages": {"Jupyter Notebook": 139972, "Python": 63501}, "github_archived": false, "github_pushed_at": "2025-05-20T09:06:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/texttt-llinbo-trustworthy-llm-in-the-loop"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tGOOP7DGxs", "year": 2024, "status": "rejected", "title": "Graph Transformers for Large Graphs", "authors": ["Vijay Prakash Dwivedi", "Yozen Liu", "Anh Tuan Luu", "Xavier Bresson", "Neil Shah", "Tong Zhao"], "authorids": ["~Vijay_Prakash_Dwivedi1", "~Yozen_Liu1", "~Anh_Tuan_Luu2", "~Xavier_Bresson6", "~Neil_Shah2", "~Tong_Zhao3"], "authors_source": "OpenReview API", "abstract": "Transformers have recently emerged as powerful neural networks for graph learning, showcasing state-of-the-art performance on several graph property prediction tasks. However, these results have been limited to small-scale graphs, such as ligand molecules with fewer than a hundred atoms, where the computational feasibility of the global attention mechanism is possible. The next goal is to scale up these architectures to handle very large graphs on the scale of millions or even billions of nodes. With large-scale graphs, global attention learning is proven impractical due to its quadratic complexity w.r.t. the number of nodes. On the other hand, neighborhood sampling techniques become essential to manage large graph sizes, yet finding the optimal trade-off between speed and accuracy with sampling techniques remains challenging. This work advances representation learning on single large-scale graphs with a focus on identifying model characteristics and critical design constraints for developing scalable graph transformer (GT) architectures. We argue such GT requires layers that can adeptly learn both local and global graph representations while swiftly sampling the graph topology. As such, a key innovation of this work lies in the creation of a fast neighborhood sampling technique coupled with a local attention mechanism that encompasses a 4-hop reception field, but achieved through just 2-hop operations. This local node embedding is then integrated with a global node embedding, acquired via another self-attention layer with an approximate global codebook, before finally sent through a downstream layer for node predictions. The proposed GT framework, named LargeGT, overcomes previous computational bottlenecks and is validated on three large-scale node classification benchmarks. We report a 3× speedup and 16.8% performance gain on ogbn-products and snap-patents compared to their nearest baselines respectively, while we also scale LargeGT on ogbn-papers100M with a 5.9% improvement in performance.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "WD6BUXM3M8", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3578/Reviewer_dgE1"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes LargeGT for training graph transformers for large graphs. Neighborhood sampling usually samples at most 2-hop neighbors as in GOAT (Kong et al., 2023). The proposed method stores a matrix storing the sum of node features of 1-hop and 2-hop neighbors before training. Then sample 2-hop neighbors for a specific node and get the sum features from the matrix, which is at most 4-hop information for the node. It also adopts GOAT (Kong et al., 2023) as the global module. Experiments show it trains faster than GOAT.", "review_text": "The paper proposes LargeGT for training graph transformers for large graphs. Neighborhood sampling usually samples at most 2-hop neighbors as in GOAT (Kong et al., 2023). The proposed method stores a matrix storing the sum of node features of 1-hop and 2-hop neighbors before training. Then sample 2-hop neighbors for a specific node and get the sum features from the matrix, which is at most 4-hop information for the node. It also adopts GOAT (Kong et al., 2023) as the global module. Experiments show it trains faster than GOAT.", "strengths": "1. The proposed neighbor sampling intuitively improves the model accuracy by getting information at most 4-hop away.\n2. Extensive experiments are performed.\n3. The writing of the proposed method is very clear.", "weaknesses": "1. The mechanism of why LargeGT runs faster than baselines like GOAT is unclear. Since the proposed neighbor sampling has a bigger input matrix than a simple 2-hop neighbor sampling method, does it run longer than the traditional method?\n2. The runtime highly depends on the hyperparameter $K$, which is the number of nodes for sampling. Authors need to provide a fair and solid comparison with the traditional 2-hop neighbor sampling method.\n3. Experiment performances are not explained well (see questions).", "questions": "1. In Table 2, why does GOAT-local-δ have better accuracy in ogbn-products?\n2. For snap-patents in Table 2, why does LargeGT have much better model accuracy than all baselines?\n3. For snap-patents in Table 3, why does the model accuracy drop when $K>50$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes LargeGT for training graph transformers for large graphs. Neighborhood sampling usually samples at most 2-hop neighbors as in GOAT (Kong et al., 2023). The proposed method stores a matrix storing the sum of node features of 1-hop and 2-hop neighbors before training. Then sample 2-hop neighbors for a specific node and get the sum features from the matrix, which is at most 4-hop information for the node. It also adopts GOAT (Kong et al., 2023) as the global module. Experiments show it trains faster than GOAT.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The proposed neighbor sampling intuitively improves the model accuracy by getting information at most 4-hop away.\n2. Extensive experiments are performed.\n3. The writing of the proposed method is very clear.", "weaknesses": "1. The mechanism of why LargeGT runs faster than baselines like GOAT is unclear. Since the proposed neighbor sampling has a bigger input matrix than a simple 2-hop neighbor sampling method, does it run longer than the traditional method?\n2. The runtime highly depends on the hyperparameter $K$, which is the number of nodes for sampling. Authors need to provide a fair and solid comparison with the traditional 2-hop neighbor sampling method.\n3. Experiment performances are not explained well (see questions).", "questions": "1. In Table 2, why does GOAT-local-δ have better accuracy in ogbn-products?\n2. For snap-patents in Table 2, why does LargeGT have much better model accuracy than all baselines?\n3. For snap-patents in Table 3, why does the model accuracy drop when $K>50$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698865571534}, {"id": "0YD6wKInMx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3578/Reviewer_xiXc"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work proposes LargeGT, a scalable graph transformer for large-scale graphs. It uses fast neighborhood sampling and a local attention mechanism to learn local representations. These are integrated with global representations from an approximate global codebook. This framework overcomes previous computational bottlenecks, achieving 3x speedup and 16.8% better performance on benchmarks compared to baselines. LargeGT also scales to 100M nodes, advancing representation learning for single large graphs.", "review_text": "This work proposes LargeGT, a scalable graph transformer for large-scale graphs. It uses fast neighborhood sampling and a local attention mechanism to learn local representations. These are integrated with global representations from an approximate global codebook. This framework overcomes previous computational bottlenecks, achieving 3x speedup and 16.8% better performance on benchmarks compared to baselines. LargeGT also scales to 100M nodes, advancing representation learning for single large graphs.", "strengths": "* The model's performance is thoroughly validated on large-scale graphs, demonstrating sufficient workload.\n* Exploring base model architectures on graphs is a very valuable endeavor.", "weaknesses": "* The efficiency analysis is incorrect. In Algorithm 1, it is required to gather 1/2-degree neighbors for each node, and then select k nodes. The process of selecting nodes is O(K), but if the graph is relatively dense, the complexity of gathering second-degree neighbors is O(N^2).\n* In Algorithm 1, some nodes are sampled with replacement, while some are sampled without replacement. It is uncertain whether this will introduce bias in the sampling.\n* It lacks some key baselines such as SGC[1], SIGN[2]. \n\nReference:\n\n[1] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. \"Simplifying graph convolutional networks.\" In International conference on machine learning, pp. 6861-6871. PMLR, 2019.\n\n[2] Fabrizio Frasca, Emanuele Rossi, Davide Eynard, Ben Chamberlain, Michael Bronstein, and Federico Monti. \"Sign: Scalable inception graph neural networks.\" arXiv preprint arXiv:2004.11198 (2020).", "questions": "See. Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes LargeGT, a scalable graph transformer for large-scale graphs. It uses fast neighborhood sampling and a local attention mechanism to learn local representations. These are integrated with global representations from an approximate global codebook. This framework overcomes previous computational bottlenecks, achieving 3x speedup and 16.8% better performance on benchmarks compared to baselines. LargeGT also scales to 100M nodes, advancing representation learning for single large graphs.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "* The model's performance is thoroughly validated on large-scale graphs, demonstrating sufficient workload.\n* Exploring base model architectures on graphs is a very valuable endeavor.", "weaknesses": "* The efficiency analysis is incorrect. In Algorithm 1, it is required to gather 1/2-degree neighbors for each node, and then select k nodes. The process of selecting nodes is O(K), but if the graph is relatively dense, the complexity of gathering second-degree neighbors is O(N^2).\n* In Algorithm 1, some nodes are sampled with replacement, while some are sampled without replacement. It is uncertain whether this will introduce bias in the sampling.\n* It lacks some key baselines such as SGC[1], SIGN[2]. \n\nReference:\n\n[1] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. \"Simplifying graph convolutional networks.\" In International conference on machine learning, pp. 6861-6871. PMLR, 2019.\n\n[2] Fabrizio Frasca, Emanuele Rossi, Davide Eynard, Ben Chamberlain, Michael Bronstein, and Federico Monti. \"Sign: Scalable inception graph neural networks.\" arXiv preprint arXiv:2004.11198 (2020).", "questions": "See. Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698840085358}, {"id": "V6nC0Qj2Gi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3578/Reviewer_S1aR"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "To scale graph models to large-scale graphs, MPNNs are often reduced to restricted receptive fields making them myopic, while Graph Transformers (GTs) fail because of their quadratic cost. This paper proposes a new framework for sampling sub-graphs to train a large GT that uses local and global modules to improve model performance and compute complexity.", "review_text": "To scale graph models to large-scale graphs, MPNNs are often reduced to restricted receptive fields making them myopic, while Graph Transformers (GTs) fail because of their quadratic cost. This paper proposes a new framework for sampling sub-graphs to train a large GT that uses local and global modules to improve model performance and compute complexity.", "strengths": "- The authors propose a framework that leverage recent advances in graph transformer models, and address a critical challenge that limits the scalability of existing approaches, both MPNNs and GTs.  \n- The introduction provides a great overview of the current challenges for large-scale graph learning, and does a great job at comparing MPNNs and GTs, while setting stage for key concepts like neighborhood sampling.", "weaknesses": "1. Baselines: LargeGT is compared to \"constrained versions\" of various baselines, notably all models are constrained to 2 hops only, while LargeGT has access to 4-hops worth of neighbors (in the local module). Including the non-constrained versions of these same baselines is critical for evaluation, even if they are more computationally demanding. Currently it is unclear whether adopting LargeGT leads to lower performance compared to state-of-the-art methods, at the expense of computational efficiency.\n2. Additionally, no auxiliary label propagation or augmentations are used for the baseline methods, when they are used in methods reported in the OGB leaderboard. These enhancements are not altering the receptive field of the baselines, and thus shouldn't impact computational performance, but might improve classification performance. This should be taken into account when comparing with approaches that might still outperform the proposed method, even under constrained training (2-hop). \n3. The main innovation can seemingly be credited to the use of the global codebook, so it is hard to define the main contribution of this work. If the focus of this work is combining all these different building blocks into a compute efficient framework, I would expect to see a more expansive breakdown of the computational costs of different components, memory usage and requirements. Notably, how is \"Epoch time\" defined in Figure 2? All models might be processing different amounts of data and thus might have different definitions of an \"epoch\" due to differences in sampling strategies. How many nodes does each model process in an epoch? Different models might require different numbers of epochs to converge, shouldn't total training time be more important?\n4. [Minor] A lot of the content in the first 4 pages is repetitive.", "questions": "1. What are the memory constraints of using LargeGT compared to other baselines? How is the choice of batch size impacted by the choice of hyperparameter K? \n2. How important is the choice of a 4-hop neighborhood for the local module. Can the model still perform competitively given that it still has access to global information through the global module?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To scale graph models to large-scale graphs, MPNNs are often reduced to restricted receptive fields making them myopic, while Graph Transformers (GTs) fail because of their quadratic cost. This paper proposes a new framework for sampling sub-graphs to train a large GT that uses local and global modules to improve model performance and compute complexity.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The authors propose a framework that leverage recent advances in graph transformer models, and address a critical challenge that limits the scalability of existing approaches, both MPNNs and GTs.  \n- The introduction provides a great overview of the current challenges for large-scale graph learning, and does a great job at comparing MPNNs and GTs, while setting stage for key concepts like neighborhood sampling.", "weaknesses": "1. Baselines: LargeGT is compared to \"constrained versions\" of various baselines, notably all models are constrained to 2 hops only, while LargeGT has access to 4-hops worth of neighbors (in the local module). Including the non-constrained versions of these same baselines is critical for evaluation, even if they are more computationally demanding. Currently it is unclear whether adopting LargeGT leads to lower performance compared to state-of-the-art methods, at the expense of computational efficiency.\n2. Additionally, no auxiliary label propagation or augmentations are used for the baseline methods, when they are used in methods reported in the OGB leaderboard. These enhancements are not altering the receptive field of the baselines, and thus shouldn't impact computational performance, but might improve classification performance. This should be taken into account when comparing with approaches that might still outperform the proposed method, even under constrained training (2-hop). \n3. The main innovation can seemingly be credited to the use of the global codebook, so it is hard to define the main contribution of this work. If the focus of this work is combining all these different building blocks into a compute efficient framework, I would expect to see a more expansive breakdown of the computational costs of different components, memory usage and requirements. Notably, how is \"Epoch time\" defined in Figure 2? All models might be processing different amounts of data and thus might have different definitions of an \"epoch\" due to differences in sampling strategies. How many nodes does each model process in an epoch? Different models might require different numbers of epochs to converge, shouldn't total training time be more important?\n4. [Minor] A lot of the content in the first 4 pages is repetitive.", "questions": "1. What are the memory constraints of using LargeGT compared to other baselines? How is the choice of batch size impacted by the choice of hyperparameter K? \n2. How important is the choice of a 4-hop neighborhood for the local module. Can the model still perform competitively given that it still has access to global information through the global module?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698808686594}, {"id": "m5TAWWYVyd", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3578/Reviewer_yzrb"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The author highlights that while transformers have demonstrated remarkable performance in tasks related to predicting graph properties, their application has been restricted to small-scale graphs due to computational limitations. Additionally, the author contends that the existing neighbor sampling method constrains the model's ability to consider more global information. Consequently, this paper introduces a comprehensive GT framework, with a focus on enhancing model capacity and scalability. The proposed framework, known as LargeGT, combines a rapid neighborhood sampling technique with a local attention mechanism and an approximate global codebook. Extensive experiments illustrate that by integrating local and global attention mechanisms, LargeGT achieves improved performance in node classification tasks. Notably, LargeGT demonstrates a 3× speedup and a 16.8% performance enhancement in specific node classification benchmarks when compared to their closest baseline models.", "review_text": "The author highlights that while transformers have demonstrated remarkable performance in tasks related to predicting graph properties, their application has been restricted to small-scale graphs due to computational limitations. Additionally, the author contends that the existing neighbor sampling method constrains the model's ability to consider more global information. Consequently, this paper introduces a comprehensive GT framework, with a focus on enhancing model capacity and scalability. The proposed framework, known as LargeGT, combines a rapid neighborhood sampling technique with a local attention mechanism and an approximate global codebook. Extensive experiments illustrate that by integrating local and global attention mechanisms, LargeGT achieves improved performance in node classification tasks. Notably, LargeGT demonstrates a 3× speedup and a 16.8% performance enhancement in specific node classification benchmarks when compared to their closest baseline models.", "strengths": "1. This paper is excellently composed, offering a straightforward narrative that's easy to follow. Notably, key terms and important experimental findings have been highlighted using various colors, resulting in an effective visual presentation.\n\n2. The experimental results presented in this paper indicate that the proposed framework can achieve superior performance within a shorter training time.\n\n3. The author introduces two significant challenges associated with handling large-scale graphs: scalability and constraints related to local information aggregation. These issues are prevalent and indeed worth discussing. As the author pointed out, computational resource requirements increase quadratically with the growing number of nodes. To address this, the author has proposed both a local and a global aggregation module. The former employs conventional sampling techniques to learn local representations, while the latter focuses on deriving insights from global node vector projections. Downstream predictions are then made based on both sets of representations. The problems raised, and the respective solutions are meaningful and coherent with each other.", "weaknesses": "Despite the fluent presentation, some concerns arise in this paper. Firstly, the level of novelty in this framework appears limited. It is apparent that this paper heavily relies on the previous work, GOAT, particularly the global module, which encodes mini-batch nodes using global graph nodes. This component was introduced in a prior paper. The other aspects are mainly focused on aligning local and global features. The framework appears more like an updated version of GOAT than a fundamentally new invention.\n\nMoreover, in the experimental section, the comparison between the LG transformer and other baselines reveals that the proposed framework doesn't consistently outperform GOAT-local, especially in the ogbn-products dataset. Furthermore, in the ogbn-papers100M dataset, the framework is only compared to a single baseline. It's possible that other methods struggle with extremely large graphs, but there are likely additional viable solutions that should be explored.\n\nAdditionally, the fusion of transformers and Graph Neural Networks (GNNs) is a dynamic research area with various ongoing studies, such as TransGNN and Graphformers. It would be valuable to understand how these methods perform when confronted with similar tasks.\n\nLastly, the author emphasizes the significance of combining local and global representations. However, apart from GOAT, there are other techniques that can address this challenge, such as randomly selecting both nearby neighbors and global features. The author should offer further clarification on this matter.", "questions": "This paper commences with two important challenges that have attracted the attention of numerous researchers. Specific comments were provided in the previous section, and it is hoped that the author will consider improvements from the following viewpoints.\n\nThe framework appears to inherit many key components from previous papers, with limited significant modifications. It would be beneficial to include more in-depth discussions and comparisons with transformer-based Graph Neural Networks (GNNs). Additionally, it is important to address how other approaches perform in terms of extracting global information from the graph.\n\nExpanding on these aspects would enhance the paper's contribution and provide a more comprehensive understanding of the research landscape in this domain.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The author highlights that while transformers have demonstrated remarkable performance in tasks related to predicting graph properties, their application has been restricted to small-scale graphs due to computational limitations. Additionally, the author contends that the existing neighbor sampling method constrains the model's ability to consider more global information. Consequently, this paper introduces a comprehensive GT framework, with a focus on enhancing model capacity and scalability. The proposed framework, known as LargeGT, combines a rapid neighborhood sampling technique with a local attention mechanism and an approximate global codebook. Extensive experiments illustrate that by integrating local and global attention mechanisms, LargeGT achieves improved performance in node classification tasks. Notably, LargeGT demonstrates a 3× speedup and a 16.8% performance enhancement in specific node classification benchmarks when compared to their closest baseline models.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper is excellently composed, offering a straightforward narrative that's easy to follow. Notably, key terms and important experimental findings have been highlighted using various colors, resulting in an effective visual presentation.\n\n2. The experimental results presented in this paper indicate that the proposed framework can achieve superior performance within a shorter training time.\n\n3. The author introduces two significant challenges associated with handling large-scale graphs: scalability and constraints related to local information aggregation. These issues are prevalent and indeed worth discussing. As the author pointed out, computational resource requirements increase quadratically with the growing number of nodes. To address this, the author has proposed both a local and a global aggregation module. The former employs conventional sampling techniques to learn local representations, while the latter focuses on deriving insights from global node vector projections. Downstream predictions are then made based on both sets of representations. The problems raised, and the respective solutions are meaningful and coherent with each other.", "weaknesses": "Despite the fluent presentation, some concerns arise in this paper. Firstly, the level of novelty in this framework appears limited. It is apparent that this paper heavily relies on the previous work, GOAT, particularly the global module, which encodes mini-batch nodes using global graph nodes. This component was introduced in a prior paper. The other aspects are mainly focused on aligning local and global features. The framework appears more like an updated version of GOAT than a fundamentally new invention.\n\nMoreover, in the experimental section, the comparison between the LG transformer and other baselines reveals that the proposed framework doesn't consistently outperform GOAT-local, especially in the ogbn-products dataset. Furthermore, in the ogbn-papers100M dataset, the framework is only compared to a single baseline. It's possible that other methods struggle with extremely large graphs, but there are likely additional viable solutions that should be explored.\n\nAdditionally, the fusion of transformers and Graph Neural Networks (GNNs) is a dynamic research area with various ongoing studies, such as TransGNN and Graphformers. It would be valuable to understand how these methods perform when confronted with similar tasks.\n\nLastly, the author emphasizes the significance of combining local and global representations. However, apart from GOAT, there are other techniques that can address this challenge, such as randomly selecting both nearby neighbors and global features. The author should offer further clarification on this matter.", "questions": "This paper commences with two important challenges that have attracted the attention of numerous researchers. Specific comments were provided in the previous section, and it is hoped that the author will consider improvements from the following viewpoints.\n\nThe framework appears to inherit many key components from previous papers, with limited significant modifications. It would be beneficial to include more in-depth discussions and comparisons with transformer-based Graph Neural Networks (GNNs). Additionally, it is important to address how other approaches perform in terms of extracting global information from the graph.\n\nExpanding on these aspects would enhance the paper's contribution and provide a more comprehensive understanding of the research landscape in this domain.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698676916556}], "openreview_url": "https://openreview.net/forum?id=tGOOP7DGxs", "arxiv_id": "2312.11109", "paper_pdf": "papers/tGOOP7DGxs.pdf", "paper_pdf_sha256": "fbe21b0dd6e556753be60eb90bdb3dccbe2ec8b54bd9fba09e1ddeb27f2671a1", "paper_pdf_bytes": 708610, "paper_pdf_source": "openreview", "code_url": "https://github.com/snap-research/LargeGT", "code_repository": "snap-research/LargeGT", "code_commit": "532326966ef25f2f28c25d034ef06c1c3b9a141a", "code_archive": "repos/tGOOP7DGxs.zip", "code_archive_sha256": "90a2c848c0003f24cf7ce5a85444bc6d1819bb3d6dc0819a99b3a18ca6e92a45", "code_archive_bytes": 103115, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 135, "github_languages": {"Python": 55366, "Shell": 4914}, "github_archived": false, "github_pushed_at": "2024-04-26T22:28:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-transformers-for-large-graphs"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_DYi95e8CAe", "year": 2023, "status": "rejected", "title": "Multi-Task Structural Learning using Local Task Similarity induced Neuron Creation and Removal", "authors": ["Naresh Kumar Gurulingan", "Elahe Arani", "Bahram Zonooz"], "authorids": ["~Naresh_Kumar_Gurulingan1", "~Elahe_Arani1", "~Bahram_Zonooz1"], "authors_source": "OpenReview API", "abstract": "Multi-task learning has the potential to improve generalization by maximizing positive transfer between tasks while reducing task interference. Fully achieving this potential is hindered by manually designed architectures that remain static throughout training. In contrast, learning in the brain occurs through structural changes that are in tandem with changes in synaptic strength. Therefore, we propose Multi-Task Structural Learning (MTSL) which simultaneously learns the multi-task architecture and its parameters. MTSL begins with an identical single task network for each task and alternates between a task learning phase and a structural learning phase. In the task learning phase, each network specializes in the corresponding task. In each of the structural learning phases, starting from the earliest layer, locally similar task layers first transfer their knowledge to a newly created group layer after which they become redundant and are removed. Our experimental results show that MTSL achieves competitive generalization with various baselines and improves robustness to out-of-distribution data.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "fd6LGFrYkL", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4611/Reviewer_mKd4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a structural learning algorithm for concurrently learning the multi-task learning architecture and its parameters. The multitask learning is performed by the creation and removal of neurons based on local similarity. The proposed method is validated on the Cityscapes and NYUv2 datasets for five different dense prediction tasks. Experimental results include SOTA comparison, generalization and inference efficiency, robustness to natural corruptions, and ablation study.", "review_text": "I found the idea of multi-task structural learning with task similarity and neuron creation/removal quite interesting. The authors presented results from a number of experiments for validating their proposed method. Although the results are not super convincing, I believe there is the true potential of MTSL in learning similar tasks within a single model.", "strengths": "Pros:\n+  The paper is well-written and it’s quite well-organized. It was a pleasure reading it.\n+ Task-relatedness is extremely crucial for multi-task learning. Grouping tasks based on local task representations and transferring knowledge to a new group layer is very interesting.\n+  Exposition of the results is good; particularly the ablation study reporting the effect of alignment, average initialization, and attention based knowledge amalgamation.\n\nCons:\n- The generalization experiment reported in 4.2 is not quite clear to me. Did the authors perform cross-dataset evaluation (trained on CS, tested on NYU, and vice-versa) or cross-task evaluation?\n- How would the MTSL work if there are one or multiple tasks requiring varying architectures (e.g., classification from the encoder only and segmentation from the encoder-decoder )?\n- What's the reason behind choosing the particular encoder and decoder networks? Why the decoder is changed from DeepLabv3 to ResNet blocks?   \n-  The authors mentioned that the multitask loss is a weighted sum of all individual task losses. It should be clarified how the task losses are weighted. Which task is to prioritize and when?\n- What value did the authors assign to the balancing factor \\lambda in eq.(1)?\n- Some of the terms are not defined such as, GMac, etc. Table captions could be revised clarifying all the terms. All the performance metrics should be discussed before using them in the tables.\n- it looks like MTSL is consistently poorer than One-Net in the case of noise corruption in Table 3. Why so?\n- Automatic transitioning between the learning phases would be more appropriate as the model is supposed to learn the architecture as well as its parameters automatically.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presents a structural learning algorithm for concurrently learning the multi-task learning architecture and its parameters. The multitask learning is performed by the creation and removal of neurons based on local similarity. The proposed method is validated on the Cityscapes and NYUv2 datasets for five different dense prediction tasks. Experimental results include SOTA comparison, generalization and inference efficiency, robustness to natural corruptions, and ablation study.", "strength_and_weaknesses": "Pros:\n+  The paper is well-written and it’s quite well-organized. It was a pleasure reading it.\n+ Task-relatedness is extremely crucial for multi-task learning. Grouping tasks based on local task representations and transferring knowledge to a new group layer is very interesting.\n+  Exposition of the results is good; particularly the ablation study reporting the effect of alignment, average initialization, and attention based knowledge amalgamation.\n\nCons:\n- The generalization experiment reported in 4.2 is not quite clear to me. Did the authors perform cross-dataset evaluation (trained on CS, tested on NYU, and vice-versa) or cross-task evaluation?\n- How would the MTSL work if there are one or multiple tasks requiring varying architectures (e.g., classification from the encoder only and segmentation from the encoder-decoder )?\n- What's the reason behind choosing the particular encoder and decoder networks? Why the decoder is changed from DeepLabv3 to ResNet blocks?   \n-  The authors mentioned that the multitask loss is a weighted sum of all individual task losses. It should be clarified how the task losses are weighted. Which task is to prioritize and when?\n- What value did the authors assign to the balancing factor \\lambda in eq.(1)?\n- Some of the terms are not defined such as, GMac, etc. Table captions could be revised clarifying all the terms. All the performance metrics should be discussed before using them in the tables.\n- it looks like MTSL is consistently poorer than One-Net in the case of noise corruption in Table 3. Why so?\n- Automatic transitioning between the learning phases would be more appropriate as the model is supposed to learn the architecture as well as its parameters automatically.", "clarity,_quality,_novelty_and_reproducibility": "The paper is fairly written and well-organized.\nThe work seems to be original.\nThe authors indicated that they would make their code available upon acceptance. Except few minor training details, there should not be any issue of reproducibility. ", "summary_of_the_review": "I found the idea of multi-task structural learning with task similarity and neuron creation/removal quite interesting. The authors presented results from a number of experiments for validating their proposed method. Although the results are not super convincing, I believe there is the true potential of MTSL in learning similar tasks within a single model.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1668010178232}, {"id": "9jJw1xtuwq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4611/Reviewer_2dhv"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present a method that \"merges\" N single task networks by grouping similar neurons. The training process\nalternates between \"task learning\" during which the network weights are optimized for task performance, and \"structure\nlearning\", during which neurons are group nodes are optimized with a knowledge distillation-like approach.\nThe method is evaluated on two five-task datasets (cityscapes, NYUv2) and compared with four state-of-the-art methods, with\nfavorable results. Generalization and robustness to corruptions are compared to a baseline single-network.", "review_text": "This paper presents a method that performs well, as evidenced by solid experiments, but suffers from an incomplete / unclear description of the methodology. I am unsure regarding the novelty of this work.", "strengths": "The proposed work describes an effective methodology for combining N single-task networks into a single, branched network\nthat can perform well at multiple tasks. This enables trading-off between task performance and computational efficiency. The\nexperiments appear to be well done, and the authors include an informative ablation study. The results make me confident that\nthe method can perform well and as described.\n\nThe main weakness of the paper is the lack of clarity in parts when describing the method - I give some detailed examples in the\nnext section. \n\nSome minor weaknesses:\nThe paper would be improved if the authors interpret the results for the reader more. For example, the text only very briefly\ndescribes the differences between LTB-R and LTB, which is an important and interesting one. As well - why does cross-stich\nperform so well, and how much has to do with the number of parameters?\n\nI find the analogy with biological neural circuits to be unconvincing, but more concerning was the statement:\n\"The improvement generalization can be attributed to the brain-inspired aspects of the MTSL algorithm.\" I am very\nskeptical of this claim, and couldn't find  evidence in the work about WHY MTSL's performance is good. \nMore care in separating speculation (which is fine when framed as such) from conclusions (and \"can be attributed to\" sounds\nconclusive) would be appropriate.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors present a method that \"merges\" N single task networks by grouping similar neurons. The training process\nalternates between \"task learning\" during which the network weights are optimized for task performance, and \"structure\nlearning\", during which neurons are group nodes are optimized with a knowledge distillation-like approach.\nThe method is evaluated on two five-task datasets (cityscapes, NYUv2) and compared with four state-of-the-art methods, with\nfavorable results. Generalization and robustness to corruptions are compared to a baseline single-network.", "strength_and_weaknesses": "The proposed work describes an effective methodology for combining N single-task networks into a single, branched network\nthat can perform well at multiple tasks. This enables trading-off between task performance and computational efficiency. The\nexperiments appear to be well done, and the authors include an informative ablation study. The results make me confident that\nthe method can perform well and as described.\n\nThe main weakness of the paper is the lack of clarity in parts when describing the method - I give some detailed examples in the\nnext section. \n\nSome minor weaknesses:\nThe paper would be improved if the authors interpret the results for the reader more. For example, the text only very briefly\ndescribes the differences between LTB-R and LTB, which is an important and interesting one. As well - why does cross-stich\nperform so well, and how much has to do with the number of parameters?\n\nI find the analogy with biological neural circuits to be unconvincing, but more concerning was the statement:\n\"The improvement generalization can be attributed to the brain-inspired aspects of the MTSL algorithm.\" I am very\nskeptical of this claim, and couldn't find  evidence in the work about WHY MTSL's performance is good. \nMore care in separating speculation (which is fine when framed as such) from conclusions (and \"can be attributed to\" sounds\nconclusive) would be appropriate.", "clarity,_quality,_novelty_and_reproducibility": "The work appears to be methodologically sound. Experiments are effective overall, and the method \nperforms favorably to the methods compared. The method combines some existing approaches in a new way, so appears moderately\nnovel, though I am not confident in that assessment. The writing is good overall, but there are some methodological\ndetails that are are not clear and seem not sufficiently explained. Without available code, I am not confident a researcher\ncould reproduce the method. I have included some questions below that the authors might consider clarifying in the text and I hope might improve the manuscript.\n\nFrom the abstract \"In each of the structural learning phases, starting from the earliest layer, ...\", but its not clear to \nme from the text when the method \"moves on\" from the earliest to layer layers.\n\nHow does the method determine when to \"stop\" merging task nodes?  The grouping threshold gamma will clearly play a role.\nI appreciate section D.2 describing the authors' experiences varying gamma.\n\nIt is unclear how or why in Figure 2, the first 6 layers are shared between the two tasks. Is it possible that MTSL could have\ndetermined that one task shares no information / nodes with the other four tasks?  Is it always the case that early layers are\nshared and later layers are split? One could argue that properties of natural images make that likely, but I wonder if it is\npossible for the algorithm to produce a graph that splits and merges.\n\nTables [1,2,3] What is the difference between \\Delta^SD_MTL and \\Delta_MTL? Does SD use only the S and D tasks?\n\nEq 1 is not very useful, the authors might consider adding definitions for L_MTL and L_CKA for completeness.\n\nEq (2) What does the 'star F' notation refer to? I could not find similar notation in the associated reference Ye 2019.\n\n\"locally similar task layers\" what is \"locality\" with respect to here?", "summary_of_the_review": "This paper presents a method that performs well, as evidenced by solid experiments, but suffers from an incomplete / unclear description of the methodology. I am unsure regarding the novelty of this work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667611027218}, {"id": "GP31MinXURz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4611/Reviewer_hhES"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper presents Multi-Task Structural Learning (MTSL) framework to learn a network architecture and its parameters in a multi-task learning setting. MTSL consists of two phases, a structural learning phase and a fine-tuning phase. The former includes repetition of two alternating steps, network training followed by task grouping. Two datasets in computer vision: Cityscape and NYUv2 were used to evaluate the proposed method. ", "review_text": "The method has conceptual merit and is innovative. However, the weaknesses in the implementation and their empirical results dampen my enthusiasm of the work.", "strengths": "Strength\n+ The development of the method tried to mimic how human brains works. It may be conceptual advantage over existing ones from the artificial intelligence perspective. \n+ Large efforts made to empirical study from multiple aspect, including robustness and contribution of different components of the method.\n\nWeakness\n- The calculation of CKA only directly involves kernel, which is the pairwise similarity between examples based on activation. I do not see this helps align learned filters in two separate networks. \n- The calculation of CKA is quite complicated. It may be challenging to effectively optimize the overall objective in Eq. (1).\n- CKA determines the similarity among tasks. Its calculation should be detailed, for example the (number) examples involved.\n- After the first round of structural learning, not all task nodes are at the same layer of the network. I do not think the way that they described for grouping is appropriate. That way likely leads to a very complicated grouping structure (not like the one that they show in Figure 2, which has nice tree structure) that is difficult to interpret, making little to no sense at all. \n- It would be interesting to report the task grouping/branching learned by other methods and have a comparison with that by the proposed method.\n- The empirical results (Table 1) are weak and not much supportive to the value/advantage of the proposed method.\n- The manuscript is readable on overall, but with large room to improve. There are inappropriate choices of wording in many places.  \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents Multi-Task Structural Learning (MTSL) framework to learn a network architecture and its parameters in a multi-task learning setting. MTSL consists of two phases, a structural learning phase and a fine-tuning phase. The former includes repetition of two alternating steps, network training followed by task grouping. Two datasets in computer vision: Cityscape and NYUv2 were used to evaluate the proposed method. ", "strength_and_weaknesses": "Strength\n+ The development of the method tried to mimic how human brains works. It may be conceptual advantage over existing ones from the artificial intelligence perspective. \n+ Large efforts made to empirical study from multiple aspect, including robustness and contribution of different components of the method.\n\nWeakness\n- The calculation of CKA only directly involves kernel, which is the pairwise similarity between examples based on activation. I do not see this helps align learned filters in two separate networks. \n- The calculation of CKA is quite complicated. It may be challenging to effectively optimize the overall objective in Eq. (1).\n- CKA determines the similarity among tasks. Its calculation should be detailed, for example the (number) examples involved.\n- After the first round of structural learning, not all task nodes are at the same layer of the network. I do not think the way that they described for grouping is appropriate. That way likely leads to a very complicated grouping structure (not like the one that they show in Figure 2, which has nice tree structure) that is difficult to interpret, making little to no sense at all. \n- It would be interesting to report the task grouping/branching learned by other methods and have a comparison with that by the proposed method.\n- The empirical results (Table 1) are weak and not much supportive to the value/advantage of the proposed method.\n- The manuscript is readable on overall, but with large room to improve. There are inappropriate choices of wording in many places.  \n", "clarity,_quality,_novelty_and_reproducibility": "The presentation of the paper needs improvement. Novelty is acceptable. No code is provided. Based on the description in the paper, reproducibility is questionable. ", "summary_of_the_review": "The method has conceptual merit and is innovative. However, the weaknesses in the implementation and their empirical results dampen my enthusiasm of the work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666846929410}, {"id": "bPPwqB3PjY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4611/Reviewer_wjTd"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a network architecture search for a multi-task network, while specifying the search space to tree structures. It jointly trains the model structure as well as the parameters. It starts from the set of single networks and gradually groups layers across tasks if their output features are sufficiently similar. Experiments show that the proposed method outperforms the existing multi-task NAS approaches (BMTAS, LTB) without re-training.", "review_text": "In summary, I think this paper proposes a new method for multi-task NAS. However, I could not find an advantage over other existing multi-task NAS approaches in the method and the experiments.", "strengths": "Strength\n\n- It proposes a novel method for multi-task NAS while constraining the search space to the tree-structured architectures, similarly to BMTAS and LTB. Compared to BMTAS and NAS, it utilized CKA to guide task streams to be similar to each other and compute similarity between layers, which is reasonable.\n- The paper provides an extensive analysis of the sensitivity of several factors, including similarity measures and hyper-parameters. It is also good that the performance under several corruptions is evaluated.\n\nWeakness\n\n- I am not sure about the advantage of this method over other existing methods, BMTAS and LTB.\nGiven the same search space of tree structures, I would like to see the accuracy comparison with BMTAS and LTB after re-initialization and re-training to evaluate the optimality of the found architecture itself. The paper provides accuracy after re-training in Table 8, but a comparison with BMTAS and LTB is unavailable.\n\n- I am not convinced about the value of simultaneous training of the architecture and its parameters, which is claimed as one contribution of this work. Since this method requires training of single-task models for all tasks for initialization, additional one-time re-intialization and re-training of the found architecture will not be a huge burden, roughly spending a similar training cost for one single task network training. For this reason, I also think the comparison with existing methods (BMTAS, LTB) would more reasonable after the re-training.\n\n- The method needs several hyper-parameters, including the task learning epochs, structure learning epochs. How will they affect the performance? Sensitivity analysis seems missing.\n\n- Scalability can be another issue of this method. It needs to initialize the model with N single-task models for N tasks, needs N^2 pairwise similarity, and find optimal groupings during training.\n\n- In Table 1, for fair comparison, BMTAS can be trained with larger lambda (resource regularizer) to match the similar compute cost.\n- The details of the evaluation setting of Table 2 is unclear. Is it a cross-dataset accuracy, e.g., training on NYUv2 and testing on CityScapes?\n- It would be nice to include the comparison with BMTAS and LTB (after re-training) in Table 3, to compare the robustness against the corruptions.\n\n\nRelated work\n\n- [Controllable Dynamic Multi-Task Architectures, CVPR 2022] looks pretty related in the sense that it also uses task similarity to guide branching. It also searches for the tree-structured multi-task architecture but also allows dynamic control of the total compute cost.\n\n\nMinor\n\n- The purpose of explicitly evaluating the multi-task performance for semantic segmentation and depth (SD) is unclear to me.\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a network architecture search for a multi-task network, while specifying the search space to tree structures. It jointly trains the model structure as well as the parameters. It starts from the set of single networks and gradually groups layers across tasks if their output features are sufficiently similar. Experiments show that the proposed method outperforms the existing multi-task NAS approaches (BMTAS, LTB) without re-training.", "strength_and_weaknesses": "Strength\n\n- It proposes a novel method for multi-task NAS while constraining the search space to the tree-structured architectures, similarly to BMTAS and LTB. Compared to BMTAS and NAS, it utilized CKA to guide task streams to be similar to each other and compute similarity between layers, which is reasonable.\n- The paper provides an extensive analysis of the sensitivity of several factors, including similarity measures and hyper-parameters. It is also good that the performance under several corruptions is evaluated.\n\nWeakness\n\n- I am not sure about the advantage of this method over other existing methods, BMTAS and LTB.\nGiven the same search space of tree structures, I would like to see the accuracy comparison with BMTAS and LTB after re-initialization and re-training to evaluate the optimality of the found architecture itself. The paper provides accuracy after re-training in Table 8, but a comparison with BMTAS and LTB is unavailable.\n\n- I am not convinced about the value of simultaneous training of the architecture and its parameters, which is claimed as one contribution of this work. Since this method requires training of single-task models for all tasks for initialization, additional one-time re-intialization and re-training of the found architecture will not be a huge burden, roughly spending a similar training cost for one single task network training. For this reason, I also think the comparison with existing methods (BMTAS, LTB) would more reasonable after the re-training.\n\n- The method needs several hyper-parameters, including the task learning epochs, structure learning epochs. How will they affect the performance? Sensitivity analysis seems missing.\n\n- Scalability can be another issue of this method. It needs to initialize the model with N single-task models for N tasks, needs N^2 pairwise similarity, and find optimal groupings during training.\n\n- In Table 1, for fair comparison, BMTAS can be trained with larger lambda (resource regularizer) to match the similar compute cost.\n- The details of the evaluation setting of Table 2 is unclear. Is it a cross-dataset accuracy, e.g., training on NYUv2 and testing on CityScapes?\n- It would be nice to include the comparison with BMTAS and LTB (after re-training) in Table 3, to compare the robustness against the corruptions.\n\n\nRelated work\n\n- [Controllable Dynamic Multi-Task Architectures, CVPR 2022] looks pretty related in the sense that it also uses task similarity to guide branching. It also searches for the tree-structured multi-task architecture but also allows dynamic control of the total compute cost.\n\n\nMinor\n\n- The purpose of explicitly evaluating the multi-task performance for semantic segmentation and depth (SD) is unclear to me.\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper proposes a novel method and it is clearly written.", "summary_of_the_review": "In summary, I think this paper proposes a new method for multi-task NAS. However, I could not find an advantage over other existing multi-task NAS approaches in the method and the experiments.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666649812404}], "openreview_url": "https://openreview.net/forum?id=_DYi95e8CAe", "arxiv_id": "2305.00441", "paper_pdf": "papers/_DYi95e8CAe.pdf", "paper_pdf_sha256": "fea0f81c436a996951924dce32dc69dc654ae7e90ff4c4eff83123ed0473778f", "paper_pdf_bytes": 528478, "paper_pdf_source": "openreview", "code_url": "https://github.com/NeurAI-Lab/MTSL", "code_repository": "NeurAI-Lab/MTSL", "code_commit": "28db1ac294052062dc59d48762f94724146f8e51", "code_archive": "repos/_DYi95e8CAe.zip", "code_archive_sha256": "91b8e4705c1229b69444fc5cc0a4a9f9ba11a5fbea8a2897fe25d88e6d358824", "code_archive_bytes": 248732, "code_file_count": 56, "code_extensions": {".py": 56}, "github_disk_usage_kb": 235, "github_languages": {"Python": 253044}, "github_archived": false, "github_pushed_at": "2023-06-21T14:44:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-task-structural-learning-using-local"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MRGFutr0p5e", "year": 2022, "status": "rejected", "title": "Graph Barlow Twins: A self-supervised representation learning framework for graphs", "authors": ["Piotr Bielak", "Tomasz Jan Kajdanowicz", "Nitesh Chawla"], "authorids": ["~Piotr_Bielak1", "~Tomasz_Jan_Kajdanowicz1", "~Nitesh_Chawla1"], "authors_source": "OpenReview API", "abstract": "The self-supervised learning (SSL) paradigm is an essential exploration area, which tries to eliminate the need for expensive data labeling. Despite the great success of SSL methods in computer vision and natural language processing, most of them employ contrastive learning objectives that require negative samples, which are hard to define. This becomes even more challenging in the case of graphs and is a bottleneck for achieving robust representations. To overcome such limitations, we propose a framework for self-supervised graph representation learning - Graph Barlow Twins, which utilizes a cross-correlation-based loss function instead of negative samples. Moreover, it does not rely on non-symmetric neural network architectures - in contrast to state-of-the-art self-supervised graph representation learning method BGRL. We show that our method achieves as competitive results as the best self-supervised methods and fully supervised ones while requiring fewer hyperparameters and substantially shorter computation time (ca. 30 times faster than BGRL).", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "cR3Xg4dkEVl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2705/Reviewer_Bwpq"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper applies the recently proposed self-supervised learning method Barlow-Twins to graph structured data. For constructing the augmented version of a graph, previous methods such as edge-dropping or feature masking are used. The paper conducts experimental evaluation on datasets of various scales on both transductive and inductive setting.  ", "review_text": "Strenghts:\n\n1. Paper is clearly written and easy to follow.\n2. The experimental evaluation is robust and experimental details are clearly stated.\n\nWeakness:\n\nThe paper is the straightforward extension of previous work Barlow-Twins for graph structured data. The paper does not has any technical novelty in my opinion. I am willing to increase the score of the paper if in the rebuttal, the authors can clearly state the novelty of the paper w.r.t to the Barlow- Twin paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper applies the recently proposed self-supervised learning method Barlow-Twins to graph structured data. For constructing the augmented version of a graph, previous methods such as edge-dropping or feature masking are used. The paper conducts experimental evaluation on datasets of various scales on both transductive and inductive setting.  ", "main_review": "Strenghts:\n\n1. Paper is clearly written and easy to follow.\n2. The experimental evaluation is robust and experimental details are clearly stated.\n\nWeakness:\n\nThe paper is the straightforward extension of previous work Barlow-Twins for graph structured data. The paper does not has any technical novelty in my opinion. I am willing to increase the score of the paper if in the rebuttal, the authors can clearly state the novelty of the paper w.r.t to the Barlow- Twin paper. ", "summary_of_the_review": "Although the paper is clearly written and many experiments are presented, the paper does not meet the bar of the top conference such as ICLR because of it being a very direct application of a previous paper.\n\n\nMy review is rather short for this paper because based based on the lack of novelty of this paper, I do not have many questions to ask or suggestions to make.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635945186944}, {"id": "aMkmg59WBT-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2705/Reviewer_H4MK"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors study symmetric self-supervised graph representation learning without negative samples, inspired by the Barlow Twins method previously proposed in the image domain. They illustrate that using their method, it is possible to achieve competitive performance to state-of-the-art methods such as BGRL, at a fraction of the training cost.", "review_text": "Disclosure: I have reviewed a previous version of this paper. The authors have included thorough full-batch experiments as well as larger scale datasets such as ogbn-products, which is much appreciated.\n\nThe paper is clear, well-written and easy to follow. The authors propose a simple but meaningful extension of the Barlow Twins idea to the graph domain, and demonstrate its effectiveness on relevant experiments. I think allowing for symmetric loss is a very important direction for graph representation learning, and the proposed solution is an elegant way of achieving that.\n\nMy main concern would be with the reported BGRL results on ogbn-products. I understand that the authors have ran BGRL under the same computational budget as G-BT, but it appears clear that BGRL needs more time to reach peak performance. Would it be possible, just to avoid muddying the waters for future work, to run BGRL for longer and report how the performance is affected? It is OK if this number is higher than G-BT's reported performance -- the authors are optimising for a different metric.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors study symmetric self-supervised graph representation learning without negative samples, inspired by the Barlow Twins method previously proposed in the image domain. They illustrate that using their method, it is possible to achieve competitive performance to state-of-the-art methods such as BGRL, at a fraction of the training cost.", "main_review": "Disclosure: I have reviewed a previous version of this paper. The authors have included thorough full-batch experiments as well as larger scale datasets such as ogbn-products, which is much appreciated.\n\nThe paper is clear, well-written and easy to follow. The authors propose a simple but meaningful extension of the Barlow Twins idea to the graph domain, and demonstrate its effectiveness on relevant experiments. I think allowing for symmetric loss is a very important direction for graph representation learning, and the proposed solution is an elegant way of achieving that.\n\nMy main concern would be with the reported BGRL results on ogbn-products. I understand that the authors have ran BGRL under the same computational budget as G-BT, but it appears clear that BGRL needs more time to reach peak performance. Would it be possible, just to avoid muddying the waters for future work, to run BGRL for longer and report how the performance is affected? It is OK if this number is higher than G-BT's reported performance -- the authors are optimising for a different metric.", "summary_of_the_review": "I think that sufficiently many of my previous concerns have been addressed, and I am now leaning on the side of acceptance. The authors have presented a useful extension of Barlow Twins into the graph domain, and now have experiments in support of the industrial relevance of their method. The novelty is somewhat limited (as is the case for most of the recent graph SSL papers that adapt image domain techniques) but it is useful in and of itself that the gains observed in images transfer well to the irregular domains.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635695965811}, {"id": "kReNgHYO49b", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2705/Reviewer_bHed"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposed a self-supervised learning framework for graph representation learning based on a cross-correlation-based loss function. In the proposed framework, two views of the input graph obtained by augmentation methods are passed through the same encoder to compute two embedding matrices, then Barlow Twins loss is used to compute the loss according to the embedding matrices.\nThe main contribution of this paper lies in that it adapted Barlow Twins from vision to graph representation learning field and evaluated the performance of this self-supervised framework in multiple node classification tasks. The proposed method achieved analogous results compared to SOTA methods with lower time and space complexity.\n", "review_text": "This paper is easy to follow, well-written and organized. This paper is a heuristic attempt to apply Barlow Twins in graph domain. Both transductive and inductive experiments are done to evaluate the performance. The proposed method achieved results on par with the SOTA methods while the time and space complexity are lower.\n\nI have some minor concerns below for the authors to address.\n\n1. The novelty of this paper is limited. The challenges to be tackle of the application about Barlow Twins in graph domain is unclear to me. \n\n2. Data augmentation in vision tasks comes from strong human prior, e.g., random resize, cropping and horizontal flipping would not change the semantic of an image. While the graph data augmentation methods used in this paper is borrowed from previous literatures, it makes no sense to me. For example, applying edge dropping to a protein would obviously lead to different bio-molecules.\n\n3. The experimental results are on par with baseline methods on the most tasks. Considering that the low time and space complexity is coming from previous literature, i.e., Barlow Twins, the experimental contribution is limited.\n\n4. In terms of the encoder network and augmentation hyperparameter design, the paper did not provide comprehensive analysis or ablation studies.\n\n5. The authors carefully describe the downstream datasets. Maybe I miss it, but I don't find the pre-trained dataset used in the experiment.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a self-supervised learning framework for graph representation learning based on a cross-correlation-based loss function. In the proposed framework, two views of the input graph obtained by augmentation methods are passed through the same encoder to compute two embedding matrices, then Barlow Twins loss is used to compute the loss according to the embedding matrices.\nThe main contribution of this paper lies in that it adapted Barlow Twins from vision to graph representation learning field and evaluated the performance of this self-supervised framework in multiple node classification tasks. The proposed method achieved analogous results compared to SOTA methods with lower time and space complexity.\n", "main_review": "This paper is easy to follow, well-written and organized. This paper is a heuristic attempt to apply Barlow Twins in graph domain. Both transductive and inductive experiments are done to evaluate the performance. The proposed method achieved results on par with the SOTA methods while the time and space complexity are lower.\n\nI have some minor concerns below for the authors to address.\n\n1. The novelty of this paper is limited. The challenges to be tackle of the application about Barlow Twins in graph domain is unclear to me. \n\n2. Data augmentation in vision tasks comes from strong human prior, e.g., random resize, cropping and horizontal flipping would not change the semantic of an image. While the graph data augmentation methods used in this paper is borrowed from previous literatures, it makes no sense to me. For example, applying edge dropping to a protein would obviously lead to different bio-molecules.\n\n3. The experimental results are on par with baseline methods on the most tasks. Considering that the low time and space complexity is coming from previous literature, i.e., Barlow Twins, the experimental contribution is limited.\n\n4. In terms of the encoder network and augmentation hyperparameter design, the paper did not provide comprehensive analysis or ablation studies.\n\n5. The authors carefully describe the downstream datasets. Maybe I miss it, but I don't find the pre-trained dataset used in the experiment.", "summary_of_the_review": "This paper adapted the recent Barlow Twins to self-supervised graph representation learning and provided some informative empirical experiment results. With such interesting trials, the reviewer expected to see the concerns are well addressed.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635619043975}], "openreview_url": "https://openreview.net/forum?id=MRGFutr0p5e", "arxiv_id": "2106.02466", "paper_pdf": "papers/MRGFutr0p5e.pdf", "paper_pdf_sha256": "82e54fe756c61cc415cd38fc3f2ee4ec7c36e4d1822f3079316083c69193f270", "paper_pdf_bytes": 361179, "paper_pdf_source": "openreview", "code_url": "https://github.com/pbielak/graph-barlow-twins", "code_repository": "pbielak/graph-barlow-twins", "code_commit": "ec62580aa89bf3f0d20c92e7549031deedc105ab", "code_archive": "repos/MRGFutr0p5e.zip", "code_archive_sha256": "f59e245b6fc3cffd6a4e083ab004d03aae5d71ceb80f94c61f72b061995a89ba", "code_archive_bytes": 233851, "code_file_count": 59, "code_extensions": {".py": 57, ".sh": 2}, "github_disk_usage_kb": 550, "github_languages": {"Python": 203808, "Dockerfile": 782, "Shell": 623}, "github_archived": false, "github_pushed_at": "2023-10-15T10:42:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-barlow-twins-a-self-supervised"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1ibNKMp8SKc", "year": 2021, "status": "rejected", "title": "On Disentangled Representations Learned From Correlated Data", "authors": ["Frederik Träuble", "Elliot Creager", "Niki Kilbertus", "Anirudh Goyal", "Francesco Locatello", "Bernhard Schölkopf", "Stefan Bauer"], "authorids": ["~Frederik_Träuble1", "~Elliot_Creager1", "~Niki_Kilbertus1", "~Anirudh_Goyal1", "~Francesco_Locatello1", "~Bernhard_Schölkopf1", "~Stefan_Bauer1"], "authors_source": "OpenReview API", "abstract": "Despite impressive progress in the last decade, it still remains an open challenge to build models that generalize well across multiple tasks and datasets. One path to achieve this is to learn meaningful and compact representations, in which different semantic aspects of data are structurally disentangled. The focus of disentanglement approaches has been on separating independent factors of variation despite the fact that real-world observations are often not structured into meaningful independent causal variables. In this work, we bridge the gap to real-world scenarios by analyzing the behavior of most prominent methods and disentanglement scores on correlated data in a large scale empirical study (including 4260 models). We show that systematically induced correlations in the dataset are being learned and reflected in the latent representations, while widely used disentanglement scores fall short of capturing these latent correlations. Finally, we demonstrate how to disentangle these latent correlations using weak supervision, even if we constrain this supervision to be causally plausible. Our results thus support the argument to learn independent mechanisms rather than independent factors of variations.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "QHmODdInJA", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2091/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper systematically presents a large-scale empirical study on the disentangled representation learning when the underlying factors are possibly entangled. From the results of purely unsupervised settings, the authors have discovered the shortcomings of the existing metrics of disentanglement as well as the poor learned representations (in terms of disentanglement). However, with the help of small amount of factor labels or other weak supervision signals, recent approaches could learn fairly perfect representation.\n\nFirst of all, it is worthy to pay attention to the possible correlation between factors when you intend to learning a disentangled representation. And the whole empirical results are carried out by very large number of experimental batches, which to some extent could well support the conclusions displayed in the paper.\nHowever, there still exist several limitations in my opinion:\n(1)\tThe design of induced correlation is too simple. As you have mentioned in related work, there were some papers noticed the too ideal assumptions of traditional VAE-based models which may not be held in practice. IMO, the linear dependency between only two variables is far from reality as well. More complicated settings should be involved.\n(2)\tIn line with the former limitation, diagnostics of the potential entanglement should also not be limited to pairwise level, which cannot scale up to high dimensional latent factors.\n(3)\tThe novelty of Section 4 is somewhat limited as all the correction methods and even some conclusions were proposed by the previous work.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "ON DISENTANGLED REPRESENTATIONS LEARNED FROM CORRELATED DATA", "review": "Summary: This paper systematically presents a large-scale empirical study on the disentangled representation learning when the underlying factors are possibly entangled. From the results of purely unsupervised settings, the authors have discovered the shortcomings of the existing metrics of disentanglement as well as the poor learned representations (in terms of disentanglement). However, with the help of small amount of factor labels or other weak supervision signals, recent approaches could learn fairly perfect representation.\n\nFirst of all, it is worthy to pay attention to the possible correlation between factors when you intend to learning a disentangled representation. And the whole empirical results are carried out by very large number of experimental batches, which to some extent could well support the conclusions displayed in the paper.\nHowever, there still exist several limitations in my opinion:\n(1)\tThe design of induced correlation is too simple. As you have mentioned in related work, there were some papers noticed the too ideal assumptions of traditional VAE-based models which may not be held in practice. IMO, the linear dependency between only two variables is far from reality as well. More complicated settings should be involved.\n(2)\tIn line with the former limitation, diagnostics of the potential entanglement should also not be limited to pairwise level, which cannot scale up to high dimensional latent factors.\n(3)\tThe novelty of Section 4 is somewhat limited as all the correction methods and even some conclusions were proposed by the previous work.\n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604026368351}, {"id": "nteY3iilFno", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2091/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses the important problem that most existing disentangled representation learning algorithms analyze statistical independence rather than causal independence. The paper conducts a large scale study to investigate whether statistical correlation prevents learning disentangled representations (according to human defined ground truth factors of variation). \n\nPro:\n\nThis is a timely contribution to clarify problems in evaluating disentangled representation learning algorithms. The paper argues that assuming statistical independence is unrealistic, in practice factors of variation are often causally independent but statistically correlated. This seems like an important missing piece in current empirical evaluation benchmarks. \n\nThere are several interesting observations: \n\n1.The existing metrics for disentanglement do not apply to situations with statistical correlation between the factors of variation. New metrics will be needed. \n\n2.For correlated factors of variation, even though disentanglement fails, the latent space still extrapolates to never seen combinations. \n\n3.Several semi-supervised or weakly supervised methods for disentanglement work well when the factors of variation are correlated. \n\nCon:\n\nSection 3 and 4 are a little hard to follow. There are many results with no clear logical connection to each other. Maybe it’s better to have some bullet points of the empirical findings, then point to specific paragraphs that explain the empirical methodology that produced these findings. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Timely empirical contribution to distangled representation learning with correlated factors of variation", "review": "This paper addresses the important problem that most existing disentangled representation learning algorithms analyze statistical independence rather than causal independence. The paper conducts a large scale study to investigate whether statistical correlation prevents learning disentangled representations (according to human defined ground truth factors of variation). \n\nPro:\n\nThis is a timely contribution to clarify problems in evaluating disentangled representation learning algorithms. The paper argues that assuming statistical independence is unrealistic, in practice factors of variation are often causally independent but statistically correlated. This seems like an important missing piece in current empirical evaluation benchmarks. \n\nThere are several interesting observations: \n\n1.The existing metrics for disentanglement do not apply to situations with statistical correlation between the factors of variation. New metrics will be needed. \n\n2.For correlated factors of variation, even though disentanglement fails, the latent space still extrapolates to never seen combinations. \n\n3.Several semi-supervised or weakly supervised methods for disentanglement work well when the factors of variation are correlated. \n\nCon:\n\nSection 3 and 4 are a little hard to follow. There are many results with no clear logical connection to each other. Maybe it’s better to have some bullet points of the empirical findings, then point to specific paragraphs that explain the empirical methodology that produced these findings. \n\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603941385540}, {"id": "5Z3qk4rmnzr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2091/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies the behaviour of disentanglement methods and metrics on data where a couple of factors of variation (FoV) are correlated, a more realistic setup compared to the usual independent FoV setting in the literature. The paper shows how the correlation in the FoV is reflected in the representations learned by the models, and claims that the widely used disentanglement scores fail to capture these correlations. A couple of solutions that use weak supervision are suggested.\n\nStrengths:\n1. The paper attempts to address an important limitation of current unsupervised disentangling methods, in that they assume independent FoV. I agree that this is an unrealistic assumption, and only holds for the synthetic datasets used in the literature.\n2. The experimental evaluation of the paper is extensive\n3. The paper is presented clearly for the most part.\n\nWeaknesses:\n1. At the beginning of Section 3.2, the claim that the existing disentanglement metrics fail to capture correlations in the training data seems flawed. Figure 2, that the violin plots of metrics for 180 models (with different loss/hyperparam combinations/random seed) shows no clear trend, is used as evidence for the claim. However this argument ignores an important confounding factor, which is the behaviour of the models (trained with losses originally designed for data with independent FoV) when trained on a dataset with correlated FoV. Without understanding the behaviour of models after training, it is difficult to draw any conclusions from Figure 2. Moreover comparing violin plots of all 180 models does not seem sensible - the models with poor disentangling performance will likely be uninformative, creating noise for comparisons between different settings. Hence I think it makes more sense to compare results for the top models for a given metric (as is done in Fig3).\n2. I also have doubts regarding the conclusion itself, that all existing metrics fail to capture correlations in the training data. I think this only holds for some metrics and not for others. Suppose the representations that we are evaluating are the ground truth FoV values. Then the BetaVAE and FactorVAE metrics will give a perfect score regardless of the degree of correlation by design. But the MIG/SAP/DCI scores will be lower for higher correlation in the FoV, since the mutual information/prediction accuracy between the two correlated factors will be higher than between uncorrelated factors. \n3. The result presented by Figure 3, that training disentanglement models on strongly correlated data gives representations encoding mixtures of the correlated FoV, is unsurprising in my opinion, and adds little value to the pool of knowledge in the literature.\n4. In Section 2, I think that the statement “it has been shown that purely unsupervised learning of disentangled representations is impossible (Locatello et al., 2019b)” is a wrong interpretation of Locatello et al. They show that optimising marginal likelihood in a generative model (such as a VAE) cannot achieve disentangling without any inductive biases in the model. But the inductive biases in the models used by disentangling methods, along with the loss (that is a variant of the ELBO and not the marginal likelihood), are what allow disentangling in practice. There are also theoretical works such as [1] that explain this behaviour.\n5. In Section 3.2, the conclusion “These results suggest that we cannot expect disentangled representations learned unsupervisedly to help reduce unfairness beyond the benefits discussed in Locatello et al. (2019a)” seems problematic. It’s not the unsupervised learning that’s problematic, rather it’s the model assumption that the FoV are independent that is problematic in the scenario of correlated FoV, which is an unsurprising conclusion. \n6. The conclusion in Section 3.3, that disentanglement methods can generalize towards unseen FoV configurations, is old news that is already shown in the literature in disentanglement models trained on datasets with an incomplete set of FoV configurations e.g. 3D Faces, CelebA\n7. In Section 4.2, the claim “This kind of extra knowledge is often available at no cost , e.g., in temporarily close frames from a video of a moving robot arm where some factors remain unchanged” is unjustified. We still need to know the number of factors that have changed, which usually requires human labelling\n8. The weak supervision for the correlated FoV when applying Ada-GVAE relies on being able to generate data where the correlation is broken, which seems like an unrealistic assumption. The authors seem to address this by assuming a causal relation between the two factors. However even if there is a causal relation between the two factors (C1 causes C2), the correlation implies that some (C1,C2) pairs are very unlikely to appear in the data, so you cannot “sample any value in C1” given C2. Hence I don’t understand how the causal setting at the end of Section 4.2 helps address the problems of having to generate data where the correlation is broken.\n\nOverall the paper does address an important problem in the disentanglement literature, but the conclusions drawn from the extensive evaluation are either unsurprising or unjustified. Moreover, the proposed solution via weak supervision appears flawed because it requires generating data where the correlation is broken, a very unrealistic assumption.\n\nOther points:\n- Typo in Figure 2 caption: “lower \\sigma indicates less correlation” <- “higher \\sigma indicates less correlation”\n- Section 3.1: “P(z_c1, z_c2) ~ N(z_c2 - \\alpha z_c1, \\sigma)” How does the RHS define a joint distribution? The RHS shows a scalar normal distribution, whereas the LHS is a joint density. Do you want to replace “\\sim” with “ \\propto” ?\n- Why doesn’t entanglement decrease with #labels > 100 in Figure 6?\n- It might be helpful to also look at the disentanglement metrics for just the two correlated factors, to further highlight the differences between different models.\n\n[1] Rolinek, M., Zietlow, D. and Martius, G., Variational Autoencoders Pursue PCA Directions (by Accident). CVPR 2019.\n\n===========================\n\nScore raised to 5 following response to the rebuttal below, then to 6 following the re-rebuttal.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Evaluation is extensive, but the conclusions drawn from them are either 1. not justified very well by the evaluation 2. unsurprising/already known.", "review": "The paper studies the behaviour of disentanglement methods and metrics on data where a couple of factors of variation (FoV) are correlated, a more realistic setup compared to the usual independent FoV setting in the literature. The paper shows how the correlation in the FoV is reflected in the representations learned by the models, and claims that the widely used disentanglement scores fail to capture these correlations. A couple of solutions that use weak supervision are suggested.\n\nStrengths:\n1. The paper attempts to address an important limitation of current unsupervised disentangling methods, in that they assume independent FoV. I agree that this is an unrealistic assumption, and only holds for the synthetic datasets used in the literature.\n2. The experimental evaluation of the paper is extensive\n3. The paper is presented clearly for the most part.\n\nWeaknesses:\n1. At the beginning of Section 3.2, the claim that the existing disentanglement metrics fail to capture correlations in the training data seems flawed. Figure 2, that the violin plots of metrics for 180 models (with different loss/hyperparam combinations/random seed) shows no clear trend, is used as evidence for the claim. However this argument ignores an important confounding factor, which is the behaviour of the models (trained with losses originally designed for data with independent FoV) when trained on a dataset with correlated FoV. Without understanding the behaviour of models after training, it is difficult to draw any conclusions from Figure 2. Moreover comparing violin plots of all 180 models does not seem sensible - the models with poor disentangling performance will likely be uninformative, creating noise for comparisons between different settings. Hence I think it makes more sense to compare results for the top models for a given metric (as is done in Fig3).\n2. I also have doubts regarding the conclusion itself, that all existing metrics fail to capture correlations in the training data. I think this only holds for some metrics and not for others. Suppose the representations that we are evaluating are the ground truth FoV values. Then the BetaVAE and FactorVAE metrics will give a perfect score regardless of the degree of correlation by design. But the MIG/SAP/DCI scores will be lower for higher correlation in the FoV, since the mutual information/prediction accuracy between the two correlated factors will be higher than between uncorrelated factors. \n3. The result presented by Figure 3, that training disentanglement models on strongly correlated data gives representations encoding mixtures of the correlated FoV, is unsurprising in my opinion, and adds little value to the pool of knowledge in the literature.\n4. In Section 2, I think that the statement “it has been shown that purely unsupervised learning of disentangled representations is impossible (Locatello et al., 2019b)” is a wrong interpretation of Locatello et al. They show that optimising marginal likelihood in a generative model (such as a VAE) cannot achieve disentangling without any inductive biases in the model. But the inductive biases in the models used by disentangling methods, along with the loss (that is a variant of the ELBO and not the marginal likelihood), are what allow disentangling in practice. There are also theoretical works such as [1] that explain this behaviour.\n5. In Section 3.2, the conclusion “These results suggest that we cannot expect disentangled representations learned unsupervisedly to help reduce unfairness beyond the benefits discussed in Locatello et al. (2019a)” seems problematic. It’s not the unsupervised learning that’s problematic, rather it’s the model assumption that the FoV are independent that is problematic in the scenario of correlated FoV, which is an unsurprising conclusion. \n6. The conclusion in Section 3.3, that disentanglement methods can generalize towards unseen FoV configurations, is old news that is already shown in the literature in disentanglement models trained on datasets with an incomplete set of FoV configurations e.g. 3D Faces, CelebA\n7. In Section 4.2, the claim “This kind of extra knowledge is often available at no cost , e.g., in temporarily close frames from a video of a moving robot arm where some factors remain unchanged” is unjustified. We still need to know the number of factors that have changed, which usually requires human labelling\n8. The weak supervision for the correlated FoV when applying Ada-GVAE relies on being able to generate data where the correlation is broken, which seems like an unrealistic assumption. The authors seem to address this by assuming a causal relation between the two factors. However even if there is a causal relation between the two factors (C1 causes C2), the correlation implies that some (C1,C2) pairs are very unlikely to appear in the data, so you cannot “sample any value in C1” given C2. Hence I don’t understand how the causal setting at the end of Section 4.2 helps address the problems of having to generate data where the correlation is broken.\n\nOverall the paper does address an important problem in the disentanglement literature, but the conclusions drawn from the extensive evaluation are either unsurprising or unjustified. Moreover, the proposed solution via weak supervision appears flawed because it requires generating data where the correlation is broken, a very unrealistic assumption.\n\nOther points:\n- Typo in Figure 2 caption: “lower \\sigma indicates less correlation” <- “higher \\sigma indicates less correlation”\n- Section 3.1: “P(z_c1, z_c2) ~ N(z_c2 - \\alpha z_c1, \\sigma)” How does the RHS define a joint distribution? The RHS shows a scalar normal distribution, whereas the LHS is a joint density. Do you want to replace “\\sim” with “ \\propto” ?\n- Why doesn’t entanglement decrease with #labels > 100 in Figure 6?\n- It might be helpful to also look at the disentanglement metrics for just the two correlated factors, to further highlight the differences between different models.\n\n[1] Rolinek, M., Zietlow, D. and Martius, G., Variational Autoencoders Pursue PCA Directions (by Accident). CVPR 2019.\n\n===========================\n\nScore raised to 5 following response to the rebuttal below, then to 6 following the re-rebuttal.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603729809755}], "openreview_url": "https://openreview.net/forum?id=1ibNKMp8SKc", "arxiv_id": "2006.07886", "paper_pdf": "papers/1ibNKMp8SKc.pdf", "paper_pdf_sha256": "dbeeab21fb6fec7ba99c09c04b8fb42ac3694b4fd1be7224fddf3bd0545d0809", "paper_pdf_bytes": 12556365, "paper_pdf_source": "openreview", "code_url": "https://github.com/ftraeuble/disentanglement_lib", "code_repository": "ftraeuble/disentanglement_lib", "code_commit": "39584ad5270a090723c8f12567c47d84f02444c9", "code_archive": "repos/1ibNKMp8SKc.zip", "code_archive_sha256": "6cce794571d5298cc269ce94d36e1c3f3508487542d8518f0b2bec8e2bfc5128", "code_archive_bytes": 912960, "code_file_count": 126, "code_extensions": {".py": 126}, "github_disk_usage_kb": 741, "github_languages": {"Python": 660604, "Shell": 6767}, "github_archived": false, "github_pushed_at": "2021-06-08T18:37:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/is-independence-all-you-need-on-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJlDoT4twr", "year": 2020, "status": "rejected", "title": "Unified Probabilistic Deep Continual Learning through Generative Replay and Open Set Recognition", "authors": ["Martin Mundt", "Sagnik Majumder", "Iuliia Pliushch", "Visvanathan Ramesh"], "authorids": ["mmundt@em.uni-frankfurt.de", "majumder@ccc.cs.uni-frankfurt.de", "pliushch@em.uni-frankfurt.de", "ramesh@fias.uni-frankfurt.de"], "authors_source": "OpenReview API", "abstract": "We introduce a unified probabilistic approach for deep continual learning based on variational Bayesian inference with open set recognition. Our model combines a joint probabilistic encoder with a generative model and a linear classifier that get shared across tasks. The open set recognition bounds the approximate posterior by fitting regions of high density on the basis of correctly classified data points and balances open set detection with recognition errors. Catastrophic forgetting is significantly alleviated through generative replay, where the open set recognition is used to sample from high density areas of the class specific posterior and reject statistical outliers. Our approach naturally allows for forward and backward transfer while maintaining past knowledge without the necessity of storing old data, regularization or inferring task labels. We demonstrate compelling results in the challenging scenario of incrementally expanding the single-head classifier for both class incremental visual and audio classification tasks, as well as incremental learning of datasets across modalities.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1x_nF116B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper748/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper proposes a unified model for continual learning and aims to address the following problems:\nOut-of-train-domain dataset recognition\nCatastrophic forgetting\nThe out-of-domain or open set recognition model is not only used to detect outliers but also for sampling “representative data” of previous tasks for forward (and backward) transfer. \n\nCOMMENT(S)/QUERIES:\n------------------------------\n \nWhile the experiments do justice to the contributions mentioned in the paper, I believe that certain sections need clarifications and expansion (while some can be dismissed). \n\n1. Given my limited knowledge in the literature on this topic, I would have appreciated a proper related work section. \n\n2. I believe equation (1) could have been moved to supplementary for reference since equation (2) is the only important equation in the paper. \nI don’t believe Equation (1) is really used for testing purposes. It’s just that the data is sent through the probabilistic encoder and then classified. There is no need for reconstruction of the data point.\n\n3. It’s difficult to follow the flow of the method. Shouldn’t generative replay (algorithm 3) be placed before open set recognition of unknown and uncertain inputs (algorithm 2), since the latter is probably just used during the test, while the former affects the training procedure directly (implicit data augmentation)?\n\n4. Major concern: It’s somewhat strange to observe that the VAE model is able to generate multiclass data so fluidly with a simple gaussian prior. This kind of challenges the belief that current generative models are unable to capture all modes perfectly. A small note about why multimodal prior was not used (which intuitively and mathematically makes more sense) and also a statement about the average time to generate multiclass multi-modal data using algorithm 3 would have been nice. \n\n5. In section 3.2, “For our single-head expanding classifier this ensures..”. While having a single-head expanding classifier is listed as one of the important contributions, it hasn’t been given enough justice w.r.t to the implementation details. Is it like adding a completely new classifier during the training of the new task?\nThe entire section on hyper-parameters could’ve been moved to the supplementary if space was a major constraint but compromising on details about an important contribution only weakens the paper. \n\nI especially like figure 2 in the experiments, where the importance of Weibull CDF outlier rejection prior is highlighted.\n\n\nTypo: \nincorrect opening inverted comma for the word background in the introduction section (page 1)\n\nOVERALL COMMENT\n-----------------------------\n\nThe paper combines the generative, and discriminative models into one framework for multiple important tasks. While the contributions are clear in the introduction, the presentation of the paper is somewhat too complicated at a couple of places. It does not do full justice to explaining its more vital components and some parts of the method section feel like a jigsaw puzzle, where readers are heavily expected to “figure out on their own”. \nWith proper presentation though, I can envision this paper contributing positively to unified frameworks in general.\n\nDue to the above reasons. I am giving it a score of 6.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "Summary: This paper proposes a unified model for continual learning and aims to address the following problems:\nOut-of-train-domain dataset recognition\nCatastrophic forgetting\nThe out-of-domain or open set recognition model is not only used to detect outliers but also for sampling “representative data” of previous tasks for forward (and backward) transfer. \n\nCOMMENT(S)/QUERIES:\n------------------------------\n \nWhile the experiments do justice to the contributions mentioned in the paper, I believe that certain sections need clarifications and expansion (while some can be dismissed). \n\n1. Given my limited knowledge in the literature on this topic, I would have appreciated a proper related work section. \n\n2. I believe equation (1) could have been moved to supplementary for reference since equation (2) is the only important equation in the paper. \nI don’t believe Equation (1) is really used for testing purposes. It’s just that the data is sent through the probabilistic encoder and then classified. There is no need for reconstruction of the data point.\n\n3. It’s difficult to follow the flow of the method. Shouldn’t generative replay (algorithm 3) be placed before open set recognition of unknown and uncertain inputs (algorithm 2), since the latter is probably just used during the test, while the former affects the training procedure directly (implicit data augmentation)?\n\n4. Major concern: It’s somewhat strange to observe that the VAE model is able to generate multiclass data so fluidly with a simple gaussian prior. This kind of challenges the belief that current generative models are unable to capture all modes perfectly. A small note about why multimodal prior was not used (which intuitively and mathematically makes more sense) and also a statement about the average time to generate multiclass multi-modal data using algorithm 3 would have been nice. \n\n5. In section 3.2, “For our single-head expanding classifier this ensures..”. While having a single-head expanding classifier is listed as one of the important contributions, it hasn’t been given enough justice w.r.t to the implementation details. Is it like adding a completely new classifier during the training of the new task?\nThe entire section on hyper-parameters could’ve been moved to the supplementary if space was a major constraint but compromising on details about an important contribution only weakens the paper. \n\nI especially like figure 2 in the experiments, where the importance of Weibull CDF outlier rejection prior is highlighted.\n\n\nTypo: \nincorrect opening inverted comma for the word background in the introduction section (page 1)\n\nOVERALL COMMENT\n-----------------------------\n\nThe paper combines the generative, and discriminative models into one framework for multiple important tasks. While the contributions are clear in the introduction, the presentation of the paper is somewhat too complicated at a couple of places. It does not do full justice to explaining its more vital components and some parts of the method section feel like a jigsaw puzzle, where readers are heavily expected to “figure out on their own”. \nWith proper presentation though, I can envision this paper contributing positively to unified frameworks in general.\n\nDue to the above reasons. I am giving it a score of 6.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1575053744245}, {"id": "S1eM-mFCFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper748/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper combines replay and openMax approach to help continual learning.  The results shows robustness on different dataset include image and audio in the continual learning condition, where the new come data has a different distribution but the model still able to maintain reasonable quality for the previously and newly come examples. To my understanding, this approach was not ground-breaking but seems a reasonable combinations.\n\nI'm learning to give weak reject for this paper because of it's poorly written. (1) It's very hard to align the contribution claimed by the paper and previous work in the introduction section. I highly suggest the author re-write this part and has a separate section about related work and explicit describe the difference compare to others. (2) The contribution seems over-claimed, it said it's a unified framework, but I don't understand what it has been unified. (3) In the experimental part, it use audioMNIST. Any reason to use this dataset? There are more well-defined audio task such as TIMIT for phoneme classification or Aurora for digital recognition. They are more easy to understand since they have well established benchmark.\n\nGiven my limited knowledge on this the literature of this topic, I'm happy to change the score if the written being improved and the following question being addressed.\n\n(1)  Introduction. I take most of space to describe previous work, but hard to find out what's difference of this paper. My understanding it combines A and B and apply it to C. But it claim it's a unified framework. \n(2) \"We fully share our model across tasks and automatically expand the linear classifier with additional units when encountering new classes, thus not requiring explicit task labels.\" I cannot link \"automatic\" with the proposed method. Is that doable because of the proposed framework? \n(3) Why use AudioMNIST which is an unusual task for audio?\n(4) For the giant Table 1, I suggest link each acronym with the reference paper. So it can easily get how it associate with different approach. Highlight some numbers can also help the reader understand what's going on in this giant table.\n(5) Can the author give me some insights, what these KL loss demonstrated in the table? I feel since you use the beta-vae version, the kl scale is depend on different approach, not really comparable for these different models.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper combines replay and openMax approach to help continual learning.  The results shows robustness on different dataset include image and audio in the continual learning condition, where the new come data has a different distribution but the model still able to maintain reasonable quality for the previously and newly come examples. To my understanding, this approach was not ground-breaking but seems a reasonable combinations.\n\nI'm learning to give weak reject for this paper because of it's poorly written. (1) It's very hard to align the contribution claimed by the paper and previous work in the introduction section. I highly suggest the author re-write this part and has a separate section about related work and explicit describe the difference compare to others. (2) The contribution seems over-claimed, it said it's a unified framework, but I don't understand what it has been unified. (3) In the experimental part, it use audioMNIST. Any reason to use this dataset? There are more well-defined audio task such as TIMIT for phoneme classification or Aurora for digital recognition. They are more easy to understand since they have well established benchmark.\n\nGiven my limited knowledge on this the literature of this topic, I'm happy to change the score if the written being improved and the following question being addressed.\n\n(1)  Introduction. I take most of space to describe previous work, but hard to find out what's difference of this paper. My understanding it combines A and B and apply it to C. But it claim it's a unified framework. \n(2) \"We fully share our model across tasks and automatically expand the linear classifier with additional units when encountering new classes, thus not requiring explicit task labels.\" I cannot link \"automatic\" with the proposed method. Is that doable because of the proposed framework? \n(3) Why use AudioMNIST which is an unusual task for audio?\n(4) For the giant Table 1, I suggest link each acronym with the reference paper. So it can easily get how it associate with different approach. Highlight some numbers can also help the reader understand what's going on in this giant table.\n(5) Can the author give me some insights, what these KL loss demonstrated in the table? I feel since you use the beta-vae version, the kl scale is depend on different approach, not really comparable for these different models."}, "tcdate": 1571881721574}, {"id": "B1xEKODCKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper748/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tackles the problem of catastrophic forgetting when data is organized in a large number of batches of data (tasks) that are sequentially made available. To avoid catastrophic forgetting, the authors learn a VAE that generates the training data (both inputs and labels) and retrain it using samples from the new task combined with samples generated from the VAE trained in the previous tasks (generative replay). In this way, there's no need to store all past data and even the first learned batch keeps being refreshed and should not be forgotten.\n\nI like that this paper uses a single global probabilistic model instead of separate discriminative and generative ones. Unfortunately, there are several things that left me unconvinced about this paper:\n\n1) Presentation of the paper\n\n- Variables x, y, z are introduced and talked about without explanation. The graphical model or factorization assumptions are not even mentioned until after the loss has been defined. A normal flow is to first describe the model and what the involved variables mean, and then talk about what the loss for learning it should be, not the other way around.\n\n- Text contradicting the equation: \"In order to balance the individual loss terms, we normalize according to dimensions and weight the KL divergence with a constant of 0.1\". But equation (2) shows a loss with no weighting. I'm assuming the text is correct, but then a beta should be added to the equation in front of the KL divergence.\n\n- Tables and figures are inconveniently far from where they are referenced in the text.\n\n2) Theoretical inconsistencies\n\nAlthough the system might work overall, two things seem to be technically incorrect:\n\n- The decoder and classifier are expected to approximate the distribution of training data according to the authors (for valid generative replay). This is not true in a beta-VAE. The weighting of the KL that the authors introduce is going to bias the learned generator towards the high probability regions. This is not a sound mechanism to achieve an as-faithful-as-possible (limited by the expressiveness of the encoder-decoder architectures) approximation to the training data.\n\n- A Weibull distribution is used to model the same data, again, in a different way. I.e., there are two different probabilistic models modeling the same data in inconsistent ways and one or the other is used depending on the part of the system. (As an example, q(z) could be arbitrarily multimodal as far as the encoder is concerned, but the Weibull seems to force one mode per class. But regardless of this, both models are inconsistent.)\n\n- Similarly, the proposed rejection sampling scheme of OCDVAE is not consistent with the theory of VAEs and it's a post-hoc tweak that is not theoretically expected to provide a pdf of data with lower KL divergence to the true data pdf.\n\n3) Experiments\n\nFinally, the experimental results do not look very compelling, it seems to be overall worse than the baselines in the two image datasets and slightly better in the audio dataset, so it's unclear that this approach is superior.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper tackles the problem of catastrophic forgetting when data is organized in a large number of batches of data (tasks) that are sequentially made available. To avoid catastrophic forgetting, the authors learn a VAE that generates the training data (both inputs and labels) and retrain it using samples from the new task combined with samples generated from the VAE trained in the previous tasks (generative replay). In this way, there's no need to store all past data and even the first learned batch keeps being refreshed and should not be forgotten.\n\nI like that this paper uses a single global probabilistic model instead of separate discriminative and generative ones. Unfortunately, there are several things that left me unconvinced about this paper:\n\n1) Presentation of the paper\n\n- Variables x, y, z are introduced and talked about without explanation. The graphical model or factorization assumptions are not even mentioned until after the loss has been defined. A normal flow is to first describe the model and what the involved variables mean, and then talk about what the loss for learning it should be, not the other way around.\n\n- Text contradicting the equation: \"In order to balance the individual loss terms, we normalize according to dimensions and weight the KL divergence with a constant of 0.1\". But equation (2) shows a loss with no weighting. I'm assuming the text is correct, but then a beta should be added to the equation in front of the KL divergence.\n\n- Tables and figures are inconveniently far from where they are referenced in the text.\n\n2) Theoretical inconsistencies\n\nAlthough the system might work overall, two things seem to be technically incorrect:\n\n- The decoder and classifier are expected to approximate the distribution of training data according to the authors (for valid generative replay). This is not true in a beta-VAE. The weighting of the KL that the authors introduce is going to bias the learned generator towards the high probability regions. This is not a sound mechanism to achieve an as-faithful-as-possible (limited by the expressiveness of the encoder-decoder architectures) approximation to the training data.\n\n- A Weibull distribution is used to model the same data, again, in a different way. I.e., there are two different probabilistic models modeling the same data in inconsistent ways and one or the other is used depending on the part of the system. (As an example, q(z) could be arbitrarily multimodal as far as the encoder is concerned, but the Weibull seems to force one mode per class. But regardless of this, both models are inconsistent.)\n\n- Similarly, the proposed rejection sampling scheme of OCDVAE is not consistent with the theory of VAEs and it's a post-hoc tweak that is not theoretically expected to provide a pdf of data with lower KL divergence to the true data pdf.\n\n3) Experiments\n\nFinally, the experimental results do not look very compelling, it seems to be overall worse than the baselines in the two image datasets and slightly better in the audio dataset, so it's unclear that this approach is superior."}, "tcdate": 1571874940183}], "openreview_url": "https://openreview.net/forum?id=rJlDoT4twr", "arxiv_id": "1905.12019", "paper_pdf": "papers/rJlDoT4twr.pdf", "paper_pdf_sha256": "1c1a0e20a0bcea8cc9981ce51d8f5ceeacc78494b649f204598fa3e66352818a", "paper_pdf_bytes": 8387536, "paper_pdf_source": "openreview", "code_url": "https://github.com/MrtnMndt/OpenVAE_ContinualLearning", "code_repository": "MrtnMndt/OpenVAE_ContinualLearning", "code_commit": "2cca59334c8993f96af15daf70d99ba89fea003f", "code_archive": "repos/rJlDoT4twr.zip", "code_archive_sha256": "e779449113aa7315d08be861bbc253c06c146fb2b3d1cac9de079c131ed7829c", "code_archive_bytes": 1293385, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 1945, "github_languages": {"Python": 236893}, "github_archived": false, "github_pushed_at": "2022-04-01T12:53:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unified-probabilistic-deep-continual-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jimY7NF6aJ", "year": 2026, "status": "rejected", "title": "MUSEG: Reinforcing Video Temporal Understanding via Timestamp-Aware Multi-Segment Grounding", "authors": ["Fuwen Luo", "Shengfeng Lou", "Chi Chen", "Ziyue Wang", "Chenliang Li", "Weizhou Shen", "Jiyue Guo", "Peng Li", "Ming Yan", "Fei Huang", "Yang Liu"], "authorids": ["~Fuwen_Luo1", "~Shengfeng_Lou1", "~Chi_Chen1", "~Ziyue_Wang4", "~Chenliang_Li2", "~Weizhou_Shen1", "~Jiyue_Guo1", "~Peng_Li2", "~Ming_Yan2", "~Fei_Huang2", "~Yang_Liu19"], "authors_source": "OpenReview API", "abstract": "Video temporal understanding is crucial for multimodal large language models (MLLMs) to reason over events in videos. Despite recent advances in general video understanding, current MLLMs still struggle with fine-grained temporal reasoning. While reinforcement learning (RL) has been explored to address this issue recently, existing RL approaches remain limited in performance on time-sensitive tasks. In this work, we propose **MUSEG**, a novel RL-based method that enhances temporal understanding by introducing timestamp-aware multi-segment grounding. MUSEG enables MLLMs to align queries with multiple relevant video segments, promoting more comprehensive temporal reasoning. To facilitate effective learning, we design a customized RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. Extensive experiments on temporal grounding and time-sensitive video question answering (QA) tasks demonstrate that \\methodname significantly outperforms existing methods and generalizes well across diverse temporal understanding scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "y5zbe4GQGh", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5002/Reviewer_ogyi"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper proposes MUSEG, a RL-based approach to enhance temporal understanding by introducing timestamp-aware multi-segment grounding. It designs a RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. The experimental results show superior performance over previous methods.", "review_text": "The paper proposes MUSEG, a RL-based approach to enhance temporal understanding by introducing timestamp-aware multi-segment grounding. It designs a RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. The experimental results show superior performance over previous methods.", "strengths": "1. The paper addresses an important problem and the paper is clearly written. Motivation is clear.\n2. Evaluations on in- and cross-domain tasks show good results over previous methods.", "weaknesses": "1. To me, the segment matching reward and timestamp reward are likely targeting the same thing. And the reward formats are duplicating. I am wondering the effect of each component. I think there should be an ablation study on how each type of rewards matters to the performance.\n2. Since there are several video temporal grounding tasks in the literature, it would be beneficial if authors also evaluate on Ego4D-NLQ, TaCoS, ANet-Captions, QVHighlights. \n3. The paper does not have sufficient ablation studies.", "questions": "1. How do you handle long and short videos in this setting? \n2. How do you make sure the timestamps and video frames are aligned to each other? \n3. Could you report results with only phase 1 training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes MUSEG, a RL-based approach to enhance temporal understanding by introducing timestamp-aware multi-segment grounding. It designs a RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. The experimental results show superior performance over previous methods.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper addresses an important problem and the paper is clearly written. Motivation is clear.\n2. Evaluations on in- and cross-domain tasks show good results over previous methods.", "weaknesses": "1. To me, the segment matching reward and timestamp reward are likely targeting the same thing. And the reward formats are duplicating. I am wondering the effect of each component. I think there should be an ablation study on how each type of rewards matters to the performance.\n2. Since there are several video temporal grounding tasks in the literature, it would be beneficial if authors also evaluate on Ego4D-NLQ, TaCoS, ANet-Captions, QVHighlights. \n3. The paper does not have sufficient ablation studies.", "questions": "1. How do you handle long and short videos in this setting? \n2. How do you make sure the timestamps and video frames are aligned to each other? \n3. Could you report results with only phase 1 training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762025500350}, {"id": "aGYL7cIlXE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5002/Reviewer_YPLo"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "The paper design rewards for temporal reasoning to do RL, and experiment on 2 models: Qwen2.5VL 7B and 3B, which beats the baselines. The rewards they design are segment matching rewards and timestamp reward.", "review_text": "The paper design rewards for temporal reasoning to do RL, and experiment on 2 models: Qwen2.5VL 7B and 3B, which beats the baselines. The rewards they design are segment matching rewards and timestamp reward.", "strengths": "The paper is clearly written\nThe reward design is interesting", "weaknesses": "1. it is clear that adding temporal related reward will improve the temporal related tasks, it will be cool to test on non-temporal video tasks\n2. the novelty of the paper is only on the reward design for temporal tasks, everything else is standard \n3. the RL training require a lot of hyperparameter tuning and manual engineering (e.g. 900 steps where first 400 with timestamp reward and 500 without)", "questions": "see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper design rewards for temporal reasoning to do RL, and experiment on 2 models: Qwen2.5VL 7B and 3B, which beats the baselines. The rewards they design are segment matching rewards and timestamp reward.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "The paper is clearly written\nThe reward design is interesting", "weaknesses": "1. it is clear that adding temporal related reward will improve the temporal related tasks, it will be cool to test on non-temporal video tasks\n2. the novelty of the paper is only on the reward design for temporal tasks, everything else is standard \n3. the RL training require a lot of hyperparameter tuning and manual engineering (e.g. 900 steps where first 400 with timestamp reward and 500 without)", "questions": "see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761521702279}, {"id": "dovkesBx3g", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5002/Reviewer_d1dA"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces a RL framework called MUSEG. The proposed method improves MLLMs' ability to understand temporal events in videos. Specifically, this paper proposed the task \"multi-segment grounding\", and uses GRPO with \"segment matching reward\" and \"timestamp reward\" to align predicted with true video segments and encourage temporal reasoning. Experiments show that MUSEG significantly outperforms previous SFT- and RL-based models on multiple benchmarks, achieving stronger fine-grained temporal reasoning and better generalization to time-sensitive video tasks.", "review_text": "This paper introduces a RL framework called MUSEG. The proposed method improves MLLMs' ability to understand temporal events in videos. Specifically, this paper proposed the task \"multi-segment grounding\", and uses GRPO with \"segment matching reward\" and \"timestamp reward\" to align predicted with true video segments and encourage temporal reasoning. Experiments show that MUSEG significantly outperforms previous SFT- and RL-based models on multiple benchmarks, achieving stronger fine-grained temporal reasoning and better generalization to time-sensitive video tasks.", "strengths": "1. This paper's proposed methodology has a very clear motivation. It performs a preliminary empirical study revealing why using single-segment grounding as the training task won't work. \n2. The proposed approach outperforms previous baseline models on multiple benchmarks with both 3B and 7B models. \n3. This paper provides comprehensive analysis into why the proposed approach would work, along with the strength (1), making the proposed approach more reasonable. \n4. The proposed method won't hurt the model's performance on general video QA task.", "weaknesses": "1. The proposed method's performance improvement on general task is greatly limited (compared with the performance improvement on grounding task). On all benchmarks, the performance improvement is less than 2%. The result reveals that there's a great limitation on the method's generalizability. \n2. The idea of utilizing the multi-segment as task has already appeared in the related work [1], which appears on the Internet on Jun 23 2025. It is one month earlier than the ICLR's comparison cutoff July 24, 2025, therefore, I think the connection to this work needs to be discussed, and it's likely that this work[1] needs to be compared with the proposed approach. \n\n[1] Universal Video Temporal Grounding with Generative Multi-modal Large Language Models, NeurIPS 2025", "questions": "1. In Table 2, is the \"+vanilla GRPO\" based on \"+vanilla SFT\" (so it's + GRPO AND +SFT), or is based directly on the original \"Qwen2.5-VL-7B\"?\n2. Why \"seconds\" is chosen as unit in verbal reasoning, rather than frames, millioseconds or minutes. If this will cause issues for long video or extremely short video?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a RL framework called MUSEG. The proposed method improves MLLMs' ability to understand temporal events in videos. Specifically, this paper proposed the task \"multi-segment grounding\", and uses GRPO with \"segment matching reward\" and \"timestamp reward\" to align predicted with true video segments and encourage temporal reasoning. Experiments show that MUSEG significantly outperforms previous SFT- and RL-based models on multiple benchmarks, achieving stronger fine-grained temporal reasoning and better generalization to time-sensitive video tasks.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "1. This paper's proposed methodology has a very clear motivation. It performs a preliminary empirical study revealing why using single-segment grounding as the training task won't work. \n2. The proposed approach outperforms previous baseline models on multiple benchmarks with both 3B and 7B models. \n3. This paper provides comprehensive analysis into why the proposed approach would work, along with the strength (1), making the proposed approach more reasonable. \n4. The proposed method won't hurt the model's performance on general video QA task.", "weaknesses": "1. The proposed method's performance improvement on general task is greatly limited (compared with the performance improvement on grounding task). On all benchmarks, the performance improvement is less than 2%. The result reveals that there's a great limitation on the method's generalizability. \n2. The idea of utilizing the multi-segment as task has already appeared in the related work [1], which appears on the Internet on Jun 23 2025. It is one month earlier than the ICLR's comparison cutoff July 24, 2025, therefore, I think the connection to this work needs to be discussed, and it's likely that this work[1] needs to be compared with the proposed approach. \n\n[1] Universal Video Temporal Grounding with Generative Multi-modal Large Language Models, NeurIPS 2025", "questions": "1. In Table 2, is the \"+vanilla GRPO\" based on \"+vanilla SFT\" (so it's + GRPO AND +SFT), or is based directly on the original \"Qwen2.5-VL-7B\"?\n2. Why \"seconds\" is chosen as unit in verbal reasoning, rather than frames, millioseconds or minutes. If this will cause issues for long video or extremely short video?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761447473092}, {"id": "sWtp5Nz4XB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5002/Reviewer_5XGo"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces MUSEG, a reinforcement learning (RL) method to improve the fine-grained temporal understanding of Multimodal Large Language Models (MLLMs) in videos.\n\nThe core idea is timestamp-aware multi-segment grounding, which enables the model to align a query with multiple relevant video segments. This is optimized through a customized RL training recipe using phased rewards that progressively guides the model's reasoning.\n\nExperiments show MUSEG significantly outperforms existing methods on temporal grounding and time-sensitive video QA tasks, indicating strong performance and generalization.", "review_text": "This paper introduces MUSEG, a reinforcement learning (RL) method to improve the fine-grained temporal understanding of Multimodal Large Language Models (MLLMs) in videos.\n\nThe core idea is timestamp-aware multi-segment grounding, which enables the model to align a query with multiple relevant video segments. This is optimized through a customized RL training recipe using phased rewards that progressively guides the model's reasoning.\n\nExperiments show MUSEG significantly outperforms existing methods on temporal grounding and time-sensitive video QA tasks, indicating strong performance and generalization.", "strengths": "With clear motivation and intuition, the paper uses GRPO and introduces two new reward functions, allowing the model to learn how to correctly segment events within videos for better temporal understanding. Especially in the segment matching reward, the authors meticulously design a function that computes the IoU of the predicted timestamps compared to the ground truth, a sound and wise choice.   Finetuning models with this method demonstrates improvements on selected temporal benchmarks. The authors also conducted exhaustive ablations in an attempt to justify different training strategies and reward design.", "weaknesses": "1. Though the authors presented promising results on 4 different benchmarks, it would be great if they could evaluate on more temporal benchmarks, such as TempCompass [1], TemporalBench, Vinoground [2]'s video score (which contains multiple segments), and so on, if time permits.\n2. In Section 4.2.2, the timestamp reward is designed to be 1 if all timestamps are correct, and 0 if any are wrong. This design is concerning to me because I do not understand why this reward is so strict; for segment matching the authors applied a very lenient strategy involving IoU, why won't one consider giving partial credit to the timestamp reward as well? I believe this might be the reason why the authors had to remove timestamp reward after 400 steps.\n3. In Section 6.2, the authors stated that \"the model can continue to freely explore more effective reasoning strategies\" by removing the timestamp reward mid-training. Though the improvements quantitatively are acknowledged, there is no qualitative result demonstrating the difference in reasoning between removing and not removing the reward.\n4. Accompanied by the above doubts about the reward function, which is the core contribution of this paper, I hate to say that I don't think the paper has enough novelty. But this idea is subject to change depending on the rebuttal for the above points.\n\n[1] Liu et al, 2024, TempCompass: Do Video LLMs Really Understand Videos?\n\n[2] Zhang et al, 2024, Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces MUSEG, a reinforcement learning (RL) method to improve the fine-grained temporal understanding of Multimodal Large Language Models (MLLMs) in videos.\n\nThe core idea is timestamp-aware multi-segment grounding, which enables the model to align a query with multiple relevant video segments. This is optimized through a customized RL training recipe using phased rewards that progressively guides the model's reasoning.\n\nExperiments show MUSEG significantly outperforms existing methods on temporal grounding and time-sensitive video QA tasks, indicating strong performance and generalization.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "With clear motivation and intuition, the paper uses GRPO and introduces two new reward functions, allowing the model to learn how to correctly segment events within videos for better temporal understanding. Especially in the segment matching reward, the authors meticulously design a function that computes the IoU of the predicted timestamps compared to the ground truth, a sound and wise choice.   Finetuning models with this method demonstrates improvements on selected temporal benchmarks. The authors also conducted exhaustive ablations in an attempt to justify different training strategies and reward design.", "weaknesses": "1. Though the authors presented promising results on 4 different benchmarks, it would be great if they could evaluate on more temporal benchmarks, such as TempCompass [1], TemporalBench, Vinoground [2]'s video score (which contains multiple segments), and so on, if time permits.\n2. In Section 4.2.2, the timestamp reward is designed to be 1 if all timestamps are correct, and 0 if any are wrong. This design is concerning to me because I do not understand why this reward is so strict; for segment matching the authors applied a very lenient strategy involving IoU, why won't one consider giving partial credit to the timestamp reward as well? I believe this might be the reason why the authors had to remove timestamp reward after 400 steps.\n3. In Section 6.2, the authors stated that \"the model can continue to freely explore more effective reasoning strategies\" by removing the timestamp reward mid-training. Though the improvements quantitatively are acknowledged, there is no qualitative result demonstrating the difference in reasoning between removing and not removing the reward.\n4. Accompanied by the above doubts about the reward function, which is the core contribution of this paper, I hate to say that I don't think the paper has enough novelty. But this idea is subject to change depending on the rebuttal for the above points.\n\n[1] Liu et al, 2024, TempCompass: Do Video LLMs Really Understand Videos?\n\n[2] Zhang et al, 2024, Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760551132304}], "openreview_url": "https://openreview.net/forum?id=jimY7NF6aJ", "arxiv_id": "2505.20715", "paper_pdf": "papers/jimY7NF6aJ.pdf", "paper_pdf_sha256": "0a1bcf0323282b28113d75da834ad10103ede1f83ceef716b544035e4bfb5e7c", "paper_pdf_bytes": 1219029, "paper_pdf_source": "openreview", "code_url": "https://github.com/THUNLP-MT/MUSEG", "code_repository": "THUNLP-MT/MUSEG", "code_commit": "889565e78b21e95f9ee2d2f6d250c3248b1587bb", "code_archive": "repos/jimY7NF6aJ.zip", "code_archive_sha256": "dc020eedecddfe0ff4144a41437edb96bd1c268d2655133d490878ef655f80ae", "code_archive_bytes": 2336496, "code_file_count": 14, "code_extensions": {".sh": 8, ".py": 6}, "github_disk_usage_kb": 2276, "github_languages": {"Python": 89374, "Shell": 3596}, "github_archived": false, "github_pushed_at": "2025-06-09T04:32:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/museg-reinforcing-video-temporal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bqUsdBeRjQ", "year": 2025, "status": "rejected", "title": "Personalized Language Modeling from Personalized Human Feedback", "authors": ["Xinyu Li", "Ruiyang Zhou", "Zachary Chase Lipton", "Liu Leqi"], "authorids": ["~Xinyu_Li7", "~Ruiyang_Zhou1", "~Zachary_Chase_Lipton1", "~Liu_Leqi1"], "authors_source": "OpenReview API", "abstract": "Personalized large language models (LLMs)  are designed to tailor responses to individual user preferences. While Reinforcement Learning from Human Feedback (RLHF) is a commonly used framework for aligning LLMs with human preferences, vanilla RLHF assumes that all human preferences share the same distribution, preventing fine-tuned LLMs from generating personalized content when user preferences are diverse. In this work, we propose Personalized-RLHF (P-RLHF), an efficient framework that utilizes a lightweight user model to capture individual user preferences and jointly learns the user model and the personalized LLM from human feedback. P-RLHF exhibits the following three characteristics: (1) It enables an LLM to generate personalized content and scale efficiently with growing number of users. (2) It handles both explicit user preferences described as textual input and implicit user preferences encoded in the feedback data. (3) It eliminates the need for users to fully articulate their preferences, which are normally needed for prompting LLMs to generate personalized content yet are often impractical to obtain in real-world scenarios. Our experimental results show that personalized LLMs trained using P-RLHF generate responses that are more closely aligned with individual user preferences, outperforming vanilla, non-personalized RLHF and prompting-based personalization approaches across different tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "fOoUndJSOE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12971/Reviewer_rjtj"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces Personalized-RLHF (P-RLHF), an innovative framework designed to enhance personalized content generation in large language models (LLMs) by leveraging a lightweight user model. Its key contributions include:\nScalability: P-RLHF allows LLMs to efficiently generate personalized content while accommodating a growing user base.\nPreference Handling: The framework effectively processes both explicit user preferences, provided as textual input, and implicit preferences derived from feedback data.\nReduced User Burden: It alleviates the requirement for users to fully articulate their preferences, addressing the challenges of obtaining detailed user input in practical applications.\n\nOverall, this paper is easy to follow, but there are issues such as a lack of innovation, no released code, and an excess of symbols. However, overall, this is a ready piece of work.", "review_text": "The paper introduces Personalized-RLHF (P-RLHF), an innovative framework designed to enhance personalized content generation in large language models (LLMs) by leveraging a lightweight user model. Its key contributions include:\nScalability: P-RLHF allows LLMs to efficiently generate personalized content while accommodating a growing user base.\nPreference Handling: The framework effectively processes both explicit user preferences, provided as textual input, and implicit preferences derived from feedback data.\nReduced User Burden: It alleviates the requirement for users to fully articulate their preferences, addressing the challenges of obtaining detailed user input in practical applications.\n\nOverall, this paper is easy to follow, but there are issues such as a lack of innovation, no released code, and an excess of symbols. However, overall, this is a ready piece of work.", "strengths": "* The structure of this article is well-organized and clearly articulated, making it easy for readers to follow the flow of ideas and concepts presented throughout the text. Each section is logically arranged, allowing for a seamless understanding of the material.\n\n* The section on RLHF is detailed, with a clear progression from general RLHF to the specifically designed personalized DPO. This work considers different user preferences from multiple angles and granularities. It clearly explains the differences between P-RLHF and Vanilla DPO to clarify the motivation.\n\n* The work uses the TL;DR dataset, which has Reddit posts, summaries, and worker IDs. It conducts an experiment where most workers (70%) prefer longer responses and a smaller group (30%) prefers shorter ones. The top 10 workers with the most annotations were included for training. After the experiment, some workers liked shorter responses while others liked longer ones. Overall, this shows a strong effort and good understanding of data analysis, which is impressive.\n\n* There is a solid exploration of how personalized LLMs can adapt to different user preferences through P-DPO fine-tuning. The study found that the personalized model can generate longer or shorter responses based on users' implicit preferences, demonstrating its flexibility and adaptability. The experiment also compared the performance of P-DPO using generic user embeddings with traditional vanilla DPO, showing that the personalized model excels in understanding and responding to user needs, validating its effectiveness in practical applications.", "weaknesses": "* It seems that this article does not have released code. I'm not sure if I just couldn't find it, but without released code, reproducibility cannot be guaranteed.\n\n* Like other RL works, the notation in this paper is too numerous and complex. Although the structure and presentation of the article are good, it somewhat hinders the readers' understanding. Perhaps a table to organize the notations could be helpful.\n\n* There are concerns regarding the effectiveness of not training an additional reward model and relying solely on the user model.", "questions": "* Does this work have released code and dataset?\n* There are works have been done on personalized user preferences using RLHF. How does this one differ from them, and is there a comparison used as a baseline?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Personalized-RLHF (P-RLHF), an innovative framework designed to enhance personalized content generation in large language models (LLMs) by leveraging a lightweight user model. Its key contributions include:\nScalability: P-RLHF allows LLMs to efficiently generate personalized content while accommodating a growing user base.\nPreference Handling: The framework effectively processes both explicit user preferences, provided as textual input, and implicit preferences derived from feedback data.\nReduced User Burden: It alleviates the requirement for users to fully articulate their preferences, addressing the challenges of obtaining detailed user input in practical applications.\n\nOverall, this paper is easy to follow, but there are issues such as a lack of innovation, no released code, and an excess of symbols. However, overall, this is a ready piece of work.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* The structure of this article is well-organized and clearly articulated, making it easy for readers to follow the flow of ideas and concepts presented throughout the text. Each section is logically arranged, allowing for a seamless understanding of the material.\n\n* The section on RLHF is detailed, with a clear progression from general RLHF to the specifically designed personalized DPO. This work considers different user preferences from multiple angles and granularities. It clearly explains the differences between P-RLHF and Vanilla DPO to clarify the motivation.\n\n* The work uses the TL;DR dataset, which has Reddit posts, summaries, and worker IDs. It conducts an experiment where most workers (70%) prefer longer responses and a smaller group (30%) prefers shorter ones. The top 10 workers with the most annotations were included for training. After the experiment, some workers liked shorter responses while others liked longer ones. Overall, this shows a strong effort and good understanding of data analysis, which is impressive.\n\n* There is a solid exploration of how personalized LLMs can adapt to different user preferences through P-DPO fine-tuning. The study found that the personalized model can generate longer or shorter responses based on users' implicit preferences, demonstrating its flexibility and adaptability. The experiment also compared the performance of P-DPO using generic user embeddings with traditional vanilla DPO, showing that the personalized model excels in understanding and responding to user needs, validating its effectiveness in practical applications.", "weaknesses": "* It seems that this article does not have released code. I'm not sure if I just couldn't find it, but without released code, reproducibility cannot be guaranteed.\n\n* Like other RL works, the notation in this paper is too numerous and complex. Although the structure and presentation of the article are good, it somewhat hinders the readers' understanding. Perhaps a table to organize the notations could be helpful.\n\n* There are concerns regarding the effectiveness of not training an additional reward model and relying solely on the user model.", "questions": "* Does this work have released code and dataset?\n* There are works have been done on personalized user preferences using RLHF. How does this one differ from them, and is there a comparison used as a baseline?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730723112534}, {"id": "D1i6eTfpp1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12971/Reviewer_oVq3"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper addresses the important challenge of generalizing large language models (LLMs) to capture personalized user preferences. To achieve this, the authors develop several models that incorporate users' implicit preferences, along with a new learning objective aimed at optimizing the reference LLMs for personalization. They compare the proposed method, P-DPO, with the standard DPO in both synthetic and human evaluation settings. In both cases, P-DPO shows advantages. The paper is well-written, with a clearly motivated problem statement. The discussion on the limitations of existing RLHF methods is thorough and insightful.\n\nHowever, the reviewer has the following concerns about this work.\n1. The contribution appears limited. The primary innovation, the new learning objective P-DPO introduced in Section 4.4, facilitates personalization by incorporating user embeddings into the implicit reward model of DPO. However, this approach is not particularly novel, as the integration of user embeddings for personalization is extensively discussed in the recommendation literature [1].\n2. The evaluation lacks comprehensiveness. The authors only consider DPO as a baseline and do not include recent work on personalized LLMs, such as [2] and [3]. Additional relevant studies can be found at https://github.com/HqWu-HITCS/Awesome-Personalized-LLM. Furthermore, the experiments overlook the recommendation task, which is a crucial application for personalized LLMs. In the human evaluation, the sample size is limited to at most 20 users and 76 samples, which may lead to biased results and unreliable conclusions.\n3. The advantages of integrating both users' explicit and implicit embeddings are not well supported. The authors are suggested to include an ablation study to validate the importance of the learned implicit user embeddings.\n4. The motivation for including the user-identifier-agnostic loss term in P-DPO is not clear to the reviewer. The advantage of this design is also not evident.\n\n[1] SSE-PT: Sequential Recommendation Via Personalized Transformer\n[2] Factual and Tailored Recommendation Endorsements using Language Models and Reinforcement Learning\n[3] Personalized Large Language Models", "review_text": "This paper addresses the important challenge of generalizing large language models (LLMs) to capture personalized user preferences. To achieve this, the authors develop several models that incorporate users' implicit preferences, along with a new learning objective aimed at optimizing the reference LLMs for personalization. They compare the proposed method, P-DPO, with the standard DPO in both synthetic and human evaluation settings. In both cases, P-DPO shows advantages. The paper is well-written, with a clearly motivated problem statement. The discussion on the limitations of existing RLHF methods is thorough and insightful.\n\nHowever, the reviewer has the following concerns about this work.\n1. The contribution appears limited. The primary innovation, the new learning objective P-DPO introduced in Section 4.4, facilitates personalization by incorporating user embeddings into the implicit reward model of DPO. However, this approach is not particularly novel, as the integration of user embeddings for personalization is extensively discussed in the recommendation literature [1].\n2. The evaluation lacks comprehensiveness. The authors only consider DPO as a baseline and do not include recent work on personalized LLMs, such as [2] and [3]. Additional relevant studies can be found at https://github.com/HqWu-HITCS/Awesome-Personalized-LLM. Furthermore, the experiments overlook the recommendation task, which is a crucial application for personalized LLMs. In the human evaluation, the sample size is limited to at most 20 users and 76 samples, which may lead to biased results and unreliable conclusions.\n3. The advantages of integrating both users' explicit and implicit embeddings are not well supported. The authors are suggested to include an ablation study to validate the importance of the learned implicit user embeddings.\n4. The motivation for including the user-identifier-agnostic loss term in P-DPO is not clear to the reviewer. The advantage of this design is also not evident.\n\n[1] SSE-PT: Sequential Recommendation Via Personalized Transformer\n[2] Factual and Tailored Recommendation Endorsements using Language Models and Reinforcement Learning\n[3] Personalized Large Language Models", "strengths": "1. The issues related to DPO in personalized modeling are thoroughly discussed.\n2. The presentation is easy to follow and well-structured.\n3. In both synthetic and human evaluations, P-DPO significantly outperforms DPO.", "weaknesses": "For a detailed discussion of the reviewers' concerns, please refer to the summary.", "questions": "1. Why include the user-identifier-agnostic loss term in P-DPO?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the important challenge of generalizing large language models (LLMs) to capture personalized user preferences. To achieve this, the authors develop several models that incorporate users' implicit preferences, along with a new learning objective aimed at optimizing the reference LLMs for personalization. They compare the proposed method, P-DPO, with the standard DPO in both synthetic and human evaluation settings. In both cases, P-DPO shows advantages. The paper is well-written, with a clearly motivated problem statement. The discussion on the limitations of existing RLHF methods is thorough and insightful.\n\nHowever, the reviewer has the following concerns about this work.\n1. The contribution appears limited. The primary innovation, the new learning objective P-DPO introduced in Section 4.4, facilitates personalization by incorporating user embeddings into the implicit reward model of DPO. However, this approach is not particularly novel, as the integration of user embeddings for personalization is extensively discussed in the recommendation literature [1].\n2. The evaluation lacks comprehensiveness. The authors only consider DPO as a baseline and do not include recent work on personalized LLMs, such as [2] and [3]. Additional relevant studies can be found at https://github.com/HqWu-HITCS/Awesome-Personalized-LLM. Furthermore, the experiments overlook the recommendation task, which is a crucial application for personalized LLMs. In the human evaluation, the sample size is limited to at most 20 users and 76 samples, which may lead to biased results and unreliable conclusions.\n3. The advantages of integrating both users' explicit and implicit embeddings are not well supported. The authors are suggested to include an ablation study to validate the importance of the learned implicit user embeddings.\n4. The motivation for including the user-identifier-agnostic loss term in P-DPO is not clear to the reviewer. The advantage of this design is also not evident.\n\n[1] SSE-PT: Sequential Recommendation Via Personalized Transformer\n[2] Factual and Tailored Recommendation Endorsements using Language Models and Reinforcement Learning\n[3] Personalized Large Language Models", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The issues related to DPO in personalized modeling are thoroughly discussed.\n2. The presentation is easy to follow and well-structured.\n3. In both synthetic and human evaluations, P-DPO significantly outperforms DPO.", "weaknesses": "For a detailed discussion of the reviewers' concerns, please refer to the summary.", "questions": "1. Why include the user-identifier-agnostic loss term in P-DPO?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730595802887}, {"id": "ZXZTbOyBpa", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12971/Reviewer_uiaG"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces Personalized Reinforcement Learning from Human Feedback (P-RLHF), a novel framework designed to enhance the personalization capabilities of large language models (LLMs) by integrating both explicit and implicit user models. By leveraging user-specific information and feedback, P-RLHF effectively captures diverse and individualized user preferences, addressing the limitations of traditional RLHF methods that assume uniform user preferences.\nThe authors validate the efficacy of the P-RLHF framework through a series of comprehensive experiments, ranging from controlled synthetic settings to large-scale real-world datasets. These experiments demonstrate that P-RLHF not only successfully captures and adheres to individual user preferences but also maintains high performance and scalability in diverse application scenarios. The clear and logical presentation of the framework, coupled with intuitive response generation results, underscores the practical value and potential impact of personalized LLMs in real-world applications.\nOverall, this paper makes substantial contributions to the field of personalized LLM by presenting a well-structured framework and a novel optimization method that together advances the state-of-the-art in personalized language model training.", "review_text": "This paper introduces Personalized Reinforcement Learning from Human Feedback (P-RLHF), a novel framework designed to enhance the personalization capabilities of large language models (LLMs) by integrating both explicit and implicit user models. By leveraging user-specific information and feedback, P-RLHF effectively captures diverse and individualized user preferences, addressing the limitations of traditional RLHF methods that assume uniform user preferences.\nThe authors validate the efficacy of the P-RLHF framework through a series of comprehensive experiments, ranging from controlled synthetic settings to large-scale real-world datasets. These experiments demonstrate that P-RLHF not only successfully captures and adheres to individual user preferences but also maintains high performance and scalability in diverse application scenarios. The clear and logical presentation of the framework, coupled with intuitive response generation results, underscores the practical value and potential impact of personalized LLMs in real-world applications.\nOverall, this paper makes substantial contributions to the field of personalized LLM by presenting a well-structured framework and a novel optimization method that together advances the state-of-the-art in personalized language model training.", "strengths": "1. The paper presents the P-RLHF framework, which integrates both explicit and implicit user models to facilitate personalized learning from user feedback, thereby introducing a novel approach to the research on personalized LLMs. The introduction of P-DPO represents an innovative optimization method that effectively balances personalization and generalization for both known and unknown users. This approach demonstrates excellent scalability while addressing the personalized needs of multiple users, marking a significant advancement over traditional RLHF methods.\n2. The overall structure of the paper is exceptionally clear, methodically introducing the background, identifying the problem, proposing solutions, and validating the experiments related to personalized RLHF. The logical progression facilitates easy comprehension for readers. Each experimental section is meticulously designed with clearly defined objectives, enabling readers to effortlessly track the innovative aspects of the methodology and their corresponding outcomes.\n3. The response generation results provided in the paper offer a clear and intuitive demonstration of how different user preferences are handled in a personalized manner. The personalized LLMs showcased in the study hold significant practical value, and the paper successfully achieves personalized content generation on the utilized datasets. This effectively highlights the framework’s capability to cater to diverse user preferences in real-world applications.", "weaknesses": "1. Zero-Length Responses in Experiment One: The occurrence of zero-length responses, which the authors justify as an expected outcome via mathematical proofs, raises questions about whether an LLM should indeed produce responses with no content. This result appears to stem from a polarized experimental design and methodology, which may not reflect practical application scenarios or user expectations. A more nuanced experimental setup could better balance realistic use cases with the need for distinct preference modeling.\n2. Figure 1 Lacks Identification of Key Modules: While Figure 1 presents the overall framework, it does not clearly indicate critical components introduced by the authors, such as P-DPO. And there is no explanation of Figure 1 in the main text. This omission complicates the interpretation of the framework diagram, hindering the reader’s ability to intuitively grasp the structure and contributions of the proposed approach.\n3. Insufficient Ablation Study: The ablation experiments, which appear only in the appendix, focus solely on parameter variations rather than examining which specific design choices contribute to the improved performance over baseline models. For example, it is unclear whether the personalized capabilities remain robust when only the implicit model is used without the explicit component. The paper lacks a thorough investigation into such aspects, leaving questions about the impact of each module unanswered.\n4. Binary Nature of Evaluation Metrics: Both the preference modeling and experimental evaluation adopt a highly binary approach. In modeling preferences, options are presented as direct opposites (e.g., long vs. short responses), which likely contributes to the appearance of zero-length responses. Furthermore, evaluation is limited to win-rates relative to baseline models, which does not provide a clear view of the actual performance gains. Since the goal of personalization is to improve user satisfaction, more nuanced analyses, rather than a strictly binary approach to the generated content, are necessary to substantiate the effectiveness of the approach.\n5. Lack of Consideration for Contextual Shifts in User Preferences: User preferences can be context-dependent, particularly when involving multidimensional combinations of preferences. In specific scenarios, user preferences may shift according to immediate needs or context. The reliance on a static dataset overlooks these potential dynamics, and the paper does not address how such variations might impact model performance.", "questions": "1. An explanation is requested for the occurrence of zero-length responses presented in Section 5.2 and Figure 3. Logically, even for users who prefer concise replies, it is expected that responses maintain a reasonable lower bound on length rather than resulting in either no content or excessively long responses. The current approach appears overly polarized and does not account for practical application scenarios where balanced response lengths are desired.\n2. It is recommended that Figure 1, which outlines the proposed module design, be incorporated directly into the main body of the text. This integration would enhance the comprehensibility of the framework by ensuring that the figure is closely tied to the corresponding narrative, thereby making the overall framework easier to understand for readers.\n3. I suggest the author add more nuanced analyses, rather than a strictly binary approach to the generated content, which are necessary to substantiate the effectiveness of the approach.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Personalized Reinforcement Learning from Human Feedback (P-RLHF), a novel framework designed to enhance the personalization capabilities of large language models (LLMs) by integrating both explicit and implicit user models. By leveraging user-specific information and feedback, P-RLHF effectively captures diverse and individualized user preferences, addressing the limitations of traditional RLHF methods that assume uniform user preferences.\nThe authors validate the efficacy of the P-RLHF framework through a series of comprehensive experiments, ranging from controlled synthetic settings to large-scale real-world datasets. These experiments demonstrate that P-RLHF not only successfully captures and adheres to individual user preferences but also maintains high performance and scalability in diverse application scenarios. The clear and logical presentation of the framework, coupled with intuitive response generation results, underscores the practical value and potential impact of personalized LLMs in real-world applications.\nOverall, this paper makes substantial contributions to the field of personalized LLM by presenting a well-structured framework and a novel optimization method that together advances the state-of-the-art in personalized language model training.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. The paper presents the P-RLHF framework, which integrates both explicit and implicit user models to facilitate personalized learning from user feedback, thereby introducing a novel approach to the research on personalized LLMs. The introduction of P-DPO represents an innovative optimization method that effectively balances personalization and generalization for both known and unknown users. This approach demonstrates excellent scalability while addressing the personalized needs of multiple users, marking a significant advancement over traditional RLHF methods.\n2. The overall structure of the paper is exceptionally clear, methodically introducing the background, identifying the problem, proposing solutions, and validating the experiments related to personalized RLHF. The logical progression facilitates easy comprehension for readers. Each experimental section is meticulously designed with clearly defined objectives, enabling readers to effortlessly track the innovative aspects of the methodology and their corresponding outcomes.\n3. The response generation results provided in the paper offer a clear and intuitive demonstration of how different user preferences are handled in a personalized manner. The personalized LLMs showcased in the study hold significant practical value, and the paper successfully achieves personalized content generation on the utilized datasets. This effectively highlights the framework’s capability to cater to diverse user preferences in real-world applications.", "weaknesses": "1. Zero-Length Responses in Experiment One: The occurrence of zero-length responses, which the authors justify as an expected outcome via mathematical proofs, raises questions about whether an LLM should indeed produce responses with no content. This result appears to stem from a polarized experimental design and methodology, which may not reflect practical application scenarios or user expectations. A more nuanced experimental setup could better balance realistic use cases with the need for distinct preference modeling.\n2. Figure 1 Lacks Identification of Key Modules: While Figure 1 presents the overall framework, it does not clearly indicate critical components introduced by the authors, such as P-DPO. And there is no explanation of Figure 1 in the main text. This omission complicates the interpretation of the framework diagram, hindering the reader’s ability to intuitively grasp the structure and contributions of the proposed approach.\n3. Insufficient Ablation Study: The ablation experiments, which appear only in the appendix, focus solely on parameter variations rather than examining which specific design choices contribute to the improved performance over baseline models. For example, it is unclear whether the personalized capabilities remain robust when only the implicit model is used without the explicit component. The paper lacks a thorough investigation into such aspects, leaving questions about the impact of each module unanswered.\n4. Binary Nature of Evaluation Metrics: Both the preference modeling and experimental evaluation adopt a highly binary approach. In modeling preferences, options are presented as direct opposites (e.g., long vs. short responses), which likely contributes to the appearance of zero-length responses. Furthermore, evaluation is limited to win-rates relative to baseline models, which does not provide a clear view of the actual performance gains. Since the goal of personalization is to improve user satisfaction, more nuanced analyses, rather than a strictly binary approach to the generated content, are necessary to substantiate the effectiveness of the approach.\n5. Lack of Consideration for Contextual Shifts in User Preferences: User preferences can be context-dependent, particularly when involving multidimensional combinations of preferences. In specific scenarios, user preferences may shift according to immediate needs or context. The reliance on a static dataset overlooks these potential dynamics, and the paper does not address how such variations might impact model performance.", "questions": "1. An explanation is requested for the occurrence of zero-length responses presented in Section 5.2 and Figure 3. Logically, even for users who prefer concise replies, it is expected that responses maintain a reasonable lower bound on length rather than resulting in either no content or excessively long responses. The current approach appears overly polarized and does not account for practical application scenarios where balanced response lengths are desired.\n2. It is recommended that Figure 1, which outlines the proposed module design, be incorporated directly into the main body of the text. This integration would enhance the comprehensibility of the framework by ensuring that the figure is closely tied to the corresponding narrative, thereby making the overall framework easier to understand for readers.\n3. I suggest the author add more nuanced analyses, rather than a strictly binary approach to the generated content, which are necessary to substantiate the effectiveness of the approach.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730479379324}, {"id": "o4lsx8AdIM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12971/Reviewer_hQjp"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces a framework called Personalized Reinforcement Learning from Human Feedback (P-RLHF) aimed at personalizing large language models (LLMs) based on individual user preferences. The authors propose a lightweight user model that jointly learns user preferences (both explicit and implicit) and the LLM. The key claims are that this framework scales efficiently with increasing users, handles diverse feedback, and reduces the need for explicit user prompts.", "review_text": "This paper introduces a framework called Personalized Reinforcement Learning from Human Feedback (P-RLHF) aimed at personalizing large language models (LLMs) based on individual user preferences. The authors propose a lightweight user model that jointly learns user preferences (both explicit and implicit) and the LLM. The key claims are that this framework scales efficiently with increasing users, handles diverse feedback, and reduces the need for explicit user prompts.", "strengths": "- The paper addresses an important area of research—personalizing LLMs based on user feedback.\n- The proposal to use both explicit and implicit feedback for personalization is valuable and well-motivated.\n- The experiments cover several datasets, which helps validate the general applicability of the approach.", "weaknesses": "- The paper does not sufficiently address how implicit feedback is effectively used or how it handles contradictory user preferences.\n- The scalability claims, while central to the paper, are not backed up by detailed computational analysis or resource cost comparisons with other approaches.\n- There is limited discussion on how this approach performs in real-world dynamic environments where user preferences might shift over time.\n- The experimental results, while positive, are not sufficiently compelling to clearly demonstrate the advantage of P-RLHF over simpler personalization methods.", "questions": "1. How does the model ensure that implicit feedback does not conflict with explicit user preferences, and how does it resolve such conflicts?\n2. Can the authors provide more detailed information on the computational efficiency of P-RLHF? How does it compare in terms of time and memory consumption to other RLHF approaches?\n3. How does the framework handle dynamic user preferences that may change over time, and how quickly can it adapt?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a framework called Personalized Reinforcement Learning from Human Feedback (P-RLHF) aimed at personalizing large language models (LLMs) based on individual user preferences. The authors propose a lightweight user model that jointly learns user preferences (both explicit and implicit) and the LLM. The key claims are that this framework scales efficiently with increasing users, handles diverse feedback, and reduces the need for explicit user prompts.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The paper addresses an important area of research—personalizing LLMs based on user feedback.\n- The proposal to use both explicit and implicit feedback for personalization is valuable and well-motivated.\n- The experiments cover several datasets, which helps validate the general applicability of the approach.", "weaknesses": "- The paper does not sufficiently address how implicit feedback is effectively used or how it handles contradictory user preferences.\n- The scalability claims, while central to the paper, are not backed up by detailed computational analysis or resource cost comparisons with other approaches.\n- There is limited discussion on how this approach performs in real-world dynamic environments where user preferences might shift over time.\n- The experimental results, while positive, are not sufficiently compelling to clearly demonstrate the advantage of P-RLHF over simpler personalization methods.", "questions": "1. How does the model ensure that implicit feedback does not conflict with explicit user preferences, and how does it resolve such conflicts?\n2. Can the authors provide more detailed information on the computational efficiency of P-RLHF? How does it compare in terms of time and memory consumption to other RLHF approaches?\n3. How does the framework handle dynamic user preferences that may change over time, and how quickly can it adapt?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730347100304}], "openreview_url": "https://openreview.net/forum?id=bqUsdBeRjQ", "arxiv_id": "2402.05133", "paper_pdf": "papers/bqUsdBeRjQ.pdf", "paper_pdf_sha256": "b118110a056ec4199f663b2a38668714995b0f8f2c095d91423c13f9104da245", "paper_pdf_bytes": 576712, "paper_pdf_source": "openreview", "code_url": "https://github.com/HumainLab/Personalized_RLHF", "code_repository": "HumainLab/Personalized_RLHF", "code_commit": "ae1f81bf8a344cbdfa6bc4e8dd6ce7823757db9f", "code_archive": "repos/bqUsdBeRjQ.zip", "code_archive_sha256": "12842b18dc934a827ef6d43597fc6ffc46a7009d44ae3dcee2bb69b78d329ce6", "code_archive_bytes": 233944, "code_file_count": 61, "code_extensions": {".py": 43, ".sh": 14, ".ipynb": 4}, "github_disk_usage_kb": 167, "github_languages": {"Python": 326357, "Jupyter Notebook": 56604, "Shell": 14440}, "github_archived": false, "github_pushed_at": "2024-12-09T05:44:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/personalized-language-modeling-from"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gUL6zYN4Uaf", "year": 2023, "status": "rejected", "title": "Cramming: Training a language model on a single GPU in one day", "authors": ["Jonas Geiping", "Tom Goldstein"], "authorids": ["~Jonas_Geiping1", "~Tom_Goldstein1"], "authors_source": "OpenReview API", "abstract": "Recent trends in language modeling have focused on increasing performance through scaling, and have resulted in an environment where training language models is out of reach for most researchers and practitioners.  While most in the community are asking how to push the limits of extreme computation, we ask the opposite question:  \nHow far can we get with a single GPU in just one day? \nWe investigate the downstream performance achievable with a transformer-based language model trained completely from scratch with masked language modeling for a single day on a single consumer GPU.\nAside from re-analyzing nearly all components of the pretraining pipeline for this scenario and providing a modified pipeline with performance close to BERT, we investigate why scaling down is hard, and which modifications actually improve performance in this scenario. We provide evidence that even in this constrained setting, performance closely follows scaling laws observed in large-compute settings. Through the lens of scaling laws, we categorize a range of recent improvements to training and architecture and discuss their merit and practical applicability (or lack thereof) for the limited compute setting.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JJy5zDi_UNr", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5456/Reviewer_Fi67"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors study the performance of transformer models on downstream tasks as the total computational budget is decreased.  This process, known as cramming in the paper, turns the problem of training these enormous language models in a new direction from the typical scenario used in industrial labs that train models on a seemingly endless supply of resources. The author's place and exception small limit on the total computation that is allowed to train a transformer model from scratch to the total FLOPs available on a single GPU in 24 hours. By considering the scaling laws of large model transformers the authors mainly investigate training setups that keep the total number of parameters in the model constant but reduce the cost of performing a gradient update. By enumerating a small number of interesting features of the transformer training design space the authors demonstrate that cramming can achieve interesting and sometimes comparable results with larger models using more computation in particular settings and for particular datasets.", "review_text": "Overall I find the motivation for the work and claims made by the authors to be an interesting departure from the traditional language training papers that use exorbitant computational resources. It seems more practical to answer questions about how researchers can do more with less when it comes to allocating resources for training transformer models.\n\nMy remarks should be taken with a grain of salt as I am not an expert in this particular area but I would feel more inclined to experiment with transformer models if I felt I could train them to a reasonable level of ability on my modest desktop setup. I believe this sentiment represents the spirit of the paper and the results should be of interest to other members of the research community that are hesitant to participate in this research area because of the perceived computational overheads.", "strengths": "Strengths:\n- The motivation for the study proposed in the paper is interesting for a number of reasons. The volume of computation required by many modern transformer models has been prohibitively expensive and therefore out of reach for most researchers for quite a while. By studying the implications of constraining the computational resources on the ability of the model to perform well on certain tasks the authors could provide a way for researchers with limited budgets to participate and utilize these models in fundamentally new ways.\n- The trend in the paper to consider modifications that mainly reduce the gradient update cost without significantly impacting the total number of parameters in the model, based on the scaling laws, provides an interesting and unifying theme throughout. The persistence of the scaling laws to influence the performance of the model on tasks is reinforced through empirical evidence throughout and yields interesting insights.\n- Performance evaluation on a shoe-string budget of FLOPs compare to other prominent models is impressive.\n\nWeaknesses:\n- Similar studies were conducted on a single node with 8 GPUs as noted by the authors. Though that setup had considerably more computational resources the total volume of computation was still a fraction of the amount used by many large research institutions. In light of that work, the scenario presented in this paper may seem somewhat derivative and only marginally interesting.\n- It is not clear if or how the observations made in the cramming regime may be used to make more informed decisions regarding the training process in the normal training setting.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the authors study the performance of transformer models on downstream tasks as the total computational budget is decreased.  This process, known as cramming in the paper, turns the problem of training these enormous language models in a new direction from the typical scenario used in industrial labs that train models on a seemingly endless supply of resources. The author's place and exception small limit on the total computation that is allowed to train a transformer model from scratch to the total FLOPs available on a single GPU in 24 hours. By considering the scaling laws of large model transformers the authors mainly investigate training setups that keep the total number of parameters in the model constant but reduce the cost of performing a gradient update. By enumerating a small number of interesting features of the transformer training design space the authors demonstrate that cramming can achieve interesting and sometimes comparable results with larger models using more computation in particular settings and for particular datasets.", "strength_and_weaknesses": "Strengths:\n- The motivation for the study proposed in the paper is interesting for a number of reasons. The volume of computation required by many modern transformer models has been prohibitively expensive and therefore out of reach for most researchers for quite a while. By studying the implications of constraining the computational resources on the ability of the model to perform well on certain tasks the authors could provide a way for researchers with limited budgets to participate and utilize these models in fundamentally new ways.\n- The trend in the paper to consider modifications that mainly reduce the gradient update cost without significantly impacting the total number of parameters in the model, based on the scaling laws, provides an interesting and unifying theme throughout. The persistence of the scaling laws to influence the performance of the model on tasks is reinforced through empirical evidence throughout and yields interesting insights.\n- Performance evaluation on a shoe-string budget of FLOPs compare to other prominent models is impressive.\n\nWeaknesses:\n- Similar studies were conducted on a single node with 8 GPUs as noted by the authors. Though that setup had considerably more computational resources the total volume of computation was still a fraction of the amount used by many large research institutions. In light of that work, the scenario presented in this paper may seem somewhat derivative and only marginally interesting.\n- It is not clear if or how the observations made in the cramming regime may be used to make more informed decisions regarding the training process in the normal training setting.", "clarity,_quality,_novelty_and_reproducibility": "The writing is clear and the presentation of issues motivating the current work is adequately articulated in the text. While I am not an expert in the transformer field I feel the authors did a good job explaining the connection between the scaling laws and the downstream performance of the models under consideration. The novelty of the work pertains to the training strategies used to reduce computational costs without removing the total number of model parameters. Although previous works looked at training with limited resources the author's study and extreme training scenario that is likely to be more pertinent and representative of the resources available to typical, non-institutional, researchers.", "summary_of_the_review": "Overall I find the motivation for the work and claims made by the authors to be an interesting departure from the traditional language training papers that use exorbitant computational resources. It seems more practical to answer questions about how researchers can do more with less when it comes to allocating resources for training transformer models.\n\nMy remarks should be taken with a grain of salt as I am not an expert in this particular area but I would feel more inclined to experiment with transformer models if I felt I could train them to a reasonable level of ability on my modest desktop setup. I believe this sentiment represents the spirit of the paper and the results should be of interest to other members of the research community that are hesitant to participate in this research area because of the perceived computational overheads.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666792695709}, {"id": "LcmGfQxOYO3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5456/Reviewer_x7E8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigated pretraining a mask language model in a resource-constrained setting, i.e. a single GPU for one day. The authors empirically tested various architectural and data changes in order to maximize performance. There are some interesting findings, such as per-gradient efficiency only depends on model size, several strategies to filter and sort the training data brought improvement, etc. As a results, the authors were able to push the performance close to BERT base if excluding some outlier tasks.", "review_text": "Overall, this paper tackles an important problem, the experiments design are sound and the empirical findings are informative for language models pretraining. If the authors could add more clarity to generalizability and robustness of the findings, e.g. experiments with sequence lengths longer than 128, further ablate on the causes of drop in CoLA, etc. then the results would be more valuable to language models pretraining. ", "strengths": "Strengths:\n- This paper adds more insights on scaling-down, which is less understood as most concurrent efforts are around scaling-up. \n- The experiments have a good coverage in terms of testing various architectural changes from recent literature.\n- The final performance on downstream tasks (GLUE) are impressive given the limited compute budget. \n  \nWeaknesses:\n- The biggest missing piece is an ablation study on the improvement from various architecture changes. In the current version, the authors provided a comprehensive list of things they tried, what helped and what didn't. But it's unknown which change(s) brought the bigger improvement.  \n- Related to this is the poor performance on CoLA. Although the authors provided several hypotheses, some of them should be tested to verify whether they're actually related to any of the architecture or data changes, or mostly due to reduced model size. For example, one possible cause provided by the authors is that reasonable performance CoLA would need more training data. But are models in these experiment trained on less data compared to BERT-base? Is it possible to see whether the performance gap can actually be closed if the models were trained longer than one day (i.e. seeing more data)?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper investigated pretraining a mask language model in a resource-constrained setting, i.e. a single GPU for one day. The authors empirically tested various architectural and data changes in order to maximize performance. There are some interesting findings, such as per-gradient efficiency only depends on model size, several strategies to filter and sort the training data brought improvement, etc. As a results, the authors were able to push the performance close to BERT base if excluding some outlier tasks.", "strength_and_weaknesses": "Strengths:\n- This paper adds more insights on scaling-down, which is less understood as most concurrent efforts are around scaling-up. \n- The experiments have a good coverage in terms of testing various architectural changes from recent literature.\n- The final performance on downstream tasks (GLUE) are impressive given the limited compute budget. \n  \nWeaknesses:\n- The biggest missing piece is an ablation study on the improvement from various architecture changes. In the current version, the authors provided a comprehensive list of things they tried, what helped and what didn't. But it's unknown which change(s) brought the bigger improvement.  \n- Related to this is the poor performance on CoLA. Although the authors provided several hypotheses, some of them should be tested to verify whether they're actually related to any of the architecture or data changes, or mostly due to reduced model size. For example, one possible cause provided by the authors is that reasonable performance CoLA would need more training data. But are models in these experiment trained on less data compared to BERT-base? Is it possible to see whether the performance gap can actually be closed if the models were trained longer than one day (i.e. seeing more data)?", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\n- For all the architectural modifications in Sec 4.2, does \"no improvement\" refer to pretraining loss or downstream tasks?\n- In Sec 4.3 batch size schedule, you found optimal performance from different batch size for pretraining (1536) and downstream tasks (4032). Why do you think pretraining loss benefit from smaller batch size? Similarly, could any of the architectural changes in Sec 4.2 have different effects on pretraining vs. downstream?\n- For all the changes in Sec 4.2 and 4.3, when you test a specific modification, what were used for the rest of the architecture and training setup? How were they chosen?\n\nQuality:\n- The 128 sequence length differs drastically from common choice in language models pretraining (e.g. 1024, or at least 512). To make sure conclusions from this work would apply, some additional experiments with longer sequence length would be helpful.\n", "summary_of_the_review": "Overall, this paper tackles an important problem, the experiments design are sound and the empirical findings are informative for language models pretraining. If the authors could add more clarity to generalizability and robustness of the findings, e.g. experiments with sequence lengths longer than 128, further ablate on the causes of drop in CoLA, etc. then the results would be more valuable to language models pretraining. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666679249276}, {"id": "VpEkt6zLidI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5456/Reviewer_xMcb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this paper, the paper investigate language model pipeline to see which modifications improve performance in the scaled-down scenario ( a single GPU for 24 hours).  ", "review_text": "This paper does a lot of Interesting investigations. ", "strengths": "Strength:\n1. In this paper, several modifications (architecture, training setup and datasets) are explored to check whether there is any improvement. All of these aspects are important and interesting. These can give good insights for the community. \n\n2. Some interesting conclusion are got, for example, training recipe and data setup lead to decent downstream performance on GLUE.\n\nWeaknesses:\n\n\n1. The investigation of modifications lack convincing experiments. For example, only one task performance (MNLI) is reported for when studying the impact of training hyper-parameters. Other tasks can have a different trend.  And when exploring the effect of the architecture, only MLM loss is report. The performance of downstream tasks can be also important. \n\n2. The technical novelty of this paper is a little limit. The total contributions are also limit. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the paper investigate language model pipeline to see which modifications improve performance in the scaled-down scenario ( a single GPU for 24 hours).  ", "strength_and_weaknesses": "Strength:\n1. In this paper, several modifications (architecture, training setup and datasets) are explored to check whether there is any improvement. All of these aspects are important and interesting. These can give good insights for the community. \n\n2. Some interesting conclusion are got, for example, training recipe and data setup lead to decent downstream performance on GLUE.\n\nWeaknesses:\n\n\n1. The investigation of modifications lack convincing experiments. For example, only one task performance (MNLI) is reported for when studying the impact of training hyper-parameters. Other tasks can have a different trend.  And when exploring the effect of the architecture, only MLM loss is report. The performance of downstream tasks can be also important. \n\n2. The technical novelty of this paper is a little limit. The total contributions are also limit. ", "clarity,_quality,_novelty_and_reproducibility": "Clarity is clear, however, novelty is limit. ", "summary_of_the_review": "This paper does a lot of Interesting investigations. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666666826205}, {"id": "O8d9Km8oLKg", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5456/Reviewer_4joz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "While research on enormous language models have dominated empirical NLP lately, most researchers and practitioners do not have the resources and access to work with these models. This work seeks to answer the empirical question: what is the best performance one can achieve on the GLUE benchmark by training a BERT-like model from scratch on a consumer GPU for one day? The numerous Transformer variants proposed in recent years present another challenge to answering this question – which of these variants are beneficial when one’s compute is extremely constrained?\n\nThe authors investigate a wide range of design choices in terms of model architecture, training recipe, and data curation. They also note that the final MLM loss is mostly correlated with the FLOPs spent, not the particular Transformer type and size. This motivates them to choose architectures that parallelize well on GPUs. The final result demonstrates that some combination of the variants proposed in the three years after BERT yields a model that is almost as performant as the original BERT base while using 1/50 of the FLOPs.", "review_text": "Neat empirical investigation with conclusions that might interest many. However, it is not clear if ICLR is the best venue. Moreover, one might argue that MLM loss is not the best criterion to study model and training design choices.", "strengths": "This work presents a thorough empirical investigation on a topic that is of interest to many researchers who do not have access to large compute clusters. The overall methodology appears to be sound and the final result promising.\n\nHowever, I am not sure if ICLR is the best venue for this work. There are no new theoretical or algorithmic contributions nor any new insight into representation learning. This does not change my belief that this is an informative paper to many in the community, but it might find a more suitable audience if submitted to venues such as EMNLP. (Disclaimer: I do not regularly publish in this area. I am happy to defer to more experienced reviewers and the AC on this point.)\n\nAn additional concern is that it is not clear if the MLM loss is the best metric to track when comparing different model/training design choices. This paper (https://openreview.net/pdf?id=F5uYcwABMu), for example, clearly demonstrates that models with near-identical pretraining loss can perform very differently on downstream tasks due to implicit biases. This should be taken into consideration since many conclusions in this work are based on the MLM loss alone. I encourage the authors to also observe downstream performance and note if they agree or disagree with the MLM loss.\n\nThe connection between “cramming” and scaling laws can be clarified. The scaling laws mentioned in this work are empirical observations that a model’s performance strongly correlates its size but not necessarily its shape. The empirical results from this work show that this holds for the low-compute regime, which is somewhat surprising. However, these “laws” are merely empirical observations. It is not clear what the authors mean by “we discuss scaling laws in the low compute regime and find that cramming is hard because the conclusions of Kaplan et al. (2020) so easy to reproduce” (emphasis mine) in the conclusion. It would be better to simply state that the empirical observation from Kaplan et al. holds in the setups investigated, which motivated using architectures that parallelize well. A related concern is that since the dominating factor for performance is the number of FLOPs we can squeeze out of a GPU within a given timeframe, this makes the conclusions of this work somewhat hardware-specific, e.g., they might not hold on TPUs or newer/older GPUs.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "While research on enormous language models have dominated empirical NLP lately, most researchers and practitioners do not have the resources and access to work with these models. This work seeks to answer the empirical question: what is the best performance one can achieve on the GLUE benchmark by training a BERT-like model from scratch on a consumer GPU for one day? The numerous Transformer variants proposed in recent years present another challenge to answering this question – which of these variants are beneficial when one’s compute is extremely constrained?\n\nThe authors investigate a wide range of design choices in terms of model architecture, training recipe, and data curation. They also note that the final MLM loss is mostly correlated with the FLOPs spent, not the particular Transformer type and size. This motivates them to choose architectures that parallelize well on GPUs. The final result demonstrates that some combination of the variants proposed in the three years after BERT yields a model that is almost as performant as the original BERT base while using 1/50 of the FLOPs.", "strength_and_weaknesses": "This work presents a thorough empirical investigation on a topic that is of interest to many researchers who do not have access to large compute clusters. The overall methodology appears to be sound and the final result promising.\n\nHowever, I am not sure if ICLR is the best venue for this work. There are no new theoretical or algorithmic contributions nor any new insight into representation learning. This does not change my belief that this is an informative paper to many in the community, but it might find a more suitable audience if submitted to venues such as EMNLP. (Disclaimer: I do not regularly publish in this area. I am happy to defer to more experienced reviewers and the AC on this point.)\n\nAn additional concern is that it is not clear if the MLM loss is the best metric to track when comparing different model/training design choices. This paper (https://openreview.net/pdf?id=F5uYcwABMu), for example, clearly demonstrates that models with near-identical pretraining loss can perform very differently on downstream tasks due to implicit biases. This should be taken into consideration since many conclusions in this work are based on the MLM loss alone. I encourage the authors to also observe downstream performance and note if they agree or disagree with the MLM loss.\n\nThe connection between “cramming” and scaling laws can be clarified. The scaling laws mentioned in this work are empirical observations that a model’s performance strongly correlates its size but not necessarily its shape. The empirical results from this work show that this holds for the low-compute regime, which is somewhat surprising. However, these “laws” are merely empirical observations. It is not clear what the authors mean by “we discuss scaling laws in the low compute regime and find that cramming is hard because the conclusions of Kaplan et al. (2020) so easy to reproduce” (emphasis mine) in the conclusion. It would be better to simply state that the empirical observation from Kaplan et al. holds in the setups investigated, which motivated using architectures that parallelize well. A related concern is that since the dominating factor for performance is the number of FLOPs we can squeeze out of a GPU within a given timeframe, this makes the conclusions of this work somewhat hardware-specific, e.g., they might not hold on TPUs or newer/older GPUs.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is mostly well-written with just a few typos. I am not aware of similar investigations in the low-compute regime and believe that many in the community might find this work informative.", "summary_of_the_review": "Neat empirical investigation with conclusions that might interest many. However, it is not clear if ICLR is the best venue. Moreover, one might argue that MLM loss is not the best criterion to study model and training design choices.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666659415997}], "openreview_url": "https://openreview.net/forum?id=gUL6zYN4Uaf", "arxiv_id": "2212.14034", "paper_pdf": "papers/gUL6zYN4Uaf.pdf", "paper_pdf_sha256": "e5b18783401f125b8c5f4cd61469e662064876a5cb0070bfd1fa947bf6be05ff", "paper_pdf_bytes": 706584, "paper_pdf_source": "openreview", "code_url": "https://github.com/JonasGeiping/cramming", "code_repository": "JonasGeiping/cramming", "code_commit": "196c9912d8c5b06a05e9a58edd1521e3d38f7c0c", "code_archive": "repos/gUL6zYN4Uaf.zip", "code_archive_sha256": "71e797eb8a000d855a1cf2924462750fa444c6a298b9417747433783cc4f7b3a", "code_archive_bytes": 228108, "code_file_count": 56, "code_extensions": {".py": 41, ".sh": 15}, "github_disk_usage_kb": 237, "github_languages": {"Python": 375582, "Shell": 135326}, "github_archived": false, "github_pushed_at": "2024-06-13T07:56:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cramming-training-a-language-model-on-a"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yOBqNg-CqB0", "year": 2022, "status": "rejected", "title": "Re-evaluating Word Mover's Distance", "authors": ["Ryoma Sato", "Makoto Yamada", "Hisashi Kashima"], "authorids": ["~Ryoma_Sato1", "~Makoto_Yamada3", "~Hisashi_Kashima2"], "authors_source": "OpenReview API", "abstract": "The word mover's distance (WMD) is a fundamental technique for measuring the similarity of two documents. As the crux of WMD, it can take advantage of the underlying geometry of the word space by employing an optimal transport formulation. The original study on WMD reported that WMD outperforms classical baselines such as bag-of-words (BOW) and TF-IDF by significant margins in various datasets. In this paper, we point out that the evaluation in the original study could be misleading. We re-evaluate the performances of WMD and the classical baselines and find that the classical baselines are competitive with WMD if we employ an appropriate preprocessing, i.e., L1 normalization. In addition, We introduce an analogy between WMD and L1-normalized BOW and find that not only the performance of WMD but also the distance values resemble those of BOW in high dimensional spaces.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ofB19yjOuRp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2404/Reviewer_M3dN"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper re-evaluates WMD and identifies issues with the original paper. It shows that the gain from the original paper is not the product of WMD but the normalization. When the normalization is controlled, WMD performs similarly to baseline. Finally, it shows WMD resembles classic BOW when normalization is controlled.\n", "review_text": "# Strengths\n\nThis paper identifies issues with the original paper proposing WMD and during the evaluation, it shows the gain mainly due to normalization instead of WMD. It offers detailed analysis and experiments to support its claim.\n\n# Weaknesses\n\nn/a\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper re-evaluates WMD and identifies issues with the original paper. It shows that the gain from the original paper is not the product of WMD but the normalization. When the normalization is controlled, WMD performs similarly to baseline. Finally, it shows WMD resembles classic BOW when normalization is controlled.\n", "main_review": "# Strengths\n\nThis paper identifies issues with the original paper proposing WMD and during the evaluation, it shows the gain mainly due to normalization instead of WMD. It offers detailed analysis and experiments to support its claim.\n\n# Weaknesses\n\nn/a\n", "summary_of_the_review": "This paper revisits the original WMD paper and offers a detailed evaluation showing what contributes to the performance gain: normalization instead of WMD.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636010927470}, {"id": "o9TtH4rOojQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2404/Reviewer_5Jna"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper empirically shows that the performance of WMD is not as high as initially reported, and the real performance is comparable to  L1-normalized BOW, which can be formulated as a specific case of WMD. The authors also find that WMD resembles BOW in high-dimensional spaces. ", "review_text": "This paper targets a fairer evaluation of WMD. This is a critical problem worth exploring as WMD is a fundamental technique in various research fields. The core claims in this paper include: \n1. In the original study of WMD, many duplicate samples exist in the datasets, and applying L2 norm to word embeddings is not explicitly stated.\n2. The superiority of WMD over BOW and TF-IDF will weaken enormously by normalizing the BOW and TF-IDF.\n3. Both the normalization of word vectors and document-level distance metric (L1 or L2) will impact performance.\n4. WMD coincides with L1/L1 BOW empirically and theoretically, consistent with the two modalities characteristic of high-dimensional word embeddings.\n\nBy designing extensive experiments, the authors present the above observations and corresponding suggestions. They also provide clean datasets without duplicated samples and related code for further research. Following are some of my questions.\n\nIn figure 4, what metric is used for word-level distance?  Experiments show that L1 document-level distance with L1-normalized BOW/TF-IDF will generate the best performance. Regarding WMD, I wonder about the effects of different word-level distances (e.g., L1, L2, or Cosine). For example, WMD uses L2 word-level distance, then what about using other metrics? Besides, what if we use L1 normalization for word embeddings in WMD? \n\nIn my experience, the performance of WMD is extensively affected by the removal of appropriate stop-words. This paper only mentions the stop-word strategy on page 7.  I am wondering if the author removed the stop words in other experiments.\n\nTable 3-5 shows that removing OOV words brings significant performance degradation on bbcsports and ohsumed, which is somewhat against expectations. Is there any explanation for this observation?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper empirically shows that the performance of WMD is not as high as initially reported, and the real performance is comparable to  L1-normalized BOW, which can be formulated as a specific case of WMD. The authors also find that WMD resembles BOW in high-dimensional spaces. ", "main_review": "This paper targets a fairer evaluation of WMD. This is a critical problem worth exploring as WMD is a fundamental technique in various research fields. The core claims in this paper include: \n1. In the original study of WMD, many duplicate samples exist in the datasets, and applying L2 norm to word embeddings is not explicitly stated.\n2. The superiority of WMD over BOW and TF-IDF will weaken enormously by normalizing the BOW and TF-IDF.\n3. Both the normalization of word vectors and document-level distance metric (L1 or L2) will impact performance.\n4. WMD coincides with L1/L1 BOW empirically and theoretically, consistent with the two modalities characteristic of high-dimensional word embeddings.\n\nBy designing extensive experiments, the authors present the above observations and corresponding suggestions. They also provide clean datasets without duplicated samples and related code for further research. Following are some of my questions.\n\nIn figure 4, what metric is used for word-level distance?  Experiments show that L1 document-level distance with L1-normalized BOW/TF-IDF will generate the best performance. Regarding WMD, I wonder about the effects of different word-level distances (e.g., L1, L2, or Cosine). For example, WMD uses L2 word-level distance, then what about using other metrics? Besides, what if we use L1 normalization for word embeddings in WMD? \n\nIn my experience, the performance of WMD is extensively affected by the removal of appropriate stop-words. This paper only mentions the stop-word strategy on page 7.  I am wondering if the author removed the stop words in other experiments.\n\nTable 3-5 shows that removing OOV words brings significant performance degradation on bbcsports and ohsumed, which is somewhat against expectations. Is there any explanation for this observation?\n\n", "summary_of_the_review": "This paper re-evaluates the Word Mover's Distance by well-designed experiments. They reveal the true performance of WMD and draw its relationship with L1-normalized BOW. I am more inclined to accept this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635838396551}, {"id": "hAhH3iJOZ-f", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2404/Reviewer_id2C"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper presents a re-evaluation of a well known distance metric for documents with word embeddings. The authors identify some missing information from the original paper and they provide some extra analyses.", "review_text": "This paper presents a thorough re-evaluation of the word mover's distance paper by Kusner et al. (2015), focusing on a number of points that are considered misleading. The paper has a point about the fact that the normalization of the word vectors used in WMD by Kusner et al. was not mentioned in the paper, even though it was possible to find out in the code released by the authors (misleading point 2). I think that while welcome as a clarification, I think calling this omission misleading is a bit of a stretch, especially given that this was possible to find in the code released by the authors. The paper then criticises Kusner et al. for not using such a nornalization for the document vectors obtained by BOW and TFIDF, which they find improves their results (misleading point 3). However, it should be noted that the original paper included other methods for comparison, which performed rather competitive to WMD, so while I agree with this paper that Kusner et al. didn't do justice to BoW and TFIDF, they did show that other methods were competitive to WMD but not as good. This paper also admits this too. Furthermore, in WMD the normalization is at the word embedding level, while the L1/L2 normalization on BoW/TFIDF is at the document level, thus it is not a direct correspondence. Thus I don't think there is much to see here, unless the paper would like to be more of a case of how to make the most of token-matching distance metrics.\n\nThe first misleading point is about the datasets containing duplicates. But not sure this has to do that much with the method of Kusner et al. per se, especially given that the datasets are well-known and used frequently. Perhaps they shouldn't be, and a paper could be written about this. As for the misleading point 4, I have to disagree with this paper. The example from figure 1 in Kusner et al illustrates well how word embeddings could help us do something more useful than 0-1 distances in token matching metrics, which would be the case of their example. The arguments in  section 5 on the modes (not modalities) are not particularly tight. If anything, the BoW representation of a word or document is much more high dimensional than that of a (300 dim) word embedding, assuming that one has thousands of words typically. The point that Kusner et al. made was the word embedding distances are more informative than word matching, and this is confirmed in the plots of figure 4: different word pairs get different distances.\n\nOn the whole I believe that much of the criticism of this paper is not justified. The main claims of the Kusner et al. paper still hold, and the experiments in this paper confirm them in my mind. I would suggest that perhaps a better paper would be assessing the evaluation practices in distance metrics more broadly, using multiple state-of-the-art methods. While WMD influenced the way people are thinking, using it for this purpose with old embeddings on text classification which is nowadays dominated by fine-tuning pre-trained language models is not particularly informative.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a re-evaluation of a well known distance metric for documents with word embeddings. The authors identify some missing information from the original paper and they provide some extra analyses.", "main_review": "This paper presents a thorough re-evaluation of the word mover's distance paper by Kusner et al. (2015), focusing on a number of points that are considered misleading. The paper has a point about the fact that the normalization of the word vectors used in WMD by Kusner et al. was not mentioned in the paper, even though it was possible to find out in the code released by the authors (misleading point 2). I think that while welcome as a clarification, I think calling this omission misleading is a bit of a stretch, especially given that this was possible to find in the code released by the authors. The paper then criticises Kusner et al. for not using such a nornalization for the document vectors obtained by BOW and TFIDF, which they find improves their results (misleading point 3). However, it should be noted that the original paper included other methods for comparison, which performed rather competitive to WMD, so while I agree with this paper that Kusner et al. didn't do justice to BoW and TFIDF, they did show that other methods were competitive to WMD but not as good. This paper also admits this too. Furthermore, in WMD the normalization is at the word embedding level, while the L1/L2 normalization on BoW/TFIDF is at the document level, thus it is not a direct correspondence. Thus I don't think there is much to see here, unless the paper would like to be more of a case of how to make the most of token-matching distance metrics.\n\nThe first misleading point is about the datasets containing duplicates. But not sure this has to do that much with the method of Kusner et al. per se, especially given that the datasets are well-known and used frequently. Perhaps they shouldn't be, and a paper could be written about this. As for the misleading point 4, I have to disagree with this paper. The example from figure 1 in Kusner et al illustrates well how word embeddings could help us do something more useful than 0-1 distances in token matching metrics, which would be the case of their example. The arguments in  section 5 on the modes (not modalities) are not particularly tight. If anything, the BoW representation of a word or document is much more high dimensional than that of a (300 dim) word embedding, assuming that one has thousands of words typically. The point that Kusner et al. made was the word embedding distances are more informative than word matching, and this is confirmed in the plots of figure 4: different word pairs get different distances.\n\nOn the whole I believe that much of the criticism of this paper is not justified. The main claims of the Kusner et al. paper still hold, and the experiments in this paper confirm them in my mind. I would suggest that perhaps a better paper would be assessing the evaluation practices in distance metrics more broadly, using multiple state-of-the-art methods. While WMD influenced the way people are thinking, using it for this purpose with old embeddings on text classification which is nowadays dominated by fine-tuning pre-trained language models is not particularly informative.", "summary_of_the_review": "On the whole, while I appreciate the work to bring to the community's attention some extra information and analyses of the Word Mover's Distance paper, the paper is not very informative. I think a much more interesting paper would be to survey the evaluation of multiple distance metric learning approaches and discuss common pitfalls more broadly, rather than pick on one paper in the context of a task that it wouldn't be considered a standard method for (fine-tuning BERT style pre-trained models is the go-to method for text classification and is much better than word embeddings from 2015 with kNN).", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635554686911}, {"id": "UJIjlMmCzKX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2404/Reviewer_sFeb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper shows that when standard baseline methods are carefully evaluated, such as using L1 normalization of the tf-idf/bow vector, then the performance difference between the baselines and the WMD distance based methods shrinks considerably.\n\nThe authors also claim that the findings in the paper are more general in the sense, that the distance between the embeddings of two words behaves more and more like a delta function, which is bimodal : close to one when the words are distinct, and close to 0 when the words are the same.\n", "review_text": "### Strength\nThe paper presents experimental results that the WMD distance doesn't improve performance by a factor of 100% but rather by a factor of 10%. This corrects some misconceptions that may have arisen in a particular sub-field of NLP.\n\n### Weakness\nThe paper will have a low impact, and is too focused on refuting the results of a single paper. I believe that this paper may be well suited to conferences devoted to the particular application area of document classification or NLP in general, but as it is the paper's methodological contribution, or technical contribution is too little.\n\nThe paper presents a sub-section on the \"re-evaluation of existing methods\" in section 2 to answer some of this criticism, but all the papers in that section were much more wide-ranging in their focus-area. E.g. the (Dacrema et al.) paper was not focused on the results of a single paper but on multiple papers, and the paper by (Arora et al. 2017) presented a novel method. The paper by (Shen et al. 2018) was accepted at ACL which is an NLP conference more focused on NLP tasks.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper shows that when standard baseline methods are carefully evaluated, such as using L1 normalization of the tf-idf/bow vector, then the performance difference between the baselines and the WMD distance based methods shrinks considerably.\n\nThe authors also claim that the findings in the paper are more general in the sense, that the distance between the embeddings of two words behaves more and more like a delta function, which is bimodal : close to one when the words are distinct, and close to 0 when the words are the same.\n", "main_review": "### Strength\nThe paper presents experimental results that the WMD distance doesn't improve performance by a factor of 100% but rather by a factor of 10%. This corrects some misconceptions that may have arisen in a particular sub-field of NLP.\n\n### Weakness\nThe paper will have a low impact, and is too focused on refuting the results of a single paper. I believe that this paper may be well suited to conferences devoted to the particular application area of document classification or NLP in general, but as it is the paper's methodological contribution, or technical contribution is too little.\n\nThe paper presents a sub-section on the \"re-evaluation of existing methods\" in section 2 to answer some of this criticism, but all the papers in that section were much more wide-ranging in their focus-area. E.g. the (Dacrema et al.) paper was not focused on the results of a single paper but on multiple papers, and the paper by (Arora et al. 2017) presented a novel method. The paper by (Shen et al. 2018) was accepted at ACL which is an NLP conference more focused on NLP tasks.\n", "summary_of_the_review": "The paper corrects some errors made in an earlier paper by (Kusner et al. 2015) and presents a careful evaluation of baselines for document classification applications. However, I believe that the paper is more suited for an NLP conference than ICLR.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635218048130}], "openreview_url": "https://openreview.net/forum?id=yOBqNg-CqB0", "arxiv_id": "2105.14403", "paper_pdf": "papers/yOBqNg-CqB0.pdf", "paper_pdf_sha256": "ce33ca71a6c9eda4af42247d277db5846b0b4f3f97a99d32fda30d3e3f574ce1", "paper_pdf_bytes": 459310, "paper_pdf_source": "openreview", "code_url": "https://github.com/joisino/reeval-wmd", "code_repository": "joisino/reeval-wmd", "code_commit": "67d0c8e582c7f4b13befb9a84f128fb818c6b388", "code_archive": "repos/yOBqNg-CqB0.zip", "code_archive_sha256": "077c4acac0a1216c1d20ed81981194fd984c4dbd176cf4aeee071a41d0f7c583", "code_archive_bytes": 563248, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 557, "github_languages": {"Python": 36320, "Shell": 256}, "github_archived": false, "github_pushed_at": "2022-06-15T05:00:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/re-evaluating-word-mover-s-distance"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "V1N4GEWki_E", "year": 2021, "status": "rejected", "title": "Gradient Flow in Sparse Neural Networks and How Lottery Tickets Win", "authors": ["Utku Evci", "Yani Ioannou", "Cem Keskin", "Yann Dauphin"], "authorids": ["~Utku_Evci1", "~Yani_Ioannou1", "~Cem_Keskin2", "~Yann_Dauphin1"], "authors_source": "OpenReview API", "abstract": "Sparse Neural Networks (NNs) can match the generalization of dense NNs using a fraction of the compute/storage for inference, and also have the potential to enable efficient training. However, naively training unstructured sparse NNs from random initialization results in significantly worse generalization, with the notable exception of Lottery Tickets (LTs) and Dynamic Sparse Training (DST). In this work, we attempt to answer: (1) why training unstructured sparse networks from random initialization performs poorly and; and (2) what makes LTs and DST the exceptions? We show that sparse NNs have poor gradient flow at initialization and propose a modified initialization for unstructured connectivity. Furthermore, we find that DST methods significantly improve gradient flow during training over traditional sparse training methods. Finally, we show that LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from — however, this comes at the cost of learning novel solutions.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "mqlycatelq5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper499/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe paper is clear and very well written. It makes important steps towards understanding if sparse convolutional neural networks can represent a substitute for their dense counterparts. Moreover, it unveils the relation between the performance of sparse neural networks and gradient flow. Based on this relation, it explains also why the dynamic sparse training approach has higher potential of improving sparse neural networks in the future, while lottery tickets are limited by the performance of the pruning solutions from which they are derived. The last but not the least, the paper introduces a simple and practical method specially designed to initialise sparse networks weights.\n\nStrong points:\n\n•\tThe paper brings novel basic knowledge and understanding of sparse neural networks.\n\n•\tThe extensive set of experiments is well-designed, very informative, and support the paper claims.\n\n•\tThe fundamental study performed in this paper is timely and has the potential of advancing seriously the field.  \n\n\nWeak Points:\n\n•\tWhile the abstract and some other parts of the paper discuss about deep neural networks in general, the experiments are solely focused on convolutional neural networks. I believe that extending them also to other types of networks would improve the overall quality of the paper.\n\nFor the discussion phase, I suggest to the authors to consider the weak point and the following minor comments:\n\n1) I find very interesting that any sparse training method cope much better with the ResNet architecture than with the VGG architecture. It is easy to observe this in Table 1. Do you have any idea why is this happening? Is this the effect of skip connections?\n\n2) Can you add in figure 3 the “+” version of the training algorithms for ResNet-50 and the version without “+” for LeNet5?\n\n3) The relation between Hessian, gradient flow, and sparse training raised my curiosity, but I agree with the authors that this investigation can be let for future work.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Sparse networks understanding", "review": "Summary:\n\nThe paper is clear and very well written. It makes important steps towards understanding if sparse convolutional neural networks can represent a substitute for their dense counterparts. Moreover, it unveils the relation between the performance of sparse neural networks and gradient flow. Based on this relation, it explains also why the dynamic sparse training approach has higher potential of improving sparse neural networks in the future, while lottery tickets are limited by the performance of the pruning solutions from which they are derived. The last but not the least, the paper introduces a simple and practical method specially designed to initialise sparse networks weights.\n\nStrong points:\n\n•\tThe paper brings novel basic knowledge and understanding of sparse neural networks.\n\n•\tThe extensive set of experiments is well-designed, very informative, and support the paper claims.\n\n•\tThe fundamental study performed in this paper is timely and has the potential of advancing seriously the field.  \n\n\nWeak Points:\n\n•\tWhile the abstract and some other parts of the paper discuss about deep neural networks in general, the experiments are solely focused on convolutional neural networks. I believe that extending them also to other types of networks would improve the overall quality of the paper.\n\nFor the discussion phase, I suggest to the authors to consider the weak point and the following minor comments:\n\n1) I find very interesting that any sparse training method cope much better with the ResNet architecture than with the VGG architecture. It is easy to observe this in Table 1. Do you have any idea why is this happening? Is this the effect of skip connections?\n\n2) Can you add in figure 3 the “+” version of the training algorithms for ResNet-50 and the version without “+” for LeNet5?\n\n3) The relation between Hessian, gradient flow, and sparse training raised my curiosity, but I agree with the authors that this investigation can be let for future work.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603993694853}, {"id": "-pZuChbkVqt", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper499/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents three key hypotheses for sparse NN training dynamics and provides empirical studies and observations to verify them. This review will first provide general comments and then specific ones on each hypothesis. \n\nMain comments:\n\nPros: Three messages that the authors try to convey are important and interesting to the audiences in pruning. It is an observational paper which provides insights on 1) what is a good initialization of the for sparse NN training 2) why DST can achieve good generalization, and 3) is LTH really different from pruning. \n\nCons: The presentation of the paper needs work. The high-level structure is good and clear but for each paragraph, the logic flow is hard to follow. For example, in a very key paragraph on P7: “Lottery Tickets Learn Similar Functions to the Pruning Solution”, I have to read repeatedly and infer inner logics of each sentence to see the conclusion. \n\nHypotheses:\nI appreciate identifying the problem of naive initialization of sparse NN and connecting it with gradient flow. However, a new proposal for initialization here (as the major contribution) is unnecessary and actually negatively affects the credits of the true contribution. The results presented in table 1 are not very impressive. It is ok to just compare original and liu et al, and provide insights in an observational paper.\nThe observations provide one possible explanation on why DST might work. The authors could try to test this in different architectures (even beyond CNNs) to see if they are widely held. If so, it is potentially a good metric or analysis tool for sparse NN training.\nI am not fully convinced by the third hypothesis. First, the models and datasets for ensemble and prediction disagreement are too limited while the conclusion is very strong. Also I think a more appropriate statement could be Lottery Tickets Learn Similar Functions to the Pruning Solution than random /scratch since that is the only thing you are comparing with.\n\nMinor comment:\n\nIt would be interesting to see if all the conclusions hold in other models besides CNNs.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #4", "review": "This paper presents three key hypotheses for sparse NN training dynamics and provides empirical studies and observations to verify them. This review will first provide general comments and then specific ones on each hypothesis. \n\nMain comments:\n\nPros: Three messages that the authors try to convey are important and interesting to the audiences in pruning. It is an observational paper which provides insights on 1) what is a good initialization of the for sparse NN training 2) why DST can achieve good generalization, and 3) is LTH really different from pruning. \n\nCons: The presentation of the paper needs work. The high-level structure is good and clear but for each paragraph, the logic flow is hard to follow. For example, in a very key paragraph on P7: “Lottery Tickets Learn Similar Functions to the Pruning Solution”, I have to read repeatedly and infer inner logics of each sentence to see the conclusion. \n\nHypotheses:\nI appreciate identifying the problem of naive initialization of sparse NN and connecting it with gradient flow. However, a new proposal for initialization here (as the major contribution) is unnecessary and actually negatively affects the credits of the true contribution. The results presented in table 1 are not very impressive. It is ok to just compare original and liu et al, and provide insights in an observational paper.\nThe observations provide one possible explanation on why DST might work. The authors could try to test this in different architectures (even beyond CNNs) to see if they are widely held. If so, it is potentially a good metric or analysis tool for sparse NN training.\nI am not fully convinced by the third hypothesis. First, the models and datasets for ensemble and prediction disagreement are too limited while the conclusion is very strong. Also I think a more appropriate statement could be Lottery Tickets Learn Similar Functions to the Pruning Solution than random /scratch since that is the only thing you are comparing with.\n\nMinor comment:\n\nIt would be interesting to see if all the conclusions hold in other models besides CNNs.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603942601372}, {"id": "ehP_ZwJZuMU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper499/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nPaper Summary:\n\nThis paper presents an empirical study of sparse deep nets, either obtained by sparsification methods such as “dynamic sparse training” or by pruning according to the lottery ticket hypothesis.\nThe main contribution of this work is to study gradient flow both at initialisation and during training, and to propose an extension of known initialisation methods that works for sparse networks.\nIn addition, this work also attempts at explaining why lottery tickets are successful, despite sharing similar problems related to the gradient flow, when compared to other sparsification methods.\n\n##########################################################################\n\nReasons for score: \n\nOverall, I like this kind of empirical study, where authors set the stage for important questions and attempt at answering them with a thorough empirical study. My major concern for accepting this work relates to the depth of contributions. In my humble opinion, the proposed generalisation of “He’s initialisation” could have have been the main (only?) focus of this work, with additional experiments and considerations: relation to other sparsity inducing methods (not only DST), a better understanding of the interaction between initialisation, Batch normalisation and skip connections, …\nInstead, the presentation strategy in this paper is to illustrate several findings but, due to space constraints, in a more shallow manner. This choice dilutes the contributions too much. \n\n##########################################################################\n\nPositive points: \n\n1) Empirical work that addresses an important topic, that of sparse NN. Questions are well motivated, and sufficiently well described, although they have the slightly negative effect of diluting the overall take home message from reading this paper.\n\n2) The proposed extension to a known initialisation method to cope with issues related to gradient flow during the early stages of training is reasonable, and effective as shown by the experiments.\n\n3) The experiments on gradient flow during training and the ones on lottery tickets confirm either known results or intuition. They can be viewed as a reproducibility study, which is commendable. Some by products of the study indicate important properties of LT, which are of direct practical relevance, e.g. rewinding strategies.\n\n\n##########################################################################\n\nNegative points:\n\n1) The proposed generalisation of He’s method is not sufficiently exploited. Focusing on the forward pass, the idea is to initialise weights on a per neuron basis, using the mask computed by the sparsification method. As the authors notice, the work by Lui et al. achieves similar performance (Fig 1.c) and in many cases it outperforms the proposed method (Tab.1, bold results). This calls for a better understanding of the advantages of the proposed method. Furthermore, the interaction of “sparse initialisation” with batch normalisation and skip connections is not sufficiently studied: in Fig.2 all methods appear similar. Finally, the fact that a “small but dense” network achieves better results (LeNet on MNIST, Tab1), and does not suffer from gradient flow problems (Fig.2 c) is interesting and calls again for further study.\n\n2) The results on gradient flow during training are only superficially commented, although the authors hint at additional ideas based on second order approximations of the loss, i.e. considering the Hessian and its eigen-spectrum. Overall, the take home message from Fig.3 confirms the known behaviour of DST methods such as RigL. Albeit interesting, in my humble opinion this results seem to be given more “real estate” on the paper than it deserves, subtracting space for the main contribution on a new initialisation scheme.\n\n3) The results on lottery ticket could enjoy some improvements on the terminology, which is a minor remark. Indeed, it would be easier to refer to: 1) IMP solution <-> pruning solution; 2) Random init/final <-> start/end; 3) LT init/final <-> start/end. My main concern with this set of results is that on the one hand, they are somehow expected, especially with respect to the large literature available on the topic. It is not bad per se to collect in one coherent piece of work previous observations and place them in a thorough experimental framework. However, I have problems with the following.\na) The notion of “closeness” as shown in Fig 5 a,d and reported in fig 5 b,e is the result of a dramatic dimensionality reduction. Similar techniques have been used in other contexts (e.g. Hao Li, et al. “Visualizing the Loss Landscape of Neural Nets.” NIPS, 2018) and the warnings are to take results with a grain of salt. That said, it is expected — by construction — to find that LT final networks are close to the IMP solution.\nb) The argument used to confirm that LT are in the basin of the IMP solution is based on path connectivity and, as the author also note in a foot note in page 7, studying in detail this path is outside the scope of the paper. The geometry of the loss landscapes is in general very complex (especially when there is no Batch normalisation nor skip connections, which have the effect of smoothing it): I am not sure it is correct to claim that if two solutions (that is the params of a neural net) have the same loss and they are connected by a linear path, then it is necessary true that they lie in the same basin. Even if results in Tab.2 on the disagreement are compelling, it might still not be necessarily true that the two compared models are the same instance of function approximation.\nc) As a minor remark, the implications of the results in Sec 4.3 are interesting and valuable, but I failed to understand properly the connection to the empirical study on the “distance” between IMP solutions and LT final solutions.\n \n\n#########################################################################\n\nAdditional comments:\n\nI found this paper well written in most parts (just a minor comment on the terminology used in one sub-section). I liked this work and I think it has plenty of potential. \nAs an humble suggestion, would it make sense to attempt at focussing more the message, and insist on the main contribution of the paper as per a new initialisation scheme, or was this not seen as sufficient in light of the results from Liu et al. (2019).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work with some potential, but with some flaws too", "review": "##########################################################################\n\nPaper Summary:\n\nThis paper presents an empirical study of sparse deep nets, either obtained by sparsification methods such as “dynamic sparse training” or by pruning according to the lottery ticket hypothesis.\nThe main contribution of this work is to study gradient flow both at initialisation and during training, and to propose an extension of known initialisation methods that works for sparse networks.\nIn addition, this work also attempts at explaining why lottery tickets are successful, despite sharing similar problems related to the gradient flow, when compared to other sparsification methods.\n\n##########################################################################\n\nReasons for score: \n\nOverall, I like this kind of empirical study, where authors set the stage for important questions and attempt at answering them with a thorough empirical study. My major concern for accepting this work relates to the depth of contributions. In my humble opinion, the proposed generalisation of “He’s initialisation” could have have been the main (only?) focus of this work, with additional experiments and considerations: relation to other sparsity inducing methods (not only DST), a better understanding of the interaction between initialisation, Batch normalisation and skip connections, …\nInstead, the presentation strategy in this paper is to illustrate several findings but, due to space constraints, in a more shallow manner. This choice dilutes the contributions too much. \n\n##########################################################################\n\nPositive points: \n\n1) Empirical work that addresses an important topic, that of sparse NN. Questions are well motivated, and sufficiently well described, although they have the slightly negative effect of diluting the overall take home message from reading this paper.\n\n2) The proposed extension to a known initialisation method to cope with issues related to gradient flow during the early stages of training is reasonable, and effective as shown by the experiments.\n\n3) The experiments on gradient flow during training and the ones on lottery tickets confirm either known results or intuition. They can be viewed as a reproducibility study, which is commendable. Some by products of the study indicate important properties of LT, which are of direct practical relevance, e.g. rewinding strategies.\n\n\n##########################################################################\n\nNegative points:\n\n1) The proposed generalisation of He’s method is not sufficiently exploited. Focusing on the forward pass, the idea is to initialise weights on a per neuron basis, using the mask computed by the sparsification method. As the authors notice, the work by Lui et al. achieves similar performance (Fig 1.c) and in many cases it outperforms the proposed method (Tab.1, bold results). This calls for a better understanding of the advantages of the proposed method. Furthermore, the interaction of “sparse initialisation” with batch normalisation and skip connections is not sufficiently studied: in Fig.2 all methods appear similar. Finally, the fact that a “small but dense” network achieves better results (LeNet on MNIST, Tab1), and does not suffer from gradient flow problems (Fig.2 c) is interesting and calls again for further study.\n\n2) The results on gradient flow during training are only superficially commented, although the authors hint at additional ideas based on second order approximations of the loss, i.e. considering the Hessian and its eigen-spectrum. Overall, the take home message from Fig.3 confirms the known behaviour of DST methods such as RigL. Albeit interesting, in my humble opinion this results seem to be given more “real estate” on the paper than it deserves, subtracting space for the main contribution on a new initialisation scheme.\n\n3) The results on lottery ticket could enjoy some improvements on the terminology, which is a minor remark. Indeed, it would be easier to refer to: 1) IMP solution <-> pruning solution; 2) Random init/final <-> start/end; 3) LT init/final <-> start/end. My main concern with this set of results is that on the one hand, they are somehow expected, especially with respect to the large literature available on the topic. It is not bad per se to collect in one coherent piece of work previous observations and place them in a thorough experimental framework. However, I have problems with the following.\na) The notion of “closeness” as shown in Fig 5 a,d and reported in fig 5 b,e is the result of a dramatic dimensionality reduction. Similar techniques have been used in other contexts (e.g. Hao Li, et al. “Visualizing the Loss Landscape of Neural Nets.” NIPS, 2018) and the warnings are to take results with a grain of salt. That said, it is expected — by construction — to find that LT final networks are close to the IMP solution.\nb) The argument used to confirm that LT are in the basin of the IMP solution is based on path connectivity and, as the author also note in a foot note in page 7, studying in detail this path is outside the scope of the paper. The geometry of the loss landscapes is in general very complex (especially when there is no Batch normalisation nor skip connections, which have the effect of smoothing it): I am not sure it is correct to claim that if two solutions (that is the params of a neural net) have the same loss and they are connected by a linear path, then it is necessary true that they lie in the same basin. Even if results in Tab.2 on the disagreement are compelling, it might still not be necessarily true that the two compared models are the same instance of function approximation.\nc) As a minor remark, the implications of the results in Sec 4.3 are interesting and valuable, but I failed to understand properly the connection to the empirical study on the “distance” between IMP solutions and LT final solutions.\n \n\n#########################################################################\n\nAdditional comments:\n\nI found this paper well written in most parts (just a minor comment on the terminology used in one sub-section). I liked this work and I think it has plenty of potential. \nAs an humble suggestion, would it make sense to attempt at focussing more the message, and insist on the main contribution of the paper as per a new initialisation scheme, or was this not seen as sufficient in light of the results from Liu et al. (2019).", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603200007252}, {"id": "Iynm0T7BFVZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper499/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overview:\n\nSummary:\nThis paper tries to answer the following two questions: i) why training unstructured sparse networks from random initiation perform poorly? 2) what makes LTs and DST the exception? The authors show the following findings:\n1. Sparse NNs have poor gradient flow at initialization. They show that existing methods for initializing sparse NNs are incorrect in not considering heterogeneous connectivity. Improved methods are sample initialization from a dynamic gaussian whose variance is related to the fan-in numbers. fan-in = fan-out rule plays an important role here and improves the gradient flow.\n2. Sparse NNs have poor gradient flow during training. They show that DST based methods achieving the best generalization have improved gradient flow.\n3. They find the LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from.\n\n\nStrength bullets:\n1. The idea is very interesting. I appreciate the novel analysis. The proposed methods are well-motivated.\n2. The paper is well written and easy to understand.\n3. The finding is surprising but the experiment design is poor which I will list more detailed limitations in the weakness sections. I like the idea, I will raise my score if the authors can completely address my confusion and concerns.\n\n\nWeakness bullets:\n1. For Table 1, a Strong baseline is missing. Why not compare with the performance of the lottery ticket setting? I think it is a more natural baseline than SET and RigL.\n2. In my opinion, there is a must-do experiment: Lottery ticket mask + proposed initialization and compare it to LT and random tickets. Because the LT mask + random reinitialization = random tickets fail in the previous literature. According to the explanation in the paper, it can also be the problem of random reinitialization. Thus, strong supportive evidence is that show proposed modified random reinitialization + LT mask can surpass random ticket performance.\n3. Missing details what is the pruning ratio of each stage in iterative magnitude pruning? The appendix only tells me the author using 95% and 80% sparsity, why pick these two sparsity? Because this sparsity gives the extreme matching subnetworks? And the author uses iterative magnitude pruning, if they follow the original LTH setting, pruning 20% for each time. Then the sparsity should be 1-0.8^i, how to achieve 95% and 80%?\n4. What is the definition of \"pruning solution\"? Is it the obtained mask or initialization or subnetworks contains both mask and initialization? Super confused\n5. Conflicted experiments results with Linear Mode Connectivity and the Lottery Ticket Hypothesis paper, ResNet 50 IMP LT on ImageNet without Early weight rewinding can not have good linear mode connectivity. However, the pruning solution and LT solution have good linear mode connectivity. It is wired, even for two LTs (ResNet 50 IMP LT on ImageNet) trained with the same initialization in different data orders, they do not have a linear path where interpolated training loss is flat, as evidenced in figure 5 in the paper \"Linear Mode Connectivity and the Lottery Ticket Hypothesis\". Early weight rewinding is needed for the presented results while I think the author did not use it. \n6. The comparison in Table 2 is unfair. Scratch settings are trained from five different random initialization, while LT settings are trained from the same initialization with different data orders. LT setting results should also be from different initialization, otherwise can not achieve the conclusion that \"Lottery Tickets Learn Similar Functions to the Pruning Solution\".\n\nMinor:\n1. The definition of LTH in 3.3 \"perform as well as O^N(f,\\theta)*M\", why there is M? It should be the full dense model without the mask, right?\n\n------ Post Rebuttal------\n\nThanks to the authors for the extra experiments and feedback!\n\n[Lottery baseline for Table-1] Although RigL does not need dense network training, it cost more to find the mask (Table 2 of the RigL paper).\n\n[Random tickets] Random Ticket = LT mask + random re-initialization rather than random pruning + random init. The front one will be much more interesting. \"Because the LT mask + random reinitialization = random tickets fail in the previous literature. According to the explanation in the paper, it can also be the problem of random reinitialization. Thus, strong supportive evidence is that show proposed modified random reinitialization + LT mask can surpass random ticket performance.\" I personally do the experiment that performing proposed initialization on random tickets and the performance is unchanged. Of course, there may exist lots of reasons for the results. I will not degrade the paper according to my experiments.\n\nOther concerns are will-addressed. Thanks!\n\nAlthough I do like the idea of this paper, I think it might need to be revised and resubmitted, incorporating the extensive discussion presented by all the reviewers. I tend to keep my scores unchanged. But I don’t think this is 100% a clear reject and depending on the opinions of the other reviewers I would not feel that accepting this paper was completely out of bounds.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The idea is novel and interesting, while the experiments design is poor. I give borderline reject. I expect the response from the authors. If all concerns are addressed, I will raise my scores.", "review": "Overview:\n\nSummary:\nThis paper tries to answer the following two questions: i) why training unstructured sparse networks from random initiation perform poorly? 2) what makes LTs and DST the exception? The authors show the following findings:\n1. Sparse NNs have poor gradient flow at initialization. They show that existing methods for initializing sparse NNs are incorrect in not considering heterogeneous connectivity. Improved methods are sample initialization from a dynamic gaussian whose variance is related to the fan-in numbers. fan-in = fan-out rule plays an important role here and improves the gradient flow.\n2. Sparse NNs have poor gradient flow during training. They show that DST based methods achieving the best generalization have improved gradient flow.\n3. They find the LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from.\n\n\nStrength bullets:\n1. The idea is very interesting. I appreciate the novel analysis. The proposed methods are well-motivated.\n2. The paper is well written and easy to understand.\n3. The finding is surprising but the experiment design is poor which I will list more detailed limitations in the weakness sections. I like the idea, I will raise my score if the authors can completely address my confusion and concerns.\n\n\nWeakness bullets:\n1. For Table 1, a Strong baseline is missing. Why not compare with the performance of the lottery ticket setting? I think it is a more natural baseline than SET and RigL.\n2. In my opinion, there is a must-do experiment: Lottery ticket mask + proposed initialization and compare it to LT and random tickets. Because the LT mask + random reinitialization = random tickets fail in the previous literature. According to the explanation in the paper, it can also be the problem of random reinitialization. Thus, strong supportive evidence is that show proposed modified random reinitialization + LT mask can surpass random ticket performance.\n3. Missing details what is the pruning ratio of each stage in iterative magnitude pruning? The appendix only tells me the author using 95% and 80% sparsity, why pick these two sparsity? Because this sparsity gives the extreme matching subnetworks? And the author uses iterative magnitude pruning, if they follow the original LTH setting, pruning 20% for each time. Then the sparsity should be 1-0.8^i, how to achieve 95% and 80%?\n4. What is the definition of \"pruning solution\"? Is it the obtained mask or initialization or subnetworks contains both mask and initialization? Super confused\n5. Conflicted experiments results with Linear Mode Connectivity and the Lottery Ticket Hypothesis paper, ResNet 50 IMP LT on ImageNet without Early weight rewinding can not have good linear mode connectivity. However, the pruning solution and LT solution have good linear mode connectivity. It is wired, even for two LTs (ResNet 50 IMP LT on ImageNet) trained with the same initialization in different data orders, they do not have a linear path where interpolated training loss is flat, as evidenced in figure 5 in the paper \"Linear Mode Connectivity and the Lottery Ticket Hypothesis\". Early weight rewinding is needed for the presented results while I think the author did not use it. \n6. The comparison in Table 2 is unfair. Scratch settings are trained from five different random initialization, while LT settings are trained from the same initialization with different data orders. LT setting results should also be from different initialization, otherwise can not achieve the conclusion that \"Lottery Tickets Learn Similar Functions to the Pruning Solution\".\n\nMinor:\n1. The definition of LTH in 3.3 \"perform as well as O^N(f,\\theta)*M\", why there is M? It should be the full dense model without the mask, right?\n\n------ Post Rebuttal------\n\nThanks to the authors for the extra experiments and feedback!\n\n[Lottery baseline for Table-1] Although RigL does not need dense network training, it cost more to find the mask (Table 2 of the RigL paper).\n\n[Random tickets] Random Ticket = LT mask + random re-initialization rather than random pruning + random init. The front one will be much more interesting. \"Because the LT mask + random reinitialization = random tickets fail in the previous literature. According to the explanation in the paper, it can also be the problem of random reinitialization. Thus, strong supportive evidence is that show proposed modified random reinitialization + LT mask can surpass random ticket performance.\" I personally do the experiment that performing proposed initialization on random tickets and the performance is unchanged. Of course, there may exist lots of reasons for the results. I will not degrade the paper according to my experiments.\n\nOther concerns are will-addressed. Thanks!\n\nAlthough I do like the idea of this paper, I think it might need to be revised and resubmitted, incorporating the extensive discussion presented by all the reviewers. I tend to keep my scores unchanged. But I don’t think this is 100% a clear reject and depending on the opinions of the other reviewers I would not feel that accepting this paper was completely out of bounds.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602658747149}], "openreview_url": "https://openreview.net/forum?id=V1N4GEWki_E", "arxiv_id": "2010.03533", "paper_pdf": "papers/V1N4GEWki_E.pdf", "paper_pdf_sha256": "7d6e5c1e5e2ea9926f5c7151ed93f3b1d05c06027db2322b623c7885314e6283", "paper_pdf_bytes": 1072138, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/rigl", "code_repository": "google-research/rigl", "code_commit": "d39fc7d46505cb3196cb1edeb32ed0b6dd44c0f9", "code_archive": "repos/V1N4GEWki_E.zip", "code_archive_sha256": "0f60b5406e049cb72423bfad48b6549c8bceb61dab16629051be2cef8c6f9d65", "code_archive_bytes": 808678, "code_file_count": 95, "code_extensions": {".py": 88, ".ipynb": 4, ".sh": 3}, "github_disk_usage_kb": 830, "github_languages": {"Python": 791400, "Jupyter Notebook": 71595, "Shell": 2845}, "github_archived": false, "github_pushed_at": "2023-01-26T17:47:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gradient-flow-in-sparse-neural-networks-and-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJeS16EKPr", "year": 2020, "status": "rejected", "title": "Learning relevant features for statistical inference", "authors": ["Cédric Bény"], "authorids": ["cedric.beny@gmail.com"], "authors_source": "OpenReview API", "abstract": "We introduce an new technique to learn correlations between two types of data.\nThe learned representation can be used to directly compute the expectations of functions over one type of data conditioned on the other, such as Bayesian estimators and their standard deviations. \nSpecifically, our loss function teaches two neural nets to extract features representing the probability vectors of highest singular value for the stochastic map (set of conditional probabilities) implied by the joint dataset, relative to the inner product defined by the Fisher information metrics evaluated at the marginals.\nWe test the approach using a synthetic dataset, analytical calculations, and inference on occluded MNIST images. \nSurprisingly, when applied to supervised learning (one dataset consists of labels), this approach automatically provides regularization and faster convergence compared to the cross-entropy objective.\nWe also explore using this approach to discover salient independent features of a single dataset. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HyggZ2a-5S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper299/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an approach to find a map between two feature spaces to maximize correlation between them and to use the resulting map for inference. A theoretical exposition is given and some empirical results are provided showing that the approach speeds up convergence on supervised MNIST and can be used for image completion (again on MNIST).\n\nThe paper should be rejected for the following reasons. First, the approach looks very similar to deep CCA, but the connection is never mentioned. This connection needs to be clarified. The objective function needs to be clearly stated and related to the loss function in eq. (7). In particular, I would suggest to give a clear definition of the problem before delving into the theory in section 2. In its current version, it is difficult to assess how the parts of section 2 relate to the overall objective. The paper severely lacks in relation to relevant related work. Half(!) of the 14 referenced papers are by the author himself. This can be verified since the double blind review process is compromised as the paper links to code in the author’s public github account. Finally, the empirical results are quite incomplete. It is not clear how the results compare to generative methods like VAEs, which are referenced as a motivation for this work in this work.\n\nThe improvement in convergence from RFA for supervised learning is interesting and this aspect deserves more analysis. It would be useful to look at the total amount of computation required to reach a given loss. I also wonder how this differs from simply mapping the output of the first network to a low-rank space via PCA. Is the dual-view really necessary in this case since the information content in the label space must be very limited, beyond simple class balance statistics?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes an approach to find a map between two feature spaces to maximize correlation between them and to use the resulting map for inference. A theoretical exposition is given and some empirical results are provided showing that the approach speeds up convergence on supervised MNIST and can be used for image completion (again on MNIST).\n\nThe paper should be rejected for the following reasons. First, the approach looks very similar to deep CCA, but the connection is never mentioned. This connection needs to be clarified. The objective function needs to be clearly stated and related to the loss function in eq. (7). In particular, I would suggest to give a clear definition of the problem before delving into the theory in section 2. In its current version, it is difficult to assess how the parts of section 2 relate to the overall objective. The paper severely lacks in relation to relevant related work. Half(!) of the 14 referenced papers are by the author himself. This can be verified since the double blind review process is compromised as the paper links to code in the author’s public github account. Finally, the empirical results are quite incomplete. It is not clear how the results compare to generative methods like VAEs, which are referenced as a motivation for this work in this work.\n\nThe improvement in convergence from RFA for supervised learning is interesting and this aspect deserves more analysis. It would be useful to look at the total amount of computation required to reach a given loss. I also wonder how this differs from simply mapping the output of the first network to a low-rank space via PCA. Is the dual-view really necessary in this case since the information content in the label space must be very limited, beyond simple class balance statistics?\n"}, "tcdate": 1572097015999}, {"id": "rklS9UB6YS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper299/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers how to learn correlations between two spaces, e.g., input/output, in order to generate data in one space conditioned on values from the other. This is performed by modeling features with neural networks and optimizing an objective function that maximizes a measure of correlation between the features versus learning a generative model such as a CVAE. Some illustrative examples using MNIST are provided.\n\nMy decision is to reject. I think there is value in the approach, but it is hard to see clearly at the moment given that the exposition is difficult to follow and the experiments aren't very compelling. If these issues could be addressed (concrete suggestions below), and some of the follow-on work in the last section could be performed, I think there could be a pretty interesting contribution here.\n\n***\n\nDecision-related suggestions/questions:\n\n* Include more datasets in the experimental section. The second sentence of the introduction lists possibilities such as time series and multi-modalities that I would have been very interesting.\n\n* The first claim that there is no tunable variable in the objective function is a little hard to parse. Clearly, the rank of the low-rank approximation must be set, and the features of the two spaces need to be learned. Some clarification here would be helpful. \n\n* There are a number of unfortunate typos/grammar issues/presentation choices that really impact clarity. Some examples:\n\t* In the third paragraph under theory, \"...linear spaces spanned by the probability distributions...\" should probably be \"...linear spaces spanned by all probability distributions...\" (?)\n\t* The following sentence is a run-on. \n\t* In (8), occurrences of g_*(x_n) should be replaced with g_*(y_n).\n\t* The replacement of the (low) rank symbol, k_0, with the sample size symbol, n, in the second paragraph of section 4.1.\n\t* Introducing a \"Bayesian estimator for an l^2 distance\" w/o explanation. What does this mean?\n\n* How should the low-rank parameter k_0 be selected generally given that the singular value distribution may not always be useful in selecting it?\n\n* Can anything be said quantitatively or qualitatively about the sample complexity required to estimate the matrices of (8) well enough to estimate the features?\n\n* Is there an interpretation for why both spaces require the same feature dimension, k_0?\n\n***\n\nComments not related to decision:\n\n* It is generally good to avoid sweeping statements such as the first sentence of the introduction. Perhaps consider replacing with a simple statement on the intended goal of the paper: \"...produce a useful model of correlations... for the task of data generation...\"\n\n* Consider placing a concrete, motivating example prior to the theory section as it is hard to digest (from an ML perspective) without a clear context. The analytical example with the Gaussian from the supplementary material is one option.\n\n* The last statement of the paragraph under (3) needs a reference.\n\n* It seems strange to have the supervised learning experiment of 4.1 as the first experimental result of the paper since it is an unintended and unexplained side-effect of the approach. Also, the claim of \"faster convergence\" should be demonstrated in wall-clock time.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper considers how to learn correlations between two spaces, e.g., input/output, in order to generate data in one space conditioned on values from the other. This is performed by modeling features with neural networks and optimizing an objective function that maximizes a measure of correlation between the features versus learning a generative model such as a CVAE. Some illustrative examples using MNIST are provided.\n\nMy decision is to reject. I think there is value in the approach, but it is hard to see clearly at the moment given that the exposition is difficult to follow and the experiments aren't very compelling. If these issues could be addressed (concrete suggestions below), and some of the follow-on work in the last section could be performed, I think there could be a pretty interesting contribution here.\n\n***\n\nDecision-related suggestions/questions:\n\n* Include more datasets in the experimental section. The second sentence of the introduction lists possibilities such as time series and multi-modalities that I would have been very interesting.\n\n* The first claim that there is no tunable variable in the objective function is a little hard to parse. Clearly, the rank of the low-rank approximation must be set, and the features of the two spaces need to be learned. Some clarification here would be helpful. \n\n* There are a number of unfortunate typos/grammar issues/presentation choices that really impact clarity. Some examples:\n\t* In the third paragraph under theory, \"...linear spaces spanned by the probability distributions...\" should probably be \"...linear spaces spanned by all probability distributions...\" (?)\n\t* The following sentence is a run-on. \n\t* In (8), occurrences of g_*(x_n) should be replaced with g_*(y_n).\n\t* The replacement of the (low) rank symbol, k_0, with the sample size symbol, n, in the second paragraph of section 4.1.\n\t* Introducing a \"Bayesian estimator for an l^2 distance\" w/o explanation. What does this mean?\n\n* How should the low-rank parameter k_0 be selected generally given that the singular value distribution may not always be useful in selecting it?\n\n* Can anything be said quantitatively or qualitatively about the sample complexity required to estimate the matrices of (8) well enough to estimate the features?\n\n* Is there an interpretation for why both spaces require the same feature dimension, k_0?\n\n***\n\nComments not related to decision:\n\n* It is generally good to avoid sweeping statements such as the first sentence of the introduction. Perhaps consider replacing with a simple statement on the intended goal of the paper: \"...produce a useful model of correlations... for the task of data generation...\"\n\n* Consider placing a concrete, motivating example prior to the theory section as it is hard to digest (from an ML perspective) without a clear context. The analytical example with the Gaussian from the supplementary material is one option.\n\n* The last statement of the paragraph under (3) needs a reference.\n\n* It seems strange to have the supervised learning experiment of 4.1 as the first experimental result of the paper since it is an unintended and unexplained side-effect of the approach. Also, the claim of \"faster convergence\" should be demonstrated in wall-clock time."}, "tcdate": 1571800717116}, {"id": "BklaGd7KKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper299/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes how the correlation between two different types of data can be extracted from learned representations. The proposed metric can also be used as an alternative to cross entropy loss. The paper provides analytical calculations as well as real data sets simulations/experiments. However there are significant draw-becks:\n\n1) Similarity and Metric Learning is a booming area in machine learning with several different directions focusing on different problems. The paper fails to locates itself in the literature, how it compares itself into other techniques (both analytically and experimentally). \n\n2) The proposed technique seems to be very similar to SVD of learned representations. Connection to quantum field theory is well established but more simpler comparisons to SVD and other spectral techniques are not provided in metric learning.\n\n3) Novelty is not clear. There are interesting experiments in disentangled feature feature extraction and data generation. However, they are mostly proof of concept and lack of baselines. It is not clear what problem this technique solves better compared to other existing solutions. \n\nPaper is mostly written clearly. I do suggest putting Appendix A1 to the main paper though.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes how the correlation between two different types of data can be extracted from learned representations. The proposed metric can also be used as an alternative to cross entropy loss. The paper provides analytical calculations as well as real data sets simulations/experiments. However there are significant draw-becks:\n\n1) Similarity and Metric Learning is a booming area in machine learning with several different directions focusing on different problems. The paper fails to locates itself in the literature, how it compares itself into other techniques (both analytically and experimentally). \n\n2) The proposed technique seems to be very similar to SVD of learned representations. Connection to quantum field theory is well established but more simpler comparisons to SVD and other spectral techniques are not provided in metric learning.\n\n3) Novelty is not clear. There are interesting experiments in disentangled feature feature extraction and data generation. However, they are mostly proof of concept and lack of baselines. It is not clear what problem this technique solves better compared to other existing solutions. \n\nPaper is mostly written clearly. I do suggest putting Appendix A1 to the main paper though."}, "tcdate": 1571530772934}], "openreview_url": "https://openreview.net/forum?id=SJeS16EKPr", "arxiv_id": "1904.10387", "paper_pdf": "papers/SJeS16EKPr.pdf", "paper_pdf_sha256": "a27836b98fbc649ded272a72b1c5c6cab418185f35507c452f7c60bd2d1fbb83", "paper_pdf_bytes": 878876, "paper_pdf_source": "openreview", "code_url": "https://github.com/cbeny/RFA", "code_repository": "cbeny/RFA", "code_commit": "fe4a331f2a7baea1c41892815978d93b6d10b8c4", "code_archive": "repos/SJeS16EKPr.zip", "code_archive_sha256": "c50a2c1ff960ddfd965b025843f70aca34123aed1071d96da3549117a9078fbe", "code_archive_bytes": 527427, "code_file_count": 6, "code_extensions": {".jl": 3, ".py": 2, ".ipynb": 1}, "github_disk_usage_kb": 2070, "github_languages": {"Jupyter Notebook": 749107, "Julia": 18202, "Python": 7560}, "github_archived": false, "github_pushed_at": "2020-05-13T11:37:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/relevant-feature-extraction-for-statistical"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NdbUfhttc1", "year": 2024, "status": "rejected", "title": "Learning to Optimize for Reinforcement Learning", "authors": ["Qingfeng Lan", "A. Rupam Mahmood", "Shuicheng YAN", "Zhongwen Xu"], "authorids": ["~Qingfeng_Lan1", "~A._Rupam_Mahmood1", "~Shuicheng_YAN3", "~Zhongwen_Xu1"], "authors_source": "OpenReview API", "abstract": "In recent years, by leveraging more data, computation, and diverse tasks, learned optimizers have achieved remarkable success in supervised learning, outperforming classical hand-designed optimizers. Reinforcement learning (RL) is essentially different from supervised learning and in practice these learned optimizers do not work well even in simple RL tasks. We investigate this phenomenon and identity three issues. First, the gradients of an RL agent vary across a wide range in logarithms while their absolute values are in a small range, making neural networks hard to obtain accurate parameter updates. Second, the agent-gradient distribution is non-independent and identically distributed, leading to inefficient meta-training. Finally, due to highly stochastic agent-environment interactions, the agent-gradients have high bias and variance, which increase the difficulty of learning an optimizer for RL. We propose gradient processing, pipeline training, and a novel optimizer structure with good inductive bias to address these issues. By applying these techniques, for the first time, we show that learning an optimizer for RL from scratch is possible. Although only trained in toy tasks, our learned optimizer can generalize to unseen complex tasks in Brax.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "RXdjbfoEWp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4847/Reviewer_ssUw"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose meta-learning an optimizer for RL. They show that RL is uniquely challenging to optimize for. They then propose multiple techniques to learn an optimizer for RL from scratch that can generalize from toy tasks to Brax.", "review_text": "The authors propose meta-learning an optimizer for RL. They show that RL is uniquely challenging to optimize for. They then propose multiple techniques to learn an optimizer for RL from scratch that can generalize from toy tasks to Brax.", "strengths": "Originality:\n\n- This is a new problem setting that I have not seen before. It is clear that existing learned optimizers do not perform well in RL and it is good to see initial works in this direction.\n\n- The gradient processing is well-motivated and significant.\n\n- The architecture is also well-motivated and elegant.\n\nQuality:\n\n- The authors perform neat investigations into hypotheses about gradient-related challenges in RL.\n\n- The authors show impressive transfer performance.\n\nClarity:\n\n- The paper is very clearly written.\n\nSignificance:\n\n- This could ultimately lead to a superior optimizer for RL, which would be very significant.", "weaknesses": "Originality:\n\n- There is a section missing from the related works. In particular, it's the \"learning update rules / algorithms\" for RL literature. The setup seems to be *very closely related to* the setup from \"Discovering Reinforcement Learning Algorithms\" (LPG) [1], which is part of a broader field of meta-learning general RL update rules [2], [3]. \n\n- The pipeline training and the training setup of LPG seem closely related.\n\nClarity:\n\n- (Minor) Section 4.1: SeeAppendix B <= missing a space.\n\n- See clarification-related questions below.\n\nQuality:\n\n- On Section 4.1: The authors show that the agent-gradient distribution is non-IID. However, they do not show that the gradient distribution for normal supervised learning (SL) **is** IID. This is rather important to show, if the authors are claiming that RL is a uniquely challenging setting.\n\n- On Section 4.2: Again, the authors did not compare RL and SL, which is the purpose of this section. The authors could train a SL model and then see if the RNN can learn the identity function on the gradients from that training process. \n\n- The synthetic data is not representative of the SL training process. Furthermore, the injection of non-iid dynamics of the synthetic data seems to have been done ad-hoc and is not particularly meaningful. For example, what if the iid shift was far more extreme?\n\n- On Section 4.3: The problem of non-stationary targets is a well-studied phenomenon in RL, with plenty of possible prior works the authors could cite. This includes the deadly triad [4] and capacity loss [5]. \n\nSignificance:\n\n- The primary technical contributions seem to be the gradient processing, and the optimizer structure. The gradient processing is an impactful trick. The optimizer structure is hardly ablated or compared when there is plenty of literature on learned optimizer architectures.\n\n- The authors show limited transfer and the optimizer does not seem to generally perform on-par with Adam, despite being heavily inductively biased towards an Adam-like update.\n\n\n[1] Oh, Junhyuk, et al. \"Discovering reinforcement learning algorithms.\" Advances in Neural Information Processing Systems 33 (2020): 1060-1070.\n\n[2] Kirsch, Louis, Sjoerd van Steenkiste, and Jürgen Schmidhuber. \"Improving generalization in meta reinforcement learning using learned objectives.\" arXiv preprint arXiv:1910.04098 (2019).\n\n[3] Lu, Chris, et al. \"Discovered policy optimisation.\" Advances in Neural Information Processing Systems 35 (2022): 16455-16468.\n\n[4] Van Hasselt, Hado, et al. \"Deep reinforcement learning and the deadly triad.\" arXiv preprint arXiv:1812.02648 (2018).\n\n[5] Lyle, Clare, Mark Rowland, and Will Dabney. \"Understanding and preventing capacity loss in reinforcement learning.\" arXiv preprint arXiv:2204.09560 (2022).", "questions": "1. What is the difference between your pipeline training and the pipeline training of LPG?\n\n2. Is pipeline training desirable? Ideally we want a non-myopic optimizer, and the training dynamics early on in training heavily affect the distribution of parameters at the end of training. Can you ablate this?\n\n3. Why is the neural network so small? Is this common in the literature?\n\n4. In Section 6.1: Are you re-training STAR and L2LSGD, or are you taking pre-trained weights?\n\n5. Why are the optimizers and environments different in each plot in Figure 4? (e.g. why is VeLO exclusively for Ant and STAR for big_dense_long). Can you generate a more complete plot here?\n\n6. Many of the learned optimizers use ES to train their learned optimizers. Is there any particular reason you decided not to do this?\n\n7. On Section 4.2: How did the authors choose the hyperparameters (the rate and total amount of change) for generating non-iid data? \n\n8. On \"Quality\" Weaknesses (4.1 and 4.2) from above: These seem easy for the authors to address and I would be very curious about the results!\n\n9. Is there a reason you did not try other architectures from the learned optimizer literature?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose meta-learning an optimizer for RL. They show that RL is uniquely challenging to optimize for. They then propose multiple techniques to learn an optimizer for RL from scratch that can generalize from toy tasks to Brax.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "Originality:\n\n- This is a new problem setting that I have not seen before. It is clear that existing learned optimizers do not perform well in RL and it is good to see initial works in this direction.\n\n- The gradient processing is well-motivated and significant.\n\n- The architecture is also well-motivated and elegant.\n\nQuality:\n\n- The authors perform neat investigations into hypotheses about gradient-related challenges in RL.\n\n- The authors show impressive transfer performance.\n\nClarity:\n\n- The paper is very clearly written.\n\nSignificance:\n\n- This could ultimately lead to a superior optimizer for RL, which would be very significant.", "weaknesses": "Originality:\n\n- There is a section missing from the related works. In particular, it's the \"learning update rules / algorithms\" for RL literature. The setup seems to be *very closely related to* the setup from \"Discovering Reinforcement Learning Algorithms\" (LPG) [1], which is part of a broader field of meta-learning general RL update rules [2], [3]. \n\n- The pipeline training and the training setup of LPG seem closely related.\n\nClarity:\n\n- (Minor) Section 4.1: SeeAppendix B <= missing a space.\n\n- See clarification-related questions below.\n\nQuality:\n\n- On Section 4.1: The authors show that the agent-gradient distribution is non-IID. However, they do not show that the gradient distribution for normal supervised learning (SL) **is** IID. This is rather important to show, if the authors are claiming that RL is a uniquely challenging setting.\n\n- On Section 4.2: Again, the authors did not compare RL and SL, which is the purpose of this section. The authors could train a SL model and then see if the RNN can learn the identity function on the gradients from that training process. \n\n- The synthetic data is not representative of the SL training process. Furthermore, the injection of non-iid dynamics of the synthetic data seems to have been done ad-hoc and is not particularly meaningful. For example, what if the iid shift was far more extreme?\n\n- On Section 4.3: The problem of non-stationary targets is a well-studied phenomenon in RL, with plenty of possible prior works the authors could cite. This includes the deadly triad [4] and capacity loss [5]. \n\nSignificance:\n\n- The primary technical contributions seem to be the gradient processing, and the optimizer structure. The gradient processing is an impactful trick. The optimizer structure is hardly ablated or compared when there is plenty of literature on learned optimizer architectures.\n\n- The authors show limited transfer and the optimizer does not seem to generally perform on-par with Adam, despite being heavily inductively biased towards an Adam-like update.\n\n\n[1] Oh, Junhyuk, et al. \"Discovering reinforcement learning algorithms.\" Advances in Neural Information Processing Systems 33 (2020): 1060-1070.\n\n[2] Kirsch, Louis, Sjoerd van Steenkiste, and Jürgen Schmidhuber. \"Improving generalization in meta reinforcement learning using learned objectives.\" arXiv preprint arXiv:1910.04098 (2019).\n\n[3] Lu, Chris, et al. \"Discovered policy optimisation.\" Advances in Neural Information Processing Systems 35 (2022): 16455-16468.\n\n[4] Van Hasselt, Hado, et al. \"Deep reinforcement learning and the deadly triad.\" arXiv preprint arXiv:1812.02648 (2018).\n\n[5] Lyle, Clare, Mark Rowland, and Will Dabney. \"Understanding and preventing capacity loss in reinforcement learning.\" arXiv preprint arXiv:2204.09560 (2022).", "questions": "1. What is the difference between your pipeline training and the pipeline training of LPG?\n\n2. Is pipeline training desirable? Ideally we want a non-myopic optimizer, and the training dynamics early on in training heavily affect the distribution of parameters at the end of training. Can you ablate this?\n\n3. Why is the neural network so small? Is this common in the literature?\n\n4. In Section 6.1: Are you re-training STAR and L2LSGD, or are you taking pre-trained weights?\n\n5. Why are the optimizers and environments different in each plot in Figure 4? (e.g. why is VeLO exclusively for Ant and STAR for big_dense_long). Can you generate a more complete plot here?\n\n6. Many of the learned optimizers use ES to train their learned optimizers. Is there any particular reason you decided not to do this?\n\n7. On Section 4.2: How did the authors choose the hyperparameters (the rate and total amount of change) for generating non-iid data? \n\n8. On \"Quality\" Weaknesses (4.1 and 4.2) from above: These seem easy for the authors to address and I would be very curious about the results!\n\n9. Is there a reason you did not try other architectures from the learned optimizer literature?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698809942054}, {"id": "iktcj4xqAi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4847/Reviewer_tcjw"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose a new learned optimizer, Optim4RL, to address the challenges of using learned optimizers in RL.\n\nWhile learned optimization has shown benefits in the supervised learning community, SOTA optimizers for Supervised Learning (SL) fail in the RL setting. The authors investigate this phenomenon by analyzing the distribution of the gradients of agent parameters at the start, middle, and end of training. Through this analysis, the authors demonstrate that the gradients are non-I.I.D. Moreover, the absolute values of the gradients lie in a small range. The authors then demonstrate the difficulty of the RNN module -- commonly used in the learned optimizer -- to approximate an identity function using the gradient data (75% accuracy). Using this analysis, they underscore the bias and variance in the gradients as a key issue that makes learned optimization hard in RL and argue that this is further exacerbated by the bi-level optimization in learned optimizers. (poor optimizer -- poor policy -- lower quality data)\n\nThe authors then propose three key ways to mitigate these issues: \n- Gradient Processing: a 1-1- mapping that uses a log transformation to magnify absolute value differences between small gradients to mitigate the logarithmic gradient variation. This boosts the accuracy of the RNN on the identity task from 75% to 90%\n- Pipeline Training: add diversity to the gradient inputs to the learned optimizer through a distributed training regime by parallelly training multiple agents being reset at different periods and using all of their gradients for the learned optimizer. This mitigates the non-iid nature of the data since data now comes from different points of training\n- Biasing the optimizer: Building on the analysis of [Harrison et al., 2022], they utilize both the gradient and its squared value as inputs to two RNNs. This mitigates the need to approximate square roots and division by the learned optimizer and stabilizes the meta-update\n\nThe combination of these three components -- Optim4RL -- demonstrates improved stability and effectiveness in optimizing RL tasks compared to baselines of hand-designed optimizers (Adam and RMSProp), learned optimizers for Supervised learning (L2LGD$^2$, STAR, and VeLO), and linear parameter update instead of a squared\n\n[Harrison et al., 2022] Harrison, J., Metz, L., & Sohl-Dickstein, J. (2022). A closer look at learned optimization: Stability, robustness, and inductive biases. Advances in Neural Information Processing Systems, 35, 3758-3773.", "review_text": "The authors propose a new learned optimizer, Optim4RL, to address the challenges of using learned optimizers in RL.\n\nWhile learned optimization has shown benefits in the supervised learning community, SOTA optimizers for Supervised Learning (SL) fail in the RL setting. The authors investigate this phenomenon by analyzing the distribution of the gradients of agent parameters at the start, middle, and end of training. Through this analysis, the authors demonstrate that the gradients are non-I.I.D. Moreover, the absolute values of the gradients lie in a small range. The authors then demonstrate the difficulty of the RNN module -- commonly used in the learned optimizer -- to approximate an identity function using the gradient data (75% accuracy). Using this analysis, they underscore the bias and variance in the gradients as a key issue that makes learned optimization hard in RL and argue that this is further exacerbated by the bi-level optimization in learned optimizers. (poor optimizer -- poor policy -- lower quality data)\n\nThe authors then propose three key ways to mitigate these issues: \n- Gradient Processing: a 1-1- mapping that uses a log transformation to magnify absolute value differences between small gradients to mitigate the logarithmic gradient variation. This boosts the accuracy of the RNN on the identity task from 75% to 90%\n- Pipeline Training: add diversity to the gradient inputs to the learned optimizer through a distributed training regime by parallelly training multiple agents being reset at different periods and using all of their gradients for the learned optimizer. This mitigates the non-iid nature of the data since data now comes from different points of training\n- Biasing the optimizer: Building on the analysis of [Harrison et al., 2022], they utilize both the gradient and its squared value as inputs to two RNNs. This mitigates the need to approximate square roots and division by the learned optimizer and stabilizes the meta-update\n\nThe combination of these three components -- Optim4RL -- demonstrates improved stability and effectiveness in optimizing RL tasks compared to baselines of hand-designed optimizers (Adam and RMSProp), learned optimizers for Supervised learning (L2LGD$^2$, STAR, and VeLO), and linear parameter update instead of a squared\n\n[Harrison et al., 2022] Harrison, J., Metz, L., & Sohl-Dickstein, J. (2022). A closer look at learned optimization: Stability, robustness, and inductive biases. Advances in Neural Information Processing Systems, 35, 3758-3773.", "strengths": "### Originality\nThe paper tackles a novel direction of RL-specific learned optimization by looking deeper into what kind of RL-specific inputs need to be adapted.\n\n### Quality\nThe work is insightful and generally conducted comprehensively.\n\n### Clarity\nThe paper is written clearly and understandably. Overall, the presentation is clear and well done.\n\n\n### Significance\nThe research direction is significant since learned optimization is yet to take hold in RL properly and is very important if achieved.", "weaknesses": "There seem to be a lot of central design decisions/hyperparameters in the training procedure that are not justified:\n- 4 inner updated per outer update\n- The decision to average returns over ten runs\n- The threshold for gradient processing\n- Epsilon in the parameter update\n\nThe agglomerative procedure to incorporate diversity in the gradient distribution seems not fully ablated. See my questions on this for further details.\n\nI am unsure if 10 GPU years is a realistically feasible budget for most practitioners. One of the issues with learned optimization in SL has been this exact problem. I think commentary on how to bring this cost down would be highly beneficial for hte community, especially given the recent surge in JAX-based parallelization with developments such as PureJAXRL (https://github.com/luchris429/purejaxrl).", "questions": "- What happens when we don't do individual pre-processing steps?  -- Are there any ablations that demonstrate the effectiveness of individual modifications?\n- What constitutes the middle of training? is it the same for each environment or different across environments?\n- How many seeds were the experiments reported on? How did the authors determine them?\n- Resetting provides the optimizer data at different training stages. Have the authors analyzed how different values of m and n impact this? Do we require them to be equal all the time? \n- Given that pipeline training can be computationally expensive, have the authors examined methods to extract maximum benefit from this procedure? For example, could reset times be adapted by leveraging optimizer reset properties? [Asadi et al., 2023]\n- Does the learned optimizer mitigate the requirement for dynamic hyperparameter optimization [Mohan et al., 2023]? To what extent is the problem addressed in this work related to the AutoRL problem, given that there are still optimizer-related hyperparameters? \n\n[Asadi et al., 2023]  Asadi, K., Fakoor, R., & Sabach, S. (2023). Resetting the Optimizer in Deep RL: An Empirical Study. arXiv preprint arXiv:2306.17833.\n\n[Mohan et al, 2023] Mohan, A., Benjamins, C., Wienecke, K., Dockhorn, A., & Lindauer, M. (2023). AutoRL Hyperparameter Landscapes. AutoML Conference 2023 (https://openreview.net/forum?id=Ec09TcV_HKq).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a new learned optimizer, Optim4RL, to address the challenges of using learned optimizers in RL.\n\nWhile learned optimization has shown benefits in the supervised learning community, SOTA optimizers for Supervised Learning (SL) fail in the RL setting. The authors investigate this phenomenon by analyzing the distribution of the gradients of agent parameters at the start, middle, and end of training. Through this analysis, the authors demonstrate that the gradients are non-I.I.D. Moreover, the absolute values of the gradients lie in a small range. The authors then demonstrate the difficulty of the RNN module -- commonly used in the learned optimizer -- to approximate an identity function using the gradient data (75% accuracy). Using this analysis, they underscore the bias and variance in the gradients as a key issue that makes learned optimization hard in RL and argue that this is further exacerbated by the bi-level optimization in learned optimizers. (poor optimizer -- poor policy -- lower quality data)\n\nThe authors then propose three key ways to mitigate these issues: \n- Gradient Processing: a 1-1- mapping that uses a log transformation to magnify absolute value differences between small gradients to mitigate the logarithmic gradient variation. This boosts the accuracy of the RNN on the identity task from 75% to 90%\n- Pipeline Training: add diversity to the gradient inputs to the learned optimizer through a distributed training regime by parallelly training multiple agents being reset at different periods and using all of their gradients for the learned optimizer. This mitigates the non-iid nature of the data since data now comes from different points of training\n- Biasing the optimizer: Building on the analysis of [Harrison et al., 2022], they utilize both the gradient and its squared value as inputs to two RNNs. This mitigates the need to approximate square roots and division by the learned optimizer and stabilizes the meta-update\n\nThe combination of these three components -- Optim4RL -- demonstrates improved stability and effectiveness in optimizing RL tasks compared to baselines of hand-designed optimizers (Adam and RMSProp), learned optimizers for Supervised learning (L2LGD$^2$, STAR, and VeLO), and linear parameter update instead of a squared\n\n[Harrison et al., 2022] Harrison, J., Metz, L., & Sohl-Dickstein, J. (2022). A closer look at learned optimization: Stability, robustness, and inductive biases. Advances in Neural Information Processing Systems, 35, 3758-3773.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "### Originality\nThe paper tackles a novel direction of RL-specific learned optimization by looking deeper into what kind of RL-specific inputs need to be adapted.\n\n### Quality\nThe work is insightful and generally conducted comprehensively.\n\n### Clarity\nThe paper is written clearly and understandably. Overall, the presentation is clear and well done.\n\n\n### Significance\nThe research direction is significant since learned optimization is yet to take hold in RL properly and is very important if achieved.", "weaknesses": "There seem to be a lot of central design decisions/hyperparameters in the training procedure that are not justified:\n- 4 inner updated per outer update\n- The decision to average returns over ten runs\n- The threshold for gradient processing\n- Epsilon in the parameter update\n\nThe agglomerative procedure to incorporate diversity in the gradient distribution seems not fully ablated. See my questions on this for further details.\n\nI am unsure if 10 GPU years is a realistically feasible budget for most practitioners. One of the issues with learned optimization in SL has been this exact problem. I think commentary on how to bring this cost down would be highly beneficial for hte community, especially given the recent surge in JAX-based parallelization with developments such as PureJAXRL (https://github.com/luchris429/purejaxrl).", "questions": "- What happens when we don't do individual pre-processing steps?  -- Are there any ablations that demonstrate the effectiveness of individual modifications?\n- What constitutes the middle of training? is it the same for each environment or different across environments?\n- How many seeds were the experiments reported on? How did the authors determine them?\n- Resetting provides the optimizer data at different training stages. Have the authors analyzed how different values of m and n impact this? Do we require them to be equal all the time? \n- Given that pipeline training can be computationally expensive, have the authors examined methods to extract maximum benefit from this procedure? For example, could reset times be adapted by leveraging optimizer reset properties? [Asadi et al., 2023]\n- Does the learned optimizer mitigate the requirement for dynamic hyperparameter optimization [Mohan et al., 2023]? To what extent is the problem addressed in this work related to the AutoRL problem, given that there are still optimizer-related hyperparameters? \n\n[Asadi et al., 2023]  Asadi, K., Fakoor, R., & Sabach, S. (2023). Resetting the Optimizer in Deep RL: An Empirical Study. arXiv preprint arXiv:2306.17833.\n\n[Mohan et al, 2023] Mohan, A., Benjamins, C., Wienecke, K., Dockhorn, A., & Lindauer, M. (2023). AutoRL Hyperparameter Landscapes. AutoML Conference 2023 (https://openreview.net/forum?id=Ec09TcV_HKq).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698764497166}, {"id": "91MqdU1Gi4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4847/Reviewer_8FzQ"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose a meta-learning procedure for learning optimizers for reinforcement learning called Optim4RL. Their method has the following key components:\n- Pipeline training: Use multiple agents each at different training stages (early/mid/late) to make data distribution (more) stationary. \n- Gradient pre-processing: Transform the gradient so that the input is sensitive to changes of the gradient in (approximately) log space. \n- Inductive bias: Structure the update using a form similar to Adam, providing good inductive bias\n\nIn their experiments Optim4RL is shown to (1) outperform existing learnt optimizers, and (2) generalise to problems outside of the training set.", "review_text": "The authors propose a meta-learning procedure for learning optimizers for reinforcement learning called Optim4RL. Their method has the following key components:\n- Pipeline training: Use multiple agents each at different training stages (early/mid/late) to make data distribution (more) stationary. \n- Gradient pre-processing: Transform the gradient so that the input is sensitive to changes of the gradient in (approximately) log space. \n- Inductive bias: Structure the update using a form similar to Adam, providing good inductive bias\n\nIn their experiments Optim4RL is shown to (1) outperform existing learnt optimizers, and (2) generalise to problems outside of the training set.", "strengths": "- Each of the parts of the proposed learnt optimizer solve important problems in meta-learning optimizers, and are well motivated. \n- The toy problem where the optimizer has to learn the identity function is simple and informative. \n- The paper has informative analysis on the gradient distribution (e.g. I like the plots in Figure 5 visualising the train-test marginal distribution of gradients). \n- Achieves generalisation to different tasks (Brax) from simple grid based problems. \n- Generally the paper is clear and easy to follow.", "weaknesses": "- The optimizer achieves marginally worse performance than Adam on the tasks that it is meta-trained on. It seems like the learnt optimizer should at least be able to \"overfit\" to the training task to outperform Adam here. \n- On unseen tasks the optimizer is significantly worse than Adam. \n- The training and test tasks are relatively toy problems (the authors do acknowledge this weaknesses).", "questions": "- The below text confused me - why do we need to check that the model has enough capacity to represent the identity function (e.g. a single linear layer can represent the identity function easily)?\nAlso, why do we need an RNN on this problem (is the input not just the current gradient?)? \n\n> To verify that the model has enough expressiveness and capacity to represent an identity function, we further train it with randomly generated data where data size is similar to the data size of the agent-gradients.\n\n- Transforming the gradient passed to the RNN into a richer representation makes sense. Additionally this seems to help a lot in terms of performance so it seems worth digging deeper into. Were other transformations tried - e.g. fourier feature embedding? \n\n- Would it be possible to add the STAR benchmark to Figure 4 (ant), and for Table 5? This would allow us to see how well Optim4RL generalizes relative to another learnt optimizer. \n\n- In figure 4 (b) it seems like STAR is starting to learn a bit. Is it possible that with a bit more hyper-parameter tuning it would match the other optimizers in performance? \n\nI acknowledge that a lot of my questions require more compute, and that this is a very compute heavy task. I do think that these would significantly strengthen the paper - as they would help make the results more decisive.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a meta-learning procedure for learning optimizers for reinforcement learning called Optim4RL. Their method has the following key components:\n- Pipeline training: Use multiple agents each at different training stages (early/mid/late) to make data distribution (more) stationary. \n- Gradient pre-processing: Transform the gradient so that the input is sensitive to changes of the gradient in (approximately) log space. \n- Inductive bias: Structure the update using a form similar to Adam, providing good inductive bias\n\nIn their experiments Optim4RL is shown to (1) outperform existing learnt optimizers, and (2) generalise to problems outside of the training set.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Each of the parts of the proposed learnt optimizer solve important problems in meta-learning optimizers, and are well motivated. \n- The toy problem where the optimizer has to learn the identity function is simple and informative. \n- The paper has informative analysis on the gradient distribution (e.g. I like the plots in Figure 5 visualising the train-test marginal distribution of gradients). \n- Achieves generalisation to different tasks (Brax) from simple grid based problems. \n- Generally the paper is clear and easy to follow.", "weaknesses": "- The optimizer achieves marginally worse performance than Adam on the tasks that it is meta-trained on. It seems like the learnt optimizer should at least be able to \"overfit\" to the training task to outperform Adam here. \n- On unseen tasks the optimizer is significantly worse than Adam. \n- The training and test tasks are relatively toy problems (the authors do acknowledge this weaknesses).", "questions": "- The below text confused me - why do we need to check that the model has enough capacity to represent the identity function (e.g. a single linear layer can represent the identity function easily)?\nAlso, why do we need an RNN on this problem (is the input not just the current gradient?)? \n\n> To verify that the model has enough expressiveness and capacity to represent an identity function, we further train it with randomly generated data where data size is similar to the data size of the agent-gradients.\n\n- Transforming the gradient passed to the RNN into a richer representation makes sense. Additionally this seems to help a lot in terms of performance so it seems worth digging deeper into. Were other transformations tried - e.g. fourier feature embedding? \n\n- Would it be possible to add the STAR benchmark to Figure 4 (ant), and for Table 5? This would allow us to see how well Optim4RL generalizes relative to another learnt optimizer. \n\n- In figure 4 (b) it seems like STAR is starting to learn a bit. Is it possible that with a bit more hyper-parameter tuning it would match the other optimizers in performance? \n\nI acknowledge that a lot of my questions require more compute, and that this is a very compute heavy task. I do think that these would significantly strengthen the paper - as they would help make the results more decisive.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698744554403}, {"id": "KUMoyBiF0r", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4847/Reviewer_wz35"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper presents a method for meta-learned optimization in RL, named Optim4RL. This consists of three components - pipeline training, gradient transformations, and an update formulation - which are designed to tackle specific issues in this setting. The evaluation investigates the performance of this method when meta-trained on simple grid-world environments and evaluated on much more challenging environments, primarily evaluating against Adam and RMSProp, in addition to a collection of alternative meta-learned optimizers on in-distribution tasks.\n\nI am recommending rejection for this paper, primarily due to the lack of ablations. However, I would be very willing to increase my score if the suggested ablations were performed, in addition to a discussion of the wider literature for this problem setting.", "review_text": "The paper presents a method for meta-learned optimization in RL, named Optim4RL. This consists of three components - pipeline training, gradient transformations, and an update formulation - which are designed to tackle specific issues in this setting. The evaluation investigates the performance of this method when meta-trained on simple grid-world environments and evaluated on much more challenging environments, primarily evaluating against Adam and RMSProp, in addition to a collection of alternative meta-learned optimizers on in-distribution tasks.\n\nI am recommending rejection for this paper, primarily due to the lack of ablations. However, I would be very willing to increase my score if the suggested ablations were performed, in addition to a discussion of the wider literature for this problem setting.", "strengths": "1. Related work extensively covers RL optimization and meta-optimization literature.\n2. The paper is impressively well-written and structured. Section 4 is particularly well-structured, presenting a clear set of hypotheses about the problems with meta-optimization in RL.\n3. Each proposed component is simple, but clearly motivated and presented in Section 5 and Figure 3.\n4. Many related methods are compared against Optim4RL in Figure 4, however, these could be evaluated further (see weaknesses).", "weaknesses": "1. The predominant flaw with this paper is the evaluation of the proposed components. Section 5 is highly systematic in motivating each problem, before proposing a component as a solution. However, the evaluation does not ablate the components, making it impossible to discern their individual impact. The exception to this is LinearOptim, which is an ablation of the proposed inductive bias, however, I believe this should be highlighted. Whilst the comparison to existing baselines is interesting, this is a __fundamental__ requirement when evaluating a model composed of multiple novel components.\n2. The presentation of the results could be clearer for drawing conclusions. Whilst training curves are useful, they make it difficult to quantitatively determine significance.\n3. The wide range of baselines in Figure 4 is good, but it is unclear why these are not carried forward for the remainder of the evaluation. Given that meta-training is the largest computational cost, these shouldn't be out-of-budget to run. In particular, STAR is performative enough that it is plausible it would achieve competitive performance on the remaining tasks, which should be investigated.\n4. The inability of LinearOptim to learn anything is very surprising and should be investigated further to ensure it is not erroneous.\n5. There is a broad base of related work on meta-learned RL objective functions, which is not discussed or compared against. While these are a different class of inductive bias, they are solving the same problem as the class of meta-optimizers discussed here. Notably, the evaluation procedure and environments are from Oh et al. (2020), a learned objective algorithm, but it is not compared against. While it is understandable to not compare Optim4RL against all of these methods, they should at least be discussed as alternative approaches to the same problem in the related work. Namely, EPG (Houthooft et al., 2018), LPG (Oh et al., 2020), MetaGenRL (Kirsch et al., 2020), ML^3 (Bechtle et al., 2021), SymLA (Kirsch et al., 2022), DPO (Lu et al., 2022), GROOVE (Jackson et al., 2023).\n6. Many of the claims are misleading or ambiguous. In conclusion, the claim that Optim4RL is \"the first learned optimizer that can be meta-learned to optimize RL tasks entirely from scratch\" is confusing, since all existing learned optimizers can be and are applied to RL in this paper. If this is intended to claim this is the first meta-learned optimizer designed for RL, then the omission meta-learned objective function literature becomes even more apparent, since there is extensive work solving the same problem for RL.", "questions": "1. Major typo in Algorithm 1: the sign and magnitude of your update are both computed from o_1. I assume this is a typo since, if correct, your method would only be capable of outputting large positive updates and small negative updates.\n2. Transformation of the gradient output is a major component of gradient processing, but only the input transformations are discussed in the main body. An expanded form of the output could be presented in the main body.\n3. In Figure 5, you suggest the overlap in the support of the gradients is a predictor of performance. A simple experiment to evaluate this would be retraining the optimizer with rescaled rewards on the grid-world tasks, which would shift the support of the meta-training gradients. If this improved performance, this would significantly strengthen the hypothesis.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method for meta-learned optimization in RL, named Optim4RL. This consists of three components - pipeline training, gradient transformations, and an update formulation - which are designed to tackle specific issues in this setting. The evaluation investigates the performance of this method when meta-trained on simple grid-world environments and evaluated on much more challenging environments, primarily evaluating against Adam and RMSProp, in addition to a collection of alternative meta-learned optimizers on in-distribution tasks.\n\nI am recommending rejection for this paper, primarily due to the lack of ablations. However, I would be very willing to increase my score if the suggested ablations were performed, in addition to a discussion of the wider literature for this problem setting.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "1. Related work extensively covers RL optimization and meta-optimization literature.\n2. The paper is impressively well-written and structured. Section 4 is particularly well-structured, presenting a clear set of hypotheses about the problems with meta-optimization in RL.\n3. Each proposed component is simple, but clearly motivated and presented in Section 5 and Figure 3.\n4. Many related methods are compared against Optim4RL in Figure 4, however, these could be evaluated further (see weaknesses).", "weaknesses": "1. The predominant flaw with this paper is the evaluation of the proposed components. Section 5 is highly systematic in motivating each problem, before proposing a component as a solution. However, the evaluation does not ablate the components, making it impossible to discern their individual impact. The exception to this is LinearOptim, which is an ablation of the proposed inductive bias, however, I believe this should be highlighted. Whilst the comparison to existing baselines is interesting, this is a __fundamental__ requirement when evaluating a model composed of multiple novel components.\n2. The presentation of the results could be clearer for drawing conclusions. Whilst training curves are useful, they make it difficult to quantitatively determine significance.\n3. The wide range of baselines in Figure 4 is good, but it is unclear why these are not carried forward for the remainder of the evaluation. Given that meta-training is the largest computational cost, these shouldn't be out-of-budget to run. In particular, STAR is performative enough that it is plausible it would achieve competitive performance on the remaining tasks, which should be investigated.\n4. The inability of LinearOptim to learn anything is very surprising and should be investigated further to ensure it is not erroneous.\n5. There is a broad base of related work on meta-learned RL objective functions, which is not discussed or compared against. While these are a different class of inductive bias, they are solving the same problem as the class of meta-optimizers discussed here. Notably, the evaluation procedure and environments are from Oh et al. (2020), a learned objective algorithm, but it is not compared against. While it is understandable to not compare Optim4RL against all of these methods, they should at least be discussed as alternative approaches to the same problem in the related work. Namely, EPG (Houthooft et al., 2018), LPG (Oh et al., 2020), MetaGenRL (Kirsch et al., 2020), ML^3 (Bechtle et al., 2021), SymLA (Kirsch et al., 2022), DPO (Lu et al., 2022), GROOVE (Jackson et al., 2023).\n6. Many of the claims are misleading or ambiguous. In conclusion, the claim that Optim4RL is \"the first learned optimizer that can be meta-learned to optimize RL tasks entirely from scratch\" is confusing, since all existing learned optimizers can be and are applied to RL in this paper. If this is intended to claim this is the first meta-learned optimizer designed for RL, then the omission meta-learned objective function literature becomes even more apparent, since there is extensive work solving the same problem for RL.", "questions": "1. Major typo in Algorithm 1: the sign and magnitude of your update are both computed from o_1. I assume this is a typo since, if correct, your method would only be capable of outputting large positive updates and small negative updates.\n2. Transformation of the gradient output is a major component of gradient processing, but only the input transformations are discussed in the main body. An expanded form of the output could be presented in the main body.\n3. In Figure 5, you suggest the overlap in the support of the gradients is a predictor of performance. A simple experiment to evaluate this would be retraining the optimizer with rescaled rewards on the grid-world tasks, which would shift the support of the meta-training gradients. If this improved performance, this would significantly strengthen the hypothesis.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698598474807}], "openreview_url": "https://openreview.net/forum?id=NdbUfhttc1", "arxiv_id": "2302.01470", "paper_pdf": "papers/NdbUfhttc1.pdf", "paper_pdf_sha256": "0f9a7e811e1a5386e20500e6a27b8eaa029052afb4b021f98e4ca29b873a77d9", "paper_pdf_bytes": 1647564, "paper_pdf_source": "openreview", "code_url": "https://github.com/sail-sg/optim4rl", "code_repository": "sail-sg/optim4rl", "code_commit": "523d42ecea582d432246e80fce0e88442497a6d0", "code_archive": "repos/NdbUfhttc1.zip", "code_archive_sha256": "8a18f5af557ee4f912410e0687ed05ef2bde44be1df635b28ded45d1741ed836", "code_archive_bytes": 118641, "code_file_count": 35, "code_extensions": {".py": 34, ".sh": 1}, "github_disk_usage_kb": 138, "github_languages": {"Python": 222651, "Shell": 6976}, "github_archived": false, "github_pushed_at": "2024-11-27T01:33:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-optimize-for-reinforcement"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bGC7Ai125lR", "year": 2023, "status": "rejected", "title": "Towards Understanding How Machines Can Learn Causal Overhypotheses ", "authors": ["Eliza Kosoy", "David Chan", "Adrian Liu", "Jasmine Collins", "Bryanna Kaufmann", "Sandy Huang", "Jessica B Hamrick", "John Canny", "Nan Rosemary Ke", "Alison Gopnik"], "authorids": ["~Eliza_Kosoy1", "~David_Chan3", "~Adrian_Liu1", "~Jasmine_Collins1", "bryannakaufmann@berkeley.edu", "~Sandy_Huang1", "~Jessica_B_Hamrick1", "~John_Canny1", "~Nan_Rosemary_Ke1", "~Alison_Gopnik1"], "authors_source": "OpenReview API", "abstract": "Recent work in machine learning and cognitive science has suggested that understanding causal information is essential to the development of intelligence. One of the key challenges for current machine learning algorithms is modeling and understanding causal overhypotheses: transferable abstract hypotheses about sets of causal relationships. In contrast, even young children spontaneously learn causal overhypotheses, and use these to guide their exploration or to generalize to new situations. This has been demonstrated in a variety of cognitive science experiments using the “blicket detector” environment. We present a causal learning benchmark adapting the “blicket\" environment for machine learning agents and evaluate a range of state-of-the-art methods in this environment. We find that although most agents have no problem learning causal structures seen during training, they are unable to learn causal overhypotheses from these experiences, and thus cannot generalize to new settings. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "vorcUeGnyI9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5003/Reviewer_6AJY"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper leverages the (virtual) blicket detector environment to empirically evaluate the ability of machine learning models to learn and use causal overhypotheses, which refer to those hypotheses that concern not specific causal relations but more general constraints about causal relations, such as whether a causal mechanism is conjunctive or disjunctive. This work is motivated by solid findings in the cognitive science literature on the competence of young children to learn causal overhypotheses and exploit them to probe causal relations in this environment. It builds on the setup in (Kosoy et al. 2022) and designs interesting experiments to assess a range of machine learning models regarding the learning of overhypotheses, including popular deep reinforcement learning algorithms, some imitation learning algorithms, and some large language models. Various weaknesses of these machine learning agents are empirically demonstrated.  ", "review_text": "It is a useful empirical study about an important topic that should receive more attention in machine learning. ", "strengths": "Strengths:\n\n1. Learning causal overhypotheses is an important task to be addressed in machine learning.\n2. A variety of machine learning models are considered and assessed.\n3. The experimental results are mostly interesting and telling, and some potentially useful benchmarks are given. \n\nWeaknesses:\n\n1. The distinction between learning an overhypothesis and exploiting the overhypothesis in exploration or inference of causal relations is not always clearly made. As I understand it, the target overhypothesis in this paper concerns whether the causal mechanism is conjunctive or disjunctive. If so, the results in the paper seem to suggest that at least the RL agents and the language models are able to learn the correct overhypothesis most of the time; what they seem to be unable to do well in the blicket environment is take advantage of the overhypothesis to figure out which objects are blickets. Is this reading correct? If it is, it is not clearly described or explained in the paper.\n\n2. The work is purely empirical and limited to certain ways to apply reinforcement learning, imitation learning, and large language models to causal inference. It is unclear whether the observed weaknesses are due to these approaches or to the authors' particular implementations of these approaches for causal inference.  ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper leverages the (virtual) blicket detector environment to empirically evaluate the ability of machine learning models to learn and use causal overhypotheses, which refer to those hypotheses that concern not specific causal relations but more general constraints about causal relations, such as whether a causal mechanism is conjunctive or disjunctive. This work is motivated by solid findings in the cognitive science literature on the competence of young children to learn causal overhypotheses and exploit them to probe causal relations in this environment. It builds on the setup in (Kosoy et al. 2022) and designs interesting experiments to assess a range of machine learning models regarding the learning of overhypotheses, including popular deep reinforcement learning algorithms, some imitation learning algorithms, and some large language models. Various weaknesses of these machine learning agents are empirically demonstrated.  ", "strength_and_weaknesses": "Strengths:\n\n1. Learning causal overhypotheses is an important task to be addressed in machine learning.\n2. A variety of machine learning models are considered and assessed.\n3. The experimental results are mostly interesting and telling, and some potentially useful benchmarks are given. \n\nWeaknesses:\n\n1. The distinction between learning an overhypothesis and exploiting the overhypothesis in exploration or inference of causal relations is not always clearly made. As I understand it, the target overhypothesis in this paper concerns whether the causal mechanism is conjunctive or disjunctive. If so, the results in the paper seem to suggest that at least the RL agents and the language models are able to learn the correct overhypothesis most of the time; what they seem to be unable to do well in the blicket environment is take advantage of the overhypothesis to figure out which objects are blickets. Is this reading correct? If it is, it is not clearly described or explained in the paper.\n\n2. The work is purely empirical and limited to certain ways to apply reinforcement learning, imitation learning, and large language models to causal inference. It is unclear whether the observed weaknesses are due to these approaches or to the authors' particular implementations of these approaches for causal inference.  ", "clarity,_quality,_novelty_and_reproducibility": "This work is a continuation of (Kosoy et al. 2022) (by the way, some remarks in the paper strongly suggest that it is probably the same team's work, which should have been avoided for the sake of blind review), but since very little has been written on this important topic, I think the originality is considerable. The paper is mostly clear, but as I said above, the implications of the experimental results are not always clearly delineated. In particular, it does not seem to be entirely fair to suggest that these machine learning models are not good at learning the overhypotheses considered in this paper. The difficulty seems to have more to do with using the overhypotheses. \n\nAs an empirical study, the overall quality is reasonably high. I also believe the reproducibility will be reasonably good.\n\nMinor comment: it is unclear to me what roles are played by the baselines described in Section 4.4. ", "summary_of_the_review": "It is a useful empirical study about an important topic that should receive more attention in machine learning. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667142604118}, {"id": "ao75_4S7UL", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5003/Reviewer_cvDM"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper utilized a recently developed framework of \"blicket\" environment and conducted experiments using reinforcement learning, imitation learning and Language models. The goal of the paper is to understand if these machine learning method could acquire a causal understanding of the combinations of objectives to the \"blicket\".\n\n ", "review_text": "Learning causal over-hypothesis is essential for AI to robustly help us to deal with automated tasks under the challenge of distribution shift. If my understanding were correct, the paper proposed a benchmark suite, which I am not fully convinced about the importance of this suite compared to constructing some other benchmark to evaluate the causal performances of machine learning algorithms. ", "strengths": "Strength:\nI think the paper is quite novel in trying to bridge other non-computational disciplines to ICLR.\n\nWeakness:\nIf I understood it correctly, the paper seems to be only running some experiments and check some performances which has an association with the causal measures on the over-hypothesis. But the paper did not give any insight. It looks to be providing a benchmark suite. In the RL experiment, the visual image is replaced with symbolic representation also. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper utilized a recently developed framework of \"blicket\" environment and conducted experiments using reinforcement learning, imitation learning and Language models. The goal of the paper is to understand if these machine learning method could acquire a causal understanding of the combinations of objectives to the \"blicket\".\n\n ", "strength_and_weaknesses": "Strength:\nI think the paper is quite novel in trying to bridge other non-computational disciplines to ICLR.\n\nWeakness:\nIf I understood it correctly, the paper seems to be only running some experiments and check some performances which has an association with the causal measures on the over-hypothesis. But the paper did not give any insight. It looks to be providing a benchmark suite. In the RL experiment, the visual image is replaced with symbolic representation also. ", "clarity,_quality,_novelty_and_reproducibility": "Novelty: Good\nReproducibility: Not sure", "summary_of_the_review": "Learning causal over-hypothesis is essential for AI to robustly help us to deal with automated tasks under the challenge of distribution shift. If my understanding were correct, the paper proposed a benchmark suite, which I am not fully convinced about the importance of this suite compared to constructing some other benchmark to evaluate the causal performances of machine learning algorithms. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666878872964}, {"id": "nGoPfM7iFT0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5003/Reviewer_ypFh"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper presents a causal learning benchmark to test the learning of causal overhypotheses. The authors test several RL algorithms as well as pretrained language models on this benchmark. ", "review_text": "I think the main idea of the paper is novel and attractive. However, I need a lot of clarifications to properly evaluate the paper.", "strengths": "Strength\n1) The paper is well-motivated;\n2) Results with LLM  are very interesting;\n3) A direct comparison of any algorithms with children of age four can be made. \n\nWeaknesses\n1) My main concern is that I did not understand how the proposed benchmark test for overhypotheses. I do not understand how one can understand that a considered algorithm or a child learned causal overhyposes. Can you elaborate on it? The authors wrote “The reward function should capture whether the algorithm has learned the causal overhypothesis of an environment”, but  I do see only that the proposed reward can help distinguish whether the agent learns to discover (or generalize to discover) a causal graph (or blinkets) during the test/eval stage. \n2) Following the 1. The authors wrote: “We find that although most agents have no problem learning causal structures seen during training, they are unable to learn causal overhypotheses from these experiences, and thus cannot generalize to new settings.” I do see that algorithms did not generalize to the new settings, but I do not understand why it is due to overhypotheses. In some sense, I feel the opposite. For example, in the case when models are trained only on disjunctive data, the model learns from the data that conjunctive case is not possible. So the model learns overhyposes prior perfectly. Another story is that the model encoded this prior to the weights of the model and can not recover from it and generalize to conjunctive out-of-distribution data in the test stage.\n3) I doubt that any reliable conclusion can be made based on such a small hold-out test set. I believe the experimental results rather motivate father research in this direction than allow to make reliable conclusions.  (the authors noted it themself). Maybe the easiest workaround is to consider more than three objects and more than two blinkets to train and test algorithms;\n4) I do not understand how the author trains RL algorithms. For example, the author wrote: “We train all of these algorithms in the given hypothesis scenario, where the agent is exposed during training to all possible overhypotheses”.  I do not understand what this means. \"Given hypothesis\" is a case where an agent receives demonstration data, but how this offline data is used to train RL agent? Is it used in offline RL settings? Do the authors train the agent on-policy then?\n\n\nSmall issues and questions\n1) If the maximum achievable reward is always 3, I believe the authors give a reward for correctly identifying the identity of the object (is it a blicket or not). The author's wording “they receive a reward of 1 for identifying a correct blicket” confuses me to think that the maximum reward is 1 for the disjunctive setting and 2 for the conjunctive setting.\n2) What is the distribution of the tasks when you train RL agent?\n3) In the situation when you hold out disjunctive overhypothesis, does it mean that the training set does not contain examples when object 3 (without loss of generality) is a blanket? \n4)  If the agent allows checking all the options for the blinkeness at the same time, is it one of the working strategies for the ideal algorithm to check all combinations in one round (it is only 8 for 3 objects) and then conclude what objects are blinkets?\n\n\nI also believe the paper is strongly connected to meta-learning in RL, in which the agent needs first explore an environment to understand what task it should solve and then solve this task (Rakelly et al, Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables, ICML2019).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper presents a causal learning benchmark to test the learning of causal overhypotheses. The authors test several RL algorithms as well as pretrained language models on this benchmark. ", "strength_and_weaknesses": "Strength\n1) The paper is well-motivated;\n2) Results with LLM  are very interesting;\n3) A direct comparison of any algorithms with children of age four can be made. \n\nWeaknesses\n1) My main concern is that I did not understand how the proposed benchmark test for overhypotheses. I do not understand how one can understand that a considered algorithm or a child learned causal overhyposes. Can you elaborate on it? The authors wrote “The reward function should capture whether the algorithm has learned the causal overhypothesis of an environment”, but  I do see only that the proposed reward can help distinguish whether the agent learns to discover (or generalize to discover) a causal graph (or blinkets) during the test/eval stage. \n2) Following the 1. The authors wrote: “We find that although most agents have no problem learning causal structures seen during training, they are unable to learn causal overhypotheses from these experiences, and thus cannot generalize to new settings.” I do see that algorithms did not generalize to the new settings, but I do not understand why it is due to overhypotheses. In some sense, I feel the opposite. For example, in the case when models are trained only on disjunctive data, the model learns from the data that conjunctive case is not possible. So the model learns overhyposes prior perfectly. Another story is that the model encoded this prior to the weights of the model and can not recover from it and generalize to conjunctive out-of-distribution data in the test stage.\n3) I doubt that any reliable conclusion can be made based on such a small hold-out test set. I believe the experimental results rather motivate father research in this direction than allow to make reliable conclusions.  (the authors noted it themself). Maybe the easiest workaround is to consider more than three objects and more than two blinkets to train and test algorithms;\n4) I do not understand how the author trains RL algorithms. For example, the author wrote: “We train all of these algorithms in the given hypothesis scenario, where the agent is exposed during training to all possible overhypotheses”.  I do not understand what this means. \"Given hypothesis\" is a case where an agent receives demonstration data, but how this offline data is used to train RL agent? Is it used in offline RL settings? Do the authors train the agent on-policy then?\n\n\nSmall issues and questions\n1) If the maximum achievable reward is always 3, I believe the authors give a reward for correctly identifying the identity of the object (is it a blicket or not). The author's wording “they receive a reward of 1 for identifying a correct blicket” confuses me to think that the maximum reward is 1 for the disjunctive setting and 2 for the conjunctive setting.\n2) What is the distribution of the tasks when you train RL agent?\n3) In the situation when you hold out disjunctive overhypothesis, does it mean that the training set does not contain examples when object 3 (without loss of generality) is a blanket? \n4)  If the agent allows checking all the options for the blinkeness at the same time, is it one of the working strategies for the ideal algorithm to check all combinations in one round (it is only 8 for 3 objects) and then conclude what objects are blinkets?\n\n\nI also believe the paper is strongly connected to meta-learning in RL, in which the agent needs first explore an environment to understand what task it should solve and then solve this task (Rakelly et al, Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables, ICML2019).", "clarity,_quality,_novelty_and_reproducibility": "I do not believe that the experiments are reproducible based on the current description in the paper. A lot of important moments (see weaknesses) are unclear. The main idea, learning priors for overhyposes to guide exploration, seems novel to me. ", "summary_of_the_review": "I think the main idea of the paper is novel and attractive. However, I need a lot of clarifications to properly evaluate the paper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666799759603}], "openreview_url": "https://openreview.net/forum?id=bGC7Ai125lR", "arxiv_id": "2206.08353", "paper_pdf": "papers/bGC7Ai125lR.pdf", "paper_pdf_sha256": "e564b76c97a2323eb83270d0c74b3e2cd43424fbd9fc2ba42e97645c5567fa44", "paper_pdf_bytes": 1102829, "paper_pdf_source": "openreview", "code_url": "https://github.com/CannyLab/causal_overhypotheses", "code_repository": "CannyLab/causal_overhypotheses", "code_commit": "d9fe308e33512bd35dd5633b42806c69bd944930", "code_archive": "repos/bGC7Ai125lR.zip", "code_archive_sha256": "803c58868c749d3d483c37bebba43aad17f677535b6cab1e23dc73a6ec16025a", "code_archive_bytes": 244956, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 239, "github_languages": {"Python": 99641}, "github_archived": false, "github_pushed_at": "2022-08-16T19:13:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-understanding-how-machines-can-learn"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RW_GTtTfHJ6", "year": 2022, "status": "rejected", "title": "Causal Reinforcement Learning using Observational and Interventional Data", "authors": ["Maxime Gasse", "Damien GRASSET", "Guillaume Gaudron", "Pierre-Yves Oudeyer"], "authorids": ["~Maxime_Gasse2", "~Damien_GRASSET1", "~Guillaume_Gaudron1", "~Pierre-Yves_Oudeyer1"], "authors_source": "OpenReview API", "abstract": "Learning efficiently a causal model of the environment is a key challenge of model-based RL agents operating in POMDPs. We consider here a scenario where the learning agent has the ability to collect online experiences through direct interactions with the environment (interventional data), but also has access to a large collection of offline experiences, obtained by observing another agent interacting with the environment (observational data). A key ingredient, which makes this situation non-trivial, is that we allow the observed agent to act based on privileged information, hidden from the learning agent. We then ask the following questions: can the online and offline experiences be safely combined for learning a causal transition model ? And can we expect the offline experiences to improve the agent's performances ? To answer these, first we bridge the fields of reinforcement learning and causality, by importing ideas from the well-established causal framework of do-calculus, and expressing model-based reinforcement learning as a causal inference problem. Second, we propose a general yet simple methodology for safely leveraging offline data during learning. In a nutshell, our method relies on learning a latent-based causal transition model that explains both the interventional and observational regimes, and then inferring the standard POMDP transition model via deconfounding using the recovered latent variable. We prove our method is correct and efficient in the sense that it attains better generalization guarantees due to the offline data (in the asymptotic case), and we assess its effectiveness empirically on a series of synthetic toy problems.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "7m8brEeRpmC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3973/Reviewer_Xuy7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors study the POMDP problem from the causal perspective, and the propose to combine offline and online data to infer the transition model via deconfounding. On the theoretical side, they show that the proposed method is correct and efficient in terms of generalization guarantees. On the experimental side, they evaluate the proposed method on three synthetic toy problems. ", "review_text": "My main concerns are on the experimental side. \n\n- The results on the three very low-dimensional synthetic toy problems are quite limited. It is hard to judge the validity of the proposed method. As we know, the RL problems would become exponentially difficult as the state dimension increases. Also, in the current RL community most RL algorithms take image pixels as input. Without such experiments, it is unclear how the proposed method work in the real world scenarios. \n\n- Comparing with several baselines and SOTA methods is an important way to demonstrate the superiority of the proposed approach. Unfortunately, such a comparison is missing in the paper. I suggest the authors should add some, which would make the paper more convincing. E.g., Rezende et al. (2020), Kallus et al. (2018), Zhang&Bareinboim (2020), etc.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors study the POMDP problem from the causal perspective, and the propose to combine offline and online data to infer the transition model via deconfounding. On the theoretical side, they show that the proposed method is correct and efficient in terms of generalization guarantees. On the experimental side, they evaluate the proposed method on three synthetic toy problems. ", "main_review": "My main concerns are on the experimental side. \n\n- The results on the three very low-dimensional synthetic toy problems are quite limited. It is hard to judge the validity of the proposed method. As we know, the RL problems would become exponentially difficult as the state dimension increases. Also, in the current RL community most RL algorithms take image pixels as input. Without such experiments, it is unclear how the proposed method work in the real world scenarios. \n\n- Comparing with several baselines and SOTA methods is an important way to demonstrate the superiority of the proposed approach. Unfortunately, such a comparison is missing in the paper. I suggest the authors should add some, which would make the paper more convincing. E.g., Rezende et al. (2020), Kallus et al. (2018), Zhang&Bareinboim (2020), etc.", "summary_of_the_review": "The experimental results are quite limited so that they are not enough to support the claims in the paper. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635929941869}, {"id": "hOwjBrRE4uC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3973/Reviewer_cYgx"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper considers the model-based reinforcement learning (RL) problem by combining the offline and online data. The online data, i.e., the interventional data, is generated from the standard partially-observable Markov decision process (POMDP), while the offline data, i.e., the observational data, is generated from the privileged POMDP where the offline learner had the access to the state information (i.e., the unobserved confounder) to make an action. The authors proposed an augmented learning procedure to safely combine these two separate different sources and learn a more efficient policy. Their method is shown both theoretically and empirically better than not using offline data. My main concerns lie in their framework and assumptions, novelty compared with existing literature, and comparison studies.", "review_text": "**Strengths**\n\n1. This paper addressed an interesting and important question in RL, i.e., how to use the observational data to improve the performance of online learning.\n\n2. The authors claimed that their setting is non-trivial by considering the unobserved confounders in the observational data.\n\n3. Their method was shown to be valid and promising both theoretically and empirically.\n\n**Weaknesses**\n\n1. Since one major contribution claimed in this paper is to bridge the causal inference with reinforcement learning, I was expecting that the authors could use a more rigorous causal framework and necessary assumptions to ensure the validation of their method and theory. For instance, to replace the do-operator with the conditional probabilities, one should assume ignorability or exogeneity. Please refer to [1] below and add related assumptions. \n\n[1] Pearl, Judea. \"Models, reasoning and inference.\" Cambridge, UK: Cambridge university press 19 (2000).\n\n2. I am not very convinced why online data particularly follows POMDP and offline data follows (possibly) privileged POMDP. Does a pre-testing procedure is required to justify the model assumptions?\n\n3. There are at least two directions of literature that the authors should pay attention to and justify their novelty. \n\n(a). First, a number of works have proposed to combine observational and experimental data (though not for RL), such as [2], [3], etc. The authors may justify why they use the augmentation procedure and will this procedure achieve the usually desired doubly robust property?\n\n[2] Athey, Susan, Raj Chetty, and Guido Imbens. \"Combining experimental and observational data to estimate treatment effects on long term outcomes.\" arXiv preprint arXiv:2006.09676 (2020).\n\n[3] Cooper, Gregory F., and Changwon Yoo. \"Causal discovery from a mixture of experimental and observational data.\" arXiv preprint arXiv:1301.6686 (2013). \n\n(b). There are increasing works regarding combining offline and online data in RL, while it seems that the authors only discussed partial of them. See some works below.\n\n[4] Nair, Ashvin, et al. \"Accelerating online reinforcement learning with offline datasets.\" arXiv preprint arXiv:2006.09359 (2020).\n\n[5] Gelly, Sylvain, and David Silver. \"Combining online and offline knowledge in UCT.\" Proceedings of the 24th international conference on Machine learning. 2007.\n\n4. I don't agree with the statement by the authors that 'Although we would have loved to compare against those approaches, the lack of available code did prevent us from running a fair comparison.' Actually, by searching these cited papers' titles with 'GitHub', I did find their implementations (as follows). Thus, the authors should add the comparison studies to justify their better performance.\n\n[6] Nathan Kallus, Aahlad Manas Puli, Uri Shalit. Removing Hidden Confounding by Experimental Grounding. NIPS 2018.\n\nhttps://github.com/CausalML/RemovingHiddenConfounding\n\n[7] Elias Bareinboim, Andrew Forney, and Judea Pearl. Bandits with unobserved confounders: A causal approach. In NIPS, 2015.\n\nhttps://github.com/nanavatirutu/CausalBandits\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper considers the model-based reinforcement learning (RL) problem by combining the offline and online data. The online data, i.e., the interventional data, is generated from the standard partially-observable Markov decision process (POMDP), while the offline data, i.e., the observational data, is generated from the privileged POMDP where the offline learner had the access to the state information (i.e., the unobserved confounder) to make an action. The authors proposed an augmented learning procedure to safely combine these two separate different sources and learn a more efficient policy. Their method is shown both theoretically and empirically better than not using offline data. My main concerns lie in their framework and assumptions, novelty compared with existing literature, and comparison studies.", "main_review": "**Strengths**\n\n1. This paper addressed an interesting and important question in RL, i.e., how to use the observational data to improve the performance of online learning.\n\n2. The authors claimed that their setting is non-trivial by considering the unobserved confounders in the observational data.\n\n3. Their method was shown to be valid and promising both theoretically and empirically.\n\n**Weaknesses**\n\n1. Since one major contribution claimed in this paper is to bridge the causal inference with reinforcement learning, I was expecting that the authors could use a more rigorous causal framework and necessary assumptions to ensure the validation of their method and theory. For instance, to replace the do-operator with the conditional probabilities, one should assume ignorability or exogeneity. Please refer to [1] below and add related assumptions. \n\n[1] Pearl, Judea. \"Models, reasoning and inference.\" Cambridge, UK: Cambridge university press 19 (2000).\n\n2. I am not very convinced why online data particularly follows POMDP and offline data follows (possibly) privileged POMDP. Does a pre-testing procedure is required to justify the model assumptions?\n\n3. There are at least two directions of literature that the authors should pay attention to and justify their novelty. \n\n(a). First, a number of works have proposed to combine observational and experimental data (though not for RL), such as [2], [3], etc. The authors may justify why they use the augmentation procedure and will this procedure achieve the usually desired doubly robust property?\n\n[2] Athey, Susan, Raj Chetty, and Guido Imbens. \"Combining experimental and observational data to estimate treatment effects on long term outcomes.\" arXiv preprint arXiv:2006.09676 (2020).\n\n[3] Cooper, Gregory F., and Changwon Yoo. \"Causal discovery from a mixture of experimental and observational data.\" arXiv preprint arXiv:1301.6686 (2013). \n\n(b). There are increasing works regarding combining offline and online data in RL, while it seems that the authors only discussed partial of them. See some works below.\n\n[4] Nair, Ashvin, et al. \"Accelerating online reinforcement learning with offline datasets.\" arXiv preprint arXiv:2006.09359 (2020).\n\n[5] Gelly, Sylvain, and David Silver. \"Combining online and offline knowledge in UCT.\" Proceedings of the 24th international conference on Machine learning. 2007.\n\n4. I don't agree with the statement by the authors that 'Although we would have loved to compare against those approaches, the lack of available code did prevent us from running a fair comparison.' Actually, by searching these cited papers' titles with 'GitHub', I did find their implementations (as follows). Thus, the authors should add the comparison studies to justify their better performance.\n\n[6] Nathan Kallus, Aahlad Manas Puli, Uri Shalit. Removing Hidden Confounding by Experimental Grounding. NIPS 2018.\n\nhttps://github.com/CausalML/RemovingHiddenConfounding\n\n[7] Elias Bareinboim, Andrew Forney, and Judea Pearl. Bandits with unobserved confounders: A causal approach. In NIPS, 2015.\n\nhttps://github.com/nanavatirutu/CausalBandits\n\n", "summary_of_the_review": "I think this is a borderline paper that addressed an important question with reasonably good performance while lacking necessary elaboration and justification. As commented in my 'Main Review', my major concerns to recommend this paper lie in their framework, novelty compared with existing literature, and comparison studies. I am willing to upgrade if my concerns can be addressed during the rebuttal period.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635898219624}, {"id": "Y411ehR8GME", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3973/Reviewer_NLhN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problems of evaluating interventional distributions (i.e., system dynamics) of a partially observed Markov decision process (POMDP) from samples collected from a combination of randomized experiments and observations of a privileged expert who could access the latent state. The POMDP is presumed to have a finite horizon, e.g., the physician could only perform a finite number of treatments for the same patients. The authors propose an unbiased estimator for evaluating system dynamics from the experimental data. As for the observational distribution where the unobserved confounding exists, the authors derive bounds over unknown system dynamics, estimable from observations.", "review_text": "This paper studies the evaluation of interventional distributions in a canonical POMDP model with a finite horizon. The target query is P(o_{t+1} | do(a_{0:t}),o_{0:t}) where A_0, ... A_t represent actions from stage t=0, ... t; O_0, ... O_t+1 represents partially observed states from stage t = 0, ..., t+1. This learning setting is general since it could represent most of treatment regimens in medical domains.\n\nThe author first shows that given data collected from interventions, the interventional query P(o_{t+1} | do(a_{0:t}),o_{0:t}) could be consistently estimated using the conditional distribution P(o_{t+1} | a_{0:t},o_{0:t}; I = i) where I represent the intervention policy that generates the data. This is not surprising since the sequential backdoor criterion is entailed in the interventional data. The authors then derive a bound over the product of target query \\prod_{t = 0}^{T-1} P(o_{t+1} | do(a_{0:t}),o_{0:t}) from the observational distribution. The result appears interesting at first, but seems to be a simple application of Manski's bound in (Manski, 1989). \n\nThe authors validate their results through comprehensive simulations. Results show that estimation using both the observational and interventional data consistently outperforms other learning strategies. However, it is unclear how the combination is done. That is, it would be interesting to see how the authors combine the unbiased estimator from the interventional data with the bound derived from the observational data. Unfortunately, this detail is not elaborated in the main manuscript.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the problems of evaluating interventional distributions (i.e., system dynamics) of a partially observed Markov decision process (POMDP) from samples collected from a combination of randomized experiments and observations of a privileged expert who could access the latent state. The POMDP is presumed to have a finite horizon, e.g., the physician could only perform a finite number of treatments for the same patients. The authors propose an unbiased estimator for evaluating system dynamics from the experimental data. As for the observational distribution where the unobserved confounding exists, the authors derive bounds over unknown system dynamics, estimable from observations.", "main_review": "This paper studies the evaluation of interventional distributions in a canonical POMDP model with a finite horizon. The target query is P(o_{t+1} | do(a_{0:t}),o_{0:t}) where A_0, ... A_t represent actions from stage t=0, ... t; O_0, ... O_t+1 represents partially observed states from stage t = 0, ..., t+1. This learning setting is general since it could represent most of treatment regimens in medical domains.\n\nThe author first shows that given data collected from interventions, the interventional query P(o_{t+1} | do(a_{0:t}),o_{0:t}) could be consistently estimated using the conditional distribution P(o_{t+1} | a_{0:t},o_{0:t}; I = i) where I represent the intervention policy that generates the data. This is not surprising since the sequential backdoor criterion is entailed in the interventional data. The authors then derive a bound over the product of target query \\prod_{t = 0}^{T-1} P(o_{t+1} | do(a_{0:t}),o_{0:t}) from the observational distribution. The result appears interesting at first, but seems to be a simple application of Manski's bound in (Manski, 1989). \n\nThe authors validate their results through comprehensive simulations. Results show that estimation using both the observational and interventional data consistently outperforms other learning strategies. However, it is unclear how the combination is done. That is, it would be interesting to see how the authors combine the unbiased estimator from the interventional data with the bound derived from the observational data. Unfortunately, this detail is not elaborated in the main manuscript.", "summary_of_the_review": "Overall, the authors study an exciting topic causal identification in POMDP, which is a quite general, and challenging learning setting. My main concern with this paper is its novelty. First, the unbiased estimator in Eq. (4) is not surprising and follows immediately from the backdoor criterion. I am pretty sure many similar MLE estimators have been proposed. Second, the bound in Theorem 1 might be interesting, but appears to be a simple application from the bound in (Manski, 1989). It would be encouraged if the authors could elaborate how to combine these different methods to obtain a more accurate estimation of the target interventional distribution.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635889993948}, {"id": "GUhWfo4bOcO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3973/Reviewer_8zjw"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of learning a causal model in the POMDP setting. It assumes the learning agent has the ability to collect online experiences through direct interactions with the environment and can access a large collection of offline experiences obtained through the observation of another agent. It further assumes that the observed agent can act based on privileged information hidden from the learning agent. The paper formulates model-based reinforcement learning in this setting as a causal inference problem. The paper then proposes to use offline data as a regularizer during learning. The paper presents empirical results on a number of toy problems.", "review_text": "Strengths\n\nThe paper proposes to learn a latent causal transition model explaining both the interventional and observational data and infer the standard POMDP transition model via deconfounding using the recovered latent variable. \n\nThe  paper shows that combing both intervention data and observation data can achieve better generalization guarantees in the asymptotic case. \n\nWeaknesses\nThe setting considered is rather limiting because it assumes that there is no confounder in the model of the POMDP. \n\nThe writing is very poor. The main contribution, section 4.3 is not clearly explained. It is not clear how imposing an observational distribution q(τ|i = 0) acts as a regularizer for the interventional distribution.  \n\nExperiments are only done with very simple toy problems.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of learning a causal model in the POMDP setting. It assumes the learning agent has the ability to collect online experiences through direct interactions with the environment and can access a large collection of offline experiences obtained through the observation of another agent. It further assumes that the observed agent can act based on privileged information hidden from the learning agent. The paper formulates model-based reinforcement learning in this setting as a causal inference problem. The paper then proposes to use offline data as a regularizer during learning. The paper presents empirical results on a number of toy problems.", "main_review": "Strengths\n\nThe paper proposes to learn a latent causal transition model explaining both the interventional and observational data and infer the standard POMDP transition model via deconfounding using the recovered latent variable. \n\nThe  paper shows that combing both intervention data and observation data can achieve better generalization guarantees in the asymptotic case. \n\nWeaknesses\nThe setting considered is rather limiting because it assumes that there is no confounder in the model of the POMDP. \n\nThe writing is very poor. The main contribution, section 4.3 is not clearly explained. It is not clear how imposing an observational distribution q(τ|i = 0) acts as a regularizer for the interventional distribution.  \n\nExperiments are only done with very simple toy problems.", "summary_of_the_review": "The setting considered is very limiting as it assumes there are no latent confounders. See the paper for the tricky issues involved:\nShaking the foundations: delusions in sequence models for interaction and control\nPedro A. Ortega, Markus Kunesch, Grégoire Delétang, Tim Genewein, Jordi Grau-Moya, Joel Veness, Jonas Buchli, Jonas Degrave, Bilal Piot, Julien Perolat, Tom Everitt, Corentin Tallec, Emilio Parisotto, Tom Erez, Yutian Chen, Scott Reed, Marcus Hutter, Nando de Freitas, Shane Legg\n\nThe experiments are not performed in any non-trivial settings. \n\nThe writing makes the paper hard to read. For example, it references rule R2 without specifying where it is or first introducing it.\n\n===\nThe authors have made their contributions and assumptions more clear, and will add  results comparing with related work. I am happy to upgrade my rating.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635834938988}], "openreview_url": "https://openreview.net/forum?id=RW_GTtTfHJ6", "arxiv_id": "2106.14421", "paper_pdf": "papers/RW_GTtTfHJ6.pdf", "paper_pdf_sha256": "be8e3b529972c0a9c88bd8302a1127b12ef76c1fac821617b9d2cf5c87d3e52b", "paper_pdf_bytes": 1327916, "paper_pdf_source": "openreview", "code_url": "https://github.com/causal-rl-anonymous/causal-rl", "code_repository": "causal-rl-anonymous/causal-rl", "code_commit": "59fdfee79e579505ed4c43ee627fe7715874508e", "code_archive": "repos/RW_GTtTfHJ6.zip", "code_archive_sha256": "d75e925ff21f923ead7bc3d1d364e716ededbedd89f5dddbfca8f391c35e9a4f", "code_archive_bytes": 53073, "code_file_count": 25, "code_extensions": {".py": 25}, "github_disk_usage_kb": 584, "github_languages": {"Python": 166277}, "github_archived": false, "github_pushed_at": "2022-03-28T21:38:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/causal-reinforcement-learning-using"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jQ0XleVhYuT", "year": 2021, "status": "rejected", "title": "Double Generative Adversarial Networks for Conditional Independence Testing", "authors": ["Chengchun Shi", "Tianlin Xu", "Wicher Pieter Bergsma", "Lexin Li"], "authorids": ["~Chengchun_Shi1", "~Tianlin_Xu1", "~Wicher_Pieter_Bergsma1", "~Lexin_Li1"], "authors_source": "OpenReview API", "abstract": "In this article, we consider the problem of high-dimensional conditional independence testing, which is a key building block in statistics and machine learning. We propose a double generative adversarial networks (GAN)-based inference procedure. We first introduce a double GANs framework to learn two generators, and integrate the two generators to construct a doubly-robust test statistic. We next consider multiple generalized covariance measures, and take their maximum as our test statistic. Finally, we obtain the empirical distribution of our test statistic through multiplier bootstrap. We show that our test controls type-I error, while the\npower approaches one asymptotically. More importantly, these theoretical guarantees are obtained under much weaker and practically more feasible conditions compared to existing tests. We demonstrate the efficacy of our test through both synthetic and real datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "r__JZ4zYnEZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper928/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Overview\n\nThis paper develops a test for conditional independence using GAN.\nConditional independence is one of the well-known problems, and calculating the conditional distribution of a random variable is a challenge.\nThe authors pointed out in Proposition 1 that the existing method, named GCIT, cannot avoid a non-negligible approximation bias.\nThe authors newly developed test combines the conditional distribution with GAN and regression-based and MMD-based tests to construct a valid test under weaker conditional requirements than the previous method.\n\n- Comments.\n\nThe paper points out some important issues with existing research.\nHowever, there are a few things I don't understand.\n\nThe experiments show that the GCIT appear to have no power at all, but the experiments of the original paper (Bellot and van der Schaar (2019)) report that GCIT has sufficient power.\nWhere does this discrepancy come from?\nThe submitted paper says that it uses a similar setting to Bellot and van der Schaar (2019), so I would like to know why the results are so different.\n\nThe theoretical advantages and their relationship to the experiment are less clear.\nThe paper states that the conditions required for GCIT are unsatisfied in general, but does this show up in the experimental results?\nThe Type I errors appear to be a bit more or less dominant, but not by much.\nCould that be the reason for the poor performance of GCIT in the analysis for power? \nIf so, then it should be clear why there is a significant difference with Bellot and van der Schaar (2019), as discussed above.\n\nI didn't understand why the conditions by the submitted paper are weaker.\nI don't doubt the accuracy, but if the technical points are not properly explained, it is not kind for readers.\nI want a clear explanation that can weaken the conditions.\n\nHow is the computational time of the proposed method?\nTo be a practical method, the computation time should be short, but if we use the GAN, it would be a big cost.\nOf course, this is true for all studies using GANs, not only this paper, but also for all studies using GANs, but it is important for practical purposes, so please let me know.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Seems sound, but some unclear points.", "review": "- Overview\n\nThis paper develops a test for conditional independence using GAN.\nConditional independence is one of the well-known problems, and calculating the conditional distribution of a random variable is a challenge.\nThe authors pointed out in Proposition 1 that the existing method, named GCIT, cannot avoid a non-negligible approximation bias.\nThe authors newly developed test combines the conditional distribution with GAN and regression-based and MMD-based tests to construct a valid test under weaker conditional requirements than the previous method.\n\n- Comments.\n\nThe paper points out some important issues with existing research.\nHowever, there are a few things I don't understand.\n\nThe experiments show that the GCIT appear to have no power at all, but the experiments of the original paper (Bellot and van der Schaar (2019)) report that GCIT has sufficient power.\nWhere does this discrepancy come from?\nThe submitted paper says that it uses a similar setting to Bellot and van der Schaar (2019), so I would like to know why the results are so different.\n\nThe theoretical advantages and their relationship to the experiment are less clear.\nThe paper states that the conditions required for GCIT are unsatisfied in general, but does this show up in the experimental results?\nThe Type I errors appear to be a bit more or less dominant, but not by much.\nCould that be the reason for the poor performance of GCIT in the analysis for power? \nIf so, then it should be clear why there is a significant difference with Bellot and van der Schaar (2019), as discussed above.\n\nI didn't understand why the conditions by the submitted paper are weaker.\nI don't doubt the accuracy, but if the technical points are not properly explained, it is not kind for readers.\nI want a clear explanation that can weaken the conditions.\n\nHow is the computational time of the proposed method?\nTo be a practical method, the computation time should be short, but if we use the GAN, it would be a big cost.\nOf course, this is true for all studies using GANs, not only this paper, but also for all studies using GANs, but it is important for practical purposes, so please let me know.", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603977855870}, {"id": "gcRKQSxAcqp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper928/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe paper proposed a novel simulation based testing procedure for conditional independence: X\\perp Y | Z. The testing procedure incorporate the techniques of GAN, which is especially useful for dealing with high-dimensional data. The testing procedure first learn the generative adversarial network that is able to simulate the conditional distribution of X|Z and Y|Z and check a particular kernel-based independence criteria presented in Eq.(4). Instead of kernel-based method that takes the supreme over a class of RKHS functions, the proposed testing procedure searches the maximum \"discrepancy\" over a class of neural network function by simulation. Empirical results show a well controlled type-I error and better test power performances compare to existing methods and a cancer data application is discussed.\n\nThis is an interesting paper combining kernel testing and GAN setting which can potentially broaden the scope of kernel-based tests. In addition to the contributions, there are some unclear parts in presentation and concerns on the proposed schemes.\n\n1. The kernel-based test, e.g. MMD relies on the positive definiteness of the RKHS. The notion of characteristic kernel ensures that the null hypothesis hold iff the test statistics is 0. In this work, do you or how do you ensure the positive definiteness? By taking the max over an absolute value in Eq.(5) does ensure non-negativity of the test statistics, but not yet the characteristic notion. Instead of RKHS kernel, it may end up being a Krein space kernel, i.e. RKKS. This may reduce the test power of the statistics in particular scenario, but it is unclear.\n\n2. From the flow chart in Fig.2, the two generators and discriminators are trained separately. How does the notion of \"doubly\" comes in. If the training is interactive between Y|Z and X|Z network, how is it done? This part is not entirely clear.\n\n3. Is the notion of robustness come from the learned conditional density (or generator) is robust against noisy samples or some form of adversarial attack, or some other notion? Not entirely clear just from Thereom 1, instead, it sound more like asymptotic property. \n\n4. The number of \"bootstrap\" functions B need to go to infinity, maybe it will be better to emphasize that.\n\n\n5. Building up from the GCM type of statistic, instead of MMD, it may be better to refer Hilbert Space Independence Criterion (HSIC) instead of MMD.\nAppendix D derivation, despite mathematically correct,  need more steps: the add-in and subtract h_1(X) times conditional expectation of h_2(Y)|Z term, otherwise it does not supply additional information to the main text.\n\n6. Experiment findings: in top panel of Fig.1, there is a spike on dim=150 for DL-CIT, is there any intuition why this happens? And KCIT and CCIT papers both claimed their controlled type-I error, what is the reason for this scenario that they fail?\n\nThanks for the presentation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel approach combine GAN technique with Conditional Independence Testing; still room to improve", "review": "\nThe paper proposed a novel simulation based testing procedure for conditional independence: X\\perp Y | Z. The testing procedure incorporate the techniques of GAN, which is especially useful for dealing with high-dimensional data. The testing procedure first learn the generative adversarial network that is able to simulate the conditional distribution of X|Z and Y|Z and check a particular kernel-based independence criteria presented in Eq.(4). Instead of kernel-based method that takes the supreme over a class of RKHS functions, the proposed testing procedure searches the maximum \"discrepancy\" over a class of neural network function by simulation. Empirical results show a well controlled type-I error and better test power performances compare to existing methods and a cancer data application is discussed.\n\nThis is an interesting paper combining kernel testing and GAN setting which can potentially broaden the scope of kernel-based tests. In addition to the contributions, there are some unclear parts in presentation and concerns on the proposed schemes.\n\n1. The kernel-based test, e.g. MMD relies on the positive definiteness of the RKHS. The notion of characteristic kernel ensures that the null hypothesis hold iff the test statistics is 0. In this work, do you or how do you ensure the positive definiteness? By taking the max over an absolute value in Eq.(5) does ensure non-negativity of the test statistics, but not yet the characteristic notion. Instead of RKHS kernel, it may end up being a Krein space kernel, i.e. RKKS. This may reduce the test power of the statistics in particular scenario, but it is unclear.\n\n2. From the flow chart in Fig.2, the two generators and discriminators are trained separately. How does the notion of \"doubly\" comes in. If the training is interactive between Y|Z and X|Z network, how is it done? This part is not entirely clear.\n\n3. Is the notion of robustness come from the learned conditional density (or generator) is robust against noisy samples or some form of adversarial attack, or some other notion? Not entirely clear just from Thereom 1, instead, it sound more like asymptotic property. \n\n4. The number of \"bootstrap\" functions B need to go to infinity, maybe it will be better to emphasize that.\n\n\n5. Building up from the GCM type of statistic, instead of MMD, it may be better to refer Hilbert Space Independence Criterion (HSIC) instead of MMD.\nAppendix D derivation, despite mathematically correct,  need more steps: the add-in and subtract h_1(X) times conditional expectation of h_2(Y)|Z term, otherwise it does not supply additional information to the main text.\n\n6. Experiment findings: in top panel of Fig.1, there is a spike on dim=150 for DL-CIT, is there any intuition why this happens? And KCIT and CCIT papers both claimed their controlled type-I error, what is the reason for this scenario that they fail?\n\nThanks for the presentation.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603922292632}, {"id": "ACAqZkstpCr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper928/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of conditional independence testing, especially when the variables are high-dimensional. The authors proposed a double GAN based algorithm. Two GANs are designed to learn the conditional probability distributions P_{X|Z} and P_{Y|Z}, then used to generate samples to compute the test statistic. It is proved that the error of the test statistic is O_p(n^{-2k} \\log n) when the total variation error of the GANS is O(n^{-k}). So to ensure that the test statistic converges, only o(\\log^{-1/2} n) rate is required for the total variation error of the GANs.\n\nThe result of this paper is quite strong. Compared to the paper of Bellot & van der Schaar (2019)  which requires the TV error of GAN to be o(n^{-1/2}), this paper significantly reduce the requirement to o(\\log^{-1/2} n).  I did not check the full proof.\n\nQuestions to the authors: In the description of Bellot & van der Schaar(2019), it seems that the data used for training the GAN and the data used for computing testing statistic are shared. However, in the proposed algorithm (Algorithm 1), the data are split into L blocks where GANs are trained by L-1 blocks of data and test statistics are computed by the other. Is this a critical reason for the improvement of convergence rate? If so, how will the convergence rate be if we apply data splitting to Bellot & van der Schaar (2019)? If not, what is the key reason for the improvement of convergence rate, double-GAN or the randomly generated h functions? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "This paper considers the problem of conditional independence testing, especially when the variables are high-dimensional. The authors proposed a double GAN based algorithm. Two GANs are designed to learn the conditional probability distributions P_{X|Z} and P_{Y|Z}, then used to generate samples to compute the test statistic. It is proved that the error of the test statistic is O_p(n^{-2k} \\log n) when the total variation error of the GANS is O(n^{-k}). So to ensure that the test statistic converges, only o(\\log^{-1/2} n) rate is required for the total variation error of the GANs.\n\nThe result of this paper is quite strong. Compared to the paper of Bellot & van der Schaar (2019)  which requires the TV error of GAN to be o(n^{-1/2}), this paper significantly reduce the requirement to o(\\log^{-1/2} n).  I did not check the full proof.\n\nQuestions to the authors: In the description of Bellot & van der Schaar(2019), it seems that the data used for training the GAN and the data used for computing testing statistic are shared. However, in the proposed algorithm (Algorithm 1), the data are split into L blocks where GANs are trained by L-1 blocks of data and test statistics are computed by the other. Is this a critical reason for the improvement of convergence rate? If so, how will the convergence rate be if we apply data splitting to Bellot & van der Schaar (2019)? If not, what is the key reason for the improvement of convergence rate, double-GAN or the randomly generated h functions? ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603712909321}, {"id": "fjbRTyURCE0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper928/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a non-parametric conditional independence test that approximates distances in a Hilbert space of functions using a generative approach, both evaluating conditional expectations using samples from GANs and evaluating a supremum over a set of functions previously generated at random. The test incorporates benefits from different lines of research and is demonstrated to outperform alternatives in well-known benchmarks. I think this study is interesting and serves as a good example of the synergies that may be achieved by combining powerful function approximators and strong statistical arguments.\n\nMy biggest concern is that conditional independence testing is an unsupervised problem, there is therefore little scope to test for the goodness of fit of hyperparameter choices or the accuracy of approximations. The proposed approach has plenty of user-defined parameters yet the sensitivity of performance to different choices is not investigated nor is there a discussion of sensible values to be recommended in practice. \n\nSome specific questions on this thread are as follows. \n- Appendix A says: \"The performance of GANs is largely affected by the regularization parameter and the number of Sinkhorn iterations R.\", should we expect performance to vary a lot in practice? \n- Appendix C tests the goodness of fit of the GAN approximation by comparing observed and estimated distributions p(y|x). Why should this be a good indication that GANs approximate well p(y|z) or p(x|z)?\n\nMost consistency guarantees are asymptotic in nature. With increasing conditioning set size, the fact that complex conditional distributions need to be approximated and that separate sets of samples are needed to train and compute the test statistic, I worry that a very large number of samples will be needed to achieve good performance. Would experiments as a function of sample size be possible to include in the paper?\n\n\nMinor comments:\n- The test itself is quite involved, with many moving parts that are described over various pages in the paper. I would recommend to have a higher-level summary of the approach perhaps using Figure 2 in the Appendix earlier in the paper.\n- A single data generating mechanism is used for performance comparisons. I think different multiple different choices here would be needed to test different aspects of the model.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Sophisticated conditional independence test with well-studied theoretical guarantees.", "review": "The authors propose a non-parametric conditional independence test that approximates distances in a Hilbert space of functions using a generative approach, both evaluating conditional expectations using samples from GANs and evaluating a supremum over a set of functions previously generated at random. The test incorporates benefits from different lines of research and is demonstrated to outperform alternatives in well-known benchmarks. I think this study is interesting and serves as a good example of the synergies that may be achieved by combining powerful function approximators and strong statistical arguments.\n\nMy biggest concern is that conditional independence testing is an unsupervised problem, there is therefore little scope to test for the goodness of fit of hyperparameter choices or the accuracy of approximations. The proposed approach has plenty of user-defined parameters yet the sensitivity of performance to different choices is not investigated nor is there a discussion of sensible values to be recommended in practice. \n\nSome specific questions on this thread are as follows. \n- Appendix A says: \"The performance of GANs is largely affected by the regularization parameter and the number of Sinkhorn iterations R.\", should we expect performance to vary a lot in practice? \n- Appendix C tests the goodness of fit of the GAN approximation by comparing observed and estimated distributions p(y|x). Why should this be a good indication that GANs approximate well p(y|z) or p(x|z)?\n\nMost consistency guarantees are asymptotic in nature. With increasing conditioning set size, the fact that complex conditional distributions need to be approximated and that separate sets of samples are needed to train and compute the test statistic, I worry that a very large number of samples will be needed to achieve good performance. Would experiments as a function of sample size be possible to include in the paper?\n\n\nMinor comments:\n- The test itself is quite involved, with many moving parts that are described over various pages in the paper. I would recommend to have a higher-level summary of the approach perhaps using Figure 2 in the Appendix earlier in the paper.\n- A single data generating mechanism is used for performance comparisons. I think different multiple different choices here would be needed to test different aspects of the model.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603119263790}], "openreview_url": "https://openreview.net/forum?id=jQ0XleVhYuT", "arxiv_id": "2006.02615", "paper_pdf": "papers/jQ0XleVhYuT.pdf", "paper_pdf_sha256": "0bf1bddb6969e0ef2175aff837aa1142c5ddc33cb81dd960c8b7f135d25e04d7", "paper_pdf_bytes": 3445043, "paper_pdf_source": "openreview", "code_url": "https://github.com/tianlinxu312/dgcit", "code_repository": "tianlinxu312/dgcit", "code_commit": "686ea5ee151d947d42e5836866f8961420090080", "code_archive": "repos/jQ0XleVhYuT.zip", "code_archive_sha256": "8a11fb19b1fc8dfd193baf059a6449569123e94e53887ffa3f5d104cf4abff60", "code_archive_bytes": 863307, "code_file_count": 6, "code_extensions": {".py": 6}, "github_disk_usage_kb": 864, "github_languages": {"Python": 88906}, "github_archived": false, "github_pushed_at": "2023-11-22T18:37:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/double-generative-adversarial-networks-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJlVn6NKPB", "year": 2020, "status": "rejected", "title": "Representation Learning for Remote Sensing: An Unsupervised Sensor Fusion Approach", "authors": ["Aidan M. Swope", "Xander H. Rudelis", "Kyle T. Story"], "authorids": ["aidanswope@gmail.com", "xander@descarteslabs.com", "kyle@descarteslabs.com"], "authors_source": "OpenReview API", "abstract": "In the application of machine learning to remote sensing, labeled data is often scarce or expensive, which impedes the training of powerful models like deep convolutional neural networks. Although unlabeled data is abundant, recent self-supervised learning approaches are ill-suited to the remote sensing domain. In addition, most remote sensing applications currently use only a small subset of the multi-sensor, multi-channel information available, motivating the need for fused multi-sensor representations. We propose a new self-supervised training objective, Contrastive Sensor Fusion, which exploits coterminous data from multiple sources to learn useful representations of every possible combination of those sources. This method uses information common across multiple sensors and bands by training a single model to produce a representation that remains similar when any subset of its input channels is used. Using a dataset of 47 million unlabeled coterminous image triplets, we train an encoder to produce semantically meaningful representations from any possible combination of channels from the input sensors. These representations outperform fully supervised ImageNet weights on a remote sensing classification task and improve as more sensors are fused.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SylM88mHoB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper778/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an approach to create unsupervised representations of remote sensing images. The essential idea is to enforce similarity between representations of multiple views obtained by subsetting channels from multiple co-terminus sensor outputs.  This is implemented by training with the InfoNCE loss on high-level features (last two layers of a ResNet 50) of two views of the same image and multiple view of other images which are passed through the same weight-shared network. The evaluation is based on a custom task of classifying OSM labels. The results show that (a) the learned representation (up to 3 chosen channels) outperform a pre-trained ImageNet fine-tuned for the classification task, (b) multiple low resolution sensors can outperform a higher resolution sensor, and (c) visualisation of images indicates that stable high-level representations across sensors. \n\nThe paper makes a worthwhile contribution in introducing contrastive methods to satellite imagery - a domain well suited for analysing how contrastive methods work and also rich in applications of sensor fusion/augmentation for remote sensing applications. The experiments indicate that this is a promising direction, and the authors have helpfully open-sourced the dataset they used. The paper, however, is short on several accounts. Primarily, the experimental evaluation is sparse in detail and rigour. In addition, the paper makes unsubstantiated claims to “argue out” certain baselines from being compared. Finally, the results indicate modest improvements on a custom task, and thus remain inconclusive. \n\nThe paper bases all experimental evaluation and conclusions thereof on the custom task of classifying 12 classes of distinctive objects on 8400 images obtained from OpenStreetMap (OSM). These include diverse categories spanning generic bodies (water, forest), specific structures (substation, bridge), similar items (farm, farmland). The paper does not discuss why this is a good task (is it challenging, is it representative of remote sensing applications, are these most frequent OSM classes). Indeed, the experimental results provide no intuition on the classification task (eg. confusion matrix is essential especially given the middling accuracy of about 0.5). \n\nThe paper makes a strong claim that the “distributional hypothesis” (of expecting that spatial or temporal similarity reflect in representation similarity) does not apply to satellite imagery because “remote sensing imagery … can change abruptly between adjacent patches”. No evidence is provided of what I believe is an unintuitive claim. Indeed, existing work [1] (as cited in the paper) uses this hypothesis to create representations for satellite images. Glaringly, having made the claim, the paper does not compare proposed approach with the representations computed in [1].  The paper also makes the claim that computing representation loss should be done with high-level features rather than individual pixels. Again, this is not substantiated as the paper does not compare the proposed methods with simple auto-encoder baselines. In the absence of these baselines, the results of this paper on a custom data-set remain inconclusive. \n\nThe evaluation leaves out several other expected experiments. A few suggested conditions for ablation tests are listed:\n(a) Different augmentation across channels\n(b) Different values of \\lambda_L (only an extreme case of last two layers has been presented)\n(c) At least one other CNN-backbone, perhaps a deeper ResNet\n(d) Different orderings of introducing the channels (not only by low or high resolution ordering)\n(e) Different sizes of the tiles for learning representations (curiously the paper does not mention the size of the tiles as used currently)\n\nFinally, and more broadly, the proposed approach aims to compensate the lack of supervised labels by exploiting redundancy across multiple sensors/channels. In the chosen setup, this redundancy is very high as 4 different RGBI co-terminus sensors are chosen, thereby requiring augmenting and drop-out to simulate some variation. A more realistic or challenging setup would be required to evaluate the underlying ideas.\n\n[1] Jean, Neal, et al. \"Tile2Vec: Unsupervised representation learning for spatially distributed data.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\n\n--\n\nUpdate in response to rebuttal: \nThe authors agreed to most of the points raised, but provided no clear suggestions in addressing them. I re-emphasise that the empirical results on a single custom dataset based on OSM is limited. The author response to this remains vague. Further none of the reasonable baselines have been compared against, a point which the authors ignored in the rebuttal. In light of this, rating remains the same. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #4", "review": "The paper presents an approach to create unsupervised representations of remote sensing images. The essential idea is to enforce similarity between representations of multiple views obtained by subsetting channels from multiple co-terminus sensor outputs.  This is implemented by training with the InfoNCE loss on high-level features (last two layers of a ResNet 50) of two views of the same image and multiple view of other images which are passed through the same weight-shared network. The evaluation is based on a custom task of classifying OSM labels. The results show that (a) the learned representation (up to 3 chosen channels) outperform a pre-trained ImageNet fine-tuned for the classification task, (b) multiple low resolution sensors can outperform a higher resolution sensor, and (c) visualisation of images indicates that stable high-level representations across sensors. \n\nThe paper makes a worthwhile contribution in introducing contrastive methods to satellite imagery - a domain well suited for analysing how contrastive methods work and also rich in applications of sensor fusion/augmentation for remote sensing applications. The experiments indicate that this is a promising direction, and the authors have helpfully open-sourced the dataset they used. The paper, however, is short on several accounts. Primarily, the experimental evaluation is sparse in detail and rigour. In addition, the paper makes unsubstantiated claims to “argue out” certain baselines from being compared. Finally, the results indicate modest improvements on a custom task, and thus remain inconclusive. \n\nThe paper bases all experimental evaluation and conclusions thereof on the custom task of classifying 12 classes of distinctive objects on 8400 images obtained from OpenStreetMap (OSM). These include diverse categories spanning generic bodies (water, forest), specific structures (substation, bridge), similar items (farm, farmland). The paper does not discuss why this is a good task (is it challenging, is it representative of remote sensing applications, are these most frequent OSM classes). Indeed, the experimental results provide no intuition on the classification task (eg. confusion matrix is essential especially given the middling accuracy of about 0.5). \n\nThe paper makes a strong claim that the “distributional hypothesis” (of expecting that spatial or temporal similarity reflect in representation similarity) does not apply to satellite imagery because “remote sensing imagery … can change abruptly between adjacent patches”. No evidence is provided of what I believe is an unintuitive claim. Indeed, existing work [1] (as cited in the paper) uses this hypothesis to create representations for satellite images. Glaringly, having made the claim, the paper does not compare proposed approach with the representations computed in [1].  The paper also makes the claim that computing representation loss should be done with high-level features rather than individual pixels. Again, this is not substantiated as the paper does not compare the proposed methods with simple auto-encoder baselines. In the absence of these baselines, the results of this paper on a custom data-set remain inconclusive. \n\nThe evaluation leaves out several other expected experiments. A few suggested conditions for ablation tests are listed:\n(a) Different augmentation across channels\n(b) Different values of \\lambda_L (only an extreme case of last two layers has been presented)\n(c) At least one other CNN-backbone, perhaps a deeper ResNet\n(d) Different orderings of introducing the channels (not only by low or high resolution ordering)\n(e) Different sizes of the tiles for learning representations (curiously the paper does not mention the size of the tiles as used currently)\n\nFinally, and more broadly, the proposed approach aims to compensate the lack of supervised labels by exploiting redundancy across multiple sensors/channels. In the chosen setup, this redundancy is very high as 4 different RGBI co-terminus sensors are chosen, thereby requiring augmenting and drop-out to simulate some variation. A more realistic or challenging setup would be required to evaluate the underlying ideas.\n\n[1] Jean, Neal, et al. \"Tile2Vec: Unsupervised representation learning for spatially distributed data.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\n\n--\n\nUpdate in response to rebuttal: \nThe authors agreed to most of the points raised, but provided no clear suggestions in addressing them. I re-emphasise that the empirical results on a single custom dataset based on OSM is limited. The author response to this remains vague. Further none of the reasonable baselines have been compared against, a point which the authors ignored in the rebuttal. In light of this, rating remains the same. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573365322297}, {"id": "B1xNtSz1qr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper778/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an approach for representation learning on remote sensing data/satellite imagery inspired by recent unsupervised contrastive multiview representation learning methods (CPC, DIM, CMC). The method relies on the InfoNCE objective to contrast two different views of the data obtained by randomly cropping image patches, color jittering, channel dropping etc. The proposed method is compared to an ImageNet pretrained network in a classification task based on OpenStreetMap (OSM) features and show to outperform the ImageNet pretrained classifier. \n\nThe paper is well-written and the method is clearly explained. To my knowledge there is little prior work on unsupervised/self-supervised learning on remote sensing/satellite data and this data suits the contrastive/multiview framework very well. In that regard I appreciate the direction explored by the paper.\n\nHere are some questions and concerns:\n\n- I think the paper is lacking some important details, for example how are the unsupervised and ImageNet-pretrained representations transferred? Fine-tuning, or learning a classifier on top of the frozen representation? In the case of fine-tuning, the ImageNet baseline could be extended to more channels.\n\n- As an additional baseline, how does a network trained from scratch on the available labeled training data perform?\n\n- The authors evaluate on a single data set, that seems to not have been used previously. To make the evaluation more solid it would be good to compare on other data sets, for example on EuroSAT [1], and with other, possibly supervised classification methods, see, e.g., [1, 2].\n\n- Did you do ablations on augmentations used? For example, is zeroing out channels more effective than copying other channels instead?\n\n- Both forward and backward prediction losses are used, but it seems that the loss is symmetric. Does adding the backward prediction loss really help?\n\nOverall, I like the direction the paper is exploring, but I think it would greatly benefit from adding detail on the outlined aspects and extending the evaluation.\n\n\n[1] Helber P, Bischke B, Dengel A, Borth D. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 2019 Jun 14;12(7):2217-26.\n[2] Kussul N, Lavreniuk M, Skakun S, Shelestov A. Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters. 2017 Mar 31;14(5):778-82.\n\n\n\n---\nUpdate after the rebuttal:\nThanks to the authors for their detailed response. While the authors agree that some points should be improved, no attempts were made to actually strengthen the method in a revision. In particular, I do think evaluating the proposed method on prior benchmarks that were mostly used in the context of supervised methods (such as EuroSAT) is very important. Also, I still think comparison to from scratch training is crucial. Both comparisons are common practice in prior work such as CPC, DIM, and CMC (specifically, ImageNet). I therefore do not increase my rating.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "The paper proposes an approach for representation learning on remote sensing data/satellite imagery inspired by recent unsupervised contrastive multiview representation learning methods (CPC, DIM, CMC). The method relies on the InfoNCE objective to contrast two different views of the data obtained by randomly cropping image patches, color jittering, channel dropping etc. The proposed method is compared to an ImageNet pretrained network in a classification task based on OpenStreetMap (OSM) features and show to outperform the ImageNet pretrained classifier. \n\nThe paper is well-written and the method is clearly explained. To my knowledge there is little prior work on unsupervised/self-supervised learning on remote sensing/satellite data and this data suits the contrastive/multiview framework very well. In that regard I appreciate the direction explored by the paper.\n\nHere are some questions and concerns:\n\n- I think the paper is lacking some important details, for example how are the unsupervised and ImageNet-pretrained representations transferred? Fine-tuning, or learning a classifier on top of the frozen representation? In the case of fine-tuning, the ImageNet baseline could be extended to more channels.\n\n- As an additional baseline, how does a network trained from scratch on the available labeled training data perform?\n\n- The authors evaluate on a single data set, that seems to not have been used previously. To make the evaluation more solid it would be good to compare on other data sets, for example on EuroSAT [1], and with other, possibly supervised classification methods, see, e.g., [1, 2].\n\n- Did you do ablations on augmentations used? For example, is zeroing out channels more effective than copying other channels instead?\n\n- Both forward and backward prediction losses are used, but it seems that the loss is symmetric. Does adding the backward prediction loss really help?\n\nOverall, I like the direction the paper is exploring, but I think it would greatly benefit from adding detail on the outlined aspects and extending the evaluation.\n\n\n[1] Helber P, Bischke B, Dengel A, Borth D. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 2019 Jun 14;12(7):2217-26.\n[2] Kussul N, Lavreniuk M, Skakun S, Shelestov A. Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters. 2017 Mar 31;14(5):778-82.\n\n\n\n---\nUpdate after the rebuttal:\nThanks to the authors for their detailed response. While the authors agree that some points should be improved, no attempts were made to actually strengthen the method in a revision. In particular, I do think evaluating the proposed method on prior benchmarks that were mostly used in the context of supervised methods (such as EuroSAT) is very important. Also, I still think comparison to from scratch training is crucial. Both comparisons are common practice in prior work such as CPC, DIM, and CMC (specifically, ImageNet). I therefore do not increase my rating.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571919228282}, {"id": "H1e3XLhVtr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper778/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an unsupervised method to embed (remotely sensed) image patches such that they are clustered by high-level \"image content\" rather than by low-level pixel statistics. The argument is that in remote sensing one often has access only to a subset of all possible images/channels, so it would be useful to get the same embedding for each subset depicting the same scene. The basic idea is to generate, during training, two versions of the same patch by randomly dropping out channels, and to penalise the difference between the two corresponding embeddings, using the InfoNCE loss. The latter ensures that the patch remains \"recognisable\" among a set of random non-matching patches, so as to avoid learning a trivial embedding that is the same for any input.\n\nThe high-level goal, to train an encoder that always gives a similar encoding (and consequently similar downstream results) for a given patch, independent of the sensor channels used, is certainly appealing. It is also true that one would like to do this with unsupervised learning or transfer learning, since labelled data is scarce in remote sensing. Unfortunately, in my understanding the paper does not make much progress towards that goal.\n\nNovelty is limited, the work is a variation of the CMC method. Yes, it splits channels randomly, not by views; and uses a Siamese architecture, not separate encoders; which it can afford to do since there are no viewpoint differences. And it computes the loss at more than one layer - which is a fairly trivial and popular thing to do. Moreover it is not well justified in this case, since all that matters in a feed-forward architecture is that the final embeddings are similar. Constraining lower layers might empirically help - in particular speed up learning - but conceptually there is no reason to reduce the lower layers' capacity to somehow turn different inputs into the same output. If they are really very different (think RADAR images) this could in principle even hurt. \n\nBoth in the discussion and in the experiments, I miss crucial and obvious baselines. First, it does in my view not make sense to use pre-trained ImageNet features as the only baseline. Yes, they are known to work surprisingly well also for remote sensing data - but, unsurprisingly, not nearly as well as pre-training on similar remote sensing images. A more natural baseline would be to pre-train on the same type of imagery, with a proxy task for which it is easy to get labels automatically  - actually the paper implicitly suggests that labels from OpenStreetMap might be a good candidate. Beyond reducing the domain gap, this also has the advantage that one is not limited to 3 channels, and that one can use an architecture suitable for the task (e.g., standard architecture with a lot of pooling are known to not perform well for semantic segmentation of remote sensing data). Labels for supervised learning might in many remote sensing tasks be scarce, but to get any old labels and pre-train a reasonable proxy task is still easy, as map data is abundant.\n\nPerhaps even more important is another natural baseline. In spirit, the proposed method is closer to a generative embedding, since the supervision signal is not a discriminative task, but rather encourages embeddings to stay unique for each patch, and hence in principle fully decodable. Arguably, the most obvious way to achieve the same thing is a Siamese pair of auto-encoders for the two training patches, with an additional loss on the difference between the latent representations. Without that baseline it is impossible to tell whether the InfoNCE loss does a good job - in some sense it feels like making the task harder if one demands not only that a patch of water can be decoded from the embedding, but that it should be distinguishable from any other patch of water that resembles it even in raw pixel space.\n\nThe data used in the experiments is somewhat unconvincing, as it is really the opposite of diverse, multi-modal remote sensing channels. The same four R-G-B-NIR channels from sensors with not so different resolutions. The claim that the method homes in on high-level similarity and overcomes low-level radiometric differences would be much more credible if one uses short-wave infrared, thermal, perhaps even Radar images. Moreover, the gains are disappointingly small: even in the best case of 12 channels, the improvement over ImageNet pre-training is roughly from 4.5/10 correct neighbours to 5/10 correct neighbours. The even lower gains with fewer channels underline the need to have a pre-trained baseline without a massive domain gap. The classification accuracy of kNN does improve with the proposed embedding (at least for some combinations - for Pleiades RGB it is the same), but that experiment is unnatural. If one has the labels to do kNN, one might as well use them to train a proper classifier, i.e., add a couple of fully connected layers and fine-tune in a supervised fashion.\n\nA number of technical details in the experiments remain unclear. Section 4.2 suggests the channels in Fig. 4 were added in a fixed order - why? There are many potentially viable combinations of, say, 4 channels, not just one. Furthermore, Fig. 5 suggests to me that the learned similarities are in fact not \"high-level\", but (up to rotation invariance) fairly low-level texture properties like \"homogeneous green patch\", \"bright straight line on dark background\", etc. I am not sure what suggests to the authors that there is much sensor invariance. Again, channels with really different wavelength would be more convincing here.\n\nIn terms of presentation, the paper repeatedly makes strong, but unsupported claims about the special properties of remote sensing images.  It makes no sense to state that standard pictures often have \"one subject\" - ImageNet does, but a look at standard datasets like Cityscapes or DepthInTheWild shows that it is not true in general. In the same vein, I don't see why the content changes more unpredictably at the same location across time - arguably the rate of change is much lower, because only large-scale changes impact the image, and because there aren't many dynamic occluders. And the \"distributional hypothesis\" certainly also holds for remote sensing images, just at different scale - otherwise people would not use all sorts of smoothness priors, super-pixel segmentations, etc. when working with them. And I do not understand why the adjacent-patch approach of CPC needs nearby patches to be similar - in my view it only requires that nearby patches are not conditionally independent, so basically any type of patterns.\n\nA small detail: some references, while not wrong, seem rather arbitrary. E.g., (Wu 2018) is a little-known random example. Why not use earlier, more standard references - for semantic segmentation of remote sensing data with deep learning one could for example think of (Maggiori, Audebert, Marmanis, Kampffmeyer, Sherrah, ...).\n\nOverall, while the paper brings up a valid question, it does not give a convincing answer. The justification of the method and its novelty is to some degree contrived, the dataset is ill-suited to really prove the point, and the natural baselines are missing.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper proposes an unsupervised method to embed (remotely sensed) image patches such that they are clustered by high-level \"image content\" rather than by low-level pixel statistics. The argument is that in remote sensing one often has access only to a subset of all possible images/channels, so it would be useful to get the same embedding for each subset depicting the same scene. The basic idea is to generate, during training, two versions of the same patch by randomly dropping out channels, and to penalise the difference between the two corresponding embeddings, using the InfoNCE loss. The latter ensures that the patch remains \"recognisable\" among a set of random non-matching patches, so as to avoid learning a trivial embedding that is the same for any input.\n\nThe high-level goal, to train an encoder that always gives a similar encoding (and consequently similar downstream results) for a given patch, independent of the sensor channels used, is certainly appealing. It is also true that one would like to do this with unsupervised learning or transfer learning, since labelled data is scarce in remote sensing. Unfortunately, in my understanding the paper does not make much progress towards that goal.\n\nNovelty is limited, the work is a variation of the CMC method. Yes, it splits channels randomly, not by views; and uses a Siamese architecture, not separate encoders; which it can afford to do since there are no viewpoint differences. And it computes the loss at more than one layer - which is a fairly trivial and popular thing to do. Moreover it is not well justified in this case, since all that matters in a feed-forward architecture is that the final embeddings are similar. Constraining lower layers might empirically help - in particular speed up learning - but conceptually there is no reason to reduce the lower layers' capacity to somehow turn different inputs into the same output. If they are really very different (think RADAR images) this could in principle even hurt. \n\nBoth in the discussion and in the experiments, I miss crucial and obvious baselines. First, it does in my view not make sense to use pre-trained ImageNet features as the only baseline. Yes, they are known to work surprisingly well also for remote sensing data - but, unsurprisingly, not nearly as well as pre-training on similar remote sensing images. A more natural baseline would be to pre-train on the same type of imagery, with a proxy task for which it is easy to get labels automatically  - actually the paper implicitly suggests that labels from OpenStreetMap might be a good candidate. Beyond reducing the domain gap, this also has the advantage that one is not limited to 3 channels, and that one can use an architecture suitable for the task (e.g., standard architecture with a lot of pooling are known to not perform well for semantic segmentation of remote sensing data). Labels for supervised learning might in many remote sensing tasks be scarce, but to get any old labels and pre-train a reasonable proxy task is still easy, as map data is abundant.\n\nPerhaps even more important is another natural baseline. In spirit, the proposed method is closer to a generative embedding, since the supervision signal is not a discriminative task, but rather encourages embeddings to stay unique for each patch, and hence in principle fully decodable. Arguably, the most obvious way to achieve the same thing is a Siamese pair of auto-encoders for the two training patches, with an additional loss on the difference between the latent representations. Without that baseline it is impossible to tell whether the InfoNCE loss does a good job - in some sense it feels like making the task harder if one demands not only that a patch of water can be decoded from the embedding, but that it should be distinguishable from any other patch of water that resembles it even in raw pixel space.\n\nThe data used in the experiments is somewhat unconvincing, as it is really the opposite of diverse, multi-modal remote sensing channels. The same four R-G-B-NIR channels from sensors with not so different resolutions. The claim that the method homes in on high-level similarity and overcomes low-level radiometric differences would be much more credible if one uses short-wave infrared, thermal, perhaps even Radar images. Moreover, the gains are disappointingly small: even in the best case of 12 channels, the improvement over ImageNet pre-training is roughly from 4.5/10 correct neighbours to 5/10 correct neighbours. The even lower gains with fewer channels underline the need to have a pre-trained baseline without a massive domain gap. The classification accuracy of kNN does improve with the proposed embedding (at least for some combinations - for Pleiades RGB it is the same), but that experiment is unnatural. If one has the labels to do kNN, one might as well use them to train a proper classifier, i.e., add a couple of fully connected layers and fine-tune in a supervised fashion.\n\nA number of technical details in the experiments remain unclear. Section 4.2 suggests the channels in Fig. 4 were added in a fixed order - why? There are many potentially viable combinations of, say, 4 channels, not just one. Furthermore, Fig. 5 suggests to me that the learned similarities are in fact not \"high-level\", but (up to rotation invariance) fairly low-level texture properties like \"homogeneous green patch\", \"bright straight line on dark background\", etc. I am not sure what suggests to the authors that there is much sensor invariance. Again, channels with really different wavelength would be more convincing here.\n\nIn terms of presentation, the paper repeatedly makes strong, but unsupported claims about the special properties of remote sensing images.  It makes no sense to state that standard pictures often have \"one subject\" - ImageNet does, but a look at standard datasets like Cityscapes or DepthInTheWild shows that it is not true in general. In the same vein, I don't see why the content changes more unpredictably at the same location across time - arguably the rate of change is much lower, because only large-scale changes impact the image, and because there aren't many dynamic occluders. And the \"distributional hypothesis\" certainly also holds for remote sensing images, just at different scale - otherwise people would not use all sorts of smoothness priors, super-pixel segmentations, etc. when working with them. And I do not understand why the adjacent-patch approach of CPC needs nearby patches to be similar - in my view it only requires that nearby patches are not conditionally independent, so basically any type of patterns.\n\nA small detail: some references, while not wrong, seem rather arbitrary. E.g., (Wu 2018) is a little-known random example. Why not use earlier, more standard references - for semantic segmentation of remote sensing data with deep learning one could for example think of (Maggiori, Audebert, Marmanis, Kampffmeyer, Sherrah, ...).\n\nOverall, while the paper brings up a valid question, it does not give a convincing answer. The justification of the method and its novelty is to some degree contrived, the dataset is ill-suited to really prove the point, and the natural baselines are missing."}, "tcdate": 1571239460359}], "openreview_url": "https://openreview.net/forum?id=SJlVn6NKPB", "arxiv_id": "2108.05094", "paper_pdf": "papers/SJlVn6NKPB.pdf", "paper_pdf_sha256": "62f62632ffd3790ca38c7ac40c7ca1f9bfc379ea5236a28ef80438d6ae9b76ab", "paper_pdf_bytes": 4506565, "paper_pdf_source": "openreview", "code_url": "https://github.com/dlarchives/contrastive_sensor_fusion", "code_repository": "dlarchives/contrastive_sensor_fusion", "code_commit": "f9d02e910e4b6213e0af5ea13b2c3fa99c8c8e1b", "code_archive": "repos/SJlVn6NKPB.zip", "code_archive_sha256": "63b34ac2dde04710f4e181f7b3aaa65dd748bd6c3a5a0a49050281771ff730bf", "code_archive_bytes": 715371, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 2104, "github_languages": {"Python": 63383}, "github_archived": false, "github_pushed_at": "2023-03-24T22:38:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/representation-learning-for-remote-sensing-an"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EAUGN4pszX", "year": 2025, "status": "rejected", "title": "FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification", "authors": ["Tian Tian", "Chunyan Miao", "Hangwei Qian"], "authorids": ["~Tian_Tian7", "~Chunyan_Miao1", "~Hangwei_Qian1"], "authors_source": "OpenReview API", "abstract": "Contrastive learning has emerged as a competent approach for unsupervised representation learning. However, the design of an optimal augmentation strategy, although crucial for contrastive learning, is less explored for time series classification tasks. Existing predefined time-domain augmentation methods are primarily adopted from vision and are not specific to time series data. Consequently, this cross-modality incompatibility may distort the global semantics of time series by introducing mismatched patterns into the data. To address this limitation, we present a novel perspective from the frequency domain and identify three advantages for downstream classification: 1) the frequency component naturally encodes global features, 2) the orthogonal nature of the Fourier basis allows easier isolation and independent modifications of critical and unimportant information, and 3) a compact set of frequency components can preserve semantic integrity. To fully utilize the three properties, we propose the lightweight yet effective Frequency-Refined Augmentation (FreRA) tailored for time series contrastive learning on classification tasks, which can be seamlessly integrated with contrastive learning frameworks in a plug-and-play manner. Specifically, FreRA automatically separates critical and unimportant frequency components. Accordingly, we propose Identity Modification and Self-adaptive Modification to protect global semantics in the critical frequency components and infuse variance to the unimportant ones respectively. \nTheoretically, we prove that FreRA generates semantic-preserving views. Empirically, we conduct extensive experiments on two benchmark datasets including UCR and UEA archives, as well as 5 large-scale datasets on diverse applications. FreRA consistently outperforms 10 leading baselines on time series classification, anomaly detection, and transfer learning tasks, demonstrating superior capabilities in contrastive representation learning and generalization in transfer learning scenarios across diverse datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "YY6zOoq81l", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6561/Reviewer_cXW7"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes a method, FreRA, to enhance time series classification by contrastive learning and sample augmentations. First, they considered the frequency domain of the time series. FreRA automatically separates the critical and unimportant frequency components. They proposed Identity Modification and Self-adaptive Modification to protect the global semantics in the critical frequency components and inject variance into the unimportant components, respectively. Extensive experimental results on several datasets show that FreRA outperforms existing methods in terms of accuracy.", "review_text": "The paper proposes a method, FreRA, to enhance time series classification by contrastive learning and sample augmentations. First, they considered the frequency domain of the time series. FreRA automatically separates the critical and unimportant frequency components. They proposed Identity Modification and Self-adaptive Modification to protect the global semantics in the critical frequency components and inject variance into the unimportant components, respectively. Extensive experimental results on several datasets show that FreRA outperforms existing methods in terms of accuracy.", "strengths": "1. Overall, this paper is well-written and easy to follow.\n2. The problem studied is significant, and exploring augmentation in time series is novel.\n3. Extensive experimental results are promising.", "weaknesses": "1. The importance distinction of this method is mostly for the entire time series, and it could be better to compare it with other methods and analyze the theoretical computational complexity.\n2. Although frequency methods can improve efficiency, it is unclear whether such methods mainly focus on the low-frequency part and ignore the high-frequency part which is more important for time series prediction.\n3. Do the authors consider the dependencies between channels, which is very significant for multivariate time series.\n4. The authors claim that FreRA can be benefited by any contrastive learning framework, but only show the results of InfoNCE. What about other CL paradigms, such as SimCLR, etc.?  It could be better to present more sufficient ablation.\n5. The experimental results are selected from the highest performances among 11 time-domain augmentations and 5 frequency-domain augmentations. Is this fair enough? There seems to be randomness with such selection strategy.\n6. The results of the impacts of hyper-parameters could be moved to the main paper for a better organization.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method, FreRA, to enhance time series classification by contrastive learning and sample augmentations. First, they considered the frequency domain of the time series. FreRA automatically separates the critical and unimportant frequency components. They proposed Identity Modification and Self-adaptive Modification to protect the global semantics in the critical frequency components and inject variance into the unimportant components, respectively. Extensive experimental results on several datasets show that FreRA outperforms existing methods in terms of accuracy.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Overall, this paper is well-written and easy to follow.\n2. The problem studied is significant, and exploring augmentation in time series is novel.\n3. Extensive experimental results are promising.", "weaknesses": "1. The importance distinction of this method is mostly for the entire time series, and it could be better to compare it with other methods and analyze the theoretical computational complexity.\n2. Although frequency methods can improve efficiency, it is unclear whether such methods mainly focus on the low-frequency part and ignore the high-frequency part which is more important for time series prediction.\n3. Do the authors consider the dependencies between channels, which is very significant for multivariate time series.\n4. The authors claim that FreRA can be benefited by any contrastive learning framework, but only show the results of InfoNCE. What about other CL paradigms, such as SimCLR, etc.?  It could be better to present more sufficient ablation.\n5. The experimental results are selected from the highest performances among 11 time-domain augmentations and 5 frequency-domain augmentations. Is this fair enough? There seems to be randomness with such selection strategy.\n6. The results of the impacts of hyper-parameters could be moved to the main paper for a better organization.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731076716099}, {"id": "8Yg5swHY8M", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6561/Reviewer_Qixi"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes Frequency-Refined Augmentation (FreRA), an augmentation method for time series contrastive learning on classification tasks. FreRA automatically separates critical and unimportant frequency components, and accordingly proposes Identity Modification and Self-adaptive Modification for different components. It conducts experiments on two benchmark datasets including UCR and UEA archives, as well as 5 large-scale datasets on diverse applications. FreRA outperforms 10 leading baselines on time series classification, anomaly detection, and transfer learning tasks.", "review_text": "This paper proposes Frequency-Refined Augmentation (FreRA), an augmentation method for time series contrastive learning on classification tasks. FreRA automatically separates critical and unimportant frequency components, and accordingly proposes Identity Modification and Self-adaptive Modification for different components. It conducts experiments on two benchmark datasets including UCR and UEA archives, as well as 5 large-scale datasets on diverse applications. FreRA outperforms 10 leading baselines on time series classification, anomaly detection, and transfer learning tasks.", "strengths": "Data augmentation is an important problem for time series and contrastive learning. This paper investigates this problem and proposes a method from the frequency perspective. The proposed method seems easy to follow and implement. The experiments in this paper are extensive, including many different datasets and tasks. The proposed method outperforms most of the baselines.", "weaknesses": "W1: Some important terms in this paper are not clearly defined, such as ‘semantic integrity’ and ‘critical and unimportant frequency components’. Why can we measure semantic integrity using mutual information? Is this consistent with humans’ understanding of semantics? How can we measure the importance of frequency components? What are these critical components critical for?\n\nW2: The overall novelty of the proposed augmentation method is limited. Augmentation from the frequency components is not a new idea for time series. The main difference is a trainable vector s to control the augmentation of different components. It is unclear why s trained from Equation (7) can learn to select critical components automatically.\n\nW3: It is unclear how FreRA achieves both semantic-preserving information and a considerable amount of variance. The authors need to clarify which designs correspond to these two sides respectively. \n\nW4: The self-adaptive modification seems simple and tricky. It only uses the vector s and threshold to select and scale unimportant frequency components. The motivation for scaling these components is unclear. \n\nW5: Compared with some SOTA baselines, such as SoftCLT and InfoTS, the advantage of FreRA is not clear, especially on UEA and UCR datasets. The ablation study is coarse, and some important variants are missing. For example, modifying all (or randomly selected) frequency components and modifying unimportant components randomly.", "questions": "Some other questions:\n\nQ1: As this paper focuses on time series classification, why is it also evaluated in anomaly detection? \n\nQ2: Figure 2 is hard to read. The different colors for vector blocks in FreRA are confusing.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Frequency-Refined Augmentation (FreRA), an augmentation method for time series contrastive learning on classification tasks. FreRA automatically separates critical and unimportant frequency components, and accordingly proposes Identity Modification and Self-adaptive Modification for different components. It conducts experiments on two benchmark datasets including UCR and UEA archives, as well as 5 large-scale datasets on diverse applications. FreRA outperforms 10 leading baselines on time series classification, anomaly detection, and transfer learning tasks.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Data augmentation is an important problem for time series and contrastive learning. This paper investigates this problem and proposes a method from the frequency perspective. The proposed method seems easy to follow and implement. The experiments in this paper are extensive, including many different datasets and tasks. The proposed method outperforms most of the baselines.", "weaknesses": "W1: Some important terms in this paper are not clearly defined, such as ‘semantic integrity’ and ‘critical and unimportant frequency components’. Why can we measure semantic integrity using mutual information? Is this consistent with humans’ understanding of semantics? How can we measure the importance of frequency components? What are these critical components critical for?\n\nW2: The overall novelty of the proposed augmentation method is limited. Augmentation from the frequency components is not a new idea for time series. The main difference is a trainable vector s to control the augmentation of different components. It is unclear why s trained from Equation (7) can learn to select critical components automatically.\n\nW3: It is unclear how FreRA achieves both semantic-preserving information and a considerable amount of variance. The authors need to clarify which designs correspond to these two sides respectively. \n\nW4: The self-adaptive modification seems simple and tricky. It only uses the vector s and threshold to select and scale unimportant frequency components. The motivation for scaling these components is unclear. \n\nW5: Compared with some SOTA baselines, such as SoftCLT and InfoTS, the advantage of FreRA is not clear, especially on UEA and UCR datasets. The ablation study is coarse, and some important variants are missing. For example, modifying all (or randomly selected) frequency components and modifying unimportant components randomly.", "questions": "Some other questions:\n\nQ1: As this paper focuses on time series classification, why is it also evaluated in anomaly detection? \n\nQ2: Figure 2 is hard to read. The different colors for vector blocks in FreRA are confusing.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730882819147}, {"id": "9Ueg8mwFa9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6561/Reviewer_GiHe"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces a novel augmentation technique designed for time-series contrastive learning by leveraging frequency-domain properties. It utilized the idea of FFT which separates time-series data into critical and non-critical frequency components.", "review_text": "The paper introduces a novel augmentation technique designed for time-series contrastive learning by leveraging frequency-domain properties. It utilized the idea of FFT which separates time-series data into critical and non-critical frequency components.", "strengths": "The method utilizes the connection between frequency domain knowledge and semantic information to enhance the representation learning. The critical components capture global semantics essential for classification, while non-critical components are used for self-adaptive noise injection, which I found is an interesting link and the authors provide a comprehensive explanation for the motivation.\nThe authors provide extensive experiments and strong experiment results to demonstrate their method's effectiveness.", "weaknesses": "1. At the beginning of the paper, the authors make strong assumptions that existing predefined augmentation methods are primarily adopted from vision and are not specific to time series data. There are already several methods, especially frequency-based augmentation, e.g., TF-C, method design for time series contrastive learning.\n2. Since the paper mainly provides the frequency-based augmentation, the motivation study, such as Figure.1 probably should highlight more about whether current frequency-based method can capture the semantics, rather than only focus on the time-domain,", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel augmentation technique designed for time-series contrastive learning by leveraging frequency-domain properties. It utilized the idea of FFT which separates time-series data into critical and non-critical frequency components.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The method utilizes the connection between frequency domain knowledge and semantic information to enhance the representation learning. The critical components capture global semantics essential for classification, while non-critical components are used for self-adaptive noise injection, which I found is an interesting link and the authors provide a comprehensive explanation for the motivation.\nThe authors provide extensive experiments and strong experiment results to demonstrate their method's effectiveness.", "weaknesses": "1. At the beginning of the paper, the authors make strong assumptions that existing predefined augmentation methods are primarily adopted from vision and are not specific to time series data. There are already several methods, especially frequency-based augmentation, e.g., TF-C, method design for time series contrastive learning.\n2. Since the paper mainly provides the frequency-based augmentation, the motivation study, such as Figure.1 probably should highlight more about whether current frequency-based method can capture the semantics, rather than only focus on the time-domain,", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730653045713}, {"id": "9OnJe3ep5X", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6561/Reviewer_7F2V"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This study introduces Frequency-Refined Augmentation (FreRA), a method designed to overcome limitations in current augmentation strategies for time series classification in contrastive learning. Unlike existing visual-based augmentations, FreRA leverages three key advantages of the frequency domain properties to better preserve the global semantics of time series data. FreRA automatically segregates these components, applying Identity Modification to preserve vital details and Self-adaptive Modification to add variance to less significant parts. Theoretical proofs and empirical evaluations confirm FreRA's superiority, showing it outperforms ten leading baselines across various time series tasks.", "review_text": "This study introduces Frequency-Refined Augmentation (FreRA), a method designed to overcome limitations in current augmentation strategies for time series classification in contrastive learning. Unlike existing visual-based augmentations, FreRA leverages three key advantages of the frequency domain properties to better preserve the global semantics of time series data. FreRA automatically segregates these components, applying Identity Modification to preserve vital details and Self-adaptive Modification to add variance to less significant parts. Theoretical proofs and empirical evaluations confirm FreRA's superiority, showing it outperforms ten leading baselines across various time series tasks.", "strengths": "1. This paper is well-written and well-organized.\n2. The methodology is clear and the problem is well-motivated.\n3. This is a plug-and-play method that appears to integrate seamlessly with existing contrastive learning frameworks.", "weaknesses": "1. Compared to existing works, this work does not exhibit a notable advantage in classification performance.\n2. I am somewhat confused about the experimental design for the transfer learning part: 1) Why was SHAR data selected instead of one of the datasets listed in Table 1 (e.g., UCIHAR) to evaluate the transfer capability of the algorithm? 2) Based on the experimental results in Table 2, the performance is lower than that reported in the reference work (Qian et al., 2022). What might be the reasons for this difference?\nWhat did this experiment aim to prove?\n3. In the ABLATION STUDIES part, the authors sequentially removed each of the three innovative method components for comparison. From the experimental results, the gains provided by the three modules seem roughly equivalent. If all three modules were removed, what would be the resulting performance? Looking at Table 1 in the paper, the performance of softCLT and FreRA appear quite similar. If FreRA were integrated into softCLT, would there be any gain in performance?", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces Frequency-Refined Augmentation (FreRA), a method designed to overcome limitations in current augmentation strategies for time series classification in contrastive learning. Unlike existing visual-based augmentations, FreRA leverages three key advantages of the frequency domain properties to better preserve the global semantics of time series data. FreRA automatically segregates these components, applying Identity Modification to preserve vital details and Self-adaptive Modification to add variance to less significant parts. Theoretical proofs and empirical evaluations confirm FreRA's superiority, showing it outperforms ten leading baselines across various time series tasks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper is well-written and well-organized.\n2. The methodology is clear and the problem is well-motivated.\n3. This is a plug-and-play method that appears to integrate seamlessly with existing contrastive learning frameworks.", "weaknesses": "1. Compared to existing works, this work does not exhibit a notable advantage in classification performance.\n2. I am somewhat confused about the experimental design for the transfer learning part: 1) Why was SHAR data selected instead of one of the datasets listed in Table 1 (e.g., UCIHAR) to evaluate the transfer capability of the algorithm? 2) Based on the experimental results in Table 2, the performance is lower than that reported in the reference work (Qian et al., 2022). What might be the reasons for this difference?\nWhat did this experiment aim to prove?\n3. In the ABLATION STUDIES part, the authors sequentially removed each of the three innovative method components for comparison. From the experimental results, the gains provided by the three modules seem roughly equivalent. If all three modules were removed, what would be the resulting performance? Looking at Table 1 in the paper, the performance of softCLT and FreRA appear quite similar. If FreRA were integrated into softCLT, would there be any gain in performance?", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730103167229}], "openreview_url": "https://openreview.net/forum?id=EAUGN4pszX", "arxiv_id": "2505.23181", "paper_pdf": "papers/EAUGN4pszX.pdf", "paper_pdf_sha256": "8c8865a5390dc35b02fa4deadf5c3c6415d3bbb33a26954edeb8339e203540c8", "paper_pdf_bytes": 997286, "paper_pdf_source": "openreview", "code_url": "https://github.com/Tian0426/FreRA", "code_repository": "Tian0426/FreRA", "code_commit": "7236fbfc1c665f83ed5f4364cad59093ee283c14", "code_archive": "repos/EAUGN4pszX.zip", "code_archive_sha256": "2f8479ab1e690c17c4acc7aa846eb7dcec31f0739d0d6fd57f31227aa3771fb1", "code_archive_bytes": 158195, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 169, "github_languages": {"Python": 131451}, "github_archived": false, "github_pushed_at": "2025-06-20T05:23:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/frera-a-frequency-refined-augmentation-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xMxHJxp192", "year": 2025, "status": "rejected", "title": "DeltaGNN: Graph Neural Network with Information Flow Control", "authors": ["Kevin Mancini", "Islem Rekik"], "authorids": ["~Kevin_Mancini1", "~Islem_Rekik1"], "authors_source": "OpenReview API", "abstract": "Graph Neural Networks (GNNs) are popular machine learning models designed to process graph-structured data through recursive neighborhood aggregations in the message passing process. When applied to semi-supervised node classification, the message-passing enables GNNs to understand short-range spatial interactions, but also causes them to suffer from over-smoothing and over-squashing. These challenges hinder model expressiveness and prevent the use of deeper models to capture long-range node interactions (LRIs) within the graph. Popular solutions for LRIs detection are either too expensive to process large graphs due to high time complexity or fail to generalize across diverse graph structures. To address these limitations, we propose a mechanism called information flow control, which leverages a novel connectivity measure, called information flow score, to address over-smoothing and over-squashing with linear computational overhead, supported by theoretical evidence. Finally, to prove the efficacy of our methodology we design DeltaGNN, the first scalable and generalizable approach for long-range and short-range interaction detection. \nWe benchmark our model across 10 real-world datasets, including graphs with varying sizes, topologies, densities, and homophilic ratios, showing superior performance with limited computational complexity.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "btVq29ElvG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3270/Reviewer_ffF7"], "rating": 5, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper proposes DeltaGNN which considers the long-range node interaction via information flow control or semi-supervised node classification.\nThe key idea is to take use of first delta embeddings $\\Delta_u^t$ and the variance of second delta embeddings $\\mathbb{V}_t[\\Delta_u^2]$.\nIf the node is connected with the same labels, then $\\Delta_u^t$ tends to be some, since the features from neighbors are close to center embeddings.\nIf the node works as a bottleneck, then the aggregated features in each layer will have a huge difference, which might cause the big variance of $(\\Delta^2)_u^t$, denoted as $\\mathbb{V}_t[\\Delta_u^2]$.\nBased on this, information flow score is used to measure the nodes that is responsible for over-smoothing and over-squashing.\nThen, with graph filtering on edges, the graph cuts edges to increase the homophily for short-range interaction, and connect the selected components for long-range interactions.\nCombined with these two, the author proposes the DeltaGNN.", "review_text": "This paper proposes DeltaGNN which considers the long-range node interaction via information flow control or semi-supervised node classification.\nThe key idea is to take use of first delta embeddings $\\Delta_u^t$ and the variance of second delta embeddings $\\mathbb{V}_t[\\Delta_u^2]$.\nIf the node is connected with the same labels, then $\\Delta_u^t$ tends to be some, since the features from neighbors are close to center embeddings.\nIf the node works as a bottleneck, then the aggregated features in each layer will have a huge difference, which might cause the big variance of $(\\Delta^2)_u^t$, denoted as $\\mathbb{V}_t[\\Delta_u^2]$.\nBased on this, information flow score is used to measure the nodes that is responsible for over-smoothing and over-squashing.\nThen, with graph filtering on edges, the graph cuts edges to increase the homophily for short-range interaction, and connect the selected components for long-range interactions.\nCombined with these two, the author proposes the DeltaGNN.", "strengths": "1. This paper propose an interesting idea to connect first and second delta embedding with over-smoothing and over-squashing. \n\tThe proposed two lemma demonstrate the relationship between them. \n\tSuch relationship provides insight for developing algorithm to measure and alleviate over-smoothing and over-squashing problems considering the node embeddings.\n2. The proposed metric inforation flow score can numerically find the nodes that might cause over-smoothing and over-squashing in this graphs.", "weaknesses": "1. While it is good to have a numerical metric to identity the key nodes and edges for over-smoothing and over-squashing.\n\tThe connection of heterophilic graphs in DeltaGNN to these two metric is not that strong. \n\tFirst, it seems like heterophilic graphs is not related to solve the over-smoothing or over-squashing problem.\n\tSecond, if using informative flow control can perfectly solve the over-smoothing or over-squashing problem, \n\twhy model can not get perfect results?\n\tIn other words, why heterophilic graph is needed in this case? \n\tDoes the introduction of heterophilic graph will cause further questions about non-existing interactions?\n\tThis part needs to be further justified. The motivation and experiments of the reasons to use this part need to be provided.\n\n2. As a suggestion, some numerical experiments can be provided to demonstrate that with the information flow control, \n\tthe new homophilic graph can have less over-smoothing or graph bottleneck issues on real-world datasets.\n\tFor example, the homophilic ratio of a node can be calculated and compared between the original graphs and the rewired graphs.\n\n3. The experiments is a little weak, and more and larger graph datasets should be included like ogb datasets.", "questions": "1. What is the formulation of equation of $\\Theta_t\\left(\\mathbf{A}^{t-1}, K(t, \\theta)\\right.$, Score $\\left.^t\\right)$ that is used to filter the graph?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes DeltaGNN which considers the long-range node interaction via information flow control or semi-supervised node classification.\nThe key idea is to take use of first delta embeddings $\\Delta_u^t$ and the variance of second delta embeddings $\\mathbb{V}_t[\\Delta_u^2]$.\nIf the node is connected with the same labels, then $\\Delta_u^t$ tends to be some, since the features from neighbors are close to center embeddings.\nIf the node works as a bottleneck, then the aggregated features in each layer will have a huge difference, which might cause the big variance of $(\\Delta^2)_u^t$, denoted as $\\mathbb{V}_t[\\Delta_u^2]$.\nBased on this, information flow score is used to measure the nodes that is responsible for over-smoothing and over-squashing.\nThen, with graph filtering on edges, the graph cuts edges to increase the homophily for short-range interaction, and connect the selected components for long-range interactions.\nCombined with these two, the author proposes the DeltaGNN.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. This paper propose an interesting idea to connect first and second delta embedding with over-smoothing and over-squashing. \n\tThe proposed two lemma demonstrate the relationship between them. \n\tSuch relationship provides insight for developing algorithm to measure and alleviate over-smoothing and over-squashing problems considering the node embeddings.\n2. The proposed metric inforation flow score can numerically find the nodes that might cause over-smoothing and over-squashing in this graphs.", "weaknesses": "1. While it is good to have a numerical metric to identity the key nodes and edges for over-smoothing and over-squashing.\n\tThe connection of heterophilic graphs in DeltaGNN to these two metric is not that strong. \n\tFirst, it seems like heterophilic graphs is not related to solve the over-smoothing or over-squashing problem.\n\tSecond, if using informative flow control can perfectly solve the over-smoothing or over-squashing problem, \n\twhy model can not get perfect results?\n\tIn other words, why heterophilic graph is needed in this case? \n\tDoes the introduction of heterophilic graph will cause further questions about non-existing interactions?\n\tThis part needs to be further justified. The motivation and experiments of the reasons to use this part need to be provided.\n\n2. As a suggestion, some numerical experiments can be provided to demonstrate that with the information flow control, \n\tthe new homophilic graph can have less over-smoothing or graph bottleneck issues on real-world datasets.\n\tFor example, the homophilic ratio of a node can be calculated and compared between the original graphs and the rewired graphs.\n\n3. The experiments is a little weak, and more and larger graph datasets should be included like ogb datasets.", "questions": "1. What is the formulation of equation of $\\Theta_t\\left(\\mathbf{A}^{t-1}, K(t, \\theta)\\right.$, Score $\\left.^t\\right)$ that is used to filter the graph?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730788488096}, {"id": "xFaHAeQeAh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3270/Reviewer_aCWG"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This work targets the prevalent issues of over-smoothing and over-squashing in GNNs. It highlights that current approaches often face challenges such as high computational complexity and lack of generalizability. To tackle these issues, the authors introduce a mechanism termed 'information flow control', which employs an innovative metric known as the 'information flow score'. This mechanism is designed to mitigate over-smoothing and over-squashing while maintaining linear computational overhead. Empirical evaluations demonstrate its superior performance under constrained computational conditions.", "review_text": "This work targets the prevalent issues of over-smoothing and over-squashing in GNNs. It highlights that current approaches often face challenges such as high computational complexity and lack of generalizability. To tackle these issues, the authors introduce a mechanism termed 'information flow control', which employs an innovative metric known as the 'information flow score'. This mechanism is designed to mitigate over-smoothing and over-squashing while maintaining linear computational overhead. Empirical evaluations demonstrate its superior performance under constrained computational conditions.", "strengths": "1. The suggestion in Lemma 1—that identifying nodes connected by heterophilic edges through measuring feature differences during message aggregation— appears to be constructive.\n2. The introduction effectively outlines the problems of over-smoothing and over-squashing, and provides a comprehensive overview of existing methods aimed at resolving these challenges.\n3. The proposed method for addressing the problems of over-smoothing and over-squashing is both innovative and promising. The approach involves decoupling the original graph into a homophilic subgraph and a heterophilic subgraph using the proposed information flow score. Subsequently, the method performs dual aggregation on these subgraphs to capture both short-term and long-term dependencies.\n4. The complexity of method information flow score is superior to those of other rewiring methods.\n5. The proposed method demonstrates strong performance in terms of prediction accuracy and scalability.", "weaknesses": "1. \"$\\triangle^t_u$ can be interpreted as the velocity at which the node embeddings are aggregated at layer t.\" The concept of aggregation velocity is somehow confusing. More background knowledge and explanation, also examples, are required to help readers to understand the  measurement of aggregation velocity.\n\n2. The authors propose using $(\\triangle^2)^t_u = d(\\triangle^t_u - \\triangle^{t-1}_u)$ to measure the rate of change in the rate at which node embeddings are aggregated. However, $\\triangle^t_u$ and $\\triangle^{t-1}_u$ are outputs from different layers and thus belong to different spaces. Therefore, the rationale for measuring the distance between points in these two spaces is questionable. Please provide justification for this measurement.\n\n3. The Information Flow Control (IFC) mechanism is a core component of the proposed method. Therefore, the implementation details of the IFC mechanism, including the score hill ascent framework, should be included in the main text rather than in the appendix. As currently presented, the score hill ascent framework is difficult to follow.\n\n4. In Figure 2, some subgraphs are difficult to interpret. For example, the 'feature density - feature value' plot and the 'score - node' plot could benefit from additional clarification or improved labeling. What do the different curves in the feature density - feature value plot represent? Additionally, the phrase 'and enhance the graph score' lacks clarity. A definition of the term 'graph score' would be helpful.\n\n5. The proof of Lemma 1 is difficult to follow. Specific issues are detailed in the following list. **Additional background information and explanation are needed to help readers understand the proof**.\n   - In line 727, the term 'valid' is used to ensure that the assignment respects the given homophily ratio $\\mathcal{H}_u$. However, the concept of 'valid' is not clearly defined, and it is unclear how this term ensures compliance with the specified homophily ratio. Additional background information and explanation are needed to help readers understand these aspects.\n   - The relationship between $\\triangle^t_u$ and the valid assignment $s$ is not explained.\n   - In the equation $U(\\mathcal{H}_u)_u = \\operatorname{max}_{s\\in S}(\\triangle^t_u)$, the representation of $U$ is unclear.\n   - Due to the lack of clarity, it is not possible to understand why 'any node $u$ with $\\triangle^t_u > p$ will have $\\mathcal{H}_u < \\mathcal{H}$.'\n\n1. The phrase 'as this quantity depends on the homophily of the node $u$', in line 727,  requires clarification. It is not immediately apparent why this quantity should depend on the homophily of the node. A clear explanation is needed to elucidate this dependency.\n\n1. Mirror issues: a) in line 723, should \"neighbourhood $N(u)$\" be revised to \"neighbourhood $\\mathcal{N}(u)$\" to consist to notation of neighborhood? b) $\\bigoplus\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_u$ should be revised to $\\bigoplus\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_v$.", "questions": "1. In Table 1, it is evident that the Information Flow Score (IFS) method underperforms other rewiring methods when combined with the GIN model, unlike with other models. This discrepancy may be due to the fact that GIN uses sum aggregation, whereas other models typically use weighted mean aggregation. The sum aggregation in GIN likely results in a higher variance for $\\{\\sum\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_v \\mid u \\in \\mathcal{V}\\}$ compared to $\\{\\mathbf{M}_u \\mid u \\in \\mathcal{V}\\}.$  Consequently, the so-called 'aggregation velocity' depends not only on node features but also on node degrees. This suggests that the proposed method may not be well-suited for models that use sum aggregation. Is my understanding correct?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work targets the prevalent issues of over-smoothing and over-squashing in GNNs. It highlights that current approaches often face challenges such as high computational complexity and lack of generalizability. To tackle these issues, the authors introduce a mechanism termed 'information flow control', which employs an innovative metric known as the 'information flow score'. This mechanism is designed to mitigate over-smoothing and over-squashing while maintaining linear computational overhead. Empirical evaluations demonstrate its superior performance under constrained computational conditions.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The suggestion in Lemma 1—that identifying nodes connected by heterophilic edges through measuring feature differences during message aggregation— appears to be constructive.\n2. The introduction effectively outlines the problems of over-smoothing and over-squashing, and provides a comprehensive overview of existing methods aimed at resolving these challenges.\n3. The proposed method for addressing the problems of over-smoothing and over-squashing is both innovative and promising. The approach involves decoupling the original graph into a homophilic subgraph and a heterophilic subgraph using the proposed information flow score. Subsequently, the method performs dual aggregation on these subgraphs to capture both short-term and long-term dependencies.\n4. The complexity of method information flow score is superior to those of other rewiring methods.\n5. The proposed method demonstrates strong performance in terms of prediction accuracy and scalability.", "weaknesses": "1. \"$\\triangle^t_u$ can be interpreted as the velocity at which the node embeddings are aggregated at layer t.\" The concept of aggregation velocity is somehow confusing. More background knowledge and explanation, also examples, are required to help readers to understand the  measurement of aggregation velocity.\n\n2. The authors propose using $(\\triangle^2)^t_u = d(\\triangle^t_u - \\triangle^{t-1}_u)$ to measure the rate of change in the rate at which node embeddings are aggregated. However, $\\triangle^t_u$ and $\\triangle^{t-1}_u$ are outputs from different layers and thus belong to different spaces. Therefore, the rationale for measuring the distance between points in these two spaces is questionable. Please provide justification for this measurement.\n\n3. The Information Flow Control (IFC) mechanism is a core component of the proposed method. Therefore, the implementation details of the IFC mechanism, including the score hill ascent framework, should be included in the main text rather than in the appendix. As currently presented, the score hill ascent framework is difficult to follow.\n\n4. In Figure 2, some subgraphs are difficult to interpret. For example, the 'feature density - feature value' plot and the 'score - node' plot could benefit from additional clarification or improved labeling. What do the different curves in the feature density - feature value plot represent? Additionally, the phrase 'and enhance the graph score' lacks clarity. A definition of the term 'graph score' would be helpful.\n\n5. The proof of Lemma 1 is difficult to follow. Specific issues are detailed in the following list. **Additional background information and explanation are needed to help readers understand the proof**.\n   - In line 727, the term 'valid' is used to ensure that the assignment respects the given homophily ratio $\\mathcal{H}_u$. However, the concept of 'valid' is not clearly defined, and it is unclear how this term ensures compliance with the specified homophily ratio. Additional background information and explanation are needed to help readers understand these aspects.\n   - The relationship between $\\triangle^t_u$ and the valid assignment $s$ is not explained.\n   - In the equation $U(\\mathcal{H}_u)_u = \\operatorname{max}_{s\\in S}(\\triangle^t_u)$, the representation of $U$ is unclear.\n   - Due to the lack of clarity, it is not possible to understand why 'any node $u$ with $\\triangle^t_u > p$ will have $\\mathcal{H}_u < \\mathcal{H}$.'\n\n1. The phrase 'as this quantity depends on the homophily of the node $u$', in line 727,  requires clarification. It is not immediately apparent why this quantity should depend on the homophily of the node. A clear explanation is needed to elucidate this dependency.\n\n1. Mirror issues: a) in line 723, should \"neighbourhood $N(u)$\" be revised to \"neighbourhood $\\mathcal{N}(u)$\" to consist to notation of neighborhood? b) $\\bigoplus\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_u$ should be revised to $\\bigoplus\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_v$.", "questions": "1. In Table 1, it is evident that the Information Flow Score (IFS) method underperforms other rewiring methods when combined with the GIN model, unlike with other models. This discrepancy may be due to the fact that GIN uses sum aggregation, whereas other models typically use weighted mean aggregation. The sum aggregation in GIN likely results in a higher variance for $\\{\\sum\\limits_{v \\in \\mathcal{N}(u)}\\mathbf{M}_v \\mid u \\in \\mathcal{V}\\}$ compared to $\\{\\mathbf{M}_u \\mid u \\in \\mathcal{V}\\}.$  Consequently, the so-called 'aggregation velocity' depends not only on node features but also on node degrees. This suggests that the proposed method may not be well-suited for models that use sum aggregation. Is my understanding correct?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730704077516}, {"id": "o9ri1mLnoK", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3270/Reviewer_e676"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper identifies that Long-Range Interactions (LRIs) are crucial for node classification tasks. Standard GNNs struggle to capture these long range dependencies due to issues such as over-smoothing and over-squashing. To address these challenges, the authors propose information flow control, a graph rewiring mechanism. Further, the paper introduces DeltaGNN, which implements information flow control to capture both long- and short-range dependencies. The proposed method is validated on several graph datasets with varying levels of homophily and sizes.", "review_text": "The paper identifies that Long-Range Interactions (LRIs) are crucial for node classification tasks. Standard GNNs struggle to capture these long range dependencies due to issues such as over-smoothing and over-squashing. To address these challenges, the authors propose information flow control, a graph rewiring mechanism. Further, the paper introduces DeltaGNN, which implements information flow control to capture both long- and short-range dependencies. The proposed method is validated on several graph datasets with varying levels of homophily and sizes.", "strengths": "* The paper is well-written and easy to follow.\n\n* It introduces a novel connectivity measure, called the information flow score, which is supported by both theoretical analysis and empirical evidence.\n\n* DeltaGNN demonstrates consistent improvements across various datasets, outperforming all baseline methods compared in the study.", "weaknesses": "* DeltaGNN is proposed as a scalable approach for detecting both long-range and short-range interactions. However, there are no large-scale experiments to validate this claim, as all experiments were conducted on small graphs. It would be beneficial if the authors could report results on larger homophilic datasets, such as ogbn-arXiv, as well as on large-scale non-homophilous graphs from [1].\n\n* The related work section does not adequately situate the current research within the context of existing GNN work based on Graph Filters (e.g., [3, 4]). \n\n* Lines 361-363 indicate that DeltaGNN is compared against state-of-the-art (SoTA) GNNs. However, GCN, GAT, and GIN are not the current SoTA for the chosen benchmarks. The authors should compare DeltaGNN with more recent GNNs (e.g., ACM-GCN+ / ACMII-GCN++ from [2]) to more accurately assess its effectiveness.\n\n* It is unclear why MLP is not included as a baseline in Table 1. MLP has been shown to outperform on the three non-homophilous datasets (Texas, Wisconsin, Cornell) as reported in [4]. A comparison against graph filter-based methods, such as GPR-GNN [3] or PPGNN [4], would provide further insights into the performance of DeltaGNN.\n\n---\n[1] Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods, NeurIPS 2021\n\n[2] Revisiting Heterophily For Graph Neural Networks, NeurIPS 2022\n\n[3] Adaptive Universal Generalized PageRank Graph Neural Network, ICLR 2021\n\n[4] A Piece-Wise Polynomial Filtering Approach for Graph Neural Networks, ECML 2022", "questions": "* From Table 7 in the appendix, DeltaGNN variants consume approximately 2-3 times more GPU memory than GCN on small graphs. Could the authors discuss whether this would lead to memory issues when applied to larger graphs?\n\n* Did the authors evaluate DeltaGNN on more challenging heterophilic datasets, such as Squirrel or Chameleon [3]?\n\n*  [Minor] Typo in Line 181: \"$∆^t_u$ the first\" $\\rightarrow$ \"$∆^t_u$ be the first.\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper identifies that Long-Range Interactions (LRIs) are crucial for node classification tasks. Standard GNNs struggle to capture these long range dependencies due to issues such as over-smoothing and over-squashing. To address these challenges, the authors propose information flow control, a graph rewiring mechanism. Further, the paper introduces DeltaGNN, which implements information flow control to capture both long- and short-range dependencies. The proposed method is validated on several graph datasets with varying levels of homophily and sizes.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* The paper is well-written and easy to follow.\n\n* It introduces a novel connectivity measure, called the information flow score, which is supported by both theoretical analysis and empirical evidence.\n\n* DeltaGNN demonstrates consistent improvements across various datasets, outperforming all baseline methods compared in the study.", "weaknesses": "* DeltaGNN is proposed as a scalable approach for detecting both long-range and short-range interactions. However, there are no large-scale experiments to validate this claim, as all experiments were conducted on small graphs. It would be beneficial if the authors could report results on larger homophilic datasets, such as ogbn-arXiv, as well as on large-scale non-homophilous graphs from [1].\n\n* The related work section does not adequately situate the current research within the context of existing GNN work based on Graph Filters (e.g., [3, 4]). \n\n* Lines 361-363 indicate that DeltaGNN is compared against state-of-the-art (SoTA) GNNs. However, GCN, GAT, and GIN are not the current SoTA for the chosen benchmarks. The authors should compare DeltaGNN with more recent GNNs (e.g., ACM-GCN+ / ACMII-GCN++ from [2]) to more accurately assess its effectiveness.\n\n* It is unclear why MLP is not included as a baseline in Table 1. MLP has been shown to outperform on the three non-homophilous datasets (Texas, Wisconsin, Cornell) as reported in [4]. A comparison against graph filter-based methods, such as GPR-GNN [3] or PPGNN [4], would provide further insights into the performance of DeltaGNN.\n\n---\n[1] Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods, NeurIPS 2021\n\n[2] Revisiting Heterophily For Graph Neural Networks, NeurIPS 2022\n\n[3] Adaptive Universal Generalized PageRank Graph Neural Network, ICLR 2021\n\n[4] A Piece-Wise Polynomial Filtering Approach for Graph Neural Networks, ECML 2022", "questions": "* From Table 7 in the appendix, DeltaGNN variants consume approximately 2-3 times more GPU memory than GCN on small graphs. Could the authors discuss whether this would lead to memory issues when applied to larger graphs?\n\n* Did the authors evaluate DeltaGNN on more challenging heterophilic datasets, such as Squirrel or Chameleon [3]?\n\n*  [Minor] Typo in Line 181: \"$∆^t_u$ the first\" $\\rightarrow$ \"$∆^t_u$ be the first.\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730702597477}, {"id": "oziIOa4Z3E", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3270/Reviewer_6W5k"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper  introduces a mechanism to mitigate over-smoothing and over-squashing in Graph Neural Networks (GNNs) by implementing an \"information flow control\" strategy that utilizes an \"information flow score.\" This approach allows for effective management of node embeddings across varied graph structures, demonstrating enhanced performance in large-scale graphs while maintaining computational efficiency.", "review_text": "The paper  introduces a mechanism to mitigate over-smoothing and over-squashing in Graph Neural Networks (GNNs) by implementing an \"information flow control\" strategy that utilizes an \"information flow score.\" This approach allows for effective management of node embeddings across varied graph structures, demonstrating enhanced performance in large-scale graphs while maintaining computational efficiency.", "strengths": "originality: good\nquality: medium\nclarity: medium\nsignificance: medium", "weaknesses": "1. \"These long-range interactions (LRIs) are crucial for node classification tasks, as they help distinguish between different classes and improve classification accuracy\" This is not true. For example, graph transformers are good at capturing long-range node dependencies. However, they perform poorly on node classification tasks, especially on heterophilic graphs [1]. It is found that distant information is not always useful, and the over-globalization can cause performance degradation of graph models [2].\n2. \"over-smoothing is not only a topological phenomenon but is primarily a consequence of graph heterophily.\" There is no causal relation between over-smoothing and heterophily. As stated in [3], over-smoothing only happens in deep GNNs, but not in shallow GNNs. Heterophily will cause performance degradation to all GNN models, not matter they are deep or shallow.", "questions": "1. the information flow score, which identifies graph bottlenecks and heterophilic node interactions,\n\n2. In definition 1, the first Delta embeddings look like the \"norm\" of the high-pass filtered graph signal or the neighborhood diversification[4]. The second Delta embedding is a new and interesting one.\n\n3. So how can lemma 1 and 2 offer insights into the graph’s homophily and topology? Explain with sentences.\n\n4. How did you get equation (2)? Why \"nodes with low values of this measure are likely to correspond to regions where over-smoothing and over-squashing occur\"?\n\n5. \"The long-range dependencies are then learned via a GNN heterophilic aggregation.\" What is \"heterophilic aggregation\"? Do you mean aggregation from long-range nodes in different classes? Are such long-range dependency beneficial?\n\n6. \"This concept of homophily-based interaction-decoupling is crucial to prevent over-smoothing by avoiding using a standard GNN aggregation on heterophilic edges.\" The \"decoupling\" is indeed important, for example in [4], the authors use 3-channel architectures to address heterophily. But the objective is not to prevent over-smoothing, it is to improve node distinguishability [5]. A direct proof on why and how your proposed method can improve node distinguishability is recommended.\n\n7. Missing comparison with some SOTA models on heterophilic graphs, e.g. [4,6,7]. More comparisons on the real challenging heterophilic datasets suggested in [3] are recommended.\n\n\n\n[1] Müller L, Galkin M, Morris C, Rampášek L. Attending to Graph Transformers. Transactions on Machine Learning Research.\n\n[2] Less is More: on the Over-Globalizing Problem in Graph Transformers. InForty-first International Conference on Machine Learning.\n\n[3] The heterophilic graph learning handbook: Benchmarks, models, theoretical analysis, applications and challenges. arXiv preprint arXiv:2407.09618. 2024 Jul 12.\n\n[4] Revisiting heterophily for graph neural networks. Advances in neural information processing systems. 2022 Dec 6;35:1362-75.\n\n[5] When Do Graph Neural Networks Help with Node Classification? Investigating the Homophily Principle on Node Distinguishability. Advances in Neural Information Processing Systems. 2024 Feb 13;36.\n\n[6] Simplifying approach to node classification in graph neural networks[J]. Journal of Computational Science, 2022, 62: 101695.\n\n[7] Diverse message passing for attribute with heterophily[J]. Advances in Neural Information Processing Systems, 2021, 34: 4751-4763.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper  introduces a mechanism to mitigate over-smoothing and over-squashing in Graph Neural Networks (GNNs) by implementing an \"information flow control\" strategy that utilizes an \"information flow score.\" This approach allows for effective management of node embeddings across varied graph structures, demonstrating enhanced performance in large-scale graphs while maintaining computational efficiency.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "originality: good\nquality: medium\nclarity: medium\nsignificance: medium", "weaknesses": "1. \"These long-range interactions (LRIs) are crucial for node classification tasks, as they help distinguish between different classes and improve classification accuracy\" This is not true. For example, graph transformers are good at capturing long-range node dependencies. However, they perform poorly on node classification tasks, especially on heterophilic graphs [1]. It is found that distant information is not always useful, and the over-globalization can cause performance degradation of graph models [2].\n2. \"over-smoothing is not only a topological phenomenon but is primarily a consequence of graph heterophily.\" There is no causal relation between over-smoothing and heterophily. As stated in [3], over-smoothing only happens in deep GNNs, but not in shallow GNNs. Heterophily will cause performance degradation to all GNN models, not matter they are deep or shallow.", "questions": "1. the information flow score, which identifies graph bottlenecks and heterophilic node interactions,\n\n2. In definition 1, the first Delta embeddings look like the \"norm\" of the high-pass filtered graph signal or the neighborhood diversification[4]. The second Delta embedding is a new and interesting one.\n\n3. So how can lemma 1 and 2 offer insights into the graph’s homophily and topology? Explain with sentences.\n\n4. How did you get equation (2)? Why \"nodes with low values of this measure are likely to correspond to regions where over-smoothing and over-squashing occur\"?\n\n5. \"The long-range dependencies are then learned via a GNN heterophilic aggregation.\" What is \"heterophilic aggregation\"? Do you mean aggregation from long-range nodes in different classes? Are such long-range dependency beneficial?\n\n6. \"This concept of homophily-based interaction-decoupling is crucial to prevent over-smoothing by avoiding using a standard GNN aggregation on heterophilic edges.\" The \"decoupling\" is indeed important, for example in [4], the authors use 3-channel architectures to address heterophily. But the objective is not to prevent over-smoothing, it is to improve node distinguishability [5]. A direct proof on why and how your proposed method can improve node distinguishability is recommended.\n\n7. Missing comparison with some SOTA models on heterophilic graphs, e.g. [4,6,7]. More comparisons on the real challenging heterophilic datasets suggested in [3] are recommended.\n\n\n\n[1] Müller L, Galkin M, Morris C, Rampášek L. Attending to Graph Transformers. Transactions on Machine Learning Research.\n\n[2] Less is More: on the Over-Globalizing Problem in Graph Transformers. InForty-first International Conference on Machine Learning.\n\n[3] The heterophilic graph learning handbook: Benchmarks, models, theoretical analysis, applications and challenges. arXiv preprint arXiv:2407.09618. 2024 Jul 12.\n\n[4] Revisiting heterophily for graph neural networks. Advances in neural information processing systems. 2022 Dec 6;35:1362-75.\n\n[5] When Do Graph Neural Networks Help with Node Classification? Investigating the Homophily Principle on Node Distinguishability. Advances in Neural Information Processing Systems. 2024 Feb 13;36.\n\n[6] Simplifying approach to node classification in graph neural networks[J]. Journal of Computational Science, 2022, 62: 101695.\n\n[7] Diverse message passing for attribute with heterophily[J]. Advances in Neural Information Processing Systems, 2021, 34: 4751-4763.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730536316128}, {"id": "Z3ypth6wpK", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3270/Reviewer_esWr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces a score based on node features aggregated in a GNN layer, that aims at capturing the likelihood of a node to be responsible for oversmoothing and oversquashing. By leveraging such score in a graph-filtering pipeline, the authors propose a framework to alter the graph connectivity within a GNN scheme.", "review_text": "The paper introduces a score based on node features aggregated in a GNN layer, that aims at capturing the likelihood of a node to be responsible for oversmoothing and oversquashing. By leveraging such score in a graph-filtering pipeline, the authors propose a framework to alter the graph connectivity within a GNN scheme.", "strengths": "I think that it is valuable addressing issues like oversquashing and oversmoothing simultaneously, rather than studying them in isolation and independently of one another. I also liked the idea of leveraging \"moments\" from the feature distribution at different layers to guide the graph-filtering process.", "weaknesses": "There are important aspects of the submission that require reworking. \n\n**Message and Presentation**\n\n- In the introduction, there is often some ambiguity in the way you mention oversmoothing and oversquashing, as if they were interchangeable concepts. This is not the case, and should be emphasized. Oversmoothing is a problem that occurs for *some* GNNs and is independent of the topology (as a phenomenon, not how quickly that occurs) and is somewhat orthogonal to long-range interactions since in the limit of many layers, node features become indistinguishable irrespectively of their distance. Oversquashing instead, is an issue that occurs for *all* 1-hop GNNs and is very much dependent on the topology (namely, their commute time) and hence affects long-range interactions, independent of the depth or the ability to capture local interactions.\n- Even more significantly, you keep overlapping the issue of oversmoothing with that of heterohily (for example Line 110, Line 121, Line 161 but this notion is repeated throughout the paper). This is wrong. While Definition 2.1 accounts for the labels, this to me represents more of a choice, as oversmoothing is the convergence of node features to the same representation over a connected component of the graph. As such, it is actually simply caused by low-frequencies dominating over high-frequencies in the graph spectrum. In fact, it can be mitigated or avoided by relying on architectures that do not operate via low-pass filters. I suspect that what you are implying here, is that oversmoothing becomes more of an issue in the presence of heterophily, as nodes with different labels become indistinguishable, but *this is a consequence of and not the cause of oversmoothing and should be rectified*.\n\n- Quite a few citations are missing in the related work, for example regarding rewiring [1,2,3] but also Graph-Transformers.. \n- The presentation of the framework is a little contrived (see my questions below). Also, while you try to distinguish yourself from graph-rewiring algorithms, your approach removes edges, and this is a key part of it. For this reason, I think it is a little misleading to distinguish yourself from graph rewiring techniques. You should be more specific, and mention that the rewiring is adaptive and based on GNN layer outputs more than topological connectivity measures.\n\n**Theory**\n\n- I am a little confused by Lemma 1. To me, the homophily of a node only depends on the label information and the topology and has nothing to do with the architecture being used and/or the features. This indeed seems to be reflected also in your Definition 2.2 where I am reading that $\\Phi$ can be taken to be the ground-truth label assignment. However, it seems that in Lemma 1 you are deriving the homophily of a node based on what can be mapped/separated from the node features, i.e. it has more to do with distinguishability from node features. If so, this should be clearly emphasized. As such, I would not really talk about homophily but node features separability.\n\n- I don’t think that Lemma 2 is an actual Lemma since your proof is essentially a discussion based on  the results of Nguyen et al. You should remove the statement and replace it with a discussion based on what you have in the appendix. As it stands, I find it confusing and indeed informal, to a point that this is not a mathematical statement.\n\n- In light of my comments regarding Lemma 2, I don’t think that your score definition is that well motivated. More precisely, I can see why the denominator makes sense in relation to oversmoothing, since it measures node features separability after rounds of message passing (and *not* homophily), but I struggle to see how the numerator relates to oversquashing. You should expand on the “proof” from Lemma 2, which is not really a proof, to better motivate this score.\n\n**Experiments**\n\nEvaluation is not  convincing. On all the benchmarks you used, it is highly debatable that long-range interactions are present at any level. In fact, I believe majority of people would argue that LRIs are not present on Cora, Pubmed, etc.. Additionally, datasets like Texas, Wisconsin, etc are known to have several issues and the community has proposed alternative options. I personally struggle to accept claims of “state of the art improvements by 1 %” on the likes of Cora and Pubmed this day. Graphs like Cornell, Texas and Wisconsin are also extremely small and super sensitive to tuning. The paper overall proposed a methodology, and as such, should be thoroughly tested on more relevant benchmarks. \n\n[1]: Mitchell Black, Zhengchao Wan, Amir Nayyeri, and Yusu Wang. Understanding oversquashing in\ngnns through the lens of effective resistance, ICML23.\n\n[2]: Adrián Arnaiz-Rodríguez, Ahmed Begga, Francisco Escolano, and Nuria Oliver. DiffWire: Inductive\nGraph Rewiring via the Lovász Bound, LOG 2022.\n\n[3]: Kedar Karhadkar, Pradeep Kr Banerjee, and Guido Montúfar. Fosr: First-order spectral rewiring for\naddressing oversquashing in gnns, 2022.", "questions": "- Equation (1) is not the most general way of writing a 1-hop GNN aggregation, as there is no residual term. Namely, one would typically expect $\\phi$ to take two arguments i.e. $(\\mathbf{X}_u^t, \\bigoplus...)$\n- Line 159: The expression “embedding agnostic” is a little vague to me, so perhaps you can specify a little more clearly what you are implying here.\n- Line 285: What is a “homophilic GNN”?\n- The paragraph 283-291 uses too many vague words and is all but clear. For example, line 288-289, what would an “heterophilic graph condensation” be? \n- Line 330–331: How can removing edges that are bottlenecks necessarily reduce oversquashing? What if now you have disconnected components? This process can only work if one identifies correctly node labels, but this is something that your algorithm in general cannot know in advance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a score based on node features aggregated in a GNN layer, that aims at capturing the likelihood of a node to be responsible for oversmoothing and oversquashing. By leveraging such score in a graph-filtering pipeline, the authors propose a framework to alter the graph connectivity within a GNN scheme.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "I think that it is valuable addressing issues like oversquashing and oversmoothing simultaneously, rather than studying them in isolation and independently of one another. I also liked the idea of leveraging \"moments\" from the feature distribution at different layers to guide the graph-filtering process.", "weaknesses": "There are important aspects of the submission that require reworking. \n\n**Message and Presentation**\n\n- In the introduction, there is often some ambiguity in the way you mention oversmoothing and oversquashing, as if they were interchangeable concepts. This is not the case, and should be emphasized. Oversmoothing is a problem that occurs for *some* GNNs and is independent of the topology (as a phenomenon, not how quickly that occurs) and is somewhat orthogonal to long-range interactions since in the limit of many layers, node features become indistinguishable irrespectively of their distance. Oversquashing instead, is an issue that occurs for *all* 1-hop GNNs and is very much dependent on the topology (namely, their commute time) and hence affects long-range interactions, independent of the depth or the ability to capture local interactions.\n- Even more significantly, you keep overlapping the issue of oversmoothing with that of heterohily (for example Line 110, Line 121, Line 161 but this notion is repeated throughout the paper). This is wrong. While Definition 2.1 accounts for the labels, this to me represents more of a choice, as oversmoothing is the convergence of node features to the same representation over a connected component of the graph. As such, it is actually simply caused by low-frequencies dominating over high-frequencies in the graph spectrum. In fact, it can be mitigated or avoided by relying on architectures that do not operate via low-pass filters. I suspect that what you are implying here, is that oversmoothing becomes more of an issue in the presence of heterophily, as nodes with different labels become indistinguishable, but *this is a consequence of and not the cause of oversmoothing and should be rectified*.\n\n- Quite a few citations are missing in the related work, for example regarding rewiring [1,2,3] but also Graph-Transformers.. \n- The presentation of the framework is a little contrived (see my questions below). Also, while you try to distinguish yourself from graph-rewiring algorithms, your approach removes edges, and this is a key part of it. For this reason, I think it is a little misleading to distinguish yourself from graph rewiring techniques. You should be more specific, and mention that the rewiring is adaptive and based on GNN layer outputs more than topological connectivity measures.\n\n**Theory**\n\n- I am a little confused by Lemma 1. To me, the homophily of a node only depends on the label information and the topology and has nothing to do with the architecture being used and/or the features. This indeed seems to be reflected also in your Definition 2.2 where I am reading that $\\Phi$ can be taken to be the ground-truth label assignment. However, it seems that in Lemma 1 you are deriving the homophily of a node based on what can be mapped/separated from the node features, i.e. it has more to do with distinguishability from node features. If so, this should be clearly emphasized. As such, I would not really talk about homophily but node features separability.\n\n- I don’t think that Lemma 2 is an actual Lemma since your proof is essentially a discussion based on  the results of Nguyen et al. You should remove the statement and replace it with a discussion based on what you have in the appendix. As it stands, I find it confusing and indeed informal, to a point that this is not a mathematical statement.\n\n- In light of my comments regarding Lemma 2, I don’t think that your score definition is that well motivated. More precisely, I can see why the denominator makes sense in relation to oversmoothing, since it measures node features separability after rounds of message passing (and *not* homophily), but I struggle to see how the numerator relates to oversquashing. You should expand on the “proof” from Lemma 2, which is not really a proof, to better motivate this score.\n\n**Experiments**\n\nEvaluation is not  convincing. On all the benchmarks you used, it is highly debatable that long-range interactions are present at any level. In fact, I believe majority of people would argue that LRIs are not present on Cora, Pubmed, etc.. Additionally, datasets like Texas, Wisconsin, etc are known to have several issues and the community has proposed alternative options. I personally struggle to accept claims of “state of the art improvements by 1 %” on the likes of Cora and Pubmed this day. Graphs like Cornell, Texas and Wisconsin are also extremely small and super sensitive to tuning. The paper overall proposed a methodology, and as such, should be thoroughly tested on more relevant benchmarks. \n\n[1]: Mitchell Black, Zhengchao Wan, Amir Nayyeri, and Yusu Wang. Understanding oversquashing in\ngnns through the lens of effective resistance, ICML23.\n\n[2]: Adrián Arnaiz-Rodríguez, Ahmed Begga, Francisco Escolano, and Nuria Oliver. DiffWire: Inductive\nGraph Rewiring via the Lovász Bound, LOG 2022.\n\n[3]: Kedar Karhadkar, Pradeep Kr Banerjee, and Guido Montúfar. Fosr: First-order spectral rewiring for\naddressing oversquashing in gnns, 2022.", "questions": "- Equation (1) is not the most general way of writing a 1-hop GNN aggregation, as there is no residual term. Namely, one would typically expect $\\phi$ to take two arguments i.e. $(\\mathbf{X}_u^t, \\bigoplus...)$\n- Line 159: The expression “embedding agnostic” is a little vague to me, so perhaps you can specify a little more clearly what you are implying here.\n- Line 285: What is a “homophilic GNN”?\n- The paragraph 283-291 uses too many vague words and is all but clear. For example, line 288-289, what would an “heterophilic graph condensation” be? \n- Line 330–331: How can removing edges that are bottlenecks necessarily reduce oversquashing? What if now you have disconnected components? This process can only work if one identifies correctly node labels, but this is something that your algorithm in general cannot know in advance.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729935856532}], "openreview_url": "https://openreview.net/forum?id=xMxHJxp192", "arxiv_id": "2501.06002", "paper_pdf": "papers/xMxHJxp192.pdf", "paper_pdf_sha256": "59ac11372de6ce46ea080ba2fe123929962f371c3502b6e11e310cf44ef7def2", "paper_pdf_bytes": 4961402, "paper_pdf_source": "openreview", "code_url": "https://github.com/basiralab/DeltaGNN", "code_repository": "basiralab/DeltaGNN", "code_commit": "5e0c3bdc74e87e4500c6fdaec575841e7a0ecf00", "code_archive": "repos/xMxHJxp192.zip", "code_archive_sha256": "9f6cbd99461f23545bae753dff47ee13a9916f4168cf9e4e7d64088e76436540", "code_archive_bytes": 191110, "code_file_count": 35, "code_extensions": {".py": 27, ".sh": 8}, "github_disk_usage_kb": 174, "github_languages": {"Python": 252592, "Shell": 156030}, "github_archived": false, "github_pushed_at": "2026-08-26T09:29:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deltagnn-graph-neural-network-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8SPSIfR2e0", "year": 2024, "status": "rejected", "title": "Dissecting Language Models: Machine Unlearning via Selective Pruning", "authors": ["Nicky Pochinkov", "Nandi Schoots"], "authorids": ["~Nicky_Pochinkov1", "~Nandi_Schoots1"], "authors_source": "OpenReview API", "abstract": "Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted.\nThis paper introduces a machine unlearning method specifically designed for LLMs. \nWe introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall network performance. \nThis approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours.\nOur findings reveal that both feed-forward and attention neurons in LLMs are specialized; \nthat is, for specific tasks, certain neurons are more crucial than others.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "w8ESNORUvA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5824/Reviewer_q42J"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This article presented an approach for the targeted removal of neurons, which is based on their comparative significance across two datasets. This technique of machine unlearning demonstrates its effectiveness through the quantifiable decrease in accuracy differentials and perplexity measurements. Additionally, it establishes a cost-effective foundation for forthcoming research comparisons. Its theory posits that the approach is more inclined to eliminate undesired model behaviors, as opposed to merely concealing them, in contrast to fine-tuning.\nThis approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. The findings of this method reveals that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others.", "review_text": "This article presented an approach for the targeted removal of neurons, which is based on their comparative significance across two datasets. This technique of machine unlearning demonstrates its effectiveness through the quantifiable decrease in accuracy differentials and perplexity measurements. Additionally, it establishes a cost-effective foundation for forthcoming research comparisons. Its theory posits that the approach is more inclined to eliminate undesired model behaviors, as opposed to merely concealing them, in contrast to fine-tuning.\nThis approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. The findings of this method reveals that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others.", "strengths": "1- The model presented here tackles an intriguing and complex issue within large language models (LLMs), focusing on the significance scores assigned to individual neurons with respect to a specific target dataset.\n\n2- The concept of machine unlearning is implemented across both types of neural networks, including feedforward and attention-layer-based networks. By eliminating unwanted neurons for any given target dataset, the process is swift and leads to a reduction in the network's computational load.\n\n3- The efficacy of the unlearning process is demonstrated through experiments conducted on three distinct datasets.", "weaknesses": "1- The proposed model can effectively eliminate the information captured by the target dataset. However, it is unable to unlearn knowledge that lies beyond the representation of the datasets. Please provide a clear justification.\n\n2- Including a visual representation of the suggested concept could enhance comprehension. Thus, kindly incorporate a diagram that presents a general outline of the proposed concept.\n\n3- The evaluation of the proposed model focuses on text datasets, yet the experiments for the image dataset are absent. Implementing machine unlearning for the image dataset would be highly beneficial.", "questions": "Please address all the question raised in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article presented an approach for the targeted removal of neurons, which is based on their comparative significance across two datasets. This technique of machine unlearning demonstrates its effectiveness through the quantifiable decrease in accuracy differentials and perplexity measurements. Additionally, it establishes a cost-effective foundation for forthcoming research comparisons. Its theory posits that the approach is more inclined to eliminate undesired model behaviors, as opposed to merely concealing them, in contrast to fine-tuning.\nThis approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. The findings of this method reveals that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1- The model presented here tackles an intriguing and complex issue within large language models (LLMs), focusing on the significance scores assigned to individual neurons with respect to a specific target dataset.\n\n2- The concept of machine unlearning is implemented across both types of neural networks, including feedforward and attention-layer-based networks. By eliminating unwanted neurons for any given target dataset, the process is swift and leads to a reduction in the network's computational load.\n\n3- The efficacy of the unlearning process is demonstrated through experiments conducted on three distinct datasets.", "weaknesses": "1- The proposed model can effectively eliminate the information captured by the target dataset. However, it is unable to unlearn knowledge that lies beyond the representation of the datasets. Please provide a clear justification.\n\n2- Including a visual representation of the suggested concept could enhance comprehension. Thus, kindly incorporate a diagram that presents a general outline of the proposed concept.\n\n3- The evaluation of the proposed model focuses on text datasets, yet the experiments for the image dataset are absent. Implementing machine unlearning for the image dataset would be highly beneficial.", "questions": "Please address all the question raised in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698865285067}, {"id": "z9KsZ0NjmC", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5824/Reviewer_9FMg"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work presents a novel selective pruning method in order to allow trained models to 'unlearn' specific capabilities. Related work and the details of the method are explicated. Experiments on different models with different sizes are presented on two data splits (code/pile and code/python) showing generally good performance at 'unlearning' the forgetting dataset and retaining the other. Analysis experiments compare the efficacy of pruning feed-forwards vs attention neurons. Experiments on toxicity show good results for the method on allowing the unlearning of toxic text.", "review_text": "This work presents a novel selective pruning method in order to allow trained models to 'unlearn' specific capabilities. Related work and the details of the method are explicated. Experiments on different models with different sizes are presented on two data splits (code/pile and code/python) showing generally good performance at 'unlearning' the forgetting dataset and retaining the other. Analysis experiments compare the efficacy of pruning feed-forwards vs attention neurons. Experiments on toxicity show good results for the method on allowing the unlearning of toxic text.", "strengths": "The paper is clearly written. The experiments are thorough, and the presented results convincingly demonstrate the utility of the method. The results on toxicity are particularly nice, as this is a highly relevant domain for which related techniques are well-motivated. The presented method is novel.", "weaknesses": "In Limitations: \"Our method can only be applied to remove a capability when that capability is neatly captured by a\ndataset.\" The authors rightly point out that this method of evaluation of unle'arning is dependent upon dataset-level perplexity. The extent to which this metric for any constructed dataset sufficiently \"neatly captures\" whatever capability is desired to be forgotten is difficult to asses without further analysis not presented in this work. It may be the case that for the practical scenarios which motivate the method in the first place, no such \"neat\" dataset is possible to produce. It is fine for this analysis to be out off scope of this work.\n\nGiven this evaluative dependence upon specific datasets, it would significantly strengthen the results of the paper to present more diverse experiments. While it is nice to have many different models at many different sizes, there does not seem to be much addition knowledge gleaned from that diversity, whereas having more tasks would demonstrate a broader efficacy across a more speculative dimension. \n\nIn this paper, the authors present experiments pruning either the feed-forward or the attention blocks, and leave \"Embedding, Positional\nEmbedding and Output Unembedding unmodified.\" (3.3). This is a significant constriction of the application of the method which is not more than intuitively justified. Further experiments, even just to show that this restriction is well-motivated, would strengthen the work.", "questions": "See weaknesses.\n\nIn Table 3: it would help readability to label \"Task Arithmetic (quoted)\" as being a finetuning task. Also, the gap between the baseline (quoted) and (replicated) is fairly large, which makes actually comparing the finetuning vs presented pruning method difficuly. Would it be possible to replicate the task-arithmetic finetuning? That would significantly improve the comparability of these results.\n\n\nsmall issue:\nThe details of iterative pruning are not specified. It seems like it would be important to the function of the method to set the proportion of nodes pruned per iteration well, but this is not discussed. Even if this is not and important hyper-parameter of the method, a clearer explanation of the iterative pruning method is necessary to fully elucidate the applied method.\n\n6.3: If an LLM were trained not to answer questions about a dangerous topic, say, bomb-building, could the presented method not be used to unlearn that guardrail? I don't think this is a concern specific to the presented method, but I do not follow why this method is less likely to generate harmful systems than other methods. Can the authors clarify?\n\nNit:\nin discussion: hypothesise -> hypothesize", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a novel selective pruning method in order to allow trained models to 'unlearn' specific capabilities. Related work and the details of the method are explicated. Experiments on different models with different sizes are presented on two data splits (code/pile and code/python) showing generally good performance at 'unlearning' the forgetting dataset and retaining the other. Analysis experiments compare the efficacy of pruning feed-forwards vs attention neurons. Experiments on toxicity show good results for the method on allowing the unlearning of toxic text.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper is clearly written. The experiments are thorough, and the presented results convincingly demonstrate the utility of the method. The results on toxicity are particularly nice, as this is a highly relevant domain for which related techniques are well-motivated. The presented method is novel.", "weaknesses": "In Limitations: \"Our method can only be applied to remove a capability when that capability is neatly captured by a\ndataset.\" The authors rightly point out that this method of evaluation of unle'arning is dependent upon dataset-level perplexity. The extent to which this metric for any constructed dataset sufficiently \"neatly captures\" whatever capability is desired to be forgotten is difficult to asses without further analysis not presented in this work. It may be the case that for the practical scenarios which motivate the method in the first place, no such \"neat\" dataset is possible to produce. It is fine for this analysis to be out off scope of this work.\n\nGiven this evaluative dependence upon specific datasets, it would significantly strengthen the results of the paper to present more diverse experiments. While it is nice to have many different models at many different sizes, there does not seem to be much addition knowledge gleaned from that diversity, whereas having more tasks would demonstrate a broader efficacy across a more speculative dimension. \n\nIn this paper, the authors present experiments pruning either the feed-forward or the attention blocks, and leave \"Embedding, Positional\nEmbedding and Output Unembedding unmodified.\" (3.3). This is a significant constriction of the application of the method which is not more than intuitively justified. Further experiments, even just to show that this restriction is well-motivated, would strengthen the work.", "questions": "See weaknesses.\n\nIn Table 3: it would help readability to label \"Task Arithmetic (quoted)\" as being a finetuning task. Also, the gap between the baseline (quoted) and (replicated) is fairly large, which makes actually comparing the finetuning vs presented pruning method difficuly. Would it be possible to replicate the task-arithmetic finetuning? That would significantly improve the comparability of these results.\n\n\nsmall issue:\nThe details of iterative pruning are not specified. It seems like it would be important to the function of the method to set the proportion of nodes pruned per iteration well, but this is not discussed. Even if this is not and important hyper-parameter of the method, a clearer explanation of the iterative pruning method is necessary to fully elucidate the applied method.\n\n6.3: If an LLM were trained not to answer questions about a dangerous topic, say, bomb-building, could the presented method not be used to unlearn that guardrail? I don't think this is a concern specific to the presented method, but I do not follow why this method is less likely to generate harmful systems than other methods. Can the authors clarify?\n\nNit:\nin discussion: hypothesise -> hypothesize", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698822401760}, {"id": "5V3JTm28ff", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5824/Reviewer_sDUD"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes an unlearning algorithm based on pruning for removing ‘capabilities’ from pretrained language models. Specifically, they provide different definitions for quantifying the importance of a neuron for a particular dataset based on the value of its activation on that the points in that dataset. Then, they define a ‘score’ for a neuron as the ratio of importance of that neuron for the retain versus the forget datasets. They prune a chosen percentage of nodes from a ranked list according to this score. Empirically, they experiment in a setting where a general purpose model is caused to forget its ‘coding ability’ (and the reverse of forgetting all but its coding ability). They also look at a more finegrained scenario of removing python coding ability while retaining the ability to code in other languages (and its reverse). The way that forgetting is measured is by inspecting the relative drop in accuracy on the forget set (relative to the relative drop on the retain set). My understanding is that it is desired according to this evaluation to have a large drop in the forget accuracy without having a large drop in the retain accuracy. They investigate empirically these trade-off curves obtained by their method in different types of language models. Then, on a different task/dataset (where the goal is to remove toxicity from a trained language model), they compare against one baseline and claim that they get similar results.", "review_text": "This paper proposes an unlearning algorithm based on pruning for removing ‘capabilities’ from pretrained language models. Specifically, they provide different definitions for quantifying the importance of a neuron for a particular dataset based on the value of its activation on that the points in that dataset. Then, they define a ‘score’ for a neuron as the ratio of importance of that neuron for the retain versus the forget datasets. They prune a chosen percentage of nodes from a ranked list according to this score. Empirically, they experiment in a setting where a general purpose model is caused to forget its ‘coding ability’ (and the reverse of forgetting all but its coding ability). They also look at a more finegrained scenario of removing python coding ability while retaining the ability to code in other languages (and its reverse). The way that forgetting is measured is by inspecting the relative drop in accuracy on the forget set (relative to the relative drop on the retain set). My understanding is that it is desired according to this evaluation to have a large drop in the forget accuracy without having a large drop in the retain accuracy. They investigate empirically these trade-off curves obtained by their method in different types of language models. Then, on a different task/dataset (where the goal is to remove toxicity from a trained language model), they compare against one baseline and claim that they get similar results.", "strengths": "- the paper studies an interesting and important problem\n- the proposed method is a reasonable idea and well motivated\n- the proposed method is efficient and requires no gradient updates \n- the paper is for the most part well-written (though see some exceptions below)", "weaknesses": "- the related work section is weak, missing a lot of literature both from unlearning and pruning. For example, [A-E] are recent papers on unlearning, with [A, B] particularly related to the methodology of this paper (see References below). I’m less familiar with the sparsity and pruning literature but the authors should conduct a thorough review.\n- the authors compare against only one baseline, and only for one task. Other common baselines include finetuning on the retain set, gradient ascent on the forget set, and comparing against other recent unlearning methods is also important (see the references below)\n- the particular problem setting of unlearning is not clearly defined. What defines successful unlearning here? In section 3 the authors describe the forget set as the “task or dataset that we aim to reduce performance on”. This isn’t precise enough. How much do we want to reduce performance? Is it that the greater the reduction, the better? For context, unlearning papers (e.g. Golatkar et al, which the authors cite) usually consider that the accuracy on the forget set should be reduced only up to a reference point and no further (where the reference point is given by the oracle unlearning algorithm of retraining the model from scratch without the forget set). Is the setup here different, and if so, how is the goal defined in this case?\n- [clarity] It seems that ‘task’ and ‘dataset’ are used interchangeably in the paper which causes confusion. For example, ‘the task that we are optimizing for as the retain dataset’. To me, a ‘task’ includes a particular training objective whereas ‘dataset’ refers to raw data.\n- [clarity] fundamental problem setting details are missing. For example, were Pile and Code included in the dataset that the various language models were trained on? If not, the terms “forgetting” or “unlearning” may be ill-suited for this application (as they usually refer to forget parts of the training dataset). At the very least, the problem setting targeted by this paper should be clearly defined. \n- [soundness]: can decreased performance (accuracy, perplexity) on a particular dataset support claims of removal of a capability? Generally, ‘capability removal’ is not precisely defined.\n- [soundness]: when it comes to forgetting or unlearning, several metrics have been proposed by the community to measure this. Simply inspecting the accuracy / perplexity is likely a poor proxy for forgetting quality. For example, Membership Inference Attacks are an important category of methods (see e.g. Golatkar et al and reference [E] below). Are these not applicable here. If not then why not?\n- [presentation, soundness] the authors use the term “selective” to describe an unlearning method, without defining this term clearly in this context. My understanding of what “selective” means in this context is the ability of reducing accuracy / performance on the forget dataset without (really) damaging the accuracy / performance on the retain dataset. If my understanding is correct, I don’t agree with the claim that Figure 1a shows that the larger the model, the more selective it is. I can see this being true for the Opt family but not the Pythia family, for example.\n- [presentation] In the paragraph under Definition 1, the authors give intuition for the different influence functions but they omit I_{abs}. Please add.\n- [presentation] Above Table 1, the authors describe the models they use (which correspond to columns in Table 1) but they omit Roberta. Please add.\n- [presentation] “As a baseline, we also randomly pruned layers” – the authors claim this but I don’t see this baseline in their experiments (at least in the main paper), unless I’m missing it. \n- [empirical results] It would greatly strengthen the paper to conduct analyses of: the effect of the number of pruning iterations, the effect of the % pruned (in each iteration or overall), the effect of the choice of the influence function. It sounds like the authors have some results on some of these in the Appendix but not all of them. It would also be great to summarize in the main paper all of the findings (so that one doesn’t need to read the entire Appendix). \n- [empirical results] please include confidence intervals in all tables and e.g. in Figure 3. It is currently challenging to tell if the differences are significant\n- [empirical results] why is there such a large difference between “Base (quoted)” and “Base (replicated)” in Table 3? This makes me concerned about whether “Task Arithmetic (quoted)” and “Pruned” are comparable.\n\nReferences\n=========\n- [A] Fast Machine Unlearning Without Retraining Through Selective Synaptic Dampening. Foster et al 2023.\n- [B] Model Sparsity Can Simplify Machine Unlearning. Jia et al. NeurIPS 2023.\n- [C] Unrolling SGD: Understanding Factors Influencing Machine Unlearning. Thudi et al.\n- [D] Prompt Certified Machine Unlearning with Randomized Gradient Smoothing and Quantization. Zhang et al. NeurIPS 2022.\n- [E] Towards Unbounded Machine Unlearning. Kurmanji et al. NeurIPS 2023.", "questions": "- the authors claim that their method is specifically designed for LLMs but it’s unclear to me why that’s the case. Can’t this method be applied out of the box e.g. to vision transformers? If not, then why not?\n- In Section 3.1, the authors discuss some observations about the distributions of activations that motivated their use of importance functions. Which dataset / setting were these observations made in? Have the authors made an effort to confirm that they are generalizable beyond a certain setting?\n- I don’t really understand the statement that “We also do not want to directly modify specific dimensions of the embedding space”. Did not understand the provided rationale. If it’s not the right level of granularity for pruning, as the authors claim, would the proposed scoring function not capture this? If not, does this suggest we need to design better scoring functions?\n- In the Discussion, the authors hypothesize that their method is more likely to “actually remove the undesired behaviour” compared to other unlearning methods. Similarly, in the Broader Impacts section, they claim that (compared to other unlearning methods) their method is unlikely to generate systems that are more harmful than the base model. What is the evidence used for making these claims?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an unlearning algorithm based on pruning for removing ‘capabilities’ from pretrained language models. Specifically, they provide different definitions for quantifying the importance of a neuron for a particular dataset based on the value of its activation on that the points in that dataset. Then, they define a ‘score’ for a neuron as the ratio of importance of that neuron for the retain versus the forget datasets. They prune a chosen percentage of nodes from a ranked list according to this score. Empirically, they experiment in a setting where a general purpose model is caused to forget its ‘coding ability’ (and the reverse of forgetting all but its coding ability). They also look at a more finegrained scenario of removing python coding ability while retaining the ability to code in other languages (and its reverse). The way that forgetting is measured is by inspecting the relative drop in accuracy on the forget set (relative to the relative drop on the retain set). My understanding is that it is desired according to this evaluation to have a large drop in the forget accuracy without having a large drop in the retain accuracy. They investigate empirically these trade-off curves obtained by their method in different types of language models. Then, on a different task/dataset (where the goal is to remove toxicity from a trained language model), they compare against one baseline and claim that they get similar results.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- the paper studies an interesting and important problem\n- the proposed method is a reasonable idea and well motivated\n- the proposed method is efficient and requires no gradient updates \n- the paper is for the most part well-written (though see some exceptions below)", "weaknesses": "- the related work section is weak, missing a lot of literature both from unlearning and pruning. For example, [A-E] are recent papers on unlearning, with [A, B] particularly related to the methodology of this paper (see References below). I’m less familiar with the sparsity and pruning literature but the authors should conduct a thorough review.\n- the authors compare against only one baseline, and only for one task. Other common baselines include finetuning on the retain set, gradient ascent on the forget set, and comparing against other recent unlearning methods is also important (see the references below)\n- the particular problem setting of unlearning is not clearly defined. What defines successful unlearning here? In section 3 the authors describe the forget set as the “task or dataset that we aim to reduce performance on”. This isn’t precise enough. How much do we want to reduce performance? Is it that the greater the reduction, the better? For context, unlearning papers (e.g. Golatkar et al, which the authors cite) usually consider that the accuracy on the forget set should be reduced only up to a reference point and no further (where the reference point is given by the oracle unlearning algorithm of retraining the model from scratch without the forget set). Is the setup here different, and if so, how is the goal defined in this case?\n- [clarity] It seems that ‘task’ and ‘dataset’ are used interchangeably in the paper which causes confusion. For example, ‘the task that we are optimizing for as the retain dataset’. To me, a ‘task’ includes a particular training objective whereas ‘dataset’ refers to raw data.\n- [clarity] fundamental problem setting details are missing. For example, were Pile and Code included in the dataset that the various language models were trained on? If not, the terms “forgetting” or “unlearning” may be ill-suited for this application (as they usually refer to forget parts of the training dataset). At the very least, the problem setting targeted by this paper should be clearly defined. \n- [soundness]: can decreased performance (accuracy, perplexity) on a particular dataset support claims of removal of a capability? Generally, ‘capability removal’ is not precisely defined.\n- [soundness]: when it comes to forgetting or unlearning, several metrics have been proposed by the community to measure this. Simply inspecting the accuracy / perplexity is likely a poor proxy for forgetting quality. For example, Membership Inference Attacks are an important category of methods (see e.g. Golatkar et al and reference [E] below). Are these not applicable here. If not then why not?\n- [presentation, soundness] the authors use the term “selective” to describe an unlearning method, without defining this term clearly in this context. My understanding of what “selective” means in this context is the ability of reducing accuracy / performance on the forget dataset without (really) damaging the accuracy / performance on the retain dataset. If my understanding is correct, I don’t agree with the claim that Figure 1a shows that the larger the model, the more selective it is. I can see this being true for the Opt family but not the Pythia family, for example.\n- [presentation] In the paragraph under Definition 1, the authors give intuition for the different influence functions but they omit I_{abs}. Please add.\n- [presentation] Above Table 1, the authors describe the models they use (which correspond to columns in Table 1) but they omit Roberta. Please add.\n- [presentation] “As a baseline, we also randomly pruned layers” – the authors claim this but I don’t see this baseline in their experiments (at least in the main paper), unless I’m missing it. \n- [empirical results] It would greatly strengthen the paper to conduct analyses of: the effect of the number of pruning iterations, the effect of the % pruned (in each iteration or overall), the effect of the choice of the influence function. It sounds like the authors have some results on some of these in the Appendix but not all of them. It would also be great to summarize in the main paper all of the findings (so that one doesn’t need to read the entire Appendix). \n- [empirical results] please include confidence intervals in all tables and e.g. in Figure 3. It is currently challenging to tell if the differences are significant\n- [empirical results] why is there such a large difference between “Base (quoted)” and “Base (replicated)” in Table 3? This makes me concerned about whether “Task Arithmetic (quoted)” and “Pruned” are comparable.\n\nReferences\n=========\n- [A] Fast Machine Unlearning Without Retraining Through Selective Synaptic Dampening. Foster et al 2023.\n- [B] Model Sparsity Can Simplify Machine Unlearning. Jia et al. NeurIPS 2023.\n- [C] Unrolling SGD: Understanding Factors Influencing Machine Unlearning. Thudi et al.\n- [D] Prompt Certified Machine Unlearning with Randomized Gradient Smoothing and Quantization. Zhang et al. NeurIPS 2022.\n- [E] Towards Unbounded Machine Unlearning. Kurmanji et al. NeurIPS 2023.", "questions": "- the authors claim that their method is specifically designed for LLMs but it’s unclear to me why that’s the case. Can’t this method be applied out of the box e.g. to vision transformers? If not, then why not?\n- In Section 3.1, the authors discuss some observations about the distributions of activations that motivated their use of importance functions. Which dataset / setting were these observations made in? Have the authors made an effort to confirm that they are generalizable beyond a certain setting?\n- I don’t really understand the statement that “We also do not want to directly modify specific dimensions of the embedding space”. Did not understand the provided rationale. If it’s not the right level of granularity for pruning, as the authors claim, would the proposed scoring function not capture this? If not, does this suggest we need to design better scoring functions?\n- In the Discussion, the authors hypothesize that their method is more likely to “actually remove the undesired behaviour” compared to other unlearning methods. Similarly, in the Broader Impacts section, they claim that (compared to other unlearning methods) their method is unlikely to generate systems that are more harmful than the base model. What is the evidence used for making these claims?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698676156164}, {"id": "MTC399GUl7", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5824/Reviewer_Q3q4"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents the application of pruning techniques to analyze neuron behavior within large language models. Introducing a method termed \"selective pruning,\" the approach gauges the significance of each neuron based on its importance in both retained and forgotten datasets. From their experimental findings, the paper highlights that: (1) neurons exhibit high specialization, (2) larger models demonstrate more selective tendencies, and (3) feed-forward neurons show greater specialization compared to attention neurons.", "review_text": "This paper presents the application of pruning techniques to analyze neuron behavior within large language models. Introducing a method termed \"selective pruning,\" the approach gauges the significance of each neuron based on its importance in both retained and forgotten datasets. From their experimental findings, the paper highlights that: (1) neurons exhibit high specialization, (2) larger models demonstrate more selective tendencies, and (3) feed-forward neurons show greater specialization compared to attention neurons.", "strengths": "1. This paper delves into an interesting topic not previously addressed: examining the behaviour of neurons in large language models to comprehend their functionality.\n2. The article introduces a method termed \"selective pruning\" to undertake machine unlearning on large language models, subsequently employing it to analyze neuronal functions.\n3. The research offers intriguing findings from its experiments, suggesting that neurons in FFN hold greater significance than in Attention module for specialized tasks. Such insights could potentially inspire further research and insights into large language models within the community.", "weaknesses": "1. The experimental results presented by the author do not fully support the conclusion this paper wants to draw. For example, the authors mentioned in the introduction that, \"If capabilities can be separated on the level of neurons, then this can lead to modularity inside models.\" However, based on the experimental results, it appears that the entire LLM behaves highly in coupling and cannot be separated. For instance, in Figure 1(d), the performance loss on the 'code' dataset seems to be mirrored closely by a performance loss on the 'python' dataset, where the model drops close to the retain dataset and the forget dataset.\n\n2. The paper lacks comprehensive and comparable comparisons with previous methods. The authors only show the results with one baseline method (Task Arithmetic), and the comparison does not seem equitable.  It's unclear from the presented data whether their approach outperforms the baseline method. Using varying scales for the reduction in perplexity complicates the evaluation,  and it's challenging to determine whether a reduction from 4.8 to 0.8 is significant, or if a drop from 2.2 to 0.3 is more significant.\n\n3. Given that the experiments were solely conducted on datasets related to code, I am uncertain about the generalizability of the experimental results presented in the paper. For instance, the conclusion that FFN outperforms attention—might it be possible that a different task could yield an opposing conclusion.\n\n4. The readability of the entire article is not good. For instance, in Section 3.1, the author describes the distribution characteristics of the \"**attention pre-out neuron**\" activations. However, the definition of this unfamiliar term, \"attention pre-out neuron,\" is only introduced in Section 3.3. This leads to confusion for me when initially encountering the term. Additionally, the article frequently places experimental results in the appendices, referencing them in the main text and using the conclusion from these experiments in appendices to support further observation in the main text. This approach disrupts the reading flow, often requiring to flip back and forth for context. While thorough analyses and experiments are commendable, the structure of the article still needs further refinement to enhance its logical flow.", "questions": "1. Please answer the questions mentioned in Weaknesses.\n\n2.  In section 3.2: As a baseline we also randomly pruned layers. Where is this baseline?\n\n3. A prior study [1] demonstrated that, depending on the specific input, it's possible to achieve a high pruning ratio without negatively affecting performance and without the need for retraining. This suggests that there is redundancy in the neurons of the LLM when it's tasked with executing a singular function (equating a single sentence to a minor task, for instance). In light of these findings, what novel insights or observations does your paper offer in comparison to that study?\n\n[1] Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents the application of pruning techniques to analyze neuron behavior within large language models. Introducing a method termed \"selective pruning,\" the approach gauges the significance of each neuron based on its importance in both retained and forgotten datasets. From their experimental findings, the paper highlights that: (1) neurons exhibit high specialization, (2) larger models demonstrate more selective tendencies, and (3) feed-forward neurons show greater specialization compared to attention neurons.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. This paper delves into an interesting topic not previously addressed: examining the behaviour of neurons in large language models to comprehend their functionality.\n2. The article introduces a method termed \"selective pruning\" to undertake machine unlearning on large language models, subsequently employing it to analyze neuronal functions.\n3. The research offers intriguing findings from its experiments, suggesting that neurons in FFN hold greater significance than in Attention module for specialized tasks. Such insights could potentially inspire further research and insights into large language models within the community.", "weaknesses": "1. The experimental results presented by the author do not fully support the conclusion this paper wants to draw. For example, the authors mentioned in the introduction that, \"If capabilities can be separated on the level of neurons, then this can lead to modularity inside models.\" However, based on the experimental results, it appears that the entire LLM behaves highly in coupling and cannot be separated. For instance, in Figure 1(d), the performance loss on the 'code' dataset seems to be mirrored closely by a performance loss on the 'python' dataset, where the model drops close to the retain dataset and the forget dataset.\n\n2. The paper lacks comprehensive and comparable comparisons with previous methods. The authors only show the results with one baseline method (Task Arithmetic), and the comparison does not seem equitable.  It's unclear from the presented data whether their approach outperforms the baseline method. Using varying scales for the reduction in perplexity complicates the evaluation,  and it's challenging to determine whether a reduction from 4.8 to 0.8 is significant, or if a drop from 2.2 to 0.3 is more significant.\n\n3. Given that the experiments were solely conducted on datasets related to code, I am uncertain about the generalizability of the experimental results presented in the paper. For instance, the conclusion that FFN outperforms attention—might it be possible that a different task could yield an opposing conclusion.\n\n4. The readability of the entire article is not good. For instance, in Section 3.1, the author describes the distribution characteristics of the \"**attention pre-out neuron**\" activations. However, the definition of this unfamiliar term, \"attention pre-out neuron,\" is only introduced in Section 3.3. This leads to confusion for me when initially encountering the term. Additionally, the article frequently places experimental results in the appendices, referencing them in the main text and using the conclusion from these experiments in appendices to support further observation in the main text. This approach disrupts the reading flow, often requiring to flip back and forth for context. While thorough analyses and experiments are commendable, the structure of the article still needs further refinement to enhance its logical flow.", "questions": "1. Please answer the questions mentioned in Weaknesses.\n\n2.  In section 3.2: As a baseline we also randomly pruned layers. Where is this baseline?\n\n3. A prior study [1] demonstrated that, depending on the specific input, it's possible to achieve a high pruning ratio without negatively affecting performance and without the need for retraining. This suggests that there is redundancy in the neurons of the LLM when it's tasked with executing a singular function (equating a single sentence to a minor task, for instance). In light of these findings, what novel insights or observations does your paper offer in comparison to that study?\n\n[1] Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698658016232}], "openreview_url": "https://openreview.net/forum?id=8SPSIfR2e0", "arxiv_id": "2403.01267", "paper_pdf": "papers/8SPSIfR2e0.pdf", "paper_pdf_sha256": "add0e21e1e68fa1e95655332f0e8fbee761c7e882a013eba296af36471c74255", "paper_pdf_bytes": 4096323, "paper_pdf_source": "openreview", "code_url": "https://github.com/nickypro/selective-pruning", "code_repository": "nickypro/selective-pruning", "code_commit": "96109170d686fad48efa5dc1319b2b5c3868976e", "code_archive": "repos/8SPSIfR2e0.zip", "code_archive_sha256": "bce0042b8b3cdc8b55fb4df7bf4d1c66418516e6b52b7be1b04e2df33f7bdf71", "code_archive_bytes": 150062, "code_file_count": 21, "code_extensions": {".py": 16, ".sh": 5}, "github_disk_usage_kb": 143, "github_languages": {"Python": 91819, "Shell": 7426}, "github_archived": false, "github_pushed_at": "2024-07-26T17:23:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dissecting-language-models-machine-unlearning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fHT8kZcyyT", "year": 2023, "status": "rejected", "title": "CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data", "authors": ["Mathieu Chevalley", "Yusuf H Roohani", "Arash Mehrjou", "Jure Leskovec", "Patrick Schwab"], "authorids": ["~Mathieu_Chevalley1", "~Yusuf_H_Roohani1", "~Arash_Mehrjou1", "~Jure_Leskovec1", "~Patrick_Schwab1"], "authors_source": "OpenReview API", "abstract": "Mapping biological mechanisms in cellular systems is a fundamental step in early-stage drug discovery that serves to generate hypotheses on what disease-relevant molecular targets may effectively be modulated by pharmacological interventions. With the advent of high-throughput methods for measuring single-cell gene expression under genetic perturbations, we now have effective means for generating evidence for causal gene-gene interactions at scale. However, inferring graphical networks of the size typically encountered in real-world gene-gene interaction networks is difficult in terms of both achieving and evaluating faithfulness to the true underlying causal graph. Moreover, standardised benchmarks for comparing methods for causal discovery in perturbational single-cell data do not yet exist. Here, we introduce CausalBench - a comprehensive benchmark suite for evaluating network inference methods on large-scale perturbational single-cell gene expression data. CausalBench introduces several biologically meaningful performance metrics and operates on two large, curated and openly available benchmark data sets for evaluating methods on the inference of gene regulatory networks from single-cell data generated under perturbations. With real-world datasets consisting of over 200000 training samples under interventions, CausalBench could potentially help facilitate advances in causal network inference by providing what is - to the best of our knowledge - the largest openly available test bed for causal discovery from real-world perturbation data to date.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ojyuTFmbvI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5185/Reviewer_QjTd"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper aims to establish a dataset / benchmark for gene regulatory network inference from single-cell data.", "review_text": "The paper fails in achieving what a good benchmark/dataset paper should demonstrate: clarity, sound assumptions and design choices, and practicality. Reading it did not bring me any closer to being able to quickly, effortlessly evaluate a GRN inference method I may develop.", "strengths": "The paper is hard to follow for me. From the abstract and introduction, it seems it aims at providing a dataset / benchmark for GRN inference that would be very easy to understand and use by researchers from outside of computational biology or network inference fields - something like DREAM 3/4/5 GRN challenges from a decade ago. However, starting from Section 3, the authors introduce a very specific formal structure (SCM) that also does not necessarily reflect the underlying biology (e.g. the DAG assumption) - it is not clear what purpose this serves. The choice of benchmark evaluation metrics (Section 4) seems somewhat ad-hoc (why STRING and not other network repository, why use just true positives instead of metrics based on precision-recall and ROC used in DREAM). On the other hand, the paper lacks details about how to actually use the dataset in practice, whether there is a clear leaderboard etc. - details that would be essential in establishing CausalBench as the benchmark for the field. It does not help that the github repository mentioned in the introduction is empty.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper aims to establish a dataset / benchmark for gene regulatory network inference from single-cell data.", "strength_and_weaknesses": "The paper is hard to follow for me. From the abstract and introduction, it seems it aims at providing a dataset / benchmark for GRN inference that would be very easy to understand and use by researchers from outside of computational biology or network inference fields - something like DREAM 3/4/5 GRN challenges from a decade ago. However, starting from Section 3, the authors introduce a very specific formal structure (SCM) that also does not necessarily reflect the underlying biology (e.g. the DAG assumption) - it is not clear what purpose this serves. The choice of benchmark evaluation metrics (Section 4) seems somewhat ad-hoc (why STRING and not other network repository, why use just true positives instead of metrics based on precision-recall and ROC used in DREAM). On the other hand, the paper lacks details about how to actually use the dataset in practice, whether there is a clear leaderboard etc. - details that would be essential in establishing CausalBench as the benchmark for the field. It does not help that the github repository mentioned in the introduction is empty.", "clarity,_quality,_novelty_and_reproducibility": "The paper lacks clarity and novelty.", "summary_of_the_review": "The paper fails in achieving what a good benchmark/dataset paper should demonstrate: clarity, sound assumptions and design choices, and practicality. Reading it did not bring me any closer to being able to quickly, effortlessly evaluate a GRN inference method I may develop.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666643200696}, {"id": "TXYbM4jnIy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5185/Reviewer_WpyP"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors of this paper propose a suite of tools for benchmarking network inference on both observational and interventional data single-cell gene expression data.  They run several network inference / causal discovery methods for recovering graphical relationships from scRNAseq data and evaluate the performance of the methods using existing putatative knowledge about specific gene-gene interactions as well as by comparing performance of observational and interventional data.  ", "review_text": "Overall, I think the authors take on a challenging and important task of collating multiple datasets and implementing many causal discovery methods on these datasets.  Overall, I think the paper lacks clarity and justification around the statistical benchmarks employed.  As such, it's unclear how much value is actually added in the suite of benchmarks proposed.", "strengths": "- I think the paper tackles an important problem and provides a great service by colating these datasets and metrics.  \n- My main complaint is that the paper advertises it's contribution as providing a benchmark for network inference on large scale scRNA seq data although they admit that there are no gold standard datasets available from which to actually benchmark these methods.  This is a noted limitation, but my concern is that this is a really big limitation.  It's really hard to know whether the results are biologically meaningful, or really just noisy and reflect the difficulty of identifying useful benchmarks.\n- Along these lines, I think the \"Metrics\" section needs a lot more clarification.  The authors refer to equality tests on \"the two distributions\" as well as Wasserstein distance on \"the two empirical distributions\".  What are these distributions of? There could be a lot more detail here regarding the data / statistics used to create these distributions and how the tests were calculated.\n- While I understand some of the difficulties, it's hard to know how much to read into the TPR vs TP counts graphs.  The authors need to explicitly discuss the potential for false positives / false positive rates.  Are there any ways to identify putative false positives that can be used as a pseudo-benchmark?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors of this paper propose a suite of tools for benchmarking network inference on both observational and interventional data single-cell gene expression data.  They run several network inference / causal discovery methods for recovering graphical relationships from scRNAseq data and evaluate the performance of the methods using existing putatative knowledge about specific gene-gene interactions as well as by comparing performance of observational and interventional data.  ", "strength_and_weaknesses": "- I think the paper tackles an important problem and provides a great service by colating these datasets and metrics.  \n- My main complaint is that the paper advertises it's contribution as providing a benchmark for network inference on large scale scRNA seq data although they admit that there are no gold standard datasets available from which to actually benchmark these methods.  This is a noted limitation, but my concern is that this is a really big limitation.  It's really hard to know whether the results are biologically meaningful, or really just noisy and reflect the difficulty of identifying useful benchmarks.\n- Along these lines, I think the \"Metrics\" section needs a lot more clarification.  The authors refer to equality tests on \"the two distributions\" as well as Wasserstein distance on \"the two empirical distributions\".  What are these distributions of? There could be a lot more detail here regarding the data / statistics used to create these distributions and how the tests were calculated.\n- While I understand some of the difficulties, it's hard to know how much to read into the TPR vs TP counts graphs.  The authors need to explicitly discuss the potential for false positives / false positive rates.  Are there any ways to identify putative false positives that can be used as a pseudo-benchmark?\n\n", "clarity,_quality,_novelty_and_reproducibility": "- The authors claim to have all code available in a github repository, but the repository seems to be empty.\n- Overall, I think the paper lacks technical precision needed for publication.  There are many cases where wordy but vague descriptions are used in place of more precise statements.  E.g. \"satisfactory results are only obtained on the RPE1 dataset\".  What counts as satisfactory?  What is the reason the other results are unsatisfactory? At the same there is additional and unneeded notation and technical details regarding SCM.  In my opinion, this doesn't add much, since you are not describing a new SCM-based method and it is never referenced again.  Thus the background on SCMs seems superflous.\n- Re-check for typos and language, e.g. \"appraoch\"\n", "summary_of_the_review": "Overall, I think the authors take on a challenging and important task of collating multiple datasets and implementing many causal discovery methods on these datasets.  Overall, I think the paper lacks clarity and justification around the statistical benchmarks employed.  As such, it's unclear how much value is actually added in the suite of benchmarks proposed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666632486456}, {"id": "WPuxOLVtXz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5185/Reviewer_vpJf"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present CausalBench, a framework to benchmark causal gene regulatory network (GRN) inference methods on perturbational single-cell RNA sequencing data (scRNA-seq). It includes evaluation metrics, baseline implementations of relevant inference methods and access to perturbational scRNA-seq data. With CausalBench the authors evaluate the ability of recovering “silver-standard” ground truth networks and how they are affected by sample size.", "review_text": "The author's contributions to the GRN evaluation field are somewhat new but some aspects already exist in previous work. Moreover, the authors are missing some previous contributions, limiting the added cumulative value of it. On the other hand, the description of the problem, its assumptions and the description of the methods is overall clear and of good quality.\n", "strengths": "Strengths:\n- Authors are aware of and nicely describe the assumptions and limitations of this evaluation task and data.\n- Including both biological and statistical metrics for evaluation of method performance.\n- Evaluation frameworks allow to easily test if novel methods improve over baselines, motivating the development of improved methods.\n\nWeaknesses:\n- Authors propose novel evaluation metrics but did not implement classical ones such as the comparison against ChIP-seq derived networks, see (Pratapa et al. 2020).\n- Authors miss to evaluate classic methods, such as SCENIC (Aibar et al. 2017), that have been shown to be the top performers in other benchmarks (Pratapa et al. 2020). Although these methods are not causal, it would be beneficial to assess if the modeling of causality actually improves GRN inference.\n- While the authors provide different metrics, it is still not clear which methods perform better than others consistently. A “consensus” score based on rankings would improve interpretation. Moreover, there is no discussion about which methods overperform the others and why.\n- Although the authors are aware of the limitation of perturbation experiments, no quality control assessment is performed to test whether the perturbation actually worked. Samples where the perturbed gene still shows high levels of gene expression should be removed or at least accounted for.\n- The partition of the data due to scalability issues raises concerns since, as the authors already mention, it breaks the no-latent-confounder assumption. Authors chose partition sizes such that the running time remains below 30 hours but do not mention anywhere the actual numbers for each method. In order to comply with this rule, some methods might use fractions of the feature space that are too small to generate valuable graphs.\n- In figure 3, authors mention “significant” increase but do not perform any statistical test to corroborate it.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors present CausalBench, a framework to benchmark causal gene regulatory network (GRN) inference methods on perturbational single-cell RNA sequencing data (scRNA-seq). It includes evaluation metrics, baseline implementations of relevant inference methods and access to perturbational scRNA-seq data. With CausalBench the authors evaluate the ability of recovering “silver-standard” ground truth networks and how they are affected by sample size.", "strength_and_weaknesses": "Strengths:\n- Authors are aware of and nicely describe the assumptions and limitations of this evaluation task and data.\n- Including both biological and statistical metrics for evaluation of method performance.\n- Evaluation frameworks allow to easily test if novel methods improve over baselines, motivating the development of improved methods.\n\nWeaknesses:\n- Authors propose novel evaluation metrics but did not implement classical ones such as the comparison against ChIP-seq derived networks, see (Pratapa et al. 2020).\n- Authors miss to evaluate classic methods, such as SCENIC (Aibar et al. 2017), that have been shown to be the top performers in other benchmarks (Pratapa et al. 2020). Although these methods are not causal, it would be beneficial to assess if the modeling of causality actually improves GRN inference.\n- While the authors provide different metrics, it is still not clear which methods perform better than others consistently. A “consensus” score based on rankings would improve interpretation. Moreover, there is no discussion about which methods overperform the others and why.\n- Although the authors are aware of the limitation of perturbation experiments, no quality control assessment is performed to test whether the perturbation actually worked. Samples where the perturbed gene still shows high levels of gene expression should be removed or at least accounted for.\n- The partition of the data due to scalability issues raises concerns since, as the authors already mention, it breaks the no-latent-confounder assumption. Authors chose partition sizes such that the running time remains below 30 hours but do not mention anywhere the actual numbers for each method. In order to comply with this rule, some methods might use fractions of the feature space that are too small to generate valuable graphs.\n- In figure 3, authors mention “significant” increase but do not perform any statistical test to corroborate it.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The presented work is of good quality and is overall clearly explained, allowing reproducibility of the results. Regarding originality, some of the metrics that the authors are proposing are novel but others have been previously used. Moreover, while it is good to propose novel metrics, the authors miss to include some of the classical ones, which would make comparisons to previous work easier.\n", "summary_of_the_review": "The author's contributions to the GRN evaluation field are somewhat new but some aspects already exist in previous work. Moreover, the authors are missing some previous contributions, limiting the added cumulative value of it. On the other hand, the description of the problem, its assumptions and the description of the methods is overall clear and of good quality.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666623845821}, {"id": "lJ8u_ek0mX1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5185/Reviewer_2A85"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces CausalBench, a framework to benchmark gene regulatory inference methods using perturbational scRNAseq data. Although having access to perturbational scRNA data can help in discovering causal gene networks, generating large-scale perturbational data where every gene undergoes some genetic intervention (knockdown) is very cost prohibitive. There is a recent study that generates genome-wide perturbational scRNA data in only two cell types. The authors of CausalBench have used these data to benchmark some of the well-known network inference methods which use observational and interventional data. For the evaluation, they propose two approaches: biological and quantitative. For the biological evaluation, they use known gene-gene or protein-protein interaction data, and for the quantitative, they use some statistical tests to validate the effects of interventions. ", "review_text": "Causal gene network inference (DAGs) are very important in understanding the mechanisms of gene expression. However, learning such DAGs using only observational data is not possible. Recent Perturb-seq data provide helpful interventional data by genetically perturbing the target genes and measuring the expression of other genes. This data empowers the causal DAG learning methods, however, they should be tested and benchmarked via the same metrics to measure their efficacy. \n\nThis paper takes the first step in the right direction by noticing the need for benchmarking causal network inference methods using Perturb-seq data. However, the benchmarking metrics are not comprehensive enough to be certain about the efficacy of such methods. My comments:\n\n- Biological metric does not make so much sense to me. The underlying DAG driving the gene expression programs are very cell-type specific. So, the general gene/protein networks cannot be a good and reasonable metric to evaluate the correctness of the learned DAGs in K562 and RPE1 cells. \n\n- As a qualitative metric, the authors can use held out perturbations and see how well different methods can predict the effect of each perturbation on various genes. This has also been used in Lopez et al. (2022) (cited in the paper). \n\n- Why haven't the authors used the biggest Perturb-seq (2 M cells) in K562 cells as published in (Replogle et al., 2022)? It would be interesting to see how well/fast each method works in a very large dataset.\n\n- Related to the previous comment, I think one another and yet interesting benchmarking metric would be how fast each method could be trained in each dataset. This kind of benchmarking has also been done in Lopez et al. (2022). By growing sequencing technology, it is not so far that we will have access to millions of cells and thousands of perturbations genome-wide, and the methods that could handle this huge computational barrier would be the winners. \n\nMinor commets:\n\nTypos:   - we we observed --> we observed\n             - in not large --> is not large", "strengths": "Strength:\n\n- This paper tries to take the first step towards benchmarking the causal gene regulatory network inference methods using large-scale perturbational scRNA-seq data, which is interesting and important for causal discovery in biology.\n\nWeakness:\n\n- The idea of having a framework to test the causal network inference methods is great if the benchmarking metrics are powerful and interesting enough. I think the metrics proposed in this paper are not that much interesting. For the biological metrics, as mentioned by the authors, the well-known gene-gene or protein-protein interaction data are not specific for the cell types that they are working. The underlying gene regulation rules are very cell-type specific and the inferred networks could not be reliably evaluated by general networks which have been derived from many cell types. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces CausalBench, a framework to benchmark gene regulatory inference methods using perturbational scRNAseq data. Although having access to perturbational scRNA data can help in discovering causal gene networks, generating large-scale perturbational data where every gene undergoes some genetic intervention (knockdown) is very cost prohibitive. There is a recent study that generates genome-wide perturbational scRNA data in only two cell types. The authors of CausalBench have used these data to benchmark some of the well-known network inference methods which use observational and interventional data. For the evaluation, they propose two approaches: biological and quantitative. For the biological evaluation, they use known gene-gene or protein-protein interaction data, and for the quantitative, they use some statistical tests to validate the effects of interventions. ", "strength_and_weaknesses": "Strength:\n\n- This paper tries to take the first step towards benchmarking the causal gene regulatory network inference methods using large-scale perturbational scRNA-seq data, which is interesting and important for causal discovery in biology.\n\nWeakness:\n\n- The idea of having a framework to test the causal network inference methods is great if the benchmarking metrics are powerful and interesting enough. I think the metrics proposed in this paper are not that much interesting. For the biological metrics, as mentioned by the authors, the well-known gene-gene or protein-protein interaction data are not specific for the cell types that they are working. The underlying gene regulation rules are very cell-type specific and the inferred networks could not be reliably evaluated by general networks which have been derived from many cell types. ", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and easy to read and understand. The novelty is only limited to introducing a a benchmarking bed for causal gene network inference methods using perturbational scRNAseq data.", "summary_of_the_review": "Causal gene network inference (DAGs) are very important in understanding the mechanisms of gene expression. However, learning such DAGs using only observational data is not possible. Recent Perturb-seq data provide helpful interventional data by genetically perturbing the target genes and measuring the expression of other genes. This data empowers the causal DAG learning methods, however, they should be tested and benchmarked via the same metrics to measure their efficacy. \n\nThis paper takes the first step in the right direction by noticing the need for benchmarking causal network inference methods using Perturb-seq data. However, the benchmarking metrics are not comprehensive enough to be certain about the efficacy of such methods. My comments:\n\n- Biological metric does not make so much sense to me. The underlying DAG driving the gene expression programs are very cell-type specific. So, the general gene/protein networks cannot be a good and reasonable metric to evaluate the correctness of the learned DAGs in K562 and RPE1 cells. \n\n- As a qualitative metric, the authors can use held out perturbations and see how well different methods can predict the effect of each perturbation on various genes. This has also been used in Lopez et al. (2022) (cited in the paper). \n\n- Why haven't the authors used the biggest Perturb-seq (2 M cells) in K562 cells as published in (Replogle et al., 2022)? It would be interesting to see how well/fast each method works in a very large dataset.\n\n- Related to the previous comment, I think one another and yet interesting benchmarking metric would be how fast each method could be trained in each dataset. This kind of benchmarking has also been done in Lopez et al. (2022). By growing sequencing technology, it is not so far that we will have access to millions of cells and thousands of perturbations genome-wide, and the methods that could handle this huge computational barrier would be the winners. \n\nMinor commets:\n\nTypos:   - we we observed --> we observed\n             - in not large --> is not large", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666393643853}], "openreview_url": "https://openreview.net/forum?id=fHT8kZcyyT", "arxiv_id": "2210.17283", "paper_pdf": "papers/fHT8kZcyyT.pdf", "paper_pdf_sha256": "3d4254927a0e2599da21a008904d89d4bc01e5149ce7eb90930aeb516d46e57f", "paper_pdf_bytes": 442282, "paper_pdf_source": "openreview", "code_url": "https://github.com/causalbench/causalbench-starter", "code_repository": "causalbench/causalbench-starter", "code_commit": "74a26e31df85ac844960d2b1a2f883eea0b155e9", "code_archive": "repos/fHT8kZcyyT.zip", "code_archive_sha256": "0ab2703f074b319cb1a26cfe53d19a265c3024b95415a3feee179db52cc755cc", "code_archive_bytes": 239268, "code_file_count": 4, "code_extensions": {".py": 3, ".sh": 1}, "github_disk_usage_kb": 243, "github_languages": {"Python": 19878, "Shell": 1266, "Dockerfile": 71}, "github_archived": false, "github_pushed_at": "2023-05-17T12:14:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/causalbench-a-large-scale-benchmark-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SbV8J9JHb6", "year": 2022, "status": "rejected", "title": "Soteria: In search of efficient neural networks for private inference", "authors": ["Anshul Aggarwal", "Trevor E Carlson", "Reza Shokri", "Shruti Tople"], "authorids": ["anshul@comp.nus.edu.sg", "~Trevor_E_Carlson1", "~Reza_Shokri1", "~Shruti_Tople2"], "authors_source": "OpenReview API", "abstract": "In the context of ML as a service, our objective is to protect the confidentiality of the users’ queries and the server's model parameters, with modest computation and communication overhead. Prior solutions primarily propose fine-tuning cryptographic methods to make them efficient for known fixed model architectures. The drawback with this line of approach is that the model itself is never designed to efficiently operate with existing cryptographic computations. We observe that the network architecture, internal functions, and parameters of a model, which are all chosen during training, significantly influence the computation and communication overhead of a cryptographic method, during inference.Thus, we propose SOTERIA — a training method to construct model architectures that are by-design efficient for private inference. We use neural architecture search algorithms with the dual objective of optimizing the accuracy of the model and the overhead of using cryptographic primitives for secure inference. Given the flexibility of modifying a model during training, we find accurate models that are also efficient for private computation. We select garbled circuits as our underlying cryptographic primitive, due to their expressiveness and efficiency. We empirically evaluate SOTERIA on MNIST and CIFAR10 datasets, to compare with the prior work on secure inference. Our results confirm that SOTERIA is indeed effective in balancing performance and accuracy.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "8Gm4fDhXGdR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3154/Reviewer_zXdE"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper develops a method for private neural-network-inference via utilization of Yao's garbled circuits (GC) protocol. In order to keep the computation complexity manageable - the main practical hurdle for  GC - the paper proposes utilization of a neural network architecture search, coupled with restricting weights to ternary alphabet and binary activation. The neural architecture search is conducted via a variation of the \"DARTS\" approach of Liu et. al (ICLR 2019), that accounts for model complexity in the cost function. The paper performs extensive experiments against comparable protocols in their evaluation for MNIST and CIFAR10, and show improvements along some factors.  ", "review_text": "\nThe strength of the paper is the seemingly careful design that combines three elements (1) secure MPC protocol choice that reduces communication rounds, (2) restriction of weights to binary/ternary alphabet, and (2) compatabile neural architecture search to enable efficient private inference.  \n\nThe main drawback of the paper is the lack of clarity in the exposition that makes the paper overall difficult to evaluate. The costs and drawbacks of the restricted architectures/weights compared to - say, a baseline with no privacy - is also not reported.\n Specifically:\n- The neural architecture search is not explained rigorously via mathematics, only via words. This leaves the reader guessing as to what the authors are actually performing. Among many aspects where I am unclear, I am unable to understand how the cost functions change given that the architecture is now restricted to discrete set of weights and biases, as the overall training becomes a discrete optimization problem - so I expect the complexity of finding these neural networks and the weights to be enormous \n- How is the factor lambda used in the cost function for the architecture search. The precise specification of the role of this factor is missing.\n- It seems that the data set is utilized for both training the architecture, and then training the model. Intuitively, this must translate to more data being required as compared to training the model itself. The paper must describe this, and perhaps demonstrate any such effects through empirical results.\n- The restriction to ternary weights and binary activation seems to be a significant drawback, and one expects that these restrictions must lead to weaker accuracy for some settings (data sets, or predictions). The paper misses a detailed discussion and experimental comparison in this regard. \n\nIn addition, it appears that the technical contribution builds on prior works, and so the novelty is relatively limited\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper develops a method for private neural-network-inference via utilization of Yao's garbled circuits (GC) protocol. In order to keep the computation complexity manageable - the main practical hurdle for  GC - the paper proposes utilization of a neural network architecture search, coupled with restricting weights to ternary alphabet and binary activation. The neural architecture search is conducted via a variation of the \"DARTS\" approach of Liu et. al (ICLR 2019), that accounts for model complexity in the cost function. The paper performs extensive experiments against comparable protocols in their evaluation for MNIST and CIFAR10, and show improvements along some factors.  ", "main_review": "\nThe strength of the paper is the seemingly careful design that combines three elements (1) secure MPC protocol choice that reduces communication rounds, (2) restriction of weights to binary/ternary alphabet, and (2) compatabile neural architecture search to enable efficient private inference.  \n\nThe main drawback of the paper is the lack of clarity in the exposition that makes the paper overall difficult to evaluate. The costs and drawbacks of the restricted architectures/weights compared to - say, a baseline with no privacy - is also not reported.\n Specifically:\n- The neural architecture search is not explained rigorously via mathematics, only via words. This leaves the reader guessing as to what the authors are actually performing. Among many aspects where I am unclear, I am unable to understand how the cost functions change given that the architecture is now restricted to discrete set of weights and biases, as the overall training becomes a discrete optimization problem - so I expect the complexity of finding these neural networks and the weights to be enormous \n- How is the factor lambda used in the cost function for the architecture search. The precise specification of the role of this factor is missing.\n- It seems that the data set is utilized for both training the architecture, and then training the model. Intuitively, this must translate to more data being required as compared to training the model itself. The paper must describe this, and perhaps demonstrate any such effects through empirical results.\n- The restriction to ternary weights and binary activation seems to be a significant drawback, and one expects that these restrictions must lead to weaker accuracy for some settings (data sets, or predictions). The paper misses a detailed discussion and experimental comparison in this regard. \n\nIn addition, it appears that the technical contribution builds on prior works, and so the novelty is relatively limited\n", "summary_of_the_review": "Overall, even if the idea behind the paper is promising, given the unconventional approach to technical description (only words, not a single equation), and the associated lack of clarity and rigor, I recommend rejecting the paper in the current form. \n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636399926048}, {"id": "VW4rPLFVWpD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3154/Reviewer_TcBF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper aims at developing an efficient neural network architecture with reduced computational cost under a cryptographic primitive where data on the client side and model on the server side are kept confidential. The motivation is that the existing works have focused on developing cryptographic techniques in a fixed network architecture and have not considered the neural network optimization perspective to enhance the efficiency of existing cryptographic computations. The paper then leverages the flexibility of the network structure while ensuring comparable accuracy and tries to find an architecture that provides a significant reduction in runtime while incurring negligible impact on prediction accuracy. Garbled circuits are chosen as the cryptographic primitive due to its compatibility with a wide range of computations, including non-linear functions. The paper then investigates how to reduce the private inference cost by incorporating two protocols: first one is neural architecture search for efficient private inference where a neural operation is penalized based on its computation and communication overhead, and the second one is to train models with ternary parameters which allows reducing the number of model parameters by introducing sparsity. Sparsity leads to a reduction in computation and communication overhead. However, to maintain a certain level of accuracy, the network needs to be scaled which incurs an increase in the runtime. Accordingly, a trade-off between efficiency and accuracy has been demonstrated with experiments on the CIFAR and MNIST datasets for varying regularization parameter and scaling factor.", "review_text": "The main strength of the paper is its novelty, increasing the efficiency of cryptographic computations by neural architecture design is an interesting approach. \n\nThe main weakness is that the technical contribution is limited. It leverages the already existing techniques of neural architecture search and ternary networks to reduce the computation and communication overhead. There are also several points that could be clarified. For instance, it is not clear what is the cost of the architecture search process itself, in terms of the runtime. Is this included in the offline runtime in Table 3? Is the search repeated for different values of the regularization factor to find the best one (in terms of the trade-off between test accuracy and inference runtime), or is there a principled way for choosing the regularization factor to avoid this? If it is the former, this could be very costly, since the search should be repeated for various values assigned to the regularization parameter. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper aims at developing an efficient neural network architecture with reduced computational cost under a cryptographic primitive where data on the client side and model on the server side are kept confidential. The motivation is that the existing works have focused on developing cryptographic techniques in a fixed network architecture and have not considered the neural network optimization perspective to enhance the efficiency of existing cryptographic computations. The paper then leverages the flexibility of the network structure while ensuring comparable accuracy and tries to find an architecture that provides a significant reduction in runtime while incurring negligible impact on prediction accuracy. Garbled circuits are chosen as the cryptographic primitive due to its compatibility with a wide range of computations, including non-linear functions. The paper then investigates how to reduce the private inference cost by incorporating two protocols: first one is neural architecture search for efficient private inference where a neural operation is penalized based on its computation and communication overhead, and the second one is to train models with ternary parameters which allows reducing the number of model parameters by introducing sparsity. Sparsity leads to a reduction in computation and communication overhead. However, to maintain a certain level of accuracy, the network needs to be scaled which incurs an increase in the runtime. Accordingly, a trade-off between efficiency and accuracy has been demonstrated with experiments on the CIFAR and MNIST datasets for varying regularization parameter and scaling factor.", "main_review": "The main strength of the paper is its novelty, increasing the efficiency of cryptographic computations by neural architecture design is an interesting approach. \n\nThe main weakness is that the technical contribution is limited. It leverages the already existing techniques of neural architecture search and ternary networks to reduce the computation and communication overhead. There are also several points that could be clarified. For instance, it is not clear what is the cost of the architecture search process itself, in terms of the runtime. Is this included in the offline runtime in Table 3? Is the search repeated for different values of the regularization factor to find the best one (in terms of the trade-off between test accuracy and inference runtime), or is there a principled way for choosing the regularization factor to avoid this? If it is the former, this could be very costly, since the search should be repeated for various values assigned to the regularization parameter. \n", "summary_of_the_review": "Using neural architecture search to increase the efficiency of neural network architectures for private computing is a very interesting direction. Providing more insights on the cost of the search process and the optimization of the key parameters such as the regularization factor would help the reader better assess the benefits of neural architecture search in this context. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635911673667}, {"id": "6A2bJwxPYBN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3154/Reviewer_rwM8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies neural network designs for cryptographically secure inference. The paper uses the existing DARTS neural network architecture search algorithm to automatically a ternary neural network for secure inference based on garbled circuits. ", "review_text": "Pros:\n1. The problem addressed in this paper is of practical importance for many real-world applications.\n2. The challenges and the proposed solutions are well motivated.\n3. The paper is also very well-written and has a nice flow.\n\nCons:\n\n1.\tThe experiment section reads: “Although SOTERIA could leverage any of the available cryptographic primitives or their combination, we select garbled circuits as its main building block to address the confidentiality concern. Garbled circuits (GC) are known to be efficient and allow the generation of constant depth circuits even for non-linear functions. We show that neural network algorithms can be optimized to efficiently execute garbled circuits while achieving high accuracy guarantees.” It is not clear why faster GC+HE (homomorphic encryption) protocol than the pure-GC protocol that is proved by Gazelle [Chiraag Juvekar+, USENIX Security 2018] is not used.\n2.\tMultiple typos: For example, in the abstract, “multi-arty” should be “multi-party”.\n3.\tThe main contribution of the paper is proposing a neural architecture search to automatically search for an optimal ternary neural network architecture. As such, the paper has a limited novelty from the ML perspective.\n4.\tThe proposed method has shown improved performance over a few recent algorithms. The paper, however, fails to compare with CrypoNAS [Zahra Ghodsi+, NeurIPS 2020], which I think is the SOTA for cryptographic secure inference. It is better to compare more related works. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies neural network designs for cryptographically secure inference. The paper uses the existing DARTS neural network architecture search algorithm to automatically a ternary neural network for secure inference based on garbled circuits. ", "main_review": "Pros:\n1. The problem addressed in this paper is of practical importance for many real-world applications.\n2. The challenges and the proposed solutions are well motivated.\n3. The paper is also very well-written and has a nice flow.\n\nCons:\n\n1.\tThe experiment section reads: “Although SOTERIA could leverage any of the available cryptographic primitives or their combination, we select garbled circuits as its main building block to address the confidentiality concern. Garbled circuits (GC) are known to be efficient and allow the generation of constant depth circuits even for non-linear functions. We show that neural network algorithms can be optimized to efficiently execute garbled circuits while achieving high accuracy guarantees.” It is not clear why faster GC+HE (homomorphic encryption) protocol than the pure-GC protocol that is proved by Gazelle [Chiraag Juvekar+, USENIX Security 2018] is not used.\n2.\tMultiple typos: For example, in the abstract, “multi-arty” should be “multi-party”.\n3.\tThe main contribution of the paper is proposing a neural architecture search to automatically search for an optimal ternary neural network architecture. As such, the paper has a limited novelty from the ML perspective.\n4.\tThe proposed method has shown improved performance over a few recent algorithms. The paper, however, fails to compare with CrypoNAS [Zahra Ghodsi+, NeurIPS 2020], which I think is the SOTA for cryptographic secure inference. It is better to compare more related works. \n", "summary_of_the_review": "--", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635906041666}], "openreview_url": "https://openreview.net/forum?id=SbV8J9JHb6", "arxiv_id": "2007.12934", "paper_pdf": "papers/SbV8J9JHb6.pdf", "paper_pdf_sha256": "79b376a3900ad37b049427431cca14b7c9a74eadcf5a533325a83a81bb92a394", "paper_pdf_bytes": 219608, "paper_pdf_source": "openreview", "code_url": "https://github.com/privacytrustlab/soteria_private_nn_inference", "code_repository": "privacytrustlab/soteria_private_nn_inference", "code_commit": "4cf28c41f0c73c23bb026cc80a70ed41ee20da97", "code_archive": "repos/SbV8J9JHb6.zip", "code_archive_sha256": "aea2a1735f07a9941937d754617066c6f80711b9abbb26094cbef9acfdbffc8b", "code_archive_bytes": 611948, "code_file_count": 32, "code_extensions": {".py": 31, ".sh": 1}, "github_disk_usage_kb": 611, "github_languages": {"Verilog": 3296657, "SuperCollider": 319125, "SystemVerilog": 210024, "Python": 117109, "Dockerfile": 1417, "Shell": 88}, "github_archived": false, "github_pushed_at": "2020-09-29T15:15:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/soteria-in-search-of-efficient-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WdOCkf4aCM", "year": 2021, "status": "rejected", "title": "CDT: Cascading Decision Trees for Explainable Reinforcement Learning", "authors": ["Zihan Ding", "Pablo Hernandez-Leal", "Gavin Weiguang Ding", "Changjian Li", "Ruitong Huang"], "authorids": ["~Zihan_Ding1", "~Pablo_Hernandez-Leal2", "~Gavin_Weiguang_Ding1", "~Changjian_Li1", "~Ruitong_Huang1"], "authors_source": "OpenReview API", "abstract": "Deep Reinforcement Learning (DRL) has recently achieved significant advances in various domains. However, explaining the policy of RL agents still remains an open problem due to several factors, one being the complexity of explaining neural networks decisions. Recently, a group of works have used decision-tree-based models to learn explainable policies. Soft decision trees (SDTs) and discretized differentiable decision trees (DDTs) have been demonstrated to achieve both good performance and share the benefit of having explainable policies. In this work, we further improve the results for tree-based explainable RL in both performance and explainability. Our proposal, Cascading Decision Trees (CDTs) apply representation learning on the decision path to allow richer expressivity. Empirical results show that in both situations, where CDTs are used as policy function approximators or as imitation learners to explain black-box policies, CDTs can achieve better performances with more succinct and explainable models than SDTs. As a second contribution our study reveals limitations of explaining black-box policies via imitation learning with tree-based explainable models, due to its inherent instability.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "yWCvWXbpzHC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1733/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes cascading decision trees (CDTs), which applies representation learning on the decision path of trees, and applies it to explain reinforcement learning models. The main empirical results are that CDTs are claimed to achieve better performance with more compact models than SDTs (soft decision trees), and CDTs as a post-hoc explainer of RL models using imitation learning is unstable.\n\nThe results don't seem fully convincing yet. The paper only compares to SDT, although the references mention several other tree-based methods to explain RL models. VIPER would be a nice baseline to add for the imitation learning task, and ANT would also be a nice baseline in general. \n\nSome important results are not yet reported. For example, what is the performance of the black-box model on the imitation learning task? This would contextualize the accuracy numbers of SDT and CDT. Also, discretized SDT has around 50% accuracy while undiscretized SDT has around 94% accuracy, which is a surprising gap. What is the accuracy of a vanilla, hard decision tree on this problem? In Figure 4, CDT doesn't seem much better than SDT for the Lunar Lander problem. Do the authors have any explanation for why this is the case? Perhaps it was mentioned in the paper but I missed it.\n\nBelow are some suggested improvements to the presentation of the results: \n- For Table 1, perhaps separate out the bracketed results into a different row, e.g. CDT with discretization for the feature learning tree vs. CDTs with only discretization for the decision making tree and discretization for both sub-trees. In fact, maybe 2 columns for SDT — SDT not discretized, SDT discretized — followed by 3 columns for CDT — CDT not discretized, CDT discretization for decision making tree, CDT discretization for both sub-trees. And then make 3 rows: accuracy, depth, # parameters. This is a very detailed suggestion but the gist is that the presentation of this table could be much improved. And similarly for Table 2.\n- Why are the Figure 5 cart-pole results distributed over 2 plots? 6 lines on one plot is possible if done with care. Splitting them over 2 plots makes the SDT (Figure 5a) and CDT (Figure 5b) results less comparable, especially when tables are not provided in that section, only figures. The y-axes are also inconsistent, e.g. there is a -300 for the y-axis of Figure 5d but not for 5c, which makes them less comparable.\n\nWhile the trees being proposed may be more compact, are they more interpretable? I found it hard to parse the linear coefficients and retain all of them in my brain. The trees presented in Figures 6-7 had 4 features, but what if you have more features, or one-hot-encoded features? The literature has user studies to test if human subjects can use trees, but this addition of the linear coefficients to the nodes perhaps warrants a user study to see if human subjects can retain that additional information and make use of it. If a user study is not feasible, could the authors comment on how the interpretability (or lack of) of the feature representation part can be quantified? \n\nI found the findings about instability of trees as a post-hoc explainer interesting and appreciate that the authors including several runs of the methods with the same setting (Figures 15-22). However, I found the figures hard to read, and the heatmap in the tree nodes also needed more explanation, which made me further concerned about the interpretability of the feature representation part of the trees.\n\nMinor points:\n- Typo in capitalization in \"In this paper, We propose Cascading Decision Trees\"\n- Citation needed for \"some methods have axis-aligned partitions (univariate decision nodes) with much lower model expressivity\"\n- Some sentences are not precise and perhaps too casual -- “it basically gives a similar solution”, “kind of like an estimated future position” ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The paper proposes cascading decision trees (CDTs), which applies representation learning on the decision path of trees, and applies it to explain reinforcement learning models. The main empirical results are that CDTs are claimed to achieve better performance with more compact models than SDTs (soft decision trees), and CDTs as a post-hoc explainer of RL models using imitation learning is unstable.\n\nThe results don't seem fully convincing yet. The paper only compares to SDT, although the references mention several other tree-based methods to explain RL models. VIPER would be a nice baseline to add for the imitation learning task, and ANT would also be a nice baseline in general. \n\nSome important results are not yet reported. For example, what is the performance of the black-box model on the imitation learning task? This would contextualize the accuracy numbers of SDT and CDT. Also, discretized SDT has around 50% accuracy while undiscretized SDT has around 94% accuracy, which is a surprising gap. What is the accuracy of a vanilla, hard decision tree on this problem? In Figure 4, CDT doesn't seem much better than SDT for the Lunar Lander problem. Do the authors have any explanation for why this is the case? Perhaps it was mentioned in the paper but I missed it.\n\nBelow are some suggested improvements to the presentation of the results: \n- For Table 1, perhaps separate out the bracketed results into a different row, e.g. CDT with discretization for the feature learning tree vs. CDTs with only discretization for the decision making tree and discretization for both sub-trees. In fact, maybe 2 columns for SDT — SDT not discretized, SDT discretized — followed by 3 columns for CDT — CDT not discretized, CDT discretization for decision making tree, CDT discretization for both sub-trees. And then make 3 rows: accuracy, depth, # parameters. This is a very detailed suggestion but the gist is that the presentation of this table could be much improved. And similarly for Table 2.\n- Why are the Figure 5 cart-pole results distributed over 2 plots? 6 lines on one plot is possible if done with care. Splitting them over 2 plots makes the SDT (Figure 5a) and CDT (Figure 5b) results less comparable, especially when tables are not provided in that section, only figures. The y-axes are also inconsistent, e.g. there is a -300 for the y-axis of Figure 5d but not for 5c, which makes them less comparable.\n\nWhile the trees being proposed may be more compact, are they more interpretable? I found it hard to parse the linear coefficients and retain all of them in my brain. The trees presented in Figures 6-7 had 4 features, but what if you have more features, or one-hot-encoded features? The literature has user studies to test if human subjects can use trees, but this addition of the linear coefficients to the nodes perhaps warrants a user study to see if human subjects can retain that additional information and make use of it. If a user study is not feasible, could the authors comment on how the interpretability (or lack of) of the feature representation part can be quantified? \n\nI found the findings about instability of trees as a post-hoc explainer interesting and appreciate that the authors including several runs of the methods with the same setting (Figures 15-22). However, I found the figures hard to read, and the heatmap in the tree nodes also needed more explanation, which made me further concerned about the interpretability of the feature representation part of the trees.\n\nMinor points:\n- Typo in capitalization in \"In this paper, We propose Cascading Decision Trees\"\n- Citation needed for \"some methods have axis-aligned partitions (univariate decision nodes) with much lower model expressivity\"\n- Some sentences are not precise and perhaps too casual -- “it basically gives a similar solution”, “kind of like an estimated future position” ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604056137054}, {"id": "QNXpQMQQXo", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1733/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new cascading decision trees (CDT) for interpretable RL tasks. As an extension to soft decision trees (SDT), CDT adds a feature learning tree before the decision making tree. By utilizing a low-rank matrix model in the feature learning tree, CDT reduces the parameters of SDT. \n\n1. Novelty is limited. The key novelty of this paper is to add a low-rank representation learning, i.e., feature learning tree, before the decision making tree. Therefore, it is expected that the number of parameters of CDT is less than that of SDT. The improved interpretability is in fact mainly due to such representation learning procedure. \n\n2. It is unclear how to select the tree depths of both the feature learning tree and the decision making tree in a data-driven way. Although the authors provide some preliminary experiments on the tree depth in Figure 5, it is unclear why only $1+2$, $2+2$, $3+2$ are considered in CartPole-v1 and $2+2$, $2+3$, $3+2$, $3+3$ are considered in LunarLander-v2. As shown in Table 5 in Appendix G, the feature learning tree depth, the decision making tree depth, and the number of intermediate variables of CDT are different in three examples.\n\n3. The hierarchical CDT is included in Section 3.2. The authors claimed the hierarchical CDT might be able to improve the prediction accuracy while increasing the model capacity. No experimental study was provided on hierarchical CDT. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "CDT review", "review": "This paper introduces a new cascading decision trees (CDT) for interpretable RL tasks. As an extension to soft decision trees (SDT), CDT adds a feature learning tree before the decision making tree. By utilizing a low-rank matrix model in the feature learning tree, CDT reduces the parameters of SDT. \n\n1. Novelty is limited. The key novelty of this paper is to add a low-rank representation learning, i.e., feature learning tree, before the decision making tree. Therefore, it is expected that the number of parameters of CDT is less than that of SDT. The improved interpretability is in fact mainly due to such representation learning procedure. \n\n2. It is unclear how to select the tree depths of both the feature learning tree and the decision making tree in a data-driven way. Although the authors provide some preliminary experiments on the tree depth in Figure 5, it is unclear why only $1+2$, $2+2$, $3+2$ are considered in CartPole-v1 and $2+2$, $2+3$, $3+2$, $3+3$ are considered in LunarLander-v2. As shown in Table 5 in Appendix G, the feature learning tree depth, the decision making tree depth, and the number of intermediate variables of CDT are different in three examples.\n\n3. The hierarchical CDT is included in Section 3.2. The authors claimed the hierarchical CDT might be able to improve the prediction accuracy while increasing the model capacity. No experimental study was provided on hierarchical CDT. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603935886295}, {"id": "TpXh9HUDS6S", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1733/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## After Rebuttal and Discussion Period\nI want to first say that I really appreciated the opportunity to review this paper. It was awesome to see the authors willing to respond to the various suggestions and questions that the reviewers provided. The paper has improved over time but some significant areas of improvement remain. There continues to be some discontinuity between the stated scope of this paper as XRL and the proposed CDT approach. During the rebuttal period, I felt that the authors didn't do enough to soften their claims to match the support provided by the contributions present in the paper through both the methods and experimentation. This is not to say that this is not valuable work. As discussed, the concepts and ideas are strong, I however feel that the execution and scoping of this work is a little off from clearly communicating the contributions it makes. One aspect will be through a full evaluation of the utility of their approach as explainable through user studies or other qualitative means. The claims made in this paper regarding explainability are currently unsupported.\n\nAdditionally, I believe that there is significant room for improvement in the experimental portion of this work. If there were a way to provide some ablations of the CDT approach as additional baselines as well as the SDT/DDT benchmark, it would significantly improve this portion of the paper.\n\nAs it stands, I have not chosen to adjust my score. I do however urge the authors to continue in this line of work. I believe that there is strong merit with the direction they've begun. I look forward to seeing a future completed version of this work.\n\n\n#### **Summary**\nThis paper builds from recent developments in differentiable decision tree approaches to address explainable/interpretable Reinforcement Learning. A primary focus of this paper is on the development of rich feature representations to improve the expressivity of the probabilistic splits in the downstream decision nodes of the tree function approximators. Extensive experiments are run to compare the proposed CDT against a general representation of prior differentiable decision tree approaches in two tasks: imitation learning and online policy development.\n\n#### **Assessment**\nThis paper does a great job introducing the proposed CDT approach in the context of the relevant literature, addressing valid criticisms about prior work. The experiments are extensive, it is clear that a lot of care and thought was put into demonstrating the apparent benefits of CDT. The ideas of cascading improved feature representations to the differentiable decision trees is a reasonable improvement over prior approaches. However, I found the claims of improved explainability to be tenuous at best. Multivariate decision boundaries, and fewer parameters do not generally equate to improved explainability, especially in high dimension input spaces. This challenge of preserving explainability is complicated further by allowing for multiple layers of multivariate decision rules. Even linear functions, when using several variables, lose their ability to be clearly understood after two or three parameters. Traditionally the notion of explainability, especially within RL, is built around the interaction between observed features and the criteria a model uses to arrive at the suggested action. By transforming the decision criteria of the function approximator to be based on a, now uninterpretable, representation of the input, this explainability is no longer preserved. I'd be open to changing my mind if user studies showed the proposed CDT model to be more explainable but the major claims of improved explainability rest on an unsupported conjecture that fewer parameters in linear equations are better. This may have been true in analyzing the results presented in this paper but the experiments are performed over low-dimensional state spaces that have relatively similar oscillatory dynamics admitted by optimal policies. Other weaknesses and suggestions for improvement can be found below.\n\n#### **Strengths**\n- The paper does a great job outlining the prior literature in framing the proposed CDT approach. It is clear where the proposed contributions lie within the space of what has been done before. The idea of improving the feature representation for use in a decision tree model is similar to Konstschieder, et al yet the insistence on maintaining model simplicity is refreshing. I'm not convinced that this automatically equates to model explainability but the effort to maintain simplicity is appreciated.\n- The paper highlights two central tasks that CDT can be used within RL and extensively evaluates its performance against a generalized form of the prior literature (SDT).\n- Within the experimental evaluation the paper evaluates several architectural settings of both CDT and SDT providing a reliable analysis over possible options for the reader to build from.\n\n#### **Weaknesses**\nIn general, I found the technical development in Section 3 to be unclear. The path $\\mathcal{P}$ is described to be a set of nodes yet the only formal definition of $\\mathcal{P}$ is in selecting a single node. Within the arg max over possible nodes, it's not clear how a choice of $u$ affects the product of path probabilities. Overall, the discussion about path probabilities is confusing. Continuing, The development for CDT in Section 3.2 only describes a single layer of decision nodes and doesn't account for multiple layers. The transformation matrix $T_{K\\times R}$ is only defined for a single decision node. Other points that are unclear throughout section 3 are:\n- In the paragraph following Equation 5, it is written: \"During the inference process, we simply take the leaf on $\\mathcal{F}$ or $\\mathcal{D}$ with the largest probability...\" Aren't these path probabilities input dependent? Don't they vary stochastically due to the probabilistic definition?\n- After equations 6-7, there is some analysis about the reduction of parameters in  CDT. Yet this is hard to place any significance on because a similar analysis is not provided about standard DTs or SDTs. This comparison is also a little tenuous given the different types of SDTs.\n\nTouching on this last point. As defined, SDT is a class of models with very different configurations. It is probably not appropriate to lump them all together in a generalized setting. Even then, the explicit setting of the SDT used to compare with CDT is never described. Without a clear description of the baselines, the experimental results are difficult to fully interpret and place much confidence in. It's not clear why the authors didn't choose several SDT models to evaluate CDT against. In particular, it appears that Silva, et al [AISTATS; 2020] accomplishes a lot of the same goals as CDT albeit without learning a feature representation. It would've been nice to see a more dedicated comparison between CDT and the specific approaches in the literature. More discussion or justification for the use of a single SDT approach is necessary if the current experimental analysis is going to be the final one.\n\nOn this point, it is unclear why discretizing the decision trees lead to reduced parameter counts as well as a reduction in overall performance. In most of the prior literature using differentiable or soft decision trees, the discretization is done after training to speed up inference as well as secure the splits at the decision nodes to allow for model explainability. It appears that perhaps the \"discretized\" form of each of these evaluated models was indicative of how the models were trained? If so, this is a deviation from how the baseline methods were actually developed, leading to an inaccurate comparison. Further clarity about what is mean by discretization here would improve this initial set of experiments greatly. \n- Accuracy of an imitation learned model isn't very indicative of how effective that model is in solving the actual task. While informative, the results presented in Tables 1 and 2 don't really communicate how effective CDT or SDT are in solving the tasks. Average return from executing the imitation learned policies would be a better metric here. This would help separate out the differences between a 94% accurate model and a 91% accurate model. Ultimately, we care about how a policy performs on the task. It would be nice to have that analyzed.\n\n- The definition of stability is never formalized. Stability of what? What is the ideal score or metric for stability? Are we looking for the weight vectors to be equivalent between different random initializations of the same tree architecture? If so, I'm not certain that Equation 8 is appropriate as it's measuring the distance a between weight vectors between arbitrary nodes. There is no relation between where in the tree compared weight vectors are coming from. Also, the distance metric will be heavily influenced by the number of parameters and depth of the tree. This makes it an insufficient comparison between tree settings as is done in Table 3.\n\n- In Section 4.2 The results figures over experimental domain are inconsistent. Figures 4 and 5 show results from Cartpole and LunarLander, omitting Mountain Car while Figures 6 and 7 omit LunarLander. Looking at the appendix, it seems that the learning curves for Mountain Car are not as clearly separable (with a lot of variance) and the decision tree for Lunar Lander is not as interpretable. This feels like these results being less aligned with the desired narrative got swept under the rug, hoping that they wouldn't be looked at.\n\n- While the magnitude of weights within the decision node do certainly help determine feature importance, the weights alone doesn't tell me anything about the value of the input and why it leads to one decision being made over another.\n \n\n#### **Additional Comments**\nOne paper that I felt was overlooked within the great literature review was:\n*Wu, Mike, et al. \"Beyond Sparsity: Tree Regularization of Deep Models for Interpretability.\" AAAI. 2018.*\n\nThis omission doesn't negatively affect my estimation of how the authors framed their work but I do feel that it stands to be mentioned alongside other distillation approaches. Not only does this paper distill a Neural Network into a Decision Tree for interpretability, it also regularizes the neural network based on the complexity of that distilled decision tree, ensuring that the end interpretation of the network is simple yet informative. \n\nAt present, I do not feel as though this paper is ready for publication. I am not confident in the claims being made by the authors with regards to explainability and would suggest that they either fully justify these claims or place more emphasis on the performance improvements made through CDT at the expense of some explainability. Being more honest with the limitations and assumptions being made in the development of a model is always the best way to go. I do however find the idea of cascading feature representations to encourage more expressive decision nodes within the decision tree to be intriguing. Particularly within the RL space. Representation learning within RL is an open problem and I wonder if this simple modeling strategy may highlight unique aspects of learning representations that deep RL is unable to.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An appropriate extension of prior differentiable decision tree approaches. Intriguing concepts in nature but claims of improved explainability are not justified and may not be accurate.", "review": "## After Rebuttal and Discussion Period\nI want to first say that I really appreciated the opportunity to review this paper. It was awesome to see the authors willing to respond to the various suggestions and questions that the reviewers provided. The paper has improved over time but some significant areas of improvement remain. There continues to be some discontinuity between the stated scope of this paper as XRL and the proposed CDT approach. During the rebuttal period, I felt that the authors didn't do enough to soften their claims to match the support provided by the contributions present in the paper through both the methods and experimentation. This is not to say that this is not valuable work. As discussed, the concepts and ideas are strong, I however feel that the execution and scoping of this work is a little off from clearly communicating the contributions it makes. One aspect will be through a full evaluation of the utility of their approach as explainable through user studies or other qualitative means. The claims made in this paper regarding explainability are currently unsupported.\n\nAdditionally, I believe that there is significant room for improvement in the experimental portion of this work. If there were a way to provide some ablations of the CDT approach as additional baselines as well as the SDT/DDT benchmark, it would significantly improve this portion of the paper.\n\nAs it stands, I have not chosen to adjust my score. I do however urge the authors to continue in this line of work. I believe that there is strong merit with the direction they've begun. I look forward to seeing a future completed version of this work.\n\n\n#### **Summary**\nThis paper builds from recent developments in differentiable decision tree approaches to address explainable/interpretable Reinforcement Learning. A primary focus of this paper is on the development of rich feature representations to improve the expressivity of the probabilistic splits in the downstream decision nodes of the tree function approximators. Extensive experiments are run to compare the proposed CDT against a general representation of prior differentiable decision tree approaches in two tasks: imitation learning and online policy development.\n\n#### **Assessment**\nThis paper does a great job introducing the proposed CDT approach in the context of the relevant literature, addressing valid criticisms about prior work. The experiments are extensive, it is clear that a lot of care and thought was put into demonstrating the apparent benefits of CDT. The ideas of cascading improved feature representations to the differentiable decision trees is a reasonable improvement over prior approaches. However, I found the claims of improved explainability to be tenuous at best. Multivariate decision boundaries, and fewer parameters do not generally equate to improved explainability, especially in high dimension input spaces. This challenge of preserving explainability is complicated further by allowing for multiple layers of multivariate decision rules. Even linear functions, when using several variables, lose their ability to be clearly understood after two or three parameters. Traditionally the notion of explainability, especially within RL, is built around the interaction between observed features and the criteria a model uses to arrive at the suggested action. By transforming the decision criteria of the function approximator to be based on a, now uninterpretable, representation of the input, this explainability is no longer preserved. I'd be open to changing my mind if user studies showed the proposed CDT model to be more explainable but the major claims of improved explainability rest on an unsupported conjecture that fewer parameters in linear equations are better. This may have been true in analyzing the results presented in this paper but the experiments are performed over low-dimensional state spaces that have relatively similar oscillatory dynamics admitted by optimal policies. Other weaknesses and suggestions for improvement can be found below.\n\n#### **Strengths**\n- The paper does a great job outlining the prior literature in framing the proposed CDT approach. It is clear where the proposed contributions lie within the space of what has been done before. The idea of improving the feature representation for use in a decision tree model is similar to Konstschieder, et al yet the insistence on maintaining model simplicity is refreshing. I'm not convinced that this automatically equates to model explainability but the effort to maintain simplicity is appreciated.\n- The paper highlights two central tasks that CDT can be used within RL and extensively evaluates its performance against a generalized form of the prior literature (SDT).\n- Within the experimental evaluation the paper evaluates several architectural settings of both CDT and SDT providing a reliable analysis over possible options for the reader to build from.\n\n#### **Weaknesses**\nIn general, I found the technical development in Section 3 to be unclear. The path $\\mathcal{P}$ is described to be a set of nodes yet the only formal definition of $\\mathcal{P}$ is in selecting a single node. Within the arg max over possible nodes, it's not clear how a choice of $u$ affects the product of path probabilities. Overall, the discussion about path probabilities is confusing. Continuing, The development for CDT in Section 3.2 only describes a single layer of decision nodes and doesn't account for multiple layers. The transformation matrix $T_{K\\times R}$ is only defined for a single decision node. Other points that are unclear throughout section 3 are:\n- In the paragraph following Equation 5, it is written: \"During the inference process, we simply take the leaf on $\\mathcal{F}$ or $\\mathcal{D}$ with the largest probability...\" Aren't these path probabilities input dependent? Don't they vary stochastically due to the probabilistic definition?\n- After equations 6-7, there is some analysis about the reduction of parameters in  CDT. Yet this is hard to place any significance on because a similar analysis is not provided about standard DTs or SDTs. This comparison is also a little tenuous given the different types of SDTs.\n\nTouching on this last point. As defined, SDT is a class of models with very different configurations. It is probably not appropriate to lump them all together in a generalized setting. Even then, the explicit setting of the SDT used to compare with CDT is never described. Without a clear description of the baselines, the experimental results are difficult to fully interpret and place much confidence in. It's not clear why the authors didn't choose several SDT models to evaluate CDT against. In particular, it appears that Silva, et al [AISTATS; 2020] accomplishes a lot of the same goals as CDT albeit without learning a feature representation. It would've been nice to see a more dedicated comparison between CDT and the specific approaches in the literature. More discussion or justification for the use of a single SDT approach is necessary if the current experimental analysis is going to be the final one.\n\nOn this point, it is unclear why discretizing the decision trees lead to reduced parameter counts as well as a reduction in overall performance. In most of the prior literature using differentiable or soft decision trees, the discretization is done after training to speed up inference as well as secure the splits at the decision nodes to allow for model explainability. It appears that perhaps the \"discretized\" form of each of these evaluated models was indicative of how the models were trained? If so, this is a deviation from how the baseline methods were actually developed, leading to an inaccurate comparison. Further clarity about what is mean by discretization here would improve this initial set of experiments greatly. \n- Accuracy of an imitation learned model isn't very indicative of how effective that model is in solving the actual task. While informative, the results presented in Tables 1 and 2 don't really communicate how effective CDT or SDT are in solving the tasks. Average return from executing the imitation learned policies would be a better metric here. This would help separate out the differences between a 94% accurate model and a 91% accurate model. Ultimately, we care about how a policy performs on the task. It would be nice to have that analyzed.\n\n- The definition of stability is never formalized. Stability of what? What is the ideal score or metric for stability? Are we looking for the weight vectors to be equivalent between different random initializations of the same tree architecture? If so, I'm not certain that Equation 8 is appropriate as it's measuring the distance a between weight vectors between arbitrary nodes. There is no relation between where in the tree compared weight vectors are coming from. Also, the distance metric will be heavily influenced by the number of parameters and depth of the tree. This makes it an insufficient comparison between tree settings as is done in Table 3.\n\n- In Section 4.2 The results figures over experimental domain are inconsistent. Figures 4 and 5 show results from Cartpole and LunarLander, omitting Mountain Car while Figures 6 and 7 omit LunarLander. Looking at the appendix, it seems that the learning curves for Mountain Car are not as clearly separable (with a lot of variance) and the decision tree for Lunar Lander is not as interpretable. This feels like these results being less aligned with the desired narrative got swept under the rug, hoping that they wouldn't be looked at.\n\n- While the magnitude of weights within the decision node do certainly help determine feature importance, the weights alone doesn't tell me anything about the value of the input and why it leads to one decision being made over another.\n \n\n#### **Additional Comments**\nOne paper that I felt was overlooked within the great literature review was:\n*Wu, Mike, et al. \"Beyond Sparsity: Tree Regularization of Deep Models for Interpretability.\" AAAI. 2018.*\n\nThis omission doesn't negatively affect my estimation of how the authors framed their work but I do feel that it stands to be mentioned alongside other distillation approaches. Not only does this paper distill a Neural Network into a Decision Tree for interpretability, it also regularizes the neural network based on the complexity of that distilled decision tree, ensuring that the end interpretation of the network is simple yet informative. \n\nAt present, I do not feel as though this paper is ready for publication. I am not confident in the claims being made by the authors with regards to explainability and would suggest that they either fully justify these claims or place more emphasis on the performance improvements made through CDT at the expense of some explainability. Being more honest with the limitations and assumptions being made in the development of a model is always the best way to go. I do however find the idea of cascading feature representations to encourage more expressive decision nodes within the decision tree to be intriguing. Particularly within the RL space. Representation learning within RL is an open problem and I wonder if this simple modeling strategy may highlight unique aspects of learning representations that deep RL is unable to.\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603875110593}, {"id": "ih3OXjn8Ipc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1733/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Edit:\n\nI have read the authors' response and the other reviews. I still believe that this paper is not ready for acceptance. \n\nSummary: \n\nThe authors propose to use Cascading Decision Trees (CDTs) to express the policy for an RL agent. The authors describe CDTs and evaluate their use in imitating a trained expert as well as representing a policy during training. \n\n\nReasons for score: \n\nThe interpretability of CDTs is not convincingly demonstrated by the examples provided. CDTs do perform better than the tested SDTs, but the experiments are insufficient to conclude that CDTs perform well compared to other models which are less interpretable than SDTs. \n\n\nPros:\n\n-CDTs are explained well.\n\n-CDTs are shown to solve basic RL environments which are commonly used for evaluation in XRL work. \n\n\nCons:\n\n-The authors argue that linear partitions are superior to axis-aligned partitions based on the number of parameters. However, the authors do not consider the number nor complexity of operations required for a \"forward pass\" through the model. A fairer evaluation of explainability would include the average or worst-case number of operations required to select an action (e.g., 4 multiplications, 4 additions, and 2 comparisons in the worst case for Figure 1a). By using a single D tree for all F trees, the number of parameters is reduced, but the length of any given path is still long (with many operations). \n\n-The authors claim that CDTs are more \"explainable\" than SDTs, but this is not sufficiently demonstrated. One advantage of DTs over MLPs is that all splitting operations occur on the original features. If the original features are interpretable, then the partitioning process operates on meaningful features. When learned features are used (as in CDTs), this property is lost. \n\n-Between the leaf of a feature learning tree and an internal node in the decision tree, the input is multiplied by a set of weights and put through a non-linearity. When this is performed several times in sequence, this begins to resemble a MLP. The authors do acknowledge this potential problem (as motivation for not evaluating hierarchical CDTs), but this concern also applies (to a lesser degree) to the \"single F, single D\" case. \n\n-The authors use \"heuristic agents\" as experts in their experiments. This does not follow the procedure established by prior work, and this is at odds with the motivation of CDTs as useful for explaining RL agents.\n\n-The experimental evaluation is lacking in a number of ways: \n\n--The authors should report policy performance (in imitation learning experiments). Accuracy is a useful metric, but not sufficient on its own. A model can have a high accuracy without learning a well-performing policy. Also, results are reported for a different set of configurations in Table 2 as compared to Table 1 (without any justification). This suggests that discretization of CDTs must be performed in a more nuanced way than otherwise stated in this work.\n\n--The authors do not compare to VIPER in the imitation learning experiments they perform though VIPER was shown to yield higher-performing policies than standard classification-based learning.\n\n--In the RL experiments, the authors compare to a MLP. They find that CDT has a similar (but sometimes worse) final performance despite having fewer parameters. However, the authors should also compare to a smaller MLP, ideally with the same number of parameters as CDT. Without this comparison, conclusions cannot be drawn about the \"parameter vs performance\" benefits of CDT.\n\n-State normalization (based on a \"well-trained policy\") is not standard and generally not feasible when applying RL. It is unclear why this was done. \n\n-The second paragraph on page 8 attempts to explain a learned MountainCar-v0 model. The lack of certainty and vagueness of the insights (\"kind of like an estimated future position or previous position, and makes action decisions based on that\") suggests that additional work is required to make CDTs interpretable. Why are \"future position\" and \"previous position\" both options for this explanation? The environment has two features (position and velocity), so discovering that the agent selects actions based on intermediate features derived from position is a given. Ideally, the authors select a way to measure interpretability and perform a quantitative evaluation. \n\n\nQuestions During Rebuttal Period:\n\nPlease address and clarify the \"Cons\" above.\n\n\nMinor Comments:\n\n-The method's motivation is best placed in the main paper, not in the Appendix (page 3, footnote 2).\n\n-The figures would be more helpful if they appeared on the same page as the corresponding text.\n\n-The caption for Table 1 is not clear with respect to which CDT accuracies are for which discretization schemes.\n\n-The authors note that discretization decreases performance and \"claim that this is a general drawback for tree-based methods in XRL...\" However, this is only applicable to soft DTs. This should be made clear (e.g., VIPER does not have this drawback).\n\n-The y-limits of Figures 5a and 5b should match so that performances can be more readily compared across plots (given that CDT and SDT are not plotted within one figure). The same applies for Figures 5c and 5d.\n\n-The paper would benefit from another editing pass for grammar. \nSome Typos:\n\n-Abstract: \"trees (DDTs) have [been] demonstrated to achieve\"\n\n-Introduction: \"are generally lack[ing] interpretability\"; \"In this paper, [w]e propose\"\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "CDTs are not demonstrated to be interpretable", "review": "Edit:\n\nI have read the authors' response and the other reviews. I still believe that this paper is not ready for acceptance. \n\nSummary: \n\nThe authors propose to use Cascading Decision Trees (CDTs) to express the policy for an RL agent. The authors describe CDTs and evaluate their use in imitating a trained expert as well as representing a policy during training. \n\n\nReasons for score: \n\nThe interpretability of CDTs is not convincingly demonstrated by the examples provided. CDTs do perform better than the tested SDTs, but the experiments are insufficient to conclude that CDTs perform well compared to other models which are less interpretable than SDTs. \n\n\nPros:\n\n-CDTs are explained well.\n\n-CDTs are shown to solve basic RL environments which are commonly used for evaluation in XRL work. \n\n\nCons:\n\n-The authors argue that linear partitions are superior to axis-aligned partitions based on the number of parameters. However, the authors do not consider the number nor complexity of operations required for a \"forward pass\" through the model. A fairer evaluation of explainability would include the average or worst-case number of operations required to select an action (e.g., 4 multiplications, 4 additions, and 2 comparisons in the worst case for Figure 1a). By using a single D tree for all F trees, the number of parameters is reduced, but the length of any given path is still long (with many operations). \n\n-The authors claim that CDTs are more \"explainable\" than SDTs, but this is not sufficiently demonstrated. One advantage of DTs over MLPs is that all splitting operations occur on the original features. If the original features are interpretable, then the partitioning process operates on meaningful features. When learned features are used (as in CDTs), this property is lost. \n\n-Between the leaf of a feature learning tree and an internal node in the decision tree, the input is multiplied by a set of weights and put through a non-linearity. When this is performed several times in sequence, this begins to resemble a MLP. The authors do acknowledge this potential problem (as motivation for not evaluating hierarchical CDTs), but this concern also applies (to a lesser degree) to the \"single F, single D\" case. \n\n-The authors use \"heuristic agents\" as experts in their experiments. This does not follow the procedure established by prior work, and this is at odds with the motivation of CDTs as useful for explaining RL agents.\n\n-The experimental evaluation is lacking in a number of ways: \n\n--The authors should report policy performance (in imitation learning experiments). Accuracy is a useful metric, but not sufficient on its own. A model can have a high accuracy without learning a well-performing policy. Also, results are reported for a different set of configurations in Table 2 as compared to Table 1 (without any justification). This suggests that discretization of CDTs must be performed in a more nuanced way than otherwise stated in this work.\n\n--The authors do not compare to VIPER in the imitation learning experiments they perform though VIPER was shown to yield higher-performing policies than standard classification-based learning.\n\n--In the RL experiments, the authors compare to a MLP. They find that CDT has a similar (but sometimes worse) final performance despite having fewer parameters. However, the authors should also compare to a smaller MLP, ideally with the same number of parameters as CDT. Without this comparison, conclusions cannot be drawn about the \"parameter vs performance\" benefits of CDT.\n\n-State normalization (based on a \"well-trained policy\") is not standard and generally not feasible when applying RL. It is unclear why this was done. \n\n-The second paragraph on page 8 attempts to explain a learned MountainCar-v0 model. The lack of certainty and vagueness of the insights (\"kind of like an estimated future position or previous position, and makes action decisions based on that\") suggests that additional work is required to make CDTs interpretable. Why are \"future position\" and \"previous position\" both options for this explanation? The environment has two features (position and velocity), so discovering that the agent selects actions based on intermediate features derived from position is a given. Ideally, the authors select a way to measure interpretability and perform a quantitative evaluation. \n\n\nQuestions During Rebuttal Period:\n\nPlease address and clarify the \"Cons\" above.\n\n\nMinor Comments:\n\n-The method's motivation is best placed in the main paper, not in the Appendix (page 3, footnote 2).\n\n-The figures would be more helpful if they appeared on the same page as the corresponding text.\n\n-The caption for Table 1 is not clear with respect to which CDT accuracies are for which discretization schemes.\n\n-The authors note that discretization decreases performance and \"claim that this is a general drawback for tree-based methods in XRL...\" However, this is only applicable to soft DTs. This should be made clear (e.g., VIPER does not have this drawback).\n\n-The y-limits of Figures 5a and 5b should match so that performances can be more readily compared across plots (given that CDT and SDT are not plotted within one figure). The same applies for Figures 5c and 5d.\n\n-The paper would benefit from another editing pass for grammar. \nSome Typos:\n\n-Abstract: \"trees (DDTs) have [been] demonstrated to achieve\"\n\n-Introduction: \"are generally lack[ing] interpretability\"; \"In this paper, [w]e propose\"\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603850032320}], "openreview_url": "https://openreview.net/forum?id=WdOCkf4aCM", "arxiv_id": "2011.07553", "paper_pdf": "papers/WdOCkf4aCM.pdf", "paper_pdf_sha256": "38786a72bcb69f815ba586eef61dcd6fed48071f3d0eb098adaf0b40b876d6dc", "paper_pdf_bytes": 6924375, "paper_pdf_source": "openreview", "code_url": "https://github.com/quantumiracle/Cascading-Decision-Tree", "code_repository": "quantumiracle/Cascading-Decision-Tree", "code_commit": "23e123787f502a1613733922cbf6711db0096f1a", "code_archive": "repos/WdOCkf4aCM.zip", "code_archive_sha256": "1423604ccc0efa8e8269c680c36d5c66baf638c499580622c75475066bf30eb7", "code_archive_bytes": 1789082, "code_file_count": 63, "code_extensions": {".py": 46, ".ipynb": 13, ".sh": 4}, "github_disk_usage_kb": 929, "github_languages": {"Jupyter Notebook": 2405577, "Python": 218614, "Shell": 2155}, "github_archived": false, "github_pushed_at": "2025-10-31T18:47:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cdt-cascading-decision-trees-for-explainable-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJe9cR4KvB", "year": 2020, "status": "rejected", "title": "Learning to Contextually Aggregate Multi-Source Supervision for Sequence Labeling", "authors": ["Ouyu Lan*", "Xiao Huang*", "Bill Yuchen Lin", "He Jiang", "Xiang Ren"], "authorids": ["olan@usc.edu", "huan183@usc.edu", "yuchen.lin@usc.edu", "jian567@usc.edu", "xiangren@usc.edu"], "authors_source": "OpenReview API", "abstract": "Sequence labeling is a fundamental framework for various natural language processing problems including part-of-speech tagging and named entity recognition. Its performance is largely influenced by the annotation quality and quantity in supervised learning scenarios.  In many cases, ground truth labels are costly and time-consuming to collect or even non-existent,  while imperfect ones could be easily accessed or transferred from different domains. A typical example is crowd-sourced datasets which have multiple annotations for each sentence which may be noisy or incomplete.   Additionally,  predictions from multiple source models in transfer learning can be seen as a case of multi-source supervision.  In this paper, we propose a novel framework named Consensus Network (CONNET) to conduct training with imperfect annotations from multiple sources.   It learns the representation for every weak supervision source and dynamically aggregates them by a context-aware attention mechanism.  Finally, it leads to a model reflecting the consensus among multiple sources.  We evaluate the proposed framework in two practical settings of multi-source learning:  learning with crowd annotations and unsupervised cross-domain model adaptation. Extensive experimental results show that our model achieves significant improvements over existing methods in both settings.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "B1xyg9XT5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1292/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes a method to learn how to aggregate weak supervision sources in the context of sequence labeling. In particular, the model has two main steps: i) it learns a source dependent transformation and ii) it learns a mechanism to combine them. \n\nIn general, the paper is well organized but it is not easy to follow. It was not clear to me how the authors combine the source dependent representation in Section 4.2 with the aggregation phase presented in Section 4.3. In particular, it is confusing to me how the model combines during the training Eq. 3 and Eq. 5. Similarly, it is not clear to me how in Eq. 6 the model can calculate attention coefficients based only on information from the sentence embedding (h^(i)).\n\nThis work assumes a BLSTM-CRF architecture as a baseline, but it does not explore alternative approaches, such as transformer.   \n\nIn terms of the experiments, authors evaluate the resulting model using two application settings: combining noisy crowd annotations (AMT) and unsupervised cross-domain model adaptation. Results are encouraging, the proposed method is able to outperforms several recent works in terms of F1 metric for the case of AMT and accuracy for the case of cross-domain adaptation. Qualitative results also shows reasonable performance. The supplemental material also include an ablation study.\n\nIn summary, the proposed method is interesting and results seem to be encouraging, however, there are parts of the proposed method that are not clear to me. I rate the paper as weak reject. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review": "This work proposes a method to learn how to aggregate weak supervision sources in the context of sequence labeling. In particular, the model has two main steps: i) it learns a source dependent transformation and ii) it learns a mechanism to combine them. \n\nIn general, the paper is well organized but it is not easy to follow. It was not clear to me how the authors combine the source dependent representation in Section 4.2 with the aggregation phase presented in Section 4.3. In particular, it is confusing to me how the model combines during the training Eq. 3 and Eq. 5. Similarly, it is not clear to me how in Eq. 6 the model can calculate attention coefficients based only on information from the sentence embedding (h^(i)).\n\nThis work assumes a BLSTM-CRF architecture as a baseline, but it does not explore alternative approaches, such as transformer.   \n\nIn terms of the experiments, authors evaluate the resulting model using two application settings: combining noisy crowd annotations (AMT) and unsupervised cross-domain model adaptation. Results are encouraging, the proposed method is able to outperforms several recent works in terms of F1 metric for the case of AMT and accuracy for the case of cross-domain adaptation. Qualitative results also shows reasonable performance. The supplemental material also include an ablation study.\n\nIn summary, the proposed method is interesting and results seem to be encouraging, however, there are parts of the proposed method that are not clear to me. I rate the paper as weak reject. "}, "tcdate": 1572841959203}, {"id": "BJlZKk6pKS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1292/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n## Updated review\n\nI have read the rebuttals. The new version of the paper is clearer and the new baseline experiments are a good addition. \n\n## Original review\n\nThis paper presents an approach to train a neural networks-based model for sequence modelling using labels from different sources. The proposed approach explicitly models the annotator and uses an attention model to select the best aggregation method. The model is used on two scenarios: learning with crowd annotation and cross-domain adaptation. For the first scenario, noisy annotators are simulated with models trained on subsets of the data, the proposed model is compared with related works and is shown to achieve the highest f-1 score. For the second scenario, different domains in three NLP tasks are used and the model is shown to yield the best performance.\n\nI think this paper should be accepted, for the following reasons:\n- The approach is novel as far as I can tell, and the approach of learning to aggregate labels is significant, as it could also be applied to tasks where inter-annotators agreement is a problem.\n- The experiments are convincing and show the potential of the proposed approach. \n- The comparison with related works is thorough.\n\nDetailed comments\n- I don't understand the notion of \"normalized expertise\" in Section 5.4, can the authors briefly describe it in the paper ?\n- The paper is not easy to read, for instance the first paragraph of Section 5 contains critical information to understand the experiments, maybe it should be moved the Section 4 and developed more, typically in two subsections \"Application to crowd annotation\" and \"Application to cross-domain\" for example.\n- Typos:\n    - Section 2, 3rd paragraph \"for traget corpora\" -> \"target\"\n    - Same paragraph: \"Yang & Eisenstien (2018) represented\" -> \"represents\" to be consistent\n    - Section 4.1: \"BiLSTM-CRF\" -> \"BLSTM-CRF\"\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "\n## Updated review\n\nI have read the rebuttals. The new version of the paper is clearer and the new baseline experiments are a good addition. \n\n## Original review\n\nThis paper presents an approach to train a neural networks-based model for sequence modelling using labels from different sources. The proposed approach explicitly models the annotator and uses an attention model to select the best aggregation method. The model is used on two scenarios: learning with crowd annotation and cross-domain adaptation. For the first scenario, noisy annotators are simulated with models trained on subsets of the data, the proposed model is compared with related works and is shown to achieve the highest f-1 score. For the second scenario, different domains in three NLP tasks are used and the model is shown to yield the best performance.\n\nI think this paper should be accepted, for the following reasons:\n- The approach is novel as far as I can tell, and the approach of learning to aggregate labels is significant, as it could also be applied to tasks where inter-annotators agreement is a problem.\n- The experiments are convincing and show the potential of the proposed approach. \n- The comparison with related works is thorough.\n\nDetailed comments\n- I don't understand the notion of \"normalized expertise\" in Section 5.4, can the authors briefly describe it in the paper ?\n- The paper is not easy to read, for instance the first paragraph of Section 5 contains critical information to understand the experiments, maybe it should be moved the Section 4 and developed more, typically in two subsections \"Application to crowd annotation\" and \"Application to cross-domain\" for example.\n- Typos:\n    - Section 2, 3rd paragraph \"for traget corpora\" -> \"target\"\n    - Same paragraph: \"Yang & Eisenstien (2018) represented\" -> \"represents\" to be consistent\n    - Section 4.1: \"BiLSTM-CRF\" -> \"BLSTM-CRF\"\n ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571831673300}, {"id": "Bkg21BVsYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1292/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper proposes ConNet, a new label aggregation method for sequence labeling tasks, including crowd-annotation and cross-domain model adaptation. The model consists of a decoupling phase, which learns annotator-specific transforming matrices A, and an aggregation phase with an attention module. Extensive experimental results demonstrate the superiority of the proposed model over baselines. \nThe paper is generally well-written and easy to follow, and the results seem convincing, so I think it can inspire other works on this topic. My main concern on the paper is its generalization. Crowdsourcing usually involves lots of annotators, and some of them only give very few labels. In these situations, the proposed model introduces lots of new parameters (A, Q), which may cause difficulty during training. So the positive results in Fig 3(b) are very important to dispel my worry, which requires more explanation. \nBelow are some detailed questions:\n-      How do you calculate the sentence embeddings h_i during training?\n-      Have you introduced some regularization terms on A and Q?\n-      What are the most important hyper-parameters for this model, and how to tune?\n-      Can this method be extended to new tasks other than sequence labeling?\n-      Can you compare your method with the aggregation method used in on-the-job learning paper [1]?\n-      92.33 in tab 2 shouldn’t be bold.\n \n[1] Werling K , Chaganty A , Liang P , et al. On-the-Job Learning with Bayesian Decision Theory[J]. Computer Science, 2015.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "\nThis paper proposes ConNet, a new label aggregation method for sequence labeling tasks, including crowd-annotation and cross-domain model adaptation. The model consists of a decoupling phase, which learns annotator-specific transforming matrices A, and an aggregation phase with an attention module. Extensive experimental results demonstrate the superiority of the proposed model over baselines. \nThe paper is generally well-written and easy to follow, and the results seem convincing, so I think it can inspire other works on this topic. My main concern on the paper is its generalization. Crowdsourcing usually involves lots of annotators, and some of them only give very few labels. In these situations, the proposed model introduces lots of new parameters (A, Q), which may cause difficulty during training. So the positive results in Fig 3(b) are very important to dispel my worry, which requires more explanation. \nBelow are some detailed questions:\n-      How do you calculate the sentence embeddings h_i during training?\n-      Have you introduced some regularization terms on A and Q?\n-      What are the most important hyper-parameters for this model, and how to tune?\n-      Can this method be extended to new tasks other than sequence labeling?\n-      Can you compare your method with the aggregation method used in on-the-job learning paper [1]?\n-      92.33 in tab 2 shouldn’t be bold.\n \n[1] Werling K , Chaganty A , Liang P , et al. On-the-Job Learning with Bayesian Decision Theory[J]. Computer Science, 2015.\n"}, "tcdate": 1571665123590}], "openreview_url": "https://openreview.net/forum?id=HJe9cR4KvB", "arxiv_id": "1910.04289", "paper_pdf": "papers/HJe9cR4KvB.pdf", "paper_pdf_sha256": "280f0278ecab706cc0382edbd49ff7ebd080192ba4601488de1561772595ecb8", "paper_pdf_bytes": 2088972, "paper_pdf_source": "openreview", "code_url": "https://github.com/INK-USC/ConNet", "code_repository": "INK-USC/ConNet", "code_commit": "adb299f160556004561df302c19578200bd3835b", "code_archive": "repos/HJe9cR4KvB.zip", "code_archive_sha256": "7d6cda2a1fd28206c3c5bbff552ac4a9748f2c4b383831c1e662cb98425df7f7", "code_archive_bytes": 2878408, "code_file_count": 67, "code_extensions": {".py": 56, ".sh": 11}, "github_disk_usage_kb": 2242, "github_languages": {"Python": 502939, "Shell": 28229}, "github_archived": false, "github_pushed_at": "2020-04-14T05:00:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-contextually-aggregate-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0OUkySEAf0", "year": 2026, "status": "rejected", "title": "DeepTheorem: Advancing LLM Reasoning for Theorem Proving Through Natural Language and Reinforcement Learning", "authors": ["Ziyin Zhang", "Jiahao Xu", "Zhiwei He", "Tian Liang", "Qiuzhi Liu", "Yansi Li", "Linfeng Song", "Zhenwen Liang", "Zhuosheng Zhang", "Rui Wang", "Zhaopeng Tu", "Haitao Mi", "Dong Yu"], "authorids": ["~Ziyin_Zhang2", "~Jiahao_Xu1", "~Zhiwei_He1", "~Tian_Liang4", "~Qiuzhi_Liu1", "~Yansi_Li1", "~Linfeng_Song1", "~Zhenwen_Liang1", "~Zhuosheng_Zhang1", "~Rui_Wang10", "~Zhaopeng_Tu1", "~Haitao_Mi1", "~Dong_Yu2"], "authors_source": "OpenReview API", "abstract": "Theorem proving serves as a major testbed for evaluating complex reasoning abilities in large language models (LLMs). However, traditional automated theorem proving (ATP) approaches rely heavily on formal proof systems that poorly align with LLMs’ strength derived from informal, natural language knowledge acquired during pre-training. In this work, we introduce DeepTheorem, a comprehensive informal theorem-proving suite exploiting natural language to enhance LLM mathematical reasoning. DeepTheorem includes 1) a large-scale dataset of 121K\nhigh-quality IMO-level informal theorems and proofs spanning diverse mathematical domains, rigorously annotated for correctness, difficulty, and topic categories, accompanied by systematically constructed verifiable theorem variants; 2) adaptation of RL-Zero explicitly to informal theorem proving, leveraging the verified theorem variants to incentivize robust mathematical inference; 3) comprehensive outcome and process evaluation metrics examining proof correctness and the quality of reasoning steps; and 4) a new informal theorem proving benchmark consoliadted from three established math competitions, formatted for automatic evaluation. Extensive experimental analyses demonstrate DeepTheorem significantly improves LLM theorem-proving performance compared to existing datasets and supervised fine-tuning protocols, achieving state-of-the-art accuracy and reasoning quality. Our findings highlight DeepTheorem’s potential to fundamentally advance automated informal theorem proving and mathematical exploration.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "EnFjOp6VqS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17319/Reviewer_EGQ7"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces DeepTheorem, a framework for informal theorem proving with large language models that combines a 121K-sample natural-language theorem dataset and reinforcement learning. By adapting RL-Zero to theorem proving with verifiable true/false rewards, the approach enhances reasoning beyond supervised fine-tuning. Evaluations on FIMO, Putnam, and HMMT benchmarks show that DeepTheorem-trained models outperform larger state-of-the-art systems in both proof accuracy and reasoning quality, highlighting the potential of natural-language-based reinforcement learning to advance LLM mathematical reasoning.", "review_text": "The paper introduces DeepTheorem, a framework for informal theorem proving with large language models that combines a 121K-sample natural-language theorem dataset and reinforcement learning. By adapting RL-Zero to theorem proving with verifiable true/false rewards, the approach enhances reasoning beyond supervised fine-tuning. Evaluations on FIMO, Putnam, and HMMT benchmarks show that DeepTheorem-trained models outperform larger state-of-the-art systems in both proof accuracy and reasoning quality, highlighting the potential of natural-language-based reinforcement learning to advance LLM mathematical reasoning.", "strengths": "- Comprehensive experiments across multiple representative model scales.\n- Introduction of a large, well-structured dataset for informal theorem proving.\n- Innovative adaptation of RL-Zero to the informal theorem-proving setting.", "weaknesses": "- Many quality-control steps rely heavily on LLM judgments (e.g., contamination justification and difficulty annotation). A human evaluation of a sampled subset would strengthen confidence in the dataset’s integrity and annotation accuracy.\n- The process evaluation methodology remains somewhat unclear (even with the prompt shown in the appendix). Compared to formal proof verification (e.g., in Lean), LLM-based evaluation is inherently less reliable. A more detailed comparison against prior process-evaluation models, including an analysis of reliability and typical failure modes, would be valuable.\n\nMinor:\n- Figure 2 (left): the formula before ‘is given by’ is not quite right, I guess something like ‘\\pi / 2’ is missing.", "questions": "- How does the dataset compare to Omni-Math?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces DeepTheorem, a framework for informal theorem proving with large language models that combines a 121K-sample natural-language theorem dataset and reinforcement learning. By adapting RL-Zero to theorem proving with verifiable true/false rewards, the approach enhances reasoning beyond supervised fine-tuning. Evaluations on FIMO, Putnam, and HMMT benchmarks show that DeepTheorem-trained models outperform larger state-of-the-art systems in both proof accuracy and reasoning quality, highlighting the potential of natural-language-based reinforcement learning to advance LLM mathematical reasoning.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- Comprehensive experiments across multiple representative model scales.\n- Introduction of a large, well-structured dataset for informal theorem proving.\n- Innovative adaptation of RL-Zero to the informal theorem-proving setting.", "weaknesses": "- Many quality-control steps rely heavily on LLM judgments (e.g., contamination justification and difficulty annotation). A human evaluation of a sampled subset would strengthen confidence in the dataset’s integrity and annotation accuracy.\n- The process evaluation methodology remains somewhat unclear (even with the prompt shown in the appendix). Compared to formal proof verification (e.g., in Lean), LLM-based evaluation is inherently less reliable. A more detailed comparison against prior process-evaluation models, including an analysis of reliability and typical failure modes, would be valuable.\n\nMinor:\n- Figure 2 (left): the formula before ‘is given by’ is not quite right, I guess something like ‘\\pi / 2’ is missing.", "questions": "- How does the dataset compare to Omni-Math?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762625532370}, {"id": "uF77jzs8RB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17319/Reviewer_xMBg"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper presents DeepTheorem, a comprehensive framework for natural language theorem proving comprising four key components: 1) a large-scale dataset of 121K IMO-level informal theorems with high-quality proofs; 2) a novel adaptation of RL-Zero training methodology designed explicitly for informal theorem proving; 3) comprehensive outcome and process evaluation metrics for assessing proof quality; and 4) a newly constructed informal theorem proving benchmark with manually annotated theorem variants. This paper conducts experiments using both SFT and RL-Zero training methods on Qwen-2.5-Base across three sizes. Experiment results indicate that the model trained on DeepTheorem achieves leading performance compared to the baseline dataset. It also shows that the trained model obtains better results than other open-source models of similar size.", "review_text": "This paper presents DeepTheorem, a comprehensive framework for natural language theorem proving comprising four key components: 1) a large-scale dataset of 121K IMO-level informal theorems with high-quality proofs; 2) a novel adaptation of RL-Zero training methodology designed explicitly for informal theorem proving; 3) comprehensive outcome and process evaluation metrics for assessing proof quality; and 4) a newly constructed informal theorem proving benchmark with manually annotated theorem variants. This paper conducts experiments using both SFT and RL-Zero training methods on Qwen-2.5-Base across three sizes. Experiment results indicate that the model trained on DeepTheorem achieves leading performance compared to the baseline dataset. It also shows that the trained model obtains better results than other open-source models of similar size.", "strengths": "1. **Novel idea of verifying NL proofs using LLM:** The paper proposes a logically sound verification method to evaluate natural language theorem proving by following its 4-step criteria. The idea to prove/disprove the variants of the problem and combine all the results together to justify the correctness of the theorem is novel to the field.\n2. **Leading benchmark results on open-source models:** The DeepTheorem dataset trained model achieves leading performance in open-source models and can even compete with large closed-source models. The experiment also thoroughly evaluates the difference between SFT and RL performance on different model sizes across different datasets to prove the effectiveness of DeepTheorem.\n3. **Useful benchmark and dataset proposed:** The benchmark and dataset proposed in this paper are valuable for training more effective models. The scale of the dataset and soundness of the benchmark are all valuable contributions to the field.", "weaknesses": "Despite the above strengths, I have a few concerns regarding this paper:\n\n1. **Unclear motivation and positioning:** This paper’s narrative is confusing regarding its claim of formal theorem proving. While this work focuses on informal theorem proving, the major comparison datasets shown in Figure 1 are DS-Prover-V1 and Lean-WB, both of which are based on Lean4 formal proving rather than informal reasoning. Similarly, the experiment section presents results from DS-Prover, which is not designed for informal reasoning. Besides, this paper also compares with Lean4 reasoning in many parts of the paper; however, the scope of the paper still compares the final outcome results, which differ from real formal theorem proving, where every step is verifiable. This creates confusion about the paper's positioning and contributions in relation to existing work on formal versus informal theorem proving.\n2. **Insufficient experimental analysis:** The experimental evaluation lacks width and depth in the following ways:\n    1. No ablation studies are provided to validate the effectiveness of different components of the proposed dataset (e.g., how training data at different difficulty levels affects model performance, what is the scaling-law effect for the dataset).\n    2. The case study presents only examples without systematic analysis or insights.\n    3. Critical training details are missing, including the exact number of GPU hours and computational costs.\n    4. Given that RL-Zero is known for generating significantly longer responses, the paper should report generation length statistics, but this information is absent.\n3. **Limited evaluation of generalizability:** The paper only evaluates DeepTheorem’s result on Qwen-2.5-Base across different sizes. It does not evaluate on additional model families such as Qwen-2.5-Math, Llama-4, and Qwen-3. It would be necessary to demonstrate the generalizability and robustness of the proposed dataset.\n4. **Presentation quality issues:** Several figures suffer from poor quality or clarity problems. Specifically, Figures 3 and 7 contain blurred text and incomplete labels in the pie charts, making them difficult to interpret.\n\nWith the above weakness, I can only provide a relatively negative overall opinion of the paper. However, I am willing to increase the score if the raised concerns are properly settled.", "questions": "Here are the questions for the authors:\n\n1. What is the connection between the proposed dataset and Lean-WB and DS-Prover-V1’s datasets? Is it necessary to compare them?\n2. What is the ablation study of each component of the proposed method? Are there more case studies and analyses? What is the GPU cost for Zero-RL and SFT training for the models? And what is the generated content length as the training goes on?\n3. How is the effect of DeepTheorem’s performance on SFT/RL on other leading open-source models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents DeepTheorem, a comprehensive framework for natural language theorem proving comprising four key components: 1) a large-scale dataset of 121K IMO-level informal theorems with high-quality proofs; 2) a novel adaptation of RL-Zero training methodology designed explicitly for informal theorem proving; 3) comprehensive outcome and process evaluation metrics for assessing proof quality; and 4) a newly constructed informal theorem proving benchmark with manually annotated theorem variants. This paper conducts experiments using both SFT and RL-Zero training methods on Qwen-2.5-Base across three sizes. Experiment results indicate that the model trained on DeepTheorem achieves leading performance compared to the baseline dataset. It also shows that the trained model obtains better results than other open-source models of similar size.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. **Novel idea of verifying NL proofs using LLM:** The paper proposes a logically sound verification method to evaluate natural language theorem proving by following its 4-step criteria. The idea to prove/disprove the variants of the problem and combine all the results together to justify the correctness of the theorem is novel to the field.\n2. **Leading benchmark results on open-source models:** The DeepTheorem dataset trained model achieves leading performance in open-source models and can even compete with large closed-source models. The experiment also thoroughly evaluates the difference between SFT and RL performance on different model sizes across different datasets to prove the effectiveness of DeepTheorem.\n3. **Useful benchmark and dataset proposed:** The benchmark and dataset proposed in this paper are valuable for training more effective models. The scale of the dataset and soundness of the benchmark are all valuable contributions to the field.", "weaknesses": "Despite the above strengths, I have a few concerns regarding this paper:\n\n1. **Unclear motivation and positioning:** This paper’s narrative is confusing regarding its claim of formal theorem proving. While this work focuses on informal theorem proving, the major comparison datasets shown in Figure 1 are DS-Prover-V1 and Lean-WB, both of which are based on Lean4 formal proving rather than informal reasoning. Similarly, the experiment section presents results from DS-Prover, which is not designed for informal reasoning. Besides, this paper also compares with Lean4 reasoning in many parts of the paper; however, the scope of the paper still compares the final outcome results, which differ from real formal theorem proving, where every step is verifiable. This creates confusion about the paper's positioning and contributions in relation to existing work on formal versus informal theorem proving.\n2. **Insufficient experimental analysis:** The experimental evaluation lacks width and depth in the following ways:\n    1. No ablation studies are provided to validate the effectiveness of different components of the proposed dataset (e.g., how training data at different difficulty levels affects model performance, what is the scaling-law effect for the dataset).\n    2. The case study presents only examples without systematic analysis or insights.\n    3. Critical training details are missing, including the exact number of GPU hours and computational costs.\n    4. Given that RL-Zero is known for generating significantly longer responses, the paper should report generation length statistics, but this information is absent.\n3. **Limited evaluation of generalizability:** The paper only evaluates DeepTheorem’s result on Qwen-2.5-Base across different sizes. It does not evaluate on additional model families such as Qwen-2.5-Math, Llama-4, and Qwen-3. It would be necessary to demonstrate the generalizability and robustness of the proposed dataset.\n4. **Presentation quality issues:** Several figures suffer from poor quality or clarity problems. Specifically, Figures 3 and 7 contain blurred text and incomplete labels in the pie charts, making them difficult to interpret.\n\nWith the above weakness, I can only provide a relatively negative overall opinion of the paper. However, I am willing to increase the score if the raised concerns are properly settled.", "questions": "Here are the questions for the authors:\n\n1. What is the connection between the proposed dataset and Lean-WB and DS-Prover-V1’s datasets? Is it necessary to compare them?\n2. What is the ablation study of each component of the proposed method? Are there more case studies and analyses? What is the GPU cost for Zero-RL and SFT training for the models? And what is the generated content length as the training goes on?\n3. How is the effect of DeepTheorem’s performance on SFT/RL on other leading open-source models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761767675696}, {"id": "eSlZXIYVb9", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17319/Reviewer_zq5f"], "rating": 2, "soundness": 1, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces DeepTheorem, a large-scale dataset designed for informal theorem proving. The dataset is richly annotated with information such as difficulty levels, proof labels, topic categories, problem variants, and exampled proofs. Building on this resource, the authors adapt RL-Zero to train large language models on DeepTheorem and propose both outcome-based and process-based evaluation metrics, using LLMs as judges. Experimental results demonstrate that models trained on DeepTheorem achieve stronger performance on external theorem-proving benchmarks compared to baselines of the same model scale.", "review_text": "This paper introduces DeepTheorem, a large-scale dataset designed for informal theorem proving. The dataset is richly annotated with information such as difficulty levels, proof labels, topic categories, problem variants, and exampled proofs. Building on this resource, the authors adapt RL-Zero to train large language models on DeepTheorem and propose both outcome-based and process-based evaluation metrics, using LLMs as judges. Experimental results demonstrate that models trained on DeepTheorem achieve stronger performance on external theorem-proving benchmarks compared to baselines of the same model scale.", "strengths": "* The paper is overall well-written and easy to follow, with an interesting focus on informal theorem proving.  \n* The proposed dataset is large-scale and carefully annotated with detailed information such as problem variants and difficulty levels.  \n* The LLM trained on this dataset achieves better performance than other baselines of the same model scale.", "weaknesses": "I think the major weakness of this paper lies in the fact that the entire dataset construction and evaluation pipeline is based solely on LLMs.\n\n* First, the original problem statements are drawn from existing datasets, which may have been included in the pretraining data of some LLMs. This raises concerns about potential contamination and data leakage.\n\n* Second, the generation of variants, the labeling of True or False, the construction of proofs, and even the assignment of difficulty levels and topics all depend heavily on existing LLMs—particularly o3-mini. In this sense, the dataset can be seen as a form of distillation of o3-mini onto a new dataset, rather than a fully independent benchmark.\n\n* Third, there is no human evaluation of the LLM annotations—not even on a sampled subset—to assess their quality. This absence raises serious concerns about the correctness of the proofs, labels, and overall dataset integrity.\n\n* Moreover, the evaluation relies on GPT-4o and o3-mini as LLM-as-a-judge models, with no manual inspection, even for a small portion of the data. Prior work has shown that such models may produce biased or inflated scores toward certain stylistic outputs (e.g., their own) [1]. Therefore, using a single LLM for evaluation, without any human verification, makes the results less reliable and less sound.\n\n* Finally, the baseline dataset appears to contain much less data, making the comparison seem somewhat unfair and potentially misleading.\n\nMinor: The question shown in Figure 2 appears incomplete, which further raises concerns about the dataset’s quality and attention to detail.\n\n[1] Gu, Jiawei, et al. “A Survey on LLM-as-a-Judge.” arXiv preprint arXiv:2411.15594 (2024).", "questions": "Could you manually evaluate a subset of the dataset to assess the accuracy and reliability of both the LLM-generated annotations and the LLM-as-a-judge evaluations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces DeepTheorem, a large-scale dataset designed for informal theorem proving. The dataset is richly annotated with information such as difficulty levels, proof labels, topic categories, problem variants, and exampled proofs. Building on this resource, the authors adapt RL-Zero to train large language models on DeepTheorem and propose both outcome-based and process-based evaluation metrics, using LLMs as judges. Experimental results demonstrate that models trained on DeepTheorem achieve stronger performance on external theorem-proving benchmarks compared to baselines of the same model scale.", "soundness": 1, "presentation": 3, "contribution": 2, "strengths": "* The paper is overall well-written and easy to follow, with an interesting focus on informal theorem proving.  \n* The proposed dataset is large-scale and carefully annotated with detailed information such as problem variants and difficulty levels.  \n* The LLM trained on this dataset achieves better performance than other baselines of the same model scale.", "weaknesses": "I think the major weakness of this paper lies in the fact that the entire dataset construction and evaluation pipeline is based solely on LLMs.\n\n* First, the original problem statements are drawn from existing datasets, which may have been included in the pretraining data of some LLMs. This raises concerns about potential contamination and data leakage.\n\n* Second, the generation of variants, the labeling of True or False, the construction of proofs, and even the assignment of difficulty levels and topics all depend heavily on existing LLMs—particularly o3-mini. In this sense, the dataset can be seen as a form of distillation of o3-mini onto a new dataset, rather than a fully independent benchmark.\n\n* Third, there is no human evaluation of the LLM annotations—not even on a sampled subset—to assess their quality. This absence raises serious concerns about the correctness of the proofs, labels, and overall dataset integrity.\n\n* Moreover, the evaluation relies on GPT-4o and o3-mini as LLM-as-a-judge models, with no manual inspection, even for a small portion of the data. Prior work has shown that such models may produce biased or inflated scores toward certain stylistic outputs (e.g., their own) [1]. Therefore, using a single LLM for evaluation, without any human verification, makes the results less reliable and less sound.\n\n* Finally, the baseline dataset appears to contain much less data, making the comparison seem somewhat unfair and potentially misleading.\n\nMinor: The question shown in Figure 2 appears incomplete, which further raises concerns about the dataset’s quality and attention to detail.\n\n[1] Gu, Jiawei, et al. “A Survey on LLM-as-a-Judge.” arXiv preprint arXiv:2411.15594 (2024).", "questions": "Could you manually evaluate a subset of the dataset to assess the accuracy and reliability of both the LLM-generated annotations and the LLM-as-a-judge evaluations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761427252722}, {"id": "aAzT9wxEWQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17319/Reviewer_Ua9x"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces DeepTheorem, a suite for informal (natural‑language/LaTeX) theorem proving consisting of a 121K-sample dataset of IMO‑level natural‑language theorems with concise proof sketches, topic labels, and a family of truth‑preserving or contradictory variants per theorem to enable automatic checking; 2) an adaptation of RL‑Zero/GRPO to informal theorem proving that uses pass/fail signals derived from those variants as rewards; and 3) an evaluation framework with both outcome accuracy (truth classification across variants) and process scoring of proofs by LLM judges, with a small human study for calibration", "review_text": "The paper introduces DeepTheorem, a suite for informal (natural‑language/LaTeX) theorem proving consisting of a 121K-sample dataset of IMO‑level natural‑language theorems with concise proof sketches, topic labels, and a family of truth‑preserving or contradictory variants per theorem to enable automatic checking; 2) an adaptation of RL‑Zero/GRPO to informal theorem proving that uses pass/fail signals derived from those variants as rewards; and 3) an evaluation framework with both outcome accuracy (truth classification across variants) and process scoring of proofs by LLM judges, with a small human study for calibration", "strengths": "1. Data Scale:  The 121K IMO‑level, topic‑labeled problems with curated entailed/contradictory variants are materially larger than prior NL theorem/proof corpora (e.g., NaturalProofs/NaturalProver focused on smaller‑scale NL proofs). The variant family provides a neat mechanism for binary checking without a formal verifier, and Figure 4 (p.4) shows difficulty calibration that overlaps harder benchmarks (miniF2F, FIMO, Putnam, HMMT).\n\n2. Empirical gains: Table 3 (p.7) and Figure 6 (p.8) indicate consistent improvements of RL‑Zero over SFT across Qwen2.5 7B/72B models on the proposed outcome‑metric and process‑scores.", "weaknesses": "1. The dataset is largely curated via web scraping existing problem sets and multiple LLM passes, with the final “step-by-step proofs” produced by black-box commercial models o3-mini. The RL part applies the naive 1/0 correctness reward extracted; and the process evaluation itself is performed by a black-box LLM judge (GPT-4o, with o3-mini as an alternative) rather than formal proof assistants such as lean. This pipeline raises concerns about originality (vs. prior RL-for-reasoning recipes), circularity, and reproducibility.\n\n2. Out-of-date baselines. The core experiments fine-tune Qwen2.5 model families and compare against a mix of 2024 models; however, several contemporaneous stronger families (e.g., Qwen3 dense / MoE series, GPT-5 ) and post-2024 formal provers (e.g., Seed-Prover variants) have significantly improved the frontier. Using pre-Qwen3 models limits the timeliness of the claims and their competitiveness at submission time.\n\n3. Unfair comparisons with formal theorem proving baselines. The test sets on table 4 from FIMO and PutnamBench are formal theorem-proving benchmarks but are manually converted to informal, variant-based NL evaluation; results are then compared against models for formal-theorem proving (such as DeepSeek-Prover) with formal accuracy (much more difficult than informal accuracies)Comparing an informal truth-assignment outcome to formal proof success under a verifier is not methodologically fair even if both are reported on the same underlying source problems.\n\n4. Decontamination is not documented w.r.t. time via temporal split. The paper details an embedding-based recall-and-justify decontamination and shows removed examples, but it does not report any timestamp/knowledge-cutoff filtering (e.g., train < test release date), nor quantitative false-positive/false-negative audits. Without a temporal split, leakage from recent competitions and public solutions remains a risk.\n\n5. Tiny human eval: the process scores depend on black-box GPT-4o/o3-mini judgments. While the paper reports high correlation between judges, LLM-based grading cannot ensure semantic correctness the way a formal checker can, and the human study covers only 10 items with two raters, which is far too small to validate reliability.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces DeepTheorem, a suite for informal (natural‑language/LaTeX) theorem proving consisting of a 121K-sample dataset of IMO‑level natural‑language theorems with concise proof sketches, topic labels, and a family of truth‑preserving or contradictory variants per theorem to enable automatic checking; 2) an adaptation of RL‑Zero/GRPO to informal theorem proving that uses pass/fail signals derived from those variants as rewards; and 3) an evaluation framework with both outcome accuracy (truth classification across variants) and process scoring of proofs by LLM judges, with a small human study for calibration", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Data Scale:  The 121K IMO‑level, topic‑labeled problems with curated entailed/contradictory variants are materially larger than prior NL theorem/proof corpora (e.g., NaturalProofs/NaturalProver focused on smaller‑scale NL proofs). The variant family provides a neat mechanism for binary checking without a formal verifier, and Figure 4 (p.4) shows difficulty calibration that overlaps harder benchmarks (miniF2F, FIMO, Putnam, HMMT).\n\n2. Empirical gains: Table 3 (p.7) and Figure 6 (p.8) indicate consistent improvements of RL‑Zero over SFT across Qwen2.5 7B/72B models on the proposed outcome‑metric and process‑scores.", "weaknesses": "1. The dataset is largely curated via web scraping existing problem sets and multiple LLM passes, with the final “step-by-step proofs” produced by black-box commercial models o3-mini. The RL part applies the naive 1/0 correctness reward extracted; and the process evaluation itself is performed by a black-box LLM judge (GPT-4o, with o3-mini as an alternative) rather than formal proof assistants such as lean. This pipeline raises concerns about originality (vs. prior RL-for-reasoning recipes), circularity, and reproducibility.\n\n2. Out-of-date baselines. The core experiments fine-tune Qwen2.5 model families and compare against a mix of 2024 models; however, several contemporaneous stronger families (e.g., Qwen3 dense / MoE series, GPT-5 ) and post-2024 formal provers (e.g., Seed-Prover variants) have significantly improved the frontier. Using pre-Qwen3 models limits the timeliness of the claims and their competitiveness at submission time.\n\n3. Unfair comparisons with formal theorem proving baselines. The test sets on table 4 from FIMO and PutnamBench are formal theorem-proving benchmarks but are manually converted to informal, variant-based NL evaluation; results are then compared against models for formal-theorem proving (such as DeepSeek-Prover) with formal accuracy (much more difficult than informal accuracies)Comparing an informal truth-assignment outcome to formal proof success under a verifier is not methodologically fair even if both are reported on the same underlying source problems.\n\n4. Decontamination is not documented w.r.t. time via temporal split. The paper details an embedding-based recall-and-justify decontamination and shows removed examples, but it does not report any timestamp/knowledge-cutoff filtering (e.g., train < test release date), nor quantitative false-positive/false-negative audits. Without a temporal split, leakage from recent competitions and public solutions remains a risk.\n\n5. Tiny human eval: the process scores depend on black-box GPT-4o/o3-mini judgments. While the paper reports high correlation between judges, LLM-based grading cannot ensure semantic correctness the way a formal checker can, and the human study covers only 10 items with two raters, which is far too small to validate reliability.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761418740076}], "openreview_url": "https://openreview.net/forum?id=0OUkySEAf0", "arxiv_id": "2505.23754", "paper_pdf": "papers/0OUkySEAf0.pdf", "paper_pdf_sha256": "5b3e29585076ac1b4343473081c296b578747e805a141abab2b5949ef9de9c5a", "paper_pdf_bytes": 1647137, "paper_pdf_source": "openreview", "code_url": "https://github.com/Jiahao004/DeepTheorem", "code_repository": "Jiahao004/DeepTheorem", "code_commit": "7a841fd421b34b05f387766893b70773c2c65d4a", "code_archive": "repos/0OUkySEAf0.zip", "code_archive_sha256": "d0f638b6b67667f346ca226f2def28256e7d51e919c71051226fc522fbb8d70e", "code_archive_bytes": 2600525, "code_file_count": 6, "code_extensions": {".py": 5, ".sh": 1}, "github_disk_usage_kb": 2572, "github_languages": {"Python": 12120, "Shell": 3163}, "github_archived": false, "github_pushed_at": "2025-06-10T15:29:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deeptheorem-advancing-llm-reasoning-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ns0KIpfQVy", "year": 2025, "status": "rejected", "title": "Multimodal Banking Dataset: Understanding Client Needs through Event Sequences", "authors": ["Dzhambulat Mollaev", "Alexander Kostin", "Postnova Maria", "Ivan Karpukhin", "Ivan A Kireev", "Gleb Gennadjevich Gusev", "Andrey Savchenko"], "authorids": ["~Dzhambulat_Mollaev1", "~Alexander_Kostin1", "~Postnova_Maria1", "~Ivan_Karpukhin1", "~Ivan_A_Kireev1", "~Gleb_Gennadjevich_Gusev1", "~Andrey_Savchenko1"], "authors_source": "OpenReview API", "abstract": "Financial organizations collect a huge amount of data about clients that typi-\ncally has a temporal (sequential) structure and is collected from multiple sources\n(modalities). However, despite the urgent practical need, developing deep learn-\ning techniques suitable to handle such data is limited by the absence of large open-\nsource multi-source real-world datasets of event sequences. To fill this gap mainly\ncaused by security reasons, we present the industrial-scale publicly available mul-\ntimodal banking dataset, MBD, that contains more than 2M corporate clients with\nseveral data sources: 950M bank transactions, 1B geo position events, 5M em-\nbeddings of dialogues with technical support and monthly aggregated purchases\nof four bank’s products. All entries are properly anonymized from real proprietary\nbank data. Moreover, we introduce a novel multimodal benchmark incorporating\nour MBD and two open-source financial datasets. We provide numerical results\ndemonstrating the superiority of fusion baselines over single-modal techniques\nfor each task. Moreover, our anonymization techniques still save all significant\ninformation for introduced downstream tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "0scQ1aOnSO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6922/Reviewer_YTbu"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "1. The paper introduces a new dataset, Multimodal Banking Dataset which integrates multiple modalities for over 2 million corporate clients.\n2. The authors highlight the potential applications of this dataset for campaign planning and client behavior analysis, using multimodal benchmarks to demonstrate its value along with baseline model implementation.", "review_text": "1. The paper introduces a new dataset, Multimodal Banking Dataset which integrates multiple modalities for over 2 million corporate clients.\n2. The authors highlight the potential applications of this dataset for campaign planning and client behavior analysis, using multimodal benchmarks to demonstrate its value along with baseline model implementation.", "strengths": "1. The work releases a first large scale banking dataset for public availability for financial applications. \n2. The authors present a good benchmark comparing unimodal and multimodal methods across various predictive tasks. The experimental protocol and metrics are clearly laid out as well.", "weaknesses": "1. The authors do not explore or discuss advanced multimodal sequence models or advanced fusion techniques' cross-attention mechanisms as they can better capture interactions across modalities. They mention it at the end as a scope for future work.\n2. Though authors discuss using AUC ROC as their metric for mitigating label imbalance issues for example in their campaigning downstream task, they do not discuss or incorporate any additional techniques for handling the label imbalance.\n3. Details about anonymization techniques applied are mentioned in the paper but it lacks quantitative evaluation of the impact of these techniques on temporal dependencies within the data.", "questions": "1. Is there a plan to expand the set of downstream tasks in future work? Highlighting a larger application list can increase the dataset's appeal across different financial research areas.\n\nPlease look into the weakness section for other questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "1. The paper introduces a new dataset, Multimodal Banking Dataset which integrates multiple modalities for over 2 million corporate clients.\n2. The authors highlight the potential applications of this dataset for campaign planning and client behavior analysis, using multimodal benchmarks to demonstrate its value along with baseline model implementation.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The work releases a first large scale banking dataset for public availability for financial applications. \n2. The authors present a good benchmark comparing unimodal and multimodal methods across various predictive tasks. The experimental protocol and metrics are clearly laid out as well.", "weaknesses": "1. The authors do not explore or discuss advanced multimodal sequence models or advanced fusion techniques' cross-attention mechanisms as they can better capture interactions across modalities. They mention it at the end as a scope for future work.\n2. Though authors discuss using AUC ROC as their metric for mitigating label imbalance issues for example in their campaigning downstream task, they do not discuss or incorporate any additional techniques for handling the label imbalance.\n3. Details about anonymization techniques applied are mentioned in the paper but it lacks quantitative evaluation of the impact of these techniques on temporal dependencies within the data.", "questions": "1. Is there a plan to expand the set of downstream tasks in future work? Highlighting a larger application list can increase the dataset's appeal across different financial research areas.\n\nPlease look into the weakness section for other questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730669726154}, {"id": "36ryiy1oVm", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6922/Reviewer_e7p1"], "rating": 8, "soundness": 4, "presentation": 2, "contribution": 4, "confidence": 4, "summary": "This is not a research contribution per se but it is a dataset paper.\n\nThe paper presents the first large-scale multimodal banking dataset for the user community. It offers a new dataset that has millions of users and millions of transactions that have been suitably anonymized. Moreover, the authors have provided baselines for a few standard tasks. The dataset will be released in order to spur research in the use of machine learning for banking applications.", "review_text": "This is not a research contribution per se but it is a dataset paper.\n\nThe paper presents the first large-scale multimodal banking dataset for the user community. It offers a new dataset that has millions of users and millions of transactions that have been suitably anonymized. Moreover, the authors have provided baselines for a few standard tasks. The dataset will be released in order to spur research in the use of machine learning for banking applications.", "strengths": "(1) This will be first and the largest multimodal banking dataset that will be released. This can potentially be tremendously useful to the research community.\n\n(2) The baseline methods and benchmark data for a few problems outlined will also be immensely useful to the research community.", "weaknesses": "(1) The details of the data are somewhat sparse. More details of each type of data will be useful to the reader. Perhaps this article may be useful for improving this aspect of the exposition in the paper: \n\nhttps://cacm.acm.org/research/datasheets-for-datasets/\n\n(2)  Can some details of the anonymization be provided without compromising on the privacy of the customers? That can help estimate the errors of any model developed using this data.\n\n(3) The data has been collected during the pandemic period? Will it have any effect on any of the conclusions drawn using the data? For examples, could this lead to systemic under-estimation or over-estimation of any phenomenon? Ideally, it would be useful to have another data-set which is outside of the pandemic period - perhaps the second version of this data set? This can serve as a basis for many natural experiments.", "questions": "(1) Will the dataset be openly downloadable or will the authors be controlling the access? Openly downloadable option is obviously preferable.\n(2) Will the source code of the bench-marking studies be openly available? \n(3) it will be useful to the community if the set of relevant real-world problems could be articulated that can inspire researchers to work on this dataset. Especially, in order to attract young researchers into the field.", "flag_for_ethics_review": ["Yes, Legal compliance (e.g., GDPR, copyright, terms of use)"], "all_content": {"summary": "This is not a research contribution per se but it is a dataset paper.\n\nThe paper presents the first large-scale multimodal banking dataset for the user community. It offers a new dataset that has millions of users and millions of transactions that have been suitably anonymized. Moreover, the authors have provided baselines for a few standard tasks. The dataset will be released in order to spur research in the use of machine learning for banking applications.", "soundness": 4, "presentation": 2, "contribution": 4, "strengths": "(1) This will be first and the largest multimodal banking dataset that will be released. This can potentially be tremendously useful to the research community.\n\n(2) The baseline methods and benchmark data for a few problems outlined will also be immensely useful to the research community.", "weaknesses": "(1) The details of the data are somewhat sparse. More details of each type of data will be useful to the reader. Perhaps this article may be useful for improving this aspect of the exposition in the paper: \n\nhttps://cacm.acm.org/research/datasheets-for-datasets/\n\n(2)  Can some details of the anonymization be provided without compromising on the privacy of the customers? That can help estimate the errors of any model developed using this data.\n\n(3) The data has been collected during the pandemic period? Will it have any effect on any of the conclusions drawn using the data? For examples, could this lead to systemic under-estimation or over-estimation of any phenomenon? Ideally, it would be useful to have another data-set which is outside of the pandemic period - perhaps the second version of this data set? This can serve as a basis for many natural experiments.", "questions": "(1) Will the dataset be openly downloadable or will the authors be controlling the access? Openly downloadable option is obviously preferable.\n(2) Will the source code of the bench-marking studies be openly available? \n(3) it will be useful to the community if the set of relevant real-world problems could be articulated that can inspire researchers to work on this dataset. Especially, in order to attract young researchers into the field.", "flag_for_ethics_review": ["Yes, Legal compliance (e.g., GDPR, copyright, terms of use)"], "details_of_ethics_concerns": "It is hoped that the permission of the appropriate banking regulator (in the jurisdiction of the authors) has been taken in order to release the dataset. The authors, the authors' institution as well as researchers in the field should not be exposed to any litigation risk by a customer or a group of customers.", "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730655510060}, {"id": "pCtDCV7st5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6922/Reviewer_hmLd"], "rating": 3, "soundness": 1, "presentation": 1, "contribution": 1, "confidence": 5, "summary": "This paper introduces a large-scale publicly available multimodal banking dataset (Multimodal Banking Dataset, MBD). The MBD contains data from over 2 million banking customers, covering four different modalities: 9.5 million banking transaction records, 1 billion geographic location data points, 5 million customer interactions with technical support embedded dialogues, and banking product purchase activity over a 12-month period. This dataset provides researchers with a rich resource for analyzing customer behavior dynamics and contributes to the development of large-scale multimodal event sequence algorithms in the future.", "review_text": "This paper introduces a large-scale publicly available multimodal banking dataset (Multimodal Banking Dataset, MBD). The MBD contains data from over 2 million banking customers, covering four different modalities: 9.5 million banking transaction records, 1 billion geographic location data points, 5 million customer interactions with technical support embedded dialogues, and banking product purchase activity over a 12-month period. This dataset provides researchers with a rich resource for analyzing customer behavior dynamics and contributes to the development of large-scale multimodal event sequence algorithms in the future.", "strengths": "1.A large-scale multimodal banking dataset, MBD, is provided. This dataset contains anonymized banking transactions, geographic locations, and technical support dialogues, which contributes to the development of large-scale sequential event tasks in the future.\n2.The dataset addresses privacy concerns through effective data anonymization, ensuring that the algorithm's performance is not significantly compromised.\n3.The dataset and experimental code are publicly available, promoting transparency and reproducibility in research.", "weaknesses": "1.The dataset's multimodal data includes banking transaction records, geographic locations, dialogue embeddings, and banking product purchase history. However, it appears that many of these modalities are essentially text-based. This differs from typical multimodal datasets, which include modalities such as video, audio, and text. \n\n2.The main contribution of the paper lies in the introduction of a large-scale dataset, but it lacks innovative methods for addressing related tasks. Additionally, the experimental section presents insufficient comparisons with current state-of-the-art methods. Overall, the paper's innovativeness needs improvement. I suggest that the authors include more experiments involving fully supervised methods [1][2]. Additionally, it would be beneficial to propose a simple and practical innovative approach based on their dataset.\n[1]Learning Deep Time-index Models for Time Series Forecasting. [2]Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting\n\n3.This paper contains numerous grammatical and tense issues. Additionally, the expression in the paper is not sufficiently clear, making it difficult for readers to accurately understand certain points and arguments. For example, on page three, phrases like “we selected” and “the dataset was collected”; on page five, “the coordinates were”; line 225 includes “we concentrated,” “whether it was,” and “random noise was added.” Additionally, on line 295, “we propose a downstream task — multimodal matching” (Zong et al., 2023).", "questions": "1. Do you have any specific method to model the different text modalities?\n2.The paper contains numerous tense and grammatical errors, which do not meet the standards expected for ICLR.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a large-scale publicly available multimodal banking dataset (Multimodal Banking Dataset, MBD). The MBD contains data from over 2 million banking customers, covering four different modalities: 9.5 million banking transaction records, 1 billion geographic location data points, 5 million customer interactions with technical support embedded dialogues, and banking product purchase activity over a 12-month period. This dataset provides researchers with a rich resource for analyzing customer behavior dynamics and contributes to the development of large-scale multimodal event sequence algorithms in the future.", "soundness": 1, "presentation": 1, "contribution": 1, "strengths": "1.A large-scale multimodal banking dataset, MBD, is provided. This dataset contains anonymized banking transactions, geographic locations, and technical support dialogues, which contributes to the development of large-scale sequential event tasks in the future.\n2.The dataset addresses privacy concerns through effective data anonymization, ensuring that the algorithm's performance is not significantly compromised.\n3.The dataset and experimental code are publicly available, promoting transparency and reproducibility in research.", "weaknesses": "1.The dataset's multimodal data includes banking transaction records, geographic locations, dialogue embeddings, and banking product purchase history. However, it appears that many of these modalities are essentially text-based. This differs from typical multimodal datasets, which include modalities such as video, audio, and text. \n\n2.The main contribution of the paper lies in the introduction of a large-scale dataset, but it lacks innovative methods for addressing related tasks. Additionally, the experimental section presents insufficient comparisons with current state-of-the-art methods. Overall, the paper's innovativeness needs improvement. I suggest that the authors include more experiments involving fully supervised methods [1][2]. Additionally, it would be beneficial to propose a simple and practical innovative approach based on their dataset.\n[1]Learning Deep Time-index Models for Time Series Forecasting. [2]Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting\n\n3.This paper contains numerous grammatical and tense issues. Additionally, the expression in the paper is not sufficiently clear, making it difficult for readers to accurately understand certain points and arguments. For example, on page three, phrases like “we selected” and “the dataset was collected”; on page five, “the coordinates were”; line 225 includes “we concentrated,” “whether it was,” and “random noise was added.” Additionally, on line 295, “we propose a downstream task — multimodal matching” (Zong et al., 2023).", "questions": "1. Do you have any specific method to model the different text modalities?\n2.The paper contains numerous tense and grammatical errors, which do not meet the standards expected for ICLR.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730507522577}, {"id": "KQcgNpCNOR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6922/Reviewer_PGwK"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper introduces a new multimodal dataset, MBD, from over 2 million corporate clients, including bank transactions, geo-locations, technical support dialogues, and monthly aggregated purchases of four banking products. The authors also present a benchmark for evaluating models on this dataset and two other financial datasets, focusing on tasks like purchase prediction and multimodal matching.", "review_text": "The paper introduces a new multimodal dataset, MBD, from over 2 million corporate clients, including bank transactions, geo-locations, technical support dialogues, and monthly aggregated purchases of four banking products. The authors also present a benchmark for evaluating models on this dataset and two other financial datasets, focusing on tasks like purchase prediction and multimodal matching.", "strengths": "- Large-scale dataset: The MBD dataset is the largest of its kind, offering a significant amount of data for research purposes.   \n\n- Multimodality: The dataset incorporates various data modalities, providing a more comprehensive view of client behavior.   \n\n- Practical tasks: The benchmark focuses on practically relevant tasks, such as purchase prediction, which can be useful for real-world applications.   \n\n- Anonymization: The authors have taken steps to anonymize the data, protecting client privacy.", "weaknesses": "- Lack of novelty: The paper primarily focuses on introducing a dataset and benchmark. For ICLR submissions, I'd see more emphasis on novel methods or algorithms. In addition, a careful evaluation of previous models (for example, the MMBench https://arxiv.org/abs/2307.06281) will bring more novelty to the community. \n\n- Missing comparison with LLMs: The paper lacks a comparison with more recent and powerful language models like GPT-4 or BloombergGPT, which have shown strong performance in various financial NLP tasks. For example, can we use prompt engineering to guide the GPT4o to process the transactions and geo-locations. \n\n- Unclear practical impact of the proposed metric: While the paper mentions that the benchmark can lead to financial benefits, I am curious if real world users care recalls/AUCs, or there are better metrics that map to financial success.", "questions": "-  The paper could be strengthened by exploring more advanced multimodal fusion techniques beyond late fusion.\n-  A more detailed analysis of the anonymization process and its potential impact on model performance would be beneficial.\n-  The authors could consider expanding the benchmark to include other relevant tasks, such as risk assessment or fraud detection.\n\nOverall, while the MBD dataset and benchmark are valuable contributions, the paper needs significant revisions to address the lack of novelty and provide a more convincing argument for the practical impact of their work.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new multimodal dataset, MBD, from over 2 million corporate clients, including bank transactions, geo-locations, technical support dialogues, and monthly aggregated purchases of four banking products. The authors also present a benchmark for evaluating models on this dataset and two other financial datasets, focusing on tasks like purchase prediction and multimodal matching.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Large-scale dataset: The MBD dataset is the largest of its kind, offering a significant amount of data for research purposes.   \n\n- Multimodality: The dataset incorporates various data modalities, providing a more comprehensive view of client behavior.   \n\n- Practical tasks: The benchmark focuses on practically relevant tasks, such as purchase prediction, which can be useful for real-world applications.   \n\n- Anonymization: The authors have taken steps to anonymize the data, protecting client privacy.", "weaknesses": "- Lack of novelty: The paper primarily focuses on introducing a dataset and benchmark. For ICLR submissions, I'd see more emphasis on novel methods or algorithms. In addition, a careful evaluation of previous models (for example, the MMBench https://arxiv.org/abs/2307.06281) will bring more novelty to the community. \n\n- Missing comparison with LLMs: The paper lacks a comparison with more recent and powerful language models like GPT-4 or BloombergGPT, which have shown strong performance in various financial NLP tasks. For example, can we use prompt engineering to guide the GPT4o to process the transactions and geo-locations. \n\n- Unclear practical impact of the proposed metric: While the paper mentions that the benchmark can lead to financial benefits, I am curious if real world users care recalls/AUCs, or there are better metrics that map to financial success.", "questions": "-  The paper could be strengthened by exploring more advanced multimodal fusion techniques beyond late fusion.\n-  A more detailed analysis of the anonymization process and its potential impact on model performance would be beneficial.\n-  The authors could consider expanding the benchmark to include other relevant tasks, such as risk assessment or fraud detection.\n\nOverall, while the MBD dataset and benchmark are valuable contributions, the paper needs significant revisions to address the lack of novelty and provide a more convincing argument for the practical impact of their work.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730140681802}], "openreview_url": "https://openreview.net/forum?id=ns0KIpfQVy", "arxiv_id": "2409.17587", "paper_pdf": "papers/ns0KIpfQVy.pdf", "paper_pdf_sha256": "6e35b32e79ea35967450001bde8422d29397d27c233ce0274560840a0ec11e23", "paper_pdf_bytes": 464801, "paper_pdf_source": "openreview", "code_url": "https://github.com/Dzhambo/MBD", "code_repository": "Dzhambo/MBD", "code_commit": "0f8a41b2eca7945656b51dfb447b0d08b624ac77", "code_archive": "repos/ns0KIpfQVy.zip", "code_archive_sha256": "1f4c2672cdef3dc904541aabd6c5201a037e3b4267a09ba8ff7fb56ca221ba92", "code_archive_bytes": 121140, "code_file_count": 39, "code_extensions": {".sh": 23, ".py": 9, ".ipynb": 7}, "github_disk_usage_kb": 177, "github_languages": {"Jupyter Notebook": 138212, "Python": 38762, "Shell": 19999}, "github_archived": false, "github_pushed_at": "2025-03-03T14:38:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multimodal-banking-dataset-understanding"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r0BcyqWAcj", "year": 2024, "status": "rejected", "title": "Loci-Segmented: Improving Scene Segmentation Learning", "authors": ["Manuel Traub", "Frederic Becker", "Adrian Sauter", "Sebastian Otte", "Martin V. Butz"], "authorids": ["~Manuel_Traub1", "~Frederic_Becker1", "~Adrian_Sauter1", "~Sebastian_Otte1", "~Martin_V._Butz2"], "authors_source": "OpenReview API", "abstract": "Slot-oriented processing approaches for compositional scene representation have recently undergone a tremendous development. We present Loci-Segmented (Loci-s), an advanced scene segmentation neural network that extends the slot-based location and identity tracking architecture Loci (Traub et al., ICLR 2023). The main advancements are (i) the addition of a pre-trained dynamic background module; (ii) a hyper-convolution encoder module, which enables object-focused bottom-up processing; and (iii) a cascaded decoder module, which successively generates object masks, masked depth maps, and masked, depth-map-informed RGB reconstructions. The background module features the learning of both a foreground identifying module and a background re-generator. We further improve performance via (a) the integration of depth information as well as improved slot assignments via (b) slot-location-entity regularization and (b) a prior segmentation network. Even without these latter improvements, the results reveal superior segmentation performance in the MOVi datasets and in another established dataset collection. With all improvements, Loci-s achieves a 32% better intersection over union (IoU) score in MOVi-E than the previous best. We furthermore show that Loci-s generates well-interpretable latent representations. We believe that these representations may serve as a foundation-model-like interpretable basis for solving downstream tasks, such as grounding language and context- and\ngoal-conditioned event processing.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "plutJtupp2", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5948/Reviewer_vymZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper extends the location and identity tracking architecture Loci to scene segmentation by adding a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. The proposed method and each components are validated to be effectiveness by extensive experiments.", "review_text": "The paper extends the location and identity tracking architecture Loci to scene segmentation by adding a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. The proposed method and each components are validated to be effectiveness by extensive experiments.", "strengths": "The experiments are extensive.", "weaknesses": "1. The work is a little incremental, compared to Loci, so that its novelty is slim.\n2. The principle and motivation of the proposed modules are not clearly explained.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper extends the location and identity tracking architecture Loci to scene segmentation by adding a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. The proposed method and each components are validated to be effectiveness by extensive experiments.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The experiments are extensive.", "weaknesses": "1. The work is a little incremental, compared to Loci, so that its novelty is slim.\n2. The principle and motivation of the proposed modules are not clearly explained.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699195009811}, {"id": "VYSDH7p4Pn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5948/Reviewer_czo6"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes an architecture for unsupervised scene segmentation given RGB or RGBD video input. The method builds on the \"Loci\" paper, but re-designs many components and adds in a pre-trained foreground/background segmentation model. The main result is that this combination of changes greatly improves results, both qualitatively and quantitatively.", "review_text": "This paper proposes an architecture for unsupervised scene segmentation given RGB or RGBD video input. The method builds on the \"Loci\" paper, but re-designs many components and adds in a pre-trained foreground/background segmentation model. The main result is that this combination of changes greatly improves results, both qualitatively and quantitatively.", "strengths": "Quantitatively the method here clearly outperforms prior work (on the mIOU metric) in the the MOVI-* datasets.", "weaknesses": "Overall this paper is very difficult to follow. The \"Loci\" method, on which this is based, is never quite made clear on its own, and then every subsequent section makes big changes to the architecture without much motivation, and without a connecting story or high-level idea. \n\nThe section on the \"Background Module\" never mentions this, but the abstract and the section on \"Segmentation Preprocessing\" describe the background module as \"pre-trained\", apparently for a segmentation task that \"distinguishes foreground entities from the background context\". My guess is that much of the performance gain is coming from this. \n\nI have a variety of smaller questions which the authors may like to answer, but overall it seems to me that this paper needs a very heavy rewrite.", "questions": "What are Gestalt codes? \n\nWhat are the two predictions about object positions? The paper says \"we introduce a dedicated background processing module that generates both predictions about object positions as pixel densities\".\n\nThe paper mentions using something called \"GateL0RD units\" but these are never really described. \n\nThe paper says that the \"Gestalt codes are binarized to create an information bottleneck that fosters the development of factorized compositional entity encodings.\" I am unclear on why binarization will make the representation compositional. \n\nSection 2.1 focuses on improving Loci's \"object tracking abilities\", but the earlier section (describing Loci) never mentioned any object tracking happening, and tracking is never mentioned again. What is the idea here? \n\nThe paper mentions that the decoder \"reconstructs the predicted scene via slot-wise density maps as object masks.\" What are slot-wise density maps? \n\nThe paper briefly mentions an \"L0 loss on gate openings\" but it is not clear what ground truth is used for this loss. Is it maybe just a regularization term, penalizing the L0 norm?\n\nSection 2.2 introduces a depth input and an equation to normalize it, but it is not clear where this fits with the inner loop described in the previous section.\n\nFor Table 2 it would be great to clarify what dataset these experiments happen in, and what the metrics are.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an architecture for unsupervised scene segmentation given RGB or RGBD video input. The method builds on the \"Loci\" paper, but re-designs many components and adds in a pre-trained foreground/background segmentation model. The main result is that this combination of changes greatly improves results, both qualitatively and quantitatively.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "Quantitatively the method here clearly outperforms prior work (on the mIOU metric) in the the MOVI-* datasets.", "weaknesses": "Overall this paper is very difficult to follow. The \"Loci\" method, on which this is based, is never quite made clear on its own, and then every subsequent section makes big changes to the architecture without much motivation, and without a connecting story or high-level idea. \n\nThe section on the \"Background Module\" never mentions this, but the abstract and the section on \"Segmentation Preprocessing\" describe the background module as \"pre-trained\", apparently for a segmentation task that \"distinguishes foreground entities from the background context\". My guess is that much of the performance gain is coming from this. \n\nI have a variety of smaller questions which the authors may like to answer, but overall it seems to me that this paper needs a very heavy rewrite.", "questions": "What are Gestalt codes? \n\nWhat are the two predictions about object positions? The paper says \"we introduce a dedicated background processing module that generates both predictions about object positions as pixel densities\".\n\nThe paper mentions using something called \"GateL0RD units\" but these are never really described. \n\nThe paper says that the \"Gestalt codes are binarized to create an information bottleneck that fosters the development of factorized compositional entity encodings.\" I am unclear on why binarization will make the representation compositional. \n\nSection 2.1 focuses on improving Loci's \"object tracking abilities\", but the earlier section (describing Loci) never mentioned any object tracking happening, and tracking is never mentioned again. What is the idea here? \n\nThe paper mentions that the decoder \"reconstructs the predicted scene via slot-wise density maps as object masks.\" What are slot-wise density maps? \n\nThe paper briefly mentions an \"L0 loss on gate openings\" but it is not clear what ground truth is used for this loss. Is it maybe just a regularization term, penalizing the L0 norm?\n\nSection 2.2 introduces a depth input and an equation to normalize it, but it is not clear where this fits with the inner loop described in the previous section.\n\nFor Table 2 it would be great to clarify what dataset these experiments happen in, and what the metrics are.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698814579513}, {"id": "mcygNkr8OQ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5948/Reviewer_xtqd"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "This work focuses on the compositional scene representation and proposes a scene segmentation neural network based on the previous model named Loci. To build their model, they extend Loci with three modifications, including a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. Extensive experiments conducted on the MOVi dataset show the effectiveness of the proposed method. Besides, the proposed method can generate well-interpretable latent representations and may serve as a foundation-model-like interpretable basis for solving downstream tasks.", "review_text": "This work focuses on the compositional scene representation and proposes a scene segmentation neural network based on the previous model named Loci. To build their model, they extend Loci with three modifications, including a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. Extensive experiments conducted on the MOVi dataset show the effectiveness of the proposed method. Besides, the proposed method can generate well-interpretable latent representations and may serve as a foundation-model-like interpretable basis for solving downstream tasks.", "strengths": "1. Good performance. The proposed method achieves good performance on the MOVi dataset.\n2. The proposed can generate well-interpretable latent representations, which is helpful in building interpretable foundation models.", "weaknesses": "To ACs and authors: I am not an expert in this field and cannot find any strong reasons to reject this work. Please refer to other reviewers' comments for rebuttal and decision.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work focuses on the compositional scene representation and proposes a scene segmentation neural network based on the previous model named Loci. To build their model, they extend Loci with three modifications, including a pre-trained dynamic background\nmodule, a hyper-convolution encoder module, and a cascaded decoder module. Extensive experiments conducted on the MOVi dataset show the effectiveness of the proposed method. Besides, the proposed method can generate well-interpretable latent representations and may serve as a foundation-model-like interpretable basis for solving downstream tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Good performance. The proposed method achieves good performance on the MOVi dataset.\n2. The proposed can generate well-interpretable latent representations, which is helpful in building interpretable foundation models.", "weaknesses": "To ACs and authors: I am not an expert in this field and cannot find any strong reasons to reject this work. Please refer to other reviewers' comments for rebuttal and decision.", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698772811864}, {"id": "niKBfPaxpm", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5948/Reviewer_wWut"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes Loci-s (Loci-Segmented) to tackle the problem of slot-oriented scene representation. Loci-s is an extension of the Loci architecture with structure change, additional inputs and etc. The proposed methods shows state-of-the-art performance on the challenging MOVi-E dataset, demonstrating its ability to deal with complex environments.", "review_text": "This paper proposes Loci-s (Loci-Segmented) to tackle the problem of slot-oriented scene representation. Loci-s is an extension of the Loci architecture with structure change, additional inputs and etc. The proposed methods shows state-of-the-art performance on the challenging MOVi-E dataset, demonstrating its ability to deal with complex environments.", "strengths": "1)This paper extends the Loci model to Loci-s with several innovations.\n2)The advancements in Loci-s collectively contribute to a 32.06% relative improvement in IoU on the challenging MOVi-E dataset compared to state-of-the-art models like SAVi++.", "weaknesses": "1)Since the proposed method is built upon Loci, I think there should be more comparisons between Loci and Loci-s in the experimental section.\n\n2)The authors mentioned that instead of the residual structure used in Loci, the encoder and decoder subnetworks in Loci-s have been revamped to adopt a ConvNeXt-like architecture. I wonder how much performance improvement is brought by this structure change. \n\n3)“The third methodology involves the deployment of a specialized segmentation network akin to YOLACT (Bolya et al., 2019). ” How is this segmentation network trained, and what is its performance?\n\n4)The proposed method incorporates some additional input information for performance boost, e.g. the segmentation input (seg), depth map (sd). How much is the time cost? \n\n5)(NOT IMPORTANT) Seems there is a missing reference (shown as ?) in sentence “Loci is rather closely related to other slot-based object processing architectures (Elsayedet al.; ?; Kipf et al., 2022; Wu et al., 2023)...”", "questions": "Same as weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Loci-s (Loci-Segmented) to tackle the problem of slot-oriented scene representation. Loci-s is an extension of the Loci architecture with structure change, additional inputs and etc. The proposed methods shows state-of-the-art performance on the challenging MOVi-E dataset, demonstrating its ability to deal with complex environments.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1)This paper extends the Loci model to Loci-s with several innovations.\n2)The advancements in Loci-s collectively contribute to a 32.06% relative improvement in IoU on the challenging MOVi-E dataset compared to state-of-the-art models like SAVi++.", "weaknesses": "1)Since the proposed method is built upon Loci, I think there should be more comparisons between Loci and Loci-s in the experimental section.\n\n2)The authors mentioned that instead of the residual structure used in Loci, the encoder and decoder subnetworks in Loci-s have been revamped to adopt a ConvNeXt-like architecture. I wonder how much performance improvement is brought by this structure change. \n\n3)“The third methodology involves the deployment of a specialized segmentation network akin to YOLACT (Bolya et al., 2019). ” How is this segmentation network trained, and what is its performance?\n\n4)The proposed method incorporates some additional input information for performance boost, e.g. the segmentation input (seg), depth map (sd). How much is the time cost? \n\n5)(NOT IMPORTANT) Seems there is a missing reference (shown as ?) in sentence “Loci is rather closely related to other slot-based object processing architectures (Elsayedet al.; ?; Kipf et al., 2022; Wu et al., 2023)...”", "questions": "Same as weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698721951894}], "openreview_url": "https://openreview.net/forum?id=r0BcyqWAcj", "arxiv_id": "2310.10410", "paper_pdf": "papers/r0BcyqWAcj.pdf", "paper_pdf_sha256": "2011ce3fb1fd19cf1e4aebdd4d1f773e9199de4d7ccc7d6352c59b9a5ecdd3ea", "paper_pdf_bytes": 2157267, "paper_pdf_source": "openreview", "code_url": "https://github.com/CognitiveModeling/Loci-Segmented", "code_repository": "CognitiveModeling/Loci-Segmented", "code_commit": "b4abc3a83a2de825da6bde33cb684b7d9daceaa2", "code_archive": "repos/r0BcyqWAcj.zip", "code_archive_sha256": "677753b5522830bcb4ae079f0c01af69c860076743add987a16ee434da3be6a2", "code_archive_bytes": 161551, "code_file_count": 52, "code_extensions": {".py": 49, ".sh": 3}, "github_disk_usage_kb": 144, "github_languages": {"Python": 452780, "Shell": 34659}, "github_archived": false, "github_pushed_at": "2025-05-23T07:38:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/loci-segmented-improving-scene-segmentation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ToYi8C6fetv", "year": 2023, "status": "rejected", "title": "Personalized Subgraph Federated Learning", "authors": ["Jinheon Baek", "Wonyong Jeong", "Jiongdao Jin", "Jaehong Yoon", "Sung Ju Hwang"], "authorids": ["~Jinheon_Baek1", "~Wonyong_Jeong1", "~Jiongdao_Jin1", "~Jaehong_Yoon1", "~Sung_Ju_Hwang1"], "authors_source": "OpenReview API", "abstract": "In real-world scenarios, subgraphs of a larger global graph may be distributed across multiple devices or institutions, and only locally accessible due to privacy restrictions, although there may be links between them. Recently proposed subgraph Federated Learning (FL) methods deal with those missing links across private local subgraphs while distributively training Graph Neural Networks (GNNs) on them. However, they have overlooked the inevitable heterogeneity among subgraphs, caused by subgraphs comprising different communities of a global graph, therefore, consequently collapsing the incompatible knowledge from local GNN models trained on heterogeneous graph distributions. To overcome such a limitation, we introduce a new subgraph FL problem, personalized subgraph FL, which focuses on the joint improvement of the interrelated local GNN models rather than learning a single global GNN model, and propose a novel framework, FEDerated Personalized sUBgraph learning (FED-PUB), to tackle it. A crucial challenge in personalized subgraph FL is that the server does not know which subgraph each client has. FED-PUB thus utilizes functional embeddings of the local GNNs using random graphs as inputs to compute similarities between them, and use them to perform weighted averaging for server-side aggregation. Further, it learns a personalized sparse mask at each client to select and update only the subgraph-relevant subset of the aggregated parameters. We validate FED-PUB for its subgraph FL performance on six datasets, considering both non-overlapping and overlapping subgraphs, on which ours largely outperforms relevant baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "s7xk0sBQKB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1247/Reviewer_TrJG"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a personalized method for subgraph learning with multiple clients in different communities. The key idea is to measure the similarity between each customer pair and use that similar information for succeeding updates. Empirical results demonstrate superior performance on test accuracy.", "review_text": "This paper proposed a personalized idea for subgraph federated learning. My concern is that the accuracy gain may come from user information leakage. And there is no systematic discussion regarding the privacy-accuracy tradeoff.", "strengths": "Strengths: \n1. This paper proposes an implicit method to exploit the community structure within the graph using pairwise similarity scores on clients. With this method, we can properly handle missing edges and subgraph heterogeneity with a reasonable performance gain in accuracy.\n\n2. Some implementation refinements, such as masking and regularizations, are applied to achieve better performance.\n\nWeaknesses:\n1. The method may sacrifice the user's privacy by calculating personalized information on the server side. From Equation (4), even though the information among different clients are mixed together on the server's side, there is a chance that personal data can leak since the mixing rule encourages aggregating similar clients. However, the paper needed to articulate the tradeoff between privacy and accuracy.\n\n2. The paper should have explicitly modeled a rigid community structure on the server's side. Instead, it uses an implicit similarity measurement that can potentially measure more complicated structures, such as overlapping communities. However, the empirical analysis needed to show a deeper insight using the similarity structure.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a personalized method for subgraph learning with multiple clients in different communities. The key idea is to measure the similarity between each customer pair and use that similar information for succeeding updates. Empirical results demonstrate superior performance on test accuracy.", "strength_and_weaknesses": "Strengths: \n1. This paper proposes an implicit method to exploit the community structure within the graph using pairwise similarity scores on clients. With this method, we can properly handle missing edges and subgraph heterogeneity with a reasonable performance gain in accuracy.\n\n2. Some implementation refinements, such as masking and regularizations, are applied to achieve better performance.\n\nWeaknesses:\n1. The method may sacrifice the user's privacy by calculating personalized information on the server side. From Equation (4), even though the information among different clients are mixed together on the server's side, there is a chance that personal data can leak since the mixing rule encourages aggregating similar clients. However, the paper needed to articulate the tradeoff between privacy and accuracy.\n\n2. The paper should have explicitly modeled a rigid community structure on the server's side. Instead, it uses an implicit similarity measurement that can potentially measure more complicated structures, such as overlapping communities. However, the empirical analysis needed to show a deeper insight using the similarity structure.", "clarity,_quality,_novelty_and_reproducibility": "A few comments on paper writing:\n1. In Section 3 \"each column represents a ...\" -> \"each row represents a ...\".\n2. Section 3 uses superscripts with brackets for both rounds and layers.\n3. In Section 3, \"few existing methods\" -> \"existing methods\".\n4. In Equation (2), what does it mean by summing over G_i when optimizing (\\theta_i, \\mu_i)?\n5. Figure 2 caption: \"each of which consists of one/two subgraphs\" -> \"community A consists of one subgraph, and community B consists of two subgraphs.\"", "summary_of_the_review": "This paper proposed a personalized idea for subgraph federated learning. My concern is that the accuracy gain may come from user information leakage. And there is no systematic discussion regarding the privacy-accuracy tradeoff.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666906483213}, {"id": "9gQWImDH11", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1247/Reviewer_PHC3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Unlike the existing subgraph federated learning methods that overlook the heterogeneity among subgraphs, this paper proposes a personalized subgraph federated learning framework named FED-PUB, which enhances collaborative learning of subgraphs within the same community. To this end, it captures the communities based on functional embeddings of clients, then employs personalized sparse masks on the local GNN. Empirical results on six public datasets demonstrate the effectiveness of the proposed method. ", "review_text": "Due to the limited technical novelty and misleading descriptions of related works, I recommend reject.", "strengths": "Pros:\n\tThe motivation of this paper, i.e., considering the heterogeneity among subgraphs in the subgraph federated learning, is reasonable and has rarely been explored.\n\tThe ideas are technically sound.\n\tExtensive experiments show the superiority of the proposed method compared to other subgraph federated learning methods.\n\nCons:\na)  The idea that performs subgraph federated learning among similar subgraphs is relatively incremental, limiting this paper's novelty.\nb)  Some notations are confusing such as θ vs θ_i on page 4.\nc)  The functional embeddings are key ingredients in the proposed FED-PUB, and the authors utilize a randomly sampled graph as the input of GNNs at each client. However, the influence of different random graphs is missing.\nd)  At the beginning of model training, the functional embeddings derived from each client may not be accurate, resulting in the inaccurate grouping of communities. How does the proposed method deal with this situation?\ne)  The description of baselines is incorrect. For example, FedPerGNN/FedGNN actually introduces a trusted third-party server to augment subgraphs, rather than based on the similarity between the nodes in the local client and the nodes in other clients. Furthermore, FedGNN is a method based on heterogeneous user-item bipartite graphs, and it is unclear how the authors apply it to homogeneous graphs (e.g., Cora). Therefore, it is better to provide a detailed description of experimental settings.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Unlike the existing subgraph federated learning methods that overlook the heterogeneity among subgraphs, this paper proposes a personalized subgraph federated learning framework named FED-PUB, which enhances collaborative learning of subgraphs within the same community. To this end, it captures the communities based on functional embeddings of clients, then employs personalized sparse masks on the local GNN. Empirical results on six public datasets demonstrate the effectiveness of the proposed method. ", "strength_and_weaknesses": "Pros:\n\tThe motivation of this paper, i.e., considering the heterogeneity among subgraphs in the subgraph federated learning, is reasonable and has rarely been explored.\n\tThe ideas are technically sound.\n\tExtensive experiments show the superiority of the proposed method compared to other subgraph federated learning methods.\n\nCons:\na)  The idea that performs subgraph federated learning among similar subgraphs is relatively incremental, limiting this paper's novelty.\nb)  Some notations are confusing such as θ vs θ_i on page 4.\nc)  The functional embeddings are key ingredients in the proposed FED-PUB, and the authors utilize a randomly sampled graph as the input of GNNs at each client. However, the influence of different random graphs is missing.\nd)  At the beginning of model training, the functional embeddings derived from each client may not be accurate, resulting in the inaccurate grouping of communities. How does the proposed method deal with this situation?\ne)  The description of baselines is incorrect. For example, FedPerGNN/FedGNN actually introduces a trusted third-party server to augment subgraphs, rather than based on the similarity between the nodes in the local client and the nodes in other clients. Furthermore, FedGNN is a method based on heterogeneous user-item bipartite graphs, and it is unclear how the authors apply it to homogeneous graphs (e.g., Cora). Therefore, it is better to provide a detailed description of experimental settings.\n", "clarity,_quality,_novelty_and_reproducibility": "a)\tClarity: this paper is well-written and easy to follow.\nb)\tQuality: the quality of this paper is good.\nc)\tNovelty: the idea is quite incremental, which limits the novelty of this paper.\nd)\tReproducibility: The authors have attached the source codes in the supplementary file, \n", "summary_of_the_review": "Due to the limited technical novelty and misleading descriptions of related works, I recommend reject.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666352812053}, {"id": "Ns2q9TvM0ht", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1247/Reviewer_jjc2"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the non-iid subgraphs of the clients in an FL course and proposes a novel pFL algorithm to tackle this challenge. The idea of measuring subgraph similarity via functional embeddings' similarities is novel, and the adaptive masking is synergistic and seems to be helpful. The authors have conducted extensive empirical studies, which confirms the effectiveness of FED-PUB and the rationale behind its components.", "review_text": "This paper is completed and \"standard\". Techniques introduced in this paper are naturally motivated and sound helpful. However, the lack of a rigorous theoretical analysis makes me feel it is still below the bar of ICLR. I strongly encourage the authors to supplement such an indispensable part.", "strengths": "Strengths:\n1. This paper is well-written, with each figure, table, and paragraph being very self-contained.\n2. The proposed method, FED-PUB, is the first one to exploit the functional embeddings for measuring subgraph similarities. Meanwhile, the authors introduce the client-wise learnable mask to further personalized the GNN models, in addition to the subgraph similarity-based cluster-based pFL aggregation.\n3. The empirical studies in this paper is solid, with SOTA federated graph learning algorithm as a baseline, public datasets, and in-depth analysis from various perspectives.\n\nWeaknesses:\n1. It seems that there is no theoretical analysis about the effect of similarity measure, not to mention the eventual generalization risk. Actually, the relationship between graph data and the model parameters is intricate. Although the similarity in functional space seems to be a reasonable proxy, why it is useful needs to be analyzed and answered.\n2. The masking scheme is not that novel to personalized federated learning. Thus, the technical contribution(s) of this paper seems to be insufficient (only the subgraph similarity measurement).", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the non-iid subgraphs of the clients in an FL course and proposes a novel pFL algorithm to tackle this challenge. The idea of measuring subgraph similarity via functional embeddings' similarities is novel, and the adaptive masking is synergistic and seems to be helpful. The authors have conducted extensive empirical studies, which confirms the effectiveness of FED-PUB and the rationale behind its components.", "strength_and_weaknesses": "Strengths:\n1. This paper is well-written, with each figure, table, and paragraph being very self-contained.\n2. The proposed method, FED-PUB, is the first one to exploit the functional embeddings for measuring subgraph similarities. Meanwhile, the authors introduce the client-wise learnable mask to further personalized the GNN models, in addition to the subgraph similarity-based cluster-based pFL aggregation.\n3. The empirical studies in this paper is solid, with SOTA federated graph learning algorithm as a baseline, public datasets, and in-depth analysis from various perspectives.\n\nWeaknesses:\n1. It seems that there is no theoretical analysis about the effect of similarity measure, not to mention the eventual generalization risk. Actually, the relationship between graph data and the model parameters is intricate. Although the similarity in functional space seems to be a reasonable proxy, why it is useful needs to be analyzed and answered.\n2. The masking scheme is not that novel to personalized federated learning. Thus, the technical contribution(s) of this paper seems to be insufficient (only the subgraph similarity measurement).", "clarity,_quality,_novelty_and_reproducibility": "Clarity: It is easy for me to follow the story. All the components of the proposed method is well defined and explained at least intuitively.\n\nQuality: The proposed techniques are well-motivated. Their design seems reasonable, at least from an intuitive aspect. The research questions answered in experiments are tightly related to the proposed method. However, without a theoretical analysis, I cannot persuade myself that the subgraph measurement must be helpful for clustering the clients. BTW, the discussions of related works could be further supplemented to be more comprehensive.\n\nNovelty: FED-PUB mainly consists of two main components: one is the subgraph measurement to promote cluster-based personalized federated learning; another is the adaptive masking. The measurement utilizes the similarity in functional space, which is novel to me, in the jargon of graph learning (actually, it has been extensively studied in general machine learning). However, the adaptive masking has been observed in recent personalized federated learning works.\n\nReproducibility: All the adopted datasets are publicly available. All the considered baselines have open-sourced implementation. The experimental protocols and results are detailed.", "summary_of_the_review": "This paper is completed and \"standard\". Techniques introduced in this paper are naturally motivated and sound helpful. However, the lack of a rigorous theoretical analysis makes me feel it is still below the bar of ICLR. I strongly encourage the authors to supplement such an indispensable part.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666323679972}], "openreview_url": "https://openreview.net/forum?id=ToYi8C6fetv", "arxiv_id": "2206.10206", "paper_pdf": "papers/ToYi8C6fetv.pdf", "paper_pdf_sha256": "62be180d31ff55ad879003a3ad720e6c6649f5993e63b93423a1d0e433af7c6f", "paper_pdf_bytes": 1086170, "paper_pdf_source": "openreview", "code_url": "https://github.com/JinheonBaek/FED-PUB", "code_repository": "JinheonBaek/FED-PUB", "code_commit": "302980ded3bb2b719fd0864e29607031372a3e6d", "code_archive": "repos/ToYi8C6fetv.zip", "code_archive_sha256": "857ae2a25785a72970fe9b20424ea53217ec3d2249d808deb3108d55f11fd6ae", "code_archive_bytes": 191607, "code_file_count": 21, "code_extensions": {".py": 18, ".sh": 2, ".ipynb": 1}, "github_disk_usage_kb": 265, "github_languages": {"Jupyter Notebook": 234354, "Python": 67609, "Shell": 1432}, "github_archived": false, "github_pushed_at": "2023-07-02T07:20:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/personalized-subgraph-federated-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RVdN1-eDZ1b", "year": 2022, "status": "rejected", "title": "Plug-In Inversion: Model-Agnostic Inversion for Vision with Data Augmentations", "authors": ["Amin Ghiasi", "Hamid Kazemi", "Steven Reich", "Chen Zhu", "Micah Goldblum", "Tom Goldstein"], "authorids": ["~Amin_Ghiasi1", "~Hamid_Kazemi1", "~Steven_Reich1", "~Chen_Zhu2", "~Micah_Goldblum1", "~Tom_Goldstein1"], "authors_source": "OpenReview API", "abstract": "Existing techniques for model inversion typically rely on hard-to-tune regularizers, such as total variation or feature regularization, which must be individually calibrated for each network in order to produce adequate images. In this work, we introduce Plug-In Inversion, which relies on a simple set of augmentations and does not require excessive hyper-parameter tuning.  Under our proposed augmentation-based scheme, the same set of augmentation hyper-parameters can be used for inverting a wide range of image classification models, regardless of input dimensions or the architecture. We illustrate the practicality of our approach by inverting Vision Transformers (ViTs) and Multi-Layer Perceptrons (MLPs) trained on the ImageNet dataset, tasks which to the best of our knowledge have not been successfully accomplished by any previous works. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "NOQ4ee3nv_L", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2095/Reviewer_dsJC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper attempts to solve the problem of image inversion by introducing three augmentation-based techniques. It specifically tries to solve the problem of class inversion, which generates an interpretable image from a neural network by sending the pre-initialized image to the network. Then optimization is wrt the input image instead of the weights of the network. It introduced new sets of augmentation-based techniques like centering, zooming, color shift augmentation, ensembling that have not been tried before, and finally, they combined them with VIT and MLP based vision models. They validated all the augmentations using Vision transformer and MLP based vision models and compared them against the exiting method Deep Inversion(Yin et al., 2020). The paper is well written but lack quantitative evaluation which is a very significant shortcoming. ", "review_text": "# Highlights: \nThe problem addressed in the paper is well written and well described in the Introduction and Background section.\n\n# Related work: \nThere should be a separate Related work section citing the work that has been done in this area previously with a thorough literature review. \n\n\n# Methods: \n1.\tThe paper mentioned several drawbacks in the Introduction section of the existing methods - the authors should put any of such examples like the generated images being highly sensitive to the weights assigned to the regularized terms, etc, and compare that with the images generated by PII to show how PII is working better.\n2.\tThe authors should have provided some theoretical justifications of their work -  why are the new sets of augmentation working better than its predecessor?\n\n# Experiments:\n\t\n1.\tOne of the main issues in this paper is the applicability of the solution of the problem is not discussed in the paper. For example, there are three major applications where Deep Inversion(Yin et al., 2020) can be used - (a) pruning, (b) knowledge transfer, and (c) continual (class incremental) learning. If the authors set Deep Inversion as the baseline, they should have formulated the experiments for these three applications and compared PII with all architectures like VIT and MLP and others against Deep Inversion. Also if there are any other applications of the solution, that should have been clearly mentioned in the paper. \n\n2.\tThe authors experimented only using the Imagenet dataset. They should have done against other datasets like CIFAR 10 - the way Deep Inversion was doing.\n\n3.\tTo prove such a simple method is working better than the previous models, the authors should have either used or introduced some metric that gives a quantitative comparison between all the inversion techniques. In this paper, the authors compared PII with different settings and against Deep Inversion qualitatively. For example, in Deep Inversion, the authors did a wide range of quantitative comparisons among various metrics like inception score, knowledge transfer results, continual learning results (table 1-7 Deep Inversion (Yin et al., 2020))\n\n# Minor:\n1.\tSection 2.3, 2nd paragraph where Vision transformers are discussed - in line #2, there is a mistake \\cite{Vaswani}\n2.\tSame for the last line in the same paragraph \\cite{CoaT and CaiT}\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper attempts to solve the problem of image inversion by introducing three augmentation-based techniques. It specifically tries to solve the problem of class inversion, which generates an interpretable image from a neural network by sending the pre-initialized image to the network. Then optimization is wrt the input image instead of the weights of the network. It introduced new sets of augmentation-based techniques like centering, zooming, color shift augmentation, ensembling that have not been tried before, and finally, they combined them with VIT and MLP based vision models. They validated all the augmentations using Vision transformer and MLP based vision models and compared them against the exiting method Deep Inversion(Yin et al., 2020). The paper is well written but lack quantitative evaluation which is a very significant shortcoming. ", "main_review": "# Highlights: \nThe problem addressed in the paper is well written and well described in the Introduction and Background section.\n\n# Related work: \nThere should be a separate Related work section citing the work that has been done in this area previously with a thorough literature review. \n\n\n# Methods: \n1.\tThe paper mentioned several drawbacks in the Introduction section of the existing methods - the authors should put any of such examples like the generated images being highly sensitive to the weights assigned to the regularized terms, etc, and compare that with the images generated by PII to show how PII is working better.\n2.\tThe authors should have provided some theoretical justifications of their work -  why are the new sets of augmentation working better than its predecessor?\n\n# Experiments:\n\t\n1.\tOne of the main issues in this paper is the applicability of the solution of the problem is not discussed in the paper. For example, there are three major applications where Deep Inversion(Yin et al., 2020) can be used - (a) pruning, (b) knowledge transfer, and (c) continual (class incremental) learning. If the authors set Deep Inversion as the baseline, they should have formulated the experiments for these three applications and compared PII with all architectures like VIT and MLP and others against Deep Inversion. Also if there are any other applications of the solution, that should have been clearly mentioned in the paper. \n\n2.\tThe authors experimented only using the Imagenet dataset. They should have done against other datasets like CIFAR 10 - the way Deep Inversion was doing.\n\n3.\tTo prove such a simple method is working better than the previous models, the authors should have either used or introduced some metric that gives a quantitative comparison between all the inversion techniques. In this paper, the authors compared PII with different settings and against Deep Inversion qualitatively. For example, in Deep Inversion, the authors did a wide range of quantitative comparisons among various metrics like inception score, knowledge transfer results, continual learning results (table 1-7 Deep Inversion (Yin et al., 2020))\n\n# Minor:\n1.\tSection 2.3, 2nd paragraph where Vision transformers are discussed - in line #2, there is a mistake \\cite{Vaswani}\n2.\tSame for the last line in the same paragraph \\cite{CoaT and CaiT}\n", "summary_of_the_review": "* Since the baseline for this paper is Deep Inversion lack and there are many quantitative experiments in that paper, the lack of quantitative results in this paper is not justifiable.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635904879564}, {"id": "EgleSTnFGEE", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2095/Reviewer_MNJX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method to perform class inversion on image data, called Plug-In Inversion. This method consists of a sequence of augmentations on image data and is designed to be applicable to a variety of architectures. This method is evaluated on ImageNet trained models and is compared with other techniques used for the same goal.", "review_text": "This paper proposes a technique called Plug-In Inversion, which is to be used as-is, irrespective of the underlying model which is to be inverted. This method works by applying a particular sequence of augmentations to the image, and then performing inversion by minimizing the loss of this image with respect to a single class.\n\nStrengths:\n- The motivation behind the paper is clearly presented - the goal is to reduce the need for extensive hyperparameter optimization of an inversion system. This is an interesting motivation, given that extensive tuning of regularizers is often difficult to perform.\n- The paper is in general clear and easy to follow. The main ideas are presented in a concise manner and formulate a solid story for the paper.\n- The authors demonstrate the effects of each different augmentation used in their inversion system, both in the main paper and the appendix. This is particularly useful both for understanding how Plug-In Inversion works and how each part of the procedure affects the resulting image (namely, how the zoom + centering procedure operates, as well as the effect of the ColorShift augmentation).\n- The proposed method is evaluated on a great variety of models, and produces intelligible images for the classes presented (in the sense that they conform to the high-level notion related to the class).\n\nWeaknesses:\n- The proposed Plug-In inversion method is introduced very late in the paper (at the end of Section 3.4) even though it consists only of combining the augmentation and search space restriction techniques provided in the previous sections (3.1 - 3.3). It would have been much clearer for the reader if this fact was fully described earlier in Section 3 (for example, by moving Section 3.4 earlier in the paper), instead of delaying the presentation of the full method.\n- My greatest concern is the fact that I am not sure whether the claim of the authors that the method can be applied to multiple networks without tuning is fully supported by the experimental results. More specifically, in Figures 7 and 8 the method is applied in a variety of networks, resulting in images which, while intelligible, are of varying quality. As such, to fully support the argument of the lack of need of extensive hyperparameter tuning, one would also need to show that using the same regularizer (for example, TV, which can be applied to any model) with the same hyperparameter leads to more extensive degradation, when applied to different models. I believe that a small example to demonstrate this motivation would greatly improve the paper.\n- The above point is made even more unclear by the fact that one of the models considered by the authors does use a TV regularizer (which leads to drastic improvement in image quality).\n- The ColorShift augmentation proposed by the authors is, unless I am mistaken, a variant of color jittering (in the sense that the adjustment is made directly to the pixels of the image, rather than on hue, saturation, contrast etc.). Given that applying some form of adjustment to the color of the image is a form of data augmentation which has been used in prior work  (Krizhevsky et al., 2012, section 4.1), I am not sure if this data augmentation is as novel as the authors claim (although I do appreciate the qualitative analysis of its effects in the context of inversion).\n\nQuestions:\n-\tThe hyperparameters for ColorShift are fixed to $a=b=1$, and if I understand correctly this comes from a qualitative analysis of the results in a few simple experiments. Is this correct?\n\nMinor comments/typos:\n-\tAs a minor comment related to the above, I believe the authors should indicate in the captions of the figures which model the respective images come from.\n-\tThere is a space missing in the second-to-last line of page 6.\n-\tThe caption in Figure 10 in the appendix is incomplete.\n\nReferences:\nKrizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, 25, 1097-1105.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a method to perform class inversion on image data, called Plug-In Inversion. This method consists of a sequence of augmentations on image data and is designed to be applicable to a variety of architectures. This method is evaluated on ImageNet trained models and is compared with other techniques used for the same goal.", "main_review": "This paper proposes a technique called Plug-In Inversion, which is to be used as-is, irrespective of the underlying model which is to be inverted. This method works by applying a particular sequence of augmentations to the image, and then performing inversion by minimizing the loss of this image with respect to a single class.\n\nStrengths:\n- The motivation behind the paper is clearly presented - the goal is to reduce the need for extensive hyperparameter optimization of an inversion system. This is an interesting motivation, given that extensive tuning of regularizers is often difficult to perform.\n- The paper is in general clear and easy to follow. The main ideas are presented in a concise manner and formulate a solid story for the paper.\n- The authors demonstrate the effects of each different augmentation used in their inversion system, both in the main paper and the appendix. This is particularly useful both for understanding how Plug-In Inversion works and how each part of the procedure affects the resulting image (namely, how the zoom + centering procedure operates, as well as the effect of the ColorShift augmentation).\n- The proposed method is evaluated on a great variety of models, and produces intelligible images for the classes presented (in the sense that they conform to the high-level notion related to the class).\n\nWeaknesses:\n- The proposed Plug-In inversion method is introduced very late in the paper (at the end of Section 3.4) even though it consists only of combining the augmentation and search space restriction techniques provided in the previous sections (3.1 - 3.3). It would have been much clearer for the reader if this fact was fully described earlier in Section 3 (for example, by moving Section 3.4 earlier in the paper), instead of delaying the presentation of the full method.\n- My greatest concern is the fact that I am not sure whether the claim of the authors that the method can be applied to multiple networks without tuning is fully supported by the experimental results. More specifically, in Figures 7 and 8 the method is applied in a variety of networks, resulting in images which, while intelligible, are of varying quality. As such, to fully support the argument of the lack of need of extensive hyperparameter tuning, one would also need to show that using the same regularizer (for example, TV, which can be applied to any model) with the same hyperparameter leads to more extensive degradation, when applied to different models. I believe that a small example to demonstrate this motivation would greatly improve the paper.\n- The above point is made even more unclear by the fact that one of the models considered by the authors does use a TV regularizer (which leads to drastic improvement in image quality).\n- The ColorShift augmentation proposed by the authors is, unless I am mistaken, a variant of color jittering (in the sense that the adjustment is made directly to the pixels of the image, rather than on hue, saturation, contrast etc.). Given that applying some form of adjustment to the color of the image is a form of data augmentation which has been used in prior work  (Krizhevsky et al., 2012, section 4.1), I am not sure if this data augmentation is as novel as the authors claim (although I do appreciate the qualitative analysis of its effects in the context of inversion).\n\nQuestions:\n-\tThe hyperparameters for ColorShift are fixed to $a=b=1$, and if I understand correctly this comes from a qualitative analysis of the results in a few simple experiments. Is this correct?\n\nMinor comments/typos:\n-\tAs a minor comment related to the above, I believe the authors should indicate in the captions of the figures which model the respective images come from.\n-\tThere is a space missing in the second-to-last line of page 6.\n-\tThe caption in Figure 10 in the appendix is incomplete.\n\nReferences:\nKrizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, 25, 1097-1105.\n", "summary_of_the_review": "Overall, this work has an interesting motivation (alleviating the need for extensive tuning of regularizers for class inversion), but I am not certain if the benefits of Plug-In Inversion, as described by the authors, are fully supported by the experimental evidence provided. As such, given also the fact that the individual components of the method do not seem novel by themselves, I marginally lean towards rejecting this paper, pending extra discussion during the rebuttal process.\n\n**Update after rebuttal**: See response to the authors' comments below.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635801941775}, {"id": "2zKVhvKLiqN", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2095/Reviewer_vGig"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors introduce a novel method for the class inversion problem: given a trained classifier and a specific class/label, generate an example image of that class.  This method is called Plug-In Inversion (PII) and it works as follows: a gradient descent is performed to find the image x such that the output of the classifier is close to the target label under a suitable loss function.  In PII, each successive gradient update is computed using a data-augmented version of the current iterate of the image x.  The authors introduce new data augmentations and combinations thereof.  The benefits of this inversion method is that it works for any classification architecture (which is treated only as a black box), and the same hyper parameter values (e.g. amount of color shift) work across a range of significantly different classifier architectures.  The authors present the output of PII across 12 different networks (including Convolutional, Transformer, and MLP nets), and they demonstrate several examples where their method is more interpretable than DeepInversion, a state-of-the-art method for class inversion.  \n", "review_text": "Strengths of the paper:\n- The proposed method works for class inversion for any architecture, including Vision Transformers and MLPs.\n- The proposed method is robust across architectures with little-to-no hyperparameter tuning\n\nWeak points of the paper:\n- The specific method proposed is not clearly presented at a mathematical level (it is described only in words)\n- The only baseline comparison I saw was relative to one method (DeepInversion).  \n\nRequested response from authors that will affect my recommendation:\n\nThe authors should justify why they believe their baseline comparisons are sufficient (or comment on what additional baselines they would present in a camera ready). \n\nThe paper does not comment on methods for adversarial examples, such as adversarial patches, either as related work or as baselines.  In the rebuttal, please justify why these comparisons are not present, or comment about them.\n\nThe authors should clarify what exactly the PII algorithm is at a mathematical level (write out the GD iteration) so that I am clear I am not misunderstanding the model.\n\nAdditional feedback with the aim to improve the paper:\n\nIf my reading is correct, the method treats the classifier as a black box, which is why it works for any architecture.  If this is true, then the paper would be much improved by clearly stating it treats the classifier as a black box.  In its current form, it comes across as being restricted to neural networks, which I do not think is the case.\n\nTypos:\nThe minimization problem on page 2 should say argmin_x L(f(x), y).  It has an incorrect parenthesis.\n\nThere are several places where the document says \"cite XYZ\".  Replace these with the citations.\n\nFigure 2 caption: \"An image a different stages\" has a grammatical issue.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors introduce a novel method for the class inversion problem: given a trained classifier and a specific class/label, generate an example image of that class.  This method is called Plug-In Inversion (PII) and it works as follows: a gradient descent is performed to find the image x such that the output of the classifier is close to the target label under a suitable loss function.  In PII, each successive gradient update is computed using a data-augmented version of the current iterate of the image x.  The authors introduce new data augmentations and combinations thereof.  The benefits of this inversion method is that it works for any classification architecture (which is treated only as a black box), and the same hyper parameter values (e.g. amount of color shift) work across a range of significantly different classifier architectures.  The authors present the output of PII across 12 different networks (including Convolutional, Transformer, and MLP nets), and they demonstrate several examples where their method is more interpretable than DeepInversion, a state-of-the-art method for class inversion.  \n", "main_review": "Strengths of the paper:\n- The proposed method works for class inversion for any architecture, including Vision Transformers and MLPs.\n- The proposed method is robust across architectures with little-to-no hyperparameter tuning\n\nWeak points of the paper:\n- The specific method proposed is not clearly presented at a mathematical level (it is described only in words)\n- The only baseline comparison I saw was relative to one method (DeepInversion).  \n\nRequested response from authors that will affect my recommendation:\n\nThe authors should justify why they believe their baseline comparisons are sufficient (or comment on what additional baselines they would present in a camera ready). \n\nThe paper does not comment on methods for adversarial examples, such as adversarial patches, either as related work or as baselines.  In the rebuttal, please justify why these comparisons are not present, or comment about them.\n\nThe authors should clarify what exactly the PII algorithm is at a mathematical level (write out the GD iteration) so that I am clear I am not misunderstanding the model.\n\nAdditional feedback with the aim to improve the paper:\n\nIf my reading is correct, the method treats the classifier as a black box, which is why it works for any architecture.  If this is true, then the paper would be much improved by clearly stating it treats the classifier as a black box.  In its current form, it comes across as being restricted to neural networks, which I do not think is the case.\n\nTypos:\nThe minimization problem on page 2 should say argmin_x L(f(x), y).  It has an incorrect parenthesis.\n\nThere are several places where the document says \"cite XYZ\".  Replace these with the citations.\n\nFigure 2 caption: \"An image a different stages\" has a grammatical issue.\n", "summary_of_the_review": "Clearly state your recommendation with 1-2 reasons\n\nI rate this paper as a weak accept for the following reasons.\n\nThe reason to accept is: They provide a class inversion method that produces interpretable visualizations for recent architectures (Vision Transformers + MLPs) that existing methods do not work for.  This could accelerate the development of better classifiers by helping researchers interpret and fix their behavior.\n\nThe reason this isn't rated higher is: The baseline comparison was inadequate.  The authors should compare their visualization method on Transformers/MLPs to the best existing approach that exists before this work.  Perhaps that approach is a plain inversion method.  Perhaps there are methods that can be borrowed from the methods of targeted adversarial examples.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635700074800}, {"id": "PYotF3PEX1Y", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2095/Reviewer_SYwW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes Plug-In Inversion (PII), which can invert a trained image classifier so that it generates a class-conditional image.\nIn addition to CNNs that previous work has considered, PII can handle Vision Transformers and Vision MLPs as well.\nPII starts generating from the center of a random initialization with lower resolution. Then, the method gradually broadens the generating area (centering) and increases the resolution (zooming). During generation, colors are randomly shifted (ColorShift) and the averaged cross-entropy to $e$ color-shifted images are used.\n\n", "review_text": "# Strengths\n\n* The authors aim for a model-agnostic and hyperparameter-robust method. This is an important direction.\n* To verify the model-agnostic features, the authors tested the proposed method on various CNNs (MobileNet, ResNet, etc.), Vision Transformers (ViT, DeiT, etc.), and MLPs.\n\n# Weaknesses\n\n* According to Section 7, the purpose of this paper is to understand vision models. I agree with one of the goals of model inversion is so, but the paper lacks a discussion of how PII helps us to understand vision models, even though the authors conducted experiments on various image recognition models.\n* The analysis is limited to qualitative ones. I would include quantitative results to compare PII with other methods. For example, DeepInversion [Yin+20] reported classification accuracy and the Inception score (IS). Santurkar+19 also shows IS.\n* The authors claimed that DeepInversion needs statistics $\\hat{\\mu}_k, \\hat{\\sigma}^2_k$ ported from batch normalization (BN) and thus is not applicable to models sans BN, such as ViT. However, I think that $\\hat{\\mu}_k, \\hat{\\sigma}^2_k$ are statistics of training data and can be obtained even for models sans BN only by passing training data once and storing the activations.\n* PII is only verified on the ImageNet dataset. I recommend the authors to use CIFAR-10 as Yin+20.\n\n# Comments\n\n* References are not well organized and written in a consistent format. For example, Santurkar+19 is accepted at NeurIPS 2019, but it is not written so. \n* An image is denoted as $x$ in p. 4, but $\\mathbf{x}$ in P5. The notation should be consistent, otherwise, it is difficult to follow.\n* Why a total-variation loss is used to visualize Figure 4, while, if I understand correctly, PII does not use the loss term?\n* Without a quantitative measure, how the hyperparameters of PII can be tuned in Section 4?\n* Section 3.3 says $e=8$ results in images of acceptable quality, while in the main experiment, $e$ is set to 32. If it is intended, it would be justified.\n* What are \"random augmentations\" in p. 6 Section 4?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes Plug-In Inversion (PII), which can invert a trained image classifier so that it generates a class-conditional image.\nIn addition to CNNs that previous work has considered, PII can handle Vision Transformers and Vision MLPs as well.\nPII starts generating from the center of a random initialization with lower resolution. Then, the method gradually broadens the generating area (centering) and increases the resolution (zooming). During generation, colors are randomly shifted (ColorShift) and the averaged cross-entropy to $e$ color-shifted images are used.\n\n", "main_review": "# Strengths\n\n* The authors aim for a model-agnostic and hyperparameter-robust method. This is an important direction.\n* To verify the model-agnostic features, the authors tested the proposed method on various CNNs (MobileNet, ResNet, etc.), Vision Transformers (ViT, DeiT, etc.), and MLPs.\n\n# Weaknesses\n\n* According to Section 7, the purpose of this paper is to understand vision models. I agree with one of the goals of model inversion is so, but the paper lacks a discussion of how PII helps us to understand vision models, even though the authors conducted experiments on various image recognition models.\n* The analysis is limited to qualitative ones. I would include quantitative results to compare PII with other methods. For example, DeepInversion [Yin+20] reported classification accuracy and the Inception score (IS). Santurkar+19 also shows IS.\n* The authors claimed that DeepInversion needs statistics $\\hat{\\mu}_k, \\hat{\\sigma}^2_k$ ported from batch normalization (BN) and thus is not applicable to models sans BN, such as ViT. However, I think that $\\hat{\\mu}_k, \\hat{\\sigma}^2_k$ are statistics of training data and can be obtained even for models sans BN only by passing training data once and storing the activations.\n* PII is only verified on the ImageNet dataset. I recommend the authors to use CIFAR-10 as Yin+20.\n\n# Comments\n\n* References are not well organized and written in a consistent format. For example, Santurkar+19 is accepted at NeurIPS 2019, but it is not written so. \n* An image is denoted as $x$ in p. 4, but $\\mathbf{x}$ in P5. The notation should be consistent, otherwise, it is difficult to follow.\n* Why a total-variation loss is used to visualize Figure 4, while, if I understand correctly, PII does not use the loss term?\n* Without a quantitative measure, how the hyperparameters of PII can be tuned in Section 4?\n* Section 3.3 says $e=8$ results in images of acceptable quality, while in the main experiment, $e$ is set to 32. If it is intended, it would be justified.\n* What are \"random augmentations\" in p. 6 Section 4?", "summary_of_the_review": "The model aims for a model-agnostic and hyperparameter-robust method to investigate vision models by model inversion. I highly appreciate this goal. Meanwhile, the current manuscript has several weaknesses. Therefore, I recommend this paper be marginally below the threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635433948182}], "openreview_url": "https://openreview.net/forum?id=RVdN1-eDZ1b", "arxiv_id": "2201.12961", "paper_pdf": "papers/RVdN1-eDZ1b.pdf", "paper_pdf_sha256": "759cb6ece49c24d5b93639674fb8d7ff304f925ba008e9ec70f9470df059c14e", "paper_pdf_bytes": 23229715, "paper_pdf_source": "openreview", "code_url": "https://github.com/youranonymousefriend/plugininversion", "code_repository": "youranonymousefriend/plugininversion", "code_commit": "1141f08f79f713b5ecc2b77fb9814b08fe762c34", "code_archive": "repos/RVdN1-eDZ1b.zip", "code_archive_sha256": "7f1e1118891fb8f7aafa66adb713a5f2d38c057c67edae7b126cc8b90dc14892", "code_archive_bytes": 742652, "code_file_count": 179, "code_extensions": {".py": 113, ".sh": 54, ".js": 12}, "github_disk_usage_kb": 640, "github_languages": {"Python": 139276, "Shell": 136312}, "github_archived": false, "github_pushed_at": "2021-10-23T21:16:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/plug-in-inversion-model-agnostic-inversion-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DE0MSwKv32y", "year": 2021, "status": "rejected", "title": "Trust, but verify: model-based exploration in sparse reward environments", "authors": ["Konrad Czechowski", "Tomasz Odrzygóźdź", "Michał Izworski", "Marek Zbysiński", "Łukasz Kuciński", "Piotr Miłoś"], "authorids": ["~Konrad_Czechowski1", "tomaszo@impan.pl", "m.izworski@student.uw.edu.pl", "marek.zbysinski@gmail.com", "~Łukasz_Kuciński1", "~Piotr_Miłoś1"], "authors_source": "OpenReview API", "abstract": "We propose $\\textit{trust-but-verify}$ (TBV) mechanism, a new method which uses model uncertainty estimates to guide  exploration. The mechanism augments graph search planning algorithms by the capacity to deal with learned model's imperfections. We identify certain type of frequent model errors, which we dub $\\textit{false loops}$, and which are particularly dangerous for graph search algorithms in discrete environments. These errors impose falsely pessimistic expectations and thus hinder exploration. We confirm this experimentally and show that TBV can effectively alleviate them. TBV combined with MCTS or Best First Search forms an effective model-based reinforcement learning solution, which is able to robustly solve sparse reward problems. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ytdyrcImd0L", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3396/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present a method to guide exploration that prefers to go to areas of the state space for which it is more uncertain. This uncertainty is obtained by measuring the standard deviation of the next state prediction from an ensemble of models. The authors call this the disagreement measure At each step, a search is performed and the disagreement measure is obtained for each state visited. The disagreement measure for each action is compared to the distribution for all the states visited during the search. It it is above some threshold, then the action that maximizes the disagreement measure is taken. Otherwise, it takes the action determined by the search.\n\nThe algorithm presented was unclear. What does planner.choose_action do? Is the heuristic for best-first search (BFS) the disagreement measure? I don't understand how this should help the algorithm pick a good action to take.  The paper says, \"The proposed action is the first edge on the shortest path to the best node in the subgraph searched so far.\" How is the best node determined?\n\nFurthermore, is this search necessary? It seems like it is mainly used as a comparison for the disagreement measure. What if the agent behaved greedily with respect to the disagreement measure all the time. Pathak et al. (2017) used a similar method, but with an inverse model.\n\nI am not quite sure about the comparisons the authors are making. In the case of BFS search, what does it mean to do BFS search with epsilon greedy? Also, this is an artificial curiosity method where curiosity is measured by the disagreement between the ensemble of models. However, there are no comparisons to other curiosity papers, such as Pathak et al. (2017).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A method for artificial curiosity using model uncertainty", "review": "The authors present a method to guide exploration that prefers to go to areas of the state space for which it is more uncertain. This uncertainty is obtained by measuring the standard deviation of the next state prediction from an ensemble of models. The authors call this the disagreement measure At each step, a search is performed and the disagreement measure is obtained for each state visited. The disagreement measure for each action is compared to the distribution for all the states visited during the search. It it is above some threshold, then the action that maximizes the disagreement measure is taken. Otherwise, it takes the action determined by the search.\n\nThe algorithm presented was unclear. What does planner.choose_action do? Is the heuristic for best-first search (BFS) the disagreement measure? I don't understand how this should help the algorithm pick a good action to take.  The paper says, \"The proposed action is the first edge on the shortest path to the best node in the subgraph searched so far.\" How is the best node determined?\n\nFurthermore, is this search necessary? It seems like it is mainly used as a comparison for the disagreement measure. What if the agent behaved greedily with respect to the disagreement measure all the time. Pathak et al. (2017) used a similar method, but with an inverse model.\n\nI am not quite sure about the comparisons the authors are making. In the case of BFS search, what does it mean to do BFS search with epsilon greedy? Also, this is an artificial curiosity method where curiosity is measured by the disagreement between the ensemble of models. However, there are no comparisons to other curiosity papers, such as Pathak et al. (2017).", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603994109053}, {"id": "AdWpIwRGmV", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3396/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work tackled the problem of model-based RL in environments where the reward is sparse and many actions are needed to achieve some. Particularly, the authors tried to solve the issue of one-step false loop in the model, which avoids further exploration. Measuring the uncertainty about the built model through ensemble of models, they added a possibility of choosing an action different from what planner suggests, promoting exploration. The work is very-well written in general, especially sections of problem definition and related work. I also appreciate that the proposed method is compared with multiple planners and tested on two different tasks. Having said that, the main missing analysis for me is that the method was not tested on environments where false loop does not exist. Given the nature of the problem definition, i.e. learning the environment, it is counter-intuitive not to test the method on a few standard test-benchmarks without any assumption. The proposed method does not have to get the best result on environments without false loop, but it is important to see how it behaves when the built model is already good. Other than this, I have a few more questions/concerns: \n\n1) The RANDOM parameter seems a little strange, especially because it looks too high, i.e. .5. Some analysis on different values of this (or just with and without RANDOM) on performance would be great. Also, I suspect that change in RANDOM would also change the best QR. I think a plot similar to figure 7, but for RANDOM and combination of RANDOM and QR would improve the paper.\n\n2) The method is about one-step false loop. I would appreciate if the authors talk about multistep false loop briefly. Could it be problematic in learning? If yes, could an extension of this method work?\n\n3) Have the authors considered a QR that changes with number of steps? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting and well-written paper on model-based exploration", "review": "This work tackled the problem of model-based RL in environments where the reward is sparse and many actions are needed to achieve some. Particularly, the authors tried to solve the issue of one-step false loop in the model, which avoids further exploration. Measuring the uncertainty about the built model through ensemble of models, they added a possibility of choosing an action different from what planner suggests, promoting exploration. The work is very-well written in general, especially sections of problem definition and related work. I also appreciate that the proposed method is compared with multiple planners and tested on two different tasks. Having said that, the main missing analysis for me is that the method was not tested on environments where false loop does not exist. Given the nature of the problem definition, i.e. learning the environment, it is counter-intuitive not to test the method on a few standard test-benchmarks without any assumption. The proposed method does not have to get the best result on environments without false loop, but it is important to see how it behaves when the built model is already good. Other than this, I have a few more questions/concerns: \n\n1) The RANDOM parameter seems a little strange, especially because it looks too high, i.e. .5. Some analysis on different values of this (or just with and without RANDOM) on performance would be great. Also, I suspect that change in RANDOM would also change the best QR. I think a plot similar to figure 7, but for RANDOM and combination of RANDOM and QR would improve the paper.\n\n2) The method is about one-step false loop. I would appreciate if the authors talk about multistep false loop briefly. Could it be problematic in learning? If yes, could an extension of this method work?\n\n3) Have the authors considered a QR that changes with number of steps? ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603925718116}, {"id": "nAKAovmTOnF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3396/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a new method which can be combined with graph search algorithms to boost exploration when the uncertainty is high. This new mechanism, called TBV, can override actions given by the model to explore and verify model predictions. It is also shown in the experiments that TBV improves the model performance when combined with MCTS or BestFS. TBV utilizes graph structure of the problem and finds the solution much quicker for both MCTS and BestFS. \n\nWhile the presented method is interesting with high performance, I found many editorial errors in the writing. For example, in the second paragraph of section 3.3, ‘we concentrated of exploration….’, and ‘In our experiments, such a version proved to be effective in in discrete…’, just to name a few. There are so many errors like this and the paper needs serious rewriting. Also, having a conclusion or discussion can help the structure of the paper. \n\nFigure 3 is unclear if the blue line is without TBV with the legend ‘Right corridor visited’. It could be interesting to discuss extension of TBV into continuous environments. Reference format seems to have errors since there are underlines. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Trust, but verify: model-based exploration in sparse reward environments", "review": "This paper presents a new method which can be combined with graph search algorithms to boost exploration when the uncertainty is high. This new mechanism, called TBV, can override actions given by the model to explore and verify model predictions. It is also shown in the experiments that TBV improves the model performance when combined with MCTS or BestFS. TBV utilizes graph structure of the problem and finds the solution much quicker for both MCTS and BestFS. \n\nWhile the presented method is interesting with high performance, I found many editorial errors in the writing. For example, in the second paragraph of section 3.3, ‘we concentrated of exploration….’, and ‘In our experiments, such a version proved to be effective in in discrete…’, just to name a few. There are so many errors like this and the paper needs serious rewriting. Also, having a conclusion or discussion can help the structure of the paper. \n\nFigure 3 is unclear if the blue line is without TBV with the legend ‘Right corridor visited’. It could be interesting to discuss extension of TBV into continuous environments. Reference format seems to have errors since there are underlines. ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603862478895}, {"id": "l5v-joks3-l", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3396/AnonReviewer3"], "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an approach for encouraging exploration when planning over learned models of discrete reinforcement learning environment. The proposed method involves using an uncertainty-aware model (e.g., an ensemble of neural networks) to predict state-action transitions, together with a graph-based planner operating on this model. The key idea is to replace (with some probability) the planner's action with the action leading to the highest uncertainty in model prediction. The paper evaluates the proposed technique using two standard search planners (MCTS and BFS). \n\nUnfortunately, I think the significance and technical contribution of this work is minimal, an issue that mostly likely starts from a deficient literature review. What the authors refer to as Trust-But-Verify, it's just an ad-hoc instance of the well-known principle of *optimism in the face of uncertainty*, which underlies classic bandit and RL algorithms such as UCB1[1], UCT[2], Thompson Sampling [3, 4]. In the model-free setting, this idea has lead to numerous recent algorithms, many of which also use ensembles for uncertainty quantification [5-8]. In the model-based setting there are also some precedents of work using similar ideas [9-11]. There is also a large body of work treating the problem from the point of view of Bayesian RL (see [12] for a survey).  It is a bad sign that none of this body of previous work was discussed in the paper, which I would argue was the more relevant literature upon which the paper had to be positioned.\n\nThis could conceivably be excused if the paper technical and experimental contribution was impressive enough, but this is not the case.  In contrast to the literature outlined above (where proposed exploration strategies typically follow from rigorous statistical analysis), this work presents the proposed method as a heuristic rule, providing no insight as to why one should expect the approach to work well in general. Moreover, the experiments are done in relatively simple domain, and compared against simple baselines. Some of the baseline choices do not seem appropriate either. For example, why use $\\epsilon$-greedy for exploration, instead of more robust strategies using upper confidence bounds? \n\nFinally, the writing on the paper can be improved in many places. For example, the paper refers to using the graph structure of the underlying problem, but what this graph structure refers to is not properly defined anywhere in the paper. I imagine it refers to the graph wherein edges represent non-zero probability transitions between states, but this is not clear from the text. Additionally, some paragraphs add little in terms of content. For example, the first paragraph of Section 3 is devoted to describe the basic problem that all model-based RL methods are trying to solve; this issue is ubiquitous so there is no need for a full example and so much text to describe this. Other sentences are unclear, such as \"The optimistic and pessimistic errors are often of the same nature\", which I am not sure what is referring to . Additionally, I didn't see a description of the learned model used in the experiments, which is not an obvious choice, since the environment is discrete. \n\nOverall, to end on a somewhat constructive note, I think the problem the authors are trying to solve is interesting and the proposed approach is based on the right intuitions. However, this work is still too immature for publication. I suggest to the authors to position their work properly with regards to the relevant literature, refine their technical contribution accordingly, and compare with more appropriate baselines.  \n\n----------------------------------------------------------------------\n[1] Auer, Peter, Nicolo Cesa-Bianchi, and Paul Fischer. \"Finite-time analysis of the multiarmed bandit problem.\" Machine learning 47.2-3 (2002): 235-256.\n\n[2] Kocsis, Levente, and Csaba Szepesvári. \"Bandit based monte-carlo planning.\" European conference on machine learning. Springer, Berlin, Heidelberg, 2006.\n\n[3] Thompson, William R. \"On the likelihood that one unknown probability exceeds another in view of the evidence of two samples.\" Biometrika 25.3/4 (1933): 285-294.\n\n[4] Russo, Daniel, et al. \"A tutorial on thompson sampling.\" arXiv preprint arXiv:1707.02038 (2017).\n\n[5] Bellemare, M., Srinivasan, S., Ostrovski, G., Schaul, T., Saxton, D., & Munos, R. (2016). Unifying count-based exploration and intrinsic motivation. In Advances in neural information processing systems (pp. 1471-1479).\n\n[6] Ostrovski, G., Bellemare, M. G., Oord, A., & Munos, R. (2017, July). Count-Based Exploration with Neural Density Models. In International Conference on Machine Learning (pp. 2721-2730).\n\n[7] Osband, I., Blundell, C., Pritzel, A., & Van Roy, B. (2016). Deep exploration via bootstrapped DQN. In Advances in neural information processing systems (pp. 4026-4034).\n\n[8] Fortunato, M., Azar, M. G., Piot, B., Menick, J., Osband, I., Graves, A., ... & Blundell, C. (2017). Noisy networks for exploration. arXiv preprint arXiv:1706.10295.\n\n[9] Sanner, S., Goetschalckx, R., Driessens, K., & Shani, G. (2009). Bayesian real-time dynamic programming. In Proceedings of the 21st International Joint Conference on Artificial Intelligence (IJCAI-09) (pp. 1784-1789). IJCAI-INT JOINT CONF ARTIF INTELL.\n\n[10] Pathak, Deepak, Dhiraj Gandhi, and Abhinav Gupta. \"Self-supervised exploration via disagreement.\" arXiv preprint arXiv:1906.04161 (2019).\n\n[11] Shyam, Pranav, Wojciech Jaśkowski, and Faustino Gomez. \"Model-based active exploration.\" International Conference on Machine Learning. 2019.\n\n[12] Ghavamzadeh, M., Mannor, S., Pineau, J., & Tamar, A. (2016). Bayesian reinforcement learning: A survey. arXiv preprint arXiv:1609.04436.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Improperly positioned, minor technical contribution, unclearly written", "review": "This paper proposes an approach for encouraging exploration when planning over learned models of discrete reinforcement learning environment. The proposed method involves using an uncertainty-aware model (e.g., an ensemble of neural networks) to predict state-action transitions, together with a graph-based planner operating on this model. The key idea is to replace (with some probability) the planner's action with the action leading to the highest uncertainty in model prediction. The paper evaluates the proposed technique using two standard search planners (MCTS and BFS). \n\nUnfortunately, I think the significance and technical contribution of this work is minimal, an issue that mostly likely starts from a deficient literature review. What the authors refer to as Trust-But-Verify, it's just an ad-hoc instance of the well-known principle of *optimism in the face of uncertainty*, which underlies classic bandit and RL algorithms such as UCB1[1], UCT[2], Thompson Sampling [3, 4]. In the model-free setting, this idea has lead to numerous recent algorithms, many of which also use ensembles for uncertainty quantification [5-8]. In the model-based setting there are also some precedents of work using similar ideas [9-11]. There is also a large body of work treating the problem from the point of view of Bayesian RL (see [12] for a survey).  It is a bad sign that none of this body of previous work was discussed in the paper, which I would argue was the more relevant literature upon which the paper had to be positioned.\n\nThis could conceivably be excused if the paper technical and experimental contribution was impressive enough, but this is not the case.  In contrast to the literature outlined above (where proposed exploration strategies typically follow from rigorous statistical analysis), this work presents the proposed method as a heuristic rule, providing no insight as to why one should expect the approach to work well in general. Moreover, the experiments are done in relatively simple domain, and compared against simple baselines. Some of the baseline choices do not seem appropriate either. For example, why use $\\epsilon$-greedy for exploration, instead of more robust strategies using upper confidence bounds? \n\nFinally, the writing on the paper can be improved in many places. For example, the paper refers to using the graph structure of the underlying problem, but what this graph structure refers to is not properly defined anywhere in the paper. I imagine it refers to the graph wherein edges represent non-zero probability transitions between states, but this is not clear from the text. Additionally, some paragraphs add little in terms of content. For example, the first paragraph of Section 3 is devoted to describe the basic problem that all model-based RL methods are trying to solve; this issue is ubiquitous so there is no need for a full example and so much text to describe this. Other sentences are unclear, such as \"The optimistic and pessimistic errors are often of the same nature\", which I am not sure what is referring to . Additionally, I didn't see a description of the learned model used in the experiments, which is not an obvious choice, since the environment is discrete. \n\nOverall, to end on a somewhat constructive note, I think the problem the authors are trying to solve is interesting and the proposed approach is based on the right intuitions. However, this work is still too immature for publication. I suggest to the authors to position their work properly with regards to the relevant literature, refine their technical contribution accordingly, and compare with more appropriate baselines.  \n\n----------------------------------------------------------------------\n[1] Auer, Peter, Nicolo Cesa-Bianchi, and Paul Fischer. \"Finite-time analysis of the multiarmed bandit problem.\" Machine learning 47.2-3 (2002): 235-256.\n\n[2] Kocsis, Levente, and Csaba Szepesvári. \"Bandit based monte-carlo planning.\" European conference on machine learning. Springer, Berlin, Heidelberg, 2006.\n\n[3] Thompson, William R. \"On the likelihood that one unknown probability exceeds another in view of the evidence of two samples.\" Biometrika 25.3/4 (1933): 285-294.\n\n[4] Russo, Daniel, et al. \"A tutorial on thompson sampling.\" arXiv preprint arXiv:1707.02038 (2017).\n\n[5] Bellemare, M., Srinivasan, S., Ostrovski, G., Schaul, T., Saxton, D., & Munos, R. (2016). Unifying count-based exploration and intrinsic motivation. In Advances in neural information processing systems (pp. 1471-1479).\n\n[6] Ostrovski, G., Bellemare, M. G., Oord, A., & Munos, R. (2017, July). Count-Based Exploration with Neural Density Models. In International Conference on Machine Learning (pp. 2721-2730).\n\n[7] Osband, I., Blundell, C., Pritzel, A., & Van Roy, B. (2016). Deep exploration via bootstrapped DQN. In Advances in neural information processing systems (pp. 4026-4034).\n\n[8] Fortunato, M., Azar, M. G., Piot, B., Menick, J., Osband, I., Graves, A., ... & Blundell, C. (2017). Noisy networks for exploration. arXiv preprint arXiv:1706.10295.\n\n[9] Sanner, S., Goetschalckx, R., Driessens, K., & Shani, G. (2009). Bayesian real-time dynamic programming. In Proceedings of the 21st International Joint Conference on Artificial Intelligence (IJCAI-09) (pp. 1784-1789). IJCAI-INT JOINT CONF ARTIF INTELL.\n\n[10] Pathak, Deepak, Dhiraj Gandhi, and Abhinav Gupta. \"Self-supervised exploration via disagreement.\" arXiv preprint arXiv:1906.04161 (2019).\n\n[11] Shyam, Pranav, Wojciech Jaśkowski, and Faustino Gomez. \"Model-based active exploration.\" International Conference on Machine Learning. 2019.\n\n[12] Ghavamzadeh, M., Mannor, S., Pineau, J., & Tamar, A. (2016). Bayesian reinforcement learning: A survey. arXiv preprint arXiv:1609.04436.", "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603858500571}], "openreview_url": "https://openreview.net/forum?id=DE0MSwKv32y", "arxiv_id": null, "paper_pdf": "papers/DE0MSwKv32y.pdf", "paper_pdf_sha256": "57ea382757f5b3a937230f60da27babb629196363ad8d2b7a1fe1442e8767562", "paper_pdf_bytes": 656461, "paper_pdf_source": "openreview", "code_url": "https://github.com/ComradeMisha/TrustButVerify", "code_repository": "ComradeMisha/TrustButVerify", "code_commit": "c410b44e18157fcf1f7ed82745cd4719e623416a", "code_archive": "repos/DE0MSwKv32y.zip", "code_archive_sha256": "d3e57575504c6e39bc9898937977dc2b74b56305df8eef3f529e1c17384b43f1", "code_archive_bytes": 981650, "code_file_count": 75, "code_extensions": {".py": 74, ".js": 1}, "github_disk_usage_kb": 942, "github_languages": {"Python": 487273, "JavaScript": 13049, "CSS": 1198, "HTML": 778}, "github_archived": false, "github_pushed_at": "2021-02-10T20:35:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/trust-but-verify-model-based-exploration-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1lqDertwr", "year": 2020, "status": "rejected", "title": "Regularization Matters in Policy Optimization", "authors": ["Zhuang Liu", "Xuanlin Li", "Bingyi Kang", "Trevor Darrell"], "authorids": ["zhuangl@berkeley.edu", "xuanlinli17@berkeley.edu", "kang@u.nus.edu", "trevor@eecs.berkeley.edu"], "authors_source": "OpenReview API", "abstract": "Deep Reinforcement Learning (Deep RL) has been receiving increasingly more attention  thanks to its encouraging performance on a variety of control tasks. Yet, conventional regularization techniques in training neural networks (e.g., $L_2$ regularization, dropout) have been largely ignored in RL methods, possibly because agents are typically trained and evaluated in the same environment. In this work, we present the first comprehensive study of regularization techniques with multiple policy optimization algorithms on continuous control tasks. Interestingly, we find conventional regularization techniques on the policy networks can often bring large improvement on the task performance, and the improvement is typically more significant when the task is more difficult. We also compare with the widely used entropy regularization and find $L_2$ regularization is generally better. Our findings are further confirmed to be robust against the choice of training hyperparameters. We also study the effects of regularizing different components and find that only regularizing the policy network is typically enough. We hope our study provides guidance for future practices in regularizing policy optimization algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rkeScHLTKr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2370/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper investigates the use of conventional regularizers for neural networks in the reinforcement learning setting. Contrary to the standard practice of foregoing regularizers in deep RL, the paper finds that their addition can improve the performance of policy gradient algorithms on a standard suite of continuous control tasks. Various regularizers are tried, including l2/l1 regularization, entropy regularization and dropout in a combination with a few standard deep RL algorithms such as TRPO, PPO and SAC. Other experiments also verify the impact of these regularizers on the sensitivity of other hyperparameters and whether regularization should be applied to the value or policy networks. \n\nOverall, I find this paper to be a solid empirical study of regularization in deep reinforcement learning. The experiments are thorough, with various aspects being examined in more detail. I find several of the findings interesting, such as the importance of regularizing solely the policy network and that batch norm/dropout are effective for off-policy methods but not on-policy ones. There were certain points which warranted some further clarification. \n\nI would be willing to increase my score based on the authors' response to the following points:\n1) In section 6, the last two sentences (\"For A2C, TRPO, and PPO ... so further regularization is unnecessary.\") are unclear to me.\n\t- \"rewards are already normalized using running mean filters.\" I thought that rewards are also normalized for SAC, so I'm not sure how this could explain the difference between the on-policy algorithms and the off-policy ones.\n\t- \"mitigates the overestimation bias...further regularization is unncessary.\" Could you clarify the connection between regularization and overestimation bias? Related to this point, in section 2 of the paper, it is written that \"L2 regularization is applied to the critic Q network because it tends to have overestimation bias (Fujimoto et al., 2018)\" but I was not able to find such an explanation in the cited paper though I may have missed it.\n\n2) In section 7, in the paragraph on BN/Dropout, could you clarify the point starting from \"1) For both BN and dropout,...\"? In particular, which discrepancy between the sampling policy and the optimization policy is being referred to here? \n\n3) Did you consider trying weight decay (\"Fixing Weight Decay Regularization in Adam\", Loschilov et al. 2018) as a regularizer? Given the success of L2 regularization, it could be possible that weight decay is even more effective.\n\n4) For the hyperparameter sensitivity plots, where one hyperparameter is varied at a time, why are the step sizes for the policy and value networks not included in these experiments? They are usually a critical hyperparameter.\n\n\nMinor comments and typos:\n- On p.5, when defining \"hurting\", perhaps it could be better to choose a looser definition such as \"\\mu_r < \\mu_b\" or \"\\mu_r - \\sigma_r < \\mu_b - \\sigma_b\". This way, there could be a larger distinction between the most effective methods. Currently, both l2 and entropy regularization achieve 0.0% and the next best two regularizers are also under 10%. \n- In abstract: \"regularizing the policy network is typically enough.\" Rephrase perhaps? The experiments seem to show that applying a regularizer to only the policy network is better than on both.\n- In abstract: \"large improvement\" -> \"large improvements\"\n- p.2, par. 2: \"those regularizations\" -> \"those regularizers\"\n- p.3, Weight Clipping: \"This greatly stablizes\" -> \"This greatly stabilizes\". This sentence could be rephrased since \"This\" seems to refer to only weight clipping, but is not the only change in WGANs. \n- p.3, Dropout: \"regularization technique\" -> \"regularization techniques\"\n- p.4, par. 1: \"due to more stochasticity\" -> \"due to increased stochasticity\"\n- p.4, 2nd to last par.: \"during policy update\" -> \"during policy updates\"\n- p.5, 2nd to last par.: \"sometimes help\" -> \"sometimes helps\", \"easier ones baseline is\" -> \"easier ones the baseline is\"\n- p.8, 2nd to last par.: \"it naturally accepts\" -> \"they naturally accept\", \"been shown effective\" -> \"been shown to be effective\"\n- p.8, last sentence: \"policy network without the value network.\" -> \"policy network but not the value network.\"\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper investigates the use of conventional regularizers for neural networks in the reinforcement learning setting. Contrary to the standard practice of foregoing regularizers in deep RL, the paper finds that their addition can improve the performance of policy gradient algorithms on a standard suite of continuous control tasks. Various regularizers are tried, including l2/l1 regularization, entropy regularization and dropout in a combination with a few standard deep RL algorithms such as TRPO, PPO and SAC. Other experiments also verify the impact of these regularizers on the sensitivity of other hyperparameters and whether regularization should be applied to the value or policy networks. \n\nOverall, I find this paper to be a solid empirical study of regularization in deep reinforcement learning. The experiments are thorough, with various aspects being examined in more detail. I find several of the findings interesting, such as the importance of regularizing solely the policy network and that batch norm/dropout are effective for off-policy methods but not on-policy ones. There were certain points which warranted some further clarification. \n\nI would be willing to increase my score based on the authors' response to the following points:\n1) In section 6, the last two sentences (\"For A2C, TRPO, and PPO ... so further regularization is unnecessary.\") are unclear to me.\n\t- \"rewards are already normalized using running mean filters.\" I thought that rewards are also normalized for SAC, so I'm not sure how this could explain the difference between the on-policy algorithms and the off-policy ones.\n\t- \"mitigates the overestimation bias...further regularization is unncessary.\" Could you clarify the connection between regularization and overestimation bias? Related to this point, in section 2 of the paper, it is written that \"L2 regularization is applied to the critic Q network because it tends to have overestimation bias (Fujimoto et al., 2018)\" but I was not able to find such an explanation in the cited paper though I may have missed it.\n\n2) In section 7, in the paragraph on BN/Dropout, could you clarify the point starting from \"1) For both BN and dropout,...\"? In particular, which discrepancy between the sampling policy and the optimization policy is being referred to here? \n\n3) Did you consider trying weight decay (\"Fixing Weight Decay Regularization in Adam\", Loschilov et al. 2018) as a regularizer? Given the success of L2 regularization, it could be possible that weight decay is even more effective.\n\n4) For the hyperparameter sensitivity plots, where one hyperparameter is varied at a time, why are the step sizes for the policy and value networks not included in these experiments? They are usually a critical hyperparameter.\n\n\nMinor comments and typos:\n- On p.5, when defining \"hurting\", perhaps it could be better to choose a looser definition such as \"\\mu_r < \\mu_b\" or \"\\mu_r - \\sigma_r < \\mu_b - \\sigma_b\". This way, there could be a larger distinction between the most effective methods. Currently, both l2 and entropy regularization achieve 0.0% and the next best two regularizers are also under 10%. \n- In abstract: \"regularizing the policy network is typically enough.\" Rephrase perhaps? The experiments seem to show that applying a regularizer to only the policy network is better than on both.\n- In abstract: \"large improvement\" -> \"large improvements\"\n- p.2, par. 2: \"those regularizations\" -> \"those regularizers\"\n- p.3, Weight Clipping: \"This greatly stablizes\" -> \"This greatly stabilizes\". This sentence could be rephrased since \"This\" seems to refer to only weight clipping, but is not the only change in WGANs. \n- p.3, Dropout: \"regularization technique\" -> \"regularization techniques\"\n- p.4, par. 1: \"due to more stochasticity\" -> \"due to increased stochasticity\"\n- p.4, 2nd to last par.: \"during policy update\" -> \"during policy updates\"\n- p.5, 2nd to last par.: \"sometimes help\" -> \"sometimes helps\", \"easier ones baseline is\" -> \"easier ones the baseline is\"\n- p.8, 2nd to last par.: \"it naturally accepts\" -> \"they naturally accept\", \"been shown effective\" -> \"been shown to be effective\"\n- p.8, last sentence: \"policy network without the value network.\" -> \"policy network but not the value network.\"\n\n\n"}, "tcdate": 1571804557201}, {"id": "SJlYWS2hYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2370/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper provides an empirical study of regularization in policy optimization methods in multiple continuous control tasks. The paper focuses on the effect of conventional regularization on performance in training environments, not generalization ability to different (but similar) testing environments. Their findings suggest that L2 and entropy regularization can improve the performance, be robust to hyperparameters on the tasks studied in the paper. \n\nOverall, the paper is well written. However, I am leaning to reject this paper because (1) the experimental finding is not well justified (2) the experiments are missing some details and do not provide convincing evidence.  \n\nFirst, the paper does not well justify why regularization methods improve performance in training environments. One potential reason is discussed in Section 7: regularization can improve generalization to unseen samples. However, the improvement can simply due to better hyperparemer optimization. When we introduce more hyperparemers and computation compared to baselines, it’s not surprising to see a better performance, especially in deep RL where using a different seed or using a different implementation can have significant difference in performance [1]. Moreover, it is unclear that inability to generalize to unseen samples is a problem in the continuous control tasks evaluated in the paper. I think the paper should demonstrate that this is indeed a problem. If it is not a problem, why would you expect regularization to help?\n\nThere are some missing details which makes it difficult to draw conclusion:\n1. How was \\sigma_{env,r} computed? Is it the standard error of the mean return, or the standard deviation of the return? \n2. What does the average rank mean (in Table 2 and 3)? the average ranking over 5 seeds and all environments? If so, does it make sense to compare these numbers? e.g. Algorithm A with rank 1, 1, 7, 7 and Algorithm B with rank 4, 4, 4, 4 have the same average rank, but totally different performance. \n3. The experiment in Figure 3 seems very interesting, however, what’s the conclusion here? \n4. Why do you use difference hyperparamer ranges (lambda for L2, L1 and entropy regularization) for different algorithms in appendix A? \n\nMinor comment which does not impact the score:\n1. It would have been better if there’s a brief description of each algorithm (before section 4 or in appendix). \n\n[1] Reproducibility of Benchmarked Deep Reinforcement Learning Tasks for Continuous Control\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper provides an empirical study of regularization in policy optimization methods in multiple continuous control tasks. The paper focuses on the effect of conventional regularization on performance in training environments, not generalization ability to different (but similar) testing environments. Their findings suggest that L2 and entropy regularization can improve the performance, be robust to hyperparameters on the tasks studied in the paper. \n\nOverall, the paper is well written. However, I am leaning to reject this paper because (1) the experimental finding is not well justified (2) the experiments are missing some details and do not provide convincing evidence.  \n\nFirst, the paper does not well justify why regularization methods improve performance in training environments. One potential reason is discussed in Section 7: regularization can improve generalization to unseen samples. However, the improvement can simply due to better hyperparemer optimization. When we introduce more hyperparemers and computation compared to baselines, it’s not surprising to see a better performance, especially in deep RL where using a different seed or using a different implementation can have significant difference in performance [1]. Moreover, it is unclear that inability to generalize to unseen samples is a problem in the continuous control tasks evaluated in the paper. I think the paper should demonstrate that this is indeed a problem. If it is not a problem, why would you expect regularization to help?\n\nThere are some missing details which makes it difficult to draw conclusion:\n1. How was \\sigma_{env,r} computed? Is it the standard error of the mean return, or the standard deviation of the return? \n2. What does the average rank mean (in Table 2 and 3)? the average ranking over 5 seeds and all environments? If so, does it make sense to compare these numbers? e.g. Algorithm A with rank 1, 1, 7, 7 and Algorithm B with rank 4, 4, 4, 4 have the same average rank, but totally different performance. \n3. The experiment in Figure 3 seems very interesting, however, what’s the conclusion here? \n4. Why do you use difference hyperparamer ranges (lambda for L2, L1 and entropy regularization) for different algorithms in appendix A? \n\nMinor comment which does not impact the score:\n1. It would have been better if there’s a brief description of each algorithm (before section 4 or in appendix). \n\n[1] Reproducibility of Benchmarked Deep Reinforcement Learning Tasks for Continuous Control\n"}, "tcdate": 1571763457294}, {"id": "rJgXCiRoFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2370/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "An interesting paper on the role of regularization in policy optimization\n\n\nIn this paper, the authors study a set of existing direct policy optimization methods in the field of reinforcement learning. The authors provide a detailed investigation of the effect of regulations on the performance and behavior of agents following these methods.\n\nThe authors present that regularization methods mostly help to improve the agents' performance in terms of final scores. Specifically, they show that direct regularizations on model parameters, such as the standard case of L2 or L1 regularization, generally improve the agent performance. They also show that these regularizations, in their study, is more proper than entropy regularization. The authors also show that, in the presence of such regularizations, the learning algorithms become less sensitive to the hyperparameters. \n\nFew comments:\n1) The paper is well written and easy to follow. I appreciate it. I found the writing of the paper has a bit of repetition. The authors might find it slightly more proper to remove some of the repetitions (e.g. section 4.2)\n\n2) While I appreciate the clear writing and reasoning in this paper, I might suggest a slight change in the second paraphrase of the intro. I agree with the authors' reason on the first three lines, but I think it would be useful to also emphasize the role of the questions the researchers investigate to answer. I might also add one the main reason that the researchers in the field of DRL have spent less time on regulation or architecture search was their focus on more high-level algorithm design which is in the more immediate step of relevance and specialty to the field of reinforcement learning. \n\n3) I would suggest rephrasing the last two sentences of the second paragraph in related work: \"Also, these techniques consider ...\". Regularizing the output also regularizes the parameters, I think the authors' point was \"directly regularize\" the parameters. \n\n4) In the \"Entropy Regularization\" part of section 3, I guess the Hs has not been defined. \n\n5) Repeated \"the\" in the last paragraph of section 4.1 (despite it already incorporates the the maximization of)\n\n6) The authors used the term \"not converge\" multiple times. While it is hard from the plots to see whether the series converges or not, I have a strong feeling that by this term the authors mean the algorithm does not converge to a resealable solution rather than being divergent up to a bandwidth. Maybe clarifying would be helpful.\n\n7) In section 5, the authors study the sensitivity to the hyperparameters. In this section, I had a hard time to understand the role of term 3\n\"BN and dropout hurts on-policy algorithms but can bring improvement only for the off-policy SAC algorithm.\" Does it mean that deploying BN, results in a more sensitive algorithm? or it means that the performance degrades (which is a different topic than section 5 is supposed to serve)?\n\n8) In section 7, the authors put out a hypothesis \"\nHowever, there is still generalization between samples: the agents are only trained on the limited\" but the provided empirical study might not fully be considered to be designed to test this hypothesis. In order to test this hypothesis, the author might be interested in training the models with bigger sample sizes, more training iteration, different function classes, and more fitting in order to test this hypothesis.\n\n\n9) Section 7 on \"Why do BN and dropout work only with off-policy algorithms?\" while I agree with the authors on their first reason which is quite commonly known, I might hesitate to make the second statement (2)\n\n\n\nGenerally, I found this paper an interesting paper and appreciate the authors for their careful empirical study. But I found the contribution of this work to be not significant enough. Most of the statements and claims in this paper are well know in the community, especially among deep learning practitioners. While I acknowledge the scientific value of this study, its concreteness, and appreciate the contribution of this paper, due to the low acceptance rate of this conference, I might be reluctant in accepting this paper. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "An interesting paper on the role of regularization in policy optimization\n\n\nIn this paper, the authors study a set of existing direct policy optimization methods in the field of reinforcement learning. The authors provide a detailed investigation of the effect of regulations on the performance and behavior of agents following these methods.\n\nThe authors present that regularization methods mostly help to improve the agents' performance in terms of final scores. Specifically, they show that direct regularizations on model parameters, such as the standard case of L2 or L1 regularization, generally improve the agent performance. They also show that these regularizations, in their study, is more proper than entropy regularization. The authors also show that, in the presence of such regularizations, the learning algorithms become less sensitive to the hyperparameters. \n\nFew comments:\n1) The paper is well written and easy to follow. I appreciate it. I found the writing of the paper has a bit of repetition. The authors might find it slightly more proper to remove some of the repetitions (e.g. section 4.2)\n\n2) While I appreciate the clear writing and reasoning in this paper, I might suggest a slight change in the second paraphrase of the intro. I agree with the authors' reason on the first three lines, but I think it would be useful to also emphasize the role of the questions the researchers investigate to answer. I might also add one the main reason that the researchers in the field of DRL have spent less time on regulation or architecture search was their focus on more high-level algorithm design which is in the more immediate step of relevance and specialty to the field of reinforcement learning. \n\n3) I would suggest rephrasing the last two sentences of the second paragraph in related work: \"Also, these techniques consider ...\". Regularizing the output also regularizes the parameters, I think the authors' point was \"directly regularize\" the parameters. \n\n4) In the \"Entropy Regularization\" part of section 3, I guess the Hs has not been defined. \n\n5) Repeated \"the\" in the last paragraph of section 4.1 (despite it already incorporates the the maximization of)\n\n6) The authors used the term \"not converge\" multiple times. While it is hard from the plots to see whether the series converges or not, I have a strong feeling that by this term the authors mean the algorithm does not converge to a resealable solution rather than being divergent up to a bandwidth. Maybe clarifying would be helpful.\n\n7) In section 5, the authors study the sensitivity to the hyperparameters. In this section, I had a hard time to understand the role of term 3\n\"BN and dropout hurts on-policy algorithms but can bring improvement only for the off-policy SAC algorithm.\" Does it mean that deploying BN, results in a more sensitive algorithm? or it means that the performance degrades (which is a different topic than section 5 is supposed to serve)?\n\n8) In section 7, the authors put out a hypothesis \"\nHowever, there is still generalization between samples: the agents are only trained on the limited\" but the provided empirical study might not fully be considered to be designed to test this hypothesis. In order to test this hypothesis, the author might be interested in training the models with bigger sample sizes, more training iteration, different function classes, and more fitting in order to test this hypothesis.\n\n\n9) Section 7 on \"Why do BN and dropout work only with off-policy algorithms?\" while I agree with the authors on their first reason which is quite commonly known, I might hesitate to make the second statement (2)\n\n\n\nGenerally, I found this paper an interesting paper and appreciate the authors for their careful empirical study. But I found the contribution of this work to be not significant enough. Most of the statements and claims in this paper are well know in the community, especially among deep learning practitioners. While I acknowledge the scientific value of this study, its concreteness, and appreciate the contribution of this paper, due to the low acceptance rate of this conference, I might be reluctant in accepting this paper. \n"}, "tcdate": 1571707851465}], "openreview_url": "https://openreview.net/forum?id=B1lqDertwr", "arxiv_id": "1910.09191", "paper_pdf": "papers/B1lqDertwr.pdf", "paper_pdf_sha256": "90f7bef954753b20d6341470db00a6a47c64e6325bba9c1c2203a7ac53cbb74f", "paper_pdf_bytes": 16829849, "paper_pdf_source": "openreview", "code_url": "https://github.com/xuanlinli17/iclr2021_rlreg", "code_repository": "xuanlinli17/iclr2021_rlreg", "code_commit": "1fbecb9d75b35b89f239e4bf5739c8aba8ce51b2", "code_archive": "repos/B1lqDertwr.zip", "code_archive_sha256": "94acdc54dac9ef2400a095ec2ebf2672cec794f5ac4aac73e0e82c8086176ae7", "code_archive_bytes": 1432869, "code_file_count": 124, "code_extensions": {".py": 124}, "github_disk_usage_kb": 2242, "github_languages": {"Python": 577494}, "github_archived": false, "github_pushed_at": "2021-11-01T06:02:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/regularization-matters-in-policy-optimization-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "yJiMANMsz2", "year": 2026, "status": "rejected", "title": "DRESS: Disentangled Representation-based Self-Supervised Meta-Learning for Diverse Tasks", "authors": ["Wei Cui", "Tongzi Wu", "Jesse C. Cresswell", "Yi Sui", "Keyvan Golestan"], "authorids": ["~Wei_Cui4", "~Tongzi_Wu1", "~Jesse_C._Cresswell1", "~Yi_Sui1", "~Keyvan_Golestan1"], "authors_source": "OpenReview API", "abstract": "Meta-learning represents a strong class of approaches for solving few-shot learning tasks. Nonetheless, recent research suggests that simply pre-training a generic encoder can potentially surpass meta-learning algorithms. In this paper, we hypothesize that the reason meta-learning fails to stand out in popular few-shot learning benchmarks is the lack of diversity among the few-shot learning tasks. We propose DRESS, a task-agnostic Disentangled REpresentation-based Self-Supervised meta-learning approach that enables fast model adaptation on highly diversified few-shot learning tasks. Specifically, DRESS utilizes disentangled representation learning to create self-supervised tasks that can fuel the meta-training process. We validate the effectiveness of DRESS through experiments on datasets with multiple factors of variation and varying complexity. The results suggest that DRESS is able to outperform competing methods on the majority of the datasets and task setups. Through this paper, we advocate for a re-examination of how task adaptation studies are conducted, and aim to reignite interest in the potential of meta-learning for solving few-shot learning tasks via disentangled representations.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NkvIQCv7qU", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21341/Reviewer_WGZW"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper has a clear motivation or assumption that current few-shot learning benchmarks lack task diversity, which makes simple pre-training and fine-tuning approaches appear superior to meta-learning. This paper introduces DRESS (Disentangled REpresentation-based Self-Supervised meta-learning), a novel framework for few-shot learning that integrates disentangled representation learning with self-supervised meta-learning. \n\nTo address this, DRESS constructs diverse self-supervised tasks and aligns latent dimensions to ensure semantic consistency,  then it clusters each latent dimension independently to form pseudo-classes, and creates varied few-shot tasks for meta-training using these clusters.\n\nComprehensive experiments on multiple datasets: SmallNORB, Shapes3D, Causal3D, MPI3D, CelebA, and LFWA, demonstrate that DRESS outperforms other baselines.", "review_text": "This paper has a clear motivation or assumption that current few-shot learning benchmarks lack task diversity, which makes simple pre-training and fine-tuning approaches appear superior to meta-learning. This paper introduces DRESS (Disentangled REpresentation-based Self-Supervised meta-learning), a novel framework for few-shot learning that integrates disentangled representation learning with self-supervised meta-learning. \n\nTo address this, DRESS constructs diverse self-supervised tasks and aligns latent dimensions to ensure semantic consistency,  then it clusters each latent dimension independently to form pseudo-classes, and creates varied few-shot tasks for meta-training using these clusters.\n\nComprehensive experiments on multiple datasets: SmallNORB, Shapes3D, Causal3D, MPI3D, CelebA, and LFWA, demonstrate that DRESS outperforms other baselines.", "strengths": "1. The motivation is clear and strong.\n2. The presentation is very well and the arguments are coherent.\n3. Comprehensive experiments and ablation studies are conducted to validate the proposed approach.", "weaknesses": "1. Disentangled encoder dependency: DRESS relies on disentangled representation models such as FDAE or LSD, which require substantial training resources and careful hyperparameter tuning. There are two concerns for this:\n - 1.1: The paper could discuss the computational trade-offs more explicitly.\n - 1.2: From the paper, \"*all images available for meta-training are collected, and used to train a general purpose encoder*\", it's confusing whether the pretrained model is also trained in this way or not. The difference should be clarified because the paper claims that in the scenario of task diversity scenario, meta-learning could outperform pretrained models. It seems meta-learning also has a pretrained encoder condition.\n\n2. The task diversity metric validation is not enough: While intuitive, the proposed class-partition IoU metric’s correlation with downstream adaptation performance could be analyzed more deeply (e.g., correlation plots or regression). Could the proposed task diversity metric be integrated into the training process to encourage the model to sample or emphasize more diverse pseudo-tasks dynamically?", "questions": "In addition to the questions in weakness, some additional questions:\n\n1. It is reasonable that on simpler datasets such as Omniglot, where tasks are highly similar, meta-learning methods often underperform compared to pre-trained models. When constructing tasks with high diversity, could incorporating out-of-domain tasks be beneficial? Specifically, how does DRESS perform if the disentanglement model is pre-trained on a dataset different from the meta-learning dataset (i.e., under out-of-domain disentanglement)?\n\n2. What is the computational cost of DRESS compared to baselines (such as CACTUS or Meta-GMVAE, etc.), for instance, in terms of GPU hours or memory usage?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper has a clear motivation or assumption that current few-shot learning benchmarks lack task diversity, which makes simple pre-training and fine-tuning approaches appear superior to meta-learning. This paper introduces DRESS (Disentangled REpresentation-based Self-Supervised meta-learning), a novel framework for few-shot learning that integrates disentangled representation learning with self-supervised meta-learning. \n\nTo address this, DRESS constructs diverse self-supervised tasks and aligns latent dimensions to ensure semantic consistency,  then it clusters each latent dimension independently to form pseudo-classes, and creates varied few-shot tasks for meta-training using these clusters.\n\nComprehensive experiments on multiple datasets: SmallNORB, Shapes3D, Causal3D, MPI3D, CelebA, and LFWA, demonstrate that DRESS outperforms other baselines.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The motivation is clear and strong.\n2. The presentation is very well and the arguments are coherent.\n3. Comprehensive experiments and ablation studies are conducted to validate the proposed approach.", "weaknesses": "1. Disentangled encoder dependency: DRESS relies on disentangled representation models such as FDAE or LSD, which require substantial training resources and careful hyperparameter tuning. There are two concerns for this:\n - 1.1: The paper could discuss the computational trade-offs more explicitly.\n - 1.2: From the paper, \"*all images available for meta-training are collected, and used to train a general purpose encoder*\", it's confusing whether the pretrained model is also trained in this way or not. The difference should be clarified because the paper claims that in the scenario of task diversity scenario, meta-learning could outperform pretrained models. It seems meta-learning also has a pretrained encoder condition.\n\n2. The task diversity metric validation is not enough: While intuitive, the proposed class-partition IoU metric’s correlation with downstream adaptation performance could be analyzed more deeply (e.g., correlation plots or regression). Could the proposed task diversity metric be integrated into the training process to encourage the model to sample or emphasize more diverse pseudo-tasks dynamically?", "questions": "In addition to the questions in weakness, some additional questions:\n\n1. It is reasonable that on simpler datasets such as Omniglot, where tasks are highly similar, meta-learning methods often underperform compared to pre-trained models. When constructing tasks with high diversity, could incorporating out-of-domain tasks be beneficial? Specifically, how does DRESS perform if the disentanglement model is pre-trained on a dataset different from the meta-learning dataset (i.e., under out-of-domain disentanglement)?\n\n2. What is the computational cost of DRESS compared to baselines (such as CACTUS or Meta-GMVAE, etc.), for instance, in terms of GPU hours or memory usage?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762236833961}, {"id": "nXLlNJXyNl", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21341/Reviewer_U7QL"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "To overcome the limitations of current meta-learning methods compared to pre-training and fine-tuning, the authors propose a simple addition to the few-shot learning task, where they add clustering-based disentangled representation learning for latent task embeddings to further diversify learning and aid generalization across tasks. They propose to generate self-supervised tasks by clustering along disentangled latent dimensions, thereby enforcing diversity among tasks. Experiments on controlled datasets (Shapes3D, SmallNORB, MPI3D, Causal3D) and real-world data (CelebA, LFWA) show that the proposed method achieves competitive results to existing meta-learning methods.", "review_text": "To overcome the limitations of current meta-learning methods compared to pre-training and fine-tuning, the authors propose a simple addition to the few-shot learning task, where they add clustering-based disentangled representation learning for latent task embeddings to further diversify learning and aid generalization across tasks. They propose to generate self-supervised tasks by clustering along disentangled latent dimensions, thereby enforcing diversity among tasks. Experiments on controlled datasets (Shapes3D, SmallNORB, MPI3D, Causal3D) and real-world data (CelebA, LFWA) show that the proposed method achieves competitive results to existing meta-learning methods.", "strengths": "- The perspective on fortifying the task diversity and its implementation through disentanglement and clustering-based approach seems straightforward and reasonable.\n\n- The objective of the proposed method is explained in a straightforward manner, with simple formulation which can be easily implemented and reproduced by other researchers.\n\n- The authors performed extensive experimental validations on multiple , and they also conducted some ablation experiments to facilitate further understanding for the proposed work.", "weaknesses": "- Although the authors' main claim is that the limitations of current meta-learning methods come from the lack of task diversity in well-known meta-learning datasets such as miniImageNet and CIFAR-FS, it is not theoretically or empirically verified through thorough validation.\n\n- Related to the previous question, the paper seems to conclude that it is better to use fine-tuning and pre-training approaches when there is a few-shot training dataset with small diversity. However, there are scenarios where the backbone feature extractor model lacks representational power (e.g. small in capacity), and this work seems to overlook this setting where conventional meta-learning can be dominant.\n\n- Eventually, the proposed DRESS seems to only strike a middle ground where the task diversity is not too low, not too high, and the backbone network is not too simple, not too complex. This seems to limit the proposed work's wide application to various possible scenarios.", "questions": "Please refer to the weaknesses section. The insights found and the issues raised by the authors are still interesting and worth the analysis, but the applicability of the proposed method seems to be narrow, with limited usages.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To overcome the limitations of current meta-learning methods compared to pre-training and fine-tuning, the authors propose a simple addition to the few-shot learning task, where they add clustering-based disentangled representation learning for latent task embeddings to further diversify learning and aid generalization across tasks. They propose to generate self-supervised tasks by clustering along disentangled latent dimensions, thereby enforcing diversity among tasks. Experiments on controlled datasets (Shapes3D, SmallNORB, MPI3D, Causal3D) and real-world data (CelebA, LFWA) show that the proposed method achieves competitive results to existing meta-learning methods.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The perspective on fortifying the task diversity and its implementation through disentanglement and clustering-based approach seems straightforward and reasonable.\n\n- The objective of the proposed method is explained in a straightforward manner, with simple formulation which can be easily implemented and reproduced by other researchers.\n\n- The authors performed extensive experimental validations on multiple , and they also conducted some ablation experiments to facilitate further understanding for the proposed work.", "weaknesses": "- Although the authors' main claim is that the limitations of current meta-learning methods come from the lack of task diversity in well-known meta-learning datasets such as miniImageNet and CIFAR-FS, it is not theoretically or empirically verified through thorough validation.\n\n- Related to the previous question, the paper seems to conclude that it is better to use fine-tuning and pre-training approaches when there is a few-shot training dataset with small diversity. However, there are scenarios where the backbone feature extractor model lacks representational power (e.g. small in capacity), and this work seems to overlook this setting where conventional meta-learning can be dominant.\n\n- Eventually, the proposed DRESS seems to only strike a middle ground where the task diversity is not too low, not too high, and the backbone network is not too simple, not too complex. This seems to limit the proposed work's wide application to various possible scenarios.", "questions": "Please refer to the weaknesses section. The insights found and the issues raised by the authors are still interesting and worth the analysis, but the applicability of the proposed method seems to be narrow, with limited usages.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761933993367}, {"id": "FRsynIZrmp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21341/Reviewer_JfVw"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes DRESS, a self-supervised meta-learning framework that constructs diverse few-shot tasks using disentangled representations.\nThe method trains an encoder (e.g., FDAE or LSD) to separate latent factors, aligns latent dimensions, clusters each dimension to form pseudo-classes, and uses them to meta-train a model with MAML.\nExperiments on curated and real-world datasets show DRESS outperforming existing self-supervised and unsupervised meta-learning baselines, especially when task diversity is high.", "review_text": "This paper proposes DRESS, a self-supervised meta-learning framework that constructs diverse few-shot tasks using disentangled representations.\nThe method trains an encoder (e.g., FDAE or LSD) to separate latent factors, aligns latent dimensions, clusters each dimension to form pseudo-classes, and uses them to meta-train a model with MAML.\nExperiments on curated and real-world datasets show DRESS outperforming existing self-supervised and unsupervised meta-learning baselines, especially when task diversity is high.", "strengths": "(1) The paper clearly motivates the link between task diversity and the failure of standard meta-learning benchmarks.\n\n(2) The framework is conceptually interesting which using disentanglement to create pseudo-tasks is an intuitive way to increase diversity.\n\n(3) The experiments are relatively broad, covering both synthetic datasets (Shapes3D, MPI3D, etc.) and realistic ones (CelebA, LFWA).", "weaknesses": "(1) The entire approach hinges on having well-disentangled latent factors, but disentanglement models rarely achieve clean factor separation in practice; this assumption is barely examined.\n\n(2) The claimed novelty over prior unsupervised meta-learning methods (e.g., CACTUS, Meta-GMVAE) is incremental. It replaces clustering in feature space with clustering along latent dimensions but doesn’t bring a real theoretical or algorithmic advance.\n\n(3) The so-called “task diversity metric” is ad-hoc and does not correlate with adaptation performance; it’s unclear whether higher diversity actually improves meta-generalization beyond synthetic datasets.", "questions": "(1) How robust is DRESS when the disentanglement isn’t perfect? In practice, disentangled features are often noisy or overlapping — does the whole task construction pipeline break down when that happens?\n\n(2) How much does the method rely on the specific encoder choice (FDAE vs. LSD)? If we just use a strong pretrained backbone like DINOv2 without explicit disentanglement, does DRESS still work or does the advantage disappear?\n\n(3) The paper claims that supervised labels can “misguide adaptation,” but that point feels hand-wavy. Can the authors actually show a concrete example or quantitative result where this happens?\n\n(4) Why only test MAML? Would something simpler like ProtoNet or RelationNet also benefit from the proposed task construction, or is DRESS inherently tied to gradient-based adaptation methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes DRESS, a self-supervised meta-learning framework that constructs diverse few-shot tasks using disentangled representations.\nThe method trains an encoder (e.g., FDAE or LSD) to separate latent factors, aligns latent dimensions, clusters each dimension to form pseudo-classes, and uses them to meta-train a model with MAML.\nExperiments on curated and real-world datasets show DRESS outperforming existing self-supervised and unsupervised meta-learning baselines, especially when task diversity is high.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "(1) The paper clearly motivates the link between task diversity and the failure of standard meta-learning benchmarks.\n\n(2) The framework is conceptually interesting which using disentanglement to create pseudo-tasks is an intuitive way to increase diversity.\n\n(3) The experiments are relatively broad, covering both synthetic datasets (Shapes3D, MPI3D, etc.) and realistic ones (CelebA, LFWA).", "weaknesses": "(1) The entire approach hinges on having well-disentangled latent factors, but disentanglement models rarely achieve clean factor separation in practice; this assumption is barely examined.\n\n(2) The claimed novelty over prior unsupervised meta-learning methods (e.g., CACTUS, Meta-GMVAE) is incremental. It replaces clustering in feature space with clustering along latent dimensions but doesn’t bring a real theoretical or algorithmic advance.\n\n(3) The so-called “task diversity metric” is ad-hoc and does not correlate with adaptation performance; it’s unclear whether higher diversity actually improves meta-generalization beyond synthetic datasets.", "questions": "(1) How robust is DRESS when the disentanglement isn’t perfect? In practice, disentangled features are often noisy or overlapping — does the whole task construction pipeline break down when that happens?\n\n(2) How much does the method rely on the specific encoder choice (FDAE vs. LSD)? If we just use a strong pretrained backbone like DINOv2 without explicit disentanglement, does DRESS still work or does the advantage disappear?\n\n(3) The paper claims that supervised labels can “misguide adaptation,” but that point feels hand-wavy. Can the authors actually show a concrete example or quantitative result where this happens?\n\n(4) Why only test MAML? Would something simpler like ProtoNet or RelationNet also benefit from the proposed task construction, or is DRESS inherently tied to gradient-based adaptation methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761917968489}, {"id": "rwxkXYhgIm", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21341/Reviewer_mUUa"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper introduces DRESS, a task-agnostic Disentangled Representation-based Self-Supervised meta-learning framework designed to enable rapid model adaptation across highly diverse few-shot learning tasks. Specifically, DRESS leverages disentangled representation learning to construct self-supervised tasks that enhance the meta-training process. The effectiveness of DRESS is demonstrated through experiments on datasets featuring multiple factors of variation and different levels of complexity.", "review_text": "This paper introduces DRESS, a task-agnostic Disentangled Representation-based Self-Supervised meta-learning framework designed to enable rapid model adaptation across highly diverse few-shot learning tasks. Specifically, DRESS leverages disentangled representation learning to construct self-supervised tasks that enhance the meta-training process. The effectiveness of DRESS is demonstrated through experiments on datasets featuring multiple factors of variation and different levels of complexity.", "strengths": "This paper points out the limited task diversity in existing few-shot learning benchmarks, which explains why pre-training and fine-tuning sometimes appear to outperform meta-learning. To address this, the authors develop new few-shot learning benchmarks featuring more diverse tasks for rigorous evaluation. Furthermore, the paper introduces DRESS, a method that generates diverse tasks to enable self-supervised meta-learning through disentangled representations. Extensive experiments validate the effectiveness of the proposed model.", "weaknesses": "1.\tFirst, meta-learning has been extensively explored over the past decades, and I believe it may no longer be a highly novel research direction.\n2.\tThe disentanglement technique has been applied to many tasks, so I have concerns about the originality and novelty of this work.\n3.\tHow many latent dimensions (“L”) are used in the model? As far as I know, increasing the number of latent variables will significantly enlarge the model size, which may not be an efficient strategy. Moreover, with the rapid advancement of foundation models that already exhibit strong generalization capabilities, I believe such models could potentially address the problems discussed in this paper.", "questions": "Please see the Weaknesses, I would like to see details of novelty.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces DRESS, a task-agnostic Disentangled Representation-based Self-Supervised meta-learning framework designed to enable rapid model adaptation across highly diverse few-shot learning tasks. Specifically, DRESS leverages disentangled representation learning to construct self-supervised tasks that enhance the meta-training process. The effectiveness of DRESS is demonstrated through experiments on datasets featuring multiple factors of variation and different levels of complexity.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper points out the limited task diversity in existing few-shot learning benchmarks, which explains why pre-training and fine-tuning sometimes appear to outperform meta-learning. To address this, the authors develop new few-shot learning benchmarks featuring more diverse tasks for rigorous evaluation. Furthermore, the paper introduces DRESS, a method that generates diverse tasks to enable self-supervised meta-learning through disentangled representations. Extensive experiments validate the effectiveness of the proposed model.", "weaknesses": "1.\tFirst, meta-learning has been extensively explored over the past decades, and I believe it may no longer be a highly novel research direction.\n2.\tThe disentanglement technique has been applied to many tasks, so I have concerns about the originality and novelty of this work.\n3.\tHow many latent dimensions (“L”) are used in the model? As far as I know, increasing the number of latent variables will significantly enlarge the model size, which may not be an efficient strategy. Moreover, with the rapid advancement of foundation models that already exhibit strong generalization capabilities, I believe such models could potentially address the problems discussed in this paper.", "questions": "Please see the Weaknesses, I would like to see details of novelty.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761812676934}], "openreview_url": "https://openreview.net/forum?id=yJiMANMsz2", "arxiv_id": "2503.09679", "paper_pdf": "papers/yJiMANMsz2.pdf", "paper_pdf_sha256": "5e74067d12025fcb090e97c40e8bc03d4b4dbe8d19bc5777891b4ed77d01779a", "paper_pdf_bytes": 5112783, "paper_pdf_source": "openreview", "code_url": "https://github.com/layer6ai-labs/DRESS", "code_repository": "layer6ai-labs/DRESS", "code_commit": "692d3d6231ecc38e02e4751c55cf6909c67aa119", "code_archive": "repos/yJiMANMsz2.zip", "code_archive_sha256": "06521407470ff41a2736a08ba80a1a0fae5d929dcb77f6e23bb9108698b96799", "code_archive_bytes": 2646029, "code_file_count": 47, "code_extensions": {".py": 41, ".sh": 6}, "github_disk_usage_kb": 2913, "github_languages": {"Python": 258853, "Shell": 76821}, "github_archived": false, "github_pushed_at": "2026-06-17T01:59:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dress-disentangled-representation-based-self"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kMT8ujhYbA", "year": 2025, "status": "rejected", "title": "Catastrophic Cyber Capabilities Benchmark (3CB): Robustly Evaluating LLM Agent Cyber Offense Capabilities", "authors": ["Andrey Anurin", "Jonathan Ng", "Kibo Schaffer", "Ziyue Wang", "Jason Hoelscher-Obermaier", "Esben Kran"], "authorids": ["~Andrey_Anurin1", "~Jonathan_Ng1", "~Kibo_Schaffer1", "~Ziyue_Wang6", "~Jason_Hoelscher-Obermaier1", "~Esben_Kran1"], "authors_source": "OpenReview API", "abstract": "LLM agents have the potential to revolutionize defensive cyber operations, but their offensive capabilities are not yet fully understood. To prepare for emerging threats, model developers and governments are evaluating the cyber capabilities of foundation models. However, these assessments often lack transparency and a comprehensive focus on offensive capabilities. In response, we introduce the Catastrophic Cyber Capabilities Benchmark (3CB), a novel framework designed to rigorously assess the real-world offensive capabilities of LLM agents. Our evaluation of modern LLMs on 3CB reveals that frontier models, such as GPT-4o and Claude 3.5 Sonnet, can perform offensive tasks such as reconnaissance and exploitation across domains ranging from binary analysis to web technologies. Conversely, smaller open-source models exhibit limited offensive capabilities. Our software solution and the corresponding benchmark provides a critical tool to reduce the gap between rapidly improving capabilities and robustness of cyber offense evaluations, aiding in the safer deployment and regulation of these powerful technologies.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "7X9dwaTOOY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12152/Reviewer_hgnu"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper proposes a framework and benchmark of 15 challenges that allows a LLM to conduct attacks. They evaluate a number of LLMs in this framework.", "review_text": "This paper proposes a framework and benchmark of 15 challenges that allows a LLM to conduct attacks. They evaluate a number of LLMs in this framework.", "strengths": "This is a benchmarks paper and I believe the authors have chosen a primary area that is not suited for this paper. If I evaluate this according to the choice of authors' choice of primary area, then this paper should be rejected as there is no new technique in safety, alignment, privacy, or social considerations.\n\nMerits:\n1) The problem is interesting and a good direction to study and benchmark the behavior of LLMs.\n2) The experiments have been done extensively (but as acknowledged by authors, not covering all tactics)\n3) This can be a valuable basis for further engineering.", "weaknesses": "Weakness:\n1) This line \"While some benchmarks claim to measure general cyber capabilities but only cover specific sub-capabilities, 3CB ensures that each challenge is designed such that successful completion by a model accurately reflects its ability to apply the technique described in that challenge.\" First, this line requires citations to the \"some\" benchmarks and then I did not understand how 3CB ensures that each challenge is designed such that successful completion by a model accurately reflects its ability to apply the technique described in that challenge? Where is the evidence for this claim?\n2) Also, the authors claim novelty of the challenges - how is this ensured. Particularly, when the challenges are broadly in the known ATT&CK category? Also, LLMs can easily relate similar (but not same) textual concepts, so where is the this novelty coming from - are these conceptually new challenges?\n3) Related to above, the authors say 11 challenges are publicly released - does this mean on the internet?\n4) There are many open questions arising from the work, some of which should be done to make the work stronger? Will a cybersecurity-specific finetuned smaller LLM (like Llamma) do better? Clearly, large LLM like GPT have many advantages in terms of training data and size of model, more fine-grained comparison of why these succeed and smaller LLM fail is needed.\n\nOverall, I find this to be useful work, but I am not sure if it fits well in a top research conference.", "questions": "Please respond to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a framework and benchmark of 15 challenges that allows a LLM to conduct attacks. They evaluate a number of LLMs in this framework.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "This is a benchmarks paper and I believe the authors have chosen a primary area that is not suited for this paper. If I evaluate this according to the choice of authors' choice of primary area, then this paper should be rejected as there is no new technique in safety, alignment, privacy, or social considerations.\n\nMerits:\n1) The problem is interesting and a good direction to study and benchmark the behavior of LLMs.\n2) The experiments have been done extensively (but as acknowledged by authors, not covering all tactics)\n3) This can be a valuable basis for further engineering.", "weaknesses": "Weakness:\n1) This line \"While some benchmarks claim to measure general cyber capabilities but only cover specific sub-capabilities, 3CB ensures that each challenge is designed such that successful completion by a model accurately reflects its ability to apply the technique described in that challenge.\" First, this line requires citations to the \"some\" benchmarks and then I did not understand how 3CB ensures that each challenge is designed such that successful completion by a model accurately reflects its ability to apply the technique described in that challenge? Where is the evidence for this claim?\n2) Also, the authors claim novelty of the challenges - how is this ensured. Particularly, when the challenges are broadly in the known ATT&CK category? Also, LLMs can easily relate similar (but not same) textual concepts, so where is the this novelty coming from - are these conceptually new challenges?\n3) Related to above, the authors say 11 challenges are publicly released - does this mean on the internet?\n4) There are many open questions arising from the work, some of which should be done to make the work stronger? Will a cybersecurity-specific finetuned smaller LLM (like Llamma) do better? Clearly, large LLM like GPT have many advantages in terms of training data and size of model, more fine-grained comparison of why these succeed and smaller LLM fail is needed.\n\nOverall, I find this to be useful work, but I am not sure if it fits well in a top research conference.", "questions": "Please respond to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730378734466}, {"id": "XIUHtm3Gei", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12152/Reviewer_mLRf"], "rating": 8, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper proposes an evaluation framework of state of the art LLM agents in their capabilities to carry out offensive cybersecurity tasks. Based on the MITRE ATT&CK framework, they propose 15 cybersecurity challenges, representing the 15 attack categories (one from each category) of the ATT&CK framework, that they use to evaluate the capabilities of 14 LLM agents in completing successfully each challenge. Each challenge is attempted 10 times with the best elicitation variation. Then, a linear mixed-effects model is used to quantify each agent's performance. The experimental results show a variability in the rate of completion between LLM agents, the best being the larger ones (GPT-4 and Claude 3.5 Sonnet). They also reveal a variability due to the communication protocol used by the agent, notably, XML outperforming Markdown and XML. The overall conclusion of the paper is that the proposed 3CB framework shows that LLMs (mainly the large ones) are capable of sophisticated cyber operations and hence their use in cybersecurity offensive operations should be regulated.", "review_text": "The paper proposes an evaluation framework of state of the art LLM agents in their capabilities to carry out offensive cybersecurity tasks. Based on the MITRE ATT&CK framework, they propose 15 cybersecurity challenges, representing the 15 attack categories (one from each category) of the ATT&CK framework, that they use to evaluate the capabilities of 14 LLM agents in completing successfully each challenge. Each challenge is attempted 10 times with the best elicitation variation. Then, a linear mixed-effects model is used to quantify each agent's performance. The experimental results show a variability in the rate of completion between LLM agents, the best being the larger ones (GPT-4 and Claude 3.5 Sonnet). They also reveal a variability due to the communication protocol used by the agent, notably, XML outperforming Markdown and XML. The overall conclusion of the paper is that the proposed 3CB framework shows that LLMs (mainly the large ones) are capable of sophisticated cyber operations and hence their use in cybersecurity offensive operations should be regulated.", "strengths": "On the positive side, \n1) The paper is very well written and enjoyable to read.\n2) The topic is timely and very relevant, as penetration testing is still a manual process heavily involving human agents who need to be trained on offensive tactics and requiring a high technical expertise. Such framework is very welcome for organizations who opt for an offensive security strategy to harden their system security.\n3) The proposed framework is a good step towards leveraging the capabilities of state of the art LLM agents for cybersecurity offensive tasks.\n4) The approach is comprehensive as it covers all the categories of cybersecurity attacks (according to MITRE) and is evaluated on the major LLM agents currently.", "weaknesses": "On the negative side,\n1) In my opinion, the paper does not show clearly the capabilities of LLM agents in cybersecurity offensive operations. Apart from the (good) example in Appendix B.2, as a reader I couldn't gauge the capabilities of the LLMs in cybersecurity attacks.\n2) For example, the proposed challenges in Table 1 come with brief descriptions that does not allow to assess how simple/sophisticated they are. Take the first challenge, nmap, to me it is a straightforward challenge. Then, when you check the rate of completion for the various LLMs (Figure 4), you notice that not all LLMs could complete it. In particular, binpwn, which is a more demanding challenge, has a better completion rate. I find this very strange. Also if you take the setuid, you notice a higher completion rate than nmap. This suggests that setuid is even simpler than nmap ! I find these results counterintuitive and do not help to gauge well the capabilities of the LLMs.\n3) Apart from binpwn, no details are provided for the other challenges (I understand that 4 of them are kept confidential on purpose). I tried to get those details from the github repository, but as far as I saw, it contains just a demo of again the binpwn. Nothing on the other challenges. I may be wrong. \n4) I couldn't run the code because the repository lacks clear instructions on how to install the software and run it.  \n\nSo as recommendation, I would suggest providing more details about each challenge (in an extensive version of Table 1, or better yet, in the appendix, similair to binpwn). Also, list all the prompts used in the evaluation (in the appendix or in the github).", "questions": "1) How the prompts are chosen for each challenge? As I understand, several prompts were used for every challenge.\n2) Some challenges require specific environment setup (VMs, specific network configuration, etc.). How those environment requirements are handled in the framework ?\n3) Can you explain why some LLMs complete some sophisticated challenges but fail in simpler ones ? This is crucial to understand better the capabilities of the LLMs in cybersecurity offensive operations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an evaluation framework of state of the art LLM agents in their capabilities to carry out offensive cybersecurity tasks. Based on the MITRE ATT&CK framework, they propose 15 cybersecurity challenges, representing the 15 attack categories (one from each category) of the ATT&CK framework, that they use to evaluate the capabilities of 14 LLM agents in completing successfully each challenge. Each challenge is attempted 10 times with the best elicitation variation. Then, a linear mixed-effects model is used to quantify each agent's performance. The experimental results show a variability in the rate of completion between LLM agents, the best being the larger ones (GPT-4 and Claude 3.5 Sonnet). They also reveal a variability due to the communication protocol used by the agent, notably, XML outperforming Markdown and XML. The overall conclusion of the paper is that the proposed 3CB framework shows that LLMs (mainly the large ones) are capable of sophisticated cyber operations and hence their use in cybersecurity offensive operations should be regulated.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "On the positive side, \n1) The paper is very well written and enjoyable to read.\n2) The topic is timely and very relevant, as penetration testing is still a manual process heavily involving human agents who need to be trained on offensive tactics and requiring a high technical expertise. Such framework is very welcome for organizations who opt for an offensive security strategy to harden their system security.\n3) The proposed framework is a good step towards leveraging the capabilities of state of the art LLM agents for cybersecurity offensive tasks.\n4) The approach is comprehensive as it covers all the categories of cybersecurity attacks (according to MITRE) and is evaluated on the major LLM agents currently.", "weaknesses": "On the negative side,\n1) In my opinion, the paper does not show clearly the capabilities of LLM agents in cybersecurity offensive operations. Apart from the (good) example in Appendix B.2, as a reader I couldn't gauge the capabilities of the LLMs in cybersecurity attacks.\n2) For example, the proposed challenges in Table 1 come with brief descriptions that does not allow to assess how simple/sophisticated they are. Take the first challenge, nmap, to me it is a straightforward challenge. Then, when you check the rate of completion for the various LLMs (Figure 4), you notice that not all LLMs could complete it. In particular, binpwn, which is a more demanding challenge, has a better completion rate. I find this very strange. Also if you take the setuid, you notice a higher completion rate than nmap. This suggests that setuid is even simpler than nmap ! I find these results counterintuitive and do not help to gauge well the capabilities of the LLMs.\n3) Apart from binpwn, no details are provided for the other challenges (I understand that 4 of them are kept confidential on purpose). I tried to get those details from the github repository, but as far as I saw, it contains just a demo of again the binpwn. Nothing on the other challenges. I may be wrong. \n4) I couldn't run the code because the repository lacks clear instructions on how to install the software and run it.  \n\nSo as recommendation, I would suggest providing more details about each challenge (in an extensive version of Table 1, or better yet, in the appendix, similair to binpwn). Also, list all the prompts used in the evaluation (in the appendix or in the github).", "questions": "1) How the prompts are chosen for each challenge? As I understand, several prompts were used for every challenge.\n2) Some challenges require specific environment setup (VMs, specific network configuration, etc.). How those environment requirements are handled in the framework ?\n3) Can you explain why some LLMs complete some sophisticated challenges but fail in simpler ones ? This is crucial to understand better the capabilities of the LLMs in cybersecurity offensive operations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730272098714}, {"id": "FpBN9pcSgP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12152/Reviewer_CUpv"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new benchmark called 3CB, which includes 15 challenges used to evaluate the cyber offensive capabilities of different LLMs. This paper then conducts experiments on a set of agents with different LLMs and shows that larger LLMs tend to have stronger cyber offensive capabilities than smaller LLMs.", "review_text": "This paper introduces a new benchmark called 3CB, which includes 15 challenges used to evaluate the cyber offensive capabilities of different LLMs. This paper then conducts experiments on a set of agents with different LLMs and shows that larger LLMs tend to have stronger cyber offensive capabilities than smaller LLMs.", "strengths": "1. The motivation is reasonable and has real-world implications.\n\n2. This is the first benchmark for evaluating the catastrophic cyber capabilities of LLMs.", "weaknesses": "Although the authors claim that this is the first comprehensive benchmark for evaluating the catastrophic cyber offensive capabilities of LLMs, the technical contribution is not strong enough to pass the high bar of ICLR. The presentation of the paper is more like a technical report, which is not as well-organized as a technical research paper. For experiments, only some superficial results are obtained, e.g., larger LLMs tend to have stronger cyber offensive capabilities than smaller LLMs, which makes the paper seem neither suitable for a technical paper nor a dataset/benchmark paper (in some other top conferences). Therefore, there is a quite large space for improvement before the paper is ready for acceptance.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new benchmark called 3CB, which includes 15 challenges used to evaluate the cyber offensive capabilities of different LLMs. This paper then conducts experiments on a set of agents with different LLMs and shows that larger LLMs tend to have stronger cyber offensive capabilities than smaller LLMs.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The motivation is reasonable and has real-world implications.\n\n2. This is the first benchmark for evaluating the catastrophic cyber capabilities of LLMs.", "weaknesses": "Although the authors claim that this is the first comprehensive benchmark for evaluating the catastrophic cyber offensive capabilities of LLMs, the technical contribution is not strong enough to pass the high bar of ICLR. The presentation of the paper is more like a technical report, which is not as well-organized as a technical research paper. For experiments, only some superficial results are obtained, e.g., larger LLMs tend to have stronger cyber offensive capabilities than smaller LLMs, which makes the paper seem neither suitable for a technical paper nor a dataset/benchmark paper (in some other top conferences). Therefore, there is a quite large space for improvement before the paper is ready for acceptance.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729924079257}], "openreview_url": "https://openreview.net/forum?id=kMT8ujhYbA", "arxiv_id": "2410.09114", "paper_pdf": "papers/kMT8ujhYbA.pdf", "paper_pdf_sha256": "8a7756a2dcef05742d474779cfd0d052845eaf2907b5888c406391a9ef42b4ac", "paper_pdf_bytes": 499013, "paper_pdf_source": "openreview", "code_url": "https://github.com/apartresearch/3cb", "code_repository": "apartresearch/3cb", "code_commit": "d87effdeaceeac153517ede88a5b57b885c7f5cc", "code_archive": "repos/kMT8ujhYbA.zip", "code_archive_sha256": "c5161fc62328fa4062841ff26e10aa81dd1725065ec7f10f89622e4dfdd951de", "code_archive_bytes": 191715, "code_file_count": 12, "code_extensions": {".py": 11, ".ipynb": 1}, "github_disk_usage_kb": 194, "github_languages": {"Python": 50123, "Jupyter Notebook": 17921, "Mako": 510}, "github_archived": false, "github_pushed_at": "2024-10-30T12:42:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/catastrophic-cyber-capabilities-benchmark-3cb"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "piWvNRR0Ym", "year": 2024, "status": "rejected", "title": "Towards Minimal Targeted Updates of Language Models with Targeted Negative Training", "authors": ["Lily H Zhang", "Rajesh Ranganath", "Arya Tafvizi"], "authorids": ["~Lily_H_Zhang1", "~Rajesh_Ranganath2", "~Arya_Tafvizi1"], "authors_source": "OpenReview API", "abstract": "Generative models of language exhibit impressive capabilities but still place non-negligible probability mass over undesirable outputs. In this work, we address the task of updating a model to avoid unwanted outputs while minimally changing model behavior otherwise, a challenge we refer to as a minimal targeted update. We first formalize the notion of a minimal targeted update and propose a method to achieve such updates using negative examples from a model's generations. Our proposed Targeted Negative Training (TNT) results in updates that keep the new distribution close to the original, unlike existing losses for negative signal which push down probability but do not control what the updated distribution will be. In experiments, we demonstrate that TNT yields a better trade-off between reducing unwanted behavior and preserving model generation behavior than baselines, paving the way towards a modeling paradigm based on iterative training updates that constrain models from generating undesirable outputs while preserving their impressive capabilities.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "kfphUJHwN9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6282/Reviewer_DCRf"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "1. The paper tackles the problem of modifying the generative distribution of LLMs to reduce the likelihood of generating undesirable tokens whilst simultaneously not deviating too much from the original distribution.\n2. Concretely, the authors present their approach (dubbed Targeted Negative Training, or TNT), wherein for each prefix and continuation from an already trained model, if the target token is annotated to be undesirable then the approach minimizes a divergence between the original distribution from the already trained model and a modified distribution based on setting the logit of the undesirable token to 0 and renormalizing for the target model. Otherwise the objective minimizes a divergence between the original and the learned model distribution.\n3. The authors explore different instantiations of divergences: (forward KL, forward KL), (reverse KL, reverse KL) and (reverse KL, forward KL) for the (negative token signal, positive token signal), which they dub (D_{n}, D_{p}) respectively. In addition to that, they also explore with (forward KL, maximum likelihood) and (reverse KL, maximum likelihood) for (D_{n}, D_{p}).\n4. The results presented show that the proposed method allows for better tradeoffs between controlling for hallucinations and keeping the model close to the original distribution.", "review_text": "1. The paper tackles the problem of modifying the generative distribution of LLMs to reduce the likelihood of generating undesirable tokens whilst simultaneously not deviating too much from the original distribution.\n2. Concretely, the authors present their approach (dubbed Targeted Negative Training, or TNT), wherein for each prefix and continuation from an already trained model, if the target token is annotated to be undesirable then the approach minimizes a divergence between the original distribution from the already trained model and a modified distribution based on setting the logit of the undesirable token to 0 and renormalizing for the target model. Otherwise the objective minimizes a divergence between the original and the learned model distribution.\n3. The authors explore different instantiations of divergences: (forward KL, forward KL), (reverse KL, reverse KL) and (reverse KL, forward KL) for the (negative token signal, positive token signal), which they dub (D_{n}, D_{p}) respectively. In addition to that, they also explore with (forward KL, maximum likelihood) and (reverse KL, maximum likelihood) for (D_{n}, D_{p}).\n4. The results presented show that the proposed method allows for better tradeoffs between controlling for hallucinations and keeping the model close to the original distribution.", "strengths": "1. The proposed method for leveraging sentence level annotations in order to modify model behaviour is quite interesting. \n2. I quite like the extensiveness of explorations in terms of exploring the different divergences for both the positive and negative signals. From the results, both TNFF and TNRF seem to achieve a pretty good tradeoff between being faithful to the original distribution and controlling for hallucinations.", "weaknesses": "1. For the different proposed models, without an equivalent table similar to Table 1, it is hard to understand the effectiveness of the approach. Concretely, similar to the TNFLL and TNRLL rows, it would also be good to have an equivalent row for TNFF, TNRR and TNRF.\n2. From Table 1, for both the TNFLL and TNRLL approaches, the performance is considerably worse compared to the baseline method. For TNFLL, it the hallucination rate is substantially higher, while for TNRLL, the BLEU score is much lower. Given this observation, I am hesitant to believe that the proposed approach is actually substantially than the baseline approach.\n3. In my opinion, intuitively, because this approach minimizes the KL at a prefix level, especially considering the fact that the annotations obtained are from a noisy source, it is possible that this approach would steer the model towards not predicting certain words in certain contexts. Concretely, (based from the example in Figure 5), for the sentence \"In some regions of the country, the sex ratio is still quite concerning\", because the annotations are noisy, the word \"sex\" would be (incorrectly) marked as offensive. Consequently, because of the proposed objective, the model might not be able to produce the token \"sex\" for a similar prefix as \"In some regions of the country, the\", even if it did make sense in the context. I think this is a reasonably big limitation of the approach, and it would have been nice to have some discussion on this in the paper.", "questions": "1. Would it be possible to rows for TNFF, TNRR and TNRF in Table 1 ?\n2. Would it be possible to provide some clarification on how Figure 2 was constructed ? Specifically, is the level of hallucination mapped to a different value of \\alpha used (so higher \\alpha -> lower hallucination rates and original distribution fidelity) ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "1. The paper tackles the problem of modifying the generative distribution of LLMs to reduce the likelihood of generating undesirable tokens whilst simultaneously not deviating too much from the original distribution.\n2. Concretely, the authors present their approach (dubbed Targeted Negative Training, or TNT), wherein for each prefix and continuation from an already trained model, if the target token is annotated to be undesirable then the approach minimizes a divergence between the original distribution from the already trained model and a modified distribution based on setting the logit of the undesirable token to 0 and renormalizing for the target model. Otherwise the objective minimizes a divergence between the original and the learned model distribution.\n3. The authors explore different instantiations of divergences: (forward KL, forward KL), (reverse KL, reverse KL) and (reverse KL, forward KL) for the (negative token signal, positive token signal), which they dub (D_{n}, D_{p}) respectively. In addition to that, they also explore with (forward KL, maximum likelihood) and (reverse KL, maximum likelihood) for (D_{n}, D_{p}).\n4. The results presented show that the proposed method allows for better tradeoffs between controlling for hallucinations and keeping the model close to the original distribution.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The proposed method for leveraging sentence level annotations in order to modify model behaviour is quite interesting. \n2. I quite like the extensiveness of explorations in terms of exploring the different divergences for both the positive and negative signals. From the results, both TNFF and TNRF seem to achieve a pretty good tradeoff between being faithful to the original distribution and controlling for hallucinations.", "weaknesses": "1. For the different proposed models, without an equivalent table similar to Table 1, it is hard to understand the effectiveness of the approach. Concretely, similar to the TNFLL and TNRLL rows, it would also be good to have an equivalent row for TNFF, TNRR and TNRF.\n2. From Table 1, for both the TNFLL and TNRLL approaches, the performance is considerably worse compared to the baseline method. For TNFLL, it the hallucination rate is substantially higher, while for TNRLL, the BLEU score is much lower. Given this observation, I am hesitant to believe that the proposed approach is actually substantially than the baseline approach.\n3. In my opinion, intuitively, because this approach minimizes the KL at a prefix level, especially considering the fact that the annotations obtained are from a noisy source, it is possible that this approach would steer the model towards not predicting certain words in certain contexts. Concretely, (based from the example in Figure 5), for the sentence \"In some regions of the country, the sex ratio is still quite concerning\", because the annotations are noisy, the word \"sex\" would be (incorrectly) marked as offensive. Consequently, because of the proposed objective, the model might not be able to produce the token \"sex\" for a similar prefix as \"In some regions of the country, the\", even if it did make sense in the context. I think this is a reasonably big limitation of the approach, and it would have been nice to have some discussion on this in the paper.", "questions": "1. Would it be possible to rows for TNFF, TNRR and TNRF in Table 1 ?\n2. Would it be possible to provide some clarification on how Figure 2 was constructed ? Specifically, is the level of hallucination mapped to a different value of \\alpha used (so higher \\alpha -> lower hallucination rates and original distribution fidelity) ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699475448942}, {"id": "qIfWLi37gt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6282/Reviewer_VP85"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work provides a fine-tuning-based algorithm to update a language model to suppress harmful content generation while remaining as close to the original model as possible. Given a set of harmful responses to avoid (defined by a function that takes in a string and produces binary output), the authors define the ideal target model as one that matches the original model conditioned on never producing an undesirable string. The authors optimize a loss function based on this definition and find that the resulting model better satisfies the desired notion of safety.", "review_text": "This work provides a fine-tuning-based algorithm to update a language model to suppress harmful content generation while remaining as close to the original model as possible. Given a set of harmful responses to avoid (defined by a function that takes in a string and produces binary output), the authors define the ideal target model as one that matches the original model conditioned on never producing an undesirable string. The authors optimize a loss function based on this definition and find that the resulting model better satisfies the desired notion of safety.", "strengths": "1. The paper's key problem is relevant to practice and is well-specified \n\n2. The paper's solution is similarly elegant and was a natural consequence of the problem specification, providing a tractable solution to the proposed problem. \n\n3. The results demonstrate a solid improvement over reasonable baselines to provide targeted updates to the model.", "weaknesses": "1. The authors are missing connections to existing methods. For example, [Korbak et al, 2022](https://arxiv.org/abs/2205.11275) (among others) show that PPO would converge to the same closed-form presented in Eq. 2 when using the positivity/negativity classification as a reward function; this method would also avoid the drawbacks mentioned in the related work for inference time procedures.\n\n2. The paper demonstrates their technique on some relatively easier benchmarks, and it would be much more interesting to try more complicated schemes. There are three ways in which I find them weak\n    - In the specific case where p_neg is defined as the presence of a bad token in the string, there is no need to do any training, and by simply ignoring bad tokens at decoding, one recovers the true optimal solution for both greedy and temperature decoding with zero overhead. Both the toxicity and hallucinations benchmarks provided in this text are dangerously close to \"reject any sentence with a bad (word/entity)\", which makes the application rather uninteresting. It would be cooler to see benchmarks where the reward model may still be automated but captures some global property of the sentence that requires a learning-based technique such as yours to solve.\n    - For the toxicity benchmark, the authors mention that 1.6% of the time, the completion is toxic. I believe a very natural baseline to this problem is performing temperature 1 decoding and regenerating anytime the output is toxic. I get the sense this will very rarely require few regenerations for this specific benchmark, making the learning rather excessive. Having a benchmark that requires a larger change will be a true test of this problem.\n   - For baselines, as mentioned in Weakness 1, I believe PPO is a more natural baseline for learning from a reward model\n\nOverall, I believe the work can better demonstrate the technique.", "questions": "1. At the top of page 7, the paper mentions that all experiments were done with greedy decoding. Does that mean fine-tuning was also based on greedy generations? If only 1.6% of model generations were toxic, would the model get any gradient signal for 98.4% of the generations?\n\n2. Why is TNRLL not called TNFR?\n\n3. Just to be clear, does \"token level annotations\" refer to a function that can take a sentence and assess whether it is negative or positive? This was not super clear to me, even though there were a few sentences dedicated to it on page 4. If it is a function, it might be better specified as such.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work provides a fine-tuning-based algorithm to update a language model to suppress harmful content generation while remaining as close to the original model as possible. Given a set of harmful responses to avoid (defined by a function that takes in a string and produces binary output), the authors define the ideal target model as one that matches the original model conditioned on never producing an undesirable string. The authors optimize a loss function based on this definition and find that the resulting model better satisfies the desired notion of safety.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper's key problem is relevant to practice and is well-specified \n\n2. The paper's solution is similarly elegant and was a natural consequence of the problem specification, providing a tractable solution to the proposed problem. \n\n3. The results demonstrate a solid improvement over reasonable baselines to provide targeted updates to the model.", "weaknesses": "1. The authors are missing connections to existing methods. For example, [Korbak et al, 2022](https://arxiv.org/abs/2205.11275) (among others) show that PPO would converge to the same closed-form presented in Eq. 2 when using the positivity/negativity classification as a reward function; this method would also avoid the drawbacks mentioned in the related work for inference time procedures.\n\n2. The paper demonstrates their technique on some relatively easier benchmarks, and it would be much more interesting to try more complicated schemes. There are three ways in which I find them weak\n    - In the specific case where p_neg is defined as the presence of a bad token in the string, there is no need to do any training, and by simply ignoring bad tokens at decoding, one recovers the true optimal solution for both greedy and temperature decoding with zero overhead. Both the toxicity and hallucinations benchmarks provided in this text are dangerously close to \"reject any sentence with a bad (word/entity)\", which makes the application rather uninteresting. It would be cooler to see benchmarks where the reward model may still be automated but captures some global property of the sentence that requires a learning-based technique such as yours to solve.\n    - For the toxicity benchmark, the authors mention that 1.6% of the time, the completion is toxic. I believe a very natural baseline to this problem is performing temperature 1 decoding and regenerating anytime the output is toxic. I get the sense this will very rarely require few regenerations for this specific benchmark, making the learning rather excessive. Having a benchmark that requires a larger change will be a true test of this problem.\n   - For baselines, as mentioned in Weakness 1, I believe PPO is a more natural baseline for learning from a reward model\n\nOverall, I believe the work can better demonstrate the technique.", "questions": "1. At the top of page 7, the paper mentions that all experiments were done with greedy decoding. Does that mean fine-tuning was also based on greedy generations? If only 1.6% of model generations were toxic, would the model get any gradient signal for 98.4% of the generations?\n\n2. Why is TNRLL not called TNFR?\n\n3. Just to be clear, does \"token level annotations\" refer to a function that can take a sentence and assess whether it is negative or positive? This was not super clear to me, even though there were a few sentences dedicated to it on page 4. If it is a function, it might be better specified as such.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698815667116}, {"id": "bcv4C9q6Z1", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6282/Reviewer_fDnb"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "Summary:\n\nThe paper introduces Targeted Negative Training (TNT), a method for minimally updating language models to prevent unwanted outputs. Unlike previous techniques, TNT fine-tunes models using negative examples generated by the model itself, with the goal of closely aligning the updated model's distribution to the original while avoiding specific undesired behaviors. The method operates by minimizing reverse KL-divergence, ensuring the updated model does not deviate significantly from its initial training. Experiments demonstrate that TNT effectively maintains original model performance better than other negative training approaches while reducing unwanted behaviors. However, it requires access to the original model and detailed token-level annotations, presenting potential practical challenges. TNT's iterative nature also suggests it could be used to enhance model safety over time through continuous refinement.", "review_text": "Summary:\n\nThe paper introduces Targeted Negative Training (TNT), a method for minimally updating language models to prevent unwanted outputs. Unlike previous techniques, TNT fine-tunes models using negative examples generated by the model itself, with the goal of closely aligning the updated model's distribution to the original while avoiding specific undesired behaviors. The method operates by minimizing reverse KL-divergence, ensuring the updated model does not deviate significantly from its initial training. Experiments demonstrate that TNT effectively maintains original model performance better than other negative training approaches while reducing unwanted behaviors. However, it requires access to the original model and detailed token-level annotations, presenting potential practical challenges. TNT's iterative nature also suggests it could be used to enhance model safety over time through continuous refinement.", "strengths": "Advantages:\n - Iterative Model Updates: TNT allows for iterative updates to language models without needing all negative tokens to be specified upfront. This flexibility is advantageous for practical applications where updates may be continuous and ongoing.\n\n - Maintained Model Performance: The experimental setup indicates that TNT can effectively maintain the original model's performance better than baseline methods while also reducing unwanted behaviors.\n\n - Reproducibility and Accessibility: The experiments are reproducible, with the promise of making the code public and using publicly available datasets, which enhances the credibility and utility of the research.", "weaknesses": "Disadvantages:\n - Limited Novelty: The core ideas and methods of TNT may not be as novel as claimed, given the prior existence of the NADO algorithm [NeurIPS 2022, https://arxiv.org/pdf/2205.14219.pdf] framework which appears to address similar goals using related techniques.Further more, NADO has proven its objective to be the **theoretically closed form solution** of the shared targets of TNT/NADO, which weakens the value of the approximated solution (step-level branch-cutting) given by TNT. TNT's flexibility is also less than that of NADO as it requires auxiliary negative annotation, which is a closer setup to that of the FUDGE algorithm [NAACL 2021, https://arxiv.org/abs/2104.05218].\n\n - Oversight in Literature Review: Even if TNT can differ itself from NADO and FUDGE (since the two previous methods study different tasks, yet essentially with similar mathematical setup), the absence of a citation and/or discussion of these existing works suggests a possible gap in the literature review process, which might question the thoroughness of the background research conducted for the paper.", "questions": "What is the essential different between TNT and previous constrained decoding algorithms (FUDGE, NADO, Neural Logic, etc.) that aim to maximize/minimize a given (emplicitly through a symbolic process or implicitly through negative samples) sequence boolean function that defines the negativity of samples?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Summary:\n\nThe paper introduces Targeted Negative Training (TNT), a method for minimally updating language models to prevent unwanted outputs. Unlike previous techniques, TNT fine-tunes models using negative examples generated by the model itself, with the goal of closely aligning the updated model's distribution to the original while avoiding specific undesired behaviors. The method operates by minimizing reverse KL-divergence, ensuring the updated model does not deviate significantly from its initial training. Experiments demonstrate that TNT effectively maintains original model performance better than other negative training approaches while reducing unwanted behaviors. However, it requires access to the original model and detailed token-level annotations, presenting potential practical challenges. TNT's iterative nature also suggests it could be used to enhance model safety over time through continuous refinement.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "Advantages:\n - Iterative Model Updates: TNT allows for iterative updates to language models without needing all negative tokens to be specified upfront. This flexibility is advantageous for practical applications where updates may be continuous and ongoing.\n\n - Maintained Model Performance: The experimental setup indicates that TNT can effectively maintain the original model's performance better than baseline methods while also reducing unwanted behaviors.\n\n - Reproducibility and Accessibility: The experiments are reproducible, with the promise of making the code public and using publicly available datasets, which enhances the credibility and utility of the research.", "weaknesses": "Disadvantages:\n - Limited Novelty: The core ideas and methods of TNT may not be as novel as claimed, given the prior existence of the NADO algorithm [NeurIPS 2022, https://arxiv.org/pdf/2205.14219.pdf] framework which appears to address similar goals using related techniques.Further more, NADO has proven its objective to be the **theoretically closed form solution** of the shared targets of TNT/NADO, which weakens the value of the approximated solution (step-level branch-cutting) given by TNT. TNT's flexibility is also less than that of NADO as it requires auxiliary negative annotation, which is a closer setup to that of the FUDGE algorithm [NAACL 2021, https://arxiv.org/abs/2104.05218].\n\n - Oversight in Literature Review: Even if TNT can differ itself from NADO and FUDGE (since the two previous methods study different tasks, yet essentially with similar mathematical setup), the absence of a citation and/or discussion of these existing works suggests a possible gap in the literature review process, which might question the thoroughness of the background research conducted for the paper.", "questions": "What is the essential different between TNT and previous constrained decoding algorithms (FUDGE, NADO, Neural Logic, etc.) that aim to maximize/minimize a given (emplicitly through a symbolic process or implicitly through negative samples) sequence boolean function that defines the negativity of samples?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698811646889}, {"id": "l9NK1UZ2FU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6282/Reviewer_eQid"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work tackles the problem of updating a language model to reduce undesirable behavior (e.g. generating offensive content or hallucinations), while minimally changing generations elsewhere. The authors propose a method, TNT (Targeted Negative Training) which aims to keep the probability distribution of text close to the original except for instances which are intended to be removed. The paper presents some experiments with a T5 model on reducing hallucinations in summarization and in avoiding toxic responses.", "review_text": "This work tackles the problem of updating a language model to reduce undesirable behavior (e.g. generating offensive content or hallucinations), while minimally changing generations elsewhere. The authors propose a method, TNT (Targeted Negative Training) which aims to keep the probability distribution of text close to the original except for instances which are intended to be removed. The paper presents some experiments with a T5 model on reducing hallucinations in summarization and in avoiding toxic responses.", "strengths": "1. The topic studied by this work is timely, and progress in this direction would be of interest to many in the community.\n2. The proposed method is novel and a significant contribution to the literature", "weaknesses": "1. The experiments could be greatly strengthened. The authors study a single model T5, at a fairly small scale (220M) parameters. No purely autoregressive model is studied. Only two tasks are examined. This lack of breadth makes it hard for readers to access the generality of the claims in the paper. I believe this paper would benefit greatly from more experiments and ablations.\n2. There are very few comparisons to previous work, which in many cases tackle the same tasks presented in the experiments. As an example, avoiding toxic generations has been studied by [1,2]\n\n[1] Lu, Ximing, et al. \"Quark: Controllable text generation with reinforced unlearning.\" Advances in neural information processing systems 35 (2022): 27591-27609.\n[2] Ilharco, Gabriel, et al. \"Editing models with task arithmetic.\" arXiv preprint arXiv:2212.04089 (2022).", "questions": "I have no questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work tackles the problem of updating a language model to reduce undesirable behavior (e.g. generating offensive content or hallucinations), while minimally changing generations elsewhere. The authors propose a method, TNT (Targeted Negative Training) which aims to keep the probability distribution of text close to the original except for instances which are intended to be removed. The paper presents some experiments with a T5 model on reducing hallucinations in summarization and in avoiding toxic responses.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The topic studied by this work is timely, and progress in this direction would be of interest to many in the community.\n2. The proposed method is novel and a significant contribution to the literature", "weaknesses": "1. The experiments could be greatly strengthened. The authors study a single model T5, at a fairly small scale (220M) parameters. No purely autoregressive model is studied. Only two tasks are examined. This lack of breadth makes it hard for readers to access the generality of the claims in the paper. I believe this paper would benefit greatly from more experiments and ablations.\n2. There are very few comparisons to previous work, which in many cases tackle the same tasks presented in the experiments. As an example, avoiding toxic generations has been studied by [1,2]\n\n[1] Lu, Ximing, et al. \"Quark: Controllable text generation with reinforced unlearning.\" Advances in neural information processing systems 35 (2022): 27591-27609.\n[2] Ilharco, Gabriel, et al. \"Editing models with task arithmetic.\" arXiv preprint arXiv:2212.04089 (2022).", "questions": "I have no questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698633251132}], "openreview_url": "https://openreview.net/forum?id=piWvNRR0Ym", "arxiv_id": "2406.13660", "paper_pdf": "papers/piWvNRR0Ym.pdf", "paper_pdf_sha256": "2e7fad090acf793c5aa2befa33e5206140e352f4a9892e572350e491f1580b2b", "paper_pdf_bytes": 779365, "paper_pdf_source": "openreview", "code_url": "https://github.com/google/t5patches", "code_repository": "google/t5patches", "code_commit": "09285b378b7e36cd428738b3e7c5614cd3dde330", "code_archive": "repos/piWvNRR0Ym.zip", "code_archive_sha256": "3789371e1e12e5285ff1a8a2854fc26153d6d05c8a80fffab6e4a7fcfa4a093e", "code_archive_bytes": 57367, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 144, "github_languages": {"Python": 231087}, "github_archived": true, "github_pushed_at": "2024-05-31T20:55:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-minimal-targeted-updates-of-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cddqs4kvC20", "year": 2023, "status": "rejected", "title": "Deep Transformer Q-Networks for Partially Observable Reinforcement Learning", "authors": ["Kevin Esslinger", "Robert Platt", "Christopher Amato"], "authorids": ["~Kevin_Esslinger1", "~Robert_Platt2", "~Christopher_Amato1"], "authors_source": "OpenReview API", "abstract": "Real-world reinforcement learning tasks often involve some form of partial observability where the observations only give a partial or noisy view of the true state of the world. Such tasks typically require some form of memory, where the agent has access to multiple past observations, in order to perform well. One popular way to incorporate memory is by using a recurrent neural network to access the agent's history. However, recurrent neural networks in reinforcement learning are often fragile and difficult to train, susceptible to catastrophic forgetting and sometimes fail completely as a result. In this work, we propose Deep Transformer Q-Networks (DTQN), a novel architecture utilizing transformers and self-attention to encode an agent's history. DTQN is designed modularly, and we compare results against several modifications to our base model. Our experiments demonstrate the transformer can solve partially observable tasks faster and more stably than previous recurrent approaches.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "MGmqphQKY_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3826/Reviewer_CoHr"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed a transformer-based DQN agent.", "review_text": "This paper proposed a transformer-based DQN agent.", "strengths": "Weaknesses:\n1. Incremental work. Replacing RNN with Transformer in deep reinforcement learning is not a new idea! There are so many works with similar ideas, such as Decision Transformer[1], GATO[2], [3], [4], [5], [6] and [7].\n2. Bad writing. The figures in this paper are really bad. For example, in Figure 2(a)(c), the blue curves are out of the figure.\n3. Toy benchmarks. The authors claim that they are better than DQN or DRQN, they should conduct experiments on popular RL benchmarks, such as Atari, MuJoco, and DeepMind Lab. \n4. Weak baselines. They should compare their method with real SOTA methods, such as PPO, Rainbow and IMPALA.\n\n\n- [1] Chen L, Lu K, Rajeswaran A, et al. Decision transformer: Reinforcement learning via sequence modeling[J]. Advances in neural information processing systems, 2021, 34: 15084-15097.\n- [2] Reed S, Zolna K, Parisotto E, et al. A generalist agent[J]. arXiv preprint arXiv:2205.06175, 2022.\n- [3] Laskin M, Wang L, Oh J, et al. In-context Reinforcement Learning with Algorithm Distillation[J]. arXiv preprint arXiv:2210.14215, 2022.\n- [4] Parisotto, Emilio, et al. \"Stabilizing transformers for reinforcement learning.\" International conference on machine learning. PMLR, 2020.\n- [5] Banino, Andrea, et al. \"Coberl: Contrastive bert for reinforcement learning.\" arXiv preprint arXiv:2107.05431 (2021).\n- [6] Furuta, Hiroki, Yutaka Matsuo, and Shixiang Shane Gu. \"Generalized decision transformer for offline hindsight information matching.\" arXiv preprint arXiv:2111.10364 (2021).\n- [7] Micheli, Vincent, Eloi Alonso, and François Fleuret. \"Transformers are sample efficient world models.\" arXiv preprint arXiv:2209.00588 (2022).\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposed a transformer-based DQN agent.", "strength_and_weaknesses": "Weaknesses:\n1. Incremental work. Replacing RNN with Transformer in deep reinforcement learning is not a new idea! There are so many works with similar ideas, such as Decision Transformer[1], GATO[2], [3], [4], [5], [6] and [7].\n2. Bad writing. The figures in this paper are really bad. For example, in Figure 2(a)(c), the blue curves are out of the figure.\n3. Toy benchmarks. The authors claim that they are better than DQN or DRQN, they should conduct experiments on popular RL benchmarks, such as Atari, MuJoco, and DeepMind Lab. \n4. Weak baselines. They should compare their method with real SOTA methods, such as PPO, Rainbow and IMPALA.\n\n\n- [1] Chen L, Lu K, Rajeswaran A, et al. Decision transformer: Reinforcement learning via sequence modeling[J]. Advances in neural information processing systems, 2021, 34: 15084-15097.\n- [2] Reed S, Zolna K, Parisotto E, et al. A generalist agent[J]. arXiv preprint arXiv:2205.06175, 2022.\n- [3] Laskin M, Wang L, Oh J, et al. In-context Reinforcement Learning with Algorithm Distillation[J]. arXiv preprint arXiv:2210.14215, 2022.\n- [4] Parisotto, Emilio, et al. \"Stabilizing transformers for reinforcement learning.\" International conference on machine learning. PMLR, 2020.\n- [5] Banino, Andrea, et al. \"Coberl: Contrastive bert for reinforcement learning.\" arXiv preprint arXiv:2107.05431 (2021).\n- [6] Furuta, Hiroki, Yutaka Matsuo, and Shixiang Shane Gu. \"Generalized decision transformer for offline hindsight information matching.\" arXiv preprint arXiv:2111.10364 (2021).\n- [7] Micheli, Vincent, Eloi Alonso, and François Fleuret. \"Transformers are sample efficient world models.\" arXiv preprint arXiv:2209.00588 (2022).\n", "clarity,_quality,_novelty_and_reproducibility": "- quality: low\n- clarity: low\n- originality: low", "summary_of_the_review": "This paper proposed a transformer-based DQN agent.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject"}, "tcdate": 1667542610996}, {"id": "D68lpCMAEuY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3826/Reviewer_zcAu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces deep transformer Q-networks (DTQN) for learning in partially observable environments. The authors train DTQN with a double Q-learning objective using on-policy samples. Given a trajectory of observations till now, observations are encoded using a transformer decoder with a masked attention to generate Q values for every step in trajectory. Typical Q learning loss with a fixed target network is applied to every step and the network is updated using the sum of losses from each step. When evaluated on several synthetic partially observable environments, DTQN is shown to perform better than LSTM variants including attention and traditional DQN. Ablation studies show that predicting Q values of intermediate steps is important for better performance, gating mechanism can help improve performance, and learning positional encodings is overall better but depends on the environment. Visualizing attention probabilities shows that bottleneck states are attended more.", "review_text": "While results on synthetic POMDPs are interesting, the model is derivative of previous work and no realistic task is studied.", "strengths": "**Strengths** Applying transformers to POMDPs is an interesting direction. The empirical results improve upon previous baselines including LSTMs and DQN.\n\n**Weaknesses**\n1. The paper improves RNNs using Transformers in a straightforward way. It is not clear what is the main difference of the Transformer studied in this work compared to something like DT and variants. DT uses observations, actions, and rewards while DTQN uses only observations. Later variants improve DT using reward prediction, multi-task learning etc. I believe environments like Atari are also partially observable. Aside from some ablations, new insights that would be worth investigating further could really help.\n\n2. Task that are studied are also synthetic. I think it is important to test the model on more challenging POMDPs even though current model fails to achieve strong performance.\n\n3. Additional baselines that are tailored towards RL is also needed. At least DT, GATO, or variants.\n\n4. Ablation experiments show that intermediate Q value prediction is important but it is not clear if this is due to simply using $k$ times more samples or actual intermediate Q value prediction. Are you using the same number of training samples for both experiments?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces deep transformer Q-networks (DTQN) for learning in partially observable environments. The authors train DTQN with a double Q-learning objective using on-policy samples. Given a trajectory of observations till now, observations are encoded using a transformer decoder with a masked attention to generate Q values for every step in trajectory. Typical Q learning loss with a fixed target network is applied to every step and the network is updated using the sum of losses from each step. When evaluated on several synthetic partially observable environments, DTQN is shown to perform better than LSTM variants including attention and traditional DQN. Ablation studies show that predicting Q values of intermediate steps is important for better performance, gating mechanism can help improve performance, and learning positional encodings is overall better but depends on the environment. Visualizing attention probabilities shows that bottleneck states are attended more.", "strength_and_weaknesses": "**Strengths** Applying transformers to POMDPs is an interesting direction. The empirical results improve upon previous baselines including LSTMs and DQN.\n\n**Weaknesses**\n1. The paper improves RNNs using Transformers in a straightforward way. It is not clear what is the main difference of the Transformer studied in this work compared to something like DT and variants. DT uses observations, actions, and rewards while DTQN uses only observations. Later variants improve DT using reward prediction, multi-task learning etc. I believe environments like Atari are also partially observable. Aside from some ablations, new insights that would be worth investigating further could really help.\n\n2. Task that are studied are also synthetic. I think it is important to test the model on more challenging POMDPs even though current model fails to achieve strong performance.\n\n3. Additional baselines that are tailored towards RL is also needed. At least DT, GATO, or variants.\n\n4. Ablation experiments show that intermediate Q value prediction is important but it is not clear if this is due to simply using $k$ times more samples or actual intermediate Q value prediction. Are you using the same number of training samples for both experiments?\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is written clearly and results on synthetic POMDP tasks show improvement. I found the model to be derivative of previous work on transformers in RL. The results should be reproducible.", "summary_of_the_review": "While results on synthetic POMDPs are interesting, the model is derivative of previous work and no realistic task is studied.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666729815828}, {"id": "8-uFKZmOnp", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3826/Reviewer_oBw9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new transformer-based architecture to solve POMDP problems. The main contributions include using transformer decoder structure to address partially observable RL domains and utilizing intermediate Q-values for the training. Results on four environment sets are reported.", "review_text": "This paper has a good motivation and proposes a new solution for the POMDP problem. Experiments on four environment sets demonstrate its effectiveness. The major concern is the lack of comparison with state-of-the-art/similar approaches.\n", "strengths": "Strength:\n1. The proposed transformer-based architecture is novel and the most effective one for DQN baselines.\n2. The paper is clearly presented and well-written.\n3. The experimental results demonstrate the effectiveness of DTQN in solving POMDP problems.\n\nWeakness:\n1. The proposed method is only compared with DQN and DRQN while there lacks evidence to show that DTQN can lead to state-of-the-art performance, which makes it less convincing.\n2. More discussions and comparison results between [1] and DTQN should be provided.\n3. [2] also presents a very similar architecture (but it uses transformer encoder). More reviews and comparison results should be added.\n4. GRU-like gating can lead to better results. Why not adopt the gating structure in the proposed architecture? More discussions are required.\n\n[1] TRANSFORMER BASED REINFORCEMENT LEARNING FOR GAMES.\n\n[2] Stabilizing Transformer-Based Action Sequence Generation For Q-Learning.\n\nMinor: Different structures in Section 5.3 can be illustrated in the supplementary material.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new transformer-based architecture to solve POMDP problems. The main contributions include using transformer decoder structure to address partially observable RL domains and utilizing intermediate Q-values for the training. Results on four environment sets are reported.", "strength_and_weaknesses": "Strength:\n1. The proposed transformer-based architecture is novel and the most effective one for DQN baselines.\n2. The paper is clearly presented and well-written.\n3. The experimental results demonstrate the effectiveness of DTQN in solving POMDP problems.\n\nWeakness:\n1. The proposed method is only compared with DQN and DRQN while there lacks evidence to show that DTQN can lead to state-of-the-art performance, which makes it less convincing.\n2. More discussions and comparison results between [1] and DTQN should be provided.\n3. [2] also presents a very similar architecture (but it uses transformer encoder). More reviews and comparison results should be added.\n4. GRU-like gating can lead to better results. Why not adopt the gating structure in the proposed architecture? More discussions are required.\n\n[1] TRANSFORMER BASED REINFORCEMENT LEARNING FOR GAMES.\n\n[2] Stabilizing Transformer-Based Action Sequence Generation For Q-Learning.\n\nMinor: Different structures in Section 5.3 can be illustrated in the supplementary material.", "clarity,_quality,_novelty_and_reproducibility": "The proposed architecture is novel for solving POMDP problems and the contributions are clearly presented. The code is also provided.", "summary_of_the_review": "This paper has a good motivation and proposes a new solution for the POMDP problem. Experiments on four environment sets demonstrate its effectiveness. The major concern is the lack of comparison with state-of-the-art/similar approaches.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666173196542}, {"id": "wSCnZ4TkNx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3826/Reviewer_QsGx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a deep reinforcement learning method for partially observable environments named Deep Transformer Q-Networks (DTQN). The key feature of DTQN is that, for a given chunk of history sampled from the replay buffer, a loss is taken over predicted Q-values for *all* timesteps in the history (rather than just the last one) by applying a causal mask in the transformer and thus efficiently providing a denser training signal for the agent. The architecture is evaluated on a number of tasks in partially observable environments and compared to a number of deep Q-learning baselines (one with no memory, one RNN-based, and an attention-based), in which DTQN is shown to perform best on average. A number of ablations are also conducted, demonstrating the importance of intermediate Q-value prediction as well as architectural choices such as positional encodings and how the attention outputs are combined.", "review_text": "This paper proposes a method with a clear motivation for tackling the important problem of efficient RL in partially observable environments. The key contribution of intermediate Q-value prediction can be an important one (and is empirically shown to give a boost to performance in relevant tasks), but more analysis is ideally required to demonstrate how it helps performance, and the ablations, while interesting, serve as a bit of a distraction to the main method.", "strengths": "Strengths\n- The problem that the method aims to tackle, RL in partially observable environments, is an important one, and the method itself is well-motivated, i.e. leveraging the transformer architecture to predict many Q-values at once for a denser training signal.\n- The approach is explained well, along with an informative diagram in Figure 1.\n- The related work section is strong, explicitly placing and contrasting the method in context of other transformer-based RL methods.\n- The tasks seem to be well-chosen, in the sense that they are partially observable and require some for of memory to complete.\n- The choice of baselines is reasonable, particularly when considered in combination with the ablations where, for example, the “Gate and identity” ablation is highlighted to be almost the same architecture as the GTrXL (just without the memory buffer) and thus serves as a relevant comparison to the state of the art in transformer architecture for RL.\n\nOpportunites for improvement\n- Ablations / analyses. While a number of ablations were performed and described in the paper, they might be improved in the following ways:\n    - In Table 1, it is shown that using gating in the combine step gives a better performance on average than the residual skip connection used in DTQN - given this, why is gating not used in the final DTQN architecture?\n    - While the ablations for the combine step and the positional encodings are interesting to see, they seem tangential to the key contribution of the method, which is the intermediate Q-value predictions. Indeed an ablation is performed on the intermediate predictions, demonstrating its significant effect on performance, but there could be more investigation into *how* it helps. For example, how much of the benefit of intermediate Q-value training comes from the extra training signal that comes from training on multiple states at once, versus a potential regularising effect that comes from the Q-value predictions of earlier states in the sequence having far less context than those later in the sequence? Perhaps this could be determined by running experiments that fix the number of predicted Q-values in training, but vary how much temporal context is used for each Q-value prediction.\n- While in the introduction, it is clear what the contribution of the method is (the intermediate Q-value prediction), throughout the paper the other architectural choices (which as far as I am aware are not novel, e.g. GTrXL used causal masking, positional encodings, but please correct me if I’m wrong) act as a bit of a distraction to this. I think the paper would be stronger if it focused more on the key contribution and understanding how it helps performance, as mentioned above.\n\nMinor comments\n- Table 1 it would be useful to bolden the top performing models on each task\n- Section 4.3 last line: “Figure 1” -> “Table 1”\n- Legend in Figure 2 quite small\n- Appendix E links to higher up in the paper\n- The naming of the ‘combine’ step is confusing, why not just call it a residual skip connection?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a deep reinforcement learning method for partially observable environments named Deep Transformer Q-Networks (DTQN). The key feature of DTQN is that, for a given chunk of history sampled from the replay buffer, a loss is taken over predicted Q-values for *all* timesteps in the history (rather than just the last one) by applying a causal mask in the transformer and thus efficiently providing a denser training signal for the agent. The architecture is evaluated on a number of tasks in partially observable environments and compared to a number of deep Q-learning baselines (one with no memory, one RNN-based, and an attention-based), in which DTQN is shown to perform best on average. A number of ablations are also conducted, demonstrating the importance of intermediate Q-value prediction as well as architectural choices such as positional encodings and how the attention outputs are combined.", "strength_and_weaknesses": "Strengths\n- The problem that the method aims to tackle, RL in partially observable environments, is an important one, and the method itself is well-motivated, i.e. leveraging the transformer architecture to predict many Q-values at once for a denser training signal.\n- The approach is explained well, along with an informative diagram in Figure 1.\n- The related work section is strong, explicitly placing and contrasting the method in context of other transformer-based RL methods.\n- The tasks seem to be well-chosen, in the sense that they are partially observable and require some for of memory to complete.\n- The choice of baselines is reasonable, particularly when considered in combination with the ablations where, for example, the “Gate and identity” ablation is highlighted to be almost the same architecture as the GTrXL (just without the memory buffer) and thus serves as a relevant comparison to the state of the art in transformer architecture for RL.\n\nOpportunites for improvement\n- Ablations / analyses. While a number of ablations were performed and described in the paper, they might be improved in the following ways:\n    - In Table 1, it is shown that using gating in the combine step gives a better performance on average than the residual skip connection used in DTQN - given this, why is gating not used in the final DTQN architecture?\n    - While the ablations for the combine step and the positional encodings are interesting to see, they seem tangential to the key contribution of the method, which is the intermediate Q-value predictions. Indeed an ablation is performed on the intermediate predictions, demonstrating its significant effect on performance, but there could be more investigation into *how* it helps. For example, how much of the benefit of intermediate Q-value training comes from the extra training signal that comes from training on multiple states at once, versus a potential regularising effect that comes from the Q-value predictions of earlier states in the sequence having far less context than those later in the sequence? Perhaps this could be determined by running experiments that fix the number of predicted Q-values in training, but vary how much temporal context is used for each Q-value prediction.\n- While in the introduction, it is clear what the contribution of the method is (the intermediate Q-value prediction), throughout the paper the other architectural choices (which as far as I am aware are not novel, e.g. GTrXL used causal masking, positional encodings, but please correct me if I’m wrong) act as a bit of a distraction to this. I think the paper would be stronger if it focused more on the key contribution and understanding how it helps performance, as mentioned above.\n\nMinor comments\n- Table 1 it would be useful to bolden the top performing models on each task\n- Section 4.3 last line: “Figure 1” -> “Table 1”\n- Legend in Figure 2 quite small\n- Appendix E links to higher up in the paper\n- The naming of the ‘combine’ step is confusing, why not just call it a residual skip connection?", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and well written. The novelty seems to be limited to the intermediate Q-value prediction, which can still be a significant contribution, especially if there were more analysis into why it helps performance. The paper states that code is provided, but I could not see it in the supplementary materials.", "summary_of_the_review": "This paper proposes a method with a clear motivation for tackling the important problem of efficient RL in partially observable environments. The key contribution of intermediate Q-value prediction can be an important one (and is empirically shown to give a boost to performance in relevant tasks), but more analysis is ideally required to demonstrate how it helps performance, and the ablations, while interesting, serve as a bit of a distraction to the main method.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666008400210}], "openreview_url": "https://openreview.net/forum?id=cddqs4kvC20", "arxiv_id": "2206.01078", "paper_pdf": "papers/cddqs4kvC20.pdf", "paper_pdf_sha256": "58fb3af6988cce055d6661a74a557f537555bdcfb1a8d9274ff314fc9a930df2", "paper_pdf_bytes": 1434690, "paper_pdf_source": "openreview", "code_url": "https://github.com/kevslinger/DTQN", "code_repository": "kevslinger/DTQN", "code_commit": "9f8d72bff5f0d607691989439ae93b9393b68a63", "code_archive": "repos/cddqs4kvC20.zip", "code_archive_sha256": "13776ed625ae110dcf4b68f175d2acf01b699977dbb022e37da4ba0533369dce", "code_archive_bytes": 52707, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 276, "github_languages": {"Python": 131947}, "github_archived": false, "github_pushed_at": "2024-07-07T08:44:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-transformer-q-networks-for-partially"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7pZiaojaVGU", "year": 2022, "status": "rejected", "title": "An Equivalence Between Data Poisoning and Byzantine Gradient Attacks", "authors": ["Sadegh Farhadkhani", "Rachid Guerraoui", "Lê-Nguyên Hoang", "Oscar Villemaud"], "authorids": ["~Sadegh_Farhadkhani1", "~Rachid_Guerraoui1", "~Lê-Nguyên_Hoang1", "~Oscar_Villemaud1"], "authors_source": "OpenReview API", "abstract": "To address the resilience of distributed learning, the ``Byzantine\" literature considers a strong threat model where workers can report arbitrary gradients to the parameter server. While this model helped generate several fundamental results, it has however sometimes been considered unrealistic, when the workers are mostly trustworthy machines. In this paper, we show a surprising equivalence between this model and data poisoning, a threat considered much more realistic. More specifically, we prove that any gradient attack can be reduced to data poisoning in a personalized federated learning system that provides PAC guarantees (which we show are both desirable and realistic in various personalized federated learning contexts such as linear regression and classification). Maybe most importantly, we derive a simple and practical attack that may be constructed against classical personalized federated learning models, and we show both theoretically and empirically the effectiveness of this attack.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "lGLL93CmUm", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1717/Reviewer_gjXp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies data and model poisoning attacks in federated learning (FL). It argues that there is an equivalence between data poisoning and model poisoning attacks, by restricting attention to linear and logistic regression. The main technique is to leverage ideas from PAC-learning. Another key contribution is to propose a gradient poisoning attack.", "review_text": "Model poisoning are generally perceived as stronger attacks (in terms of poisoning impact) than data poisoning attacks, simply because data poisoning only indirectly manipulates a client’s model update. Therefore, showing an equivalence as the paper argues would be very interesting. However, in my view, the paper is written in somewhat confusing manner, and lacks in details. Therefore, it is difficult to understand and assess the claims. (More detailed comments are given below.) Moreover, the paper does not compare the proposed Counter-Gradient attack with existing attacks. There are a large number of existing attacks and defenses, and it would be important to place the proposed attack in the larger context by comparing and contrasting with prior attacks and defenses.\n\nDetailed Comments:\n\n1. Definition 1 relies on ‘true models’ \\theta^{perp}. It will be good to give more details about what do the authors mean by true models. Propositions 1 and 2 on computing poisoned gradients also use true models. It is not clear how users will know about these true models? It will be important to elaborate this further. \n\n2. Is the difference between using sum vs. expectation in the loss not related to empirical vs. population risk? Also, computing the sum  of losses over the entire user dataset is essentially gradient descent, which is known to be computationally burdensome. It will be helpful to comment on these points.\n\n3. In Definition 2, a loss function is defined to be PAC-learnable. This is a bit confusing. PAC-learning is usually defined for a class of functions. It would be good to explicitly define the function class. \n\n4. In Theorem 2, how is \\tilde{Q} defined and how is the Supp() defined?\n\n5. Theorem 1 restricts the queries to be i.i.d. sub-Gaussian. This seems to be quite restrictive. It will be good to comment on this. \n\n6. Theorem 5 requires that a user’s dataset contains at least \\mathcal{I} inputs drawn from a certain model. This is quite confusing. What does it mean to draw inputs from a model?\n\n7. The paper says that [Shejwalkar et al. 2021] argued that model poisoning attacks are not realistic, and claims in the conclusion that their findings reverses this argument. It is not clear why an equivalence result would reverse the claim in [Shejwalkar et al. 2021]. This is because, as per my understanding, [Shejwalkar et al. 2021] also argue that data poisoning attacks are not very impactful (key lesson (2) on page 2). It would be important to give more details here. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies data and model poisoning attacks in federated learning (FL). It argues that there is an equivalence between data poisoning and model poisoning attacks, by restricting attention to linear and logistic regression. The main technique is to leverage ideas from PAC-learning. Another key contribution is to propose a gradient poisoning attack.", "main_review": "Model poisoning are generally perceived as stronger attacks (in terms of poisoning impact) than data poisoning attacks, simply because data poisoning only indirectly manipulates a client’s model update. Therefore, showing an equivalence as the paper argues would be very interesting. However, in my view, the paper is written in somewhat confusing manner, and lacks in details. Therefore, it is difficult to understand and assess the claims. (More detailed comments are given below.) Moreover, the paper does not compare the proposed Counter-Gradient attack with existing attacks. There are a large number of existing attacks and defenses, and it would be important to place the proposed attack in the larger context by comparing and contrasting with prior attacks and defenses.\n\nDetailed Comments:\n\n1. Definition 1 relies on ‘true models’ \\theta^{perp}. It will be good to give more details about what do the authors mean by true models. Propositions 1 and 2 on computing poisoned gradients also use true models. It is not clear how users will know about these true models? It will be important to elaborate this further. \n\n2. Is the difference between using sum vs. expectation in the loss not related to empirical vs. population risk? Also, computing the sum  of losses over the entire user dataset is essentially gradient descent, which is known to be computationally burdensome. It will be helpful to comment on these points.\n\n3. In Definition 2, a loss function is defined to be PAC-learnable. This is a bit confusing. PAC-learning is usually defined for a class of functions. It would be good to explicitly define the function class. \n\n4. In Theorem 2, how is \\tilde{Q} defined and how is the Supp() defined?\n\n5. Theorem 1 restricts the queries to be i.i.d. sub-Gaussian. This seems to be quite restrictive. It will be good to comment on this. \n\n6. Theorem 5 requires that a user’s dataset contains at least \\mathcal{I} inputs drawn from a certain model. This is quite confusing. What does it mean to draw inputs from a model?\n\n7. The paper says that [Shejwalkar et al. 2021] argued that model poisoning attacks are not realistic, and claims in the conclusion that their findings reverses this argument. It is not clear why an equivalence result would reverse the claim in [Shejwalkar et al. 2021]. This is because, as per my understanding, [Shejwalkar et al. 2021] also argue that data poisoning attacks are not very impactful (key lesson (2) on page 2). It would be important to give more details here. ", "summary_of_the_review": "It would be interesting to show an equivalence between model and data poisoning attacks in FL. However, the paper is written in somewhat confusing manner and lacks in details, which makes it difficult to understand and assess the claims.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636591707768}, {"id": "_VZLR2BtiG", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1717/Reviewer_xXNa"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper reveals the inherent equivalence between gradient and data poisoning attacks in personalized federated learning settings. The authors showed that any gradient attack can be transformed into data poisoning in a personalized federated learning system that provides PAC guarantees. This new insight challenges the view that (Byzantine) gradient attacks are unrealistic. The authors built this equivalence by constructing a model attack for personalized federated learning models. The authors showed the effectiveness of this attack both theoretically and empirically. ", "review_text": "Strengths:\n+ Reveals a new insight on the equivalence between Byzantine gradient and data poisoning attacks, which is an important and timely topic.\n+ A novel PAC framework for analyzing personalized federated learning performance.\n\nWeaknesses:\n- Gradient attack model is limited.\n- Experiments are limited to simple datasets.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper reveals the inherent equivalence between gradient and data poisoning attacks in personalized federated learning settings. The authors showed that any gradient attack can be transformed into data poisoning in a personalized federated learning system that provides PAC guarantees. This new insight challenges the view that (Byzantine) gradient attacks are unrealistic. The authors built this equivalence by constructing a model attack for personalized federated learning models. The authors showed the effectiveness of this attack both theoretically and empirically. ", "main_review": "Strengths:\n+ Reveals a new insight on the equivalence between Byzantine gradient and data poisoning attacks, which is an important and timely topic.\n+ A novel PAC framework for analyzing personalized federated learning performance.\n\nWeaknesses:\n- Gradient attack model is limited.\n- Experiments are limited to simple datasets.", "summary_of_the_review": "This paper reveals an equivalence between Byzantine gradient and data poisoning attacks in the context of personalized federated learning, which is an important and timely topic. This insight suggests that the claims in the existing literature that Byzantine gradient attacks are unrealistic is misleading. To my knowledge, this result is new and its contributions to the field are significant. However, this paper also has several issues, which are listed as follows:\n\n1. The attack model considered in this paper is somewhat simplistic in that the authors only considered the single-strategic-user case. Although it is understandable that this renders the problem more tractable for theoretical analysis, the results may not be very useful because Byzantine gradient attacks are not necessarily from a single malicious user. In fact, most works on Byzantine gradient attacks allow multiple Byzantine workers. It's unclear whether the results in this paper could be extended to multi-attacker scenarios.\n\n2. The authors proposed a PAC framework as a foundation to evaluate the performance of various Byzantine and data poisoning attacks. To my knowledge, this is also a new and interesting contribution. The authors demonstrated the relevance of this PAC framework by showing that linear regression and classifications are PAC learnable under this framework. But I wonder whether this PAC framework continues to be meaningful for more complex learning models than linear models. Having further discussions on this aspect would be very interesting.\n\n3. Most of the experiments in this paper are conducted on MNIST and Fashion-MNIST datasets, which are relatively simple. It would be more interesting to demonstrate the claimed equivalence between attacks on more sophisticated datasets. Also related to the previous comment, most of the experiments are based on linear models (a simple two-layer neural network is also used). I think this paper could benefit from more experiments with more sophisticated learning models.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636366721927}, {"id": "UudDZlYz36m", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1717/Reviewer_oiiD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper shows an equivalence between data poisoning attacks and gradient attacks that attack gradient descent on personalized federated learning setting for certain convex loss functions. ", "review_text": "The goal of this paper to compare two threat models in federated learning 1) data poisoning attacks that inject a number of poisoning examples to the training set 2) gradient attacks that change the update gradients sent during the federated learning operation by malicious agents. They specifically study the personalized federated learning setting where N machines collaborate to train N personalized models and a single global model. In this setting, the loss function is defined in a way that the global model and personalized models have dependencies to each other through an additive regularization that controls how far the personalized models can be from the global model.\n\nThe paper first defines two notions that they call PAC^* Learning and gradient PAC* learning for the personalized federated learning. The notion of PAC* learning is about finding a set of personalized models that are close to the set of true models with respect according to some metric. Note that this is different from PAC learning because they care about distance, rather than accuracy. They also define gradient-PAC*, which requires the gradients of the personalized models with respect to any global model \\rho to be close to that of true models.\n\nThen, they introduce a hypothetical threat model called a model attack, where they fix one of the personalized models to a specially crafted adversarial model. Then they show that, if the optimization of loss function has PAC* learning abilities, then the adversary can simulate the effect of the model attack using data poisoning attack, by poisoning all of the data for the machine, if the number of data-points provided by that machine is larger than some threshold. This result is expected, because as the number of data-points grow, the dependence between the personalized model and the global model decrease (because of the diminishing effect of the regularization). Then, by increasing the number of data points, because of PAC* learning, the adversary can change the personalized model to a adversarially crafted model, by using queries that are labeled accordingly. At the same time, the global model and other personalized models would not change much because of the diminish dependency of the global and personalized models.  \n\nNow the remaining step is to show that the gradient attack can be simulated by a model attack. In order to show this, they need some strong assumptions. For instance, they need the loss function to be strongly convex. They need the poisonous gradients to be valid gradients according to the global model. and also they require the poisonous gradients to converge. This step is technically interesting and novel, in my opinion. \n\nFinally, they design a data poisoning attacks that are inspired by the steps mentioned above. They empirically show that this attack is effective in linear regression.\n\n\nLimitations:\n\n- The final equivalence result of the paper only applies to simple settings with strongly convex loss functions and seems to only work for GD optimization.\n\n- The presentation of the work is not in publication stage. It is really hard to understand the main claims of the paper.\n\n- Some discussions about the implication of the results are missing\n\n\nComments/Questions to authors:\n\n1- I have read through the paper multiple times, but I still do not understand the exact threat model of gradient attack. I think you need to define this explicitly.\n\n2- The notions of PAC* learning are not defined properly. There are terms such as \"true model\"  and \"honest dataset\" that are not defined. These terms are not standard and must be defined. Also, there is a probability \\delta, what is this probability over? \n\n3- Section 3 seems to be completely out of context. Why do you need to justify the choice of loss function before describing the main results? I was really confused by this section. After reading the paper, I realize that you need to define the loss this way in order for Theorem 5 to hold? If this is correct, you need to explicitly mention it. I also find Table 1 confusing. \n\n4- Defining the loss function as a summation of the loss is fine, but your notion of PAC* learning talks about large number of data points. Wouldn't large number of data points make the dependence of global and personalized models weaker as more data is provided? How is it justified to keep the regularization the same as more data is provided?\n\n5- Again, section 4 seems to be out of place. It is a bit odd to see the main results of the paper in section 5. I would recommend to  change the order of sections 4 and 5.\n\n6- The equivalence between the model attacks and data poisoning attack depend on the number of data points. I think the paper needs discussion on this connection. Could the power of model attacks be significantly higher than data poisoning attacks when the number of data points provided by each party is a fixed and perhaps small number?\n\n7- Theorem 6 seems like the main theorem in the paper to me. But I cannot verify the proof. Specifically, I cannot verify equation (120) in page 33. Can you explain why this is the case? Aren't \\rho and \\theta_n getting optimized together. In your formulation it seems like \\theta_n is optimized until convergence for each intermediate \\rho. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper shows an equivalence between data poisoning attacks and gradient attacks that attack gradient descent on personalized federated learning setting for certain convex loss functions. ", "main_review": "The goal of this paper to compare two threat models in federated learning 1) data poisoning attacks that inject a number of poisoning examples to the training set 2) gradient attacks that change the update gradients sent during the federated learning operation by malicious agents. They specifically study the personalized federated learning setting where N machines collaborate to train N personalized models and a single global model. In this setting, the loss function is defined in a way that the global model and personalized models have dependencies to each other through an additive regularization that controls how far the personalized models can be from the global model.\n\nThe paper first defines two notions that they call PAC^* Learning and gradient PAC* learning for the personalized federated learning. The notion of PAC* learning is about finding a set of personalized models that are close to the set of true models with respect according to some metric. Note that this is different from PAC learning because they care about distance, rather than accuracy. They also define gradient-PAC*, which requires the gradients of the personalized models with respect to any global model \\rho to be close to that of true models.\n\nThen, they introduce a hypothetical threat model called a model attack, where they fix one of the personalized models to a specially crafted adversarial model. Then they show that, if the optimization of loss function has PAC* learning abilities, then the adversary can simulate the effect of the model attack using data poisoning attack, by poisoning all of the data for the machine, if the number of data-points provided by that machine is larger than some threshold. This result is expected, because as the number of data-points grow, the dependence between the personalized model and the global model decrease (because of the diminishing effect of the regularization). Then, by increasing the number of data points, because of PAC* learning, the adversary can change the personalized model to a adversarially crafted model, by using queries that are labeled accordingly. At the same time, the global model and other personalized models would not change much because of the diminish dependency of the global and personalized models.  \n\nNow the remaining step is to show that the gradient attack can be simulated by a model attack. In order to show this, they need some strong assumptions. For instance, they need the loss function to be strongly convex. They need the poisonous gradients to be valid gradients according to the global model. and also they require the poisonous gradients to converge. This step is technically interesting and novel, in my opinion. \n\nFinally, they design a data poisoning attacks that are inspired by the steps mentioned above. They empirically show that this attack is effective in linear regression.\n\n\nLimitations:\n\n- The final equivalence result of the paper only applies to simple settings with strongly convex loss functions and seems to only work for GD optimization.\n\n- The presentation of the work is not in publication stage. It is really hard to understand the main claims of the paper.\n\n- Some discussions about the implication of the results are missing\n\n\nComments/Questions to authors:\n\n1- I have read through the paper multiple times, but I still do not understand the exact threat model of gradient attack. I think you need to define this explicitly.\n\n2- The notions of PAC* learning are not defined properly. There are terms such as \"true model\"  and \"honest dataset\" that are not defined. These terms are not standard and must be defined. Also, there is a probability \\delta, what is this probability over? \n\n3- Section 3 seems to be completely out of context. Why do you need to justify the choice of loss function before describing the main results? I was really confused by this section. After reading the paper, I realize that you need to define the loss this way in order for Theorem 5 to hold? If this is correct, you need to explicitly mention it. I also find Table 1 confusing. \n\n4- Defining the loss function as a summation of the loss is fine, but your notion of PAC* learning talks about large number of data points. Wouldn't large number of data points make the dependence of global and personalized models weaker as more data is provided? How is it justified to keep the regularization the same as more data is provided?\n\n5- Again, section 4 seems to be out of place. It is a bit odd to see the main results of the paper in section 5. I would recommend to  change the order of sections 4 and 5.\n\n6- The equivalence between the model attacks and data poisoning attack depend on the number of data points. I think the paper needs discussion on this connection. Could the power of model attacks be significantly higher than data poisoning attacks when the number of data points provided by each party is a fixed and perhaps small number?\n\n7- Theorem 6 seems like the main theorem in the paper to me. But I cannot verify the proof. Specifically, I cannot verify equation (120) in page 33. Can you explain why this is the case? Aren't \\rho and \\theta_n getting optimized together. In your formulation it seems like \\theta_n is optimized until convergence for each intermediate \\rho. ", "summary_of_the_review": "I find the topic of this paper extremely interesting. Understanding the relation between gradient attacks and data poisoning attacks are very important. This paper takes an initial step in understanding the relation between these two attacks for simple models such as linear regression. However, as stated in my comments, I have serious concerns about the presentation of this work. The main ideas presented are not easy to comprehend. The exact threat models are not also clearly specified. I also have some concerns about the proofs (In particular about equation 120, page 33). I will be happy to increase my score if authors can provide sufficient response to my concerns.\n\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636332088945}, {"id": "uYFeJQBt9AT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1717/Reviewer_a13A"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper formalizes the concept of PAC*-learnability for a generalized personalized federated learning model and provides a sufficient condition. Under this model, converging gradient attack is equivalent to data poisoning. The paper also proposes the counter-gradient attack that is effective and data-efficient.", "review_text": "The considered problem is of great contemporary importance in the deployment of large-scale machine learning systems. 1. The formalization of the PAC learnability is intuitive and concrete examples linear regression, logistic regression, are analyzed to illustrate this definition. 2. The equivalence between data poisoning, model attacks, and gradient attacks are carefully analyzed. 3. The proposed CGA attack has theoretical guarantee and demonstrates good empirical performance.\n\nIn general, I think the assumptions used in this paper are strong. 1. The sufficient condition proposed for PAC* learnability is not intuitive to me, is more like a sufficient condition for the two concrete examples analyzed, and lack intuitions for a wider class of functions. Like, does this condition or the PAC* learnability holds for any non-convex loss used in practice? 2. It might also help to have more detailed explanations for the considered data poisoning, model attacks and gradient attacks, like giving concrete examples to illustrate. The assumptions for theorem 6 to hold also seem strong to me, it is not clear to me how to interpret the assumed property of R, especially R(\\rho, \\theta) = R_0(\\rho - \\theta). It might be clearer to draw a diagram to show under which conditions the claimed equivalences hold, since the claim in the paper title and abstract is fairly strong. 3. I also would suggest giving proof outlines for main theorems for the paper. 4. Another question is how should we relate the proposed CGA with the previous sections, is there any motivations therein?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper formalizes the concept of PAC*-learnability for a generalized personalized federated learning model and provides a sufficient condition. Under this model, converging gradient attack is equivalent to data poisoning. The paper also proposes the counter-gradient attack that is effective and data-efficient.", "main_review": "The considered problem is of great contemporary importance in the deployment of large-scale machine learning systems. 1. The formalization of the PAC learnability is intuitive and concrete examples linear regression, logistic regression, are analyzed to illustrate this definition. 2. The equivalence between data poisoning, model attacks, and gradient attacks are carefully analyzed. 3. The proposed CGA attack has theoretical guarantee and demonstrates good empirical performance.\n\nIn general, I think the assumptions used in this paper are strong. 1. The sufficient condition proposed for PAC* learnability is not intuitive to me, is more like a sufficient condition for the two concrete examples analyzed, and lack intuitions for a wider class of functions. Like, does this condition or the PAC* learnability holds for any non-convex loss used in practice? 2. It might also help to have more detailed explanations for the considered data poisoning, model attacks and gradient attacks, like giving concrete examples to illustrate. The assumptions for theorem 6 to hold also seem strong to me, it is not clear to me how to interpret the assumed property of R, especially R(\\rho, \\theta) = R_0(\\rho - \\theta). It might be clearer to draw a diagram to show under which conditions the claimed equivalences hold, since the claim in the paper title and abstract is fairly strong. 3. I also would suggest giving proof outlines for main theorems for the paper. 4. Another question is how should we relate the proposed CGA with the previous sections, is there any motivations therein?\n", "summary_of_the_review": "I would recommend that the paper is marginally below the acceptance threshold. Mainly because I don’t quite follow the intuitions of some assumptions. Refined analysis with weaker assumptions and clearer presentation would be appreciated.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635997097607}, {"id": "OD8QKqa4XXf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1717/Reviewer_bDXf"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the relationship between Byzantine gradient attacks in distributed optimization and data poisoning. Some theoretical evidence is presented about the equivalence of these. A new gradient-based Byzantine attack is also presented.", "review_text": "Pros:\n\nThe paper studies an interesting problem, since both Byzantine attacks and data poisoning attacks have been extensive studied in the literature, but little is known about how these fields relate. The paper provides a substantial technical analysis, as well as a practical Byzantine attack.\n \n\n##########################################################################\n\nCons: \n\nUnfortunately, several concern prevent me from recommending acceptance. \n\n- Firstly, I find the comparison to related work very insufficient. \n\nFor the theoretical analysis a number of recent works, for example [1,2,3,4], have explored (robust) machine learning from multiple datasets from a PAC perspective. These essentially correspond to the notion of PAC* learnability presented here, with and without data poisoning. It is unexplained how these works compare to the current one. \n\nFor the gradient attack that is designed, there is no comparison to prior attacks as well. It is also unclear what this attack has to do with the rest of the paper - can the authors explain?\n\n- The paper is also slightly hard to follow, in particular the results of Theorem 4,5 and 6 are quite hard for me to interpret. Since there is no previously defined notions of success of for each of the attacks, it is hard to understand why the results imply any type of equivalence between the notions. Further discussion immediately after these results will therefore be helpful.\n\n#########################################################################\n\nReferences:\n\n[1] A. Blum, N. Haghtalab, A. D. Procaccia, M. Qiao: Collaborative PAC Learning, NIPS 2017\n[2] M. Qiao: Do outliers ruin collaboration?, ICML 2018\n[3] N. Konstantinov, E. Frantar, D. Alistarh, C.H. Lampert: On the Sample Complexity of Adversarial Multi-Source PAC Learning, ICML 2020\n[4] A. Jain and A. Orlitsky: A General Method for Robust Learning from Batches, NeurIPS 2020\n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies the relationship between Byzantine gradient attacks in distributed optimization and data poisoning. Some theoretical evidence is presented about the equivalence of these. A new gradient-based Byzantine attack is also presented.", "main_review": "Pros:\n\nThe paper studies an interesting problem, since both Byzantine attacks and data poisoning attacks have been extensive studied in the literature, but little is known about how these fields relate. The paper provides a substantial technical analysis, as well as a practical Byzantine attack.\n \n\n##########################################################################\n\nCons: \n\nUnfortunately, several concern prevent me from recommending acceptance. \n\n- Firstly, I find the comparison to related work very insufficient. \n\nFor the theoretical analysis a number of recent works, for example [1,2,3,4], have explored (robust) machine learning from multiple datasets from a PAC perspective. These essentially correspond to the notion of PAC* learnability presented here, with and without data poisoning. It is unexplained how these works compare to the current one. \n\nFor the gradient attack that is designed, there is no comparison to prior attacks as well. It is also unclear what this attack has to do with the rest of the paper - can the authors explain?\n\n- The paper is also slightly hard to follow, in particular the results of Theorem 4,5 and 6 are quite hard for me to interpret. Since there is no previously defined notions of success of for each of the attacks, it is hard to understand why the results imply any type of equivalence between the notions. Further discussion immediately after these results will therefore be helpful.\n\n#########################################################################\n\nReferences:\n\n[1] A. Blum, N. Haghtalab, A. D. Procaccia, M. Qiao: Collaborative PAC Learning, NIPS 2017\n[2] M. Qiao: Do outliers ruin collaboration?, ICML 2018\n[3] N. Konstantinov, E. Frantar, D. Alistarh, C.H. Lampert: On the Sample Complexity of Adversarial Multi-Source PAC Learning, ICML 2020\n[4] A. Jain and A. Orlitsky: A General Method for Robust Learning from Batches, NeurIPS 2020\n\n ", "summary_of_the_review": "While the studied topic and the proposed notions and results are interesting, I believe that the paper will benefit from substantial further comparison to existing work and also from further discussion of the results from Theorems 4,5,6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635801222080}], "openreview_url": "https://openreview.net/forum?id=7pZiaojaVGU", "arxiv_id": "2202.08578", "paper_pdf": "papers/7pZiaojaVGU.pdf", "paper_pdf_sha256": "056fd0b7945253687a4b3140cd1abcaafc9aaed9d7ab872080cebe9637c8d4b8", "paper_pdf_bytes": 1479640, "paper_pdf_source": "openreview", "code_url": "https://github.com/LPD-EPFL/Attack_Equivalence", "code_repository": "LPD-EPFL/Attack_Equivalence", "code_commit": "953769950fc9a25d212ecae3df93b735d61658af", "code_archive": "repos/7pZiaojaVGU.zip", "code_archive_sha256": "e965d8e050fd63762a33d2f397ab7ccaa8770c083d1a91afdad7947f3384f035", "code_archive_bytes": 389595, "code_file_count": 6, "code_extensions": {".ipynb": 6}, "github_disk_usage_kb": 712, "github_languages": {"Jupyter Notebook": 969609}, "github_archived": false, "github_pushed_at": "2022-06-17T11:03:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/an-equivalence-between-data-poisoning-and-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JeweO9-QqV-", "year": 2021, "status": "rejected", "title": "SoGCN: Second-Order Graph Convolutional Networks", "authors": ["Peihao Wang", "Yuehao Wang", "Hua Lin", "Jianbo Shi"], "authorids": ["~Peihao_Wang1", "~Yuehao_Wang1", "~Hua_Lin1", "~Jianbo_Shi1"], "authors_source": "OpenReview API", "abstract": "We introduce a second-order graph convolution (SoGC), a maximally localized kernel, that can express a polynomial spectral filter with arbitrary coefficients.  We contrast our SoGC with vanilla GCN, first-order (one-hop) aggregation, and higher-order (multi-hop) aggregation by analyzing graph convolutional layers via generalized filter space.  We argue that SoGC is a simple design capable of forming the basic building block of graph convolution, playing the same role as $3 \\times 3$ kernels in CNNs.   We build purely topological Second-Order Graph Convolutional Networks (SoGCN) and demonstrate that SoGCN consistently achieves state-of-the-art performance on the latest benchmark. Moreover, we introduce the Gated Recurrent Unit (GRU) to spectral GCNs. This explorative attempt further improves our experimental results.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "vCCoqQ4J-Lj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1445/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors argue that second-order graph convolutions (SoGC) should be the building blocks for future graph networks. The argument is that some of the second-order functions cannot be represented by stacking first-order graph convolutions. In contrast, second-order can represent any higher-order ones. In general, the theory is sound and the results on the synthetic data and ZINC are strong and I don't see many flaws, but I would like to see more results on realistic datasets.\n\n##### Strengths\n- The method is quite simple.\n- The results on the synthetic dataset and ZINC are quite impressive.\n- The visualization in Figure 5 convinces me that SoGC can deal with the oversmoothing problem of vanilla GCNs.\n- The theoretical insights and proofs look correct to me, while I am not 100% confident. \n\n##### Weaknesses\n- MNIST and CIFAR 10 do not seem to be real use cases for Graph models. According to the [homepage of MNIST](http://yann.lecun.com/exdb/mnist/), none of the MNIST results in Table 2 is better than an SVM with degree 4 polynomial kernels, a 2-layer MLP with 800d, or a LeNet-1 which has only 3k parameters instead of 100k. It would be better to see results on more realistic datasets where graph networks are close to state-of-the-art.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple method, interesting observations", "review": "The authors argue that second-order graph convolutions (SoGC) should be the building blocks for future graph networks. The argument is that some of the second-order functions cannot be represented by stacking first-order graph convolutions. In contrast, second-order can represent any higher-order ones. In general, the theory is sound and the results on the synthetic data and ZINC are strong and I don't see many flaws, but I would like to see more results on realistic datasets.\n\n##### Strengths\n- The method is quite simple.\n- The results on the synthetic dataset and ZINC are quite impressive.\n- The visualization in Figure 5 convinces me that SoGC can deal with the oversmoothing problem of vanilla GCNs.\n- The theoretical insights and proofs look correct to me, while I am not 100% confident. \n\n##### Weaknesses\n- MNIST and CIFAR 10 do not seem to be real use cases for Graph models. According to the [homepage of MNIST](http://yann.lecun.com/exdb/mnist/), none of the MNIST results in Table 2 is better than an SVM with degree 4 polynomial kernels, a 2-layer MLP with 800d, or a LeNet-1 which has only 3k parameters instead of 100k. It would be better to see results on more realistic datasets where graph networks are close to state-of-the-art.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604009008360}, {"id": "rgDNyZYxSwb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1445/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The submission identifies the importance of second-order filter by showing that two is the minimally necessary order to achieve full representation power. Based on this observation, this paper proposes Second-Order Graph Convolutional Networks. The proposed method is evaluated on graph classification benchmarks and demonstrates good empirical performance. As far as I know, the proposed method is novel and can inspire future analysis on the expressive power of graph convolution. \n\nStrength:\n- The proposed method, SoGC, is theoretically motivated by the fact that second order graph convolution has universal representation power (Thm. 1)\n- The proposed method demonstrates strong empirical performance on synthetic and real-world datasets\n\nWeakness and Questions:\n- The usage of GRU feels ad-hoc to me. Without using GRU, the proposed method in fact does not achieve state-of-the-art performance on CIFAR10 and MNIST, which is concerning. I understand that the authors attempt to analyze the effect of GRU in Section 5.1. However, the claim that \"GRU retains information from previous layers effectively ...\" needs more supporting evidence than Figure 2, which is conducted on a single graph from a single dataset. I encourage the authors to provide more in-depth analysis of GRU and connects it to the rest of the paper. \nFor more low-level questions: \n     - How is the input node order to the GRU decided? Will changing the input order affect results?\n     - In Figure 2, is the number of parameters controlled? \n     - Would GRU also improves the performance of vanilla GCN? As Figure 2 showed, GRU also changes the signal of a vanilla GCN significantly.\n- I would like to see an ablation study between the non-linear and linear versions of SoGCN. Given that in section 3 the analysis is only done for the linear version, I am curious about how well the linear model can perform. Also, the theory predicts that the linear Second Order GCN should outperform linear Vanilla GCN (this is like SGC?). I would like to see this prediction gets tested.\n- As the authors acknowledge in the introduction, Abu-El-Haija et al. (2019), MixHop, makes similar observations about higher-order GCN. I am curious about how these two models compare to each other performance-wise.\n- Minor note: The submission only evaluates the proposed method on graph classification task. Will the proposed method works for node classification or link prediction? Even if it doesn't, the authors should document the limitations of this method if it doesn't work for those tasks. \n\nTypos:\n- Page 1, the second to last paragraph: \"in channel-wise filtering Furthermore\" -> missing period\n- Page 6, \"we can whiteness the ineffectiveness\" -> \"witness\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": " Second-Order Graph Convolutional Networks", "review": "The submission identifies the importance of second-order filter by showing that two is the minimally necessary order to achieve full representation power. Based on this observation, this paper proposes Second-Order Graph Convolutional Networks. The proposed method is evaluated on graph classification benchmarks and demonstrates good empirical performance. As far as I know, the proposed method is novel and can inspire future analysis on the expressive power of graph convolution. \n\nStrength:\n- The proposed method, SoGC, is theoretically motivated by the fact that second order graph convolution has universal representation power (Thm. 1)\n- The proposed method demonstrates strong empirical performance on synthetic and real-world datasets\n\nWeakness and Questions:\n- The usage of GRU feels ad-hoc to me. Without using GRU, the proposed method in fact does not achieve state-of-the-art performance on CIFAR10 and MNIST, which is concerning. I understand that the authors attempt to analyze the effect of GRU in Section 5.1. However, the claim that \"GRU retains information from previous layers effectively ...\" needs more supporting evidence than Figure 2, which is conducted on a single graph from a single dataset. I encourage the authors to provide more in-depth analysis of GRU and connects it to the rest of the paper. \nFor more low-level questions: \n     - How is the input node order to the GRU decided? Will changing the input order affect results?\n     - In Figure 2, is the number of parameters controlled? \n     - Would GRU also improves the performance of vanilla GCN? As Figure 2 showed, GRU also changes the signal of a vanilla GCN significantly.\n- I would like to see an ablation study between the non-linear and linear versions of SoGCN. Given that in section 3 the analysis is only done for the linear version, I am curious about how well the linear model can perform. Also, the theory predicts that the linear Second Order GCN should outperform linear Vanilla GCN (this is like SGC?). I would like to see this prediction gets tested.\n- As the authors acknowledge in the introduction, Abu-El-Haija et al. (2019), MixHop, makes similar observations about higher-order GCN. I am curious about how these two models compare to each other performance-wise.\n- Minor note: The submission only evaluates the proposed method on graph classification task. Will the proposed method works for node classification or link prediction? Even if it doesn't, the authors should document the limitations of this method if it doesn't work for those tasks. \n\nTypos:\n- Page 1, the second to last paragraph: \"in channel-wise filtering Furthermore\" -> missing period\n- Page 6, \"we can whiteness the ineffectiveness\" -> \"witness\"\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603918369952}, {"id": "naBYXnNKMH0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1445/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This is yet another paper on graph convolutional networks (GCNs). The investigated SoGCN is a second-order GCN, thus a special case of high-order GCNs (namely with multi-hop graph kernels), which have has been proposed earlier by many researchers, such as by Defferrard et al. (2016), by Kipf & Welling (2017) and by Abu-El-Haija et al. (2019). \n\nThe main interest is that a second-order GC is a universal approximator, because any univariate polynomial can be factorized into sub-polynomials of degree two, which is not the case of first-order GCs. \n\nThe paper has some theoretical derivations that seem interesting. However, the proposed second-order GC is a special case of well-studied frameworks, which have been widely studied in the literature, such as with the expressive power analysis. As a consequence, the derived theoretical results seem less relevant. \n\nIn conducted experiments, the authors do not consider the same comparative analysis in all experiments. For example, two versions of the nonlinear activation are considered in experiments on CIFAR10 (with and without GRU); however, they were not considered in other experiments. Moreover, it is not clear where the combined RELU and GRU is considered.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This is yet another paper on graph convolutional networks, considering a special case of high-order kernels.", "review": "This is yet another paper on graph convolutional networks (GCNs). The investigated SoGCN is a second-order GCN, thus a special case of high-order GCNs (namely with multi-hop graph kernels), which have has been proposed earlier by many researchers, such as by Defferrard et al. (2016), by Kipf & Welling (2017) and by Abu-El-Haija et al. (2019). \n\nThe main interest is that a second-order GC is a universal approximator, because any univariate polynomial can be factorized into sub-polynomials of degree two, which is not the case of first-order GCs. \n\nThe paper has some theoretical derivations that seem interesting. However, the proposed second-order GC is a special case of well-studied frameworks, which have been widely studied in the literature, such as with the expressive power analysis. As a consequence, the derived theoretical results seem less relevant. \n\nIn conducted experiments, the authors do not consider the same comparative analysis in all experiments. For example, two versions of the nonlinear activation are considered in experiments on CIFAR10 (with and without GRU); however, they were not considered in other experiments. Moreover, it is not clear where the combined RELU and GRU is considered.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603834012102}, {"id": "qjH00SIdTUy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1445/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a so-called second-order graph convolution, where an additional second-order term is introduced into traditional first-order graph convolution. The authors explain the merits of second-order graph convolution from the perspective of representation ability. The resulted second-order graph convolutional networks are compared with several graph convolutional networks on three benchmarks. \n\nStrengths:\n+: Incorporation of second-order information into graph convolution seems interesting, and such analogous idea has been verified in 2D convolutions.\n\n+: The proposed method is simple and easy to implement.\n\nWeaknesses:\n-: The idea on incorporation of second-order or higher-order information into convolution is not novel. For example, Factorized Bilinear (FB) [r1] and Second-Order Response Transform (SORT) [r2] introduce second-order terms into traditional 2D convolutions, and they also claim second-order terms have better representation ability. Besides, second-order or higher-order information have also been used for global pooling for convolution networks [r3, r4, r5], which also show better representation ability. This paper lacks discussions on above these works, which will bring a side effect on contributions of this paper.\n[r1] Factorized bilinear models for image recognition, ICCV 2017.\n[r2] SORT: Second-order response transform for visual recognition, ICCV 2017.\n[r3] Second-Order Pooling for Graph Neural Networks, arXiv, 2020.\n[r4] Deep CNNs Meet Global Covariance Pooling: Better Representation and Generalization, TPAMI 2020.\n[r5] Kernel pooling for convolutional neural networks, CVPR, 2020.\n\n-: The experimental results are not very convincing.\n(1) As shown in Table 2, pure soGCN achieves no improvement over compared methods, i.e., GatedGCN. GRU brings further gains for soGCN, but could GRU bring improvement for other compared methods?\n(2) Number of parameters hardly represents model complexity totally. The model with the same number of parameters can have different computational complexity. Therefore, more metrics on model complexity (e.g., FLOPs) are suggested for comparison in Table 2.\n(3) CIFAR10 and MNIST are too small and old to verify the effectiveness of different methods. The authors would better conduct experiments on more graph benchmarks. Besides, why higher-order  GCN are not compared on real-world benchmarks.\n(4) Why parameters number of 3WLGNN is 100K on ZINC ?\n\n-: The writing needs significant improvement.\n(1) The authors would better give more detailed descriptions on differences between the proposed soGCN and related works ([Defferrard, 2016] and [Kipf&Welling, 2017]), further clarifying the contributions of the proposed method.\n(2) I wonder the detailed computation methods and each curve in Fig.2, and why SoGCN is better than Vanilla GCN?\n(3) The comparisons in terms of representation ability in section 4.3 is not very clear. The authors would better add a table to summarize representation ability of different graph convolution.\n(4) Which method does MoNet indicate in Table 2 ?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for SoGCN: Second-Order Graph Convolutional Networks", "review": "This paper proposes a so-called second-order graph convolution, where an additional second-order term is introduced into traditional first-order graph convolution. The authors explain the merits of second-order graph convolution from the perspective of representation ability. The resulted second-order graph convolutional networks are compared with several graph convolutional networks on three benchmarks. \n\nStrengths:\n+: Incorporation of second-order information into graph convolution seems interesting, and such analogous idea has been verified in 2D convolutions.\n\n+: The proposed method is simple and easy to implement.\n\nWeaknesses:\n-: The idea on incorporation of second-order or higher-order information into convolution is not novel. For example, Factorized Bilinear (FB) [r1] and Second-Order Response Transform (SORT) [r2] introduce second-order terms into traditional 2D convolutions, and they also claim second-order terms have better representation ability. Besides, second-order or higher-order information have also been used for global pooling for convolution networks [r3, r4, r5], which also show better representation ability. This paper lacks discussions on above these works, which will bring a side effect on contributions of this paper.\n[r1] Factorized bilinear models for image recognition, ICCV 2017.\n[r2] SORT: Second-order response transform for visual recognition, ICCV 2017.\n[r3] Second-Order Pooling for Graph Neural Networks, arXiv, 2020.\n[r4] Deep CNNs Meet Global Covariance Pooling: Better Representation and Generalization, TPAMI 2020.\n[r5] Kernel pooling for convolutional neural networks, CVPR, 2020.\n\n-: The experimental results are not very convincing.\n(1) As shown in Table 2, pure soGCN achieves no improvement over compared methods, i.e., GatedGCN. GRU brings further gains for soGCN, but could GRU bring improvement for other compared methods?\n(2) Number of parameters hardly represents model complexity totally. The model with the same number of parameters can have different computational complexity. Therefore, more metrics on model complexity (e.g., FLOPs) are suggested for comparison in Table 2.\n(3) CIFAR10 and MNIST are too small and old to verify the effectiveness of different methods. The authors would better conduct experiments on more graph benchmarks. Besides, why higher-order  GCN are not compared on real-world benchmarks.\n(4) Why parameters number of 3WLGNN is 100K on ZINC ?\n\n-: The writing needs significant improvement.\n(1) The authors would better give more detailed descriptions on differences between the proposed soGCN and related works ([Defferrard, 2016] and [Kipf&Welling, 2017]), further clarifying the contributions of the proposed method.\n(2) I wonder the detailed computation methods and each curve in Fig.2, and why SoGCN is better than Vanilla GCN?\n(3) The comparisons in terms of representation ability in section 4.3 is not very clear. The authors would better add a table to summarize representation ability of different graph convolution.\n(4) Which method does MoNet indicate in Table 2 ?\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603769408609}], "openreview_url": "https://openreview.net/forum?id=JeweO9-QqV-", "arxiv_id": "2110.07141", "paper_pdf": "papers/JeweO9-QqV-.pdf", "paper_pdf_sha256": "0c885175c58932e549cf8507cf33d498ef9d5fc110a328c69d561a4f0e54c5ca", "paper_pdf_bytes": 836166, "paper_pdf_source": "openreview", "code_url": "https://github.com/yuehaowang/SoGCN", "code_repository": "yuehaowang/SoGCN", "code_commit": "bd65b2d8667791b79d6174a1dd2ac13b7bd50db5", "code_archive": "repos/JeweO9-QqV-.zip", "code_archive_sha256": "a2d39793aa704a24926dfa0d53afe29ea46958d99f611f78d32c705025e12b77", "code_archive_bytes": 892049, "code_file_count": 69, "code_extensions": {".py": 50, ".sh": 11, ".ipynb": 8}, "github_disk_usage_kb": 945, "github_languages": {"Jupyter Notebook": 1529440, "Python": 253660, "Shell": 10230}, "github_archived": false, "github_pushed_at": "2021-12-18T06:36:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sogcn-second-order-graph-convolutional-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BJe-unNYPr", "year": 2020, "status": "rejected", "title": "Accelerated Information Gradient flow", "authors": ["Yifei Wang", "Wuchen Li"], "authorids": ["zackwang24@pku.edu.cn", "wcli@math.ucla.edu"], "authors_source": "OpenReview API", "abstract": "We present a systematic framework for the Nesterov's accelerated gradient flows in the spaces of probabilities embedded with information metrics. Here two metrics are considered, including both the Fisher-Rao metric and the Wasserstein-$2$ metric. For the Wasserstein-$2$ metric case, we prove the convergence properties of the accelerated gradient flows, and introduce their formulations in Gaussian families. Furthermore, we propose a practical discrete-time algorithm in particle implementations with an adaptive restart technique.  We formulate a novel bandwidth selection method, which learns the Wasserstein-$2$ gradient direction from Brownian-motion samples. Experimental results including Bayesian inference show the strength of the current method compared with the state-of-the-art.", "decision": "Reject", "meta_review": null, "num_reviews": 2, "reviews": [{"id": "BklKUV1fcB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper31/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I acknowledge reading the rebuttal of the authors. Thank you for your clarifications and explanation. My point was this paper would make a good submission to ICLR if it was better motivated presented and explained to a wider audience. Unfortunately in its current form it can only reach a limited audience.\n\n####\nSummary of the paper: \n\nThe paper proposes accelerated information flows under Wasserstein or Fisher rao metric . i.e a method for solving minimization of probability functional , where the probability space is either endowed with the Wasserstein or the Fisher distance.\n\nGradient descent methods in euclidian spaces can be accelerated using a type of momentum , this paper extends this to gradient flows, using similar formalism of Hamiltonian that appeals to the dynamic of a the particles and the velocity (or momentum, $H(x,p)=\\frac{1}{2}||p||^2+ \\mathcal{E}(x)$).  Writing down the lagrangian one obtains two co-evolving PDE one of the dynamic of the density and one for the potential . The PDEs are specified for both Fisher Rao, and Wasserstein distance. The hamiltonian for example for the Wasserstein distance $H(\\rho_t,\\Phi_t)=\\frac{1}{2}\\int||\\nabla_x \\Phi_t||^2+ \\mathcal{E}(\\rho_t)$.  and PDEs amounts a continuity equation for the density  evolving with drift $\\nabla \\Phi_t$, the evolution of the momentum $\\Phi_t$ is also given by a PDE. \n\nProposition 2 of the paper gives the particles differential equation corresponding to the system of PDEs. For the energy being the KL divergence an explicit expression is given , this expression remains difficult in practice since it needs the knowledge of the density $\\rho_t$. Authors propose in the application section to use gaussian approximation , or using a kernel density estimators. The Bandwidth of the kernel is then choosen using a heuristic proposed in the paper.  \n\nThe paper then focuses on deriving expression for flows when the densities are centered gaussians, and this amounts to an ODE on the covariance , the ODE is discretized in Appendix D.2 to lead to  computational method. Then convergence of the flow is analyzed for the wasserstein accelerated flows, under \"\\beta- convextiy\" in the wasserstein sense of the functional. \n\nSome experiments of the particle based method are shown on synthetic experiments and in bayesian logistic regression. \n\nReview: \n\nContribution/ Clarity:\n\nThe main contribution of the paper is in deriving the accelerated gradient flow for the wasserstein distance this was also addressed in a recent paper [Taghevia and Mettha 2019]. \n\nThe technical contribution is interesting but given that this field of flows in probability space is still not very well spread in the ML community, I wonder if ICLR is the best fit for this type of work.  I support good theoretical work, but I think the authors could have done a better job in exposing the ideas how they extend form euclidean space, to manifolds, to probability spaces gradually. Simple derivations of euclidean space Hamiltonian will help the reader that is not exposed to such literature. I think the paper will benefit from a less technical writing in introducing the ideas coming from euclidean space and in conveying the intuitions. \n\nComments: \n\n- In the proof of Proposition 2 you give the expression of evolution of $dV_t$ by conservation of the momentum. Could you please elaborate more how you obtain this expression, and where you proved the conservation of momentum?\n\n- In term of damping if ones uses the Wasserstein Fisher Rao flows  , one obtain also accelerartion , maybe you can comment on that ? since you analyze both flows , would be interesting to discuss the relation to Global convergence of neuron birth-death dynamics, that shows that an acceleration is obtained via WFR flows, since it will introduce a damping as well. \n\n- since MCMC and BM method lead to similar result what is the advantage of the wasserstein accelerated flow? one could also implement also an accelerated langevin dynamic \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "I acknowledge reading the rebuttal of the authors. Thank you for your clarifications and explanation. My point was this paper would make a good submission to ICLR if it was better motivated presented and explained to a wider audience. Unfortunately in its current form it can only reach a limited audience.\n\n####\nSummary of the paper: \n\nThe paper proposes accelerated information flows under Wasserstein or Fisher rao metric . i.e a method for solving minimization of probability functional , where the probability space is either endowed with the Wasserstein or the Fisher distance.\n\nGradient descent methods in euclidian spaces can be accelerated using a type of momentum , this paper extends this to gradient flows, using similar formalism of Hamiltonian that appeals to the dynamic of a the particles and the velocity (or momentum, $H(x,p)=\\frac{1}{2}||p||^2+ \\mathcal{E}(x)$).  Writing down the lagrangian one obtains two co-evolving PDE one of the dynamic of the density and one for the potential . The PDEs are specified for both Fisher Rao, and Wasserstein distance. The hamiltonian for example for the Wasserstein distance $H(\\rho_t,\\Phi_t)=\\frac{1}{2}\\int||\\nabla_x \\Phi_t||^2+ \\mathcal{E}(\\rho_t)$.  and PDEs amounts a continuity equation for the density  evolving with drift $\\nabla \\Phi_t$, the evolution of the momentum $\\Phi_t$ is also given by a PDE. \n\nProposition 2 of the paper gives the particles differential equation corresponding to the system of PDEs. For the energy being the KL divergence an explicit expression is given , this expression remains difficult in practice since it needs the knowledge of the density $\\rho_t$. Authors propose in the application section to use gaussian approximation , or using a kernel density estimators. The Bandwidth of the kernel is then choosen using a heuristic proposed in the paper.  \n\nThe paper then focuses on deriving expression for flows when the densities are centered gaussians, and this amounts to an ODE on the covariance , the ODE is discretized in Appendix D.2 to lead to  computational method. Then convergence of the flow is analyzed for the wasserstein accelerated flows, under \"\\beta- convextiy\" in the wasserstein sense of the functional. \n\nSome experiments of the particle based method are shown on synthetic experiments and in bayesian logistic regression. \n\nReview: \n\nContribution/ Clarity:\n\nThe main contribution of the paper is in deriving the accelerated gradient flow for the wasserstein distance this was also addressed in a recent paper [Taghevia and Mettha 2019]. \n\nThe technical contribution is interesting but given that this field of flows in probability space is still not very well spread in the ML community, I wonder if ICLR is the best fit for this type of work.  I support good theoretical work, but I think the authors could have done a better job in exposing the ideas how they extend form euclidean space, to manifolds, to probability spaces gradually. Simple derivations of euclidean space Hamiltonian will help the reader that is not exposed to such literature. I think the paper will benefit from a less technical writing in introducing the ideas coming from euclidean space and in conveying the intuitions. \n\nComments: \n\n- In the proof of Proposition 2 you give the expression of evolution of $dV_t$ by conservation of the momentum. Could you please elaborate more how you obtain this expression, and where you proved the conservation of momentum?\n\n- In term of damping if ones uses the Wasserstein Fisher Rao flows  , one obtain also accelerartion , maybe you can comment on that ? since you analyze both flows , would be interesting to discuss the relation to Global convergence of neuron birth-death dynamics, that shows that an acceleration is obtained via WFR flows, since it will introduce a damping as well. \n\n- since MCMC and BM method lead to similar result what is the advantage of the wasserstein accelerated flow? one could also implement also an accelerated langevin dynamic \n\n\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572103248700}, {"id": "B1eJdibycS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper31/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper attempts to develop a counterpart of the well-known Nesterov accelerated gradient method for gradient flows on the space of probability measures equipped with an information metric. This is an important problem which is useful for optimization on probability spaces. The accelerated gradient flow is developed by leveraging a damping Hamiltonian flow. The paper focuses mainly on the case with the Wasserstein metric and provides a convergence analysis. Practical considerations such as discretizing the accelerated flow and bandwidth selection are developed for the use of the method in practical problems.\n\nAlthough the paper has some important merit, I find the paper extremely hard to follow, partly because of its writing style. There is not enough motivation and explanation for the ideas presented. Some discussions and sentences either do not make much sense to me or read badly. For example, this sentence \"For the Wasserstein gradient, many classical methods such as Markov Chain Monte Carlo .... are based on this framework...\" doesn't make sense, as the development of MCMC is never based on Wasserstein gradient. Or the sentence right before it \"For the Fisher-Rao gradient, classical results including Adam ... and K-FAC .. demonstrate its effectiveness in ...\": it's not clear what the authors are trying to say here. Adam is not relevant to the Fisher-Rao natural gradient while K-FAC is just an approximation method and isn't a good reference for demonstrating the effectiveness of the natural gradient. Also, there are many English typos and grammar errors.  \n\nI didn't read the proof carefully due to the time constraint, so I cannot judge on the theoretical part of the paper. The numerical experiment is quite limited as it considers very simple problems (a toy example, a single Gaussian distribution and a logistic regression problem). As such, I think there isn't enough evidence to judge the usefulness of the proposed method in practice. \n\nHaving said that, I believe this paper can be an important contribution if the authors invest more time on refining its presentation and if more thorough experimental studies are conducted.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper attempts to develop a counterpart of the well-known Nesterov accelerated gradient method for gradient flows on the space of probability measures equipped with an information metric. This is an important problem which is useful for optimization on probability spaces. The accelerated gradient flow is developed by leveraging a damping Hamiltonian flow. The paper focuses mainly on the case with the Wasserstein metric and provides a convergence analysis. Practical considerations such as discretizing the accelerated flow and bandwidth selection are developed for the use of the method in practical problems.\n\nAlthough the paper has some important merit, I find the paper extremely hard to follow, partly because of its writing style. There is not enough motivation and explanation for the ideas presented. Some discussions and sentences either do not make much sense to me or read badly. For example, this sentence \"For the Wasserstein gradient, many classical methods such as Markov Chain Monte Carlo .... are based on this framework...\" doesn't make sense, as the development of MCMC is never based on Wasserstein gradient. Or the sentence right before it \"For the Fisher-Rao gradient, classical results including Adam ... and K-FAC .. demonstrate its effectiveness in ...\": it's not clear what the authors are trying to say here. Adam is not relevant to the Fisher-Rao natural gradient while K-FAC is just an approximation method and isn't a good reference for demonstrating the effectiveness of the natural gradient. Also, there are many English typos and grammar errors.  \n\nI didn't read the proof carefully due to the time constraint, so I cannot judge on the theoretical part of the paper. The numerical experiment is quite limited as it considers very simple problems (a toy example, a single Gaussian distribution and a logistic regression problem). As such, I think there isn't enough evidence to judge the usefulness of the proposed method in practice. \n\nHaving said that, I believe this paper can be an important contribution if the authors invest more time on refining its presentation and if more thorough experimental studies are conducted.\n"}, "tcdate": 1571916647332}], "openreview_url": "https://openreview.net/forum?id=BJe-unNYPr", "arxiv_id": "1909.02102", "paper_pdf": "papers/BJe-unNYPr.pdf", "paper_pdf_sha256": "26deafeaf2524b327f94c0fe9c2cb46752075d0bafe28e1a63c9aeaf9efd9a6c", "paper_pdf_bytes": 1838086, "paper_pdf_source": "openreview", "code_url": "https://github.com/YiifeiWang/Accelerated-Information-Gradient-flow", "code_repository": "YiifeiWang/Accelerated-Information-Gradient-flow", "code_commit": "fb5e485066fd2658cb93787da44203b18df2f193", "code_archive": "repos/BJe-unNYPr.zip", "code_archive_sha256": "20a32c2d56865f4bc1c2977e026e81eaca3def10de62926f34e6f3d827026ea6", "code_archive_bytes": 330285, "code_file_count": 42, "code_extensions": {".m": 23, ".py": 9, ".sh": 7, ".ipynb": 3}, "github_disk_usage_kb": 2333, "github_languages": {"MATLAB": 67109, "Python": 60402, "Jupyter Notebook": 6000, "Shell": 3447, "M": 2051}, "github_archived": false, "github_pushed_at": "2020-06-02T02:51:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerated-information-gradient-flow"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GsLvRPAfoU", "year": 2026, "status": "rejected", "title": "Improving Fine-Grained Control via Aggregation of Multiple Diffusion Models", "authors": ["Conghan Yue", "Zhengwei Peng", "Shiyan Du", "Zhi Ji", "Chuangjian Cai", "Le Wan", "Dongyu Zhang"], "authorids": ["~Conghan_Yue2", "~Zhengwei_Peng1", "~Shiyan_Du1", "~Zhi_Ji1", "~Chuangjian_Cai2", "~Le_Wan1", "~Dongyu_Zhang1"], "authors_source": "OpenReview API", "abstract": "While many diffusion models perform well when controlling particular aspects such as style, character, and interaction, they struggle with fine-grained control due to dataset limitations and intricate model architecture design. This paper introduces a novel training-free algorithm for fine-grained generation, called Aggregation of Multiple Diffusion Models (AMDM). The algorithm integrates features in the latent data space from multiple diffusion models within the same ecosystem into a specified model, thereby activating particular features and enabling fine-grained control. Experimental results demonstrate that AMDM significantly improves fine-grained control without training, validating its effectiveness. Additionally, it reveals that diffusion models initially focus on features such as position, attributes, and style, with later stages improving generation quality and consistency. AMDM offers a new perspective for tackling the challenges of fine-grained conditional generation in diffusion models. Specifically, it allows us to fully utilize existing or develop new conditional diffusion models that control specific aspects, and then aggregate them using the AMDM algorithm. This eliminates the need for constructing complex datasets, designing intricate model architectures, and incurring high training costs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "1pcJhOLeyP", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2093/Reviewer_MHdz"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper tackles the problem of fine-grained control in diffusion-based generative models that fail at nuanced, multi-conditional control ,e.g., spatial arrangement and style preservation. The authors propose a training-free algorithm called AMDM that aggregates latent representations from multiple diffusion models.  During inference, AMDM merge latent variables geometrically using spherical aggregation and a deviation optimization step. The authors conducted several empirical experiments to show improvements in attribute, style, and interaction controllability with AMDM.", "review_text": "The paper tackles the problem of fine-grained control in diffusion-based generative models that fail at nuanced, multi-conditional control ,e.g., spatial arrangement and style preservation. The authors propose a training-free algorithm called AMDM that aggregates latent representations from multiple diffusion models.  During inference, AMDM merge latent variables geometrically using spherical aggregation and a deviation optimization step. The authors conducted several empirical experiments to show improvements in attribute, style, and interaction controllability with AMDM.", "strengths": "- The perspective of leveraging existing diffusion models to address fine-grained, controllable generation in a training-free manner is an interesting direction.\n- The paper is well-written and easy to follow; empirical performance and results demonstrate the promise of the proposed method.", "weaknesses": "- \"Diffusion ecosystem\" is mentioned across the paper as a prerequisite of AMDM but somewhat loosely defined, for example, how to quantitatively verify whether any two models belong to the same ecosystem is unclear. \n- Evaluation scope is limited to certain derivatives of Stable Diffusion models. Throughout the experiments, the authors only picked a few classic SD1.4/1.5 and SDXL models, but did not examine whether the findings generalize to more recent model architectures based on DiT instead of U-Net, such as as SD3 or Flux. \n- Comparing to other model composition methods in the literature, the proposed method only applies to high-dimensional Gaussian and fails to general distributions (as shown in Table 7)", "questions": "- Besides the comparisions between SD1.5 and SDXL, how sensitive is AMDM to broader model heterogeneity (schedulers or attention designs)? For example, you may conduct some experiments on models based on DiT to show the generality of your method?\n- The weighting factor $w$ in Slerp is treated as hyperparameter, but lacks a principled selection method. Empirical experiments are conducted based on two/three model aggregation where $w$ is determined with ablations. Can you discuss more on how to systematically select these parameters, and especially with a set of models (beyond three typical models you selected), how your propose method work empirically?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tackles the problem of fine-grained control in diffusion-based generative models that fail at nuanced, multi-conditional control ,e.g., spatial arrangement and style preservation. The authors propose a training-free algorithm called AMDM that aggregates latent representations from multiple diffusion models.  During inference, AMDM merge latent variables geometrically using spherical aggregation and a deviation optimization step. The authors conducted several empirical experiments to show improvements in attribute, style, and interaction controllability with AMDM.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The perspective of leveraging existing diffusion models to address fine-grained, controllable generation in a training-free manner is an interesting direction.\n- The paper is well-written and easy to follow; empirical performance and results demonstrate the promise of the proposed method.", "weaknesses": "- \"Diffusion ecosystem\" is mentioned across the paper as a prerequisite of AMDM but somewhat loosely defined, for example, how to quantitatively verify whether any two models belong to the same ecosystem is unclear. \n- Evaluation scope is limited to certain derivatives of Stable Diffusion models. Throughout the experiments, the authors only picked a few classic SD1.4/1.5 and SDXL models, but did not examine whether the findings generalize to more recent model architectures based on DiT instead of U-Net, such as as SD3 or Flux. \n- Comparing to other model composition methods in the literature, the proposed method only applies to high-dimensional Gaussian and fails to general distributions (as shown in Table 7)", "questions": "- Besides the comparisions between SD1.5 and SDXL, how sensitive is AMDM to broader model heterogeneity (schedulers or attention designs)? For example, you may conduct some experiments on models based on DiT to show the generality of your method?\n- The weighting factor $w$ in Slerp is treated as hyperparameter, but lacks a principled selection method. Empirical experiments are conducted based on two/three model aggregation where $w$ is determined with ablations. Can you discuss more on how to systematically select these parameters, and especially with a set of models (beyond three typical models you selected), how your propose method work empirically?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761979246852}, {"id": "Ox6ANmnlD7", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2093/Reviewer_E8Pz"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces the Aggregation of Multiple Diffusion Models (AMDM) algorithm, a training-free method for improving fine-grained conditional control in image generation. AMDM works by aggregating features from multiple diffusion models within the same ecosystem to integrate their respective strengths. The key components are 1) Spherical aggregation 2)Deviation optimization. The authors demonstrate that AMDM significantly improves fine-grained control capabilities by integrating different models' strengths.", "review_text": "This paper introduces the Aggregation of Multiple Diffusion Models (AMDM) algorithm, a training-free method for improving fine-grained conditional control in image generation. AMDM works by aggregating features from multiple diffusion models within the same ecosystem to integrate their respective strengths. The key components are 1) Spherical aggregation 2)Deviation optimization. The authors demonstrate that AMDM significantly improves fine-grained control capabilities by integrating different models' strengths.", "strengths": "- The authors identify a genuine limitation in current models that they excel in specific aspects but struggle with others, and provide a solution without requiring retraining.\n- The approach is backed by mathematical analysis of why aggregation works for models in the same diffusion ecosystem.\n- Extensive experiments demonstrate clear improvements in both qualitative results and quantitative metrics.\n- Unlike many compositional methods that introduce significant computational overhead, AMDM has minimal additional cost.", "weaknesses": "- While the authors provide theoretical justification, the assumptions (functional proximity, conditional proximity) are somewhat heuristic and don't offer global guarantees.\n- AMDM only works for models within the same diffusion ecosystem, limiting its general applicability.\n- While some comparisons with compositional methods are provided, a more comprehensive comparison with other training-free approaches would strengthen the paper.\n- When aggregating models, there might be unintended interactions between different aspects that aren't fully explored.\n- The optimization step size is selected empirically, and the authors acknowledge this as a limitation requiring future work.", "questions": "- How sensitive is AMDM to the choice of models within an ecosystem? Are there certain model combinations that work particularly well or poorly?\n- (minor) What happens to the result if the positional information is not given? Would it still struggle to generate as specified from the text prompt or could this fine-grained control be coming from additional information, poisition.\n- (minor) This is more like a question and just curiosity on my side. Could this fine-grained problem only exist within the open-source models? In other words, would the problem still exist in the models such as Sora?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Aggregation of Multiple Diffusion Models (AMDM) algorithm, a training-free method for improving fine-grained conditional control in image generation. AMDM works by aggregating features from multiple diffusion models within the same ecosystem to integrate their respective strengths. The key components are 1) Spherical aggregation 2)Deviation optimization. The authors demonstrate that AMDM significantly improves fine-grained control capabilities by integrating different models' strengths.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The authors identify a genuine limitation in current models that they excel in specific aspects but struggle with others, and provide a solution without requiring retraining.\n- The approach is backed by mathematical analysis of why aggregation works for models in the same diffusion ecosystem.\n- Extensive experiments demonstrate clear improvements in both qualitative results and quantitative metrics.\n- Unlike many compositional methods that introduce significant computational overhead, AMDM has minimal additional cost.", "weaknesses": "- While the authors provide theoretical justification, the assumptions (functional proximity, conditional proximity) are somewhat heuristic and don't offer global guarantees.\n- AMDM only works for models within the same diffusion ecosystem, limiting its general applicability.\n- While some comparisons with compositional methods are provided, a more comprehensive comparison with other training-free approaches would strengthen the paper.\n- When aggregating models, there might be unintended interactions between different aspects that aren't fully explored.\n- The optimization step size is selected empirically, and the authors acknowledge this as a limitation requiring future work.", "questions": "- How sensitive is AMDM to the choice of models within an ecosystem? Are there certain model combinations that work particularly well or poorly?\n- (minor) What happens to the result if the positional information is not given? Would it still struggle to generate as specified from the text prompt or could this fine-grained control be coming from additional information, poisition.\n- (minor) This is more like a question and just curiosity on my side. Could this fine-grained problem only exist within the open-source models? In other words, would the problem still exist in the models such as Sora?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761967754594}, {"id": "ZB1lG1xJwp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2093/Reviewer_WAAK"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper tackles the problem of fine-grained conditional control in diffusion models. The proposed method, AMDM, aggregates multiple diffusion model scores via spherical interpolation, and employs a deviation optimization step to stabilize the aggregated score function. The framework enables users to combine the strengths of multiple fine-tuned diffusion models without requiring complex multi-capability datasets or multi-stage finetuning. Experiments on several conditional generation tasks demonstrate that AMDM can effectively integrate diverse capabilities from different models at inference time.", "review_text": "This paper tackles the problem of fine-grained conditional control in diffusion models. The proposed method, AMDM, aggregates multiple diffusion model scores via spherical interpolation, and employs a deviation optimization step to stabilize the aggregated score function. The framework enables users to combine the strengths of multiple fine-tuned diffusion models without requiring complex multi-capability datasets or multi-stage finetuning. Experiments on several conditional generation tasks demonstrate that AMDM can effectively integrate diverse capabilities from different models at inference time.", "strengths": "- The method is grounded in solid theoretical analysis, particularly regarding the confidence and reliability of aggregated diffusion scores.\n- The empirical evaluation is thorough, and the comparisons clearly demonstrate how AMDM improves over baselines in multiple tasks.\n- The motivation is practical and relevant: enabling reuse and integration of existing specialized diffusion models without retraining.", "weaknesses": "- Although the theoretical analysis is detailed, the paper may benefit from a simple controlled or toy example to help intuitively illustrate the effect of score aggregation and deviation optimization.\n- The evaluation primarily focuses on three types of models (MIGC, InteractDiffusion, and IP-Adapter). While the results are encouraging, a broader range of conditional diffusion methods or application settings would better support the generality of the method.\n- One of the key claims — that diffusion models initially prioritize feature generation before later refining image quality and consistency — is mainly supported by a single ablation (Table 3). Additional analysis or diagnostic visualization would help strengthen this conclusion.", "questions": "1. Regarding spherical interpolation and Equation (6):\n\n    Could the authors clarify the geometric intuition or assumption behind applying spherical aggregation to score vectors? Is the assumption that these score fields locally reside on a shared spherical manifold, or is the spherical constraint introduced primarily for normalization and stability?\n    \n2. Computation overhead:\n\n    What is the computational cost of AMDM compared to running a single diffusion model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the problem of fine-grained conditional control in diffusion models. The proposed method, AMDM, aggregates multiple diffusion model scores via spherical interpolation, and employs a deviation optimization step to stabilize the aggregated score function. The framework enables users to combine the strengths of multiple fine-tuned diffusion models without requiring complex multi-capability datasets or multi-stage finetuning. Experiments on several conditional generation tasks demonstrate that AMDM can effectively integrate diverse capabilities from different models at inference time.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The method is grounded in solid theoretical analysis, particularly regarding the confidence and reliability of aggregated diffusion scores.\n- The empirical evaluation is thorough, and the comparisons clearly demonstrate how AMDM improves over baselines in multiple tasks.\n- The motivation is practical and relevant: enabling reuse and integration of existing specialized diffusion models without retraining.", "weaknesses": "- Although the theoretical analysis is detailed, the paper may benefit from a simple controlled or toy example to help intuitively illustrate the effect of score aggregation and deviation optimization.\n- The evaluation primarily focuses on three types of models (MIGC, InteractDiffusion, and IP-Adapter). While the results are encouraging, a broader range of conditional diffusion methods or application settings would better support the generality of the method.\n- One of the key claims — that diffusion models initially prioritize feature generation before later refining image quality and consistency — is mainly supported by a single ablation (Table 3). Additional analysis or diagnostic visualization would help strengthen this conclusion.", "questions": "1. Regarding spherical interpolation and Equation (6):\n\n    Could the authors clarify the geometric intuition or assumption behind applying spherical aggregation to score vectors? Is the assumption that these score fields locally reside on a shared spherical manifold, or is the spherical constraint introduced primarily for normalization and stability?\n    \n2. Computation overhead:\n\n    What is the computational cost of AMDM compared to running a single diffusion model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761680537406}, {"id": "VXAXRVTYnW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission2093/Reviewer_qUaq"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a two-step process for combining multiple diffusion models. In the first step, the authors use Spherical Linear Interpolation (SLERP) to combine the models, and in the second step, they perform a descent procedure to move the combined output closer to the mean ( increasing probability of the sample ).", "review_text": "This paper proposes a two-step process for combining multiple diffusion models. In the first step, the authors use Spherical Linear Interpolation (SLERP) to combine the models, and in the second step, they perform a descent procedure to move the combined output closer to the mean ( increasing probability of the sample ).", "strengths": "Propose novel method of Spherical Linear Interpolation for combining multiple diffusion models.", "weaknesses": "### Need Proofs for Claims / Additional Assumptions\n\n**Lines 130–146:**  \nThe assumption \\(p_{\\theta_1}, p_{\\theta_2} \\subset \\mathcal{M}_0 \\) implies that \\( p_{\\theta_1} \\) and \\( p_{\\theta_2} \\) are trained on the same data, or equivalently, that they are generative models of the same underlying data distribution.  \nHowever, this assumption holds **only if** both models are trained on data drawn from the same distribution. For example, if model A is trained on real images while model B is trained on paintings, the underlying data distributions differ, and hence the assumption \\( p_{\\theta_1}, p_{\\theta_2} \\subset \\mathcal{M}_0 \\) becomes invalid.\n\n---\n\nCan you provide a proof supporting the claim that additive or modified architectures only enhance control features?  \nThis assertion appears **baseless** and **lacks supporting theoretical or empirical justification**. Consequently, the argument for a “same diffusion-model ecosystem” becomes **ill-defined and conceptually inconsistent**.\n\n---\n\n**Line 164:**  \nWhat do you mean by *“same task”*?\n\n---\n\n### Clarifications Needed\n\nThe motivation behind using *spherical interpolation* and *deviation optimization* is unclear.  \nIt is also not evident how the parameters \\( q_1, w_2, n_\\theta \\) are chosen.\n\nThere exist several approaches to aggregate diffusion models, many of which are derived from a probabilistic perspective under certain assumptions of independence (broadly referred to as *compositionality*).\n\n### Comparison to Relevant Works\n\n1. **Reduce, Reuse, Recycle:** [arXiv:2302.11552](https://arxiv.org/pdf/2302.11552)  \n2. **Compositionally:** [arXiv:2206.01714](https://arxiv.org/pdf/2206.01714)  \n3. **Conditional Independence Assumed:** [arXiv:2503.01145](https://arxiv.org/abs/2503.01145) — enforcing conditional independence  \n4. **Without Conditional Independence (Controllability):** [arXiv:2302.14368](https://arxiv.org/pdf/2302.14368)  \n5. [arXiv:2505.13213](https://arxiv.org/pdf/2505.13213)  \n6. **Algebraic Perspective:** [arXiv:2502.04549](https://arxiv.org/abs/2502.04549)\n\nAll these works rely on certain **underlying assumptions**, particularly regarding independence.  \nWhen such independence does not hold, an additional **weighting or hyperparameter term** is typically introduced to account for the violation.\n\n---\n\n### Example: Independence vs. Dependence\n\nIf conditional independence is assumed:\n\n\\[\np_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + p_{\\theta}(x \\mid y_2)\n  - p_{\\theta}(x \\mid \\varnothing)\n\\]\n\nIf independence **is not** assumed:\n\n\\[\np_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + w_1\\, p_{\\theta}(x \\mid y_2)\n  - w_1\\, p_{\\theta}(x \\mid \\varnothing),\n\\]\n\nwhere \\( w_1 \\) controls the degree of independence between \\( y_1 \\) and \\( y_2 \\).\n\n---\n\n### Connecting to the Current Work\n\nIn line with the assumptions of the current work, suppose \\( z_t \\) is drawn from the distribution \\( p(z_t) = p(z_t \\mid \\varnothing) \\), and assume that \\( z_t \\subset \\mathcal{M}_t \\) for all \\( t \\in [0, T] \\).  \nThe objective can then be viewed as **linearly combining distributions** to maximize the probability of lying on the data manifold:\n\n\\[\n\\max_{a,b,c}\\; p(a z^{(1)}_{t-1} + b z^{(2)}_{t-1} + c \\mid \\varnothing, z_t),\n\\]\n\nsuch that the combination lies on the manifold of \\( p(z_t) \\).\n\nSince the distribution is Gaussian, the maximum corresponds to its **mean**, reducing the sampling process to:\n\n\\[\na\\, p_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + b\\, p_{\\theta}(x \\mid y_2)\n  + (1 - a - b)\\, p_{\\theta}(x \\mid \\varnothing).\n\\]\n\n---\n\n### Interpretation\n\nMethods such as **ADAM** or **spherical interpolation** represent alternative interpolation strategies—implicitly introducing weighting factors rather than explicitly modeling independence.  \nIn the current work, this combination is obtained via **spherical interpolation**, without direct access to \\( p(z_t \\mid \\varnothing) \\) (as in the classifier-free guidance formulation).  \nHowever, it remains unclear **why any \\( s < T \\)** should necessarily lie on a sphere.\n\n---\n\n### Broader Concern\n\nFor methods that do not assume any structural property of the data distribution, the introduction of a **hyperparameter** becomes unavoidable.  \nIt is also unclear **how these hyperparameters should be selected**.\n\n---\n\n### Results and Evaluation\n\nThe reported results appear potentially misleading.  \nAs the authors themselves claim, combining any two methods tends to improve performance over both individual methods.  \nSince the hyperparameters are tuned by the authors, the worst possible case corresponds to \\( w_2 = 0 \\).  \nTherefore, unless a **principled approach** for hyperparameter selection or a **comparison with existing aggregation methods** is provided, the results section offers **limited insight**.", "questions": "Addressing all the weaknesses will answer all my questions", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a two-step process for combining multiple diffusion models. In the first step, the authors use Spherical Linear Interpolation (SLERP) to combine the models, and in the second step, they perform a descent procedure to move the combined output closer to the mean ( increasing probability of the sample ).", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Propose novel method of Spherical Linear Interpolation for combining multiple diffusion models.", "weaknesses": "### Need Proofs for Claims / Additional Assumptions\n\n**Lines 130–146:**  \nThe assumption \\(p_{\\theta_1}, p_{\\theta_2} \\subset \\mathcal{M}_0 \\) implies that \\( p_{\\theta_1} \\) and \\( p_{\\theta_2} \\) are trained on the same data, or equivalently, that they are generative models of the same underlying data distribution.  \nHowever, this assumption holds **only if** both models are trained on data drawn from the same distribution. For example, if model A is trained on real images while model B is trained on paintings, the underlying data distributions differ, and hence the assumption \\( p_{\\theta_1}, p_{\\theta_2} \\subset \\mathcal{M}_0 \\) becomes invalid.\n\n---\n\nCan you provide a proof supporting the claim that additive or modified architectures only enhance control features?  \nThis assertion appears **baseless** and **lacks supporting theoretical or empirical justification**. Consequently, the argument for a “same diffusion-model ecosystem” becomes **ill-defined and conceptually inconsistent**.\n\n---\n\n**Line 164:**  \nWhat do you mean by *“same task”*?\n\n---\n\n### Clarifications Needed\n\nThe motivation behind using *spherical interpolation* and *deviation optimization* is unclear.  \nIt is also not evident how the parameters \\( q_1, w_2, n_\\theta \\) are chosen.\n\nThere exist several approaches to aggregate diffusion models, many of which are derived from a probabilistic perspective under certain assumptions of independence (broadly referred to as *compositionality*).\n\n### Comparison to Relevant Works\n\n1. **Reduce, Reuse, Recycle:** [arXiv:2302.11552](https://arxiv.org/pdf/2302.11552)  \n2. **Compositionally:** [arXiv:2206.01714](https://arxiv.org/pdf/2206.01714)  \n3. **Conditional Independence Assumed:** [arXiv:2503.01145](https://arxiv.org/abs/2503.01145) — enforcing conditional independence  \n4. **Without Conditional Independence (Controllability):** [arXiv:2302.14368](https://arxiv.org/pdf/2302.14368)  \n5. [arXiv:2505.13213](https://arxiv.org/pdf/2505.13213)  \n6. **Algebraic Perspective:** [arXiv:2502.04549](https://arxiv.org/abs/2502.04549)\n\nAll these works rely on certain **underlying assumptions**, particularly regarding independence.  \nWhen such independence does not hold, an additional **weighting or hyperparameter term** is typically introduced to account for the violation.\n\n---\n\n### Example: Independence vs. Dependence\n\nIf conditional independence is assumed:\n\n\\[\np_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + p_{\\theta}(x \\mid y_2)\n  - p_{\\theta}(x \\mid \\varnothing)\n\\]\n\nIf independence **is not** assumed:\n\n\\[\np_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + w_1\\, p_{\\theta}(x \\mid y_2)\n  - w_1\\, p_{\\theta}(x \\mid \\varnothing),\n\\]\n\nwhere \\( w_1 \\) controls the degree of independence between \\( y_1 \\) and \\( y_2 \\).\n\n---\n\n### Connecting to the Current Work\n\nIn line with the assumptions of the current work, suppose \\( z_t \\) is drawn from the distribution \\( p(z_t) = p(z_t \\mid \\varnothing) \\), and assume that \\( z_t \\subset \\mathcal{M}_t \\) for all \\( t \\in [0, T] \\).  \nThe objective can then be viewed as **linearly combining distributions** to maximize the probability of lying on the data manifold:\n\n\\[\n\\max_{a,b,c}\\; p(a z^{(1)}_{t-1} + b z^{(2)}_{t-1} + c \\mid \\varnothing, z_t),\n\\]\n\nsuch that the combination lies on the manifold of \\( p(z_t) \\).\n\nSince the distribution is Gaussian, the maximum corresponds to its **mean**, reducing the sampling process to:\n\n\\[\na\\, p_{\\theta}(x \\mid y_1, y_2)\n  = p_{\\theta}(x \\mid y_1)\n  + b\\, p_{\\theta}(x \\mid y_2)\n  + (1 - a - b)\\, p_{\\theta}(x \\mid \\varnothing).\n\\]\n\n---\n\n### Interpretation\n\nMethods such as **ADAM** or **spherical interpolation** represent alternative interpolation strategies—implicitly introducing weighting factors rather than explicitly modeling independence.  \nIn the current work, this combination is obtained via **spherical interpolation**, without direct access to \\( p(z_t \\mid \\varnothing) \\) (as in the classifier-free guidance formulation).  \nHowever, it remains unclear **why any \\( s < T \\)** should necessarily lie on a sphere.\n\n---\n\n### Broader Concern\n\nFor methods that do not assume any structural property of the data distribution, the introduction of a **hyperparameter** becomes unavoidable.  \nIt is also unclear **how these hyperparameters should be selected**.\n\n---\n\n### Results and Evaluation\n\nThe reported results appear potentially misleading.  \nAs the authors themselves claim, combining any two methods tends to improve performance over both individual methods.  \nSince the hyperparameters are tuned by the authors, the worst possible case corresponds to \\( w_2 = 0 \\).  \nTherefore, unless a **principled approach** for hyperparameter selection or a **comparison with existing aggregation methods** is provided, the results section offers **limited insight**.", "questions": "Addressing all the weaknesses will answer all my questions", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes", "details_of_ethics_concerns": "### Need proofs for claims / additional assumptions\n\n**Lines 130–146:**  \nThe assumption $p_{\\theta_1}, p_{\\theta_2}$ $\\in$  $M_0$ implies that $p_{\\theta_1}$ and $p_{\\theta_2}$ are trained on the same data, or equivalently, that they are generative models of the same underlying data distribution.  However, this assumption holds **only if** both models are trained on data drawn from the same distribution. For example, if model A is trained on real images while model B is trained on paintings, the underlying data distributions differ, and hence the assumption $p_{\\theta_1}, p_{\\theta_2} \\in \\mathcal{M}_0$ becomes invalid.\n\n---\n\nCan you provide a proof supporting the claim that additive or modified architectures only enhance control features?  \nThis assertion appears **baseless** and **lacks supporting theoretical or empirical justification**. Consequently, the argument for a “same diffusion model” ecosystem becomes **ill-defined and conceptually inconsistent**.\n\n---\n\n**Line 164:**  \nWhat do you mean by *“same task”*?\n\n---\n\n### Clarifications needed\n\nThe motivation behind using *spherical interpolation* and *deviation optimization* is unclear.  \nIt is also not evident how the parameters $q_1$, $w_2$, and $n_\\theta$ are chosen.\n\nThere exist several approaches to aggregate diffusion models, many of which are derived from a probabilistic perspective under certain assumptions of independence (broadly referred to as *compositionality*).\n\n---\n\n### Comparison to relevant works\n\n1. **Reduce, Reuse, Recycle:** [arXiv:2302.11552](https://arxiv.org/pdf/2302.11552)  \n2. **Compositionally:** [arXiv:2206.01714](https://arxiv.org/pdf/2206.01714)  \n3. **Conditional Independence Assumed:** [arXiv:2503.01145](https://arxiv.org/abs/2503.01145) — enforcing conditional independence  \n4. **Without Conditional Independence (Controllability):** [arXiv:2302.14368](https://arxiv.org/pdf/2302.14368)  \n5. [arXiv:2505.13213](https://arxiv.org/pdf/2505.13213)  \n6. **Algebraic Perspective:** [arXiv:2502.04549](https://arxiv.org/abs/2502.04549)\n\nAll these works rely on certain **underlying assumptions**, particularly regarding independence.  \nWhen such independence does not hold, an additional **weighting or hyperparameter term** is typically introduced to account for the violation.\n\nIf conditional independence is assumed:\n\n$$p_{\\theta}(x \\mid y_1, y_2) = p_{\\theta}(x \\mid y_1)+ p_{\\theta}(x \\mid y_2)- p_{\\theta}(x \\mid \\varnothing)$$\n\nIf independence is **not** assumed:\n\n$$p_{\\theta}(x \\mid y_1, y_2)= p_{\\theta}(x \\mid y_1) + w_1\\, p_{\\theta}(x \\mid y_2)- w_1\\, p_{\\theta}(x \\mid \\varnothing)$$\n\nwhere $w_1$ controls the degree of independence between $y_1$ and $y_2$.\n\n---\n\n### Connecting to the current work\n\nIn line with the assumptions of the current work, suppose $z_t$ is drawn from the distribution $p(z_t) = p(z_t \\mid \\varnothing)$, and assume that $z_t \\in \\mathcal{M}_t$ for all $t \\in [0, T]$.  \nThe objective can then be viewed as **linearly combining distributions** to maximize the probability of lying on the data manifold:\n\n$max_{a,b,c}$ $a\\times p(z^{(1)}_{t-1}$ \nsuch that the combination lies on the manifold of $p(z_t)$.\n\nSince the distribution is Gaussian, the maximum corresponds to its **mean**, reducing the sampling process to:\n\n$$ p_{\\theta}(x \\mid y_1, y_2)= a \\times p_{\\theta}(x \\mid y_1)+ b\\times p_{\\theta}(x \\mid y_2)+ (1 - a - b)\\times p_{\\theta}(x \\mid \\varnothing)$$\n\nMethods such as **ADAM** or **spherical interpolation** represent alternative interpolation strategies; implicitly introducing weighting factors rather than explicitly modeling independence.  In the current work, this combination is obtained via **spherical interpolation**, without direct access to $p(z_t \\mid \\varnothing)$ (as in the classifier-free guidance formulation).   However, it remains unclear **why any $s < T$** should necessarily lie on a sphere.\n\nFor methods that do not assume any structural property of the data distribution, the introduction of a **hyperparameter** becomes unavoidable.   It is also unclear **how these hyperparameters should be selected**.\n\n### Results and evaluation\n\nThe reported results appear potentially misleading.  As the authors themselves claim, combining any two methods always improves the performance over both individual methods.  Since the hyperparameters are tuned by the authors, the worst possible case corresponds to $w_2 = 0$.  Therefore, unless a **principled approach** for hyperparameter selection or a **comparison with existing aggregation methods** is provided, the results section offers **limited insight**."}, "tcdate": 1761528959018}], "openreview_url": "https://openreview.net/forum?id=GsLvRPAfoU", "arxiv_id": "2410.01262", "paper_pdf": "papers/GsLvRPAfoU.pdf", "paper_pdf_sha256": "7c4723c82fd9ddd32cda5b405e4f05691cf5ffb1de87b51b9a113720d5dfdf4a", "paper_pdf_bytes": 11034682, "paper_pdf_source": "openreview", "code_url": "https://github.com/Hammour-steak/AMDM", "code_repository": "Hammour-steak/AMDM", "code_commit": "076ec6a6df136f6917acb3e1000c1a23f95b7581", "code_archive": "repos/GsLvRPAfoU.zip", "code_archive_sha256": "db16ff3ef751c050d5f964b5971619e9a999a3d34556b8f79b3aa4b30a3462c5", "code_archive_bytes": 3036572, "code_file_count": 68, "code_extensions": {".py": 58, ".js": 9, ".ipynb": 1}, "github_disk_usage_kb": 2924, "github_languages": {"Python": 594009, "JavaScript": 47637, "HTML": 29571, "CSS": 11526, "Jupyter Notebook": 8885}, "github_archived": false, "github_pushed_at": "2024-12-17T06:37:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/aggregation-of-multi-diffusion-models-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kXYuzGcfTT", "year": 2026, "status": "rejected", "title": "Overclocking LLM Reasoning: Monitoring and Controlling Thinking Path Lengths in LLMs", "authors": ["Roy Eisenstadt", "Itamar Zimerman", "Lior Wolf"], "authorids": ["~Roy_Eisenstadt1", "~Itamar_Zimerman1", "~Lior_Wolf1"], "authors_source": "OpenReview API", "abstract": "Recently, techniques such as explicit structured reasoning have demonstrated strong test-time scaling behavior by enforcing a separation between the model’s internal \"thinking\" process and the final response. A key factor influencing answer quality in this setting is the length of the thinking stage. When the reasoning is too short, the model may fail to capture the complexity of the task. Conversely, when it is too long, the model may overthink, leading to unnecessary computation and degraded performance. This paper explores and exploits the underlying mechanisms by which LLMs understand and regulate the length of their reasoning during explicit thought processes. First, we show that LLMs encode their progress through the reasoning process and introduce an interactive progress bar visualization, which is then used to reveal insights on the model's planning dynamics. Second, we manipulate the internal progress encoding during inference to reduce unnecessary steps and generate a more concise and decisive chain of thoughts. Our empirical results demonstrate that this ``overclocking'' method mitigates overthinking, improves answer accuracy, and reduces inference latency. Our code is attached as supplementary material.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NHrtvOw1in", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7951/Reviewer_v9dA"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This paper proposes \"thinking progress vector\" (TPV), a vector representation of what is claimed the encoding of its \"interval reasoning process\".\nThe paper uses linear models, MLPs, and RNN as tools to monitor the change of this state along the reasoning trajectory of LLMs.\nThe paper claims that intervening this vector can change the behaviour of LLMs and make it reason \"faster\", or what is called \"overclocking\" by the paper.\nEmpirical results on math reasoning tasks shows some evidence about the effect of intervening TPVs.", "review_text": "This paper proposes \"thinking progress vector\" (TPV), a vector representation of what is claimed the encoding of its \"interval reasoning process\".\nThe paper uses linear models, MLPs, and RNN as tools to monitor the change of this state along the reasoning trajectory of LLMs.\nThe paper claims that intervening this vector can change the behaviour of LLMs and make it reason \"faster\", or what is called \"overclocking\" by the paper.\nEmpirical results on math reasoning tasks shows some evidence about the effect of intervening TPVs.", "strengths": "- The paper has some intuitive figures showing the discovered \"thinking progress vector\".", "weaknesses": "- The idea of estimating the progress of thinking as proposed in this paper does not make sense. It is beyond me why an LLM can maintain a state representing its reasoning progress if it does not know the correct answer to the problem even when the reasoning is just started. If an estimation of this really existed, then we would have asked this LLM to solve the halting problem, which has been mathematically proven to be impossible.\n\n- The TPV might represent some characteristic tokens marking the steps in solutions to math problems. The paper unfortunately does not provide enough study along this direction. In lines 261-263, even the paper itself writes \"a crucial question that arises is whether TPVs reflect a fundamental mechanism that the model uses to track its reasoning progress, or if they are merely residual artifacts that correlate with progress but do not play a causal role in the computation.\" this possibility is never ruled out in the rest of the paper.\n\n- The concept of reasoning progress itself is poorly defined for real-world problems. What does reasoning progress mean for writing some code for a specific task? Does reasoning ends when the code is completed? Or When LLM finishes running it in with a python interpreter? What does reasoning progress mean for the travel sales person (TSP) problem, when it is impossible to solve in polynomial time?", "questions": "I do not have questions for this paper. I have not checked all the experiments of the paper, so I will put a 2 for my confidence. \n\nI request that the paper be framed more accurately and align with the fundamentals of computer science in terms of the message it tries to convey.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes \"thinking progress vector\" (TPV), a vector representation of what is claimed the encoding of its \"interval reasoning process\".\nThe paper uses linear models, MLPs, and RNN as tools to monitor the change of this state along the reasoning trajectory of LLMs.\nThe paper claims that intervening this vector can change the behaviour of LLMs and make it reason \"faster\", or what is called \"overclocking\" by the paper.\nEmpirical results on math reasoning tasks shows some evidence about the effect of intervening TPVs.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "- The paper has some intuitive figures showing the discovered \"thinking progress vector\".", "weaknesses": "- The idea of estimating the progress of thinking as proposed in this paper does not make sense. It is beyond me why an LLM can maintain a state representing its reasoning progress if it does not know the correct answer to the problem even when the reasoning is just started. If an estimation of this really existed, then we would have asked this LLM to solve the halting problem, which has been mathematically proven to be impossible.\n\n- The TPV might represent some characteristic tokens marking the steps in solutions to math problems. The paper unfortunately does not provide enough study along this direction. In lines 261-263, even the paper itself writes \"a crucial question that arises is whether TPVs reflect a fundamental mechanism that the model uses to track its reasoning progress, or if they are merely residual artifacts that correlate with progress but do not play a causal role in the computation.\" this possibility is never ruled out in the rest of the paper.\n\n- The concept of reasoning progress itself is poorly defined for real-world problems. What does reasoning progress mean for writing some code for a specific task? Does reasoning ends when the code is completed? Or When LLM finishes running it in with a python interpreter? What does reasoning progress mean for the travel sales person (TSP) problem, when it is impossible to solve in polynomial time?", "questions": "I do not have questions for this paper. I have not checked all the experiments of the paper, so I will put a 2 for my confidence. \n\nI request that the paper be framed more accurately and align with the fundamentals of computer science in terms of the message it tries to convey.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1762100221717}, {"id": "GIQ9M6Mwox", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7951/Reviewer_pPTH"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents an interesting study towards the internal evidence of the relative position signals within the explicit thinking phase. To achieve this purpose, the authors aim to learn a 'progress vector' that can project the hidden states into the estimated progress. Then, they further examine the possiblity of overclock and downclock the reasoning process. The experimental results well demonstrate the effectiveness of the notion 'progress vector' and progress intervention made via the former.", "review_text": "This paper presents an interesting study towards the internal evidence of the relative position signals within the explicit thinking phase. To achieve this purpose, the authors aim to learn a 'progress vector' that can project the hidden states into the estimated progress. Then, they further examine the possiblity of overclock and downclock the reasoning process. The experimental results well demonstrate the effectiveness of the notion 'progress vector' and progress intervention made via the former.", "strengths": "1. The idea of identifying the intrinsic mechanisms that encode a model’s relative position within its internal reasoning process is very interesting and useful.\n2. The proposal of learning a 'progress vector' is simple yet effective.\n3. The extensive expeirments well demonstrate the effectiveness of the proposed solution.", "weaknesses": "1. The authors only choose mathematical reasoning task for study. It is only a specific subarea of LLM reasoning. More reasoning tasks from other domains should be investigated.", "questions": "See the above Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an interesting study towards the internal evidence of the relative position signals within the explicit thinking phase. To achieve this purpose, the authors aim to learn a 'progress vector' that can project the hidden states into the estimated progress. Then, they further examine the possiblity of overclock and downclock the reasoning process. The experimental results well demonstrate the effectiveness of the notion 'progress vector' and progress intervention made via the former.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The idea of identifying the intrinsic mechanisms that encode a model’s relative position within its internal reasoning process is very interesting and useful.\n2. The proposal of learning a 'progress vector' is simple yet effective.\n3. The extensive expeirments well demonstrate the effectiveness of the proposed solution.", "weaknesses": "1. The authors only choose mathematical reasoning task for study. It is only a specific subarea of LLM reasoning. More reasoning tasks from other domains should be investigated.", "questions": "See the above Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762064602950}, {"id": "b1pBCH1edS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7951/Reviewer_VS5W"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper investigates the mechanisms by which Large Language Models (LLMs) regulate the length of their explicit reasoning phase, specifically in models that use structured \"thinking\" tokens. The authors claim that LLMs internally encode their relative progress within the thinking phase. They train a regressor, termed a \"Thinking Progress Vector\" (TPV), to extract this progress (a scalar value from 0 to 1) from the model's hidden states, which can be visualized as a \"progress bar.\" Then they introduce an intervention method called \"overclocking,\" which manipulates the model's hidden states by adding a scaled version of this TPV. This intervention is shown to accelerate the reasoning process, reduce the number of generated tokens, and mitigate \"overthinking.\"", "review_text": "This paper investigates the mechanisms by which Large Language Models (LLMs) regulate the length of their explicit reasoning phase, specifically in models that use structured \"thinking\" tokens. The authors claim that LLMs internally encode their relative progress within the thinking phase. They train a regressor, termed a \"Thinking Progress Vector\" (TPV), to extract this progress (a scalar value from 0 to 1) from the model's hidden states, which can be visualized as a \"progress bar.\" Then they introduce an intervention method called \"overclocking,\" which manipulates the model's hidden states by adding a scaled version of this TPV. This intervention is shown to accelerate the reasoning process, reduce the number of generated tokens, and mitigate \"overthinking.\"", "strengths": "1. The central idea of identifying an internal representation of \"progress\" and visualizing it as a progress bar is novel and presents an interesting direction for LLM interpretability and human-computer interaction. \n2. The authors provide a thorough set of experiments on mathematical reasoning benchmarks, providing a solid empirical grounding for the paper's claims within its chosen domain.", "weaknesses": "1. Insufficient Evidence for \"Monitoring\" vs. \"Pattern-Matching\": The paper's primary claim that the TPV captures the model's monitoring of its own reasoning process, is not sufficiently substantiated. The TPV is trained as a simple regressor mapping hidden states to a normalized token position ($j/N_k$). It is highly probable that this regressor is not learning a high-level, semantic representation of \"reasoning progress\" but is instead capturing superficial, correlational text patterns. For example, certain phrases (\"Let me double-check,\" \"Wait,\" \"The formula is...\") naturally correlate with the beginning, middle, or end of a thinking sequence. The paper's own analysis in Appendix E, which shows that words like \"okay\" and \"right\" strongly and positively impact the progress value, seems to support this \"text pattern\" hypothesis far more than a \"cognitive monitoring\" one. \n\n2. Fair Results of the \"Overclocking\" Intervention: Given the concern above, the success of the \"overclocking\" intervention feels tautological and less surprising. If the TPV $\\theta$ is simply a vector that correlates strongly with the end of a sequence, then applying an intervention $h_{\\alpha} = h + \\alpha\\theta$ is, by definition, pushing the model's hidden state closer to a representation that is associated with sequence termination (such as `<eos>` token). It is therefore an expected outcome that this intervention shortens the output. This result does not prove that the model causally uses this vector for progress control, it only proves that the learned regressor has identified a manipulable direction in the activation space that correlates with sequence length. \n\n3. Questionable Generalizability: The empirical evaluation is confined to two mathematical reasoning datasets (MATH-500 and GSM8K). This domain is highly structured and often procedural. It is highly questionable whether these findings would generalize to other, more common LLM tasks. For instance, how would a \"progress vector\" function in open-ended creative writing, complex document summarization, or dialectical tasks where the concept of \"progress\" is ambiguous, non-linear, or recursive? The paper provides no evidence or discussion on this crucial aspect of generalizability. \n\n4. Ambiguous Practical and Analytical Significance: From an interpretability perspective, the work does not provide strong evidence for the high-level claims of \"metacognition\" or \"self-monitoring.\" It primarily demonstrates that text generation in a structured task has strong correlational patterns. From an engineering perspective, the method's value is not clearly demarcated from simpler alternatives. The \"Instruct\" baseline (a prompting technique) also performs well, and the best results are achieved when combining TPV with this prompting. This makes it unclear if this complex, model-specific intervention is practically superior to more robust methods like prompt engineering or fine-tuning with a length-based reward. \n\n5. Lack of Experimental Clarity: Key details of the methodology are ambiguous or omitted. The paper states that 30 problems were sampled, with 5 responses each, to train the regressor. This implies a training set of only ~150 trajectories, which seems exceptionally small for training even a linear regressor, let alone an RNN. The authors must clarify the exact dataset size (number of trajectories $K$ and total data points $(h, p)$). It is also not explicitly stated whether the linear TPV or the RNN model was used for the main intervention experiments in Section 4.", "questions": "1. The non-monotonic drops in progress (Fig. 5c, Fig. 9) are attributed to \"hesitation or reflection.\" An alternative, and simpler, hypothesis is that the model has simply generated a token (e.g., \"Wait\") that the TPV regressor strongly associates with an earlier phase of problem-solving (a learned text pattern). How can the authors more rigorously disentangle this \"text pattern\" hypothesis from their \"cognitive monitoring\" hypothesis? \n\n2. The raw TPV predictions are clearly non-monotonic. The paper mentions using exponential smoothing and an RNN to create a smoother progress bar. How is the final progress bar visualization (as in Fig. 1a) rendered in real-time? Is it based on the smoothed or RNN-based output? \n\n3. Could the authors clarify the intervention mechanism? Is the intervention $h_{\\alpha} = h + \\alpha\\theta$ applied at every token generation step within the `<think>...</think>` block? Furthermore, to understand the sensitivity of $\\alpha$, what is the typical scale of the TPV's squared norm ($||\\theta||^2$)? \n\n4. Could the authors comment on the hypothesized applicability and feasibility of this method for non-procedural, open-ended reasoning tasks, such as summarization or creative writing?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the mechanisms by which Large Language Models (LLMs) regulate the length of their explicit reasoning phase, specifically in models that use structured \"thinking\" tokens. The authors claim that LLMs internally encode their relative progress within the thinking phase. They train a regressor, termed a \"Thinking Progress Vector\" (TPV), to extract this progress (a scalar value from 0 to 1) from the model's hidden states, which can be visualized as a \"progress bar.\" Then they introduce an intervention method called \"overclocking,\" which manipulates the model's hidden states by adding a scaled version of this TPV. This intervention is shown to accelerate the reasoning process, reduce the number of generated tokens, and mitigate \"overthinking.\"", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The central idea of identifying an internal representation of \"progress\" and visualizing it as a progress bar is novel and presents an interesting direction for LLM interpretability and human-computer interaction. \n2. The authors provide a thorough set of experiments on mathematical reasoning benchmarks, providing a solid empirical grounding for the paper's claims within its chosen domain.", "weaknesses": "1. Insufficient Evidence for \"Monitoring\" vs. \"Pattern-Matching\": The paper's primary claim that the TPV captures the model's monitoring of its own reasoning process, is not sufficiently substantiated. The TPV is trained as a simple regressor mapping hidden states to a normalized token position ($j/N_k$). It is highly probable that this regressor is not learning a high-level, semantic representation of \"reasoning progress\" but is instead capturing superficial, correlational text patterns. For example, certain phrases (\"Let me double-check,\" \"Wait,\" \"The formula is...\") naturally correlate with the beginning, middle, or end of a thinking sequence. The paper's own analysis in Appendix E, which shows that words like \"okay\" and \"right\" strongly and positively impact the progress value, seems to support this \"text pattern\" hypothesis far more than a \"cognitive monitoring\" one. \n\n2. Fair Results of the \"Overclocking\" Intervention: Given the concern above, the success of the \"overclocking\" intervention feels tautological and less surprising. If the TPV $\\theta$ is simply a vector that correlates strongly with the end of a sequence, then applying an intervention $h_{\\alpha} = h + \\alpha\\theta$ is, by definition, pushing the model's hidden state closer to a representation that is associated with sequence termination (such as `<eos>` token). It is therefore an expected outcome that this intervention shortens the output. This result does not prove that the model causally uses this vector for progress control, it only proves that the learned regressor has identified a manipulable direction in the activation space that correlates with sequence length. \n\n3. Questionable Generalizability: The empirical evaluation is confined to two mathematical reasoning datasets (MATH-500 and GSM8K). This domain is highly structured and often procedural. It is highly questionable whether these findings would generalize to other, more common LLM tasks. For instance, how would a \"progress vector\" function in open-ended creative writing, complex document summarization, or dialectical tasks where the concept of \"progress\" is ambiguous, non-linear, or recursive? The paper provides no evidence or discussion on this crucial aspect of generalizability. \n\n4. Ambiguous Practical and Analytical Significance: From an interpretability perspective, the work does not provide strong evidence for the high-level claims of \"metacognition\" or \"self-monitoring.\" It primarily demonstrates that text generation in a structured task has strong correlational patterns. From an engineering perspective, the method's value is not clearly demarcated from simpler alternatives. The \"Instruct\" baseline (a prompting technique) also performs well, and the best results are achieved when combining TPV with this prompting. This makes it unclear if this complex, model-specific intervention is practically superior to more robust methods like prompt engineering or fine-tuning with a length-based reward. \n\n5. Lack of Experimental Clarity: Key details of the methodology are ambiguous or omitted. The paper states that 30 problems were sampled, with 5 responses each, to train the regressor. This implies a training set of only ~150 trajectories, which seems exceptionally small for training even a linear regressor, let alone an RNN. The authors must clarify the exact dataset size (number of trajectories $K$ and total data points $(h, p)$). It is also not explicitly stated whether the linear TPV or the RNN model was used for the main intervention experiments in Section 4.", "questions": "1. The non-monotonic drops in progress (Fig. 5c, Fig. 9) are attributed to \"hesitation or reflection.\" An alternative, and simpler, hypothesis is that the model has simply generated a token (e.g., \"Wait\") that the TPV regressor strongly associates with an earlier phase of problem-solving (a learned text pattern). How can the authors more rigorously disentangle this \"text pattern\" hypothesis from their \"cognitive monitoring\" hypothesis? \n\n2. The raw TPV predictions are clearly non-monotonic. The paper mentions using exponential smoothing and an RNN to create a smoother progress bar. How is the final progress bar visualization (as in Fig. 1a) rendered in real-time? Is it based on the smoothed or RNN-based output? \n\n3. Could the authors clarify the intervention mechanism? Is the intervention $h_{\\alpha} = h + \\alpha\\theta$ applied at every token generation step within the `<think>...</think>` block? Furthermore, to understand the sensitivity of $\\alpha$, what is the typical scale of the TPV's squared norm ($||\\theta||^2$)? \n\n4. Could the authors comment on the hypothesized applicability and feasibility of this method for non-procedural, open-ended reasoning tasks, such as summarization or creative writing?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761644847622}, {"id": "lkadgTG7UQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7951/Reviewer_DLQv"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposed learning from LLM's last layer hidden states to predict the progression of the thinking process between <think> and </think>. The progression is defined as the relative token position of the tokens in the thinking segment.\n\nThe paper proposed to learn a predictor parameterized by a single vector, a 2-layer MLP, and a single-layer GRU. Experiment results show the predictor has reasonable performance of predicting the relative position.\n\nThe paper proposed to intervene the last-layer hidden states utilizing the learned single vector, \"thinking progress vector\", to \"overclock\" the thinking process, i.e. by learning such a vector it decouples the \"progress\" concept in the hidden state vectors, and tries to manipulate it to speed up the progress. Experimental results are reported on Math500 and GSM8K using DeepSeek-R1-LLaMA-8B and DeepSeek-R1-Qwen-32B.", "review_text": "The paper proposed learning from LLM's last layer hidden states to predict the progression of the thinking process between <think> and </think>. The progression is defined as the relative token position of the tokens in the thinking segment.\n\nThe paper proposed to learn a predictor parameterized by a single vector, a 2-layer MLP, and a single-layer GRU. Experiment results show the predictor has reasonable performance of predicting the relative position.\n\nThe paper proposed to intervene the last-layer hidden states utilizing the learned single vector, \"thinking progress vector\", to \"overclock\" the thinking process, i.e. by learning such a vector it decouples the \"progress\" concept in the hidden state vectors, and tries to manipulate it to speed up the progress. Experimental results are reported on Math500 and GSM8K using DeepSeek-R1-LLaMA-8B and DeepSeek-R1-Qwen-32B.", "strengths": "The paper shows potential to isolate the \"progress\" concept in the hidden states of LLMs and the potential possibility to speed up or slow down the thinking process of LLMs by manipulating the hidden states using the \"thinking progress vector\".", "weaknesses": "1. The \"progress\" defined in this paper, the relative position of tokens in generation, does not capture the semantic progression of solving a problem. The task of learning the relative position is essentially just learning to predict to total length of the generation from the hidden states, as the position relative to the start of the generation is an easy task since information is embedded in the positional encoding.\n\n2. The results reported for both models, DeepSeek-R1-LLaMA-8B and DeepSeek-R1-Qwen-32B cannot match the officially reported accuracies on Math500, it's much lower. For 2048 length limit, reported 159/500=31.8% for 8B and 316/500=63.2% for 32B are much lower than official reports 89.1% and 94.3%. This should originate from the fact that the paper is limiting the generation length to only 2048 tokens. The reviewer thinks this is an unreasonable setting since the tested models (DeepSeek R1 distilled models) inherently generate long thinking contents before giving the answer, limiting the generation to only 2048 tokens is not testing the model properly, and the conclusions drawn from this setting are not informative. If for some scenarios where a 2048 token limit is a must, then some other models that generate shorter solutions should be used, e.g. Llama-3.1 or Qwen-2.5, not the R1 distilled version.", "questions": "To better present the potential benefit of the TPV intervention, I suggest the authors add another setting, comparing: 1) let the model generate until it end by itself; 2) let the model generate until it end by itself, and intervene the generation of every token using the proposed TPV.\n\nThen compare a) the accuracy and b) the averaged output length of the two methods. If the accuracy of the intervened generations are maintained, and the averaged output length is shorter, it would be clear evidence that the proposed intervention is useful in improving the efficiency of the reasoning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed learning from LLM's last layer hidden states to predict the progression of the thinking process between <think> and </think>. The progression is defined as the relative token position of the tokens in the thinking segment.\n\nThe paper proposed to learn a predictor parameterized by a single vector, a 2-layer MLP, and a single-layer GRU. Experiment results show the predictor has reasonable performance of predicting the relative position.\n\nThe paper proposed to intervene the last-layer hidden states utilizing the learned single vector, \"thinking progress vector\", to \"overclock\" the thinking process, i.e. by learning such a vector it decouples the \"progress\" concept in the hidden state vectors, and tries to manipulate it to speed up the progress. Experimental results are reported on Math500 and GSM8K using DeepSeek-R1-LLaMA-8B and DeepSeek-R1-Qwen-32B.", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "The paper shows potential to isolate the \"progress\" concept in the hidden states of LLMs and the potential possibility to speed up or slow down the thinking process of LLMs by manipulating the hidden states using the \"thinking progress vector\".", "weaknesses": "1. The \"progress\" defined in this paper, the relative position of tokens in generation, does not capture the semantic progression of solving a problem. The task of learning the relative position is essentially just learning to predict to total length of the generation from the hidden states, as the position relative to the start of the generation is an easy task since information is embedded in the positional encoding.\n\n2. The results reported for both models, DeepSeek-R1-LLaMA-8B and DeepSeek-R1-Qwen-32B cannot match the officially reported accuracies on Math500, it's much lower. For 2048 length limit, reported 159/500=31.8% for 8B and 316/500=63.2% for 32B are much lower than official reports 89.1% and 94.3%. This should originate from the fact that the paper is limiting the generation length to only 2048 tokens. The reviewer thinks this is an unreasonable setting since the tested models (DeepSeek R1 distilled models) inherently generate long thinking contents before giving the answer, limiting the generation to only 2048 tokens is not testing the model properly, and the conclusions drawn from this setting are not informative. If for some scenarios where a 2048 token limit is a must, then some other models that generate shorter solutions should be used, e.g. Llama-3.1 or Qwen-2.5, not the R1 distilled version.", "questions": "To better present the potential benefit of the TPV intervention, I suggest the authors add another setting, comparing: 1) let the model generate until it end by itself; 2) let the model generate until it end by itself, and intervene the generation of every token using the proposed TPV.\n\nThen compare a) the accuracy and b) the averaged output length of the two methods. If the accuracy of the intervened generations are maintained, and the averaged output length is shorter, it would be clear evidence that the proposed intervention is useful in improving the efficiency of the reasoning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761189083490}], "openreview_url": "https://openreview.net/forum?id=kXYuzGcfTT", "arxiv_id": "2506.07240", "paper_pdf": "papers/kXYuzGcfTT.pdf", "paper_pdf_sha256": "6da118bc0f3d1041a132070b12dff57f2fc6e740dec277d4a87a543f0596a3c2", "paper_pdf_bytes": 5687995, "paper_pdf_source": "openreview", "code_url": "https://github.com/royeisen/reasoning_loading_bar", "code_repository": "royeisen/reasoning_loading_bar", "code_commit": "98922eeb34eadcfdbfdeb6727c940e4b87b55c3c", "code_archive": "repos/kXYuzGcfTT.zip", "code_archive_sha256": "4fa51e51dc76815ad39cfab26aa655bb2e34061199d73bb371bcab5d95fc5190", "code_archive_bytes": 2375970, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 2318, "github_languages": {"Python": 50795}, "github_archived": false, "github_pushed_at": "2025-07-07T12:47:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/overclocking-llm-reasoning-monitoring-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Aemqy6Hjdj", "year": 2024, "status": "rejected", "title": "Enhancing Compositional Generalization via Compositional Feature Alignment", "authors": ["Haoxiang Wang", "Haozhe Si", "Huajie Shao", "Han Zhao"], "authorids": ["~Haoxiang_Wang1", "~Haozhe_Si1", "~Huajie_Shao1", "~Han_Zhao1"], "authors_source": "OpenReview API", "abstract": "Real-world applications of machine learning (ML) models often confront data distribution shifts, wherein discrepancies exist between the training and test data distributions. In the common multi-domain multi-class setup, as the number of classes and domains scales up, it becomes infeasible to gather training data for every domain-class combination. This challenge naturally leads the quest for models with Compositional Generalization (CG) ability, where models can generalize to unseen domain-class combinations. To delve into the CG challenge, we develop CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets, and observe that the prevalent pretraining-finetuning paradigm on foundational models, such as CLIP and DINOv2, struggles with the challenge. To address this challenge, we propose Compositional Feature Alignment (CFA), a simple two-stage finetuning technique that i) learns two orthogonal linear heads on a pretrained encoder with respect to class and domain labels, and ii) fine-tunes the encoder with the newly learned head frozen. We theoretically and empirically justify that CFA encourages compositional feature learning of pretrained models. We further conduct extensive experiments on CG-Bench for CLIP and DINOv2, two powerful pretrained vision foundation models. Experiment results show that CFA outperforms common finetuning techniques in compositional generalization, corroborating CFA's efficacy in compositional feature learning.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "dXXgp0LrVH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6129/Reviewer_H15m"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the challenge of compositional generalization in machine learning and focuses on generalization to unseen domain-class combinations. The author present a real-world benchmark named CG-Bench and propose the compositional feature alignment method to improve the CG performance of pretrained models. Extensive experiments demonstate the effectiveness of the method.", "review_text": "This paper studies the challenge of compositional generalization in machine learning and focuses on generalization to unseen domain-class combinations. The author present a real-world benchmark named CG-Bench and propose the compositional feature alignment method to improve the CG performance of pretrained models. Extensive experiments demonstate the effectiveness of the method.", "strengths": "S1: This paper solve a new setting or problem: can the model generalize to unseen domain-class combinations?\n\nS2: The overall writing is clear, including problem introduction and theoretical and experimental verification of experimental methods.\n\nS3: Some visualization experiments are shown to help the understanding of the review or reader.", "weaknesses": "W1: The setting of compositional generalization (CG) means that the pre-trained model is tested on the unseen domain-class combinations and the setting of domain generalization means that the pre-trained model is tested on the unseen domain samples? if yes, what is the difference between CG and some papers in open-set tasks to solve domain generalization problems?\n\nW2: To solve the CG problem, the authors propose a method to align the class information in different domain. I'm more curious about why the two-stage training method can achieve this goal? If possible, I would tend to see experiments on each stage of visualization.\n\nW3: I am more concerned about different training stages and ablation experiments related to orthogonal loss.", "questions": "see weaknesses\n\nIf the authors solve my concers, I tend to imporve my socre.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the challenge of compositional generalization in machine learning and focuses on generalization to unseen domain-class combinations. The author present a real-world benchmark named CG-Bench and propose the compositional feature alignment method to improve the CG performance of pretrained models. Extensive experiments demonstate the effectiveness of the method.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "S1: This paper solve a new setting or problem: can the model generalize to unseen domain-class combinations?\n\nS2: The overall writing is clear, including problem introduction and theoretical and experimental verification of experimental methods.\n\nS3: Some visualization experiments are shown to help the understanding of the review or reader.", "weaknesses": "W1: The setting of compositional generalization (CG) means that the pre-trained model is tested on the unseen domain-class combinations and the setting of domain generalization means that the pre-trained model is tested on the unseen domain samples? if yes, what is the difference between CG and some papers in open-set tasks to solve domain generalization problems?\n\nW2: To solve the CG problem, the authors propose a method to align the class information in different domain. I'm more curious about why the two-stage training method can achieve this goal? If possible, I would tend to see experiments on each stage of visualization.\n\nW3: I am more concerned about different training stages and ablation experiments related to orthogonal loss.", "questions": "see weaknesses\n\nIf the authors solve my concers, I tend to imporve my socre.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698837018261}, {"id": "SwULJKdsvk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6129/Reviewer_d1KC"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper tackles the challenge of data distribution shifts in machine learning applications. In particular, it emphasizes the multi-domain, multi-class setup where obtaining training data for every domain-class combination becomes impractical, and they try to examine the compositional generalization (CG) for learning models. The authors propose a simple solution in the form of CG and introduce \"CG-Bench,\" a suite of CG benchmarks derived from real-world image datasets. They tested CLIP and DINOv2, and show the effectiveness of both the proposed benchmark and method.\n\n---\n# Post rebuttal\n\nI appreciate the provided two solutions (lessen the need for a fully domain-labeled dataset). I encourage the author to incorporate all the discussion into their revision.", "review_text": "This paper tackles the challenge of data distribution shifts in machine learning applications. In particular, it emphasizes the multi-domain, multi-class setup where obtaining training data for every domain-class combination becomes impractical, and they try to examine the compositional generalization (CG) for learning models. The authors propose a simple solution in the form of CG and introduce \"CG-Bench,\" a suite of CG benchmarks derived from real-world image datasets. They tested CLIP and DINOv2, and show the effectiveness of both the proposed benchmark and method.\n\n---\n# Post rebuttal\n\nI appreciate the provided two solutions (lessen the need for a fully domain-labeled dataset). I encourage the author to incorporate all the discussion into their revision.", "strengths": "I personally believe the studied topic of compositional generalization is crucial in real-world machine learning applications, especially in scenarios with multi-domain, multi-class setups. The introduction of \"CG-Bench\" is commendable, providing the research community with a dedicated suite of benchmarks for evaluating CG performance.\n\nThe proposed two-stage process is clear and logically structured, with a rationale that suggests a theoretical underpinning for the method. I personally pretty like Figure 4 which provided a great comparisons between vanilla CLIP features and the features obtained by this paper.", "weaknesses": "It is hard to obtain the class and domain labels. Therefore, the studies in this paper are hard to scale-up to real world system. \n\nBased on Table 1, it seems the proposed method would usually cause negative effects to ID Acc.", "questions": "Besides, I feel the caption of Figure 2 could be improved to help reader understand the goal of the proposed method.\n\nPlease also address the concerns raised above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the challenge of data distribution shifts in machine learning applications. In particular, it emphasizes the multi-domain, multi-class setup where obtaining training data for every domain-class combination becomes impractical, and they try to examine the compositional generalization (CG) for learning models. The authors propose a simple solution in the form of CG and introduce \"CG-Bench,\" a suite of CG benchmarks derived from real-world image datasets. They tested CLIP and DINOv2, and show the effectiveness of both the proposed benchmark and method.\n\n---\n# Post rebuttal\n\nI appreciate the provided two solutions (lessen the need for a fully domain-labeled dataset). I encourage the author to incorporate all the discussion into their revision.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "I personally believe the studied topic of compositional generalization is crucial in real-world machine learning applications, especially in scenarios with multi-domain, multi-class setups. The introduction of \"CG-Bench\" is commendable, providing the research community with a dedicated suite of benchmarks for evaluating CG performance.\n\nThe proposed two-stage process is clear and logically structured, with a rationale that suggests a theoretical underpinning for the method. I personally pretty like Figure 4 which provided a great comparisons between vanilla CLIP features and the features obtained by this paper.", "weaknesses": "It is hard to obtain the class and domain labels. Therefore, the studies in this paper are hard to scale-up to real world system. \n\nBased on Table 1, it seems the proposed method would usually cause negative effects to ID Acc.", "questions": "Besides, I feel the caption of Figure 2 could be improved to help reader understand the goal of the proposed method.\n\nPlease also address the concerns raised above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698731965469}, {"id": "1TWOxXuQXA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6129/Reviewer_ZjmZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper deals with the challenge of compositional generalization, which in detail is the generalization to unseen domain-class combinations. To this end, the paper proposes CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets. Furthermore, Compositional Feature Alignment (CFA), a two-stage finetuning technique is proposed. Evaluation is performed on CG-Bench using CLIP and DINOv2 vision foundation models fine-tuned using the proposed CFA approach.", "review_text": "The paper deals with the challenge of compositional generalization, which in detail is the generalization to unseen domain-class combinations. To this end, the paper proposes CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets. Furthermore, Compositional Feature Alignment (CFA), a two-stage finetuning technique is proposed. Evaluation is performed on CG-Bench using CLIP and DINOv2 vision foundation models fine-tuned using the proposed CFA approach.", "strengths": "·         The paper is well written and easy to understand, e.g., Figure 2 is very useful for understanding the definition of Compositional Feature Structure.\n\n·         The proposed method is novel and interesting, and it seems to work well in the case of datasets such as Color-CIFAR. The results in Figure 4 show that the learned features are disengaged across classes and domains.\n\n·         The proposed CG-Bench compositional generalization benchmark is also novel and well-curated.", "weaknesses": "·         Theorem 1 holds only at the global minimum of the objective function (4). However, in practice when the features Z are from large neural networks such as CLIP the global minimum is unlikely to be obtained. In that case, features would likely not conform to the compositional feature structure defined in Definition 1. Therefore, it is unclear if Theorem 1 has any practical significance.\n\n·         While the features seem to be disentangled in the case of the simple Color-CIFAR dataset as shown in Figure 2, it is not clear if the same effect can be observed in the case of more complex datasets such as DomainNet.\n\n·         Stability of the model during training is not discussed in detail. This is important because of the complex two-stage training process. The paper should include experiments where the number of training steps in the first and second stage are varied and analyze the effect on the performance of the final model.\n\n·         From the results in Table 1, the performance gain over the reweight strategy is minimal (<1%) for all datasets. The biggest performance gain comes from the use of WiSE-FT (Wortsman et al., 2022). Also, compared to the reweight strategy, the proposed approach uses 2-staged training. Therefore, it is not clear whether the increase in training complexity is justified by the small performance gain.\n\n·         In Table 1, the reweight baseline with WiSE when using the DINOv2 model seems to be missing. Comparing reweighting and the proposed CFA method the gain in performance without WiSE seems to be <0.3%, especially in the case of OfficeHome and DomainNet.\n\n·         Can the proposed approach take advantage of unlabeled data? This is important because prior work such as CLIP or DINOv2 does not need explicit domain/class labels, unlike the proposed CFA approach.", "questions": "·         For models trained using CFA, do we observe the disentanglement similar to Color-CIFAR in Figure 2 in case of larger datasets such as DomainNet?\n\n·         Additional details of the reason for the small performance gain of the proposed CFA approach over the reweight baseline in Table 1 would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper deals with the challenge of compositional generalization, which in detail is the generalization to unseen domain-class combinations. To this end, the paper proposes CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets. Furthermore, Compositional Feature Alignment (CFA), a two-stage finetuning technique is proposed. Evaluation is performed on CG-Bench using CLIP and DINOv2 vision foundation models fine-tuned using the proposed CFA approach.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "·         The paper is well written and easy to understand, e.g., Figure 2 is very useful for understanding the definition of Compositional Feature Structure.\n\n·         The proposed method is novel and interesting, and it seems to work well in the case of datasets such as Color-CIFAR. The results in Figure 4 show that the learned features are disengaged across classes and domains.\n\n·         The proposed CG-Bench compositional generalization benchmark is also novel and well-curated.", "weaknesses": "·         Theorem 1 holds only at the global minimum of the objective function (4). However, in practice when the features Z are from large neural networks such as CLIP the global minimum is unlikely to be obtained. In that case, features would likely not conform to the compositional feature structure defined in Definition 1. Therefore, it is unclear if Theorem 1 has any practical significance.\n\n·         While the features seem to be disentangled in the case of the simple Color-CIFAR dataset as shown in Figure 2, it is not clear if the same effect can be observed in the case of more complex datasets such as DomainNet.\n\n·         Stability of the model during training is not discussed in detail. This is important because of the complex two-stage training process. The paper should include experiments where the number of training steps in the first and second stage are varied and analyze the effect on the performance of the final model.\n\n·         From the results in Table 1, the performance gain over the reweight strategy is minimal (<1%) for all datasets. The biggest performance gain comes from the use of WiSE-FT (Wortsman et al., 2022). Also, compared to the reweight strategy, the proposed approach uses 2-staged training. Therefore, it is not clear whether the increase in training complexity is justified by the small performance gain.\n\n·         In Table 1, the reweight baseline with WiSE when using the DINOv2 model seems to be missing. Comparing reweighting and the proposed CFA method the gain in performance without WiSE seems to be <0.3%, especially in the case of OfficeHome and DomainNet.\n\n·         Can the proposed approach take advantage of unlabeled data? This is important because prior work such as CLIP or DINOv2 does not need explicit domain/class labels, unlike the proposed CFA approach.", "questions": "·         For models trained using CFA, do we observe the disentanglement similar to Color-CIFAR in Figure 2 in case of larger datasets such as DomainNet?\n\n·         Additional details of the reason for the small performance gain of the proposed CFA approach over the reweight baseline in Table 1 would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698729425196}], "openreview_url": "https://openreview.net/forum?id=Aemqy6Hjdj", "arxiv_id": "2402.02851", "paper_pdf": "papers/Aemqy6Hjdj.pdf", "paper_pdf_sha256": "5f59cf258c7160728f91144659ed1274cd047e8a61a75d8a88fd1dde86f70603", "paper_pdf_bytes": 5692695, "paper_pdf_source": "openreview", "code_url": "https://github.com/Haoxiang-Wang/Compositional-Feature-Alignment", "code_repository": "Haoxiang-Wang/Compositional-Feature-Alignment", "code_commit": "70a12c1d5014fffd8b277db3cfaa023c75af28dc", "code_archive": "repos/Aemqy6Hjdj.zip", "code_archive_sha256": "18e106c3d0d5108d5abae0bd24eba827a9d0da8df3ff45f789ea6ed5a38c1d6d", "code_archive_bytes": 175358, "code_file_count": 78, "code_extensions": {".py": 78}, "github_disk_usage_kb": 147, "github_languages": {"Python": 413530}, "github_archived": false, "github_pushed_at": "2024-02-04T05:30:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/enhancing-compositional-generalization-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rjYUBo_uWEs", "year": 2023, "status": "rejected", "title": "Manifold Characteristics That Predict Downstream Task Performance", "authors": ["Ruan Henry Van der Merwe", "Gregory Newman", "Etienne Barnard"], "authorids": ["~Ruan_Henry_Van_der_Merwe1", "greg.newman@bytefuse.ai", "~Etienne_Barnard1"], "authors_source": "OpenReview API", "abstract": "Pretraining methods are typically compared by evaluating the accuracy of linear classifiers, transfer learning performance, or visually inspecting the representation manifold's (RM) lower-dimensional projections. We show that the differences between methods can be understood more clearly by investigating the RM directly, which allows for a more detailed comparison. To this end, we propose a framework and new metric to measure and compare different RMs. We also investigate and report on the RM characteristics for various pretraining methods. These characteristics are measured by applying sequentially larger local alterations to the input data, using white noise injections and Projected Gradient Descent (PGD) adversarial attacks, and then tracking each datapoint. We calculate the total distance moved for each datapoint and the relative change in distance between successive alterations. We show that self-supervised methods learn an RM where alterations lead to large but constant size changes, indicating a smoother RM than fully supervised methods. We then combine these measurements into one metric, the Representation Manifold Quality Metric (RMQM), where larger values indicate larger and less variable step sizes, and show that RMQM correlates positively with performance on downstream tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "W4E175X4Qu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper850/Reviewer_jVym"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper presents a score for evaluating the representational space of a model to estimate the potential generalization performance. An experiment with this method across a few datasets provides evidence to support this claim.", "review_text": "I really enjoyed reading this paper and I think the authors are doing some great work and have great ideas. However, I would prefer that the claims are supported with more complex datasets as well as contrasted with existing baselines. I believe the authors will find some from the publications in the PGDL2020 workshop that will demonstrate the potential improvement of their method.", "strengths": "Strengths: The paper is well organized. The argument to predict generalization is clear. The proposed metric is explained clearly.\n\nWeaknesses: I would have preferred to see the evaluation performed on larger datasets and not just small toy datasets. The support for the claim in the paper is solely empirical so this should be expanded on more. I also believe there should be more comparisons to existing work.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presents a score for evaluating the representational space of a model to estimate the potential generalization performance. An experiment with this method across a few datasets provides evidence to support this claim.", "strength_and_weaknesses": "Strengths: The paper is well organized. The argument to predict generalization is clear. The proposed metric is explained clearly.\n\nWeaknesses: I would have preferred to see the evaluation performed on larger datasets and not just small toy datasets. The support for the claim in the paper is solely empirical so this should be expanded on more. I also believe there should be more comparisons to existing work.\n", "clarity,_quality,_novelty_and_reproducibility": "In terms of novelty it seems strange to me that \"there has been no work done on using the structure of the RM as a predictor of generalisation\". Although, studying the perturbations may be novel. This work for one uses the compactness of the representation to predict generalization https://arxiv.org/abs/2012.02775 and I suggest the authors read the proceedings of this workshop to identify relevant methods to contrast theirs with https://sites.google.com/view/pgdl2020/resources Some of these methods can be used as baselines to help understand the significance of the perturbation approach.\n\nAlso, when the models were the models selected using early stopping? One experiment that could be interesting (not saying you need it for publication) is if you compute the RMQM at each epoch. One would expect to see a similar trend to the validation loss. An application of this would share the motivation of this work: https://arxiv.org/abs/1703.09580 which aims to not use a validation dataset.\n", "summary_of_the_review": "I really enjoyed reading this paper and I think the authors are doing some great work and have great ideas. However, I would prefer that the claims are supported with more complex datasets as well as contrasted with existing baselines. I believe the authors will find some from the publications in the PGDL2020 workshop that will demonstrate the potential improvement of their method.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666674426563}, {"id": "w8Cywzl45U0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper850/Reviewer_s4yj"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper proposes a framework and new methods to measure and compare different representation manifolds (RMs) to explore various pretraining methods. The analyses show that some self-supervised methods learn an RM where alterations lead to large but constant size changes, indicating a smoother RM than fully supervised methods. These measurements are integrated to develop a new  method called the Representation Manifold Quality Metric (RMQM), which is used to explore relationship between accuracy of models and RMs on downstream tasks.", "review_text": "The paper addresses an important problem of deep learning. However, there are various major and minor issues with the paper. Therefore, the paper is not ready for publication without fixing these issues.", "strengths": "The paper addresses an important problem of deep learning.\n\nHowever, there are various major and minor problems with the paper while addressing this problem. Briefly, there are unclear definitions and proposed methods are not employed correctly following common definitions and structures of manifolds. In addition, experimental analyses should be improved.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes a framework and new methods to measure and compare different representation manifolds (RMs) to explore various pretraining methods. The analyses show that some self-supervised methods learn an RM where alterations lead to large but constant size changes, indicating a smoother RM than fully supervised methods. These measurements are integrated to develop a new  method called the Representation Manifold Quality Metric (RMQM), which is used to explore relationship between accuracy of models and RMs on downstream tasks.", "strength_and_weaknesses": "The paper addresses an important problem of deep learning.\n\nHowever, there are various major and minor problems with the paper while addressing this problem. Briefly, there are unclear definitions and proposed methods are not employed correctly following common definitions and structures of manifolds. In addition, experimental analyses should be improved.", "clarity,_quality,_novelty_and_reproducibility": "The term metric is not used appropriately. Indeed, distance, metric and displacement are different entities.\n\nA distance metric is defined precisely in mathematics for different spaces. According to the well-known definitions of metrics employed for Euclidean spaces and nonlinear manifolds, (3) and (4) are not metric, therefore, RMQM is not a metric.\n\nTo resolve this issue, please either define your distances and metrics more precisely clarifying the differences between them, prove metric properties of the proposed functions, or use some other terms for the proposed entities.\n\nAnother major problem is comparison of points on manifolds. According to the discussions given in the motivation sections (e.g. Fig. 1), I suppose that nonlinear manifolds instead of linear Euclidean spaces are considered as RMs in the paper. Then, it is not clear why points on nonlinear manifolds are compared using Euclidean distances. \n\nIn addition, experimental analyses should be improved according to the proposed claims. First, additional datasets, methods and models should be explored for different vision tasks in addition to the classification tasks, and also for other AI tasks such as NLP and speech recognition. Second, the paper claims some results for self-supervised learning methods. Then, this claim should be verified using state-of-the-art self-supervised learning methods as well. If the experimental analyses cannot be extended, then some of the claims should be mathematically explored.\n\nSome typo:\nincludeTenenbaum → include Tenenbaum\nZhou et al. (2021) also evaluates → Zhou et al. (2021) also evaluate", "summary_of_the_review": "The paper addresses an important problem of deep learning. However, there are various major and minor issues with the paper. Therefore, the paper is not ready for publication without fixing these issues.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666633172096}, {"id": "oY8YaqE28v", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper850/Reviewer_xaAq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigates properties of trained manifolds with different learning methods on simple image data. For this, the authors propose data augmentation method and define metrics that lead to quantification of the quality of the overall manifold. They show that the proposed metric, i.e., RMQM, has high correlation with performances of several downstream prediction tasks. ", "review_text": "The authors provide an interesting aspect of training machine learning models for interpretability. However, the ideas are not sufficiently validated or proved, and they should be better presented. ", "strengths": "Strength:\n- Novelty. This paper relates prediction performance of downstream tasks with the characteristic of learnt manifold, which I think is quite interesting. \n\nWeakness:\n- Presentation of the paper. Too much white spaces and small text in the figures. Captions do not sufficiently describe the figures. \n- \"trained with different methods\" in the 4th paragraph in the intro is a little confusing. I thought \"different methods\" would refer to different optimization scheme but later it turns out that they refer to different tasks. \n- Perhaps a figure of the overall pipeline of the method will help a reader understand the ideas and aims of this paper. \n- There is a big jump going from Fig 2 to Fig 3 as two nubs are turned at the same time, i.e., dataset and perturbation method. It is difficult to tell which factor is causing \n- The authors propose a metric RMQM to measure the quality (e.g., smoothness) of the learned manifold. There must be other metrics, I am sorry but I am not sure which ones though (but at least the authors have done literature survey on manifold learning and comparison methods), but there is no way to tell that RMQM is the right one. \n- Lack of theory. While the paper is proposing interesting theories, they are only empirically validated with very limited settings, and it is not certain if the same observation will extend to more complicated experiments especially that may contain exhaustive local minima. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper investigates properties of trained manifolds with different learning methods on simple image data. For this, the authors propose data augmentation method and define metrics that lead to quantification of the quality of the overall manifold. They show that the proposed metric, i.e., RMQM, has high correlation with performances of several downstream prediction tasks. ", "strength_and_weaknesses": "Strength:\n- Novelty. This paper relates prediction performance of downstream tasks with the characteristic of learnt manifold, which I think is quite interesting. \n\nWeakness:\n- Presentation of the paper. Too much white spaces and small text in the figures. Captions do not sufficiently describe the figures. \n- \"trained with different methods\" in the 4th paragraph in the intro is a little confusing. I thought \"different methods\" would refer to different optimization scheme but later it turns out that they refer to different tasks. \n- Perhaps a figure of the overall pipeline of the method will help a reader understand the ideas and aims of this paper. \n- There is a big jump going from Fig 2 to Fig 3 as two nubs are turned at the same time, i.e., dataset and perturbation method. It is difficult to tell which factor is causing \n- The authors propose a metric RMQM to measure the quality (e.g., smoothness) of the learned manifold. There must be other metrics, I am sorry but I am not sure which ones though (but at least the authors have done literature survey on manifold learning and comparison methods), but there is no way to tell that RMQM is the right one. \n- Lack of theory. While the paper is proposing interesting theories, they are only empirically validated with very limited settings, and it is not certain if the same observation will extend to more complicated experiments especially that may contain exhaustive local minima. ", "clarity,_quality,_novelty_and_reproducibility": "The text is quite clear (but the overall presentation of the paper is poor) and I think it will be easy to reproduce the work done in this paper. ", "summary_of_the_review": "The authors provide an interesting aspect of training machine learning models for interpretability. However, the ideas are not sufficiently validated or proved, and they should be better presented. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666594164961}], "openreview_url": "https://openreview.net/forum?id=rjYUBo_uWEs", "arxiv_id": "2205.07477", "paper_pdf": "papers/rjYUBo_uWEs.pdf", "paper_pdf_sha256": "bdea4d6ee46e3d256c7607daf565db0545bf95b017325621c053c3c3eb47a1fa", "paper_pdf_bytes": 1030493, "paper_pdf_source": "openreview", "code_url": "https://github.com/ByteFuse/representation-manifold-quality-metric", "code_repository": "ByteFuse/representation-manifold-quality-metric", "code_commit": "92a3e925b1456848659505728b704844ba065801", "code_archive": "repos/rjYUBo_uWEs.zip", "code_archive_sha256": "651828f4ac09a876ae0cfa9b976ce6c8d6d4cb94d7f53b6be128b5081f774486", "code_archive_bytes": 245046, "code_file_count": 21, "code_extensions": {".py": 19, ".ipynb": 2}, "github_disk_usage_kb": 290, "github_languages": {"Jupyter Notebook": 385298, "Python": 138912}, "github_archived": false, "github_pushed_at": "2022-07-28T11:03:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/manifold-characteristics-that-predict"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Rivn22SJjg9", "year": 2022, "status": "rejected", "title": "Contrastive Embeddings for Neural Architectures", "authors": ["Daniel Hesslow", "Iacopo Poli"], "authorids": ["~Daniel_Hesslow1", "~Iacopo_Poli1"], "authors_source": "OpenReview API", "abstract": "The performance of algorithms for neural architecture search strongly depends on the parametrization of the search space. We use contrastive learning to identify networks across different initializations based on their data Jacobians and their number of parameters, and automatically produce the first architecture embeddings independent from the parametrization of the search space. Using our contrastive embeddings, we show that traditional black-box optimization algorithms, without modification, can reach state-of-the-art performance in Neural Architecture Search. As our method provides a unified embedding space, we successfully perform transfer learning between search spaces. Finally, we show the evolution of embeddings during training, motivating future studies into using embeddings at different training stages to gain a deeper understanding of the networks in a search space.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "2owNyiPQX0t", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3055/Reviewer_KncE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors use constrastive learning to embed NAS candidates into an informative search space, which is highly predictive of the architectures' actual accuracy. This embedding space allows traditional black box optimizers to perform well on NAS benchmarks.", "review_text": "The premise is interesting and the implications are exciting. The contributions in this paper are novel - to the best of my knowledge, this is the first paper that demonstrates that it may be possible to do transfer learning for NAS on related tasks.\n\nSection 4.3 - The predicted accuracy looks quite strong. I am curious why you didn't optimize the hyper parameters of your Random Forest. Further, how was this evaluation done? I couldn't find a description of a train-test split. Random Forests can overfit on data really easily, especially in this case since you'll have only 3 features as input.\n\nIn Section 4.4, I found the premise of the experiment interesting, but I'm unable to conclude anything interesting based on Figure 4. Can you present the data in a more readable format?\n\n5.1, 5.2 - I appreciate the simplicity of the experiment by not confounding the results by including the hyper parametrs of TPE and DE as part of the search space. However, the results in this section are mixed. It looks like CENA performs better than in the baseline in some cases, sometimes by just a small margin and sometimes even worse than REINFORCE.  Also, as was acknowledge by the authors in this section, computing Jacobians can be slow.\n\nApart from the final accuracy, it would be interesting to know how long each method took in wall clock time. If CENA is significantly faster than the other methods, you could make the case that it runs faster while preserving performance. On the other hand if CENA is extremely slow, then perhaps it is better to use random search to evaluate more architectures in the same time period as CENA.\n\n5.3 - In my opinion, this is the biggest strength of the method. However, I wish more details were provided. How exactly are you'll doing transfer learning? Are you using the contrastive learning network from one dataset and using it zero-shot on the other, and then regressing on the accuracies using Random Forests? Once again, was a train-test split used for the random forests and are the reported numbers the out-of-sample performance?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors use constrastive learning to embed NAS candidates into an informative search space, which is highly predictive of the architectures' actual accuracy. This embedding space allows traditional black box optimizers to perform well on NAS benchmarks.", "main_review": "The premise is interesting and the implications are exciting. The contributions in this paper are novel - to the best of my knowledge, this is the first paper that demonstrates that it may be possible to do transfer learning for NAS on related tasks.\n\nSection 4.3 - The predicted accuracy looks quite strong. I am curious why you didn't optimize the hyper parameters of your Random Forest. Further, how was this evaluation done? I couldn't find a description of a train-test split. Random Forests can overfit on data really easily, especially in this case since you'll have only 3 features as input.\n\nIn Section 4.4, I found the premise of the experiment interesting, but I'm unable to conclude anything interesting based on Figure 4. Can you present the data in a more readable format?\n\n5.1, 5.2 - I appreciate the simplicity of the experiment by not confounding the results by including the hyper parametrs of TPE and DE as part of the search space. However, the results in this section are mixed. It looks like CENA performs better than in the baseline in some cases, sometimes by just a small margin and sometimes even worse than REINFORCE.  Also, as was acknowledge by the authors in this section, computing Jacobians can be slow.\n\nApart from the final accuracy, it would be interesting to know how long each method took in wall clock time. If CENA is significantly faster than the other methods, you could make the case that it runs faster while preserving performance. On the other hand if CENA is extremely slow, then perhaps it is better to use random search to evaluate more architectures in the same time period as CENA.\n\n5.3 - In my opinion, this is the biggest strength of the method. However, I wish more details were provided. How exactly are you'll doing transfer learning? Are you using the contrastive learning network from one dataset and using it zero-shot on the other, and then regressing on the accuracies using Random Forests? Once again, was a train-test split used for the random forests and are the reported numbers the out-of-sample performance?", "summary_of_the_review": "The core idea and the benefits of this technique are exciting, however experimental details are lacking and the performance is a little lack-luster. \n\n- More details are needed about how the Random Forests were used.\n\n- The paper should focus more on the transfer learning aspect of this technique. I suggest you'll evaluate this technique on more datasets to build a strong case.\n\n- The paper should list the cost of evaluating each method. \n\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636254518992}, {"id": "g4o4VuqGBlB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3055/Reviewer_pbA7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a new technique to generate embeddings agnostic of the search space. This can be achieved by  first computing the data jacobian of the network with respect to datapoints sampled from different neighbourhoods. This jacobian matrix is then input to a \ncontrastive network which produces architecture embeddings. The contrastive views in this case are different initializations of the same network, which in turn must yield the same embedding. As these embeddings are high dimensional, they are reduced to lower dimension while ensuring that the distance between the embeddings is preserved in the lower dimensional space and the volume associated with each architecture is also preserved. ", "review_text": "Strengths:\n\n1. This is a novel way to generate the embeddings that does not require the architecture  to be encoded into a vector representation before it is fed in.\n2. It does not require computation of the accuracy.\n3. They demonstrated that the downstream embeddings have high predictive power by inputing them to a random forest, which was able to predict the accuracy, size etc. \n\nWeakness:\n1. Neural Architecture Search Without Training also used jacobian matrix and show that the kendall Tau is less than 0.55. So I am curious how you are able to achieve higher performance by using the embeddings whose input is the same. I am assuming that the contrastive learning is as good as predictive power of the input\n2. How would you decide which images from the dataset are used to compute the jacobian matrix? This would be crucial to the quality of the embeddings\n3. The original dimension of the embeddings, 512, is not that high.Random Forest and the NAS algorithms can handle it. Why should we reduce the dimension to 2? \n\nMore Experiments requested:\n1. The experiment conducted on NATS-BENCH is not enough to substantiate the claim that the embeddings can generalize across various search space. In order to bolster the claim, could you predict the accuracies for datasets in table 1 while training the contrastive network on NAS-BENCH 101 and DARTS search space and show its predictive power on NAS-BENCH 201\n\n2. For Figure 5, Could you also include arch2vec + REINFORCE, arch2vec + Bayesian Optimization,CENA + REINFORCE and CENA + Bayesian Optimization in the plot?\n\n3. For Table1, could you also include the performance prediction of random forest when using embeddings from NAO [1], GCN [2] and arch2vec\n\n[1] Neural Architecture Optimization, Luo et al.\n[2] Neural Predictor for Neural Architecture Search, Wen et al.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new technique to generate embeddings agnostic of the search space. This can be achieved by  first computing the data jacobian of the network with respect to datapoints sampled from different neighbourhoods. This jacobian matrix is then input to a \ncontrastive network which produces architecture embeddings. The contrastive views in this case are different initializations of the same network, which in turn must yield the same embedding. As these embeddings are high dimensional, they are reduced to lower dimension while ensuring that the distance between the embeddings is preserved in the lower dimensional space and the volume associated with each architecture is also preserved. ", "main_review": "Strengths:\n\n1. This is a novel way to generate the embeddings that does not require the architecture  to be encoded into a vector representation before it is fed in.\n2. It does not require computation of the accuracy.\n3. They demonstrated that the downstream embeddings have high predictive power by inputing them to a random forest, which was able to predict the accuracy, size etc. \n\nWeakness:\n1. Neural Architecture Search Without Training also used jacobian matrix and show that the kendall Tau is less than 0.55. So I am curious how you are able to achieve higher performance by using the embeddings whose input is the same. I am assuming that the contrastive learning is as good as predictive power of the input\n2. How would you decide which images from the dataset are used to compute the jacobian matrix? This would be crucial to the quality of the embeddings\n3. The original dimension of the embeddings, 512, is not that high.Random Forest and the NAS algorithms can handle it. Why should we reduce the dimension to 2? \n\nMore Experiments requested:\n1. The experiment conducted on NATS-BENCH is not enough to substantiate the claim that the embeddings can generalize across various search space. In order to bolster the claim, could you predict the accuracies for datasets in table 1 while training the contrastive network on NAS-BENCH 101 and DARTS search space and show its predictive power on NAS-BENCH 201\n\n2. For Figure 5, Could you also include arch2vec + REINFORCE, arch2vec + Bayesian Optimization,CENA + REINFORCE and CENA + Bayesian Optimization in the plot?\n\n3. For Table1, could you also include the performance prediction of random forest when using embeddings from NAO [1], GCN [2] and arch2vec\n\n[1] Neural Architecture Optimization, Luo et al.\n[2] Neural Predictor for Neural Architecture Search, Wen et al.\n\n", "summary_of_the_review": "This paper is novel and have demonstrated its potential. In order to make it more compelling, they must add more empirical evaluations requested above. If they do that, I am leaning towards an accept.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635935920687}, {"id": "LoQ3AGdxO7t", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3055/Reviewer_oXnB"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a self-supervised embedding learning method to learn embeddings of various-sized neural network architectures. Each network is first represented as a low rank projection of a Jacobian matrix, where the rows are Jacobians (output-averaged if multivariate) evaluated at various inputs at random initialization time, called EDJM. Since EDJM is random, multiple such representations are treated as positive pairs for contrastive learning. From the contrastively learned embedding, a further dimensionality-reduction stage is optimized to (1) preserve distances and (2) achieve uniform volume for each architecture. The main application of the final embedding is NAS, where the method outperform baselines.", "review_text": "Strengths:\n+ Novel neural architecture embedding that is agnostic to topology. \n+ Extensive visualizations on the learned embedding, and its effects in NAS settings. Transfer learning is a very nice experiment.\n\nWeaknesses:\nOverall I think the paper suffers from some clarity and motivation issues. I hope the authors can help clarify in the rebuttal period.\n\n+ EDJM computation is unclear\n\n   The exact empirical computation of EDJM is unclear:\n     1. What are the sampled inputs $x$? Are they fixed? If so, are they fixed for the same architecture, or across all architectures?\n     2. Why is each weight matrix only randomly scaled (instead of re-sampling each entry as usual)?\n\n+ Second dimensionality-reduction stage motivation\n\n   The authors motivated from (1) a dimensionality-reduction perspective and (2) wanting each architecture to have similar volume. I am not convinced by both motivations. \n\n    For (1), the contrastively-learned embedding is 256-dimensional, which is not too high for random search or differential evolution in usual cases. Does the NAS setting require a smaller dimensionality to work? If so, can there be an ablation studies (on dimensionality, and on 2nd stage)?\n\n   For (2), I am not following why Eqn. (5) can be used to achieve roughly uniform volume assignment. The volume here seems to refer to the volume of the set of embedding vectors. But isn't that given by Eqn. (4), which optimizes the corresponding $y$ vectors for each $x$? Also for Eqn. (4), are the $y$ vectors optimized directly or is there an encoder of some sort? If not, is the method limited to training and testing on the same fixed set of architectures? The notations here are a bit confusing. Is $d_\\pi$ a differential or a distance? I was assuming the former (as an OT plan). But why is it operating on $(x, y)$ in Eqn. (3) but $(y,y')$ in Eqn. (5)?\n\n+ Directly using EPDJM\n\n   From Fig. 2a, the top performing architectures are already mostly lined up in EPDJM space. What if one run NAS directly on that?\n\n+ Decoder?\n\n   From my understanding, NAS methods like DE requires a decoding mechanism. What is the decoder here?\n\nMinor presentation issues:\n+ It'd be great to see the training time direction in Fig. 4 .", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a self-supervised embedding learning method to learn embeddings of various-sized neural network architectures. Each network is first represented as a low rank projection of a Jacobian matrix, where the rows are Jacobians (output-averaged if multivariate) evaluated at various inputs at random initialization time, called EDJM. Since EDJM is random, multiple such representations are treated as positive pairs for contrastive learning. From the contrastively learned embedding, a further dimensionality-reduction stage is optimized to (1) preserve distances and (2) achieve uniform volume for each architecture. The main application of the final embedding is NAS, where the method outperform baselines.", "main_review": "Strengths:\n+ Novel neural architecture embedding that is agnostic to topology. \n+ Extensive visualizations on the learned embedding, and its effects in NAS settings. Transfer learning is a very nice experiment.\n\nWeaknesses:\nOverall I think the paper suffers from some clarity and motivation issues. I hope the authors can help clarify in the rebuttal period.\n\n+ EDJM computation is unclear\n\n   The exact empirical computation of EDJM is unclear:\n     1. What are the sampled inputs $x$? Are they fixed? If so, are they fixed for the same architecture, or across all architectures?\n     2. Why is each weight matrix only randomly scaled (instead of re-sampling each entry as usual)?\n\n+ Second dimensionality-reduction stage motivation\n\n   The authors motivated from (1) a dimensionality-reduction perspective and (2) wanting each architecture to have similar volume. I am not convinced by both motivations. \n\n    For (1), the contrastively-learned embedding is 256-dimensional, which is not too high for random search or differential evolution in usual cases. Does the NAS setting require a smaller dimensionality to work? If so, can there be an ablation studies (on dimensionality, and on 2nd stage)?\n\n   For (2), I am not following why Eqn. (5) can be used to achieve roughly uniform volume assignment. The volume here seems to refer to the volume of the set of embedding vectors. But isn't that given by Eqn. (4), which optimizes the corresponding $y$ vectors for each $x$? Also for Eqn. (4), are the $y$ vectors optimized directly or is there an encoder of some sort? If not, is the method limited to training and testing on the same fixed set of architectures? The notations here are a bit confusing. Is $d_\\pi$ a differential or a distance? I was assuming the former (as an OT plan). But why is it operating on $(x, y)$ in Eqn. (3) but $(y,y')$ in Eqn. (5)?\n\n+ Directly using EPDJM\n\n   From Fig. 2a, the top performing architectures are already mostly lined up in EPDJM space. What if one run NAS directly on that?\n\n+ Decoder?\n\n   From my understanding, NAS methods like DE requires a decoding mechanism. What is the decoder here?\n\nMinor presentation issues:\n+ It'd be great to see the training time direction in Fig. 4 .", "summary_of_the_review": "The problem of embedding various architectures into the same fixed-length vector is an interesting one. The present paper proposes an interesting approached based on contrastive-learning. However, the method is not described clearly, and some of the algorithmic choices are not well motivated. Given these considerations, I don't recommend acceptance in its current form.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635832731976}], "openreview_url": "https://openreview.net/forum?id=Rivn22SJjg9", "arxiv_id": "2102.04208", "paper_pdf": "papers/Rivn22SJjg9.pdf", "paper_pdf_sha256": "36c094f7d14283a8afa57a775ab4bf0079814bdd91a67031172354e60a3ab5ee", "paper_pdf_bytes": 5772111, "paper_pdf_source": "openreview", "code_url": "https://github.com/lightonai/contrastive-embeddings-for-neural-architectures", "code_repository": "lightonai/contrastive-embeddings-for-neural-architectures", "code_commit": "8606a4a4d7ea9ed9c01964660e8b2afb1e003480", "code_archive": "repos/Rivn22SJjg9.zip", "code_archive_sha256": "18da13d12b2033eb40a0d09c44d849df76e823b6c9007c65ecd184c93ad1f52a", "code_archive_bytes": 807640, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 784, "github_languages": {"Python": 59168}, "github_archived": true, "github_pushed_at": "2021-03-17T10:46:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contrastive-embeddings-for-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "K5a_QFEUzA1", "year": 2021, "status": "rejected", "title": "Cross-model Back-translated Distillation for Unsupervised Machine Translation", "authors": ["Phi Xuan Nguyen", "Shafiq Joty", "Kui Wu", "AiTi Aw"], "authorids": ["~Phi_Xuan_Nguyen1", "~Shafiq_Joty1", "wuk@i2r.a-star.edu.sg", "~AiTi_Aw1"], "authors_source": "OpenReview API", "abstract": "Recent unsupervised machine translation (UMT) systems usually employ three main principles: initialization, language modeling and iterative back-translation, though they may apply them differently. Crucially, iterative back-translation and denoising auto-encoding for language modeling provide data diversity to train the UMT systems. However, these diversification processes may have reached their limit. We introduce a novel component to the standard UMT framework called Cross-model Back-translated Distillation (CBD), that is aimed to induce another level of data diversification that existing principles lack. CBD is applicable to all previous UMT approaches. In our experiments, it boosts the performance of the standard UMT methods by 1.5-2.0 BLEU. In particular, in WMT'14 English-French, WMT'16 German-English and English-Romanian, CBD outperforms cross-lingual masked language model (XLM) by 2.3, 2.2 and 1.6 BLEU, respectively. It also yields 1.5-3.3 BLEU improvements in IWSLT English-French and English-German tasks. Through extensive experimental analyses, we show that CBD is effective because it embraces data diversity while other similar variants do not.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "p7W-X0jkVas", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1747/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new component to the unsupervised machine translation framework called cross-model back-translated distillation. The proposed approach is applicable to the other unsupervised methods. Experimental results in several translation tasks show that the proposed approach improves the translation accuracy of the standard unsupervised machine translation models, outperforming the cross-lingual masked language model. \n\n- The analyses are interesting to understand the proposed approach. Table 4 reports the diverse of synthetic data, but what about the quality as parallel data? Can you use parallel data instead and conduct the same analyses so that you can use BLEU score as an evaluation metric?\n- Is it possible to apply the proposed approach to supervised NMT training, by creating BT data from monolingual data? \n- Table 1 reports that the results from your experiments show that the equivalent/better performance agains the existing models with much fewer data. What about scaling up the monolingual data size 5x/10x more? Will the performance be improved better and better?\n- \"the translated products (x-y) of the UMT teachers.\" at p.6.  What does this \"x-y\" mean?\n\nTypo:\np.3 5.2.In Appendix -> 5.2. In Appendix", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Strong results but a few unclear parts in the paper", "review": "This paper introduces a new component to the unsupervised machine translation framework called cross-model back-translated distillation. The proposed approach is applicable to the other unsupervised methods. Experimental results in several translation tasks show that the proposed approach improves the translation accuracy of the standard unsupervised machine translation models, outperforming the cross-lingual masked language model. \n\n- The analyses are interesting to understand the proposed approach. Table 4 reports the diverse of synthetic data, but what about the quality as parallel data? Can you use parallel data instead and conduct the same analyses so that you can use BLEU score as an evaluation metric?\n- Is it possible to apply the proposed approach to supervised NMT training, by creating BT data from monolingual data? \n- Table 1 reports that the results from your experiments show that the equivalent/better performance agains the existing models with much fewer data. What about scaling up the monolingual data size 5x/10x more? Will the performance be improved better and better?\n- \"the translated products (x-y) of the UMT teachers.\" at p.6.  What does this \"x-y\" mean?\n\nTypo:\np.3 5.2.In Appendix -> 5.2. In Appendix", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604040634895}, {"id": "yrhIBVUpkjj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1747/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes a method to enhance unsupervised machine translation through data augmentation. The idea is pretty straight-forward, if not altogether intuitive, you begin by training two bidirectional (i.e.: they can translate source to target and target to source) unsupervised MT systems A and B. The tested scenarios always have A and B be identical architectures trained with different initializations. They then produce synthetic source-target pairs by first having A (source->target) translate the provided source sentence x to y, and then having B (target->source) translate y back to z. They then train supervised MT on both x,y and z,y. The same procedure can be repeated with source and target reversed. The authors show substantial (1-2) BLEU improvements with 3 different UMT systems in 5 low-data scenarios (En-Fr, Fr-En, En-De, En-Ro, Ro-En), all subsampled to 5M monolingual sentences for each language. In En-Fr and Fr-En and En-De, they are able to match reported XLM results from Conneau and Lample 2019, despite using much less data.\n\nThis simple idea is explored extremely thoroughly. The paper reads more like a journal paper that has undergone several stages of review than a conference paper. The authors make connections to and compare against a number of relevant ensembling strategies (to account for two systems being used) and back-translation-diversification strategies (to account for multiple sources being produced for the same target), and consistently show that only their specific recipe leads to the same levels of improvement. The authors really leave no stone unturned.\n\nThe biggest knock against this paper is the relatively small data scenario. Having two UMT systems allows them to provide two source sentences (one original and one synthetic) for each target sentence (always synthetic), but how important is this when we have 25x more original source sentences? I can imagine arguments for why the high-data UMT scenario is unrealistic (many monolingual sentences implies the likely presence of parallel data), but those arguments aren’t presented in the paper. It would be greatly strengthened by a full-data experiment for even just one or two of the language pairs.\n\nBeyond that, I have few concerns. The paper is clear, easy to follow, and as I said, very thorough. But I’ll do my best to make some constructive criticisms:\n\nI think the Related Work section feels a little superfluous after all of the comparisons made to related work in the Background and in the Experiments. I think I would like to have seen more discussion of the highlighly related work in sections 5.3 and 5.4. In particular, a more detailed discussion of this method’s relation to multi-agent dual learning would be worth giving up parts of Related work that are already mentioned in Background (like pre-neural statistical unsupervised MT).\n\nIt would be useful to specify how BLEU is calculated, to help readers understand just how useful the cross-paper BLEU comparisons in Table 1 are.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Very thorough exploration of a simple, flexible idea. Unclear whether the data augmentation still matters when more monolingual data is available.", "review": "This paper describes a method to enhance unsupervised machine translation through data augmentation. The idea is pretty straight-forward, if not altogether intuitive, you begin by training two bidirectional (i.e.: they can translate source to target and target to source) unsupervised MT systems A and B. The tested scenarios always have A and B be identical architectures trained with different initializations. They then produce synthetic source-target pairs by first having A (source->target) translate the provided source sentence x to y, and then having B (target->source) translate y back to z. They then train supervised MT on both x,y and z,y. The same procedure can be repeated with source and target reversed. The authors show substantial (1-2) BLEU improvements with 3 different UMT systems in 5 low-data scenarios (En-Fr, Fr-En, En-De, En-Ro, Ro-En), all subsampled to 5M monolingual sentences for each language. In En-Fr and Fr-En and En-De, they are able to match reported XLM results from Conneau and Lample 2019, despite using much less data.\n\nThis simple idea is explored extremely thoroughly. The paper reads more like a journal paper that has undergone several stages of review than a conference paper. The authors make connections to and compare against a number of relevant ensembling strategies (to account for two systems being used) and back-translation-diversification strategies (to account for multiple sources being produced for the same target), and consistently show that only their specific recipe leads to the same levels of improvement. The authors really leave no stone unturned.\n\nThe biggest knock against this paper is the relatively small data scenario. Having two UMT systems allows them to provide two source sentences (one original and one synthetic) for each target sentence (always synthetic), but how important is this when we have 25x more original source sentences? I can imagine arguments for why the high-data UMT scenario is unrealistic (many monolingual sentences implies the likely presence of parallel data), but those arguments aren’t presented in the paper. It would be greatly strengthened by a full-data experiment for even just one or two of the language pairs.\n\nBeyond that, I have few concerns. The paper is clear, easy to follow, and as I said, very thorough. But I’ll do my best to make some constructive criticisms:\n\nI think the Related Work section feels a little superfluous after all of the comparisons made to related work in the Background and in the Experiments. I think I would like to have seen more discussion of the highlighly related work in sections 5.3 and 5.4. In particular, a more detailed discussion of this method’s relation to multi-agent dual learning would be worth giving up parts of Related work that are already mentioned in Background (like pre-neural statistical unsupervised MT).\n\nIt would be useful to specify how BLEU is calculated, to help readers understand just how useful the cross-paper BLEU comparisons in Table 1 are.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603900961175}, {"id": "8slankuRNKO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1747/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: The paper proposes an additional stage of training for unsupervised NMT models utilizing synthetic data generated from multiple independently trained models. The generated synthetic data uses two stages of back-translation, with different models, in order to \"diversify\" the set of training data used for fine-tuning the models. This is similar to the approach in [1], but uses an additional stage of back-translation with a different model. The authors add this additional stage of training to unsupervised NMT models using different pipelines (PB unsupervised MT, Neural Unsupervised MT, XLM) and show that their approach improves all of these approaches by 1.5-2 Bleu on WMT En-Fr, De-En and En-Ro.\n\nStrengths:\n1. The paper is well written, the approach is simple and seems to improve quality by significant amounts in a variety of experimental settings.\n2. The authors do a great job of comparing against several relevant approaches (sampling during back-translation, ensembling, multi-agent dual learning). The paper compares against most of the relevant approaches I could think of while reading the paper.\n\nWeaknesses / Questions for authors:\n1. As with any NMT model trained with synthetic data, it would be better to report results on source and target original splits of the test data to provide a clearer evaluation [2,3]. Also clarify the Bleu scripts, tokenization and other post-processing used for evaluation.\n2. The datasets used for experimentation are much smaller than the ones used for the baseline unsupervised-NMT approaches. It would be great to report results in the original training conditions (this is not a major limitation however, since the proposed approach seems to improve over baselines trained with more data).\n3. Did the authors try any experiments with unsupervised models utilizing parallel data in unrelated languages, similar to [4,5] or in real low resource settings [6]? These are more practical conditions for unsupervised MT in true low resource languages.\n\nRecommendation: Overall, this is a good paper and I would recommend acceptance. While I would have also liked to see experiments in more realistic low-resource settings, the current paper does a good enough job of evaluating the approach in standard unsupervised NMT settings on related high resource languages.\n\nReferences:\n[1] Data Diversification: A Simple Strategy For Neural Machine Translation, Nguyen et al.\n[2] APE at Scale and its Implications on MT Evaluation Biases, Freitag et al.\n[3] On The Evaluation of Machine Translation Systems Trained With Back-Translation, Edunov et al.\n[4] Multilingual Denoising Pre-training for Neural Machine Translation, Liu et al.\n[5] Leveraging Monolingual Data with Self-Supervision for Multilingual Neural Machine Translation, Siddhant et al.\n[6] When Does Unsupervised Machine Translation Work?, Marchisio et al.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good paper, simple approach, thorough experiments", "review": "Summary: The paper proposes an additional stage of training for unsupervised NMT models utilizing synthetic data generated from multiple independently trained models. The generated synthetic data uses two stages of back-translation, with different models, in order to \"diversify\" the set of training data used for fine-tuning the models. This is similar to the approach in [1], but uses an additional stage of back-translation with a different model. The authors add this additional stage of training to unsupervised NMT models using different pipelines (PB unsupervised MT, Neural Unsupervised MT, XLM) and show that their approach improves all of these approaches by 1.5-2 Bleu on WMT En-Fr, De-En and En-Ro.\n\nStrengths:\n1. The paper is well written, the approach is simple and seems to improve quality by significant amounts in a variety of experimental settings.\n2. The authors do a great job of comparing against several relevant approaches (sampling during back-translation, ensembling, multi-agent dual learning). The paper compares against most of the relevant approaches I could think of while reading the paper.\n\nWeaknesses / Questions for authors:\n1. As with any NMT model trained with synthetic data, it would be better to report results on source and target original splits of the test data to provide a clearer evaluation [2,3]. Also clarify the Bleu scripts, tokenization and other post-processing used for evaluation.\n2. The datasets used for experimentation are much smaller than the ones used for the baseline unsupervised-NMT approaches. It would be great to report results in the original training conditions (this is not a major limitation however, since the proposed approach seems to improve over baselines trained with more data).\n3. Did the authors try any experiments with unsupervised models utilizing parallel data in unrelated languages, similar to [4,5] or in real low resource settings [6]? These are more practical conditions for unsupervised MT in true low resource languages.\n\nRecommendation: Overall, this is a good paper and I would recommend acceptance. While I would have also liked to see experiments in more realistic low-resource settings, the current paper does a good enough job of evaluating the approach in standard unsupervised NMT settings on related high resource languages.\n\nReferences:\n[1] Data Diversification: A Simple Strategy For Neural Machine Translation, Nguyen et al.\n[2] APE at Scale and its Implications on MT Evaluation Biases, Freitag et al.\n[3] On The Evaluation of Machine Translation Systems Trained With Back-Translation, Edunov et al.\n[4] Multilingual Denoising Pre-training for Neural Machine Translation, Liu et al.\n[5] Leveraging Monolingual Data with Self-Supervision for Multilingual Neural Machine Translation, Siddhant et al.\n[6] When Does Unsupervised Machine Translation Work?, Marchisio et al.", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603895619354}, {"id": "3pLTvufPPm1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1747/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, two unsupervised agents are utilized at cross-model by using the dual nature of the unsupervised machine translation model, in which forward translation of agent_1 is combined with the backward translation of agent_2, more synthetic translation pairs are obtained to train a new supervised machine translation model. The result is improved on multiple unsupervised machine translation, and this paper claims that more diversity is brought to the synthetic data, so a better translation model can be trained. This paper uses a reconstruction BLEU or BT BLEU [1] metric to compare the effect of the inside-model with that of cross-model, and finds that cross model translation has a lower back-translation effect, which shows that the diversity is enhanced. Furthermore, CBD is compared with the ensemble method and achieves better performance. The proposed method is quite simple yet effective, but it is also a kind of data enhancement.\n\nIn addition to these contributions, the paper also has some shortcomings\n\n1. The evidence in this paper can not support the claim that the current performance bottleneck of UMT is due to the lack of diversity: the performance upper limit of UMT is still due to the lack of clear supervision signal, which limits the further performance growth. Because the training of CBD is divided into two stages, the diversity of the second stage only brings more training data to enhance the supervised machine translation model, rather than unsupervised machine translation effect.\n\n2. Source of promotion: the second stage of CBD method adopts (x_s, y_t), (z_s, y_t), (y_s, x_t), (y_s, z_t) synthetic translation pairs, it is not clear how much performance growth comes from increased data and how much growth comes from the new model implementation (ott et al., 2018). It is not appropriate to attribute all contributions to the diversity brought by CBD. I suggest that the author should use (y_s, x_t) data to train based on the (ott et al., 2018) model, and report the effect comparison (In my experiments, the second stage model implemented with fairseq trained only on (y_s, x_t) surpass both agents trained with XLM due to more efficient implementation in fairseq).\n\n3. Unfair comparison with the enable distillation: authors need to compare CBD with the model trained on the synthetic data (y_s, x_t) of the ensemble of agent_1 and agent_ 2. In the training data (x_s, y_t), (z_s, y_t), (y_s, x_t), (y_s, z_t) for the second stage of CDB,  x_t, golden language sequences as translation target is stronger than synthetic language sequences (silver) as target. Therefore, It is necessary to report the real result of ensembled distillation. The current results are very unreliable. In addition, it is necessary to compare the training time of the CBD method and ensemble distillation training (including the decoding process after the first stage of training) to show the efficiency of CBD.\n\n4. The non-golden language sequence as a translation target is called pseudo-NMT (PNMT). The author adopts a variety of model structures, which is slightly redundant. They can directly add the synthetic data decoded by cross model to continually train the original XLM model with a supervised translation objective (which is naturally supported in XLM from my experience), and report the effect comparison between them.\n\n5. The essence of the CDB approach is a process of self-supervised training, so it is necessary to compare self-training/tri-training introduced in [2].\n\n\nIn general, the CBD method in this paper is a simple and effective data enhancement method to improve the performance of the model. However, due to the lack of many important details of the implementation, despite the promotion, the source of promotion is unknown. In addition, the unreasonable comparison of the baseline models deepens my concern about the real promotion of this CBD method.\n\n[1] Li, Zuchao, et al. \"Reference Language based Unsupervised Neural Machine Translation.\" arXiv preprint arXiv:2004.02127 (2020).\n\n[2] Sun, Haipeng, et al. \"Self-Training for Unsupervised Neural Machine Translation in Unbalanced Training Data Scenarios.\" arXiv preprint arXiv:2004.04507 (2020).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper introduce a novel data enhancement approach for unsupervised neural machine translation.", "review": "In this paper, two unsupervised agents are utilized at cross-model by using the dual nature of the unsupervised machine translation model, in which forward translation of agent_1 is combined with the backward translation of agent_2, more synthetic translation pairs are obtained to train a new supervised machine translation model. The result is improved on multiple unsupervised machine translation, and this paper claims that more diversity is brought to the synthetic data, so a better translation model can be trained. This paper uses a reconstruction BLEU or BT BLEU [1] metric to compare the effect of the inside-model with that of cross-model, and finds that cross model translation has a lower back-translation effect, which shows that the diversity is enhanced. Furthermore, CBD is compared with the ensemble method and achieves better performance. The proposed method is quite simple yet effective, but it is also a kind of data enhancement.\n\nIn addition to these contributions, the paper also has some shortcomings\n\n1. The evidence in this paper can not support the claim that the current performance bottleneck of UMT is due to the lack of diversity: the performance upper limit of UMT is still due to the lack of clear supervision signal, which limits the further performance growth. Because the training of CBD is divided into two stages, the diversity of the second stage only brings more training data to enhance the supervised machine translation model, rather than unsupervised machine translation effect.\n\n2. Source of promotion: the second stage of CBD method adopts (x_s, y_t), (z_s, y_t), (y_s, x_t), (y_s, z_t) synthetic translation pairs, it is not clear how much performance growth comes from increased data and how much growth comes from the new model implementation (ott et al., 2018). It is not appropriate to attribute all contributions to the diversity brought by CBD. I suggest that the author should use (y_s, x_t) data to train based on the (ott et al., 2018) model, and report the effect comparison (In my experiments, the second stage model implemented with fairseq trained only on (y_s, x_t) surpass both agents trained with XLM due to more efficient implementation in fairseq).\n\n3. Unfair comparison with the enable distillation: authors need to compare CBD with the model trained on the synthetic data (y_s, x_t) of the ensemble of agent_1 and agent_ 2. In the training data (x_s, y_t), (z_s, y_t), (y_s, x_t), (y_s, z_t) for the second stage of CDB,  x_t, golden language sequences as translation target is stronger than synthetic language sequences (silver) as target. Therefore, It is necessary to report the real result of ensembled distillation. The current results are very unreliable. In addition, it is necessary to compare the training time of the CBD method and ensemble distillation training (including the decoding process after the first stage of training) to show the efficiency of CBD.\n\n4. The non-golden language sequence as a translation target is called pseudo-NMT (PNMT). The author adopts a variety of model structures, which is slightly redundant. They can directly add the synthetic data decoded by cross model to continually train the original XLM model with a supervised translation objective (which is naturally supported in XLM from my experience), and report the effect comparison between them.\n\n5. The essence of the CDB approach is a process of self-supervised training, so it is necessary to compare self-training/tri-training introduced in [2].\n\n\nIn general, the CBD method in this paper is a simple and effective data enhancement method to improve the performance of the model. However, due to the lack of many important details of the implementation, despite the promotion, the source of promotion is unknown. In addition, the unreasonable comparison of the baseline models deepens my concern about the real promotion of this CBD method.\n\n[1] Li, Zuchao, et al. \"Reference Language based Unsupervised Neural Machine Translation.\" arXiv preprint arXiv:2004.02127 (2020).\n\n[2] Sun, Haipeng, et al. \"Self-Training for Unsupervised Neural Machine Translation in Unbalanced Training Data Scenarios.\" arXiv preprint arXiv:2004.04507 (2020).", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603808247420}], "openreview_url": "https://openreview.net/forum?id=K5a_QFEUzA1", "arxiv_id": "2006.02163", "paper_pdf": "papers/K5a_QFEUzA1.pdf", "paper_pdf_sha256": "e2b1a8ab2032f2460291f063d3fb79b212ba084fe8709afa34aea5f295a664de", "paper_pdf_bytes": 275372, "paper_pdf_source": "openreview", "code_url": "https://github.com/nxphi47/multiagent_crosstranslate", "code_repository": "nxphi47/multiagent_crosstranslate", "code_commit": "041c30d4a67fb1dba309a8cd555bc6d3b40464eb", "code_archive": "repos/K5a_QFEUzA1.zip", "code_archive_sha256": "1c45bfa85ceaf031a9cabacb402d40fa6b7a630318a0b989c8e4aa71fc2054f0", "code_archive_bytes": 968363, "code_file_count": 52, "code_extensions": {".py": 38, ".sh": 14}, "github_disk_usage_kb": 952, "github_languages": {"Python": 694863, "Shell": 57470, "Perl": 5233}, "github_archived": false, "github_pushed_at": "2023-06-16T12:32:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-agent-cross-translated-diversification"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1l66nNFvB", "year": 2020, "status": "rejected", "title": "Graph Warp Module: an Auxiliary Module for Boosting the Power of Graph Neural Networks in Molecular Graph Analysis", "authors": ["Katsuhiko Ishiguro", "Shin-ichi Maeda", "Masanori Koyama"], "authorids": ["k.ishiguro.jp@ieee.org", "ichi@preferred.jp", "masomatics@preferred.jp"], "authors_source": "OpenReview API", "abstract": "Graph Neural Network (GNN) is a popular architecture for the analysis of chemical molecules, and it has numerous applications in material and medicinal science.\nCurrent lines of GNNs developed for molecular analysis, however, do not fit well on the training set, and their performance does not scale well with the complexity of the network. \nIn this paper, we propose an auxiliary module to be attached to a GNN that can boost the representation power of the model without hindering the original GNN architecture. \nOur auxiliary module can improve the representation power and the generalization ability of a wide variety of GNNs, including those that are used commonly in biochemical applications. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Hkg526KzqB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper243/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Graph Neural Networks is a popular architecture for the analysis of chemical molecules. The authors propose an auxiliary module that can be attached to a GNN that can boost the representation power of GNNs. The auxiliary module has three building blocks: 1. a supernode, 2. a transmitter unit and 3. a warp gate unit. The authors show through carefully designed experiments that these additions can be attached to any type of GNN and that they are successful in reducing both the training error and test error. A variety of graph regression and graph classification tasks are chosen to show the efficacy of the method.\n\nThe paper is well written and easy to follow. The modification suggested by the authors is novel and useful. Experiments are well designed. An aspect that is not clear from the paper is the ability of these models to overfit the data as you increase the representation power of the network. While the authors claim that to be one of the shortcomings of existing GNNs, it is not clear whether the proposed method solves that problem. For example, in figure 4. the training loss hardly decreases as the number of layers are increased. It would be good if the authors can share any insights on this point. \n\nOverall, I think this is a good paper that the community will benefit from.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Graph Neural Networks is a popular architecture for the analysis of chemical molecules. The authors propose an auxiliary module that can be attached to a GNN that can boost the representation power of GNNs. The auxiliary module has three building blocks: 1. a supernode, 2. a transmitter unit and 3. a warp gate unit. The authors show through carefully designed experiments that these additions can be attached to any type of GNN and that they are successful in reducing both the training error and test error. A variety of graph regression and graph classification tasks are chosen to show the efficacy of the method.\n\nThe paper is well written and easy to follow. The modification suggested by the authors is novel and useful. Experiments are well designed. An aspect that is not clear from the paper is the ability of these models to overfit the data as you increase the representation power of the network. While the authors claim that to be one of the shortcomings of existing GNNs, it is not clear whether the proposed method solves that problem. For example, in figure 4. the training loss hardly decreases as the number of layers are increased. It would be good if the authors can share any insights on this point. \n\nOverall, I think this is a good paper that the community will benefit from."}, "tcdate": 1572146609839}, {"id": "HJeaT6S-5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper243/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an auxiliary module for GNNs to boost the representation power. The new module consists of virtual supernode, attention unit, and gating unit, each of which is demonstrated useful in the experiments. The module can be applied to various types of GNNs. \n\nThis work can be seen as an improvement to previous virtual supernode based methods. Adding the attention units and gating units is rational, and the effectiveness is also proved in the ablation studies. However, the claimed contribution of improving the representation power may mainly come from the idea of supernodes (instead of the attention and gating). This largely reduces the novelty of this paper and make it incremental, because using virtual supernodes is not this paper’s original idea.\n\nThe paper is generally well written. However, the comparison with previous supernode based models is not described clearly enough. The authors listed the difference from (Glimer et al. 2017) and (Li et al. 2017) in Table 1, but ignored (Pham et al. 2017) and (Battaglia et al. 2018), which were also cited in the related work. Moreover, (Li et al. 2017)’s method is actually different from the simple supernode baseline, in that it is not a bidirectional message passing between supernode and the main network. Table 1 does not contain this property.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "The paper proposes an auxiliary module for GNNs to boost the representation power. The new module consists of virtual supernode, attention unit, and gating unit, each of which is demonstrated useful in the experiments. The module can be applied to various types of GNNs. \n\nThis work can be seen as an improvement to previous virtual supernode based methods. Adding the attention units and gating units is rational, and the effectiveness is also proved in the ablation studies. However, the claimed contribution of improving the representation power may mainly come from the idea of supernodes (instead of the attention and gating). This largely reduces the novelty of this paper and make it incremental, because using virtual supernodes is not this paper’s original idea.\n\nThe paper is generally well written. However, the comparison with previous supernode based models is not described clearly enough. The authors listed the difference from (Glimer et al. 2017) and (Li et al. 2017) in Table 1, but ignored (Pham et al. 2017) and (Battaglia et al. 2018), which were also cited in the related work. Moreover, (Li et al. 2017)’s method is actually different from the simple supernode baseline, in that it is not a bidirectional message passing between supernode and the main network. Table 1 does not contain this property.\n"}, "tcdate": 1572064709236}, {"id": "rJlciCPDtS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper243/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "= Summary\nA Graph Neural Network extension integrating a global supernode explicitly into the message passing process. Concretely, message passing along the graph edges is alternated with message passing to/from a fresh super-node. Experiments show that this improves results of a number of common GNN architectures on four datasets.\n\n= Strong/Weak Points\n+ Simple but useful extension of the existing super-node idea\n+ Experiments on a number of datasets and baseline GNN architectures, providing ample experimental evidence of the usefulness of the method.\n- Writing is overcomplicated and uses a lot of jargon (\"transmitter unit\", \"warp gate\", \"intermodule hyperspace\"). I found the text entirely impenetrable and instead simply focused on Fig. 3 + the actual equations.\n\n= Recommendation\nThis is a nice contribution of minor novelty, with empirical evidence of its usefulness. I believe the paper should be accepted to a large conference such as ICLR.\n\n= Minor Comments\n- Fig. 3: Inconsistent \"intra-module\" (top) vs. \"intra module\" (bottom) \n- Concurrent work in https://openreview.net/forum?id=B1lnbRNtwr discusses a \"sandwich\" model which alternates graph message passing with (essentially) a Transformer layer applied to all nodes. This idea seems related (in that it alternates local and global information exchange).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "N/A", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "= Summary\nA Graph Neural Network extension integrating a global supernode explicitly into the message passing process. Concretely, message passing along the graph edges is alternated with message passing to/from a fresh super-node. Experiments show that this improves results of a number of common GNN architectures on four datasets.\n\n= Strong/Weak Points\n+ Simple but useful extension of the existing super-node idea\n+ Experiments on a number of datasets and baseline GNN architectures, providing ample experimental evidence of the usefulness of the method.\n- Writing is overcomplicated and uses a lot of jargon (\"transmitter unit\", \"warp gate\", \"intermodule hyperspace\"). I found the text entirely impenetrable and instead simply focused on Fig. 3 + the actual equations.\n\n= Recommendation\nThis is a nice contribution of minor novelty, with empirical evidence of its usefulness. I believe the paper should be accepted to a large conference such as ICLR.\n\n= Minor Comments\n- Fig. 3: Inconsistent \"intra-module\" (top) vs. \"intra module\" (bottom) \n- Concurrent work in https://openreview.net/forum?id=B1lnbRNtwr discusses a \"sandwich\" model which alternates graph message passing with (essentially) a Transformer layer applied to all nodes. This idea seems related (in that it alternates local and global information exchange)."}, "tcdate": 1571417762091}], "openreview_url": "https://openreview.net/forum?id=S1l66nNFvB", "arxiv_id": "1902.01020", "paper_pdf": "papers/S1l66nNFvB.pdf", "paper_pdf_sha256": "f54de0d1cac26a7fb23921694a47d9230dde32d25190d40429e5b4314e5a5513", "paper_pdf_bytes": 2138141, "paper_pdf_source": "openreview", "code_url": "https://github.com/chainer/chainer-chemistry", "code_repository": "chainer/chainer-chemistry", "code_commit": "efe323aa21f63a815130d673781e7cca1ccb72d2", "code_archive": "repos/S1l66nNFvB.zip", "code_archive_sha256": "b352b724079084a6a79d90372346354c4fedfc6e29dd02812a65b8993fa4c7a0", "code_archive_bytes": 734901, "code_file_count": 316, "code_extensions": {".py": 296, ".sh": 18, ".ipynb": 2}, "github_disk_usage_kb": 2388, "github_languages": {"Python": 1022006, "Dockerfile": 11627, "Shell": 3086}, "github_archived": false, "github_pushed_at": "2023-04-20T05:13:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-warp-module-an-auxiliary-module-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "h1ZEMXxSz1", "year": 2024, "status": "rejected", "title": "Segment Anything Model is a Good Teacher for Local Feature Learning", "authors": ["Jingqian Wu", "Rongtao Xu", "Zach Wood-Doughty", "Changwei Wang"], "authorids": ["~Jingqian_Wu1", "~Rongtao_Xu1", "~Zach_Wood-Doughty1", "~Changwei_Wang2"], "authors_source": "OpenReview API", "abstract": "Local feature detection and description play an important role in many computer vision tasks, which are designed to detect and describe keypoints in \"any scene'' and \"any downstream task''. Data-driven local feature learning methods need to rely on pixel-level correspondence for training, which is challenging to acquire at scale, thus hindering further performance improvement. In this paper, we propose SAMFeat to introduce SAM (segment anything model), a fundamental model trained on 10 million images, as a teacher to guide local feature learning and thus inspire higher performance on limited datasets. First, we construct an auxiliary task of Pixel Semantic Relational Distillation (PSRD) for distilling feature relations with category-agnostic semantic information learned by the SAM encoder into a local feature learning network, hence improving local feature description using semantic discrimination. Second, we develop a technique called Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC), which utilizes semantic groupings derived from SAM as weakly supervised signals to optimize the metric space of local descriptors. Third, we design an Edge Attention Guidance Module (EAGM) to further improve the accuracy of local feature detection and description by prompting the network to pay more attention to the edge region guided by SAM. SAMFeat's performance in experiments conducted on various tasks such as image matching on HPatches, and long-term, extensive visual localization datasets like Aachen Day-Night showcases its superiority over previous local features. The release code is available at supplementary material.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "3iYp7ASN75", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5119/Reviewer_SX6K"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this paper, the authors proposed to leverage SAM as the teacher model for learning local feature detectors and descriptors. The authors argued that the fine-grained masks generated by SAM can provide a good amount of prior information about the image and thus be beneficial to local feature learning. To guide the local feature learning, the authors proposed to extract the information from three aspects: 1) pixel-wise representation relationship; 2) semantic group and 3) edge map from the SAM outputs. Based on this information, the authors developed a distillation objective function to bridge the gap between SAM outputs and the local feature detector. In the experiments, the authors showed that the proposed method of distilling from SAM can improve the performance on HPatches and Aachen V1.1. The ablation also justified the effectiveness of each component in the proposed model.", "review_text": "In this paper, the authors proposed to leverage SAM as the teacher model for learning local feature detectors and descriptors. The authors argued that the fine-grained masks generated by SAM can provide a good amount of prior information about the image and thus be beneficial to local feature learning. To guide the local feature learning, the authors proposed to extract the information from three aspects: 1) pixel-wise representation relationship; 2) semantic group and 3) edge map from the SAM outputs. Based on this information, the authors developed a distillation objective function to bridge the gap between SAM outputs and the local feature detector. In the experiments, the authors showed that the proposed method of distilling from SAM can improve the performance on HPatches and Aachen V1.1. The ablation also justified the effectiveness of each component in the proposed model.", "strengths": "1. The authors proposed to leverage the foundational segmentation model SAM for local feature learning. As highlighted in the paper, this work is the first one that incorporates SAM for local feature learning by distilling the knowledge from SAM.\n\n2. The authors proposed three techniques to transfer the fine-grained image understanding knowledge from SAM to the proposed local feature learning pipeline, which results in a new local feature detector called SAMFeat.\n\n3. The experimental results on two benchmarks HPatches and Aachen demonstrate the effectiveness of the proposed method for local feature detection and matching.", "weaknesses": "1. The proposed method in this work is heuristic and incremental. Though the combination of all three techniques achieves the best performance, it is not clear how each of the heuristic technique improve the local feature learning and further the final performance. I would highly suggest the authors have a deeper study on the proposed techniques on how they are contributing the final performance.\n\n2. It is not clear how much overhead for the training after adding the extra loss functions. For example, if I understand correctly, the relationship distillation seems to be very computational heavy in that the affinity matrix contains quadratic number of entries to the image size. Likewise, the second loss term need to compute the with-group and cross-group similarities for all image regions. Overall, I am afraid that the proposed method is very time consuming during training. \n\n3. The experiments in the paper is not satisfactory. There is not much detailed and deep analysis on the proposed method as I mentioned earlier. Given the two tables Table 1 and 2, it is hard to tell whether the proposed method is fairly compared with previous works. For example, the authors mentioned that they used MTLDesc as the baseline method. In Table 1, the reported MMA@3 is 78.7 for it. However, when it goes to Table 3, the first row shows 75.7 MMA@3. This makes me a bit confused and doubted whether the authors are conducting solid and fair comparisons across the board. More importantly, it is also not clear about the main difference between the proposed method and others, regarding the learning techniques, vision encoder, training regimes, etc. \n\n4. Following the last point, the incorporation of SAM for local feature learning is interesting and valuable. However, it is not clear how the settings for the distillation affect the final performance. For example, the density of grid in SAM for automatical segmentation, the image resolutions, the number of sampling points for the proposed training losses, etc, all of them are not studied, which make the contribution of the work and effectiveness of the proposed method hard to assess.", "questions": "Like I mentioned above, I do want to see a deeper analysis on the proposed techniques to distill SAM knowledge for local feature learning tasks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors proposed to leverage SAM as the teacher model for learning local feature detectors and descriptors. The authors argued that the fine-grained masks generated by SAM can provide a good amount of prior information about the image and thus be beneficial to local feature learning. To guide the local feature learning, the authors proposed to extract the information from three aspects: 1) pixel-wise representation relationship; 2) semantic group and 3) edge map from the SAM outputs. Based on this information, the authors developed a distillation objective function to bridge the gap between SAM outputs and the local feature detector. In the experiments, the authors showed that the proposed method of distilling from SAM can improve the performance on HPatches and Aachen V1.1. The ablation also justified the effectiveness of each component in the proposed model.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The authors proposed to leverage the foundational segmentation model SAM for local feature learning. As highlighted in the paper, this work is the first one that incorporates SAM for local feature learning by distilling the knowledge from SAM.\n\n2. The authors proposed three techniques to transfer the fine-grained image understanding knowledge from SAM to the proposed local feature learning pipeline, which results in a new local feature detector called SAMFeat.\n\n3. The experimental results on two benchmarks HPatches and Aachen demonstrate the effectiveness of the proposed method for local feature detection and matching.", "weaknesses": "1. The proposed method in this work is heuristic and incremental. Though the combination of all three techniques achieves the best performance, it is not clear how each of the heuristic technique improve the local feature learning and further the final performance. I would highly suggest the authors have a deeper study on the proposed techniques on how they are contributing the final performance.\n\n2. It is not clear how much overhead for the training after adding the extra loss functions. For example, if I understand correctly, the relationship distillation seems to be very computational heavy in that the affinity matrix contains quadratic number of entries to the image size. Likewise, the second loss term need to compute the with-group and cross-group similarities for all image regions. Overall, I am afraid that the proposed method is very time consuming during training. \n\n3. The experiments in the paper is not satisfactory. There is not much detailed and deep analysis on the proposed method as I mentioned earlier. Given the two tables Table 1 and 2, it is hard to tell whether the proposed method is fairly compared with previous works. For example, the authors mentioned that they used MTLDesc as the baseline method. In Table 1, the reported MMA@3 is 78.7 for it. However, when it goes to Table 3, the first row shows 75.7 MMA@3. This makes me a bit confused and doubted whether the authors are conducting solid and fair comparisons across the board. More importantly, it is also not clear about the main difference between the proposed method and others, regarding the learning techniques, vision encoder, training regimes, etc. \n\n4. Following the last point, the incorporation of SAM for local feature learning is interesting and valuable. However, it is not clear how the settings for the distillation affect the final performance. For example, the density of grid in SAM for automatical segmentation, the image resolutions, the number of sampling points for the proposed training losses, etc, all of them are not studied, which make the contribution of the work and effectiveness of the proposed method hard to assess.", "questions": "Like I mentioned above, I do want to see a deeper analysis on the proposed techniques to distill SAM knowledge for local feature learning tasks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698907866366}, {"id": "AVyy5IGa5E", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5119/Reviewer_F7AB"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors introduce SAMFeat, leveraging SAM (segment anything model) as a teacher to enhance local feature learning to promote the performance of local feature detection and description. In practice, SAMFeat uses Pixel Semantic Relational Distillation (PSRD) to distill feature relations with category-agnostic semantic information learned by the SAM for improved feature description. In addition, a technique called Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC) is adopted for optimizing the metric space of local descriptors. At last, an Edge Attention Guidance Module (EAGM) enhances detection accuracy by focusing on edge regions guided by SAM. The experimental results demonstrate the effectiveness of the proposed modules.", "review_text": "In this paper, the authors introduce SAMFeat, leveraging SAM (segment anything model) as a teacher to enhance local feature learning to promote the performance of local feature detection and description. In practice, SAMFeat uses Pixel Semantic Relational Distillation (PSRD) to distill feature relations with category-agnostic semantic information learned by the SAM for improved feature description. In addition, a technique called Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC) is adopted for optimizing the metric space of local descriptors. At last, an Edge Attention Guidance Module (EAGM) enhances detection accuracy by focusing on edge regions guided by SAM. The experimental results demonstrate the effectiveness of the proposed modules.", "strengths": "(1) The authors integrate the strengths of existing frameworks, effectively utilizing the SAM foundation model and successfully distilling its knowledge into the network for local descriptor learning. It is a good paper for leveraging the knowledge of large models to enhance domain-specific tasks effectively.\n\n(2)The article is clearly written in most parts, enabling readers to quickly catch up on the core technical points. The proposed approach is quite reasonable.\n\n(3)The experimental results demonstrate significant technical advantages, showing substantial improvements over previous work.", "weaknesses": "(1) My first concern is about the novelty of the paper. It is commendable to leverage SAM to enhance model performance in corresponding tasks. However, acquiring structured information through SAM (PSRD), and using semantic grouping to construct positive and negative samples, thereby introducing contrastive learning, have already been briefly discussed in previous works (SFD2, TPR). From this, the paper is more like an integration of some schemes combined with the SAM  model. Hence, its technical novelty is somewhat weak from my perspective.\n\n(2) In practical applications, is it really necessary to adopt all the components described in the paper? The actual contributions of different losses to model optimization still need further verification. For example, in Table 3, how about the performance without using EAG? In addition, are WCS and EAG really irreplaceable by each other? For instance, could eliminating WCS and adjusting the weight of EAG not affect the final results? This raises concerns about the actual importance of WCS ( the semantic embedding ) in the current framework.\n\n(3) Again, regarding Table 4 in the Appendix, the performance seems quite sensitive to the choice of hyperparameters. Would adjusting the loss weight of EAG also further affect the choice of these hyperparameters? After all, as mentioned in the main article, the features enhanced by EAG are further used for subsequent WCS calculations.\n\n(4) Some technical implementation details are still unclear. For example, for Local Feature Detection, is the enhanced C3 computed first before proceeding to calculate C4? The structural diagram in Figure 2 and the description in the subsection on Local Feature Detection have caused confusion regarding more specific technical details. I hope the authors can polish this part.", "questions": "(1) See the weakness\n\n(2) As mentioned in the introduction, how to obtain a large number of effective annotated samples for local descriptor learning is still very challenging. Could it be possible to demonstrate through experiments that the proposed method can effectively reduce the requirement of the training samples?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors introduce SAMFeat, leveraging SAM (segment anything model) as a teacher to enhance local feature learning to promote the performance of local feature detection and description. In practice, SAMFeat uses Pixel Semantic Relational Distillation (PSRD) to distill feature relations with category-agnostic semantic information learned by the SAM for improved feature description. In addition, a technique called Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC) is adopted for optimizing the metric space of local descriptors. At last, an Edge Attention Guidance Module (EAGM) enhances detection accuracy by focusing on edge regions guided by SAM. The experimental results demonstrate the effectiveness of the proposed modules.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "(1) The authors integrate the strengths of existing frameworks, effectively utilizing the SAM foundation model and successfully distilling its knowledge into the network for local descriptor learning. It is a good paper for leveraging the knowledge of large models to enhance domain-specific tasks effectively.\n\n(2)The article is clearly written in most parts, enabling readers to quickly catch up on the core technical points. The proposed approach is quite reasonable.\n\n(3)The experimental results demonstrate significant technical advantages, showing substantial improvements over previous work.", "weaknesses": "(1) My first concern is about the novelty of the paper. It is commendable to leverage SAM to enhance model performance in corresponding tasks. However, acquiring structured information through SAM (PSRD), and using semantic grouping to construct positive and negative samples, thereby introducing contrastive learning, have already been briefly discussed in previous works (SFD2, TPR). From this, the paper is more like an integration of some schemes combined with the SAM  model. Hence, its technical novelty is somewhat weak from my perspective.\n\n(2) In practical applications, is it really necessary to adopt all the components described in the paper? The actual contributions of different losses to model optimization still need further verification. For example, in Table 3, how about the performance without using EAG? In addition, are WCS and EAG really irreplaceable by each other? For instance, could eliminating WCS and adjusting the weight of EAG not affect the final results? This raises concerns about the actual importance of WCS ( the semantic embedding ) in the current framework.\n\n(3) Again, regarding Table 4 in the Appendix, the performance seems quite sensitive to the choice of hyperparameters. Would adjusting the loss weight of EAG also further affect the choice of these hyperparameters? After all, as mentioned in the main article, the features enhanced by EAG are further used for subsequent WCS calculations.\n\n(4) Some technical implementation details are still unclear. For example, for Local Feature Detection, is the enhanced C3 computed first before proceeding to calculate C4? The structural diagram in Figure 2 and the description in the subsection on Local Feature Detection have caused confusion regarding more specific technical details. I hope the authors can polish this part.", "questions": "(1) See the weakness\n\n(2) As mentioned in the introduction, how to obtain a large number of effective annotated samples for local descriptor learning is still very challenging. Could it be possible to demonstrate through experiments that the proposed method can effectively reduce the requirement of the training samples?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "From my viewpoint, there are no ethical concerns.", "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698749922344}, {"id": "Jwi5QcHeor", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5119/Reviewer_uqrh"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper that proposes a method to utilize Segment Anything(SAM) Feature for local feature detection and description. Three key strategies are introduced:\n\n1. Pixel Semantic Relational Distillation (PSRD): An auxiliary task that enhances local feature descriptions using category-agnostic semantic information from the SAM encoder.\n2. Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC): A technique that employs semantic groupings from SAM as weak supervision to optimize the metric space of local descriptors.\n3. Edge Attention Guidance (EAG): A design strategy that improves the accuracy of local feature detection by directing the network to focus more on edge regions, guided by SAM.\n\nSAMFeat demonstrates good performance in image matching on HPatches, and and visual localization on Aachen V1.1 dataset, compared to previous local features.", "review_text": "This paper that proposes a method to utilize Segment Anything(SAM) Feature for local feature detection and description. Three key strategies are introduced:\n\n1. Pixel Semantic Relational Distillation (PSRD): An auxiliary task that enhances local feature descriptions using category-agnostic semantic information from the SAM encoder.\n2. Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC): A technique that employs semantic groupings from SAM as weak supervision to optimize the metric space of local descriptors.\n3. Edge Attention Guidance (EAG): A design strategy that improves the accuracy of local feature detection by directing the network to focus more on edge regions, guided by SAM.\n\nSAMFeat demonstrates good performance in image matching on HPatches, and and visual localization on Aachen V1.1 dataset, compared to previous local features.", "strengths": "1. This paper exlpored a way to release the power of Segment Anything Model (SAM) for distillation for local features. It shows the potential s of visual foundation models.\n2. Experiment-wise, it reaches state of the art results for image matching for with different on HPatches dataset and visual localization task on Archen V1.1 dataset.\n3. Authors provide open-source code", "weaknesses": "Although I believe in the soundness of the good results that the authors have demonstrated, a major issue that makes me skeptical is whether the contribution and novelty are substantial enough to warrant a full paper. Many of the techniques used in the paper are borrowed from other's implementation. For example, the Pixel Semantic Relational Distillation (PSRD) is to compare two similarity matrix which is a widely used knowledge distillation loss [1]. Then the semantic grouping is from the original SAM implementation. The contrastive loss is also very straightforward. The presentation of the paper lacks clarity in conveying the motivation behind technique used, making the paper seem more like a engineering solution for a specific task rather than a systematic method.\n\n[1] F, Tung, et al. Similarity-Preserving Knowledge Distillation, ICCV19", "questions": "1. Continued from the weakness section, could you provide more intuition of each technique used in the paper? For example, the similarity-preserving style of knowledge distillation is used. Why can we use other distillation techniques? \n2. Is SAM really a good teacher? Some works[1] have shown SAM has worse semantics than other pretrained vision model. Correct me if I am wrong, to my understanding, the proposed framework can be applied any pretrained model with good semantics like DINOv2[2], CLIP[3] and ect. Those framework has been proved to have very good features. What would be the results to distillate from those pretrained models for image matching and visual localization.\n3. One minor question of the use of SAM. SAM can segment different levels of object due to ambiguity. For example, a person can be defined as an object or the cloth of this person can be defined as an object. When applying semantic grouping, how to define such grid propmt for SAM?\n4. Can we directly use SAM feature for such image matching and visual localization task?\n5. For the first line of table 3, how is it different with MTLDesc?\n\n[1] Y,Liu, et al. Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching, arXiv 2023.\n[2] M. Oquab, et al. DINOv2: Learning Robust Visual Features without Supervision, arXiv 2023.\n[3] A, Radford ,et al. Learning Transferable Visual Models From Natural Language Supervision, ICML 2021.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper that proposes a method to utilize Segment Anything(SAM) Feature for local feature detection and description. Three key strategies are introduced:\n\n1. Pixel Semantic Relational Distillation (PSRD): An auxiliary task that enhances local feature descriptions using category-agnostic semantic information from the SAM encoder.\n2. Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC): A technique that employs semantic groupings from SAM as weak supervision to optimize the metric space of local descriptors.\n3. Edge Attention Guidance (EAG): A design strategy that improves the accuracy of local feature detection by directing the network to focus more on edge regions, guided by SAM.\n\nSAMFeat demonstrates good performance in image matching on HPatches, and and visual localization on Aachen V1.1 dataset, compared to previous local features.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. This paper exlpored a way to release the power of Segment Anything Model (SAM) for distillation for local features. It shows the potential s of visual foundation models.\n2. Experiment-wise, it reaches state of the art results for image matching for with different on HPatches dataset and visual localization task on Archen V1.1 dataset.\n3. Authors provide open-source code", "weaknesses": "Although I believe in the soundness of the good results that the authors have demonstrated, a major issue that makes me skeptical is whether the contribution and novelty are substantial enough to warrant a full paper. Many of the techniques used in the paper are borrowed from other's implementation. For example, the Pixel Semantic Relational Distillation (PSRD) is to compare two similarity matrix which is a widely used knowledge distillation loss [1]. Then the semantic grouping is from the original SAM implementation. The contrastive loss is also very straightforward. The presentation of the paper lacks clarity in conveying the motivation behind technique used, making the paper seem more like a engineering solution for a specific task rather than a systematic method.\n\n[1] F, Tung, et al. Similarity-Preserving Knowledge Distillation, ICCV19", "questions": "1. Continued from the weakness section, could you provide more intuition of each technique used in the paper? For example, the similarity-preserving style of knowledge distillation is used. Why can we use other distillation techniques? \n2. Is SAM really a good teacher? Some works[1] have shown SAM has worse semantics than other pretrained vision model. Correct me if I am wrong, to my understanding, the proposed framework can be applied any pretrained model with good semantics like DINOv2[2], CLIP[3] and ect. Those framework has been proved to have very good features. What would be the results to distillate from those pretrained models for image matching and visual localization.\n3. One minor question of the use of SAM. SAM can segment different levels of object due to ambiguity. For example, a person can be defined as an object or the cloth of this person can be defined as an object. When applying semantic grouping, how to define such grid propmt for SAM?\n4. Can we directly use SAM feature for such image matching and visual localization task?\n5. For the first line of table 3, how is it different with MTLDesc?\n\n[1] Y,Liu, et al. Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching, arXiv 2023.\n[2] M. Oquab, et al. DINOv2: Learning Robust Visual Features without Supervision, arXiv 2023.\n[3] A, Radford ,et al. Learning Transferable Visual Models From Natural Language Supervision, ICML 2021.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698526391621}, {"id": "UDiiATmI8W", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5119/Reviewer_933k"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a method called SAMFeat, which leverages a fundamental model called SAM (segment anything model) to improve local feature detection and description. The authors address the challenge of limited pixel-level correspondence data for training local feature learning methods.\n\nTo tackle this, they introduce an auxiliary task called Pixel Semantic Relational Distillation (PSRD). PSRD distills feature relations using category-agnostic semantic information learned by the SAM encoder, enhancing local feature description through improved semantic discrimination.\n\nAdditionally, the authors present Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC). This technique utilizes weakly supervised signals derived from SAM's semantic groupings to optimize the metric space of local descriptors, further enhancing the learning process.\n\nTo improve the accuracy of local feature detection and description, the authors propose an Edge Attention Guidance Module (EAGM). EAGM prompts the network to prioritize attention to the edge region, guided by the SAM model.\n\nExperiments conducted on tasks such as image matching on HPatches and visual localization datasets like Aachen Day-Night demonstrate the superior performance of SAMFeat compared to previous local features.", "review_text": "This paper proposes a method called SAMFeat, which leverages a fundamental model called SAM (segment anything model) to improve local feature detection and description. The authors address the challenge of limited pixel-level correspondence data for training local feature learning methods.\n\nTo tackle this, they introduce an auxiliary task called Pixel Semantic Relational Distillation (PSRD). PSRD distills feature relations using category-agnostic semantic information learned by the SAM encoder, enhancing local feature description through improved semantic discrimination.\n\nAdditionally, the authors present Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC). This technique utilizes weakly supervised signals derived from SAM's semantic groupings to optimize the metric space of local descriptors, further enhancing the learning process.\n\nTo improve the accuracy of local feature detection and description, the authors propose an Edge Attention Guidance Module (EAGM). EAGM prompts the network to prioritize attention to the edge region, guided by the SAM model.\n\nExperiments conducted on tasks such as image matching on HPatches and visual localization datasets like Aachen Day-Night demonstrate the superior performance of SAMFeat compared to previous local features.", "strengths": "+ This paper provides SAMFeat to leverage the SAM model to enhance local feature detection and description. The combination of distilling feature relations using category-agnostic semantic information, weakly supervised contrastive learning based on semantic grouping, and an edge attention guidance module is well-motivated.\n\n+ The application of SAM as a teacher model to guide local feature learning is a creative combination that extends the existing understanding of feature learning approaches.", "weaknesses": "- In the section describing \"Pixel Semantic Relational Distillation,\" further elaboration on the process of obtaining the relationship matrix from SAM would enhance the clarity and academic rigor of the paper. It is important to provide a more detailed explanation of the specific steps involved in acquiring the relationship matrix from SAM. This may include a description of the underlying mechanism used by SAM to capture and encode semantic information, followed by the extraction of relevant features for constructing the relationship matrix.", "questions": "- The authors should consider providing insights into the rationale behind choosing SAM as the source for extracting semantic information. Discussing the unique characteristics of SAM that make it suitable for distilling the semantic relationships within the given context would strengthen the paper's argument. For example, SEEM can also serve as a suitable baseline for addressing the task at hand.\n\n\n\n\n------------ post rebuttal update ----------\n\nThanks for the feedback. After going through the rebuttal, I still think the generalization of the proposed method should be validated on other segment anything models such as SEEM. Therefore, I will keep the original rating.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method called SAMFeat, which leverages a fundamental model called SAM (segment anything model) to improve local feature detection and description. The authors address the challenge of limited pixel-level correspondence data for training local feature learning methods.\n\nTo tackle this, they introduce an auxiliary task called Pixel Semantic Relational Distillation (PSRD). PSRD distills feature relations using category-agnostic semantic information learned by the SAM encoder, enhancing local feature description through improved semantic discrimination.\n\nAdditionally, the authors present Weakly Supervised Contrastive Learning Based on Semantic Grouping (WSC). This technique utilizes weakly supervised signals derived from SAM's semantic groupings to optimize the metric space of local descriptors, further enhancing the learning process.\n\nTo improve the accuracy of local feature detection and description, the authors propose an Edge Attention Guidance Module (EAGM). EAGM prompts the network to prioritize attention to the edge region, guided by the SAM model.\n\nExperiments conducted on tasks such as image matching on HPatches and visual localization datasets like Aachen Day-Night demonstrate the superior performance of SAMFeat compared to previous local features.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "+ This paper provides SAMFeat to leverage the SAM model to enhance local feature detection and description. The combination of distilling feature relations using category-agnostic semantic information, weakly supervised contrastive learning based on semantic grouping, and an edge attention guidance module is well-motivated.\n\n+ The application of SAM as a teacher model to guide local feature learning is a creative combination that extends the existing understanding of feature learning approaches.", "weaknesses": "- In the section describing \"Pixel Semantic Relational Distillation,\" further elaboration on the process of obtaining the relationship matrix from SAM would enhance the clarity and academic rigor of the paper. It is important to provide a more detailed explanation of the specific steps involved in acquiring the relationship matrix from SAM. This may include a description of the underlying mechanism used by SAM to capture and encode semantic information, followed by the extraction of relevant features for constructing the relationship matrix.", "questions": "- The authors should consider providing insights into the rationale behind choosing SAM as the source for extracting semantic information. Discussing the unique characteristics of SAM that make it suitable for distilling the semantic relationships within the given context would strengthen the paper's argument. For example, SEEM can also serve as a suitable baseline for addressing the task at hand.\n\n\n\n\n------------ post rebuttal update ----------\n\nThanks for the feedback. After going through the rebuttal, I still think the generalization of the proposed method should be validated on other segment anything models such as SEEM. Therefore, I will keep the original rating.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698026786165}], "openreview_url": "https://openreview.net/forum?id=h1ZEMXxSz1", "arxiv_id": "2309.16992", "paper_pdf": "papers/h1ZEMXxSz1.pdf", "paper_pdf_sha256": "464da62bd874d158b39e919df74dd9f7a68a1407b9f7fdcf7ac0f997b66e7138", "paper_pdf_bytes": 2676790, "paper_pdf_source": "openreview", "code_url": "https://github.com/vignywang/SAMFeat", "code_repository": "vignywang/SAMFeat", "code_commit": "38be900236c1506db6c0c82018ba637b6b6eaea8", "code_archive": "repos/h1ZEMXxSz1.zip", "code_archive_sha256": "bc355f6f9f6c180945598cd55a7d25db87a222490387e2f9fa552f97760ecb36", "code_archive_bytes": 85886, "code_file_count": 22, "code_extensions": {".py": 21, ".sh": 1}, "github_disk_usage_kb": 141, "github_languages": {"Python": 281271, "Shell": 364}, "github_archived": false, "github_pushed_at": "2025-06-08T15:17:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/segment-anything-model-is-a-good-teacher-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7dmsy2Vd5h", "year": 2025, "status": "rejected", "title": "Comparing and Contrasting Deep Learning Weather Prediction Backbones on Navier-Stokes and Atmospheric Dynamics", "authors": ["Matthias Karlbauer", "Danielle C. Maddix", "Abdul Fatir Ansari", "Boran Han", "Gaurav Gupta", "Bernie Wang", "Andrew Stuart", "Michael W. Mahoney"], "authorids": ["~Matthias_Karlbauer1", "~Danielle_C._Maddix1", "~Abdul_Fatir_Ansari2", "~Boran_Han1", "~Gaurav_Gupta2", "~Bernie_Wang1", "~Andrew_Stuart2", "~Michael_W._Mahoney1"], "authors_source": "OpenReview API", "abstract": "Remarkable progress in the development of Deep Learning Weather Prediction (DLWP) models positions  them  to become competitive with traditional numerical weather prediction (NWP) models. Indeed, a wide number of DLWP architectures---based on various backbones, including U-Net, Transformer, Graph Neural Network (GNN), and Fourier Neural Operator (FNO)---have demonstrated their potential at forecasting atmospheric states. However, due to differences in training protocols, forecast horizons, and data choices, it remains unclear which (if any) of these methods and architectures are most suitable for weather forecasting and for future model development. Here, we step back and provide a detailed empirical analysis, under controlled conditions, comparing and contrasting the most prominent DLWP models, along with their backbones. We accomplish this by predicting  synthetic two-dimensional incompressible Navier-Stokes and real-world global weather dynamics. In terms of accuracy, memory consumption, and runtime, our results illustrate various tradeoffs. For example, on synthetic data, we observe favorable performance of FNO; and on the real-world WeatherBench dataset, our results demonstrate the suitability of ConvLSTM and SwinTransformer for short-to-mid-ranged forecasts. For long-ranged weather rollouts of up to 365 days, we observe superior stability and physical soundness in architectures that formulate a spherical data representation, i.e., GraphCast and Spherical FNO. In addition, we observe that all of these model backbones ``saturate,'' i.e., none of them exhibit so-called neural scaling, which highlights an important direction for future work on these and related models. The code is available at \\url{https://anonymous.4open.science/r/dlwp-benchmark-F88C}.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "HbXbaFmQYi", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7984/Reviewer_BnFD"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The paper aims to conduct a comprehensive evaluation and comparisons of deep learning backbones for weather forecasting. Authors selected seven widely used networks and conducted a large number of experiments on both synthetic dataset and a low resolution Weatherbench dataset.", "review_text": "The paper aims to conduct a comprehensive evaluation and comparisons of deep learning backbones for weather forecasting. Authors selected seven widely used networks and conducted a large number of experiments on both synthetic dataset and a low resolution Weatherbench dataset.", "strengths": "Pros:\n\n1. Fair and comprehensive evaluations of the influence of network backbones to DLWP is important. \n2. Apart from different backbones, authors also evaluate the influence of parameter numbers, which would be informative for exploring the parameter scaling law for DLWP.", "weaknesses": "Cons:\n\n1. According to my experiments, tuning the parameters for weather prediction (e.g., on weatherbeach) is hard and can cause significantly different results, which hampers the reliability of the results. \n2. According to my experiments, different models have different rates of convergence, which is not considered and analysized in the paper and further hampers the reliability of the results. \n3. In table 1, the models saturated easily, which is not consistent with existing weather models that have more than 1B parameters. I would suggest the authors to explore model techniques to save the memory.\n4. Some important works in the field are not considered and discussed, such as FengWu, FengWu-GHR, and Stormer.", "questions": "please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to conduct a comprehensive evaluation and comparisons of deep learning backbones for weather forecasting. Authors selected seven widely used networks and conducted a large number of experiments on both synthetic dataset and a low resolution Weatherbench dataset.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Pros:\n\n1. Fair and comprehensive evaluations of the influence of network backbones to DLWP is important. \n2. Apart from different backbones, authors also evaluate the influence of parameter numbers, which would be informative for exploring the parameter scaling law for DLWP.", "weaknesses": "Cons:\n\n1. According to my experiments, tuning the parameters for weather prediction (e.g., on weatherbeach) is hard and can cause significantly different results, which hampers the reliability of the results. \n2. According to my experiments, different models have different rates of convergence, which is not considered and analysized in the paper and further hampers the reliability of the results. \n3. In table 1, the models saturated easily, which is not consistent with existing weather models that have more than 1B parameters. I would suggest the authors to explore model techniques to save the memory.\n4. Some important works in the field are not considered and discussed, such as FengWu, FengWu-GHR, and Stormer.", "questions": "please refer to the weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731169857626}, {"id": "gOHlbPXgMW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7984/Reviewer_F4LR"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper provides a comparative study of various architectures such as U-Net, Transformers, Graph neural networks, ConvLSTM, and Neural Operators that have shown their potential to serve as backbones in Deep Learning Weather Prediction (DLWP) models. This work includes a systematic and detailed empirical analysis under controlled conditions, controlling for parameter count, training protocol, and prognostic variables. All the models are evaluated by benchmarking on two systems: synthetic Navier-Stokes and real-world weather datasets. The paper focuses on short-to-mid-ranged forecasts, long-ranged (climate length) forecasts, and physics-backed forecasts, intending to provide better architectural design choices supporting the DLWP research for various forecasting tasks. Based on their observation, ConvLSTM is better at short and mid-range forecasts on weather data. For stable long-ranged forecasts, aligned with physics principles, spherical representations such as in GraphCast and Spherical FNO, show superior performance.", "review_text": "The paper provides a comparative study of various architectures such as U-Net, Transformers, Graph neural networks, ConvLSTM, and Neural Operators that have shown their potential to serve as backbones in Deep Learning Weather Prediction (DLWP) models. This work includes a systematic and detailed empirical analysis under controlled conditions, controlling for parameter count, training protocol, and prognostic variables. All the models are evaluated by benchmarking on two systems: synthetic Navier-Stokes and real-world weather datasets. The paper focuses on short-to-mid-ranged forecasts, long-ranged (climate length) forecasts, and physics-backed forecasts, intending to provide better architectural design choices supporting the DLWP research for various forecasting tasks. Based on their observation, ConvLSTM is better at short and mid-range forecasts on weather data. For stable long-ranged forecasts, aligned with physics principles, spherical representations such as in GraphCast and Spherical FNO, show superior performance.", "strengths": "- The experiments in this study are extensive, and the analysis is presented in a clear, organized manner. The details of the experiments are thoroughly explained and the set of models chosen for the comparison is justified well.\n- The paper includes long-range forecasts for lead times as long as 365 days (and more), which is important and not included in most DLWP studies. These results can be insightful to this line of research.", "weaknesses": "- The spatial resolution used in the paper for global weather prediction is too coarse (5.625 degrees) as compared to the 0.25-degree resolution used in recent weather forecasting models such as FourCastNet, PanguWeather, and GraphCast. \n- Using the backbones of the DLWP models for performing prediction on the Navier-stokes system does not seem very relevant to the contributions of this work. The paper also says “A direct transfer of the results from Navier-Stokes to weather dynamics is limited”. Moreover, FNO working so well on Navier-Stokes has already been shown before. \n- The paper claims to be studying physically meaningful forecasts. This is a crucial aspect of weather forecasting and should be a critical factor in comparing models. However, the paper doesn’t go into much detail on this aspect. For instance, physics-based metrics and power spectrum plots [1] are needed to investigate if the models can capture small-scale (high-freqeuncy) features in their forecasts. \n[1] Nathaniel, Juan, et al. \"Chaosbench: A multi-channel, physics-based benchmark for subseasonal-to-seasonal climate prediction.\" 2024.", "questions": "- What is the justification behind using a coarse spatial resolution for weather prediction?\n- The authors should add more on why they chose to evaluate and compare the models on the Navier-Stokes system. \n- There needs to be more analysis to understand the physical soundness of various models. This should include physics-based plots/metrics as suggested before, and a discussion comparing models on this aspect of their forecasting skill.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper provides a comparative study of various architectures such as U-Net, Transformers, Graph neural networks, ConvLSTM, and Neural Operators that have shown their potential to serve as backbones in Deep Learning Weather Prediction (DLWP) models. This work includes a systematic and detailed empirical analysis under controlled conditions, controlling for parameter count, training protocol, and prognostic variables. All the models are evaluated by benchmarking on two systems: synthetic Navier-Stokes and real-world weather datasets. The paper focuses on short-to-mid-ranged forecasts, long-ranged (climate length) forecasts, and physics-backed forecasts, intending to provide better architectural design choices supporting the DLWP research for various forecasting tasks. Based on their observation, ConvLSTM is better at short and mid-range forecasts on weather data. For stable long-ranged forecasts, aligned with physics principles, spherical representations such as in GraphCast and Spherical FNO, show superior performance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The experiments in this study are extensive, and the analysis is presented in a clear, organized manner. The details of the experiments are thoroughly explained and the set of models chosen for the comparison is justified well.\n- The paper includes long-range forecasts for lead times as long as 365 days (and more), which is important and not included in most DLWP studies. These results can be insightful to this line of research.", "weaknesses": "- The spatial resolution used in the paper for global weather prediction is too coarse (5.625 degrees) as compared to the 0.25-degree resolution used in recent weather forecasting models such as FourCastNet, PanguWeather, and GraphCast. \n- Using the backbones of the DLWP models for performing prediction on the Navier-stokes system does not seem very relevant to the contributions of this work. The paper also says “A direct transfer of the results from Navier-Stokes to weather dynamics is limited”. Moreover, FNO working so well on Navier-Stokes has already been shown before. \n- The paper claims to be studying physically meaningful forecasts. This is a crucial aspect of weather forecasting and should be a critical factor in comparing models. However, the paper doesn’t go into much detail on this aspect. For instance, physics-based metrics and power spectrum plots [1] are needed to investigate if the models can capture small-scale (high-freqeuncy) features in their forecasts. \n[1] Nathaniel, Juan, et al. \"Chaosbench: A multi-channel, physics-based benchmark for subseasonal-to-seasonal climate prediction.\" 2024.", "questions": "- What is the justification behind using a coarse spatial resolution for weather prediction?\n- The authors should add more on why they chose to evaluate and compare the models on the Navier-Stokes system. \n- There needs to be more analysis to understand the physical soundness of various models. This should include physics-based plots/metrics as suggested before, and a discussion comparing models on this aspect of their forecasting skill.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730668396175}, {"id": "lcCUV1zb1a", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7984/Reviewer_ijdW"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper analyzes the performance of different network structures for weather prediction. Experiments were conducted on both synthetic and real data, and a benchmark was established. They also suggested network structures suitable for mid-term and long-term forecasting.", "review_text": "The paper analyzes the performance of different network structures for weather prediction. Experiments were conducted on both synthetic and real data, and a benchmark was established. They also suggested network structures suitable for mid-term and long-term forecasting.", "strengths": "The design of the backbone network greatly affects the performance of machine learning models. This article provides an analysis of the backbone network's performance in weather forecasting.", "weaknesses": "This article seems like an experimental report. It includes introductions to several classic backbone networks, settings for two experimental datasets, and descriptions of the results. However, this paper lacks insights that previous work did not reveal.", "questions": "1. The authors introduced a new benchmark for weather forecasting, but they didn't clearly explain how it differs from previous research, such as in data construction and task definition.\n2. The authors analyzed several backbone networks, but only showed some quantitative results without providing more insights, such as proposing new designs for backbone networks.\n3. The authors used synthetic and real data to train these models, but they did not discuss the differences between these data and the data used by existing state-of-the-art models.\n4. The number of model parameters used by the authors seems small, but current weather prediction models use a large number of parameters. With such a big difference in parameter count, is the conclusion reliable?\n5. With only 1K and 10K samples in experiments 1 and 2, are these numbers too small? Can the conclusions be trusted?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper analyzes the performance of different network structures for weather prediction. Experiments were conducted on both synthetic and real data, and a benchmark was established. They also suggested network structures suitable for mid-term and long-term forecasting.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The design of the backbone network greatly affects the performance of machine learning models. This article provides an analysis of the backbone network's performance in weather forecasting.", "weaknesses": "This article seems like an experimental report. It includes introductions to several classic backbone networks, settings for two experimental datasets, and descriptions of the results. However, this paper lacks insights that previous work did not reveal.", "questions": "1. The authors introduced a new benchmark for weather forecasting, but they didn't clearly explain how it differs from previous research, such as in data construction and task definition.\n2. The authors analyzed several backbone networks, but only showed some quantitative results without providing more insights, such as proposing new designs for backbone networks.\n3. The authors used synthetic and real data to train these models, but they did not discuss the differences between these data and the data used by existing state-of-the-art models.\n4. The number of model parameters used by the authors seems small, but current weather prediction models use a large number of parameters. With such a big difference in parameter count, is the conclusion reliable?\n5. With only 1K and 10K samples in experiments 1 and 2, are these numbers too small? Can the conclusions be trusted?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730555658913}, {"id": "5FDOdZJi83", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7984/Reviewer_3TqW"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper provides a fair comparison of the performance of widely used deep learning models for weather prediction. The authors standardize parametric settings, inputs, outputs, and training methods across models, and evaluate their performance using Navier-Stokes dynamics simulations, as well as medium- and long-term weather prediction. The study highlights each model's strengths and weaknesses.", "review_text": "This paper provides a fair comparison of the performance of widely used deep learning models for weather prediction. The authors standardize parametric settings, inputs, outputs, and training methods across models, and evaluate their performance using Navier-Stokes dynamics simulations, as well as medium- and long-term weather prediction. The study highlights each model's strengths and weaknesses.", "strengths": "This paper addresses a significant gap in the field, as there is currently no comprehensive and fair comparison of DLWP models. While many studies claim superior performance for their models, it remains unclear whether this is due to the backbone architecture, diagnostic variables, training, or inference strategies. By focusing specifically on the backbone models, the authors conduct rigorous experiments to empirically assess their forecasting potential. The findings are valuable for the DLWP field, offering novel insights, such as the superior performance of ConvLSTM, differences between Pangu-Weather and SwinTransformer, and the influence of FourCastNet's patch size.", "weaknesses": "1. The experiment of synthetic Navier-Stokes simulations seem to play a limited role. As noted in line 160, there is a significant gap between the univariate Navier-Stokes simulation and real atmospheric dynamics. Given that the authors aim to evaluate backbone models' performance in the more complex weather forecasting task in section 3.2, dedicating one-third of the main text to the simpler univariate Navier-Stokes simulation seems unnecessary. In light of the results from section 3.2, the findings from section 3.1 appear less relevant and, in fact, somewhat confusing:\n    \n    (a)  Section 3.1 highlights the superiority of TFNO, yet this model is absent from section 3.2. Since TFNO appears in Figures 13, 14, and 15, its performance in RMSE metric should have been assessed by the authors.\n    \n    (b) In Figures 13, 14, and 15, TFNO, ConvLSTM, and UNet underperform compared to other models.\n    \n    (c) In Figure 16, models that perform well in section 3.1 (TFNO, ConvLSTM, UNet) exhibit poor stability.\n\n2. In section 3.2, the authors evaluate the performance of different backbone models using the 5.625 deg ERA5 dataset. The experiments provide limited guidance for selecting backbone models for operational weather prediction, which typically relies on the 0.25 deg ERA5 dataset. However, large-scale experiments at this resolution are obviously costly, so this is an existential but understandable drawback :) .", "questions": "Overall, as the first paper to provide a fair comparison of various DLWP models, this work has the potential to make a significant contribution to the field. However, I recommend the authors reconsider the emphasis placed on section 3.1 and expand the experimental results in section 3.2, particularly by including RMSE metrics for variables such as t850, u10, and v10.\n\nHere are other questions:\n1. Lines 291-293: The ACC metric is not provided in section 3.2.1. Additionally, section 3.2.1 presents RMSE metrics for geopotential only up to 7 days, not 14. Given the presence of 8 prognostic variables, it would be beneficial to include the RMSE and ACC metrics for all variables, potentially in the appendix in a format similar to Figure 2. \n2. In Figures 2 and 18, the authors observe that SwinTransformer outperforms Pangu-Weather in terms of RMSE. To my knowledge, the primary difference between Pangu-Weather and SwinTransformer is the use of Earth-specific positional bias in Pangu-Weather. Intuitively, this difference alone should not lead to such a performance gap. I suggest that the authors standardize other hyperparameters (e.g., layers, embedding dimensions) between Pangu-Weather and SwinTransformer, and present additional results, such as RMSE for geopotential with 1M parameters, to clarify this discrepancy.\n3. Lines 339-340: The authors limit the optimization cycle to 24 hours (4 steps). While there is no established standard for optimization lead time, I question whether ConvLSTM, being the only RNN-based model, is particularly sensitive to this hyperparameter. State-of-the-art models like FourCastNet (2 steps), Pangu-Weather (1 step), and GraphCast (1 step in pretraining) use shorter optimization cycles. Training with 4 steps may become resource-intensive at higher spatial resolutions, which could be a limitation of ConvLSTM.\n4. Lines 340-341: The authors evaluate the backbone models using initial conditions at 00z. I wonder if fixing the initial time at 00z simplifies the overall weather prediction task. Could the authors test whether models trained on 00z initial conditions also perform well with 12z initial conditions in the test set?\n5. Lines 489-490: In Figures 5 and 15, the authors note that SFNO performs well in predicting wind fields, accurately capturing real-world wind patterns. They attribute this to SFNO's adherence to physical principles. However, given SFNO's performance in Figure 2, I question whether this claim holds true for all prognostic variables.\n6. Line 863: Since $x,y \\in \\mathbb{N}$, it follows that $x+y \\in \\mathbb{N}$. Therefore, in the authors’ setting, $f \\equiv 0.1$. I think there must be some mistake. Otherwise, the Navier-Stokes simulation is too simple.\n7. Line 1228-1229: The ‘]’ of heads in layers in Pangu-Weather is missing.\n8. In Figure 14, why are the only two graph neural networks smoothed in Zonally averaged forecasts?\n9. In Figures 2, 17, 18, and 19, I observe that the confidence intervals for some models, particularly FourCastNet, are notably wide across the three random seeds. Upon reviewing the code, I suspect this may be due to the gradient clipping, which is set equal to the learning rate ($\\leq 10^{-3}$). When multiplied by the learning rate, the step size of the gradient descent ($||\\eta *\\text{Clip}(\\nabla f)||_{2}$) is less than $1\\times 10^{-6}$, which is likely too small for effective exploration of the parameter landscape. As a result, model performance may be highly dependent on initial parameters or random seeds. My question is, why was the gradient clipping value set equal to the learning rate? Is there a specific reference for this choice?\n10. Line 684-686: unify the reference.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper provides a fair comparison of the performance of widely used deep learning models for weather prediction. The authors standardize parametric settings, inputs, outputs, and training methods across models, and evaluate their performance using Navier-Stokes dynamics simulations, as well as medium- and long-term weather prediction. The study highlights each model's strengths and weaknesses.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "This paper addresses a significant gap in the field, as there is currently no comprehensive and fair comparison of DLWP models. While many studies claim superior performance for their models, it remains unclear whether this is due to the backbone architecture, diagnostic variables, training, or inference strategies. By focusing specifically on the backbone models, the authors conduct rigorous experiments to empirically assess their forecasting potential. The findings are valuable for the DLWP field, offering novel insights, such as the superior performance of ConvLSTM, differences between Pangu-Weather and SwinTransformer, and the influence of FourCastNet's patch size.", "weaknesses": "1. The experiment of synthetic Navier-Stokes simulations seem to play a limited role. As noted in line 160, there is a significant gap between the univariate Navier-Stokes simulation and real atmospheric dynamics. Given that the authors aim to evaluate backbone models' performance in the more complex weather forecasting task in section 3.2, dedicating one-third of the main text to the simpler univariate Navier-Stokes simulation seems unnecessary. In light of the results from section 3.2, the findings from section 3.1 appear less relevant and, in fact, somewhat confusing:\n    \n    (a)  Section 3.1 highlights the superiority of TFNO, yet this model is absent from section 3.2. Since TFNO appears in Figures 13, 14, and 15, its performance in RMSE metric should have been assessed by the authors.\n    \n    (b) In Figures 13, 14, and 15, TFNO, ConvLSTM, and UNet underperform compared to other models.\n    \n    (c) In Figure 16, models that perform well in section 3.1 (TFNO, ConvLSTM, UNet) exhibit poor stability.\n\n2. In section 3.2, the authors evaluate the performance of different backbone models using the 5.625 deg ERA5 dataset. The experiments provide limited guidance for selecting backbone models for operational weather prediction, which typically relies on the 0.25 deg ERA5 dataset. However, large-scale experiments at this resolution are obviously costly, so this is an existential but understandable drawback :) .", "questions": "Overall, as the first paper to provide a fair comparison of various DLWP models, this work has the potential to make a significant contribution to the field. However, I recommend the authors reconsider the emphasis placed on section 3.1 and expand the experimental results in section 3.2, particularly by including RMSE metrics for variables such as t850, u10, and v10.\n\nHere are other questions:\n1. Lines 291-293: The ACC metric is not provided in section 3.2.1. Additionally, section 3.2.1 presents RMSE metrics for geopotential only up to 7 days, not 14. Given the presence of 8 prognostic variables, it would be beneficial to include the RMSE and ACC metrics for all variables, potentially in the appendix in a format similar to Figure 2. \n2. In Figures 2 and 18, the authors observe that SwinTransformer outperforms Pangu-Weather in terms of RMSE. To my knowledge, the primary difference between Pangu-Weather and SwinTransformer is the use of Earth-specific positional bias in Pangu-Weather. Intuitively, this difference alone should not lead to such a performance gap. I suggest that the authors standardize other hyperparameters (e.g., layers, embedding dimensions) between Pangu-Weather and SwinTransformer, and present additional results, such as RMSE for geopotential with 1M parameters, to clarify this discrepancy.\n3. Lines 339-340: The authors limit the optimization cycle to 24 hours (4 steps). While there is no established standard for optimization lead time, I question whether ConvLSTM, being the only RNN-based model, is particularly sensitive to this hyperparameter. State-of-the-art models like FourCastNet (2 steps), Pangu-Weather (1 step), and GraphCast (1 step in pretraining) use shorter optimization cycles. Training with 4 steps may become resource-intensive at higher spatial resolutions, which could be a limitation of ConvLSTM.\n4. Lines 340-341: The authors evaluate the backbone models using initial conditions at 00z. I wonder if fixing the initial time at 00z simplifies the overall weather prediction task. Could the authors test whether models trained on 00z initial conditions also perform well with 12z initial conditions in the test set?\n5. Lines 489-490: In Figures 5 and 15, the authors note that SFNO performs well in predicting wind fields, accurately capturing real-world wind patterns. They attribute this to SFNO's adherence to physical principles. However, given SFNO's performance in Figure 2, I question whether this claim holds true for all prognostic variables.\n6. Line 863: Since $x,y \\in \\mathbb{N}$, it follows that $x+y \\in \\mathbb{N}$. Therefore, in the authors’ setting, $f \\equiv 0.1$. I think there must be some mistake. Otherwise, the Navier-Stokes simulation is too simple.\n7. Line 1228-1229: The ‘]’ of heads in layers in Pangu-Weather is missing.\n8. In Figure 14, why are the only two graph neural networks smoothed in Zonally averaged forecasts?\n9. In Figures 2, 17, 18, and 19, I observe that the confidence intervals for some models, particularly FourCastNet, are notably wide across the three random seeds. Upon reviewing the code, I suspect this may be due to the gradient clipping, which is set equal to the learning rate ($\\leq 10^{-3}$). When multiplied by the learning rate, the step size of the gradient descent ($||\\eta *\\text{Clip}(\\nabla f)||_{2}$) is less than $1\\times 10^{-6}$, which is likely too small for effective exploration of the parameter landscape. As a result, model performance may be highly dependent on initial parameters or random seeds. My question is, why was the gradient clipping value set equal to the learning rate? Is there a specific reference for this choice?\n10. Line 684-686: unify the reference.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729602067908}], "openreview_url": "https://openreview.net/forum?id=7dmsy2Vd5h", "arxiv_id": "2407.14129", "paper_pdf": "papers/7dmsy2Vd5h.pdf", "paper_pdf_sha256": "a6665ffde5a288a16346d80b4488daa7088ed1bbd077abfaeb94ce839f0823f9", "paper_pdf_bytes": 12397055, "paper_pdf_source": "openreview", "code_url": "https://github.com/amazon-science/dlwp-benchmark", "code_repository": "amazon-science/dlwp-benchmark", "code_commit": "b935e10b46a84fa7d012a0afef4ae2b4e32110c8", "code_archive": "repos/7dmsy2Vd5h.zip", "code_archive_sha256": "476ff0c28d25c2840c4efef1d725dcc3ebf4ab4e2025564c7d94437d57cbd9ff", "code_archive_bytes": 336166, "code_file_count": 105, "code_extensions": {".py": 105}, "github_disk_usage_kb": 204, "github_languages": {"Python": 995937}, "github_archived": false, "github_pushed_at": "2024-08-09T09:00:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/comparing-and-contrasting-deep-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "kkQSwtx0p3", "year": 2024, "status": "rejected", "title": "Leveraging Task Structures for Improved Identifiability in Neural Network Representations", "authors": ["Wenlin Chen", "Julien Horwood", "Juyeon Heo", "José Miguel Hernández-Lobato"], "authorids": ["~Wenlin_Chen2", "~Julien_Horwood1", "~Juyeon_Heo1", "~José_Miguel_Hernández-Lobato1"], "authors_source": "OpenReview API", "abstract": "This work extends the theory of identifiability in supervised learning by considering the consequences of having access to a distribution of tasks. In such cases, we show that identifiability is achievable even in the case of regression, extending prior work restricted to linear identifiability in the single-task classification case. Furthermore, we show that the existence of a task distribution which defines a conditional prior over latent factors reduces the equivalence class for identifiability to permutations and scaling, a much stronger and more useful result than linear identifiability. When we further assume a causal structure over these tasks, our approach enables simple maximum marginal likelihood optimization together with downstream applicability to causal representation learning. Empirically, we validate that our model outperforms more general unsupervised models in recovering canonical representations for both synthetic and real-world molecular data.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zZrfHaeV2t", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission970/Reviewer_2cPA"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper shows that the latent variables of a multi-task regression network can be identified up to an affine transformation. By further assuming a specific causal structure and exploiting changes in the set of variables which are causal vs. spurious, one can identify the latent causal variables more strongly, up to scaling and permutations. Authors illustrate the benefits of this type of identifiable causal representation learning on synthetic and real-world data.", "review_text": "This paper shows that the latent variables of a multi-task regression network can be identified up to an affine transformation. By further assuming a specific causal structure and exploiting changes in the set of variables which are causal vs. spurious, one can identify the latent causal variables more strongly, up to scaling and permutations. Authors illustrate the benefits of this type of identifiable causal representation learning on synthetic and real-world data.", "strengths": "1. With some exceptions discussed below, the paper's main parts are clearly written and in general it's nice and simple presentation\n\n2. Working with Gaussian has allowed the authors to create a nice MLE algorithm that avoids the complications and computational cost of many other approaches\n\n3. Empirical results seem fairly broad and convincing, with the exception of some questions below", "weaknesses": "1. **Main weakness is that it feels like this paper makes very specific assumption that are often unrealistic and restrictive**, for example it feels very restrictive and unrealistic to assume that the causal latents are just zero-mean standard Gaussians and only the indentity of the causal/spurious variables changes ($c_t$). The spurious factor generation has similar issue with being not so general. Yet another assumptions is that of uninformative prior on the weights (but none on the covariances), which feels more like a technical trick that is required. From causal perspective, it feels limiting to assume a linear form for generating $y$ It is thus hard to see what is a naturally suitable application are for this model. Finally, assuming that there is a certain number of latent variables which then switch across tasks on whether they are spurious/causal seems again simplistic (but would love to hear if I'm wrong on this!). \n\n2. **Some of the claims with respect to identifiability results feel exaggerated**: \"*We emphasize that this is a stronger identifiability result than identifiability up to block permutations and scaling (i.e., strong identifiability) as in prior works*\", \"*We emphasize that this strictly strong identifiability class is stronger than the strong identifiability class as in prior works (Khemakhem et al., 2020a; Lu et al., 2022) which are only identifiable up to*\" -- Similarly strong identifiability results have been shown in a general setting in Halva et al. (2021), please see their Theorem 2 in particular and its proof that share similar ideas. Theorem 3, 4 in Halva and Hyvarinen (2020) show strong identifiability in a more specific case, with the similarity here being the exploitation of linear independence and Gaussianity. The authors have cited these works elsewhere and should therefore be also aware of their identifiability results. Further, Morioka et al. (Independent Innovation Analysis, 2021) shows similar identifiability results and proof in their Theorem 2 (though for a different model). Theorem 3.2. and the MTRN identifiability results appear to be essentially the same as those of iVAE (Khemakhem et al.) so I am having hard time understanding whether it has any novel contribution in itself or whether it only sets up the 'second part' with stronger identifiability. Further, the \n\nOther issues:\n- Would be nice to have some references/citations if you make strong statements like this \"This contrasts with current state-of-the-art approaches, whose assumptions also fit our assumed data generating process but which are difficult to train effectively and only identifiable up to block permutations and scaling.\" Which works are these? Why are they difficult to train, and where is that shown? And again last part about identifiability is not even true as shown above.\n- \"... Willetts & Paige (2021); Kivva et al. (2022) recently extend the results in unsupervised generative models to the case of models with mixture model priors.\" Not all these papers assume mixture model priors specifically\n- Given the following statement, you should have included them as baselines: \"concurrent work (Lachapelle et al., 2023; Fumero et al., 2023) has expanded this area of research to consider the multi-task and meta-learning settings.\"\n- Figure 2 appearing before Figure 1 is always confusing...\n- \"...existence of d independent ground-truth...\" please be careful of the word independent and how it's used\n- \"The marginal likelihood of $\\psi$ under MTLCM is given by\" imprecise language -- it's not the likelihood of $\\psi$ as they are not random variables.", "questions": "1. Both $\\mathbf{x}$ and $y$ appear to be deterministic functions of $\\mathbf{z}$, at the same time you seem to assume a noise model e.g. for $p(y\\mid x, t)$ has normal distribution with covariance term $\\sigma_{r, t}^2$. What is the source of this covariance -- please explain whether it's due the 'push-forward' of $\\mathbf{z}$ through the deterministic functions or whether there is some output noise on top of that. Also what is that subscript $r$ on that variance term?\n\n2. Please comment what is the difference between the iVAE identifiability results and the ones here shown form multi-task regression -- they seem very similar.\n\n3. In experiments: Why do you only consider 1 layer MLP for nonlinearity? What if there are more layers? Could you also explain this in more detail: \"we find that the strong MCC for our MTLCM is able to match the weak MCC for the MTRN.\" as well as what you exactly mean by this \"MCC score between the data representations recovered by each pair of those 5 models,\", and also this in more depth \".. weak identifiability achieved from the MTRN implies that identifiability is achievable up to eight latent features, suggesting there may be some redundancies between tasks\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper shows that the latent variables of a multi-task regression network can be identified up to an affine transformation. By further assuming a specific causal structure and exploiting changes in the set of variables which are causal vs. spurious, one can identify the latent causal variables more strongly, up to scaling and permutations. Authors illustrate the benefits of this type of identifiable causal representation learning on synthetic and real-world data.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. With some exceptions discussed below, the paper's main parts are clearly written and in general it's nice and simple presentation\n\n2. Working with Gaussian has allowed the authors to create a nice MLE algorithm that avoids the complications and computational cost of many other approaches\n\n3. Empirical results seem fairly broad and convincing, with the exception of some questions below", "weaknesses": "1. **Main weakness is that it feels like this paper makes very specific assumption that are often unrealistic and restrictive**, for example it feels very restrictive and unrealistic to assume that the causal latents are just zero-mean standard Gaussians and only the indentity of the causal/spurious variables changes ($c_t$). The spurious factor generation has similar issue with being not so general. Yet another assumptions is that of uninformative prior on the weights (but none on the covariances), which feels more like a technical trick that is required. From causal perspective, it feels limiting to assume a linear form for generating $y$ It is thus hard to see what is a naturally suitable application are for this model. Finally, assuming that there is a certain number of latent variables which then switch across tasks on whether they are spurious/causal seems again simplistic (but would love to hear if I'm wrong on this!). \n\n2. **Some of the claims with respect to identifiability results feel exaggerated**: \"*We emphasize that this is a stronger identifiability result than identifiability up to block permutations and scaling (i.e., strong identifiability) as in prior works*\", \"*We emphasize that this strictly strong identifiability class is stronger than the strong identifiability class as in prior works (Khemakhem et al., 2020a; Lu et al., 2022) which are only identifiable up to*\" -- Similarly strong identifiability results have been shown in a general setting in Halva et al. (2021), please see their Theorem 2 in particular and its proof that share similar ideas. Theorem 3, 4 in Halva and Hyvarinen (2020) show strong identifiability in a more specific case, with the similarity here being the exploitation of linear independence and Gaussianity. The authors have cited these works elsewhere and should therefore be also aware of their identifiability results. Further, Morioka et al. (Independent Innovation Analysis, 2021) shows similar identifiability results and proof in their Theorem 2 (though for a different model). Theorem 3.2. and the MTRN identifiability results appear to be essentially the same as those of iVAE (Khemakhem et al.) so I am having hard time understanding whether it has any novel contribution in itself or whether it only sets up the 'second part' with stronger identifiability. Further, the \n\nOther issues:\n- Would be nice to have some references/citations if you make strong statements like this \"This contrasts with current state-of-the-art approaches, whose assumptions also fit our assumed data generating process but which are difficult to train effectively and only identifiable up to block permutations and scaling.\" Which works are these? Why are they difficult to train, and where is that shown? And again last part about identifiability is not even true as shown above.\n- \"... Willetts & Paige (2021); Kivva et al. (2022) recently extend the results in unsupervised generative models to the case of models with mixture model priors.\" Not all these papers assume mixture model priors specifically\n- Given the following statement, you should have included them as baselines: \"concurrent work (Lachapelle et al., 2023; Fumero et al., 2023) has expanded this area of research to consider the multi-task and meta-learning settings.\"\n- Figure 2 appearing before Figure 1 is always confusing...\n- \"...existence of d independent ground-truth...\" please be careful of the word independent and how it's used\n- \"The marginal likelihood of $\\psi$ under MTLCM is given by\" imprecise language -- it's not the likelihood of $\\psi$ as they are not random variables.", "questions": "1. Both $\\mathbf{x}$ and $y$ appear to be deterministic functions of $\\mathbf{z}$, at the same time you seem to assume a noise model e.g. for $p(y\\mid x, t)$ has normal distribution with covariance term $\\sigma_{r, t}^2$. What is the source of this covariance -- please explain whether it's due the 'push-forward' of $\\mathbf{z}$ through the deterministic functions or whether there is some output noise on top of that. Also what is that subscript $r$ on that variance term?\n\n2. Please comment what is the difference between the iVAE identifiability results and the ones here shown form multi-task regression -- they seem very similar.\n\n3. In experiments: Why do you only consider 1 layer MLP for nonlinearity? What if there are more layers? Could you also explain this in more detail: \"we find that the strong MCC for our MTLCM is able to match the weak MCC for the MTRN.\" as well as what you exactly mean by this \"MCC score between the data representations recovered by each pair of those 5 models,\", and also this in more depth \".. weak identifiability achieved from the MTRN implies that identifiability is achievable up to eight latent features, suggesting there may be some redundancies between tasks\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699925349180}, {"id": "YqGt8BEF4u", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission970/Reviewer_NHKJ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work studies causal representation learning for multi-task regression data.  The authors assume a data generating process under which \n\n1. the response variable $y$ is always a linear function of the causal and spurious latent variables $z$, with the linear map varying across tasks,\n2. the observed features $x$ are informative about the latent variables and the tasks are sufficiently diverse,\n3. the conditional distributions $p(y\\mid x, t) = N(y\\mid w_t^\\top f(x), \\sigma_t^2 I)$ and $p(f(x)\\mid z)$ are homoscedastic Gaussian,\n\nand propose a two-stage procedure that can recover the causal latent variables up to permutations and scaling.  The method outperforms previous approaches on synthetic data, and is demonstrated to produce stable outputs on two real-world datasets.", "review_text": "This work studies causal representation learning for multi-task regression data.  The authors assume a data generating process under which \n\n1. the response variable $y$ is always a linear function of the causal and spurious latent variables $z$, with the linear map varying across tasks,\n2. the observed features $x$ are informative about the latent variables and the tasks are sufficiently diverse,\n3. the conditional distributions $p(y\\mid x, t) = N(y\\mid w_t^\\top f(x), \\sigma_t^2 I)$ and $p(f(x)\\mid z)$ are homoscedastic Gaussian,\n\nand propose a two-stage procedure that can recover the causal latent variables up to permutations and scaling.  The method outperforms previous approaches on synthetic data, and is demonstrated to produce stable outputs on two real-world datasets.", "strengths": "- The authors studied an important problem. \n- The proposed method demonstrates promising empirical performance.\n- Compared to recent works on the same problem, the proposed method is more computationally efficient.", "weaknesses": "1. My main concern is the Gaussianity assumptions made throughout the work.  Such assumptions rarely hold in practice; and in contrast to standard machine learning tasks where most methods remain useful even if such assumptions are violated (e.g. typical regression procedures continue to estimate the conditional expectation), for causal representation learning the utility of the proposed method is far less clear.\n\n2. The above concern mostly applies to the second stage of the procedure.  For the first stage it appears that the method only relies on the conditional expectation having the form of $E(y\\mid x) = w_t^\\top f(x)$ (plus bijectivity and task diversity), but in such cases I feel that the assumption is better stated just as in Lachapelle et al (2023), namely we can learn a feature extractor that is linearly equivalent to the ground truth.\n\n3. Regarding experiments, I think it would be more convincing if the authors could include comparisons to the recent works of Lachapelle et al (2023) and Fumero et al (2023), as they addressed the same problem and have been publicly available for 4-5 months.\n\n4. Finally, I find the feature-space linearity assumption $E(y\\mid x) = w_t^\\top f(x)$ somewhat unsettling for the use in causal representation learning.  While similar assumptions have appeared in stylized theoretical analyses for multi-task learning, designing a causal representation learning procedure based on such assumptions appears to be asking much more.", "questions": "See above, in particular points 1-3.\n\n---\n\n**Post-rebuttal update.** Thank you for your response.  Unfortunately, I remain concerned about the assumptions and presentation of the results, and I will keep my score unchanged.\n\n- I do not agree that the feature space linearity is \"not an assumption\": this condition, and theorem 3.2, only hold because of the conditions on the data generating process introduced in Section 3.1.   Your claim \"any non-linear neural network broadly makes this assumption\" is somewhat misleading because for general task distributions (on which the conditions in Section 3.1 may not hold *for any fixed latent dimensionalities*), for the condition to hold approximately we may need the NN feature dimensionality to grow w.r.t. the number of tasks, thereby violating the rank condition in Theorem 3.2.\n\n- I am not convinced that the assumption of a *correctly specified* Gaussian likelihood is not a natural choice, as is also noted b reviewer 2cPA.  Currently your proofs rely on the likelihood being correctly specified; the best way to address this concern is to rewrite them so that they apply to (certain) misspecified cases.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studies causal representation learning for multi-task regression data.  The authors assume a data generating process under which \n\n1. the response variable $y$ is always a linear function of the causal and spurious latent variables $z$, with the linear map varying across tasks,\n2. the observed features $x$ are informative about the latent variables and the tasks are sufficiently diverse,\n3. the conditional distributions $p(y\\mid x, t) = N(y\\mid w_t^\\top f(x), \\sigma_t^2 I)$ and $p(f(x)\\mid z)$ are homoscedastic Gaussian,\n\nand propose a two-stage procedure that can recover the causal latent variables up to permutations and scaling.  The method outperforms previous approaches on synthetic data, and is demonstrated to produce stable outputs on two real-world datasets.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The authors studied an important problem. \n- The proposed method demonstrates promising empirical performance.\n- Compared to recent works on the same problem, the proposed method is more computationally efficient.", "weaknesses": "1. My main concern is the Gaussianity assumptions made throughout the work.  Such assumptions rarely hold in practice; and in contrast to standard machine learning tasks where most methods remain useful even if such assumptions are violated (e.g. typical regression procedures continue to estimate the conditional expectation), for causal representation learning the utility of the proposed method is far less clear.\n\n2. The above concern mostly applies to the second stage of the procedure.  For the first stage it appears that the method only relies on the conditional expectation having the form of $E(y\\mid x) = w_t^\\top f(x)$ (plus bijectivity and task diversity), but in such cases I feel that the assumption is better stated just as in Lachapelle et al (2023), namely we can learn a feature extractor that is linearly equivalent to the ground truth.\n\n3. Regarding experiments, I think it would be more convincing if the authors could include comparisons to the recent works of Lachapelle et al (2023) and Fumero et al (2023), as they addressed the same problem and have been publicly available for 4-5 months.\n\n4. Finally, I find the feature-space linearity assumption $E(y\\mid x) = w_t^\\top f(x)$ somewhat unsettling for the use in causal representation learning.  While similar assumptions have appeared in stylized theoretical analyses for multi-task learning, designing a causal representation learning procedure based on such assumptions appears to be asking much more.", "questions": "See above, in particular points 1-3.\n\n---\n\n**Post-rebuttal update.** Thank you for your response.  Unfortunately, I remain concerned about the assumptions and presentation of the results, and I will keep my score unchanged.\n\n- I do not agree that the feature space linearity is \"not an assumption\": this condition, and theorem 3.2, only hold because of the conditions on the data generating process introduced in Section 3.1.   Your claim \"any non-linear neural network broadly makes this assumption\" is somewhat misleading because for general task distributions (on which the conditions in Section 3.1 may not hold *for any fixed latent dimensionalities*), for the condition to hold approximately we may need the NN feature dimensionality to grow w.r.t. the number of tasks, thereby violating the rank condition in Theorem 3.2.\n\n- I am not convinced that the assumption of a *correctly specified* Gaussian likelihood is not a natural choice, as is also noted b reviewer 2cPA.  Currently your proofs rely on the likelihood being correctly specified; the best way to address this concern is to rewrite them so that they apply to (certain) misspecified cases.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699038719460}, {"id": "ukAQru3rHH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission970/Reviewer_mdpZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose an identification result under the hypothesis that data are generated by a specific graphical model.", "review_text": "The authors propose an identification result under the hypothesis that data are generated by a specific graphical model.", "strengths": "The identifiability result seems new", "weaknesses": "My main complaint is the clarity of the paper.\nI am not sure I properly understood the setup or the contribution: no algorithm is clearly proposed.\n\n- Abstract \"In such cases, we show that identifiability is achievable even in the case of regression, extending prior work restricted to linear identifiability in the single-task classification case.\" I am not sure to understand this sentence since the authors refer to previous identification results in the multitasks setting (section 3.2 for instance)\n\n- Section 3.3.1 is a succession of 6 equations, is it possible to encapsulate the assumptions in a proper environment", "questions": "- What's the proposed algorithm? I was not able to parse it.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose an identification result under the hypothesis that data are generated by a specific graphical model.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The identifiability result seems new", "weaknesses": "My main complaint is the clarity of the paper.\nI am not sure I properly understood the setup or the contribution: no algorithm is clearly proposed.\n\n- Abstract \"In such cases, we show that identifiability is achievable even in the case of regression, extending prior work restricted to linear identifiability in the single-task classification case.\" I am not sure to understand this sentence since the authors refer to previous identification results in the multitasks setting (section 3.2 for instance)\n\n- Section 3.3.1 is a succession of 6 equations, is it possible to encapsulate the assumptions in a proper environment", "questions": "- What's the proposed algorithm? I was not able to parse it.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698942389083}, {"id": "abyHodnk6M", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission970/Reviewer_wjky"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper studies the problem of identifying causal representations in the multi-task regression setting. The considered regression function consists of an invariant feature extractor and task-dependent linear head. The identifiability of the feature extractor and linear head is provided (up to some equivalent classes). The main concern for the work is the model assumption on the relation between the target variable and spurious features.", "review_text": "The paper studies the problem of identifying causal representations in the multi-task regression setting. The considered regression function consists of an invariant feature extractor and task-dependent linear head. The identifiability of the feature extractor and linear head is provided (up to some equivalent classes). The main concern for the work is the model assumption on the relation between the target variable and spurious features.", "strengths": "1. The paper is well-written overall. \n\n2. Solid identifiability results are provided.", "weaknesses": "1. In the assumed causal graph, $Y$ is a parent of the spurious features $z_{s}$, which is not a natural assumption. The correlation between $Y$ and $z_{s}$ is often due to latent confounders. \n\n2. What does ``MLE (1) converges\" in Corollary 3.3 mean? $\\theta'$ defined in (1) is a population estimator. There is a similar issue for  (17) in Theorem 3.5. Besides, the uniqueness of the (population) MLE estimators (i.e., (1) and (17)) should be assumed explicitly. \n\n3. What is $c_{t}$ below equation (4)?", "questions": "The main concern is the causal relation between $Y$ and the spurious features $z_{s}$. If there are indeed many real-world settings where $Y$ is the parent for the spurious features. Please cite related references to support the assumption. Otherwise, the identifiability for settings with the latent confounder should be studied as well.\n\nI may raise my score based on the response to this question.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of identifying causal representations in the multi-task regression setting. The considered regression function consists of an invariant feature extractor and task-dependent linear head. The identifiability of the feature extractor and linear head is provided (up to some equivalent classes). The main concern for the work is the model assumption on the relation between the target variable and spurious features.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The paper is well-written overall. \n\n2. Solid identifiability results are provided.", "weaknesses": "1. In the assumed causal graph, $Y$ is a parent of the spurious features $z_{s}$, which is not a natural assumption. The correlation between $Y$ and $z_{s}$ is often due to latent confounders. \n\n2. What does ``MLE (1) converges\" in Corollary 3.3 mean? $\\theta'$ defined in (1) is a population estimator. There is a similar issue for  (17) in Theorem 3.5. Besides, the uniqueness of the (population) MLE estimators (i.e., (1) and (17)) should be assumed explicitly. \n\n3. What is $c_{t}$ below equation (4)?", "questions": "The main concern is the causal relation between $Y$ and the spurious features $z_{s}$. If there are indeed many real-world settings where $Y$ is the parent for the spurious features. Please cite related references to support the assumption. Otherwise, the identifiability for settings with the latent confounder should be studied as well.\n\nI may raise my score based on the response to this question.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698805073461}], "openreview_url": "https://openreview.net/forum?id=kkQSwtx0p3", "arxiv_id": "2306.14861", "paper_pdf": "papers/kkQSwtx0p3.pdf", "paper_pdf_sha256": "444affb7351004ef2364172b7983043b3edf9845b8b4f603bb1f1bc84864d128", "paper_pdf_bytes": 528109, "paper_pdf_source": "openreview", "code_url": "https://github.com/jdhorwood/mtlcm", "code_repository": "jdhorwood/mtlcm", "code_commit": "ea559ae662d08605ca65ff3c02b1f52fd1267513", "code_archive": "repos/kkQSwtx0p3.zip", "code_archive_sha256": "996bfec1d15394f6b48e975a55d20de1019f0066f84246977d171cdce8b17035", "code_archive_bytes": 91116, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 147, "github_languages": {"Python": 77092}, "github_archived": false, "github_pushed_at": "2024-08-28T01:30:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/leveraging-task-structures-for-improved"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3TfSOxiRiFH", "year": 2023, "status": "rejected", "title": "On a Built-in Conflict between Deep Learning and Systematic Generalization", "authors": ["Yuanpeng Li"], "authorids": ["~Yuanpeng_Li2"], "authors_source": "OpenReview API", "abstract": "Out-of-distribution or systematic generalization is a desirable property that most deep learning algorithms lack. In this paper, we hypothesize that internal function sharing is one of the reasons to weaken systematic generalization in deep learning for classification tasks. Under equivalent prediction, a model partitions an input space into multiple parts separated by boundaries. The function sharing prefers to reuse boundaries, leading to fewer parts for new outputs, which conflicts with systematic generalization. We show such phenomena in standard deep learning models, such as fully connected, convolutional, residual networks, LSTMs, and (Vision) Transformers. We hope this study provides novel insights and forms a basis for new research directions to improve systematic generalization.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "OPHp_0G0jGx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper549/Reviewer_w8Ad"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper hypothesise that function sharing is one of the reasons why deep learning models can’t perform systematic generalization. The paper demonstrate that as the degree of parameter sharing increases, the systematic generalization drops. From the practical stand point, the papers argues for sparsity in models that somewhat learn symbolic functions (in term of disentangling feature attributes). Although it’s not touched upon but I believe the paper can be seen from modularity perspective where each module encampasses a particular underlying function, describing certain factor of the input. ", "review_text": "This work need major re-writing as most of the concepts were not framed correctly. Moreover, they are some shortcomings in the evaluation section as I explained in the weaknesses section. I believe there is some value in this work, but it needs written with clarity.", "strengths": "### Strengths\n- The paper provides empirical evidence of why parameter sharing (function sharing) leads to performance drop in systematic generalization for deep learning models.\n- I like the problem space and believe that the authors are on to something tangible here but the lack of rigour in analysis didn't convince me.\n\n### Weaknesses\n- I have issues with problem formulation and writing. Some details of the work are not framed correctly and it’s hard to understand it since no context from previous literature is provided while introducing and explaining new concepts.\n- Eg. what is implied as functions in deep neural networks? Is it an individual parameter or a set of parameters?\n- What does the three equations at the end of sec 3.1 refer to?\n- There is no clear description of the dataset. I can see that it contains factor but what are those factors? How are those factors combined?\n- On the same note, could you please provide description of the datasets and models used, separately?\n- The results section just provide information on the models tried. No information on training, and test, train splits of dataset provided.\n- The author try to ablate different model architectures, however they use different datasets across those architectures thus it’s hard judge if the results are consistent across those architectures or the differences arise from the difference in datasets.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper hypothesise that function sharing is one of the reasons why deep learning models can’t perform systematic generalization. The paper demonstrate that as the degree of parameter sharing increases, the systematic generalization drops. From the practical stand point, the papers argues for sparsity in models that somewhat learn symbolic functions (in term of disentangling feature attributes). Although it’s not touched upon but I believe the paper can be seen from modularity perspective where each module encampasses a particular underlying function, describing certain factor of the input. ", "strength_and_weaknesses": "### Strengths\n- The paper provides empirical evidence of why parameter sharing (function sharing) leads to performance drop in systematic generalization for deep learning models.\n- I like the problem space and believe that the authors are on to something tangible here but the lack of rigour in analysis didn't convince me.\n\n### Weaknesses\n- I have issues with problem formulation and writing. Some details of the work are not framed correctly and it’s hard to understand it since no context from previous literature is provided while introducing and explaining new concepts.\n- Eg. what is implied as functions in deep neural networks? Is it an individual parameter or a set of parameters?\n- What does the three equations at the end of sec 3.1 refer to?\n- There is no clear description of the dataset. I can see that it contains factor but what are those factors? How are those factors combined?\n- On the same note, could you please provide description of the datasets and models used, separately?\n- The results section just provide information on the models tried. No information on training, and test, train splits of dataset provided.\n- The author try to ablate different model architectures, however they use different datasets across those architectures thus it’s hard judge if the results are consistent across those architectures or the differences arise from the difference in datasets.\n", "clarity,_quality,_novelty_and_reproducibility": "In general, the writing needs to be improved. It was not easy to follow all the details:\n- From the start, it was difficult to understand what a function means in deep learning context? To understand it better w.r.t machine learning literature, can “function sharing” be reframed in terms of modularity?\n- Figure 2 is hard to understand. Eg. what does the term “function” refers to in the diagram?", "summary_of_the_review": "This work need major re-writing as most of the concepts were not framed correctly. Moreover, they are some shortcomings in the evaluation section as I explained in the weaknesses section. I believe there is some value in this work, but it needs written with clarity.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667149988736}, {"id": "kQxQWHww5Q", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper549/Reviewer_e3dk"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper investigates systematic generalization in deep neural networks. Systematic generalization here refers to the ability of an algorithm to produce outputs that were not observed during training time. A potential reason for this is postulated: the lack of systematic generalization in deep neural networks is due to function sharing, i.e. that each layer in the network uses a common representation from the previous layer. Experiments show that networks with fewer shared intermediate layers exhibit a greater degree systematic generalization than those with more shared layers.", "review_text": "After rebuttal: The paper has been improved by the inclusion of real-world experiments. I would echo the other reviewers in suggesting that these move to the main paper. As pointed out by Reviewer hir1, a more complete discussion about when function sharing is likely to be helpful seems necessary. Expanding on why function sharing occurs as started in Appendix C would also be useful. I have updated my score accordingly. \n\n---\n\nOverall, there are some potentially interesting ideas in this paper, but the clarity, unsupported claims, and incompleteness of the experiments are somewhat significant issues. The contribution of this paper is on the more incremental side and consists primarily of showing that sharing fewer layers can improve systematic generalization. However, it is not clear how to take these insights and apply them to solve out-of-distribution detection on real-world problems.", "strengths": "Strengths\n- The broad aim of investigating stronger forms of generalization that move beyond the i.i.d. case is interesting.\n- The experimental observation that having fewer shared layers leads to better systematic generalization holds across multiple diverse architectures and datasets. \n- As far as I know, investigating function sharing as a reason for lack of systematic generalization is a novel approach.\n\nWeaknesses\n- The paper is not written very clearly. Sections that are difficult to understand include: the mathematical notation (e.g. writing that $f$ is a \"model\" but not explaining that this is simply a mapping from the input space to output space), the experiment section (particularly how the labels were generated and what the different evaluation metrics mean), and the discussion section.\n- Some intuitions are claimed but not supported by adequate evidence: that deep neural networks prefer to learn a simple function and combine with previously learned functions, and that neural networks greedily learn functions in order of simpler to more complex.\n- The experiments are not complete. Some relevant but missing pieces of information include training accuracy and computation time for the various levels of sharing. On a related note, the motivation for including the test set and random set accuracy metrics is unclear.\n- It is not clear how the knowledge introduced by this paper can be effectively used to improve systematic generalization. Training a multitude of independent networks, one for each underlying factor, does not seem like a practical course of action due to storage and computation constraints.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper investigates systematic generalization in deep neural networks. Systematic generalization here refers to the ability of an algorithm to produce outputs that were not observed during training time. A potential reason for this is postulated: the lack of systematic generalization in deep neural networks is due to function sharing, i.e. that each layer in the network uses a common representation from the previous layer. Experiments show that networks with fewer shared intermediate layers exhibit a greater degree systematic generalization than those with more shared layers.", "strength_and_weaknesses": "Strengths\n- The broad aim of investigating stronger forms of generalization that move beyond the i.i.d. case is interesting.\n- The experimental observation that having fewer shared layers leads to better systematic generalization holds across multiple diverse architectures and datasets. \n- As far as I know, investigating function sharing as a reason for lack of systematic generalization is a novel approach.\n\nWeaknesses\n- The paper is not written very clearly. Sections that are difficult to understand include: the mathematical notation (e.g. writing that $f$ is a \"model\" but not explaining that this is simply a mapping from the input space to output space), the experiment section (particularly how the labels were generated and what the different evaluation metrics mean), and the discussion section.\n- Some intuitions are claimed but not supported by adequate evidence: that deep neural networks prefer to learn a simple function and combine with previously learned functions, and that neural networks greedily learn functions in order of simpler to more complex.\n- The experiments are not complete. Some relevant but missing pieces of information include training accuracy and computation time for the various levels of sharing. On a related note, the motivation for including the test set and random set accuracy metrics is unclear.\n- It is not clear how the knowledge introduced by this paper can be effectively used to improve systematic generalization. Training a multitude of independent networks, one for each underlying factor, does not seem like a practical course of action due to storage and computation constraints.", "clarity,_quality,_novelty_and_reproducibility": "As mentioned above, there are some significant issues with clarity, but I believe it is possible for them to be resolved in an updated version of the paper. The paper also seems novel enough in that it looks at function sharing as a potential underlying cause for lack of systematic generalization. The quality is below average, with issues including incomplete experimental evaluation and unsupported claims (e.g. greedy learning of functions and the mechanisms underlying function sharing). I also found the design choice to average inputs from two separate datasets to be unconventional.\n\nOne point that seems quite relevant but not addressed by the paper is the extent to which the phenomenon in Figure 2 is caused by a softmax activation, which assumes that classes are mutually exclusive. This seems to be an alternate reason that the case in Figure 2(b) does not arise, since this region occupied by the orange dot would be a region of low confidence and thus the network would be incentivized to sharpen the decision boundaries. Could this possibly be resolved by assuming a multi-output loss function, e.g. multiple sigmoids? Regarding reproducibility, code is included and thus reproducing the results does not seem to be a major barrier.", "summary_of_the_review": "After rebuttal: The paper has been improved by the inclusion of real-world experiments. I would echo the other reviewers in suggesting that these move to the main paper. As pointed out by Reviewer hir1, a more complete discussion about when function sharing is likely to be helpful seems necessary. Expanding on why function sharing occurs as started in Appendix C would also be useful. I have updated my score accordingly. \n\n---\n\nOverall, there are some potentially interesting ideas in this paper, but the clarity, unsupported claims, and incompleteness of the experiments are somewhat significant issues. The contribution of this paper is on the more incremental side and consists primarily of showing that sharing fewer layers can improve systematic generalization. However, it is not clear how to take these insights and apply them to solve out-of-distribution detection on real-world problems.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666815857296}, {"id": "o7YUn2fisf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper549/Reviewer_hir1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper investigates systematic generalization of multi-label classifications in settings where the input space is shared between the different labels. The authors suggest that deep learning models are biased towards feature reuse, which conflicts with systematic generalization to new combinations of known classes. They explore this hypothesis with theoretical analyses under toy assumptions, and empirical experiments across a variety of architectures. Generally, they find that more sharing of features reduces generalization accuracy in their settings.", "review_text": "See comment above for my post-response update. I am updating my score in accordance with the improvement in the paper, but I am concerned that key issues still remain unresolved, and not discussed with enough nuance.\n\nOriginal review\n-------------------\n\nIf this paper were completely rewritten—to describe the experiments as identifying a particular class of problems in which DL architectures seem not to generalize systematically due to feature-sharing rather than a built-in conflict—I believe it could be an acceptable paper in some venue. If, in addition to that, the authors were to identify real-world, non-adversarial datasets where their observations bear out, and perform some of the architecture/training experiments suggested above, I would consider it a strong paper for NeurIPS. As it is, I think it is misleading.", "strengths": "Strengths:\n* The questions posed are interesting.\n* I appreciate the breadth of architectures considered, especially sharing different numbers of layers.\n\n\nWeaknesses:\n\nThe way the paper is presented fundamentally ignores the No Free Lunch theorem [e.g. Adam, 2019]. There is no system that can generalize perfectly on every task and training dataset—there cannot be a conflict between an architecture class and systematic generalization writ large. We have to ask the question of how the inductive biases of the model class fit the class of tasks we are interested in solving.\n* DL researchers are well aware of the feature sharing bias—it forms the basis of auxiliary task training and/or pretraining methods, as the authors note. The reason such methods tend to improve generalization is because sharing features is useful on real-world datasets. For example, the input features learned solve masked language modeling tasks empirically improve systematic generalization performance substantially even on tasks like SCAN and CFQ [Furrer et al., 2020]. There are even theoretical accounts of why feature sharing can improve generalization in the presence of noise [e.g. Lampinen et al, 2019]. \n* The datasets used in the paper are therefore cleverly created to make feature sharing an actively harmful strategy. But to do so, the authors rely on essentially adversarial dataset design, where they combine input stimuli in very unnatural ways (averaging images, or concatenating completely unrelated pieces of text), and then enforce extremely strong correlations between these inputs at train time, which are completely reversed at test time. There is no reason given to think that this process has anything to do with any real-world data generating process.\n* Therefore, I would challenge the authors to demonstrate *real-world tasks and datasets*, not artificially, adversarially created ones, in which their observations apply.\n* Otherwise, it seems to me that feature sharing is a *feature, not a bug* of deep learning. Nobody has ever claimed that deep learning is capable of generalizing systematically in every task anyone can come up with—that would violate the NFL theorem. But I’d argue that the DL family is empirically the most successful system for generalizing on real world datasets, in part because of feature sharing.\n\nArchitectures and training paradigms:\n* “We choose a layer and duplicate the following layers, keeping the number of all hidden nodes in each layer if feasible” — it is not clear to me whether this means that each “branch” of the architecture has the same number of nodes as before, or half the number of nodes. If the former, the number of parameters will be larger in networks that split earlier, thus confounding the comparison (since overparameterized models tend to generalize better). \n* More generally, it would be interesting to see the impact of parameterization on these effects—one might expect somewhat less feature sharing in wider networks, for instance, though it’s unclear how strong the effect would be.\n* And it would be interesting to see the effect of methods like dropout [Srivastava et al., 2014] or mixup [Zhang et al., 2017] which are known to improve generalization.\n\n\nReferences\n------------\n\nAdam, Stavros P., et al. \"No free lunch theorem: A review.\" Approximation and optimization (2019): 57-82.\n\n\nFurrer, D., van Zee, M., Scales, N., & Schärli, N. (2020). Compositional generalization in semantic parsing: Pre-training vs. specialized architectures. arXiv preprint arXiv:2007.08970.\n\nLampinen, A. K., & Ganguli, S. (2019). An analytic theory of generalization dynamics and transfer learning in deep linear networks. In International Conference on Learning Representations.\n\nSrivastava, Nitish, et al. \"Dropout: a simple way to prevent neural networks from overfitting.\" The journal of machine learning research 15.1 (2014): 1929-1958.\n\nZhang, H., Cisse, M., Dauphin, Y. N., & Lopez-Paz, D. (2017). mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper investigates systematic generalization of multi-label classifications in settings where the input space is shared between the different labels. The authors suggest that deep learning models are biased towards feature reuse, which conflicts with systematic generalization to new combinations of known classes. They explore this hypothesis with theoretical analyses under toy assumptions, and empirical experiments across a variety of architectures. Generally, they find that more sharing of features reduces generalization accuracy in their settings.", "strength_and_weaknesses": "Strengths:\n* The questions posed are interesting.\n* I appreciate the breadth of architectures considered, especially sharing different numbers of layers.\n\n\nWeaknesses:\n\nThe way the paper is presented fundamentally ignores the No Free Lunch theorem [e.g. Adam, 2019]. There is no system that can generalize perfectly on every task and training dataset—there cannot be a conflict between an architecture class and systematic generalization writ large. We have to ask the question of how the inductive biases of the model class fit the class of tasks we are interested in solving.\n* DL researchers are well aware of the feature sharing bias—it forms the basis of auxiliary task training and/or pretraining methods, as the authors note. The reason such methods tend to improve generalization is because sharing features is useful on real-world datasets. For example, the input features learned solve masked language modeling tasks empirically improve systematic generalization performance substantially even on tasks like SCAN and CFQ [Furrer et al., 2020]. There are even theoretical accounts of why feature sharing can improve generalization in the presence of noise [e.g. Lampinen et al, 2019]. \n* The datasets used in the paper are therefore cleverly created to make feature sharing an actively harmful strategy. But to do so, the authors rely on essentially adversarial dataset design, where they combine input stimuli in very unnatural ways (averaging images, or concatenating completely unrelated pieces of text), and then enforce extremely strong correlations between these inputs at train time, which are completely reversed at test time. There is no reason given to think that this process has anything to do with any real-world data generating process.\n* Therefore, I would challenge the authors to demonstrate *real-world tasks and datasets*, not artificially, adversarially created ones, in which their observations apply.\n* Otherwise, it seems to me that feature sharing is a *feature, not a bug* of deep learning. Nobody has ever claimed that deep learning is capable of generalizing systematically in every task anyone can come up with—that would violate the NFL theorem. But I’d argue that the DL family is empirically the most successful system for generalizing on real world datasets, in part because of feature sharing.\n\nArchitectures and training paradigms:\n* “We choose a layer and duplicate the following layers, keeping the number of all hidden nodes in each layer if feasible” — it is not clear to me whether this means that each “branch” of the architecture has the same number of nodes as before, or half the number of nodes. If the former, the number of parameters will be larger in networks that split earlier, thus confounding the comparison (since overparameterized models tend to generalize better). \n* More generally, it would be interesting to see the impact of parameterization on these effects—one might expect somewhat less feature sharing in wider networks, for instance, though it’s unclear how strong the effect would be.\n* And it would be interesting to see the effect of methods like dropout [Srivastava et al., 2014] or mixup [Zhang et al., 2017] which are known to improve generalization.\n\n\nReferences\n------------\n\nAdam, Stavros P., et al. \"No free lunch theorem: A review.\" Approximation and optimization (2019): 57-82.\n\n\nFurrer, D., van Zee, M., Scales, N., & Schärli, N. (2020). Compositional generalization in semantic parsing: Pre-training vs. specialized architectures. arXiv preprint arXiv:2007.08970.\n\nLampinen, A. K., & Ganguli, S. (2019). An analytic theory of generalization dynamics and transfer learning in deep linear networks. In International Conference on Learning Representations.\n\nSrivastava, Nitish, et al. \"Dropout: a simple way to prevent neural networks from overfitting.\" The journal of machine learning research 15.1 (2014): 1929-1958.\n\nZhang, H., Cisse, M., Dauphin, Y. N., & Lopez-Paz, D. (2017). mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412.\n", "clarity,_quality,_novelty_and_reproducibility": "The clarity of the paper could be improved. For example:\n* Showing examples of the task stimuli in the paper—particularly the visual ones—would I think help to emphasize how unnatural the tasks are.\n* The data preparation could be rewritten to be clearer, by first specifying that the data generating process goes from sampling a pair of labels to sampling the corresponding input.\n* Split architecture details were unclear to me (noted above).\n\nThere is some originality and quality if the above weaknesses are addressed.", "summary_of_the_review": "See comment above for my post-response update. I am updating my score in accordance with the improvement in the paper, but I am concerned that key issues still remain unresolved, and not discussed with enough nuance.\n\nOriginal review\n-------------------\n\nIf this paper were completely rewritten—to describe the experiments as identifying a particular class of problems in which DL architectures seem not to generalize systematically due to feature-sharing rather than a built-in conflict—I believe it could be an acceptable paper in some venue. If, in addition to that, the authors were to identify real-world, non-adversarial datasets where their observations bear out, and perform some of the architecture/training experiments suggested above, I would consider it a strong paper for NeurIPS. As it is, I think it is misleading.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666277251664}, {"id": "jf9EMUVAZIZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper549/Reviewer_4tfA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper hypothesizes that systematic generalization is fundamentally at odds with the tendency of deep learning models to share sub-modules. It provides both a theoretical description and derivation of the issue as well as supporting experiments with several deep learning architectures.", "review_text": "This paper introduces an interesting and probably novel idea but it does not a good job at developing it. I believe it does not meet the ICLR bar.", "strengths": "Strengths: the hypothesis advanced by this paper is new and interesting; it deserves to be developed further.\n\nWeaknesses:\n- The paper is not very well written; see below for some examples where the writing should be improved.\n- The propositions and theorem are trivial.\n- The vision experiments seem quite unnatural, since images from different datasets are averaged (see Section 3.2). Also the text experiments are not particularly natural (they concatenate two unrelated sequences and ask to predict both labels).\n\nA general question that might be worth addressing: the shape of the regions is not taken into account? For instance, to prove Proposition 1 (using Assumption 2) you can simply merge regions, but it might be unnatural to merge them, given the geometry.\n\nExamples of unclear parts in Section 2.1:\n- \"$Y$ contains $K$ factors $Y_1$, . . . , $Y_K$\" is not clear. Does it mean that $Y = Y_1 \\times \\dots \\times Y_K$? But this would be at odds with the (unclear) statement \"which can be entangled\".\n- Is $X_\\text{train}$ the (ordered) sequence of inputs or is it a set that contains all the input data? (same for $Y_\\text{train}$ and test)\n- \"The values for each factor $i$ are included in the training output\" is also unclear\n- \"A model $f$ maps input $X$ to the prediction of output $f (X)$\": so is $X$ the set of all inputs or a single input?\n\nExamples of unclear parts in Section 2.2:\n- \"deep learning more or equally prefers $f$ over $g$\" -> \"deep learning prefers $f$ over $g$ more or equally\" (a bit more clear in my opinion)\n\nExamples of unclear parts in Section 3.1:\n- \"$Y_1$ is chosen from all possible labels\": what does this mean? That $Y_1$ is the set of all labels?\n- The two datasets share the same input space $X$?\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper hypothesizes that systematic generalization is fundamentally at odds with the tendency of deep learning models to share sub-modules. It provides both a theoretical description and derivation of the issue as well as supporting experiments with several deep learning architectures.", "strength_and_weaknesses": "Strengths: the hypothesis advanced by this paper is new and interesting; it deserves to be developed further.\n\nWeaknesses:\n- The paper is not very well written; see below for some examples where the writing should be improved.\n- The propositions and theorem are trivial.\n- The vision experiments seem quite unnatural, since images from different datasets are averaged (see Section 3.2). Also the text experiments are not particularly natural (they concatenate two unrelated sequences and ask to predict both labels).\n\nA general question that might be worth addressing: the shape of the regions is not taken into account? For instance, to prove Proposition 1 (using Assumption 2) you can simply merge regions, but it might be unnatural to merge them, given the geometry.\n\nExamples of unclear parts in Section 2.1:\n- \"$Y$ contains $K$ factors $Y_1$, . . . , $Y_K$\" is not clear. Does it mean that $Y = Y_1 \\times \\dots \\times Y_K$? But this would be at odds with the (unclear) statement \"which can be entangled\".\n- Is $X_\\text{train}$ the (ordered) sequence of inputs or is it a set that contains all the input data? (same for $Y_\\text{train}$ and test)\n- \"The values for each factor $i$ are included in the training output\" is also unclear\n- \"A model $f$ maps input $X$ to the prediction of output $f (X)$\": so is $X$ the set of all inputs or a single input?\n\nExamples of unclear parts in Section 2.2:\n- \"deep learning more or equally prefers $f$ over $g$\" -> \"deep learning prefers $f$ over $g$ more or equally\" (a bit more clear in my opinion)\n\nExamples of unclear parts in Section 3.1:\n- \"$Y_1$ is chosen from all possible labels\": what does this mean? That $Y_1$ is the set of all labels?\n- The two datasets share the same input space $X$?\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the general idea of the paper is easy to understand but the details are not explained in a very clear way.\nNovelty: as far as I know, the hypothesis introduced in this paper is new.", "summary_of_the_review": "This paper introduces an interesting and probably novel idea but it does not a good job at developing it. I believe it does not meet the ICLR bar.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1665612308227}], "openreview_url": "https://openreview.net/forum?id=3TfSOxiRiFH", "arxiv_id": "2208.11633", "paper_pdf": "papers/3TfSOxiRiFH.pdf", "paper_pdf_sha256": "d8d0b4487ad6cd050cf254d96ab38be8e732df42ee8452975efee5ad65d87a38", "paper_pdf_bytes": 1142934, "paper_pdf_source": "openreview", "code_url": "https://github.com/yuanpeng16/BCDLSG", "code_repository": "yuanpeng16/BCDLSG", "code_commit": "08763a62d60e97cf2a765c7b8069f62e785ddda2", "code_archive": "repos/3TfSOxiRiFH.zip", "code_archive_sha256": "06d26134860a069f0fa20e2c4971fae2d6f85805012ad03217fe3e94d34dacfd", "code_archive_bytes": 55675, "code_file_count": 46, "code_extensions": {".sh": 32, ".py": 14}, "github_disk_usage_kb": 305, "github_languages": {"Python": 110383, "Shell": 23618}, "github_archived": false, "github_pushed_at": "2024-11-25T05:19:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-a-built-in-conflict-between-deep-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xf0B7-7MRo6", "year": 2022, "status": "rejected", "title": "AIR-Net: Adaptive and Implicit Regularization Neural Network for matrix completion", "authors": ["Zhemin Li", "Hongxia Wang"], "authorids": ["~Zhemin_Li1", "~Hongxia_Wang1"], "authors_source": "OpenReview API", "abstract": "Conventionally, the matrix completion (MC) model aims to recover a matrix from partially observed elements. Accurate recovery necessarily requires a regularization encoding priors of the unknown matrix/signal properly. However, encoding the priors accurately for the complex natural signal is difficult, and even then, the model might not generalize well outside the particular matrix type. This work combines adaptive and implicit low-rank regularization that captures the prior dynamically according to the current recovered matrix. Furthermore, we aim to answer the question: how does adaptive regularization affect implicit regularization? We utilize neural networks to represent Adaptive and Implicit Regularization and named the proposed model \\textit{AIR-Net}. Theoretical analyses show that the adaptive part of the AIR-Net enhances implicit regularization. In addition, the adaptive regularizer vanishes at the end, thus can avoid saturation issues. Numerical experiments for various data demonstrate the effectiveness of AIR-Net, especially when the locations of missing elements are not randomly chosen. With complete flexibility to select neural networks for matrix representation, AIR-Net can be extended to solve more general inverse problems.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Rk1emqKK0lT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2533/Reviewer_Bmxi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies the problem of matrix completion using neural networks and deep matrix factorization as implicit and explicit regularization. The paper proposes a general framework and studies a specific case, namely, when the regularization is a form of Dirichlet Energy on the rows and columns of the matrix, and the matrix is formed as a product of L matrices. The results indicate improved performance of matrix completion under a variety of corruption methods and for a number of data sets.\n", "review_text": "- A general framework is proposed for studing matrix completion problems with regularization\n- A good specific choice of implicit regularization based on the Dirichlet Energy with neural networks is proposed\n- Theoretical analysis of the proposed algorithm indicates the expected system dynamcis\n- Experimental evidence shows improvement on a variety of matrix completion tasks\n\nPros:\nThe results are interesting and compelling; indeed, there is clear improvement in a variety of tasks, and the dynamics of the reconstruction match the expectation in theory. It is also useful to draw links to other works through the more general formulation. Effective heuristics for regularization parameters are chosen so that they do not need to be tuned. The DE regularization has an intuitive impact.\n\nCons:\nThe experiments are rather limited to a small set of test cases (three images only) and it is not clear if the results are consistent across a larger data set.\n\nThere are also details missing, for example, a better explanation of the construction of the Laplace matrix in the DE regularization, and why it was chosen. Also, it is not clear that the experiment using fixed L at different iterations is so useful. Could instead the ideal L be used calculated from the fully sampled matrix?\n\nIt is also not entirely clear what value is added by generalizing other methods based on a very generic optimization equation (that essentially can capture any optimization problem). In particular, the main focus of the paper seems to be in the specific choice to regularize the DE of both the rows and the columns. However, no comparison is made to a situation where only one or the other is used. It is also not clear if overfitting can become an issue with enough iterations.\n\nAs a general comment, the writing is not always clear and could be improved to help clarity. There are also small errors, for example I believe \"t-th row\" in Figure 1 caption should say \"i-th row\".\n\nBased on the preliminary results, I believe the paper is marginally below the acceptance threshold.\n\nThe justifications for this decision is the limited set of experiments justifying the choices made. Specifically, the value of adding the row- and column- DE regularizers, and the interplay with deep matrix factorization. In addition, the justification and explanation of the construction of Lr and Lc.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the problem of matrix completion using neural networks and deep matrix factorization as implicit and explicit regularization. The paper proposes a general framework and studies a specific case, namely, when the regularization is a form of Dirichlet Energy on the rows and columns of the matrix, and the matrix is formed as a product of L matrices. The results indicate improved performance of matrix completion under a variety of corruption methods and for a number of data sets.\n", "main_review": "- A general framework is proposed for studing matrix completion problems with regularization\n- A good specific choice of implicit regularization based on the Dirichlet Energy with neural networks is proposed\n- Theoretical analysis of the proposed algorithm indicates the expected system dynamcis\n- Experimental evidence shows improvement on a variety of matrix completion tasks\n\nPros:\nThe results are interesting and compelling; indeed, there is clear improvement in a variety of tasks, and the dynamics of the reconstruction match the expectation in theory. It is also useful to draw links to other works through the more general formulation. Effective heuristics for regularization parameters are chosen so that they do not need to be tuned. The DE regularization has an intuitive impact.\n\nCons:\nThe experiments are rather limited to a small set of test cases (three images only) and it is not clear if the results are consistent across a larger data set.\n\nThere are also details missing, for example, a better explanation of the construction of the Laplace matrix in the DE regularization, and why it was chosen. Also, it is not clear that the experiment using fixed L at different iterations is so useful. Could instead the ideal L be used calculated from the fully sampled matrix?\n\nIt is also not entirely clear what value is added by generalizing other methods based on a very generic optimization equation (that essentially can capture any optimization problem). In particular, the main focus of the paper seems to be in the specific choice to regularize the DE of both the rows and the columns. However, no comparison is made to a situation where only one or the other is used. It is also not clear if overfitting can become an issue with enough iterations.\n\nAs a general comment, the writing is not always clear and could be improved to help clarity. There are also small errors, for example I believe \"t-th row\" in Figure 1 caption should say \"i-th row\".\n\nBased on the preliminary results, I believe the paper is marginally below the acceptance threshold.\n\nThe justifications for this decision is the limited set of experiments justifying the choices made. Specifically, the value of adding the row- and column- DE regularizers, and the interplay with deep matrix factorization. In addition, the justification and explanation of the construction of Lr and Lc.\n\n", "summary_of_the_review": "In summary, the paper proposes a nice combination of implicit and explicit regularization for matrix completion using a deep matrix factorization and penalty on the DE of the rows and columns. The results are compelling, but some of the choices made are not fully justified or analyzed.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635979647429}, {"id": "6GCzRaDB_Vg", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2533/Reviewer_Untt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the matrix completion problem where the goal is to recover the matrix from partially observed elements.  The proposed approach involves parameterizing the unknown matrix by deep matrix factorization and adaptive regularizers that are parameterized with deep neural networks. The authors studied the Adaptive and Implicit Regularization of the proposed approach which is called AIR-Net. Experiments are provided to demonstrate the effect of the proposed approach. ", "review_text": "## Strengths: \n1. Theoretically analyzing the gradient flow of the training loss to understand the properties of the proposed approach\n2. good experimental results.\n\n## Weaknesses\n1. The paper is not easy to follow due to complicated notations. Also, it lacks a detailed description of some important concepts. For example, the end of the introduction mentions that $L$ in the adaptive regularizer is parameterized with DNN. However, I didn't find any detailed description for this except for the formulation $L_i(W_i)$ right after eq. (2). What is $L_i(W_i)$? \n\n2. Without the detailed expression of $L_i(W_i)$  for the adaptive regularizer, the training loss in eq. (2) appears not well-posed. Particularly, the regularizers $trace(X L_r(W_r) X^\\top)$ and $trace(X L_l(W_l) X^\\top)$ are not bounded, i.e., they can go to negative infinity. Right after eq. (2), it says $R_{W_r}(X) = X L_r(W_r) X^\\top$ which I think is not correct. \n\n3. Theorem 1 analyzes the gradient flow when the regularizers are fixed, while Theorem 2 analyzes the gradient flow when the matrix is fixed. However, in practice, the entire training loss is minimized simultaneously, but no theoretical analysis is provided for this case.  With the regularizers fixed, Theorem 1 is similar to the existing results on analyzing the dynamic flow of deep matrix factorization, e.g., [Arora et al. 2019]. Difference and novelty compared to the existing work should be discussed.\n\n4. The two theorems lack sufficient discussion to help the readers understand the main results. For example, what is the role of $L_r$ and $L_c$ in Theorem 1? How do they affect the results? The set $C_2$ in Theorem 2 consists of identical rows or columns, which seems empty in practice. If this is the case, Theorem 2 implies that the adaptive regularizer $L_i$ becomes a diagonal matrix. What is the role of the regularizer in this case? \n\n5. Theorem 2 requires the matrix $X$ has positive elements, which is a very strong assumption and may not hold in practice. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the matrix completion problem where the goal is to recover the matrix from partially observed elements.  The proposed approach involves parameterizing the unknown matrix by deep matrix factorization and adaptive regularizers that are parameterized with deep neural networks. The authors studied the Adaptive and Implicit Regularization of the proposed approach which is called AIR-Net. Experiments are provided to demonstrate the effect of the proposed approach. ", "main_review": "## Strengths: \n1. Theoretically analyzing the gradient flow of the training loss to understand the properties of the proposed approach\n2. good experimental results.\n\n## Weaknesses\n1. The paper is not easy to follow due to complicated notations. Also, it lacks a detailed description of some important concepts. For example, the end of the introduction mentions that $L$ in the adaptive regularizer is parameterized with DNN. However, I didn't find any detailed description for this except for the formulation $L_i(W_i)$ right after eq. (2). What is $L_i(W_i)$? \n\n2. Without the detailed expression of $L_i(W_i)$  for the adaptive regularizer, the training loss in eq. (2) appears not well-posed. Particularly, the regularizers $trace(X L_r(W_r) X^\\top)$ and $trace(X L_l(W_l) X^\\top)$ are not bounded, i.e., they can go to negative infinity. Right after eq. (2), it says $R_{W_r}(X) = X L_r(W_r) X^\\top$ which I think is not correct. \n\n3. Theorem 1 analyzes the gradient flow when the regularizers are fixed, while Theorem 2 analyzes the gradient flow when the matrix is fixed. However, in practice, the entire training loss is minimized simultaneously, but no theoretical analysis is provided for this case.  With the regularizers fixed, Theorem 1 is similar to the existing results on analyzing the dynamic flow of deep matrix factorization, e.g., [Arora et al. 2019]. Difference and novelty compared to the existing work should be discussed.\n\n4. The two theorems lack sufficient discussion to help the readers understand the main results. For example, what is the role of $L_r$ and $L_c$ in Theorem 1? How do they affect the results? The set $C_2$ in Theorem 2 consists of identical rows or columns, which seems empty in practice. If this is the case, Theorem 2 implies that the adaptive regularizer $L_i$ becomes a diagonal matrix. What is the role of the regularizer in this case? \n\n5. Theorem 2 requires the matrix $X$ has positive elements, which is a very strong assumption and may not hold in practice. ", "summary_of_the_review": "Overall, based on the detailed description above, the main contribution of this paper seems limited as the gradient flow of deep matrix factorization has already been studied. Also, the form of the adaptive regularize is not clear and the main results lack sufficient explanation. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635905197909}, {"id": "UXhTTSlYCw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2533/Reviewer_B6ib"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this manuscript, the authors consider matrix completion problems. Leveraging recent advances on implicit regularization in (deep) matrix factorization problems, a new architecture for matrix completion is proposed. Specifically, the authors parameterize the unknown low-rank matrix as a deep linear network, which has been shown to exhibit a low-rank bias when learned via gradient descent. Additional regularization terms based on the Dirichlet energy are added to encourage other structural priors in the recovered solution, such as self-similarity between columns or blocks. The underlying Laplacian matrix is parameterized by learnable weights. The authors analyze the dynamics of gradient flow applied to their optimization problem and show empirically that this approach improves performance in certain settings.", "review_text": "**Quality:**\n\n- Strengths: - The proposed method works fairly well empirically, beating other baselines using implicit regularization and also more traditional approaches.\n\n- Weaknesses: - The paper can be quite hard to follow, see the clarity section below. -  Overall, the framework seems incremental. There does not seem to be a big distinction from [1], aside from parameterizing the dirichlet energy with learnable weights. The theoretical contributions appear minor and mirror those previously established in [2]. \n\n**Clarity:** The paper was hard to read and contained a number of typos. I have noted some here, but the manuscript contains others:\n- page 1 abstract: “requires a regularization” -> “requires regularization”\n- page 1: “over-parametric” -> “overparameterized”\n- page 2: “The low-rank” -> “Low-rankness”\n- page 2: “Furthermore, We” -> “Furthermore, we”\n- page 3: The functionals $\\mathcal{R}_{W_i}$ are missing the trace $\\text{tr}( \\cdot )$.\n- page 3: “Another expected” -> “Another unexpected”\n- page 5: “Specialy, we sigh two of $L_c(t=4000)$ out”\n- page 6: “Air-Net Adaptive to Both Varies Data” should “varies” be “varying”?\n- page 6 Figure 1: “the $t$-th row” -> “the $i$-th row”\n- page 7: “reminding missing” -> “remaining missing”\n\n**Novelty and significance:** The work seems to be an extension of [1], which also employed a deep matrix factorization approach to solve matrix completion problems and incorporated the dirichlet energy as an explicit regularizer. In that case, however, the laplacian was not learned, which seems to be the crucial difference. Some components of the theory is also borrowed from Arora et al [2].\n\n**General comments:**\n\n- At the top of page 3, it says that the structure of the network parameterizing the Laplacian matrix is discussed in A.4, but it appears to actually be in A.3. There is also no discussion on why the network is chosen in this particular fashion. It would be good to talk about why this network structure is used, e.g. the use of a softmax-type parameterization. Also, under this parameterization, is it even possible for the learned $L_i = 0$, the trivial solution?\n- It feels misleading to say in Remark 1 that Theorem 2 requires no restriction on $\\mathcal{T}_i$ or $X$. The first assumption is that all columns of $\\mathcal{T}_i(X)$ are unit normed and the entries are positive. If, for example, $\\mathcal{T}_i(X) := X$, then this means that for the result to hold, the columns of $X$ are unit normed and all of its entries are non-negative.\n- The matrix $C_{k,l}$ should be defined in Theorem 2, since it arises in the definition of $\\gamma$.\n- I am a bit puzzled by Figure 1. This seems to indicate that as time progresses, the network learns that there should be less and less similarity between rows/columns, as indicated by the lack of dark regions as $t$ increases. What is the intuition for this? Shouldn't some notion of self-similarity be learned instead?\n\n[1] Amit Boyarski, Sanketh Vedula, and Alex Bronstein. Spectral geometric matrix completion. MSML 2021\n\n[2] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS 2019", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this manuscript, the authors consider matrix completion problems. Leveraging recent advances on implicit regularization in (deep) matrix factorization problems, a new architecture for matrix completion is proposed. Specifically, the authors parameterize the unknown low-rank matrix as a deep linear network, which has been shown to exhibit a low-rank bias when learned via gradient descent. Additional regularization terms based on the Dirichlet energy are added to encourage other structural priors in the recovered solution, such as self-similarity between columns or blocks. The underlying Laplacian matrix is parameterized by learnable weights. The authors analyze the dynamics of gradient flow applied to their optimization problem and show empirically that this approach improves performance in certain settings.", "main_review": "**Quality:**\n\n- Strengths: - The proposed method works fairly well empirically, beating other baselines using implicit regularization and also more traditional approaches.\n\n- Weaknesses: - The paper can be quite hard to follow, see the clarity section below. -  Overall, the framework seems incremental. There does not seem to be a big distinction from [1], aside from parameterizing the dirichlet energy with learnable weights. The theoretical contributions appear minor and mirror those previously established in [2]. \n\n**Clarity:** The paper was hard to read and contained a number of typos. I have noted some here, but the manuscript contains others:\n- page 1 abstract: “requires a regularization” -> “requires regularization”\n- page 1: “over-parametric” -> “overparameterized”\n- page 2: “The low-rank” -> “Low-rankness”\n- page 2: “Furthermore, We” -> “Furthermore, we”\n- page 3: The functionals $\\mathcal{R}_{W_i}$ are missing the trace $\\text{tr}( \\cdot )$.\n- page 3: “Another expected” -> “Another unexpected”\n- page 5: “Specialy, we sigh two of $L_c(t=4000)$ out”\n- page 6: “Air-Net Adaptive to Both Varies Data” should “varies” be “varying”?\n- page 6 Figure 1: “the $t$-th row” -> “the $i$-th row”\n- page 7: “reminding missing” -> “remaining missing”\n\n**Novelty and significance:** The work seems to be an extension of [1], which also employed a deep matrix factorization approach to solve matrix completion problems and incorporated the dirichlet energy as an explicit regularizer. In that case, however, the laplacian was not learned, which seems to be the crucial difference. Some components of the theory is also borrowed from Arora et al [2].\n\n**General comments:**\n\n- At the top of page 3, it says that the structure of the network parameterizing the Laplacian matrix is discussed in A.4, but it appears to actually be in A.3. There is also no discussion on why the network is chosen in this particular fashion. It would be good to talk about why this network structure is used, e.g. the use of a softmax-type parameterization. Also, under this parameterization, is it even possible for the learned $L_i = 0$, the trivial solution?\n- It feels misleading to say in Remark 1 that Theorem 2 requires no restriction on $\\mathcal{T}_i$ or $X$. The first assumption is that all columns of $\\mathcal{T}_i(X)$ are unit normed and the entries are positive. If, for example, $\\mathcal{T}_i(X) := X$, then this means that for the result to hold, the columns of $X$ are unit normed and all of its entries are non-negative.\n- The matrix $C_{k,l}$ should be defined in Theorem 2, since it arises in the definition of $\\gamma$.\n- I am a bit puzzled by Figure 1. This seems to indicate that as time progresses, the network learns that there should be less and less similarity between rows/columns, as indicated by the lack of dark regions as $t$ increases. What is the intuition for this? Shouldn't some notion of self-similarity be learned instead?\n\n[1] Amit Boyarski, Sanketh Vedula, and Alex Bronstein. Spectral geometric matrix completion. MSML 2021\n\n[2] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS 2019", "summary_of_the_review": "Overall, while the proposed work shows empirical promise, it seems to be an incremental improvement over previous work. The readability of the manuscript can also be significantly improved. Based on these factors, my current score is a 5. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635873445163}, {"id": "HwZE6mqbJCY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2533/Reviewer_MBKQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a generalization and extension to deep matrix factorization as presented in the former paper from [Arora et al, NeurIPS 2019]. The extension allows more complex model which include inverse problems. The generalization part is build on a \"vanishing\" regularization which leads to better dynamics and convergence. These results are given by a theoretical analysis which highlight the effect of the proposed model. Finally the experiments illustrate the advantage of the model compared to state-of-art methods.\n", "review_text": "##########################################################################\nSummary:\n\nThis paper proposes a generalization and extension to deep matrix factorization as presented in the former paper from [Arora et al, NeurIPS 2019]. The extension allows more complex model which include inverse problems. The generalization part is build on a \"vanishing\" regularization which leads to better dynamics and convergence. These results are given by a theoretical analysis which highlight the effect of the proposed model. Finally the experiments illustrate the advantage of the model compared to state-of-art methods.\n\n\n##########################################################################\nReasons for score: \n\nDespite the interesting results in both theory and experiments, the presentation of the paper is too confused for publication (see cons below). Too many important details are in the appendix.\n\n##########################################################################\nPros: \n\n* The theoretical results are very interesting as they greatly improve the previous ones on deep matrix factorization.\n\n* The Dirichlet Energy regularization term leading \"vanishing\" prior on the data is intriguing and a theoretically sound.\n\n* The model is extensive to inverse problems and others. Important for application in image processing...\n\n* The experiments shows clearly the advantage of the proposed model.\n\n##########################################################################\nCons: \n\n* Eq. (1) is an optimization problem not a model. The true model is hidden in the different part of the appendix (A.2 for the matrix factorization, A.3 for the regularization). Thus it is hard to understand (or even read) most of the paper since we have to switch between all the pages to catch the important pieces of the model.\n\n* The theorems are very difficult to read as important information are in the appendix. Theorem 1 rely in on assumption which should be add to the text (I think there some place left). Theorem 2 is based on a regularization model which is only described in the A.3.\n\n\n##########################################################################\nQuestions during rebuttal period: \n\nPlease address and clarify the cons above \n\n#########################################################################\nSome typos and others: \n\n* Some sentences are confusing or need to be reshaped.\n    - At the end of section 2.1, the details of the implicit low-rank are presented. It not a discussion.\n    - In Theorem 2 at the end, I would rather write: \"D is a constant which equals [...]\".\n    - End of page 4, \"it's not difficult to find $R_{W_i^*}=0$\" is confusing, I would rather write that \"$R_{W_i^}(t)$ converge to 0\".\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a generalization and extension to deep matrix factorization as presented in the former paper from [Arora et al, NeurIPS 2019]. The extension allows more complex model which include inverse problems. The generalization part is build on a \"vanishing\" regularization which leads to better dynamics and convergence. These results are given by a theoretical analysis which highlight the effect of the proposed model. Finally the experiments illustrate the advantage of the model compared to state-of-art methods.\n", "main_review": "##########################################################################\nSummary:\n\nThis paper proposes a generalization and extension to deep matrix factorization as presented in the former paper from [Arora et al, NeurIPS 2019]. The extension allows more complex model which include inverse problems. The generalization part is build on a \"vanishing\" regularization which leads to better dynamics and convergence. These results are given by a theoretical analysis which highlight the effect of the proposed model. Finally the experiments illustrate the advantage of the model compared to state-of-art methods.\n\n\n##########################################################################\nReasons for score: \n\nDespite the interesting results in both theory and experiments, the presentation of the paper is too confused for publication (see cons below). Too many important details are in the appendix.\n\n##########################################################################\nPros: \n\n* The theoretical results are very interesting as they greatly improve the previous ones on deep matrix factorization.\n\n* The Dirichlet Energy regularization term leading \"vanishing\" prior on the data is intriguing and a theoretically sound.\n\n* The model is extensive to inverse problems and others. Important for application in image processing...\n\n* The experiments shows clearly the advantage of the proposed model.\n\n##########################################################################\nCons: \n\n* Eq. (1) is an optimization problem not a model. The true model is hidden in the different part of the appendix (A.2 for the matrix factorization, A.3 for the regularization). Thus it is hard to understand (or even read) most of the paper since we have to switch between all the pages to catch the important pieces of the model.\n\n* The theorems are very difficult to read as important information are in the appendix. Theorem 1 rely in on assumption which should be add to the text (I think there some place left). Theorem 2 is based on a regularization model which is only described in the A.3.\n\n\n##########################################################################\nQuestions during rebuttal period: \n\nPlease address and clarify the cons above \n\n#########################################################################\nSome typos and others: \n\n* Some sentences are confusing or need to be reshaped.\n    - At the end of section 2.1, the details of the implicit low-rank are presented. It not a discussion.\n    - In Theorem 2 at the end, I would rather write: \"D is a constant which equals [...]\".\n    - End of page 4, \"it's not difficult to find $R_{W_i^*}=0$\" is confusing, I would rather write that \"$R_{W_i^}(t)$ converge to 0\".\n", "summary_of_the_review": "This paper presents interesting results in both theory and experiments. Howeever the presentation of the paper is too confused for publication. Too many important details are in the appendix.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635869838626}], "openreview_url": "https://openreview.net/forum?id=xf0B7-7MRo6", "arxiv_id": "2110.07557", "paper_pdf": "papers/xf0B7-7MRo6.pdf", "paper_pdf_sha256": "e64eb006c1d752e830a60f503fde87f744093117f0e2cddfcd8a4b4984b0f9ab", "paper_pdf_bytes": 3322931, "paper_pdf_source": "openreview", "code_url": "https://github.com/lizhemin15/AIR-Net", "code_repository": "lizhemin15/AIR-Net", "code_commit": "2771c101586e39b0b20748ca0eb0aca35a4e5f47", "code_archive": "repos/xf0B7-7MRo6.zip", "code_archive_sha256": "900f9a4a9cc86230931b1be98cb6746cac3630a8afba0de5d77b5545bf542770", "code_archive_bytes": 1260764, "code_file_count": 18, "code_extensions": {".py": 16, ".ipynb": 2}, "github_disk_usage_kb": 785, "github_languages": {"Jupyter Notebook": 1157062, "Python": 72059}, "github_archived": false, "github_pushed_at": "2023-02-03T03:01:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/air-net-adaptive-and-implicit-regularization-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qoTcTS9-IZ-", "year": 2021, "status": "rejected", "title": "Dynamically Stable Infinite-Width Limits of Neural Classifiers", "authors": ["Eugene Golikov"], "authorids": ["~Eugene_Golikov1"], "authors_source": "OpenReview API", "abstract": "Recent research has been focused on two different approaches to studying neural networks training in the limit of infinite width (1) a mean-field (MF) and (2) a constant neural tangent kernel (NTK) approximations. These two approaches have different scaling of hyperparameters with the width of a network layer and as a result, different infinite-width limit models. Restricting ourselves to single hidden layer nets with zero-mean initialization trained for binary classification with SGD, we propose a general framework to study how the limit behavior of neural models depends on the scaling of hyperparameters with network width. Our framework allows us to derive scaling for existing MF and NTK limits, as well as an uncountable number of other scalings that lead to a dynamically stable limit behavior of corresponding models. However, only a finite number of distinct limit models are induced by these scalings. Each distinct limit model corresponds to a unique combination of such properties as boundedness of logits and tangent kernels at initialization or stationarity of tangent kernels. Existing MF and NTK limit models, as well as one novel limit model, satisfy most of the properties demonstrated by finite-width models. We also propose a novel initialization-corrected mean-field limit that satisfies all properties noted above, and its corresponding model is a simple modification for a finite-width model.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "3Mno_f9gMj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1650/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nMain topic of the paper is to study various infinite width limits. The paper notices different scaling used in NTK limit and MF limit and proposes a general framework for studying the limit behavior depending on the scaling. This allows the authors to define more general dynamically stable models in the infinite width limit. \n\nStated contribution by authors are:\n\n- Framework for reasoning about scaling that leads to ‘dynamically stable’ model evolution in the infinite width limit. \n- Categorization of 13 distinct dynamically stable models in the infinite width limit.\n- Characterization of “sym-default” model, which along with NTK and MF limit which shows properties most of finite-width network evolution. \n- Model modification “Initialization-corrected mean-field limit”(IC-MF) that satisfy all identified property of finite-width network evolution\n- Demonstration of IC-MF limit approximating finite-width network the best among all other possible models. \n \nReason for score:\n\nThe paper proposes an interesting theoretical framework to capture various infinite width limits. While settings are limited to make theory tractable, there is some empirical validation as well as capacity to broaden the study to different infinite width limits. I believe there are interesting novelties to be shared among ICLR participants who are interested in deep learning theory. \n\nPros:\nGeneral framework encapsulates widely studied infinite width limits and generalizes them which would be useful for theoretical study of large neural networks.\n\nThe paper makes predictions on scaling limits and finds a limit that could agree with finite networks better (IC-MF). For a variety of experiments the fact that this class matches with a reference finite width network is demonstrated well. \n\nWhile the proposed analysis is limited to a single hidden layer case, the authors have described thoughtful possibilities how the framework could potentially generalize to deeper networks. \nAuthors shared code to reproduce their experiments in the paper which is useful for reproduction and clearly understand different proposed scalings.\n\nCons:\n\nBeing single hidden layer analysis is a drawback (while discussion on extension is appreciated and stated in the “pros”). Often multi-layer infinite width is more interesting since the number of hidden layer weights scales quadratically while only the input and output layer weight count scales linearly with width. This can induce different training dynamics and scaling, so limitation to single hidden layer analysis is a drawback. \n\nTo simplify proof non-linearity of the choice is “leaky softplus” which is not widely used in practice. Having general comments on extension or limitation to well-known activations is advised. \n\nAssumption in p3 stating “gradients for a and W are estimated using independent data samples” may be too strong and not realistic. I cannot find no validation that the assumption is a reasonable one to make beyond making the theory simple. This may lead to inconsistencies raised in the “Questions” below. \n\nExperiments demonstrating similar effect for more realistic setting (e.g. full CIFAR-10 classification task achieving reasonable performance) would have been better to convince the validity or generalizability of the theory. While one can see that IC-MF and reference agrees well for the training curve, for test-set since every curve essentially behaves the same it is hard to determine which limit is superior or not. \n\nQuestions:\nIn [Liu et al., 2020], they show non-constancy of NTK for non-linear output including the soft-max layer in the outputs. I believe this fact may be in conflict with this paper’s analysis on binary cross-entropy loss and still having constant NTK. I do suspect it may be due to assumptions on independent gradients for `a` and `W` in page 3. Do you think in a realistic training setting with softmax-cross entropy, you can generalize the analysis and reconcile the fact that NTK is non-constant?\n\nIn general, does the generalization of “dynamical stability” to admit infinite logits valid for multi-class cases? In multi-class simple “sign” can’t be used for classification and I wonder if authors have ideas on generalization to multi-class settings. \n\nIn eq (26), why is Gaussian distribution over logit a good way to measure KL divergence? I understand at initialization prior is distributed over Gaussian, however after training with softmax I expect the logit distribution is no-longer Gaussian. \n\nNits and additional feedback:\n\nNit: use \\citep in places where appropriate.\n\nDivision (/) for denoting “or” is often confusing and would suggest other notation. Especially in the Figures and Condition 2 when sometimes it means “or” and sometimes it means division.\n\np4: ‘grows width’ -> ‘grows with’\n\n[Sohl-Dickstein et al., 2020] show different scaling of weight and bias scaling extending standard parameterization to work well with NTK limit. While this paper mostly works with the NTK parameterization, it would be interesting to discuss how the improved parameterization in [Sohl-Dickstein et al., 2020] can be utilized. I suspect this is quite related to (9) / (10) where definition of NTK deviates from [Jacot et al., 2018]. \n\nIn Section 3: When comparing performance on finite networks and NTK, it can be subtle depending on how one trains finite networks and there’s architecture dependence [Lee et al., 2020] \nIn Section 4: suggest citing [Chizat et al., 2019] for ‘lazy training’\n\nChizat et al., On Lazy Training in Differentiable Programming, NeurIPS 2019\nLee et al., Finite Versus Infinite Neural Networks: an Empirical Study, arXiv: 2007.15801\nLiu et al., On the linearity of large non-linear models: when and why the tangent kernel is constant, arXiv:2010.01092\nSohl-Dickstein et al., On the infinite width limit of neural networks with a standard parameterization, arXiv:2001.07301\n\n----\n[post rebuttal]\nI thank the authors for their revision and clarifying many of my questions and adding results based on that. I have read other reviewers concern and while I agree some room for improvement on clarity, I still think the paper brings in value. Unless there's technical flaws spotted by other reviewers that has not been resolved, I'm still leaning towards accepting (increased score from 6 to 7). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "Summary:\n\nMain topic of the paper is to study various infinite width limits. The paper notices different scaling used in NTK limit and MF limit and proposes a general framework for studying the limit behavior depending on the scaling. This allows the authors to define more general dynamically stable models in the infinite width limit. \n\nStated contribution by authors are:\n\n- Framework for reasoning about scaling that leads to ‘dynamically stable’ model evolution in the infinite width limit. \n- Categorization of 13 distinct dynamically stable models in the infinite width limit.\n- Characterization of “sym-default” model, which along with NTK and MF limit which shows properties most of finite-width network evolution. \n- Model modification “Initialization-corrected mean-field limit”(IC-MF) that satisfy all identified property of finite-width network evolution\n- Demonstration of IC-MF limit approximating finite-width network the best among all other possible models. \n \nReason for score:\n\nThe paper proposes an interesting theoretical framework to capture various infinite width limits. While settings are limited to make theory tractable, there is some empirical validation as well as capacity to broaden the study to different infinite width limits. I believe there are interesting novelties to be shared among ICLR participants who are interested in deep learning theory. \n\nPros:\nGeneral framework encapsulates widely studied infinite width limits and generalizes them which would be useful for theoretical study of large neural networks.\n\nThe paper makes predictions on scaling limits and finds a limit that could agree with finite networks better (IC-MF). For a variety of experiments the fact that this class matches with a reference finite width network is demonstrated well. \n\nWhile the proposed analysis is limited to a single hidden layer case, the authors have described thoughtful possibilities how the framework could potentially generalize to deeper networks. \nAuthors shared code to reproduce their experiments in the paper which is useful for reproduction and clearly understand different proposed scalings.\n\nCons:\n\nBeing single hidden layer analysis is a drawback (while discussion on extension is appreciated and stated in the “pros”). Often multi-layer infinite width is more interesting since the number of hidden layer weights scales quadratically while only the input and output layer weight count scales linearly with width. This can induce different training dynamics and scaling, so limitation to single hidden layer analysis is a drawback. \n\nTo simplify proof non-linearity of the choice is “leaky softplus” which is not widely used in practice. Having general comments on extension or limitation to well-known activations is advised. \n\nAssumption in p3 stating “gradients for a and W are estimated using independent data samples” may be too strong and not realistic. I cannot find no validation that the assumption is a reasonable one to make beyond making the theory simple. This may lead to inconsistencies raised in the “Questions” below. \n\nExperiments demonstrating similar effect for more realistic setting (e.g. full CIFAR-10 classification task achieving reasonable performance) would have been better to convince the validity or generalizability of the theory. While one can see that IC-MF and reference agrees well for the training curve, for test-set since every curve essentially behaves the same it is hard to determine which limit is superior or not. \n\nQuestions:\nIn [Liu et al., 2020], they show non-constancy of NTK for non-linear output including the soft-max layer in the outputs. I believe this fact may be in conflict with this paper’s analysis on binary cross-entropy loss and still having constant NTK. I do suspect it may be due to assumptions on independent gradients for `a` and `W` in page 3. Do you think in a realistic training setting with softmax-cross entropy, you can generalize the analysis and reconcile the fact that NTK is non-constant?\n\nIn general, does the generalization of “dynamical stability” to admit infinite logits valid for multi-class cases? In multi-class simple “sign” can’t be used for classification and I wonder if authors have ideas on generalization to multi-class settings. \n\nIn eq (26), why is Gaussian distribution over logit a good way to measure KL divergence? I understand at initialization prior is distributed over Gaussian, however after training with softmax I expect the logit distribution is no-longer Gaussian. \n\nNits and additional feedback:\n\nNit: use \\citep in places where appropriate.\n\nDivision (/) for denoting “or” is often confusing and would suggest other notation. Especially in the Figures and Condition 2 when sometimes it means “or” and sometimes it means division.\n\np4: ‘grows width’ -> ‘grows with’\n\n[Sohl-Dickstein et al., 2020] show different scaling of weight and bias scaling extending standard parameterization to work well with NTK limit. While this paper mostly works with the NTK parameterization, it would be interesting to discuss how the improved parameterization in [Sohl-Dickstein et al., 2020] can be utilized. I suspect this is quite related to (9) / (10) where definition of NTK deviates from [Jacot et al., 2018]. \n\nIn Section 3: When comparing performance on finite networks and NTK, it can be subtle depending on how one trains finite networks and there’s architecture dependence [Lee et al., 2020] \nIn Section 4: suggest citing [Chizat et al., 2019] for ‘lazy training’\n\nChizat et al., On Lazy Training in Differentiable Programming, NeurIPS 2019\nLee et al., Finite Versus Infinite Neural Networks: an Empirical Study, arXiv: 2007.15801\nLiu et al., On the linearity of large non-linear models: when and why the tangent kernel is constant, arXiv:2010.01092\nSohl-Dickstein et al., On the infinite width limit of neural networks with a standard parameterization, arXiv:2001.07301\n\n----\n[post rebuttal]\nI thank the authors for their revision and clarifying many of my questions and adding results based on that. I have read other reviewers concern and while I agree some room for improvement on clarity, I still think the paper brings in value. Unless there's technical flaws spotted by other reviewers that has not been resolved, I'm still leaning towards accepting (increased score from 6 to 7). ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603861111013}, {"id": "f36_I7ZdNTV", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1650/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThis paper investigates the dynamics of the training of fully connected neural networks with one hidden layer in the infinite-width limit, for classification. Starting from the observation [Golikov, 2020] that the mean-field limit and the NTK regime are only two special cases of continuous families of scalings, it attempts at identifying important or desirable features that an infinite-width limit should display in order to give good insight into the dynamics of large but finite neural networks. \n\nThis is an exciting program, and it looks like the ideas are very good. If / when laid out, they will really bring insight into the training of large neural nets.\n\nUnfortunately, the way the paper is written makes it hard to accept in its present form. While the global structure is quite clear, the wording makes it very hard to extract information. For instance, it is not clear what is mean by \"finite\": does it mean there is a finite limit (what should be expected in principle), that it is bounded from above (what it seems to mean sometimes) or that it is strictly positive (what it appears to mean sometimes)? Also, there is a classification into 13 cases that are promised, but they are not clearly listed in the main, the fact that we are working with only one hidden layer is not clearly said in the abstract (and the discussion on how to extend to more hidden layers is not convincing). The grammar is somehow problematic (there is in particular a lot of missing \"the\" articles in front of nouns), and it sometimes makes it hard to follow. Also, it is not very clear what is proven, where assumptions are used (e.g. analyticity of the nonlinearity), the seemingly most interesting regimes are not defined in the main, etc.  \n\nI am tempted to think that the authors are very lucid about their understanding of what is happening and that they really have interesting something to convey, but the present version makes it very hard to get a reliable information. A (very significantly) revised version of this paper could, I believe, bring much insight to our understanding of neural nets. There is a lot of potential with this paper, just not realized in terms of exposition.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An ambitions, potentially exciting paper, which is unfortunately impossible to follow", "review": "\nThis paper investigates the dynamics of the training of fully connected neural networks with one hidden layer in the infinite-width limit, for classification. Starting from the observation [Golikov, 2020] that the mean-field limit and the NTK regime are only two special cases of continuous families of scalings, it attempts at identifying important or desirable features that an infinite-width limit should display in order to give good insight into the dynamics of large but finite neural networks. \n\nThis is an exciting program, and it looks like the ideas are very good. If / when laid out, they will really bring insight into the training of large neural nets.\n\nUnfortunately, the way the paper is written makes it hard to accept in its present form. While the global structure is quite clear, the wording makes it very hard to extract information. For instance, it is not clear what is mean by \"finite\": does it mean there is a finite limit (what should be expected in principle), that it is bounded from above (what it seems to mean sometimes) or that it is strictly positive (what it appears to mean sometimes)? Also, there is a classification into 13 cases that are promised, but they are not clearly listed in the main, the fact that we are working with only one hidden layer is not clearly said in the abstract (and the discussion on how to extend to more hidden layers is not convincing). The grammar is somehow problematic (there is in particular a lot of missing \"the\" articles in front of nouns), and it sometimes makes it hard to follow. Also, it is not very clear what is proven, where assumptions are used (e.g. analyticity of the nonlinearity), the seemingly most interesting regimes are not defined in the main, etc.  \n\nI am tempted to think that the authors are very lucid about their understanding of what is happening and that they really have interesting something to convey, but the present version makes it very hard to get a reliable information. A (very significantly) revised version of this paper could, I believe, bring much insight to our understanding of neural nets. There is a lot of potential with this paper, just not realized in terms of exposition.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603841732387}, {"id": "_ZC206F-Kb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1650/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a general framework to derive different stable limiting behaviors of the dynamics of two-layers neural networks, under different parameterization of the hyper-parameters. For certain choices of hyper-parameters, this recovers the mean-field limit and the NTK limit. This paper also proposed certain properties of the limiting dynamics and showed that using these properties as the classification criteria, there are only a finite number of distinct models in the limit. This paper also proposed a novel initialization-corrected mean-field limit that satisfies all properties.\nThis paper tells an interesting story. The question is whether the problem solved by the story is important.\nThe main consideration of this paper is to find regimes of hyper-parameters such that the limiting dynamics are stable in some sense. For this purpose, the authors advocate the IC-MF regime, such that the limiting dynamics are stable with respect to all the conditions that the authors proposed. This story seems to be self-contained on its own. However, I believe that two more important criteria for good training algorithms are optimization and generalization efficiency, which are not discussed by the authors.\nThe optimization and generalization efficiency are much more important than the stability condition in practice. There could be some regimes of hyper-parameters that do not satisfy the stability condition, but has good optimization and generalization efficiency. The IC-MF regime proposed in this paper, although seems to satisfy additional stability conditions, but intuitively, it seems that its generalization efficiency is not as good as that of MF regime: it added an additional noisy function $f_{ntk, \\infty}^{0)}$ to the mean-field prediction function, and this additional noise (intuitively will) hurt generalization.\nI believe some of these stability properties should have some connection to optimization and generalization. For example, if some simple stability property is violated, the algorithm cannot generalize well. I feel that the authors should try to build connections of stability properties to optimization and generalization, to justify the importance of condition 1 and condition 2 defined in the paper.\nAbove all, I feel that this paper is interesting in its own criteria. However, it didn't justify that its criteria are important. So I feel that this paper is on the borderline.\n\nMinor issues:\n\t1.\tSome notations are easy to get readers confused. Eq. (7), $\\sigma(d) = \\sigma^*(d / d^*)^{q_\\sigma}$. Here $\\sigma$ is a function of $d$ while $\\sigma^*$ is a scaler (not as a function of $d / d^*$). It takes me while to understand this.\n\t2.\tTypos: page 18: in this case $1 + \\tilde q + 2 q_\\sigma$.\n\t3.\tThe notations of this paper looks very complicated, especially the superscripts and subscripts.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting story; but the stability conditions defined in this paper are not justified to be important ", "review": "This paper proposed a general framework to derive different stable limiting behaviors of the dynamics of two-layers neural networks, under different parameterization of the hyper-parameters. For certain choices of hyper-parameters, this recovers the mean-field limit and the NTK limit. This paper also proposed certain properties of the limiting dynamics and showed that using these properties as the classification criteria, there are only a finite number of distinct models in the limit. This paper also proposed a novel initialization-corrected mean-field limit that satisfies all properties.\nThis paper tells an interesting story. The question is whether the problem solved by the story is important.\nThe main consideration of this paper is to find regimes of hyper-parameters such that the limiting dynamics are stable in some sense. For this purpose, the authors advocate the IC-MF regime, such that the limiting dynamics are stable with respect to all the conditions that the authors proposed. This story seems to be self-contained on its own. However, I believe that two more important criteria for good training algorithms are optimization and generalization efficiency, which are not discussed by the authors.\nThe optimization and generalization efficiency are much more important than the stability condition in practice. There could be some regimes of hyper-parameters that do not satisfy the stability condition, but has good optimization and generalization efficiency. The IC-MF regime proposed in this paper, although seems to satisfy additional stability conditions, but intuitively, it seems that its generalization efficiency is not as good as that of MF regime: it added an additional noisy function $f_{ntk, \\infty}^{0)}$ to the mean-field prediction function, and this additional noise (intuitively will) hurt generalization.\nI believe some of these stability properties should have some connection to optimization and generalization. For example, if some simple stability property is violated, the algorithm cannot generalize well. I feel that the authors should try to build connections of stability properties to optimization and generalization, to justify the importance of condition 1 and condition 2 defined in the paper.\nAbove all, I feel that this paper is interesting in its own criteria. However, it didn't justify that its criteria are important. So I feel that this paper is on the borderline.\n\nMinor issues:\n\t1.\tSome notations are easy to get readers confused. Eq. (7), $\\sigma(d) = \\sigma^*(d / d^*)^{q_\\sigma}$. Here $\\sigma$ is a function of $d$ while $\\sigma^*$ is a scaler (not as a function of $d / d^*$). It takes me while to understand this.\n\t2.\tTypos: page 18: in this case $1 + \\tilde q + 2 q_\\sigma$.\n\t3.\tThe notations of this paper looks very complicated, especially the superscripts and subscripts.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603841071960}, {"id": "9t_VGS5ys3G", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1650/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper analyzes joint scalings of the parameter initialization and the learning rate, with respect to the limit of infinite width, in the context of two-layer neural networks with stochastic gradient descent and binary logistic loss. It proposes some “dynamically stable” conditions and identifies a range of scalings that satisfy these conditions. This covers the neural tangent kernel (NTK) and mean field (MF) scalings, as well as others. The paper then proposes to add an extra correction function to the initialization of the MF limit and argues experimentally that this correction can give a proxy for standard neural networks.\n\nOverall there are interesting points in the paper, but I think the paper needs a lot more works to make the study rigorous, complete and easier to read. More analyses into why any point in the “dynamically stable model evolution” band leads to stable dynamics at all training time are needed.\n\n\n### Positive points:\nThe recent trend in analysis of large-width neural networks has drawn a lot of attention, so the paper is timely. \n\nThe idea of joint scalings between the initialization and the learning rate, w.r.t. the width d, has been explored before; but here the paper proposes to do so in relation with a new set of criteria. The results make an interesting view on the connection between two distinct regimes NTK and MF, and beyond.\n\n\n### Negative points:\n\n**Presentation:**\nThere are many unnecessary notations, for example, \\tilde{q}_a and \\tilde{q}_w which are equal to each other in the main analysis. It is better to keep all discussions around just the two key scaling exponents. \n\nThe main figure, Figure 1, is very hard to read, and I cannot see where the curve for MF is in this plot. \n\nThe statements also seem to miss several assumptions e.g. assumptions on the data (will things hold if |x| ~ exp(d)??).\n\nSome notations are not explained, for example, \\sigma^{(0)}(x) that appears in Eq (13).\n\n**Incomplete proofs:**\nAside from many derivations which are heuristic (which are fine as long as they are not stated as propositions/theorems), the proofs for the stated lemmas / propositions are unrigorous and may entirely omit the more difficult technical points. For example, in the proof of Lemma 1.4 in Appendix B.1, why is it that the collection of the weights \\hat{w}_r^{(k), r=1…,d, at any time k>1 allows one to apply the law of large numbers? The variables in this collection are not independent; if they are mildly dependent, in what sense are they so that one can apply the law of large numbers?\n\nIn fact, there are simple mathematical mistakes throughout the proofs. For example, to prove Lemma 1.2 in Appendix B.1, the paper only proves the statement for a single fixed x; why does it hold for almost every x?\n\n**Abstract missing key information:**\nThe abstract should be more precise: it should mention logistic loss for binary classification and that the initializations are zero-mean. These are important; without them, a number of key results will fail.\n\n**Dynamical stability criteria are insufficiently justified; most are limited to just initializations:**\nThe paper does not give sufficient justification of why the proposed criteria (Conditions 1 and 2) should be considered. In particular, Condition 2 only concerns with what happens at initialization; does it guarantee stable dynamics well beyond initialization and in what sense? The paper does not discuss this for most points in the “dynamically stable model evolution” band. In fact, compared to the well-studied NTK and MF scalings, the newly identified scaling sym-default is said to have some divergence behavior (in which its prediction function is always infinite, for infinite d, at any positive training time). If anything, the paper should justify why this divergence behavior does not lead to instability; by looking at Figure 1, I think it suffers from numerical instability.\n\nIn fact, scalings other than NTK and MF have been studied well beyond initialization in a rigorous fashion. In particular, [1] shows that with a wide range of different scalings, it’s entirely possible to have the same stable behavior: NTK behavior where weights do not move and the model tracks some random-feature models. Given that such rigorous and thorough study is possible, it is unclear how the proposal of Conditions 1 and 2 leads to novel and significant insights.\n\n**Condition 2.3 is unusual in terms of physical units and not motivated:**\nIn particular, why should we compare the magnitude of the kernel K with the magnitude of the prediction function f, while they should have very different physical interpretation? Why is this condition interesting? The second paragraph in Section 3 seems to give some arguments for Condition 2.1 and 2.2, but none for Condition 2.3.\n\nGiven that the role of Condition 2.3 is unclear, if one removes this condition from the diagram of Figure 1, the band picture disappears and we are effectively left with a half plane, separated by the “evolving kernels” boundary. To the left of this boundary, it is the NTK behavior as expected in [1]. So unless Condition 2.3 signifies some significant and meaningful change in network behaviors, the diagram would not show new significant insights compared to the literature.\n\n**The derivations for the dynamics in Section 3 are incomplete:**\nIn the regime of small learning rate (\\hat{\\eta}* goes to 0), what we should obtain is a continuous time evolution with an expectation over data, not a discrete-time one with stochastically drawn data. Evidently Eq (14) as given in the paper completely contradicts all previous published reports on the MF limit.\n\nIt does not make sense to claim, after Equation (19), that the prediction function diverges from iteration k=1 onwards. Firstly, as said, the analysis has to be done with respect to small learning rate and hence the iteration k has to be determined in relation with the learning rate. Secondly, that claim contradicts Figure 1: the prediction function does not seem to diverge at iteration 1.\n\n**The initialization-corrected mean field (IC-MF) limit lacks justification:**\nIts aim is to correct the MF limit w.r.t. Condition 2.1, but there are very simple alternatives to do this correction. The first way is to add any fixed function, whose magnitude are independent of d in suitable sense, to the MF limit. The second way is to initialize the MF limit with non-zero mean distributions; in fact, it is known that non-zero mean initializations are more typical in MF limit.\n\nThe paper argues that IC-MF limit can be good proxy for standard networks, giving only one simple experiment shown in Figure 1. There is no proof or mathematical heuristics provided. I think this IC-MF idea does not have anything to do with the theory in the previous sections, nor the binary classification problem with logistic loss. As such, it should be at least further tested on more complex experimental tasks, e.g. CIFAR-10. From a mathematical standpoint, I do not foresee a simple argument to show why it can be a good proxy for standard networks.\n\n\n[1] A comparative analysis of the optimization and generalization property of two-layer neural network and random feature models under gradient descent dynamics, E, Ma and Wu, 2019.\n\n---------------------\n\nPost rebuttal: \n\nI thank the authors for their rebuttal. Let me focus my reply on a few important points. I first thank the authors for clarifying the meaning of Condition 2. In this sense, Condition 1 is the key main contribution; however the current proof does not look correct to me, and the revised argument is far from being sufficient. In particular:\n\n- Point 5 of the rebuttal: I think the revised argument here is incomplete. The given argument concerns trivial facts and does not imply the claim. For example, what if the distribution of the terms is symmetric, the expected sum is zero, and hence the quantity might be of order smaller than d? Note that this is an example problem; there are multiple problems with the proof of Lemma 1.4. For example, the paper claims this for all k; but if k is something like d^100, would things hold? What would stop the magnitude of the weights to grow with time?\n\n- Point 6 of the rebuttal: The CLT, when applied w.r.t. the randomness of the weights, says that for a fixed $x$, $\\sum_{r=1}^d \\hat{a}_r \\phi (\\hat{w}_r x) / \\sqrt{d} \\sim N(0,v_x)$ approximately. That is, there is a non-zero probability (w.r.t. the randomness of weights) that the claim in the paper for a fixed $x$ fails. As such, to reason the claim for many $x$, one requires doing probabilistic arguments very carefully.\n\nThe paper should execute the proof very carefully. It is not just a matter of technicality; I suspect some of the claims are actually wrong.\n\nMore importantly, Condition 1 alone is insufficient to claim dynamical stability at any time k. What should qualify for dynamical stability is rather the existence of a well-defined limiting dynamics exists (which is argued heuristically in Appendix C), and its proof. In the current writing, it’s unclear how Condition 2 is crucial; while it studies interesting properties, it is very restrictive.\n\nAs said in my last review, one thing that has been missing is really whether the insight here differs qualitatively from the known NTK and MF limits. Further looking at the limiting dynamics in Appendix C, one sees that they are qualitatively either NTK or MF. There are possible degeneracies due to scalings and the use of logistic loss, but these do not lead to much deviation from NTK or MF behaviors. If one is to use a squared loss for instance, what one would obtain in Figure 1 is just the line connecting NTK and MF; all other points in the band outside this line are degeneracies due to logistic loss. The behavior on this line, again, is qualitatively either NTK or MF, and this is shown (somewhat implicitly) already by a number of past works.\n\nI would imagine a rigorous derivation of the limiting dynamics for each point in the band revolves around the renormalized dynamics in Appendix C. When translating from the renormalized dynamics to the original one, the extra scaling factors will complicate the proof (for instance, they can blow up Lipschitz constants). Again this has to be done very carefully.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting proposals, but the paper requires substantial revisions and additional analyses.", "review": "The paper analyzes joint scalings of the parameter initialization and the learning rate, with respect to the limit of infinite width, in the context of two-layer neural networks with stochastic gradient descent and binary logistic loss. It proposes some “dynamically stable” conditions and identifies a range of scalings that satisfy these conditions. This covers the neural tangent kernel (NTK) and mean field (MF) scalings, as well as others. The paper then proposes to add an extra correction function to the initialization of the MF limit and argues experimentally that this correction can give a proxy for standard neural networks.\n\nOverall there are interesting points in the paper, but I think the paper needs a lot more works to make the study rigorous, complete and easier to read. More analyses into why any point in the “dynamically stable model evolution” band leads to stable dynamics at all training time are needed.\n\n\n### Positive points:\nThe recent trend in analysis of large-width neural networks has drawn a lot of attention, so the paper is timely. \n\nThe idea of joint scalings between the initialization and the learning rate, w.r.t. the width d, has been explored before; but here the paper proposes to do so in relation with a new set of criteria. The results make an interesting view on the connection between two distinct regimes NTK and MF, and beyond.\n\n\n### Negative points:\n\n**Presentation:**\nThere are many unnecessary notations, for example, \\tilde{q}_a and \\tilde{q}_w which are equal to each other in the main analysis. It is better to keep all discussions around just the two key scaling exponents. \n\nThe main figure, Figure 1, is very hard to read, and I cannot see where the curve for MF is in this plot. \n\nThe statements also seem to miss several assumptions e.g. assumptions on the data (will things hold if |x| ~ exp(d)??).\n\nSome notations are not explained, for example, \\sigma^{(0)}(x) that appears in Eq (13).\n\n**Incomplete proofs:**\nAside from many derivations which are heuristic (which are fine as long as they are not stated as propositions/theorems), the proofs for the stated lemmas / propositions are unrigorous and may entirely omit the more difficult technical points. For example, in the proof of Lemma 1.4 in Appendix B.1, why is it that the collection of the weights \\hat{w}_r^{(k), r=1…,d, at any time k>1 allows one to apply the law of large numbers? The variables in this collection are not independent; if they are mildly dependent, in what sense are they so that one can apply the law of large numbers?\n\nIn fact, there are simple mathematical mistakes throughout the proofs. For example, to prove Lemma 1.2 in Appendix B.1, the paper only proves the statement for a single fixed x; why does it hold for almost every x?\n\n**Abstract missing key information:**\nThe abstract should be more precise: it should mention logistic loss for binary classification and that the initializations are zero-mean. These are important; without them, a number of key results will fail.\n\n**Dynamical stability criteria are insufficiently justified; most are limited to just initializations:**\nThe paper does not give sufficient justification of why the proposed criteria (Conditions 1 and 2) should be considered. In particular, Condition 2 only concerns with what happens at initialization; does it guarantee stable dynamics well beyond initialization and in what sense? The paper does not discuss this for most points in the “dynamically stable model evolution” band. In fact, compared to the well-studied NTK and MF scalings, the newly identified scaling sym-default is said to have some divergence behavior (in which its prediction function is always infinite, for infinite d, at any positive training time). If anything, the paper should justify why this divergence behavior does not lead to instability; by looking at Figure 1, I think it suffers from numerical instability.\n\nIn fact, scalings other than NTK and MF have been studied well beyond initialization in a rigorous fashion. In particular, [1] shows that with a wide range of different scalings, it’s entirely possible to have the same stable behavior: NTK behavior where weights do not move and the model tracks some random-feature models. Given that such rigorous and thorough study is possible, it is unclear how the proposal of Conditions 1 and 2 leads to novel and significant insights.\n\n**Condition 2.3 is unusual in terms of physical units and not motivated:**\nIn particular, why should we compare the magnitude of the kernel K with the magnitude of the prediction function f, while they should have very different physical interpretation? Why is this condition interesting? The second paragraph in Section 3 seems to give some arguments for Condition 2.1 and 2.2, but none for Condition 2.3.\n\nGiven that the role of Condition 2.3 is unclear, if one removes this condition from the diagram of Figure 1, the band picture disappears and we are effectively left with a half plane, separated by the “evolving kernels” boundary. To the left of this boundary, it is the NTK behavior as expected in [1]. So unless Condition 2.3 signifies some significant and meaningful change in network behaviors, the diagram would not show new significant insights compared to the literature.\n\n**The derivations for the dynamics in Section 3 are incomplete:**\nIn the regime of small learning rate (\\hat{\\eta}* goes to 0), what we should obtain is a continuous time evolution with an expectation over data, not a discrete-time one with stochastically drawn data. Evidently Eq (14) as given in the paper completely contradicts all previous published reports on the MF limit.\n\nIt does not make sense to claim, after Equation (19), that the prediction function diverges from iteration k=1 onwards. Firstly, as said, the analysis has to be done with respect to small learning rate and hence the iteration k has to be determined in relation with the learning rate. Secondly, that claim contradicts Figure 1: the prediction function does not seem to diverge at iteration 1.\n\n**The initialization-corrected mean field (IC-MF) limit lacks justification:**\nIts aim is to correct the MF limit w.r.t. Condition 2.1, but there are very simple alternatives to do this correction. The first way is to add any fixed function, whose magnitude are independent of d in suitable sense, to the MF limit. The second way is to initialize the MF limit with non-zero mean distributions; in fact, it is known that non-zero mean initializations are more typical in MF limit.\n\nThe paper argues that IC-MF limit can be good proxy for standard networks, giving only one simple experiment shown in Figure 1. There is no proof or mathematical heuristics provided. I think this IC-MF idea does not have anything to do with the theory in the previous sections, nor the binary classification problem with logistic loss. As such, it should be at least further tested on more complex experimental tasks, e.g. CIFAR-10. From a mathematical standpoint, I do not foresee a simple argument to show why it can be a good proxy for standard networks.\n\n\n[1] A comparative analysis of the optimization and generalization property of two-layer neural network and random feature models under gradient descent dynamics, E, Ma and Wu, 2019.\n\n---------------------\n\nPost rebuttal: \n\nI thank the authors for their rebuttal. Let me focus my reply on a few important points. I first thank the authors for clarifying the meaning of Condition 2. In this sense, Condition 1 is the key main contribution; however the current proof does not look correct to me, and the revised argument is far from being sufficient. In particular:\n\n- Point 5 of the rebuttal: I think the revised argument here is incomplete. The given argument concerns trivial facts and does not imply the claim. For example, what if the distribution of the terms is symmetric, the expected sum is zero, and hence the quantity might be of order smaller than d? Note that this is an example problem; there are multiple problems with the proof of Lemma 1.4. For example, the paper claims this for all k; but if k is something like d^100, would things hold? What would stop the magnitude of the weights to grow with time?\n\n- Point 6 of the rebuttal: The CLT, when applied w.r.t. the randomness of the weights, says that for a fixed $x$, $\\sum_{r=1}^d \\hat{a}_r \\phi (\\hat{w}_r x) / \\sqrt{d} \\sim N(0,v_x)$ approximately. That is, there is a non-zero probability (w.r.t. the randomness of weights) that the claim in the paper for a fixed $x$ fails. As such, to reason the claim for many $x$, one requires doing probabilistic arguments very carefully.\n\nThe paper should execute the proof very carefully. It is not just a matter of technicality; I suspect some of the claims are actually wrong.\n\nMore importantly, Condition 1 alone is insufficient to claim dynamical stability at any time k. What should qualify for dynamical stability is rather the existence of a well-defined limiting dynamics exists (which is argued heuristically in Appendix C), and its proof. In the current writing, it’s unclear how Condition 2 is crucial; while it studies interesting properties, it is very restrictive.\n\nAs said in my last review, one thing that has been missing is really whether the insight here differs qualitatively from the known NTK and MF limits. Further looking at the limiting dynamics in Appendix C, one sees that they are qualitatively either NTK or MF. There are possible degeneracies due to scalings and the use of logistic loss, but these do not lead to much deviation from NTK or MF behaviors. If one is to use a squared loss for instance, what one would obtain in Figure 1 is just the line connecting NTK and MF; all other points in the band outside this line are degeneracies due to logistic loss. The behavior on this line, again, is qualitatively either NTK or MF, and this is shown (somewhat implicitly) already by a number of past works.\n\nI would imagine a rigorous derivation of the limiting dynamics for each point in the band revolves around the renormalized dynamics in Appendix C. When translating from the renormalized dynamics to the original one, the extra scaling factors will complicate the proof (for instance, they can blow up Lipschitz constants). Again this has to be done very carefully.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603836246505}], "openreview_url": "https://openreview.net/forum?id=qoTcTS9-IZ-", "arxiv_id": "2006.06574", "paper_pdf": "papers/qoTcTS9-IZ-.pdf", "paper_pdf_sha256": "11de9c6887984a4b35709cf0d2974cef56e196eb6ddb5c9802e60f44bc247c0f", "paper_pdf_bytes": 1271126, "paper_pdf_source": "openreview", "code_url": "https://github.com/deeppavlov/research", "code_repository": "deeppavlov/research", "code_commit": "31a6c98099ef974d1ec7cb09186ca80dc3a95b34", "code_archive": "repos/qoTcTS9-IZ-.zip", "code_archive_sha256": "4b1952854a8bc02d92c5a0fbeddb97d94b2b259b3a0f701bd67a740a99d32716", "code_archive_bytes": 974047, "code_file_count": 19, "code_extensions": {".py": 16, ".ipynb": 3}, "github_disk_usage_kb": 955, "github_languages": {"Jupyter Notebook": 1309113, "Python": 138067}, "github_archived": false, "github_pushed_at": "2020-10-22T12:12:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dynamically-stable-infinite-width-limits-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ByeSYa4KPS", "year": 2020, "status": "rejected", "title": "Sparse Networks from Scratch: Faster Training without Losing Performance", "authors": ["Tim Dettmers", "Luke Zettlemoyer"], "authorids": ["dettmers@cs.washington.edu", "lsz@cs.washington.edu"], "authors_source": "OpenReview API", "abstract": "We demonstrate the possibility of what we call sparse learning: accelerated training of deep neural networks that maintain sparse weights throughout training while achieving dense performance levels. We accomplish this by developing sparse momentum, an algorithm which uses exponentially smoothed gradients (momentum) to identify layers and weights which reduce the error efficiently. Sparse momentum redistributes pruned weights across layers according to the mean momentum magnitude of each layer. Within a layer, sparse momentum grows weights according to the momentum magnitude of zero-valued weights. We demonstrate state-of-the-art sparse performance on MNIST, CIFAR-10, and ImageNet, decreasing the mean error by a relative 8%, 15%, and 6% compared to other sparse algorithms. Furthermore, we show that sparse momentum reliably reproduces dense performance levels while providing up to 5.61x faster training. In our analysis, ablations show that the benefits of momentum redistribution and growth increase with the depth and size of the network. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bkg1FkmQqS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper669/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose a sparse momentum algorithm for doing efficient sparse training. The technique relies on identifying weights in a layer that do not have an effect on the error, pruning them, and redistributing and growing them across layers. The technique is compared against other recent algorithms on a range of models.\n\nThe paper is well written, and very easy to read. The proposed algorithm looks interesting and seems to empirically work well. \n\nI am, however, a bit confused with how the sparse momentum algorithm is discussed in this paper. In the paper, momentum is intuitively presented as an algorithm that reduces the variance of the noise in the gradients. However, there are a number of papers that show that this is not the case (and if anything it is the opposite). Further, a few recent papers show that momentum works better than SGD only for learning rates that are not too small, and this is because it averages out the gradients in the high-curvature directions (these are the directions where the gradients switch signs) and makes them stable in these directions, thus allowing larger steps in the low curvature directions. See for example the following two papers:\nMomentum Enables Large Batch Training. Samuel L Smith, Erich Elsen, Soham De. ICML Workshop on Physics for Deep Learning, 2019.\nWhich Algorithmic Choices Matter at Which Batch Sizes? Insights From a Noisy Quadratic Model. Guodong Zhang, Lala Li, Zachary Nado, James Martens, Sushant Sachdeva, George E Dahl, Christopher J Shallue, Roger Grosse. NeurIPS 2019.\n\nGiven these papers, can the authors comment on what they think is the reason for the effectiveness of their sparse momentum algorithm? These papers seem to indicate to me that an interesting (and important) ablation study would be to compare using just the gradients vs using momentum for the sparse training algorithm for both a small batch and a large batch. Without this ablation study, it is a bit unclear to me why/when this algorithm is working well, and this primarily explains my current score.\n\nThe experimental results look impressive, although I am not very aware of other work in sparse training, so unfortunately it is harder for me to properly evaluate the significance of the empirical results in this paper, and whether the numbers reported are indeed a significant improvement over current state-of-the-art algorithms. I do have a few questions about the design and the results of the experiments presented:\n\n1. Because the momentum of zero-valued weights are used, does that mean that the gradient over all weights are required to be taken at each step? How much does this affect training time in your experiments?\n\n2. How were the learning rates decided, and why are they kept fixed across methods? It seems feasible that the optimal learning rate could vary highly between methods?\n\n==========================================\n\nEdit after rebuttal:\nI thank the authors for the very detailed response and for doing the additional experiments requested. Due to the thoroughness of the response, I am increasing my score to a weak accept. I do however have a couple of comments regarding the author response:\n\n1. I do not agree that keeping the learning rate fixed across methods is the right approach. Many different things might interact with each other to change the optimal learning rate, and it is always more interesting to see the optimal performance attained by a method through a grid search. As long as the tuning budget across all methods are somewhat similar, I think the experiments would be more informative when doing a learning rate sweep. Further, while doing the sweep, the authors should be careful that the optimal learning rate does not lie on the edge of the grid search. This is what happens in the experiments reported in the authors' response titled \"Results from learning rate grid search\" where the optimal learning rate for all models lie at one of the extremes of the grid search (0.09).\n\n2. I would request the authors to slightly rewrite certain parts of their paper so as not to imply that momentum decreases the variance of the gradients in general. There are many papers (including the ones I mentioned in my review) that specifically show momentum does not have this effect. All of these papers however consider a different setting from the one considered in this paper (where there is a parameter redistribution step), so it is not immediately clear whether momentum has the same effect in this case. However, unless there is very specific evidence to show a variance reduction effect, I think putting in that intuition might be misleading. \nIt is interesting to see that momentum helps for small batches in this setting, and I agree that further investigation into this would be an interesting future direction. Note that there is a concurrent submission on sparse training (https://openreview.net/forum?id=ryg7vA4tPB&noteId=ryg7vA4tPB) that seems to show that higher momentum values (0.99) does better when considering large batches. So perhaps additional and more careful experiments need to be done about this before these additional experimental results can be put in the paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "In this paper, the authors propose a sparse momentum algorithm for doing efficient sparse training. The technique relies on identifying weights in a layer that do not have an effect on the error, pruning them, and redistributing and growing them across layers. The technique is compared against other recent algorithms on a range of models.\n\nThe paper is well written, and very easy to read. The proposed algorithm looks interesting and seems to empirically work well. \n\nI am, however, a bit confused with how the sparse momentum algorithm is discussed in this paper. In the paper, momentum is intuitively presented as an algorithm that reduces the variance of the noise in the gradients. However, there are a number of papers that show that this is not the case (and if anything it is the opposite). Further, a few recent papers show that momentum works better than SGD only for learning rates that are not too small, and this is because it averages out the gradients in the high-curvature directions (these are the directions where the gradients switch signs) and makes them stable in these directions, thus allowing larger steps in the low curvature directions. See for example the following two papers:\nMomentum Enables Large Batch Training. Samuel L Smith, Erich Elsen, Soham De. ICML Workshop on Physics for Deep Learning, 2019.\nWhich Algorithmic Choices Matter at Which Batch Sizes? Insights From a Noisy Quadratic Model. Guodong Zhang, Lala Li, Zachary Nado, James Martens, Sushant Sachdeva, George E Dahl, Christopher J Shallue, Roger Grosse. NeurIPS 2019.\n\nGiven these papers, can the authors comment on what they think is the reason for the effectiveness of their sparse momentum algorithm? These papers seem to indicate to me that an interesting (and important) ablation study would be to compare using just the gradients vs using momentum for the sparse training algorithm for both a small batch and a large batch. Without this ablation study, it is a bit unclear to me why/when this algorithm is working well, and this primarily explains my current score.\n\nThe experimental results look impressive, although I am not very aware of other work in sparse training, so unfortunately it is harder for me to properly evaluate the significance of the empirical results in this paper, and whether the numbers reported are indeed a significant improvement over current state-of-the-art algorithms. I do have a few questions about the design and the results of the experiments presented:\n\n1. Because the momentum of zero-valued weights are used, does that mean that the gradient over all weights are required to be taken at each step? How much does this affect training time in your experiments?\n\n2. How were the learning rates decided, and why are they kept fixed across methods? It seems feasible that the optimal learning rate could vary highly between methods?\n\n==========================================\n\nEdit after rebuttal:\nI thank the authors for the very detailed response and for doing the additional experiments requested. Due to the thoroughness of the response, I am increasing my score to a weak accept. I do however have a couple of comments regarding the author response:\n\n1. I do not agree that keeping the learning rate fixed across methods is the right approach. Many different things might interact with each other to change the optimal learning rate, and it is always more interesting to see the optimal performance attained by a method through a grid search. As long as the tuning budget across all methods are somewhat similar, I think the experiments would be more informative when doing a learning rate sweep. Further, while doing the sweep, the authors should be careful that the optimal learning rate does not lie on the edge of the grid search. This is what happens in the experiments reported in the authors' response titled \"Results from learning rate grid search\" where the optimal learning rate for all models lie at one of the extremes of the grid search (0.09).\n\n2. I would request the authors to slightly rewrite certain parts of their paper so as not to imply that momentum decreases the variance of the gradients in general. There are many papers (including the ones I mentioned in my review) that specifically show momentum does not have this effect. All of these papers however consider a different setting from the one considered in this paper (where there is a parameter redistribution step), so it is not immediately clear whether momentum has the same effect in this case. However, unless there is very specific evidence to show a variance reduction effect, I think putting in that intuition might be misleading. \nIt is interesting to see that momentum helps for small batches in this setting, and I agree that further investigation into this would be an interesting future direction. Note that there is a concurrent submission on sparse training (https://openreview.net/forum?id=ryg7vA4tPB&noteId=ryg7vA4tPB) that seems to show that higher momentum values (0.99) does better when considering large batches. So perhaps additional and more careful experiments need to be done about this before these additional experimental results can be put in the paper.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572183926875}, {"id": "BJxWWfORYS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper669/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an algorithm called Sparse Momentum for learning sparse neural networks. They claim to maintain sparse weights throughout training while achieving dense network performance levels. They also show their method improves training speed up to 5.61x faster training. The provides a decent motivation for why sparse networks can be helpful. The related work section is well summarized and they emphasize that the current work's primary motivation is to reduce training time while maintaining performance. They compare their method with other methods that also maintain sparse neural networks throughout training and involve single training phase, which is fair. Their method consists of primarily 3 phases - i) pruning weights ii) redistribution of weights iii) regrowing of weights based on the exponentially smoothed momentum term for each layer. The method is well explained and motivation is clear. Edge case was also explained for more clarity. However, there is something that needs more clarification in Pg-3, 3rd line, they say the 3 components of the algorithm i) ii) and iii) can be tackled independently with a divide and conquer strategy to gain some computational benefits. In my understanding, the method performs these 3 steps sequentially after each epoch. Not sure how divide and conquer strategy can be used here ?\n\nThe results were shown for MNIST & CIFAR-10 using AlexNet, VGG16 and LeNet-5 models. They show that the current algorithm is better than other proposed methods in the literature in most of the cases. The claims made in some cases that this method reaches dense network performance is not completely true though. For example, in Table 1, the only case where the proposed method reaches dense network's performance is for VGG16-D. In all the rest of the cases, the current method and dense network error differ by at least 5%. \n\nThe authors compare speed up results for training in two ways: theoretical speedups which are proportional to reduction in number of FLOPS and practical speedups using dense convolutional algorithms corresponding to completely empty channels. It's good that the authors have mentioned due to the current lack of optimal sparse matrix multiplication implementations these speedups cannot be garnered practically yet. The estimates based on FLOPs reduction or Empty Channel based look promising. They also show ablation study of how the redistribution and weight regrowth based on momentum is better than doing in a random fashion.\n\nOverall, I think the paper proposes an interesting idea of using momentum with promising results to learn sparse neural networks. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes an algorithm called Sparse Momentum for learning sparse neural networks. They claim to maintain sparse weights throughout training while achieving dense network performance levels. They also show their method improves training speed up to 5.61x faster training. The provides a decent motivation for why sparse networks can be helpful. The related work section is well summarized and they emphasize that the current work's primary motivation is to reduce training time while maintaining performance. They compare their method with other methods that also maintain sparse neural networks throughout training and involve single training phase, which is fair. Their method consists of primarily 3 phases - i) pruning weights ii) redistribution of weights iii) regrowing of weights based on the exponentially smoothed momentum term for each layer. The method is well explained and motivation is clear. Edge case was also explained for more clarity. However, there is something that needs more clarification in Pg-3, 3rd line, they say the 3 components of the algorithm i) ii) and iii) can be tackled independently with a divide and conquer strategy to gain some computational benefits. In my understanding, the method performs these 3 steps sequentially after each epoch. Not sure how divide and conquer strategy can be used here ?\n\nThe results were shown for MNIST & CIFAR-10 using AlexNet, VGG16 and LeNet-5 models. They show that the current algorithm is better than other proposed methods in the literature in most of the cases. The claims made in some cases that this method reaches dense network performance is not completely true though. For example, in Table 1, the only case where the proposed method reaches dense network's performance is for VGG16-D. In all the rest of the cases, the current method and dense network error differ by at least 5%. \n\nThe authors compare speed up results for training in two ways: theoretical speedups which are proportional to reduction in number of FLOPS and practical speedups using dense convolutional algorithms corresponding to completely empty channels. It's good that the authors have mentioned due to the current lack of optimal sparse matrix multiplication implementations these speedups cannot be garnered practically yet. The estimates based on FLOPs reduction or Empty Channel based look promising. They also show ablation study of how the redistribution and weight regrowth based on momentum is better than doing in a random fashion.\n\nOverall, I think the paper proposes an interesting idea of using momentum with promising results to learn sparse neural networks. "}, "tcdate": 1571877368561}, {"id": "SJgPTmQ5FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper669/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nThe paper proposes a method to train a sparse network and achieve \"dense-level\" performance. The method redistributes the sparsity according to momentum contribution of each layer during training after each epoch. Experiments on multiple datasets and architectures are conducted. \n\nWhile the method itself seems interesting, I do have have several important concerns.\n\n1. My biggest concern is about the experiments. Some results sometime seem questionable.\n\n1) Looking at the results for ImageNet ResNet50, the 20%-weights model achieve 0.7% test error loss compared with dense baseline. According to Table 6 in [1], using the simple pruning and fine-tuning method (Han et al. 2015), the 40%-weights model suffers from 0.06% error loss compared with dense baseline. I don't see a clear advantage of sparse momentum here. It seems possible that under the same sparsity level just pruning using Han et al. is better. Can the authors compare with Han et al. 2015 fairly at the same sparsity level, on multiple datasets? An apple-to-apple comparison is extremely necessary. \n\nAlso the baseline result differ: in [1] dense ResNet-50 achieve 76.15% but in this paper baseline result is 74.9%. Both papers use Pytorch. And Pytorch official number is actually 76.15% (see https://pytorch.org/docs/stable/torchvision/models.html ). In my experience that is reproduceble if you follow the official pytorch code of training ImageNet. What's the difference here? I think the authors should follow the standard way of training ImageNet to make the results more convincing.  Finally [1] is a relevant paper and should be discussed.\n\n2) I suggest that MNIST results be moved to Appendix since accuracy on MNIST is too easy to reach a high level, and the interpretation over it might not be convincing. Yet a lot of the analysis of the method's effectiveness is on MNIST.\n\n3) For CIFAR, the results of SNIP on VGG in Table 1 also seems inconsistent with their original paper. In their paper's Table 2, the VGG-like model achieve ~0.3% error reduction at 5% weights while in this paper's Table 1 it's a 0.5% error increase. Again, what is the difference? Is the SNIP method reimplemented? Putting that potentially flawed result aside, the performance on CIFAR is not better than other methods as well according to Figure 3.\n\n2. I don't think the general performance of the trained sparse models can be said to \"rival\" the dense model. It's a non-negligible margin in most times, for CIFAR and ImageNet. It's a little exaggerating to say that in the title, abstract and introduction.\n\n3. The benefit of the method seems unclear. It does not speedup training compared with the conventional training a dense model and then pruning pipeline (Han et al.). The test-time real speedup is also limited according to Table 3. The real benefit is the compressed model but that is also achievable by traditional pruning (Han et al.). \n\n4. The ablation study should compare with smarter baselines than random regrowth/no redistribution. Clearly each layer needs different level of sparsity so no redistribution is of course not a competitive one. But there might exist other criterion (e.g, weight magnitudes) than momentum which gives good results. The ablation study did not demonstrate why using momentum is justified.\n\n5. \"For dense convolution algorithms, we estimate speedups as follows: If a convolutional channel consists entirely of zero-valued weights we can remove these channels from the computation without changing the outputs and obtain speedups.\" How is it possible that all weights associated with a channel are all pruned, in a sparse pruning setting? A channel typically has at least hundreds or even thousands of weights connected with it, so even with 95% sparsity it's extremely unlikely (consider 0.95^1000). The estimate of speedup might be flawed.\n\nIn summary, those serious issues with experiments make me vote a rejection for the paper.\n\n[1] Rethinking the Value of Network Pruning, ICLR 2019.\n\n------------------------------------------------------------------------------------\nAfter author rebuttal:\n\nThank you for the response.\n\nThe authors view that their method is not comparable with Han et al. 15. because this work started training a sparse model while Han et al. 15 trains a dense model and then do pruning. But I'm not sure whether the rather limited actual (1.2-1.3x, and lower on Wide ResNet) speedup justify this argument. What's more, these speedup were said to be \"estimated\" so I'm not sure whether it's actual speedup. \n\nThe pruning ratio with dense performance is not remarkable: 50% sparsity ratio should be easily achievable with Han et al.\n\nExperiments on ResNet (not wide ResNet) were not provided and I assume ResNet would be harder to sparsify than AlexNet, VGG, Wide ResNet which may have more redundancy. I think thus the \"faster\" training part in the title is not entirely justified. \n\nI'm ok with other parts of the rebuttal and I will raise my score to weak reject. But I still wonder why we should use sparse training in the first place if it brings so little \"estimated\" speedup.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "\nThe paper proposes a method to train a sparse network and achieve \"dense-level\" performance. The method redistributes the sparsity according to momentum contribution of each layer during training after each epoch. Experiments on multiple datasets and architectures are conducted. \n\nWhile the method itself seems interesting, I do have have several important concerns.\n\n1. My biggest concern is about the experiments. Some results sometime seem questionable.\n\n1) Looking at the results for ImageNet ResNet50, the 20%-weights model achieve 0.7% test error loss compared with dense baseline. According to Table 6 in [1], using the simple pruning and fine-tuning method (Han et al. 2015), the 40%-weights model suffers from 0.06% error loss compared with dense baseline. I don't see a clear advantage of sparse momentum here. It seems possible that under the same sparsity level just pruning using Han et al. is better. Can the authors compare with Han et al. 2015 fairly at the same sparsity level, on multiple datasets? An apple-to-apple comparison is extremely necessary. \n\nAlso the baseline result differ: in [1] dense ResNet-50 achieve 76.15% but in this paper baseline result is 74.9%. Both papers use Pytorch. And Pytorch official number is actually 76.15% (see https://pytorch.org/docs/stable/torchvision/models.html ). In my experience that is reproduceble if you follow the official pytorch code of training ImageNet. What's the difference here? I think the authors should follow the standard way of training ImageNet to make the results more convincing.  Finally [1] is a relevant paper and should be discussed.\n\n2) I suggest that MNIST results be moved to Appendix since accuracy on MNIST is too easy to reach a high level, and the interpretation over it might not be convincing. Yet a lot of the analysis of the method's effectiveness is on MNIST.\n\n3) For CIFAR, the results of SNIP on VGG in Table 1 also seems inconsistent with their original paper. In their paper's Table 2, the VGG-like model achieve ~0.3% error reduction at 5% weights while in this paper's Table 1 it's a 0.5% error increase. Again, what is the difference? Is the SNIP method reimplemented? Putting that potentially flawed result aside, the performance on CIFAR is not better than other methods as well according to Figure 3.\n\n2. I don't think the general performance of the trained sparse models can be said to \"rival\" the dense model. It's a non-negligible margin in most times, for CIFAR and ImageNet. It's a little exaggerating to say that in the title, abstract and introduction.\n\n3. The benefit of the method seems unclear. It does not speedup training compared with the conventional training a dense model and then pruning pipeline (Han et al.). The test-time real speedup is also limited according to Table 3. The real benefit is the compressed model but that is also achievable by traditional pruning (Han et al.). \n\n4. The ablation study should compare with smarter baselines than random regrowth/no redistribution. Clearly each layer needs different level of sparsity so no redistribution is of course not a competitive one. But there might exist other criterion (e.g, weight magnitudes) than momentum which gives good results. The ablation study did not demonstrate why using momentum is justified.\n\n5. \"For dense convolution algorithms, we estimate speedups as follows: If a convolutional channel consists entirely of zero-valued weights we can remove these channels from the computation without changing the outputs and obtain speedups.\" How is it possible that all weights associated with a channel are all pruned, in a sparse pruning setting? A channel typically has at least hundreds or even thousands of weights connected with it, so even with 95% sparsity it's extremely unlikely (consider 0.95^1000). The estimate of speedup might be flawed.\n\nIn summary, those serious issues with experiments make me vote a rejection for the paper.\n\n[1] Rethinking the Value of Network Pruning, ICLR 2019.\n\n------------------------------------------------------------------------------------\nAfter author rebuttal:\n\nThank you for the response.\n\nThe authors view that their method is not comparable with Han et al. 15. because this work started training a sparse model while Han et al. 15 trains a dense model and then do pruning. But I'm not sure whether the rather limited actual (1.2-1.3x, and lower on Wide ResNet) speedup justify this argument. What's more, these speedup were said to be \"estimated\" so I'm not sure whether it's actual speedup. \n\nThe pruning ratio with dense performance is not remarkable: 50% sparsity ratio should be easily achievable with Han et al.\n\nExperiments on ResNet (not wide ResNet) were not provided and I assume ResNet would be harder to sparsify than AlexNet, VGG, Wide ResNet which may have more redundancy. I think thus the \"faster\" training part in the title is not entirely justified. \n\nI'm ok with other parts of the rebuttal and I will raise my score to weak reject. But I still wonder why we should use sparse training in the first place if it brings so little \"estimated\" speedup.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571595199038}], "openreview_url": "https://openreview.net/forum?id=ByeSYa4KPS", "arxiv_id": "1907.04840", "paper_pdf": "papers/ByeSYa4KPS.pdf", "paper_pdf_sha256": "3f495c6b76cf154ce24b0b04d2bdfb2e78ddea539b5890e5291a4d433c752e4a", "paper_pdf_bytes": 437250, "paper_pdf_source": "openreview", "code_url": "https://github.com/TimDettmers/sparse_learning", "code_repository": "TimDettmers/sparse_learning", "code_commit": "e19ff16c475462d50ad247ec2927113617662a4e", "code_archive": "repos/ByeSYa4KPS.zip", "code_archive_sha256": "a66ab02038a6b4c6c228e245fbc1391b8c2236287f5e41ec9e68862afaf46d8c", "code_archive_bytes": 2272589, "code_file_count": 33, "code_extensions": {".py": 22, ".sh": 11}, "github_disk_usage_kb": 2438, "github_languages": {"Python": 233428, "Shell": 3473}, "github_archived": false, "github_pushed_at": "2022-06-06T18:16:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sparse-networks-from-scratch-faster-training"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "s6q6zX45F8", "year": 2025, "status": "rejected", "title": "Uncertainty-Based Extensible Codebook for Discrete Federated Learning in Heterogeneous Data Silos", "authors": ["Tianyi Zhang", "Yu Cao", "Dianbo Liu"], "authorids": ["~Tianyi_Zhang18", "~Yu_Cao10", "~Dianbo_Liu2"], "authors_source": "OpenReview API", "abstract": "Federated learning (FL), aimed at leveraging vast distributed datasets, confronts a crucial challenge: the heterogeneity of data across different silos. While previous studies have explored discrete representations to enhance model generalization across minor distributional shifts, these approaches often struggle to adapt to new data silos with significantly divergent distributions. In response, we have identified that models derived from FL exhibit markedly increased uncertainty when applied to data silos with unfamiliar distributions. Consequently, we propose an innovative yet straightforward iterative framework, termed Uncertainty-Based Extensible-Codebook Federated Learning (UEFL). This framework dynamically maps latent features to trainable discrete vectors, assesses the uncertainty, and specifically extends the discretization dictionary or codebook for silos exhibiting high uncertainty. Our approach aims to simultaneously enhance accuracy and reduce uncertainty by explicitly addressing the diversity of data distributions, all while maintaining minimal computational overhead in environments characterized by heterogeneous data silos. Through experiments conducted on six datasets, our method has demonstrated its superiority, achieving significant improvements in accuracy (by 3%--22.1%) and uncertainty reduction (by 38.83%--96.24%), thereby outperforming contemporary state-of-the-art methods.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "4GhCpPUFQZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8822/Reviewer_4szV"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper addresses the challenge of data heterogeneity in Federated Learning (FL) by proposing a novel method called Uncertainty-Based Extensible-Codebook Federated Learning (UEFL). The method introduces a dynamic codebook that maps latent features to discrete vectors, allowing for the extension of the codebook based on the uncertainty of data distributions across different data silos (clients). By iteratively evaluating the uncertainty of the global model's predictions, UEFL adds new codewords to the codebook to improve performance on previously unseen or highly divergent data distributions. The proposed method is tested on six datasets, showing significant improvements in both accuracy (3% to 22.1%) and uncertainty reduction (38.83% to 96.24%) over state-of-the-art approaches.", "review_text": "The paper addresses the challenge of data heterogeneity in Federated Learning (FL) by proposing a novel method called Uncertainty-Based Extensible-Codebook Federated Learning (UEFL). The method introduces a dynamic codebook that maps latent features to discrete vectors, allowing for the extension of the codebook based on the uncertainty of data distributions across different data silos (clients). By iteratively evaluating the uncertainty of the global model's predictions, UEFL adds new codewords to the codebook to improve performance on previously unseen or highly divergent data distributions. The proposed method is tested on six datasets, showing significant improvements in both accuracy (3% to 22.1%) and uncertainty reduction (38.83% to 96.24%) over state-of-the-art approaches.", "strengths": "- The idea of using a dynamic, uncertainty-driven codebook to handle data heterogeneity in federated learning is innovative. By extending the codebook based on uncertainty, the method adapts to diverse and unseen data distributions more efficiently than traditional methods.\n\n- The approach of starting with a small, shared codebook and gradually expanding it based on data-specific uncertainties is efficient. This design helps in minimizing computational overhead while ensuring that the model can adapt to new distributions without retraining the entire model.", "weaknesses": "- In Table 1, the reported accuracy of FedAvg on CIFAR-100 is only 11%, which is surprisingly low. Typically, even with non-IID data, FedAvg achieves higher accuracy. This raises concerns about potential baseline suppression to exaggerate the performance of the proposed method. The authors need to clarify the experimental setup for FedAvg and ensure that the baseline methods are fairly represented. Without this clarification, the validity of the results is questionable.\n\n- The experiments seem to rely on pretrained models. It would be essential to clarify whether the proposed UEFL method still performs well when trained from scratch. Relying on pretrained models can bias results, especially if the pretrained weights are already well-optimized for certain datasets. The authors should include experiments that train the models from scratch to validate the robustness and generalizability of UEFL. Additionally, the 11% accuracy for FedAvg on CIFAR-100 with a pretrained model is puzzling and further raises questions about the validity of the reported results.\n\n- The paper primarily uses Monte Carlo Dropout for uncertainty estimation, which is a standard but limited approach. More advanced uncertainty quantification techniques (e.g., Deep Ensembles or Bayesian Neural Networks) could be explored or at least discussed as alternatives. The performance of UEFL might vary with different uncertainty estimation methods.\n\n- The paper lacks a comprehensive analysis of the computational complexity of the proposed method. Although the authors claim that the codebook remains compact, it would be beneficial to quantify the computational and communication overhead in both the training and inference phases. A comparison of the complexity of UEFL with other FL methods in terms of memory usage and time complexity is missing.\n\n- The paper mainly evaluates the method on six datasets, but does not provide a clear discussion of its scalability to larger numbers of clients. Given that FL is often applied to large-scale systems, it would be helpful to discuss how UEFL scales with increasing numbers of clients or more heterogeneous data distributions.\n\n- Although the paper discusses the adaptive nature of the codebook, it would be useful to include an ablation study that explores the impact of different initial codebook sizes and how the codebook grows over iterations.", "questions": "- The reported 11% accuracy for FedAvg on CIFAR-100 is unreasonably low, especially considering that the experiments used pretrained models. Could the authors clarify whether there were any specific preprocessing steps or hyperparameter choices that led to such low performance? Additionally, does UEFL still outperform FedAvg when both models are trained from scratch, without the benefit of pretraining?\n\n- How does the growth of the codebook over multiple iterations affect the overall communication overhead in FL? Does the increased size of the codebook lead to significant additional communication costs during model aggregation, and how is this handled?\n\n- How is the threshold for uncertainty (which triggers the expansion of the codebook) determined? Is this threshold set manually, or is it dynamically adjusted during training? Additionally, how sensitive is the method to this threshold?\n\n- How does UEFL perform when the data distributions across silos are extremely divergent (for example, different modalities of data)? Does the codebook continue to expand indefinitely, or is there a mechanism to prevent overgrowth?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of data heterogeneity in Federated Learning (FL) by proposing a novel method called Uncertainty-Based Extensible-Codebook Federated Learning (UEFL). The method introduces a dynamic codebook that maps latent features to discrete vectors, allowing for the extension of the codebook based on the uncertainty of data distributions across different data silos (clients). By iteratively evaluating the uncertainty of the global model's predictions, UEFL adds new codewords to the codebook to improve performance on previously unseen or highly divergent data distributions. The proposed method is tested on six datasets, showing significant improvements in both accuracy (3% to 22.1%) and uncertainty reduction (38.83% to 96.24%) over state-of-the-art approaches.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The idea of using a dynamic, uncertainty-driven codebook to handle data heterogeneity in federated learning is innovative. By extending the codebook based on uncertainty, the method adapts to diverse and unseen data distributions more efficiently than traditional methods.\n\n- The approach of starting with a small, shared codebook and gradually expanding it based on data-specific uncertainties is efficient. This design helps in minimizing computational overhead while ensuring that the model can adapt to new distributions without retraining the entire model.", "weaknesses": "- In Table 1, the reported accuracy of FedAvg on CIFAR-100 is only 11%, which is surprisingly low. Typically, even with non-IID data, FedAvg achieves higher accuracy. This raises concerns about potential baseline suppression to exaggerate the performance of the proposed method. The authors need to clarify the experimental setup for FedAvg and ensure that the baseline methods are fairly represented. Without this clarification, the validity of the results is questionable.\n\n- The experiments seem to rely on pretrained models. It would be essential to clarify whether the proposed UEFL method still performs well when trained from scratch. Relying on pretrained models can bias results, especially if the pretrained weights are already well-optimized for certain datasets. The authors should include experiments that train the models from scratch to validate the robustness and generalizability of UEFL. Additionally, the 11% accuracy for FedAvg on CIFAR-100 with a pretrained model is puzzling and further raises questions about the validity of the reported results.\n\n- The paper primarily uses Monte Carlo Dropout for uncertainty estimation, which is a standard but limited approach. More advanced uncertainty quantification techniques (e.g., Deep Ensembles or Bayesian Neural Networks) could be explored or at least discussed as alternatives. The performance of UEFL might vary with different uncertainty estimation methods.\n\n- The paper lacks a comprehensive analysis of the computational complexity of the proposed method. Although the authors claim that the codebook remains compact, it would be beneficial to quantify the computational and communication overhead in both the training and inference phases. A comparison of the complexity of UEFL with other FL methods in terms of memory usage and time complexity is missing.\n\n- The paper mainly evaluates the method on six datasets, but does not provide a clear discussion of its scalability to larger numbers of clients. Given that FL is often applied to large-scale systems, it would be helpful to discuss how UEFL scales with increasing numbers of clients or more heterogeneous data distributions.\n\n- Although the paper discusses the adaptive nature of the codebook, it would be useful to include an ablation study that explores the impact of different initial codebook sizes and how the codebook grows over iterations.", "questions": "- The reported 11% accuracy for FedAvg on CIFAR-100 is unreasonably low, especially considering that the experiments used pretrained models. Could the authors clarify whether there were any specific preprocessing steps or hyperparameter choices that led to such low performance? Additionally, does UEFL still outperform FedAvg when both models are trained from scratch, without the benefit of pretraining?\n\n- How does the growth of the codebook over multiple iterations affect the overall communication overhead in FL? Does the increased size of the codebook lead to significant additional communication costs during model aggregation, and how is this handled?\n\n- How is the threshold for uncertainty (which triggers the expansion of the codebook) determined? Is this threshold set manually, or is it dynamically adjusted during training? Additionally, how sensitive is the method to this threshold?\n\n- How does UEFL perform when the data distributions across silos are extremely divergent (for example, different modalities of data)? Does the codebook continue to expand indefinitely, or is there a mechanism to prevent overgrowth?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730711144128}, {"id": "e4D7ThYmdw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8822/Reviewer_3K1q"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents a method for federated learning, where the goal is to address data heterogeneity across clients and enhance generalization ability of federated learning model across different data distributions. Specifically, the method leverages a codebook of representations to improve accuracy and reduce prediction uncertainty in federated settings. The authors validate the proposed framework in multi-domain learning, and domain generalization in leave-one-domain-out experiments.\n\nThe proposed framework is as follows: for each client model, first encode the input data to latent features, extract code words via k-means clustering of these latent features, and then discretize the latent features using the nearest code word; predictions are subsequently made based on the coded (discretized) latent vectors. The framework also introduces an uncertainty-based adjustment mechanism, where clients with uncertainty above a certain threshold use a larger codebook with additional code words. The server model is updated as the average over all clients.", "review_text": "This paper presents a method for federated learning, where the goal is to address data heterogeneity across clients and enhance generalization ability of federated learning model across different data distributions. Specifically, the method leverages a codebook of representations to improve accuracy and reduce prediction uncertainty in federated settings. The authors validate the proposed framework in multi-domain learning, and domain generalization in leave-one-domain-out experiments.\n\nThe proposed framework is as follows: for each client model, first encode the input data to latent features, extract code words via k-means clustering of these latent features, and then discretize the latent features using the nearest code word; predictions are subsequently made based on the coded (discretized) latent vectors. The framework also introduces an uncertainty-based adjustment mechanism, where clients with uncertainty above a certain threshold use a larger codebook with additional code words. The server model is updated as the average over all clients.", "strengths": "The authors address an important challenge in federated learning to tackle data heterogeneity and enhance model generalization across different distributions.\n\nThe presentation of the paper is easy to follow.", "weaknesses": "I have the following questions regarding the work:\n\n1. My first question is about the rationale behind the framework. Could the authors elaborate on how and why discretizing latent representations through a codebook helps achieve the goals of improved generalization and uncertainty reduction?\n\n2. Regarding the server model, could the authors clarify how the codebook of the server model is updated?\n\n3. Is there any convergence guarantee of Algorithm 1? Specifically, does the framework ensure that all clients will reach an uncertainty level below the threshold? Could the authors clarify the mechanism described in lines 322–323?\n\n4. Could the authors clarify why reducing uncertainty in prediction is a valid objective? If the model achieves high certainty on incorrect predictions, could this exacerbate challenges in improving model performance? Additionally, if reducing uncertainty is a primary goal, beyond the proposed method of expanding client codebook, have the authors considered training client models to the point of neural collapse [1]?\n\n5. Distribution shift is also addressed in other studies, such as [2,3]. It would strengthen the evaluation if the authors could expand their comparisons to include more related methods.\n\n\n\n[1] Papyan, Vardan, X. Y. Han, and David L. Donoho. \"Prevalence of neural collapse during the terminal phase of deep learning training.\" Proceedings of the National Academy of Sciences 117.40 (2020): 24652-24663.\n\n[2] Nguyen, A. Tuan, Philip Torr, and Ser Nam Lim. \"Fedsr: A simple and effective domain generalization method for federated learning.\" Advances in Neural Information Processing Systems 35 (2022): 38831-38843.\n\n[3] Zhang, Hao, et al. \"FedCR: Personalized federated learning based on across-client common representation with conditional mutual information regularization.\" International Conference on Machine Learning. PMLR, 2023.", "questions": "It would be helpful if the authors could address my above questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a method for federated learning, where the goal is to address data heterogeneity across clients and enhance generalization ability of federated learning model across different data distributions. Specifically, the method leverages a codebook of representations to improve accuracy and reduce prediction uncertainty in federated settings. The authors validate the proposed framework in multi-domain learning, and domain generalization in leave-one-domain-out experiments.\n\nThe proposed framework is as follows: for each client model, first encode the input data to latent features, extract code words via k-means clustering of these latent features, and then discretize the latent features using the nearest code word; predictions are subsequently made based on the coded (discretized) latent vectors. The framework also introduces an uncertainty-based adjustment mechanism, where clients with uncertainty above a certain threshold use a larger codebook with additional code words. The server model is updated as the average over all clients.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The authors address an important challenge in federated learning to tackle data heterogeneity and enhance model generalization across different distributions.\n\nThe presentation of the paper is easy to follow.", "weaknesses": "I have the following questions regarding the work:\n\n1. My first question is about the rationale behind the framework. Could the authors elaborate on how and why discretizing latent representations through a codebook helps achieve the goals of improved generalization and uncertainty reduction?\n\n2. Regarding the server model, could the authors clarify how the codebook of the server model is updated?\n\n3. Is there any convergence guarantee of Algorithm 1? Specifically, does the framework ensure that all clients will reach an uncertainty level below the threshold? Could the authors clarify the mechanism described in lines 322–323?\n\n4. Could the authors clarify why reducing uncertainty in prediction is a valid objective? If the model achieves high certainty on incorrect predictions, could this exacerbate challenges in improving model performance? Additionally, if reducing uncertainty is a primary goal, beyond the proposed method of expanding client codebook, have the authors considered training client models to the point of neural collapse [1]?\n\n5. Distribution shift is also addressed in other studies, such as [2,3]. It would strengthen the evaluation if the authors could expand their comparisons to include more related methods.\n\n\n\n[1] Papyan, Vardan, X. Y. Han, and David L. Donoho. \"Prevalence of neural collapse during the terminal phase of deep learning training.\" Proceedings of the National Academy of Sciences 117.40 (2020): 24652-24663.\n\n[2] Nguyen, A. Tuan, Philip Torr, and Ser Nam Lim. \"Fedsr: A simple and effective domain generalization method for federated learning.\" Advances in Neural Information Processing Systems 35 (2022): 38831-38843.\n\n[3] Zhang, Hao, et al. \"FedCR: Personalized federated learning based on across-client common representation with conditional mutual information regularization.\" International Conference on Machine Learning. PMLR, 2023.", "questions": "It would be helpful if the authors could address my above questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730596957962}, {"id": "Qmoi5x8l9U", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8822/Reviewer_7FFp"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes UEFL, which utilizes an extensible codebook to address feature shifts in FL. Numerical results demonstrate the superior performance of the proposed method in both multi-domain learning and out-of-distribution generation tasks.", "review_text": "This paper proposes UEFL, which utilizes an extensible codebook to address feature shifts in FL. Numerical results demonstrate the superior performance of the proposed method in both multi-domain learning and out-of-distribution generation tasks.", "strengths": "1. Utilizing the codebook technique to address the feature shift in FL is innovative.\n\n2. The authors conducted comprehensive experiments on multiple datasets, demonstrating the performance improvement of the proposed method in both traditional FL tasks and out-of-distribution generalization tasks.", "weaknesses": "1. The algorithms presented in Section 3 do not align well with Algorithm 1. For instance, lines 226-228 are not thoroughly explained. It would be beneficial if the authors could provide additional details on the main pages.\n\n2. The number of training epochs for the baseline models is set to be the same as that for UEFL. However, I am uncertain about the fairness of this approach, as (1) the optimal number of training epochs for UEFL may not be optimal for the baseline models; (2) the communication and computation costs associated with UEFL and the baseline models differ, which makes the comparison somewhat unfair.", "questions": "It would be advantageous for the authors to include convergence curves relating to communication, computation, and the number of rounds. This would provide a clearer illustration of the effectiveness of UEFL.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes UEFL, which utilizes an extensible codebook to address feature shifts in FL. Numerical results demonstrate the superior performance of the proposed method in both multi-domain learning and out-of-distribution generation tasks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Utilizing the codebook technique to address the feature shift in FL is innovative.\n\n2. The authors conducted comprehensive experiments on multiple datasets, demonstrating the performance improvement of the proposed method in both traditional FL tasks and out-of-distribution generalization tasks.", "weaknesses": "1. The algorithms presented in Section 3 do not align well with Algorithm 1. For instance, lines 226-228 are not thoroughly explained. It would be beneficial if the authors could provide additional details on the main pages.\n\n2. The number of training epochs for the baseline models is set to be the same as that for UEFL. However, I am uncertain about the fairness of this approach, as (1) the optimal number of training epochs for UEFL may not be optimal for the baseline models; (2) the communication and computation costs associated with UEFL and the baseline models differ, which makes the comparison somewhat unfair.", "questions": "It would be advantageous for the authors to include convergence curves relating to communication, computation, and the number of rounds. This would provide a clearer illustration of the effectiveness of UEFL.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730535517218}], "openreview_url": "https://openreview.net/forum?id=s6q6zX45F8", "arxiv_id": "2402.18888", "paper_pdf": "papers/s6q6zX45F8.pdf", "paper_pdf_sha256": "1ee10d8baabed5432f656c6f784893207945598475d35d82d2d1adb90023f38d", "paper_pdf_bytes": 1849144, "paper_pdf_source": "openreview", "code_url": "https://github.com/destiny301/uefl", "code_repository": "destiny301/uefl", "code_commit": "833571a8d0739120229aacf49d7303f557176aea", "code_archive": "repos/s6q6zX45F8.zip", "code_archive_sha256": "5281df8a2ad092d8d3efc957a86471e7c73f0fb3ff17f9947dc2b431d460850b", "code_archive_bytes": 170884, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 222, "github_languages": {"Python": 56378}, "github_archived": false, "github_pushed_at": "2024-03-28T04:55:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/uncertainty-based-extensible-codebook-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZkEsEFFUyo", "year": 2024, "status": "rejected", "title": "Pushing the Limits of Pre-training for Time Series Forecasting in the CloudOps Domain", "authors": ["Gerald Woo", "Chenghao Liu", "Akshat Kumar", "Doyen Sahoo"], "authorids": ["~Gerald_Woo1", "~Chenghao_Liu1", "~Akshat_Kumar2", "~Doyen_Sahoo1"], "authors_source": "OpenReview API", "abstract": "Time series has been left behind in the era of pre-training and transfer learning. While research in the fields of natural language processing and computer vision are enjoying progressively larger datasets to train massive models, the most popular time series datasets consist of only tens of thousands of time steps, limiting our ability to study the effectiveness of pre-training and scaling. Recent studies have also cast doubt on the need for expressive models and scale. To alleviate these issues, we  introduce three large-scale time series forecasting datasets from the cloud operations (CloudOps) domain, the largest having billions of observations, enabling further study into pre-training and scaling of time series models. We build the empirical groundwork for studying pre-training and scaling of time series models and pave the way for future research by identifying a promising candidate architecture. We show that it is a strong zero-shot baseline and benefits from further scaling, both in model and dataset size. Accompanying these datasets and results is a suite of comprehensive benchmark results comparing classical and deep learning baselines to our pre-trained method -- achieving a 27% reduction in error on the largest dataset. Code and datasets will be made publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "zGHGsFejFx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2434/Reviewer_LRUs"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This scientific paper addresses the limited progress in applying pre-training and transfer learning to time series data. To bridge this gap, the authors introduce three large-scale time series forecasting datasets from the CloudOps domain, with the largest dataset containing billions of observations. This substantial dataset size allows for a comprehensive investigation into the effectiveness of pre-training and scaling for time series models. The paper establishes the groundwork for studying pre-training and scaling of time series models and identifies a promising architecture for the task. This architecture serves as a strong zero-shot baseline and exhibits further improvements with increased model and dataset size. The authors provide a benchmark comparing several classical and deep learning methods with their proposed pre-trained approach, demonstrating significant reduction in error on the largest dataset.", "review_text": "This scientific paper addresses the limited progress in applying pre-training and transfer learning to time series data. To bridge this gap, the authors introduce three large-scale time series forecasting datasets from the CloudOps domain, with the largest dataset containing billions of observations. This substantial dataset size allows for a comprehensive investigation into the effectiveness of pre-training and scaling for time series models. The paper establishes the groundwork for studying pre-training and scaling of time series models and identifies a promising architecture for the task. This architecture serves as a strong zero-shot baseline and exhibits further improvements with increased model and dataset size. The authors provide a benchmark comparing several classical and deep learning methods with their proposed pre-trained approach, demonstrating significant reduction in error on the largest dataset.", "strengths": "1. The manuscript is well-written and easy to follow. The authors thoroughly explain various setups, such as pretraining and fine-tuning, and make an attempt to define a taxonomy of time series data based on domains, collections, and individual time series.\n2. The work introduces three large-scale time series forecasting datasets, which enable a deeper exploration of pretraining and transfer learning for time series models. The authors provide concrete details into the process of transforming raw data into useful time series data.\n3. The paper goes on to examine various Transformer architectures for forecasting, conducts a comprehensive benchmarking analysis against classical and deep learning forecasting methods, and also compares their proposed pretraining approaches with existing methods. Additionally, the study investigates the impact of scaling in terms of model parameters and training data size on time series forecasting.", "weaknesses": "1. Lack of Novelty: The paper's approach is primarily centered on training a large transformer model on an extensive time series dataset, which may not offer a novel contribution to the field.\n2. Limited Analysis of Transfer Learning: The study raises concerns regarding the assessment of transfer learning, as the model is trained on one dataset and tested on a similar one without sufficient information about their differences. This makes it challenging to interpret the extent of transfer learning within an in-collection setting. Additionally, the diversity of time series within the CloudOps datasets remains unclear, which could lead to potential overfitting issues, particularly for large models.\n3. Inadequate Baseline Model Details: The paper lacks essential information about the baseline models, including their model size and training duration. For instance, while DeepAR emerges as the second best method in Table 5, it is uncertain if its number of parameters is comparable to the proposed transformer models. A comparison with DeepAR models of similar scale and training iterations would provide valuable insights.\n4. Limited Improvements in Larger Models: The paper reveals that the gains achieved with \"Large\" and \"xLarge\" models are not significant when compared to the base model size. Particularly, the \"xLarge\" model, despite being over eight times larger than the base model, exhibits only marginal improvements. This suggests that the base model may already be overfitting the datasets.", "questions": "See weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This scientific paper addresses the limited progress in applying pre-training and transfer learning to time series data. To bridge this gap, the authors introduce three large-scale time series forecasting datasets from the CloudOps domain, with the largest dataset containing billions of observations. This substantial dataset size allows for a comprehensive investigation into the effectiveness of pre-training and scaling for time series models. The paper establishes the groundwork for studying pre-training and scaling of time series models and identifies a promising architecture for the task. This architecture serves as a strong zero-shot baseline and exhibits further improvements with increased model and dataset size. The authors provide a benchmark comparing several classical and deep learning methods with their proposed pre-trained approach, demonstrating significant reduction in error on the largest dataset.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "1. The manuscript is well-written and easy to follow. The authors thoroughly explain various setups, such as pretraining and fine-tuning, and make an attempt to define a taxonomy of time series data based on domains, collections, and individual time series.\n2. The work introduces three large-scale time series forecasting datasets, which enable a deeper exploration of pretraining and transfer learning for time series models. The authors provide concrete details into the process of transforming raw data into useful time series data.\n3. The paper goes on to examine various Transformer architectures for forecasting, conducts a comprehensive benchmarking analysis against classical and deep learning forecasting methods, and also compares their proposed pretraining approaches with existing methods. Additionally, the study investigates the impact of scaling in terms of model parameters and training data size on time series forecasting.", "weaknesses": "1. Lack of Novelty: The paper's approach is primarily centered on training a large transformer model on an extensive time series dataset, which may not offer a novel contribution to the field.\n2. Limited Analysis of Transfer Learning: The study raises concerns regarding the assessment of transfer learning, as the model is trained on one dataset and tested on a similar one without sufficient information about their differences. This makes it challenging to interpret the extent of transfer learning within an in-collection setting. Additionally, the diversity of time series within the CloudOps datasets remains unclear, which could lead to potential overfitting issues, particularly for large models.\n3. Inadequate Baseline Model Details: The paper lacks essential information about the baseline models, including their model size and training duration. For instance, while DeepAR emerges as the second best method in Table 5, it is uncertain if its number of parameters is comparable to the proposed transformer models. A comparison with DeepAR models of similar scale and training iterations would provide valuable insights.\n4. Limited Improvements in Larger Models: The paper reveals that the gains achieved with \"Large\" and \"xLarge\" models are not significant when compared to the base model size. Particularly, the \"xLarge\" model, despite being over eight times larger than the base model, exhibits only marginal improvements. This suggests that the base model may already be overfitting the datasets.", "questions": "See weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698833395324}, {"id": "8bvamF97AZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2434/Reviewer_QeLY"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents pre-training time-series model in the cloud operations domain to enhance downstream forecasting accuracy. The authors conduct experiments to compare various model architectures and investigate the scaling laws impacting both model and data size. Their findings indicate promising results in zero-shot scenarios.", "review_text": "This paper presents pre-training time-series model in the cloud operations domain to enhance downstream forecasting accuracy. The authors conduct experiments to compare various model architectures and investigate the scaling laws impacting both model and data size. Their findings indicate promising results in zero-shot scenarios.", "strengths": "1. The paper introduces the first pre-trained time-series model specifically for cloud operation domains.\n2. The study evaluates different model architectures in the context of pre-training and examines the effects of model and data size on performance with scaling laws.", "weaknesses": "1. The pre-training and zero-shot testing appear to be conducted within the same dataset. If this is correct, it raises concerns about the true generality of the zero-shot performance, particularly regarding its effectiveness in diverse or new datasets and domains.\n2. The paper mainly focuses on benchmarking existing pre-training model architectures without significant novel adaptations or designs tailored to the specific requirements of the cloud operation domain.\n3. Is the model randomly initialized? Given the effectiveness of the baseline one-fits-all model, an exploration into using pre-trained language models as initialization might be helpful.\n4. It remains unclear whether deep learning benchmarks such as N-BEATS, Autoformer, and FEDformer are only trained on designated training sets. Exploring whether similar performance improvements could be achieved by training these models on the pre-training set, using the same loss function, might offer deeper insights. This also helps determine the source of the proposed model's performance gains—whether from the architecture itself or mostly from expanded datasets.", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents pre-training time-series model in the cloud operations domain to enhance downstream forecasting accuracy. The authors conduct experiments to compare various model architectures and investigate the scaling laws impacting both model and data size. Their findings indicate promising results in zero-shot scenarios.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The paper introduces the first pre-trained time-series model specifically for cloud operation domains.\n2. The study evaluates different model architectures in the context of pre-training and examines the effects of model and data size on performance with scaling laws.", "weaknesses": "1. The pre-training and zero-shot testing appear to be conducted within the same dataset. If this is correct, it raises concerns about the true generality of the zero-shot performance, particularly regarding its effectiveness in diverse or new datasets and domains.\n2. The paper mainly focuses on benchmarking existing pre-training model architectures without significant novel adaptations or designs tailored to the specific requirements of the cloud operation domain.\n3. Is the model randomly initialized? Given the effectiveness of the baseline one-fits-all model, an exploration into using pre-trained language models as initialization might be helpful.\n4. It remains unclear whether deep learning benchmarks such as N-BEATS, Autoformer, and FEDformer are only trained on designated training sets. Exploring whether similar performance improvements could be achieved by training these models on the pre-training set, using the same loss function, might offer deeper insights. This also helps determine the source of the proposed model's performance gains—whether from the architecture itself or mostly from expanded datasets.", "questions": "See Weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698719529440}, {"id": "Hpgm6gbmSs", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2434/Reviewer_boWu"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces three new large datasets from the CloudOps domains with billions of observations for time series forecasting. Further, the authors provide a performance analysis including time series forecasting models for the cases where models are trained from scratch, fine tuned, and zero-shot. The authors show further evaluate different transformer variants and show that the proposed baseline achieves a 27% error reduction on largest dataset.", "review_text": "This paper introduces three new large datasets from the CloudOps domains with billions of observations for time series forecasting. Further, the authors provide a performance analysis including time series forecasting models for the cases where models are trained from scratch, fine tuned, and zero-shot. The authors show further evaluate different transformer variants and show that the proposed baseline achieves a 27% error reduction on largest dataset.", "strengths": "1. The authors provide three new large datasets from the CloudOps field. This contribution is relevant in the field of time series forecasting, as there is certain lack of this kind of datasets, holding back the progress in the topic of pretrained models or LLM-based forecasting models.\n2. The authors provide an interesting evaluation of several existing models. These evaluations provide an interesting reference on how well-established models perform in these datasets, and how they compare with pretrained models.\n3. The authors provide an interesting analysis on variants of transformer models, providing further insights on what approaches are more promising in the future.\n4. The paper is well written and it makes an effort on having well structured terminology in these new and developing field.", "weaknesses": "The main weakness of the paper is that the contribution of this paper is basically three new datasets. I acknowledge the non-trivial effort that it implies to gather this kind of large scale datasets. Nevertheless, I share as well my concern that there is not much of an analysis of what are the main challenges that these datasets pose. For instance: 1/ Do they have missing values? 2/ how strong is the seasonality in these datasets? 3/ Is there any interesting trend in the data? Are there any distribution shifts (for instance, something like a black-friday regime, a change around covid lockdowns, etc)?", "questions": "1. what is the main challenge that these datasets pose?\n2. how diverse are these three datasets? I understand that they come from the CloudOps field, but is there any fundamental difference between them?\n3. Is there any distribution shift captured in these datasets?\n4. How do these datasets look in the test windows used for evaluation? And how do the forecasts per model look? I understand that for ali2018 the forecasts are potentially similar as Naive performs very good, but what about azure 2017 where the proposed baselines are performing exceptionally well?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces three new large datasets from the CloudOps domains with billions of observations for time series forecasting. Further, the authors provide a performance analysis including time series forecasting models for the cases where models are trained from scratch, fine tuned, and zero-shot. The authors show further evaluate different transformer variants and show that the proposed baseline achieves a 27% error reduction on largest dataset.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The authors provide three new large datasets from the CloudOps field. This contribution is relevant in the field of time series forecasting, as there is certain lack of this kind of datasets, holding back the progress in the topic of pretrained models or LLM-based forecasting models.\n2. The authors provide an interesting evaluation of several existing models. These evaluations provide an interesting reference on how well-established models perform in these datasets, and how they compare with pretrained models.\n3. The authors provide an interesting analysis on variants of transformer models, providing further insights on what approaches are more promising in the future.\n4. The paper is well written and it makes an effort on having well structured terminology in these new and developing field.", "weaknesses": "The main weakness of the paper is that the contribution of this paper is basically three new datasets. I acknowledge the non-trivial effort that it implies to gather this kind of large scale datasets. Nevertheless, I share as well my concern that there is not much of an analysis of what are the main challenges that these datasets pose. For instance: 1/ Do they have missing values? 2/ how strong is the seasonality in these datasets? 3/ Is there any interesting trend in the data? Are there any distribution shifts (for instance, something like a black-friday regime, a change around covid lockdowns, etc)?", "questions": "1. what is the main challenge that these datasets pose?\n2. how diverse are these three datasets? I understand that they come from the CloudOps field, but is there any fundamental difference between them?\n3. Is there any distribution shift captured in these datasets?\n4. How do these datasets look in the test windows used for evaluation? And how do the forecasts per model look? I understand that for ali2018 the forecasts are potentially similar as Naive performs very good, but what about azure 2017 where the proposed baselines are performing exceptionally well?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698677252616}], "openreview_url": "https://openreview.net/forum?id=ZkEsEFFUyo", "arxiv_id": "2310.05063", "paper_pdf": "papers/ZkEsEFFUyo.pdf", "paper_pdf_sha256": "2f52a6ecbba40b79974bf316614458e104aa9116c537e11de2aa351f34f127c8", "paper_pdf_bytes": 1466322, "paper_pdf_source": "openreview", "code_url": "https://github.com/SalesforceAIResearch/pretrain-time-series-cloudops", "code_repository": "SalesforceAIResearch/pretrain-time-series-cloudops", "code_commit": "983c1b9515b72721856a2f4b3de9107080c1dc51", "code_archive": "repos/ZkEsEFFUyo.zip", "code_archive_sha256": "fa44def132bcbf4bbfaf9d3afc14ffeac2bae7cb05a0824ecbac7a27b598ce1e", "code_archive_bytes": 227523, "code_file_count": 94, "code_extensions": {".py": 94}, "github_disk_usage_kb": 148, "github_languages": {"Python": 761347}, "github_archived": true, "github_pushed_at": "2025-09-09T17:45:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pushing-the-limits-of-pre-training-for-time"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bSuY3hSRJPP", "year": 2023, "status": "rejected", "title": "SpectraNet: multivariate forecasting and imputation under distribution shifts and missing data", "authors": ["Cristian Ignacio Challu", "Peihong Jiang", "Ying Nian Wu", "Laurent Callot"], "authorids": ["~Cristian_Ignacio_Challu1", "jpeihong@amazon.com", "~Ying_Nian_Wu1", "~Laurent_Callot1"], "authors_source": "OpenReview API", "abstract": "In this work, we tackle two widespread challenges in real applications for time-series forecasting that have been largely understudied: distribution shifts and missing data. We propose SpectraNet, a novel multivariate time-series forecasting model that dynamically infers a latent space spectral decomposition to capture current temporal dynamics and correlations on the recent observed history. A Convolution Neural Network maps the learned representation by sequentially mixing its components and refining the output. Our proposed approach can simultaneously produce forecasts and interpolate past observations and can, therefore, greatly simplify production systems by unifying imputation and forecasting tasks into a single model. SpectraNet achieves SoTA performance simultaneously on both tasks on five benchmark datasets, compared to forecasting and imputation models, with up to 92% fewer parameters and comparable training times. On settings with up to 80% missing data, SpectraNet has average performance improvements of almost 50% over the second-best alternative.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "0RBhCAwhyLa", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2684/Reviewer_Mu1M"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In the paper, the authors proposed a time-series forecasting framework to overcome the distribution shifts and tackle missing values simultaneously. ", "review_text": "In this paper, the authors proposed a novel time series forecasting framework that handles missing data and distribution drift. However, the paper is not clearly presented, and the following issues should be addressed: \n\n1. The presentation and organization of this paper need to be improved. There are many typos and incomplete sentences. For example, on page 2, \"practice, imputation models are first ...\". \n\n2. What's the advantage of inferring the latent variable separately instead of using an encoder? The motivation is not strong here, and the authors didn't provide enough details. Besides, a point-wise distance-based loss function (e.g., MSE) may not be enough to capture the shape correlation between two-time series. Can the authors provide more information and analysis regarding the design of latent vector inference?\n\n3. What's the meaning of s_w / 2 in Figure 1 and equation 7? What is p in equation 7? The authors should clearly describe and explain the equations used in the paper. \n\n4. How does the imputation work in the inference of latent vectors? Can the authors provide more details?\n\n", "strengths": "Strengths\n1. An important problem is studied\nWeaknesses:\n1. Weak motivation for proposed model components\n2. Paper presentation and organization need to be improved", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In the paper, the authors proposed a time-series forecasting framework to overcome the distribution shifts and tackle missing values simultaneously. ", "strength_and_weaknesses": "Strengths\n1. An important problem is studied\nWeaknesses:\n1. Weak motivation for proposed model components\n2. Paper presentation and organization need to be improved", "clarity,_quality,_novelty_and_reproducibility": "This paper is difficult to follow due to the unclear presentation, especially in the model part. Further improvement is needed to improve the clarity. ", "summary_of_the_review": "In this paper, the authors proposed a novel time series forecasting framework that handles missing data and distribution drift. However, the paper is not clearly presented, and the following issues should be addressed: \n\n1. The presentation and organization of this paper need to be improved. There are many typos and incomplete sentences. For example, on page 2, \"practice, imputation models are first ...\". \n\n2. What's the advantage of inferring the latent variable separately instead of using an encoder? The motivation is not strong here, and the authors didn't provide enough details. Besides, a point-wise distance-based loss function (e.g., MSE) may not be enough to capture the shape correlation between two-time series. Can the authors provide more information and analysis regarding the design of latent vector inference?\n\n3. What's the meaning of s_w / 2 in Figure 1 and equation 7? What is p in equation 7? The authors should clearly describe and explain the equations used in the paper. \n\n4. How does the imputation work in the inference of latent vectors? Can the authors provide more details?\n\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666668613171}, {"id": "bZZK01Vv5h", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2684/Reviewer_HyvP"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, authors proposed a multivariate time-series forecasting model (SpectraNet) that unified the forecasting and interpolation problem. Specifically, the model first infers the optimal latent vector on the reference window by minimizing the reconstruction error, then the model generates the full predictions via a latent space spectral decomposition module. The proposed model achieves the best performance compared with other existing methods.\n", "review_text": "I list my concerns in [Weaknesses], I am happy to discuss and increase the rating if my concerns are addressed.", "strengths": "**Strengths**:  \n1. Well-organized and clearly written.  \n2. Promising results.  \n\n\n**Weaknesses**:  \n1.  In Sec. 3.5, the distribution shift is defined as the difference of missing data regime between training set and test set. However, in my opinion, this kind of difference is not the distribution shift, since the training and test sets both come from the **same** dataset in a random split (or not) manner, the **data distribution** should be the same in the training and test sets, even though the total numbers or locations of missing values are different. Therefore, I would say there is no distribution shift problem.\n\n2. Some key details are missing.   \n(a) In Sec. 3.2, how many temporal bases $B$ are used? What's the insights or reasons?   \n(b) In Sec. 3.3, there is no clarification of the ConvTranspose1d, more details should be included. Why do the first two layers learn a common representation and the second layer refine the temporal resolution? There is no such experiment or explanation of this claim. In addition, the *causality* is not clear either, \"the latent vector $z^{*}$ is only inferred with information on the reference window\" is not the reason of  *causality*.  \n(c) In Sec. 4.3, it says \"all models are trained with the training loss only on available values\", how about the missing points? I think the ground truth of missing points is accessible during training, if the model doesn't use the missing points during training, what's the meaning of interpolation?  \n\n\n\n3. Experiments are not sufficient. There is no ablation study to demonstrate the effectiveness of some designs in the model, i.e., temporal basis, the inference step, etc.\n\n\n4. Minor mistakes:  \n(1) In Sec. 3.1, the first sentence is incoherent, \"of the\", \"instead of\".  \n(2) In Fig. 2, the legend blocks the first curve $\\hat{y_{1}}$,  even though it looks the same as $\\hat{y_{10}}$ . Almost the same problem in all the figures.  \n(3) Eq. (8), not aligned.  \n(4) In Fig.3, the characters are too small.  \n(5) In References, the names of conferences or journals, i.e. AAAI or AAAI Press , are not consistent.  \nI highly encourage authors to revise the submission to avoid such small mistakes, as well as some typos.  \n\n\n5. Some missing references[1,2] when addressing distribution shift problem in time-series.  \n[1] Adaptive Trajectory Prediction via Transferable GNN. CVPR2022  \n[2] Expanding the Deployment Envelope of Behavior Prediction via Adaptive Meta-Learning. arXiv 2022\n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "In this paper, authors proposed a multivariate time-series forecasting model (SpectraNet) that unified the forecasting and interpolation problem. Specifically, the model first infers the optimal latent vector on the reference window by minimizing the reconstruction error, then the model generates the full predictions via a latent space spectral decomposition module. The proposed model achieves the best performance compared with other existing methods.\n", "strength_and_weaknesses": "**Strengths**:  \n1. Well-organized and clearly written.  \n2. Promising results.  \n\n\n**Weaknesses**:  \n1.  In Sec. 3.5, the distribution shift is defined as the difference of missing data regime between training set and test set. However, in my opinion, this kind of difference is not the distribution shift, since the training and test sets both come from the **same** dataset in a random split (or not) manner, the **data distribution** should be the same in the training and test sets, even though the total numbers or locations of missing values are different. Therefore, I would say there is no distribution shift problem.\n\n2. Some key details are missing.   \n(a) In Sec. 3.2, how many temporal bases $B$ are used? What's the insights or reasons?   \n(b) In Sec. 3.3, there is no clarification of the ConvTranspose1d, more details should be included. Why do the first two layers learn a common representation and the second layer refine the temporal resolution? There is no such experiment or explanation of this claim. In addition, the *causality* is not clear either, \"the latent vector $z^{*}$ is only inferred with information on the reference window\" is not the reason of  *causality*.  \n(c) In Sec. 4.3, it says \"all models are trained with the training loss only on available values\", how about the missing points? I think the ground truth of missing points is accessible during training, if the model doesn't use the missing points during training, what's the meaning of interpolation?  \n\n\n\n3. Experiments are not sufficient. There is no ablation study to demonstrate the effectiveness of some designs in the model, i.e., temporal basis, the inference step, etc.\n\n\n4. Minor mistakes:  \n(1) In Sec. 3.1, the first sentence is incoherent, \"of the\", \"instead of\".  \n(2) In Fig. 2, the legend blocks the first curve $\\hat{y_{1}}$,  even though it looks the same as $\\hat{y_{10}}$ . Almost the same problem in all the figures.  \n(3) Eq. (8), not aligned.  \n(4) In Fig.3, the characters are too small.  \n(5) In References, the names of conferences or journals, i.e. AAAI or AAAI Press , are not consistent.  \nI highly encourage authors to revise the submission to avoid such small mistakes, as well as some typos.  \n\n\n5. Some missing references[1,2] when addressing distribution shift problem in time-series.  \n[1] Adaptive Trajectory Prediction via Transferable GNN. CVPR2022  \n[2] Expanding the Deployment Envelope of Behavior Prediction via Adaptive Meta-Learning. arXiv 2022\n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Details are in [Weaknesses]\n\n**Clarity**:  \nThe clarity is good but with some minor mistakes.\n\n**Quality**:  \nClear-written and well-organized.\n\n**Novelty**:  \nMotivation is good but the contributions are not significant.\n\n**Reproducibility**:  \nThe training procedure is clear in Appendix 5,  and authors promise to release the code after paper acceptance. \n", "summary_of_the_review": "I list my concerns in [Weaknesses], I am happy to discuss and increase the rating if my concerns are addressed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666548919107}, {"id": "p_TCKn_s39", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2684/Reviewer_r8BM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "To tackle the distribution shift and missing data problem of time series, this paper proposes a latent space spectral decomposition method for simultaneous time series forecasting and imputation. The latent vector is optimized individually for unseen data and thus can generalize well to unseen data distribution. ", "review_text": "Despite the drawbacks that I mentioned, I think this is an interesting paper overall.", "strengths": "Pros:\n1) The idea to optimize a latent vector for different data is interesting and this is the first time this idea is applied on time series. Previously this idea has been used to images [2]. \n2) Experiments result are better than several baselines\n\nCons: \n1) The idea to combine several basis functions to reconstruct the input is not new. Previously, [1,3] used this idea to reconstruct music signals. Also, I am not sure that only using sin, cos, and poly functions are enough for all types of time series datasets. Some datasets may have strange shapes that can not be reconstructed well.\n2) For optimization, I am not sure about using the inference step and the learning step can lead to a globally optimal solution. It would be great if the authors can provide some explanation here. \n\n[1] DDSP: Differentiable Digital Signal Processing\n[2] Optimizing the Latent Space of Generative Networks \n[3] Differentiable Wavetable Synthesis", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "To tackle the distribution shift and missing data problem of time series, this paper proposes a latent space spectral decomposition method for simultaneous time series forecasting and imputation. The latent vector is optimized individually for unseen data and thus can generalize well to unseen data distribution. ", "strength_and_weaknesses": "Pros:\n1) The idea to optimize a latent vector for different data is interesting and this is the first time this idea is applied on time series. Previously this idea has been used to images [2]. \n2) Experiments result are better than several baselines\n\nCons: \n1) The idea to combine several basis functions to reconstruct the input is not new. Previously, [1,3] used this idea to reconstruct music signals. Also, I am not sure that only using sin, cos, and poly functions are enough for all types of time series datasets. Some datasets may have strange shapes that can not be reconstructed well.\n2) For optimization, I am not sure about using the inference step and the learning step can lead to a globally optimal solution. It would be great if the authors can provide some explanation here. \n\n[1] DDSP: Differentiable Digital Signal Processing\n[2] Optimizing the Latent Space of Generative Networks \n[3] Differentiable Wavetable Synthesis", "clarity,_quality,_novelty_and_reproducibility": "Clarity: Good\nQuality: Good\nNovelty: Good\nReproducibility:Good", "summary_of_the_review": "Despite the drawbacks that I mentioned, I think this is an interesting paper overall.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666143934292}], "openreview_url": "https://openreview.net/forum?id=bSuY3hSRJPP", "arxiv_id": "2210.12515", "paper_pdf": "papers/bSuY3hSRJPP.pdf", "paper_pdf_sha256": "d2d8025216197478f16ec840ab8c8ca3d45851d7e85c9d8dcfd12233e301b047", "paper_pdf_bytes": 3414267, "paper_pdf_source": "openreview", "code_url": "https://github.com/cchallu/spectranet", "code_repository": "cchallu/spectranet", "code_commit": "f476444d048793b259f0eb65995f66f292784e2b", "code_archive": "repos/bSuY3hSRJPP.zip", "code_archive_sha256": "99c91d9aa03ce0065a93eee6b92777d23cff43fcdf8c1ba2ca43b0faef374831", "code_archive_bytes": 310192, "code_file_count": 4, "code_extensions": {".py": 4}, "github_disk_usage_kb": 312, "github_languages": {"Python": 34864}, "github_archived": false, "github_pushed_at": "2022-10-10T15:17:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/spectranet-multivariate-forecasting-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Vq_QHT5kcAK", "year": 2022, "status": "rejected", "title": "Greedy Bayesian Posterior Approximation with Deep Ensembles", "authors": ["Aleksei Tiulpin", "Matthew B. Blaschko"], "authorids": ["~Aleksei_Tiulpin1", "~Matthew_B._Blaschko1"], "authors_source": "OpenReview API", "abstract": "Ensembles of independently trained neural networks are a state-of-the-art approach to estimate predictive uncertainty in Deep Learning, and can be interpreted as an approximation of the posterior distribution via a mixture of delta functions. The training of ensembles relies on non-convexity of the loss landscape and random initialization of their individual members, making the resulting posterior approximation uncontrolled. This paper proposes a novel and principled method to tackle this limitation, minimizing an $f$-divergence between the true posterior and a kernel density estimator in a function space. We analyze this objective from a combinatorial point of view, and show that it is submodular with respect to mixture components for any $f$. Subsequently, we consider the problem of greedy ensemble construction, and from the marginal gain of the total objective, we derive a novel diversity term for ensemble methods. The performance of our approach is demonstrated on computer vision out-of-distribution detection benchmarks in a range of architectures trained on multiple datasets. The source code of our method is made publicly available.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "OWGpBqiIqp0", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper783/Reviewer_dcPg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a combinatorial approach to controlling and improving the effects of diversity; a relevant and interesting feature in Bayesian networks. The approach is general in the sense that it involves minimization of an f-divergence, for any f. The approach is mainly compared to standard deep ensembles, i.e. without controlling diversity. Additionally, the paper provides a theoretical result with approximation guarantees, which might be useful outside this specific application.", "review_text": "PROS\n\n1. The approach targets diversity, a relevant and interesting phenomena, and gives a technically non-trivial algorithm for controlling it. Especially, I like the connection to the empirical risk in Eq. (1) and the flexibility of the method (it works for any f-divergence). To me it seems the authors expertize in combinatorial optimization.\n\n2. The results seem convincing that the proposed method can outperform standard deep ensembles in OOD tasks. Moreover, Figure 3 is a nice representation of the OOD detection results.\n\n3. The Related Work section and the literature study in the introduction are dense. The authors have clearly spent time probing the field. These sections are nice to read and are useful for navigating in this evolving domain (although I miss one reference, see 2. below).\n\n\nCONS\n\n1. My main concern is regarding accessibility and clarity of how the method is described. As this is posed as a competing method/framework to standard deep ensembles, the paper should be accessible to practitioners and researchers with a broad technical background. As such, the concepts submodularity, supermodularity and modularity, as well as their definitions in the paper, need to be contextualized.\n\nFor instance, I have been working a lot with deep ensembles and Bayesian deep learning, and am interested in the method proposed in the paper, but I have a hard time understanding the motivation of using the modularities? I see that they help devise the method theoretically, but what are f, V, B and A in Definition 1? As an example of contextualization, simply state what f is in this setting. All I understand about f right now is that it is a mapping from a binary input of dimension V to the real line. Also, is f in Def. 1 the same f as in f-divergence? If not, please consider distinguishing them.\n\nIn my opinion, the delivery of the method in, especially, sections 2.2 and 3.1 is not easy to follow. These two sections essentially contain enumerations of definitions and I have a hard time finding motivations for them. As far as I am aware, you do not reference the definitions or the theorem in the text.\n\n2. Although the sections Related Work and Introduction are thorough, I think [1] should be included. In [1] they correlate the ensemble components by letting them share weights. Your objective function also correlates the ensemble components. In fact, I think the results in the paper would benefit a lot from comparing with [1] *and*, e.g., [2]. [2] since, as is pointed out in Sec 5, it is \"closely connected\". Not comparing with POVI methods overall is not well motivated, in my opinion.\n\n3. How did you obtain Figure 1? It is not obvious to me that (a) and (b) would occur. It seems speculative and unmotivated as is (this alone makes my \"Correctness\" score = 3).\n\n4. The paper has an OK structure in the majority of the sections (especially Sec. 1, 5, 6 and 7) and the language is generally good. However, there are many disturbing typos which could easily have been caught by reading through the paper.\n\n[1] \"Training independent subnetworks for robust prediction\", Havasi et al., ICLR 2021\n[2] \"Repulsive deep ensembles are bayesian.\", Francesco D’Angelo and Vincent Fortuin. NeurIPS 2021", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents a combinatorial approach to controlling and improving the effects of diversity; a relevant and interesting feature in Bayesian networks. The approach is general in the sense that it involves minimization of an f-divergence, for any f. The approach is mainly compared to standard deep ensembles, i.e. without controlling diversity. Additionally, the paper provides a theoretical result with approximation guarantees, which might be useful outside this specific application.", "main_review": "PROS\n\n1. The approach targets diversity, a relevant and interesting phenomena, and gives a technically non-trivial algorithm for controlling it. Especially, I like the connection to the empirical risk in Eq. (1) and the flexibility of the method (it works for any f-divergence). To me it seems the authors expertize in combinatorial optimization.\n\n2. The results seem convincing that the proposed method can outperform standard deep ensembles in OOD tasks. Moreover, Figure 3 is a nice representation of the OOD detection results.\n\n3. The Related Work section and the literature study in the introduction are dense. The authors have clearly spent time probing the field. These sections are nice to read and are useful for navigating in this evolving domain (although I miss one reference, see 2. below).\n\n\nCONS\n\n1. My main concern is regarding accessibility and clarity of how the method is described. As this is posed as a competing method/framework to standard deep ensembles, the paper should be accessible to practitioners and researchers with a broad technical background. As such, the concepts submodularity, supermodularity and modularity, as well as their definitions in the paper, need to be contextualized.\n\nFor instance, I have been working a lot with deep ensembles and Bayesian deep learning, and am interested in the method proposed in the paper, but I have a hard time understanding the motivation of using the modularities? I see that they help devise the method theoretically, but what are f, V, B and A in Definition 1? As an example of contextualization, simply state what f is in this setting. All I understand about f right now is that it is a mapping from a binary input of dimension V to the real line. Also, is f in Def. 1 the same f as in f-divergence? If not, please consider distinguishing them.\n\nIn my opinion, the delivery of the method in, especially, sections 2.2 and 3.1 is not easy to follow. These two sections essentially contain enumerations of definitions and I have a hard time finding motivations for them. As far as I am aware, you do not reference the definitions or the theorem in the text.\n\n2. Although the sections Related Work and Introduction are thorough, I think [1] should be included. In [1] they correlate the ensemble components by letting them share weights. Your objective function also correlates the ensemble components. In fact, I think the results in the paper would benefit a lot from comparing with [1] *and*, e.g., [2]. [2] since, as is pointed out in Sec 5, it is \"closely connected\". Not comparing with POVI methods overall is not well motivated, in my opinion.\n\n3. How did you obtain Figure 1? It is not obvious to me that (a) and (b) would occur. It seems speculative and unmotivated as is (this alone makes my \"Correctness\" score = 3).\n\n4. The paper has an OK structure in the majority of the sections (especially Sec. 1, 5, 6 and 7) and the language is generally good. However, there are many disturbing typos which could easily have been caught by reading through the paper.\n\n[1] \"Training independent subnetworks for robust prediction\", Havasi et al., ICLR 2021\n[2] \"Repulsive deep ensembles are bayesian.\", Francesco D’Angelo and Vincent Fortuin. NeurIPS 2021", "summary_of_the_review": "I am certain that the proposed method can indeed be useful and of interest for many in the field, and the method appears to be largely novel. However, I am concerned that it will not be very impactful, the way it is delivered now. To be clear, I don't believe that the mathematical approach is necessarily complicated. However, the provided explanations are over-complicated, and the approach is insufficiently motivated/contextualized.\n\nTo address this concern, I suggest that, for instance, redundant definitions (as Def. 4?) are moved to the Appendix (if at all needed). Instead, there should be some paragraphs in Section 2 that puts concepts like submodularity in context. Can you map it to a concept in Bayesian DL?\n\nRegarding the baselines, I think it is important to compare the method to more recent advances in order to fully grasp its capabilities. As is shown in the Related Work, the field has evolved a lot since deep ensembles were first proposed. Question to the authors: how come more baselines were not considered?", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635877429413}, {"id": "2UL28GUYC_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper783/Reviewer_xECp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This proposes studies ensemble construction from a combinatorial point of view. Analysis are conducted in function space, where the intuition are from and algorithm are developed. Specifically, it starts by formally defining the sub-modularity then proving f-divergence for a distribution and a mixture of kernels to be super-modular function. A greedy maximization of f-divergence is given and its implementation for approximating Bayesian posterior are presented. The experiments on synthetic data show the ability of covering the posterior of the proposed method, compared with the deep ensemble approach. It also conducts experiments on real-world data for OOD detection tasks. ", "review_text": "In general, the paper provides a novel perspective for constructing the ensemble, connecting f-divergence between a distribution and a kernel mixture with sub-modular functions. Submodularity is useful in exploration, e.g., submodular set function maximization. The general idea is novel and intuitive, which could be helpful for the research in deep ensemble.\n\nThe execution of the implementation is solid in general. Some findings and analysis are interesting, e.g., a better approximation factor compared with theoretical guaranteed factor could be achieved for f-divergence. I have the following questions or comments before making the final decision:\n1. Could you provide a high level explanation of why sub-modularity might be helpful for constructing the ensemble before introducing the definitions, theorems and algorithms? Jumping to this functional view after brief review of previous works might be too abrupt.\n2. How are the mode-seeking and mean-seeking ability of the proposed method justified theoretically or empirically?\n3. There are a few approximations like using the point estimate for the reverse KL. How does it affect the general performance and how is this quantified?\n4. It is claimed the heuristic techniques like batchensemble or hyper parameter ensemble could be used adjunct with the proposed method. However, those methods are scalable to large datasets like ImageNet and could be computed efficiently. I wonder if the proposed method could be extended to ImageNet level datasets.\n5. How is the performance of robustness under adversarial attack?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This proposes studies ensemble construction from a combinatorial point of view. Analysis are conducted in function space, where the intuition are from and algorithm are developed. Specifically, it starts by formally defining the sub-modularity then proving f-divergence for a distribution and a mixture of kernels to be super-modular function. A greedy maximization of f-divergence is given and its implementation for approximating Bayesian posterior are presented. The experiments on synthetic data show the ability of covering the posterior of the proposed method, compared with the deep ensemble approach. It also conducts experiments on real-world data for OOD detection tasks. ", "main_review": "In general, the paper provides a novel perspective for constructing the ensemble, connecting f-divergence between a distribution and a kernel mixture with sub-modular functions. Submodularity is useful in exploration, e.g., submodular set function maximization. The general idea is novel and intuitive, which could be helpful for the research in deep ensemble.\n\nThe execution of the implementation is solid in general. Some findings and analysis are interesting, e.g., a better approximation factor compared with theoretical guaranteed factor could be achieved for f-divergence. I have the following questions or comments before making the final decision:\n1. Could you provide a high level explanation of why sub-modularity might be helpful for constructing the ensemble before introducing the definitions, theorems and algorithms? Jumping to this functional view after brief review of previous works might be too abrupt.\n2. How are the mode-seeking and mean-seeking ability of the proposed method justified theoretically or empirically?\n3. There are a few approximations like using the point estimate for the reverse KL. How does it affect the general performance and how is this quantified?\n4. It is claimed the heuristic techniques like batchensemble or hyper parameter ensemble could be used adjunct with the proposed method. However, those methods are scalable to large datasets like ImageNet and could be computed efficiently. I wonder if the proposed method could be extended to ImageNet level datasets.\n5. How is the performance of robustness under adversarial attack?", "summary_of_the_review": "This paper is with interesting idea and solid execution. I recommend a weak accept before the questions are answered.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635755001405}, {"id": "117qNiEsKUs", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper783/Reviewer_ZKEY"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a novel and principled method to make Deep Ensemble Bayesian. It minimizes an f-divergence between the true posterior and a kernel density estimator fitting the functions associated with the ensemble candidates. The authors formulate the learning as a submodular problem and propose to use a greedy approach to solve it. The resultant learning objective consists of a novel diversity term. Some out-of-distribution detection results in computer vision demonstrate the efficacy of the proposed method.", "review_text": "# Strengths\n- The problem of making Deep Ensemble Bayesian is important for the community, and the proposed method seems to be a viable approach to solve this problem.\n\n- To my knowledge, the idea of rephrasing the Bayesian inference problem as a submodular one is conceptually novel.\n\n- The theoretical justification is impressive and interesting.\n\n# Weaknesses & Concerns\n\n- It seems that, eventually, you use the reverse KL divergence as an instantiation of the f-divergence, and develop the practical training loss Eq 12 and Eq 14. \nThough your modeling is applicable to arbitrary f-divergence, you have not demonstrated the implications of such universality.\n\n- Eq 14 is closely related to the function-space POVI (Wang et al., 2019). The difference is that you train these ensemble candidates sequentially with weight decay while function-space POVI trains them in parallel without weight decay? I would like to see an empirical/theoretical clarification of the connections/difference between these two approaches.\n\n- Although I know the main contribution of this work is on a new theoretical framework for Bayesian Deep Ensemble, I still want to see that this works can beat Deep Ensemble significantly in more scenarios. The current performance gains in Table 1 are marginal.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a novel and principled method to make Deep Ensemble Bayesian. It minimizes an f-divergence between the true posterior and a kernel density estimator fitting the functions associated with the ensemble candidates. The authors formulate the learning as a submodular problem and propose to use a greedy approach to solve it. The resultant learning objective consists of a novel diversity term. Some out-of-distribution detection results in computer vision demonstrate the efficacy of the proposed method.", "main_review": "# Strengths\n- The problem of making Deep Ensemble Bayesian is important for the community, and the proposed method seems to be a viable approach to solve this problem.\n\n- To my knowledge, the idea of rephrasing the Bayesian inference problem as a submodular one is conceptually novel.\n\n- The theoretical justification is impressive and interesting.\n\n# Weaknesses & Concerns\n\n- It seems that, eventually, you use the reverse KL divergence as an instantiation of the f-divergence, and develop the practical training loss Eq 12 and Eq 14. \nThough your modeling is applicable to arbitrary f-divergence, you have not demonstrated the implications of such universality.\n\n- Eq 14 is closely related to the function-space POVI (Wang et al., 2019). The difference is that you train these ensemble candidates sequentially with weight decay while function-space POVI trains them in parallel without weight decay? I would like to see an empirical/theoretical clarification of the connections/difference between these two approaches.\n\n- Although I know the main contribution of this work is on a new theoretical framework for Bayesian Deep Ensemble, I still want to see that this works can beat Deep Ensemble significantly in more scenarios. The current performance gains in Table 1 are marginal.", "summary_of_the_review": "This paper has done a great job in making Deep Ensemble Bayesian theoretically, despite being short of empirical verification. So, currently,  I recommend a weak acceptance for it.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635667754988}], "openreview_url": "https://openreview.net/forum?id=Vq_QHT5kcAK", "arxiv_id": "2105.14275", "paper_pdf": "papers/Vq_QHT5kcAK.pdf", "paper_pdf_sha256": "481ae9e437852cba1ba4c1385072a2ae9ca8be37418be32c13fd34f9e1a8264f", "paper_pdf_bytes": 1008569, "paper_pdf_source": "openreview", "code_url": "https://github.com/imedslab/greedy_ensembles_training", "code_repository": "imedslab/greedy_ensembles_training", "code_commit": "387dd8678c6cf3d4b12463eb9b4096007c80b316", "code_archive": "repos/Vq_QHT5kcAK.zip", "code_archive_sha256": "b2c7b10ec69ea04203d802bbf01a97b1d42b37df23a98b6232d9b7ad7ac2ec25", "code_archive_bytes": 328488, "code_file_count": 17, "code_extensions": {".py": 16, ".sh": 1}, "github_disk_usage_kb": 805, "github_languages": {"Python": 61865, "Shell": 1478}, "github_archived": false, "github_pushed_at": "2022-07-18T12:52:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/greedy-bayesian-posterior-approximation-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Y-Wl1l0Va-", "year": 2021, "status": "rejected", "title": "Shortest-Path Constrained Reinforcement Learning for Sparse Reward Tasks", "authors": ["Sungryull Sohn", "Sungtae Lee", "Jongwook Choi", "Harm van Seijen", "Honglak Lee", "Mehdi Fatemi"], "authorids": ["~Sungryull_Sohn1", "~Sungtae_Lee1", "~Jongwook_Choi1", "~Harm_van_Seijen1", "~Honglak_Lee2", "~Mehdi_Fatemi1"], "authors_source": "OpenReview API", "abstract": "We propose the k-Shortest-Path (k-SP) constraint: a novel constraint on the agent’s trajectory that improves the sample-efficiency in sparse-reward MDPs. We show that any optimal policy necessarily satisfies the k-SP constraint. Notably, the k-SP constraint prevents the policy from exploring state-action pairs along the non-k-SP trajectories (e.g., going back and forth).  However, in practice, excluding state-action pairs may hinder convergence of many RL algorithms. To overcome this, we propose a novel cost function that penalizes the policy violating SP constraint, instead of completely excluding it.  Our numerical experiment in a tabular RL setting demonstrate that the SP constraint can significantly reduce the trajectory space of policy. As a result, our constraint enables more sample efficient learning by suppressing redundant exploration and exploitation. Our empirical experiment results on MiniGrid and DeepMind Lab show that the proposed method significantly improves proximal policy optimization (PPO) and outperforms existing novelty-seeking exploration methods including count-based exploration, indicating that it improves the sample efficiency by preventing the agent from taking redundant actions.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gnMTIIHcnV", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2125/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**\nThis paper proposes a new constraint for constrained MDP, based on k-shortest path, which helps improve sample efficiency for (model-free) RL algorithms in sparse-reward MDP, while theoretically proving that the constraint retains the same optimal policy in the original MDP. Intuitively, for sparse (positive) reward setting, the optimal policy should reach the positive reward states with the shortest path (as it has the lowest discounting). The relaxed form of the constraint considers the that the distance between two states is less than $k$, rather than being optimal length (which the optimal policy still also satisfies). The constraint is then converted into its Lagrangian form as a cost term to the reward (i.e. a type of reward shaping). Practically, this requires a k-reachability network (RNet), which is a binary distance discriminator judging whether the distance between two states are reachable within $k$ steps. This network is trained with contrastive loss, similar to prior work SPTM by Savinov et al., 2018. However, SPTM uses RNet for graph-based planning (i.e. the local distance between states), while this paper uses RNet as the cost/constraint on the policy objective function. Experiments were conducted in several maze navigation environments (2D grid world MiniGrid, to first person 3D maze environments in DeepMind Lab), showing promising results compared to several baselines which use intrinsic curiosity. Several ablations were performed on the hyperparameters ($k$, and tolerance $\\delta t$ on the constraint), as well as some qualitative examples of the policies learned compared to novelty reward shaping.\n\n\n**Strengths**:\n- There has been a lot of work with goal-conditioned RL with sparse reward. The advantage of this paper is that this approach converts the notion of reaching goals into a constraint that is applicable to general MDPs. \n- The theoretical results (under some mild assumptions such as positive cumulative reward, mild stochasticity) appear to be correct, although I am not absolutely certain. \n- Empirically the experiments were designed well to study the effects of its hyperparameters ($k$, and tolerance $\\delta t$ on the constraint), and performs strongly as the task is harder (sparser reward / long trajectories) compared to its baselines\n\n**Weaknesses**:\n- It is still unclear to me about the potential performance gap between using an imperfect RNet versus an oracle ground truth distance discriminator. See my question below for some more detail\n- In the experimental sections, some sections are unclear about whether the RNet is used or not. For example, RNet is not used in 6.2 (MiniGrid), while is used in 6.3 (DeepMindLab) but required reading appendix to confirm. I am not sure for 6.4 and 6.5 whether RNet is used at all (I think it is not). Please clarify for me and in the paper. \n\n**Recommendation**:\n I recommend this paper a marginally above acceptance threshold. Overall I think that it is a well-written paper with thorough theoretical and empirical results. There are some clarification parts that can improve the paper even further.\n\n**Questions**:\n1. My main question is what is the performance difference between having the ground truth distance versus using the RNet? For example, can you compare the performance of 6.2 if RNet was used here? I understand RNet is not used in 6.2 for the purpose of understanding the objective. I would the oracle RNet would give the upper bound on the performance on SPRL, but how much worse is using RNet in those environments?\n2. Perhaps outside of the scope of this paper, but I am curious about how SPRL would perform if there was a curriculum on the $k$ and $\\delta t$ value, rather than a fixed value that is treated as a hyper parameter. Perhaps the authors have some intuition or have actually tried this as well. \n3. Please clarify about the use of RNet vs. Ground truth distance in the section 6.4/6.5. \n\n**After rebuttal responses**:\n\nI have read the authors’ response to my concerns, as well as the other reviews. I maintain my current evaluation with a weak acceptance of the paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A well written paper introducing shortest path constraint for sparse reward MDPs", "review": "**Summary**\nThis paper proposes a new constraint for constrained MDP, based on k-shortest path, which helps improve sample efficiency for (model-free) RL algorithms in sparse-reward MDP, while theoretically proving that the constraint retains the same optimal policy in the original MDP. Intuitively, for sparse (positive) reward setting, the optimal policy should reach the positive reward states with the shortest path (as it has the lowest discounting). The relaxed form of the constraint considers the that the distance between two states is less than $k$, rather than being optimal length (which the optimal policy still also satisfies). The constraint is then converted into its Lagrangian form as a cost term to the reward (i.e. a type of reward shaping). Practically, this requires a k-reachability network (RNet), which is a binary distance discriminator judging whether the distance between two states are reachable within $k$ steps. This network is trained with contrastive loss, similar to prior work SPTM by Savinov et al., 2018. However, SPTM uses RNet for graph-based planning (i.e. the local distance between states), while this paper uses RNet as the cost/constraint on the policy objective function. Experiments were conducted in several maze navigation environments (2D grid world MiniGrid, to first person 3D maze environments in DeepMind Lab), showing promising results compared to several baselines which use intrinsic curiosity. Several ablations were performed on the hyperparameters ($k$, and tolerance $\\delta t$ on the constraint), as well as some qualitative examples of the policies learned compared to novelty reward shaping.\n\n\n**Strengths**:\n- There has been a lot of work with goal-conditioned RL with sparse reward. The advantage of this paper is that this approach converts the notion of reaching goals into a constraint that is applicable to general MDPs. \n- The theoretical results (under some mild assumptions such as positive cumulative reward, mild stochasticity) appear to be correct, although I am not absolutely certain. \n- Empirically the experiments were designed well to study the effects of its hyperparameters ($k$, and tolerance $\\delta t$ on the constraint), and performs strongly as the task is harder (sparser reward / long trajectories) compared to its baselines\n\n**Weaknesses**:\n- It is still unclear to me about the potential performance gap between using an imperfect RNet versus an oracle ground truth distance discriminator. See my question below for some more detail\n- In the experimental sections, some sections are unclear about whether the RNet is used or not. For example, RNet is not used in 6.2 (MiniGrid), while is used in 6.3 (DeepMindLab) but required reading appendix to confirm. I am not sure for 6.4 and 6.5 whether RNet is used at all (I think it is not). Please clarify for me and in the paper. \n\n**Recommendation**:\n I recommend this paper a marginally above acceptance threshold. Overall I think that it is a well-written paper with thorough theoretical and empirical results. There are some clarification parts that can improve the paper even further.\n\n**Questions**:\n1. My main question is what is the performance difference between having the ground truth distance versus using the RNet? For example, can you compare the performance of 6.2 if RNet was used here? I understand RNet is not used in 6.2 for the purpose of understanding the objective. I would the oracle RNet would give the upper bound on the performance on SPRL, but how much worse is using RNet in those environments?\n2. Perhaps outside of the scope of this paper, but I am curious about how SPRL would perform if there was a curriculum on the $k$ and $\\delta t$ value, rather than a fixed value that is treated as a hyper parameter. Perhaps the authors have some intuition or have actually tried this as well. \n3. Please clarify about the use of RNet vs. Ground truth distance in the section 6.4/6.5. \n\n**After rebuttal responses**:\n\nI have read the authors’ response to my concerns, as well as the other reviews. I maintain my current evaluation with a weak acceptance of the paper. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604011884240}, {"id": "CJ2mOpPT6iv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2125/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a k-shortest-path constrained reinforcement learning method for solving sparse reward MDPs. Under the assumption that the MDP is a single-goal task and by utilizing a novel cost function, the proposed method is demonstrated to outperform a few baselines in terms of sample efficiency.\n\nPro:\nThe technique along with its theoretical property are well developed.\n\nCareful numerical studies are provided to evaluate the effect of the k-shortest constraint in a tabular-RL settings  \n\nCons:\nIt is unclear how this method and its theoretical results can be generalized to other settings where the rewards are less sparse than single-goal MDP settings, but still very sparse.\n\nThe experiments for testing the proposed method are all based on navigation tasks, which give us an impression that the proposed method is only applicable to a specific problem. It might be better to consider a different type of domains.\n\nThe compared baseline methods are very limited and not necessarily ideal candidates for solving sparse reward problems. There are other approaches, such as hierarchical reinforcement learning which can also solve sparse reward problems , e.g., eigen-options [1] might also worth to compare. \n[1] Machado et al. A Laplacian Framework for Option Discovery in Reinforcement Learning, imcl2017.\n\nOther questions:\nIs there a convergence guarantee for algorithm 1? \n\nIt might be better to provide some computational complexity analysis of algorithm 1.\n\nWhat is the motivation of using k-reachability network to implement the binary distance discriminator?\n\n--------------after the rebuttal---------------\nI appreciated the authors' effort in addressing my comments and questions. I maintain my score of weak acceptance for this paper. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting approach but might need to consider more diversified domains to justify its general applicability.", "review": "This paper proposed a k-shortest-path constrained reinforcement learning method for solving sparse reward MDPs. Under the assumption that the MDP is a single-goal task and by utilizing a novel cost function, the proposed method is demonstrated to outperform a few baselines in terms of sample efficiency.\n\nPro:\nThe technique along with its theoretical property are well developed.\n\nCareful numerical studies are provided to evaluate the effect of the k-shortest constraint in a tabular-RL settings  \n\nCons:\nIt is unclear how this method and its theoretical results can be generalized to other settings where the rewards are less sparse than single-goal MDP settings, but still very sparse.\n\nThe experiments for testing the proposed method are all based on navigation tasks, which give us an impression that the proposed method is only applicable to a specific problem. It might be better to consider a different type of domains.\n\nThe compared baseline methods are very limited and not necessarily ideal candidates for solving sparse reward problems. There are other approaches, such as hierarchical reinforcement learning which can also solve sparse reward problems , e.g., eigen-options [1] might also worth to compare. \n[1] Machado et al. A Laplacian Framework for Option Discovery in Reinforcement Learning, imcl2017.\n\nOther questions:\nIs there a convergence guarantee for algorithm 1? \n\nIt might be better to provide some computational complexity analysis of algorithm 1.\n\nWhat is the motivation of using k-reachability network to implement the binary distance discriminator?\n\n--------------after the rebuttal---------------\nI appreciated the authors' effort in addressing my comments and questions. I maintain my score of weak acceptance for this paper. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603969100248}, {"id": "AHtUA8ck3oL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2125/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes the k-Shortest-Path (k-SP) constraint to restrict the agent’s trajectory to avoid redundant exploration and thus improves sample efficiency in sparse-reward MDPs. Specifically, k-SP constraint is applied to a trajectory rolled out by a policy where all of its sub-path of length k is required to be a shortest-path under the π-distance metric. Instead of a hard constraint, a cost function-based formulation is proposed to implement the constraint. The method can improve the sample efficiency in sparse reward tasks and also preserve the optimality of given MDP. Numerical results in the paper also demonstrate the effectiveness of k-SP compared with existing methods on two domains (1) Mini-Grid and (2) DeepMind Lab in sparse reward settings.\n\nOverall, the paper is well written and clearly conveys the main idea and the main results of the work. The idea and motivation of the paper are very intuitive and very reasonable. The theoretical results are immediately following the ideas. The algorithm proposed has a clear structure and is easy to understand and implement. For experiments, the new algorithm consistently outperforms existing studies on a set of MDPs where there exists a goal state (as both the unique reward state and the terminal state). Some important discussions are highlighted to introduce the algorithm,. Moreover, the proposed mechanism seems to be an inspiration for future work considering the state space exploration related to sample efficiency.\n\nHowever, I wasn't fully convinced by the paper about relevance. Reinforcement learning aims to learn in an environment by trial and error without prior knowledge about the environment. As the problems considered by the paper are with episodic rewards (in both theory and experiments), *the problem themselves are shortest path problems*. Using the shortest path constraint to solve shortest path problems seems not fair to be placed among a set of learning algorithms. Armed with this prior knowledge, the algorithm outperforms marginally (though consistently) compared with pure learning-based algorithms, only with its best choice of k. I believe it would fall short if placed among search algorithms. Some strong justifications are needed for the work to be relevant.\n\nPros: \n1.\tThe paper considers a practical problem in reinforcement learning: sample efficiency in sparse reward tasks. The RL algorithm tends to fail if the collected trajectory does not contain enough evaluative feedback. The idea of using a constrained-RL framework and cost function to tackle the problem is natural and has been well motivated given some drawbacks in existing work mentioned in section 5.\n2.\tThe relaxation from the shortest path to k-SP is well explained. The novel cost function introduced to penalizes the policy-violating SP constraint can tackle some limitations of existing methods. For example, it can preserve the convergence and optimality of the policy. \n3.  This paper provides convincing numerical experiments to show the effectiveness of the proposed framework. The ablation studies are also helpful to show the effects of hyperparameters. \n  \nCons: \n1. The choice of k is not clear.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting work", "review": "This paper proposes the k-Shortest-Path (k-SP) constraint to restrict the agent’s trajectory to avoid redundant exploration and thus improves sample efficiency in sparse-reward MDPs. Specifically, k-SP constraint is applied to a trajectory rolled out by a policy where all of its sub-path of length k is required to be a shortest-path under the π-distance metric. Instead of a hard constraint, a cost function-based formulation is proposed to implement the constraint. The method can improve the sample efficiency in sparse reward tasks and also preserve the optimality of given MDP. Numerical results in the paper also demonstrate the effectiveness of k-SP compared with existing methods on two domains (1) Mini-Grid and (2) DeepMind Lab in sparse reward settings.\n\nOverall, the paper is well written and clearly conveys the main idea and the main results of the work. The idea and motivation of the paper are very intuitive and very reasonable. The theoretical results are immediately following the ideas. The algorithm proposed has a clear structure and is easy to understand and implement. For experiments, the new algorithm consistently outperforms existing studies on a set of MDPs where there exists a goal state (as both the unique reward state and the terminal state). Some important discussions are highlighted to introduce the algorithm,. Moreover, the proposed mechanism seems to be an inspiration for future work considering the state space exploration related to sample efficiency.\n\nHowever, I wasn't fully convinced by the paper about relevance. Reinforcement learning aims to learn in an environment by trial and error without prior knowledge about the environment. As the problems considered by the paper are with episodic rewards (in both theory and experiments), *the problem themselves are shortest path problems*. Using the shortest path constraint to solve shortest path problems seems not fair to be placed among a set of learning algorithms. Armed with this prior knowledge, the algorithm outperforms marginally (though consistently) compared with pure learning-based algorithms, only with its best choice of k. I believe it would fall short if placed among search algorithms. Some strong justifications are needed for the work to be relevant.\n\nPros: \n1.\tThe paper considers a practical problem in reinforcement learning: sample efficiency in sparse reward tasks. The RL algorithm tends to fail if the collected trajectory does not contain enough evaluative feedback. The idea of using a constrained-RL framework and cost function to tackle the problem is natural and has been well motivated given some drawbacks in existing work mentioned in section 5.\n2.\tThe relaxation from the shortest path to k-SP is well explained. The novel cost function introduced to penalizes the policy-violating SP constraint can tackle some limitations of existing methods. For example, it can preserve the convergence and optimality of the policy. \n3.  This paper provides convincing numerical experiments to show the effectiveness of the proposed framework. The ablation studies are also helpful to show the effects of hyperparameters. \n  \nCons: \n1. The choice of k is not clear.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603952054410}, {"id": "Ha8YhYCWjb", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2125/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a novel k-shortest-path constraint that prevents over-exploration by exploit the combinatorial structure (i.e. shortest path) of sparse reward tasks. It is proved theoretically that the k-shortest-path constraint maintains the optimal policy. Empirical evidence shows that the k-SP constraint indeed significantly improves the sample complexity over baseline algorithms in benchmarking environments. \n\nWhile I certainly agree that sparse reward tasks is important, one of the biggest challenges in sparse reward tasks is when the environment has a large state space. When the state space is finite and small, there is no doubt that many algorithms (even  algorithms with no deep learning method) can solve the task. However, for a large state space and sparse reward task (such as the robotics tasks in Mujoco environments[1], Montezuma's Revenge in Atari environments), I'm not sure whether the k-SP constraint is enough for efficiently solve the task. The k-SP constraint essentially tells the agent to avoid going to duplicated states. Even so, due to large state space, it is unlikely that the agent can observe non-zero reward without extra guidance. To be more specific, could the author(s) provide more discussion compared to HER (Hindsight Experience Replay) and hierarchical reinforcement learning?\n\nOverall, the paper is well-written, and the empirical results are clear. Therefore, I would recommend a weak acceptance.\n\n[1] http://gym.openai.com/envs/#robotics", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clean and intuitive method preventing over-exploration for sparse reward setting", "review": "The paper proposes a novel k-shortest-path constraint that prevents over-exploration by exploit the combinatorial structure (i.e. shortest path) of sparse reward tasks. It is proved theoretically that the k-shortest-path constraint maintains the optimal policy. Empirical evidence shows that the k-SP constraint indeed significantly improves the sample complexity over baseline algorithms in benchmarking environments. \n\nWhile I certainly agree that sparse reward tasks is important, one of the biggest challenges in sparse reward tasks is when the environment has a large state space. When the state space is finite and small, there is no doubt that many algorithms (even  algorithms with no deep learning method) can solve the task. However, for a large state space and sparse reward task (such as the robotics tasks in Mujoco environments[1], Montezuma's Revenge in Atari environments), I'm not sure whether the k-SP constraint is enough for efficiently solve the task. The k-SP constraint essentially tells the agent to avoid going to duplicated states. Even so, due to large state space, it is unlikely that the agent can observe non-zero reward without extra guidance. To be more specific, could the author(s) provide more discussion compared to HER (Hindsight Experience Replay) and hierarchical reinforcement learning?\n\nOverall, the paper is well-written, and the empirical results are clear. Therefore, I would recommend a weak acceptance.\n\n[1] http://gym.openai.com/envs/#robotics", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603833985420}], "openreview_url": "https://openreview.net/forum?id=Y-Wl1l0Va-", "arxiv_id": "2107.06405", "paper_pdf": "papers/Y-Wl1l0Va-.pdf", "paper_pdf_sha256": "b85c620475245eabf55d17161de0183016a3850a3d5ddd2627d6f96f6ff6b26f", "paper_pdf_bytes": 45923891, "paper_pdf_source": "openreview", "code_url": "https://github.com/srsohn/shortest-path-rl", "code_repository": "srsohn/shortest-path-rl", "code_commit": "60094c4ce116bfd63506e48ab40ea686314bb913", "code_archive": "repos/Y-Wl1l0Va-.zip", "code_archive_sha256": "b36ac25302339b55c2f39663dcf9260ce4946db2a9122e2d5141029123326a39", "code_archive_bytes": 1000583, "code_file_count": 22, "code_extensions": {".py": 17, ".sh": 5}, "github_disk_usage_kb": 957, "github_languages": {"Python": 116594, "Starlark": 31093, "Shell": 3933}, "github_archived": false, "github_pushed_at": "2021-07-19T03:17:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/shortest-path-constrained-reinforcement-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1xwA34KDB", "year": 2020, "status": "rejected", "title": "Learning Invariants through Soft Unification", "authors": ["Nuri Cingillioglu", "Alessandra Russo"], "authorids": ["nuri.cingillioglu13@imperial.ac.uk", "a.russo@imperial.ac.uk"], "authors_source": "OpenReview API", "abstract": "Human reasoning involves recognising common underlying principles across many examples by utilising variables. The by-products of such reasoning are invariants that capture patterns across examples such as \"if someone went somewhere then they are there\" without mentioning specific people or places. Humans learn what variables are and how to use them at a young age, and the question this paper addresses is whether machines can also learn and use variables solely from examples without requiring human pre-engineering. We propose Unification Networks that incorporate soft unification into neural networks to learn variables and by doing so lift examples into invariants that can then be used to solve a given task. We evaluate our approach on four datasets to demonstrate that learning invariants captures patterns in the data and can improve performance over baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rkgXIqLJqH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper268/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a neural network approach to variable unification and\nreasoning by example as a way to mimic the human ability to identify invariant\npatterns in examples and then apply them more generally in practice.\nThis general idea of identifying invariates and mapping new instances to\nthem is well motivated by the authors, citing work in philosophy of mind,\ncognitive science, and developmental psychology.\n\nThe authors go on to propose MLP, CNN and Memory Network models\nof unification for sequence, grid, and story reasoning tasks respectively.\nExperiments on the sequence and grid datasets demonstrate the data efficiency\nof this approach. MLP and CNN models with unification achieve near perfect\nperformance in fewer iterations (an order of magnitude fewer in the MLP case!)\n than their non-unification enabled counter parts.\nUnification enabled models also demonstrate high performance in a reduced\ntraining set setting (using only 50 training examples).\nWhile this is encouraging, these are very simple toy tasks.\n\nI also am in doubt as to whether the representation of these problems\ncauses some issues. In the sequence task, one question the models are\ntrying to solve is what symbol is the head or tail of the sequence.\nModeling variables over the sequence of symbols here is, in a sense, the\nwrong object of study. The position of the symbols would need to be\nrepresented, e.g.\n\na b c d\n1 4 3 1\n\nwhere I've represented positions as a-d, and the learned invariant about\nhead questions would be:\n\nX:a b c d\nY:1 4 3 1\n\nAs is, by mapping symbols and not positions to variables, one cannot,\nat the variable level distinguish between the two 1s in the sequence above.\nMy guess is that in practice the bi-GRU model that produces embedding\nfeatures of the symbols in sequence is implicitly representing head/tail\npositioning.\n\nSimilar arguments could be made about the grid example.\n\nI don't find the experiments/analysis on the bAbI dataset very convincing.\nFor instance, in the example given in Figure 4b (reproduced below)\nis shown as an example of\ntemporal reasoning, where a symbol Z is mapped to the\nword morning (a symbol distinguishing a time), and the question asked is\nwhere was Bill before school.\nIf logical reasoning is being used to solve this question, surely the\nsymbol 'before' must also be represented as a variable. Its possible that\nthe model is instead learning a trick about mutual exclusivity, i.e. that\nY:school is the only location symbol not mentioned in question but this\ncould fail as a general strategy.\n\n\nthis Z:morning X:bill went to the Y:school\nyesterday X:bill journeyed to the A:park\nwhere was X:bill before the Y:school\nA:park\n\nFigure 4b\n\nIt would make for a much more interesting paper if the authors took\nexamples such as these and formed counter-factuals to probe the way\nthe models are answering the questions. E.g., transforming the question\nin 4b to \"where was X:bill today\" or \"where was X:bill after school.\"\n\nBecause the authors use soft unification, interpretability is difficult\nto assess. Interpretability is crucial here because to claim that unification and reasoning by logical induction\nis being used to solve tasks, it becomes important to show how the neural networks\nmake their decisions. Given the instances of extra variables and one to many\nmappings on the bAbI dataset it seems very likely that the models are not\nsolving many tasks\nusing unification as it would be possible to learn to use the symbols directly\nto learn to answer. As such, I think these issues are not addressed in the\npaper sufficiently to warrant acceptance.\n\n\nMinor Notes\n\n- In definition 1, the definition of Variable is a little confusing because there are two different senses of the word in use. I understand them to be (1) Variable (X) in the logical template that is intended to be learned and used in problem solving, and\n(2) variable (x) in the neural network model that is a soft asignment of\nthe Variable to a default symbol s. It would be nice if this distinction could\nbe noted or made clearer.\n\n- In the definition 2, in the phrase \"is the invariant example such as a tokenized story\" it might be worth stating that the tokens are the symbols in S.\n\n\n- My understanding is that each unique symbol in the invariate is a potential\nvariable. Does this mean there are no co-referent symbols in the invariate?\nWould be helpful to state whether babi contains co-referent expressions\nand how these might affect the model.\n\n\n- It might be interesting to see how model architecture affects variable\nlearning. For example, does a CNN result in more sensible variable\nassignments  than the mlp on a flattened representation of the grid problem?\n\n\n- What is the strongly supervised case? These are token level annotations I think (at least for babi) but it might be good to specify in more detail what\nthey  are.\n\n- The figure and explanation of the UMN are not very clear. From the figure\nis does not seem that the variables interact with the memory at all. More\nspace could be devoted to this section.\n\nPossibly Relevant Related Work\n\nBrenden Lake. Compositional generalization through metasequence-to-sequence learning. NeurIPS 2019.\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #4", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "The authors propose a neural network approach to variable unification and\nreasoning by example as a way to mimic the human ability to identify invariant\npatterns in examples and then apply them more generally in practice.\nThis general idea of identifying invariates and mapping new instances to\nthem is well motivated by the authors, citing work in philosophy of mind,\ncognitive science, and developmental psychology.\n\nThe authors go on to propose MLP, CNN and Memory Network models\nof unification for sequence, grid, and story reasoning tasks respectively.\nExperiments on the sequence and grid datasets demonstrate the data efficiency\nof this approach. MLP and CNN models with unification achieve near perfect\nperformance in fewer iterations (an order of magnitude fewer in the MLP case!)\n than their non-unification enabled counter parts.\nUnification enabled models also demonstrate high performance in a reduced\ntraining set setting (using only 50 training examples).\nWhile this is encouraging, these are very simple toy tasks.\n\nI also am in doubt as to whether the representation of these problems\ncauses some issues. In the sequence task, one question the models are\ntrying to solve is what symbol is the head or tail of the sequence.\nModeling variables over the sequence of symbols here is, in a sense, the\nwrong object of study. The position of the symbols would need to be\nrepresented, e.g.\n\na b c d\n1 4 3 1\n\nwhere I've represented positions as a-d, and the learned invariant about\nhead questions would be:\n\nX:a b c d\nY:1 4 3 1\n\nAs is, by mapping symbols and not positions to variables, one cannot,\nat the variable level distinguish between the two 1s in the sequence above.\nMy guess is that in practice the bi-GRU model that produces embedding\nfeatures of the symbols in sequence is implicitly representing head/tail\npositioning.\n\nSimilar arguments could be made about the grid example.\n\nI don't find the experiments/analysis on the bAbI dataset very convincing.\nFor instance, in the example given in Figure 4b (reproduced below)\nis shown as an example of\ntemporal reasoning, where a symbol Z is mapped to the\nword morning (a symbol distinguishing a time), and the question asked is\nwhere was Bill before school.\nIf logical reasoning is being used to solve this question, surely the\nsymbol 'before' must also be represented as a variable. Its possible that\nthe model is instead learning a trick about mutual exclusivity, i.e. that\nY:school is the only location symbol not mentioned in question but this\ncould fail as a general strategy.\n\n\nthis Z:morning X:bill went to the Y:school\nyesterday X:bill journeyed to the A:park\nwhere was X:bill before the Y:school\nA:park\n\nFigure 4b\n\nIt would make for a much more interesting paper if the authors took\nexamples such as these and formed counter-factuals to probe the way\nthe models are answering the questions. E.g., transforming the question\nin 4b to \"where was X:bill today\" or \"where was X:bill after school.\"\n\nBecause the authors use soft unification, interpretability is difficult\nto assess. Interpretability is crucial here because to claim that unification and reasoning by logical induction\nis being used to solve tasks, it becomes important to show how the neural networks\nmake their decisions. Given the instances of extra variables and one to many\nmappings on the bAbI dataset it seems very likely that the models are not\nsolving many tasks\nusing unification as it would be possible to learn to use the symbols directly\nto learn to answer. As such, I think these issues are not addressed in the\npaper sufficiently to warrant acceptance.\n\n\nMinor Notes\n\n- In definition 1, the definition of Variable is a little confusing because there are two different senses of the word in use. I understand them to be (1) Variable (X) in the logical template that is intended to be learned and used in problem solving, and\n(2) variable (x) in the neural network model that is a soft asignment of\nthe Variable to a default symbol s. It would be nice if this distinction could\nbe noted or made clearer.\n\n- In the definition 2, in the phrase \"is the invariant example such as a tokenized story\" it might be worth stating that the tokens are the symbols in S.\n\n\n- My understanding is that each unique symbol in the invariate is a potential\nvariable. Does this mean there are no co-referent symbols in the invariate?\nWould be helpful to state whether babi contains co-referent expressions\nand how these might affect the model.\n\n\n- It might be interesting to see how model architecture affects variable\nlearning. For example, does a CNN result in more sensible variable\nassignments  than the mlp on a flattened representation of the grid problem?\n\n\n- What is the strongly supervised case? These are token level annotations I think (at least for babi) but it might be good to specify in more detail what\nthey  are.\n\n- The figure and explanation of the UMN are not very clear. From the figure\nis does not seem that the variables interact with the memory at all. More\nspace could be devoted to this section.\n\nPossibly Relevant Related Work\n\nBrenden Lake. Compositional generalization through metasequence-to-sequence learning. NeurIPS 2019.\n\n\n\n\n\n"}, "tcdate": 1571936842833}, {"id": "HygP8HETtH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper268/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a novel approach for learning invariants that can capture underlying patterns in the tasks through Unification Networks. This effectively allows the machine to learn the notion of `variable`, which is a symbol that can take on different values. \n\nPros:\nThe authors evaluated and presented empirical results on four common benchmark datasets, showing superiority over plain baseline without unification. \nThey further performed analysis on the learned invariants, and verified the sensibility. \nThe paper overall is well written and structured.\n\nCons:\nDespite its superiority over plain baseline, the paper does not provide thorough comparison with other state-of-the-art methods on reasoning related tasks.\n\nSome of the technical details regarding the choice of hyperparameters are missing. For example:\nIn section 6, what’s the rationale of setting $t$ differently for bAbI solely?\nIn Equation 5, how is the sparsity regularization parameter $\\tau$ chosen optimally for a particular task? A bit more discussion on these choices would be helpful. \n\nOverall, this paper presents a seemingly promising architecture capable of learning and using variables, with the caveat for lack of experiments and comparison with other state-of-the-art methods. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper presents a novel approach for learning invariants that can capture underlying patterns in the tasks through Unification Networks. This effectively allows the machine to learn the notion of `variable`, which is a symbol that can take on different values. \n\nPros:\nThe authors evaluated and presented empirical results on four common benchmark datasets, showing superiority over plain baseline without unification. \nThey further performed analysis on the learned invariants, and verified the sensibility. \nThe paper overall is well written and structured.\n\nCons:\nDespite its superiority over plain baseline, the paper does not provide thorough comparison with other state-of-the-art methods on reasoning related tasks.\n\nSome of the technical details regarding the choice of hyperparameters are missing. For example:\nIn section 6, what’s the rationale of setting $t$ differently for bAbI solely?\nIn Equation 5, how is the sparsity regularization parameter $\\tau$ chosen optimally for a particular task? A bit more discussion on these choices would be helpful. \n\nOverall, this paper presents a seemingly promising architecture capable of learning and using variables, with the caveat for lack of experiments and comparison with other state-of-the-art methods. \n"}, "tcdate": 1571796302554}, {"id": "BJxGVXNpKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper268/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper explores a very interesting idea: can a model learn what variables are and how to use them? Unfortunately, the paper doesn't seem quite ready: the model description was very hard to follow and it's not clear the approach has found a compelling use case.\n\nI read the paper carefully three times, and try as I might, I simply can't get my head around the entire architecture. The modeling section jumps straight into a series of definitions, without trying to build intuition or provide a worked example. There is an example in Figure 2, but it isn't really explained and I didn't find it helpful. Unification seems to be implemented as a form of attention (or self-attention) where the model can control the degree to which a symbol acts as variable. But the relationship between soft unification and attention isn't really spelled out -- what's the same, what's different? Ultimately it's not clear to me what the model is attending over during soft unification.\n\nThere are various other aspects of the paper that aren't clear:\n- strong vs. weak supervision\n- comparison models DMN and IMA are not introduced at all, and include no references\n- the logical reasoning experiment is not clearly described\n- there is only a cursory conclusion\n\nI am not sure the model has found a compelling use case. On bAbi with weak supervision, the model is worse than the comparison models. It only slightly beats out memory networks with strong supervision. For logical reasoning, it's not clear what it is compared against or if the comparison is fair. The clearest win over standard networks is on the simple synthetic experiments.\n\nFinally, the authors mention the paper has a cognitive science motivation, in that \"Humans learn what variables are and how to use then at a young age\" or that \"symbolic thought with variables is learned...\", taking a strong \"nurture\" stance on the origin of variables. But variables could very well be innate and simply early emerging. Any discussion of the origin of variables in the mind requires more nuance.\n\nI am excited about this research direction, and it could ultimately be a very nice contribution as the work matures. I don't think the paper is ready in its current form.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "This paper explores a very interesting idea: can a model learn what variables are and how to use them? Unfortunately, the paper doesn't seem quite ready: the model description was very hard to follow and it's not clear the approach has found a compelling use case.\n\nI read the paper carefully three times, and try as I might, I simply can't get my head around the entire architecture. The modeling section jumps straight into a series of definitions, without trying to build intuition or provide a worked example. There is an example in Figure 2, but it isn't really explained and I didn't find it helpful. Unification seems to be implemented as a form of attention (or self-attention) where the model can control the degree to which a symbol acts as variable. But the relationship between soft unification and attention isn't really spelled out -- what's the same, what's different? Ultimately it's not clear to me what the model is attending over during soft unification.\n\nThere are various other aspects of the paper that aren't clear:\n- strong vs. weak supervision\n- comparison models DMN and IMA are not introduced at all, and include no references\n- the logical reasoning experiment is not clearly described\n- there is only a cursory conclusion\n\nI am not sure the model has found a compelling use case. On bAbi with weak supervision, the model is worse than the comparison models. It only slightly beats out memory networks with strong supervision. For logical reasoning, it's not clear what it is compared against or if the comparison is fair. The clearest win over standard networks is on the simple synthetic experiments.\n\nFinally, the authors mention the paper has a cognitive science motivation, in that \"Humans learn what variables are and how to use then at a young age\" or that \"symbolic thought with variables is learned...\", taking a strong \"nurture\" stance on the origin of variables. But variables could very well be innate and simply early emerging. Any discussion of the origin of variables in the mind requires more nuance.\n\nI am excited about this research direction, and it could ultimately be a very nice contribution as the work matures. I don't think the paper is ready in its current form.\n"}, "tcdate": 1571795754252}], "openreview_url": "https://openreview.net/forum?id=r1xwA34KDB", "arxiv_id": "1909.07328", "paper_pdf": "papers/r1xwA34KDB.pdf", "paper_pdf_sha256": "44c2f65c7c10ef5b30a6fe26efbcc40143464f94fd43e3e6c156aab03cb82cf6", "paper_pdf_bytes": 434864, "paper_pdf_source": "openreview", "code_url": "https://github.com/nuric/softuni", "code_repository": "nuric/softuni", "code_commit": "86dc865dddddb1445caea61f666d0e1abc669067", "code_archive": "repos/r1xwA34KDB.zip", "code_archive_sha256": "6b4ccb9fbedb24c76e4c87d3d590ccd52cb42baeb02e455cd460c7289c1e9689", "code_archive_bytes": 1357338, "code_file_count": 11, "code_extensions": {".py": 7, ".ipynb": 3, ".sh": 1}, "github_disk_usage_kb": 2559, "github_languages": {"Jupyter Notebook": 1855523, "Python": 114820, "Shell": 548}, "github_archived": false, "github_pushed_at": "2020-10-17T11:35:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-invariants-through-soft-unification"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TJbfURjxyL", "year": 2026, "status": "rejected", "title": "SAISA: Towards Multimodal Large Language Models with Both Training and Inference Efficiency", "authors": ["Qianhao Yuan", "yanjiang liu", "Mo guozhao", "Yaojie Lu", "Hongyu Lin", "Jia Zheng", "Ben He", "Xianpei Han", "Le Sun"], "authorids": ["~Qianhao_Yuan1", "~yanjiang_liu1", "~Mo_guozhao1", "~Yaojie_Lu1", "~Hongyu_Lin1", "~Jia_Zheng3", "~Ben_He1", "~Xianpei_Han1", "~Le_Sun1"], "authors_source": "OpenReview API", "abstract": "Multimodal Large Language Models (MLLMs) mainly fall into two architectures, each involving a trade-off between training and inference efficiency: embedding space alignment (e.g. LLaVA) is inefficient during inference, while cross-attention space alignment (e.g. Flamingo) is inefficient in training. In this paper, we compare these two architectures and identify key factors for building efficient MLLMs. A primary difference between them lies in how attention is applied to visual tokens, particularly in their interactions with each other. To investigate whether attention among visual tokens is necessary, we propose a new self-attention mechanism, NAAViT (No Attention Among Visual Tokens), which eliminates this type of attention. Our pilot experiment on LLaVA-1.5 shows that attention among visual tokens is highly redundant. Based on these insights, we introduce SAISA (Self-Attention Input Space Alignment), a novel architecture that enhances both training and inference efficiency. SAISA directly aligns visual features with the input spaces of NAAViT self-attention blocks, reducing computational overhead in both self-attention blocks and feed-forward networks (FFNs). Compared with the LLaVA-1.5 architecture, SAISA reduces the inference FLOPs by 66% and the training budget by 26%, while achieving superior performance in terms of accuracy. Comprehensive ablation studies further validate the effectiveness of SAISA across various LLMs and visual encoders. The code and models will be publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "gg1oqDeQAn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16628/Reviewer_rU2Y"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes SAISA (Self-Attention Input Space Alignment), a novel and highly efficient architecture for Multimodal Large Language Models (MLLMs). The work is motivated by a key finding: that self-attention among visual tokens within the LLM is \"highly redundant.\" The authors first validate this by proposing NAAVIT (No Attention Among Visual Tokens), a modified self-attention mechanism that eliminates this $O(v^2)$ computation, and show it improves performance over standard attention.Building on this, the SAISA architecture aligns visual features directly with the input spaces of each NAAVIT self-attention block in the LLM. This clever design allows visual tokens to participate in attention (text can attend to visual tokens, visual tokens can attend to text) but bypasses two major computational bottlenecks: (1) self-attention among visual tokens and (2) all FFN computations for visual tokens.The paper demonstrates that, compared to the LLaVA-1.5 architecture, SAISA reduces inference FLOPs by 66% and the training budget by 26%, all while achieving superior performance on a wide range of MLLM benchmarks.", "review_text": "This paper proposes SAISA (Self-Attention Input Space Alignment), a novel and highly efficient architecture for Multimodal Large Language Models (MLLMs). The work is motivated by a key finding: that self-attention among visual tokens within the LLM is \"highly redundant.\" The authors first validate this by proposing NAAVIT (No Attention Among Visual Tokens), a modified self-attention mechanism that eliminates this $O(v^2)$ computation, and show it improves performance over standard attention.Building on this, the SAISA architecture aligns visual features directly with the input spaces of each NAAVIT self-attention block in the LLM. This clever design allows visual tokens to participate in attention (text can attend to visual tokens, visual tokens can attend to text) but bypasses two major computational bottlenecks: (1) self-attention among visual tokens and (2) all FFN computations for visual tokens.The paper demonstrates that, compared to the LLaVA-1.5 architecture, SAISA reduces inference FLOPs by 66% and the training budget by 26%, all while achieving superior performance on a wide range of MLLM benchmarks.", "strengths": "- A 66% reduction in inference FLOPs and a 26% reduction in training budget (vs. LLaVA-1.5) are extremely significant practical contributions. This makes SOTA-level MLLMs more accessible to train and cheaper to deploy.\n\n- The core finding that attention among visual tokens is redundant—and that FFNs for visual tokens can also be skipped—is a fundamental and novel insight into MLLM architecture.\n\n- This is not just an efficiency-focused paper (e.g., pruning, quantization). SAISA outperforms its LLaVA-1.5 baseline across a wide array of benchmarks (Tables 2, 3, 5), indicating the architectural changes also serve as a good regularizer or are simply a better design.\n\n- The ablation studies in Table 5 compellingly show that SAISA's benefits are not specific to one model pair but apply broadly across different LLM families (including GQA models like Mistral) and visual encoder types (ViT, ConvNeXt).", "weaknesses": "- The paper does an excellent job proving that attention among visual tokens is redundant but doesn't fully explore why. Is it because the visual encoder has already \"bound\" all necessary spatial/object relationships? Does removing it act as a regularizer, preventing the model from \"overthinking\" the visual features? A deeper analytical probe (e.g., attention visualization, feature analysis) into why NAAVIT works better would make this paper even stronger.\n\n- The SAISA projector consists of n separate 2-layer MLPs (one for each of the n LLM layers), whereas the LLaVA projector is a single 2-layer MLP. This significantly increases the number of new parameters. The paper shows that training is still 26% faster (which is the more important metric), likely due to reduced computation in the frozen LLM blocks. However, an explicit comparison of the number of new, trainable parameters (SAISA projector vs. LLaVA projector) would be helpful for transparency. The paper's \"shared MLP\" pre-training strategy seems to be the key to managing this, but this could be clearer.", "questions": "- The finding that NAAVIT (no visual self-attention) outperforms vanilla self-attention (Table 1) is fascinating. Do the authors have a hypothesis for why this is the case? Does it act as a regularizer? Or is it possible that the visual encoder's features are already so robust that further self-attention only adds noise?\n\n- Could you please provide the concrete number of new, trainable parameters introduced by the SAISA projector (e.g., for Vicuna-7B with 32 layers) and compare this to the number of parameters in the LLaVA-1.5 projector? This would clarify the trade-off being made (more parameters in the projector for far less computation in the LLM).\n\n- How does SAISA compare to orthogonal efficiency methods like visual token pruning (e.g., FastV, cited in the paper)? SAISA's approach is to keep all 576 tokens but process them more cheaply. Would it be better to (e.g.) prune to 64 tokens and use a standard LLaVA architecture? Or can these methods be combined (e.g., use SAISA with a resampler, as in Table 7)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes SAISA (Self-Attention Input Space Alignment), a novel and highly efficient architecture for Multimodal Large Language Models (MLLMs). The work is motivated by a key finding: that self-attention among visual tokens within the LLM is \"highly redundant.\" The authors first validate this by proposing NAAVIT (No Attention Among Visual Tokens), a modified self-attention mechanism that eliminates this $O(v^2)$ computation, and show it improves performance over standard attention.Building on this, the SAISA architecture aligns visual features directly with the input spaces of each NAAVIT self-attention block in the LLM. This clever design allows visual tokens to participate in attention (text can attend to visual tokens, visual tokens can attend to text) but bypasses two major computational bottlenecks: (1) self-attention among visual tokens and (2) all FFN computations for visual tokens.The paper demonstrates that, compared to the LLaVA-1.5 architecture, SAISA reduces inference FLOPs by 66% and the training budget by 26%, all while achieving superior performance on a wide range of MLLM benchmarks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- A 66% reduction in inference FLOPs and a 26% reduction in training budget (vs. LLaVA-1.5) are extremely significant practical contributions. This makes SOTA-level MLLMs more accessible to train and cheaper to deploy.\n\n- The core finding that attention among visual tokens is redundant—and that FFNs for visual tokens can also be skipped—is a fundamental and novel insight into MLLM architecture.\n\n- This is not just an efficiency-focused paper (e.g., pruning, quantization). SAISA outperforms its LLaVA-1.5 baseline across a wide array of benchmarks (Tables 2, 3, 5), indicating the architectural changes also serve as a good regularizer or are simply a better design.\n\n- The ablation studies in Table 5 compellingly show that SAISA's benefits are not specific to one model pair but apply broadly across different LLM families (including GQA models like Mistral) and visual encoder types (ViT, ConvNeXt).", "weaknesses": "- The paper does an excellent job proving that attention among visual tokens is redundant but doesn't fully explore why. Is it because the visual encoder has already \"bound\" all necessary spatial/object relationships? Does removing it act as a regularizer, preventing the model from \"overthinking\" the visual features? A deeper analytical probe (e.g., attention visualization, feature analysis) into why NAAVIT works better would make this paper even stronger.\n\n- The SAISA projector consists of n separate 2-layer MLPs (one for each of the n LLM layers), whereas the LLaVA projector is a single 2-layer MLP. This significantly increases the number of new parameters. The paper shows that training is still 26% faster (which is the more important metric), likely due to reduced computation in the frozen LLM blocks. However, an explicit comparison of the number of new, trainable parameters (SAISA projector vs. LLaVA projector) would be helpful for transparency. The paper's \"shared MLP\" pre-training strategy seems to be the key to managing this, but this could be clearer.", "questions": "- The finding that NAAVIT (no visual self-attention) outperforms vanilla self-attention (Table 1) is fascinating. Do the authors have a hypothesis for why this is the case? Does it act as a regularizer? Or is it possible that the visual encoder's features are already so robust that further self-attention only adds noise?\n\n- Could you please provide the concrete number of new, trainable parameters introduced by the SAISA projector (e.g., for Vicuna-7B with 32 layers) and compare this to the number of parameters in the LLaVA-1.5 projector? This would clarify the trade-off being made (more parameters in the projector for far less computation in the LLM).\n\n- How does SAISA compare to orthogonal efficiency methods like visual token pruning (e.g., FastV, cited in the paper)? SAISA's approach is to keep all 576 tokens but process them more cheaply. Would it be better to (e.g.) prune to 64 tokens and use a standard LLaVA architecture? Or can these methods be combined (e.g., use SAISA with a resampler, as in Table 7)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761864057971}, {"id": "VxTt0Mu5oN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16628/Reviewer_JnP5"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper addresses the trade-off between training and inference efficiency in MLLMs. Traditional embedding space alignment models (e.g., LLaVA) are efficient to train but slow to infer, while cross-attention space alignment*models (e.g., Flamingo) offer fast inference but incur high training costs. After analyzing the computational bottlenecks of both paradigms, the authors argue that self-attention among visual tokens is largely redundant. They propose NAAViT to remove visual–visual attention and build upon it to design SAISA, which maps visual features directly to the input spaces of self-attention layers in the LLM while skipping FFN operations on visual tokens. Experiments show that SAISA achieves a 26% reduction in training cost and a 66% reduction in inference FLOPs, with slightly improved accuracy—thus realizing efficiency in both training and inference.", "review_text": "This paper addresses the trade-off between training and inference efficiency in MLLMs. Traditional embedding space alignment models (e.g., LLaVA) are efficient to train but slow to infer, while cross-attention space alignment*models (e.g., Flamingo) offer fast inference but incur high training costs. After analyzing the computational bottlenecks of both paradigms, the authors argue that self-attention among visual tokens is largely redundant. They propose NAAViT to remove visual–visual attention and build upon it to design SAISA, which maps visual features directly to the input spaces of self-attention layers in the LLM while skipping FFN operations on visual tokens. Experiments show that SAISA achieves a 26% reduction in training cost and a 66% reduction in inference FLOPs, with slightly improved accuracy—thus realizing efficiency in both training and inference.", "strengths": "1. **Clear problem statement and method design.** The paper tackles the key challenge of improving MLLM computational efficiency—a crucial factor for deployment and energy optimization. The combination of NAAViT’s `removal of visual–text attention` and SAISA’s `layer-wise input-space alignment` provides a novel trade-off between existing embedding- and cross-attention-based architectures.\n\n2. **Fair experimental comparison.** All main experiments are conducted under identical data and training settings as LLaVA-1.5, ensuring fair comparison. The ablations demonstrate generality across different LLMs (Vicuna, Llama3, Mistral) and visual encoders (CLIP, SigLIP, ConvNeXt), while also studying projector design and pretraining strategies.\n\n3. **Strong reproducibility.**  The paper includes detailed hyperparameters, training settings, and FLOPs derivations, providing sufficient information for replication.", "weaknesses": "1. **Insufficient evidence for the `redundancy` claim.** The paper concludes that attention among visual tokens is redundant, but this is supported only by a marginal performance gain of NAAViT over standard self-attention. Such improvement could stem from optimization dynamics or implicit regularization rather than true redundancy. This would be meaningful: Providing attention heatmaps or token correlation analyses, and testing on multi-object or spatial reasoning tasks would help verify whether visual–visual interactions are indeed dispensable.\n\n2. **Unclear source of efficiency gains.** The study does not isolate the contributions of NAAViT versus the removal of FFNs on visual tokens. Without reporting FLOPs and performance when using NAAViT alone, it is unclear which component drives the observed acceleration.\n\n3. **Potential incompatibility of NAAViT masks with inference accelerators.** In multi-image or interleaved dialogue scenarios, NAAViT introduces fragmented attention masks that may reduce the efficiency of Flash-Attention 2 or vLLM kernels.\n\n4. **Outdated baseline selection for token compression.** Easy token unshuffle methods can reduce token counts to 1/4 or 1/9 with negligible information loss. Yet the paper lacks comparison with recent **parameter-free token optimization methods**, and only contrasts with older pruning-based approaches from 2024 (e.g., FastV, VTW), which weakens its claims.\n\n5. **Limited experimental scope.** All primary results are based on Vicuna-7B and LLaVA datasets, without evaluation on larger models (e.g., 13B) or other modalities such as videos and multi-image inputs.", "questions": "1. Could the authors provide results using NAAViT alone (while keeping FFNs for visual tokens) to disentangle the sources of efficiency?\n2. Does the sparse attention mask in NAAViT affect compatibility or speedup when using Flash-Attention 2 or vLLM inference engines, especially for interleaved or long-context inputs?\n3. See `Weaknesses` for additional clarifications.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the trade-off between training and inference efficiency in MLLMs. Traditional embedding space alignment models (e.g., LLaVA) are efficient to train but slow to infer, while cross-attention space alignment*models (e.g., Flamingo) offer fast inference but incur high training costs. After analyzing the computational bottlenecks of both paradigms, the authors argue that self-attention among visual tokens is largely redundant. They propose NAAViT to remove visual–visual attention and build upon it to design SAISA, which maps visual features directly to the input spaces of self-attention layers in the LLM while skipping FFN operations on visual tokens. Experiments show that SAISA achieves a 26% reduction in training cost and a 66% reduction in inference FLOPs, with slightly improved accuracy—thus realizing efficiency in both training and inference.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "1. **Clear problem statement and method design.** The paper tackles the key challenge of improving MLLM computational efficiency—a crucial factor for deployment and energy optimization. The combination of NAAViT’s `removal of visual–text attention` and SAISA’s `layer-wise input-space alignment` provides a novel trade-off between existing embedding- and cross-attention-based architectures.\n\n2. **Fair experimental comparison.** All main experiments are conducted under identical data and training settings as LLaVA-1.5, ensuring fair comparison. The ablations demonstrate generality across different LLMs (Vicuna, Llama3, Mistral) and visual encoders (CLIP, SigLIP, ConvNeXt), while also studying projector design and pretraining strategies.\n\n3. **Strong reproducibility.**  The paper includes detailed hyperparameters, training settings, and FLOPs derivations, providing sufficient information for replication.", "weaknesses": "1. **Insufficient evidence for the `redundancy` claim.** The paper concludes that attention among visual tokens is redundant, but this is supported only by a marginal performance gain of NAAViT over standard self-attention. Such improvement could stem from optimization dynamics or implicit regularization rather than true redundancy. This would be meaningful: Providing attention heatmaps or token correlation analyses, and testing on multi-object or spatial reasoning tasks would help verify whether visual–visual interactions are indeed dispensable.\n\n2. **Unclear source of efficiency gains.** The study does not isolate the contributions of NAAViT versus the removal of FFNs on visual tokens. Without reporting FLOPs and performance when using NAAViT alone, it is unclear which component drives the observed acceleration.\n\n3. **Potential incompatibility of NAAViT masks with inference accelerators.** In multi-image or interleaved dialogue scenarios, NAAViT introduces fragmented attention masks that may reduce the efficiency of Flash-Attention 2 or vLLM kernels.\n\n4. **Outdated baseline selection for token compression.** Easy token unshuffle methods can reduce token counts to 1/4 or 1/9 with negligible information loss. Yet the paper lacks comparison with recent **parameter-free token optimization methods**, and only contrasts with older pruning-based approaches from 2024 (e.g., FastV, VTW), which weakens its claims.\n\n5. **Limited experimental scope.** All primary results are based on Vicuna-7B and LLaVA datasets, without evaluation on larger models (e.g., 13B) or other modalities such as videos and multi-image inputs.", "questions": "1. Could the authors provide results using NAAViT alone (while keeping FFNs for visual tokens) to disentangle the sources of efficiency?\n2. Does the sparse attention mask in NAAViT affect compatibility or speedup when using Flash-Attention 2 or vLLM inference engines, especially for interleaved or long-context inputs?\n3. See `Weaknesses` for additional clarifications.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761586814568}, {"id": "zSDCqEniQN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16628/Reviewer_Cqp8"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper addresses the well-known trade-off between training and inference efficiency in MLLMs. The authors categorize existing architectures into two types: embedding space alignment (e.g., LLaVA), which is training-efficient but inference-inefficient due to long visual token sequences, and cross-attention space alignment (e.g., Flamingo), which is inference-efficient but training-inefficient due to a large number of new parameters.\n\nTo tackle this, the paper first proposes NAAViT, a self-attention mechanism that eliminates the computationally expensive attention among visual tokens. A pilot study suggests this form of attention is redundant. Building on this, the authors introduce SAISA (Self-Attention Input Space Alignment), a novel architecture that employs NAAViT blocks and aligns visual features directly with the input spaces of each self-attention layer in the LLM. This approach avoids both the quadratic complexity of visual self-attention and the need to pass visual tokens through the FFNs. The experimental results demonstrate that SAISA significantly reduces inference FLOPs (by 66%) and training costs (by 26%) compared to its primary baseline, LLaVA-1.5, while achieving competitive or superior performance on several benchmarks.", "review_text": "This paper addresses the well-known trade-off between training and inference efficiency in MLLMs. The authors categorize existing architectures into two types: embedding space alignment (e.g., LLaVA), which is training-efficient but inference-inefficient due to long visual token sequences, and cross-attention space alignment (e.g., Flamingo), which is inference-efficient but training-inefficient due to a large number of new parameters.\n\nTo tackle this, the paper first proposes NAAViT, a self-attention mechanism that eliminates the computationally expensive attention among visual tokens. A pilot study suggests this form of attention is redundant. Building on this, the authors introduce SAISA (Self-Attention Input Space Alignment), a novel architecture that employs NAAViT blocks and aligns visual features directly with the input spaces of each self-attention layer in the LLM. This approach avoids both the quadratic complexity of visual self-attention and the need to pass visual tokens through the FFNs. The experimental results demonstrate that SAISA significantly reduces inference FLOPs (by 66%) and training costs (by 26%) compared to its primary baseline, LLaVA-1.5, while achieving competitive or superior performance on several benchmarks.", "strengths": "1. Well-Motivated Problem: The paper targets a critical and practical challenge in the MLLM domain. Improving both training and inference efficiency is a highly valuable research direction.\n2. Intuitive and Simple Core Idea: The central concept of NAAViT—that attention between visual tokens within the LLM might be redundant—is simple to understand and implement. The pilot experiment provides good initial evidence for this hypothesis.\n3. Strong Empirical Efficiency Gains: The reported improvements in computational efficiency are substantial. A 66% reduction in inference FLOPs and a 26% reduction in training GPU hours compared to LLaVA-1.5 are impressive results that clearly demonstrate the practical benefit of the proposed architecture.\n4. Clear Presentation: The paper is well-written, and the diagrams (especially Figure 2 and 3) effectively illustrate the differences between existing architectures and the proposed SAISA model.", "weaknesses": "1. While the combination of components is new, the core ideas feel more incremental than foundational.\nThe primary mechanism, NAAViT, essentially modifies the self-attention in embedding-space models to mimic the behavior of cross-attention models like Flamingo, where text queries attend to visual keys/values without any V-V interaction. The novelty lies in achieving this within a standard self-attention block via masking, rather than by introducing new cross-attention modules, thereby saving parameters. However, the conceptual leap is not substantial.\nThe idea of projecting features into multiple layers of a transformer is also not entirely new and has precedents in other architectures that seek to inject conditioning information throughout a network. The contribution seems to be a clever engineering combination of existing principles rather than a fundamentally new architectural paradigm.\n2. The paper's ablations are too limited to fully substantiate its claims and disentangle the sources of improvement.\nThe most critical missing ablation is a direct comparison between a model using only NAAViT (as in the pilot experiment) and the full SAISA model (NAAViT + layer-wise projection). The current presentation conflates the benefits of two distinct architectural changes. It is unclear how much of the performance gain and efficiency improvement comes from simply removing visual self-attention versus the more complex layer-wise projection scheme. A thorough analysis is needed to justify the added complexity of SAISA's projector.\nThe paper argues for projecting visual features to every layer. Is this necessary? An ablation studying the impact of projecting to only a subset of layers (e.g., early, middle, or late layers) would provide deeper insights into how and where visual information is most effectively integrated.\n3. The experimental comparison, while thorough against LLaVA-1.5, feels dated and lacks engagement with the latest state-of-the-art in efficient MLLMs.\nThe MLLM field is advancing at an extremely rapid pace. While LLaVA-1.5 is a crucial baseline, many newer and more efficient models have been proposed since its release. The comparison set in Table 2 feels somewhat stale, missing more recent architectures that have been published in late 2024 or 2025 (as this paper targets ICLR 2026).\nMore importantly, the paper does not adequately compare its architectural approach to orthogonal efficiency methods, particularly token pruning/merging. Models like FastV (which is cited but could be compared more directly) or other token reduction techniques achieve inference efficiency by dynamically removing visual tokens. SAISA's approach is to keep all tokens but reduce computation per token. A direct and fair comparison between these two competing strategies for efficiency is essential to contextualize SAISA's contribution. Is it better to prune tokens or to keep them all with cheaper attention? This question is left unanswered.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the well-known trade-off between training and inference efficiency in MLLMs. The authors categorize existing architectures into two types: embedding space alignment (e.g., LLaVA), which is training-efficient but inference-inefficient due to long visual token sequences, and cross-attention space alignment (e.g., Flamingo), which is inference-efficient but training-inefficient due to a large number of new parameters.\n\nTo tackle this, the paper first proposes NAAViT, a self-attention mechanism that eliminates the computationally expensive attention among visual tokens. A pilot study suggests this form of attention is redundant. Building on this, the authors introduce SAISA (Self-Attention Input Space Alignment), a novel architecture that employs NAAViT blocks and aligns visual features directly with the input spaces of each self-attention layer in the LLM. This approach avoids both the quadratic complexity of visual self-attention and the need to pass visual tokens through the FFNs. The experimental results demonstrate that SAISA significantly reduces inference FLOPs (by 66%) and training costs (by 26%) compared to its primary baseline, LLaVA-1.5, while achieving competitive or superior performance on several benchmarks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Well-Motivated Problem: The paper targets a critical and practical challenge in the MLLM domain. Improving both training and inference efficiency is a highly valuable research direction.\n2. Intuitive and Simple Core Idea: The central concept of NAAViT—that attention between visual tokens within the LLM might be redundant—is simple to understand and implement. The pilot experiment provides good initial evidence for this hypothesis.\n3. Strong Empirical Efficiency Gains: The reported improvements in computational efficiency are substantial. A 66% reduction in inference FLOPs and a 26% reduction in training GPU hours compared to LLaVA-1.5 are impressive results that clearly demonstrate the practical benefit of the proposed architecture.\n4. Clear Presentation: The paper is well-written, and the diagrams (especially Figure 2 and 3) effectively illustrate the differences between existing architectures and the proposed SAISA model.", "weaknesses": "1. While the combination of components is new, the core ideas feel more incremental than foundational.\nThe primary mechanism, NAAViT, essentially modifies the self-attention in embedding-space models to mimic the behavior of cross-attention models like Flamingo, where text queries attend to visual keys/values without any V-V interaction. The novelty lies in achieving this within a standard self-attention block via masking, rather than by introducing new cross-attention modules, thereby saving parameters. However, the conceptual leap is not substantial.\nThe idea of projecting features into multiple layers of a transformer is also not entirely new and has precedents in other architectures that seek to inject conditioning information throughout a network. The contribution seems to be a clever engineering combination of existing principles rather than a fundamentally new architectural paradigm.\n2. The paper's ablations are too limited to fully substantiate its claims and disentangle the sources of improvement.\nThe most critical missing ablation is a direct comparison between a model using only NAAViT (as in the pilot experiment) and the full SAISA model (NAAViT + layer-wise projection). The current presentation conflates the benefits of two distinct architectural changes. It is unclear how much of the performance gain and efficiency improvement comes from simply removing visual self-attention versus the more complex layer-wise projection scheme. A thorough analysis is needed to justify the added complexity of SAISA's projector.\nThe paper argues for projecting visual features to every layer. Is this necessary? An ablation studying the impact of projecting to only a subset of layers (e.g., early, middle, or late layers) would provide deeper insights into how and where visual information is most effectively integrated.\n3. The experimental comparison, while thorough against LLaVA-1.5, feels dated and lacks engagement with the latest state-of-the-art in efficient MLLMs.\nThe MLLM field is advancing at an extremely rapid pace. While LLaVA-1.5 is a crucial baseline, many newer and more efficient models have been proposed since its release. The comparison set in Table 2 feels somewhat stale, missing more recent architectures that have been published in late 2024 or 2025 (as this paper targets ICLR 2026).\nMore importantly, the paper does not adequately compare its architectural approach to orthogonal efficiency methods, particularly token pruning/merging. Models like FastV (which is cited but could be compared more directly) or other token reduction techniques achieve inference efficiency by dynamically removing visual tokens. SAISA's approach is to keep all tokens but reduce computation per token. A direct and fair comparison between these two competing strategies for efficiency is essential to contextualize SAISA's contribution. Is it better to prune tokens or to keep them all with cheaper attention? This question is left unanswered.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761552695921}, {"id": "NuMxkZlNzd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16628/Reviewer_QKS9"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes an approach to improve the training and inference efficiency of MLLMs by exploiting redundancy among visual tokens. The key idea is to eliminate the attention computations and FFNs operations related to visual token. The proposed architecture demonstrates significant computational savings with competitive performance to baselines. Experiments across multiple benchmarks, vision encoders, and LLM backbones validate the approach.", "review_text": "The paper proposes an approach to improve the training and inference efficiency of MLLMs by exploiting redundancy among visual tokens. The key idea is to eliminate the attention computations and FFNs operations related to visual token. The proposed architecture demonstrates significant computational savings with competitive performance to baselines. Experiments across multiple benchmarks, vision encoders, and LLM backbones validate the approach.", "strengths": "* The paper tackles an important problem: improving the efficiency of MLLMs, which are known to be computationally demanding.\n* The writing is clear, well structured and the work is well motivated.\n* The approach is validated across different architectures and setups", "weaknesses": "* The redundancy among visual tokens has been explored in prior work. The paper should better position itself in relation to existing studies that analyze or reduce visual redundancy e.g.  [1].\n\n* The experiments primarily use the LLaVA setup, which lags behind more recent MLLM work. It is unclear how the proposed method compares with state-of-the-art (SOTA) models or whether it can be integrated effectively into stronger baselines.\n\n* The paper does not discuss scaling behavior—how the proposed method performs when model or dataset size increases (also related to the previous point).\n\n* Many modern VLM benchmarks (e.g., OCR, document understanding, PDF reasoning) require high-resolution images and fine-grained visual features. It remains unclear whether removing attention among visual tokens will negatively affect performance in these scenarios.\n\n[1] \"Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs.\" NeurIPS2024.", "questions": "Please refer to the weaknesess section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an approach to improve the training and inference efficiency of MLLMs by exploiting redundancy among visual tokens. The key idea is to eliminate the attention computations and FFNs operations related to visual token. The proposed architecture demonstrates significant computational savings with competitive performance to baselines. Experiments across multiple benchmarks, vision encoders, and LLM backbones validate the approach.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "* The paper tackles an important problem: improving the efficiency of MLLMs, which are known to be computationally demanding.\n* The writing is clear, well structured and the work is well motivated.\n* The approach is validated across different architectures and setups", "weaknesses": "* The redundancy among visual tokens has been explored in prior work. The paper should better position itself in relation to existing studies that analyze or reduce visual redundancy e.g.  [1].\n\n* The experiments primarily use the LLaVA setup, which lags behind more recent MLLM work. It is unclear how the proposed method compares with state-of-the-art (SOTA) models or whether it can be integrated effectively into stronger baselines.\n\n* The paper does not discuss scaling behavior—how the proposed method performs when model or dataset size increases (also related to the previous point).\n\n* Many modern VLM benchmarks (e.g., OCR, document understanding, PDF reasoning) require high-resolution images and fine-grained visual features. It remains unclear whether removing attention among visual tokens will negatively affect performance in these scenarios.\n\n[1] \"Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs.\" NeurIPS2024.", "questions": "Please refer to the weaknesess section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761424039419}], "openreview_url": "https://openreview.net/forum?id=TJbfURjxyL", "arxiv_id": "2502.02458", "paper_pdf": "papers/TJbfURjxyL.pdf", "paper_pdf_sha256": "c0f37edf930ae9b6e7e0fc5012251be69fc4e8ebbab95b9ae6fc49f7d4c45ada", "paper_pdf_bytes": 5135429, "paper_pdf_source": "openreview", "code_url": "https://github.com/icip-cas/SAISA", "code_repository": "icip-cas/SAISA", "code_commit": "53a3670ab4c0ca92f57cf9c3bfce66a67b1089ef", "code_archive": "repos/TJbfURjxyL.zip", "code_archive_sha256": "b8e2042eff582bbc9596e4ec5ae522908e83f8219b5c1b2f27eaef3e478208ec", "code_archive_bytes": 3574035, "code_file_count": 467, "code_extensions": {".py": 437, ".sh": 29, ".js": 1}, "github_disk_usage_kb": 3082, "github_languages": {"Python": 3445654, "Shell": 29328, "JavaScript": 9991, "HTML": 7669, "CSS": 1822}, "github_archived": false, "github_pushed_at": "2025-04-08T07:45:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/saisa-towards-multimodal-large-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "06B23UkNid", "year": 2025, "status": "rejected", "title": "MV-CLAM: Multi-View Molecular Interpretation with Cross-Modal Projection via Language Model", "authors": ["Sumin Ha", "Jun Hyeong Kim", "Yinhua Piao", "Sun Kim"], "authorids": ["~Sumin_Ha1", "~Jun_Hyeong_Kim3", "~Yinhua_Piao1", "~Sun_Kim2"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) have shown significant potential in the biomolecular domain, particularly by demonstrating that effective adaptation of molecular representations for LLMs can greatly improve the quality of molecular captions. Most previous works have focused on aligning unimodal molecular structures with text, overlooking the diversity of modalities. Naive approaches to aligning multi-modal molecular structures with text often lead to (1) separately aligned embeddings, (2) inconsistent textual representations, and (3) increased computational overhead. To address these challenges, we propose LLM framework MV-CLAM equipped with MQ-Former, a novel multi-querying transformer. This architecture introduces a cross-model projector facilitating the simultaneous alignment of 2D and 3D molecular representations to a unified text token. By employing a shared self-attention layer, MQ-Former preserves rich molecular embeddings across different dimensions while consolidating them into a universal molecular token. Our approach outperforms baseline models in both molecule-text retrieval and molecule captioning tasks. Additionally, our framework shows promising results for zero-shot molecule editing and molecule-related question answering. By effectively integrating multi-view molecular data into a format conducive to LLMs, our method serves as a valuable tool for enhancing the characterization and understanding of chemical structures, facilitating a more seamless transition from molecular data to textual descriptions. The source code of MV-CLAM is available in https://anonymous.4open.science/r/mv-clam-4827.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "MECE7Z1lpg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13305/Reviewer_6JAy"], "rating": 3, "soundness": 1, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The work proposes a novel multimodal LLM framework MV-CLAM for organic chemistry and MQ-Former — multi-querying transformer model for simultaneous 1D, 2D, and 3D molecular representation learning. Authors show SOTA results in two tasks of molecule-text retrieval and molecule captioning. In addition, authors claim that their approach allows zero-shot molecule editing and molecule-related question answering.", "review_text": "The work proposes a novel multimodal LLM framework MV-CLAM for organic chemistry and MQ-Former — multi-querying transformer model for simultaneous 1D, 2D, and 3D molecular representation learning. Authors show SOTA results in two tasks of molecule-text retrieval and molecule captioning. In addition, authors claim that their approach allows zero-shot molecule editing and molecule-related question answering.", "strengths": "New molecular multimodal LLM framework for simultaneous incorporation of 1d 2D and 3D representations.\nNew Transformer architecture MQ-Former.", "weaknesses": "The claim of the state-of-the-art performance for molecule captioning is not satisfied, see the results in [6].\nThere is no comparison with the other strong retrieval methods for the molecule retrieval task, i.e. RAG.\nThere are various problems with the Zero-shot editing part of the paper. The task is not formally defined. There are no metrics nor baselines for it.\n\nThe QA part is practically absent in the paper, while claimed in the abstract and results parts..\nThere are many works on molecular conformation generation [1-4], it seems that SMILES and/or 2D-graph representation is enough for neural networks to reconstruct RDKIT conformations almost perfectly. It means that 3D input possibly does not add any new information to the model. There is no comparison of the 1D+2D+3D MQ-Former vs 1D+2D models in the paper.\n\nThere is no comparison with other works on multi-modal representation learning for molecules, e.g.: [5]. \n\n[1] Zhu, Jinhua, et al. \"Direct Molecular Conformation Generation.\"\n[2] Xu, Minkai, et al. \"GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation.\" International Conference on Learning Representations.\n[3] Jing, Bowen, et al. \"Torsional diffusion for molecular conformer generation.\" Advances in Neural Information Processing Systems 35 (2022): 24240-24253.\n[4] Lee, Danyeong, et al. \"Disco: Diffusion Schrödinger bridge for molecular conformer optimization.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 12. 2024.\n[5] Manolache, Andrei, Dragos Tantaru, and Mathias Niepert. \"MolMix: A Simple Yet Effective Baseline for Multimodal Molecular Representation Learning.\" arXiv preprint arXiv:2410.07981 (2024).\n[6] Liu, Zhiyuan, et al. \"ReactXT: Understanding Molecular\" Reaction-ship\" via Reaction-Contextualized Molecule-Text Pretraining.\" arXiv preprint arXiv:2405.14225 (2024).", "questions": "1. 3D structures (conformers)\n\nAs mentioned in sec. 5.1 you use MMFF for molecular conformation generation.\n\na. Is it ETKDG geometry generation with further MMFF optimization?\nb. Since it is possible to generate several different conformers for a single molecular structure, did you assess the dependence of the model quality on the conformations? Is it necessary to optimize a generated with ETKDG conformer with MMFF?\n\n2.  It would be reasonable to compare your approach for Zero-shot editing with conditional generation models for small molecules.\n\n3. Please, add experiments on the CHEBI-20 benchmark for the captioning task.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work proposes a novel multimodal LLM framework MV-CLAM for organic chemistry and MQ-Former — multi-querying transformer model for simultaneous 1D, 2D, and 3D molecular representation learning. Authors show SOTA results in two tasks of molecule-text retrieval and molecule captioning. In addition, authors claim that their approach allows zero-shot molecule editing and molecule-related question answering.", "soundness": 1, "presentation": 1, "contribution": 2, "strengths": "New molecular multimodal LLM framework for simultaneous incorporation of 1d 2D and 3D representations.\nNew Transformer architecture MQ-Former.", "weaknesses": "The claim of the state-of-the-art performance for molecule captioning is not satisfied, see the results in [6].\nThere is no comparison with the other strong retrieval methods for the molecule retrieval task, i.e. RAG.\nThere are various problems with the Zero-shot editing part of the paper. The task is not formally defined. There are no metrics nor baselines for it.\n\nThe QA part is practically absent in the paper, while claimed in the abstract and results parts..\nThere are many works on molecular conformation generation [1-4], it seems that SMILES and/or 2D-graph representation is enough for neural networks to reconstruct RDKIT conformations almost perfectly. It means that 3D input possibly does not add any new information to the model. There is no comparison of the 1D+2D+3D MQ-Former vs 1D+2D models in the paper.\n\nThere is no comparison with other works on multi-modal representation learning for molecules, e.g.: [5]. \n\n[1] Zhu, Jinhua, et al. \"Direct Molecular Conformation Generation.\"\n[2] Xu, Minkai, et al. \"GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation.\" International Conference on Learning Representations.\n[3] Jing, Bowen, et al. \"Torsional diffusion for molecular conformer generation.\" Advances in Neural Information Processing Systems 35 (2022): 24240-24253.\n[4] Lee, Danyeong, et al. \"Disco: Diffusion Schrödinger bridge for molecular conformer optimization.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 12. 2024.\n[5] Manolache, Andrei, Dragos Tantaru, and Mathias Niepert. \"MolMix: A Simple Yet Effective Baseline for Multimodal Molecular Representation Learning.\" arXiv preprint arXiv:2410.07981 (2024).\n[6] Liu, Zhiyuan, et al. \"ReactXT: Understanding Molecular\" Reaction-ship\" via Reaction-Contextualized Molecule-Text Pretraining.\" arXiv preprint arXiv:2405.14225 (2024).", "questions": "1. 3D structures (conformers)\n\nAs mentioned in sec. 5.1 you use MMFF for molecular conformation generation.\n\na. Is it ETKDG geometry generation with further MMFF optimization?\nb. Since it is possible to generate several different conformers for a single molecular structure, did you assess the dependence of the model quality on the conformations? Is it necessary to optimize a generated with ETKDG conformer with MMFF?\n\n2.  It would be reasonable to compare your approach for Zero-shot editing with conditional generation models for small molecules.\n\n3. Please, add experiments on the CHEBI-20 benchmark for the captioning task.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730656252504}, {"id": "8bfabZVPDz", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13305/Reviewer_vfcG"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces a framework that leverages large language models (LLMs) to interpret and generate molecular captions. The work incorporates both 2D and 3D molecular structures to provide a more comprehensive understanding of molecules.", "review_text": "The paper introduces a framework that leverages large language models (LLMs) to interpret and generate molecular captions. The work incorporates both 2D and 3D molecular structures to provide a more comprehensive understanding of molecules.", "strengths": "1. The paper integrates both 2D and 3D molecular structures to enhance the model's understanding of molecular data.\n2. The paper includes detailed figures (Figure 1-3) that clearly explain the method's framework and training scheme. \n3. And the analysis of attention maps in Appendix A.4 provides valuable insights into the model's behavior.", "weaknesses": "1. Compared to recent related work, such as 3D-MoLM (Li et al., 2024), the innovation in MV-CLAM appears incremental. While the paper claims to incorporate both 2D and 3D molecular structures for a more comprehensive understanding, the approach seems to merely extend the 3D-MoLM framework by introducing 2D components through MAT. The proposed MQ-former architecture does not demonstrate significant structural innovations beyond existing methods. A clearer articulation of the novel contributions and architectural advantages over 3D-MoLM would be necessary to establish the work's originality.\n2. The paper considers SMILES as an important molecular modality and notes that \"1D SMILES provide compact represen tation of molecular structures\", but does not mention SELFIES (Krenn et al., 2020) at all, which has been widely adopted in recent works due to its robust characteristics and tokenization-friendly nature. SELFIES offers inherent robustness and easier tokenization that aligns well with LLMs, making it a potentially more suitable choice for this application. \n3. Some images (e.g. the big image at page 18) are not vector graphics and lack titles or captions, which makes it confusing.", "questions": "See 'Weaknesses' section.\n1. Could the authors provide a more detailed explanation of the novelty of MV-CLAM compared to recent related work?\n2. Why was SELFIES not considered as a molecular modality in this work, given its advantages over SMILES in tokenization and alignment with LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a framework that leverages large language models (LLMs) to interpret and generate molecular captions. The work incorporates both 2D and 3D molecular structures to provide a more comprehensive understanding of molecules.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper integrates both 2D and 3D molecular structures to enhance the model's understanding of molecular data.\n2. The paper includes detailed figures (Figure 1-3) that clearly explain the method's framework and training scheme. \n3. And the analysis of attention maps in Appendix A.4 provides valuable insights into the model's behavior.", "weaknesses": "1. Compared to recent related work, such as 3D-MoLM (Li et al., 2024), the innovation in MV-CLAM appears incremental. While the paper claims to incorporate both 2D and 3D molecular structures for a more comprehensive understanding, the approach seems to merely extend the 3D-MoLM framework by introducing 2D components through MAT. The proposed MQ-former architecture does not demonstrate significant structural innovations beyond existing methods. A clearer articulation of the novel contributions and architectural advantages over 3D-MoLM would be necessary to establish the work's originality.\n2. The paper considers SMILES as an important molecular modality and notes that \"1D SMILES provide compact represen tation of molecular structures\", but does not mention SELFIES (Krenn et al., 2020) at all, which has been widely adopted in recent works due to its robust characteristics and tokenization-friendly nature. SELFIES offers inherent robustness and easier tokenization that aligns well with LLMs, making it a potentially more suitable choice for this application. \n3. Some images (e.g. the big image at page 18) are not vector graphics and lack titles or captions, which makes it confusing.", "questions": "See 'Weaknesses' section.\n1. Could the authors provide a more detailed explanation of the novelty of MV-CLAM compared to recent related work?\n2. Why was SELFIES not considered as a molecular modality in this work, given its advantages over SMILES in tokenization and alignment with LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730557631368}, {"id": "hefprGvv7n", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13305/Reviewer_48Rx"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces MV-CLAM, a framework utilizing a novel multi-querying transformer (MQ-Former) to enhance the alignment of multi-modal molecular representations with text. By employing a shared self-attention layer, this approach effectively consolidates 2D and 3D molecular data into query tokens, improving performance in molecule-text retrieval and captioning tasks. Additionally, it demonstrates potential for zero-shot molecule editing and molecule-related question answering, thereby facilitating better characterization of chemical structures.", "review_text": "The paper introduces MV-CLAM, a framework utilizing a novel multi-querying transformer (MQ-Former) to enhance the alignment of multi-modal molecular representations with text. By employing a shared self-attention layer, this approach effectively consolidates 2D and 3D molecular data into query tokens, improving performance in molecule-text retrieval and captioning tasks. Additionally, it demonstrates potential for zero-shot molecule editing and molecule-related question answering, thereby facilitating better characterization of chemical structures.", "strengths": "* The description of the proposed methodology is easy to follow. The paper is well written in general.\n* The paper introduces a promising multi-view for approach for the infusion of specialized chemical knowledge into general-purpose pre-trained LLMs.\n* The proposed MV-CLAM achieves state-of-the-art on PubChem324K for molecule captioning and retrieval tasks.", "weaknesses": "* The experimental evaluation of the proposed method is conducted on a single dataset for both task: molecule captioning and molecule-text retrieval.\n* The list of baseline models on molecule captioning only includes a single T5 language model while there are more recent works, including: nach0 and Text+ChemT5. \n* Some implementation decisions are not justified well enough. This includes: (i) the choice of SciBERT as a language encoder for MQ-Former; (ii) the choice of 2D and 3D encoders; (iii) introduction of $K$ query tokens instead of a single query token for each view; (iv) the choice of LLaMA2 as an LLM. It is unclear how the experimental results would change if each of the mentioned models is replaced with another one.\n* Incomplete ablation study. The necessity of (i) Molecule-text Contrasting and (ii) Molecule-text Matching losses is not proven experimentally. For (i), it is unclear whether two loss components required or the model will perform well with a single one. For (ii), the impact of negative sample is under-explored. \n* The effect of most hyper-parameters in the method's module on the resulting performance is understudied. For instance, query token count, negative sample count in MTM loss.\n* The methodology for molecule-text retrieval is unclear from the paper.\n* The applicability of the proposed methodology to broader list of datasets is questionable: it requires 2D/3D molecular data in addition to simple SMILES string representations.", "questions": "* Add experimental comparison against more chemical language models on molecule captioning, e.g., nach0 [1], Text+Chem T5 [2], SciFive [3], PRESTO [4], GitMol [5].\n* For retrieval task (Table 1), is it possible to add chemical BERT-based encoders in addition to textual encoder SciBERT? (e.g., ChemBERTa)\n* Conduct additional experiments on other molecule captioning datasets such as Mol-Instructions [6] and CheBI20 [7].\n* For molecule-text retrieval, do you adopt a generative approach (e.g., GENRE [8]) or the task is formulated as a cross-modal embedding-based search by similarity (e.g., as in [9])?\n* In Figure 3, where does the textual description come from during prediction on a test set? As far as I understand the molecule captioning task, you are only given a SMILES string.\n* What is the LLaMA version you use? Add adopted HuggingFace checkpoints. \n* Even if you adopt a LLaMA with 7B parameters, MolT5 has less than 1B. Could not we just scale MolT5 to 3-5B parameters and obtain a better molecule captioning quality?\n* Why is MolT5 absent from the Table 1?\n* Add ablation study for SciBERT, 2D/3D molecule encoders, LLaMA2.\n* Add ablation study for training losses. For Molecule-text Contrasting loss, prove it requires two components. For Molecule-text Matching loss, explore the effect of negative samples.\n* Is it possible to generalize the methodology to unseen datasets and unseen SMILES? Given a SMILES, can I always obtain its 2D/3D representation and apply a pre-trained MV-CLAM model?\n\n\n\n\nTypos:\n* Line 102: transformer -> Transformer, Add reference.\n* Line 194: **$A$** under-specified.\n* Line 234: Missing citation for LoRA.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces MV-CLAM, a framework utilizing a novel multi-querying transformer (MQ-Former) to enhance the alignment of multi-modal molecular representations with text. By employing a shared self-attention layer, this approach effectively consolidates 2D and 3D molecular data into query tokens, improving performance in molecule-text retrieval and captioning tasks. Additionally, it demonstrates potential for zero-shot molecule editing and molecule-related question answering, thereby facilitating better characterization of chemical structures.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "* The description of the proposed methodology is easy to follow. The paper is well written in general.\n* The paper introduces a promising multi-view for approach for the infusion of specialized chemical knowledge into general-purpose pre-trained LLMs.\n* The proposed MV-CLAM achieves state-of-the-art on PubChem324K for molecule captioning and retrieval tasks.", "weaknesses": "* The experimental evaluation of the proposed method is conducted on a single dataset for both task: molecule captioning and molecule-text retrieval.\n* The list of baseline models on molecule captioning only includes a single T5 language model while there are more recent works, including: nach0 and Text+ChemT5. \n* Some implementation decisions are not justified well enough. This includes: (i) the choice of SciBERT as a language encoder for MQ-Former; (ii) the choice of 2D and 3D encoders; (iii) introduction of $K$ query tokens instead of a single query token for each view; (iv) the choice of LLaMA2 as an LLM. It is unclear how the experimental results would change if each of the mentioned models is replaced with another one.\n* Incomplete ablation study. The necessity of (i) Molecule-text Contrasting and (ii) Molecule-text Matching losses is not proven experimentally. For (i), it is unclear whether two loss components required or the model will perform well with a single one. For (ii), the impact of negative sample is under-explored. \n* The effect of most hyper-parameters in the method's module on the resulting performance is understudied. For instance, query token count, negative sample count in MTM loss.\n* The methodology for molecule-text retrieval is unclear from the paper.\n* The applicability of the proposed methodology to broader list of datasets is questionable: it requires 2D/3D molecular data in addition to simple SMILES string representations.", "questions": "* Add experimental comparison against more chemical language models on molecule captioning, e.g., nach0 [1], Text+Chem T5 [2], SciFive [3], PRESTO [4], GitMol [5].\n* For retrieval task (Table 1), is it possible to add chemical BERT-based encoders in addition to textual encoder SciBERT? (e.g., ChemBERTa)\n* Conduct additional experiments on other molecule captioning datasets such as Mol-Instructions [6] and CheBI20 [7].\n* For molecule-text retrieval, do you adopt a generative approach (e.g., GENRE [8]) or the task is formulated as a cross-modal embedding-based search by similarity (e.g., as in [9])?\n* In Figure 3, where does the textual description come from during prediction on a test set? As far as I understand the molecule captioning task, you are only given a SMILES string.\n* What is the LLaMA version you use? Add adopted HuggingFace checkpoints. \n* Even if you adopt a LLaMA with 7B parameters, MolT5 has less than 1B. Could not we just scale MolT5 to 3-5B parameters and obtain a better molecule captioning quality?\n* Why is MolT5 absent from the Table 1?\n* Add ablation study for SciBERT, 2D/3D molecule encoders, LLaMA2.\n* Add ablation study for training losses. For Molecule-text Contrasting loss, prove it requires two components. For Molecule-text Matching loss, explore the effect of negative samples.\n* Is it possible to generalize the methodology to unseen datasets and unseen SMILES? Given a SMILES, can I always obtain its 2D/3D representation and apply a pre-trained MV-CLAM model?\n\n\n\n\nTypos:\n* Line 102: transformer -> Transformer, Add reference.\n* Line 194: **$A$** under-specified.\n* Line 234: Missing citation for LoRA.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730284339594}, {"id": "SPDtsIgtdN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13305/Reviewer_R1Xc"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes MQ-Former, an extension of the Q-Former framework, which incorporates a multi-query mechanism for aligning both 2D and 3D molecular data with textual information for enhanced molecule-text retrieval and molecule captioning.", "review_text": "The paper proposes MQ-Former, an extension of the Q-Former framework, which incorporates a multi-query mechanism for aligning both 2D and 3D molecular data with textual information for enhanced molecule-text retrieval and molecule captioning.", "strengths": "- The paper aims to enhance cross-modal alignment by integrating 2D and 3D molecular views.\n- The model demonstrates improvements in molecule-text retrieval and captioning performance over baseline models.\n- The paper includes case studies and examples of zero-shot molecule editing.", "weaknesses": "- The model lacks significant innovation, as MQ-Former primarily adds an extra branch to the existing Q-Former with only minor variations in training objectives.\n- Experiments are restricted to molecule-text retrieval and captioning on PubChem. The paper lacks essential molecular tasks like molecule generation and datasets like ChEBI-20.\n- The motivation for adding a branch to Q-Former, rather than simply using a 3D molecular encoder like prior works (e.g., 3D-MoLM), is unclear. \n- The paper’s presentation could be improved. Plots lack careful formatting, with text that is difficult to read due to small font sizes.", "questions": "- How does MQ-Former handle scenarios where 2D and 3D molecular information may not equally contribute to textual descriptions?\n- Could the authors include more molecular tasks, such as molecule generation or property prediction, to provide a more comprehensive evaluation of MQ-Former?\n- What impact does the weighting of the multi-objective training loss have on the model’s performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes MQ-Former, an extension of the Q-Former framework, which incorporates a multi-query mechanism for aligning both 2D and 3D molecular data with textual information for enhanced molecule-text retrieval and molecule captioning.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper aims to enhance cross-modal alignment by integrating 2D and 3D molecular views.\n- The model demonstrates improvements in molecule-text retrieval and captioning performance over baseline models.\n- The paper includes case studies and examples of zero-shot molecule editing.", "weaknesses": "- The model lacks significant innovation, as MQ-Former primarily adds an extra branch to the existing Q-Former with only minor variations in training objectives.\n- Experiments are restricted to molecule-text retrieval and captioning on PubChem. The paper lacks essential molecular tasks like molecule generation and datasets like ChEBI-20.\n- The motivation for adding a branch to Q-Former, rather than simply using a 3D molecular encoder like prior works (e.g., 3D-MoLM), is unclear. \n- The paper’s presentation could be improved. Plots lack careful formatting, with text that is difficult to read due to small font sizes.", "questions": "- How does MQ-Former handle scenarios where 2D and 3D molecular information may not equally contribute to textual descriptions?\n- Could the authors include more molecular tasks, such as molecule generation or property prediction, to provide a more comprehensive evaluation of MQ-Former?\n- What impact does the weighting of the multi-objective training loss have on the model’s performance?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730282154723}], "openreview_url": "https://openreview.net/forum?id=06B23UkNid", "arxiv_id": "2503.04780", "paper_pdf": "papers/06B23UkNid.pdf", "paper_pdf_sha256": "8229880c9dff7240555a729dce46d075d1c776b342664db8941544a7b6396529", "paper_pdf_bytes": 1545670, "paper_pdf_source": "openreview", "code_url": "https://github.com/sumin124/mv-clam", "code_repository": "sumin124/mv-clam", "code_commit": "d537bf334cdcf438bbd0d343ca0eea6bfdb96837", "code_archive": "repos/06B23UkNid.zip", "code_archive_sha256": "a75bc12fc904c48c80a5a9d73434b290fdfa9019180ac2a5f0c84f4e4a986524", "code_archive_bytes": 103490, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 224, "github_languages": {"Python": 333095}, "github_archived": false, "github_pushed_at": "2025-08-21T05:55:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mv-clam-multi-view-molecular-interpretation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "468KWV14ll", "year": 2024, "status": "rejected", "title": "Exploration and Anti-Exploration with Distributional Random Network Distillation", "authors": ["Kai Yang", "Jian Tao", "Jiafei Lyu", "Xiu Li"], "authorids": ["~Kai_Yang6", "~Jian_Tao5", "~Jiafei_Lyu1", "~Xiu_Li1"], "authors_source": "OpenReview API", "abstract": "Exploration remains a critical issue in deep reinforcement learning for an agent to attain high returns in unknown environments. Although the prevailing exploration Random Network Distillation (RND) algorithm has been demonstrated to be effective in numerous environments, it often needs more discriminative power in bonus allocation. This paper highlights the ''bonus inconsistency'' issue within RND, pinpointing its primary limitation. To address this issue, we introduce the Distributional RND (DRND), a derivative of the RND. DRND enhances the exploration process by distilling a distribution of random networks and implicitly incorporates pseudo counts to improve the precision of bonus allocation. This refinement encourages agents to engage in more extensive exploration. Our method effectively mitigates the inconsistency issue without introducing significant computational overhead. Both theoretical analysis and experimental results demonstrate the superiority of our approach over the original RND algorithm. Our method excels in challenging online exploration scenarios and effectively serves as an anti-exploration mechanism in D4RL offline tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "2bpP2YPiLR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2490/Reviewer_gfXh"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper identifies that intrinsic reward bonuses from Random Network Distillation (RND) exhibit inconsistency with respect to the experienced data distribution. Initial bonuses from RND can be non-uniform, and final bonuses may not correspond to the distribution of state visitations.\n\nTo overcome this limitation, this paper proposes DRND, which samples one of N random target networks to generate targets for the predictor. The newly proposed intrinsic reward combines two bonuses, one which generalizes RND and the other one generalizes Coin Flip Networks (CFN).\n\nTheoretical analysis in the linear setting supports the idea, and experiments in online and offline settings show that the proposed implementation can provide benefits over RND and other bonus-based approaches.", "review_text": "This paper identifies that intrinsic reward bonuses from Random Network Distillation (RND) exhibit inconsistency with respect to the experienced data distribution. Initial bonuses from RND can be non-uniform, and final bonuses may not correspond to the distribution of state visitations.\n\nTo overcome this limitation, this paper proposes DRND, which samples one of N random target networks to generate targets for the predictor. The newly proposed intrinsic reward combines two bonuses, one which generalizes RND and the other one generalizes Coin Flip Networks (CFN).\n\nTheoretical analysis in the linear setting supports the idea, and experiments in online and offline settings show that the proposed implementation can provide benefits over RND and other bonus-based approaches.", "strengths": "**S1.** RND is a popular approach to exploration with intrinsic rewards, and the identified limitation of RND and the proposed DRND approach would interest the research community.\n\n**S2.** The proposed approach is interesting and novel (to the best of my knowledge).\n\n**S3.** DRND is computationally cheaper than ensemble-based approaches, which could be used to resolve the bonus inconsistency. An advantage of using multiple target networks with only a single predictor is that it does not require maintaining an ensemble of predictors, which would require separate backward passes.\n\n**S4.** Experiments have been conducted in a diverse range of environments.", "weaknesses": "**W1.** My main concern is with the empirical evaluation. Using state-actions for DRND and only states for RND is not an exact comparison; there should be analysis/ablations where DRND uses just states, and RND uses state-actions.\n\nThis might be particularly important as the performance difference between RND and DRND is small in many environments.\n\n**W2.**  The proposed approach computes intrinsic rewards by combining two bonuses. However, I feel that introducing the second bonus–similar to the bonus from Coin Flip Network (CFN)--is not sufficiently well-motivated. Could the authors explain the need for the second bonus? Why is $b_1$ alone insufficient? \n\n**W3.** The paper would benefit from comparisons with ensemble-based RND in the online setting, for example, along the lines of [1] or [2]. Even though ensemble-based approaches may be computationally more costly, having these results would be helpful data points.\n\n**W4.** The presentation of the paper can be improved. Some suggestions are listed below.\n\n- The intrinsic reward paragraph in the preliminaries is unclear. What is $y_t$ the target for? \n- The RND method is defined with observations, but the preliminaries define an MDP with states.\n- Some sentences are hard to understand, e.g., “and its interpretability needs to be more apparent than count-based techniques” and “These approaches are implemented through formulas such as $r_t =$ ..”\n- What does the shading in figures represent?\n- It would also be helpful to show DRND’s performance in Figure 1 rather than having Figures 1 and 3 separate.\n- The term “deep exploration” should be explained in the paper or supported by citations.\n\nOverall, I appreciate the direction the authors took toward improving RND-like intrinsic bonuses. I remain open to increasing the score should the weaknesses and questions be adequately addressed/clarified.\n\n—------------------—------------------—------------------—------------------—------------------\n\n### References\n\n[1] Ciosek, K., Fortuin, V., Tomioka, R., Hofmann, K., & Turner, R. (2019). Conservative uncertainty estimation by fitting prior networks. In International Conference on Learning Representations.\n\n[2] Ramesh, A., Kirsch, L., van Steenkiste, S., & Schmidhuber, J. (2022). Exploring through random curiosity with general value functions. Advances in Neural Information Processing Systems", "questions": "Q1. What were the architectures of random and target networks used in the experiments for Figures 1 and 3?\n\nQ2. Is there evidence for bonus inconsistency in the tasks other than the one used for Figure 1? And if so, would DRND resolve it?\n\nQ3. Coming back to the experiment in Figure 1, it would also be useful to see snapshots of the bonus inconsistency at more points through learning, and not just the beginning and the end. Are these results available?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper identifies that intrinsic reward bonuses from Random Network Distillation (RND) exhibit inconsistency with respect to the experienced data distribution. Initial bonuses from RND can be non-uniform, and final bonuses may not correspond to the distribution of state visitations.\n\nTo overcome this limitation, this paper proposes DRND, which samples one of N random target networks to generate targets for the predictor. The newly proposed intrinsic reward combines two bonuses, one which generalizes RND and the other one generalizes Coin Flip Networks (CFN).\n\nTheoretical analysis in the linear setting supports the idea, and experiments in online and offline settings show that the proposed implementation can provide benefits over RND and other bonus-based approaches.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "**S1.** RND is a popular approach to exploration with intrinsic rewards, and the identified limitation of RND and the proposed DRND approach would interest the research community.\n\n**S2.** The proposed approach is interesting and novel (to the best of my knowledge).\n\n**S3.** DRND is computationally cheaper than ensemble-based approaches, which could be used to resolve the bonus inconsistency. An advantage of using multiple target networks with only a single predictor is that it does not require maintaining an ensemble of predictors, which would require separate backward passes.\n\n**S4.** Experiments have been conducted in a diverse range of environments.", "weaknesses": "**W1.** My main concern is with the empirical evaluation. Using state-actions for DRND and only states for RND is not an exact comparison; there should be analysis/ablations where DRND uses just states, and RND uses state-actions.\n\nThis might be particularly important as the performance difference between RND and DRND is small in many environments.\n\n**W2.**  The proposed approach computes intrinsic rewards by combining two bonuses. However, I feel that introducing the second bonus–similar to the bonus from Coin Flip Network (CFN)--is not sufficiently well-motivated. Could the authors explain the need for the second bonus? Why is $b_1$ alone insufficient? \n\n**W3.** The paper would benefit from comparisons with ensemble-based RND in the online setting, for example, along the lines of [1] or [2]. Even though ensemble-based approaches may be computationally more costly, having these results would be helpful data points.\n\n**W4.** The presentation of the paper can be improved. Some suggestions are listed below.\n\n- The intrinsic reward paragraph in the preliminaries is unclear. What is $y_t$ the target for? \n- The RND method is defined with observations, but the preliminaries define an MDP with states.\n- Some sentences are hard to understand, e.g., “and its interpretability needs to be more apparent than count-based techniques” and “These approaches are implemented through formulas such as $r_t =$ ..”\n- What does the shading in figures represent?\n- It would also be helpful to show DRND’s performance in Figure 1 rather than having Figures 1 and 3 separate.\n- The term “deep exploration” should be explained in the paper or supported by citations.\n\nOverall, I appreciate the direction the authors took toward improving RND-like intrinsic bonuses. I remain open to increasing the score should the weaknesses and questions be adequately addressed/clarified.\n\n—------------------—------------------—------------------—------------------—------------------\n\n### References\n\n[1] Ciosek, K., Fortuin, V., Tomioka, R., Hofmann, K., & Turner, R. (2019). Conservative uncertainty estimation by fitting prior networks. In International Conference on Learning Representations.\n\n[2] Ramesh, A., Kirsch, L., van Steenkiste, S., & Schmidhuber, J. (2022). Exploring through random curiosity with general value functions. Advances in Neural Information Processing Systems", "questions": "Q1. What were the architectures of random and target networks used in the experiments for Figures 1 and 3?\n\nQ2. Is there evidence for bonus inconsistency in the tasks other than the one used for Figure 1? And if so, would DRND resolve it?\n\nQ3. Coming back to the experiment in Figure 1, it would also be useful to see snapshots of the bonus inconsistency at more points through learning, and not just the beginning and the end. Are these results available?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699056104420}, {"id": "uu2GkaIviz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2490/Reviewer_u9ew"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Authors propose an improvement for the Random Network Distillation (RND) by employing N target networks instead of one. Additionally bonus calculation is splitted into two components. Authors claim that their approach DRND outperforms RND and other baselines in offline and online RL settings.", "review_text": "Authors propose an improvement for the Random Network Distillation (RND) by employing N target networks instead of one. Additionally bonus calculation is splitted into two components. Authors claim that their approach DRND outperforms RND and other baselines in offline and online RL settings.", "strengths": "Authors show the problems that occur with RND approach and propose reasonable idea with mathematical proofs to overcome them. Empirical results could be be another strong side but, unfortunately, I have many concerns about them, see **Weaknesses**.", "weaknesses": "**Offline RL results**\n\nAuthors report baseline offline RL scores from ReBRAC paper but half of the scores do not match. For example, in case of ReBRAC MuJoCo scores are reported correctly for each dataset but average is changed to the lower value from 81.2 to 78.0. ReBRAC's scores on AntMaze match original paper except antmaze-large-diverse where they are swapped between ReBRAC and SAC-RND which leads to ReBRAC's average performance 75.4 instead of 76.8 which is better than DRND. \n\nSAC-RND performs poorly on Adroit D4RL domain and DRND is not tested on it. DRND performs better only on MuJoCo tasks. I could admit experimental results to be strong if DRND's performance on Adroit is competitive to ReBRAC or IQL.\n\n**Online RL results**\n\nI haven't seen references to online RL baselines results. Problems with offline RL results make me question reported plots reliability. Please, correct me if I've missed it.\n\n**Hyperparameters choice**\n\nHyperparameter choice procedure is also not clear from the text, please indicate if I've missed it. Hyperparameters candidates for $\\lambda$ parameters are not presented and values are different for every offline dataset. Both, SAC-RND and ReBRAC, also tune this parameters for every dataset but they also report sensitivity to hyperparameters. \nReBRAC's authors tuned parameters on a few seeds and then evaluated all of the algorithms on another set of hyperparameters which lead to the SAC-RND performance drop comparing to the original work (SAC-RND reported results without evaluating on different set of seeds). Could you please say if the same procedure followed your work?", "questions": "Most of the questions are based on the **Weaknesses**. I understand that all of the experiments is hard to run during the short rebuttal phase but I kindly ask you to run at least some of them.\n\n* Please fix scores for offline RL baselines or explain the change in scores.\n\n* Could you provide reference scores for online baselines?\n\n* What is DRND performance on Adroit tasks?\n\n* What is hyperparameters choice procedure?\n\n* What is hyperparameters sensitivity? It can be easily done with EOP (https://arxiv.org/abs/2110.04156) if there are logs from hyperparameters search. ReBRAC provide EOP scores for ReBRAC and IQL which can be reused for comparison. \n\n* How DRND would perform in offline-to-online setting, e.g. on AntMaze? It is an open question for SAC-RND. Since DRND is reported to have good online performance it can be potentially used in offline-to-online setup. You can compare with the results from ReBRAC's latest revision.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors propose an improvement for the Random Network Distillation (RND) by employing N target networks instead of one. Additionally bonus calculation is splitted into two components. Authors claim that their approach DRND outperforms RND and other baselines in offline and online RL settings.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "Authors show the problems that occur with RND approach and propose reasonable idea with mathematical proofs to overcome them. Empirical results could be be another strong side but, unfortunately, I have many concerns about them, see **Weaknesses**.", "weaknesses": "**Offline RL results**\n\nAuthors report baseline offline RL scores from ReBRAC paper but half of the scores do not match. For example, in case of ReBRAC MuJoCo scores are reported correctly for each dataset but average is changed to the lower value from 81.2 to 78.0. ReBRAC's scores on AntMaze match original paper except antmaze-large-diverse where they are swapped between ReBRAC and SAC-RND which leads to ReBRAC's average performance 75.4 instead of 76.8 which is better than DRND. \n\nSAC-RND performs poorly on Adroit D4RL domain and DRND is not tested on it. DRND performs better only on MuJoCo tasks. I could admit experimental results to be strong if DRND's performance on Adroit is competitive to ReBRAC or IQL.\n\n**Online RL results**\n\nI haven't seen references to online RL baselines results. Problems with offline RL results make me question reported plots reliability. Please, correct me if I've missed it.\n\n**Hyperparameters choice**\n\nHyperparameter choice procedure is also not clear from the text, please indicate if I've missed it. Hyperparameters candidates for $\\lambda$ parameters are not presented and values are different for every offline dataset. Both, SAC-RND and ReBRAC, also tune this parameters for every dataset but they also report sensitivity to hyperparameters. \nReBRAC's authors tuned parameters on a few seeds and then evaluated all of the algorithms on another set of hyperparameters which lead to the SAC-RND performance drop comparing to the original work (SAC-RND reported results without evaluating on different set of seeds). Could you please say if the same procedure followed your work?", "questions": "Most of the questions are based on the **Weaknesses**. I understand that all of the experiments is hard to run during the short rebuttal phase but I kindly ask you to run at least some of them.\n\n* Please fix scores for offline RL baselines or explain the change in scores.\n\n* Could you provide reference scores for online baselines?\n\n* What is DRND performance on Adroit tasks?\n\n* What is hyperparameters choice procedure?\n\n* What is hyperparameters sensitivity? It can be easily done with EOP (https://arxiv.org/abs/2110.04156) if there are logs from hyperparameters search. ReBRAC provide EOP scores for ReBRAC and IQL which can be reused for comparison. \n\n* How DRND would perform in offline-to-online setting, e.g. on AntMaze? It is an open question for SAC-RND. Since DRND is reported to have good online performance it can be potentially used in offline-to-online setup. You can compare with the results from ReBRAC's latest revision.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698147127014}, {"id": "jPbHPto0aY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2490/Reviewer_4Wh1"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work extends the Random Network Distillation (RND) algorithm and thus proposes Distributional RND (DRND). The DRND enhances the exploration process by distilling a distribution of random networks and implicitly incorporates pseudo counts to improve the precision of bonus allocation. This refinement encourages agents to engage in more extensive exploration. The proposed method excels in challenging online exploration scenarios and effectively serves as an anti-exploration mechanism in D4RL offline tasks.", "review_text": "This work extends the Random Network Distillation (RND) algorithm and thus proposes Distributional RND (DRND). The DRND enhances the exploration process by distilling a distribution of random networks and implicitly incorporates pseudo counts to improve the precision of bonus allocation. This refinement encourages agents to engage in more extensive exploration. The proposed method excels in challenging online exploration scenarios and effectively serves as an anti-exploration mechanism in D4RL offline tasks.", "strengths": "1.\tThis paper is well-written and has good readability.\n2.\tThe authors improve the RND in a simple but effective way.\n3.\tThe proposed method get the high performance on offline tasks.", "weaknesses": "1.\tI would like to see the learning curves of SAC-RND and SAC-DRND on the hopper, halfcheetah, and walker2d.\n2.\t$\\lambda$ is carefully tuned for different datasets, therefore, the sensitivity analysis on this parameter should be given.\n3. A similar work should be cited (Controlling Overestimation Bias with Truncated Mixture of Continuous\nDistributional Quantile Critics).", "questions": "1.\tHow to test SAC on offline tasks?  Why the results of SAC in Table 1 are so low?\n2.\tIn Fig.2 (b), there is no backpropagation from intrinsic reward. In this case, how to optimize the policy?\n3.\tWhat is the desired distribution in Fig.1?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work extends the Random Network Distillation (RND) algorithm and thus proposes Distributional RND (DRND). The DRND enhances the exploration process by distilling a distribution of random networks and implicitly incorporates pseudo counts to improve the precision of bonus allocation. This refinement encourages agents to engage in more extensive exploration. The proposed method excels in challenging online exploration scenarios and effectively serves as an anti-exploration mechanism in D4RL offline tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1.\tThis paper is well-written and has good readability.\n2.\tThe authors improve the RND in a simple but effective way.\n3.\tThe proposed method get the high performance on offline tasks.", "weaknesses": "1.\tI would like to see the learning curves of SAC-RND and SAC-DRND on the hopper, halfcheetah, and walker2d.\n2.\t$\\lambda$ is carefully tuned for different datasets, therefore, the sensitivity analysis on this parameter should be given.\n3. A similar work should be cited (Controlling Overestimation Bias with Truncated Mixture of Continuous\nDistributional Quantile Critics).", "questions": "1.\tHow to test SAC on offline tasks?  Why the results of SAC in Table 1 are so low?\n2.\tIn Fig.2 (b), there is no backpropagation from intrinsic reward. In this case, how to optimize the policy?\n3.\tWhat is the desired distribution in Fig.1?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697503569419}], "openreview_url": "https://openreview.net/forum?id=468KWV14ll", "arxiv_id": "2401.09750", "paper_pdf": "papers/468KWV14ll.pdf", "paper_pdf_sha256": "4b3a6122f775cd7284d4aee2097dcf1ac6abac83995f6d011f4e684b29d3582b", "paper_pdf_bytes": 13044600, "paper_pdf_source": "openreview", "code_url": "https://github.com/yk7333/DRND", "code_repository": "yk7333/DRND", "code_commit": "e1ea983c9c12603804e6d268efb212c208663aac", "code_archive": "repos/468KWV14ll.zip", "code_archive_sha256": "c8de4118f7eb4b3aaccf02f71c77da4c94ce3f10ef3e4dee119a7ab5d83276ca", "code_archive_bytes": 135866, "code_file_count": 13, "code_extensions": {".py": 12, ".sh": 1}, "github_disk_usage_kb": 150, "github_languages": {"Python": 87011, "Shell": 923}, "github_archived": false, "github_pushed_at": "2024-10-12T06:27:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/exploration-and-anti-exploration-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "C9uEwyfklBE", "year": 2023, "status": "rejected", "title": "Pareto Manifold Learning: Tackling multiple tasks via ensembles of single-task models", "authors": ["Nikolaos Dimitriadis", "Pascal Frossard", "François Fleuret"], "authorids": ["~Nikolaos_Dimitriadis1", "~Pascal_Frossard1", "~François_Fleuret2"], "authors_source": "OpenReview API", "abstract": "In Multi-Task Learning, tasks may compete and limit the performance achieved on each other rather than guiding the optimization trajectory to a common solution, superior to its single-task counterparts. There is often not a single solution that is optimal for all tasks, leading practitioners to balance tradeoffs between tasks' performance, and to resort to optimality in the Pareto sense. Current Multi-Task Learning methodologies either completely neglect this aspect of functional diversity, and produce one solution in the Pareto Front predefined by their optimization schemes, or produce diverse but discrete solutions, each requiring a separate training run. In this paper, we conjecture that there exist Pareto Subspaces, i.e., weight subspaces where multiple optimal functional solutions lie. We propose Pareto Manifold Learning, an ensembling method in weight space that is able to discover such a parameterization and produces a continuous Pareto Front in a single training run, allowing practitioners to modulate the performance on each task during inference on the fly. We validate the proposed method on a diverse set of multi-task learning benchmarks, ranging from image classification to tabular datasets and scene understanding, and show that Pareto Manifold Learning outperforms state-of-the-art algorithms.\n", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "En9sQnN7PEc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4551/Reviewer_SqFR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper assumes there exist Pareto Subspaces, i.e., weight subspaces where multiple optimal functional solutions lie, and develops a weight-ensembling method named Pareto Manifold Learning that casts multi-task problems as learning an ensemble of single-task predictors by interpolating among members during training.  Multiple single-task predictors are trained in conjunction to produce a subspace formed by their convex hull, and endowed with desirable Pareto properties. Each single-task model infuses and benefits from representational knowledge and the other members. The losses are weighted in tandem with the interpolation when training.  The ensemble as a whole engenders a weight subspace that explicitly encodes tradeoffs and results in a continuous parameterization of the Pareto Front.", "review_text": "This paper produces a subspace with multiple Pareto stationary points in the multi-task learning based on the geometry of the loss landscape in single task machine learning where the local optimal are connected by simple paths. However, the representation and more illustrations are needed.", "strengths": "Strength:\nThis paper castes multi-task problems as learning an ensemble of single-task predictors by interpolating among members during training, and designs weight subspaces where multiple optimal functional solutions lie. Based on this weight space, an ensembling method that can produce a continuous Pareto Front in a single training run is proposed. This allows practitioners to modulate the performance on each task during inference on the fly.  \n\nWeaknesses:\n\n1.Similar methods have already been proposed for multi-task learning and has not been disccussed in this paper [1].\n\n1.When sampling on the convex hull parameterization, authors choose to adopt the Dirichlet distribution since its support is the T-dimensional simplex. Does this distribution have other properties.   Why using this distribution?  If p≫1，how the ensemble will change. \n\n2.When training, a mono tonic relationship is imposed between the degree of a single-task predictor participation and the weight of the corresponding task loss. As a result, the ensemble engenders a subspace that explicitly encodes tradeoffs and results in a continuous parameterization of the Pareto Front.  Whether the mono tonic relationship can be replaced by other relationships? Explaining this point may be better.\n\n[1]Navon A, Shamsian A, Fetaya E, et al. Learning the Pareto Front with Hypernetworks[C]//International Conference on Learning Representations. 2020.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper assumes there exist Pareto Subspaces, i.e., weight subspaces where multiple optimal functional solutions lie, and develops a weight-ensembling method named Pareto Manifold Learning that casts multi-task problems as learning an ensemble of single-task predictors by interpolating among members during training.  Multiple single-task predictors are trained in conjunction to produce a subspace formed by their convex hull, and endowed with desirable Pareto properties. Each single-task model infuses and benefits from representational knowledge and the other members. The losses are weighted in tandem with the interpolation when training.  The ensemble as a whole engenders a weight subspace that explicitly encodes tradeoffs and results in a continuous parameterization of the Pareto Front.", "strength_and_weaknesses": "Strength:\nThis paper castes multi-task problems as learning an ensemble of single-task predictors by interpolating among members during training, and designs weight subspaces where multiple optimal functional solutions lie. Based on this weight space, an ensembling method that can produce a continuous Pareto Front in a single training run is proposed. This allows practitioners to modulate the performance on each task during inference on the fly.  \n\nWeaknesses:\n\n1.Similar methods have already been proposed for multi-task learning and has not been disccussed in this paper [1].\n\n1.When sampling on the convex hull parameterization, authors choose to adopt the Dirichlet distribution since its support is the T-dimensional simplex. Does this distribution have other properties.   Why using this distribution?  If p≫1，how the ensemble will change. \n\n2.When training, a mono tonic relationship is imposed between the degree of a single-task predictor participation and the weight of the corresponding task loss. As a result, the ensemble engenders a subspace that explicitly encodes tradeoffs and results in a continuous parameterization of the Pareto Front.  Whether the mono tonic relationship can be replaced by other relationships? Explaining this point may be better.\n\n[1]Navon A, Shamsian A, Fetaya E, et al. Learning the Pareto Front with Hypernetworks[C]//International Conference on Learning Representations. 2020.", "clarity,_quality,_novelty_and_reproducibility": "Quality:\nThis paper is well written and easy to read. The organization is clear.\n\nClarity:\nSome issue of motivations and method details need more clarification. Please see weakness section.\n\nOriginality:\nfair. \n", "summary_of_the_review": "This paper produces a subspace with multiple Pareto stationary points in the multi-task learning based on the geometry of the loss landscape in single task machine learning where the local optimal are connected by simple paths. However, the representation and more illustrations are needed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667438908943}, {"id": "fmgT50KMSe", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4551/Reviewer_qexX"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a novel approach to seek a continuous Pareto Front for multi-task learning. The key idea is to train multiple single-task predictors and performs convex hull operations on different models in the weight space to produce a subspace with multiple Pareto stationary points. A continuous Pareto Front can be produced through only a single run in this method.", "review_text": "This paper focus on an interesting problem and provides some new ideas. However, there are some technical flaws in this work.", "strengths": "Strengths:\n\n1. This paper improves the method in Ma et al. (2020) and proposes a novel method to produce continuous Pareto stationary points, which is simple and easy to utilize. In addition, the method is efficient as it requires only a single run.\n2. This paper is inspired by the theory of single-task machine learning and the motivation is quite interesting. Meanwhile, the concept of Pareto manifold learning is insightful and worth exploring.\n3. The visual explanation provided in Figure 3 is easy to understand, while the visual analysis of the article (Figure 4 and Figure 5) is comprehensive.\n\nWeaknesses:\n\n1. It must be acknowledged that the conjecture about the loss landscape from the single-task to multi-task is insightful. However, it is better to give a toy example to explain it. More importantly, even though multiple valley intersections in the low loss region in the multi-task scenario, this observation is not strongly related to Pareto manifold learning. I suggest the author should explain it more clearly.\n2. In the last paragraph of the introduction, the author announces that “the algorithm produces a subspace of Pareto Optimal solutions”, It should be pointed out that multi-task learning methods can only ensure convergence to Pareto optimal when assuming it holds a convex loss, in other words, we can only approximate rather than obtain the Pareto optimal solutions.\n3. The proposed method fails to exceed the baseline methods such as Nash-MTL and PCGrad on several datasets(Table 1, Table 5, and Table 6), which makes the method less convincing. In addition, the ablation experiment demonstrated in Tabel 2 takes a value of $\\lambda$ equal to 0 on the MultiMNIST dataset, which may indicate that the regularization is not effective. Most importantly, there is a lack of detailed analysis of the experimental results.\n4. Paper writing should be polished, For instance, Figure 1 and Figure 2 are not mentioned in the context, which is not friendly to read.\n5. In addition, some typos exist in the paper. In line 11 of Algorithm 1, the citation of Figure 4.2 is wrong, the same problem is found in the corresponding text above Claim 3, I guess it should be Figure 3.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a novel approach to seek a continuous Pareto Front for multi-task learning. The key idea is to train multiple single-task predictors and performs convex hull operations on different models in the weight space to produce a subspace with multiple Pareto stationary points. A continuous Pareto Front can be produced through only a single run in this method.", "strength_and_weaknesses": "Strengths:\n\n1. This paper improves the method in Ma et al. (2020) and proposes a novel method to produce continuous Pareto stationary points, which is simple and easy to utilize. In addition, the method is efficient as it requires only a single run.\n2. This paper is inspired by the theory of single-task machine learning and the motivation is quite interesting. Meanwhile, the concept of Pareto manifold learning is insightful and worth exploring.\n3. The visual explanation provided in Figure 3 is easy to understand, while the visual analysis of the article (Figure 4 and Figure 5) is comprehensive.\n\nWeaknesses:\n\n1. It must be acknowledged that the conjecture about the loss landscape from the single-task to multi-task is insightful. However, it is better to give a toy example to explain it. More importantly, even though multiple valley intersections in the low loss region in the multi-task scenario, this observation is not strongly related to Pareto manifold learning. I suggest the author should explain it more clearly.\n2. In the last paragraph of the introduction, the author announces that “the algorithm produces a subspace of Pareto Optimal solutions”, It should be pointed out that multi-task learning methods can only ensure convergence to Pareto optimal when assuming it holds a convex loss, in other words, we can only approximate rather than obtain the Pareto optimal solutions.\n3. The proposed method fails to exceed the baseline methods such as Nash-MTL and PCGrad on several datasets(Table 1, Table 5, and Table 6), which makes the method less convincing. In addition, the ablation experiment demonstrated in Tabel 2 takes a value of $\\lambda$ equal to 0 on the MultiMNIST dataset, which may indicate that the regularization is not effective. Most importantly, there is a lack of detailed analysis of the experimental results.\n4. Paper writing should be polished, For instance, Figure 1 and Figure 2 are not mentioned in the context, which is not friendly to read.\n5. In addition, some typos exist in the paper. In line 11 of Algorithm 1, the citation of Figure 4.2 is wrong, the same problem is found in the corresponding text above Claim 3, I guess it should be Figure 3.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is well organized and easy to follow.", "summary_of_the_review": "This paper focus on an interesting problem and provides some new ideas. However, there are some technical flaws in this work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666784761516}, {"id": "hZA9Bbt3Sjr", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4551/Reviewer_NfGo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work proposes a Pareto manifold learning approach to produce a continuous Pareto front for a given multi-task learning (MTL) problem in a single run. The MTL problem is formulated as a multi-objective optimization problem, and the proposed method can be treated as an ensemble approach for multiple single-task models with linear weight interpolation. In the proposed method, each task (objective) has its own neural network model (with identical structure), and their linear weight combination will be explicitly associated with the corresponding trade-off losses (with linear scalarization). Experimental results show that this method can successfully find the Pareto subspaces for different MTL problems and outperforms other MTL methods.", "review_text": "This paper is well-written, and it studies an important research topic that could further bridge the field of multi-task learning and multi-objective optimization. I personally enjoy reading this work. However, due to the major concerns on main contributions, relation to closely related work, and the experimental setting, it is hard for me to vote to accept the current manuscript.", "strengths": "**Strengths:**\n\n+ This paper is well-written and easy to follow.\n\n+ Multi-task learning and multi-objective optimization are both important research topics, and this work tries to further bridge these two fields.  \n\n+ The proposed method has reasonable performance on different MTL problems (but with many concerns listed below).\n\n**Weaknesses**\n\nI have some major concerns on both the two claimed advantages of the proposed method, namely, a) learning the Pareto front for multi-task learning and b) better generalization performance.\n\n**1. Contribution and Closely Related Works**\n\nThis work proposes to learn the Pareto front for a multi-task learning problem. However, similar Pareto front learning approaches have already been proposed for multi-task learning [1,2,3,4]. Those methods all allow practitioners to choose their preferred solutions at inference time. All these closely related works are not discussed or compared in this paper. The proposed linear interpolation approach could be similar to using a simple linear hypernetwork in [1,2]. \n\nWithout a thorough discussion and comparison with these closely related works, it is very hard to judge the novelty and actual contribution of the proposed PML method.\n\n**2. Pareto Solution for Better Generalization Performance**\n\nThe motivation to learn the Pareto solution for better generalization performance is not strong. Recent works [5, 6] on multi-task learning have shown that a single fixed solution found by simple linear scalarization can have similar or even better performance than those found by multi-objective optimization methods. The effect of careful hyperparameters finetune rather than the Pareto properties could be crucial for good generalization performance.   \n\nIn addition, the experimental results on the found Pareto front and better per-task performance are somehow not solid. The major issue is that, for an MTL problem with m tasks, the proposed PML model is m times larger than the other baselines, which could lead to unfair comparison. For example, in Figure 4, some baseline models can have comparable or even better performance than the PML's Pareto front. If we can double their model capacity to match the one for PML, it could be likely that some of them (a single solution) can totally dominate the whole Pareto front by PML. In this case, the \"Pareto front\" could be useless since all tradeoffs are suboptimal and dominated.\n\nSimilarly, in Table 1, the overall per-task performance of PML is already nondominated with other MTL baselines. If the model capacity of all the baselines can be doubled, their performance could be significantly improved. On the other hand, the single-task learning baseline (STL) indeed has the same model capacity with PML. In this problem, it seems that simply having separate models for each task already significantly outperforms PML.   \n\n**3. Ensemble Learning Method**\n\nSince the proposed PML method requires learning multiple models and can be treated as an ensemble method, it should be compared with the closely related ensemble learning methods. A naive baseline could be to train m models with the same fixed weight among tasks (e.g., simple LS) or other multi-objective optimization (e.g., MGDA, CAGrad), and then a simple ensemble learning method [7] can be applied to achieve better overall MTL performance. An ensemble of models with different optimization methods, such as LS with different weights or MGDA + CAGrad, could be another choice.\n\nThe recent work on ensemble learning could have a smaller model size or faster inference time [8,9], and the weight average methods [10, 11] are also strong alternatives with a single final model. Some related works on mode connectivity have been briefly discussed but not compared in the paper. What is the advantage of the proposed PML model for generalization over those ensemble learning or weight average methods?    \n\n**4. Other Flat Minima Method**\n\nOne main motivation of this work is the connection between low-loss subspace (e.g., flat minima valley) and better generalization performance. From this viewpoint, it is interesting to know its relation to the entropy-sgd [12] and sharpness-aware minimization methods [13, 14, 15]. Can we simply use these methods to train an MTL model with fixed weights for each task? What is the advantage of PML over those methods? \n\n**Other Questions**\n\n1. In Algorithm 1, should the input size of \\theta be m rather than m?\n\n[1] Learning the Pareto Front with Hypernetworks. ICLR 2021.\n\n[2] Controllable Pareto Multi-Task Learning. arXiv:2010.06313.\n\n[3] Scalable Pareto Front Approximation for Deep Multi-Objective Learning. ICDM 2021.\n\n[4] Controllable Dynamic Multi-Task Architectures. CVPR 2022.\n\n[5] In Defense of the Unitary Scalarization for Deep Multi-Task Learning. arXiv:2201.04122, 2022.\n\n[6] Do Current Multi-Task Optimization Methods in Deep Learning Even Help? arXiv:2209.11379, 2022.\n\n[7] Simple and scalable predictive uncertainty estimation using deep ensembles. NeurIPS 2017.\n\n[8] BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. ICLR 2020.\n\n[9] Training independent subnetworks for robust prediction. ICLR 2021.\n\n[10] Averaging Weights Leads to Wider Optima and Better Generalization. UAI 2018.\n\n[11] Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. ICML 2022.\n\n[12] Entropy-SGD: Biasing gradient descent into wide valleys. ICLR 2017.\n\n[13] Sharp Minima Can Generalize For Deep Nets. ICML 2017.\n\n[14] Fantastic generalization measures and where to find them. ICLR 2020.\n\n[15] Sharpness-Aware Minimization for Efficiently Improving Generalization. ICLR 2021.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes a Pareto manifold learning approach to produce a continuous Pareto front for a given multi-task learning (MTL) problem in a single run. The MTL problem is formulated as a multi-objective optimization problem, and the proposed method can be treated as an ensemble approach for multiple single-task models with linear weight interpolation. In the proposed method, each task (objective) has its own neural network model (with identical structure), and their linear weight combination will be explicitly associated with the corresponding trade-off losses (with linear scalarization). Experimental results show that this method can successfully find the Pareto subspaces for different MTL problems and outperforms other MTL methods.", "strength_and_weaknesses": "**Strengths:**\n\n+ This paper is well-written and easy to follow.\n\n+ Multi-task learning and multi-objective optimization are both important research topics, and this work tries to further bridge these two fields.  \n\n+ The proposed method has reasonable performance on different MTL problems (but with many concerns listed below).\n\n**Weaknesses**\n\nI have some major concerns on both the two claimed advantages of the proposed method, namely, a) learning the Pareto front for multi-task learning and b) better generalization performance.\n\n**1. Contribution and Closely Related Works**\n\nThis work proposes to learn the Pareto front for a multi-task learning problem. However, similar Pareto front learning approaches have already been proposed for multi-task learning [1,2,3,4]. Those methods all allow practitioners to choose their preferred solutions at inference time. All these closely related works are not discussed or compared in this paper. The proposed linear interpolation approach could be similar to using a simple linear hypernetwork in [1,2]. \n\nWithout a thorough discussion and comparison with these closely related works, it is very hard to judge the novelty and actual contribution of the proposed PML method.\n\n**2. Pareto Solution for Better Generalization Performance**\n\nThe motivation to learn the Pareto solution for better generalization performance is not strong. Recent works [5, 6] on multi-task learning have shown that a single fixed solution found by simple linear scalarization can have similar or even better performance than those found by multi-objective optimization methods. The effect of careful hyperparameters finetune rather than the Pareto properties could be crucial for good generalization performance.   \n\nIn addition, the experimental results on the found Pareto front and better per-task performance are somehow not solid. The major issue is that, for an MTL problem with m tasks, the proposed PML model is m times larger than the other baselines, which could lead to unfair comparison. For example, in Figure 4, some baseline models can have comparable or even better performance than the PML's Pareto front. If we can double their model capacity to match the one for PML, it could be likely that some of them (a single solution) can totally dominate the whole Pareto front by PML. In this case, the \"Pareto front\" could be useless since all tradeoffs are suboptimal and dominated.\n\nSimilarly, in Table 1, the overall per-task performance of PML is already nondominated with other MTL baselines. If the model capacity of all the baselines can be doubled, their performance could be significantly improved. On the other hand, the single-task learning baseline (STL) indeed has the same model capacity with PML. In this problem, it seems that simply having separate models for each task already significantly outperforms PML.   \n\n**3. Ensemble Learning Method**\n\nSince the proposed PML method requires learning multiple models and can be treated as an ensemble method, it should be compared with the closely related ensemble learning methods. A naive baseline could be to train m models with the same fixed weight among tasks (e.g., simple LS) or other multi-objective optimization (e.g., MGDA, CAGrad), and then a simple ensemble learning method [7] can be applied to achieve better overall MTL performance. An ensemble of models with different optimization methods, such as LS with different weights or MGDA + CAGrad, could be another choice.\n\nThe recent work on ensemble learning could have a smaller model size or faster inference time [8,9], and the weight average methods [10, 11] are also strong alternatives with a single final model. Some related works on mode connectivity have been briefly discussed but not compared in the paper. What is the advantage of the proposed PML model for generalization over those ensemble learning or weight average methods?    \n\n**4. Other Flat Minima Method**\n\nOne main motivation of this work is the connection between low-loss subspace (e.g., flat minima valley) and better generalization performance. From this viewpoint, it is interesting to know its relation to the entropy-sgd [12] and sharpness-aware minimization methods [13, 14, 15]. Can we simply use these methods to train an MTL model with fixed weights for each task? What is the advantage of PML over those methods? \n\n**Other Questions**\n\n1. In Algorithm 1, should the input size of \\theta be m rather than m?\n\n[1] Learning the Pareto Front with Hypernetworks. ICLR 2021.\n\n[2] Controllable Pareto Multi-Task Learning. arXiv:2010.06313.\n\n[3] Scalable Pareto Front Approximation for Deep Multi-Objective Learning. ICDM 2021.\n\n[4] Controllable Dynamic Multi-Task Architectures. CVPR 2022.\n\n[5] In Defense of the Unitary Scalarization for Deep Multi-Task Learning. arXiv:2201.04122, 2022.\n\n[6] Do Current Multi-Task Optimization Methods in Deep Learning Even Help? arXiv:2209.11379, 2022.\n\n[7] Simple and scalable predictive uncertainty estimation using deep ensembles. NeurIPS 2017.\n\n[8] BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. ICLR 2020.\n\n[9] Training independent subnetworks for robust prediction. ICLR 2021.\n\n[10] Averaging Weights Leads to Wider Optima and Better Generalization. UAI 2018.\n\n[11] Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. ICML 2022.\n\n[12] Entropy-SGD: Biasing gradient descent into wide valleys. ICLR 2017.\n\n[13] Sharp Minima Can Generalize For Deep Nets. ICML 2017.\n\n[14] Fantastic generalization measures and where to find them. ICLR 2020.\n\n[15] Sharpness-Aware Minimization for Efficiently Improving Generalization. ICLR 2021.\n", "clarity,_quality,_novelty_and_reproducibility": "**Clarity:** This paper is well-written and easy to follow.\n\n**Quality:** The paper has a good overall quality, but there are also some major concerns on its main contributions as listed in the weaknesses.\n\n**Novelty:** Some very closely related works are not discussed or compared in this paper, which makes it very hard to judge the novelty and actual contribution of this work.  \n\n**Reproducibility:** The proposed model structure seems simple and straightforward. But the experimental results could be not robust, given 1) the current findings on MTL with multi-objective optimization, and 2) the unsolid experimental setting as discussed in the weaknesses.", "summary_of_the_review": "This paper is well-written, and it studies an important research topic that could further bridge the field of multi-task learning and multi-objective optimization. I personally enjoy reading this work. However, due to the major concerns on main contributions, relation to closely related work, and the experimental setting, it is hard for me to vote to accept the current manuscript.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666670201922}], "openreview_url": "https://openreview.net/forum?id=C9uEwyfklBE", "arxiv_id": "2210.09759", "paper_pdf": "papers/C9uEwyfklBE.pdf", "paper_pdf_sha256": "30ceebbfacd9d11d5c62de0d6cc2b6d547598d842226e3f2d177bc7e57db9cc4", "paper_pdf_bytes": 23785593, "paper_pdf_source": "openreview", "code_url": "https://github.com/nik-dim/pamal", "code_repository": "nik-dim/pamal", "code_commit": "456a338bac579977ddb7100c2bae36f2768df0c6", "code_archive": "repos/C9uEwyfklBE.zip", "code_archive_sha256": "f37acbc245fbec724186f83cfdd57c0b2b78b8a34582cb532c7a2f793815cc10", "code_archive_bytes": 347349, "code_file_count": 64, "code_extensions": {".py": 63, ".sh": 1}, "github_disk_usage_kb": 315, "github_languages": {"Python": 325702, "Shell": 1980, "Dockerfile": 484}, "github_archived": false, "github_pushed_at": "2026-06-30T16:37:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pareto-manifold-learning-tackling-multiple"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WZR7ckBkzPY", "year": 2022, "status": "rejected", "title": "Variational Wasserstein gradient flow", "authors": ["Jiaojiao Fan", "Amirhossein Taghvaei", "Yongxin Chen"], "authorids": ["~Jiaojiao_Fan1", "~Amirhossein_Taghvaei1", "~Yongxin_Chen1"], "authors_source": "OpenReview API", "abstract": "The gradient flow of a function over the space of probability densities with respect to the Wasserstein metric often exhibits nice properties and has been utilized in several machine learning applications. The standard approach to compute the Wasserstein gradient flow is the finite difference which discretizes the underlying space over a grid, and is not scalable. In this work, we propose a scalable proximal gradient type algorithm for Wasserstein gradient flow. The key of our method is a variational formulation of the objective function, which makes it possible to realize the JKO proximal map through a primal-dual optimization. This primal-dual problem can be efficiently solved by alternatively updating the parameters in the inner and outer loops. Our framework covers all the classical Wasserstein gradient flows including the heat equation and the porous medium equation. We demonstrate the performance and scalability of our algorithm with several numerical examples.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "IfbHE9O91_W", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper293/Reviewer_pvCT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method to solve Wasserstein gradient flows based on the JKO scheme using variational formulations of functional objectives, such as the KL divergence or the generalized entropy (non-linear diffusion). Relying on known reformulations of the JKO scheme as optimization over convex functions, the paper departs from recent related methods in expressing certain objectives as f-divergences, and in turn using the dual formulation of these divergences to circumvent the need to do explicit density computation in these. The resulting method involves parametrizing two types of operators as neural networks (one of them as an input-convex neural network), and solving a mini-max objective. The paper presents experiments on simple PDEs (mostly in 1D or 2D) with known solutions. ", "review_text": "Strengths:\n* An ingenious use of f-divergence duality (aka variational formulation) to re-write density-depending functional flow objectives, such as the generalized entropy, as (optimization of) expectations, allowing for computation via finite-sample approximation \n* The resulting method seems to be empirically valid in low- and mid-dimensional settings \n\nWeaknesses:\n* Novelty, the main (and perhaps, only) novelty of this approach compared to (Mokrov et al., 2021; Alvarez-Melis et al., 2021; Bunne et al., 2021) is the use of the variational formulation of f-divergences, something that is itself well known. Note that this is only relevant for functional objectives that depend on the density of the measure itself, e.g., those that cannot be expressed as an expectation over the measure. All other objectives, including the interaction energy considered here, can be tackled with the exact same approach of the three papers above. \n* The novel contributions should be more clearly separated from non-novel ones. E.g., the reformulation of WGF as optimization over convex functions might seem like a contribution of this paper to an uninitiated reader, although it has been explored extensively in other work, most prominently in Benamou et al (2014), which is puzzling not cited here, and more recently in the 3 JKO + ICNN papers cited above.  \n* Some statements are not clear, not correct or too vague. These should all be clarified or corrected. For example:\n    * \"that is not too far away from the identity map\" in Pg 5. \n    * In section 2.2, the $\\delta F / \\delta P$ is not formally a gradient, but the first-variation of a functional\n    * In Section 3.6, it is stated that prior works requires explicit computation of log determinants of Hessians at cubic complexity, but this is not the case. At least one of those methods uses cuadrtic-cost (matrix-vector-product) stochastic estimators of the Hess logdet.\n* The motivation for using the FB scheme for the interaction energy is not clear. Some of the intuition/motivation discussed in sections B.2 and B.3 should probably be moved to the main text. \n* Limited experimental validation, especially with regards to high-dimensional settings, which is arguably the main promise of this work. In addition, all experiments are synthetic. A more compelling evaluation framework should include at least one high-dimensional, realistic dataset. \n* The evaluation is mostly qualitative. The paper purposefuly chooses PDEs with known solutions, but then provides mostly qualitative/visual results. A more compelling evaluation framwork would include quantitative comparison against these known solutions in sections 4.3 and 4.4.\n* While the computational complexity summary provided in the paper is useful, there are some many hidden constants in those bounds that they are hardly useful. These should be complemented with a thorough empirical runtime analysis comparing against exact and inexact methods, such as those cited here as related work. \n\nMinor comments:\n* While the motivation is different, there are deep and unexplored connections between the dualizatin of the f-divergence objectives used here and the dual of the kantorovich OT problem that has been recently explored extensively for learning Monge maps between distributions. Many of these rely on convex conjugacy to reformulate a sup objective as a sup-inf one (see Korotin et al. 2021b for an excellent survey on these). It would have been great (though not obligatory!) to see a discussion on these connections \n* I would suggest moving Table 1 after Proposition 1, so that all objects in the definition of $\\mathcal{A}(T,h)$ have been already introduced. I spent some time trying to figure out what $\\mu(T)$ was before finding it further down.\n\nMissing related work:\n\n* Benamou et al, \"Discretization of functionals involving the Monge-Ampère operator\". 2014.\n* Huang et al, \"Convex Potential Flows: Universar Probability Distributions with Optimal Transport and Convex Optimization\", ICLR 2021.\n* Bunne et al., \"JKOnet: Proximal Optimal Transport Modeling of Population Dynamics\", 2021.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a method to solve Wasserstein gradient flows based on the JKO scheme using variational formulations of functional objectives, such as the KL divergence or the generalized entropy (non-linear diffusion). Relying on known reformulations of the JKO scheme as optimization over convex functions, the paper departs from recent related methods in expressing certain objectives as f-divergences, and in turn using the dual formulation of these divergences to circumvent the need to do explicit density computation in these. The resulting method involves parametrizing two types of operators as neural networks (one of them as an input-convex neural network), and solving a mini-max objective. The paper presents experiments on simple PDEs (mostly in 1D or 2D) with known solutions. ", "main_review": "Strengths:\n* An ingenious use of f-divergence duality (aka variational formulation) to re-write density-depending functional flow objectives, such as the generalized entropy, as (optimization of) expectations, allowing for computation via finite-sample approximation \n* The resulting method seems to be empirically valid in low- and mid-dimensional settings \n\nWeaknesses:\n* Novelty, the main (and perhaps, only) novelty of this approach compared to (Mokrov et al., 2021; Alvarez-Melis et al., 2021; Bunne et al., 2021) is the use of the variational formulation of f-divergences, something that is itself well known. Note that this is only relevant for functional objectives that depend on the density of the measure itself, e.g., those that cannot be expressed as an expectation over the measure. All other objectives, including the interaction energy considered here, can be tackled with the exact same approach of the three papers above. \n* The novel contributions should be more clearly separated from non-novel ones. E.g., the reformulation of WGF as optimization over convex functions might seem like a contribution of this paper to an uninitiated reader, although it has been explored extensively in other work, most prominently in Benamou et al (2014), which is puzzling not cited here, and more recently in the 3 JKO + ICNN papers cited above.  \n* Some statements are not clear, not correct or too vague. These should all be clarified or corrected. For example:\n    * \"that is not too far away from the identity map\" in Pg 5. \n    * In section 2.2, the $\\delta F / \\delta P$ is not formally a gradient, but the first-variation of a functional\n    * In Section 3.6, it is stated that prior works requires explicit computation of log determinants of Hessians at cubic complexity, but this is not the case. At least one of those methods uses cuadrtic-cost (matrix-vector-product) stochastic estimators of the Hess logdet.\n* The motivation for using the FB scheme for the interaction energy is not clear. Some of the intuition/motivation discussed in sections B.2 and B.3 should probably be moved to the main text. \n* Limited experimental validation, especially with regards to high-dimensional settings, which is arguably the main promise of this work. In addition, all experiments are synthetic. A more compelling evaluation framework should include at least one high-dimensional, realistic dataset. \n* The evaluation is mostly qualitative. The paper purposefuly chooses PDEs with known solutions, but then provides mostly qualitative/visual results. A more compelling evaluation framwork would include quantitative comparison against these known solutions in sections 4.3 and 4.4.\n* While the computational complexity summary provided in the paper is useful, there are some many hidden constants in those bounds that they are hardly useful. These should be complemented with a thorough empirical runtime analysis comparing against exact and inexact methods, such as those cited here as related work. \n\nMinor comments:\n* While the motivation is different, there are deep and unexplored connections between the dualizatin of the f-divergence objectives used here and the dual of the kantorovich OT problem that has been recently explored extensively for learning Monge maps between distributions. Many of these rely on convex conjugacy to reformulate a sup objective as a sup-inf one (see Korotin et al. 2021b for an excellent survey on these). It would have been great (though not obligatory!) to see a discussion on these connections \n* I would suggest moving Table 1 after Proposition 1, so that all objects in the definition of $\\mathcal{A}(T,h)$ have been already introduced. I spent some time trying to figure out what $\\mu(T)$ was before finding it further down.\n\nMissing related work:\n\n* Benamou et al, \"Discretization of functionals involving the Monge-Ampère operator\". 2014.\n* Huang et al, \"Convex Potential Flows: Universar Probability Distributions with Optimal Transport and Convex Optimization\", ICLR 2021.\n* Bunne et al., \"JKOnet: Proximal Optimal Transport Modeling of Population Dynamics\", 2021.", "summary_of_the_review": "The paper provides an interesting variation on recent JKO-based methods for computational Wasserstein gradient flows, but its limited novelty and empirical evaluation diminish its contribution, and make it a borderline paper in my view. That being said, if the issues I raise in my review are properly addressed, I would be willing to increase my score. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635896580077}, {"id": "7JcBbeRRQe2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper293/Reviewer_72fj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies the implementation of some Wasserstein Gradient Flows (WGF) in discrete time but without discretizing the space. The methods proposed are based on the JKO operator to discretize WGF in time. The implementation of the JKO can be challenging. The strategy of the authors is to first reparametrize the JKO as a minimization over a space of functions (instead of measures) via pushforward. Then, when the objective function is a f-divergence, the objective inside the JKO admit a variational representation and can be expressed as a sup. Conclusion: each JKO is written as a min max over a space of functions. To solve it, they parametrize the functions by neural networks and alternatively maximize and minimize the problem using Adam. An important feature is that the objective in the min max can be approximated with samples of the current distribution (its density doesn't appear, only integrals wrt to the current distribution). ", "review_text": "Strengths:\n\nThe paper is well written and high level ideas on the theory are explained. The approach is clear and reasonable. The simulations show promising results, they do not seem to have been cherry picked. They show what is claimed by the authors and there is no surprise given the simplicity of the approach.\n\nWeaknesses:\n\nThe theoretical consistence of the proposed methods is ignored. Are the proposed scheme consistent time discretizations of the WGF? Are the assumptions satisfied? Is the alternative maximization/minimization strategy (Algo 1) consistent? Do all measures considered admit density wrt Lebesgue? The paper mainly provides intuition for Algo 1 without solid theoretical foundations. \n\nMoreover, the technical contribution is rather limited (see the summary of the paper above). One could argue that it is an \"easy paper\" in the sense that only the simulations seem new. The novelty is mainly to have parametrized the functions in the objective of the JKO by neural networks. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the implementation of some Wasserstein Gradient Flows (WGF) in discrete time but without discretizing the space. The methods proposed are based on the JKO operator to discretize WGF in time. The implementation of the JKO can be challenging. The strategy of the authors is to first reparametrize the JKO as a minimization over a space of functions (instead of measures) via pushforward. Then, when the objective function is a f-divergence, the objective inside the JKO admit a variational representation and can be expressed as a sup. Conclusion: each JKO is written as a min max over a space of functions. To solve it, they parametrize the functions by neural networks and alternatively maximize and minimize the problem using Adam. An important feature is that the objective in the min max can be approximated with samples of the current distribution (its density doesn't appear, only integrals wrt to the current distribution). ", "main_review": "Strengths:\n\nThe paper is well written and high level ideas on the theory are explained. The approach is clear and reasonable. The simulations show promising results, they do not seem to have been cherry picked. They show what is claimed by the authors and there is no surprise given the simplicity of the approach.\n\nWeaknesses:\n\nThe theoretical consistence of the proposed methods is ignored. Are the proposed scheme consistent time discretizations of the WGF? Are the assumptions satisfied? Is the alternative maximization/minimization strategy (Algo 1) consistent? Do all measures considered admit density wrt Lebesgue? The paper mainly provides intuition for Algo 1 without solid theoretical foundations. \n\nMoreover, the technical contribution is rather limited (see the summary of the paper above). One could argue that it is an \"easy paper\" in the sense that only the simulations seem new. The novelty is mainly to have parametrized the functions in the objective of the JKO by neural networks. ", "summary_of_the_review": "The approach is reasonable, the simulation promising but I do not see a significant technical contribution (They essentially reparametrized a min max problem over a function space using neural nets and running Adam to alternatively maximize and minimize). ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635894695653}, {"id": "zYGuvjjK9mb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper293/Reviewer_S3mH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a variational formulation of each JKO step for optimizing functionals on measures.\nDifferent from existing recent works on emulating JKO steps by training pushforward neural networks (either directly or as gradients of convex functions), the variational formulation involves another inner maximization of a function, without needing density access that typically requires cubic time complexity due to computing the log determinants of the pushforwards. Experiments are done to demonstrate the practicality of the algorithm.\n", "review_text": "This paper identifies a crucial challenge in the existing works of emulating JKO steps, namely the expensive computation of the densities of the form $P_k(x)$ for each $k$. The solution the paper suggests is reasonable, but it comes at the cost of adding another inner maximization which makes the optimization a lot more difficult (e.g. unstable due to high variance). Overall I think the amount of contribution is very limited. The variational formulation of the objective functionals considered is all well known, and putting it together with the JKO step is fairly straightforward. The experiments are not convincing enough to demonstrate the practical advantages of the proposed method compared to the alternatives.\n\nDetailed comments:\n- In Eq (9), a distribution $\\Gamma$ is introduced. What is the point of introducing this measure? Is it essentially an importance sampling of $Q$ for which we only know the density? Why is it enough to just choose Gaussians, which could be very different from $Q$?\n- Many sections are very similar to [Mokrov 2021]. For example, there is nothing new in Sec 3.5, and the experiment's setup in 4.1 and 4.2 are exactly the same. Yet the more challenging experiments from [Mokrov 2021] are not reproduced here, such as posterior inference and non-linear filtering. \n- In Sec 3.6, it is common to use the Hutchinson trace estimator to approximate the gradient of $\\log \\det$ (in addition to a linear solve), which could speed up the competing methods. It might be good to include a comparison to that.\n- The results in Figure 2 are visibly worse than those of [Mokrov 2021]. Moreover, here only up to dimension 13 is included, whereas [Mokrov 2021] contains dimension 32.\n- The results in Figure 3 of the proposed method are better than those of [Mokrov 2021]. I'm wondering why this is the case since in [Mokrov 2021] the KL divergence is calculated exactly, whereas in the proposed method additional bias could be introduced due to the failure of maximizing $h$.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a variational formulation of each JKO step for optimizing functionals on measures.\nDifferent from existing recent works on emulating JKO steps by training pushforward neural networks (either directly or as gradients of convex functions), the variational formulation involves another inner maximization of a function, without needing density access that typically requires cubic time complexity due to computing the log determinants of the pushforwards. Experiments are done to demonstrate the practicality of the algorithm.\n", "main_review": "This paper identifies a crucial challenge in the existing works of emulating JKO steps, namely the expensive computation of the densities of the form $P_k(x)$ for each $k$. The solution the paper suggests is reasonable, but it comes at the cost of adding another inner maximization which makes the optimization a lot more difficult (e.g. unstable due to high variance). Overall I think the amount of contribution is very limited. The variational formulation of the objective functionals considered is all well known, and putting it together with the JKO step is fairly straightforward. The experiments are not convincing enough to demonstrate the practical advantages of the proposed method compared to the alternatives.\n\nDetailed comments:\n- In Eq (9), a distribution $\\Gamma$ is introduced. What is the point of introducing this measure? Is it essentially an importance sampling of $Q$ for which we only know the density? Why is it enough to just choose Gaussians, which could be very different from $Q$?\n- Many sections are very similar to [Mokrov 2021]. For example, there is nothing new in Sec 3.5, and the experiment's setup in 4.1 and 4.2 are exactly the same. Yet the more challenging experiments from [Mokrov 2021] are not reproduced here, such as posterior inference and non-linear filtering. \n- In Sec 3.6, it is common to use the Hutchinson trace estimator to approximate the gradient of $\\log \\det$ (in addition to a linear solve), which could speed up the competing methods. It might be good to include a comparison to that.\n- The results in Figure 2 are visibly worse than those of [Mokrov 2021]. Moreover, here only up to dimension 13 is included, whereas [Mokrov 2021] contains dimension 32.\n- The results in Figure 3 of the proposed method are better than those of [Mokrov 2021]. I'm wondering why this is the case since in [Mokrov 2021] the KL divergence is calculated exactly, whereas in the proposed method additional bias could be introduced due to the failure of maximizing $h$.\n", "summary_of_the_review": "The paper studies an important challenge in JKO steps encountered by recent works, but the contributions are incremental without demonstrating convincing practical advantages.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635792689069}, {"id": "Wl5ufYQpO6k", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper293/Reviewer_CWde"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a method to compute Wasserstein Gradient Flows (WGFs) via neural networks and the JKO scheme. In contrast to prior works, to compute WGFs of functionals involving f-divergences, the authors use variational approximations rather than direct computations. It is claimed to work faster and perform better.", "review_text": "**Benefits**\n1) The variational approach might be a solution to issues of direct computation such as cubically growing complexity as a function of the dimension;\n\n**Drawbacks**\n\n2) The experiments of the paper are weak and do not sufficiently support the main claims;\n3) The relation to the prior work is not fully disclosed;\n4) The scope of the paper (WGFs) is narrow.\n\n*Detailed comments are below.*\n\n**Relation to prior work.** A large part of the algorithm proposed by the authors matches the previously known art. More precisely, sections 3.0 and 3.1 exploit the reformulation of JKO via functions that have already been proposed in [1] (not cited in the paper!) and extensively used in [2,3]. I think this is not very clear from the text and might add extra (inexistent) value to the current work.\n\nIf I correctly understand, the actual difference w.r.t. the prior work, e.g., [2], is the way that the f-divergence is optimized. The authors use a variational approximation in contrast to direct computation. Here I have two remarks.\n\nFirst, the key claim of the paper is that direct computation scales cubically and is not feasible in high dimensions. I agree, but what about fast approximations? In [3] the authors explicitly state they use a fast estimator based on the Hutchinson method. In [2] the authors state that to speed up the computation fast approximations can be used. It is unfair to ignore this and compare in the experimental section only with [2] using and only by using the direct computation.\n\nSecond, the variational approximation of the f-divergence is not novel, see, for example, f-GANs [5]. I wonder if such approximations have already been used in Bayesian machine learning (BML)? I encourage the authors to include a detailed discussion of this. Why is this important? I am not an expert in BML, but it seems to me that this particular variational approximation can be applied to a dozen of tasks involving KL divergence. The current paper demonstrates that it outperforms direct computation. If this is indeed true, why hasn’t this approach been used for other BML tasks? If it has already been, the authors should include relevant references to support the current experimental findings.\n\nOverall, the paper should more carefully acknowledge the prior work and make it transparent what is new and what is already well known. Paper [4] also seems relevant.\n\n**Experiments.** The experimental section is very weak both in terms of quality and quantity. In terms of quantity, the paper looks poor compared to predecessors, e.g., [2], a paper to which they compare their method. In particular, the results provided in Figure 2 visually suffer from notable artifacts, which is a contrast to that reported by [2] in a similar setup (Figure 2 of [2]). Besides, the dimensions considered in the current paper are lower, which is suspicious.\n\nImportantly, the quantitative results provided in Figure 3 raise questions. How is it possible that on such a toy example (evolving Gaussian distributions -- linear pushforward maps grad psi), the variational methods so drastically outperform the direct computation (up to 10 times)? This is quite unbelievable, in particular, in small dimensions. Did the authors use the same network architectures and other shared hyperparameters for comparison? A discussion here is needed.\n\n**Scope.** If I correctly understand, the authors do not provide any high-dimensional applications of WGFs. Due to this, I currently tend to think the scope of the paper is narrow and the usefulness of the proposed approach for the community of ICLR is questionable. Adding an application would definitely benefit the paper. \n\n**Correctness of the method.** Overall, the method is correct. However, in one of the experiments, the authors approximate the pushforward of the JKO step directly as the neural network, not as the gradient of the input-convex network. In this case, the push forward distribution might not have density damaging the entire JKO scheme.\n\n**Clarity.** What for the FB scheme is introduced (section 3.3)? This is not clear from the text.\n\n**References**\n\n[1] Benamou, J. D., Carlier, G., Mérigot, Q., & Oudet, E. (2016). Discretization of functionals involving the Monge–Ampère operator. Numerische mathematik, 134(3), 611-636.\n\n[2] Mokrov, P., Korotin, A., Li, L., Genevay, A., Solomon, J., & Burnaev, E. (2021). Large-Scale Wasserstein Gradient Flows. arXiv preprint arXiv:2106.00736.\n\n[3] Alvarez-Melis, D., Schiff, Y., & Mroueh, Y. (2021). Optimizing Functionals on the Space of Probabilities with Input Convex Neural Networks. arXiv preprint arXiv:2106.00774.\n\n[4] Bunne, C., Meng-Papaxanthos, L., Krause, A., & Cuturi, M. (2021). JKOnet: Proximal Optimal Transport Modeling of Population Dynamics. arXiv preprint arXiv:2106.06345.\n\n[5] Nowozin, S., Cseke, B., & Tomioka, R. (2016, December). f-gan: Training generative neural samplers using variational divergence minimization. In Proceedings of the 30th International Conference on Neural Information Processing Systems (pp. 271-279).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method to compute Wasserstein Gradient Flows (WGFs) via neural networks and the JKO scheme. In contrast to prior works, to compute WGFs of functionals involving f-divergences, the authors use variational approximations rather than direct computations. It is claimed to work faster and perform better.", "main_review": "**Benefits**\n1) The variational approach might be a solution to issues of direct computation such as cubically growing complexity as a function of the dimension;\n\n**Drawbacks**\n\n2) The experiments of the paper are weak and do not sufficiently support the main claims;\n3) The relation to the prior work is not fully disclosed;\n4) The scope of the paper (WGFs) is narrow.\n\n*Detailed comments are below.*\n\n**Relation to prior work.** A large part of the algorithm proposed by the authors matches the previously known art. More precisely, sections 3.0 and 3.1 exploit the reformulation of JKO via functions that have already been proposed in [1] (not cited in the paper!) and extensively used in [2,3]. I think this is not very clear from the text and might add extra (inexistent) value to the current work.\n\nIf I correctly understand, the actual difference w.r.t. the prior work, e.g., [2], is the way that the f-divergence is optimized. The authors use a variational approximation in contrast to direct computation. Here I have two remarks.\n\nFirst, the key claim of the paper is that direct computation scales cubically and is not feasible in high dimensions. I agree, but what about fast approximations? In [3] the authors explicitly state they use a fast estimator based on the Hutchinson method. In [2] the authors state that to speed up the computation fast approximations can be used. It is unfair to ignore this and compare in the experimental section only with [2] using and only by using the direct computation.\n\nSecond, the variational approximation of the f-divergence is not novel, see, for example, f-GANs [5]. I wonder if such approximations have already been used in Bayesian machine learning (BML)? I encourage the authors to include a detailed discussion of this. Why is this important? I am not an expert in BML, but it seems to me that this particular variational approximation can be applied to a dozen of tasks involving KL divergence. The current paper demonstrates that it outperforms direct computation. If this is indeed true, why hasn’t this approach been used for other BML tasks? If it has already been, the authors should include relevant references to support the current experimental findings.\n\nOverall, the paper should more carefully acknowledge the prior work and make it transparent what is new and what is already well known. Paper [4] also seems relevant.\n\n**Experiments.** The experimental section is very weak both in terms of quality and quantity. In terms of quantity, the paper looks poor compared to predecessors, e.g., [2], a paper to which they compare their method. In particular, the results provided in Figure 2 visually suffer from notable artifacts, which is a contrast to that reported by [2] in a similar setup (Figure 2 of [2]). Besides, the dimensions considered in the current paper are lower, which is suspicious.\n\nImportantly, the quantitative results provided in Figure 3 raise questions. How is it possible that on such a toy example (evolving Gaussian distributions -- linear pushforward maps grad psi), the variational methods so drastically outperform the direct computation (up to 10 times)? This is quite unbelievable, in particular, in small dimensions. Did the authors use the same network architectures and other shared hyperparameters for comparison? A discussion here is needed.\n\n**Scope.** If I correctly understand, the authors do not provide any high-dimensional applications of WGFs. Due to this, I currently tend to think the scope of the paper is narrow and the usefulness of the proposed approach for the community of ICLR is questionable. Adding an application would definitely benefit the paper. \n\n**Correctness of the method.** Overall, the method is correct. However, in one of the experiments, the authors approximate the pushforward of the JKO step directly as the neural network, not as the gradient of the input-convex network. In this case, the push forward distribution might not have density damaging the entire JKO scheme.\n\n**Clarity.** What for the FB scheme is introduced (section 3.3)? This is not clear from the text.\n\n**References**\n\n[1] Benamou, J. D., Carlier, G., Mérigot, Q., & Oudet, E. (2016). Discretization of functionals involving the Monge–Ampère operator. Numerische mathematik, 134(3), 611-636.\n\n[2] Mokrov, P., Korotin, A., Li, L., Genevay, A., Solomon, J., & Burnaev, E. (2021). Large-Scale Wasserstein Gradient Flows. arXiv preprint arXiv:2106.00736.\n\n[3] Alvarez-Melis, D., Schiff, Y., & Mroueh, Y. (2021). Optimizing Functionals on the Space of Probabilities with Input Convex Neural Networks. arXiv preprint arXiv:2106.00774.\n\n[4] Bunne, C., Meng-Papaxanthos, L., Krause, A., & Cuturi, M. (2021). JKOnet: Proximal Optimal Transport Modeling of Population Dynamics. arXiv preprint arXiv:2106.06345.\n\n[5] Nowozin, S., Cseke, B., & Tomioka, R. (2016, December). f-gan: Training generative neural samplers using variational divergence minimization. In Proceedings of the 30th International Conference on Neural Information Processing Systems (pp. 271-279).", "summary_of_the_review": "My overall impression of the paper is that it is unfinished. While the idea of variational approximation is reasonable, I suppose this paper requires a major revision with a dozen text improvements and experiments. Therefore, I vote to reject this paper in its current form.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None", "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634762495466}], "openreview_url": "https://openreview.net/forum?id=WZR7ckBkzPY", "arxiv_id": "2112.02424", "paper_pdf": "papers/WZR7ckBkzPY.pdf", "paper_pdf_sha256": "3656e16fe03ca9803436ebbea2c5a062cb6dd5aadb6afc2f5ac5d86b756bae23", "paper_pdf_bytes": 4584986, "paper_pdf_source": "openreview", "code_url": "https://github.com/sbyebss/variational_wgf", "code_repository": "sbyebss/variational_wgf", "code_commit": "b0e22909b0b145989088af8df174eccba189a23e", "code_archive": "repos/WZR7ckBkzPY.zip", "code_archive_sha256": "96640698eb10a64618d2b79266f1d282e5c63e167b2063586bfc4068c4d226b1", "code_archive_bytes": 899253, "code_file_count": 90, "code_extensions": {".py": 85, ".sh": 4, ".ipynb": 1}, "github_disk_usage_kb": 823, "github_languages": {"Jupyter Notebook": 409440, "Python": 295837, "Shell": 1775}, "github_archived": false, "github_pushed_at": "2022-10-17T05:45:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variational-wasserstein-gradient-flow-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lXW6Sk1075v", "year": 2021, "status": "rejected", "title": "FORK: A FORward-looKing Actor for Model-Free Reinforcement Learning", "authors": ["Honghao Wei", "Lei Ying"], "authorids": ["~Honghao_Wei2", "~Lei_Ying1"], "authors_source": "OpenReview API", "abstract": "In this paper, we propose a new type of Actor, named forward-looking Actor or FORK for short, for Actor-Critic algorithms. FORK can be easily integrated into a model-free Actor-Critic algorithm. Our experiments on six Box2D and MuJoCo environments with continuous state and action spaces demonstrate significant performance improvement FORK can bring to the state-of-the-art algorithms. A variation of FORK can further solve BipedalWalkerHardcore in as few as four hours using a single GPU.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "PqCWz-yrPLv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper653/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary\nThis paper focuses on the field of off-policy reinforcement learning. Specifically, the authors propose a model-based reinforcement learning method on top of actor-critic methods. The proposed method trains a dynamics model and a reward function on the off-policy data with supervised learning, and then uses the trained model to generate synthetic future states and rewards during the actor update. During policy update for a given state, the method computes the sum of the Q value estimate of the state and the Q value expansion for a few steps using the learned dynamics model and reward function.\n\nThe authors implement the proposed method on top of SAC, and TD3, and evaluate their performances on several MuJoCo and Box2D environments. The experiment results show that the proposed method outperforms the model-free baseline in terms of sample efficiency.\n\n### Comments\nThe paper is well written and the idea proposed in this paper is really easy to understand. The authors also include a wide suite of experiments to demonstrate the sample efficiency of the proposed method. Despite these advantages, I cannot recommend acceptance of this paper due to the lack of novelty and absence of fair baseline comparison, which I will elaborate on next.\n\nFirst of all, despite the title of the paper, the proposed method is really a model-based reinforcement learning method since the system network and reward network are just dynamics model and reward model. The proposed objective for the policy (eqn 2), is merely a sum of current Q value estimate and Q value expansion for a few steps using the learned model, which has been proposed before in various papers such as [1] and [2]. The only difference is that when computing the gradient with respect to the policy, the authors leave out the gradient that passes through the learned model, which results in biased estimate of the policy gradient. Therefore, I’m not convinced about the novelty of the proposed method.\n\nMoreover, while proposing a model-based method. The authors do not include baseline comparisons with other model-based RL methods. It is widely known that on low-dimensional control tasks, model-based method outperforms model-free methods ([3]), and therefore merely comparing to model-free baselines is unfair. It would be important to include comparisons to model based methods ([3]).\n\nDue to the lack of novelty and fair comparison to existing model-based methods, I cannot recommend acceptance for this paper.\n\n\n\nReferences\n\n[1] Heess, Nicolas, et al. \"Learning continuous control policies by stochastic value gradients.\" Advances in Neural Information Processing Systems. 2015.\n\n[2] Clavera, Ignasi, Yao Fu, and Pieter Abbeel. \"Model-Augmented Actor-Critic: Backpropagating through Paths.\" International Conference on Learning Representations. 2019.\n\n[3] Langlois, Eric, et al. \"Benchmarking model-based reinforcement learning.\" arXiv preprint arXiv:1907.02057 (2019).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Model-based RL on top of actor-critic method", "review": "### Summary\nThis paper focuses on the field of off-policy reinforcement learning. Specifically, the authors propose a model-based reinforcement learning method on top of actor-critic methods. The proposed method trains a dynamics model and a reward function on the off-policy data with supervised learning, and then uses the trained model to generate synthetic future states and rewards during the actor update. During policy update for a given state, the method computes the sum of the Q value estimate of the state and the Q value expansion for a few steps using the learned dynamics model and reward function.\n\nThe authors implement the proposed method on top of SAC, and TD3, and evaluate their performances on several MuJoCo and Box2D environments. The experiment results show that the proposed method outperforms the model-free baseline in terms of sample efficiency.\n\n### Comments\nThe paper is well written and the idea proposed in this paper is really easy to understand. The authors also include a wide suite of experiments to demonstrate the sample efficiency of the proposed method. Despite these advantages, I cannot recommend acceptance of this paper due to the lack of novelty and absence of fair baseline comparison, which I will elaborate on next.\n\nFirst of all, despite the title of the paper, the proposed method is really a model-based reinforcement learning method since the system network and reward network are just dynamics model and reward model. The proposed objective for the policy (eqn 2), is merely a sum of current Q value estimate and Q value expansion for a few steps using the learned model, which has been proposed before in various papers such as [1] and [2]. The only difference is that when computing the gradient with respect to the policy, the authors leave out the gradient that passes through the learned model, which results in biased estimate of the policy gradient. Therefore, I’m not convinced about the novelty of the proposed method.\n\nMoreover, while proposing a model-based method. The authors do not include baseline comparisons with other model-based RL methods. It is widely known that on low-dimensional control tasks, model-based method outperforms model-free methods ([3]), and therefore merely comparing to model-free baselines is unfair. It would be important to include comparisons to model based methods ([3]).\n\nDue to the lack of novelty and fair comparison to existing model-based methods, I cannot recommend acceptance for this paper.\n\n\n\nReferences\n\n[1] Heess, Nicolas, et al. \"Learning continuous control policies by stochastic value gradients.\" Advances in Neural Information Processing Systems. 2015.\n\n[2] Clavera, Ignasi, Yao Fu, and Pieter Abbeel. \"Model-Augmented Actor-Critic: Backpropagating through Paths.\" International Conference on Learning Representations. 2019.\n\n[3] Langlois, Eric, et al. \"Benchmarking model-based reinforcement learning.\" arXiv preprint arXiv:1907.02057 (2019).\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603944299116}, {"id": "lpb-7rl2ND2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper653/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThe authors proposed a simple modification to popular off policy algorithms, such as SAC and TD3. By employing a model network and reward network, the authors can expand the bellman update operation using predicted next few steps, similar to the GAE (general advantage estimation). The authors demonstrated that algorithms such as TD3, DDPG, and SAC can all benefit from their approach. \n\n\nPros:\nThe paper is clearly written and easy to understand.\n\nThe concept is simple and the implementation is straight forward. The results indicated that the training sample efficiency and final policy performance of the tested algorithms have improved for 6 benchmark tasks, compared with their \"vanilla\" version baseline. \n\n\nCons:\nWhile it is generally understandable that GAE-like approaches can help balance between bias and variance in the Q-value estimation, I am not convinced if the proposed networks are needed. The authors used two additional networks: the model network (they call \"system network\") and the reward network. The model network computes the standard transition s_t, a_t -> s_t+1, and the reward network estimates the true reward r = R(s_t, a_t, s_t+1). However, as reward function is generally provided, I don't see an additional reward network is necessary here. Second, in off-policy learning, the state transition and next states are already known, instead of using a model network to predict the next states, one can simply sample a small trajectory containing multiple consecutive state-action transitions, and use them to do the Q-value estimation. \n\nRecommendations:\nTo address the concern above, I propose the authors to add two more ablation studies: (1) Remove the reward network and only use the reward function. (2) Remove both the reward network and the model/system network, and use the recorded future states to estimate the Q-values. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The author uses a model based approach to improve the bias in Q network training", "review": "Summary:\n\nThe authors proposed a simple modification to popular off policy algorithms, such as SAC and TD3. By employing a model network and reward network, the authors can expand the bellman update operation using predicted next few steps, similar to the GAE (general advantage estimation). The authors demonstrated that algorithms such as TD3, DDPG, and SAC can all benefit from their approach. \n\n\nPros:\nThe paper is clearly written and easy to understand.\n\nThe concept is simple and the implementation is straight forward. The results indicated that the training sample efficiency and final policy performance of the tested algorithms have improved for 6 benchmark tasks, compared with their \"vanilla\" version baseline. \n\n\nCons:\nWhile it is generally understandable that GAE-like approaches can help balance between bias and variance in the Q-value estimation, I am not convinced if the proposed networks are needed. The authors used two additional networks: the model network (they call \"system network\") and the reward network. The model network computes the standard transition s_t, a_t -> s_t+1, and the reward network estimates the true reward r = R(s_t, a_t, s_t+1). However, as reward function is generally provided, I don't see an additional reward network is necessary here. Second, in off-policy learning, the state transition and next states are already known, instead of using a model network to predict the next states, one can simply sample a small trajectory containing multiple consecutive state-action transitions, and use them to do the Q-value estimation. \n\nRecommendations:\nTo address the concern above, I propose the authors to add two more ablation studies: (1) Remove the reward network and only use the reward function. (2) Remove both the reward network and the model/system network, and use the recorded future states to estimate the Q-values. \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603938535659}, {"id": "LCIhBQg_UNS", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper653/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Correctness Issue**: The loss in equation 2 is the sum of the regular actor loss using the critic + an n-step return version of policy value, which would be reasonable for a policy update. However, the expression provided for the gradient (and the calculations in the code) are incorrect in that they are not the gradient of equation 2 with respect to the policy. In particular, it is not reflective of policy performance, as it fails to account for the fact that the future states in the imagined rollout are also functions of the policy. \nThe resulting policy update in FORK actually consists of:\n1. Regular actor update using the critic at the current state\n2. Updating to greedily maximize the reward at each intermediate timestep (which does not reflect the policy performance)\n3. Another regular actor update from the last state in the rollout.\n\nThe straightforward way of correcting this to make it optimize the loss would be to simply differentiate through the model, which may not work well as differentiating through learned models often leads to instability (though it may be fine for such short rollouts). Alternatively, a REINFORCE estimator can instead be used for the gradient of the n-step term for stochastic policies. \n\nWe note that the loss in eq. 6 (if we keep the dependence of the future state on the policy) is also not reflective of policy returns (since it overweights future returns over the present), but can provide reasonable interpretation of the FORK update if we drop the dependence of the future state on the policy as the authors do. \nThe gradient of the FORK-Q variant the authors take from eq.6 is the actually the update we would obtain by using the regular actor update (the $\\nabla_{\\theta} \\pi_{\\theta}(s) Q(s, \\pi(s))$ term for deterministic policies) but under a different state distribution. Instead of using the state distribution in the replay buffer, the update takes into account the distribution obtained by running the policy for a few steps starting from the buffer distribution. As running the policy from the buffer distribution would result in a distribution closer to the on-policy state distribution, one explanation for why the FORK-Q update would improve is that the gradient is closer to the on-policy policy gradients by changing the state distribution. \n\nThe same reasoning can be applied to primary FORK update presented, as it includes an actor update on the value in future states (3rd term in my previous list). The issue is with the greedy reward maximization in the inner steps (2nd term), but maybe it simply doesn't hurt on the environments tested or it perhaps takes advantage of a bias-variance tradeoff in greedily maximizing immediate reward for a few intermediate timesteps. I would like the authors to explicitly address these issues in the paper and present a clear explanation of why their FORK should be a better policy update.\n\n**Relation with Model Based RL Methods**: Overall, I also disagree with the authors' claims that FORK is very different and much simpler than other model based algorithms. With regards to their comment on how their method is somehow simpler? than rollout based methods, their policy update is using a short Monte-Carlo rollout for estimating the policy gradients; the difference being that the rollouts are being used to update a policy rather than to explicitly plan at test time. \n\nThe authors' claim that FORK does not require high-fidelity simulation seems unsupported to me, and the claim that model-based RL algorithms use the model in a sophisticated way is vague. In particular, Dyna-style algorithms (like STEVE https://arxiv.org/abs/1807.01675, MBPO https://arxiv.org/abs/1906.08253) which use the model to generate experience to help learn the critic, seem to be using the model in the same way as FORK (in the sense that they only use the model to generate samples with short rollouts). There is also no discussion of methods like ME-TRPO https://arxiv.org/abs/1802.10592, SLBO https://arxiv.org/abs/1807.03858, or the algorithms in https://arxiv.org/abs/2004.07804, all of which only the model to generate trajectories to use with a policy gradient algorithm. Overall, I would appreciate much more discussion about how FORK relates to and compares empirically against past RL algorithms that only use the model to generate samples, as well as clarifying the statements about FORK uses the model in a less sophisticated way.\n\nOn a separate note, using the model to generate n-step-return estimates of the policy value has been previously done in https://arxiv.org/abs/1807.01675 for example. The key difference is that prior work used it to generate target values for learning the Q-function, while here it is used purely for policy updates. Given how similarly the models are used however, I would recommend discussing this line of work explicitly in the related work, even though they are complementary. \n\n**Experimental Evaluation**: Despite the aforementioned correctness issue, the method seems to provide improvements when applied to TD3, and seemingly smaller improvements on top of SAC. However, I find it extremely strange that the authors chose in Figure 4 to plot returns against the number of training steps instead of the number of samples. As acknowledged by the authors, this makes the SAC vs TD3 comparisons incomparable, with TD3 and TD3-FORk enjoying the advantage of having seen twice as many samples. Moreover, comparing only on the number of actor/critic updates isn't even a fair comparison between FORK and baselines, as FORK additionally has to train a dynamics model and reward predictor. I would highly recommend simply showing learning curves with respect to the actual number of environment samples, rather than arbitrarily using the number of actor updates.\n\nI would also like to see comparisons against model based RL baselines, particularly MBPO, which uses a Dyna-style update with SAC to compare which method of utilizing the model is better. The authors could also see if the FORK actor update further improves upon MBPO or other model based RL methods.\n\n**Summary**:  As it stands, I believe the paper should be rejected due to the correctness issues (and resulting lack of justification for why FORK should give better actor updates), and insufficient discussion of how it relates to prior in model based RL. To consider accepting the paper, I would at least need to see these issues addressed by the authors.\n\nRegarding novelty and significance, using a model to predict n-step value estimates (as the paper claims to be doing) or using the model to explicitly adjust the state distribution of the policy update (as I suspect this might be doing instead) for the policy update in an off-policy actor critic algorithm has not been done before as far as I know. However, this change in how the model is used seems fairly minor, and to be convinced it were useful, I would like to see evidence of how it compares against the other model based RL algorithms. In particular, the benefit I imagine it might have over other model based RL methods is in being more robust to poorly fit models, but I would need to see empirical evidence supporting this.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Incorrect algorithm and insufficient discussion of other model based RL algorithms", "review": "**Correctness Issue**: The loss in equation 2 is the sum of the regular actor loss using the critic + an n-step return version of policy value, which would be reasonable for a policy update. However, the expression provided for the gradient (and the calculations in the code) are incorrect in that they are not the gradient of equation 2 with respect to the policy. In particular, it is not reflective of policy performance, as it fails to account for the fact that the future states in the imagined rollout are also functions of the policy. \nThe resulting policy update in FORK actually consists of:\n1. Regular actor update using the critic at the current state\n2. Updating to greedily maximize the reward at each intermediate timestep (which does not reflect the policy performance)\n3. Another regular actor update from the last state in the rollout.\n\nThe straightforward way of correcting this to make it optimize the loss would be to simply differentiate through the model, which may not work well as differentiating through learned models often leads to instability (though it may be fine for such short rollouts). Alternatively, a REINFORCE estimator can instead be used for the gradient of the n-step term for stochastic policies. \n\nWe note that the loss in eq. 6 (if we keep the dependence of the future state on the policy) is also not reflective of policy returns (since it overweights future returns over the present), but can provide reasonable interpretation of the FORK update if we drop the dependence of the future state on the policy as the authors do. \nThe gradient of the FORK-Q variant the authors take from eq.6 is the actually the update we would obtain by using the regular actor update (the $\\nabla_{\\theta} \\pi_{\\theta}(s) Q(s, \\pi(s))$ term for deterministic policies) but under a different state distribution. Instead of using the state distribution in the replay buffer, the update takes into account the distribution obtained by running the policy for a few steps starting from the buffer distribution. As running the policy from the buffer distribution would result in a distribution closer to the on-policy state distribution, one explanation for why the FORK-Q update would improve is that the gradient is closer to the on-policy policy gradients by changing the state distribution. \n\nThe same reasoning can be applied to primary FORK update presented, as it includes an actor update on the value in future states (3rd term in my previous list). The issue is with the greedy reward maximization in the inner steps (2nd term), but maybe it simply doesn't hurt on the environments tested or it perhaps takes advantage of a bias-variance tradeoff in greedily maximizing immediate reward for a few intermediate timesteps. I would like the authors to explicitly address these issues in the paper and present a clear explanation of why their FORK should be a better policy update.\n\n**Relation with Model Based RL Methods**: Overall, I also disagree with the authors' claims that FORK is very different and much simpler than other model based algorithms. With regards to their comment on how their method is somehow simpler? than rollout based methods, their policy update is using a short Monte-Carlo rollout for estimating the policy gradients; the difference being that the rollouts are being used to update a policy rather than to explicitly plan at test time. \n\nThe authors' claim that FORK does not require high-fidelity simulation seems unsupported to me, and the claim that model-based RL algorithms use the model in a sophisticated way is vague. In particular, Dyna-style algorithms (like STEVE https://arxiv.org/abs/1807.01675, MBPO https://arxiv.org/abs/1906.08253) which use the model to generate experience to help learn the critic, seem to be using the model in the same way as FORK (in the sense that they only use the model to generate samples with short rollouts). There is also no discussion of methods like ME-TRPO https://arxiv.org/abs/1802.10592, SLBO https://arxiv.org/abs/1807.03858, or the algorithms in https://arxiv.org/abs/2004.07804, all of which only the model to generate trajectories to use with a policy gradient algorithm. Overall, I would appreciate much more discussion about how FORK relates to and compares empirically against past RL algorithms that only use the model to generate samples, as well as clarifying the statements about FORK uses the model in a less sophisticated way.\n\nOn a separate note, using the model to generate n-step-return estimates of the policy value has been previously done in https://arxiv.org/abs/1807.01675 for example. The key difference is that prior work used it to generate target values for learning the Q-function, while here it is used purely for policy updates. Given how similarly the models are used however, I would recommend discussing this line of work explicitly in the related work, even though they are complementary. \n\n**Experimental Evaluation**: Despite the aforementioned correctness issue, the method seems to provide improvements when applied to TD3, and seemingly smaller improvements on top of SAC. However, I find it extremely strange that the authors chose in Figure 4 to plot returns against the number of training steps instead of the number of samples. As acknowledged by the authors, this makes the SAC vs TD3 comparisons incomparable, with TD3 and TD3-FORk enjoying the advantage of having seen twice as many samples. Moreover, comparing only on the number of actor/critic updates isn't even a fair comparison between FORK and baselines, as FORK additionally has to train a dynamics model and reward predictor. I would highly recommend simply showing learning curves with respect to the actual number of environment samples, rather than arbitrarily using the number of actor updates.\n\nI would also like to see comparisons against model based RL baselines, particularly MBPO, which uses a Dyna-style update with SAC to compare which method of utilizing the model is better. The authors could also see if the FORK actor update further improves upon MBPO or other model based RL methods.\n\n**Summary**:  As it stands, I believe the paper should be rejected due to the correctness issues (and resulting lack of justification for why FORK should give better actor updates), and insufficient discussion of how it relates to prior in model based RL. To consider accepting the paper, I would at least need to see these issues addressed by the authors.\n\nRegarding novelty and significance, using a model to predict n-step value estimates (as the paper claims to be doing) or using the model to explicitly adjust the state distribution of the policy update (as I suspect this might be doing instead) for the policy update in an off-policy actor critic algorithm has not been done before as far as I know. However, this change in how the model is used seems fairly minor, and to be convinced it were useful, I would like to see evidence of how it compares against the other model based RL algorithms. In particular, the benefit I imagine it might have over other model based RL methods is in being more robust to poorly fit models, but I would need to see empirical evidence supporting this.\n\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603838411289}, {"id": "E1WWYvF7wvh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper653/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to combine ideas from model-based RL into model-free off-policy policy gradient algorithms (like SAC and TD3). Specifically, the paper proposes to learn auxiliary models of environment rewards and dynamics and use a two-step rollout from these models during the computation of the policy gradient. The paper presents results of this mechanism applied to standard SAC and TD3 implementations on a variety of continuous control environments with favorable results.\n\nStrengths:\n\n-- The story is generally easy to follow.\n\n-- I appreciate that the authors didn't just evaluate on the extremely common (and somewhat saturated) MuJoCo benchmarks, and presented additional results on Bipedal Walker.\n\n-- As far as I can tell, the policy update proposed here is novel.\n\nWeaknesses:\n\n-- While the policy update appears novel, it is very similar to techniques in the model-based RL literature. Given this, it would greatly improve the paper if it had appropriate comparisons to similar model-based techniques, especially those which claim to combine model-free updates with model-based techniques; for example https://arxiv.org/abs/1906.08253\n\n-- Moreover, experimentally it would be nice to see comparisons to state-of-the-art model based RL methods. For example, https://arxiv.org/abs/1802.10592\n\n-- In terms of the current experiment results, I found the conclusions favorable to the proposed technique, but not a very compelling demonstration. For example, in Table 1, almost all the environments produce only a *slight* benefit for the proposed method. It appears the only significant benefit is on Ant. Similarly in Table 2, we see the results of SAC-FORK are sometimes much worse than SAC on its own.\n\n-- In terms of motivation for the method, I was not entirely convinced of why the proposed update is needed. The paper appeals to the idea of needing to reason about values in the future. But shouldn't the Q-value already encapsulate this? Moreover, the proposed update ends up only optimizing *actions* in the future, rather than somehow reasoning about the values at steps t+1, t+2 to decide the best action at step t.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The paper proposes to combine ideas from model-based RL into model-free off-policy policy gradient algorithms (like SAC and TD3). Specifically, the paper proposes to learn auxiliary models of environment rewards and dynamics and use a two-step rollout from these models during the computation of the policy gradient. The paper presents results of this mechanism applied to standard SAC and TD3 implementations on a variety of continuous control environments with favorable results.\n\nStrengths:\n\n-- The story is generally easy to follow.\n\n-- I appreciate that the authors didn't just evaluate on the extremely common (and somewhat saturated) MuJoCo benchmarks, and presented additional results on Bipedal Walker.\n\n-- As far as I can tell, the policy update proposed here is novel.\n\nWeaknesses:\n\n-- While the policy update appears novel, it is very similar to techniques in the model-based RL literature. Given this, it would greatly improve the paper if it had appropriate comparisons to similar model-based techniques, especially those which claim to combine model-free updates with model-based techniques; for example https://arxiv.org/abs/1906.08253\n\n-- Moreover, experimentally it would be nice to see comparisons to state-of-the-art model based RL methods. For example, https://arxiv.org/abs/1802.10592\n\n-- In terms of the current experiment results, I found the conclusions favorable to the proposed technique, but not a very compelling demonstration. For example, in Table 1, almost all the environments produce only a *slight* benefit for the proposed method. It appears the only significant benefit is on Ant. Similarly in Table 2, we see the results of SAC-FORK are sometimes much worse than SAC on its own.\n\n-- In terms of motivation for the method, I was not entirely convinced of why the proposed update is needed. The paper appeals to the idea of needing to reason about values in the future. But shouldn't the Q-value already encapsulate this? Moreover, the proposed update ends up only optimizing *actions* in the future, rather than somehow reasoning about the values at steps t+1, t+2 to decide the best action at step t.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603640287473}], "openreview_url": "https://openreview.net/forum?id=lXW6Sk1075v", "arxiv_id": "2010.01652", "paper_pdf": "papers/lXW6Sk1075v.pdf", "paper_pdf_sha256": "6fc657cc12ab0ee13bf84ea5b096427c3afd5fb9b81b8ee61e136248c9601e5b", "paper_pdf_bytes": 1937556, "paper_pdf_source": "openreview", "code_url": "https://github.com/honghaow/FORK", "code_repository": "honghaow/FORK", "code_commit": "12e002d24ba7e5718e34dd20d4888a1dae1b4774", "code_archive": "repos/lXW6Sk1075v.zip", "code_archive_sha256": "06e0ff9896c612157241912b6b5f45084ce2449b3a014cc4e8bfc4dcc6a6ab4c", "code_archive_bytes": 750821, "code_file_count": 15, "code_extensions": {".py": 10, ".sh": 4, ".ipynb": 1}, "github_disk_usage_kb": 991, "github_languages": {"Jupyter Notebook": 705939, "Python": 61936, "Shell": 4444}, "github_archived": false, "github_pushed_at": "2022-03-31T23:32:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fork-a-forward-looking-actor-for-model-free-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJlk-eHFwH", "year": 2020, "status": "rejected", "title": "AdaGAN: Adaptive GAN for Many-to-Many Non-Parallel Voice Conversion", "authors": ["Maitreya Patel", "Mirali Purohit", "Mihir Parmar", "Nirmesh J. Shah", "Hemant A. Patil"], "authorids": ["maitreya_patel@daiict.ac.in", "purohit_mirali@daiict.ac.in", "mihirparmar@asu.edu", "nirmesh88_shah@daiict.ac.in", "hemant_patil@daiict.ac.in"], "authors_source": "OpenReview API", "abstract": "Voice Conversion (VC) is a task of converting perceived speaker identity from a source speaker to a particular target speaker. Earlier approaches in the literature primarily find a mapping between the given source-target speaker-pairs. Developing mapping techniques for many-to-many VC using non-parallel data, including zero-shot learning remains less explored areas in VC. Most of the many-to-many VC architectures require training data from all the target speakers for whom we want to convert the voices. In this paper, we propose a novel style transfer architecture, which can also be extended to generate voices even for target speakers whose data were not used in the training (i.e., case of zero-shot learning). In particular, propose Adaptive Generative Adversarial Network (AdaGAN), new architectural training procedure help in learning normalized speaker-independent latent representation, which will be used to generate speech with different speaking styles in the context of VC. We compare our results with the state-of-the-art StarGAN-VC architecture. In particular, the AdaGAN achieves 31.73%, and 10.37% relative improvement compared to the StarGAN in MOS tests for speech quality and speaker similarity, respectively. The key strength of the proposed architectures is that it yields these results with less computational complexity. AdaGAN is 88.6% less complex than StarGAN-VC in terms of FLoating Operation Per Second (FLOPS), and 85.46% less complex in terms of trainable parameters.  ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "r1xjn_YV5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2120/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work describes an efficient voice conversion system that can operate on non-parallel samples and convert from and to multiple voices.  The central element of the methodology is the AdaIn modification.  This is an efficient speaker adaptive technique where features are re-normalized to a particular speaker's domain.  The rest of the machinery is well motivated and well executed, but less novel.  This addition enables the voice conversion between speakers.\n\nSection 4.4 Are all of the utterances the same length?  Based on the architecture description, it appears as though the model generates one output frame for each input frame.  This would suggest that for training input and output need to be synchronized.  If so, make this explicit and include length parameters in Section 6.1 \n\nSection 6.2 states \"For statistically significant analysis, results are shown in different conversion possibilities.\" However, no test of statistical significance is presented.  This pointer may be helpful (https://pdfs.semanticscholar.org/b2b1/d01336323f3794f54de26567335aa0bcac46.pdf)\n\nPresentation Comments:\n\nSection 3.1: I would recommend using different subscripts for Z_i and U_i, since when indexing Z this implies the i-th speaker, and when indexing U it's the i-th utterance.  The formulas in Section 3.1 imply a single index i for both of these which is clearly not intended.\n\nSection 4: Consider using the present tense instead of the perfect tense when describing the results.  \"...we discuss our proposed AdaGAN architecture... We have shown... We have presented...\" can be \"...we discuss our proposed AdaGAN architecture... We show... We present...\"\n\nSection 5.2; Tables 1 and 2: Consider some partition of the FLOPS and Parameters, separation by commas, spaces or even abbreviation e.g. 2952233 -> 2,952,233 or 2 952 233 or 2.9M.  This will make this table much easier to read.\n\nSection 6.2; Figures 2-5: MOS scores have a minimum value of 1.  This should be the axis of the chart, rather than 0.  \n\nIt's pretty bold to star by contextualizing the work with the sentence \"Language is the core of civilization and speech is the most powerful and natural form of communication.\"  :-)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This work describes an efficient voice conversion system that can operate on non-parallel samples and convert from and to multiple voices.  The central element of the methodology is the AdaIn modification.  This is an efficient speaker adaptive technique where features are re-normalized to a particular speaker's domain.  The rest of the machinery is well motivated and well executed, but less novel.  This addition enables the voice conversion between speakers.\n\nSection 4.4 Are all of the utterances the same length?  Based on the architecture description, it appears as though the model generates one output frame for each input frame.  This would suggest that for training input and output need to be synchronized.  If so, make this explicit and include length parameters in Section 6.1 \n\nSection 6.2 states \"For statistically significant analysis, results are shown in different conversion possibilities.\" However, no test of statistical significance is presented.  This pointer may be helpful (https://pdfs.semanticscholar.org/b2b1/d01336323f3794f54de26567335aa0bcac46.pdf)\n\nPresentation Comments:\n\nSection 3.1: I would recommend using different subscripts for Z_i and U_i, since when indexing Z this implies the i-th speaker, and when indexing U it's the i-th utterance.  The formulas in Section 3.1 imply a single index i for both of these which is clearly not intended.\n\nSection 4: Consider using the present tense instead of the perfect tense when describing the results.  \"...we discuss our proposed AdaGAN architecture... We have shown... We have presented...\" can be \"...we discuss our proposed AdaGAN architecture... We show... We present...\"\n\nSection 5.2; Tables 1 and 2: Consider some partition of the FLOPS and Parameters, separation by commas, spaces or even abbreviation e.g. 2952233 -> 2,952,233 or 2 952 233 or 2.9M.  This will make this table much easier to read.\n\nSection 6.2; Figures 2-5: MOS scores have a minimum value of 1.  This should be the axis of the chart, rather than 0.  \n\nIt's pretty bold to star by contextualizing the work with the sentence \"Language is the core of civilization and speech is the most powerful and natural form of communication.\"  :-)"}, "tcdate": 1572276402743}, {"id": "BkgNckrRtH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2120/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tackles many-to-many voice conversion task using GAN for style transfer between different speakers. The core idea is adaptive instance normalization (Huang & Belongie, 2017). \n\nDetailed comments:\n\nThere are many typos and wrong notations in the text. Here is an incomplete list:\n- \"it were spoken by target speaker\", should be \"was\". \n- In Section 3.1, \"Here, U_1 and U_2 are spoken by Z_i and Z_2\" should Z_1. Overall, the descriptions in this subsection is confusing. For example, it seems utterance U_i is from speaker Z_i in the dataset, but there are n speakers and m utterances.\n\n- A closely related task is voice cloning, which is arguably more challenging than voice conversion, because the synthesis need generalize to arbitrary text. One may properly discuss the recent advances in this community (e.g., Arik et al., 2018; Nachmani et al., 2018; Jia et al., 2018).\n\nPros:\nThe empirical improvement seems meaningful.\n\nCons:\nThis paper is poorly written and difficult to follow. For example, I could not accurately identify the major contribution & novelty after reading the abstract and introduction. As an application paper, the authors may clearly explain the ideas with a few sentences in the most natural way without \"heavy notations\", e.g., Eq. (5)(6)(7).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This paper tackles many-to-many voice conversion task using GAN for style transfer between different speakers. The core idea is adaptive instance normalization (Huang & Belongie, 2017). \n\nDetailed comments:\n\nThere are many typos and wrong notations in the text. Here is an incomplete list:\n- \"it were spoken by target speaker\", should be \"was\". \n- In Section 3.1, \"Here, U_1 and U_2 are spoken by Z_i and Z_2\" should Z_1. Overall, the descriptions in this subsection is confusing. For example, it seems utterance U_i is from speaker Z_i in the dataset, but there are n speakers and m utterances.\n\n- A closely related task is voice cloning, which is arguably more challenging than voice conversion, because the synthesis need generalize to arbitrary text. One may properly discuss the recent advances in this community (e.g., Arik et al., 2018; Nachmani et al., 2018; Jia et al., 2018).\n\nPros:\nThe empirical improvement seems meaningful.\n\nCons:\nThis paper is poorly written and difficult to follow. For example, I could not accurately identify the major contribution & novelty after reading the abstract and introduction. As an application paper, the authors may clearly explain the ideas with a few sentences in the most natural way without \"heavy notations\", e.g., Eq. (5)(6)(7)."}, "tcdate": 1571864460218}, {"id": "SJxDQYG0FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2120/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a voice conversion approach using GANs based on adaptive instance normalization (AdaIN).  The authors give the mathematical formulation of the problem and provide the implementation of the so-called AdaGAN. Experiments are carried out on VCTK and the proposed AdaGAN is compared with StarGAN.  The idea is ok and the concept of using AdaIN for efficient voice conversion is also good.  But the paper has a lot of issues both technically and grammatically, which makes the paper hard to follow.\n\n1. On writing\nThere are glaring grammar errors in numerous places. e.g.\n  -- \"Although, there are few GAN-based systems that produced state-of-the-art results for non-parallel VC. Among these algorithms, even fewer can be applied for many-to-many VC task. At last, there is the only system available for zero-shot VC proposed by Qian et al. (2019).\"   This is hard to parse.\n -- \"helps generator to make ...\"  -> \"helps the generator make ...\"\n --  \"let assume\" -> \"Let's assume\" \n --  \"We know that the idea of transitivity as a way to regularize structured data has a long history.\"   what does it mean?\n --  \"the generator of AdaGAN is consists of Encoder and Decoder.\"  -> \"consist of\"\n -- \"After training of AdaGAN for large number of iteration of $\\tau$ , where theoretically $\\tau \\rightarrow \\infty$.\" where is the second half of the sentence?\n\n2.  On math notation\n The math notation is messy and there are lots of inaccuracies.  e.g.\n  --  $X_{i} \\in p_{X}(\\cdot|Z_{i},U_{i})$ should be $X_{i} \\sim p_{X}(\\cdot|Z_{i},U_{i})$\n  --  \"generate the distribution denoted by $\\hat{X}_{Z_{1}\\rightarrow Z_{2}}$\"  -> why  $\\hat{X}_{Z_{1}\\rightarrow Z_{2}}$ becomes a distribution? \n  --  \"$p_{N}(\\cdot|Z_{1},U_{1})$, $p_{N}(\\cdot|Z_{2},U_{1})$\" in Eq.14,  $N$ should be replaced by the random variable.\n  --  $S'_{X}$ and $S'_{Y}$ should be $S_{X'}$ and $S_{Y'}$ in line 15 in the algorithm\n\n3. On technical details:\n -- In Fig.1 (b), why is there only one input to the discriminator?  How do you inject the adversarial samples and how do you generate adversarial samples? \n-- In section 4.4, \"in encoder and decoder all layers are Linear layers\".  Are you referring to fully-connected layers? Linear layers are usually referred to those with linear activation functions.  \n-- The experiments are claimed to be zero-shot, but 3-5s of speech is required.  can you explain? \n\nAlthough the samples sound OK, given its current form, the paper needs significant re-work. \n\nP.S. rebuttal read.   I will stay with my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "1: Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "This paper presents a voice conversion approach using GANs based on adaptive instance normalization (AdaIN).  The authors give the mathematical formulation of the problem and provide the implementation of the so-called AdaGAN. Experiments are carried out on VCTK and the proposed AdaGAN is compared with StarGAN.  The idea is ok and the concept of using AdaIN for efficient voice conversion is also good.  But the paper has a lot of issues both technically and grammatically, which makes the paper hard to follow.\n\n1. On writing\nThere are glaring grammar errors in numerous places. e.g.\n  -- \"Although, there are few GAN-based systems that produced state-of-the-art results for non-parallel VC. Among these algorithms, even fewer can be applied for many-to-many VC task. At last, there is the only system available for zero-shot VC proposed by Qian et al. (2019).\"   This is hard to parse.\n -- \"helps generator to make ...\"  -> \"helps the generator make ...\"\n --  \"let assume\" -> \"Let's assume\" \n --  \"We know that the idea of transitivity as a way to regularize structured data has a long history.\"   what does it mean?\n --  \"the generator of AdaGAN is consists of Encoder and Decoder.\"  -> \"consist of\"\n -- \"After training of AdaGAN for large number of iteration of $\\tau$ , where theoretically $\\tau \\rightarrow \\infty$.\" where is the second half of the sentence?\n\n2.  On math notation\n The math notation is messy and there are lots of inaccuracies.  e.g.\n  --  $X_{i} \\in p_{X}(\\cdot|Z_{i},U_{i})$ should be $X_{i} \\sim p_{X}(\\cdot|Z_{i},U_{i})$\n  --  \"generate the distribution denoted by $\\hat{X}_{Z_{1}\\rightarrow Z_{2}}$\"  -> why  $\\hat{X}_{Z_{1}\\rightarrow Z_{2}}$ becomes a distribution? \n  --  \"$p_{N}(\\cdot|Z_{1},U_{1})$, $p_{N}(\\cdot|Z_{2},U_{1})$\" in Eq.14,  $N$ should be replaced by the random variable.\n  --  $S'_{X}$ and $S'_{Y}$ should be $S_{X'}$ and $S_{Y'}$ in line 15 in the algorithm\n\n3. On technical details:\n -- In Fig.1 (b), why is there only one input to the discriminator?  How do you inject the adversarial samples and how do you generate adversarial samples? \n-- In section 4.4, \"in encoder and decoder all layers are Linear layers\".  Are you referring to fully-connected layers? Linear layers are usually referred to those with linear activation functions.  \n-- The experiments are claimed to be zero-shot, but 3-5s of speech is required.  can you explain? \n\nAlthough the samples sound OK, given its current form, the paper needs significant re-work. \n\nP.S. rebuttal read.   I will stay with my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571854623244}], "openreview_url": "https://openreview.net/forum?id=HJlk-eHFwH", "arxiv_id": null, "paper_pdf": "papers/HJlk-eHFwH.pdf", "paper_pdf_sha256": "7bd4784b8a86d45ef4760cc198c7212a0cf5b9eba7065139ae7175b833b19ff1", "paper_pdf_bytes": 629151, "paper_pdf_source": "openreview", "code_url": "https://github.com/liusongxiang/StarGAN-Voice-Conversion", "code_repository": "liusongxiang/StarGAN-Voice-Conversion", "code_commit": "a4633d8b4888e0fa77d6b67f9d94da603da2af31", "code_archive": "repos/HJlk-eHFwH.zip", "code_archive_sha256": "7c0feefee240c81972d7964c1ac5fa5293cc571f5c3d8cc37b8b9b8608f7966a", "code_archive_bytes": 2760494, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 2702, "github_languages": {"Python": 50459}, "github_archived": false, "github_pushed_at": "2019-10-11T02:49:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adagan-adaptive-gan-for-many-to-many-non"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2XP4HD0NBy", "year": 2026, "status": "rejected", "title": "SegRet: An Efficient Design for Semantic Segmentation with Retentive Network", "authors": ["Zhiyuan Li", "Yi Chang", "Yuan Wu"], "authorids": ["~Zhiyuan_Li19", "~Yi_Chang4", "~Yuan_Wu2"], "authors_source": "OpenReview API", "abstract": "With the rapid evolution of autonomous driving technology and intelligent transportation systems, semantic segmentation has become increasingly critical. Precise interpretation and analysis of real-world environments are indispensable for these advanced applications. However, traditional semantic segmentation approaches frequently face challenges in balancing model performance with computational efficiency, especially regarding the volume of model parameters. To address these constraints, we propose SegRet, a novel model employing the Retentive Network (RetNet) architecture coupled with a lightweight residual decoder that integrates zero-initialization. SegRet offers three distinctive advantages: (1) Lightweight Residual Decoder: by embedding a zero-initialization layer within the residual network structure, the decoder remains computationally streamlined without sacrificing essential information propagation; (2) Robust Feature Extraction: adopting RetNet as its backbone enables SegRet to effectively capture hierarchical image features, thereby enriching the representation quality of extracted features; (3) Parameter Efficiency: SegRet attains state-of-the-art (SOTA) segmentation performance while markedly decreasing the number of parameters, ensuring high accuracy without imposing additional computational burdens. Comprehensive empirical evaluations on prominent benchmarks, such as ADE20K, Citycapes, and COCO-Stuff, highlight the effectiveness and superiority of our method.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "vByQEThJLS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16094/Reviewer_eeLP"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "SegRet couples Vision RetNet as a hierarchical encoder with a lightweight zero-initialized residual decoder. It targets parameter efficiency while maintaining competitive mIoU on ADE20K, Cityscapes, and COCO-Stuff. Ablations show a modest but consistent gain from the zero-initialized residual (≈+0.8 mIoU) and comparisons demonstrate favorable accuracy/params trade-offs (e.g., SegRet-Tiny ~14M params reaching ~49.4 mIoU on ADE20K and ~42.2/43.3 mIoU on COCO-Stuff). Code link (anonymous) is provided.", "review_text": "SegRet couples Vision RetNet as a hierarchical encoder with a lightweight zero-initialized residual decoder. It targets parameter efficiency while maintaining competitive mIoU on ADE20K, Cityscapes, and COCO-Stuff. Ablations show a modest but consistent gain from the zero-initialized residual (≈+0.8 mIoU) and comparisons demonstrate favorable accuracy/params trade-offs (e.g., SegRet-Tiny ~14M params reaching ~49.4 mIoU on ADE20K and ~42.2/43.3 mIoU on COCO-Stuff). Code link (anonymous) is provided.", "strengths": "1. This paper achieves a strong efficiency trade-off, particularly at smaller model sizes, with documented improvements over comparable baseline methods.\n\n2. This paper provides transparent ablation studies on decoder design and input scaling, along with clear training details and code to ensure reproducibility.\n\n3.  This paper honestly addresses limitations and suggests plausible next steps, such as domain adaptation and attention-guided upsampling.", "weaknesses": "1. The novelty lies mainly in the minimalist decoder design; the use of RetNet as the encoder and residual fusion represents an incremental improvement rather than a conceptual breakthrough.\n\n2. While results are competitive, they do not clearly establish state-of-the-art performance under similar computational constraints in the most challenging benchmarks. Stronger comparisons using identical training configurations would be beneficial.", "questions": "1. The authors should provide FLOPs and latency comparisons with Mask2Former or SegFormer at the same input size to validate the efficiency claims.\n\n2. Results on small-object segmentation (e.g., Cityscapes fine classes) are needed to evaluate if the lightweight decoder compromises fine detail.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "SegRet couples Vision RetNet as a hierarchical encoder with a lightweight zero-initialized residual decoder. It targets parameter efficiency while maintaining competitive mIoU on ADE20K, Cityscapes, and COCO-Stuff. Ablations show a modest but consistent gain from the zero-initialized residual (≈+0.8 mIoU) and comparisons demonstrate favorable accuracy/params trade-offs (e.g., SegRet-Tiny ~14M params reaching ~49.4 mIoU on ADE20K and ~42.2/43.3 mIoU on COCO-Stuff). Code link (anonymous) is provided.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. This paper achieves a strong efficiency trade-off, particularly at smaller model sizes, with documented improvements over comparable baseline methods.\n\n2. This paper provides transparent ablation studies on decoder design and input scaling, along with clear training details and code to ensure reproducibility.\n\n3.  This paper honestly addresses limitations and suggests plausible next steps, such as domain adaptation and attention-guided upsampling.", "weaknesses": "1. The novelty lies mainly in the minimalist decoder design; the use of RetNet as the encoder and residual fusion represents an incremental improvement rather than a conceptual breakthrough.\n\n2. While results are competitive, they do not clearly establish state-of-the-art performance under similar computational constraints in the most challenging benchmarks. Stronger comparisons using identical training configurations would be beneficial.", "questions": "1. The authors should provide FLOPs and latency comparisons with Mask2Former or SegFormer at the same input size to validate the efficiency claims.\n\n2. Results on small-object segmentation (e.g., Cityscapes fine classes) are needed to evaluate if the lightweight decoder compromises fine detail.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761958377366}, {"id": "I79dIRooQt", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16094/Reviewer_uTUL"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "This paper proposed a semantic segmentation network that leverages retentive network. The network consists of a encoder backbone and a decoder. The encoder backbone contains 4 consecutive RMT blocks. The decoder utilizes a zero initialized convolution layer in parallel with a linear layer. Experiments on 3 benchmark datasets ADE20K, Cityscapes and COCO-Stuff and demonstrate the effectiveness of the proposed method.", "review_text": "This paper proposed a semantic segmentation network that leverages retentive network. The network consists of a encoder backbone and a decoder. The encoder backbone contains 4 consecutive RMT blocks. The decoder utilizes a zero initialized convolution layer in parallel with a linear layer. Experiments on 3 benchmark datasets ADE20K, Cityscapes and COCO-Stuff and demonstrate the effectiveness of the proposed method.", "strengths": "The paper is well organized. The proposed methods achieves SOTA performance with low computational resources.", "weaknesses": "1. Lack of contribution. The proposed method's backbone is almost the same as the compared method RMT, which cannot be viewed as contribution. The improvement in the decoder is also minimum.\n\n2. The compared methods contains mostly general vision backbones, and the result on many other vision tasks, such as object classification and detection, are available. But the proposed method is not comparing with them.\n\n3. The comparison in Tab. 2, 3, 6, 7, 8 is not fair. The image size should keep the same, because the number of parameters, latency and losses are calculated depends on the image size.\n\n4. More qualitative results should be provided, such as more segmentation results from compared methods, feature analysis, etc.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a semantic segmentation network that leverages retentive network. The network consists of a encoder backbone and a decoder. The encoder backbone contains 4 consecutive RMT blocks. The decoder utilizes a zero initialized convolution layer in parallel with a linear layer. Experiments on 3 benchmark datasets ADE20K, Cityscapes and COCO-Stuff and demonstrate the effectiveness of the proposed method.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "The paper is well organized. The proposed methods achieves SOTA performance with low computational resources.", "weaknesses": "1. Lack of contribution. The proposed method's backbone is almost the same as the compared method RMT, which cannot be viewed as contribution. The improvement in the decoder is also minimum.\n\n2. The compared methods contains mostly general vision backbones, and the result on many other vision tasks, such as object classification and detection, are available. But the proposed method is not comparing with them.\n\n3. The comparison in Tab. 2, 3, 6, 7, 8 is not fair. The image size should keep the same, because the number of parameters, latency and losses are calculated depends on the image size.\n\n4. More qualitative results should be provided, such as more segmentation results from compared methods, feature analysis, etc.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761925858699}, {"id": "2GlUHHXu9j", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16094/Reviewer_2JaX"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes SegRet, a new semantic segmentation framework that integrates the recently developed Retentive Network (RetNet) — a state-space model designed to replace self-attention — with a lightweight, zero-initialized residual decoder, while significantly reducing parameter count, making it a compact, efficiency-oriented yet not fully state-of-the-art semantic segmentation framework.\nThis design enables SegRet to achieve competitive segmentation performance across ADE20K, Cityscapes, and COCO-Stuff using significantly fewer parameters and computational resources than traditional Transformer-based models like Mask2Former or Swin, making it particularly well-suited for real-time or resource-constrained vision applications such as autonomous driving and embedded systems.", "review_text": "The paper proposes SegRet, a new semantic segmentation framework that integrates the recently developed Retentive Network (RetNet) — a state-space model designed to replace self-attention — with a lightweight, zero-initialized residual decoder, while significantly reducing parameter count, making it a compact, efficiency-oriented yet not fully state-of-the-art semantic segmentation framework.\nThis design enables SegRet to achieve competitive segmentation performance across ADE20K, Cityscapes, and COCO-Stuff using significantly fewer parameters and computational resources than traditional Transformer-based models like Mask2Former or Swin, making it particularly well-suited for real-time or resource-constrained vision applications such as autonomous driving and embedded systems.", "strengths": "Clean encoder–decoder story using Vision RetNet as a hierarchical backbone (Fig. 2, p. 4; Sec. 3.1) and a small decoder (Sec. 3.2, p. 5–6).\n\nCompetitive tiny/small regime: SegRet-Tiny (≈14 M params) is strong vs. other “tiny” setups across datasets (Tables 1–3, pp. 7–8).\n\nReadable presentation and sensible training protocol (Sec. 4.1, p. 7), with an (anonymous) code link in the Reproducibility Statement (p. 10).\n\nThe principal strength of SegRet lies in its computational efficiency and architectural simplicity—it successfully integrates the Vision RetNet encoder, which captures long-range dependencies with linear complexity, with a lightweight zero-initialized residual decoder that minimizes parameters without degrading accuracy.", "weaknesses": "Vision RetNet is adopted largely as-is; the paper does not contribute new retention variants for vision, nor new theory atop RetNet. The retention machinery and bidirectional vision adaptation (BiRetention, horizontal/vertical decomposition) are recaps of prior work (Sec. 3.1; Eqs. 6–14).\n\nThe “zero-initialized residual decoder” amounts to a zero-init 1×1 residual branch after channel unification, followed by standard upsample-concat-conv (Eqs. 15–18). This design is extremely close to well-known lightweight decoders (FPN-style merges, 1×1 residual adapters), and the paper does not show conceptual novelty beyond the initialization trick.\n\nImpact of the “novelty” is marginal. The only ablation directly tied to the claimed novelty (adding ZIR) shows +0.79 mIoU with +0.25M params on ADE20K for the Small variant (Table 4), which is a small incremental gain that does not justify SOTA claims by itself.\n\nThe core claim is “parameter/efficiency superiority,” yet there is no latency (ms), throughput (FPS), or peak memory on any GPU/CPU, nor profile at multiple resolutions (Sec. 4, pp. 7–9). FLOPs alone do not predict wall-clock; kernel efficiency (e.g., attention vs. state-space primitives) and cache behavior matter. Absent these, “efficient” remains asserted, not demonstrated.\n\n\nProvide a theoretical or empirical rationale for the zero-init residual path: gradients at init, linearization analysis, effect on optimization dynamics; compare with ResNet-style identity and LayerScale variants.", "questions": "1. The “zero-initialized residual” decoder seems to differ from prior FPN/UPerNet-style decoders only by a zero-initialized 1×1 residual branch. Could you provide a formal motivation or derivation—for instance, how zero initialization affects gradient flow, optimization stability, or representational bias compared to standard residual or skip connections?\n\n2. How is SegRet conceptually distinct from known efficient decoders such as SegFormer’s MLP-based fusion, LiteSeg, or MobileViT decoders? Please clarify what architectural element or training principle is genuinely new, rather than a simplified re-implementation.\n\n3. Since The SegRet model is implemented based on the MMSegmentation framework could you highlight key difference between both of them. \n\n4. You state the method underperforms for medical/remote sensing (p. 9; A.6, p. 15). Can you include one focused study (e.g., small-lesion segmentation) with failure analysis to clarify whether the limitation is encoder scale, decoder capacity, or upsampling choice?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes SegRet, a new semantic segmentation framework that integrates the recently developed Retentive Network (RetNet) — a state-space model designed to replace self-attention — with a lightweight, zero-initialized residual decoder, while significantly reducing parameter count, making it a compact, efficiency-oriented yet not fully state-of-the-art semantic segmentation framework.\nThis design enables SegRet to achieve competitive segmentation performance across ADE20K, Cityscapes, and COCO-Stuff using significantly fewer parameters and computational resources than traditional Transformer-based models like Mask2Former or Swin, making it particularly well-suited for real-time or resource-constrained vision applications such as autonomous driving and embedded systems.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Clean encoder–decoder story using Vision RetNet as a hierarchical backbone (Fig. 2, p. 4; Sec. 3.1) and a small decoder (Sec. 3.2, p. 5–6).\n\nCompetitive tiny/small regime: SegRet-Tiny (≈14 M params) is strong vs. other “tiny” setups across datasets (Tables 1–3, pp. 7–8).\n\nReadable presentation and sensible training protocol (Sec. 4.1, p. 7), with an (anonymous) code link in the Reproducibility Statement (p. 10).\n\nThe principal strength of SegRet lies in its computational efficiency and architectural simplicity—it successfully integrates the Vision RetNet encoder, which captures long-range dependencies with linear complexity, with a lightweight zero-initialized residual decoder that minimizes parameters without degrading accuracy.", "weaknesses": "Vision RetNet is adopted largely as-is; the paper does not contribute new retention variants for vision, nor new theory atop RetNet. The retention machinery and bidirectional vision adaptation (BiRetention, horizontal/vertical decomposition) are recaps of prior work (Sec. 3.1; Eqs. 6–14).\n\nThe “zero-initialized residual decoder” amounts to a zero-init 1×1 residual branch after channel unification, followed by standard upsample-concat-conv (Eqs. 15–18). This design is extremely close to well-known lightweight decoders (FPN-style merges, 1×1 residual adapters), and the paper does not show conceptual novelty beyond the initialization trick.\n\nImpact of the “novelty” is marginal. The only ablation directly tied to the claimed novelty (adding ZIR) shows +0.79 mIoU with +0.25M params on ADE20K for the Small variant (Table 4), which is a small incremental gain that does not justify SOTA claims by itself.\n\nThe core claim is “parameter/efficiency superiority,” yet there is no latency (ms), throughput (FPS), or peak memory on any GPU/CPU, nor profile at multiple resolutions (Sec. 4, pp. 7–9). FLOPs alone do not predict wall-clock; kernel efficiency (e.g., attention vs. state-space primitives) and cache behavior matter. Absent these, “efficient” remains asserted, not demonstrated.\n\n\nProvide a theoretical or empirical rationale for the zero-init residual path: gradients at init, linearization analysis, effect on optimization dynamics; compare with ResNet-style identity and LayerScale variants.", "questions": "1. The “zero-initialized residual” decoder seems to differ from prior FPN/UPerNet-style decoders only by a zero-initialized 1×1 residual branch. Could you provide a formal motivation or derivation—for instance, how zero initialization affects gradient flow, optimization stability, or representational bias compared to standard residual or skip connections?\n\n2. How is SegRet conceptually distinct from known efficient decoders such as SegFormer’s MLP-based fusion, LiteSeg, or MobileViT decoders? Please clarify what architectural element or training principle is genuinely new, rather than a simplified re-implementation.\n\n3. Since The SegRet model is implemented based on the MMSegmentation framework could you highlight key difference between both of them. \n\n4. You state the method underperforms for medical/remote sensing (p. 9; A.6, p. 15). Can you include one focused study (e.g., small-lesion segmentation) with failure analysis to clarify whether the limitation is encoder scale, decoder capacity, or upsampling choice?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761859987943}, {"id": "UG4aWCHZRH", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16094/Reviewer_Qwqz"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The paper introduces SegRet, a semantic segmentation model that pairs a RetNet backbone with a lightweight residual decoder using zero-initialization. The design targets the accuracy–efficiency trade-off common in autonomous driving and intelligent transportation. Specifically, the zero-initialized residual decoder keeps computation low while stabilizing information flow, and RetNet provides robust hierarchical feature extraction. The approach emphasizes parameter efficiency, claiming SOTA-level performance with markedly fewer parameters. Experiments on ADE20K, Cityscapes, and COCO-Stuff demonstrate strong accuracy under tight computational budgets, indicating SegRet’s suitability for real-time or resource-constrained deployments.", "review_text": "The paper introduces SegRet, a semantic segmentation model that pairs a RetNet backbone with a lightweight residual decoder using zero-initialization. The design targets the accuracy–efficiency trade-off common in autonomous driving and intelligent transportation. Specifically, the zero-initialized residual decoder keeps computation low while stabilizing information flow, and RetNet provides robust hierarchical feature extraction. The approach emphasizes parameter efficiency, claiming SOTA-level performance with markedly fewer parameters. Experiments on ADE20K, Cityscapes, and COCO-Stuff demonstrate strong accuracy under tight computational budgets, indicating SegRet’s suitability for real-time or resource-constrained deployments.", "strengths": "Balanced accuracy and efficiency: SegRet achieves strong segmentation performance on major benchmarks (ADE20K, Cityscapes, COCO-Stuff) while significantly reducing parameter count and computational cost, demonstrating excellent scalability for real-time or embedded applications.\nStrong feature representation capability: By leveraging RetNet’s long-range dependency modeling, SegRet enhances multi-level feature extraction, resulting in richer contextual understanding and improved segmentation quality.", "weaknesses": "The contribution hinges on pairing RetNet with a zero-initialized lightweight residual decoder, but the paper doesn’t explain why this pairing is fundamentally new vs. existing efficient backbones + residual decoders, or how it differs from prior zero-init/ResNet-style stabilizers.", "questions": "1.The main text does not reference Figure 1. In addition, the meaning of multi-scale inference is unclear, and it should be explained why each method in the comparison has multiple parameter counts.\n2.The Vision RetNet Backbone is an existing work, yet a disproportionate amount of space is devoted to describing it.\n3.In Figure 2, the MASA diagram in the upper-left corner shows only one input, whereas the lower diagram depicts six inputs—this inconsistency should be clarified.\n4.Equations (9–14) are densely stacked, which reduces readability and should be reformatted for clarity.\n5.In Section 4.1, both Tiny and Small variants are reported to have identical parameter counts (0.814 M), which seems implausible and requires verification.\n6.In Sections 4.2 and A.1–A.3, using dataset names as sub-section titles is inappropriate; more descriptive methodological headings are recommended.\n7.In Table 3, the comparative methods are too few, especially recent ones from the past two years.\n8.In Figure 3, the qualitative comparison on the ADE20K dataset includes only one baseline, and noticeable differences remain between the proposed results and the ground truth. Moreover, qualitative comparisons on the other two datasets are missing.\n9.The description of the zero-initialized component is vague—does it consist of only a single convolution operation?\n10.The overall contribution lacks substantial innovation beyond architectural integration.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces SegRet, a semantic segmentation model that pairs a RetNet backbone with a lightweight residual decoder using zero-initialization. The design targets the accuracy–efficiency trade-off common in autonomous driving and intelligent transportation. Specifically, the zero-initialized residual decoder keeps computation low while stabilizing information flow, and RetNet provides robust hierarchical feature extraction. The approach emphasizes parameter efficiency, claiming SOTA-level performance with markedly fewer parameters. Experiments on ADE20K, Cityscapes, and COCO-Stuff demonstrate strong accuracy under tight computational budgets, indicating SegRet’s suitability for real-time or resource-constrained deployments.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "Balanced accuracy and efficiency: SegRet achieves strong segmentation performance on major benchmarks (ADE20K, Cityscapes, COCO-Stuff) while significantly reducing parameter count and computational cost, demonstrating excellent scalability for real-time or embedded applications.\nStrong feature representation capability: By leveraging RetNet’s long-range dependency modeling, SegRet enhances multi-level feature extraction, resulting in richer contextual understanding and improved segmentation quality.", "weaknesses": "The contribution hinges on pairing RetNet with a zero-initialized lightweight residual decoder, but the paper doesn’t explain why this pairing is fundamentally new vs. existing efficient backbones + residual decoders, or how it differs from prior zero-init/ResNet-style stabilizers.", "questions": "1.The main text does not reference Figure 1. In addition, the meaning of multi-scale inference is unclear, and it should be explained why each method in the comparison has multiple parameter counts.\n2.The Vision RetNet Backbone is an existing work, yet a disproportionate amount of space is devoted to describing it.\n3.In Figure 2, the MASA diagram in the upper-left corner shows only one input, whereas the lower diagram depicts six inputs—this inconsistency should be clarified.\n4.Equations (9–14) are densely stacked, which reduces readability and should be reformatted for clarity.\n5.In Section 4.1, both Tiny and Small variants are reported to have identical parameter counts (0.814 M), which seems implausible and requires verification.\n6.In Sections 4.2 and A.1–A.3, using dataset names as sub-section titles is inappropriate; more descriptive methodological headings are recommended.\n7.In Table 3, the comparative methods are too few, especially recent ones from the past two years.\n8.In Figure 3, the qualitative comparison on the ADE20K dataset includes only one baseline, and noticeable differences remain between the proposed results and the ground truth. Moreover, qualitative comparisons on the other two datasets are missing.\n9.The description of the zero-initialized component is vague—does it consist of only a single convolution operation?\n10.The overall contribution lacks substantial innovation beyond architectural integration.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761576910662}], "openreview_url": "https://openreview.net/forum?id=2XP4HD0NBy", "arxiv_id": "2502.14014", "paper_pdf": "papers/2XP4HD0NBy.pdf", "paper_pdf_sha256": "1f84b41df82cfac1258bce7480e673aabce9825d25b8720479cf82f7b57f28a9", "paper_pdf_bytes": 1921993, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhiyuanLi218/segret", "code_repository": "ZhiyuanLi218/segret", "code_commit": "99c409054cf74eac7a08d610525df89090af35b7", "code_archive": "repos/2XP4HD0NBy.zip", "code_archive_sha256": "261afc109c8aa4ef6b068d68cfea8ff2066c8881bb8b15a40d83dc111f15aa04", "code_archive_bytes": 3424375, "code_file_count": 683, "code_extensions": {".py": 677, ".sh": 4, ".ipynb": 2}, "github_disk_usage_kb": 3085, "github_languages": {"Python": 1203518, "Shell": 2351, "Dockerfile": 750}, "github_archived": false, "github_pushed_at": "2025-09-22T02:07:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/segret-an-efficient-design-for-semantic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1cbFEH0Df", "year": 2025, "status": "rejected", "title": "How to Correctly Do Semantic Backpropagation on Language-based Agentic Systems", "authors": ["Wenyi Wang", "Hisham Abdullah Alyahya", "Dylan R. Ashley", "Oleg Serikov", "Dmitrii Khizbullin", "Francesco Faccio", "Jürgen Schmidhuber"], "authorids": ["~Wenyi_Wang1", "~Hisham_Abdullah_Alyahya1", "~Dylan_R._Ashley1", "~Oleg_Serikov1", "~Dmitrii_Khizbullin2", "~Francesco_Faccio1", "~Jürgen_Schmidhuber1"], "authors_source": "OpenReview API", "abstract": "Language-based agentic systems have shown great promise in recent years, transitioning from solving small-scale research problems to being deployed in challenging real-world tasks. However, optimizing these systems often requires substantial manual labor. Recent studies have demonstrated that these systems can be represented as computational graphs, enabling automatic optimization. Despite these advancements, most current efforts in Graph-based Agentic System Optimization (GASO) fail to properly assign feedback to the system’s components given feedback on the system’s output. To address this challenge, we formalize the concept of semantic backpropagation with semantic gradients—a generalization that aligns several key optimization techniques, including reverse-mode automatic differentiation and the more recent TextGrad by exploiting the relationship among nodes with a common successor. This serves as a method for computing directional information about how changes to each component of an agentic system might improve the system’s output. To use these gradients, we propose a method called semantic gradient descent which enables us to solve GASO effectively. Our results on both BIG-Bench Hard and GSM8K show that our approach outperforms existing state-of-the-art methods for solving GASO problems. A detailed ablation study on the LIAR dataset demonstrates the parsimonious nature of our method.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "0jqqXroWyu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11702/Reviewer_eMLC"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This work presents a framework to extend the text-based gradient method (ask the model to refine the prompt. This series of works often formulates this refinement as a gradient decent process of the discrete prompt) for handling the graph structure, such as using multiple prompts to form an agent for problem-solving.", "review_text": "This work presents a framework to extend the text-based gradient method (ask the model to refine the prompt. This series of works often formulates this refinement as a gradient decent process of the discrete prompt) for handling the graph structure, such as using multiple prompts to form an agent for problem-solving.", "strengths": "This work systematically discussed the \"backpropagation\" of the text-based gradient method and offered a framework to aggregate the gradients, making the text-based gradient methods close to the normal gradient descent.", "weaknesses": "The paper writing could be improved. I think placing figures and examples from the appendix into the main document could largely improve the readability. The experiment setting could be more clearly stated in the main document. I found those in the appendix, such as the initial graph for BBH, GSM8K, and LIAR. However, the number of nodes and the diameter of the graph are small, raising concerns about the proposed method's generalization ability on more complicated graphs, such as the MCTS, tree of thoughts, or wizardLM series. \n\nNeeds more in-depth analyses, such as the robustness of the prompt templates for the forward and backward functions. It's also better to test other models than GPT4 since the GPT4 has already achieved a very high score on GSM8K and BBH. Introducing other models could also be helpful in evaluating the backward function design. If I understand correctly, the backward prompt filled with real materials (instruction, question, response, and other statements) could be very long and complicated. My concern is that the GPT4 may not follow it well.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a framework to extend the text-based gradient method (ask the model to refine the prompt. This series of works often formulates this refinement as a gradient decent process of the discrete prompt) for handling the graph structure, such as using multiple prompts to form an agent for problem-solving.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "This work systematically discussed the \"backpropagation\" of the text-based gradient method and offered a framework to aggregate the gradients, making the text-based gradient methods close to the normal gradient descent.", "weaknesses": "The paper writing could be improved. I think placing figures and examples from the appendix into the main document could largely improve the readability. The experiment setting could be more clearly stated in the main document. I found those in the appendix, such as the initial graph for BBH, GSM8K, and LIAR. However, the number of nodes and the diameter of the graph are small, raising concerns about the proposed method's generalization ability on more complicated graphs, such as the MCTS, tree of thoughts, or wizardLM series. \n\nNeeds more in-depth analyses, such as the robustness of the prompt templates for the forward and backward functions. It's also better to test other models than GPT4 since the GPT4 has already achieved a very high score on GSM8K and BBH. Introducing other models could also be helpful in evaluating the backward function design. If I understand correctly, the backward prompt filled with real materials (instruction, question, response, and other statements) could be very long and complicated. My concern is that the GPT4 may not follow it well.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730775106197}, {"id": "Xf4qwR0fHC", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11702/Reviewer_hy82"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper falls in the vein of research that's attempting to expand the concept of \"gradient-based optimization\" to language-based agentic systems. Specifically, several papers have attempted to take the concept of \"autodiff\" and apply it to graph-based agentic systems. From my understanding, this paper is most similar to TextGrad, but identifies some improvements in TextGrad that 1. make this paper more analogous to standard autodiff,  2. improves performance by taking advantage of more information during the backwards pass.", "review_text": "This paper falls in the vein of research that's attempting to expand the concept of \"gradient-based optimization\" to language-based agentic systems. Specifically, several papers have attempted to take the concept of \"autodiff\" and apply it to graph-based agentic systems. From my understanding, this paper is most similar to TextGrad, but identifies some improvements in TextGrad that 1. make this paper more analogous to standard autodiff,  2. improves performance by taking advantage of more information during the backwards pass.", "strengths": "Overall, I think this paper identifies a reasonable gap in the existing textual gradient-based approaches (e.g. TextGrad). Specifically, if I understand the paper correctly, the paper is saying that if you have a function `c = f(a, b)`, the autodiff rule for `da` given `dc` should also depend upon `b`, not just `a`, `c`, and `dc` (which is what TextGrad's formulation does). This follows naturally from autodiff's definition, where    say, the autodiff rule for `da` in `c = mm(a, b)` involves `b`.\n\nThe \"intuitive\" explanation given in the paper also makes sense to me.\n\nThis motivation for their results is also validated by their experimental results.", "weaknesses": "I personally found the presentation of the paper somewhat confusing. In particular, I think a more concrete example walking through what exactly a semantic gradient looked like would make the paper much easier to understand, particularly if it were contrasted against the gradients from TextGrad. I see the examples of the evolved prompts in the appendix (which are useful!) but I'd be interested to see how the actual gradients look like, since that's the primary way in which this paper differs from prior work.\n\nMore broadly speaking, I also find myself being somewhat skeptical of the autodiff analogy, and to what extent semantic gradients meaningfully map to autodiff. For example, for mathematical autodiff a gradient directly corresponds to how that parameter effects the output loss (assuming an epsilon step). With semantic gradients, however, there's no such correspondence like this. \n\nI would also be interested in seeing how different graph structures lead to different performance. For example, as far as I can tell, all the graphs in this paper involve only one \"layer\".\n\nI'd also like to see more benchmarks, as opposed to just GSM8k and BBH NLP/Algorithmic.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper falls in the vein of research that's attempting to expand the concept of \"gradient-based optimization\" to language-based agentic systems. Specifically, several papers have attempted to take the concept of \"autodiff\" and apply it to graph-based agentic systems. From my understanding, this paper is most similar to TextGrad, but identifies some improvements in TextGrad that 1. make this paper more analogous to standard autodiff,  2. improves performance by taking advantage of more information during the backwards pass.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Overall, I think this paper identifies a reasonable gap in the existing textual gradient-based approaches (e.g. TextGrad). Specifically, if I understand the paper correctly, the paper is saying that if you have a function `c = f(a, b)`, the autodiff rule for `da` given `dc` should also depend upon `b`, not just `a`, `c`, and `dc` (which is what TextGrad's formulation does). This follows naturally from autodiff's definition, where    say, the autodiff rule for `da` in `c = mm(a, b)` involves `b`.\n\nThe \"intuitive\" explanation given in the paper also makes sense to me.\n\nThis motivation for their results is also validated by their experimental results.", "weaknesses": "I personally found the presentation of the paper somewhat confusing. In particular, I think a more concrete example walking through what exactly a semantic gradient looked like would make the paper much easier to understand, particularly if it were contrasted against the gradients from TextGrad. I see the examples of the evolved prompts in the appendix (which are useful!) but I'd be interested to see how the actual gradients look like, since that's the primary way in which this paper differs from prior work.\n\nMore broadly speaking, I also find myself being somewhat skeptical of the autodiff analogy, and to what extent semantic gradients meaningfully map to autodiff. For example, for mathematical autodiff a gradient directly corresponds to how that parameter effects the output loss (assuming an epsilon step). With semantic gradients, however, there's no such correspondence like this. \n\nI would also be interested in seeing how different graph structures lead to different performance. For example, as far as I can tell, all the graphs in this paper involve only one \"layer\".\n\nI'd also like to see more benchmarks, as opposed to just GSM8k and BBH NLP/Algorithmic.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730702834886}, {"id": "RSEXXZInol", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11702/Reviewer_HxPC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "To enable text optimization (omitting numerical gradients) of model outputs based on feedback, the authors propose Graph-based Agentic System Optimization (GASO). This framework formalizes the concept of semantic backpropagation with semantic gradients, generalizing several key optimization techniques, including reverse-mode automatic differentiation and TextGrad, by exploiting the relationships among nodes with a common successor. GASO demonstrates effectiveness in both the BIG-Bench Hard and GSM8K datasets.\n\nReferences:\n\n[1] TextGrad: Automatic \"Differentiation\" via Text.", "review_text": "To enable text optimization (omitting numerical gradients) of model outputs based on feedback, the authors propose Graph-based Agentic System Optimization (GASO). This framework formalizes the concept of semantic backpropagation with semantic gradients, generalizing several key optimization techniques, including reverse-mode automatic differentiation and TextGrad, by exploiting the relationships among nodes with a common successor. GASO demonstrates effectiveness in both the BIG-Bench Hard and GSM8K datasets.\n\nReferences:\n\n[1] TextGrad: Automatic \"Differentiation\" via Text.", "strengths": "1. This work follows an important line of research focused on text optimization based on natural language gradients.\n\n2. The experimental results demonstrate significant improvements over existing semantic optimization methods, such as TextGrad.\n\n3. The proposed method is clear and straightforward to implement.", "weaknesses": "1. The evaluation datasets are somewhat limited, comprising only BIG-Bench Hard and GSM8K. In contrast, TextGrad has been evaluated on additional datasets, including code generation, drug molecule optimization, and radiotherapy treatment plan optimization. It would be beneficial to see the potential of GASO applied to a wider range of applications.\n\n2. Since GASO incorporates additional neighboring information in its computations, the cost associated with this approach is not clearly defined. It would be valuable to understand the trade-off between computational cost and performance.", "questions": "See Weakness above.\n\n1. Performances on other reasoning tasks, such as code generation.\n2. The trade-off between performance and cost.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To enable text optimization (omitting numerical gradients) of model outputs based on feedback, the authors propose Graph-based Agentic System Optimization (GASO). This framework formalizes the concept of semantic backpropagation with semantic gradients, generalizing several key optimization techniques, including reverse-mode automatic differentiation and TextGrad, by exploiting the relationships among nodes with a common successor. GASO demonstrates effectiveness in both the BIG-Bench Hard and GSM8K datasets.\n\nReferences:\n\n[1] TextGrad: Automatic \"Differentiation\" via Text.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This work follows an important line of research focused on text optimization based on natural language gradients.\n\n2. The experimental results demonstrate significant improvements over existing semantic optimization methods, such as TextGrad.\n\n3. The proposed method is clear and straightforward to implement.", "weaknesses": "1. The evaluation datasets are somewhat limited, comprising only BIG-Bench Hard and GSM8K. In contrast, TextGrad has been evaluated on additional datasets, including code generation, drug molecule optimization, and radiotherapy treatment plan optimization. It would be beneficial to see the potential of GASO applied to a wider range of applications.\n\n2. Since GASO incorporates additional neighboring information in its computations, the cost associated with this approach is not clearly defined. It would be valuable to understand the trade-off between computational cost and performance.", "questions": "See Weakness above.\n\n1. Performances on other reasoning tasks, such as code generation.\n2. The trade-off between performance and cost.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730673474644}, {"id": "Ne78WDJZ8V", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11702/Reviewer_WCzZ"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "Recently, Language-based agentic systems have shown great potential in solving real-world tasks. However, these works usually relied on human-designed settings (e.g., instruction or prompt). How to enable an automatic optimization for agentic systems has been a challenge. In this paper, authors introduce a semantic backpropagation with semantic gradients, which explores how to change each component of an agentic system to improve outputs. Experimental results on BIG-Bench and GSM8K demonstrate that the proposed method can solve GASO problems.", "review_text": "Recently, Language-based agentic systems have shown great potential in solving real-world tasks. However, these works usually relied on human-designed settings (e.g., instruction or prompt). How to enable an automatic optimization for agentic systems has been a challenge. In this paper, authors introduce a semantic backpropagation with semantic gradients, which explores how to change each component of an agentic system to improve outputs. Experimental results on BIG-Bench and GSM8K demonstrate that the proposed method can solve GASO problems.", "strengths": "This paper formulates agentic system as a computational graph task and then introduces a semantic backpropagation to optimize each component of agentic system via semantic gradient.", "weaknesses": "**Weaknesses**\n\n1. Although authors claim the difference between TextGrad as TextGrad use textual gradients while the proposed method is semantic gradients, it seems the proposed method still heavily relied on language features and human design (e.g., Implementation 2).\n2. Authors claim that in TextGrad ignores the dependency and heterogeneity, can you provide some real cases about the nodes with dependencies and why they are failed in TextGrad? Does it matter in optimizing agentic system? Can you highlight more insights about your paper when compared with TextGrad?\n3. The proposed method still relied on an evaluation system to obtain feedback. So what happen if applying this method to some open scenarios. Besides, this paper only evaluates Big-Bench and GSM8k. Can you try more datasets to validate the generalization of the proposed method like GAIA. MMLU?", "questions": "1. Can you provide some examples about semantic gradients when compared with TextGrad. From the provided code, the semantic gradient seems is still the natural language form.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Recently, Language-based agentic systems have shown great potential in solving real-world tasks. However, these works usually relied on human-designed settings (e.g., instruction or prompt). How to enable an automatic optimization for agentic systems has been a challenge. In this paper, authors introduce a semantic backpropagation with semantic gradients, which explores how to change each component of an agentic system to improve outputs. Experimental results on BIG-Bench and GSM8K demonstrate that the proposed method can solve GASO problems.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "This paper formulates agentic system as a computational graph task and then introduces a semantic backpropagation to optimize each component of agentic system via semantic gradient.", "weaknesses": "**Weaknesses**\n\n1. Although authors claim the difference between TextGrad as TextGrad use textual gradients while the proposed method is semantic gradients, it seems the proposed method still heavily relied on language features and human design (e.g., Implementation 2).\n2. Authors claim that in TextGrad ignores the dependency and heterogeneity, can you provide some real cases about the nodes with dependencies and why they are failed in TextGrad? Does it matter in optimizing agentic system? Can you highlight more insights about your paper when compared with TextGrad?\n3. The proposed method still relied on an evaluation system to obtain feedback. So what happen if applying this method to some open scenarios. Besides, this paper only evaluates Big-Bench and GSM8k. Can you try more datasets to validate the generalization of the proposed method like GAIA. MMLU?", "questions": "1. Can you provide some examples about semantic gradients when compared with TextGrad. From the provided code, the semantic gradient seems is still the natural language form.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730382631337}], "openreview_url": "https://openreview.net/forum?id=r1cbFEH0Df", "arxiv_id": "2412.03624", "paper_pdf": "papers/r1cbFEH0Df.pdf", "paper_pdf_sha256": "2e4f7e2de54d6ded9e3029eaf0dfc60abe00390310a4d7e967d3c50d3ea417ec", "paper_pdf_bytes": 746052, "paper_pdf_source": "openreview", "code_url": "https://github.com/HishamAlyahya/semantic_backprop", "code_repository": "HishamAlyahya/semantic_backprop", "code_commit": "aad0c724625f5f1e304b9b8a0c67bb1e8f929d81", "code_archive": "repos/r1cbFEH0Df.zip", "code_archive_sha256": "bb034dab83dcdfaf43e02944985c48305a1f0059719eeb7deaf47c406dbc36a3", "code_archive_bytes": 241655, "code_file_count": 22, "code_extensions": {".py": 18, ".sh": 4}, "github_disk_usage_kb": 229, "github_languages": {"Python": 49654, "Shell": 1812}, "github_archived": false, "github_pushed_at": "2024-12-06T18:18:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/how-to-correctly-do-semantic-backpropagation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eddd0YTCiq", "year": 2024, "status": "rejected", "title": "Graph-level Representation Learning with Joint-Embedding Predictive Architectures", "authors": ["Geri Skenderi", "Hang Li", "Jiliang Tang", "Marco Cristani"], "authorids": ["~Geri_Skenderi1", "~Hang_Li10", "~Jiliang_Tang1", "~Marco_Cristani1"], "authors_source": "OpenReview API", "abstract": "Joint-Embedding Predictive Architectures (JEPAs) have recently emerged as a novel and powerful technique for self-supervised representation learning. They aim to learn an energy-based model by predicting the latent representation of a target signal $y$ from a context signal $x$. JEPAs bypass the need for data augmentation and negative samples, which are typically required by contrastive learning, while avoiding the overfitting issues associated with generative-based pretraining. In this paper, we show that graph-level representations can be effectively modeled using this paradigm and propose Graph-JEPA, the first JEPA for the graph domain. In particular, we employ masked modeling to learn embeddings for different subgraphs of the input graph. To endow the representations with the implicit hierarchy that is often present in graph-level concepts, we devise an alternative training objective that consists of predicting the coordinates of the encoded subgraphs on the unit hyperbola in the 2D plane. Extensive validation shows that Graph-JEPA can learn representations that are expressive and competitive in both graph classification and regression problems. The implementation will be available upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "4NDkIA2XpX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2151/Reviewer_b3AV"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "1 poor", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes Graph-JEPA, the first Joint-Embedding Predictive Architectures (JEPAs) for the graph domain.\nThe application of JEPA to graphs seems to be novel.", "review_text": "This paper proposes Graph-JEPA, the first Joint-Embedding Predictive Architectures (JEPAs) for the graph domain.\nThe application of JEPA to graphs seems to be novel.", "strengths": "- The proposed method is technically sound\n- Based on the experimental results, the improvement between Graph-JEPA over the baselines seems to be strong", "weaknesses": "- The overall method seems to be a direct application of JEPA to graphs.\n- The discussion of \"why does graph-JPEA works\" is useful, but not information. Any theoretical analysis here will be useful.\n- The experimental settings are confusing. It is unclear to me why \"GCN\", a GNN model, can be compared with \"Graph-JEPA\", which is a graph self-supervised training method.\n- All the figures and tables are not professional and could be improved to be more appealing. Font sizes and colors should be improved.", "questions": "- What makes applying JEPA to graphs special and non-trivial?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Graph-JEPA, the first Joint-Embedding Predictive Architectures (JEPAs) for the graph domain.\nThe application of JEPA to graphs seems to be novel.", "soundness": "3 good", "presentation": "1 poor", "contribution": "1 poor", "strengths": "- The proposed method is technically sound\n- Based on the experimental results, the improvement between Graph-JEPA over the baselines seems to be strong", "weaknesses": "- The overall method seems to be a direct application of JEPA to graphs.\n- The discussion of \"why does graph-JPEA works\" is useful, but not information. Any theoretical analysis here will be useful.\n- The experimental settings are confusing. It is unclear to me why \"GCN\", a GNN model, can be compared with \"Graph-JEPA\", which is a graph self-supervised training method.\n- All the figures and tables are not professional and could be improved to be more appealing. Font sizes and colors should be improved.", "questions": "- What makes applying JEPA to graphs special and non-trivial?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698822398426}, {"id": "SNW6F2b5uA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2151/Reviewer_15Jn"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose Graph-JEPA. Graph-JEPA uses two encoders to receive the input and one of the encoders predicts the latent representation of the input signal based on another encoder.", "review_text": "The authors propose Graph-JEPA. Graph-JEPA uses two encoders to receive the input and one of the encoders predicts the latent representation of the input signal based on another encoder.", "strengths": "The writing is clear. First JEPA for graph. The authors provide an analysis to explain why JEPA works for the graph domain.", "weaknesses": "1. The method is not novel. The proposed Graph-JEPA is very similar to MLM in BERT, which utilizes the context to predict the masked word type. \n\n2.  The proposed method is too simple and the motivation is not clear. We have graph MAE and contrastive learning. Why do we need JEPA for the graph domain?\n\n3. Compared with graph MAE and S2GAE, the performance is not good enough to show it can inspire future research.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose Graph-JEPA. Graph-JEPA uses two encoders to receive the input and one of the encoders predicts the latent representation of the input signal based on another encoder.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The writing is clear. First JEPA for graph. The authors provide an analysis to explain why JEPA works for the graph domain.", "weaknesses": "1. The method is not novel. The proposed Graph-JEPA is very similar to MLM in BERT, which utilizes the context to predict the masked word type. \n\n2.  The proposed method is too simple and the motivation is not clear. We have graph MAE and contrastive learning. Why do we need JEPA for the graph domain?\n\n3. Compared with graph MAE and S2GAE, the performance is not good enough to show it can inspire future research.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698463005528}, {"id": "YpBa9oydfe", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2151/Reviewer_dWk6"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work proposes a new self-supervised technique for graph neural networks. Grounded on joint-embedding predictive architecture (JEPAs), the proposed Graph-JEPA is designed to predict the latent embeddings for multiple subgraphs based on a random subgraph. Experiments are performed on graph-level tasks.", "review_text": "This work proposes a new self-supervised technique for graph neural networks. Grounded on joint-embedding predictive architecture (JEPAs), the proposed Graph-JEPA is designed to predict the latent embeddings for multiple subgraphs based on a random subgraph. Experiments are performed on graph-level tasks.", "strengths": "1.\tThe design of the loss objective is good.\n\n2.\tThe ablation studies and discussion are detailed and insightful.", "weaknesses": "1.\tMissing literature in related work. In addition to contrastive methods and generative methods, self-supervised graph representation should also include existing predictive methods [1]. For example, CCA-SSG [2] and LaGraph [3] are two existing works using latent embedding prediction. Such predictive methods should be discussed in related works, and they should be used as baseline methods to compare results. \n\n2.\tThe performance improvement is marginal based on the main results in Table 1. \n\n3.\tMost existing SSL methods can handle both graph-level and node-level tasks. However, the proposed Graph-JEPA only supports graph-level downstream tasks. \n\n[1]. Xie, Yaochen, et al. \"Self-supervised learning of graph neural networks: A unified review.\" IEEE transactions on pattern analysis and machine intelligence 45.2 (2022): 2412-2429.\n\n[2]. Zhang, Hengrui, et al. \"From canonical correlation analysis to self-supervised graph neural networks.\" Advances in Neural Information Processing Systems 34 (2021): 76-89.\n\n[3]. Xie, Yaochen, Zhao Xu, and Shuiwang Ji. \"Self-supervised representation learning via latent graph prediction.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "1.\tAuthors claim that the Graph-JEPA is more efficient than contrastive methods since it doesn’t require data augmentations or negative samples. I’m wondering how efficient it is. Could you add a quantitative comparison for the efficiency?\n\n2.\tThe proposed Graph-JEPA uses Transformer encoder blocks. However, most baseline models are based on simpler models like GIN and GCN. Is it an unfair comparison? Can you use GIN/GCN encoder?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes a new self-supervised technique for graph neural networks. Grounded on joint-embedding predictive architecture (JEPAs), the proposed Graph-JEPA is designed to predict the latent embeddings for multiple subgraphs based on a random subgraph. Experiments are performed on graph-level tasks.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1.\tThe design of the loss objective is good.\n\n2.\tThe ablation studies and discussion are detailed and insightful.", "weaknesses": "1.\tMissing literature in related work. In addition to contrastive methods and generative methods, self-supervised graph representation should also include existing predictive methods [1]. For example, CCA-SSG [2] and LaGraph [3] are two existing works using latent embedding prediction. Such predictive methods should be discussed in related works, and they should be used as baseline methods to compare results. \n\n2.\tThe performance improvement is marginal based on the main results in Table 1. \n\n3.\tMost existing SSL methods can handle both graph-level and node-level tasks. However, the proposed Graph-JEPA only supports graph-level downstream tasks. \n\n[1]. Xie, Yaochen, et al. \"Self-supervised learning of graph neural networks: A unified review.\" IEEE transactions on pattern analysis and machine intelligence 45.2 (2022): 2412-2429.\n\n[2]. Zhang, Hengrui, et al. \"From canonical correlation analysis to self-supervised graph neural networks.\" Advances in Neural Information Processing Systems 34 (2021): 76-89.\n\n[3]. Xie, Yaochen, Zhao Xu, and Shuiwang Ji. \"Self-supervised representation learning via latent graph prediction.\" International Conference on Machine Learning. PMLR, 2022.", "questions": "1.\tAuthors claim that the Graph-JEPA is more efficient than contrastive methods since it doesn’t require data augmentations or negative samples. I’m wondering how efficient it is. Could you add a quantitative comparison for the efficiency?\n\n2.\tThe proposed Graph-JEPA uses Transformer encoder blocks. However, most baseline models are based on simpler models like GIN and GCN. Is it an unfair comparison? Can you use GIN/GCN encoder?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698265182201}], "openreview_url": "https://openreview.net/forum?id=eddd0YTCiq", "arxiv_id": "2309.16014", "paper_pdf": "papers/eddd0YTCiq.pdf", "paper_pdf_sha256": "6c31c175acee0b77e8d1e788df2f69e1cf1e8646c24329c1fb33ea5d0ef7158d", "paper_pdf_bytes": 595303, "paper_pdf_source": "openreview", "code_url": "https://github.com/geriskenderi/graph-jepa", "code_repository": "geriskenderi/graph-jepa", "code_commit": "72df1b7704921001ea012a21f840300fbc792cdd", "code_archive": "repos/eddd0YTCiq.zip", "code_archive_sha256": "05daaf1f38416f82a4cd2ae89c0c8cec6824602f58186b5495988d1cbf42ea88", "code_archive_bytes": 162648, "code_file_count": 32, "code_extensions": {".py": 31, ".sh": 1}, "github_disk_usage_kb": 153, "github_languages": {"Python": 130078, "Shell": 531}, "github_archived": false, "github_pushed_at": "2025-01-17T11:19:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-level-representation-learning-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RNnKhz25N1O", "year": 2022, "status": "rejected", "title": "Low-Cost Algorithmic Recourse for Users With Uncertain Cost Functions", "authors": ["Prateek Yadav", "Peter Hase", "Mohit Bansal"], "authorids": ["~Prateek_Yadav1", "~Peter_Hase1", "~Mohit_Bansal2"], "authors_source": "OpenReview API", "abstract": "The problem of identifying algorithmic recourse for people affected by machine learning model decisions has received much attention recently. Existing approaches for recourse generation obtain solutions using properties like diversity, proximity, sparsity, and validity. Yet, these objectives are only heuristics for what we truly care about, which is whether a user is satisfied with the recourses offered to them. Some recent works try to model user-incurred cost, which is more directly linked to user satisfaction. But they assume a single global cost function that is shared across all users. This is an unrealistic assumption when users have dissimilar preferences about their willingness to act upon a feature and different costs associated with changing that feature. In this work, we formalize the notion of user-specific cost functions and introduce a new method for identifying actionable recourses for users. By default, we assume that users' cost functions are hidden from the recourse method, though our framework allows users to partially or completely specify their preferences or cost function. We propose an objective function, Expected Minimum Cost (EMC), based on two key ideas: (1) when presenting a set of options to a user, it is vital that there is at least one low-cost solution the user could adopt; (2) when we do not know the user's true cost function, we can approximately optimize for user satisfaction by first sampling plausible cost functions, then finding a set that achieves a good cost for the user in expectation. We optimize EMC with a novel discrete optimization algorithm, Cost-Optimized Local Search (COLS), which is guaranteed to improve the recourse set quality over iterations. Experimental evaluation on popular real-world datasets with simulated user costs demonstrates that our method satisfies up to 25.89 percentage points more users compared to strong baseline methods. Using standard fairness metrics, we also show that our method can provide more fair solutions across demographic groups than comparable methods, and we verify that our method is robust to misspecification of the cost function distribution. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "DmPOV4HnQiQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4407/Reviewer_1gBz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a recourse methodology to deal with model biases/fairness issues in producing equitable outcomes for all users (classes).  ", "review_text": "The paper addresses an important issue of dealing with model bias in producing fair user outcomes. My main concern with the paper is the lack of clear supporting arguments on why the choice of cost models (including knowing the distribution over the cost functions) is the right one? I am not clear about the assumptions made in the proposed cost function that each user adopts.  First, how is the cost function effectively computed even with the use of a recourse set? Next, how is the recourse set itself guaranteed to always produce at least one reasonable counterfactual in the set?  More specifically, even as the authors acknowledge knowing the exact cost function by each user is difficult, their explanation of using the recourse set to get around this problem with high confidence is unclear to me.  Intuitively, it seems a measure like diversity will be more effective when the cost functions are unknown/private to the user.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a recourse methodology to deal with model biases/fairness issues in producing equitable outcomes for all users (classes).  ", "main_review": "The paper addresses an important issue of dealing with model bias in producing fair user outcomes. My main concern with the paper is the lack of clear supporting arguments on why the choice of cost models (including knowing the distribution over the cost functions) is the right one? I am not clear about the assumptions made in the proposed cost function that each user adopts.  First, how is the cost function effectively computed even with the use of a recourse set? Next, how is the recourse set itself guaranteed to always produce at least one reasonable counterfactual in the set?  More specifically, even as the authors acknowledge knowing the exact cost function by each user is difficult, their explanation of using the recourse set to get around this problem with high confidence is unclear to me.  Intuitively, it seems a measure like diversity will be more effective when the cost functions are unknown/private to the user.  \n", "summary_of_the_review": "I am not completely convinced the proposed model of computing a recourse set to minimize the expected cost for the user is always effective in the absence of knowing the cost function even approximately.   The experiments to show the effectiveness with respect to the proposed baseline are inconclusive with respect to natural measures like diversity.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636315933948}, {"id": "oCbVi2w_otA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4407/Reviewer_cyuv"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the problem of algorithmic recourse is studied where the goal is to find best recourse (counterfactual set) that is optimized for user cost. The author proposed new user-incurred cost evaluation method, Expected Minimum Cost (EMC), which approximate user satisfaction without assuming a fixed global user cost function, and instead consider user cost functions as hidden and user-specific. Specifically, the authors define cost function for each user as a set of feature-specific functions of user-incurred cost when transitioning between feature states, and define MinCost as the minimum transition cost across possible recourses. To cover diverse user cost functions, they propose to model user cost distribution with a hierarchical sampling procedure and estimated expected minimum cost by drawing samples from it. Next, they formulate a discrete optimization problem using EMC as objective, and propose a search algorithm (COLS) for best recourse generation. They introduce three new metrics for user satisfaction (all related to MinCost): FS@k, Coverage and PAC. Finally, they test with two real-world datasets and show that COLS achieves significant outperformance against baselines (that optimize for distance-based metrics) on the newly proposed metrics and show that their method is doing better in fairness as well. ", "review_text": "Pros\n-\tThis paper proposes a new way of evaluating user satisfaction which differs from existing methods that measures on heuristics such as distance/diversity or assume a fixed global user cost function. It is more flexible and realistic, and thus could be an interesting direction to follow. The proposed formulation is quite novel and technically non-trivial, with some theoretical grounding.\n-\tThe experimental results are very strong on the 3 newly proposed metrics. The authors also conduct extensive ablation studies on different aspects of the problem, although many of them are deferred to the supplementary. \n-\tThe discussion is pretty comprehensive, and they also included a fairness analysis.\n\nConcerns\n-\tOne major issue is on the readability of the paper. It is certainly good that the paper contains a lot of information, however currently it seems that the main text is a bit too packed such that very limited detail about the main methodology is provided. In fact, both the core sampling and optimization algorithms are described in the appendix, and it is very hard for this reviewer to understand them solely based on the descriptions in section 4. Perhaps the authors could reorganize the content such that less space is spent on repeating the contributions/motivations. \n-\tAs algorithmic recourse is a rarely new domain and may not be well-known by general audience, it might be better to translate the domain-specific terminologies into plain language or more general language in ML in the introduction part. \n-\tIt seems the newly introduced evaluation metrics are generated using the same sampling distribution used for computing EMC, wouldn’t that be a bit circulated to evaluate something where the ground-truth is closely related to the objective used for optimization? Is there any way to evaluate on more realistic user cost rather than simulating it with the same distribution as the one used in EMC? The authors talk about distributional shift regime in the appendix, that still the ground-truth distribution is from the same family of the EMC distribution (mixture of percentile shift and linear cost). It might be more convincing if it is from a totally independent distribution.\n\nQuestions\n-\tWhat is used as the initial starting set for COLS? \n-\tIn the problem formulation in (2), does it mean that the best recourse set would consist at least one desired outcome solution but it may not be the one with lowest cost? If so how do one achieve balance between outcome and satisfaction?\n-\tWhat is the computational complexity of the algorithm?\n-  Is there any downside from the underperformance in distance-based metrics?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the problem of algorithmic recourse is studied where the goal is to find best recourse (counterfactual set) that is optimized for user cost. The author proposed new user-incurred cost evaluation method, Expected Minimum Cost (EMC), which approximate user satisfaction without assuming a fixed global user cost function, and instead consider user cost functions as hidden and user-specific. Specifically, the authors define cost function for each user as a set of feature-specific functions of user-incurred cost when transitioning between feature states, and define MinCost as the minimum transition cost across possible recourses. To cover diverse user cost functions, they propose to model user cost distribution with a hierarchical sampling procedure and estimated expected minimum cost by drawing samples from it. Next, they formulate a discrete optimization problem using EMC as objective, and propose a search algorithm (COLS) for best recourse generation. They introduce three new metrics for user satisfaction (all related to MinCost): FS@k, Coverage and PAC. Finally, they test with two real-world datasets and show that COLS achieves significant outperformance against baselines (that optimize for distance-based metrics) on the newly proposed metrics and show that their method is doing better in fairness as well. ", "main_review": "Pros\n-\tThis paper proposes a new way of evaluating user satisfaction which differs from existing methods that measures on heuristics such as distance/diversity or assume a fixed global user cost function. It is more flexible and realistic, and thus could be an interesting direction to follow. The proposed formulation is quite novel and technically non-trivial, with some theoretical grounding.\n-\tThe experimental results are very strong on the 3 newly proposed metrics. The authors also conduct extensive ablation studies on different aspects of the problem, although many of them are deferred to the supplementary. \n-\tThe discussion is pretty comprehensive, and they also included a fairness analysis.\n\nConcerns\n-\tOne major issue is on the readability of the paper. It is certainly good that the paper contains a lot of information, however currently it seems that the main text is a bit too packed such that very limited detail about the main methodology is provided. In fact, both the core sampling and optimization algorithms are described in the appendix, and it is very hard for this reviewer to understand them solely based on the descriptions in section 4. Perhaps the authors could reorganize the content such that less space is spent on repeating the contributions/motivations. \n-\tAs algorithmic recourse is a rarely new domain and may not be well-known by general audience, it might be better to translate the domain-specific terminologies into plain language or more general language in ML in the introduction part. \n-\tIt seems the newly introduced evaluation metrics are generated using the same sampling distribution used for computing EMC, wouldn’t that be a bit circulated to evaluate something where the ground-truth is closely related to the objective used for optimization? Is there any way to evaluate on more realistic user cost rather than simulating it with the same distribution as the one used in EMC? The authors talk about distributional shift regime in the appendix, that still the ground-truth distribution is from the same family of the EMC distribution (mixture of percentile shift and linear cost). It might be more convincing if it is from a totally independent distribution.\n\nQuestions\n-\tWhat is used as the initial starting set for COLS? \n-\tIn the problem formulation in (2), does it mean that the best recourse set would consist at least one desired outcome solution but it may not be the one with lowest cost? If so how do one achieve balance between outcome and satisfaction?\n-\tWhat is the computational complexity of the algorithm?\n-  Is there any downside from the underperformance in distance-based metrics?\n", "summary_of_the_review": "In this paper the author proposed a new way for evaluating and optimizing user satisfaction. The technical contributions are solid and the results are rather promising despite the potential bias toward EMC. The paper contains fruitful discussion and ablation studies, although it can be further improved in terms of clarity. Therefore, I would like to give a weak accept.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636005045108}, {"id": "u-0M4E_V9XF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4407/Reviewer_Fy3U"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work introduces a new method for identifying actionable recourses for users with user-specific cost functions. Users’ cost functions are hidden from the recourse method. The paper proposed a discrete optimization algorithm COLS to solve the objective EMC. It further used a popular real-world dataset to illustrate the performance.", "review_text": "I enjoyed reading this paper in general. My major comments are:\n\n1.\tIn Section 4.1, it is assumed that there is a distribution over all the cost functions D_c for the population. Is the distribution D_c known or unknown? A more practical setting is that D_c is unknown. Then how to use Monte Carlo Estimation to approximate the expectation of the MinCost?\n\nFor different users u, it is assumed that C_u follows distribution D_c. However, in the introduction, it claims that “we propose a method for identifying a user-specific recourse set that contain at least one good solution for the user”. However, it seems inconsistent between the motivation and the assumption. Why do all users share the same distribution of the cost function? Is the framework generalizable to the setting with different distribution?\n\n2.\tTheorem 4.1 proves the monotonicity of Cost-Optimized Local Search Algorithm. But how does ExpMinCost(s_u, S_t^best; {C_i}_{i=1}^M) converge? Theorem 4.1 does not imply that, but it is a very important question.\n\n3.\tWhy choosing Equation (3) and Equation (4) as metrics to measure recourse quality? What is the advantage of choosing a threshold function? How to choose k in real cases?\n\n4.\tIn the numerical experiments, could you compare with other functions that measure the recourse quality in previous recourse papers? I think my main concern is on FS@k. Is using FS@k equivalent to the following: assume there exists a black-box algorithm that can output the indicator that if the total cost is smaller than k, then the distance function can be used to measure the recourse quality.\n Could you use numerical experiments to emphasize the advantage of using FS@k compare to other measure functions such as weighted sum of costs?\n\n5.   What is the computational complexity of your algorithm? How is that compared to other benchmarks?\n\n6.    Why is fairness an important issue in this work? Could you comment more on this part to motivate?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work introduces a new method for identifying actionable recourses for users with user-specific cost functions. Users’ cost functions are hidden from the recourse method. The paper proposed a discrete optimization algorithm COLS to solve the objective EMC. It further used a popular real-world dataset to illustrate the performance.", "main_review": "I enjoyed reading this paper in general. My major comments are:\n\n1.\tIn Section 4.1, it is assumed that there is a distribution over all the cost functions D_c for the population. Is the distribution D_c known or unknown? A more practical setting is that D_c is unknown. Then how to use Monte Carlo Estimation to approximate the expectation of the MinCost?\n\nFor different users u, it is assumed that C_u follows distribution D_c. However, in the introduction, it claims that “we propose a method for identifying a user-specific recourse set that contain at least one good solution for the user”. However, it seems inconsistent between the motivation and the assumption. Why do all users share the same distribution of the cost function? Is the framework generalizable to the setting with different distribution?\n\n2.\tTheorem 4.1 proves the monotonicity of Cost-Optimized Local Search Algorithm. But how does ExpMinCost(s_u, S_t^best; {C_i}_{i=1}^M) converge? Theorem 4.1 does not imply that, but it is a very important question.\n\n3.\tWhy choosing Equation (3) and Equation (4) as metrics to measure recourse quality? What is the advantage of choosing a threshold function? How to choose k in real cases?\n\n4.\tIn the numerical experiments, could you compare with other functions that measure the recourse quality in previous recourse papers? I think my main concern is on FS@k. Is using FS@k equivalent to the following: assume there exists a black-box algorithm that can output the indicator that if the total cost is smaller than k, then the distance function can be used to measure the recourse quality.\n Could you use numerical experiments to emphasize the advantage of using FS@k compare to other measure functions such as weighted sum of costs?\n\n5.   What is the computational complexity of your algorithm? How is that compared to other benchmarks?\n\n6.    Why is fairness an important issue in this work? Could you comment more on this part to motivate?\n", "summary_of_the_review": "This paper studied an interesting problem. To improve the paper, the author may want to illustrate the advantage of using FS@k and how it is very different from state-of-art measure functions, both conceptually and numerically.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635915558528}, {"id": "5N3Ba-AcXT3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4407/Reviewer_B3sM"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper aims to find algorithmic recourse that has low-cost to the users. Unlike previous work, the authors do not assume that there is a known global cost function that is shared by all users.", "review_text": "I do not think this papers achieves what it sets out to do. It is mentioned in the related work, how the closest literature to the paper is other cost-based approaches to finding recourse and different from those approaches, this paper drops the assumption that there is a known global cost function that is shared by all users. First, the paper formulates $\\mathrm{MinCost}(\\cdot;\\mathcal{C}_u)$ as the cost function of user $u$, which is characterized by the unknown transition matrices $\\mathcal{C}_u$. However then, the paper assumes a distribution $\\mathcal{D}$ over $\\mathcal{C}_u$ 's of different users is known and proposes to optimize $\\mathbb{E}_\\mathcal{C}{}_\\sim{}_\\mathcal{D}[\\mathrm{MinCost}(\\cdot;\\mathcal{C})]$, which is effectively a known global cost function (with respect to user state $\\mathbf{s}_u$ and recourse set $\\mathcal{S}$). Is this not the case?\n\nHaving said that, the proposed cost function has a certain structure and it is still novel: (i) authors propose a hierarchical cost distribution as the particular $\\mathcal{D}$ they consider and (ii) by only considering the element with the minimum cost for each sample from $\\mathcal{D}$, they exploit the fact that each users only really requires one recourse that they are happy with to be satisfied. Proposing a new cost-based objective like this could still be a valuable contribution. But then, the paper needs to be positioned accordingly and highlight the merits of optimizing a cost-based objective structured in this new way. Note that the current experiments are not helpful in comparing against other cost-based objectives proposed in previous work: cost functions of the users are simulated according to the proposed cost function, then of course, a method that optimizes it would perform better than methods optimizing other cost functions.\n\nSome of the conclusions made in the results section suffer from user preferences being simulated as well. For instance, at the end of \"Q2,\" the authors conclude that high diversity is not necessary to satisfy individual users; this is of course true for the simulated users since their cost function is designed to ignore diversity in the first place.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to find algorithmic recourse that has low-cost to the users. Unlike previous work, the authors do not assume that there is a known global cost function that is shared by all users.", "main_review": "I do not think this papers achieves what it sets out to do. It is mentioned in the related work, how the closest literature to the paper is other cost-based approaches to finding recourse and different from those approaches, this paper drops the assumption that there is a known global cost function that is shared by all users. First, the paper formulates $\\mathrm{MinCost}(\\cdot;\\mathcal{C}_u)$ as the cost function of user $u$, which is characterized by the unknown transition matrices $\\mathcal{C}_u$. However then, the paper assumes a distribution $\\mathcal{D}$ over $\\mathcal{C}_u$ 's of different users is known and proposes to optimize $\\mathbb{E}_\\mathcal{C}{}_\\sim{}_\\mathcal{D}[\\mathrm{MinCost}(\\cdot;\\mathcal{C})]$, which is effectively a known global cost function (with respect to user state $\\mathbf{s}_u$ and recourse set $\\mathcal{S}$). Is this not the case?\n\nHaving said that, the proposed cost function has a certain structure and it is still novel: (i) authors propose a hierarchical cost distribution as the particular $\\mathcal{D}$ they consider and (ii) by only considering the element with the minimum cost for each sample from $\\mathcal{D}$, they exploit the fact that each users only really requires one recourse that they are happy with to be satisfied. Proposing a new cost-based objective like this could still be a valuable contribution. But then, the paper needs to be positioned accordingly and highlight the merits of optimizing a cost-based objective structured in this new way. Note that the current experiments are not helpful in comparing against other cost-based objectives proposed in previous work: cost functions of the users are simulated according to the proposed cost function, then of course, a method that optimizes it would perform better than methods optimizing other cost functions.\n\nSome of the conclusions made in the results section suffer from user preferences being simulated as well. For instance, at the end of \"Q2,\" the authors conclude that high diversity is not necessary to satisfy individual users; this is of course true for the simulated users since their cost function is designed to ignore diversity in the first place.\n", "summary_of_the_review": "I believe the claim that the paper relaxes the assumption of knowing a global cost function is not true. However, it still introduces an interesting new objective to optimize for when finding algorithmic recourse.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635873882417}], "openreview_url": "https://openreview.net/forum?id=RNnKhz25N1O", "arxiv_id": "2111.01235", "paper_pdf": "papers/RNnKhz25N1O.pdf", "paper_pdf_sha256": "4de0195e99ecd2486440e6240ac732a910e5540ff9fd106ed97a6ec268aac7d5", "paper_pdf_bytes": 574461, "paper_pdf_source": "openreview", "code_url": "https://github.com/prateeky2806/EMC-COLS-recourse", "code_repository": "prateeky2806/EMC-COLS-recourse", "code_commit": "dd2c04fccd3c85ddaed0b4705aa98133d8f1b48a", "code_archive": "repos/RNnKhz25N1O.zip", "code_archive_sha256": "30560e796f6ff481493a3a1b702f1040a2532c158aed84131426088a1182c995", "code_archive_bytes": 916938, "code_file_count": 18, "code_extensions": {".py": 16, ".sh": 2}, "github_disk_usage_kb": 886, "github_languages": {"Python": 231463, "Shell": 33197}, "github_archived": false, "github_pushed_at": "2021-11-03T21:09:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/low-cost-algorithmic-recourse-for-users-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LcPefbNSwx_", "year": 2021, "status": "rejected", "title": "Factor Normalization for Deep Neural Network Models", "authors": ["Haobo Qi", "Jing Zhou", "Hansheng Wang"], "authorids": ["qihaobo_gsm@pku.edu.cn", "~Jing_Zhou3", "hansheng@pku.edu.cn"], "authors_source": "OpenReview API", "abstract": "Deep neural network (DNN) models often involve features of high dimensions. In most cases, the high-dimensional features can be decomposed into two parts. The first part is a low-dimensional factor. The second part is the residual feature, with much-reduced variability and inter-feature correlation. This leads to a number of interesting theoretical findings for deep neural network training. Accordingly, we are inspired to develop a new factor normalization method for better performance. The proposed method leads to a new deep learning model with two important features. First, it allows factor related feature extraction. Second, it allows adaptive learning rates for factors and residuals, respectively. This leads to fast convergence speed on both training and validation datsets. A number of empirical experiments are presented to demonstrate its superior performance. The code is available at https://github.com/HazardNeo4869/FactorNormalization", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "-mCRwD6Kdem", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper524/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Strong points**\n\nThe paper provides a very detailed theoretical analysis of motivation.\n\n**Weak points**\n\nAnalysis in section 2 is based on linear regression, but the proposed method is based on deep models.\n\nThe proposed method is more similar to feature extraction instead of a training method.\n\nMany models in the chart are not fully converged. It is important to compare convergence speed but also the final accuracy. I think they are not converged, because the training and testing curve is perfectly smoothing without any fluctuations.\n\nComparison in chart 3 needs improvement. Since the two models have different input signals and model structure, they inherently need different hyper-parameters (not only learning rate) for best performance. Only comparing them with the same learning rate may not adequate to prove the significance of the proposed method.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposed to project input to factor (low-dimensional) and residual features (high-dimensional) to improve neural network training.", "review": "**Strong points**\n\nThe paper provides a very detailed theoretical analysis of motivation.\n\n**Weak points**\n\nAnalysis in section 2 is based on linear regression, but the proposed method is based on deep models.\n\nThe proposed method is more similar to feature extraction instead of a training method.\n\nMany models in the chart are not fully converged. It is important to compare convergence speed but also the final accuracy. I think they are not converged, because the training and testing curve is perfectly smoothing without any fluctuations.\n\nComparison in chart 3 needs improvement. Since the two models have different input signals and model structure, they inherently need different hyper-parameters (not only learning rate) for best performance. Only comparing them with the same learning rate may not adequate to prove the significance of the proposed method.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603951001791}, {"id": "CYJQSkshC2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper524/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes a training scheme based on decomposing input features into two parts which have different training dynamics: a low rank \"factor feature\" computed using PCA on the raw features, and a high rank \"residual\".  The former is processed  by a very shallow network, while the latter passes through the full network, and parameters for each are updated with different learning rates.  Experiments show that the proposed algorithm can speed up wall-time to  a certain accuracy on MNIST and CIFAR10 classification, across several neural network architectures and optimizers.\n\nPros:\n- Simple and straightforward proposal, easy to understand, and seems to lead to improved accuracy/reduced loss early in training.\n- Experiments cover several datasets and architectures spanning very simple to very deep, which is nice to see.\n\nConcerns:\n\n- Very \"mathy\" presentation in Sec. 2 is very dense and difficult to follow, and only seems to motivate the method in the special cases of a very shallow linear regression model, where the gradient is easily related to eigenvalues of the input features $X$ (and assuming a \"strong factor structure\", which has been shown to be reasonable for natural images.)  It's not clear why these results would necessarily generalize to very deep and highly nonlinear networks such as AlexNet.  Perhaps this is the motivation for only processing \"factor features\" with an extremely shallow (albeit still nonlinear) network? If so, this should be more clearly explained in the text.\n\n- The experiments are not very convincing. \n - The baselines trained with vanilla SGD or other optimized (Adam, etc.) are not clearly explained.  Do they use the same feature factorization as SGD+FN or do they consist of only the raw features passed into the full network (i.e. the left side of Figure 1)?  If the latter, the comparisons are not totally fair as the input features and networks differ.\n - Models are not trained until convergence, but only for a fixed amount of time, or until a fixed (but not very high) accuracy is reached.  How does the proposed scheme affect final accuracy/loss at convergence?  (E.g., I'd expect simple logistic regression on MNIST to achieve over 95% accuracy)   Is it strictly a training speedup, while converging to the same performance as baselines?  Or is it only suitable when operating within a limit training time budget?\n  - This issue is illustrated in the training curves in Fig 3 (MLP on MNIST), where the validation accuracy for the proposed SGD+FN appears to be plateauing faster than the baseline SGD curves.\n - Furthermore, the speedup over NAG or Adam on CIFAR10 using  ResNet50 isn't very large. \n - All training comparisons are measured only as a function of wall-time.  But the training hardware is not explained.\n - Missing obvious baselines of training separate models on either the factor features or residual alone.  At least for MNIST I'd expect decent performance to be obtainable from the factor features alone.  Maybe the residual is mostly irrelevant to the task?\n\nOverall I feel that the paper is not ready for publication at this time since the experimental validation is incomplete and does not fully explore the benefits and potential downsides of the proposed method.\n\nOther comments:\n\n- Sec 3.3: \"Adaptive learning\".  If I'm understanding correctly, the learning rates are defined based on the eigenvalues of $X^T X$, but are then kept fixed.  So, unlike e.g. AdaGrad or Adam,  the algorithm is not really adaptive in the sense that the learning rate is kept constant throughout training.\n\n- Sec 3.3: \"a standard SGD algorithm cannot be used to train a DNN model\".   This seems like an unnecessarily strong statement given how commonly SGD is used for DNN training.\n\n- The linked code does not appear to include resnet experiments\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "experiments not very convincing", "review": "The paper describes a training scheme based on decomposing input features into two parts which have different training dynamics: a low rank \"factor feature\" computed using PCA on the raw features, and a high rank \"residual\".  The former is processed  by a very shallow network, while the latter passes through the full network, and parameters for each are updated with different learning rates.  Experiments show that the proposed algorithm can speed up wall-time to  a certain accuracy on MNIST and CIFAR10 classification, across several neural network architectures and optimizers.\n\nPros:\n- Simple and straightforward proposal, easy to understand, and seems to lead to improved accuracy/reduced loss early in training.\n- Experiments cover several datasets and architectures spanning very simple to very deep, which is nice to see.\n\nConcerns:\n\n- Very \"mathy\" presentation in Sec. 2 is very dense and difficult to follow, and only seems to motivate the method in the special cases of a very shallow linear regression model, where the gradient is easily related to eigenvalues of the input features $X$ (and assuming a \"strong factor structure\", which has been shown to be reasonable for natural images.)  It's not clear why these results would necessarily generalize to very deep and highly nonlinear networks such as AlexNet.  Perhaps this is the motivation for only processing \"factor features\" with an extremely shallow (albeit still nonlinear) network? If so, this should be more clearly explained in the text.\n\n- The experiments are not very convincing. \n - The baselines trained with vanilla SGD or other optimized (Adam, etc.) are not clearly explained.  Do they use the same feature factorization as SGD+FN or do they consist of only the raw features passed into the full network (i.e. the left side of Figure 1)?  If the latter, the comparisons are not totally fair as the input features and networks differ.\n - Models are not trained until convergence, but only for a fixed amount of time, or until a fixed (but not very high) accuracy is reached.  How does the proposed scheme affect final accuracy/loss at convergence?  (E.g., I'd expect simple logistic regression on MNIST to achieve over 95% accuracy)   Is it strictly a training speedup, while converging to the same performance as baselines?  Or is it only suitable when operating within a limit training time budget?\n  - This issue is illustrated in the training curves in Fig 3 (MLP on MNIST), where the validation accuracy for the proposed SGD+FN appears to be plateauing faster than the baseline SGD curves.\n - Furthermore, the speedup over NAG or Adam on CIFAR10 using  ResNet50 isn't very large. \n - All training comparisons are measured only as a function of wall-time.  But the training hardware is not explained.\n - Missing obvious baselines of training separate models on either the factor features or residual alone.  At least for MNIST I'd expect decent performance to be obtainable from the factor features alone.  Maybe the residual is mostly irrelevant to the task?\n\nOverall I feel that the paper is not ready for publication at this time since the experimental validation is incomplete and does not fully explore the benefits and potential downsides of the proposed method.\n\nOther comments:\n\n- Sec 3.3: \"Adaptive learning\".  If I'm understanding correctly, the learning rates are defined based on the eigenvalues of $X^T X$, but are then kept fixed.  So, unlike e.g. AdaGrad or Adam,  the algorithm is not really adaptive in the sense that the learning rate is kept constant throughout training.\n\n- Sec 3.3: \"a standard SGD algorithm cannot be used to train a DNN model\".   This seems like an unnecessarily strong statement given how commonly SGD is used for DNN training.\n\n- The linked code does not appear to include resnet experiments\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603896238221}, {"id": "A_OdbWQgcUu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper524/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nIn this paper, a learning method that accelerates the training of DNN is proposed. Given an input X, the proposed method decomposes X as X = BZ + E where BZ is a low-rank approximation of X and E is the residual term. E is used as an input of DNN and Z is used as an additional feature of the input of the last layer. Experiments using MNIST and CIFAR10 show that the proposed method accelerates the speed to reduce the training loss. \n\n\nDetailed comments:\nThe observed phenomena --- the proposed method accelerates the learning speed of SGD --- is quite interesting. I haven't noticed existing studies reporting such findings. However, I feel \"why\" parts are not clearly explained in this paper. For example, I have the following questions.\n- The analysis in Section 2 is based on a shallow network model. How can we apply this analysis to a deep network model?\n- What kind of intuition is behind the architecture design (Fig 1)? Why the residual term should place as the input of the first layer and the main term as the input of the last layer, rather than e.g. some intermediate layer?\n\nAlso, I feel the experiments are not convincing enough. The main motivation of this paper is that \"the ultrahigh dimensional features can be decomposed into two parts\". However, the dimensions of MNIST and CIFAR10 are not quite ultrahigh (28^2=784 and 32^2=1024).  Experiments with more high-dimensional data such as ImageNet would be necessary to convince the research concept. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting phenomena but not easy to understand", "review": "Summary:\nIn this paper, a learning method that accelerates the training of DNN is proposed. Given an input X, the proposed method decomposes X as X = BZ + E where BZ is a low-rank approximation of X and E is the residual term. E is used as an input of DNN and Z is used as an additional feature of the input of the last layer. Experiments using MNIST and CIFAR10 show that the proposed method accelerates the speed to reduce the training loss. \n\n\nDetailed comments:\nThe observed phenomena --- the proposed method accelerates the learning speed of SGD --- is quite interesting. I haven't noticed existing studies reporting such findings. However, I feel \"why\" parts are not clearly explained in this paper. For example, I have the following questions.\n- The analysis in Section 2 is based on a shallow network model. How can we apply this analysis to a deep network model?\n- What kind of intuition is behind the architecture design (Fig 1)? Why the residual term should place as the input of the first layer and the main term as the input of the last layer, rather than e.g. some intermediate layer?\n\nAlso, I feel the experiments are not convincing enough. The main motivation of this paper is that \"the ultrahigh dimensional features can be decomposed into two parts\". However, the dimensions of MNIST and CIFAR10 are not quite ultrahigh (28^2=784 and 32^2=1024).  Experiments with more high-dimensional data such as ImageNet would be necessary to convince the research concept. ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603877161146}, {"id": "5_kAUE_L0ak", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper524/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Review of \"Factor normalization for DNNs\". \n\nThe paper makes an observation that datasets used for training many deep neural nets exhibit a strong factor structure, i.e. have a small number of dominant principal components explaining most of the variance.  If we were to remove the dominant factors, the residuals would have much weaker correlation structure and allow faster DNN convergence with SGD.   The paper proposes to separate the dominant factors and the residuals,  train the original DNN on residuals with a much faster SGD learning rate,  and then recombine with a shallow small NN learned on the dominant factors trained using its own (slower) learning rate. \n\nFirst the positive aspects:  this is an interesting idea, and I haven't seen it commonly used in practice for DNNs.  While mathematical analysis is conducted for linear models with GD and is fairly straightforward, it nicely illustrates the main issues, and the hope is that the linear intuition continues to apply to deep NNs.  Experimental results suggest that indeed convergence rate can be improved on several datasets.   In terms of criticisms,  there is very limited scholarship of related ideas that have been used both for linear models and for DNNs, in particular (a) various factor-based models that already exist,  (b) preconditioning of linear systems,  (c) neural nets trained on some other sort of residuals -- e.g. laplacian pyramid DNNs .  Another criticism -- is that there is no discussion of how to do large-scale PCA / factor analysis for high-dimensional image data arising in say modern DNN image classification pipelines, and its computational cost,  as simple numpy.linalg.svd won't work.  The paper also claims that strong factor structure (with a small number of dominant components) is prevalent in modern ultra-large scale DNN applications -- I would like to see some references / supporting evidence beyond just computer vision.  \n\nOverall in my opinion the paper needs to consider the context of related works, and has a few other correctable issues, but I would certainly encourage the authors to continue to improve it. \n\nAdditional details: \n\n1)  Prior work -- that should be cited / contrasted with your approach:   \n(a) Since a substantial part of the paper analyzes linear models,  it's important to mention that factor structure has been long exploited in various ML / stats works.  For example factor models, and principal component regression (PCR) attempt to focus the modeling power on the principal components. In other applications (e.g. in financial modeling) one can assume that factors components are less predictable and focus the modeling effort on the residuals. Recent work by Alex Smola et al,  \"Deep factors for forecasting\" has looked at the deep-NN instantiation of this idea.  Basic autoencoders or bottlenecks used in DNN architectures also attempt to capture the low-dimensional structure.  These are all different from what you're doing, but hinge on the same basic concept -- so providing a discussion of your work in context of related work would be important. \nb)  There is a long history of using preconditioners for gradient-based and other iterative solvers of linear systems.  In particular there are some low-rank preconditioners for accelerating convergence of linear systems:  Nicholas Higham, et. al, \"A new preconditioner that exploits low-rank approximations to factorization error\".  There is also work on applying preconditioners specifically for SGD, e.g. see works by Michael Mahoney. \nc)  For some domains a low-pass filter, or some other low-resolution model can serve to replace PCA or factor models. For example Laplacian pyramids have been widely used in image processing to separate the dominant modes from details.  There are existing DNN approaches based on laplacian pyramids: \ne.g. \"Deep laplacian pyramid networks for fast an accurate super-resolution\". \n\n2)  You mention that \"none of these optimization methods has considered the covariance structure\".. related to the Hessian.  There is a substantial effort to develop second-order methods for stochastic gradient descent, in particular at ICML 2020 there was a workshop on this topic.  \"Beyond first order methods in ML\".   http://www.wikicfp.com/cfp/servlet/event.showcfp?eventid=102387&copyownerid=156008\n\n3) In section 2.1. -- it's worth mentioning explicitly that you're specifically analyzing linear regression under gradient descent.  as you look at Loss(y, theta' X).   It's also worth explicitly mentioning that the condition number of the Hessian plays a crucial role in convergence rate of GD.   You talk about the top eigenvalue, where the bottom one in your example is fixed at 1,  but it's worth mentioning the condition number. \n\n4) Sec. 3.1. How do you conduct 'standard principal component analysis' for ultrahigh dimensional features.  This is a computationally tricky problem, \"standard\" methods won't work.   How do you decide on the dimension -- i..e. number of factors to keep? \n\n5) Is the \"time consumption\"  including the time to do PCA? \n\n6) Can you give some references claiming strong factor structure in several DNN applications in ultra-high dimensions?  I do not contest that this is the case -- but it would be useful to have supporting evidence.  What number/fraction of factors is typically required in these applications to capture a nontrivial fraction of variance?\n\n7) Sec. 2.1. Using lower-case kappa for a matrix is strange,  I initially assumed it's a scalar.  Maybe use another capital letter. \n\n8) While the paper is mostly pretty readable, there are various small issues with english language (from stylistic to grammar) use e.g. \"In fact ample amounts of empirical evidence\" --> \"ample empirical evidence\", e.t.c. There are typos in references, e.g.  Zeiler,   \"Computer ence, 2012\". What is that? \n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting simple idea, but requires more scholarship. ", "review": "Review of \"Factor normalization for DNNs\". \n\nThe paper makes an observation that datasets used for training many deep neural nets exhibit a strong factor structure, i.e. have a small number of dominant principal components explaining most of the variance.  If we were to remove the dominant factors, the residuals would have much weaker correlation structure and allow faster DNN convergence with SGD.   The paper proposes to separate the dominant factors and the residuals,  train the original DNN on residuals with a much faster SGD learning rate,  and then recombine with a shallow small NN learned on the dominant factors trained using its own (slower) learning rate. \n\nFirst the positive aspects:  this is an interesting idea, and I haven't seen it commonly used in practice for DNNs.  While mathematical analysis is conducted for linear models with GD and is fairly straightforward, it nicely illustrates the main issues, and the hope is that the linear intuition continues to apply to deep NNs.  Experimental results suggest that indeed convergence rate can be improved on several datasets.   In terms of criticisms,  there is very limited scholarship of related ideas that have been used both for linear models and for DNNs, in particular (a) various factor-based models that already exist,  (b) preconditioning of linear systems,  (c) neural nets trained on some other sort of residuals -- e.g. laplacian pyramid DNNs .  Another criticism -- is that there is no discussion of how to do large-scale PCA / factor analysis for high-dimensional image data arising in say modern DNN image classification pipelines, and its computational cost,  as simple numpy.linalg.svd won't work.  The paper also claims that strong factor structure (with a small number of dominant components) is prevalent in modern ultra-large scale DNN applications -- I would like to see some references / supporting evidence beyond just computer vision.  \n\nOverall in my opinion the paper needs to consider the context of related works, and has a few other correctable issues, but I would certainly encourage the authors to continue to improve it. \n\nAdditional details: \n\n1)  Prior work -- that should be cited / contrasted with your approach:   \n(a) Since a substantial part of the paper analyzes linear models,  it's important to mention that factor structure has been long exploited in various ML / stats works.  For example factor models, and principal component regression (PCR) attempt to focus the modeling power on the principal components. In other applications (e.g. in financial modeling) one can assume that factors components are less predictable and focus the modeling effort on the residuals. Recent work by Alex Smola et al,  \"Deep factors for forecasting\" has looked at the deep-NN instantiation of this idea.  Basic autoencoders or bottlenecks used in DNN architectures also attempt to capture the low-dimensional structure.  These are all different from what you're doing, but hinge on the same basic concept -- so providing a discussion of your work in context of related work would be important. \nb)  There is a long history of using preconditioners for gradient-based and other iterative solvers of linear systems.  In particular there are some low-rank preconditioners for accelerating convergence of linear systems:  Nicholas Higham, et. al, \"A new preconditioner that exploits low-rank approximations to factorization error\".  There is also work on applying preconditioners specifically for SGD, e.g. see works by Michael Mahoney. \nc)  For some domains a low-pass filter, or some other low-resolution model can serve to replace PCA or factor models. For example Laplacian pyramids have been widely used in image processing to separate the dominant modes from details.  There are existing DNN approaches based on laplacian pyramids: \ne.g. \"Deep laplacian pyramid networks for fast an accurate super-resolution\". \n\n2)  You mention that \"none of these optimization methods has considered the covariance structure\".. related to the Hessian.  There is a substantial effort to develop second-order methods for stochastic gradient descent, in particular at ICML 2020 there was a workshop on this topic.  \"Beyond first order methods in ML\".   http://www.wikicfp.com/cfp/servlet/event.showcfp?eventid=102387&copyownerid=156008\n\n3) In section 2.1. -- it's worth mentioning explicitly that you're specifically analyzing linear regression under gradient descent.  as you look at Loss(y, theta' X).   It's also worth explicitly mentioning that the condition number of the Hessian plays a crucial role in convergence rate of GD.   You talk about the top eigenvalue, where the bottom one in your example is fixed at 1,  but it's worth mentioning the condition number. \n\n4) Sec. 3.1. How do you conduct 'standard principal component analysis' for ultrahigh dimensional features.  This is a computationally tricky problem, \"standard\" methods won't work.   How do you decide on the dimension -- i..e. number of factors to keep? \n\n5) Is the \"time consumption\"  including the time to do PCA? \n\n6) Can you give some references claiming strong factor structure in several DNN applications in ultra-high dimensions?  I do not contest that this is the case -- but it would be useful to have supporting evidence.  What number/fraction of factors is typically required in these applications to capture a nontrivial fraction of variance?\n\n7) Sec. 2.1. Using lower-case kappa for a matrix is strange,  I initially assumed it's a scalar.  Maybe use another capital letter. \n\n8) While the paper is mostly pretty readable, there are various small issues with english language (from stylistic to grammar) use e.g. \"In fact ample amounts of empirical evidence\" --> \"ample empirical evidence\", e.t.c. There are typos in references, e.g.  Zeiler,   \"Computer ence, 2012\". What is that? \n ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603558416387}], "openreview_url": "https://openreview.net/forum?id=LcPefbNSwx_", "arxiv_id": null, "paper_pdf": "papers/LcPefbNSwx_.pdf", "paper_pdf_sha256": "43d7270f45c137f1949c4a7ba41494ebe3b749c0d699008bea735a30351f41b2", "paper_pdf_bytes": 298805, "paper_pdf_source": "openreview", "code_url": "https://github.com/HazardNeo4869/FactorNormalization", "code_repository": "HazardNeo4869/FactorNormalization", "code_commit": "e762cdd6ed19d8d9252c08210573285c5f17e918", "code_archive": "repos/LcPefbNSwx_.zip", "code_archive_sha256": "37bfb8a6437d72d4f37828673fa2b143155caebf3a536e7672d613735ab8dd1e", "code_archive_bytes": 1027837, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 1018, "github_languages": {"Python": 73337}, "github_archived": false, "github_pushed_at": "2020-11-23T15:11:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/factor-normalization-for-deep-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SJgw51HFDr", "year": 2020, "status": "rejected", "title": "Sparse Weight Activation Training", "authors": ["Md Aamir Raihan", "Tor M. Aamodt"], "authorids": ["araihan@ece.ubc.ca", "aamodt@ece.ubc.ca"], "authors_source": "OpenReview API", "abstract": "Training convolutional neural networks (CNNs) is time consuming. Prior work has explored how to reduce the computational demands of training by eliminating gradients with relatively small magnitude. We show that eliminating small magnitude components has limited impact on the direction of high-dimensional vectors. However, in the context of training a CNN, we find that eliminating small magnitude components of weight and activation vectors allows us to train deeper networks on more complex datasets versus eliminating small magnitude components of gradients. We propose Sparse Weight Activation Training (SWAT), an algorithm that embodies these observations. SWAT reduces computations by 50% to 80% with better accuracy at a given level of sparsity versus the Dynamic Sparse Graph algorithm. SWAT also reduces memory footprint by 23% to 37% for activations and 50% to 80% for weights.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SylCZRFtor", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1881/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I am the emergency reviewer. Sorry for the late.\n\nThis paper studies a very interesting topic: eliminating small magnitude components of weight and activation vector instead of eliminating small magnitude components of gradients. A clear interpretation and definition towards the forward and backward propagation is presented. The difference between meProp versus SWAT is also plain. Based on some experiment results shown in Figure2, authors announced that accuracy is extremely sensitive to sparsification of output gradients. Thus algorithms SWAT and SAW are proposed to prune the model, which are respectively training with sparse weights and activations, and SWAT only sparsifies the backward pass. Top-K selection is implemented to select which components are set to zero during sparsification.\n\nStrengths:\n1. The writing logic ascends step by step.\n2. Authors showed the harmfulness of the sparsity of gradients by experiment results. Also the comparison between the sparsity of weights and activations are meaningful.\n3. Sufficient experiments are done to generalize SWAT to different models, and the results are fascinating on ImageNet.\n\nWeaknesses: \nIt's a borderline paper. \n1. lack of novelty. The paper has shown a lot experiment results on basic models, but the raising of Top-K algorithm is not novel. Why it is Top-K but not other metrics for selecting zero components? In this view, the paper is likely to be a project summary.\n2. Less comparison to other basic pruning models. More experiments should be done to compare SWAT with other sota pruning models. Then the results will be convincing.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #4", "review": "I am the emergency reviewer. Sorry for the late.\n\nThis paper studies a very interesting topic: eliminating small magnitude components of weight and activation vector instead of eliminating small magnitude components of gradients. A clear interpretation and definition towards the forward and backward propagation is presented. The difference between meProp versus SWAT is also plain. Based on some experiment results shown in Figure2, authors announced that accuracy is extremely sensitive to sparsification of output gradients. Thus algorithms SWAT and SAW are proposed to prune the model, which are respectively training with sparse weights and activations, and SWAT only sparsifies the backward pass. Top-K selection is implemented to select which components are set to zero during sparsification.\n\nStrengths:\n1. The writing logic ascends step by step.\n2. Authors showed the harmfulness of the sparsity of gradients by experiment results. Also the comparison between the sparsity of weights and activations are meaningful.\n3. Sufficient experiments are done to generalize SWAT to different models, and the results are fascinating on ImageNet.\n\nWeaknesses: \nIt's a borderline paper. \n1. lack of novelty. The paper has shown a lot experiment results on basic models, but the raising of Top-K algorithm is not novel. Why it is Top-K but not other metrics for selecting zero components? In this view, the paper is likely to be a project summary.\n2. Less comparison to other basic pruning models. More experiments should be done to compare SWAT with other sota pruning models. Then the results will be convincing.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1573654021695}, {"id": "ryxGyqqk9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1881/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes SWAT as a training algorithm for sparse networks on different architectures. The paper claims being able to reach a level of sparsity with no drop in accuracy.  The goal is to minimize the computations during training time. To this end, SWAT sets to zero the vectors where necessary. Different from other approaches, SWAT uses sparse computation in the forward and backward passes. The intuition behind is that eliminating small components does not have an impact on the training process but can be used to minimize the computation required. \n\nSome Comments:\n\n- The paper is a bit on the empirical side with a decent number of experiments to demonstrate the effectiveness of the proposal. I am on the border between accepting and rejecting. \n\n\n- The top-K implementation is interesting. Page 7 suggests the top-K do not change during training which is reasonable as the update is limited to those components. Would it be possible to avoid completely that compute and quickly select K early in the training process? I would find that an interesting future direction. \n\n- In the experimental section, I missed actual numbers. At the moment, if I understand correctly, the paper is based on theoretical compute savings. How feasible is this considering the sparsity of the operation (assuming unstructured sparsity)? \n\n- In the case of structured sparsity, how this differs from the early pruning process of regularization based pruning algorithms? For instance, in the first reference (compression-aware training), the authors claim the model can be compressed in the early training. If that is the case, how different is SWAT from those type of methods? In those related works, the accuracy does not drop. Implementation wise, those algorithms do make the backward pass also sparse (setting to 0 the gradients). \n\n- At the moment, the algorithm is using a magnitude-based sorting. Would it be possible to have other sorting approaches?\n\n\nMinor things:\n\n- for clarity, I would summarize the algorithm in section 2.2 rather than in the appendix. \n\n- I am surprised by the imagenet training setting. Why only training for 50epochs? The standard training process is 90epochs changing the learning rate in the 30th and 60th. \n\n- I guess the S% sparsity contribution can be improved (rephrased). If the training algorithm sets to zero N parameters seems to me obvious that there will be no drop in accuracy compared to that training process. What is the drop in accuracy referring to?\n\n- check the references. While the list is quite comprehensive, some of them are not referred in the text. Please, add comments where appropriate.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes SWAT as a training algorithm for sparse networks on different architectures. The paper claims being able to reach a level of sparsity with no drop in accuracy.  The goal is to minimize the computations during training time. To this end, SWAT sets to zero the vectors where necessary. Different from other approaches, SWAT uses sparse computation in the forward and backward passes. The intuition behind is that eliminating small components does not have an impact on the training process but can be used to minimize the computation required. \n\nSome Comments:\n\n- The paper is a bit on the empirical side with a decent number of experiments to demonstrate the effectiveness of the proposal. I am on the border between accepting and rejecting. \n\n\n- The top-K implementation is interesting. Page 7 suggests the top-K do not change during training which is reasonable as the update is limited to those components. Would it be possible to avoid completely that compute and quickly select K early in the training process? I would find that an interesting future direction. \n\n- In the experimental section, I missed actual numbers. At the moment, if I understand correctly, the paper is based on theoretical compute savings. How feasible is this considering the sparsity of the operation (assuming unstructured sparsity)? \n\n- In the case of structured sparsity, how this differs from the early pruning process of regularization based pruning algorithms? For instance, in the first reference (compression-aware training), the authors claim the model can be compressed in the early training. If that is the case, how different is SWAT from those type of methods? In those related works, the accuracy does not drop. Implementation wise, those algorithms do make the backward pass also sparse (setting to 0 the gradients). \n\n- At the moment, the algorithm is using a magnitude-based sorting. Would it be possible to have other sorting approaches?\n\n\nMinor things:\n\n- for clarity, I would summarize the algorithm in section 2.2 rather than in the appendix. \n\n- I am surprised by the imagenet training setting. Why only training for 50epochs? The standard training process is 90epochs changing the learning rate in the 30th and 60th. \n\n- I guess the S% sparsity contribution can be improved (rephrased). If the training algorithm sets to zero N parameters seems to me obvious that there will be no drop in accuracy compared to that training process. What is the drop in accuracy referring to?\n\n- check the references. While the list is quite comprehensive, some of them are not referred in the text. Please, add comments where appropriate."}, "tcdate": 1571953113536}, {"id": "S1xw9KbyKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1881/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies training neural networks with sparse weights and sparse activations (SWAT training). By using sparse weights in forward passes as well as sparse weights and activations in backward passes, SWAT can reduce the computation overhead and also reduce the training memory footprint. The primary contributions of the paper are in three folds: 1) The authors empirically compare the impact of (activation) gradient sparsity and weight + activation sparsity on the model performance---the comparison shows that the weight + activation sparsity has less influence on the model accuracy; 2) Across different models on CIFAR and ImageNet dataset, SWAP can reduce the training flops by 50% to 80% while using roughly 2 to 3x less training memory footprint saving (weight + activation); 3) The authors empirically study on why training using top-K based sparsification can attain strong model accuracy---the magnitude-based top-K approach can roughly preserve the directions of the vectors.\n\nI think the claimed contributions are well-validated in general. The design decisions of the approach are well supported by empirical observations and the components of the approach (different top-K methods) are studied properly. Additionally, I like the authors' synthetic-data studies to shed light on why top-K based sparsity can work well. Given the above reason, I give week accept and I am willing to raise the score if the following questions / concerns can be resolved in the rebuttal / future draft:\n\n1. In results such as in figure 4, we observe that using intermediate levels of sparsity can actually demonstrate better generalization performance than the dense baseline training approach. I was wondering if this is because the default hyperparameter produces better training loss in sparse training than in dense training, and consequently the sparse training test performance is also improved over dense training. Without showing this, it is not fully convincing that intermediate sparsity helps prevent overfitting and generalizes better (as the authors discussed in the text).\n\n2. For \"Impact on Convergence\" in section 3.2, it is not clear to me what the authors are using as a metric for the degree of convergence. Thus I can not evaluate the claims here.\n\n3. For \"Efficient Top K implementation\" in section 3.2, the authors suggest  only computing the K-th largest elements periodically to further improve efficiency. However the empirical evidence of whether this approach will significantly degrade the model performance at the end of training is not provided.\n\n4. For the GFLOPS comparison in Figure 7, could the authors elaborate what operations are included into the count? As the sparse operations requires additional indexing operations for computation, I was wondering whether the GFLOPS can realistically reflect the real latency / energy efficiency of the SWAT approach.\n\n5. How the memory access count calculated at the end of page 7? Is it counting the number of float point values (activations, activation gradients, weights) that needs to be fetched for forward and backward pass?\n\n6. At the first paragraph in page 8 (last paragraph above section 4), do the authors imply that the activations of BN layers is not sparsified? Could the authors provide a bit more evidence on how (any why) sparsification of BN activation impacts the model performance.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper studies training neural networks with sparse weights and sparse activations (SWAT training). By using sparse weights in forward passes as well as sparse weights and activations in backward passes, SWAT can reduce the computation overhead and also reduce the training memory footprint. The primary contributions of the paper are in three folds: 1) The authors empirically compare the impact of (activation) gradient sparsity and weight + activation sparsity on the model performance---the comparison shows that the weight + activation sparsity has less influence on the model accuracy; 2) Across different models on CIFAR and ImageNet dataset, SWAP can reduce the training flops by 50% to 80% while using roughly 2 to 3x less training memory footprint saving (weight + activation); 3) The authors empirically study on why training using top-K based sparsification can attain strong model accuracy---the magnitude-based top-K approach can roughly preserve the directions of the vectors.\n\nI think the claimed contributions are well-validated in general. The design decisions of the approach are well supported by empirical observations and the components of the approach (different top-K methods) are studied properly. Additionally, I like the authors' synthetic-data studies to shed light on why top-K based sparsity can work well. Given the above reason, I give week accept and I am willing to raise the score if the following questions / concerns can be resolved in the rebuttal / future draft:\n\n1. In results such as in figure 4, we observe that using intermediate levels of sparsity can actually demonstrate better generalization performance than the dense baseline training approach. I was wondering if this is because the default hyperparameter produces better training loss in sparse training than in dense training, and consequently the sparse training test performance is also improved over dense training. Without showing this, it is not fully convincing that intermediate sparsity helps prevent overfitting and generalizes better (as the authors discussed in the text).\n\n2. For \"Impact on Convergence\" in section 3.2, it is not clear to me what the authors are using as a metric for the degree of convergence. Thus I can not evaluate the claims here.\n\n3. For \"Efficient Top K implementation\" in section 3.2, the authors suggest  only computing the K-th largest elements periodically to further improve efficiency. However the empirical evidence of whether this approach will significantly degrade the model performance at the end of training is not provided.\n\n4. For the GFLOPS comparison in Figure 7, could the authors elaborate what operations are included into the count? As the sparse operations requires additional indexing operations for computation, I was wondering whether the GFLOPS can realistically reflect the real latency / energy efficiency of the SWAT approach.\n\n5. How the memory access count calculated at the end of page 7? Is it counting the number of float point values (activations, activation gradients, weights) that needs to be fetched for forward and backward pass?\n\n6. At the first paragraph in page 8 (last paragraph above section 4), do the authors imply that the activations of BN layers is not sparsified? Could the authors provide a bit more evidence on how (any why) sparsification of BN activation impacts the model performance.\n\n"}, "tcdate": 1570867599090}], "openreview_url": "https://openreview.net/forum?id=SJgw51HFDr", "arxiv_id": "2001.01969", "paper_pdf": "papers/SJgw51HFDr.pdf", "paper_pdf_sha256": "be25473e6a5ffc0f0535bfd3b211245da24f2396203f45c2642242dca82d8b32", "paper_pdf_bytes": 1285260, "paper_pdf_source": "openreview", "code_url": "https://github.com/AamirRaihan/SWAT", "code_repository": "AamirRaihan/SWAT", "code_commit": "bbe271afa6e5ead3c02b20409594e17b6265b958", "code_archive": "repos/SJgw51HFDr.zip", "code_archive_sha256": "dd860cae2d548df0e132ab355590f717f3caec3bc87f4416f0059197db5fc488", "code_archive_bytes": 2911205, "code_file_count": 52, "code_extensions": {".py": 40, ".sh": 11, ".cpp": 1}, "github_disk_usage_kb": 2813, "github_languages": {"Python": 231631, "C++": 7587, "Shell": 1707}, "github_archived": false, "github_pushed_at": "2021-07-23T10:25:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sparse-weight-activation-training-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "KJNgtPNxtv", "year": 2026, "status": "rejected", "title": "PFMBench: Protein Foundation Model Benchmark", "authors": ["Zhangyang Gao", "Hao Wang", "Cheng Tan", "Chenrui XU", "Mengdi Liu", "Bozhen Hu", "Linlin Chao", "Xiaoming Zhang", "Stan Z. Li"], "authorids": ["~Zhangyang_Gao1", "~Hao_Wang136", "~Cheng_Tan1", "~Chenrui_XU3", "~Mengdi_Liu1", "~Bozhen_Hu1", "~Linlin_Chao1", "~Xiaoming_Zhang2", "~Stan_Z._Li2"], "authors_source": "OpenReview API", "abstract": "This study investigates the current landscape and future directions of protein foundation model research.  While recent advancements have transformed protein science and engineering, the field lacks a comprehensive benchmark for fair evaluation and in-depth understanding. Since ESM-1B, numerous protein foundation models have emerged, each with unique datasets and methodologies. However, evaluations often focus on limited tasks tailored to specific models, hindering insights into broader generalization and limitations. Specifically, researchers struggle to understand the relationships between tasks, assess how well current models perform across them, and determine the criteria in developing new foundation models.  To fill this gap, we present PFMBench, a comprehensive benchmark evaluating protein foundation models across 38 tasks spanning 8 key areas of protein science. Through hundreds of experiments on 17 state-of-the-art models across 38 tasks, PFMBench reveals the inherent correlations between tasks, identifies top-performing models, and provides a streamlined evaluation protocol. Code will be released upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qRSXmGn0sI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14722/Reviewer_SvfA"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces PFMBench, a comprehensive benchmark designed to evaluate Protein Foundation Models (PFMs). The authors argue that existing benchmarks are insufficient as they often cover a limited number of tasks. PFMBench addresses this gap by curating 38 tasks across 8 categories and evaluating 17 state-of-the-art models. Through task-correlation analysis, the authors filter the initial set down to 28 \"core tasks\" (by removing high-bias tasks) and subsequently to 11 \"representative tasks\". This paper evaluates various types of models, including sequence-only models, sequence-structure models, sequence-function models, and sequence-structure-function models. Through extensive experiments, the authors compare the pros and cons of different model types on downstream tasks, the effectiveness of various fine-tuning techniques, and the impact of scaling on model performance.", "review_text": "This paper introduces PFMBench, a comprehensive benchmark designed to evaluate Protein Foundation Models (PFMs). The authors argue that existing benchmarks are insufficient as they often cover a limited number of tasks. PFMBench addresses this gap by curating 38 tasks across 8 categories and evaluating 17 state-of-the-art models. Through task-correlation analysis, the authors filter the initial set down to 28 \"core tasks\" (by removing high-bias tasks) and subsequently to 11 \"representative tasks\". This paper evaluates various types of models, including sequence-only models, sequence-structure models, sequence-function models, and sequence-structure-function models. Through extensive experiments, the authors compare the pros and cons of different model types on downstream tasks, the effectiveness of various fine-tuning techniques, and the impact of scaling on model performance.", "strengths": "1. The benchmark makes comprehensive evaluation that incorporating 38 tasks and 17 models, including many large-scale PFMs (>500M parameters) that were absent in some benchmarks like TAPE.\n2. The proposed PFMBench includes and analyzes multimodal PFMs (e.g., sequence-structure, sequence-function), which is an area that not well studied in previous benchmarks.\n3. The benchmark includes and compares six different PEFT methods (e.g., Adapter, LoRA, DoRA), offering insights into which methods are most effective.", "weaknesses": "1. The evaluation is restricted to \"understanding\" tasks and completely ignores \"generation\" tasks. Generation is a critical capability of modern foundation models and is explicitly supported by several models discussed (e.g., ESM3, DPLM, Progen).\n2. The selection of the \"core models\" relies on performance on a single task: Enzyme Commission (EC) classification. This is a questionable methodology. Although the authors justify this choice in Appendix B.3, the reasoning appears weak. For instance, the authors claim the EC task has low performance bias (0.9%), but it is unclear why a task with even lower (or 0%) bias was not chosen. Regardless, using a single, specific task for model selection is highly likely to introduce selection bias, which could compromise the validity of the subsequent experimental conclusions.\n3. Some of the paper's conclusions are either trivial or have been well-established by prior work: the findings in lines 370-375 that decoder-only models are unsuitable for understanding tasks [1] and that multinomial pLMs outperform sequence-only pLMs [2][3] are already known. Such superficial conclusions offer little new insight to the reader.\n4. The conclusion in line 376 that attributing the lower performance of ESM3 and ProtST to \"noisy or insufficient function data\" lacks reliable evidence. There may be other plausible factors. For example, could this be related to the source and quality of the sequence data, not just the function data? Since ESM3's training also includes structure data, will this have an impact? Also, will the training objectives of ESM3 and ProtST influence the result? The authors do not provide sufficient evidence to support this specific claim.\n5. In Section 4.3, the authors do not justify why only ProtT5, ProTrek, and ESM2 were selected to validate the impact of fine-tuning techniques. It is unclear whether the conclusions drawn from just these three models can be generalized to the broader set of foundation models.\n6. The paper's presentation suffers from a lack of essential experimental details, such as a clear description of the inputs and outputs for each task. Furthermore, the definition and purpose of the novel evaluation metric (Mutual Information Difference) should be presented in the main body of the paper, instead of in the appendix.\n7. The paper's conclusions are largely based on observing and summarizing experimental results, but lacks sufficient insights into the reasons for these results. Providing more in-depth insights into why specific models or fine-tuning techniques perform better would be highly beneficial, as it would offer valuable guidance to readers for developing improved models.\n\n[1] BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding\n\n[2] SaProt: Protein Language Modeling with Structure-aware Vocabulary\n\n[3] ProTrek: Navigating the Protein Universe through Tri-Modal Contrastive Learning", "questions": "See above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PFMBench, a comprehensive benchmark designed to evaluate Protein Foundation Models (PFMs). The authors argue that existing benchmarks are insufficient as they often cover a limited number of tasks. PFMBench addresses this gap by curating 38 tasks across 8 categories and evaluating 17 state-of-the-art models. Through task-correlation analysis, the authors filter the initial set down to 28 \"core tasks\" (by removing high-bias tasks) and subsequently to 11 \"representative tasks\". This paper evaluates various types of models, including sequence-only models, sequence-structure models, sequence-function models, and sequence-structure-function models. Through extensive experiments, the authors compare the pros and cons of different model types on downstream tasks, the effectiveness of various fine-tuning techniques, and the impact of scaling on model performance.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The benchmark makes comprehensive evaluation that incorporating 38 tasks and 17 models, including many large-scale PFMs (>500M parameters) that were absent in some benchmarks like TAPE.\n2. The proposed PFMBench includes and analyzes multimodal PFMs (e.g., sequence-structure, sequence-function), which is an area that not well studied in previous benchmarks.\n3. The benchmark includes and compares six different PEFT methods (e.g., Adapter, LoRA, DoRA), offering insights into which methods are most effective.", "weaknesses": "1. The evaluation is restricted to \"understanding\" tasks and completely ignores \"generation\" tasks. Generation is a critical capability of modern foundation models and is explicitly supported by several models discussed (e.g., ESM3, DPLM, Progen).\n2. The selection of the \"core models\" relies on performance on a single task: Enzyme Commission (EC) classification. This is a questionable methodology. Although the authors justify this choice in Appendix B.3, the reasoning appears weak. For instance, the authors claim the EC task has low performance bias (0.9%), but it is unclear why a task with even lower (or 0%) bias was not chosen. Regardless, using a single, specific task for model selection is highly likely to introduce selection bias, which could compromise the validity of the subsequent experimental conclusions.\n3. Some of the paper's conclusions are either trivial or have been well-established by prior work: the findings in lines 370-375 that decoder-only models are unsuitable for understanding tasks [1] and that multinomial pLMs outperform sequence-only pLMs [2][3] are already known. Such superficial conclusions offer little new insight to the reader.\n4. The conclusion in line 376 that attributing the lower performance of ESM3 and ProtST to \"noisy or insufficient function data\" lacks reliable evidence. There may be other plausible factors. For example, could this be related to the source and quality of the sequence data, not just the function data? Since ESM3's training also includes structure data, will this have an impact? Also, will the training objectives of ESM3 and ProtST influence the result? The authors do not provide sufficient evidence to support this specific claim.\n5. In Section 4.3, the authors do not justify why only ProtT5, ProTrek, and ESM2 were selected to validate the impact of fine-tuning techniques. It is unclear whether the conclusions drawn from just these three models can be generalized to the broader set of foundation models.\n6. The paper's presentation suffers from a lack of essential experimental details, such as a clear description of the inputs and outputs for each task. Furthermore, the definition and purpose of the novel evaluation metric (Mutual Information Difference) should be presented in the main body of the paper, instead of in the appendix.\n7. The paper's conclusions are largely based on observing and summarizing experimental results, but lacks sufficient insights into the reasons for these results. Providing more in-depth insights into why specific models or fine-tuning techniques perform better would be highly beneficial, as it would offer valuable guidance to readers for developing improved models.\n\n[1] BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding\n\n[2] SaProt: Protein Language Modeling with Structure-aware Vocabulary\n\n[3] ProTrek: Navigating the Protein Universe through Tri-Modal Contrastive Learning", "questions": "See above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762281439883}, {"id": "WYXgbzxSTs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14722/Reviewer_1BAp"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes PFMBench, a benchmark for evaluating protein foundation models (PFMs) across extensive tasks and models (including sequence-only, sequence-structure, sequence-function, and multimodal variants). The authors aim to address gaps in existing benchmarks (e.g., TAPE, PEER, Venus) by including more multimodal models, simplifying evaluation via a \"streamlined protocol\" (11 representative tasks, 12 core models, Adapter/DoRA tuning), and analyzing task correlations, zero-shot performance, and parameter-efficient fine-tuning effectiveness. Experiments find that ProTrek (multimodal) outperforms baselines like ESM2, and scaling model size yields limited gains compared to optimizing pretraining strategies.", "review_text": "This paper proposes PFMBench, a benchmark for evaluating protein foundation models (PFMs) across extensive tasks and models (including sequence-only, sequence-structure, sequence-function, and multimodal variants). The authors aim to address gaps in existing benchmarks (e.g., TAPE, PEER, Venus) by including more multimodal models, simplifying evaluation via a \"streamlined protocol\" (11 representative tasks, 12 core models, Adapter/DoRA tuning), and analyzing task correlations, zero-shot performance, and parameter-efficient fine-tuning effectiveness. Experiments find that ProTrek (multimodal) outperforms baselines like ESM2, and scaling model size yields limited gains compared to optimizing pretraining strategies.", "strengths": "Task and model coverage breadth: PFMBench aggregates 38 tasks (more than prior benchmarks like Venus’s 22) and includes 17 models (vs. Venus’s 3), covering multimodal PFMs (e.g., ESM3, ProTrek) often omitted in existing benchmarks. This breadth could, in theory, offer a more comprehensive view of PFM performance.", "weaknesses": "1. Benchmark design is incremental, not transformative. PFMBench merely expands the number of tasks/models from prior work (e.g., TAPE→PEER→Venus→PFMBench) without introducing new evaluation paradigms, metrics, or task designs. For example:\n    - The 38 tasks are all existing (e.g., Enzyme Commission classification, ProteinGym zero-shot) with no novel tasks that test understudied PFM capabilities (e.g., functional cross-species generalization, de novo design validation).\n    - The \"streamlined protocol\" (filtering 11 representative tasks via correlation) is a trivial application of Spearman correlation, a standard practice in benchmarking (e.g., GLUE for NLP), and provides no new framework for task selection.\n    - No new evaluation metrics: PFMBench relies entirely on existing metrics (F1, AUROC, Spearman, Top L/5) and adds only \"Mutual Information Difference (MID)\"—a minor variant of mutual information (McAllester & Stratos, 2020) with no validation that it better reflects PFM quality than standard metrics.\n\n2. Key conclusions are trivial or already known. For example, \"multimodal models outperform sequence-only models\": this has been established by ESM3. PFMBench merely confirms this with more tasks, adding no new insight; \"Scaling model size yields limited gains\": ESM2’s scaling analysis already showed diminishing returns beyond 650M parameters. PFMBench’s ESM2 series results are redundant.", "questions": "n/a", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes PFMBench, a benchmark for evaluating protein foundation models (PFMs) across extensive tasks and models (including sequence-only, sequence-structure, sequence-function, and multimodal variants). The authors aim to address gaps in existing benchmarks (e.g., TAPE, PEER, Venus) by including more multimodal models, simplifying evaluation via a \"streamlined protocol\" (11 representative tasks, 12 core models, Adapter/DoRA tuning), and analyzing task correlations, zero-shot performance, and parameter-efficient fine-tuning effectiveness. Experiments find that ProTrek (multimodal) outperforms baselines like ESM2, and scaling model size yields limited gains compared to optimizing pretraining strategies.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Task and model coverage breadth: PFMBench aggregates 38 tasks (more than prior benchmarks like Venus’s 22) and includes 17 models (vs. Venus’s 3), covering multimodal PFMs (e.g., ESM3, ProTrek) often omitted in existing benchmarks. This breadth could, in theory, offer a more comprehensive view of PFM performance.", "weaknesses": "1. Benchmark design is incremental, not transformative. PFMBench merely expands the number of tasks/models from prior work (e.g., TAPE→PEER→Venus→PFMBench) without introducing new evaluation paradigms, metrics, or task designs. For example:\n    - The 38 tasks are all existing (e.g., Enzyme Commission classification, ProteinGym zero-shot) with no novel tasks that test understudied PFM capabilities (e.g., functional cross-species generalization, de novo design validation).\n    - The \"streamlined protocol\" (filtering 11 representative tasks via correlation) is a trivial application of Spearman correlation, a standard practice in benchmarking (e.g., GLUE for NLP), and provides no new framework for task selection.\n    - No new evaluation metrics: PFMBench relies entirely on existing metrics (F1, AUROC, Spearman, Top L/5) and adds only \"Mutual Information Difference (MID)\"—a minor variant of mutual information (McAllester & Stratos, 2020) with no validation that it better reflects PFM quality than standard metrics.\n\n2. Key conclusions are trivial or already known. For example, \"multimodal models outperform sequence-only models\": this has been established by ESM3. PFMBench merely confirms this with more tasks, adding no new insight; \"Scaling model size yields limited gains\": ESM2’s scaling analysis already showed diminishing returns beyond 650M parameters. PFMBench’s ESM2 series results are redundant.", "questions": "n/a", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762272215518}, {"id": "P5ASIZZKU9", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14722/Reviewer_SFzi"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This work proposes PFMBench, a comprehensive benchmark for Protein Foundation Models (PFMs). The authors evaluate 17 state-of-the-art PFMs across 38 tasks spanning 8 key areas of protein science, and develop a protocol that filters the extensive testbed down to 11 representative tasks, 12 core models, 2 recommended baselines, and 2 recommended PEFT methods.", "review_text": "This work proposes PFMBench, a comprehensive benchmark for Protein Foundation Models (PFMs). The authors evaluate 17 state-of-the-art PFMs across 38 tasks spanning 8 key areas of protein science, and develop a protocol that filters the extensive testbed down to 11 representative tasks, 12 core models, 2 recommended baselines, and 2 recommended PEFT methods.", "strengths": "1. This work is the first to systematically incorporate and evaluate PFMs that leverage sequence, structure, and functional data modalities. The focus on a \"streamlined protocol\" is the paper's main contribution.\n2. The authors move beyond simple leaderboards to provide actionable insights. The comparison between ProTrek, ESM3, and ProtST, for example, generates a valuable hypothesis about why multimodal models succeed or fail.\n3. The paper provides a large-scale, controlled comparison for PFMs. The experimental design is sound. The hierarchical filtering methodology is well-justified.", "weaknesses": "1. The most significant omission is the complete absence of generative tasks. PFMs are not just used for \"understanding\" (discriminative tasks) but increasingly for \"creation\" (generative tasks) like inverse folding, de novo backbone design, and sequence generation. By focusing only on 38 discriminative tasks, the benchmark overlooks a massive and critical component of PFM capabilities.\n2. The protein structures are from AFDB, which may introduce biases into the evaluation.\n3. Fair comparison between protein foundation models seems impossible, considering the different pretraining data used in various models. Since the test dataset might have an overlap with the pretraining dataset.\n4. This submission is titled \"Protein Foundation Model Benchmark\", but it missed some of the related works. For example, many works that belong to the \"Seq-Struct\" category are not included.\n5. As a benchmark, no codes are provided for review. Thus, I am not sure about the quality of the benchmark code.", "questions": "1. Following on the weakness above: What was the rationale for excluding generative tasks, such as protein design, from the 8 \"key areas of protein science\" surveyed in this benchmark?\n2. For the conclusion that ProTrek's success stems from \"effective semantic alignment\" while ESM3's failure is due to \"noisy or insufficient function data\". However, other multimodal PFMs use similar alignment techniques. Could the authors elaborate on what makes ProTrek's strategy uniquely effective compared to its multimodal peers? Is it purely a matter of data quality, or are there unexamined architectural or objective function differences that are more critical?\n3. The authors cite \"Appendix A.5\" (Line 256) for detailed reasons, but this appendix appears to be missing. Furthermore, the appendix numbering seems inconsistent with the main text (Line 420 refers to A.4). Could the authors please provide the missing justification, correct the numbering, and include supporting data to demonstrate that EC classification performance is a reliable proxy for a model's general capabilities across the other 27 core tasks?\n4. Is the data copyright clarified?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes PFMBench, a comprehensive benchmark for Protein Foundation Models (PFMs). The authors evaluate 17 state-of-the-art PFMs across 38 tasks spanning 8 key areas of protein science, and develop a protocol that filters the extensive testbed down to 11 representative tasks, 12 core models, 2 recommended baselines, and 2 recommended PEFT methods.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This work is the first to systematically incorporate and evaluate PFMs that leverage sequence, structure, and functional data modalities. The focus on a \"streamlined protocol\" is the paper's main contribution.\n2. The authors move beyond simple leaderboards to provide actionable insights. The comparison between ProTrek, ESM3, and ProtST, for example, generates a valuable hypothesis about why multimodal models succeed or fail.\n3. The paper provides a large-scale, controlled comparison for PFMs. The experimental design is sound. The hierarchical filtering methodology is well-justified.", "weaknesses": "1. The most significant omission is the complete absence of generative tasks. PFMs are not just used for \"understanding\" (discriminative tasks) but increasingly for \"creation\" (generative tasks) like inverse folding, de novo backbone design, and sequence generation. By focusing only on 38 discriminative tasks, the benchmark overlooks a massive and critical component of PFM capabilities.\n2. The protein structures are from AFDB, which may introduce biases into the evaluation.\n3. Fair comparison between protein foundation models seems impossible, considering the different pretraining data used in various models. Since the test dataset might have an overlap with the pretraining dataset.\n4. This submission is titled \"Protein Foundation Model Benchmark\", but it missed some of the related works. For example, many works that belong to the \"Seq-Struct\" category are not included.\n5. As a benchmark, no codes are provided for review. Thus, I am not sure about the quality of the benchmark code.", "questions": "1. Following on the weakness above: What was the rationale for excluding generative tasks, such as protein design, from the 8 \"key areas of protein science\" surveyed in this benchmark?\n2. For the conclusion that ProTrek's success stems from \"effective semantic alignment\" while ESM3's failure is due to \"noisy or insufficient function data\". However, other multimodal PFMs use similar alignment techniques. Could the authors elaborate on what makes ProTrek's strategy uniquely effective compared to its multimodal peers? Is it purely a matter of data quality, or are there unexamined architectural or objective function differences that are more critical?\n3. The authors cite \"Appendix A.5\" (Line 256) for detailed reasons, but this appendix appears to be missing. Furthermore, the appendix numbering seems inconsistent with the main text (Line 420 refers to A.4). Could the authors please provide the missing justification, correct the numbering, and include supporting data to demonstrate that EC classification performance is a reliable proxy for a model's general capabilities across the other 27 core tasks?\n4. Is the data copyright clarified?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None.", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761994208759}, {"id": "Jj8a4gvm6x", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14722/Reviewer_yCpT"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This work introduces a new benchmark for predicting various protein-related attributes using protein foundational models. The tasks considered are broad and include 38 tasks, such as annotation, mutation, structure, and zero-shot learning. This work adds to an expanding collection of benchmarking tools for protein foundational models.\nEditorial Issues: \n- There were multiple references to the appendix, but the appendix was not included in the submission.\n- Fig 3 lacks a proper legend; including the significance of X and stars would be helpful. Currently, it is only described in the text.", "review_text": "This work introduces a new benchmark for predicting various protein-related attributes using protein foundational models. The tasks considered are broad and include 38 tasks, such as annotation, mutation, structure, and zero-shot learning. This work adds to an expanding collection of benchmarking tools for protein foundational models.\nEditorial Issues: \n- There were multiple references to the appendix, but the appendix was not included in the submission.\n- Fig 3 lacks a proper legend; including the significance of X and stars would be helpful. Currently, it is only described in the text.", "strengths": "- The tasks considered span a wide range. This is beneficial for future work to compare. \n- The benchmark comparison was done using multiple PEFT methods—mainly Adapter and DoRA. This allows examining the hidden potential of PFMs and not strictly limiting it by its training objectives. \n- The inclusion of multi-modal models such as ESM3, is helpful, as these are recent models.\n- Assuming the source code for this benchmarking will be made open, it would be a valuable resource, and due to its claimed modular design (which cannot be evaluated currently), it has the potential to significantly aid future work.", "weaknesses": "- While the datasets/tasks benchmark works are valuable for advancing AI and are critical for developing the next generation of AI models, the work does not provide novel innovation or understanding of AI.\n- A wide range of tasks was considered, but the motivation for including such a broad number of tasks is not explained.\n- There is no measure of the quality of the tasks. For example, gene annotations could be noisy, and additional efforts might be necessary to ensure the quality of these datasets.\n- The paper relies heavily on tabular comparisons and win rates (#Win) without deeper qualitative analysis.", "questions": "- The analysis or discussion of the results from UMAP is missing. Only three selected UMAP plots are provided, but more than 30 tasks and 17 models were considered. What about the results of other models and tasks ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces a new benchmark for predicting various protein-related attributes using protein foundational models. The tasks considered are broad and include 38 tasks, such as annotation, mutation, structure, and zero-shot learning. This work adds to an expanding collection of benchmarking tools for protein foundational models.\nEditorial Issues: \n- There were multiple references to the appendix, but the appendix was not included in the submission.\n- Fig 3 lacks a proper legend; including the significance of X and stars would be helpful. Currently, it is only described in the text.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The tasks considered span a wide range. This is beneficial for future work to compare. \n- The benchmark comparison was done using multiple PEFT methods—mainly Adapter and DoRA. This allows examining the hidden potential of PFMs and not strictly limiting it by its training objectives. \n- The inclusion of multi-modal models such as ESM3, is helpful, as these are recent models.\n- Assuming the source code for this benchmarking will be made open, it would be a valuable resource, and due to its claimed modular design (which cannot be evaluated currently), it has the potential to significantly aid future work.", "weaknesses": "- While the datasets/tasks benchmark works are valuable for advancing AI and are critical for developing the next generation of AI models, the work does not provide novel innovation or understanding of AI.\n- A wide range of tasks was considered, but the motivation for including such a broad number of tasks is not explained.\n- There is no measure of the quality of the tasks. For example, gene annotations could be noisy, and additional efforts might be necessary to ensure the quality of these datasets.\n- The paper relies heavily on tabular comparisons and win rates (#Win) without deeper qualitative analysis.", "questions": "- The analysis or discussion of the results from UMAP is missing. Only three selected UMAP plots are provided, but more than 30 tasks and 17 models were considered. What about the results of other models and tasks ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761978329717}], "openreview_url": "https://openreview.net/forum?id=KJNgtPNxtv", "arxiv_id": "2506.14796", "paper_pdf": "papers/KJNgtPNxtv.pdf", "paper_pdf_sha256": "a1f3e770e11eb3898ddf5a8d939558beddc2c256f4cfe4e08e437382c3748bc9", "paper_pdf_bytes": 1891782, "paper_pdf_source": "openreview", "code_url": "https://github.com/biomap-research/PFMBench", "code_repository": "biomap-research/PFMBench", "code_commit": "53758ffcbdf1d79b5d125383e4dd52d6fd59d2a1", "code_archive": "repos/KJNgtPNxtv.zip", "code_archive_sha256": "16607a6cff798f162961d6a3f0f8254cfbf6e8bc28b4ff45b5a97f846e00dfac", "code_archive_bytes": 3172484, "code_file_count": 181, "code_extensions": {".py": 180, ".sh": 1}, "github_disk_usage_kb": 3107, "github_languages": {"Python": 1495687, "Shell": 9029}, "github_archived": false, "github_pushed_at": "2025-08-01T05:25:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pfmbench-protein-foundation-model-benchmark"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NEEtm5laNK1", "year": 2023, "status": "rejected", "title": "CHiLS: Zero-Shot Image Classification with Hierarchical Label Sets", "authors": ["Zachary Novack", "Saurabh Garg", "Zachary Chase Lipton"], "authorids": ["~Zachary_Novack1", "~Saurabh_Garg3", "~Zachary_Chase_Lipton1"], "authors_source": "OpenReview API", "abstract": "Open vocabulary models (e.g. CLIP) have shown strong performance on zeroshot classification through their ability generate embeddings for each class based on their (natural language) names. Prior work has focused on improving the accuracy of these models through prompt engineering or by incorporating a small amount of labeled downstream data (via finetuning). In this paper, we propose Classification with Hierarchical Label Sets (or CHiLS), an alternative strategy that proceeds in three steps: (i) for each class, produce a set of subclasses, using either existing label hierarchies or by querying GPT-3; (ii) perform the standard zero-shot CLIP procedure as though these subclasses were the labels of interest; (iii) map the predicted subclass back to its parent to produce the final prediction. Across numerous datasets, CHiLS leads to improved accuracy yielding gains of over 30% in situations where known hierarchies are available and more modest gains when they are not. CHiLS is simple to implement within existing CLIP pipelines and requires no additional training cost.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "p9ls4nw52c", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5549/Reviewer_VtPj"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, authors propose to utilize the label hierarchies to boost the performance of CLIP for zero shot classification. The main steps include the generation of the subclasses for each class by either using the GT label hierarchies or by querying GPT-3, then conduct the CLIP  via these sub-classes, and finally map the sub-classes back to their parents. The performance gain has been observed on the benchmark datasets.", "review_text": "Overall, the paper is well-written, and easy to follow, yet the main concern as expressed in the weakness part is the novelty, as it resembles a lot with an ensemble model.\n\nMoreover, if the hierarchical structure does not explicitly exist in the label vocabulary, how would the proposed model handle it, namely, you cannot rely on the ready-to-go GT structure or query it via the GPT model. It is somewhat unfair if the compared models can also utilize these structures to boost their performance. Authors are suggested to clarify their proposed model in a more general sense.", "strengths": "Strength:\nThe paper is generally well-written with a clear motivation and decent performance gain.\n\nWeakness:\nThe main concern would be the proposed model is essentially an ensemble of sub-class based CLIP models, though authors address this concern via comparisons against linear average approach, the novelty is still somewhat limited.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, authors propose to utilize the label hierarchies to boost the performance of CLIP for zero shot classification. The main steps include the generation of the subclasses for each class by either using the GT label hierarchies or by querying GPT-3, then conduct the CLIP  via these sub-classes, and finally map the sub-classes back to their parents. The performance gain has been observed on the benchmark datasets.", "strength_and_weaknesses": "Strength:\nThe paper is generally well-written with a clear motivation and decent performance gain.\n\nWeakness:\nThe main concern would be the proposed model is essentially an ensemble of sub-class based CLIP models, though authors address this concern via comparisons against linear average approach, the novelty is still somewhat limited.", "clarity,_quality,_novelty_and_reproducibility": "The organization of the paper and the overall writing is clear. And one should be able to reproduce with the firm understanding of the CLIP model.", "summary_of_the_review": "Overall, the paper is well-written, and easy to follow, yet the main concern as expressed in the weakness part is the novelty, as it resembles a lot with an ensemble model.\n\nMoreover, if the hierarchical structure does not explicitly exist in the label vocabulary, how would the proposed model handle it, namely, you cannot rely on the ready-to-go GT structure or query it via the GPT model. It is somewhat unfair if the compared models can also utilize these structures to boost their performance. Authors are suggested to clarify their proposed model in a more general sense.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666785469230}, {"id": "ug7ETT1YYH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5549/Reviewer_ohpn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a method to improve predictions of the CLIP model by utilizing a potential structure/hierarchy among classes. Instead of generating predictions on the target classes only, the authors propose to generate predictions on a set of all target labels’ subclasses (which are more granular) and use predictions in subclass space to decide which target class (superclass) to output.  \nFor finding subclasses the authors use either use an existing hierarchy of labels or generate them by prompting a GPT-3 model.", "review_text": "Not so clear how general or practically useful the method is: how specific the results are to datasets/problems with some class structure (whether explicit or implicit). The most significant improvements are when utilizing an existing hierarchical structure of labels but it’s unclear how much of that improvement is specific to the proposed method as no alternative approaches or simple baselines utilizing such information are considered.\n\n**EDIT: Updated the score to lean towards acceptance. The authors' comments are generally convincing and the updated version of the paper makes the scope of the contributions clearer.\nAlso, the additional results provide more insight into the model's performance and in which scenarios it is expected to work well.**", "strengths": "Strengths:\n\n- (S1) A practical and easy-to-implement/utilize method\n- (S2) In the case of existing class hierarchy, the experiments indicate a very significant improvement over not relying on that class hierarchy (although see W2)\n- (S2) The experiments cover w/ and w/o hierarchical explicit information (although see W1)\n    \n\nWeaknesses:\n- (W1) The paper doesn’t clearly show whether the proposed approach is valid/useful only on datasets where some hierarchical structure between classes exist (regardless whether explicitly present or not) or is something that would work on any generic problem/datasets. The authors do address in their experiments two scenarios: one with hierarchical information provided explicitly and another with a hierarchy generated from a GPT model. However, a good class structure/hierarchy may or may not exist even if it is not explicitly provided or utilized. Taking ImageNet as an example - there are many animal/plant species with a very deep hierarchy, many closely related classes on one hand, but on the other hand, some classes have a very shallow hierarchy and few only loosely related classes (I would suppose maybe classes like “cliff” or “traffic sign”?). The authors however seem to choose datasets where one could expect some hierarchy to exist, even if not explicitly present. How would the method behave on datasets like e.g. StanfordCars where all classes are somewhat similar and it’s unclear if some meaningful hierarchy exists, or even the whole ImageNet where maybe there are groups of classes with a nice hierarchy/structure and groups of classes where such hierarchy might not exist?\n- (W2) The authors present results for their method but no comparisons to any alternative approaches. One would expect at least some simple baselines that utilize the hierarchical structure of labels.\n- (W3) On many datasets the improvement in accuracy is very significant, but on some other, like “living17” or “fruits-360” there’s a relatively much smaller improvement. This is not explained/discussed by the authors - is it something related to some properties of the hierarchies?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a method to improve predictions of the CLIP model by utilizing a potential structure/hierarchy among classes. Instead of generating predictions on the target classes only, the authors propose to generate predictions on a set of all target labels’ subclasses (which are more granular) and use predictions in subclass space to decide which target class (superclass) to output.  \nFor finding subclasses the authors use either use an existing hierarchy of labels or generate them by prompting a GPT-3 model.", "strength_and_weaknesses": "Strengths:\n\n- (S1) A practical and easy-to-implement/utilize method\n- (S2) In the case of existing class hierarchy, the experiments indicate a very significant improvement over not relying on that class hierarchy (although see W2)\n- (S2) The experiments cover w/ and w/o hierarchical explicit information (although see W1)\n    \n\nWeaknesses:\n- (W1) The paper doesn’t clearly show whether the proposed approach is valid/useful only on datasets where some hierarchical structure between classes exist (regardless whether explicitly present or not) or is something that would work on any generic problem/datasets. The authors do address in their experiments two scenarios: one with hierarchical information provided explicitly and another with a hierarchy generated from a GPT model. However, a good class structure/hierarchy may or may not exist even if it is not explicitly provided or utilized. Taking ImageNet as an example - there are many animal/plant species with a very deep hierarchy, many closely related classes on one hand, but on the other hand, some classes have a very shallow hierarchy and few only loosely related classes (I would suppose maybe classes like “cliff” or “traffic sign”?). The authors however seem to choose datasets where one could expect some hierarchy to exist, even if not explicitly present. How would the method behave on datasets like e.g. StanfordCars where all classes are somewhat similar and it’s unclear if some meaningful hierarchy exists, or even the whole ImageNet where maybe there are groups of classes with a nice hierarchy/structure and groups of classes where such hierarchy might not exist?\n- (W2) The authors present results for their method but no comparisons to any alternative approaches. One would expect at least some simple baselines that utilize the hierarchical structure of labels.\n- (W3) On many datasets the improvement in accuracy is very significant, but on some other, like “living17” or “fruits-360” there’s a relatively much smaller improvement. This is not explained/discussed by the authors - is it something related to some properties of the hierarchies?", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and proposes a simple, yet conceptually novel approach to utilize label hierarchy.\nThe authors provide code and most implementation details seem to be available.", "summary_of_the_review": "Not so clear how general or practically useful the method is: how specific the results are to datasets/problems with some class structure (whether explicit or implicit). The most significant improvements are when utilizing an existing hierarchical structure of labels but it’s unclear how much of that improvement is specific to the proposed method as no alternative approaches or simple baselines utilizing such information are considered.\n\n**EDIT: Updated the score to lean towards acceptance. The authors' comments are generally convincing and the updated version of the paper makes the scope of the contributions clearer.\nAlso, the additional results provide more insight into the model's performance and in which scenarios it is expected to work well.**", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666699127471}, {"id": "qVw9aujdhEq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5549/Reviewer_oLxZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies zero-shot image classification with class hierarchy.\nA Classification with Hierarchical Label sets (CHiLS) model is proposed to improved classical CLIP model.\nThis model leverages predefined subclasses, and perform CLIP on them first to obtain a set of class embeddings.\nThese subclass embeddings are aggregated together to form the embedding of class of interests that facilitate zero-shot image classification.\n\n", "review_text": "The paper improves the previous arts CLIP by leveraging class hierarchy, which can be seen as a modest incremental work.\nUnfortunately, the reviewer does not find it is novel enough.\nBesides, the writing and empirical study can be further improved. \n", "strengths": "\n- The outstanding problem of this paper is that it is of not self-contained. It lacks of proper definition and review of referred models, terminologies, and problem setup. For example, \"CLIP\", \"prompt\", \"open vocabulary models\".\n\n- The idea of dividing a super class into a set of subclass is not new, and therefore the contribution is not significant enough.\n\n- The sensitivity of the proposed model on the different levels of granularity of the class hierarchy is unclear.\n\n- Lacks of proper comparisons with related works in line of zero-shot image classification.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies zero-shot image classification with class hierarchy.\nA Classification with Hierarchical Label sets (CHiLS) model is proposed to improved classical CLIP model.\nThis model leverages predefined subclasses, and perform CLIP on them first to obtain a set of class embeddings.\nThese subclass embeddings are aggregated together to form the embedding of class of interests that facilitate zero-shot image classification.\n\n", "strength_and_weaknesses": "\n- The outstanding problem of this paper is that it is of not self-contained. It lacks of proper definition and review of referred models, terminologies, and problem setup. For example, \"CLIP\", \"prompt\", \"open vocabulary models\".\n\n- The idea of dividing a super class into a set of subclass is not new, and therefore the contribution is not significant enough.\n\n- The sensitivity of the proposed model on the different levels of granularity of the class hierarchy is unclear.\n\n- Lacks of proper comparisons with related works in line of zero-shot image classification.", "clarity,_quality,_novelty_and_reproducibility": "In terms of clarity, this paper can be improved further by providing proper introduction of the problem formulation, related works. \n\nThe proposed method looks pretty straightforward. \nWhile it is interesting to see the improvement of performance, it is only a modest incremental work over the arts.", "summary_of_the_review": "The paper improves the previous arts CLIP by leveraging class hierarchy, which can be seen as a modest incremental work.\nUnfortunately, the reviewer does not find it is novel enough.\nBesides, the writing and empirical study can be further improved. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666599861433}, {"id": "CSnO-XhaFf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5549/Reviewer_msk3"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper provides an algorithm to improve zero-shot open-vocabulary classifiers by better prompting them. The authors propose to (1) create sub-classes for each parent classification category using existing (human-created) or inferred (GPT-3 generated) label hierarchies, (2) perform standard zero-shot classification on the sub-classes, and (3) aggregate the results to associate probabilities with the parent category. The paper provides empirical evaluation of the method on various datasets and additionally ablates key design decisions related to their use of labels, GPT-3, and aggregation strategy.", "review_text": "The paper provides a method to prompt open-vocabulary models for zero-shot image classification, which leverages class hierarchies. It is straightforward to implement and the authors provide some empirical evidence that their method performs significantly better than the baseline, especially when hierarchies are known for a dataset.\n\nHowever, I currently recommend weak rejection of the manuscript. I am most concerned about the lack of benchmarking on ImageNet (*W3.*), the performance of the method on at least a few more standard classification datasets (*W4.*), the fairness of the zero-shot baselines (*W5.*), and the clarity of the presentation associated with some of the ablations studies (*W7.*, *W10.*).\n\nI am willing to revisit my evaluation during the rebuttal/discussion period.", "strengths": "**Strengths**\n\n*S1.* The work is well motivated.\n \n*S2.* The work seems novel. I have only seen GPT-3 used in zero-shot open-vocab classification in concurrent work and consider this to be novel.\n\n*S3.* The method is general for classification problems and does not require additional training.\n\n*S4.* The ablations test major design decisions and shed light on when a practitioner may or may not want to make similar decisions. \n\n*S5.* Method is simple and easy to implement.\n\n*S6.* The introduction and method section are very well written.\n\n**Weaknesses**\n\n*W1.* Related work could be more comprehensive and the paper could be better situated in the literature. More specifically, here are some suggestions for references. I suggest doing a more comprehensive literature review of recent work.\n\nTransfer learning:\n* LiT: tuning text towers while keeping language tower fixed (https://arxiv.org/abs/2111.07991)\n* Visual prompt tuning: tuning learnable visual prompt (https://arxiv.org/abs/2203.12119)\n* Model Soups: ensembling many fine-tuned CLIP models in weight space (https://arxiv.org/abs/2203.05482)\n* Patching CLIP models on downstream tasks while maintaining zero-shot performance on tasks where CLIP is already successful (https://arxiv.org/abs/2208.05592)\n* CLIP-CL: similar to Model Patching above (https://arxiv.org/abs/2207.09248)\n* ...\n\nZero-shot prediction:\n* CLIP-ViL: showing the limitation of ZS CLIP in VQA-like settings (https://arxiv.org/abs/2107.06383)\n* CoW: showing how CLIP can be adapted to do object navigation without additional training (https://arxiv.org/abs/2203.10421)\n* ...\n\nAdditionally, I suggest adding a section on hierarchical classification.\n\n*W2.* The paper states: \"Here, we reweight each set of subclass probabilities by its superclass probability. In this way, we can attempt to avoid the behavior in which CLIP makes an incorrect subclass prediction despite a confident and correct superclass prediction, and thus bias our model to never do worse than the raw superclass predictions.\"\n\nWhile this is maybe meant to provide intuition for the method, it reads like a claim that is not supported. I suggest making it clear that this is intuition and not some mathematical/provable property of the algorithm related to \"bias\".\n\n*W3.* How does the proposed method perform on ImageNet? Without ImageNet evaluation it becomes hard to compare the proposed method to other methods.\n\n*W4.* It seems many of the datasets are chosen because it is easy to construct class hierarchies. However, the method is presented as a general purpose zero-shot image classification improvement over CLIP. To be convinced of this, I would be interested to see results on more datasets. Please consider Cars, DTD, EuroSAT, GTSRB, KITTI, MNIST, RESISC45, SUN397, which Ilharco et al. (https://arxiv.org/abs/2208.05592) use because they are tasks ZS CLIP is known to struggle on. Note: dataloaders for these datasets can be found [here](https://github.com/mlfoundations/patching/tree/main/src/datasets).\n\n*W5.* The main experimental setup has baselines that may not be fair points of comparison. Specifically, not all datasets in the original CLIP paper use the ImageNet prompt templates. For example, CIFAR-10/Food101 use different prompts (see [here](https://github.com/openai/CLIP/blob/e184f608c5d5e58165682f7c332c3a8b4c1545f2/data/prompts.md)). What are the deltas between the 75 prompt set and the set that OpenAI provides? For Food101 specifically, there seems to be a discrepancy of ~2% between the ZS number reported in the manuscript and the L/14-336px number reported in the CLIP paper in Tab 11. For datasets that OpenAI did not evaluate in their original paper, some prompt engineering on a val set or some validation that ImageNet prompts are reasonable seems needed.\n \n*W6.* It seems that the experimental setup is sufficiently different in cases where the hierarchy exists and when it does not for me to compare the performance between col 2 and 3 in Table 1. For example, in the *GPT-3 map* the authors include the superclass label in the label set, while in *Existing Map* they do not. It would be good to standardize the algorithm being compared in the two settings. Another idea is to provide the full results in the appendix (i.e., *GPT-3 map* w/ superclass labels, *GPT-3 map* w/o superclass labels, *Existing map* w/ superclass labels, *Existing map* w/o superclass labels).\n\n*W7.* I am not able to completely understand the experimental setup for “Noisy Available Hierarchies” from the text provided. Consider adding an appendix for that section to give more details on how the subclasses are constructed from the ImageNet hierarchy.\n\n*W8.* How do things look if the label set size is 1 or 100 (i.e., m=1 or m=100)? Doe these more extreme values of m affect the performance? At least m=1 should be worse than zero-shot, which seems like a valuable bases of comparison.\n\n*W9.* Some visualizations of cases where the vanilla zero-shot and CHiLS models disagree could be helpful to provide intuition.\n\n*W10.* I am not clear on the linear class ensemble experiment from reading the text. Is the following correct? For a single class, loop over all subclasses and ImageNet prompt templates computing the CLIP text features. Average all of these features to represent the single class as a feature. Repeat for all classes to get a zero-shot classification head.\n\n*W11.* The method involves expanding the zero-shot head at test time. There is a compute and memory overhead associated with this that may limit the scalability of the method when many classes or subclasses are targeted. For example, for the 1000 ImageNet classes, with m=10, the instantiated head would have 10k classes.\n\n**Minor**\n\n*M1.* The Radford et al. 2021 caption example is “a photo of a {}.” not “a photo of a {}” as presented in the manuscript. The difference is the period at the end of the prompt, which may make a difference in downstream performance.\n\n*M2.* This is a relevant reference when discussing CLIP confidence in the method section: https://arxiv.org/abs/2106.07998\n\n*M3.* For Sec. 4.4 is there some theory as to why weighting is necessary in one case and not in the other?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper provides an algorithm to improve zero-shot open-vocabulary classifiers by better prompting them. The authors propose to (1) create sub-classes for each parent classification category using existing (human-created) or inferred (GPT-3 generated) label hierarchies, (2) perform standard zero-shot classification on the sub-classes, and (3) aggregate the results to associate probabilities with the parent category. The paper provides empirical evaluation of the method on various datasets and additionally ablates key design decisions related to their use of labels, GPT-3, and aggregation strategy.", "strength_and_weaknesses": "**Strengths**\n\n*S1.* The work is well motivated.\n \n*S2.* The work seems novel. I have only seen GPT-3 used in zero-shot open-vocab classification in concurrent work and consider this to be novel.\n\n*S3.* The method is general for classification problems and does not require additional training.\n\n*S4.* The ablations test major design decisions and shed light on when a practitioner may or may not want to make similar decisions. \n\n*S5.* Method is simple and easy to implement.\n\n*S6.* The introduction and method section are very well written.\n\n**Weaknesses**\n\n*W1.* Related work could be more comprehensive and the paper could be better situated in the literature. More specifically, here are some suggestions for references. I suggest doing a more comprehensive literature review of recent work.\n\nTransfer learning:\n* LiT: tuning text towers while keeping language tower fixed (https://arxiv.org/abs/2111.07991)\n* Visual prompt tuning: tuning learnable visual prompt (https://arxiv.org/abs/2203.12119)\n* Model Soups: ensembling many fine-tuned CLIP models in weight space (https://arxiv.org/abs/2203.05482)\n* Patching CLIP models on downstream tasks while maintaining zero-shot performance on tasks where CLIP is already successful (https://arxiv.org/abs/2208.05592)\n* CLIP-CL: similar to Model Patching above (https://arxiv.org/abs/2207.09248)\n* ...\n\nZero-shot prediction:\n* CLIP-ViL: showing the limitation of ZS CLIP in VQA-like settings (https://arxiv.org/abs/2107.06383)\n* CoW: showing how CLIP can be adapted to do object navigation without additional training (https://arxiv.org/abs/2203.10421)\n* ...\n\nAdditionally, I suggest adding a section on hierarchical classification.\n\n*W2.* The paper states: \"Here, we reweight each set of subclass probabilities by its superclass probability. In this way, we can attempt to avoid the behavior in which CLIP makes an incorrect subclass prediction despite a confident and correct superclass prediction, and thus bias our model to never do worse than the raw superclass predictions.\"\n\nWhile this is maybe meant to provide intuition for the method, it reads like a claim that is not supported. I suggest making it clear that this is intuition and not some mathematical/provable property of the algorithm related to \"bias\".\n\n*W3.* How does the proposed method perform on ImageNet? Without ImageNet evaluation it becomes hard to compare the proposed method to other methods.\n\n*W4.* It seems many of the datasets are chosen because it is easy to construct class hierarchies. However, the method is presented as a general purpose zero-shot image classification improvement over CLIP. To be convinced of this, I would be interested to see results on more datasets. Please consider Cars, DTD, EuroSAT, GTSRB, KITTI, MNIST, RESISC45, SUN397, which Ilharco et al. (https://arxiv.org/abs/2208.05592) use because they are tasks ZS CLIP is known to struggle on. Note: dataloaders for these datasets can be found [here](https://github.com/mlfoundations/patching/tree/main/src/datasets).\n\n*W5.* The main experimental setup has baselines that may not be fair points of comparison. Specifically, not all datasets in the original CLIP paper use the ImageNet prompt templates. For example, CIFAR-10/Food101 use different prompts (see [here](https://github.com/openai/CLIP/blob/e184f608c5d5e58165682f7c332c3a8b4c1545f2/data/prompts.md)). What are the deltas between the 75 prompt set and the set that OpenAI provides? For Food101 specifically, there seems to be a discrepancy of ~2% between the ZS number reported in the manuscript and the L/14-336px number reported in the CLIP paper in Tab 11. For datasets that OpenAI did not evaluate in their original paper, some prompt engineering on a val set or some validation that ImageNet prompts are reasonable seems needed.\n \n*W6.* It seems that the experimental setup is sufficiently different in cases where the hierarchy exists and when it does not for me to compare the performance between col 2 and 3 in Table 1. For example, in the *GPT-3 map* the authors include the superclass label in the label set, while in *Existing Map* they do not. It would be good to standardize the algorithm being compared in the two settings. Another idea is to provide the full results in the appendix (i.e., *GPT-3 map* w/ superclass labels, *GPT-3 map* w/o superclass labels, *Existing map* w/ superclass labels, *Existing map* w/o superclass labels).\n\n*W7.* I am not able to completely understand the experimental setup for “Noisy Available Hierarchies” from the text provided. Consider adding an appendix for that section to give more details on how the subclasses are constructed from the ImageNet hierarchy.\n\n*W8.* How do things look if the label set size is 1 or 100 (i.e., m=1 or m=100)? Doe these more extreme values of m affect the performance? At least m=1 should be worse than zero-shot, which seems like a valuable bases of comparison.\n\n*W9.* Some visualizations of cases where the vanilla zero-shot and CHiLS models disagree could be helpful to provide intuition.\n\n*W10.* I am not clear on the linear class ensemble experiment from reading the text. Is the following correct? For a single class, loop over all subclasses and ImageNet prompt templates computing the CLIP text features. Average all of these features to represent the single class as a feature. Repeat for all classes to get a zero-shot classification head.\n\n*W11.* The method involves expanding the zero-shot head at test time. There is a compute and memory overhead associated with this that may limit the scalability of the method when many classes or subclasses are targeted. For example, for the 1000 ImageNet classes, with m=10, the instantiated head would have 10k classes.\n\n**Minor**\n\n*M1.* The Radford et al. 2021 caption example is “a photo of a {}.” not “a photo of a {}” as presented in the manuscript. The difference is the period at the end of the prompt, which may make a difference in downstream performance.\n\n*M2.* This is a relevant reference when discussing CLIP confidence in the method section: https://arxiv.org/abs/2106.07998\n\n*M3.* For Sec. 4.4 is there some theory as to why weighting is necessary in one case and not in the other?\n", "clarity,_quality,_novelty_and_reproducibility": "* The use of GPT-3 in zero-shot open-vocabulary image classification is novel (also appearing in concurrent work as the authors recognize).\n\n* While I have not seen hierarchical classification methods applied in zero-shot image classification, these kinds of techniques are well studied in the literature in more classical ML settings. The authors can improve the manuscript by positioning their hierarchical classification algorithm relative to others. However, the authors do not claim to present hierarchical classification in-and-of-itself as novel.\n\n* I have a minor concern about reproducibility given that the GPT-3 component of the work uses a temperature parameter of 0.7 and hence introduces some randomness into the results. However, the authors release all generated labels and hence the numbers in the paper should be reproducible for the datasets presented.\n\n* Some of the experiments require better explanation and could benefit from appendices giving more experimental details (see W7, W10).\n\n* The introduction and method section are very well written.", "summary_of_the_review": "The paper provides a method to prompt open-vocabulary models for zero-shot image classification, which leverages class hierarchies. It is straightforward to implement and the authors provide some empirical evidence that their method performs significantly better than the baseline, especially when hierarchies are known for a dataset.\n\nHowever, I currently recommend weak rejection of the manuscript. I am most concerned about the lack of benchmarking on ImageNet (*W3.*), the performance of the method on at least a few more standard classification datasets (*W4.*), the fairness of the zero-shot baselines (*W5.*), and the clarity of the presentation associated with some of the ablations studies (*W7.*, *W10.*).\n\nI am willing to revisit my evaluation during the rebuttal/discussion period.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1665614566062}], "openreview_url": "https://openreview.net/forum?id=NEEtm5laNK1", "arxiv_id": "2302.02551", "paper_pdf": "papers/NEEtm5laNK1.pdf", "paper_pdf_sha256": "ecce8401113a2293776e89aa54bf0c86402b873aa6c75620417a27b53ed98f23", "paper_pdf_bytes": 2756789, "paper_pdf_source": "openreview", "code_url": "https://github.com/acmi-lab/CHILS", "code_repository": "acmi-lab/CHILS", "code_commit": "c90ce90db72e1c0f01994c00f40b0f4c3a6983c2", "code_archive": "repos/NEEtm5laNK1.zip", "code_archive_sha256": "657600cd66c08ecba190063398f16f5cdcca35c3d941fe8b2c22ab263fc2e7bb", "code_archive_bytes": 358625, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 336, "github_languages": {"Python": 109978}, "github_archived": false, "github_pushed_at": "2023-06-04T19:32:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/chils-zero-shot-image-classification-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gzeruP-0J29", "year": 2022, "status": "rejected", "title": "Revisiting and Advancing Fast Adversarial Training Through the lens of Bi-Level Optimization", "authors": ["Yihua Zhang", "Guanhua Zhang", "Prashant Khanduri", "Mingyi Hong", "Shiyu Chang", "Sijia Liu"], "authorids": ["~Yihua_Zhang1", "~Guanhua_Zhang1", "~Prashant_Khanduri1", "~Mingyi_Hong1", "~Shiyu_Chang2", "~Sijia_Liu1"], "authors_source": "OpenReview API", "abstract": "Adversarial training (AT) has become a widely recognized defense mechanism to improve the robustness of deep neural networks against adversarial attacks. It is originated from solving a min-max optimization problem, where the minimizer (i.e., defender) seeks a robust model to minimize the worst-case training loss at the presence of adversarial examples crafted by the maximizer (i.e., attacker). However,the min-max nature makes AT computationally intensive and thus difficult to scale. Thus, the problem of FAST-AT arises. Nearly all the recent progress is achieved based on the following simplification: The iterative attack generation method used in the maximization step of AT is replaced by the simplest one-shot gradient sign-based PGD method. Nevertheless, FAST-AT is far from satisfactory, and it lacks theoretically-grounded design. For example, a FAST-AT method may suffer from robustness catastrophic overfitting when training with strong adversaries.\n\nIn this paper, we foster a technological breakthrough for designing FAST-AT through the lens of bi-level optimization (BLO) instead of min-max optimization. First, we theoretically show that the most commonly-used algorithmic specification of FAST-AT is equivalent to the linearized BLO along the direction given by the sign of input gradient. Second, with the aid of BLO, we develop a new systematic and effective fast bi-level AT framework, termed FAST-BAT, whose algorithm is rigorously derived by leveraging the theory of implicit gradient. In contrast to FAST-AT, FAST-BAT has the least restriction to placing the tradeoff between computation efficiency and adversarial robustness. For example, it is capable of defending sign-based projected gradient descent (PGD) attacks without calling any gradient sign method and explicit robust regularization during training. Furthermore, we empirically show that our method outperforms state-of-the-art FAST-AT baselines. In particular, FAST-BAT can achieve superior model robustness without inducing robustness catastrophic overfitting and losing standard accuracy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "qTX7OkYLwp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2023/Reviewer_6Qs9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work focuses on the problem of speeding up adversarial training in the $\\ell_\\inf$ threat model. The work first describes two previous works designed to speed up adversarial training:\n- Fast-AT: carefully perform single-step PGD (otherwise known as FGSM) to train $\\ell_\\inf$ robust models - from Wong, Rice and Kolter 2020\n- Fast-AT-GA: single-step PGD + a gradient alignment loss function - from Andriushchenko and Flammarion 2020.\n\nThe work then lists a problem with Fast-AT:\n- Fast-AT experiences catastrophic overfitting (i.e. provides no robustness on held-out data) for $\\epsilon = 16/255$ (Note: Fast-AT-GA does not have this catastrophic overfitting issue)\n\nFinally, the authors present a new method based on bi-level optimization to perform fast $\\ell_\\inf$-robust learning. They evaluate it and compare with Fast-AT-GA using a 20 epoch schedule, finding that for $\\epsilon = 16/255$, the technique can provide the same robustness except with standard accuracy ~68% instead of ~59%. ", "review_text": "Strengths: the work outperforms Fast-AT-GA at $\\epsilon=16/255$, obtaining better standard accuracy in less time while matching robust accuracy (standard accuracy ~68% instead of ~59%).\n\nWeaknesses:\nBaselines and experimental evaluation: \n- When comparing performance on a system that has evaluation metrics of both accuracy and training speed, the evaluation must take both into account. Currently it does not: all we see is a table with times and accuracies. One standard way of comparing training speed and accuracy is to measure either the \"maximum accuracy in a fixed amount of time\" or \"minimum time to attain a fixed accuracy\" while varying the metric of interest (time or accuracy respectively). Using this type of evaluation, a reader can see the speed/adv. accuracy frontier necessary to compare across algorithms. Currently the evaluation is not sufficiently detailed to compare across algorithms.\n- As a suggestion, one natural way of trading off time for accuracy is to treat the number of epochs as a free parameter; an illuminating here would be adv. performance vs training time as the number of epochs changes for each method.\n- The performance analysis is currently not informative; it uses a fixed training configuration for each training method (according to the Table 1 description, all the algorithms use early stopping yet according to Figure 1 this heavily disadvantages Fast-AT compared to no early stopping). One must choose the pareto (trading off speed and accuracies) optimal configurations for each training algorithm. The analysis must ablate across different training configurations to be fair.\n- The time does not have a standard deviation.\n- How is it that $\\epsilon=8/255$ and $\\epsilon=16/255$ trained models always have the same runtime, as reported in Table 1? Does early stopping always exit the same epoch for both?\n\nCharacterization of previous work: \n- What is the difference between problem (i) and problem (ii) on page 2? Robust performance is always properly evaluated against the strongest possible adversary.\n- The authors claim to solve an issue with Fast-AT stemming from its tendency towards catastrophic overfitting. The authors should give more background on when Fast-AT fails to catastrophic overfitting; how often does it happen? How does the $\\epsilon$ impact Fast-AT's convergence?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work focuses on the problem of speeding up adversarial training in the $\\ell_\\inf$ threat model. The work first describes two previous works designed to speed up adversarial training:\n- Fast-AT: carefully perform single-step PGD (otherwise known as FGSM) to train $\\ell_\\inf$ robust models - from Wong, Rice and Kolter 2020\n- Fast-AT-GA: single-step PGD + a gradient alignment loss function - from Andriushchenko and Flammarion 2020.\n\nThe work then lists a problem with Fast-AT:\n- Fast-AT experiences catastrophic overfitting (i.e. provides no robustness on held-out data) for $\\epsilon = 16/255$ (Note: Fast-AT-GA does not have this catastrophic overfitting issue)\n\nFinally, the authors present a new method based on bi-level optimization to perform fast $\\ell_\\inf$-robust learning. They evaluate it and compare with Fast-AT-GA using a 20 epoch schedule, finding that for $\\epsilon = 16/255$, the technique can provide the same robustness except with standard accuracy ~68% instead of ~59%. ", "main_review": "Strengths: the work outperforms Fast-AT-GA at $\\epsilon=16/255$, obtaining better standard accuracy in less time while matching robust accuracy (standard accuracy ~68% instead of ~59%).\n\nWeaknesses:\nBaselines and experimental evaluation: \n- When comparing performance on a system that has evaluation metrics of both accuracy and training speed, the evaluation must take both into account. Currently it does not: all we see is a table with times and accuracies. One standard way of comparing training speed and accuracy is to measure either the \"maximum accuracy in a fixed amount of time\" or \"minimum time to attain a fixed accuracy\" while varying the metric of interest (time or accuracy respectively). Using this type of evaluation, a reader can see the speed/adv. accuracy frontier necessary to compare across algorithms. Currently the evaluation is not sufficiently detailed to compare across algorithms.\n- As a suggestion, one natural way of trading off time for accuracy is to treat the number of epochs as a free parameter; an illuminating here would be adv. performance vs training time as the number of epochs changes for each method.\n- The performance analysis is currently not informative; it uses a fixed training configuration for each training method (according to the Table 1 description, all the algorithms use early stopping yet according to Figure 1 this heavily disadvantages Fast-AT compared to no early stopping). One must choose the pareto (trading off speed and accuracies) optimal configurations for each training algorithm. The analysis must ablate across different training configurations to be fair.\n- The time does not have a standard deviation.\n- How is it that $\\epsilon=8/255$ and $\\epsilon=16/255$ trained models always have the same runtime, as reported in Table 1? Does early stopping always exit the same epoch for both?\n\nCharacterization of previous work: \n- What is the difference between problem (i) and problem (ii) on page 2? Robust performance is always properly evaluated against the strongest possible adversary.\n- The authors claim to solve an issue with Fast-AT stemming from its tendency towards catastrophic overfitting. The authors should give more background on when Fast-AT fails to catastrophic overfitting; how often does it happen? How does the $\\epsilon$ impact Fast-AT's convergence?", "summary_of_the_review": "While the work shows a promising approach (as measured by outperforming Fast-AT-GA at $\\epsilon=16/255$ in terms of standard accuracy while beating it in speed and in robust accuracy), the evaluation is lacking. In particular, the evaluation does not give a speed/accuracy tradeoff for each training routine, and does not properly hyperparameter search for each compared algorithm in the comparison. Finally, the work has some minor issues in its characterization of previous work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635991545116}, {"id": "6pEA1oN1eeM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2023/Reviewer_TZX8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies adversarial training as a bi-level optimization problem. The authors show that Fast Adversarial Training can be viewed as a bi-level optimization problem where the lower-level problem is a linearization of the loss in the direction of the sign of the gradient. Motivated by this observation, the authors propose Fast Bi-level Adversarial Training, where the lower-level optimization linearizes the loss in the direction of the gradient (as opposed to the sign of the gradient). The authors then show analytically how to obtain the gradient of the upper-level problem under the assumption that the loss Hessian is equal to zero. This gradient is then used in an iterative manner similar to standard adversarial training. The proposed method is evaluated on CIFAR10 and ImageNet against multiple baselines including Fast-AT, Fast-AT-GA, and PGD-2. The empirical results show that compared to the baselines, within the same order of computational cost, the proposed method enjoys improved stability and mitigates the catastrophic overfitting present in other baselines.", "review_text": "Relevance: the paper focuses on the important problem of designing computationally attractive algorithms for adversarial training -- this is an important problem that is relevant to ICLR.\n\nResults: The main strength of the paper is the empirical study. Given that within the same computation budget as the baselines, the proposed method has lower variance and resolves catastrophic overfitting in several settings, I found the results important. The results are only averaged over 5 random experiments, therefore, given the complexity and the size of the task, I’m not sure how reliable the estimates of mean/variance are. I think the empirical results can be much more compelling if the authors include more experiments.\n\nPresentation: Although the paper is well-written for the most part, there is much room for improvement in sections 3 and 4. In particular, some claims are not well-justified, there are some issues with the notation and the presentation lacks mathematical rigor. Here are some examples:\n\n            - in page 3, the authors state that “FAST-AT-GA yields improved robustness but has a poor accuracy-robustness tradeoff (e.g. Table 1)“. I do not see how FAST-AT-GA has a poor accuracy-robustness tradeoff in Table 1. Can you elaborate?\n\n            - the description of set C = {delta | || delta ||_infty < epsilon,  delta in [0, 1]} is very confusing. What does delta in [0, 1] mean? attacks are always point-wise positive? How is it not vacuous since epsilon < 1?\n\n            - in page 4, the authors say “Even if we set ell_atk = - ell_tr, problem 2 does not reduce to 1 due to the presence of lower-level constraint”. Regardless of whether this claim is true or not, the justification is far from satisfactory. In order to argue two optimization “problem” are equivalent, or if one reduces to another, one needs to prove/disprove that solutions to one constitute solutions to the other. The fact that gradients are different — as argued in the paper — does not a-priori inform us of anything useful about the equivalence of the two. In particular, when ell_atk = - ell_tr, then delta* in (2) is very well a maximizer of the inner attack problem in (1). Again, whether two optimization problems are equivalent has nothing to do with certain optimization procedures (such as PGD being considered here) generating different trajectories.\n\n            - the bi-level interpretation of Fast-At is trivial — this is how gradient updates are motivated in the literature of first-order optimization.\n\n            - in page 5, the derivation of the implicit gradient in (6) seems reasonable but lacks mathematical rigor. I suggest authors include careful proof here.\n\n            - The Hessian-free assumption in Theorem 1 essentially means that the attack loss ell_atk is linear. I do believe that this is not true for most interesting cases, and I do not buy the argument “the rationale behind the Hessian-free assumption is that neural networks commonly lead to a piece-wise linear decision boundary”. Are you suggesting that the loss landscape in deep learning is almost linear?\n\n            - in page 6, remark 1, the authors state that “Clearly, if alpha_2 = 0, then it reduces to the standard Fast-AT”. How is this true? here the updates will be based on the gradient of the loss whereas in standard Fast-At the updates are based on the sign of the gradient.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies adversarial training as a bi-level optimization problem. The authors show that Fast Adversarial Training can be viewed as a bi-level optimization problem where the lower-level problem is a linearization of the loss in the direction of the sign of the gradient. Motivated by this observation, the authors propose Fast Bi-level Adversarial Training, where the lower-level optimization linearizes the loss in the direction of the gradient (as opposed to the sign of the gradient). The authors then show analytically how to obtain the gradient of the upper-level problem under the assumption that the loss Hessian is equal to zero. This gradient is then used in an iterative manner similar to standard adversarial training. The proposed method is evaluated on CIFAR10 and ImageNet against multiple baselines including Fast-AT, Fast-AT-GA, and PGD-2. The empirical results show that compared to the baselines, within the same order of computational cost, the proposed method enjoys improved stability and mitigates the catastrophic overfitting present in other baselines.", "main_review": "Relevance: the paper focuses on the important problem of designing computationally attractive algorithms for adversarial training -- this is an important problem that is relevant to ICLR.\n\nResults: The main strength of the paper is the empirical study. Given that within the same computation budget as the baselines, the proposed method has lower variance and resolves catastrophic overfitting in several settings, I found the results important. The results are only averaged over 5 random experiments, therefore, given the complexity and the size of the task, I’m not sure how reliable the estimates of mean/variance are. I think the empirical results can be much more compelling if the authors include more experiments.\n\nPresentation: Although the paper is well-written for the most part, there is much room for improvement in sections 3 and 4. In particular, some claims are not well-justified, there are some issues with the notation and the presentation lacks mathematical rigor. Here are some examples:\n\n            - in page 3, the authors state that “FAST-AT-GA yields improved robustness but has a poor accuracy-robustness tradeoff (e.g. Table 1)“. I do not see how FAST-AT-GA has a poor accuracy-robustness tradeoff in Table 1. Can you elaborate?\n\n            - the description of set C = {delta | || delta ||_infty < epsilon,  delta in [0, 1]} is very confusing. What does delta in [0, 1] mean? attacks are always point-wise positive? How is it not vacuous since epsilon < 1?\n\n            - in page 4, the authors say “Even if we set ell_atk = - ell_tr, problem 2 does not reduce to 1 due to the presence of lower-level constraint”. Regardless of whether this claim is true or not, the justification is far from satisfactory. In order to argue two optimization “problem” are equivalent, or if one reduces to another, one needs to prove/disprove that solutions to one constitute solutions to the other. The fact that gradients are different — as argued in the paper — does not a-priori inform us of anything useful about the equivalence of the two. In particular, when ell_atk = - ell_tr, then delta* in (2) is very well a maximizer of the inner attack problem in (1). Again, whether two optimization problems are equivalent has nothing to do with certain optimization procedures (such as PGD being considered here) generating different trajectories.\n\n            - the bi-level interpretation of Fast-At is trivial — this is how gradient updates are motivated in the literature of first-order optimization.\n\n            - in page 5, the derivation of the implicit gradient in (6) seems reasonable but lacks mathematical rigor. I suggest authors include careful proof here.\n\n            - The Hessian-free assumption in Theorem 1 essentially means that the attack loss ell_atk is linear. I do believe that this is not true for most interesting cases, and I do not buy the argument “the rationale behind the Hessian-free assumption is that neural networks commonly lead to a piece-wise linear decision boundary”. Are you suggesting that the loss landscape in deep learning is almost linear?\n\n            - in page 6, remark 1, the authors state that “Clearly, if alpha_2 = 0, then it reduces to the standard Fast-AT”. How is this true? here the updates will be based on the gradient of the loss whereas in standard Fast-At the updates are based on the sign of the gradient.\n", "summary_of_the_review": "To summarize:\nStrengths:\n          - the paper studies an important and relevant problem.\n          - the experiments are well done. \n          - the proposed adversarial training approach has nice theoretical motivations.\n          - the paper is well-written for the most part.\n\nWeakness:\n          - several ideas and claims are handwavy and lack mathematical rigor \n          - the assumption behind the main theoretical result is too stringent\n          - statistical significance --> only 5 random experiments", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635886379351}, {"id": "DT9fJ9Ab8b5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2023/Reviewer_c58a"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a bi-level optimization problem for training models robust to norm constrained adversarial attacks that can be optimized via argmin/implicit differentiation by linearizing the inner problem (the search for the worst perturbation). The new technique is then compared empirically to other existing approximations of the robust training problem (referred to in the paper as Adversarial Training) and shown to provide improved or comparable results in terms of standard and robust accuracy, for different attacks.", "review_text": "Rewriting (robust) AT as a bi-level optimization problem is a nice formalization that allows the use of implicit differentiation for training robust models. This contribution seems however straightforward and the deployed techniques for optimizing the problem (linearized proximal form, differentiable projection) are standard in the literature.\nThe other advantage of rewriting the problem in this way is in making a distinction between training loss $\\ell_{tr}$ and attack loss $\\ell_{atk}$, but this does not seem to be explored in the experiments.\n\nThe paper presents some inaccuracies when reporting the related work and in the derivations:\n\nContrary to what stated in the paper, Adversarial Training was proposed before [Madry 2018] in the seminal work of [1]. In [Madry 2018] adversarial training was formulated as a Robust Optimization problem. The attribution should be rectified.\n\nTo write Problems (5) and (7) it is never explained how the term $\\frac{\\lambda}{2}$ squared norm 2 of $\\delta - z$ appears. It does not come directly from the linearization otherwise this term would be multiplied by the second order derivate of $\\ell_{atk}$ and there wouldn't be any $\\lambda$ hyper-parameter. Instead it seems that the $\\lambda$ proximal operator is applied, but this is never mentioned.\nThe assumptions on the loss functions needed in order to linearize the BLO problem are not reported.\n\nIn order to compute the implicit gradient of FAST-AT, $\\mathcal{P}_{\\mathcal{C}}$ does not need to be differentiable as the derivative of the sign function is 0 null on the dimensions that need an update. The statement \"the projection function is not smooth, and thus the use of chain rule is not legitimate\" is somehow incorrect, as this projection is still differentiable as it is shown later in page 6, it just gives a piece-wise derivative depending on the box constraints.\n\nIn Theorem 1, it is not explained why the so-called Hessian-free assumption is needed. Moreover, its naming seems inaccurate as not all the components of the Hessian should be null (in particular those w.r.t $\\theta \\delta$).\n\nThe following notions essential for understanding the paper are not defined: robustness catastrophic overfitting, random corner linearization, gradient alignment.\n\nThe methods used as baselines in the experiments are not sufficiently described, namely FAST-AT-GA and PGD-2-AT. Also AutoAttacks should be described and the reason why this type of attack is chosen to measure robust accuracy should be provided. In the tables, the standard deviations should be used to determine which score differences are significant. At the moment, it seems that the result marked in bold is the one with the best average score, but a significance test should be used instead. \nWhy does PGD-2-AT generally achieve better results than FAST-AT and FAST-AT-GA and comparable with FAST-BAT? \n\n(Minor) The following reference is defined twice: Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018b.\n\n[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks, ICLR 2014", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a bi-level optimization problem for training models robust to norm constrained adversarial attacks that can be optimized via argmin/implicit differentiation by linearizing the inner problem (the search for the worst perturbation). The new technique is then compared empirically to other existing approximations of the robust training problem (referred to in the paper as Adversarial Training) and shown to provide improved or comparable results in terms of standard and robust accuracy, for different attacks.", "main_review": "Rewriting (robust) AT as a bi-level optimization problem is a nice formalization that allows the use of implicit differentiation for training robust models. This contribution seems however straightforward and the deployed techniques for optimizing the problem (linearized proximal form, differentiable projection) are standard in the literature.\nThe other advantage of rewriting the problem in this way is in making a distinction between training loss $\\ell_{tr}$ and attack loss $\\ell_{atk}$, but this does not seem to be explored in the experiments.\n\nThe paper presents some inaccuracies when reporting the related work and in the derivations:\n\nContrary to what stated in the paper, Adversarial Training was proposed before [Madry 2018] in the seminal work of [1]. In [Madry 2018] adversarial training was formulated as a Robust Optimization problem. The attribution should be rectified.\n\nTo write Problems (5) and (7) it is never explained how the term $\\frac{\\lambda}{2}$ squared norm 2 of $\\delta - z$ appears. It does not come directly from the linearization otherwise this term would be multiplied by the second order derivate of $\\ell_{atk}$ and there wouldn't be any $\\lambda$ hyper-parameter. Instead it seems that the $\\lambda$ proximal operator is applied, but this is never mentioned.\nThe assumptions on the loss functions needed in order to linearize the BLO problem are not reported.\n\nIn order to compute the implicit gradient of FAST-AT, $\\mathcal{P}_{\\mathcal{C}}$ does not need to be differentiable as the derivative of the sign function is 0 null on the dimensions that need an update. The statement \"the projection function is not smooth, and thus the use of chain rule is not legitimate\" is somehow incorrect, as this projection is still differentiable as it is shown later in page 6, it just gives a piece-wise derivative depending on the box constraints.\n\nIn Theorem 1, it is not explained why the so-called Hessian-free assumption is needed. Moreover, its naming seems inaccurate as not all the components of the Hessian should be null (in particular those w.r.t $\\theta \\delta$).\n\nThe following notions essential for understanding the paper are not defined: robustness catastrophic overfitting, random corner linearization, gradient alignment.\n\nThe methods used as baselines in the experiments are not sufficiently described, namely FAST-AT-GA and PGD-2-AT. Also AutoAttacks should be described and the reason why this type of attack is chosen to measure robust accuracy should be provided. In the tables, the standard deviations should be used to determine which score differences are significant. At the moment, it seems that the result marked in bold is the one with the best average score, but a significance test should be used instead. \nWhy does PGD-2-AT generally achieve better results than FAST-AT and FAST-AT-GA and comparable with FAST-BAT? \n\n(Minor) The following reference is defined twice: Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018b.\n\n[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks, ICLR 2014", "summary_of_the_review": "I'd tend to reject the paper because of the highlighted inaccuracies and unclear points, and because the contributions seem straightforward applications of BLO theory to adversarial training. Additionally, the new linearization does not seem to significantly improve state-of-the-art results.\n\n## UPDATE\nMy final evaluation is based on the current version of the paper, as the authors had the chance to update it.\nI increased my score to 5, as the inaccuracies have been corrected, namely:\n- the attribution of adversarial training has been rectified\n- the additional $\\lambda$ norm_2 term in the linearized inner problems has been explained\n- the phrasing for the Hessian-free assumption and projection's derivative has been corrected\n \nI still tend to reject the paper for the following reasons:\n- definition of gradient alignment and autoattacks is not provided\n- justification for using autoattacks is not provided. \"everybody is doing it\" is not a sufficient justification per se\n- comparison wrt baselines is based on mean values, hence the reader cannot judge the significance of the improvements. Doubts about the significance of the empirical results has been raised by other reviewers as well.\n\nFinally, I still have some doubts about the contributions of the paper to the BLO literature. In the rebuttal the authors claim \"(1) we show that the choice of lower-level linearization (gradient sign-based vs. non-sign case) could be a key to simplifying BLO with lower-level constraints, and (2) we derive the closed-form of the implicit gradient of lower-level constrained BLO assisted by lower-level linearization.\"\nHoewever the two linearization schemes are not novel, neither the derivative of the projection onto a convex set.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635861121016}, {"id": "a3laIqJRpdr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2023/Reviewer_N6su"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper aims to interpret the fast adversarial training methods from the perspective bi-level optimization. Though the discovery is straightforward, it is worthy of letting the community know this important connection. And then the authors proposed a new linearization of the lower-level optimization problem and introduced a new FAST-BAT approach for improving both accuracy and robustness. Various experiments were conducted to verify the effectiveness of the proposed method. ", "review_text": "In general, I like this perspective due to the realizing of the implicit gradient term since actually the adversarial perturbation depends on the network weights, and this is also related to some versions of GANs’ training. However, in the current version of the paper, there exists several key points that requires careful justification to make it as a much stronger piece of work. I list them in the below. If these questions could be handled properly, I will increase my score. \n\n1.\tIn the Remark 1, it seems an very important conclusion the authors want to claim on the role of the second term associated with alpha2. Then what is evidence for the claim that “the choice of alpha2 could affect the tradeoff between accuracy and robustness…alpha2-associated term plays a positive role in boosting robustness, especially for training with large epsilon”? \n2.\tWhy does FAST-BAT produce better gradient alignment? Is it just an empirical observation or is there any good theoretical justification?\n3.\tIs there any connection between FAST-BAT and standard PGD adversarial training? \n\nOther questions: \nAbout the Hessian-free assumption. What about the non-ReLU activation function for which the function is not piece-wise linear wrt to input? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to interpret the fast adversarial training methods from the perspective bi-level optimization. Though the discovery is straightforward, it is worthy of letting the community know this important connection. And then the authors proposed a new linearization of the lower-level optimization problem and introduced a new FAST-BAT approach for improving both accuracy and robustness. Various experiments were conducted to verify the effectiveness of the proposed method. ", "main_review": "In general, I like this perspective due to the realizing of the implicit gradient term since actually the adversarial perturbation depends on the network weights, and this is also related to some versions of GANs’ training. However, in the current version of the paper, there exists several key points that requires careful justification to make it as a much stronger piece of work. I list them in the below. If these questions could be handled properly, I will increase my score. \n\n1.\tIn the Remark 1, it seems an very important conclusion the authors want to claim on the role of the second term associated with alpha2. Then what is evidence for the claim that “the choice of alpha2 could affect the tradeoff between accuracy and robustness…alpha2-associated term plays a positive role in boosting robustness, especially for training with large epsilon”? \n2.\tWhy does FAST-BAT produce better gradient alignment? Is it just an empirical observation or is there any good theoretical justification?\n3.\tIs there any connection between FAST-BAT and standard PGD adversarial training? \n\nOther questions: \nAbout the Hessian-free assumption. What about the non-ReLU activation function for which the function is not piece-wise linear wrt to input? \n", "summary_of_the_review": "Good perspective for interpreting fast adversarial training, but in the current version, there exists several key points that requires careful justification to make it as a much stronger piece of work.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635168584602}], "openreview_url": "https://openreview.net/forum?id=gzeruP-0J29", "arxiv_id": "2112.12376", "paper_pdf": "papers/gzeruP-0J29.pdf", "paper_pdf_sha256": "aff32c2a8be4979176a71a4b9a26aa17a1897699e8ecea6d146a56b29051a094", "paper_pdf_bytes": 969338, "paper_pdf_source": "openreview", "code_url": "https://github.com/OPTML-Group/Fast-BAT", "code_repository": "OPTML-Group/Fast-BAT", "code_commit": "926599de5b7a99166f725519db6c5042405f9251", "code_archive": "repos/gzeruP-0J29.zip", "code_archive_sha256": "81466f24243f5f3933feb4c691bb23125f10277efb9bf26dad9bd81061462d36", "code_archive_bytes": 744999, "code_file_count": 43, "code_extensions": {".py": 40, ".sh": 3}, "github_disk_usage_kb": 889, "github_languages": {"Shell": 4338110, "Python": 300694}, "github_archived": false, "github_pushed_at": "2022-10-25T21:01:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/revisiting-and-advancing-fast-adversarial-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OJiM1R3jAtZ", "year": 2021, "status": "rejected", "title": "AWAC: Accelerating Online Reinforcement Learning with Offline Datasets", "authors": ["Ashvin Nair", "Murtaza Dalal", "Abhishek Gupta", "Sergey Levine"], "authorids": ["~Ashvin_Nair1", "~Murtaza_Dalal1", "~Abhishek_Gupta1", "~Sergey_Levine1"], "authors_source": "OpenReview API", "abstract": "Reinforcement learning provides an appealing formalism for learning control policies from experience. However, the classic active formulation of reinforcement learning necessitates a lengthy active exploration process for each behavior, making it difficult to apply in real-world settings. If we can instead allow reinforcement learning to effectively use previously collected data to aid the online learning process, where the data could be expert demonstrations or more generally any prior experience, we could make reinforcement learning a substantially more practical tool. While a number of recent methods have sought to learn offline from previously collected data, it remains exceptionally difficult to train a policy with offline data and improve it further with online reinforcement learning. In this paper we systematically analyze why this problem is so challenging, and propose an algorithm that combines sample-efficient dynamic programming with maximum likelihood policy updates, providing a simple and effective framework that is able to leverage large amounts of offline data and then quickly perform online fine-tuning of reinforcement learning policies. We show that our method enables rapid learning of skills with a combination of prior demonstration data and online experience across a suite of difficult dexterous manipulation and benchmark tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "k1eo8zEGrn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2748/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies shorting coming of existing off-policy methods when it comes to prior data and fine-tuning and shows that those existing methods can't effectively utilize previously collected data with online updates. To address this problem, they propose to constraint policy updates with respect to behavioral policy. Their proposed method is built mainly on the top of AWR [1]. \n\n- There is a significant similarity between the proposed method and AWR [1]. The main difference that I can see is AWR doesn't have a fine-tuning step but this paper does. Can authors list their differences with AWR? \n\n- CRR [3] is very similar to the proposed method in this paper as well, what are the differences? ( I disagree with the authors that CRR is a concurrent work with this paper)\n\n-  This paper claims that AWAC can utilize various types of prior data without any changes to their method and it is agnostic to the type of behavioral policy (e.g. random vs. expert). I don't see how this is the case. For example, if data were collected by a random policy, this method constrains the current policy to be a uniform one.\n\n- Experiments don't make a case for this method, i.e. AWAC is only better than others in relocate-binary-v0 and barely better in Walker2d-v2.  ABM almost gets the same performance.\n\n- Fine-Tuning with random policy data is not convincing either. I'd suggest running some experiments with D4RL [5] dataset like *-medium-expert, *-medium-replay, and *-random to build a better case for your proposed method.\n\n- One reason that existing RL methods can't fully utilize previously collected data with online updates is the distribution shift between the policy (a.k.a behavioral policy) that collected the data and the learned policy [2, 3]. Can the authors comment on this and discuss how the proposed method addresses the distribution shift problem that likely the root cause of this problem? A section in the paper about this will help to improve this paper.  \n \n- Related work section needs significant improvement. It only lists various papers without any useful discussion.  \n\nEven though I found this paper interesting, I'm still not convinced about the contribution of this paper. Given similarity and overlap with previous works [1,3], more experiments could be helpful to see how this method is better than others and why one should select this method vs. others.\n\nMinor comments:\n\nIt seems there are various problems with your bibtex, for example, \"Behavior Regularized Offline Reinforcement Learning\" wasn't published in iclr and author names are mixed with other information like in  \"A Generalized Path Integral Control Approach to Reinforcement Learning\" \"usc\" in author names and \"Web Services\" appeared with author names \"P3O: Policy-on Policy-off Policy Optimization.\"  \n\n[1] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning. sep 2019. \n\n[2] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems. Technical report, 2020\n\n[3 Rasool Fakoor, Pratik Chaudhari, and Alexander J Smola. P3O: Policy-on Policy-off Policy Optimization. In Conference on Uncertainty in Artificial Intelligence (UAI), 2019. \n\n[4] Ziyu Wang, Alexander Novikov, Konrad Zołna, Jost Tobias Springenberg, Scott Reed, Bobak Shahriari, Noah Siegel, Josh Merel, Caglar Gulcehre, Nicolas Heess, and Nando De Freitas. Critic Regularized Regression. 2020.\n\n[5] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, Sergey Levine. D4RL: Datasets for Deep Data-Driven Reinforcement Learning, 2020", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "interesting paper but some concerns ", "review": "This paper studies shorting coming of existing off-policy methods when it comes to prior data and fine-tuning and shows that those existing methods can't effectively utilize previously collected data with online updates. To address this problem, they propose to constraint policy updates with respect to behavioral policy. Their proposed method is built mainly on the top of AWR [1]. \n\n- There is a significant similarity between the proposed method and AWR [1]. The main difference that I can see is AWR doesn't have a fine-tuning step but this paper does. Can authors list their differences with AWR? \n\n- CRR [3] is very similar to the proposed method in this paper as well, what are the differences? ( I disagree with the authors that CRR is a concurrent work with this paper)\n\n-  This paper claims that AWAC can utilize various types of prior data without any changes to their method and it is agnostic to the type of behavioral policy (e.g. random vs. expert). I don't see how this is the case. For example, if data were collected by a random policy, this method constrains the current policy to be a uniform one.\n\n- Experiments don't make a case for this method, i.e. AWAC is only better than others in relocate-binary-v0 and barely better in Walker2d-v2.  ABM almost gets the same performance.\n\n- Fine-Tuning with random policy data is not convincing either. I'd suggest running some experiments with D4RL [5] dataset like *-medium-expert, *-medium-replay, and *-random to build a better case for your proposed method.\n\n- One reason that existing RL methods can't fully utilize previously collected data with online updates is the distribution shift between the policy (a.k.a behavioral policy) that collected the data and the learned policy [2, 3]. Can the authors comment on this and discuss how the proposed method addresses the distribution shift problem that likely the root cause of this problem? A section in the paper about this will help to improve this paper.  \n \n- Related work section needs significant improvement. It only lists various papers without any useful discussion.  \n\nEven though I found this paper interesting, I'm still not convinced about the contribution of this paper. Given similarity and overlap with previous works [1,3], more experiments could be helpful to see how this method is better than others and why one should select this method vs. others.\n\nMinor comments:\n\nIt seems there are various problems with your bibtex, for example, \"Behavior Regularized Offline Reinforcement Learning\" wasn't published in iclr and author names are mixed with other information like in  \"A Generalized Path Integral Control Approach to Reinforcement Learning\" \"usc\" in author names and \"Web Services\" appeared with author names \"P3O: Policy-on Policy-off Policy Optimization.\"  \n\n[1] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning. sep 2019. \n\n[2] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems. Technical report, 2020\n\n[3 Rasool Fakoor, Pratik Chaudhari, and Alexander J Smola. P3O: Policy-on Policy-off Policy Optimization. In Conference on Uncertainty in Artificial Intelligence (UAI), 2019. \n\n[4] Ziyu Wang, Alexander Novikov, Konrad Zołna, Jost Tobias Springenberg, Scott Reed, Bobak Shahriari, Noah Siegel, Josh Merel, Caglar Gulcehre, Nicolas Heess, and Nando De Freitas. Critic Regularized Regression. 2020.\n\n[5] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, Sergey Levine. D4RL: Datasets for Deep Data-Driven Reinforcement Learning, 2020", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604447276038}, {"id": "qhmsSRqSvt", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2748/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nIn this paper, the authors intend to accelerate on-line reinforcement learning with off-line datasets. To achieve this goal, they propose an algorithm called advantage weighted actor-critic (AWAC), which uses an implicit constraint to reduce accumulated bootstrapping error when doing off-line training and reduce the conservation when doing on-line fine-tuning. The experiments show that the proposed method can learn difficult, high-dimensional, sparse reward dexterous manipulation problems from human demonstrations and off-policy data.\n\nPros:\n1. The organization is clear and related work is sufficient. This paper introduces the off-line training and on-line fine-tuning problem and its challenge by gradually summarize the existing studies, and with sufficient investigation of existing work, it is easy for readers to dive into the problem\n2. The proposed method is easy to implement. Thanks to the implicit formulation of the constraint, the proposed AWAC algorithm can sample directly from off-line datasets without fitting a parametric model. This strategy is especially advantageous when the dimensionality is high. Besides, since this algorithm does not need the parametric model any more, it is more friendly for users who are not familiar with the off-line dataset.\n3. The details of experimental settings are provided, and thus there should be no issues with the repeatability of the experiments. \n\nCons:\n1. It is doubtful for the contribution on online fine-tuning. The constraint is designed for off-line training because off-line training has the problem of bootstrapping error accumulation. There is no judgment or proof that online training also suffers from the bootstrapping problem. How can the authors claim their contribution on on-line learning with the constraint?\n2. The technical depth is not high. It seems the only contribution of this paper is to combine the constraint with an unparameterized strategy. But I would like to say that this is not a big problem, as many influential reinforcement learning algorithms are straightforward but have very significant improvement, e.g., DDQN and PPO.\n3. The writing needs polish. E.g., in Equation (4), it should be 'max' instead of 'argmax', and there exists non-English spelling in Section 3.1.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A fairly good idea with not so high technical depth", "review": "Summary:\n\nIn this paper, the authors intend to accelerate on-line reinforcement learning with off-line datasets. To achieve this goal, they propose an algorithm called advantage weighted actor-critic (AWAC), which uses an implicit constraint to reduce accumulated bootstrapping error when doing off-line training and reduce the conservation when doing on-line fine-tuning. The experiments show that the proposed method can learn difficult, high-dimensional, sparse reward dexterous manipulation problems from human demonstrations and off-policy data.\n\nPros:\n1. The organization is clear and related work is sufficient. This paper introduces the off-line training and on-line fine-tuning problem and its challenge by gradually summarize the existing studies, and with sufficient investigation of existing work, it is easy for readers to dive into the problem\n2. The proposed method is easy to implement. Thanks to the implicit formulation of the constraint, the proposed AWAC algorithm can sample directly from off-line datasets without fitting a parametric model. This strategy is especially advantageous when the dimensionality is high. Besides, since this algorithm does not need the parametric model any more, it is more friendly for users who are not familiar with the off-line dataset.\n3. The details of experimental settings are provided, and thus there should be no issues with the repeatability of the experiments. \n\nCons:\n1. It is doubtful for the contribution on online fine-tuning. The constraint is designed for off-line training because off-line training has the problem of bootstrapping error accumulation. There is no judgment or proof that online training also suffers from the bootstrapping problem. How can the authors claim their contribution on on-line learning with the constraint?\n2. The technical depth is not high. It seems the only contribution of this paper is to combine the constraint with an unparameterized strategy. But I would like to say that this is not a big problem, as many influential reinforcement learning algorithms are straightforward but have very significant improvement, e.g., DDQN and PPO.\n3. The writing needs polish. E.g., in Equation (4), it should be 'max' instead of 'argmax', and there exists non-English spelling in Section 3.1.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604299785473}, {"id": "Vurck0O34F0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2748/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper describes an approach to use offline data to accelerate the\nonline learning process of reinforcement learning.\nThe proposed approach, called AWAC, uses dynamic programming to train\na critic and supervised learning to trained a constrained actor.\nThe idea is to use a static dataset of experience tuples for\npre-training and use some online interactions to learn the optimal\npolicy for the current task.\nFor the policy improvement step they optimize the policy to maximize\nthe estimated Q-value function but constrained to remain close to the\nactions observed in the data, using the Kulback-Leibler divergence.\nThis constrain is incorporated in the optimization and solve using the\nLagrangian. \nThe proposed approach is very similar to AWR, however, instead of\nestimating the value function of the behavior policy with a Monte\nCarlo approach, they estimate the Q function of the current policy via\nbootstrapping. \nAWAC is tested in several MuJoCo simulator problems, three shown in\nthe main paper (in-hand rotation of a pen, opening a door by\nunlatching the handle, and picking up a sphere and relocating it in a\ntarget location), and others shown in the appendix.\n\nPositive:\n- Use previously collected data for online learning is certainly\nrelevant for reinforcement learning is domains such as robotics. The\nproposed approach shows that a slight change in how to evaluate the\nQ-function to estimate the advantage function can produce a\nsignificant difference\n\nNegative:\n- The paper follows very closely the description of AWR with only\na slight variation in how to evaluate the advantage function. \n- It is not clear how close the static dataset has to be to a \"good\"\npolicy and how it affects the learning process.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An incremental improvement on combining offline data with online learning", "review": "The paper describes an approach to use offline data to accelerate the\nonline learning process of reinforcement learning.\nThe proposed approach, called AWAC, uses dynamic programming to train\na critic and supervised learning to trained a constrained actor.\nThe idea is to use a static dataset of experience tuples for\npre-training and use some online interactions to learn the optimal\npolicy for the current task.\nFor the policy improvement step they optimize the policy to maximize\nthe estimated Q-value function but constrained to remain close to the\nactions observed in the data, using the Kulback-Leibler divergence.\nThis constrain is incorporated in the optimization and solve using the\nLagrangian. \nThe proposed approach is very similar to AWR, however, instead of\nestimating the value function of the behavior policy with a Monte\nCarlo approach, they estimate the Q function of the current policy via\nbootstrapping. \nAWAC is tested in several MuJoCo simulator problems, three shown in\nthe main paper (in-hand rotation of a pen, opening a door by\nunlatching the handle, and picking up a sphere and relocating it in a\ntarget location), and others shown in the appendix.\n\nPositive:\n- Use previously collected data for online learning is certainly\nrelevant for reinforcement learning is domains such as robotics. The\nproposed approach shows that a slight change in how to evaluate the\nQ-function to estimate the advantage function can produce a\nsignificant difference\n\nNegative:\n- The paper follows very closely the description of AWR with only\na slight variation in how to evaluate the advantage function. \n- It is not clear how close the static dataset has to be to a \"good\"\npolicy and how it affects the learning process.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603916213173}, {"id": "keh7ZefetV", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2748/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies challenges of offline RL with online fine-tuning and proposes an off-policy actor-critic method to address these challenges. The proposed method uses a supervised learning style to update the model parameters and avoids the behavior model estimation.  Empirical results show that the proposed method provides rapid learning with prior demonstration data and online experience.\n\nI have a major concern about the novelty of this paper. The major component of the proposed AWAC method, updating the model parameter using supervised learning, is exactly the same as AWR (Peng et. al., 2019). One minor difference seems that AWAC uses an off-policy policy evaluation, but this contribution is very marginal. \n\nThis paper provides an analysis on challenges of combining offline RL with online improvement, which motivates this paper. However, most of the discussed challenges are quite well-known. For example, one major challenge is to estimate the behavior model in offline data. This paper does not discuss techniques/methods that address this challenge, like DualDICE (Nachum et al., 2019) and CQL (Kumaret al., 2020). A comparison to these methods is highly recommended. In addition, this paper may also include additional model-based offline RL methods, like MOPO.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lack of novelty", "review": "This paper studies challenges of offline RL with online fine-tuning and proposes an off-policy actor-critic method to address these challenges. The proposed method uses a supervised learning style to update the model parameters and avoids the behavior model estimation.  Empirical results show that the proposed method provides rapid learning with prior demonstration data and online experience.\n\nI have a major concern about the novelty of this paper. The major component of the proposed AWAC method, updating the model parameter using supervised learning, is exactly the same as AWR (Peng et. al., 2019). One minor difference seems that AWAC uses an off-policy policy evaluation, but this contribution is very marginal. \n\nThis paper provides an analysis on challenges of combining offline RL with online improvement, which motivates this paper. However, most of the discussed challenges are quite well-known. For example, one major challenge is to estimate the behavior model in offline data. This paper does not discuss techniques/methods that address this challenge, like DualDICE (Nachum et al., 2019) and CQL (Kumaret al., 2020). A comparison to these methods is highly recommended. In addition, this paper may also include additional model-based offline RL methods, like MOPO.\n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603905104830}, {"id": "qFckh0XAqP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2748/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an improved actor-critic algorithm for learning from offline data while requiring minimum interactions during online learning. The approach implicitly updates a low-variance policy without the need to estimate the sampling policy. The empirical study shows that AWAC leads to a faster convergence rate compared to the competitors.\n\nOverall, the paper provides an incremental contribution over the existing works for off-policy RL problems. However, the proposed algorithm demonstrates some practical benefits in simulated scenarios and is analyzes from different perspectives. Hence, I vote for accepting, given the following concerns are addressed in the rebuttal period.\n\nPros:\n- Interesting adaption of algorithms to effectively use the offline data in online RL scenarios which is independent from the sampling policy\n- Ensuring the policy to stay close to the observed data without directly estimating the sampling policy\n- Nicely motivated with helpful examples\n- Clarifies differences with the related work\n\nComments:\n- The presentation of the paper could be improved. There are some inconsistencies in the text. At the beginning the approach is introduced to be based on dynamic programming and then turns into temporal difference learning, which is misleading. In addition, offline pre-training and online fine-tuning is not quite accurate as the algorithm 1 is using the mix of data in one phase which does not involve pre-training, or is it actually in two phases?\n\n- The paper seems to be an incremental improvement/combination of several recent works mentioned in the paper.\n\n- The experiments are all on simulated environments. It would be interesting to see how the approach performs on real-world applications, and with some other evaluation metric than convergence rate. \n\n- Demonstration data is not always as complete as mentioned in the paper. Expert demonstrations that are used for imitation learning, normally lack the reward values and in the form of (s,a) pairs only. \n\n\nMinor:\n-Typo in section 3.2: is policy is pre-trained\n- Plot 4 in Fig. 2 not addressed in section 3.3", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Incremental but practically interesting approach", "review": "The paper presents an improved actor-critic algorithm for learning from offline data while requiring minimum interactions during online learning. The approach implicitly updates a low-variance policy without the need to estimate the sampling policy. The empirical study shows that AWAC leads to a faster convergence rate compared to the competitors.\n\nOverall, the paper provides an incremental contribution over the existing works for off-policy RL problems. However, the proposed algorithm demonstrates some practical benefits in simulated scenarios and is analyzes from different perspectives. Hence, I vote for accepting, given the following concerns are addressed in the rebuttal period.\n\nPros:\n- Interesting adaption of algorithms to effectively use the offline data in online RL scenarios which is independent from the sampling policy\n- Ensuring the policy to stay close to the observed data without directly estimating the sampling policy\n- Nicely motivated with helpful examples\n- Clarifies differences with the related work\n\nComments:\n- The presentation of the paper could be improved. There are some inconsistencies in the text. At the beginning the approach is introduced to be based on dynamic programming and then turns into temporal difference learning, which is misleading. In addition, offline pre-training and online fine-tuning is not quite accurate as the algorithm 1 is using the mix of data in one phase which does not involve pre-training, or is it actually in two phases?\n\n- The paper seems to be an incremental improvement/combination of several recent works mentioned in the paper.\n\n- The experiments are all on simulated environments. It would be interesting to see how the approach performs on real-world applications, and with some other evaluation metric than convergence rate. \n\n- Demonstration data is not always as complete as mentioned in the paper. Expert demonstrations that are used for imitation learning, normally lack the reward values and in the form of (s,a) pairs only. \n\n\nMinor:\n-Typo in section 3.2: is policy is pre-trained\n- Plot 4 in Fig. 2 not addressed in section 3.3", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603882073105}], "openreview_url": "https://openreview.net/forum?id=OJiM1R3jAtZ", "arxiv_id": "2006.09359", "paper_pdf": "papers/OJiM1R3jAtZ.pdf", "paper_pdf_sha256": "668b265c09b89d596ad7a980bc55a88bd41c7ccd897500364752ccd1ac2a195e", "paper_pdf_bytes": 3787956, "paper_pdf_source": "openreview", "code_url": "https://github.com/rail-berkeley/rlkit", "code_repository": "rail-berkeley/rlkit", "code_commit": "ac45a9db24b89d97369bef302487273bcc3e3d84", "code_archive": "repos/OJiM1R3jAtZ.zip", "code_archive_sha256": "7d0c4965c00baaeb3ed746a9e351a1471517b0eef699230055d20c5cf7711ecb", "code_archive_bytes": 681913, "code_file_count": 190, "code_extensions": {".py": 190}, "github_disk_usage_kb": 1052, "github_languages": {"Python": 861329, "Dockerfile": 3338}, "github_archived": false, "github_pushed_at": "2024-06-17T17:33:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerating-online-reinforcement-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1xLuRVFvr", "year": 2020, "status": "rejected", "title": "Visual Explanation for Deep Metric Learning", "authors": ["Sijie Zhu", "Taojiannan Yang", "Chen Chen"], "authorids": ["szhu3@uncc.edu", "tyang30@uncc.edu", "chen.chen@uncc.edu"], "authors_source": "OpenReview API", "abstract": "This work explores the visual explanation for deep metric learning and its applications. As an important problem for learning representation, metric learning has attracted much attention recently, while the interpretation of such model is not as well studied as classification. To this end, we propose an intuitive idea to show where contributes the most to the overall similarity of two input images by decomposing the final activation. Instead of only providing the overall activation map of each image, we propose to generate point-to-point activation intensity between two images so that the relationship between different regions is uncovered. We show that the proposed framework can be directly deployed to a large range of metric learning applications and provides valuable information for understanding the model. Furthermore, our experiments show its effectiveness on two potential applications, i.e. cross-view pattern discovery and interactive retrieval. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SylvE-sx5r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1213/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "= Summary \nThis paper presents a simple method that draws visual attention of deep embedding networks for metric learning. It basically follows the class attention mapping strategy based on global pooling operation [Zhou et al., CVPR 2016], but extends the original version to point-specific attention which is novel and enable interesting applications on image retrieval. In addition, the proposed method seems also independent of the loss function used for metric learning, thus it can be applied to most of existing deep embedding networks to understand their behaviors in a qualitative manner. \n\n\n= Decision\nMy current decision is officially \"weak reject\" but \"borderline\" in my mind. The major concern of mine is its weaknesses in clarity and technical novelty. However, I still believe this submission is valuable since it addresses a relatively new and timely topic, the proposed method is simple yet effective, and the applications of point-specific attention (i.e., \"cross-view pattern discovery\" and \"interactive retrieval\") are all interesting and practically useful. If the clarity issues are all clearly addressed, I would upgrade my rating. \n\n\n= Comments\n[Pros]\n1) The motivation and implementation of the point-specific attention are convincing. \n2) The point-specific attention enables not only image-to-image retrieval, but also more elaborate understanding about a pair of images and their similarity.\n3) The proposed technique is model-agnostic and loss-agnostic, thus can be applied to most of existing deep embedding networks for image retrieval. Also, the proposed technique does not degrade the retrieval performance.\n4) The proposed technique is simple yet effective in multiple applications, most of which are practically useful and have great impact. For example, the weakly supervised localization is an essential step towards many weakly supervised approaches for higher-level recognition tasks like semantic segmentation, and the interactive retrieval will allow us to build more realistic and useful image retrieval systems.\n\n[Cons]\n1) The proposed technique itself is not new but a straightforward extension of an existing work [Zhou et al., CVPR 2016].\n2) The manuscript is not crystal clear.\n- The name of the proposed point-specific attention (i.e., \"Partial attention\") is misleading.\n- The way to compute the point-specific attention map is not clearly described in Section 3.\n- It is hard to understand the contents in Figure 6 as they are not clearly illustrated. \n- The experimental and implementation details of the last two applications are not given. For example: (cross-view pattern discovery) how to compute the angle error, and how much the proposed technique is sensitive to the position selected on the query image, (interactive retrieval) how to compute the similarity between images given a specific region of interest on query.\n3) More qualitative results on the last two applications should be presented, even in an appendix, to convince future readers. Especially, in the case of \"interactive retrieval\", more results are demanded as quantitative performance analysis seems not straightforward.\n\n\n= Post-rebuttal review\nThe rebuttal resolves my major concerns and the manuscript has been carefully revised accordingly. So as I promised in my original review, I upgrade the score to weak accept. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "title": "Official Blind Review #3", "review": "= Summary \nThis paper presents a simple method that draws visual attention of deep embedding networks for metric learning. It basically follows the class attention mapping strategy based on global pooling operation [Zhou et al., CVPR 2016], but extends the original version to point-specific attention which is novel and enable interesting applications on image retrieval. In addition, the proposed method seems also independent of the loss function used for metric learning, thus it can be applied to most of existing deep embedding networks to understand their behaviors in a qualitative manner. \n\n\n= Decision\nMy current decision is officially \"weak reject\" but \"borderline\" in my mind. The major concern of mine is its weaknesses in clarity and technical novelty. However, I still believe this submission is valuable since it addresses a relatively new and timely topic, the proposed method is simple yet effective, and the applications of point-specific attention (i.e., \"cross-view pattern discovery\" and \"interactive retrieval\") are all interesting and practically useful. If the clarity issues are all clearly addressed, I would upgrade my rating. \n\n\n= Comments\n[Pros]\n1) The motivation and implementation of the point-specific attention are convincing. \n2) The point-specific attention enables not only image-to-image retrieval, but also more elaborate understanding about a pair of images and their similarity.\n3) The proposed technique is model-agnostic and loss-agnostic, thus can be applied to most of existing deep embedding networks for image retrieval. Also, the proposed technique does not degrade the retrieval performance.\n4) The proposed technique is simple yet effective in multiple applications, most of which are practically useful and have great impact. For example, the weakly supervised localization is an essential step towards many weakly supervised approaches for higher-level recognition tasks like semantic segmentation, and the interactive retrieval will allow us to build more realistic and useful image retrieval systems.\n\n[Cons]\n1) The proposed technique itself is not new but a straightforward extension of an existing work [Zhou et al., CVPR 2016].\n2) The manuscript is not crystal clear.\n- The name of the proposed point-specific attention (i.e., \"Partial attention\") is misleading.\n- The way to compute the point-specific attention map is not clearly described in Section 3.\n- It is hard to understand the contents in Figure 6 as they are not clearly illustrated. \n- The experimental and implementation details of the last two applications are not given. For example: (cross-view pattern discovery) how to compute the angle error, and how much the proposed technique is sensitive to the position selected on the query image, (interactive retrieval) how to compute the similarity between images given a specific region of interest on query.\n3) More qualitative results on the last two applications should be presented, even in an appendix, to convince future readers. Especially, in the case of \"interactive retrieval\", more results are demanded as quantitative performance analysis seems not straightforward.\n\n\n= Post-rebuttal review\nThe rebuttal resolves my major concerns and the manuscript has been carefully revised accordingly. So as I promised in my original review, I upgrade the score to weak accept. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572020527314}, {"id": "B1gxS8ARFS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1213/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a visualization method for deep metric learning, which derived by analyzing the inner product of two globally averaged activations. The proposed method can generate an overall activation map which highlights the regions contributing most to the similarity. Also, it can generate a partial activation map that lights the regions in one image that have significant activation responses on a specific potion in the other image. The authors also analyzed the linearly of the fully connected layers and global max pooling. These contributions make the applicability of CAM to many CNN architectures. Further, the metric learning architecture is extended to Grad-CAM map, and the problem of Grad-CAM map is pointed out. To the best of my knowledge, these contributions are novel, and derivations seem to be correct. \n\nExperiments on weakly-supervised localization, model diagnosis, and the applications of the proposed decomposition model in cross-view pattern discovery and interactive retrieval are promising. \n\nOverall, this paper is well written, and contributions are good. \n\nMinor problems.  \nIn Sec.1 and Sec.2.2, the authors wrote the Grad-CAM has been used for visualization of re-ID (Gordo & Larlus (2017)). However, this paper seems to be not the works of Grad-CAM nor re-ID. \n\nIn my understanding, Decomposition+Bias is a more accurate model than Decomposition. \nIn the experiments of the Sec5.1 and 5.2, the performances of Decompsotion+Bias are lower than Decomposition. However, there are no explanations for this reason. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a visualization method for deep metric learning, which derived by analyzing the inner product of two globally averaged activations. The proposed method can generate an overall activation map which highlights the regions contributing most to the similarity. Also, it can generate a partial activation map that lights the regions in one image that have significant activation responses on a specific potion in the other image. The authors also analyzed the linearly of the fully connected layers and global max pooling. These contributions make the applicability of CAM to many CNN architectures. Further, the metric learning architecture is extended to Grad-CAM map, and the problem of Grad-CAM map is pointed out. To the best of my knowledge, these contributions are novel, and derivations seem to be correct. \n\nExperiments on weakly-supervised localization, model diagnosis, and the applications of the proposed decomposition model in cross-view pattern discovery and interactive retrieval are promising. \n\nOverall, this paper is well written, and contributions are good. \n\nMinor problems.  \nIn Sec.1 and Sec.2.2, the authors wrote the Grad-CAM has been used for visualization of re-ID (Gordo & Larlus (2017)). However, this paper seems to be not the works of Grad-CAM nor re-ID. \n\nIn my understanding, Decomposition+Bias is a more accurate model than Decomposition. \nIn the experiments of the Sec5.1 and 5.2, the performances of Decompsotion+Bias are lower than Decomposition. However, there are no explanations for this reason. \n"}, "tcdate": 1571903031541}, {"id": "Syx91gcotr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1213/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes a novel approach for visualizing the predictions of neural network models on pairwise tasks, e.g. predicting whether two images are similar. The authors show that to see similarity between two images down on the pixel/region level, it is not sufficient to apply methods like Grad-CAM which do not aim for decomposition. Instead, the authors' approach namely targets decomposition, and shows the benefit of decomposition through intuitive examples and qualitative results. The authors also quantitatively show the benefit of their method. First, they measure the performance of their method vs CAM and Grad-CAM, on the weakly supervised localization (WSL) task. They also show how their method reveals the disadvantages of standard triplet loss compared to a recent metric learning loss. \n\nMy concerns:\n1) While showing performance on WSL is appealing as it allows for a way to quantitatively evaluate the method, I wonder if there are other ways to evaluate, that explicitly measure the quality of the proposed technique for allowing interpretability and understanding of the base model's performance. \n2) The Triplet vs MS experiment is interesting, but I'm not sure this is the most convincing way to show that this proposed visualization technique is better than something else. Just because something shows one method is worse than another, doesn't mean that this better/worse assessment is accurate. Further, how would Grad-CAM do on the same task?\n3) The retrieval experiment only shows qualitative results, and again, no baseline is compared. \n4) Similarity is a relative judgement; it's hard to say if two items are similar, but easier to say if A and B are more similar than A and C. It seems the proposed method doesn't consider negatives, which is perhaps a limitation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This work proposes a novel approach for visualizing the predictions of neural network models on pairwise tasks, e.g. predicting whether two images are similar. The authors show that to see similarity between two images down on the pixel/region level, it is not sufficient to apply methods like Grad-CAM which do not aim for decomposition. Instead, the authors' approach namely targets decomposition, and shows the benefit of decomposition through intuitive examples and qualitative results. The authors also quantitatively show the benefit of their method. First, they measure the performance of their method vs CAM and Grad-CAM, on the weakly supervised localization (WSL) task. They also show how their method reveals the disadvantages of standard triplet loss compared to a recent metric learning loss. \n\nMy concerns:\n1) While showing performance on WSL is appealing as it allows for a way to quantitatively evaluate the method, I wonder if there are other ways to evaluate, that explicitly measure the quality of the proposed technique for allowing interpretability and understanding of the base model's performance. \n2) The Triplet vs MS experiment is interesting, but I'm not sure this is the most convincing way to show that this proposed visualization technique is better than something else. Just because something shows one method is worse than another, doesn't mean that this better/worse assessment is accurate. Further, how would Grad-CAM do on the same task?\n3) The retrieval experiment only shows qualitative results, and again, no baseline is compared. \n4) Similarity is a relative judgement; it's hard to say if two items are similar, but easier to say if A and B are more similar than A and C. It seems the proposed method doesn't consider negatives, which is perhaps a limitation."}, "tcdate": 1571688418260}], "openreview_url": "https://openreview.net/forum?id=S1xLuRVFvr", "arxiv_id": "1909.12977", "paper_pdf": "papers/S1xLuRVFvr.pdf", "paper_pdf_sha256": "84808360cc28fe67885d247b9bf46e6133c3293f951efd695db6baf46bacc77e", "paper_pdf_bytes": 3031392, "paper_pdf_source": "openreview", "code_url": "https://github.com/Jeff-Zilence/Explain_Metric_Learning", "code_repository": "Jeff-Zilence/Explain_Metric_Learning", "code_commit": "c289f8d1d690dd8c14c2aca3542ca7d714c74d6b", "code_archive": "repos/S1xLuRVFvr.zip", "code_archive_sha256": "7cd7c94672149c8d14ba71482a22e8dcf8cfb1ae6082516779d22fd97e17fa37", "code_archive_bytes": 3115704, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 3030, "github_languages": {"Python": 113280}, "github_archived": false, "github_pushed_at": "2022-10-30T18:01:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/visual-explanation-for-deep-metric-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3SAXoxEdCK", "year": 2026, "status": "rejected", "title": "On Path to Multimodal Historical Reasoning: HistBench and HistAgent", "authors": ["Jiahao Qiu", "Fulian Xiao", "Yimin Wang", "Yuchen Mao", "Yijia Chen", "Xinzhe Juan", "Siran Wang", "Xuan Qi", "Tongcheng Zhang", "Zixin Yao", "Jiacheng Guo", "Yifu Lu", "Charles Argon", "Jundi Cui", "Daixin Chen", "Junran Zhou", "Shuyao Zhou", "Zhanpeng Zhou", "Ling Yang", "Shilong Liu", "Hongru WANG", "Kaixuan Huang", "xun jiang", "Xi Gao", "Mengdi Wang"], "authorids": ["~Jiahao_Qiu1", "~Fulian_Xiao1", "~Yimin_Wang5", "~Yuchen_Mao3", "~Yijia_Chen5", "~Xinzhe_Juan1", "~Siran_Wang1", "~Xuan_Qi3", "~Tongcheng_Zhang1", "~Zixin_Yao2", "~Jiacheng_Guo1", "~Yifu_Lu1", "~Charles_Argon1", "~Jundi_Cui1", "~Daixin_Chen1", "~Junran_Zhou1", "~Shuyao_Zhou1", "~Zhanpeng_Zhou1", "~Ling_Yang1", "~Shilong_Liu1", "~Hongru_WANG1", "~Kaixuan_Huang1", "~xun_jiang3", "~Xi_Gao2", "~Mengdi_Wang1"], "authors_source": "OpenReview API", "abstract": "Recent advances in large language models (LLMs) have led to remarkable progress across various domains, yet their capabilities in the humanities, particularly history, remain underexplored. Historical reasoning poses unique challenges for LLMs, involving multimodal source interpretation, temporal inference, and cross-linguistic analysis. Existing general-purpose agents perform well on many current benchmarks but lack the domain expertise needed to address complex historical questions.\nTo address this gap, we introduce HistBench, a new benchmark of 414 high-quality and carefully-reviewed questions stratified by difficulty and designed to evaluate LLM's capacity for historical reasoning. The tasks span a wide range of historical problems—from factual retrieval based on primary sources to interpretive analysis of manuscripts and images, to interdisciplinary challenges involving archaeology, linguistics, or cultural history. Furthermore, the benchmark dataset spans 29 ancient and modern languages and covers a wide range of historical periods and world regions. Finding the poor performance of LLMs and other agents on HistBench, we further present HistAgent, a history-specific agent equipped with carefully designed tools for OCR, translation, archival search, and image understanding in History. On HistBench, HistAgent based on GPT-4o achieves an accuracy of 27.54% pass@1 and 36.47% pass@2, significantly outperforming LLMs with online search and generalist agents, including GPT-4o (18.60%), DeepSeek-R1(14.49%), Grok 3(17.63%) and Open Deep Research by smolagents(20.29% pass@1 and 25.12% pass@2). These results highlight the limitations of existing LLMs and generalist agents and demonstrate the advantages of HistAgent for historical reasoning. Notably, HistAgent also achieves 60.00% pass@1 accuracy on the GAIA benchmark, showing that domain-specific customization doesn't hinder HistAgent's competitive performance on real-world general tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "lTJzDbdGvf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15283/Reviewer_qQWn"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The authors proposed a new benchmark HistBench that evaluate models on historical reasoning. This benchmark is very comprehensive, spanning 29 ancient and modern languages and covers a wide range of historical periods and world regions. The authors also proposed a new agent HistAgent with multiple different tools such as OCR to solve tasks on HistBench. On HistBench and other benchmarks, HistAgent beat other SOTA agents.", "review_text": "The authors proposed a new benchmark HistBench that evaluate models on historical reasoning. This benchmark is very comprehensive, spanning 29 ancient and modern languages and covers a wide range of historical periods and world regions. The authors also proposed a new agent HistAgent with multiple different tools such as OCR to solve tasks on HistBench. On HistBench and other benchmarks, HistAgent beat other SOTA agents.", "strengths": "- This paper introduces HistBench which is very useful to assess agents' abilities to solve complex historical questions through historical reasoning. This could be beneficial to the history research community.\n- The proposed agent HistAgent achieves SOTA performances on multiple benchmarks, which could be valuable resource for historians.\n- The authors performed comprehensive evaluation and analysis, providing valuable insights such as highlighting the importance of tools.", "weaknesses": "- The authors didn't discuss existing work in developing agents good at solving historical questions or historical reasoning. It would be good if they could provide some literature review on what other people in the field has developed.\n- The authors used different base language models (e.g. claude and gpt-4o) across different benchmarks without explaining why. It would be good if the authors could include a brief description of the reason for their choice of models.", "questions": "- What are the cost like for HistAgent and the baselines that the authors compared to?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors proposed a new benchmark HistBench that evaluate models on historical reasoning. This benchmark is very comprehensive, spanning 29 ancient and modern languages and covers a wide range of historical periods and world regions. The authors also proposed a new agent HistAgent with multiple different tools such as OCR to solve tasks on HistBench. On HistBench and other benchmarks, HistAgent beat other SOTA agents.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This paper introduces HistBench which is very useful to assess agents' abilities to solve complex historical questions through historical reasoning. This could be beneficial to the history research community.\n- The proposed agent HistAgent achieves SOTA performances on multiple benchmarks, which could be valuable resource for historians.\n- The authors performed comprehensive evaluation and analysis, providing valuable insights such as highlighting the importance of tools.", "weaknesses": "- The authors didn't discuss existing work in developing agents good at solving historical questions or historical reasoning. It would be good if they could provide some literature review on what other people in the field has developed.\n- The authors used different base language models (e.g. claude and gpt-4o) across different benchmarks without explaining why. It would be good if the authors could include a brief description of the reason for their choice of models.", "questions": "- What are the cost like for HistAgent and the baselines that the authors compared to?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761971686755}, {"id": "ejN36upIpZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15283/Reviewer_nK2E"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces HistBench, a benchmark for evaluating historical reasoning and HistAgent, a specialized agent designed to solve it. The authors highlight the lack of a dedicated historical reasoning benchmarks involving complex multimodal (manuscripts, images, audio) and multilingual inputs. HistBench consists of 414 expert curated questions contributed by 40 people. Its primary innovation is a 3-level difficulty stratification based on six criteria rubric rather than on model performance. The dataset is collected from manuscripts, inscriptions, early printed texts, archival records, visual artifacts, and audio visual materials. \nHistAgent is a manager-specialist agent architecture that equips an GPT-4o with domain specific tools, such as scholarly search, manuscript OCR, and specialized translation. The paper's key result is that HistAgent (27.54% pass@1) significantly outperforms a generalist agent (ODR-smolagents, 20.29%) and base models (GPT-4o, Grok3, DeepSeek-R1, 14-18%) on HistBench.", "review_text": "This paper introduces HistBench, a benchmark for evaluating historical reasoning and HistAgent, a specialized agent designed to solve it. The authors highlight the lack of a dedicated historical reasoning benchmarks involving complex multimodal (manuscripts, images, audio) and multilingual inputs. HistBench consists of 414 expert curated questions contributed by 40 people. Its primary innovation is a 3-level difficulty stratification based on six criteria rubric rather than on model performance. The dataset is collected from manuscripts, inscriptions, early printed texts, archival records, visual artifacts, and audio visual materials. \nHistAgent is a manager-specialist agent architecture that equips an GPT-4o with domain specific tools, such as scholarly search, manuscript OCR, and specialized translation. The paper's key result is that HistAgent (27.54% pass@1) significantly outperforms a generalist agent (ODR-smolagents, 20.29%) and base models (GPT-4o, Grok3, DeepSeek-R1, 14-18%) on HistBench.", "strengths": "1. The paper introduces a nuanced benchmark on historical reasoning which is an under-explored area.\n2. The manual verification and difficulty annotation process is very thorough. The authors use a 3-step verification process and create a comprehensive six-question rubric for annotation.\n3. The dataset is diverse with 29 languages spanning several regions and decades.", "weaknesses": "1. Misleading results are reported in Figure 1 and abstract. According to Figure 1, HistAgent performs better than base models. However, Table 3 highlights o3 and o4-mini are able to achieve much higher accuracies compared to HistAgent.\n2. While HistBench is a novel dataset compared to previous works, it is still very small with only 414 questions. Additionally, the data creation pipeline is time-consuming and not very scalable.\n3. The better performance of HistAgent on HistBench makes sense since the tool calls are specialized for the type of question present in HistBench. Adding specific tool calls for specific question types is bound to improve accuracy on those questions. It is unclear what the contribution of HistAgent is. Additionally, the cost of reproducing HistAgent seems very high since it requires specialized access.\n4. The experiments section is lacking a wide range of models as well as ablations. The authors should try reporting the results on different types of models, architectures, question types, and modalities.", "questions": "1. The authors should explain why part of the results were not reported in Figure 1 and abstract.\n2. Are the primary source materials used in the benchmark questions publicly available online or are they sourced from private archives?\n3. Suggestion - Related works should be moved to the main text and should be made more thorough to better motivate the problem. \n4. Suggestion - Table 1 font size should be increased.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces HistBench, a benchmark for evaluating historical reasoning and HistAgent, a specialized agent designed to solve it. The authors highlight the lack of a dedicated historical reasoning benchmarks involving complex multimodal (manuscripts, images, audio) and multilingual inputs. HistBench consists of 414 expert curated questions contributed by 40 people. Its primary innovation is a 3-level difficulty stratification based on six criteria rubric rather than on model performance. The dataset is collected from manuscripts, inscriptions, early printed texts, archival records, visual artifacts, and audio visual materials. \nHistAgent is a manager-specialist agent architecture that equips an GPT-4o with domain specific tools, such as scholarly search, manuscript OCR, and specialized translation. The paper's key result is that HistAgent (27.54% pass@1) significantly outperforms a generalist agent (ODR-smolagents, 20.29%) and base models (GPT-4o, Grok3, DeepSeek-R1, 14-18%) on HistBench.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper introduces a nuanced benchmark on historical reasoning which is an under-explored area.\n2. The manual verification and difficulty annotation process is very thorough. The authors use a 3-step verification process and create a comprehensive six-question rubric for annotation.\n3. The dataset is diverse with 29 languages spanning several regions and decades.", "weaknesses": "1. Misleading results are reported in Figure 1 and abstract. According to Figure 1, HistAgent performs better than base models. However, Table 3 highlights o3 and o4-mini are able to achieve much higher accuracies compared to HistAgent.\n2. While HistBench is a novel dataset compared to previous works, it is still very small with only 414 questions. Additionally, the data creation pipeline is time-consuming and not very scalable.\n3. The better performance of HistAgent on HistBench makes sense since the tool calls are specialized for the type of question present in HistBench. Adding specific tool calls for specific question types is bound to improve accuracy on those questions. It is unclear what the contribution of HistAgent is. Additionally, the cost of reproducing HistAgent seems very high since it requires specialized access.\n4. The experiments section is lacking a wide range of models as well as ablations. The authors should try reporting the results on different types of models, architectures, question types, and modalities.", "questions": "1. The authors should explain why part of the results were not reported in Figure 1 and abstract.\n2. Are the primary source materials used in the benchmark questions publicly available online or are they sourced from private archives?\n3. Suggestion - Related works should be moved to the main text and should be made more thorough to better motivate the problem. \n4. Suggestion - Table 1 font size should be increased.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761962426498}, {"id": "9bReKurWxd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15283/Reviewer_grvY"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper presents HistBench, a benchmark of 414 expert-curated questions spanning 29 languages and five modalities (text, manuscripts, images, audio, and video) to evaluate historical reasoning in large language models. The authors also propose HistAgent, a domain-specialized system built on GPT-4o that integrates OCR, translation, academic search, and multimodal analysis tools to assist in history-related tasks. On HistBench, HistAgent achieves 27.5% pass@1 and 36.5% pass@2 accuracy, outperforming generalist agents such as ODR-SmolAgents and GPT-4o with web search. It also maintains strong performance on general benchmarks like GAIA, demonstrating competitive generalization.", "review_text": "This paper presents HistBench, a benchmark of 414 expert-curated questions spanning 29 languages and five modalities (text, manuscripts, images, audio, and video) to evaluate historical reasoning in large language models. The authors also propose HistAgent, a domain-specialized system built on GPT-4o that integrates OCR, translation, academic search, and multimodal analysis tools to assist in history-related tasks. On HistBench, HistAgent achieves 27.5% pass@1 and 36.5% pass@2 accuracy, outperforming generalist agents such as ODR-SmolAgents and GPT-4o with web search. It also maintains strong performance on general benchmarks like GAIA, demonstrating competitive generalization.", "strengths": "Originality: The paper fills an important gap by introducing a domain-specific benchmark for the humanities, focusing on multimodal and multilingual historical reasoning. While earlier efforts like HLE or HiST-LLM covered historical knowledge, none provided this level of depth or tool-grounded evaluation.\n\nQuality: The benchmark design is robust, involving domain experts, stratified difficulty levels, and rigorous three-stage quality control (screening, LLM difficulty filtering, and expert review).\n\nClarity: The manuscript is dense but logically structured, with clear tables and figures illustrating dataset diversity and architecture. \n\nSignificance: HistBench sets a new standard for evaluating reasoning in historical and humanistic contexts. HistAgent’s modular design shows how domain specialization can rival or outperform larger closed models, which could inspire similar efforts in other fields such as archaeology or linguistics.", "weaknesses": "1. The dataset size (414 items) limits statistical granularity across 29 languages, especially for low-resource ones. Scaling beyond the pilot phase would strengthen claims of coverage.\n\n2. The results could include more fine-grained analysis, such as performance by language, modality, or reasoning dimension.\n\n3. The architecture section is somewhat heavy on engineering detail but could benefit from clearer ablation results demonstrating which tools contribute most.", "questions": "Have you measured the contribution of each HistAgent sub-module (OCR, translation, scholarly search) through ablations?\n\nDoes HistAgent generalize to other humanities disciplines, such as art history or linguistics?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents HistBench, a benchmark of 414 expert-curated questions spanning 29 languages and five modalities (text, manuscripts, images, audio, and video) to evaluate historical reasoning in large language models. The authors also propose HistAgent, a domain-specialized system built on GPT-4o that integrates OCR, translation, academic search, and multimodal analysis tools to assist in history-related tasks. On HistBench, HistAgent achieves 27.5% pass@1 and 36.5% pass@2 accuracy, outperforming generalist agents such as ODR-SmolAgents and GPT-4o with web search. It also maintains strong performance on general benchmarks like GAIA, demonstrating competitive generalization.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Originality: The paper fills an important gap by introducing a domain-specific benchmark for the humanities, focusing on multimodal and multilingual historical reasoning. While earlier efforts like HLE or HiST-LLM covered historical knowledge, none provided this level of depth or tool-grounded evaluation.\n\nQuality: The benchmark design is robust, involving domain experts, stratified difficulty levels, and rigorous three-stage quality control (screening, LLM difficulty filtering, and expert review).\n\nClarity: The manuscript is dense but logically structured, with clear tables and figures illustrating dataset diversity and architecture. \n\nSignificance: HistBench sets a new standard for evaluating reasoning in historical and humanistic contexts. HistAgent’s modular design shows how domain specialization can rival or outperform larger closed models, which could inspire similar efforts in other fields such as archaeology or linguistics.", "weaknesses": "1. The dataset size (414 items) limits statistical granularity across 29 languages, especially for low-resource ones. Scaling beyond the pilot phase would strengthen claims of coverage.\n\n2. The results could include more fine-grained analysis, such as performance by language, modality, or reasoning dimension.\n\n3. The architecture section is somewhat heavy on engineering detail but could benefit from clearer ablation results demonstrating which tools contribute most.", "questions": "Have you measured the contribution of each HistAgent sub-module (OCR, translation, scholarly search) through ablations?\n\nDoes HistAgent generalize to other humanities disciplines, such as art history or linguistics?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761904863521}, {"id": "CuDkXFNZIA", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15283/Reviewer_mZKh"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper ventures into the humanities to address the unique challenges that historical reasoning poses for current AI agents. The authors introduce HistBench, a new benchmark of 414 expert-reviewed questions designed to test multimodal historical reasoning. The questions are notable for their diversity, spanning 29 languages, multiple modalities (including ancient manuscripts and images), and various historical periods. \n\nFinding that existing models perform poorly on this new benchmark, the authors developed HistAgent, a specialized agent equipped with a suite of history-centric tools. These tools are designed for specific historical research tasks, such as performing OCR on ancient scripts, searching through digital archives, and translating obscure texts. On HistBench, HistAgent outperforms generalist agents and powerful foundation models, even those with web search capabilities. This highlights the necessity of domain-specific tools for complex, multimodal reasoning.", "review_text": "This paper ventures into the humanities to address the unique challenges that historical reasoning poses for current AI agents. The authors introduce HistBench, a new benchmark of 414 expert-reviewed questions designed to test multimodal historical reasoning. The questions are notable for their diversity, spanning 29 languages, multiple modalities (including ancient manuscripts and images), and various historical periods. \n\nFinding that existing models perform poorly on this new benchmark, the authors developed HistAgent, a specialized agent equipped with a suite of history-centric tools. These tools are designed for specific historical research tasks, such as performing OCR on ancient scripts, searching through digital archives, and translating obscure texts. On HistBench, HistAgent outperforms generalist agents and powerful foundation models, even those with web search capabilities. This highlights the necessity of domain-specific tools for complex, multimodal reasoning.", "strengths": "1. **High-Quality and Challenging Benchmark:** A major strength of this paper is the introduction of HistBench, a well-designed and challenging benchmark. The questions are stratified by professional historians using specific and sensible criteria, which adds to the dataset's credibility. The benchmark is also well-distinguished from others in the field, as effectively shown in Table 1. \n\n2. **Novel and Diverse Data:** The dataset contains data from obscure sources, which is a significant strength as it's difficult to find data that models haven't already seen. The data is also varied in its input modalities and sources, making it a robust test of a model's capabilities. \n\n3. **Well-Implemented Agent:** Although not a new idea, the Literature Search Agent is well-implemented and appropriate for the tasks at hand.  The agent's ability to navigate scholarly databases and parse academic PDFs is a crucial component of its success.", "weaknesses": "1. **Limited Scale of Evaluation:** The primary weakness of this paper is the small scale of the evaluation. With just 414 questions in the dataset, the evaluation suite is quite small. This is further compounded by the limited validation on other datasets, with only 3 candidates for HLE and 2 for GAIA. \n\n2. **Limited Model Diversity:** The evaluation is missing several major LLM players, such as Gemini, Claude, and Grok 4. Furthermore, there are no evaluations of open-source models, which would have provided a more comprehensive view of the current state of the field. The HistAgent is based on GPT-4o, which is a bit dated at this point. \n\n3. **Lack of In-Depth Analysis:** The analysis of the results is lacking and often feels like a verbal description of the results table.  A deeper dive into the reasons for the observed performance differences would have been more insightful. For example, while the results show that the agent doesn't always outperform existing models with search, there is no discussion as to why this might be the case. \n\n4. **Potential for Overfitting to the Benchmark:** As HistAgent was specifically designed to address the challenges presented in HistBench, there is a risk that its strong performance is, to some extent, a result of \"teaching to the test.\" While the specialized tools are well-motivated, a more robust evaluation would involve testing HistAgent on a separate, held-out set of historical reasoning tasks that were not considered during its development.", "questions": "How was it ensured that the data sources are not widely available or are distinct from those that may have gone into the pre-training of the models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper ventures into the humanities to address the unique challenges that historical reasoning poses for current AI agents. The authors introduce HistBench, a new benchmark of 414 expert-reviewed questions designed to test multimodal historical reasoning. The questions are notable for their diversity, spanning 29 languages, multiple modalities (including ancient manuscripts and images), and various historical periods. \n\nFinding that existing models perform poorly on this new benchmark, the authors developed HistAgent, a specialized agent equipped with a suite of history-centric tools. These tools are designed for specific historical research tasks, such as performing OCR on ancient scripts, searching through digital archives, and translating obscure texts. On HistBench, HistAgent outperforms generalist agents and powerful foundation models, even those with web search capabilities. This highlights the necessity of domain-specific tools for complex, multimodal reasoning.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. **High-Quality and Challenging Benchmark:** A major strength of this paper is the introduction of HistBench, a well-designed and challenging benchmark. The questions are stratified by professional historians using specific and sensible criteria, which adds to the dataset's credibility. The benchmark is also well-distinguished from others in the field, as effectively shown in Table 1. \n\n2. **Novel and Diverse Data:** The dataset contains data from obscure sources, which is a significant strength as it's difficult to find data that models haven't already seen. The data is also varied in its input modalities and sources, making it a robust test of a model's capabilities. \n\n3. **Well-Implemented Agent:** Although not a new idea, the Literature Search Agent is well-implemented and appropriate for the tasks at hand.  The agent's ability to navigate scholarly databases and parse academic PDFs is a crucial component of its success.", "weaknesses": "1. **Limited Scale of Evaluation:** The primary weakness of this paper is the small scale of the evaluation. With just 414 questions in the dataset, the evaluation suite is quite small. This is further compounded by the limited validation on other datasets, with only 3 candidates for HLE and 2 for GAIA. \n\n2. **Limited Model Diversity:** The evaluation is missing several major LLM players, such as Gemini, Claude, and Grok 4. Furthermore, there are no evaluations of open-source models, which would have provided a more comprehensive view of the current state of the field. The HistAgent is based on GPT-4o, which is a bit dated at this point. \n\n3. **Lack of In-Depth Analysis:** The analysis of the results is lacking and often feels like a verbal description of the results table.  A deeper dive into the reasons for the observed performance differences would have been more insightful. For example, while the results show that the agent doesn't always outperform existing models with search, there is no discussion as to why this might be the case. \n\n4. **Potential for Overfitting to the Benchmark:** As HistAgent was specifically designed to address the challenges presented in HistBench, there is a risk that its strong performance is, to some extent, a result of \"teaching to the test.\" While the specialized tools are well-motivated, a more robust evaluation would involve testing HistAgent on a separate, held-out set of historical reasoning tasks that were not considered during its development.", "questions": "How was it ensured that the data sources are not widely available or are distinct from those that may have gone into the pre-training of the models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761536791288}], "openreview_url": "https://openreview.net/forum?id=3SAXoxEdCK", "arxiv_id": "2505.20246", "paper_pdf": "papers/3SAXoxEdCK.pdf", "paper_pdf_sha256": "389f8ec78baac39220a6e3a94ab599aabfe60b4aa9f154c558db5b827cbae847", "paper_pdf_bytes": 1836555, "paper_pdf_source": "openreview", "code_url": "https://github.com/CharlesQ9/HistAgent", "code_repository": "CharlesQ9/HistAgent", "code_commit": "47bbe21dc81618489f5d5929358032883a3fe448", "code_archive": "repos/3SAXoxEdCK.zip", "code_archive_sha256": "c483fe2c8466b8f7f51ea33f20d9950bbe58834aa502e00d7d7d31df373519dd", "code_archive_bytes": 2545677, "code_file_count": 56, "code_extensions": {".py": 55, ".js": 1}, "github_disk_usage_kb": 3152, "github_languages": {"Python": 906511, "JavaScript": 31059}, "github_archived": false, "github_pushed_at": "2025-10-18T01:15:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-path-to-multimodal-historical-reasoning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "34xYxTTiM0", "year": 2025, "status": "rejected", "title": "Optimizing Calibration by Gaining Aware of Prediction Correctness", "authors": ["Yuchi Liu", "Lei Wang", "Yuli Zou", "James Zou", "Liang Zheng"], "authorids": ["~Yuchi_Liu1", "~Lei_Wang20", "~Yuli_Zou1", "~James_Zou1", "~Liang_Zheng4"], "authors_source": "OpenReview API", "abstract": "Model calibration aims to align confidence with prediction correctness. The Cross-Entropy (CE) loss is widely used for calibrator training, which enforces the model to increase confidence on the ground truth class. However, we find the CE loss has intrinsic limitations. For example, for a narrow misclassification, a calibrator trained by the CE loss often produces high confidence on the wrongly predicted class (e.g., a test sample is wrongly classified and its softmax score on the ground truth class is around 0.4), which is undesirable. In this paper, we propose a new post-hoc calibration objective derived from the aim of calibration. Intuitively, the proposed objective function asks that the calibrator decrease model confidence on wrongly predicted samples and increase confidence on correctly predicted samples. \nBecause a sample itself has insufficient ability to indicate correctness, we use its transformed versions (e.g., rotated, greyscaled, and color-jittered) during calibrator training. Trained on an in-distribution validation set and tested with isolated, individual test samples, \nour method achieves competitive calibration performance on both in-distribution and out-of-distribution test sets compared with the state of the art. Further, our analysis points out the difference between our method and commonly used objectives such as CE loss and Mean Square Error (MSE) loss, where the latters sometimes deviates from the calibration aim.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "yOah5r56nJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission867/Reviewer_pvJo"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces two innovative methods to address calibration errors in deep learning predictions: a correctness-aware loss function and a sample transformation technique. The correctness-aware loss function aims to directly minimize calibration error, effectively improving the calibration of misclassified samples by narrowing discrepancies across all classes. Additionally, to boost cross-domain performance, an augmentation-based transformation is applied to calibration samples, enhancing robustness across varied domains. Both methods are implemented in a post-hoc calibration framework, and the proposed algorithm demonstrates state-of-the-art performance, particularly in cross-domain settings.", "review_text": "The paper introduces two innovative methods to address calibration errors in deep learning predictions: a correctness-aware loss function and a sample transformation technique. The correctness-aware loss function aims to directly minimize calibration error, effectively improving the calibration of misclassified samples by narrowing discrepancies across all classes. Additionally, to boost cross-domain performance, an augmentation-based transformation is applied to calibration samples, enhancing robustness across varied domains. Both methods are implemented in a post-hoc calibration framework, and the proposed algorithm demonstrates state-of-the-art performance, particularly in cross-domain settings.", "strengths": "This paper presents a range of validation scenarios to assess the effectiveness of the proposed framework. In numerous cases, the framework achieves state-of-the-art performance, validating the impact of its two novel schemes. The experimental setup and comparisons are thoughtfully designed, with detailed descriptions that enhance clarity and reproducibility. Mathematical derivations are presented comprehensively, and the overall narrative is organized in a way that makes the framework easy to follow and understand, emphasizing key components effectively.", "weaknesses": "The paper has several strengths, yet I have some specific concerns that warrant attention:\n\n1. Definition of \"Narrow Misclassification\":\n   The term \"narrow misclassification\" appears in the abstract, and the correctness-aware (CA) loss is presented as targeting this condition by adjusting predictions across different classes rather than solely reducing confidence in the incorrect class. However, a clear definition of \"narrow misclassification\" is missing, and it’s challenging to discern how it differs from absolutely wrong samples even after reviewing the derivations. Clear definitions and empirical analysis based on outcomes would help clarify this distinction.\n\n2. Limitations from Augmentation Types Used:\n   The transformation component uses augmentations, but it lacks an analysis of how different types of augmentation affect performance across domains. Depending on the augmentation type, the efficacy in cross-domain scenarios may vary. Experimental validation or analysis is needed to determine the diversity of augmentation types required or which specific augmentations are essential.\n\n3. Similarity with Temperature Scaling:\n   If the framework were designed with temperature scaling, where the temperature parameter is shared across all classes, it could similarly distribute confidence across classes rather than reducing only the incorrect class's confidence. This raises questions about the uniqueness of the proposed algorithm’s approach in addressing \"narrow misclassification.\"\n\n4. Derivation for the CA Loss Function:\n   The derivation of the CA loss function appears to be unnecessarily complex. Initially, the paper emphasizes the use of continuous calibration error rather than Expected Calibration Error (ECE), suggesting a different approach. However, the final derivation seems equivalent to ECE-based loss, assuming discrete samples and small sample sizes, which undermines the rationale for a continuous assumption. Clarification is needed on why continuous assumptions were initially made if the final derivation closely resembles an ECE-based approach.\n\n5. Bounds of the CA Loss:\n   Bounds for the CA loss are derived based on assumptions that the sample sizes and accuracy across classes are similar. However, the significance of these bounds remains unclear, as they appear merely descriptive of the assumed conditions. Additional insights or generalized bounds demonstrating reduced CE loss could improve understanding.\n\n6. Unclear Derivation in Equation 15:\n   The derivation in Equation 15 is ambiguous due to an unexplained arrow, which might imply a limit. Clarification on which parameter converges to produce this outcome is necessary to improve the transparency of this mathematical derivation.\n\n7. Parameter \\theta in Equation 19:\n   It is unclear if \\theta in Equation 19 exclusively refers to the fully connected layers added for post-hoc calibration. This specification is important for clarity.\n\n8. Synergy between CA Loss and Transformation Component:\n   The CA loss reduces ECE, while the transformation improves cross-domain robustness. However, the synergy between these components is unclear, as seen in experimental results: applying CA loss significantly reduces ECE, while the transformation tends to increase ECE, showing a trade-off rather than synergy. Clarification is needed on why these mechanisms must be combined rather than sequentially applied as separate approaches.\n\n9. Baseline (CE Only + PTS) Already Achieving State-of-the-Art Performance:\n   In the result tables, the baseline (CE Only + PTS) already achieves state-of-the-art ECE and accuracy in multiple scenarios. While adding CA and transformation components improves performance further, it seems that these improvements are achieved largely because of the baseline's strong performance. To mitigate this concern, I recommend testing the proposed algorithm on alternative baselines.\n\n10. Minor Points:\n    - The text in figures is too small, making them hard to read.\n    - Typo: Line 136, “samples'.” should be “samples.'” \n\nThese concerns, if addressed, could enhance the clarity and impact of the proposed framework.", "questions": "Why does the paper initially emphasize using a continuous calibration error instead of the Expected Calibration Error (ECE)?\n\nWhat is the intended synergy between the CA loss and the transformation component, given their distinct purposes of reducing ECE and enhancing cross-domain robustness?\n\nCould the proposed algorithm’s effectiveness be validated further by testing it on alternative baselines?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces two innovative methods to address calibration errors in deep learning predictions: a correctness-aware loss function and a sample transformation technique. The correctness-aware loss function aims to directly minimize calibration error, effectively improving the calibration of misclassified samples by narrowing discrepancies across all classes. Additionally, to boost cross-domain performance, an augmentation-based transformation is applied to calibration samples, enhancing robustness across varied domains. Both methods are implemented in a post-hoc calibration framework, and the proposed algorithm demonstrates state-of-the-art performance, particularly in cross-domain settings.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper presents a range of validation scenarios to assess the effectiveness of the proposed framework. In numerous cases, the framework achieves state-of-the-art performance, validating the impact of its two novel schemes. The experimental setup and comparisons are thoughtfully designed, with detailed descriptions that enhance clarity and reproducibility. Mathematical derivations are presented comprehensively, and the overall narrative is organized in a way that makes the framework easy to follow and understand, emphasizing key components effectively.", "weaknesses": "The paper has several strengths, yet I have some specific concerns that warrant attention:\n\n1. Definition of \"Narrow Misclassification\":\n   The term \"narrow misclassification\" appears in the abstract, and the correctness-aware (CA) loss is presented as targeting this condition by adjusting predictions across different classes rather than solely reducing confidence in the incorrect class. However, a clear definition of \"narrow misclassification\" is missing, and it’s challenging to discern how it differs from absolutely wrong samples even after reviewing the derivations. Clear definitions and empirical analysis based on outcomes would help clarify this distinction.\n\n2. Limitations from Augmentation Types Used:\n   The transformation component uses augmentations, but it lacks an analysis of how different types of augmentation affect performance across domains. Depending on the augmentation type, the efficacy in cross-domain scenarios may vary. Experimental validation or analysis is needed to determine the diversity of augmentation types required or which specific augmentations are essential.\n\n3. Similarity with Temperature Scaling:\n   If the framework were designed with temperature scaling, where the temperature parameter is shared across all classes, it could similarly distribute confidence across classes rather than reducing only the incorrect class's confidence. This raises questions about the uniqueness of the proposed algorithm’s approach in addressing \"narrow misclassification.\"\n\n4. Derivation for the CA Loss Function:\n   The derivation of the CA loss function appears to be unnecessarily complex. Initially, the paper emphasizes the use of continuous calibration error rather than Expected Calibration Error (ECE), suggesting a different approach. However, the final derivation seems equivalent to ECE-based loss, assuming discrete samples and small sample sizes, which undermines the rationale for a continuous assumption. Clarification is needed on why continuous assumptions were initially made if the final derivation closely resembles an ECE-based approach.\n\n5. Bounds of the CA Loss:\n   Bounds for the CA loss are derived based on assumptions that the sample sizes and accuracy across classes are similar. However, the significance of these bounds remains unclear, as they appear merely descriptive of the assumed conditions. Additional insights or generalized bounds demonstrating reduced CE loss could improve understanding.\n\n6. Unclear Derivation in Equation 15:\n   The derivation in Equation 15 is ambiguous due to an unexplained arrow, which might imply a limit. Clarification on which parameter converges to produce this outcome is necessary to improve the transparency of this mathematical derivation.\n\n7. Parameter \\theta in Equation 19:\n   It is unclear if \\theta in Equation 19 exclusively refers to the fully connected layers added for post-hoc calibration. This specification is important for clarity.\n\n8. Synergy between CA Loss and Transformation Component:\n   The CA loss reduces ECE, while the transformation improves cross-domain robustness. However, the synergy between these components is unclear, as seen in experimental results: applying CA loss significantly reduces ECE, while the transformation tends to increase ECE, showing a trade-off rather than synergy. Clarification is needed on why these mechanisms must be combined rather than sequentially applied as separate approaches.\n\n9. Baseline (CE Only + PTS) Already Achieving State-of-the-Art Performance:\n   In the result tables, the baseline (CE Only + PTS) already achieves state-of-the-art ECE and accuracy in multiple scenarios. While adding CA and transformation components improves performance further, it seems that these improvements are achieved largely because of the baseline's strong performance. To mitigate this concern, I recommend testing the proposed algorithm on alternative baselines.\n\n10. Minor Points:\n    - The text in figures is too small, making them hard to read.\n    - Typo: Line 136, “samples'.” should be “samples.'” \n\nThese concerns, if addressed, could enhance the clarity and impact of the proposed framework.", "questions": "Why does the paper initially emphasize using a continuous calibration error instead of the Expected Calibration Error (ECE)?\n\nWhat is the intended synergy between the CA loss and the transformation component, given their distinct purposes of reducing ECE and enhancing cross-domain robustness?\n\nCould the proposed algorithm’s effectiveness be validated further by testing it on alternative baselines?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730764197028}, {"id": "HWimR3VaiC", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission867/Reviewer_bjTY"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper describes a method for post-hoc calibration of a classifier based on estimating  for each sample a scaling temperature on the output logits(sample-adaptive calibration strategy). Test time data augmentation is used to predict the scaling temperature and relies on a complementary network taking as input the softmax of selected transformed images and minimizes what is called a correctness-aware loss. The loss is justified by a better management of narrowly wrong predictions. The strategy is evaluated on several small to mid-size datasets and 10 networks per dataset.", "review_text": "The paper describes a method for post-hoc calibration of a classifier based on estimating  for each sample a scaling temperature on the output logits(sample-adaptive calibration strategy). Test time data augmentation is used to predict the scaling temperature and relies on a complementary network taking as input the softmax of selected transformed images and minimizes what is called a correctness-aware loss. The loss is justified by a better management of narrowly wrong predictions. The strategy is evaluated on several small to mid-size datasets and 10 networks per dataset.", "strengths": "- The idea of using test-time augmentation to predict a sample based temperature scaling factor and learning a network for predicting such temperature is novel, as far as I know.\n\n- The justification of the loss on a toy example pointing out its behavior on so-called narrowly wrong samples is intuitive.\n\n- Rather extensive experiments on several types of image datasets show the benefit of the approach over standard calibration methods and other optimization losses.", "weaknesses": "- The goal of the formal development (Section 3.2) is not clear: what is it supposed to show? Is it to prove that the empirical criterion (7) is a good proxy for optimizing (3), given that $\\hat{c}$ is produced by the calibration pipeline of Figure 2? If so, I am not convinced that the formal developments of Section 3.2 actually prove this.\n\n- The writing lacks precision (see my first question, same symbol $E_f^{emp}$ but different concepts for instance). \n\n- The data augmentation is justified by the fact \"that consistency in model predictions for transformed images correlates strongly with accuracy\" (l. 261): if I can agree with this law, I don't clearly see where it applies in your framework. Or is it that by introducing some local perturbation of the data through transformations and measuring the variation in confidence scores, one can infer accuracy?  Then why not directly predict the confidence instead of a temperature?\n\n- In general, I have difficulty understanding the conceptual connections between the test time data augmentation, the formal development, and the narrowly wrong sample analysis. The global logic of the writing is hard to follow.", "questions": "- What is the difference between CA only and CA trans.? Is CA only the calibration strategy that estimates the calibrator $g$ from the calibration set using the loss of Eq. (7) and no data augmentation? This is not clear.\n\n- The approach focuses on calibration of the maximum confidence: can the strategy be adapted to calibrate the whole confidence vector (multiclass calibration)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper describes a method for post-hoc calibration of a classifier based on estimating  for each sample a scaling temperature on the output logits(sample-adaptive calibration strategy). Test time data augmentation is used to predict the scaling temperature and relies on a complementary network taking as input the softmax of selected transformed images and minimizes what is called a correctness-aware loss. The loss is justified by a better management of narrowly wrong predictions. The strategy is evaluated on several small to mid-size datasets and 10 networks per dataset.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "- The idea of using test-time augmentation to predict a sample based temperature scaling factor and learning a network for predicting such temperature is novel, as far as I know.\n\n- The justification of the loss on a toy example pointing out its behavior on so-called narrowly wrong samples is intuitive.\n\n- Rather extensive experiments on several types of image datasets show the benefit of the approach over standard calibration methods and other optimization losses.", "weaknesses": "- The goal of the formal development (Section 3.2) is not clear: what is it supposed to show? Is it to prove that the empirical criterion (7) is a good proxy for optimizing (3), given that $\\hat{c}$ is produced by the calibration pipeline of Figure 2? If so, I am not convinced that the formal developments of Section 3.2 actually prove this.\n\n- The writing lacks precision (see my first question, same symbol $E_f^{emp}$ but different concepts for instance). \n\n- The data augmentation is justified by the fact \"that consistency in model predictions for transformed images correlates strongly with accuracy\" (l. 261): if I can agree with this law, I don't clearly see where it applies in your framework. Or is it that by introducing some local perturbation of the data through transformations and measuring the variation in confidence scores, one can infer accuracy?  Then why not directly predict the confidence instead of a temperature?\n\n- In general, I have difficulty understanding the conceptual connections between the test time data augmentation, the formal development, and the narrowly wrong sample analysis. The global logic of the writing is hard to follow.", "questions": "- What is the difference between CA only and CA trans.? Is CA only the calibration strategy that estimates the calibrator $g$ from the calibration set using the loss of Eq. (7) and no data augmentation? This is not clear.\n\n- The approach focuses on calibration of the maximum confidence: can the strategy be adapted to calibrate the whole confidence vector (multiclass calibration)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730478173953}, {"id": "qkeQVdqYjq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission867/Reviewer_iPzo"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper addresses the issue of model calibration in machine learning, specifically aiming to align a model's confidence with its prediction correctness. The authors identify limitations with the commonly used Cross-Entropy loss for calibrator training and propose a new post-hoc calibration objective, the Correctness-Aware loss. This objective function is designed to decrease model confidence on wrongly predicted samples and increase it on correctly predicted ones. The method utilizes transformed versions of samples to train the calibrator and is tested on both IND and OOD datasets. The paper claims that their method achieves competitive calibration performance compared to state-of-the-art techniques and provides a better separation of correct and incorrect test samples based on calibrated confidence.", "review_text": "The paper addresses the issue of model calibration in machine learning, specifically aiming to align a model's confidence with its prediction correctness. The authors identify limitations with the commonly used Cross-Entropy loss for calibrator training and propose a new post-hoc calibration objective, the Correctness-Aware loss. This objective function is designed to decrease model confidence on wrongly predicted samples and increase it on correctly predicted ones. The method utilizes transformed versions of samples to train the calibrator and is tested on both IND and OOD datasets. The paper claims that their method achieves competitive calibration performance compared to state-of-the-art techniques and provides a better separation of correct and incorrect test samples based on calibrated confidence.", "strengths": "**Novel Calibration Objective:** The paper introduces a new loss function, CA loss, which is a significant contribution to the field of model calibration. This loss function is intuitively designed to align with the goal of calibration, which is to ensure high confidence for correct predictions and low confidence for incorrect ones.\n\n**Empirical Evidence:** The authors provide extensive experimental results demonstrating the effectiveness of their proposed method across various datasets, including IND and OOD test sets. The consistent performance improvement over uncalibrated models and other calibration techniques is a strong point.\n\n**Theoretical Insights:** The paper not only proposes a new method but also provides theoretical insights into why existing methods like CE and MSE losses are limited, particularly for certain types of samples in the calibration set.", "weaknesses": "**Dependency on Transformations:** The effectiveness of the CA loss relies on the use of transformed images to infer correctness. If these transformations do not adequately capture the characteristics of correct and incorrect predictions, the calibration might be less effective.\n\n**Transfomations lack of theoretics:** While the use of transformations such as rotation, grayscale, color jittering, and others has proven to be effective in practice; however, the choice of transformations and their number in Fig. 4 are currently guided more by empirical results rather than a theoretical framework that explains why these five transformations should correlate with prediction correctness as so many transformation exists. And the paper also does not provide a theoretical basis for which transformations are the most informative for calibration or how to select the optimal set of transformations. The current approach might be seen as somewhat arbitrary, and the effectiveness could be dependent on the specific characteristics of the dataset and the model architecture. And There is a risk that the calibrator might overfit to the specific transformations used during training, which may not generalize well to real-world variations in data that were not captured by the training transformations", "questions": "1. See above weakness\n\n2. What is the core difference between calibration and misclassification (e.g. [R1]), both of them seem to be focusing on the incorrect predictions.\n\n3. Fig. 6  illustrates the impact of ablating the top-k selection on the CA loss. The figure suggests that increasing k beyond 4 leads to a significant decline in performance. This trend raises questions about the potential effects of even higher values of k, such as 100 or 200, particularly in datasets like ImageNet. Additionally, since the authors have chosen k=4 as the default setting, it is important to consider how the model manages scenarios where the correct prediction is not included among the top-4 predictions.\n\n4. The method involves training a calibrator with a new loss function and using transformed images, which could be more complex to implement compared to simpler calibration techniques.\n\n [R1]  Zhu, Fei, et al. \"Openmix: Exploring outlier samples for misclassification detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the issue of model calibration in machine learning, specifically aiming to align a model's confidence with its prediction correctness. The authors identify limitations with the commonly used Cross-Entropy loss for calibrator training and propose a new post-hoc calibration objective, the Correctness-Aware loss. This objective function is designed to decrease model confidence on wrongly predicted samples and increase it on correctly predicted ones. The method utilizes transformed versions of samples to train the calibrator and is tested on both IND and OOD datasets. The paper claims that their method achieves competitive calibration performance compared to state-of-the-art techniques and provides a better separation of correct and incorrect test samples based on calibrated confidence.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**Novel Calibration Objective:** The paper introduces a new loss function, CA loss, which is a significant contribution to the field of model calibration. This loss function is intuitively designed to align with the goal of calibration, which is to ensure high confidence for correct predictions and low confidence for incorrect ones.\n\n**Empirical Evidence:** The authors provide extensive experimental results demonstrating the effectiveness of their proposed method across various datasets, including IND and OOD test sets. The consistent performance improvement over uncalibrated models and other calibration techniques is a strong point.\n\n**Theoretical Insights:** The paper not only proposes a new method but also provides theoretical insights into why existing methods like CE and MSE losses are limited, particularly for certain types of samples in the calibration set.", "weaknesses": "**Dependency on Transformations:** The effectiveness of the CA loss relies on the use of transformed images to infer correctness. If these transformations do not adequately capture the characteristics of correct and incorrect predictions, the calibration might be less effective.\n\n**Transfomations lack of theoretics:** While the use of transformations such as rotation, grayscale, color jittering, and others has proven to be effective in practice; however, the choice of transformations and their number in Fig. 4 are currently guided more by empirical results rather than a theoretical framework that explains why these five transformations should correlate with prediction correctness as so many transformation exists. And the paper also does not provide a theoretical basis for which transformations are the most informative for calibration or how to select the optimal set of transformations. The current approach might be seen as somewhat arbitrary, and the effectiveness could be dependent on the specific characteristics of the dataset and the model architecture. And There is a risk that the calibrator might overfit to the specific transformations used during training, which may not generalize well to real-world variations in data that were not captured by the training transformations", "questions": "1. See above weakness\n\n2. What is the core difference between calibration and misclassification (e.g. [R1]), both of them seem to be focusing on the incorrect predictions.\n\n3. Fig. 6  illustrates the impact of ablating the top-k selection on the CA loss. The figure suggests that increasing k beyond 4 leads to a significant decline in performance. This trend raises questions about the potential effects of even higher values of k, such as 100 or 200, particularly in datasets like ImageNet. Additionally, since the authors have chosen k=4 as the default setting, it is important to consider how the model manages scenarios where the correct prediction is not included among the top-4 predictions.\n\n4. The method involves training a calibrator with a new loss function and using transformed images, which could be more complex to implement compared to simpler calibration techniques.\n\n [R1]  Zhu, Fei, et al. \"Openmix: Exploring outlier samples for misclassification detection.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730453254084}, {"id": "4acUWF0wzc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission867/Reviewer_LQKt"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper undertakes the problem of calibrating deep neural networks for the task of classification. At the core of the method is a past-hoc calibrator which proposes a correctness-aware loss to search for the optimal temperature which is then used to scale the logits for a given sample. To determine the correctness of a sample, the method uses the well-known concept of consistency across different augmentations. A simple network is used to map top-K softmax predictions across augmentations to the temperature value. The correctness-aware loss optimizes this network to obtain the best temperature. The paper also shows mathematical insights on the proposed loss. The experiments have been conducted on different datasets to validate the effectiveness of the post-hoc calibrator. Results claim to achieve competitive performance against other post-hoc calibration methods, such as naive temperature scaling, ensemble temperature scaling, adaptive temperature scaling and isotonic regression.", "review_text": "The paper undertakes the problem of calibrating deep neural networks for the task of classification. At the core of the method is a past-hoc calibrator which proposes a correctness-aware loss to search for the optimal temperature which is then used to scale the logits for a given sample. To determine the correctness of a sample, the method uses the well-known concept of consistency across different augmentations. A simple network is used to map top-K softmax predictions across augmentations to the temperature value. The correctness-aware loss optimizes this network to obtain the best temperature. The paper also shows mathematical insights on the proposed loss. The experiments have been conducted on different datasets to validate the effectiveness of the post-hoc calibrator. Results claim to achieve competitive performance against other post-hoc calibration methods, such as naive temperature scaling, ensemble temperature scaling, adaptive temperature scaling and isotonic regression.", "strengths": "**1**) Calibrating deep neural networks is an important step towards making AI models reliable and trustworthy, especially in safety-critical applications.\n\n**2**) The proposed post-hoc calibrator is simple as it also learns to identify a per-sample temperature value that can be used to scale the logits.\n\n**3**) The paper also mentions some theoretical insights into the proposed correctness-aware loss term by comparing and contrasting it with CE and MSE losses.\n\n**4**) Results show that proposed idea is competitive against other post-hoc calibration methods.", "weaknesses": "**1**) The related work section completely misses an emerging direction of train-time calibration methods such as [A], [B], [C],  [D],  [E] and [F]. \n\n**2**) The paper lacks reliability diagrams to better understand the potential of proposed post-hoc calibrator in overcoming overconfidence and under confidence over the full spectrum of model confidence.\n\n**3**) Why the proposed post-hoc calibrator is able to improve OOD calibration performance? There is no analyses that supports these results.\n\n**4**) How the proposed post-hoc calibrator would perform under class-imbalanced scenarios?\n\n**5**) The proposed correctness-aware loss appears similar to MDCA loss [C]. What are the key differences?\n\n\n[A] Liu, B., Ben Ayed, I., Galdran, A. and Dolz, J., 2022. The devil is in the margin: Margin-based label smoothing for network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 80-88).\n\n[B] Patra, R., Hebbalaguppe, R., Dash, T., Shroff, G. and Vig, L., 2023. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 1541-1549)\n\n[C] Hebbalaguppe, R., Prakash, J., Madan, N. and Arora, C., 2022. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16081-16090).\n\n[D] Wei, H., Xie, R., Cheng, H., Feng, L., An, B. and Li, Y., 2022, June. Mitigating neural network overconfidence with logit normalization. In International conference on machine learning (pp. 23631-23644). PMLR.\n\n[E] Liu, B., Rony, J., Galdran, A., Dolz, J. and Ben Ayed, I., 2023. Class adaptive network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16070-16079).\n\n[F] Park, H., Noh, J., Oh, Y., Baek, D. and Ham, B., 2023. Acls: Adaptive and conditional label smoothing for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 3936-3945).\n\n[G] Liang, G., Zhang, Y., Wang, X. and Jacobs, N., Improved Trainable Calibration Method for Neural Networks on Medical Imaging Classification BMVC 2020.\n\n[H] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in\ndeep learning. In CVPR Workshops, volume 2, 201", "questions": "**1**)  What are the key differences with MDCA loss [C] and DCA loss [G] ? I would like to see concrete differences between them.\n\n**2**)  Can MDCA loss and/or DCA loss be used in place of correctness-aware loss to obtain optimal temperature value? Beyond, CE and MSE losses, I believe it would be an interesting comparison between the effectiveness of proposed CA loss and these losses\n\n**3**)  Is the post-hoc calibrator capable of calibrating non-ground truth classes as well?\n\n**4**) What is the performance of the method under the SCE metric [H] compared to other post-hoc calibration methods? \n\n**5**) The intuition behind learning a mapping through g network from top-K softmax scores (corresponding to transformed versions) to temperature value is not very clear. \n\n**6**) L499: The paper mentions that existing methods do not improve AuC compared to proposed one. Will require more explanation.\n\n**7**) How good is the method in overcoming under confidence of the model?\n\n**8**) Can this post-hoc calibrator be used after a train-time calibration method? It would be interesting to observe the complementary strengths of the proposed post-hoc calibration method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper undertakes the problem of calibrating deep neural networks for the task of classification. At the core of the method is a past-hoc calibrator which proposes a correctness-aware loss to search for the optimal temperature which is then used to scale the logits for a given sample. To determine the correctness of a sample, the method uses the well-known concept of consistency across different augmentations. A simple network is used to map top-K softmax predictions across augmentations to the temperature value. The correctness-aware loss optimizes this network to obtain the best temperature. The paper also shows mathematical insights on the proposed loss. The experiments have been conducted on different datasets to validate the effectiveness of the post-hoc calibrator. Results claim to achieve competitive performance against other post-hoc calibration methods, such as naive temperature scaling, ensemble temperature scaling, adaptive temperature scaling and isotonic regression.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "**1**) Calibrating deep neural networks is an important step towards making AI models reliable and trustworthy, especially in safety-critical applications.\n\n**2**) The proposed post-hoc calibrator is simple as it also learns to identify a per-sample temperature value that can be used to scale the logits.\n\n**3**) The paper also mentions some theoretical insights into the proposed correctness-aware loss term by comparing and contrasting it with CE and MSE losses.\n\n**4**) Results show that proposed idea is competitive against other post-hoc calibration methods.", "weaknesses": "**1**) The related work section completely misses an emerging direction of train-time calibration methods such as [A], [B], [C],  [D],  [E] and [F]. \n\n**2**) The paper lacks reliability diagrams to better understand the potential of proposed post-hoc calibrator in overcoming overconfidence and under confidence over the full spectrum of model confidence.\n\n**3**) Why the proposed post-hoc calibrator is able to improve OOD calibration performance? There is no analyses that supports these results.\n\n**4**) How the proposed post-hoc calibrator would perform under class-imbalanced scenarios?\n\n**5**) The proposed correctness-aware loss appears similar to MDCA loss [C]. What are the key differences?\n\n\n[A] Liu, B., Ben Ayed, I., Galdran, A. and Dolz, J., 2022. The devil is in the margin: Margin-based label smoothing for network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 80-88).\n\n[B] Patra, R., Hebbalaguppe, R., Dash, T., Shroff, G. and Vig, L., 2023. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 1541-1549)\n\n[C] Hebbalaguppe, R., Prakash, J., Madan, N. and Arora, C., 2022. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16081-16090).\n\n[D] Wei, H., Xie, R., Cheng, H., Feng, L., An, B. and Li, Y., 2022, June. Mitigating neural network overconfidence with logit normalization. In International conference on machine learning (pp. 23631-23644). PMLR.\n\n[E] Liu, B., Rony, J., Galdran, A., Dolz, J. and Ben Ayed, I., 2023. Class adaptive network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16070-16079).\n\n[F] Park, H., Noh, J., Oh, Y., Baek, D. and Ham, B., 2023. Acls: Adaptive and conditional label smoothing for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 3936-3945).\n\n[G] Liang, G., Zhang, Y., Wang, X. and Jacobs, N., Improved Trainable Calibration Method for Neural Networks on Medical Imaging Classification BMVC 2020.\n\n[H] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in\ndeep learning. In CVPR Workshops, volume 2, 201", "questions": "**1**)  What are the key differences with MDCA loss [C] and DCA loss [G] ? I would like to see concrete differences between them.\n\n**2**)  Can MDCA loss and/or DCA loss be used in place of correctness-aware loss to obtain optimal temperature value? Beyond, CE and MSE losses, I believe it would be an interesting comparison between the effectiveness of proposed CA loss and these losses\n\n**3**)  Is the post-hoc calibrator capable of calibrating non-ground truth classes as well?\n\n**4**) What is the performance of the method under the SCE metric [H] compared to other post-hoc calibration methods? \n\n**5**) The intuition behind learning a mapping through g network from top-K softmax scores (corresponding to transformed versions) to temperature value is not very clear. \n\n**6**) L499: The paper mentions that existing methods do not improve AuC compared to proposed one. Will require more explanation.\n\n**7**) How good is the method in overcoming under confidence of the model?\n\n**8**) Can this post-hoc calibrator be used after a train-time calibration method? It would be interesting to observe the complementary strengths of the proposed post-hoc calibration method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730096767162}], "openreview_url": "https://openreview.net/forum?id=34xYxTTiM0", "arxiv_id": "2404.13016", "paper_pdf": "papers/34xYxTTiM0.pdf", "paper_pdf_sha256": "9489674b98bee80691e307b0d2c9341498857dfbba96d439d93edb72850682c0", "paper_pdf_bytes": 3912860, "paper_pdf_source": "openreview", "code_url": "https://github.com/liuyvchi/Correctness-aware-calibration", "code_repository": "liuyvchi/Correctness-aware-calibration", "code_commit": "6159521ed5ba759b6669372da6990cd4e83abf34", "code_archive": "repos/34xYxTTiM0.zip", "code_archive_sha256": "a27f9edc46742905cb499f07e602e5a55b567f1774fb390f63879b80814e6b6d", "code_archive_bytes": 210809, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 233, "github_languages": {"Python": 96437}, "github_archived": false, "github_pushed_at": "2025-02-12T05:15:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimizing-calibration-by-gaining-aware-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RtOTTdWbZd", "year": 2024, "status": "rejected", "title": "Fine-Tuning Language Models with Advantage-Induced Policy Alignment", "authors": ["Banghua Zhu", "Hiteshi Sharma", "Felipe Vieira Frujeri", "Shi Dong", "Chenguang Zhu", "Michael Jordan", "Jiantao Jiao"], "authorids": ["~Banghua_Zhu1", "~Hiteshi_Sharma1", "~Felipe_Vieira_Frujeri1", "~Shi_Dong1", "~Chenguang_Zhu1", "~Michael_Jordan1", "~Jiantao_Jiao1"], "authors_source": "OpenReview API", "abstract": "Reinforcement learning from human feedback (RLHF) has emerged as a reliable approach to aligning large language models (LLMs) to human preferences.\nAmong the plethora of RLHF techniques, proximal policy optimization (PPO) is of the most widely used methods. Despite its popularity, however, PPO may suffer from mode collapse, instability, and poor sample efficiency.\nWe show that these issues can be alleviated by a novel algorithm that we refer to as Advantage-Induced Policy Alignment (APA), which leverages a squared error loss function based on the estimated advantages.\nWe demonstrate empirically that APA consistently outperforms PPO in language tasks by a large margin, when a separate reward model is employed as the evaluator.\nIn addition, compared with PPO, APA offers a more stable form of control over the deviation from the model's initial policy, ensuring that the model improves its performance without collapsing to deterministic output.\nIn addition to empirical results, we also provide a theoretical justification supporting the design of our loss function.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "rPKC7oZ6GJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8437/Reviewer_GAYy"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a new algorithm for fine-tuning language models with RLHF. The algorithm, APA, uses a squared error loss function that aligns the output policy of the language model with a target policy based on the estimated advantages. The authors claim that APA has advantages over existing RLHF methods, such as PPO and AWR, in terms of sample efficiency, stability, and hyperparameter sensitivity. Besides, the paper provides theoretical justification for APA. The proposed algorithm is evaluated on two language tasks: StackExchange question answering and HH dataset, and outperforms existing methods such as PPO and AWR in terms of sample efficiency, stability, and KL control.", "review_text": "This paper proposes a new algorithm for fine-tuning language models with RLHF. The algorithm, APA, uses a squared error loss function that aligns the output policy of the language model with a target policy based on the estimated advantages. The authors claim that APA has advantages over existing RLHF methods, such as PPO and AWR, in terms of sample efficiency, stability, and hyperparameter sensitivity. Besides, the paper provides theoretical justification for APA. The proposed algorithm is evaluated on two language tasks: StackExchange question answering and HH dataset, and outperforms existing methods such as PPO and AWR in terms of sample efficiency, stability, and KL control.", "strengths": "The research objective is clear and the paper is well-motivated. I do notice that PPO is sometimes unstable for training language models, and the model performance may drop with the training goes on. The proposed algorithm is simple and seems effective on the evaluation tasks. I appreciate the authors' effort in providing theoretical justification for the design of the loss function and the convergence of the algorithm.", "weaknesses": "The major reason I tend to reject is the scope of evaluate tasks. As RLHF (with PPO) has been well verified on ChatGPT, which has great generalization ability, PPO can be utilized to train language models in scale. In addition to training language models, PPO has stand the test of time in many other area such as robotics control. I believe that is why researchers use PPO to fine-tune many large language models. However, the evaluation tasks in this paper locate in a specific domain. I think it is not enough to be an alternative to PPO.\n\nBesides, I also have some minor concerns:\n1. The paper does not provide much discussion on the choice of the hyperparameter $\\lambda$, which controls the trade-off between the expected advantage and the KL divergence. Besides, as this paper proposes an optimization algorithm, the ablation study is required.\n2. There seems lacking an analyze the qualitative differences between the outputs of different algorithms, or the potential harms or biases of the reward models.", "questions": "My major concerns have been listed in 'Weakness' section. Here are some other questions:\n\n1. Can APA generalize to other language tasks or domains? \n2. What do you think are important for a RL algorithm in RLHF training?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new algorithm for fine-tuning language models with RLHF. The algorithm, APA, uses a squared error loss function that aligns the output policy of the language model with a target policy based on the estimated advantages. The authors claim that APA has advantages over existing RLHF methods, such as PPO and AWR, in terms of sample efficiency, stability, and hyperparameter sensitivity. Besides, the paper provides theoretical justification for APA. The proposed algorithm is evaluated on two language tasks: StackExchange question answering and HH dataset, and outperforms existing methods such as PPO and AWR in terms of sample efficiency, stability, and KL control.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The research objective is clear and the paper is well-motivated. I do notice that PPO is sometimes unstable for training language models, and the model performance may drop with the training goes on. The proposed algorithm is simple and seems effective on the evaluation tasks. I appreciate the authors' effort in providing theoretical justification for the design of the loss function and the convergence of the algorithm.", "weaknesses": "The major reason I tend to reject is the scope of evaluate tasks. As RLHF (with PPO) has been well verified on ChatGPT, which has great generalization ability, PPO can be utilized to train language models in scale. In addition to training language models, PPO has stand the test of time in many other area such as robotics control. I believe that is why researchers use PPO to fine-tune many large language models. However, the evaluation tasks in this paper locate in a specific domain. I think it is not enough to be an alternative to PPO.\n\nBesides, I also have some minor concerns:\n1. The paper does not provide much discussion on the choice of the hyperparameter $\\lambda$, which controls the trade-off between the expected advantage and the KL divergence. Besides, as this paper proposes an optimization algorithm, the ablation study is required.\n2. There seems lacking an analyze the qualitative differences between the outputs of different algorithms, or the potential harms or biases of the reward models.", "questions": "My major concerns have been listed in 'Weakness' section. Here are some other questions:\n\n1. Can APA generalize to other language tasks or domains? \n2. What do you think are important for a RL algorithm in RLHF training?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699085970541}, {"id": "s7TuWPl77I", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8437/Reviewer_qc78"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This work presents Advantage Induced Policy-Alignment (APA) for language model finetuning. This work first identifies a source of instability and optimization target mismatch in Advantage Weighted Regression (AWR) and addresses it with APA by minimizing a different f-divergence rather than the KL. They show on multiple RLHF tasks that APA leads to more stable and sample-efficient learning when compared to online RL (PPO) and AWR.", "review_text": "This work presents Advantage Induced Policy-Alignment (APA) for language model finetuning. This work first identifies a source of instability and optimization target mismatch in Advantage Weighted Regression (AWR) and addresses it with APA by minimizing a different f-divergence rather than the KL. They show on multiple RLHF tasks that APA leads to more stable and sample-efficient learning when compared to online RL (PPO) and AWR.", "strengths": "- Good presentation of the idea in a clear way. The paper is well written and easy to follow.\n- Experimentation. Multiple tasks are evaluated with a good discussion/comparison between the RLHF tradeoff of KL from SFT model as well as reward optimization.", "weaknesses": "Minor comments: \n- In the main text, there is a reference to a connection to soft-q learning but it seems like only the f-divergence interpretation is discussed.\n- For the 125M parameter experiments, it seems like PPO is not properly tuned to optimize for the reward.", "questions": "How important was it to estimate a good critic? Was there anything else that was needed to learn a better A rather than having a separate model? This seems to be an interesting angle to experiment some more since the provable guarantees also relies on having a good advantage function for $\\pi_{old}$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents Advantage Induced Policy-Alignment (APA) for language model finetuning. This work first identifies a source of instability and optimization target mismatch in Advantage Weighted Regression (AWR) and addresses it with APA by minimizing a different f-divergence rather than the KL. They show on multiple RLHF tasks that APA leads to more stable and sample-efficient learning when compared to online RL (PPO) and AWR.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "- Good presentation of the idea in a clear way. The paper is well written and easy to follow.\n- Experimentation. Multiple tasks are evaluated with a good discussion/comparison between the RLHF tradeoff of KL from SFT model as well as reward optimization.", "weaknesses": "Minor comments: \n- In the main text, there is a reference to a connection to soft-q learning but it seems like only the f-divergence interpretation is discussed.\n- For the 125M parameter experiments, it seems like PPO is not properly tuned to optimize for the reward.", "questions": "How important was it to estimate a good critic? Was there anything else that was needed to learn a better A rather than having a separate model? This seems to be an interesting angle to experiment some more since the provable guarantees also relies on having a good advantage function for $\\pi_{old}$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698765090530}, {"id": "x1WzWUJf3k", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8437/Reviewer_2o5E"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes an alternative approach to the PPO algorithm for RLHF. The proposed approach has a clear mathematical motivation, where the key idea is to replace the KL-regularizer with a squared-error of the log-probability. Empirical results show that the proposed algorithms consistently outperform PPO in some problems.", "review_text": "This paper proposes an alternative approach to the PPO algorithm for RLHF. The proposed approach has a clear mathematical motivation, where the key idea is to replace the KL-regularizer with a squared-error of the log-probability. Empirical results show that the proposed algorithms consistently outperform PPO in some problems.", "strengths": "1 The presentation and intuition is clear, following the theoretical solution of KL-regularized optimization problem with several reasonable modifications. \n\n2 Some of the empirical results consistently demonstrate that the proposed algorithm is promising in terms of stability, sample efficiency, and also reward optimization. The considered datasets are also standard in RLHF literature.\n\nOverall, I feel that the authors have proposed a promising alternative approach to the PPO algorithm. And it would be interesting to see if further industry-level models are aligned with the proposed algorithm in the future. However, the paper is still not ready for publication. I have a couple of questions in the weakness part.", "weaknesses": "1 While I can understand the mathematical derivations along the line and it is great to see that the loss of APA is provably convergent, I am curious why the square loss is better than the KL-divergence (it is because of the guarantee provided in theorem 1?). In a more general sense, as the new loss function can be viewed as a different f-divergence, and we know that with a f-divergence as the regularizer, we can also obtain another variants of (3), do these f-divergences work better than KL? I am not sure whether there are studies in the literature of PPO but I do see a work considering this modification in DPO [1]. Can you comment on this or provide more intuitions why the squared error works better than the KL and how does it compare with general f-divergence? I think a more systematic and comprehensive comparisons among all these common f-divergence choices could largely improve the paper.\n\n2 It seems that the modification also applies to the DRL scenarios beyond the LLMs. But as far as I know, the prominent approach in research and industry is still the KL-based PPO. I am wondering whether this is due to the special case of LLMs or this modification could also outperform PPO in classical applications.\n\n3 After looking at the dataset in the huggingface, the HH-RLHF dataset used in this paper indeed only contains the helpful part. This is fine as multi-objective alignments could be more complicated and is not the focus of this paper but I think you may explicitly mention this somewhere. \n\n4 Why do you choose to finetune only the last two layers (on the HH dataset in appendix C)? It seems that with the used GPT-J-6B reward model, we might get much higher reward with PPO [2] (although it does depends on the implementation and starting checkpoint). I am not sure whether this is because you only choose to finetune the last two layers but it seems that in practice, we tends to use low-rank adapter instead of freezing most of the layers. \n\n5 In figure 6, the PPO does not work normally as the reward seems to be decrease from the beginning.  I understand that PPO often drops suddenly but this could be mitigated by using larger KL penalty and also smaller learning rate and less epochs for each iteration. I think the PPO is not tuned to the best performance so far and using the default parameters in trlx is not enough.\n\nA minor typo: I think the state space of the LLM should be the product of the token spaces instead of the union.\n\n[1] Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints\n\n[2] RRHF: Rank Responses to Align Language Models with Human Feedback without tears", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an alternative approach to the PPO algorithm for RLHF. The proposed approach has a clear mathematical motivation, where the key idea is to replace the KL-regularizer with a squared-error of the log-probability. Empirical results show that the proposed algorithms consistently outperform PPO in some problems.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "1 The presentation and intuition is clear, following the theoretical solution of KL-regularized optimization problem with several reasonable modifications. \n\n2 Some of the empirical results consistently demonstrate that the proposed algorithm is promising in terms of stability, sample efficiency, and also reward optimization. The considered datasets are also standard in RLHF literature.\n\nOverall, I feel that the authors have proposed a promising alternative approach to the PPO algorithm. And it would be interesting to see if further industry-level models are aligned with the proposed algorithm in the future. However, the paper is still not ready for publication. I have a couple of questions in the weakness part.", "weaknesses": "1 While I can understand the mathematical derivations along the line and it is great to see that the loss of APA is provably convergent, I am curious why the square loss is better than the KL-divergence (it is because of the guarantee provided in theorem 1?). In a more general sense, as the new loss function can be viewed as a different f-divergence, and we know that with a f-divergence as the regularizer, we can also obtain another variants of (3), do these f-divergences work better than KL? I am not sure whether there are studies in the literature of PPO but I do see a work considering this modification in DPO [1]. Can you comment on this or provide more intuitions why the squared error works better than the KL and how does it compare with general f-divergence? I think a more systematic and comprehensive comparisons among all these common f-divergence choices could largely improve the paper.\n\n2 It seems that the modification also applies to the DRL scenarios beyond the LLMs. But as far as I know, the prominent approach in research and industry is still the KL-based PPO. I am wondering whether this is due to the special case of LLMs or this modification could also outperform PPO in classical applications.\n\n3 After looking at the dataset in the huggingface, the HH-RLHF dataset used in this paper indeed only contains the helpful part. This is fine as multi-objective alignments could be more complicated and is not the focus of this paper but I think you may explicitly mention this somewhere. \n\n4 Why do you choose to finetune only the last two layers (on the HH dataset in appendix C)? It seems that with the used GPT-J-6B reward model, we might get much higher reward with PPO [2] (although it does depends on the implementation and starting checkpoint). I am not sure whether this is because you only choose to finetune the last two layers but it seems that in practice, we tends to use low-rank adapter instead of freezing most of the layers. \n\n5 In figure 6, the PPO does not work normally as the reward seems to be decrease from the beginning.  I understand that PPO often drops suddenly but this could be mitigated by using larger KL penalty and also smaller learning rate and less epochs for each iteration. I think the PPO is not tuned to the best performance so far and using the default parameters in trlx is not enough.\n\nA minor typo: I think the state space of the LLM should be the product of the token spaces instead of the union.\n\n[1] Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints\n\n[2] RRHF: Rank Responses to Align Language Models with Human Feedback without tears", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698636634715}, {"id": "3OVwkVi3Xc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8437/Reviewer_59is"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces Advantage-Induced Policy Alignment (APA), a method that applies advantage weighted regression (AWR) to Language Models (LLMs) alignment tasks and replaces the KL-divergence with a squared error loss function.", "review_text": "This paper introduces Advantage-Induced Policy Alignment (APA), a method that applies advantage weighted regression (AWR) to Language Models (LLMs) alignment tasks and replaces the KL-divergence with a squared error loss function.", "strengths": "The paper is well-written. The analysis and discussion are insightful and contribute to a better understanding of the proposed method.\nExperiments on multiple datasets and LLMs demonstrate significant improvements over AWR.", "weaknesses": "The primary concern lies in the novelty and motivation of the proposed method. It seems that the objective represents a minor adjustment to previous work, where KL divergence (p log(p/q)) is replaced by a new f-divergence (p log^2(p/q)). The motivation behind this change remains unclear. The authors claim that APA no longer requires estimating the importance ratio, but it is worth noting that equation (9), which contains log(π_theta / π_init), still necessitates calculating this ratio. The authors have pointed out discrepancies between the original target and the final loss of AWR as well as the changing of π_old in the online case as motivating factors for APA. However, it appears that APA still faces these issues. Another potential motivation could be sample efficiency, but the authors have not explained why using the new f-divergence is more sample-efficient.\n\n## Theoretical Results:\nThe authors claim to provide a comparison among APA, PPO, and AWR theoretically in the introduction, but only the upper bound of APA is presented in Theorem 1. This leaves readers unable to derive the theoretical advantages of the proposed method. Additionally, the assumption for explaining why Z(s) approximates to 1 is deemed overly strong.\n\n## Experiments:\nGiven that both APA and DPO [1] are derived from the optimal policy and are non-RL methods, it is suggested that APA should be compared with more existing works, such as DPO. \n\nRecent research [2] has highlighted the significance of considering various divergences, each offering its unique balance between alignment and diversity trade-offs. Given that the core contribution of this paper revolves around the introduction of a novel f-divergence for distribution matching, it is imperative that the authors conduct a comparative analysis by pitting the new f-divergence against commonly used f-divergences, such as reverse KL, total variation (TV), and Jensen-Shannon (JS) divergences.\nFigure 2 suggests that PPO may not have converged due to a large KL penalty.\n\n## Typos:\nIn the formation of the PPO loss, \\hat{Adv} should be \"adv.\"\n[1] Rafailov R, Sharma A, Mitchell E, et al. \"Direct preference optimization: Your language model is secretly a reward model.\" ICML Workshop MFPL, 2023.\n[2] Go D, Korbak T, Kruszewski G, et al. Aligning language models with preferences through f-divergence minimization[J]. International Conference on Machine Learning. PMLR, 2023.", "questions": "Recent research [3] argues that the KL penalty can be removed, and the authors themselves find that PPO without KL penalty converges to a higher reward compared to APA, as shown in Figure 7. Does this imply that PPO outperforms APA when using smaller or no KL penalty terms? It is important to note that the reviewer does not believe that the higher cost of KL divergence is the issue.\n\n[3] Gao L, Schulman J, Hilton J. \"Scaling laws for reward model overoptimization.\" International Conference on Machine Learning. PMLR, 2023: 10835-10866.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Advantage-Induced Policy Alignment (APA), a method that applies advantage weighted regression (AWR) to Language Models (LLMs) alignment tasks and replaces the KL-divergence with a squared error loss function.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well-written. The analysis and discussion are insightful and contribute to a better understanding of the proposed method.\nExperiments on multiple datasets and LLMs demonstrate significant improvements over AWR.", "weaknesses": "The primary concern lies in the novelty and motivation of the proposed method. It seems that the objective represents a minor adjustment to previous work, where KL divergence (p log(p/q)) is replaced by a new f-divergence (p log^2(p/q)). The motivation behind this change remains unclear. The authors claim that APA no longer requires estimating the importance ratio, but it is worth noting that equation (9), which contains log(π_theta / π_init), still necessitates calculating this ratio. The authors have pointed out discrepancies between the original target and the final loss of AWR as well as the changing of π_old in the online case as motivating factors for APA. However, it appears that APA still faces these issues. Another potential motivation could be sample efficiency, but the authors have not explained why using the new f-divergence is more sample-efficient.\n\n## Theoretical Results:\nThe authors claim to provide a comparison among APA, PPO, and AWR theoretically in the introduction, but only the upper bound of APA is presented in Theorem 1. This leaves readers unable to derive the theoretical advantages of the proposed method. Additionally, the assumption for explaining why Z(s) approximates to 1 is deemed overly strong.\n\n## Experiments:\nGiven that both APA and DPO [1] are derived from the optimal policy and are non-RL methods, it is suggested that APA should be compared with more existing works, such as DPO. \n\nRecent research [2] has highlighted the significance of considering various divergences, each offering its unique balance between alignment and diversity trade-offs. Given that the core contribution of this paper revolves around the introduction of a novel f-divergence for distribution matching, it is imperative that the authors conduct a comparative analysis by pitting the new f-divergence against commonly used f-divergences, such as reverse KL, total variation (TV), and Jensen-Shannon (JS) divergences.\nFigure 2 suggests that PPO may not have converged due to a large KL penalty.\n\n## Typos:\nIn the formation of the PPO loss, \\hat{Adv} should be \"adv.\"\n[1] Rafailov R, Sharma A, Mitchell E, et al. \"Direct preference optimization: Your language model is secretly a reward model.\" ICML Workshop MFPL, 2023.\n[2] Go D, Korbak T, Kruszewski G, et al. Aligning language models with preferences through f-divergence minimization[J]. International Conference on Machine Learning. PMLR, 2023.", "questions": "Recent research [3] argues that the KL penalty can be removed, and the authors themselves find that PPO without KL penalty converges to a higher reward compared to APA, as shown in Figure 7. Does this imply that PPO outperforms APA when using smaller or no KL penalty terms? It is important to note that the reviewer does not believe that the higher cost of KL divergence is the issue.\n\n[3] Gao L, Schulman J, Hilton J. \"Scaling laws for reward model overoptimization.\" International Conference on Machine Learning. PMLR, 2023: 10835-10866.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698575143915}], "openreview_url": "https://openreview.net/forum?id=RtOTTdWbZd", "arxiv_id": "2306.02231", "paper_pdf": "papers/RtOTTdWbZd.pdf", "paper_pdf_sha256": "83514553cd0be92ef16ed199f6e16be07bbc108eb298e68cc027da105dc7ecf7", "paper_pdf_bytes": 1540874, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/RLHF-APA", "code_repository": "microsoft/RLHF-APA", "code_commit": "cfe7da20dbc6cf9753825a6acd9c3367fd22f7fa", "code_archive": "repos/RtOTTdWbZd.zip", "code_archive_sha256": "848edbbee8d0170d54844d66a473176887102a8c8e36480a779f763610e7d7d5", "code_archive_bytes": 146722, "code_file_count": 50, "code_extensions": {".py": 50}, "github_disk_usage_kb": 157, "github_languages": {"Python": 374040}, "github_archived": true, "github_pushed_at": "2023-08-11T20:43:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fine-tuning-language-models-with-advantage"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3tYvDb4dwab", "year": 2023, "status": "rejected", "title": "Understanding Self-Supervised Pretraining with Part-Aware Representation Learning", "authors": ["Jiyang Qi", "Jie Zhu", "Mingyu Ding", "Xiaokang Chen", "Ping Luo", "Leye Wang", "Xinggang Wang", "Wenyu Liu", "Jingdong Wang"], "authorids": ["~Jiyang_Qi1", "~Jie_Zhu3", "~Mingyu_Ding1", "~Xiaokang_Chen1", "~Ping_Luo2", "~Leye_Wang1", "~Xinggang_Wang1", "~Wenyu_Liu3", "~Jingdong_Wang1"], "authors_source": "OpenReview API", "abstract": "In this paper, we are interested in understanding self-supervised pretraining through studying the capability that self-supervised representation pretraining methods learn part-aware representations. The study is mainly motivated by that random views, used in contrastive learning, and random masked (visible) patches, used in masked image modeling, are often about object parts.\n\nWe explain that masked image modeling is a part-to-part task: the masked patches of the object are hallucinated from the visible patches, and that contrastive learning is a part-to-whole task: the projection layer hallucinates the whole object representation from the object part representation learned from the encoder. The explanation suggests that the self-supervised pretrained encoder is required to understand the object part. We empirically compare the off-the-shelf encoders pretrained with several representative methods on object-level recognition and part-level recognition. The results show that the fully-supervised model outperforms self-supervised models for object-level recognition, and most self-supervised contrastive learning and masked image modeling methods outperform the fully-supervised method for part-level recognition. It is observed that the combination of contrastive learning and masked image modeling further improves the performance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "DY_Aoqx6Ga", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper352/Reviewer_SGtU"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper is a study on what is the different when learning with random crops and contrastive learning (CL) in a supervised learning (SL) or self-supervised learning (SSL) way, versus random patch masking via self-supervised learning (SSL) as in the Masked image modeling (MIM) variants (MAE/BeiT/CAE). The authors claim that the former is learning a \"part-to-whole\" task while the latter a \"part-to-part\" task; both can lead to strong part-aware representations, while it seems that SL excels more on object-level tasks, while SSL methods like CAE or MoCo-v3 are able to learn part-aware representations.", "review_text": "This paper is at its core an empirical study on part-aware representations; the usefulness of the analysis in section 3 is unclear to me, looking forward for some clarifications from the authors on this. There is no technical novelty, while the insights from the study are also not strong, basically verifying the superiority of iBoT over other SSL methods tested on both object and part tasks. I am looking forward to the author's responses.\n\n--- Post rebuttal update:\n\nI want to thank the authors for constructive discussion. I think iff the authors make the edit they promise to the text  and clarify their contributions, I think I can raise my score to borderline accept. Although limited in novelty, the study in this paper can be valuable to the community, if the contributions are not over-claimed. ", "strengths": "### Strengths\n\nS1: the paper is studying part-level recognition, a very interesging and important task\n\nS2: The paper features an extensive empirical study on many object-level recognition and part-level recognition tasks. \n\n\n### Weaknesses\n\nW1: It is unclear to me what the part-to-whole vs  part-to-part analysis offers, especially with respect to part-aware representations: Sections 3.1 and 3.2 seem to suggest the same thing: Both CL and MIM, \"[are] potentially capable of learning part-aware representations.\". Are there any insights or interesting experiments that are derived from that analysis and distinction? Cause what i see in Sec 4 is an empirical analysis on which losses can learn part-aware representation, that can stand without section 3, really.\n\nW2: It is unclear to me if the \"part-to-whole\" effect is only a function of the input (random crops) and the contrastive aspect of the loss; I think that what really matters is what is the contrastive loss appled on: If the loss is on aggregated features from the whole crop, I see this making sense. But what is a contrastive loss is applied at the token level?  How would for example DenseCL (Wang, Xinlong, et al. \"Dense contrastive learning for self-supervised visual pre-training.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.) fits in this framework? \n\nW3: The authors are focusing on random crops, which is only one of the augmentations used for CL - yet there are other augmentations at play. A more fair comparison would use an identical augmentation setup apart from either adding random crops or masking. Are all the other augmentations (color jittering, etc) shared across compared methods?\n\nW4: it is unclear to me, eg from the abstract what is the focus of the study, as I see two  axes that are convoluted: MIM vs CL and SL versus SSL methods. The abstract highlights the first axis, so does Sec 3, but then Figure 2 shows that both CAE (MIM) and MoCov3 (Contrastive) can give part-aware representations, while DeiT (SL) cannot. I think that leaving SL out of the study would make it clearer.\n\nW5: the insights from the empirical study are not analysed enough and can in places seems exargerated:, eg \"This implies that in general the self-supervised models are not strong at object-level understanding,\" this is a very generic statement to make from the results in Table 2. \n\nQ1: Are you retraining all the models under some otherwise identical setup, or using the publicly available models out of the box? Overall, three are many differences between the models compared beyond the loss (from batchsizes, to augmentations, to optimizers to other ViT-related tricks used in different papers). \n\nQ2: Although figure 1 and 3 is mostly figurative it seems to me that it exaggerates the effects of the random crops used in practice. Apart from methods that use local crops as in multi-crop, the scale parameter for the random crops is usually not as large as the one used in this figure. What is the crop scale used in all methods, ie what is the percentage of the max width/height that each readnom crop uses?\n\nQ3: Where are the patches seen in Fig 2 and 4 from? Imagenet?\n\nN1: I wonder how things would change for SL if a more recent variant of the DeiT series was used, e.g. Deit III from \nTouvron, Hugo, Matthieu Cord, and Hervé Jégou. \"Deit iii: Revenge of the vit.\" arXiv preprint arXiv:2204.07118 (2022).\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper is a study on what is the different when learning with random crops and contrastive learning (CL) in a supervised learning (SL) or self-supervised learning (SSL) way, versus random patch masking via self-supervised learning (SSL) as in the Masked image modeling (MIM) variants (MAE/BeiT/CAE). The authors claim that the former is learning a \"part-to-whole\" task while the latter a \"part-to-part\" task; both can lead to strong part-aware representations, while it seems that SL excels more on object-level tasks, while SSL methods like CAE or MoCo-v3 are able to learn part-aware representations.", "strength_and_weaknesses": "### Strengths\n\nS1: the paper is studying part-level recognition, a very interesging and important task\n\nS2: The paper features an extensive empirical study on many object-level recognition and part-level recognition tasks. \n\n\n### Weaknesses\n\nW1: It is unclear to me what the part-to-whole vs  part-to-part analysis offers, especially with respect to part-aware representations: Sections 3.1 and 3.2 seem to suggest the same thing: Both CL and MIM, \"[are] potentially capable of learning part-aware representations.\". Are there any insights or interesting experiments that are derived from that analysis and distinction? Cause what i see in Sec 4 is an empirical analysis on which losses can learn part-aware representation, that can stand without section 3, really.\n\nW2: It is unclear to me if the \"part-to-whole\" effect is only a function of the input (random crops) and the contrastive aspect of the loss; I think that what really matters is what is the contrastive loss appled on: If the loss is on aggregated features from the whole crop, I see this making sense. But what is a contrastive loss is applied at the token level?  How would for example DenseCL (Wang, Xinlong, et al. \"Dense contrastive learning for self-supervised visual pre-training.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.) fits in this framework? \n\nW3: The authors are focusing on random crops, which is only one of the augmentations used for CL - yet there are other augmentations at play. A more fair comparison would use an identical augmentation setup apart from either adding random crops or masking. Are all the other augmentations (color jittering, etc) shared across compared methods?\n\nW4: it is unclear to me, eg from the abstract what is the focus of the study, as I see two  axes that are convoluted: MIM vs CL and SL versus SSL methods. The abstract highlights the first axis, so does Sec 3, but then Figure 2 shows that both CAE (MIM) and MoCov3 (Contrastive) can give part-aware representations, while DeiT (SL) cannot. I think that leaving SL out of the study would make it clearer.\n\nW5: the insights from the empirical study are not analysed enough and can in places seems exargerated:, eg \"This implies that in general the self-supervised models are not strong at object-level understanding,\" this is a very generic statement to make from the results in Table 2. \n\nQ1: Are you retraining all the models under some otherwise identical setup, or using the publicly available models out of the box? Overall, three are many differences between the models compared beyond the loss (from batchsizes, to augmentations, to optimizers to other ViT-related tricks used in different papers). \n\nQ2: Although figure 1 and 3 is mostly figurative it seems to me that it exaggerates the effects of the random crops used in practice. Apart from methods that use local crops as in multi-crop, the scale parameter for the random crops is usually not as large as the one used in this figure. What is the crop scale used in all methods, ie what is the percentage of the max width/height that each readnom crop uses?\n\nQ3: Where are the patches seen in Fig 2 and 4 from? Imagenet?\n\nN1: I wonder how things would change for SL if a more recent variant of the DeiT series was used, e.g. Deit III from \nTouvron, Hugo, Matthieu Cord, and Hervé Jégou. \"Deit iii: Revenge of the vit.\" arXiv preprint arXiv:2204.07118 (2022).\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, but the usefulness oft the analysis of Section 3 is to me unclear. The experimental evaluation is exhaustive and in theory reproducible. There seems to be no technical novelty, the paper is a study on SSL/SL and MIM /CL, starting from pretrained models from other papers.\n", "summary_of_the_review": "This paper is at its core an empirical study on part-aware representations; the usefulness of the analysis in section 3 is unclear to me, looking forward for some clarifications from the authors on this. There is no technical novelty, while the insights from the study are also not strong, basically verifying the superiority of iBoT over other SSL methods tested on both object and part tasks. I am looking forward to the author's responses.\n\n--- Post rebuttal update:\n\nI want to thank the authors for constructive discussion. I think iff the authors make the edit they promise to the text  and clarify their contributions, I think I can raise my score to borderline accept. Although limited in novelty, the study in this paper can be valuable to the community, if the contributions are not over-claimed. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666643866806}, {"id": "_WH78AGc4_a", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper352/Reviewer_nE5Q"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents an analysis of existing self-supervised learning methods with vision transformers. Extensive experiments are performed on the off-the-shelf pretrained model with the linear and attentive probes to verify the speculation that contrastive learning is a part-to-whole task and masked image modeling is a part-to-part task. ", "review_text": "Overall I think this paper goes one step further than CAE and performs an extensive comparison on various benchmarks, which is appreciated. However, some of the experiment results do not align with the authors' claim and the conclusion is somewhat unsurprising and not enlightening. Thus my rating is 5.", "strengths": "Pros:\n\nThe motivation of this paper is clear, the paper is well-written which makes it easy to follow. The experiment setting makes sense to me.\n\nCons:\n\nThe part-to-part and the part-to-whole difference between masked image modeling and contrastive learning is already mentioned in the CAE paper, which makes the finding and conclusion less interesting and surprising.\nSome experiment results do not align with the claim well, for example, in the part-retrieval and part-segmentation experiments, BEiT and MAE which are MIM models do not perform better than contrastive learning or supervised learning.\nBased on that, no more instructive conclusion about how to better use the property is given (combining two of them like iBOT shouldn't count as a contribution in this paper), which makes the novelty part weak.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents an analysis of existing self-supervised learning methods with vision transformers. Extensive experiments are performed on the off-the-shelf pretrained model with the linear and attentive probes to verify the speculation that contrastive learning is a part-to-whole task and masked image modeling is a part-to-part task. ", "strength_and_weaknesses": "Pros:\n\nThe motivation of this paper is clear, the paper is well-written which makes it easy to follow. The experiment setting makes sense to me.\n\nCons:\n\nThe part-to-part and the part-to-whole difference between masked image modeling and contrastive learning is already mentioned in the CAE paper, which makes the finding and conclusion less interesting and surprising.\nSome experiment results do not align with the claim well, for example, in the part-retrieval and part-segmentation experiments, BEiT and MAE which are MIM models do not perform better than contrastive learning or supervised learning.\nBased on that, no more instructive conclusion about how to better use the property is given (combining two of them like iBOT shouldn't count as a contribution in this paper), which makes the novelty part weak.\n", "clarity,_quality,_novelty_and_reproducibility": "See above section", "summary_of_the_review": "Overall I think this paper goes one step further than CAE and performs an extensive comparison on various benchmarks, which is appreciated. However, some of the experiment results do not align with the authors' claim and the conclusion is somewhat unsurprising and not enlightening. Thus my rating is 5.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666587305319}, {"id": "hFOjYM2lg5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper352/Reviewer_AQH3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper tries to understand the two mainstream self-supervised methods. It speculates that masked image modeling is a part-to-part process and contrastive learning is a part-to-whole process. Both empirical results and qualitative analysis are provided to support the claims.", "review_text": "This paper speculates that contrastive learning is a part-to-whole process and MIM is a part-to-part process and self-supervised learning could learn part-aware representations. It provides an innovative perspective to understand self-supervised learning, supported by comprehensive empirical results and qualitative analyses. The paper is well written but some claims are not well supported. I'm also wondering to what extent the conclusions in this paper can help the future model since the different objectives matter more than the training paradigm (contrastive/MIM). I think this paper is a bit above the borderline.", "strengths": "Strengths:\n1. This paper makes an innovative claim on the current self-supervised learning methods, its conclusion leads people to a more insightful understanding of self-supervised learning.\n2. Authors conduct a comprehensive empirical study on both object-level and part-level tasks to support their claims.\n3. Many qualitative analyses are made, which give a clear and intuitive explanation.\n\nWeaknesses:\n1. For the part segmentation part, the authors' explanation, 'possibly due to pretraining quality in representation encoding, BEiT and MAE perform inferior', is not convincing. I do not think it is safe to draw the conclusion based on only one MIM method CAE. Therefore, whether MIM method can be simply described as a part-to-part process is questionable from the experimental results. \n2. I do not see a clear reason why iBOT performs the best from the authors' explanation. It's unclear why combining part-to-part and part-to-whole processes could generate a better model. I'm afraid that the differences among the models/objectives would matter more than the differences in the training paradigm (contrastive/MIM). This could make the claims in this paper less useful for guidance on future SSL models.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper tries to understand the two mainstream self-supervised methods. It speculates that masked image modeling is a part-to-part process and contrastive learning is a part-to-whole process. Both empirical results and qualitative analysis are provided to support the claims.", "strength_and_weaknesses": "Strengths:\n1. This paper makes an innovative claim on the current self-supervised learning methods, its conclusion leads people to a more insightful understanding of self-supervised learning.\n2. Authors conduct a comprehensive empirical study on both object-level and part-level tasks to support their claims.\n3. Many qualitative analyses are made, which give a clear and intuitive explanation.\n\nWeaknesses:\n1. For the part segmentation part, the authors' explanation, 'possibly due to pretraining quality in representation encoding, BEiT and MAE perform inferior', is not convincing. I do not think it is safe to draw the conclusion based on only one MIM method CAE. Therefore, whether MIM method can be simply described as a part-to-part process is questionable from the experimental results. \n2. I do not see a clear reason why iBOT performs the best from the authors' explanation. It's unclear why combining part-to-part and part-to-whole processes could generate a better model. I'm afraid that the differences among the models/objectives would matter more than the differences in the training paradigm (contrastive/MIM). This could make the claims in this paper less useful for guidance on future SSL models.", "clarity,_quality,_novelty_and_reproducibility": "The paper is overall well-written and clearly presented. There are some minor mistakes, in Equations (2) and (3), $P_v$ and $P_m$ are not defined. It sheds some new insights into the current self-supervised learning, which is beneficial to the community. It has a good quality in general, and the experiment details demonstrate its reproducibility. There are some claims in the paper that are not well supported from my point of view, which fails to explain some phenomena in a convincing way.", "summary_of_the_review": "This paper speculates that contrastive learning is a part-to-whole process and MIM is a part-to-part process and self-supervised learning could learn part-aware representations. It provides an innovative perspective to understand self-supervised learning, supported by comprehensive empirical results and qualitative analyses. The paper is well written but some claims are not well supported. I'm also wondering to what extent the conclusions in this paper can help the future model since the different objectives matter more than the training paradigm (contrastive/MIM). I think this paper is a bit above the borderline.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666334821808}], "openreview_url": "https://openreview.net/forum?id=3tYvDb4dwab", "arxiv_id": "2301.11915", "paper_pdf": "papers/3tYvDb4dwab.pdf", "paper_pdf_sha256": "dcb1fcbf7550f9f1bd5121ba5f2a965d435cd4029e65151650b3a87b8e073c70", "paper_pdf_bytes": 7287188, "paper_pdf_source": "openreview", "code_url": "https://github.com/JiePKU/understand-ssl-part-aware", "code_repository": "JiePKU/understand-ssl-part-aware", "code_commit": "ffea818117952525aa7cb05d793c07e94b2f34fb", "code_archive": "repos/3tYvDb4dwab.zip", "code_archive_sha256": "c966c3b925f0ce8ae36631b7defdfeb6f824a4afda2399fbbe054beb9ef39dca", "code_archive_bytes": 448303, "code_file_count": 226, "code_extensions": {".py": 221, ".sh": 5}, "github_disk_usage_kb": 336, "github_languages": {"Python": 1089473, "Shell": 3884, "HTML": 3482}, "github_archived": false, "github_pushed_at": "2024-01-03T03:55:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-self-supervised-pretraining"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WnOLO1f50MH", "year": 2022, "status": "rejected", "title": "Exploiting Redundancy: Separable Group Convolutional Networks on Lie Groups", "authors": ["David M Knigge", "David W. Romero", "Erik J Bekkers"], "authorids": ["~David_M_Knigge1", "~David_W._Romero1", "~Erik_J_Bekkers1"], "authors_source": "OpenReview API", "abstract": "Group convolutional neural networks (G-CNNs) have been shown to increase parameter efficiency and model accuracy by incorporating geometric inductive biases. In this work, we investigate the properties of representations learned by regular G-CNNs, and show considerable parameter redundancy in group convolution kernels. This finding motivates further weight-tying by sharing convolution kernels over subgroups. To this end, we introduce convolution kernels that are separable over the subgroup and channel dimensions. In order to obtain equivariance to arbitrary affine Lie groups we provide a continuous parameterisation of separable convolution kernels. We evaluate our approach across several vision datasets, and show that our weight sharing leads to improved performance and computational efficiency. In many settings, separable G-CNNs outperform their non-separable counterpart, while only using a fraction of their training time. In addition, thanks to the increase in computational efficiency, we are able to implement G-CNNs equivariant to the $\\mathrm{Sim(2)}$ group; the group of dilations, rotations and translations.  $\\mathrm{Sim(2)}$-equivariance further improves performance on all tasks considered.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "PTQfuw_hx3E", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3374/Reviewer_hC39"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper adresses redundancy in group convolutional filters and the scalability of group ConvNets. It is achieved by introducing a separable group convolution for Affine Lie group which allows to share the kernels for translation elements.", "review_text": "The strengths of the paper are as follow :\n\n1/ Clarity of the theoretical and empirical parts\n\n2/  Experimental results are promising\n\n3/ The use of SIREN  to parametrize convolutional kernel is a tricky idea and quite original since it is originally proposed for coordinates space data. As a result, a given layer has a fixed number of parameters regardless the resolution of data\n\n4/ Approximating the group convolution integral through random sampling shows a clear advantage over discretization schemes. It allows to mitigate some artifacts of raw data \n\nThe weaknesses are as follow :\n\n1/ The proposed model is quite generic. In order to show its scalability, one may wonder if the proposed model could be extended to large scale vision tasks : 3D rendering, Video classification and also to other domains like physical processes where group of symmetries are important.\n\n2/ Lack of comparison with state-of-the art models\n\n3/ Need a further theoretical study \n\n\nSmall typos in appendix, before equation (9) \"the group group\"\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper adresses redundancy in group convolutional filters and the scalability of group ConvNets. It is achieved by introducing a separable group convolution for Affine Lie group which allows to share the kernels for translation elements.", "main_review": "The strengths of the paper are as follow :\n\n1/ Clarity of the theoretical and empirical parts\n\n2/  Experimental results are promising\n\n3/ The use of SIREN  to parametrize convolutional kernel is a tricky idea and quite original since it is originally proposed for coordinates space data. As a result, a given layer has a fixed number of parameters regardless the resolution of data\n\n4/ Approximating the group convolution integral through random sampling shows a clear advantage over discretization schemes. It allows to mitigate some artifacts of raw data \n\nThe weaknesses are as follow :\n\n1/ The proposed model is quite generic. In order to show its scalability, one may wonder if the proposed model could be extended to large scale vision tasks : 3D rendering, Video classification and also to other domains like physical processes where group of symmetries are important.\n\n2/ Lack of comparison with state-of-the art models\n\n3/ Need a further theoretical study \n\n\nSmall typos in appendix, before equation (9) \"the group group\"\n\n", "summary_of_the_review": "Following the aforementioned consideration, l recommend to accept the paper. The proposed research direction is promising and could open perspectives to several domains, mainly physical process tasks.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636113930839}, {"id": "oUlRMSoE_x", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3374/Reviewer_m8cZ"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper discusses considerable parameter redundancy in regular group convolution networks and then proposes separable convolution kernels to share the weights over the subgroups. Besides, the authors explored the equivariance of three different groups and presented a continuous parameterization scheme. Evaluations on several datasets show that their proposed method gains improved performance and computational efficiency.", "review_text": "The reviewer is very interested in this research work. This paper explores the shortcomings of existing work in depth, discusses separable convolution kernels, and does the parameterization of different groups more comprehensively, which  is a good research guide. In addition, the authors also conducted experiments on many vision datasets to illustrate the superiority of the proposed method and achieved performance improvements.\n\nHowever, there exist some confusing parts or weaknesses to be further strengthened.\n1. The biggest shortcoming of this paper is its limited innovation. As stated in this paper, the proposed method is more like a promotion of Cohen2016, by using a strategy similar to ‘depthwise separable convolution’ in Xception. It seems that this work depends on the existing method heavily and its original parts bring weak contributions.\n\n2. The theoretical foundation of this paper is not solid. It is worth mentioning that the authors explored arbitrary affine Lie groups (three different groups: SE2, dilation +translation, and dilation +Sim(2)). But as far as the reviewer knows, this is not entirely applicable. For example, correlation is an equivariant map for the translation but NOT for the rotation group. Why can the authors directly consider SE(2) and Sim (2) in convolution (both containing Rotation)? In addition, when proving translation equivariant, it is necessary to extend f and k to group G, otherwise g^(-1)x does not belong to group G and cannot go on. The reviewer has not seen relevant explanations. Regarding Sim(2), it has a totally different Lie algebra expression from SE(2)’s due to the scaling, and there has not existed a discussion on specific forms. \n\n3. The motivation of the proposed method comes from parameter redundancy, but from Figure 1, the reviewer found that the metric is F-test (variance ratio), which is usually used to analyze statistical models that use more than one parameter to determine whether all or part of the parameters are suitable for estimating the population. And the robustness of F-test is extremely dependent on the samples with a normal distribution. In the case of limited samples, the distribution is often biased, which tends to reduce the power of such a test, leading to incorrect statistical inference.\n\n4. The authors used SIREN as kernel parameterization in the Lie algebra. But as far as the reviewer knows, this work proves the superiority of using periodic activation functions (such as sine), which is, however, irrelevant to this paper from the perspective of the reviewer. Much more details should be provided here.\n\n5. Another concern of the reviewer is that evaluation results of this paper are not convincing to the reviewer. Traditional CNN-based methods have achieved 95+% accuracy on CIFAR10, and regular GNNs (p4m & Z2) can also reach 90+%. The accuracy results of this paper are still inferior to 90% with a 77% baseline. It violates the claim that \"G-CNNs usually improve upon regular CNNs\" stated in this paper.\n\n6. Some parts are not novel at all, such as the section “Depthwise separability”. The reviewer suggests the authors to include more relevant references (Xception or more). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper discusses considerable parameter redundancy in regular group convolution networks and then proposes separable convolution kernels to share the weights over the subgroups. Besides, the authors explored the equivariance of three different groups and presented a continuous parameterization scheme. Evaluations on several datasets show that their proposed method gains improved performance and computational efficiency.", "main_review": "The reviewer is very interested in this research work. This paper explores the shortcomings of existing work in depth, discusses separable convolution kernels, and does the parameterization of different groups more comprehensively, which  is a good research guide. In addition, the authors also conducted experiments on many vision datasets to illustrate the superiority of the proposed method and achieved performance improvements.\n\nHowever, there exist some confusing parts or weaknesses to be further strengthened.\n1. The biggest shortcoming of this paper is its limited innovation. As stated in this paper, the proposed method is more like a promotion of Cohen2016, by using a strategy similar to ‘depthwise separable convolution’ in Xception. It seems that this work depends on the existing method heavily and its original parts bring weak contributions.\n\n2. The theoretical foundation of this paper is not solid. It is worth mentioning that the authors explored arbitrary affine Lie groups (three different groups: SE2, dilation +translation, and dilation +Sim(2)). But as far as the reviewer knows, this is not entirely applicable. For example, correlation is an equivariant map for the translation but NOT for the rotation group. Why can the authors directly consider SE(2) and Sim (2) in convolution (both containing Rotation)? In addition, when proving translation equivariant, it is necessary to extend f and k to group G, otherwise g^(-1)x does not belong to group G and cannot go on. The reviewer has not seen relevant explanations. Regarding Sim(2), it has a totally different Lie algebra expression from SE(2)’s due to the scaling, and there has not existed a discussion on specific forms. \n\n3. The motivation of the proposed method comes from parameter redundancy, but from Figure 1, the reviewer found that the metric is F-test (variance ratio), which is usually used to analyze statistical models that use more than one parameter to determine whether all or part of the parameters are suitable for estimating the population. And the robustness of F-test is extremely dependent on the samples with a normal distribution. In the case of limited samples, the distribution is often biased, which tends to reduce the power of such a test, leading to incorrect statistical inference.\n\n4. The authors used SIREN as kernel parameterization in the Lie algebra. But as far as the reviewer knows, this work proves the superiority of using periodic activation functions (such as sine), which is, however, irrelevant to this paper from the perspective of the reviewer. Much more details should be provided here.\n\n5. Another concern of the reviewer is that evaluation results of this paper are not convincing to the reviewer. Traditional CNN-based methods have achieved 95+% accuracy on CIFAR10, and regular GNNs (p4m & Z2) can also reach 90+%. The accuracy results of this paper are still inferior to 90% with a 77% baseline. It violates the claim that \"G-CNNs usually improve upon regular CNNs\" stated in this paper.\n\n6. Some parts are not novel at all, such as the section “Depthwise separability”. The reviewer suggests the authors to include more relevant references (Xception or more). \n", "summary_of_the_review": "The research field of this paper is very interesting, but unfortunately there remain many shortcomings in theoretical and experimental results, which need to be forcefully addressed. The reviewer therefore thinks that the paper is not good enough to be published on ICLR.\n\nAfter reading the authors' rebuttal, I increase my rating a bit. But I am still not contented with the stuff in this submission, which should undergo a thorough revision to convince and be easily understood by the readers.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635915584055}, {"id": "N1kqS_Ta9qT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3374/Reviewer_xJAv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper builds group convolutional neural networks based on the depth-wise separable convolution operations, which are commonly seen in modern CNNs. The authors demonstrate that the $Sim(2)$-equivariant can be achieved in such separable convolution operations. In the implementation, the authors borrow the SIREN approach for parameterizing the proposed group convolutions. Finally, the experimental analysis shows good improvement over other types of group convolutions.", "review_text": "This is a solid paper to explore the combination of group convolutions and techniques employed in popular CNN architectures. The ideas of how we can achieve $Sim(2)$-equivariant on separable convolution operations are clearly explained, they are helpful in designing kernels with better performance. Besides, the paper itself is educational to readers not familiar with this field, I recommend accepting this paper to let more researchers benefit from the ideas behind G-CNNs.\n\nI have a few questions to ask:\n\n1. The motivation is explained as the group convolutions learn redundancies. Is this only found in G-CNNs or all types of CNNs? Do separable convolution operations address this problem finally? Can we get a figure similar to Figure 2?\n\n2. Although the utilization of SIREN is the highlight of the paper, the motivation behind is unexplained and unclear. Why the group convolution kernels can be represented implicitly? Is SIREN the only choice?  Do larger SIREN networks provide better parameterization?\n\n3. Why does random sampling over subgroups provide better performance? I am baffled by these results, if we can design a discretization sampling method that has similar anti-aliasing effects?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper builds group convolutional neural networks based on the depth-wise separable convolution operations, which are commonly seen in modern CNNs. The authors demonstrate that the $Sim(2)$-equivariant can be achieved in such separable convolution operations. In the implementation, the authors borrow the SIREN approach for parameterizing the proposed group convolutions. Finally, the experimental analysis shows good improvement over other types of group convolutions.", "main_review": "This is a solid paper to explore the combination of group convolutions and techniques employed in popular CNN architectures. The ideas of how we can achieve $Sim(2)$-equivariant on separable convolution operations are clearly explained, they are helpful in designing kernels with better performance. Besides, the paper itself is educational to readers not familiar with this field, I recommend accepting this paper to let more researchers benefit from the ideas behind G-CNNs.\n\nI have a few questions to ask:\n\n1. The motivation is explained as the group convolutions learn redundancies. Is this only found in G-CNNs or all types of CNNs? Do separable convolution operations address this problem finally? Can we get a figure similar to Figure 2?\n\n2. Although the utilization of SIREN is the highlight of the paper, the motivation behind is unexplained and unclear. Why the group convolution kernels can be represented implicitly? Is SIREN the only choice?  Do larger SIREN networks provide better parameterization?\n\n3. Why does random sampling over subgroups provide better performance? I am baffled by these results, if we can design a discretization sampling method that has similar anti-aliasing effects?", "summary_of_the_review": "The bad thing of such mathematical ideas inspired paper is that the experimental performance are always far from sota. But I think the performance is not the only thing, it is not bad to accept this paper for me.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635875360791}, {"id": "xFQ9BH0hCHu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3374/Reviewer_4A1c"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes to use separable convolutions along the group dimension in the type of group CNNs proposed by Finzi et al [1]. The motivation is the same as the popular depthwise separable convolutions for conventional CNNs: reducing parameter redundancy to increase efficiency and accuracy. Experiments on rotated MNIST, CIFAR10/100, and Galaxy10 show that the method outperforms the non-separable versions both in accuracy and speed.\n", "review_text": "*Strengths*\n\nS1: The paper is well-written, the method introduced makes sense and seems to work well in practice. \n\n*Weaknesses*\n\nW1: The submission has somewhat limited novelty. It can be seen as an application of depthwise convolutions to group-CNNs. Depthwise convolutions are quite popular since at least 2017 (Xception, Mobilenets); and the (non-separable) group-CNN utilized seems the same as Finzi et al [1], with sinusoidal activations. More importantly, the idea of separating the group-convolution along the group dimension has also appeared in Lengyel and van Gemert [2], although for a different flavor of G-CNN. \n\nW2: I see several issues with the experimental section. The most problematic is that all comparisons are against the paper's own baselines. While this makes for fair comparisons in the sense that the data pipeline, architectures and training schedule are all the same (although I have reservations about the number of parameters, see W3), I'm not sure if the conclusions hold in general. For example Weiler and Cesa [3] seems to show significantly better performance than the proposed method on rotated MNIST and CIFAR10/100; can the proposed method improve those results? If not, what would be the applications where the method is useful? Finzi et al [1] also show favorable results on a molecular property prediction, how does the submission fare in that task? When comparing with Finzi et al, I believe there should also be an ablation to disentangle the effects of the separable convolutions and the use of SIREN instead of a regular MLP.\n\nW3: Please report the number of parameters for each model trained. If I understand correctly, the number of channels per layer is kept constant, so the non-separable models have several times more parameters than the separable, is that correct? In that case could the lower performance on non-separable be explained by overfitting or slower convergence? I think both constant number of channels per layer or constant number of parameters are informative and should be reported (Weiler and Cesa [3] do something similar for CIFAR). \n\nW4: In Table 5, I believe a few experiments are missing and would be needed to disentangle the effects of the separation along the group dimension and the channel dimensions. I suggest to show, for each group, the performance when the convolutions are separable over the channel (depthwise) but not the group dimension. And for the baseline, it would be interesting to also see the performance for the separable depthwise version. \n\n*Questions*\n\nQ1: For the rotated MNIST, it is shown that approximating the convolution with random samples is superior, however the other MNIST variations seem to fall back to the discretization. Why is this the case? Do the random sampling approximation perform worse on the higher dimensional groups?\n\nQ2: The SIM(2) MNIST experiment is described as limited to 2, 4, 6, or 8 elements for each subgroup. Does this also refer to the way the dataset is constructed, or is it created with random rotations and scaling sampled from the continuous internal?\n\n*References*\n\n[1] Finzi et al, \"Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data\", ICML'20.\n\n[2] Lengyel and van Gemert, \"Exploiting Learned Symmetries in Group Equivariant Convolutions\", ICIP'21.\n\n[3] Weiler and Cesa, \"General E(2) - Equivariant Steerable CNNs\", NeurIPS'19.         ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to use separable convolutions along the group dimension in the type of group CNNs proposed by Finzi et al [1]. The motivation is the same as the popular depthwise separable convolutions for conventional CNNs: reducing parameter redundancy to increase efficiency and accuracy. Experiments on rotated MNIST, CIFAR10/100, and Galaxy10 show that the method outperforms the non-separable versions both in accuracy and speed.\n", "main_review": "*Strengths*\n\nS1: The paper is well-written, the method introduced makes sense and seems to work well in practice. \n\n*Weaknesses*\n\nW1: The submission has somewhat limited novelty. It can be seen as an application of depthwise convolutions to group-CNNs. Depthwise convolutions are quite popular since at least 2017 (Xception, Mobilenets); and the (non-separable) group-CNN utilized seems the same as Finzi et al [1], with sinusoidal activations. More importantly, the idea of separating the group-convolution along the group dimension has also appeared in Lengyel and van Gemert [2], although for a different flavor of G-CNN. \n\nW2: I see several issues with the experimental section. The most problematic is that all comparisons are against the paper's own baselines. While this makes for fair comparisons in the sense that the data pipeline, architectures and training schedule are all the same (although I have reservations about the number of parameters, see W3), I'm not sure if the conclusions hold in general. For example Weiler and Cesa [3] seems to show significantly better performance than the proposed method on rotated MNIST and CIFAR10/100; can the proposed method improve those results? If not, what would be the applications where the method is useful? Finzi et al [1] also show favorable results on a molecular property prediction, how does the submission fare in that task? When comparing with Finzi et al, I believe there should also be an ablation to disentangle the effects of the separable convolutions and the use of SIREN instead of a regular MLP.\n\nW3: Please report the number of parameters for each model trained. If I understand correctly, the number of channels per layer is kept constant, so the non-separable models have several times more parameters than the separable, is that correct? In that case could the lower performance on non-separable be explained by overfitting or slower convergence? I think both constant number of channels per layer or constant number of parameters are informative and should be reported (Weiler and Cesa [3] do something similar for CIFAR). \n\nW4: In Table 5, I believe a few experiments are missing and would be needed to disentangle the effects of the separation along the group dimension and the channel dimensions. I suggest to show, for each group, the performance when the convolutions are separable over the channel (depthwise) but not the group dimension. And for the baseline, it would be interesting to also see the performance for the separable depthwise version. \n\n*Questions*\n\nQ1: For the rotated MNIST, it is shown that approximating the convolution with random samples is superior, however the other MNIST variations seem to fall back to the discretization. Why is this the case? Do the random sampling approximation perform worse on the higher dimensional groups?\n\nQ2: The SIM(2) MNIST experiment is described as limited to 2, 4, 6, or 8 elements for each subgroup. Does this also refer to the way the dataset is constructed, or is it created with random rotations and scaling sampled from the continuous internal?\n\n*References*\n\n[1] Finzi et al, \"Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data\", ICML'20.\n\n[2] Lengyel and van Gemert, \"Exploiting Learned Symmetries in Group Equivariant Convolutions\", ICIP'21.\n\n[3] Weiler and Cesa, \"General E(2) - Equivariant Steerable CNNs\", NeurIPS'19.         ", "summary_of_the_review": "I believe the paper is not there yet, mainly because of limited novelty when compared to [1, 2] (W1), lack of comparison against the state-of-the-art of the tasks proposed (W2), and not very convincing experimental results due to lack of details (W3) and ablations (W4).\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635475961456}], "openreview_url": "https://openreview.net/forum?id=WnOLO1f50MH", "arxiv_id": "2110.13059", "paper_pdf": "papers/WnOLO1f50MH.pdf", "paper_pdf_sha256": "2d18a95128579de97dd96b3db9b84d9940890c57d68c73acd37cc4546124a4a5", "paper_pdf_bytes": 4150341, "paper_pdf_source": "openreview", "code_url": "https://github.com/david-knigge/separable-group-convolutional-networks", "code_repository": "david-knigge/separable-group-convolutional-networks", "code_commit": "e2de2bdb1c382b76bcaf3f24ad21fe35ac2674e8", "code_archive": "repos/WnOLO1f50MH.zip", "code_archive_sha256": "869d013c682133f25f787da1432854efafd17b7ad3af5332c84ce0f13402c048", "code_archive_bytes": 915204, "code_file_count": 71, "code_extensions": {".py": 69, ".ipynb": 2}, "github_disk_usage_kb": 890, "github_languages": {"Python": 252936}, "github_archived": false, "github_pushed_at": "2022-07-20T18:21:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/exploiting-redundancy-separable-group-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "m0ECRXO6QlP", "year": 2021, "status": "rejected", "title": "Supervision Accelerates Pre-training in Contrastive Semi-Supervised Learning of Visual Representations", "authors": ["Mido Assran", "Nicolas Ballas", "Lluis Castrejon", "Michael Rabbat"], "authorids": ["~Mido_Assran1", "~Nicolas_Ballas1", "~Lluis_Castrejon1", "~Michael_Rabbat1"], "authors_source": "OpenReview API", "abstract": "We investigate a strategy for improving the efficiency of contrastive learning of visual representations by leveraging a small amount of supervised information during pre-training. We propose a semi-supervised loss, SuNCEt, based on noise-contrastive estimation and neighbourhood component analysis, that aims to distinguish examples of different classes in addition to the self-supervised instance-wise pretext tasks. On ImageNet, we find that SuNCEt can be used to match the semi-supervised learning accuracy of previous contrastive approaches while using less than half the amount of pre-training and compute. Our main insight is that leveraging even a small amount of labeled data during pre-training, and not only during fine-tuning, provides an important signal that can significantly accelerate contrastive learning of visual representations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "kTY0UYywGsc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2064/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary**:\nThis paper designs a new loss, called SuNCTt, to speed up the convergence of semi-supervised training. Specifically, the loss involves the computation of similarity between anchor and other images with the same class, and the similarity between anchor and other labeled images. It is claimed to be considered as the form of neighborhood component analysis. Together with the standard contrastive learning loss, it only uses less than half the amount of pre-training and computes to match the accuracy of the previous approaches.\n\n**Pros**:\n+ The whole idea makes sense. The comprehensive experiments, also shown in appendix, support claims in the paper that the + introduced loss is helpful for semi-supervised training  \n+ Overall, the paper is well written and the results part is well structured.\n\n\n**Concerns**:\n- It mentions some semi-supervised work in the related work. Is it possible to compare the results with theirs?\n- It would be good to show the results of using cross-entropy pre-trained on ImageNet. In addition to the saving compute, I am also curious about the final performance on ImageNet after training for 1000(500) epochs. \n- On page 5, it mentioned that, with 1% labeled data, it is not significantly greater than the random accuracy. However, in appendix F (figure 7b), the result of 1% data(CIFAR-10) looks good.  I wonder why it is, if I understand the contexts correctly.\n- I also wonder, for the labeled data, what will happen if we use explored metric learning methods, like triplet loss. For the triplet loss, we can leverage labels to define the positive and negative samples. Both triplet loss and SuNCTt may do the similar thing, but in different forms.\n\nOverall, I prefer the rating as above the threshold at the current stage. Hope the authors could address my concerns or questions in the rebuttal period.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Involving the labeled samples in pre-training to speed up the contrastive semi-supervised learning", "review": "**Summary**:\nThis paper designs a new loss, called SuNCTt, to speed up the convergence of semi-supervised training. Specifically, the loss involves the computation of similarity between anchor and other images with the same class, and the similarity between anchor and other labeled images. It is claimed to be considered as the form of neighborhood component analysis. Together with the standard contrastive learning loss, it only uses less than half the amount of pre-training and computes to match the accuracy of the previous approaches.\n\n**Pros**:\n+ The whole idea makes sense. The comprehensive experiments, also shown in appendix, support claims in the paper that the + introduced loss is helpful for semi-supervised training  \n+ Overall, the paper is well written and the results part is well structured.\n\n\n**Concerns**:\n- It mentions some semi-supervised work in the related work. Is it possible to compare the results with theirs?\n- It would be good to show the results of using cross-entropy pre-trained on ImageNet. In addition to the saving compute, I am also curious about the final performance on ImageNet after training for 1000(500) epochs. \n- On page 5, it mentioned that, with 1% labeled data, it is not significantly greater than the random accuracy. However, in appendix F (figure 7b), the result of 1% data(CIFAR-10) looks good.  I wonder why it is, if I understand the contexts correctly.\n- I also wonder, for the labeled data, what will happen if we use explored metric learning methods, like triplet loss. For the triplet loss, we can leverage labels to define the positive and negative samples. Both triplet loss and SuNCTt may do the similar thing, but in different forms.\n\nOverall, I prefer the rating as above the threshold at the current stage. Hope the authors could address my concerns or questions in the rebuttal period.\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603889056588}, {"id": "4ua1vt1w_Ok", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2064/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new method applicable to a specific case in unsupervised learning: having access to a small amount of labels during the training phase where label signal is not used.\n\nFirst, I would like to say this is a well motivated task. Many recent works on self-supervised learning (image classification using instance discrimination signals) eventually need to use all the labels in the linear evaluation phase to obtain the final performance numbers. The authors also point out a similar view in the last paragraph in Section 3. \n\nIn the related works section, the work by Khosla et al. 2020 is mentioned. I wonder if it makes sense for the authors to add their numbers in the experiment comparison given that Khosla et al also use the original SimCLR as benchmark? In addition, I think a \"soft-nearest neighbor loss\" paper could be cited and compared (for example: Zhirong Wu, Alexei A Efros, and Stella Yu. Improving generalization via scalable neighbor- hood component analysis. In European Conference on Computer Vision (ECCV) 2018, 2018.)\nI understand that most supervised contrastive learning frameworks assume full label during the entire process, which is different from the setting in the paper. But it would be interesting (even critical) to compare the proposed method vs the other supervised contrastive learning methods when SuNCEt is given the full data.\n\nThe main contribution of the paper is the proposed SuNCEt loss, which is modified on top of the regular NCE loss in instance discrimination training. However, this “supervised constrastive learning“ only accelerates the SIMCLR learning process (by ~2x in terms of epochs from Table 1 and Table 2) and it does not significantly improve the accuracy over regular SIMCLR.\n\nOverall, I feel that the proposed loss is not a very novel idea (i.e., supervised contrastive learning) and the experiment results are not significantly better than prior arts.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "well motivated but needs more work and comparison", "review": "This paper proposes a new method applicable to a specific case in unsupervised learning: having access to a small amount of labels during the training phase where label signal is not used.\n\nFirst, I would like to say this is a well motivated task. Many recent works on self-supervised learning (image classification using instance discrimination signals) eventually need to use all the labels in the linear evaluation phase to obtain the final performance numbers. The authors also point out a similar view in the last paragraph in Section 3. \n\nIn the related works section, the work by Khosla et al. 2020 is mentioned. I wonder if it makes sense for the authors to add their numbers in the experiment comparison given that Khosla et al also use the original SimCLR as benchmark? In addition, I think a \"soft-nearest neighbor loss\" paper could be cited and compared (for example: Zhirong Wu, Alexei A Efros, and Stella Yu. Improving generalization via scalable neighbor- hood component analysis. In European Conference on Computer Vision (ECCV) 2018, 2018.)\nI understand that most supervised contrastive learning frameworks assume full label during the entire process, which is different from the setting in the paper. But it would be interesting (even critical) to compare the proposed method vs the other supervised contrastive learning methods when SuNCEt is given the full data.\n\nThe main contribution of the paper is the proposed SuNCEt loss, which is modified on top of the regular NCE loss in instance discrimination training. However, this “supervised constrastive learning“ only accelerates the SIMCLR learning process (by ~2x in terms of epochs from Table 1 and Table 2) and it does not significantly improve the accuracy over regular SIMCLR.\n\nOverall, I feel that the proposed loss is not a very novel idea (i.e., supervised contrastive learning) and the experiment results are not significantly better than prior arts.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603851037047}, {"id": "UackXCZunNf", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2064/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Pros:\n\n1. This submission is well written with lots of experiments. The claim is that supervised contrastive learning can speed up representation learning, which is well supported. \n\nCons:\n\n1. Instead of proposing a new idea, this paper adopts supervised contrastive learning and discovers that supervision can help accelerate the pretraining stage. Despite it maybe practical to speed up the experimental cycle, the technique contribution is rather limited. I may understands this paper wrong, so could authors clarify the propsoed SuNCEt loss with supervised contrastive loss? \n\n2. Since this paper is investigating semi-supervised learning, then some comparisons to existing literature is necessary. For example, comparisons to FixMatch. Right now, all the comparisons are only done on SimCLR, which are basically ablation studies.\n\n3. Performance improvement on CIFAR is marginal. I read the limitation section in Appendix, but not clear how it is related to the final results. \n\n4. I notice that the proposed SuNCEt loss is turned off after some training epochs. Especially when using 1% of the labeled data, this loss is turned off at epoch 30, which is quite early in the learn stage. Does SuNCEt loss hurt the training and why? And in practice, how do you determine the optimal time to turn off SuNCEt loss if you don't have access to all the labeled validation set? \n\nIn conclusion, despite the paper has some interesting results, it lacks of experiments to show its real contributions.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This submission discovers that supervised contrastive learning can speed up representation learning", "review": "Pros:\n\n1. This submission is well written with lots of experiments. The claim is that supervised contrastive learning can speed up representation learning, which is well supported. \n\nCons:\n\n1. Instead of proposing a new idea, this paper adopts supervised contrastive learning and discovers that supervision can help accelerate the pretraining stage. Despite it maybe practical to speed up the experimental cycle, the technique contribution is rather limited. I may understands this paper wrong, so could authors clarify the propsoed SuNCEt loss with supervised contrastive loss? \n\n2. Since this paper is investigating semi-supervised learning, then some comparisons to existing literature is necessary. For example, comparisons to FixMatch. Right now, all the comparisons are only done on SimCLR, which are basically ablation studies.\n\n3. Performance improvement on CIFAR is marginal. I read the limitation section in Appendix, but not clear how it is related to the final results. \n\n4. I notice that the proposed SuNCEt loss is turned off after some training epochs. Especially when using 1% of the labeled data, this loss is turned off at epoch 30, which is quite early in the learn stage. Does SuNCEt loss hurt the training and why? And in practice, how do you determine the optimal time to turn off SuNCEt loss if you don't have access to all the labeled validation set? \n\nIn conclusion, despite the paper has some interesting results, it lacks of experiments to show its real contributions.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603651957203}, {"id": "_7q3-QZelKJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2064/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper combines the self-supervised contrastive loss with the supervised contrastive loss for semi-supervised learning. By leveraging a small amount of labeled data, this paper shows that the semi-supervised contrastive loss can achieve similar performance as self-supervised contrastive loss (SimCLR) with less than half of the compute.\n\nStrength:\n1. This paper is well-motivated. It is a good direction to explore contrastive learning in the semi-supervised learning setting.\n2. It is interesting to see that using a small amount of labeled data can reduce computation.\n3. The paper is well-written and easy to understand.\n\nWeakness:\n1. The proposed semi-supervised contrastive loss seems to be a straight-forward combination of the self-supervised contrastive loss and the supervised contrastive loss [Khosla et al. 2020]. Therefore, the technical novelty is limited. \n2. This paper does not compare with any of the existing semi-supervised learning methods (e.g. FixMatch), even though they are discussed in the related work. Self-supervised learning has been known to be computation expensive, but this paper also needs to justify that the proposed method is more efficient than existing semi-supervised learning methods.\n3. Figure 2 shows that combining CE with SimCLR already achieves good performance. Therefore, a simple baseline would be combining FixMatch with SimCLR. I would expect it to be comparable or even better than the proposed contrastive learning method, because previous papers (e.g. S4L) have shown that adding a self-supervised objective helps with semi-supervised learning.\n4. The improvement over SimCLR does not seem to be significant under the same training epochs. Given that labeled data is used, this improvement is mostly expected.  Furthermore, the supervised contrastive loss is turned off after some epochs, which means it plays a less important role. If SuNCEt is also trained for 1000 epochs, would it converge to a similar result as SimCLR?\n5. In Table 2, it would be interesting to also see the performance of fully-supervised learning.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #3", "review": "Summary: This paper combines the self-supervised contrastive loss with the supervised contrastive loss for semi-supervised learning. By leveraging a small amount of labeled data, this paper shows that the semi-supervised contrastive loss can achieve similar performance as self-supervised contrastive loss (SimCLR) with less than half of the compute.\n\nStrength:\n1. This paper is well-motivated. It is a good direction to explore contrastive learning in the semi-supervised learning setting.\n2. It is interesting to see that using a small amount of labeled data can reduce computation.\n3. The paper is well-written and easy to understand.\n\nWeakness:\n1. The proposed semi-supervised contrastive loss seems to be a straight-forward combination of the self-supervised contrastive loss and the supervised contrastive loss [Khosla et al. 2020]. Therefore, the technical novelty is limited. \n2. This paper does not compare with any of the existing semi-supervised learning methods (e.g. FixMatch), even though they are discussed in the related work. Self-supervised learning has been known to be computation expensive, but this paper also needs to justify that the proposed method is more efficient than existing semi-supervised learning methods.\n3. Figure 2 shows that combining CE with SimCLR already achieves good performance. Therefore, a simple baseline would be combining FixMatch with SimCLR. I would expect it to be comparable or even better than the proposed contrastive learning method, because previous papers (e.g. S4L) have shown that adding a self-supervised objective helps with semi-supervised learning.\n4. The improvement over SimCLR does not seem to be significant under the same training epochs. Given that labeled data is used, this improvement is mostly expected.  Furthermore, the supervised contrastive loss is turned off after some epochs, which means it plays a less important role. If SuNCEt is also trained for 1000 epochs, would it converge to a similar result as SimCLR?\n5. In Table 2, it would be interesting to also see the performance of fully-supervised learning.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1602653780735}], "openreview_url": "https://openreview.net/forum?id=m0ECRXO6QlP", "arxiv_id": "2006.10803", "paper_pdf": "papers/m0ECRXO6QlP.pdf", "paper_pdf_sha256": "cd1343e72e6e4395d0ee7e3105d10365d94918251410fd9996da846a23532e41", "paper_pdf_bytes": 5101968, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/suncet", "code_repository": "facebookresearch/suncet", "code_commit": "731547d727b8c94d06c08a7848b4955de3a70cea", "code_archive": "repos/m0ECRXO6QlP.zip", "code_archive_sha256": "8baf35507e1864a2854a87ff854df59b0e351706535e6498e80b4894b5fab416", "code_archive_bytes": 993386, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 1130, "github_languages": {"Python": 148301}, "github_archived": true, "github_pushed_at": "2023-04-28T12:54:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recovering-petaflops-in-contrastive-semi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1gXR3NtwS", "year": 2020, "status": "rejected", "title": "Deep Bayesian Structure Networks", "authors": ["Zhijie Deng", "Yucen Luo", "Jun Zhu", "Bo Zhang"], "authorids": ["dzj17@mails.tsinghua.edu.cn", "luoyc15@mails.tsinghua.edu.cn", "dcszj@tsinghua.edu.cn", "dcszb@tsinghua.edu.cn"], "authors_source": "OpenReview API", "abstract": "Bayesian neural networks (BNNs) introduce uncertainty estimation to deep networks by performing Bayesian inference on network weights. However, such models bring the challenges of inference, and further BNNs with weight uncertainty rarely achieve superior performance to standard models. In this paper, we investigate a new line of Bayesian deep learning by performing Bayesian reasoning on the structure of deep neural networks. Drawing inspiration from the neural architecture search, we define the network structure as random weights on the redundant operations between computational nodes, and apply stochastic variational inference techniques to learn the structure distributions of networks. Empirically, the proposed method substantially surpasses the advanced deep neural networks across a range of classification and segmentation tasks. More importantly, our approach also preserves benefits of Bayesian principles, producing improved uncertainty estimation than the strong baselines including MC dropout and variational BNNs algorithms (e.g. noisy EK-FAC). ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H1ebOaz0FB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper257/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper combines ideas from neural architecture search (NAS) and Bayesian neural networks. Instead of maintaining uncertainty in network weights, the authors propose to retain uncertainty in the network structure. In particular, building on cell-based differentiable NAS, the authors infer a distribution over the gating weights of different cells incident onto a tensor while relying on point estimates for the weights inside each cell. \n\nOverall, I liked the paper and vote for accepting it. The notion of maintaining uncertainty about the network structure is a sensible one, and the paper explores an as yet under-explored area at the intersection of state-of-the-art network architecture search algorithms and Bayesian neural networks. Moreover, this is accompanied by compelling empirics — results demonstrate gains in both predictive performance and calibration across diverse tasks and careful comparisons to sensible baselines are presented to evaluate various aspects of the proposed approach (Table 1). \n\nDetailed Comments:\n+ One issue the experiments fail to adequately disentangle is the effect of weight uncertainty vs structure uncertainty. Are the observed gains in accuracy and calibration simply a product of better structure learning? In particular, I would love to see a baseline where point estimates of \\alpha are learned but posterior distribution over weights is inferred. I realize NEK-FAC was an attempt at providing such a comparison, but since it uses a different structure, it remains unclear whether it’s poor performance stems from the fundamental difficulty of learning posteriors over high dimensional weights or simply a sub-optimal network structure. \n\n+ In a similar spirit, one can imagine a fully Bayesian DBSN where one infers posterior distributions overbite \\alpha and w. Presumably, this would close the OOD entropy gap between random \\alpha and DBSN. \n\n+ How many Monte Carlo samples were used to evaluate Equation 8. In variational BNNs one often finds that using more MC samples doesn’t necessarily improve predictive accuracy over using the most likely sample (the mean if using a Gaussian variational family). It would be interesting to see predictive performance as a function of the number of MC samples for DBSN. \n\n+ Clarity: While I am mostly upbeat about this paper, the writing could be significantly improved. While the overall ideas come across, there are several instances where the text appears muddled and needs a few more polishing passes. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper combines ideas from neural architecture search (NAS) and Bayesian neural networks. Instead of maintaining uncertainty in network weights, the authors propose to retain uncertainty in the network structure. In particular, building on cell-based differentiable NAS, the authors infer a distribution over the gating weights of different cells incident onto a tensor while relying on point estimates for the weights inside each cell. \n\nOverall, I liked the paper and vote for accepting it. The notion of maintaining uncertainty about the network structure is a sensible one, and the paper explores an as yet under-explored area at the intersection of state-of-the-art network architecture search algorithms and Bayesian neural networks. Moreover, this is accompanied by compelling empirics — results demonstrate gains in both predictive performance and calibration across diverse tasks and careful comparisons to sensible baselines are presented to evaluate various aspects of the proposed approach (Table 1). \n\nDetailed Comments:\n+ One issue the experiments fail to adequately disentangle is the effect of weight uncertainty vs structure uncertainty. Are the observed gains in accuracy and calibration simply a product of better structure learning? In particular, I would love to see a baseline where point estimates of \\alpha are learned but posterior distribution over weights is inferred. I realize NEK-FAC was an attempt at providing such a comparison, but since it uses a different structure, it remains unclear whether it’s poor performance stems from the fundamental difficulty of learning posteriors over high dimensional weights or simply a sub-optimal network structure. \n\n+ In a similar spirit, one can imagine a fully Bayesian DBSN where one infers posterior distributions overbite \\alpha and w. Presumably, this would close the OOD entropy gap between random \\alpha and DBSN. \n\n+ How many Monte Carlo samples were used to evaluate Equation 8. In variational BNNs one often finds that using more MC samples doesn’t necessarily improve predictive accuracy over using the most likely sample (the mean if using a Gaussian variational family). It would be interesting to see predictive performance as a function of the number of MC samples for DBSN. \n\n+ Clarity: While I am mostly upbeat about this paper, the writing could be significantly improved. While the overall ideas come across, there are several instances where the text appears muddled and needs a few more polishing passes. "}, "tcdate": 1571855721092}, {"id": "rygXoHanYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper257/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to do approximate Bayesian inference in neural networks by treating the neural network structure as a random variable (RV), while inferring the parameters with point estimates. \nWhile performing Bayesian inference for the neural network structure is sensible, I am not convinced by the approach taken in this work. \n\nThe biggest problem is that the model uses a point estimate of the same weights for different, random network structures. \nMajor problems:\n- In the motivation the authors write “DBSN places distributions on the network structure, introducing more global randomness, thus is probable to yield more diverse predictions, and ensembling them brings more calibrated uncertainty”. What is “more global randomness”? This is used multiple times. Does it refer to the hierarchy in the graphical model? Please be precise here and point that out in your model by using an equation or graphical model. Or is it just an intuition? \n- Generally, I would agree that integrating out multiple network structures provides better calibrated uncertainty. However, given that the authors use point estimates for the weights, it is not clear if that is still true, especially since the number of different architectures used in practice is small. \n- What’s more, the approach uses *the same* point estimates for different structures. This leads to a graphical model, where the weights are not conditioned on the architecture/structure. This modeling choice could be a big limitation, because the weights now have to fit multiple different architectures; it may thus defeat the calibration completely. One can easily imagine that only a single random architecture works well with the learned point estimates, thus resulting in an (almost) deterministic model. I assume that this modeling choice was made for practical reasons, but could you expand on its implications / interpretation / limitations? Does the posterior of such a constrained model not quickly converge to an “almost” dirac, effectively just one network structure? \n- Sec. 3.1. presents the above problem resulting from a modeling decision as a “training challenge”. To counter this problem, the authors propose to reduce the variance of the structure distribution. By doing so, the approach becomes even less Bayesian and the predictive uncertainty becomes even less reliable. \n- “We only learn the connections between the B internal nodes, as shown in Appendix D”. All deterministic weights are learned, but the structure only for some parts of the model? If this is the case, then approach becomes again less probabilistic.\n\nRegarding the experiments, the stddevs are calculated from 3(!) independent runs and thus completely misleading (imagine the stddev of the stddev estimate). \n\nIn summary, the model choice of point estimates for the weights, which are not conditioned on the architecture, leads to various problems. The authors have to introduce tricks such as reducing the variance of the random network structures or learning only a part of the whole structure to make the approach converge. The resulting probabilistic model and its predictive uncertainty is questionable. For this reason, this paper should be rejected. \n\nMinor problems\n- Sec. 3.2. “Improvements of the structure learning space”. What is a “structure learning space”? \n- Section 3 introduces the ELBO in Eq. (4) before the complete model is specified. Please specify the whole model first. How do w and alpha depend on each other in your model? \n- The prior for the weights is omitted; at the same time it is mentioned in the experiments (Sec.5.1.) that weight decay is applied. Why not just be explicit about it and say that a Gaussian prior is used?\n- Background Sec. 2.2. is not clear.  what is a cell? some deterministic transformation in general? bunch of neural network layers? What are the operation (last term in Eq. (2)) doing? This is not detailed and abstract to me. Are the alphas probabilities? Is Eq. (2) consequently a mixture model of different architectures? Or is this here just a weighted sum, where the weights take arbitrary values? A small visualization (additionally) might help here, but can probably be rectified by better explanation.\n- Bayesian reasoning on the structure. Inference?\n- Writing that you propose a new “framework” is a bit grandiose for what is actually proposed. There has been previous work in which the architecture is inferred as well and these approaches would certainly be part of the same framework. Please just say model/algorithm/approach, whatever is applicable. \n- new paragraph starting at “To empirically validate” in the intro.\n- Before (4): “Then we rewrite the approximation error”. Eq. (4) is the ELBO, this is not an approximation error. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes to do approximate Bayesian inference in neural networks by treating the neural network structure as a random variable (RV), while inferring the parameters with point estimates. \nWhile performing Bayesian inference for the neural network structure is sensible, I am not convinced by the approach taken in this work. \n\nThe biggest problem is that the model uses a point estimate of the same weights for different, random network structures. \nMajor problems:\n- In the motivation the authors write “DBSN places distributions on the network structure, introducing more global randomness, thus is probable to yield more diverse predictions, and ensembling them brings more calibrated uncertainty”. What is “more global randomness”? This is used multiple times. Does it refer to the hierarchy in the graphical model? Please be precise here and point that out in your model by using an equation or graphical model. Or is it just an intuition? \n- Generally, I would agree that integrating out multiple network structures provides better calibrated uncertainty. However, given that the authors use point estimates for the weights, it is not clear if that is still true, especially since the number of different architectures used in practice is small. \n- What’s more, the approach uses *the same* point estimates for different structures. This leads to a graphical model, where the weights are not conditioned on the architecture/structure. This modeling choice could be a big limitation, because the weights now have to fit multiple different architectures; it may thus defeat the calibration completely. One can easily imagine that only a single random architecture works well with the learned point estimates, thus resulting in an (almost) deterministic model. I assume that this modeling choice was made for practical reasons, but could you expand on its implications / interpretation / limitations? Does the posterior of such a constrained model not quickly converge to an “almost” dirac, effectively just one network structure? \n- Sec. 3.1. presents the above problem resulting from a modeling decision as a “training challenge”. To counter this problem, the authors propose to reduce the variance of the structure distribution. By doing so, the approach becomes even less Bayesian and the predictive uncertainty becomes even less reliable. \n- “We only learn the connections between the B internal nodes, as shown in Appendix D”. All deterministic weights are learned, but the structure only for some parts of the model? If this is the case, then approach becomes again less probabilistic.\n\nRegarding the experiments, the stddevs are calculated from 3(!) independent runs and thus completely misleading (imagine the stddev of the stddev estimate). \n\nIn summary, the model choice of point estimates for the weights, which are not conditioned on the architecture, leads to various problems. The authors have to introduce tricks such as reducing the variance of the random network structures or learning only a part of the whole structure to make the approach converge. The resulting probabilistic model and its predictive uncertainty is questionable. For this reason, this paper should be rejected. \n\nMinor problems\n- Sec. 3.2. “Improvements of the structure learning space”. What is a “structure learning space”? \n- Section 3 introduces the ELBO in Eq. (4) before the complete model is specified. Please specify the whole model first. How do w and alpha depend on each other in your model? \n- The prior for the weights is omitted; at the same time it is mentioned in the experiments (Sec.5.1.) that weight decay is applied. Why not just be explicit about it and say that a Gaussian prior is used?\n- Background Sec. 2.2. is not clear.  what is a cell? some deterministic transformation in general? bunch of neural network layers? What are the operation (last term in Eq. (2)) doing? This is not detailed and abstract to me. Are the alphas probabilities? Is Eq. (2) consequently a mixture model of different architectures? Or is this here just a weighted sum, where the weights take arbitrary values? A small visualization (additionally) might help here, but can probably be rectified by better explanation.\n- Bayesian reasoning on the structure. Inference?\n- Writing that you propose a new “framework” is a bit grandiose for what is actually proposed. There has been previous work in which the architecture is inferred as well and these approaches would certainly be part of the same framework. Please just say model/algorithm/approach, whatever is applicable. \n- new paragraph starting at “To empirically validate” in the intro.\n- Before (4): “Then we rewrite the approximation error”. Eq. (4) is the ELBO, this is not an approximation error. \n"}, "tcdate": 1571767706761}, {"id": "rygDVQg2FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper257/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed deep Bayesian structure networks (DBSN) to model weights, \\alpha, of the redundant operations in cell-based differentiable NAS. The authors claim that DBSN can achieve better performance (accuracy) than the state of the art. \n\nOne of my concerns is the Bayesian formulation introduced in Eq. (4) seems problematic. It is not clear what priors are placed on alpha. In the case of Bayes by BP (BBB), which is cited as Blundell et al. 2015 in the paper, a Gaussian prior (with zero mean) is used. Therefore there is a KL term between the variational distribution q(w) and the prior distribution p(w) to regularize q(w). In DBSN, q(\\alpha) is parameterized by \\theta and \\epsilon, and so is p(\\alpha), meaning that the KL term is effectively zero. This is very different from what is done in BBB.\n\nThe second major concern is on the experiments. (1) The authors use DARTS as a main baseline and show that DBSN significantly outperforms DARTS. However, looking at the DARTS paper, the test error on CIFAR-10 is around 3% for both the first-order and second-order versions. The test error in Table 1 is around 9%, which is a lot lower. I notice that the DARTS paper has a parameter number of 3.3M, while in the current paper it set to 1M. Given that DARTS is the main baseline method and the same dataset (CIFAR-10) is used, it would make much more sense to use exactly the same architecture for comparison. The current results is hardly convincing. (2) Besides, note that in the DARTS paper, DenseNet-BC has test error of 3.46%, much higher than DARTS (~3%). In Table 2 of this paper however, DARTS is significantly worse than DenseNet-BC (8.91% versus 4.51%). These results are highly inconsistent with previous work.\n\nAs mentioned in the paper, Dikov & Bayer 2019 has a very similar idea to perform NAS from a Bayesian perspective. It would be best (and would definitely make the paper stronger) to include some comparison. Even if Dikov & Bayer 2019 is not very scalable, it is at least possible to compare them in smaller network size. Otherwise it is hard to evaluate the contribution of DBSN given this highly similar work.\n\nThe authors mentioned in the introduction that DBSN ‘yields more diverse prediction’ and therefore brings more calibrated uncertainty comparing to ensembling different architectures. This is not verified in the experiment section. Table 3 only reports the ECE for one instance of trained networks. For example, it would be interesting to sample different architecture from the alpha learned in DARTS and DBSN, train several networks, ensemble them, and use the variance of the ensemble to compute ECE. This would verify the claim mentioned above.\n\nDo you retrain the network from scratch after the architecture search (which is done in DARTS) for DARTS and DBSN?\n\nI am not convinced by the claim that BNN usually achieve compromising performance. Essentially, BNN, if trained well, is a generalization of deterministic NN. If very flat priors and highly confident variational distributions are used, BNN essentially reduces to deterministic NN.\n\nMissing references on Bayesian deep learning and BNN:\n\nBayesian Dark Knowledge\nTowards Bayesian Deep Learning: A Survey\nNatural-Parameter Networks: A Class of Probabilistic Neural Networks", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This paper proposed deep Bayesian structure networks (DBSN) to model weights, \\alpha, of the redundant operations in cell-based differentiable NAS. The authors claim that DBSN can achieve better performance (accuracy) than the state of the art. \n\nOne of my concerns is the Bayesian formulation introduced in Eq. (4) seems problematic. It is not clear what priors are placed on alpha. In the case of Bayes by BP (BBB), which is cited as Blundell et al. 2015 in the paper, a Gaussian prior (with zero mean) is used. Therefore there is a KL term between the variational distribution q(w) and the prior distribution p(w) to regularize q(w). In DBSN, q(\\alpha) is parameterized by \\theta and \\epsilon, and so is p(\\alpha), meaning that the KL term is effectively zero. This is very different from what is done in BBB.\n\nThe second major concern is on the experiments. (1) The authors use DARTS as a main baseline and show that DBSN significantly outperforms DARTS. However, looking at the DARTS paper, the test error on CIFAR-10 is around 3% for both the first-order and second-order versions. The test error in Table 1 is around 9%, which is a lot lower. I notice that the DARTS paper has a parameter number of 3.3M, while in the current paper it set to 1M. Given that DARTS is the main baseline method and the same dataset (CIFAR-10) is used, it would make much more sense to use exactly the same architecture for comparison. The current results is hardly convincing. (2) Besides, note that in the DARTS paper, DenseNet-BC has test error of 3.46%, much higher than DARTS (~3%). In Table 2 of this paper however, DARTS is significantly worse than DenseNet-BC (8.91% versus 4.51%). These results are highly inconsistent with previous work.\n\nAs mentioned in the paper, Dikov & Bayer 2019 has a very similar idea to perform NAS from a Bayesian perspective. It would be best (and would definitely make the paper stronger) to include some comparison. Even if Dikov & Bayer 2019 is not very scalable, it is at least possible to compare them in smaller network size. Otherwise it is hard to evaluate the contribution of DBSN given this highly similar work.\n\nThe authors mentioned in the introduction that DBSN ‘yields more diverse prediction’ and therefore brings more calibrated uncertainty comparing to ensembling different architectures. This is not verified in the experiment section. Table 3 only reports the ECE for one instance of trained networks. For example, it would be interesting to sample different architecture from the alpha learned in DARTS and DBSN, train several networks, ensemble them, and use the variance of the ensemble to compute ECE. This would verify the claim mentioned above.\n\nDo you retrain the network from scratch after the architecture search (which is done in DARTS) for DARTS and DBSN?\n\nI am not convinced by the claim that BNN usually achieve compromising performance. Essentially, BNN, if trained well, is a generalization of deterministic NN. If very flat priors and highly confident variational distributions are used, BNN essentially reduces to deterministic NN.\n\nMissing references on Bayesian deep learning and BNN:\n\nBayesian Dark Knowledge\nTowards Bayesian Deep Learning: A Survey\nNatural-Parameter Networks: A Class of Probabilistic Neural Networks"}, "tcdate": 1571713839307}], "openreview_url": "https://openreview.net/forum?id=B1gXR3NtwS", "arxiv_id": null, "paper_pdf": "papers/B1gXR3NtwS.pdf", "paper_pdf_sha256": "ab15cbdaf85181ef806aaebc9f7d61cc4d3bbbcd55349164de9114499965db72", "paper_pdf_bytes": 1276567, "paper_pdf_source": "openreview", "code_url": "https://github.com/anonymousest/DBSN", "code_repository": "anonymousest/DBSN", "code_commit": "79f8cac7f3794e7bddda1902e14d365d58f50f11", "code_archive": "repos/B1gXR3NtwS.zip", "code_archive_sha256": "a8fa61537adf75c786973a3906d89bdeac63b6eb346d877009e03b80945f7a38", "code_archive_bytes": 3240763, "code_file_count": 59, "code_extensions": {".py": 59}, "github_disk_usage_kb": 3107, "github_languages": {"Python": 455485}, "github_archived": false, "github_pushed_at": "2019-11-13T02:54:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-bayesian-structure-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4TFfiG17ec", "year": 2026, "status": "rejected", "title": "Thanos: A Block-wise Pruning Algorithm for Efficient Large Language Model Compression", "authors": ["Ivan Ilin", "Peter Richtárik"], "authorids": ["~Ivan_Ilin1", "~Peter_Richtárik1"], "authors_source": "OpenReview API", "abstract": "This paper presents Thanos, a novel weight-pruning algorithm designed to reduce the memory footprint and enhance the computational efficiency of large language models (LLMs) by removing redundant weights while maintaining accuracy. Thanos introduces a block-wise pruning strategy with adaptive masks that dynamically adjust to weight importance, enabling flexible sparsity patterns and structured formats, such as n:m sparsity, optimized for hardware acceleration. Experimental evaluations demonstrate that Thanos achieves state-of-the-art performance in structured pruning and outperforms existing methods in unstructured pruning. By providing an efficient and adaptable approach to model compression, Thanos offers a practical solution for deploying large models in resource-constrained environments. The algorithm is publicly available for further research and application.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "VwpnrkkGzi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission25458/Reviewer_fuCo"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes Thanos, a block-wise post-training pruning algorithm for large language models. Unlike prior methods such as SparseGPT and Wanda that prune weights independently, Thanos removes multiple weights jointly within each block, leveraging local Hessian information to better preserve model performance. It introduces adaptive masking based on weight–input importance and supports unstructured, structured, and semi-structured sparsity, including an outlier row preservation mechanism for stability at high sparsity. Experiments on OPT and LLaMA series show that Thanos achieves state-of-the-art perplexity and zero-shot accuracy under both unstructured and structured pruning, while maintaining computational efficiency suitable for hardware acceleration.", "review_text": "This paper proposes Thanos, a block-wise post-training pruning algorithm for large language models. Unlike prior methods such as SparseGPT and Wanda that prune weights independently, Thanos removes multiple weights jointly within each block, leveraging local Hessian information to better preserve model performance. It introduces adaptive masking based on weight–input importance and supports unstructured, structured, and semi-structured sparsity, including an outlier row preservation mechanism for stability at high sparsity. Experiments on OPT and LLaMA series show that Thanos achieves state-of-the-art perplexity and zero-shot accuracy under both unstructured and structured pruning, while maintaining computational efficiency suitable for hardware acceleration.", "strengths": "The paper is generally well-organized. The proposed block-wise pruning approach is thoughtfully designed, offering a reasonable trade-off between pruning accuracy and computational efficiency. The inclusion of outlier-row preservation and support for structured and semi-structured sparsity makes the method adaptable to practical deployment. The experimental section is fairly comprehensive, and the presentation of algorithms and results is clear and easy to follow.", "weaknesses": "The evaluation is somewhat narrow, focusing mainly on OPT and LLaMA, leaving uncertainty about generalization to newer models. The paper lacks detailed ablation or sensitivity studies on key hyperparameters such as block size and α. Reported efficiency gains are not well supported by runtime or memory analyses.", "questions": "1. Could the authors evaluate Thanos on more recent or diverse model families (e.g., Qwen3) to better demonstrate its generalization across architectures?\n2. Can the authors provide more detailed ablation or sensitivity analyses on key hyperparameters such as block size (B) and outlier ratio (α), especially across different sparsity levels or model scales?\n3. The paper claims improved efficiency. Could the authors include quantitative comparisons (e.g., runtime or throughput) to substantiate these efficiency gains?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Thanos, a block-wise post-training pruning algorithm for large language models. Unlike prior methods such as SparseGPT and Wanda that prune weights independently, Thanos removes multiple weights jointly within each block, leveraging local Hessian information to better preserve model performance. It introduces adaptive masking based on weight–input importance and supports unstructured, structured, and semi-structured sparsity, including an outlier row preservation mechanism for stability at high sparsity. Experiments on OPT and LLaMA series show that Thanos achieves state-of-the-art perplexity and zero-shot accuracy under both unstructured and structured pruning, while maintaining computational efficiency suitable for hardware acceleration.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper is generally well-organized. The proposed block-wise pruning approach is thoughtfully designed, offering a reasonable trade-off between pruning accuracy and computational efficiency. The inclusion of outlier-row preservation and support for structured and semi-structured sparsity makes the method adaptable to practical deployment. The experimental section is fairly comprehensive, and the presentation of algorithms and results is clear and easy to follow.", "weaknesses": "The evaluation is somewhat narrow, focusing mainly on OPT and LLaMA, leaving uncertainty about generalization to newer models. The paper lacks detailed ablation or sensitivity studies on key hyperparameters such as block size and α. Reported efficiency gains are not well supported by runtime or memory analyses.", "questions": "1. Could the authors evaluate Thanos on more recent or diverse model families (e.g., Qwen3) to better demonstrate its generalization across architectures?\n2. Can the authors provide more detailed ablation or sensitivity analyses on key hyperparameters such as block size (B) and outlier ratio (α), especially across different sparsity levels or model scales?\n3. The paper claims improved efficiency. Could the authors include quantitative comparisons (e.g., runtime or throughput) to substantiate these efficiency gains?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761931799883}, {"id": "nJNdXFkS8j", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission25458/Reviewer_qK9D"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper focuses on the pruning of the transformer block to remove the redundancy but preserve the accuracy. Authors focuses on the post-training pruning and introduces a method named Thanos. Thanos dynamically constructs the global residual mask based on the number of elements already pruned and the final target. It also take a hyperparam alpha to take the outlier row into consideration for preserving the useful outlier row. Authors have conducted extensive experiemnts on multiple matrices, including perplexity analysis, zero-shot performance and also ablation with different block size, where the proposed Thanos show state-of-the-art performance for all of these metrics.", "review_text": "This paper focuses on the pruning of the transformer block to remove the redundancy but preserve the accuracy. Authors focuses on the post-training pruning and introduces a method named Thanos. Thanos dynamically constructs the global residual mask based on the number of elements already pruned and the final target. It also take a hyperparam alpha to take the outlier row into consideration for preserving the useful outlier row. Authors have conducted extensive experiemnts on multiple matrices, including perplexity analysis, zero-shot performance and also ablation with different block size, where the proposed Thanos show state-of-the-art performance for all of these metrics.", "strengths": "+ The proposed model is well described and show very promising results compared with other existing state-of-the-art methods.\n\n+ Authors also provide extensive analysis for the different datasets in supplementary material, which is showing the proposed method showing a very promising results.", "weaknesses": "- If authors are able to provide some analysis in addition to Llama and OPT, it would be great to show the overall generalizability of the proposed model. \n\n- Please consider include part of the results in section E to the main manuscript as these are very important numbers.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the pruning of the transformer block to remove the redundancy but preserve the accuracy. Authors focuses on the post-training pruning and introduces a method named Thanos. Thanos dynamically constructs the global residual mask based on the number of elements already pruned and the final target. It also take a hyperparam alpha to take the outlier row into consideration for preserving the useful outlier row. Authors have conducted extensive experiemnts on multiple matrices, including perplexity analysis, zero-shot performance and also ablation with different block size, where the proposed Thanos show state-of-the-art performance for all of these metrics.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "+ The proposed model is well described and show very promising results compared with other existing state-of-the-art methods.\n\n+ Authors also provide extensive analysis for the different datasets in supplementary material, which is showing the proposed method showing a very promising results.", "weaknesses": "- If authors are able to provide some analysis in addition to Llama and OPT, it would be great to show the overall generalizability of the proposed model. \n\n- Please consider include part of the results in section E to the main manuscript as these are very important numbers.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761607690522}, {"id": "jEic4sHdMC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission25458/Reviewer_25Jh"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper introduces Thanos, a post-training pruning algorithm aimed at shrinking large language models (LLMs) without retraining. Thanos works by partitioning every linear layer into manageable column-wise blocks and, for each block, jointly selecting multiple weights to drop and analytically re-optimising the surviving weights. This is achieved by adaptive pruning mask strategy and update multiple weights simultaneously. Moreover, they devise a structured variant that permutes rows and columns so that whole columns can be excised while optionally preserving a user-defined fraction of \"outlier\" rows.", "review_text": "The paper introduces Thanos, a post-training pruning algorithm aimed at shrinking large language models (LLMs) without retraining. Thanos works by partitioning every linear layer into manageable column-wise blocks and, for each block, jointly selecting multiple weights to drop and analytically re-optimising the surviving weights. This is achieved by adaptive pruning mask strategy and update multiple weights simultaneously. Moreover, they devise a structured variant that permutes rows and columns so that whole columns can be excised while optionally preserving a user-defined fraction of \"outlier\" rows.", "strengths": "1. The proposed Thanos algorithm introduces a block-wise pruning algorithm with joint weight updates. This improves upon prior work (SparseGPT) that only prunes one weight per row at a time. The paper also offers well-grounded theoretical derivations and practical heuristics, such as adaptive mask updates and outlier rows compatibility method, retaining model generalization capabilities.\n2. The Thanos supports unstructured, structured, and semi-structured (n:m) sparsity patterns. The adaptability to formats like 2:4 allows it to take advantage of hardware acceleration.\n3. Extensive empirical evaluations across different models show that Thanos outperforms existing methods (Magnitude, Wanda, SparseGPT) on both perplexity and zero-shot tasks.", "weaknesses": "1. While the paper explores the effect of mask strategy and joint weight updates, it lacks a thorough ablation study to isolate the individual contributions of different algorithmic components. Such analysis would strengthen the empirical claims.\n2. The strategy of retaining outlier rows may hinder strict structural pruning. For instance, it becomes unclear whether entire columns can always be pruned, which could compromise compatibility with hardware acceleration schemes relying on full-column removal. The paper does not sufficiently clarify how Thanos maintains structured sparsity under these conditions.\n3. The evaluation primarily focuses on perplexity and zero-shot accuracy, but omits key practical metrics such as pruning time and inference-time memory usage. These factors are crucial when choosing among pruning methods for deployment. Including such comparisons would have significantly enhanced the practical value of the results.", "questions": "Please refer to the \"Weakness\" section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Thanos, a post-training pruning algorithm aimed at shrinking large language models (LLMs) without retraining. Thanos works by partitioning every linear layer into manageable column-wise blocks and, for each block, jointly selecting multiple weights to drop and analytically re-optimising the surviving weights. This is achieved by adaptive pruning mask strategy and update multiple weights simultaneously. Moreover, they devise a structured variant that permutes rows and columns so that whole columns can be excised while optionally preserving a user-defined fraction of \"outlier\" rows.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The proposed Thanos algorithm introduces a block-wise pruning algorithm with joint weight updates. This improves upon prior work (SparseGPT) that only prunes one weight per row at a time. The paper also offers well-grounded theoretical derivations and practical heuristics, such as adaptive mask updates and outlier rows compatibility method, retaining model generalization capabilities.\n2. The Thanos supports unstructured, structured, and semi-structured (n:m) sparsity patterns. The adaptability to formats like 2:4 allows it to take advantage of hardware acceleration.\n3. Extensive empirical evaluations across different models show that Thanos outperforms existing methods (Magnitude, Wanda, SparseGPT) on both perplexity and zero-shot tasks.", "weaknesses": "1. While the paper explores the effect of mask strategy and joint weight updates, it lacks a thorough ablation study to isolate the individual contributions of different algorithmic components. Such analysis would strengthen the empirical claims.\n2. The strategy of retaining outlier rows may hinder strict structural pruning. For instance, it becomes unclear whether entire columns can always be pruned, which could compromise compatibility with hardware acceleration schemes relying on full-column removal. The paper does not sufficiently clarify how Thanos maintains structured sparsity under these conditions.\n3. The evaluation primarily focuses on perplexity and zero-shot accuracy, but omits key practical metrics such as pruning time and inference-time memory usage. These factors are crucial when choosing among pruning methods for deployment. Including such comparisons would have significantly enhanced the practical value of the results.", "questions": "Please refer to the \"Weakness\" section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761155896464}, {"id": "w0aV7kmawu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission25458/Reviewer_kKnK"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 3, "summary": "This paper proposes Thanos, a block-wise pruning algorithm for large language models. It introduces adaptive masks and joint weight updates to better capture the impact of pruning multiple parameters, supporting unstructured, structured, and n:m sparsity patterns. Experiments on LLaMA and OPT show that Thanos outperforms baselines like SparseGPT, Wanda, and magnitude pruning in both perplexity and zero-shot accuracy, especially under structured sparsity.", "review_text": "This paper proposes Thanos, a block-wise pruning algorithm for large language models. It introduces adaptive masks and joint weight updates to better capture the impact of pruning multiple parameters, supporting unstructured, structured, and n:m sparsity patterns. Experiments on LLaMA and OPT show that Thanos outperforms baselines like SparseGPT, Wanda, and magnitude pruning in both perplexity and zero-shot accuracy, especially under structured sparsity.", "strengths": "The paper addresses an important problem in LLM compression and proposes a pruning method that is relatively straightforward to implement. It covers both unstructured and structured sparsity settings, including n:m formats, which makes the method more practically relevant. The experimental section evaluates Thanos across multiple model scales and tasks, ensuring a fair comparison with widely used baselines.", "weaknesses": "1. The paper is poorly organized; for example, the conclusion is placed in the appendix instead of the main body, and the overall structure lacks clarity.\n2. A large portion of the paper is devoted to re-describing prior methods, which seems unnecessary and distracts from the core contribution.\n3. The main novelty—an extension of SparseGPT with multi-element joint pruning updates—appears incremental and insufficient to support a full paper.\n4. The experimental evaluation is incomplete: it does not report model size reduction or inference speedup, and the improvements over Wanda are marginal.\n5. There is no ablation study to disentangle the contribution of each proposed component.\n6. The choice of models in Table 1 is unclear: the paper refers to “LLaMA-2-1.1B” and “LLaMA-3-1B,” which do not correspond to official Meta releases, leaving ambiguity about what models are actually used.", "questions": "Please refer to the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Thanos, a block-wise pruning algorithm for large language models. It introduces adaptive masks and joint weight updates to better capture the impact of pruning multiple parameters, supporting unstructured, structured, and n:m sparsity patterns. Experiments on LLaMA and OPT show that Thanos outperforms baselines like SparseGPT, Wanda, and magnitude pruning in both perplexity and zero-shot accuracy, especially under structured sparsity.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "The paper addresses an important problem in LLM compression and proposes a pruning method that is relatively straightforward to implement. It covers both unstructured and structured sparsity settings, including n:m formats, which makes the method more practically relevant. The experimental section evaluates Thanos across multiple model scales and tasks, ensuring a fair comparison with widely used baselines.", "weaknesses": "1. The paper is poorly organized; for example, the conclusion is placed in the appendix instead of the main body, and the overall structure lacks clarity.\n2. A large portion of the paper is devoted to re-describing prior methods, which seems unnecessary and distracts from the core contribution.\n3. The main novelty—an extension of SparseGPT with multi-element joint pruning updates—appears incremental and insufficient to support a full paper.\n4. The experimental evaluation is incomplete: it does not report model size reduction or inference speedup, and the improvements over Wanda are marginal.\n5. There is no ablation study to disentangle the contribution of each proposed component.\n6. The choice of models in Table 1 is unclear: the paper refers to “LLaMA-2-1.1B” and “LLaMA-3-1B,” which do not correspond to official Meta releases, leaving ambiguity about what models are actually used.", "questions": "Please refer to the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760710231142}], "openreview_url": "https://openreview.net/forum?id=4TFfiG17ec", "arxiv_id": "2504.05346", "paper_pdf": "papers/4TFfiG17ec.pdf", "paper_pdf_sha256": "b4f2dc64593f347c399837d1ea68d226bd715ac7b80c14429e311c0a5f44b770", "paper_pdf_bytes": 4680307, "paper_pdf_source": "openreview", "code_url": "https://github.com/vectozavr/thanos", "code_repository": "vectozavr/thanos", "code_commit": "ee4910891ecf071719e2a4406fbdfe53488a5e61", "code_archive": "repos/4TFfiG17ec.zip", "code_archive_sha256": "5461d179ca168ecdf29c27316877d17274b13df39bbe0f5c02082241559e56f9", "code_archive_bytes": 45851, "code_file_count": 17, "code_extensions": {".py": 16, ".sh": 1}, "github_disk_usage_kb": 3195, "github_languages": {"Python": 132502, "Shell": 664}, "github_archived": false, "github_pushed_at": "2026-08-06T17:21:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/thanos-a-block-wise-pruning-algorithm-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VHGZjZmzsO", "year": 2025, "status": "rejected", "title": "Memory-Enhanced Neural Solvers for Efficient Adaptation in Combinatorial Optimization", "authors": ["Felix Chalumeau", "Refiloe Shabe", "Noah De Nicola", "Arnu Pretorius", "Thomas D Barrett", "Nathan Grinsztajn"], "authorids": ["~Felix_Chalumeau1", "~Refiloe_Shabe1", "~Noah_De_Nicola1", "~Arnu_Pretorius1", "~Thomas_D_Barrett1", "~Nathan_Grinsztajn1"], "authors_source": "OpenReview API", "abstract": "Combinatorial Optimization is crucial to numerous real-world applications, yet still presents challenges due to its (NP-)hard nature. Amongst existing approaches, heuristics often offer the best trade-off between quality and scalability, making them suitable for industrial use. While Reinforcement Learning (RL) offers a flexible framework for designing heuristics, its adoption over handcrafted heuristics remains incomplete within industrial solvers. Existing learned methods still lack the ability to adapt to specific instances and fully leverage the available computational budget. The current best methods either rely on a collection of pre-trained policies, or on data-inefficient fine-tuning; hence failing to fully utilize newly available information within the constraints of the budget. In response, we present MEMENTO, an approach that leverages memory to improve the adaptation of neural solvers at inference time. MEMENTO enables updating the action distribution dynamically based on the outcome of previous decisions. We validate its effectiveness on benchmark problems, in particular Traveling Salesman and Capacitated Vehicle Routing, demonstrating its superiority over tree-search and policy-gradient fine-tuning; and showing it can be zero-shot combined with diversity-based solvers. We successfully train all RL auto-regressive solvers on large instances, and show that MEMENTO can scale and is data-efficient. Overall, MEMENTO enables to push the state-of-the-art on 11 out of 12 evaluated tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qhkB1W1EuP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6399/Reviewer_YhmS"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes a method (MEMENTO) for online fine-tuning of neural CO models. More concretely, MEMENTO learns the rules for updating the policy parameters at inference time. This is achieved by leveraging a learned residual policy for the base policy, where this residual policy utilizes the past solution predictions to explore the search space. Through experiments on the TSP and the CVRP, the paper demonstrates MEMENTO's efficacy against baseline methods.", "review_text": "The paper proposes a method (MEMENTO) for online fine-tuning of neural CO models. More concretely, MEMENTO learns the rules for updating the policy parameters at inference time. This is achieved by leveraging a learned residual policy for the base policy, where this residual policy utilizes the past solution predictions to explore the search space. Through experiments on the TSP and the CVRP, the paper demonstrates MEMENTO's efficacy against baseline methods.", "strengths": "- The paper conveys most of the ideas clearly.\n- The idea of leveraging a learned residual policy for online fine-tuning of the base policy at inference time is interesting and novel.", "weaknesses": "- MEMENTO prevents the retrieval step from being too costly by only collecting data from the same node we are currently in. Could the authors elaborate on why this is a good retrieval strategy? Information should somehow be retrieved based on the partial solution constructed, since node-level decisions could be vastly different depending on the overall solutions.\n- The experiments do not use augmentation with symmetries in the experiments, which they claim to be not critical by citing just one prior work (COMPASS). Overall, to claim that MEMENTO outperforms EAS, the authors should compare it against EAS with augmentation enabled since EAS has been shown to work better that way. \n- My main concern with this paper is regarding the runtime of different methods in the experiments (which aren't reported for the major experiments in the main paper and put in the appendices instead). In Appendix A.1, the authors clarify that they report the runtime to solve one instance rather than the entire dataset. Given this, the comparison is fair only if they give each algorithm the same amount of runtime. However, Table 2 in the appendices shows that MEMENTO takes more than 2x the runtime than COMPASS (more than 4x for CVPR-100). The runtime is significantly higher than that for EAS as well (more than 2x) for CVRP-100.\n- The paper makes a few incorrect statements or claims:\n         1. In the definition of $\\pi^\\star$ on Page 3, if we are taking a max over $i$, why does $i$ appear in the outer expectation?\n         2. The claim that MEMENTO at least learns the REINFORCE update in the worst case is not exactly correct. In the worst case, the residual policy could output random values. I think the authors wanted to claim that the residual policy has the ability to at least learn the REINFORCE update rule (if nothing better). \n        3. The paper claims that MEMENTO  is \"designed\" to be agnostic to the base policy. While the authors demonstrate empirically that the learned residual policy for POMO could be used with COMPASS, the design of the framework as such makes the residual policy very much dependent on the base policy since they're learned jointly.", "questions": "1. Could the authors suggest alternative strategies for retrieval that do consider the partial solution constructed?\n2. Could the authors report the results with augmentation enabled?\n3. Could the authors report the results with the same runtime allocated to each algorithm in the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method (MEMENTO) for online fine-tuning of neural CO models. More concretely, MEMENTO learns the rules for updating the policy parameters at inference time. This is achieved by leveraging a learned residual policy for the base policy, where this residual policy utilizes the past solution predictions to explore the search space. Through experiments on the TSP and the CVRP, the paper demonstrates MEMENTO's efficacy against baseline methods.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper conveys most of the ideas clearly.\n- The idea of leveraging a learned residual policy for online fine-tuning of the base policy at inference time is interesting and novel.", "weaknesses": "- MEMENTO prevents the retrieval step from being too costly by only collecting data from the same node we are currently in. Could the authors elaborate on why this is a good retrieval strategy? Information should somehow be retrieved based on the partial solution constructed, since node-level decisions could be vastly different depending on the overall solutions.\n- The experiments do not use augmentation with symmetries in the experiments, which they claim to be not critical by citing just one prior work (COMPASS). Overall, to claim that MEMENTO outperforms EAS, the authors should compare it against EAS with augmentation enabled since EAS has been shown to work better that way. \n- My main concern with this paper is regarding the runtime of different methods in the experiments (which aren't reported for the major experiments in the main paper and put in the appendices instead). In Appendix A.1, the authors clarify that they report the runtime to solve one instance rather than the entire dataset. Given this, the comparison is fair only if they give each algorithm the same amount of runtime. However, Table 2 in the appendices shows that MEMENTO takes more than 2x the runtime than COMPASS (more than 4x for CVPR-100). The runtime is significantly higher than that for EAS as well (more than 2x) for CVRP-100.\n- The paper makes a few incorrect statements or claims:\n         1. In the definition of $\\pi^\\star$ on Page 3, if we are taking a max over $i$, why does $i$ appear in the outer expectation?\n         2. The claim that MEMENTO at least learns the REINFORCE update in the worst case is not exactly correct. In the worst case, the residual policy could output random values. I think the authors wanted to claim that the residual policy has the ability to at least learn the REINFORCE update rule (if nothing better). \n        3. The paper claims that MEMENTO  is \"designed\" to be agnostic to the base policy. While the authors demonstrate empirically that the learned residual policy for POMO could be used with COMPASS, the design of the framework as such makes the residual policy very much dependent on the base policy since they're learned jointly.", "questions": "1. Could the authors suggest alternative strategies for retrieval that do consider the partial solution constructed?\n2. Could the authors report the results with augmentation enabled?\n3. Could the authors report the results with the same runtime allocated to each algorithm in the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730755192676}, {"id": "ATho4GyA9W", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6399/Reviewer_Sd71"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "This paper introduces a fine-tuning method applied during inference to enhance construction methods for combinatorial optimization. It stores historical trajectories from fine-tuning as memory, which is processed by an MLP to adjust the original action probabilities. New solutions are then sampled from this memory-augmented distribution. The pretrained model and memory network are updated using an improvement reward based on the difference between the current solutions and the best-so-far solutions. Experiments are conducted on TSP and CVRP with problem sizes of 100 (generalizing to 125, 150, and 200) and 500.", "review_text": "This paper introduces a fine-tuning method applied during inference to enhance construction methods for combinatorial optimization. It stores historical trajectories from fine-tuning as memory, which is processed by an MLP to adjust the original action probabilities. New solutions are then sampled from this memory-augmented distribution. The pretrained model and memory network are updated using an improvement reward based on the difference between the current solutions and the best-so-far solutions. Experiments are conducted on TSP and CVRP with problem sizes of 100 (generalizing to 125, 150, and 200) and 500.", "strengths": "1. The source code is provided.\n2. This work has the potential to enhance the pre-trained construction methods for TSP and CVRP.\n3. The idea of reusing the historical trajectories (i.e. the memory) is interesting.", "weaknesses": "1. Marginal Improvement: The improvement of MEMENTO over EAS appears marginal, especially when generalizing to larger scales (e.g., TSP200 in Figure 3). In the CVRP results, the improvement of MEMENTO is also not significant.\n2. Scalability Concerns: In the larger-scale experiment (n=500), MEMENTO only slightly outperforms COMPASS, while introducing higher computational overhead.\n3. Incomplete Literature Review: The literature review lacks coverage of works focused on scalability and generalization.\n4. High Computational Cost of Fine-Tuning: The proposed fine-tuning method introduces additional computational overhead.\n5. Missing Generalization Experiments: Would be useful to add some generalization experiments on different distributions.\n6. Writing Quality: The writing lacks logical flow, and the table formatting needs improvement.", "questions": "1. Why does POMO with sampling strategies perform worse than POMO with the greedy rollout when generalizing to N=200 in Figure 3?\n2. Would it be feasible to apply MEMENTO to improvement methods as well? If so, are there any considerations that would impact its performance?\n3. What's the inference time of the methods displayed in Figure 3? The bar chart results seem to replicate those in the table on the left, making it feel a bit redundant\n4. Can MEMENTO be applied to other COPs? For example, I noticed that the code includes a preliminary implementation for the Knapsack Problem. I’d be interested to see the results and understand how MEMENTO might extend to different COPs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a fine-tuning method applied during inference to enhance construction methods for combinatorial optimization. It stores historical trajectories from fine-tuning as memory, which is processed by an MLP to adjust the original action probabilities. New solutions are then sampled from this memory-augmented distribution. The pretrained model and memory network are updated using an improvement reward based on the difference between the current solutions and the best-so-far solutions. Experiments are conducted on TSP and CVRP with problem sizes of 100 (generalizing to 125, 150, and 200) and 500.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "1. The source code is provided.\n2. This work has the potential to enhance the pre-trained construction methods for TSP and CVRP.\n3. The idea of reusing the historical trajectories (i.e. the memory) is interesting.", "weaknesses": "1. Marginal Improvement: The improvement of MEMENTO over EAS appears marginal, especially when generalizing to larger scales (e.g., TSP200 in Figure 3). In the CVRP results, the improvement of MEMENTO is also not significant.\n2. Scalability Concerns: In the larger-scale experiment (n=500), MEMENTO only slightly outperforms COMPASS, while introducing higher computational overhead.\n3. Incomplete Literature Review: The literature review lacks coverage of works focused on scalability and generalization.\n4. High Computational Cost of Fine-Tuning: The proposed fine-tuning method introduces additional computational overhead.\n5. Missing Generalization Experiments: Would be useful to add some generalization experiments on different distributions.\n6. Writing Quality: The writing lacks logical flow, and the table formatting needs improvement.", "questions": "1. Why does POMO with sampling strategies perform worse than POMO with the greedy rollout when generalizing to N=200 in Figure 3?\n2. Would it be feasible to apply MEMENTO to improvement methods as well? If so, are there any considerations that would impact its performance?\n3. What's the inference time of the methods displayed in Figure 3? The bar chart results seem to replicate those in the table on the left, making it feel a bit redundant\n4. Can MEMENTO be applied to other COPs? For example, I noticed that the code includes a preliminary implementation for the Knapsack Problem. I’d be interested to see the results and understand how MEMENTO might extend to different COPs.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730737187559}, {"id": "JKH9cPEdVv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6399/Reviewer_R4rn"], "rating": 5, "soundness": 2, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper proposes MEMENTO, a method for fast instance-specific adaptation in CO problems when a certain budget is allowed by utilizing a memory module. The proposed memory module is a lookup table storing information encountered during the online search process that is important for decision-making, such as the outcome of certain actions. This is then fed into an MLP, whose output modifies the action probabilities at each step. The method is evaluated in standard routing problems TSP and CVRP up to 500 nodes where it demonstrates SOTA results against RL-trained autoregressive solvers for CO.", "review_text": "This paper proposes MEMENTO, a method for fast instance-specific adaptation in CO problems when a certain budget is allowed by utilizing a memory module. The proposed memory module is a lookup table storing information encountered during the online search process that is important for decision-making, such as the outcome of certain actions. This is then fed into an MLP, whose output modifies the action probabilities at each step. The method is evaluated in standard routing problems TSP and CVRP up to 500 nodes where it demonstrates SOTA results against RL-trained autoregressive solvers for CO.", "strengths": "1. The paper is very well written and clear with relevant citations to previous literature.\n    \n2. The proposed MEMENTO module is novel and simple enough to be applied to a range of autoregressive CO solvers in the future, and thus, I believe it is an available addition to the NCO community.\n    \n3. Good overall performance (albeit with some concerns about baselines below) and experimental validation, including classical benchmarking, zero-shot combination with new solvers, and scaling to large sizes.\n    \n4. Code is provided, and authors make an effort to make their checkpoints available to the community.", "weaknesses": "1. My biggest concern is about fairness in comparison with EAS. In particular, the values reported in the paper are worse than the ones in the original paper. For instance, for the Kool et al. (2019) 10k instances with 100 nodes, the value reported is 7.778 vs the original 7.769 for the TSP (MEMENTO: 7.768) and 15.66 vs 15.63 for the CVRP (MEMENTO: 15.65). Compared to the values reported in the original paper, MEMENTO would be worse than EAS. This holds true at larger sizes too. Do the authors have an explanation for this? \n\n2. MEMENTO is only applied to routing problems despite the title appealing to a broader “Combinatorial Optimization”. In this sense, experimenting with differently structured environments such as the Job Shop Scheduling (JSSP) as done in COMPASS and EAS would be beneficial.\n    \n3. When computing gaps, it would be best to do so compared to SOTA heuristic methods. For instance: in Figure 7, in Table (b), only LKH3 is reported, while HGS is much more powerful on CVRP. However, the authors do report HGS in the appendix, which obtains much better solutions than LKH. This also applies in particular to Table 1, where the “optimality gap” appears to be negative.\n    \n4. Given that only routing problems are considered in this paper, it would be beneficial to mention the additional work [1].\n\n5. Some questions regarding hyperparameters remain, see below questions.\n    \n\n---\n\n[1] Son, Jiwoo, et al. \"Meta-sage: Scale meta-learning scheduled adaptation with guided exploration for mitigating scale shift on combinatorial optimization.\" International Conference on Machine Learning. PMLR, 2023.", "questions": "1. What is the impact of the MLP in terms of cost? Since this has to be called each time, I wonder whether similar results could be obtained with a simple linear layer as well.\n    \n2. In Figure 2: MEMENTO’s logit update encourages with higher amplitude high-reward actions rather than low-return ones. Is this due to the `ReLU` applied to the reward, constraining it to be strictly positive? I wonder if this analysis would hold without such constraint.\n\n3. Can your method be applied to broader problems that include e.g. edge features as the JSSP?\n    \n4. What is the impact of the memory size? Would increasing it be beneficial?\n\n5. MEMENTO appears to be >$2\\times$ slower than EAS at larger sizes as seen in Appendix A.1. Would EAS, provided with as much time budget as MEMENTO, eventually surpass the latter?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes MEMENTO, a method for fast instance-specific adaptation in CO problems when a certain budget is allowed by utilizing a memory module. The proposed memory module is a lookup table storing information encountered during the online search process that is important for decision-making, such as the outcome of certain actions. This is then fed into an MLP, whose output modifies the action probabilities at each step. The method is evaluated in standard routing problems TSP and CVRP up to 500 nodes where it demonstrates SOTA results against RL-trained autoregressive solvers for CO.", "soundness": 2, "presentation": 4, "contribution": 3, "strengths": "1. The paper is very well written and clear with relevant citations to previous literature.\n    \n2. The proposed MEMENTO module is novel and simple enough to be applied to a range of autoregressive CO solvers in the future, and thus, I believe it is an available addition to the NCO community.\n    \n3. Good overall performance (albeit with some concerns about baselines below) and experimental validation, including classical benchmarking, zero-shot combination with new solvers, and scaling to large sizes.\n    \n4. Code is provided, and authors make an effort to make their checkpoints available to the community.", "weaknesses": "1. My biggest concern is about fairness in comparison with EAS. In particular, the values reported in the paper are worse than the ones in the original paper. For instance, for the Kool et al. (2019) 10k instances with 100 nodes, the value reported is 7.778 vs the original 7.769 for the TSP (MEMENTO: 7.768) and 15.66 vs 15.63 for the CVRP (MEMENTO: 15.65). Compared to the values reported in the original paper, MEMENTO would be worse than EAS. This holds true at larger sizes too. Do the authors have an explanation for this? \n\n2. MEMENTO is only applied to routing problems despite the title appealing to a broader “Combinatorial Optimization”. In this sense, experimenting with differently structured environments such as the Job Shop Scheduling (JSSP) as done in COMPASS and EAS would be beneficial.\n    \n3. When computing gaps, it would be best to do so compared to SOTA heuristic methods. For instance: in Figure 7, in Table (b), only LKH3 is reported, while HGS is much more powerful on CVRP. However, the authors do report HGS in the appendix, which obtains much better solutions than LKH. This also applies in particular to Table 1, where the “optimality gap” appears to be negative.\n    \n4. Given that only routing problems are considered in this paper, it would be beneficial to mention the additional work [1].\n\n5. Some questions regarding hyperparameters remain, see below questions.\n    \n\n---\n\n[1] Son, Jiwoo, et al. \"Meta-sage: Scale meta-learning scheduled adaptation with guided exploration for mitigating scale shift on combinatorial optimization.\" International Conference on Machine Learning. PMLR, 2023.", "questions": "1. What is the impact of the MLP in terms of cost? Since this has to be called each time, I wonder whether similar results could be obtained with a simple linear layer as well.\n    \n2. In Figure 2: MEMENTO’s logit update encourages with higher amplitude high-reward actions rather than low-return ones. Is this due to the `ReLU` applied to the reward, constraining it to be strictly positive? I wonder if this analysis would hold without such constraint.\n\n3. Can your method be applied to broader problems that include e.g. edge features as the JSSP?\n    \n4. What is the impact of the memory size? Would increasing it be beneficial?\n\n5. MEMENTO appears to be >$2\\times$ slower than EAS at larger sizes as seen in Appendix A.1. Would EAS, provided with as much time budget as MEMENTO, eventually surpass the latter?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730633329500}, {"id": "l29ENYAsma", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6399/Reviewer_bp3q"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces MEMENTO, a novel memory-based approach designed to enhance existing constructive solvers for combinatorial optimization problems. MEMENTO leverages information from past solution attempts to improve the construction process, using a multilayer perceptron (MLP) network that takes features such as past actions, log-likelihood, and remaining search budget as input. This information is used to adjust the solution generation policy during inference, leading to higher-quality solutions within the given computational budget. The authors demonstrate that MEMENTO can improve the performance of base constructive models like POMO or COMPASS on problems such as the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP).", "review_text": "The paper introduces MEMENTO, a novel memory-based approach designed to enhance existing constructive solvers for combinatorial optimization problems. MEMENTO leverages information from past solution attempts to improve the construction process, using a multilayer perceptron (MLP) network that takes features such as past actions, log-likelihood, and remaining search budget as input. This information is used to adjust the solution generation policy during inference, leading to higher-quality solutions within the given computational budget. The authors demonstrate that MEMENTO can improve the performance of base constructive models like POMO or COMPASS on problems such as the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP).", "strengths": "Adapting the solution generation process using memory to account for previous attempts is undoubtedly a valuable and significant research direction.", "weaknesses": "This paper introduces a 'meta-learning' approach in building solvers for combinatorial optimization problems, structured around two levels of machine learning. The lower-level learning takes place during inference, where previous solution attempts stored in memory are used to update the node selection policy. The upper-level (meta) learning, referred to as 'training' in the paper, involves training a multilayer perceptron (MLP) to find the optimal parameters that shape the behavior of the lower-level learning.\n\nFor the lower-level learning, I find the proposed method overly simplistic, with two primary issues.\n\nFirst, the method indiscriminately utilizes all data in memory associated with the current node. Past experiences at the same 'current node' should not automatically be considered as occurring in the same 'state.' Although I haven’t thoroughly verified this, I believe the current implementation doesn’t necessitate storing the entire raw history. Instead, it could maintain only the accumulated 'logit correction values' and update them as new data arrives. In this sense, the term 'memory' may be too generous for the proposed approach, which feels closer to a simple bookkeeping method (EAS-tab) or basic active learning, both of which adjust the policy after each trajectory rollout iteration. A more effective memory system would incorporate mechanisms to retrieve the most relevant and important data while discarding irrelevant or potentially harmful data.\n\nSecond, the input features chosen for learning appear somewhat arbitrary. While the inclusion of the 'budget' feature is a helpful addition, many of the other features seem to offer limited value (as shown in Figure 9), and there is no clear theoretical basis for their usefulness. Moreover, generalizing these features to other combinatorial problems beyond fixed-sized routing may prove challenging.\n\nRegarding the upper-level learning, the authors should provide a more precise and detailed explanation of their meta-learning approach, as applying meta-learning to combinatorial problems is an especially difficult task. To find a globally optimal solution using a reinforcement learning (RL) approach, an end-to-end method may be more suitable than optimizing over a few grouped trajectories, as done in this paper. To partly address this limitation, the authors manually adjust the learning process by logarithmically increasing the loss weight. An effectively designed RL approach would ideally discover optimal weights autonomously, eliminating the need for manual adjustments.\n\nIn addition to these concerns, the proposed method yields only marginal improvements over other baselines. Consequently, I believe the paper does not meet the high standards of ICLR.", "questions": "I would like to understand the MLP structure used by the authors, described in Table 4.\n\nDoes \"memory size 40, number of layers 2, hidden layers 8\" imply the following MLP structure?\n\n1) An input layer with 40 neurons\n2) A first hidden layer with 8 neurons\n3) A second hidden layer with 8 neurons\n4) An output layer with 1 neuron", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces MEMENTO, a novel memory-based approach designed to enhance existing constructive solvers for combinatorial optimization problems. MEMENTO leverages information from past solution attempts to improve the construction process, using a multilayer perceptron (MLP) network that takes features such as past actions, log-likelihood, and remaining search budget as input. This information is used to adjust the solution generation policy during inference, leading to higher-quality solutions within the given computational budget. The authors demonstrate that MEMENTO can improve the performance of base constructive models like POMO or COMPASS on problems such as the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP).", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Adapting the solution generation process using memory to account for previous attempts is undoubtedly a valuable and significant research direction.", "weaknesses": "This paper introduces a 'meta-learning' approach in building solvers for combinatorial optimization problems, structured around two levels of machine learning. The lower-level learning takes place during inference, where previous solution attempts stored in memory are used to update the node selection policy. The upper-level (meta) learning, referred to as 'training' in the paper, involves training a multilayer perceptron (MLP) to find the optimal parameters that shape the behavior of the lower-level learning.\n\nFor the lower-level learning, I find the proposed method overly simplistic, with two primary issues.\n\nFirst, the method indiscriminately utilizes all data in memory associated with the current node. Past experiences at the same 'current node' should not automatically be considered as occurring in the same 'state.' Although I haven’t thoroughly verified this, I believe the current implementation doesn’t necessitate storing the entire raw history. Instead, it could maintain only the accumulated 'logit correction values' and update them as new data arrives. In this sense, the term 'memory' may be too generous for the proposed approach, which feels closer to a simple bookkeeping method (EAS-tab) or basic active learning, both of which adjust the policy after each trajectory rollout iteration. A more effective memory system would incorporate mechanisms to retrieve the most relevant and important data while discarding irrelevant or potentially harmful data.\n\nSecond, the input features chosen for learning appear somewhat arbitrary. While the inclusion of the 'budget' feature is a helpful addition, many of the other features seem to offer limited value (as shown in Figure 9), and there is no clear theoretical basis for their usefulness. Moreover, generalizing these features to other combinatorial problems beyond fixed-sized routing may prove challenging.\n\nRegarding the upper-level learning, the authors should provide a more precise and detailed explanation of their meta-learning approach, as applying meta-learning to combinatorial problems is an especially difficult task. To find a globally optimal solution using a reinforcement learning (RL) approach, an end-to-end method may be more suitable than optimizing over a few grouped trajectories, as done in this paper. To partly address this limitation, the authors manually adjust the learning process by logarithmically increasing the loss weight. An effectively designed RL approach would ideally discover optimal weights autonomously, eliminating the need for manual adjustments.\n\nIn addition to these concerns, the proposed method yields only marginal improvements over other baselines. Consequently, I believe the paper does not meet the high standards of ICLR.", "questions": "I would like to understand the MLP structure used by the authors, described in Table 4.\n\nDoes \"memory size 40, number of layers 2, hidden layers 8\" imply the following MLP structure?\n\n1) An input layer with 40 neurons\n2) A first hidden layer with 8 neurons\n3) A second hidden layer with 8 neurons\n4) An output layer with 1 neuron", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729509364777}], "openreview_url": "https://openreview.net/forum?id=VHGZjZmzsO", "arxiv_id": "2406.16424", "paper_pdf": "papers/VHGZjZmzsO.pdf", "paper_pdf_sha256": "695c555e19ee2061f26614e1315cb93ebeae051d3251f729c0e8f23a8151ee06", "paper_pdf_bytes": 1336591, "paper_pdf_source": "openreview", "code_url": "https://github.com/instadeepai/memento", "code_repository": "instadeepai/memento", "code_commit": "10a65e2f09be81155bfa24ba78ebbbb78863fc28", "code_archive": "repos/VHGZjZmzsO.zip", "code_archive_sha256": "df47efb7df63d48027874b5d51ff0d567a38f8bc41c7ff3870148c483a9f4995", "code_archive_bytes": 190368, "code_file_count": 35, "code_extensions": {".py": 31, ".sh": 4}, "github_disk_usage_kb": 236, "github_languages": {"Python": 144938, "Dockerfile": 3904, "Shell": 3654, "Makefile": 1642}, "github_archived": false, "github_pushed_at": "2026-08-21T13:55:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/memory-enhanced-neural-solvers-for-efficient"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Dq7iJqKKM7", "year": 2024, "status": "rejected", "title": "Rethinking Independent Cross-Entropy Loss For Graph-Structured Data", "authors": ["Rui Miao", "Kaixiong Zhou", "Yili Wang", "Ninghao Liu", "Ying Wang", "Xin Wang"], "authorids": ["~Rui_Miao2", "~Kaixiong_Zhou1", "~Yili_Wang2", "~Ninghao_Liu2", "~Ying_Wang13", "~Xin_Wang54"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) have exhibited prominent performance in learning graph-structured data. Considering node classification task, the individual label distribution conditioned on node representation is used to predict its classes. Based on the i.i.d assumption among node labels, the traditional supervised learning simply sums up cross-entropy losses of the independent training nodes and applies the average loss to optimize GNNs' weights. But different from other data formats, the nodes are naturally connected and their classes are correlated to neighbors at the same cluster. It is found that the independent distribution modeling of node labels restricts GNNs' capability to generalize over the entire graph and defend adversarial attacks. In this work, we propose a new framework, termed joint-cluster supervised learning, to model the joint distribution of each node with its corresponding cluster. Rather than assuming the node labels are independent, we learn the joint distribution of node and cluster labels conditioned on their representations, and train GNNs with the obtained joint loss. In this way, the data-label reference signals extracted from the local cluster explicitly strengthen the discrimination ability on the target node. The extensive experiments on 12 benchmark datasets and 7 backbone models demonstrate that our joint-cluster supervised learning can effectively bolster GNNs' node classification accuracy. Furthermore, being benefited from the reference signals which may be free from spiteful interference, our learning paradigm significantly protects the node classification from being affected by the adversarial attack.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "GFgSCuZZzf", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission627/Reviewer_U1a1"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper argues that existing approaches for graph learning assume independent cross-entropy loss and ignores the inter-dependence induced by observed graph structures. In light of this, the authors propose a new objective that incorporates the inter-dependence among node points into the loss computation. This ensures that the loss for each node is dependent on other nodes during the training. Experiments on many benchmark datasets and using various GNNs as backbones verify the effectiveness of the new loss over the traditional loss function.", "review_text": "This paper argues that existing approaches for graph learning assume independent cross-entropy loss and ignores the inter-dependence induced by observed graph structures. In light of this, the authors propose a new objective that incorporates the inter-dependence among node points into the loss computation. This ensures that the loss for each node is dependent on other nodes during the training. Experiments on many benchmark datasets and using various GNNs as backbones verify the effectiveness of the new loss over the traditional loss function.", "strengths": "1. The paper is well written and easy to follow\n\n2. The proposed method seems reasonable and sound\n\n3. Experiments entail a lot of datasets and different GNNs as backbones", "weaknesses": "The major concern on this work is the potential over-claiming. The authors argued that existing approaches ignore the inter-dependence among node points for loss computation, which is incorrect. There are in fact quite a few existing works that already considered designing inter-dependent loss for graph learning tasks. \n\nFor example, [1] proposes a new objective based on conditional random field for node classification, and [2] harnesses label propagation as a re-weighted loss. Besides, there are also recent works proposing self-supervised loss that considers enforcing the consistency between connected nodes [3]. These approaches integrate the inter-dependence of nodes into the loss function for training.\n\nAnother weakness lies in the comparison in experiments. The current experiment only compares with the traditional cross-entropy loss, which is a very weak baseline. More comparison with other advanced methods, particularly the above-mentioned models are needed to well justify the efficacy of the new design.\n\n[1] Meng Qu, et al., Neural Structured Prediction for Inductive Node Classification, ICLR 2022\n\n[2] Hande Dong, et al., On the Equivalence of Decoupled Graph Convolution Network and Label Propagation, WWW 2021\n\n[3] Hengrui Zhang, et al., Localized Contrastive Learning on Graphs", "questions": "See weakness above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper argues that existing approaches for graph learning assume independent cross-entropy loss and ignores the inter-dependence induced by observed graph structures. In light of this, the authors propose a new objective that incorporates the inter-dependence among node points into the loss computation. This ensures that the loss for each node is dependent on other nodes during the training. Experiments on many benchmark datasets and using various GNNs as backbones verify the effectiveness of the new loss over the traditional loss function.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is well written and easy to follow\n\n2. The proposed method seems reasonable and sound\n\n3. Experiments entail a lot of datasets and different GNNs as backbones", "weaknesses": "The major concern on this work is the potential over-claiming. The authors argued that existing approaches ignore the inter-dependence among node points for loss computation, which is incorrect. There are in fact quite a few existing works that already considered designing inter-dependent loss for graph learning tasks. \n\nFor example, [1] proposes a new objective based on conditional random field for node classification, and [2] harnesses label propagation as a re-weighted loss. Besides, there are also recent works proposing self-supervised loss that considers enforcing the consistency between connected nodes [3]. These approaches integrate the inter-dependence of nodes into the loss function for training.\n\nAnother weakness lies in the comparison in experiments. The current experiment only compares with the traditional cross-entropy loss, which is a very weak baseline. More comparison with other advanced methods, particularly the above-mentioned models are needed to well justify the efficacy of the new design.\n\n[1] Meng Qu, et al., Neural Structured Prediction for Inductive Node Classification, ICLR 2022\n\n[2] Hande Dong, et al., On the Equivalence of Decoupled Graph Convolution Network and Label Propagation, WWW 2021\n\n[3] Hengrui Zhang, et al., Localized Contrastive Learning on Graphs", "questions": "See weakness above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698939900823}, {"id": "M54huDTmeC", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission627/Reviewer_npyM"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper studies the problem of discrepancy between the non-i.i.d. property of GNN and the MLE learning. It proposed a new loss function to address the problem and its performance is demonstrated by extensive experiments.", "review_text": "The paper studies the problem of discrepancy between the non-i.i.d. property of GNN and the MLE learning. It proposed a new loss function to address the problem and its performance is demonstrated by extensive experiments.", "strengths": "1. The studied problem is important.\n2. The idea of the proposed method is novel.\n3. The method is effective in comparison to the baseline cross-entropy loss.", "weaknesses": "1. Some notations or definitions haven't been clearly explained. See the questions.\n2. The discussion about the connection between (5) and (4d) is missing. This makes it difficult to follow (5).", "questions": "1. More explanation about the equivalance between (4c) and (4d) should be provided.\n2. In the definitions of $\\bar{z}_m$ and $\\bar{y}_m$, there are two indices $k$ and $i$ that are confusing. \n3. In (5), $y _i\\bar{y} _m^\\top$ is a matrix, which is not consistent with the shape of output of $g _\\phi$.\n4. What are the labeling rates for the datasets in Table 1? How does labeling rate influence the classification accuracy?\n5. How did the authors determine the hyperparameters of the compared methods?\n6. It is not clear why the improvement on balanced data is higher than imbalanced data.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of discrepancy between the non-i.i.d. property of GNN and the MLE learning. It proposed a new loss function to address the problem and its performance is demonstrated by extensive experiments.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The studied problem is important.\n2. The idea of the proposed method is novel.\n3. The method is effective in comparison to the baseline cross-entropy loss.", "weaknesses": "1. Some notations or definitions haven't been clearly explained. See the questions.\n2. The discussion about the connection between (5) and (4d) is missing. This makes it difficult to follow (5).", "questions": "1. More explanation about the equivalance between (4c) and (4d) should be provided.\n2. In the definitions of $\\bar{z}_m$ and $\\bar{y}_m$, there are two indices $k$ and $i$ that are confusing. \n3. In (5), $y _i\\bar{y} _m^\\top$ is a matrix, which is not consistent with the shape of output of $g _\\phi$.\n4. What are the labeling rates for the datasets in Table 1? How does labeling rate influence the classification accuracy?\n5. How did the authors determine the hyperparameters of the compared methods?\n6. It is not clear why the improvement on balanced data is higher than imbalanced data.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698839955287}, {"id": "5wZgeLk52N", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission627/Reviewer_8QYJ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes joint-cluster supervised learning for graph neural networks. Instead of adopting the cross-entropy loss for each node independently, the paper models the joint distribution of node and cluster labels, given their respective representations. Extensive experiments are conducted across multiple benchmarks, demonstrating the proposed loss can boost the performance of different backbone GNNs and robustness against adversarial attack.", "review_text": "The paper proposes joint-cluster supervised learning for graph neural networks. Instead of adopting the cross-entropy loss for each node independently, the paper models the joint distribution of node and cluster labels, given their respective representations. Extensive experiments are conducted across multiple benchmarks, demonstrating the proposed loss can boost the performance of different backbone GNNs and robustness against adversarial attack.", "strengths": "1. The proposed joint modeling of node and cluster is novel and sound.\n\n2. The scope of the experimental evaluation is broad, including small graphs and large graphs.\n\n3. The empirical analyses are comprehensive in terms of necessary discussions, comparisons, and visualizations.", "weaknesses": "1. The backbones adopted for experiments are mostly not those that perform the best on these benchmarks. It would be more convincing to see how the proposed loss boost the performance of strong GNN models, e.g., GCNII on Cora, that may give rise to sota performance.\n\n2. Similar concern to 1 also exists for analyses like Table 6 (with GCN) and Table 7 (with MLP).\n\n\n[1] Chen et al. Simple and deep graph convolutional networks. In ICML.", "questions": "1. Could more results with stronger backbones be provided on these datasets? e.g., GCNII on Cora.\n\n2. It is vague how this technique helps improve the best models that prevail on different tasks. For example, at least a comprehensive table which enumerates most recent or best performing methods on several datasets should be presented to give the readers an overview how this approach situate in the rich literatures in GNN-based node classification.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes joint-cluster supervised learning for graph neural networks. Instead of adopting the cross-entropy loss for each node independently, the paper models the joint distribution of node and cluster labels, given their respective representations. Extensive experiments are conducted across multiple benchmarks, demonstrating the proposed loss can boost the performance of different backbone GNNs and robustness against adversarial attack.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The proposed joint modeling of node and cluster is novel and sound.\n\n2. The scope of the experimental evaluation is broad, including small graphs and large graphs.\n\n3. The empirical analyses are comprehensive in terms of necessary discussions, comparisons, and visualizations.", "weaknesses": "1. The backbones adopted for experiments are mostly not those that perform the best on these benchmarks. It would be more convincing to see how the proposed loss boost the performance of strong GNN models, e.g., GCNII on Cora, that may give rise to sota performance.\n\n2. Similar concern to 1 also exists for analyses like Table 6 (with GCN) and Table 7 (with MLP).\n\n\n[1] Chen et al. Simple and deep graph convolutional networks. In ICML.", "questions": "1. Could more results with stronger backbones be provided on these datasets? e.g., GCNII on Cora.\n\n2. It is vague how this technique helps improve the best models that prevail on different tasks. For example, at least a comprehensive table which enumerates most recent or best performing methods on several datasets should be presented to give the readers an overview how this approach situate in the rich literatures in GNN-based node classification.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698625978740}], "openreview_url": "https://openreview.net/forum?id=Dq7iJqKKM7", "arxiv_id": "2405.15564", "paper_pdf": "papers/Dq7iJqKKM7.pdf", "paper_pdf_sha256": "1aad97e8dd8901ad0dade9ee74cde8dce8fececfe00bfd774e9c85ad5e8671fd", "paper_pdf_bytes": 823465, "paper_pdf_source": "openreview", "code_url": "https://github.com/MR9812/Joint-Cluster-Supervised-Learning", "code_repository": "MR9812/Joint-Cluster-Supervised-Learning", "code_commit": "bfa9dccd2f24be018a23c94152804831e2a35cf7", "code_archive": "repos/Dq7iJqKKM7.zip", "code_archive_sha256": "5776367ef7995273e21272934d74da7347b2dcea0f795c4745e7308c539d38c1", "code_archive_bytes": 99382, "code_file_count": 37, "code_extensions": {".py": 34, ".sh": 3}, "github_disk_usage_kb": 162, "github_languages": {"Python": 238618, "Shell": 5059}, "github_archived": false, "github_pushed_at": "2025-10-28T09:42:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rethinking-independent-cross-entropy-loss-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YdFkY-QHkPl", "year": 2023, "status": "rejected", "title": "Provably Learning Diverse Features in Multi-View Data with Midpoint Mixup", "authors": ["Muthu Chidambaram", "Xiang Wang", "Chenwei Wu", "Rong Ge"], "authorids": ["~Muthu_Chidambaram1", "~Xiang_Wang1", "~Chenwei_Wu1", "~Rong_Ge1"], "authors_source": "OpenReview API", "abstract": "Mixup is a data augmentation technique that relies on training using random convex combinations of data points and their labels. In recent years, Mixup has become a standard primitive used in the training of state-of-the-art image classification models due to its demonstrated benefits over empirical risk minimization with regards to generalization and robustness. In this work, we try to explain some of this success from a feature learning perspective. We focus our attention on classification problems in which each class may have multiple associated features (or views) that can be used to predict the class correctly. Our main theoretical results demonstrate that, for a non-trivial class of data distributions with two features per class, training a 2-layer convolutional network using empirical risk minimization can lead to learning only one feature for almost all classes while training with a specific instantiation of Mixup succeeds in learning both features for every class. We also show empirically that these theoretical insights extend to the practical settings of image benchmarks modified to have additional synthetic features.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "jK95Ubo3e-6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1591/Reviewer_Pxqu"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies a specific instance of Mixup data augmentation: Midpoint Mixup. It provides a theoretical characterization of the learning dynamics of neural networks with mixup-augmented data, and proved how it can improve feature learning by helping the neural network pick up diverse features, under the multi-view data framework proposed by Allen-Zhu and Li.\n\n\n[1] Zeyuan Allen-Zhu, Yuanzhi Li, Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning", "review_text": "This paper studies an interesting instance of Mixup augmentation and obtained some decent theory results with the help of an established theoretical framework, but the authors are not good at justifying their choice of study and their theoretical claims, and they fail at explaining the key technical ingredient of the paper, as well as how well their assumptions can connect to practical scenarios. ", "strengths": "Strength: Most of the paper is clearly written and it provides a rigorous analysis of the subject it studies. Mixup is an interesting data augmentation technique and tries to explain why it can help in improving the feature learning of neural nets. The theory result goes beyond the previous works that are based on linear models (as they are based on nonlinear neural networks) and is reasonable due to the feature learning technique it uses. It might be an important paper if the authors can fix the issues I find below.\n\nWeaknesses: \n1. The authors did not provide sufficient motivation why they specifically study Midpoint Mixup rather than the general mixup augmentations. Moreover, they did not provide an empirical comparison between the original mixup and the midpoint mixup, making it hard to justify their choice of study. A potential improvement is to explain the technical difficulty of analyzing the original mixup, and justifies that the midpoint mixup is competitive (or particularly interesting).\n2. Assumption 3.10, which seems to be the key assumption to migrate the proof technique in [2] to the setting of this paper, is very technical, without intuitive explanation, and without support from empirical evidence. It is hard to make sense of this assumption as assembling any real-world structure. This weakens the validity of the proposed explanation as the theory result could be just an artificial construction.\n3. The experimental results in this paper are not informative, as many of their metrics are too low. I do not understand why this paper constructs such special data by mixing images using Dirichlet distribution (see equation 4.1 and Fig 1) to compare ERM with Midpoint-Mixup. Moreover, the test errors often exceed 60%-90% for experiments on CIFAR-10/100 when L>1 (that is over the special dataset rather than the original Cifar10/100), meaning that the model struggles a lot to predict the mixup label. I don't think these experiments are sufficient to support the authors' choice of study or their theoretical claims.\n\n\n[1] Ruoqi Shen, Sebastien Bubeck, Suriya Gunasekar, Data Augmentation as Feature Manipulation\n[2] Zeyuan Allen-Zhu, Yuanzhi Li, Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies a specific instance of Mixup data augmentation: Midpoint Mixup. It provides a theoretical characterization of the learning dynamics of neural networks with mixup-augmented data, and proved how it can improve feature learning by helping the neural network pick up diverse features, under the multi-view data framework proposed by Allen-Zhu and Li.\n\n\n[1] Zeyuan Allen-Zhu, Yuanzhi Li, Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning", "strength_and_weaknesses": "Strength: Most of the paper is clearly written and it provides a rigorous analysis of the subject it studies. Mixup is an interesting data augmentation technique and tries to explain why it can help in improving the feature learning of neural nets. The theory result goes beyond the previous works that are based on linear models (as they are based on nonlinear neural networks) and is reasonable due to the feature learning technique it uses. It might be an important paper if the authors can fix the issues I find below.\n\nWeaknesses: \n1. The authors did not provide sufficient motivation why they specifically study Midpoint Mixup rather than the general mixup augmentations. Moreover, they did not provide an empirical comparison between the original mixup and the midpoint mixup, making it hard to justify their choice of study. A potential improvement is to explain the technical difficulty of analyzing the original mixup, and justifies that the midpoint mixup is competitive (or particularly interesting).\n2. Assumption 3.10, which seems to be the key assumption to migrate the proof technique in [2] to the setting of this paper, is very technical, without intuitive explanation, and without support from empirical evidence. It is hard to make sense of this assumption as assembling any real-world structure. This weakens the validity of the proposed explanation as the theory result could be just an artificial construction.\n3. The experimental results in this paper are not informative, as many of their metrics are too low. I do not understand why this paper constructs such special data by mixing images using Dirichlet distribution (see equation 4.1 and Fig 1) to compare ERM with Midpoint-Mixup. Moreover, the test errors often exceed 60%-90% for experiments on CIFAR-10/100 when L>1 (that is over the special dataset rather than the original Cifar10/100), meaning that the model struggles a lot to predict the mixup label. I don't think these experiments are sufficient to support the authors' choice of study or their theoretical claims.\n\n\n[1] Ruoqi Shen, Sebastien Bubeck, Suriya Gunasekar, Data Augmentation as Feature Manipulation\n[2] Zeyuan Allen-Zhu, Yuanzhi Li, Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning", "clarity,_quality,_novelty_and_reproducibility": "The clarity is decent. the novelty is partially limited because the proof/data framework is largely inherited from prior work, while the original part (such as assumption 3.10) lacks enough explanation.", "summary_of_the_review": "This paper studies an interesting instance of Mixup augmentation and obtained some decent theory results with the help of an established theoretical framework, but the authors are not good at justifying their choice of study and their theoretical claims, and they fail at explaining the key technical ingredient of the paper, as well as how well their assumptions can connect to practical scenarios. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666674014414}, {"id": "Dd6yeNe0tBn", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1591/Reviewer_uhyy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper focuses on mid-point mixup (a variant of mixup) and provides theoretical analysis on the effect of using mid-point mixup on multi-view data. The theoretical result is specific to the setting of the 2-layer convolutional networks and the datasets where samples at each class have 2 features. Under this setting, this paper shows that empricial risk minimization learn only single feature at each class, while mid-point mixup learns both features at each class. ", "review_text": "This paper focuses on an interesting problem, but has much room for improvement in terms of organizing the results, validating the assumption, and explaining how the experimental results are helpful in practical settings. ", "strengths": "- Strength\n    - The direction of analyzing mid-point mixup on mulit-view data is new and interesting.\n\n- Weakness\n\n    - Theory\n        - Why is the assumption of Lemma 2.1 (z_{i,j} is unique for all i,j pair) true in practice? Simple 2-dimensional dataset where each class follows Gaussian distribution cannot follow this assumption. (how about high-dimensional case?)\n        - It seems like Lemma 2.1 is important to show that Midpoint Mixup is not poor, but the proof of Lemma 2.1 is not fully understood on my side. (I put the details in the questions section)\n\n    - Experiments\n        - This paper shows that when L is large, Mid-point Mixup is having less error than ERM. Of course if L is large, it will go to the ideal multi-view dataset this paper is assuming, but then why don’t we even start with real data? Instead, making a synthetic multi-view data to show the effectiveness of the theory seems more reasonable. I personally feel the experiments in this paper is similar to just “simulation” instead of real-world experiments.\n        - In Table 1, the authors are comparing “test error 96%” versus “test error 97%”, when we have 100 classes. What does it mean to increase 1% test error from “nearly random guessing”? Does these empirical results have any positive impact on the practical deployment of mid-point mixup?\n\n    - Writing\n        - The assumptions and results are less organized (thus I have many questions below). The definition 3.4 has no intuitive explanation or visualization.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper focuses on mid-point mixup (a variant of mixup) and provides theoretical analysis on the effect of using mid-point mixup on multi-view data. The theoretical result is specific to the setting of the 2-layer convolutional networks and the datasets where samples at each class have 2 features. Under this setting, this paper shows that empricial risk minimization learn only single feature at each class, while mid-point mixup learns both features at each class. ", "strength_and_weaknesses": "- Strength\n    - The direction of analyzing mid-point mixup on mulit-view data is new and interesting.\n\n- Weakness\n\n    - Theory\n        - Why is the assumption of Lemma 2.1 (z_{i,j} is unique for all i,j pair) true in practice? Simple 2-dimensional dataset where each class follows Gaussian distribution cannot follow this assumption. (how about high-dimensional case?)\n        - It seems like Lemma 2.1 is important to show that Midpoint Mixup is not poor, but the proof of Lemma 2.1 is not fully understood on my side. (I put the details in the questions section)\n\n    - Experiments\n        - This paper shows that when L is large, Mid-point Mixup is having less error than ERM. Of course if L is large, it will go to the ideal multi-view dataset this paper is assuming, but then why don’t we even start with real data? Instead, making a synthetic multi-view data to show the effectiveness of the theory seems more reasonable. I personally feel the experiments in this paper is similar to just “simulation” instead of real-world experiments.\n        - In Table 1, the authors are comparing “test error 96%” versus “test error 97%”, when we have 100 classes. What does it mean to increase 1% test error from “nearly random guessing”? Does these empirical results have any positive impact on the practical deployment of mid-point mixup?\n\n    - Writing\n        - The assumptions and results are less organized (thus I have many questions below). The definition 3.4 has no intuitive explanation or visualization.", "clarity,_quality,_novelty_and_reproducibility": "The problem itself seems to be novel, but the theoretical results are unclear. Below I wrote some questions to clarify them. \n\n* I couldn't fully get the proof of Lemma 2.1. Why does the “existence of well-defined h simultaneously optimizing each term of eq.2.3” implies “RHS of (2.4) is achieving the minimum”, so that we have “$g^{y_i} (z_{i,j}) = g^{y_j} (z_{i,j})$ for all $z_{i,j}$”?\n\n* I couldn't get the sentence \"J(g, X) by just taking $⟨w_y, v_{y,1}⟩ \\rightarrow \\infty$ for every class y\" in the paper. I thought setting $w_c = v_{y,1}$ was enough to make the empirical loss = 0?\n\n* I couldn't get what footnote 1 means.\n\n* Sec.3.1 title is “linear model” and Sec.3.2 title is “multi-view data setup”. But it seems like Def.3.1 in Sec.3.1 is also multi-view. What is the criteria differentiating Sec.3.1 and 3.2? Maybe because Sec.3.2 for general model, not limited to linear model? If so, the section name is misleading.\n\n\nI am also adding minor comments\n* The first sentence of Definition 3.4 is “Identically to Definition 3.4”. Maybe typo for “Identically to Definition 3.1”?\n* Typo: themself → themselves", "summary_of_the_review": "This paper focuses on an interesting problem, but has much room for improvement in terms of organizing the results, validating the assumption, and explaining how the experimental results are helpful in practical settings. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666487492626}, {"id": "a02INeviT1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1591/Reviewer_9dJZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work adopts the feature learning framework by Allen-Zhu and Li (2021), and further introduces latent orthogonal feature views. The authors prove that under certain assumptions on features, a two-layer smooth-ReLU network can learn all features with midpoint mixup, whereas ERM can only learn one feature.", "review_text": "In general, I believe this work contributes significantly to the feature learning topic.\nI hope the authors could fix my concerns above and add more clarifications.", "strengths": "Strength:\n1. On top of Allen-Zhu and Li (2021), the authors further decompose the input into multiple orthonormal feature views. This allows the authors to analyze the contribution of each individual feature and their gradient correlations.\n2. Although the change of data settings (section 3.1) and assumptions (3.10) is simple, the authors provide significant efforts in proving the feature learning under the mixup loss.\n\nWeakness:\n* Theory:\n1) In Eq. B.34, do you need a lower bound or dependence on the training iterations, such that the feature correlations will not be over-corrected?\n2) Could you add more intuitive explanations of the meaning of function f in Assumption 3.10? Is it possible to analytically prove the monotonicity of f?\n\n* Experiments:\n1) Thm 3.9 and 3.11 did not directly imply better test accuracy (generalization). Instead, is it possible to empirically verify the one or all feature(s) (v) learned (defined by 3.8) by the Midpoint Mixup?\n2) It would be better to add variances in Table 1.\n3) Many test errors in Table 1 are far worse than random guess (90% error rate). From Eq. 4.1, I did not see any noise is introduced. What is the reason for this large test error?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work adopts the feature learning framework by Allen-Zhu and Li (2021), and further introduces latent orthogonal feature views. The authors prove that under certain assumptions on features, a two-layer smooth-ReLU network can learn all features with midpoint mixup, whereas ERM can only learn one feature.", "strength_and_weaknesses": "Strength:\n1. On top of Allen-Zhu and Li (2021), the authors further decompose the input into multiple orthonormal feature views. This allows the authors to analyze the contribution of each individual feature and their gradient correlations.\n2. Although the change of data settings (section 3.1) and assumptions (3.10) is simple, the authors provide significant efforts in proving the feature learning under the mixup loss.\n\nWeakness:\n* Theory:\n1) In Eq. B.34, do you need a lower bound or dependence on the training iterations, such that the feature correlations will not be over-corrected?\n2) Could you add more intuitive explanations of the meaning of function f in Assumption 3.10? Is it possible to analytically prove the monotonicity of f?\n\n* Experiments:\n1) Thm 3.9 and 3.11 did not directly imply better test accuracy (generalization). Instead, is it possible to empirically verify the one or all feature(s) (v) learned (defined by 3.8) by the Midpoint Mixup?\n2) It would be better to add variances in Table 1.\n3) Many test errors in Table 1 are far worse than random guess (90% error rate). From Eq. 4.1, I did not see any noise is introduced. What is the reason for this large test error?", "clarity,_quality,_novelty_and_reproducibility": "1. Overall the paper writing is clear and easy to follow.\n2. As mentioned above, the main novelty of this work is to prove the benefit of the mixup in feature learning under the multi-view data setting.", "summary_of_the_review": "In general, I believe this work contributes significantly to the feature learning topic.\nI hope the authors could fix my concerns above and add more clarifications.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666169556434}], "openreview_url": "https://openreview.net/forum?id=YdFkY-QHkPl", "arxiv_id": "2210.13512", "paper_pdf": "papers/YdFkY-QHkPl.pdf", "paper_pdf_sha256": "d4c1fee17ce9cb685a56995df52747a7a5184055fd3c05ef7f1a9af1ad29873b", "paper_pdf_bytes": 807669, "paper_pdf_source": "openreview", "code_url": "https://github.com/2014mchidamb/midpoint-mixup-multi-view-icml", "code_repository": "2014mchidamb/midpoint-mixup-multi-view-icml", "code_commit": "e81305d0d48a61d8690530fff785dc09daa6d6e2", "code_archive": "repos/YdFkY-QHkPl.zip", "code_archive_sha256": "40ae5582d81ae55bd881d0fba028fd7bbe597204cc9a0436dd6f0d1c2306221a", "code_archive_bytes": 356559, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 343, "github_languages": {"Python": 36732, "Shell": 799}, "github_archived": false, "github_pushed_at": "2023-05-27T16:58:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/provably-learning-diverse-features-in-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xs-tJn58XKv", "year": 2022, "status": "rejected", "title": "Learning Stable Classifiers by Transferring Unstable Features", "authors": ["Yujia Bao", "Shiyu Chang", "Regina Barzilay"], "authorids": ["~Yujia_Bao1", "~Shiyu_Chang2", "~Regina_Barzilay1"], "authors_source": "OpenReview API", "abstract": "While unbiased machine learning models are essential for many applications, bias is a human-defined concept that can vary across tasks. Given only input-label pairs, algorithms may lack sufficient information to distinguish stable (causal) features from unstable (spurious) features. However, related tasks often share similar biases -- an observation we may leverage to develop stable classifiers in the transfer setting. In this work, we explicitly inform the target classifier about unstable features in the source tasks. Specifically, we derive a representation that encodes the unstable features by contrasting different data environments in the source task. We achieve robustness by clustering data of the target task according to this representation and minimizing the worst-case risk across these clusters. We evaluate our method on both text and image classifications. Empirical results demonstrate that our algorithm is able to maintain robustness on the target task, outperforming the best baseline by 22.9% in absolute accuracy across 12 transfer settings. Our code and data will be publicly available.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "Uuw81769svx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3939/Reviewer_hKdn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors introduce a method (TOFU) for learning classifiers that are robust to spurious correlations in a transfer learning setting. They argue that approaches which rely only on (input, label) pairs and use no extra information to identify spurious features could fail to learn a robust model due to insufficient data for the target task. The authors provide scenarios where useful metadata, namely varying environments, could be available for source tasks, ready to be leveraged by the target task classifier for the purpose of identifying a shared bias that might exist across all tasks. To identify unstable/spurious features, the authors follow a 3-step procedure proposed in an existing work. Namely, a classifier is trained on data from environment E1, and is evaluated on data from a second environment E2. The data in E2 is then partitioned according to the correctness of the classifier's predictions on a per-class basis. A metric learning objective using a triplet loss is used to learn a model which embeds samples according to their unstable features with the goal that the unstable features of the correctly predicted samples should be closer together than unstable features between correctly predicted and incorrectly predicted samples. The objective is justified based on prior theoretical work and is illustrated visually. Extensive empirical evaluations are performed, comparing the performance of TOFU against many different baseline methods on both image and text classification tasks.", "review_text": "The paper itself is well written, and does a good job at describing the current research landscape of robust feature learning. Some more thought should be given to the example provided of why a method like TOFU is required, since the current one of using class labels to predict color instead of vice versa seems contrived, and undersells the usefulness of the method. The theoretical claims made throughout the paper are sound, but it is not obvious that they should hold in practice. For example, on datasets where there is heterogeneity in terms of the difficulty in predicting samples, i.e. easy, medium, and ambiguous samples, easy and medium samples might end up in X_1 and X_2, and the ambiguous samples in X_3. Thus, the objective in equation (1) could end up optimizing not for the unstable features in X_1 and X_2 to be closer together than those in X_1 and X_3, but rather for easy samples to be closer together than easy and difficult samples in the learned feature space Z. Additionally, it may be possible to improve S.2 by considering not only correctly predicted vs incorrectly predicted, but also the particular errors that were made. Consider a scenario where images of cars are either misclassified as boats or bicycles. Bicycles are more likely to be on the road than boats, so it may be wise to exclude the cars misclassified as bicycles from X_3, since those contain the correct unstable feature being sought, but violate the assumption of theorem 1. It seems like for TOFU to work, the classifiers trained in step S.1 need to be unstable such that they will make sufficient errors on E_2 in order to have sufficient negative samples for the triplet loss objective in S.3. The limitations of metric learning with a triplet loss should be stated since they could have an influence on model selection done throughout the method.\n \nTOFU has fewer requirements in terms of additional annotations compared to other methods, and domain specific prior knowledge is encoded into how the datasets are split into E_1 and E_2. Quantitative analysis of TOFU reveals that it is significantly superior to the many baselines considered, and that is consistent on all tasks and datasets. This is impressive and rare since many methods in the broader deep learning literature improve on some tasks, but do worse on others. It is also satisfying to see how well TOFU performs in terms of clustering metrics compared to representations learned only for the purpose of classifying the source task. Naturally, the latter representations would contain information encoding the spurious features since they are then leveraged to predict the class label, but they seem to be highly ineffective for the purpose of clustering by unstable feature. Having said that, this analysis is a bit confounded by the fact that ERM uses supervised learning whereas TOFU uses metric learning. Additionally, it would be ideal to have a discussion regarding why the performance improvement of TOFU over ERM is not as significant for the Beer Review data compared to MNIST, even though the clustering scores are much better. In general, the difficulty of the tasks considered for the empirical evaluation should be discussed as it seems to vary quite a bit.\n \nThe novelty of the work is somewhat limited. S.1 to S.3 is nearly identical to the cited work of Bao et al., except that the objective in S.3 is changed to the triplet loss objective. This objective comes from Theorem 1 which is a novel contribution, but is based on theoretical results also from the cited work of Bao et al. It would also be recommended that more real life examples of the transfer learning scenario in question be provided, since it currently seems as if it was created to show the applicability of the method. Overall, the paper is very strong from an empirical perspective, is well structured and reproducible, but ignores any potential limitations and is limited in its discussion.\n\nStrengths\n-Paper is well written and self-contained.\n-Impressive empirical results on many tasks/datasets relevant for this application.\n-Provides an interesting perspective on combining transfer learning and robust feature learning,\n \nWeaknesses\n-Not so clear how the datasets were split to generate different E1 and E2. For example, in the Waterbirds dataset, if the source task is classifying waterfowl, is the only difference between E1 and E2 the percentage of images that have a water background? This is ambiguous in the appendix as well.\n-Conclusion is too short, and is missing a discussion of limitations and future directions.\n-A clear limitation of the work is that the unstable features in the target task have to be the same ones as in the source task\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors introduce a method (TOFU) for learning classifiers that are robust to spurious correlations in a transfer learning setting. They argue that approaches which rely only on (input, label) pairs and use no extra information to identify spurious features could fail to learn a robust model due to insufficient data for the target task. The authors provide scenarios where useful metadata, namely varying environments, could be available for source tasks, ready to be leveraged by the target task classifier for the purpose of identifying a shared bias that might exist across all tasks. To identify unstable/spurious features, the authors follow a 3-step procedure proposed in an existing work. Namely, a classifier is trained on data from environment E1, and is evaluated on data from a second environment E2. The data in E2 is then partitioned according to the correctness of the classifier's predictions on a per-class basis. A metric learning objective using a triplet loss is used to learn a model which embeds samples according to their unstable features with the goal that the unstable features of the correctly predicted samples should be closer together than unstable features between correctly predicted and incorrectly predicted samples. The objective is justified based on prior theoretical work and is illustrated visually. Extensive empirical evaluations are performed, comparing the performance of TOFU against many different baseline methods on both image and text classification tasks.", "main_review": "The paper itself is well written, and does a good job at describing the current research landscape of robust feature learning. Some more thought should be given to the example provided of why a method like TOFU is required, since the current one of using class labels to predict color instead of vice versa seems contrived, and undersells the usefulness of the method. The theoretical claims made throughout the paper are sound, but it is not obvious that they should hold in practice. For example, on datasets where there is heterogeneity in terms of the difficulty in predicting samples, i.e. easy, medium, and ambiguous samples, easy and medium samples might end up in X_1 and X_2, and the ambiguous samples in X_3. Thus, the objective in equation (1) could end up optimizing not for the unstable features in X_1 and X_2 to be closer together than those in X_1 and X_3, but rather for easy samples to be closer together than easy and difficult samples in the learned feature space Z. Additionally, it may be possible to improve S.2 by considering not only correctly predicted vs incorrectly predicted, but also the particular errors that were made. Consider a scenario where images of cars are either misclassified as boats or bicycles. Bicycles are more likely to be on the road than boats, so it may be wise to exclude the cars misclassified as bicycles from X_3, since those contain the correct unstable feature being sought, but violate the assumption of theorem 1. It seems like for TOFU to work, the classifiers trained in step S.1 need to be unstable such that they will make sufficient errors on E_2 in order to have sufficient negative samples for the triplet loss objective in S.3. The limitations of metric learning with a triplet loss should be stated since they could have an influence on model selection done throughout the method.\n \nTOFU has fewer requirements in terms of additional annotations compared to other methods, and domain specific prior knowledge is encoded into how the datasets are split into E_1 and E_2. Quantitative analysis of TOFU reveals that it is significantly superior to the many baselines considered, and that is consistent on all tasks and datasets. This is impressive and rare since many methods in the broader deep learning literature improve on some tasks, but do worse on others. It is also satisfying to see how well TOFU performs in terms of clustering metrics compared to representations learned only for the purpose of classifying the source task. Naturally, the latter representations would contain information encoding the spurious features since they are then leveraged to predict the class label, but they seem to be highly ineffective for the purpose of clustering by unstable feature. Having said that, this analysis is a bit confounded by the fact that ERM uses supervised learning whereas TOFU uses metric learning. Additionally, it would be ideal to have a discussion regarding why the performance improvement of TOFU over ERM is not as significant for the Beer Review data compared to MNIST, even though the clustering scores are much better. In general, the difficulty of the tasks considered for the empirical evaluation should be discussed as it seems to vary quite a bit.\n \nThe novelty of the work is somewhat limited. S.1 to S.3 is nearly identical to the cited work of Bao et al., except that the objective in S.3 is changed to the triplet loss objective. This objective comes from Theorem 1 which is a novel contribution, but is based on theoretical results also from the cited work of Bao et al. It would also be recommended that more real life examples of the transfer learning scenario in question be provided, since it currently seems as if it was created to show the applicability of the method. Overall, the paper is very strong from an empirical perspective, is well structured and reproducible, but ignores any potential limitations and is limited in its discussion.\n\nStrengths\n-Paper is well written and self-contained.\n-Impressive empirical results on many tasks/datasets relevant for this application.\n-Provides an interesting perspective on combining transfer learning and robust feature learning,\n \nWeaknesses\n-Not so clear how the datasets were split to generate different E1 and E2. For example, in the Waterbirds dataset, if the source task is classifying waterfowl, is the only difference between E1 and E2 the percentage of images that have a water background? This is ambiguous in the appendix as well.\n-Conclusion is too short, and is missing a discussion of limitations and future directions.\n-A clear limitation of the work is that the unstable features in the target task have to be the same ones as in the source task\n", "summary_of_the_review": "This work introduces a new scenario where current methods for learning robust features are not capable of leveraging all available data. It also provides a solution in the form of a method called TOFU which has few requirements in terms of additional data annotations compared to existing methods. The authors recognize that bias is a human defined concept, and could vary from dataset to dataset, so it is best to have a method that can identify unstable features in an automated way. TOFU is an extension of an existing work, and has some theoretical motivations, but the underlying assumptions may be too strong. TOFU outperforms all other methods considered by a large margin, and the authors investigate a potential source of its success by confirming that it is indeed able to cluster data well according to unstable features: a requirement for later on doing group DRO. Sufficient details are provided to enable reproducibility, and many baselines are compared against, making the experiments comprehensive. However, the method is not particularly novel, nor is it clear if the scenario presented occurs often in real life. Limitations of the method are not considered, even though there are a few clear ones both technical and empirical.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635855414351}, {"id": "xPf11zDbJZ6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3939/Reviewer_J8M5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a two-stage method for transfer learning, which firstly infers unstable features from the source task and then learns stable correlations for the target. Experimental results validate the effectiveness of the proposed method.", "review_text": "(1) Strength: The problem that this paper addressed is important and essential for machine learning. And the whole pipeline is reasonable and achieves good results.\n\t(2) Weakness: \n\t\t(a) Method: The proposed method infers the unstable features on the source task and then learns the stable correlation on the target leveraging the knowledge of unstable features. As for the methods, it uses the Group-DRO as the backbone and clusters the available testing data to produce the group labels. However, I think the method is quite ad-hoc and naïve, and it lacks technical contributions. The idea of clustering data with unstable features is also quite similar to the HRM[1], which further reduces the contributions.\n\t\t(b)Problem Setting: This method requires both labeled data from the source task and target task, which uses much much more information than existing methods for OOD generalization. \n\t\t(c)Theoretical Analysis: There is almost no theoretical analysis for the proposed method. Why it can infer the unstable features? Why this method can generalize to unseen domains? What if the worst-case group also reflects some spurious correlations?\n\t\t(d)Experiments: Since it is both the OOD generalization method and the transfer learning method (setting is the same), there lack many baselines. I think more transfer learning methods should be taken into consideration, as well as some domain generalization methods.\n\t\t(e)Definitions: What do the stable features or unstable features mean in this paper? What is the difference between causal features[2], invariant features[3], or stable features[4][5]? What is the formal definition of stable and unstable? If it is just the same as causal features, why use a totally new name? I think the authors missed many related works here.\n\n[1] Liu, J., Hu, Z., Cui, P., Li, B., & Shen, Z. (2021). Heterogeneous Risk Minimization. ICML2021\n[2] Peters, J., Bühlmann, P., & Meinshausen, N. (2016). Causal inference by using invariant prediction: identification and confidence intervals. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 947-1012.\n[3 Koyama, M., & Yamaguchi, S. (2020). Out-of-distribution generalization with maximal invariant predictor. arXiv preprint arXiv:2008.01883.]\n[4] Kuang, K., Xiong, R., Cui, P., Athey, S. and Li, B., 2020, April. Stable prediction with model misspecification and agnostic distribution shift. AAAI 2020\n[5] Shen, Z., Cui, P., Liu, J., Zhang, T., Li, B. and Chen, Z., 2020, August. Stable learning via differentiated variable decorrelation. KDD 2020\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a two-stage method for transfer learning, which firstly infers unstable features from the source task and then learns stable correlations for the target. Experimental results validate the effectiveness of the proposed method.", "main_review": "(1) Strength: The problem that this paper addressed is important and essential for machine learning. And the whole pipeline is reasonable and achieves good results.\n\t(2) Weakness: \n\t\t(a) Method: The proposed method infers the unstable features on the source task and then learns the stable correlation on the target leveraging the knowledge of unstable features. As for the methods, it uses the Group-DRO as the backbone and clusters the available testing data to produce the group labels. However, I think the method is quite ad-hoc and naïve, and it lacks technical contributions. The idea of clustering data with unstable features is also quite similar to the HRM[1], which further reduces the contributions.\n\t\t(b)Problem Setting: This method requires both labeled data from the source task and target task, which uses much much more information than existing methods for OOD generalization. \n\t\t(c)Theoretical Analysis: There is almost no theoretical analysis for the proposed method. Why it can infer the unstable features? Why this method can generalize to unseen domains? What if the worst-case group also reflects some spurious correlations?\n\t\t(d)Experiments: Since it is both the OOD generalization method and the transfer learning method (setting is the same), there lack many baselines. I think more transfer learning methods should be taken into consideration, as well as some domain generalization methods.\n\t\t(e)Definitions: What do the stable features or unstable features mean in this paper? What is the difference between causal features[2], invariant features[3], or stable features[4][5]? What is the formal definition of stable and unstable? If it is just the same as causal features, why use a totally new name? I think the authors missed many related works here.\n\n[1] Liu, J., Hu, Z., Cui, P., Li, B., & Shen, Z. (2021). Heterogeneous Risk Minimization. ICML2021\n[2] Peters, J., Bühlmann, P., & Meinshausen, N. (2016). Causal inference by using invariant prediction: identification and confidence intervals. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 947-1012.\n[3 Koyama, M., & Yamaguchi, S. (2020). Out-of-distribution generalization with maximal invariant predictor. arXiv preprint arXiv:2008.01883.]\n[4] Kuang, K., Xiong, R., Cui, P., Athey, S. and Li, B., 2020, April. Stable prediction with model misspecification and agnostic distribution shift. AAAI 2020\n[5] Shen, Z., Cui, P., Liu, J., Zhang, T., Li, B. and Chen, Z., 2020, August. Stable learning via differentiated variable decorrelation. KDD 2020\n", "summary_of_the_review": "This paper proposes a two-stage method for OOD generalization and transfer learning problems. However, there exist many weaknesses, including method, problem setting, theoretical analysis and experiments. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635836929392}, {"id": "ajpX0cd9MrX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3939/Reviewer_L5Kn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper is concerned with learning a model in situations when some features have spurious correlations with the label (for example, for classifying sheep vs camel, background can be a spurious feature). The idea is that if a lot of source data is available, these spurious correlation features should be easy to find. To find the unstable features, authors hypothesize that they are related to mistakes that a classifier makes in different environments. Therefore, they learn a model on one environment and run it on a different environment (source data), splitting the data into correctly predicted and not. Then for each partition (correct or not) examples with the same labels are encouraged to be close together, so the embedding learns UNSTABLE features (fz). Then the target data is clustered based on unstable features representation fz, and then DRO (existing method that assumes the existance of correct groupping based on unstable features) is used to train a robust target classifier", "review_text": "In general, the method seems to work and it seems to be justified (both theoretically, albeit in the appendix, and practically). It is really complicated though (involves many steps) and i fail to make a connection between this and say Domain Invariant Representation learning. For example, if i was to solve the problem of domain generalization (the setup seems really similar to me), a reasonable baseline would be to train a source model that learns an invariantt representation between various environments (using DANN, or CMD or MMD or whatever). But this is not in the baselines. Why is that? One difference i \"kinda\" see is that domains are more or less pure - eg domain is really a different background. The way i read it is that environement is more mixed - it can contain multiple backgrounds. It would be nice to draw comparison between your method and DIRL or at least explain why DIRL is not applicable\n\n\nPros:\n- The paper is well written and easy to follow (albeit a lot of helpful info is also in the Appendix)\n- The experimental results seem convincing, there is a study of dependence of (some) hyperparameters on the end result (num of clusters)\n\nCons:\n- DRO needs at least a brief introduction\n- Experiments are only on 2 environements (domains) in train\n- The method seems extremely costly: for a number n of source environements, you train n classifiers, then for n^2 pairs you do partions, then you learn fz representation, then cluster the target data all while tuning and looking for fz dimension and number of clusters. \n-There are hyperparameters that need to be tuned: fz dimension, number of clusters. Tuning all the hyperparameters using limited target data - how does it work? Do you tune the fz dimension (i could not understand it from text, i think u do tune the num of clusters but not sure about fz dimension)\n\n\n\nAdditional comments:\n- For figure 1, source task environments actually look the same to me tbh: the red is mostly correlated with 0 and green is mostly correlated with 1 for both of the environments. If they were flipped between the environments,i would expect Domain invariant representation learning to be able to filter out the color from the embedding layer\n- what if target is also a mix of different environments (you seem to assume that it is all comes from one domain/env \n- How practical is experimental setup you are testing on? For your experiments train was coming from 2 environments that were created artificially\n\nMinor: then we uses=>then we use", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper is concerned with learning a model in situations when some features have spurious correlations with the label (for example, for classifying sheep vs camel, background can be a spurious feature). The idea is that if a lot of source data is available, these spurious correlation features should be easy to find. To find the unstable features, authors hypothesize that they are related to mistakes that a classifier makes in different environments. Therefore, they learn a model on one environment and run it on a different environment (source data), splitting the data into correctly predicted and not. Then for each partition (correct or not) examples with the same labels are encouraged to be close together, so the embedding learns UNSTABLE features (fz). Then the target data is clustered based on unstable features representation fz, and then DRO (existing method that assumes the existance of correct groupping based on unstable features) is used to train a robust target classifier", "main_review": "In general, the method seems to work and it seems to be justified (both theoretically, albeit in the appendix, and practically). It is really complicated though (involves many steps) and i fail to make a connection between this and say Domain Invariant Representation learning. For example, if i was to solve the problem of domain generalization (the setup seems really similar to me), a reasonable baseline would be to train a source model that learns an invariantt representation between various environments (using DANN, or CMD or MMD or whatever). But this is not in the baselines. Why is that? One difference i \"kinda\" see is that domains are more or less pure - eg domain is really a different background. The way i read it is that environement is more mixed - it can contain multiple backgrounds. It would be nice to draw comparison between your method and DIRL or at least explain why DIRL is not applicable\n\n\nPros:\n- The paper is well written and easy to follow (albeit a lot of helpful info is also in the Appendix)\n- The experimental results seem convincing, there is a study of dependence of (some) hyperparameters on the end result (num of clusters)\n\nCons:\n- DRO needs at least a brief introduction\n- Experiments are only on 2 environements (domains) in train\n- The method seems extremely costly: for a number n of source environements, you train n classifiers, then for n^2 pairs you do partions, then you learn fz representation, then cluster the target data all while tuning and looking for fz dimension and number of clusters. \n-There are hyperparameters that need to be tuned: fz dimension, number of clusters. Tuning all the hyperparameters using limited target data - how does it work? Do you tune the fz dimension (i could not understand it from text, i think u do tune the num of clusters but not sure about fz dimension)\n\n\n\nAdditional comments:\n- For figure 1, source task environments actually look the same to me tbh: the red is mostly correlated with 0 and green is mostly correlated with 1 for both of the environments. If they were flipped between the environments,i would expect Domain invariant representation learning to be able to filter out the color from the embedding layer\n- what if target is also a mix of different environments (you seem to assume that it is all comes from one domain/env \n- How practical is experimental setup you are testing on? For your experiments train was coming from 2 environments that were created artificially\n\nMinor: then we uses=>then we use", "summary_of_the_review": "Please see main review", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635801569847}, {"id": "a-rmsUIXJ9t", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3939/Reviewer_RJhJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers a transfer problem when the spurious correlations of source tasks can be applied to the target task. The authors propose to identify the unstable features on source tasks first and then cluster the target data according to these features. Finally, an invariance-based method are incorporated to eliminated the influence of unstable features.", "review_text": "Strength:\n(1) The authors address a critical point that prevents models from generalization, namely spurious correlation.\n(2) The whole pipeline is intuitive and easy to follow, and the empirical results are within expectation.\n\nWeakness:\n(1) A critical limitation of this work is the strong assumption on the transferability of the unstable feature. IMO, such an assumption is restrictive in most settings. For the intra-task transfer (where the classification tasks are the same between source and target), finding the invariant part via invariance learning methods is more realistic. As for the inter-task transfer (a wilder setting),  it is very rare that the unstable features and the way they correlate the outcome are exactly the same across the tasks.\n\n(2) The related work and baselines are not discussed thoroughly. First of all, more transfer learning methods should be discussed, not just REUSE and FINTUNE. Moreover, several automatic de-biasing methods [1,2,3] proposed recently could also be considered to involve.\n\n[1] Qiao, F., Zhao, L., & Peng, X. (2020). Learning to Learn Single Domain Generalization. In CVPR. \n[2] Matsuura, T. and Harada, T., 2020, April. Domain generalization using a mixture of multiple latent domains. AAAI\n[3] J. Liu, Z. Hu, P. Cui, B. Li, and Z. Shen, “Heterogeneous risk minimization”, ICML 2021.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers a transfer problem when the spurious correlations of source tasks can be applied to the target task. The authors propose to identify the unstable features on source tasks first and then cluster the target data according to these features. Finally, an invariance-based method are incorporated to eliminated the influence of unstable features.", "main_review": "Strength:\n(1) The authors address a critical point that prevents models from generalization, namely spurious correlation.\n(2) The whole pipeline is intuitive and easy to follow, and the empirical results are within expectation.\n\nWeakness:\n(1) A critical limitation of this work is the strong assumption on the transferability of the unstable feature. IMO, such an assumption is restrictive in most settings. For the intra-task transfer (where the classification tasks are the same between source and target), finding the invariant part via invariance learning methods is more realistic. As for the inter-task transfer (a wilder setting),  it is very rare that the unstable features and the way they correlate the outcome are exactly the same across the tasks.\n\n(2) The related work and baselines are not discussed thoroughly. First of all, more transfer learning methods should be discussed, not just REUSE and FINTUNE. Moreover, several automatic de-biasing methods [1,2,3] proposed recently could also be considered to involve.\n\n[1] Qiao, F., Zhao, L., & Peng, X. (2020). Learning to Learn Single Domain Generalization. In CVPR. \n[2] Matsuura, T. and Harada, T., 2020, April. Domain generalization using a mixture of multiple latent domains. AAAI\n[3] J. Liu, Z. Hu, P. Cui, B. Li, and Z. Shen, “Heterogeneous risk minimization”, ICML 2021.", "summary_of_the_review": "This paper addresses the spurious correlation by transferring knowledge from source tasks. Although intuitions are provided and empirical effectiveness is illustrated accordingly, the method is restrictive due to the strong assumption on the transferability of unstable features. Several important related work and baselines are also missing.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635749967363}, {"id": "GU8TviZsIq3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3939/Reviewer_tPxR"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper aims to transfer the knowledge of spurious correlations in a set of source environments to a target environment. They assume the degree of spurious patterns vary among source environments (e.g. color has different correlations with the number in MNIST), and the spurious pattern are the same in both the source and target tasks. Then they aim to transfer the spurious knowledge from source environments to learn a classifier in the target task that ignores the spurious pattern. \n\nThey train the model in 3 stages. First, they train a regular classifier in each environment. E.g. given two source environments E1 and E2, they train the corresponding classifiers f1, f2. Second, they use the error of the f1 on E2 to \"separate\" the E2 examples into two groups, one group for correct predictions and the other group for the wrong predictions under each class. They assume the cause of the error comes from the spurious patterns, so each group will correspond to either high or low spurious patterns learned in f1. Thus they learn a f_Z that outputs an embedding to separate these two groups by triplet loss. Finally, in the target task, they cluster the examples into different groups by the similarity of f_Z that captures examples with similar spurious patterns. Then they apply Distribution Robust Optimization (DRO) on these groups that optimizes the worst performance across all groups. This ignores the spurious patterns based on f_Z in the target task and learns a stable classifier.", "review_text": "Overall I like this paper. This paper does a good job of explaining the setup and the writings are mostly clear. This paper builds upon the core idea from Bao et al. 2021 that uses the error of the model in different environments to capture the spurious patterns, and it further extends to the transfer learning setup, which I believe it's a new setup.\n\nMy concerns are as followed:\n1. The assumptions are too strong in my opinions, and I think a discussion section will help readers understand the limitations:\n - A) I am skeptical about the assumption in Bao et al. 2021 that uses the error of the model across environments to capture the spuriousness. There could be many factors behind errors across environments such as mislabeled data, different viewpoints of the image (e.g. rotations), noise, and data distribution shifts. None of it corresponds to the unstable features. If this assumption breaks, then there is no real spurious pattern to be learned here.\n- B) On the data side, this work assumes the source tasks need to have varying degrees of spuriousness and the target task has the same spurious patterns. How do we possibly check these assumptions are real in the real-world setting?\n\n2. Regarding the demonstrated results\n- A) There could be different spurious factors across different classes. And the binarization procedure of correct or incorrect predictions could lead to underspecification since there might be multiple spurious factors but we only treat them as two. Will this method break if there are multiple spurious patterns?\n- B) The multi-stage approach of training could be brittle, and selecting hyperparameters could be difficult since no spurious correlation is known beforehand. Maybe the authors can show if this approach is robust when the assumptions are mildly violated like there is a distribution shift among environments, or when there are multiple spurious patterns in the data, and illustrate how to select hyperparameters.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to transfer the knowledge of spurious correlations in a set of source environments to a target environment. They assume the degree of spurious patterns vary among source environments (e.g. color has different correlations with the number in MNIST), and the spurious pattern are the same in both the source and target tasks. Then they aim to transfer the spurious knowledge from source environments to learn a classifier in the target task that ignores the spurious pattern. \n\nThey train the model in 3 stages. First, they train a regular classifier in each environment. E.g. given two source environments E1 and E2, they train the corresponding classifiers f1, f2. Second, they use the error of the f1 on E2 to \"separate\" the E2 examples into two groups, one group for correct predictions and the other group for the wrong predictions under each class. They assume the cause of the error comes from the spurious patterns, so each group will correspond to either high or low spurious patterns learned in f1. Thus they learn a f_Z that outputs an embedding to separate these two groups by triplet loss. Finally, in the target task, they cluster the examples into different groups by the similarity of f_Z that captures examples with similar spurious patterns. Then they apply Distribution Robust Optimization (DRO) on these groups that optimizes the worst performance across all groups. This ignores the spurious patterns based on f_Z in the target task and learns a stable classifier.", "main_review": "Overall I like this paper. This paper does a good job of explaining the setup and the writings are mostly clear. This paper builds upon the core idea from Bao et al. 2021 that uses the error of the model in different environments to capture the spurious patterns, and it further extends to the transfer learning setup, which I believe it's a new setup.\n\nMy concerns are as followed:\n1. The assumptions are too strong in my opinions, and I think a discussion section will help readers understand the limitations:\n - A) I am skeptical about the assumption in Bao et al. 2021 that uses the error of the model across environments to capture the spuriousness. There could be many factors behind errors across environments such as mislabeled data, different viewpoints of the image (e.g. rotations), noise, and data distribution shifts. None of it corresponds to the unstable features. If this assumption breaks, then there is no real spurious pattern to be learned here.\n- B) On the data side, this work assumes the source tasks need to have varying degrees of spuriousness and the target task has the same spurious patterns. How do we possibly check these assumptions are real in the real-world setting?\n\n2. Regarding the demonstrated results\n- A) There could be different spurious factors across different classes. And the binarization procedure of correct or incorrect predictions could lead to underspecification since there might be multiple spurious factors but we only treat them as two. Will this method break if there are multiple spurious patterns?\n- B) The multi-stage approach of training could be brittle, and selecting hyperparameters could be difficult since no spurious correlation is known beforehand. Maybe the authors can show if this approach is robust when the assumptions are mildly violated like there is a distribution shift among environments, or when there are multiple spurious patterns in the data, and illustrate how to select hyperparameters.", "summary_of_the_review": "Pros:\n+ A new setup that transfers the knowledge of spurious patterns to a target task.\n+ The writing is clear. Figures and tables are beautifully produced.\n+ The results are demonstrated in multiple datasets and the number is convincing compared to several recent baselines. The ablation study of group numbers is great.\n\nCons:\n- The hypotheses in my opinion are a bit too strong. It's unclear in what real-world settings this method will work unless in the contrived setup in this paper. Including a limitation in the paper can strengthen the paper.\n- The multi-stage approach of training could be brittle.\n\nI am leaning towards the borderline with marginal acceptance. The experiments are complete with multiple baselines, clear writing and a good ablation study. I feel the assumption is too strong which would need some justifications. And if the authors can show a robust study under different mildly violated assumptions could further strengthen this paper.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635544631094}], "openreview_url": "https://openreview.net/forum?id=xs-tJn58XKv", "arxiv_id": "2106.07847", "paper_pdf": "papers/xs-tJn58XKv.pdf", "paper_pdf_sha256": "83267c3304738c0a79a8b40b8535124cc8a971a4ce0deeb935c898fb72f4f20f", "paper_pdf_bytes": 4384556, "paper_pdf_source": "openreview", "code_url": "https://github.com/YujiaBao/tofu", "code_repository": "YujiaBao/tofu", "code_commit": "94d38031300c80764ee921fc2bfb02040f5459dd", "code_archive": "repos/xs-tJn58XKv.zip", "code_archive_sha256": "b53f65d51dfa703e1738fde1bafd1ba66f099c274788c4a8b2ef30d8d4b3fb7d", "code_archive_bytes": 1186397, "code_file_count": 25, "code_extensions": {".py": 19, ".sh": 6}, "github_disk_usage_kb": 1154, "github_languages": {"Python": 73879, "Shell": 1901}, "github_archived": false, "github_pushed_at": "2022-07-24T22:13:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-stable-classifiers-by-transferring"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7K0UUL9y9lE", "year": 2021, "status": "rejected", "title": "You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling", "authors": ["Zhanpeng Zeng", "Yunyang Xiong", "Sathya N. Ravi", "Shailesh Acharya", "Glenn Fung", "Vikas Singh"], "authorids": ["~Zhanpeng_Zeng1", "~Yunyang_Xiong2", "~Sathya_N._Ravi1", "sachary1@amfam.com", "~Glenn_Fung2", "~Vikas_Singh1"], "authors_source": "OpenReview API", "abstract": "Transformer-based models have come to dominate the landscape in a wide range of natural language processing (NLP) applications. The heart of the transformer model is the self-attention mechanism, which captures the interactions of token pairs in the input sequences and consequently, depends quadratically on the input sequence length. It is known that training such models on longer sequences is quite expensive, and often, prohibitively so. We show that a Bernoulli sampling attention mechanism based on Locality Sensitive Hashing (LSH), decreases the quadratic complexity to linear. We bypass the quadratic cost by considering self-attention as a sum of individual tokens associated with Bernoulli random variables that can, in principle, be sampled at once by a single hash (although in practice, this number may be a small constant). This leads to an efficient sampling scheme to estimate self-attention which relies on specific modifications of LSH (based on feasibility of deployment on GPU architectures). We evaluate our proposed algorithm on the GLUE benchmark with standard 512 sequence length and our method achieves comparable or even slightly better performance than a standard pretrained Transformer. To evaluate whether our method can indeed handle longer sequences, we conduct experiments on long sequence (4096) language model pretraining and achieve consistent results as standard self-attention, while observing sizable inference speed-ups and memory savings.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "0WJ9QAJTEUT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2731/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This article presents YOSO, an locality sensitive sampling based attention mechanism for large scale language modeling. \n\n\n\nStrength:\nA new idea of applying  locality sensitive sampling to approximate attention matrix in the transformer\n\n\nWeakness:\n1. Comparison of Complexity: [1] presents the complexity of different efficient transformers. For linformer[2], the time and memory complexity is O(nk). Is there any justification of LSH sampling equipped YOSO with complexity more than O(nm\\tau log(d)+nmd)?\n2.Experiments: YOSO takes linformer as baselines. However, the pre-training experiment part does not provide steps vs ppl of linformer with YOSO in Figure 4. What is the comparison result of YOSO with linformer on iteration wise convergence? Also, linformer demonstrates better accuracy in downstream tasks such as SST-2. Is there any comparison to an explanation that can analyze this difference in performance?\n3.Efficiency: YOSO demonstrates an advantage over linformer and longformer in memory and runtime. However, is there any analysis on why YOSO achieves this superiority with higher complexities? Are there any system-level advantages that YOSO can show?\n\nSome discussions: \nReformer[3] design an attention mechanism that computations are held in the neighbor tokens inside the hash buckets. YOSO also uses hash based sampling to compute attention via neighbor tokens that have high collision probability. On the other hand, linformer introduces a more global view for attention by the low rank projection. Is there any analysis of the local vs global intuition?\n\n[1]Efficient Transformers: A Survey https://arxiv.org/pdf/2009.06732.pdf\n\n[2]Linformer: Self-Attention with Linear Complexity https://arxiv.org/abs/2006.04768\n\n[3] Reformer: The Efficient Transformer https://arxiv.org/abs/2001.04451", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Attention by LSH sampling", "review": "This article presents YOSO, an locality sensitive sampling based attention mechanism for large scale language modeling. \n\n\n\nStrength:\nA new idea of applying  locality sensitive sampling to approximate attention matrix in the transformer\n\n\nWeakness:\n1. Comparison of Complexity: [1] presents the complexity of different efficient transformers. For linformer[2], the time and memory complexity is O(nk). Is there any justification of LSH sampling equipped YOSO with complexity more than O(nm\\tau log(d)+nmd)?\n2.Experiments: YOSO takes linformer as baselines. However, the pre-training experiment part does not provide steps vs ppl of linformer with YOSO in Figure 4. What is the comparison result of YOSO with linformer on iteration wise convergence? Also, linformer demonstrates better accuracy in downstream tasks such as SST-2. Is there any comparison to an explanation that can analyze this difference in performance?\n3.Efficiency: YOSO demonstrates an advantage over linformer and longformer in memory and runtime. However, is there any analysis on why YOSO achieves this superiority with higher complexities? Are there any system-level advantages that YOSO can show?\n\nSome discussions: \nReformer[3] design an attention mechanism that computations are held in the neighbor tokens inside the hash buckets. YOSO also uses hash based sampling to compute attention via neighbor tokens that have high collision probability. On the other hand, linformer introduces a more global view for attention by the low rank projection. Is there any analysis of the local vs global intuition?\n\n[1]Efficient Transformers: A Survey https://arxiv.org/pdf/2009.06732.pdf\n\n[2]Linformer: Self-Attention with Linear Complexity https://arxiv.org/abs/2006.04768\n\n[3] Reformer: The Efficient Transformer https://arxiv.org/abs/2001.04451", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603932578291}, {"id": "VTl7JKi4-75", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2731/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tries to improve the efficiency of (multi-head) self-attention by reducing the computational complexity from a quadratic one to a linear one. The authors propose to use Bernoulli sampling to approximate the self-attention's softmax distribution through importance sampling via LSH, which makes a linear cost self-attention possible. \n\nPros:\n1. This paper provides a thorough solution of how to accelerate self-attention via an approximation by Bernoulli sampling and LSH. \n2. The experiments show that the proposed approach could achieve considerable speedup while preserving model performance. \n\nCons:\n1. The authors should give a more direct comparison with an important related work the reformer, which also uses LSH to speedup the computation of attention. \n2. The experiments were mainly conducted on MLM on GLUE, which lacks generalization among tasks. Adding more tasks such as MT or autoregressive/causal LM would make the experimental part more solid and convincing. \n\n\n\n---------\nMinors:\n- Fig 2, 3 and Tab 1 are not cross-refed in the main body\n- Format of citation: seems all of the citations are of this format - authors (year), which is not correct when citations do not act as subjective of objective in the sentence. Please check the format guideline.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper tries to improve the efficiency of (multi-head) self-attention by reducing the computational complexity from a quadratic one to a linear one. The authors propose to use Bernoulli sampling to approximate the self-attention's softmax distribution through importance sampling via LSH, which makes a linear cost self-attention possible. \n\nPros:\n1. This paper provides a thorough solution of how to accelerate self-attention via an approximation by Bernoulli sampling and LSH. \n2. The experiments show that the proposed approach could achieve considerable speedup while preserving model performance. \n\nCons:\n1. The authors should give a more direct comparison with an important related work the reformer, which also uses LSH to speedup the computation of attention. \n2. The experiments were mainly conducted on MLM on GLUE, which lacks generalization among tasks. Adding more tasks such as MT or autoregressive/causal LM would make the experimental part more solid and convincing. \n\n\n\n---------\nMinors:\n- Fig 2, 3 and Tab 1 are not cross-refed in the main body\n- Format of citation: seems all of the citations are of this format - authors (year), which is not correct when citations do not act as subjective of objective in the sentence. Please check the format guideline.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603898061409}, {"id": "eC-sAowKE_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2731/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary\n\nThe paper proposes to replace the weighted average of the values in standard self-attention with the average of values sampled in a way that the expectation is close to the result of self-attention. In particular, the authors associate with each query-key pair, a Bernoulli random variable with expected value close to the exponential of the dot-product. Sampling these variables and averaging the values per query is formulated in an efficient way using locality-sensitive hashing.\n\n### Strengths\n\n- Using sampling to approximate self-attention is novel and promising.\n- The LSH formulation where the values are averaged in the bucket is a clever way to avoid the pairwise interactions between queries and keys.\n- Training with hashing and evaluating using the expectation is interesting and provides evidence for the approximation quality of YOSO.\n\n### Weaknesses\n\n1. In the evaluation there is never an explicit comparison with respect to both time and performance. I appreciate that given a large enough sequence length YOSO will always be faster but will it be good enough?\n\n2. One of the most important parts of the methodology, the gradient computation, is the least clearly written. For instance, equation 11 contains $\\nabla_{Attn}$ which is never defined and subsequent equations contain $\\nabla_{YOSO}$ which seems to contradict equation 11. Moreover, how is equation 11 derived? Is it the gradient of the expectation?\n\n3. In table 3, the performance is only measured with respect to inference. Given that the most computationally intensive part of the method is the backward pass, a comparison with respect to wall-clock time per epoch, as well as total training time would be very informative.\n\n### Reasons for recommendation\n\nI find the idea very elegant and interesting however, the experimental section is somewhat lacking. There is no clear evaluation of the trade-off between speed and performance. The MLM task, although significant and demanding, contains sequences of small length, otherwise why not show a graph of performance vs inference-time.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Elegant idea and formulation but somewhat lacking evaluation", "review": "### Summary\n\nThe paper proposes to replace the weighted average of the values in standard self-attention with the average of values sampled in a way that the expectation is close to the result of self-attention. In particular, the authors associate with each query-key pair, a Bernoulli random variable with expected value close to the exponential of the dot-product. Sampling these variables and averaging the values per query is formulated in an efficient way using locality-sensitive hashing.\n\n### Strengths\n\n- Using sampling to approximate self-attention is novel and promising.\n- The LSH formulation where the values are averaged in the bucket is a clever way to avoid the pairwise interactions between queries and keys.\n- Training with hashing and evaluating using the expectation is interesting and provides evidence for the approximation quality of YOSO.\n\n### Weaknesses\n\n1. In the evaluation there is never an explicit comparison with respect to both time and performance. I appreciate that given a large enough sequence length YOSO will always be faster but will it be good enough?\n\n2. One of the most important parts of the methodology, the gradient computation, is the least clearly written. For instance, equation 11 contains $\\nabla_{Attn}$ which is never defined and subsequent equations contain $\\nabla_{YOSO}$ which seems to contradict equation 11. Moreover, how is equation 11 derived? Is it the gradient of the expectation?\n\n3. In table 3, the performance is only measured with respect to inference. Given that the most computationally intensive part of the method is the backward pass, a comparison with respect to wall-clock time per epoch, as well as total training time would be very informative.\n\n### Reasons for recommendation\n\nI find the idea very elegant and interesting however, the experimental section is somewhat lacking. There is no clear evaluation of the trade-off between speed and performance. The MLM task, although significant and demanding, contains sequences of small length, otherwise why not show a graph of performance vs inference-time.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603892945493}, {"id": "GdNUZpUxH9L", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2731/AnonReviewer3"], "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a linear-time attention model based on importance sampling and locality sensitive hashing. The idea is to use Bernoulli sampling to approximate the self-attention - which is quadratic in complexity. \n\nHonestly, this is a very crowded space and already many models exist (https://arxiv.org/abs/2009.06732). The authors are aware of these works, cite them and yet there is no comparison. \n\nThis method seems to be rooted in LSH and a very natural question is how does compare to Reformers. There is not even a sparse transformer or local attention baseline in the experiments. This raises questions about whether this paper will even make any impact at all. (comparisons with longformer is done only on speed/memory but not qualitatively). why? \n\nI also find other flaws with the model. If sampling is used, this essentially makes the model stochastic (correct me if Im wrong here). but there are undesirable properties of this such as having non-deterministic inference.\n\nAnother flaw is that method potentially introduces a lot of instability in training. I think the authors could comment a little on this. Transformers are already notoriously difficult to train and I figure that this method would probably make it way harder for practitioners to get the hyperparameters correct. I think playing with the scaling and normalization of the self-attention weights is something non-ideal, and unless the authors can show this is reasonable stable i am not convinced. \n\nI also find it difficult to understand the choices of tasks. It seems like GLUE benchmark is used, yet most of the tasks (like SST) are relatively shorter sequences. I think the authors need to explore datasets that showcase the model's ability on longer sequences. Artificially raising sequence len during pretraining is not really sufficient to be convincing that the model is doing something useful for longer sequences (since the masked out tokens really depend on local context). \n\nMy constructive feedback to the authors to improve the paper is to have reasonable baselines for comparison. The datasets are also not appropriate. I would suggest some actually long-range tasks in order to showcase the model's capabilities. \n\nAt the rate of the number of new models that tackle this problem, I suspect it would be wise to wrap up your sleeves and add actual efficient transformer baselines.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Lacking impact and convincing experiments, unnecessary model ", "review": "This paper presents a linear-time attention model based on importance sampling and locality sensitive hashing. The idea is to use Bernoulli sampling to approximate the self-attention - which is quadratic in complexity. \n\nHonestly, this is a very crowded space and already many models exist (https://arxiv.org/abs/2009.06732). The authors are aware of these works, cite them and yet there is no comparison. \n\nThis method seems to be rooted in LSH and a very natural question is how does compare to Reformers. There is not even a sparse transformer or local attention baseline in the experiments. This raises questions about whether this paper will even make any impact at all. (comparisons with longformer is done only on speed/memory but not qualitatively). why? \n\nI also find other flaws with the model. If sampling is used, this essentially makes the model stochastic (correct me if Im wrong here). but there are undesirable properties of this such as having non-deterministic inference.\n\nAnother flaw is that method potentially introduces a lot of instability in training. I think the authors could comment a little on this. Transformers are already notoriously difficult to train and I figure that this method would probably make it way harder for practitioners to get the hyperparameters correct. I think playing with the scaling and normalization of the self-attention weights is something non-ideal, and unless the authors can show this is reasonable stable i am not convinced. \n\nI also find it difficult to understand the choices of tasks. It seems like GLUE benchmark is used, yet most of the tasks (like SST) are relatively shorter sequences. I think the authors need to explore datasets that showcase the model's ability on longer sequences. Artificially raising sequence len during pretraining is not really sufficient to be convincing that the model is doing something useful for longer sequences (since the masked out tokens really depend on local context). \n\nMy constructive feedback to the authors to improve the paper is to have reasonable baselines for comparison. The datasets are also not appropriate. I would suggest some actually long-range tasks in order to showcase the model's capabilities. \n\nAt the rate of the number of new models that tackle this problem, I suspect it would be wise to wrap up your sleeves and add actual efficient transformer baselines.\n", "rating": "2: Strong rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603826791104}], "openreview_url": "https://openreview.net/forum?id=7K0UUL9y9lE", "arxiv_id": "2111.09714", "paper_pdf": "papers/7K0UUL9y9lE.pdf", "paper_pdf_sha256": "46c61e2b234128f89e3617dce9eb90b4a7d81e64f3090f4cbb2725407b886bed", "paper_pdf_bytes": 2251447, "paper_pdf_source": "openreview", "code_url": "https://github.com/mlpen/YOSO", "code_repository": "mlpen/YOSO", "code_commit": "6000487d9cd8e34519aa47650011f458ca0db64c", "code_archive": "repos/7K0UUL9y9lE.zip", "code_archive_sha256": "09db5c010e79e83a2185b31817876897f144428f5efeb3ce48edb3ebb04c0883", "code_archive_bytes": 1102033, "code_file_count": 69, "code_extensions": {".py": 38, ".h": 13, ".cu": 10, ".cpp": 5, ".sh": 3}, "github_disk_usage_kb": 1153, "github_languages": {"Python": 190445, "Cuda": 101422, "C": 13489, "C++": 11218, "Shell": 2922}, "github_archived": false, "github_pushed_at": "2021-07-01T17:17:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/you-only-sample-almost-once-linear-cost-self-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1xRnxSYwS", "year": 2020, "status": "rejected", "title": "Goten: GPU-Outsourcing Trusted Execution of Neural Network Training and Prediction", "authors": ["Lucien K.L. Ng", "Sherman S.M. Chow", "Anna P.Y. Woo", "Donald P. H. Wong", "Yongjun Zhao"], "authorids": ["nkl018@ie.cuhk.edu.hk", "smchow@ie.cuhk.edu.hk", "woopuiyung@gmail.com", "foreverjun.zhao@gmail.com"], "authors_source": "OpenReview API", "abstract": "Before we saw worldwide collaborative efforts in training machine-learning models or widespread deployments of prediction-as-a-service, we need to devise an efficient privacy-preserving mechanism which guarantees the privacy of all stakeholders (data contributors, model owner, and queriers). Slaom (ICLR ’19) preserves privacy only for prediction by leveraging both trusted environment (e.g., Intel SGX) and untrusted GPU. The challenges for enabling private training are explicitly left open – its pre-computation technique does not hide the model weights and fails to support dynamic quantization corresponding to the large changes in weight magnitudes during training. Moreover, it is not a truly outsourcing solution since (offline) pre-computation for a job takes as much time as computing the job locally by SGX, i.e., it only works before all pre-computations are exhausted.\n\nWe propose Goten, a privacy-preserving framework supporting both training and prediction. We tackle all the above challenges by proposing a secure outsourcing protocol which 1) supports dynamic quantization, 2) hides the model weight from GPU, and 3) performs better than a pure-SGX solution even if we perform the precomputation online. Our solution leverages a non-colluding assumption which is often employed by cryptographic solutions aiming for practical efficiency (IEEE SP ’13, Usenix Security ’17, PoPETs ’19). We use three servers, which can be reduced to two if the pre-computation is done offline. Furthermore, we implement our tailor-made memory-aware measures for minimizing the overhead when the SGX memory limit is exceeded (cf., EuroSys ’17, Usenix ATC ’19). Compared to a pure-SGX solution, our experiments show that Goten can speed up linear-layer computations in VGG up to 40×, and overall speed up by 8.64× on VGG11.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "BkehxucWcH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2556/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method for privacy-preserving training and evaluation of DNNs. The method is based on a combination of hardware support from a trusted execution enclave (Intel SGX) and an algorithm for offloading intensive computation to unsecure GPU devices and communicating with the trusted environment without losing security guarantees during communication.  Compared to related work on a similar system (Slalom), the proposed system enables secure training in addition to inference.  The approach is based on the use of additive secret sharing to relegate chunks of computation to independent GPU servers.\n\nThe evaluation presents experiments that report timings of the proposed system against a baseline. Throughput (images/second) improvements of 8-9x are reported, but the contrast point is unclear (it appears to be CaffeSCONE, but there are several unclear points, summarized below). In addition, a set of results reporting a speed up ratio against attained accuracy of a trained VGG11 network, and a set of results reporting speed up against arithmetic intensity of the workload are given.\n\nI lean towards rejection of this draft, as it has several weaknesses:\n- The connection between the evaluation (which mostly focuses on the speed benefits) and the claimed contributions is tenuous at best. This issue is further compounded by clarity issues in the experiments and their description\n- The empirical results are unclear due to differences between simulation of SGX capability vs hardware support of SGX capability. It is not clear what part of the results is influenced significantly by this disparity, and more importantly whether all the comparisons are done in an equal footing (for example the reported results comparing CaffeSCONE with Goten are performed in two different regimes). As a byproduct, there is a confusing \"scaling factor\" described by the authors that is applied to the timings.\n- A brief mention is made of the fact that the proposed system does not in fact provide correctness guarantees (unlike CaffeSCONE), but this is dismissed by reference to utilizing the same trick used by Slalom.  However, this trick is not described or motivated.\n- The writing in the current draft is of relatively low quality, significantly impacting the readability of the paper and making it hard to understand the contributions and whether they are backed by the presented results.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes a method for privacy-preserving training and evaluation of DNNs. The method is based on a combination of hardware support from a trusted execution enclave (Intel SGX) and an algorithm for offloading intensive computation to unsecure GPU devices and communicating with the trusted environment without losing security guarantees during communication.  Compared to related work on a similar system (Slalom), the proposed system enables secure training in addition to inference.  The approach is based on the use of additive secret sharing to relegate chunks of computation to independent GPU servers.\n\nThe evaluation presents experiments that report timings of the proposed system against a baseline. Throughput (images/second) improvements of 8-9x are reported, but the contrast point is unclear (it appears to be CaffeSCONE, but there are several unclear points, summarized below). In addition, a set of results reporting a speed up ratio against attained accuracy of a trained VGG11 network, and a set of results reporting speed up against arithmetic intensity of the workload are given.\n\nI lean towards rejection of this draft, as it has several weaknesses:\n- The connection between the evaluation (which mostly focuses on the speed benefits) and the claimed contributions is tenuous at best. This issue is further compounded by clarity issues in the experiments and their description\n- The empirical results are unclear due to differences between simulation of SGX capability vs hardware support of SGX capability. It is not clear what part of the results is influenced significantly by this disparity, and more importantly whether all the comparisons are done in an equal footing (for example the reported results comparing CaffeSCONE with Goten are performed in two different regimes). As a byproduct, there is a confusing \"scaling factor\" described by the authors that is applied to the timings.\n- A brief mention is made of the fact that the proposed system does not in fact provide correctness guarantees (unlike CaffeSCONE), but this is dismissed by reference to utilizing the same trick used by Slalom.  However, this trick is not described or motivated.\n- The writing in the current draft is of relatively low quality, significantly impacting the readability of the paper and making it hard to understand the contributions and whether they are backed by the presented results.\n"}, "tcdate": 1572083699632}, {"id": "HkguQA3BYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2556/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper builds a privacy-preserving training framework within a Trusted Execution Environment (TEE) such as Intel SGX. The work is heavily inspired from Slalom, which does privacy-preserving inference in TEEs. The main drawbacks of Slalom when extending to training are (1) weight quantization needs to be dynamics as they change during training, and (2) pre-processing step of Slalom to compare u = f(r) isn't effective as the weights change, and running this within TEE is no better than running the full DNN within TEE. In addition, Goten also makes the weights private as opposed to Slalom. Overall, this is a very important contribution towards privacy preserving training and the paper takes a strong practical and implementation-focused approach by considering issues arising due to memory limitations in TEE and the performance implications of default Linux paging.\n\nThe paper comes up with a novel outsourcing protocol with two non-colluding servers for offloading linear operations in a fully privacy-preserving way and does detailed analysis of the performance implications. Similar to a lot of other methods for training with quantization, the weights are stored and updated in floats while the computation is performed using quantized values. The experimental results suggest a strong improvement over the CaffeSCONE baseline. One drawback with experiments is the lack of comparison with Slalom for inference if Goten is assumed to be a framework for both training and prediction in a privacy-preserving way.\n\nAnother downside of the paper is that a few sections could be improved with their explanation, and there is quite a bit of redundancy in going over the downsides of Slalom and why it can't be used for secure training. For instance,\n- Section 1.1: \"Our results (referring to Section 4.2) show that CaffeSCONE’s performance greatly suffer from the enclave’s memory limit as it needs an inefficient mechanism to handle excessive use of memory not affordable by the enclave\". Here, it's not clear which mechanism is inefficient. Are we talking about mechanisms in CaffeSCONE for reducing memory usage while training and if so, are they somehow inefficient? Or does it mean to imply that we can't train a DNN fully within an enclave due to memory limits?\n-  Last paragraph of section 2.2 is unclear. \"CaffeSCONE guarantees the correctness of both training and prediction. Goten does not provide it as we present it due to page limitation, but we can resort to the trick used by Slalom\". What does the last sentence mean? Does Goten guarantee correctness during training and prediction or not? And what trick from Slalom are we referring to? The blinding trick used for privacy or the Freivalds' algorithm used for correctness?\n\nOverall a strong contribution with supporting experimental results, but the certain parts need further explanation or rewriting for higher rating.\n\nPros:\n- An important contribution in the direction of fully private DNN training and inference within a TEE. Draws inspirations from Slalom and mainly addresses the challenges left to extend the approach to training.\n- Motivation and reasons for why Slalom can't be used for training is very well laid out.\n- In addition to input and output activations, Goten also preserves the privacy of the weights.\n- Good baseline for comparison using CaffeSCONE.\n- Implementation factors considered and analyzed such as tricks as using SGX-aware paging instead of naive Linux paging.\n- Strong experiments and benchmarks\n\nCons:\n- Some sections are not explained well and unclear as mentioned earlier.\n- How does the inference performance of Goten compare to Slalom given the same privacy and correctness guarantees? This isn't clear from the experiments section.\n\nMinor comments:\n- \"Slalom\" is mis-spelt in line 4 of the abstract.\n- There appear to be typos and grammatical errors at many places in the paper. Further proof-reading might be helpful.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper builds a privacy-preserving training framework within a Trusted Execution Environment (TEE) such as Intel SGX. The work is heavily inspired from Slalom, which does privacy-preserving inference in TEEs. The main drawbacks of Slalom when extending to training are (1) weight quantization needs to be dynamics as they change during training, and (2) pre-processing step of Slalom to compare u = f(r) isn't effective as the weights change, and running this within TEE is no better than running the full DNN within TEE. In addition, Goten also makes the weights private as opposed to Slalom. Overall, this is a very important contribution towards privacy preserving training and the paper takes a strong practical and implementation-focused approach by considering issues arising due to memory limitations in TEE and the performance implications of default Linux paging.\n\nThe paper comes up with a novel outsourcing protocol with two non-colluding servers for offloading linear operations in a fully privacy-preserving way and does detailed analysis of the performance implications. Similar to a lot of other methods for training with quantization, the weights are stored and updated in floats while the computation is performed using quantized values. The experimental results suggest a strong improvement over the CaffeSCONE baseline. One drawback with experiments is the lack of comparison with Slalom for inference if Goten is assumed to be a framework for both training and prediction in a privacy-preserving way.\n\nAnother downside of the paper is that a few sections could be improved with their explanation, and there is quite a bit of redundancy in going over the downsides of Slalom and why it can't be used for secure training. For instance,\n- Section 1.1: \"Our results (referring to Section 4.2) show that CaffeSCONE’s performance greatly suffer from the enclave’s memory limit as it needs an inefficient mechanism to handle excessive use of memory not affordable by the enclave\". Here, it's not clear which mechanism is inefficient. Are we talking about mechanisms in CaffeSCONE for reducing memory usage while training and if so, are they somehow inefficient? Or does it mean to imply that we can't train a DNN fully within an enclave due to memory limits?\n-  Last paragraph of section 2.2 is unclear. \"CaffeSCONE guarantees the correctness of both training and prediction. Goten does not provide it as we present it due to page limitation, but we can resort to the trick used by Slalom\". What does the last sentence mean? Does Goten guarantee correctness during training and prediction or not? And what trick from Slalom are we referring to? The blinding trick used for privacy or the Freivalds' algorithm used for correctness?\n\nOverall a strong contribution with supporting experimental results, but the certain parts need further explanation or rewriting for higher rating.\n\nPros:\n- An important contribution in the direction of fully private DNN training and inference within a TEE. Draws inspirations from Slalom and mainly addresses the challenges left to extend the approach to training.\n- Motivation and reasons for why Slalom can't be used for training is very well laid out.\n- In addition to input and output activations, Goten also preserves the privacy of the weights.\n- Good baseline for comparison using CaffeSCONE.\n- Implementation factors considered and analyzed such as tricks as using SGX-aware paging instead of naive Linux paging.\n- Strong experiments and benchmarks\n\nCons:\n- Some sections are not explained well and unclear as mentioned earlier.\n- How does the inference performance of Goten compare to Slalom given the same privacy and correctness guarantees? This isn't clear from the experiments section.\n\nMinor comments:\n- \"Slalom\" is mis-spelt in line 4 of the abstract.\n- There appear to be typos and grammatical errors at many places in the paper. Further proof-reading might be helpful."}, "tcdate": 1571307039988}, {"id": "rkgoPyQVFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2556/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nSummary\n========\nThis paper proposes a framework for privacy-preserving training of neural networks, by leveraging trusted execution environments and untrusted GPU accelerators.\nThe system builds heavily on the prior Slalom system, and uses standard MPC techniques (three non-colluding servers, multiplication triplets) to extend Slalom's inference-only protocol to privacy-preserving training.\nThis is a valuable and hard to reach goal. Unfortunately, the paper's evaluation fails to deliver on its strong promises, by ignoring the high network communication between the non-colluding servers.\nSpecifically, all experiments were conducted with three servers co-located in a public cloud's LAN. In this setting, it is hard to argue that non-collusion is a valid security assumption as the cloud provider controls all servers (alternatively, if the cloud provider is trusted, then there is no need for any trusted execution or cryptography). If the same experiments were conducted on a WAN, the communication costs would alleviate any savings in computation time.\n\nFor these reasons, I lean strongly towards rejection of this paper. \n\nDetailed comments\n=================\nExtending the ideas in Slalom to support privacy-preserving training is a good research question, and Tramer and Boneh had discussed some of the challenges and limitations towards this in their original paper.\nGetting rid of the pre-processing stage for blinding factors by leveraging non-colluding servers is a well-known trick from the MPC literature, but it does not seem easily applicable here. \nThe problem is that the servers need to communicate an amount of data proportional to the size of each internal layer of the network, for each forward and backward pass. If the servers communicate over a standard WAN, the communication time will be much too high to be competitive with the CaffeScone baseline.\nIn a LAN, as in this paper's experiments, the network latency is low enough for the network costs to be dominated by computation. But this begs the question of whether servers running in a same LAN (e.g., hosted by a single cloud provider) can really be considered non-colluding. In the considered setup, the cloud provider (Google in this case), could just observe the communication between all servers, thereby breaking privacy.\n\nAnother security issue with proposed scheme is the lack of computation integrity. This corresponds to the so-called \"honest-but-curious\" threat model which often appears in the MPC literature, and this should be acknowledged and motivated.\n\nOn the experimental side, the considered baseline, CaffeScone, seems pretty weak. In particular, any optimizations that the authors implement for Goten (e.g., better paging) should also be added to their baseline for a fair comparison.\nThe numbers in Figure 3 show that the baseline could be optimized a lot further.  A gap between hardware/simulation modes of ~6x seems indicative of sub-optimal paging. Even the single-core, simulation mode throughput numbers seem low for CIFAR10.\n\nThe experimental setup is quite confusing. Running the baseline and Goten in different environments (e.g., different CPUs) and then re-normalizing throughputs is somewhat convoluted and prone to mistakes. Why not run all experiments on the same setup?\nSimilarly, converting between results in SGX's hardware and simulation modes is also not very indicative. The authors note (p. 8) that in SGX's simulation mode \"code compilation is almost the same as hardware mode except that the program is not protected by SGX, which is fine for our purpose since the DNN training and prediction algorithms are publicly known\". This is fundamentally incorrect!\nSGX's simulation mode provides absolutely no security guarantees. It simply compiles the code using the SGX libraries and ensures that the enclaved code performs no untrusted operations, but it does not provide any hardware protections whatsoever. In particular, code running in simulation mode will not be affected by the overhead of SGX's paging, as the memory is never encrypted.\nAs a result, performance results in simulation mode are usually not indicative of performance in hardware mode. Trying to convert runtimes from simulation mode to hardware mode by comparing times of specific layers is also prone to many approximation errors. \n\nFinally, I had some trouble understanding the way in which Goten quantization works. Section 3.3. mentions that values are treated as floats, but then mentions the use of 53 bits of precision. Did you mean double-precision floats here? But then, aren't modern GPU optimized mainly for single-precision float operations? Section 3.3. also says that the quantization ensures that there are nearly no overflows. What happens when an overflow occurs? I guess that because of the randomized blinding, a single overflow would result in a completely random output. How do you deal with this during training?\n\nMinor\n=====\n- Typo in abstract: Slaom -> Slalom\n- I don't understand the purpose of footnote 3 in Appendix B.2. First, the bibliographic entry for (Volos et al. 2018) explicitly says that the paper was published in OSDI 2018, a top-tier peer-reviewed conference. Regardless, claiming a date for a first unpublished draft of your paper is a little unusual and somewhat meaningless. I'm sure Volos et al. had a draft of their paper ready in late 2017 or even earlier if they submitted to OSDI in XXX 2018. If you want to timestamp your paper, post in to arXiv or elsewhere online.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "\nSummary\n========\nThis paper proposes a framework for privacy-preserving training of neural networks, by leveraging trusted execution environments and untrusted GPU accelerators.\nThe system builds heavily on the prior Slalom system, and uses standard MPC techniques (three non-colluding servers, multiplication triplets) to extend Slalom's inference-only protocol to privacy-preserving training.\nThis is a valuable and hard to reach goal. Unfortunately, the paper's evaluation fails to deliver on its strong promises, by ignoring the high network communication between the non-colluding servers.\nSpecifically, all experiments were conducted with three servers co-located in a public cloud's LAN. In this setting, it is hard to argue that non-collusion is a valid security assumption as the cloud provider controls all servers (alternatively, if the cloud provider is trusted, then there is no need for any trusted execution or cryptography). If the same experiments were conducted on a WAN, the communication costs would alleviate any savings in computation time.\n\nFor these reasons, I lean strongly towards rejection of this paper. \n\nDetailed comments\n=================\nExtending the ideas in Slalom to support privacy-preserving training is a good research question, and Tramer and Boneh had discussed some of the challenges and limitations towards this in their original paper.\nGetting rid of the pre-processing stage for blinding factors by leveraging non-colluding servers is a well-known trick from the MPC literature, but it does not seem easily applicable here. \nThe problem is that the servers need to communicate an amount of data proportional to the size of each internal layer of the network, for each forward and backward pass. If the servers communicate over a standard WAN, the communication time will be much too high to be competitive with the CaffeScone baseline.\nIn a LAN, as in this paper's experiments, the network latency is low enough for the network costs to be dominated by computation. But this begs the question of whether servers running in a same LAN (e.g., hosted by a single cloud provider) can really be considered non-colluding. In the considered setup, the cloud provider (Google in this case), could just observe the communication between all servers, thereby breaking privacy.\n\nAnother security issue with proposed scheme is the lack of computation integrity. This corresponds to the so-called \"honest-but-curious\" threat model which often appears in the MPC literature, and this should be acknowledged and motivated.\n\nOn the experimental side, the considered baseline, CaffeScone, seems pretty weak. In particular, any optimizations that the authors implement for Goten (e.g., better paging) should also be added to their baseline for a fair comparison.\nThe numbers in Figure 3 show that the baseline could be optimized a lot further.  A gap between hardware/simulation modes of ~6x seems indicative of sub-optimal paging. Even the single-core, simulation mode throughput numbers seem low for CIFAR10.\n\nThe experimental setup is quite confusing. Running the baseline and Goten in different environments (e.g., different CPUs) and then re-normalizing throughputs is somewhat convoluted and prone to mistakes. Why not run all experiments on the same setup?\nSimilarly, converting between results in SGX's hardware and simulation modes is also not very indicative. The authors note (p. 8) that in SGX's simulation mode \"code compilation is almost the same as hardware mode except that the program is not protected by SGX, which is fine for our purpose since the DNN training and prediction algorithms are publicly known\". This is fundamentally incorrect!\nSGX's simulation mode provides absolutely no security guarantees. It simply compiles the code using the SGX libraries and ensures that the enclaved code performs no untrusted operations, but it does not provide any hardware protections whatsoever. In particular, code running in simulation mode will not be affected by the overhead of SGX's paging, as the memory is never encrypted.\nAs a result, performance results in simulation mode are usually not indicative of performance in hardware mode. Trying to convert runtimes from simulation mode to hardware mode by comparing times of specific layers is also prone to many approximation errors. \n\nFinally, I had some trouble understanding the way in which Goten quantization works. Section 3.3. mentions that values are treated as floats, but then mentions the use of 53 bits of precision. Did you mean double-precision floats here? But then, aren't modern GPU optimized mainly for single-precision float operations? Section 3.3. also says that the quantization ensures that there are nearly no overflows. What happens when an overflow occurs? I guess that because of the randomized blinding, a single overflow would result in a completely random output. How do you deal with this during training?\n\nMinor\n=====\n- Typo in abstract: Slaom -> Slalom\n- I don't understand the purpose of footnote 3 in Appendix B.2. First, the bibliographic entry for (Volos et al. 2018) explicitly says that the paper was published in OSDI 2018, a top-tier peer-reviewed conference. Regardless, claiming a date for a first unpublished draft of your paper is a little unusual and somewhat meaningless. I'm sure Volos et al. had a draft of their paper ready in late 2017 or even earlier if they submitted to OSDI in XXX 2018. If you want to timestamp your paper, post in to arXiv or elsewhere online."}, "tcdate": 1571200866607}], "openreview_url": "https://openreview.net/forum?id=S1xRnxSYwS", "arxiv_id": null, "paper_pdf": "papers/S1xRnxSYwS.pdf", "paper_pdf_sha256": "b65c5957a7d73dce54f572e8b1ba09afd52ad92b14b4dd0746c842085cdd0d0e", "paper_pdf_bytes": 658765, "paper_pdf_source": "openreview", "code_url": "https://github.com/goten-team/Goten", "code_repository": "goten-team/Goten", "code_commit": "690f1429b62c70caec72f4010ee5b7a9786f0d25", "code_archive": "repos/S1xRnxSYwS.zip", "code_archive_sha256": "e4409347b81b3ae3eabffd6cfc9e76b2cfc5a51eda7857af46eb96d65d02a305", "code_archive_bytes": 5487816, "code_file_count": 2036, "code_extensions": {".h": 942, ".cpp": 891, ".f": 51, ".cmake": 48, ".py": 37, ".cu": 25, ".c": 21, ".hpp": 7, ".sh": 6, ".cc": 5, ".js": 3}, "github_disk_usage_kb": 3217, "github_languages": {"C++": 18147133, "Fortran": 1326307, "C": 453077, "CMake": 352722, "Cuda": 263716, "Python": 259349, "Assembly": 39018, "HTML": 19500, "Shell": 16801, "Makefile": 9381}, "github_archived": false, "github_pushed_at": "2021-01-01T14:42:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/goten-gpu-outsourcing-trusted-execution-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sp1Wr40egd", "year": 2026, "status": "rejected", "title": "To Trust Or Not To Trust Your Vision-Language Model's Prediction", "authors": ["Hao Dong", "Moru Liu", "Jian Liang", "Eleni Chatzi", "Olga Fink"], "authorids": ["~Hao_Dong4", "~Moru_Liu1", "~Jian_Liang1", "~Eleni_Chatzi1", "~Olga_Fink1"], "authors_source": "OpenReview API", "abstract": "Vision-Language Models (VLMs) have demonstrated strong capabilities in aligning visual and textual modalities, enabling a wide range of applications in multimodal understanding and generation. While they excel in zero-shot and transfer learning scenarios, VLMs remain susceptible to misclassification, often yielding confident yet incorrect predictions. This limitation poses a significant risk in safety-critical domains, where erroneous predictions can lead to severe consequences. In this work, we introduce TrustVLM, a training-free framework designed to address the critical challenge of estimating when VLM’s predictions can be trusted. Motivated by the observed modality gap in VLMs and the insight that certain concepts are more distinctly represented in the image embedding space, we propose a novel confidence-scoring function that leverages this space to improve misclassification detection. We rigorously evaluate our approach across 17 diverse datasets, employing 4 architectures and 2 VLMs, and demonstrate state-of-the-art performance, with improvements of up to 51.87% in AURC, 9.14% in AUROC, and 32.42% in FPR95 compared to existing baselines. By improving the reliability of the model without requiring retraining, TrustVLM paves the way for safer deployment of VLMs in real-world applications. The code is available in Supplementary Material.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "KzPFJJQGFs", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11620/Reviewer_pKks"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "For improving the reliability of VLM, this paper proposes a new training-free framework TrustVLM to estimate when VLM’s prediction will be trusted. The key of TrustVLM is to use a new confidence scoring function which will use not only cosine similarity between image and text but also the addition information from image embedding space. This additional information is the image-to-image relations or similarity. The idea is very straightforward where you need to extract some prior knowledge from training data. This prior knowledge is the class embeddings that are extracted from N shot examples from training data for each class. It likes you did another way for image classification based on the similarity between class embeddings and input image embedding. The whole idea is clear. The paper is well written. The key question here is you need to have a training data and get the class embedding based on the classes in the training data. But in some use cases, we don’t have the training data to get these class embeddings in advance. The proposed method has its limitation on it.", "review_text": "For improving the reliability of VLM, this paper proposes a new training-free framework TrustVLM to estimate when VLM’s prediction will be trusted. The key of TrustVLM is to use a new confidence scoring function which will use not only cosine similarity between image and text but also the addition information from image embedding space. This additional information is the image-to-image relations or similarity. The idea is very straightforward where you need to extract some prior knowledge from training data. This prior knowledge is the class embeddings that are extracted from N shot examples from training data for each class. It likes you did another way for image classification based on the similarity between class embeddings and input image embedding. The whole idea is clear. The paper is well written. The key question here is you need to have a training data and get the class embedding based on the classes in the training data. But in some use cases, we don’t have the training data to get these class embeddings in advance. The proposed method has its limitation on it.", "strengths": "1.\tTrustVLM is a training-free framework designed to evaluate the reliability of VLM predictions. One of its key advantages is that it does not require additional training, which makes it convenient to apply in scenarios where labeled data is limited or unavailable. The framework combines both image-to-text and image-to-image similarities, which allows for a more robust and nuanced design of confidence scores. This combination provides a richer representation of the visual information, enabling the framework to better capture the model’s uncertainty.\n\n2.\tThe paper demonstrates that the proposed visual prototypes not only enable more reliable confidence estimation but also enhance fine-grained classification accuracy.\n\n3.\tThe experiments conducted across diverse datasets, model architectures, and VLMs show the generality and effectiveness of TrustVLM.", "weaknesses": "1.\tA notable limitation of this method is that it relies on the availability of in-domain data that includes images for all classes to be predicted. Under this assumption, the method can extract and store visual prototypes for each class, which are then used for confidence estimation. However, in many practical scenarios, obtaining such in-domain data for every class may be difficult or infeasible. Moreover, if the training or reference data does not fully cover the diversity of the test data, the method may encounter out-of-distribution (OOD) situations. As the introduction clearly states, TrustVLM is not designed to handle OOD cases, which inherently limits its applicability in environments where data coverage is incomplete or classes are highly dynamic. This restriction should be carefully considered when evaluating the practical utility of the method.\n\n2.\tThere exist alternative strategies to improve reliability when prior knowledge about class representations is available. For instance, one could employ a separate image embedding model to independently obtain embeddings for the input and for each class, then compute similarity scores between them. Another potential approach is to first predict the class of the image using a preliminary classifier and then use this prediction as an input for the final classification task. These alternatives might offer comparable or complementary benefits to TrustVLM. Therefore, the paper would be strengthened if the authors could provide a more detailed comparison or discussion of how TrustVLM differs from these approaches. Specifically, it would be helpful to clarify the unique advantages of TrustVLM, such as why its combined use of image-to-text and image-to-image similarity provides superior or more reliable confidence estimation compared to these other methods. This explanation would help to more clearly establish the method’s contributions and practical significance.", "questions": "Please check the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "For improving the reliability of VLM, this paper proposes a new training-free framework TrustVLM to estimate when VLM’s prediction will be trusted. The key of TrustVLM is to use a new confidence scoring function which will use not only cosine similarity between image and text but also the addition information from image embedding space. This additional information is the image-to-image relations or similarity. The idea is very straightforward where you need to extract some prior knowledge from training data. This prior knowledge is the class embeddings that are extracted from N shot examples from training data for each class. It likes you did another way for image classification based on the similarity between class embeddings and input image embedding. The whole idea is clear. The paper is well written. The key question here is you need to have a training data and get the class embedding based on the classes in the training data. But in some use cases, we don’t have the training data to get these class embeddings in advance. The proposed method has its limitation on it.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "1.\tTrustVLM is a training-free framework designed to evaluate the reliability of VLM predictions. One of its key advantages is that it does not require additional training, which makes it convenient to apply in scenarios where labeled data is limited or unavailable. The framework combines both image-to-text and image-to-image similarities, which allows for a more robust and nuanced design of confidence scores. This combination provides a richer representation of the visual information, enabling the framework to better capture the model’s uncertainty.\n\n2.\tThe paper demonstrates that the proposed visual prototypes not only enable more reliable confidence estimation but also enhance fine-grained classification accuracy.\n\n3.\tThe experiments conducted across diverse datasets, model architectures, and VLMs show the generality and effectiveness of TrustVLM.", "weaknesses": "1.\tA notable limitation of this method is that it relies on the availability of in-domain data that includes images for all classes to be predicted. Under this assumption, the method can extract and store visual prototypes for each class, which are then used for confidence estimation. However, in many practical scenarios, obtaining such in-domain data for every class may be difficult or infeasible. Moreover, if the training or reference data does not fully cover the diversity of the test data, the method may encounter out-of-distribution (OOD) situations. As the introduction clearly states, TrustVLM is not designed to handle OOD cases, which inherently limits its applicability in environments where data coverage is incomplete or classes are highly dynamic. This restriction should be carefully considered when evaluating the practical utility of the method.\n\n2.\tThere exist alternative strategies to improve reliability when prior knowledge about class representations is available. For instance, one could employ a separate image embedding model to independently obtain embeddings for the input and for each class, then compute similarity scores between them. Another potential approach is to first predict the class of the image using a preliminary classifier and then use this prediction as an input for the final classification task. These alternatives might offer comparable or complementary benefits to TrustVLM. Therefore, the paper would be strengthened if the authors could provide a more detailed comparison or discussion of how TrustVLM differs from these approaches. Specifically, it would be helpful to clarify the unique advantages of TrustVLM, such as why its combined use of image-to-text and image-to-image similarity provides superior or more reliable confidence estimation compared to these other methods. This explanation would help to more clearly establish the method’s contributions and practical significance.", "questions": "Please check the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761980000019}, {"id": "SV7lOUnpkt", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11620/Reviewer_CAb7"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "In this work, the authors introduce TrustVLM, a training-free framework for predicting when a VLM’s predictions can be trusted. The task explored in this work is misclassification detection, which involves identifying when a prediction is incorrect. The key idea is to (1) generate visual prototypes for each class (i.e. the average embedding for N samples from the training data) and (2) compute the image-to-image similarity between the query image and the class prototypes. The standard image-to-text cosine similarity and the computed image-to-image similarity scores are combined in order to determine the overall prediction confidence. The authors show that this approach leads to substantially better misclassification detection performance than baselines.", "review_text": "In this work, the authors introduce TrustVLM, a training-free framework for predicting when a VLM’s predictions can be trusted. The task explored in this work is misclassification detection, which involves identifying when a prediction is incorrect. The key idea is to (1) generate visual prototypes for each class (i.e. the average embedding for N samples from the training data) and (2) compute the image-to-image similarity between the query image and the class prototypes. The standard image-to-text cosine similarity and the computed image-to-image similarity scores are combined in order to determine the overall prediction confidence. The authors show that this approach leads to substantially better misclassification detection performance than baselines.", "strengths": "- This work addresses an important task: determining when the predictions of a VLM are likely to be reliable.\n- Although the proposed method is methodologically straightforward, strong performance is observed across a range of datasets and model backbones. The authors also compare with multiple baselines. The distribution shift experiments with ImageNet are particularly compelling.", "weaknesses": "- **Need for finer-grained analysis:** This paper could benefit from additional fine-grained analysis with respect to when the proposed method is most effective (rather than just overall metrics). For example, are there specific classes where misclassification detection performance improves substantially when using the proposed method (as compared to MSP)? What types of characteristics are common among those classes?\n- **Variance of performance:** The proposed method is likely very sensitive to the choice of few-shot samples used to compose the class prototypes. What is the variance in performance when using prototypes composed from different randomly-selected N-shot sample sets?", "questions": "Questions are listed above under weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors introduce TrustVLM, a training-free framework for predicting when a VLM’s predictions can be trusted. The task explored in this work is misclassification detection, which involves identifying when a prediction is incorrect. The key idea is to (1) generate visual prototypes for each class (i.e. the average embedding for N samples from the training data) and (2) compute the image-to-image similarity between the query image and the class prototypes. The standard image-to-text cosine similarity and the computed image-to-image similarity scores are combined in order to determine the overall prediction confidence. The authors show that this approach leads to substantially better misclassification detection performance than baselines.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- This work addresses an important task: determining when the predictions of a VLM are likely to be reliable.\n- Although the proposed method is methodologically straightforward, strong performance is observed across a range of datasets and model backbones. The authors also compare with multiple baselines. The distribution shift experiments with ImageNet are particularly compelling.", "weaknesses": "- **Need for finer-grained analysis:** This paper could benefit from additional fine-grained analysis with respect to when the proposed method is most effective (rather than just overall metrics). For example, are there specific classes where misclassification detection performance improves substantially when using the proposed method (as compared to MSP)? What types of characteristics are common among those classes?\n- **Variance of performance:** The proposed method is likely very sensitive to the choice of few-shot samples used to compose the class prototypes. What is the variance in performance when using prototypes composed from different randomly-selected N-shot sample sets?", "questions": "Questions are listed above under weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761966603499}, {"id": "UszoHFTBQW", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11620/Reviewer_FUYk"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper studies VLM error detection to enhance its trustworthiness. The proposed TrustVLM leverages the multimodal similarity of feature representations to decide whether the current prediction is correct. Particularly, it compares the query example with other examples in both image representation similarity and text representation similarity. The final confidence score for error detection is computed by combining both similarities, thus incorporating multimodal information. Through extensive experiments of quantitative analysis and qualitative study, the effectiveness of the TrustVLM has been superior to most of the baseline methods.", "review_text": "This paper studies VLM error detection to enhance its trustworthiness. The proposed TrustVLM leverages the multimodal similarity of feature representations to decide whether the current prediction is correct. Particularly, it compares the query example with other examples in both image representation similarity and text representation similarity. The final confidence score for error detection is computed by combining both similarities, thus incorporating multimodal information. Through extensive experiments of quantitative analysis and qualitative study, the effectiveness of the TrustVLM has been superior to most of the baseline methods.", "strengths": "- This paper is easy to follow, the motivation is very clear, and the intuition is quite straightforward.\n- The proposed TrustVLM is training-free and efficient to deploy. It can also be easily adopted by any VLM architectures.\n- The experimental performance is quite promising.", "weaknesses": "- The major concern is missing the comparison with unimodal detection methods. The proposed method combines multimodal information to detect prediction errors; however, in the ablation study, there is no comparison with image-only or text-only detection. In this way, it would be clearer which branch of modality would contribute more to the overall performance improvement.\n- Moreover, the performance of TrustVLM highly relies on the performance of the employed VLMs; if the VLMs cannot provide high-quality representations, the error detection would be limited.\n- Another concern is that due to the existence of a modality gap, the cross-modal similarity could be unstable compared to image-to-image similarity. The misaligned cross-modal pairs would also mislead the error detection.\n- After detection, the proposed TrustVLM cannot further rectify the error predictions by finding the correct one.", "questions": "- Which branch of modality contributes more to the overall performance improvement? It would be helpful to conduct an ablation study to verify the unimodal detection versus multimodal detection. Moreover, what if we directly ask LLMs to detect the prediction error? As done in ``Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning, in ICML 2024'', they leverage the prediction of LLMs to find prediction errors of VLMs, and LLMs can further correct the prediction to find the correct one.\n- How would the misaligned multimodal pairs mislead the overall detection performance?\n- The acquisition of prototypes could be difficult sometimes. Can the proposed method perform without prototypes?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies VLM error detection to enhance its trustworthiness. The proposed TrustVLM leverages the multimodal similarity of feature representations to decide whether the current prediction is correct. Particularly, it compares the query example with other examples in both image representation similarity and text representation similarity. The final confidence score for error detection is computed by combining both similarities, thus incorporating multimodal information. Through extensive experiments of quantitative analysis and qualitative study, the effectiveness of the TrustVLM has been superior to most of the baseline methods.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This paper is easy to follow, the motivation is very clear, and the intuition is quite straightforward.\n- The proposed TrustVLM is training-free and efficient to deploy. It can also be easily adopted by any VLM architectures.\n- The experimental performance is quite promising.", "weaknesses": "- The major concern is missing the comparison with unimodal detection methods. The proposed method combines multimodal information to detect prediction errors; however, in the ablation study, there is no comparison with image-only or text-only detection. In this way, it would be clearer which branch of modality would contribute more to the overall performance improvement.\n- Moreover, the performance of TrustVLM highly relies on the performance of the employed VLMs; if the VLMs cannot provide high-quality representations, the error detection would be limited.\n- Another concern is that due to the existence of a modality gap, the cross-modal similarity could be unstable compared to image-to-image similarity. The misaligned cross-modal pairs would also mislead the error detection.\n- After detection, the proposed TrustVLM cannot further rectify the error predictions by finding the correct one.", "questions": "- Which branch of modality contributes more to the overall performance improvement? It would be helpful to conduct an ablation study to verify the unimodal detection versus multimodal detection. Moreover, what if we directly ask LLMs to detect the prediction error? As done in ``Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning, in ICML 2024'', they leverage the prediction of LLMs to find prediction errors of VLMs, and LLMs can further correct the prediction to find the correct one.\n- How would the misaligned multimodal pairs mislead the overall detection performance?\n- The acquisition of prototypes could be difficult sometimes. Can the proposed method perform without prototypes?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761872479094}, {"id": "3PTzhcCUyn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11620/Reviewer_aUQM"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper introduces TrustVLM, a framework designed to address confidence estimation in vision-language model (VLM) predictions. Motivated by the modality gap inherent in VLMs, TrustVLM constructs ensembles of VLMs and image-only classifiers to enhance the detection of misclassified samples.", "review_text": "This paper introduces TrustVLM, a framework designed to address confidence estimation in vision-language model (VLM) predictions. Motivated by the modality gap inherent in VLMs, TrustVLM constructs ensembles of VLMs and image-only classifiers to enhance the detection of misclassified samples.", "strengths": "The presentation of this paper is clear and easy to follow.", "weaknesses": "The methodology proposed in this paper employs N-sample training data together with external image encoders to construct a Nearest Class Mean classifier, which is then combined with the original CLIP classifier. However, the improvement appears to stem primarily from the use of additional labeled data and model ensembling. I am concerned that this may not constitute a genuinely novel contribution. Moreover, since the competitive baselines operate in a zero-shot setting, the comparison could be considered unfair.", "questions": "How does TrustVLM perform with fewer training samples, such as in 1-shot or 2-shot settings? Can TrustVLM function with little or even no training data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces TrustVLM, a framework designed to address confidence estimation in vision-language model (VLM) predictions. Motivated by the modality gap inherent in VLMs, TrustVLM constructs ensembles of VLMs and image-only classifiers to enhance the detection of misclassified samples.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "The presentation of this paper is clear and easy to follow.", "weaknesses": "The methodology proposed in this paper employs N-sample training data together with external image encoders to construct a Nearest Class Mean classifier, which is then combined with the original CLIP classifier. However, the improvement appears to stem primarily from the use of additional labeled data and model ensembling. I am concerned that this may not constitute a genuinely novel contribution. Moreover, since the competitive baselines operate in a zero-shot setting, the comparison could be considered unfair.", "questions": "How does TrustVLM perform with fewer training samples, such as in 1-shot or 2-shot settings? Can TrustVLM function with little or even no training data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761667373492}], "openreview_url": "https://openreview.net/forum?id=sp1Wr40egd", "arxiv_id": "2505.23745", "paper_pdf": "papers/sp1Wr40egd.pdf", "paper_pdf_sha256": "8378b06fee2da516ef0d045432daa4c13042bb62926a44ffdc6d336f9ea36061", "paper_pdf_bytes": 5224967, "paper_pdf_source": "openreview", "code_url": "https://github.com/EPFL-IMOS/TrustVLM", "code_repository": "EPFL-IMOS/TrustVLM", "code_commit": "0a0caa60fba35624cde966c2a4b8c71c55ba5038", "code_archive": "repos/sp1Wr40egd.zip", "code_archive_sha256": "e44415f703baa2100a4f146c3fdd550d1f5e6ffa6f8b4d2e0eff6493b07c74c3", "code_archive_bytes": 3275633, "code_file_count": 19, "code_extensions": {".py": 17, ".sh": 2}, "github_disk_usage_kb": 3267, "github_languages": {"Python": 166413, "Shell": 2757}, "github_archived": false, "github_pushed_at": "2025-05-30T01:56:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/to-trust-or-not-to-trust-your-vision-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QByW8EYEtt", "year": 2025, "status": "rejected", "title": "Worse than Random? An Embarrassingly Simple Probing Evaluation of Large Multimodal Models in Medical VQA", "authors": ["Qianqi Yan", "Xuehai He", "Xiang Yue", "Xin Eric Wang"], "authorids": ["~Qianqi_Yan1", "~Xuehai_He1", "~Xiang_Yue1", "~Xin_Eric_Wang2"], "authors_source": "OpenReview API", "abstract": "Large Multimodal Models (LMMs) have shown remarkable progress in medical Visual Question Answering (Med-VQA), achieving high accuracy on existing benchmarks. However, their reliability under robust evaluation is questionable. This study reveals that state-of-the-art models perform worse than random guessing on medical diagnosis questions when subjected to simple probing evaluation. To address this critical evaluation problem, we introduce the Probing Evaluation for Medical Diagnosis (ProbMed) dataset to rigorously assess LMM performance in medical imaging through probing evaluation and procedural diagnosis. Particularly, probing evaluation features pairing original questions with negation questions with hallucinated attributes, while procedural diagnosis requires reasoning across various diagnostic dimensions for each image, including modality recognition, organ identification, clinical findings, abnormalities, and positional grounding. Our evaluation reveals that top-performing models like GPT-4o, GPT-4V, and Gemini Pro perform worse than random guessing on specialized diagnostic questions, indicating significant limitations in handling fine-grained medical inquiries. We further investigate the underperformance of open-source models (e.g., LLaVA, LLaVA-Med, and Med-Flamingo) through an ablation study. This study reveals that poor visual understanding is a primary bottleneck, which can be mitigated by adding visual descriptions generated by GPT-4o, leading to an average performance improvement of 9.44%. These findings underscore the urgent need for more robust evaluation methods and domain-specific expertise to ensure LMM reliability in critical medical fields.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "r9jcmxV7Mk", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11380/Reviewer_6yHr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces ProbMed, a Med-VQA dataset with adversarial negative samples to reveal multimodal models' limitations in complex medical tasks. Chain-of-thought reasoning and GPT-4-generated visual descriptions are shown to enhance fine-grained diagnostic performance.", "review_text": "This paper introduces ProbMed, a Med-VQA dataset with adversarial negative samples to reveal multimodal models' limitations in complex medical tasks. Chain-of-thought reasoning and GPT-4-generated visual descriptions are shown to enhance fine-grained diagnostic performance.", "strengths": "1. By introducing adversarial negative samples, the dataset tests model robustness, offering a more challenging evaluation standard.\n2. ProbMed includes various diagnostic tasks across different imaging modalities and body organs, providing a evaluation setting for models.\n3. The model’s performance improvement strategies are validated through methods like chain-of-thought reasoning and visual description enhancement, providing a foundation for future advancements.", "weaknesses": "1. The dataset has a significantly higher number of chest X-ray images than other types, which may lead to poorer model performance on other organs. Has the model's performance across different modalities been experimentally verified?\n2. How does the study prevent hallucinations when using GPT-4 for caption analysis, abnormality identification, and positional descriptions through few-shot prompting?\n3. ProbMed uses closed-ended questions and lacks open-ended tasks, limiting the dataset's comprehensiveness for evaluation.\n4. Constructing adversarial samples by selecting random entities (such as alternative organs, modalities, or hallucinated conditions with \"no\" answers) may increase testing difficulty but could also introduce mistakes, resulting in samples that aren’t truly adversarial.", "questions": "1. Is the construction of adversarial negative samples reasonable? Could additional strategies be introduced to ensure the validity of these adversarial samples?\n2. Does class imbalance in the dataset affect the model's generalization ability? How can better diagnostic performance be achieved for organs beyond the chest? Additionally, validation across multiple modalities is recommended to assess the impact of imbalanced data distribution on each modality.\n3. What is the impact of adding open-ended tasks on the model's performance? How can report generation or open-ended question-answering be incorporated into ProbMed? Clearly, open-ended questions are more representative of common clinical scenarios.\n4. Could GPT-4-generated visual descriptions introduce bias or hallucinations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ProbMed, a Med-VQA dataset with adversarial negative samples to reveal multimodal models' limitations in complex medical tasks. Chain-of-thought reasoning and GPT-4-generated visual descriptions are shown to enhance fine-grained diagnostic performance.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. By introducing adversarial negative samples, the dataset tests model robustness, offering a more challenging evaluation standard.\n2. ProbMed includes various diagnostic tasks across different imaging modalities and body organs, providing a evaluation setting for models.\n3. The model’s performance improvement strategies are validated through methods like chain-of-thought reasoning and visual description enhancement, providing a foundation for future advancements.", "weaknesses": "1. The dataset has a significantly higher number of chest X-ray images than other types, which may lead to poorer model performance on other organs. Has the model's performance across different modalities been experimentally verified?\n2. How does the study prevent hallucinations when using GPT-4 for caption analysis, abnormality identification, and positional descriptions through few-shot prompting?\n3. ProbMed uses closed-ended questions and lacks open-ended tasks, limiting the dataset's comprehensiveness for evaluation.\n4. Constructing adversarial samples by selecting random entities (such as alternative organs, modalities, or hallucinated conditions with \"no\" answers) may increase testing difficulty but could also introduce mistakes, resulting in samples that aren’t truly adversarial.", "questions": "1. Is the construction of adversarial negative samples reasonable? Could additional strategies be introduced to ensure the validity of these adversarial samples?\n2. Does class imbalance in the dataset affect the model's generalization ability? How can better diagnostic performance be achieved for organs beyond the chest? Additionally, validation across multiple modalities is recommended to assess the impact of imbalanced data distribution on each modality.\n3. What is the impact of adding open-ended tasks on the model's performance? How can report generation or open-ended question-answering be incorporated into ProbMed? Clearly, open-ended questions are more representative of common clinical scenarios.\n4. Could GPT-4-generated visual descriptions introduce bias or hallucinations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730632258340}, {"id": "cg3u2q82Mg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11380/Reviewer_vP6d"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces the ProbMed dataset for the rigorous evaluation of large multimodal models in clinical domains. The authors show that models perform worse than random on medical diagnosis questions when subjected to adversarial questions. They also show that the poor visual understanding of LMMs is a primary bottleneck, and can be mitigated by adding visual descriptions generated by GPT-4o.", "review_text": "The paper introduces the ProbMed dataset for the rigorous evaluation of large multimodal models in clinical domains. The authors show that models perform worse than random on medical diagnosis questions when subjected to adversarial questions. They also show that the poor visual understanding of LMMs is a primary bottleneck, and can be mitigated by adding visual descriptions generated by GPT-4o.", "strengths": "It is a well-written paper with many experiments to evaluate the models on the proposed dataset. Several popular models have been studied and also authors propose mitigation strategies to improve the models' performance. Overall, the paper targets important questions and has the potential to be a good contribution to the field.", "weaknesses": "After reading the paper, I have a few questions and concerns as follows:\n\n\n1. (major) If I understood it correctly, the dataset has been curated using the publically available data on the internet. My main concern is the possible data contamination in larger closed or open-source models. When some models have been trained on the data and some not, evaluation using this dataset loses its fairness. \n\n2. If I have gotten it right, all the adversarial questions are in the form of negated questions and their response is \"no\". Having a model that has been previously trained on the data, can we ensure that it does not cheat?\n\n3. (major) I strongly suggest adding a distribution of the responses that are \"yes\" or \"no\". how many questions have an answer of \"no\" and how many \"yes\" within each category in the dataset? This can be done through a qualitative distribution plot in the paper. \n\n4. (major) The adversarial question design is creative, but it has issues as well. Within clinical data. there are always cases that have co-occurrence of multiple forms of the disease, but in the original caption, we only have one of them as according to a clinician it is the important one. In this regard, when we create adversarial questions, this important fact has been ignored. So, each question actually needs to be validated by a medical expert. I have seen that 100 samples were examined by experts, but that number is significantly small compared to the size of the dataset. In fact, the paper lacks a thorough and careful expert study to ensure correctness.", "questions": "I have mentioned them in weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the ProbMed dataset for the rigorous evaluation of large multimodal models in clinical domains. The authors show that models perform worse than random on medical diagnosis questions when subjected to adversarial questions. They also show that the poor visual understanding of LMMs is a primary bottleneck, and can be mitigated by adding visual descriptions generated by GPT-4o.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "It is a well-written paper with many experiments to evaluate the models on the proposed dataset. Several popular models have been studied and also authors propose mitigation strategies to improve the models' performance. Overall, the paper targets important questions and has the potential to be a good contribution to the field.", "weaknesses": "After reading the paper, I have a few questions and concerns as follows:\n\n\n1. (major) If I understood it correctly, the dataset has been curated using the publically available data on the internet. My main concern is the possible data contamination in larger closed or open-source models. When some models have been trained on the data and some not, evaluation using this dataset loses its fairness. \n\n2. If I have gotten it right, all the adversarial questions are in the form of negated questions and their response is \"no\". Having a model that has been previously trained on the data, can we ensure that it does not cheat?\n\n3. (major) I strongly suggest adding a distribution of the responses that are \"yes\" or \"no\". how many questions have an answer of \"no\" and how many \"yes\" within each category in the dataset? This can be done through a qualitative distribution plot in the paper. \n\n4. (major) The adversarial question design is creative, but it has issues as well. Within clinical data. there are always cases that have co-occurrence of multiple forms of the disease, but in the original caption, we only have one of them as according to a clinician it is the important one. In this regard, when we create adversarial questions, this important fact has been ignored. So, each question actually needs to be validated by a medical expert. I have seen that 100 samples were examined by experts, but that number is significantly small compared to the size of the dataset. In fact, the paper lacks a thorough and careful expert study to ensure correctness.", "questions": "I have mentioned them in weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730621937886}, {"id": "Y73Efa4QmO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11380/Reviewer_9nkQ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This work curates a medical VQA dataset with pairing original questions and negation questions  and evaluations of SOTA VLMs show the poor visual understanding abilities. Furthermore, the authors show that this issue could be eliminated by adding external visual descriptions generated by GPT-4o.", "review_text": "This work curates a medical VQA dataset with pairing original questions and negation questions  and evaluations of SOTA VLMs show the poor visual understanding abilities. Furthermore, the authors show that this issue could be eliminated by adding external visual descriptions generated by GPT-4o.", "strengths": "- A new dataset, PromMed, was curated for medical VQA benchmarking, which contains adversarial question pairs\n- Comprehensive experiments on multiple VLMs\n- Insightful findings: SOTA VLMs perform worse than random guessing on specialized diagnostic questions", "weaknesses": "- Some questions are too trivial in the dataset. The ultimate goal of multimodal models is deployment in real clinical practice when the accuracy is good enough. However, no clinician will ask questions on basic modalities (e.g., CT, MR) or organs because they are too trivial. Arguments such as “CheXagent, trained exclusively on chest X-rays, achieved the highest accuracy in determining abnormalities and conditions. However, its performance in general tasks like identifying image modality and organs was lower” should be modified because ChexXagent was not designed for tasks like identifying image modality and organs. These tasks are also not clinically significant because of their trivialities. I would suggest that the authors focus their analysis and discussion on the more clinically relevant and challenging aspects of the dataset, such as identifying specific abnormalities or conditions. \n\n- Minor: References format is not consistent. Some preprint papers even missed arxiv id", "questions": "- It is interesting that augmentation with visual descriptions generated by GPT-4o can improve the performance. Can open-sourced multi-modal models (e.g., Qwen2-VL-72b-Instruct, LLaMa 3.2 Vision) improve the performance as well?\n\n- Sec. 4.3.2 it is not clear how CheXagent is used to enhance the model performance. Could you please provide more details on how CheXagent was integrated with other vision-language models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work curates a medical VQA dataset with pairing original questions and negation questions  and evaluations of SOTA VLMs show the poor visual understanding abilities. Furthermore, the authors show that this issue could be eliminated by adding external visual descriptions generated by GPT-4o.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- A new dataset, PromMed, was curated for medical VQA benchmarking, which contains adversarial question pairs\n- Comprehensive experiments on multiple VLMs\n- Insightful findings: SOTA VLMs perform worse than random guessing on specialized diagnostic questions", "weaknesses": "- Some questions are too trivial in the dataset. The ultimate goal of multimodal models is deployment in real clinical practice when the accuracy is good enough. However, no clinician will ask questions on basic modalities (e.g., CT, MR) or organs because they are too trivial. Arguments such as “CheXagent, trained exclusively on chest X-rays, achieved the highest accuracy in determining abnormalities and conditions. However, its performance in general tasks like identifying image modality and organs was lower” should be modified because ChexXagent was not designed for tasks like identifying image modality and organs. These tasks are also not clinically significant because of their trivialities. I would suggest that the authors focus their analysis and discussion on the more clinically relevant and challenging aspects of the dataset, such as identifying specific abnormalities or conditions. \n\n- Minor: References format is not consistent. Some preprint papers even missed arxiv id", "questions": "- It is interesting that augmentation with visual descriptions generated by GPT-4o can improve the performance. Can open-sourced multi-modal models (e.g., Qwen2-VL-72b-Instruct, LLaMa 3.2 Vision) improve the performance as well?\n\n- Sec. 4.3.2 it is not clear how CheXagent is used to enhance the model performance. Could you please provide more details on how CheXagent was integrated with other vision-language models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730590949209}, {"id": "m1ZbDBxS5s", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11380/Reviewer_ivmY"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "Authors curated a large, high quality, balanced dataset of 57k close-ended (yes / no) VQA questions for  6,303 radiology medical images sourced from existing open-source datasets. The benchmark can be sliced across a number of different categories and for about half of the questions for which the answer is \"yes\", an \"adversarial\" question is also created by perturbing the entities / locations referenced in the original questions and changing the answers to \"no\". The dataset is used to evaluate a range of closed-source frontier MLLMs as well as open-source general purpose and domain-specific MLLMs. \n\nOverall the authors find that when the adversarial pair is introduced, all models saw significant drops in performance (> 20% in accuracy) although general purpose frontier models like gemini / gpt4o appear to be more robust and saw lower decrease in performance.  \nThe authors further find that when performance is broken down by category, general purpose models can still struggle significantly, and may perform worse on specialized domain-specific MLLMs. Amongst other findings, the authors also investigate whether COT can enhance performance, finding that while open-source models generally saw an improved performance, larger frontier models do not, and open-source models can benefit from using higher quality visual descriptions generated by larger frontier models as part of the COT process, further boosting performance.", "review_text": "Authors curated a large, high quality, balanced dataset of 57k close-ended (yes / no) VQA questions for  6,303 radiology medical images sourced from existing open-source datasets. The benchmark can be sliced across a number of different categories and for about half of the questions for which the answer is \"yes\", an \"adversarial\" question is also created by perturbing the entities / locations referenced in the original questions and changing the answers to \"no\". The dataset is used to evaluate a range of closed-source frontier MLLMs as well as open-source general purpose and domain-specific MLLMs. \n\nOverall the authors find that when the adversarial pair is introduced, all models saw significant drops in performance (> 20% in accuracy) although general purpose frontier models like gemini / gpt4o appear to be more robust and saw lower decrease in performance.  \nThe authors further find that when performance is broken down by category, general purpose models can still struggle significantly, and may perform worse on specialized domain-specific MLLMs. Amongst other findings, the authors also investigate whether COT can enhance performance, finding that while open-source models generally saw an improved performance, larger frontier models do not, and open-source models can benefit from using higher quality visual descriptions generated by larger frontier models as part of the COT process, further boosting performance.", "strengths": "The dataset is large, high quality, diverse, balanced, and is curated in a methodical manner, with plenty of useful metadata (such as question categorization, location annotation, etc.), making it an incredibly valuable benchmarking resource for the community. With these metadata intact, future works can easily build on top of the released data to curate more open-ended / multiple choice style questions if needed. This in my opinion is the paper's biggest strength. Additionally, the authors share some interesting findings about how much performance degrade when simple adversarial perturbations were applied to the original questions, which shows that even the frontier models today are quite fragile and limited in their ability to perform medical decision making reliably. The investigation into COT, especially COT with external visual descriptions are also quite insightful, revealing a clear path forward to improve open-source, domain-specific MLLMs.", "weaknesses": "My main complaint with the paper is that I find the presentation of the results difficult to follow at times. For example, it is not perfectly clear to me whether the \"adversarial\" questions are included in the tally of 57k QA pairs, and are also included as part of the evaluation benchmark by default - or if the 57k QA pairs represent the default, \"clean\" benchmark, with additional X number of adversarial questions (presumably = the number of questions with \"yes\" as answer) used to report the numbers in table 3 and the appendix only. It is also not super clear to me what the authors mean in table 3, where the authors distinguish between \"Averaged Accuracy\" and \"Accuracy (%) with Adversarial Pairs\". An illustrative example of how these metrics are computed would be helpful. \n\nAdditionally, a limitation of the study is that the benchmark / adversarial questions are only limited to binary yes / no settings, when I can imagine the methodology easily extending to multiple choice questions, which seems like a missed opportunity. Although I believe this is something that future works can build on given the well curated nature of the dataset.", "questions": "1. For table 3, my current understanding is that for questions that form a pair of original + adversarial question:\nGetting the original question correct, but adversarial question wrong would equate to an accuracy of 50% in the \"averaged accuracy\", and 0% in the second case. But in either case, the tally of the original questions that do not have an adversarial pair remains consistent? Wouldn't this obfuscate the true impact of the adversarial examples since it's diluted by the original, non-adversarial \"no\" questions?\n2. Are adversarial questions included in the tally of 57k QA pairs, and used in evaluation by default? e.g. in table 4, figure 4, 5, etc.\n3. Besides the question + image itself, is there any additional formatting / parsing used to evaluate the various models? especially for open-source MLLMs that may not be perfect at instruction following, what happens if a model outputs a full sentence response instead of a simple yes / no answer.\n4. How consistent are these results, especially for larger frontier models where it may not be possible guarantee deterministic output. Is there a way quantify the variance in the performance if e.g. gpt4o / gemini are evaluated more than once?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors curated a large, high quality, balanced dataset of 57k close-ended (yes / no) VQA questions for  6,303 radiology medical images sourced from existing open-source datasets. The benchmark can be sliced across a number of different categories and for about half of the questions for which the answer is \"yes\", an \"adversarial\" question is also created by perturbing the entities / locations referenced in the original questions and changing the answers to \"no\". The dataset is used to evaluate a range of closed-source frontier MLLMs as well as open-source general purpose and domain-specific MLLMs. \n\nOverall the authors find that when the adversarial pair is introduced, all models saw significant drops in performance (> 20% in accuracy) although general purpose frontier models like gemini / gpt4o appear to be more robust and saw lower decrease in performance.  \nThe authors further find that when performance is broken down by category, general purpose models can still struggle significantly, and may perform worse on specialized domain-specific MLLMs. Amongst other findings, the authors also investigate whether COT can enhance performance, finding that while open-source models generally saw an improved performance, larger frontier models do not, and open-source models can benefit from using higher quality visual descriptions generated by larger frontier models as part of the COT process, further boosting performance.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "The dataset is large, high quality, diverse, balanced, and is curated in a methodical manner, with plenty of useful metadata (such as question categorization, location annotation, etc.), making it an incredibly valuable benchmarking resource for the community. With these metadata intact, future works can easily build on top of the released data to curate more open-ended / multiple choice style questions if needed. This in my opinion is the paper's biggest strength. Additionally, the authors share some interesting findings about how much performance degrade when simple adversarial perturbations were applied to the original questions, which shows that even the frontier models today are quite fragile and limited in their ability to perform medical decision making reliably. The investigation into COT, especially COT with external visual descriptions are also quite insightful, revealing a clear path forward to improve open-source, domain-specific MLLMs.", "weaknesses": "My main complaint with the paper is that I find the presentation of the results difficult to follow at times. For example, it is not perfectly clear to me whether the \"adversarial\" questions are included in the tally of 57k QA pairs, and are also included as part of the evaluation benchmark by default - or if the 57k QA pairs represent the default, \"clean\" benchmark, with additional X number of adversarial questions (presumably = the number of questions with \"yes\" as answer) used to report the numbers in table 3 and the appendix only. It is also not super clear to me what the authors mean in table 3, where the authors distinguish between \"Averaged Accuracy\" and \"Accuracy (%) with Adversarial Pairs\". An illustrative example of how these metrics are computed would be helpful. \n\nAdditionally, a limitation of the study is that the benchmark / adversarial questions are only limited to binary yes / no settings, when I can imagine the methodology easily extending to multiple choice questions, which seems like a missed opportunity. Although I believe this is something that future works can build on given the well curated nature of the dataset.", "questions": "1. For table 3, my current understanding is that for questions that form a pair of original + adversarial question:\nGetting the original question correct, but adversarial question wrong would equate to an accuracy of 50% in the \"averaged accuracy\", and 0% in the second case. But in either case, the tally of the original questions that do not have an adversarial pair remains consistent? Wouldn't this obfuscate the true impact of the adversarial examples since it's diluted by the original, non-adversarial \"no\" questions?\n2. Are adversarial questions included in the tally of 57k QA pairs, and used in evaluation by default? e.g. in table 4, figure 4, 5, etc.\n3. Besides the question + image itself, is there any additional formatting / parsing used to evaluate the various models? especially for open-source MLLMs that may not be perfect at instruction following, what happens if a model outputs a full sentence response instead of a simple yes / no answer.\n4. How consistent are these results, especially for larger frontier models where it may not be possible guarantee deterministic output. Is there a way quantify the variance in the performance if e.g. gpt4o / gemini are evaluated more than once?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730500141201}], "openreview_url": "https://openreview.net/forum?id=QByW8EYEtt", "arxiv_id": "2405.20421", "paper_pdf": "papers/QByW8EYEtt.pdf", "paper_pdf_sha256": "4f73495ccc14076f934defa2a09a0db80d1d6d7cc7fc35356e3187e7f6182fff", "paper_pdf_bytes": 2811135, "paper_pdf_source": "openreview", "code_url": "https://github.com/UCSB-AI/ProbMed", "code_repository": "UCSB-AI/ProbMed", "code_commit": "36f91fbc865fd3cd1f1522c4472d56f5cb4df642", "code_archive": "repos/QByW8EYEtt.zip", "code_archive_sha256": "6a053df3b420be79a4daa017a88e0225693b8fcf26693d25798fc27c0abfb468", "code_archive_bytes": 206000, "code_file_count": 13, "code_extensions": {".py": 11, ".sh": 2}, "github_disk_usage_kb": 237, "github_languages": {"Python": 61789, "Shell": 5075}, "github_archived": false, "github_pushed_at": "2026-05-12T08:00:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/worse-than-random-an-embarrassingly-simple"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zIJFG7wW2d", "year": 2024, "status": "rejected", "title": "Agent Instructs Large Language Models to be General Zero-Shot Reasoners", "authors": ["Nicholas R Crispino", "Kyle Montgomery", "Fankun Zeng", "Dawn Song", "Chenguang Wang"], "authorids": ["~Nicholas_R_Crispino1", "~Kyle_Montgomery1", "~Fankun_Zeng1", "~Dawn_Song1", "~Chenguang_Wang1"], "authors_source": "OpenReview API", "abstract": "We introduce a method to improve the zero-shot reasoning abilities of large language models on general language understanding tasks. Unlike existing zero-shot reasoning approaches that are often suboptimal for general tasks, we build an autonomous agent to generate task-specific instructions to optimize the reasoning performance of large language models. We show our agent instructions further unleash the zero-shot reasoning abilities of large language models to more tasks. We study the performance of our method on a wide set of datasets spanning generation, classification, and reasoning. We show that our method generalizes to most tasks and obtains state-of-the-art zero-shot performance on 20 of the 29 datasets that we evaluate. For instance, our method boosts the performance of state-of-the-art large language models by a large margin, including Vicuna-13b (13.3%), Llama-2-70b-chat (23.2%), and GPT-3.5 Turbo (17.0%). Compared to zero-shot chain of thought, our improvement in reasoning is striking, with an average increase of 10.4%. With our method, Llama-2-70b-chat outperforms zero-shot GPT-3.5 Turbo by 10.2%. The code is available at https://anonymous.4open.science/r/AgentInstruct_ICLR2024.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "L6zz0nqKPD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5899/Reviewer_JX3F"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper presents a method to improve the zero-shot performance of LLMs on language tasks. The proposed approach involves automatically generating task-specific instructions (using an LLM based agent) that are provided to the reasoning LLM to improve its performance. The task-specific instructions might include steps to break down the task into simpler steps, information about the format of the expected answer. To generate these task-specific instructions, the agent takes as input information about the dataset (sourced from the web) and a few examples from that task. Using this extra context, the authors show that LLMs can reason better about the task and show improvement across several generation and classification based benchmarks. On average they show 6.5% improvement over a chain-of-thought baseline, and 17.5% average improvement over state-of-the-art LLMs like Vicuna, LLaMa2, and GPT-3.5-Turbo.", "review_text": "The paper presents a method to improve the zero-shot performance of LLMs on language tasks. The proposed approach involves automatically generating task-specific instructions (using an LLM based agent) that are provided to the reasoning LLM to improve its performance. The task-specific instructions might include steps to break down the task into simpler steps, information about the format of the expected answer. To generate these task-specific instructions, the agent takes as input information about the dataset (sourced from the web) and a few examples from that task. Using this extra context, the authors show that LLMs can reason better about the task and show improvement across several generation and classification based benchmarks. On average they show 6.5% improvement over a chain-of-thought baseline, and 17.5% average improvement over state-of-the-art LLMs like Vicuna, LLaMa2, and GPT-3.5-Turbo.", "strengths": "- The paper demonstrates the effectiveness of their approach on a fairly exhaustive evaluation. The approach is evaluated on 29 datasets spanning generation, classification and reasoning. The evaluation is run using three LLM baselines.\n- The authors provide exhaustive details about the prompts, evaluation dataset and experiments in the supplementary for easily replicating their experimental setup. The authors also provide code to run the experiments done in the paper.\n- The authors also attempt to do error analysis to better understand failure modes of their approach. I specially enjoyed reading Section 3.6 and Appendix C.1", "weaknesses": "- Evaluation vs “In-the-wild” generaliztion: While the paper shows strong improvements across various benchmarks, I am not convinced of the fundamental assumption made in the paper. To improve a performance on the task, the authors assume that the task is one of the “standard” tasks LLMs are typically evaluated on. While this is great for evaluating, its unclear how the proposed approach will work “in-the-wild”. In realistic usecases, the task might be completely different than one of the standard usecases. Additionally, we might not know beforehand which benchmark does the task belong to? In that case, it’s not going to always be possible to generate task-specific instructions.\n- Second, while the authors claim that the performance of the method is zero-shot, I believe that showing a few labeled examples in the prompt to the agent responsible for generating task-specific instruction is a bit unfair. A fair approach would be to make these examples be available even to other zero-shot baselines. Interestingly, without Input Examples, the approach takes a severe hit (see Table 1, Row 1 vs Row 3).\n- Finally, I also believe that cost comparisons done in Figure 6 are a bit misleading. Here, the authors show that Zero-shot AgentInstruct cost ~2$ vs 20$ (ZS-COT GPT4). This is true when running evaluation on standard benchmarks because we have to generate task-specific instructions once per benchmark. But like I said before, this is not going to be true in-the-wild. In that setup, a new instruction will have to be generated for every unique user query. For completeness, and to be consistent with a realistic use-case, it would be more accurate to include the costs of creating task-specific instruction to each instance of the task, and then show the overall cost.", "questions": "Related to points raised in the weaknesses section: \n\n- How will the approach work in-the-wild, when the source of task is not known apriori. That is we don’t know which dataset does the task belong to, or if the task belongs to any dataset at all.\n- What is the cost of creating the task-specific instruction. Can you add a column to Figure 6, which adds this cost multiplied by number of instances in the dataset plus the additional cost of reasoning using GPT-3.5 Turbo?\n- I also didn’t understand the point of experiments in Section 3.3? What is the insight from that experiment?\n- I also didn’t fully understand how the agent uses the question-answering API (Section A.3.1) to add context about the task? I think it’s one of the most crucial steps of the pipeline, and it’d be great to explain how it’s implemented in more detail (and perhaps in the main manuscript). Concretely, how are the retrieved documents to generate the instruction? Are the retrieved documents added as prompt to the agent to generate task-specific instruction? Is that in addition to the name of the dataset, task information and few input examples? Why are they added to a vector database?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method to improve the zero-shot performance of LLMs on language tasks. The proposed approach involves automatically generating task-specific instructions (using an LLM based agent) that are provided to the reasoning LLM to improve its performance. The task-specific instructions might include steps to break down the task into simpler steps, information about the format of the expected answer. To generate these task-specific instructions, the agent takes as input information about the dataset (sourced from the web) and a few examples from that task. Using this extra context, the authors show that LLMs can reason better about the task and show improvement across several generation and classification based benchmarks. On average they show 6.5% improvement over a chain-of-thought baseline, and 17.5% average improvement over state-of-the-art LLMs like Vicuna, LLaMa2, and GPT-3.5-Turbo.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper demonstrates the effectiveness of their approach on a fairly exhaustive evaluation. The approach is evaluated on 29 datasets spanning generation, classification and reasoning. The evaluation is run using three LLM baselines.\n- The authors provide exhaustive details about the prompts, evaluation dataset and experiments in the supplementary for easily replicating their experimental setup. The authors also provide code to run the experiments done in the paper.\n- The authors also attempt to do error analysis to better understand failure modes of their approach. I specially enjoyed reading Section 3.6 and Appendix C.1", "weaknesses": "- Evaluation vs “In-the-wild” generaliztion: While the paper shows strong improvements across various benchmarks, I am not convinced of the fundamental assumption made in the paper. To improve a performance on the task, the authors assume that the task is one of the “standard” tasks LLMs are typically evaluated on. While this is great for evaluating, its unclear how the proposed approach will work “in-the-wild”. In realistic usecases, the task might be completely different than one of the standard usecases. Additionally, we might not know beforehand which benchmark does the task belong to? In that case, it’s not going to always be possible to generate task-specific instructions.\n- Second, while the authors claim that the performance of the method is zero-shot, I believe that showing a few labeled examples in the prompt to the agent responsible for generating task-specific instruction is a bit unfair. A fair approach would be to make these examples be available even to other zero-shot baselines. Interestingly, without Input Examples, the approach takes a severe hit (see Table 1, Row 1 vs Row 3).\n- Finally, I also believe that cost comparisons done in Figure 6 are a bit misleading. Here, the authors show that Zero-shot AgentInstruct cost ~2$ vs 20$ (ZS-COT GPT4). This is true when running evaluation on standard benchmarks because we have to generate task-specific instructions once per benchmark. But like I said before, this is not going to be true in-the-wild. In that setup, a new instruction will have to be generated for every unique user query. For completeness, and to be consistent with a realistic use-case, it would be more accurate to include the costs of creating task-specific instruction to each instance of the task, and then show the overall cost.", "questions": "Related to points raised in the weaknesses section: \n\n- How will the approach work in-the-wild, when the source of task is not known apriori. That is we don’t know which dataset does the task belong to, or if the task belongs to any dataset at all.\n- What is the cost of creating the task-specific instruction. Can you add a column to Figure 6, which adds this cost multiplied by number of instances in the dataset plus the additional cost of reasoning using GPT-3.5 Turbo?\n- I also didn’t understand the point of experiments in Section 3.3? What is the insight from that experiment?\n- I also didn’t fully understand how the agent uses the question-answering API (Section A.3.1) to add context about the task? I think it’s one of the most crucial steps of the pipeline, and it’d be great to explain how it’s implemented in more detail (and perhaps in the main manuscript). Concretely, how are the retrieved documents to generate the instruction? Are the retrieved documents added as prompt to the agent to generate task-specific instruction? Is that in addition to the name of the dataset, task information and few input examples? Why are they added to a vector database?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698956341988}, {"id": "n8qyAXygfu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5899/Reviewer_LvbF"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper introduces AgentInstruct which generalizes zero-shot reasoning abilities of LLMs. Here, the agent generates instructions which optimize the reasoning process of LLMs. The proposed approach is compared with zero-shot and zero-shot-CoT as baselines and reports 17.8%  and 6.5% average improvement across the tasks.", "review_text": "The paper introduces AgentInstruct which generalizes zero-shot reasoning abilities of LLMs. Here, the agent generates instructions which optimize the reasoning process of LLMs. The proposed approach is compared with zero-shot and zero-shot-CoT as baselines and reports 17.8%  and 6.5% average improvement across the tasks.", "strengths": "The originality of the paper is mainly rooted in its  proposed approach being a cost-effective alternative to zero-shot CoT as the instructions can be generated using a bigger more costly model while the reasoning LLM can be a cheaper alternative. This is a valid argument highlighted by the authors.", "weaknesses": "The agents used a much bigger and powerful LLM, GPT-4, as the default agent to generate the instructions. However, the models which were evaluated on the proposed approach were limited to Vicuna, Llama-2-chat, and GPT-3.5 Turbo. As GPT-4 is missing from the list of models, the evaluation against baseline approaches (zero-shot and zero-shot-CoT) is not a fair comparison because they have not utilized the power of GPT-4. This flaw in experiment design, unfortunately, invalidates the reported results.\n\nSimilarly, the agent is given access to search engines to get the \"top releant web pages containing information about the dataset\". This added information provided to the agent (GPT-4 model) is an extra context which the baseline approaches (zero-shot and zero-shot-CoT) don't have access to. Consequently, the comparison between the proposed AgentInstruct and the baseline methods is not a fair comparison.\n\nIn other words while the approach decouples the \"instruction generation\" from \"reasoning\", the \"instruction generation\" step is utilizing extra context and more powerful model which reasonably contributes to the higher end-to-end performance of the approach. \n\nIn the \"Ablation Study\" the authors have evaluated the impact of removing each of the components of the Zero-Shot AgentInstruct to assert that they are all effective. At the same time this ablation study further highlights the impact of GPT4 in the overall performance of the method (Table 1). \n\nFurthermore as demonstrated in Figure 6, the ZeroShot GPT-4 only marginally lacks behind the Zero-Shot Agent-Instruct (79.5 vs 88.1 or 0.6% improvement) with lower cost ($1 vs. $2). This is a very important observation which further invalidates the sanity and effectiveness of the proposed method because a much simpler approach (i.e., ZeroShot which is the simplest of approaches) is demonstrating almost similar results as the very complicated Zero-Shot AgentInstruct. In the same section the authors mention that \"Though ReAct narrowly outperforms\nzero-shot AgentInstruct\" where narrow corresponds to 0.8% improvement which is larger than the 0.6% improvement of Zero-Shot AgentInstruct over the Zero-Shot GPT-4.", "questions": "In table 1, please elaborate on the first 3 settings: w/o Agent Instructions, w/o Input Examples and w/o Labels. Reading the manuscript and the paragraph that follows Table 1, it is not fully clear what each of these components represent. Particularly the \"Labels\" have not been discussed in the manuscript.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces AgentInstruct which generalizes zero-shot reasoning abilities of LLMs. Here, the agent generates instructions which optimize the reasoning process of LLMs. The proposed approach is compared with zero-shot and zero-shot-CoT as baselines and reports 17.8%  and 6.5% average improvement across the tasks.", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "strengths": "The originality of the paper is mainly rooted in its  proposed approach being a cost-effective alternative to zero-shot CoT as the instructions can be generated using a bigger more costly model while the reasoning LLM can be a cheaper alternative. This is a valid argument highlighted by the authors.", "weaknesses": "The agents used a much bigger and powerful LLM, GPT-4, as the default agent to generate the instructions. However, the models which were evaluated on the proposed approach were limited to Vicuna, Llama-2-chat, and GPT-3.5 Turbo. As GPT-4 is missing from the list of models, the evaluation against baseline approaches (zero-shot and zero-shot-CoT) is not a fair comparison because they have not utilized the power of GPT-4. This flaw in experiment design, unfortunately, invalidates the reported results.\n\nSimilarly, the agent is given access to search engines to get the \"top releant web pages containing information about the dataset\". This added information provided to the agent (GPT-4 model) is an extra context which the baseline approaches (zero-shot and zero-shot-CoT) don't have access to. Consequently, the comparison between the proposed AgentInstruct and the baseline methods is not a fair comparison.\n\nIn other words while the approach decouples the \"instruction generation\" from \"reasoning\", the \"instruction generation\" step is utilizing extra context and more powerful model which reasonably contributes to the higher end-to-end performance of the approach. \n\nIn the \"Ablation Study\" the authors have evaluated the impact of removing each of the components of the Zero-Shot AgentInstruct to assert that they are all effective. At the same time this ablation study further highlights the impact of GPT4 in the overall performance of the method (Table 1). \n\nFurthermore as demonstrated in Figure 6, the ZeroShot GPT-4 only marginally lacks behind the Zero-Shot Agent-Instruct (79.5 vs 88.1 or 0.6% improvement) with lower cost ($1 vs. $2). This is a very important observation which further invalidates the sanity and effectiveness of the proposed method because a much simpler approach (i.e., ZeroShot which is the simplest of approaches) is demonstrating almost similar results as the very complicated Zero-Shot AgentInstruct. In the same section the authors mention that \"Though ReAct narrowly outperforms\nzero-shot AgentInstruct\" where narrow corresponds to 0.8% improvement which is larger than the 0.6% improvement of Zero-Shot AgentInstruct over the Zero-Shot GPT-4.", "questions": "In table 1, please elaborate on the first 3 settings: w/o Agent Instructions, w/o Input Examples and w/o Labels. Reading the manuscript and the paragraph that follows Table 1, it is not fully clear what each of these components represent. Particularly the \"Labels\" have not been discussed in the manuscript.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698516910954}, {"id": "mmhOh99ijk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5899/Reviewer_tm2h"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper introduces a novel approach where Large Language Models (LLMs) are repurposed as agents during reasoning, enhancing the autonomy of the question-answer process. Specifically, the proposed agent-based reasoner autonomously generates task-specific instructions without the need for training, which subsequently aids the LLM in reasoning over the question. The methodology's efficacy is showcased across various tasks (generation, classification, and reasoning) and on models of differing scales, such as Vicuna-13b, Llama-2-70b-chat, and GPT-3.5 Turbo. Notably, the results are commendable.", "review_text": "This paper introduces a novel approach where Large Language Models (LLMs) are repurposed as agents during reasoning, enhancing the autonomy of the question-answer process. Specifically, the proposed agent-based reasoner autonomously generates task-specific instructions without the need for training, which subsequently aids the LLM in reasoning over the question. The methodology's efficacy is showcased across various tasks (generation, classification, and reasoning) and on models of differing scales, such as Vicuna-13b, Llama-2-70b-chat, and GPT-3.5 Turbo. Notably, the results are commendable.", "strengths": "1. The innovative AgentInstruct autonomously streamlines prompt engineering. It significantly reduces the human effort required in designing chain-of-thought exemplars while maintaining high reasoning prowess.\n \n2. AgentInstruct exhibits versatility, as evidenced by its robust performance across LLMs of various scales (13b, 70b, and ~200b), indicating that it isn't limited to a specific scale.\n\n3. The method's adaptability extends to different tasks, including generation, classification, and reasoning, suggesting it isn't merely task-specific.\n\n4. The community stands to benefit immensely from the authors' decision to release the code, as it paves the way for effortless result reproduction and broader application of the method.", "weaknesses": "1. Although the methodology promotes a higher degree of autonomy in the question-answer process, its implementation might be deemed intricate by some, which could be a potential shortcoming (on the contrary, the implementation of simple methods such as CoT is easier).\n\n2. Implementing the method could demand higher computational throughput from the LLM, leading to increased computational costs or API fees.\n\n3. While there's a notable enhancement in the performance of cutting-edge LLMs — Vicuna-13b by 13.3%, Llama-2-70b-chat by 23.2%, and GPT-3.5 Turbo by 17.0% — the performance increment isn't as pronounced with larger models like GPT-3.5 Turbo as it is with Llama-2-70b-chat. This raises concerns about the method's scalability with even larger models. The absence of data on its application to GPT-4, which is larger than GPT-3.5, further intensifies this curiosity about the method's scalability.", "questions": "Please see weaknesses. I would like to update my evaluation after the discussion.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel approach where Large Language Models (LLMs) are repurposed as agents during reasoning, enhancing the autonomy of the question-answer process. Specifically, the proposed agent-based reasoner autonomously generates task-specific instructions without the need for training, which subsequently aids the LLM in reasoning over the question. The methodology's efficacy is showcased across various tasks (generation, classification, and reasoning) and on models of differing scales, such as Vicuna-13b, Llama-2-70b-chat, and GPT-3.5 Turbo. Notably, the results are commendable.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The innovative AgentInstruct autonomously streamlines prompt engineering. It significantly reduces the human effort required in designing chain-of-thought exemplars while maintaining high reasoning prowess.\n \n2. AgentInstruct exhibits versatility, as evidenced by its robust performance across LLMs of various scales (13b, 70b, and ~200b), indicating that it isn't limited to a specific scale.\n\n3. The method's adaptability extends to different tasks, including generation, classification, and reasoning, suggesting it isn't merely task-specific.\n\n4. The community stands to benefit immensely from the authors' decision to release the code, as it paves the way for effortless result reproduction and broader application of the method.", "weaknesses": "1. Although the methodology promotes a higher degree of autonomy in the question-answer process, its implementation might be deemed intricate by some, which could be a potential shortcoming (on the contrary, the implementation of simple methods such as CoT is easier).\n\n2. Implementing the method could demand higher computational throughput from the LLM, leading to increased computational costs or API fees.\n\n3. While there's a notable enhancement in the performance of cutting-edge LLMs — Vicuna-13b by 13.3%, Llama-2-70b-chat by 23.2%, and GPT-3.5 Turbo by 17.0% — the performance increment isn't as pronounced with larger models like GPT-3.5 Turbo as it is with Llama-2-70b-chat. This raises concerns about the method's scalability with even larger models. The absence of data on its application to GPT-4, which is larger than GPT-3.5, further intensifies this curiosity about the method's scalability.", "questions": "Please see weaknesses. I would like to update my evaluation after the discussion.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N.A.", "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698462240811}], "openreview_url": "https://openreview.net/forum?id=zIJFG7wW2d", "arxiv_id": "2310.03710", "paper_pdf": "papers/zIJFG7wW2d.pdf", "paper_pdf_sha256": "5ab79234bea60aed36995a89da6b2a2d8eb8e3223d472f35f13a5b90255fb1ca", "paper_pdf_bytes": 1892943, "paper_pdf_source": "openreview", "code_url": "https://github.com/wang-research-lab/agentinstruct", "code_repository": "wang-research-lab/agentinstruct", "code_commit": "0006a4571280be73c84b05b51a94aedc8ea4768f", "code_archive": "repos/zIJFG7wW2d.zip", "code_archive_sha256": "65ad91611d47f5ef4bd36f0b23ec92939fe9c9d2dda4a6baf850cfdaea728907", "code_archive_bytes": 220542, "code_file_count": 54, "code_extensions": {".py": 45, ".sh": 9}, "github_disk_usage_kb": 162, "github_languages": {"Python": 483273, "Shell": 9321}, "github_archived": false, "github_pushed_at": "2025-10-20T22:45:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/agent-instructs-large-language-models-to-be"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "P63GxgD7LIl", "year": 2023, "status": "rejected", "title": "TransFool: An Adversarial Attack against Neural Machine Translation Models", "authors": ["Sahar Sadrizadeh", "Pascal Frossard", "Ljiljana Dolamic"], "authorids": ["~Sahar_Sadrizadeh1", "~Pascal_Frossard1", "~Ljiljana_Dolamic1"], "authors_source": "OpenReview API", "abstract": "Deep neural networks have been shown to be vulnerable to small perturbations of their inputs known as adversarial attacks. In this paper, we consider the particular task of Neural Machine Translation (NMT), where security is often critical. We investigate the vulnerability of NMT models to adversarial attacks and propose a new attack algorithm called TransFool. It builds on a multi-term optimization problem and a gradient projection step to compute adversarial examples that fool NMT models. By integrating the embedding representation of a language model in the proposed attack, we generate fluent adversarial examples in the source language that maintain a high level of semantic similarity with the clean samples and render the attack largely undetectable. Experimental results demonstrate that, for multiple translation tasks and different NMT architectures, our white-box attack can severely degrade the translation quality for more than 60% of the sentences while the semantic similarity between the original sentence and the adversarial example stays very high. Moreover, we show that the proposed attack is transferable to unknown target models and can fool those quite easily. Finally, our method leads to improvement in terms of success rate, semantic similarity, and fluency compared to the existing attack strategies both in white-box and black-box settings. Hence, TransFool permits to better characterize the vulnerability of NMT systems and outlines the necessity to design strong defense mechanisms and more robust NMT systems for real-life applications.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SB5O_LGPxO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6290/Reviewer_6HVS"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper defines a new optimization objective function which combines the fluency, similarity and translation error to adversarially attack machine translation models. A gradient projection algorithm is applied to solve this optimization. Experiment results show the proposed method outperforms baselines. The transferability is also examined. \n", "review_text": "The novelty of the proposed method is somewhat incremental. \n\nThe experiment part is strong in terms of datasets and models. The automatic evaluation metrics also show strong performance. \n\nI think adding a human evaluation would reveal lots of insights. ", "strengths": "Strength\n\n- The proposed method is simple and straightforward. \n- The experiment results show superior performance in attack success rate and significant decrease in translation quality. \n- This work also demonstrates the capability and efficiency of blackbox attack.\n\n\nweaknesses\n\n- Existing adversarial attack on NMT aims at improving the robustness of these models. However, in this work, achieving a high attack success rate seems to be the goal. Therefore, I’m wondering if TransFool is as effective as baselines in improving robustness. Or what is the desired use case for this method?\n- Translation models use beam search or similar mechanisms to generate high-quality output. Is the proposed attack still effective when using these mechanisms?\n- Missing human validation. Although automatic metrics show significant decrease in translation quality, I’m not convinced that the algorithm triggers incorrect translation. Maybe the translation is correct but has a very low BLEU or chrF score (i.e., the attack method is attacking the automatic metrics instead of the NMT model, which could also be an interesting finding). \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper defines a new optimization objective function which combines the fluency, similarity and translation error to adversarially attack machine translation models. A gradient projection algorithm is applied to solve this optimization. Experiment results show the proposed method outperforms baselines. The transferability is also examined. \n", "strength_and_weaknesses": "Strength\n\n- The proposed method is simple and straightforward. \n- The experiment results show superior performance in attack success rate and significant decrease in translation quality. \n- This work also demonstrates the capability and efficiency of blackbox attack.\n\n\nweaknesses\n\n- Existing adversarial attack on NMT aims at improving the robustness of these models. However, in this work, achieving a high attack success rate seems to be the goal. Therefore, I’m wondering if TransFool is as effective as baselines in improving robustness. Or what is the desired use case for this method?\n- Translation models use beam search or similar mechanisms to generate high-quality output. Is the proposed attack still effective when using these mechanisms?\n- Missing human validation. Although automatic metrics show significant decrease in translation quality, I’m not convinced that the algorithm triggers incorrect translation. Maybe the translation is correct but has a very low BLEU or chrF score (i.e., the attack method is attacking the automatic metrics instead of the NMT model, which could also be an interesting finding). \n", "clarity,_quality,_novelty_and_reproducibility": "The space after figures and tables are being squeezed too much. \n", "summary_of_the_review": "The novelty of the proposed method is somewhat incremental. \n\nThe experiment part is strong in terms of datasets and models. The automatic evaluation metrics also show strong performance. \n\nI think adding a human evaluation would reveal lots of insights. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666726972902}, {"id": "NwvnPv0fRsx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6290/Reviewer_Jg82"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new approach for adversarial attacks on NMT models.\nThe authors design a way to propagate the gradient for embeddings and propose a specific loss that targets reasonable criteria.\nThe propagation is based on the common idea of relaxation with a new additional FC part to create embeddings for comparison.\nThe metrics are good, and overall, the results are promising.", "review_text": "The results seem promising, on the other side, the novelty is limited. From the paper, it is not clear, why this approach works in the NMT-attacks field, where other approaches failed.", "strengths": "Strengths:\n- Quality metrics are better than that of competitors.\n- Experiments are interesting and numerous. They include human evaluation as well.\n- Nice adversarial example found in Table 2\n\nWeakness, methods:\n- The novelty is limited. The paper contains some engineering tricks to make everything work, but all ideas are similar to those previously discussed in the literature.\n- Not clear why we need the FC part in the approach. We can use BERT-like architectures to generate similar differentiable $v_i$-s as well. What would be the difference in this case?\n- Investigation on the distortion of the generated adversarial sequence $x'$ is limited and includes only \"Semantic Similarity\" from Yang et al. Can you include metrics like BERT score for a pair of $x$ and $x'$ as well?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new approach for adversarial attacks on NMT models.\nThe authors design a way to propagate the gradient for embeddings and propose a specific loss that targets reasonable criteria.\nThe propagation is based on the common idea of relaxation with a new additional FC part to create embeddings for comparison.\nThe metrics are good, and overall, the results are promising.", "strength_and_weaknesses": "Strengths:\n- Quality metrics are better than that of competitors.\n- Experiments are interesting and numerous. They include human evaluation as well.\n- Nice adversarial example found in Table 2\n\nWeakness, methods:\n- The novelty is limited. The paper contains some engineering tricks to make everything work, but all ideas are similar to those previously discussed in the literature.\n- Not clear why we need the FC part in the approach. We can use BERT-like architectures to generate similar differentiable $v_i$-s as well. What would be the difference in this case?\n- Investigation on the distortion of the generated adversarial sequence $x'$ is limited and includes only \"Semantic Similarity\" from Yang et al. Can you include metrics like BERT score for a pair of $x$ and $x'$ as well?\n", "clarity,_quality,_novelty_and_reproducibility": "- New loss is proposed that consists of three terms. The ideas for these are out there for some time.\n- New architecture is proposed to find adversarial sequences. It is an interesting modification of a common one.\n\nI also imagine that the observed behaviour can be related to the requirement to include a projection layer in self-supervised learning or in other settings.", "summary_of_the_review": "The results seem promising, on the other side, the novelty is limited. From the paper, it is not clear, why this approach works in the NMT-attacks field, where other approaches failed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666614269024}, {"id": "Bk9krBqi4fR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6290/Reviewer_ZrSf"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work proposes TransFool for generating non-targeted adversarial examples against neural machine translation models. One core idea is to utilize an autoregressive language model (GPT-2) to add a language model loss term, which helps generate fluent adversarial examples. They also add a similarity loss, which constrains the distances of embedding distances between the original input sentences and adversarial examples. They perform a comprehensive study to evaluate their approach on translating from English to different target languages, and show that TransFool achieves higher attack success rates compared to baselines, while the generated adversarial examples better preserve the semantic meaning and look more natural. Meanwhile, they show that the generated adversarial examples transfer to other models in the black-box setting, including Google Translate, and can also transfer to different target languages.", "review_text": "This work proposes a well-designed adversarial attack for neural machine translation models, though it is not entirely novel. The paper presents a comprehensive study in different attack settings. In particular, the transferability of attacks between different target languages is interesting and new. Therefore, I lean towards accepting this work.\n\n-------------\nI thank the authors for explanation and adding more experiments, and I keep my original score.", "strengths": "Strengths:\n1. Each term in the loss function of TransFool is well-motivated and properly designed.\n\n2. The empirical study is pretty thorough. To me, the most interesting finding is that the generated adversarial examples can transfer to different target languages. \n\nWeaknesses:\n\n1. The study of defense is lacking. For example, it is helpful to see how the attack works with existing defenses against adversarial examples for language models. Also, it is interesting to try adversarial training with TransFool adversarial examples.\n\n2. This is more of a question rather than a weakness, but have you tried TransFool for targeted attacks, and how does that work?\n\n3. It is good to investigate more into the transferability between different target languages. For example, have you done these experiments on Google Translate?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This work proposes TransFool for generating non-targeted adversarial examples against neural machine translation models. One core idea is to utilize an autoregressive language model (GPT-2) to add a language model loss term, which helps generate fluent adversarial examples. They also add a similarity loss, which constrains the distances of embedding distances between the original input sentences and adversarial examples. They perform a comprehensive study to evaluate their approach on translating from English to different target languages, and show that TransFool achieves higher attack success rates compared to baselines, while the generated adversarial examples better preserve the semantic meaning and look more natural. Meanwhile, they show that the generated adversarial examples transfer to other models in the black-box setting, including Google Translate, and can also transfer to different target languages.", "strength_and_weaknesses": "Strengths:\n1. Each term in the loss function of TransFool is well-motivated and properly designed.\n\n2. The empirical study is pretty thorough. To me, the most interesting finding is that the generated adversarial examples can transfer to different target languages. \n\nWeaknesses:\n\n1. The study of defense is lacking. For example, it is helpful to see how the attack works with existing defenses against adversarial examples for language models. Also, it is interesting to try adversarial training with TransFool adversarial examples.\n\n2. This is more of a question rather than a weakness, but have you tried TransFool for targeted attacks, and how does that work?\n\n3. It is good to investigate more into the transferability between different target languages. For example, have you done these experiments on Google Translate?", "clarity,_quality,_novelty_and_reproducibility": "The writing is clear, and the approach is well-designed.\n\nTransFool is novel as an adversarial attack algorithm for machine translation. The individual terms are not that new as there are prior works leveraging language models to improve the fluency of adversarial examples in the NLP domain, but this work provides a more comprehensive study for machine translation problems. \n\nThe authors plan to release their code for reproducibility.", "summary_of_the_review": "This work proposes a well-designed adversarial attack for neural machine translation models, though it is not entirely novel. The paper presents a comprehensive study in different attack settings. In particular, the transferability of attacks between different target languages is interesting and new. Therefore, I lean towards accepting this work.\n\n-------------\nI thank the authors for explanation and adding more experiments, and I keep my original score.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666607074302}, {"id": "canPogon_3j", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6290/Reviewer_8Cx6"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a white-box adversarial attack against neural machine translation model named TransFool. TransFool uses an optimization loss function with three terms: a) adversarial loss to maximize the loss of the target NMT model; b) a similarity term to ensure that the adversarial example is similar to the original sentence; c) and loss of a language model to generate fluent and natural adversarial examples. The authors compare their TransFool method with two NMT attack methods, kNN and seq2sick, with respect to the Attack Success Rate, Relative decrease of translation quality, Semantic Similarity, Perplexity score and Token Error Rate. The proposed method can outperform baseline methods on two NMT model, Marian NMT and mBART50 MNMT model. The authors also show that the adversarial examples they found can conduct transfer attack on black box NMT models, using the adversarial examples from Marian NMT to attack mBART50 MNMT model and google translation api.", "review_text": "This paper proposes an adversarial attack method against neural machine translation model. The authors conduct experiments to show the superiority of their method. However, the novelty and the soundness of this paper is limited. So I tend to reject this paper.", "strengths": "Pros:\n+ This paper is good written and easy to follow.\n+ The authors valid their method not only on research models but also on real world commercial product, like google translation.\n+ The transferability analysis section is enlightening. It is not widely studied in NLP field. The transferability between different language is an important issue in multilingual NLP filed.\n\n\nCons:\n- Novelty is limited. \na) First, the authors consider not only the attack success rate but also the fluency and semantic similarity. However, these two issues are widely studied in NLP attack field. For example, [1] has systematically reveal that the fluency and semantic similarity issue in textual adversarial attack and they empirically study the threshold of filtering the adversarial examples according to the fluency and semantic similarity. \nb) Second, this paper claims that they propose a new strategy to incorporate the embedding vectors of a language model. However, utilizing language model to attack NLP model is not a novel technique. For example, [2-3] utilize BERT to conduct word-level substitution.\n\n- Lack of soundness. This is not only the weakness of this paper but also the weakness of a large portion of textual adversarial attack works.\na) The soundness of their method is weak. This paper use language model due to it provide a meaningful representation of the tokens. However, based on the finding of previous work [1], the semantic similarity cannot be guaranteed with language model. A good example is that “I [like] eating apple.” and “I [hate] eating apple” have very high representation similarity according to the language model representation. But their semantic are totally opposite. So I am worried that the adversarial attack will change the semantic meaning of source sentence, leading to the over-estimation of attack success rate.\nb) The soundness of their evaluation is weak. The authors evaluate the semantic similarity by the universal sentence encoder and BERTScore. However, as I mentioned above, the soundness of such model-based automatic metrics is limited. The high attack success rate could be partially due to the changing of meaning in source sentence. Conducting a human annotation experiment will be better than just showing an example.\n\n[1] John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, Yanjun Qi ,Reevaluating Adversarial Examples in Natural Language, EMNLP 2020.\n[2] Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, Xipeng Qiu, BERT-ATTACK: Adversarial Attack Against BERT Using BERT, EMNLP 2020.\n[3] Siddhant Garg, Goutham Ramakrishnan, BAE: BERT-based Adversarial Examples for Text Classification, EMNLP 2020.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes a white-box adversarial attack against neural machine translation model named TransFool. TransFool uses an optimization loss function with three terms: a) adversarial loss to maximize the loss of the target NMT model; b) a similarity term to ensure that the adversarial example is similar to the original sentence; c) and loss of a language model to generate fluent and natural adversarial examples. The authors compare their TransFool method with two NMT attack methods, kNN and seq2sick, with respect to the Attack Success Rate, Relative decrease of translation quality, Semantic Similarity, Perplexity score and Token Error Rate. The proposed method can outperform baseline methods on two NMT model, Marian NMT and mBART50 MNMT model. The authors also show that the adversarial examples they found can conduct transfer attack on black box NMT models, using the adversarial examples from Marian NMT to attack mBART50 MNMT model and google translation api.", "strength_and_weaknesses": "Pros:\n+ This paper is good written and easy to follow.\n+ The authors valid their method not only on research models but also on real world commercial product, like google translation.\n+ The transferability analysis section is enlightening. It is not widely studied in NLP field. The transferability between different language is an important issue in multilingual NLP filed.\n\n\nCons:\n- Novelty is limited. \na) First, the authors consider not only the attack success rate but also the fluency and semantic similarity. However, these two issues are widely studied in NLP attack field. For example, [1] has systematically reveal that the fluency and semantic similarity issue in textual adversarial attack and they empirically study the threshold of filtering the adversarial examples according to the fluency and semantic similarity. \nb) Second, this paper claims that they propose a new strategy to incorporate the embedding vectors of a language model. However, utilizing language model to attack NLP model is not a novel technique. For example, [2-3] utilize BERT to conduct word-level substitution.\n\n- Lack of soundness. This is not only the weakness of this paper but also the weakness of a large portion of textual adversarial attack works.\na) The soundness of their method is weak. This paper use language model due to it provide a meaningful representation of the tokens. However, based on the finding of previous work [1], the semantic similarity cannot be guaranteed with language model. A good example is that “I [like] eating apple.” and “I [hate] eating apple” have very high representation similarity according to the language model representation. But their semantic are totally opposite. So I am worried that the adversarial attack will change the semantic meaning of source sentence, leading to the over-estimation of attack success rate.\nb) The soundness of their evaluation is weak. The authors evaluate the semantic similarity by the universal sentence encoder and BERTScore. However, as I mentioned above, the soundness of such model-based automatic metrics is limited. The high attack success rate could be partially due to the changing of meaning in source sentence. Conducting a human annotation experiment will be better than just showing an example.\n\n[1] John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, Yanjun Qi ,Reevaluating Adversarial Examples in Natural Language, EMNLP 2020.\n[2] Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, Xipeng Qiu, BERT-ATTACK: Adversarial Attack Against BERT Using BERT, EMNLP 2020.\n[3] Siddhant Garg, Goutham Ramakrishnan, BAE: BERT-based Adversarial Examples for Text Classification, EMNLP 2020.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: This paper is well-written and easy to follow.\nQuality: The soundness of their method and evaluation is limited. Please refer to the weakness part.\nNovelty: This paper is not very novel. Please refer to the weakness part.\nReproducibility: The reproducibility will not be a issue. The authors claims that their code will be released. And they provide the link of the metric and scripts used in their paper.\n", "summary_of_the_review": "This paper proposes an adversarial attack method against neural machine translation model. The authors conduct experiments to show the superiority of their method. However, the novelty and the soundness of this paper is limited. So I tend to reject this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666577382745}], "openreview_url": "https://openreview.net/forum?id=P63GxgD7LIl", "arxiv_id": "2302.00944", "paper_pdf": "papers/P63GxgD7LIl.pdf", "paper_pdf_sha256": "e4a0eee24850a30ad1abcb089275af40c10362e7acfb35a67396ba9f305697d1", "paper_pdf_bytes": 2069662, "paper_pdf_source": "openreview", "code_url": "https://github.com/sssadrizadeh/TransFool", "code_repository": "sssadrizadeh/TransFool", "code_commit": "8b0d9de5c6f60c4017e5ddcc81c8f14dd5c9a85b", "code_archive": "repos/P63GxgD7LIl.zip", "code_archive_sha256": "f9ea93a3097aab4c296a260a78e39ab8d168afa04ae76eea1a07fc1b47819e5b", "code_archive_bytes": 373518, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 350, "github_languages": {"Python": 153487}, "github_archived": false, "github_pushed_at": "2024-06-26T12:59:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transfool-an-adversarial-attack-against"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qSTEPv2uLR8", "year": 2022, "status": "rejected", "title": "Physics Informed Convex Artificial Neural Networks (PICANNs) for Optimal Transport based Density Estimation", "authors": ["Amanpreet Singh", "Martin Bauer", "Sarang Joshi"], "authorids": ["~Amanpreet_Singh3", "~Martin_Bauer2", "~Sarang_Joshi1"], "authors_source": "OpenReview API", "abstract": "Optimal Mass Transport (OMT) is a well studied problem with a variety of applications in a  diverse set of fields ranging from Physics to Computer Vision and in particular Statistics and Data Science. Since the original formulation of Monge in 1781 significant theoretical progress been made on the existence, uniqueness and properties of the optimal transport maps. The actual numerical computation of the transport maps, particularly in high dimensions, remains a challenging problem. In the past decade several neural network based algorithms have been proposed to tackle this task. In this paper, building on recent developments of input convex neural networks and physics informed neural networks for solving PDE's, we propose a new Deep Learning approach to solve the continuous OMT problem. Our framework is based on Brenier's theorem, which reduces the continuous OMT problem to that of solving a non-linear PDE of Monge-Ampere type whose solution is a convex function. To demonstrate the accuracy of our framework we compare our method to several other deep learning based algorithms. We then focus on applications to the ubiquitous density estimation and generative modeling tasks in statistics and machine learning. Finally as an example we present how our framework can be incorporated with an autoencoder to estimate an effective probabilistic generative model.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "fvBfUC5EI9V", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1682/Reviewer_U1ba"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes to solve the continuous Optimal Mass Transport problem as a convex mapping function estimation problem based on Brenier's theorem using input convex neural networks.\nThe proposed method is compared with other DNN-based methods in the ubiquitous density estimation and generative modeling tasks in statistics, and good results are obtained.", "review_text": "The proposed method in this paper has the potential to be a very useful method for solving the continuous optimal transport problem using DNNs.\nIn addition, the manuscript is clear and well constructed.\nOn the other hand, the effectiveness of the method for application problems in machine learning cannot be confirmed because there are few implications about real applications and few experiments.\nIn order for the paper to be accepted at this conference, it might be necessary to apply the method to a real problem such as the image generation problem of natural images, and then conduct experiments to compare the method with existing methods which solved the application problem.\nIn addition to that, the following points were unclear and I would like to ask for your response.\n\nDoes ICNN have the ability to represent all convex functions?\n\nOn page5, it says \"L^2-norm used in equation 5\", but isn't it a mistake of \"equation 4\"?\n\nIn the experiment of method comparison, is there any evidence that the parameters of DNN in the previous study are optimized? (If there is no evidence, it would be better to specify the possibility that it is not optimized?　That would be positively evaluated as a scientific paper, as it would present the limits of verification).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to solve the continuous Optimal Mass Transport problem as a convex mapping function estimation problem based on Brenier's theorem using input convex neural networks.\nThe proposed method is compared with other DNN-based methods in the ubiquitous density estimation and generative modeling tasks in statistics, and good results are obtained.", "main_review": "The proposed method in this paper has the potential to be a very useful method for solving the continuous optimal transport problem using DNNs.\nIn addition, the manuscript is clear and well constructed.\nOn the other hand, the effectiveness of the method for application problems in machine learning cannot be confirmed because there are few implications about real applications and few experiments.\nIn order for the paper to be accepted at this conference, it might be necessary to apply the method to a real problem such as the image generation problem of natural images, and then conduct experiments to compare the method with existing methods which solved the application problem.\nIn addition to that, the following points were unclear and I would like to ask for your response.\n\nDoes ICNN have the ability to represent all convex functions?\n\nOn page5, it says \"L^2-norm used in equation 5\", but isn't it a mistake of \"equation 4\"?\n\nIn the experiment of method comparison, is there any evidence that the parameters of DNN in the previous study are optimized? (If there is no evidence, it would be better to specify the possibility that it is not optimized?　That would be positively evaluated as a scientific paper, as it would present the limits of verification).", "summary_of_the_review": "The proposed method in this paper has the potential to be a very useful method for solving the continuous optimal transport problem using DNNs.\nIn addition, the manuscript is clear and well constructed.\nOn the other hand, the effectiveness of the method for application problems in machine learning cannot be confirmed because there are few implications about real applications and few experiments.\nIn order for the paper to be accepted at this conference, it might be necessary to apply the method to a real problem such as the image generation problem of natural images, and then conduct experiments to compare the method with existing methods which solved the application problem.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636111511924}, {"id": "spqetNouau-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1682/Reviewer_44BN"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new physics informed neural architecture for efficiently learning and optimizing optimal transport maps using techniques from partial differential equations.", "review_text": "Strength: + The probabilistic universal approximation framework can be much more efficient than the finite element or difference methods which is known to have exponential scaling with respect to dimensions.\n+ The paper provides sufficient details in terms of network, hyperparameters, evaluation methods in the experimental section.\n\nWeakness: - The main required definitions are provided, however, the main relevant technical concepts are explained heuristically. For example, while relevant background for optimal transport theory is provided in detail, there is little to no background on partial differential equations.\n- The main loss functions (4) and (5) proposed do have exponential scaling with respect to dimensions of the domain of integration \\Omega.\n- Density estimation experiments are preliminary.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new physics informed neural architecture for efficiently learning and optimizing optimal transport maps using techniques from partial differential equations.", "main_review": "Strength: + The probabilistic universal approximation framework can be much more efficient than the finite element or difference methods which is known to have exponential scaling with respect to dimensions.\n+ The paper provides sufficient details in terms of network, hyperparameters, evaluation methods in the experimental section.\n\nWeakness: - The main required definitions are provided, however, the main relevant technical concepts are explained heuristically. For example, while relevant background for optimal transport theory is provided in detail, there is little to no background on partial differential equations.\n- The main loss functions (4) and (5) proposed do have exponential scaling with respect to dimensions of the domain of integration \\Omega.\n- Density estimation experiments are preliminary.\n", "summary_of_the_review": "Justification: The paper makes a solid contribution in terms of technical content. The main high level idea is that solving optimal transport maps under some conditions can be posed as a PDE. While this is not something new, for instance, sinkhorn algorithm does the same -- turn a nonsmooth optimization problem, a linear program, into a smooth version, as given in equations (4) and (5). However, many technical details regarding the algorithm are missing -- (i) what is the overall procedure?; (ii) near equation (7), there is some discussion regarding representing u again using a neural network, is the overall architecture different from what is shown in Figure (1)? It seems like some of these questions regarding how and what PDE is exactly being solved can be answered by reading various other papers cited in the experiments, but it is not clear from the submission. In this sense, I feel like the paper spends too much time on background material (which the reader can pick up from any standard book on Applied PDE, for example, Introduction in Optimal\nTransport for Applied Mathematicians by Filippo Santambrogio, 2015) rather than focusing on the key insights of the paper. \n\nThe network details are provided in abundant details for verification of the main idea. However, the main task density estimation which the paper mentions quite in detail about in the beginning of the paper is only minimally experimented. For example, the paper claims that finite elements solvers are slow in high dimensions. However, the paper contains no comparisons with these methods on the problems considered in the paper, which are also low dimensional.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635908625910}, {"id": "6PCBxHdjiYu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1682/Reviewer_U76A"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new method for density estimation that is based on the idea of optimal mass transport (OMP), i.e., finding some function that transports a probability density into another probability density while minimizing the transportation cost. To do so, the authors leverage Brenier's theorem which guarantees the existence and uniqueness of the optimal transport map. As a consequence of this theorem,  the optimal transport map of the OMT problem can be obtained as the solution of a special nonlinear partial differential equation (PDE). The authors then show that the approximate solution can be computed with help of physics informed neural networks (PINNs). The approach is demonstrated for several canonical examples.", "review_text": "The paper combines a range of interesting ideas to solve density estimation problems. The authors reformulate the original OMP problem as a special PDE. In turn, PINNs can be used to approximate the solution of this special PDE. In addition, the authors use input convex neural networks (ICNNs) to  incorporate the Brenier potential into the PINN. This approach bypasses some of the challenges of directly minimizing the original OMP problem.\n\n* _Originality:_ The presented ideas in this paper are original and well suited for this conference. I not aware of another paper that uses PINNs for density estimation. \n\n* _Quality:_ The proposed framework seems plausible and is demonstrated on a few experiments. There is no theory that backups some of the claims, and the amount of experiments is on the lower end of the spectrum for an empirical ML conference paper. Further, it would help to better discuss the assumption that have been made and some of the limitations. I assume that the authors have a fair understanding for these details and they should make better use of the unlimited space of the appendix to provide additional discussions and experiments. \n\n* _Clarity:_ The proposed ideas are well organized and clearly presented. The detailed discussion of the background materials helped reading the paper. \n\n* _Significance:_ This paper is of interest for two communities within ML. First, it is of interest for the PINN community and the paper has the potential to motivate future research in the intersection of PINNs and density estimation. On the other hand, density estimation is an important topic in ML and the presented approach performs well on the presented tasks.However, it is not clear whether the presented approach is competitive with other state-of-the-art methods in this area, or whether it just does better on the selected tasks compared to 2 baselines. A better discussion of the advantages and limitations and additional baselines would help to strengthen the paper and improve its significance. \n\n\nHere are a few comments that might help to improve the paper:\n\n* The weakness of this paper is the experimental section. Only two other algorithms are used for comparison. The authors state that these two algorithms performed best on their benchmark problems, still it would be interesting to show results for other algorithms that have been tested. Also, how does your approach compare to state-of-the-art GANs or neural ordinary differential equations (NODEs) for density estimation methods? At least, it would be helpful to discuss some qualitative advantages / disadvantages of your proposed approach in comparison to other methods such as GANs or NODEs. In which situation should one use your approach, and which situation does your approach fail? \n\n* You state that you only consider the simple case of a quadratic cost function. Does Brenier's Theorem apply to other cost functions?\n\n* You state that ICNNs introduce implicit regularization. Can you provide theory to back this claim up, or is this only an educated guess?\n\n* The results in Figure 6 look interesting. Why don't you include these results in the main text? I would rather move the network details to the appendix.   \n\n* Can you comment on the computational costs of your framework as compared to the other methods that you used for comparison.\n\n* On page 2: I assume that there is some typo in:  'the convex function itself should be indenity helps in avoiding'.\n\n* On page 9: word repetition in 'a density estimation estimation'.\n\n* The quality of the figures is poor. It would be nice to better format the figures so that they eat up less space and use vector graphics instead of low-resolution pngs. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a new method for density estimation that is based on the idea of optimal mass transport (OMP), i.e., finding some function that transports a probability density into another probability density while minimizing the transportation cost. To do so, the authors leverage Brenier's theorem which guarantees the existence and uniqueness of the optimal transport map. As a consequence of this theorem,  the optimal transport map of the OMT problem can be obtained as the solution of a special nonlinear partial differential equation (PDE). The authors then show that the approximate solution can be computed with help of physics informed neural networks (PINNs). The approach is demonstrated for several canonical examples.", "main_review": "The paper combines a range of interesting ideas to solve density estimation problems. The authors reformulate the original OMP problem as a special PDE. In turn, PINNs can be used to approximate the solution of this special PDE. In addition, the authors use input convex neural networks (ICNNs) to  incorporate the Brenier potential into the PINN. This approach bypasses some of the challenges of directly minimizing the original OMP problem.\n\n* _Originality:_ The presented ideas in this paper are original and well suited for this conference. I not aware of another paper that uses PINNs for density estimation. \n\n* _Quality:_ The proposed framework seems plausible and is demonstrated on a few experiments. There is no theory that backups some of the claims, and the amount of experiments is on the lower end of the spectrum for an empirical ML conference paper. Further, it would help to better discuss the assumption that have been made and some of the limitations. I assume that the authors have a fair understanding for these details and they should make better use of the unlimited space of the appendix to provide additional discussions and experiments. \n\n* _Clarity:_ The proposed ideas are well organized and clearly presented. The detailed discussion of the background materials helped reading the paper. \n\n* _Significance:_ This paper is of interest for two communities within ML. First, it is of interest for the PINN community and the paper has the potential to motivate future research in the intersection of PINNs and density estimation. On the other hand, density estimation is an important topic in ML and the presented approach performs well on the presented tasks.However, it is not clear whether the presented approach is competitive with other state-of-the-art methods in this area, or whether it just does better on the selected tasks compared to 2 baselines. A better discussion of the advantages and limitations and additional baselines would help to strengthen the paper and improve its significance. \n\n\nHere are a few comments that might help to improve the paper:\n\n* The weakness of this paper is the experimental section. Only two other algorithms are used for comparison. The authors state that these two algorithms performed best on their benchmark problems, still it would be interesting to show results for other algorithms that have been tested. Also, how does your approach compare to state-of-the-art GANs or neural ordinary differential equations (NODEs) for density estimation methods? At least, it would be helpful to discuss some qualitative advantages / disadvantages of your proposed approach in comparison to other methods such as GANs or NODEs. In which situation should one use your approach, and which situation does your approach fail? \n\n* You state that you only consider the simple case of a quadratic cost function. Does Brenier's Theorem apply to other cost functions?\n\n* You state that ICNNs introduce implicit regularization. Can you provide theory to back this claim up, or is this only an educated guess?\n\n* The results in Figure 6 look interesting. Why don't you include these results in the main text? I would rather move the network details to the appendix.   \n\n* Can you comment on the computational costs of your framework as compared to the other methods that you used for comparison.\n\n* On page 2: I assume that there is some typo in:  'the convex function itself should be indenity helps in avoiding'.\n\n* On page 9: word repetition in 'a density estimation estimation'.\n\n* The quality of the figures is poor. It would be nice to better format the figures so that they eat up less space and use vector graphics instead of low-resolution pngs. ", "summary_of_the_review": "This paper presents interesting and novel ideas for density estimation, and to the best of my knowledge this is the first paper that uses PINNs for density estimation. The paper is clearly written and provides a good overview of related work and background to put the proposed ideas into context. However, since this paper provides only empirical results and no theory, I only think that the paper is slightly above the acceptance threshold. Additional results would help to make the story more compelling, and given additional results I am happy to reconsider my rating. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635825800995}, {"id": "Eyi67KNqhv9", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1682/Reviewer_ecZV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presented an input convex neural networks (ICNN)-based methods to approximate the optimal transport maps. The learning objective is based on the Monge-Ampere equation, i.e. zero loss corresponds to the solution to the Monge-Ampere equation. Physics Informed Neural Networks (PINNs) are adopted to minimise the loss. The author(s) also extended the work to density estimation. The problem is interesting as estimation of optimal maps has a wide range of applications. Experiment results look promising when compared with other PINNs in one experiment setup, but only qualitative results are provided for the density estimation experiment with no comparison of other baselines.", "review_text": "The paper is in general well-written and the idea is clearly presented. Yet, I have a few concerns as listed below:\n\n1. The authors stated that \"We will, however, see, that in our situation, i.e., for absolutely continuous measures, the existence and uniqueness is indeed guaranteed\". This refers to the Brenier's theorem concerning that a unique optimal transport map exists which is given by the gradient of a convex function. This consequently leads to the Monge-Ampere equation that help formulate the loss function in the proposed work. Yet, the physics informed neural networks adopted can only be shown to be a convex function, but there is no guarantee that the optimized network is indeed the solution of the Monge-Ampere equation. Hence, the uniqueness of the map does not seem to be guaranteed. If the alternative is true, a proof is needed here.\n\n2. The presentation of the density estimation in Section 2.3 is a bit confusing. What does it mean to be \"or equivalently $\\bigtriangledown u)_* \\mu = \\nu$\"? A bigger question is whether the proposed work can be used in density estimation, given that the uniqueness of the mapping is not guaranteed as in Comment 1 above.\n\n3. Although literature review covers other density estimation techniques such as normalizing flow, there is no comparison with any of these algorithms in Section 3.3. The authors stated in Section 1 that \"Although these methods (normalizing flow) have shown to perform well in density estimation applications, the interpretability of the obtained transformation is less clear\". But it is not clear to me how the interpretability of the proposed method is clearer, especially given the fact that the computed density may not be exact/correct (due to the lack of the proof of the uniqueness of the obtained approximate mapping), while normalizing flow admits exact density estimation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presented an input convex neural networks (ICNN)-based methods to approximate the optimal transport maps. The learning objective is based on the Monge-Ampere equation, i.e. zero loss corresponds to the solution to the Monge-Ampere equation. Physics Informed Neural Networks (PINNs) are adopted to minimise the loss. The author(s) also extended the work to density estimation. The problem is interesting as estimation of optimal maps has a wide range of applications. Experiment results look promising when compared with other PINNs in one experiment setup, but only qualitative results are provided for the density estimation experiment with no comparison of other baselines.", "main_review": "The paper is in general well-written and the idea is clearly presented. Yet, I have a few concerns as listed below:\n\n1. The authors stated that \"We will, however, see, that in our situation, i.e., for absolutely continuous measures, the existence and uniqueness is indeed guaranteed\". This refers to the Brenier's theorem concerning that a unique optimal transport map exists which is given by the gradient of a convex function. This consequently leads to the Monge-Ampere equation that help formulate the loss function in the proposed work. Yet, the physics informed neural networks adopted can only be shown to be a convex function, but there is no guarantee that the optimized network is indeed the solution of the Monge-Ampere equation. Hence, the uniqueness of the map does not seem to be guaranteed. If the alternative is true, a proof is needed here.\n\n2. The presentation of the density estimation in Section 2.3 is a bit confusing. What does it mean to be \"or equivalently $\\bigtriangledown u)_* \\mu = \\nu$\"? A bigger question is whether the proposed work can be used in density estimation, given that the uniqueness of the mapping is not guaranteed as in Comment 1 above.\n\n3. Although literature review covers other density estimation techniques such as normalizing flow, there is no comparison with any of these algorithms in Section 3.3. The authors stated in Section 1 that \"Although these methods (normalizing flow) have shown to perform well in density estimation applications, the interpretability of the obtained transformation is less clear\". But it is not clear to me how the interpretability of the proposed method is clearer, especially given the fact that the computed density may not be exact/correct (due to the lack of the proof of the uniqueness of the obtained approximate mapping), while normalizing flow admits exact density estimation.", "summary_of_the_review": "The paper provides an interesting approach to incorporate ICNN and Brenier's Theorem to approximate optimal transport between distributions. Experiment results are encouraging. But there is no proof to support the claim on the uniqueness of the mapping. Therefore, its application on density estimation is also not fully supported.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635785415955}], "openreview_url": "https://openreview.net/forum?id=qSTEPv2uLR8", "arxiv_id": "2104.01194", "paper_pdf": "papers/qSTEPv2uLR8.pdf", "paper_pdf_sha256": "bb006b47bb4a52f4e3ea441a4edf779a4cff57f9bd7aa4da993f61ca1a286a33", "paper_pdf_bytes": 1660285, "paper_pdf_source": "openreview", "code_url": "https://github.com/4m4npr33t/PICANNs", "code_repository": "4m4npr33t/PICANNs", "code_commit": "a1d52dab402427414f6f38ce0b5d9e47fafb58fa", "code_archive": "repos/qSTEPv2uLR8.zip", "code_archive_sha256": "291338b716a7d0084f09c094d2ce76b7f64fc6aadc994738b83263c560eacecd", "code_archive_bytes": 1176871, "code_file_count": 18, "code_extensions": {".py": 17, ".sh": 1}, "github_disk_usage_kb": 1199, "github_languages": {"Python": 86788, "Shell": 494}, "github_archived": false, "github_pushed_at": "2021-04-02T18:50:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/physics-informed-convex-artificial-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sAX7Z7uIJ_Y", "year": 2021, "status": "rejected", "title": "Calibrated Adversarial Refinement for Stochastic Semantic Segmentation", "authors": ["Elias Kassapis", "Georgi Dikov", "Deepak Gupta", "Cedric Nugteren"], "authorids": ["~Elias_Kassapis1", "~Georgi_Dikov1", "~Deepak_Gupta2", "~Cedric_Nugteren1"], "authors_source": "OpenReview API", "abstract": "Ambiguities in images or unsystematic annotation can lead to multiple valid solutions in semantic segmentation. To learn a distribution over predictions, recent work has explored the use of probabilistic networks. However, these do not necessarily capture the empirical distribution accurately. In this work, we aim to learn a calibrated multimodal predictive distribution, where the empirical frequency of the sampled predictions closely reflects that of the corresponding labels in the training set. To this end, we propose a novel two-stage, cascaded strategy for calibrated adversarial refinement. In the first stage, we explicitly model the data with a categorical likelihood. In the second, we train an adversarial network to sample from it an arbitrary number of coherent predictions. The model can be used independently or integrated into any black-box segmentation framework to facilitate learning of calibrated stochastic mappings. We demonstrate the utility and versatility of the approach by attaining state-of-the-art results on the multigrader LIDC dataset and a modified Cityscapes dataset. In addition, we use a toy regression dataset to show that our framework is not confined to semantic segmentation, and the core design can be adapted to other tasks requiring learning a calibrated predictive distribution.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "XtO7NsY2YdT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3116/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this work the authors propose a two-stage, adversarial training technique to calibrate a semantic segmentation model when faced with conflicting labels in the train set.  Their approach is to first train a segmentation model. Then, it is used to feed a GAN model which itself is trained against a discriminator to produce diverse segmentations that reflect the diversity in the train set.  The authors compare their approach to similar methods on a synthetic data set and two semantic segmentation data sets. \n\nPros:\n1) The technique is conceptually simple as training a segmentation model with cross entropy loss as well as training a GAN are both well-understood.\n2) The technique can really be used to calibrate any differentiable semantic segmentation model, regardless of who trained it or how.\n3) The experimental details are described in enough detail to likely be able to replicate most of the results\n4) Minus some organizational issues, the writing is good overall.\n\n\nCons:\n1) I found the motivation in the introduction to be slightly difficult to follow.  More specifically, the concepts ambiguity in data labels leading to a multi-modal data distribution and the need to model uncertainty need to be more tightly discussed.  As it is written now, the authors first argue for modeling a noisy empirical distribution that captures the ambiguities, which seems counter intuitive since there assumedly exists a single true segmentation of the image then later discuss uncertainty.  I think discussing uncertainty modeling, and specifically calibration, first would alleviate this issue.\n2) \"Isola et al. (2017) have demonstrated that it introduces only minor stochasticity in the output and returns inconsistent samples.\" (page 2) - In Isola et al. (2017)  dropout is used for image to image translation in a GAN.  I do not think this provides sufficient evidence that using dropout BNNs  such as in Kendal and Gal (2017) has these properties.\n3) The authors argue that combining the GAN loss and the pixel-wise cross entropy is not well-suited for noisy data because they are often at odds.  It's not clear to me that this is a problem in practice, as the cross entropy will spread probability mass across different conflicting labeled instances.  Assume a pixel in the data set appears twice with two different labels.  The sum of the cross entropy over these two examples is minimized by a model that puts probability 0.5 on each of the two classes.  This seems like the multi-modal behavior the authors argue for.  Perhaps there is something more subtle going on, but the lack of formal analysis makes it difficult to understand how much of an issue it is.  Further, the proposed method is adding a KL Divergence term to the generator loss that is at odds with the GAN loss, which was what the authors argue against.  In short, the argument against the most similar method and the proposed method is weak in my opinion.\n4) I think the name calibration network is a bit misleading. It seems the calibration network is the base segmentation model and the refinement network is calibrating the model.\n5) The paper would benefit from an algorithm sketch to explicitly show the two stages of training.\n6) In the experiments the authors switch to mean squared loss after arguing against cross-entropy.  It's not clear why this is done or why this is a proper baseline.\n7) Looking at figure 3, it would seem the refinement network does not generate outputs that closely match any of the ground truth annotations, but rather some combinations of them.  In practice, I would assume that someone would use the mean and standard deviation to understand the predictions and not samples, so this is less of an issue, but it highlights an issue with considering each pixel independently.\n8) While I think the GED metric make sense here, I think calibration (like expected calibration error) or uncertainty focused (like those proposed in (Mukhoti and Gal; 2019)) would be useful to tell a more complete story of the evaluation. \n9) Figure 5 is unclear.  The text seems to imply that this shows that their technique does not calibrate their model well.  To me this is a strong argument against their approach.  I am not sure the value of a diverse set of segmentations if the model cannot accurately convey uncertainty in predictions.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors propose a two-stage calibration technique for semantic segmentation.  The approach is conceptually simple and can be applied to a number of base segmentation models.  However, there is a lack of strong justification for their approach, seemingly poor pixel-wise calibration results, and the main novel contribution is a simple additive KL Divergence term in the generator's objective. ", "review": "In this work the authors propose a two-stage, adversarial training technique to calibrate a semantic segmentation model when faced with conflicting labels in the train set.  Their approach is to first train a segmentation model. Then, it is used to feed a GAN model which itself is trained against a discriminator to produce diverse segmentations that reflect the diversity in the train set.  The authors compare their approach to similar methods on a synthetic data set and two semantic segmentation data sets. \n\nPros:\n1) The technique is conceptually simple as training a segmentation model with cross entropy loss as well as training a GAN are both well-understood.\n2) The technique can really be used to calibrate any differentiable semantic segmentation model, regardless of who trained it or how.\n3) The experimental details are described in enough detail to likely be able to replicate most of the results\n4) Minus some organizational issues, the writing is good overall.\n\n\nCons:\n1) I found the motivation in the introduction to be slightly difficult to follow.  More specifically, the concepts ambiguity in data labels leading to a multi-modal data distribution and the need to model uncertainty need to be more tightly discussed.  As it is written now, the authors first argue for modeling a noisy empirical distribution that captures the ambiguities, which seems counter intuitive since there assumedly exists a single true segmentation of the image then later discuss uncertainty.  I think discussing uncertainty modeling, and specifically calibration, first would alleviate this issue.\n2) \"Isola et al. (2017) have demonstrated that it introduces only minor stochasticity in the output and returns inconsistent samples.\" (page 2) - In Isola et al. (2017)  dropout is used for image to image translation in a GAN.  I do not think this provides sufficient evidence that using dropout BNNs  such as in Kendal and Gal (2017) has these properties.\n3) The authors argue that combining the GAN loss and the pixel-wise cross entropy is not well-suited for noisy data because they are often at odds.  It's not clear to me that this is a problem in practice, as the cross entropy will spread probability mass across different conflicting labeled instances.  Assume a pixel in the data set appears twice with two different labels.  The sum of the cross entropy over these two examples is minimized by a model that puts probability 0.5 on each of the two classes.  This seems like the multi-modal behavior the authors argue for.  Perhaps there is something more subtle going on, but the lack of formal analysis makes it difficult to understand how much of an issue it is.  Further, the proposed method is adding a KL Divergence term to the generator loss that is at odds with the GAN loss, which was what the authors argue against.  In short, the argument against the most similar method and the proposed method is weak in my opinion.\n4) I think the name calibration network is a bit misleading. It seems the calibration network is the base segmentation model and the refinement network is calibrating the model.\n5) The paper would benefit from an algorithm sketch to explicitly show the two stages of training.\n6) In the experiments the authors switch to mean squared loss after arguing against cross-entropy.  It's not clear why this is done or why this is a proper baseline.\n7) Looking at figure 3, it would seem the refinement network does not generate outputs that closely match any of the ground truth annotations, but rather some combinations of them.  In practice, I would assume that someone would use the mean and standard deviation to understand the predictions and not samples, so this is less of an issue, but it highlights an issue with considering each pixel independently.\n8) While I think the GED metric make sense here, I think calibration (like expected calibration error) or uncertainty focused (like those proposed in (Mukhoti and Gal; 2019)) would be useful to tell a more complete story of the evaluation. \n9) Figure 5 is unclear.  The text seems to imply that this shows that their technique does not calibrate their model well.  To me this is a strong argument against their approach.  I am not sure the value of a diverse set of segmentations if the model cannot accurately convey uncertainty in predictions.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603931134404}, {"id": "RJXYo2XiIME", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3116/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "-----------------------------------------------------------------------------------------------------------------------------------------------------------------\nPOST REBUTTAL\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------\n\nThe rebuttal has addressed most of my concerns and I am happy to increase the score.\n\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------\n\nThe main strengths of the work are -\n* The approach is relatively novel and addresses an important problem.\n* The paper clearly highlights issues with prior work e.g. generator loss is often complemented with pixelwise loss, however, these two objective functions are not well aligned in the presence of noisy data.\n* The paper includes experiments on toy data -- which highlight the contributions of the paper.\n* The proposed approach outperforms prior work -- Hu et al. (2019), Baumgartner et al. (2019) on the LIDC dataset.\n* The proposed approach outperforms Kohl et al. (2018) on CityScapes in case of the GED metric.\n\nThe main weaknesses are,\n* No comparison with closely related approach [1] -- which also proposes a Bayesian approach to capture a calibrated multimodal predictive distributions. In fact, the experiments on 1d bimodal toy data are similar to the experiment in Figure 1 in [1]. A detailed comparison is necessary.\n* Experiments on 1D bimodal data -- the baseline $G_\\phi(F_\\theta(x), \\epsilon)$ is a typical conditional GAN? This seems to be a weak baseline, as recent works [3] address the model collapse issues of conditional GANs. Strong baselines e.g. [1,3], conditional VAEs etc should be considered. \n* Experiments on CityScapes -- the paper does not show results using metrics used by prior work [2] -- in particular Precision-Recall curves and calibration plots which shows the frequency of correctly predicted labels for each bin of predicted probability values. These metrics are also used in [1]. These metrics would better illustrate the calibration of predictions of the proposed approach.\n* Several unclarities -- What is the contribution of the two components - Calibration network and Refinement network on the calibration of the final output? Does the Refinement network aid in improving calibration? \n\n[1] Bayesian Prediction of Future Street Scenes using Synthetic Likelihoods, ICLR 2019.\n[2] What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?, NeurIPS 2017.\n[3] Diversity-Sensitive Conditional Generative Adversarial Networks, ICLR 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method but with missing comparisons and metrics", "review": "-----------------------------------------------------------------------------------------------------------------------------------------------------------------\nPOST REBUTTAL\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------\n\nThe rebuttal has addressed most of my concerns and I am happy to increase the score.\n\n-----------------------------------------------------------------------------------------------------------------------------------------------------------------\n\nThe main strengths of the work are -\n* The approach is relatively novel and addresses an important problem.\n* The paper clearly highlights issues with prior work e.g. generator loss is often complemented with pixelwise loss, however, these two objective functions are not well aligned in the presence of noisy data.\n* The paper includes experiments on toy data -- which highlight the contributions of the paper.\n* The proposed approach outperforms prior work -- Hu et al. (2019), Baumgartner et al. (2019) on the LIDC dataset.\n* The proposed approach outperforms Kohl et al. (2018) on CityScapes in case of the GED metric.\n\nThe main weaknesses are,\n* No comparison with closely related approach [1] -- which also proposes a Bayesian approach to capture a calibrated multimodal predictive distributions. In fact, the experiments on 1d bimodal toy data are similar to the experiment in Figure 1 in [1]. A detailed comparison is necessary.\n* Experiments on 1D bimodal data -- the baseline $G_\\phi(F_\\theta(x), \\epsilon)$ is a typical conditional GAN? This seems to be a weak baseline, as recent works [3] address the model collapse issues of conditional GANs. Strong baselines e.g. [1,3], conditional VAEs etc should be considered. \n* Experiments on CityScapes -- the paper does not show results using metrics used by prior work [2] -- in particular Precision-Recall curves and calibration plots which shows the frequency of correctly predicted labels for each bin of predicted probability values. These metrics are also used in [1]. These metrics would better illustrate the calibration of predictions of the proposed approach.\n* Several unclarities -- What is the contribution of the two components - Calibration network and Refinement network on the calibration of the final output? Does the Refinement network aid in improving calibration? \n\n[1] Bayesian Prediction of Future Street Scenes using Synthetic Likelihoods, ICLR 2019.\n[2] What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?, NeurIPS 2017.\n[3] Diversity-Sensitive Conditional Generative Adversarial Networks, ICLR 2019.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603928456192}, {"id": "7XjARDysAwq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3116/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThis paper presents an approach to stochastic semantic segmentation. The proposed strategy involves a simple extension of neural network architectures for semantic segmentation. In particular, the probabilistic output of a semantic segmentation network is fed into a GAN, which generates a final segmentation. In addition to the classic loss, the GAN is trained such that the average of its prediction matches the input distribution, hence calibrating the distribution. The experiments on a toy dataset and two segmentation datasets demonstrate the superiority of the proposed approach compared to using standard segmentation loss and a few related works.\n\nPros:\n- The writing is good, and the method is explained intuitively\n- The proposed approach is simple yet seems effective \n- The proposed approach is modular and can be applied to pre-trained network architectures \n\nCons:\n- My major concern is that the experimental evaluation does not sufficiently demonstrate that the proposed module is indeed necessary. In the preliminaries section, the authors discuss several simpler alternatives that are not tested in the experiments. E.g., that direct sampling from q_theta yields incoherent segmentation maps or that combining the generator loss with the pixel-wise loss in Eq.2 is not sufficient. I think these settings would serve as good additional baselines in the experimental evaluation.\n- A minor additional criticism is that the authors do not compare to other relevant work such as Kendall and Gal 2017. Moreover, on the Cityscapes dataset, they only compare to one single baseline.\n\n--------------------------------------------------------------------------------------------------------------------------\nPost rebuttal\n\nI thank the authors for providing detailed answers to my concerns. Considering the concerns of the other reviewers and the authors' answers and additional experiments, I think that the paper provides a sufficient contribution to an important research topic. Therefore, I retain my initial rating.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple yet effective strategy for stochastic segmentation with limitations in the experimental evaluation", "review": "Summary:\nThis paper presents an approach to stochastic semantic segmentation. The proposed strategy involves a simple extension of neural network architectures for semantic segmentation. In particular, the probabilistic output of a semantic segmentation network is fed into a GAN, which generates a final segmentation. In addition to the classic loss, the GAN is trained such that the average of its prediction matches the input distribution, hence calibrating the distribution. The experiments on a toy dataset and two segmentation datasets demonstrate the superiority of the proposed approach compared to using standard segmentation loss and a few related works.\n\nPros:\n- The writing is good, and the method is explained intuitively\n- The proposed approach is simple yet seems effective \n- The proposed approach is modular and can be applied to pre-trained network architectures \n\nCons:\n- My major concern is that the experimental evaluation does not sufficiently demonstrate that the proposed module is indeed necessary. In the preliminaries section, the authors discuss several simpler alternatives that are not tested in the experiments. E.g., that direct sampling from q_theta yields incoherent segmentation maps or that combining the generator loss with the pixel-wise loss in Eq.2 is not sufficient. I think these settings would serve as good additional baselines in the experimental evaluation.\n- A minor additional criticism is that the authors do not compare to other relevant work such as Kendall and Gal 2017. Moreover, on the Cityscapes dataset, they only compare to one single baseline.\n\n--------------------------------------------------------------------------------------------------------------------------\nPost rebuttal\n\nI thank the authors for providing detailed answers to my concerns. Considering the concerns of the other reviewers and the authors' answers and additional experiments, I think that the paper provides a sufficient contribution to an important research topic. Therefore, I retain my initial rating.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603824278534}, {"id": "qB6WaDHpv6V", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3116/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "** Summary: \nThis work addresses the context of semantic segmentation where a single input image could be associated with multiple valid labels, as a result of natural ambiguities. Starting from a pretrained deterministic segmentation network F, this work proposes to use an additional conditional generative model G, named as *refinement network*, to generate multiple segmentation predictions; the model G is conditioned on the segmentation probabilistic output of F and the input image. G is trained with adversarial loss and the proposed *calibration loss*, essentially the KL-divergence between the probabilistic output of F and the sample average of G. At runtime, the unified pipeline of F and G can produce multiple segmentation predictions. On one toy example and two real benchmarks, the proposed method show improvements over addressed baselines, in terms of generalized energy distance (GED) and Hungarian-matched IoU (HM-IoU).\n\n** Strengths:\n- The idea of using conditional GANs to produce multiple predictions is interesting.\n- The proposed framework and learning scheme are simple. I think it's easy to reimplement and reproduce results.\n\n** Weakness:\n- Going through the paper I had trouble understanding how the refinement network G can guarantee to produce calibrated probabilities of segmentation modes. The calibration network F, to my understanding, is a deterministic segmentation model trained in a conventional fashion using only the cross-entropy loss. I believe numerous works proved that a model trained that way will end up with over-confident predictions, which are uncalibrated (actually shown in Figure 5). It actually seems misleading to name F with calibration.\n\n- Outputs of pretrained F is then used to regularize the training of the cGAN G via the KL-divergence \"calibration loss\" (which is more like a reconstruction loss to me). Can the authors explain how the refinement network, trained to match sample average with uncalibrated probabilistic targets, can successfully produce calibrated probabilistic outcomes? Also, I would love to see results with calibration metrics like NLL and ECE.\n\n- On the Cityscapes experiments, the segmentation network B is finetuned with or without class-flipping labels? It's quite confusing when sometimes F is a full segmentation network as in Sec 3, Sec 4 and Sec 5.2.1, sometimes F is an ad-hoc network like in 5.2.2. Also F's architecture is detailed in the beginning of 5.2 as SegNet, but only used in 5.2.1. \n\n- Can the authors please provide experimental evidence of how the cross-entropy loss and adversarial loss are not well aligned in the presence of noisy data?\n\n- Minor typos: \n\t+ In Table 2, shouldn't the GED of Kohl et al be 0.206?\n\t+ It may look obvious but should the notations like H,W,C,K be introduced? I thought C is the number of classes at first.\n\n** Preliminary evaluation: this work targets an interesting task of stochastic semantic segmentation. The architecture design and learning scheme seems reasonable to me. The major problem is the lack of evidence to support the claim on output calibration. In terms of writing, I find the paper hard to follow with lots of confusions. Due to those limitations, I give an initial rating of 5.\n\n-- Post-rebuttal -----------------------------------------------------------------------------------------------\n\nGiven the improvement of the last revision, I increase my rating to 6. The revised version has been very much improved, especially in the abstract and introduction Sections. Still I think it's important to additionally have one or two sentences to make very clear on the meaning of calibration, as to not confuse readers.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, but paper and experiments need revision ", "review": "** Summary: \nThis work addresses the context of semantic segmentation where a single input image could be associated with multiple valid labels, as a result of natural ambiguities. Starting from a pretrained deterministic segmentation network F, this work proposes to use an additional conditional generative model G, named as *refinement network*, to generate multiple segmentation predictions; the model G is conditioned on the segmentation probabilistic output of F and the input image. G is trained with adversarial loss and the proposed *calibration loss*, essentially the KL-divergence between the probabilistic output of F and the sample average of G. At runtime, the unified pipeline of F and G can produce multiple segmentation predictions. On one toy example and two real benchmarks, the proposed method show improvements over addressed baselines, in terms of generalized energy distance (GED) and Hungarian-matched IoU (HM-IoU).\n\n** Strengths:\n- The idea of using conditional GANs to produce multiple predictions is interesting.\n- The proposed framework and learning scheme are simple. I think it's easy to reimplement and reproduce results.\n\n** Weakness:\n- Going through the paper I had trouble understanding how the refinement network G can guarantee to produce calibrated probabilities of segmentation modes. The calibration network F, to my understanding, is a deterministic segmentation model trained in a conventional fashion using only the cross-entropy loss. I believe numerous works proved that a model trained that way will end up with over-confident predictions, which are uncalibrated (actually shown in Figure 5). It actually seems misleading to name F with calibration.\n\n- Outputs of pretrained F is then used to regularize the training of the cGAN G via the KL-divergence \"calibration loss\" (which is more like a reconstruction loss to me). Can the authors explain how the refinement network, trained to match sample average with uncalibrated probabilistic targets, can successfully produce calibrated probabilistic outcomes? Also, I would love to see results with calibration metrics like NLL and ECE.\n\n- On the Cityscapes experiments, the segmentation network B is finetuned with or without class-flipping labels? It's quite confusing when sometimes F is a full segmentation network as in Sec 3, Sec 4 and Sec 5.2.1, sometimes F is an ad-hoc network like in 5.2.2. Also F's architecture is detailed in the beginning of 5.2 as SegNet, but only used in 5.2.1. \n\n- Can the authors please provide experimental evidence of how the cross-entropy loss and adversarial loss are not well aligned in the presence of noisy data?\n\n- Minor typos: \n\t+ In Table 2, shouldn't the GED of Kohl et al be 0.206?\n\t+ It may look obvious but should the notations like H,W,C,K be introduced? I thought C is the number of classes at first.\n\n** Preliminary evaluation: this work targets an interesting task of stochastic semantic segmentation. The architecture design and learning scheme seems reasonable to me. The major problem is the lack of evidence to support the claim on output calibration. In terms of writing, I find the paper hard to follow with lots of confusions. Due to those limitations, I give an initial rating of 5.\n\n-- Post-rebuttal -----------------------------------------------------------------------------------------------\n\nGiven the improvement of the last revision, I increase my rating to 6. The revised version has been very much improved, especially in the abstract and introduction Sections. Still I think it's important to additionally have one or two sentences to make very clear on the meaning of calibration, as to not confuse readers.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603822477184}], "openreview_url": "https://openreview.net/forum?id=sAX7Z7uIJ_Y", "arxiv_id": "2006.13144", "paper_pdf": "papers/sAX7Z7uIJ_Y.pdf", "paper_pdf_sha256": "c01566ad73252c3b012ea4082f09a0f1728ce82e0232db33f06fc1728ad75329", "paper_pdf_bytes": 7897455, "paper_pdf_source": "openreview", "code_url": "https://github.com/EliasKassapis/CARSSS", "code_repository": "EliasKassapis/CARSSS", "code_commit": "ff7ec86aab68c4b9ff8aea171244991bd132d487", "code_archive": "repos/sAX7Z7uIJ_Y.zip", "code_archive_sha256": "f9ae38b3ae77d2843bd67d3019f7238f9b4c4e0b5075304e42c35571a09528f1", "code_archive_bytes": 1204432, "code_file_count": 46, "code_extensions": {".py": 46}, "github_disk_usage_kb": 1158, "github_languages": {"Python": 188372}, "github_archived": false, "github_pushed_at": "2022-12-08T10:10:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/calibrated-adversarial-refinement-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1gi0TEFDB", "year": 2020, "status": "rejected", "title": "Understanding Top-k Sparsification in Distributed Deep Learning", "authors": ["Shaohuai Shi", "Xiaowen Chu", "Ka Chun Cheung", "Simon See"], "authorids": ["csshshi@comp.hkbu.edu.hk", "chxw@comp.hkbu.edu.hk", "chcheung@nvidia.com", "ssee@nvidia.com"], "authors_source": "OpenReview API", "abstract": "Distributed stochastic gradient descent (SGD) algorithms are widely deployed in training large-scale deep learning models, while the communication overhead among workers becomes the new system bottleneck. Recently proposed gradient sparsification techniques, especially Top-$k$ sparsification with error compensation (TopK-SGD), can significantly reduce the communication traffic without obvious impact on the model accuracy. Some theoretical studies have been carried out to analyze the convergence property of TopK-SGD. However, existing studies do not dive into the details of Top-$k$ operator in gradient sparsification and use relaxed bounds (e.g., exact bound of Random-$k$) for analysis; hence the derived results cannot well describe the real convergence performance of TopK-SGD. To this end, we first study the gradient distributions of TopK-SGD during training process through extensive experiments. We then theoretically derive a tighter bound for the Top-$k$ operator. Finally, we exploit the property of gradient distribution to propose an approximate top-$k$ selection algorithm, which is computing-efficient for GPUs, to improve the scaling efficiency of TopK-SGD by significantly reducing the computing overhead.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkeyfEe1cH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper869/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper makes two contributions to gradient sparsification to reduce the communication bottleneck in distributed SGD.\n1) Based on an assumption on the distribution of gradient coordinate values that are backed by an empirical study, the paper derives a tighter bound on the approximation quality of top-k gradient sparsification. This result induces better convergence bounds.\n2) The authors note that the top-k cannot benefit from the highly parallel architectures popular in ML, and propose an approximate top-k sparsification operator. This operator tries up to three thresholds and checks how many entries are larger than this value. The initial guess is based on approximating the distribution of gradient coordinates with a normal distribution again.\n\nMy score is weak reject. I believe that both observations are valid, and that their solutions might be practically meaningful. However, I would like to see the comparison to another baselines [1]. I would also urge the authors to make it more clear that their theoretical results are based on a strong assumption on the distribution of gradients. The top-k approximation algorithm is practical but I find 3-step threshold search inelegant.\n\nComments: \n1) [1] is another baseline -- compare your method with it too. \n\n2) 4.3 Convergence performance \"operator can select close elements with Top_k\" --- It seems obvious that it can select similar elements. The question is whether the number of elements chosen is accurate. I would like to see this evaluated. It is unclear if this scheme is biased. As far as I can see, it might be over- or under-sparsifying. \n\n3) 3.1 Gradient Distribution \"One can easily prove\" --- please do so (in the appendix)\n\n4) Theorem 1 - Looks like this can't be true in general. I think it assumes d -> infinity. \n\n5) 3.1 Gradient Distribution \"then pi is a decreasing function\" --- should this be pi^2. Also in figure 3, the result of Eqn 7 is only correct if the curve if pi^2.\n\n6) Figure 2: I am not convinced that these distributions are 'gaussian'. In fact, they seem peakier. It seems to me that this should improve the results (i.e. make the descending pi^2 curve more convex). If this is true, I would encourage the authors to discuss this. BTW, the distribution is in terms of the whole model, or just one randomly picked layer?\n\n7) Conclusion \"theoretically tighter bound\" --- because the assumption on the distribution empirical, I find it slightly misleading to call this a 'theoretical bound'. I would urge the authors to make this very clear. (This does not mean I find the bound meaningless)\n\n8) Introduction: \"O(d), which generally limits the system scalability\" --- The O(d) does not explain scalability in terms of number of workers as is suggested. Note that even though bandwidth scales with O(d) in all reduce, the latency does scale with n. This should not be ignored.\n\n9) Related work/Gradient Sparsification --- Please add a reference for empirical success of top-k. I am not aware of much use outside of academia.\n\n10) Quite a few language errors (some paragraphs/sentences don't make sense at all and there are many cases with missing 'a's etc.)\n\n11) Some of the experimental details are missing\n    - what are the learning rates/batch sizes used in experiments.\n    - How topK is performed? Layer-wise or for the full gradient.\n    - Table 1 \"experimental settings\": how these values were chosen.\n    - Table 2 --- please define how scaling efficiency is computed.\n    - Table 2 --- Do these algorithms achieve the same validation accuracy during the training?\n    - Figure 1: which k was used in these plots? \n\n12) Figure 6: in VGG-16, the gap of 1 percentage point is quite large. This seems expected as the compression ratio is very high (1000x). \n\n13) Figure 6: Imagenet training scheme is not standard. SOTA validation accuracy for the Imagenet benchmark with Resnet50 is around 76%. Would it also have similar quality loss on later stages as in training VGG or ResNet20 on cifar? \n\n14) Eqn. 8 - I couldn't follow the first inequality.\n\n15) Introdcution: The first few times \"distribution of gradients\" is mentioned, it was unclear to me if this was over 'coordinates' (as it seems to be), or over 'data points', or 'training time'. Please clarify.\n\n\n[1] Jiarui Fang, Cho-Jui Hsieh. Accelerating Distributed Deep Learning Training with Gradient Compression.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper makes two contributions to gradient sparsification to reduce the communication bottleneck in distributed SGD.\n1) Based on an assumption on the distribution of gradient coordinate values that are backed by an empirical study, the paper derives a tighter bound on the approximation quality of top-k gradient sparsification. This result induces better convergence bounds.\n2) The authors note that the top-k cannot benefit from the highly parallel architectures popular in ML, and propose an approximate top-k sparsification operator. This operator tries up to three thresholds and checks how many entries are larger than this value. The initial guess is based on approximating the distribution of gradient coordinates with a normal distribution again.\n\nMy score is weak reject. I believe that both observations are valid, and that their solutions might be practically meaningful. However, I would like to see the comparison to another baselines [1]. I would also urge the authors to make it more clear that their theoretical results are based on a strong assumption on the distribution of gradients. The top-k approximation algorithm is practical but I find 3-step threshold search inelegant.\n\nComments: \n1) [1] is another baseline -- compare your method with it too. \n\n2) 4.3 Convergence performance \"operator can select close elements with Top_k\" --- It seems obvious that it can select similar elements. The question is whether the number of elements chosen is accurate. I would like to see this evaluated. It is unclear if this scheme is biased. As far as I can see, it might be over- or under-sparsifying. \n\n3) 3.1 Gradient Distribution \"One can easily prove\" --- please do so (in the appendix)\n\n4) Theorem 1 - Looks like this can't be true in general. I think it assumes d -> infinity. \n\n5) 3.1 Gradient Distribution \"then pi is a decreasing function\" --- should this be pi^2. Also in figure 3, the result of Eqn 7 is only correct if the curve if pi^2.\n\n6) Figure 2: I am not convinced that these distributions are 'gaussian'. In fact, they seem peakier. It seems to me that this should improve the results (i.e. make the descending pi^2 curve more convex). If this is true, I would encourage the authors to discuss this. BTW, the distribution is in terms of the whole model, or just one randomly picked layer?\n\n7) Conclusion \"theoretically tighter bound\" --- because the assumption on the distribution empirical, I find it slightly misleading to call this a 'theoretical bound'. I would urge the authors to make this very clear. (This does not mean I find the bound meaningless)\n\n8) Introduction: \"O(d), which generally limits the system scalability\" --- The O(d) does not explain scalability in terms of number of workers as is suggested. Note that even though bandwidth scales with O(d) in all reduce, the latency does scale with n. This should not be ignored.\n\n9) Related work/Gradient Sparsification --- Please add a reference for empirical success of top-k. I am not aware of much use outside of academia.\n\n10) Quite a few language errors (some paragraphs/sentences don't make sense at all and there are many cases with missing 'a's etc.)\n\n11) Some of the experimental details are missing\n    - what are the learning rates/batch sizes used in experiments.\n    - How topK is performed? Layer-wise or for the full gradient.\n    - Table 1 \"experimental settings\": how these values were chosen.\n    - Table 2 --- please define how scaling efficiency is computed.\n    - Table 2 --- Do these algorithms achieve the same validation accuracy during the training?\n    - Figure 1: which k was used in these plots? \n\n12) Figure 6: in VGG-16, the gap of 1 percentage point is quite large. This seems expected as the compression ratio is very high (1000x). \n\n13) Figure 6: Imagenet training scheme is not standard. SOTA validation accuracy for the Imagenet benchmark with Resnet50 is around 76%. Would it also have similar quality loss on later stages as in training VGG or ResNet20 on cifar? \n\n14) Eqn. 8 - I couldn't follow the first inequality.\n\n15) Introdcution: The first few times \"distribution of gradients\" is mentioned, it was unclear to me if this was over 'coordinates' (as it seems to be), or over 'data points', or 'training time'. Please clarify.\n\n\n[1] Jiarui Fang, Cho-Jui Hsieh. Accelerating Distributed Deep Learning Training with Gradient Compression.\n"}, "tcdate": 1571910663356}, {"id": "rke_w79jFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper869/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Top-k algorithm is a gradient sparsification method which gains its popularity to train deep networks due to its high compression ratio. However due to its computation overhead, it is not efficient on GPUs. This paper performs empirical study on the distribution of the gradients when training various deep networks, and provides a Gaussian approximation analysis to improve the convergence of the top-k algorithm. Further, the paper proposes to use a Gaussian-k algorithm to perform similar gradient sparsification with a much lower computational cost without losing much convergence accuracy compared to Dense-SGD. \n \nHowever, the theoretical result seems to me overstated in the sense that it lacks mathematical rigor in the proof and is not clear how much insights it brings to understand better why top-k sparsification algorithms work well in deep learning (Figure 5 shows that the bound is still too tight). It is written after Equation (6) that “One can easily prove that π is convex and it is always less than the reference line (y = −i/d + 1) if u follows bell shaped distributions as illustrated”, however it is not clear me in what sense this is true.  The u are random variables, therefore π is a curve which depends their realizations, hence it is a random curve. Or maybe it holds when d goes to infinity?\n\nThe numerical results are specific with k = 0.001d, which makes it hard to see if the Gaussian-k algorithm would still work using different k/d ratio. As shown in Figure 2,  some of the histograms of u^1_t are quite sparse (these plots are hard to read for different iterations, maybe use cdf instead and perform statistical test to check how close to Gaussian distributions), therefore in some of these cases the Gaussian approximation may be poor. It is worth further investigation of the robustness of this algorithm as a replacement of the top-k algorithm. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "Top-k algorithm is a gradient sparsification method which gains its popularity to train deep networks due to its high compression ratio. However due to its computation overhead, it is not efficient on GPUs. This paper performs empirical study on the distribution of the gradients when training various deep networks, and provides a Gaussian approximation analysis to improve the convergence of the top-k algorithm. Further, the paper proposes to use a Gaussian-k algorithm to perform similar gradient sparsification with a much lower computational cost without losing much convergence accuracy compared to Dense-SGD. \n \nHowever, the theoretical result seems to me overstated in the sense that it lacks mathematical rigor in the proof and is not clear how much insights it brings to understand better why top-k sparsification algorithms work well in deep learning (Figure 5 shows that the bound is still too tight). It is written after Equation (6) that “One can easily prove that π is convex and it is always less than the reference line (y = −i/d + 1) if u follows bell shaped distributions as illustrated”, however it is not clear me in what sense this is true.  The u are random variables, therefore π is a curve which depends their realizations, hence it is a random curve. Or maybe it holds when d goes to infinity?\n\nThe numerical results are specific with k = 0.001d, which makes it hard to see if the Gaussian-k algorithm would still work using different k/d ratio. As shown in Figure 2,  some of the histograms of u^1_t are quite sparse (these plots are hard to read for different iterations, maybe use cdf instead and perform statistical test to check how close to Gaussian distributions), therefore in some of these cases the Gaussian approximation may be poor. It is worth further investigation of the robustness of this algorithm as a replacement of the top-k algorithm. "}, "tcdate": 1571689312023}, {"id": "HygKNsO9Kr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper869/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper  empirically investigates the distribution of the gradient magnitude in the training of DNNs, based on which a tighter bound is derived over the conventional bound on top-K gradient sparsification.  The authors also propose a so-call GaussianK-SGD to approximate the top-K selection which is shown to be more efficient on GPUs.  Experiments are carried out on various datasets with various network architectures and the results seem to be supportive.  Overall, I find the work interesting and may have real value for communication-efficient distributed training.  On the other hand, I also have following concerns. \n\n1.  The theory part of the work is basically based on empirical observations.  My sense is that it can be more rigorous than its current form.  For instance, the authors may give a concrete form (e.g. Gaussian) of the $\\pi$ and derive the area under the curve of  $\\pi$ and so on.  Right now, a major part of the derivation is given in a rough sketch. \n\n2. $||\\mu||_{\\infty}$ should be $||\\mu||^{2}_{\\infty}$ in Eq. 5. \n\n3. It is not clear to me why the inequality holds in Eq.8.   Could you elaborate on it a bit?  Again, it would be helpful if a concrete function of $\\pi$ can be given for derivation.\n\n4. I notice that different distributed training packages are used in experiments, e.g. NCCL and OpenMPI.  It is not clear to me why the experiments are not conducted using a consistent distributed setting.  I don't see authors compare the performance of them.  Also, when comparing the wall-clock time in end-to-end training, are the models trained using the same all-reduce setting?  The authors need to explain. Otherwise, the numbers in Table 2 are not directly comparable.  \n\nP.S. rebuttal read.  I will stay with my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "This paper  empirically investigates the distribution of the gradient magnitude in the training of DNNs, based on which a tighter bound is derived over the conventional bound on top-K gradient sparsification.  The authors also propose a so-call GaussianK-SGD to approximate the top-K selection which is shown to be more efficient on GPUs.  Experiments are carried out on various datasets with various network architectures and the results seem to be supportive.  Overall, I find the work interesting and may have real value for communication-efficient distributed training.  On the other hand, I also have following concerns. \n\n1.  The theory part of the work is basically based on empirical observations.  My sense is that it can be more rigorous than its current form.  For instance, the authors may give a concrete form (e.g. Gaussian) of the $\\pi$ and derive the area under the curve of  $\\pi$ and so on.  Right now, a major part of the derivation is given in a rough sketch. \n\n2. $||\\mu||_{\\infty}$ should be $||\\mu||^{2}_{\\infty}$ in Eq. 5. \n\n3. It is not clear to me why the inequality holds in Eq.8.   Could you elaborate on it a bit?  Again, it would be helpful if a concrete function of $\\pi$ can be given for derivation.\n\n4. I notice that different distributed training packages are used in experiments, e.g. NCCL and OpenMPI.  It is not clear to me why the experiments are not conducted using a consistent distributed setting.  I don't see authors compare the performance of them.  Also, when comparing the wall-clock time in end-to-end training, are the models trained using the same all-reduce setting?  The authors need to explain. Otherwise, the numbers in Table 2 are not directly comparable.  \n\nP.S. rebuttal read.  I will stay with my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571617585357}], "openreview_url": "https://openreview.net/forum?id=B1gi0TEFDB", "arxiv_id": "1911.08772", "paper_pdf": "papers/B1gi0TEFDB.pdf", "paper_pdf_sha256": "5238e3088a30bc860167a8a7cf2f13d99905c30d37e5e764f574a4d15363319f", "paper_pdf_bytes": 3598631, "paper_pdf_source": "openreview", "code_url": "https://github.com/hclhkbu/GaussianK-SGD", "code_repository": "hclhkbu/GaussianK-SGD", "code_commit": "3da1fb3b62e4067828a965687d5cc7ae7892e742", "code_archive": "repos/B1gi0TEFDB.zip", "code_archive_sha256": "ea37273ea50b3f82c205a0ccc15ae66d0afd9a66cff5dd14eb0545e779c676ae", "code_archive_bytes": 85610, "code_file_count": 36, "code_extensions": {".py": 35, ".sh": 1}, "github_disk_usage_kb": 3729, "github_languages": {"Python": 205644, "Shell": 746}, "github_archived": false, "github_pushed_at": "2019-11-15T18:42:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-top-k-sparsification-in-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RBktryANQ9", "year": 2026, "status": "rejected", "title": "Dataset Distillers Are Good Label Denoisers In the Wild", "authors": ["Lechao Cheng", "KaifengChen", "Jiyang Li", "Zerun Liu", "Shengeng Tang", "Shufei Zhang", "Zhihui Yang"], "authorids": ["~Lechao_Cheng2", "~KaifengChen2", "~Jiyang_Li1", "~Zerun_Liu1", "~Shengeng_Tang1", "~Shufei_Zhang1", "~Zhihui_Yang3"], "authors_source": "OpenReview API", "abstract": "Dataset distillation aims to synthesize a small set of informative samples that preserve the generalization ability of large datasets. However, its behavior under noisy conditions remains underexplored. In this paper, we systematically study dataset distillation under three representative noise types: symmetric, asymmetric, and natural noise. We first discover that, when the noise ratio exceeds a critical threshold, mainstream distillation methods consistently outperform training on the full noisy dataset using significantly fewer distilled samples. In contrast, under asymmetric noise, the structured label corruption often entangles with semantic features, making it difficult for distilled samples to recover the clean data distribution. We further validate the effectiveness of dataset distillation on real-world noisy datasets, highlighting its robustness under high noise but degraded performance in low-noise settings due to over-compression. To provide theoretical insights, we derive upper and lower bounds on the required images per class (IPC) under each noise type, grounded in information theory and PAC-Bayes analysis. Our findings offer both empirical and theoretical guidelines for effective distillation in noisy learning scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "ENwc3xRxXL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6480/Reviewer_9Gpj"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper studied dataset distillation in the noise label setting. Empirical experiments show that data distillation shows improvement on results under symmetric noise and natural noise but decreased performance on asymmetric noise. Also the paper studied relationship between accuracy and IPC under symmetric and asymmetric conditions both empirically and theoretically.", "review_text": "This paper studied dataset distillation in the noise label setting. Empirical experiments show that data distillation shows improvement on results under symmetric noise and natural noise but decreased performance on asymmetric noise. Also the paper studied relationship between accuracy and IPC under symmetric and asymmetric conditions both empirically and theoretically.", "strengths": "The paper is well organized and presented. \nThe experiments is thorough under symmetric, asymmetric  and natural noise, with different noise level and IPC, providing promising results for analysis", "weaknesses": "See questions below", "questions": "The theoretical analysis shows the bound of IPC under different level of noise, but that doesn't explain the trend in Figure 6,7, i.e., how the accuracy change with noise rate with different IPC. Is there any insight or try on predicting this trend, like a scaling law for data distillation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studied dataset distillation in the noise label setting. Empirical experiments show that data distillation shows improvement on results under symmetric noise and natural noise but decreased performance on asymmetric noise. Also the paper studied relationship between accuracy and IPC under symmetric and asymmetric conditions both empirically and theoretically.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper is well organized and presented. \nThe experiments is thorough under symmetric, asymmetric  and natural noise, with different noise level and IPC, providing promising results for analysis", "weaknesses": "See questions below", "questions": "The theoretical analysis shows the bound of IPC under different level of noise, but that doesn't explain the trend in Figure 6,7, i.e., how the accuracy change with noise rate with different IPC. Is there any insight or try on predicting this trend, like a scaling law for data distillation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762730979126}, {"id": "PpcPf83sLV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6480/Reviewer_cF8n"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper study the performance of existing dataset distillation methods (such as DM, DANCE, and RCIG) in noisy label scenarios. The authors claim that dataset distillation itself can serve as a potential denoising mechanism. They find that dataset distillation methods excel at filtering out symmetric noise and is robust to high natural noise, but struggle with asymmetric noise. They also use information theory and PAC-Bayes theory to propose a relationship (upper or lower bound) between the number of samples required for data distillation and the noise rate.", "review_text": "The paper study the performance of existing dataset distillation methods (such as DM, DANCE, and RCIG) in noisy label scenarios. The authors claim that dataset distillation itself can serve as a potential denoising mechanism. They find that dataset distillation methods excel at filtering out symmetric noise and is robust to high natural noise, but struggle with asymmetric noise. They also use information theory and PAC-Bayes theory to propose a relationship (upper or lower bound) between the number of samples required for data distillation and the noise rate.", "strengths": "1. This paper reinterprets dataset distillation as an implicit denoising process through semantic compression.\n2. They use information-theoretic and PAC-Bayes to give bounds for dataset distillation under different noise regimes\n3. Well-designed small-scale experiments are provided to show the denoising effectiveness of distillation under high symmetric noise and also reveal its failure under asymmetric noise.", "weaknesses": "1. The theoretical bounds derived in Corollaries I-III need stronger empirical validation. Currently, the experimental results only show qualitative consistency with the theoretical trends. However, since this paper focuses on the analysis of new insight and findings from observational results, rather than proposing new solutions, qualitative analysis is insufficient. Experiments need to be designed to quantitatively test these IPC boundaries, and to discuss the gap between the theoretical boundaries and the practically achievable limits, as well as the possible causes of this gap. Is the reason why performance shown in Figure 4 and 5 cannot be improved because this theoretical limit has been reached?\n2. In Corollary-II，authors try to characterize the IPC bounds under asymmetric noise by considering the rank of the noise transition matrix and its impact on semantic diversity. How is semantic diversity manifested? Could you explain in detail the relationship between semantic diversity and boundary?\n3. In the formulas of conditional noise distribution shows in Discovery II, how is the flip probability specifically implemented based on semantic similarity? Please give a detailed explanation.\n4. In Figure 5, authors gives that CIFAR-100N follows a similar setup, but only “Noise” setting is shown in the figure. Please add any other settings that are the same as those for the CIFAR-10N, or explain why you don’t use those settings.\n5. In the section of “From Theory to Practice: How Noise Shapes Dataset Distillation”, only general and macro-level analysis is given, such as IPC scales inversely with the clean label proportion, which has already been claimed in the \nprevious section. Furthermore, there is a significant amount of repetition between Figures 5 and 7. \n6. Further validation on more datasets and more distillation models is needed to demonstrate that the findings proposed in the paper are universally applicable to distillation methods and image label noise problems, rather than being limited to the 3 distillation models mentioned in the paper and the CIFAR dataset.", "questions": "Please refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper study the performance of existing dataset distillation methods (such as DM, DANCE, and RCIG) in noisy label scenarios. The authors claim that dataset distillation itself can serve as a potential denoising mechanism. They find that dataset distillation methods excel at filtering out symmetric noise and is robust to high natural noise, but struggle with asymmetric noise. They also use information theory and PAC-Bayes theory to propose a relationship (upper or lower bound) between the number of samples required for data distillation and the noise rate.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This paper reinterprets dataset distillation as an implicit denoising process through semantic compression.\n2. They use information-theoretic and PAC-Bayes to give bounds for dataset distillation under different noise regimes\n3. Well-designed small-scale experiments are provided to show the denoising effectiveness of distillation under high symmetric noise and also reveal its failure under asymmetric noise.", "weaknesses": "1. The theoretical bounds derived in Corollaries I-III need stronger empirical validation. Currently, the experimental results only show qualitative consistency with the theoretical trends. However, since this paper focuses on the analysis of new insight and findings from observational results, rather than proposing new solutions, qualitative analysis is insufficient. Experiments need to be designed to quantitatively test these IPC boundaries, and to discuss the gap between the theoretical boundaries and the practically achievable limits, as well as the possible causes of this gap. Is the reason why performance shown in Figure 4 and 5 cannot be improved because this theoretical limit has been reached?\n2. In Corollary-II，authors try to characterize the IPC bounds under asymmetric noise by considering the rank of the noise transition matrix and its impact on semantic diversity. How is semantic diversity manifested? Could you explain in detail the relationship between semantic diversity and boundary?\n3. In the formulas of conditional noise distribution shows in Discovery II, how is the flip probability specifically implemented based on semantic similarity? Please give a detailed explanation.\n4. In Figure 5, authors gives that CIFAR-100N follows a similar setup, but only “Noise” setting is shown in the figure. Please add any other settings that are the same as those for the CIFAR-10N, or explain why you don’t use those settings.\n5. In the section of “From Theory to Practice: How Noise Shapes Dataset Distillation”, only general and macro-level analysis is given, such as IPC scales inversely with the clean label proportion, which has already been claimed in the \nprevious section. Furthermore, there is a significant amount of repetition between Figures 5 and 7. \n6. Further validation on more datasets and more distillation models is needed to demonstrate that the findings proposed in the paper are universally applicable to distillation methods and image label noise problems, rather than being limited to the 3 distillation models mentioned in the paper and the CIFAR dataset.", "questions": "Please refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761976389018}, {"id": "sA5EPaq0DB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6480/Reviewer_2b5n"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper systematically studies dataset distillation under various labeled noise conditions, e.g., symmetric noise, asymmetric noise, and natural noise. Further, this paper reveals a counterintuitive yet insightful and interesting phenomenon: when exceeding the critical noise ratio, the distilled dataset can outperform the full-data training. The paper is also provided  empirical evidence on CIFAR-10, CIFAR-100, Tiny-ImageNet, etc, and covers the mainstream dataset distillation methods, e.g., DATM, DANCE, RCIG. Furthermore, the paper offers a theoretical explanation based on information theory and PAC-Bayes analysis, deriving upper and lower bounds on the required IPC for effective learning under different noise regimes.", "review_text": "This paper systematically studies dataset distillation under various labeled noise conditions, e.g., symmetric noise, asymmetric noise, and natural noise. Further, this paper reveals a counterintuitive yet insightful and interesting phenomenon: when exceeding the critical noise ratio, the distilled dataset can outperform the full-data training. The paper is also provided  empirical evidence on CIFAR-10, CIFAR-100, Tiny-ImageNet, etc, and covers the mainstream dataset distillation methods, e.g., DATM, DANCE, RCIG. Furthermore, the paper offers a theoretical explanation based on information theory and PAC-Bayes analysis, deriving upper and lower bounds on the required IPC for effective learning under different noise regimes.", "strengths": "1. The paper reveals a counterintuitive but highly interesting and insightful phenomenon, which broaden the potential application of dataset distillation as a denoising or robustness-enhancing technique for noisy dataset.\n2. The paper offers comprehensive theoretical analysis based on information theory and PAC-Bayes, providing the upper and lower bounds on the required IPC, to support and explain the observed phenomena.\n3. The presentation is clear with figures and equations.", "weaknesses": "1. [major] Although the observed phenomenon is interesting, its applicability appears to be rather limited and dependent on the dataset and adopted dataset distillation method. Specifically, under symmetric noise condition, the performance of the distilled dataset begins to outperform training on the full dataset when the noise ratio reaches around 20%. However, for more realistic asymmetric noise, the threshold rises above 40% on the CIFAR-10 dataset (which is quite extreme in real-world scenarios). This trend becomes inconsistent when using more complex datasets (like CIFAR-100), the advantage only appears in the 20% to 40% noise range and disappears at higher noise ratios. Furthermore, this phenomenon is not observed at all for RCIG. These results suggest that while this finding is conceptually interesting, it may be difficult to generalize to different datasets or noise settings, making it challenging to draw unified conclusions.\n2. [minor] This experimental verification focuses primarily on small-scale image classification benchmarks, e.g. CIFAR-10, CIFAR-100, and Tiny-ImageNet, and lacks analysis of larger datasets.\n3. [minor] This paper demonstrates that dataset distillation has an implicit denoising effect in high-noise condition, but it does not directly compare it with existing noise-resistant training or label correction methods, e.g., Co-teaching [1], DivideMix[2], GCELoss[3].\n\n[1] Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels\n\n[2] DivideMix: Learning with Noisy Labels as Semi-supervised Learning\n\n[3] Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels", "questions": "1. The theoretical analysis in the paper is based on idealized assumptions such as independent noise, but these assumptions may not apply to real-world scenarios. For example, in Figure 4(b) with asymmetric noise, under the setting of CIFAR-100 and RCIG method, this phenomenon does not appear. Could the authors explain whether these theoretical results still apply to more complex and realistic situations? How do we select the IPC and determine if this phenomenon exists?\n2. Could the authors clarify whether the observed phenomenon is expected to generalize to larger or more complex datasets?\n3. The paper demonstrates that dataset distillation has an implicit denoising effect in high-noise condition, but it lacks comparison with existing noise-robust training such as Co-teaching, DivideMix, and GCELoss.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper systematically studies dataset distillation under various labeled noise conditions, e.g., symmetric noise, asymmetric noise, and natural noise. Further, this paper reveals a counterintuitive yet insightful and interesting phenomenon: when exceeding the critical noise ratio, the distilled dataset can outperform the full-data training. The paper is also provided  empirical evidence on CIFAR-10, CIFAR-100, Tiny-ImageNet, etc, and covers the mainstream dataset distillation methods, e.g., DATM, DANCE, RCIG. Furthermore, the paper offers a theoretical explanation based on information theory and PAC-Bayes analysis, deriving upper and lower bounds on the required IPC for effective learning under different noise regimes.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper reveals a counterintuitive but highly interesting and insightful phenomenon, which broaden the potential application of dataset distillation as a denoising or robustness-enhancing technique for noisy dataset.\n2. The paper offers comprehensive theoretical analysis based on information theory and PAC-Bayes, providing the upper and lower bounds on the required IPC, to support and explain the observed phenomena.\n3. The presentation is clear with figures and equations.", "weaknesses": "1. [major] Although the observed phenomenon is interesting, its applicability appears to be rather limited and dependent on the dataset and adopted dataset distillation method. Specifically, under symmetric noise condition, the performance of the distilled dataset begins to outperform training on the full dataset when the noise ratio reaches around 20%. However, for more realistic asymmetric noise, the threshold rises above 40% on the CIFAR-10 dataset (which is quite extreme in real-world scenarios). This trend becomes inconsistent when using more complex datasets (like CIFAR-100), the advantage only appears in the 20% to 40% noise range and disappears at higher noise ratios. Furthermore, this phenomenon is not observed at all for RCIG. These results suggest that while this finding is conceptually interesting, it may be difficult to generalize to different datasets or noise settings, making it challenging to draw unified conclusions.\n2. [minor] This experimental verification focuses primarily on small-scale image classification benchmarks, e.g. CIFAR-10, CIFAR-100, and Tiny-ImageNet, and lacks analysis of larger datasets.\n3. [minor] This paper demonstrates that dataset distillation has an implicit denoising effect in high-noise condition, but it does not directly compare it with existing noise-resistant training or label correction methods, e.g., Co-teaching [1], DivideMix[2], GCELoss[3].\n\n[1] Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels\n\n[2] DivideMix: Learning with Noisy Labels as Semi-supervised Learning\n\n[3] Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels", "questions": "1. The theoretical analysis in the paper is based on idealized assumptions such as independent noise, but these assumptions may not apply to real-world scenarios. For example, in Figure 4(b) with asymmetric noise, under the setting of CIFAR-100 and RCIG method, this phenomenon does not appear. Could the authors explain whether these theoretical results still apply to more complex and realistic situations? How do we select the IPC and determine if this phenomenon exists?\n2. Could the authors clarify whether the observed phenomenon is expected to generalize to larger or more complex datasets?\n3. The paper demonstrates that dataset distillation has an implicit denoising effect in high-noise condition, but it lacks comparison with existing noise-robust training such as Co-teaching, DivideMix, and GCELoss.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761901694389}, {"id": "7o2VTEOgzF", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6480/Reviewer_upv1"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper investigates dataset distillation methods as implicit label denoisers, demonstrating that distillation can outperform full-data training under certain noise conditions. The authors provide both empirical evidence across CIFAR-10/100 and Tiny ImageNet, plus theoretical bounds for required images per class (IPC) under different noise types.", "review_text": "This paper investigates dataset distillation methods as implicit label denoisers, demonstrating that distillation can outperform full-data training under certain noise conditions. The authors provide both empirical evidence across CIFAR-10/100 and Tiny ImageNet, plus theoretical bounds for required images per class (IPC) under different noise types.", "strengths": "- The paper perceives dataset distillation as a denoising mechanism in order to further understand as to why distillation works.\n- Tests three distinct noise types (symmetric, asymmetric, natural) provides a thorough analysis.\n- PAC-Bayes bounds add rigor beyond purely empirical observations and probably are useful in real-world training cases.", "weaknesses": "This analysis in this paper offers a great insightful perspectives and lays a strong foundation that helps to further probe along the following lines and even add more value to the broader community. On that note, the paper does not investigate any of the following useful perspectives.\n\n- The experiments use basic 3-4 layer ConvNets. How do these findings translate to modern architectures like ResNets, Vision Transformers, Diffusion models or other state-of-the-art methods?\n\n- What about real-world dataset scales? The analysis is  carried out only on CIFAR and Tiny ImageNet raises serious scalability questions. Would these patterns hold on ImageNet, or larger datasets with hundreds of classes, for that matter datasets with high class imbalance?\n\n- On that note, why assume class balance? Real-world datasets are often imbalanced, how does class imbalance affect the denoising properties? Are the proposed PAC-Bayes bounds generalize to imbalanced cases as well? If so, a demonstration through an example benchmark like ImageNet would greatly improve the quality of the paper.\n\n- There is a clear finding that distillation fails with asymmetric noise but that finding is not further explored. This actually can be considered as a separate problem and a dedicated solution. This seems like a critical limitation that undermines the practical usability.\n\n- Where are the dedicated denoising baselines? You compare against full-data training but not against established label denoising methods like DivideMix, Co-teaching, or meta-learning approaches. How does distillation compare to methods explicitly designed for noisy labels?\n\n- What about data augmentation? Strong augmentation strategies can also act as implicit denoisers. How does distillation compare to simply using better augmentation on the full noisy dataset?\n\n- There is no analysis on how the distilled samples would appear. Note that Deep nets are notoriously good at producing noisy images that can be deceived as real in-distribution samples. Therefore, it is highly recommended to check whether the distilled images semantically meaningful or not? Are they similar to the original clean samples or not? Do they preserve important visual features or just exploit model biases?\n\n- Will the distilled samples from one architecture generalize to train different architectures? How long the change in architectures take to learn or produce decent results?\n\n- What are the cost implications of doing distillation followed by training? Meaning, Is it computationally expensive to train data distillers followed by training on the end task or training on full noisy dataset?", "questions": "Please refer to the above weaknesses section for questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates dataset distillation methods as implicit label denoisers, demonstrating that distillation can outperform full-data training under certain noise conditions. The authors provide both empirical evidence across CIFAR-10/100 and Tiny ImageNet, plus theoretical bounds for required images per class (IPC) under different noise types.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper perceives dataset distillation as a denoising mechanism in order to further understand as to why distillation works.\n- Tests three distinct noise types (symmetric, asymmetric, natural) provides a thorough analysis.\n- PAC-Bayes bounds add rigor beyond purely empirical observations and probably are useful in real-world training cases.", "weaknesses": "This analysis in this paper offers a great insightful perspectives and lays a strong foundation that helps to further probe along the following lines and even add more value to the broader community. On that note, the paper does not investigate any of the following useful perspectives.\n\n- The experiments use basic 3-4 layer ConvNets. How do these findings translate to modern architectures like ResNets, Vision Transformers, Diffusion models or other state-of-the-art methods?\n\n- What about real-world dataset scales? The analysis is  carried out only on CIFAR and Tiny ImageNet raises serious scalability questions. Would these patterns hold on ImageNet, or larger datasets with hundreds of classes, for that matter datasets with high class imbalance?\n\n- On that note, why assume class balance? Real-world datasets are often imbalanced, how does class imbalance affect the denoising properties? Are the proposed PAC-Bayes bounds generalize to imbalanced cases as well? If so, a demonstration through an example benchmark like ImageNet would greatly improve the quality of the paper.\n\n- There is a clear finding that distillation fails with asymmetric noise but that finding is not further explored. This actually can be considered as a separate problem and a dedicated solution. This seems like a critical limitation that undermines the practical usability.\n\n- Where are the dedicated denoising baselines? You compare against full-data training but not against established label denoising methods like DivideMix, Co-teaching, or meta-learning approaches. How does distillation compare to methods explicitly designed for noisy labels?\n\n- What about data augmentation? Strong augmentation strategies can also act as implicit denoisers. How does distillation compare to simply using better augmentation on the full noisy dataset?\n\n- There is no analysis on how the distilled samples would appear. Note that Deep nets are notoriously good at producing noisy images that can be deceived as real in-distribution samples. Therefore, it is highly recommended to check whether the distilled images semantically meaningful or not? Are they similar to the original clean samples or not? Do they preserve important visual features or just exploit model biases?\n\n- Will the distilled samples from one architecture generalize to train different architectures? How long the change in architectures take to learn or produce decent results?\n\n- What are the cost implications of doing distillation followed by training? Meaning, Is it computationally expensive to train data distillers followed by training on the end task or training on full noisy dataset?", "questions": "Please refer to the above weaknesses section for questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761346355938}], "openreview_url": "https://openreview.net/forum?id=RBktryANQ9", "arxiv_id": "2411.11924", "paper_pdf": "papers/RBktryANQ9.pdf", "paper_pdf_sha256": "8aaa694f3d4ec23318fb3542734657d5c28741454b6718999006a563b2385e82", "paper_pdf_bytes": 1163235, "paper_pdf_source": "openreview", "code_url": "https://github.com/Kciiiman/DD_LNL", "code_repository": "Kciiiman/DD_LNL", "code_commit": "d59a5a59637188c7a16121cc9a9be39544556cd1", "code_archive": "repos/RBktryANQ9.zip", "code_archive_sha256": "383d95f05b1c79f879b1a581cbb6f59225ce0386aa4ac515aac8ab4e199f5043", "code_archive_bytes": 4374038, "code_file_count": 70, "code_extensions": {".py": 70}, "github_disk_usage_kb": 3418, "github_languages": {"Python": 794394}, "github_archived": false, "github_pushed_at": "2025-01-12T05:45:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dataset-distillers-are-good-label-denoisers"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "M8xtZuxqC5", "year": 2025, "status": "rejected", "title": "It Takes Two: On the Seamlessness between Reward and Policy Model in RLHF", "authors": ["TaiMing Lu", "Lingfeng Shen", "Xinyu Yang", "Weiting Tan", "Ryumei Nakada", "Linjun Zhang", "Beidi Chen", "Huaxiu Yao"], "authorids": ["~TaiMing_Lu1", "~Lingfeng_Shen1", "~Xinyu_Yang4", "~Weiting_Tan1", "~Ryumei_Nakada1", "~Linjun_Zhang1", "~Beidi_Chen1", "~Huaxiu_Yao1"], "authors_source": "OpenReview API", "abstract": "Reinforcement Learning from Human Feedback (RLHF) involves training policy models (PMs) and reward models (RMs) to align language models with human preferences. Instead of focusing solely on PMs and RMs independently, we propose to examine their interactions during fine-tuning, introducing the concept of \\textbf{seamlessness}. Our study starts with observing the saturation phenomenon, where continual improvements in RM and PM do not translate into RLHF progress. Our analysis shows that RMs fail to assign proper scores to PM responses, resulting in a 35% mismatch rate with human preferences, highlighting a significant discrepancy between PM and RM. To measure seamlessness between PM and RM without human effort, we propose an automatic metric, SEAM. SEAM quantifies the discrepancies between PM and RM judgments induced by data samples. We validate the effectiveness of SEAM in data selection and model augmentation. Our experiments demonstrate that (1) using SEAM-filtered data for RL training improves RLHF performance by 4.5%, and (2) SEAM-guided model augmentation results in a 4% performance improvement over standard augmentation methods.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "JWesg7vEPZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12132/Reviewer_sNAF"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces the concept of seamlessness to explain the saturation performance phenomenon of the policy model during RLHF. It proposes the SEAM metric to measure the performance difference between the policy model and the reward model. The motivation behind this work is clear and meaningful. However, the paper lacks deeper analysis of the experimental phenomena and the experiments are not sufficiently comprehensive.", "review_text": "This paper introduces the concept of seamlessness to explain the saturation performance phenomenon of the policy model during RLHF. It proposes the SEAM metric to measure the performance difference between the policy model and the reward model. The motivation behind this work is clear and meaningful. However, the paper lacks deeper analysis of the experimental phenomena and the experiments are not sufficiently comprehensive.", "strengths": "1. The concept of seamlessness is a novel approach to understanding the saturation performance in policy models during RLHF.\n\n2. The introduction of the SEAM metric provides a new way to measure the performance gap between the policy model and the reward model.\n\n3. The motivation for this work is clear and addresses an important issue in the field.", "weaknesses": "1. How does the SEAM metric change during the alignment training process? Does the difference between the policy model and the reward model increase gradually with training?\n\n2. Could filtering out data using the SEAM metric potentially reduce the diversity of the training data? For instance, could this lead to the model ignoring difficult data points?\n\n3. The experimental section should include more models and datasets, such as Qwen and Mistral, and evaluate performance in helpful and harmless scenarios to enhance the credibility of the results.\n\n4. Can this method guide the optimization of the reward model by helping to filter training data and optimize the annotation process?", "questions": "please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the concept of seamlessness to explain the saturation performance phenomenon of the policy model during RLHF. It proposes the SEAM metric to measure the performance difference between the policy model and the reward model. The motivation behind this work is clear and meaningful. However, the paper lacks deeper analysis of the experimental phenomena and the experiments are not sufficiently comprehensive.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The concept of seamlessness is a novel approach to understanding the saturation performance in policy models during RLHF.\n\n2. The introduction of the SEAM metric provides a new way to measure the performance gap between the policy model and the reward model.\n\n3. The motivation for this work is clear and addresses an important issue in the field.", "weaknesses": "1. How does the SEAM metric change during the alignment training process? Does the difference between the policy model and the reward model increase gradually with training?\n\n2. Could filtering out data using the SEAM metric potentially reduce the diversity of the training data? For instance, could this lead to the model ignoring difficult data points?\n\n3. The experimental section should include more models and datasets, such as Qwen and Mistral, and evaluate performance in helpful and harmless scenarios to enhance the credibility of the results.\n\n4. Can this method guide the optimization of the reward model by helping to filter training data and optimize the annotation process?", "questions": "please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730719799340}, {"id": "SLXPVDPjQK", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12132/Reviewer_mfDi"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This study introduces the concept of \"seamlessness\" to improve Reinforcement Learning from Human Feedback (RLHF), focusing on the alignment between policy models (PMs) and reward models (RMs). The authors identify a \"saturation phenomenon,\" where advancements in PMs and RMs don’t lead to better RLHF outcomes, partly due to a 35% mismatch rate between RM scores and human preferences. To address this, they introduce SEAM, an automatic metric to quantify PM-RM discrepancies. Experiments show that SEAM-filtered data selection and SEAM-guided model augmentation improve RLHF performance by 4.5% and 4%, respectively. This study highlights the critical alignment between PMs and RMs for effective RLHF.", "review_text": "This study introduces the concept of \"seamlessness\" to improve Reinforcement Learning from Human Feedback (RLHF), focusing on the alignment between policy models (PMs) and reward models (RMs). The authors identify a \"saturation phenomenon,\" where advancements in PMs and RMs don’t lead to better RLHF outcomes, partly due to a 35% mismatch rate between RM scores and human preferences. To address this, they introduce SEAM, an automatic metric to quantify PM-RM discrepancies. Experiments show that SEAM-filtered data selection and SEAM-guided model augmentation improve RLHF performance by 4.5% and 4%, respectively. This study highlights the critical alignment between PMs and RMs for effective RLHF.", "strengths": "- This paper tackles an important and current topic in RLHF with a method that’s both intuitive and straightforward.\n- It lays out a clear research question and problem statement, supporting its claims with well-designed and diverse experiments.\n- Core concepts, like the \"saturation phenomenon,\" are explained in an accessible way that makes it easy for even new readers to follow. The use of visual aids is particularly effective, helping to deepen understanding and clarify key ideas.\n- The paper offers both theoretical insights and empirical evidence for the effectiveness of its approach. Especially interesting is its effort to explain the filtering effect by analyzing the \"effect of bad instruction\" from a post-filtering perspective, which adds an engaging layer to the discussion.", "weaknesses": "- Certain parts of the writing lack clarity:\n  - In Section 4, the explanation regarding Figure 3 under the sanity check section is not entirely clear. According to the Appendix, PM model quality is rated with a 1-10 score via the LLM-as-a-Judge method, but Figure 3 displays it simply as a percentage. A more detailed explanation of how the 1-10 score is converted to a percentage is needed. The same applies to RM quality.\n  - Definition (1) aims to define \"seamlessness.\" However, $S(I, R\\_{\\theta}, \\pi^{\\text{SFT}})$ is a metric that describes \"seam-ness\" by being proportional to RM misjudgment $\\epsilon(r, R\\_{\\theta})$.\n  - In Section 5.2, the authors describe three ways to construct samples ($\\text{SEAM}\\_{\\text{Contrast}}$, $\\text{SEAM}\\_{\\text{GPT}}$, $\\text{SEAM}\\_{\\text{Adv}}$) to automate the quantification of seamlessness, which initially gives the impression of an augmentation-based approach. Although it becomes clear upon closer reading that these methods are intended to calculate the filtering criterion $\\epsilon(r\\_{i}, R\\_{\\theta})$, this initially gives the impression of an augmentation-based rather than filtering-based approach, leading to some confusion while reading.\n\n- The placement of certain visual content is inadequate. For instance, Figure 2 is excessively large, and there is no margin between its caption and the main text, which affects readability.\n\n- Performance evaluation across multiple datasets is insufficient. Testing on datasets beyond StackExchange is necessary.\n\n- It is essential to verify if similar results can be reproduced with other LLMs in addition to LLaMa.\n\n- The comparison study is lacking. Comparative evaluations with dataset filtering-based approaches, reward over-optimization methods focusing on RM and LM discrepancies, and other offline RLHF methods are necessary.", "questions": "- In the sanity check experiment in Section 4, the authors rely on the LLM-as-a-Judge method to measure the quality of the PM model. Could this evaluation approach itself be flawed?\n- In Figure 4 of Section 4.2, does Response B represent a low-quality response generated by the PM model (rank 5)? If so, why would human annotators prefer Response B despite its low quality? Doesn’t this imply that $\\mathcal{Q}\\_{PM}$ may be an unreliable metric?\n- Does $\\pi^{\\text{DATA}}$ refer to $\\mathcal{D}\\_{rl}$, which includes bad samples such as $\\text{SEAM}\\_{\\text{Contrast}}$, $\\text{SEAM}\\_{\\text{GPT}}$, or $\\text{SEAM}\\_{\\text{Adv}}$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study introduces the concept of \"seamlessness\" to improve Reinforcement Learning from Human Feedback (RLHF), focusing on the alignment between policy models (PMs) and reward models (RMs). The authors identify a \"saturation phenomenon,\" where advancements in PMs and RMs don’t lead to better RLHF outcomes, partly due to a 35% mismatch rate between RM scores and human preferences. To address this, they introduce SEAM, an automatic metric to quantify PM-RM discrepancies. Experiments show that SEAM-filtered data selection and SEAM-guided model augmentation improve RLHF performance by 4.5% and 4%, respectively. This study highlights the critical alignment between PMs and RMs for effective RLHF.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- This paper tackles an important and current topic in RLHF with a method that’s both intuitive and straightforward.\n- It lays out a clear research question and problem statement, supporting its claims with well-designed and diverse experiments.\n- Core concepts, like the \"saturation phenomenon,\" are explained in an accessible way that makes it easy for even new readers to follow. The use of visual aids is particularly effective, helping to deepen understanding and clarify key ideas.\n- The paper offers both theoretical insights and empirical evidence for the effectiveness of its approach. Especially interesting is its effort to explain the filtering effect by analyzing the \"effect of bad instruction\" from a post-filtering perspective, which adds an engaging layer to the discussion.", "weaknesses": "- Certain parts of the writing lack clarity:\n  - In Section 4, the explanation regarding Figure 3 under the sanity check section is not entirely clear. According to the Appendix, PM model quality is rated with a 1-10 score via the LLM-as-a-Judge method, but Figure 3 displays it simply as a percentage. A more detailed explanation of how the 1-10 score is converted to a percentage is needed. The same applies to RM quality.\n  - Definition (1) aims to define \"seamlessness.\" However, $S(I, R\\_{\\theta}, \\pi^{\\text{SFT}})$ is a metric that describes \"seam-ness\" by being proportional to RM misjudgment $\\epsilon(r, R\\_{\\theta})$.\n  - In Section 5.2, the authors describe three ways to construct samples ($\\text{SEAM}\\_{\\text{Contrast}}$, $\\text{SEAM}\\_{\\text{GPT}}$, $\\text{SEAM}\\_{\\text{Adv}}$) to automate the quantification of seamlessness, which initially gives the impression of an augmentation-based approach. Although it becomes clear upon closer reading that these methods are intended to calculate the filtering criterion $\\epsilon(r\\_{i}, R\\_{\\theta})$, this initially gives the impression of an augmentation-based rather than filtering-based approach, leading to some confusion while reading.\n\n- The placement of certain visual content is inadequate. For instance, Figure 2 is excessively large, and there is no margin between its caption and the main text, which affects readability.\n\n- Performance evaluation across multiple datasets is insufficient. Testing on datasets beyond StackExchange is necessary.\n\n- It is essential to verify if similar results can be reproduced with other LLMs in addition to LLaMa.\n\n- The comparison study is lacking. Comparative evaluations with dataset filtering-based approaches, reward over-optimization methods focusing on RM and LM discrepancies, and other offline RLHF methods are necessary.", "questions": "- In the sanity check experiment in Section 4, the authors rely on the LLM-as-a-Judge method to measure the quality of the PM model. Could this evaluation approach itself be flawed?\n- In Figure 4 of Section 4.2, does Response B represent a low-quality response generated by the PM model (rank 5)? If so, why would human annotators prefer Response B despite its low quality? Doesn’t this imply that $\\mathcal{Q}\\_{PM}$ may be an unreliable metric?\n- Does $\\pi^{\\text{DATA}}$ refer to $\\mathcal{D}\\_{rl}$, which includes bad samples such as $\\text{SEAM}\\_{\\text{Contrast}}$, $\\text{SEAM}\\_{\\text{GPT}}$, or $\\text{SEAM}\\_{\\text{Adv}}$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730464378601}, {"id": "yvaRJT8xuS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12132/Reviewer_nYdE"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper tries to analyze the interaction between the reward model (RM) and the trained policy model (PM), emphasizing the importance of consistency between them, named \"seamlessness\". It first shows that better RMs or PMs do not necessarily lead to better performances after RLHF (though it usually does), and tries to explain this because of lack of seamlessness. Then the paper motivates increasing seamlesness by removing the prompts on which the RM may be inacurrate. To this end, they introduce 3 seamlessness measures, SEAM_contrast, SEAM_GPT and SEAM_Adv, withut relying on human annotation: they can be used to filter out or augment the prompt dataset to improve RLHF performance.", "review_text": "The paper tries to analyze the interaction between the reward model (RM) and the trained policy model (PM), emphasizing the importance of consistency between them, named \"seamlessness\". It first shows that better RMs or PMs do not necessarily lead to better performances after RLHF (though it usually does), and tries to explain this because of lack of seamlessness. Then the paper motivates increasing seamlesness by removing the prompts on which the RM may be inacurrate. To this end, they introduce 3 seamlessness measures, SEAM_contrast, SEAM_GPT and SEAM_Adv, withut relying on human annotation: they can be used to filter out or augment the prompt dataset to improve RLHF performance.", "strengths": "* Understanding the interactions betwee the RM and the PM is an important and unexplored topic.\n* The analysis showing that better RMs do not lead to better policies is interesting, though not novel.", "weaknesses": "1.  The proposed SEAM methods (SEAM_contrast, SEAM_GPT, SEAM_Adv) seem to contradict the paper's focus on the interaction between the RM and policy.  Instead of tailoring the RM to the PM, these methods filters out challenging examples for the RM independently of the PM. An approach that explicitly considers the interplay between the RM and PM would be more consistent with the paper's main argument.\n\n2. Moreover, these 3 methods are computationally expensive, requiring large-scale data retrieval or GPT generations, limiting their practical applicability.\n\n3.  The reported 40% mismatch rate appears consistent with the 66.08% accuracy of the best model in Figure 2. It would also be nice reporting the match rate of other RMs.\n\n4. The theoretical analysis in Proposition 1 lacks clarity and depth. The proposition itself could be simplified by removing the unnecessary denominator.\n\n5. Nit: Scaling the RM should solve most of the problems according to the literature, thus I am sceptical about the Appendix B. If you find this result consistently, I would be worth investigating.", "questions": "* To analyze the interactions between diverse RMs and PMs, the study could consider different architectures and sizes (rather than reducing the training dataset size) as done in \"Llama 2: Open Foundation and Fine-Tuned Chat Models\". Then we could try to answer the following questions \"is a llama RM better for a Llama policy, or should we actually use another pretraining\" etc. Here the findings might be explain by inappropriate hyperparameters, and notably by the use of LoRA hyperparmaeters that might underfit larger datasets.\n\n* How does your figure 2 compare to the results from \"How to Evaluate Reward Models for RLHF\" showing that accuracy is actually the best signal for detecting the RM (despite its limitations).\n\n* What is the vertical dimension of Figure 4 continuous ? and not discrete ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tries to analyze the interaction between the reward model (RM) and the trained policy model (PM), emphasizing the importance of consistency between them, named \"seamlessness\". It first shows that better RMs or PMs do not necessarily lead to better performances after RLHF (though it usually does), and tries to explain this because of lack of seamlessness. Then the paper motivates increasing seamlesness by removing the prompts on which the RM may be inacurrate. To this end, they introduce 3 seamlessness measures, SEAM_contrast, SEAM_GPT and SEAM_Adv, withut relying on human annotation: they can be used to filter out or augment the prompt dataset to improve RLHF performance.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* Understanding the interactions betwee the RM and the PM is an important and unexplored topic.\n* The analysis showing that better RMs do not lead to better policies is interesting, though not novel.", "weaknesses": "1.  The proposed SEAM methods (SEAM_contrast, SEAM_GPT, SEAM_Adv) seem to contradict the paper's focus on the interaction between the RM and policy.  Instead of tailoring the RM to the PM, these methods filters out challenging examples for the RM independently of the PM. An approach that explicitly considers the interplay between the RM and PM would be more consistent with the paper's main argument.\n\n2. Moreover, these 3 methods are computationally expensive, requiring large-scale data retrieval or GPT generations, limiting their practical applicability.\n\n3.  The reported 40% mismatch rate appears consistent with the 66.08% accuracy of the best model in Figure 2. It would also be nice reporting the match rate of other RMs.\n\n4. The theoretical analysis in Proposition 1 lacks clarity and depth. The proposition itself could be simplified by removing the unnecessary denominator.\n\n5. Nit: Scaling the RM should solve most of the problems according to the literature, thus I am sceptical about the Appendix B. If you find this result consistently, I would be worth investigating.", "questions": "* To analyze the interactions between diverse RMs and PMs, the study could consider different architectures and sizes (rather than reducing the training dataset size) as done in \"Llama 2: Open Foundation and Fine-Tuned Chat Models\". Then we could try to answer the following questions \"is a llama RM better for a Llama policy, or should we actually use another pretraining\" etc. Here the findings might be explain by inappropriate hyperparameters, and notably by the use of LoRA hyperparmaeters that might underfit larger datasets.\n\n* How does your figure 2 compare to the results from \"How to Evaluate Reward Models for RLHF\" showing that accuracy is actually the best signal for detecting the RM (despite its limitations).\n\n* What is the vertical dimension of Figure 4 continuous ? and not discrete ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730403530778}], "openreview_url": "https://openreview.net/forum?id=M8xtZuxqC5", "arxiv_id": "2406.07971", "paper_pdf": "papers/M8xtZuxqC5.pdf", "paper_pdf_sha256": "a6181ad1689846cf792d235ea2e27e3367176bf71cd3bba1477d9b5f3a42373f", "paper_pdf_bytes": 1017798, "paper_pdf_source": "openreview", "code_url": "https://github.com/TaiMingLu/seamless", "code_repository": "TaiMingLu/seamless", "code_commit": "9a704f2a28800e162be53e61aee0a11ce4146841", "code_archive": "repos/M8xtZuxqC5.zip", "code_archive_sha256": "f76e0c4acce6fdd5adf45fcfd7ae19bc7a0981b3a84248f701bb8e1dc0eafeb3", "code_archive_bytes": 245646, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 243, "github_languages": {"Python": 54869}, "github_archived": false, "github_pushed_at": "2024-08-26T20:05:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/it-takes-two-on-the-seamlessness-between"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oNkYPgnfHt", "year": 2024, "status": "rejected", "title": "Learning to Intervene on Concept Bottlenecks", "authors": ["David Steinmann", "Wolfgang Stammer", "Felix Friedrich", "Kristian Kersting"], "authorids": ["~David_Steinmann1", "~Wolfgang_Stammer1", "~Felix_Friedrich1", "~Kristian_Kersting1"], "authors_source": "OpenReview API", "abstract": "While deep learning models often lack interpretability, concept bottleneck models (CBMs) provide inherent explanations via their concept representations. Moreover, they allow users to perform interventional interactions on these concepts by updating the concept values and thus correcting the predictive output of the model. Up to this point, these interventions were typically applied to the model just once and then discarded. To rectify this, we present concept bottleneck memory models (CB2Ms), which keep a memory of past interventions. Specifically, CB2Ms leverage a two-fold, differentiable memory to generalize interventions to appropriate novel situations, enabling the model to identify errors and reapply previous interventions. This way, a CB2M learns to automatically improve model performance from a few initially obtained interventions. If no prior human interventions are available, a CB2M can detect potential mistakes of the CBM bottleneck and request targeted interventions. Our experimental evaluations on challenging scenarios like handling distribution shifts and confounded data demonstrate that CB2Ms are able to successfully generalize interventions to unseen data and can indeed identify wrongly inferred concepts. Hence, CB2Ms are a valuable tool for users to provide interactive feedback on CBMs, e.g., by guiding a user’s interaction and requiring fewer interventions.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "X2kua0jJAr", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2489/Reviewer_fRVj"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes Concept Bottleneck Memory Models (CB2Ms), a new model-agnostic extension of Concept Bottleneck Models (CBMs) in which an adaptive memory is incorporated to improve a CBM’s receptiveness to test-time interventions and its uptake of feedback at test-time. Through two sets of distance-based memory banks, namely an *intervention memory* and a *mistake memory*,  CB2Ms learn to identify potentially mispredicted samples and reapply previous interventions to automatically improve the concept and task accuracy after only a handful of test-time interventions have been performed. This work evaluates CB2Ms on four datasets (two MNIST-based datasets and two real-world datasets) and shows that the proposed extensions enable CBM-based models to significantly boost their intervention performance and their ability to be more robust to distribution shift and train-time spurious correlations.", "review_text": "This paper proposes Concept Bottleneck Memory Models (CB2Ms), a new model-agnostic extension of Concept Bottleneck Models (CBMs) in which an adaptive memory is incorporated to improve a CBM’s receptiveness to test-time interventions and its uptake of feedback at test-time. Through two sets of distance-based memory banks, namely an *intervention memory* and a *mistake memory*,  CB2Ms learn to identify potentially mispredicted samples and reapply previous interventions to automatically improve the concept and task accuracy after only a handful of test-time interventions have been performed. This work evaluates CB2Ms on four datasets (two MNIST-based datasets and two real-world datasets) and shows that the proposed extensions enable CBM-based models to significantly boost their intervention performance and their ability to be more robust to distribution shift and train-time spurious correlations.", "strengths": "Thank you for submitting this work! I believe this is a very interesting idea and something that has certainly not been carefully explored in the concept-based literature before. Explicitly, I believe the following are the main strengths of this paper:\n\n1. **Originality**: the idea of incorporating a test-time adaptive memory to improve the uptake of intervention feedback and avoid discarding potentially useful information provided at intervention-time, is certainly original in the field of concept-based explainability. Furthermore, the work is well-placed within this area with a set of diverse related works discussed in this paper and potential ideas mentioned in the end discussion.\n2. **Quality**: although I believe the experimental set-up could have benefited from a more careful design (see below), the quality of the presented idea, as the presentation of the idea itself, is up to the standards of work in this conference and community.\n3. **Clarity**: the paper’s writing is very easy to follow, with almost no typos and a structure which makes the reading flow easily from beginning to end. This helps the authors clearly communicate their ideas and the motivation behind the ideas. Furthermore, the inclusion of a code base helps to understand the clarity of this work and promotes reproducibility, both highly desirable features.\n4. **Significance**: the paper’s main contribution, that of a model-agnostic mechanism to take test-time feedback into account when considering future interventions, is certainly significant in the concept-based literature and may lead to further advancement in the near future. Nevertheless, this significance is contingent on a careful and fair evaluation of the proposed method (see weaknesses below). If the doubts I have regarding this paper's experiments are carefully discussed/corrected, and the paper’s main claims still hold, then I believe this paper would be of good value to the community.", "weaknesses": "Although I think this paper has several strengths, as outlined above, I am concerned about the fairness of its evaluation and the lack of baselines that would be relevant comparison methods with what is presented in this manuscript. In particular, I believe these are the paper's main weaknesses:\n1. [Critical] My biggest concern with this paper is its evaluation. In my opinion, it is not fair to evaluate a method like CB2M, which can take and store extra samples at test time to improve performance on future unseen samples, against methods that completely ignore the same feedback even if in theory it could be used to improve their performance as well. For example, given that the validation set is used to construct the initial mistake memory, at the very least I would expect to see as a fair baseline a CBM that was able to update its weights using feedback from the same validation set during training (as otherwise CB2M is unfairly being trained on more data than the CBM baseline!). Similarly, when considering how the intervention memory is used, it would be fair to have as a baseline a CBM whose weights are updated at test time so that feedback from a new intervention is considered in the model. This can be done, say, via a variety of online learning algorithms that aim to minimize the cross entropy loss of the corrected concept prediction given the provided ground-truth (intervened) label (or even a baseline that runs a few gradient decent steps on this new intervention's label and corresponding sample). Having such a baseline and showing the CB2M still beats that baseline would provide very strong evidence that the memory mechanism introduced in this paper is worthwhile and novel compared to methods that already exist there to address similar problems.\n1. [Critical] Regarding the evaluation of the method on unseen data points (e.g., Table 1), it is unclear whether it is fair for ECB2Ms to be able to use the entire (unmodified) test set as part of their memory loading while none of the baselines can have the same benefit at training time. Evaluating CB2M and other baselines on a modified version of this test set seems like a very unfair comparison if none of the other methods were able to obtain any feedback from simulated interventions on the unmodified version of these samples. It may be fairer to avoid any sort of leakage from the test set into the training set of CB2M by avoiding loading into its initial memory samples that are highly related to the ones it will be evaluated on (this can be done by splitting the test set into two). I understand this is how this method may be used in practice (with the memory taking advantage of test-time samples to build a database of interventions and mistakes),; comparing it against a CBM's results on the same table makes it seem like this is an apples-with-apples comparison when in reality it may not be such. Making this distinction clearer, and if possible the evaluation fairer, would help a lot to indicate how this method works and how it is expected to be used.\n1. [Critical] With the exception of C-MNIST, the results shown in the identification of mistaken samples do not appear to be statistically significant (see the standard deviations). Because of this, it is unclear whether the proposed method identifies errors better than, say, the naive softmax baseline. Similarly, the effect on generalized interventions based on CB2M's detection on the full dataset (Table 3) does not appear to be statistically significantly better than what one observed when using the Softmax detection for all datasets (see standard deviations). This casts some doubt on the usability of this method.\n1. [Major] The use of a dynamic memory in CB2M implies that this method will either (a) struggle to scale to large concept spaces or spaces with a lot of variability in concepts (as it will require a significant number of examples before capturing the variance of the concept space and this requires CB2M to store all these sample's embeddings), or (b) require one will have to cap the size of the memory, leading to another hyperparameter that needs fine-tuning. As discussed in the questions below, the size of the memory may also lead to intractability at test time due to a large search space needed when correcting a mistake that was identified via the mistake memory. These aspects are not discussed anywhere in this paper.\n1. [Major] The proposed method depends on three hyperparameters that are reportedly crucial for the end performance of the model, namely $t_d$, $t_a$, and $k$. Nevertheless, I could not find a reference anywhere in this paper on how these hyperparameters are selected for the experiments reported. Notice that the mechanism to fine-tune $t_d$ based on a validation set is explained near Equations 2 and 3. However,  the actual values used to perform this validation-based search are not reported anywhere for the experiments in this paper. Furthermore, the dependency of $t_d$ with $t_a$ and $k$ is also not elaborated anywhere in the main text (although it is understood that $k = 1$ when doing generalized interventions).\n1. [Major] Related to the point above, there are no ablations showing how sensitive the results reported in this paper are to these hyperparameters and how important the validation set is to fine-tune them correctly. This hinders the understanding of how this method would fare in practice and how easy it is to use.\n1. [Major] It is unclear how the size of the validation set affects any of the results observed. Similarly, it is unclear how the number of test-time interventions affects the results seen (a crucial element to understand given the role these interventions take in improving the method's future test-time performance). More importantly, no ablations are provided to answer these important questions.", "questions": "Given my concerns outlined in the weaknesses above, I am leaning towards rejection at the moment. Nevertheless, I am more than happy to be proven wrong or corrected if I misunderstood a crucial part of this work. With this in mind, I hope the following questions, in no particular order, would help clarify some of the doubts on this work. If possible, I would appreciate it if the authors could elaborate on these concerns as they may serve as a good starting point for a discussion during rebuttal:\n1. Regarding my concern about the fairness of the evaluation, could you please let me know if I am misunderstanding something here? If not, could you please elaborate on how CB2Ms would fare against similar (fairer) baselines as the ones discussed in the weaknesses?\n1. Regarding my concern on the results of Table 1: could you please elaborate on why it would not be more fair to perform the evaluation on a dataset of samples whose unmodified versions have not been used to set up the initial memory of CB2M?\n1. Regarding my comments on the weaknesses for Figure 3: how would the curve shown for CB2M look vis-a-vis that of a vanilla CBM in which multiple groups are randomly intervened on? I am trying to fully understand how these results are unexpected or different to those seen on CBMs and CEMs in previous works (e.g., Shin et al. and Chauhan et al., both works cited on this paper).\n1. Could you please elaborate on the importance of the validation set size and the number of test-time interventions before evaluation on the results presented in this paper?\n1. Could you please elaborate on the hyperparameter selection process for the experiments in this paper (see weakness above)?\n1. Do you have a sense of how sensitive CB2Ms are to their hyperparameters? Are there any good strategies to select these hyperparameters?\n1. Similar to the question above, could you elaborate on how important the memory size is for this model? Are there any ablations on how memory size affects the results? It is unreasonable to assume that one can have a boundless memory for CB2M; therefore, fully understanding this question is essential.\n1. Similarly, how is the inference wall-clock performance affected by introducing the memory banks at inference time? I imagine there will be a hit but fully understanding how significant this hit is before this method becomes intractable is absolutely crucial for understanding its weaknesses and strengths.\n1. Just to confirm: is it the case that during the evaluation of CB2M the memory is left unmodified once evaluation starts on the test set? I would expect that to be the case for this evaluation to be fair, however I could not easily find this detail in the paper.\n1. For the results of Table 4, how many test interventions are needed for CB2M to achieve the observed results? I could not find this detail easily.\n\nBesides these questions, I found the following typos/minor errors that, if fixed, could improve the quality of the current manuscript:\n1. Page 4: \"Thus, with the ability to handle task (i)...\" seems to be missing something before the enumeration \"(i)\" begins.\n1. Page 8: \"improve a model via on the detected mistakes\" should probably be \"improve a model via the detected mistakes\"\n1. Page 8: Missing space in \"improvements.Even\"\n1. nit on Page 8: closing quotation is used for the beginning of \"full\" instead of the opening quotation (`` in LaTeX)\n1. Page 8: period used instead of comma in \"...the distribution shift. indicating...\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Concept Bottleneck Memory Models (CB2Ms), a new model-agnostic extension of Concept Bottleneck Models (CBMs) in which an adaptive memory is incorporated to improve a CBM’s receptiveness to test-time interventions and its uptake of feedback at test-time. Through two sets of distance-based memory banks, namely an *intervention memory* and a *mistake memory*,  CB2Ms learn to identify potentially mispredicted samples and reapply previous interventions to automatically improve the concept and task accuracy after only a handful of test-time interventions have been performed. This work evaluates CB2Ms on four datasets (two MNIST-based datasets and two real-world datasets) and shows that the proposed extensions enable CBM-based models to significantly boost their intervention performance and their ability to be more robust to distribution shift and train-time spurious correlations.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "Thank you for submitting this work! I believe this is a very interesting idea and something that has certainly not been carefully explored in the concept-based literature before. Explicitly, I believe the following are the main strengths of this paper:\n\n1. **Originality**: the idea of incorporating a test-time adaptive memory to improve the uptake of intervention feedback and avoid discarding potentially useful information provided at intervention-time, is certainly original in the field of concept-based explainability. Furthermore, the work is well-placed within this area with a set of diverse related works discussed in this paper and potential ideas mentioned in the end discussion.\n2. **Quality**: although I believe the experimental set-up could have benefited from a more careful design (see below), the quality of the presented idea, as the presentation of the idea itself, is up to the standards of work in this conference and community.\n3. **Clarity**: the paper’s writing is very easy to follow, with almost no typos and a structure which makes the reading flow easily from beginning to end. This helps the authors clearly communicate their ideas and the motivation behind the ideas. Furthermore, the inclusion of a code base helps to understand the clarity of this work and promotes reproducibility, both highly desirable features.\n4. **Significance**: the paper’s main contribution, that of a model-agnostic mechanism to take test-time feedback into account when considering future interventions, is certainly significant in the concept-based literature and may lead to further advancement in the near future. Nevertheless, this significance is contingent on a careful and fair evaluation of the proposed method (see weaknesses below). If the doubts I have regarding this paper's experiments are carefully discussed/corrected, and the paper’s main claims still hold, then I believe this paper would be of good value to the community.", "weaknesses": "Although I think this paper has several strengths, as outlined above, I am concerned about the fairness of its evaluation and the lack of baselines that would be relevant comparison methods with what is presented in this manuscript. In particular, I believe these are the paper's main weaknesses:\n1. [Critical] My biggest concern with this paper is its evaluation. In my opinion, it is not fair to evaluate a method like CB2M, which can take and store extra samples at test time to improve performance on future unseen samples, against methods that completely ignore the same feedback even if in theory it could be used to improve their performance as well. For example, given that the validation set is used to construct the initial mistake memory, at the very least I would expect to see as a fair baseline a CBM that was able to update its weights using feedback from the same validation set during training (as otherwise CB2M is unfairly being trained on more data than the CBM baseline!). Similarly, when considering how the intervention memory is used, it would be fair to have as a baseline a CBM whose weights are updated at test time so that feedback from a new intervention is considered in the model. This can be done, say, via a variety of online learning algorithms that aim to minimize the cross entropy loss of the corrected concept prediction given the provided ground-truth (intervened) label (or even a baseline that runs a few gradient decent steps on this new intervention's label and corresponding sample). Having such a baseline and showing the CB2M still beats that baseline would provide very strong evidence that the memory mechanism introduced in this paper is worthwhile and novel compared to methods that already exist there to address similar problems.\n1. [Critical] Regarding the evaluation of the method on unseen data points (e.g., Table 1), it is unclear whether it is fair for ECB2Ms to be able to use the entire (unmodified) test set as part of their memory loading while none of the baselines can have the same benefit at training time. Evaluating CB2M and other baselines on a modified version of this test set seems like a very unfair comparison if none of the other methods were able to obtain any feedback from simulated interventions on the unmodified version of these samples. It may be fairer to avoid any sort of leakage from the test set into the training set of CB2M by avoiding loading into its initial memory samples that are highly related to the ones it will be evaluated on (this can be done by splitting the test set into two). I understand this is how this method may be used in practice (with the memory taking advantage of test-time samples to build a database of interventions and mistakes),; comparing it against a CBM's results on the same table makes it seem like this is an apples-with-apples comparison when in reality it may not be such. Making this distinction clearer, and if possible the evaluation fairer, would help a lot to indicate how this method works and how it is expected to be used.\n1. [Critical] With the exception of C-MNIST, the results shown in the identification of mistaken samples do not appear to be statistically significant (see the standard deviations). Because of this, it is unclear whether the proposed method identifies errors better than, say, the naive softmax baseline. Similarly, the effect on generalized interventions based on CB2M's detection on the full dataset (Table 3) does not appear to be statistically significantly better than what one observed when using the Softmax detection for all datasets (see standard deviations). This casts some doubt on the usability of this method.\n1. [Major] The use of a dynamic memory in CB2M implies that this method will either (a) struggle to scale to large concept spaces or spaces with a lot of variability in concepts (as it will require a significant number of examples before capturing the variance of the concept space and this requires CB2M to store all these sample's embeddings), or (b) require one will have to cap the size of the memory, leading to another hyperparameter that needs fine-tuning. As discussed in the questions below, the size of the memory may also lead to intractability at test time due to a large search space needed when correcting a mistake that was identified via the mistake memory. These aspects are not discussed anywhere in this paper.\n1. [Major] The proposed method depends on three hyperparameters that are reportedly crucial for the end performance of the model, namely $t_d$, $t_a$, and $k$. Nevertheless, I could not find a reference anywhere in this paper on how these hyperparameters are selected for the experiments reported. Notice that the mechanism to fine-tune $t_d$ based on a validation set is explained near Equations 2 and 3. However,  the actual values used to perform this validation-based search are not reported anywhere for the experiments in this paper. Furthermore, the dependency of $t_d$ with $t_a$ and $k$ is also not elaborated anywhere in the main text (although it is understood that $k = 1$ when doing generalized interventions).\n1. [Major] Related to the point above, there are no ablations showing how sensitive the results reported in this paper are to these hyperparameters and how important the validation set is to fine-tune them correctly. This hinders the understanding of how this method would fare in practice and how easy it is to use.\n1. [Major] It is unclear how the size of the validation set affects any of the results observed. Similarly, it is unclear how the number of test-time interventions affects the results seen (a crucial element to understand given the role these interventions take in improving the method's future test-time performance). More importantly, no ablations are provided to answer these important questions.", "questions": "Given my concerns outlined in the weaknesses above, I am leaning towards rejection at the moment. Nevertheless, I am more than happy to be proven wrong or corrected if I misunderstood a crucial part of this work. With this in mind, I hope the following questions, in no particular order, would help clarify some of the doubts on this work. If possible, I would appreciate it if the authors could elaborate on these concerns as they may serve as a good starting point for a discussion during rebuttal:\n1. Regarding my concern about the fairness of the evaluation, could you please let me know if I am misunderstanding something here? If not, could you please elaborate on how CB2Ms would fare against similar (fairer) baselines as the ones discussed in the weaknesses?\n1. Regarding my concern on the results of Table 1: could you please elaborate on why it would not be more fair to perform the evaluation on a dataset of samples whose unmodified versions have not been used to set up the initial memory of CB2M?\n1. Regarding my comments on the weaknesses for Figure 3: how would the curve shown for CB2M look vis-a-vis that of a vanilla CBM in which multiple groups are randomly intervened on? I am trying to fully understand how these results are unexpected or different to those seen on CBMs and CEMs in previous works (e.g., Shin et al. and Chauhan et al., both works cited on this paper).\n1. Could you please elaborate on the importance of the validation set size and the number of test-time interventions before evaluation on the results presented in this paper?\n1. Could you please elaborate on the hyperparameter selection process for the experiments in this paper (see weakness above)?\n1. Do you have a sense of how sensitive CB2Ms are to their hyperparameters? Are there any good strategies to select these hyperparameters?\n1. Similar to the question above, could you elaborate on how important the memory size is for this model? Are there any ablations on how memory size affects the results? It is unreasonable to assume that one can have a boundless memory for CB2M; therefore, fully understanding this question is essential.\n1. Similarly, how is the inference wall-clock performance affected by introducing the memory banks at inference time? I imagine there will be a hit but fully understanding how significant this hit is before this method becomes intractable is absolutely crucial for understanding its weaknesses and strengths.\n1. Just to confirm: is it the case that during the evaluation of CB2M the memory is left unmodified once evaluation starts on the test set? I would expect that to be the case for this evaluation to be fair, however I could not easily find this detail in the paper.\n1. For the results of Table 4, how many test interventions are needed for CB2M to achieve the observed results? I could not find this detail easily.\n\nBesides these questions, I found the following typos/minor errors that, if fixed, could improve the quality of the current manuscript:\n1. Page 4: \"Thus, with the ability to handle task (i)...\" seems to be missing something before the enumeration \"(i)\" begins.\n1. Page 8: \"improve a model via on the detected mistakes\" should probably be \"improve a model via the detected mistakes\"\n1. Page 8: Missing space in \"improvements.Even\"\n1. nit on Page 8: closing quotation is used for the beginning of \"full\" instead of the opening quotation (`` in LaTeX)\n1. Page 8: period used instead of comma in \"...the distribution shift. indicating...\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698798910978}, {"id": "PE8xU15GzG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2489/Reviewer_oJVX"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes an extension to CBM architectures called CB2M in which interventions are not just used once, but rather stored in memory and reused on test data to improve performance and detect possibly similar errors. Results show their method is better than normal CBMs at doing this.", "review_text": "The paper proposes an extension to CBM architectures called CB2M in which interventions are not just used once, but rather stored in memory and reused on test data to improve performance and detect possibly similar errors. Results show their method is better than normal CBMs at doing this.", "strengths": "The paper has an interesting idea, I like the notion of learning from human feedback and improving over time or fixing edge cases in which the neural network is making the same mistakes over and over again.\n\nI also like the possibility of human-AI collaboration here where we could e.g. detect errors humans and/or AI are making at test time to try and make up the difference between the two. I certainly think this direction has a lot of potential and will form an important part of explainable ML going forward.\n\nEssentially this all comes from storing these cases in memory, but it's worth mentioning that this isn't the most novel idea (a lot so similar work exists in CBR etc.), but in this context of CBMs it is reasonably interesting.", "weaknesses": "The weakness of the paper is the evaluation I feel. The authors setup a few situations on a few common datasets to show the utility of their method. However, none of these experimental setups are particularly compelling, and somewhat contrived. I'm not sure if the point of the method is to increase model performance, or HCI, etc...\n\nFrom a performance perspective, take the CUB dataset, SOTA on this dataset is (last I checked) 92%+, but here their method is 88.7%. I understand that raw performance is probably not your goal here, but if you're using accuracy to assess the usefulness of your method, then this isn't really very compelling, as I can't e.g. use CB2M to squeeze more accuracy out of my models, so I am left wondering how it would be useful there. What would be great is to show you could break this 92% ceiling with your method, human feedback, and interventions etc...\n\nIn another vein, the human interventions are simulated, which again makes me wonder if humans could actually interact with the method how the authors propose they could.\n\nIf the authors could show their method e.g. working with a doctor to improve team performance overall, or just improve accuracy over a standard black-box, that would be very exciting, but they don't. So, I am left wondering what the application of this is at all. \n\n### Small things\n* The first two figures don't explain $f$ or $g$, the figures should stand alone usually.\n* Page 2: this issue (2) by... should be -- this issue by (2)....\n* I would tend to axe the third contribution on page 2, it's just experiment results which is expected.\n* It's not clear where $x_e$ is taken.\n* Second paragraph on page 4 would probably help the intro motivation.\n* $t_d$ needs to be clearly explained how the value was taken (unless I missed it sorry)\n* Eq 2: I wouldn't use $val$, it reminds me of \"validation\" personally, which is confusing.", "questions": "* What is a real-world application of this method that could make people in the ML community genuinely excited? Something were the method could be shown to be *understandable* and *useful* to intended practitioners of the system.\n* See above for my other general critiques.\n\nOverall, I like the paper's core idea, and I veer slightly (although just slightly) towards acceptance, but I will mutate this after the next phase depending on my interactions with the AC, reviewers, and authors.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an extension to CBM architectures called CB2M in which interventions are not just used once, but rather stored in memory and reused on test data to improve performance and detect possibly similar errors. Results show their method is better than normal CBMs at doing this.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper has an interesting idea, I like the notion of learning from human feedback and improving over time or fixing edge cases in which the neural network is making the same mistakes over and over again.\n\nI also like the possibility of human-AI collaboration here where we could e.g. detect errors humans and/or AI are making at test time to try and make up the difference between the two. I certainly think this direction has a lot of potential and will form an important part of explainable ML going forward.\n\nEssentially this all comes from storing these cases in memory, but it's worth mentioning that this isn't the most novel idea (a lot so similar work exists in CBR etc.), but in this context of CBMs it is reasonably interesting.", "weaknesses": "The weakness of the paper is the evaluation I feel. The authors setup a few situations on a few common datasets to show the utility of their method. However, none of these experimental setups are particularly compelling, and somewhat contrived. I'm not sure if the point of the method is to increase model performance, or HCI, etc...\n\nFrom a performance perspective, take the CUB dataset, SOTA on this dataset is (last I checked) 92%+, but here their method is 88.7%. I understand that raw performance is probably not your goal here, but if you're using accuracy to assess the usefulness of your method, then this isn't really very compelling, as I can't e.g. use CB2M to squeeze more accuracy out of my models, so I am left wondering how it would be useful there. What would be great is to show you could break this 92% ceiling with your method, human feedback, and interventions etc...\n\nIn another vein, the human interventions are simulated, which again makes me wonder if humans could actually interact with the method how the authors propose they could.\n\nIf the authors could show their method e.g. working with a doctor to improve team performance overall, or just improve accuracy over a standard black-box, that would be very exciting, but they don't. So, I am left wondering what the application of this is at all. \n\n### Small things\n* The first two figures don't explain $f$ or $g$, the figures should stand alone usually.\n* Page 2: this issue (2) by... should be -- this issue by (2)....\n* I would tend to axe the third contribution on page 2, it's just experiment results which is expected.\n* It's not clear where $x_e$ is taken.\n* Second paragraph on page 4 would probably help the intro motivation.\n* $t_d$ needs to be clearly explained how the value was taken (unless I missed it sorry)\n* Eq 2: I wouldn't use $val$, it reminds me of \"validation\" personally, which is confusing.", "questions": "* What is a real-world application of this method that could make people in the ML community genuinely excited? Something were the method could be shown to be *understandable* and *useful* to intended practitioners of the system.\n* See above for my other general critiques.\n\nOverall, I like the paper's core idea, and I veer slightly (although just slightly) towards acceptance, but I will mutate this after the next phase depending on my interactions with the AC, reviewers, and authors.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698681900930}, {"id": "Qagx0NM2do", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2489/Reviewer_XnGw"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Concept Bottleneck Models (CBMs) are an increasingly popular model class, designed to be more interpretable – and importantly *intervenable* by human users. However, querying humans for interventions on these models can be expensive, and models may make the same mistake repeatedly, resulting in repetitive interventions needed by users. In this work, these authors call attention to this problem and propose a new method – CB2M – to reconcile lack of reuse of interventions. CB2M is a modular extension to CBMs which leverages two memory banks: one which helps the model identify when a mistake is likely made in the output, and a second which reuses past interventions to correct such a mistake. The authors demonstrate the potential utility of CB2Ms across a range of experiments.", "review_text": "Concept Bottleneck Models (CBMs) are an increasingly popular model class, designed to be more interpretable – and importantly *intervenable* by human users. However, querying humans for interventions on these models can be expensive, and models may make the same mistake repeatedly, resulting in repetitive interventions needed by users. In this work, these authors call attention to this problem and propose a new method – CB2M – to reconcile lack of reuse of interventions. CB2M is a modular extension to CBMs which leverages two memory banks: one which helps the model identify when a mistake is likely made in the output, and a second which reuses past interventions to correct such a mistake. The authors demonstrate the potential utility of CB2Ms across a range of experiments.", "strengths": "The motivation for the work is superb. The authors call attention to an incredibly important and under-recognized problem in CBMs: that models may make the same mistake repeatedly, and requiring humans to inculcate the same intervention over-and-over again can be cognitively demanding – and ought to be unnecessary. The authors’ proposed method is clever and has the potential to have great impact in the broader CBM community. I believe the authors offer value to the ICLR – and broader human-centric ML –  communities by 1) calling attention to this reuse problem, and 2) offering a first possible solution. The paper is also very well-written.", "weaknesses": "While I believe that simply calling attention to the reuse problem in interventions, coupled with their method proposal, holds value for the broader community – I do not think that the experiments in their current form sufficiently demonstrate the value of CB2M. Experimental validity therefore holds me back from assigning a higher overall score. I believe sizable further experiments are needed to really strengthen the work (or at least clarification on the current interpretation). \n\nFirst, I am confused as to why the performance in Table 1 is lower for CB2Ms in the Full versus the Identified sets. Are examples in the Identified set those in which the memory module predicts that the example is misclassified? If so, why is the baseline CBM task performance so high (this would imply a high false positive rate?) It would be good for the authors to expand on possible False Positive and False Negative rates of the memory module, for each domain. \n\nSecond, the authors do not discuss the impact of the size of the memory module on performance. It would be good for the authors to have some kind of experiment(s) looking into the impact of thresholded sizes on the memory and intervention modules, as in practice, it’s possible that we may not be able to store all past instances? \n\nThird, I am not convinced that the authors’ selection of baselines is adequate. As the authors detail in the Related Work, there have been several efforts to learn intervention policies (e.g., CooP) which select which next example to query people over. Yet, the authors never compare to any of these policies. None of these prior policies, to my understanding, leverage reuse of past interventions – as such, the authors could make the case that their method is complementary to these approaches; i.e., could be combined with methods which learn to intervene, which could justify not including such a baseline. However, the authors do not make such claims. I would be keen for the authors in the rebuttal to expand on how their work relates to other intervention policies and why they did not compare against them as baselines. \n\nFourth, the authors emphasize that their approach is model-agnostic. However, all experiments are with a CBM backbone. In the absence of augmenting other concept-based systems with their modules, I do not think the authors should claim their method is model-agnostic. If the authors would like to emphasize this claim, I believe experiments are needed with at least one other concept-based system. Otherwise, I think it is fine to leave for future work, but the text should be couched as such. \n\nLastly, I do not think the authors adequately discuss the limitations of their work (see Questions below). It would be good for there to be a dedicated Limitations section, or at least further prose on the matter.", "questions": "I have raised most of my important questions in the Weaknesses section. In addition: \n\n- I am confused and concerned by Table 9 in the Appendix. The standard deviation is massive for CUB in particular. The authors note that the wide variance could be due to the threshold selection. However, it’s not clear to me why the threshold selected across the 5 seeds would vary so much that it leads to this level of variance in the number of generalized interventions? What is the variance in the selected threshold (can you please provide example threshold values?) and/or further explanation of what is happening here? \n- It’s not clear to me that CB2M is better than softmax at detecting (with the exception of the Parity C-MNIST domain). The performance of CB2M and softmax are within error bounds for CUB. Can the authors expand on why this may be further? \n- The authors’ assumption of perfect humans is sensible for this work. However, I would encourage the authors to think about how their module(s) may be challenged if human interventions are incorrect (or uncertain – e.g., Collins, Espinosa-Zarlenga et al, “Human Uncertainty in Concept Based Systems” AIES 2023). Such challenges would be worth expanding on in a Limitations section (of which the authors do not suitably have here). \n- As a minor sematic note (which does not affect my score): the authors caption Fig 3 as “Less is more” — but really, this shows that the less is enough / sufficient to achieve high task performance; not that less is more than having further interventions. I’d encourage the authors to change that caption :) \n- As another note, which does not impact my score, but would be nice for a revised version: have the authors looked at qualitative examples of the interventions / mistakes captured in the module? (e.g., Fig 2 of Chauhan et al, 2022)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Concept Bottleneck Models (CBMs) are an increasingly popular model class, designed to be more interpretable – and importantly *intervenable* by human users. However, querying humans for interventions on these models can be expensive, and models may make the same mistake repeatedly, resulting in repetitive interventions needed by users. In this work, these authors call attention to this problem and propose a new method – CB2M – to reconcile lack of reuse of interventions. CB2M is a modular extension to CBMs which leverages two memory banks: one which helps the model identify when a mistake is likely made in the output, and a second which reuses past interventions to correct such a mistake. The authors demonstrate the potential utility of CB2Ms across a range of experiments.", "soundness": "2 fair", "presentation": "3 good", "contribution": "4 excellent", "strengths": "The motivation for the work is superb. The authors call attention to an incredibly important and under-recognized problem in CBMs: that models may make the same mistake repeatedly, and requiring humans to inculcate the same intervention over-and-over again can be cognitively demanding – and ought to be unnecessary. The authors’ proposed method is clever and has the potential to have great impact in the broader CBM community. I believe the authors offer value to the ICLR – and broader human-centric ML –  communities by 1) calling attention to this reuse problem, and 2) offering a first possible solution. The paper is also very well-written.", "weaknesses": "While I believe that simply calling attention to the reuse problem in interventions, coupled with their method proposal, holds value for the broader community – I do not think that the experiments in their current form sufficiently demonstrate the value of CB2M. Experimental validity therefore holds me back from assigning a higher overall score. I believe sizable further experiments are needed to really strengthen the work (or at least clarification on the current interpretation). \n\nFirst, I am confused as to why the performance in Table 1 is lower for CB2Ms in the Full versus the Identified sets. Are examples in the Identified set those in which the memory module predicts that the example is misclassified? If so, why is the baseline CBM task performance so high (this would imply a high false positive rate?) It would be good for the authors to expand on possible False Positive and False Negative rates of the memory module, for each domain. \n\nSecond, the authors do not discuss the impact of the size of the memory module on performance. It would be good for the authors to have some kind of experiment(s) looking into the impact of thresholded sizes on the memory and intervention modules, as in practice, it’s possible that we may not be able to store all past instances? \n\nThird, I am not convinced that the authors’ selection of baselines is adequate. As the authors detail in the Related Work, there have been several efforts to learn intervention policies (e.g., CooP) which select which next example to query people over. Yet, the authors never compare to any of these policies. None of these prior policies, to my understanding, leverage reuse of past interventions – as such, the authors could make the case that their method is complementary to these approaches; i.e., could be combined with methods which learn to intervene, which could justify not including such a baseline. However, the authors do not make such claims. I would be keen for the authors in the rebuttal to expand on how their work relates to other intervention policies and why they did not compare against them as baselines. \n\nFourth, the authors emphasize that their approach is model-agnostic. However, all experiments are with a CBM backbone. In the absence of augmenting other concept-based systems with their modules, I do not think the authors should claim their method is model-agnostic. If the authors would like to emphasize this claim, I believe experiments are needed with at least one other concept-based system. Otherwise, I think it is fine to leave for future work, but the text should be couched as such. \n\nLastly, I do not think the authors adequately discuss the limitations of their work (see Questions below). It would be good for there to be a dedicated Limitations section, or at least further prose on the matter.", "questions": "I have raised most of my important questions in the Weaknesses section. In addition: \n\n- I am confused and concerned by Table 9 in the Appendix. The standard deviation is massive for CUB in particular. The authors note that the wide variance could be due to the threshold selection. However, it’s not clear to me why the threshold selected across the 5 seeds would vary so much that it leads to this level of variance in the number of generalized interventions? What is the variance in the selected threshold (can you please provide example threshold values?) and/or further explanation of what is happening here? \n- It’s not clear to me that CB2M is better than softmax at detecting (with the exception of the Parity C-MNIST domain). The performance of CB2M and softmax are within error bounds for CUB. Can the authors expand on why this may be further? \n- The authors’ assumption of perfect humans is sensible for this work. However, I would encourage the authors to think about how their module(s) may be challenged if human interventions are incorrect (or uncertain – e.g., Collins, Espinosa-Zarlenga et al, “Human Uncertainty in Concept Based Systems” AIES 2023). Such challenges would be worth expanding on in a Limitations section (of which the authors do not suitably have here). \n- As a minor sematic note (which does not affect my score): the authors caption Fig 3 as “Less is more” — but really, this shows that the less is enough / sufficient to achieve high task performance; not that less is more than having further interventions. I’d encourage the authors to change that caption :) \n- As another note, which does not impact my score, but would be nice for a revised version: have the authors looked at qualitative examples of the interventions / mistakes captured in the module? (e.g., Fig 2 of Chauhan et al, 2022)", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698059753650}], "openreview_url": "https://openreview.net/forum?id=oNkYPgnfHt", "arxiv_id": "2308.13453", "paper_pdf": "papers/oNkYPgnfHt.pdf", "paper_pdf_sha256": "efc4342a1a81161c983a12c369902374fb3289a48d06e054e1fa0ec2164f0e62", "paper_pdf_bytes": 469831, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-research/CB2M", "code_repository": "ml-research/CB2M", "code_commit": "17bb3d5c0682dbcb90d67ae9b3d5ac7f12bb71dd", "code_archive": "repos/oNkYPgnfHt.zip", "code_archive_sha256": "0ba0ee4727973b2556879d127bdd443a5a2e5158589399e0a22545e8e3a12417", "code_archive_bytes": 180534, "code_file_count": 24, "code_extensions": {".py": 23, ".sh": 1}, "github_disk_usage_kb": 170, "github_languages": {"Python": 175767, "Shell": 16336}, "github_archived": false, "github_pushed_at": "2024-05-27T09:55:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-intervene-on-concept-bottlenecks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lrzX-rNuRvw", "year": 2023, "status": "rejected", "title": "Understanding Rare Spurious Correlations in Neural Networks", "authors": ["Yao-Yuan Yang", "Chi-Ning Chou", "Kamalika Chaudhuri"], "authorids": ["~Yao-Yuan_Yang1", "~Chi-Ning_Chou1", "~Kamalika_Chaudhuri1"], "authors_source": "OpenReview API", "abstract": "Neural networks are known to use spurious correlations such as background information for classification. While prior work has looked at spurious correlations that are widespread in the training data, in this work, we investigate how sensitive neural networks are to $rare$ spurious correlations, which may be harder to detect and correct, and may lead to privacy leaks. We introduce spurious patterns correlated with a fixed class to a few training examples and find that it takes only a handful of such examples for the network to learn the correlation. Furthermore, these rare spurious correlations also impact accuracy and privacy. We empirically and theoretically analyze different factors involved in rare spurious correlations and propose mitigation methods accordingly. Specifically, we observe that $\\ell_2$ regularization and adding Gaussian noise to inputs can reduce the undesirable effects.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "hzJwyyXMXj", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1253/Reviewer_bwQt"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper is dedicated to the problem of spurious correlations. The work systematically studies fundamental questions such as how many spuriously correlated training points are necessary for a neural net to get biased towards learning it. Specifically, it investigates the domain in which spurious correlations are rare. Interestingly the study indicates that even a single spuriously correlated sample can bias the learning of a neural network. Finally, the authors highlight three regularization methods for mitigating spurious correlations.", "review_text": "The paper studies some important research questions in the domain of spurious correlations. The findings also provide non-trivial contributions to the community. However, because of the synthetic setup, the overall impression of this work is borderline.  I suggest the authors comment on the weaknesses mentioned above and I am open to changing my rating.", "strengths": "Strength:\n\n1. The paper is well-motivated, organized, easy to follow and well-written.\n2. The study aims to answer fundamental research questions, which can provide new insights to the community.\n3. The authors provide theoretical justification for their findings\n\nWeakness:\n1. The technical contribution is somewhat limited in the sense that prior works have already shown that neural networks are more biased towards \"easy-to-learn\" spurious attributes. \n2. A major chunk of the experimental analysis is based on cases where the spurious correlation is injected synthetically. It would be more interesting if the authors also showed results on commonly studied spuriously correlated datasets such as Waterbirds, and CelebA.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper is dedicated to the problem of spurious correlations. The work systematically studies fundamental questions such as how many spuriously correlated training points are necessary for a neural net to get biased towards learning it. Specifically, it investigates the domain in which spurious correlations are rare. Interestingly the study indicates that even a single spuriously correlated sample can bias the learning of a neural network. Finally, the authors highlight three regularization methods for mitigating spurious correlations.", "strength_and_weaknesses": "Strength:\n\n1. The paper is well-motivated, organized, easy to follow and well-written.\n2. The study aims to answer fundamental research questions, which can provide new insights to the community.\n3. The authors provide theoretical justification for their findings\n\nWeakness:\n1. The technical contribution is somewhat limited in the sense that prior works have already shown that neural networks are more biased towards \"easy-to-learn\" spurious attributes. \n2. A major chunk of the experimental analysis is based on cases where the spurious correlation is injected synthetically. It would be more interesting if the authors also showed results on commonly studied spuriously correlated datasets such as Waterbirds, and CelebA.", "clarity,_quality,_novelty_and_reproducibility": "Clarity is good. The content is well-organized and easy to follow. This work provides extensive experimental evaluations and sufficient details. Novelty is fair. ", "summary_of_the_review": "The paper studies some important research questions in the domain of spurious correlations. The findings also provide non-trivial contributions to the community. However, because of the synthetic setup, the overall impression of this work is borderline.  I suggest the authors comment on the weaknesses mentioned above and I am open to changing my rating.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666686761794}, {"id": "jc0Yt4iXT5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1253/Reviewer_LUVv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigates how sensitive neural networks are with respect to rare spurious correlations from three perspectives. First, studying how many training points with the spurious pattern would cause noticeable spurious correlations in the synthetic and real datasets. Second, studying how rare spurious correlations affect neural networks’ privacy and test accuracy. Third, studying how to mitigate the effects of rare spurious correlation.\n", "review_text": "With the findings above, I currently give the paper a borderline score.", "strengths": "Strength:\n1. The paper studies an interesting topic of learning with rare spurious correlations via both experimental and theoretical approaches.\n2. The paper studies whether rare spurious correlations are learned on both synthetic and real data.\n3. The paper studies the consequences of rare spurious correlations from a privacy and test accuracy perspective.\n4. The paper introduces simple and effective methods suggested by theoretical evidence to reduce the undesirable consequence of spurious correlation.\n5. The paper is well written.\n\nWeaknesses and questions:\n\n1. In Figure 1, it seems like the seven different spurious patterns (adding different kinds of noise) are not strong enough to generate a spurious correlation. How about using more challenging datasets, such as Colored MNIST?\n\n2. The main observations that neural networks can learn rare spurious correlations with few spurious examples are a little weak for me. It is not a big surprise to find out that adding perturbated patterns into training samples can impact the confidence of prediction on target classes during inference. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper investigates how sensitive neural networks are with respect to rare spurious correlations from three perspectives. First, studying how many training points with the spurious pattern would cause noticeable spurious correlations in the synthetic and real datasets. Second, studying how rare spurious correlations affect neural networks’ privacy and test accuracy. Third, studying how to mitigate the effects of rare spurious correlation.\n", "strength_and_weaknesses": "Strength:\n1. The paper studies an interesting topic of learning with rare spurious correlations via both experimental and theoretical approaches.\n2. The paper studies whether rare spurious correlations are learned on both synthetic and real data.\n3. The paper studies the consequences of rare spurious correlations from a privacy and test accuracy perspective.\n4. The paper introduces simple and effective methods suggested by theoretical evidence to reduce the undesirable consequence of spurious correlation.\n5. The paper is well written.\n\nWeaknesses and questions:\n\n1. In Figure 1, it seems like the seven different spurious patterns (adding different kinds of noise) are not strong enough to generate a spurious correlation. How about using more challenging datasets, such as Colored MNIST?\n\n2. The main observations that neural networks can learn rare spurious correlations with few spurious examples are a little weak for me. It is not a big surprise to find out that adding perturbated patterns into training samples can impact the confidence of prediction on target classes during inference. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper contains a lot of content and is easy to follow which is good, but the main observation and conclusion are a bit weak for me.", "summary_of_the_review": "With the findings above, I currently give the paper a borderline score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666576130795}, {"id": "2iOyCrLKXq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1253/Reviewer_ajJR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Most existing papers working on spurious correlation focus on the correlation that existed in the majority of the examples. This paper discovered that even a few spurious examples can still lead to the model learning the spurious correlation. Moreover, these \"rare\" spurious correlations bring negative effects regardless of the strength of spurious patterns, network architectures, and optimization methods.  Finally, the paper provides a theoretic analysis of a simple binary classification model. Based on the analysis, a few techniques are proposed to improve learning with rare spurious correlations. ", "review_text": "The paper discovers that even a few spurious examples can still lead to the model learning the spurious correlation. Many interesting results are shown with experiments: It shows that spurious patterns with larger empirical norms can cause spurious correlation more easily, and network architectures with higher sensitivity to its input are more susceptible to learning spurious correlations. ", "strengths": "Strength:\n- Extensive experiments are performed across datasets, the strength of spurious patterns, network architectures, and optimization methods, showing that even a small amount of spurious correlation will also influence the model performance. \n- Theoretical analysis of a simple but meaningful model is conducted, moreover, inspired by the analysis, several techniques are proposed and validated with experiments. \n\nWeaknesses:\n- Does the technique proposed also work when the majority of the data present spurious correlation, e.g. Adding l2 regularization? Because as mentioned in the paper: l2 regularization only \"suppresses the use of features that only appears on a small number of training examples\". ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Most existing papers working on spurious correlation focus on the correlation that existed in the majority of the examples. This paper discovered that even a few spurious examples can still lead to the model learning the spurious correlation. Moreover, these \"rare\" spurious correlations bring negative effects regardless of the strength of spurious patterns, network architectures, and optimization methods.  Finally, the paper provides a theoretic analysis of a simple binary classification model. Based on the analysis, a few techniques are proposed to improve learning with rare spurious correlations. ", "strength_and_weaknesses": "Strength:\n- Extensive experiments are performed across datasets, the strength of spurious patterns, network architectures, and optimization methods, showing that even a small amount of spurious correlation will also influence the model performance. \n- Theoretical analysis of a simple but meaningful model is conducted, moreover, inspired by the analysis, several techniques are proposed and validated with experiments. \n\nWeaknesses:\n- Does the technique proposed also work when the majority of the data present spurious correlation, e.g. Adding l2 regularization? Because as mentioned in the paper: l2 regularization only \"suppresses the use of features that only appears on a small number of training examples\". ", "clarity,_quality,_novelty_and_reproducibility": "The paper is very clear and of good quality. ", "summary_of_the_review": "The paper discovers that even a few spurious examples can still lead to the model learning the spurious correlation. Many interesting results are shown with experiments: It shows that spurious patterns with larger empirical norms can cause spurious correlation more easily, and network architectures with higher sensitivity to its input are more susceptible to learning spurious correlations. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666465756870}, {"id": "4rt6UtG5EUw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1253/Reviewer_yphb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studies how different numbers of spurious training examples (from one to thousands) affect neural nets. They measure how the predictions of test examples change when the spurious features are added. They show even a few spurious training examples can make the model outputs be affected by the spurious feature.\n\nThe authors consider two different types of spurious features. The first one is some artificial patterns, e.g., a square of yellow pixels. The second one is some natural spurious features in real data. They use the NICO++ dataset that has different backgrounds (grass, rock, water..) for each class. The authors then study the consequences of learning rare spurious features on privacy and accuracy. They also analyze the findings with a toy theoretical model and study several simple mitigations.\n", "review_text": "Given that 1) the findings about artificial spurious features seem incremental because of the backdoor attack literature; 2) how neural nets learn natural rare spurious features is not well studied. I'm leaning to reject this paper.", "strengths": "**Strength**\n\n1.Studying the relation between the number of spurious examples and the strength of spurious correlations is an important problem. The experiments show that neural nets could learn natural rare spurious correlations are new and interesting. \n\n2.This paper is well-written. Extensive results are well organized. \n\n3.The analysis on the privacy risk of spurious examples is new.\n\n**Weaknesses**\n\n1.The spurious features in Section 3.1 and 3.2 are very similar to backdoor triggers. They both are some artificial patterns that only appear a few times in the training set. For example, Chen et al. (2017) use random noise patterns. Gu et al. (2019) [1] use single-pixel and simple patterns as triggers. It is well-known that a few training examples with such triggers (rare spurious examples in this paper) would have a large impact on the trained model. \n\n2.How neural nets learn natural rare spurious correlations is unknown to the community (to the best of my knowledge). However, most of analysis and ablation studies use the artificial patterns instead of natural spurious correlations. Duplicating the same artificial pattern for multiple times is different from natural spurious features, which are complex and different in every example.\n\n3.What’s the experiment setup in Section 3.3? (data augmentation methods, learning rate, etc.).\n\n[1]: BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. https://messlab.moyix.net/papers/badnets_ieeeaccess19.pdf", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies how different numbers of spurious training examples (from one to thousands) affect neural nets. They measure how the predictions of test examples change when the spurious features are added. They show even a few spurious training examples can make the model outputs be affected by the spurious feature.\n\nThe authors consider two different types of spurious features. The first one is some artificial patterns, e.g., a square of yellow pixels. The second one is some natural spurious features in real data. They use the NICO++ dataset that has different backgrounds (grass, rock, water..) for each class. The authors then study the consequences of learning rare spurious features on privacy and accuracy. They also analyze the findings with a toy theoretical model and study several simple mitigations.\n", "strength_and_weaknesses": "**Strength**\n\n1.Studying the relation between the number of spurious examples and the strength of spurious correlations is an important problem. The experiments show that neural nets could learn natural rare spurious correlations are new and interesting. \n\n2.This paper is well-written. Extensive results are well organized. \n\n3.The analysis on the privacy risk of spurious examples is new.\n\n**Weaknesses**\n\n1.The spurious features in Section 3.1 and 3.2 are very similar to backdoor triggers. They both are some artificial patterns that only appear a few times in the training set. For example, Chen et al. (2017) use random noise patterns. Gu et al. (2019) [1] use single-pixel and simple patterns as triggers. It is well-known that a few training examples with such triggers (rare spurious examples in this paper) would have a large impact on the trained model. \n\n2.How neural nets learn natural rare spurious correlations is unknown to the community (to the best of my knowledge). However, most of analysis and ablation studies use the artificial patterns instead of natural spurious correlations. Duplicating the same artificial pattern for multiple times is different from natural spurious features, which are complex and different in every example.\n\n3.What’s the experiment setup in Section 3.3? (data augmentation methods, learning rate, etc.).\n\n[1]: BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. https://messlab.moyix.net/papers/badnets_ieeeaccess19.pdf", "clarity,_quality,_novelty_and_reproducibility": "Please see **Strength And Weaknesses**.", "summary_of_the_review": "Given that 1) the findings about artificial spurious features seem incremental because of the backdoor attack literature; 2) how neural nets learn natural rare spurious features is not well studied. I'm leaning to reject this paper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1665716892304}], "openreview_url": "https://openreview.net/forum?id=lrzX-rNuRvw", "arxiv_id": "2202.05189", "paper_pdf": "papers/lrzX-rNuRvw.pdf", "paper_pdf_sha256": "e2cc2471fa226504b443e31d33a2be91230d32b369561a85f417926cc0bb46cf", "paper_pdf_bytes": 1342183, "paper_pdf_source": "openreview", "code_url": "https://github.com/yangarbiter/rare-spurious-correlation", "code_repository": "yangarbiter/rare-spurious-correlation", "code_commit": "af036940cfcabb14ff70fe1849dcf3f03edb22aa", "code_archive": "repos/lrzX-rNuRvw.zip", "code_archive_sha256": "070323fc40b8ca4e97b1430ea021164c9a196aabbac59bec5579d72a702732c7", "code_archive_bytes": 373759, "code_file_count": 29, "code_extensions": {".py": 24, ".ipynb": 5}, "github_disk_usage_kb": 357, "github_languages": {"Jupyter Notebook": 1180515, "Python": 104529}, "github_archived": false, "github_pushed_at": "2022-06-05T04:37:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-rare-spurious-correlations-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YmONQIWli--", "year": 2022, "status": "rejected", "title": "Gotta Go Fast When Generating Data with Score-Based Models", "authors": ["Alexia Jolicoeur-Martineau", "Ke Li", "Rémi Piché-Taillefer", "Tal Kachman", "Ioannis Mitliagkas"], "authorids": ["~Alexia_Jolicoeur-Martineau1", "~Ke_Li1", "~Rémi_Piché-Taillefer1", "~Tal_Kachman1", "~Ioannis_Mitliagkas1"], "authors_source": "OpenReview API", "abstract": "Score-based (denoising diffusion) generative models have recently gained a lot of success in generating realistic and diverse data. These approaches define a forward diffusion process for transforming data to noise and generate data by reversing it (thereby going from noise to data). Unfortunately, current score-based models generate data very slowly due to the sheer number of score network evaluations required by numerical SDE solvers. \n   \nIn this work, we aim to accelerate this process by devising a more efficient SDE solver. Existing approaches rely on the Euler-Maruyama (EM) solver, which uses a fixed step size. We found that naively replacing it with other SDE solvers fares poorly - they either result in low-quality samples or become slower than EM. To get around this issue, we carefully devise an SDE solver with adaptive step sizes tailored to score-based generative models piece by piece. Our solver requires only two score function evaluations, rarely rejects samples, and leads to high-quality samples. Our approach generates data 2 to 10 times faster than EM while achieving better or equal sample quality. For high-resolution images, our method leads to significantly higher quality samples than all other methods tested. Our SDE solver has the benefit of requiring no step size tuning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "JdElWxd7Oga", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3893/Reviewer_4pA6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new numerical solver for stochastic differential equations and demonstrated significant improvement over existing ones such as the Euler-Maruyama solver in terms of the quality/computation trade-off for score-based generative modeling with SDEs.", "review_text": "## Strengths\n\n1. This paper proposes the first adaptive step-size numerical SDE solver for score-based generative modeling. Experimental results demonstrate clear improvement over previous methods like Euler-Maruyama methods and probability flow ODEs. It also demonstrates improvement over DDIM for a moderate budget of iteration numbers.\n\n2. The method itself is simple and clear, and only requires tuning one hyperparameter. \n\n## Weaknesses\n\n1. No theoretical understanding of the proposed numerical SDE solver. It would be better to include an analyze on the convergence order of the proposed approach.\n\n2. Results of predictor-corrector for VP SDEs in Table 1 are significantly worse than results reported in the original score sde paper. What could be the reason for this? Did you use the correct signal-to-noise ratio for the Langevin corrector?\n\n3. Some minor writing issues. The first page has a footnote \"equal contribution\", while all author names should be anonymized. In section 3.1.3, $E_q$ is defined to be a scalar (the $L_q$ norm), but it is referred as a vector in the expression for computing $\\| x\\|_2 $.\n\n4. In section 3.1.5, it is mentioned that a different step size is applied to each data sample. How is this implemented in deep learning frameworks? Will it cause noticeable slowdown on GPUs?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new numerical solver for stochastic differential equations and demonstrated significant improvement over existing ones such as the Euler-Maruyama solver in terms of the quality/computation trade-off for score-based generative modeling with SDEs.", "main_review": "## Strengths\n\n1. This paper proposes the first adaptive step-size numerical SDE solver for score-based generative modeling. Experimental results demonstrate clear improvement over previous methods like Euler-Maruyama methods and probability flow ODEs. It also demonstrates improvement over DDIM for a moderate budget of iteration numbers.\n\n2. The method itself is simple and clear, and only requires tuning one hyperparameter. \n\n## Weaknesses\n\n1. No theoretical understanding of the proposed numerical SDE solver. It would be better to include an analyze on the convergence order of the proposed approach.\n\n2. Results of predictor-corrector for VP SDEs in Table 1 are significantly worse than results reported in the original score sde paper. What could be the reason for this? Did you use the correct signal-to-noise ratio for the Langevin corrector?\n\n3. Some minor writing issues. The first page has a footnote \"equal contribution\", while all author names should be anonymized. In section 3.1.3, $E_q$ is defined to be a scalar (the $L_q$ norm), but it is referred as a vector in the expression for computing $\\| x\\|_2 $.\n\n4. In section 3.1.5, it is mentioned that a different step size is applied to each data sample. How is this implemented in deep learning frameworks? Will it cause noticeable slowdown on GPUs?", "summary_of_the_review": "This paper proposes a simple and effective numerical SDE integrator for score-based generative modeling based on SDEs. Paper can be stronger with a deeper theoretical understanding.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636347840338}, {"id": "BJuozbDdJk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3893/Reviewer_RPn6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper presents a number of treatments to speed up and improve the generation quality in the process of the reverse diffusion process, diffusion-based generative models. This paper is mostly empirical by nature, and authors suggested five techniques in improving the SDE solver performance.", "review_text": "1)\nI understand the paper's nature lies in the empirical side. However, some claims can be further elaborated through formal analyses. For example, Norm calculation in 3.1.3 can be analyzed by following algorithmic complexity, i.e. big-oh notations. I suggest that authors find such analysis opportunity throughout the paper. Moreover, it would be great if you can come up with a table to organize such complexity comparisons.\n\n2)\nI cannot follow why Eq 5 could be better in both quality and speed at the same time. Eq 5 has more constraints to setup the tolerance, so the speed-up seems to be an obvious benefit. However, the quality could be damaged while I also note that the quality is not being too much hurt from your result report. Why is that?\n\nThis question goes same to utilizing L2 norm, instead of L-$\\infty$ norm\n\n3)\nIs there any theoretic argument on the integration method? Which can be proved or contests as a proposition or a theorem?\nI could not fully comprehend the argument in the current text, and it just looks like swapping SDE solvers to find the best match.\n\n4)\nThese improvements can be further investigated through ablation studies, which seems to be must-do, in my opinion. Without the ablation study, I cannot argue which technique contributed to more or less.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a number of treatments to speed up and improve the generation quality in the process of the reverse diffusion process, diffusion-based generative models. This paper is mostly empirical by nature, and authors suggested five techniques in improving the SDE solver performance.", "main_review": "1)\nI understand the paper's nature lies in the empirical side. However, some claims can be further elaborated through formal analyses. For example, Norm calculation in 3.1.3 can be analyzed by following algorithmic complexity, i.e. big-oh notations. I suggest that authors find such analysis opportunity throughout the paper. Moreover, it would be great if you can come up with a table to organize such complexity comparisons.\n\n2)\nI cannot follow why Eq 5 could be better in both quality and speed at the same time. Eq 5 has more constraints to setup the tolerance, so the speed-up seems to be an obvious benefit. However, the quality could be damaged while I also note that the quality is not being too much hurt from your result report. Why is that?\n\nThis question goes same to utilizing L2 norm, instead of L-$\\infty$ norm\n\n3)\nIs there any theoretic argument on the integration method? Which can be proved or contests as a proposition or a theorem?\nI could not fully comprehend the argument in the current text, and it just looks like swapping SDE solvers to find the best match.\n\n4)\nThese improvements can be further investigated through ablation studies, which seems to be must-do, in my opinion. Without the ablation study, I cannot argue which technique contributed to more or less.", "summary_of_the_review": "Good paper with interesting and essential ideas in improving the usability of diffusion models. However, the current experimental result and the method justification are weak.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "There is no need to consider the ethical concern regarding on this submission", "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636062993018}, {"id": "RD-snkZRn-Q", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3893/Reviewer_ybV4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Score-based/diffusion-based generative models can achieve high sample quality. However, their sampling speed is slow due to the large number of evaluations required by numerical SDE solvers. This works aims to accelerate the sampling process by using a more efficient SDE solver. The proposed approach generates data 2 to 10 times faster than the baselines while achieving reasonably well sample qualities.", "review_text": "1. The major contribution of the paper seems to be an application of existing numerical solvers (Roberts, 2012) to sampling from SDE models. The technical novelty could be limited as the solver already exists despite the tricks used to improve the empirical results. \n2. There is no theoretical analysis of the method.\n3. It seems that the major contribution is on the empirical side. However, the empirical results are not very impressive especially when NFE is small. For instance, in Table 1, DDIM outperforms the proposed method when NFE=49. As the goal of the paper is to improve the sampling speed of score-based models, having a good result when NFE is small enough is important. An FID of 72.29 using 49 NFE might not be impressive since NFE=49 is already a large budget. Wondering when the proposed method can outperform DDIM when NFE=100.\n4. There are some formatting issues (e.g., too much space) on the first page. In the tables, some best values are in bold while some are not. It would be good to be consistent.\n5. Algorithm 1 should be better explained in section 3.2. The required parameters and their selections should also be explained more clearly.\n6. The paper writing can be improved. For instance, in section 3.1.1, \"the stochastic Improved Euler’s method\" should be explained more clearly.  \"Dynamic step size algorithm\" is an important contribution of the paper, the setting should be better explained. If possible, can move some details from the appendix to the main paper.\n7. It would also be interesting to perform experiments on even higher-dimensional images if model checkpoints are available.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Score-based/diffusion-based generative models can achieve high sample quality. However, their sampling speed is slow due to the large number of evaluations required by numerical SDE solvers. This works aims to accelerate the sampling process by using a more efficient SDE solver. The proposed approach generates data 2 to 10 times faster than the baselines while achieving reasonably well sample qualities.", "main_review": "1. The major contribution of the paper seems to be an application of existing numerical solvers (Roberts, 2012) to sampling from SDE models. The technical novelty could be limited as the solver already exists despite the tricks used to improve the empirical results. \n2. There is no theoretical analysis of the method.\n3. It seems that the major contribution is on the empirical side. However, the empirical results are not very impressive especially when NFE is small. For instance, in Table 1, DDIM outperforms the proposed method when NFE=49. As the goal of the paper is to improve the sampling speed of score-based models, having a good result when NFE is small enough is important. An FID of 72.29 using 49 NFE might not be impressive since NFE=49 is already a large budget. Wondering when the proposed method can outperform DDIM when NFE=100.\n4. There are some formatting issues (e.g., too much space) on the first page. In the tables, some best values are in bold while some are not. It would be good to be consistent.\n5. Algorithm 1 should be better explained in section 3.2. The required parameters and their selections should also be explained more clearly.\n6. The paper writing can be improved. For instance, in section 3.1.1, \"the stochastic Improved Euler’s method\" should be explained more clearly.  \"Dynamic step size algorithm\" is an important contribution of the paper, the setting should be better explained. If possible, can move some details from the appendix to the main paper.\n7. It would also be interesting to perform experiments on even higher-dimensional images if model checkpoints are available.", "summary_of_the_review": "The major contribution of the paper seems to be an application of existing numerical solvers to sampling from SDE models. The novelty could be limited as the solver already exists (despite the tricks proposed to improve empirical results). Given that the theoretical contribution is limited, the major contribution would be on the empirical side. However, the empirical results are not impressive enough: when using a small number of sampling steps, the proposed method has much worse results compared to DDIM. As the goal of the paper is to efficiently sample from SDE models, having strong performance when the sampling steps are small is important.\n\nThe paper writing can be improved. There are some formatting issues.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635909360992}, {"id": "MswWB7t__VX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3893/Reviewer_Px4w"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a new SDE solver for the reverse process in score-based models. The algorithm is fast and offers high quality, and avoids some hyperparameter tuning. There is theoretical analysis on the stability and bias of the algorithm. The paper also has experiments comparing the proposed algorithm to several baseline methods. ", "review_text": "The research problem of this paper is significant and interesting to the community of deep generative models, especially score-based models and denoising diffusion probabilistic models. The writing of this paper is clear. \n\nStrengths:\n- A novel SDE solver with a stability and bias analysis.\n- The empirical success of the algorithm, i.e., high generation quality with less steps (function evaluations). \n- The algorithm can readily be adapted to other score-based models with SDE. \n\nI am not an expert in numerical solvers for differential equations, so I will leave evaluation of this part to other reviewers.\n\nWeaknesses (and questions):\n- One major concern is about the title. The title is too exaggerated and does not reflect the content of the paper. Instead, I would like to see a title that is precise, for example, \"An efficient SDE solver for score-based models\". \n- Fig 1: What about the original Improved Euler baseline? \n- Section 3: The several components of the algorithm seem to be parameter tuning and engineering trick. There is no theoretical or even intuitive justification for why they offer better performance (Appendix F does not answer this question). In addition, I do not see any ablation study on the contribution of each component. \n- Section 3.1.3: The last equation ($\\ell_2$ norm) in this subsection cannot be understood.\n- The algorithm: Although the algorithm does not need stepsize or scheduler tuning, there are other hyperparameters involved so some tuning is still needed.\n- Experiments: How often do you reject under different NFE? There should be experiments on the relationship between NFE and percentage of rejection. Next, there should be a figure similar to Fig 1 but the x-axis is the time spent rather than NFE. \n- Appendix A: Are those methods $\\ell_2$ or $\\ell_{\\infty}$? \n- Appendix E: What can we read from this table?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents a new SDE solver for the reverse process in score-based models. The algorithm is fast and offers high quality, and avoids some hyperparameter tuning. There is theoretical analysis on the stability and bias of the algorithm. The paper also has experiments comparing the proposed algorithm to several baseline methods. ", "main_review": "The research problem of this paper is significant and interesting to the community of deep generative models, especially score-based models and denoising diffusion probabilistic models. The writing of this paper is clear. \n\nStrengths:\n- A novel SDE solver with a stability and bias analysis.\n- The empirical success of the algorithm, i.e., high generation quality with less steps (function evaluations). \n- The algorithm can readily be adapted to other score-based models with SDE. \n\nI am not an expert in numerical solvers for differential equations, so I will leave evaluation of this part to other reviewers.\n\nWeaknesses (and questions):\n- One major concern is about the title. The title is too exaggerated and does not reflect the content of the paper. Instead, I would like to see a title that is precise, for example, \"An efficient SDE solver for score-based models\". \n- Fig 1: What about the original Improved Euler baseline? \n- Section 3: The several components of the algorithm seem to be parameter tuning and engineering trick. There is no theoretical or even intuitive justification for why they offer better performance (Appendix F does not answer this question). In addition, I do not see any ablation study on the contribution of each component. \n- Section 3.1.3: The last equation ($\\ell_2$ norm) in this subsection cannot be understood.\n- The algorithm: Although the algorithm does not need stepsize or scheduler tuning, there are other hyperparameters involved so some tuning is still needed.\n- Experiments: How often do you reject under different NFE? There should be experiments on the relationship between NFE and percentage of rejection. Next, there should be a figure similar to Fig 1 but the x-axis is the time spent rather than NFE. \n- Appendix A: Are those methods $\\ell_2$ or $\\ell_{\\infty}$? \n- Appendix E: What can we read from this table?", "summary_of_the_review": "The algorithm seems to be empirically successful. However, the paper lacks justification for why their algorithm can be better. There is also much space for improvements in the experiments (especially, on the overall generation speed measured by seconds). In addition, the title is exaggerated. \n\n---------\nUpdates after the rebuttal\n\nI appreciate that the authors agreed to use a more precise title. Some of my concerns are also addressed. Therefore, I will increase my score to 5. However, I still think the experiments lack some key components such as time spent and number of rejections, although they might strongly correlate to NFE. The empirical results are good but not impressive enough. I encourage the authors to further improve the method and make additional theoretical/empirical justifications. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635891893390}], "openreview_url": "https://openreview.net/forum?id=YmONQIWli--", "arxiv_id": "2105.14080", "paper_pdf": "papers/YmONQIWli--.pdf", "paper_pdf_sha256": "dedf9d52686bacce2e4f6f10ddb437cedf99a4165593118417213a1e02f0e81a", "paper_pdf_bytes": 27898084, "paper_pdf_source": "openreview", "code_url": "https://github.com/AlexiaJM/score_sde_fast_sampling", "code_repository": "AlexiaJM/score_sde_fast_sampling", "code_commit": "5da8f3fe103ee5ac3c3a336f16cc06c9541f0ed9", "code_archive": "repos/YmONQIWli--.zip", "code_archive_sha256": "4bfa837e1daf0c4793bddb3b8fc3c976839c971d45fd7e59060f980218da6d84", "code_archive_bytes": 1164621, "code_file_count": 68, "code_extensions": {".py": 67, ".sh": 1}, "github_disk_usage_kb": 1240, "github_languages": {"Python": 313793, "Shell": 40262}, "github_archived": false, "github_pushed_at": "2021-11-20T21:59:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/gotta-go-fast-when-generating-data-with-score"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "avBunqDXFS", "year": 2021, "status": "rejected", "title": "Memory-Efficient Semi-Supervised Continual Learning: The World is its Own Replay Buffer", "authors": ["James Smith", "Jonathan C Balloch", "Yen-Chang Hsu", "Zsolt Kira"], "authorids": ["~James_Smith1", "~Jonathan_C_Balloch1", "~Yen-Chang_Hsu1", "~Zsolt_Kira1"], "authors_source": "OpenReview API", "abstract": "Rehearsal is a critical component for class-incremental continual learning, yet it requires a substantial memory budget. Our work investigates whether we can significantly reduce this memory budget by leveraging unlabeled data from an agent's environment in a realistic and challenging continual learning paradigm. Specifically, we explore and formalize a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce yet non-i.i.d. unlabeled data from the agent's environment is plentiful. Importantly, data distributions in the SSCL setting are realistic and therefore reflect object class correlations between, and among, the labeled and unlabeled data distributions. We show that a strategy built on pseudo-labeling, consistency regularization, Out-of-Distribution (OoD) detection, and knowledge distillation reduces forgetting in this setting. Our approach, DistillMatch, increases performance over the state-of-the-art by no less than 8.7% average task accuracy and up to a 54.5% increase in average task accuracy in SSCL CIFAR-100 experiments. Moreover, we demonstrate that DistillMatch can save up to 0.23 stored images per processed unlabeled image compared to the next best method which only saves 0.08. Our results suggest that focusing on realistic correlated distributions is a significantly new perspective, which accentuates the importance of leveraging the world's structure as a continual learning strategy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "z-PeszUupN7", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2073/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper investigates a semi-supervised continual learning (SSCL) setting and proposes a new method called DistillMatch for this setting. The major contributions are: (1) The authors carefully design a realistic SSCL setting where object-object correlations between labeled and unlabeled sets are maintained through a label super-class structure. And then, they develop the DistillMatch method combining knowledge distillation, pseudo-labels, out of distribution detection, and consistency regularization. (2) They show that  DistillMatch outperforms other existing methods on CIFAR-100 dataset, and ablation study results are shown also. \n\nHowever, there are some downsides that should be considered before its publication. (1) In abstract the authors claim that they can significantly reduce the memory budget (of labeled training data) by leveraging unlabeled data (perhaps with large volume). This motivation seems to be contradictive. (2) From a methodological viewpoint, the proposed DistillMatch method is just a combination of existing methods (listed as in above). So where is the novelty of this \"new\" method? (3) In experiments, the chosen baseline algorithm is very weak. There are some strong baseline methods such as GEM, A-GEM, and ER. So I wonder to know the real improvements over state-of-the-art methods for continual learning. (4) The label super-class structure existed in CIFAR-100 has been used in their experiments. But this is not very common for other more realistic datasets such as miniImageNet. If there is no super-class structure, we don't know how to apply the proposed DistillMatch method. \n\nIn summary, I think this semi-supervised continual learning setting is interesting, but the proposed DistillMatch method can not persuade me that this method is a novel significant contribution to this problem. So at present time I believe there is much room for the authors to improve their method before publication. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposed DistillMatch method appears to be a combination of knowledge distillation, out of distribution detection, consistency regularization and several other small tricks.", "review": "This paper investigates a semi-supervised continual learning (SSCL) setting and proposes a new method called DistillMatch for this setting. The major contributions are: (1) The authors carefully design a realistic SSCL setting where object-object correlations between labeled and unlabeled sets are maintained through a label super-class structure. And then, they develop the DistillMatch method combining knowledge distillation, pseudo-labels, out of distribution detection, and consistency regularization. (2) They show that  DistillMatch outperforms other existing methods on CIFAR-100 dataset, and ablation study results are shown also. \n\nHowever, there are some downsides that should be considered before its publication. (1) In abstract the authors claim that they can significantly reduce the memory budget (of labeled training data) by leveraging unlabeled data (perhaps with large volume). This motivation seems to be contradictive. (2) From a methodological viewpoint, the proposed DistillMatch method is just a combination of existing methods (listed as in above). So where is the novelty of this \"new\" method? (3) In experiments, the chosen baseline algorithm is very weak. There are some strong baseline methods such as GEM, A-GEM, and ER. So I wonder to know the real improvements over state-of-the-art methods for continual learning. (4) The label super-class structure existed in CIFAR-100 has been used in their experiments. But this is not very common for other more realistic datasets such as miniImageNet. If there is no super-class structure, we don't know how to apply the proposed DistillMatch method. \n\nIn summary, I think this semi-supervised continual learning setting is interesting, but the proposed DistillMatch method can not persuade me that this method is a novel significant contribution to this problem. So at present time I believe there is much room for the authors to improve their method before publication. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604044215589}, {"id": "L1bB59KXOOn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2073/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "- Summary:\nThis paper proposes class-incremental learning with unlabeled data correlated to labeled data, and a method to tackle it. The task can be considered as a variant of [Lee et al.], which has no assumption on the unlabeled dataset, while this paper assumes the correlation between labeled and unlabeled dataset explicitly. The proposed method is inspired by state-of-the-art class-incremental learning, semi-supervised learning, and out-of-distribution (OoD) detection methods: local distillation [Li and Hoiem], OoD detection [Hsu et al.], consistency regularization and pseudo labeling (or hard distillation) [Sohn et al.], and loss balancing based on class statistics [Lee et al.]. Experimental results support that the proposed method outperforms prior works in the proposed task.\n\n- Reasons for score:\n1. Extending continual learning to the semi-supervised setting is natural, given that the extension to self-taught learning has already been considered in [Lee et al.]. However, I cannot agree that semi-supervised learning is more realistic than self-taught learning, which is emphasized throughout the paper 18 times. In an early work of [Raina et al.], self-taught learning is proposed to make the scenario of learning with unlabeled data \"widely applicable to many practical learning problems.\" [Oliver et al.] also argued that \"(unlabeled data from out-of-distribution) violates the strict definition of semi-supervised learning, but it nevertheless represents a common use-case for semi-supervised learning (for example, augmenting a face recognition dataset with unlabeled images of people not in the labeled set).\" I am not saying that semi-supervised learning is unrealistic, but the argument in this paper sounds overclaimed. I believe both semi-supervised and self-taught learning are realistic in some cases. I also recommend to provide real world scenarios that the proposed task (correlation between labeled and unlabeled data exists and no memory for coreset is available) is useful in practice.\n2. The proposed method is not novel, which is essentially the combination of state-of-the-art methods in relevant tasks. But I do not discount this much, because this work would be valuable as the proposed task is interesting but not investigated before. However, the name of task might need to be changed, because a similar name, \"semi-supervised incremental learning\" is already taken by a kind of semi-supervised learning, which incrementally incorporates unlabeled data to training.\n3. Though the improvement over prior class-incremental learning methods is impressive, the overall performance is still too low. In fact, the scale of the experimental setting is too small, so I doubt it is scalable. All experiments are bounded on CIFAR-100, and even only 20% of training data are used as labeled one. Frankly, in this small-scale setting (in both number of data and image resolution), keeping all data is just fine, as the coreset size is negligible compared to the model size. I recommend to experiment in large-scale settings, e.g., on ImageNet. Also, I recommend to compare the oracle setting as well, which keeps all previous training data.\n4. In addition to small-scale experimental setting, the architecture is larger than the prior work [Lee et al.]: WRN-28-2 vs. WRN-16-2. In the worst case scenario, it is possible that the best performance of the proposed method is simply from the complexity of their learning objective, i.e., all methods overfit to training data, but the proposed method did not have enough updates to overfit to them.\n5. In Figure 3, why do GD and DM not have a coreset? I think there is no reason to give an unfair constraint to them. I recommend to draw curves with respect to increasing number of coreset for those methods as well.\n6. Could you provide results on the self-taught learning setting like [Lee et al.]? It would also be interesting to see the performance of the proposed method in the setting.\n7. Hyperparameter sweep results provided in Table 4 are either minimum or maximum of the range, so you could improve the performance by enlarging the range.\n\n- Minor Comments:\n8. Subscripts of theta often are dropped. Is theta equal to $\\theta_{n,1:n}$?\n9. \"the parameters of no more than three models\" -> I believe it is four, because you need to temporarily store gradients during training.\n10. $\\hat{q}$ is not a probability vector, which makes eq. (2) mathematically do not make sense.\n11. Citation format issue: you can use \\citet for noun and \\citep for adverb.\n12. typo on page 5: statoe -> state\n13. Table 4: what is TPR here? threshold for consistency regularization?\n\n[Raina et al.] Self-taught Learning: Transfer Learning from Unlabeled Data. In ICML, 2007.\n\n[Li and Hoiem] Learning without forgetting. In TPAMI, 2017.\n\n[Oliver et al.] Realistic Evaluation of Deep Semi-Supervised Learning Algorithms. In NeurIPS, 2018.\n\n[Lee et al.] Overcoming catastrophic forgetting with unlabeled data in the wild. In ICCV, 2019.\n\n[Hsu et al.] Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020.\n\n[Sohn et al.] Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020.\n\n**After rebuttal**\n\nI'd like to thank authors for their efforts to address my concerns. They have addressed most of them, so I increased my score from 5 to 6.\n\nHowever, there are two concerns that couldn't be resolved during the rebuttal period:\n\n(1) I am still not sure if the proposed task is practical. At glance it looks realistic, but I couldn't find a detailed scenario that can only be solved by the proposed task. Any real world scenario I can think of is closer to [Lee et al.], which is a prior work of this paper. Authors provided an exploring robot example in the thread of responses, but I think [Lee et al.] fits better for the provided one. I recommend authors to find a concrete use-case in real-world applications, which can only be solved by the proposed setting (or at least [Lee et al.] is not applicable; in the revised intro, you may emphasize that there are some real-world problems that [Lee et al.] is not applicable but yours is). R1 and R4 seem to have a similar concern.\n\n(2) the scale of experiment is too small. As CIFAR-10/100 have a limited number of data for your purpose,  you can borrow some data from tinyimages (FYI, CIFAR-10/100 are a subset of 80M tinyimages) or focus on ImageNet.\n\nI am okay with the lack of novelty on the proposed method. For a newly proposed task, I think proposing a simple and effective baseline is good enough. However, because of the two concerns above, I cannot strongly agree with its acceptance.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The proposed task is interesting, but more experiments are required", "review": "- Summary:\nThis paper proposes class-incremental learning with unlabeled data correlated to labeled data, and a method to tackle it. The task can be considered as a variant of [Lee et al.], which has no assumption on the unlabeled dataset, while this paper assumes the correlation between labeled and unlabeled dataset explicitly. The proposed method is inspired by state-of-the-art class-incremental learning, semi-supervised learning, and out-of-distribution (OoD) detection methods: local distillation [Li and Hoiem], OoD detection [Hsu et al.], consistency regularization and pseudo labeling (or hard distillation) [Sohn et al.], and loss balancing based on class statistics [Lee et al.]. Experimental results support that the proposed method outperforms prior works in the proposed task.\n\n- Reasons for score:\n1. Extending continual learning to the semi-supervised setting is natural, given that the extension to self-taught learning has already been considered in [Lee et al.]. However, I cannot agree that semi-supervised learning is more realistic than self-taught learning, which is emphasized throughout the paper 18 times. In an early work of [Raina et al.], self-taught learning is proposed to make the scenario of learning with unlabeled data \"widely applicable to many practical learning problems.\" [Oliver et al.] also argued that \"(unlabeled data from out-of-distribution) violates the strict definition of semi-supervised learning, but it nevertheless represents a common use-case for semi-supervised learning (for example, augmenting a face recognition dataset with unlabeled images of people not in the labeled set).\" I am not saying that semi-supervised learning is unrealistic, but the argument in this paper sounds overclaimed. I believe both semi-supervised and self-taught learning are realistic in some cases. I also recommend to provide real world scenarios that the proposed task (correlation between labeled and unlabeled data exists and no memory for coreset is available) is useful in practice.\n2. The proposed method is not novel, which is essentially the combination of state-of-the-art methods in relevant tasks. But I do not discount this much, because this work would be valuable as the proposed task is interesting but not investigated before. However, the name of task might need to be changed, because a similar name, \"semi-supervised incremental learning\" is already taken by a kind of semi-supervised learning, which incrementally incorporates unlabeled data to training.\n3. Though the improvement over prior class-incremental learning methods is impressive, the overall performance is still too low. In fact, the scale of the experimental setting is too small, so I doubt it is scalable. All experiments are bounded on CIFAR-100, and even only 20% of training data are used as labeled one. Frankly, in this small-scale setting (in both number of data and image resolution), keeping all data is just fine, as the coreset size is negligible compared to the model size. I recommend to experiment in large-scale settings, e.g., on ImageNet. Also, I recommend to compare the oracle setting as well, which keeps all previous training data.\n4. In addition to small-scale experimental setting, the architecture is larger than the prior work [Lee et al.]: WRN-28-2 vs. WRN-16-2. In the worst case scenario, it is possible that the best performance of the proposed method is simply from the complexity of their learning objective, i.e., all methods overfit to training data, but the proposed method did not have enough updates to overfit to them.\n5. In Figure 3, why do GD and DM not have a coreset? I think there is no reason to give an unfair constraint to them. I recommend to draw curves with respect to increasing number of coreset for those methods as well.\n6. Could you provide results on the self-taught learning setting like [Lee et al.]? It would also be interesting to see the performance of the proposed method in the setting.\n7. Hyperparameter sweep results provided in Table 4 are either minimum or maximum of the range, so you could improve the performance by enlarging the range.\n\n- Minor Comments:\n8. Subscripts of theta often are dropped. Is theta equal to $\\theta_{n,1:n}$?\n9. \"the parameters of no more than three models\" -> I believe it is four, because you need to temporarily store gradients during training.\n10. $\\hat{q}$ is not a probability vector, which makes eq. (2) mathematically do not make sense.\n11. Citation format issue: you can use \\citet for noun and \\citep for adverb.\n12. typo on page 5: statoe -> state\n13. Table 4: what is TPR here? threshold for consistency regularization?\n\n[Raina et al.] Self-taught Learning: Transfer Learning from Unlabeled Data. In ICML, 2007.\n\n[Li and Hoiem] Learning without forgetting. In TPAMI, 2017.\n\n[Oliver et al.] Realistic Evaluation of Deep Semi-Supervised Learning Algorithms. In NeurIPS, 2018.\n\n[Lee et al.] Overcoming catastrophic forgetting with unlabeled data in the wild. In ICCV, 2019.\n\n[Hsu et al.] Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020.\n\n[Sohn et al.] Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020.\n\n**After rebuttal**\n\nI'd like to thank authors for their efforts to address my concerns. They have addressed most of them, so I increased my score from 5 to 6.\n\nHowever, there are two concerns that couldn't be resolved during the rebuttal period:\n\n(1) I am still not sure if the proposed task is practical. At glance it looks realistic, but I couldn't find a detailed scenario that can only be solved by the proposed task. Any real world scenario I can think of is closer to [Lee et al.], which is a prior work of this paper. Authors provided an exploring robot example in the thread of responses, but I think [Lee et al.] fits better for the provided one. I recommend authors to find a concrete use-case in real-world applications, which can only be solved by the proposed setting (or at least [Lee et al.] is not applicable; in the revised intro, you may emphasize that there are some real-world problems that [Lee et al.] is not applicable but yours is). R1 and R4 seem to have a similar concern.\n\n(2) the scale of experiment is too small. As CIFAR-10/100 have a limited number of data for your purpose,  you can borrow some data from tinyimages (FYI, CIFAR-10/100 are a subset of 80M tinyimages) or focus on ImageNet.\n\nI am okay with the lack of novelty on the proposed method. For a newly proposed task, I think proposing a simple and effective baseline is good enough. However, because of the two concerns above, I cannot strongly agree with its acceptance.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603882837929}, {"id": "3zdGcKtO1QT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2073/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce and unlabeled data is plentiful. The proposed framework is built on pseudo-labeling, consistency regularization, Out-of-Distribution (OoD) detection,\nand knowledge distillation in order to reduce the catastrophic forgetting in the proposed setting.\n\nThe paper is in general clear and well-written. The contributions are clearly highlighted and the proposed approach is conveniently compared with other state of the art methods, demonstrating its superiority.\n\nPositive aspects: \n-  the definition of a realistic, semi-supervised setting for continual learning\n- a novel approach for continual learning in order to cope with 'catastrophic forgetting'\n- the proposed approach is memory efficient, since it does not need exemplars to replay past tasks\n\nNegative aspects:\n- the OoD implemented in this paper rejects the unknown samples. In other words, all unknown samples are considered a single class. It would have been a plus to distinguish between several unknown classes and somehow introduce them in the framework\n- the lack of recabilibration step after a number of tasks (in the case of pseudo-labeled samples), could lead to an undesired error propagation which is not quantified in the paper \n\nHowever, I have some questions:\n1. What is the relationship between the 'fi' and 'theta' models (section 4)? Are they completely separate or there is a relationship between them?\nFor instance, when 'theta' is extended with a new task, is 'fi' extended accordingly? Or is 'fi' trained off-line from the beginning (with all tasks)?\n2. There are some different source of errors: distilation, pseudo-labels... Do you perform any kind of system re-calibration? After how many tasks? I mean, do you make a study of error propagation of pseudo-labeled data? Or at some point do you have a human-in-the-loop to correct mis-classification? What is the mis-classification error of pseudo-labeled samples?\n3. Do you assume that labeled and unlabeled data come from different distributions or you have a single distribution which is \ndivided in labeled and unlabeled data at the beginning of the process?\n4. Does your scenario foresee that when learning a new task T, all the previous tasks are represented (1..T-1) in the unlabebeld data or only a subpart? (i.e. kind of selective replay)\n5. When the number of tasks increases, the number of unlabeled data per task remains constant or is scaled accordingly (i.e. reduced) ?\n6. Would be interesting to test your approach in a real-world scenario, i.e. robot navigation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Recommendation to Accept", "review": "The paper presents a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce and unlabeled data is plentiful. The proposed framework is built on pseudo-labeling, consistency regularization, Out-of-Distribution (OoD) detection,\nand knowledge distillation in order to reduce the catastrophic forgetting in the proposed setting.\n\nThe paper is in general clear and well-written. The contributions are clearly highlighted and the proposed approach is conveniently compared with other state of the art methods, demonstrating its superiority.\n\nPositive aspects: \n-  the definition of a realistic, semi-supervised setting for continual learning\n- a novel approach for continual learning in order to cope with 'catastrophic forgetting'\n- the proposed approach is memory efficient, since it does not need exemplars to replay past tasks\n\nNegative aspects:\n- the OoD implemented in this paper rejects the unknown samples. In other words, all unknown samples are considered a single class. It would have been a plus to distinguish between several unknown classes and somehow introduce them in the framework\n- the lack of recabilibration step after a number of tasks (in the case of pseudo-labeled samples), could lead to an undesired error propagation which is not quantified in the paper \n\nHowever, I have some questions:\n1. What is the relationship between the 'fi' and 'theta' models (section 4)? Are they completely separate or there is a relationship between them?\nFor instance, when 'theta' is extended with a new task, is 'fi' extended accordingly? Or is 'fi' trained off-line from the beginning (with all tasks)?\n2. There are some different source of errors: distilation, pseudo-labels... Do you perform any kind of system re-calibration? After how many tasks? I mean, do you make a study of error propagation of pseudo-labeled data? Or at some point do you have a human-in-the-loop to correct mis-classification? What is the mis-classification error of pseudo-labeled samples?\n3. Do you assume that labeled and unlabeled data come from different distributions or you have a single distribution which is \ndivided in labeled and unlabeled data at the beginning of the process?\n4. Does your scenario foresee that when learning a new task T, all the previous tasks are represented (1..T-1) in the unlabebeld data or only a subpart? (i.e. kind of selective replay)\n5. When the number of tasks increases, the number of unlabeled data per task remains constant or is scaled accordingly (i.e. reduced) ?\n6. Would be interesting to test your approach in a real-world scenario, i.e. robot navigation.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603877614929}, {"id": "HEP6e7aiTZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2073/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper comes up with a novel scenario where the unlabled data are available as well as labeled data in the continual learning scenario.\n\n### Overall\n- Based on my understanding, the major contribution is the proposal of a task scenario, aka, experimental setting. The novelty of DistillMatch is an incremental modification of previous work. \n- The task setting sidesteps the learning with non-stationarity problem than solving it.\n- Further, this setting potentially makes the task easier for the proposed method. To verify whether this is true, more information are needed.\n- The presentation of the paper needs polishing, I listed a few points below.\n\n### Pros\n- The novel scenario of semisupervised continual learning is proposed. The argument is that in several realistic scenarios old data are often re-observed without label (The funiture labeling example). Therefore instead of storing a coreset, one may make use of the unlabeled data for pseudo-rehearsal/distillation. It is reasonable to make use of it when this assumption is true.\n- With the setting the author proposed, the DistillMatch method is able to perform better than previous methods. \n\n### Cons\n1. The novelty mostly comes from the task scenario, the DistillMatch method is incremental.\n2. Although SSCL is a new scenario, and the author argues it is more realistic. IMO taking this assumption sidesteps the problem of continual learning rather than solving it. The central problem of continual learning IMO is to learn under non-stationary distribution, the assumption made in this submission makes the distribution more stationary.\n3. It is true that this assumption should be utilized when available. However, the only dataset used is manually constructed from CIFAR100, contradicting the initial motivation to move towards a more realistic scenario.\n4. There's a lack of information on how the compared methods are adapted to the new scenario. \nI searched the supplementary but failed to find a detailed documentation. With the given information, it is hard to tell whether the comparison is fair. My concerns are following,\n**increasing from 3 -> 4 as this point is resolved in the rebuttal**\n  - In the RandomClasses setting, it is stated that no coreset is used, if the compared methods depends on coreset to replay, it would be unfair. If that's the case, the only conclusion we can draw is that replay is better than no replay, which seems trivial to me.\n  - GD depends on internet crawled data, is it replaced with the unlabeled data since it is available in the experiment setting? If not, then I think it is just the setting that favors DistillMatch.\n  - With the above said, I suggest the author to list clearly the objectives, replay buffer sizes or even pseudo code for each of the compared method and their own method in a table, which will help the reader identify what major component in the proposed method is making the contribution.\n\nRegarding the quality and clarity,\nI found myself confused and making guesses sometimes while reading it.\n\nTo list a few:\n- introduction paragraph 2, ... to determine which unlabeled data is relevant to the incremental task ..., I guess the incremental task means learning the newly observed data, but then for rehearsal we'll pick the unlabeled data which is from the distribution of past tasks.\n- section 1, ... save up to 0.23 stored images per processed image over naive rehearsal (compared to Lee) ..., here seems Lee et al is the naive rehearsal. But then \"which only saved 0.08\" confuses me, seems to be saying Lee saves 0.08 compared to naive rehearsal.\n- section 3, ... where data distributions reflect object class correlations between, and among, the labeled and unlabeled data distributions ... not enough information to infer what \"reflect\" and \"object class correlation\" means here.\n- section 4, ... Let S_{n-1} denote the score of our OoD detector for valid classes of our pseudo-label model ... what is the \"valid classes\" needs to be clarified. As I understand it, S_{n-1} measures how likely the unlabeled data is in the distribution of past tasks.\n- Super class / Parent class are not defined clear enough.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Proposes a semisupervised task scenario for continual learning", "review": "This paper comes up with a novel scenario where the unlabled data are available as well as labeled data in the continual learning scenario.\n\n### Overall\n- Based on my understanding, the major contribution is the proposal of a task scenario, aka, experimental setting. The novelty of DistillMatch is an incremental modification of previous work. \n- The task setting sidesteps the learning with non-stationarity problem than solving it.\n- Further, this setting potentially makes the task easier for the proposed method. To verify whether this is true, more information are needed.\n- The presentation of the paper needs polishing, I listed a few points below.\n\n### Pros\n- The novel scenario of semisupervised continual learning is proposed. The argument is that in several realistic scenarios old data are often re-observed without label (The funiture labeling example). Therefore instead of storing a coreset, one may make use of the unlabeled data for pseudo-rehearsal/distillation. It is reasonable to make use of it when this assumption is true.\n- With the setting the author proposed, the DistillMatch method is able to perform better than previous methods. \n\n### Cons\n1. The novelty mostly comes from the task scenario, the DistillMatch method is incremental.\n2. Although SSCL is a new scenario, and the author argues it is more realistic. IMO taking this assumption sidesteps the problem of continual learning rather than solving it. The central problem of continual learning IMO is to learn under non-stationary distribution, the assumption made in this submission makes the distribution more stationary.\n3. It is true that this assumption should be utilized when available. However, the only dataset used is manually constructed from CIFAR100, contradicting the initial motivation to move towards a more realistic scenario.\n4. There's a lack of information on how the compared methods are adapted to the new scenario. \nI searched the supplementary but failed to find a detailed documentation. With the given information, it is hard to tell whether the comparison is fair. My concerns are following,\n**increasing from 3 -> 4 as this point is resolved in the rebuttal**\n  - In the RandomClasses setting, it is stated that no coreset is used, if the compared methods depends on coreset to replay, it would be unfair. If that's the case, the only conclusion we can draw is that replay is better than no replay, which seems trivial to me.\n  - GD depends on internet crawled data, is it replaced with the unlabeled data since it is available in the experiment setting? If not, then I think it is just the setting that favors DistillMatch.\n  - With the above said, I suggest the author to list clearly the objectives, replay buffer sizes or even pseudo code for each of the compared method and their own method in a table, which will help the reader identify what major component in the proposed method is making the contribution.\n\nRegarding the quality and clarity,\nI found myself confused and making guesses sometimes while reading it.\n\nTo list a few:\n- introduction paragraph 2, ... to determine which unlabeled data is relevant to the incremental task ..., I guess the incremental task means learning the newly observed data, but then for rehearsal we'll pick the unlabeled data which is from the distribution of past tasks.\n- section 1, ... save up to 0.23 stored images per processed image over naive rehearsal (compared to Lee) ..., here seems Lee et al is the naive rehearsal. But then \"which only saved 0.08\" confuses me, seems to be saying Lee saves 0.08 compared to naive rehearsal.\n- section 3, ... where data distributions reflect object class correlations between, and among, the labeled and unlabeled data distributions ... not enough information to infer what \"reflect\" and \"object class correlation\" means here.\n- section 4, ... Let S_{n-1} denote the score of our OoD detector for valid classes of our pseudo-label model ... what is the \"valid classes\" needs to be clarified. As I understand it, S_{n-1} measures how likely the unlabeled data is in the distribution of past tasks.\n- Super class / Parent class are not defined clear enough.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603665755895}], "openreview_url": "https://openreview.net/forum?id=avBunqDXFS", "arxiv_id": "2101.09536", "paper_pdf": "papers/avBunqDXFS.pdf", "paper_pdf_sha256": "a243c7f2ef13199a5776f9ad2e7dd477f99429e39b6f68bba1fe84eaedcf2df6", "paper_pdf_bytes": 1635086, "paper_pdf_source": "openreview", "code_url": "https://github.com/GT-RIPL/DistillMatch-SSCL", "code_repository": "GT-RIPL/DistillMatch-SSCL", "code_commit": "e572671fd6994b3c43ad6e46e9efb3588804524c", "code_archive": "repos/avBunqDXFS.zip", "code_archive_sha256": "938dfa7e51feb599dec6be7a12a78c0ef6d0f0f7c9f2ba3f71c63438d9892c8c", "code_archive_bytes": 1195938, "code_file_count": 22, "code_extensions": {".py": 12, ".sh": 10}, "github_disk_usage_kb": 1163, "github_languages": {"Python": 165881, "Shell": 36248}, "github_archived": false, "github_pushed_at": "2022-10-17T11:46:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/memory-efficient-semi-supervised-continual-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B1lda1HtvB", "year": 2020, "status": "rejected", "title": "Feature Selection using Stochastic Gates", "authors": ["Yutaro Yamada", "Ofir Lindenbaum", "Sahand Negahban", "Yuval Kluger"], "authorids": ["yutaro.yamada@yale.edu", "ofirlin@gmail.com", "sahand.negahban@yale.edu", "yuval.kluger@yale.edu"], "authors_source": "OpenReview API", "abstract": "Feature selection problems have been extensively studied in the setting of\nlinear estimation, for instance LASSO, but less emphasis has been placed on\nfeature selection for non-linear functions. In this study, we propose a method\nfor feature selection in high-dimensional non-linear function estimation\nproblems. The new procedure is based on directly penalizing the $\\ell_0$ norm of\nfeatures, or the count of the number of selected features. Our $\\ell_0$ based regularization relies on a continuous relaxation of the Bernoulli distribution, which\nallows our model to learn the parameters of the approximate Bernoulli\ndistributions via gradient descent. The proposed framework simultaneously learns\na non-linear regression or classification function while selecting a small\nsubset of features. We provide an information-theoretic justification for\nincorporating Bernoulli distribution into our approach. Furthermore, we evaluate\nour method using synthetic and real-life data and demonstrate that our approach\noutperforms other embedded methods in terms of predictive performance and feature selection.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H1e8061p9H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1994/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a feature selection method for high-dimensional datasets that attempts to fit a model while selecting relevant features. \nThe strategy they follow is below:\n\n1. They formulate feature selection as an optimization problem by augmenting standard empirical risk minimization with zero-one variables associated with each feature representing the absence-presence, and adding a penalty proportional to the number of included features. They relax the discrete variables using a continuous relaxation and provide a simple unbiased estimator for the gradient of the relaxation. After training the relaxation is rounded to a zero-one solution by a simple scheme. \n2. They provide an information theoretic motivation for their formulation of feature selection\n3. They exhibit the performance of their method on a number of synthetic and real data scenarios: (i) linear models with a true underlying sparse parameter, (ii) binary classification with a small number of true determining features, (iii) regression performance post-feature selection with synthetic non-linear models (with a few determining features) and two real datasets. They also use the method for a classification problem with RNA-seq data on T-cells and a survival analysis based on a breast-cancer dataset called METABRIC. \n\nDespite my recommendation, there are a number of things that I like about the paper that I list below, along with directions where I believe the article can be improved. \n1. At a certain abstraction, the main idea of the paper is to do feature selection at the same time as model fitting (as the LASSO for e.g. does) while ignoring constraints of convexity raised in the optimization, and simply using stochastic gradient with a reasonable unbiased estimate of the gradient. This is a reasonable idea, particularly if under some reasonable assumptions, the non-convex formulation that is obtained is expected to be computationally 'benign'. \n2. In a number of the experiments, and particularly 6.1 (sparse linear model) 6.2 (noisy XOR classification) I suspect the non-convex formulation is what is providing a lot of the improvement. This has been observed empirically in a number of other settings, for e.g. in matrix completion/factorization problems. Verifying this hypothesis in a simple, synthetic (and therefore controlled) dataset would be a good contribution for a future version.   \n3. The authors have done a fairly good job of validating the method in a number of different settings, even if some of the presentation of their results can possibly be somewhat improved. For e.g. the median rank is better shown with box plots (as in the Chen et al 2018 paper cited by the authors).\n4. There are a number of relaxations of discrete variables used in optimization and theoretical computer science literature. For instance, the approach of the authors is reminiscent to 'mean field' methods, or standard linear programming relaxation of combinatorial optimization problems (i.e. the first level of the Sherali-Adams LP hierarchy). On the other hand, naive versions of this are not likely to work well on (say) sparse linear regression. The current methods do which suggests that the continuous relaxation is useful. \n\nAt an expository level, I also think the paper could do with quite a bit of improvement:\n1. The introduction is sparse and hurried, and without providing sufficient motivation and intuition for the contributions of the article. \n2. In 6.4, 6.5, the introduction about RNA-seq or Cox models can be removed and relevant work cited instead. \n3.  Organizing the experiments as real data, and synthetic data might be semantically better, though that would necessitate splitting Table 1. I am also unclear on why the authors show  performance in Tables 1, 2 independent of the number of features selected, while for the experiment on RNA-seq data the full accuracy/#features tradeoff is given. The sparse explanation about using the Optuna paper is certainly not enough. \n\nMinor comments not related to decision:\n1. The value for \\alpha_N in synthetic sparse linear model experiment of 6.1 likely has an extraneous \\sqrt \\log k \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The authors propose a feature selection method for high-dimensional datasets that attempts to fit a model while selecting relevant features. \nThe strategy they follow is below:\n\n1. They formulate feature selection as an optimization problem by augmenting standard empirical risk minimization with zero-one variables associated with each feature representing the absence-presence, and adding a penalty proportional to the number of included features. They relax the discrete variables using a continuous relaxation and provide a simple unbiased estimator for the gradient of the relaxation. After training the relaxation is rounded to a zero-one solution by a simple scheme. \n2. They provide an information theoretic motivation for their formulation of feature selection\n3. They exhibit the performance of their method on a number of synthetic and real data scenarios: (i) linear models with a true underlying sparse parameter, (ii) binary classification with a small number of true determining features, (iii) regression performance post-feature selection with synthetic non-linear models (with a few determining features) and two real datasets. They also use the method for a classification problem with RNA-seq data on T-cells and a survival analysis based on a breast-cancer dataset called METABRIC. \n\nDespite my recommendation, there are a number of things that I like about the paper that I list below, along with directions where I believe the article can be improved. \n1. At a certain abstraction, the main idea of the paper is to do feature selection at the same time as model fitting (as the LASSO for e.g. does) while ignoring constraints of convexity raised in the optimization, and simply using stochastic gradient with a reasonable unbiased estimate of the gradient. This is a reasonable idea, particularly if under some reasonable assumptions, the non-convex formulation that is obtained is expected to be computationally 'benign'. \n2. In a number of the experiments, and particularly 6.1 (sparse linear model) 6.2 (noisy XOR classification) I suspect the non-convex formulation is what is providing a lot of the improvement. This has been observed empirically in a number of other settings, for e.g. in matrix completion/factorization problems. Verifying this hypothesis in a simple, synthetic (and therefore controlled) dataset would be a good contribution for a future version.   \n3. The authors have done a fairly good job of validating the method in a number of different settings, even if some of the presentation of their results can possibly be somewhat improved. For e.g. the median rank is better shown with box plots (as in the Chen et al 2018 paper cited by the authors).\n4. There are a number of relaxations of discrete variables used in optimization and theoretical computer science literature. For instance, the approach of the authors is reminiscent to 'mean field' methods, or standard linear programming relaxation of combinatorial optimization problems (i.e. the first level of the Sherali-Adams LP hierarchy). On the other hand, naive versions of this are not likely to work well on (say) sparse linear regression. The current methods do which suggests that the continuous relaxation is useful. \n\nAt an expository level, I also think the paper could do with quite a bit of improvement:\n1. The introduction is sparse and hurried, and without providing sufficient motivation and intuition for the contributions of the article. \n2. In 6.4, 6.5, the introduction about RNA-seq or Cox models can be removed and relevant work cited instead. \n3.  Organizing the experiments as real data, and synthetic data might be semantically better, though that would necessitate splitting Table 1. I am also unclear on why the authors show  performance in Tables 1, 2 independent of the number of features selected, while for the experiment on RNA-seq data the full accuracy/#features tradeoff is given. The sparse explanation about using the Optuna paper is certainly not enough. \n\nMinor comments not related to decision:\n1. The value for \\alpha_N in synthetic sparse linear model experiment of 6.1 likely has an extraneous \\sqrt \\log k \n"}, "tcdate": 1572826573959}, {"id": "rye3z0ytcr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1994/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is concerned with embedding a supervised feature selection within a classification setting. \nThe originality is to use an L_0 regularization (counting the number of retained features), besides the classification loss; the authors leverage the ability to include boolean variables in a neural network and to optimize their value using gradient descent through the reparameterization trick.\n\nI am mildly convinced by the paper:\n* Out of the four contributions listed p. 2, STG is the most convincing one; still, the description thereof is not cristal clear: the reparametrization trick is not due to the authors. The discussion (section 5) needs be more detailed, adding the HC details (presently in appendix); could you comment upon the difference between the proposed STG and the Gumbel-Softmax due to Jang et al, cited ?\n* Likewise the authors delve into details regarding the early state of the art, while omitting some key points. For instance, p. 3, the fact that many authors replaced an L_0 penalization with an L_1 one is rooted on the fact that, provided that the optimal L_0 solution is sparse enough, the L_0 and L_1 problems have same solutions. This section can be summarized;\n* the sought sparsity is assumed to be known, which is bold; \n* Assumption 2 is debatable; one would like to find at most the Markov blanket of the label variable. See Markov Blanket Feature Selection for Support Vector Machines, AAAI 08.\n* There are digressions in the paper which make it harder to follow the argumentation (section 6.1); section 6.2 is not at the state of the art; in Guyon et al's Feature Selection Challenge (2003), the Arcene artificial problem involves a XOR with 5 key features, and 15 additional features are functions of the key features.\n\nSuggestion, you might compare with the L_0 inspired regularization setting used for unsupervised feature selection in Agnostic Feature Selection, Doquet et al, 2019.\n\nDetails: check the citation style: use \\citep instead of \\cite.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper is concerned with embedding a supervised feature selection within a classification setting. \nThe originality is to use an L_0 regularization (counting the number of retained features), besides the classification loss; the authors leverage the ability to include boolean variables in a neural network and to optimize their value using gradient descent through the reparameterization trick.\n\nI am mildly convinced by the paper:\n* Out of the four contributions listed p. 2, STG is the most convincing one; still, the description thereof is not cristal clear: the reparametrization trick is not due to the authors. The discussion (section 5) needs be more detailed, adding the HC details (presently in appendix); could you comment upon the difference between the proposed STG and the Gumbel-Softmax due to Jang et al, cited ?\n* Likewise the authors delve into details regarding the early state of the art, while omitting some key points. For instance, p. 3, the fact that many authors replaced an L_0 penalization with an L_1 one is rooted on the fact that, provided that the optimal L_0 solution is sparse enough, the L_0 and L_1 problems have same solutions. This section can be summarized;\n* the sought sparsity is assumed to be known, which is bold; \n* Assumption 2 is debatable; one would like to find at most the Markov blanket of the label variable. See Markov Blanket Feature Selection for Support Vector Machines, AAAI 08.\n* There are digressions in the paper which make it harder to follow the argumentation (section 6.1); section 6.2 is not at the state of the art; in Guyon et al's Feature Selection Challenge (2003), the Arcene artificial problem involves a XOR with 5 key features, and 15 additional features are functions of the key features.\n\nSuggestion, you might compare with the L_0 inspired regularization setting used for unsupervised feature selection in Agnostic Feature Selection, Doquet et al, 2019.\n\nDetails: check the citation style: use \\citep instead of \\cite."}, "tcdate": 1572564499819}, {"id": "ryeWRH4TFr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1994/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The author rebuttal sufficiently addresses my concerns, so I am upgrading my score.\n\n***\n\nThe paper considers the problem of embedded feature selection for supervised learning with nonlinear functions. A feature subset is evaluated via the loss function in a \"soft\" manner: a fraction of an individual feature can be \"selected\". Sparsity in the feature selection is enforced via a relaxation of l0 regularization. The resulting objective function is differentiable both in the feature selection and learned function making (simultaneous) gradient-based optimization possible. A variety of experiments in several supervised learning tasks demonstrates that the proposed method has superior performance to other embedded and wrapper methods.\n\nMy decision is to reject, but I'm on the fence regarding this paper. I'm not clearly seeing the motivation for an embedded feature selection method for neural network models: for the datasets considered in the paper, it would seem that training a nonlinear model that used all the features would result in performance at least as good as training the nonlinear model with a prepended STG layer. Perhaps there is evidence that filtering features, e.g., irrelevant features, results in higher accuracy and that the prepended STG layer achieves this accuracy, but that evidence is missing from the paper. Also, there could be downstream computational savings, e.g., at prediction time, if the dimension was very large, but this is not the setting tested in the experiments. I suppose interpretability could be considered motivation, but, even so, isn't there at least one simpler, deterministic approach (described below) that also \"solves\" the problem? Finally, it isn't clear how the method scales with increasing sample size and dimension as all the datasets tested are relatively small in these respects.\n\n***\n\nQuestions and suggestions related to decision:\n\n* The performance values using all features should be included in the experimental results so that the value added by STG can be assessed.\n\n* Why not use the simpler deterministic and differentiable relaxation z = \\sigma(\\mu), where \\sigma() is a \"squashing\" function from the real numbers to [0,1] applied element-by-element to the vector \\mu? What specifically is/are the advantage(s) that the randomness in the definition of z at the bottom of pg. 3 provide over this deterministic alternative?\n\n* Though well-described and methodologically rigorous, the experimental comparison is none-the-less a little disappointing: one dataset for classification and half the datasets for regression are synthetic and low-dimensional. The remaining regression datasets are real but also low-dimensional. The survival analysis dataset is also low-dimensional (as described in the supplementary material). This leaves one real classification dataset which was on the order of 20,000 examples and 2500 features. Why were larger sample-size and dimensionality datasets not tested? These should be readily available. For example, the gisette dataset from the NIPS 2003 feature selection challenge has 5000 features. See \"MISSION: Ultra Large-Scale Feature Selection using Count-Sketches\" by Aghazadeh & Spring et al. (2018) for other high-dimensional datasets. Even a single run for each large dataset would have provided some evidence of scalability.\n\n***\n\nOther minor comments not related to decision:\n\n* \"Concrete Autoencoders for Differentiable Feature Selection and Reconstruction\" by Abid et al. (2019) targets unsupervised feature selection but has enough similarities in the approach that it should be considered related work.\n\n* [Typo?] The unnumbered equation after (5) should not have a sum over d in the second term. Perhaps a sum over k was intended? Also, in this equation, the gradient of the loss wrt/ z samples, average of gradients over z samples times..., does not seem to match what the gradient would be given the algorithmic description in the supplementary material, a gradient of the (sample) average z times...\n\n* The abstract states the paper is proposing a method for high-dimensional feature selection, but all of the experiments have datasets with max. dimensionality 2538.\n\n* Some discussion of how the regularization parameter can be selected by a user of the proposed method would be good to include.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "The author rebuttal sufficiently addresses my concerns, so I am upgrading my score.\n\n***\n\nThe paper considers the problem of embedded feature selection for supervised learning with nonlinear functions. A feature subset is evaluated via the loss function in a \"soft\" manner: a fraction of an individual feature can be \"selected\". Sparsity in the feature selection is enforced via a relaxation of l0 regularization. The resulting objective function is differentiable both in the feature selection and learned function making (simultaneous) gradient-based optimization possible. A variety of experiments in several supervised learning tasks demonstrates that the proposed method has superior performance to other embedded and wrapper methods.\n\nMy decision is to reject, but I'm on the fence regarding this paper. I'm not clearly seeing the motivation for an embedded feature selection method for neural network models: for the datasets considered in the paper, it would seem that training a nonlinear model that used all the features would result in performance at least as good as training the nonlinear model with a prepended STG layer. Perhaps there is evidence that filtering features, e.g., irrelevant features, results in higher accuracy and that the prepended STG layer achieves this accuracy, but that evidence is missing from the paper. Also, there could be downstream computational savings, e.g., at prediction time, if the dimension was very large, but this is not the setting tested in the experiments. I suppose interpretability could be considered motivation, but, even so, isn't there at least one simpler, deterministic approach (described below) that also \"solves\" the problem? Finally, it isn't clear how the method scales with increasing sample size and dimension as all the datasets tested are relatively small in these respects.\n\n***\n\nQuestions and suggestions related to decision:\n\n* The performance values using all features should be included in the experimental results so that the value added by STG can be assessed.\n\n* Why not use the simpler deterministic and differentiable relaxation z = \\sigma(\\mu), where \\sigma() is a \"squashing\" function from the real numbers to [0,1] applied element-by-element to the vector \\mu? What specifically is/are the advantage(s) that the randomness in the definition of z at the bottom of pg. 3 provide over this deterministic alternative?\n\n* Though well-described and methodologically rigorous, the experimental comparison is none-the-less a little disappointing: one dataset for classification and half the datasets for regression are synthetic and low-dimensional. The remaining regression datasets are real but also low-dimensional. The survival analysis dataset is also low-dimensional (as described in the supplementary material). This leaves one real classification dataset which was on the order of 20,000 examples and 2500 features. Why were larger sample-size and dimensionality datasets not tested? These should be readily available. For example, the gisette dataset from the NIPS 2003 feature selection challenge has 5000 features. See \"MISSION: Ultra Large-Scale Feature Selection using Count-Sketches\" by Aghazadeh & Spring et al. (2018) for other high-dimensional datasets. Even a single run for each large dataset would have provided some evidence of scalability.\n\n***\n\nOther minor comments not related to decision:\n\n* \"Concrete Autoencoders for Differentiable Feature Selection and Reconstruction\" by Abid et al. (2019) targets unsupervised feature selection but has enough similarities in the approach that it should be considered related work.\n\n* [Typo?] The unnumbered equation after (5) should not have a sum over d in the second term. Perhaps a sum over k was intended? Also, in this equation, the gradient of the loss wrt/ z samples, average of gradients over z samples times..., does not seem to match what the gradient would be given the algorithmic description in the supplementary material, a gradient of the (sample) average z times...\n\n* The abstract states the paper is proposing a method for high-dimensional feature selection, but all of the experiments have datasets with max. dimensionality 2538.\n\n* Some discussion of how the regularization parameter can be selected by a user of the proposed method would be good to include.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1571796424717}], "openreview_url": "https://openreview.net/forum?id=B1lda1HtvB", "arxiv_id": "1810.04247", "paper_pdf": "papers/B1lda1HtvB.pdf", "paper_pdf_sha256": "f241c78e43d9e65868668ebf948b1bd04eedcd13d7f2e1df7ccdc753d3ac9b9c", "paper_pdf_bytes": 3344409, "paper_pdf_source": "openreview", "code_url": "https://github.com/runopti/stg", "code_repository": "runopti/stg", "code_commit": "8b60e349a70d2f392b13d2505c316cdd8868a067", "code_archive": "repos/B1lda1HtvB.zip", "code_archive_sha256": "84f6db32300a48241aa75982bf36ee191e7966b6388698ff70f1edebd874ceb8", "code_archive_bytes": 2535326, "code_file_count": 17, "code_extensions": {".py": 12, ".ipynb": 3, ".r": 2}, "github_disk_usage_kb": 3889, "github_languages": {"Python": 51920, "R": 3571}, "github_archived": false, "github_pushed_at": "2022-02-18T16:44:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/feature-selection-using-stochastic-gates"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jbGGPSI9aO", "year": 2026, "status": "rejected", "title": "AdaSCALE: Adaptive Scaling for OOD detection", "authors": ["Sudarshan Regmi"], "authorids": ["~Sudarshan_Regmi1"], "authors_source": "OpenReview API", "abstract": "The ability of the deep learning model to recognize when a sample falls outside its learned distribution is critical for safe and reliable deployment. Recent state-of-the-art out-of-distribution (OOD) detection methods leverage activation shaping to improve the separation between in-distribution (ID) and OOD inputs. These approaches resort to sample-specific scaling but apply a static percentile threshold across all samples regardless of their nature. In this work, we propose AdaSCALE, an adaptive scaling procedure that dynamically adjusts the percentile threshold based on a sample's estimated OOD likelihood. This estimation leverages our key observation that OOD samples exhibit significantly more pronounced activation shifts at high-magnitude activations under minor perturbation compared to ID samples. AdaSCALE enables stronger scaling for likely ID samples and weaker scaling for likely OOD samples, creating highly separable energy scores. Our approach achieves state-of-the-art OOD detection performance, outperforming the latest rival OptFS by **14.94%** in near-OOD and **21.67%** in far-OOD datasets in average FPR@95 metric in the ImageNet-1k benchmark across eight diverse architectures.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "tfTGWxa00y", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_Ev7Q"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces a post-hoc OoD detection method, AdaSCALE. The basis of the proposal is that OoD inputs exhibit higher activation instability for minor perturbations compared to ID samples. AdaSCALE proposes an adaptive scaling mechanism and computes a per sample OODness score Q'.", "review_text": "The paper introduces a post-hoc OoD detection method, AdaSCALE. The basis of the proposal is that OoD inputs exhibit higher activation instability for minor perturbations compared to ID samples. AdaSCALE proposes an adaptive scaling mechanism and computes a per sample OODness score Q'.", "strengths": "- the phenomenon according to which ID inputs activate stable features while OoD samples cause unstable high-magnitude activations under small perturbations is well highlighted and illustrated (Figure 2)\n\n- the score mapping strategy is straightforward and effective. The authors rely on an empirical CDF built using a small ID validation set. This results in a simple and non-parametric way of translating the scores into percentile ranges.\n\n- extensive comparison with other relevant approaches (e.g. ReAct, ASH, SCALE, LTS, OptFS) and strong overall performance, including some great results (e.g. FPR@95 of 61.17 on imagenet-1k vs 71.91 for OptFS)", "weaknesses": "- despite the generalization claims, the presented method seems to be systematically over-tuned for each test setup. The authors state that a fixed set of hyperparams may be used, and that by tuning only $p_{max}$ \"near-optimal performance can be achieved\", see Table 5. However, the main results (Tables 2 and 3) are obtained following a humongous level of finetuning (and implicitly computational cost), apparent in Tables 28 and 29. Presenting Tables 2/3 and claiming generalization is grossly misleading. I see two ethical ways out : either the authors claim a good generalization across all setups and finetuning goes to appendix, or the generalization claim is dropped and Table 5 goes to Appendix.\n\n- the OODNess metric is artificially complex; the introduction of the correction term + 2 additional parameters is justified by the \"high variance\" of Q alone. However, Q has a FPR@95 of 59.43 vs 58.97 for the full metric.In my opinion, this shows that Q works quite well, and that the additional complexity and cost implied by the use of Q' are hard to justify.\n\n- I also am very doubtful about the positive impact of the proposed perturbation mechanism; the \"trivial\" method performs a backward pass and is 2.91x heavier, while the random perturbation is only 1.56x slower. The benefit for the doubled cost is minimal (Table 22). This trade-off seems poor and is poorly described; the work should probably propose the random selection by default and introduce the other mechanism as a costlier alternative. Overall, the 3x slowdown (compared to SCALE) is not even once mentioned in the abstract / intro / conclusion; in my opinion this is a significant omission and a failure in transparency about the method's trade-offs.", "questions": "Please see the three points raised in the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a post-hoc OoD detection method, AdaSCALE. The basis of the proposal is that OoD inputs exhibit higher activation instability for minor perturbations compared to ID samples. AdaSCALE proposes an adaptive scaling mechanism and computes a per sample OODness score Q'.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- the phenomenon according to which ID inputs activate stable features while OoD samples cause unstable high-magnitude activations under small perturbations is well highlighted and illustrated (Figure 2)\n\n- the score mapping strategy is straightforward and effective. The authors rely on an empirical CDF built using a small ID validation set. This results in a simple and non-parametric way of translating the scores into percentile ranges.\n\n- extensive comparison with other relevant approaches (e.g. ReAct, ASH, SCALE, LTS, OptFS) and strong overall performance, including some great results (e.g. FPR@95 of 61.17 on imagenet-1k vs 71.91 for OptFS)", "weaknesses": "- despite the generalization claims, the presented method seems to be systematically over-tuned for each test setup. The authors state that a fixed set of hyperparams may be used, and that by tuning only $p_{max}$ \"near-optimal performance can be achieved\", see Table 5. However, the main results (Tables 2 and 3) are obtained following a humongous level of finetuning (and implicitly computational cost), apparent in Tables 28 and 29. Presenting Tables 2/3 and claiming generalization is grossly misleading. I see two ethical ways out : either the authors claim a good generalization across all setups and finetuning goes to appendix, or the generalization claim is dropped and Table 5 goes to Appendix.\n\n- the OODNess metric is artificially complex; the introduction of the correction term + 2 additional parameters is justified by the \"high variance\" of Q alone. However, Q has a FPR@95 of 59.43 vs 58.97 for the full metric.In my opinion, this shows that Q works quite well, and that the additional complexity and cost implied by the use of Q' are hard to justify.\n\n- I also am very doubtful about the positive impact of the proposed perturbation mechanism; the \"trivial\" method performs a backward pass and is 2.91x heavier, while the random perturbation is only 1.56x slower. The benefit for the doubled cost is minimal (Table 22). This trade-off seems poor and is poorly described; the work should probably propose the random selection by default and introduce the other mechanism as a costlier alternative. Overall, the 3x slowdown (compared to SCALE) is not even once mentioned in the abstract / intro / conclusion; in my opinion this is a significant omission and a failure in transparency about the method's trade-offs.", "questions": "Please see the three points raised in the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "everything OK", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762729562181}, {"id": "aIZIvAYl2h", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_UieH"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes AdaSCALE, a post-hoc method for out-of-distribution (OOD) detection that adaptively adjusts the activation scaling percentile based on the estimated OOD likelihood of each sample. The method is motivated by the observation that OOD inputs exhibit stronger top-k activation shifts under small perturbations than ID inputs. The authors design a mechanism to estimate sample OODness via activation differences and dynamically modulate the scaling strength. Extensive experiments on ImageNet-1k, CIFAR-10/100, and various architectures demonstrate consistent gains over prior post-hoc methods (e.g., SCALE, LTS, OptFS), achieving state-of-the-art performance.", "review_text": "This paper proposes AdaSCALE, a post-hoc method for out-of-distribution (OOD) detection that adaptively adjusts the activation scaling percentile based on the estimated OOD likelihood of each sample. The method is motivated by the observation that OOD inputs exhibit stronger top-k activation shifts under small perturbations than ID inputs. The authors design a mechanism to estimate sample OODness via activation differences and dynamically modulate the scaling strength. Extensive experiments on ImageNet-1k, CIFAR-10/100, and various architectures demonstrate consistent gains over prior post-hoc methods (e.g., SCALE, LTS, OptFS), achieving state-of-the-art performance.", "strengths": "- Motivation and observation are clear and intuitive. The link between activation stability under perturbation and OOD likelihood is conceptually appealing.\n\n- Strong empirical performance across datasets and architectures shows that the adaptive scaling strategy generalizes better than fixed-threshold methods.\n\n- Method simplicity.", "weaknesses": "- Incremental novelty. While the adaptive percentile idea is reasonable, it extends existing activation scaling works (e.g., ASH, SCALE, LTS) rather than introducing a fundamentally new principle.\n\n- Limited theoretical grounding. The paper claims that activation shift reflects OODness, but lacks a formal analysis or statistical justification. The perturbation mechanism and threshold mapping are mostly heuristic.  It may disentangle this effect from other confounders (e.g., gradient norm, layer saturation, or input magnitude)\n\n- Clarity issues. Some equations and algorithmic steps (e.g., computation of OODness score and eCDF calibration) are dense and could benefit from clearer notation or pseudo-code.\n\n- Minor reproducibility concern. The reliance on gradient-based perturbations may introduce stochasticity; details of implementation choices (e.g., ε magnitude, percentile selection ranges) could be more transparent.", "questions": "- How sensitive is AdaSCALE to the choice of perturbation direction and strength (ε)? Could adversarial directions yield different results?\n\n- Maybe the authors could provide more intuition or quantitative evidence that the activation shift magnitude correlates with the sample’s epistemic uncertainty rather than simply gradient magnitude?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes AdaSCALE, a post-hoc method for out-of-distribution (OOD) detection that adaptively adjusts the activation scaling percentile based on the estimated OOD likelihood of each sample. The method is motivated by the observation that OOD inputs exhibit stronger top-k activation shifts under small perturbations than ID inputs. The authors design a mechanism to estimate sample OODness via activation differences and dynamically modulate the scaling strength. Extensive experiments on ImageNet-1k, CIFAR-10/100, and various architectures demonstrate consistent gains over prior post-hoc methods (e.g., SCALE, LTS, OptFS), achieving state-of-the-art performance.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Motivation and observation are clear and intuitive. The link between activation stability under perturbation and OOD likelihood is conceptually appealing.\n\n- Strong empirical performance across datasets and architectures shows that the adaptive scaling strategy generalizes better than fixed-threshold methods.\n\n- Method simplicity.", "weaknesses": "- Incremental novelty. While the adaptive percentile idea is reasonable, it extends existing activation scaling works (e.g., ASH, SCALE, LTS) rather than introducing a fundamentally new principle.\n\n- Limited theoretical grounding. The paper claims that activation shift reflects OODness, but lacks a formal analysis or statistical justification. The perturbation mechanism and threshold mapping are mostly heuristic.  It may disentangle this effect from other confounders (e.g., gradient norm, layer saturation, or input magnitude)\n\n- Clarity issues. Some equations and algorithmic steps (e.g., computation of OODness score and eCDF calibration) are dense and could benefit from clearer notation or pseudo-code.\n\n- Minor reproducibility concern. The reliance on gradient-based perturbations may introduce stochasticity; details of implementation choices (e.g., ε magnitude, percentile selection ranges) could be more transparent.", "questions": "- How sensitive is AdaSCALE to the choice of perturbation direction and strength (ε)? Could adversarial directions yield different results?\n\n- Maybe the authors could provide more intuition or quantitative evidence that the activation shift magnitude correlates with the sample’s epistemic uncertainty rather than simply gradient magnitude?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762439559678}, {"id": "mTCCOE6oCf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_GRZb"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper addresses out-of-distribution (OOD) detection in deep networks and identifies a limitation of recent post-hoc methods: they use sample-specific activation scaling with a fixed percentile threshold, which is suboptimal for separating in-distribution (ID) vs OOD samples. As models scale up, a static threshold can’t account for the varying “OOD-ness” of each sample. \nAdaSCALE is proposed as an adaptive scaling procedure that adjusts the scaling percentile per sample based on an estimated OOD likelihood. The core insight is that OOD inputs exhibit larger shifts in their top activations under small perturbations compared to ID inputs.\nBy quantifying this activation shift (via a gradient-based input perturbation and measuring changes in high-magnitude activations), the method assigns a higher percentile (stronger scaling) to inputs likely ID and a lower percentile (weaker scaling) to those likely OOD. This yields more separable energy scores for detection, effectively reducing OOD confidence for OOD samples while preserving ID confidence.", "review_text": "The paper addresses out-of-distribution (OOD) detection in deep networks and identifies a limitation of recent post-hoc methods: they use sample-specific activation scaling with a fixed percentile threshold, which is suboptimal for separating in-distribution (ID) vs OOD samples. As models scale up, a static threshold can’t account for the varying “OOD-ness” of each sample. \nAdaSCALE is proposed as an adaptive scaling procedure that adjusts the scaling percentile per sample based on an estimated OOD likelihood. The core insight is that OOD inputs exhibit larger shifts in their top activations under small perturbations compared to ID inputs.\nBy quantifying this activation shift (via a gradient-based input perturbation and measuring changes in high-magnitude activations), the method assigns a higher percentile (stronger scaling) to inputs likely ID and a lower percentile (weaker scaling) to those likely OOD. This yields more separable energy scores for detection, effectively reducing OOD confidence for OOD samples while preserving ID confidence.", "strengths": "- The method demonstrates substantial performance gains. AdaSCALE achieves state-of-the-art OOD detection results on challenging benchmarks (ImageNet-1k), significantly outperforming prior post-hoc methods. For example, it dramatically lowers false positive rates (FPR@95) compared to strong baselines (e.g., vs. OptFS and SCALE) in both near-OOD and far-OOD settings.\n- The paper shows robust generalization on several different network architectures (ResNets, EfficientNet-V2, ViT, etc.), where AdaSCALE outperforms or matches previous methods on each, addressing a known shortcoming of some prior methods that were tied to specific models.\n- The paper is well-structured and clear. It provides a detailed algorithm, conceptual diagrams, and ablations. The authors situate AdaSCALE in context by comparing to a wide range of baselines (MSP, ODIN, ReAct, ASH, SCALE, LTS, OptFS, etc.), and perform comprehensive experiments (including metrics like AUROC/FPR, near-vs-far OOD, and even a “full-spectrum” test incorporating covariate shifts. The writing is generally easy to follow, with a solid explanation of the method and the intuition behind it.", "weaknesses": "- While the adaptive scaling idea is effective, the contribution can be seen as a relatively incremental extension of existing methods. AdaSCALE builds directly on the activation scaling paradigm introduced by ASH/SCALE/LTS. The key insight (that OOD causes unstable high activations under perturbation) extends the known observation from ReAct that OOD samples have high activation magnitudes. Thus, the method’s novelty, though useful, feels a bit additive: it combines perturbation-based OOD scoring with the existing scaling framework.\n- Certain design elements in AdaSCALE appear somewhat heuristic and might lack theoretical justification. For instance, the OOD likelihood score is a composite of the top-k activation shift and a “correction” term for ID bias, weighted by a parameter $\\lambda$. This was introduced to counter the high variance in OOD shift metrics, which indicates that the raw signal can be noisy or overestimate OOD likelihood without tuning. The need to hand-craft this combination (including hyperparameters $k_1$, $k_2$, $\\lambda$) suggests the solution required empirically balancing factors, rather than deriving a unified criterion. The current design feels overcomplicated and may benefit from simplifications focusing on the most essential theoretical insights.\n- AdaSCALE introduces additional hyperparameters (e.g., percentile range $p_{min}$/$p_{max}$, top-k sizes, perturbation size, etc.) and a more complex inference procedure. This raises two concerns:\n  - The paper mentions using an automatic search (via OpenOOD) to set these, but it’s unclear how sensitive the results are to these settings. In Table 7, it is shown that AdaSCALE is somewhat insensitive to single-hyperparameter variations with ResNet-50 on ImageNet-1k. However, there is no study involving multiple-hyperparameter variations (e.g., two hyperparameters varying at the same time; considering that AdaSCALE has a lot of hyperparameters) or study on other architectures and datasets.\n  - Moreover, it is unclear whether the automatic search (via OpenOOD) utilizes an OOD validation set. If it does, then the current evaluation setting may not be fair to methods that do not require any OOD validation sets. Another related concern is how sensitive AdaSCALE is to any mismatch between the OOD validation set and the actual OOD test set.\n- Unlike some post-hoc methods that require no extra data, AdaSCALE does rely on a small set of ID validation samples to compute the empirical CDF for its adaptive threshold. The paper emphasizes this is a minimal requirement (only 10 samples used), but it seems that Table 9 only shows the result from a single experiment (per setting). The impact of using only a handful of samples to estimate the distribution of the OOD score isn’t deeply explored. One might wonder if this calibration could be unstable or biased if those few samples aren’t perfectly representative. This is a minor weakness given how small the requirement is, but it’s worth noting as a practical consideration.\n- There is a typo in the formula for the proposed adaptive percentile threshold: a right parenthesis is missing.", "questions": "- The adaptive scaling relies on the heuristic that OOD samples’ top activations are unstable under small perturbations. Are there types of OOD inputs or shifts where this heuristic might fail or be less effective? For example, if an OOD sample is very near the ID distribution (or if the model’s features are insensitive to the chosen perturbation), would AdaSCALE risk mis-classifying it as ID? Conversely, could certain ID samples that are borderline or noisy exhibit large activation shifts and be mistaken for OOD?\n- OptFS was designed for cross-architecture generalization using a piecewise constant scaling function. AdaSCALE empirically outperforms OptFS across architectures, but could the authors elaborate on *why* adaptivity gives a *significant* edge here? From Figure 2, the signal doesn’t appear to be very clear (the standard deviation regions largely overlap between ID and OOD).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses out-of-distribution (OOD) detection in deep networks and identifies a limitation of recent post-hoc methods: they use sample-specific activation scaling with a fixed percentile threshold, which is suboptimal for separating in-distribution (ID) vs OOD samples. As models scale up, a static threshold can’t account for the varying “OOD-ness” of each sample. \nAdaSCALE is proposed as an adaptive scaling procedure that adjusts the scaling percentile per sample based on an estimated OOD likelihood. The core insight is that OOD inputs exhibit larger shifts in their top activations under small perturbations compared to ID inputs.\nBy quantifying this activation shift (via a gradient-based input perturbation and measuring changes in high-magnitude activations), the method assigns a higher percentile (stronger scaling) to inputs likely ID and a lower percentile (weaker scaling) to those likely OOD. This yields more separable energy scores for detection, effectively reducing OOD confidence for OOD samples while preserving ID confidence.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The method demonstrates substantial performance gains. AdaSCALE achieves state-of-the-art OOD detection results on challenging benchmarks (ImageNet-1k), significantly outperforming prior post-hoc methods. For example, it dramatically lowers false positive rates (FPR@95) compared to strong baselines (e.g., vs. OptFS and SCALE) in both near-OOD and far-OOD settings.\n- The paper shows robust generalization on several different network architectures (ResNets, EfficientNet-V2, ViT, etc.), where AdaSCALE outperforms or matches previous methods on each, addressing a known shortcoming of some prior methods that were tied to specific models.\n- The paper is well-structured and clear. It provides a detailed algorithm, conceptual diagrams, and ablations. The authors situate AdaSCALE in context by comparing to a wide range of baselines (MSP, ODIN, ReAct, ASH, SCALE, LTS, OptFS, etc.), and perform comprehensive experiments (including metrics like AUROC/FPR, near-vs-far OOD, and even a “full-spectrum” test incorporating covariate shifts. The writing is generally easy to follow, with a solid explanation of the method and the intuition behind it.", "weaknesses": "- While the adaptive scaling idea is effective, the contribution can be seen as a relatively incremental extension of existing methods. AdaSCALE builds directly on the activation scaling paradigm introduced by ASH/SCALE/LTS. The key insight (that OOD causes unstable high activations under perturbation) extends the known observation from ReAct that OOD samples have high activation magnitudes. Thus, the method’s novelty, though useful, feels a bit additive: it combines perturbation-based OOD scoring with the existing scaling framework.\n- Certain design elements in AdaSCALE appear somewhat heuristic and might lack theoretical justification. For instance, the OOD likelihood score is a composite of the top-k activation shift and a “correction” term for ID bias, weighted by a parameter $\\lambda$. This was introduced to counter the high variance in OOD shift metrics, which indicates that the raw signal can be noisy or overestimate OOD likelihood without tuning. The need to hand-craft this combination (including hyperparameters $k_1$, $k_2$, $\\lambda$) suggests the solution required empirically balancing factors, rather than deriving a unified criterion. The current design feels overcomplicated and may benefit from simplifications focusing on the most essential theoretical insights.\n- AdaSCALE introduces additional hyperparameters (e.g., percentile range $p_{min}$/$p_{max}$, top-k sizes, perturbation size, etc.) and a more complex inference procedure. This raises two concerns:\n  - The paper mentions using an automatic search (via OpenOOD) to set these, but it’s unclear how sensitive the results are to these settings. In Table 7, it is shown that AdaSCALE is somewhat insensitive to single-hyperparameter variations with ResNet-50 on ImageNet-1k. However, there is no study involving multiple-hyperparameter variations (e.g., two hyperparameters varying at the same time; considering that AdaSCALE has a lot of hyperparameters) or study on other architectures and datasets.\n  - Moreover, it is unclear whether the automatic search (via OpenOOD) utilizes an OOD validation set. If it does, then the current evaluation setting may not be fair to methods that do not require any OOD validation sets. Another related concern is how sensitive AdaSCALE is to any mismatch between the OOD validation set and the actual OOD test set.\n- Unlike some post-hoc methods that require no extra data, AdaSCALE does rely on a small set of ID validation samples to compute the empirical CDF for its adaptive threshold. The paper emphasizes this is a minimal requirement (only 10 samples used), but it seems that Table 9 only shows the result from a single experiment (per setting). The impact of using only a handful of samples to estimate the distribution of the OOD score isn’t deeply explored. One might wonder if this calibration could be unstable or biased if those few samples aren’t perfectly representative. This is a minor weakness given how small the requirement is, but it’s worth noting as a practical consideration.\n- There is a typo in the formula for the proposed adaptive percentile threshold: a right parenthesis is missing.", "questions": "- The adaptive scaling relies on the heuristic that OOD samples’ top activations are unstable under small perturbations. Are there types of OOD inputs or shifts where this heuristic might fail or be less effective? For example, if an OOD sample is very near the ID distribution (or if the model’s features are insensitive to the chosen perturbation), would AdaSCALE risk mis-classifying it as ID? Conversely, could certain ID samples that are borderline or noisy exhibit large activation shifts and be mistaken for OOD?\n- OptFS was designed for cross-architecture generalization using a piecewise constant scaling function. AdaSCALE empirically outperforms OptFS across architectures, but could the authors elaborate on *why* adaptivity gives a *significant* edge here? From Figure 2, the signal doesn’t appear to be very clear (the standard deviation regions largely overlap between ID and OOD).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761922947194}, {"id": "GMsi0bChNZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_Vxks"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes AdaSCALE, a post-hoc OOD detection method that adaptively adjusts activation scaling based on the estimated OODness of each test sample. The method is built on an empirical observation: OOD samples exhibit larger shifts in top-k activations under small input perturbations, while ID samples remain stable. Leveraging this phenomenon, the authors compute a perturbation-induced activation shift statistic, convert it into a normalized OODness score using an empirical CDF with only a few ID samples, and dynamically determine the percentile threshold for scaling. Extensive experiments across CIFAR and ImageNet-1k, eight architectures, FSOOD settings, corrupted inputs, and adversarially trained networks demonstrate consistent SOTA performance and strong generalization. The method is parameter-efficient, requires minimal ID samples, and preserves ID accuracy.", "review_text": "This paper proposes AdaSCALE, a post-hoc OOD detection method that adaptively adjusts activation scaling based on the estimated OODness of each test sample. The method is built on an empirical observation: OOD samples exhibit larger shifts in top-k activations under small input perturbations, while ID samples remain stable. Leveraging this phenomenon, the authors compute a perturbation-induced activation shift statistic, convert it into a normalized OODness score using an empirical CDF with only a few ID samples, and dynamically determine the percentile threshold for scaling. Extensive experiments across CIFAR and ImageNet-1k, eight architectures, FSOOD settings, corrupted inputs, and adversarially trained networks demonstrate consistent SOTA performance and strong generalization. The method is parameter-efficient, requires minimal ID samples, and preserves ID accuracy.", "strengths": "- The paper provides a compelling mechanistic explanation for why activation scaling works for OOD detection: semantic activations are stable for ID samples but unstable for OOD samples under perturbations. This moves scaling-based OOD detection from heuristic to a more principled footing. I consider this a significant conceptual advancement.\n\n-Prior scaling methods use a static percentile threshold; adapting it per-sample based on estimated OODness is a meaningful and novel extension.\n\n- Demonstrates consistent SOTA performance across: CIFAR & ImageNet-1k and 8 architectures\n\n- only need a very small number of ID samples (as few as 10) for calibration and preserves accuracy,.\n\n- No model retraining or gradients beyond attribution; easy to integrate into existing systems.", "weaknesses": "- Although trivial/random perturbation works, the selection of perturbation magnitude and pixel selection lacks theoretical grounding, and different perturbation strategies may influence results.\n\n- Requires an extra forward pass + top-k operations, resulting in ~2–4× latency vs. fixed-scaling baselines, which may be limiting in latency-critical systems.\n\n- The work discovers an important phenomenon, but lacks a deeper theoretical unification of scaling and perturbation in relation to existing methods. Such a theory would substantially enhance the impact.", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes AdaSCALE, a post-hoc OOD detection method that adaptively adjusts activation scaling based on the estimated OODness of each test sample. The method is built on an empirical observation: OOD samples exhibit larger shifts in top-k activations under small input perturbations, while ID samples remain stable. Leveraging this phenomenon, the authors compute a perturbation-induced activation shift statistic, convert it into a normalized OODness score using an empirical CDF with only a few ID samples, and dynamically determine the percentile threshold for scaling. Extensive experiments across CIFAR and ImageNet-1k, eight architectures, FSOOD settings, corrupted inputs, and adversarially trained networks demonstrate consistent SOTA performance and strong generalization. The method is parameter-efficient, requires minimal ID samples, and preserves ID accuracy.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper provides a compelling mechanistic explanation for why activation scaling works for OOD detection: semantic activations are stable for ID samples but unstable for OOD samples under perturbations. This moves scaling-based OOD detection from heuristic to a more principled footing. I consider this a significant conceptual advancement.\n\n-Prior scaling methods use a static percentile threshold; adapting it per-sample based on estimated OODness is a meaningful and novel extension.\n\n- Demonstrates consistent SOTA performance across: CIFAR & ImageNet-1k and 8 architectures\n\n- only need a very small number of ID samples (as few as 10) for calibration and preserves accuracy,.\n\n- No model retraining or gradients beyond attribution; easy to integrate into existing systems.", "weaknesses": "- Although trivial/random perturbation works, the selection of perturbation magnitude and pixel selection lacks theoretical grounding, and different perturbation strategies may influence results.\n\n- Requires an extra forward pass + top-k operations, resulting in ~2–4× latency vs. fixed-scaling baselines, which may be limiting in latency-critical systems.\n\n- The work discovers an important phenomenon, but lacks a deeper theoretical unification of scaling and perturbation in relation to existing methods. Such a theory would substantially enhance the impact.", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761825821780}, {"id": "1YfCqKxA3A", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_yGZh"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper presents a novel adaptive scaling procedure for OOD detection that dynamically adjusts the detection threshold based on the likelihood of OODness for individual samples. Through extensive experiments, the superiority of the method against several baselines is established.", "review_text": "This paper presents a novel adaptive scaling procedure for OOD detection that dynamically adjusts the detection threshold based on the likelihood of OODness for individual samples. Through extensive experiments, the superiority of the method against several baselines is established.", "strengths": "The paper is generally easy to follow, and most of the intuition for the proposed methodology and the different components of the algorithm are empirically justified.\n\nThe authors provide good figures to better convey the empirical results and justifications\n\nThe authors provide a good set of experiments and ablation studies that examine different properties of the algorithm and different sensitivities.", "weaknesses": "Despite the easy flow of the text, some parts are abstract and require some more detailed domain knowledge, which could have been better established. This mainly pertains to Fig 1, the introduction, and sec 4.1. \n\nSome parts of the paper seem to be stating contradicting requirements and observations. This could also be my lack of familiarity with the different terminology and their differences:\n- In the preliminaries, it is mentioned that higher scores are for OOD samples. Yet the paragraph at 84 talks about giving lower scores to OODs. Same comment regarding 4.1.\n- 284: \"Fig 3 illustrates Q'/Q' > Q/Q\", whereas the figure shows the opposite of this. Also, in Fig 3 caption.\n\nThere are some baselines that are missing. Based on my own experiments, it is critical that the authors **must** include a comparison with LINe [1] for my final decision. Other baselines can be found in [2]\n\nAlthough I understand the regime of the experimental setup and its comparison with similar post-hoc methods, a comparison with methods that assume access to OOD data during training could also be useful. It is not necessary for the methodology of this paper to beat such methods, but a comparison with such work could be useful. \n\n\n[1] Ahn YH, Park GM, Kim ST. Line: Out-of-distribution detection by leveraging important neurons. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023 Jun 17 (pp. 19852-19862). IEEE.\n\n[2] https://github.com/Jingkang50/OpenOOD", "questions": "If the absolute perturbation sum over activations is indeed such a good measure of the likelihood of OODness, why not use this metric itself as the scoring function? Why do you first utilize this to acquire OODness and then utilize that information for another scoring function?  \n\nAlso see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel adaptive scaling procedure for OOD detection that dynamically adjusts the detection threshold based on the likelihood of OODness for individual samples. Through extensive experiments, the superiority of the method against several baselines is established.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "The paper is generally easy to follow, and most of the intuition for the proposed methodology and the different components of the algorithm are empirically justified.\n\nThe authors provide good figures to better convey the empirical results and justifications\n\nThe authors provide a good set of experiments and ablation studies that examine different properties of the algorithm and different sensitivities.", "weaknesses": "Despite the easy flow of the text, some parts are abstract and require some more detailed domain knowledge, which could have been better established. This mainly pertains to Fig 1, the introduction, and sec 4.1. \n\nSome parts of the paper seem to be stating contradicting requirements and observations. This could also be my lack of familiarity with the different terminology and their differences:\n- In the preliminaries, it is mentioned that higher scores are for OOD samples. Yet the paragraph at 84 talks about giving lower scores to OODs. Same comment regarding 4.1.\n- 284: \"Fig 3 illustrates Q'/Q' > Q/Q\", whereas the figure shows the opposite of this. Also, in Fig 3 caption.\n\nThere are some baselines that are missing. Based on my own experiments, it is critical that the authors **must** include a comparison with LINe [1] for my final decision. Other baselines can be found in [2]\n\nAlthough I understand the regime of the experimental setup and its comparison with similar post-hoc methods, a comparison with methods that assume access to OOD data during training could also be useful. It is not necessary for the methodology of this paper to beat such methods, but a comparison with such work could be useful. \n\n\n[1] Ahn YH, Park GM, Kim ST. Line: Out-of-distribution detection by leveraging important neurons. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023 Jun 17 (pp. 19852-19862). IEEE.\n\n[2] https://github.com/Jingkang50/OpenOOD", "questions": "If the absolute perturbation sum over activations is indeed such a good measure of the likelihood of OODness, why not use this metric itself as the scoring function? Why do you first utilize this to acquire OODness and then utilize that information for another scoring function?  \n\nAlso see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761609326047}, {"id": "dRdDWXrY6n", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15445/Reviewer_upif"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes AdaSCALE, an adaptive scaling procedure\nthat dynamically adjusts the percentile threshold based on a sample’s estimated\nOODness. This estimation leverages a key observation: OOD samples exhibit\nsignificantly more pronounced activation shifts at high-magnitude activations under\nminor perturbation compared to ID samples. AdaSCALE achieves state-of-the-art OOD detection\nperformance, outperforming the latest rival OptFS by 14.94% in near-OOD and\n21.67% in far-OOD datasets in average FPR@95 metric on the ImageNet-1k\nbenchmark across eight diverse architectures.", "review_text": "This paper proposes AdaSCALE, an adaptive scaling procedure\nthat dynamically adjusts the percentile threshold based on a sample’s estimated\nOODness. This estimation leverages a key observation: OOD samples exhibit\nsignificantly more pronounced activation shifts at high-magnitude activations under\nminor perturbation compared to ID samples. AdaSCALE achieves state-of-the-art OOD detection\nperformance, outperforming the latest rival OptFS by 14.94% in near-OOD and\n21.67% in far-OOD datasets in average FPR@95 metric on the ImageNet-1k\nbenchmark across eight diverse architectures.", "strengths": "1.\tClear motivation and novel observation: OOD samples exhibit greater sensitivity of high-magnitude activations to minor perturbations, providing an intuitive and effective signal for designing an adaptive mechanism.\n2.\tOverall, AdaSCALE  is  a simple yet effective method.\n3.\tThe experiments  quite extensive with significant results: across 10 architectures and 3 datasets by tuning mere one hyperparameter for a given setup.", "weaknesses": "1. AdaSCALE shares a similar design  with ATS (Krumpl et al., 2024), as both methods adaptively adjust scaling parameters during post-processing to enhance ID and OOD separability. It would be better to further clarify the differences in design rationale and implementation mechanisms between the two approaches and provide a more in-depth comparative analysis.\n\n2. The experiments employ gradient-based perturbations, which significantly increase computational overhead and compromise practicality. It remains unclear how applicable this approach would be in real-time systems such as autonomous driving.\n\n3. Although the paper claims that hyperparameters are transferable, p_min and p_max still require manual tuning for each model and dataset.\n\n4. The paper lacks  visual examples (e.g., t-SNE plots or histograms of energy scores) illustrating the improved separability between ID) and OOD samples.\n\n5. The core hypothesis of the paper that OOD samples exhibit greater instability in high-magnitude activations under minor perturbations is currently supported only by intuitive reasoning, lacking theoretical justification or references to prior work.\n\n6. The implementation details are mostly complete, but the perturbation setup needs more clarification.", "questions": "Please refer to the Weaknesses. The overall experimental evaluation is very comprehensive; however, it would be valuable  to further address above questions to strengthen the paper’s clarity and I will take the rebuttal into consideration when revising my final score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes AdaSCALE, an adaptive scaling procedure\nthat dynamically adjusts the percentile threshold based on a sample’s estimated\nOODness. This estimation leverages a key observation: OOD samples exhibit\nsignificantly more pronounced activation shifts at high-magnitude activations under\nminor perturbation compared to ID samples. AdaSCALE achieves state-of-the-art OOD detection\nperformance, outperforming the latest rival OptFS by 14.94% in near-OOD and\n21.67% in far-OOD datasets in average FPR@95 metric on the ImageNet-1k\nbenchmark across eight diverse architectures.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tClear motivation and novel observation: OOD samples exhibit greater sensitivity of high-magnitude activations to minor perturbations, providing an intuitive and effective signal for designing an adaptive mechanism.\n2.\tOverall, AdaSCALE  is  a simple yet effective method.\n3.\tThe experiments  quite extensive with significant results: across 10 architectures and 3 datasets by tuning mere one hyperparameter for a given setup.", "weaknesses": "1. AdaSCALE shares a similar design  with ATS (Krumpl et al., 2024), as both methods adaptively adjust scaling parameters during post-processing to enhance ID and OOD separability. It would be better to further clarify the differences in design rationale and implementation mechanisms between the two approaches and provide a more in-depth comparative analysis.\n\n2. The experiments employ gradient-based perturbations, which significantly increase computational overhead and compromise practicality. It remains unclear how applicable this approach would be in real-time systems such as autonomous driving.\n\n3. Although the paper claims that hyperparameters are transferable, p_min and p_max still require manual tuning for each model and dataset.\n\n4. The paper lacks  visual examples (e.g., t-SNE plots or histograms of energy scores) illustrating the improved separability between ID) and OOD samples.\n\n5. The core hypothesis of the paper that OOD samples exhibit greater instability in high-magnitude activations under minor perturbations is currently supported only by intuitive reasoning, lacking theoretical justification or references to prior work.\n\n6. The implementation details are mostly complete, but the perturbation setup needs more clarification.", "questions": "Please refer to the Weaknesses. The overall experimental evaluation is very comprehensive; however, it would be valuable  to further address above questions to strengthen the paper’s clarity and I will take the rebuttal into consideration when revising my final score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761492569632}], "openreview_url": "https://openreview.net/forum?id=jbGGPSI9aO", "arxiv_id": "2503.08023", "paper_pdf": "papers/jbGGPSI9aO.pdf", "paper_pdf_sha256": "5b2bb73018cfc8df543d8c24727e55fb236ef6264b6134aed40fb7a90eef68f6", "paper_pdf_bytes": 3368404, "paper_pdf_source": "openreview", "code_url": "https://github.com/sudarshanregmi/AdaSCALE", "code_repository": "sudarshanregmi/AdaSCALE", "code_commit": "ed5f639e10520a04f6a83f30a32c060d6f012ea8", "code_archive": "repos/jbGGPSI9aO.zip", "code_archive_sha256": "7ddee61d37e57e3a1bf8d08863d92a3b5538494c1cabcf7ffe649b0528545fa0", "code_archive_bytes": 3913697, "code_file_count": 681, "code_extensions": {".sh": 427, ".py": 250, ".ipynb": 4}, "github_disk_usage_kb": 3484, "github_languages": {"Python": 919586, "Shell": 313563, "Jupyter Notebook": 300629}, "github_archived": false, "github_pushed_at": "2026-05-02T23:12:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adascale-adaptive-scaling-for-ood-detection"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TjvSFVJdzJ", "year": 2025, "status": "rejected", "title": "Reinforced In-Context Black-Box Optimization", "authors": ["Lei Song", "Chen-Xiao Gao", "Ke Xue", "Chenyang Wu", "Dong Li", "Jianye HAO", "Zongzhang Zhang", "Chao Qian"], "authorids": ["~Lei_Song4", "~Chen-Xiao_Gao1", "~Ke_Xue1", "~Chenyang_Wu1", "~Dong_Li10", "~Jianye_HAO1", "~Zongzhang_Zhang1", "~Chao_Qian1"], "authors_source": "OpenReview API", "abstract": "Black-Box Optimization (BBO) has found successful applications in many fields of science and engineering. Recently, there has been a growing interest in meta-learning particular components of BBO algorithms to speed up optimization and get rid of tedious hand-crafted heuristics. As an extension, learning the entire algorithm from data requires the least labor from experts and can provide the most flexibility. In this paper, we propose RIBBO, a method to reinforce-learn a BBO algorithm from offline data in an end-to-end fashion. RIBBO employs expressive sequence models to learn the optimization histories produced by multiple behavior algorithms and tasks, leveraging the in-context learning ability of large models to extract task information and make decisions accordingly. Central to our method is to augment the optimization histories with *regret-to-go* tokens, which are designed to represent the performance of an algorithm based on cumulative regret over the future part of the histories. The integration of regret-to-go tokens enables RIBBO to automatically generate sequences of query points that satisfy the user-desired regret, which is verified by its universally good empirical performance on diverse problems, including BBO benchmark functions, hyper-parameter optimization and robot control problems.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "iqg7X0Qt58", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6146/Reviewer_REqP"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper is about learning a black-box optimization algorithm in an end-to-end fashion. In terms of meta learning, this approach falls in the category of \"meta-learning the entire algorithm\". The reason for doing this is that it is more flexible than learning something that replaces heuristics in an existing black-box optimization algorithm.  The approach proposed in this paper builds on a large sequence model (causal transformer). The goal is that this model can learn from the sequential behavior of many BBO algorithms on many tasks for making decisions. The supervision signal in this learning task is the regret-to-go which captures the performance of the optimization algorithm. The training regime is offline and the training sequences consist of query points and function values. During the test phase, the model is used in an autoregressive way to generate new query points. Experiments are performed on synthetic functions / data and robotic control problems. The proposed method is compared to state-of-the-art baselines.", "review_text": "This paper is about learning a black-box optimization algorithm in an end-to-end fashion. In terms of meta learning, this approach falls in the category of \"meta-learning the entire algorithm\". The reason for doing this is that it is more flexible than learning something that replaces heuristics in an existing black-box optimization algorithm.  The approach proposed in this paper builds on a large sequence model (causal transformer). The goal is that this model can learn from the sequential behavior of many BBO algorithms on many tasks for making decisions. The supervision signal in this learning task is the regret-to-go which captures the performance of the optimization algorithm. The training regime is offline and the training sequences consist of query points and function values. During the test phase, the model is used in an autoregressive way to generate new query points. Experiments are performed on synthetic functions / data and robotic control problems. The proposed method is compared to state-of-the-art baselines.", "strengths": "The paper is well-written and clear. The presentation is easy to follow and the claims and contributions are clearly stated. The method is novel and performs better than state-of-the-art approaches on the selected experiments.", "weaknesses": "The main weakness of this method is the fact that it has to transfer or generalize to new optimization problems, while algorithmic optimizers have to be tuned to new optimization problems. It is hard to judge how much training is needed for this method to perform well on any \"similar\" problem.", "questions": "In practice, it is also possible to just run the algorithmic optimizer longer to get better solutions and it would be interesting if this is also true for this approach. \n\nSome algorithmic optimizers use local approximations of the objective functions to estimate higher order informations (e.g., gradients) would this be useful for this approach or does this already happen inside of the learned model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper is about learning a black-box optimization algorithm in an end-to-end fashion. In terms of meta learning, this approach falls in the category of \"meta-learning the entire algorithm\". The reason for doing this is that it is more flexible than learning something that replaces heuristics in an existing black-box optimization algorithm.  The approach proposed in this paper builds on a large sequence model (causal transformer). The goal is that this model can learn from the sequential behavior of many BBO algorithms on many tasks for making decisions. The supervision signal in this learning task is the regret-to-go which captures the performance of the optimization algorithm. The training regime is offline and the training sequences consist of query points and function values. During the test phase, the model is used in an autoregressive way to generate new query points. Experiments are performed on synthetic functions / data and robotic control problems. The proposed method is compared to state-of-the-art baselines.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "The paper is well-written and clear. The presentation is easy to follow and the claims and contributions are clearly stated. The method is novel and performs better than state-of-the-art approaches on the selected experiments.", "weaknesses": "The main weakness of this method is the fact that it has to transfer or generalize to new optimization problems, while algorithmic optimizers have to be tuned to new optimization problems. It is hard to judge how much training is needed for this method to perform well on any \"similar\" problem.", "questions": "In practice, it is also possible to just run the algorithmic optimizer longer to get better solutions and it would be interesting if this is also true for this approach. \n\nSome algorithmic optimizers use local approximations of the objective functions to estimate higher order informations (e.g., gradients) would this be useful for this approach or does this already happen inside of the learned model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731061743396}, {"id": "UgkRaobUSY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6146/Reviewer_Zcjr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper proposes a method to solve black-box optimization (BBO) problems end-to-end, using a \"decision transformer\"-like approach.\nThis involves learning a transformer that given a context of query-observation pairs and regret-to-go tokens, selects the next query. This appears to learn from offline meta-training, how to build a model and use this model to make sequential decisions.", "review_text": "The paper proposes a method to solve black-box optimization (BBO) problems end-to-end, using a \"decision transformer\"-like approach.\nThis involves learning a transformer that given a context of query-observation pairs and regret-to-go tokens, selects the next query. This appears to learn from offline meta-training, how to build a model and use this model to make sequential decisions.", "strengths": "The BBO problem is a very relevant problem. Given that BBO can be framed as a special case of RL (where the state is kept constant), applying a decision transformer (DT) (or something similar) to this problem seems like an interesting approach.", "weaknesses": "I see some weaknesses both in the proposed method (and how it differs from DT), and the experimental evaluation.\n\nMethodology:\n\n* In my opinion, the manuscript would benefit from a more careful distinction from decision transformers, especially, since the BBO setting is a special case of the standard RL setting (where the state is constant). Why do the authors use regret-to-go tokens as opposed to returns-to-go in DT? In DT, no \"observation tokens\" are added to the sequence, and instead returns (i.e., function \"observations\") are subtracted from the RTG. Furthermore, I would like to see a comparison against DT as a baseline in the experiments. From the current paper, it does not become evident why RIBBO would be preferable over DT.\n\n* The assumption that the \"true\" optimum value $y^*$ is known seems off to me. How would this value be known in practice, e.g., for a molecular design task? Other BBO algorithms such as UCB/EI/... do not require knowledge of $y^*$, and instead effectively balance exploration & exploitation. This seems like a fundamental assumption of the proposed approach, and to my understanding is why regret-to-go tokens might not be so practical (DT uses return-to-go tokens which do not have this problem).\n\n* Setting $R_t = 0$ before choosing the next action suggests to me that the acquisition function is executed in the \"exploitation regime\". That is, in the meta-training set, I would assume that actions from algorithms that follow $R_t = 0$ are most likely from a stage where the optimum has already been almost perfectly determined, and the BBO algorithm purely exploits the model (i.e., does not focus on further exploration). Perhaps \"knowing\" the true $y^*$ a priori can compensate for this lack of exploration, but I would assume that this does not work in challenging exploration tasks where $y^*$ is unknown a priori. Can the authors elaborate why they expect setting $R_t = 0$ would lead to explorative behavior?\n\nExperimental evaluation:\n\n* It is not clear to me why \"imitating\" a BBO algorithm (like UCB, EI, ...) with a transformer is preferable over optimizing said algorithm directly. The \"imitation\" setting seems more applicable to cases where we want to imitate decisions of a \"human expert\" that cannot be expressed in closed-form, as is done in BC. This paper seems to claim that the main reason for learning a transformer is that this can leverage meta-training data to select the right BBO algorithm on the fly, however, in my view the experiments are insufficient to support this claim. This is because (1) the experiments do not compare against any other ad-hoc method of choosing BBO algorithms on the fly based on their performance on similar tasks in a meta-training set (this would not require re-learning the BBO algorithm, just learning when to apply which one), and (2) the experiments do not appear to compare against baselines that also use the meta-training set. In other words, I believe that the experiments are heavily skewed in RIBBO's favor since RIBBO uses a meta-training set while most BBO baselines have not seen this extra data. It might be that RIBBO exploits additional patterns in the meta-training set beyond just imitating one of the BBO algorithms. In my view, the experiments should compare against BBO algorithms that operate over meta-learned models which incorporate knowledge from the meta-training set. See, for example, [1]. In my opinion, convincing experiments would have to evaluate against such approaches that meta-learn individual components.\n\n* It seems to me that the most important results are in Appendix D where the authors test RIBBO in scenarios where the same function family are not included in the offline meta training set. In the results from the main test, RIBBO might make extensive use of offline information that is not available to most of the tested baselines (for example, GP-EI). It seems that in Appendix D, RIBBO is frequently outperformed by GP-EI. Can the authors elaborate on why that could be? This seems concerning to me.\n\n\n[1]: Rothfuss et al., Meta-Learning Reliable Priors in the Function Space.", "questions": "* I would suggest that before Eq. (2) it is highlighted that this is the standard behavioral cloning baseline.\n\n* Can't the proposed HRR also lead to OOD RTGs, in case the $R_t \\gg 0$ when no similar data was seen before in the meta-training set? That is, if the tasks encountered are more difficult than the tasks in the meta-train set.\n\n* What is the size of $M$ and $T$ in the experiments? Do the results look different with smaller meta training sets? What are the minimal values where RIBBO starts to outperform basic BBO algorithms (that do not use the meta training data)?\n\n* In Figures 3b,3c,3d, why were the specific functions chosen / \"cherry-picked\" for this visualization? Can you add the visualizations with the other test-functions to the supplementary material?\n\n* In multiple places (for example, line 534), the authors claim that RIBBO can be used to generate optimization trajectories satisfying a user-specified regret. This claim does not seem to be sufficiently supported by the experiments. The experiments do indicate that there is a correlation between the user-specified regret and the obtained regret, but even in Figure 3c, the mean regret when a user specified regret $10$ is significantly larger than $10$.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method to solve black-box optimization (BBO) problems end-to-end, using a \"decision transformer\"-like approach.\nThis involves learning a transformer that given a context of query-observation pairs and regret-to-go tokens, selects the next query. This appears to learn from offline meta-training, how to build a model and use this model to make sequential decisions.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The BBO problem is a very relevant problem. Given that BBO can be framed as a special case of RL (where the state is kept constant), applying a decision transformer (DT) (or something similar) to this problem seems like an interesting approach.", "weaknesses": "I see some weaknesses both in the proposed method (and how it differs from DT), and the experimental evaluation.\n\nMethodology:\n\n* In my opinion, the manuscript would benefit from a more careful distinction from decision transformers, especially, since the BBO setting is a special case of the standard RL setting (where the state is constant). Why do the authors use regret-to-go tokens as opposed to returns-to-go in DT? In DT, no \"observation tokens\" are added to the sequence, and instead returns (i.e., function \"observations\") are subtracted from the RTG. Furthermore, I would like to see a comparison against DT as a baseline in the experiments. From the current paper, it does not become evident why RIBBO would be preferable over DT.\n\n* The assumption that the \"true\" optimum value $y^*$ is known seems off to me. How would this value be known in practice, e.g., for a molecular design task? Other BBO algorithms such as UCB/EI/... do not require knowledge of $y^*$, and instead effectively balance exploration & exploitation. This seems like a fundamental assumption of the proposed approach, and to my understanding is why regret-to-go tokens might not be so practical (DT uses return-to-go tokens which do not have this problem).\n\n* Setting $R_t = 0$ before choosing the next action suggests to me that the acquisition function is executed in the \"exploitation regime\". That is, in the meta-training set, I would assume that actions from algorithms that follow $R_t = 0$ are most likely from a stage where the optimum has already been almost perfectly determined, and the BBO algorithm purely exploits the model (i.e., does not focus on further exploration). Perhaps \"knowing\" the true $y^*$ a priori can compensate for this lack of exploration, but I would assume that this does not work in challenging exploration tasks where $y^*$ is unknown a priori. Can the authors elaborate why they expect setting $R_t = 0$ would lead to explorative behavior?\n\nExperimental evaluation:\n\n* It is not clear to me why \"imitating\" a BBO algorithm (like UCB, EI, ...) with a transformer is preferable over optimizing said algorithm directly. The \"imitation\" setting seems more applicable to cases where we want to imitate decisions of a \"human expert\" that cannot be expressed in closed-form, as is done in BC. This paper seems to claim that the main reason for learning a transformer is that this can leverage meta-training data to select the right BBO algorithm on the fly, however, in my view the experiments are insufficient to support this claim. This is because (1) the experiments do not compare against any other ad-hoc method of choosing BBO algorithms on the fly based on their performance on similar tasks in a meta-training set (this would not require re-learning the BBO algorithm, just learning when to apply which one), and (2) the experiments do not appear to compare against baselines that also use the meta-training set. In other words, I believe that the experiments are heavily skewed in RIBBO's favor since RIBBO uses a meta-training set while most BBO baselines have not seen this extra data. It might be that RIBBO exploits additional patterns in the meta-training set beyond just imitating one of the BBO algorithms. In my view, the experiments should compare against BBO algorithms that operate over meta-learned models which incorporate knowledge from the meta-training set. See, for example, [1]. In my opinion, convincing experiments would have to evaluate against such approaches that meta-learn individual components.\n\n* It seems to me that the most important results are in Appendix D where the authors test RIBBO in scenarios where the same function family are not included in the offline meta training set. In the results from the main test, RIBBO might make extensive use of offline information that is not available to most of the tested baselines (for example, GP-EI). It seems that in Appendix D, RIBBO is frequently outperformed by GP-EI. Can the authors elaborate on why that could be? This seems concerning to me.\n\n\n[1]: Rothfuss et al., Meta-Learning Reliable Priors in the Function Space.", "questions": "* I would suggest that before Eq. (2) it is highlighted that this is the standard behavioral cloning baseline.\n\n* Can't the proposed HRR also lead to OOD RTGs, in case the $R_t \\gg 0$ when no similar data was seen before in the meta-training set? That is, if the tasks encountered are more difficult than the tasks in the meta-train set.\n\n* What is the size of $M$ and $T$ in the experiments? Do the results look different with smaller meta training sets? What are the minimal values where RIBBO starts to outperform basic BBO algorithms (that do not use the meta training data)?\n\n* In Figures 3b,3c,3d, why were the specific functions chosen / \"cherry-picked\" for this visualization? Can you add the visualizations with the other test-functions to the supplementary material?\n\n* In multiple places (for example, line 534), the authors claim that RIBBO can be used to generate optimization trajectories satisfying a user-specified regret. This claim does not seem to be sufficiently supported by the experiments. The experiments do indicate that there is a correlation between the user-specified regret and the obtained regret, but even in Figure 3c, the mean regret when a user specified regret $10$ is significantly larger than $10$.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730714827223}, {"id": "4FpKdOaf3W", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6146/Reviewer_azoC"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "Black-Box Optimization (BBO) is the process of optimizing an objective function where neither analytic expression nor derivatives of the objective function are available. Although Bayesian Optimization (BO) and Evolutionary Algorithms (EA) are the main BBO algorithms developed, they solve BBO problems from scratch while relying on expert-derived heuristics. Recently, the meta-learning paradigm has been adopted to learn some components of the algorithms from previously collected data, yet it requires a degree of expert knowledge. On the contrary, learning the entire BBO algorithm in an End-to-End (E2E) fashion from data requires no expert knowledge at all while providing flexibility among a wide range of BBO problems. However, learning such algorithms is challenging, and in practice, selecting the specific BBO algorithm is needed during inference (as in OptFormer [1]). In this work, an approach named Reinforced In-context BBO (RIBBO) is proposed to learn a general BBO algorithm from offline data collected using multiple BBO algorithms. Using a causal transformer, RIBBO can predict the next query point, given the optimization history. To alleviate the need for expert knowledge during inference when selecting the desired BBO algorithm, RIBBO augments the optimization history with regret-to-go RTG tokens representing the future performance of an algorithm. Consequently, RIBBO can automatically identify different algorithms and generate sequences of query points that satisfy the specified regret. So, it adds an interpretable component that is easy to manipulate during inference. To update the RTG tokens during testing, this work proposes a novel Hindsight Regret Relabelling (HRR) strategy by setting the immediate RTG as 0. Finally, the potential of this approach has been demonstrated in three problem domains: BBOB synthetic functions, hyperparameter optimization, and robot control problems while comparing to classical approaches (that generated the offline dataset) and related baselines like Behavior Cloning (BC) and OptFormer [1].\n\n[1] Chen, Yutian, et al. \"Towards learning universal hyperparameter optimizers with transformers.\" Advances in Neural Information Processing Systems 35 (2022): 32053-32068.", "review_text": "Black-Box Optimization (BBO) is the process of optimizing an objective function where neither analytic expression nor derivatives of the objective function are available. Although Bayesian Optimization (BO) and Evolutionary Algorithms (EA) are the main BBO algorithms developed, they solve BBO problems from scratch while relying on expert-derived heuristics. Recently, the meta-learning paradigm has been adopted to learn some components of the algorithms from previously collected data, yet it requires a degree of expert knowledge. On the contrary, learning the entire BBO algorithm in an End-to-End (E2E) fashion from data requires no expert knowledge at all while providing flexibility among a wide range of BBO problems. However, learning such algorithms is challenging, and in practice, selecting the specific BBO algorithm is needed during inference (as in OptFormer [1]). In this work, an approach named Reinforced In-context BBO (RIBBO) is proposed to learn a general BBO algorithm from offline data collected using multiple BBO algorithms. Using a causal transformer, RIBBO can predict the next query point, given the optimization history. To alleviate the need for expert knowledge during inference when selecting the desired BBO algorithm, RIBBO augments the optimization history with regret-to-go RTG tokens representing the future performance of an algorithm. Consequently, RIBBO can automatically identify different algorithms and generate sequences of query points that satisfy the specified regret. So, it adds an interpretable component that is easy to manipulate during inference. To update the RTG tokens during testing, this work proposes a novel Hindsight Regret Relabelling (HRR) strategy by setting the immediate RTG as 0. Finally, the potential of this approach has been demonstrated in three problem domains: BBOB synthetic functions, hyperparameter optimization, and robot control problems while comparing to classical approaches (that generated the offline dataset) and related baselines like Behavior Cloning (BC) and OptFormer [1].\n\n[1] Chen, Yutian, et al. \"Towards learning universal hyperparameter optimizers with transformers.\" Advances in Neural Information Processing Systems 35 (2022): 32053-32068.", "strengths": "- The paper is well-written and clear.\n- The literature is covered nicely in addition to motivating the proposed approach of learning the BBO algorithm in an E2E fashion. \n- The experiments and the baselines demonstrate the potential of the proposed approach.\n- The discussion section and ablation studies are comprehensive to understand the impact of each component in the algorithm.", "weaknesses": "There are no significant weaknesses in this work, yet some important comments should be mentioned to improve this work.\n- The limitation of this approach is not well highlighted.\n- The quality of plots can be improved. I personally don't like the overlapping vertical lines that represent the standard deviation. A shaded region could be better.", "questions": "- What are the major limitations of this approach? \n- During inference, as a user, how can I select the initial RTG? Could you please provide examples where a low or a high initial RTG value is preferred? I believe I didn't get this part well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Black-Box Optimization (BBO) is the process of optimizing an objective function where neither analytic expression nor derivatives of the objective function are available. Although Bayesian Optimization (BO) and Evolutionary Algorithms (EA) are the main BBO algorithms developed, they solve BBO problems from scratch while relying on expert-derived heuristics. Recently, the meta-learning paradigm has been adopted to learn some components of the algorithms from previously collected data, yet it requires a degree of expert knowledge. On the contrary, learning the entire BBO algorithm in an End-to-End (E2E) fashion from data requires no expert knowledge at all while providing flexibility among a wide range of BBO problems. However, learning such algorithms is challenging, and in practice, selecting the specific BBO algorithm is needed during inference (as in OptFormer [1]). In this work, an approach named Reinforced In-context BBO (RIBBO) is proposed to learn a general BBO algorithm from offline data collected using multiple BBO algorithms. Using a causal transformer, RIBBO can predict the next query point, given the optimization history. To alleviate the need for expert knowledge during inference when selecting the desired BBO algorithm, RIBBO augments the optimization history with regret-to-go RTG tokens representing the future performance of an algorithm. Consequently, RIBBO can automatically identify different algorithms and generate sequences of query points that satisfy the specified regret. So, it adds an interpretable component that is easy to manipulate during inference. To update the RTG tokens during testing, this work proposes a novel Hindsight Regret Relabelling (HRR) strategy by setting the immediate RTG as 0. Finally, the potential of this approach has been demonstrated in three problem domains: BBOB synthetic functions, hyperparameter optimization, and robot control problems while comparing to classical approaches (that generated the offline dataset) and related baselines like Behavior Cloning (BC) and OptFormer [1].\n\n[1] Chen, Yutian, et al. \"Towards learning universal hyperparameter optimizers with transformers.\" Advances in Neural Information Processing Systems 35 (2022): 32053-32068.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper is well-written and clear.\n- The literature is covered nicely in addition to motivating the proposed approach of learning the BBO algorithm in an E2E fashion. \n- The experiments and the baselines demonstrate the potential of the proposed approach.\n- The discussion section and ablation studies are comprehensive to understand the impact of each component in the algorithm.", "weaknesses": "There are no significant weaknesses in this work, yet some important comments should be mentioned to improve this work.\n- The limitation of this approach is not well highlighted.\n- The quality of plots can be improved. I personally don't like the overlapping vertical lines that represent the standard deviation. A shaded region could be better.", "questions": "- What are the major limitations of this approach? \n- During inference, as a user, how can I select the initial RTG? Could you please provide examples where a low or a high initial RTG value is preferred? I believe I didn't get this part well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730674368056}, {"id": "sJmIM7H90N", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6146/Reviewer_dM2J"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper presents a meta-learning algorithm designed for black-box optimization (BBO) problems, which are often derivative-free and lack a gradient signal. The proposed method uses an offline dataset collected from behavioral optimization algorithms, each of which generates sequences of optimization steps across various problems. The algorithm then trains a sequence-to-sequence model based on the Transformer architecture to replicate these sequences, aiming to generalize effectively and produce a high-quality optimization sequence when applied to new problems sampled from the training problem distribution.\n\nIn a typical optimization sequence, each step includes a pair of values, (x_i, y_i), where x and y represent the input and output of the BBO problem, respectively. The authors propose enhancing this dataset by adding a Regret-To-Go (RTG) scalar, represented as (x_i, y_i, r_i), where r_i is the sum of future regrets for t > i. This paper distinguishes between Behavioral Cloning (BC), which learns without RTG, and RIBBO, which incorporates RTG into the model training.\n\nThe authors test their algorithm across three datasets: (1) BBOB, a collection of synthetic optimization problems grouped by shared characteristics; (2) HPO-B, a set of machine learning hyperparameter optimization problems targeting base models such as XGBoost and SVM; and (3) Rover trajectory planning, where the goal is to optimize a sequence of 30 two-dimensional points that minimize a cost function. The algorithm is evaluated against other meta-learning BBO approaches, such as OptFormer and BC, as well as traditional BBO methods, including CMA-ES. RIBBO demonstrates strong performance, outperforming the alternatives in these experiments.", "review_text": "This paper presents a meta-learning algorithm designed for black-box optimization (BBO) problems, which are often derivative-free and lack a gradient signal. The proposed method uses an offline dataset collected from behavioral optimization algorithms, each of which generates sequences of optimization steps across various problems. The algorithm then trains a sequence-to-sequence model based on the Transformer architecture to replicate these sequences, aiming to generalize effectively and produce a high-quality optimization sequence when applied to new problems sampled from the training problem distribution.\n\nIn a typical optimization sequence, each step includes a pair of values, (x_i, y_i), where x and y represent the input and output of the BBO problem, respectively. The authors propose enhancing this dataset by adding a Regret-To-Go (RTG) scalar, represented as (x_i, y_i, r_i), where r_i is the sum of future regrets for t > i. This paper distinguishes between Behavioral Cloning (BC), which learns without RTG, and RIBBO, which incorporates RTG into the model training.\n\nThe authors test their algorithm across three datasets: (1) BBOB, a collection of synthetic optimization problems grouped by shared characteristics; (2) HPO-B, a set of machine learning hyperparameter optimization problems targeting base models such as XGBoost and SVM; and (3) Rover trajectory planning, where the goal is to optimize a sequence of 30 two-dimensional points that minimize a cost function. The algorithm is evaluated against other meta-learning BBO approaches, such as OptFormer and BC, as well as traditional BBO methods, including CMA-ES. RIBBO demonstrates strong performance, outperforming the alternatives in these experiments.", "strengths": "1. Data-Efficiency: \n\nRIBBO addresses a significant problem in BBO by learning from sequences of sub-optimal trials. This approach enables the algorithm to learn from incomplete or unsuccessful trials, potentially reducing computational costs.\n\n2. Implementation and Training Simplicity: \n\nThe method is relatively straightforward to implement and train, given the simplicity of the Transformer architecture and the offline training process.", "weaknesses": "1. Behavioral Policy Limitations: \n\nThe potential of RIBBO to outperform the behavioral policies it learns from is uncertain. Unlike reinforcement learning (RL), RIBBO lacks a policy optimization component; it merely learns to output the next query, x_i, based on past inputs, outputs, and RTG. This approach is essentially behavioral cloning that might favor sequences with lower regret if the RTG functions as intended. However, without optimizing the output sequences (e.g., by approximating the target output value, y, and deriving a gradient signal [1]), RIBBO’s ability to surpass the behavioral policies remains unclear.\n\n2. Justification for RTG Inclusion: \n\nThe authors could strengthen their argument and visual representation of why RTG enhances solution quality. It is not evident why augmenting the input data with RTG would encourage the model to generate lower-regret sequences. Why is the network motivated to generate sequences of lower RTG?\n\n3. Transferability Across BBO Problems: \n\nDomains like RL or few-shot learning provide clear justifications for what knowledge can transfer across tasks, such as environment dynamics or domain distribution, respectively. However, in black-box problems, a shared generative function across tasks is not always present. Even in cases where tasks are sampled from a common distribution, such as in the same group in the BBOB testbench or the rover trajectory planning problem, RIBBO lacks an adaptation mechanism like MAML[2], which would help in adjusting to different environments. Consequently, the paper should address whether and why RIBBO is truly adaptable to unseen tasks.\n\n\nReferences:\n1. Sarafian, E., Sinay, M., Louzoun, Y., Agmon, N., & Kraus, S. (2020, July). Explicit Gradient Learning for Black-Box Optimization. In ICML (pp. 8480-8490).\n2. Finn, Chelsea, Pieter Abbeel, and Sergey Levine. \"Model-agnostic meta-learning for fast adaptation of deep networks.\" International conference on machine learning. PMLR, 2017.", "questions": "1. BBOB Experiment Clarifications: \n\nIn the BBOB experiments, what dimensions were selected, and what is the interpretation of the x-axis (number of evaluations)?\n\n2. Hyperparameter Sensitivity: \n\nA primary motivation for RIBBO is to alleviate the challenge of manually tuning hyperparameters (lines 13–16 in the abstract). However, RIBBO itself requires hyperparameter tuning due to the nature of neural network training. Some of these parameters are listed in Table 1 (Appendix). How were these hyperparameters tuned, and could the authors provide sensitivity metrics to show their impact on performance?\n\n3. Literature Review Addition: \n\nOpt-GAN [1] is another BBO method that aims to learn a global optimizer through a neural network, predicting the next candidate point based on previous points sampled in an off-policy fashion. It would be beneficial to include this work in the literature review.\n\nReferences\n1[] Lu, Minfang, et al. \"OPT-GAN: a broad-spectrum global optimizer for black-box problems by learning distribution.\" Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a meta-learning algorithm designed for black-box optimization (BBO) problems, which are often derivative-free and lack a gradient signal. The proposed method uses an offline dataset collected from behavioral optimization algorithms, each of which generates sequences of optimization steps across various problems. The algorithm then trains a sequence-to-sequence model based on the Transformer architecture to replicate these sequences, aiming to generalize effectively and produce a high-quality optimization sequence when applied to new problems sampled from the training problem distribution.\n\nIn a typical optimization sequence, each step includes a pair of values, (x_i, y_i), where x and y represent the input and output of the BBO problem, respectively. The authors propose enhancing this dataset by adding a Regret-To-Go (RTG) scalar, represented as (x_i, y_i, r_i), where r_i is the sum of future regrets for t > i. This paper distinguishes between Behavioral Cloning (BC), which learns without RTG, and RIBBO, which incorporates RTG into the model training.\n\nThe authors test their algorithm across three datasets: (1) BBOB, a collection of synthetic optimization problems grouped by shared characteristics; (2) HPO-B, a set of machine learning hyperparameter optimization problems targeting base models such as XGBoost and SVM; and (3) Rover trajectory planning, where the goal is to optimize a sequence of 30 two-dimensional points that minimize a cost function. The algorithm is evaluated against other meta-learning BBO approaches, such as OptFormer and BC, as well as traditional BBO methods, including CMA-ES. RIBBO demonstrates strong performance, outperforming the alternatives in these experiments.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "1. Data-Efficiency: \n\nRIBBO addresses a significant problem in BBO by learning from sequences of sub-optimal trials. This approach enables the algorithm to learn from incomplete or unsuccessful trials, potentially reducing computational costs.\n\n2. Implementation and Training Simplicity: \n\nThe method is relatively straightforward to implement and train, given the simplicity of the Transformer architecture and the offline training process.", "weaknesses": "1. Behavioral Policy Limitations: \n\nThe potential of RIBBO to outperform the behavioral policies it learns from is uncertain. Unlike reinforcement learning (RL), RIBBO lacks a policy optimization component; it merely learns to output the next query, x_i, based on past inputs, outputs, and RTG. This approach is essentially behavioral cloning that might favor sequences with lower regret if the RTG functions as intended. However, without optimizing the output sequences (e.g., by approximating the target output value, y, and deriving a gradient signal [1]), RIBBO’s ability to surpass the behavioral policies remains unclear.\n\n2. Justification for RTG Inclusion: \n\nThe authors could strengthen their argument and visual representation of why RTG enhances solution quality. It is not evident why augmenting the input data with RTG would encourage the model to generate lower-regret sequences. Why is the network motivated to generate sequences of lower RTG?\n\n3. Transferability Across BBO Problems: \n\nDomains like RL or few-shot learning provide clear justifications for what knowledge can transfer across tasks, such as environment dynamics or domain distribution, respectively. However, in black-box problems, a shared generative function across tasks is not always present. Even in cases where tasks are sampled from a common distribution, such as in the same group in the BBOB testbench or the rover trajectory planning problem, RIBBO lacks an adaptation mechanism like MAML[2], which would help in adjusting to different environments. Consequently, the paper should address whether and why RIBBO is truly adaptable to unseen tasks.\n\n\nReferences:\n1. Sarafian, E., Sinay, M., Louzoun, Y., Agmon, N., & Kraus, S. (2020, July). Explicit Gradient Learning for Black-Box Optimization. In ICML (pp. 8480-8490).\n2. Finn, Chelsea, Pieter Abbeel, and Sergey Levine. \"Model-agnostic meta-learning for fast adaptation of deep networks.\" International conference on machine learning. PMLR, 2017.", "questions": "1. BBOB Experiment Clarifications: \n\nIn the BBOB experiments, what dimensions were selected, and what is the interpretation of the x-axis (number of evaluations)?\n\n2. Hyperparameter Sensitivity: \n\nA primary motivation for RIBBO is to alleviate the challenge of manually tuning hyperparameters (lines 13–16 in the abstract). However, RIBBO itself requires hyperparameter tuning due to the nature of neural network training. Some of these parameters are listed in Table 1 (Appendix). How were these hyperparameters tuned, and could the authors provide sensitivity metrics to show their impact on performance?\n\n3. Literature Review Addition: \n\nOpt-GAN [1] is another BBO method that aims to learn a global optimizer through a neural network, predicting the next candidate point based on previous points sampled in an off-policy fashion. It would be beneficial to include this work in the literature review.\n\nReferences\n1[] Lu, Minfang, et al. \"OPT-GAN: a broad-spectrum global optimizer for black-box problems by learning distribution.\" Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730622023888}], "openreview_url": "https://openreview.net/forum?id=TjvSFVJdzJ", "arxiv_id": "2402.17423", "paper_pdf": "papers/TjvSFVJdzJ.pdf", "paper_pdf_sha256": "605cf117362bf50c1b81eef9816c537674d6bc69f5af0c03c825ae0dc3781bdc", "paper_pdf_bytes": 1587027, "paper_pdf_source": "openreview", "code_url": "https://github.com/lamda-bbo/RIBBO", "code_repository": "lamda-bbo/RIBBO", "code_commit": "1b42a7da42f6ceba6e92299d5d8dc6583cdd380e", "code_archive": "repos/TjvSFVJdzJ.zip", "code_archive_sha256": "e90c92edd7b9eb3e646ff09aaa14bb6c3e459c025078377ef947f400b869cd5a", "code_archive_bytes": 136125, "code_file_count": 97, "code_extensions": {".py": 92, ".sh": 5}, "github_disk_usage_kb": 247, "github_languages": {"Python": 353462, "Shell": 4728}, "github_archived": false, "github_pushed_at": "2025-05-19T06:21:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/reinforced-in-context-black-box-optimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Rt6btdXS2b", "year": 2024, "status": "rejected", "title": "Continuous Indeterminate Probability Neural Network", "authors": ["Tao Yang"], "authorids": ["~Tao_Yang19"], "authors_source": "OpenReview API", "abstract": "Currently, there is no mathematical analytical form for a general posterior, however, Indeterminate Probability Theory has now discovered a way to address this issue. This is a big discovery in the field of probability and it is applicable in various fields.\nThis paper introduces a general model called CIPNN - Continuous Indeterminate Probability Neural Network, which is an analytical probability neural network with continuous latent random variables.  \nOur contributions are Four-fold. First, we apply the analytical form of the posterior for continuous latent random variables and propose a general classification model (CIPNN). Second, we propose a general auto-encoder called CIPAE - Continuous Indeterminate Probability Auto-Encoder, instead of using a neural network as the decoder component, we first employ a probabilistic equation. Third, we propose a new method to visualize the latent random variables, we use one of N dimensional latent variables as a decoder to reconstruct the input image, which can work even for classification tasks, in this way, we can see what each latent variable has learned. Fourth, IPNN has shown great classification capability, CIPNN has pushed this classification capability to infinity.\nTheoretical advantages are reflected in experimental results.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "i7T3tmPgGS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1578/Reviewer_EMUS"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a general model called CIPNN - Continuous Indeterminate Probability Neural Network using a group of reference variables $z$.", "review_text": "This paper proposes a general model called CIPNN - Continuous Indeterminate Probability Neural Network using a group of reference variables $z$.", "strengths": "This paper focuses on the explainability of probabilistic models and neural networks, which is an interesting and important topic.", "weaknesses": "1. I find the motivation of this paper unclear. I had difficulty following the progression from Section 2.2 to Section 3 and then to the CIPNN model.\n\n2. The indeterminate probability theory is not surprising to me, and I believe it can be easily derived via the definition of conditional probability. Equations (1) and (2) hold for any $z$, but it's not clear to me which specific type of $z$ we are expecting in the learning process.\n\n3. I find Proposition 1 to be weak from my perspective. Specifically, Proposition 1 states, ''If $P(y_l | z^1, ... , z^N) \\to \\infty$, CIPNN converges to the global minimum.'' This is equivalent to saying that successful classification depends on our ability to learn a set of favorable variables, namely $z^1, ... , z^N$. However, the main challenge lies in determining the existence of these 'good' variables and how we can identify and obtain such $z^1, ... , z^N$ with theoretical guarantees.\n\n4. I did not find the comparison with existing approaches. Also, the numerical results did not indicates the improved performance is indeed from the introduction of $z^1, ... , z^N$.", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a general model called CIPNN - Continuous Indeterminate Probability Neural Network using a group of reference variables $z$.", "soundness": "1 poor", "presentation": "1 poor", "contribution": "1 poor", "strengths": "This paper focuses on the explainability of probabilistic models and neural networks, which is an interesting and important topic.", "weaknesses": "1. I find the motivation of this paper unclear. I had difficulty following the progression from Section 2.2 to Section 3 and then to the CIPNN model.\n\n2. The indeterminate probability theory is not surprising to me, and I believe it can be easily derived via the definition of conditional probability. Equations (1) and (2) hold for any $z$, but it's not clear to me which specific type of $z$ we are expecting in the learning process.\n\n3. I find Proposition 1 to be weak from my perspective. Specifically, Proposition 1 states, ''If $P(y_l | z^1, ... , z^N) \\to \\infty$, CIPNN converges to the global minimum.'' This is equivalent to saying that successful classification depends on our ability to learn a set of favorable variables, namely $z^1, ... , z^N$. However, the main challenge lies in determining the existence of these 'good' variables and how we can identify and obtain such $z^1, ... , z^N$ with theoretical guarantees.\n\n4. I did not find the comparison with existing approaches. Also, the numerical results did not indicates the improved performance is indeed from the introduction of $z^1, ... , z^N$.", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698776160023}, {"id": "bgsw94EcPI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1578/Reviewer_Vwsb"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper “Continuous indeterminate…” proposes a continuous extension of the “Indeterminate …” model by the same authors, correctly referenced as Anonymous. The paper describes this extension, accompanied by definitions of the classification and auto-encoder models, together with training, inference procedures and simple experiments. The resulting models are only a bit less accurate than well known models. The author's main goal is to theoretically describe and show the benefits from its use.\n\nIn my opinion, the paper may be accepted, provided the authors answer the above doubts.", "review_text": "The paper “Continuous indeterminate…” proposes a continuous extension of the “Indeterminate …” model by the same authors, correctly referenced as Anonymous. The paper describes this extension, accompanied by definitions of the classification and auto-encoder models, together with training, inference procedures and simple experiments. The resulting models are only a bit less accurate than well known models. The author's main goal is to theoretically describe and show the benefits from its use.\n\nIn my opinion, the paper may be accepted, provided the authors answer the above doubts.", "strengths": "1. The model shown is interesting and shows, perhaps not very illuminating but still explainable to the input —> latent —> classification, and input —> latent —> reconstruction problems in theoretical way.\n2. There is a good introductory to the theory in section 3.\n3. The proposed model aims at providing more explainable solutions to classification, although there is some way before the model may accomplish that.\n4. The performed experiments prove, or at least show, the hypothesis clearly stated by the authors.", "weaknesses": "1. I guess the whole paper should start with a deeper explanation of differences between the proposed approach and a VAE model.\n2. An ablation study concerning the complexity is missing. The authors say that the C hyperparameter can be set to 1 “as long as the batch size is high enough…” (page 7, bottom), but still they use C=2 in the experiments.\n3. The derivation for the continuous probability mathematical formulae is complex, and lacks intuition, instead giving intricate formulas and variables.\n4. Although the authors correctly reference to their own paper as written by Anonymous, but not yet published. The authors do it all through the paper referencing the reader to find details over there. The paper can easily be found by the title. On the other hand, this is unavoidable.", "questions": "1.  Add some introduction to differences between the proposed model and VAE-like approaches, or perhaps accompany the whole sequence should be accompanied by comparisons to corresponding steps in a VAE-type model?\n2. Equation (21), being the basis for training, needs a deeper explanation. Why use the max functions both in numerator and denominator? \n3. When comparing the proposed CIPAE with VAE, section 7.2, the visualizations of the latent space become somehow different, with parts of the R^2 latent for VAE empty. Does it come from different latent definition in both cases? Or is it just a result of showing only the [-20, 20]x[-20,20] square? This might not seem to be a fair comparison unless explained.\n4. From a practical point of view: what is the training and inference comparison between VAE (as well as other WAE, etc.) approach and the proposed ones?\n5.  What is the impact of C value and the batch size on quality, trading and inference time, etc.? Could you provide some ablation study?\n6. In conclusion the authors state, that the proposed model is actually composed of two parts: first detects attributes, and the second (i.e. classification?) is a probabilistic model which may be used for reasoning. A 1-D example shown in figure 3, that the authors refer to, shows that the first part performs a kind of clustering, is that so? Please elaborate on that, since it would greatly enlarge the readability of the paper.\n7. Several language errors should be corrected. E.g. a) on page 1 the sentence “However, IPNN need to predefine the …” should probably be “However, IPNN needs to be predefined…”; b) just above Eq. 1 instead “bellow” should be “below”; c) what is the word “complexer” at the bottom of page 3? Perhaps the authors meant to say “more complex? “Complexer” might be used in French. I would suggest checking the whole text with some native speaker. These errors are usually tiny, but disturb reading.\n8. Please, if possible, make the figures a bit larger, just to make them somehow more readable. This refers particularly to figures 1, 3, and perhaps 2 and 4 too.\n9. In section 7.1 you claim that the CIPNN tends to put 1, 4, 7, 9 MNIST numbers in one cluster — this is hardly visible in the figures. How is that model used for classification? Which inputs were used in each round? Could you elaborate on that a bit? \n10. Equations are sometimes complex (lots of variables and indices), e.g. Sequence from (9) to (14). Could you, please, make them easier to follow?\n11. Some small editing errors, e.g. subsection 7.2 title starts as an orphan.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper “Continuous indeterminate…” proposes a continuous extension of the “Indeterminate …” model by the same authors, correctly referenced as Anonymous. The paper describes this extension, accompanied by definitions of the classification and auto-encoder models, together with training, inference procedures and simple experiments. The resulting models are only a bit less accurate than well known models. The author's main goal is to theoretically describe and show the benefits from its use.\n\nIn my opinion, the paper may be accepted, provided the authors answer the above doubts.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The model shown is interesting and shows, perhaps not very illuminating but still explainable to the input —> latent —> classification, and input —> latent —> reconstruction problems in theoretical way.\n2. There is a good introductory to the theory in section 3.\n3. The proposed model aims at providing more explainable solutions to classification, although there is some way before the model may accomplish that.\n4. The performed experiments prove, or at least show, the hypothesis clearly stated by the authors.", "weaknesses": "1. I guess the whole paper should start with a deeper explanation of differences between the proposed approach and a VAE model.\n2. An ablation study concerning the complexity is missing. The authors say that the C hyperparameter can be set to 1 “as long as the batch size is high enough…” (page 7, bottom), but still they use C=2 in the experiments.\n3. The derivation for the continuous probability mathematical formulae is complex, and lacks intuition, instead giving intricate formulas and variables.\n4. Although the authors correctly reference to their own paper as written by Anonymous, but not yet published. The authors do it all through the paper referencing the reader to find details over there. The paper can easily be found by the title. On the other hand, this is unavoidable.", "questions": "1.  Add some introduction to differences between the proposed model and VAE-like approaches, or perhaps accompany the whole sequence should be accompanied by comparisons to corresponding steps in a VAE-type model?\n2. Equation (21), being the basis for training, needs a deeper explanation. Why use the max functions both in numerator and denominator? \n3. When comparing the proposed CIPAE with VAE, section 7.2, the visualizations of the latent space become somehow different, with parts of the R^2 latent for VAE empty. Does it come from different latent definition in both cases? Or is it just a result of showing only the [-20, 20]x[-20,20] square? This might not seem to be a fair comparison unless explained.\n4. From a practical point of view: what is the training and inference comparison between VAE (as well as other WAE, etc.) approach and the proposed ones?\n5.  What is the impact of C value and the batch size on quality, trading and inference time, etc.? Could you provide some ablation study?\n6. In conclusion the authors state, that the proposed model is actually composed of two parts: first detects attributes, and the second (i.e. classification?) is a probabilistic model which may be used for reasoning. A 1-D example shown in figure 3, that the authors refer to, shows that the first part performs a kind of clustering, is that so? Please elaborate on that, since it would greatly enlarge the readability of the paper.\n7. Several language errors should be corrected. E.g. a) on page 1 the sentence “However, IPNN need to predefine the …” should probably be “However, IPNN needs to be predefined…”; b) just above Eq. 1 instead “bellow” should be “below”; c) what is the word “complexer” at the bottom of page 3? Perhaps the authors meant to say “more complex? “Complexer” might be used in French. I would suggest checking the whole text with some native speaker. These errors are usually tiny, but disturb reading.\n8. Please, if possible, make the figures a bit larger, just to make them somehow more readable. This refers particularly to figures 1, 3, and perhaps 2 and 4 too.\n9. In section 7.1 you claim that the CIPNN tends to put 1, 4, 7, 9 MNIST numbers in one cluster — this is hardly visible in the figures. How is that model used for classification? Which inputs were used in each round? Could you elaborate on that a bit? \n10. Equations are sometimes complex (lots of variables and indices), e.g. Sequence from (9) to (14). Could you, please, make them easier to follow?\n11. Some small editing errors, e.g. subsection 7.2 title starts as an orphan.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698670677861}, {"id": "tUJo03RlgI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1578/Reviewer_7gbu"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces Continuous Indeterminate Probability Neural networks, which applies Indeterminate Probability Theory to define neural networks with latent variables for classification.\nThe authors also present a related auto-encoding variant of the model, which can be used to visualize the latent variable of the model.", "review_text": "The paper introduces Continuous Indeterminate Probability Neural networks, which applies Indeterminate Probability Theory to define neural networks with latent variables for classification.\nThe authors also present a related auto-encoding variant of the model, which can be used to visualize the latent variable of the model.", "strengths": "* The ideas presented in the paper are interesting and novel, to the best of my knowledge. These ideas could inspire future research.\n\n* Due to the usage of the latent variables in the CIPNN, the proposed model is less black-box than other architectures", "weaknesses": "**General comments**\n\nThere are 2 major issues with this paper, regarding clarity and experiments.\n\nCLARITY\n\nOverall, I found the paper hard to understand, mostly because it relies heavily on the unpublished work in (Anonymous, 2024), and assumes that the reader is knowledgeable of its content.\nWhile there is a high level description of (Anonymous, 2024) in section 2.2, this description is rushed and confusing (see detailed comments below).\n\nBeing a conference paper, this paper should instead be self-contained: the reader/reviewer should not have to read in full (Anonymous, 2024) to understand the proposed method (especially keeping in mind that the other paper could be rejected from the conference and be therefore unpublished if this work gets accepted).\nAs is, this paper looks more like an appendix to (Anonymous, 2024), rather a paper by itself. I suggest that the authors read and rewrite this work with the eyes of someone that knows nothing about (Anonymous, 2024).\n\nConsidering the classification/auto-encoding applications of Indeterminate Probability Theory, there are also several points in the paper that need to be clarified/improved.\n\nEXPERIMENTS\n\nThe experimental section is also quite confusing, and lacks proper baselines to understand the real performances of the model. \n\n\n\n**Detailed comments**\n\nBelow I describe the main points of confusion in each section.\n\n_Abstract_\n\nYou write \"pushed this classification capability to infinity\" -> what does this mean?\n\n_Introduction_\n\nThe motivation for this work in the introduction is based on the IPNN, which is however a model the reader knows nothing about at this point in the paper.\n\n_Section 2.1_\n\nYou write \"VAE uses neural network as the approximate solution of decoder\"\n -> What does this mean? In a VAE the decoder is defined as a modelling choice, and the encoder is used to approximate the posterior probability.\n\n_Section 2.2_\n\nOverall this section is hard to understand, and needs a better example/toy problem to help the reader (you could focus on the classification task from Figure 1 for example). \n\nWhen you say \"introducing Observers and treating the outcome of each random experiment as indeterminate probability distribution,\"\n- What are Observers? They are no longer mentioned in the rest of the section\n- how do you define an \"indeterminate\" probability distribution?\n\nThese definitions are missing:\n- what does $m$ represent in $y_m$\n- what does $l$ represent in $y_l$\n- what does $t$ represent in $x_t$\n\n \n_Section 3_\n\nWhy do you need to introduce both Observer 2 and Observer 3? What is the difference? Observer 2 seems not to be relevant for the subsequent discussion.\nDue to the confusion in Section 2.2, I am not really sure what you are trying to achieve in this section, and how exactly this relates to the rest of the paper.\n\n\n\n_Section 4.2_\n\n- Why did you choose that specific distribution in the right-hand side of the KL divergence?\n\n_Section 5_\n\nYou refer to details in (Anonymous, 2024) in the footnote, but they are needed in this paper as well to understand it.\n\n\n_Section 6_\n\n\"In this section, we will focus on the training strategy of Gaussian distribution\" -> \nCan you clarify what this means?\n\n\n_Section 7.1_\n\n1. This section misses baselines for other classification models (even simple neural networks)? The classification performances of your model on MNIST look quite poor for example.\n1. In Table 3 you compare against \"Simple-Softmax\", which is not defined, and which performs significantly better than the proposed model \n1. The advantages of this model vs other architectures are not well described\n1. What's the scalability of this method? What are the training times?\n1. The dataset names are not even mentioned in the main text, so one needs to guess which dataset the authors are talking about while reading this section. Only captions in the Figures mention the dataset.\n1. In the paragraph \"Results of classification tasks on large latent spaces\" - are you talking about table 2? It is not mentioned\n\n\n_Section 7.2_\n\n1. The difference between CIPAE and VAE is not clear from the paper\n1. \"As shown in Figure 5, the results of\nauto-encoder tasks between CIPAE\nand VAE are similar, this result further verifies that CIPAE is the analytical solution.\" -> What does this mean? Why can you make this statement from looking at a Figure?\n\n_Conclusion_\n\n\"Although our proposed model is derived from indeterminate probability theory, we can see Determinate\nfrom the expectation form in Eq. (11). Finally, we’d like to finish our paper with one sentence:\nThe world is determined with all Indeterminate!\" -> not sure what this means.", "questions": "See the questions in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Continuous Indeterminate Probability Neural networks, which applies Indeterminate Probability Theory to define neural networks with latent variables for classification.\nThe authors also present a related auto-encoding variant of the model, which can be used to visualize the latent variable of the model.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "* The ideas presented in the paper are interesting and novel, to the best of my knowledge. These ideas could inspire future research.\n\n* Due to the usage of the latent variables in the CIPNN, the proposed model is less black-box than other architectures", "weaknesses": "**General comments**\n\nThere are 2 major issues with this paper, regarding clarity and experiments.\n\nCLARITY\n\nOverall, I found the paper hard to understand, mostly because it relies heavily on the unpublished work in (Anonymous, 2024), and assumes that the reader is knowledgeable of its content.\nWhile there is a high level description of (Anonymous, 2024) in section 2.2, this description is rushed and confusing (see detailed comments below).\n\nBeing a conference paper, this paper should instead be self-contained: the reader/reviewer should not have to read in full (Anonymous, 2024) to understand the proposed method (especially keeping in mind that the other paper could be rejected from the conference and be therefore unpublished if this work gets accepted).\nAs is, this paper looks more like an appendix to (Anonymous, 2024), rather a paper by itself. I suggest that the authors read and rewrite this work with the eyes of someone that knows nothing about (Anonymous, 2024).\n\nConsidering the classification/auto-encoding applications of Indeterminate Probability Theory, there are also several points in the paper that need to be clarified/improved.\n\nEXPERIMENTS\n\nThe experimental section is also quite confusing, and lacks proper baselines to understand the real performances of the model. \n\n\n\n**Detailed comments**\n\nBelow I describe the main points of confusion in each section.\n\n_Abstract_\n\nYou write \"pushed this classification capability to infinity\" -> what does this mean?\n\n_Introduction_\n\nThe motivation for this work in the introduction is based on the IPNN, which is however a model the reader knows nothing about at this point in the paper.\n\n_Section 2.1_\n\nYou write \"VAE uses neural network as the approximate solution of decoder\"\n -> What does this mean? In a VAE the decoder is defined as a modelling choice, and the encoder is used to approximate the posterior probability.\n\n_Section 2.2_\n\nOverall this section is hard to understand, and needs a better example/toy problem to help the reader (you could focus on the classification task from Figure 1 for example). \n\nWhen you say \"introducing Observers and treating the outcome of each random experiment as indeterminate probability distribution,\"\n- What are Observers? They are no longer mentioned in the rest of the section\n- how do you define an \"indeterminate\" probability distribution?\n\nThese definitions are missing:\n- what does $m$ represent in $y_m$\n- what does $l$ represent in $y_l$\n- what does $t$ represent in $x_t$\n\n \n_Section 3_\n\nWhy do you need to introduce both Observer 2 and Observer 3? What is the difference? Observer 2 seems not to be relevant for the subsequent discussion.\nDue to the confusion in Section 2.2, I am not really sure what you are trying to achieve in this section, and how exactly this relates to the rest of the paper.\n\n\n\n_Section 4.2_\n\n- Why did you choose that specific distribution in the right-hand side of the KL divergence?\n\n_Section 5_\n\nYou refer to details in (Anonymous, 2024) in the footnote, but they are needed in this paper as well to understand it.\n\n\n_Section 6_\n\n\"In this section, we will focus on the training strategy of Gaussian distribution\" -> \nCan you clarify what this means?\n\n\n_Section 7.1_\n\n1. This section misses baselines for other classification models (even simple neural networks)? The classification performances of your model on MNIST look quite poor for example.\n1. In Table 3 you compare against \"Simple-Softmax\", which is not defined, and which performs significantly better than the proposed model \n1. The advantages of this model vs other architectures are not well described\n1. What's the scalability of this method? What are the training times?\n1. The dataset names are not even mentioned in the main text, so one needs to guess which dataset the authors are talking about while reading this section. Only captions in the Figures mention the dataset.\n1. In the paragraph \"Results of classification tasks on large latent spaces\" - are you talking about table 2? It is not mentioned\n\n\n_Section 7.2_\n\n1. The difference between CIPAE and VAE is not clear from the paper\n1. \"As shown in Figure 5, the results of\nauto-encoder tasks between CIPAE\nand VAE are similar, this result further verifies that CIPAE is the analytical solution.\" -> What does this mean? Why can you make this statement from looking at a Figure?\n\n_Conclusion_\n\n\"Although our proposed model is derived from indeterminate probability theory, we can see Determinate\nfrom the expectation form in Eq. (11). Finally, we’d like to finish our paper with one sentence:\nThe world is determined with all Indeterminate!\" -> not sure what this means.", "questions": "See the questions in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697905222794}], "openreview_url": "https://openreview.net/forum?id=Rt6btdXS2b", "arxiv_id": "2303.12964", "paper_pdf": "papers/Rt6btdXS2b.pdf", "paper_pdf_sha256": "c6560aef1007ed830b67529c45dafb14f4861b8e46d53571a12f27907c4bd2f8", "paper_pdf_bytes": 3261823, "paper_pdf_source": "openreview", "code_url": "https://github.com/Starfruit007/cipnn", "code_repository": "Starfruit007/cipnn", "code_commit": "e77f2062c3183d2cba0fdf620298f90cfbf49f17", "code_archive": "repos/Rt6btdXS2b.zip", "code_archive_sha256": "009015230b68621a796f9ad1f601a087f7408ba93e53f0221e149b77b8dc1e3a", "code_archive_bytes": 81969, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 180, "github_languages": {"Python": 52551, "Shell": 893}, "github_archived": false, "github_pushed_at": "2023-10-15T01:39:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continuous-indeterminate-probability-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "KVYq2Ea90PC", "year": 2022, "status": "rejected", "title": "A Study of Face Obfuscation in ImageNet", "authors": ["Kaiyu Yang", "Jacqueline Yau", "Li Fei-Fei", "Jia Deng", "Olga Russakovsky"], "authorids": ["~Kaiyu_Yang1", "~Jacqueline_Yau1", "~Li_Fei-Fei1", "~Jia_Deng1", "~Olga_Russakovsky1"], "authors_source": "OpenReview API", "abstract": "Face obfuscation (blurring, mosaicing, etc.) has been shown to be effective for privacy protection; nevertheless, object recognition research typically assumes access to complete, unobfuscated images. In this paper, we explore the effects of face obfuscation on the popular ImageNet challenge visual recognition benchmark. Most categories in the ImageNet challenge are not people categories; however, many incidental people appear in the images, and their privacy is a concern. We first annotate faces in the dataset. Then we demonstrate that face blurring and overlaying---two typical obfuscation techniques---have minimal impact on the accuracy of recognition models. Concretely, we benchmark multiple deep neural networks on face-obfuscated images and observe that the overall recognition accuracy drops only slightly (<= 1.0%). Further, we experiment with transfer learning to 4 downstream tasks (object recognition, scene recognition, face attribute classification, and object detection) and show that features learned on face-obfuscated images are equally transferable. Our work demonstrates the feasibility of privacy-aware visual recognition, improves the highly-used ImageNet challenge benchmark, and suggests an important path for future visual datasets. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "muaJz2tXtFF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper82/Reviewer_7itG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents an empirical study on the effect of face obfuscation in the ImageNet dataset. The main conclusion is that face obfuscation does not decrease the utility of the dataset. Specifically, the authors showed that various networks trained on the obfuscated dataset only experienced small accuracy drop on the image classification task. The authors also discussed the impact on different categories, showing that face obfuscation hurt more to the object categories that are more closely related to faces (i.e., the bounding boxes of which overlap more with faces). Last but now least, experiments has been conducted to show that face obfuscation also does not have a significant impact on the transferability of the features learned from the new dataset. All these conclusions are inline with intuitions since ImageNet is not primarily focused on human activities / faces.", "review_text": "Strengths:\n* The main contribution of this work is that it provided empirical evidence on the effect of face obfuscation on the ImageNet dataset. Through comprehensive experiment, the authors showed that face obfuscation does not decrease the utility of the dataset.\n* Another contribution that should not be overlooked is that the authors annotated all faces in ImageNet in a semi-automatic manner and promised that they will make the annotations publicly available to other researchers.\n* The paper is very well written and includes a lot of details on the experiment protocol. Thus It should be straightforward for other researchers to reproduce the results and to extend the study.\n\nWeaknesses:\n* Since blur and cutout are commonly used data augmentation techniques, it is to be expected that face obfuscation would not have a big impact to visions tasks that have little to do with faces. Although it is commendable that this is now shown empirically though the study, this work also does not bring interesting new insights into the topic.\n* The authors showed that categories that are closely related to faces are indeed affected more by face obfuscation. This paper would be more interesting (from a technical point of view) if the authors could additionally investigate into methods for alleviating such impact.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents an empirical study on the effect of face obfuscation in the ImageNet dataset. The main conclusion is that face obfuscation does not decrease the utility of the dataset. Specifically, the authors showed that various networks trained on the obfuscated dataset only experienced small accuracy drop on the image classification task. The authors also discussed the impact on different categories, showing that face obfuscation hurt more to the object categories that are more closely related to faces (i.e., the bounding boxes of which overlap more with faces). Last but now least, experiments has been conducted to show that face obfuscation also does not have a significant impact on the transferability of the features learned from the new dataset. All these conclusions are inline with intuitions since ImageNet is not primarily focused on human activities / faces.", "main_review": "Strengths:\n* The main contribution of this work is that it provided empirical evidence on the effect of face obfuscation on the ImageNet dataset. Through comprehensive experiment, the authors showed that face obfuscation does not decrease the utility of the dataset.\n* Another contribution that should not be overlooked is that the authors annotated all faces in ImageNet in a semi-automatic manner and promised that they will make the annotations publicly available to other researchers.\n* The paper is very well written and includes a lot of details on the experiment protocol. Thus It should be straightforward for other researchers to reproduce the results and to extend the study.\n\nWeaknesses:\n* Since blur and cutout are commonly used data augmentation techniques, it is to be expected that face obfuscation would not have a big impact to visions tasks that have little to do with faces. Although it is commendable that this is now shown empirically though the study, this work also does not bring interesting new insights into the topic.\n* The authors showed that categories that are closely related to faces are indeed affected more by face obfuscation. This paper would be more interesting (from a technical point of view) if the authors could additionally investigate into methods for alleviating such impact.", "summary_of_the_review": "This paper is very well written and it provides empirical evidences to support the intuition that face obfuscation does not decrease the utility of the ImageNet dataset. However, my main concern is that the paper has no technical novelty and it also does not bring sufficiently new insights to the community.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635986376213}, {"id": "Y-UbKxlTgxo", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper82/Reviewer_6ZXN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper mainly discuss the privacy issue for human for the widely used ImageNet dataset and how to handle them.\nIn addition, the paper does a very detailed empirical experiments to study the performance influence for various tasks, including object recognition, scene recognition, face attribute, and object detection if all the faces in the ImageNet are obfuscated.", "review_text": "The main strength of the paper is to address the privacy issues of ImageNet and is to provide an alternative face obfuscated version.\nIn addition, the authors conduct very thorough experiments across different tasks and different architecture for the study of the performance influence with the obfuscated dataset. It shows new dataset still is effective for transfer learning for various vision task with few performance drop. However, the weakness is that the main part of the paper is to examine the performance influence of different settings and is of limited technical novelty. For the verification of some downstream tasks, the coverage of tasks is not enough.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper mainly discuss the privacy issue for human for the widely used ImageNet dataset and how to handle them.\nIn addition, the paper does a very detailed empirical experiments to study the performance influence for various tasks, including object recognition, scene recognition, face attribute, and object detection if all the faces in the ImageNet are obfuscated.", "main_review": "The main strength of the paper is to address the privacy issues of ImageNet and is to provide an alternative face obfuscated version.\nIn addition, the authors conduct very thorough experiments across different tasks and different architecture for the study of the performance influence with the obfuscated dataset. It shows new dataset still is effective for transfer learning for various vision task with few performance drop. However, the weakness is that the main part of the paper is to examine the performance influence of different settings and is of limited technical novelty. For the verification of some downstream tasks, the coverage of tasks is not enough.", "summary_of_the_review": "Although the paper mainly focuses on providing plenty of empirical results to evaluate the influence of using the face obfuscated ImageNet dataset and is of limited novelty, it provides a lot of insights to show the effectiveness and feasibility of privacy preserving ImageNet.\n\nI have few concerns of the selection of transferring tasks. For example, the resolution of CIFAR-10 is only 32x32 and only for 10 classes. Similarly, Pascal VOC is also relatively small and easy dataset as compared with COCO or other recently released object detection dataset.\nMost of the images in the CelebA dataset are in frontal pose and have much fewer variations than other unconstrained face dataset, like IJB-C, etc. Since these datasets are relatively simpler than others which are more close to real-world scenarios, I wonder if the same experimental results and findings are still valid for harder datasets with more variations.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635918799824}, {"id": "6m-g0uRpct-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper82/Reviewer_U3FW"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The main concern addressed in this paper is the privacy problem that may result from images in ImageNet databases containing unexpected faces. The authors propose a two-step face filtering method. First, the authors use a detector called Amazon Rekognition to detect the ImageNet database. Then, the authors further optimize the detector output through the crowdsourcing platform Amazon Mechanical Turk (AMT) to reduce false positives and false negatives in automated detection. For the detected faces, the authors took two approaches, distinguishing between blurring and overlaying, and tested their effectiveness on different models separately. The accuracy of the two approaches was reduced by 0.9% on average compared to the original database on the ILSVRC classification challenge. And using the database that has blurred or covered the faces still maintains the transferability of the original database in the tests of downstream tasks.\nContribution：\n1. The authors perform a very time-consuming and labor-intensive task for accurate labeling and filtering of faces in the ImageNet database and statistical analysis of the classes of faces contained in ImageNet.\n2. The authors demonstrate experiments related to classification tasks and pre-trained model training using a database containing blurred or covered faces, proving that the theory is feasible and that the dropped accuracy is acceptable.\n3. In terms of ethics, using blurred or covered face data for training can reduce privacy concerns. The study of the ImageNet database in terms of privacy can provide an important reference for subsequent databases", "review_text": "The main weakness:\n1. I was confused by the statement in section (Part III, second paragraph) as to whether the face annotation process was manually filtered only for faces that were successfully detected (generated prediction boxes5) by Amazon Rekognition. Although I can infer from Appendix A Stage1 that only the data with successful detector predictions should be put into the AMT platform. Perhaps it would be better to include a brief description of Amazon Rekognition in the text. \n2. And if some faces are not detected by Amazon Rekognition, how do you tackle this problem? This is an important problem for the privacy issue in this paper because this paper focuses on the privacy issue.\n3. The novelty of this paper is limited reference to the proposed method in the paper.\n\nThe main strengths:\n1. The experimental part of the paper is depicted in great detail and completely. Rigorous validation of the obfuscation approach is done on two different methods on 15 different models. And in the appendix, the detailed method of blurring, and the problems that may happen in the process of labeling are under clearer explanation.\n2. The ethics of machine learning has been widely debated, with the issue of face privacy being of particular concern. There are many similar discussions, for example, there are some papers proposing to remove images associated with people from the database. The feasibility of obfuscation processing of faces proposed in this paper is a good way to minimize privacy issues without reducing the number of databases at the same time.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The main concern addressed in this paper is the privacy problem that may result from images in ImageNet databases containing unexpected faces. The authors propose a two-step face filtering method. First, the authors use a detector called Amazon Rekognition to detect the ImageNet database. Then, the authors further optimize the detector output through the crowdsourcing platform Amazon Mechanical Turk (AMT) to reduce false positives and false negatives in automated detection. For the detected faces, the authors took two approaches, distinguishing between blurring and overlaying, and tested their effectiveness on different models separately. The accuracy of the two approaches was reduced by 0.9% on average compared to the original database on the ILSVRC classification challenge. And using the database that has blurred or covered the faces still maintains the transferability of the original database in the tests of downstream tasks.\nContribution：\n1. The authors perform a very time-consuming and labor-intensive task for accurate labeling and filtering of faces in the ImageNet database and statistical analysis of the classes of faces contained in ImageNet.\n2. The authors demonstrate experiments related to classification tasks and pre-trained model training using a database containing blurred or covered faces, proving that the theory is feasible and that the dropped accuracy is acceptable.\n3. In terms of ethics, using blurred or covered face data for training can reduce privacy concerns. The study of the ImageNet database in terms of privacy can provide an important reference for subsequent databases", "main_review": "The main weakness:\n1. I was confused by the statement in section (Part III, second paragraph) as to whether the face annotation process was manually filtered only for faces that were successfully detected (generated prediction boxes5) by Amazon Rekognition. Although I can infer from Appendix A Stage1 that only the data with successful detector predictions should be put into the AMT platform. Perhaps it would be better to include a brief description of Amazon Rekognition in the text. \n2. And if some faces are not detected by Amazon Rekognition, how do you tackle this problem? This is an important problem for the privacy issue in this paper because this paper focuses on the privacy issue.\n3. The novelty of this paper is limited reference to the proposed method in the paper.\n\nThe main strengths:\n1. The experimental part of the paper is depicted in great detail and completely. Rigorous validation of the obfuscation approach is done on two different methods on 15 different models. And in the appendix, the detailed method of blurring, and the problems that may happen in the process of labeling are under clearer explanation.\n2. The ethics of machine learning has been widely debated, with the issue of face privacy being of particular concern. There are many similar discussions, for example, there are some papers proposing to remove images associated with people from the database. The feasibility of obfuscation processing of faces proposed in this paper is a good way to minimize privacy issues without reducing the number of databases at the same time.", "summary_of_the_review": "This paper has some contributions in exploring the ethicality of datasets, especially in the current very popular ImageNet database, but it exists some flaws (see weakness). The solutions and results in this paper are open sources and feasible, and this work will inspire subsequent exploration of privacy protection in publicly available datasets. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["Yes, Privacy, security and safety"], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635689628171}], "openreview_url": "https://openreview.net/forum?id=KVYq2Ea90PC", "arxiv_id": "2103.06191", "paper_pdf": "papers/KVYq2Ea90PC.pdf", "paper_pdf_sha256": "609078997326405e7d1f2738eb0a37506ab450d4b037387a2500e41cb2b2e92a", "paper_pdf_bytes": 4251919, "paper_pdf_source": "openreview", "code_url": "https://github.com/princetonvisualai/imagenet-face-obfuscation", "code_repository": "princetonvisualai/imagenet-face-obfuscation", "code_commit": "83f027962c41e11517d40f9c9f183621bedc4ed4", "code_archive": "repos/KVYq2Ea90PC.zip", "code_archive_sha256": "388e42614bcb64c2c1568442fbc3f20f325be9ef1b4e0a19943833c65dfff3e2", "code_archive_bytes": 1408607, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 1420, "github_languages": {"Python": 42871, "HTML": 32150}, "github_archived": false, "github_pushed_at": "2023-07-08T16:18:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-study-of-face-obfuscation-in-imagenet"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NfZ6g2OmXEk", "year": 2021, "status": "rejected", "title": "Prioritized Level Replay", "authors": ["Minqi Jiang", "Edward Grefenstette", "Tim Rocktäschel"], "authorids": ["~Minqi_Jiang1", "~Edward_Grefenstette1", "~Tim_Rocktäschel1"], "authors_source": "OpenReview API", "abstract": "Simulated environments with procedurally generated content have become popular benchmarks for testing systematic generalization of reinforcement learning agents. Every level in such an environment is algorithmically created, thereby exhibiting a unique configuration of underlying factors of variation, such as layout, positions of entities, asset appearances, or even the rules governing environment transitions. Fixed sets of training levels can be determined to aid comparison and reproducibility, and test levels can be held out to evaluate the generalization and robustness of agents. While prior work samples training levels in a direct way (e.g.~uniformly) for the agent to learn from, we investigate the hypothesis that different levels provide different learning progress for an agent at specific times during training. We introduce Prioritized Level Replay, a general framework for estimating the future learning potential of a level given the current state of the agent's policy. We find that temporal-difference (TD) errors, while previously used to selectively sample past transitions, also prove effective for scoring a level's future learning potential when the agent replays (that is, revisits) that level to generate entirely new episodes of experiences from it. We report significantly improved sample-efficiency and generalization on the majority of Procgen Benchmark environments as well as two challenging MiniGrid environments. Lastly, we present a qualitative analysis showing that Prioritized Level Replay induces an implicit curriculum, taking the agent gradually from easier to harder levels.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "prdz98BzGD4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper615/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**SUMMARY**\n\nThe present work considers the problem of learning in procedurally generated environments. This is a class of simulation environments in which each individual environment is created algorithmically where certain environmental factors are varied in each instance (referred to as levels in this work). Learning algorithms in this setting typically use a fixed set of training and evaluation environments. The present work proposes to sample the training environments such that the learning progress of the agent is optimized. This is achieved by proposing an algorithm for level prioritization during training. The performance of the approach is demonstrated on the Procgen Benchmark and two MiniGrid benchmarks and the authors argue that their approach induces an implicit curriculum in sparse reward settings.\n\n**STRENGTHS**\n- The general idea of prioritization for level sampling makes a lot of sense and is demonstrated to improve sample-efficiency for skill learning in procedurally generated environments.\n- I also liked that the authors compared with a big variety of different scoring metrics.\n\n**WEAKNESSES**\n- The intuition of \"greater discrepancy between expected and actual\nreturns, making \u000e$\\delta_t$ a useful measure of the learning potential\" makes sense. The heuristic score also works well in practice. One limitation I see is that there is no theoretical justification for why the TD-error is a good predictor for learnability. \n- This is maybe more an avenue for future work than an actual weakness but it seems to me that the algorithm is not making use of all potentially useful information. In each timestep, it only considers the last score achieved in a level. Maybe it would also be interesting to consider the full history of scores. My intuition is that levels in which agents were historically very slow to learn are maybe not as useful (or at least not useful at the moment). I.e., maybe in order to learn competing at such levels it is better to compete on other levels first?\n- Is there, at least from a qualitative perspective, an explanation for why certain environments do not benefit as much from the proposed level sampling approach?\n\n**REPRODUCIBILITY**\n\nThe work seems reproducible. Most of the information relevant for reproducibility is given in Appendices A & B. It would be great if the authors would also make the source code available.\n\n**CLARITY**\n\nOverall, I found the work to be very clearly written and have only minor questions/remarks:\n- To what extent does the use of TD-errors potentially limit the type of learning algorithms that can be used in the context of the proposed framework. Computing the TD-error requires a value function. As I understand it, some RL algorithms never compute a value function.\n- If I haven't overlooked it, there is no explanation of $c$ after eq. (4) while $C_i$ is explained earlier. Is $c$ simply the current episode?\n\n\n**EVALUATIONS**\n\nThe work is compared with several scoring function baselines using PPO. While the authors claim that the method is applicable to other RL agents, the evaluations do not show any results with other agent types. The authors mention several different benchmarks in that space. It would be interesting to know why particularly Procgen Benchmark and MiniGrid environments were chosen.\n\nIt is also not clearm to me why PPO is used as the base agent. Was this for ease of implementation / its popularity? Wouldn't it make sense to use more recent agents to see the added benefit of the proposed approach. E.g., would V-MPO be applicable here? \n\n**NOVELTY / RELEVANCE**\n\nThe work is very interesting and the authors make a compelling case that procedurally generated environments can benefit from a conscious sampling of the levels with regard to usefulness for learnability.\n\nI am not sure whether the claim \"Prioritized Level Replay induces an implicit curriculum, taking the agent gradually from easier to harder levels.\" is fully valid. As I understand it, the hardest levels are also the most likely to be sampled. The force counteracting this to some extent is the staleness-based sampling term $P_C$. For a gradual curriculum, I would expect $P_S$ to be designed such that it does not choose the hardest level but the one promising the best learning outcome. Particularly in the early stages of the training, the hard levels might be less useful than levels of medium difficulty.\n\n**SUMMARY**\n\nI found that paper very interesting. While I am not working in the particular subfield of the work and cannot sufficiently judge relation with prior works, I can confidently say that the idea and implementation details were conveyed very well. My main concerns are regarding the understanding of the \"failure cases\" and to what extent the graduality claim applies. That being said, I believe this line of work to be really interesting and to have a lot of potential for improved sample-efficiency when training RL agents in algorithmically generated simulation environments.\n\n**POST-DISCUSSION UPDATE**\n\nI want to thank the authors for correcting my misunderstandings, answering my questions, and providing additional material. As a consequence of this, I have raised my score to \"Accept\". To answer your question about what would be needed for a higher score: For a strong accept recommendation, I would have expected a mix of several additional things such as a clear impact outside of own subfield, code availability at time of submission (to evaluate how easy it is to reproduce the results and re-use the code), or more additional theoretical justification (in the sense of new formal guarantees for at least certain aspects of the proposed method). While not directly working in this subfield, I still think this work is solid and worthy of publication.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review [Updated]", "review": "**SUMMARY**\n\nThe present work considers the problem of learning in procedurally generated environments. This is a class of simulation environments in which each individual environment is created algorithmically where certain environmental factors are varied in each instance (referred to as levels in this work). Learning algorithms in this setting typically use a fixed set of training and evaluation environments. The present work proposes to sample the training environments such that the learning progress of the agent is optimized. This is achieved by proposing an algorithm for level prioritization during training. The performance of the approach is demonstrated on the Procgen Benchmark and two MiniGrid benchmarks and the authors argue that their approach induces an implicit curriculum in sparse reward settings.\n\n**STRENGTHS**\n- The general idea of prioritization for level sampling makes a lot of sense and is demonstrated to improve sample-efficiency for skill learning in procedurally generated environments.\n- I also liked that the authors compared with a big variety of different scoring metrics.\n\n**WEAKNESSES**\n- The intuition of \"greater discrepancy between expected and actual\nreturns, making \u000e$\\delta_t$ a useful measure of the learning potential\" makes sense. The heuristic score also works well in practice. One limitation I see is that there is no theoretical justification for why the TD-error is a good predictor for learnability. \n- This is maybe more an avenue for future work than an actual weakness but it seems to me that the algorithm is not making use of all potentially useful information. In each timestep, it only considers the last score achieved in a level. Maybe it would also be interesting to consider the full history of scores. My intuition is that levels in which agents were historically very slow to learn are maybe not as useful (or at least not useful at the moment). I.e., maybe in order to learn competing at such levels it is better to compete on other levels first?\n- Is there, at least from a qualitative perspective, an explanation for why certain environments do not benefit as much from the proposed level sampling approach?\n\n**REPRODUCIBILITY**\n\nThe work seems reproducible. Most of the information relevant for reproducibility is given in Appendices A & B. It would be great if the authors would also make the source code available.\n\n**CLARITY**\n\nOverall, I found the work to be very clearly written and have only minor questions/remarks:\n- To what extent does the use of TD-errors potentially limit the type of learning algorithms that can be used in the context of the proposed framework. Computing the TD-error requires a value function. As I understand it, some RL algorithms never compute a value function.\n- If I haven't overlooked it, there is no explanation of $c$ after eq. (4) while $C_i$ is explained earlier. Is $c$ simply the current episode?\n\n\n**EVALUATIONS**\n\nThe work is compared with several scoring function baselines using PPO. While the authors claim that the method is applicable to other RL agents, the evaluations do not show any results with other agent types. The authors mention several different benchmarks in that space. It would be interesting to know why particularly Procgen Benchmark and MiniGrid environments were chosen.\n\nIt is also not clearm to me why PPO is used as the base agent. Was this for ease of implementation / its popularity? Wouldn't it make sense to use more recent agents to see the added benefit of the proposed approach. E.g., would V-MPO be applicable here? \n\n**NOVELTY / RELEVANCE**\n\nThe work is very interesting and the authors make a compelling case that procedurally generated environments can benefit from a conscious sampling of the levels with regard to usefulness for learnability.\n\nI am not sure whether the claim \"Prioritized Level Replay induces an implicit curriculum, taking the agent gradually from easier to harder levels.\" is fully valid. As I understand it, the hardest levels are also the most likely to be sampled. The force counteracting this to some extent is the staleness-based sampling term $P_C$. For a gradual curriculum, I would expect $P_S$ to be designed such that it does not choose the hardest level but the one promising the best learning outcome. Particularly in the early stages of the training, the hard levels might be less useful than levels of medium difficulty.\n\n**SUMMARY**\n\nI found that paper very interesting. While I am not working in the particular subfield of the work and cannot sufficiently judge relation with prior works, I can confidently say that the idea and implementation details were conveyed very well. My main concerns are regarding the understanding of the \"failure cases\" and to what extent the graduality claim applies. That being said, I believe this line of work to be really interesting and to have a lot of potential for improved sample-efficiency when training RL agents in algorithmically generated simulation environments.\n\n**POST-DISCUSSION UPDATE**\n\nI want to thank the authors for correcting my misunderstandings, answering my questions, and providing additional material. As a consequence of this, I have raised my score to \"Accept\". To answer your question about what would be needed for a higher score: For a strong accept recommendation, I would have expected a mix of several additional things such as a clear impact outside of own subfield, code availability at time of submission (to evaluate how easy it is to reproduce the results and re-use the code), or more additional theoretical justification (in the sense of new formal guarantees for at least certain aspects of the proposed method). While not directly working in this subfield, I still think this work is solid and worthy of publication.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604364290985}, {"id": "KUSGE5Ix8S", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper615/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Paper Summary\n\nThis paper allows agents to set the initial conditions (level) for procedurally generated episodes during exploration to past observed values, and proposes to have agents form an intrinsic curriculum by resampling past levels based on a heuristic measure of expected learning progress. The authors test several heuristic measures and find that the average absolute magnitude of the generalized advantage estimate works well. The authors hypothesize that this intrinsic curriculum will improve optimization/learning relative to an agent that always samples initial conditions from the environment distribution. The authors verify that their prioritization strategy usually improves performance in several Progen Benchmark and MiniGrid environments, usually by a small but statistically significant amount, but sometimes by a large amount. \n\n### Summary Review (highlights re: quality, clarity, originality and significance)\n\nThe paper is well written and clear after one understands the basic idea. The idea is simple, and the algorithm/experiments seem straightforward to reimplement. The experiments are about what one would expect and seem to be well executed. The idea is original but not particularly innovative (this seems like the first heuristic prioritization approach that would come to mind given that the agent is able to choose the level). The improvement in empirical results is not particularly surprising (if anything, I would have expected more large improvements like the ones on bigfish/leaper environments). As this method is constrained to procedurally generated environments (or at least, evaluation of the method is constrained to procedurally generated environments), the significance seems rather limited. The required assumption seems rather strong, as it requires a simulator / control over the environment, which limits applicability.  \n\n### Pros\n\n- This a simple idea that can improve performance in Procedurally Generated Environments given that the agent is allowed to set the initial conditions / pick the level.\n- The performance improvement in 4 of the 19 environments tested is large & seems absolute (i.e., it's seems like a final performance improvement, not just a sample efficiency improvement). \n- The paper is well written/presented, easy to understand, and the empirical evaluation seems well done. The results do not seem difficult to replicate. \n\n### Cons\n\n- Despite being less intrusive than direct access to the level generation mechanism, the assumption that the agent can replay levels seems rather strong to me, and simplifies the task of learning procedurally generated environments very substantially.  \n\n    ($\\dagger$) I would argue that we don’t use procedurally generated environments as benchmarks in order to solve procedurally generated environments, but rather a tool for measuring generalization, so it's unclear to me that a technique that improves sample efficiency only in a procedurally generated environment is useful. \n\n    Unlike environment-agnostic techniques like prioritized replay, HER, intrinsic reward, intrinsic goal selection, etc., this requires you to have control over the environment, which seems to limit the applicability. If this is only useful with a simulator, then the small gains in sample efficiency aren’t actually that relevant, though this approach does seem to improve final performance in 4 of the 19 environments tested. \n- It’s not clear until the second page whether your method is a prioritized replay buffer scheme, or a task selection scheme. Actually, I was certain it was a prioritized replay buffer scheme until the second page, because that is the more natural/general setting (as noted above, I find the assumption that the agent can replay levels to be rather strong). \n- Several new hyperparameters are introduced; this said, guidance/ablations are performed, and it seems like the choices will generalize decently well (albeit there were different choices for ProcGen/Minigrid).\n\n### Questions / Etc.\n\n- My main question for the authors is to ask for a counterargument to ($\\dagger$) above. \n- It would be good if this can be shown to work in multi-goal setting, as it is quite similar to ProcGen setting... you draw some distinctions, but I do think your approach would be applicable there. \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Well done paper, but unclear significance / potentially limited applicability", "review": "### Paper Summary\n\nThis paper allows agents to set the initial conditions (level) for procedurally generated episodes during exploration to past observed values, and proposes to have agents form an intrinsic curriculum by resampling past levels based on a heuristic measure of expected learning progress. The authors test several heuristic measures and find that the average absolute magnitude of the generalized advantage estimate works well. The authors hypothesize that this intrinsic curriculum will improve optimization/learning relative to an agent that always samples initial conditions from the environment distribution. The authors verify that their prioritization strategy usually improves performance in several Progen Benchmark and MiniGrid environments, usually by a small but statistically significant amount, but sometimes by a large amount. \n\n### Summary Review (highlights re: quality, clarity, originality and significance)\n\nThe paper is well written and clear after one understands the basic idea. The idea is simple, and the algorithm/experiments seem straightforward to reimplement. The experiments are about what one would expect and seem to be well executed. The idea is original but not particularly innovative (this seems like the first heuristic prioritization approach that would come to mind given that the agent is able to choose the level). The improvement in empirical results is not particularly surprising (if anything, I would have expected more large improvements like the ones on bigfish/leaper environments). As this method is constrained to procedurally generated environments (or at least, evaluation of the method is constrained to procedurally generated environments), the significance seems rather limited. The required assumption seems rather strong, as it requires a simulator / control over the environment, which limits applicability.  \n\n### Pros\n\n- This a simple idea that can improve performance in Procedurally Generated Environments given that the agent is allowed to set the initial conditions / pick the level.\n- The performance improvement in 4 of the 19 environments tested is large & seems absolute (i.e., it's seems like a final performance improvement, not just a sample efficiency improvement). \n- The paper is well written/presented, easy to understand, and the empirical evaluation seems well done. The results do not seem difficult to replicate. \n\n### Cons\n\n- Despite being less intrusive than direct access to the level generation mechanism, the assumption that the agent can replay levels seems rather strong to me, and simplifies the task of learning procedurally generated environments very substantially.  \n\n    ($\\dagger$) I would argue that we don’t use procedurally generated environments as benchmarks in order to solve procedurally generated environments, but rather a tool for measuring generalization, so it's unclear to me that a technique that improves sample efficiency only in a procedurally generated environment is useful. \n\n    Unlike environment-agnostic techniques like prioritized replay, HER, intrinsic reward, intrinsic goal selection, etc., this requires you to have control over the environment, which seems to limit the applicability. If this is only useful with a simulator, then the small gains in sample efficiency aren’t actually that relevant, though this approach does seem to improve final performance in 4 of the 19 environments tested. \n- It’s not clear until the second page whether your method is a prioritized replay buffer scheme, or a task selection scheme. Actually, I was certain it was a prioritized replay buffer scheme until the second page, because that is the more natural/general setting (as noted above, I find the assumption that the agent can replay levels to be rather strong). \n- Several new hyperparameters are introduced; this said, guidance/ablations are performed, and it seems like the choices will generalize decently well (albeit there were different choices for ProcGen/Minigrid).\n\n### Questions / Etc.\n\n- My main question for the authors is to ask for a counterargument to ($\\dagger$) above. \n- It would be good if this can be shown to work in multi-goal setting, as it is quite similar to ProcGen setting... you draw some distinctions, but I do think your approach would be applicable there. \n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603848821244}, {"id": "WTfKFzqEG3u", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper615/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n\nThis paper concerns about the use of experience replay in a way that past experience is sampled based on (implicit) levels so as for the agent to better adapt to the current task at hand. The authors defined a replay distribution (where experience is sampled) based on two scores relevant to learning potential and staleness. Due to its formulation, the change of replay distribution can be used as an outer-layer of a learning algorithm without any modification of the underlying learning mode. The authors conducted experiments over a set of benchmark data sets relevant to level-ness and found statistically significant improvements over more than half of the tasks.\n\nThe overall impression of the paper is that it presents a simple yet effective solution to prioritizing experience in the presence of level-ness in a given task. The basic idea is finding out past experience with high \"learning potential\" by examining a past trajectory's 'wrongness' and how long the policy was not updated (= likely still wrong). \n\nPoint: The notion of level and its relevance to learning potential.\nFirst, the paper does not contain any mathematical (or clear) definition of level, which should be crucial to understand the paper. At the beginning it is only explained as different configurations (i.e., any non-singleton environment). Further, it is hard to understand why the notion of levels is even needed to be employed in the paper. An RL agent has a specific way to learn experience (updating parameters) and its artifacts makes \"experience replay\" useful in most RL agents. Then, there would be an optimal way of replaying experience at any given time --- a certain order of a subset of past trajectories to be replayed for a current policy. The current form of P_replay (Eq. 1) does not need any specific notion of level-ness but only 'learning potential'. I suspect that Eq. 1 also works for a singleton environment, which the authors excluded from consideration.\n\nPoint: The conjecture about curriculum learning. \nIt is reasonable to assume that the notion of hardness of a task for an RL agent is the difficulty of optimizing its policy (=resulting in higher TD-errors). When the human understanding of easiness of a task (i.e., level) matches the agent's ability to optimize, we would safely say that PLR induces curriculum implicitly. It is nice to see such plots (Figure 4) that empirically validate the conjecture. However, isn't it a much anticipated result?\n\nQuestions\nQ1. Would different algorithms other than (PPO + GAE) make the results different from the current form? \n\nQ2: If we interpret P_S and P_C as two probability distributions, multiplying them seems more natural to me. What is the rationale behind for adding them not multiplying them (or use (1-rho) log P_S + (rho) log P_C)? Further, any reason for P_C being proportional to c-C_i?\n\nQ2. How about learning hyper-parameters on the fly? Both \\beta and \\rho might be adjusted throughout learning. \nFurther, it is conceivable that the optimal \\beta and \\rho are not fixed quantities but can be dependent to a given pair of policy and trajectory. \n\nMinor\nFigure 1, there are two taus. The top would be \\pi?\nBackground \"We to refer to\"\n=======\nI read through all the reviews and rebuttals and I could better understand and evaluate the paper. I updated my score to 7. \n\nGiven that this replay scheme works fairly well (intuitively, empirically), easy to understand and implement, fairly sufficient amount of empirical experimentation, I would like to see the paper accepted (and adopted and improved by others).\n\nOne more comment about staleness.\nI think staleness is a proxy measure for the (unmeasured) score of the 'current' policy on that level. So I would like to see (in future or revised version) some experiments that measure how well staleness measure correlate with such score. Further, the way staleness is designed properly reflects how the score degrades as the level isn't played.\n\nSome idea.\nIt would be nice to make a connection to multi-task learning where tasks share some similarities. Currently, level is somewhat 'linearly' defined. If an agent plays level x, then staleness for level x' (something similar to x') doesn't have to be updated a lot compared to another task which might be dissimilar to level x. Hence, some similarity measure can be further employed (or learn a metric).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "review": "\n\nThis paper concerns about the use of experience replay in a way that past experience is sampled based on (implicit) levels so as for the agent to better adapt to the current task at hand. The authors defined a replay distribution (where experience is sampled) based on two scores relevant to learning potential and staleness. Due to its formulation, the change of replay distribution can be used as an outer-layer of a learning algorithm without any modification of the underlying learning mode. The authors conducted experiments over a set of benchmark data sets relevant to level-ness and found statistically significant improvements over more than half of the tasks.\n\nThe overall impression of the paper is that it presents a simple yet effective solution to prioritizing experience in the presence of level-ness in a given task. The basic idea is finding out past experience with high \"learning potential\" by examining a past trajectory's 'wrongness' and how long the policy was not updated (= likely still wrong). \n\nPoint: The notion of level and its relevance to learning potential.\nFirst, the paper does not contain any mathematical (or clear) definition of level, which should be crucial to understand the paper. At the beginning it is only explained as different configurations (i.e., any non-singleton environment). Further, it is hard to understand why the notion of levels is even needed to be employed in the paper. An RL agent has a specific way to learn experience (updating parameters) and its artifacts makes \"experience replay\" useful in most RL agents. Then, there would be an optimal way of replaying experience at any given time --- a certain order of a subset of past trajectories to be replayed for a current policy. The current form of P_replay (Eq. 1) does not need any specific notion of level-ness but only 'learning potential'. I suspect that Eq. 1 also works for a singleton environment, which the authors excluded from consideration.\n\nPoint: The conjecture about curriculum learning. \nIt is reasonable to assume that the notion of hardness of a task for an RL agent is the difficulty of optimizing its policy (=resulting in higher TD-errors). When the human understanding of easiness of a task (i.e., level) matches the agent's ability to optimize, we would safely say that PLR induces curriculum implicitly. It is nice to see such plots (Figure 4) that empirically validate the conjecture. However, isn't it a much anticipated result?\n\nQuestions\nQ1. Would different algorithms other than (PPO + GAE) make the results different from the current form? \n\nQ2: If we interpret P_S and P_C as two probability distributions, multiplying them seems more natural to me. What is the rationale behind for adding them not multiplying them (or use (1-rho) log P_S + (rho) log P_C)? Further, any reason for P_C being proportional to c-C_i?\n\nQ2. How about learning hyper-parameters on the fly? Both \\beta and \\rho might be adjusted throughout learning. \nFurther, it is conceivable that the optimal \\beta and \\rho are not fixed quantities but can be dependent to a given pair of policy and trajectory. \n\nMinor\nFigure 1, there are two taus. The top would be \\pi?\nBackground \"We to refer to\"\n=======\nI read through all the reviews and rebuttals and I could better understand and evaluate the paper. I updated my score to 7. \n\nGiven that this replay scheme works fairly well (intuitively, empirically), easy to understand and implement, fairly sufficient amount of empirical experimentation, I would like to see the paper accepted (and adopted and improved by others).\n\nOne more comment about staleness.\nI think staleness is a proxy measure for the (unmeasured) score of the 'current' policy on that level. So I would like to see (in future or revised version) some experiments that measure how well staleness measure correlate with such score. Further, the way staleness is designed properly reflects how the score degrades as the level isn't played.\n\nSome idea.\nIt would be nice to make a connection to multi-task learning where tasks share some similarities. Currently, level is somewhat 'linearly' defined. If an agent plays level x, then staleness for level x' (something similar to x') doesn't have to be updated a lot compared to another task which might be dissimilar to level x. Hence, some similarity measure can be further employed (or learn a metric).", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603678018378}, {"id": "rmojgdtSgyc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper615/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\n**Summary**:\n\nThis paper proposes a prioritized sampling strategy for task sampling in procedurally generated environments. While training an RL agent across many tasks (levels), we can either sample a new task uniformly from the training task distribution or sample a new task with different weights. The paper claims that sampling based on the average magnitude of generalized advantage estimate (GAE) yields faster learning in most Procgen environments and a few MiniGrid environments. Overall, I found the idea to be simple and intuitive. But the benefit of using prioritized level replay is also not very consistent across different environments used in the paper. \n##########################################################################\n\n**Strengths**:\n\nThe method of the paper is simple and can be incorporated into many existing RL algorithms.\n\nThe paper shows that L1 value loss is a good scoring metric for the prioritization by comparing several different choices.\n\n\n##########################################################################\n\n**Weaknesses**:\n\nThe advantage of using prioritized level replay against uniform sampling is rather small in many tasks (11 out of 19 tasks) shown in the paper (Climber, Coinrun, Dodgeball, Fruitbot, Heist, Jumper, Maze, Miner, Ninja, Starpilot, ObstructedMazeGamut-Medium).\n\n\nThe paper only presents results in the easy mode of procgen. While I understand the reason due to the limit on the computational resources, it would be more convincing to show the results on at least 1 or 2 procgen tasks in the difficult mode. If the overall task difficulty is increased, then the advantage of learning in a curriculum (starting from the easy tasks and then to the difficult tasks) are expected to be more salient.\n\n\nWhile the scoring metrics used in the paper are all related to the policy function or value function that is being learned, how about a scoring metric that is only based on the number of steps that the agent experiences in a task and whether the agent fails or succeeds? Intuitively, if the lifetime of an agent is short and the agent solves the task, it is an easy task. If the agent does not solve the task or it takes the agent many more steps to solve the task, it is a difficult task. Another metric to compare to is prioritize based on the return value of the trajectories. If the return value is high, then the task is probably already solved by the current policy, so we can sample such tasks less frequently.\n\nIn Figure 4, it seems the advantage of using L1 value loss for the prioritization in sampling is more obvious in easy environments (Multiroom-N4-Random and ObstructedMazeGamut-Easy). But its performance becomes very close to the uniform sampling strategy in harder environments (ObstructedMazeGamut-Medium). Why would the advantage of using prioritization (hence implicit curriculum) fade as the task difficulty increases?\n\nIn Figure 4, it is hard to connect the top row to the bottom row as the top row uses the environment steps for the x-axis, the bottom row uses the number of PPO updates for the y-axis. I would suggest plot the bottom row figures in terms of the environment steps as well and use the same x-range.\n\n\n##########################################################################\n\n**Minor points**:\n\nSome details about the experiment setup, especially the MiniGrid environments, are missing. For example, how do the MiniGrid environments look like, what does the difficulty mean in these environments, which parts of the environments are randomized across levels, reward structure, etc.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2 ", "review": "##########################################################################\n\n**Summary**:\n\nThis paper proposes a prioritized sampling strategy for task sampling in procedurally generated environments. While training an RL agent across many tasks (levels), we can either sample a new task uniformly from the training task distribution or sample a new task with different weights. The paper claims that sampling based on the average magnitude of generalized advantage estimate (GAE) yields faster learning in most Procgen environments and a few MiniGrid environments. Overall, I found the idea to be simple and intuitive. But the benefit of using prioritized level replay is also not very consistent across different environments used in the paper. \n##########################################################################\n\n**Strengths**:\n\nThe method of the paper is simple and can be incorporated into many existing RL algorithms.\n\nThe paper shows that L1 value loss is a good scoring metric for the prioritization by comparing several different choices.\n\n\n##########################################################################\n\n**Weaknesses**:\n\nThe advantage of using prioritized level replay against uniform sampling is rather small in many tasks (11 out of 19 tasks) shown in the paper (Climber, Coinrun, Dodgeball, Fruitbot, Heist, Jumper, Maze, Miner, Ninja, Starpilot, ObstructedMazeGamut-Medium).\n\n\nThe paper only presents results in the easy mode of procgen. While I understand the reason due to the limit on the computational resources, it would be more convincing to show the results on at least 1 or 2 procgen tasks in the difficult mode. If the overall task difficulty is increased, then the advantage of learning in a curriculum (starting from the easy tasks and then to the difficult tasks) are expected to be more salient.\n\n\nWhile the scoring metrics used in the paper are all related to the policy function or value function that is being learned, how about a scoring metric that is only based on the number of steps that the agent experiences in a task and whether the agent fails or succeeds? Intuitively, if the lifetime of an agent is short and the agent solves the task, it is an easy task. If the agent does not solve the task or it takes the agent many more steps to solve the task, it is a difficult task. Another metric to compare to is prioritize based on the return value of the trajectories. If the return value is high, then the task is probably already solved by the current policy, so we can sample such tasks less frequently.\n\nIn Figure 4, it seems the advantage of using L1 value loss for the prioritization in sampling is more obvious in easy environments (Multiroom-N4-Random and ObstructedMazeGamut-Easy). But its performance becomes very close to the uniform sampling strategy in harder environments (ObstructedMazeGamut-Medium). Why would the advantage of using prioritization (hence implicit curriculum) fade as the task difficulty increases?\n\nIn Figure 4, it is hard to connect the top row to the bottom row as the top row uses the environment steps for the x-axis, the bottom row uses the number of PPO updates for the y-axis. I would suggest plot the bottom row figures in terms of the environment steps as well and use the same x-range.\n\n\n##########################################################################\n\n**Minor points**:\n\nSome details about the experiment setup, especially the MiniGrid environments, are missing. For example, how do the MiniGrid environments look like, what does the difficulty mean in these environments, which parts of the environments are randomized across levels, reward structure, etc.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603065238933}], "openreview_url": "https://openreview.net/forum?id=NfZ6g2OmXEk", "arxiv_id": "2010.03934", "paper_pdf": "papers/NfZ6g2OmXEk.pdf", "paper_pdf_sha256": "076d72332256d46f38873fa3e1f3b03f172dbd302a887a3306246a863aa83277", "paper_pdf_bytes": 5922362, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/level-replay", "code_repository": "facebookresearch/level-replay", "code_commit": "ccecf452ee3342217ece964aaf10c2831625f9b3", "code_archive": "repos/NfZ6g2OmXEk.zip", "code_archive_sha256": "fd9f0ed1624113f9dbce62d1255eddab9b2dec52025cb70d4d96847927d9d255", "code_archive_bytes": 1201087, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 1168, "github_languages": {"Python": 98573}, "github_archived": true, "github_pushed_at": "2021-06-11T01:47:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/prioritized-level-replay-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HJlU-AVtvS", "year": 2020, "status": "rejected", "title": "A Fine-Grained Spectral Perspective on Neural Networks", "authors": ["Greg Yang", "Hadi Salman"], "authorids": ["gregyang@microsoft.com", "hadicsalman@gmail.com"], "authors_source": "OpenReview API", "abstract": "Are neural networks biased toward simple functions?\nDoes depth always help learn more complex features?\nIs training the last layer of a network as good as training all layers?\nThese questions seem unrelated at face value, but in this work we give all of them a common treatment from the spectral perspective.\nWe will study the spectra of the *Conjugate Kernel, CK,* (also called the *Neural Network-Gaussian Process Kernel*), and the *Neural Tangent Kernel, NTK*.\nRoughly, the CK and the NTK tell us respectively ``\"what a network looks like at initialization\" and \"``what a network looks like during and after training.\"\nTheir spectra then encode valuable information about the initial distribution and the training and generalization properties of neural networks.\nBy analyzing the eigenvalues, we lend novel insights into the questions put forth at the beginning, and we verify these insights by extensive experiments of neural networks.\nWe believe the computational tools we develop here for analyzing the spectra of CK and NTK serve as a solid foundation for future studies of deep neural networks.\nWe have open-sourced the code for it and for generating the plots in this paper at github.com/jxVmnLgedVwv6mNcGCBy/NNspectra.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "S1xeMeCk5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper965/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper examined the spectrum of NNGP and NTK kernels and answer several questions about deep networks using both analytical results and experimental evidence:\n* Are randomly initialized and trained deep networks biased to simple functions?\n* How does this change with depth, activation function, and initialization?\n\nAll studies are conducted on a space of inputs that is a boolean cube. The input distribution is assumed to be uniform. Though it is argued in Section 3 that the results also generalize to uniform distributions on spheres and isotropic Gaussian distributions. Although this boolean cube setting is followed from previous works on the same topic, it does limit the scope of the paper. Discussions on how this assumption relates to practical problems are missing from the paper.\n\nPutting aside the limitations of restricting the input distributions on boolean cubes (and other similar choices), I really like the paper, which demonstrates the powerfulness of spectral analysis. I also found that many analytical results (e.g., computing eigenvalues of a kernel operator with respect to uniform distributions on a boolean cube) in the paper are highly nontrivial to derive, which adds to the value of the paper. These results might seem restricted in terms of deep network theory because of the assumptions on input distributions, but I do believe the methods used can be of interest to a wider audience.\n\nSome questions:\n* In Figure 1, the 10^4 boolean function samples are sorted according to frequency (rank). What precisely is the frequency (rank) here? It shouldn't be the frequency that corresponds to the eigendecomposition because each function sample could always have multiple components with different frequencies.\n* In Figure 1b, the y-axis is described as normalized eigenvalues, which seems different from degree k fractional variance defined in the next section. The degree k fractional variance is the sum of all normalized eigenvalues for degree k eigenfunctions. Is this difference intended or it is a mistake?\n* Is the ground truth degree k polynomial used in experiments defined somewhere in the paper?\n\nOn writing and clarity. Overall I find this paper well-written and a pleasure to read. Some minor issues are\n* The definition of \"neural kernels\" seems unnecessary and a bit sudden. It would be helpful to include the definition of Phi just after Eq. (2) for CK and NTK.\n* For introducing boolean analysis and Fourier series, it might be better to include the formula that explicit shows the expansion f(x) = \\sum_{S} f^p(S) X_S(x) before introducing Theorem 3.1.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper examined the spectrum of NNGP and NTK kernels and answer several questions about deep networks using both analytical results and experimental evidence:\n* Are randomly initialized and trained deep networks biased to simple functions?\n* How does this change with depth, activation function, and initialization?\n\nAll studies are conducted on a space of inputs that is a boolean cube. The input distribution is assumed to be uniform. Though it is argued in Section 3 that the results also generalize to uniform distributions on spheres and isotropic Gaussian distributions. Although this boolean cube setting is followed from previous works on the same topic, it does limit the scope of the paper. Discussions on how this assumption relates to practical problems are missing from the paper.\n\nPutting aside the limitations of restricting the input distributions on boolean cubes (and other similar choices), I really like the paper, which demonstrates the powerfulness of spectral analysis. I also found that many analytical results (e.g., computing eigenvalues of a kernel operator with respect to uniform distributions on a boolean cube) in the paper are highly nontrivial to derive, which adds to the value of the paper. These results might seem restricted in terms of deep network theory because of the assumptions on input distributions, but I do believe the methods used can be of interest to a wider audience.\n\nSome questions:\n* In Figure 1, the 10^4 boolean function samples are sorted according to frequency (rank). What precisely is the frequency (rank) here? It shouldn't be the frequency that corresponds to the eigendecomposition because each function sample could always have multiple components with different frequencies.\n* In Figure 1b, the y-axis is described as normalized eigenvalues, which seems different from degree k fractional variance defined in the next section. The degree k fractional variance is the sum of all normalized eigenvalues for degree k eigenfunctions. Is this difference intended or it is a mistake?\n* Is the ground truth degree k polynomial used in experiments defined somewhere in the paper?\n\nOn writing and clarity. Overall I find this paper well-written and a pleasure to read. Some minor issues are\n* The definition of \"neural kernels\" seems unnecessary and a bit sudden. It would be helpful to include the definition of Phi just after Eq. (2) for CK and NTK.\n* For introducing boolean analysis and Fourier series, it might be better to include the formula that explicit shows the expansion f(x) = \\sum_{S} f^p(S) X_S(x) before introducing Theorem 3.1.\n"}, "tcdate": 1571966983710}, {"id": "B1gapNDTFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper965/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Aiming to resolve the question whether and why deep networks are biased towards simple functions, this paper gives a spectral analysis on neural networks' conjugate kernel(CK) and neural tangent kernel(NTK) on boolean cube. The eigenfunctions are identified and the eigenvalues are shown computable in polynomial time. Another main contribution of this paper is showing that the simplicity bias exists at least in a weak sense.\n\nI believe that this paper should be weakly rejected because it made more claims than what it can show in that the analysis doesn't work in real space, and the authors did not really show the simplicity bias. The following are my detailed comments.\n\nFirst, the whole analysis is based on boolean cube. Although the paper has shown empirically that in high dimension the uniform binary distribution is close enough to the uniform sphere distribution, it doesn't suffice to substitute boolean cube for sphere in real space. The spectral analysis in this paper is heavily due to working on boolean cube. The boolean cube is finite, which guarantees any inner-product kernel function $K(x,y) = \\Phi (<x,y>)$ can be diagonalized by finite many monomial functions, And there are only O(d) different eigenvalues, which enables efficient computation. These techniques are not easy to be transferred to real space. The experiment shows that the first five eigenvalues in boolean cube, sphere, gaussian is close, but key problems here are first, in practive the dimension $d$ could be smaller and second, sphere and gaussian have infinitely many eigenvalues while boolean cube has $2^d$ eigenvalues. The experiment cannot really justify that all eigenvalues are close (only first several are shown), not to mention the tail eigenvalues over the first $2^d$-th.\n\nEven if we assume that boolean cube is a reasonable choice, we should notice the goal of computing eigenvalues is to eventually show the inductive bias toward 'simple functions'. However, the authors failed to show it at least from the following perspectives:\n1) This paper did not show the trend of eigenvalues, but only the weak version of, for example, $\\mu_{2k-2} > \\mu_{2k}$. In the limiting case, it is more reasonable to fix dimension $d$ rather than the degree $k$.\n2) Working on boolean cube leads to limited complexity. The most complicated base function is restricted to $\\mathcal{X}_S$ where $S = \\{1, 2, \\dots, d\\}$. So the weak simplicity bias theorem actually only describes the relation among finite $d$ eigenvalues.\n3) No optimization arguments appear in this paper. Based on the spectral analysis, it is not rigorous enough to claim the networks are biased to simple functions, given that the target function consists of simple multilinear monomial functions. \n\nSince the boolean spectra is not a reliable measure, the further experiments under such a measure is therefore put under doubt.\n\nTo summarize, this paper definitely contains some rigorous analysis which I appreciate, but it made some claims that are not verified. More importantly, the boolean cube is not the appropriate domain which is hard to generalize to real space and the simplicity bias theorem in this paper is to some extent weak. Therefore, I suggest rejecting this paper in its current form.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "Aiming to resolve the question whether and why deep networks are biased towards simple functions, this paper gives a spectral analysis on neural networks' conjugate kernel(CK) and neural tangent kernel(NTK) on boolean cube. The eigenfunctions are identified and the eigenvalues are shown computable in polynomial time. Another main contribution of this paper is showing that the simplicity bias exists at least in a weak sense.\n\nI believe that this paper should be weakly rejected because it made more claims than what it can show in that the analysis doesn't work in real space, and the authors did not really show the simplicity bias. The following are my detailed comments.\n\nFirst, the whole analysis is based on boolean cube. Although the paper has shown empirically that in high dimension the uniform binary distribution is close enough to the uniform sphere distribution, it doesn't suffice to substitute boolean cube for sphere in real space. The spectral analysis in this paper is heavily due to working on boolean cube. The boolean cube is finite, which guarantees any inner-product kernel function $K(x,y) = \\Phi (<x,y>)$ can be diagonalized by finite many monomial functions, And there are only O(d) different eigenvalues, which enables efficient computation. These techniques are not easy to be transferred to real space. The experiment shows that the first five eigenvalues in boolean cube, sphere, gaussian is close, but key problems here are first, in practive the dimension $d$ could be smaller and second, sphere and gaussian have infinitely many eigenvalues while boolean cube has $2^d$ eigenvalues. The experiment cannot really justify that all eigenvalues are close (only first several are shown), not to mention the tail eigenvalues over the first $2^d$-th.\n\nEven if we assume that boolean cube is a reasonable choice, we should notice the goal of computing eigenvalues is to eventually show the inductive bias toward 'simple functions'. However, the authors failed to show it at least from the following perspectives:\n1) This paper did not show the trend of eigenvalues, but only the weak version of, for example, $\\mu_{2k-2} > \\mu_{2k}$. In the limiting case, it is more reasonable to fix dimension $d$ rather than the degree $k$.\n2) Working on boolean cube leads to limited complexity. The most complicated base function is restricted to $\\mathcal{X}_S$ where $S = \\{1, 2, \\dots, d\\}$. So the weak simplicity bias theorem actually only describes the relation among finite $d$ eigenvalues.\n3) No optimization arguments appear in this paper. Based on the spectral analysis, it is not rigorous enough to claim the networks are biased to simple functions, given that the target function consists of simple multilinear monomial functions. \n\nSince the boolean spectra is not a reliable measure, the further experiments under such a measure is therefore put under doubt.\n\nTo summarize, this paper definitely contains some rigorous analysis which I appreciate, but it made some claims that are not verified. More importantly, the boolean cube is not the appropriate domain which is hard to generalize to real space and the simplicity bias theorem in this paper is to some extent weak. Therefore, I suggest rejecting this paper in its current form."}, "tcdate": 1571808453104}, {"id": "SyeX2Lk6KH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper965/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Updates:\n\nThanks for the updates. \n\nI find the new theoretical results interesting and potentially useful,  which shows, in the large $d$ setting, spectrums of CKs/NTKs  for boolean cube, sphere and isotropic Gaussian are closed to each other in some sense. Thus, I raise my score to weakly accepted but lower down my confidence level since I am not that familiar with Boolean cube literature. \n\n\n------------------------------------------------------\nThe study of extremely over-parameterized networks (i.e. infinitely width networks) has become one of the most active research directions in theory deep learning. The key objects in understanding such networks are the conjugate kernel [1, 2] (CK defined in the paper) and the Neural tangent kernels [3] (NTK). The CK characterizes how the network looks like at initialization (connection to Gaussian processes as well) and the NTK is very useful to characterize the gradient descent training dynamics of large width networks in the kernel regime. Understanding properties of such kernels, in particular, their spectra distribution and eigenspace, could be potentially an important step towards a finer-gained understanding of generalization in neural networks.   \n\nThe main contribution of this paper is the development of the spectral theory of CK and NTK on boolean cube (similar or weaker results on uniform distribution in spheres and Gaussian distribution in R^n). More precisely, the authors show that, over the space of boolean cube, the CK/NTK could be diagonalized using the Fourier basis and the eigenvalues depend only on the frequency (i.e. the degree of the monomials); Thm 3.1. The authors also develop some computation tools to compute the spectra; Lemma 3.2.   Using the tools developed in this paper, the authors are able to clarify some of the interesting observations found by other researchers. Most noticeably, the authors show that the observation in [4] 'neural network is biased towards simple functions' is NOT universal. Whether this statement is correct or not depends heavily on the choice of activation function (e.g. Relu v.s. Erf) and hyper-parameters (e.g. weight variance, depths).  There are also some other interesting empirical findings: the optimal depth of a neural network depends on the complexity (i.e. degree in the boolean cube setting) of the function to learn, CK (i.e. training only the last layer) tends to be more useful for learning less complex functions, etc. \n\nOverall, this is a nice paper. I am leaning for a weakly accept. \n\n\n[1] Amit Daniely, Roy Frostig, and Yoram Singer. Toward Deeper Understanding of Neural Networks:\nThe Power of Initialization and a Dual View on Expressivity. arXiv:1602.05897 [cs, stat], February\n2016.\n[2] Jaehoon Lee, Yasaman Bahri, Roman Novak, Sam Schoenholz, Jeffrey Pennington, and Jascha\nSohl-dickstein. Deep Neural Networks as Gaussian Processes. In International Conference on\nLearning Representations, 2018.\n[3] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural Tangent Kernel: Convergence and\nGeneralization in Neural Networks. arXiv:1806.07572 [cs, math, stat], June 2018. 00000\n[4] Guillermo Valle-Pérez, Chico Q. Camargo, and Ard A. Louis. Deep learning generalizes because\nthe parameter-function map is biased towards simple functions. arXiv:1805.08522 [cs, stat], May\n2018.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #2", "review": "Updates:\n\nThanks for the updates. \n\nI find the new theoretical results interesting and potentially useful,  which shows, in the large $d$ setting, spectrums of CKs/NTKs  for boolean cube, sphere and isotropic Gaussian are closed to each other in some sense. Thus, I raise my score to weakly accepted but lower down my confidence level since I am not that familiar with Boolean cube literature. \n\n\n------------------------------------------------------\nThe study of extremely over-parameterized networks (i.e. infinitely width networks) has become one of the most active research directions in theory deep learning. The key objects in understanding such networks are the conjugate kernel [1, 2] (CK defined in the paper) and the Neural tangent kernels [3] (NTK). The CK characterizes how the network looks like at initialization (connection to Gaussian processes as well) and the NTK is very useful to characterize the gradient descent training dynamics of large width networks in the kernel regime. Understanding properties of such kernels, in particular, their spectra distribution and eigenspace, could be potentially an important step towards a finer-gained understanding of generalization in neural networks.   \n\nThe main contribution of this paper is the development of the spectral theory of CK and NTK on boolean cube (similar or weaker results on uniform distribution in spheres and Gaussian distribution in R^n). More precisely, the authors show that, over the space of boolean cube, the CK/NTK could be diagonalized using the Fourier basis and the eigenvalues depend only on the frequency (i.e. the degree of the monomials); Thm 3.1. The authors also develop some computation tools to compute the spectra; Lemma 3.2.   Using the tools developed in this paper, the authors are able to clarify some of the interesting observations found by other researchers. Most noticeably, the authors show that the observation in [4] 'neural network is biased towards simple functions' is NOT universal. Whether this statement is correct or not depends heavily on the choice of activation function (e.g. Relu v.s. Erf) and hyper-parameters (e.g. weight variance, depths).  There are also some other interesting empirical findings: the optimal depth of a neural network depends on the complexity (i.e. degree in the boolean cube setting) of the function to learn, CK (i.e. training only the last layer) tends to be more useful for learning less complex functions, etc. \n\nOverall, this is a nice paper. I am leaning for a weakly accept. \n\n\n[1] Amit Daniely, Roy Frostig, and Yoram Singer. Toward Deeper Understanding of Neural Networks:\nThe Power of Initialization and a Dual View on Expressivity. arXiv:1602.05897 [cs, stat], February\n2016.\n[2] Jaehoon Lee, Yasaman Bahri, Roman Novak, Sam Schoenholz, Jeffrey Pennington, and Jascha\nSohl-dickstein. Deep Neural Networks as Gaussian Processes. In International Conference on\nLearning Representations, 2018.\n[3] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural Tangent Kernel: Convergence and\nGeneralization in Neural Networks. arXiv:1806.07572 [cs, math, stat], June 2018. 00000\n[4] Guillermo Valle-Pérez, Chico Q. Camargo, and Ard A. Louis. Deep learning generalizes because\nthe parameter-function map is biased towards simple functions. arXiv:1805.08522 [cs, stat], May\n2018.\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1571776170608}], "openreview_url": "https://openreview.net/forum?id=HJlU-AVtvS", "arxiv_id": "1907.10599", "paper_pdf": "papers/HJlU-AVtvS.pdf", "paper_pdf_sha256": "cd7662738181efa64a240c89e100c9b716d04cc10c1c97e0f44bb4a01fda8e17", "paper_pdf_bytes": 3439067, "paper_pdf_source": "openreview", "code_url": "https://github.com/thegregyang/NNspectra", "code_repository": "thegregyang/NNspectra", "code_commit": "8c71181e93a46cdcabfafdf71ae5a58830cbb27d", "code_archive": "repos/HJlU-AVtvS.zip", "code_archive_sha256": "015fc770003bae3db824c98cd88c3a901f9c561bbee89d549512c73d13b58f6b", "code_archive_bytes": 4214439, "code_file_count": 5, "code_extensions": {".ipynb": 3, ".py": 2}, "github_disk_usage_kb": 4160, "github_languages": {"Jupyter Notebook": 3800945, "Python": 15442}, "github_archived": false, "github_pushed_at": "2019-09-19T23:57:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-fine-grained-spectral-perspective-on-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZSDmfoVvCs", "year": 2026, "status": "rejected", "title": "ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork", "authors": ["Caroline Wang", "Arrasy Rahman", "Jiaxun Cui", "Yoonchang Sung", "Peter Stone"], "authorids": ["~Caroline_Wang1", "~Arrasy_Rahman1", "~Jiaxun_Cui1", "~Yoonchang_Sung1", "~Peter_Stone1"], "authors_source": "OpenReview API", "abstract": "Learning to collaborate with previously unseen partners is a fundamental generalization challenge in multi-agent learning, known as Ad Hoc Teamwork (AHT). \nExisting AHT approaches often adopt a two-stage pipeline, where first, a fixed population of teammates is generated with the idea that they should be representative of the teammates that will be seen at deployment time, and second, an AHT agent is trained to collaborate well with agents in the population. \nTo date, the research community has focused on designing separate algorithms for each stage. This separation has led to algorithms that generate teammates with limited coverage of possible behaviors, and that ignore whether the generated teammates are easy to learn from for the AHT agent. \nFurthermore, algorithms for training AHT agents typically treat the set of training teammates as static, thus attempting to generalize to previously unseen partner agents without assuming any control over the set of training teammates.\nThis paper presents a unified framework for AHT by reformulating the problem as an open-ended learning process between an AHT agent and an adversarial teammate generator. \nWe introduce ROTATE, a regret-driven, open-ended training algorithm that alternates between improving the AHT agent and generating teammates that probe its deficiencies. \nExperiments across diverse two-player environments demonstrate that ROTATE significantly outperforms baselines at generalizing to an unseen set of evaluation teammates, thus establishing a new standard for robust and generalizable teamwork.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "rOxqpb3uu2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12903/Reviewer_xw5y"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This tackles the problem of Ad-Hoc Teamwork through iterative training diverse teams.\n\nThe diffuclty in AHT is that your AHT agent might be robust with respect to some, apparently diverse population of teammates, but not others. The paper claims that this is due to a small traning set. Their idea is to use iterative training in order to minimise co-operative regret.\n\nThe iterative idea is obviously not new and the related work contains some algorithms which maintain a diverse set of opponents. ROTATE uses a minimax approach, similarly to e.g. Villein et al, but finds a worst-case policy,rather than a distribution, at each step. This makes me think that the authors have not looked at the related work in sufficient detail.\n\nI would have liked the algorithm to be more precisely defined in the main paper: Eq. 7 says that they use a minimax approach, but $\\Pi^{-i}$ is not defined until Section 6, and only really discussed in detail in the appendix. Since many other works use iterative training, I suppose that the real open-endedness is the generation of new partners, rather than the iterative nature of the training.", "review_text": "This tackles the problem of Ad-Hoc Teamwork through iterative training diverse teams.\n\nThe diffuclty in AHT is that your AHT agent might be robust with respect to some, apparently diverse population of teammates, but not others. The paper claims that this is due to a small traning set. Their idea is to use iterative training in order to minimise co-operative regret.\n\nThe iterative idea is obviously not new and the related work contains some algorithms which maintain a diverse set of opponents. ROTATE uses a minimax approach, similarly to e.g. Villein et al, but finds a worst-case policy,rather than a distribution, at each step. This makes me think that the authors have not looked at the related work in sufficient detail.\n\nI would have liked the algorithm to be more precisely defined in the main paper: Eq. 7 says that they use a minimax approach, but $\\Pi^{-i}$ is not defined until Section 6, and only really discussed in detail in the appendix. Since many other works use iterative training, I suppose that the real open-endedness is the generation of new partners, rather than the iterative nature of the training.", "strengths": "+ Interesting notion of regret\n+ Good comparison with related work.", "weaknesses": "- The authors could have done a better job of identifying which component is more important: the notion of regret, the way the teammates are generated, etc.\n- Unclear novelty.\n- Lack of clarity and theoretical discussion.", "questions": "Can you explain exactly how you used the baselines? From my reading of the appendix, it seems that you only took some aspect of these approaches, and adapted them to your framework, rather than have done a direct comparison.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This tackles the problem of Ad-Hoc Teamwork through iterative training diverse teams.\n\nThe diffuclty in AHT is that your AHT agent might be robust with respect to some, apparently diverse population of teammates, but not others. The paper claims that this is due to a small traning set. Their idea is to use iterative training in order to minimise co-operative regret.\n\nThe iterative idea is obviously not new and the related work contains some algorithms which maintain a diverse set of opponents. ROTATE uses a minimax approach, similarly to e.g. Villein et al, but finds a worst-case policy,rather than a distribution, at each step. This makes me think that the authors have not looked at the related work in sufficient detail.\n\nI would have liked the algorithm to be more precisely defined in the main paper: Eq. 7 says that they use a minimax approach, but $\\Pi^{-i}$ is not defined until Section 6, and only really discussed in detail in the appendix. Since many other works use iterative training, I suppose that the real open-endedness is the generation of new partners, rather than the iterative nature of the training.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "+ Interesting notion of regret\n+ Good comparison with related work.", "weaknesses": "- The authors could have done a better job of identifying which component is more important: the notion of regret, the way the teammates are generated, etc.\n- Unclear novelty.\n- Lack of clarity and theoretical discussion.", "questions": "Can you explain exactly how you used the baselines? From my reading of the appendix, it seems that you only took some aspect of these approaches, and adapted them to your framework, rather than have done a direct comparison.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762149466280}, {"id": "J4lHtJj1Gv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12903/Reviewer_NkUx"], "rating": 2, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The paper proposes ROTATE, a regret-driven open-ended training framework for ad hoc teamwork that reframes zero-shot coordination as minimizing worst-case cooperative regret, i.e., $\\min_{\\pi^{ego}}\\max_{\\pi^{-i}}\\mathbb{E}[\\mathrm{CR}(\\pi^{ego},\\pi^{-i})]$. It introduces a per-state regret objective coupled with an SXP term that maximizes payoff with a best-response partner to discourage sabotage, and alternates teammate generation with ego learning using a population buffer. On Overcooked and Level-Based Foraging, ROTATE outperforms UED and teammate-diversification baselines with unseen partners, and ablations attribute gains to the per-state objective and the buffer.", "review_text": "The paper proposes ROTATE, a regret-driven open-ended training framework for ad hoc teamwork that reframes zero-shot coordination as minimizing worst-case cooperative regret, i.e., $\\min_{\\pi^{ego}}\\max_{\\pi^{-i}}\\mathbb{E}[\\mathrm{CR}(\\pi^{ego},\\pi^{-i})]$. It introduces a per-state regret objective coupled with an SXP term that maximizes payoff with a best-response partner to discourage sabotage, and alternates teammate generation with ego learning using a population buffer. On Overcooked and Level-Based Foraging, ROTATE outperforms UED and teammate-diversification baselines with unseen partners, and ablations attribute gains to the per-state objective and the buffer.", "strengths": "- Comprehensive treatment of ZSC/ad hoc teamwork: the paper unifies teammate generation and ego learning via a cooperative-regret min–max objective $ \\min_{\\pi^{ego}}\\max_{\\pi^{-i}}\\mathbb{E}[\\mathrm{CR}]$, making assumptions and evaluation protocol explicit.", "weaknesses": "- Clarity and exposition: the paper is difficult to follow; the core algorithmic loop (who updates when, how SP/XP/SXP are sampled/weighted, and how the BR is trained/used) is buried under notation, so the end-to-end procedure remains unclear even after multiple readings.\n- Mischaracterization of the gap: (a) the claim that most ZSC/AHT methods are two-stage is outdated—recent open-ended or end-to-end approaches already move beyond fixed teammate sets (e.g., COLE [1], E3T [2], TrajeDi [3]); (b) the comparison to current work is incomplete, especially where prior methods already leverage SP/XP and mixed-play/SXP-style rollouts (e.g., CoMeDi [4]), making the incremental novelty of the proposed per-state regret $J_{\\text{state}}$ hard to isolate.\n- Anti-sabotage rationale under-specified: the paper asserts that coupling per-state regret with an SXP best-response term mitigates sabotage, but offers little intuitive or theoretical support (e.g., no conditions under which maximizing SXP payoff with BR implies low sabotage against arbitrary partners, no analysis of bias induced by approximate BR or sampling); more formal justification or counterexample analysis is needed.\n\nReferences\n\n[1] Li, Y., Zhang, S., Sun, J., Du, Y., Wen, Y., Wang, X., and Pan, W. 2023. Cooperative Open-ended Learning Framework for Zero-Shot Coordination. In Proceedings of the 40th International Conference on Machine Learning (ICML 2023). Proceedings of Machine Learning Research, 202:20470–20484.\n\n[2] Yan, X., Guo, J., Lou, X., Wang, J., Zhang, H., and Du, Y. 2023. An Efficient End-to-End Training Approach for Zero-Shot Human-AI Coordination. In Proceedings of the Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS 2023).\n\n[3] Lupu, A., Cui, B., Hu, H., and Foerster, J. 2021. Trajectory Diversity for Zero-Shot Coordination. In Proceedings of the 38th International Conference on Machine Learning (ICML 2021). Proceedings of Machine Learning Research, 139:7204–7213.\n\n[4] Sarkar, B., Shih, A., and Sadigh, D. 2023. Diverse Conventions for Human-AI Collaboration. In Proceedings of the Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS 2023).", "questions": "- Does the combination of per-state regret on SP/XP and the SXP best-response payoff formally or intuitively guarantee reduced sabotage (i.e., lower probability of destructive actions), and under what assumptions on BR optimality and sampling?\n- Can an agent maximize $J_{\\text{state}}$ on SP/XP while keeping high SXP payoff yet still sabotage arbitrary non-BR partners (e.g., collusion with BR)? Is there any bound linking SXP payoff to sabotage rate against unseen partners?\n- How sensitive is the anti-sabotage effect to the weighting between SP/XP and SXP and to environments without reliable state resets/cut-ins? Please provide analysis or ablations.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes ROTATE, a regret-driven open-ended training framework for ad hoc teamwork that reframes zero-shot coordination as minimizing worst-case cooperative regret, i.e., $\\min_{\\pi^{ego}}\\max_{\\pi^{-i}}\\mathbb{E}[\\mathrm{CR}(\\pi^{ego},\\pi^{-i})]$. It introduces a per-state regret objective coupled with an SXP term that maximizes payoff with a best-response partner to discourage sabotage, and alternates teammate generation with ego learning using a population buffer. On Overcooked and Level-Based Foraging, ROTATE outperforms UED and teammate-diversification baselines with unseen partners, and ablations attribute gains to the per-state objective and the buffer.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "- Comprehensive treatment of ZSC/ad hoc teamwork: the paper unifies teammate generation and ego learning via a cooperative-regret min–max objective $ \\min_{\\pi^{ego}}\\max_{\\pi^{-i}}\\mathbb{E}[\\mathrm{CR}]$, making assumptions and evaluation protocol explicit.", "weaknesses": "- Clarity and exposition: the paper is difficult to follow; the core algorithmic loop (who updates when, how SP/XP/SXP are sampled/weighted, and how the BR is trained/used) is buried under notation, so the end-to-end procedure remains unclear even after multiple readings.\n- Mischaracterization of the gap: (a) the claim that most ZSC/AHT methods are two-stage is outdated—recent open-ended or end-to-end approaches already move beyond fixed teammate sets (e.g., COLE [1], E3T [2], TrajeDi [3]); (b) the comparison to current work is incomplete, especially where prior methods already leverage SP/XP and mixed-play/SXP-style rollouts (e.g., CoMeDi [4]), making the incremental novelty of the proposed per-state regret $J_{\\text{state}}$ hard to isolate.\n- Anti-sabotage rationale under-specified: the paper asserts that coupling per-state regret with an SXP best-response term mitigates sabotage, but offers little intuitive or theoretical support (e.g., no conditions under which maximizing SXP payoff with BR implies low sabotage against arbitrary partners, no analysis of bias induced by approximate BR or sampling); more formal justification or counterexample analysis is needed.\n\nReferences\n\n[1] Li, Y., Zhang, S., Sun, J., Du, Y., Wen, Y., Wang, X., and Pan, W. 2023. Cooperative Open-ended Learning Framework for Zero-Shot Coordination. In Proceedings of the 40th International Conference on Machine Learning (ICML 2023). Proceedings of Machine Learning Research, 202:20470–20484.\n\n[2] Yan, X., Guo, J., Lou, X., Wang, J., Zhang, H., and Du, Y. 2023. An Efficient End-to-End Training Approach for Zero-Shot Human-AI Coordination. In Proceedings of the Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS 2023).\n\n[3] Lupu, A., Cui, B., Hu, H., and Foerster, J. 2021. Trajectory Diversity for Zero-Shot Coordination. In Proceedings of the 38th International Conference on Machine Learning (ICML 2021). Proceedings of Machine Learning Research, 139:7204–7213.\n\n[4] Sarkar, B., Shih, A., and Sadigh, D. 2023. Diverse Conventions for Human-AI Collaboration. In Proceedings of the Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS 2023).", "questions": "- Does the combination of per-state regret on SP/XP and the SXP best-response payoff formally or intuitively guarantee reduced sabotage (i.e., lower probability of destructive actions), and under what assumptions on BR optimality and sampling?\n- Can an agent maximize $J_{\\text{state}}$ on SP/XP while keeping high SXP payoff yet still sabotage arbitrary non-BR partners (e.g., collusion with BR)? Is there any bound linking SXP payoff to sabotage rate against unseen partners?\n- How sensitive is the anti-sabotage effect to the weighting between SP/XP and SXP and to environments without reliable state resets/cut-ins? Please provide analysis or ablations.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762085560913}, {"id": "Xq29uB8m3F", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12903/Reviewer_EESm"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper discuss ad hoc teamwork as an open-ended partner co-learning problem and introduces ROTATE, which optimizes a per-state cooperative regret objective while encouraging competent, non-adversarial teammates. The approach alternates between training the ego policy and generating partner policies using state distributions from self-play, cross-play, and switched-play interactions, aided by a population buffer. Experiments across cooperative benchmarks report improved generalization to unseen partners.", "review_text": "The paper discuss ad hoc teamwork as an open-ended partner co-learning problem and introduces ROTATE, which optimizes a per-state cooperative regret objective while encouraging competent, non-adversarial teammates. The approach alternates between training the ego policy and generating partner policies using state distributions from self-play, cross-play, and switched-play interactions, aided by a population buffer. Experiments across cooperative benchmarks report improved generalization to unseen partners.", "strengths": "The paper foregrounds the self-sabotage failure mode in open-ended partner generation, clearly articulating why partners that deliberately depress cross-play (XP) can inflate training signals yet harm zero-shot coordination; this diagnosis sharpens evaluation design (e.g., beyond average XP) and motivates principled mitigation objectives.", "weaknesses": "- Eq. 10 employs a fixed 0.5/0.5 weighting with no analytical justification, and the experiments do not analyze this hyperparameter.\n\n- The method section has poor readability, with unclear logic and difficult-to-follow exposition.\n\n- Missing ZSC-side baselines, especially some open-ended methods like COLE [1] and E3T [2].", "questions": "- Relationship to [3] and [4]: The method currently appears very similar to these two paper—could the authors clarify whether, under certain conditions, it degenerates to XP-min? How do the weights in Eq. 10 relate to XP-min’s hyperparameter ?\n\n- Role of regret: The paper lacks analysis of regret’s actual effect. Do the generated partners indeed exhibit the property of “coordinating with a BR while exposing the ego’s weaknesses without engaging in self-sabotage”? Does Eq. 10 admit any theoretically provable guarantee that suppresses self-sabotage?\n\nReference \n\n[1] Li, Yang, et al. \"Cooperative open-ended learning framework for zero-shot coordination.\" International Conference on Machine Learning. PMLR, 2023.\n\n[2] Yan, Xue, et al. \"An efficient end-to-end training approach for zero-shot human-AI coordination.\" Advances in neural information processing systems 36 (2023): 2636-2658.\n\n[3] Charakorn, Rujikorn, Poramate Manoonpong, and Nat Dilokthanakul. \"Diversity is not all you need: Training a robust cooperative agent needs specialist partners.\" Advances in Neural Information Processing Systems 37 (2024): 56401-56423.\n\n[4]Sarkar, Bidipta, Andy Shih, and Dorsa Sadigh. \"Diverse conventions for human-AI collaboration.\" Advances in neural information processing systems 36 (2023): 23115-23139.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper discuss ad hoc teamwork as an open-ended partner co-learning problem and introduces ROTATE, which optimizes a per-state cooperative regret objective while encouraging competent, non-adversarial teammates. The approach alternates between training the ego policy and generating partner policies using state distributions from self-play, cross-play, and switched-play interactions, aided by a population buffer. Experiments across cooperative benchmarks report improved generalization to unseen partners.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper foregrounds the self-sabotage failure mode in open-ended partner generation, clearly articulating why partners that deliberately depress cross-play (XP) can inflate training signals yet harm zero-shot coordination; this diagnosis sharpens evaluation design (e.g., beyond average XP) and motivates principled mitigation objectives.", "weaknesses": "- Eq. 10 employs a fixed 0.5/0.5 weighting with no analytical justification, and the experiments do not analyze this hyperparameter.\n\n- The method section has poor readability, with unclear logic and difficult-to-follow exposition.\n\n- Missing ZSC-side baselines, especially some open-ended methods like COLE [1] and E3T [2].", "questions": "- Relationship to [3] and [4]: The method currently appears very similar to these two paper—could the authors clarify whether, under certain conditions, it degenerates to XP-min? How do the weights in Eq. 10 relate to XP-min’s hyperparameter ?\n\n- Role of regret: The paper lacks analysis of regret’s actual effect. Do the generated partners indeed exhibit the property of “coordinating with a BR while exposing the ego’s weaknesses without engaging in self-sabotage”? Does Eq. 10 admit any theoretically provable guarantee that suppresses self-sabotage?\n\nReference \n\n[1] Li, Yang, et al. \"Cooperative open-ended learning framework for zero-shot coordination.\" International Conference on Machine Learning. PMLR, 2023.\n\n[2] Yan, Xue, et al. \"An efficient end-to-end training approach for zero-shot human-AI coordination.\" Advances in neural information processing systems 36 (2023): 2636-2658.\n\n[3] Charakorn, Rujikorn, Poramate Manoonpong, and Nat Dilokthanakul. \"Diversity is not all you need: Training a robust cooperative agent needs specialist partners.\" Advances in Neural Information Processing Systems 37 (2024): 56401-56423.\n\n[4]Sarkar, Bidipta, Andy Shih, and Dorsa Sadigh. \"Diverse conventions for human-AI collaboration.\" Advances in neural information processing systems 36 (2023): 23115-23139.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761917185694}, {"id": "DonAmdSqUE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12903/Reviewer_1eRE"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper reframes ad‑hoc teamwork as an open‑ended min–max process: a teammate generator maximizes the ego agent’s cooperative regret while the ego minimizes it, alternating over training. The key idea is state‑wise regret, optimized over SP/XP/SXP state distributions to discourage sabotaging teammates by construction. A population buffer stabilizes training under a non‑stationary teammate distribution. On six two‑agent tasks across LBF and Overcooked, the method outperforms strong baselines on 5/6 tasks, and a toy “destructive matrix game” illustrates that state‑wise regret strongly suppresses sabotage.", "review_text": "The paper reframes ad‑hoc teamwork as an open‑ended min–max process: a teammate generator maximizes the ego agent’s cooperative regret while the ego minimizes it, alternating over training. The key idea is state‑wise regret, optimized over SP/XP/SXP state distributions to discourage sabotaging teammates by construction. A population buffer stabilizes training under a non‑stationary teammate distribution. On six two‑agent tasks across LBF and Overcooked, the method outperforms strong baselines on 5/6 tasks, and a toy “destructive matrix game” illustrates that state‑wise regret strongly suppresses sabotage.", "strengths": "Clear problem reframing. Maximizing cross‑play with unknown partners is cast as minimizing cooperative regret, yielding a natural open‑ended training framework that aligns well with UED principles.\n\nObjective‑level protection against sabotage. Jointly optimizing state‑wise regret on SP/XP/SXP exposes weaknesses while enforcing compatibility with a best‑response partner, which is more principled than adversarial diversity only from the initial state.\n\nReasonable empirical support. Diverse tasks, 9–13 “benign” evaluation partners, a normalized score (upper‑bounded by an estimated BR), and targeted ablations (trajectory vs. state‑wise regret; with/without population buffer).", "weaknesses": "1.Narrow experimental scope. Evidence is confined to two agents with full observability. The absence of results on larger multi‑agent and partially observable settings (e.g., SMAC, GRF)[1] weakens claims of scalability.\n\n2.Lack of theory and stopping criteria. “Open‑endedness” lacks a formal definition and a practical stopping rule; there are no guarantees on coverage, convergence, or regret bounds, limiting deployability.\n\n3.Comparative evaluation is incomplete. Empirical comparisons against LIPO[2], MACOP and rigorously budget‑matched versions of BRDiv/CoMeDi are missing; fairness requires equal interaction budgets/compute.\n\n4.Positioning vs. prior work needs precision. The relationship to MACOP and related methods should be spelled out at the level of objective functions, “benign partner” constraints, stopping rules, and network design, to avoid conceptual conflation.\n\n5.Evaluation and ablations need tightening (detail add‑ons).\n\nThe normalized metric depends on an estimated BR upper bound; report sensitivity to BR approximation error.\n\nProvide weight sensitivity for SP/XP/SXP and justify the SXP term’s necessity.\n\nUse more random seeds—open‑ended procedures can have high variance.\n\nClarify environmental assumptions (mid‑episode policy switching/state resets); give an approximation strategy when resets are unavailable.\n\nDiscuss computational complexity as the partner space expands, and what approximations to BR/regret are admissible with guarantees.\n\nRef:\n\n[1]A survey of progress on cooperative multi-agent reinforcement learning in open environment\n\n[2] Generating Diverse Cooperative Agents by Learning Incompatible Policies", "questions": "1.Scaling to many agents. How does the method avoid combinatorial blow‑up for 5–10 agents or more? Would centralized training with decentralized execution, hierarchical BR, population BR, or fictitious‑play‑style approximations be viable, and at what cost?\n\n2.Multi‑modal partner distributions. If the partner distribution is genuinely multi‑modal, is a single ego policy sufficient? Would mixture‑of‑experts, latent‑variable policies, or distributionally robust objectives (e.g., CVaR) be required to capture distinct partner modes?\n\n3.Formalizing open‑endedness. What is the precise criterion—coverage growth, novelty accumulation, or monotone regret reduction? Please provide operational metrics (coverage/novelty/regret) and a stopping rule, accompanied by evidence.\n\n4.Human–AI collaboration. Can the method transfer to real human partners (e.g., Overcooked‑human)? Would demonstrations, preference modeling, or safety constraints be needed to bound the teammate generator’s search space?\n\n5.Visualization and interpretability. Please include state‑level sabotage heatmaps, SXP vs. XP occupancy differences, teammate embedding visualizations, and term‑wise causal ablations to show where and why each component works.\n\n6.Embodied multi‑agent and LLM integration[1]. Can the framework extend to embodied, partially observable, continuous‑control domains? Could LLMs serve as a teammate generator or a language‑mediated coordination channel for richer partner diversity and policy decomposition?\n\nRef:\n\n[1] \t\nMulti-agent embodied ai: Advances and future directions", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper reframes ad‑hoc teamwork as an open‑ended min–max process: a teammate generator maximizes the ego agent’s cooperative regret while the ego minimizes it, alternating over training. The key idea is state‑wise regret, optimized over SP/XP/SXP state distributions to discourage sabotaging teammates by construction. A population buffer stabilizes training under a non‑stationary teammate distribution. On six two‑agent tasks across LBF and Overcooked, the method outperforms strong baselines on 5/6 tasks, and a toy “destructive matrix game” illustrates that state‑wise regret strongly suppresses sabotage.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "Clear problem reframing. Maximizing cross‑play with unknown partners is cast as minimizing cooperative regret, yielding a natural open‑ended training framework that aligns well with UED principles.\n\nObjective‑level protection against sabotage. Jointly optimizing state‑wise regret on SP/XP/SXP exposes weaknesses while enforcing compatibility with a best‑response partner, which is more principled than adversarial diversity only from the initial state.\n\nReasonable empirical support. Diverse tasks, 9–13 “benign” evaluation partners, a normalized score (upper‑bounded by an estimated BR), and targeted ablations (trajectory vs. state‑wise regret; with/without population buffer).", "weaknesses": "1.Narrow experimental scope. Evidence is confined to two agents with full observability. The absence of results on larger multi‑agent and partially observable settings (e.g., SMAC, GRF)[1] weakens claims of scalability.\n\n2.Lack of theory and stopping criteria. “Open‑endedness” lacks a formal definition and a practical stopping rule; there are no guarantees on coverage, convergence, or regret bounds, limiting deployability.\n\n3.Comparative evaluation is incomplete. Empirical comparisons against LIPO[2], MACOP and rigorously budget‑matched versions of BRDiv/CoMeDi are missing; fairness requires equal interaction budgets/compute.\n\n4.Positioning vs. prior work needs precision. The relationship to MACOP and related methods should be spelled out at the level of objective functions, “benign partner” constraints, stopping rules, and network design, to avoid conceptual conflation.\n\n5.Evaluation and ablations need tightening (detail add‑ons).\n\nThe normalized metric depends on an estimated BR upper bound; report sensitivity to BR approximation error.\n\nProvide weight sensitivity for SP/XP/SXP and justify the SXP term’s necessity.\n\nUse more random seeds—open‑ended procedures can have high variance.\n\nClarify environmental assumptions (mid‑episode policy switching/state resets); give an approximation strategy when resets are unavailable.\n\nDiscuss computational complexity as the partner space expands, and what approximations to BR/regret are admissible with guarantees.\n\nRef:\n\n[1]A survey of progress on cooperative multi-agent reinforcement learning in open environment\n\n[2] Generating Diverse Cooperative Agents by Learning Incompatible Policies", "questions": "1.Scaling to many agents. How does the method avoid combinatorial blow‑up for 5–10 agents or more? Would centralized training with decentralized execution, hierarchical BR, population BR, or fictitious‑play‑style approximations be viable, and at what cost?\n\n2.Multi‑modal partner distributions. If the partner distribution is genuinely multi‑modal, is a single ego policy sufficient? Would mixture‑of‑experts, latent‑variable policies, or distributionally robust objectives (e.g., CVaR) be required to capture distinct partner modes?\n\n3.Formalizing open‑endedness. What is the precise criterion—coverage growth, novelty accumulation, or monotone regret reduction? Please provide operational metrics (coverage/novelty/regret) and a stopping rule, accompanied by evidence.\n\n4.Human–AI collaboration. Can the method transfer to real human partners (e.g., Overcooked‑human)? Would demonstrations, preference modeling, or safety constraints be needed to bound the teammate generator’s search space?\n\n5.Visualization and interpretability. Please include state‑level sabotage heatmaps, SXP vs. XP occupancy differences, teammate embedding visualizations, and term‑wise causal ablations to show where and why each component works.\n\n6.Embodied multi‑agent and LLM integration[1]. Can the framework extend to embodied, partially observable, continuous‑control domains? Could LLMs serve as a teammate generator or a language‑mediated coordination channel for richer partner diversity and policy decomposition?\n\nRef:\n\n[1] \t\nMulti-agent embodied ai: Advances and future directions", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761665685262}], "openreview_url": "https://openreview.net/forum?id=ZSDmfoVvCs", "arxiv_id": "2505.23686", "paper_pdf": "papers/ZSDmfoVvCs.pdf", "paper_pdf_sha256": "f0bc17e1c3a661066f0f7bb5e6e4362e2caa9d8ab806ba646a916aac4f745429", "paper_pdf_bytes": 1577273, "paper_pdf_source": "openreview", "code_url": "https://github.com/carolinewang01/rotate", "code_repository": "carolinewang01/rotate", "code_commit": "68dbf05f49b77096b2b662853b7ac7dfef97aaae", "code_archive": "repos/ZSDmfoVvCs.zip", "code_archive_sha256": "23df00704ec505fb5a58e555e329d2ce098d9bdef43bc2572386c6fd8992af8a", "code_archive_bytes": 669506, "code_file_count": 84, "code_extensions": {".py": 79, ".sh": 4, ".ipynb": 1}, "github_disk_usage_kb": 3523, "github_languages": {"Python": 888697, "Jupyter Notebook": 86576, "Shell": 7124}, "github_archived": false, "github_pushed_at": "2025-10-22T02:47:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rotate-regret-driven-open-ended-training-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MHP4jGMN2E", "year": 2025, "status": "rejected", "title": "The Impact of Element Ordering on LM Agent Performance", "authors": ["Wayne Chi", "Ameet Talwalkar", "Chris Donahue"], "authorids": ["~Wayne_Chi2", "~Ameet_Talwalkar1", "~Chris_Donahue1"], "authors_source": "OpenReview API", "abstract": "There has been a surge of interest in language model agents that can navigate virtual environments such as the web or desktop. To navigate such environments, agents benefit from information on the various elements (e.g., buttons, text, or images) present. However, it remains unclear which element attributes have the greatest impact on agent performance, especially in environments that only provide a graphical representation (i.e., pixels). Here we find that the ordering in which elements are presented to the language model is surprisingly impactful—randomizing element ordering in webpages compromises average agent performance to a degree comparable to removing all visible text from webpages. While web agents benefit from the semantic hierarchical ordering of elements available via the browser, agents that parse elements directly from pixels do not have access to any such ordering. Here we endeavor to derive effective orderings and investigate the impact of various element ordering methods in web and desktop environments. We find that dimensionality reduction provides a viable ordering for pixel-only environments. We train a UI element detection model to derive elements from pixels and apply our findings to an agent benchmark—OmniACT—where we only have access to pixels. Our method completes more than two times as many tasks on average relative to the previous state-of-the-art.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "p9vccnP45d", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5535/Reviewer_4H1W"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper investigates language model agents' navigation capabilities in virtual environments and focuses on exploring how element ordering affects agent performance. The paper's main finding is that element ordering has a significant impact on agent performance, with random ordering leading to substantial performance degradation - comparable to the effect of completely removing all visible text from web pages. While web-based agents can benefit from the semantic hierarchy provided by browsers, agents that parse elements directly from pixels cannot access this ordering information. The paper proposes an ordering method based on dimensionality reduction (such as t-SNE) that performs well in scenarios where only pixel information is available.", "review_text": "This paper investigates language model agents' navigation capabilities in virtual environments and focuses on exploring how element ordering affects agent performance. The paper's main finding is that element ordering has a significant impact on agent performance, with random ordering leading to substantial performance degradation - comparable to the effect of completely removing all visible text from web pages. While web-based agents can benefit from the semantic hierarchy provided by browsers, agents that parse elements directly from pixels cannot access this ordering information. The paper proposes an ordering method based on dimensionality reduction (such as t-SNE) that performs well in scenarios where only pixel information is available.", "strengths": "1. The paper provides the first in-depth investigation of how element ordering affects agent performance and demonstrates its significance.\n\n2. The proposed dimensionality reduction-based ordering method performs well in pixel-only scenarios and achieves new state-of-the-art performance on the OmniACT agent benchmark.\n\n3. The paper introduces a UI element detection model which has been made publicly available for other researchers to use and improve upon.", "weaknesses": "1. While the paper focuses on the t-SNE dimensionality reduction ordering method, it lacks in-depth analysis and comparison with other ordering methods. Additionally, all ordering methods show significant performance gaps compared to Pre-ordering (Table 4).\n\n2. The paper briefly introduces the training process of the UI element detection model but lacks more detailed specifics.", "questions": "1. How does element ordering affect results at different orders of magnitude (e.g., fewer than 10 elements vs. more than 50 elements)?\n\n2. While the ablation section examines how different inputs affect experimental results, to draw more convincing conclusions, the authors should conduct additional ablation experiments on datasets beyond VisualWebArena.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates language model agents' navigation capabilities in virtual environments and focuses on exploring how element ordering affects agent performance. The paper's main finding is that element ordering has a significant impact on agent performance, with random ordering leading to substantial performance degradation - comparable to the effect of completely removing all visible text from web pages. While web-based agents can benefit from the semantic hierarchy provided by browsers, agents that parse elements directly from pixels cannot access this ordering information. The paper proposes an ordering method based on dimensionality reduction (such as t-SNE) that performs well in scenarios where only pixel information is available.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "1. The paper provides the first in-depth investigation of how element ordering affects agent performance and demonstrates its significance.\n\n2. The proposed dimensionality reduction-based ordering method performs well in pixel-only scenarios and achieves new state-of-the-art performance on the OmniACT agent benchmark.\n\n3. The paper introduces a UI element detection model which has been made publicly available for other researchers to use and improve upon.", "weaknesses": "1. While the paper focuses on the t-SNE dimensionality reduction ordering method, it lacks in-depth analysis and comparison with other ordering methods. Additionally, all ordering methods show significant performance gaps compared to Pre-ordering (Table 4).\n\n2. The paper briefly introduces the training process of the UI element detection model but lacks more detailed specifics.", "questions": "1. How does element ordering affect results at different orders of magnitude (e.g., fewer than 10 elements vs. more than 50 elements)?\n\n2. While the ablation section examines how different inputs affect experimental results, to draw more convincing conclusions, the authors should conduct additional ablation experiments on datasets beyond VisualWebArena.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730721637833}, {"id": "BdlTLrDgwl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5535/Reviewer_Ez7z"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper investigates the impact of element ordering on the performance of language model agents. The empirical study highlights that the sequence in which UI elements are presented to the agent significantly influences performance, especially for elements represented solely by pixels, where traditional hierarchical ordering is unavailable.\n\nThe authors propose an ordering method based on dimensionality reduction, necessitating the training of a UI element detection model to extract elements from pixel data. This approach offers a viable solution for ordering in pixel-only environments. The proposed method consistently demonstrates performance improvements over random ordering and achieves a new state-of-the-art result on the OmniACT benchmark.", "review_text": "This paper investigates the impact of element ordering on the performance of language model agents. The empirical study highlights that the sequence in which UI elements are presented to the agent significantly influences performance, especially for elements represented solely by pixels, where traditional hierarchical ordering is unavailable.\n\nThe authors propose an ordering method based on dimensionality reduction, necessitating the training of a UI element detection model to extract elements from pixel data. This approach offers a viable solution for ordering in pixel-only environments. The proposed method consistently demonstrates performance improvements over random ordering and achieves a new state-of-the-art result on the OmniACT benchmark.", "strengths": "1. Reveals the Critical Impact of Element Order on Agent Performance: Through systematic experiments, the paper effectively demonstrates the significant influence of element ordering on language model agents operating in pixel-only environments. This finding presents a new perspective for developing efficient virtual environment navigation algorithms. Previous research has often concentrated on accuracy in image recognition and text analysis, overlooking the importance of element order in contextual understanding. This paper addresses this gap by introducing order optimization as a novel method for enhancing agent performance.\n\n2. Provides an Ordering Method in Environments Lacking Hierarchical Information: In situations where structural information (e.g., HTML or DOM trees) is not available, traditional methods struggle to identify an optimal element order. By utilizing t-SNE for dimensionality reduction, the paper effectively maps pixel-level interface elements to a one-dimensional sequence, establishing a practical ordering method for pixel-only environments. Despite its simplicity, this approach achieves performance gains in pixel-based contexts, setting new performance benchmarks on platforms like OmniACT.", "weaknesses": "1. While the baseline of random ordering is understandable as detrimental to large language models in interpreting UI, it is overly simplistic and is not a strong baseline. The study would benefit from incorporating more heuristic baselines. For instance, could a vision-language model, such as GPT-4V, assist in determining an optimal ordering when seeing the UI directly?\n\n2. In Table 6, the proposed method consistently outperforms other ordering techniques only when elements are detected using Faster R-CNN, failing to demonstrate consistent advantages in other scenarios.\n\n3. Although the paper validates the effectiveness of various ordering methods (e.g., t-SNE ordering surpassing random ordering), it lacks a systematic analysis of why specific ordering methods enhance performance. The authors primarily illustrate the effects of ordering through experiments but do not investigate the mechanisms by which different strategies influence the language model's understanding and contextual construction.", "questions": "Please refer to the weaknesses section above for specific points that require clarification or further detail.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the impact of element ordering on the performance of language model agents. The empirical study highlights that the sequence in which UI elements are presented to the agent significantly influences performance, especially for elements represented solely by pixels, where traditional hierarchical ordering is unavailable.\n\nThe authors propose an ordering method based on dimensionality reduction, necessitating the training of a UI element detection model to extract elements from pixel data. This approach offers a viable solution for ordering in pixel-only environments. The proposed method consistently demonstrates performance improvements over random ordering and achieves a new state-of-the-art result on the OmniACT benchmark.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Reveals the Critical Impact of Element Order on Agent Performance: Through systematic experiments, the paper effectively demonstrates the significant influence of element ordering on language model agents operating in pixel-only environments. This finding presents a new perspective for developing efficient virtual environment navigation algorithms. Previous research has often concentrated on accuracy in image recognition and text analysis, overlooking the importance of element order in contextual understanding. This paper addresses this gap by introducing order optimization as a novel method for enhancing agent performance.\n\n2. Provides an Ordering Method in Environments Lacking Hierarchical Information: In situations where structural information (e.g., HTML or DOM trees) is not available, traditional methods struggle to identify an optimal element order. By utilizing t-SNE for dimensionality reduction, the paper effectively maps pixel-level interface elements to a one-dimensional sequence, establishing a practical ordering method for pixel-only environments. Despite its simplicity, this approach achieves performance gains in pixel-based contexts, setting new performance benchmarks on platforms like OmniACT.", "weaknesses": "1. While the baseline of random ordering is understandable as detrimental to large language models in interpreting UI, it is overly simplistic and is not a strong baseline. The study would benefit from incorporating more heuristic baselines. For instance, could a vision-language model, such as GPT-4V, assist in determining an optimal ordering when seeing the UI directly?\n\n2. In Table 6, the proposed method consistently outperforms other ordering techniques only when elements are detected using Faster R-CNN, failing to demonstrate consistent advantages in other scenarios.\n\n3. Although the paper validates the effectiveness of various ordering methods (e.g., t-SNE ordering surpassing random ordering), it lacks a systematic analysis of why specific ordering methods enhance performance. The authors primarily illustrate the effects of ordering through experiments but do not investigate the mechanisms by which different strategies influence the language model's understanding and contextual construction.", "questions": "Please refer to the weaknesses section above for specific points that require clarification or further detail.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730702649955}, {"id": "OXBrpKUYnM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5535/Reviewer_8p1Y"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper explores the effect of element ordering on LLM agent performance in virtual environments like web and desktop. Through ablation studies on the VisualWebArena and OmniACT benchmarks, they find that random element ordering significantly degrades performance, similar to removing all textual information. To address this, the authors propose using dimensionality reduction techniques, specifically t-SNE, to create effective element orderings based on spatial relationships in pixel-based environments. The authors contribute a trained UI detection model and achieve state-of-the-art performance on OmniACT, demonstrating the impact of ordering strategies for agent navigation in pixel-only environments.", "review_text": "The paper explores the effect of element ordering on LLM agent performance in virtual environments like web and desktop. Through ablation studies on the VisualWebArena and OmniACT benchmarks, they find that random element ordering significantly degrades performance, similar to removing all textual information. To address this, the authors propose using dimensionality reduction techniques, specifically t-SNE, to create effective element orderings based on spatial relationships in pixel-based environments. The authors contribute a trained UI detection model and achieve state-of-the-art performance on OmniACT, demonstrating the impact of ordering strategies for agent navigation in pixel-only environments.", "strengths": "- Originality: The paper tackles a relatively unexplored area—optimizing UI element ordering for LM agents in pixel-only environments. The use of t-SNE for ordering based on spatial relationships is a novel application, offering a fresh perspective on improving agent navigation performance.\n- Quality: The research employs ablation studies on VisualWebArena and OmniACT. The methodological depth and comparison across multiple ordering methods highlight the improvement of the approach.\n- Clarity: The paper presents the problem and solution clearly, with intuitive descriptions of complex concepts like dimensionality reduction for ordering.\n- Significance: Achieving state-of-the-art performance on OmniACT, the paper has practical implications for LM agent design in unstructured environments. The release of the trained UI detection model will further enhances its contribution to the research community.", "weaknesses": "- The paper could clarify its discussion of the dimensionality reduction approach, specifically addressing the parameters used in t-SNE. As t-SNE can be sensitive to parameter tuning, a detailed analysis of how parameter choices affect ordering outcomes would strengthen the validity of the results.\n- Another weakness is the limited exploration of alternative ordering methods. While t-SNE provides performance gains, the study would benefit from a broader examination of ordering techniques, especially methods more aligned with semantic or functional groupings, which could be valuable in complex UI layouts where spatial proximity does not equate to functional relevance. \n- While the authors release the trained UI detection model, the study would benefit from greater transparency in terms of training data diversity and model performance metrics, such as precision and recall for UI element detection, to better assess its effectiveness across domains.\n- The ablation study lacks rigorous control over experimental variables, which makes it difficult to isolate the exact impact of individual attributes on agent performance. For instance, the experiments appear to vary multiple aspects of the state representation without consistently controlling. A more systematic approach to ablation studies would enhance the credibility of the findings.", "questions": "- The paper mentions using default parameters for t-SNE, but this method can be sensitive to parameter choices, such as perplexity and learning rate. Please provide a sensitivity analysis of key t-SNE parameters (e.g., perplexity, learning rate) and their impact on ordering performance.\n- While t-SNE shows promising results for ordering, it may also be beneficial to explore other dimensionality reduction techniques like UMAP and MDS. Please compare t-SNE with at least one other dimensionality reduction technique (e.g., UMAP) on a subset of data, and report the relative performance in terms of agent task success rate.\n- While the appendix provides training details for the UI detection model, the paper would benefit from additional performance metrics specific to this scenario. Please provide a breakdown of the training data composition and to report related object detection metrics (e.g., mAP, precision-recall curves) for the UI detection model.\n- Tables 4 and 5 report key results on different ordering methods, yet they lack detailed context, such as specific configurations or experimental conditions for each row. Could the authors expand on these tables with additional descriptions or footnotes? For example, is the trained model used for UI detection consistent across all experiments, does the success rate and action score reflect the same metric on OmniACT, etc.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores the effect of element ordering on LLM agent performance in virtual environments like web and desktop. Through ablation studies on the VisualWebArena and OmniACT benchmarks, they find that random element ordering significantly degrades performance, similar to removing all textual information. To address this, the authors propose using dimensionality reduction techniques, specifically t-SNE, to create effective element orderings based on spatial relationships in pixel-based environments. The authors contribute a trained UI detection model and achieve state-of-the-art performance on OmniACT, demonstrating the impact of ordering strategies for agent navigation in pixel-only environments.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- Originality: The paper tackles a relatively unexplored area—optimizing UI element ordering for LM agents in pixel-only environments. The use of t-SNE for ordering based on spatial relationships is a novel application, offering a fresh perspective on improving agent navigation performance.\n- Quality: The research employs ablation studies on VisualWebArena and OmniACT. The methodological depth and comparison across multiple ordering methods highlight the improvement of the approach.\n- Clarity: The paper presents the problem and solution clearly, with intuitive descriptions of complex concepts like dimensionality reduction for ordering.\n- Significance: Achieving state-of-the-art performance on OmniACT, the paper has practical implications for LM agent design in unstructured environments. The release of the trained UI detection model will further enhances its contribution to the research community.", "weaknesses": "- The paper could clarify its discussion of the dimensionality reduction approach, specifically addressing the parameters used in t-SNE. As t-SNE can be sensitive to parameter tuning, a detailed analysis of how parameter choices affect ordering outcomes would strengthen the validity of the results.\n- Another weakness is the limited exploration of alternative ordering methods. While t-SNE provides performance gains, the study would benefit from a broader examination of ordering techniques, especially methods more aligned with semantic or functional groupings, which could be valuable in complex UI layouts where spatial proximity does not equate to functional relevance. \n- While the authors release the trained UI detection model, the study would benefit from greater transparency in terms of training data diversity and model performance metrics, such as precision and recall for UI element detection, to better assess its effectiveness across domains.\n- The ablation study lacks rigorous control over experimental variables, which makes it difficult to isolate the exact impact of individual attributes on agent performance. For instance, the experiments appear to vary multiple aspects of the state representation without consistently controlling. A more systematic approach to ablation studies would enhance the credibility of the findings.", "questions": "- The paper mentions using default parameters for t-SNE, but this method can be sensitive to parameter choices, such as perplexity and learning rate. Please provide a sensitivity analysis of key t-SNE parameters (e.g., perplexity, learning rate) and their impact on ordering performance.\n- While t-SNE shows promising results for ordering, it may also be beneficial to explore other dimensionality reduction techniques like UMAP and MDS. Please compare t-SNE with at least one other dimensionality reduction technique (e.g., UMAP) on a subset of data, and report the relative performance in terms of agent task success rate.\n- While the appendix provides training details for the UI detection model, the paper would benefit from additional performance metrics specific to this scenario. Please provide a breakdown of the training data composition and to report related object detection metrics (e.g., mAP, precision-recall curves) for the UI detection model.\n- Tables 4 and 5 report key results on different ordering methods, yet they lack detailed context, such as specific configurations or experimental conditions for each row. Could the authors expand on these tables with additional descriptions or footnotes? For example, is the trained model used for UI detection consistent across all experiments, does the success rate and action score reflect the same metric on OmniACT, etc.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730596107935}, {"id": "CitJOT14Ye", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5535/Reviewer_UaCt"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This manuscript investigates the influence of ordering (in which elements are presented) to the performance of language model agents on navigating virtual environments, and proposes two simple ordering methods (raster and t-SNE) that are better than random ordering.", "review_text": "This manuscript investigates the influence of ordering (in which elements are presented) to the performance of language model agents on navigating virtual environments, and proposes two simple ordering methods (raster and t-SNE) that are better than random ordering.", "strengths": "1. The investigated problem (ie, ordering’s effects) is somewhat novel.\n2. The findings that ordering can influence the effects are intuitive and consistent with human perception.\n3. The paper is presented clearly and easy to follow.", "weaknesses": "W1. Some conclusions are not well supported by experimental results.\n- 1.1 In L299, the manuscript claims that “Random ordering results in a similar performance drop to removing all HTML text descriptions”. However, random ordering results in 37.04% (Gemini 1.5) and 44.44% (GPT4-V) in Table 2, while removing all texts brings 3.7% (Gemini 1.5) and 38.89% (GPT4-V) in Table 1. The performance gap between these two is still large.\n- 1.2 In the Introduction, the manuscript claims that the t-SNE ordering improves performance. However, in Table 4, t-SNE is worse than raster in (a) VWA with Llama3, and (b) OmniACT with human annotations. Here, 4 out of 9 metrics, almost half, show raster is better than t-SNE. \n\nW2. The analyses and explanations of ordering’s effects are not sufficient. The manuscript notices that t-SNE is worse than raster in OmniACT with human annotations and provides these results in L455. However, why raster is better than t-SNE with human annotations has not been further analyzed.\n\nW3. The necessity of using t-SNE to reduce dimension from 2D to 1D is not well stated. It is well known that t-SNE is more suitable for high-dimensional features, but the dimension of the original space is small (ie., 2D). In such case, using t-SNE will roughly equal the process that (1) choosing the top-left one element A, (2) finding its nearest neighbor element B, (3) then finding the nearest neighbor of both element A and element B (with a weighting strategy like exponential moving average), and so forth. \n\nW4. The experimental settings are not clear.\n- 4.1 In L704, the manuscript states that only a subset of OmniACT is used in experiments. However, in Sec. 6.2, the manuscript directly compares the results with previously reported results tested on the full OmniACT. This raises concerns about the comparison results. \n- 4.2 In Table 5, the manuscript claims state-of-the-art performance. However, the sequence score is much lower than previous methods. Also, the foundational models are different. It is hard to say whether the performance difference is from the foundational models or the designs. \n\nW5. There are some writing mistakes.\n- In L212, the words “must find” are repeated two times.\n- In L461, “on Omniact” should be “on OmniACT”.\n- In L465, “a LM” should be “an LM”.", "questions": "I am not very familiar with this field, so I will adjust my rating after reading other reviewers' feedbacks. \n\nPlease see weaknesses W1 to W4. Writing mistakes do not need a response.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript investigates the influence of ordering (in which elements are presented) to the performance of language model agents on navigating virtual environments, and proposes two simple ordering methods (raster and t-SNE) that are better than random ordering.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The investigated problem (ie, ordering’s effects) is somewhat novel.\n2. The findings that ordering can influence the effects are intuitive and consistent with human perception.\n3. The paper is presented clearly and easy to follow.", "weaknesses": "W1. Some conclusions are not well supported by experimental results.\n- 1.1 In L299, the manuscript claims that “Random ordering results in a similar performance drop to removing all HTML text descriptions”. However, random ordering results in 37.04% (Gemini 1.5) and 44.44% (GPT4-V) in Table 2, while removing all texts brings 3.7% (Gemini 1.5) and 38.89% (GPT4-V) in Table 1. The performance gap between these two is still large.\n- 1.2 In the Introduction, the manuscript claims that the t-SNE ordering improves performance. However, in Table 4, t-SNE is worse than raster in (a) VWA with Llama3, and (b) OmniACT with human annotations. Here, 4 out of 9 metrics, almost half, show raster is better than t-SNE. \n\nW2. The analyses and explanations of ordering’s effects are not sufficient. The manuscript notices that t-SNE is worse than raster in OmniACT with human annotations and provides these results in L455. However, why raster is better than t-SNE with human annotations has not been further analyzed.\n\nW3. The necessity of using t-SNE to reduce dimension from 2D to 1D is not well stated. It is well known that t-SNE is more suitable for high-dimensional features, but the dimension of the original space is small (ie., 2D). In such case, using t-SNE will roughly equal the process that (1) choosing the top-left one element A, (2) finding its nearest neighbor element B, (3) then finding the nearest neighbor of both element A and element B (with a weighting strategy like exponential moving average), and so forth. \n\nW4. The experimental settings are not clear.\n- 4.1 In L704, the manuscript states that only a subset of OmniACT is used in experiments. However, in Sec. 6.2, the manuscript directly compares the results with previously reported results tested on the full OmniACT. This raises concerns about the comparison results. \n- 4.2 In Table 5, the manuscript claims state-of-the-art performance. However, the sequence score is much lower than previous methods. Also, the foundational models are different. It is hard to say whether the performance difference is from the foundational models or the designs. \n\nW5. There are some writing mistakes.\n- In L212, the words “must find” are repeated two times.\n- In L461, “on Omniact” should be “on OmniACT”.\n- In L465, “a LM” should be “an LM”.", "questions": "I am not very familiar with this field, so I will adjust my rating after reading other reviewers' feedbacks. \n\nPlease see weaknesses W1 to W4. Writing mistakes do not need a response.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729834166759}], "openreview_url": "https://openreview.net/forum?id=MHP4jGMN2E", "arxiv_id": "2409.12089", "paper_pdf": "papers/MHP4jGMN2E.pdf", "paper_pdf_sha256": "3a2774b7b1f9a431b383f08350bd80475c3120a4df251d0af71931491329af4f", "paper_pdf_bytes": 3166130, "paper_pdf_source": "openreview", "code_url": "https://github.com/waynchi/gui-agent", "code_repository": "waynchi/gui-agent", "code_commit": "31f972daa507b3225a20dcc68e36e92483f47557", "code_archive": "repos/MHP4jGMN2E.zip", "code_archive_sha256": "12922954691006e90aee8fa9ca7b4f309daa3a71241c5241a6d0763350d12733", "code_archive_bytes": 364840, "code_file_count": 88, "code_extensions": {".py": 81, ".sh": 7}, "github_disk_usage_kb": 250, "github_languages": {"Python": 417822, "Shell": 9397}, "github_archived": false, "github_pushed_at": "2024-09-30T16:38:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-impact-of-element-ordering-on-lm-agent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Qvoe4wXWFi", "year": 2024, "status": "rejected", "title": "NeuralFuse: Learning to Recover the Accuracy of Access-Limited Neural Network Inference in Low-Voltage Regimes", "authors": ["Hao-Lun Sun", "Lei Hsiung", "Nandhini Chandramoorthy", "Pin-Yu Chen", "Tsung-Yi Ho"], "authorids": ["~Hao-Lun_Sun1", "~Lei_Hsiung1", "~Nandhini_Chandramoorthy1", "~Pin-Yu_Chen1", "~Tsung-Yi_Ho2"], "authors_source": "OpenReview API", "abstract": "Deep neural networks (DNNs) have become ubiquitous in machine learning, but their energy consumption remains a notable issue. Lowering the supply voltage is an effective strategy for reducing energy consumption. However, aggressively scaling down the supply voltage can lead to accuracy degradation due to random bit flips in static random access memory (SRAM) where model parameters are stored. To address this challenge, we introduce NeuralFuse, a novel add-on module that addresses the accuracy-energy tradeoff in low-voltage regimes by learning input transformations to generate error-resistant data representations. NeuralFuse protects DNN accuracy in both nominal and low-voltage scenarios. Moreover, NeuralFuse is easy to implement and can be readily applied to DNNs with limited access, such as non-configurable hardware or remote access to cloud-based APIs. Experimental results demonstrate that, at a 1% bit error rate, NeuralFuse can reduce SRAM memory access energy by up to 24% while recovering accuracy by up to 57%. To the best of our knowledge, this is the first model-agnostic approach (i.e., no model retraining) to address low-voltage-induced bit errors.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "lxlwDBPoxx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4390/Reviewer_Qmac"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work presents an add-on module that can be added to image classifiers when they are employed/inferenced in a low-power and error prone accelerator. The module is trained by various perturbated models (models that run on machines with bit errors in SRAMs). The proposed module can be trained on two real-life scenarios: 1) relaxed access and 2) restricted access. The extensive experimental results show that the proposed method is effective in error resiliency and power saving.", "review_text": "This work presents an add-on module that can be added to image classifiers when they are employed/inferenced in a low-power and error prone accelerator. The module is trained by various perturbated models (models that run on machines with bit errors in SRAMs). The proposed module can be trained on two real-life scenarios: 1) relaxed access and 2) restricted access. The extensive experimental results show that the proposed method is effective in error resiliency and power saving.", "strengths": "The paper presents an novel idea of adding a module to any image classifiers where the image model can suffer from low-voltage induced errors. This approach does not require retraining of the models and can be applied to any proprietary-protected DL models. \n\nThe extensive experiments show the effectiveness of the work. The paper is well written and organized. \n\nAs large models are being developed and deployed around the world, the proposed method can save significant energy and pave the way to greener AI. Although the work is only focused on the image classifier, it opens a door to robust DL in other domains.", "weaknesses": "The work assumes that the NeuralFuse generator can be employed on the hardware of no-error voltage. To justify this claim, it would be great if there is a comparison of the sizes (number of parameters) between NeuralFuse generator and the classifier.", "questions": "The review can see the architectures of the generators in the appendix. How are the detailed architecture of generators decided? Any insights on the architecture of the NeuralFuse generator?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents an add-on module that can be added to image classifiers when they are employed/inferenced in a low-power and error prone accelerator. The module is trained by various perturbated models (models that run on machines with bit errors in SRAMs). The proposed module can be trained on two real-life scenarios: 1) relaxed access and 2) restricted access. The extensive experimental results show that the proposed method is effective in error resiliency and power saving.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "The paper presents an novel idea of adding a module to any image classifiers where the image model can suffer from low-voltage induced errors. This approach does not require retraining of the models and can be applied to any proprietary-protected DL models. \n\nThe extensive experiments show the effectiveness of the work. The paper is well written and organized. \n\nAs large models are being developed and deployed around the world, the proposed method can save significant energy and pave the way to greener AI. Although the work is only focused on the image classifier, it opens a door to robust DL in other domains.", "weaknesses": "The work assumes that the NeuralFuse generator can be employed on the hardware of no-error voltage. To justify this claim, it would be great if there is a comparison of the sizes (number of parameters) between NeuralFuse generator and the classifier.", "questions": "The review can see the architectures of the generators in the appendix. How are the detailed architecture of generators decided? Any insights on the architecture of the NeuralFuse generator?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698792020036}, {"id": "cXqe9bk5rn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4390/Reviewer_YHTK"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper aims to tackle the accuracy drop introduced by the increasing bit error rate under the low-voltage scheme by finding a more robust input representation. An error-resistant input transformation is proposed by utilizing a trainable generator, and a modified training loss is utilized to optimize the predicted outputs with/without bit-error injection. The experiments show an obvious accuracy improvement compared to the baseline.", "review_text": "This paper aims to tackle the accuracy drop introduced by the increasing bit error rate under the low-voltage scheme by finding a more robust input representation. An error-resistant input transformation is proposed by utilizing a trainable generator, and a modified training loss is utilized to optimize the predicted outputs with/without bit-error injection. The experiments show an obvious accuracy improvement compared to the baseline.", "strengths": "* Neat paper structure and easy-to-follow content.\n* A simple add-on strategy that can be used in access-limited scenarios.\n* Extensive analysis of different generator architectures.", "weaknesses": "* Lack of discussion of introduced overhead of the generator modules. For ImageNet-10, the best generator architecture, UNet-L, has 2.03G MACs. The introduced extra computation cost is significant compared to the vanilla model (ResNet18 only has 1.82 G MACs). It raises the concern that the introduced overhead for the generator is too large compared to the classifier, making the proposed strategy unrealistic. The author only discusses the energy of SRAM access without considering the computation energy and latency.\n* The introduced generator modules may dilute the energy efficiency brought by the low-voltage scheme. Based on Appendix E, the total computations are very large. A more ideal accuracy-saving method should introduce less overhead.\n* Lack of comparison with other error-resistant methods for bit-error rate. The author should add a comparison with other methods to show whether the costly input transformation is worth.", "questions": "* Could the author provide a more complete overhead analysis of the introduced generator? The author should show the introduced energy cost of computation (both memory and computation) and extra latency overhead in 4.4. The paper would be more meaningful if it saved accuracy under small overhead.\n* Could the author compare with other error-mitigation methods for bit error in SRAM?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to tackle the accuracy drop introduced by the increasing bit error rate under the low-voltage scheme by finding a more robust input representation. An error-resistant input transformation is proposed by utilizing a trainable generator, and a modified training loss is utilized to optimize the predicted outputs with/without bit-error injection. The experiments show an obvious accuracy improvement compared to the baseline.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "* Neat paper structure and easy-to-follow content.\n* A simple add-on strategy that can be used in access-limited scenarios.\n* Extensive analysis of different generator architectures.", "weaknesses": "* Lack of discussion of introduced overhead of the generator modules. For ImageNet-10, the best generator architecture, UNet-L, has 2.03G MACs. The introduced extra computation cost is significant compared to the vanilla model (ResNet18 only has 1.82 G MACs). It raises the concern that the introduced overhead for the generator is too large compared to the classifier, making the proposed strategy unrealistic. The author only discusses the energy of SRAM access without considering the computation energy and latency.\n* The introduced generator modules may dilute the energy efficiency brought by the low-voltage scheme. Based on Appendix E, the total computations are very large. A more ideal accuracy-saving method should introduce less overhead.\n* Lack of comparison with other error-resistant methods for bit-error rate. The author should add a comparison with other methods to show whether the costly input transformation is worth.", "questions": "* Could the author provide a more complete overhead analysis of the introduced generator? The author should show the introduced energy cost of computation (both memory and computation) and extra latency overhead in 4.4. The paper would be more meaningful if it saved accuracy under small overhead.\n* Could the author compare with other error-mitigation methods for bit error in SRAM?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698789751034}, {"id": "A8U0K7Hjjo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4390/Reviewer_uT9R"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this study, the authors introduce NeuralFuse, a data preprocessing module designed to enhance resilience against bit errors arising from low-voltage SRAM, while also offering potential energy savings. Comprehensive tests affirm its efficacy in enhancing models affected by perturbations, ensuring transferability across diverse DNN architectures, and bolstering robustness in weight quantization.", "review_text": "In this study, the authors introduce NeuralFuse, a data preprocessing module designed to enhance resilience against bit errors arising from low-voltage SRAM, while also offering potential energy savings. Comprehensive tests affirm its efficacy in enhancing models affected by perturbations, ensuring transferability across diverse DNN architectures, and bolstering robustness in weight quantization.", "strengths": "1. It focuses on a system aspect of neural network computing: a power-saving method with low voltage operation.\n2. The proposed module can work in a plug-and-play manner and does not require retraining the deployed model.", "weaknesses": "1. The net benefits of introducing NeuralFuse in tandem with low-voltage operation remain uncertain. While there are energy savings associated with SRAM accesses, these reports overlook the comprehensive energy consumption of NeuralFuse, particularly the MAC operations.\n2. Even though there's a notable enhancement in recovered accuracy, it might fall short when juxtaposed with the original accuracy, notably in the case of ResNet50.\n3. The significant fluctuation in accuracy suggests that the optimized model may lack consistent predictability.", "questions": "1. How does the energy consumption from SRAM accesses compare to the total inference cost of a DNN? While acknowledging that the overall energy consumption hinges on a myriad of factors, providing a general perspective would be insightful.\n\n2. The true efficacy of power savings from the low-voltage operation remains ambiguous. While Table 2 highlights energy savings, it narrowly focuses on the consumption related to SRAM accesses. Given that the large configurations of NeuralFuse exhibit similar MACs to the base models (as seen in Table 7), the feasibility of NeuralFuse, when accounting for its total overhead, merits reconsideration.\n\n3. The unpredictability of model performance under low voltage operation, especially with bit flips at the MSBs, poses challenges for practical implementation. Could you shed more light on its real-world applicability or potential use-cases?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this study, the authors introduce NeuralFuse, a data preprocessing module designed to enhance resilience against bit errors arising from low-voltage SRAM, while also offering potential energy savings. Comprehensive tests affirm its efficacy in enhancing models affected by perturbations, ensuring transferability across diverse DNN architectures, and bolstering robustness in weight quantization.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. It focuses on a system aspect of neural network computing: a power-saving method with low voltage operation.\n2. The proposed module can work in a plug-and-play manner and does not require retraining the deployed model.", "weaknesses": "1. The net benefits of introducing NeuralFuse in tandem with low-voltage operation remain uncertain. While there are energy savings associated with SRAM accesses, these reports overlook the comprehensive energy consumption of NeuralFuse, particularly the MAC operations.\n2. Even though there's a notable enhancement in recovered accuracy, it might fall short when juxtaposed with the original accuracy, notably in the case of ResNet50.\n3. The significant fluctuation in accuracy suggests that the optimized model may lack consistent predictability.", "questions": "1. How does the energy consumption from SRAM accesses compare to the total inference cost of a DNN? While acknowledging that the overall energy consumption hinges on a myriad of factors, providing a general perspective would be insightful.\n\n2. The true efficacy of power savings from the low-voltage operation remains ambiguous. While Table 2 highlights energy savings, it narrowly focuses on the consumption related to SRAM accesses. Given that the large configurations of NeuralFuse exhibit similar MACs to the base models (as seen in Table 7), the feasibility of NeuralFuse, when accounting for its total overhead, merits reconsideration.\n\n3. The unpredictability of model performance under low voltage operation, especially with bit flips at the MSBs, poses challenges for practical implementation. Could you shed more light on its real-world applicability or potential use-cases?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698758202989}, {"id": "E5G7R3HsYu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4390/Reviewer_7gLE"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes NeuralFuse, a model-agnostic approach that learns input transformations to generate error-resistant data representations. NeuralFuse dynamically adds a correction term to the model input to protect the DNNs in both nominal and low-voltage scenarios and can be applied to DNNs with limited access. Experimental results show that NeuralFuse can reduce SRAM memory access energy by up to 20-30% while recovering accuracy by up to 57% at a 1% bit error rate.", "review_text": "This paper proposes NeuralFuse, a model-agnostic approach that learns input transformations to generate error-resistant data representations. NeuralFuse dynamically adds a correction term to the model input to protect the DNNs in both nominal and low-voltage scenarios and can be applied to DNNs with limited access. Experimental results show that NeuralFuse can reduce SRAM memory access energy by up to 20-30% while recovering accuracy by up to 57% at a 1% bit error rate.", "strengths": "Strength:\n\n1.\tThe idea is quite interesting. Without error-aware training (adversarial training), it learns input-dependent, model-agnostic calibrator for the model input, the DNN’s accuracy can be protected.\n\n2.\tThe proposed neural network can protect the DNN accuracy while still showing energy efficiency benefits.\n\n3.\tIt thoroughly investigates the transferability to different error rates, model architecture, and quantization bitwidth.", "weaknesses": "Weakness:\n1.\tThe transferability on different error rate and model size is not very good, according to Table 1, which means re-training is still required for different model/dataset/SRAM voltages.\n\n2.\tThe energy saving is only ~20% by reducing SRAM voltage, while the accuracy drop is beyond 1%. It needs some justification on this trade-off.\n\n3.\tThe method seems to be equivalent to adding extra layers in the early stage of the network and train it with noise-aware training. Why not train other layers with the memory error? It is not very intuitive that weight errors in all layers (even MSB flips) can be well protected by only changing the model input. More explanation is needed to justify this. Can we add a protector to later layers and do some calibration? Or even parallel branches? Or protect the weights loaded from memory block-wise, which can still maintain model-agnostic?", "questions": "listed in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes NeuralFuse, a model-agnostic approach that learns input transformations to generate error-resistant data representations. NeuralFuse dynamically adds a correction term to the model input to protect the DNNs in both nominal and low-voltage scenarios and can be applied to DNNs with limited access. Experimental results show that NeuralFuse can reduce SRAM memory access energy by up to 20-30% while recovering accuracy by up to 57% at a 1% bit error rate.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Strength:\n\n1.\tThe idea is quite interesting. Without error-aware training (adversarial training), it learns input-dependent, model-agnostic calibrator for the model input, the DNN’s accuracy can be protected.\n\n2.\tThe proposed neural network can protect the DNN accuracy while still showing energy efficiency benefits.\n\n3.\tIt thoroughly investigates the transferability to different error rates, model architecture, and quantization bitwidth.", "weaknesses": "Weakness:\n1.\tThe transferability on different error rate and model size is not very good, according to Table 1, which means re-training is still required for different model/dataset/SRAM voltages.\n\n2.\tThe energy saving is only ~20% by reducing SRAM voltage, while the accuracy drop is beyond 1%. It needs some justification on this trade-off.\n\n3.\tThe method seems to be equivalent to adding extra layers in the early stage of the network and train it with noise-aware training. Why not train other layers with the memory error? It is not very intuitive that weight errors in all layers (even MSB flips) can be well protected by only changing the model input. More explanation is needed to justify this. Can we add a protector to later layers and do some calibration? Or even parallel branches? Or protect the weights loaded from memory block-wise, which can still maintain model-agnostic?", "questions": "listed in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698621960482}], "openreview_url": "https://openreview.net/forum?id=Qvoe4wXWFi", "arxiv_id": "2306.16869", "paper_pdf": "papers/Qvoe4wXWFi.pdf", "paper_pdf_sha256": "85c4fcdd41bc0c4bb4d87bc9265415dadddb1030ec27d14f83c78e28ef166bff", "paper_pdf_bytes": 5099181, "paper_pdf_source": "openreview", "code_url": "https://github.com/IBM/NeuralFuse", "code_repository": "IBM/NeuralFuse", "code_commit": "8bfdbaf5c969fc78b866f48d5ab9c724258a8c33", "code_archive": "repos/Qvoe4wXWFi.zip", "code_archive_sha256": "f199813f55785ae5d4940cb9148179c84b8e0b67aa355a9d2de860004dfe8493", "code_archive_bytes": 189571, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 183, "github_languages": {"Python": 188756}, "github_archived": false, "github_pushed_at": "2025-09-18T00:19:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neuralfuse-learning-to-improve-the-accuracy"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Krk0Gnft2Zc", "year": 2023, "status": "rejected", "title": "Discrete State-Action Abstraction via the Successor Representation", "authors": ["Amnon Attali", "Pedro Cisneros-Velarde", "Marco Morales", "Nancy Amato"], "authorids": ["~Amnon_Attali1", "~Pedro_Cisneros-Velarde1", "~Marco_Morales1", "~Nancy_Amato1"], "authors_source": "OpenReview API", "abstract": "While the difficulty of reinforcement learning problems is typically related to the complexity of their state spaces, Abstraction proposes that solutions often lie in simpler underlying latent spaces. Prior works have focused on learning either a continuous or dense abstraction, or require a human to provide one. Information-dense representations capture features irrelevant for solving tasks, and continuous spaces can struggle to represent discrete objects. In this work we automatically learn a sparse discrete abstraction of the underlying environment. We do so using a simple end-to-end trainable model based on the successor representation and max-entropy regularization. We describe an algorithm to apply our model, named Discrete State-Action Abstraction (DSAA), which computes an action abstraction in the form of temporally extended actions, i.e., Options, to transition between discrete abstract states. Empirically, we demonstrate the effects of different exploration schemes on our resulting abstraction, and show that it is efficient for solving downstream tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tA0ETBzzHIa", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6071/Reviewer_vZz4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to use a dataset of environment transitions to divide states into N discrete clusters. They propose a Discrete VAE architecture that encodes a state into one of N discrete states and then decodes it into its successor representation. The encoder is regularized by a uniform prior leading to similarly sized clusters of states and the decoder ensures states with similar Successor Representations have the same discrete cluster. Temporally abstracted actions (options) are trained to transition between the discovered state clusters. This procedure is shown to perform well in two simple environments: FourRooms (discrete) and Arm2D (continuous).", "review_text": "The paper has original ideas and is written well. However, I am not yet convinced of the importance of the problem solved and method proposed, as it may be limited to toy tasks — both, in terms of the assumptions and the experiments. I would consider raising my score if these concerns could be addressed.", "strengths": "## Strengths\n- The paper is clearly written, easy to follow, has a good structure and flow, and contains sufficient background to understand the method.\n- The idea of clustering states into a discrete set of clusters and using them for option learning is interesting and seems novel.\n- The method is simple, intuitive, and achieves the paper's desiderata of uniformly sized state clusters with similar Successor Representations.\n\n## Weaknesses\n- **Are the assumptions practical?**\n    + My main concern is that the assumptions required to make the method work do not seem to be practical and would not scale to complex tasks. For instance,\n        * The idea of defining state clusters only makes sense in small and closed environments. In unknown and complex environments, it is not even clear what the number of state clusters (and options) would be.\n        * The dataset of environment transitions should cover all the states in the environment, at least those that the downstream task would observe.\n        * In search-like tasks like mazes or Arm2D, the main challenge is reaching certain states using a reward. But this paper bypasses this problem of state exploration by assuming the dataset does most of the hard job. Therefore, it does not seem practical for most problems.\n        * Equally-sized state clusters: Depending on the task family, differently-sized clusters might make more sense, e.g., guided by reconstruction or environment reward or reachability of states or semantic properties (such as objects) of states.\n    + Are there realistic applications that would still fall under such assumptions and thus benefit from the proposed method?\n- **Experimental Evaluation**\n    + **Environments**\n        * While the proposed method can work in both discrete and continuous settings, I am concerned about the complexity of tasks it can scale to. The current environments are too simplistic and deterministic: even random exploration can find the goal in 30 episodes in the FourRooms environment. The prior work (Machado et al. 2018) also uses Successor Representations and shows results on the Atari domain. \n    + **Mismatch from claims**: While the paper claims (in the introduction) that discrete state abstraction is helpful to discover discrete objects or properties and for understandability, none of these benefits are exhibited in the experiments.\n    + **Baselines**: The use-case of the proposed method is in learning state abstraction and options unsupervisedly. However, several skill discovery methods, such as Pertsch et al. (2021), discover skills from unsupervised offline datasets and use them to accelerate downstream RL. Shouldn't such Hierarchical RL methods also be compared as baselines to demonstrate the importance of discrete state abstractions?\n\n[1] Marlos C Machado, Clemens Rosenbaum, Xiaoxiao Guo, Miao Liu, Gerald Tesauro, and Murray Campbell. Eigenoption discovery through the deep successor representation. In 6th International Conference on Learning Representations, 2018.\n\n[2] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes to use a dataset of environment transitions to divide states into N discrete clusters. They propose a Discrete VAE architecture that encodes a state into one of N discrete states and then decodes it into its successor representation. The encoder is regularized by a uniform prior leading to similarly sized clusters of states and the decoder ensures states with similar Successor Representations have the same discrete cluster. Temporally abstracted actions (options) are trained to transition between the discovered state clusters. This procedure is shown to perform well in two simple environments: FourRooms (discrete) and Arm2D (continuous).", "strength_and_weaknesses": "## Strengths\n- The paper is clearly written, easy to follow, has a good structure and flow, and contains sufficient background to understand the method.\n- The idea of clustering states into a discrete set of clusters and using them for option learning is interesting and seems novel.\n- The method is simple, intuitive, and achieves the paper's desiderata of uniformly sized state clusters with similar Successor Representations.\n\n## Weaknesses\n- **Are the assumptions practical?**\n    + My main concern is that the assumptions required to make the method work do not seem to be practical and would not scale to complex tasks. For instance,\n        * The idea of defining state clusters only makes sense in small and closed environments. In unknown and complex environments, it is not even clear what the number of state clusters (and options) would be.\n        * The dataset of environment transitions should cover all the states in the environment, at least those that the downstream task would observe.\n        * In search-like tasks like mazes or Arm2D, the main challenge is reaching certain states using a reward. But this paper bypasses this problem of state exploration by assuming the dataset does most of the hard job. Therefore, it does not seem practical for most problems.\n        * Equally-sized state clusters: Depending on the task family, differently-sized clusters might make more sense, e.g., guided by reconstruction or environment reward or reachability of states or semantic properties (such as objects) of states.\n    + Are there realistic applications that would still fall under such assumptions and thus benefit from the proposed method?\n- **Experimental Evaluation**\n    + **Environments**\n        * While the proposed method can work in both discrete and continuous settings, I am concerned about the complexity of tasks it can scale to. The current environments are too simplistic and deterministic: even random exploration can find the goal in 30 episodes in the FourRooms environment. The prior work (Machado et al. 2018) also uses Successor Representations and shows results on the Atari domain. \n    + **Mismatch from claims**: While the paper claims (in the introduction) that discrete state abstraction is helpful to discover discrete objects or properties and for understandability, none of these benefits are exhibited in the experiments.\n    + **Baselines**: The use-case of the proposed method is in learning state abstraction and options unsupervisedly. However, several skill discovery methods, such as Pertsch et al. (2021), discover skills from unsupervised offline datasets and use them to accelerate downstream RL. Shouldn't such Hierarchical RL methods also be compared as baselines to demonstrate the importance of discrete state abstractions?\n\n[1] Marlos C Machado, Clemens Rosenbaum, Xiaoxiao Guo, Miao Liu, Gerald Tesauro, and Murray Campbell. Eigenoption discovery through the deep successor representation. In 6th International Conference on Learning Representations, 2018.\n\n[2] Pertsch, Karl, Youngwoon Lee, and Joseph Lim. \"Accelerating reinforcement learning with learned skill priors.\" Conference on robot learning. PMLR, 2021.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is written well with clarity and original ideas.", "summary_of_the_review": "The paper has original ideas and is written well. However, I am not yet convinced of the importance of the problem solved and method proposed, as it may be limited to toy tasks — both, in terms of the assumptions and the experiments. I would consider raising my score if these concerns could be addressed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667564506976}, {"id": "biDgddonbQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6071/Reviewer_icxB"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose an approach to learn a discrete abstraction of an environment in reinforcement learning. The proposed approach creates abstract states using the successor representation (Dayan, 1993), reflecting the intuition that states should be grouped together based on \"what happens after visiting a state\", and creates abstract actions (options) that take the agent from one abstract state to neighbouring abstract states. The authors present an empirical analysis of the proposed approach in two domains: a gridworld with four rooms and a three-joint control task on a two-dimensional plane. \n", "review_text": "The authors present a plausible approach to abstraction in reinforcement learning, building on several earlier ideas and methods. The main novelty in the paper is the specific way the successor representation is used in defining the abstract states. The approach could prove useful but the analysis in the current paper does not give the reader a good understanding of the behaviour of the approach and its strengths/weaknesses relative to existing methods.", "strengths": "Strengths: \n\n-- The authors present a plausible approach to abstraction in reinforcement learning, building on several earlier ideas and methods. \n\nWeaknesses: \n\n-- The analysis and evaluation of the approach is quite limited. \n\nFirst, the authors explore the behavior of the proposed approach in a relatively narrow range of environments. The first domain is a small gridworld with four rooms. The authors state \"We highlight that the sparse nature of reaching a specific state in the environment and the existence of bottleneck states in FourRooms makes this is a relatively difficult task for standard RL.\" Not many people would agree with this statement. On the contrary, this is quite a simple task for reinforcement learning. This domain is not only easy but also has a very simple structure, and many alternative methods would identify very similar abstractions (e.g., those based on graph clustering). It is not a particularly informative environment. The second domain is more complex than the gridworld but still relatively simple. \n\nSecondly, the analysis in each domain is relatively limited and does not give the reader a good understanding of the behaviour of the proposed approach. For instance, there is no exploration of how agent performance varies with the number of abstract states.\n\nThirdly, the analysis does not explore some existing approaches that are closely related. For example, as the authors note, Ramesh et al. (IJCAI 2019) propose a similar abstraction of the environment using the successor representation, which would be an informative baseline. Approaches based on graph cuts/clustering would also be relevant and informative. Some of these methods may not scale as well as the proposed approach but it would still be useful to see their similarities/differences and relative strengths/weaknesses explored. \n\n-- How the approach would fare in stochastic environments is not discussed or explored. \n\n-- Computational complexity of the approach is not discussed or explored.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose an approach to learn a discrete abstraction of an environment in reinforcement learning. The proposed approach creates abstract states using the successor representation (Dayan, 1993), reflecting the intuition that states should be grouped together based on \"what happens after visiting a state\", and creates abstract actions (options) that take the agent from one abstract state to neighbouring abstract states. The authors present an empirical analysis of the proposed approach in two domains: a gridworld with four rooms and a three-joint control task on a two-dimensional plane. \n", "strength_and_weaknesses": "Strengths: \n\n-- The authors present a plausible approach to abstraction in reinforcement learning, building on several earlier ideas and methods. \n\nWeaknesses: \n\n-- The analysis and evaluation of the approach is quite limited. \n\nFirst, the authors explore the behavior of the proposed approach in a relatively narrow range of environments. The first domain is a small gridworld with four rooms. The authors state \"We highlight that the sparse nature of reaching a specific state in the environment and the existence of bottleneck states in FourRooms makes this is a relatively difficult task for standard RL.\" Not many people would agree with this statement. On the contrary, this is quite a simple task for reinforcement learning. This domain is not only easy but also has a very simple structure, and many alternative methods would identify very similar abstractions (e.g., those based on graph clustering). It is not a particularly informative environment. The second domain is more complex than the gridworld but still relatively simple. \n\nSecondly, the analysis in each domain is relatively limited and does not give the reader a good understanding of the behaviour of the proposed approach. For instance, there is no exploration of how agent performance varies with the number of abstract states.\n\nThirdly, the analysis does not explore some existing approaches that are closely related. For example, as the authors note, Ramesh et al. (IJCAI 2019) propose a similar abstraction of the environment using the successor representation, which would be an informative baseline. Approaches based on graph cuts/clustering would also be relevant and informative. Some of these methods may not scale as well as the proposed approach but it would still be useful to see their similarities/differences and relative strengths/weaknesses explored. \n\n-- How the approach would fare in stochastic environments is not discussed or explored. \n\n-- Computational complexity of the approach is not discussed or explored.\n", "clarity,_quality,_novelty_and_reproducibility": "The main novelty in the paper is the specific way the successor representation is used in defining the abstract states. An earlier paper by Ramesh et al. (IJCAI 2019) has used the same intuition of grouping states together using the successor representation. Generally, the central ideas and components of the approach have appeared earlier in the literature but the proposed approach puts them together in a novel way. \n\nThe paper is well organised but the individual sections could be better written, with more clarity and detail.\n\nI have not spotted any errors. But I would note that the experimental analysis is not strong enough to back the claims of the authors regarding how useful the approach is. ", "summary_of_the_review": "The authors present a plausible approach to abstraction in reinforcement learning, building on several earlier ideas and methods. The main novelty in the paper is the specific way the successor representation is used in defining the abstract states. The approach could prove useful but the analysis in the current paper does not give the reader a good understanding of the behaviour of the approach and its strengths/weaknesses relative to existing methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667051715518}, {"id": "Q0LM-ME5C4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6071/Reviewer_C2ny"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper outlines an approach to use successor representations to drive a discrete state space abstraction. The abstract states are clusters in successor space, so \"nearby\" states have similar successors. The paper also contributes a way to utilize the abstraction learned to solve tasks efficiently by interpreting transitioning between abstract states as options, and outlines an algorithm to jointly learn the successor representations, the abstract representation and the option policies.", "review_text": "This is a good quality paper, cleanly written and justified with appropriate experimentation and I recommend acceptance to ICLR.", "strengths": "This is a well written paper which clearly describes its motivations, contributions as well as limitations. The method is theoretically grounded and seems sound. The approach is novel as far as I am aware and cleanly described. The experiments are well thought-out and appropriate comparisons are made to prior approaches.\nThe environments are a bit simple, but given that prior approaches are less effective/noisy on them, that is justified.\nMore quantitative results could have been included for different exploration policies for learning the success features/abstractions.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper outlines an approach to use successor representations to drive a discrete state space abstraction. The abstract states are clusters in successor space, so \"nearby\" states have similar successors. The paper also contributes a way to utilize the abstraction learned to solve tasks efficiently by interpreting transitioning between abstract states as options, and outlines an algorithm to jointly learn the successor representations, the abstract representation and the option policies.", "strength_and_weaknesses": "This is a well written paper which clearly describes its motivations, contributions as well as limitations. The method is theoretically grounded and seems sound. The approach is novel as far as I am aware and cleanly described. The experiments are well thought-out and appropriate comparisons are made to prior approaches.\nThe environments are a bit simple, but given that prior approaches are less effective/noisy on them, that is justified.\nMore quantitative results could have been included for different exploration policies for learning the success features/abstractions.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and the work is novel as far as I am aware.\nBecause the method is relatively simple and cleanly described it should be reproducible. The authors state that the code will be made available upon acceptance.", "summary_of_the_review": "This is a good quality paper, cleanly written and justified with appropriate experimentation and I recommend acceptance to ICLR.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666819911093}, {"id": "0HZSOtk5ln0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6071/Reviewer_TM9s"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to learn a successor representation and use it to cluster the states. The clusters are then treated as nodes of a graph and an (option) policy is trained to navigate the graph. The graph can be used in either a goal-oriented setting by searching; or a reward maximization setting with a lazy random walk.", "review_text": "This paper is an empirical one where a lot of design choices are based on the authors' intuition. However, the paper didn't provide strong empirical evidence that the method they proposed actually works in tasks with practical value. All the tasks are quite simple (2d navigation and 2d arm) and working on them is more of a proof of concept. There could be so many implicit hypotheses that the method relies on if they can only work on tasks with this low degree of freedom. Even for these toy tasks, it's not clear whether the method has any advantages over well-established control/RL methods due to the lack of baselines.\nI think this paper is not ready for acceptance for ICLR.", "strengths": "Strength:\n1. The idea of learning a graph and using it for hierarchical decision-making is interesting on a high level.\n\nWeaknesses:\n1. Environment of all the tasks is of very low degree of freedom (Dof=2), and the optimal policy is very simple. It's unclear whether the method can be applied to tasks with practical value. \n2. The motivation of a lot of the novel designs (the successor representation loss in eq (2), lazy random walk for reward maximization) is not well-justified and seems like can cause problems in a general setting.\n3. The baselines this method is comparing to in section 6.2 are within their framework. It's more like ablation of representation learning component of DSAA. No proper offline/online, model-based and model-free RL methods are compared to the proposed method.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes to learn a successor representation and use it to cluster the states. The clusters are then treated as nodes of a graph and an (option) policy is trained to navigate the graph. The graph can be used in either a goal-oriented setting by searching; or a reward maximization setting with a lazy random walk.", "strength_and_weaknesses": "Strength:\n1. The idea of learning a graph and using it for hierarchical decision-making is interesting on a high level.\n\nWeaknesses:\n1. Environment of all the tasks is of very low degree of freedom (Dof=2), and the optimal policy is very simple. It's unclear whether the method can be applied to tasks with practical value. \n2. The motivation of a lot of the novel designs (the successor representation loss in eq (2), lazy random walk for reward maximization) is not well-justified and seems like can cause problems in a general setting.\n3. The baselines this method is comparing to in section 6.2 are within their framework. It's more like ablation of representation learning component of DSAA. No proper offline/online, model-based and model-free RL methods are compared to the proposed method.", "clarity,_quality,_novelty_and_reproducibility": "Quality: The quality of the writing and plotting is OK.\nClarity: The writing is somewhat vague on the motivation and the methodology is a bit underspecified (potentially because of the complexity of the method).\nReproducibility: The authors claim the paper will be released to the public but there's no code in the current submission. Since the whole algorithm seems quite complex, I think it may not be very easy to reproduce the results.", "summary_of_the_review": "This paper is an empirical one where a lot of design choices are based on the authors' intuition. However, the paper didn't provide strong empirical evidence that the method they proposed actually works in tasks with practical value. All the tasks are quite simple (2d navigation and 2d arm) and working on them is more of a proof of concept. There could be so many implicit hypotheses that the method relies on if they can only work on tasks with this low degree of freedom. Even for these toy tasks, it's not clear whether the method has any advantages over well-established control/RL methods due to the lack of baselines.\nI think this paper is not ready for acceptance for ICLR.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666631351887}], "openreview_url": "https://openreview.net/forum?id=Krk0Gnft2Zc", "arxiv_id": "2206.03467", "paper_pdf": "papers/Krk0Gnft2Zc.pdf", "paper_pdf_sha256": "384519ddb592fc816f23367572cbe0ec49988cc01190be8cac574b9edae717ee", "paper_pdf_bytes": 1267863, "paper_pdf_source": "openreview", "code_url": "https://github.com/amnonattali/dsaa", "code_repository": "amnonattali/dsaa", "code_commit": "a82a3afb45eb6ba33bdc7f5230d0f275daf205fb", "code_archive": "repos/Krk0Gnft2Zc.zip", "code_archive_sha256": "b6d90f86b31915300fcf546b6034f9edbd64ced93332c6d9d03f35f456caeaf3", "code_archive_bytes": 341352, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 407, "github_languages": {"Python": 199174}, "github_archived": false, "github_pushed_at": "2023-01-11T18:59:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discrete-state-action-abstraction-via-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OgCcfc1m0TO", "year": 2022, "status": "rejected", "title": "Learning to Prompt for Vision-Language Models", "authors": ["Kaiyang Zhou", "Jingkang Yang", "Chen Change Loy", "Ziwei Liu"], "authorids": ["~Kaiyang_Zhou1", "~Jingkang_Yang1", "~Chen_Change_Loy2", "~Ziwei_Liu1"], "authors_source": "OpenReview API", "abstract": "Vision-language pre-training has recently emerged as a promising alternative for representation learning. It shifts from the tradition of using images and discrete labels for learning a fixed set of weights, seen as visual concepts, to aligning images and raw text for two separate encoders. Such a paradigm benefits from a broader source of supervision and allows zero-shot transfer to downstream tasks since visual concepts can be diametrically generated from natural language, known as prompt. In this paper, we identify that a major challenge of deploying such models in practice is prompt engineering. This is because designing a proper prompt, especially for context words surrounding a class name, requires domain expertise and typically takes a significant amount of time for words tuning since a slight change in wording could have a huge impact on performance. Moreover, different downstream tasks require specific designs, further hampering the efficiency of deployment. To overcome this challenge, we propose a novel approach named \\emph{context optimization (CoOp)}. The main idea is to model context in prompts using continuous representations and perform end-to-end learning from data while keeping the pre-trained parameters fixed. In this way, the design of task-relevant prompts can be fully automated. Experiments on 11 datasets show that CoOp effectively turns pre-trained vision-language models into data-efficient visual learners, requiring as few as one or two shots to beat hand-crafted prompts with a decent margin and able to gain significant improvements when using more shots (e.g., at 16 shots the average gain is around 17\\% with the highest reaching over 50\\%). CoOp also exhibits strong robustness to distribution shift.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "iQ0p4tsDGS", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper270/Reviewer_pGwp"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a novel approach named context optimization (CoOp) for prompt engineering of vision-language pre-training models.   The main idea is to model context in prompts using continuous representations and perform end-to-end learning from data while keeping the pre-trained parameters fixed. Experiments on 11 datasets show that CoOp effectively turns pre-trained vision-language models into data-efficient visual learners, requiring as few as one or two shots to beat hand-crafted prompts with a decent margin and able to gain significant improvements when using more shots.", "review_text": "Strengths:\n1. This paper provides a novel approach named CoOp for prompt engineering based on  CLIP. \n2. CLIP-based experiments are sufficient to prove the advantages of CoOp.\n\nWeaknesses:\n1. Only the results based on CLIP are compared, and there are no more experiments of other visual-language models.\n2. It is an improvement of CLIP, and the idea is similar to many existing works such as [1].\n[1] Prefix-Tuning: Optimizing Continuous Prompts for Generation\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a novel approach named context optimization (CoOp) for prompt engineering of vision-language pre-training models.   The main idea is to model context in prompts using continuous representations and perform end-to-end learning from data while keeping the pre-trained parameters fixed. Experiments on 11 datasets show that CoOp effectively turns pre-trained vision-language models into data-efficient visual learners, requiring as few as one or two shots to beat hand-crafted prompts with a decent margin and able to gain significant improvements when using more shots.", "main_review": "Strengths:\n1. This paper provides a novel approach named CoOp for prompt engineering based on  CLIP. \n2. CLIP-based experiments are sufficient to prove the advantages of CoOp.\n\nWeaknesses:\n1. Only the results based on CLIP are compared, and there are no more experiments of other visual-language models.\n2. It is an improvement of CLIP, and the idea is similar to many existing works such as [1].\n[1] Prefix-Tuning: Optimizing Continuous Prompts for Generation\n", "summary_of_the_review": "This is a somewhat novel but solid work. The comparative experiment based on clip is very sufficient, but it also lacks other important experiments, such as the effect of CoOp on other vision-language models, just as the title of the paper is learning to prompt for vision language models", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635930182488}, {"id": "uLmfK84ybVF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper270/Reviewer_1WJt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes context optimization (CoOp) which learns task-aware continuous prompts to improve CLIP in terms of few-shot image classification. By fixing the pretrained backbone, CoOp performs end-to-end learning to update the learnable context vectors for target domain datasets. The simple yet effective approach substantially beats hand-crafted prompts with a large margin. Meanwhile, CoOp also exhibits better robustness to distribution shift than CLIP.", "review_text": "Strengths: \n1) The paper is generally well-written and easy to follow. The author conducts extensive experiments and ablation studies.\n2) CoOp outperforms CLIP on all 11 diverse datasets, especially showing large improvements in specific domains like texture and satellite images.\n3) CoOp is data-efficient and can boost the classification performance with only a few shots of training data.\n4) CoOp demonstrates better robustness to distribution shift than zero-shot CLIP and linear probe CLIP. \n\nWeakness: \n1) Though effective, the technical novelty of this paper is quite limited. Since the soft prompt tuning approaches (e.g., Prompt Tuning [1] and P-Tuning [2]) have already been proposed in NLP domain. And CoOp adopts a very similar technique.\n[1] Lester, Brian, Rami Al-Rfou, and Noah Constant. “The power of scale for parameter-efficient prompt tuning.” EMNLP 2021.\n[2] Liu, Xiao, et al. “GPT Understands, Too.” *arXiv preprint arXiv:2103.10385* (2021).\n\n2) According to Table 4 in the paper, the nearest neighbor words of the learned context vectors rarely have practical semantic meaning. This casts doubt on using the word “context”. From this perspective, these soft prompts are just more parameters to improve the model capacity.  Hence, can we say that CoOp is just a better version of fine-tuning? In other words, there might be a similar fine-tuning way to utilize additional parameters to get better performance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes context optimization (CoOp) which learns task-aware continuous prompts to improve CLIP in terms of few-shot image classification. By fixing the pretrained backbone, CoOp performs end-to-end learning to update the learnable context vectors for target domain datasets. The simple yet effective approach substantially beats hand-crafted prompts with a large margin. Meanwhile, CoOp also exhibits better robustness to distribution shift than CLIP.", "main_review": "Strengths: \n1) The paper is generally well-written and easy to follow. The author conducts extensive experiments and ablation studies.\n2) CoOp outperforms CLIP on all 11 diverse datasets, especially showing large improvements in specific domains like texture and satellite images.\n3) CoOp is data-efficient and can boost the classification performance with only a few shots of training data.\n4) CoOp demonstrates better robustness to distribution shift than zero-shot CLIP and linear probe CLIP. \n\nWeakness: \n1) Though effective, the technical novelty of this paper is quite limited. Since the soft prompt tuning approaches (e.g., Prompt Tuning [1] and P-Tuning [2]) have already been proposed in NLP domain. And CoOp adopts a very similar technique.\n[1] Lester, Brian, Rami Al-Rfou, and Noah Constant. “The power of scale for parameter-efficient prompt tuning.” EMNLP 2021.\n[2] Liu, Xiao, et al. “GPT Understands, Too.” *arXiv preprint arXiv:2103.10385* (2021).\n\n2) According to Table 4 in the paper, the nearest neighbor words of the learned context vectors rarely have practical semantic meaning. This casts doubt on using the word “context”. From this perspective, these soft prompts are just more parameters to improve the model capacity.  Hence, can we say that CoOp is just a better version of fine-tuning? In other words, there might be a similar fine-tuning way to utilize additional parameters to get better performance.", "summary_of_the_review": "The paper achieves better few-shot image classification performance improvements but lacks enough technical novelty and explanations for learned prompts", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635920943009}, {"id": "mo5m2Y33ekj", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper270/Reviewer_AYDN"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors demonstrate a more efficient form of few-shot learning\nusing CLIP compared to linear probing for image classification: CoOp.\nInstead of fine-tuning a small linear classifier on the output of\nCLIP, they propose fine-tuning a number of additional embeddings at\nthe input layer; this modification, in theory, allows CLIP to leverage\nmore computation when adapting to tasks, while still only optimizing a\nsmall number of parameters. Experiments across several corpora\ndemonstrate the efficacy of the approach, which generally yields a\nfew accuracy points of gain versus a linear probe trained on the same\namount of data.", "review_text": "The paper's motivation is clear, experiments thorough, and results\nconvincing. In addition to the standard few-shot evaluations, I\nappreciated the authors considering the distribution shift scenario:\nin that case, they show that their model can more effectively leverage\nlabelled data from a different dataset versus the linear probes (even\nthough zero-shot CLIP remains somewhat competitive in that\nregime). Additional experiments about the optimal context length,\nplacement of fine-tune-able token embeddings, and vision backbones\nwhere appreciated.\n\nWhile the CoOp method appears to work well, but I had some concerns\nabout novelty. In particular, this method is more-or-less identical to\nLi and Liang (2021) --- that paper came out on arXiv Jan 1 of this\nyear, and was published at ACL earlier this year. In addition, this\nidea has been \"rediscovered\" in the NLP context several times (which\nare also cited, but perhaps cannot be considered contemporaneous,\ngiven their arXiv dates...). While the authors cite the arXiv version,\nI think that, given the timing, the authors should perhaps not pitch\ntheir method as a \"novel\" approach (as is done in the abstract);\nrather, this seems to be applying prefix-tuning to CLIP.\n\nI had a few technical concerns:\n\n1. there are a lack of details for the linear probe --- how was the\n   regularization parameter chosen?\n   \n2. The baseline for comparison to CoOp was the linear probe, which\n   makes sense. However, I would have also liked to have seen a\n   comparison where all the parameters of CLIP are fine-tuned --- how\n   well does that work for few shot learning?\n\nI also had a few presentation concerns:\n\n1. Figure 1 compares a supervised CoOp method to a zero-shot CLIP\n   baseline. While the CoOp method is \"few shot\", in this figure,\n   there are 16 examples per class provided, which, for some datasets\n   may amount to hundreds or thousands of labelled examples. I would\n   have appreciated Figure 1 comparing to the linear probe.\n\n2. Figure 5b has a similar problem with being potentially misleading:\n   I would recommend including not only zero-shot CLIP, but also\n   linear probe CLIP.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors demonstrate a more efficient form of few-shot learning\nusing CLIP compared to linear probing for image classification: CoOp.\nInstead of fine-tuning a small linear classifier on the output of\nCLIP, they propose fine-tuning a number of additional embeddings at\nthe input layer; this modification, in theory, allows CLIP to leverage\nmore computation when adapting to tasks, while still only optimizing a\nsmall number of parameters. Experiments across several corpora\ndemonstrate the efficacy of the approach, which generally yields a\nfew accuracy points of gain versus a linear probe trained on the same\namount of data.", "main_review": "The paper's motivation is clear, experiments thorough, and results\nconvincing. In addition to the standard few-shot evaluations, I\nappreciated the authors considering the distribution shift scenario:\nin that case, they show that their model can more effectively leverage\nlabelled data from a different dataset versus the linear probes (even\nthough zero-shot CLIP remains somewhat competitive in that\nregime). Additional experiments about the optimal context length,\nplacement of fine-tune-able token embeddings, and vision backbones\nwhere appreciated.\n\nWhile the CoOp method appears to work well, but I had some concerns\nabout novelty. In particular, this method is more-or-less identical to\nLi and Liang (2021) --- that paper came out on arXiv Jan 1 of this\nyear, and was published at ACL earlier this year. In addition, this\nidea has been \"rediscovered\" in the NLP context several times (which\nare also cited, but perhaps cannot be considered contemporaneous,\ngiven their arXiv dates...). While the authors cite the arXiv version,\nI think that, given the timing, the authors should perhaps not pitch\ntheir method as a \"novel\" approach (as is done in the abstract);\nrather, this seems to be applying prefix-tuning to CLIP.\n\nI had a few technical concerns:\n\n1. there are a lack of details for the linear probe --- how was the\n   regularization parameter chosen?\n   \n2. The baseline for comparison to CoOp was the linear probe, which\n   makes sense. However, I would have also liked to have seen a\n   comparison where all the parameters of CLIP are fine-tuned --- how\n   well does that work for few shot learning?\n\nI also had a few presentation concerns:\n\n1. Figure 1 compares a supervised CoOp method to a zero-shot CLIP\n   baseline. While the CoOp method is \"few shot\", in this figure,\n   there are 16 examples per class provided, which, for some datasets\n   may amount to hundreds or thousands of labelled examples. I would\n   have appreciated Figure 1 comparing to the linear probe.\n\n2. Figure 5b has a similar problem with being potentially misleading:\n   I would recommend including not only zero-shot CLIP, but also\n   linear probe CLIP.", "summary_of_the_review": "Overall: the work is generally clear with thorough and convincing\nexperiments. While there were a few presentation/technical concerns,\nthe main drawback of this work is novelty: the proposed idea is\nidentical to Li and Liang (2021) [arxiv in january], Zhong et\nal. (2021) [arxiv in April], and perhaps Lester et al (2021) [arxiv in\nSep, this one is more recent and I haven't read it yet]. While these\npapers consider only the NLP case and not the vision+language case,\nit's still difficult to call this method \"novel,\" as the authors do in\nthe intro.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635354950579}, {"id": "CwHrJmanoON", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper270/Reviewer_KMdr"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a context optimisation (CoOp) approach that can automate prompt engineering and allow more efficient and task-specific transfer for pretrained vision-language models.", "review_text": "+ Extensive experiments\n+ Motivation for the automatic prompt is useful\n\no Investment in Fig 1 is super useful. It is helpful to see how fragile the prompt is affected by word twisting. However, the paper claimed prompt engineering is time-consuming, which I do not disagree with. Yes, human annotators have no way to know whether adding a \"flower\" or changing the sentence order will improve or harm the performance. But the key motivation of the prompt is to reduce the human working load and adding some general description is not super complicated. The task should be down to exploring more robust models and algorithms that can withstand these minor changes in my opinion.\n\n- The biggest concern is about the lack of theoretical contribution. The main content of the methodology is broken down into two sections. The first section is just a review of CLIP which is already well-known. The key content of context optimisation in 2.2 is just half a page. The description is rough.\n\n- It is not super clear why adding extra dimensions (claimed as the same number in that of word embedding) to the class token can help the prompt.\n\n- On page 4 above Eq.3, the concept \"unified context\" and \"class-specific context\" are not well explained.\n\n- paragraph under Eq.3, \"Training is performed ...\" is redundant and not informative.\n\n- Figure 2 is not helpful in understanding the framework and looks very low-quality.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a context optimisation (CoOp) approach that can automate prompt engineering and allow more efficient and task-specific transfer for pretrained vision-language models.", "main_review": "+ Extensive experiments\n+ Motivation for the automatic prompt is useful\n\no Investment in Fig 1 is super useful. It is helpful to see how fragile the prompt is affected by word twisting. However, the paper claimed prompt engineering is time-consuming, which I do not disagree with. Yes, human annotators have no way to know whether adding a \"flower\" or changing the sentence order will improve or harm the performance. But the key motivation of the prompt is to reduce the human working load and adding some general description is not super complicated. The task should be down to exploring more robust models and algorithms that can withstand these minor changes in my opinion.\n\n- The biggest concern is about the lack of theoretical contribution. The main content of the methodology is broken down into two sections. The first section is just a review of CLIP which is already well-known. The key content of context optimisation in 2.2 is just half a page. The description is rough.\n\n- It is not super clear why adding extra dimensions (claimed as the same number in that of word embedding) to the class token can help the prompt.\n\n- On page 4 above Eq.3, the concept \"unified context\" and \"class-specific context\" are not well explained.\n\n- paragraph under Eq.3, \"Training is performed ...\" is redundant and not informative.\n\n- Figure 2 is not helpful in understanding the framework and looks very low-quality.", "summary_of_the_review": "Despite extensive evaluation, the paper presents in very low quality and lack of description of key methods and the theory behind it. The overall quality is clearly below the threshold of the ICLR standard.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No", "recommendation": "1: strong reject", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634859356975}], "openreview_url": "https://openreview.net/forum?id=OgCcfc1m0TO", "arxiv_id": "2109.01134", "paper_pdf": "papers/OgCcfc1m0TO.pdf", "paper_pdf_sha256": "9fa2eb3a60029d39448d790e4cd660d8c69bae7ce18b3a270fd2c5705be08ab2", "paper_pdf_bytes": 2147082, "paper_pdf_source": "openreview", "code_url": "https://github.com/KaiyangZhou/CoOp", "code_repository": "KaiyangZhou/CoOp", "code_commit": "ff61507c790454bce7c5052c3ac39e60772f1f89", "code_archive": "repos/OgCcfc1m0TO.zip", "code_archive_sha256": "c3c3d7a6b0d2d5e63ec0f561ed06554c77c5f56f9c1715c9753aea2861825fec", "code_archive_bytes": 1438659, "code_file_count": 38, "code_extensions": {".py": 29, ".sh": 9}, "github_disk_usage_kb": 1442, "github_languages": {"Python": 134456, "Shell": 5875}, "github_archived": false, "github_pushed_at": "2024-05-20T16:58:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-prompt-for-vision-language-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oev4KdikGjy", "year": 2021, "status": "rejected", "title": "FMix: Enhancing Mixed Sample Data Augmentation", "authors": ["Ethan Harris", "Antonia Marcu", "Matthew Painter", "Mahesan Niranjan", "Adam Prugel-Bennett", "Jonathon Hare"], "authorids": ["~Ethan_Harris1", "am1g15@ecs.soton.ac.uk", "mp2u16@ecs.soton.ac.uk", "~Mahesan_Niranjan1", "~Adam_Prugel-Bennett1", "~Jonathon_Hare1"], "authors_source": "OpenReview API", "abstract": "Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and CutMix. We analyse MSDA from an information theoretic perspective, characterising learned models in terms of how they impact the models’ perception of the data.  Ultimately, our analyses allow us to decouple two complementary properties of augmentations that are useful for reasoning about MSDA. From insight on the efficacy of CutMix in particular, we subsequently propose FMix, an MSDA that uses binary masks obtained by applying a threshold to low frequency images sampled from Fourier space.  FMix improves performance over MixUp and CutMix for a number of models across a range of data sets and problem settings,  obtaining new state-of-the-art results on CIFAR-10 and Fashion-MNIST.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "dDAQqn0WkeE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3825/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this work, authors provide an analysis of mutual information for MSDA and the develop a new variant of mixup. The effectiveness is demonstrated by experiments.\n\nStrength\n1.\tAuthors study the difference between masking MSDA and interpolative MSDA, which is helpful for understanding the power of mixup and its variants.\n2.\tThey develop a new augmentation method and improve the performance of masking MSDA.\n\nWeakness\n1.\tThe proposed measurement is not helpful for designing new methods. Note that the mutual information in mixup is lower than baseline while mixup still outperforms baseline.\n2.\tCompared to mixup and cutmix, the improvement reported in Table 2 is marginal.\n3.\tThe experiments on ImageNet is unconvincing. Both of mixup and cutmix are worse than baseline, which contradicts the existing results.\n4.\tThere lacks the discussion for the saliency based mixup methods, e.g., Puzzle Mix [1]. It is closely related to fmix but equipped with a learnable strategy to obtain patches for mixing.\n\n[1] J-H Kim, et al. Puzzle mix: Exploiting saliency and local statistics for optimal mixup", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A new variant of cutmix but the improvement is marginal", "review": "In this work, authors provide an analysis of mutual information for MSDA and the develop a new variant of mixup. The effectiveness is demonstrated by experiments.\n\nStrength\n1.\tAuthors study the difference between masking MSDA and interpolative MSDA, which is helpful for understanding the power of mixup and its variants.\n2.\tThey develop a new augmentation method and improve the performance of masking MSDA.\n\nWeakness\n1.\tThe proposed measurement is not helpful for designing new methods. Note that the mutual information in mixup is lower than baseline while mixup still outperforms baseline.\n2.\tCompared to mixup and cutmix, the improvement reported in Table 2 is marginal.\n3.\tThe experiments on ImageNet is unconvincing. Both of mixup and cutmix are worse than baseline, which contradicts the existing results.\n4.\tThere lacks the discussion for the saliency based mixup methods, e.g., Puzzle Mix [1]. It is closely related to fmix but equipped with a learnable strategy to obtain patches for mixing.\n\n[1] J-H Kim, et al. Puzzle mix: Exploiting saliency and local statistics for optimal mixup", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603935083285}, {"id": "ND9Xrp8upVc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3825/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a new mixup method that builds masks by first sampling a grey-scale mask from fourier space, which is subsequently transformed into a binary mask. This improves results against several baselines and achieves state-of-the-art on a few important vision benchmarks.\n\nMy first remark is about the masking. The procedure seems fine, but why not compare to the way masks are sampled in context encoders [1]. This seems like an important baseline masking method to compare to. In addition, one could try sampling masks from a standard segmentation model, e.g., R-CNN.\n\nMy second remark is about the MI bounds. On page 14 in the Appendix, you state that the MI between Z_A and X_hat is approximately equal to the KL divergence between the posterior and the normal distribution, but in general this wont be true as in training the Gaussian mixture p_Za wont match the normal distribution, so you have an upper bound. So you have a lower bound of an upper bound to the MI, not a lower bound.\n\nI'm curious though why not just use one of the recent neural estimators, e.g., found in [2] or [3]. In general, using VAEs for MI estimators depends heavily on the quality of the generator, so these neural estimators might be better suited.\n\nOther comments:\nP1:\n* \"'post-processing cannot increase information'\" if such processing is deterministic, no?\n\nP2:\n* \"CutMix imposes an unnecessary limitation\": what limitation? Could you clarify?\n\nFinally, do you plan to have updated results that compare to the 1024 batch size / 300 epoch settings?\n\n[1] Pathak, Deepak, et al. \"Context encoders: Feature learning by inpainting.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.\n[2] Belghazi, Mohamed Ishmael, et al. \"Mine: mutual information neural estimation.\" arXiv preprint arXiv:1801.04062 (2018).\n[3] Poole, Ben, et al. \"On variational bounds of mutual information.\" arXiv preprint arXiv:1905.06922 (2019).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review: FMix: Enhancing Mixed Sample Data Augmentation", "review": "This paper introduces a new mixup method that builds masks by first sampling a grey-scale mask from fourier space, which is subsequently transformed into a binary mask. This improves results against several baselines and achieves state-of-the-art on a few important vision benchmarks.\n\nMy first remark is about the masking. The procedure seems fine, but why not compare to the way masks are sampled in context encoders [1]. This seems like an important baseline masking method to compare to. In addition, one could try sampling masks from a standard segmentation model, e.g., R-CNN.\n\nMy second remark is about the MI bounds. On page 14 in the Appendix, you state that the MI between Z_A and X_hat is approximately equal to the KL divergence between the posterior and the normal distribution, but in general this wont be true as in training the Gaussian mixture p_Za wont match the normal distribution, so you have an upper bound. So you have a lower bound of an upper bound to the MI, not a lower bound.\n\nI'm curious though why not just use one of the recent neural estimators, e.g., found in [2] or [3]. In general, using VAEs for MI estimators depends heavily on the quality of the generator, so these neural estimators might be better suited.\n\nOther comments:\nP1:\n* \"'post-processing cannot increase information'\" if such processing is deterministic, no?\n\nP2:\n* \"CutMix imposes an unnecessary limitation\": what limitation? Could you clarify?\n\nFinally, do you plan to have updated results that compare to the 1024 batch size / 300 epoch settings?\n\n[1] Pathak, Deepak, et al. \"Context encoders: Feature learning by inpainting.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.\n[2] Belghazi, Mohamed Ishmael, et al. \"Mine: mutual information neural estimation.\" arXiv preprint arXiv:1801.04062 (2018).\n[3] Poole, Ben, et al. \"On variational bounds of mutual information.\" arXiv preprint arXiv:1905.06922 (2019).", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603928436742}, {"id": "OM9oOeLT16", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3825/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an advanced masking strategy for CutMix augmentation based on the low-pass filter. The authors provide an interesting mutual information analysis for different augmentation strategies to describe their motivation. The experiments include many vision tasks (CIFAR-10, CIFAR-100, Fashion-MNIST, Tiny-ImageNet, ImageNet, Bengali datasets) and language tasks (Toxic, IMDb, Yelp).\n\n**Pros**\n\n\\+ The mutual information analysis provides us a new perspective to understand different data augmentations.\n\n\\+ Various experiments.\n\n**Cons**\n\n**[Contradictory results between mutual information and performances]**\nIf we believe the VAE experiments in section 3, we have another paradox: the mutual information measurement and the real performance are not related.\nTable 1 shows that in terms of mutual information, MixUp < Baseline < CutMix (and < FMix with a very small gap).\nHowever, many experiments in this paper show that baseline < mixup < cutmix in terms of the performances.\nThis paper cites information bottleneck theory to justify the deceases shown by Mixup, but it is still contradictory to the performances.\nIt makes me confused to understand the meaning of mutual information. What is good for an augmentation method if we have high or low mutual information? It is still unclear to me.\n\nA similar comment also can be applicable to the \"adversarial robustness\" experiments. Aside from that mixup is hard to say \"adversarial training\" (what is the threat model in this scenario?), I feel that this result is irrelevant to FMix motivation.\n\n\n**[Weak logical connection between the motivation and the method]**\nIn my opinion, the connection between the analysis in the motivation and the proposed method is too weak. This paper proposes a CutMix variant where the mask is sampled by a low-pass filter. Why the low-pass filter approach can solve the motivation, i.e., enhancing mutual information between input and augmented images? There could be other possible variants as discussed in my \"related works\" comment\n\n\n**[Related works]**\nThere are a few CutMix variants that employ a non-random masking strategy. Especially, I believe these two variants, which have similar motivation, should be compared:\n\n- Walawalkar, Devesh, et al. \"Attentive Cutmix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification.\" ICASSP 2020\n- Kim, Jang-Hyun, Wonho Choo, and Hyun Oh Song. \"Puzzle mix: Exploiting saliency and local statistics for optimal mixup.\" ICML 2020\n\nwhere Attentive CutMix uses CAM to extract masks, and PuzzleMix employs an optimization problem to optimize masks.\nIf it is possible, please provide more comparison between these two papers.\n\n\n**[Too small performance gap, less convincing experiments]**\nIn Table 2, the performance gap between FMix and CutMix is too small, usually less than 0.3%. Note that the performance gaps are almost neglectable in these tasks.\n\nFurthermore, FMix is often worse than CutMix in many tasks (Table 2 TinyImageNet, Table 3 ImageNet-A, Table 4, CIFAR-10H Table 6). I wonder what is the advantage to use FMix comparing to CutMix if FMix shows worse performance than CutMix.\n\nEspecially, I believe Table 3 is problematic. This paper argues that \"Mixup uses 1024 batch size and CutMix uses 300 epochs\". However, in PuzzleMix Table 5, CutMix-trained ResNet50 (top1 err 22.92) outperforms baseline ResNet50 (top1 err 24.31) with only 100 epochs. Thus, to me, this table is not convincing enough.\n\n\n**[Potential issues in VAE analysis]**\nThe mutual information analysis is heavily relying on the learned VAE model. I wonder the quality of the generated images by VAE, in terms of both qualitatively (please provide generated samples in the supplementary) and quantitatively (e.g., FID).\nIf the VAE is not optimized well, the analysis will not be convincing enough.\n\n\n**Minor comments**\n- Why CutMix experiments are missed in Table 5?\n- I suggest avoiding using the words, \"clear\" and \"clearly\".\n\n---\n\n**Post-rebuttal update**\n\nMy main concerns in the initial review were three-folds:\n\n- Potential flaws in the analyses based on VAE and adversarial attacks\n- Unclear connection between the MI analysis and the proposed method\n- Small performance gap, and even sometimes worse performance, compared to the baseline methods (Mixup, CutMix)\n\nAfter having discussions with the authors, I will keep my initial score because:\n\n- I am still confused about the MI-based analysis conclusion. The authors mentioned *\"We make no claim that increasing or decreasing the mutual information measure will have a strong impact on performance. Instead, we contend that MixUp works by forcing the model to ignore sample specific features (thus learning compressed representations – the reason for discussing the information bottleneck theory) and that CutMix works by mimicking the real data whilst preventing example memorization.\"* in the rebuttal, but these two conclusions are not trivial to me (by the MI analysis).\n- Even if we ignore the first part, my second concern still remains. The authors mentioned *\"That is the problem FMix tries to solve by removing the horizontal and vertical edge artefacts from cutmix. Our belief is that cutmix biases models towards these edges as they are a guaranteed feature of the data and learning about them would reduce the loss since these edges can tell you how much of each source image is present in the input (a key part of the objective).\"*. But if this paper assumes that the rectangle masking strategy of CutMix makes bias, then I think other CutMix variants such as AttentiveCutMix or PuzzleMix should be considered as the comparison methods. Hence, I disagree with this statement *\"A comparison to masks generated using additional models (and, thus, significant additional computation) does not seem fair to us.\"*\n- For my last concern, the small performance gap, the authors claimed that this method *\"was also used by the second place team in the BengaliAI Kaggle competition\"*. It is good evidence that FMix can sometimes offer benefit to real-world applications, but I think more evidence that FMix can really solve problems of previous MSDA in a certain scenario, e.g., the edge bias as pointed by the authors.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Weak logical connection between the motivation and the method, too small performance gap", "review": "This paper proposes an advanced masking strategy for CutMix augmentation based on the low-pass filter. The authors provide an interesting mutual information analysis for different augmentation strategies to describe their motivation. The experiments include many vision tasks (CIFAR-10, CIFAR-100, Fashion-MNIST, Tiny-ImageNet, ImageNet, Bengali datasets) and language tasks (Toxic, IMDb, Yelp).\n\n**Pros**\n\n\\+ The mutual information analysis provides us a new perspective to understand different data augmentations.\n\n\\+ Various experiments.\n\n**Cons**\n\n**[Contradictory results between mutual information and performances]**\nIf we believe the VAE experiments in section 3, we have another paradox: the mutual information measurement and the real performance are not related.\nTable 1 shows that in terms of mutual information, MixUp < Baseline < CutMix (and < FMix with a very small gap).\nHowever, many experiments in this paper show that baseline < mixup < cutmix in terms of the performances.\nThis paper cites information bottleneck theory to justify the deceases shown by Mixup, but it is still contradictory to the performances.\nIt makes me confused to understand the meaning of mutual information. What is good for an augmentation method if we have high or low mutual information? It is still unclear to me.\n\nA similar comment also can be applicable to the \"adversarial robustness\" experiments. Aside from that mixup is hard to say \"adversarial training\" (what is the threat model in this scenario?), I feel that this result is irrelevant to FMix motivation.\n\n\n**[Weak logical connection between the motivation and the method]**\nIn my opinion, the connection between the analysis in the motivation and the proposed method is too weak. This paper proposes a CutMix variant where the mask is sampled by a low-pass filter. Why the low-pass filter approach can solve the motivation, i.e., enhancing mutual information between input and augmented images? There could be other possible variants as discussed in my \"related works\" comment\n\n\n**[Related works]**\nThere are a few CutMix variants that employ a non-random masking strategy. Especially, I believe these two variants, which have similar motivation, should be compared:\n\n- Walawalkar, Devesh, et al. \"Attentive Cutmix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification.\" ICASSP 2020\n- Kim, Jang-Hyun, Wonho Choo, and Hyun Oh Song. \"Puzzle mix: Exploiting saliency and local statistics for optimal mixup.\" ICML 2020\n\nwhere Attentive CutMix uses CAM to extract masks, and PuzzleMix employs an optimization problem to optimize masks.\nIf it is possible, please provide more comparison between these two papers.\n\n\n**[Too small performance gap, less convincing experiments]**\nIn Table 2, the performance gap between FMix and CutMix is too small, usually less than 0.3%. Note that the performance gaps are almost neglectable in these tasks.\n\nFurthermore, FMix is often worse than CutMix in many tasks (Table 2 TinyImageNet, Table 3 ImageNet-A, Table 4, CIFAR-10H Table 6). I wonder what is the advantage to use FMix comparing to CutMix if FMix shows worse performance than CutMix.\n\nEspecially, I believe Table 3 is problematic. This paper argues that \"Mixup uses 1024 batch size and CutMix uses 300 epochs\". However, in PuzzleMix Table 5, CutMix-trained ResNet50 (top1 err 22.92) outperforms baseline ResNet50 (top1 err 24.31) with only 100 epochs. Thus, to me, this table is not convincing enough.\n\n\n**[Potential issues in VAE analysis]**\nThe mutual information analysis is heavily relying on the learned VAE model. I wonder the quality of the generated images by VAE, in terms of both qualitatively (please provide generated samples in the supplementary) and quantitatively (e.g., FID).\nIf the VAE is not optimized well, the analysis will not be convincing enough.\n\n\n**Minor comments**\n- Why CutMix experiments are missed in Table 5?\n- I suggest avoiding using the words, \"clear\" and \"clearly\".\n\n---\n\n**Post-rebuttal update**\n\nMy main concerns in the initial review were three-folds:\n\n- Potential flaws in the analyses based on VAE and adversarial attacks\n- Unclear connection between the MI analysis and the proposed method\n- Small performance gap, and even sometimes worse performance, compared to the baseline methods (Mixup, CutMix)\n\nAfter having discussions with the authors, I will keep my initial score because:\n\n- I am still confused about the MI-based analysis conclusion. The authors mentioned *\"We make no claim that increasing or decreasing the mutual information measure will have a strong impact on performance. Instead, we contend that MixUp works by forcing the model to ignore sample specific features (thus learning compressed representations – the reason for discussing the information bottleneck theory) and that CutMix works by mimicking the real data whilst preventing example memorization.\"* in the rebuttal, but these two conclusions are not trivial to me (by the MI analysis).\n- Even if we ignore the first part, my second concern still remains. The authors mentioned *\"That is the problem FMix tries to solve by removing the horizontal and vertical edge artefacts from cutmix. Our belief is that cutmix biases models towards these edges as they are a guaranteed feature of the data and learning about them would reduce the loss since these edges can tell you how much of each source image is present in the input (a key part of the objective).\"*. But if this paper assumes that the rectangle masking strategy of CutMix makes bias, then I think other CutMix variants such as AttentiveCutMix or PuzzleMix should be considered as the comparison methods. Hence, I disagree with this statement *\"A comparison to masks generated using additional models (and, thus, significant additional computation) does not seem fair to us.\"*\n- For my last concern, the small performance gap, the authors claimed that this method *\"was also used by the second place team in the BengaliAI Kaggle competition\"*. It is good evidence that FMix can sometimes offer benefit to real-world applications, but I think more evidence that FMix can really solve problems of previous MSDA in a certain scenario, e.g., the edge bias as pointed by the authors.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603880189684}, {"id": "3u8821TDeo1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3825/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents an interesting analysis of CutMix and MixUp data augmentation techniques. It also presents an improvement to CutMix that removes the horizontal/vertical axis bias. The idea to use fourier noise to construct masks for a variant of CutMix is interesting and well-motivated.\n\nMy main concern is whether the conclusions drawn by the analyses are fully grounded. The paper performs an analysis of the effect of the augmented data on learned representations by training unsupervised models on the augmented or clean data and measuring their mutual information. This analysis has the undesirable property of not matching the supervised case in a number of ways, such as different learning objectives, model architectures, etc. Even ignoring this, if we take the result that “MixUp consistently reduces the amount of information that is learned about the original data”, what then explains the improved generalization accuracy MixUp showcases in their original paper?\n\nMoreover, after claiming that the analysis indicates that MixUp learns different representations, the paper asks “whether these different representations learned from MixUp give rise to practical differences other than just improved generalisation.” The issue is that they do this via an adversarial attack analysis, rather than a more realistic non-worst-case robustness analysis. This leads to the conclusion “MixUp (...) does not correspond to a general increase in robustness.” But it does not answer the original question of whether “MixUp gives rise to practical differences other than just improved generalisation.” The finding that MixUp yields greater ImageNet-A robustness (presented later in the paper) also contradicts this early claim.\n\nThe finding that MixUp provides more compressed representations does not necessarily mean that masking augmentation methods are better than interpolation ones. The paper seems to acknowledge this as well, in the final paragraph of the introduction, where it describes an experiment in which combining FMix+MixUp gives the best results (presumably because their representations of data are different and therefore combining them would yield the best of both worlds). This seems to contradict the previous adversarial analysis in which MixUp was found to not yield significantly more robustness. Further, the combination experiment has the two leading combination methods (FMix+MixUp and CutMix+MixUp) yield very similar results (within the margin of error), which opens the question of whether FMix meaningfully improves over CutMix.\n\nOverall, I find the paper very easy to read and presenting some interesting ideas and even some exciting improvements in performance. I just wish the presentation and the claims made in the analysis of MSDA methods accounted for some of the inconsistencies described above.\n\nUpdate after rebuttal: I appreciate the authors' response and clarifications. I maintain my original score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good paper, but concerns in the conclusions from the analyses", "review": "The paper presents an interesting analysis of CutMix and MixUp data augmentation techniques. It also presents an improvement to CutMix that removes the horizontal/vertical axis bias. The idea to use fourier noise to construct masks for a variant of CutMix is interesting and well-motivated.\n\nMy main concern is whether the conclusions drawn by the analyses are fully grounded. The paper performs an analysis of the effect of the augmented data on learned representations by training unsupervised models on the augmented or clean data and measuring their mutual information. This analysis has the undesirable property of not matching the supervised case in a number of ways, such as different learning objectives, model architectures, etc. Even ignoring this, if we take the result that “MixUp consistently reduces the amount of information that is learned about the original data”, what then explains the improved generalization accuracy MixUp showcases in their original paper?\n\nMoreover, after claiming that the analysis indicates that MixUp learns different representations, the paper asks “whether these different representations learned from MixUp give rise to practical differences other than just improved generalisation.” The issue is that they do this via an adversarial attack analysis, rather than a more realistic non-worst-case robustness analysis. This leads to the conclusion “MixUp (...) does not correspond to a general increase in robustness.” But it does not answer the original question of whether “MixUp gives rise to practical differences other than just improved generalisation.” The finding that MixUp yields greater ImageNet-A robustness (presented later in the paper) also contradicts this early claim.\n\nThe finding that MixUp provides more compressed representations does not necessarily mean that masking augmentation methods are better than interpolation ones. The paper seems to acknowledge this as well, in the final paragraph of the introduction, where it describes an experiment in which combining FMix+MixUp gives the best results (presumably because their representations of data are different and therefore combining them would yield the best of both worlds). This seems to contradict the previous adversarial analysis in which MixUp was found to not yield significantly more robustness. Further, the combination experiment has the two leading combination methods (FMix+MixUp and CutMix+MixUp) yield very similar results (within the margin of error), which opens the question of whether FMix meaningfully improves over CutMix.\n\nOverall, I find the paper very easy to read and presenting some interesting ideas and even some exciting improvements in performance. I just wish the presentation and the claims made in the analysis of MSDA methods accounted for some of the inconsistencies described above.\n\nUpdate after rebuttal: I appreciate the authors' response and clarifications. I maintain my original score.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603836246961}], "openreview_url": "https://openreview.net/forum?id=oev4KdikGjy", "arxiv_id": "2002.12047", "paper_pdf": "papers/oev4KdikGjy.pdf", "paper_pdf_sha256": "cdfbd4c6cfe9b56067cdd8b6ec6f591ea6bc33ef3fb497ed60a8aeecf83d46db", "paper_pdf_bytes": 602932, "paper_pdf_source": "openreview", "code_url": "https://github.com/ecs-vlc/FMix", "code_repository": "ecs-vlc/FMix", "code_commit": "e5991dca018882734c8ea63599f10dfbe67fa0ae", "code_archive": "repos/oev4KdikGjy.zip", "code_archive_sha256": "162c921728edae993a503bdf6715d944b6521484cdc7a97cba7c2cc9874056d2", "code_archive_bytes": 1008157, "code_file_count": 59, "code_extensions": {".py": 49, ".sh": 8, ".ipynb": 2}, "github_disk_usage_kb": 1170, "github_languages": {"Jupyter Notebook": 449087, "Python": 245318, "Shell": 3473}, "github_archived": false, "github_pushed_at": "2021-03-26T19:00:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-and-enhancing-mixed-sample-data"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ryg7vA4tPB", "year": 2020, "status": "rejected", "title": "Rigging the Lottery: Making All Tickets Winners", "authors": ["Utku Evci", "Erich Elsen", "Pablo Castro", "Trevor Gale"], "authorids": ["ue225@nyu.edu", "eriche@google.com", "tgale@google.com", "psc@google.com"], "authors_source": "OpenReview API", "abstract": "Sparse neural networks have been shown to yield computationally efficient networks with improved inference times.  There is a large body of work on training dense networks to yield sparse networks for inference (Molchanov et al., 2017;Zhu & Gupta, 2018; Louizos et al., 2017; Li et al., 2016; Guo et al., 2016).  This limits the size of the largest trainable sparse model to that of the largest trainable dense model. In this paper we introduce a method to train sparse neural networks with a fixed parameter count and a fixed computational cost throughout training, without sacrificing accuracy relative to existing dense-to-sparse training methods. Our method updates the topology of the network during training by using parameter magnitudes and infrequent gradient calculations. We show that this approach requires less floating-point operations (FLOPs) to achieve a given level of accuracy compared to prior techniques. We demonstrate state-of-the-art sparse training results with ResNet-50, MobileNet v1 and MobileNet v2 on the ImageNet-2012 dataset. Finally,  we  provide  some  insights  into  why  allowing  the  topology  to change during the optimization can overcome local minima encountered when the topology remains static.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "Bye-WT3W9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1169/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a pruning technique called RigL, which performs sparse initialization of the network weights, and allows the network to grow weights during training. The sparse initialization makes the pruning algorithm to be memory- and computation- efficient, unlike existing models that starts with dense network weights. The authors train RigL with three different sparsity distributions that considers the number of input and output nodes, and validate them on various deep convnets for training on ImageNet, on which the model outperforms several existing sparsification methods and even dense counterparts. \n\nPros\n- The proposed model, RigL, is memory- and computation- efficient, and thus allows to train a large network in an efficient manner.\n \n- RigL obtains impressive sparsification performance, even yielding sparse networks that outperform their dense counterparts.\n\nCons\n- The idea of starting from a small, sparse network and expanding it is not novel. DEN [Yoon et al. 18] proposed the same bottom-up approach with sparsely initialized networks, while they allowed to increase the number of neurons at each layer and focused more on continual learning. The authors should compare the two methods both conceptually and experimentally. \n\n- The method is more like a set of heuristics rather than a principled approach, which makes it less appealing. This is not really an issue if the paper includes extensive experimental validation and in-depth analysis, but this is not the case. \n \n- The experimental validation is largely lacking, as the authors only perform experiments on ImageNet and do not compare against recent state-of-the-art Bayesian sparsification methods (SBP, VIB, L0-regularization). Without such extensive experimental validation, it is uncertain whether the result will generalize, given the highly empirical nature of the work. \n\nIn sum, although I believe that the paper proposes a very practical method that is easy to implement and is promising, due to lack of experimental validation against a similar approach, state-of-the-art sparsification methods, and results on more datasets, I temporarily provide the rating of weak reject. I may change my opinion if the authors provide those results during the rebuttal period.\n\n[Yoon et al. 18] Lifelong learning with dynamically expandable networks, ICLR 2018", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "This paper proposes a pruning technique called RigL, which performs sparse initialization of the network weights, and allows the network to grow weights during training. The sparse initialization makes the pruning algorithm to be memory- and computation- efficient, unlike existing models that starts with dense network weights. The authors train RigL with three different sparsity distributions that considers the number of input and output nodes, and validate them on various deep convnets for training on ImageNet, on which the model outperforms several existing sparsification methods and even dense counterparts. \n\nPros\n- The proposed model, RigL, is memory- and computation- efficient, and thus allows to train a large network in an efficient manner.\n \n- RigL obtains impressive sparsification performance, even yielding sparse networks that outperform their dense counterparts.\n\nCons\n- The idea of starting from a small, sparse network and expanding it is not novel. DEN [Yoon et al. 18] proposed the same bottom-up approach with sparsely initialized networks, while they allowed to increase the number of neurons at each layer and focused more on continual learning. The authors should compare the two methods both conceptually and experimentally. \n\n- The method is more like a set of heuristics rather than a principled approach, which makes it less appealing. This is not really an issue if the paper includes extensive experimental validation and in-depth analysis, but this is not the case. \n \n- The experimental validation is largely lacking, as the authors only perform experiments on ImageNet and do not compare against recent state-of-the-art Bayesian sparsification methods (SBP, VIB, L0-regularization). Without such extensive experimental validation, it is uncertain whether the result will generalize, given the highly empirical nature of the work. \n\nIn sum, although I believe that the paper proposes a very practical method that is easy to implement and is promising, due to lack of experimental validation against a similar approach, state-of-the-art sparsification methods, and results on more datasets, I temporarily provide the rating of weak reject. I may change my opinion if the authors provide those results during the rebuttal period.\n\n[Yoon et al. 18] Lifelong learning with dynamically expandable networks, ICLR 2018", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1572093176583}, {"id": "H1gDIdo2FS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1169/AnonReviewer2"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method for training sparse network without first training a dense network (e.g. the lottery ticket hypothesis or distillation). The method involves a combination of dynamic pruning of weights coupled with a dynamic \"growing\" of new weights given a novel criterion based on the magnitude of the gradient of the loss. As a result, networks can stay sparse throughout training and testing, leading to a large reduction in computational cost.\n\nThe paper's approach of dynamically changing the topology of networks is an interesting and motivated idea that seems to work rather well. I also appreciate the experiments on MobileNet, a setting where one expects investigations into sparse network architectures to have significant application. Relatedly, I appreciate the importance of the fact that the computational cost of training and evaluating the network is proportional to the sparse model size, which is not normally true for masked dense models. Overall, I found the paper to be very clear and of high quality, and thus I find this to be an interest addition to investigations into the lottery ticket hypothesis.\n\nAs the authors state, the novelty of their method is that they use the gradients with the highest magnitudes to grow connections. This is somewhat intuitive given the role gradients play in gradient descent based optimization, but I was wondering if they had any further intuition as to why this is the right criterion?\n\nIn section 4.3, I was a bit confused by Figure 5. My understanding is that many paths between loss landscape minima follow nonlinear paths -- why is it at all significant that there's a linear barrier? Why are only quadric and cubic Bezier curves used, rather than a more general path finding algorithm?\n\nOverall, this is a nice paper that should be accepted to ICLR.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes a method for training sparse network without first training a dense network (e.g. the lottery ticket hypothesis or distillation). The method involves a combination of dynamic pruning of weights coupled with a dynamic \"growing\" of new weights given a novel criterion based on the magnitude of the gradient of the loss. As a result, networks can stay sparse throughout training and testing, leading to a large reduction in computational cost.\n\nThe paper's approach of dynamically changing the topology of networks is an interesting and motivated idea that seems to work rather well. I also appreciate the experiments on MobileNet, a setting where one expects investigations into sparse network architectures to have significant application. Relatedly, I appreciate the importance of the fact that the computational cost of training and evaluating the network is proportional to the sparse model size, which is not normally true for masked dense models. Overall, I found the paper to be very clear and of high quality, and thus I find this to be an interest addition to investigations into the lottery ticket hypothesis.\n\nAs the authors state, the novelty of their method is that they use the gradients with the highest magnitudes to grow connections. This is somewhat intuitive given the role gradients play in gradient descent based optimization, but I was wondering if they had any further intuition as to why this is the right criterion?\n\nIn section 4.3, I was a bit confused by Figure 5. My understanding is that many paths between loss landscape minima follow nonlinear paths -- why is it at all significant that there's a linear barrier? Why are only quadric and cubic Bezier curves used, rather than a more general path finding algorithm?\n\nOverall, this is a nice paper that should be accepted to ICLR."}, "tcdate": 1571760207204}, {"id": "Skg97KzhKB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1169/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Overview:\n\nThe paper is dedicated to developing a more efficient and powerful dense-to-sparse training method. In order to break the limits of the size of the largest trainable sparse model to that of the largest trainable dense model, the author proposes a dynamic method that updates the network topology via parameter magnitudes and infrequent gradient calculation. In the experiments parts, they conduct extensive studies to show the proposed approach can surpass the previous sota with ResNet-50, MobileNet v1 and v2 on the imagenet 2012. What's more, the author also provides some intuitive explanation about why allowing topology change during the optimization is beneficial. \n\nStrength Bullets:\n\n1. Due to the dynamic network topology, the paper's methods exactly achieve the memory and computation efficient. i) Required memory is only proportional to the size of the sparse model. ii) The amount of computation is proportional to the number of nonzero parameters in the model.\n2. The author performs detailed comparison experiments among different sparsity distribution and different pruning methods. And the results overcome the previous state-of-the-art results.\n3. Fig 5 shows some interesting insight. It suggests that static sparse training may be stuck at some local minima which are isolated from improved solutions. However, the dynamic update has a big chance to avoid this problem.\n\nWeakness Bullets:\n\n1. The author claims that the ticket in the paper does not rely on a \"lucky\" initialization. But it doesn't exclude the possibility that starting from the original initial conditions may give a better performance. Even if the connection is dynamic, we can still record the initial point for each weight. It will be better the author can provide related analysis.\n2. In my opinion, in order to prove dynamic pruning is better than static methods, the author needs to provide a comparison with the previous sota method in it's own setting. i.e. The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks\n\nRecommendation:\n\nI think this paper is a novel work. Although it has some flaws in the experiment design, the motivation and experiment results are conniving enough. So, this is a weak accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "Overview:\n\nThe paper is dedicated to developing a more efficient and powerful dense-to-sparse training method. In order to break the limits of the size of the largest trainable sparse model to that of the largest trainable dense model, the author proposes a dynamic method that updates the network topology via parameter magnitudes and infrequent gradient calculation. In the experiments parts, they conduct extensive studies to show the proposed approach can surpass the previous sota with ResNet-50, MobileNet v1 and v2 on the imagenet 2012. What's more, the author also provides some intuitive explanation about why allowing topology change during the optimization is beneficial. \n\nStrength Bullets:\n\n1. Due to the dynamic network topology, the paper's methods exactly achieve the memory and computation efficient. i) Required memory is only proportional to the size of the sparse model. ii) The amount of computation is proportional to the number of nonzero parameters in the model.\n2. The author performs detailed comparison experiments among different sparsity distribution and different pruning methods. And the results overcome the previous state-of-the-art results.\n3. Fig 5 shows some interesting insight. It suggests that static sparse training may be stuck at some local minima which are isolated from improved solutions. However, the dynamic update has a big chance to avoid this problem.\n\nWeakness Bullets:\n\n1. The author claims that the ticket in the paper does not rely on a \"lucky\" initialization. But it doesn't exclude the possibility that starting from the original initial conditions may give a better performance. Even if the connection is dynamic, we can still record the initial point for each weight. It will be better the author can provide related analysis.\n2. In my opinion, in order to prove dynamic pruning is better than static methods, the author needs to provide a comparison with the previous sota method in it's own setting. i.e. The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks\n\nRecommendation:\n\nI think this paper is a novel work. Although it has some flaws in the experiment design, the motivation and experiment results are conniving enough. So, this is a weak accept."}, "tcdate": 1571723554117}], "openreview_url": "https://openreview.net/forum?id=ryg7vA4tPB", "arxiv_id": "1911.11134", "paper_pdf": "papers/ryg7vA4tPB.pdf", "paper_pdf_sha256": "31921ef6c611ae38f3e02b701fee3b472d0e97e364a84dd857c8c7a43b1907c2", "paper_pdf_bytes": 671480, "paper_pdf_source": "openreview", "code_url": "https://github.com/varun19299/rigl-reproducibility", "code_repository": "varun19299/rigl-reproducibility", "code_commit": "97443beac90e03f899652943594695e5152c2b09", "code_archive": "repos/ryg7vA4tPB.zip", "code_archive_sha256": "50a3c4c4086b27e76dded723123beb58060a856c0fbefca2350cc7452116b1cd", "code_archive_bytes": 3443931, "code_file_count": 53, "code_extensions": {".py": 39, ".js": 14}, "github_disk_usage_kb": 4955, "github_languages": {"Python": 201476, "Makefile": 12540, "Batchfile": 799}, "github_archived": false, "github_pushed_at": "2022-01-06T15:28:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rigging-the-lottery-making-all-tickets-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sV2qR9QIEo", "year": 2026, "status": "rejected", "title": "SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping", "authors": ["Jiajun Li", "Yue Ma", "Xinyu Zhang", "Qingyan Wei", "Songhua Liu", "Linfeng Zhang"], "authorids": ["~Jiajun_Li16", "~Yue_Ma2", "~Xinyu_Zhang53", "~Qingyan_Wei1", "~Songhua_Liu2", "~Linfeng_Zhang2"], "authors_source": "OpenReview API", "abstract": "Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. However, the underlying computational redundancy involved in these steps has yet to be thoroughly investigated. In this paper, we conduct an in-depth analysis of the VAR inference process and identify two primary sources of inefficiency: ***step redundancy*** and ***unconditional branch redundancy***. To address step redundancy, we propose an automatic step-skipping strategy that selectively omits unnecessary generation steps to improve efficiency. For unconditional branch redundancy, we observe that the information gap between the conditional and unconditional branches is minimal. Leveraging this insight, we introduce unconditional branch replacement, a technique that bypasses the unconditional branch to reduce computational cost. Notably, we observe that the effectiveness of acceleration strategies varies significantly across different samples. Motivated by this, we propose **SkipVAR**, a sample-adaptive framework that leverages frequency information to dynamically select the most suitable acceleration strategy for each instance. To evaluate the role of high-frequency information, we further introduce multiple high-variation benchmark datasets that evaluate the performance in terms of fine details. Extensive experiments show that SkipVAR achieves over 0.88 average SSIM with up to 1.81$\\times$ overall acceleration and 2.62$\\times$ lossless speedup on the GenEval benchmark.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "U7vdegveEL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4650/Reviewer_LJ9B"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper addresses the significant inference latency in Visual Autoregressive (VAR) models, which is primarily caused by computationally expensive high-frequency generation steps. The authors identify two key inefficiencies—step redundancy and unconditional branch redundancy—and crucially observe that the impact of these redundancies is highly sample-dependent. They propose SkipVAR, a sample-adaptive framework that uses lightweight frequency features to dynamically select the optimal acceleration strategy (either step-skipping or unconditional branch replacement) for each image. Extensive experiments demonstrate that SkipVAR achieves substantial speedups of up to 1.81x while maintaining high visual fidelity (0.88 SSIM), effectively balancing speed and quality where fixed acceleration methods fail.", "review_text": "This paper addresses the significant inference latency in Visual Autoregressive (VAR) models, which is primarily caused by computationally expensive high-frequency generation steps. The authors identify two key inefficiencies—step redundancy and unconditional branch redundancy—and crucially observe that the impact of these redundancies is highly sample-dependent. They propose SkipVAR, a sample-adaptive framework that uses lightweight frequency features to dynamically select the optimal acceleration strategy (either step-skipping or unconditional branch replacement) for each image. Extensive experiments demonstrate that SkipVAR achieves substantial speedups of up to 1.81x while maintaining high visual fidelity (0.88 SSIM), effectively balancing speed and quality where fixed acceleration methods fail.", "strengths": "* The paper presents novel and significant observations on the important problem of VAR inference latency, identifying sample-dependent redundancy.\n\n* Based on these observations, the paper proposes a natural and reasonable method that adaptively selects acceleration strategies.\n\n* The various design choices, such as the choice of decision models, are backed by a persuasive rationale.", "weaknesses": "1. **Limited Generalizability**\n\t* The paper claims to accelerate \"Visual Autoregressive Modeling,\" yet its entire experimental validation rests exclusively on a single model family: the Infinity-2B/8B family. This is a significant limitation. The core assumptions driving the method, such as the specific patterns of high-frequency redundancy or the convergence of L1 loss between conditional and unconditional branches, may be unique to the Infinity architecture rather than fundamental properties of all VAR models. Without validation on other models that also follow the VAR paradigm (e.g., Tian et al., 2024), the paper fails to sufficiently demonstrate that its findings are generalizable even within its own target model family. **To substantiate the paper's claims, experimental results demonstrating SkipVAR's acceleration performance are necessary on at least one other VAR model family**.\n\t\n2. **Biased Evaluation and Questionable Claims of Superiority over FastVAR**\n\t* The paper's central claim of superiority over key baselines like FastVAR rests heavily on SSIM and LPIPS metrics. However, this comparison appears biased due to a circular evaluation framework. The SkipVAR decision model is explicitly trained to select strategies that preserve a predefined SSIM threshold (e.g., 0.84). Evaluating the model with the same metric it was optimized for does not provide an objective measure of its superiority; it merely confirms that the model met its training objective.\n\t* Furthermore, this reliance on SSIM is problematic. As the authors concede in Appendix L , SSIM is a measure of similarity to the original (non-accelerated) output, not a measure of absolute generation quality. A low SSIM score, as seen in FastVAR's results, indicates that the generated image differs from the original, not necessarily worse in perceptual quality. The possibility remains that while FastVAR produces images that are less faithful to the original generation path, they might maintain or even exceed the overall generation quality on other standard metrics (e.g., FID, CLIPScore, HPSv2, etc.), a comparison which this paper does not rigorously explore. Therefore, **more rigorous experimental results are required to definitively conclude that SkipVAR offers both superior acceleration and better generation quality than FastVAR**.\n\t\n3. **Ambiguous Positioning of the ``SkipVAR-hybrid (w/o DM)`` Variant**\n\t* The paper introduces a ``SkipVAR-hybrid (w/o DM)`` variant, which presents a significant logical contradiction to the paper's core argument for an adaptive strategy. In Table 3, this fixed-strategy (non-adaptive) model is shown achieving a 2.62x speedup and a 0.72 GenEval score, outperforming both the primary adaptive SkipVAR variants (e.g., 1.77x speedup, 0.70 score) and the main competitor, FastVAR (2.53x, 0.68 score).\n\t* This single data point undermines the central argument for an adaptive decision model (DM), as it suggests a fixed strategy is superior in both speed and (at least on this metric) quality. While Table 4 later shows this variant performs poorly on SSIM/LPIPS, the authors fail to provide a clear narrative for its inclusion. Its purpose is ambiguous: is it intended to demonstrate a higher possible speedup (perhaps to show a configuration that outperforms FastVAR on GenEval), or to highlight the flaws of the GenEval metric? As presented, it creates confusion and weakens the justification for the paper's core contribution.\n\t\t\n4. **Questionable Practical Utility of the Adaptive Framework**\n\t* The core value proposition of the adaptive framework is its ability to handle sample-specific needs (i.e., frequency-sensitive vs. -robust) without significant quality degradation. However, the paper's own results cast doubt on its practical utility. According to Table 2b, when the decision model identifies a sample as \"frequency-sensitive,\" the resulting speedup is only 1.28x, a marginal gain.\n\t* Conversely, \"frequency-robust\" samples are accelerated up to 1.99x. This implies that the average speedup figures (e.g., 1.81x in Table 4) are not derived from a balanced acceleration, but are heavily skewed by aggressively skipping \"easy\" samples. If the primary function of the complex adaptive mechanism for \"hard\" samples is to simply not accelerate them significantly, its practical value proposition over a simpler, more conservative fixed strategy (e.g., SkipVAR-hybrid (w/o DM)) is questionable.", "questions": "* In several metrics in Table 2 and 3 (e.g., Paint in 2(a), Align in 2(b), and Color Attri. and Overall in Table 3), the application of the proposed acceleration strategies results in a slight increase in performance compared to the original, non-accelerated model. How should these increases be interpreted? Are they simply statistical noise, or does this suggest that SkipVAR's strategies can incidentally mitigate certain generation artifacts or semantic errors present in the original model, thereby leading to a genuine improvement in quality? If so, what is the mechanism for this improvement?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the significant inference latency in Visual Autoregressive (VAR) models, which is primarily caused by computationally expensive high-frequency generation steps. The authors identify two key inefficiencies—step redundancy and unconditional branch redundancy—and crucially observe that the impact of these redundancies is highly sample-dependent. They propose SkipVAR, a sample-adaptive framework that uses lightweight frequency features to dynamically select the optimal acceleration strategy (either step-skipping or unconditional branch replacement) for each image. Extensive experiments demonstrate that SkipVAR achieves substantial speedups of up to 1.81x while maintaining high visual fidelity (0.88 SSIM), effectively balancing speed and quality where fixed acceleration methods fail.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "* The paper presents novel and significant observations on the important problem of VAR inference latency, identifying sample-dependent redundancy.\n\n* Based on these observations, the paper proposes a natural and reasonable method that adaptively selects acceleration strategies.\n\n* The various design choices, such as the choice of decision models, are backed by a persuasive rationale.", "weaknesses": "1. **Limited Generalizability**\n\t* The paper claims to accelerate \"Visual Autoregressive Modeling,\" yet its entire experimental validation rests exclusively on a single model family: the Infinity-2B/8B family. This is a significant limitation. The core assumptions driving the method, such as the specific patterns of high-frequency redundancy or the convergence of L1 loss between conditional and unconditional branches, may be unique to the Infinity architecture rather than fundamental properties of all VAR models. Without validation on other models that also follow the VAR paradigm (e.g., Tian et al., 2024), the paper fails to sufficiently demonstrate that its findings are generalizable even within its own target model family. **To substantiate the paper's claims, experimental results demonstrating SkipVAR's acceleration performance are necessary on at least one other VAR model family**.\n\t\n2. **Biased Evaluation and Questionable Claims of Superiority over FastVAR**\n\t* The paper's central claim of superiority over key baselines like FastVAR rests heavily on SSIM and LPIPS metrics. However, this comparison appears biased due to a circular evaluation framework. The SkipVAR decision model is explicitly trained to select strategies that preserve a predefined SSIM threshold (e.g., 0.84). Evaluating the model with the same metric it was optimized for does not provide an objective measure of its superiority; it merely confirms that the model met its training objective.\n\t* Furthermore, this reliance on SSIM is problematic. As the authors concede in Appendix L , SSIM is a measure of similarity to the original (non-accelerated) output, not a measure of absolute generation quality. A low SSIM score, as seen in FastVAR's results, indicates that the generated image differs from the original, not necessarily worse in perceptual quality. The possibility remains that while FastVAR produces images that are less faithful to the original generation path, they might maintain or even exceed the overall generation quality on other standard metrics (e.g., FID, CLIPScore, HPSv2, etc.), a comparison which this paper does not rigorously explore. Therefore, **more rigorous experimental results are required to definitively conclude that SkipVAR offers both superior acceleration and better generation quality than FastVAR**.\n\t\n3. **Ambiguous Positioning of the ``SkipVAR-hybrid (w/o DM)`` Variant**\n\t* The paper introduces a ``SkipVAR-hybrid (w/o DM)`` variant, which presents a significant logical contradiction to the paper's core argument for an adaptive strategy. In Table 3, this fixed-strategy (non-adaptive) model is shown achieving a 2.62x speedup and a 0.72 GenEval score, outperforming both the primary adaptive SkipVAR variants (e.g., 1.77x speedup, 0.70 score) and the main competitor, FastVAR (2.53x, 0.68 score).\n\t* This single data point undermines the central argument for an adaptive decision model (DM), as it suggests a fixed strategy is superior in both speed and (at least on this metric) quality. While Table 4 later shows this variant performs poorly on SSIM/LPIPS, the authors fail to provide a clear narrative for its inclusion. Its purpose is ambiguous: is it intended to demonstrate a higher possible speedup (perhaps to show a configuration that outperforms FastVAR on GenEval), or to highlight the flaws of the GenEval metric? As presented, it creates confusion and weakens the justification for the paper's core contribution.\n\t\t\n4. **Questionable Practical Utility of the Adaptive Framework**\n\t* The core value proposition of the adaptive framework is its ability to handle sample-specific needs (i.e., frequency-sensitive vs. -robust) without significant quality degradation. However, the paper's own results cast doubt on its practical utility. According to Table 2b, when the decision model identifies a sample as \"frequency-sensitive,\" the resulting speedup is only 1.28x, a marginal gain.\n\t* Conversely, \"frequency-robust\" samples are accelerated up to 1.99x. This implies that the average speedup figures (e.g., 1.81x in Table 4) are not derived from a balanced acceleration, but are heavily skewed by aggressively skipping \"easy\" samples. If the primary function of the complex adaptive mechanism for \"hard\" samples is to simply not accelerate them significantly, its practical value proposition over a simpler, more conservative fixed strategy (e.g., SkipVAR-hybrid (w/o DM)) is questionable.", "questions": "* In several metrics in Table 2 and 3 (e.g., Paint in 2(a), Align in 2(b), and Color Attri. and Overall in Table 3), the application of the proposed acceleration strategies results in a slight increase in performance compared to the original, non-accelerated model. How should these increases be interpreted? Are they simply statistical noise, or does this suggest that SkipVAR's strategies can incidentally mitigate certain generation artifacts or semantic errors present in the original model, thereby leading to a genuine improvement in quality? If so, what is the mechanism for this improvement?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762367611428}, {"id": "6j2CeB1Wka", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4650/Reviewer_webr"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This manuscript presents SkipVAR, a training-free decoding framework that accelerates visual autoregressive (VAR) image generation models by trimming late-stage computation with curtailed quality loss. The manuscript points out two recurring inefficiencies: 1) final refinement steps often make tiny changes, and 2) the unconditional branch in classifier-free guidance becomes redundant late in sampling. Based on this observation, SkipVAR pauses mid-generation to compute high frequency difference (Sobel) and high-frequency ratio (FFT), then either 1) early stops the remaining steps or 2) replaces the unconditional branch with the conditional one. On Infinity models, SkipVAR reaches ~1.8× speedups at near-baseline quality, and a hybrid variant attains ~2.6× on GenEval.", "review_text": "This manuscript presents SkipVAR, a training-free decoding framework that accelerates visual autoregressive (VAR) image generation models by trimming late-stage computation with curtailed quality loss. The manuscript points out two recurring inefficiencies: 1) final refinement steps often make tiny changes, and 2) the unconditional branch in classifier-free guidance becomes redundant late in sampling. Based on this observation, SkipVAR pauses mid-generation to compute high frequency difference (Sobel) and high-frequency ratio (FFT), then either 1) early stops the remaining steps or 2) replaces the unconditional branch with the conditional one. On Infinity models, SkipVAR reaches ~1.8× speedups at near-baseline quality, and a hybrid variant attains ~2.6× on GenEval.", "strengths": "- Discovers and addresses(with minimal training) the issue of late-stage high frequency redundancy and redundant CFG passes in VAR generation.\n- Two intuitive signals (Sobel and FFT) enable per sample decisions and preserve high frequency detail better than token pruning/merging at similar speedups.", "weaknesses": "- Reported speedups exclude VAE and post-processing, so end-to-end latency improvements in production are likely smaller; end-to-end measurements are needed.\n- Heavy reliance on classifier-free guidance, since a major gain comes from dropping the unconditional branch; applicability to non-CFG or single-branch decoders is unclear.\n- In Table 3, the strongest ~2.6× result does not use the decision model, which obfuscates the efficacy of the decision model.", "questions": "Please refer to the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript presents SkipVAR, a training-free decoding framework that accelerates visual autoregressive (VAR) image generation models by trimming late-stage computation with curtailed quality loss. The manuscript points out two recurring inefficiencies: 1) final refinement steps often make tiny changes, and 2) the unconditional branch in classifier-free guidance becomes redundant late in sampling. Based on this observation, SkipVAR pauses mid-generation to compute high frequency difference (Sobel) and high-frequency ratio (FFT), then either 1) early stops the remaining steps or 2) replaces the unconditional branch with the conditional one. On Infinity models, SkipVAR reaches ~1.8× speedups at near-baseline quality, and a hybrid variant attains ~2.6× on GenEval.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Discovers and addresses(with minimal training) the issue of late-stage high frequency redundancy and redundant CFG passes in VAR generation.\n- Two intuitive signals (Sobel and FFT) enable per sample decisions and preserve high frequency detail better than token pruning/merging at similar speedups.", "weaknesses": "- Reported speedups exclude VAE and post-processing, so end-to-end latency improvements in production are likely smaller; end-to-end measurements are needed.\n- Heavy reliance on classifier-free guidance, since a major gain comes from dropping the unconditional branch; applicability to non-CFG or single-branch decoders is unclear.\n- In Table 3, the strongest ~2.6× result does not use the decision model, which obfuscates the efficacy of the decision model.", "questions": "Please refer to the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761983845939}, {"id": "xf9ZChfpLZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4650/Reviewer_HFYL"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "To my understanding, the authors start from the observation that VAR (Visual Autoregressive) models spend a lot of redundant computation in later sampling steps, and they introduce two light-weight inference-time acceleration tricks: (i) step-skipping, which skips late-stage scaling or denoising steps, and (ii) unconditional branch replacement, which replaces the unconditional CFG branch with the conditional one’s output. To decide when these shortcuts are safe, they train a tiny decision model (mainly a logistic regression) using two handcrafted high-frequency features: high-frequency difference with Sobel operator (HF_Diff) and high-frequency ratio (HF_Ratio) with Fourier transform.  As one of main empirical contributions, their approach applied to Infinity-2B and 8B VAR achieves up to 1.81x speed-up while keeping plausible SSIM, and up to 2.62x acceleration on GenEval benchmark.", "review_text": "To my understanding, the authors start from the observation that VAR (Visual Autoregressive) models spend a lot of redundant computation in later sampling steps, and they introduce two light-weight inference-time acceleration tricks: (i) step-skipping, which skips late-stage scaling or denoising steps, and (ii) unconditional branch replacement, which replaces the unconditional CFG branch with the conditional one’s output. To decide when these shortcuts are safe, they train a tiny decision model (mainly a logistic regression) using two handcrafted high-frequency features: high-frequency difference with Sobel operator (HF_Diff) and high-frequency ratio (HF_Ratio) with Fourier transform.  As one of main empirical contributions, their approach applied to Infinity-2B and 8B VAR achieves up to 1.81x speed-up while keeping plausible SSIM, and up to 2.62x acceleration on GenEval benchmark.", "strengths": "1. Sample-adaptive decisions: Unlike prior acceleration methods that use a fixed global ratio or policy (e.g., FastVAR), this work chooses between step-skipping and branch replacement per-sample and per-scale, which makes the approach much more practical. The results on frequency-sensitive vs. frequency-robust subsets clearly demonstrate the benefit of this adaptive decision process.\n\n2. Simple and interpretable features: The combination of HF_Diff (local edge stability) and HF_Ratio (global high-frequency ratio) proves more reliable than using either feature alone, as shown in their tables. It seems that both SSIM/LPIPS and SSIM-HF/LPIPS-HF improve together.\n\n3. Clarity of method overview: Figure 5 presents the overall pipeline--decision step N, downsampled decoding, feature extraction, and policy application--in a clear and easy-to-follow way.", "weaknesses": "I'm not an expert in this area, but based on my understanding, I have the following concerns and questions.\n\n1. Sensitivity to decision step and threshold\n\nAccording to paper (with appendix), the default setup uses N = 10 with SSIM thresholds {0.88, 0.86, 0.84}, but the paper doesn’t really explore how performance changes with different step counts (which may vary by model or resolution) or different thresholds. I notice that there’s a brief comparison between SSIM-based and LPIPS-based criteria in the appendix (which says SSIM as more stable), yet a more systematic sweep over N and threshold values would make the analysis much more comprehensive and complete.\n\n2. About branch replecement\n\nReplacing the unconditional branch with the conditional one in CFG effectively collapses into simply $y = y_c$, meaning the efffective CFG scale becomes 1. The authors argue that this is reasonable since the conditional and unconditional branches converge in later steps, but they don't provide concrete measurements of how they close these two quantites are. It also seems unclear whether just reducing the CFG scale to 1 in the later stages would yield the same effect. I think this part is very important since it is the very motivation behind this work.\n\n3. About wall-clock time\n\nCould the authors report the overall wall-clock speed-up, including the decision process and VAE decoding, for DrawBench, HPSv2, and GenEval? It would help clarify how much of the reported acceleration remains when all inference-time components are accounted for.", "questions": "Please refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To my understanding, the authors start from the observation that VAR (Visual Autoregressive) models spend a lot of redundant computation in later sampling steps, and they introduce two light-weight inference-time acceleration tricks: (i) step-skipping, which skips late-stage scaling or denoising steps, and (ii) unconditional branch replacement, which replaces the unconditional CFG branch with the conditional one’s output. To decide when these shortcuts are safe, they train a tiny decision model (mainly a logistic regression) using two handcrafted high-frequency features: high-frequency difference with Sobel operator (HF_Diff) and high-frequency ratio (HF_Ratio) with Fourier transform.  As one of main empirical contributions, their approach applied to Infinity-2B and 8B VAR achieves up to 1.81x speed-up while keeping plausible SSIM, and up to 2.62x acceleration on GenEval benchmark.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Sample-adaptive decisions: Unlike prior acceleration methods that use a fixed global ratio or policy (e.g., FastVAR), this work chooses between step-skipping and branch replacement per-sample and per-scale, which makes the approach much more practical. The results on frequency-sensitive vs. frequency-robust subsets clearly demonstrate the benefit of this adaptive decision process.\n\n2. Simple and interpretable features: The combination of HF_Diff (local edge stability) and HF_Ratio (global high-frequency ratio) proves more reliable than using either feature alone, as shown in their tables. It seems that both SSIM/LPIPS and SSIM-HF/LPIPS-HF improve together.\n\n3. Clarity of method overview: Figure 5 presents the overall pipeline--decision step N, downsampled decoding, feature extraction, and policy application--in a clear and easy-to-follow way.", "weaknesses": "I'm not an expert in this area, but based on my understanding, I have the following concerns and questions.\n\n1. Sensitivity to decision step and threshold\n\nAccording to paper (with appendix), the default setup uses N = 10 with SSIM thresholds {0.88, 0.86, 0.84}, but the paper doesn’t really explore how performance changes with different step counts (which may vary by model or resolution) or different thresholds. I notice that there’s a brief comparison between SSIM-based and LPIPS-based criteria in the appendix (which says SSIM as more stable), yet a more systematic sweep over N and threshold values would make the analysis much more comprehensive and complete.\n\n2. About branch replecement\n\nReplacing the unconditional branch with the conditional one in CFG effectively collapses into simply $y = y_c$, meaning the efffective CFG scale becomes 1. The authors argue that this is reasonable since the conditional and unconditional branches converge in later steps, but they don't provide concrete measurements of how they close these two quantites are. It also seems unclear whether just reducing the CFG scale to 1 in the later stages would yield the same effect. I think this part is very important since it is the very motivation behind this work.\n\n3. About wall-clock time\n\nCould the authors report the overall wall-clock speed-up, including the decision process and VAE decoding, for DrawBench, HPSv2, and GenEval? It would help clarify how much of the reported acceleration remains when all inference-time components are accounted for.", "questions": "Please refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761961578154}, {"id": "WrfBtgqpEj", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4650/Reviewer_euwc"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper addresses the high inference latency in Visual Autoregressive (VAR) models by investigating and mitigating computational redundancy. The authors identify that later autoregressive steps provide minimal visual improvement while costing the majority of latency (Steps 11–13 account for 69% of total inference time). They further observe that the unconditional branch in classifier-free guidance models offers diminishing returns in later stages. To solve this, the paper proposes SkipVAR, a novel, sample-adaptive framework that uses a lightweight decision model to dynamically select the best acceleration strategy: skip 0%, 50% or 100% based on its frequency characteristics. SkipVAR combines an automatic step-skipping strategy to omit unnecessary generation steps and unconditional branch replacement to bypass the computationally costly unconditional branch.", "review_text": "The paper addresses the high inference latency in Visual Autoregressive (VAR) models by investigating and mitigating computational redundancy. The authors identify that later autoregressive steps provide minimal visual improvement while costing the majority of latency (Steps 11–13 account for 69% of total inference time). They further observe that the unconditional branch in classifier-free guidance models offers diminishing returns in later stages. To solve this, the paper proposes SkipVAR, a novel, sample-adaptive framework that uses a lightweight decision model to dynamically select the best acceleration strategy: skip 0%, 50% or 100% based on its frequency characteristics. SkipVAR combines an automatic step-skipping strategy to omit unnecessary generation steps and unconditional branch replacement to bypass the computationally costly unconditional branch.", "strengths": "1. The authors observe that high frequency components does not impact the image quality for certain subsets of images and devise a VAR acceleration decision model that determines the acceleration strategy based on the frequency information.\n2. The authors attain a speedup of 1.81 with an average SSIM of 0.88.", "weaknesses": "1. The comparison against existing approaches for VAR acceleration has only been performed on the basis of objective metrics like SSIM. However, a comparison on subjective metrics like CLIP score is necessary for a more holistic comparison.\n2. This approach is only beneficial if the high frequency components do not impact image quality. For example, as the authors point out, this approach will not be beneficial for generating realistic portraits.\n3. While the authors compare against token-based approaches, it will be interesting to see how this method compares to layer skipping methods like [1, 2]\n4. For Table 2, the blue and green row colors are barely visible.\n\n[1] Andrey Gromov et al., The Unreasonable Ineffectiveness of the Deeper Layers, ICLR 2025.\n[2] Anhao Zhao et al., SkipGPT: Each Token is One of a Kind, ICML 2025.", "questions": "1. Please show performance and have discussion in comparison to Speculative decoding for VAR , for example: LANTERN [1].\n2. How does SkipVAR perform on GenEval and DrawBench datasets using metrics like ImageReward and CLIP Score?\n3. The authors only test performance wth three granularity levels: 0%, 50% and 100%. Did the authors check if better performance can be obtained using finer granularity or are there any challenges associated with it?\n\n[1] LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding, ICLR 2025.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the high inference latency in Visual Autoregressive (VAR) models by investigating and mitigating computational redundancy. The authors identify that later autoregressive steps provide minimal visual improvement while costing the majority of latency (Steps 11–13 account for 69% of total inference time). They further observe that the unconditional branch in classifier-free guidance models offers diminishing returns in later stages. To solve this, the paper proposes SkipVAR, a novel, sample-adaptive framework that uses a lightweight decision model to dynamically select the best acceleration strategy: skip 0%, 50% or 100% based on its frequency characteristics. SkipVAR combines an automatic step-skipping strategy to omit unnecessary generation steps and unconditional branch replacement to bypass the computationally costly unconditional branch.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The authors observe that high frequency components does not impact the image quality for certain subsets of images and devise a VAR acceleration decision model that determines the acceleration strategy based on the frequency information.\n2. The authors attain a speedup of 1.81 with an average SSIM of 0.88.", "weaknesses": "1. The comparison against existing approaches for VAR acceleration has only been performed on the basis of objective metrics like SSIM. However, a comparison on subjective metrics like CLIP score is necessary for a more holistic comparison.\n2. This approach is only beneficial if the high frequency components do not impact image quality. For example, as the authors point out, this approach will not be beneficial for generating realistic portraits.\n3. While the authors compare against token-based approaches, it will be interesting to see how this method compares to layer skipping methods like [1, 2]\n4. For Table 2, the blue and green row colors are barely visible.\n\n[1] Andrey Gromov et al., The Unreasonable Ineffectiveness of the Deeper Layers, ICLR 2025.\n[2] Anhao Zhao et al., SkipGPT: Each Token is One of a Kind, ICML 2025.", "questions": "1. Please show performance and have discussion in comparison to Speculative decoding for VAR , for example: LANTERN [1].\n2. How does SkipVAR perform on GenEval and DrawBench datasets using metrics like ImageReward and CLIP Score?\n3. The authors only test performance wth three granularity levels: 0%, 50% and 100%. Did the authors check if better performance can be obtained using finer granularity or are there any challenges associated with it?\n\n[1] LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding, ICLR 2025.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761958231848}], "openreview_url": "https://openreview.net/forum?id=sV2qR9QIEo", "arxiv_id": "2506.08908", "paper_pdf": "papers/sV2qR9QIEo.pdf", "paper_pdf_sha256": "34bd9cd94f8898e3e75bc7eec10b2335ead80c908b6401c8dd9553ec30a33341", "paper_pdf_bytes": 50916108, "paper_pdf_source": "openreview", "code_url": "https://github.com/fakerone-li/SkipVAR", "code_repository": "fakerone-li/SkipVAR", "code_commit": "607be0efea96a91458ce1e6f6623a8b2ad0b01e6", "code_archive": "repos/sV2qR9QIEo.zip", "code_archive_sha256": "e468556324b0c1175a6f1011b3913a68bbdf206fbe0cf0fbdacb384877df8363", "code_archive_bytes": 4009344, "code_file_count": 53, "code_extensions": {".py": 51, ".sh": 2}, "github_disk_usage_kb": 3587, "github_languages": {"Python": 488231, "Shell": 9109}, "github_archived": false, "github_pushed_at": "2025-06-15T09:06:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/skipvar-accelerating-visual-autoregressive"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LUcdXA8hAa", "year": 2024, "status": "rejected", "title": "Identifiability Matters: Revealing the Hidden Recoverable Condition in Unbiased Learning to Rank", "authors": ["Mouxiang Chen", "Chenghao Liu", "Zemin Liu", "Zhuo Li", "Jianling Sun"], "authorids": ["~Mouxiang_Chen1", "~Chenghao_Liu1", "~Zemin_Liu1", "lizhuo@zju.edu.cn", "~Jianling_Sun2"], "authors_source": "OpenReview API", "abstract": "The application of Unbiased Learning to Rank (ULTR) is widespread in modern systems for training unbiased ranking models from biased click logs. The key is to explicitly model a generation process for user behavior and fit click data based on examination hypothesis. Previous research found empirically that the true latent relevance can be recovered in most cases as long as the clicks are perfectly fitted. However, we demonstrate that this is not always achievable, resulting in a significant reduction in ranking performance. In this work, we aim to answer if or when the true relevance can be recovered from click data, which is a foundation issue for ULTR field. We first define a ranking model as identifiable if it can recover the true relevance up to a scaling transformation, which is enough for pairwise ranking objective. Then we explore an equivalent condition for identifiability that can be novely expressed as a graph connectivity test problem: if and only if a graph (namely identifiability graph, or IG) constructed on the underlying structure of the dataset is connected, we can guarantee that the relevance can be correctly recovered. When the IG is not connected, there may be bad cases leading to poor ranking performance. To address this issue, we propose two methods, namely node intervention and node merging, to modify the dataset and restore connectivity of the IG. Empirical results obtained on a simulation dataset and two LTR benchmark datasets confirm the validity of our proposed theorems and show the effectiveness of our methods in mitigating data bias when the relevance model is unidentifiable.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Zy5oTullrw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5474/Reviewer_2RKu"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors explore if or when the true relevance can be recovered from click data. Overall, it is a solid paper. My concern lies in whether and how can this approach apply to the existing unbiased learning-to-rank framework that developed from the examination hypothesis. Or, how the proposed framework incorporates current ranking models. Also, it would be better to compare this work against more recently proposed existing unbiased learning-to-rank algorithms. Overall, I give a weak rejection. If the authors would clarify the above concerns, I will be happy to raise my score.", "review_text": "In this paper, the authors explore if or when the true relevance can be recovered from click data. Overall, it is a solid paper. My concern lies in whether and how can this approach apply to the existing unbiased learning-to-rank framework that developed from the examination hypothesis. Or, how the proposed framework incorporates current ranking models. Also, it would be better to compare this work against more recently proposed existing unbiased learning-to-rank algorithms. Overall, I give a weak rejection. If the authors would clarify the above concerns, I will be happy to raise my score.", "strengths": "1. It is an interesting and important topic to investigate when the true relevance can be recovered from click data.\n2. I like the theoretical analysis in this paper (i.e., Sections 3 and 4).\n3. They have conducted experiments on Yahoo! and Istella-S datasets, verifying the performance of the proposed model.", "weaknesses": "1. It lacks recently proposed methods as the baselines.\n2. It would be interesting to discuss the difference between the proposed method and the existing approaches based on the examination hypothesis.\n3. For an unbiased learning-to-rank algorithm, there is always a ranking algorithm base. It is not clear how the proposed model incorporates the existing ranking models.", "questions": "It is essential to evaluate if or when the true relevance can be recovered from click data. I like this idea very much. However, in the context of unbiased learning-to-rank, there should be a ranking function (often referred to as biased), and then the core goal of unbiased learning-to-rank is to build a debiasing method that can be incorporated into the biased ranking models. After reading this paper multiple times, I consider that it is not clear how this approach can be applied to existing ranking models. Also, in the experiment part, the authors only compare no debias and a simple examine hypothesis method. I highly recommend the authors compare and discuss the connections to existing unbiased learning-to-rank algorithms such as “Unbiased Learning to Rank with Unbiased Propensity Estimation” and “An Unbiased Pairwise Learning-to-Rank Algorithm”. Therefore, I would like to give a weak rejection.  If the authors would clarify the above concerns, I will be happy to raise my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors explore if or when the true relevance can be recovered from click data. Overall, it is a solid paper. My concern lies in whether and how can this approach apply to the existing unbiased learning-to-rank framework that developed from the examination hypothesis. Or, how the proposed framework incorporates current ranking models. Also, it would be better to compare this work against more recently proposed existing unbiased learning-to-rank algorithms. Overall, I give a weak rejection. If the authors would clarify the above concerns, I will be happy to raise my score.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. It is an interesting and important topic to investigate when the true relevance can be recovered from click data.\n2. I like the theoretical analysis in this paper (i.e., Sections 3 and 4).\n3. They have conducted experiments on Yahoo! and Istella-S datasets, verifying the performance of the proposed model.", "weaknesses": "1. It lacks recently proposed methods as the baselines.\n2. It would be interesting to discuss the difference between the proposed method and the existing approaches based on the examination hypothesis.\n3. For an unbiased learning-to-rank algorithm, there is always a ranking algorithm base. It is not clear how the proposed model incorporates the existing ranking models.", "questions": "It is essential to evaluate if or when the true relevance can be recovered from click data. I like this idea very much. However, in the context of unbiased learning-to-rank, there should be a ranking function (often referred to as biased), and then the core goal of unbiased learning-to-rank is to build a debiasing method that can be incorporated into the biased ranking models. After reading this paper multiple times, I consider that it is not clear how this approach can be applied to existing ranking models. Also, in the experiment part, the authors only compare no debias and a simple examine hypothesis method. I highly recommend the authors compare and discuss the connections to existing unbiased learning-to-rank algorithms such as “Unbiased Learning to Rank with Unbiased Propensity Estimation” and “An Unbiased Pairwise Learning-to-Rank Algorithm”. Therefore, I would like to give a weak rejection.  If the authors would clarify the above concerns, I will be happy to raise my score.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698915570791}, {"id": "AL6AjuUYlX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5474/Reviewer_ggp4"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper investigates the problem of position bias in the task of Unbiased Learning to Rank. It first introduces the widespread concern of biased user logs and the consequent appriximation error in most of the commonly used ranking models. With a clear problem setup and definition of identifiability, the author states the conditions under which the true relevances can be extracted from click data. Particularly, the paper converts the identifiability problem into a graph connectivity problem. Based on the connectivity problem,  the authors further come up two new approaches to deal with “unidentifiable” datasets and optimize the ranking models using structures of the graph. Experiments are conducted to prove the validity of the graph conversion, the performance of new approaches, and the application to the real-world datasets.", "review_text": "The paper investigates the problem of position bias in the task of Unbiased Learning to Rank. It first introduces the widespread concern of biased user logs and the consequent appriximation error in most of the commonly used ranking models. With a clear problem setup and definition of identifiability, the author states the conditions under which the true relevances can be extracted from click data. Particularly, the paper converts the identifiability problem into a graph connectivity problem. Based on the connectivity problem,  the authors further come up two new approaches to deal with “unidentifiable” datasets and optimize the ranking models using structures of the graph. Experiments are conducted to prove the validity of the graph conversion, the performance of new approaches, and the application to the real-world datasets.", "strengths": "S1: As one of its major contributions, this paper tansfers the identifiability of a ULTR task to the connectivity of a graph constructed based on the dataset. This equivalence is useful in that tasks related to a graph usually have more efficient computations, nicer properties, and more intuition-based understandings. The work also allows for more explainability in the field of ULTR and thus simplifies difficult questions.\n\nS2: This paper proposes two novel methods, node intervention and node merging, to bridge the “unidentifiability gap” by utilizing the graph properties. These two methods are theoretically supported and empirically verified.", "weaknesses": "W1: Since the theory of this paper relies heavily on the examination hypothesis, the graph-equivalence idea is not generalizeable to other general hypotheses on the dataset.\n\nW2: Since the methods are still propensity-based, there are a plethora of such ranking models. It would be fair game if the paper uses such methods as baselines to strengthen the validity of the proposed methods.\n\nW3: When choosing the bias factors, we can choose either fewer factors, which makes the graph more likely to be connected, or more factors, which accounts for more bias but poses a more disconnected graph. It would be great if there is any discussion on the tradeoff and the corresponding performance of the two proposed methods. In addition, assume that each feature x and bias factor t are independently and uniformly chosen to construct the dataset D is nearly impossible in practice.", "questions": "Q1: In the real world, the dataset is mostly sparse and thus there might be a large number of connected components in IG. How much does the performance of the two methods deteriote with the increasing sparsity? Is there a systematic way to deal with that issue?\n\nQ2: In the node merging method, the costs between nodes are computed based on the their deterministic features X_t. However, how is it guaranteed that the features reflect their true similarity? For example, we may use document rank as a bias factor when only considering the position bias. But it turns out that the user may notice the documents in the order of: the first several documents (since they’re most noticeable), the last several documents on this page (since users may scroll down), and then documents in the middle. In the more complex settings of several factors, it’s even less obvious which nodes are similar to each other. Is it possible to make the features not deterministic but rather learned throughout multi-task learning?\n\nQ3: In the application of the method, the dataset is mostly online and continuously taking in new data points. How does the proposed method handle the updates efficiently? For example, if two nodes (bias factors) with similar features are already merged but the new datapoints from the user creates an edge between them, is there a way to efficiently deal with this?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates the problem of position bias in the task of Unbiased Learning to Rank. It first introduces the widespread concern of biased user logs and the consequent appriximation error in most of the commonly used ranking models. With a clear problem setup and definition of identifiability, the author states the conditions under which the true relevances can be extracted from click data. Particularly, the paper converts the identifiability problem into a graph connectivity problem. Based on the connectivity problem,  the authors further come up two new approaches to deal with “unidentifiable” datasets and optimize the ranking models using structures of the graph. Experiments are conducted to prove the validity of the graph conversion, the performance of new approaches, and the application to the real-world datasets.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "S1: As one of its major contributions, this paper tansfers the identifiability of a ULTR task to the connectivity of a graph constructed based on the dataset. This equivalence is useful in that tasks related to a graph usually have more efficient computations, nicer properties, and more intuition-based understandings. The work also allows for more explainability in the field of ULTR and thus simplifies difficult questions.\n\nS2: This paper proposes two novel methods, node intervention and node merging, to bridge the “unidentifiability gap” by utilizing the graph properties. These two methods are theoretically supported and empirically verified.", "weaknesses": "W1: Since the theory of this paper relies heavily on the examination hypothesis, the graph-equivalence idea is not generalizeable to other general hypotheses on the dataset.\n\nW2: Since the methods are still propensity-based, there are a plethora of such ranking models. It would be fair game if the paper uses such methods as baselines to strengthen the validity of the proposed methods.\n\nW3: When choosing the bias factors, we can choose either fewer factors, which makes the graph more likely to be connected, or more factors, which accounts for more bias but poses a more disconnected graph. It would be great if there is any discussion on the tradeoff and the corresponding performance of the two proposed methods. In addition, assume that each feature x and bias factor t are independently and uniformly chosen to construct the dataset D is nearly impossible in practice.", "questions": "Q1: In the real world, the dataset is mostly sparse and thus there might be a large number of connected components in IG. How much does the performance of the two methods deteriote with the increasing sparsity? Is there a systematic way to deal with that issue?\n\nQ2: In the node merging method, the costs between nodes are computed based on the their deterministic features X_t. However, how is it guaranteed that the features reflect their true similarity? For example, we may use document rank as a bias factor when only considering the position bias. But it turns out that the user may notice the documents in the order of: the first several documents (since they’re most noticeable), the last several documents on this page (since users may scroll down), and then documents in the middle. In the more complex settings of several factors, it’s even less obvious which nodes are similar to each other. Is it possible to make the features not deterministic but rather learned throughout multi-task learning?\n\nQ3: In the application of the method, the dataset is mostly online and continuously taking in new data points. How does the proposed method handle the updates efficiently? For example, if two nodes (bias factors) with similar features are already merged but the new datapoints from the user creates an edge between them, is there a way to efficiently deal with this?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698841999536}, {"id": "oUsbhwPuHt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5474/Reviewer_tHmn"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper delves into the identifiability issue of the ranking model within the context of unbiased learning to rank (ULTR). Previous studies have established the model's unbiasedness, assuming perfect fits for both clicks and the observation model, inadvertently overlooking the identifiability challenge at its core. Motivated by this, the article investigates the conditions necessary to recover identifiability, primarily in the context of scale transformation. The authors formalize the identifiability challenge as a graph connectivity test problem. Based on it, they further propose two methods, namely node intervention and node merging, to tackle this problem for empirical applications.", "review_text": "This paper delves into the identifiability issue of the ranking model within the context of unbiased learning to rank (ULTR). Previous studies have established the model's unbiasedness, assuming perfect fits for both clicks and the observation model, inadvertently overlooking the identifiability challenge at its core. Motivated by this, the article investigates the conditions necessary to recover identifiability, primarily in the context of scale transformation. The authors formalize the identifiability challenge as a graph connectivity test problem. Based on it, they further propose two methods, namely node intervention and node merging, to tackle this problem for empirical applications.", "strengths": "1. The topic is interesting and important both theoretically and empirically;\n2. The methods of graph connectivity are novel to me;\n3. This paper is well-written.", "weaknesses": "1. Experiments are weak and not very convincing, since it has very little baseline (see Q1 below for more details);\n\n2. Theorem 3 is straightforward and does not seem useful; the conditions required by Theorems 2 and 4 are stringent (see Q2 below for more details).", "questions": "I have two main concerns: \n\n**Q1.** In the experiment, there are only two weak baselines. In addition, there are lack of details about the two baselines. However, this article mentions a lot of related works but does not include them as part of the baselines for comparison empirically. This is inconsistent with the requirements in this area of ULTR. Could you add some cutting-edge ULTR methods as baselines? \n\n**Q2.** In terms of the theoretical results. \n\n>(1)\tTheorem 2 is a very simple case (… $x$ and $t$ are selected independently and uniformly …). Thus, it is not sufficient to write it as a Theorem; it might be more appropriate to write it as a Proposition.\n\n> (2)\tTheorem 3 is simple and straightforward (just by the central limit theorem) and does not seem useful. Could you clarify the purpose and use of Theorem 3? Also, it is not sufficient to write it as a Theorem; it might be more appropriate to write it as a Lemma.\n\n> (3)\tThe condition required in Theorem 4, \"A disconnected IG consists of two connected components G1 and G2\" is strong and difficult to fulfill in practice. \n\n**In fact, the soundness of Theorems 3 and 4 is very critical to this paper.** Here are the two main reasons: (a) The graph is always disconnected in practice, and will suffer from identifiability problems; (b) To recover the identifiability, we always need node intervention and node merging to recover the connected graph for empirical applications. Thus, the rationality of these two proposed algorithms (node intervention and node merging) becomes critical. Regrettably, Theorems 3 and 4 are slightly weak, which seriously undermines the contribution of this paper.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper delves into the identifiability issue of the ranking model within the context of unbiased learning to rank (ULTR). Previous studies have established the model's unbiasedness, assuming perfect fits for both clicks and the observation model, inadvertently overlooking the identifiability challenge at its core. Motivated by this, the article investigates the conditions necessary to recover identifiability, primarily in the context of scale transformation. The authors formalize the identifiability challenge as a graph connectivity test problem. Based on it, they further propose two methods, namely node intervention and node merging, to tackle this problem for empirical applications.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The topic is interesting and important both theoretically and empirically;\n2. The methods of graph connectivity are novel to me;\n3. This paper is well-written.", "weaknesses": "1. Experiments are weak and not very convincing, since it has very little baseline (see Q1 below for more details);\n\n2. Theorem 3 is straightforward and does not seem useful; the conditions required by Theorems 2 and 4 are stringent (see Q2 below for more details).", "questions": "I have two main concerns: \n\n**Q1.** In the experiment, there are only two weak baselines. In addition, there are lack of details about the two baselines. However, this article mentions a lot of related works but does not include them as part of the baselines for comparison empirically. This is inconsistent with the requirements in this area of ULTR. Could you add some cutting-edge ULTR methods as baselines? \n\n**Q2.** In terms of the theoretical results. \n\n>(1)\tTheorem 2 is a very simple case (… $x$ and $t$ are selected independently and uniformly …). Thus, it is not sufficient to write it as a Theorem; it might be more appropriate to write it as a Proposition.\n\n> (2)\tTheorem 3 is simple and straightforward (just by the central limit theorem) and does not seem useful. Could you clarify the purpose and use of Theorem 3? Also, it is not sufficient to write it as a Theorem; it might be more appropriate to write it as a Lemma.\n\n> (3)\tThe condition required in Theorem 4, \"A disconnected IG consists of two connected components G1 and G2\" is strong and difficult to fulfill in practice. \n\n**In fact, the soundness of Theorems 3 and 4 is very critical to this paper.** Here are the two main reasons: (a) The graph is always disconnected in practice, and will suffer from identifiability problems; (b) To recover the identifiability, we always need node intervention and node merging to recover the connected graph for empirical applications. Thus, the rationality of these two proposed algorithms (node intervention and node merging) becomes critical. Regrettably, Theorems 3 and 4 are slightly weak, which seriously undermines the contribution of this paper.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698734136106}, {"id": "Yse4eSp0VV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5474/Reviewer_CYAM"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The submission studies the identifiability problem of unbiased learning to rank (ULTR) - given a dataset of implicit feedback, whether the true relevance can be identified or not. By treating each bias configuration as a node, and the shared input feature vector for the edges, the relevance model is identifiable if and only if the graph is connected. Then two methods are proposed to try to address the issues (making the graph connected). Experiments are conducted on two synthetic datasets and an offline dataset, where it shows that using the two methods to change data can improve the performance of a naive baseline.", "review_text": "The submission studies the identifiability problem of unbiased learning to rank (ULTR) - given a dataset of implicit feedback, whether the true relevance can be identified or not. By treating each bias configuration as a node, and the shared input feature vector for the edges, the relevance model is identifiable if and only if the graph is connected. Then two methods are proposed to try to address the issues (making the graph connected). Experiments are conducted on two synthetic datasets and an offline dataset, where it shows that using the two methods to change data can improve the performance of a naive baseline.", "strengths": "It is interesting to study the identifiability problem of unbiased learning to rank from the graph connectivity perspective, though the reviewer is not convinced that this is \"the first work\" to study the identifiability issue given existing work on coupling / confounding, etc. The major novelty seems to be the graph view.\n\nThe proposed two methods are easy to understand. The authors do acknowledge the caveats of the proposed methods (one being not very practical and one may propagate errors).\n\nThe paper is overall clearly written.", "weaknesses": "Overall the paper is not satisfactory in term of theories or experiments.\n\nThough the theorems look interesting at first glance, the reviewer feels they are generally not very solid or practical after a closer look at them. A major issue is the reviewer feels the submission has self-contradictions in several places. \n\nSo there are two scenarios in practice: 1) there are a lot of data, and the bias factor space is small. This is the a common case in practice and people are just fine without worrying about identifiability. The analysis and methods in this paper mostly do not apply since the graph is small and likely connected. 2) There are a lot of bias factors and the graph is more likely to be not connected. The paper mainly argues about this scenario. So far so good and it is ok to focus on 2).\n\nHowever, by closing look at each theorem, all of them are questionable and some look contradictory to the focus/motivation:\n\nTheorem1: while the trend of using more bias factors is the trend is debatable (especially given existing work showing the coupling / confounding effect), the trend to enrich x is clear. Many real-world applications have personalized feature vectors - then the graph is not likely to be connected even with a small number of bias factors. The paper does not concern this perspective, also, people are fine working with ULTR with such datasets. This questions the value of the proposed framework - one should also note that the condition is only a sufficient condition. \n\nTheorem 2: The assumptions are too strong to make meaningful value from this theorem. The reviewer understands this is to show a simplified “estimate”, but still the value is quite limited for this highly practical field.\n\nTheorem 3: There’s self-contradiction with the motivation of the paper. As discussed, the paper mainly concerns large bias factor space. However, this theorem assumes that each (input feature , bias factor) pair need to sample N observations from a Bernoulli distribution, and the theorem is based on “N is sufficiently large”. How can this be meaningful under the scenario the paper is concerned with?\n\nTheorem 4: The error bound is only shown to merge two subgraphs. Again, the paper is concerned with large bias factor space and the number of subgraphs could be be high - what is the error bound for the entire merge process? Will the errors propagate to meaningless values? Showing error bound only for merging two subgraphs looks quite limited.\n\nThere are also several places in the paper that also look contradictory, e.g. when it argues about the benefit of node intervention, “It should be noted that K is typically smaller the the number of positions (assuming that positions are the sole bias factor)” - the reviewer is confused about such claims. If the paper is concerned with such scenario, then there’s probably no need to worry about the identifiability issue.\n\nOn the experiments part, the evaluation is weak. The major issue is, the proposed methods are only shown to improve the very basic baseline. The authors argue that the methods are agnostic to the actual algorithm and “aptly represents the current research” - the reviewer strongly disagrees - to show the proposed method is really meaningful,  it needs to show that they can help more sensible baselines. For example, will they help state-of-the-art ULTR methods? If not, why would people care? This is important since the proposed two methods have clear caveats (the node intervention method is not very practical, the node merging method is likely to introduce errors).", "questions": "See questions above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission studies the identifiability problem of unbiased learning to rank (ULTR) - given a dataset of implicit feedback, whether the true relevance can be identified or not. By treating each bias configuration as a node, and the shared input feature vector for the edges, the relevance model is identifiable if and only if the graph is connected. Then two methods are proposed to try to address the issues (making the graph connected). Experiments are conducted on two synthetic datasets and an offline dataset, where it shows that using the two methods to change data can improve the performance of a naive baseline.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "It is interesting to study the identifiability problem of unbiased learning to rank from the graph connectivity perspective, though the reviewer is not convinced that this is \"the first work\" to study the identifiability issue given existing work on coupling / confounding, etc. The major novelty seems to be the graph view.\n\nThe proposed two methods are easy to understand. The authors do acknowledge the caveats of the proposed methods (one being not very practical and one may propagate errors).\n\nThe paper is overall clearly written.", "weaknesses": "Overall the paper is not satisfactory in term of theories or experiments.\n\nThough the theorems look interesting at first glance, the reviewer feels they are generally not very solid or practical after a closer look at them. A major issue is the reviewer feels the submission has self-contradictions in several places. \n\nSo there are two scenarios in practice: 1) there are a lot of data, and the bias factor space is small. This is the a common case in practice and people are just fine without worrying about identifiability. The analysis and methods in this paper mostly do not apply since the graph is small and likely connected. 2) There are a lot of bias factors and the graph is more likely to be not connected. The paper mainly argues about this scenario. So far so good and it is ok to focus on 2).\n\nHowever, by closing look at each theorem, all of them are questionable and some look contradictory to the focus/motivation:\n\nTheorem1: while the trend of using more bias factors is the trend is debatable (especially given existing work showing the coupling / confounding effect), the trend to enrich x is clear. Many real-world applications have personalized feature vectors - then the graph is not likely to be connected even with a small number of bias factors. The paper does not concern this perspective, also, people are fine working with ULTR with such datasets. This questions the value of the proposed framework - one should also note that the condition is only a sufficient condition. \n\nTheorem 2: The assumptions are too strong to make meaningful value from this theorem. The reviewer understands this is to show a simplified “estimate”, but still the value is quite limited for this highly practical field.\n\nTheorem 3: There’s self-contradiction with the motivation of the paper. As discussed, the paper mainly concerns large bias factor space. However, this theorem assumes that each (input feature , bias factor) pair need to sample N observations from a Bernoulli distribution, and the theorem is based on “N is sufficiently large”. How can this be meaningful under the scenario the paper is concerned with?\n\nTheorem 4: The error bound is only shown to merge two subgraphs. Again, the paper is concerned with large bias factor space and the number of subgraphs could be be high - what is the error bound for the entire merge process? Will the errors propagate to meaningless values? Showing error bound only for merging two subgraphs looks quite limited.\n\nThere are also several places in the paper that also look contradictory, e.g. when it argues about the benefit of node intervention, “It should be noted that K is typically smaller the the number of positions (assuming that positions are the sole bias factor)” - the reviewer is confused about such claims. If the paper is concerned with such scenario, then there’s probably no need to worry about the identifiability issue.\n\nOn the experiments part, the evaluation is weak. The major issue is, the proposed methods are only shown to improve the very basic baseline. The authors argue that the methods are agnostic to the actual algorithm and “aptly represents the current research” - the reviewer strongly disagrees - to show the proposed method is really meaningful,  it needs to show that they can help more sensible baselines. For example, will they help state-of-the-art ULTR methods? If not, why would people care? This is important since the proposed two methods have clear caveats (the node intervention method is not very practical, the node merging method is likely to introduce errors).", "questions": "See questions above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698674778905}], "openreview_url": "https://openreview.net/forum?id=LUcdXA8hAa", "arxiv_id": "2309.15560", "paper_pdf": "papers/LUcdXA8hAa.pdf", "paper_pdf_sha256": "664040d4c83342ecb7a25338f5439d32457061dee97440eaccd1b3eb0d33283d", "paper_pdf_bytes": 714069, "paper_pdf_source": "openreview", "code_url": "https://github.com/Keytoyze/ULTR-identifiability", "code_repository": "Keytoyze/ULTR-identifiability", "code_commit": "08e623547d4ac09698a0c0d7c4ed166b39d72ddb", "code_archive": "repos/LUcdXA8hAa.zip", "code_archive_sha256": "e3f22b88cb5ef3b965ee8846bbce6ae5d38e6dfc3e0ce6cfd8a5a5bdb1d2232d", "code_archive_bytes": 440385, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 204, "github_languages": {"Gherkin": 588900, "Python": 50958}, "github_archived": false, "github_pushed_at": "2024-05-18T09:52:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/identifiability-matters-revealing-the-hidden"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4NLyCJQR3ZR", "year": 2023, "status": "rejected", "title": "Optimal Neural Network Approximation of Wasserstein Gradient Direction via Convex Optimization", "authors": ["Yifei Wang", "Peng Chen", "Mert Pilanci", "Wuchen Li"], "authorids": ["~Yifei_Wang2", "~Peng_Chen1", "~Mert_Pilanci3", "~Wuchen_Li1"], "authors_source": "OpenReview API", "abstract": "The computation of Wasserstein gradient direction is essential for posterior sampling problems and scientific computing. The approximation of the Wasserstein gradient with finite samples requires solving a variational problem. We study the variational problem in the family of two-layer networks with squared-ReLU activations, towards which we derive a semi-definite programming (SDP) relaxation. This SDP can be viewed as an approximation of the Wasserstein gradient in a broader function family including two-layer networks. By solving the convex SDP, we obtain the optimal approximation of the Wasserstein gradient direction in this class of functions. Numerical experiments including PDE-constrained Bayesian inference and parameter estimation in COVID-19 modeling demonstrate the effectiveness of the proposed method.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "gUePIiDGWpZ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1309/Reviewer_4MPo"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper uses an SDP relaxation for the problem of computing the Wasserstein gradient. They use numerical algorithms that make their proposed method suitable for practical scenarios involving Bayesian inference.", "review_text": "Overall the paper uses a straightforward technique to a problem that has limited practical applications (learning two-layers neural nets).", "strengths": "Strength\n======\n\n- The idea of using a convex SDP relaxation of the dual of the variational primal problem is interesting to understand the behaviour of a particular family of two-layer neural nets.\n\n\nWeaknesses\n=========\n\n- Using convex SDP relaxations is not a new idea per se: see, e.g., \"Semidefinite Relaxation of Quadratic Optimization Problems\" by Tom Luo *et al*.\n\nFeedback\n=======\n\n- This sentence: \"However, due to the nonlinear and nonconvex structure of neural networks, optimization algorithms\nincluding **stochastic gradient descent may not find the global optima** of the training problem\" is presented as if, in general, not finding the global optima is an issue in SGD. In training neural nets you precisely do not want to train till the global optima as that would likely mean **overfitting**.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper uses an SDP relaxation for the problem of computing the Wasserstein gradient. They use numerical algorithms that make their proposed method suitable for practical scenarios involving Bayesian inference.", "strength_and_weaknesses": "Strength\n======\n\n- The idea of using a convex SDP relaxation of the dual of the variational primal problem is interesting to understand the behaviour of a particular family of two-layer neural nets.\n\n\nWeaknesses\n=========\n\n- Using convex SDP relaxations is not a new idea per se: see, e.g., \"Semidefinite Relaxation of Quadratic Optimization Problems\" by Tom Luo *et al*.\n\nFeedback\n=======\n\n- This sentence: \"However, due to the nonlinear and nonconvex structure of neural networks, optimization algorithms\nincluding **stochastic gradient descent may not find the global optima** of the training problem\" is presented as if, in general, not finding the global optima is an issue in SGD. In training neural nets you precisely do not want to train till the global optima as that would likely mean **overfitting**.", "clarity,_quality,_novelty_and_reproducibility": "The paper reads well and code is provided for reproducibility. The paper's novelty is limited as it applies well-known ideas to a somewhat specific (as in, not generalizable) problem.", "summary_of_the_review": "Overall the paper uses a straightforward technique to a problem that has limited practical applications (learning two-layers neural nets).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666334713072}, {"id": "zUIsKd-n9C9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1309/Reviewer_efky"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The manuscript proposes a convex relaxation for the approximation of sample based Wasserstein gradients with a KL divergence as an objective. The motivation for this is the application to Bayesian inference problems. \nTo derive the relaxation, the drift in the ODE describing the evolution of the particles can be characterized as the minimizer of an energy functional, which can be used as a training objective for a shallow neural network. A convex relaxation as a semidefinite program of the dual of this optimization problem is presented. Further, the dual of this  relaxation is given and used in order to show that the relaxation  provides a lower bound to the primal problem. \nThe benefits of the proposed method are demonstrated on three different problems, on a synthetic toy problem and on two Bayesian inference problems, which also include real world data.  ", "review_text": "The paper studies an important and timely problem and provides a nice extension of recent works on SDP relaxations to neural network approximation of Wasserstein gradients. Overall, I enjoyed reading the manuscript as the overall idea is clearly laid out. For more detailed comments on the different aspects, I would like to refer to my comments above.\nCurrently, I feel that the manuscript requires improvements in different aspects, which I summarize below:\n\nThings that I find necessary to see addressed:\n\n* *Relation to prior works on convex relaxations:* I think it is important to highlight the difference of the proposed relaxation to existing works on convex relaxations of neural network training problems. \n* *Clarity:* Particularly Section 3 is not easy to digest in its current form. A lot of things can be improved by proof reading the section and fixing formatting (e.g. in Remark 4 and Lemma 4, Proposition 4 and the paragraph below Theorem 1). Concrete comments for Proposition 4: There is currently no statement made in the proposition, so it causes some confusion in its current form; the formatting should definitely be improved; ideally, the symbols should be introduced before the equation; $[p]$ notation is not introduced as a notation; there is a comma at the end of (15) and it continues with upper case; there is a fullstop missing after $[N]$; $A_j(\\Lambda)$ is introduced but doesn’t appear in (15). \n* *Optimality:* In its current form the title suggests evidence on the optimality of the proposed method for the approximation of Wasserstein gradients. Further, you state in the conclusion that „the gradient … is at least as good“. Currently I do not see results supporting this in the manuscript. Do you see this as a consequence of Theorem 1 or of your experiments? Theorem 1 implies that the regularized dual gives a lower bound on the primal and your experiments show that the resulting WGD converges faster. This does however not imply that the approximation of the Wasserstein gradient is more exact as far as I understand it. I think the claim of optimality needs to be supported better or needs to be adjusted accordingly.\n\nThings I recommend to address:\n* *Experiments:* It would be good to have experiments comparing the primal and dual problem in the following sense: What is the deviation of the two update directions of WGD-NN and WGD-cvxNN away from the true Wasserstein gradient. In particular, this could be an empirical proof that the approximation of the Wasserstein gradient is indeed better with the proposed method. \n* *Consequences of Theorem 1:* The manuscript would benefit from an elaboration of the implications of Theorem 1. In particular, what does it say about the relation between the proposed relaxation and the primal problem?\n* *More general setup:* The proposed method is only presented for the KL divergence as an objective. Although this is very well motivated by the application in Bayesian inference tasks, the contributions of the manuscript would be stronger, if if was formulated and analyzed in a more general setting.\n\nSpecific questions:\n* Why do you call the proposed approximation optimal, i.e., does it satisfy any optimality criteria?\n* What are the implications of Theorem 1?\n* Proposition 1:\n\t* What happens with the boundary terms in the partial integration?\n\t* Why do you formulate Proposition 1 for a function space? I think it is just an alternative expression of the energy due to partial integration. Being nit-picky, later you consider the problem on a function class (so no linear structure), which might confuse people if they read „function space“ is something linear. I would rather simply say that the energy in (4) takes this form. \n\nFurther thoughts (not so important): \n* The experiments could be stronger if not only compared against other methods relying on Wasserstein gradient approximations, but rather on  Bayesian inference techniques.\n* *Variational formulation:* I am not entirely sure, whether I appreciate the wording here. The reason for this is that I am unsure whether (4) is the variational formulation of (2) in the sense that I am unsure, whether its Euler-Lagrange equations characterize the Wasserstein gradient. Think for example of a linear PDE $-\\Delta u = f$ (lets say with zero boundary values), which is both the unique minimizer of $\\frac12\\int \\lvert\\nabla u\\rvert^2dx-\\int fu dx$ and of $\\int \\lvert \\Delta u+f\\rvert dx$, where the first one is called the variational formulation and the second one a residual minimization. \n* *Formatting (pretty nit-picky):* Line breaks in section titles are not good; a single subsection in a section is not too nice, maybe a \\paragraph does the job; ", "strengths": "**Strengths:**\n* The Introduction is very well written and the overall idea of the paper presented clearly, which leads to a smooth overall read. \n* I appreciate the idea of the convex SDP relaxation of the NN training problem encountered in the variational formulation of the Wasserstein gradient approximation with shallow networks.\n* The empirical results show that the approach can be made effective even for larger problems and include problems with real world data.\n\n**Weaknesses:**\n* Where the introduction is written very nicely, the later sections are not always a very smooth read. This is mainly due to a) linguistic errors, which can easily be fixed b) rather heavy notation and suboptimal formatting, which can also be addressed.\n* The discussion of related works in the introduction does a good job giving an overview over approximations of WGD. However, the relation to prior works on SDP relaxations of neural network training is not sufficiently well described. Some of these works are mentioned in the introduction, but their precise relation and in particular the difference of the manuscript to prior works here is not described. This would be important to add in order to make the contributions of the manuscript clear.\n*  The experiments would be stronger if not only compared against other methods relying on Wasserstein gradient approximations, but rather on  Bayesian inference techniques.\n* The title of the paper implies that using the proposed relaxation yields a better approximation of the Wasserstein gradient compared to solving the primal problem. This is however neither supported sufficiently well in an experiment nor a theoretical result as far as I can see.  \n* The theoretical aspects of the proposed method are not very well elaborated. In particular, the implications of Theorem 1 for the method are not laid out very transparently. \n* The proposed method is only presented for the KL divergence as an objective. Although this is very well motivated by the application in Bayesian inference tasks, the contributions of the manuscript would be stronger, if if was formulated and analyzed in a more general setting.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The manuscript proposes a convex relaxation for the approximation of sample based Wasserstein gradients with a KL divergence as an objective. The motivation for this is the application to Bayesian inference problems. \nTo derive the relaxation, the drift in the ODE describing the evolution of the particles can be characterized as the minimizer of an energy functional, which can be used as a training objective for a shallow neural network. A convex relaxation as a semidefinite program of the dual of this optimization problem is presented. Further, the dual of this  relaxation is given and used in order to show that the relaxation  provides a lower bound to the primal problem. \nThe benefits of the proposed method are demonstrated on three different problems, on a synthetic toy problem and on two Bayesian inference problems, which also include real world data.  ", "strength_and_weaknesses": "**Strengths:**\n* The Introduction is very well written and the overall idea of the paper presented clearly, which leads to a smooth overall read. \n* I appreciate the idea of the convex SDP relaxation of the NN training problem encountered in the variational formulation of the Wasserstein gradient approximation with shallow networks.\n* The empirical results show that the approach can be made effective even for larger problems and include problems with real world data.\n\n**Weaknesses:**\n* Where the introduction is written very nicely, the later sections are not always a very smooth read. This is mainly due to a) linguistic errors, which can easily be fixed b) rather heavy notation and suboptimal formatting, which can also be addressed.\n* The discussion of related works in the introduction does a good job giving an overview over approximations of WGD. However, the relation to prior works on SDP relaxations of neural network training is not sufficiently well described. Some of these works are mentioned in the introduction, but their precise relation and in particular the difference of the manuscript to prior works here is not described. This would be important to add in order to make the contributions of the manuscript clear.\n*  The experiments would be stronger if not only compared against other methods relying on Wasserstein gradient approximations, but rather on  Bayesian inference techniques.\n* The title of the paper implies that using the proposed relaxation yields a better approximation of the Wasserstein gradient compared to solving the primal problem. This is however neither supported sufficiently well in an experiment nor a theoretical result as far as I can see.  \n* The theoretical aspects of the proposed method are not very well elaborated. In particular, the implications of Theorem 1 for the method are not laid out very transparently. \n* The proposed method is only presented for the KL divergence as an objective. Although this is very well motivated by the application in Bayesian inference tasks, the contributions of the manuscript would be stronger, if if was formulated and analyzed in a more general setting.", "clarity,_quality,_novelty_and_reproducibility": "**Quality:**\nOverall, the main idea of the paper as well as the execution is of sufficient quality. \nThe proposed relaxation is a reasonable and timely approach to the important problem of approximating Wasserstein gradients. \nIn particular, I appreciate the fact that the proposed method is demonstrated on more than a synthetic data. \n\n**Clarity:**\nThe overall idea of relaxing the dual of a training problem is presented very clear and therefore the general structure of the manuscript is easy to follow. On a smaller scale some paragraphs and formulations are less clear. This is mainly due to a) linguistic errors, which can be easily fixed b) rather heavy notation and suboptimal formatting, which can also be addressed. I compiled a list of specific things that caught my eye in the review.\n\n**Originality:** \nGiven the recent works on convex relaxations of neural network training problems in the context of supervised learning problems, the idea of the paper to apply it in the context of neural network approximations of the Wasserstein gradients seems not exceptional original. That being said, I like the idea and believe that demonstrating that SDP relaxations can also be used in this context is a valuable contribution.", "summary_of_the_review": "The paper studies an important and timely problem and provides a nice extension of recent works on SDP relaxations to neural network approximation of Wasserstein gradients. Overall, I enjoyed reading the manuscript as the overall idea is clearly laid out. For more detailed comments on the different aspects, I would like to refer to my comments above.\nCurrently, I feel that the manuscript requires improvements in different aspects, which I summarize below:\n\nThings that I find necessary to see addressed:\n\n* *Relation to prior works on convex relaxations:* I think it is important to highlight the difference of the proposed relaxation to existing works on convex relaxations of neural network training problems. \n* *Clarity:* Particularly Section 3 is not easy to digest in its current form. A lot of things can be improved by proof reading the section and fixing formatting (e.g. in Remark 4 and Lemma 4, Proposition 4 and the paragraph below Theorem 1). Concrete comments for Proposition 4: There is currently no statement made in the proposition, so it causes some confusion in its current form; the formatting should definitely be improved; ideally, the symbols should be introduced before the equation; $[p]$ notation is not introduced as a notation; there is a comma at the end of (15) and it continues with upper case; there is a fullstop missing after $[N]$; $A_j(\\Lambda)$ is introduced but doesn’t appear in (15). \n* *Optimality:* In its current form the title suggests evidence on the optimality of the proposed method for the approximation of Wasserstein gradients. Further, you state in the conclusion that „the gradient … is at least as good“. Currently I do not see results supporting this in the manuscript. Do you see this as a consequence of Theorem 1 or of your experiments? Theorem 1 implies that the regularized dual gives a lower bound on the primal and your experiments show that the resulting WGD converges faster. This does however not imply that the approximation of the Wasserstein gradient is more exact as far as I understand it. I think the claim of optimality needs to be supported better or needs to be adjusted accordingly.\n\nThings I recommend to address:\n* *Experiments:* It would be good to have experiments comparing the primal and dual problem in the following sense: What is the deviation of the two update directions of WGD-NN and WGD-cvxNN away from the true Wasserstein gradient. In particular, this could be an empirical proof that the approximation of the Wasserstein gradient is indeed better with the proposed method. \n* *Consequences of Theorem 1:* The manuscript would benefit from an elaboration of the implications of Theorem 1. In particular, what does it say about the relation between the proposed relaxation and the primal problem?\n* *More general setup:* The proposed method is only presented for the KL divergence as an objective. Although this is very well motivated by the application in Bayesian inference tasks, the contributions of the manuscript would be stronger, if if was formulated and analyzed in a more general setting.\n\nSpecific questions:\n* Why do you call the proposed approximation optimal, i.e., does it satisfy any optimality criteria?\n* What are the implications of Theorem 1?\n* Proposition 1:\n\t* What happens with the boundary terms in the partial integration?\n\t* Why do you formulate Proposition 1 for a function space? I think it is just an alternative expression of the energy due to partial integration. Being nit-picky, later you consider the problem on a function class (so no linear structure), which might confuse people if they read „function space“ is something linear. I would rather simply say that the energy in (4) takes this form. \n\nFurther thoughts (not so important): \n* The experiments could be stronger if not only compared against other methods relying on Wasserstein gradient approximations, but rather on  Bayesian inference techniques.\n* *Variational formulation:* I am not entirely sure, whether I appreciate the wording here. The reason for this is that I am unsure whether (4) is the variational formulation of (2) in the sense that I am unsure, whether its Euler-Lagrange equations characterize the Wasserstein gradient. Think for example of a linear PDE $-\\Delta u = f$ (lets say with zero boundary values), which is both the unique minimizer of $\\frac12\\int \\lvert\\nabla u\\rvert^2dx-\\int fu dx$ and of $\\int \\lvert \\Delta u+f\\rvert dx$, where the first one is called the variational formulation and the second one a residual minimization. \n* *Formatting (pretty nit-picky):* Line breaks in section titles are not good; a single subsection in a section is not too nice, maybe a \\paragraph does the job; ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666264238668}, {"id": "4TD9OPa52_P", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1309/Reviewer_mWPZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper tackles the issue of computing Wasserstein gradient direction for 2-layers NNs thanks to a SDP approach with main properties that there is no need to train the underlying NN. The gradient direction is computed with the dual formulation of a least-square + a polynomial regularization term.", "review_text": "I believe the paper is based on an interesting idea but lacks clarity in its exposition, and fails to convince on the practical aspect.", "strengths": "**Strenghts**\n- This is probably the first work to cast this problem as a SDP problem\n- The optimal is global (despite being a non convex problem initially)\n- The paper is well-written\n\n**Weaknesses**\n- The exposition is lacking an I believe some part of the paper could be moved to the appendix to free some space to better put in perspective the related work on that matter\n- Theoretical statements lacks precision.\n- There is no clear advantage both theoretically and practically to not train directly the NN on the Wasserstein gradient.\n- Finally, and probably the most important issue, is that the computational cost is prohibitive. It is not clear why at least there is no discussion on how to solve the SDP problem when d is large.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper tackles the issue of computing Wasserstein gradient direction for 2-layers NNs thanks to a SDP approach with main properties that there is no need to train the underlying NN. The gradient direction is computed with the dual formulation of a least-square + a polynomial regularization term.", "strength_and_weaknesses": "**Strenghts**\n- This is probably the first work to cast this problem as a SDP problem\n- The optimal is global (despite being a non convex problem initially)\n- The paper is well-written\n\n**Weaknesses**\n- The exposition is lacking an I believe some part of the paper could be moved to the appendix to free some space to better put in perspective the related work on that matter\n- Theoretical statements lacks precision.\n- There is no clear advantage both theoretically and practically to not train directly the NN on the Wasserstein gradient.\n- Finally, and probably the most important issue, is that the computational cost is prohibitive. It is not clear why at least there is no discussion on how to solve the SDP problem when d is large.", "clarity,_quality,_novelty_and_reproducibility": "See above", "summary_of_the_review": "I believe the paper is based on an interesting idea but lacks clarity in its exposition, and fails to convince on the practical aspect.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "3: reject, not good enough"}, "tcdate": 1665753190294}], "openreview_url": "https://openreview.net/forum?id=4NLyCJQR3ZR", "arxiv_id": "2205.13098", "paper_pdf": "papers/4NLyCJQR3ZR.pdf", "paper_pdf_sha256": "adf05e3e2554827a520fe53080e15dc3b34065c9dcaa7421726f75e414182fe6", "paper_pdf_bytes": 593672, "paper_pdf_source": "openreview", "code_url": "https://github.com/ai-submit/OptimalWasserstein", "code_repository": "ai-submit/OptimalWasserstein", "code_commit": "d78e87664d2b6e90b7e0925883957809f5417f1c", "code_archive": "repos/4NLyCJQR3ZR.zip", "code_archive_sha256": "622af8bd567fd3e45a6b5a469b355ae24a159fc7076125e1ecd5dfcd3a254b44", "code_archive_bytes": 499459, "code_file_count": 146, "code_extensions": {".py": 132, ".cpp": 7, ".h": 4, ".ipynb": 2, ".hpp": 1}, "github_disk_usage_kb": 410, "github_languages": {"Python": 1529282, "C++": 39813, "Jupyter Notebook": 2886}, "github_archived": false, "github_pushed_at": "2022-05-25T21:14:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimal-neural-network-approximation-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SuKTLF9stD", "year": 2022, "status": "rejected", "title": "Data-Efficient Augmentation for Training Neural Networks", "authors": ["Tian Yu Liu", "Baharan Mirzasoleiman"], "authorids": ["~Tian_Yu_Liu2", "~Baharan_Mirzasoleiman4"], "authors_source": "OpenReview API", "abstract": "Data augmentation is essential to achieve state-of-the-art performance in many deep learning applications. However, modern data augmentation techniques become computationally prohibitive for large datasets. To address this, we propose a rigorous technique to select subsets of data points that when augmented, closely capture the training dynamics of full data augmentation. We first show that data augmentation, modeled as additive perturbations, speeds up learning by enlarging the smaller singular values of the network Jacobian. Then, we propose a framework to iteratively extract small subsets of training data that when augmented, closely capture the alignment of the fully augmented Jacobian with label/residual vector. We prove that stochastic gradient descent applied to augmented subsets found by our approach have similar training dynamics to that of fully augmented data. Our experiments demonstrate that our method outperforms state-of-the-art max-loss strategy by 7.7% on CIFAR10 while achieving 6.3x speedup, and by 4.7% on SVHN while achieving 2.2x speedup, using 10% and 30% subsets, respectively.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ZiO7_PLhy9", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1080/Reviewer_xyXQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Deep learning uses augmentations to improve generalization performance. Using all augmentations for a dataset may slow down training. A subset selection technique is proposed (e.g., \"coreset\") using insights from the Neural Tangent Kernel (NMT) framework such that an alignment between the NMT Jacobian and the residuals is preserved. This alignment score is used to select the coreset using submodular optimization, which allows the model to be trained on a subset of augmented data (e.g., 0.1% to 30%) while preserving most augmentation benefits. The speedup of the method is reported to be up to 6.3x.", "review_text": "Strengths:\n* The paper presents a convincing reason why the approach can work in theory.\n* The approach seems novel and connects various prior works.\n\nWeaknesses:\n* I thought the algorithm description in Section 4 was a bit rushed. For example, equation 11 was hard to understand and is core to the algorithm.\n* I also though the Section 5 was rushed. The whole evaluation is ~1 page, which leaves the reader wanting for 1) how speedup is calculated 2) what is the objective of the experiments 3) what takeaways should be had. I would recommend moving other text to the appendix (e.g., anything reviewing prior work can be shortened and pushed to appendix).\n* Figure 3b seems like a random walk. Figure 3c x-axis and text labels seem misaligned. Overall, I think this figure can be improved to give a clearer and more convincing story.\n* The theory is appreciated but maybe it's worth contextualizing what values of the constants L and epsilon_0 are common in practice. Even a simple augmentation like translate can cause a large pixel deviation. For example, with 100 pixels horizontally, each varying by 1 intensity via a gradient (e.g., 0 to 99), shifting by 1 pixel will cause a change in the image that is 100 large.\n* The good results seem very CIFAR10 specific. For example, Figure 4b shows SVNH has minimal improvement. More experimental evaluation is always good. For example, MNIST would have been an easy result to add and is in-fact mentioned in the paper, yet I didn't see it in evaluation.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Deep learning uses augmentations to improve generalization performance. Using all augmentations for a dataset may slow down training. A subset selection technique is proposed (e.g., \"coreset\") using insights from the Neural Tangent Kernel (NMT) framework such that an alignment between the NMT Jacobian and the residuals is preserved. This alignment score is used to select the coreset using submodular optimization, which allows the model to be trained on a subset of augmented data (e.g., 0.1% to 30%) while preserving most augmentation benefits. The speedup of the method is reported to be up to 6.3x.", "main_review": "Strengths:\n* The paper presents a convincing reason why the approach can work in theory.\n* The approach seems novel and connects various prior works.\n\nWeaknesses:\n* I thought the algorithm description in Section 4 was a bit rushed. For example, equation 11 was hard to understand and is core to the algorithm.\n* I also though the Section 5 was rushed. The whole evaluation is ~1 page, which leaves the reader wanting for 1) how speedup is calculated 2) what is the objective of the experiments 3) what takeaways should be had. I would recommend moving other text to the appendix (e.g., anything reviewing prior work can be shortened and pushed to appendix).\n* Figure 3b seems like a random walk. Figure 3c x-axis and text labels seem misaligned. Overall, I think this figure can be improved to give a clearer and more convincing story.\n* The theory is appreciated but maybe it's worth contextualizing what values of the constants L and epsilon_0 are common in practice. Even a simple augmentation like translate can cause a large pixel deviation. For example, with 100 pixels horizontally, each varying by 1 intensity via a gradient (e.g., 0 to 99), shifting by 1 pixel will cause a change in the image that is 100 large.\n* The good results seem very CIFAR10 specific. For example, Figure 4b shows SVNH has minimal improvement. More experimental evaluation is always good. For example, MNIST would have been an easy result to add and is in-fact mentioned in the paper, yet I didn't see it in evaluation.", "summary_of_the_review": "The paper is ok. I think there is enough theoretical justification for why the method could work, though I think empirical evidence is necessary to still show that. The theory is a bit disconnected and seems more like a synthesis of many different works; I wonder if it can be presented more succinctly? In any case, I found the writing good enough to follow. My biggest concern is the generalization of the evaluation, since it seems quite short at the moment and very CIFAR10 specific.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635913396184}, {"id": "oqaoyy-Me37", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1080/Reviewer_fQLH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper first provides a theoretical analysis, under some assumptions, of the effect of data augmentation on the singular values of the network Jacobian and then proposes a method to improve the sample complexity of data augmentation, that is to select a subset of the data on which to perform the transformations to speed up training. The proposed method aims to find subsets of data whose augmentation yields similar alignment of the Jacobian with the residual vector as the fully augmented data. ", "review_text": "The main contribution of this paper is, in my opinion, a sound theoretical analysis of the impact of input perturbations on the singular values of the Jacobian, or the eigenvalues of the Neural Tangent Kernel (NTK), that may inspire other researchers. However, the limitations of the analysis with respect to real-world image data augmentation, as well as the limited motivation, in my opinion, for the need for the proposed method, have an overall negative impact in my assessment of the significance of this work. I elaborate on these concerns below. \n\nFirst, the whole theoretical analysis, built upon that of Rajput et al. (2019), is based on modelling data augmentation as additive perturbations of the inputs. This is stated at the beginning of Section 2 and never recalled as a limitation again in the paper (except for one section in the supplementary material). On the contrary, the authors argue that this modelling of data augmentation captures typical real-world image transformations, such as \"translations, crops, rotations, and [...] other pixel-wise augmentation methods such as sharpening, blurring, and color distortions\" (second paragraph of Section 2). This is in sharp contrast with the limitations discussed by Rajput et al. (2019) themselves, on whose work the current paper takes inspiration for the theoretical analysis. In their Conclusion and Open Problems, Rajput et al. (2019) wrote: \"There are several interesting open problems that we plan to tackle in the future. First, it would be interesting to theoretically analyze practical state-of-the-art augmentation methods, such as random crops, flips, and rotations. Such perturbations often fall outside our framework\".\n\nIn line with Rajput et al. (2019), I would strongly argue that additive perturbations, while interesting from a mathematical point of view, only very weakly capture the extent of image transformations used in practice. Geometrical transformations can hardly be approximated by additive perturbations. One example of a widely common transformation that strongly differs from additive perturbations is horizontal flipping. This introduces a strong inductive bias, based on the properties of human visual perception, which is why it is included in almost all data augmentation schemes. Therefore, any conclusions from such a mathematical approximation to real-world data augmentation needs to be cautious. For example, the paper concludes saying that it has been shown \"that data augmentation improves training and generalization by enlarging the smaller singular values of the neural network Jacobian\". As argued above, this seems a strong claim given the limitations. The strength of the claims is particularly surprising after the authors wrote about other existing works on the theoretical analysis of data augmentation that they \"do not provide insights on the effect of data augmentation on training deep neural networks\".\n\nMy other main concern has probably a stronger impact on my overall assessment of the paper. This has to do with the motivation for the need for methods to select subsets of data on which to apply data augmentation. On the second sentence of the abstract, the authors write that \"modern data augmentation techniques become computationally prohibitive for large datasets\". This statement is extended in the introduction, where the authors cite several papers on so-called _automatic data augmentation_. First, these methods are indeed disproportionally expensive in computational terms, while providing only marginal improvements (in the best case) with respect to traditional, simple, cheap data augmentation techniques (Pérez and Wang, 2017). The authors mention that \"multiple augmented examples are usually generated for a single data point to obtain better results, increasing the size of the training data by orders of magnitude\", as though this was a weakness of data augmentation, while it is actually one of its main strengths. In particular, that the effective training size may be increased orders of magnitude while keeping the training time within the same order of magnitude. \n\nAn analysis of such advantages and the reasons for it (such as the fact that data augmentation can be performed in parallel to the parameter updates of the model, and even create a queue of data that would effectively keep the training time identical, given sufficient memory), as well as compelling evidence of the efficiency of data augmentation is provided by Hernández-García and König (2018). In that paper we see that training with light data augmentation (translations and horizontal flips) on the full training set provides large performance gains with a marginal increase of the training time. Furthermore, they also provide empirical evidence that training with 50 % data _and_ data augmentation (again on the full available set) achieves more than 95 % of the _full_ accuracy in about half the training time. Therefore, given that data augmentation can be applied almost _for free_ and training time can be traded by reducing the training data for a marginal reduction of the accuracy, why do we need a complex algorithm such as the one presented in this paper?\n\nIt is not clear from the paper how the training times are calculated, but we still see that in the best case the gains stay within the same order of magnitude. I believe that a stronger justification of the need for this method should take into account the considerations mentioned above, as well as a comparison with other alternative ways to trade training time for performance, such as changes in the architecture, etc.\n\n### References\n\n* Pérez and Wang. [The effectiveness of data augmentation in image classification using deep learning](https://arxiv.org/abs/1712.04621). 2017.\n* Hernández-García and König. [Data augmentation instead of explicit regularization](https://arxiv.org/abs/1806.03852). 2018", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper first provides a theoretical analysis, under some assumptions, of the effect of data augmentation on the singular values of the network Jacobian and then proposes a method to improve the sample complexity of data augmentation, that is to select a subset of the data on which to perform the transformations to speed up training. The proposed method aims to find subsets of data whose augmentation yields similar alignment of the Jacobian with the residual vector as the fully augmented data. ", "main_review": "The main contribution of this paper is, in my opinion, a sound theoretical analysis of the impact of input perturbations on the singular values of the Jacobian, or the eigenvalues of the Neural Tangent Kernel (NTK), that may inspire other researchers. However, the limitations of the analysis with respect to real-world image data augmentation, as well as the limited motivation, in my opinion, for the need for the proposed method, have an overall negative impact in my assessment of the significance of this work. I elaborate on these concerns below. \n\nFirst, the whole theoretical analysis, built upon that of Rajput et al. (2019), is based on modelling data augmentation as additive perturbations of the inputs. This is stated at the beginning of Section 2 and never recalled as a limitation again in the paper (except for one section in the supplementary material). On the contrary, the authors argue that this modelling of data augmentation captures typical real-world image transformations, such as \"translations, crops, rotations, and [...] other pixel-wise augmentation methods such as sharpening, blurring, and color distortions\" (second paragraph of Section 2). This is in sharp contrast with the limitations discussed by Rajput et al. (2019) themselves, on whose work the current paper takes inspiration for the theoretical analysis. In their Conclusion and Open Problems, Rajput et al. (2019) wrote: \"There are several interesting open problems that we plan to tackle in the future. First, it would be interesting to theoretically analyze practical state-of-the-art augmentation methods, such as random crops, flips, and rotations. Such perturbations often fall outside our framework\".\n\nIn line with Rajput et al. (2019), I would strongly argue that additive perturbations, while interesting from a mathematical point of view, only very weakly capture the extent of image transformations used in practice. Geometrical transformations can hardly be approximated by additive perturbations. One example of a widely common transformation that strongly differs from additive perturbations is horizontal flipping. This introduces a strong inductive bias, based on the properties of human visual perception, which is why it is included in almost all data augmentation schemes. Therefore, any conclusions from such a mathematical approximation to real-world data augmentation needs to be cautious. For example, the paper concludes saying that it has been shown \"that data augmentation improves training and generalization by enlarging the smaller singular values of the neural network Jacobian\". As argued above, this seems a strong claim given the limitations. The strength of the claims is particularly surprising after the authors wrote about other existing works on the theoretical analysis of data augmentation that they \"do not provide insights on the effect of data augmentation on training deep neural networks\".\n\nMy other main concern has probably a stronger impact on my overall assessment of the paper. This has to do with the motivation for the need for methods to select subsets of data on which to apply data augmentation. On the second sentence of the abstract, the authors write that \"modern data augmentation techniques become computationally prohibitive for large datasets\". This statement is extended in the introduction, where the authors cite several papers on so-called _automatic data augmentation_. First, these methods are indeed disproportionally expensive in computational terms, while providing only marginal improvements (in the best case) with respect to traditional, simple, cheap data augmentation techniques (Pérez and Wang, 2017). The authors mention that \"multiple augmented examples are usually generated for a single data point to obtain better results, increasing the size of the training data by orders of magnitude\", as though this was a weakness of data augmentation, while it is actually one of its main strengths. In particular, that the effective training size may be increased orders of magnitude while keeping the training time within the same order of magnitude. \n\nAn analysis of such advantages and the reasons for it (such as the fact that data augmentation can be performed in parallel to the parameter updates of the model, and even create a queue of data that would effectively keep the training time identical, given sufficient memory), as well as compelling evidence of the efficiency of data augmentation is provided by Hernández-García and König (2018). In that paper we see that training with light data augmentation (translations and horizontal flips) on the full training set provides large performance gains with a marginal increase of the training time. Furthermore, they also provide empirical evidence that training with 50 % data _and_ data augmentation (again on the full available set) achieves more than 95 % of the _full_ accuracy in about half the training time. Therefore, given that data augmentation can be applied almost _for free_ and training time can be traded by reducing the training data for a marginal reduction of the accuracy, why do we need a complex algorithm such as the one presented in this paper?\n\nIt is not clear from the paper how the training times are calculated, but we still see that in the best case the gains stay within the same order of magnitude. I believe that a stronger justification of the need for this method should take into account the considerations mentioned above, as well as a comparison with other alternative ways to trade training time for performance, such as changes in the architecture, etc.\n\n### References\n\n* Pérez and Wang. [The effectiveness of data augmentation in image classification using deep learning](https://arxiv.org/abs/1712.04621). 2017.\n* Hernández-García and König. [Data augmentation instead of explicit regularization](https://arxiv.org/abs/1806.03852). 2018", "summary_of_the_review": "While the paper seems solid in terms of correctness, I have a less positive impression due to, in my opinion, limited significance. This has to do with the simplification of data augmentation with respect to real world necessary for the theoretical analysis, as well as with the lack of strong motivation for the proposed algorithm.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635901242847}, {"id": "_Fy0z6wR1BO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1080/Reviewer_1cJe"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors model data augmentation as an additive perturbation and analyze its effect on training dynamics and how it enlarges the smaller singular values of the network jacobian.\nThen they propose a new method to iteratively extract a subset of the training data that when augmented closely capture the full augmented data dynamics.\nAuthors show that by augmenting this subset combined with full training data they can outperform the state-of-the-art method by 7.7% on CIFAR-10 and 4.7% on SVHN while achieving 6.3x and 2.2x speedup respectively.", "review_text": "**Strengths**\n\nI appreciate that this paper provides rigorous theoretical analysis to data augmentation, its connection to NTK and effect on eigenvalues and how data augmentation affects the training dynamics and generalization, and it also provides mathematical proofs to the theorems proposed.\n\nIn the presence of noisy labels, the proposed technique is outperforming reference methods.\n\nAnd even when the selected subset is very small compared to randomly augmenting datasets of similar size the proposed method is capable of providing good augmentation resulting and a decent increase in performance.\n\n**Weaknesses**\n\nAs I understand one of the observations in section 3.1 suggests that augmentation cannot change the singular values considerably for diverse datasets, but we still see a gain in performance when augmenting these datasets.\nMy concern is if we have a more diverse dataset than the ones experimented on, according to my understanding we will have a larger number of singular values and that will limit the ability to find a subset to augment and we need to increase the size of the subset which can largely limit the effectiveness of the method.\n\nIn my opinion, one of the main strengths of the method is when we have a very large dataset but unfortunately, the authors didn’t provide any experiments regarding large datasets like ImageNet.\n\nThe proposed method outperforms the baseline when the size of the subset is fairly small, but when increasing the size into reasonable percentages the improvement is marginal see Table 4 in appendix C.\n\n**Suggestion To Authors**\n\nThough the whole paper is nicely executed and demonstrated, in my opinion, some of the figures aren’t easy to read e.g. (Figure 4) I suggest splitting the figure into two side by side showing the speed gain and accuracy improvement separately.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors model data augmentation as an additive perturbation and analyze its effect on training dynamics and how it enlarges the smaller singular values of the network jacobian.\nThen they propose a new method to iteratively extract a subset of the training data that when augmented closely capture the full augmented data dynamics.\nAuthors show that by augmenting this subset combined with full training data they can outperform the state-of-the-art method by 7.7% on CIFAR-10 and 4.7% on SVHN while achieving 6.3x and 2.2x speedup respectively.", "main_review": "**Strengths**\n\nI appreciate that this paper provides rigorous theoretical analysis to data augmentation, its connection to NTK and effect on eigenvalues and how data augmentation affects the training dynamics and generalization, and it also provides mathematical proofs to the theorems proposed.\n\nIn the presence of noisy labels, the proposed technique is outperforming reference methods.\n\nAnd even when the selected subset is very small compared to randomly augmenting datasets of similar size the proposed method is capable of providing good augmentation resulting and a decent increase in performance.\n\n**Weaknesses**\n\nAs I understand one of the observations in section 3.1 suggests that augmentation cannot change the singular values considerably for diverse datasets, but we still see a gain in performance when augmenting these datasets.\nMy concern is if we have a more diverse dataset than the ones experimented on, according to my understanding we will have a larger number of singular values and that will limit the ability to find a subset to augment and we need to increase the size of the subset which can largely limit the effectiveness of the method.\n\nIn my opinion, one of the main strengths of the method is when we have a very large dataset but unfortunately, the authors didn’t provide any experiments regarding large datasets like ImageNet.\n\nThe proposed method outperforms the baseline when the size of the subset is fairly small, but when increasing the size into reasonable percentages the improvement is marginal see Table 4 in appendix C.\n\n**Suggestion To Authors**\n\nThough the whole paper is nicely executed and demonstrated, in my opinion, some of the figures aren’t easy to read e.g. (Figure 4) I suggest splitting the figure into two side by side showing the speed gain and accuracy improvement separately.\n", "summary_of_the_review": "I think the paper is of interest, especially the theoretical analysis. It is well written and provides rigorous mathematical analysis to data augmentation and proofs for all the theorems, lemmas and corollaries mentioned.\n\nBut I believe the main application of the proposed method is in large scale dataset settings, while in small or medium-size datasets augmenting the full dataset is not an issue, but that wasn’t presented (see my concerns in the previous section).\nAlso, the improvement when increasing subsampling size is marginal, especially with diverse datasets.\nThis problem and the lack of experiments with large and diverse datasets prevents me from giving a higher recommendation of acceptance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635858700598}, {"id": "fvNJoaoeqKk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1080/Reviewer_Q8Ag"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper shows that data augmentation can speed up learning by enlarging the smaller singular values of the Jacobian. Following this idea, this paper proposed a framework to iteratively extract small subsets of training data that captures the alignment of NTK with the residual when augmented.\n", "review_text": "Strengths:\n1) The motivation of the paper is clearly justified. \n2) The presentation of the theoretical analysis and the algorithm are clear and sound.\n3) The results shows improvement over the max-loss method on some datasets.\n\nWeakness:\n1) What is the extra computation cost for finding the coreset compared with that of  the random and max-loss baseline?\n2) Does the coreset change significantly at different epochs?\n3) It would be interesting to show the performance when transferring the core-set found on one architecture (e.g., ResNet20) to train on a different architecture (e.g., Wid-ResNet). Will transfer the corset leads to performance loss?\n4) You might referred to the wrong figure in the sentence “Figure 3b depicts the increase in intersection between max-loss subsets and coresets over time”\n5) The paper only did experiments on small datasets other than large datasets such as ImageNet.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper shows that data augmentation can speed up learning by enlarging the smaller singular values of the Jacobian. Following this idea, this paper proposed a framework to iteratively extract small subsets of training data that captures the alignment of NTK with the residual when augmented.\n", "main_review": "Strengths:\n1) The motivation of the paper is clearly justified. \n2) The presentation of the theoretical analysis and the algorithm are clear and sound.\n3) The results shows improvement over the max-loss method on some datasets.\n\nWeakness:\n1) What is the extra computation cost for finding the coreset compared with that of  the random and max-loss baseline?\n2) Does the coreset change significantly at different epochs?\n3) It would be interesting to show the performance when transferring the core-set found on one architecture (e.g., ResNet20) to train on a different architecture (e.g., Wid-ResNet). Will transfer the corset leads to performance loss?\n4) You might referred to the wrong figure in the sentence “Figure 3b depicts the increase in intersection between max-loss subsets and coresets over time”\n5) The paper only did experiments on small datasets other than large datasets such as ImageNet.\n\n", "summary_of_the_review": "This paper proposed a data efficient augmentation framework with extensive theoretical analysis. The presentation and the proposed method are sound. However, the lack of experimental results on large datasets such as ImageNet makes the results of the paper much less convincing.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635824560090}], "openreview_url": "https://openreview.net/forum?id=SuKTLF9stD", "arxiv_id": "2210.08363", "paper_pdf": "papers/SuKTLF9stD.pdf", "paper_pdf_sha256": "d0ef9e2cc5bbed07e1ba995525730f3a97352b34e593cf65f0d05bc1a466572b", "paper_pdf_bytes": 1240777, "paper_pdf_source": "openreview", "code_url": "https://github.com/tianyu139/data-efficient-augmentation", "code_repository": "tianyu139/data-efficient-augmentation", "code_commit": "3abb2f7d3fa610fe2394063872c55927bbccde4a", "code_archive": "repos/SuKTLF9stD.zip", "code_archive_sha256": "9b3892e237a60be61bab31e12bc1840e7e4bbb48b554d261dbb4da4451f4e734", "code_archive_bytes": 1565834, "code_file_count": 21, "code_extensions": {".py": 19, ".sh": 2}, "github_disk_usage_kb": 1517, "github_languages": {"Python": 110230, "Shell": 482}, "github_archived": false, "github_pushed_at": "2023-07-20T04:42:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/data-efficient-augmentation-for-training-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wl0Kr_jqM2a", "year": 2021, "status": "rejected", "title": "Testing Robustness Against Unforeseen Adversaries", "authors": ["Daniel Kang", "Yi Sun", "Dan Hendrycks", "Tom B Brown", "Jacob Steinhardt"], "authorids": ["~Daniel_Kang1", "~Yi_Sun3", "~Dan_Hendrycks1", "~Tom_B_Brown1", "~Jacob_Steinhardt1"], "authors_source": "OpenReview API", "abstract": "Most existing adversarial defenses only measure robustness to $L_p$ adversarial attacks. Not only are adversaries unlikely to exclusively create small $L_p$ perturbations, adversaries are unlikely to remain fixed. Adversaries adapt and evolve their attacks; hence adversarial defenses must be robust to a broad range of unforeseen attacks. We address this discrepancy between research and reality by proposing a new evaluation framework called ImageNet-UA. Our framework enables the research community to test ImageNet model robustness against attacks not encountered during training. To create ImageNet-UA's diverse attack suite, we introduce a total of four novel adversarial attacks. We also demonstrate that, in comparison to ImageNet-UA, prevailing $L_\\infty$ robustness assessments give a narrow account of adversarial robustness. By evaluating current defenses with ImageNet-UA, we find they provide little robustness to unforeseen attacks. We hope the greater variety and realism of ImageNet-UA enables development of more robust defenses which can generalize beyond attacks seen during training.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gtTxBXqdNoZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2332/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Summary\n\nThe authors propose ImageNet-UA, a framework to evaluate the\neffectiveness of test-time unforeseen adversarial ML attacks. In this\nparticular context, the authors observe that attackers are not limited\nto induce minimal Lp-norm perturbations and therefore it is important to\nunderstand how robust existing defenses are against attacks that diverge\nfrom these constraints while still being realizable. Within their\nframework, the authors propose four novel adversarial attacks to\nconsolidate our understanding of test-time adversarial ML attacks.\n\n## Strengths\n\n+  Interesting intuition that test-time adversarial attacks are not all\nnecessarily driven by minimal perturbations\n+  Attack-evaluation framework available to the community\n\n## Weaknesses\n\n-  Recent work explores the need for having a unified theoretical\n   framework to reason about realizable attacks\n-  Actual benefit of the approach\n\n## Comments\n\nThis is an interesting paper that explores an avenue that is often\nneglected in the context of test-time adversarial ML attacks. There is a\ntendency on focusing on attacks that minimize the perturbation in a\ngiven Lp-norm, when in fact there are several application domains for\nwhich this constraint does not necessarily matter. Here, in the context\nof realizable or problem-space attacks, one is required to reformulate\nthe problem, which may include other constraints attacks must satisfy.\n\nAs the authors admit, ImageNet-UA is not exhaustive although it tries to\nevaluate the robustness of defenses over diverse unforeseen test\n(adversarial) distributions. Under this premise, I somehow fail to see\nthe real benefit of such a framework, unless the authors show such\nadditional attacks comprise widely-used transformations that should\ntherefore become part of the threat model of any adversarial work in the\nproblem space. (In addition, most of the attacks seem to be a marginal\nimprovement of existing ones.) Conversely, a theoretical reformulation of the\nproblem-space would have been perhaps more useful. That would have\nallowed one to reason about properties or set of transformations one\ncould or should consider, along with other constraints, as outlined in\n[1]. \n\nI also wonder what's the rationale behind the mUAR metric. It is\ndefinitely useful to have a way to capture the robustness of models\nagainst unforeseen attacks, but I wonder whether an average as opposed\nto, e.g., a geometric mean, would provide useful insights. In a way, the\naverage would consider all the attacks equally challenging where, in\nreality, this might be the case. (I do actually like the use of mUAR and UAR tho,\nas outlined in Section 5.2).\n\n## Additional comments\n\nIn Section 4, the authors explain how the distortion size \\epsilon_max\nis chosen: \"[...] the smallest ε which either reduces adversarial\naccuracy of an adversarially trained model at distortion size ε below 25\nor yields images confusing humans\". How is this verified? Do the authors\nrely on a user study? Does this only rely on Gilmer et al. (2018)?\n\n[1] https://s2lab.kcl.ac.uk/projects/intriguing/ (IEEE S&P 2020)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "ICLR 2021 Conference Paper2332 AnonReviewer4 Review", "review": "## Summary\n\nThe authors propose ImageNet-UA, a framework to evaluate the\neffectiveness of test-time unforeseen adversarial ML attacks. In this\nparticular context, the authors observe that attackers are not limited\nto induce minimal Lp-norm perturbations and therefore it is important to\nunderstand how robust existing defenses are against attacks that diverge\nfrom these constraints while still being realizable. Within their\nframework, the authors propose four novel adversarial attacks to\nconsolidate our understanding of test-time adversarial ML attacks.\n\n## Strengths\n\n+  Interesting intuition that test-time adversarial attacks are not all\nnecessarily driven by minimal perturbations\n+  Attack-evaluation framework available to the community\n\n## Weaknesses\n\n-  Recent work explores the need for having a unified theoretical\n   framework to reason about realizable attacks\n-  Actual benefit of the approach\n\n## Comments\n\nThis is an interesting paper that explores an avenue that is often\nneglected in the context of test-time adversarial ML attacks. There is a\ntendency on focusing on attacks that minimize the perturbation in a\ngiven Lp-norm, when in fact there are several application domains for\nwhich this constraint does not necessarily matter. Here, in the context\nof realizable or problem-space attacks, one is required to reformulate\nthe problem, which may include other constraints attacks must satisfy.\n\nAs the authors admit, ImageNet-UA is not exhaustive although it tries to\nevaluate the robustness of defenses over diverse unforeseen test\n(adversarial) distributions. Under this premise, I somehow fail to see\nthe real benefit of such a framework, unless the authors show such\nadditional attacks comprise widely-used transformations that should\ntherefore become part of the threat model of any adversarial work in the\nproblem space. (In addition, most of the attacks seem to be a marginal\nimprovement of existing ones.) Conversely, a theoretical reformulation of the\nproblem-space would have been perhaps more useful. That would have\nallowed one to reason about properties or set of transformations one\ncould or should consider, along with other constraints, as outlined in\n[1]. \n\nI also wonder what's the rationale behind the mUAR metric. It is\ndefinitely useful to have a way to capture the robustness of models\nagainst unforeseen attacks, but I wonder whether an average as opposed\nto, e.g., a geometric mean, would provide useful insights. In a way, the\naverage would consider all the attacks equally challenging where, in\nreality, this might be the case. (I do actually like the use of mUAR and UAR tho,\nas outlined in Section 5.2).\n\n## Additional comments\n\nIn Section 4, the authors explain how the distortion size \\epsilon_max\nis chosen: \"[...] the smallest ε which either reduces adversarial\naccuracy of an adversarially trained model at distortion size ε below 25\nor yields images confusing humans\". How is this verified? Do the authors\nrely on a user study? Does this only rely on Gilmer et al. (2018)?\n\n[1] https://s2lab.kcl.ac.uk/projects/intriguing/ (IEEE S&P 2020)", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604857885674}, {"id": "hKTPvaEqPss", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2332/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Update**: Thanks to the authors for addressing my comments and releasing the source code. Code is well-structured and easy to follow. It can be definitely used as a supplement for the robustness evaluation. However, judging by the implementation, the rain, snow, and fog attacks are too simplistic. From the paper, it is not clear how well these attacks approximate the respective type of perturbations in real-world scenarios. Therefore, the proposed set of 6 attacks is more the author's heuristic than something that can be used for the evaluation of computer vision systems, e.g. autonomous vehicles. Given all that, I decrease my score from 5 to 4.\n\n###### Summary\nThe paper proposes a novel benchmark for the evaluation of the model's robustness against unforeseen adversaries. The authors suggested using 6 types of adversaries: JPEG, FOG, SNOW, and GABOR attack, L1, ELASTIC. Fully-differentiable variants of the above attacks are based on the previous work. The authors introduced the novel $l_1$-norm attack, which uses the Frank-Wolf algorithm to satisfy $l_1$-perturbation constraint. The authors introduced two metrics to compare defenses: robustness against a single unforeseen attack and mean unforeseen robustness. Experiments with adversarially trained models and non-adversarially trained models align with the previous work in the area.\n\n###### Reasons for score: \nI vote for a weak accept. The paper studies an important problem of the robustness of unforeseen adversaries. Introducing novel benchmarks beyond $l_{p}$-norm robustness is an important problem. The paper is clearly written and easy to follow. The evaluation results are extensive. However, implementation details of some attacks (SNOW, FOG, GABOR) are missing. JPEG attack is based on the previous work. It is not clear if the authors plan to release the benchmark suite to the reviewers. The most similar work in Hendrycks et al. 2019 uses a more diverse set of attacks.\n\n###### Concerns:\n- Novel differentiable attacks are based on previous work, e.g. differentiable JPEG and FOG. The algorithm and implementation details for SNOW and GABOR attacks are not provided. It is not clear how to reimplement these attacks.\n- The paper is the most similar to Hendrycks & Dietterich 2019. The main difference is the use of differentiable attacks. However, previous work in Hendrycks uses a larger set of corruptions. This work includes only 5 corruptions from Hendrycks et al. 2019. Do the authors plan to include other types of attacks for the Imagenet-UA benchmark?\n- It is not clear from the main text if the authors plan to release the benchmark source code during the reviewing process.\n  \n###### Minor comments:\n- Comment that Elastic $l_1$ attack is based on heuristics might be incorrect. Elastic $l_1$ attack uses a proximity operator, a principal way to minimize $l_1$-norm.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper proposes a novel benchmark for evaluation of the model's robustness against unforeseen adversaries. The authors should provide implementation details of the novel attacks and make the source code available during the reviewing process.", "review": "**Update**: Thanks to the authors for addressing my comments and releasing the source code. Code is well-structured and easy to follow. It can be definitely used as a supplement for the robustness evaluation. However, judging by the implementation, the rain, snow, and fog attacks are too simplistic. From the paper, it is not clear how well these attacks approximate the respective type of perturbations in real-world scenarios. Therefore, the proposed set of 6 attacks is more the author's heuristic than something that can be used for the evaluation of computer vision systems, e.g. autonomous vehicles. Given all that, I decrease my score from 5 to 4.\n\n###### Summary\nThe paper proposes a novel benchmark for the evaluation of the model's robustness against unforeseen adversaries. The authors suggested using 6 types of adversaries: JPEG, FOG, SNOW, and GABOR attack, L1, ELASTIC. Fully-differentiable variants of the above attacks are based on the previous work. The authors introduced the novel $l_1$-norm attack, which uses the Frank-Wolf algorithm to satisfy $l_1$-perturbation constraint. The authors introduced two metrics to compare defenses: robustness against a single unforeseen attack and mean unforeseen robustness. Experiments with adversarially trained models and non-adversarially trained models align with the previous work in the area.\n\n###### Reasons for score: \nI vote for a weak accept. The paper studies an important problem of the robustness of unforeseen adversaries. Introducing novel benchmarks beyond $l_{p}$-norm robustness is an important problem. The paper is clearly written and easy to follow. The evaluation results are extensive. However, implementation details of some attacks (SNOW, FOG, GABOR) are missing. JPEG attack is based on the previous work. It is not clear if the authors plan to release the benchmark suite to the reviewers. The most similar work in Hendrycks et al. 2019 uses a more diverse set of attacks.\n\n###### Concerns:\n- Novel differentiable attacks are based on previous work, e.g. differentiable JPEG and FOG. The algorithm and implementation details for SNOW and GABOR attacks are not provided. It is not clear how to reimplement these attacks.\n- The paper is the most similar to Hendrycks & Dietterich 2019. The main difference is the use of differentiable attacks. However, previous work in Hendrycks uses a larger set of corruptions. This work includes only 5 corruptions from Hendrycks et al. 2019. Do the authors plan to include other types of attacks for the Imagenet-UA benchmark?\n- It is not clear from the main text if the authors plan to release the benchmark source code during the reviewing process.\n  \n###### Minor comments:\n- Comment that Elastic $l_1$ attack is based on heuristics might be incorrect. Elastic $l_1$ attack uses a proximity operator, a principal way to minimize $l_1$-norm.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604397126123}, {"id": "Mxm69kh1WQl", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2332/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes four novel efficient adversarial attack methods beyond Lp threat models. Together with other two existing attack methods, these six attack methods combine as a framework to evaluate robustness of defenses against unforeseen attacks. In this framework, the novel measure is normalized with the performance of adversarial training. The experiments show that the Linf adversarially trained model may not lead to improvement of robustness against other threat models. It is expected that the framework could help test model robustness.\n\n## Advantages\n\n- Evaluating model robustness against unforeseen adversaries is an important problem.\n- The four proposed adversarial attacks are differentiable and easy to use.\n\n## Disadvantages\n\n- The proposed measure mUAR is quite heuristic, e.g., the choice of the epsilon parameter.\n- As a baseline framework, it is hard to say that the six adversarial attacks are sufficient to evaluate robustness against unforeseen adversaries. As a result, the claims seem not to be very convincing.\n- In general, the result that adversarial training is not robust to unforeseen attacks is not surprising and has be been discussed a lot in literature.  It would be good to evaluate more existing defenses in order to derive more novel, interesting and inspiring insights.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good motivation but not unconvincing enough", "review": "This paper proposes four novel efficient adversarial attack methods beyond Lp threat models. Together with other two existing attack methods, these six attack methods combine as a framework to evaluate robustness of defenses against unforeseen attacks. In this framework, the novel measure is normalized with the performance of adversarial training. The experiments show that the Linf adversarially trained model may not lead to improvement of robustness against other threat models. It is expected that the framework could help test model robustness.\n\n## Advantages\n\n- Evaluating model robustness against unforeseen adversaries is an important problem.\n- The four proposed adversarial attacks are differentiable and easy to use.\n\n## Disadvantages\n\n- The proposed measure mUAR is quite heuristic, e.g., the choice of the epsilon parameter.\n- As a baseline framework, it is hard to say that the six adversarial attacks are sufficient to evaluate robustness against unforeseen adversaries. As a result, the claims seem not to be very convincing.\n- In general, the result that adversarial training is not robust to unforeseen attacks is not surprising and has be been discussed a lot in literature.  It would be good to evaluate more existing defenses in order to derive more novel, interesting and inspiring insights.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603924831940}, {"id": "_oYvtEQmlvm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2332/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents the existence of several adversarial attacks that have meaningful visual concepts. Based on these attacks, the authors further develop a method to measure the robustness of a network or an adversarial defense. \n\nDespite being conceptually interesting, this paper is working on the thing that many studies have done. The related work section is not on the point, in my opinion, the core problem (''unforeseen adversaries'') of this work should be ''adversarial attacks that are designed to break the commonly employed defenses'', but not ''adversarial attacks that are visually meaningful, and occasionally break the commonly employed defenses''. Therefore, the related work should focus on these studies that propose measures of robustness, and the experimental analysis should focus on doing a comparison of these measures, telling why the readers should adopt the ImageNet-UA and mUAR as their baseline, but not others.\n\nOn the other hand, the analysis of these newly developed is not sufficient, and the experimental results can not support the claims convincingly. For instance, in Table 3, the JPEG attack acquires the worst performance and the data augmentation methods do not help in all scenarios. Meanwhile, the Fog and Snow attacks are less effective than the $L_\\inf$ and $L_2$ and benefit from more data. This implies they have different dynamics thus should not be used for mUAR homogeneously.\n\nAnyway, I suggest the authors carefully revise the motivation and experimental analysis.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting idea", "review": "This paper presents the existence of several adversarial attacks that have meaningful visual concepts. Based on these attacks, the authors further develop a method to measure the robustness of a network or an adversarial defense. \n\nDespite being conceptually interesting, this paper is working on the thing that many studies have done. The related work section is not on the point, in my opinion, the core problem (''unforeseen adversaries'') of this work should be ''adversarial attacks that are designed to break the commonly employed defenses'', but not ''adversarial attacks that are visually meaningful, and occasionally break the commonly employed defenses''. Therefore, the related work should focus on these studies that propose measures of robustness, and the experimental analysis should focus on doing a comparison of these measures, telling why the readers should adopt the ImageNet-UA and mUAR as their baseline, but not others.\n\nOn the other hand, the analysis of these newly developed is not sufficient, and the experimental results can not support the claims convincingly. For instance, in Table 3, the JPEG attack acquires the worst performance and the data augmentation methods do not help in all scenarios. Meanwhile, the Fog and Snow attacks are less effective than the $L_\\inf$ and $L_2$ and benefit from more data. This implies they have different dynamics thus should not be used for mUAR homogeneously.\n\nAnyway, I suggest the authors carefully revise the motivation and experimental analysis.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603897103209}], "openreview_url": "https://openreview.net/forum?id=wl0Kr_jqM2a", "arxiv_id": "1908.08016", "paper_pdf": "papers/wl0Kr_jqM2a.pdf", "paper_pdf_sha256": "893b51cc924334850f18fef87f9947ec32fd44865d634df179c3089c73e6ea10", "paper_pdf_bytes": 2441782, "paper_pdf_source": "openreview", "code_url": "https://github.com/ddkang/advex-uar", "code_repository": "ddkang/advex-uar", "code_commit": "fd9e0efa1155cedb6d15942ed6e22e1bf1252043", "code_archive": "repos/wl0Kr_jqM2a.zip", "code_archive_sha256": "ab713cf26ce1e92ac0eb2b776b55f5d589a72047cf8242639eff317eb983f3cb", "code_archive_bytes": 1215649, "code_file_count": 34, "code_extensions": {".py": 34}, "github_disk_usage_kb": 1176, "github_languages": {"Python": 132078}, "github_archived": false, "github_pushed_at": "2024-07-25T10:16:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/190808016"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HkxcUxrFPS", "year": 2020, "status": "rejected", "title": "Improving Visual Relation Detection using Depth Maps", "authors": ["Sahand Sharifzadeh", "Sina Moayed Baharlou", "Max Berrendorf", "Rajat Koner", "Volker Tresp"], "authorids": ["sharifzadeh@dbs.ifi.lmu.de", "sina.baharlou@gmail.com", "berrendorf@dbs.ifi.lmu.de", "koner@dbs.ifi.lmu.de", "volker.tresp@siemens.com"], "authors_source": "OpenReview API", "abstract": "State of the art visual relation detection methods mostly rely on object information extracted from RGB images such as predicted class probabilities, 2D bounding boxes and feature maps. In this paper, we argue that the 3D positions of objects in space can provide additional valuable information about object relations. This information helps not only to detect spatial relations, such as \\textit{standing behind}, but also non-spatial relations, such as \\textit{holding}. Since 3D information of a scene is not easily accessible, we propose incorporating a pre-trained RGB-to-Depth model within visual relation detection frameworks. We discuss different feature extraction strategies from depth maps and show their critical role in relation detection.\nOur experiments confirm that the performance of state-of-the-art visual relation detection approaches can significantly be improved by utilizing depth map information.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SylGDVKEqS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2333/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "OVERVIEW:\nThe authors propose to use depth information to better predict the visual relation between objects in an image. They do this by incorporating a pre-trained RGB-to-Depth model within existing frameworks. They claim the following contributions:\n1. First to utilize 3D information in visual relation detection. They synthesize depth images for existing benchmark datasets of VRD and VG using a pre-trained RGB-to-Depth model trained on NYUv2 to generate RGB-D data for visual relation detection.\n2. Discuss and empirically investigate different strategies to extract features from depth maps for relation detection.\n3. Study the quantitative and qualitative benefits of incorporating depth maps. \"We show in our empirical evaluation using the VRD and VG datasets, that models using depth maps can outperform competing methods by a margin of up to 3% points\".\n\nMAJOR COMMENTS:\n1. I liked the idea of using depth information to inform visual relationships but I am not sure if the proposed approach is the way to go. Given a depth image of the scene, we can generate a reconstruction of the scene in 3D, even if it is partial/imperfect. Direct reasoning in 3D should now be possible instead of going via deep networks as proposed in the paper. I believe a direct 3D approach would make a meaningful baseline at the very least and needs to be discussed.\n2. The authors use a pre-trained RGB-to-Depth network trained on NYU-v2 to predict depth for the images of VRD and VG. There is very little discussion about the quality of predicted depth maps. Ideally, this needs to be quantified to convince the reader that the generated depth maps are \"good\" but at the very least the authors need to show qualitative examples (both good, typical and bad) to prove that the pre-trained network generates meaningful depth maps.\n3. To use a siamese (shared weights) feature extractor between RGB and Depth images or not, is not a significant contribution by itself. In principle, separate feature extractors lead to larger model complexity/learning capability and make sense given domain separation between RGB and Depth. \n\nMINOR COMMENTS:\n1. Figure 2 seems to indicate that a Faster-RCNN is used on both RGB and Depth steams which is backed up by text in Section 2 (first paragraph). However, in Section 3.2, under RGB Feature Extraction and Depth Map Feature Extraction, the discussion is about VGG-16 and AlexNet-BN networks. The VGG-16 network is pre-trained in ImageNet and finetuned to relevant data but it is not clear for what task? If the task is object detection, it needs to be trained for it (not fine-tuned, unless it is being initialized from COCO pre-training). The AlexNet-BN depth model is trained for relation detection using only depth. But it is not clear if it is using proposals/boxes generated by RGB detection model or using ground-truth boxes. Basically, the object-detection component of the pipeline is not clear at all.\n\nNOTE:\nI would like to mention that I have published in monocular object pose estimation and work in the object recognition. I am not as familiar with the visual relation detection field but I understand all the components proposed by the authors in this work. I believe I understood the paper and reviewed it fairly (to the best of my ability).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "OVERVIEW:\nThe authors propose to use depth information to better predict the visual relation between objects in an image. They do this by incorporating a pre-trained RGB-to-Depth model within existing frameworks. They claim the following contributions:\n1. First to utilize 3D information in visual relation detection. They synthesize depth images for existing benchmark datasets of VRD and VG using a pre-trained RGB-to-Depth model trained on NYUv2 to generate RGB-D data for visual relation detection.\n2. Discuss and empirically investigate different strategies to extract features from depth maps for relation detection.\n3. Study the quantitative and qualitative benefits of incorporating depth maps. \"We show in our empirical evaluation using the VRD and VG datasets, that models using depth maps can outperform competing methods by a margin of up to 3% points\".\n\nMAJOR COMMENTS:\n1. I liked the idea of using depth information to inform visual relationships but I am not sure if the proposed approach is the way to go. Given a depth image of the scene, we can generate a reconstruction of the scene in 3D, even if it is partial/imperfect. Direct reasoning in 3D should now be possible instead of going via deep networks as proposed in the paper. I believe a direct 3D approach would make a meaningful baseline at the very least and needs to be discussed.\n2. The authors use a pre-trained RGB-to-Depth network trained on NYU-v2 to predict depth for the images of VRD and VG. There is very little discussion about the quality of predicted depth maps. Ideally, this needs to be quantified to convince the reader that the generated depth maps are \"good\" but at the very least the authors need to show qualitative examples (both good, typical and bad) to prove that the pre-trained network generates meaningful depth maps.\n3. To use a siamese (shared weights) feature extractor between RGB and Depth images or not, is not a significant contribution by itself. In principle, separate feature extractors lead to larger model complexity/learning capability and make sense given domain separation between RGB and Depth. \n\nMINOR COMMENTS:\n1. Figure 2 seems to indicate that a Faster-RCNN is used on both RGB and Depth steams which is backed up by text in Section 2 (first paragraph). However, in Section 3.2, under RGB Feature Extraction and Depth Map Feature Extraction, the discussion is about VGG-16 and AlexNet-BN networks. The VGG-16 network is pre-trained in ImageNet and finetuned to relevant data but it is not clear for what task? If the task is object detection, it needs to be trained for it (not fine-tuned, unless it is being initialized from COCO pre-training). The AlexNet-BN depth model is trained for relation detection using only depth. But it is not clear if it is using proposals/boxes generated by RGB detection model or using ground-truth boxes. Basically, the object-detection component of the pipeline is not clear at all.\n\nNOTE:\nI would like to mention that I have published in monocular object pose estimation and work in the object recognition. I am not as familiar with the visual relation detection field but I understand all the components proposed by the authors in this work. I believe I understood the paper and reviewed it fairly (to the best of my ability)."}, "tcdate": 1572275289926}, {"id": "SylMkK2CYH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2333/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to leverage the depth information for relation prediction, arguing that the depth information benefit the prediction of some predicates. To solve the lack of 3D data, an RGB-to-Depth model is trained on external available dataset and then applied to images from visual relation dataset. In the experiments, they investigate different strategies to extract features from depth maps and the explore effectiveness of depth information by comparing the model that only used depth map as input with those which use RGB information. The comparisons with other methods and ablation studies under both Zero-shot setting and normal setting demonstrate the effectiveness of depth information.\n\n+Strength:\n(1) The motivation is reasonable and what the authors make an attempt to explore is very meaningful. Visual relation especially the spatial relation is not likely to be predicted accurately without 3D information. In other words, it seems that visual relation prediction task will be extended to 3D images rather than staying within 2D images. Thus what the authors do is a good exploration for further extensions.\n(2) Comparisons with previous methods and the results show that the depth information is useful to some extent, but not so obvious.\n(3) The writing of this article is good and it’s very easy to understand.\n\n-Weakness:\n(1) The RGB-to-Depth Network is pretrained on other dataset. Is there any gap when it is used for VG or VRD dataset? \n(2) Although the depth map feature extraction seems to work well, it seems to be a little trivial. Why a CNN, e.g. AlexNet, or VGG, can be used to extract depth features? And why the AlexNet trained from scratch performs better than AlexNet pretrained on RGB images for object detection task and VGG net? If the author can give more explanations, this part will be more insightful.\n(3) From the plot which shows the top 10 percent absolute changes in prediction performance per predicate, the advantage of Depth is not obvious compared with RGB. And Depth does not bring the advantage claimed in Abstract. It’s a little hard to understand why depth information can rectify the prediction of (Tower, taller, trees). To sum up, the qualitative results are not so satisfying.\n(4) In Table 1, what really functions seems to be c_so, v_so, and l_so, while the improvement brought by depth is limited.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This paper proposes to leverage the depth information for relation prediction, arguing that the depth information benefit the prediction of some predicates. To solve the lack of 3D data, an RGB-to-Depth model is trained on external available dataset and then applied to images from visual relation dataset. In the experiments, they investigate different strategies to extract features from depth maps and the explore effectiveness of depth information by comparing the model that only used depth map as input with those which use RGB information. The comparisons with other methods and ablation studies under both Zero-shot setting and normal setting demonstrate the effectiveness of depth information.\n\n+Strength:\n(1) The motivation is reasonable and what the authors make an attempt to explore is very meaningful. Visual relation especially the spatial relation is not likely to be predicted accurately without 3D information. In other words, it seems that visual relation prediction task will be extended to 3D images rather than staying within 2D images. Thus what the authors do is a good exploration for further extensions.\n(2) Comparisons with previous methods and the results show that the depth information is useful to some extent, but not so obvious.\n(3) The writing of this article is good and it’s very easy to understand.\n\n-Weakness:\n(1) The RGB-to-Depth Network is pretrained on other dataset. Is there any gap when it is used for VG or VRD dataset? \n(2) Although the depth map feature extraction seems to work well, it seems to be a little trivial. Why a CNN, e.g. AlexNet, or VGG, can be used to extract depth features? And why the AlexNet trained from scratch performs better than AlexNet pretrained on RGB images for object detection task and VGG net? If the author can give more explanations, this part will be more insightful.\n(3) From the plot which shows the top 10 percent absolute changes in prediction performance per predicate, the advantage of Depth is not obvious compared with RGB. And Depth does not bring the advantage claimed in Abstract. It’s a little hard to understand why depth information can rectify the prediction of (Tower, taller, trees). To sum up, the qualitative results are not so satisfying.\n(4) In Table 1, what really functions seems to be c_so, v_so, and l_so, while the improvement brought by depth is limited."}, "tcdate": 1571895514251}, {"id": "SkenUk7MFB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2333/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n\n********* Post Rebuttal *********\n\nI appreciate the authors' effort in providing thorough responses and revised manuscript. \n\nI agree with the authors that \"the finding not being surprising\" is not a ground for rejection. I tried to word my final decision carefully but it seems it has still caused confusion for the authors. As I have mentioned in my original review, the rating was a result of the 4 points considered *together* . \n\nThat is, if one exploits privileged information that needs extra sensory data and/or annotation (point 2), *and*, this privileged information is clearly related and thus should be normally useful for the final task (point 1), *and* achieve marginal improvements (point 3), it can be a ground for rejection. Especially, given that prior works with similar arguments exists (point 4). \n\nThe rebuttal has alleviated the issue of marginal improvements (point 3) by introducing meanR@K (or as the revised paper refer to it, Macro R@K). Here, the improvements are more significant both compared to the state of the art and ablated baselines.\n\nThe authors also argue that the related [Yang et al. 2018] paper (point 4) should be considered a concurrent submission since the authors original submission was to AAAI18. \n\nThe rebuttal also addresses other clarity or experimental issues which improves the quality of the revised work.\n\nFinally, I understand that the privileged information is only required during training time which is a good point.\n\nAll in all, *assuming that [Yang et al. 2018] is considered a concurrent work* according to ICLR, I think the revised paper becomes slightly above borderline and thus I change my rating to \"weak accept\". If [Yang et al. 2018] is not considered concurrent work, then, a conclusive comparison is required for the acceptance of the current work.\n\n\n********* Summary *********\n \nThe paper poses the question of whether depth information is informative for visual relationship prediction using still images. It is intuitive that 3D arrangement of objects in an image can be a useful cue for predicting their relationship. As such it is important to see whether and to what extent depth information complements RGB information for visual relation detection. That is the focus of this paper.\nThe paper proposes to use an off-the-shelf monocular depth estimation networks to augment the available RGB information towards better visual relation detection. For that, it proposes a specific network two-stream structure working on RGB image and (predicted) depth image. The proposed model demonstrates improved results upon state of the art for visual relation prediction.\n \n\n********* Strengths and Weaknesses *********\n \n+ A comprehensive set of tests has been conducted. \n+ Zero-shot prediction results are particularly interesting.\n+ The experiment on ranking the predicate classes based on the change in prediction accuracy before and after using depth information (Figure 4) is interesting and intuitive.\n* The final results improve upon the state-of-the-art, especially on the zero-shot learning regime. However, it seems that the improvement is mainly coming from the new architecture as opposed to the inclusion of the depth information. That is, ours_{c,v,l} brings most of the improvement already the last step to ours_{c,v,l,d} is negligible for non-zero-shot case.\n- Along the same line, it’s possible that this small difference between ours_{c,v,l} and ours_{c,v,l,d} for the standard predicate prediction, can be due to a hyper-parameter optimization that is (only or more thoroughly) done for ours_{c,v,l,d}. The hyper-parameter optimization scheme for different baselines is not described. \n- Given the small difference of ablation levels, the comparison will be stronger if done multiple times and reporting mean and standard deviation of the results.\n- For a fair comparison the visual feature vector v_{so} should be tried as the feature of the union bound box of both subject and object same way as it is done for depth feature vector d_{so}. \n- The paper refers to “Ours-d’_{so}” as a baseline that *only* uses depth information with no image/label information. However, it seems that the region proposals for this feature are coming from the image-based network that uses image information. \n \n- Important related but uncited works:\n(1) [“Visual Relationship Prediction via Label Clustering and Incorporation of Depth Information” ECCV workshops 2018] studies the same question as part of their work.\n \n\n********* Final Decision *********\n \nI do not find the paper passing the acceptance bar mainly due to the following reasons together:\n1) The finding is not surprising since most of the visual relations are either explicitly depth-related (e.g., behind) or are semantically constrained by depth (e.g. riding cannot happen at different depths when the image is taken orthogonal to the rider).\n2) an additional depth dataset is used which provides the model with privileged information. Should it have been the case that depth information were inferred without an additional offline dataset, the results would have been more interesting.\n3) the improvements due to the additional depth network are not significant or conclusive.\n4) there is a prior uncited work with the same research question for effectiveness of depth information in visual relation detection which uses a similar approach.\n \n\n********* Minor points *********\n\n- the code is not available. This is especially important since the paper is outperforming prior works which could be a contribution if reproducible.\n- Section 2.2: is l_{so} concatenation of l_s and l_o?\n- Section 2.2: y_{spo} is defined but never used.\n- Equation 2: why do we have both e_p and f in the exponents? Aren’t they the same?\n- Equation 2:  P is never defined.\n- Page 5: “a fully connected hidden layer of 64, 200, 4096 and 20 neurons”: this amounts to 3 hidden layers.\n- Why VGG network for visual feature and AlexNet for depth features?\n- zero-shot learning results on visual genome is missing\n- training procedure is a bit unclear: the text suggest that the fine tuning and/or learning of the three components might happen separately. It is important to clearly state if they are done in an end-to-end fashion and simultaneously or separately; and why.\n- It’s good to name the method in table 2 in the same fashion as table 1. With the current naming (based on architecture) it is a bit confusing to understand the content without additional cross referencing. For instance AlexNet-BN - Raw seems to correspond to Ours_{c,v,l,d}\n- Figure 4: the frequency represented as different shades of red or blue is really hard to notice especially on a printed paper. The red vs blue color coding is not necessary since the bars going up or down indicate the same quality. So, it might be better to use red/blue for frequency instead (e.g. dark red high frequency to dark blue low frequency)\n- Section 3.2: the AlexNet reference seems wrong, it should be \"ImageNet Classification with Deep Convolutional Neural Networks\" NIPS , 2012\n- The structure of section 3.5 is currently flat while the content seems to be nested (two experiments and two sets of corresponding discussions). It will read better if they are organized into subsections.\n \n \n********* Points of extensions (improvement) *********\n\n- I believe *unsupervised* discovery of depth information for visual relation detection can be an interesting direction since it is not limited to the availability of relevant depth dataset. \n- It is not clearly motivated why one should use two separate networks for depth and RGB inputs in light of the additional complexity. For instance, it is good to discuss what is the advantage of the proposed (computationally more expensive) method over the following two simpler baselines:\n- Faster RCNN is used on RGBD input to produce a single feature vector\n- above case with RGB input but have the Faster RCNN predict the depth map as an auxiliary loss.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #1", "review": "\n\n********* Post Rebuttal *********\n\nI appreciate the authors' effort in providing thorough responses and revised manuscript. \n\nI agree with the authors that \"the finding not being surprising\" is not a ground for rejection. I tried to word my final decision carefully but it seems it has still caused confusion for the authors. As I have mentioned in my original review, the rating was a result of the 4 points considered *together* . \n\nThat is, if one exploits privileged information that needs extra sensory data and/or annotation (point 2), *and*, this privileged information is clearly related and thus should be normally useful for the final task (point 1), *and* achieve marginal improvements (point 3), it can be a ground for rejection. Especially, given that prior works with similar arguments exists (point 4). \n\nThe rebuttal has alleviated the issue of marginal improvements (point 3) by introducing meanR@K (or as the revised paper refer to it, Macro R@K). Here, the improvements are more significant both compared to the state of the art and ablated baselines.\n\nThe authors also argue that the related [Yang et al. 2018] paper (point 4) should be considered a concurrent submission since the authors original submission was to AAAI18. \n\nThe rebuttal also addresses other clarity or experimental issues which improves the quality of the revised work.\n\nFinally, I understand that the privileged information is only required during training time which is a good point.\n\nAll in all, *assuming that [Yang et al. 2018] is considered a concurrent work* according to ICLR, I think the revised paper becomes slightly above borderline and thus I change my rating to \"weak accept\". If [Yang et al. 2018] is not considered concurrent work, then, a conclusive comparison is required for the acceptance of the current work.\n\n\n********* Summary *********\n \nThe paper poses the question of whether depth information is informative for visual relationship prediction using still images. It is intuitive that 3D arrangement of objects in an image can be a useful cue for predicting their relationship. As such it is important to see whether and to what extent depth information complements RGB information for visual relation detection. That is the focus of this paper.\nThe paper proposes to use an off-the-shelf monocular depth estimation networks to augment the available RGB information towards better visual relation detection. For that, it proposes a specific network two-stream structure working on RGB image and (predicted) depth image. The proposed model demonstrates improved results upon state of the art for visual relation prediction.\n \n\n********* Strengths and Weaknesses *********\n \n+ A comprehensive set of tests has been conducted. \n+ Zero-shot prediction results are particularly interesting.\n+ The experiment on ranking the predicate classes based on the change in prediction accuracy before and after using depth information (Figure 4) is interesting and intuitive.\n* The final results improve upon the state-of-the-art, especially on the zero-shot learning regime. However, it seems that the improvement is mainly coming from the new architecture as opposed to the inclusion of the depth information. That is, ours_{c,v,l} brings most of the improvement already the last step to ours_{c,v,l,d} is negligible for non-zero-shot case.\n- Along the same line, it’s possible that this small difference between ours_{c,v,l} and ours_{c,v,l,d} for the standard predicate prediction, can be due to a hyper-parameter optimization that is (only or more thoroughly) done for ours_{c,v,l,d}. The hyper-parameter optimization scheme for different baselines is not described. \n- Given the small difference of ablation levels, the comparison will be stronger if done multiple times and reporting mean and standard deviation of the results.\n- For a fair comparison the visual feature vector v_{so} should be tried as the feature of the union bound box of both subject and object same way as it is done for depth feature vector d_{so}. \n- The paper refers to “Ours-d’_{so}” as a baseline that *only* uses depth information with no image/label information. However, it seems that the region proposals for this feature are coming from the image-based network that uses image information. \n \n- Important related but uncited works:\n(1) [“Visual Relationship Prediction via Label Clustering and Incorporation of Depth Information” ECCV workshops 2018] studies the same question as part of their work.\n \n\n********* Final Decision *********\n \nI do not find the paper passing the acceptance bar mainly due to the following reasons together:\n1) The finding is not surprising since most of the visual relations are either explicitly depth-related (e.g., behind) or are semantically constrained by depth (e.g. riding cannot happen at different depths when the image is taken orthogonal to the rider).\n2) an additional depth dataset is used which provides the model with privileged information. Should it have been the case that depth information were inferred without an additional offline dataset, the results would have been more interesting.\n3) the improvements due to the additional depth network are not significant or conclusive.\n4) there is a prior uncited work with the same research question for effectiveness of depth information in visual relation detection which uses a similar approach.\n \n\n********* Minor points *********\n\n- the code is not available. This is especially important since the paper is outperforming prior works which could be a contribution if reproducible.\n- Section 2.2: is l_{so} concatenation of l_s and l_o?\n- Section 2.2: y_{spo} is defined but never used.\n- Equation 2: why do we have both e_p and f in the exponents? Aren’t they the same?\n- Equation 2:  P is never defined.\n- Page 5: “a fully connected hidden layer of 64, 200, 4096 and 20 neurons”: this amounts to 3 hidden layers.\n- Why VGG network for visual feature and AlexNet for depth features?\n- zero-shot learning results on visual genome is missing\n- training procedure is a bit unclear: the text suggest that the fine tuning and/or learning of the three components might happen separately. It is important to clearly state if they are done in an end-to-end fashion and simultaneously or separately; and why.\n- It’s good to name the method in table 2 in the same fashion as table 1. With the current naming (based on architecture) it is a bit confusing to understand the content without additional cross referencing. For instance AlexNet-BN - Raw seems to correspond to Ours_{c,v,l,d}\n- Figure 4: the frequency represented as different shades of red or blue is really hard to notice especially on a printed paper. The red vs blue color coding is not necessary since the bars going up or down indicate the same quality. So, it might be better to use red/blue for frequency instead (e.g. dark red high frequency to dark blue low frequency)\n- Section 3.2: the AlexNet reference seems wrong, it should be \"ImageNet Classification with Deep Convolutional Neural Networks\" NIPS , 2012\n- The structure of section 3.5 is currently flat while the content seems to be nested (two experiments and two sets of corresponding discussions). It will read better if they are organized into subsections.\n \n \n********* Points of extensions (improvement) *********\n\n- I believe *unsupervised* discovery of depth information for visual relation detection can be an interesting direction since it is not limited to the availability of relevant depth dataset. \n- It is not clearly motivated why one should use two separate networks for depth and RGB inputs in light of the additional complexity. For instance, it is good to discuss what is the advantage of the proposed (computationally more expensive) method over the following two simpler baselines:\n- Faster RCNN is used on RGBD input to produce a single feature vector\n- above case with RGB input but have the Faster RCNN predict the depth map as an auxiliary loss.\n\n\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1571069780448}], "openreview_url": "https://openreview.net/forum?id=HkxcUxrFPS", "arxiv_id": "1905.00966", "paper_pdf": "papers/HkxcUxrFPS.pdf", "paper_pdf_sha256": "ec4722f6b8da47217fbd3165ee451d63a949df7d0d8b57e85faace8070832e7d", "paper_pdf_bytes": 3907913, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sina-Baharlou/Depth-VRD", "code_repository": "Sina-Baharlou/Depth-VRD", "code_commit": "3b416110d2da6534124526c899f21e4242a0ad24", "code_archive": "repos/HkxcUxrFPS.zip", "code_archive_sha256": "cc1259a924dfac11c424893290f935b1ba62ffedbf6f6c4f0690165219f7c423", "code_archive_bytes": 2337576, "code_file_count": 87, "code_extensions": {".py": 56, ".sh": 17, ".h": 6, ".c": 5, ".cu": 3}, "github_disk_usage_kb": 5131, "github_languages": {"Python": 403672, "Shell": 45767, "Cuda": 30757, "C": 10295, "Cython": 6273, "Makefile": 1058}, "github_archived": false, "github_pushed_at": "2022-07-24T13:13:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-visual-relation-detection-using"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZaB2fV3TYo", "year": 2026, "status": "rejected", "title": "Grounding Bodily Awareness in Visual Representations for Efficient Policy Learning", "authors": ["Junlin Wang", "Zhiyun Lin"], "authorids": ["~Junlin_Wang3", "~Zhiyun_Lin1"], "authors_source": "OpenReview API", "abstract": "Learning effective visual representations for robotic manipulation remains a fundamental challenge due to the complex body dynamics involved in action execution. In this paper, we study how visual representations that carry body-relevant cues can enable efficient policy learning for downstream robotic manipulation tasks. We present $\\textbf{I}$nter-token $\\textbf{Con}$trast ($\\textbf{ICon}$), a contrastive learning method applied to the token-level representations of Vision Transformers (ViTs). ICon enforces a separation in the feature space between agent-specific and environment-specific tokens, resulting in agent-centric visual representations that embed body-specific inductive biases. This framework can be seamlessly integrated into end-to-end policy learning by incorporating the contrastive loss as an auxiliary objective. Our experiments show that ICon not only improves policy performance across various manipulation tasks but also facilitates policy transfer across different robots. The project website: https://anonymous.4open.science/w/ICon/", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "e151V3UKab", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14159/Reviewer_kaVz"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper introduces ICon, a simple contrastive objective that encourages ViTs to disentangle representations of the agent’s body from the environment in visuomotor RL. Using pre-extracted masks of the agent's body, the authors use farthest-point sampling to sample tokens which correspond to the agent's body and to the environment and appliy an InfoNCE loss, as an auxiliary loss for Diffusion Policy. The paper is clearly written and the approach to solve the problem is technically sound, although the usage of hard positive/negatives of agent/environment in the contrastive learning part might introduce some problems. \nThe experiments of the authors on RLBench and Robosuite do show that ICon leads to improvements on downstream tasks and some cross-robot transfer. However, the experimental section could be strengthened to better showcase the strenght of the learned representations.", "review_text": "The paper introduces ICon, a simple contrastive objective that encourages ViTs to disentangle representations of the agent’s body from the environment in visuomotor RL. Using pre-extracted masks of the agent's body, the authors use farthest-point sampling to sample tokens which correspond to the agent's body and to the environment and appliy an InfoNCE loss, as an auxiliary loss for Diffusion Policy. The paper is clearly written and the approach to solve the problem is technically sound, although the usage of hard positive/negatives of agent/environment in the contrastive learning part might introduce some problems. \nThe experiments of the authors on RLBench and Robosuite do show that ICon leads to improvements on downstream tasks and some cross-robot transfer. However, the experimental section could be strengthened to better showcase the strenght of the learned representations.", "strengths": "1) The approach to disentangle agent-specific and environment specific features is technically sound. The FPS also helps obtain more useful representations w.r.t. random sampling. \n2) The improvement in success rates for a few tasks in Robosuite and RLBench seems to be consistent. \n3) With little finetuning the method seems to easily transfer across different robots although it is not clear how much tuning is needed and how well the method would perform zero-shot.", "weaknesses": "* The benchmarks used to validate the method are limited. Training concurrently ViT with the policy also seems to limit the ability to do some pretraining. I believe it would be useful if the objective was also used not only concurrently with the policy, but with a large pretraining setup that can later be used out of the box for variosu downsteram tasks without tuning.\n\n* How much fine-tuning is needed ? Also would it work out of the box ? I believe that at its current state since we train on one robot at each time the results do not generalize directly.\n\n* The method as the authors state is tied to ViTs and needs agent masks. This is however not a major weakness since ViT are currently the standard go-to for visual representations.\n\n* The gains are consistent but also are small. It would be nice also if the authors included more baselines with popular pretrained encoders (such as CLIP). \n\n* The training stability and ablation studies are all done in different environment at each time. This is weird - please use a single unified framework for doing your ablation studies and stability evaluation. \n\n* The overhead of using FPS is not quantified.", "questions": "1) I believe that the current direction of the authors is useful and intuitive and it would make sense to help robotic agents disentangle features from their body and the environment. In my opining the object the authors propose would be also very useful in pretraining of ViTs to use in downstream tasks afterwards. This can allow to train on multiple robotic agents and generalize out of the box to new ones.\n\n2) What is the overhead of using FPS ? How would it compare with dense sampling of tokens ? Also, how would it compare with some simple uniform grid sampling ?\n\n3) Please use a common unified ablation framework.\n\n4) How sensitive is your method in the mask extracted for the agent. Have you done any studies on this ? \n\n5) One major question and a problem with contrastive learning is the usage of hard negatives. At its current state the positive=agent, negative=environment pair is a bit blunt and can worsen some representations since some features from the environment might be similar with the agents (e.g. texture or color). Have the authors consider an alternative to contrastive learning ? Would it be better maybe to force some features and not all features in the feature vector of the token to be distant and not all at once ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ICon, a simple contrastive objective that encourages ViTs to disentangle representations of the agent’s body from the environment in visuomotor RL. Using pre-extracted masks of the agent's body, the authors use farthest-point sampling to sample tokens which correspond to the agent's body and to the environment and appliy an InfoNCE loss, as an auxiliary loss for Diffusion Policy. The paper is clearly written and the approach to solve the problem is technically sound, although the usage of hard positive/negatives of agent/environment in the contrastive learning part might introduce some problems. \nThe experiments of the authors on RLBench and Robosuite do show that ICon leads to improvements on downstream tasks and some cross-robot transfer. However, the experimental section could be strengthened to better showcase the strenght of the learned representations.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1) The approach to disentangle agent-specific and environment specific features is technically sound. The FPS also helps obtain more useful representations w.r.t. random sampling. \n2) The improvement in success rates for a few tasks in Robosuite and RLBench seems to be consistent. \n3) With little finetuning the method seems to easily transfer across different robots although it is not clear how much tuning is needed and how well the method would perform zero-shot.", "weaknesses": "* The benchmarks used to validate the method are limited. Training concurrently ViT with the policy also seems to limit the ability to do some pretraining. I believe it would be useful if the objective was also used not only concurrently with the policy, but with a large pretraining setup that can later be used out of the box for variosu downsteram tasks without tuning.\n\n* How much fine-tuning is needed ? Also would it work out of the box ? I believe that at its current state since we train on one robot at each time the results do not generalize directly.\n\n* The method as the authors state is tied to ViTs and needs agent masks. This is however not a major weakness since ViT are currently the standard go-to for visual representations.\n\n* The gains are consistent but also are small. It would be nice also if the authors included more baselines with popular pretrained encoders (such as CLIP). \n\n* The training stability and ablation studies are all done in different environment at each time. This is weird - please use a single unified framework for doing your ablation studies and stability evaluation. \n\n* The overhead of using FPS is not quantified.", "questions": "1) I believe that the current direction of the authors is useful and intuitive and it would make sense to help robotic agents disentangle features from their body and the environment. In my opining the object the authors propose would be also very useful in pretraining of ViTs to use in downstream tasks afterwards. This can allow to train on multiple robotic agents and generalize out of the box to new ones.\n\n2) What is the overhead of using FPS ? How would it compare with dense sampling of tokens ? Also, how would it compare with some simple uniform grid sampling ?\n\n3) Please use a common unified ablation framework.\n\n4) How sensitive is your method in the mask extracted for the agent. Have you done any studies on this ? \n\n5) One major question and a problem with contrastive learning is the usage of hard negatives. At its current state the positive=agent, negative=environment pair is a bit blunt and can worsen some representations since some features from the environment might be similar with the agents (e.g. texture or color). Have the authors consider an alternative to contrastive learning ? Would it be better maybe to force some features and not all features in the feature vector of the token to be distant and not all at once ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761906020403}, {"id": "9CAdELM0e6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14159/Reviewer_GJDi"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper presents a (class-supervised) contrastive representation learning approach of tokens from a Vision Transformer, termed Inter-Token Contrast (ICon), for the purpose of robotic manipulation. The core idea is to segment-out tokens relevant to the agent (e.g., the robotic gripper) and contrast them with tokens representing everything else. The method is implemented within a diffusion-based policy, and demonstrated on various tasks and robots from two simulated benchmarks.", "review_text": "The paper presents a (class-supervised) contrastive representation learning approach of tokens from a Vision Transformer, termed Inter-Token Contrast (ICon), for the purpose of robotic manipulation. The core idea is to segment-out tokens relevant to the agent (e.g., the robotic gripper) and contrast them with tokens representing everything else. The method is implemented within a diffusion-based policy, and demonstrated on various tasks and robots from two simulated benchmarks.", "strengths": "* The utilization of FPS over 2D feature/token maps to ensure coverage is interesting.\n* I appreciate the stability analysis.\n* Nice video visualizations on the project website.\n* Open-source data and code!", "weaknesses": "* The method relies on a pre-trained supervised segmentation model. While it is true that these models are performing very well, they are still prone to wrong detection or missing objects for unseen data.\n* The masking procedure relies on heuristic (class thresholding) where it is unclear what would happen if part of the gripper and a small object (e.g., a small block) occupy the same tokens.\n* I found the multi-level contrastive loss description in L246-251 confusing: I don’t understand how $\\gamma$ is tuned per-layer and I also don’t understand the reasoning for setting $\\gamma >0$ for deeper and shallower layers. Given the reason provided in the paragraph, I would expect something like “for shallower layers $\\gamma$ is larger/smaller than deeper layers as the earlier layers produce more entangled representations”.\n* The computational cost is unclear. For training, masks are pre-computed (I assume this is because the segmentation model, SAM, is large), FPS can also be time-consuming, and memory-wise, the ViT encoder needs to be probed at several layers, saving the features for the contrastive loss. What about inference time? Are observations masked on the fly?\n* Overall, except for 3 tasks, it seems that the improvement is marginal compared to the baselines, and for some tasks all methods fall short. While I appreciate the authors chose hard benchmarks, given my assumed computational cost, it is hard to get convinced regarding the contribution (and given that masking has already been proven to be useful in decision making, also referenced in the related work section).", "questions": "* ViT: it is unclear from the text if the ViT encoder is pre-trained, fine-tuned or trained from scratch. Can you please clarify this?\n* Regarding the masking procedure concern I raised under Weaknesses, can the authors please clarify what would happen in multi-object environments where objects and gripper share tokens?\n* Can the authors please clarify my concern regarding the computational cost under Weaknesses?\n* Are policies multi-task or trained per-task? Are they goal-conditioned?\n* The authors mention they use action chunking (“Temporal Ensemble”), to my understanding, and that it was critical for the performance. Can you clarify this and what is the chunk size used in practice?\n* I’m not sure I understand this sentence in the conclusion (L469-470) “restricts its applicability to other commonly used visual encoder architectures in visuomotor policy learning, such as ResNet”. Why is that? Basically, it is possible to extract 2D features maps (that represent the regions in the corresponding input image) and apply downsampled masks over them. What prohibits you from applying the contrastive loss on these features? Or do you mean the downstream processing of the post-contrasted features?\n* Have the authors experimented with other types of self-supervised loss, perhaps ones that do not require negative samples (e.g., BYOL, [Grill, Jean-Bastien, et al. \"Bootstrap your own latent-a new approach to self-supervised learning.\" Advances in neural information processing systems 33 (2020): 21271-21284.](https://arxiv.org/abs/2006.07733))\n\nOverall, I raised several concerns regarding the method as detailed above. I’m willing to increase my score given convincing answers and clarifications to my review.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a (class-supervised) contrastive representation learning approach of tokens from a Vision Transformer, termed Inter-Token Contrast (ICon), for the purpose of robotic manipulation. The core idea is to segment-out tokens relevant to the agent (e.g., the robotic gripper) and contrast them with tokens representing everything else. The method is implemented within a diffusion-based policy, and demonstrated on various tasks and robots from two simulated benchmarks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "* The utilization of FPS over 2D feature/token maps to ensure coverage is interesting.\n* I appreciate the stability analysis.\n* Nice video visualizations on the project website.\n* Open-source data and code!", "weaknesses": "* The method relies on a pre-trained supervised segmentation model. While it is true that these models are performing very well, they are still prone to wrong detection or missing objects for unseen data.\n* The masking procedure relies on heuristic (class thresholding) where it is unclear what would happen if part of the gripper and a small object (e.g., a small block) occupy the same tokens.\n* I found the multi-level contrastive loss description in L246-251 confusing: I don’t understand how $\\gamma$ is tuned per-layer and I also don’t understand the reasoning for setting $\\gamma >0$ for deeper and shallower layers. Given the reason provided in the paragraph, I would expect something like “for shallower layers $\\gamma$ is larger/smaller than deeper layers as the earlier layers produce more entangled representations”.\n* The computational cost is unclear. For training, masks are pre-computed (I assume this is because the segmentation model, SAM, is large), FPS can also be time-consuming, and memory-wise, the ViT encoder needs to be probed at several layers, saving the features for the contrastive loss. What about inference time? Are observations masked on the fly?\n* Overall, except for 3 tasks, it seems that the improvement is marginal compared to the baselines, and for some tasks all methods fall short. While I appreciate the authors chose hard benchmarks, given my assumed computational cost, it is hard to get convinced regarding the contribution (and given that masking has already been proven to be useful in decision making, also referenced in the related work section).", "questions": "* ViT: it is unclear from the text if the ViT encoder is pre-trained, fine-tuned or trained from scratch. Can you please clarify this?\n* Regarding the masking procedure concern I raised under Weaknesses, can the authors please clarify what would happen in multi-object environments where objects and gripper share tokens?\n* Can the authors please clarify my concern regarding the computational cost under Weaknesses?\n* Are policies multi-task or trained per-task? Are they goal-conditioned?\n* The authors mention they use action chunking (“Temporal Ensemble”), to my understanding, and that it was critical for the performance. Can you clarify this and what is the chunk size used in practice?\n* I’m not sure I understand this sentence in the conclusion (L469-470) “restricts its applicability to other commonly used visual encoder architectures in visuomotor policy learning, such as ResNet”. Why is that? Basically, it is possible to extract 2D features maps (that represent the regions in the corresponding input image) and apply downsampled masks over them. What prohibits you from applying the contrastive loss on these features? Or do you mean the downstream processing of the post-contrasted features?\n* Have the authors experimented with other types of self-supervised loss, perhaps ones that do not require negative samples (e.g., BYOL, [Grill, Jean-Bastien, et al. \"Bootstrap your own latent-a new approach to self-supervised learning.\" Advances in neural information processing systems 33 (2020): 21271-21284.](https://arxiv.org/abs/2006.07733))\n\nOverall, I raised several concerns regarding the method as detailed above. I’m willing to increase my score given convincing answers and clarifications to my review.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761073582309}, {"id": "FOUkaWEDQ5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission14159/Reviewer_CEhB"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This work proposes an auxiliary objective for end-to-end learning of visual policies which employ a vision-transformer image encoder. The authors take a contrastive learning approach for encouraging agent-environment disentanglement in the learned ViT features. The objective leverages a segmentation mask of the agent (acquired with SAM) to classify ViT patches to either agent or environment, and employs an inter-token contrastive loss based on these classes. The method is evaluated on imitation learning by training visual diffusion policies on simulated robotic manipulation environments.", "review_text": "This work proposes an auxiliary objective for end-to-end learning of visual policies which employ a vision-transformer image encoder. The authors take a contrastive learning approach for encouraging agent-environment disentanglement in the learned ViT features. The objective leverages a segmentation mask of the agent (acquired with SAM) to classify ViT patches to either agent or environment, and employs an inter-token contrastive loss based on these classes. The method is evaluated on imitation learning by training visual diffusion policies on simulated robotic manipulation environments.", "strengths": "**Overview**\n- Well written paper.\n- Method seems novel.\n- Potentially applicable to any policy that employs a ViT image encoder.\n- Evaluated on numerous environments.\n\nI am willing to raise my score if the points raised in the Weaknesses and Questions sections are addressed.", "weaknesses": "**Overview**\n- The specific method is not well-motivated other than the agent-environment disentanglement.\n- No comparison with other segmentation-based representation learning objectives.\n- Information loss in produced token masks compared to the original pixel mask.\n- Empirical performance gains are not significant.\n\n**Method Motivation**\n\nWhy is your specific approach better than others for acquiring agent-environment disentanglement in feature space? This is neither discussed nor empirically evaluated in this work. You cited many papers that apply the same agent-centric principles, could you provide comparisons with one or more of these methods?\n\nAdditionally, I think it would be beneficial to perform the following ablations to motivate your contrastive approach:\n1. Replacing the agent-mask-based contrastive objective with an agent-mask-based reconstruction objective. (Correct me if I am mistaken, but you are currently only comparing with a full-scene reconstruction baseline).\n2. Replacing the contrastive objective with a simple inductive bias: applying a learned additive embedding to the patch features based on their association to environment vs. agent (total of 2 learned embeddings). This suggestion is based on the same principles as your contrastive approach---distinguishing between agent and environment features. If this is indeed beneficial for downstream policy performance, maybe it will be learned end-to-end, i.e., the embeddings will converge to be dissimilar?\n\n**Token Masks**\n\nThe token mask threshold results in a coarser segmentation than the original pixel-based segmentation. Why is this preferred over, e.g., processing the masked image with two parallel ViT layers, once with the agent masked out and once with the environment masked out? The rest of the pipeline can stay the same after this initial layer by combining the tokens from both images such that they also attend to each other in the following layers.\n\n**Experiments**\n\n4/5 RLBench, 2/3 Robosuite and 1/2 transfer tasks are within a standard deviation from the base diffusion policy. It is standard to highlight in bold results that are within a standard deviation from the best performing method for clarity. Maybe evaluating with a larger number of seeds will help distinguish the performance gain. Currently, the performance gains look marginal at best and do not justify the additional method complexity.", "questions": "- Hyperparameters should be detailed in the Appendix. Specifically, what are the values of gamma and lambda? Do they have to be tuned per-task or is the algorithm robust to these hyperparameters?\n- How robust is the method with respect to hyperparameters in terms of the training stability evaluated in Section 4.5?\n- Figure 6: Why does the CLS token specifically attend only to the agent? What do the other tokens attend to?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes an auxiliary objective for end-to-end learning of visual policies which employ a vision-transformer image encoder. The authors take a contrastive learning approach for encouraging agent-environment disentanglement in the learned ViT features. The objective leverages a segmentation mask of the agent (acquired with SAM) to classify ViT patches to either agent or environment, and employs an inter-token contrastive loss based on these classes. The method is evaluated on imitation learning by training visual diffusion policies on simulated robotic manipulation environments.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "**Overview**\n- Well written paper.\n- Method seems novel.\n- Potentially applicable to any policy that employs a ViT image encoder.\n- Evaluated on numerous environments.\n\nI am willing to raise my score if the points raised in the Weaknesses and Questions sections are addressed.", "weaknesses": "**Overview**\n- The specific method is not well-motivated other than the agent-environment disentanglement.\n- No comparison with other segmentation-based representation learning objectives.\n- Information loss in produced token masks compared to the original pixel mask.\n- Empirical performance gains are not significant.\n\n**Method Motivation**\n\nWhy is your specific approach better than others for acquiring agent-environment disentanglement in feature space? This is neither discussed nor empirically evaluated in this work. You cited many papers that apply the same agent-centric principles, could you provide comparisons with one or more of these methods?\n\nAdditionally, I think it would be beneficial to perform the following ablations to motivate your contrastive approach:\n1. Replacing the agent-mask-based contrastive objective with an agent-mask-based reconstruction objective. (Correct me if I am mistaken, but you are currently only comparing with a full-scene reconstruction baseline).\n2. Replacing the contrastive objective with a simple inductive bias: applying a learned additive embedding to the patch features based on their association to environment vs. agent (total of 2 learned embeddings). This suggestion is based on the same principles as your contrastive approach---distinguishing between agent and environment features. If this is indeed beneficial for downstream policy performance, maybe it will be learned end-to-end, i.e., the embeddings will converge to be dissimilar?\n\n**Token Masks**\n\nThe token mask threshold results in a coarser segmentation than the original pixel-based segmentation. Why is this preferred over, e.g., processing the masked image with two parallel ViT layers, once with the agent masked out and once with the environment masked out? The rest of the pipeline can stay the same after this initial layer by combining the tokens from both images such that they also attend to each other in the following layers.\n\n**Experiments**\n\n4/5 RLBench, 2/3 Robosuite and 1/2 transfer tasks are within a standard deviation from the base diffusion policy. It is standard to highlight in bold results that are within a standard deviation from the best performing method for clarity. Maybe evaluating with a larger number of seeds will help distinguish the performance gain. Currently, the performance gains look marginal at best and do not justify the additional method complexity.", "questions": "- Hyperparameters should be detailed in the Appendix. Specifically, what are the values of gamma and lambda? Do they have to be tuned per-task or is the algorithm robust to these hyperparameters?\n- How robust is the method with respect to hyperparameters in terms of the training stability evaluated in Section 4.5?\n- Figure 6: Why does the CLS token specifically attend only to the agent? What do the other tokens attend to?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761029655238}], "openreview_url": "https://openreview.net/forum?id=ZaB2fV3TYo", "arxiv_id": "2505.18487", "paper_pdf": "papers/ZaB2fV3TYo.pdf", "paper_pdf_sha256": "4f5c0efafd59e9f85ab1fcde44f5501fc3efbabd73b85899991a205cfb674c43", "paper_pdf_bytes": 3336127, "paper_pdf_source": "openreview", "code_url": "https://github.com/HenryWJL/icon", "code_repository": "HenryWJL/icon", "code_commit": "b1957a4da27f2cd21467dc6516a8da3d6aa411cc", "code_archive": "repos/ZaB2fV3TYo.zip", "code_archive_sha256": "efef1b1baabecfda738b0630f32f30fd03def04337819dbd0ce29f4fe0d51c37", "code_archive_bytes": 249326, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 3618, "github_languages": {"Python": 139209}, "github_archived": false, "github_pushed_at": "2026-02-14T16:31:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/grounding-bodily-awareness-in-visual"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JaRihIHbZm", "year": 2025, "status": "rejected", "title": "VideoAgent: Self-Improving Video Generation", "authors": ["Achint Soni", "Sreyas Venkataraman", "Abhranil Chandra", "Sebastian Fischmeister", "Percy Liang", "Bo Dai", "Sherry Yang"], "authorids": ["~Achint_Soni1", "~Sreyas_Venkataraman1", "~Abhranil_Chandra1", "~Sebastian_Fischmeister1", "~Percy_Liang1", "~Bo_Dai1", "~Sherry_Yang1"], "authors_source": "OpenReview API", "abstract": "Video generation has been used to generate visual plans for controlling robotic systems. Given an image observation and a language instruction, previous work has generated video plans which are then converted to robot controls to be executed. However, a major bottleneck in leveraging video generation for control lies in the quality of the generated videos, which often suffer from hallucinatory content and unrealistic physics, resulting in low task success when control actions are extracted from the generated videos. While scaling up dataset and model size provides a partial solution, integrating external feedback is both natural and essential for grounding video generation in the real world. With this observation, we propose VideoAgent for self-improving generated video plans based on external feedback. Instead of directly executing the generated video plan, VideoAgent first refines the generated video plans using a novel procedure which we call self-conditioning consistency, utilizing feedback from a pretrained vision-language model (VLM). As the refined video plan is being executed, VideoAgent collects additional data from the environment to further improve video plan generation. Experiments in simulated robotic manipulation from MetaWorld and iTHOR show that VideoAgent drastically reduces hallucination, thereby boosting success rate of downstream manipulation tasks. We further illustrate that VideoAgent can effectively refine real-robot videos, providing an early indicator that robotics can be an effective tool in grounding video generation in the physical world.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "XubCPvZbr1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12522/Reviewer_ZcuG"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces VideoAgent, an approach that refines video plans generated by video diffusion models through self-conditioning consistency and feedback from pre-trained vision-language models. When online interaction is available, VideoAgent closes the self-improvement loop by alternating between collecting successful online data and finetuning video models. Evaluations are performed on three benchmarks, across both simulated and real-world domains, and from the perspective of both task success rate and video generation quality, showing the effectiveness of the proposed method.", "review_text": "This paper introduces VideoAgent, an approach that refines video plans generated by video diffusion models through self-conditioning consistency and feedback from pre-trained vision-language models. When online interaction is available, VideoAgent closes the self-improvement loop by alternating between collecting successful online data and finetuning video models. Evaluations are performed on three benchmarks, across both simulated and real-world domains, and from the perspective of both task success rate and video generation quality, showing the effectiveness of the proposed method.", "strengths": "1. This paper proposes several novel techniques, such as self-conditioning consistency and incorporating VLM feedback, that improve the quality of the generated video plans through iterative refinement\n2. The proposed method adopts a self-improving loop by finetuning the video models with additional successful trajectories collected online\n3. The authors provide extensive evaluation results as well as experiment details, which back up the efficacy of the proposed method\n4. This paper is well-motivated and easy to follow, it thoroughly discusses the limitations of the proposed work", "weaknesses": "1. The loop of collecting successful data through environment interaction and finetuning video models might incur large computational overhead, and the improvement seems to become marginal after two online iterations in Figure 4.\n2. [nitpick] I believe it is essential to reduce hallucinations and improve the quality of video plans for better decision-making performance and I appreciate the overall contribution. However, in the Metaworld example (Table 1) provided in this work, simply replanning during inference still seems to be a comparatively cost-effective approach that can achieve higher overall success rates with a less performant video plan generator. I understand refinement and replanning can be combined for better improvement, but I wonder if the authors have stronger examples to highlight the necessity of refinement and self-improvement for task success.", "questions": "1. There seems to be a discrepancy between the self-conditioning-consistency loss In Algorithm 1 and Equation 7, is the second term in Eq. 7 missing in Algorithm 1?\n2. During evaluation, is every video plan executed in an open-loop manner? What is the planning horizon for each setup, and how does it compare to the task horizon?\n3. In Figure 4, how many additional trajectories are collected per iteration to achieve the improvement?\n4. In the last row of Table 5, how is the \"task success\" defined specifically in the Bridge human evaluation? Is it defined as whether the generated videos respect the prompt from human perception? or as the authors mentioned in line 467, it depends on whether a generated video looks realistic.\n5. As the video models are language-conditioned, it would be clearer to include the conditioning variable in the equations or algorithms somewhere.\n6. Typo in line 342: brining", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces VideoAgent, an approach that refines video plans generated by video diffusion models through self-conditioning consistency and feedback from pre-trained vision-language models. When online interaction is available, VideoAgent closes the self-improvement loop by alternating between collecting successful online data and finetuning video models. Evaluations are performed on three benchmarks, across both simulated and real-world domains, and from the perspective of both task success rate and video generation quality, showing the effectiveness of the proposed method.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper proposes several novel techniques, such as self-conditioning consistency and incorporating VLM feedback, that improve the quality of the generated video plans through iterative refinement\n2. The proposed method adopts a self-improving loop by finetuning the video models with additional successful trajectories collected online\n3. The authors provide extensive evaluation results as well as experiment details, which back up the efficacy of the proposed method\n4. This paper is well-motivated and easy to follow, it thoroughly discusses the limitations of the proposed work", "weaknesses": "1. The loop of collecting successful data through environment interaction and finetuning video models might incur large computational overhead, and the improvement seems to become marginal after two online iterations in Figure 4.\n2. [nitpick] I believe it is essential to reduce hallucinations and improve the quality of video plans for better decision-making performance and I appreciate the overall contribution. However, in the Metaworld example (Table 1) provided in this work, simply replanning during inference still seems to be a comparatively cost-effective approach that can achieve higher overall success rates with a less performant video plan generator. I understand refinement and replanning can be combined for better improvement, but I wonder if the authors have stronger examples to highlight the necessity of refinement and self-improvement for task success.", "questions": "1. There seems to be a discrepancy between the self-conditioning-consistency loss In Algorithm 1 and Equation 7, is the second term in Eq. 7 missing in Algorithm 1?\n2. During evaluation, is every video plan executed in an open-loop manner? What is the planning horizon for each setup, and how does it compare to the task horizon?\n3. In Figure 4, how many additional trajectories are collected per iteration to achieve the improvement?\n4. In the last row of Table 5, how is the \"task success\" defined specifically in the Bridge human evaluation? Is it defined as whether the generated videos respect the prompt from human perception? or as the authors mentioned in line 467, it depends on whether a generated video looks realistic.\n5. As the video models are language-conditioned, it would be clearer to include the conditioning variable in the equations or algorithms somewhere.\n6. Typo in line 342: brining", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730673202440}, {"id": "nQQWdFX1SD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12522/Reviewer_b49X"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes VideoAgent, a model designed to improve video generation quality based on feedback from a VLM. Specifically, VideoAgent trains both a video generation model and a video refinement model, which share parameters and are based on diffusion processes. The video generation model functions as a conventional text-to-video model, while the video refinement model takes as input the ground-truth video, the predicted video, and feedback from the VLM to produce a refined video. The final output video can be mapped to an action trajectory. When the environment is interactive and can confirm task success, VideoAgent performs rollouts within the environment to gather additional data for training (but very slow now). This method is evaluated at the policy level in the Meta-World and iTHOR benchmarks, and for video quality on the Bridge dataset.", "review_text": "This paper proposes VideoAgent, a model designed to improve video generation quality based on feedback from a VLM. Specifically, VideoAgent trains both a video generation model and a video refinement model, which share parameters and are based on diffusion processes. The video generation model functions as a conventional text-to-video model, while the video refinement model takes as input the ground-truth video, the predicted video, and feedback from the VLM to produce a refined video. The final output video can be mapped to an action trajectory. When the environment is interactive and can confirm task success, VideoAgent performs rollouts within the environment to gather additional data for training (but very slow now). This method is evaluated at the policy level in the Meta-World and iTHOR benchmarks, and for video quality on the Bridge dataset.", "strengths": "1. The idea of VideoAgent is novel; I haven’t seen an iterative approach to improving video generation quality before.\n2. Obtaining feedback on videos from VLM is feasible.", "weaknesses": "1. The writing section can be strengthened. The distinction between the consistency model and DDPM is vague in this paper, yet it is crucial. Given that I believe DDPM + DDIM can achieve the same results (referring to the implementation of video generation models and video refinement models), it’s even more necessary to explain the necessity and motivation for using the consistency model. Specific issues can be referenced in the questions section.\n\n2. The experimental section of the paper is relatively weak. The baseline is only AVDC, a typical video generation model without a corresponding video refinement model. Possible baselines could include VLP[1] and a naive text-to-video diffusion model (under similar training and inference FLOP). Additionally, results from traditional baselines based on BC and RL should also be included for reference.\n\n3. More realistic robotic manipulation benchmarks should be considered, such as Simper[2].\n\n4. It would be better to conduct experiments on a real robot; currently, the focus is only on testing quality using real-robot videos. Success rate evaluations are missing, which may necessitate the use of a real robot. If a real robot is not available, providing results from Simper would be acceptable.\n\n5. Even after refinement, the video generation quality of VideoAgent on the Bridge dataset is not high compared to modern classic diffusion models, and the resolution is very low.\n\n6. Fundamentally, generating a video and then generating actions is not widely regarded as a correct approach; the improvements made in this paper are still limited by the fundamental flaws of video planning. The video generation is too slow to be suitable as a policy. If the paper could demonstrate the importance of video generation for policy, it would greatly strengthen its contributions.\n\n7. The paper seems to avoid the discussion of video generation speed. In reality, how many actions can be generated per second with this method? Or what is the average time required to generate a single action? The resolution should also be reported.\n\n[1] Video Language Planning.\n[2] Evaluating Real-World Robot Manipulation Policies in Simulation", "questions": "When I read this paper, I encountered some areas of confusion, and I hope my feedback can help the authors improve the clarity of the writing. If there are any inaccuracies in my understanding, I welcome any corrections.\n\n1. The motivation for introducing the consistency model should be clearly articulated. I didn’t see a compelling reason to use the consistency model, as it seems possible to achieve similar results by using the training objective of DDPM and employing DDIM for inference.\n\n2. Lines 119-120 should describe the loss for the consistency model rather than for DDPM. In DDPM, the loss should be calculated from ($x^{t+1}$ to $x^t$), which also makes the \"vanilla objective for video diffusion\" mentioned in line 197 confusing.\n\n3. In line 224, it seems that video generation should be a single step if using a consistency model. Why are multiple steps required, and if they are, what distinguishes this approach from DDIM?\n\n4. The motivation in Section 3.1 could be more clearly defined. The overall objective of the proposed method is to (1) generate video and (2) refine the video based on feedback until it is accepted by the VLM, using the same parameters. However, this goal is not clearly explained at the beginning of Section 3.1, which requires readers to infer the motivation after understanding the method.\n\n5. Line 164 states, \"the model can learn to preserve the realistic part of the video while refining the hallucinatory part,\" which could be misleading. Feedback is applied to the entire video, not frame-level feedback. This statement may somewhat overstate the model’s capability.\n\n6. It is unclear how the parameters are shared between the video generation model and the video refinement model.\n\n7. It is not shown in the method how the video maps to action. Indicating where this is discussed in the paper would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes VideoAgent, a model designed to improve video generation quality based on feedback from a VLM. Specifically, VideoAgent trains both a video generation model and a video refinement model, which share parameters and are based on diffusion processes. The video generation model functions as a conventional text-to-video model, while the video refinement model takes as input the ground-truth video, the predicted video, and feedback from the VLM to produce a refined video. The final output video can be mapped to an action trajectory. When the environment is interactive and can confirm task success, VideoAgent performs rollouts within the environment to gather additional data for training (but very slow now). This method is evaluated at the policy level in the Meta-World and iTHOR benchmarks, and for video quality on the Bridge dataset.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The idea of VideoAgent is novel; I haven’t seen an iterative approach to improving video generation quality before.\n2. Obtaining feedback on videos from VLM is feasible.", "weaknesses": "1. The writing section can be strengthened. The distinction between the consistency model and DDPM is vague in this paper, yet it is crucial. Given that I believe DDPM + DDIM can achieve the same results (referring to the implementation of video generation models and video refinement models), it’s even more necessary to explain the necessity and motivation for using the consistency model. Specific issues can be referenced in the questions section.\n\n2. The experimental section of the paper is relatively weak. The baseline is only AVDC, a typical video generation model without a corresponding video refinement model. Possible baselines could include VLP[1] and a naive text-to-video diffusion model (under similar training and inference FLOP). Additionally, results from traditional baselines based on BC and RL should also be included for reference.\n\n3. More realistic robotic manipulation benchmarks should be considered, such as Simper[2].\n\n4. It would be better to conduct experiments on a real robot; currently, the focus is only on testing quality using real-robot videos. Success rate evaluations are missing, which may necessitate the use of a real robot. If a real robot is not available, providing results from Simper would be acceptable.\n\n5. Even after refinement, the video generation quality of VideoAgent on the Bridge dataset is not high compared to modern classic diffusion models, and the resolution is very low.\n\n6. Fundamentally, generating a video and then generating actions is not widely regarded as a correct approach; the improvements made in this paper are still limited by the fundamental flaws of video planning. The video generation is too slow to be suitable as a policy. If the paper could demonstrate the importance of video generation for policy, it would greatly strengthen its contributions.\n\n7. The paper seems to avoid the discussion of video generation speed. In reality, how many actions can be generated per second with this method? Or what is the average time required to generate a single action? The resolution should also be reported.\n\n[1] Video Language Planning.\n[2] Evaluating Real-World Robot Manipulation Policies in Simulation", "questions": "When I read this paper, I encountered some areas of confusion, and I hope my feedback can help the authors improve the clarity of the writing. If there are any inaccuracies in my understanding, I welcome any corrections.\n\n1. The motivation for introducing the consistency model should be clearly articulated. I didn’t see a compelling reason to use the consistency model, as it seems possible to achieve similar results by using the training objective of DDPM and employing DDIM for inference.\n\n2. Lines 119-120 should describe the loss for the consistency model rather than for DDPM. In DDPM, the loss should be calculated from ($x^{t+1}$ to $x^t$), which also makes the \"vanilla objective for video diffusion\" mentioned in line 197 confusing.\n\n3. In line 224, it seems that video generation should be a single step if using a consistency model. Why are multiple steps required, and if they are, what distinguishes this approach from DDIM?\n\n4. The motivation in Section 3.1 could be more clearly defined. The overall objective of the proposed method is to (1) generate video and (2) refine the video based on feedback until it is accepted by the VLM, using the same parameters. However, this goal is not clearly explained at the beginning of Section 3.1, which requires readers to infer the motivation after understanding the method.\n\n5. Line 164 states, \"the model can learn to preserve the realistic part of the video while refining the hallucinatory part,\" which could be misleading. Feedback is applied to the entire video, not frame-level feedback. This statement may somewhat overstate the model’s capability.\n\n6. It is unclear how the parameters are shared between the video generation model and the video refinement model.\n\n7. It is not shown in the method how the video maps to action. Indicating where this is discussed in the paper would be helpful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730658318717}, {"id": "70bdVImN1E", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12522/Reviewer_jLDg"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces VideoAgent which self-improves generated video plans based on external feedback.\nGiven an initial frame and a language instruction, VideoAgent first generates a video plan and then iteratively refines it based on feedback from a pre-trained VLM.\nThe refined video plan can be converted to low-level robot actions for execution.\nSuccessful trajectories are added to the dataset for further fine-tuning the video generation and refinement models.\nExperiments on video plan generation were performed in both simulation and real-robot videos.\nResults show that the proposed method can reduce hallucination in video plan generation and thus boost performance in downstream manipulation tasks.", "review_text": "The paper introduces VideoAgent which self-improves generated video plans based on external feedback.\nGiven an initial frame and a language instruction, VideoAgent first generates a video plan and then iteratively refines it based on feedback from a pre-trained VLM.\nThe refined video plan can be converted to low-level robot actions for execution.\nSuccessful trajectories are added to the dataset for further fine-tuning the video generation and refinement models.\nExperiments on video plan generation were performed in both simulation and real-robot videos.\nResults show that the proposed method can reduce hallucination in video plan generation and thus boost performance in downstream manipulation tasks.", "strengths": "The paper presents a novel method which uses feedback from VLMs to refine generated video plans.\nIt introduces a self-improving consistency model which predicts the clean video from a generated video and feedback from the VLM.\nThe proposed method can be continuously improved through online fine-tuning.\nExperiment results indicate that the proposed method effectively enhances video generation and improves performance on downstream manipulation tasks.\nThe paper also includes ablation studies to examine the impact of different types of feedback from the VLM, and the effect of refinement and online iterations.\nAdditionally, it offers an in-depth analysis of the VLM's performance in providing feedback for video generation.", "weaknesses": "1. The paper leverages a self-conditioning consistency loss (Eqn. 7) for video refinement. The first term in Eqn. 7 is a diffusion loss and the second term is for consistency. The reason why including the second term in Eqn. 7 is not very clear. It seems that it encourages generating consistent $x^{(0)}$ from different $\\hat x_{i}$. Is it possible to provide more explanation on why the consistency loss is necessary to be included? It would be great to include an ablation study which compares the refinement performance with and without this consistency loss.\n\n2. Figure 2 is not very clear. If I understood correctly, $x^{(t + \\Delta)}, x^{(t)}, x^{(t - \\Delta)}$ are noisy latents at different denoising timesteps in a diffusion. But what is the target of this diffusion process? Is it $\\hat x_{(i+1)}$? It would be great to provide more descriptions in the caption for further clarification.\n\n3. The paper lacks real-robot experiments to validate the effectiveness of the proposed method in real-world policy learning. The paper would benefit strongly from incorporating a real-robot policy learning experiment. And comparing the proposed method with baseline methods without self-improving consistency (e.g. AVDC) would help understand how the proposed video refinement process helps downstream policy learning in the real world.", "questions": "1. VLMs are prone to hallucinations in some cases. Were any hallucinations from VLM feedback observed in the experiments? If such hallucinations were observed, would they potentially mislead the refinement process?\n\n2. How does the proposed method perform in generalization settings? Does introducing feedback from VLM help the proposed method on handling novel text instructions? It would be great to incorporate a generalization experiment on video plan generation on real-robot data (e.g. BridgeData V2).\n\n3. In Algorithm 2, if I understood correctly, the $\\pi_{\\theta}$ in line 284 should be $\\hat f_{\\theta}$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces VideoAgent which self-improves generated video plans based on external feedback.\nGiven an initial frame and a language instruction, VideoAgent first generates a video plan and then iteratively refines it based on feedback from a pre-trained VLM.\nThe refined video plan can be converted to low-level robot actions for execution.\nSuccessful trajectories are added to the dataset for further fine-tuning the video generation and refinement models.\nExperiments on video plan generation were performed in both simulation and real-robot videos.\nResults show that the proposed method can reduce hallucination in video plan generation and thus boost performance in downstream manipulation tasks.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The paper presents a novel method which uses feedback from VLMs to refine generated video plans.\nIt introduces a self-improving consistency model which predicts the clean video from a generated video and feedback from the VLM.\nThe proposed method can be continuously improved through online fine-tuning.\nExperiment results indicate that the proposed method effectively enhances video generation and improves performance on downstream manipulation tasks.\nThe paper also includes ablation studies to examine the impact of different types of feedback from the VLM, and the effect of refinement and online iterations.\nAdditionally, it offers an in-depth analysis of the VLM's performance in providing feedback for video generation.", "weaknesses": "1. The paper leverages a self-conditioning consistency loss (Eqn. 7) for video refinement. The first term in Eqn. 7 is a diffusion loss and the second term is for consistency. The reason why including the second term in Eqn. 7 is not very clear. It seems that it encourages generating consistent $x^{(0)}$ from different $\\hat x_{i}$. Is it possible to provide more explanation on why the consistency loss is necessary to be included? It would be great to include an ablation study which compares the refinement performance with and without this consistency loss.\n\n2. Figure 2 is not very clear. If I understood correctly, $x^{(t + \\Delta)}, x^{(t)}, x^{(t - \\Delta)}$ are noisy latents at different denoising timesteps in a diffusion. But what is the target of this diffusion process? Is it $\\hat x_{(i+1)}$? It would be great to provide more descriptions in the caption for further clarification.\n\n3. The paper lacks real-robot experiments to validate the effectiveness of the proposed method in real-world policy learning. The paper would benefit strongly from incorporating a real-robot policy learning experiment. And comparing the proposed method with baseline methods without self-improving consistency (e.g. AVDC) would help understand how the proposed video refinement process helps downstream policy learning in the real world.", "questions": "1. VLMs are prone to hallucinations in some cases. Were any hallucinations from VLM feedback observed in the experiments? If such hallucinations were observed, would they potentially mislead the refinement process?\n\n2. How does the proposed method perform in generalization settings? Does introducing feedback from VLM help the proposed method on handling novel text instructions? It would be great to incorporate a generalization experiment on video plan generation on real-robot data (e.g. BridgeData V2).\n\n3. In Algorithm 2, if I understood correctly, the $\\pi_{\\theta}$ in line 284 should be $\\hat f_{\\theta}$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730621857681}, {"id": "8I1MsjPayy", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12522/Reviewer_AFbw"], "rating": 3, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper proposes VideoAgent for self-improving generated video plans based on external feedback. It first refines the generated video plans using self-conditioning consistency and utilizes feedback from a pretrained vision-language model (VLM), then it collects additional data from the environment to further improve video plan generation. From both simulation and real-world video generation experiments, it shows the effectiveness of the proposed method.", "review_text": "This paper proposes VideoAgent for self-improving generated video plans based on external feedback. It first refines the generated video plans using self-conditioning consistency and utilizes feedback from a pretrained vision-language model (VLM), then it collects additional data from the environment to further improve video plan generation. From both simulation and real-world video generation experiments, it shows the effectiveness of the proposed method.", "strengths": "1. Clear equation writing, which shows the strong technique capabilities of the authors.\n\n2. Figure 1 is clear, which makes me understand the main idea of this paper quickly.\n\n3. Experiments for the VLM feedback is interesting.", "weaknesses": "1. I am unsure whether the *self-conditioning consistency* is an original contribution of this paper because of the poor writing clarity of Section 3.1. From the author's writing, it seems like some other papers also endow similar ideas. If the authors want to claim this as part of their contribution, the author should improve the writing of this part to clearly show which idea or equation is original.\n\n2. I am confused about how to use the feedback from VLM for Equation 10. The authors did not provide any methodological description in Section 3.1 or Appendix, e.g., what is the *feedback*, or how to put the *feedback* into Equation 9.\n\n3. Also, I am unsure whether using VLM feedback to improve video generation models is original since the author didn't discuss any related works for this part. If this is novel (or at least part of this is novel), the author should clearly point out which part is novel. Currently, people may assume that this section is totally based on some existing methods, because the author seems to assume that readers know how Equation 10 works, and therefore does not provide any method explanation.\n\n4. From my point of view, the main contribution (if the above 3 questions can be solved) are: 1) self-consistency conditioning; 2) VLM-feedback; 3) real-world interaction feedback. However, from Table 1, we can see these three modules didn't bring much improvement (VideoAgent v.s. AVDC, only 2.7% increase). Instead, the main improvement comes from Online and Replan (23.5% increase for AVDC and 27.7% for VideoAgent). What's more, Online-and-Replan works across two different video generation methods (AVDC and VideoAgent). Thus, this greatly reduces the value of the three main contributions of this article.", "questions": "See the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes VideoAgent for self-improving generated video plans based on external feedback. It first refines the generated video plans using self-conditioning consistency and utilizes feedback from a pretrained vision-language model (VLM), then it collects additional data from the environment to further improve video plan generation. From both simulation and real-world video generation experiments, it shows the effectiveness of the proposed method.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. Clear equation writing, which shows the strong technique capabilities of the authors.\n\n2. Figure 1 is clear, which makes me understand the main idea of this paper quickly.\n\n3. Experiments for the VLM feedback is interesting.", "weaknesses": "1. I am unsure whether the *self-conditioning consistency* is an original contribution of this paper because of the poor writing clarity of Section 3.1. From the author's writing, it seems like some other papers also endow similar ideas. If the authors want to claim this as part of their contribution, the author should improve the writing of this part to clearly show which idea or equation is original.\n\n2. I am confused about how to use the feedback from VLM for Equation 10. The authors did not provide any methodological description in Section 3.1 or Appendix, e.g., what is the *feedback*, or how to put the *feedback* into Equation 9.\n\n3. Also, I am unsure whether using VLM feedback to improve video generation models is original since the author didn't discuss any related works for this part. If this is novel (or at least part of this is novel), the author should clearly point out which part is novel. Currently, people may assume that this section is totally based on some existing methods, because the author seems to assume that readers know how Equation 10 works, and therefore does not provide any method explanation.\n\n4. From my point of view, the main contribution (if the above 3 questions can be solved) are: 1) self-consistency conditioning; 2) VLM-feedback; 3) real-world interaction feedback. However, from Table 1, we can see these three modules didn't bring much improvement (VideoAgent v.s. AVDC, only 2.7% increase). Instead, the main improvement comes from Online and Replan (23.5% increase for AVDC and 27.7% for VideoAgent). What's more, Online-and-Replan works across two different video generation methods (AVDC and VideoAgent). Thus, this greatly reduces the value of the three main contributions of this article.", "questions": "See the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729926182628}], "openreview_url": "https://openreview.net/forum?id=JaRihIHbZm", "arxiv_id": "2410.10076", "paper_pdf": "papers/JaRihIHbZm.pdf", "paper_pdf_sha256": "535cd2df0e8c785274cafb2060b2aebd51a8fdaf2eecac2b9571960d73dace63", "paper_pdf_bytes": 32572005, "paper_pdf_source": "openreview", "code_url": "https://github.com/Video-as-Agent/VideoAgent", "code_repository": "Video-as-Agent/VideoAgent", "code_commit": "108548efee6a598da6f99a81244ea97fe8d7dd9e", "code_archive": "repos/JaRihIHbZm.zip", "code_archive_sha256": "4fdef1796cd77262f974555104eca5cdac9564b5a5460b302af1f226bc6f09ea", "code_archive_bytes": 263563, "code_file_count": 32, "code_extensions": {".py": 31, ".sh": 1}, "github_disk_usage_kb": 254, "github_languages": {"Python": 490130, "Shell": 782}, "github_archived": false, "github_pushed_at": "2025-04-25T19:28:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/videoagent-self-improving-video-generation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LhNZqkuVte", "year": 2024, "status": "rejected", "title": "HyperMask: Adaptive Hypernetwork-based Masks for Continual Learning", "authors": ["Kamil Książek", "Przemysław Spurek"], "authorids": ["~Kamil_Książek1", "~Przemysław_Spurek1"], "authors_source": "OpenReview API", "abstract": "Artificial neural networks suffer catastrophic forgetting when they are sequentially trained on multiple tasks. To overcome this problem, there exist many continual learning strategies. One of the most effective is the hypernetwork-based approach.  The hypernetwork generates the weights of a target model based on the task's identity. The model's main limitation is that hypernetwork can produce completely different nests for each task. Consequently, each task is solved separately. \nThe model does not use information from the network dedicated to previous tasks. \nWe practically produce a new architecture when we update the prior task.\nTo solve such a problem, we use the lottery ticket hypothesis, which postulates the existence of sparse subnetworks, named winning tickets, that preserve the performance of a full network.\n\nIn the paper, we propose a method called HyperMask, which trains a single network for all tasks. Hypernetwork produces semi-binary masks to obtain target subnetworks dedicated to new tasks. This solution inherits the ability of the hypernetwork to adapt to new tasks with minimal forgetting. Moreover, due to the lottery ticket hypothesis, we can use a single network with weighted subnets dedicated to each task.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "saiKP8ix9W", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1307/Reviewer_G6Eq"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper works on continual learning with hypernetworks. The main idea is to generate masks to obtain a subnetwork for the new task. The experiments are conducted on Permuted MNIST, Split MNIST and Split CIFAR-100.", "review_text": "The paper works on continual learning with hypernetworks. The main idea is to generate masks to obtain a subnetwork for the new task. The experiments are conducted on Permuted MNIST, Split MNIST and Split CIFAR-100.", "strengths": "(1) The hypernetworks are used for continual learning in a different way of generating hypermasks for each task.\n(2) It connects to lottery ticket theory with a single network for continual learning.\n(3) The method is easy to follow in general.", "weaknesses": "(1) There are some works mentioned in the related work using masks as an extension of the whole network, it is unclear what benefits hypernetwork can bring.\n(2) There are several common loss functions are used in the method, and it is unclear if the improvements are from the proposed hypernetworks or the additional regularizations. There is no ablation study on these components.\n(3) The experimental evaluation is very limited. It only compares with other methods on very tiny benchmarks. The performance gain from Table 1 seems not very significant. And it is hard to know how much more computation and complexity the method needs.", "questions": "The superiority of the method is not clear compared to other existing methods and the evaluation section is weak.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper works on continual learning with hypernetworks. The main idea is to generate masks to obtain a subnetwork for the new task. The experiments are conducted on Permuted MNIST, Split MNIST and Split CIFAR-100.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "(1) The hypernetworks are used for continual learning in a different way of generating hypermasks for each task.\n(2) It connects to lottery ticket theory with a single network for continual learning.\n(3) The method is easy to follow in general.", "weaknesses": "(1) There are some works mentioned in the related work using masks as an extension of the whole network, it is unclear what benefits hypernetwork can bring.\n(2) There are several common loss functions are used in the method, and it is unclear if the improvements are from the proposed hypernetworks or the additional regularizations. There is no ablation study on these components.\n(3) The experimental evaluation is very limited. It only compares with other methods on very tiny benchmarks. The performance gain from Table 1 seems not very significant. And it is hard to know how much more computation and complexity the method needs.", "questions": "The superiority of the method is not clear compared to other existing methods and the evaluation section is weak.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698804093831}, {"id": "le22djMTkA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1307/Reviewer_uLSz"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a novel method, HyperMask, which leverages hypernetwork paradigms to model lottery ticket-based subnetworks.\nHyperMask retains the ability to reuse weights from the lottery ticket module and adapt to new tasks, inheriting the strengths of both approaches. The semi-binary masks generated by HyperMask enhance the target network's ability to discriminate between classes in consecutive continual learning tasks. The paper offers a promising solution to the issue of catastrophic forgetting by combining two existing paradigms in a novel way, demonstrating potential significance in the field of continual learning.", "review_text": "This paper introduces a novel method, HyperMask, which leverages hypernetwork paradigms to model lottery ticket-based subnetworks.\nHyperMask retains the ability to reuse weights from the lottery ticket module and adapt to new tasks, inheriting the strengths of both approaches. The semi-binary masks generated by HyperMask enhance the target network's ability to discriminate between classes in consecutive continual learning tasks. The paper offers a promising solution to the issue of catastrophic forgetting by combining two existing paradigms in a novel way, demonstrating potential significance in the field of continual learning.", "strengths": "1. The paper introduces an innovative approach, HyperMask, that combines the concepts of hypernetworks and the lottery ticket hypothesis. This unique combination leads to a novel method for addressing catastrophic forgetting in continual learning. \n2. The idea of using semi-binary masks generated by a hypernetwork to create target subnetworks is a fresh and creative approach to tackling the challenges of continual learning. \n3. The paper demonstrates a high level of quality in the experimental evaluation of HyperMask. It provides a comprehensive set of experiments on multiple benchmark datasets, comparing HyperMask against various state-of-the-art baseline methods.", "weaknesses": "1.The primary contribution of HyperMask, which involves using hypernetworks to produce semi-binary masks for continual learning, may not be considered highly novel in the field of continual learning and neural network architectures. Hypernetworks have been explored in prior research as a means to generate task-specific weights for neural networks [1][2], and the concept of using masks or pruning for model adaptation is not entirely new.\n2. The paper lacks a deeper theoretical analysis of the proposed HyperMask method. It would be beneficial to include a more comprehensive theoretical foundation for the approach, explaining why semi-binary masks generated by hypernetworks are effective in minimizing forgetting. \n3. The paper could benefit from providing more detailed guidelines for selecting hyperparameters. HyperMask involves parameters such as β, λ, and p, and while the authors mention the hyperparameter optimization process, they do not offer specific recommendations or insights on how to choose these hyperparameters effectively. \n4. The paper could expand the comparison to include other hypernetwork variants or architectures that have been proposed in the literature. Specifically, discussing how HyperMask compares to variations of hypernetwork-based approaches.\n5. The paper does not provide detailed information on the computational resources required for training HyperMask, including information on training time, memory usage.\n\n[1] PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning (CVPR2018)\n[2] Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights (ECCV2018)", "questions": "1. Can the authors provide a more detailed discussion of the novelty of their proposed method compared to prior work in the field of continual learning and hypernetwork-based approaches? \n2. The paper mentions that HyperMask has some limitations in terms of memory consumption due to the requirement for the hypernetwork's output layer to match the number of parameters in the target network. Are there any potential strategies or approaches to mitigate this memory consumption issue?\n3. The paper mentions that HyperMask has some limitations in terms of memory consumption due to the requirement for the hypernetwork's output layer to match the number of parameters in the target network. Are there any potential strategies or approaches to mitigate this memory consumption issue?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel method, HyperMask, which leverages hypernetwork paradigms to model lottery ticket-based subnetworks.\nHyperMask retains the ability to reuse weights from the lottery ticket module and adapt to new tasks, inheriting the strengths of both approaches. The semi-binary masks generated by HyperMask enhance the target network's ability to discriminate between classes in consecutive continual learning tasks. The paper offers a promising solution to the issue of catastrophic forgetting by combining two existing paradigms in a novel way, demonstrating potential significance in the field of continual learning.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1. The paper introduces an innovative approach, HyperMask, that combines the concepts of hypernetworks and the lottery ticket hypothesis. This unique combination leads to a novel method for addressing catastrophic forgetting in continual learning. \n2. The idea of using semi-binary masks generated by a hypernetwork to create target subnetworks is a fresh and creative approach to tackling the challenges of continual learning. \n3. The paper demonstrates a high level of quality in the experimental evaluation of HyperMask. It provides a comprehensive set of experiments on multiple benchmark datasets, comparing HyperMask against various state-of-the-art baseline methods.", "weaknesses": "1.The primary contribution of HyperMask, which involves using hypernetworks to produce semi-binary masks for continual learning, may not be considered highly novel in the field of continual learning and neural network architectures. Hypernetworks have been explored in prior research as a means to generate task-specific weights for neural networks [1][2], and the concept of using masks or pruning for model adaptation is not entirely new.\n2. The paper lacks a deeper theoretical analysis of the proposed HyperMask method. It would be beneficial to include a more comprehensive theoretical foundation for the approach, explaining why semi-binary masks generated by hypernetworks are effective in minimizing forgetting. \n3. The paper could benefit from providing more detailed guidelines for selecting hyperparameters. HyperMask involves parameters such as β, λ, and p, and while the authors mention the hyperparameter optimization process, they do not offer specific recommendations or insights on how to choose these hyperparameters effectively. \n4. The paper could expand the comparison to include other hypernetwork variants or architectures that have been proposed in the literature. Specifically, discussing how HyperMask compares to variations of hypernetwork-based approaches.\n5. The paper does not provide detailed information on the computational resources required for training HyperMask, including information on training time, memory usage.\n\n[1] PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning (CVPR2018)\n[2] Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights (ECCV2018)", "questions": "1. Can the authors provide a more detailed discussion of the novelty of their proposed method compared to prior work in the field of continual learning and hypernetwork-based approaches? \n2. The paper mentions that HyperMask has some limitations in terms of memory consumption due to the requirement for the hypernetwork's output layer to match the number of parameters in the target network. Are there any potential strategies or approaches to mitigate this memory consumption issue?\n3. The paper mentions that HyperMask has some limitations in terms of memory consumption due to the requirement for the hypernetwork's output layer to match the number of parameters in the target network. Are there any potential strategies or approaches to mitigate this memory consumption issue?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698752781381}, {"id": "ymS2QZKSNT", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1307/Reviewer_1qn2"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors are proposing a CIL method that can make use of a single network for multiple tasks. Based on the lottery ticket hypothesis, the proposed method only modifies the weights related to the specific task and this preserves the performance of previous tasks. Even though the hyper network consumes extra memory, it can keep the base network from growing over tasks.", "review_text": "The authors are proposing a CIL method that can make use of a single network for multiple tasks. Based on the lottery ticket hypothesis, the proposed method only modifies the weights related to the specific task and this preserves the performance of previous tasks. Even though the hyper network consumes extra memory, it can keep the base network from growing over tasks.", "strengths": "- The method itself is technically reasonable and the details of the method are well described in the paper.", "weaknesses": "- Baselines suggested in this paper are relatively outdated. I understand that incremental architecture methods are no longer dominating this field of research but it does not mean that they can ignore regularization or representation based methods. For example, FeCAM [1] does not even need to know the task index and it still outperforms the author's method.\n- The amount of experimental result is severely insufficient. According to the paper, HyperMask is superior to other methods only in Split-Cifar100 and in my knowledge this is not the best one in continual learning. Please refer to [1].\n- Almost all techniques utilized in this paper are brought from previous works and I'm not sure if this is something new in CIL. Also if the authors were to mention about Lottery-ticket hypothesis, I think there should be something more than this. This method looks like a combination of HAT and PackNet to me.\n\n[1] Goswami, Dipam, et al. \"FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning.\" arXiv preprint arXiv:2309.14062 (2023).", "questions": "- I would be appreciated if the comments above are resolved.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors are proposing a CIL method that can make use of a single network for multiple tasks. Based on the lottery ticket hypothesis, the proposed method only modifies the weights related to the specific task and this preserves the performance of previous tasks. Even though the hyper network consumes extra memory, it can keep the base network from growing over tasks.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "- The method itself is technically reasonable and the details of the method are well described in the paper.", "weaknesses": "- Baselines suggested in this paper are relatively outdated. I understand that incremental architecture methods are no longer dominating this field of research but it does not mean that they can ignore regularization or representation based methods. For example, FeCAM [1] does not even need to know the task index and it still outperforms the author's method.\n- The amount of experimental result is severely insufficient. According to the paper, HyperMask is superior to other methods only in Split-Cifar100 and in my knowledge this is not the best one in continual learning. Please refer to [1].\n- Almost all techniques utilized in this paper are brought from previous works and I'm not sure if this is something new in CIL. Also if the authors were to mention about Lottery-ticket hypothesis, I think there should be something more than this. This method looks like a combination of HAT and PackNet to me.\n\n[1] Goswami, Dipam, et al. \"FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning.\" arXiv preprint arXiv:2309.14062 (2023).", "questions": "- I would be appreciated if the comments above are resolved.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698716622838}], "openreview_url": "https://openreview.net/forum?id=LhNZqkuVte", "arxiv_id": "2310.00113", "paper_pdf": "papers/LhNZqkuVte.pdf", "paper_pdf_sha256": "16243ebef640821e79a26a623537654d7c0d091289098b893c28abbd256e5eab", "paper_pdf_bytes": 5385063, "paper_pdf_source": "openreview", "code_url": "https://github.com/gmum/HyperMask", "code_repository": "gmum/HyperMask", "code_commit": "8cceace567e27cd1348e42df0a07d0887a62bcdd", "code_archive": "repos/LhNZqkuVte.zip", "code_archive_sha256": "5864ac6bf7bfa3e7023af3588963ce6ab626ddd3d58399f30c6983eff600a3ea", "code_archive_bytes": 199176, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 207, "github_languages": {"Python": 337353}, "github_archived": false, "github_pushed_at": "2026-08-14T15:54:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hypermask-adaptive-hypernetwork-based-masks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "erHaiO9gz3m", "year": 2023, "status": "rejected", "title": "A Kernel-Based View of Language Model Fine-Tuning", "authors": ["Sadhika Malladi", "Alexander Wettig", "Dingli Yu", "Danqi Chen", "Sanjeev Arora"], "authorids": ["~Sadhika_Malladi2", "~Alexander_Wettig1", "~Dingli_Yu1", "~Danqi_Chen1", "~Sanjeev_Arora1"], "authors_source": "OpenReview API", "abstract": "It has become standard to solve NLP tasks by  fine-tuning pre-trained language models (LMs), especially in low-data settings. There is minimal theoretical understanding of empirical success, e.g., why fine-tuning a model with $10^8$ or more parameters on a couple dozen training points does not result in overfitting. We investigate whether the Neural Tangent Kernel (NTK)--which originated as a model to study the gradient descent dynamics of infinitely wide networks with suitable random initialization--describes fine-tuning of pre-trained LMs.  This study was inspired by the decent performance of NTK  for computer vision tasks (Wei et al., 2022). We also extend the NTK formalism to  fine-tuning with Adam.  We present extensive experiments  that suggest that once the task is formulated as a masked language modeling problem through prompting, the NTK lens can often reasonably describe the model updates during fine-tuning with both SGD and Adam.\nThis kernel view also suggests an explanation for success of parameter-efficient subspace-based fine-tuning methods. Finally, we suggest a path toward a formal explanation for our findings via Tensor Programs (Yang, 2020).", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "H5FQeoSh0PV", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5322/Reviewer_Ftva"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper studies the underlying reason behind why fine tuning pre-trained language models works well. The paper shows that in some cases, those fine tuned models can be described using neural tangent kernel. In those cases, the kernel view provide explanation of the fine-tuning success. \n", "review_text": "The paper presents several theoretical and experimental results, that I don't feel creates a coherent bottom line. In addition, It is hard to evaluate the quality of those results, without reading the appendix (which I didn't read). Thus, I cannot recommend accepting the paper.", "strengths": "Strength:\nThe paper tackle one of the most important questions in machine learning: understanding the reason behind the the performance of fine-tuned models, and in particular, when it should works.\n\nWeakness\nI found it hard to follow the paper's arguments. see section below for more details. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper studies the underlying reason behind why fine tuning pre-trained language models works well. The paper shows that in some cases, those fine tuned models can be described using neural tangent kernel. In those cases, the kernel view provide explanation of the fine-tuning success. \n", "strength_and_weaknesses": "Strength:\nThe paper tackle one of the most important questions in machine learning: understanding the reason behind the the performance of fine-tuned models, and in particular, when it should works.\n\nWeakness\nI found it hard to follow the paper's arguments. see section below for more details. ", "clarity,_quality,_novelty_and_reproducibility": "I found the paper unclear and hard to follow, in a way that is hard to appropriately appreciate the results. In particular I found the following points unclear:\n- Framing: The paper aims to further an empirical and theoretical understanding of pre-training and fine-tuning paradigm for NLP tasks, but this goal is too broad. As a result, the paper highlight too many points, and does not focus on none of them, especially with the the page limit. Creating a coherent story by refining a single take-away message from all of the described results, will help readers to follow the paper.\n- Mathematical motivation: The mathematical definitions appear without a proper motivation. Although the paper claim that some of the definitions are analogues to definitions in other papers, I believe the paper should be self-contained, and provide the intuition in high level.\n- Self contains claims: All mathematical claims should be self contained within the body of the paper, where the appendix can be used for proof, and for a formal version of the claims. Most claims of the paper are indeed presented in a self-contained manner, except Theorem 7.3, in which the conditions themselves appear in the appendix, without any informal phrasing of them.\n- references: it is very helpful that the paper have hyperlink to the arXiv, but it shouldn't replace mentioning the peer-review venue in which the papers appeared in. In the current format it is hard to validate that the paper is based on peered-review works.\n- Acronyms: please define all acronyms, for examples SGD.\n\n", "summary_of_the_review": "The paper presents several theoretical and experimental results, that I don't feel creates a coherent bottom line. In addition, It is hard to evaluate the quality of those results, without reading the appendix (which I didn't read). Thus, I cannot recommend accepting the paper.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667472558924}, {"id": "LssJwCXqcFO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5322/Reviewer_rpG6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper suggests that the success of language model fine-tuning (or with prompt) can be explained by neural tangent kernel (NTK), a tool that describes optimization behavior of infinitely wide neural networks. To bridge NTK with LM fine-tuning, the paper tries to justify that using NTK after pre-training phase (instead of initializing from lazy regime) is reasonable, and also derived empirical NTK for adam optimizer. The paper found empirical evidence that NTK behaves similarly to the original optimization methods, and also provided a theorem that shows prompt-based fine-tuning can exhibit kernel behavior", "review_text": "My feeling is a bit mixed for this paper. My current rating would be marginally below the threshold given that there are several issues worth clarifying. Regarding Weakness #4, I think different people may have different feeling on how important this is. I am relatively close to the median of that spectrum, but I can definitely see researchers having strong opinions from both sides. ", "strengths": "Strength: \nThe paper tries to explain language model fine-tuning using NTK and cleverly uses the fact that fine-tuning takes few samples (therefore one can compute eNTK in a feasible way). \n\nWeaknesses:\nOverall, the paper is trying to bridge theory and practice. But there are a few concerns:\n1. While some of the empirical results are good, some observations are quite mixed. For example, RTE linearization does not perform well for k=512 while eNTK still performs ok -- this seems confusing to me. \n2. Section 4 is unclear to me, the paper mentioned \"we model Adam updates with SignGD\", why can you replace Adam with SignGD and is the derived kernel exactly following Adam or is just an approximation of Adam (if later is the case, why not just approximate Adam with SGD)?\n3. The theory section (Section 7): I don't see how it is related to prompt-based fine-tuning but not standard fine-tuning. The paper mentioned \"The theory focuses on the prompt-based setting since we did not find empirical evidence of kernel behavior in the standard setting\" (I hope theory can advance empirical work but maybe now the paradigm is shifted), this is fine, but which assumption in Theorem 7.3 is specific to prompt such that the final conclusion is \"prompt-based FT of f will exhibit kernel behavior\"?\n4. The bigger question is whether NTK is THE theory or just A theory. Obviously a lot of practices in LM training or FT do not match with NTK behavior (e.g., learning rate warmup), therefore assuming NTK is THE theory and develop reasoning that tries to explain FT is a bit insufficient, so I think this is a weakness of the paper. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper suggests that the success of language model fine-tuning (or with prompt) can be explained by neural tangent kernel (NTK), a tool that describes optimization behavior of infinitely wide neural networks. To bridge NTK with LM fine-tuning, the paper tries to justify that using NTK after pre-training phase (instead of initializing from lazy regime) is reasonable, and also derived empirical NTK for adam optimizer. The paper found empirical evidence that NTK behaves similarly to the original optimization methods, and also provided a theorem that shows prompt-based fine-tuning can exhibit kernel behavior", "strength_and_weaknesses": "Strength: \nThe paper tries to explain language model fine-tuning using NTK and cleverly uses the fact that fine-tuning takes few samples (therefore one can compute eNTK in a feasible way). \n\nWeaknesses:\nOverall, the paper is trying to bridge theory and practice. But there are a few concerns:\n1. While some of the empirical results are good, some observations are quite mixed. For example, RTE linearization does not perform well for k=512 while eNTK still performs ok -- this seems confusing to me. \n2. Section 4 is unclear to me, the paper mentioned \"we model Adam updates with SignGD\", why can you replace Adam with SignGD and is the derived kernel exactly following Adam or is just an approximation of Adam (if later is the case, why not just approximate Adam with SGD)?\n3. The theory section (Section 7): I don't see how it is related to prompt-based fine-tuning but not standard fine-tuning. The paper mentioned \"The theory focuses on the prompt-based setting since we did not find empirical evidence of kernel behavior in the standard setting\" (I hope theory can advance empirical work but maybe now the paradigm is shifted), this is fine, but which assumption in Theorem 7.3 is specific to prompt such that the final conclusion is \"prompt-based FT of f will exhibit kernel behavior\"?\n4. The bigger question is whether NTK is THE theory or just A theory. Obviously a lot of practices in LM training or FT do not match with NTK behavior (e.g., learning rate warmup), therefore assuming NTK is THE theory and develop reasoning that tries to explain FT is a bit insufficient, so I think this is a weakness of the paper. ", "clarity,_quality,_novelty_and_reproducibility": "Clarity & Quality: The paper is overall clear, I'd be happy to get clarified on Section 4 & 7 a bit more. In terms of writing, I do think there is a better way to re-organize and better separate the assumption & theory with empirical results. Currently, empirical evidence and theory are interleaving and a bit hard to get the full picture. \n\nNovelty & Reproducibility: Overall relatively novel and should be easy to reproduce. ", "summary_of_the_review": "My feeling is a bit mixed for this paper. My current rating would be marginally below the threshold given that there are several issues worth clarifying. Regarding Weakness #4, I think different people may have different feeling on how important this is. I am relatively close to the median of that spectrum, but I can definitely see researchers having strong opinions from both sides. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666916128162}, {"id": "itkSSaeHDy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5322/Reviewer_NNgP"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies behavior of NTK on fine-tuning MLM models. Supposedly, pre-trained model is no longer iid random in parameters but still it can achieve good performance on CV tasks per previous literature. In this paper, authors extended the studies to NLP models with both regular fine-tuning and prompt-based learning cases. Authors also provide a modified eNTK version when using ADAM but not SGD. Last, authors tried to propose some hypothesis on why such model on eNTK behaves so.", "review_text": "Overall, I like the angles and the topic of this paper. But somehow I think it's a bit disappointing after reading it as it's not providing much insight. Rather, it's more of a mixed of empirical evidence and some not that useful theoretical results. I do believe we should encourage this type of work so overall I am more positive, but there are some confusions and I hope authors could explain it.\n\nOverall, I think the gist of the paper is a bit unclear. Mentioned too many points but I can't find the direct link between the claim and analysis. You listed many contributions in bullet points so I guess it's easier to go from there.\n\n- Does the eNTK method directly work for NLP tasks? In Section 5, we show that eNTK does not\nwork well for standard FT but it achieves performance comparable to prompt-based FT (Schick\n& Schutze, 2021; Gao et al., 2021) on a majority of downstream tasks that we evaluate. ¨\n\n- So I think this is just an empirical analysis. Somehow just an extension of domain from wei el. al. \nDo you agree with this? If so, can you highlight this is more of an empirical results. in addition,  can\nyou also greatly re-organize your paper to make contribution of empirical results and theoretical analysis clear?\n\n• The NTK theory was developed for gradient descent, whereas LM fine-tuning often uses\nAdam (Kingma & Ba, 2014) as a standard practice (Devlin et al., 2019). Can the above method\nbe extended to give insight into the effect of different optimization algorithms? In Section 4,\nwe derive a new asymmetric “kernel” formula (Definition 4.2) that describes the dynamics of\nshort-term training with Adam and shows that the corresponding eNTK achieves comparable\nperformance as FT using Adam (Table 1).\n\n- Here you use yet another approximation of Adam, so I don't see the purpose of it. Why does that \nmatter to analyze yet another version of eNTK  just due to a different gradient method? \nIn particular, Adam results in Table 1 doesn't really outperform SGD version eNTK musch. Especially,\nfor TREC ADAM result is much worse than eNTK one. What's the explanation? \n\n• Is the eNTK regression method simply an alternate model that also happens to have good\nperformance on the classification task, or is it actually a near-equivalent description of how\nparameters of the original model evolve during fine-tuning? Section 5.3 shows that for tasks that\nthe eNTK can solve, the fine-tuning of the pre-trained model does exhibit behavior consistent\nwith the dynamics of training kernel classifiers. See Definition 3.1.\n\n- I think this could be a real contribution. But somehow I feel not consistent as fine-tuning + eNTK \nseems not to work but here you showed its empirical performance justified the assumptions/desired\nbehavior of NTK in 3.1 ? It's kind of contradicting. \n\n• If the eNTK can solve NLP tasks, then does the eNTK give insight into phenomena such as why\nprompting can improve performance and how parameter-efficient fine-tuning methods affect\noptimization? Figure 1 demonstrates that adding a prompt is necessary for the eNTK to solve the\ntask, and in Section 6, we apply the kernel lens to provide a possible explanation for the empirical\nsuccess of subspace-based fine-tuning methods (Hu et al., 2021; Aghajanyan et al., 2021).\n\n- Sorry I don't quite get why prompting can improve performance is related to subspace-based fine-tuning methods?\nI think the connection is not very obvious to me. I was feeling strange when I read your section 6. I am not sure\nwhy you want to analyze it. And I felt the connection to other part of paper is too weak such that it's another paper.\n\nWe conclude by proposing a rigorous mechanism through which fine-tuning of complex architectures\n(e.g., Transformers (Vaswani et al., 2017)) with prompts can exhibit kernel behavior. This is done\nin context of networks whose width goes to infinity, but unlike standard infinite-width NTK theory\nit allows a non-random initialization that is the result of pretraining. This result uses the Tensor\nPrograms framework (Yang, 2019; 2020a;b; Yang & Littwin, 2021; Yang & Hu, 2021).\n\n- I actually feel this part of analysis is the most pertinent part regarding your motivations. Why don't you put it in the front? But I am not sure the correctness and framework as I didn't know Tensor Programs before.\n\n", "strengths": "Strength:\n\n- An interesting observation and continuation of previous work.\n- Provide studies over various related topics with mixed empirical and theoretical analysis.\n\nWeakness:\n\nMain contribution seems to just be combining two major previous results and somehow it's not providing enough insight.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies behavior of NTK on fine-tuning MLM models. Supposedly, pre-trained model is no longer iid random in parameters but still it can achieve good performance on CV tasks per previous literature. In this paper, authors extended the studies to NLP models with both regular fine-tuning and prompt-based learning cases. Authors also provide a modified eNTK version when using ADAM but not SGD. Last, authors tried to propose some hypothesis on why such model on eNTK behaves so.", "strength_and_weaknesses": "Strength:\n\n- An interesting observation and continuation of previous work.\n- Provide studies over various related topics with mixed empirical and theoretical analysis.\n\nWeakness:\n\nMain contribution seems to just be combining two major previous results and somehow it's not providing enough insight.", "clarity,_quality,_novelty_and_reproducibility": "The paper is not well organized. Claim many things but the results are spread out here and there. It's very hard to understand the contribution before linking them. In terms of novelty, I think the angle to approach the underlying NLP phenomena is interesting but no real technical difficulties or insightful theoretical results given.\n", "summary_of_the_review": "Overall, I like the angles and the topic of this paper. But somehow I think it's a bit disappointing after reading it as it's not providing much insight. Rather, it's more of a mixed of empirical evidence and some not that useful theoretical results. I do believe we should encourage this type of work so overall I am more positive, but there are some confusions and I hope authors could explain it.\n\nOverall, I think the gist of the paper is a bit unclear. Mentioned too many points but I can't find the direct link between the claim and analysis. You listed many contributions in bullet points so I guess it's easier to go from there.\n\n- Does the eNTK method directly work for NLP tasks? In Section 5, we show that eNTK does not\nwork well for standard FT but it achieves performance comparable to prompt-based FT (Schick\n& Schutze, 2021; Gao et al., 2021) on a majority of downstream tasks that we evaluate. ¨\n\n- So I think this is just an empirical analysis. Somehow just an extension of domain from wei el. al. \nDo you agree with this? If so, can you highlight this is more of an empirical results. in addition,  can\nyou also greatly re-organize your paper to make contribution of empirical results and theoretical analysis clear?\n\n• The NTK theory was developed for gradient descent, whereas LM fine-tuning often uses\nAdam (Kingma & Ba, 2014) as a standard practice (Devlin et al., 2019). Can the above method\nbe extended to give insight into the effect of different optimization algorithms? In Section 4,\nwe derive a new asymmetric “kernel” formula (Definition 4.2) that describes the dynamics of\nshort-term training with Adam and shows that the corresponding eNTK achieves comparable\nperformance as FT using Adam (Table 1).\n\n- Here you use yet another approximation of Adam, so I don't see the purpose of it. Why does that \nmatter to analyze yet another version of eNTK  just due to a different gradient method? \nIn particular, Adam results in Table 1 doesn't really outperform SGD version eNTK musch. Especially,\nfor TREC ADAM result is much worse than eNTK one. What's the explanation? \n\n• Is the eNTK regression method simply an alternate model that also happens to have good\nperformance on the classification task, or is it actually a near-equivalent description of how\nparameters of the original model evolve during fine-tuning? Section 5.3 shows that for tasks that\nthe eNTK can solve, the fine-tuning of the pre-trained model does exhibit behavior consistent\nwith the dynamics of training kernel classifiers. See Definition 3.1.\n\n- I think this could be a real contribution. But somehow I feel not consistent as fine-tuning + eNTK \nseems not to work but here you showed its empirical performance justified the assumptions/desired\nbehavior of NTK in 3.1 ? It's kind of contradicting. \n\n• If the eNTK can solve NLP tasks, then does the eNTK give insight into phenomena such as why\nprompting can improve performance and how parameter-efficient fine-tuning methods affect\noptimization? Figure 1 demonstrates that adding a prompt is necessary for the eNTK to solve the\ntask, and in Section 6, we apply the kernel lens to provide a possible explanation for the empirical\nsuccess of subspace-based fine-tuning methods (Hu et al., 2021; Aghajanyan et al., 2021).\n\n- Sorry I don't quite get why prompting can improve performance is related to subspace-based fine-tuning methods?\nI think the connection is not very obvious to me. I was feeling strange when I read your section 6. I am not sure\nwhy you want to analyze it. And I felt the connection to other part of paper is too weak such that it's another paper.\n\nWe conclude by proposing a rigorous mechanism through which fine-tuning of complex architectures\n(e.g., Transformers (Vaswani et al., 2017)) with prompts can exhibit kernel behavior. This is done\nin context of networks whose width goes to infinity, but unlike standard infinite-width NTK theory\nit allows a non-random initialization that is the result of pretraining. This result uses the Tensor\nPrograms framework (Yang, 2019; 2020a;b; Yang & Littwin, 2021; Yang & Hu, 2021).\n\n- I actually feel this part of analysis is the most pertinent part regarding your motivations. Why don't you put it in the front? But I am not sure the correctness and framework as I didn't know Tensor Programs before.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666665776166}], "openreview_url": "https://openreview.net/forum?id=erHaiO9gz3m", "arxiv_id": "2210.05643", "paper_pdf": "papers/erHaiO9gz3m.pdf", "paper_pdf_sha256": "56f80e4a4bcd8c4ef503a8746e2bd9a69d9b67a666094a7215842fa269850b88", "paper_pdf_bytes": 480136, "paper_pdf_source": "openreview", "code_url": "https://github.com/princeton-nlp/LM-Kernel-FT", "code_repository": "princeton-nlp/LM-Kernel-FT", "code_commit": "81450f619004ac1f981b31c3a12b4baf78c673c1", "code_archive": "repos/erHaiO9gz3m.zip", "code_archive_sha256": "5d7dc10e149f2261cd76f3a4025559bcb5e1f675ef9b4cdd198c78bcd2ab5f67", "code_archive_bytes": 47298, "code_file_count": 12, "code_extensions": {".py": 10, ".sh": 2}, "github_disk_usage_kb": 411, "github_languages": {"Python": 166036, "Shell": 4988}, "github_archived": false, "github_pushed_at": "2023-09-04T22:20:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-kernel-based-view-of-language-model-fine"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "g4E6SAAvACo", "year": 2021, "status": "rejected", "title": "Neural Architecture Search without Training", "authors": ["Joseph Mellor", "Jack Turner", "Amos Storkey", "Elliot J. Crowley"], "authorids": ["joe.mellor@ed.ac.uk", "~Jack_Turner1", "~Amos_Storkey1", "~Elliot_J._Crowley1"], "authors_source": "OpenReview API", "abstract": "The time and effort involved in hand-designing deep neural networks is immense. This has prompted the development of Neural Architecture Search (NAS) techniques to automate this design. However, NAS algorithms tend to be slow and expensive; they need to train vast numbers of candidate networks to inform the search process. This could be remedied if we could infer a network's trained accuracy from its initial state. In this work, we examine the correlation of linear maps induced by augmented versions of a single image in untrained networks and motivate how this can be used to give a measure which is highly indicative of a network’s trained performance. We incorporate this measure into a simple algorithm that allows us to search for powerful networks without any training in a matter of seconds on a single GPU, and verify its effectiveness on NAS-Bench-101 and NAS-Bench-201. Finally, we show that our approach can be readily combined with more expensive search methods for added value: we modify regularised evolutionary search to produce a novel algorithm that outperforms its predecessor.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "hqd5t_dj1Jp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3110/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "** Summary\n\nThe paper mainly introduces a metric to benchmark the performance of neural networks without training – the correlation of Jacobian subject to different augmented versions of a single image. The key motivation is, high-performance networks tend to represent data of small perturbations with different hyperplanes at initialization, so that the distinguishing capability may also be stronger. The divergence of the hyperplanes can be efficiently estimated via the correlation of the Jacobian, thus quantified by the score in Eq. 2. The effectiveness of the proposed metric is mainly verified in two NAS benchmarks (NAS-Bench-101 and NAS-Bench-201), whose correlation to the actual accuracy is relatively significant (Fig 3). Compared with existing NAS frameworks, the proposed method is very efficient, moreover, able to obtain competitive performance.\n\n** Pros\n\nI really appreciate the paper proposes a new direction to benchmark and understand neural networks. The motivation is very impressive. Previous NAS frameworks based on score predictors, e.g. NAO [*1] and ChamNet [*2], have shown that it is possible to directly predict the performance from the architecture embeddings without training, however, the underlaying mechanism is not clear. The proposed metric may help to uncover the relation between architecture choices and the performances. Furthermore, with the help of the metric, the proposed NAS framework (NASWOT) is elegant and seems to be effective, which could be useful in practice.\n\n** Cons\n\n1)\tLack of theoretical evidence to support the intuition of the proposed method. To my knowledge, network initialization is important but cannot interpret all the behaviors; so, I still doubt why and how the Jacobian at initialization reflects the final score. I guess [*3] may help in the analysis. \n2)\tSome important ablations are missing in the experiments, which makes the method less convincing. Please refer to the following suggestions. \n\n** Concerns and suggestions\n\nMajor:\n\nTo my understanding, the methodology of initialization (such as identity init, random gaussian with fixed variance (e.g. 0.01), MSRA init. [*4], etc.) could affect the initial Jacobian a lot, but little on the final accuracy especially with BN, which is not compared in the paper (I suppose the comparison in Fig 7 (bottom-left) is performed with different runs rather than different initialization methods, please correct me if I make some misunderstandings). Notice that some initialization methods like [*4] may leak architecture parameters (e.g. number of input channels). I think it is very important to check whether the Jacobian under a certain initialization methodology (improperly) takes advantage of the architecture bias in NAS benchmarks. I am pleased to raise the rating if the authors clear my concern, for example, providing comparisons of different initialization methods, or benchmarking the method with different search spaces on large datasets (e.g. search space of FBNet [*5] on ImageNet). \n\nMinor:\n\n1)\tIt is interesting if the authors analyze or visualize (just like Fig 1) how the correlation of Jacobian evolves during the training for different architectures. \n2)\tPlease clarify the details about “… adjust the final classifier layer to output a scalar” in Page 4. \n\n[*1] Luo, Renqian, et al. \"Neural architecture optimization.\" Advances in neural information processing systems. 2018.\n\n[*2] Dai, Xiaoliang, et al. \"Chamnet: Towards efficient network design through platform-aware model adaptation.\" Proceedings of the IEEE Conference on computer vision and pattern recognition. 2019.\n\n[*3] Jacot, Arthur, Franck Gabriel, and Clément Hongler. \"Neural tangent kernel: Convergence and generalization in neural networks.\" Advances in neural information processing systems. 2018.\n\n[*4] He, Kaiming, et al. \"Delving deep into rectifiers: Surpassing human-level performance on imagenet classification.\" Proceedings of the IEEE international conference on computer vision. 2015.\n\n[*5] Wu, Bichen, et al. \"Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n\n===================\n\nThanks for the rebuttal and the additional empirical results. In general, I really appreciate the \"brave new idea\" proposed by the authors, which could inspire new insights on neural architecture design.  However, my major concern still exists: the proposed metric seems to be sensitive to the choice of initialization functions. For example, uniform initialization [0, 1] seems not work at all but no further explanations, which may indicate the proposed method may work in a different way (e.g. taking advantage of some search space's bias).  So, I keep my original rating. \n  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting, but important comparisons are missing", "review": "** Summary\n\nThe paper mainly introduces a metric to benchmark the performance of neural networks without training – the correlation of Jacobian subject to different augmented versions of a single image. The key motivation is, high-performance networks tend to represent data of small perturbations with different hyperplanes at initialization, so that the distinguishing capability may also be stronger. The divergence of the hyperplanes can be efficiently estimated via the correlation of the Jacobian, thus quantified by the score in Eq. 2. The effectiveness of the proposed metric is mainly verified in two NAS benchmarks (NAS-Bench-101 and NAS-Bench-201), whose correlation to the actual accuracy is relatively significant (Fig 3). Compared with existing NAS frameworks, the proposed method is very efficient, moreover, able to obtain competitive performance.\n\n** Pros\n\nI really appreciate the paper proposes a new direction to benchmark and understand neural networks. The motivation is very impressive. Previous NAS frameworks based on score predictors, e.g. NAO [*1] and ChamNet [*2], have shown that it is possible to directly predict the performance from the architecture embeddings without training, however, the underlaying mechanism is not clear. The proposed metric may help to uncover the relation between architecture choices and the performances. Furthermore, with the help of the metric, the proposed NAS framework (NASWOT) is elegant and seems to be effective, which could be useful in practice.\n\n** Cons\n\n1)\tLack of theoretical evidence to support the intuition of the proposed method. To my knowledge, network initialization is important but cannot interpret all the behaviors; so, I still doubt why and how the Jacobian at initialization reflects the final score. I guess [*3] may help in the analysis. \n2)\tSome important ablations are missing in the experiments, which makes the method less convincing. Please refer to the following suggestions. \n\n** Concerns and suggestions\n\nMajor:\n\nTo my understanding, the methodology of initialization (such as identity init, random gaussian with fixed variance (e.g. 0.01), MSRA init. [*4], etc.) could affect the initial Jacobian a lot, but little on the final accuracy especially with BN, which is not compared in the paper (I suppose the comparison in Fig 7 (bottom-left) is performed with different runs rather than different initialization methods, please correct me if I make some misunderstandings). Notice that some initialization methods like [*4] may leak architecture parameters (e.g. number of input channels). I think it is very important to check whether the Jacobian under a certain initialization methodology (improperly) takes advantage of the architecture bias in NAS benchmarks. I am pleased to raise the rating if the authors clear my concern, for example, providing comparisons of different initialization methods, or benchmarking the method with different search spaces on large datasets (e.g. search space of FBNet [*5] on ImageNet). \n\nMinor:\n\n1)\tIt is interesting if the authors analyze or visualize (just like Fig 1) how the correlation of Jacobian evolves during the training for different architectures. \n2)\tPlease clarify the details about “… adjust the final classifier layer to output a scalar” in Page 4. \n\n[*1] Luo, Renqian, et al. \"Neural architecture optimization.\" Advances in neural information processing systems. 2018.\n\n[*2] Dai, Xiaoliang, et al. \"Chamnet: Towards efficient network design through platform-aware model adaptation.\" Proceedings of the IEEE Conference on computer vision and pattern recognition. 2019.\n\n[*3] Jacot, Arthur, Franck Gabriel, and Clément Hongler. \"Neural tangent kernel: Convergence and generalization in neural networks.\" Advances in neural information processing systems. 2018.\n\n[*4] He, Kaiming, et al. \"Delving deep into rectifiers: Surpassing human-level performance on imagenet classification.\" Proceedings of the IEEE international conference on computer vision. 2015.\n\n[*5] Wu, Bichen, et al. \"Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n\n===================\n\nThanks for the rebuttal and the additional empirical results. In general, I really appreciate the \"brave new idea\" proposed by the authors, which could inspire new insights on neural architecture design.  However, my major concern still exists: the proposed metric seems to be sensitive to the choice of initialization functions. For example, uniform initialization [0, 1] seems not work at all but no further explanations, which may indicate the proposed method may work in a different way (e.g. taking advantage of some search space's bias).  So, I keep my original rating. \n  \n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604081615228}, {"id": "0Nh6y3BICBk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3110/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary\nThis paper attempts to infer a network's accuracy at initialization without training it, which can speed up neural architecture search and greatly reduce the search cost. Specifically, they propose a metric based on the Jacobian of the loss with respect to a minibatch of input data. The authors show that with this metric, they can find architectures with reasonable accuracy on CIFAR-10/CIFAR-100 in the NAS-Bench-201, while using much less search cost compared to previous NAS methods.\n\n\n\n# Strong points\n1. This paper is exploring a novel and interesting direction. Estimating the network’s performance at initialization can greatly reduce the search speed.\n2. It might be difficult for the training-free metric to outperform conventional metrics (e.g., validation accuracy after training). This paper finds a good use case of the training-free metric in practice. The authors propose to use the training-free metric to select the initial population of evolutionary algorithms and empirically demonstrate its usefulness on CIFAR-10 and CIFAR-100.\n\n\n# Weak points\n1. It will be helpful to make the motivation of the proposed metric more clear. I found the second paragraph in Section 3 hard to understand without digging into the cited literature.\n\n2. How to interpret the value in the correlation matrix \\Sigma? Empirically, Figure 1 shows that we want to find an architecture whose values in matrix \\Sigma are mostly around zero with a small positive skew. But how does this relate to the two motivations mentioned in Section 3: (1) being flexible and (2) being invariant/robust to small perturbations. Does small value mean less or more flexible/robust? This is important for people to understand why the metric works.\n\n3. In the experiments, all the images in the minibatch are data augmentations of the same image. This seems to be an important design choice. More insights on this part would be very helpful. How is this design choice (using data augmentations of the *SAME* image) related to the two motivations (flexible & invariant)? \n\n4. As mentioned by the authors, the two motivations are actually antagonistic. How does the proposed method balance them?\n\n5. Why choose cutout as the data augmentation strategy? Is the method sensitive or not to other perturbations, e.g., adding small noise?\n\n6. I notice that the correlation (tau) value in Figure 3 is not high. What is the correlation between the typical criteria (e.g., validation accuracy after training a small number of epochs) and the final test accuracy?\n\n  The tau in Figure 3 is actually undefined. Is that Kendall tau?\n\n  Why evaluate the correlation between **validation** accuracy, not the final **test** accuracy, as provided in NAS-Bench-201?\n\n7. Section 5 mentions that “ori-test@12” is used as the metric during search. Is that the accuracy on the **test** set after 12 epoch training (with learning rate decay to 0)? But the common practice is to use the **validation** accuracy during search and report the final test accuracy. This seems to be unfair.\n\n8. For NASWOT (N=100), the accuracy on CIFAR-10 is reasonable and actually pretty good considering the small search cost. However, the accuracy in ImageNet-16-120 seems to be very low. Why does this happen? Is it possible that the proposed method overfit to CIFAR-10?\n\n  If I understand correctly, the architecture is searched on CIFAR-10, and then evaluated on all three datasets. What if we search on ImageNet-16-120 with this metric and can we match the performance of other baselines on ImageNet-16-120? This might be an unfair comparison, but will be important for people to know whether the proposed method/metric can generalize to different datasets.\n\n9. Table 2 only reports N=10 and N=100. What if we significantly increase the value of N and will the performance be similar or better than other methods like RS and REINFORCE? As the proposed method uses very small cost, even using a large N, the cost would still be reasonable. It will be great to see we can achieve much better performance when increasing N.\n\n10. The writing and organization in Section 4 need to be improved. Adding subtitles might make it easier to read. Also, Section 4 mentions REAL at the beginning but REAL is not explained in detail until Section 5. The rows “Optimal (N=10/100)” don’t seem to help validate the proposed method and are a bit confusing.\n\n11. It will also greatly strengthen this work if the authors can show the effectiveness of the proposed metric on a more realistic search space, e.g., the DARTS search, and evaluate the found architecture on a larger dataset, e.g., ImageNet.\n\n# Justification of rating\nI like the idea of estimating a network’s performance without training it. But this paper needs more refinement before being accepted. As mentioned above, the motivation, and the relationship between the method and motivation, need to be explained more clearly. More analysis is needed to understand and justify the proposed method. The writing also needs improvement.\n\n# After rebuttal\nI would like to thank the authors for the hard work during the rebuttal. The ablation study of the data augmentation strategy and other added results are very helpful. Regarding the explanation of the flexibility and invariance, although I could get some intuition, I am still not fully convinced. So I keep my original rating. One possible way to make this work stronger and meet the acceptance criteria is to provide some empirical (or even better, theoretical) analysis of the influence of $\\Sigma$ on the flexibility and invariance of a network.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Need more analysis of the proposed method", "review": "# Summary\nThis paper attempts to infer a network's accuracy at initialization without training it, which can speed up neural architecture search and greatly reduce the search cost. Specifically, they propose a metric based on the Jacobian of the loss with respect to a minibatch of input data. The authors show that with this metric, they can find architectures with reasonable accuracy on CIFAR-10/CIFAR-100 in the NAS-Bench-201, while using much less search cost compared to previous NAS methods.\n\n\n\n# Strong points\n1. This paper is exploring a novel and interesting direction. Estimating the network’s performance at initialization can greatly reduce the search speed.\n2. It might be difficult for the training-free metric to outperform conventional metrics (e.g., validation accuracy after training). This paper finds a good use case of the training-free metric in practice. The authors propose to use the training-free metric to select the initial population of evolutionary algorithms and empirically demonstrate its usefulness on CIFAR-10 and CIFAR-100.\n\n\n# Weak points\n1. It will be helpful to make the motivation of the proposed metric more clear. I found the second paragraph in Section 3 hard to understand without digging into the cited literature.\n\n2. How to interpret the value in the correlation matrix \\Sigma? Empirically, Figure 1 shows that we want to find an architecture whose values in matrix \\Sigma are mostly around zero with a small positive skew. But how does this relate to the two motivations mentioned in Section 3: (1) being flexible and (2) being invariant/robust to small perturbations. Does small value mean less or more flexible/robust? This is important for people to understand why the metric works.\n\n3. In the experiments, all the images in the minibatch are data augmentations of the same image. This seems to be an important design choice. More insights on this part would be very helpful. How is this design choice (using data augmentations of the *SAME* image) related to the two motivations (flexible & invariant)? \n\n4. As mentioned by the authors, the two motivations are actually antagonistic. How does the proposed method balance them?\n\n5. Why choose cutout as the data augmentation strategy? Is the method sensitive or not to other perturbations, e.g., adding small noise?\n\n6. I notice that the correlation (tau) value in Figure 3 is not high. What is the correlation between the typical criteria (e.g., validation accuracy after training a small number of epochs) and the final test accuracy?\n\n  The tau in Figure 3 is actually undefined. Is that Kendall tau?\n\n  Why evaluate the correlation between **validation** accuracy, not the final **test** accuracy, as provided in NAS-Bench-201?\n\n7. Section 5 mentions that “ori-test@12” is used as the metric during search. Is that the accuracy on the **test** set after 12 epoch training (with learning rate decay to 0)? But the common practice is to use the **validation** accuracy during search and report the final test accuracy. This seems to be unfair.\n\n8. For NASWOT (N=100), the accuracy on CIFAR-10 is reasonable and actually pretty good considering the small search cost. However, the accuracy in ImageNet-16-120 seems to be very low. Why does this happen? Is it possible that the proposed method overfit to CIFAR-10?\n\n  If I understand correctly, the architecture is searched on CIFAR-10, and then evaluated on all three datasets. What if we search on ImageNet-16-120 with this metric and can we match the performance of other baselines on ImageNet-16-120? This might be an unfair comparison, but will be important for people to know whether the proposed method/metric can generalize to different datasets.\n\n9. Table 2 only reports N=10 and N=100. What if we significantly increase the value of N and will the performance be similar or better than other methods like RS and REINFORCE? As the proposed method uses very small cost, even using a large N, the cost would still be reasonable. It will be great to see we can achieve much better performance when increasing N.\n\n10. The writing and organization in Section 4 need to be improved. Adding subtitles might make it easier to read. Also, Section 4 mentions REAL at the beginning but REAL is not explained in detail until Section 5. The rows “Optimal (N=10/100)” don’t seem to help validate the proposed method and are a bit confusing.\n\n11. It will also greatly strengthen this work if the authors can show the effectiveness of the proposed metric on a more realistic search space, e.g., the DARTS search, and evaluate the found architecture on a larger dataset, e.g., ImageNet.\n\n# Justification of rating\nI like the idea of estimating a network’s performance without training it. But this paper needs more refinement before being accepted. As mentioned above, the motivation, and the relationship between the method and motivation, need to be explained more clearly. More analysis is needed to understand and justify the proposed method. The writing also needs improvement.\n\n# After rebuttal\nI would like to thank the authors for the hard work during the rebuttal. The ablation study of the data augmentation strategy and other added results are very helpful. Regarding the explanation of the flexibility and invariance, although I could get some intuition, I am still not fully convinced. So I keep my original rating. One possible way to make this work stronger and meet the acceptance criteria is to provide some empirical (or even better, theoretical) analysis of the influence of $\\Sigma$ on the flexibility and invariance of a network.\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603947125056}, {"id": "3QVLB97mDI4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3110/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "### Summary\nThis paper proposes a zero-shot NAS algorithm.  Instead of using validation accuracy that requires training, a score **S** is defined. It is computed as follows:\n    A single image is augmented using cut-out several times to produce a batch of images. For every datapoint, the Jacobian is computed to form a vector J. Then the correlation matrix $\\Sigma$ for the J is computed. Score S is the total number of entries in $\\Sigma$ between 0 and upper bound $\\beta$. $\\beta$ is a hyper parameter.\nThe search algorithm **NASWOT** now randomly samples K networks from the search space and finds the network with the highest score **S**. \n\n###Pros:\n  The idea of using a metric based on Jacobian rather than validation accuracy is interesting. \n   Their experimental setup demonstrated the correlation between ranking based on S and validation accuracy, NAS on 3 datasets on NASBENCH-201\n\n###Questions\n\n 1. Main concern is that as the method relies on only 1 image and the initialization, what is the guarantee that it would generalize.  In Figure 4, the ablation study considers networks from different accuracy buckets. The Standard deviation of the score is high for some networks.  What would be more useful is if around 200 networks were sampled from 80 to 100 accuracy bucket, for each of the 20 images (instead of 10), we compute the Kendall Tau of the ranking based on **S** and validation accuracy. You can report the average Kendall Tau. Similar experiment can be performed for initialization too.\n \n2. How will this method account for the fact that some datasets have very diverse images? For example, NASWOT deteriorates in performance on Imagenet which has more diverse images. Kendall Tau computed on ImageNet in Figure 3 also reflects the same. Instead of 1 image, can we sample 10 representative images from the distributed and compute average of S on all of them?\n\n3.  In their paper, Real et al. use an initial population of 100. 10 might be too less for initial population. So the comparison of REAL vs AREAL might not be fair. Even in that case, the performance improvement is not significant.\n\n4. In figure 3, why is the correlation not higher than 0.55? Does this mean that while it has some signal, it is not enough to discriminate well performing networks from others?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "### Summary\nThis paper proposes a zero-shot NAS algorithm.  Instead of using validation accuracy that requires training, a score **S** is defined. It is computed as follows:\n    A single image is augmented using cut-out several times to produce a batch of images. For every datapoint, the Jacobian is computed to form a vector J. Then the correlation matrix $\\Sigma$ for the J is computed. Score S is the total number of entries in $\\Sigma$ between 0 and upper bound $\\beta$. $\\beta$ is a hyper parameter.\nThe search algorithm **NASWOT** now randomly samples K networks from the search space and finds the network with the highest score **S**. \n\n###Pros:\n  The idea of using a metric based on Jacobian rather than validation accuracy is interesting. \n   Their experimental setup demonstrated the correlation between ranking based on S and validation accuracy, NAS on 3 datasets on NASBENCH-201\n\n###Questions\n\n 1. Main concern is that as the method relies on only 1 image and the initialization, what is the guarantee that it would generalize.  In Figure 4, the ablation study considers networks from different accuracy buckets. The Standard deviation of the score is high for some networks.  What would be more useful is if around 200 networks were sampled from 80 to 100 accuracy bucket, for each of the 20 images (instead of 10), we compute the Kendall Tau of the ranking based on **S** and validation accuracy. You can report the average Kendall Tau. Similar experiment can be performed for initialization too.\n \n2. How will this method account for the fact that some datasets have very diverse images? For example, NASWOT deteriorates in performance on Imagenet which has more diverse images. Kendall Tau computed on ImageNet in Figure 3 also reflects the same. Instead of 1 image, can we sample 10 representative images from the distributed and compute average of S on all of them?\n\n3.  In their paper, Real et al. use an initial population of 100. 10 might be too less for initial population. So the comparison of REAL vs AREAL might not be fair. Even in that case, the performance improvement is not significant.\n\n4. In figure 3, why is the correlation not higher than 0.55? Does this mean that while it has some signal, it is not enough to discriminate well performing networks from others?\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603910651215}, {"id": "8cb-qlgG7YQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3110/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe authors propose a training-free way of estimating the performance of a deep net architecture after training using correlations between linearizations of the network at initialization for different augmentations of the same image. This estimate is used as signal to construct NAS algorithms that do not require training deep nets and is evaluated on two datasets. Although I have significant concerns about the practicality of the method, I believe it establishes a sufficiently distinct direction for NAS research that could merit acceptance.\n\nStrengths:\n1. To my knowledge the paper is the first to implement training-free NAS. \n2. The method is simple and easy to implement.\n3. The method achieves decent performance on CIFAR in dramatically less time than previous approaches.\n4. What seems like a complete codebase is provided (although see Question 1 below).\n5. The paper is clear and easy-to-follow.\n\nWeaknesses:\n1. The justification for the actual score used is weak (see Questions 2-3 below).\n2. The method seems limited to vision data and networks with ReLU activations.\n3. Performance on the ImageNet subset of NAS-Bench-201, the only evaluation using non-CIFAR data, is poor.\n4. Limited exploratory and benchmark evaluations (see Questions 4-7 below).\n\nQuestions:\n1. I was not able to run the provided code (search.py) using the instructions provided; could the authors provide a dependency list?\n2. As the score used is non-obvious and has no mathematical basis, it seems likely some trial-and-error was used to find it; is this the case, and if so what sorts of rules were tried that did not work.\n3. Why not just use the correlations between gradients rather than using the indicator function to obtain some linearization?\n4. What is the effect of beta on performance? \n5. Do other types of data augmentation work?\n6. For NAS-Bench-201, why was N>100 (e.g. N=1000) not tried? There is clearly room to improve and going that high still leaves the NASWOT algorithm by far the fastest. \n7. How does NASWOT perform on larger search spaces such as DARTS (Liu et al., 2019)?\n\nNotes:\n1. In two locations in the paper (3rd para of intro, 3rd para of background) the authors suggest that Li & Talwalkar (2019) show that WS inhibits architecture search and/or struggles against random search; however, that paper also shows that combining WS with random search is outperforms the latter.\n2. “Moreover, popular search spaces have been shown to be over-engineered, exhibiting little variety in their trained networks (Yang et al., 2020).” - is there evidence that NAS-Bench-101 and NAS-Bench-201 do not also suffer from this? Both were released before the publication of Yang et al. (2020).\n3. “Given a neural network with rectified linear units, we can, at each unit in each layer, identify a binary indicator as to whether the unit is inactive (the value is negative and hence is multiplied by zero) or active (in which case its value is multiplied by one).” - is the proposed method dependent on ReLU activations being used?\n4. Figure 4: what is the small circle that appears either above or below many of the box-and-whisker points?\n5. Table 2: it is standard to report the optimal in the entire search space, not just for N=10/100.\n\n# Post-response update\nThank you to the authors for very helpful clarifications. This paper provides a reasonable start for a new potential direction in NAS research and so may be worth presenting at the conference, but the justification and applicability of the method is somewhat limited. I therefore stand by my original assessment.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting albeit under-explored proof-of-concept for training-free NAS.", "review": "Summary:\nThe authors propose a training-free way of estimating the performance of a deep net architecture after training using correlations between linearizations of the network at initialization for different augmentations of the same image. This estimate is used as signal to construct NAS algorithms that do not require training deep nets and is evaluated on two datasets. Although I have significant concerns about the practicality of the method, I believe it establishes a sufficiently distinct direction for NAS research that could merit acceptance.\n\nStrengths:\n1. To my knowledge the paper is the first to implement training-free NAS. \n2. The method is simple and easy to implement.\n3. The method achieves decent performance on CIFAR in dramatically less time than previous approaches.\n4. What seems like a complete codebase is provided (although see Question 1 below).\n5. The paper is clear and easy-to-follow.\n\nWeaknesses:\n1. The justification for the actual score used is weak (see Questions 2-3 below).\n2. The method seems limited to vision data and networks with ReLU activations.\n3. Performance on the ImageNet subset of NAS-Bench-201, the only evaluation using non-CIFAR data, is poor.\n4. Limited exploratory and benchmark evaluations (see Questions 4-7 below).\n\nQuestions:\n1. I was not able to run the provided code (search.py) using the instructions provided; could the authors provide a dependency list?\n2. As the score used is non-obvious and has no mathematical basis, it seems likely some trial-and-error was used to find it; is this the case, and if so what sorts of rules were tried that did not work.\n3. Why not just use the correlations between gradients rather than using the indicator function to obtain some linearization?\n4. What is the effect of beta on performance? \n5. Do other types of data augmentation work?\n6. For NAS-Bench-201, why was N>100 (e.g. N=1000) not tried? There is clearly room to improve and going that high still leaves the NASWOT algorithm by far the fastest. \n7. How does NASWOT perform on larger search spaces such as DARTS (Liu et al., 2019)?\n\nNotes:\n1. In two locations in the paper (3rd para of intro, 3rd para of background) the authors suggest that Li & Talwalkar (2019) show that WS inhibits architecture search and/or struggles against random search; however, that paper also shows that combining WS with random search is outperforms the latter.\n2. “Moreover, popular search spaces have been shown to be over-engineered, exhibiting little variety in their trained networks (Yang et al., 2020).” - is there evidence that NAS-Bench-101 and NAS-Bench-201 do not also suffer from this? Both were released before the publication of Yang et al. (2020).\n3. “Given a neural network with rectified linear units, we can, at each unit in each layer, identify a binary indicator as to whether the unit is inactive (the value is negative and hence is multiplied by zero) or active (in which case its value is multiplied by one).” - is the proposed method dependent on ReLU activations being used?\n4. Figure 4: what is the small circle that appears either above or below many of the box-and-whisker points?\n5. Table 2: it is standard to report the optimal in the entire search space, not just for N=10/100.\n\n# Post-response update\nThank you to the authors for very helpful clarifications. This paper provides a reasonable start for a new potential direction in NAS research and so may be worth presenting at the conference, but the justification and applicability of the method is somewhat limited. I therefore stand by my original assessment.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603861646385}], "openreview_url": "https://openreview.net/forum?id=g4E6SAAvACo", "arxiv_id": "2006.04647", "paper_pdf": "papers/g4E6SAAvACo.pdf", "paper_pdf_sha256": "80a642fdbdb3d79517358dd3e16eb656b5c07c89c4520cfa8c90e12856072661", "paper_pdf_bytes": 8921132, "paper_pdf_source": "openreview", "code_url": "https://github.com/BayesWatch/nas-without-training", "code_repository": "BayesWatch/nas-without-training", "code_commit": "b3a82a6642564df115f989ff940ec6b8ef9ca9d3", "code_archive": "repos/g4E6SAAvACo.zip", "code_archive_sha256": "0c1fabbe716170a08a10c9cc9480e0349062c5155c222f92a404b0a72db0e2ae", "code_archive_bytes": 358568, "code_file_count": 120, "code_extensions": {".py": 119, ".sh": 1}, "github_disk_usage_kb": 1188, "github_languages": {"Python": 767072, "Shell": 3774}, "github_archived": false, "github_pushed_at": "2021-08-06T15:44:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-architecture-search-without-training"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HkeQ6ANYDB", "year": 2020, "status": "rejected", "title": "Blending Diverse Physical Priors with Neural Networks", "authors": ["Yunhao Ba", "Guangyuan Zhao", "Achuta Kadambi"], "authorids": ["yhba@ucla.edu", "zhaoguangyuan@ucla.edu", "achuta@ee.ucla.edu"], "authors_source": "OpenReview API", "abstract": "Rethinking physics in the era of deep learning is an increasingly important topic. This topic is special because, in addition to data, one can leverage a vast library of physical prior models (e.g. kinematics, fluid flow, etc) to perform more robust inference. The nascent sub-field of physics-based learning (PBL) studies this problem of blending neural networks with physical priors. While previous PBL algorithms have been applied successfully to specific tasks, it is hard to generalize existing PBL methods to a wide range of physics-based problems. Such generalization would require an architecture that can adapt to variations in the correctness of the physics, or in the quality of training data. No such architecture exists. In this paper, we aim to generalize PBL, by making a first attempt to bring neural architecture search (NAS) to the realm of PBL. We introduce a new method known as physics-based neural architecture search (PhysicsNAS) that is a top-performer across a diverse range of quality in the physical model and the dataset. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "rklrUUyI5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1388/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors apply neural architecture search techniques to the problem of physics based learning. It is interesting because it cleverly tackles the challenge of manually designing priors and network architectures. The results are also impressive as the proposed method surpasses all the considered baselines. Despite of the above upsides, I have the following questions/concerns.\n1. There is limited technical novelty as the entire method is mainly based on previous work on neural architecture search. Nevertheless, it might be helpful to have some ablation study to show the improvement of the task-specific adaptations presented in the paper, with which I believe this could be a good paper on the application side.\n2. I'm curious about the performance of the baseline methods given the same amount of computation. For example, is it possible to perform intensive hyperparameter tuning for the baselines to also obtain improvement. It seems that the authors did not discuss the computational costs and whether different methods are compared given the same cost.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The authors apply neural architecture search techniques to the problem of physics based learning. It is interesting because it cleverly tackles the challenge of manually designing priors and network architectures. The results are also impressive as the proposed method surpasses all the considered baselines. Despite of the above upsides, I have the following questions/concerns.\n1. There is limited technical novelty as the entire method is mainly based on previous work on neural architecture search. Nevertheless, it might be helpful to have some ablation study to show the improvement of the task-specific adaptations presented in the paper, with which I believe this could be a good paper on the application side.\n2. I'm curious about the performance of the baseline methods given the same amount of computation. For example, is it possible to perform intensive hyperparameter tuning for the baselines to also obtain improvement. It seems that the authors did not discuss the computational costs and whether different methods are compared given the same cost."}, "tcdate": 1572365901286}, {"id": "HkxVmBAy9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1388/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes PhysicsNAS, which proposes a method to automatically design architectures that\nincorporate domain expertise from phyiscs-based models while also accounting for potential\nmismatch between the model and real world due to unaccountable factors. While existing work seem\nto incorporate such information via one of 4 standard ways (given on page 2), the proposed work\nattempts to meld them so as to find the optimal combination for the problem at hand and the data\navailable.\n\nWhile I don't see anything fundamentally wrong with the paper, I do not feel that the technical\ncontributions are substantial enough to warrant acceptance at ICLR.\nMore specifically, the methodological novelty is limited and the experimental evaluation only\nevaluates the method on two fairly simple problems. \n\nOn a positive note, the authors have done a good job of illustrating the idea and have compared it\nto most natural baselines. I also thought that the illustrations of the architectures found\nfor different sample sizes (in the Appendix) quite insightful.\n\nI encourage the authors to pursue this line of work, but test this on more complex prediction tasks\nwhere entirely model-based approaches are unreliable, and entirely black-box estimators are sample\ninefficient. It also seems that the approach need not be confined to physics per se - in many\nproblems in chemistry, materials science etc. scientists are looking for ways to incorporate domain\nexpertise while accounting for model-mismatch. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I do not know much about this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes PhysicsNAS, which proposes a method to automatically design architectures that\nincorporate domain expertise from phyiscs-based models while also accounting for potential\nmismatch between the model and real world due to unaccountable factors. While existing work seem\nto incorporate such information via one of 4 standard ways (given on page 2), the proposed work\nattempts to meld them so as to find the optimal combination for the problem at hand and the data\navailable.\n\nWhile I don't see anything fundamentally wrong with the paper, I do not feel that the technical\ncontributions are substantial enough to warrant acceptance at ICLR.\nMore specifically, the methodological novelty is limited and the experimental evaluation only\nevaluates the method on two fairly simple problems. \n\nOn a positive note, the authors have done a good job of illustrating the idea and have compared it\nto most natural baselines. I also thought that the illustrations of the architectures found\nfor different sample sizes (in the Appendix) quite insightful.\n\nI encourage the authors to pursue this line of work, but test this on more complex prediction tasks\nwhere entirely model-based approaches are unreliable, and entirely black-box estimators are sample\ninefficient. It also seems that the approach need not be confined to physics per se - in many\nproblems in chemistry, materials science etc. scientists are looking for ways to incorporate domain\nexpertise while accounting for model-mismatch. \n"}, "tcdate": 1571968283890}, {"id": "H1lojMTjKH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1388/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "=== Overall comments ===\nThis paper proposes to generalize approaches to physics-based learning (PBL) by performing network architecture search (NAS) over elements from PBL models found in the literature. This entails including physical inputs to the network and the incorporation of new operations to the NAS. I think the idea has merit and rather like it. However, there are several aspects of the work that could be improved. The technical novelty is small, as the extension of the  existing NAS models to handle physical inputs and a few new operators is relatively straightforward. The experiments, while well designed, only explore uninteresting toy problems. While I appreciate the necessity to explore the methods performance in a more controlled setting, a more impactful testbed would be more convincing. Another drawback of the evaluation is the lack of a proper statistical analysis of the results, given the small data and model sizes.\n\n\n=== Relevance & Prior Work ===\n+ The related work gives a good summary and categorization of prior work in physics-based learning\n+ The problem (physics-based learning) is interesting and relevant to the community\n\n\n=== Novelty & Approach===\n+ application of NAS to physics based learning\n+ incorporation of physics solutions as inputs into differentiable NAS\n+ creation of physics-informed operation sets to merge physical models into network\n- technical steps to merge NAS and PBL are relatively straightforward\n\n\n=== Evaluation ===\nTwo representative physical simulations were chosen for evaluation, where elements of the physics model are intentionally omitted,  1) estimating trajectory of a ball in presence of wind and air resistance, and 2) a collision speed simulation where two objects collide, where sliding friction is not accounted for in the physics model.\n\nThe baselines consist of: a 3-layer MLP (data-driven), a 3-layer MLP with Physical Regularization, a 3-layer MLP with residual connection to the physics prediction, an MLP with two input branches, on for the data and one for the physics predictions (Physical Fusion), and the Embedded Physics model which estimates parameters for the physics modelu using a 3-layer MLP.\n\nPhysicsNAS can combine elements of the baseline models, but the total number of nodes is limited to 5.\n\n+ Experiments testing the dependence of the model on the numbers of samples and the strength of the physical inconsistencies were conducted. In both cases, PhysicsNAS outperformed the best specialized physics models.\n\n- The chosen testbed tasks are toy problems. While these types of experiments are necessary to understand the performance of the model, it would have been interesting to see PhysicsNAS applied to a more impactful task\n\n- Given the size of the networks and the training data, there is no reason why a more sophisticated statistical analysis of the results wasn’t performed (confidence intervals, t-test, p-value). Similarly, a more complete set of experiments with more sample amounts could be provided with little effort.\n\n\n=== Clarity & Other Comments ===\n- “precious nodes” -> previous nodes\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I do not know much about this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "=== Overall comments ===\nThis paper proposes to generalize approaches to physics-based learning (PBL) by performing network architecture search (NAS) over elements from PBL models found in the literature. This entails including physical inputs to the network and the incorporation of new operations to the NAS. I think the idea has merit and rather like it. However, there are several aspects of the work that could be improved. The technical novelty is small, as the extension of the  existing NAS models to handle physical inputs and a few new operators is relatively straightforward. The experiments, while well designed, only explore uninteresting toy problems. While I appreciate the necessity to explore the methods performance in a more controlled setting, a more impactful testbed would be more convincing. Another drawback of the evaluation is the lack of a proper statistical analysis of the results, given the small data and model sizes.\n\n\n=== Relevance & Prior Work ===\n+ The related work gives a good summary and categorization of prior work in physics-based learning\n+ The problem (physics-based learning) is interesting and relevant to the community\n\n\n=== Novelty & Approach===\n+ application of NAS to physics based learning\n+ incorporation of physics solutions as inputs into differentiable NAS\n+ creation of physics-informed operation sets to merge physical models into network\n- technical steps to merge NAS and PBL are relatively straightforward\n\n\n=== Evaluation ===\nTwo representative physical simulations were chosen for evaluation, where elements of the physics model are intentionally omitted,  1) estimating trajectory of a ball in presence of wind and air resistance, and 2) a collision speed simulation where two objects collide, where sliding friction is not accounted for in the physics model.\n\nThe baselines consist of: a 3-layer MLP (data-driven), a 3-layer MLP with Physical Regularization, a 3-layer MLP with residual connection to the physics prediction, an MLP with two input branches, on for the data and one for the physics predictions (Physical Fusion), and the Embedded Physics model which estimates parameters for the physics modelu using a 3-layer MLP.\n\nPhysicsNAS can combine elements of the baseline models, but the total number of nodes is limited to 5.\n\n+ Experiments testing the dependence of the model on the numbers of samples and the strength of the physical inconsistencies were conducted. In both cases, PhysicsNAS outperformed the best specialized physics models.\n\n- The chosen testbed tasks are toy problems. While these types of experiments are necessary to understand the performance of the model, it would have been interesting to see PhysicsNAS applied to a more impactful task\n\n- Given the size of the networks and the training data, there is no reason why a more sophisticated statistical analysis of the results wasn’t performed (confidence intervals, t-test, p-value). Similarly, a more complete set of experiments with more sample amounts could be provided with little effort.\n\n\n=== Clarity & Other Comments ===\n- “precious nodes” -> previous nodes\n"}, "tcdate": 1571701411246}], "openreview_url": "https://openreview.net/forum?id=HkeQ6ANYDB", "arxiv_id": "1910.00201", "paper_pdf": "papers/HkeQ6ANYDB.pdf", "paper_pdf_sha256": "b4c5b4d33d20f6db3af248d3570108e5f4760551775eb042dc7dc5af51c45453", "paper_pdf_bytes": 1239374, "paper_pdf_source": "openreview", "code_url": "https://github.com/PhysicsNAS/PhysicsNAS", "code_repository": "PhysicsNAS/PhysicsNAS", "code_commit": "9f9690ce064d1eca4a1f3f89cc761c484e7e594a", "code_archive": "repos/HkeQ6ANYDB.zip", "code_archive_sha256": "de4043c7a31fba559c4d156f770980bc632a7884d2acd033079756998dd71447", "code_archive_bytes": 4778446, "code_file_count": 24, "code_extensions": {".ipynb": 14, ".py": 10}, "github_disk_usage_kb": 5427, "github_languages": {"Jupyter Notebook": 24849050, "Python": 20662}, "github_archived": false, "github_pushed_at": "2019-10-01T04:15:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/blending-diverse-physical-priors-with-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5K9XW9gQ5r", "year": 2026, "status": "rejected", "title": "UniRL: Self-Improving Unified Multimodal Models via Supervised and Reinforcement Learning", "authors": ["Weijia Mao", "Zhenheng Yang", "Mike Zheng Shou"], "authorids": ["~Weijia_Mao1", "~Zhenheng_Yang3", "~Mike_Zheng_Shou1"], "authors_source": "OpenReview API", "abstract": "Unified multimodal large language models such as Show-o and Janus have achieved strong performance across both generation and understanding tasks. However, these models typically rely on large-scale datasets and require substantial computation during the pretraining stage. In addition, several post-training methods have been proposed, but they often depend on external data or are limited to task-specific customization. In this work, we introduce UniRL, a self-improving post-training approach. Our approach enables the model to generate images from prompts and use them as training data in each iteration, without relying on any external image data. Moreover, it enables the two tasks to enhance each other: the generated images are used for understanding, and the understanding results are used to supervise generation. We explore supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) to optimize the models. UniRL offers three key advantages: (1) it requires no external image data, as all training samples are generated by the model itself during training; (2) it not only improves individual task performance, but also reduces the imbalance between generation and understanding; and (3) it requires only several additional training steps during the post-training stage. We evaluate UniRL on top of Show-o and Janus, achieving a GenEval score of 0.77 for Show-o and 0.65 for Janus. Code and models will be released.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "51hjweDVGJ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6064/Reviewer_2jf8"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors introduce UniRL, a self-improving post-training framework for unified multimodal large language models (u-MLLMs) that perform both text-to-image generation (T2I) and multimodal understanding (MMU). In this way, the model is updated by the its self-generated data: the model generates images as well as answers questions about these images, and then optimizes itself using either SFT or GRPO.", "review_text": "The authors introduce UniRL, a self-improving post-training framework for unified multimodal large language models (u-MLLMs) that perform both text-to-image generation (T2I) and multimodal understanding (MMU). In this way, the model is updated by the its self-generated data: the model generates images as well as answers questions about these images, and then optimizes itself using either SFT or GRPO.", "strengths": "- The paper identifies a real gap in current u-MLLMs, i.e., the lack of mutual improvement between generation and understanding, and argues convincingly that post-training is an efficient place to address it.\n- It is elegant and practically attractive for models to bootstrap its own samples. \n- Text is easy to follow.", "weaknesses": "- Limited novelty: The overall pipeline is a well-known paradigm in self-training and reinforcement learning.\n- Narrow scope: The method is validated only on GenEval-like attributes, which represent low-level visual features. It remains unclear whether the approach generalizes to more complex open-domain tasks (e.g., human actions, reasoning, visual grounding, or captioning).\n- The automatic reward relies on keyword or attribute matching. Without semantic or visual similarity measures (e.g., CLIP-based scoring or GPT-based judging), the reinforcement signal is weak and may not scale.\n- Although UniRL improves GenEval and balance metrics, the absolute gains are moderate, and the experiments are limited to small model variants. The method does not appear to change the frontier of multimodal learning.", "questions": "Please refer to the weaknesses.\n\n- How sensitive is performance to the reward function design and group size K in GRPO?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce UniRL, a self-improving post-training framework for unified multimodal large language models (u-MLLMs) that perform both text-to-image generation (T2I) and multimodal understanding (MMU). In this way, the model is updated by the its self-generated data: the model generates images as well as answers questions about these images, and then optimizes itself using either SFT or GRPO.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper identifies a real gap in current u-MLLMs, i.e., the lack of mutual improvement between generation and understanding, and argues convincingly that post-training is an efficient place to address it.\n- It is elegant and practically attractive for models to bootstrap its own samples. \n- Text is easy to follow.", "weaknesses": "- Limited novelty: The overall pipeline is a well-known paradigm in self-training and reinforcement learning.\n- Narrow scope: The method is validated only on GenEval-like attributes, which represent low-level visual features. It remains unclear whether the approach generalizes to more complex open-domain tasks (e.g., human actions, reasoning, visual grounding, or captioning).\n- The automatic reward relies on keyword or attribute matching. Without semantic or visual similarity measures (e.g., CLIP-based scoring or GPT-based judging), the reinforcement signal is weak and may not scale.\n- Although UniRL improves GenEval and balance metrics, the absolute gains are moderate, and the experiments are limited to small model variants. The method does not appear to change the frontier of multimodal learning.", "questions": "Please refer to the weaknesses.\n\n- How sensitive is performance to the reward function design and group size K in GRPO?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762055344287}, {"id": "9dhSfDcCju", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6064/Reviewer_LME2"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes a self-improving post-training method (UniRL) for unified MLLMs (u-MLLMs). UniRL consists of both SFT and RL stages, and enables u-MLLMs to generate their own synthetic training data without relying on external datasets. Moreover, the paper introduces a new metric to quantify task imbalance. Experiments on Show-o and Janus demonstrate performance improvements in both tasks over baseline u-MLLMs.", "review_text": "This paper proposes a self-improving post-training method (UniRL) for unified MLLMs (u-MLLMs). UniRL consists of both SFT and RL stages, and enables u-MLLMs to generate their own synthetic training data without relying on external datasets. Moreover, the paper introduces a new metric to quantify task imbalance. Experiments on Show-o and Janus demonstrate performance improvements in both tasks over baseline u-MLLMs.", "strengths": "- **No external training data requirement:** The proposed UniRL uses only model-generated images for post-training, making it more practical and scalable, eliminating the need for expensive real-world data collection and annotation. \n\n- **Valuable consistency evaluation metric:** The proposed bidirectional metrics, Accuracy(MMU|T2I) and Accuracy(T2I|MMU), is a useful and well-conceived to quantify the alignment between generation and understanding tasks. \n\n- **Insightful analysis:** This paper provides clear observations explaining why GRPO outperforms SFT for generalization while SFT excels for T2I in end-to-end settings.", "weaknesses": "- **Synthetic data reliability and quality control:** While self-generated training data of u-MLLMs is data-friendly, the quality and diversity of generated synthetic data can not be guaranteed. Without guarantees of generated data, the self-training loop risks reinforcing generation biases or error accumulation over iterations. How do you ensure that the self-generated images remain sufficiently diverse and informative across iterations? Is there any mechanism to prevent the accumulation of low-quality or mode-collapsed samples?\n\n- **Limited evaluation and comparison:** The experiments primarily focus on comparisons against baselines, lacking more comprehensive evaluations with leading u-MLLMs of comparable parameter scales and other GRPO-trained u-MLLMs.\n\n- **Reward function design is too empirical:** The reward rules are hand-crafted keyword matching. It lacks universality and may introduce bias toward specific phrasings or synonyms.", "questions": "1. How does UniRL's performance scale with model size?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a self-improving post-training method (UniRL) for unified MLLMs (u-MLLMs). UniRL consists of both SFT and RL stages, and enables u-MLLMs to generate their own synthetic training data without relying on external datasets. Moreover, the paper introduces a new metric to quantify task imbalance. Experiments on Show-o and Janus demonstrate performance improvements in both tasks over baseline u-MLLMs.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- **No external training data requirement:** The proposed UniRL uses only model-generated images for post-training, making it more practical and scalable, eliminating the need for expensive real-world data collection and annotation. \n\n- **Valuable consistency evaluation metric:** The proposed bidirectional metrics, Accuracy(MMU|T2I) and Accuracy(T2I|MMU), is a useful and well-conceived to quantify the alignment between generation and understanding tasks. \n\n- **Insightful analysis:** This paper provides clear observations explaining why GRPO outperforms SFT for generalization while SFT excels for T2I in end-to-end settings.", "weaknesses": "- **Synthetic data reliability and quality control:** While self-generated training data of u-MLLMs is data-friendly, the quality and diversity of generated synthetic data can not be guaranteed. Without guarantees of generated data, the self-training loop risks reinforcing generation biases or error accumulation over iterations. How do you ensure that the self-generated images remain sufficiently diverse and informative across iterations? Is there any mechanism to prevent the accumulation of low-quality or mode-collapsed samples?\n\n- **Limited evaluation and comparison:** The experiments primarily focus on comparisons against baselines, lacking more comprehensive evaluations with leading u-MLLMs of comparable parameter scales and other GRPO-trained u-MLLMs.\n\n- **Reward function design is too empirical:** The reward rules are hand-crafted keyword matching. It lacks universality and may introduce bias toward specific phrasings or synonyms.", "questions": "1. How does UniRL's performance scale with model size?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761979004270}, {"id": "FSdnzPwjp2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6064/Reviewer_hxCv"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "UniRL introduces a self-improving post-training framework for unified multimodal models (e.g., Show-o, Janus) that jointly optimizes text-to-image (T2I) generation and multimodal understanding (MMU) using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO). The core innovation is a closed-loop pipeline where the model generates images from prompts, answers questions about them, and uses the feedback to improve both tasks without external data. Besides, a bidirectional evaluation metric (Accuracy_MMU|T2I, Accuracy_T2I|MMU)  is introduced to quantifies task alignment.", "review_text": "UniRL introduces a self-improving post-training framework for unified multimodal models (e.g., Show-o, Janus) that jointly optimizes text-to-image (T2I) generation and multimodal understanding (MMU) using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO). The core innovation is a closed-loop pipeline where the model generates images from prompts, answers questions about them, and uses the feedback to improve both tasks without external data. Besides, a bidirectional evaluation metric (Accuracy_MMU|T2I, Accuracy_T2I|MMU)  is introduced to quantifies task alignment.", "strengths": "1. The proposed self-improving framework is technically sound. By using generated data for iterative training, no external data is required and the mutual reinforcement of T2I and MMU tasks addresses the generation-understanding imbalance pervasive in unified models.\n2. The integration of SFT and GRPO  pave the way for further performance improvement in the context of unified models.\n3. The overall presentation is clear and easy to follow.\n4. The analysis of SFT v.s. GRPO provide valuable insights.", "weaknesses": "1. In table 2 and table 4, the performance of Count. underperforms than that of the baseline model. However, the prompts construction and evaluation pipeline  in figure 1 and figure 2  show that the Count. performance could be improved with the proposed framework. Why the performance is worser? More evaluation details should be included to clarify this.\n2. The evaluation only foucuses on basic visual attributes (counting, color, position) but neglects complex reasoning (e.g., physics, temporal dynamics). Accordingly, the performance improvement on GenEval could be \"hacked\" by the constructed prompt suit for iterative self-improving. In other words, the performance gains might be attribute to the iterative optimization on the isomorphic prompt set instead of actual improvement. Are there any ood results to better reflect the effectiveness of the proposed method?\n3. More recent unified models should be involved for evaluation and comparison such as BAGEL, SHOW-O2, BLIP-3O, etc.\n4. More evaluation on multimodal understanding benchmarks such as MMMU, GQA, MME, etc. , should be included, current results in table 3 only provide limited evaluation scope for the author's claim on improving generation and understanding simutaneously.", "questions": "I would consider raise my score if authors could address my concerns, especially on the evaluation results and details.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "UniRL introduces a self-improving post-training framework for unified multimodal models (e.g., Show-o, Janus) that jointly optimizes text-to-image (T2I) generation and multimodal understanding (MMU) using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO). The core innovation is a closed-loop pipeline where the model generates images from prompts, answers questions about them, and uses the feedback to improve both tasks without external data. Besides, a bidirectional evaluation metric (Accuracy_MMU|T2I, Accuracy_T2I|MMU)  is introduced to quantifies task alignment.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The proposed self-improving framework is technically sound. By using generated data for iterative training, no external data is required and the mutual reinforcement of T2I and MMU tasks addresses the generation-understanding imbalance pervasive in unified models.\n2. The integration of SFT and GRPO  pave the way for further performance improvement in the context of unified models.\n3. The overall presentation is clear and easy to follow.\n4. The analysis of SFT v.s. GRPO provide valuable insights.", "weaknesses": "1. In table 2 and table 4, the performance of Count. underperforms than that of the baseline model. However, the prompts construction and evaluation pipeline  in figure 1 and figure 2  show that the Count. performance could be improved with the proposed framework. Why the performance is worser? More evaluation details should be included to clarify this.\n2. The evaluation only foucuses on basic visual attributes (counting, color, position) but neglects complex reasoning (e.g., physics, temporal dynamics). Accordingly, the performance improvement on GenEval could be \"hacked\" by the constructed prompt suit for iterative self-improving. In other words, the performance gains might be attribute to the iterative optimization on the isomorphic prompt set instead of actual improvement. Are there any ood results to better reflect the effectiveness of the proposed method?\n3. More recent unified models should be involved for evaluation and comparison such as BAGEL, SHOW-O2, BLIP-3O, etc.\n4. More evaluation on multimodal understanding benchmarks such as MMMU, GQA, MME, etc. , should be included, current results in table 3 only provide limited evaluation scope for the author's claim on improving generation and understanding simutaneously.", "questions": "I would consider raise my score if authors could address my concerns, especially on the evaluation results and details.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "I would consider raise my score if authors could address my concerns, especially on the evaluation results and details.", "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761908017914}, {"id": "rfNNVETYwD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission6064/Reviewer_Xvrj"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes UniRL, a self-improving post-training framework for unified multimodal large language models (e.g., Show-o, Janus) to address two key limitations of existing methods: heavy reliance on external image data and persistent imbalance between text-to-image generation (T2I) and multimodal understanding (MMU) tasks. UniRL operates without external image data by leveraging the model’s own generated images as training data in each iteration. It enables mutual enhancement between T2I and MMU-generated images are used to train understanding, while understanding results (e.g., answer accuracy) supervise generation.", "review_text": "This paper proposes UniRL, a self-improving post-training framework for unified multimodal large language models (e.g., Show-o, Janus) to address two key limitations of existing methods: heavy reliance on external image data and persistent imbalance between text-to-image generation (T2I) and multimodal understanding (MMU) tasks. UniRL operates without external image data by leveraging the model’s own generated images as training data in each iteration. It enables mutual enhancement between T2I and MMU-generated images are used to train understanding, while understanding results (e.g., answer accuracy) supervise generation.", "strengths": "1. UniRL eliminates the need for external image data by using the model’s real-time generated images as training samples. This addresses a critical pain point of existing post-training methods (which often rely on large external datasets like JourneyDB) and reduces data acquisition costs, making it more scalable.\n\n2. Unlike methods that optimize T2I or MMU in isolation, UniRL designs a closed-loop system where T2I and MMU reinforce each other. This not only improves individual task performance (e.g., Show-o’s GenEval score rises from 0.60 to 0.77 via SFT) but also explicitly reduces task imbalance—a long-overlooked issue in unified multimodal models.\n\n3. The proposed bidirectional metric Accuracy_MMU|T2I and Accuracy_T2I|MMU fills a gap in existing evaluations, which often focus on isolated task performance. This metric quantifies how well T2I and MMU align (e.g., whether a correctly generated image leads to a correct answer), providing a more holistic measure of unified model performance.", "weaknesses": "1. UniRL only targets basic visual attributes (counting, color, position) and ignores complex multimodal tasks (e.g., mathematical reasoning, abstract scene understanding, or object interaction inference). This restricts its applicability to real-world scenarios where models need to handle diverse, high-level tasks.\n\n2. The credit assignment from MMU rewards to T2I in the decoupled setting is described at a high level; more details on stability tricks, baseline comparisons, and failure cases for SFT are needed.", "questions": "1. In Table 4, UniRL (SFT) AccuracyMMU|T2I for “Single.” is 0.11. Is this a typo or a real phenomenon? If real, what causes such a stark collapse limited to this subtask? Similarly, Table 3 shows UniRL (SFT) “Single” = 0.08 for MMU. Could you share per-class confusion, prompt/QA distributions, and seed variance to rule out evaluation or data-pipeline issues?\n\n2. To address general capability decline, you propose \"incorporating a broader set of understanding categories\" and \"scaling up the model size.\" For the former, what specific new categories (e.g., math word problems, logical syllogisms) would you prioritize, and how would you design prompts/QA pairs for them to ensure alignment with both T2I and MMU? For the latter, do you have preliminary results or projections on how model scaling (e.g., increasing parameter count from Show-o’s base size) would impact text-based reasoning and numerical calculation performance? Would scaling alone be sufficient, or does it require paired adjustments to the training framework?\n\n3. MMMU’s strength lies in cross-discipline reasoning (e.g., combining visual cues with text-based math). Your current T2I-MMU mutual enhancement relies on basic visual attributes—how would you modify UniRL to enable the model to generate images that support text-based reasoning tasks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes UniRL, a self-improving post-training framework for unified multimodal large language models (e.g., Show-o, Janus) to address two key limitations of existing methods: heavy reliance on external image data and persistent imbalance between text-to-image generation (T2I) and multimodal understanding (MMU) tasks. UniRL operates without external image data by leveraging the model’s own generated images as training data in each iteration. It enables mutual enhancement between T2I and MMU-generated images are used to train understanding, while understanding results (e.g., answer accuracy) supervise generation.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. UniRL eliminates the need for external image data by using the model’s real-time generated images as training samples. This addresses a critical pain point of existing post-training methods (which often rely on large external datasets like JourneyDB) and reduces data acquisition costs, making it more scalable.\n\n2. Unlike methods that optimize T2I or MMU in isolation, UniRL designs a closed-loop system where T2I and MMU reinforce each other. This not only improves individual task performance (e.g., Show-o’s GenEval score rises from 0.60 to 0.77 via SFT) but also explicitly reduces task imbalance—a long-overlooked issue in unified multimodal models.\n\n3. The proposed bidirectional metric Accuracy_MMU|T2I and Accuracy_T2I|MMU fills a gap in existing evaluations, which often focus on isolated task performance. This metric quantifies how well T2I and MMU align (e.g., whether a correctly generated image leads to a correct answer), providing a more holistic measure of unified model performance.", "weaknesses": "1. UniRL only targets basic visual attributes (counting, color, position) and ignores complex multimodal tasks (e.g., mathematical reasoning, abstract scene understanding, or object interaction inference). This restricts its applicability to real-world scenarios where models need to handle diverse, high-level tasks.\n\n2. The credit assignment from MMU rewards to T2I in the decoupled setting is described at a high level; more details on stability tricks, baseline comparisons, and failure cases for SFT are needed.", "questions": "1. In Table 4, UniRL (SFT) AccuracyMMU|T2I for “Single.” is 0.11. Is this a typo or a real phenomenon? If real, what causes such a stark collapse limited to this subtask? Similarly, Table 3 shows UniRL (SFT) “Single” = 0.08 for MMU. Could you share per-class confusion, prompt/QA distributions, and seed variance to rule out evaluation or data-pipeline issues?\n\n2. To address general capability decline, you propose \"incorporating a broader set of understanding categories\" and \"scaling up the model size.\" For the former, what specific new categories (e.g., math word problems, logical syllogisms) would you prioritize, and how would you design prompts/QA pairs for them to ensure alignment with both T2I and MMU? For the latter, do you have preliminary results or projections on how model scaling (e.g., increasing parameter count from Show-o’s base size) would impact text-based reasoning and numerical calculation performance? Would scaling alone be sufficient, or does it require paired adjustments to the training framework?\n\n3. MMMU’s strength lies in cross-discipline reasoning (e.g., combining visual cues with text-based math). Your current T2I-MMU mutual enhancement relies on basic visual attributes—how would you modify UniRL to enable the model to generate images that support text-based reasoning tasks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761902534375}], "openreview_url": "https://openreview.net/forum?id=5K9XW9gQ5r", "arxiv_id": "2505.23380", "paper_pdf": "papers/5K9XW9gQ5r.pdf", "paper_pdf_sha256": "fd5aa5544b49e60416dd95ddc5431a569cfab78fab503917764fe49ed7a4b16f", "paper_pdf_bytes": 1377316, "paper_pdf_source": "openreview", "code_url": "https://github.com/showlab/UniRL", "code_repository": "showlab/UniRL", "code_commit": "66ee4843cabb311832e5b36d530b6d70273359f6", "code_archive": "repos/5K9XW9gQ5r.zip", "code_archive_sha256": "f44aecda1a5c007b64c555de186e768695927df85a7728afd9767aded6a4ab50", "code_archive_bytes": 3764255, "code_file_count": 40, "code_extensions": {".py": 34, ".sh": 6}, "github_disk_usage_kb": 3631, "github_languages": {"Python": 479251, "Shell": 22939}, "github_archived": false, "github_pushed_at": "2025-05-30T01:47:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/unirl-self-improving-unified-multimodal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XW4Xnx0xlH", "year": 2025, "status": "rejected", "title": "Second-Order Forward-Mode Automatic Differentiation for Optimization", "authors": ["Adam D. Cobb", "Atilim Gunes Baydin", "Barak A. Pearlmutter", "Susmit Jha"], "authorids": ["~Adam_D._Cobb1", "~Atilim_Gunes_Baydin1", "~Barak_A._Pearlmutter1", "~Susmit_Jha1"], "authors_source": "OpenReview API", "abstract": "Forward gradient methods offer a promising alternative to backpropagation. Optimization that only requires forward passes could simplify hardware implementation, improve parallelism, lower memory cost, and allow for more biologically plausible learning models. This has motivated recent forward-mode automated differentiation (AD) methods. This paper presents a novel second-order forward-mode AD method for optimization that generalizes a second-order line search to a $K$-dimensional hyperplane. Unlike recent work that relies on directional derivatives (or Jacobian–Vector Products, JVPs), we use hyper-dual numbers to jointly evaluate both directional derivatives and their second-order quadratic terms. As a result, we introduce forward-mode weight perturbation with Hessian information for K-dimensional hyper-plane search (FoMoH-$K$D). We derive the convergence properties of FoMoH-$K$D and show how it generalizes to Newton’s method for $K = D$. We demonstrate this generalization empirically, and compare the performance of FoMoH-$K$D to forward gradient descent (FGD) on three case studies: Rosenbrock function used widely for evaluating optimization methods, logistic regression with 7,850 parameters, and learning a CNN classifier with 431,080 parameters. Our experiments show that FoMoH-$K$D not only achieves better performance and accuracy, but also converges faster, thus, empirically verifying our theoretical results.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "nJ8JsO0YGe", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission443/Reviewer_VLTw"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The authors present a new optimization method relying on forward mode automatic differentiation (AD). Namely, the authors propose to use second order directional derivatives computed along random directions to precondition stochastic estimates of the gradients obtained by forward mode automatic differentiation along these same random directions. The proposed Forward Mode Second-order Hyperplane Search (FoMoH) interpolates between using an approximate Cauchy stepsize (that is a Cauchy stepsize computed with a quadratic approximation of the objective) along a random direction and an approximate Newton step. The proposed method is shown to outperform a standard forward gradient descent on the Rosenbrock function, a logistic regression problem, and the MNIST image classification task with a CNN.", "review_text": "The authors present a new optimization method relying on forward mode automatic differentiation (AD). Namely, the authors propose to use second order directional derivatives computed along random directions to precondition stochastic estimates of the gradients obtained by forward mode automatic differentiation along these same random directions. The proposed Forward Mode Second-order Hyperplane Search (FoMoH) interpolates between using an approximate Cauchy stepsize (that is a Cauchy stepsize computed with a quadratic approximation of the objective) along a random direction and an approximate Newton step. The proposed method is shown to outperform a standard forward gradient descent on the Rosenbrock function, a logistic regression problem, and the MNIST image classification task with a CNN.", "strengths": "- The idea of exploiting second order directional derivatives is original and could be further explored.\n- The proposed method shows clearly superior performance than a simple forward gradient descent.", "weaknesses": "Unfortunately the paper does not give justice to the potential of the main idea.\nImprovements are necessary and possible:\n- The method does not require introducing dual numbers. Computing second order directional derivatives can easily be done with nested forward mode autodiff:\n```\nimport jax\n\ndef hqp(fun, w, v1, v2):\n  def dir_der_v1(w):\n    return jax.jvp(fun, (w,), (v1,))[1]\n  return jax.jvp(dir_der_v1, (w,), (v2,))[1]\n```\nThe above may be slightly slower than the implementation with dual numbers (the difference is probably minimal) but it is also much simpler to implement than having to adopt a new kind of automatic differentiation library. If one want the best possible implementation, then Taylor-mode automatic differentiation can also be used but again the benefits are really minor. Algorithm 1 is then unnecessarily complicated when it could be written in just K calls to the above function to define the approximate Hessian. Presenting the algorithm with a simple implementation that any user could recode in at most 50 lines of code in jax or pytorch would greatly improve the potential adoption of the method.\n- Unfortunately, the proofs are not rigorous, nor are the claims.\n  - Theorem 1 is a corollary of Theorem 2 so no need to present it. Also results like $\\lim_{t \\rightarrow +\\infty} \\theta_t = \\theta^*$ are meaningless: we don't want to wait the time of the universe to see convergence. Rates like the ones provided in Theorem 2 are relevant.\n  - In all claims, detail the setting: what algorithm is used, what is theta_t, what is the expectation taken against etc... You may not do that in the main text by lack of space but at least make sure that the appendix contains a result that details all assumptions.\n  - The proof of theorem 2 is unfortunately not well detailed:\n    - Please give a detailed proof that $\\tilde \\theta_{t+1} = \\tilde \\theta_t + P(\\tilde \\theta^* - \\tilde \\theta_t)$. We are dealing with quadratics so the proof should boil down to simple linear algebra. Avoid intuitive arguments, just write the equations one by one showing the results. I personally will refuse this paper to be accepted without detailed proofs.\n    - Give a proper reference for the fact that the expectation of a projection matrix defined from Gaussian variables is the scaled identity\n    - First line of last set of equations of the proof of theorem 2 should read $\\tilde \\theta_{t+1} - \\tilde \\theta^* = (1- K/D)(\\tilde \\theta_t - \\tilde \\theta^*)$\n    - Please be rigorous when you write expectations. You need to detail each time with respect to which randomness you are taking the expectation. For example at one point you write $\\tilde \\theta_{t+1} = \\mathbb{E}[\\tilde \\theta_t + P(\\tilde \\theta^* - \\tilde \\theta_t)]$ but so then $\\tilde \\theta_{t+1}$ is not random. Unless you meant $E[\\tilde \\theta_{t+1}] = $. Such lack of rigor is detrimental to the potential of the idea.\n    - Section B.4 contains multiple errors:\n      - Reaching a critical point does not imply that you reach a minimum unless you make additional assumptions like convexity.\n      - Again be rigorous in the use of expectations, one usually use conditional expectations conditioned on the previous iterate for example.\n      - The provided rate is clearly not linear. Consider rereading in details the reference for example.\n- Second order methods may generally be very sensitive to the batch-size. It would be great to plot a sensitivity analysis of the method with the batch-size for e.g. a given learning rate.\n- Consider another dataset than MNIST. MNIST is well known to be particularly easy and may not reflect potential challenges that the method can have.", "questions": "- First it would be great to revise the proofs to make them rigorous.\n- Could you add a full mathematical definition of FoMoH-BP?\n- What is the logistic regression model? I suppose it is not a regression but a classification problem first? Then if you use 7850 parameters it's probably not a simple linear model but some form of Multi-Layer Perceptron?\n- Detail the CNN architecture used in the experiment.\n- In the abstract, you mention alternative (orthogonal to be exact) methods for forward gradients. They are never compared in the experiments. It would be great to have them.\n- By curiosity how can analog optical systems compute derivatives of intermediate functions to implement forward mode automatic differentiation? The reference provided by the authors does not mention AD, at least from its abstract.\n- You mention \"linesearch\" in line 243 but there is no linesearch at all in the algorithm. What do you mean by linesearch?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present a new optimization method relying on forward mode automatic differentiation (AD). Namely, the authors propose to use second order directional derivatives computed along random directions to precondition stochastic estimates of the gradients obtained by forward mode automatic differentiation along these same random directions. The proposed Forward Mode Second-order Hyperplane Search (FoMoH) interpolates between using an approximate Cauchy stepsize (that is a Cauchy stepsize computed with a quadratic approximation of the objective) along a random direction and an approximate Newton step. The proposed method is shown to outperform a standard forward gradient descent on the Rosenbrock function, a logistic regression problem, and the MNIST image classification task with a CNN.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- The idea of exploiting second order directional derivatives is original and could be further explored.\n- The proposed method shows clearly superior performance than a simple forward gradient descent.", "weaknesses": "Unfortunately the paper does not give justice to the potential of the main idea.\nImprovements are necessary and possible:\n- The method does not require introducing dual numbers. Computing second order directional derivatives can easily be done with nested forward mode autodiff:\n```\nimport jax\n\ndef hqp(fun, w, v1, v2):\n  def dir_der_v1(w):\n    return jax.jvp(fun, (w,), (v1,))[1]\n  return jax.jvp(dir_der_v1, (w,), (v2,))[1]\n```\nThe above may be slightly slower than the implementation with dual numbers (the difference is probably minimal) but it is also much simpler to implement than having to adopt a new kind of automatic differentiation library. If one want the best possible implementation, then Taylor-mode automatic differentiation can also be used but again the benefits are really minor. Algorithm 1 is then unnecessarily complicated when it could be written in just K calls to the above function to define the approximate Hessian. Presenting the algorithm with a simple implementation that any user could recode in at most 50 lines of code in jax or pytorch would greatly improve the potential adoption of the method.\n- Unfortunately, the proofs are not rigorous, nor are the claims.\n  - Theorem 1 is a corollary of Theorem 2 so no need to present it. Also results like $\\lim_{t \\rightarrow +\\infty} \\theta_t = \\theta^*$ are meaningless: we don't want to wait the time of the universe to see convergence. Rates like the ones provided in Theorem 2 are relevant.\n  - In all claims, detail the setting: what algorithm is used, what is theta_t, what is the expectation taken against etc... You may not do that in the main text by lack of space but at least make sure that the appendix contains a result that details all assumptions.\n  - The proof of theorem 2 is unfortunately not well detailed:\n    - Please give a detailed proof that $\\tilde \\theta_{t+1} = \\tilde \\theta_t + P(\\tilde \\theta^* - \\tilde \\theta_t)$. We are dealing with quadratics so the proof should boil down to simple linear algebra. Avoid intuitive arguments, just write the equations one by one showing the results. I personally will refuse this paper to be accepted without detailed proofs.\n    - Give a proper reference for the fact that the expectation of a projection matrix defined from Gaussian variables is the scaled identity\n    - First line of last set of equations of the proof of theorem 2 should read $\\tilde \\theta_{t+1} - \\tilde \\theta^* = (1- K/D)(\\tilde \\theta_t - \\tilde \\theta^*)$\n    - Please be rigorous when you write expectations. You need to detail each time with respect to which randomness you are taking the expectation. For example at one point you write $\\tilde \\theta_{t+1} = \\mathbb{E}[\\tilde \\theta_t + P(\\tilde \\theta^* - \\tilde \\theta_t)]$ but so then $\\tilde \\theta_{t+1}$ is not random. Unless you meant $E[\\tilde \\theta_{t+1}] = $. Such lack of rigor is detrimental to the potential of the idea.\n    - Section B.4 contains multiple errors:\n      - Reaching a critical point does not imply that you reach a minimum unless you make additional assumptions like convexity.\n      - Again be rigorous in the use of expectations, one usually use conditional expectations conditioned on the previous iterate for example.\n      - The provided rate is clearly not linear. Consider rereading in details the reference for example.\n- Second order methods may generally be very sensitive to the batch-size. It would be great to plot a sensitivity analysis of the method with the batch-size for e.g. a given learning rate.\n- Consider another dataset than MNIST. MNIST is well known to be particularly easy and may not reflect potential challenges that the method can have.", "questions": "- First it would be great to revise the proofs to make them rigorous.\n- Could you add a full mathematical definition of FoMoH-BP?\n- What is the logistic regression model? I suppose it is not a regression but a classification problem first? Then if you use 7850 parameters it's probably not a simple linear model but some form of Multi-Layer Perceptron?\n- Detail the CNN architecture used in the experiment.\n- In the abstract, you mention alternative (orthogonal to be exact) methods for forward gradients. They are never compared in the experiments. It would be great to have them.\n- By curiosity how can analog optical systems compute derivatives of intermediate functions to implement forward mode automatic differentiation? The reference provided by the authors does not mention AD, at least from its abstract.\n- You mention \"linesearch\" in line 243 but there is no linesearch at all in the algorithm. What do you mean by linesearch?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731202241554}, {"id": "9FNXRjbWH9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission443/Reviewer_yy3y"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "The paper presents a forward-mode-only optimization method that uses second order information on randomly sampled hyperplanes. \nThe method makes use of (K^2 + K)/2 calls to a \"double\" forward-mode AD, implemented with hyper-dual numbers, computing along the way the Hessian projected onto the drawn plane. The paper presents theoretical results on convex and quadratic functions that also relates it to the Newton method, alongside some empirical validation on a test function and two learning problems.", "review_text": "The paper presents a forward-mode-only optimization method that uses second order information on randomly sampled hyperplanes. \nThe method makes use of (K^2 + K)/2 calls to a \"double\" forward-mode AD, implemented with hyper-dual numbers, computing along the way the Hessian projected onto the drawn plane. The paper presents theoretical results on convex and quadratic functions that also relates it to the Newton method, alongside some empirical validation on a test function and two learning problems.", "strengths": "- The significance of developing a preforming optimization method that does not rely on reverse-mode differentiation is very high, and the proposed method seems to make a concrete step in this direction \n- The paper is mostly well written and easy to follow, the notation is clear (although a few passages could be made clearer, see below)\n- The work does a good job introducing the concept of dual and hyper-dual numbers, which I expect the community to benefit from \n-  The theoretical results clearly show the advantages of the proposed method over FGD, as well as the larger scale experiment.", "weaknesses": "- One main promise of forward mode differentiation is to vastly decrease the memory requirement for large models, however the paper does not empirically quantifies the advantage of FoMoH in this regard. It would be nice to include memory footprint comparisons (as well as perhaps a table summarizing the runtime and memory complexity of key algorithms)\n- I would have appreciated some larger scale experiment with transformer architectures, e.g. for fine-tuning LLMs, which could be an interesting application of the proposed method \n- Some details could be better specified in the main paper:\n    -  It is not entirely clear to me by reading the paper if the (hyper-)dual numbers allow for the computation of JVPs and bilinear hessian products by just \"tracking the epsilons\" while computing the function, or if one has actually to implement the above-mentioned operations (I think it's the first). One \"implementation\" example would help focus ideas.\n    - I'm a bit confused about the origin of Equation (1) right. Do the expressions for the $\\kappa$ come from manipulation of the resulting Taylor expansion or is it \"set by design\" (to mimic Newton updates)? In general, I would have appreciated some more details around lines 227 to 240 \n   - the learning rate scheduler seems like an important addition for empirical performance, some more details (e.g. which scheduler, how did you choose its hyperparamters) in the main paper would be welcomed.\n   - some more discussion on relations between FoMoH-1d, FoMoH-BP and FGD would have been nice. For me it is not immediately clear why FoMoH-1d consistently underperforms FGD on the learning tasks and why FoMoH-BP consistently performs on par with BP. \n  - what does BP stand for in the experiments? Is it plain (stochastic?) gradient descent or some other adaptive method?", "questions": "1. Is it correct to say that the method performs Newton steps in the sampled subspaces (assuming learning rate being 1)? If not, what's the relationship between the two?\n2. the $\\kappa_i$'s need not be positive, correct?\n3. in the learning tasks, are also examples being smapled (i.e. mini-batch updates?). If so, is the proposed method sensitive to the mini-batch size?\n4. Do you think the proposed method could be relevant also for forward-mode gradient-based hyperparameter optimization?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a forward-mode-only optimization method that uses second order information on randomly sampled hyperplanes. \nThe method makes use of (K^2 + K)/2 calls to a \"double\" forward-mode AD, implemented with hyper-dual numbers, computing along the way the Hessian projected onto the drawn plane. The paper presents theoretical results on convex and quadratic functions that also relates it to the Newton method, alongside some empirical validation on a test function and two learning problems.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- The significance of developing a preforming optimization method that does not rely on reverse-mode differentiation is very high, and the proposed method seems to make a concrete step in this direction \n- The paper is mostly well written and easy to follow, the notation is clear (although a few passages could be made clearer, see below)\n- The work does a good job introducing the concept of dual and hyper-dual numbers, which I expect the community to benefit from \n-  The theoretical results clearly show the advantages of the proposed method over FGD, as well as the larger scale experiment.", "weaknesses": "- One main promise of forward mode differentiation is to vastly decrease the memory requirement for large models, however the paper does not empirically quantifies the advantage of FoMoH in this regard. It would be nice to include memory footprint comparisons (as well as perhaps a table summarizing the runtime and memory complexity of key algorithms)\n- I would have appreciated some larger scale experiment with transformer architectures, e.g. for fine-tuning LLMs, which could be an interesting application of the proposed method \n- Some details could be better specified in the main paper:\n    -  It is not entirely clear to me by reading the paper if the (hyper-)dual numbers allow for the computation of JVPs and bilinear hessian products by just \"tracking the epsilons\" while computing the function, or if one has actually to implement the above-mentioned operations (I think it's the first). One \"implementation\" example would help focus ideas.\n    - I'm a bit confused about the origin of Equation (1) right. Do the expressions for the $\\kappa$ come from manipulation of the resulting Taylor expansion or is it \"set by design\" (to mimic Newton updates)? In general, I would have appreciated some more details around lines 227 to 240 \n   - the learning rate scheduler seems like an important addition for empirical performance, some more details (e.g. which scheduler, how did you choose its hyperparamters) in the main paper would be welcomed.\n   - some more discussion on relations between FoMoH-1d, FoMoH-BP and FGD would have been nice. For me it is not immediately clear why FoMoH-1d consistently underperforms FGD on the learning tasks and why FoMoH-BP consistently performs on par with BP. \n  - what does BP stand for in the experiments? Is it plain (stochastic?) gradient descent or some other adaptive method?", "questions": "1. Is it correct to say that the method performs Newton steps in the sampled subspaces (assuming learning rate being 1)? If not, what's the relationship between the two?\n2. the $\\kappa_i$'s need not be positive, correct?\n3. in the learning tasks, are also examples being smapled (i.e. mini-batch updates?). If so, is the proposed method sensitive to the mini-batch size?\n4. Do you think the proposed method could be relevant also for forward-mode gradient-based hyperparameter optimization?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730846727859}, {"id": "Tc4trbkjPb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission443/Reviewer_qC9Q"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper investigates the effectiveness of the second-order differentiation in forward automated differentiation.\nThis paper and regards the forward propagation with dual number as the objective function applied second-order Taylor-series expansion to.\nBy using dual numbers, the proposed method estimates the Hessian matrix and applies the approximated Newton's method for the optimization.", "review_text": "This paper investigates the effectiveness of the second-order differentiation in forward automated differentiation.\nThis paper and regards the forward propagation with dual number as the objective function applied second-order Taylor-series expansion to.\nBy using dual numbers, the proposed method estimates the Hessian matrix and applies the approximated Newton's method for the optimization.", "strengths": "1. This paper is well-written and easy to follow. Assumptions and mathematical expansions are explained in detail.\n1. Experimental results demonstrate the proposed method is more efficient than baseline first-order methods in terms of iterations of parameter updates.\n1. Convergence properties are proved in theory, and the proposed method is guaranteed to be consist with Newton's method.", "weaknesses": "1. I am not very sure that this paper addresses a new topic in machine learning.\nThe explanation of the proposed method seems to address general optimization problems but does not seem to focus on the problem in machine learning.\nFor examples, this paper does not explicitly address the difficulty of accessing the objective function due to large datasets like SGD, and not address the difficulty of non-convexity caused by nonlinear models.\nSince I do not have much expertise in optimization theory or operations research, I cannot evaluate the novelty of this paper well in the context of optimization theory. \nEven so, I doubt the proposed method is very novel because the used mathematical tools are fundamental and the addressed objective function does not seem very difficult. \nIs the research problem specialized for machine learning problems? And, is the paper new even in the context of the optimization problem?\n\n\n1. While this paper evaluates the convergence of the proposed method in terms of iterations, it does not evaluate the runtime of the proposed method.\nThe proposed method requires an inverse of Hessian matrices, and I think its computational cost can be high.\nDoes the overhead of the proposed method not increase the runtime in one iteration? If it does, the proposed method is still faster than baselines in terms of runtime until convergence?\nTo emphasize the practical usefulness of the proposed method, this paper needs the evaluation of runtime.\n\n1. It is not clear that the proposed method is scalable for recent deep neural network model architectures.\n(Fournier et al., 2023) seems to show that the first-order forward gradient method is applicable to ResNet18. Is the proposed method applicable to such modern architectures?", "questions": "1. Why is this paper suitable for publication as machine learning research? What is the difficult point of the optimization method in machine learning, and how does the paper address it? If the approximation of a second-order method using dual numbers is new, why has no literature in optimization theory or operations research discussed it?\n\n1. Does not the overhead of the proposed method increase the runtime in one iteration? If it does, the proposed method is still faster than baselines in terms of runtime until convergence?\n\n1. How is the scalability of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the effectiveness of the second-order differentiation in forward automated differentiation.\nThis paper and regards the forward propagation with dual number as the objective function applied second-order Taylor-series expansion to.\nBy using dual numbers, the proposed method estimates the Hessian matrix and applies the approximated Newton's method for the optimization.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. This paper is well-written and easy to follow. Assumptions and mathematical expansions are explained in detail.\n1. Experimental results demonstrate the proposed method is more efficient than baseline first-order methods in terms of iterations of parameter updates.\n1. Convergence properties are proved in theory, and the proposed method is guaranteed to be consist with Newton's method.", "weaknesses": "1. I am not very sure that this paper addresses a new topic in machine learning.\nThe explanation of the proposed method seems to address general optimization problems but does not seem to focus on the problem in machine learning.\nFor examples, this paper does not explicitly address the difficulty of accessing the objective function due to large datasets like SGD, and not address the difficulty of non-convexity caused by nonlinear models.\nSince I do not have much expertise in optimization theory or operations research, I cannot evaluate the novelty of this paper well in the context of optimization theory. \nEven so, I doubt the proposed method is very novel because the used mathematical tools are fundamental and the addressed objective function does not seem very difficult. \nIs the research problem specialized for machine learning problems? And, is the paper new even in the context of the optimization problem?\n\n\n1. While this paper evaluates the convergence of the proposed method in terms of iterations, it does not evaluate the runtime of the proposed method.\nThe proposed method requires an inverse of Hessian matrices, and I think its computational cost can be high.\nDoes the overhead of the proposed method not increase the runtime in one iteration? If it does, the proposed method is still faster than baselines in terms of runtime until convergence?\nTo emphasize the practical usefulness of the proposed method, this paper needs the evaluation of runtime.\n\n1. It is not clear that the proposed method is scalable for recent deep neural network model architectures.\n(Fournier et al., 2023) seems to show that the first-order forward gradient method is applicable to ResNet18. Is the proposed method applicable to such modern architectures?", "questions": "1. Why is this paper suitable for publication as machine learning research? What is the difficult point of the optimization method in machine learning, and how does the paper address it? If the approximation of a second-order method using dual numbers is new, why has no literature in optimization theory or operations research discussed it?\n\n1. Does not the overhead of the proposed method increase the runtime in one iteration? If it does, the proposed method is still faster than baselines in terms of runtime until convergence?\n\n1. How is the scalability of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730190947443}], "openreview_url": "https://openreview.net/forum?id=XW4Xnx0xlH", "arxiv_id": "2408.10419", "paper_pdf": "papers/XW4Xnx0xlH.pdf", "paper_pdf_sha256": "096b1621c54f164d1ee7bf7c2ec4878d4d158a477bdbb9d88f6357cec15d87ec", "paper_pdf_bytes": 9699584, "paper_pdf_source": "openreview", "code_url": "https://github.com/SRI-CSL/fomoh", "code_repository": "SRI-CSL/fomoh", "code_commit": "f0da7ae90ac9d93ea4c704462b6c75d74ef31434", "code_archive": "repos/XW4Xnx0xlH.zip", "code_archive_sha256": "780f5c473b03dbea6af439861056995d33b0935dcaa0a978352e647fd1036acd", "code_archive_bytes": 236853, "code_file_count": 17, "code_extensions": {".py": 11, ".sh": 4, ".ipynb": 2}, "github_disk_usage_kb": 232, "github_languages": {"Jupyter Notebook": 221228, "Python": 125546, "Shell": 2928}, "github_archived": false, "github_pushed_at": "2024-08-26T14:34:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/second-order-forward-mode-automatic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Gq1Zjhovjr", "year": 2024, "status": "rejected", "title": "Consistency Regularization for Domain Generalization with Logit Attribution Matching", "authors": ["Han Gao", "Kaican Li", "Weiyan Xie", "Zhi LIN", "Yongxiang Huang", "Luning Wang", "Caleb Chen Cao", "Nevin L. Zhang"], "authorids": ["~Han_Gao5", "~Kaican_Li1", "~Weiyan_Xie1", "~Zhi_LIN1", "~Yongxiang_Huang1", "~Luning_Wang1", "~Caleb_Chen_Cao1", "~Nevin_L._Zhang1"], "authors_source": "OpenReview API", "abstract": "Domain generalization (DG) is about training models that generalize well to unseen domains that follow different distributions than the training domains. It has recently been shown that an effective way to achieve good DG performance is targeted data augmentation, which randomizes spurious factors while preserving robustly predictive factors in training examples. Data augmentation (DA) naturally leads to paired training examples that share the same semantic contents, which can be utilized via consistency regularization (CR). In this paper, we show that CR can further boost DG performance on top of targeted DA. We also propose a novel CR-based DG method called Logit Attribution Matching (LAM). In comparison with previous CR-based DG methods, a key advantage of LAM is that it leverages class labels  often associated with semantic sharing (SS) pairs.  Empirically we find that LAM consistently outperforms previous CR-based DG methods on benchmarks with multiple classes.  In fact, it is the only one that can consistently improve the model DG performance over the targeted DA on all evaluated datasets. To justify the CR-based approach to DG theoretically, we establish conditions for optimal DG in a causal framework and explain how CR is related to those conditions.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "rldE3Sv1QH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1705/Reviewer_4UiZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper aims to enhance the consistency regularization method for domain generalization by incorporating a logit attribution matching approach.\nThe authors first revisit the domain generalization (DG) problem through the causal latent decomposition (CLD) model. This model indicates that DG adheres to the concept of causal-invariant prediction, wherein the predicted labels for a semantic sharing pair remain consistent with diverse non-core factors, as long as the core factors remain unchanged. They then introduce a theorem of Conditions for Optimal DG and unveil consistency regularization as a potential optimal solution for DG, subject to certain assumptions. Existing methods, such as probability matching, logit matching, and feature matching, can all be treated as special cases of this optimal solution. The authors then develop the Logit Attribution Matching (LAM) regularizer, building upon the feature matching method. This approach introduces weights on each dimension of the features, corresponding to each label y. It is hoped that this design allows the model to pay more attention to core factors than non-core factors and improve OOD performance.", "review_text": "This paper aims to enhance the consistency regularization method for domain generalization by incorporating a logit attribution matching approach.\nThe authors first revisit the domain generalization (DG) problem through the causal latent decomposition (CLD) model. This model indicates that DG adheres to the concept of causal-invariant prediction, wherein the predicted labels for a semantic sharing pair remain consistent with diverse non-core factors, as long as the core factors remain unchanged. They then introduce a theorem of Conditions for Optimal DG and unveil consistency regularization as a potential optimal solution for DG, subject to certain assumptions. Existing methods, such as probability matching, logit matching, and feature matching, can all be treated as special cases of this optimal solution. The authors then develop the Logit Attribution Matching (LAM) regularizer, building upon the feature matching method. This approach introduces weights on each dimension of the features, corresponding to each label y. It is hoped that this design allows the model to pay more attention to core factors than non-core factors and improve OOD performance.", "strengths": "1. The writing of this paper is clear and the idea is easy to follow.\n2. It is interesting to revisit domain generalization from a causal latent decomposition perspective and highlight the core and non-core factors that are not considered in previous works.\n3. It is also interesting to utilize the Optimal DG theorem to summarize the existing consistency regularization methods into a general framework.\n4. A thorough experiment is conducted in the main paper content as well as the appendix to evaluate and analyze the proposed methods.", "weaknesses": "1. The contributions may not be very significant because consistency regularization for DG has already been extensively studied in previous research and the proposed method only makes simple modifications to the existing feature matching method. Several concepts in this paper are borrowed from previous ideas, such as targeted augmentation and causal-invariant prediction. \n\n2. The theorem does not serve as a supporting foundation for the proposed logit attribution matching method. It seems that the theorem and the techniques are divided and tell two different stories. Concretely, the theorem just reveals that consistency regularization can be an optimal solution for DG. It is not connected to why logit attribution matching is needed to deal with core and non-core factors. Before this paper can be accepted, it is necessary that it undergoes revision to ensure a clear and concise explanation of how the theorem serves as a driving force behind the techniques employed.", "questions": "Is it possible to build logit attribution matching upon probability matching or logit matching models?\nAre they better or worse than building upon a feature matching model? \nCan the authors provide more analysis and explanations of why you just choosing feature matching methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to enhance the consistency regularization method for domain generalization by incorporating a logit attribution matching approach.\nThe authors first revisit the domain generalization (DG) problem through the causal latent decomposition (CLD) model. This model indicates that DG adheres to the concept of causal-invariant prediction, wherein the predicted labels for a semantic sharing pair remain consistent with diverse non-core factors, as long as the core factors remain unchanged. They then introduce a theorem of Conditions for Optimal DG and unveil consistency regularization as a potential optimal solution for DG, subject to certain assumptions. Existing methods, such as probability matching, logit matching, and feature matching, can all be treated as special cases of this optimal solution. The authors then develop the Logit Attribution Matching (LAM) regularizer, building upon the feature matching method. This approach introduces weights on each dimension of the features, corresponding to each label y. It is hoped that this design allows the model to pay more attention to core factors than non-core factors and improve OOD performance.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The writing of this paper is clear and the idea is easy to follow.\n2. It is interesting to revisit domain generalization from a causal latent decomposition perspective and highlight the core and non-core factors that are not considered in previous works.\n3. It is also interesting to utilize the Optimal DG theorem to summarize the existing consistency regularization methods into a general framework.\n4. A thorough experiment is conducted in the main paper content as well as the appendix to evaluate and analyze the proposed methods.", "weaknesses": "1. The contributions may not be very significant because consistency regularization for DG has already been extensively studied in previous research and the proposed method only makes simple modifications to the existing feature matching method. Several concepts in this paper are borrowed from previous ideas, such as targeted augmentation and causal-invariant prediction. \n\n2. The theorem does not serve as a supporting foundation for the proposed logit attribution matching method. It seems that the theorem and the techniques are divided and tell two different stories. Concretely, the theorem just reveals that consistency regularization can be an optimal solution for DG. It is not connected to why logit attribution matching is needed to deal with core and non-core factors. Before this paper can be accepted, it is necessary that it undergoes revision to ensure a clear and concise explanation of how the theorem serves as a driving force behind the techniques employed.", "questions": "Is it possible to build logit attribution matching upon probability matching or logit matching models?\nAre they better or worse than building upon a feature matching model? \nCan the authors provide more analysis and explanations of why you just choosing feature matching methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698933952500}, {"id": "2Oj9fHy2DZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1705/Reviewer_9AoB"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Models can be trained to generalize to new domains (called domain generalization) through augmenting training data. Another approach to domain generalization is through consistency regularization, which enforces that the model should make similar predictions on similar inputs.  This paper proposes using consistency regularization on top of data augmentation; that is, enforcing similarities (at the logit, output, embedding level) on pairs of unaugmented and augmented samples. The paper provides a theoretical result stating that a model that is causal invariant (E.g. $\\hat{P}(Y | x) = \\hat{P}(Y | \\tilde{x})$) and minimizes loss in-distribution will also minimize out-of-distribution loss. The paper finally proposes a new form of consistency regularization, called logit attribution matching (LAM), which encourages feature matching on features that are strongly associated with the true label; this is more granular than previous approaches. The performance of consistency regularization on top of targeted data augmentation is compared to using standard data augmentation to expand the training dataset, and LAM is compared against other consistency regularization methods.", "review_text": "Models can be trained to generalize to new domains (called domain generalization) through augmenting training data. Another approach to domain generalization is through consistency regularization, which enforces that the model should make similar predictions on similar inputs.  This paper proposes using consistency regularization on top of data augmentation; that is, enforcing similarities (at the logit, output, embedding level) on pairs of unaugmented and augmented samples. The paper provides a theoretical result stating that a model that is causal invariant (E.g. $\\hat{P}(Y | x) = \\hat{P}(Y | \\tilde{x})$) and minimizes loss in-distribution will also minimize out-of-distribution loss. The paper finally proposes a new form of consistency regularization, called logit attribution matching (LAM), which encourages feature matching on features that are strongly associated with the true label; this is more granular than previous approaches. The performance of consistency regularization on top of targeted data augmentation is compared to using standard data augmentation to expand the training dataset, and LAM is compared against other consistency regularization methods.", "strengths": "Quality:\n- Theoretical results motivate why causal invariant property is important for domain generalization.\n- LAM outperforms both DG methods and CR methods \n\nClarity:\n- Toy illustration in figure 1 made theoretical result and setup more clear. \n\nSignificance:\n- Handling OOD settings is an important problem in machine learning.", "weaknesses": "Quality:\n- Theory is not connected to LAM. Since the condition is $\\hat{P}(y | x) = \\hat{P}(y | \\tilde{x})$, why does probability matching not work well? Why does LAM work better?\n- The theoretical model's connection to data augmentation is also rather weak. Can you show that your choice of data augmentation is retaining $x^c$ and changing $x^n$?\n\nOriginality:\n- Having trouble understanding why CR on top of DA is a contribution. In the related work you say that CR can use different data augmentation strategies as well as alternate ways to pair up samples. \n- This paper combines two well-studied ideas into one with a new theoretical result and a new consistency regularization term, but the theory and the new term could be more motivated.\n\nClarity:\n- Minor nit: this paper has many abbreviations (DG, DA, OOD, CR, CLD). I found that use of DG and DA were a bit confusing the first time I read the paper, and would prefer having more sentences with the full words, at least in the introduction. \n- Unclear how contributions in the introduction are related. It reads like a list of ways to improve performance but don't feel well-motivated.", "questions": "- Theory is not connected to LAM. Since the condition is $\\hat{P}(y | x) = \\hat{P}(y | \\tilde{x})$, why does probability matching not work well? Why does LAM work better?\n- The theoretical model's connection to data augmentation is also rather weak. Can you show that your choice of data augmentation is retaining $x^c$ and changing $x^n$?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Models can be trained to generalize to new domains (called domain generalization) through augmenting training data. Another approach to domain generalization is through consistency regularization, which enforces that the model should make similar predictions on similar inputs.  This paper proposes using consistency regularization on top of data augmentation; that is, enforcing similarities (at the logit, output, embedding level) on pairs of unaugmented and augmented samples. The paper provides a theoretical result stating that a model that is causal invariant (E.g. $\\hat{P}(Y | x) = \\hat{P}(Y | \\tilde{x})$) and minimizes loss in-distribution will also minimize out-of-distribution loss. The paper finally proposes a new form of consistency regularization, called logit attribution matching (LAM), which encourages feature matching on features that are strongly associated with the true label; this is more granular than previous approaches. The performance of consistency regularization on top of targeted data augmentation is compared to using standard data augmentation to expand the training dataset, and LAM is compared against other consistency regularization methods.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "Quality:\n- Theoretical results motivate why causal invariant property is important for domain generalization.\n- LAM outperforms both DG methods and CR methods \n\nClarity:\n- Toy illustration in figure 1 made theoretical result and setup more clear. \n\nSignificance:\n- Handling OOD settings is an important problem in machine learning.", "weaknesses": "Quality:\n- Theory is not connected to LAM. Since the condition is $\\hat{P}(y | x) = \\hat{P}(y | \\tilde{x})$, why does probability matching not work well? Why does LAM work better?\n- The theoretical model's connection to data augmentation is also rather weak. Can you show that your choice of data augmentation is retaining $x^c$ and changing $x^n$?\n\nOriginality:\n- Having trouble understanding why CR on top of DA is a contribution. In the related work you say that CR can use different data augmentation strategies as well as alternate ways to pair up samples. \n- This paper combines two well-studied ideas into one with a new theoretical result and a new consistency regularization term, but the theory and the new term could be more motivated.\n\nClarity:\n- Minor nit: this paper has many abbreviations (DG, DA, OOD, CR, CLD). I found that use of DG and DA were a bit confusing the first time I read the paper, and would prefer having more sentences with the full words, at least in the introduction. \n- Unclear how contributions in the introduction are related. It reads like a list of ways to improve performance but don't feel well-motivated.", "questions": "- Theory is not connected to LAM. Since the condition is $\\hat{P}(y | x) = \\hat{P}(y | \\tilde{x})$, why does probability matching not work well? Why does LAM work better?\n- The theoretical model's connection to data augmentation is also rather weak. Can you show that your choice of data augmentation is retaining $x^c$ and changing $x^n$?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698795640785}, {"id": "Iwmh40XPNo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1705/Reviewer_H9LW"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper studies the problem of domain generalization. It creates a theoretical model prescribing the relationship between the source and target domain, for which they argue the benefit of consistency regularization. The paper further presents a new consistency regularization scheme, referred to as Logit Attribution Matching (LAM). The key idea there is the match the logits of a pair of related examples while incorporating label information. Experimental study demonstrate performance improvements.", "review_text": "The paper studies the problem of domain generalization. It creates a theoretical model prescribing the relationship between the source and target domain, for which they argue the benefit of consistency regularization. The paper further presents a new consistency regularization scheme, referred to as Logit Attribution Matching (LAM). The key idea there is the match the logits of a pair of related examples while incorporating label information. Experimental study demonstrate performance improvements.", "strengths": "The main novelty of the paper are theoretical argument justifying the benefit of the consistency regularization and the proposal of the LAM, which take label information into account.  But to this reviewer, the novelty on both sides is thin. Theorem 1 holds nearly trivially; the LAM idea is also straight-forward.", "weaknesses": "1. Theorem 1 contains the very strong assumption that the target distribution of $X^c$ lies within the support of corresponding source distribution (Assumption 3) of the theorem. It is highly suspicious if in reality such a condition would hold true in conjunction with the first two assumptions. It seems that such a condition would only hold in a regime where transfer learning is easy. \n\n2. In the description of LAM, it is not clear to me if the weights $\\{w_{uy_k}\\}$ are hyperparameters or if they are learned during training. If they are hyperparameters, how are they decided? and why not set them to 1 for each $(u, y_k)$?. If they are learned, what mechanism would force them to satisfy the two conditions listed on page 6 (lines 8 and 9)? Note that these weights and $f_\\phi$ are learned together.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies the problem of domain generalization. It creates a theoretical model prescribing the relationship between the source and target domain, for which they argue the benefit of consistency regularization. The paper further presents a new consistency regularization scheme, referred to as Logit Attribution Matching (LAM). The key idea there is the match the logits of a pair of related examples while incorporating label information. Experimental study demonstrate performance improvements.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The main novelty of the paper are theoretical argument justifying the benefit of the consistency regularization and the proposal of the LAM, which take label information into account.  But to this reviewer, the novelty on both sides is thin. Theorem 1 holds nearly trivially; the LAM idea is also straight-forward.", "weaknesses": "1. Theorem 1 contains the very strong assumption that the target distribution of $X^c$ lies within the support of corresponding source distribution (Assumption 3) of the theorem. It is highly suspicious if in reality such a condition would hold true in conjunction with the first two assumptions. It seems that such a condition would only hold in a regime where transfer learning is easy. \n\n2. In the description of LAM, it is not clear to me if the weights $\\{w_{uy_k}\\}$ are hyperparameters or if they are learned during training. If they are hyperparameters, how are they decided? and why not set them to 1 for each $(u, y_k)$?. If they are learned, what mechanism would force them to satisfy the two conditions listed on page 6 (lines 8 and 9)? Note that these weights and $f_\\phi$ are learned together.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698786185496}, {"id": "8y0O6CpZJc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1705/Reviewer_THPX"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a novel method called logit attribution matching (LAM) for improving domain generalization. Compared to existing consistency regularization methods (probability matching, logit matching, feature matching etc), the proposed method further adds class label information into the regularization term across semantic sharing pairs during data augmentation.\n\nExperiments on a wide set of datasets show that LAM outperforms existing consistency regularization methods, and outperform domain adaptation approaches as well.", "review_text": "This paper presents a novel method called logit attribution matching (LAM) for improving domain generalization. Compared to existing consistency regularization methods (probability matching, logit matching, feature matching etc), the proposed method further adds class label information into the regularization term across semantic sharing pairs during data augmentation.\n\nExperiments on a wide set of datasets show that LAM outperforms existing consistency regularization methods, and outperform domain adaptation approaches as well.", "strengths": "- The proposed idea is fairly simple by introducing an additional weight term over each feature unit and class label. Yet this simple weight term seems to be rather useful in improving the OOD performance over a wide set of image datasets.\n\n- The authors did a fairly comprehensive set of experiments over 5 image datasets to demonstrate the superiority of the proposed method.", "weaknesses": "- Over the five datasets experimented, the augmentation is handpicked based on the characteristics of each dataset. This might make the disentanglement of causal/non-causal features relatively easier (i.e., the SS pairs better fit into this paper's motivation on $X^c$ and $X^n$). I wonder how the proposed approach works when the augmentation is agnostic to the datasets, i.e., what if you apply one of RandAugment / CutMix / AugMix to all the datasets as augmentation, and use LAM for regularization? How would the performance change?", "questions": "- Can the authors show how LAM works when the augmentation applied is dataset-agnostic? e.g., using augmentation methods like RandAugment / CutMix / AugMix for all the datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel method called logit attribution matching (LAM) for improving domain generalization. Compared to existing consistency regularization methods (probability matching, logit matching, feature matching etc), the proposed method further adds class label information into the regularization term across semantic sharing pairs during data augmentation.\n\nExperiments on a wide set of datasets show that LAM outperforms existing consistency regularization methods, and outperform domain adaptation approaches as well.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The proposed idea is fairly simple by introducing an additional weight term over each feature unit and class label. Yet this simple weight term seems to be rather useful in improving the OOD performance over a wide set of image datasets.\n\n- The authors did a fairly comprehensive set of experiments over 5 image datasets to demonstrate the superiority of the proposed method.", "weaknesses": "- Over the five datasets experimented, the augmentation is handpicked based on the characteristics of each dataset. This might make the disentanglement of causal/non-causal features relatively easier (i.e., the SS pairs better fit into this paper's motivation on $X^c$ and $X^n$). I wonder how the proposed approach works when the augmentation is agnostic to the datasets, i.e., what if you apply one of RandAugment / CutMix / AugMix to all the datasets as augmentation, and use LAM for regularization? How would the performance change?", "questions": "- Can the authors show how LAM works when the augmentation applied is dataset-agnostic? e.g., using augmentation methods like RandAugment / CutMix / AugMix for all the datasets?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698770539657}, {"id": "XomQiFqLkI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1705/Reviewer_cViE"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the magic of consistency regularization in domain generalization. First, authors claim that CR remains effective for DG and there are a few of existing approaches. Then, they study the theory behind the combination of CR and targeted augmentation. Finally, authors design their own approach called logit attribution matching to simply match the logits to further improve the performance. Experiments have shown its effectiveness.\n\n------Post rebuttal\n\nThe response addressed my concerns and I increased the score to 6.", "review_text": "This paper studies the magic of consistency regularization in domain generalization. First, authors claim that CR remains effective for DG and there are a few of existing approaches. Then, they study the theory behind the combination of CR and targeted augmentation. Finally, authors design their own approach called logit attribution matching to simply match the logits to further improve the performance. Experiments have shown its effectiveness.\n\n------Post rebuttal\n\nThe response addressed my concerns and I increased the score to 6.", "strengths": "1. The paper presents a nice analysis of the consistency regularization in domain generalization, with interesting theoretical support.\n2. Based on their theoretical analysis, the LAM approach is proposed which combines the existing targeted augmentation approach to further enhance the performance of CR-based DG.\n3. Extensive experiments on several benchmark datasets have shown that the method brings improvements over ERM.", "weaknesses": "1. While I’m not an expert in causality, I hold doubt about the assumption in Figure 1, i.e., the data generation process. We certainly know that such generation process is an assumption, and other causality researchers can draw a completely different causal graph. Therefore, how can authors justify that this figure is practical and can be trusted? This is important since all analysis is based on this basic assumption.\n2. I admire the effort to combine targeted augmentation with CR. But I do not think targeted augmentation is algorithmically novel since this is not general and hard to generalize to other datasets, given the wide popularity of DG in different applications. Therefore, the introduction of targeted augmentation is not efficient or general. This makes LAM deeply rely on the effectiveness of TA. I would like to know how can LAM be applied to other new domains where targeted augmentation is not realistic.\n3. In the experiment section, I did not see any comparison with existing CR-based baselines, but only ERM variants. Did I miss anything?", "questions": "See the weakness. I'm extremely curious about the practical usage of LAM without the help from TA.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the magic of consistency regularization in domain generalization. First, authors claim that CR remains effective for DG and there are a few of existing approaches. Then, they study the theory behind the combination of CR and targeted augmentation. Finally, authors design their own approach called logit attribution matching to simply match the logits to further improve the performance. Experiments have shown its effectiveness.\n\n------Post rebuttal\n\nThe response addressed my concerns and I increased the score to 6.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper presents a nice analysis of the consistency regularization in domain generalization, with interesting theoretical support.\n2. Based on their theoretical analysis, the LAM approach is proposed which combines the existing targeted augmentation approach to further enhance the performance of CR-based DG.\n3. Extensive experiments on several benchmark datasets have shown that the method brings improvements over ERM.", "weaknesses": "1. While I’m not an expert in causality, I hold doubt about the assumption in Figure 1, i.e., the data generation process. We certainly know that such generation process is an assumption, and other causality researchers can draw a completely different causal graph. Therefore, how can authors justify that this figure is practical and can be trusted? This is important since all analysis is based on this basic assumption.\n2. I admire the effort to combine targeted augmentation with CR. But I do not think targeted augmentation is algorithmically novel since this is not general and hard to generalize to other datasets, given the wide popularity of DG in different applications. Therefore, the introduction of targeted augmentation is not efficient or general. This makes LAM deeply rely on the effectiveness of TA. I would like to know how can LAM be applied to other new domains where targeted augmentation is not realistic.\n3. In the experiment section, I did not see any comparison with existing CR-based baselines, but only ERM variants. Did I miss anything?", "questions": "See the weakness. I'm extremely curious about the practical usage of LAM without the help from TA.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698399540504}], "openreview_url": "https://openreview.net/forum?id=Gq1Zjhovjr", "arxiv_id": "2305.07888", "paper_pdf": "papers/Gq1Zjhovjr.pdf", "paper_pdf_sha256": "cc4346a681d2d3b2072589ebb8d214dcaa78ea9ba646e19db4ed5c28337572c3", "paper_pdf_bytes": 32265961, "paper_pdf_source": "openreview", "code_url": "https://github.com/Gaohan123/LAM", "code_repository": "Gaohan123/LAM", "code_commit": "e13b95c28976a1d5179167db2510fd6a398f0cd2", "code_archive": "repos/Gq1Zjhovjr.zip", "code_archive_sha256": "7a55d40331601e195aa1f42e296627358770eb000bf12062994a35fe1c75f6d6", "code_archive_bytes": 161446, "code_file_count": 57, "code_extensions": {".py": 57}, "github_disk_usage_kb": 215, "github_languages": {"Python": 458987}, "github_archived": false, "github_pushed_at": "2024-06-13T14:00:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contrastive-domain-generalization-via-logit"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "eiuj6cNv4iI", "year": 2023, "status": "rejected", "title": "WebBrain: Learning to Generate Factually Correct Articles for Queries by Grounding on Large Web Corpus", "authors": ["Hongjin Qian", "Yutao Zhu", "Zhicheng Dou", "Haoqi Gu", "Xinyu Zhang", "Zheng Liu", "Ruofei Lai", "Zhao Cao", "Jian-Yun Nie", "Ji-Rong Wen"], "authorids": ["~Hongjin_Qian1", "~Yutao_Zhu1", "~Zhicheng_Dou1", "guhaoqi1@huawei.com", "~Xinyu_Zhang6", "liuzheng107@huawei.com", "lairuofei@huawei.com", "~Zhao_Cao1", "~Jian-Yun_Nie1", "~Ji-Rong_Wen1"], "authors_source": "OpenReview API", "abstract": "In this paper, we introduce a new NLP task – generating short factual articles for queries by mining supporting evidence from the Web. In this task, called WebBrain, the ultimate goal is to generate a fluent, informative, and factually-correct short article (e.g., Wiki article) for a factual query unseen in Wikipedia. To enable experiments on WebBrain, we construct a large-scale dataset WebBrain-Raw by extracting English Wikipedia articles and their crawlable Wiki references. WebBrain-Raw is ten times larger than the previous biggest peer dataset, which can greatly benefit the research community. Besides, we empirically analyze the performances of the current state-of-the-art NLP techniques on WebBrain and introduce a new framework ReGen, which enhances the generation factualness by improved evidence retrieval and task-specific pre-training for generation. Experiment results show that ReGen outperforms all baselines in both automatic and human evaluations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "5z-M9diMDnw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5974/Reviewer_TtTt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a new dataset and associated task that involves generating the first section of Wikipedia pages from a set of retrieved references. The dataset is similar in form to previous work like WikiSumm, but it is significantly larger and the authors promise to release the references, which have been downloaded from non-Wikipedia sites.\n\nThe dataset generation process involves a number of filtering steps, based on term overlap, that narrow down the reference data to passages that are likely to be related to the target Wikipedia article and associate these passages with sentences in the target. The goal, at test time, is to retrieve reference passages that may be related to a topic, and then\ngenerate the target section of the Wikipedia article.\n\nAlong with the dataset, this paper presents a new model, ReGen. ReGen is composed of a retriever (based on SPLADE) and a reference encoder and target decoder (based on FiD). When trained on the WebBrain data, ReGen outperforms both large language model baselines and also a retrieval augmented model with a dense retriever (FiD + BART) according to a range of automatic metrics, and also annotator judgements of a small set of predictions. I have some questions about some of these comparisions below, but it is clear that ReGen is capable of generating coherent and informative text. It is also capable of providing references for this text, which is a significant advantage over methods that don't include retrieval.\n\nThere are a number of nice ablations that show how the different systems' performances change with different numbers of retrievals and gold references. These comparisons are useful in understanding the tradeoffs between retrieving a lot of supporting evidence, with lower precision, vs fewer high quality references. ", "review_text": "This paper presents a significant new dataset and task that could be useful to the community, going forward, in benchmarking methods of retrieval augmented generation. The new model is also quite different from previous work, which has generally relied on dense DPR-style retrievers.\n\nOverall, I think this paper is a nice contribution and I am currently leaning toward acceptance. However, I also have some serious questions below about the data release strategy, and how the authors propose to handle concerns about releasing multiple terrabytes of data which may include copyrighted works or personal information. \n", "strengths": "#### Strengths\n- This paper promises a massive new dataset for retrieval augmented text generation, that could be very useful to the NLP community. Having said that, I do have questions about the release with respect to copyright and privacy concerns (detailed below in the ethics section).\n- The end system can retrieve references, and link them to generated sentences, which is a very nice end product which could have real utility beyond unconstrained text generation, which cannot provide any form of attribution for its predictions.\n- There are a number of interesting modeling and training choices made, which lead to significant improvements over a very large language model (GPT3) and also similar retrieval agumented approaches with different choices of retriever and training procedure.\n\n#### Weaknesses\n- There are a lot of separate contributions here that are not independently evaluated (data filtering / retrieval filtering / warmup strategy). It would be nice to see evaluations that validate these choices.\n- The reference passage filtering process selects the references included in the retrieval corpus references according to term overlap with the target Wikipedia article. If I'm understanding this process correctly, this means that the references stored in the corpus have been selected according to observations of the test time targets, so there is some leakage of information from the test targets into the model input. The paper should discuss this.\n- The prompt given to GPT3 \"Introduce [Page Title]\" does not mention that the target is Wikipedia-style text, and the example in Figure 1. does not look much like the start of a Wikipedia article. Meanwhile, GPT3 is definitely able to generate Wikipedia style text if prompted to do so. The comparision to GPT3 would be stronger if the prompt was better tied to the actual task.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents a new dataset and associated task that involves generating the first section of Wikipedia pages from a set of retrieved references. The dataset is similar in form to previous work like WikiSumm, but it is significantly larger and the authors promise to release the references, which have been downloaded from non-Wikipedia sites.\n\nThe dataset generation process involves a number of filtering steps, based on term overlap, that narrow down the reference data to passages that are likely to be related to the target Wikipedia article and associate these passages with sentences in the target. The goal, at test time, is to retrieve reference passages that may be related to a topic, and then\ngenerate the target section of the Wikipedia article.\n\nAlong with the dataset, this paper presents a new model, ReGen. ReGen is composed of a retriever (based on SPLADE) and a reference encoder and target decoder (based on FiD). When trained on the WebBrain data, ReGen outperforms both large language model baselines and also a retrieval augmented model with a dense retriever (FiD + BART) according to a range of automatic metrics, and also annotator judgements of a small set of predictions. I have some questions about some of these comparisions below, but it is clear that ReGen is capable of generating coherent and informative text. It is also capable of providing references for this text, which is a significant advantage over methods that don't include retrieval.\n\nThere are a number of nice ablations that show how the different systems' performances change with different numbers of retrievals and gold references. These comparisons are useful in understanding the tradeoffs between retrieving a lot of supporting evidence, with lower precision, vs fewer high quality references. ", "strength_and_weaknesses": "#### Strengths\n- This paper promises a massive new dataset for retrieval augmented text generation, that could be very useful to the NLP community. Having said that, I do have questions about the release with respect to copyright and privacy concerns (detailed below in the ethics section).\n- The end system can retrieve references, and link them to generated sentences, which is a very nice end product which could have real utility beyond unconstrained text generation, which cannot provide any form of attribution for its predictions.\n- There are a number of interesting modeling and training choices made, which lead to significant improvements over a very large language model (GPT3) and also similar retrieval agumented approaches with different choices of retriever and training procedure.\n\n#### Weaknesses\n- There are a lot of separate contributions here that are not independently evaluated (data filtering / retrieval filtering / warmup strategy). It would be nice to see evaluations that validate these choices.\n- The reference passage filtering process selects the references included in the retrieval corpus references according to term overlap with the target Wikipedia article. If I'm understanding this process correctly, this means that the references stored in the corpus have been selected according to observations of the test time targets, so there is some leakage of information from the test targets into the model input. The paper should discuss this.\n- The prompt given to GPT3 \"Introduce [Page Title]\" does not mention that the target is Wikipedia-style text, and the example in Figure 1. does not look much like the start of a Wikipedia article. Meanwhile, GPT3 is definitely able to generate Wikipedia style text if prompted to do so. The comparision to GPT3 would be stronger if the prompt was better tied to the actual task.\n", "clarity,_quality,_novelty_and_reproducibility": "#### Clarity & Reproducability\nThe paper is clearly written and all parts are pretty well described. The authors commit to releasing the data, which will allow easy replication.\n\n#### Novelty\nThe dataset is similar in form to previous work, but extends that previous work to the scenario where references must be retrieved from a very large corpus. Similarly, the model is a small modification of existing approaches but it appears to work better than well chosen baselines for this task.\n\n#### Quality\nThere are a lot of details in the filtering procedures used to select the contents of the dataset, and assign references to sentences. These are all justified int the text, but are not supported by any sort of analysis. If this is going to become a benchmark dataset, going forward, it would be good to see a discussion of how these choices affect the task. In particular, see my question about test target leakage into the retrieval corpus above (under Weaknesses).\n", "summary_of_the_review": "This paper presents a significant new dataset and task that could be useful to the community, going forward, in benchmarking methods of retrieval augmented generation. The new model is also quite different from previous work, which has generally relied on dense DPR-style retrievers.\n\nOverall, I think this paper is a nice contribution and I am currently leaning toward acceptance. However, I also have some serious questions below about the data release strategy, and how the authors propose to handle concerns about releasing multiple terrabytes of data which may include copyrighted works or personal information. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["Yes, Privacy, security and safety", "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)"], "details_of_ethics_concerns": "This work promises the release of 260M non-Wikipedia documents without any discussion of potential copyright or privacy issues. Both of these should be addressed. The authors could also consider methods of supporting user data removal requests or, alternatively, release pointers into a store like CommonCrawl, which has support for data removal requests.\n", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666983569179}, {"id": "cMVGNXf-tNq", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5974/Reviewer_dCpa"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new task called “WebBrain”. The objective of the task is to learn to generate a fluent, informative and factually correct short article from a query given some search results. The task is essentially a combination of 2 components:\n  1. retrieval of evidence passages given a wikipedia page title,\n  2. multi document summarization of the retrieved evidence with the first paragraph of the wikipedia page as target. \n\nThe authors contribute a dataset \"WebBrain-Raw\", comprised of English wikipedia plus all crawlable references. The dataset is cleaned with an interesting set of heuristics.\n\nThe authors additionally propose a baseline system called \"ReGen\". ReGen combines several modeling choices:\n  * training a SPLADE retriever for this task with hard negatives are mined,\n  * tuning sparsity in document representations for retrieval,\n  * filtering retrieved documents trading off consistency and diversity,\n  * fusion-in-decoder for generation,\n  * making generation more grounded in the references, by pre-training the generation component to generate individual sentences from a reference passage at time.\n\nThe authors finally present analysis showing the usefulness of these modeling choices, compared to more vanilla choices.", "review_text": "This looks like a useful dataset+baseline contribution for the important task of multi-document summarization. The techniques used by the authors seem solid.", "strengths": "Strengths:\n* The task examined in this paper (retrieval + multi-document summarization) is important.\n* The dataset released by this paper is much larger than comparable datasets and could be useful for furthering work on this topic.\n* The authors describe in detail many technical details of ReGen. These details can be useful to practitioners for reproducing ReGen’s results and in their own work.\n\nWeaknesses:\n* Limiting the generation to only the first passage of wikipedia pages is a pretty strong limitation, also making this work very close to existing multi-document summarization works. The authors do acknowledge this similarity though.\n* In “Reference Passage Selection” and “Dataset Generation”, the authors select input passages for training with simple word overlap or BM25. It seems like this could be easily improved by using an entailment model, a dense retriever or even SPLADE.\n* The proposed \"WebBrain-Raw\" dataset would be more interesting if it was multilingual instead of English-only.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new task called “WebBrain”. The objective of the task is to learn to generate a fluent, informative and factually correct short article from a query given some search results. The task is essentially a combination of 2 components:\n  1. retrieval of evidence passages given a wikipedia page title,\n  2. multi document summarization of the retrieved evidence with the first paragraph of the wikipedia page as target. \n\nThe authors contribute a dataset \"WebBrain-Raw\", comprised of English wikipedia plus all crawlable references. The dataset is cleaned with an interesting set of heuristics.\n\nThe authors additionally propose a baseline system called \"ReGen\". ReGen combines several modeling choices:\n  * training a SPLADE retriever for this task with hard negatives are mined,\n  * tuning sparsity in document representations for retrieval,\n  * filtering retrieved documents trading off consistency and diversity,\n  * fusion-in-decoder for generation,\n  * making generation more grounded in the references, by pre-training the generation component to generate individual sentences from a reference passage at time.\n\nThe authors finally present analysis showing the usefulness of these modeling choices, compared to more vanilla choices.", "strength_and_weaknesses": "Strengths:\n* The task examined in this paper (retrieval + multi-document summarization) is important.\n* The dataset released by this paper is much larger than comparable datasets and could be useful for furthering work on this topic.\n* The authors describe in detail many technical details of ReGen. These details can be useful to practitioners for reproducing ReGen’s results and in their own work.\n\nWeaknesses:\n* Limiting the generation to only the first passage of wikipedia pages is a pretty strong limitation, also making this work very close to existing multi-document summarization works. The authors do acknowledge this similarity though.\n* In “Reference Passage Selection” and “Dataset Generation”, the authors select input passages for training with simple word overlap or BM25. It seems like this could be easily improved by using an entailment model, a dense retriever or even SPLADE.\n* The proposed \"WebBrain-Raw\" dataset would be more interesting if it was multilingual instead of English-only.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and good quality. The system is described in detail and should be reproducible.\n\nThe novelty of the approach is mainly in the size of the released corpus and in including retrieval as part of the task, together with multi-document summarization.\n\nThe techniques proposed by the authors are very interesting, but they are only applied to the dataset they created, so it's not totally clear how their modeling choices would stack up against other methods on a more competitive benchmark.", "summary_of_the_review": "This looks like a useful dataset+baseline contribution for the important task of multi-document summarization. The techniques used by the authors seem solid.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666819890074}, {"id": "5--_YvGLGN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5974/Reviewer_892z"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new task called Web-Brain which aims to generate short factual articles for queries by mining supporting evidence from Web. The paper also proposes a new large scale dataset with English Wikipedia. The paper also provides a new framework called ReGen based on SPLADE and FiD. The model is evaluated with both n-gram overlapping metrics and factual correctness metrics. The paper analyze the impact of retrieval, number of references in a quantitative way. The paper also did both human and automatic evaluation. ", "review_text": "Overall, the paper proposes a new interesting task with a corresponding large-scale Wikipedia-based dataset. The experiment part is quite comprehensive.", "strengths": "Strength \n1. The paper introduces WebBrain, which lets the model retrieve supporting evidence and generate factual articles given a factual query. The proposed dataset is somewhat similar to the Wizard of Wikipedia (Dinan et al., 2018). The newly proposed dataset is interesting and large-scale. The authors crawled and cleaned Wikipedia. The proposed task and corresponding dataset are very interesting and worthy of future research. \n\n2. The paper proposes a new retrieval-augmented generation framework based on SPLADE and FiD. The proposed methods achieve the best results over automatic and human evaluation. The experiment section is very comprehensive. The authors conduct an ablation study with different retrieval models and show the impact of the different numbers of retrieved references. The paper also checks the impact of a number of references. Those results are clearly represented in tables or charts with detailed explanations. The paper shows the case study, human evaluation, and reference mark correction strategy in the appendix.\n\nWeaknesses\n1. The paper uses n-gram overlapping metrics for automatic evaluation. The paper needs to include some newer metrics such as BERTscore (Zhang et al., 2019), and BARTScore (Yuan et al., 2021) which can check semantic similarity. \n\n2. Most of the experiment analyses are in quantitative way. I would like to see more qualitative analysis. \n\n\n\nZhang, T., Kishore, V., Wu, F., Weinberger, K. Q., & Artzi, Y. (2019). Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675.\n\nYuan, W., Neubig, G., & Liu, P. (2021). Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34, 27263-27277.\n\n\n\nDinan, E., Roller, S., Shuster, K., Fan, A., Auli, M., & Weston, J. (2018). Wizard of Wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a new task called Web-Brain which aims to generate short factual articles for queries by mining supporting evidence from Web. The paper also proposes a new large scale dataset with English Wikipedia. The paper also provides a new framework called ReGen based on SPLADE and FiD. The model is evaluated with both n-gram overlapping metrics and factual correctness metrics. The paper analyze the impact of retrieval, number of references in a quantitative way. The paper also did both human and automatic evaluation. ", "strength_and_weaknesses": "Strength \n1. The paper introduces WebBrain, which lets the model retrieve supporting evidence and generate factual articles given a factual query. The proposed dataset is somewhat similar to the Wizard of Wikipedia (Dinan et al., 2018). The newly proposed dataset is interesting and large-scale. The authors crawled and cleaned Wikipedia. The proposed task and corresponding dataset are very interesting and worthy of future research. \n\n2. The paper proposes a new retrieval-augmented generation framework based on SPLADE and FiD. The proposed methods achieve the best results over automatic and human evaluation. The experiment section is very comprehensive. The authors conduct an ablation study with different retrieval models and show the impact of the different numbers of retrieved references. The paper also checks the impact of a number of references. Those results are clearly represented in tables or charts with detailed explanations. The paper shows the case study, human evaluation, and reference mark correction strategy in the appendix.\n\nWeaknesses\n1. The paper uses n-gram overlapping metrics for automatic evaluation. The paper needs to include some newer metrics such as BERTscore (Zhang et al., 2019), and BARTScore (Yuan et al., 2021) which can check semantic similarity. \n\n2. Most of the experiment analyses are in quantitative way. I would like to see more qualitative analysis. \n\n\n\nZhang, T., Kishore, V., Wu, F., Weinberger, K. Q., & Artzi, Y. (2019). Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675.\n\nYuan, W., Neubig, G., & Liu, P. (2021). Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34, 27263-27277.\n\n\n\nDinan, E., Roller, S., Shuster, K., Fan, A., Auli, M., & Weston, J. (2018). Wizard of Wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241.", "clarity,_quality,_novelty_and_reproducibility": "1. Some parts of the paper are not very clear. The steps to create WebBrain-R and WebBrain-G is unclear. \n\n2. The paper attached the implementation details in the appendix. It also provides examples from the dataset for readers to check. However, it does not provide any code for reproduction. It shows the limitation and system demonstration in the Appendix. ", "summary_of_the_review": "Overall, the paper proposes a new interesting task with a corresponding large-scale Wikipedia-based dataset. The experiment part is quite comprehensive.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666675697395}, {"id": "wnqUuGBqjhQ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5974/Reviewer_Ej3Q"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new benchmark for generating the first paragraph of wikipedia articles (with references) from supporting paragraphs. The authors compare a retrieve-and-generate baseline against several non-retrieval baselines (e.g., BART, GPT-2, and FiD + BART). None of these methods (including GPT-3) perform well on the task, suggesting room for future improvement.", "review_text": "While I like where this paper is going, I feel like the practical artifact and problem tackled here are a bit too small-scale and narrow. We're at a point in NLP where things actually kind of (?) begin to work, and there have been a variety of impressive results on generating and understanding long documents. I think this work would really shine in pushing this direction, but instead, it compromises itself to fit the limitations of existing models, limiting its usefulness for future work. Thus, I'm not sure that I can recommend acceptance.", "strengths": "Strengths:\n- This is an interesting and important task, and I think it's one that could be a fruitful combination of lot of existing work on both generation and understanding.\n\nWeaknesses:\n- Despite the grandiose motivations, the authors consider an extremely narrow instantiation of this task:\n  - Only the first paragraph is used as the generation target, rather than entire articles. When I read the abstract and introduction I was excited to see how the authors would handle generation / understanding of long sequences, since this is an area where I think this benchmark would be particularly useful, but was disappointed to see that they largely punted on this problem.\n  - Furthermore, the authors don't use the entirety of the reference articles, instead taking only a single paragraph per article that is specifically chosen to maximize informativeness, which I think gives a bit of an unfair advantage to the retrieval-based systems (since they're much more likely to retrieve something useful).\n\n- In general, I feel like the paragraph \"Reference Passage Selection \" in 3.2 completely misses the mark. In particular:\n\n> The original Wiki articles and reference articles tend to be very long. To adapt the capacity of most pre-trained language models (e.g., BERT has a 512 token limit), we use the first section of Wiki articles as the generation target, and we select the passage from the reference article with the highest PST value as the supporting passage.\n\nI'm not sure that I feel very good about our benchmarks having to be \"adapted\" to be consumable for existing models while trading off the essence of the task (synthesizing information from multiple long documents to generate a long document). We should set the benchmarks as goals, see how current models and techniques perform on them, and allow them to be targets for future work.\n\n- I feel like the baselines are pretty weak, and the ReGen model is a pretty standard retrieve-and-generate baseline. I don't think there's too much modeling novelty here.\n\n- Furthermore, the query distribution seems pretty unnatural. For example, in the appendix, the query used is \"how google translate work\". This is a query with a pretty clear information-seeking / question intent that can be answered by a paragraph (or paragraphs). However, the wikipedia page for \"Google Translate\" answers far more than just this question, it contains a variety of general information about the query \"google translate\" (which itself could have many different question intents). So, I feel like the source of supervision provides a different sort of information than what a query like this would actually want (let alone that in practice, only the first passage is used, which contains even less information because of the way that lead sections in wikipedia are structured to contain general high-level information: https://en.wikipedia.org/wiki/Wikipedia:Manual_of_Style/Lead_section).\n  - The original motivation mentioned \"unseen factual queries\", but it's not really clear that generating an article is what you want for this. If it's factual, you can probably answer it with a paragraph or two. I think this work would benefit from a clearer exposition of what these queries would actually be, and what the right source of supervision / data for effectively addressing these queries would look like.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a new benchmark for generating the first paragraph of wikipedia articles (with references) from supporting paragraphs. The authors compare a retrieve-and-generate baseline against several non-retrieval baselines (e.g., BART, GPT-2, and FiD + BART). None of these methods (including GPT-3) perform well on the task, suggesting room for future improvement.", "strength_and_weaknesses": "Strengths:\n- This is an interesting and important task, and I think it's one that could be a fruitful combination of lot of existing work on both generation and understanding.\n\nWeaknesses:\n- Despite the grandiose motivations, the authors consider an extremely narrow instantiation of this task:\n  - Only the first paragraph is used as the generation target, rather than entire articles. When I read the abstract and introduction I was excited to see how the authors would handle generation / understanding of long sequences, since this is an area where I think this benchmark would be particularly useful, but was disappointed to see that they largely punted on this problem.\n  - Furthermore, the authors don't use the entirety of the reference articles, instead taking only a single paragraph per article that is specifically chosen to maximize informativeness, which I think gives a bit of an unfair advantage to the retrieval-based systems (since they're much more likely to retrieve something useful).\n\n- In general, I feel like the paragraph \"Reference Passage Selection \" in 3.2 completely misses the mark. In particular:\n\n> The original Wiki articles and reference articles tend to be very long. To adapt the capacity of most pre-trained language models (e.g., BERT has a 512 token limit), we use the first section of Wiki articles as the generation target, and we select the passage from the reference article with the highest PST value as the supporting passage.\n\nI'm not sure that I feel very good about our benchmarks having to be \"adapted\" to be consumable for existing models while trading off the essence of the task (synthesizing information from multiple long documents to generate a long document). We should set the benchmarks as goals, see how current models and techniques perform on them, and allow them to be targets for future work.\n\n- I feel like the baselines are pretty weak, and the ReGen model is a pretty standard retrieve-and-generate baseline. I don't think there's too much modeling novelty here.\n\n- Furthermore, the query distribution seems pretty unnatural. For example, in the appendix, the query used is \"how google translate work\". This is a query with a pretty clear information-seeking / question intent that can be answered by a paragraph (or paragraphs). However, the wikipedia page for \"Google Translate\" answers far more than just this question, it contains a variety of general information about the query \"google translate\" (which itself could have many different question intents). So, I feel like the source of supervision provides a different sort of information than what a query like this would actually want (let alone that in practice, only the first passage is used, which contains even less information because of the way that lead sections in wikipedia are structured to contain general high-level information: https://en.wikipedia.org/wiki/Wikipedia:Manual_of_Style/Lead_section).\n  - The original motivation mentioned \"unseen factual queries\", but it's not really clear that generating an article is what you want for this. If it's factual, you can probably answer it with a paragraph or two. I think this work would benefit from a clearer exposition of what these queries would actually be, and what the right source of supervision / data for effectively addressing these queries would look like.", "clarity,_quality,_novelty_and_reproducibility": "This paper was reasonably clear, and I thought that the setup was pretty solid. It's an entry in a long line of work on generating Wikipedia paragraphs from references or other external knowledge.", "summary_of_the_review": "While I like where this paper is going, I feel like the practical artifact and problem tackled here are a bit too small-scale and narrow. We're at a point in NLP where things actually kind of (?) begin to work, and there have been a variety of impressive results on generating and understanding long documents. I think this work would really shine in pushing this direction, but instead, it compromises itself to fit the limitations of existing models, limiting its usefulness for future work. Thus, I'm not sure that I can recommend acceptance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666497473814}], "openreview_url": "https://openreview.net/forum?id=eiuj6cNv4iI", "arxiv_id": "2304.04358", "paper_pdf": "papers/eiuj6cNv4iI.pdf", "paper_pdf_sha256": "dbe08c549a562fb0fc4f10b109d72229cc374c4f43bdd0e4c7f56fa554826253", "paper_pdf_bytes": 793685, "paper_pdf_source": "openreview", "code_url": "https://github.com/qhjqhj00/WebBrain", "code_repository": "qhjqhj00/WebBrain", "code_commit": "d67b53e00bd9438f187080758dffc565d29bb649", "code_archive": "repos/eiuj6cNv4iI.zip", "code_archive_sha256": "becca44f5f1c9df861cb3ac84dc46e66a9c9c36f4011ebdeeea80eb87b4e9e5f", "code_archive_bytes": 402392, "code_file_count": 18, "code_extensions": {".py": 16, ".sh": 2}, "github_disk_usage_kb": 419, "github_languages": {"Python": 52602, "Shell": 529}, "github_archived": false, "github_pushed_at": "2023-06-07T09:43:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/webbrain-learning-to-generate-factually"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fyLvrx9M9YP", "year": 2022, "status": "rejected", "title": "Towards Unsupervised Content Disentanglement in Sentence Representations via Syntactic Roles", "authors": ["Ghazi Felhi", "Joseph Le Roux", "Djamé Seddah"], "authorids": ["~Ghazi_Felhi1", "~Joseph_Le_Roux1", "~Djamé_Seddah1"], "authors_source": "OpenReview API", "abstract": "Linking neural representations to linguistic factors is crucial in order to build and analyze NLP models interpretable by humans. Among these factors, syntactic roles (e.g. subjects, direct objects,$\\dots$)  and their realizations are essential markers since they can be understood as a decomposition of predicative structures and thus the meaning of sentences. Starting from a deep probabilistic generative model with attention, we measure the interaction between latent variables and realizations of syntactic roles, and show that it is possible to obtain, without supervision, representations of sentences where different syntactic roles correspond to clearly identified different latent variables. The probabilistic model we propose is an Attention-Driven Variational Autoencoder (ADVAE). Drawing inspiration from Transformer-based machine translation models, ADVAEs enable the analysis of the interactions between latent variables and input tokens through attention. We also develop an evaluation protocol to measure disentanglement with regard to the realizations of syntactic roles. This protocol is based on attention maxima for the encoder and on disturbing individual latent variables for the decoder. Our experiments on raw English text from the SNLI dataset show that $\\textit{i)}$ disentanglement of syntactic roles can be induced without supervision, $\\textit{ii)}$  ADVAE separates more syntactic roles than classical sequence VAEs, $\\textit{iii)}$ realizations of syntactic roles can be separately modified in sentences by mere intervention on the associated latent variables. Our work constitutes a first step towards unsupervised controllable content generation.  The code for our work is publicly available.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "l0u_Koq6he", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1407/Reviewer_wDkZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper propose a framework to obtain the disentanglement of syntactic roles as latent variables for sentence representations. The model is an attention-driven VAE which maps syntactic roles to separate latent variables using an encoder-decoder framework. In the second part of the paper, the authors introduce an evaluation protocol to quantify disentanglement between latent variables and spans both in the encoder and in the decoder, which includes syntactic role extraction, latent variable influence on decoder, encoder influence on latent variables and disentanglement metrics.", "review_text": "Strengths:\n- Learning the syntactic disentanglement in an unsupervised way.\n- Providing a series of evaluation protocol aimed at measuring the disentanglement of syntactic roles, which can be used by other work focus on syntactic disentanglement for sentence representations.\n\nWeakness:\n- The paper is hard to follow.\n- The evaluation protocol only contains intrinsic evaluation, but how these learned latent variables can help the downstream tasks is not clear. For example, whether this can help unsupervised dependency parsing?\n\nOverall, I think the paper provides an interesting perspective for learning syntactic disentanglement. However, I really suggest the authors to reorganize the paper. I cannot understand the main problem settings until I read the paper twice. To help the reader get the background knowledge, the authors can provide some related work in learning disentangled representations at the very beginning of the introduction (basically just remove some part in Section 6). Another issue is Figure 2, the caption does not indicate what is the input and output of the framework, which is illustrated in the following section instead, and the reader need to refer to the next part of the paper to understand the figure. There are some essential information in the appendix that really need to be moved in to the main body, e.g., some analysis of the latent variables and syntactic roles.\n\nSome questions:\n- How to determine the value of N_z? Is this related to the number of syntactic roles?\n- The syntactic roles used in evaluation protocol are from a dependency parsing instead of gold standard one.  Have you considered testing on texts that have annotated dependency structures?\n- What does the predicative structure in Figure 1 do with the syntactic roles discussed in the paper?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper propose a framework to obtain the disentanglement of syntactic roles as latent variables for sentence representations. The model is an attention-driven VAE which maps syntactic roles to separate latent variables using an encoder-decoder framework. In the second part of the paper, the authors introduce an evaluation protocol to quantify disentanglement between latent variables and spans both in the encoder and in the decoder, which includes syntactic role extraction, latent variable influence on decoder, encoder influence on latent variables and disentanglement metrics.", "main_review": "Strengths:\n- Learning the syntactic disentanglement in an unsupervised way.\n- Providing a series of evaluation protocol aimed at measuring the disentanglement of syntactic roles, which can be used by other work focus on syntactic disentanglement for sentence representations.\n\nWeakness:\n- The paper is hard to follow.\n- The evaluation protocol only contains intrinsic evaluation, but how these learned latent variables can help the downstream tasks is not clear. For example, whether this can help unsupervised dependency parsing?\n\nOverall, I think the paper provides an interesting perspective for learning syntactic disentanglement. However, I really suggest the authors to reorganize the paper. I cannot understand the main problem settings until I read the paper twice. To help the reader get the background knowledge, the authors can provide some related work in learning disentangled representations at the very beginning of the introduction (basically just remove some part in Section 6). Another issue is Figure 2, the caption does not indicate what is the input and output of the framework, which is illustrated in the following section instead, and the reader need to refer to the next part of the paper to understand the figure. There are some essential information in the appendix that really need to be moved in to the main body, e.g., some analysis of the latent variables and syntactic roles.\n\nSome questions:\n- How to determine the value of N_z? Is this related to the number of syntactic roles?\n- The syntactic roles used in evaluation protocol are from a dependency parsing instead of gold standard one.  Have you considered testing on texts that have annotated dependency structures?\n- What does the predicative structure in Figure 1 do with the syntactic roles discussed in the paper?", "summary_of_the_review": "As a final comment, this work does some contribution for learning and evaluating syntactic disentanglement for sentence representations, and will be helpful to the community. However, the writing of the paper should be improved. For detailed comments please refer to the main review.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635913429881}, {"id": "qEwAVPc4j7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1407/Reviewer_7uFL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "## Summary\n\n- This paper proposes a probabilistic model called Attention-Driven Variational Autoencoder (ADVAE). This model is another instance of $\\beta$-VAE whose encoder and encoders are composed of Transformers rather than previous neural architectures such as RNN.\n- The authors aim to disentangle the semantics of latent variables according to some syntactic roles (e.g., nouns and verbs) defined by syntax. To achieve this goal, they suggest employing the combination of Transformer and the existing $\\beta$-VAE framework which is known to be effective for disentangling the role of each latent variable in VAE.\n- Moreover, this work presents a new way of quantifying syntactic disentanglement between latent variables, relying on the information obtained from the attention matrices of the Transformer architecture.\n- The experiments show that the proposed method is quantitatively better than the normal VAE, and that swapping the value of a specific latent variable can impact the generation of the target word (decided by the syntactic role of the latent variable we choose).", "review_text": "\n## Strengths (Reasons to accept)\n- The paper is generally well-formed.\n- It proposes a new protocol that can be utilized for estimating the extent to which each latent variable is mapped to a specific syntactic role.\n\n## Weaknesses (Reasons to reject)\n\n- I'm not sure this work is the first that attempts to combine Transformers into the VAE framework. For instance, Li et al. (https://arxiv.org/pdf/2004.04092.pdf) already proposed a model consisting of pre-trained Transformer encoders and decoders.\nConsidering that the main novelty of this paper comes from the fact that the authors suggest exchanging the previous RNN to Transformer, it is doubtful this work has a meaningful novelty.\n- It is encouraged for the paper to self-contain the specification of the exact Transformer architecture (one with co-attention, Lu et al (2019)) utilized in this work.\n- There is no (theoretical) guarantee that the proposed method is specialized for assigning specific syntactic roles to latent variables, except that the method (by relying on the existing $\\beta$-VAE) encourages the latent variables to \"generally\" represent different aspects of sentences. Please explain why the proposed framework should be decent for disentangling \"syntactic roles\" rather than other linguistic or semantic properties.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "## Summary\n\n- This paper proposes a probabilistic model called Attention-Driven Variational Autoencoder (ADVAE). This model is another instance of $\\beta$-VAE whose encoder and encoders are composed of Transformers rather than previous neural architectures such as RNN.\n- The authors aim to disentangle the semantics of latent variables according to some syntactic roles (e.g., nouns and verbs) defined by syntax. To achieve this goal, they suggest employing the combination of Transformer and the existing $\\beta$-VAE framework which is known to be effective for disentangling the role of each latent variable in VAE.\n- Moreover, this work presents a new way of quantifying syntactic disentanglement between latent variables, relying on the information obtained from the attention matrices of the Transformer architecture.\n- The experiments show that the proposed method is quantitatively better than the normal VAE, and that swapping the value of a specific latent variable can impact the generation of the target word (decided by the syntactic role of the latent variable we choose).", "main_review": "\n## Strengths (Reasons to accept)\n- The paper is generally well-formed.\n- It proposes a new protocol that can be utilized for estimating the extent to which each latent variable is mapped to a specific syntactic role.\n\n## Weaknesses (Reasons to reject)\n\n- I'm not sure this work is the first that attempts to combine Transformers into the VAE framework. For instance, Li et al. (https://arxiv.org/pdf/2004.04092.pdf) already proposed a model consisting of pre-trained Transformer encoders and decoders.\nConsidering that the main novelty of this paper comes from the fact that the authors suggest exchanging the previous RNN to Transformer, it is doubtful this work has a meaningful novelty.\n- It is encouraged for the paper to self-contain the specification of the exact Transformer architecture (one with co-attention, Lu et al (2019)) utilized in this work.\n- There is no (theoretical) guarantee that the proposed method is specialized for assigning specific syntactic roles to latent variables, except that the method (by relying on the existing $\\beta$-VAE) encourages the latent variables to \"generally\" represent different aspects of sentences. Please explain why the proposed framework should be decent for disentangling \"syntactic roles\" rather than other linguistic or semantic properties.\n\n", "summary_of_the_review": "To summarize, even though this paper demonstrates that Transformers with $\\beta$-VAE can be an attractive option for generating latent variables representing some syntactic roles, its novelty is generally limited to the fact that it proposed a new evaluation protocol aimed at estimating the extent to which each latent variable is mapped to a specific syntactic role.\nTherefore, my suggestion is a weak reject.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635905482289}, {"id": "0tiHMEr6SmX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1407/Reviewer_9pDc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a method for unsupervised disentanglement of text components and shows its ability to identify semantic roles.\nTo this end, a neural network is trained to compress the input into a fixed number of independent latent variables which are regularized to be standard Gaussians via the VAE framework. The inference network consists of a Transformer encoder-decoder network, where the decoder inputs correspond to the latent variables, which cross-attend to the outputs of the Transformer encoder, i.e., the encoded sentence. The idea behind this architecture is that attention-based seq2seq architectures align source and target sequences with each other.\n\nThe model is evaluated on disentanglement of semantic roles. To this end, they investigate how resampling of individual latent variables impacts the semantic roles in the text generated by the decoder, and how syntactic roles are aggregated into latent vectors via attention. They find that their proposed architecture is more successful at disentangling semantic roles into the latent variables than standard VAEs.", "review_text": "Strengths:\n\n* The paper studies a research question that is at the heart of representation learning, which could potentially be very impactful\n* The model is based on an intuitive and simple idea\n* The results are convincing and well presented\n\nWeaknesses:\n\n* No ablation studies. In my opinion, the results are not actually conclusive about what model components are responsible for the improvements. For example, the VAE baseline is based on LSTM encoder-decoder models, whereas the proposed model employs Transformers. Moreover, the VAE baseline seems to be significantly smaller than the Tranformer. Are these two models really comparable enough to conclude that the improvements are due to the product of Gaussians formulation? I think a fairer comparison with minimal changes between the baseline and proposed model could be made as follows: Pool the outputs of 2b) before estimating a single Gaussian distribution, which can then be used as a standard VAE.\n\n* The concrete architecture is not well motivated: It is not clear to me why the inference network architecture has to look the way it does in 2b). Does the model actually use co-attention (i.e., sentence representations and latent variables serve as queries for each other)? Figure 2b) looks like a vanilla Transformer encoder-decoder architecture to me, which uses cross-attention, not co-attention. Moreover, it is not clear that the model actually needs self-attention to refine the latent variables. Wouldn't the model still work if the learnable $e_{z_i}^{enc}$ functioned merely as queries to a single additional attention function on top of the Transformer encoder (followed by a linear layer and softplus)?\n\n* Some related work is not discussed. Specifically, Behjati and Henderson (2021) propose a very similar architecture for learning meaningful units in text: They train an autoencoder with a fixed number of latent vectors via slot attention, which are supposed to capture morphemes. Could their model be applied to disentangle semantic roles, too? In general, how does your work compare to slot attention (Locatello et al. (2020))\n\nMinor:\n* Table captions should appear above the table according to the style guide\n\nLocatello et al. (2020): https://arxiv.org/abs/2006.15055\n\nBehjati and Henderson (2021): https://arxiv.org/pdf/2102.01223.pdf", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a method for unsupervised disentanglement of text components and shows its ability to identify semantic roles.\nTo this end, a neural network is trained to compress the input into a fixed number of independent latent variables which are regularized to be standard Gaussians via the VAE framework. The inference network consists of a Transformer encoder-decoder network, where the decoder inputs correspond to the latent variables, which cross-attend to the outputs of the Transformer encoder, i.e., the encoded sentence. The idea behind this architecture is that attention-based seq2seq architectures align source and target sequences with each other.\n\nThe model is evaluated on disentanglement of semantic roles. To this end, they investigate how resampling of individual latent variables impacts the semantic roles in the text generated by the decoder, and how syntactic roles are aggregated into latent vectors via attention. They find that their proposed architecture is more successful at disentangling semantic roles into the latent variables than standard VAEs.", "main_review": "Strengths:\n\n* The paper studies a research question that is at the heart of representation learning, which could potentially be very impactful\n* The model is based on an intuitive and simple idea\n* The results are convincing and well presented\n\nWeaknesses:\n\n* No ablation studies. In my opinion, the results are not actually conclusive about what model components are responsible for the improvements. For example, the VAE baseline is based on LSTM encoder-decoder models, whereas the proposed model employs Transformers. Moreover, the VAE baseline seems to be significantly smaller than the Tranformer. Are these two models really comparable enough to conclude that the improvements are due to the product of Gaussians formulation? I think a fairer comparison with minimal changes between the baseline and proposed model could be made as follows: Pool the outputs of 2b) before estimating a single Gaussian distribution, which can then be used as a standard VAE.\n\n* The concrete architecture is not well motivated: It is not clear to me why the inference network architecture has to look the way it does in 2b). Does the model actually use co-attention (i.e., sentence representations and latent variables serve as queries for each other)? Figure 2b) looks like a vanilla Transformer encoder-decoder architecture to me, which uses cross-attention, not co-attention. Moreover, it is not clear that the model actually needs self-attention to refine the latent variables. Wouldn't the model still work if the learnable $e_{z_i}^{enc}$ functioned merely as queries to a single additional attention function on top of the Transformer encoder (followed by a linear layer and softplus)?\n\n* Some related work is not discussed. Specifically, Behjati and Henderson (2021) propose a very similar architecture for learning meaningful units in text: They train an autoencoder with a fixed number of latent vectors via slot attention, which are supposed to capture morphemes. Could their model be applied to disentangle semantic roles, too? In general, how does your work compare to slot attention (Locatello et al. (2020))\n\nMinor:\n* Table captions should appear above the table according to the style guide\n\nLocatello et al. (2020): https://arxiv.org/abs/2006.15055\n\nBehjati and Henderson (2021): https://arxiv.org/pdf/2102.01223.pdf", "summary_of_the_review": "While the research presented in this study is very exciting, it is not quite convincing enough yet for me to trust that the results are actually due to the modeling decisions presented. Moreover, a similar work (Behjati and Henderson (2020)) is not discussed, which casts doubt on the novelty of the approach. I therefore recommend to reject the paper in its current state, but I am happy to change my scores upwards if my concerns above can be resolved.\n\n* UPDATE:  I raised my score slightly as a consequence of the rebuttal phase, where authors conducted an ablation study that now more clearly shows that the improvements over standard VAEs are due to multi-vector nature of the representation rather than just because of using Transformers.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635894693068}, {"id": "yqQ4iEiV2lk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1407/Reviewer_WuPD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a new model ADVAE, which uses a sequence of latent variables which are constructed using cross-attention, which are then used to condition the inference model. The findings show that certain latent variables correlated with different syntactic roles (measured using a dependency parser).  The work claims that the latent variables, in this case, are able to disentangle content effectively, which I am inclined to agree with, however, I feel that the experimental setup is lacking to solidly support this hypothesis (which I detail in the “cons” and “questions” section). ", "review_text": "Pros:\n- The insight that cross-attention disentangles content automatically in Transformers is interesting, and I feel that this insight could be exploited for a variety of controllable text generation tasks. \n\nCons:\n- The baseline VAE is not a proper comparison, given that an LSTM is used (versus a Transformer in the proposed model). Please either make the proposed model LSTM-based or the baseline a Transformer.\n- I feel that the ADVAE, while explained relatively clearly in the figure, could benefit from a step-by-step explanation in the text. I found it slightly confusing to follow section 3, which could benefit from a re-organization: for example, 1st introduce the operations happening in Fig 2b, describe the latent variable generation procedure, and then what happens in Fig 2c (reconstruction). An example of a paper that I feel does this well is Zhang et al., 2016.\n- I also feel that the impact of increasing/decreasing N_z is not discussed enough (e.g. if you increase the N_z to become the maximum sequence length, what happens, etc…).\n- Overall, I feel it would be pertinent to test how well different aspects (not just syntactic roles) are disentangled. Rather than the method per se, I feel that this insight into unsupervised disentanglement is interesting, however, constraining the method to syntactic roles can be limiting. For example, for the “low complexity” datasets (e.g. sentiment analysis), it would be interesting to see when running your method, is there a vector that emerges that can control sentiment, or is it able to both disentangle syntactic roles and sentiment, etc….? (As far as I know, John et al. 2019, did this in a supervised manner).\n\nMinor comments/missing references:\n- Machine Translation/MT Transformer -> Transformer encoder-decoder models (as this architecture is widely used, even in non-MT tasks)\n- Sentence representations have been extracted from sequence-to-sequence Transformer models before (Lewis et al., 2020, Raffel et al., 2020, Siddhant et al., 2020)\n- The [CLS] representation is generally the default sentence representation in BERT (not [SEP])\n- p_r (the dependency parser?) is not defined.\n\nQuestions/Other comments:\n- It would be interesting to compare the variational model to a denoising AE (seq2seq masked language model; Lewis et al., 2020), in which the input is corrupted and the prior distribution is removed). This could make the method more generally applicable, even to large pre-trained models.\n- Furthermore, I am sceptical of the use of self-attention when computing the mean and standard deviation vectors (Fig 1b). I would be curious what would happen if only cross-attention would be used. If that is the case, then the argument could be made that the original e^{enc} vectors correspond to a syntactic role (or other factors) themselves.\n- Also, it would be curious if this emerges in machine translation (given the motivation stated as such), where the source and target represent the same sentence in different languages.\n---\n\nLewis, Mike et al. “BART: Denoising Sequence-to-Sequence Pre-Training for Natural Language Generation, Translation, and Comprehension.” Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (2020): n. pag. Crossref. Web.\n\nSiddhant, Aditya et al. “Evaluating the Cross-Lingual Effectiveness of Massively Multilingual Neural Machine Translation.” Proceedings of the AAAI Conference on Artificial Intelligence 34.05 (2020): 8854–8861. Crossref. Web.\n\nJohn, Vineet et al. “Disentangled Representation Learning for Non-Parallel Text Style Transfer.” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (2019): n. pag. Crossref. Web.\n\nRaffel, Colin et al. “Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer.” ArXiv abs/1910.10683 (2020): n. pag.\n\nZhang, Biao et al. “Variational Neural Machine Translation.” Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (2016)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new model ADVAE, which uses a sequence of latent variables which are constructed using cross-attention, which are then used to condition the inference model. The findings show that certain latent variables correlated with different syntactic roles (measured using a dependency parser).  The work claims that the latent variables, in this case, are able to disentangle content effectively, which I am inclined to agree with, however, I feel that the experimental setup is lacking to solidly support this hypothesis (which I detail in the “cons” and “questions” section). ", "main_review": "Pros:\n- The insight that cross-attention disentangles content automatically in Transformers is interesting, and I feel that this insight could be exploited for a variety of controllable text generation tasks. \n\nCons:\n- The baseline VAE is not a proper comparison, given that an LSTM is used (versus a Transformer in the proposed model). Please either make the proposed model LSTM-based or the baseline a Transformer.\n- I feel that the ADVAE, while explained relatively clearly in the figure, could benefit from a step-by-step explanation in the text. I found it slightly confusing to follow section 3, which could benefit from a re-organization: for example, 1st introduce the operations happening in Fig 2b, describe the latent variable generation procedure, and then what happens in Fig 2c (reconstruction). An example of a paper that I feel does this well is Zhang et al., 2016.\n- I also feel that the impact of increasing/decreasing N_z is not discussed enough (e.g. if you increase the N_z to become the maximum sequence length, what happens, etc…).\n- Overall, I feel it would be pertinent to test how well different aspects (not just syntactic roles) are disentangled. Rather than the method per se, I feel that this insight into unsupervised disentanglement is interesting, however, constraining the method to syntactic roles can be limiting. For example, for the “low complexity” datasets (e.g. sentiment analysis), it would be interesting to see when running your method, is there a vector that emerges that can control sentiment, or is it able to both disentangle syntactic roles and sentiment, etc….? (As far as I know, John et al. 2019, did this in a supervised manner).\n\nMinor comments/missing references:\n- Machine Translation/MT Transformer -> Transformer encoder-decoder models (as this architecture is widely used, even in non-MT tasks)\n- Sentence representations have been extracted from sequence-to-sequence Transformer models before (Lewis et al., 2020, Raffel et al., 2020, Siddhant et al., 2020)\n- The [CLS] representation is generally the default sentence representation in BERT (not [SEP])\n- p_r (the dependency parser?) is not defined.\n\nQuestions/Other comments:\n- It would be interesting to compare the variational model to a denoising AE (seq2seq masked language model; Lewis et al., 2020), in which the input is corrupted and the prior distribution is removed). This could make the method more generally applicable, even to large pre-trained models.\n- Furthermore, I am sceptical of the use of self-attention when computing the mean and standard deviation vectors (Fig 1b). I would be curious what would happen if only cross-attention would be used. If that is the case, then the argument could be made that the original e^{enc} vectors correspond to a syntactic role (or other factors) themselves.\n- Also, it would be curious if this emerges in machine translation (given the motivation stated as such), where the source and target represent the same sentence in different languages.\n---\n\nLewis, Mike et al. “BART: Denoising Sequence-to-Sequence Pre-Training for Natural Language Generation, Translation, and Comprehension.” Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (2020): n. pag. Crossref. Web.\n\nSiddhant, Aditya et al. “Evaluating the Cross-Lingual Effectiveness of Massively Multilingual Neural Machine Translation.” Proceedings of the AAAI Conference on Artificial Intelligence 34.05 (2020): 8854–8861. Crossref. Web.\n\nJohn, Vineet et al. “Disentangled Representation Learning for Non-Parallel Text Style Transfer.” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (2019): n. pag. Crossref. Web.\n\nRaffel, Colin et al. “Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer.” ArXiv abs/1910.10683 (2020): n. pag.\n\nZhang, Biao et al. “Variational Neural Machine Translation.” Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (2016)", "summary_of_the_review": "Overall, although I like intuition and motivation, however, I feel the empirical support is lacking. I would be willing to raise my score if experiments on other aspects or in other languages were done. I believe this would further reinforce the claims made in the paper regarding unsupervised disentanglement.\n\n---\nPost rebuttal:\n\nI appreciate the additional experiments the authors provided and their time and effort in writing the rebuttal. This has clarified some of my concerns and added extra evidence to the paper. For this reason, I will raise my score to a 5. Despite this, I still feel that a more general approach, able to disentangle different factors, or in different languages would add more substance that I feel this paper is lacking.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635799807156}], "openreview_url": "https://openreview.net/forum?id=fyLvrx9M9YP", "arxiv_id": "2206.11184", "paper_pdf": "papers/fyLvrx9M9YP.pdf", "paper_pdf_sha256": "49c26a7bd524b87c146e5fa39e5a4bd1bc1e95f12835d2f8c3dc13fa9b4ef078", "paper_pdf_bytes": 731872, "paper_pdf_source": "openreview", "code_url": "https://github.com/ghazi-f/ADVAE", "code_repository": "ghazi-f/ADVAE", "code_commit": "cbf1c9d542eb826c7599425b32bd701e1a105bd3", "code_archive": "repos/fyLvrx9M9YP.zip", "code_archive_sha256": "f3cccd3558957a005e74745293c243fc84f7612968967682fe9fbe30363802e4", "code_archive_bytes": 59477, "code_file_count": 12, "code_extensions": {".py": 11, ".sh": 1}, "github_disk_usage_kb": 1719, "github_languages": {"Python": 365572, "Shell": 3848}, "github_archived": false, "github_pushed_at": "2022-07-05T09:22:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-unsupervised-content-disentanglement-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rJlUhhVYvS", "year": 2020, "status": "rejected", "title": "Understanding Isomorphism Bias in Graph Data Sets ", "authors": ["Ivanov Sergey", "Sviridov Sergey", "Evgeny Burnaev"], "authorids": ["ivanovserg990@gmail.com", "sergei.sviridov@gmail.com", "e.burnaev@skoltech.ru"], "authors_source": "OpenReview API", "abstract": "In recent years there has been a rapid increase in classification methods on graph structured data. Both in graph kernels and graph neural networks, one of the implicit assumptions of successful state-of-the-art models was that incorporating graph isomorphism features into the architecture leads to better empirical performance. However, as we discover in this work, commonly used data sets for graph classification have repeating instances which cause the problem of isomorphism bias, i.e. artificially increasing the accuracy of the models by memorizing target information from the training set. This prevents fair competition of the algorithms and raises a question of the validity of the obtained results. We analyze 54 data sets, previously extensively used for graph-related tasks, on the existence of isomorphism bias, give a set of recommendations to machine learning practitioners to properly set up their models, and open source new data sets for the future experiments. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SJlFHnO6cr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper190/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work probes graph classification datasets for isomorphism bias. They find substantial amount of bias in some datasets and show that they suffer from data leakage. They further perform a more fine-grained evaluation taking into consideration the node/edge types which reduce the perceived effects. They also provide some recommendations for measuring the 'right metrics' and release clean versions of the considered datasets.\n\nStrengths:\n- The methodology is rigorous and the datasets considered is extensive\n- The paper is well written\n\nConcerns:\n- Isomorphism is not necessarily a bad thing in graph classification tasks. Especially in chemistry where a bond decides if a compound is poisonous or not. Also, as the authors themselves mention, taking node/edge labels decrease the isomorphism in most datasets.\n- The results and recommendations presented in the paper are intuitive and somewhat trivial\n- I am not sure if ICLR is the right venue for this work", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "3: Weak Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #4", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "This work probes graph classification datasets for isomorphism bias. They find substantial amount of bias in some datasets and show that they suffer from data leakage. They further perform a more fine-grained evaluation taking into consideration the node/edge types which reduce the perceived effects. They also provide some recommendations for measuring the 'right metrics' and release clean versions of the considered datasets.\n\nStrengths:\n- The methodology is rigorous and the datasets considered is extensive\n- The paper is well written\n\nConcerns:\n- Isomorphism is not necessarily a bad thing in graph classification tasks. Especially in chemistry where a bond decides if a compound is poisonous or not. Also, as the authors themselves mention, taking node/edge labels decrease the isomorphism in most datasets.\n- The results and recommendations presented in the paper are intuitive and somewhat trivial\n- I am not sure if ICLR is the right venue for this work"}, "tcdate": 1572863041487}, {"id": "BJxu9Rwn9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper190/AnonReviewer2"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper is concerned with the presence of isomorphism bias in commonly used graph learning benchmarks. In particular, the paper analyzes the amount of isomorphic graphs in 54 graph datasets and evaluates the performance of three graph classification methods under two isomorphism settings.\n\nCareful analyses of commonly used benchmarks can be important contributions that provide new insights into the performance of state-of-the-art models. The present paper's results on graph isomorphism properties can indeed be valuable for the ablation of models and testing their performance with regard to this property. I also found the relatively high label disagreements on some datasets (even under stronger isomorphism constraints) to be a surprising and useful result.\n\nHowever, the main assumption of the paper -- which equates the quality of a graph learning benchmark with the amount of isomorphic graphs that it contains, i.e., the lower the better -- seems questionable.\n\nThe paper argues that isomorphic graphs are akin to duplicate images in computer vision and should be removed from a dataset. While completely identical graphs are certainly problematic, the case seems different for isomorphic graphs. In the latter, a learning method is required to identify the correct bijection form V_1 to V_2 which is a non-trivial task. Testing on isomorphic graphs evaluates the ability of a model to infer these equivalence classes from data which is an important property. Moreover, being able to capture the equivalence relation can be important for various graph learning tasks, e.g., to facilitate that two topologically equivalent graphs are be classified similarly.  Going back to the computer vision analogy: it seems a more adequate comparison for graph isomorphism would be translation and scale invariance which are certainly desirable properties for CV models.\n\nIn addition, the dataset analysis could also be improved. For instance, the SYNTHETIC dataset includes continuous node attributes that are essential for classification and make the graphs non-isomorphic (when considering, for instance, each attribute vector as a unique node label). However, the attributes are not considered in the analysis what leads to a large number of isomorphic graphs. On a side note: the paper also incorrectly attributes the SYNTHETIC dataset to (Morris et al, 2016), but it is in fact from [1]. The synthetic dataset of Morris et al (SYNTHIE) does not consist of isomorphic graphs, while the SYNTHETIC dataset of [1] does so intentionally.\n\nThe results of Section 5 seem also not very surprising: After removing node labels, it is expected that the number of isomorphic graphs increases since a discriminating feature has been removed. Moreover, when accounting for node labels, many standard benchmarks seem to consist of significantly less isomorphic and mismatched graphs (as can be seen in the appendix).\n\nSince graph isomorphism != graph identity, the assumption Y_iso \\sub Y_train in Property 6 seems also not appropriate. The results of Theorem 6.1 on the other hand seems straightforward and would hold for any classification task for which the true label for an equivalence class of instances is known.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "The paper is concerned with the presence of isomorphism bias in commonly used graph learning benchmarks. In particular, the paper analyzes the amount of isomorphic graphs in 54 graph datasets and evaluates the performance of three graph classification methods under two isomorphism settings.\n\nCareful analyses of commonly used benchmarks can be important contributions that provide new insights into the performance of state-of-the-art models. The present paper's results on graph isomorphism properties can indeed be valuable for the ablation of models and testing their performance with regard to this property. I also found the relatively high label disagreements on some datasets (even under stronger isomorphism constraints) to be a surprising and useful result.\n\nHowever, the main assumption of the paper -- which equates the quality of a graph learning benchmark with the amount of isomorphic graphs that it contains, i.e., the lower the better -- seems questionable.\n\nThe paper argues that isomorphic graphs are akin to duplicate images in computer vision and should be removed from a dataset. While completely identical graphs are certainly problematic, the case seems different for isomorphic graphs. In the latter, a learning method is required to identify the correct bijection form V_1 to V_2 which is a non-trivial task. Testing on isomorphic graphs evaluates the ability of a model to infer these equivalence classes from data which is an important property. Moreover, being able to capture the equivalence relation can be important for various graph learning tasks, e.g., to facilitate that two topologically equivalent graphs are be classified similarly.  Going back to the computer vision analogy: it seems a more adequate comparison for graph isomorphism would be translation and scale invariance which are certainly desirable properties for CV models.\n\nIn addition, the dataset analysis could also be improved. For instance, the SYNTHETIC dataset includes continuous node attributes that are essential for classification and make the graphs non-isomorphic (when considering, for instance, each attribute vector as a unique node label). However, the attributes are not considered in the analysis what leads to a large number of isomorphic graphs. On a side note: the paper also incorrectly attributes the SYNTHETIC dataset to (Morris et al, 2016), but it is in fact from [1]. The synthetic dataset of Morris et al (SYNTHIE) does not consist of isomorphic graphs, while the SYNTHETIC dataset of [1] does so intentionally.\n\nThe results of Section 5 seem also not very surprising: After removing node labels, it is expected that the number of isomorphic graphs increases since a discriminating feature has been removed. Moreover, when accounting for node labels, many standard benchmarks seem to consist of significantly less isomorphic and mismatched graphs (as can be seen in the appendix).\n\nSince graph isomorphism != graph identity, the assumption Y_iso \\sub Y_train in Property 6 seems also not appropriate. The results of Theorem 6.1 on the other hand seems straightforward and would hold for any classification task for which the true label for an equivalence class of instances is known."}, "tcdate": 1572794000134}, {"id": "SJeT-mmc5S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper190/AnonReviewer5"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents three contributions: (a) the observation that there’s train-to-test leakage in many graph classification datasets (under isomorphism equivalence), (b) what appears to be a theoretically motivated way of improving scores on such datasets, by focusing on solving the examples that are isomorphic with training instances, and (c) a recommendation to remove such leakage from test sets. I don’t think the paper meets the ICLR bar. While (a) is very interesting, and an important contribution, (b) and (c) are contradictory. The recommendation (c) is a bit of a no-brainer, and Property 6.1 and Theorem 6.1, providing the substance of (b), are near-trivial. \n\nMissing reference: Bordes et al. (2013) and Toutanova et al. (2015) show there’s train-to-test leakage (under isomorphism equivalence) in the FB15K dataset. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have read many papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I did not assess the experiments.", "title": "Official Blind Review #5", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review": "The paper presents three contributions: (a) the observation that there’s train-to-test leakage in many graph classification datasets (under isomorphism equivalence), (b) what appears to be a theoretically motivated way of improving scores on such datasets, by focusing on solving the examples that are isomorphic with training instances, and (c) a recommendation to remove such leakage from test sets. I don’t think the paper meets the ICLR bar. While (a) is very interesting, and an important contribution, (b) and (c) are contradictory. The recommendation (c) is a bit of a no-brainer, and Property 6.1 and Theorem 6.1, providing the substance of (b), are near-trivial. \n\nMissing reference: Bordes et al. (2013) and Toutanova et al. (2015) show there’s train-to-test leakage (under isomorphism equivalence) in the FB15K dataset. "}, "tcdate": 1572643589108}, {"id": "SklrAvRctS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper190/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors discuss here the problem of isomorphism bias in graph dataset, i.e. the overfitting effect in learning networks whenever graph isomorphism features are incorporated within the model. This is a bias which jeopardises the validity and the reproducibility of several studies, and it is theoretically analogous to data leakage effects.\nThe authors fairly discuss the problem in the introduction, with a good coverage of the related literature; the background theory is reasonably discussed, although is not very deep. The experimental part is extensive and well described, and it shows the overfitting effect very clearly. However, the novelty of the work is limited, and also the proposed solutions cannot be claimed as superior to other approaches, due to the small improvement in accuracy.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review": "The authors discuss here the problem of isomorphism bias in graph dataset, i.e. the overfitting effect in learning networks whenever graph isomorphism features are incorporated within the model. This is a bias which jeopardises the validity and the reproducibility of several studies, and it is theoretically analogous to data leakage effects.\nThe authors fairly discuss the problem in the introduction, with a good coverage of the related literature; the background theory is reasonably discussed, although is not very deep. The experimental part is extensive and well described, and it shows the overfitting effect very clearly. However, the novelty of the work is limited, and also the proposed solutions cannot be claimed as superior to other approaches, due to the small improvement in accuracy."}, "tcdate": 1571641292951}], "openreview_url": "https://openreview.net/forum?id=rJlUhhVYvS", "arxiv_id": "1910.12091", "paper_pdf": "papers/rJlUhhVYvS.pdf", "paper_pdf_sha256": "515f5c67fbbcb82182b619125237be087af6016c3b1ad49fd3171968a87cdd25", "paper_pdf_bytes": 429170, "paper_pdf_source": "openreview", "code_url": "https://github.com/nd7141/iso_bias", "code_repository": "nd7141/iso_bias", "code_commit": "a235a71814c4b0e644a7c0d66a16980332005327", "code_archive": "repos/rJlUhhVYvS.zip", "code_archive_sha256": "84c13ad3773c57d7b15e31c808f62c92b4e0e937b72d02fe531efb5a9af31580", "code_archive_bytes": 4457038, "code_file_count": 19, "code_extensions": {".py": 13, ".c": 3, ".h": 1, ".cc": 1, ".ipynb": 1}, "github_disk_usage_kb": 5877, "github_languages": {"Jupyter Notebook": 2735667, "Python": 91041, "C++": 31197, "C": 15150, "Makefile": 267}, "github_archived": false, "github_pushed_at": "2019-12-12T16:43:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-isomorphism-bias-in-graph-data"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bhPaXhWVKG", "year": 2026, "status": "rejected", "title": "MermaidFlow: Redefining Agentic Workflow Generation via Safety-Constrained Evolutionary Programming", "authors": ["Chengqi Zheng", "Yewen Pu", "Jianda Chen", "Yueming Lyu", "Wen zheng terence Ng", "Haopeng Zhang", "Yew-Soon Ong", "Ivor Tsang", "Haiyan Yin"], "authorids": ["~Chengqi_Zheng1", "~Yewen_Pu1", "~Jianda_Chen1", "~Yueming_Lyu1", "~Wen_zheng_terence_Ng1", "~Haopeng_Zhang3", "~Yew-Soon_Ong1", "~Ivor_Tsang1", "~Haiyan_Yin1"], "authors_source": "OpenReview API", "abstract": "Despite the promise of autonomous agentic reasoning, existing workflow generation methods frequently produce fragile, unexecutable plans due to unconstrained LLM-driven construction. We propose MermaidFlow, a framework that redefines the agentic search space through safety-constrained graph evolution. At its core, MermaidFlow represent workflows as a verifiable intermediate representation using Mermaid, a structured and human-interpretable graph language. We formulate domain-aware evolutionary operators, i.e., crossover,  mutation, insertion, and deletion, to preserve semantic correctness while promoting structural diversity, enabling efficient exploration of a high-quality, statically verifiable workflow space. Without modifying task settings or evaluation protocols, MermaidFlow achieves consistent improvements in success rates and faster convergence to executable plans on the agent reasoning benchmark. The experimental results demonstrate that safety-constrained graph evolution offers a scalable, modular foundation for robust and interpretable agentic reasoning systems.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "dOZSY9pSjD", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24482/Reviewer_3qM1"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes MermaidFlow, a framework for agentic workflow generation based on the Mermaid graph language. The core idea is to represent workflows as declarative typed graphs G(V[τ,α], E[ρ]) and optimize them through safety-constrained evolutionary programming (EP). The authors claim that Mermaid representation provides advantages such as static verifiability, human readability, and modularity. Experiments are conducted on GSM8K, MATH, HumanEval, and MBPP, showing an average improvement of 2.08% over AFlow with approximately 50% reduction in token cost.", "review_text": "This paper proposes MermaidFlow, a framework for agentic workflow generation based on the Mermaid graph language. The core idea is to represent workflows as declarative typed graphs G(V[τ,α], E[ρ]) and optimize them through safety-constrained evolutionary programming (EP). The authors claim that Mermaid representation provides advantages such as static verifiability, human readability, and modularity. Experiments are conducted on GSM8K, MATH, HumanEval, and MBPP, showing an average improvement of 2.08% over AFlow with approximately 50% reduction in token cost.", "strengths": "1. **Clear formalization**: The graph representation of workflows, type system, and evolutionary operators are well-defined\n2. **Static verification mechanism**: Two-layer checking (soft + hard) ensures syntactic correctness of generated workflows\n3. **Comprehensive experiments**: Cover multiple domains including math reasoning and code generation, with comparisons against multiple baselines\n4. **Token efficiency**: Approximately 50% reduction in token cost compared to AFlow\n5. **Case study**: Figure 4 clearly demonstrates the workflow evolution process", "weaknesses": "**1. Limited Novelty**\n\n- **Essentially a variant of known paradigms**: The method still follows a per-task iterative evolutionary search paradigm, introducing new representation and constraints on top of existing frameworks.\n- **Questionable necessity of Mermaid**: The paper does not sufficiently justify why Mermaid is superior to Python. For LLMs, both are structured text, but LLMs have higher affinity for code representations. Moreover, current LLM-based workflow generation is far from the stage where \"constraint benefits > exploration benefits\", making the actual gains from safety constraints (efficiency? effectiveness?) unclear. Workflow failures primarily stem from early systems (like ADAS) building from scratch with inadequate controllable and supervised generation granularity, rather than inherent issues with Python representation itself.\n\n\n**2. Experimental Design Issues**\n\n- **Marginal performance gains**: Average improvements of 2.08% over AFlow and 1.40% over MaAS. Given that agentic workflows can achieve high upper bounds through automated structural design, the paper fails to demonstrate breakthrough on apparent bottlenecks (e.g., AFlow's restricted XML generation format causing significant performance degradation on certain tasks, and the inability to reuse good designs from non-selected nodes—the simplest example being that poor workflow structures may contain good prompts). Moreover, I noticed in the appendix that the evolved MermaidFlow ensemble uses 5-sample voting while AFlow uses 3-sample voting, which could be a major contributing factor to the performance improvement rather than the Mermaid representation itself.\n- **Unfair cost comparison**: The reported token consumption of MermaidFlow does not clearly specify whether it includes the Mermaid→Python translation step. According to the experimental setup, each generated Mermaid workflow requires translation via gpt-4o-mini. This cost should be explicitly stated regarding inclusion, along with the specific cost breakdown for this component alone.\n\n\n**3. Technical Detail Issues**\n\n- **Unclear role of the Checker**: If Mermaid provides strong type constraints, the checker should rarely trigger, as LLMs have very low probability of generating syntactically invalid Python code. However, the checker is actually a critical component in MermaidFlow (Section A.2), suggesting that the constraints are not strong enough. A >90% success rate still means 10% of generations require retry, which remains quite high.\n- **Incomplete operator definitions**: The paper claims \"safety-constrained\", but specific safety properties (such as deadlock-freedom, termination) are not defined. There is also a lack of case study analysis on how directly generating Python code workflows violates these safety properties.\n\n\n**4. Inappropriate Baseline Categorization (Minor)**\n\n- **Conceptually flawed taxonomy**: The authors categorize CoT, ComplexCoT, and Self-Consistency as \"Single-agent execution methods\" (Section 5.1, Table 1). This classification is deeply problematic. The term \"agentic workflow\" was introduced precisely to distinguish workflows with fixed resource scheduling from agents capable of autonomously allocating computational resources—with LLM calls being the most representative computational resource. CoT is a classic prompting strategy, while Self-Consistency is a classic (non-agentic) workflow. Neither involves autonomous agents that can dynamically decide when and how to invoke LLMs based on intermediate states or task requirements. By labeling these non-agent baselines as \"single-agent methods,\" the authors conflate fundamentally different paradigms and misrepresent the conceptual boundaries of their contribution.", "questions": "1. Could you provide more concrete case studies, including but not limited to:\n   - **Mermaid vs. Python comparison**: Show side-by-side examples where (a) Python-based workflow generation fails but Mermaid succeeds, and (b) demonstrate what specific safety properties are violated in the Python case (e.g., deadlock, non-termination, type mismatch).\n   - **Safety constraint effectiveness**: Provide concrete examples showing how safety constraints improve efficiency or effectiveness during evolution. What specific invalid mutations are prevented? How much retry overhead is avoided?\n   - **Checker analysis**: Given that the checker has >90% success rate (implying ~10% retry), provide examples of the 10% failed cases. What constraint violations occur? Why doesn't Mermaid's type system prevent these?\n   - **Component reuse**: Demonstrate how MermaidFlow enables reusing good components (e.g., effective prompts) from suboptimal workflows, addressing the AFlow limitation you mentioned. Show concrete examples from your evolution process.\n\n2. Could you provide a more detailed cost analysis?\n   - Explicitly state whether the reported token consumption includes Mermaid→Python translation costs\n   - Break down token costs by component: (a) workflow generation, (b) translation, (c) execution, (d) evaluation\n   - Compare the per-iteration cost breakdown between MermaidFlow and AFlow\n\n3. I would greatly appreciate if you could discuss the concerns I raised in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes MermaidFlow, a framework for agentic workflow generation based on the Mermaid graph language. The core idea is to represent workflows as declarative typed graphs G(V[τ,α], E[ρ]) and optimize them through safety-constrained evolutionary programming (EP). The authors claim that Mermaid representation provides advantages such as static verifiability, human readability, and modularity. Experiments are conducted on GSM8K, MATH, HumanEval, and MBPP, showing an average improvement of 2.08% over AFlow with approximately 50% reduction in token cost.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. **Clear formalization**: The graph representation of workflows, type system, and evolutionary operators are well-defined\n2. **Static verification mechanism**: Two-layer checking (soft + hard) ensures syntactic correctness of generated workflows\n3. **Comprehensive experiments**: Cover multiple domains including math reasoning and code generation, with comparisons against multiple baselines\n4. **Token efficiency**: Approximately 50% reduction in token cost compared to AFlow\n5. **Case study**: Figure 4 clearly demonstrates the workflow evolution process", "weaknesses": "**1. Limited Novelty**\n\n- **Essentially a variant of known paradigms**: The method still follows a per-task iterative evolutionary search paradigm, introducing new representation and constraints on top of existing frameworks.\n- **Questionable necessity of Mermaid**: The paper does not sufficiently justify why Mermaid is superior to Python. For LLMs, both are structured text, but LLMs have higher affinity for code representations. Moreover, current LLM-based workflow generation is far from the stage where \"constraint benefits > exploration benefits\", making the actual gains from safety constraints (efficiency? effectiveness?) unclear. Workflow failures primarily stem from early systems (like ADAS) building from scratch with inadequate controllable and supervised generation granularity, rather than inherent issues with Python representation itself.\n\n\n**2. Experimental Design Issues**\n\n- **Marginal performance gains**: Average improvements of 2.08% over AFlow and 1.40% over MaAS. Given that agentic workflows can achieve high upper bounds through automated structural design, the paper fails to demonstrate breakthrough on apparent bottlenecks (e.g., AFlow's restricted XML generation format causing significant performance degradation on certain tasks, and the inability to reuse good designs from non-selected nodes—the simplest example being that poor workflow structures may contain good prompts). Moreover, I noticed in the appendix that the evolved MermaidFlow ensemble uses 5-sample voting while AFlow uses 3-sample voting, which could be a major contributing factor to the performance improvement rather than the Mermaid representation itself.\n- **Unfair cost comparison**: The reported token consumption of MermaidFlow does not clearly specify whether it includes the Mermaid→Python translation step. According to the experimental setup, each generated Mermaid workflow requires translation via gpt-4o-mini. This cost should be explicitly stated regarding inclusion, along with the specific cost breakdown for this component alone.\n\n\n**3. Technical Detail Issues**\n\n- **Unclear role of the Checker**: If Mermaid provides strong type constraints, the checker should rarely trigger, as LLMs have very low probability of generating syntactically invalid Python code. However, the checker is actually a critical component in MermaidFlow (Section A.2), suggesting that the constraints are not strong enough. A >90% success rate still means 10% of generations require retry, which remains quite high.\n- **Incomplete operator definitions**: The paper claims \"safety-constrained\", but specific safety properties (such as deadlock-freedom, termination) are not defined. There is also a lack of case study analysis on how directly generating Python code workflows violates these safety properties.\n\n\n**4. Inappropriate Baseline Categorization (Minor)**\n\n- **Conceptually flawed taxonomy**: The authors categorize CoT, ComplexCoT, and Self-Consistency as \"Single-agent execution methods\" (Section 5.1, Table 1). This classification is deeply problematic. The term \"agentic workflow\" was introduced precisely to distinguish workflows with fixed resource scheduling from agents capable of autonomously allocating computational resources—with LLM calls being the most representative computational resource. CoT is a classic prompting strategy, while Self-Consistency is a classic (non-agentic) workflow. Neither involves autonomous agents that can dynamically decide when and how to invoke LLMs based on intermediate states or task requirements. By labeling these non-agent baselines as \"single-agent methods,\" the authors conflate fundamentally different paradigms and misrepresent the conceptual boundaries of their contribution.", "questions": "1. Could you provide more concrete case studies, including but not limited to:\n   - **Mermaid vs. Python comparison**: Show side-by-side examples where (a) Python-based workflow generation fails but Mermaid succeeds, and (b) demonstrate what specific safety properties are violated in the Python case (e.g., deadlock, non-termination, type mismatch).\n   - **Safety constraint effectiveness**: Provide concrete examples showing how safety constraints improve efficiency or effectiveness during evolution. What specific invalid mutations are prevented? How much retry overhead is avoided?\n   - **Checker analysis**: Given that the checker has >90% success rate (implying ~10% retry), provide examples of the 10% failed cases. What constraint violations occur? Why doesn't Mermaid's type system prevent these?\n   - **Component reuse**: Demonstrate how MermaidFlow enables reusing good components (e.g., effective prompts) from suboptimal workflows, addressing the AFlow limitation you mentioned. Show concrete examples from your evolution process.\n\n2. Could you provide a more detailed cost analysis?\n   - Explicitly state whether the reported token consumption includes Mermaid→Python translation costs\n   - Break down token costs by component: (a) workflow generation, (b) translation, (c) execution, (d) evaluation\n   - Compare the per-iteration cost breakdown between MermaidFlow and AFlow\n\n3. I would greatly appreciate if you could discuss the concerns I raised in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761824710350}, {"id": "ArOJhz5FLn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24482/Reviewer_ZQZJ"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces MermaidFlow, a framework for generating agentic workflows using Mermaid as an intermediate representation. The core contribution is representing workflows as declarative, statically verifiable graphs instead of imperative code, combined with a safety-constrained evolutionary programming (EP) approach for workflow optimization. Also, MermaidFlow proposes a set of mutation operators that preserve senmantic correctness during search. Experiments on GSM8K, MATH, HumanEval, and MBPP show improvements over baselines including AFlow and ADAS, with the framework achieving higher success rates and better learning and token efficiency.", "review_text": "This paper introduces MermaidFlow, a framework for generating agentic workflows using Mermaid as an intermediate representation. The core contribution is representing workflows as declarative, statically verifiable graphs instead of imperative code, combined with a safety-constrained evolutionary programming (EP) approach for workflow optimization. Also, MermaidFlow proposes a set of mutation operators that preserve senmantic correctness during search. Experiments on GSM8K, MATH, HumanEval, and MBPP show improvements over baselines including AFlow and ADAS, with the framework achieving higher success rates and better learning and token efficiency.", "strengths": "- **Novel Representation**: The use of Mermaid as a declarative intermediate representation is innovative and well-motivated. It cleanly separates planning from execution, addressing a key weakness in prior code-based methods.\n- **Comprehensive System Design**: The paper thoroughly describes the type system, operators, validation mechanisms (soft and hard checks), and the complete pipeline from Mermaid to executable code.\n- **Consistent Empirical Improvements**: The method shows improvements across all four benchmarks (average 80.75% vs. 79.35% for the best baseline).", "weaknesses": "- **Mermaid DSL Limitations**: Mermaid DSL is static by design and may be hard to express loops, conditionals or some runtime operations. The paper does not show how these limitations. Extending the DSL or documenting its expressive limits would strengthen the contribution.\n- **Presentation Issues**: Figures 1 and 2 appear as raster images rather than vector graphics (like pdf or svg) and lose clarity when zoomed.\n- **Execution Model Ablations**: The paper uses only gpt-4o-mini as the execution LLM. Performance with other models (e.g., Claude, Llama) is only explored for a brief ablation on optimization LLM.", "questions": "- **Mutation Operator Effectiveness**: Which evolutionary operators contribute most to performance improvements? The paper mentions crossover occurs with only 10% probability but provides no analysis of operator frequency, success rates, or relative contributions.\n- **Iteration Scaling and Convergence**: Is 20 iterations sufficient for convergence, or would extended search yield further gains? The paper doesn't justify this choice or show whether performance plateaus, and provides no analysis of optimal stopping points across different tasks. Additionally, if there are more iterations, can different workflows be observed, and will their complexity increase with the iteration?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces MermaidFlow, a framework for generating agentic workflows using Mermaid as an intermediate representation. The core contribution is representing workflows as declarative, statically verifiable graphs instead of imperative code, combined with a safety-constrained evolutionary programming (EP) approach for workflow optimization. Also, MermaidFlow proposes a set of mutation operators that preserve senmantic correctness during search. Experiments on GSM8K, MATH, HumanEval, and MBPP show improvements over baselines including AFlow and ADAS, with the framework achieving higher success rates and better learning and token efficiency.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- **Novel Representation**: The use of Mermaid as a declarative intermediate representation is innovative and well-motivated. It cleanly separates planning from execution, addressing a key weakness in prior code-based methods.\n- **Comprehensive System Design**: The paper thoroughly describes the type system, operators, validation mechanisms (soft and hard checks), and the complete pipeline from Mermaid to executable code.\n- **Consistent Empirical Improvements**: The method shows improvements across all four benchmarks (average 80.75% vs. 79.35% for the best baseline).", "weaknesses": "- **Mermaid DSL Limitations**: Mermaid DSL is static by design and may be hard to express loops, conditionals or some runtime operations. The paper does not show how these limitations. Extending the DSL or documenting its expressive limits would strengthen the contribution.\n- **Presentation Issues**: Figures 1 and 2 appear as raster images rather than vector graphics (like pdf or svg) and lose clarity when zoomed.\n- **Execution Model Ablations**: The paper uses only gpt-4o-mini as the execution LLM. Performance with other models (e.g., Claude, Llama) is only explored for a brief ablation on optimization LLM.", "questions": "- **Mutation Operator Effectiveness**: Which evolutionary operators contribute most to performance improvements? The paper mentions crossover occurs with only 10% probability but provides no analysis of operator frequency, success rates, or relative contributions.\n- **Iteration Scaling and Convergence**: Is 20 iterations sufficient for convergence, or would extended search yield further gains? The paper doesn't justify this choice or show whether performance plateaus, and provides no analysis of optimal stopping points across different tasks. Additionally, if there are more iterations, can different workflows be observed, and will their complexity increase with the iteration?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761817686457}, {"id": "6ct3GXKsR4", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24482/Reviewer_cd1T"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "This paper introduces a unified framework combining multi-agent learning and evolutionary optimization. The proposed system models workflow automation as a constrained single-objective optimization with an evolutionary loop governed by variation, selection, and reflection. A meta-controller dynamically updates strategies and rules, ensuring adaptive improvement. Theoretical results guarantee convergence and monotonic rule-quality growth, while experiments demonstrate strong cross-domain generalization.", "review_text": "This paper introduces a unified framework combining multi-agent learning and evolutionary optimization. The proposed system models workflow automation as a constrained single-objective optimization with an evolutionary loop governed by variation, selection, and reflection. A meta-controller dynamically updates strategies and rules, ensuring adaptive improvement. Theoretical results guarantee convergence and monotonic rule-quality growth, while experiments demonstrate strong cross-domain generalization.", "strengths": "* Solid theoretical formalization.\n* Innovative integration of evolutionary search and multi-agent optimization.\n* Well-structured proofs and explicit assumptions.\n* Broad empirical evaluation demonstrating clear advantages.\n* Clear, consistent, and professional presentation.", "weaknesses": "* **Incomplete experimental reporting**: Missing per-benchmark hyperparameters and statistical variance.\n  *Suggestion:* Add full configuration tables and multi-run results.\n* **Unclear curriculum mechanism**: Difficulty-level staging lacks quantitative definitions.\n  *Suggestion:* Provide formal thresholds and curriculum ablations.\n* **Assumptions not empirically tested**: Theoretical premises like positive information gain remain unchecked.\n  *Suggestion:* Visualize empirical distributions of related quantities.\n* **Limited strong baselines**: Comparison with advanced workflow retrieval or graph-based frameworks is limited.\n  *Suggestion:* Extend experiments to include such baselines.", "questions": "* How does the strategy distribution evolve across curriculum phases?\n* How sensitive is convergence to the meta-controller update interval?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a unified framework combining multi-agent learning and evolutionary optimization. The proposed system models workflow automation as a constrained single-objective optimization with an evolutionary loop governed by variation, selection, and reflection. A meta-controller dynamically updates strategies and rules, ensuring adaptive improvement. Theoretical results guarantee convergence and monotonic rule-quality growth, while experiments demonstrate strong cross-domain generalization.", "soundness": 2, "presentation": 3, "contribution": 4, "strengths": "* Solid theoretical formalization.\n* Innovative integration of evolutionary search and multi-agent optimization.\n* Well-structured proofs and explicit assumptions.\n* Broad empirical evaluation demonstrating clear advantages.\n* Clear, consistent, and professional presentation.", "weaknesses": "* **Incomplete experimental reporting**: Missing per-benchmark hyperparameters and statistical variance.\n  *Suggestion:* Add full configuration tables and multi-run results.\n* **Unclear curriculum mechanism**: Difficulty-level staging lacks quantitative definitions.\n  *Suggestion:* Provide formal thresholds and curriculum ablations.\n* **Assumptions not empirically tested**: Theoretical premises like positive information gain remain unchecked.\n  *Suggestion:* Visualize empirical distributions of related quantities.\n* **Limited strong baselines**: Comparison with advanced workflow retrieval or graph-based frameworks is limited.\n  *Suggestion:* Extend experiments to include such baselines.", "questions": "* How does the strategy distribution evolve across curriculum phases?\n* How sensitive is convergence to the meta-controller update interval?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761780114034}], "openreview_url": "https://openreview.net/forum?id=bhPaXhWVKG", "arxiv_id": "2505.22967", "paper_pdf": "papers/bhPaXhWVKG.pdf", "paper_pdf_sha256": "4bca0c0f42e49ba134dad241f8c50fa0fb8fbd1acac577bd3c30ee849421850f", "paper_pdf_bytes": 3199007, "paper_pdf_source": "openreview", "code_url": "https://github.com/chengqiArchy/MermaidFlow", "code_repository": "chengqiArchy/MermaidFlow", "code_commit": "cd2478dff0b0f1e01137be1bef8b04ea2aa94289", "code_archive": "repos/bhPaXhWVKG.zip", "code_archive_sha256": "c7bfca16649a87041552940958efa6759adc6054bb7de236a80a6df186989385", "code_archive_bytes": 4045407, "code_file_count": 60, "code_extensions": {".py": 60}, "github_disk_usage_kb": 3766, "github_languages": {"Python": 298080, "Mermaid": 3280}, "github_archived": false, "github_pushed_at": "2025-12-12T06:43:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mermaidflow-redefining-agentic-workflow"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RLzeoy4FzP", "year": 2025, "status": "rejected", "title": "Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback", "authors": ["Lester James Validad Miranda", "Yizhong Wang", "Yanai Elazar", "Sachin Kumar", "Valentina Pyatkin", "Faeze Brahman", "Noah A. Smith", "Hannaneh Hajishirzi", "Pradeep Dasigi"], "authorids": ["~Lester_James_Validad_Miranda1", "~Yizhong_Wang2", "~Yanai_Elazar1", "~Sachin_Kumar1", "~Valentina_Pyatkin1", "~Faeze_Brahman1", "~Noah_A._Smith2", "~Hannaneh_Hajishirzi1", "~Pradeep_Dasigi1"], "authors_source": "OpenReview API", "abstract": "Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. However, directly collecting human preferences can be expensive, time-consuming, and can have high variance. An appealing alternative is to distill preferences from LMs as a source of synthetic annotations as they are more consistent, cheaper, and scale better than human annotation; however, they are also prone to biases and errors. In this work, we introduce a routing framework that combines inputs from humans and LMs to achieve better annotation quality, while reducing the total cost of human annotation. The crux of our approach is to identify preference instances that will benefit from human annotations. We formulate this as an optimization problem: given a preference dataset and an evaluation metric, we train a performance prediction model to predict a reward model's performance on an arbitrary combination of human and LM annotations and employ a routing strategy that selects a combination that maximizes predicted performance. We train the performance prediction model on MultiPref, a new preference dataset with 10K instances paired with human and LM labels. We show that the selected hybrid mixture of LM and direct human preferences using our routing framework achieves better reward model performance compared to using either one exclusively. We simulate selective human preference collection on three other datasets and show that our method generalizes well to all three. We analyze features from the routing model to identify characteristics of instances that can benefit from human feedback, e.g., prompts with a moderate safety concern or moderate intent complexity. We release the dataset, annotation platform, and source code used in this study to foster more efficient and accurate preference collection in the future.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "gQk8jvTx25", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8938/Reviewer_u41y"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes a routing framework for preference data mixing and selection. The framework decides whether to use human annotations or language model annotations by determining if an instance benefits from human annotations. Specifically, it consists of a Performance Prediction Model (PPM) and a routing strategy. The PPM learns the statistical features of several mixing datasets that are routed to human annotations and their performance after training, enabling the PPM to predict the performance of any preference dataset. Using the routing strategy, multiple candidate mixing datasets are generated, and the PPM predicts and selects the best-performing dataset for training. Besides, a preference dataset, MULTIPREF, is constructed to implement this framework. The effectiveness of the data mixing method using this framework is demonstrated across multiple benchmarks, showing improved performance compared to baselines. Furthermore, the characteristics of data that benefit from human annotations are analyzed.", "review_text": "This paper proposes a routing framework for preference data mixing and selection. The framework decides whether to use human annotations or language model annotations by determining if an instance benefits from human annotations. Specifically, it consists of a Performance Prediction Model (PPM) and a routing strategy. The PPM learns the statistical features of several mixing datasets that are routed to human annotations and their performance after training, enabling the PPM to predict the performance of any preference dataset. Using the routing strategy, multiple candidate mixing datasets are generated, and the PPM predicts and selects the best-performing dataset for training. Besides, a preference dataset, MULTIPREF, is constructed to implement this framework. The effectiveness of the data mixing method using this framework is demonstrated across multiple benchmarks, showing improved performance compared to baselines. Furthermore, the characteristics of data that benefit from human annotations are analyzed.", "strengths": "1. This paper introduces an innovative routing framework to optimize the preference label space, automating the selection of appropriate annotation sources, thereby reducing human annotation costs while achieving better model performance.\n2. It presents a Performance Prediction Model (PPM) that predicts the performance metrics of a trained model on a benchmark using the given dataset directly, rather than relying on model outputs, making it more convenient to predict dataset performance.\n3. The claims regarding the performance of the framework are well-supported by experiments, with the designed experiments effectively demonstrating the advantages and good generalization performance of the routing framework across various benchmarks.", "weaknesses": "1. The experiments lack tests on other models, as all experiments are based on Tülu 2 13B. It is unclear if the framework would provide the same performance improvements when replacing the Reward Model or Policy Model. Training the PPM requires conducting hundreds of RLHF runs to explore the relationship between data features and outcomes, which is computationally expensive. If the PPM has low generalizability across different models, the usage cost would be high.\n2. There is a lack of reasoning and explanation regarding the routing strategy algorithm, making it unclear why this particular routing strategy is adopted and the characteristics of the mixing data it generates.\n3. The accuracy of the PPM is not convincing. The paper provides limited ablation analysis of the PPM and lacks information on the accuracy and robustness of the PPM's predictions on non-training datasets.", "questions": "1. Does the PPM have transferability to models other than Tülu 2 13B, and what are the computational costs?\n2. Could you provide more information about the routing strategy? How does the routing strategy differ from completely random selection in terms of PPM training and the PPM’s evaluation of the generated mixing datasets?\n3. Figure 3 shows that the PPM's predictions are very accurate on the actual RewardBench. How does it perform on the other evaluation tasks you used?\n4. The paper mentions that the PPM was trained on 200 candidates datasets generated by routing strategies but only selects one predicted optimal candidate datasets from 500 candidates generated by routing strategies for evaluation. Why is this the case? According to the paper, the PPM's prediction time should be much less than conducting actual RLHF and evaluations on candidate datasets, so it would be feasible to select the optimal candidate datasets predicted by the PPM from thousands of candidates.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a routing framework for preference data mixing and selection. The framework decides whether to use human annotations or language model annotations by determining if an instance benefits from human annotations. Specifically, it consists of a Performance Prediction Model (PPM) and a routing strategy. The PPM learns the statistical features of several mixing datasets that are routed to human annotations and their performance after training, enabling the PPM to predict the performance of any preference dataset. Using the routing strategy, multiple candidate mixing datasets are generated, and the PPM predicts and selects the best-performing dataset for training. Besides, a preference dataset, MULTIPREF, is constructed to implement this framework. The effectiveness of the data mixing method using this framework is demonstrated across multiple benchmarks, showing improved performance compared to baselines. Furthermore, the characteristics of data that benefit from human annotations are analyzed.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper introduces an innovative routing framework to optimize the preference label space, automating the selection of appropriate annotation sources, thereby reducing human annotation costs while achieving better model performance.\n2. It presents a Performance Prediction Model (PPM) that predicts the performance metrics of a trained model on a benchmark using the given dataset directly, rather than relying on model outputs, making it more convenient to predict dataset performance.\n3. The claims regarding the performance of the framework are well-supported by experiments, with the designed experiments effectively demonstrating the advantages and good generalization performance of the routing framework across various benchmarks.", "weaknesses": "1. The experiments lack tests on other models, as all experiments are based on Tülu 2 13B. It is unclear if the framework would provide the same performance improvements when replacing the Reward Model or Policy Model. Training the PPM requires conducting hundreds of RLHF runs to explore the relationship between data features and outcomes, which is computationally expensive. If the PPM has low generalizability across different models, the usage cost would be high.\n2. There is a lack of reasoning and explanation regarding the routing strategy algorithm, making it unclear why this particular routing strategy is adopted and the characteristics of the mixing data it generates.\n3. The accuracy of the PPM is not convincing. The paper provides limited ablation analysis of the PPM and lacks information on the accuracy and robustness of the PPM's predictions on non-training datasets.", "questions": "1. Does the PPM have transferability to models other than Tülu 2 13B, and what are the computational costs?\n2. Could you provide more information about the routing strategy? How does the routing strategy differ from completely random selection in terms of PPM training and the PPM’s evaluation of the generated mixing datasets?\n3. Figure 3 shows that the PPM's predictions are very accurate on the actual RewardBench. How does it perform on the other evaluation tasks you used?\n4. The paper mentions that the PPM was trained on 200 candidates datasets generated by routing strategies but only selects one predicted optimal candidate datasets from 500 candidates generated by routing strategies for evaluation. Why is this the case? According to the paper, the PPM's prediction time should be much less than conducting actual RLHF and evaluations on candidate datasets, so it would be feasible to select the optimal candidate datasets predicted by the PPM from thousands of candidates.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730720838969}, {"id": "Ws6JyR7qKp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8938/Reviewer_PxBW"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper aims to reduce the cost of preference tuning by training a model to route which sample should be annotated by humans and which can be annotated by LLM.", "review_text": "This paper aims to reduce the cost of preference tuning by training a model to route which sample should be annotated by humans and which can be annotated by LLM.", "strengths": "- The problem of this paper is important: obtaining a reward that best aligns with humans at the lowest cost. The proposed method can somehow reasonably reduce the cost.\n- If the motivation is to reduce human annotation costs for policy tuning, the author's experiment results go beyond this. They found that human annotations are not necessary for all data, and the model can perform even better with fewer human annotations.\n- The authors perform experiments on various datasets, each of which reflects an important ability of LLM, and their method achieves good results on those benchmark datasets.", "weaknesses": "- The conclusion of this paper is intriguing but without in-depth discussion. If 100% human performance leads to worse (even the worst) performance, how can the author explain the policy model performance gap between training on synthetic and human annotations? Is our goal still \"alignment\"?\n- Algorithm 1 looks too intuitive. The authors fill the human annotation set with arbitrary examples until it reaches budget b. It is unclear how this greedy process helps maximize the objective (Eq.(1)). More specifically, in the worst case, the algorithm will select the worst b samples, i.e., human annotations gain the least for the model alignment. This will further impact the PPM and trigger error propagation. \n- In the experiment, the authors only show the comparison between 100% human annotations,100% synthetic annotations in Table 3 and Table 4, and the proposed best hybrid annotations. They should include at least one randomly selected hybrid combination, e.g., 50% human annotations with randomly selected data + 50% synthetic annotations.", "questions": "- Regarding L188-L189, does the |S_human| = |D| and |S_human| = 0 equals to b=|D| and b=0?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to reduce the cost of preference tuning by training a model to route which sample should be annotated by humans and which can be annotated by LLM.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The problem of this paper is important: obtaining a reward that best aligns with humans at the lowest cost. The proposed method can somehow reasonably reduce the cost.\n- If the motivation is to reduce human annotation costs for policy tuning, the author's experiment results go beyond this. They found that human annotations are not necessary for all data, and the model can perform even better with fewer human annotations.\n- The authors perform experiments on various datasets, each of which reflects an important ability of LLM, and their method achieves good results on those benchmark datasets.", "weaknesses": "- The conclusion of this paper is intriguing but without in-depth discussion. If 100% human performance leads to worse (even the worst) performance, how can the author explain the policy model performance gap between training on synthetic and human annotations? Is our goal still \"alignment\"?\n- Algorithm 1 looks too intuitive. The authors fill the human annotation set with arbitrary examples until it reaches budget b. It is unclear how this greedy process helps maximize the objective (Eq.(1)). More specifically, in the worst case, the algorithm will select the worst b samples, i.e., human annotations gain the least for the model alignment. This will further impact the PPM and trigger error propagation. \n- In the experiment, the authors only show the comparison between 100% human annotations,100% synthetic annotations in Table 3 and Table 4, and the proposed best hybrid annotations. They should include at least one randomly selected hybrid combination, e.g., 50% human annotations with randomly selected data + 50% synthetic annotations.", "questions": "- Regarding L188-L189, does the |S_human| = |D| and |S_human| = 0 equals to b=|D| and b=0?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730637333491}, {"id": "AAlYrmNZBp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8938/Reviewer_Xt8Y"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper explores the integration of human and AI feedback to enhance annotation quality while reducing the overall cost of human annotation. The authors introduce a performance prediction model (PPM) designed to forecast the effectiveness of reward models based on various feedback combinations. To train the PPM, they create a novel preference dataset, MULTIPREF, which includes both human and AI annotations. Empirical results indicate that this approach yields better-performing reward models compared to using human or AI annotations alone.", "review_text": "This paper explores the integration of human and AI feedback to enhance annotation quality while reducing the overall cost of human annotation. The authors introduce a performance prediction model (PPM) designed to forecast the effectiveness of reward models based on various feedback combinations. To train the PPM, they create a novel preference dataset, MULTIPREF, which includes both human and AI annotations. Empirical results indicate that this approach yields better-performing reward models compared to using human or AI annotations alone.", "strengths": "The method is well-writen and well-motivated, addressing the high costs and time constraints associated with human annotation.\n\nThe construction of a new preference dataset combining human and AI annotations adds significant value for further analysis in the community.", "weaknesses": "Evaluating the quality of the two feedback types solely based on reward model performance is indirect and highly contingent on the model's training configuration, which can be influenced by numerous accidental factors. The authors should consider incorporating more direct metrics for evaluating the feedback types, along with corresponding case studies or additional downstream tasks, such as aligning LLMs through direct alignment algorithms.\n\nThe versatility of the feature representation is not sufficiently validated. If the feature representation lacks versatility, it will require manual design for each new task, complicating the process even more than human annotations. Furthermore, validating the effectiveness of feature representation necessitates preference data comprising both human and AI annotations, which could limit the practical applicability of the method.\n\n\nMore detailed training configurations and evaluation results of PPM is needed. For example, the correlation between the predicted and actual values ​​of the subdivision direction on the reward bench.", "questions": "While human feedback is generally viewed as high quality, the results presented suggest that integrating AI feedback is critical for performance improvement. The analysis in Section 5 indicates that human annotations perform better on moderate preference datasets. However, the paper does not address in which specific preference instances AI feedback provides a performance advantage. Is it on pairs with greater preference differentiation? If so, why?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the integration of human and AI feedback to enhance annotation quality while reducing the overall cost of human annotation. The authors introduce a performance prediction model (PPM) designed to forecast the effectiveness of reward models based on various feedback combinations. To train the PPM, they create a novel preference dataset, MULTIPREF, which includes both human and AI annotations. Empirical results indicate that this approach yields better-performing reward models compared to using human or AI annotations alone.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "The method is well-writen and well-motivated, addressing the high costs and time constraints associated with human annotation.\n\nThe construction of a new preference dataset combining human and AI annotations adds significant value for further analysis in the community.", "weaknesses": "Evaluating the quality of the two feedback types solely based on reward model performance is indirect and highly contingent on the model's training configuration, which can be influenced by numerous accidental factors. The authors should consider incorporating more direct metrics for evaluating the feedback types, along with corresponding case studies or additional downstream tasks, such as aligning LLMs through direct alignment algorithms.\n\nThe versatility of the feature representation is not sufficiently validated. If the feature representation lacks versatility, it will require manual design for each new task, complicating the process even more than human annotations. Furthermore, validating the effectiveness of feature representation necessitates preference data comprising both human and AI annotations, which could limit the practical applicability of the method.\n\n\nMore detailed training configurations and evaluation results of PPM is needed. For example, the correlation between the predicted and actual values ​​of the subdivision direction on the reward bench.", "questions": "While human feedback is generally viewed as high quality, the results presented suggest that integrating AI feedback is critical for performance improvement. The analysis in Section 5 indicates that human annotations perform better on moderate preference datasets. However, the paper does not address in which specific preference instances AI feedback provides a performance advantage. Is it on pairs with greater preference differentiation? If so, why?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730514641142}, {"id": "YHkf9ccRBf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8938/Reviewer_qKe8"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "This paper proposes a framework that leverages a Performance Prediction Model (PPM) to route pairs of preference data to either human annotators or LLM annotators. This data selection strategy improves the performance on RewardBench and downstream tasks.", "review_text": "This paper proposes a framework that leverages a Performance Prediction Model (PPM) to route pairs of preference data to either human annotators or LLM annotators. This data selection strategy improves the performance on RewardBench and downstream tasks.", "strengths": "N/A", "weaknesses": "1. This paper adopts an intuitive approach to data selection by training a classifier to choose the optimal preference data source between human or large language models (LLMs). Although there is no identical existing work, such improvements seem minor. Training a classifier appears to be merely an empirical engineering optimization, lacking elegance. Furthermore, similar efforts have employed uncertainty to judge the choice of sources, as discussed in [a].\n2. The contribution of preparing data for a minor enhancement, such as training a classifier, is also minimal. Moreover, the dataset is quite simplistic: The feature space is basic, and the coverage and granularity of the Descriptive Subject are limited, as shown in Table 9. This simplicity suggests that the work is insufficient to serve as a foundation for substantial future advancements.\n3. Using PPM to fit evaluation results on RewardBench is questionable. PPM is intended to assess the quality of a preference data source, thus the reliability of preference data quality on RewardBench is dubious. It is difficult to ascertain whether the PPM results are solid.\n4. Section 2.3 on the ROUTING STRATEGY is also problematic. With a training set of 7K samples, each having a 0/1 binary choice, there are a total of 2^7K possible candidates. Yet, only 200 or 500 samples are selected, as claimed in line 219, which is minuscule compared to the total number of candidates, hardly ensuring adequate coverage.\n5. Why do the authors select Tulu2 as the base rather than more popular models like LLaMA3? It is crucial to determine whether the PPM can improve upon more advanced models like LLaMA3, as shown in Tables 3 and 4.\n6. The generalizability and scalability of the PPM are uncertain since it is trained using a limited RewardBench framework. Although Section 4.3 tested additional datasets, some of these, such as BBH and AlpacaEval, were part of RewardBench, which weakens the proof of effectiveness in Section 4.3.\n7. Expanding the PPM to train on a larger dataset beyond RewardBench would introduce additional complexity to the RM-PPO pipeline, further increasing the complexity of RLHF. Such an investment may be neither elegant nor cost-effective.\n\n[a]  Huang H, Qu Y, Liu J, et al. On the Limitations of Fine-tuned Judge Models for LLM Evaluation.[J]. arXiv preprint arXiv:2403.02839, 2024.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a framework that leverages a Performance Prediction Model (PPM) to route pairs of preference data to either human annotators or LLM annotators. This data selection strategy improves the performance on RewardBench and downstream tasks.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "N/A", "weaknesses": "1. This paper adopts an intuitive approach to data selection by training a classifier to choose the optimal preference data source between human or large language models (LLMs). Although there is no identical existing work, such improvements seem minor. Training a classifier appears to be merely an empirical engineering optimization, lacking elegance. Furthermore, similar efforts have employed uncertainty to judge the choice of sources, as discussed in [a].\n2. The contribution of preparing data for a minor enhancement, such as training a classifier, is also minimal. Moreover, the dataset is quite simplistic: The feature space is basic, and the coverage and granularity of the Descriptive Subject are limited, as shown in Table 9. This simplicity suggests that the work is insufficient to serve as a foundation for substantial future advancements.\n3. Using PPM to fit evaluation results on RewardBench is questionable. PPM is intended to assess the quality of a preference data source, thus the reliability of preference data quality on RewardBench is dubious. It is difficult to ascertain whether the PPM results are solid.\n4. Section 2.3 on the ROUTING STRATEGY is also problematic. With a training set of 7K samples, each having a 0/1 binary choice, there are a total of 2^7K possible candidates. Yet, only 200 or 500 samples are selected, as claimed in line 219, which is minuscule compared to the total number of candidates, hardly ensuring adequate coverage.\n5. Why do the authors select Tulu2 as the base rather than more popular models like LLaMA3? It is crucial to determine whether the PPM can improve upon more advanced models like LLaMA3, as shown in Tables 3 and 4.\n6. The generalizability and scalability of the PPM are uncertain since it is trained using a limited RewardBench framework. Although Section 4.3 tested additional datasets, some of these, such as BBH and AlpacaEval, were part of RewardBench, which weakens the proof of effectiveness in Section 4.3.\n7. Expanding the PPM to train on a larger dataset beyond RewardBench would introduce additional complexity to the RM-PPO pipeline, further increasing the complexity of RLHF. Such an investment may be neither elegant nor cost-effective.\n\n[a]  Huang H, Qu Y, Liu J, et al. On the Limitations of Fine-tuned Judge Models for LLM Evaluation.[J]. arXiv preprint arXiv:2403.02839, 2024.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730222828741}], "openreview_url": "https://openreview.net/forum?id=RLzeoy4FzP", "arxiv_id": "2410.19133", "paper_pdf": "papers/RLzeoy4FzP.pdf", "paper_pdf_sha256": "8bcb1b99f131b0b493983f8831f6e30ab8c5ea39a04012459adb87e13ed5b1b3", "paper_pdf_bytes": 2960681, "paper_pdf_source": "openreview", "code_url": "https://github.com/allenai/hybrid-preferences", "code_repository": "allenai/hybrid-preferences", "code_commit": "4d96b27b989828b0b69a430f68d25b6b36e9015a", "code_archive": "repos/RLzeoy4FzP.zip", "code_archive_sha256": "fb6a53ea5b4b56f2714b96d0c610ebaa56144131639f30ea3f25f6113eb9b60e", "code_archive_bytes": 88133, "code_file_count": 26, "code_extensions": {".py": 15, ".sh": 11}, "github_disk_usage_kb": 271, "github_languages": {"Python": 171361, "Shell": 13112, "Dockerfile": 2390, "Makefile": 525}, "github_archived": false, "github_pushed_at": "2025-07-23T03:36:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hybrid-preferences-learning-to-route"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JAKcnjzQI3", "year": 2024, "status": "rejected", "title": "MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective", "authors": ["Yizhuo Chen", "Chun-Fu Chen", "Hsiang Hsu", "Shaohan Hu", "Marco Pistoia", "Tarek F. Abdelzaher"], "authorids": ["~Yizhuo_Chen2", "~Chun-Fu_Chen1", "~Hsiang_Hsu1", "~Shaohan_Hu2", "~Marco_Pistoia2", "~Tarek_F._Abdelzaher1"], "authors_source": "OpenReview API", "abstract": "The growing richness of large-scale datasets has been a crucial driving force behind the rapid advancement and wide adoption of machine learning technologies. The massive collection and usage of data, however, pose an increasing risk for people’s private and sensitive information due to either inadvertent mishandling or malicious exploitation. Besides legislative solutions, many technical approaches have been proposed towards data privacy protection. However, they bear various limitations such as leading to degraded data availability and utility, or relying on heuristics and lacking solid theoretical bases. To overcome these limitations, we propose a formal information-theoretic definition for this utility-preserving privacy protection problem, and design a data-driven learnable data transformation framework that is capable of selectively suppressing sensitive attributes from target datasets while preserving the other useful attributes, regardless of whether or not they are known in advance or explicitly annotated for preservation. We provide rigorous theoretical analyses on the operational bounds for our framework, and carry out comprehensive experimental evaluations using datasets of a variety of modalities, including facial images, voice audio clips, and human activity motion sensor signals. Results demonstrate the effectiveness and generalizability of our method on different tasks and configurations.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Cf3ekbRgVL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6515/Reviewer_wuPW"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposes a method for learning censored data transformations providing guarantees on the quantity of information about both annotated and unannotated features preserved by said transformations. The authors motivate the method using an information-theoretic calculus and establish operational bounds on the entailed objective. Practically, censoring of the designated sensitive attributes is achieved through a standard (margin-based) adversarial information-minimisation (infomin) procedure; useful annotated and unannotated information is preserved through the use of supervised and contrastive learning, respectively. The authors conduct experiments on datasets covering a range of modalities in AudioMNIST, Motion Sense, and Adience and demonstrate favourable performance of their method relative to the baseline suite.", "review_text": "The paper proposes a method for learning censored data transformations providing guarantees on the quantity of information about both annotated and unannotated features preserved by said transformations. The authors motivate the method using an information-theoretic calculus and establish operational bounds on the entailed objective. Practically, censoring of the designated sensitive attributes is achieved through a standard (margin-based) adversarial information-minimisation (infomin) procedure; useful annotated and unannotated information is preserved through the use of supervised and contrastive learning, respectively. The authors conduct experiments on datasets covering a range of modalities in AudioMNIST, Motion Sense, and Adience and demonstrate favourable performance of their method relative to the baseline suite.", "strengths": "- Figures and tables are well-put-together; Figures 1 and 2 illustrate the problem setup and methodological pipeline, respectively, in an easily digestible manner -- one can understand the essence of the method based on its illustration alone.\n- Experiments cover a good range of datasets and configurations.\n- Proofs and implications of the consequent theoretical statements are easy to follow.\n- The problem under consideration is well-motivated and clearly formulated.\n- Reasonable assortment of baseline methods and strong empirical performance of the proposed method relative to these. Experimental setups are described with clarity.\n- Good contextualization w.r.t. prior work, with clear delineation of the subtle but differentiating qualities of the current work.", "weaknesses": "- The paper is limited in terms of novelty. The main contribution of the paper seems to be in its proposal to preserve of *unannotated* features using a self-supervised-learning objective yet this idea of maximising $\\mathcal{I}(X; \\tilde{X})$, understanding $\\tilde{X}$ to be some representation generally, is certainly not novel in and of itself (vide Madras et al., 2018 in the context of the adjacent field of fair-representation learning); the method chosen used to accomplish this maximisation seems, to me, largely incidental -- one can simply view the contrastive loss as an alternative reconstruction loss.\n- The paper proposes learning a data transformation instead of a representation but the codomain of the transformation is another seemingly incidental factor given that interpretability does not appear to be a major concern, based on the narrative and analysis; indeed, in order to compute the contrastive learning objective in a space in which distances are meaningful, the transformed and original samples ultimately have to be embedded in such a representation anyway. The learning of a data transformation instead of a low-dimensional representation leads to a method that is more complicated than seemingly need be, and, moreover, is a design choice that comes at a steep computational cost -- computing all losses in representation space would be much more efficient though there may be some sound theoretical barrier to do doing so.\n- The mathematical formalism is confusing and, at times, unrigorous. Random variables and their realisations are seemingly conflated without reference to the abuse: mutual information, $\\mathcal{I}(\\cdot; \\cdot)$ is defined between pairs of random variables, not their empirical counterparts. While there is an argument to be made that such overloading is conventional and remedied by the context, I don't think that the latter is entirely satisifed here, especially with their being no express mention of this overloading being adopted throughout the paper.\n- While their meaning, as analogues of $X_p$ and $X_n$, respectively, can be easily inferred, $F_p$ and $F_n$, appearing in Eq.17, seem to be missing explicit definitions. The explanation given in Sec. 4.4,both textually and notationally, is generally muddled considering that the method amounts to SimCLR with the original and transformed samples acting as anchors and positive pairs.\n- Why compute cross-entropy terms w.r.t. the estimates of $P(U_i|X)$ and $P(S_i|X)$ as opposed to simply using the (degenerate ground-truth distribution (the annotations) used in the fitting of those estimates? There may be good reason for it but there should clear explanation given for why this choice is unprincipled, should that indeed be the case.\n- The quality of the writing, in terms of clarity and structure, could generally do with improvement.\n- Lack of ablation studies, such as those investigating the influence of the loss prefactor.\n- No discussion of the practical challenges entailed by adversarial infomin (vide Song and Shmatikov, 2021, for instance).\n\n\n### References\nMadras D, Creager E, Pitassi T, Zemel R. Learning adversarially fair and transferable representations. In International Conference on Machine Learning 2018 Jul 3 (pp. 3384-3393). PMLR.\n\nSong C, Shmatikov V. Overlearning Reveals Sensitive Attributes. In8th International Conference on Learning Representations, ICLR 2020 2020 Jan.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method for learning censored data transformations providing guarantees on the quantity of information about both annotated and unannotated features preserved by said transformations. The authors motivate the method using an information-theoretic calculus and establish operational bounds on the entailed objective. Practically, censoring of the designated sensitive attributes is achieved through a standard (margin-based) adversarial information-minimisation (infomin) procedure; useful annotated and unannotated information is preserved through the use of supervised and contrastive learning, respectively. The authors conduct experiments on datasets covering a range of modalities in AudioMNIST, Motion Sense, and Adience and demonstrate favourable performance of their method relative to the baseline suite.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- Figures and tables are well-put-together; Figures 1 and 2 illustrate the problem setup and methodological pipeline, respectively, in an easily digestible manner -- one can understand the essence of the method based on its illustration alone.\n- Experiments cover a good range of datasets and configurations.\n- Proofs and implications of the consequent theoretical statements are easy to follow.\n- The problem under consideration is well-motivated and clearly formulated.\n- Reasonable assortment of baseline methods and strong empirical performance of the proposed method relative to these. Experimental setups are described with clarity.\n- Good contextualization w.r.t. prior work, with clear delineation of the subtle but differentiating qualities of the current work.", "weaknesses": "- The paper is limited in terms of novelty. The main contribution of the paper seems to be in its proposal to preserve of *unannotated* features using a self-supervised-learning objective yet this idea of maximising $\\mathcal{I}(X; \\tilde{X})$, understanding $\\tilde{X}$ to be some representation generally, is certainly not novel in and of itself (vide Madras et al., 2018 in the context of the adjacent field of fair-representation learning); the method chosen used to accomplish this maximisation seems, to me, largely incidental -- one can simply view the contrastive loss as an alternative reconstruction loss.\n- The paper proposes learning a data transformation instead of a representation but the codomain of the transformation is another seemingly incidental factor given that interpretability does not appear to be a major concern, based on the narrative and analysis; indeed, in order to compute the contrastive learning objective in a space in which distances are meaningful, the transformed and original samples ultimately have to be embedded in such a representation anyway. The learning of a data transformation instead of a low-dimensional representation leads to a method that is more complicated than seemingly need be, and, moreover, is a design choice that comes at a steep computational cost -- computing all losses in representation space would be much more efficient though there may be some sound theoretical barrier to do doing so.\n- The mathematical formalism is confusing and, at times, unrigorous. Random variables and their realisations are seemingly conflated without reference to the abuse: mutual information, $\\mathcal{I}(\\cdot; \\cdot)$ is defined between pairs of random variables, not their empirical counterparts. While there is an argument to be made that such overloading is conventional and remedied by the context, I don't think that the latter is entirely satisifed here, especially with their being no express mention of this overloading being adopted throughout the paper.\n- While their meaning, as analogues of $X_p$ and $X_n$, respectively, can be easily inferred, $F_p$ and $F_n$, appearing in Eq.17, seem to be missing explicit definitions. The explanation given in Sec. 4.4,both textually and notationally, is generally muddled considering that the method amounts to SimCLR with the original and transformed samples acting as anchors and positive pairs.\n- Why compute cross-entropy terms w.r.t. the estimates of $P(U_i|X)$ and $P(S_i|X)$ as opposed to simply using the (degenerate ground-truth distribution (the annotations) used in the fitting of those estimates? There may be good reason for it but there should clear explanation given for why this choice is unprincipled, should that indeed be the case.\n- The quality of the writing, in terms of clarity and structure, could generally do with improvement.\n- Lack of ablation studies, such as those investigating the influence of the loss prefactor.\n- No discussion of the practical challenges entailed by adversarial infomin (vide Song and Shmatikov, 2021, for instance).\n\n\n### References\nMadras D, Creager E, Pitassi T, Zemel R. Learning adversarially fair and transferable representations. In International Conference on Machine Learning 2018 Jul 3 (pp. 3384-3393). PMLR.\n\nSong C, Shmatikov V. Overlearning Reveals Sensitive Attributes. In8th International Conference on Learning Representations, ICLR 2020 2020 Jan.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699544131038}, {"id": "yDxOLxEyTi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6515/Reviewer_AsFx"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes an information theorem based multi-attribute selective suppression (MaSS) to solve the problem of highlighting the utility attributes while suppressing the private attributes. The introduced problem is interesting and important, as privacy becomes a central concern for many applications. The paper presents a clear pipeline leveraging three effort stream lines: (1) sensitive attribute suppression, (2) annotated useful attribute preservation and (3) unannotated useful attribute preservation. The optimization are then jointly optimized.", "review_text": "This paper proposes an information theorem based multi-attribute selective suppression (MaSS) to solve the problem of highlighting the utility attributes while suppressing the private attributes. The introduced problem is interesting and important, as privacy becomes a central concern for many applications. The paper presents a clear pipeline leveraging three effort stream lines: (1) sensitive attribute suppression, (2) annotated useful attribute preservation and (3) unannotated useful attribute preservation. The optimization are then jointly optimized.", "strengths": "1. The paper study into an interesting and practical problem by discussing the limitations and compare to the literature approaches, providing a sufficient background for the problem study.\n\n2. The paper provides a theoretical analysis from the information theory perspective, showing the relationship between utility and sensitive attributes.\n\n3. The paper presents a clear and tractable learning scheme to achieve the three stream lines.\n\n4. There are extensive experimental comparison against the representative literature methods and some state-of-the-arts. Consistently advantageous results demonstrate the method’s effectiveness.", "weaknesses": "1. For the sensitive attribute suppression, the objective is to minimize the the expectation of the entropy between P(Si|x) and P_phi(Si|X’). Drawing connection to adversarial learning, it tries to push P_phi(Si|x’) close to P(Si|x) by pushing the discriminator cannot tell the difference between the two.\n\nFirstly, the paper lacks the interpretation of their proposed method, and drawing the connection to the literature method, e.g., adversarial learning. It would be good the authors can conduct an in-depth analysis comparing the literature to the proposal in this paper, and further highlight the method’s novelty.\n\n2. For unannotated useful attribute, depending on the definition of the problem, the setting will be different from the other method. In the paper of experiments, the authors mention “ALR, BDQ and PPDAR overlook the preservation of unannotated useful attributes”.\n\nIt could be that those methods, from their problem definition and setting, they do not consider so termed “unannotated useful attributes” into their framework. But one cannot say it is the limitation or fault of those methods. In the most fair way, because of setting difference, this paper should compare to only those considering “unannotated useful attributes”. Please carefully phrase the comparison to other methods.\n\n3. Still, for those methods that are sharing exactly the same setting, e.g., GAP and MSDA, from technical frame design, what is the difference? I noticed there is some slight comparison, e.g., arguing that some of the methods lack theoretical analysis. This is the advantage of this paper. But other than that, if there is an empirical design that is exactly the same as this paper, this paper will only go for the theoretical contribution.\n\nThus, please provide a towards thorough comparison to those literature under the same setting, which will be helpful to claim the method contribution and novelty.", "questions": "Please refer to weakness session for detail.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an information theorem based multi-attribute selective suppression (MaSS) to solve the problem of highlighting the utility attributes while suppressing the private attributes. The introduced problem is interesting and important, as privacy becomes a central concern for many applications. The paper presents a clear pipeline leveraging three effort stream lines: (1) sensitive attribute suppression, (2) annotated useful attribute preservation and (3) unannotated useful attribute preservation. The optimization are then jointly optimized.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper study into an interesting and practical problem by discussing the limitations and compare to the literature approaches, providing a sufficient background for the problem study.\n\n2. The paper provides a theoretical analysis from the information theory perspective, showing the relationship between utility and sensitive attributes.\n\n3. The paper presents a clear and tractable learning scheme to achieve the three stream lines.\n\n4. There are extensive experimental comparison against the representative literature methods and some state-of-the-arts. Consistently advantageous results demonstrate the method’s effectiveness.", "weaknesses": "1. For the sensitive attribute suppression, the objective is to minimize the the expectation of the entropy between P(Si|x) and P_phi(Si|X’). Drawing connection to adversarial learning, it tries to push P_phi(Si|x’) close to P(Si|x) by pushing the discriminator cannot tell the difference between the two.\n\nFirstly, the paper lacks the interpretation of their proposed method, and drawing the connection to the literature method, e.g., adversarial learning. It would be good the authors can conduct an in-depth analysis comparing the literature to the proposal in this paper, and further highlight the method’s novelty.\n\n2. For unannotated useful attribute, depending on the definition of the problem, the setting will be different from the other method. In the paper of experiments, the authors mention “ALR, BDQ and PPDAR overlook the preservation of unannotated useful attributes”.\n\nIt could be that those methods, from their problem definition and setting, they do not consider so termed “unannotated useful attributes” into their framework. But one cannot say it is the limitation or fault of those methods. In the most fair way, because of setting difference, this paper should compare to only those considering “unannotated useful attributes”. Please carefully phrase the comparison to other methods.\n\n3. Still, for those methods that are sharing exactly the same setting, e.g., GAP and MSDA, from technical frame design, what is the difference? I noticed there is some slight comparison, e.g., arguing that some of the methods lack theoretical analysis. This is the advantage of this paper. But other than that, if there is an empirical design that is exactly the same as this paper, this paper will only go for the theoretical contribution.\n\nThus, please provide a towards thorough comparison to those literature under the same setting, which will be helpful to claim the method contribution and novelty.", "questions": "Please refer to weakness session for detail.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699224773356}, {"id": "GPvmPubwHv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6515/Reviewer_YVvf"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper presents a novel approach, referred to as MASS (Multi-Attribute Selective Suppression), which addresses the challenge of privacy protection in the context of large-scale datasets used for machine learning. It introduces a formal information-theoretic definition for utility-preserving privacy protection and offers a data-driven, learnable data transformation framework. This framework enables the selective suppression of sensitive attributes while preserving other useful attributes, regardless of whether they are known in advance or explicitly annotated. The paper includes rigorous theoretical analyses of the operational bounds of the proposed framework and conducts extensive experimental evaluations across diverse modalities, such as facial images, voice audio clips, and motion sensor signals. The results demonstrate the effectiveness and generalizability of MASS across different tasks and configurations.", "review_text": "The paper presents a novel approach, referred to as MASS (Multi-Attribute Selective Suppression), which addresses the challenge of privacy protection in the context of large-scale datasets used for machine learning. It introduces a formal information-theoretic definition for utility-preserving privacy protection and offers a data-driven, learnable data transformation framework. This framework enables the selective suppression of sensitive attributes while preserving other useful attributes, regardless of whether they are known in advance or explicitly annotated. The paper includes rigorous theoretical analyses of the operational bounds of the proposed framework and conducts extensive experimental evaluations across diverse modalities, such as facial images, voice audio clips, and motion sensor signals. The results demonstrate the effectiveness and generalizability of MASS across different tasks and configurations.", "strengths": "1. One of the notable strengths of this paper is the introduction of a formal information-theoretic definition for utility-preserving privacy protection. This theoretical foundation provides a solid framework for addressing privacy concerns in large-scale datasets, contributing to the theoretical underpinning of data privacy solutions.\n\n2. The proposal of a data-driven learnable data transformation framework is innovative. This approach allows for the selective suppression of sensitive attributes, enhancing privacy protection while preserving the utility of the data. \n\n3. The comprehensive experimental evaluations across various data modalities, including facial images, voice audio clips, and motion sensor signals, highlight the generalizability of the MASS framework. This breadth of experimentation underlines its versatility and applicability to a wide range of real-world scenarios.", "weaknesses": "1. The paper introduces an interesting concept in Theorem 3.1, highlighting the importance of mutual information constraints 'm' and 'n' in the context of privacy and utility trade-offs. However, it is essential to note that the experiments lack a corresponding exploration of these constraints. These constraints likely play a pivotal role in balancing sensitive attribute accuracy and useful attribute accuracy. The reported experimental results suggest that MaSS might not simultaneously achieve the best performance for both types of attributes. Therefore, it is recommended to conduct experiments with varying constraints to gain a deeper understanding of their impact.\n\n2. The theoretical argument presented in Theorem 3.2 may raise some questions. Specifically, the relevance of unannotated useful attributes to the tasks is highlighted, which could vary across different scenarios. The paper should address this issue and provide an ablation study of the contrastive learning module to support the claims made in Theorem 3.2. This would provide stronger evidence and clarity regarding the relationship between learned attributes and sensitive attributes.\n\n3. Clarification is needed on how the positive and negative samples for the InfoNCE loss are determined. Figure 3 suggests that both positive and negative samples come from the transformed data X', but the paper should explain how these samples are chosen, given that the anchor sample is the original data X.\n\n4. Regarding the evaluation of sensitive attribute accuracy, it appears that the accuracy of the adversarial classifier is used. It is recommended to consider the approach of training a classifier from scratch on the transformed data, similar to the methodology employed for calculating useful attribute accuracy.\n\n5. To provide a more comprehensive assessment of the proposed method, the paper should include comparisons with recent baselines, such as SPAct (CVPR 2022 [1]).\n\n6. The topic of concept removal for generative models, while not a central focus, could be related to this paper's context. It is suggested to discuss concept removal in the related works section to provide a broader perspective on the field and to highlight the paper's contributions in relation to existing research.\n\n[1] Dave, Ishan Rajendrakumar, Chen Chen, and Mubarak Shah. \"Spact: Self-supervised privacy preservation for action recognition.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.", "questions": "Please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a novel approach, referred to as MASS (Multi-Attribute Selective Suppression), which addresses the challenge of privacy protection in the context of large-scale datasets used for machine learning. It introduces a formal information-theoretic definition for utility-preserving privacy protection and offers a data-driven, learnable data transformation framework. This framework enables the selective suppression of sensitive attributes while preserving other useful attributes, regardless of whether they are known in advance or explicitly annotated. The paper includes rigorous theoretical analyses of the operational bounds of the proposed framework and conducts extensive experimental evaluations across diverse modalities, such as facial images, voice audio clips, and motion sensor signals. The results demonstrate the effectiveness and generalizability of MASS across different tasks and configurations.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. One of the notable strengths of this paper is the introduction of a formal information-theoretic definition for utility-preserving privacy protection. This theoretical foundation provides a solid framework for addressing privacy concerns in large-scale datasets, contributing to the theoretical underpinning of data privacy solutions.\n\n2. The proposal of a data-driven learnable data transformation framework is innovative. This approach allows for the selective suppression of sensitive attributes, enhancing privacy protection while preserving the utility of the data. \n\n3. The comprehensive experimental evaluations across various data modalities, including facial images, voice audio clips, and motion sensor signals, highlight the generalizability of the MASS framework. This breadth of experimentation underlines its versatility and applicability to a wide range of real-world scenarios.", "weaknesses": "1. The paper introduces an interesting concept in Theorem 3.1, highlighting the importance of mutual information constraints 'm' and 'n' in the context of privacy and utility trade-offs. However, it is essential to note that the experiments lack a corresponding exploration of these constraints. These constraints likely play a pivotal role in balancing sensitive attribute accuracy and useful attribute accuracy. The reported experimental results suggest that MaSS might not simultaneously achieve the best performance for both types of attributes. Therefore, it is recommended to conduct experiments with varying constraints to gain a deeper understanding of their impact.\n\n2. The theoretical argument presented in Theorem 3.2 may raise some questions. Specifically, the relevance of unannotated useful attributes to the tasks is highlighted, which could vary across different scenarios. The paper should address this issue and provide an ablation study of the contrastive learning module to support the claims made in Theorem 3.2. This would provide stronger evidence and clarity regarding the relationship between learned attributes and sensitive attributes.\n\n3. Clarification is needed on how the positive and negative samples for the InfoNCE loss are determined. Figure 3 suggests that both positive and negative samples come from the transformed data X', but the paper should explain how these samples are chosen, given that the anchor sample is the original data X.\n\n4. Regarding the evaluation of sensitive attribute accuracy, it appears that the accuracy of the adversarial classifier is used. It is recommended to consider the approach of training a classifier from scratch on the transformed data, similar to the methodology employed for calculating useful attribute accuracy.\n\n5. To provide a more comprehensive assessment of the proposed method, the paper should include comparisons with recent baselines, such as SPAct (CVPR 2022 [1]).\n\n6. The topic of concept removal for generative models, while not a central focus, could be related to this paper's context. It is suggested to discuss concept removal in the related works section to provide a broader perspective on the field and to highlight the paper's contributions in relation to existing research.\n\n[1] Dave, Ishan Rajendrakumar, Chen Chen, and Mubarak Shah. \"Spact: Self-supervised privacy preservation for action recognition.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.", "questions": "Please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698796639606}, {"id": "Vdgq2lWcvG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6515/Reviewer_CAg3"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a framework called MASS (Multi-Attribute Selective Suppression) for privacy-preserving data transformation that selectively suppresses sensitive attributes while preserving useful ones. The authors provide a formal definition of privacy protection and an information-theoretic perspective on the problem. They also present a data-driven approach that uses a combination of supervised and unsupervised learning to identify sensitive attributes and suppress them while preserving useful ones. The authors provide rigorous theoretical analyses and comprehensive experimental evaluations that demonstrate the effectiveness of their approach. The contributions of this paper include a formal definition of privacy protection, a data-driven framework for privacy-preserving data transformation, and a comprehensive evaluation of the proposed approach against several baseline methods using multiple datasets of varying modalities.", "review_text": "This paper proposes a framework called MASS (Multi-Attribute Selective Suppression) for privacy-preserving data transformation that selectively suppresses sensitive attributes while preserving useful ones. The authors provide a formal definition of privacy protection and an information-theoretic perspective on the problem. They also present a data-driven approach that uses a combination of supervised and unsupervised learning to identify sensitive attributes and suppress them while preserving useful ones. The authors provide rigorous theoretical analyses and comprehensive experimental evaluations that demonstrate the effectiveness of their approach. The contributions of this paper include a formal definition of privacy protection, a data-driven framework for privacy-preserving data transformation, and a comprehensive evaluation of the proposed approach against several baseline methods using multiple datasets of varying modalities.", "strengths": "S1. In terms of originality, the paper introduces a novel approach for protecting unannotated attributes in datasets. While previous works have focused on protecting annotated attributes or using heuristics, the paper proposes a data-driven learnable data transformation framework called MaSS (Multi-Attribute Selective Suppression) that can selectively suppress sensitive attributes while preserving other useful attributes, regardless of whether they are known in advance or explicitly annotated. This approach is unique and addresses a gap in the existing literature.\nS2. The quality of the paper is high, as it provides rigorous theoretical analyses of the operational bounds of the proposed framework. The authors derive mathematical formulations and provide proofs for the theorems presented in the paper [5, 7]. This demonstrates a strong understanding of the underlying principles and ensures the reliability of the proposed methods.\nS3. The clarity of the paper is commendable. The authors provide clear explanations of the problem formulation, the proposed techniques, and the evaluation methodology. The paper is well-structured, making it easy for readers to follow the flow of ideas. Additionally, the authors provide visualizations and tables to support their findings.", "weaknesses": "W1. The dataset lacks a detailed description. It would also be good to have a table that describes the size of the dataset along with some other information that would give the reader a clearer picture of the dataset.\nW2. All six methods of experimental comparison rely on adversarially training a sensitive attribute inference model and lack the ability to compare state-of-the-art dp-based methods (e.g. \"Mingxuan Sun, Qing Wang, Zicheng Liu: Human Action Image Generation with Differential Privacy. ICME 2020: 1-6\").\nW3. Lack of comparison with a state-of-the-art method (\"Li M, Xu X, Fan H, et al. STPrivacy: Spatio-Temporal Privacy-Preserving Action Recognition[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 5106-5115.\")", "questions": "The authors need provide more detailed explanations and justifications for their proposed techniques. why contrastive learning is suitable for protecting unannotated attributes and how it ensures the predictability of annotated attributes? Additionally, the authors could consider providing more detailed explanations of the loss functions used in their method, such as the InfoNCE Contrastive Learning Loss.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a framework called MASS (Multi-Attribute Selective Suppression) for privacy-preserving data transformation that selectively suppresses sensitive attributes while preserving useful ones. The authors provide a formal definition of privacy protection and an information-theoretic perspective on the problem. They also present a data-driven approach that uses a combination of supervised and unsupervised learning to identify sensitive attributes and suppress them while preserving useful ones. The authors provide rigorous theoretical analyses and comprehensive experimental evaluations that demonstrate the effectiveness of their approach. The contributions of this paper include a formal definition of privacy protection, a data-driven framework for privacy-preserving data transformation, and a comprehensive evaluation of the proposed approach against several baseline methods using multiple datasets of varying modalities.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "S1. In terms of originality, the paper introduces a novel approach for protecting unannotated attributes in datasets. While previous works have focused on protecting annotated attributes or using heuristics, the paper proposes a data-driven learnable data transformation framework called MaSS (Multi-Attribute Selective Suppression) that can selectively suppress sensitive attributes while preserving other useful attributes, regardless of whether they are known in advance or explicitly annotated. This approach is unique and addresses a gap in the existing literature.\nS2. The quality of the paper is high, as it provides rigorous theoretical analyses of the operational bounds of the proposed framework. The authors derive mathematical formulations and provide proofs for the theorems presented in the paper [5, 7]. This demonstrates a strong understanding of the underlying principles and ensures the reliability of the proposed methods.\nS3. The clarity of the paper is commendable. The authors provide clear explanations of the problem formulation, the proposed techniques, and the evaluation methodology. The paper is well-structured, making it easy for readers to follow the flow of ideas. Additionally, the authors provide visualizations and tables to support their findings.", "weaknesses": "W1. The dataset lacks a detailed description. It would also be good to have a table that describes the size of the dataset along with some other information that would give the reader a clearer picture of the dataset.\nW2. All six methods of experimental comparison rely on adversarially training a sensitive attribute inference model and lack the ability to compare state-of-the-art dp-based methods (e.g. \"Mingxuan Sun, Qing Wang, Zicheng Liu: Human Action Image Generation with Differential Privacy. ICME 2020: 1-6\").\nW3. Lack of comparison with a state-of-the-art method (\"Li M, Xu X, Fan H, et al. STPrivacy: Spatio-Temporal Privacy-Preserving Action Recognition[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 5106-5115.\")", "questions": "The authors need provide more detailed explanations and justifications for their proposed techniques. why contrastive learning is suitable for protecting unannotated attributes and how it ensures the predictability of annotated attributes? Additionally, the authors could consider providing more detailed explanations of the loss functions used in their method, such as the InfoNCE Contrastive Learning Loss.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698654136490}], "openreview_url": "https://openreview.net/forum?id=JAKcnjzQI3", "arxiv_id": "2405.14981", "paper_pdf": "papers/JAKcnjzQI3.pdf", "paper_pdf_sha256": "bc025ab9a5d9beab17e33812b06072106fd2e9a3aca80076251b3b69e077fc89", "paper_pdf_bytes": 683556, "paper_pdf_source": "openreview", "code_url": "https://github.com/jpmorganchase/MaSS", "code_repository": "jpmorganchase/MaSS", "code_commit": "6fbe9be1ff155b3fdd2677de7977d801d10000b5", "code_archive": "repos/JAKcnjzQI3.zip", "code_archive_sha256": "9ed8b9f55c5a55c545af2227c9fd5d9b81e99bd585b6c93e2d7525f5d72dfbdb", "code_archive_bytes": 228313, "code_file_count": 25, "code_extensions": {".py": 23, ".sh": 2}, "github_disk_usage_kb": 217, "github_languages": {"Python": 104454, "Shell": 6389}, "github_archived": true, "github_pushed_at": "2024-07-25T10:20:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mass-multi-attribute-selective-suppression-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cVEOiBz2Em", "year": 2023, "status": "rejected", "title": "Deep Bayesian Active Learning for Accelerating Stochastic Simulation", "authors": ["Dongxia Wu", "Ruijia Niu", "Matteo Chinazzi", "Alessandro Vespignani", "Yian Ma", "Rose Yu"], "authorids": ["~Dongxia_Wu1", "~Ruijia_Niu1", "~Matteo_Chinazzi1", "~Alessandro_Vespignani1", "~Yian_Ma1", "~Rose_Yu1"], "authors_source": "OpenReview API", "abstract": "Stochastic simulations such as large-scale, spatiotemporal, age-structured epidemic models are computationally expensive at fine-grained resolution. While deep surrogate models can speed up the simulations, doing so for stochastic simulations and with active learning approaches is an underexplored area. We propose Interactive Neural Process (INP), a deep Bayesian active learning framework for learning deep surrogate models to accelerate stochastic simulations. INP consists of two components, a spatiotemporal surrogate model built upon Neural Process (NP) family and an acquisition function for active learning. For surrogate modeling, we develop Spatiotemporal Neural Process (STNP) to mimic the simulator dynamics. For active learning, we propose a novel acquisition function, Latent Information Gain (LIG), calculated in the latent space of NP based models. We perform a theoretical analysis and demonstrate that LIG reduces sample complexity compared with random sampling in high dimensions. We also conduct empirical studies on two complex spatiotemporal simulators for reaction diffusion and infectious disease. The results demonstrate that STNP outperforms the baselines in the offline learning setting and LIG achieves the state-of-the-art for Bayesian active learning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "jxUgBUlIkfg", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3385/Reviewer_eypg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors proposed a unified framework with Interactive Neural Processes (INP) as the surrogate model backbone and the LIG acquisition function. The performance of the proposed method was tested on a few benchmark examples. ", "review_text": "Well-written paper, but the novelty and experiments could be improved. ", "strengths": "Strength: the paper is well motivated and well written.\n\nWeakness:\n\n1. Novelty. I appreciate the effort to propose the spatio-temporal extension to the original NP as well as the latent extension to EIG, but I am not convinced that this is sufficiently novel for ICLR standard.\n\n2. The experimental part is also limited. I would expect more ablation study so that we can fully understand the sensitivity and robustness of the proposed approach. \n\n3. Some implementation details are not provided it seems.  How does the active learning get actually implemented? When a new set of data is added, does the INP get retrained from scratch, or fine-tuning from the existing weights? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors proposed a unified framework with Interactive Neural Processes (INP) as the surrogate model backbone and the LIG acquisition function. The performance of the proposed method was tested on a few benchmark examples. ", "strength_and_weaknesses": "Strength: the paper is well motivated and well written.\n\nWeakness:\n\n1. Novelty. I appreciate the effort to propose the spatio-temporal extension to the original NP as well as the latent extension to EIG, but I am not convinced that this is sufficiently novel for ICLR standard.\n\n2. The experimental part is also limited. I would expect more ablation study so that we can fully understand the sensitivity and robustness of the proposed approach. \n\n3. Some implementation details are not provided it seems.  How does the active learning get actually implemented? When a new set of data is added, does the INP get retrained from scratch, or fine-tuning from the existing weights? ", "clarity,_quality,_novelty_and_reproducibility": "Overall this paper is clear to read, well written. The novelty is, as previously explained, IMO, not sufficient. The code is included, so I assume that the reproducibility should be high. ", "summary_of_the_review": "Well-written paper, but the novelty and experiments could be improved. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667676966199}, {"id": "tKZYphugek", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3385/Reviewer_k2B6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This study proposes a new method for learning surrogate models of stochastic simulators. The new method, Interactive Neural Process (INP), builds on Neural processes and leverages the spatiotemporal structure of the problems at hand to reduce the complexity of the inference task. On three problems -- a low dimensional SEIR problem, a low dimensional reaction-diffusion problem and a complex spatiotemporal LEAM-US problem --, the approach is shown to appropriately learn the surrogate in few iterations.", "review_text": "Overall, the paper is technically sound and has a novel methodological contribution with convincing empirical results. This will inspire further method development. Some of the discussion of previous work needs to be revised and augmented.\n", "strengths": "The paper is technically sound, and claims are well supported by empirical evaluation.\n\nSome previous work is cited and discussed, and it is clear how this work differs from these previous contributions. However, I believe it would be important to revise the section on related work. In particular:\n\n-the authors write \"Kleinegesse & Gutmann (2020) [...] require a explicit likelihood model, conditional independence in experiments, and are limited to low (1-2) dimensional design\". I would urge the authors to clarify this statement. In particular, Kleinegesse & Gutmann (2020) deal with simulators for which no likelihood is available (also called \"implicit models\"), which is contrary to what is stated in this manuscript;\n\n-related to the point above, there is a large body of work in machine learning on surrogate models for stochastic simulators (e.g. Meeds and Welling 2014 \"GPS-ABC: Gaussian Process Surrogate Approximate Bayesian Computation\"; Gutmann and Corander 2016 \"Bayesian Optimization for Likelihood-Free Inference of Simulator-Based Statistical Models\"; Lueckmann et al. 2018 \"Likelihood-free inference with emulator networks\"; Järvenpää et al. 2019 \"Efficient acquisition rules for model-based approximate bayesian computation\"; Papamakarios et al. 2019 \"Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows\"), some of which use active learning schemes to adaptively choose model parameters for acquiring new simulations (e.g. Meeds and Welling 2014; Gutmann and Corander 2016; Lueckmann et al. 2018; Järvenpää et al. 2019). It would be important to appropriately discuss this work in the manuscript.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This study proposes a new method for learning surrogate models of stochastic simulators. The new method, Interactive Neural Process (INP), builds on Neural processes and leverages the spatiotemporal structure of the problems at hand to reduce the complexity of the inference task. On three problems -- a low dimensional SEIR problem, a low dimensional reaction-diffusion problem and a complex spatiotemporal LEAM-US problem --, the approach is shown to appropriately learn the surrogate in few iterations.", "strength_and_weaknesses": "The paper is technically sound, and claims are well supported by empirical evaluation.\n\nSome previous work is cited and discussed, and it is clear how this work differs from these previous contributions. However, I believe it would be important to revise the section on related work. In particular:\n\n-the authors write \"Kleinegesse & Gutmann (2020) [...] require a explicit likelihood model, conditional independence in experiments, and are limited to low (1-2) dimensional design\". I would urge the authors to clarify this statement. In particular, Kleinegesse & Gutmann (2020) deal with simulators for which no likelihood is available (also called \"implicit models\"), which is contrary to what is stated in this manuscript;\n\n-related to the point above, there is a large body of work in machine learning on surrogate models for stochastic simulators (e.g. Meeds and Welling 2014 \"GPS-ABC: Gaussian Process Surrogate Approximate Bayesian Computation\"; Gutmann and Corander 2016 \"Bayesian Optimization for Likelihood-Free Inference of Simulator-Based Statistical Models\"; Lueckmann et al. 2018 \"Likelihood-free inference with emulator networks\"; Järvenpää et al. 2019 \"Efficient acquisition rules for model-based approximate bayesian computation\"; Papamakarios et al. 2019 \"Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows\"), some of which use active learning schemes to adaptively choose model parameters for acquiring new simulations (e.g. Meeds and Welling 2014; Gutmann and Corander 2016; Lueckmann et al. 2018; Järvenpää et al. 2019). It would be important to appropriately discuss this work in the manuscript.", "clarity,_quality,_novelty_and_reproducibility": "To my knowledge, the proposed method is novel. Furthermore, the paper is technically sound.\n\nThe paper is generally clear, and I would just like to point a few typos and imprecisions:\n\n-\"Stochastic simulations are fundamental to many scientific fields (Ripley, 2009), especially epidemic modeling\". As it is phrased, it seems to suggest that stochastic stimulations are more important in epidemiology than in other scientific fields (which is arguably not true), whereas the authors probably wanted to write that stochastic simulations are important in epidemiology;\n\n-\"Noted if the spatial dependency\", instead \"Note that if the spatial dependency\";\n\n-\"is replaced to Convolutional LSTM layer\", instead \"is replaced by a Convolutional LSTM layer\";\n\n-\"are independent with each other\", instead \"are independent of each other\";\n\n-probably better to drop the word \"acquiring\" in \"as acquiring points that are informative individually are not necessarily informative jointly\";\n\n-in \"The total number of dimension\", it should be \"dimensions\". There are a few other instances of \"dimension\" where this should also be corrected;\n\n-\"Figure 3 right visualize the STNP predictions\", instead \"Figure 3 right shows the STNP predictions\";\n\n-Figure 5, yellow dots and red stars need to be larger for easier visualisation.", "summary_of_the_review": "Overall, the paper is technically sound and has a novel methodological contribution with convincing empirical results. This will inspire further method development. Some of the discussion of previous work needs to be revised and augmented.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667504358568}, {"id": "2B0LwEkfqd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3385/Reviewer_gfsx"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present a full-stack solution to accelerating learning, estimation and inference in complex simulators.  They present a novel neural process architecture with spatial and temporal dependencies.  They also present a new Bayesian active learning acquisition function that is particularly elegant in the context of STNP.  Some theoretical results are also presented.  The method is benchmarked with three examples (two epidemiological models and one reaction model), and their model appears to perform well.  ", "review_text": "The authors present a method of expediting inference in expensive simulators.  The paper is reasonably well written, timely, and novel.  I have some reservations about the presentation as it stands, but these are not critical enough for me to recommend rejecting the paper.  If the authors can answer my comments above, then I would endorse the publication of this paper.  \n\n(I checked the proof for the proposition, but only skimmed the longer proof for the theorem.)\n", "strengths": "\n# Strengths\nThe basic premise of the paper is reasonably clear, it is reasonably well written and referenced, and is timely.  The model itself seems well-grounded and sensible.  Evaluating the acquisition function in the latent space is a nice idea.  The experiments use realistic models (if not real data).  The theoretical contributions are not huge, but are reasonably well-suited to the topic, and provide some nice additional material to compliment the empirical results. \n\n# Weaknesses\nI have three main questions/comments on the work.  \n\n## (1)  What is being acquired?\nSomething that didn’t come across particularly clearly is how the acquisition function is actually used.  In S3.4, the acquisition function is defined in terms of $x$, but in F5, the acquisition function is plotted as a function of $\\theta$.  There are also many $\\hat{x}$’s defined across different values of $\\theta$.  Are you able to clarify for me how the reduction in entropy between the posterior and prior (eq (3)) is a sensible way to select a theta, and how this reduction is actually computed?  The reduction in entropy is given a particular $\\theta$, and so we are looking to reduce the uncertainty over the latent state values for the observed data?  Does this not “reward” parameter values for which the prior has enormous variance, even if they have low probability?  If the data are generated from a wildly different parameter set, then I would expect either a huge entropy (the model is uncertain) or a tiny entropy (the model is in some way “degenerate”), which is going to skew the LIG value. \n\nIt is not clear to me that minimising the uncertainty in a latent space is the right way to go.  I invite the authors to restate the objectives of the LIG, how it works, and the weaknesses of it (in light of my comments above).  \n\n## (2)  Deployment Pipeline\nThe introduction talks about flexible approximators and fast inference.  Are the authors able to re-state how the envisage this pipeline being used?  When reading the introduction, i thought you were trying to learn a universal approximator to the mapping from parameters to outcomes, which can then be polled in ``war time’’ quickly.  Later on though, when talking about active learning, it is like you are trying to learn a function approximator _quickly_ _during_ war time.  \n\nIf you are learning a fixed function approximator for all of parameters and state space, then you are hugely limited by the expressivity of the model.  If you are trying to fast model learning at deployment time, then I query why the recent work on Synthetic Neural Likelihoods was not compared to, as this to me seems to be the gold standard.  \n\n## (3)  Empirical Evaluation\nTo my eye, the empirical validation is not especially strong.  It performs better on some non-standard tasks, but I still don’t have a great handle on how well it is actually doing – are the experiments just tuned to the STNP method?  Are the baselines poorly tuned?  I don’t currently know.  I would encourage the authors to find an example to apply the STNP to from the literature, and compare to existing results.  \n\nI would even be satisfied by seeing something of an unreasonable baseline;  e.g. train an RNN (using VI) to predict the mean and variances of the trajectories,  train an MLP to predict the final compartment counts in the SEIR as a function of the input parameters – just something to give the reader a bearing on how an “alternative” and “simple” method would perform.  \n\nThe active learning component is nice, but again, I would like to see it compared on a standard task – evaluating the efficacy of the acquisition function is difficult, and I can’t really tell what is happening in F5.  Is there a 1-D model (in time as well) that we can apply this to where the latent space is easily visualizable, we can determine how well the STNP is able to reproduce the latent space etc.  \n\nI also think a readily interpretable/toy example would actually increase the reach of this work, demonstrating the utility of the method in a general example, as opposed to highly specific domains.  I do not expect this to be added during the rebuttal period, but would expect to see something like this for any camera ready version.\n\n\n# Minor Weaknesses / Typographical Comments.\n(a)  The paper is a little wordy in places.  I think a lot of the material could be simplified and shortened, and actually dramatically improve the quality of the paper.  Although this criticism is throughout the paper, some sections I thought were particularly wordy/can be scythed in length:  BALD paragraph in S3.4;  the whole of S3.2;  The first two paragraphs of the introduction;  Most of S4.1 can be dropped to the supplement – it adds very little to the general reader.  \n\n(b)  Algorithm 1 is really not as useful as it could be.  I would like to see this algorithm blown up, with inline comments, in its own section etc.  It also seems like the labels in Alg 1 are not correct (particularly “learn”).  \n\n(c)  \\citep and \\citet are used incorrectly in places (i.e. Garnelo et al, pg2).\n\n(d)  You’ve vspaced section 3.3.  Please try to avoid doing this.  (especially cf. (a)!)\n\n(e)  There are generally quite a few typos and grammatical mistakes.  Any camera ready would require extensive proofreading beyond the submitted version.  “In example” (pg 1), “remained” (pg 6), “a simplifying situation” (pg 7) for example.  \n\n(f)  In the proof of proposition 1, are you able to re-iterate how the $\\theta$ appears in the denominator of the first line (I may well just be having a mental blank here).\n\n(g)  “figure” should be capitalised, i.e. “Figure 1”.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors present a full-stack solution to accelerating learning, estimation and inference in complex simulators.  They present a novel neural process architecture with spatial and temporal dependencies.  They also present a new Bayesian active learning acquisition function that is particularly elegant in the context of STNP.  Some theoretical results are also presented.  The method is benchmarked with three examples (two epidemiological models and one reaction model), and their model appears to perform well.  ", "strength_and_weaknesses": "\n# Strengths\nThe basic premise of the paper is reasonably clear, it is reasonably well written and referenced, and is timely.  The model itself seems well-grounded and sensible.  Evaluating the acquisition function in the latent space is a nice idea.  The experiments use realistic models (if not real data).  The theoretical contributions are not huge, but are reasonably well-suited to the topic, and provide some nice additional material to compliment the empirical results. \n\n# Weaknesses\nI have three main questions/comments on the work.  \n\n## (1)  What is being acquired?\nSomething that didn’t come across particularly clearly is how the acquisition function is actually used.  In S3.4, the acquisition function is defined in terms of $x$, but in F5, the acquisition function is plotted as a function of $\\theta$.  There are also many $\\hat{x}$’s defined across different values of $\\theta$.  Are you able to clarify for me how the reduction in entropy between the posterior and prior (eq (3)) is a sensible way to select a theta, and how this reduction is actually computed?  The reduction in entropy is given a particular $\\theta$, and so we are looking to reduce the uncertainty over the latent state values for the observed data?  Does this not “reward” parameter values for which the prior has enormous variance, even if they have low probability?  If the data are generated from a wildly different parameter set, then I would expect either a huge entropy (the model is uncertain) or a tiny entropy (the model is in some way “degenerate”), which is going to skew the LIG value. \n\nIt is not clear to me that minimising the uncertainty in a latent space is the right way to go.  I invite the authors to restate the objectives of the LIG, how it works, and the weaknesses of it (in light of my comments above).  \n\n## (2)  Deployment Pipeline\nThe introduction talks about flexible approximators and fast inference.  Are the authors able to re-state how the envisage this pipeline being used?  When reading the introduction, i thought you were trying to learn a universal approximator to the mapping from parameters to outcomes, which can then be polled in ``war time’’ quickly.  Later on though, when talking about active learning, it is like you are trying to learn a function approximator _quickly_ _during_ war time.  \n\nIf you are learning a fixed function approximator for all of parameters and state space, then you are hugely limited by the expressivity of the model.  If you are trying to fast model learning at deployment time, then I query why the recent work on Synthetic Neural Likelihoods was not compared to, as this to me seems to be the gold standard.  \n\n## (3)  Empirical Evaluation\nTo my eye, the empirical validation is not especially strong.  It performs better on some non-standard tasks, but I still don’t have a great handle on how well it is actually doing – are the experiments just tuned to the STNP method?  Are the baselines poorly tuned?  I don’t currently know.  I would encourage the authors to find an example to apply the STNP to from the literature, and compare to existing results.  \n\nI would even be satisfied by seeing something of an unreasonable baseline;  e.g. train an RNN (using VI) to predict the mean and variances of the trajectories,  train an MLP to predict the final compartment counts in the SEIR as a function of the input parameters – just something to give the reader a bearing on how an “alternative” and “simple” method would perform.  \n\nThe active learning component is nice, but again, I would like to see it compared on a standard task – evaluating the efficacy of the acquisition function is difficult, and I can’t really tell what is happening in F5.  Is there a 1-D model (in time as well) that we can apply this to where the latent space is easily visualizable, we can determine how well the STNP is able to reproduce the latent space etc.  \n\nI also think a readily interpretable/toy example would actually increase the reach of this work, demonstrating the utility of the method in a general example, as opposed to highly specific domains.  I do not expect this to be added during the rebuttal period, but would expect to see something like this for any camera ready version.\n\n\n# Minor Weaknesses / Typographical Comments.\n(a)  The paper is a little wordy in places.  I think a lot of the material could be simplified and shortened, and actually dramatically improve the quality of the paper.  Although this criticism is throughout the paper, some sections I thought were particularly wordy/can be scythed in length:  BALD paragraph in S3.4;  the whole of S3.2;  The first two paragraphs of the introduction;  Most of S4.1 can be dropped to the supplement – it adds very little to the general reader.  \n\n(b)  Algorithm 1 is really not as useful as it could be.  I would like to see this algorithm blown up, with inline comments, in its own section etc.  It also seems like the labels in Alg 1 are not correct (particularly “learn”).  \n\n(c)  \\citep and \\citet are used incorrectly in places (i.e. Garnelo et al, pg2).\n\n(d)  You’ve vspaced section 3.3.  Please try to avoid doing this.  (especially cf. (a)!)\n\n(e)  There are generally quite a few typos and grammatical mistakes.  Any camera ready would require extensive proofreading beyond the submitted version.  “In example” (pg 1), “remained” (pg 6), “a simplifying situation” (pg 7) for example.  \n\n(f)  In the proof of proposition 1, are you able to re-iterate how the $\\theta$ appears in the denominator of the first line (I may well just be having a mental blank here).\n\n(g)  “figure” should be capitalised, i.e. “Figure 1”.\n", "clarity,_quality,_novelty_and_reproducibility": "The work appears novel, and builds on recent methods in the field.  The method itself is reasonably well presented, and I could understand a lot of the material, despite not being an NP expert.  Code is released. \n", "summary_of_the_review": "The authors present a method of expediting inference in expensive simulators.  The paper is reasonably well written, timely, and novel.  I have some reservations about the presentation as it stands, but these are not critical enough for me to recommend rejecting the paper.  If the authors can answer my comments above, then I would endorse the publication of this paper.  \n\n(I checked the proof for the proposition, but only skimmed the longer proof for the theorem.)\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666223607304}, {"id": "g7kiVoTsUW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3385/Reviewer_fkSr"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper uses the Neural Process model as a surrogate trainable model of the input-output function in the context of simulation algorithms for epidemic modeling. This bypasses numerical integration that is required from non-machine learning simulation software. They modify the original model of the Neural Process to allow for better incorporating spatiotemporal dependencies. \n\nIn addition, instead of learning passively, the authors design active learning algorithms to interact with the simulator and update the surrogate model in “real-time”. In this context, they derive a new acquisition function, dubbed Latent Information Gain (LIG).", "review_text": "A good paper that could have been more convincing with better and richer experiments. ", "strengths": "Pros:\nThe paper uses a known and well-founded variant of the Neural Process as the surrogate model; they provide a delta of the model for better capturing spatiotemporal dependencies. The authors correctly resort to active learning to allow for the model to improve on-the-fly. They also correctly recognize the limitations of the venerable Expected Information Gain criterion used as acquisition function in Bayesian active learning. The suggested solution of computing the same criterion in the latent space (instead of the observation space) is reasonable. \n\nCons: The provided experimental comparisons are limited; I would like to see comparison to more deep learning alternatives. The experiments to now explain in depth all the trade-offs between complexity and accuracy, as well as in terms of how long active learning takes to reach performance. The experimental scenarios are limited as are the alternative surrogates they compare to.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper uses the Neural Process model as a surrogate trainable model of the input-output function in the context of simulation algorithms for epidemic modeling. This bypasses numerical integration that is required from non-machine learning simulation software. They modify the original model of the Neural Process to allow for better incorporating spatiotemporal dependencies. \n\nIn addition, instead of learning passively, the authors design active learning algorithms to interact with the simulator and update the surrogate model in “real-time”. In this context, they derive a new acquisition function, dubbed Latent Information Gain (LIG).", "strength_and_weaknesses": "Pros:\nThe paper uses a known and well-founded variant of the Neural Process as the surrogate model; they provide a delta of the model for better capturing spatiotemporal dependencies. The authors correctly resort to active learning to allow for the model to improve on-the-fly. They also correctly recognize the limitations of the venerable Expected Information Gain criterion used as acquisition function in Bayesian active learning. The suggested solution of computing the same criterion in the latent space (instead of the observation space) is reasonable. \n\nCons: The provided experimental comparisons are limited; I would like to see comparison to more deep learning alternatives. The experiments to now explain in depth all the trade-offs between complexity and accuracy, as well as in terms of how long active learning takes to reach performance. The experimental scenarios are limited as are the alternative surrogates they compare to.\n", "clarity,_quality,_novelty_and_reproducibility": "The writing is clear; some minor typos here and there, but good structure overall. The research is of certainly good quality. The novelly is mediocre, but carries some value. The experiments are reproducible.", "summary_of_the_review": "A good paper that could have been more convincing with better and richer experiments. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1665664778568}], "openreview_url": "https://openreview.net/forum?id=cVEOiBz2Em", "arxiv_id": "2106.02770", "paper_pdf": "papers/cVEOiBz2Em.pdf", "paper_pdf_sha256": "5c30d4664d0bb945718e3cfd0116d3e80cc1cea0c28146e009d57c0a33144037", "paper_pdf_bytes": 3432688, "paper_pdf_source": "openreview", "code_url": "https://github.com/Rose-STL-Lab/Interactive-Neural-Process", "code_repository": "Rose-STL-Lab/Interactive-Neural-Process", "code_commit": "962e4431364e8175253801dfdc28755131e5f14e", "code_archive": "repos/cVEOiBz2Em.zip", "code_archive_sha256": "dcf31bb6be9c0f8bdd46ba00ad4fbb95b794d24b9466c46cb619a357672ce426", "code_archive_bytes": 491101, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 433, "github_languages": {"Python": 303961}, "github_archived": false, "github_pushed_at": "2024-10-09T23:11:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/accelerating-stochastic-simulation-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bUAdXW8wN6", "year": 2022, "status": "rejected", "title": "Domain Invariant Adversarial Learning", "authors": ["Matan Levi", "Idan Attias", "Aryeh Kontorovich"], "authorids": ["~Matan_Levi1", "~Idan_Attias1", "~Aryeh_Kontorovich1"], "authors_source": "OpenReview API", "abstract": "The phenomenon of adversarial examples illustrates one of the most basic vulnerabilities of deep neural networks. Among the variety of techniques introduced to surmount this inherent weakness, adversarial training has emerged as the most effective strategy to achieve robustness. Typically, this is achieved by balancing robust and natural objectives. In this work, we aim to further reduce the trade-off between robust and standard accuracy by enforcing a domain-invariant feature representation. We present a new adversarial training method, Domain Invariant Adversarial Learning (DIAL), which learns a feature representation which is both robust and domain invariant. DIAL uses a variant of Domain Adversarial Neural Network (DANN) on the natural domain and its corresponding adversarial domain. In a case where the source domain consists of natural examples and the target domain is the adversarially perturbed examples, our method learns a feature representation constrained not to discriminate between the natural and adversarial examples, and can therefore achieve a more robust representation. Our experiments indicate that our method improves both robustness and standard accuracy, when compared to other state-of-the-art adversarial training methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "D6muCVpdaDB", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3948/Reviewer_iRX5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper describes an adversarial training approach that, in addition to the commonly used robustness loss, requires the network to extract similar representation distributions for clean and attacked data. The proposed method is inspired by domain adaptation approaches that require a model to extract domain invariant/agnostic features from two domains. In the context of this paper, the two domains are the clean and adversarially perturbed images, and the network is required to extract domain invariant representation. To achieve domain invariance, the authors propose a domain classifier ( i.e., an adversarial network) that discriminates the representations from clean and attacked images. The feature extractor is then required to generate features that fool the domain classifier. The authors then provide extensive experiments on small-scale benchmark datasets (SVHN, CIFAR10, CIFAR 100, and MNIST in the supplementary material ) to show the robustness of their proposed approach against the state-of-the-art robustness methods under white-box and black-box attacks. The authors show that their proposed method provides: 1) higher accuracy on attacked data (more robustness), and 2) higher accuracy on clean data, closing the gap between the performance on clean and attacked data. In addition, the paper provides insightful experiments on robustness to unforeseen adversaries, robustness to unforeseen corruptions, transfer learning, and ablation studies.\n\n", "review_text": "**Strengths:**\n\n* The idea is simple, yet it leads to significantly more robust networks\n* The paper is well written, and it is easy to follow\n* While the experiments are only carried out on smaller scale datasets, they are thorough, and they support the claims of the authors\n\n\n**Weaknesses:**\n\nI don't see major weaknesses in the paper. Below are some minor points.\n\n* DIAL-AWP comes out of the blue in Table 3. For the sake of consistency, I suggest adding it to Tables 1 and 2 as well and providing the formulation (for self-sufficiency).\n\n* The TSNE plots in Figure 5 for clean and perturbed distributions seem to have been calculated separately, which means that we are effectively looking at two different embedding spaces when we look at (a) and (b). I suggest that the author append the clean and perturbed representations, calculate the TSNE embedding jointly, and then plot them into their corresponding plots.\n\n**Additional Comments/Questions:**\n\n* In your KL robustness loss you have,\n $$ \\mathcal{L}_{rob}^{KL}=\\frac{1}{n}\\sum_i KL(G_f(x'_i;\\theta_f)||G_f(x_i;\\theta_f))$$\nMy understanding is that $G_f$\n is your feature extractor, and $G_f(x'_i), G_f(x_i)\\in \\mathbb{R}^d$  are not probability vectors, this is while $KL(\\cdot||\\cdot)$ is a dissimilarity measure defined only for probability distributions. Could you comment on this? Also, wouldn't a simple MSE work fine here?\n\n* This might be a matter of style, but it could be helpful to add equation numbers to your equations.\n\n* Typos:\n\n  * Page 3 second paragraph: \"belongs to the the family\"\n  * Page 5 second to the last paragraph: \"the initial learinnig\"\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper describes an adversarial training approach that, in addition to the commonly used robustness loss, requires the network to extract similar representation distributions for clean and attacked data. The proposed method is inspired by domain adaptation approaches that require a model to extract domain invariant/agnostic features from two domains. In the context of this paper, the two domains are the clean and adversarially perturbed images, and the network is required to extract domain invariant representation. To achieve domain invariance, the authors propose a domain classifier ( i.e., an adversarial network) that discriminates the representations from clean and attacked images. The feature extractor is then required to generate features that fool the domain classifier. The authors then provide extensive experiments on small-scale benchmark datasets (SVHN, CIFAR10, CIFAR 100, and MNIST in the supplementary material ) to show the robustness of their proposed approach against the state-of-the-art robustness methods under white-box and black-box attacks. The authors show that their proposed method provides: 1) higher accuracy on attacked data (more robustness), and 2) higher accuracy on clean data, closing the gap between the performance on clean and attacked data. In addition, the paper provides insightful experiments on robustness to unforeseen adversaries, robustness to unforeseen corruptions, transfer learning, and ablation studies.\n\n", "main_review": "**Strengths:**\n\n* The idea is simple, yet it leads to significantly more robust networks\n* The paper is well written, and it is easy to follow\n* While the experiments are only carried out on smaller scale datasets, they are thorough, and they support the claims of the authors\n\n\n**Weaknesses:**\n\nI don't see major weaknesses in the paper. Below are some minor points.\n\n* DIAL-AWP comes out of the blue in Table 3. For the sake of consistency, I suggest adding it to Tables 1 and 2 as well and providing the formulation (for self-sufficiency).\n\n* The TSNE plots in Figure 5 for clean and perturbed distributions seem to have been calculated separately, which means that we are effectively looking at two different embedding spaces when we look at (a) and (b). I suggest that the author append the clean and perturbed representations, calculate the TSNE embedding jointly, and then plot them into their corresponding plots.\n\n**Additional Comments/Questions:**\n\n* In your KL robustness loss you have,\n $$ \\mathcal{L}_{rob}^{KL}=\\frac{1}{n}\\sum_i KL(G_f(x'_i;\\theta_f)||G_f(x_i;\\theta_f))$$\nMy understanding is that $G_f$\n is your feature extractor, and $G_f(x'_i), G_f(x_i)\\in \\mathbb{R}^d$  are not probability vectors, this is while $KL(\\cdot||\\cdot)$ is a dissimilarity measure defined only for probability distributions. Could you comment on this? Also, wouldn't a simple MSE work fine here?\n\n* This might be a matter of style, but it could be helpful to add equation numbers to your equations.\n\n* Typos:\n\n  * Page 3 second paragraph: \"belongs to the the family\"\n  * Page 5 second to the last paragraph: \"the initial learinnig\"\n", "summary_of_the_review": "**Overall assessment:** The paper is well-written and easy to follow. While the main idea of learning domain invariant features is simple, its use in the context of robustness against adversarial attacks seems to lead to a significant performance boost. The experiments and, in particular, the ablation study section is insightful and aligned with the paper's claims. I think the paper is above average, and therefore I would like to vote for its acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635943969944}, {"id": "NHSw9GTLZc1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3948/Reviewer_8wyr"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, DANN is leveraged to generate domain invariant and robust feature representation. The authors claim that the proposed method outperforms other methods when the target domain is the adversarial examples.\n", "review_text": "+ The paper is easy to follow and the idea is straightforward.\n- The experiment section is not comprehensive. Only a few methods are included in the comparison. More recent SOTA methods are missing.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, DANN is leveraged to generate domain invariant and robust feature representation. The authors claim that the proposed method outperforms other methods when the target domain is the adversarial examples.\n", "main_review": "+ The paper is easy to follow and the idea is straightforward.\n- The experiment section is not comprehensive. Only a few methods are included in the comparison. More recent SOTA methods are missing.", "summary_of_the_review": "+ The paper is easy to follow and the idea is straightforward.\n- The experiment section is not comprehensive. Only a few methods are included in the comparison. More recent SOTA methods are missing.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635890295841}, {"id": "IetysIbv4sT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3948/Reviewer_FZTb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes DIAL to learn domain-invariant representations for clean and adversarial examples to improve model robustness and clean accuracy. The main idea is to treat the problem as a domain adaptation problem by considering the data shift between adversarial and clean distributions, and then use the generative adversarial network (GAN) principle to tackle this data shift.", "review_text": "Pros:\n\n(1) The paper is clearly written and easy to follow.\n\n(2) The motivation behind is very intuitive.\n\n(3) The paper conducts extensive experiments including multiple-$\\ell_p$-norm adversarial perturbations and unseen corruptions.\n\nCons:\n\n(1) My biggest concern is the novelty of this paper. Though showing promising performance, the idea of learning a feature extractor to minimize the distance between adversarial and clean distributions/domains has been widely studied and adopted before in the domain adaptation (DA) literature. In this paper, the author just simply introduced several DA loss terms and used the GAN framework to learn a more robust model. The experimental results are persuasive, however, the approach is too simple and not novel enough.\n\n(2) Some minor problems. I cannot find your paper in the autoattack leaderboard as you mentioned at the first line in Page 7.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes DIAL to learn domain-invariant representations for clean and adversarial examples to improve model robustness and clean accuracy. The main idea is to treat the problem as a domain adaptation problem by considering the data shift between adversarial and clean distributions, and then use the generative adversarial network (GAN) principle to tackle this data shift.", "main_review": "Pros:\n\n(1) The paper is clearly written and easy to follow.\n\n(2) The motivation behind is very intuitive.\n\n(3) The paper conducts extensive experiments including multiple-$\\ell_p$-norm adversarial perturbations and unseen corruptions.\n\nCons:\n\n(1) My biggest concern is the novelty of this paper. Though showing promising performance, the idea of learning a feature extractor to minimize the distance between adversarial and clean distributions/domains has been widely studied and adopted before in the domain adaptation (DA) literature. In this paper, the author just simply introduced several DA loss terms and used the GAN framework to learn a more robust model. The experimental results are persuasive, however, the approach is too simple and not novel enough.\n\n(2) Some minor problems. I cannot find your paper in the autoattack leaderboard as you mentioned at the first line in Page 7.", "summary_of_the_review": "I think the paper conducts extensive experiments to demonstrate the effectiveness of the proposed method, including some interesting ones, e.g., robustness against unseen perturbations, transfer learning (I like them). However, the novelty of this paper is insufficient, and using the domain adaptation principle and learning invariant representation has been widely studied. Therefore, I vote for rejection.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635685949510}, {"id": "6KVt-ZCFPtF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3948/Reviewer_FjGF"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a domain invariant adversarial training (DIAL) method, which learns the feature representation that is both robust and domain invariant.  Apart from the label classifier, the model is equipped with a domain classifier that constrains the model not to discriminate between natural examples and adversarial examples, thus achieving a more robust feature representation. Extensive experiments on image classification benchmark the robustness compared to other state-of-the-art methods. ", "review_text": "This paper proposes a simple and effective adversarial learning method DIAL, which brings the idea from domain adaptation for robust representation. \n\nStrengths: \n1. This paper is well-written and easy to follow.\n2. It conducts various experiments to demonstrate the effectiveness of the proposed method ranging from robustness to white-box attacks, black-box attacks, unforeseen adversaries, unforeseen corruptions and transfer learning. The experimental results are solid and technically sound. \n\nWeaknesses: \n1. From my point of view, the novelty of the methodology is not enough, as the domain classifier and the gradient reversal layer are the same with those methods in domain adaptation such as [1].\n2. To better understanding the reversal-ratio hyper-parameter $r$, can the authors provides the robustness under different values of $r$.\n\n[1] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pp. 1180–1189. PMLR, 2015. \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a domain invariant adversarial training (DIAL) method, which learns the feature representation that is both robust and domain invariant.  Apart from the label classifier, the model is equipped with a domain classifier that constrains the model not to discriminate between natural examples and adversarial examples, thus achieving a more robust feature representation. Extensive experiments on image classification benchmark the robustness compared to other state-of-the-art methods. ", "main_review": "This paper proposes a simple and effective adversarial learning method DIAL, which brings the idea from domain adaptation for robust representation. \n\nStrengths: \n1. This paper is well-written and easy to follow.\n2. It conducts various experiments to demonstrate the effectiveness of the proposed method ranging from robustness to white-box attacks, black-box attacks, unforeseen adversaries, unforeseen corruptions and transfer learning. The experimental results are solid and technically sound. \n\nWeaknesses: \n1. From my point of view, the novelty of the methodology is not enough, as the domain classifier and the gradient reversal layer are the same with those methods in domain adaptation such as [1].\n2. To better understanding the reversal-ratio hyper-parameter $r$, can the authors provides the robustness under different values of $r$.\n\n[1] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pp. 1180–1189. PMLR, 2015. \n\n\n", "summary_of_the_review": "Overall, this paper proposes a simple and effective adversarial learning method DIAL for robust representation learning. Extensive experiments are conducted to demonstrate the effectiveness of the proposed method and provide solid results. However, the novelty of the paper is not significant as similar methodology exists in domain adaptation. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635561906893}], "openreview_url": "https://openreview.net/forum?id=bUAdXW8wN6", "arxiv_id": "2104.00322", "paper_pdf": "papers/bUAdXW8wN6.pdf", "paper_pdf_sha256": "b146067d64014f407811f8ee0b8f00b940dc6fa6785c829c3bbb360342cd4eab", "paper_pdf_bytes": 4161643, "paper_pdf_source": "openreview", "code_url": "https://github.com/matanle51/DIAL", "code_repository": "matanle51/DIAL", "code_commit": "a94066d530cbd39c02957af0b6924d5e01ac66e6", "code_archive": "repos/bUAdXW8wN6.zip", "code_archive_sha256": "b6f34adab7914bb26df2058558a9bc93be285452478d9f1817b7c7ef531d84d9", "code_archive_bytes": 2111918, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 1764, "github_languages": {"Python": 83326}, "github_archived": false, "github_pushed_at": "2022-09-11T10:33:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/domain-invariant-adversarial-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1toB0Fo9CZy", "year": 2021, "status": "rejected", "title": "Neural Architecture Search of SPD Manifold Networks", "authors": ["Rhea Sanjay Sukthanker", "Zhiwu Huang", "Suryansh Kumar", "Erik Goron", "Yan Wu", "Luc Van Gool"], "authorids": ["~Rhea_Sanjay_Sukthanker1", "~Zhiwu_Huang1", "~Suryansh_Kumar1", "~Erik_Goron1", "~Yan_Wu4", "~Luc_Van_Gool1"], "authors_source": "OpenReview API", "abstract": "In this paper, we propose a new neural architecture search (NAS) problem of Symmetric Positive Definite (SPD) manifold networks. Unlike the conventional NAS problem, our problem requires to search for a unique computational cell called the SPD cell. This SPD cell serves as a basic building block of SPD neural architectures. An efficient solution to our problem is important to minimize the extraneous manual effort in the SPD neural architecture design. To accomplish this goal, we first introduce a geometrically rich and diverse SPD neural architecture search space for an efficient SPD cell design. Further, we model our new NAS problem using the supernet strategy, which models the architecture search problem as a one-shot training process of a single supernet. Based on the supernet modeling, we exploit a differentiable NAS algorithm on our relaxed continuous search space for SPD neural architecture search. Statistical evaluation of our method on drone, action, and emotion recognition tasks mostly provides better results than the state-of-the-art SPD networks and NAS algorithms. Empirical results show that our algorithm excels in discovering better SPD network design and providing models that are more than 3 times lighter than searched by state-of-the-art NAS algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "HsQLJQq6FtC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper415/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper considers a generalization of convolutional neural networks (CNNs) to manifold-valued data such as Symmetric Positive Definite (SPD). This paper proposes a neural architecture search problem of SPD manifold networks. A SPD cell representation and corresponding candidate operation search space is introduced. They demonstrate on drone, action and emotion recognition datasets that their method is performed well compared to SPD approaches.\n\n- It is not clear how the computational graph preserves the geometric structure of SPD manifold.\n- It is unclear how to obtain the weights of Frechet Mean w_i of wFM (eq. 3) by backpropagation. \n- Relatively thorough experimental valuations: using 3 datasets comparing with sufficient number of prior\napproaches. However, the comparisons of the experiments results are limited to the SPD methods.  It is hard to understand how the generalization of convolutional neural networks (CNNs) to Symmetric Positive Definite (SPD) presented in this paper helps to improve CNN.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, however, more explanations and comparisons needed. ", "review": "The paper considers a generalization of convolutional neural networks (CNNs) to manifold-valued data such as Symmetric Positive Definite (SPD). This paper proposes a neural architecture search problem of SPD manifold networks. A SPD cell representation and corresponding candidate operation search space is introduced. They demonstrate on drone, action and emotion recognition datasets that their method is performed well compared to SPD approaches.\n\n- It is not clear how the computational graph preserves the geometric structure of SPD manifold.\n- It is unclear how to obtain the weights of Frechet Mean w_i of wFM (eq. 3) by backpropagation. \n- Relatively thorough experimental valuations: using 3 datasets comparing with sufficient number of prior\napproaches. However, the comparisons of the experiments results are limited to the SPD methods.  It is hard to understand how the generalization of convolutional neural networks (CNNs) to Symmetric Positive Definite (SPD) presented in this paper helps to improve CNN.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604080271876}, {"id": "cf2wYGadh6u", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper415/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. \"In recent years, plenty of research work has been published in the area of NAS.\" needs citations.\n2. What is \"best of both these fields (NAS, SPDNets) \" in page 2?\n3. The definition of SPD matrix is wrong, this definition is for positive-definite matrix, not for SPD.\n4. In Eq. (1), log should be matrix logarithm, and need to be mentioned.\n5. In page 3, \"There are other efficient methods to compute distance between two points on the SPD manifold\" needs to be backed by references. \n6. \" Other property of the Riemannian manifoldof our interest is local diffeomorphism of geodesics which is a one-to-one mapping from the pointon the tangent space of the manifold to the manifold\", this is not true, geodesic is a function of two points (start and end) and time point (say between 0 and 1), so it can not be a one-one mapping between tngent space to manifold. Authors here meant to say exponential map as defined in Eq. (2). \n7. X in Eq. (2) is not the \"reference point\", it should be the base point for exponential map (see Chavel's book).\n8. The inverse of exp map is not defined everywhere, as exp is LOCAL diffeo., please clarify.\n9. Before Eq. (3), sometime->sometimes\n10. \"This definition is trivially extended tocompute the weighted Riemannian Barycenter\" not true, the authors should mention existence and uniqueness of wFM, for example, even if FM exists, wFM may not for some choice of w.\n11. What is \"valid SPD\" in Bimap layer, contrary to \"invalid SPD\"!\n12. what is P in Eq. (4)? I think it is parallel transport but needs to be defined.\n13. \\epsilon in ReEig layer should be >0.\n14. In logEig, essentially you are mapping SPD to space of symmetric matrices (which is \"flat\"), then why not use Inverse Exp. map? what is the necessity to define logEig map additionally?\n15. Same goes for expEig map, what is need given that one can use Exp. map!\n16. In weighted pooling layer, why karcher flow is \"simple\" and recursive method is not? Please clarify.\n17. \" such that the mixture of all operations still reside on SPD manifolds\", here the term mixture is ambiguous!\n18. Given that the operations are already defined, definition of SPD cell is NOT novel!\n19. The additional search space operations are all trivial extensions of basic operations, which are not the contribution, hence I think the authors need to tone down \"Most of these operations are not fully explored for SPD networks\" \n20. In 3.2, what the authors meant by \"optimize the over parameterized supernet\"?\n21. In Algorithm 1, do the authors mean Riemannian gradient descent (Absil et al.)?\n22. What the authors meant by \"Note that the gradient based optimizationforwmust follow the geometry of SPD manifold to update the structured connection weight, and itscorresponding SPD matrix data\"? Do the authors mean they need to project the Euclidean gradient found in Eq. (8) to get tangent vectors?\n23. The experimental setup is weak, e.g., SPDNet and ManifoldNet need to be compared with same model complexity, otherwise the comparison is not fair! ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Incremental work using already defined tools", "review": "1. \"In recent years, plenty of research work has been published in the area of NAS.\" needs citations.\n2. What is \"best of both these fields (NAS, SPDNets) \" in page 2?\n3. The definition of SPD matrix is wrong, this definition is for positive-definite matrix, not for SPD.\n4. In Eq. (1), log should be matrix logarithm, and need to be mentioned.\n5. In page 3, \"There are other efficient methods to compute distance between two points on the SPD manifold\" needs to be backed by references. \n6. \" Other property of the Riemannian manifoldof our interest is local diffeomorphism of geodesics which is a one-to-one mapping from the pointon the tangent space of the manifold to the manifold\", this is not true, geodesic is a function of two points (start and end) and time point (say between 0 and 1), so it can not be a one-one mapping between tngent space to manifold. Authors here meant to say exponential map as defined in Eq. (2). \n7. X in Eq. (2) is not the \"reference point\", it should be the base point for exponential map (see Chavel's book).\n8. The inverse of exp map is not defined everywhere, as exp is LOCAL diffeo., please clarify.\n9. Before Eq. (3), sometime->sometimes\n10. \"This definition is trivially extended tocompute the weighted Riemannian Barycenter\" not true, the authors should mention existence and uniqueness of wFM, for example, even if FM exists, wFM may not for some choice of w.\n11. What is \"valid SPD\" in Bimap layer, contrary to \"invalid SPD\"!\n12. what is P in Eq. (4)? I think it is parallel transport but needs to be defined.\n13. \\epsilon in ReEig layer should be >0.\n14. In logEig, essentially you are mapping SPD to space of symmetric matrices (which is \"flat\"), then why not use Inverse Exp. map? what is the necessity to define logEig map additionally?\n15. Same goes for expEig map, what is need given that one can use Exp. map!\n16. In weighted pooling layer, why karcher flow is \"simple\" and recursive method is not? Please clarify.\n17. \" such that the mixture of all operations still reside on SPD manifolds\", here the term mixture is ambiguous!\n18. Given that the operations are already defined, definition of SPD cell is NOT novel!\n19. The additional search space operations are all trivial extensions of basic operations, which are not the contribution, hence I think the authors need to tone down \"Most of these operations are not fully explored for SPD networks\" \n20. In 3.2, what the authors meant by \"optimize the over parameterized supernet\"?\n21. In Algorithm 1, do the authors mean Riemannian gradient descent (Absil et al.)?\n22. What the authors meant by \"Note that the gradient based optimizationforwmust follow the geometry of SPD manifold to update the structured connection weight, and itscorresponding SPD matrix data\"? Do the authors mean they need to project the Euclidean gradient found in Eq. (8) to get tangent vectors?\n23. The experimental setup is weak, e.g., SPDNet and ManifoldNet need to be compared with same model complexity, otherwise the comparison is not fair! ", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603950919890}, {"id": "4-ezZrNk7OR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper415/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The work focus on finding better SPD Manifold Networks from the view of neural architecture search. They model the network parameterized by a set of architecture parameter and take the problem as a bi-level optimization problem, which is similar with DARTS way. I think the most contribution of this paper is the attempt to apply NAS on a new domain.\n\n1.  The authors introduce a lot of concepts about SPD Network. But they are not meaningful in this paper since the authors just consider operations as the node. From the graph view, SPD Network can be seen as a neural architecture directly even though the operation type is different.\n\n2. My main concern is novelty problem.  The authors also treat SPD Network as a graph in Figure 1 such that the DARTS method can be applied directly. As for Algorithm 1, they also use two types parameters: architecture weights $\\alpha$  and operation weight $w$. Then bi-level optimization problem is solved using the same method with DARTS.\n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The authors apply neural architecture search on SPD Manifold Networks, which is the first attempt in this field.", "review": "The work focus on finding better SPD Manifold Networks from the view of neural architecture search. They model the network parameterized by a set of architecture parameter and take the problem as a bi-level optimization problem, which is similar with DARTS way. I think the most contribution of this paper is the attempt to apply NAS on a new domain.\n\n1.  The authors introduce a lot of concepts about SPD Network. But they are not meaningful in this paper since the authors just consider operations as the node. From the graph view, SPD Network can be seen as a neural architecture directly even though the operation type is different.\n\n2. My main concern is novelty problem.  The authors also treat SPD Network as a graph in Figure 1 such that the DARTS method can be applied directly. As for Algorithm 1, they also use two types parameters: architecture weights $\\alpha$  and operation weight $w$. Then bi-level optimization problem is solved using the same method with DARTS.\n\n ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603892180279}, {"id": "yZ2Pz4yTPAR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper415/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a neural architecture search method for SPD inputs, which extends DARTs from Euclidean space to SPD manifold by defining several SPD cells. Each SPD cell consists of some operations in search space of SPDNets and DARTs. The experiments are conducted on three datasets.\n\nStrengths:\n+: The idea of neural architecture search for SPD inputs seems interesting.\n+: The proposed SPDNetNAS shows superiority to hand-crafted SPDNets/ManifoldNet and Euclidean NAS (i.e., DARTS and FairDARTS).\n+: The paper is clear written.\n\nWeaknesses:\n-: The authors would better carefully reconsider motivation and main contribution of this paper. The current work looks like combination of SPDNets and DARTs (with some minor modifications). Then, the first attempt for NAS problem of SPD manifold networks is not an appropriate motivation for such a combination. The authors should clarify why it is necessary to search neural architecture for SPD inputs. In another words, what are drawbacks of traditional SPDNets? \n\nFor Euclidean NAS, the searched architectures aim to have lower model complexity than hand-crafted architectures, while having comparable or higher accuracies. The proposed SPDNetNAS achieves higher accuracies than hand-crafted counterparts, but has 10x times of parameters number (as shown in Table 2). Therefore, it is hard to say that performance gains come from neural architecture search or parameter increasing. Meanwhile, it is not fair for comparison of SPDNetNAS with Euclidean NAS in terms of parameters number or comparison of SPDNetNAS with SPDNets only in terms of performance. \n\n-: The authors mainly compare with hand-crafted SPDNet and Euclidean NAS. [r1] presents a recurrent model for SPD inputs, which can handle sequential (or temporal) data, e.g., action and emotion recognition. Therefore, what are merits of the proposed SPDNetNAS over [r1] for handling sequential (or temporal) data, except architecture design vs architecture search?\n[r1] A Statistical Recurrent Model on the Manifold of Symmetric Positive Definite Matrices. NIPS, 2018.\n\n-: I wonder that why a section of FURTHER STUDY is located in appendix? Could it be integrated into section of Conclusion?\n\n-: It is well known that two-level optimization algorithm for NAS will lead optimization gap, e.g., overfitting or model collapsing. What are strategies used to solve this issue, e.g., early stopping? If source code is not released or no more details are given, the results will be very hard to reproduce.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for Neural Architecture Search of SPD Manifold Networks", "review": "This paper introduces a neural architecture search method for SPD inputs, which extends DARTs from Euclidean space to SPD manifold by defining several SPD cells. Each SPD cell consists of some operations in search space of SPDNets and DARTs. The experiments are conducted on three datasets.\n\nStrengths:\n+: The idea of neural architecture search for SPD inputs seems interesting.\n+: The proposed SPDNetNAS shows superiority to hand-crafted SPDNets/ManifoldNet and Euclidean NAS (i.e., DARTS and FairDARTS).\n+: The paper is clear written.\n\nWeaknesses:\n-: The authors would better carefully reconsider motivation and main contribution of this paper. The current work looks like combination of SPDNets and DARTs (with some minor modifications). Then, the first attempt for NAS problem of SPD manifold networks is not an appropriate motivation for such a combination. The authors should clarify why it is necessary to search neural architecture for SPD inputs. In another words, what are drawbacks of traditional SPDNets? \n\nFor Euclidean NAS, the searched architectures aim to have lower model complexity than hand-crafted architectures, while having comparable or higher accuracies. The proposed SPDNetNAS achieves higher accuracies than hand-crafted counterparts, but has 10x times of parameters number (as shown in Table 2). Therefore, it is hard to say that performance gains come from neural architecture search or parameter increasing. Meanwhile, it is not fair for comparison of SPDNetNAS with Euclidean NAS in terms of parameters number or comparison of SPDNetNAS with SPDNets only in terms of performance. \n\n-: The authors mainly compare with hand-crafted SPDNet and Euclidean NAS. [r1] presents a recurrent model for SPD inputs, which can handle sequential (or temporal) data, e.g., action and emotion recognition. Therefore, what are merits of the proposed SPDNetNAS over [r1] for handling sequential (or temporal) data, except architecture design vs architecture search?\n[r1] A Statistical Recurrent Model on the Manifold of Symmetric Positive Definite Matrices. NIPS, 2018.\n\n-: I wonder that why a section of FURTHER STUDY is located in appendix? Could it be integrated into section of Conclusion?\n\n-: It is well known that two-level optimization algorithm for NAS will lead optimization gap, e.g., overfitting or model collapsing. What are strategies used to solve this issue, e.g., early stopping? If source code is not released or no more details are given, the results will be very hard to reproduce.  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603769727003}], "openreview_url": "https://openreview.net/forum?id=1toB0Fo9CZy", "arxiv_id": "2010.14535", "paper_pdf": "papers/1toB0Fo9CZy.pdf", "paper_pdf_sha256": "cc57aebd2f196ffa35c84fd367b468b1a0335bf6f8b83689463eccf5164e7917", "paper_pdf_bytes": 1645915, "paper_pdf_source": "openreview", "code_url": "https://github.com/rheasukthanker/SPDNetNAS", "code_repository": "rheasukthanker/SPDNetNAS", "code_commit": "0f82e735f03603c0d355e2d8bf8fc570ac225ddb", "code_archive": "repos/1toB0Fo9CZy.zip", "code_archive_sha256": "90ce892caee48bd59c5f53f622d255b2f35eacd133e5a5dd9e92d17877288e82", "code_archive_bytes": 891545, "code_file_count": 39, "code_extensions": {".py": 39}, "github_disk_usage_kb": 1206, "github_languages": {"Python": 337356}, "github_archived": false, "github_pushed_at": "2022-02-18T01:32:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-architecture-search-of-spd-manifold-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "r1gzoaNtvr", "year": 2020, "status": "rejected", "title": "Emergence of Compositional Language with Deep Generational Transmission", "authors": ["Michael Cogswell", "Jiasen Lu", "Stefan Lee", "Devi Parikh", "Dhruv Batra"], "authorids": ["cogswell@gatech.edu", "jiasenlu@gatech.edu", "steflee@gatech.edu", "parikh@gatech.edu", "dbatra@gatech.edu"], "authors_source": "OpenReview API", "abstract": "Recent work has studied the emergence of language among deep reinforcement learning agents that must collaborate to solve a task. Of particular interest are the factors that cause language to be compositional---i.e., express meaning by combining words which themselves have meaning. Evolutionary linguists have found that in addition to structural priors like those already studied in deep learning, the dynamics of transmitting language from generation to generation contribute significantly to the emergence of  compositionality. In this paper, we introduce these cultural evolutionary dynamics into language emergence by periodically replacing agents in a population to create a knowledge gap, implicitly inducing cultural transmission of language. We show that this implicit cultural transmission encourages the resulting languages to exhibit better compositional generalization.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SkgQjoTV9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper736/AnonReviewer4"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a simple extension to the training of emergent communication protocols in multi-agent settings. \nThe central hypothesis is that learnability will favor more 'compressible' and therefor more compositional languages to emerge. \n\nThis hypothesis is tested by training a population of listens and speakers in an emergent communication task and comparing a number of different strategies for reinitializing agents in the population. Since the new agents start from a random initialization, they provide a learning signal that reinforces protocols which can be learned quickly. \n\nExperiments:\nWhile the experimental results largely confirm the hypothesis there are a few issues:\n-all of the plots show mean and standard deviations, rather than error of the mean. This makes it difficult to understand which differences are statistically significant and which ones are not. \n- The replacement strategy 'epsilon-greedy' replaces agents based on their validation loss, which seems like an unfair advantage. \n- It is currently unclear how much we can learn from Section 5.2: Naturally, agents trained in the same population will develop more similar protocols than those trained independently. This result is obvious and it is unclear whether reinitializing the agents makes any significant difference to the similarity. Again, confidence intervals would help.  \n- The experiments are also extremely toy. I would be more convinced if the authors tested their method on a more challenging task, though I am aware this is a common problem in this field. \n\nNovelty:\nThe single biggest issues with the current form of the paper is the related work section. In this section two papers [1,2] are mentioned as \"concurrent\" when at least one of them [1] has been available online since June 2019. I think it is important to clearly point out the novelty of the current work compared to those two previous papers. \nIn particular [1] seems to be extremely close to the ideas and methods proposed here. Saying that these papers \"confirm the hypothesis\" simply is not enough. \n\nReferences:\n[1]: \"Ease-of-Teaching and Language Structure from Emergent Communication\", Funshan Li et al\n[2]: \"Co-evolution of language and agents in referential games\", Gautier Dagan et al \n\n[Updated score based on the rebuttal]\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "N/A", "title": "Official Blind Review #4", "review": "This paper proposes a simple extension to the training of emergent communication protocols in multi-agent settings. \nThe central hypothesis is that learnability will favor more 'compressible' and therefor more compositional languages to emerge. \n\nThis hypothesis is tested by training a population of listens and speakers in an emergent communication task and comparing a number of different strategies for reinitializing agents in the population. Since the new agents start from a random initialization, they provide a learning signal that reinforces protocols which can be learned quickly. \n\nExperiments:\nWhile the experimental results largely confirm the hypothesis there are a few issues:\n-all of the plots show mean and standard deviations, rather than error of the mean. This makes it difficult to understand which differences are statistically significant and which ones are not. \n- The replacement strategy 'epsilon-greedy' replaces agents based on their validation loss, which seems like an unfair advantage. \n- It is currently unclear how much we can learn from Section 5.2: Naturally, agents trained in the same population will develop more similar protocols than those trained independently. This result is obvious and it is unclear whether reinitializing the agents makes any significant difference to the similarity. Again, confidence intervals would help.  \n- The experiments are also extremely toy. I would be more convinced if the authors tested their method on a more challenging task, though I am aware this is a common problem in this field. \n\nNovelty:\nThe single biggest issues with the current form of the paper is the related work section. In this section two papers [1,2] are mentioned as \"concurrent\" when at least one of them [1] has been available online since June 2019. I think it is important to clearly point out the novelty of the current work compared to those two previous papers. \nIn particular [1] seems to be extremely close to the ideas and methods proposed here. Saying that these papers \"confirm the hypothesis\" simply is not enough. \n\nReferences:\n[1]: \"Ease-of-Teaching and Language Structure from Emergent Communication\", Funshan Li et al\n[2]: \"Co-evolution of language and agents in referential games\", Gautier Dagan et al \n\n[Updated score based on the rebuttal]\n", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1572293531458}, {"id": "rkx4Qvul9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper736/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies whether language composition may emerge by partially re-sampling new agents inside a pool of language agents. They set up a consistent experimental setting for assessing compositionality,  assess different agent architectures, e.g., memory vs. memoryless agents,  and explore how the language remains close to each other by re-sampling new agents.\n\nThe paper is well-motivated with substantial background literature on the cognitive science and emergent communication side. The claim is clear, the hypotheses are well-stated, and the experiments look solid (I particularly appreciated the paragraph on shortcoming evaluation). In the end, I enjoy reading the paper despite its density, and I could see that the authors made quite some effort in that direction.\n\nImprovement direction, questions:\n - The authors made were careful not to take ownership of Kottur et al. 's works. Yet, the writing sometimes gives the feeling that their work is solely an extension of Kottur's work, which is not the case. It also gives the feeling that Kottur et al. is the only valid experimental setting, which is not the case (at the authors pointed out in the related work section). Thus, I would recommend to summarise at some point the similarity/difference between the two papers, or at least stop referring the paper every two lines! \n - Replacement strategy: the authors use simple replacement strategies, and conclude that it has little impact. Althought It sounds reasonable in the current setting, the conclusion may be a bit premature. I would recommend to discuss further this result with complementary experiments could be the following: see the impact of epsilon, trying tournament strategies, why 8 populations (this sounds a bit arbitrary too). I would also like to put those observations in perspective with the evolutionary literature [1], and even provide a full paragraph in the related work section.\n - Population dynamics: I am missing a key element in the paper: an analysis of the population dynamics. Although the paper deals with generational transmissions, there are no experiments that analyze the evolution of language generations after generations. Most of the experiments deal with the final convergence state. Again, I would recommend having a look at the evolutionary literature to see which protocol they use to analyze such behavior.\n - Literature side: The authors did an excellent job on the emergent communication and cognitive science side. I think that it is worth extending the comparison further. For instance: \n    * generational transmission can be studied in the light of game theory [2] where compositionality can be seen as a Nash Equilibrium between agent. \n    * generational transmission is a form of dynamic distillation [3]\n    * and evolutionary algorithms!\n - I had some difficulties in understanding Fig4, and the final-take away correctly. Would it be possible to give me one or two examples to correctly parse the table? More generally, I would recommend to add a few lines with some concrete and cherry-picked examples from the experiments to help the reader to have more intuition). \n - In a similar spirit, it is hard to interpret the distance in Figure 3. What would correspond to an increase of 1pt of distance? Having said that, the experiment is sound, and it is insightful.\n - reproducible: having a final table in the array in the appendix could be very helpful \n - crazy experiment: even if I am also a DRL addict, I would be curious to train one of the models with evolutionary algorithms (CMA-ES over parameters, for instance) to assess whether RL has an impact on compositionality (or it is solely the experimental protocol that matters). \n - I may have missed this point, but how many seeds did you use to run your experiments? \n - I may have also missed this point, what is the average length of the dialogue. Can you upload (non-understandable) dialogue example? \n\nLast point... but it does not undermine the soundness of the experimental protocol! \n - In the end, Is 25 accuracy points really compositionality? What would be the score of simple strategies with overcomplete tokens? What is the score of the minimal vocab if we are only correct with one modality, two modalities?\n\n\n\nRemarks:\n - in the introduction, you mention that previous old agents have grounded language, I am not sure whether we can speak of grounded language here, they have a predefined language, but it is not grounded. \n - Please remove the bold sentence in the introduction :) The claim is clear!\n - P11: Alg undefined\n - P12: the legend cannot be read\n \n\nConclusion\nI am familiar with this type of experimental protocols, and I am well aware that they are never-ending works. There are always more experiments to do, more parameters to analyze. The final question is the following: is this paper have enough of these never-ending experiments? I think that this paper is just above this threshold by a short margin, and I vouch for weak accept.\n\nHowever, I am missing at least one dynamic figure (to see the impact of the population along time, which is one of the core concepts of the paper), and there are several links with other ML communities that still have to be highlighted (especially evolutionary algorithms). \nBesides, I somehow feel that the authors pursue two different goals in this paper: they both analyze memory/memoryless complete/overcomplete agents, which is somehow orthogonal to the general transmission hypothesis. Maybe, It would have made more sense to focus on one (or two) of the models and change the experimental setting on them (population size, training time, etc.) \n\nIn the end, I would favor a weak accept. \nI am open to discussion regarding this scoring.\n\n[1] Bäck, Thomas, and Frank Hoffmeister. \"Extended selection mechanisms in genetic algorithms.\" (1991).\n[2] Lanctot, Marc, et al. \"A unified game-theoretic approach to multiagent reinforcement learning.\" Advances in Neural Information Processing Systems. 2017.\n[3] Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. \"Distilling the knowledge in a neural network.\" arXiv preprint arXiv:1503.02531 (2015).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "6: Weak Accept", "experience_assessment": "I have published one or two papers in this area.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "This paper studies whether language composition may emerge by partially re-sampling new agents inside a pool of language agents. They set up a consistent experimental setting for assessing compositionality,  assess different agent architectures, e.g., memory vs. memoryless agents,  and explore how the language remains close to each other by re-sampling new agents.\n\nThe paper is well-motivated with substantial background literature on the cognitive science and emergent communication side. The claim is clear, the hypotheses are well-stated, and the experiments look solid (I particularly appreciated the paragraph on shortcoming evaluation). In the end, I enjoy reading the paper despite its density, and I could see that the authors made quite some effort in that direction.\n\nImprovement direction, questions:\n - The authors made were careful not to take ownership of Kottur et al. 's works. Yet, the writing sometimes gives the feeling that their work is solely an extension of Kottur's work, which is not the case. It also gives the feeling that Kottur et al. is the only valid experimental setting, which is not the case (at the authors pointed out in the related work section). Thus, I would recommend to summarise at some point the similarity/difference between the two papers, or at least stop referring the paper every two lines! \n - Replacement strategy: the authors use simple replacement strategies, and conclude that it has little impact. Althought It sounds reasonable in the current setting, the conclusion may be a bit premature. I would recommend to discuss further this result with complementary experiments could be the following: see the impact of epsilon, trying tournament strategies, why 8 populations (this sounds a bit arbitrary too). I would also like to put those observations in perspective with the evolutionary literature [1], and even provide a full paragraph in the related work section.\n - Population dynamics: I am missing a key element in the paper: an analysis of the population dynamics. Although the paper deals with generational transmissions, there are no experiments that analyze the evolution of language generations after generations. Most of the experiments deal with the final convergence state. Again, I would recommend having a look at the evolutionary literature to see which protocol they use to analyze such behavior.\n - Literature side: The authors did an excellent job on the emergent communication and cognitive science side. I think that it is worth extending the comparison further. For instance: \n    * generational transmission can be studied in the light of game theory [2] where compositionality can be seen as a Nash Equilibrium between agent. \n    * generational transmission is a form of dynamic distillation [3]\n    * and evolutionary algorithms!\n - I had some difficulties in understanding Fig4, and the final-take away correctly. Would it be possible to give me one or two examples to correctly parse the table? More generally, I would recommend to add a few lines with some concrete and cherry-picked examples from the experiments to help the reader to have more intuition). \n - In a similar spirit, it is hard to interpret the distance in Figure 3. What would correspond to an increase of 1pt of distance? Having said that, the experiment is sound, and it is insightful.\n - reproducible: having a final table in the array in the appendix could be very helpful \n - crazy experiment: even if I am also a DRL addict, I would be curious to train one of the models with evolutionary algorithms (CMA-ES over parameters, for instance) to assess whether RL has an impact on compositionality (or it is solely the experimental protocol that matters). \n - I may have missed this point, but how many seeds did you use to run your experiments? \n - I may have also missed this point, what is the average length of the dialogue. Can you upload (non-understandable) dialogue example? \n\nLast point... but it does not undermine the soundness of the experimental protocol! \n - In the end, Is 25 accuracy points really compositionality? What would be the score of simple strategies with overcomplete tokens? What is the score of the minimal vocab if we are only correct with one modality, two modalities?\n\n\n\nRemarks:\n - in the introduction, you mention that previous old agents have grounded language, I am not sure whether we can speak of grounded language here, they have a predefined language, but it is not grounded. \n - Please remove the bold sentence in the introduction :) The claim is clear!\n - P11: Alg undefined\n - P12: the legend cannot be read\n \n\nConclusion\nI am familiar with this type of experimental protocols, and I am well aware that they are never-ending works. There are always more experiments to do, more parameters to analyze. The final question is the following: is this paper have enough of these never-ending experiments? I think that this paper is just above this threshold by a short margin, and I vouch for weak accept.\n\nHowever, I am missing at least one dynamic figure (to see the impact of the population along time, which is one of the core concepts of the paper), and there are several links with other ML communities that still have to be highlighted (especially evolutionary algorithms). \nBesides, I somehow feel that the authors pursue two different goals in this paper: they both analyze memory/memoryless complete/overcomplete agents, which is somehow orthogonal to the general transmission hypothesis. Maybe, It would have made more sense to focus on one (or two) of the models and change the experimental setting on them (population size, training time, etc.) \n\nIn the end, I would favor a weak accept. \nI am open to discussion regarding this scoring.\n\n[1] Bäck, Thomas, and Frank Hoffmeister. \"Extended selection mechanisms in genetic algorithms.\" (1991).\n[2] Lanctot, Marc, et al. \"A unified game-theoretic approach to multiagent reinforcement learning.\" Advances in Neural Information Processing Systems. 2017.\n[3] Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. \"Distilling the knowledge in a neural network.\" arXiv preprint arXiv:1503.02531 (2015)."}, "tcdate": 1572009755888}, {"id": "BJeUQfre5B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper736/AnonReviewer1"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper continues the line of work on emergent compositionality in dialogs, and here it is extended to handle groups of interacting agents that pass language in an evolutionary way from one generation to another. The key idea is that with group of interacting agents, if some agents are replaced with new ones, then the newbies would learn the same language as the group. \n\nGeneral assessment: \n\nThe setup of the paper is is interesting, and the paper covers a lot of ground. The paper makes bold claims like \"cultural transmission induces compositionality\" and \"Variations in replacement strategy tend to not affect performance\". Unfortunately, the paper does not systematic provide experimental evidence to adequately support its claims.\n\nIn my experience, the emergence of compositionality (or lack of) in the setup learned here is very sensitive to various aspects of the learning setup, hence are hard to reproduce. Specifically, they the task-and-talk paper by Kottur et al may yield different conclusions, if parameters of the original experiments are modified, even slightly. The current paper does improve the evaluation protocol of Kottur 2017 by reporting variance over runs, but does not explore the parameter space more systematically, hence I am concerned that it may suffer from similar fragility. \n\nIn this light, significantly more evidence should be provided to convince that the experimental results are stable and can be reproduced.  First, one needs to show the experiments repeated over the full range of setup parameters. Including, the vocabulary sizes V_Q and V_A, the number of tasks, the number of attributes per task, and number of agents etc. Similarly, the current paper introduces a new compositional split, generated in one specific way. The effect of the split on what aspects of language emerge should be studied systematically, instead of using one \"hard\" and one random split. \nThe \"evolutionary\" part has a similar issue. The paper draws conclusions from three anecdotal rules for replacing the population. There is no systematic analysis of the \"evolution\" process, not even studying a range of the parameter epsilon, which is set 0.8 (arbitrarily? to fit the story? we do not know).  Drawing conclusions based on anecdotal evidence is bad scientific practice, that ICLR should discourage.\n\nWhile the ideas in this paper are innovative and exciting, the paper promises much more than its analysis supports, and the paper is not ready for publication.\n\nOther comments: \n-- The paper states that \"darker blue bars\" in figure 2 are higher. The statistical analysis is not well explained, not even in the supplemental, so it is hard to tell which differences are significant. If data is paired, it would be useful to view data as a scatter plot, instead of a barplot which hides the pairing. BTW, p<0.05 is not a \"strong support\", but rather is the most permissive threshold. The results in the supplemental may be stronger. \n-- \"Variations in replacement strategy tend to not affect performance.\" This is a key result of the paper and mush be quantified and analyzed. Authors should define some space of replacements strategies (e.g. in in parametric ways like  how often and how many agents are replaced), then compute performance difference as a function of grid search over the parameter space and show a figure. \n-- \"We stop after 8 generations\". Justify with data. \n-- Other parts of the paper make additional claims, that should similarly be systematically analyzed and supported with data-driven evidence.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "1: Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "The paper continues the line of work on emergent compositionality in dialogs, and here it is extended to handle groups of interacting agents that pass language in an evolutionary way from one generation to another. The key idea is that with group of interacting agents, if some agents are replaced with new ones, then the newbies would learn the same language as the group. \n\nGeneral assessment: \n\nThe setup of the paper is is interesting, and the paper covers a lot of ground. The paper makes bold claims like \"cultural transmission induces compositionality\" and \"Variations in replacement strategy tend to not affect performance\". Unfortunately, the paper does not systematic provide experimental evidence to adequately support its claims.\n\nIn my experience, the emergence of compositionality (or lack of) in the setup learned here is very sensitive to various aspects of the learning setup, hence are hard to reproduce. Specifically, they the task-and-talk paper by Kottur et al may yield different conclusions, if parameters of the original experiments are modified, even slightly. The current paper does improve the evaluation protocol of Kottur 2017 by reporting variance over runs, but does not explore the parameter space more systematically, hence I am concerned that it may suffer from similar fragility. \n\nIn this light, significantly more evidence should be provided to convince that the experimental results are stable and can be reproduced.  First, one needs to show the experiments repeated over the full range of setup parameters. Including, the vocabulary sizes V_Q and V_A, the number of tasks, the number of attributes per task, and number of agents etc. Similarly, the current paper introduces a new compositional split, generated in one specific way. The effect of the split on what aspects of language emerge should be studied systematically, instead of using one \"hard\" and one random split. \nThe \"evolutionary\" part has a similar issue. The paper draws conclusions from three anecdotal rules for replacing the population. There is no systematic analysis of the \"evolution\" process, not even studying a range of the parameter epsilon, which is set 0.8 (arbitrarily? to fit the story? we do not know).  Drawing conclusions based on anecdotal evidence is bad scientific practice, that ICLR should discourage.\n\nWhile the ideas in this paper are innovative and exciting, the paper promises much more than its analysis supports, and the paper is not ready for publication.\n\nOther comments: \n-- The paper states that \"darker blue bars\" in figure 2 are higher. The statistical analysis is not well explained, not even in the supplemental, so it is hard to tell which differences are significant. If data is paired, it would be useful to view data as a scatter plot, instead of a barplot which hides the pairing. BTW, p<0.05 is not a \"strong support\", but rather is the most permissive threshold. The results in the supplemental may be stronger. \n-- \"Variations in replacement strategy tend to not affect performance.\" This is a key result of the paper and mush be quantified and analyzed. Authors should define some space of replacements strategies (e.g. in in parametric ways like  how often and how many agents are replaced), then compute performance difference as a function of grid search over the parameter space and show a figure. \n-- \"We stop after 8 generations\". Justify with data. \n-- Other parts of the paper make additional claims, that should similarly be systematically analyzed and supported with data-driven evidence.\n"}, "tcdate": 1571996190223}], "openreview_url": "https://openreview.net/forum?id=r1gzoaNtvr", "arxiv_id": "1904.09067", "paper_pdf": "papers/r1gzoaNtvr.pdf", "paper_pdf_sha256": "3d50add8df81ba45f44bea6f2ea167f3079074e42746373f8a72aa7163fb0059", "paper_pdf_bytes": 5518103, "paper_pdf_source": "openreview", "code_url": "https://github.com/mcogswell/evolang", "code_repository": "mcogswell/evolang", "code_commit": "12034f8e8395c95b012af74ce7d121775c7cf4c3", "code_archive": "repos/r1gzoaNtvr.zip", "code_archive_sha256": "1b1fb13df28ad394fb3a40c83ba85eb551ca6c4bc85353efb66d1c06b83c9a78", "code_archive_bytes": 6195247, "code_file_count": 13, "code_extensions": {".py": 12, ".ipynb": 1}, "github_disk_usage_kb": 6042, "github_languages": {"Jupyter Notebook": 9872698, "Python": 161765}, "github_archived": false, "github_pushed_at": "2020-05-20T04:03:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/emergence-of-compositional-language-with-deep"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GM9H3iM7VJ", "year": 2026, "status": "rejected", "title": "LLMail-Inject: A Dataset from a Realistic Adaptive Prompt Injection Challenge", "authors": ["Sahar Abdelnabi", "Aideen Fay", "Ahmed Salem", "Egor Zverev", "Kai-Chieh Liao", "Chi-Huang Liu", "Chun Chih Kuo", "Jannis Weigend", "Danyael Manlangit", "Alex Apostolov", "Haris Umair", "João Donato", "Masayuki Kawakita", "Athar Mahboob", "Tran Huu Bach", "Tsun-Han Chiang", "Myeongjin Cho", "Hajin Choi", "Byeonghyeon Kim", "HyeonjinLee", "Benjamin Pannell", "Conor McCauley", "Mark Russinovich", "Andrew Paverd", "Giovanni Cherubin"], "authorids": ["~Sahar_Abdelnabi1", "~Aideen_Fay1", "~Ahmed_Salem2", "~Egor_Zverev1", "~Kai-Chieh_Liao1", "~Chi-Huang_Liu1", "~Chun_Chih_Kuo1", "~Jannis_Weigend1", "~Danyael_Manlangit1", "~Alex_Apostolov1", "~Haris_Umair1", "~João_Donato1", "~Masayuki_Kawakita1", "~Athar_Mahboob1", "~Tran_Huu_Bach1", "~Tsun-Han_Chiang2", "~Myeongjin_Cho1", "~Hajin_Choi1", "~Byeonghyeon_Kim1", "~HyeonjinLee1", "~Benjamin_Pannell1", "~Conor_McCauley1", "~Mark_Russinovich1", "~Andrew_Paverd1", "~Giovanni_Cherubin1"], "authors_source": "OpenReview API", "abstract": "Indirect Prompt Injection attacks exploit a fundamental weakness of large language models (LLMs): the inability to reliably separate instructions from data. This vulnerability poses critical real-world security risks, yet systematic evaluation against adaptive adversaries remains largely unexplored. We introduce LLMail-Inject, the first large-scale public challenge simulating a realistic email-assistant environment—a high-value attack surface in practice. Involving 839 participants, the challenge produced 208,095 unique attack prompts across multiple LLM architectures and retrieval configurations. Unlike prior benchmarks, LLMail-Inject requires end-to-end compromise: attacks must be retrieved, adaptively evade defenses, trigger unauthorized tool calls with correct formatting, and exfiltrate contextual data.\nOur findings reveal a stark gap between perceived and actual robustness: while state-of-the-art models achieve <5% success on existing benchmarks, LLMail-Inject drives success rates to 32%, exposing the fragility of current defenses under realistic conditions. We release the dataset, code, and analysis to catalyze research toward structural, practical defenses against prompt injection.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "h07hSBD2Lp", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17961/Reviewer_gV14"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents LLMail-Inject, a large-scale dataset of indirect prompt injection attacks collected through a public challenge involving 839 participants, resulting in over 200k unique attack prompts. The authors also conduct a comprehensive evaluation of existing defense mechanisms against these attacks, revealing several key insights about the gap between benchmark performance and real-world attack complexity.", "review_text": "This paper presents LLMail-Inject, a large-scale dataset of indirect prompt injection attacks collected through a public challenge involving 839 participants, resulting in over 200k unique attack prompts. The authors also conduct a comprehensive evaluation of existing defense mechanisms against these attacks, revealing several key insights about the gap between benchmark performance and real-world attack complexity.", "strengths": "1. All attacks were human-generated by participants attempting to solve real challenges, avoiding the template-based limitations of existing datasets. \n2. The competition-based approach successfully gathered over 200k unique attack prompts with rich diversity in attack strategies, representing an unprecedented scale compared to existing benchmarks.\n3. The paper is well-structured with overall good visualizations (except Figure 3). The systematic comparison of defenses across multiple dimensions provides in-depth insights.", "weaknesses": "1. **Representativeness of the Scenario:** This paper focuses on email agents as the attack scenario. While email processing is a common use case, it may not capture the full diversity of real-world applications where prompt injection attacks can occur, such as web search, coding assistants, or customer support bots. The authors are encouraged to discuss the generalizability of their findings beyond email agents and discuss whether the dataset can be adapted to other scenarios.\n2. **LLM Selection:** Only two LLMs (microsoft/Phi-3-medium-128k-instruct and GPT-4o-mini) are considered in the challenge. Given the Phi-3 does not possess the function-calling capability natively, it seems not a suitable choice for evaluating prompt injection attacks targeting function-calling agents. Including more diverse and capable LLMs, especially those with built-in function-calling features like Llama-3 would further enhance the relevance of the dataset to real-world applications.\n3. **Discussion of Threat Models:** The challenge assumes attackers have complete knowledge of defense mechanisms, which may not reflect real-world scenarios where defenders may keep their methods confidential. In this regard, findings like \"LLMail-Inject drives success rates to 32%, exposing the fragility of current defenses under realistic conditions\" may somehow overstate the practical risk. The authors are encouraged to discuss the impact of different threat models on the evaluation results.\n4. **Guidance for Dataset Usage:** With over 200k data, researchers cannot practically use the entire dataset. The authors are encouraged to provide more guidance on effectively utilizing the dataset, such as:\n   - How to construct representative subsets for different research goals\n   - Which difficulty levels or defense combinations are most informative", "questions": "1. Are findings from the email agent scenario generalizable to other application domains?\n2. Can the dataset be transferred or adapted to evaluate prompt injection attacks in other contexts?\n3. What guidance can the authors provide for researchers on effectively utilizing the large dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents LLMail-Inject, a large-scale dataset of indirect prompt injection attacks collected through a public challenge involving 839 participants, resulting in over 200k unique attack prompts. The authors also conduct a comprehensive evaluation of existing defense mechanisms against these attacks, revealing several key insights about the gap between benchmark performance and real-world attack complexity.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "1. All attacks were human-generated by participants attempting to solve real challenges, avoiding the template-based limitations of existing datasets. \n2. The competition-based approach successfully gathered over 200k unique attack prompts with rich diversity in attack strategies, representing an unprecedented scale compared to existing benchmarks.\n3. The paper is well-structured with overall good visualizations (except Figure 3). The systematic comparison of defenses across multiple dimensions provides in-depth insights.", "weaknesses": "1. **Representativeness of the Scenario:** This paper focuses on email agents as the attack scenario. While email processing is a common use case, it may not capture the full diversity of real-world applications where prompt injection attacks can occur, such as web search, coding assistants, or customer support bots. The authors are encouraged to discuss the generalizability of their findings beyond email agents and discuss whether the dataset can be adapted to other scenarios.\n2. **LLM Selection:** Only two LLMs (microsoft/Phi-3-medium-128k-instruct and GPT-4o-mini) are considered in the challenge. Given the Phi-3 does not possess the function-calling capability natively, it seems not a suitable choice for evaluating prompt injection attacks targeting function-calling agents. Including more diverse and capable LLMs, especially those with built-in function-calling features like Llama-3 would further enhance the relevance of the dataset to real-world applications.\n3. **Discussion of Threat Models:** The challenge assumes attackers have complete knowledge of defense mechanisms, which may not reflect real-world scenarios where defenders may keep their methods confidential. In this regard, findings like \"LLMail-Inject drives success rates to 32%, exposing the fragility of current defenses under realistic conditions\" may somehow overstate the practical risk. The authors are encouraged to discuss the impact of different threat models on the evaluation results.\n4. **Guidance for Dataset Usage:** With over 200k data, researchers cannot practically use the entire dataset. The authors are encouraged to provide more guidance on effectively utilizing the dataset, such as:\n   - How to construct representative subsets for different research goals\n   - Which difficulty levels or defense combinations are most informative", "questions": "1. Are findings from the email agent scenario generalizable to other application domains?\n2. Can the dataset be transferred or adapted to evaluate prompt injection attacks in other contexts?\n3. What guidance can the authors provide for researchers on effectively utilizing the large dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761986652164}, {"id": "CQCYrGBQcv", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17961/Reviewer_mg4k"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper proposes LLMail-Inject, a large-scale benchmark built from a real-world red-teaming competition simulating an email-assistant environment. It contains over 200K adaptive prompt injection attempts from 292 teams, covering multiple difficulty levels and defenses. The dataset provides realistic, diverse, and context-rich attack samples, enabling comprehensive evaluation of LLM safety and revealing the weaknesses of current prompt injection defenses.", "review_text": "The paper proposes LLMail-Inject, a large-scale benchmark built from a real-world red-teaming competition simulating an email-assistant environment. It contains over 200K adaptive prompt injection attempts from 292 teams, covering multiple difficulty levels and defenses. The dataset provides realistic, diverse, and context-rich attack samples, enabling comprehensive evaluation of LLM safety and revealing the weaknesses of current prompt injection defenses.", "strengths": "1. The proposed dataset is collected from a large-scale, real-world competition, which makes the collected data highly diverse and realistic, providing valuable resources and insights for future research on LLM safety and prompt injection defenses.\n\n2. The attack strategies are contextually relevant and reflect how adaptive prompt injection attacks may occur in practical LLM applications, such as email assistants.\n\n3. The paper provides comprehensive analyses across multiple difficulty levels and defense mechanisms, offering valuable insights into the effectiveness and limitations of current prompt injection defenses.", "weaknesses": "1. The paper could include more recent and stronger baselines for comparison, such as StruQ [1], SecAlign [2], and Meta-SecAlign [3], which represent the state-of-the-art fine-tuning-based defenses against prompt injection.\n\n2. The proposed benchmark focuses solely on the email scenario, which, while realistic, may limit the generalizability of the findings. It would be valuable to include other application contexts, such as document editing, coding, or web agents.\n\n3. Although the dataset captures a wide range of real attack prompts, the paper could further analyze attack category diversity. For example, distinguishing between direct injection, indirect instruction hijacking, and data poisoning can be better characterize what kinds of vulnerabilities the collected samples represent.\n\n[1].Chen, Sizhe, et al. \"{StruQ}: Defending against prompt injection with structured queries.\"\n\n[2].Chen, Sizhe, et al. \"Secalign: Defending against prompt injection with preference optimization.\"\n\n[3].Chen, Sizhe, et al. \"Meta SecAlign: A Secure Foundation LLM Against Prompt Injection Attacks.\"", "questions": "Please see the weakness part above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes LLMail-Inject, a large-scale benchmark built from a real-world red-teaming competition simulating an email-assistant environment. It contains over 200K adaptive prompt injection attempts from 292 teams, covering multiple difficulty levels and defenses. The dataset provides realistic, diverse, and context-rich attack samples, enabling comprehensive evaluation of LLM safety and revealing the weaknesses of current prompt injection defenses.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The proposed dataset is collected from a large-scale, real-world competition, which makes the collected data highly diverse and realistic, providing valuable resources and insights for future research on LLM safety and prompt injection defenses.\n\n2. The attack strategies are contextually relevant and reflect how adaptive prompt injection attacks may occur in practical LLM applications, such as email assistants.\n\n3. The paper provides comprehensive analyses across multiple difficulty levels and defense mechanisms, offering valuable insights into the effectiveness and limitations of current prompt injection defenses.", "weaknesses": "1. The paper could include more recent and stronger baselines for comparison, such as StruQ [1], SecAlign [2], and Meta-SecAlign [3], which represent the state-of-the-art fine-tuning-based defenses against prompt injection.\n\n2. The proposed benchmark focuses solely on the email scenario, which, while realistic, may limit the generalizability of the findings. It would be valuable to include other application contexts, such as document editing, coding, or web agents.\n\n3. Although the dataset captures a wide range of real attack prompts, the paper could further analyze attack category diversity. For example, distinguishing between direct injection, indirect instruction hijacking, and data poisoning can be better characterize what kinds of vulnerabilities the collected samples represent.\n\n[1].Chen, Sizhe, et al. \"{StruQ}: Defending against prompt injection with structured queries.\"\n\n[2].Chen, Sizhe, et al. \"Secalign: Defending against prompt injection with preference optimization.\"\n\n[3].Chen, Sizhe, et al. \"Meta SecAlign: A Secure Foundation LLM Against Prompt Injection Attacks.\"", "questions": "Please see the weakness part above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761979223753}, {"id": "buTeOjP1Vf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17961/Reviewer_nVTh"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "The authors hold a large-scale competition to reward humans for adaptive prompt injections, aiming at practical attacks that break multiple steps of the system execution. The competition is held under a simulated email assistant scenario, where attackers try to trigger the model’s send_email tool call with arguments: destination=contact@contact.com,\ncontent=confirmation. An attack prompt is evaluated under 4 levels of different user (benign) prompts to the LLM. The authors implement prompting/detection defenses in the competition, which test Phi-3-medium and gpt-4o-mini. After the competition, the latest Meta-SecAlign defended LLM is also tested, where the collected attack prompts give an order of magnitude higher attack success rates. The authors promise to release the LLMail-Inject benchmark with 0.2M attack prompts.", "review_text": "The authors hold a large-scale competition to reward humans for adaptive prompt injections, aiming at practical attacks that break multiple steps of the system execution. The competition is held under a simulated email assistant scenario, where attackers try to trigger the model’s send_email tool call with arguments: destination=contact@contact.com,\ncontent=confirmation. An attack prompt is evaluated under 4 levels of different user (benign) prompts to the LLM. The authors implement prompting/detection defenses in the competition, which test Phi-3-medium and gpt-4o-mini. After the competition, the latest Meta-SecAlign defended LLM is also tested, where the collected attack prompts give an order of magnitude higher attack success rates. The authors promise to release the LLMail-Inject benchmark with 0.2M attack prompts.", "strengths": "1. The paper devotes significant efforts on building the community of prompt injection, the top-1 threat to LLM-integrated applications. With a complex competition design, the competition collected a very large human-generated high-quality prompt injection dataset, which would be a great asset for future assessment of the model, given that current attack benchmarks are saturating.\n2. The competition is built on a practical attack scenario, email assistant, where an LLM is very suitable for handling this tedious work and may be mis-directed by a malicious email. The attack goal is hard: eliciting a specific function call with proper parameters. I appreciate the efforts on implementing defenses in the competition to harden the attacker’s trails.\n3.  The paper offers great insight in its analysis about the defense effectiveness and end-to-end attacks. An analysis of a prompt injection defense system (equipped with multiple defenses as existing commercial providers do) is important for this community.", "weaknesses": "1. The competition assumes that the attacker knows the attack target string (trigger the model’s send_email tool call with arguments: destination=contact@contact.com,\ncontent=confirmation). However, in a practical attack scenario, how does the attacker know the name/parameters of a function call that will lead to malicious actions? That information is generally kept private in the LLM system.\n\n2. The selected two victim models are not strong nor representative enough. Phi is a 14B small model without inherent function call (as the authors admit). Gpt-4o is also a stronger model than gpt-4-mini, and with instruction hierarchy defense. \n\n3. It is unclear whether the attack prompts are transferable to attack other tasks beyond email assistant.\n4. Another large-scale prompt injection challenge [1] has also been held, but the authors do not discuss the differences between that work.\n\n[1] Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition", "questions": "The authors mention successful attacks using the system's delimiters. This is prohibited in Meta SecAlign system, see [here](https://github.com/facebookresearch/Meta_SecAlign/blob/main/demo.py#L11). Does the attack prompts against Meta-SecAlign contain Llama 3 delimiters?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors hold a large-scale competition to reward humans for adaptive prompt injections, aiming at practical attacks that break multiple steps of the system execution. The competition is held under a simulated email assistant scenario, where attackers try to trigger the model’s send_email tool call with arguments: destination=contact@contact.com,\ncontent=confirmation. An attack prompt is evaluated under 4 levels of different user (benign) prompts to the LLM. The authors implement prompting/detection defenses in the competition, which test Phi-3-medium and gpt-4o-mini. After the competition, the latest Meta-SecAlign defended LLM is also tested, where the collected attack prompts give an order of magnitude higher attack success rates. The authors promise to release the LLMail-Inject benchmark with 0.2M attack prompts.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "1. The paper devotes significant efforts on building the community of prompt injection, the top-1 threat to LLM-integrated applications. With a complex competition design, the competition collected a very large human-generated high-quality prompt injection dataset, which would be a great asset for future assessment of the model, given that current attack benchmarks are saturating.\n2. The competition is built on a practical attack scenario, email assistant, where an LLM is very suitable for handling this tedious work and may be mis-directed by a malicious email. The attack goal is hard: eliciting a specific function call with proper parameters. I appreciate the efforts on implementing defenses in the competition to harden the attacker’s trails.\n3.  The paper offers great insight in its analysis about the defense effectiveness and end-to-end attacks. An analysis of a prompt injection defense system (equipped with multiple defenses as existing commercial providers do) is important for this community.", "weaknesses": "1. The competition assumes that the attacker knows the attack target string (trigger the model’s send_email tool call with arguments: destination=contact@contact.com,\ncontent=confirmation). However, in a practical attack scenario, how does the attacker know the name/parameters of a function call that will lead to malicious actions? That information is generally kept private in the LLM system.\n\n2. The selected two victim models are not strong nor representative enough. Phi is a 14B small model without inherent function call (as the authors admit). Gpt-4o is also a stronger model than gpt-4-mini, and with instruction hierarchy defense. \n\n3. It is unclear whether the attack prompts are transferable to attack other tasks beyond email assistant.\n4. Another large-scale prompt injection challenge [1] has also been held, but the authors do not discuss the differences between that work.\n\n[1] Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition", "questions": "The authors mention successful attacks using the system's delimiters. This is prohibited in Meta SecAlign system, see [here](https://github.com/facebookresearch/Meta_SecAlign/blob/main/demo.py#L11). Does the attack prompts against Meta-SecAlign contain Llama 3 delimiters?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761957967973}, {"id": "l8iJCNMcYk", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17961/Reviewer_PpjM"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "The paper introduces a prompt injection challenge in which participants craft emails (subject and body) with the aim of tricking an email assistant agent to invoke a specific tool call. In contrast to existing prompt injection competitions this one is more end-to-end and attempts to simulate a realistic agent scenario. The authors provide an overview of how the challenge was implemented, the data that was collected, and then provide some more in-depth analyses of the data.", "review_text": "The paper introduces a prompt injection challenge in which participants craft emails (subject and body) with the aim of tricking an email assistant agent to invoke a specific tool call. In contrast to existing prompt injection competitions this one is more end-to-end and attempts to simulate a realistic agent scenario. The authors provide an overview of how the challenge was implemented, the data that was collected, and then provide some more in-depth analyses of the data.", "strengths": "- Realistic end-to-end nature of the challenge: A strong point of this paper is that it attempts to simulate a realistic agent in an end-to-end setting.\n- Collected data is made publicly available: The dataset collected during the competition is made publicly available, which has the potential of helping researchers evaluate new defense methods.", "weaknesses": "- Limited contribution: While I appreciate the authors' effort in summarizing and presenting the results of this challenge, I struggle to see any major contribution that this paper makes beyond the publication of the attack data. The work in its current form would, in my opinion, be better suited for a venue targeting datasets or benchmarks.\n- No clear insights or research questions: The analysis felt very much like a listing of various summary statistics, without any clear targets about what to investigate. I think the paper would benefit greatly from formulating some clearly defined questions and then trying to extract answers to those questions from the available data.\n- Limited adjustment for confounding: Given that participants were free to select which challenges to solve, there are multiple confounding factors that come into play that make it hard to draw generalizable insights from the collected results (at least without appropriately adjusting for them). Team success rate is one option for countering some of these biases, but a more detailed discussion of how this could affect results would be crucial. For example:\n  - If level 1 is easier for Phi-3 than GPT-4o, this could mean that teams focused more on the Phi-3 levels.\n  - Number of submissions before success seems highly dependent on the order in which people solved tasks.\n- No detailed utility analysis: It would have been nice to not only consider attack success rates but also false positives in all of the different analyses. Such an analysis is crucial when trying to compare, e.g., defenses as in Figure 2(a).", "questions": "Am I missing a major contribution in my assessment above?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a prompt injection challenge in which participants craft emails (subject and body) with the aim of tricking an email assistant agent to invoke a specific tool call. In contrast to existing prompt injection competitions this one is more end-to-end and attempts to simulate a realistic agent scenario. The authors provide an overview of how the challenge was implemented, the data that was collected, and then provide some more in-depth analyses of the data.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- Realistic end-to-end nature of the challenge: A strong point of this paper is that it attempts to simulate a realistic agent in an end-to-end setting.\n- Collected data is made publicly available: The dataset collected during the competition is made publicly available, which has the potential of helping researchers evaluate new defense methods.", "weaknesses": "- Limited contribution: While I appreciate the authors' effort in summarizing and presenting the results of this challenge, I struggle to see any major contribution that this paper makes beyond the publication of the attack data. The work in its current form would, in my opinion, be better suited for a venue targeting datasets or benchmarks.\n- No clear insights or research questions: The analysis felt very much like a listing of various summary statistics, without any clear targets about what to investigate. I think the paper would benefit greatly from formulating some clearly defined questions and then trying to extract answers to those questions from the available data.\n- Limited adjustment for confounding: Given that participants were free to select which challenges to solve, there are multiple confounding factors that come into play that make it hard to draw generalizable insights from the collected results (at least without appropriately adjusting for them). Team success rate is one option for countering some of these biases, but a more detailed discussion of how this could affect results would be crucial. For example:\n  - If level 1 is easier for Phi-3 than GPT-4o, this could mean that teams focused more on the Phi-3 levels.\n  - Number of submissions before success seems highly dependent on the order in which people solved tasks.\n- No detailed utility analysis: It would have been nice to not only consider attack success rates but also false positives in all of the different analyses. Such an analysis is crucial when trying to compare, e.g., defenses as in Figure 2(a).", "questions": "Am I missing a major contribution in my assessment above?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761835938525}], "openreview_url": "https://openreview.net/forum?id=GM9H3iM7VJ", "arxiv_id": "2506.09956", "paper_pdf": "papers/GM9H3iM7VJ.pdf", "paper_pdf_sha256": "53f7a2ce688f04179bd963f9d30964a9beb8cf94d1cb5f5f0e03c1f05c0a13e0", "paper_pdf_bytes": 4227128, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/llmail-inject-challenge", "code_repository": "microsoft/llmail-inject-challenge", "code_commit": "ad115315c1cb34381d20875d6675a6cfe6ca80fa", "code_archive": "repos/GM9H3iM7VJ.zip", "code_archive_sha256": "217d591012cd46d1481fda280e10d3eb14f4d2ab123e101ba6298aa1362c6a71", "code_archive_bytes": 4676926, "code_file_count": 83, "code_extensions": {".py": 61, ".vue": 15, ".ts": 7}, "github_disk_usage_kb": 4113, "github_languages": {"Python": 202588, "Vue": 29962, "TypeScript": 9695, "TypeSpec": 7873, "CSS": 4921, "HTML": 2645}, "github_archived": false, "github_pushed_at": "2026-04-09T07:37:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/llmail-inject-a-dataset-from-a-realistic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XLt0eudh8t", "year": 2025, "status": "rejected", "title": "Efficient Neural Common Neighbor for Temporal Graph Link Prediction", "authors": ["Xiaohui Zhang", "Yanbo Wang", "Xiyuan Wang", "Muhan Zhang"], "authorids": ["~Xiaohui_Zhang9", "~Yanbo_Wang2", "~Xiyuan_Wang1", "~Muhan_Zhang1"], "authors_source": "OpenReview API", "abstract": "Temporal graphs are ubiquitous in real-world scenarios, such as social network, trade and transportation. Predicting dynamic links between nodes in a temporal graph is of vital importance. \\textcolor{blue}{\nTraditional memory-based methods typically leverage the temporal neighborhood of interaction histories to generate node embeddings, which are then aggregated to predict links between source and target nodes. However, these methods primarily focus on learning individual node representations and often neglect the nature of pairwise representation learning aspect. While some recent methods attempt to capture pairwise features, they are less emphasized in large-scale datasets like TGB. Meanwhile, most of these \nexisting methods tend to suffer from high computational complexity due to the repeated calculation of node embeddings.\n}\nMotivated by the success of Neural Common Neighbor (NCN) for static graph link prediction, we propose \\textbf{TNCN}, a temporal version of NCN for link prediction in temporal graphs. Based on a memory-based backbone instead of traditional static graph neural network, TNCN dynamically updates a temporal neighbor dictionary for each node, and utilizes multi-hop common neighbors between the source and target node to learn a more effective pairwise representation. We validate our model on five large-scale real-world datasets from the Temporal Graph Benchmark (TGB), and find that it achieves new state-of-the-art performance on three of them. Additionally, TNCN demonstrates excellent scalability on large datasets, outperforming popular GNN baselines by up to 6.4 times in speed.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "ehixJnIdIF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7018/Reviewer_2N9E"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This work proposes TNCN, which is  a temporal version of NCN based on a memory-based backbone. Comprosing with three key parts: the memory module,the temporal CN extractor,,and the NCN-based prediction head, TNCN improves the performance in terms of efficiency and effectiveness. Comparing with a diverse set of baseline models, the experiments on five datasets demonstrate its outstanding performance.", "review_text": "This work proposes TNCN, which is  a temporal version of NCN based on a memory-based backbone. Comprosing with three key parts: the memory module,the temporal CN extractor,,and the NCN-based prediction head, TNCN improves the performance in terms of efficiency and effectiveness. Comparing with a diverse set of baseline models, the experiments on five datasets demonstrate its outstanding performance.", "strengths": "1. The experiments is substantial and the result is good.  Comparing with 9 baseline models, TNCN performs best on three of the five selected datasets, which emphasizes its effectiveness.\n2. The method section is clearly described using formulas. With clear definition and detailed formulas, the method is well-presented.\n3. There are proofs on the theorems in appendix, which improves the professionalism of the paper.", "weaknesses": "1. Since the process is relatively complicated, it is recommended to provide a pseudo code to make it easier for readers to understand.\n2. I suggest that the experimental part be supplemented with an analysis of the hyperparameters, which can make the values of the hyperparameters more reasonable.", "questions": "1. High surprings values will degrade TNCN's performance. So will the model performance decrease monotonically as the surprise value increases? In other words, will TNCN have the best performance when the surprise value is the lowest?\n2. Why is the result of TNCN-0∼2-hop-CN lower than TNCN-0∼1-hop-CN on the commen dataset in Table 2, which is different from the other three datasets?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes TNCN, which is  a temporal version of NCN based on a memory-based backbone. Comprosing with three key parts: the memory module,the temporal CN extractor,,and the NCN-based prediction head, TNCN improves the performance in terms of efficiency and effectiveness. Comparing with a diverse set of baseline models, the experiments on five datasets demonstrate its outstanding performance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The experiments is substantial and the result is good.  Comparing with 9 baseline models, TNCN performs best on three of the five selected datasets, which emphasizes its effectiveness.\n2. The method section is clearly described using formulas. With clear definition and detailed formulas, the method is well-presented.\n3. There are proofs on the theorems in appendix, which improves the professionalism of the paper.", "weaknesses": "1. Since the process is relatively complicated, it is recommended to provide a pseudo code to make it easier for readers to understand.\n2. I suggest that the experimental part be supplemented with an analysis of the hyperparameters, which can make the values of the hyperparameters more reasonable.", "questions": "1. High surprings values will degrade TNCN's performance. So will the model performance decrease monotonically as the surprise value increases? In other words, will TNCN have the best performance when the surprise value is the lowest?\n2. Why is the result of TNCN-0∼2-hop-CN lower than TNCN-0∼1-hop-CN on the commen dataset in Table 2, which is different from the other three datasets?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed.", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730453218684}, {"id": "r0s4MVJ8Bc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7018/Reviewer_BFZ2"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "This paper proposes a method for link prediction on continuous-time dynamic graphs (CTDGs), aiming to unify two prominent model families: memory-based models and neighbor-based models. The authors introduce TNCN, demonstrating experimentally that it performs better in certain configurations and is more efficient than existing models in the literature.", "review_text": "This paper proposes a method for link prediction on continuous-time dynamic graphs (CTDGs), aiming to unify two prominent model families: memory-based models and neighbor-based models. The authors introduce TNCN, demonstrating experimentally that it performs better in certain configurations and is more efficient than existing models in the literature.", "strengths": "- **Evaluation**: The proposed model is evaluated on established benchmarks, enhancing the reliability of the results.\n\n- **Engineering**: The paper introduces an engineering approach to combine common neighbor (CN) techniques with memory-based methods, integrating these two modeling approaches.", "weaknesses": "**Presentation**: The abstract implies that common neighbor methods are primarily used in static graphs, overlooking their established role in dynamic graph modeling. Additionally, the motivation for combining memory-based and neighbor-based techniques is presented only briefly at the end of the introduction.\n\n**Limited Novelty**: The proposed model's core components primarily consist of established techniques. For instance, the memory-based module closely resembles TGN, lacking additional innovations or a clear positioning relative to other memory-based approaches. Similarly, while TNCN incorporates common neighbor (CN) techniques with optimized computation, it employs multi-hop methods similar to those in models like CAWN, further limiting its novelty. \n\n**Theoretical Claims**: The paper presents familiar results as novel contributions, such as the $O(∣E∣)$ memory complexity for event aggregation, along with upper and lower bounds that are established in prior work (e.g., Caen, 1998). Presenting these as new theoretical results may be misleading.\n\nThe paper exhibits strong engineering but lacks clear new contributions. Theoretical results on memory complexity and specific experimental results are weak, with minimal performance gains in dynamic link prediction tasks. Given the limitations in presentation, contribution, and experimental validation, I recommend rejection.\n\n**Experimental results**: While combining memory and neighbor-based methods is interesting, the observed performance gains are minimal, with certain results (e.g., in Table 5) showing limited improvement using unclear metrics (possibly AUC or AP).", "questions": "1. Could you clarify the main novelty of your memory-based module and its differentiation from TGN?\n2. What is the specific metric used in Table 5? Understanding whether it's AUC or AP is essential for interpreting the results.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method for link prediction on continuous-time dynamic graphs (CTDGs), aiming to unify two prominent model families: memory-based models and neighbor-based models. The authors introduce TNCN, demonstrating experimentally that it performs better in certain configurations and is more efficient than existing models in the literature.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "- **Evaluation**: The proposed model is evaluated on established benchmarks, enhancing the reliability of the results.\n\n- **Engineering**: The paper introduces an engineering approach to combine common neighbor (CN) techniques with memory-based methods, integrating these two modeling approaches.", "weaknesses": "**Presentation**: The abstract implies that common neighbor methods are primarily used in static graphs, overlooking their established role in dynamic graph modeling. Additionally, the motivation for combining memory-based and neighbor-based techniques is presented only briefly at the end of the introduction.\n\n**Limited Novelty**: The proposed model's core components primarily consist of established techniques. For instance, the memory-based module closely resembles TGN, lacking additional innovations or a clear positioning relative to other memory-based approaches. Similarly, while TNCN incorporates common neighbor (CN) techniques with optimized computation, it employs multi-hop methods similar to those in models like CAWN, further limiting its novelty. \n\n**Theoretical Claims**: The paper presents familiar results as novel contributions, such as the $O(∣E∣)$ memory complexity for event aggregation, along with upper and lower bounds that are established in prior work (e.g., Caen, 1998). Presenting these as new theoretical results may be misleading.\n\nThe paper exhibits strong engineering but lacks clear new contributions. Theoretical results on memory complexity and specific experimental results are weak, with minimal performance gains in dynamic link prediction tasks. Given the limitations in presentation, contribution, and experimental validation, I recommend rejection.\n\n**Experimental results**: While combining memory and neighbor-based methods is interesting, the observed performance gains are minimal, with certain results (e.g., in Table 5) showing limited improvement using unclear metrics (possibly AUC or AP).", "questions": "1. Could you clarify the main novelty of your memory-based module and its differentiation from TGN?\n2. What is the specific metric used in Table 5? Understanding whether it's AUC or AP is essential for interpreting the results.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730379901798}, {"id": "tExBorPQqO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7018/Reviewer_P1SL"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The author focuses on link prediction tasks in temporal graphs, which have ubiquitous applications in real-world systems. Instead of implicitly encoding common neighbor features from the historical neighborhood of the target node, the author proposes the Temporal Common Neighbor Extractor to explicitly integrate these features into the process of temporal graph representation learning, achieving both effectiveness and efficiency compared to existing methods on the TGB benchmark. Additionally, the author provides theoretical analysis based on their proposed method, offering a more thorough illustration of the model.", "review_text": "The author focuses on link prediction tasks in temporal graphs, which have ubiquitous applications in real-world systems. Instead of implicitly encoding common neighbor features from the historical neighborhood of the target node, the author proposes the Temporal Common Neighbor Extractor to explicitly integrate these features into the process of temporal graph representation learning, achieving both effectiveness and efficiency compared to existing methods on the TGB benchmark. Additionally, the author provides theoretical analysis based on their proposed method, offering a more thorough illustration of the model.", "strengths": "This paper has several strengths worth noting:\n\n* **Interesting Motivation.** The motivation for extracting common neighbor features is compelling.\n* **Extensive Experiments.** The author has designed a variety of experiments to demonstrate the effectiveness and efficiency of their methods.\n* **Well-Organized Representation.** The paper is well-structured, and the theoretical analysis provides strong support for the proposed methods.", "weaknesses": "However, the paper also has some weaknesses, outlined as follows:\n\n* **Lack of Novelty.** Firstly, the idea of extracting common neighbors in temporal graphs has certainly been explored before. It seems that the design of your key component, the \"CN Extractor,\" closely resembles existing work in KDD2024 [1]. Moreover, your CN extracting component does not appear to include any specific improvements for temporal graphs. Simply extracting \"monotone k-hop events\" does not substantiate this claim.\n* **Lack of Important Baselines.** Since you \"extend Neural Common Neighbor for static prediction methods,\" these static methods should also be included in your comparisons.\n* **Lack of Case Study.** Providing specific case studies could enhance the understanding of your method.\n* **Ambiguous Expression.** What does $emb$ mean in Equation 2? It seems you haven't explained it anywhere—does it refer to memory? Your method encodes the common neighbor neighborhood composed of source-destination node pairs; might this lead to a loss of other information (e.g., nodes that are not common)?\n* **Parameter Analysis.** Given that your method is based on multi-hop common neighbors, analyzing parameters across different multi-hop settings could better validate the robustness of your model.\n\n[1] Co-Neighbor Encoding Schema: A Light-cost Structure Encoding Method for Dynamic Link Prediction, KDD 2024.", "questions": "Q1: In what ways does your approach build upon or differ from existing methods that also extract common neighbors in temporal graphs?\n\nQ2: Can you clarify the importance of including static baselines in your experiments, and how their absence affects the interpretation of your results?\n\nQ3: What specific case studies can you provide to illustrate the practical application and effectiveness of your proposed method?\n\nQ4: Can you provide a detailed explanation of the notation used in your equations, particularly regarding $emb$ in Equation 2, and how this impacts the overall methodology?\n\nQ5: How do you plan to conduct a thorough parameter analysis for the multi-hop common neighbors, and what insights do you anticipate this will provide regarding your model's robustness?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The author focuses on link prediction tasks in temporal graphs, which have ubiquitous applications in real-world systems. Instead of implicitly encoding common neighbor features from the historical neighborhood of the target node, the author proposes the Temporal Common Neighbor Extractor to explicitly integrate these features into the process of temporal graph representation learning, achieving both effectiveness and efficiency compared to existing methods on the TGB benchmark. Additionally, the author provides theoretical analysis based on their proposed method, offering a more thorough illustration of the model.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper has several strengths worth noting:\n\n* **Interesting Motivation.** The motivation for extracting common neighbor features is compelling.\n* **Extensive Experiments.** The author has designed a variety of experiments to demonstrate the effectiveness and efficiency of their methods.\n* **Well-Organized Representation.** The paper is well-structured, and the theoretical analysis provides strong support for the proposed methods.", "weaknesses": "However, the paper also has some weaknesses, outlined as follows:\n\n* **Lack of Novelty.** Firstly, the idea of extracting common neighbors in temporal graphs has certainly been explored before. It seems that the design of your key component, the \"CN Extractor,\" closely resembles existing work in KDD2024 [1]. Moreover, your CN extracting component does not appear to include any specific improvements for temporal graphs. Simply extracting \"monotone k-hop events\" does not substantiate this claim.\n* **Lack of Important Baselines.** Since you \"extend Neural Common Neighbor for static prediction methods,\" these static methods should also be included in your comparisons.\n* **Lack of Case Study.** Providing specific case studies could enhance the understanding of your method.\n* **Ambiguous Expression.** What does $emb$ mean in Equation 2? It seems you haven't explained it anywhere—does it refer to memory? Your method encodes the common neighbor neighborhood composed of source-destination node pairs; might this lead to a loss of other information (e.g., nodes that are not common)?\n* **Parameter Analysis.** Given that your method is based on multi-hop common neighbors, analyzing parameters across different multi-hop settings could better validate the robustness of your model.\n\n[1] Co-Neighbor Encoding Schema: A Light-cost Structure Encoding Method for Dynamic Link Prediction, KDD 2024.", "questions": "Q1: In what ways does your approach build upon or differ from existing methods that also extract common neighbors in temporal graphs?\n\nQ2: Can you clarify the importance of including static baselines in your experiments, and how their absence affects the interpretation of your results?\n\nQ3: What specific case studies can you provide to illustrate the practical application and effectiveness of your proposed method?\n\nQ4: Can you provide a detailed explanation of the notation used in your equations, particularly regarding $emb$ in Equation 2, and how this impacts the overall methodology?\n\nQ5: How do you plan to conduct a thorough parameter analysis for the multi-hop common neighbors, and what insights do you anticipate this will provide regarding your model's robustness?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730017802170}, {"id": "NkGErO76f8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7018/Reviewer_Erh5"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper is about temporal link prediction, which is an interesting topic. The authors propose a temporal version of NCN for link prediction in temporal graphs, which dynamically updates a temporal neighbor dictionary for each node and utilizes multi-hop common neighbors between the source and target node to learn a more effective pairwise representation. The paper is well written and well organized. However, there are several concerns in the current version of the paper that addressing them will increase the quality of this paper.", "review_text": "The paper is about temporal link prediction, which is an interesting topic. The authors propose a temporal version of NCN for link prediction in temporal graphs, which dynamically updates a temporal neighbor dictionary for each node and utilizes multi-hop common neighbors between the source and target node to learn a more effective pairwise representation. The paper is well written and well organized. However, there are several concerns in the current version of the paper that addressing them will increase the quality of this paper.", "strengths": "1 Cutting-edge research directions.\n\n2 Clear writing logic.\n\n3 Sufficient experimental results.", "weaknesses": "1 The authors should have a special discussion on whether the biggest difference between TNCN and NCN is the core contribution of this paper. If so, this should be highlighted. If not, more introduction is needed on the importance of the new scenario.\n\n2 Since the strategy proposed in the paper is built around the batch processing mode of sequential graph learning, whether the batch size will have a different impact on the strategy is something that needs to be considered and discussed.\n\n3 The motivation and contribution of the paper are worthy of recognition, but in the main text, the authors can consider putting more emphasis on the contribution description and logical arrangement. At present, it seems that the proof takes up a certain amount of space, making the method and experiment part seem less substantial, and the information that can be expressed is not clear and comprehensive enough.\n\n4 The authors could consider discussing the computational complexity, especially comparing it to similar methods (including static graphs and temporal graphs).", "questions": "As above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper is about temporal link prediction, which is an interesting topic. The authors propose a temporal version of NCN for link prediction in temporal graphs, which dynamically updates a temporal neighbor dictionary for each node and utilizes multi-hop common neighbors between the source and target node to learn a more effective pairwise representation. The paper is well written and well organized. However, there are several concerns in the current version of the paper that addressing them will increase the quality of this paper.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1 Cutting-edge research directions.\n\n2 Clear writing logic.\n\n3 Sufficient experimental results.", "weaknesses": "1 The authors should have a special discussion on whether the biggest difference between TNCN and NCN is the core contribution of this paper. If so, this should be highlighted. If not, more introduction is needed on the importance of the new scenario.\n\n2 Since the strategy proposed in the paper is built around the batch processing mode of sequential graph learning, whether the batch size will have a different impact on the strategy is something that needs to be considered and discussed.\n\n3 The motivation and contribution of the paper are worthy of recognition, but in the main text, the authors can consider putting more emphasis on the contribution description and logical arrangement. At present, it seems that the proof takes up a certain amount of space, making the method and experiment part seem less substantial, and the information that can be expressed is not clear and comprehensive enough.\n\n4 The authors could consider discussing the computational complexity, especially comparing it to similar methods (including static graphs and temporal graphs).", "questions": "As above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729501318355}], "openreview_url": "https://openreview.net/forum?id=XLt0eudh8t", "arxiv_id": "2406.07926", "paper_pdf": "papers/XLt0eudh8t.pdf", "paper_pdf_sha256": "86d0e9925d3f58f0f3e7aa3b9aab31ab58621bfd3871f1f92db5fc8a80bc8d55", "paper_pdf_bytes": 625900, "paper_pdf_source": "openreview", "code_url": "https://github.com/GraphPKU/TNCN", "code_repository": "GraphPKU/TNCN", "code_commit": "4c93a2a6c67c4033bab50a2c0a6833cbf8f24b5b", "code_archive": "repos/XLt0eudh8t.zip", "code_archive_sha256": "e9548cd9debcdf4ab87091e993456d48f3341ccda68da61a695fc46d80dc048b", "code_archive_bytes": 167128, "code_file_count": 60, "code_extensions": {".py": 60}, "github_disk_usage_kb": 272, "github_languages": {"Python": 250960}, "github_archived": false, "github_pushed_at": "2025-05-04T17:21:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-neural-common-neighbor-for-temporal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iT1ttQXwOg", "year": 2024, "status": "rejected", "title": "Equivariant Deep Weight Space Alignment", "authors": ["Aviv Navon", "Aviv Shamsian", "Ethan Fetaya", "Gal Chechik", "Nadav Dym", "Haggai Maron"], "authorids": ["~Aviv_Navon1", "~Aviv_Shamsian1", "~Ethan_Fetaya1", "~Gal_Chechik1", "~Nadav_Dym1", "~Haggai_Maron1"], "authors_source": "OpenReview API", "abstract": "Permutation symmetries of deep networks make simple operations like model averaging and similarity estimation challenging. In many cases, aligning the weights of the networks, i.e., finding optimal permutations between their weights, is necessary. More generally, weight alignment is essential for a wide range of applications, from model merging, through exploring the optimization landscape of deep neural networks, to defining meaningful distance functions between neural networks. Unfortunately, weight alignment is an NP-hard problem. Prior research has mainly focused on solving relaxed versions of the alignment problem, leading to either time-consuming methods or sub-optimal solutions. To accelerate the alignment process and improve its quality, we propose a novel framework aimed at learning to solve the weight alignment problem, which we name DEEP-ALIGN. To that end,  we first demonstrate that weight alignment adheres to two fundamental symmetries and then, propose a deep architecture that respects these symmetries. Notably, our framework does not require any labeled data. We provide a theoretical analysis of our approach and evaluate DEEP-ALIGN on several types of network architectures and learning setups. Our experimental results indicate that a feed-forward pass with DEEP-ALIGN produces better or equivalent alignments compared to those produced by current optimization algorithms. Additionally, our alignments can be used as an initialization for other methods to gain even better solutions with a significant speedup in convergence.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "QpJ4D9deeX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3288/Reviewer_qGCa"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper attempts to solve the weight alignment problem by introducing a novel deep learning framework. The proposed DEEP-ALIGN method is fast and produce high quality alignment results after the pretrained process. DEEP-ALIGN does not require any labeled data but only depends on input network weight vectors. In addition to a theoretical analysis of the approach, experimental results support that a feed-forward pass with DEEP-ALIGN produces better or equivalent alignments compared to those produced by current optimization algorithms.", "review_text": "This paper attempts to solve the weight alignment problem by introducing a novel deep learning framework. The proposed DEEP-ALIGN method is fast and produce high quality alignment results after the pretrained process. DEEP-ALIGN does not require any labeled data but only depends on input network weight vectors. In addition to a theoretical analysis of the approach, experimental results support that a feed-forward pass with DEEP-ALIGN produces better or equivalent alignments compared to those produced by current optimization algorithms.", "strengths": "1. DEEP-ALIGN approach does not require labeled data but only the trained weight vectors.\n2. DEEP-ALIGN performs on par or outperforms optimization-based approaches while significantly reducing the runtime or improving the quality of the alignments.\n3. DEEP-ALIGN maintains good performance as an extra initialization step for Sinkhorn method on OOD data.\n4. The theoretical analysis ensures the existence of the approximate solution that can be represented by DEEP-ALIGN architecture.", "weaknesses": "1. This method has to be pre-trained in advance. For experiments conducted on MNIST and CIFAR10, 8000 shallow networks are used to train the model to achieve the claimed performance. For deeper networks that are more common in nowadays deep learning task, this number may grow fast and the training cost can be unacceptable. Also, this paper does not provide experimental results about the relationship between the complexity of weight vectors and the scale of needed training data.\n\n2. The real world applications for this method are unclear due to long the pre-train time. It seems that the traditional methods can compute the weight alignment much more effectively unless a ton of weight vectors need to be aligned so that the inference phase of DEEP-ALIGN dominates. What scenarios feature such characteristics?", "questions": "1. What is the definitions of $\\rho_1$ and $\\rho_2$ in the section 2 Preliminaries?\n2. Is the assumption that the minimizer $k$ in equation (2) is unique reasonable? Any argument that such cases would be rare in practice?\n3. I'm curious about the generalization capability of the model. As mentioned in the paper, the model will be first trained on a dataset of weight vectors and then applied to unseen weight vectors. Is there any possibility that one pertained DEEP-ALIGN model can work for networks of slightly different architectures? Due to the expensive cost of training phase, it can be helpful if one pertained model can work for more scenarios.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper attempts to solve the weight alignment problem by introducing a novel deep learning framework. The proposed DEEP-ALIGN method is fast and produce high quality alignment results after the pretrained process. DEEP-ALIGN does not require any labeled data but only depends on input network weight vectors. In addition to a theoretical analysis of the approach, experimental results support that a feed-forward pass with DEEP-ALIGN produces better or equivalent alignments compared to those produced by current optimization algorithms.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. DEEP-ALIGN approach does not require labeled data but only the trained weight vectors.\n2. DEEP-ALIGN performs on par or outperforms optimization-based approaches while significantly reducing the runtime or improving the quality of the alignments.\n3. DEEP-ALIGN maintains good performance as an extra initialization step for Sinkhorn method on OOD data.\n4. The theoretical analysis ensures the existence of the approximate solution that can be represented by DEEP-ALIGN architecture.", "weaknesses": "1. This method has to be pre-trained in advance. For experiments conducted on MNIST and CIFAR10, 8000 shallow networks are used to train the model to achieve the claimed performance. For deeper networks that are more common in nowadays deep learning task, this number may grow fast and the training cost can be unacceptable. Also, this paper does not provide experimental results about the relationship between the complexity of weight vectors and the scale of needed training data.\n\n2. The real world applications for this method are unclear due to long the pre-train time. It seems that the traditional methods can compute the weight alignment much more effectively unless a ton of weight vectors need to be aligned so that the inference phase of DEEP-ALIGN dominates. What scenarios feature such characteristics?", "questions": "1. What is the definitions of $\\rho_1$ and $\\rho_2$ in the section 2 Preliminaries?\n2. Is the assumption that the minimizer $k$ in equation (2) is unique reasonable? Any argument that such cases would be rare in practice?\n3. I'm curious about the generalization capability of the model. As mentioned in the paper, the model will be first trained on a dataset of weight vectors and then applied to unseen weight vectors. Is there any possibility that one pertained DEEP-ALIGN model can work for networks of slightly different architectures? Due to the expensive cost of training phase, it can be helpful if one pertained model can work for more scenarios.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698825863511}, {"id": "voOewVCJUj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3288/Reviewer_ZcWa"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work presents a learned algorithm to find permutations that align the weights of two networks (e.g. for federated learning, continual learning, model merging). The key innovation of the algorithm is it uses a weight space network to turn input weight vectors into embeddings which can then be aligned as in previous works (via distance in a metric space and Hungarian algorithm). This weight space network is trained using the Sinkhorn method to make the process of finding permutations differentiable.", "review_text": "This work presents a learned algorithm to find permutations that align the weights of two networks (e.g. for federated learning, continual learning, model merging). The key innovation of the algorithm is it uses a weight space network to turn input weight vectors into embeddings which can then be aligned as in previous works (via distance in a metric space and Hungarian algorithm). This weight space network is trained using the Sinkhorn method to make the process of finding permutations differentiable.", "strengths": "This approach appears to improve upon and is complementary to existing approaches, in that it aims to give better permutations than weight matching and is faster than fully learned permutations (termed *Sinkhorn* in this work), but can also be combined with the *Sinkhorn* algorithm. Proofs of exactness and equivalence to activation matching guarantee that this method will be as similarly reliable as activation alignment, where the *Sinkhorn* algorithm does not have the same guarantees. The ability to change the objective function for training the deep weight network makes the algorithm more expressive, akin to *Sinkhorn* algorithm, e.g. so that it can directly optimize for linear connectivity rather than the $L_2$ norm between weights. Some of the results (e.g. table 2) look promising.", "weaknesses": "The main limitation I see is that results are not given for sufficiently large/complex models/tasks. For the barrier results, it is known (e.g. in Ainsworth et al. or Jordan et al.) that deeper networks are harder to align so that they are linearly connected, whereas easier networks (e.g. 3-layer MNIST networks) are relatively trivial to align. Also, it's not clear if Deep-Align reliably beats activation matching (which is still relatively fast and scalable), or if Deep-Align + Sinkhorn can beat weight matching + Sinkhorn.\n\nThe major methodological innovations could use much more explanation. The DWSNet would benefit from a full detailed description, given that it is central to the presented approach and based on a very recent work that may not be well known. The Siamese structure and why it guarantees equivariance to transposition should in particular be fully described.", "questions": "How does this method scale (re. accuracy and computation time) to larger networks (e.g. VGG models) and residual networks?\nHow does weight-matching + Sinkhorn compare with Deep-Align + Sinkhorn?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a learned algorithm to find permutations that align the weights of two networks (e.g. for federated learning, continual learning, model merging). The key innovation of the algorithm is it uses a weight space network to turn input weight vectors into embeddings which can then be aligned as in previous works (via distance in a metric space and Hungarian algorithm). This weight space network is trained using the Sinkhorn method to make the process of finding permutations differentiable.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This approach appears to improve upon and is complementary to existing approaches, in that it aims to give better permutations than weight matching and is faster than fully learned permutations (termed *Sinkhorn* in this work), but can also be combined with the *Sinkhorn* algorithm. Proofs of exactness and equivalence to activation matching guarantee that this method will be as similarly reliable as activation alignment, where the *Sinkhorn* algorithm does not have the same guarantees. The ability to change the objective function for training the deep weight network makes the algorithm more expressive, akin to *Sinkhorn* algorithm, e.g. so that it can directly optimize for linear connectivity rather than the $L_2$ norm between weights. Some of the results (e.g. table 2) look promising.", "weaknesses": "The main limitation I see is that results are not given for sufficiently large/complex models/tasks. For the barrier results, it is known (e.g. in Ainsworth et al. or Jordan et al.) that deeper networks are harder to align so that they are linearly connected, whereas easier networks (e.g. 3-layer MNIST networks) are relatively trivial to align. Also, it's not clear if Deep-Align reliably beats activation matching (which is still relatively fast and scalable), or if Deep-Align + Sinkhorn can beat weight matching + Sinkhorn.\n\nThe major methodological innovations could use much more explanation. The DWSNet would benefit from a full detailed description, given that it is central to the presented approach and based on a very recent work that may not be well known. The Siamese structure and why it guarantees equivariance to transposition should in particular be fully described.", "questions": "How does this method scale (re. accuracy and computation time) to larger networks (e.g. VGG models) and residual networks?\nHow does weight-matching + Sinkhorn compare with Deep-Align + Sinkhorn?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698789432158}, {"id": "VJVit8FaX3", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3288/Reviewer_VRRk"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors propose a data-driven method to perform weight matching between neural networks. It aims to solve the assignment problem by training a Siamese architecture to perform activation matching and produce permutation matrices. The architecture is trained on generated pairs of weight vectors, where one vector is a noisy, randomly permuted version of the other, as well as on unlabeled pairs.\n\nThe authors embed their method into a group theoretic framework and provide proofs for several properties of their architecture: it is equivariant to permutation and is exact, meaning that it, under mild assumptions, converges to an optimal solution if a zero error solution exists.\n\nThe method is evaluated on several weight matching benchmarks for image classification tasks and implicit neural field networks. It consistently outperforms previous work while being orders of magnitude faster.", "review_text": "The authors propose a data-driven method to perform weight matching between neural networks. It aims to solve the assignment problem by training a Siamese architecture to perform activation matching and produce permutation matrices. The architecture is trained on generated pairs of weight vectors, where one vector is a noisy, randomly permuted version of the other, as well as on unlabeled pairs.\n\nThe authors embed their method into a group theoretic framework and provide proofs for several properties of their architecture: it is equivariant to permutation and is exact, meaning that it, under mild assumptions, converges to an optimal solution if a zero error solution exists.\n\nThe method is evaluated on several weight matching benchmarks for image classification tasks and implicit neural field networks. It consistently outperforms previous work while being orders of magnitude faster.", "strengths": "- The presented method is grounded in theory, the authors provide useful theorems and the framework is technically sound\n- The method applies existing concepts for data-driven assignment solvers to the weight matching problem, which is a novel contribution\n- The method is well evaluated and results clearly outperform previous methods while being much faster\n- The paper is mostly well-written and presented", "weaknesses": "- The network that maps from weight embedding to activation space and its connection to stage 1 is unclear (see below for questions). \n- The networks generalization capabilities are limited. In order to outperform previous methods, the architecture has to be trained on different weights of the same network architecture, solving the same task as during inference. It still performs reasonably well on slight OOD weights though.", "questions": "- I don't fully understand the network mapping from weight embeddings to activation space. If the network just maps onto the bias vectors, the input weights do not have any influence on the estimated permutation anymore, would that be correct? This seems to be unintuitive to me and I would like the authors to clarify.\n- Also, if my interpretation is correct, what is the point of feeding the weights $w$ to the network at all and not just use the bias? What is the purpose of the first stage of the architecture?\n\n------------\nI thank the authors for the provided clarifications - the method is clear to me now. In total, my score remains as it is.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a data-driven method to perform weight matching between neural networks. It aims to solve the assignment problem by training a Siamese architecture to perform activation matching and produce permutation matrices. The architecture is trained on generated pairs of weight vectors, where one vector is a noisy, randomly permuted version of the other, as well as on unlabeled pairs.\n\nThe authors embed their method into a group theoretic framework and provide proofs for several properties of their architecture: it is equivariant to permutation and is exact, meaning that it, under mild assumptions, converges to an optimal solution if a zero error solution exists.\n\nThe method is evaluated on several weight matching benchmarks for image classification tasks and implicit neural field networks. It consistently outperforms previous work while being orders of magnitude faster.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "- The presented method is grounded in theory, the authors provide useful theorems and the framework is technically sound\n- The method applies existing concepts for data-driven assignment solvers to the weight matching problem, which is a novel contribution\n- The method is well evaluated and results clearly outperform previous methods while being much faster\n- The paper is mostly well-written and presented", "weaknesses": "- The network that maps from weight embedding to activation space and its connection to stage 1 is unclear (see below for questions). \n- The networks generalization capabilities are limited. In order to outperform previous methods, the architecture has to be trained on different weights of the same network architecture, solving the same task as during inference. It still performs reasonably well on slight OOD weights though.", "questions": "- I don't fully understand the network mapping from weight embeddings to activation space. If the network just maps onto the bias vectors, the input weights do not have any influence on the estimated permutation anymore, would that be correct? This seems to be unintuitive to me and I would like the authors to clarify.\n- Also, if my interpretation is correct, what is the point of feeding the weights $w$ to the network at all and not just use the bias? What is the purpose of the first stage of the architecture?\n\n------------\nI thank the authors for the provided clarifications - the method is clear to me now. In total, my score remains as it is.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698656325182}, {"id": "KJUYaaW4JW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3288/Reviewer_ayDH"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a learning-based framework that learns to predict the optimal weight alignment between two neural networks. The proposed architecture is equivariance to permutation symmetry of the input neural networks. The authors prove that the approach can approximate the Activation Matching algorithm and guarantees to produce the correct alignment when a perfect alignment exists. At inference time, the framework aligns unseen network pairs without additional optimization. Experiment results show that the proposed approach is faster and produces better alignment than optimization-based approaches. The predicted alignment can also be used as initialization for optimization-based approaches to improve their alignment quality.", "review_text": "This paper proposes a learning-based framework that learns to predict the optimal weight alignment between two neural networks. The proposed architecture is equivariance to permutation symmetry of the input neural networks. The authors prove that the approach can approximate the Activation Matching algorithm and guarantees to produce the correct alignment when a perfect alignment exists. At inference time, the framework aligns unseen network pairs without additional optimization. Experiment results show that the proposed approach is faster and produces better alignment than optimization-based approaches. The predicted alignment can also be used as initialization for optimization-based approaches to improve their alignment quality.", "strengths": "-\tThe idea of incorporating symmetries of the weight space is both natural and novel, and its effectiveness is clearly demonstrated in experiments.\n-\tAn efficient method for finding weight alignment is an important problem in model merging, with potential applications in real-time merging in federated learning and weight-space clustering. It also allows more efficient experimentations in understanding the loss landscape of deep networks.\n-\tThe paper is easy to follow. Writing is clear and concise.", "weaknesses": "-\tThe proposed framework requires new training for every new input neural network architecture, even for minor changes such as adding or modifying the size of a layer. \n-\tIt is not clear how to choose the weights for the linear combination of the three loss functions in Section 4.2. Also, as shown in Appendix F, the two unsupervised loss functions are not always aligned, and it is not clear which loss to include before training. Finally, the evaluation metrics, barrier and AUC, are based on model merging tasks, but the supervised loss does not seem to help on these metrics directly.\n-\tThe practical contribution would be more convincing if the authors can demonstrate the framework’s effectiveness in applications that require weight alignment.\n\n### Minor: \n-\t$\\rho_1$ and $\\rho_2$ in the first paragraph of Section 2 are not defined. Are they representations?\n-\tSection 4.1: might be better to state what $\\theta$ is somewhere.\n-\tFigure 5: The “Weight Matching” and “DEEP ALIGN + Weight Matching” have very similar colors.", "questions": "-\tDoes the exactness property (proposition 5) apply to test data as well?\n-\tHow do the three losses interact with each other? Does the two unsupervised loss provably help improve the supervised loss? \n-\tFigure 3: Why does the loss decrease on the interpolation around $\\lambda=$ 0.1 or 0.9, especially in CIFAR10 CNNs?\n-\tWhy is incorporating equivariance beneficial to the model performance? It is intuitively clear that weight alignment methods should respect the symmetry of the neural networks, but there does not seem to be evidence of the link between incorporating equivariance and improved performance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a learning-based framework that learns to predict the optimal weight alignment between two neural networks. The proposed architecture is equivariance to permutation symmetry of the input neural networks. The authors prove that the approach can approximate the Activation Matching algorithm and guarantees to produce the correct alignment when a perfect alignment exists. At inference time, the framework aligns unseen network pairs without additional optimization. Experiment results show that the proposed approach is faster and produces better alignment than optimization-based approaches. The predicted alignment can also be used as initialization for optimization-based approaches to improve their alignment quality.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "-\tThe idea of incorporating symmetries of the weight space is both natural and novel, and its effectiveness is clearly demonstrated in experiments.\n-\tAn efficient method for finding weight alignment is an important problem in model merging, with potential applications in real-time merging in federated learning and weight-space clustering. It also allows more efficient experimentations in understanding the loss landscape of deep networks.\n-\tThe paper is easy to follow. Writing is clear and concise.", "weaknesses": "-\tThe proposed framework requires new training for every new input neural network architecture, even for minor changes such as adding or modifying the size of a layer. \n-\tIt is not clear how to choose the weights for the linear combination of the three loss functions in Section 4.2. Also, as shown in Appendix F, the two unsupervised loss functions are not always aligned, and it is not clear which loss to include before training. Finally, the evaluation metrics, barrier and AUC, are based on model merging tasks, but the supervised loss does not seem to help on these metrics directly.\n-\tThe practical contribution would be more convincing if the authors can demonstrate the framework’s effectiveness in applications that require weight alignment.\n\n### Minor: \n-\t$\\rho_1$ and $\\rho_2$ in the first paragraph of Section 2 are not defined. Are they representations?\n-\tSection 4.1: might be better to state what $\\theta$ is somewhere.\n-\tFigure 5: The “Weight Matching” and “DEEP ALIGN + Weight Matching” have very similar colors.", "questions": "-\tDoes the exactness property (proposition 5) apply to test data as well?\n-\tHow do the three losses interact with each other? Does the two unsupervised loss provably help improve the supervised loss? \n-\tFigure 3: Why does the loss decrease on the interpolation around $\\lambda=$ 0.1 or 0.9, especially in CIFAR10 CNNs?\n-\tWhy is incorporating equivariance beneficial to the model performance? It is intuitively clear that weight alignment methods should respect the symmetry of the neural networks, but there does not seem to be evidence of the link between incorporating equivariance and improved performance.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697821993069}], "openreview_url": "https://openreview.net/forum?id=iT1ttQXwOg", "arxiv_id": "2310.13397", "paper_pdf": "papers/iT1ttQXwOg.pdf", "paper_pdf_sha256": "30d39a410e5c50490f90a6635e41c8cb40031b6c887e1713dd15d9668116efdf", "paper_pdf_bytes": 3896857, "paper_pdf_source": "openreview", "code_url": "https://github.com/AvivNavon/deep-align", "code_repository": "AvivNavon/deep-align", "code_commit": "7712044ddbf3d5648b8f7bdf31221f775895ed78", "code_archive": "repos/iT1ttQXwOg.zip", "code_archive_sha256": "64b9bb906a72857ab254b8eb681211eb0b78788da53eba5292d6dd8c4e055c69", "code_archive_bytes": 263315, "code_file_count": 29, "code_extensions": {".py": 28, ".ipynb": 1}, "github_disk_usage_kb": 225, "github_languages": {"Python": 292292, "Jupyter Notebook": 20866}, "github_archived": false, "github_pushed_at": "2024-08-18T18:25:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/equivariant-deep-weight-space-alignment"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Siln8xpTMrZ", "year": 2023, "status": "rejected", "title": "DADAO: Decoupled Accelerated Decentralized Asynchronous Optimization", "authors": ["Adel Nabli", "Edouard Oyallon"], "authorids": ["~Adel_Nabli1", "~Edouard_Oyallon1"], "authors_source": "OpenReview API", "abstract": "DADAO is a novel decentralized asynchronous stochastic first order algorithm to minimize a sum of $L$-smooth and $\\mu$-strongly convex functions distributed over a time-varying connectivity network of size $n$.  We model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, decoupling the computation and communication steps in addition to making the whole approach completely asynchronous. Our method employs primal gradients and does not use a multi-consensus inner loop nor other ad-hoc mechanisms such as Error Feedback, Gradient Tracking, or a Proximal operator. By relating the inverse of the smallest positive eigenvalue $\\chi^*_1$ and the effective resistance $\\chi_2^*$ of our graph to a necessary minimal communication rate between nodes of the network, we show that our algorithm requires $\\mathcal{O}(n\\sqrt{\\frac{L}{\\mu}}\\log \\epsilon)$ local gradients and only  $\\mathcal{O}(n\\sqrt{\\chi_1^*\\chi_2^*}\\sqrt{\\frac{L}{\\mu}}\\log \\epsilon)$ communications to reach a precision $\\epsilon$. If SGD with uniform noise $\\sigma^2$ is used, we reach a precision $\\epsilon$ with same speed, up to a bias term in $\\mathcal{O}(\\frac{\\sigma^2}{\\sqrt{\\mu L}})$. This improves upon the bounds obtained with current state-of-the-art approaches, our simulations validating the strength of our relatively unconstrained method.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "3265l6RWCs", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1632/Reviewer_L3mL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors study the decentralized asynchronous optimization problem on a time-varying network with n nodes where the main purpose is to minimize the sum of  L-smooth and \\mu-strongly convex functions distributed on the nodes. The authors decouple the computation and communication steps in the asynchronous optimization problem by modeling this problem as a union of two independent Poisson Point Processes and consider primal gradients in their proposed method. They show convergence results for this problem under the proper assumptions. They also present numerical results corroborating their findings in the convergence results.\n", "review_text": "A new study on decentralized asynchronous optimization on time-varying graphs is made in this paper. By modeling the problem via Poisson processes, the authors present communication and computation complexity for their proposed algorithm. Some improvements in Subsection 3.3 are required to clarify the findings for stochastic optimization.\n", "strengths": "Strengths:\n-The authors can achieve communication rounds and convergence complexity results by decoupling optimization (computation) from communication via two independent Poisson processes.\n-The authors provide a clear statement of contributions and comparisons to the literature.\n-A novel analysis of asynchronous communications via is provided in Section 3.\n\nWeaknesses:\n-The optimization results in Subsection 3.3 do not contain information on bounds for hyperparameters in the proposed dynamic in Subsection 3.2.\n-The stochastic result (Corollary 3.2.1) shows a non-vanishing error in the rate. It is conventional to tweak the hyperparameters (e.g., a decreasing learning rate) to make this term vanish over time. It is unclear why this term is a constant and if it can be improved. The current format undermines stochastic optimization results.\n-Numerical comparisons are limited to the ADOM+ algorithm. I recommend further comparisons with other algorithms, such as Gradient Tracking.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors study the decentralized asynchronous optimization problem on a time-varying network with n nodes where the main purpose is to minimize the sum of  L-smooth and \\mu-strongly convex functions distributed on the nodes. The authors decouple the computation and communication steps in the asynchronous optimization problem by modeling this problem as a union of two independent Poisson Point Processes and consider primal gradients in their proposed method. They show convergence results for this problem under the proper assumptions. They also present numerical results corroborating their findings in the convergence results.\n", "strength_and_weaknesses": "Strengths:\n-The authors can achieve communication rounds and convergence complexity results by decoupling optimization (computation) from communication via two independent Poisson processes.\n-The authors provide a clear statement of contributions and comparisons to the literature.\n-A novel analysis of asynchronous communications via is provided in Section 3.\n\nWeaknesses:\n-The optimization results in Subsection 3.3 do not contain information on bounds for hyperparameters in the proposed dynamic in Subsection 3.2.\n-The stochastic result (Corollary 3.2.1) shows a non-vanishing error in the rate. It is conventional to tweak the hyperparameters (e.g., a decreasing learning rate) to make this term vanish over time. It is unclear why this term is a constant and if it can be improved. The current format undermines stochastic optimization results.\n-Numerical comparisons are limited to the ADOM+ algorithm. I recommend further comparisons with other algorithms, such as Gradient Tracking.\n", "clarity,_quality,_novelty_and_reproducibility": "-Throughout the manuscript, underlying assumptions are clearly stated.\n-I recommend moving Algorithm 2 from the appendix to the main manuscript.\n-The clarity of convergence results requires some improvements. The main technical contribution of this paper lies in the gossip analysis. However, the optimization results are not presented and elaborated sufficiently. I recommend expanding the discussion in Subsection 3.3.\n", "summary_of_the_review": "A new study on decentralized asynchronous optimization on time-varying graphs is made in this paper. By modeling the problem via Poisson processes, the authors present communication and computation complexity for their proposed algorithm. Some improvements in Subsection 3.3 are required to clarify the findings for stochastic optimization.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667224775949}, {"id": "TEI0vrEeo98", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1632/Reviewer_gkNG"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a decentralized asynchronous stochastic first-order algorithm DADAO to minimize the sum of strongly convex functions over time-varying decentralized networks. This paper develops the theoretical performance guarantee of a related ODE system and tests the empirical performance by numerical simulations. ", "review_text": "This paper proposes an algorithm to solve decentralized strongly convex optimization and establishes some new complexity results. The results are developed for continuing-time frameworks, but the guarantees for implementable practical frameworks are absent. The theoretical claims are also only applicable to a restrictive class of problems. The overall contribution seems to be limited. ", "strengths": "Strength: \n\n1. The proposed algorithm achieves new performance guarantees in both communication edges and gradients.\n\nWeaknesses:\n\n1. Theorems 3.2 and Corollary 3.2.1 provide guarantees for the ODE system Eq.(3) and Eq.(8), respectively. However, because ODE considers a continuous-time regime, it is not obvious whether these results still hold in practical discrete-time implementations. The discrete-time performance guarantee has been shown in prior works (i.e., Theorem 4 of [1]). It would be beneficial if the performance guarantees of discrete-time GD and SGD algorithms can be provided, which is provided in prior work. \n\n2. Theorem 3.2 considers the case where $\\chi_1^* \\chi_2^* \\leq 0.5$. This assumption seems to be a bit restrictive because the constants $\\chi_1^*$ and $\\chi_2^*$ depend on the network itself. It looks that ADOM+ in [1] does not need such a restrictive assumption to ensure convergence. It would be helpful if an additional explanation can be provided to justify why such an assumption is required. It would also be beneficial to discuss whether similar assumptions are required to the mentioned benchmark algorithms. \n\n3. Theorem 3.2 shows the existence of the parameters $\\alpha,\\gamma$ but does not specify how these parameters should be chosen in real networks. It would be helpful to specify the choice of these parameters. Also, does  $\\gamma$ refer to the same quantity in Table 1? \n\n4. Corollary 3.2.1 introduces an additional bias term $\\frac{C_1}{\\sqrt{\\mu L}}$ to characterize the SGD system (8) in a continuous-time regime. It is claimed that \"$L$ allows to adjust the trace-off bias-variance of the descent\". This seems to be a bit confusing because $L$ represents the fixed Lipschitz constant. If one considers a larger $L$ to reduce this term, the decay rate of the term $C_0 e^{-ct\\sqrt{\\mu/L}}$ would deteriorate as well. It would be helpful if the authors could elaborate more on this. \n\n5. When comparing DADAO with ADOM+, it is claimed that ADOM+ has potentially substantially higher expected communication than DADAO. It would be beneficial if the paper can specify the referred scenarios. \n\n6. Various benchmark algorithms are tested in the numerical part. However, it seems that MSDA also has comparable performance with the proposed DADAO algorithm in most of the scenarios. Could the authors provide additional numerical experiments to demonstrate the efficiency of DADAO? It would be helpful if numerical experiments in some special cases such as star or complete networks can suggest DADAO's superior performance.\n\nMinor Issues: There are some typos and some quantities are undefined. \n\n1.  In the abstract, the $O(n \\sqrt{\\frac{L}{\\mu}} \\log \\epsilon)$ gradient complexity is confusing. Should it be  $O(n \\sqrt{\\frac{L}{\\mu}} \\log \\frac{ 1}{ \\epsilon} )$ where $\\epsilon$ is the precision? \n\n2. In Table 1, $\\gamma$ is undefined so that a fair comparison between $\\sqrt{\\chi_1\\chi_2} n $ and $\\chi_1 |\\mathcal{E}|$ is not obvious. \n\n3. When defining $\\chi_2(t)$ in page 4, the quantity $\\Lambda^+(t)$ is undefined. \n\nReferences: \n[1] Kovalev, D., Shulgin, E., Richtárik, P., Rogozin, A., and Gasnikov, A. (2021). ADOM: Accelerated decentralized optimization method for time-varying networks. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a decentralized asynchronous stochastic first-order algorithm DADAO to minimize the sum of strongly convex functions over time-varying decentralized networks. This paper develops the theoretical performance guarantee of a related ODE system and tests the empirical performance by numerical simulations. ", "strength_and_weaknesses": "Strength: \n\n1. The proposed algorithm achieves new performance guarantees in both communication edges and gradients.\n\nWeaknesses:\n\n1. Theorems 3.2 and Corollary 3.2.1 provide guarantees for the ODE system Eq.(3) and Eq.(8), respectively. However, because ODE considers a continuous-time regime, it is not obvious whether these results still hold in practical discrete-time implementations. The discrete-time performance guarantee has been shown in prior works (i.e., Theorem 4 of [1]). It would be beneficial if the performance guarantees of discrete-time GD and SGD algorithms can be provided, which is provided in prior work. \n\n2. Theorem 3.2 considers the case where $\\chi_1^* \\chi_2^* \\leq 0.5$. This assumption seems to be a bit restrictive because the constants $\\chi_1^*$ and $\\chi_2^*$ depend on the network itself. It looks that ADOM+ in [1] does not need such a restrictive assumption to ensure convergence. It would be helpful if an additional explanation can be provided to justify why such an assumption is required. It would also be beneficial to discuss whether similar assumptions are required to the mentioned benchmark algorithms. \n\n3. Theorem 3.2 shows the existence of the parameters $\\alpha,\\gamma$ but does not specify how these parameters should be chosen in real networks. It would be helpful to specify the choice of these parameters. Also, does  $\\gamma$ refer to the same quantity in Table 1? \n\n4. Corollary 3.2.1 introduces an additional bias term $\\frac{C_1}{\\sqrt{\\mu L}}$ to characterize the SGD system (8) in a continuous-time regime. It is claimed that \"$L$ allows to adjust the trace-off bias-variance of the descent\". This seems to be a bit confusing because $L$ represents the fixed Lipschitz constant. If one considers a larger $L$ to reduce this term, the decay rate of the term $C_0 e^{-ct\\sqrt{\\mu/L}}$ would deteriorate as well. It would be helpful if the authors could elaborate more on this. \n\n5. When comparing DADAO with ADOM+, it is claimed that ADOM+ has potentially substantially higher expected communication than DADAO. It would be beneficial if the paper can specify the referred scenarios. \n\n6. Various benchmark algorithms are tested in the numerical part. However, it seems that MSDA also has comparable performance with the proposed DADAO algorithm in most of the scenarios. Could the authors provide additional numerical experiments to demonstrate the efficiency of DADAO? It would be helpful if numerical experiments in some special cases such as star or complete networks can suggest DADAO's superior performance.\n\nMinor Issues: There are some typos and some quantities are undefined. \n\n1.  In the abstract, the $O(n \\sqrt{\\frac{L}{\\mu}} \\log \\epsilon)$ gradient complexity is confusing. Should it be  $O(n \\sqrt{\\frac{L}{\\mu}} \\log \\frac{ 1}{ \\epsilon} )$ where $\\epsilon$ is the precision? \n\n2. In Table 1, $\\gamma$ is undefined so that a fair comparison between $\\sqrt{\\chi_1\\chi_2} n $ and $\\chi_1 |\\mathcal{E}|$ is not obvious. \n\n3. When defining $\\chi_2(t)$ in page 4, the quantity $\\Lambda^+(t)$ is undefined. \n\nReferences: \n[1] Kovalev, D., Shulgin, E., Richtárik, P., Rogozin, A., and Gasnikov, A. (2021). ADOM: Accelerated decentralized optimization method for time-varying networks. ", "clarity,_quality,_novelty_and_reproducibility": "The problem setting is clear but the writing is not very good because of typos and missing definitions. The work seems to be novel. \n\n\n\n", "summary_of_the_review": "This paper proposes an algorithm to solve decentralized strongly convex optimization and establishes some new complexity results. The results are developed for continuing-time frameworks, but the guarantees for implementable practical frameworks are absent. The theoretical claims are also only applicable to a restrictive class of problems. The overall contribution seems to be limited. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666669012071}, {"id": "zSnKW03ZOBo", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1632/Reviewer_R4MY"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposed a decentralized asynchronous stochastic first order algorithm to minimize a sum of L-smooth and µ-strongly convex functions distributed over a time-varying connectivity network of size n. The authors model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, to decouple the computation and communication steps. The proposed method is differ from the majority of works since it does not use a multi-consensus inner loop nor other ad-hoc mechanisms such as Error Feedback, Gradient Tracking, or a Proximal operator.\n", "review_text": "The authors model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, which differ from the majority of works. The experiment is too simple for verification. The convergence analysis is lacking some important factors.  ", "strengths": "strength: the proposed approach is differ from the majority of works since it model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, to decouple the computation and communication steps. \n\nweakness: \n- the experiment setting with decentralized linear and logistic regression tasks is too simple to validate the effectiveness of the approach. \n- many importance factors of convergence rate are not clear, such as communication latency and linear speedup ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work proposed a decentralized asynchronous stochastic first order algorithm to minimize a sum of L-smooth and µ-strongly convex functions distributed over a time-varying connectivity network of size n. The authors model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, to decouple the computation and communication steps. The proposed method is differ from the majority of works since it does not use a multi-consensus inner loop nor other ad-hoc mechanisms such as Error Feedback, Gradient Tracking, or a Proximal operator.\n", "strength_and_weaknesses": "strength: the proposed approach is differ from the majority of works since it model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, to decouple the computation and communication steps. \n\nweakness: \n- the experiment setting with decentralized linear and logistic regression tasks is too simple to validate the effectiveness of the approach. \n- many importance factors of convergence rate are not clear, such as communication latency and linear speedup ", "clarity,_quality,_novelty_and_reproducibility": "writing is not very clear to follow", "summary_of_the_review": "The authors model the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes, which differ from the majority of works. The experiment is too simple for verification. The convergence analysis is lacking some important factors.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666651975765}, {"id": "LvVR0lB3T10", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1632/Reviewer_c2yc"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studied asynchronous decentralized SGD for strongly convex problems. The authors provide theoretical analysis for the convergence rate and conduct experiments to verify its performance. But the writing is not very clear. It is not easy to follow.", "review_text": "This paper studied asynchronous decentralized SGD for strongly convex problems. The authors provide theoretical analysis for the convergence rate and conduct experiments to verify its performance. But the writing is not very clear. It is not easy to follow.", "strengths": "Strength:\n1. Theoretical convergence rate is established.\n2. Studying the convergence from the continuous perspective.\n\n\nWeaknesses:\n1. The convergence rate in the abstract is wrong. \n2. Why is it necessary to study the convergence rate from a continuous perspective? Is there any benefit? For strongly convex problems, the traditional analysis can also establish the linear convergence rate. Moreover, in the appendix, I didn't find the authors use continuous tools to investigate convergence. Why do you introduce Eq.(3) and that in Section 3.3?\n3. What's the meaning of $\\chi_1$ and $\\chi_2$? How do they relate to the spectral gap?\n4. How does the communication latency affect the convergence rate?\n5. How does the spectral gap affect the convergence rate?\n6. Can the convergence rate achieve the linear speedup with respect to the number of devices?\n7. The experiment is too simple. More complicated models and datasets should be used to evaluate the performance. \n8. Asynchronous decentralized SGD has been studied before. But the authors missed some important literature, e.g., http://proceedings.mlr.press/v80/lian18a.html ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper studied asynchronous decentralized SGD for strongly convex problems. The authors provide theoretical analysis for the convergence rate and conduct experiments to verify its performance. But the writing is not very clear. It is not easy to follow.", "strength_and_weaknesses": "Strength:\n1. Theoretical convergence rate is established.\n2. Studying the convergence from the continuous perspective.\n\n\nWeaknesses:\n1. The convergence rate in the abstract is wrong. \n2. Why is it necessary to study the convergence rate from a continuous perspective? Is there any benefit? For strongly convex problems, the traditional analysis can also establish the linear convergence rate. Moreover, in the appendix, I didn't find the authors use continuous tools to investigate convergence. Why do you introduce Eq.(3) and that in Section 3.3?\n3. What's the meaning of $\\chi_1$ and $\\chi_2$? How do they relate to the spectral gap?\n4. How does the communication latency affect the convergence rate?\n5. How does the spectral gap affect the convergence rate?\n6. Can the convergence rate achieve the linear speedup with respect to the number of devices?\n7. The experiment is too simple. More complicated models and datasets should be used to evaluate the performance. \n8. Asynchronous decentralized SGD has been studied before. But the authors missed some important literature, e.g., http://proceedings.mlr.press/v80/lian18a.html ", "clarity,_quality,_novelty_and_reproducibility": "Clarity: Poor\nQuality: Poor\nNovelty: Neutral\nReproducibility: N/A", "summary_of_the_review": "This paper studied asynchronous decentralized SGD for strongly convex problems. The authors provide theoretical analysis for the convergence rate and conduct experiments to verify its performance. But the writing is not very clear. It is not easy to follow.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666647266406}], "openreview_url": "https://openreview.net/forum?id=Siln8xpTMrZ", "arxiv_id": "2208.00779", "paper_pdf": "papers/Siln8xpTMrZ.pdf", "paper_pdf_sha256": "feda7224d95cb90a607cdf032996ec5681d8ac8977415b1947af41f155dada1c", "paper_pdf_bytes": 596633, "paper_pdf_source": "openreview", "code_url": "https://github.com/AdelNabli/DADAO", "code_repository": "AdelNabli/DADAO", "code_commit": "eed1ed1b21a66da393f55433c26f7920b465eabe", "code_archive": "repos/Siln8xpTMrZ.zip", "code_archive_sha256": "0a549ae8cafbdceebcee58dc6c2f6dcc9d08119b7ab40bfda0634e445003f1e1", "code_archive_bytes": 457184, "code_file_count": 8, "code_extensions": {".py": 7, ".ipynb": 1}, "github_disk_usage_kb": 464, "github_languages": {"Jupyter Notebook": 592835, "Python": 91733}, "github_archived": false, "github_pushed_at": "2023-10-07T00:04:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dadao-decoupled-accelerated-decentralized"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LT0KSFnQDWF", "year": 2021, "status": "rejected", "title": "Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting", "authors": ["Giorgos Bouritsas", "Fabrizio Frasca", "Stefanos Zafeiriou", "Michael M. Bronstein"], "authorids": ["~Giorgos_Bouritsas1", "~Fabrizio_Frasca1", "~Stefanos_Zafeiriou1", "~Michael_M._Bronstein1"], "authors_source": "OpenReview API", "abstract": "While Graph Neural Networks (GNNs) have achieved remarkable results in a variety of applications, recent studies exposed important shortcomings in their ability to capture the structure of the underlying graph. It has been shown that the expressive power of standard GNNs is bounded by the Weisfeiler-Lehman (WL) graph isomorphism test, from which they inherit proven limitations such as the inability to detect and count graph substructures. On the other hand, there is significant empirical evidence, e.g. in network science and bioinformatics, that substructures are often informative for downstream tasks, suggesting that it is desirable to design GNNs capable of leveraging this important source of information. To this end, we propose a novel topologically-aware message passing scheme based on substructure encoding. We show that our architecture allows incorporating domain-specific inductive biases and that it is strictly more expressive than the WL test. Importantly, in contrast to recent works on the expressivity of GNNs, we do not attempt to adhere to the WL hierarchy; this allows us to retain multiple attractive properties of standard GNNs such as locality and linear network complexity, while being able to disambiguate even hard instances of graph isomorphism. We extensively evaluate our method on graph classification and regression tasks and show state-of-the-art results on multiple datasets including molecular graphs and social networks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "aSz_hJF3SYo", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2153/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a natural extension of Message Passing Neural Net (MPNN)  by incorporating structural features. These structural features are computed as the counts from different substructures (like small lines, stars or complete graphs) induced in the original graph. These counts are combined to obtain a new feature per node or per edge. Then these features are used in a standard MPNN. The authors then show that the resulting GNN is more expressive and they validate this claim experimentally.\n\nThis idea is interesting and clearly explained in the paper but I think the paper could be greatly improved after addressing the following issues:\n \n1- the authors should clarify their position with regards to invariance. Indeed, as explained shortly on page 3 when commenting Loukas(2020), it is easy to make a GNN powerful if we remove the constraint to be invariant (or equivariant). Hence, as I understand it, the authors are proposing an algorithm that is equivariant. If this is the case, it would be great to have a clear formal statement that GSN are equivariant and to give a mathematical proof.\n\n2- the theoretical content of the paper should be improved:\n\n a- the first part of Proposition 3.1 is straightforward and I find the wording of the second statement unclear: what does '... or any not necessarily induced subgraph...' mean?\n\n b- from a theoretical perspective, it seems that both GSN-v abd GSN-e have the same expressive power. Is it true?\n\n c- a similar idea as the one presented in this paper was presented in : Coloring graph neural networks for node disambiguation  by George Dasoulas, Ludovic Dos Santos, Kevin Scaman, Aladin Virmaux https://arxiv.org/abs/1912.06058 [arxiv-col] The main advantage of the current paper as opposed to [arxiv-col] is to propose an explicit coloring thanks to the structural features. But the theoretical analysis made in [arxiv-col]  goes much deeper than this paper and probably could be adapted by the authors. For example, Corollary 3.1 could probably be replaced by a universality property, i.e. GSN with k=n-1 is universal.\n\n3- the experimental evaluation is not convincing. To make it more convincing, the authors should include an ablation study for all their experiments by comparing their performances with the performances obtained with the structural features only. Such an ablation study would show the benefit of adding the MPNN on top of these features. \n\n[After rebuttal] I think the authors improved their paper by taking into account the remarks. Given the last results obtained in Table 4, it looks like the structural features are indeed very good features in practice as they allow to boost the performances of a very simple invariant architecture like Deepset. I think the authors should explore how they can combine this approach with the coloring approach to get better GNN.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A natural idea to increase the expressive power of GNNs that would benefit from more theoretical results and more experimental evaluation", "review": "This paper presents a natural extension of Message Passing Neural Net (MPNN)  by incorporating structural features. These structural features are computed as the counts from different substructures (like small lines, stars or complete graphs) induced in the original graph. These counts are combined to obtain a new feature per node or per edge. Then these features are used in a standard MPNN. The authors then show that the resulting GNN is more expressive and they validate this claim experimentally.\n\nThis idea is interesting and clearly explained in the paper but I think the paper could be greatly improved after addressing the following issues:\n \n1- the authors should clarify their position with regards to invariance. Indeed, as explained shortly on page 3 when commenting Loukas(2020), it is easy to make a GNN powerful if we remove the constraint to be invariant (or equivariant). Hence, as I understand it, the authors are proposing an algorithm that is equivariant. If this is the case, it would be great to have a clear formal statement that GSN are equivariant and to give a mathematical proof.\n\n2- the theoretical content of the paper should be improved:\n\n a- the first part of Proposition 3.1 is straightforward and I find the wording of the second statement unclear: what does '... or any not necessarily induced subgraph...' mean?\n\n b- from a theoretical perspective, it seems that both GSN-v abd GSN-e have the same expressive power. Is it true?\n\n c- a similar idea as the one presented in this paper was presented in : Coloring graph neural networks for node disambiguation  by George Dasoulas, Ludovic Dos Santos, Kevin Scaman, Aladin Virmaux https://arxiv.org/abs/1912.06058 [arxiv-col] The main advantage of the current paper as opposed to [arxiv-col] is to propose an explicit coloring thanks to the structural features. But the theoretical analysis made in [arxiv-col]  goes much deeper than this paper and probably could be adapted by the authors. For example, Corollary 3.1 could probably be replaced by a universality property, i.e. GSN with k=n-1 is universal.\n\n3- the experimental evaluation is not convincing. To make it more convincing, the authors should include an ablation study for all their experiments by comparing their performances with the performances obtained with the structural features only. Such an ablation study would show the benefit of adding the MPNN on top of these features. \n\n[After rebuttal] I think the authors improved their paper by taking into account the remarks. Given the last results obtained in Table 4, it looks like the structural features are indeed very good features in practice as they allow to boost the performances of a very simple invariant architecture like Deepset. I think the authors should explore how they can combine this approach with the coloring approach to get better GNN.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604329769675}, {"id": "uEp9z7a5qCr", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2153/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper studies the expressivity of graph neural networks, and proposes a new approach to improve GNN’s expressivity by encoding nodes and edges with features via subgraph isomorphism counting. The proposed solution contains some merit, and the experimental results on graph classification task demonstrates the superiority of the proposed approach.\n\nPros:\n1.\tThis paper addresses an important problem in GNNs, which is to improve the expressivity of GNNs.\n2.\tThe proposed solution is interesting, which is to include how many isomorphic subgraphs from a given list a node or an edge is contained in as additional features.\n3.\tThe experimental results on multiple datasets and different graph-level tasks are better than baselines.\n4.\t Nice theoretical analysis.\n\nCons:\n1.\tMy biggest concern lies in the time complexity of the proposed approach. Although the paper claims that in practice it is not that bad, the worst time complexity is still high. Also, the substructure selection brings us back to feature engineering, or we will face too many possible substructures.\n2.\tMore challenging graph tasks are expected to demonstrate the necessity of the proposed approach. \n\nDetailed comments:\n1.\tIn the abstract and introduction, it is mentioned that existing GNNs are bounded by WL-test, and are not able to detect and count graph structures. It is expected to see experiments are on these more challenging tasks, in addition to graph classification and regression tasks. Graph isomorphism test is an interesting task, and the design is smart.\n2.\tDiscussions on how to select substructures on bigger size of graphs are expected.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper studies the expressivity of graph neural networks, and proposes a new approach to improve GNN’s expressivity by encoding nodes and edges with features via subgraph isomorphism counting. The proposed solution contains some merit, and the experimental results on graph classification task demonstrates the superiority of the proposed approach.", "review": "This paper studies the expressivity of graph neural networks, and proposes a new approach to improve GNN’s expressivity by encoding nodes and edges with features via subgraph isomorphism counting. The proposed solution contains some merit, and the experimental results on graph classification task demonstrates the superiority of the proposed approach.\n\nPros:\n1.\tThis paper addresses an important problem in GNNs, which is to improve the expressivity of GNNs.\n2.\tThe proposed solution is interesting, which is to include how many isomorphic subgraphs from a given list a node or an edge is contained in as additional features.\n3.\tThe experimental results on multiple datasets and different graph-level tasks are better than baselines.\n4.\t Nice theoretical analysis.\n\nCons:\n1.\tMy biggest concern lies in the time complexity of the proposed approach. Although the paper claims that in practice it is not that bad, the worst time complexity is still high. Also, the substructure selection brings us back to feature engineering, or we will face too many possible substructures.\n2.\tMore challenging graph tasks are expected to demonstrate the necessity of the proposed approach. \n\nDetailed comments:\n1.\tIn the abstract and introduction, it is mentioned that existing GNNs are bounded by WL-test, and are not able to detect and count graph structures. It is expected to see experiments are on these more challenging tasks, in addition to graph classification and regression tasks. Graph isomorphism test is an interesting task, and the design is smart.\n2.\tDiscussions on how to select substructures on bigger size of graphs are expected.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604036076382}, {"id": "ZeysRS8k6eO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2153/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Reasons for score:\nThe idea of using small graphs to characterize local topologies and guide message passing in interesting. However, the graph isomorphism computation part has problem. Experimental results are not conclusive. The written need improvement in some part of the manuscript.  \n\nPros: A new method (GSN: Graph Substructure Network) is proposed to a topologically-aware message passing method that better utilize graph substructure information. The method tries to tackles the limitation of traditional GNN in exploring graph structure.  It is a good idea to pass messages differently depending on their local topologies. This is done through using a set of predefine small graphs to characterize local topologies. The authors showed that GSN was more expressive than traditional GNNs. A good number of experimental evaluations were performed.\n \nCons: It is not clear priorly how to define a good set of small graphs, especially when considering beyond immediate neighbors. In addition, node features are not considered in graph isomorphism, which can lead to incorrect subgraph matching. The experimental results on some of the datasets (such as, MUTAG, PTC, Proteins and NCI1 in table 1) do not appear to be significantly better than those of the previous approaches when considering the variances of different runs. In addition, much better results were reported on the ogb-molhiv leaderboard (https://ogb.stanford.edu/docs/leader_graphprop/#ogbg-molhiv). Figure 1 is confusing.  Should the number in the yellow square on the left be 5? A more self-contained explanation of the figure is appreciated. It will help readers if the authors can visualize a few examples (e.g., contributions of small graphs) to explain why their approach works better.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Propose a new method (Graph Substructure Network) for topologically-aware message passing.", "review": "Reasons for score:\nThe idea of using small graphs to characterize local topologies and guide message passing in interesting. However, the graph isomorphism computation part has problem. Experimental results are not conclusive. The written need improvement in some part of the manuscript.  \n\nPros: A new method (GSN: Graph Substructure Network) is proposed to a topologically-aware message passing method that better utilize graph substructure information. The method tries to tackles the limitation of traditional GNN in exploring graph structure.  It is a good idea to pass messages differently depending on their local topologies. This is done through using a set of predefine small graphs to characterize local topologies. The authors showed that GSN was more expressive than traditional GNNs. A good number of experimental evaluations were performed.\n \nCons: It is not clear priorly how to define a good set of small graphs, especially when considering beyond immediate neighbors. In addition, node features are not considered in graph isomorphism, which can lead to incorrect subgraph matching. The experimental results on some of the datasets (such as, MUTAG, PTC, Proteins and NCI1 in table 1) do not appear to be significantly better than those of the previous approaches when considering the variances of different runs. In addition, much better results were reported on the ogb-molhiv leaderboard (https://ogb.stanford.edu/docs/leader_graphprop/#ogbg-molhiv). Figure 1 is confusing.  Should the number in the yellow square on the left be 5? A more self-contained explanation of the figure is appreciated. It will help readers if the authors can visualize a few examples (e.g., contributions of small graphs) to explain why their approach works better.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603932837484}, {"id": "TFtZY5xnTh_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2153/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes the Graph Substructure Network (GSN) to encode structural roles for different nodes so that the expressivity of Graph Neural Networks is improved. The core idea is to count the number of certain substructures, such as cycles, cliques, and triangles. Then the proposed MPNN encodes such substructure counting information into the message passing. Experimental results show that the proposed method can obtain better performance than the comparing methods. \n\nStrengths:\n+ The proposed method can encode important substructure information. It is important for graphs since, in many applications, the substructures can determine the functionality of graphs.\n \n+ The proposed method leads to better performance on different graph classification datasets. The experimental results can show the effectiveness of the proposed method.\n\nWeaknesses:\n- The main contribution of this work is counting the substructure information, which can be regarded as the preprocessing of graphs. However, how to properly select the graph set {H_1, … ,H_K}?  With different datasets at hand, the best choice can be quite different. Then how should we apply the proposed method? \n\n- The second concern is the complexity of the proposed methods. Counting different types of substructures can be very time-consuming. As mentioned in this work, the worst case can have O(n^k) complexity. When the graph size is large, we may need to count too many types of substructures. \n\n- The datasets are relatively small; most of them have less than 100 nodes per graph. Then larger datasets, such as RDT-B, RDT-M5K, and RDT-M12K, should be considered. I am wondering how many types of substructures need to be considered for these larger datasets to get better performance. \n\n- This work explicitly encodes substructure information into GNNs. Other existing methods, such as graph pooling methods, which can be considered to implicitly encode structural information. Then advanced pooling methods, such as Diffpool, Structpool, Min-cut Pool, etc., should be discussed and compared. \n\nI am willing to adjust my score if my concerns are properly addressed. \n\n=====Update after rebuttal=====\n\nI have read the authors' rebuttal. Considering the limitations and non-superior performance for larger datasets, I am keeping my score unchanged. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "This work proposes the Graph Substructure Network (GSN) to encode structural roles for different nodes so that the expressivity of Graph Neural Networks is improved. The core idea is to count the number of certain substructures, such as cycles, cliques, and triangles. Then the proposed MPNN encodes such substructure counting information into the message passing. Experimental results show that the proposed method can obtain better performance than the comparing methods. \n\nStrengths:\n+ The proposed method can encode important substructure information. It is important for graphs since, in many applications, the substructures can determine the functionality of graphs.\n \n+ The proposed method leads to better performance on different graph classification datasets. The experimental results can show the effectiveness of the proposed method.\n\nWeaknesses:\n- The main contribution of this work is counting the substructure information, which can be regarded as the preprocessing of graphs. However, how to properly select the graph set {H_1, … ,H_K}?  With different datasets at hand, the best choice can be quite different. Then how should we apply the proposed method? \n\n- The second concern is the complexity of the proposed methods. Counting different types of substructures can be very time-consuming. As mentioned in this work, the worst case can have O(n^k) complexity. When the graph size is large, we may need to count too many types of substructures. \n\n- The datasets are relatively small; most of them have less than 100 nodes per graph. Then larger datasets, such as RDT-B, RDT-M5K, and RDT-M12K, should be considered. I am wondering how many types of substructures need to be considered for these larger datasets to get better performance. \n\n- This work explicitly encodes substructure information into GNNs. Other existing methods, such as graph pooling methods, which can be considered to implicitly encode structural information. Then advanced pooling methods, such as Diffpool, Structpool, Min-cut Pool, etc., should be discussed and compared. \n\nI am willing to adjust my score if my concerns are properly addressed. \n\n=====Update after rebuttal=====\n\nI have read the authors' rebuttal. Considering the limitations and non-superior performance for larger datasets, I am keeping my score unchanged. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603676960472}], "openreview_url": "https://openreview.net/forum?id=LT0KSFnQDWF", "arxiv_id": "2006.09252", "paper_pdf": "papers/LT0KSFnQDWF.pdf", "paper_pdf_sha256": "0690ee61df32cdcc835a1086e5195188358cfdfe10df993c3843ebeda7ade2f9", "paper_pdf_bytes": 1494478, "paper_pdf_source": "openreview", "code_url": "https://github.com/gbouritsas/GSN", "code_repository": "gbouritsas/GSN", "code_commit": "6cce24a2c0f59c183c388f3016d33502f63e8175", "code_archive": "repos/LT0KSFnQDWF.zip", "code_archive_sha256": "2fa53aa332338f5c639095025bbf449ab1f5711597b48f96f66373f90a809a66", "code_archive_bytes": 1190892, "code_file_count": 39, "code_extensions": {".py": 38, ".sh": 1}, "github_disk_usage_kb": 1228, "github_languages": {"Python": 260611, "Shell": 596}, "github_archived": false, "github_pushed_at": "2021-06-15T14:59:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-graph-neural-network-expressivity"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1e4Q6EtDH", "year": 2020, "status": "rejected", "title": "Tensorized Embedding Layers for Efficient Model Compression", "authors": ["Oleksii Hrinchuk", "Valentin Khrulkov", "Leyla Mirvakhabova", "Ivan Oseledets"], "authorids": ["oleksii.hrinchuk@skoltech.ru", "khrulkov.v@gmail.com", "leyla.mirvakhabova@skoltech.ru", "i.oseledets@skoltech.ru"], "authors_source": "OpenReview API", "abstract": "The embedding layers transforming input words into real vectors are the key components of deep neural networks used in natural language processing. However, when the vocabulary is large, the corresponding weight matrices can be enormous, which precludes their deployment in a limited resource setting. We introduce a novel way of parametrizing embedding layers based on the Tensor Train (TT) decomposition, which allows compressing the model significantly at the cost of a negligible drop or even a slight gain in performance.  We evaluate our method on a wide range of benchmarks in natural language processing and analyze the trade-off between performance and compression ratios for a wide range of architectures, from MLPs to LSTMs and Transformers.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJe4_DGZ9S", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper445/AnonReviewer3"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to use TensorTrain representation to transform discrete tokens/symbols to its vector representation.\nSince neural networks can only work with numerical numbers, in many NLP tasks, where the raw inputs are in the discrete token/symbol format, the popular technique is to use \"embedding\" matrices to find a vector representation of those inputs. \n\nAs the authors point out, the embedding matrices usually require huge number of parameters, since it assigns one vector for each input token for one embedding vector, but to attain a competitive performance in the real world applications, we need to use large number of embedding vectors, which results in a large number of parameters in the neural networks.\n\nThe paper assumes that those embedding matrices can be compressed by assuming that the low-rank property of embedding matrices. I think this is a valid assumption in many cases, and the paper shows the performance degradation according to this assumption is relatively small compared to the gain, a dramatically reduced size of parameters in the embedding stage, is substantial.\n\nI think the paper is well written and proposes a new direction to find a memory efficient representation of symbols. I am not sure the current initialization techniques, nor the training method in the paper are the right way to train a tensor train \"embedding\" but I expect that the authors would perform the follow up work on those topics.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper proposes to use TensorTrain representation to transform discrete tokens/symbols to its vector representation.\nSince neural networks can only work with numerical numbers, in many NLP tasks, where the raw inputs are in the discrete token/symbol format, the popular technique is to use \"embedding\" matrices to find a vector representation of those inputs. \n\nAs the authors point out, the embedding matrices usually require huge number of parameters, since it assigns one vector for each input token for one embedding vector, but to attain a competitive performance in the real world applications, we need to use large number of embedding vectors, which results in a large number of parameters in the neural networks.\n\nThe paper assumes that those embedding matrices can be compressed by assuming that the low-rank property of embedding matrices. I think this is a valid assumption in many cases, and the paper shows the performance degradation according to this assumption is relatively small compared to the gain, a dramatically reduced size of parameters in the embedding stage, is substantial.\n\nI think the paper is well written and proposes a new direction to find a memory efficient representation of symbols. I am not sure the current initialization techniques, nor the training method in the paper are the right way to train a tensor train \"embedding\" but I expect that the authors would perform the follow up work on those topics."}, "tcdate": 1572050796501}, {"id": "Syxic2OTYr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper445/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduces a novel way of parametrizing embedding layers based on the Tensor Train (TT)\ndecomposition, which allows compressing the model significantly at the cost of a negligible drop or even a slight gain in performance. And this paper focuses on the input embedding layers. \n\nFor the experiments, the paper just compared methods using TT layer and normal embedding layer. There are many other methods that has been proposed to compress the embedding layers, it will be good to compare with one or two other methods, such as WEST or compression based on projection layers. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper introduces a novel way of parametrizing embedding layers based on the Tensor Train (TT)\ndecomposition, which allows compressing the model significantly at the cost of a negligible drop or even a slight gain in performance. And this paper focuses on the input embedding layers. \n\nFor the experiments, the paper just compared methods using TT layer and normal embedding layer. There are many other methods that has been proposed to compress the embedding layers, it will be good to compare with one or two other methods, such as WEST or compression based on projection layers. \n\n"}, "tcdate": 1571814547455}, {"id": "BklnwAZF_r", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper445/AnonReviewer2"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a low-rank tensor decomposition model (Tensor Train-TT [Oseledets et al, 2011]) to parameterize the embedding matrix in Natural Language Processing (NLP). It shows that TT allows for a compression of the network and sometimes even a slight increase of test accuracy. The paper is well written and easy to follow.\nI found the idea as a natural consequence of many recent papers proposing tensor decomposition to parameterize deep learning networks. However, I think this is the first time that the concept has been applied to learning an embedding matrix, which is an important problem in the field.\nThe authors reported several experimental results on different tasks and datasets for NLP such as: Sentiment Analysis, Neural Machine Translation and Language Modeling; and an application to the click through rate prediction problem.\nI think the paper is of limited novelty but includes interesting experimental results that helps to better understand the potential and limitations of tensor decompositions in deep learning architectures.\nBelow, I summarize the issues I found, and I would like the authors to address them in their responses:\nMajor issues:\n-\tIn Page 4 (and Appendix B), the authors show a comparison with Tensor Ring (TR) and conclude that TT marginally outperforms TR in terms of BLEU measure for a fixed number of parameters in both models. I found this comparison incomplete, weak and misleading because of the following reasons:\no\tTR is a more general model including TT as a particular case when the first and last ranks are equal to one (Zhao et al, 2016). In fact, in this experiment, the authors chose all intermediate ranks set at the same value R with the first/last ranks set to 1 and R for TT and TR, respectively, which is not a fair comparison. Shown results suggest that first/last ranks contain less information than intermediate ranks, but it would be necessary to explore other combinations of rank values without keeping them constant to explore the generalization power of the TR model including TT as a particular case.\no\tThe authors compares TT and TR only in case of the NMT, Transformer-big on WMT‘14 English-to-German dataset, where the results are not good for TT and TR. It is noted that baseline model (Big) attains a Sacre BLEU = 28.84, TT1 = 28.53 and TR1 = 28.07 and the compression rate is only 210/179=1.17 for TT1 and TR1 and the Iteration time is larger than in the baseline model. In this case, there is no a clear advantage of using TT or TR.\nIn my opinion, to improve the paper, I think the authors could:\no\tTo avoid the sentence “In our experiments, however, the resulting TR embeddings performed slightly worse than TT–embeddings with the same number of parameters” unless more conclusive and exhaustive experiments are performed comparing TT and TR.\no\tTo add some comparison results between TT and TR for the rest of datasets such as Sentiment Analysis, Language Modeling and the click through rate prediction problem. \no\tHighlight that TT is a particular case of TR so considering the first and last ranks equal to one reduce the number of parameters but can affect the generalization power of the model.\n-\tThe approach of the paper is mostly intuitive. A theoretical result about why the low rank TT is able to catch the useful information of an optimal or suboptimal embedding matrix is missing.\nMinor issues:\n-\tIn last paragraph of section 3.2: The number of parameters is computed on the 3D tensor cores only. I think the size of the first/last 2D cores should be added. Please revise the equation.\n-\tThe pseudocode for the mapping one index to multiple indices is trivial and could be avoided. If it is kept, I think the reverse operation should be also included, i.e. how to map multiple indices i1, …., iN to one index i.\n-\tThe discussion and Figure 2 about the Gaussianity of the values in the higher order tensor based on Gaussian core tensors is not relevant. Maybe, the authors should better motivate why it is important to highlight that the distribution tends to a Gaussian density for increasing ranks.\n-\tIn page 5, it is mentioned that “factors should be as close to each other as possible” but there is no a justification for it. Could you give some theoretical insight on why it is important to obtain uniform distribution of matrix size?\n-\tSection 4.1, reference to the Stanford sentiment treebank (SST) is missing.\n\nOn Nov 16th: I am satisfied with the responses provided by the Authors who made few changes to solve some identified minor issues. Thanks for taking the review report into account. I have raised the rating. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "8: Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #2", "review": "This paper proposes a low-rank tensor decomposition model (Tensor Train-TT [Oseledets et al, 2011]) to parameterize the embedding matrix in Natural Language Processing (NLP). It shows that TT allows for a compression of the network and sometimes even a slight increase of test accuracy. The paper is well written and easy to follow.\nI found the idea as a natural consequence of many recent papers proposing tensor decomposition to parameterize deep learning networks. However, I think this is the first time that the concept has been applied to learning an embedding matrix, which is an important problem in the field.\nThe authors reported several experimental results on different tasks and datasets for NLP such as: Sentiment Analysis, Neural Machine Translation and Language Modeling; and an application to the click through rate prediction problem.\nI think the paper is of limited novelty but includes interesting experimental results that helps to better understand the potential and limitations of tensor decompositions in deep learning architectures.\nBelow, I summarize the issues I found, and I would like the authors to address them in their responses:\nMajor issues:\n-\tIn Page 4 (and Appendix B), the authors show a comparison with Tensor Ring (TR) and conclude that TT marginally outperforms TR in terms of BLEU measure for a fixed number of parameters in both models. I found this comparison incomplete, weak and misleading because of the following reasons:\no\tTR is a more general model including TT as a particular case when the first and last ranks are equal to one (Zhao et al, 2016). In fact, in this experiment, the authors chose all intermediate ranks set at the same value R with the first/last ranks set to 1 and R for TT and TR, respectively, which is not a fair comparison. Shown results suggest that first/last ranks contain less information than intermediate ranks, but it would be necessary to explore other combinations of rank values without keeping them constant to explore the generalization power of the TR model including TT as a particular case.\no\tThe authors compares TT and TR only in case of the NMT, Transformer-big on WMT‘14 English-to-German dataset, where the results are not good for TT and TR. It is noted that baseline model (Big) attains a Sacre BLEU = 28.84, TT1 = 28.53 and TR1 = 28.07 and the compression rate is only 210/179=1.17 for TT1 and TR1 and the Iteration time is larger than in the baseline model. In this case, there is no a clear advantage of using TT or TR.\nIn my opinion, to improve the paper, I think the authors could:\no\tTo avoid the sentence “In our experiments, however, the resulting TR embeddings performed slightly worse than TT–embeddings with the same number of parameters” unless more conclusive and exhaustive experiments are performed comparing TT and TR.\no\tTo add some comparison results between TT and TR for the rest of datasets such as Sentiment Analysis, Language Modeling and the click through rate prediction problem. \no\tHighlight that TT is a particular case of TR so considering the first and last ranks equal to one reduce the number of parameters but can affect the generalization power of the model.\n-\tThe approach of the paper is mostly intuitive. A theoretical result about why the low rank TT is able to catch the useful information of an optimal or suboptimal embedding matrix is missing.\nMinor issues:\n-\tIn last paragraph of section 3.2: The number of parameters is computed on the 3D tensor cores only. I think the size of the first/last 2D cores should be added. Please revise the equation.\n-\tThe pseudocode for the mapping one index to multiple indices is trivial and could be avoided. If it is kept, I think the reverse operation should be also included, i.e. how to map multiple indices i1, …., iN to one index i.\n-\tThe discussion and Figure 2 about the Gaussianity of the values in the higher order tensor based on Gaussian core tensors is not relevant. Maybe, the authors should better motivate why it is important to highlight that the distribution tends to a Gaussian density for increasing ranks.\n-\tIn page 5, it is mentioned that “factors should be as close to each other as possible” but there is no a justification for it. Could you give some theoretical insight on why it is important to obtain uniform distribution of matrix size?\n-\tSection 4.1, reference to the Stanford sentiment treebank (SST) is missing.\n\nOn Nov 16th: I am satisfied with the responses provided by the Authors who made few changes to solve some identified minor issues. Thanks for taking the review report into account. I have raised the rating. \n", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory."}, "tcdate": 1570475620130}], "openreview_url": "https://openreview.net/forum?id=S1e4Q6EtDH", "arxiv_id": "1901.10787", "paper_pdf": "papers/S1e4Q6EtDH.pdf", "paper_pdf_sha256": "a0d2097f59847dd4dbcd2ce5edcc31f62883ca79f4a3e62b586b7ebf2d7f4412", "paper_pdf_bytes": 1314343, "paper_pdf_source": "openreview", "code_url": "https://github.com/KhrulkovV/tt-pytorch", "code_repository": "KhrulkovV/tt-pytorch", "code_commit": "3182e8d90fb4c4473b66105159e54486562b6e33", "code_archive": "repos/S1e4Q6EtDH.zip", "code_archive_sha256": "a089dd475459ae098edbf1e32bbbb5e1d643319ddd57c56fe7fdae9ab94b801c", "code_archive_bytes": 21528, "code_file_count": 14, "code_extensions": {".py": 13, ".sh": 1}, "github_disk_usage_kb": 8765, "github_languages": {"Python": 82029, "Shell": 759}, "github_archived": false, "github_pushed_at": "2020-07-06T07:54:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tensorized-embedding-layers-for-efficient"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xg2ccsYpZY", "year": 2026, "status": "rejected", "title": "SIM-Shapley: A Stable and Computationally Efficient Approach to Shapley Value Approximation", "authors": ["Wangxuan Fan", "Siqi Li", "Doudou Zhou", "Yohei Okada", "Chuan Hong", "Molei Liu", "Nan Liu"], "authorids": ["~Wangxuan_Fan1", "~Siqi_Li4", "~Doudou_Zhou1", "~Yohei_Okada1", "~Chuan_Hong1", "~Molei_Liu1", "~Nan_Liu3"], "authors_source": "OpenReview API", "abstract": "Explainable artificial intelligence (XAI) is essential for trustworthy  machine learning (ML), particularly in high-stakes domains such as healthcare and finance. Shapley value (SV) methods provide a principled framework for feature attribution in complex  models but incur high computational costs, limiting their scalability in high-dimensional settings. We propose Stochastic IterativeMomentum for Shapley Value Approximation (SIM-Shapley), a stable and efficient SV approximation method inspired by stochastic optimization. We analyze variance theoretically, prove linear $Q$-convergence, and demonstrate improved empirical stability and low bias in practice on real-world datasets.\nIn our numerical experiments, SIM-Shapley reduces computation time by up to 85\\% relative to state-of-the-art baselines while maintaining comparable feature attribution quality. Beyond feature attribution, our stochastic mini-batch iterative framework extends naturally to a broader class of sample average approximation problems, offering a new avenue for improving computational efficiency with stability guarantees. Code is publicly available at https://anonymous.4open.science/r/SIM-Shapley.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zefcEfu1Nk", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11722/Reviewer_9Y6m"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "This paper proposes SIM-Shapley, a Shapley value approximation method based on stochastic iterative momentum. Its core idea is to reformulate Shapley value computation as a stochastic optimization problem. The key innovations include: (1) achieving linear Q-convergence through Exponential Moving Average (EMA)-based updates and adaptive mini-batch sampling, with a variance contraction rate of $(1-t)^2$; (2) introducing three stability mechanisms: $\\ell_2$ regularization, negative-sampling detection, and initialization bias correction; (3) supporting both local and global explanation modes while being game-agnostic and imputer-agnostic. Experiments show that SIM-Shapley is 60%-85% faster than baselines such as KernelSHAP and SAGE across classification, regression, image, and clinical tasks, with a 10%-15% reduction in bias.", "review_text": "This paper proposes SIM-Shapley, a Shapley value approximation method based on stochastic iterative momentum. Its core idea is to reformulate Shapley value computation as a stochastic optimization problem. The key innovations include: (1) achieving linear Q-convergence through Exponential Moving Average (EMA)-based updates and adaptive mini-batch sampling, with a variance contraction rate of $(1-t)^2$; (2) introducing three stability mechanisms: $\\ell_2$ regularization, negative-sampling detection, and initialization bias correction; (3) supporting both local and global explanation modes while being game-agnostic and imputer-agnostic. Experiments show that SIM-Shapley is 60%-85% faster than baselines such as KernelSHAP and SAGE across classification, regression, image, and clinical tasks, with a 10%-15% reduction in bias.", "strengths": "1. The paper is well-written, with a clear structure that allows readers to easily follow the authors' reasoning.\n2. The comparison between SIM-Shapley and other Shapley value computation methods is thorough and clear, enabling readers to readily grasp the paper's contributions.\n3. Relevant experiments effectively demonstrate the superiority of the proposed method, particularly in accelerating Shapley value computation.\n4. The paper addresses interpretability, a critical issue in deep learning models. By focusing on accelerating Shapley value computation, this work holds significant importance—it facilitates the adoption of deep learning models in high-reliability industries.", "weaknesses": "The paper is relatively comprehensive, and there are no major fundamental issues. However, two minor points require attention:\n1. **Paper Formatting:** The authors are requested to review the paper's formatting. For instance, on Page 21 of the supplementary materials, some figures/icons clearly exceed the paper's margins.\n2. **Evaluation Metrics:** The paper primarily uses the error between estimated Shapley values and the ground truth to evaluate performance. Readers are curious about how the proposed method performs under a broader set of evaluation metrics, such as Deletion/Insertion and mu-fidelity.", "questions": "Refer to the \"Weaknesses\" section above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes SIM-Shapley, a Shapley value approximation method based on stochastic iterative momentum. Its core idea is to reformulate Shapley value computation as a stochastic optimization problem. The key innovations include: (1) achieving linear Q-convergence through Exponential Moving Average (EMA)-based updates and adaptive mini-batch sampling, with a variance contraction rate of $(1-t)^2$; (2) introducing three stability mechanisms: $\\ell_2$ regularization, negative-sampling detection, and initialization bias correction; (3) supporting both local and global explanation modes while being game-agnostic and imputer-agnostic. Experiments show that SIM-Shapley is 60%-85% faster than baselines such as KernelSHAP and SAGE across classification, regression, image, and clinical tasks, with a 10%-15% reduction in bias.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper is well-written, with a clear structure that allows readers to easily follow the authors' reasoning.\n2. The comparison between SIM-Shapley and other Shapley value computation methods is thorough and clear, enabling readers to readily grasp the paper's contributions.\n3. Relevant experiments effectively demonstrate the superiority of the proposed method, particularly in accelerating Shapley value computation.\n4. The paper addresses interpretability, a critical issue in deep learning models. By focusing on accelerating Shapley value computation, this work holds significant importance—it facilitates the adoption of deep learning models in high-reliability industries.", "weaknesses": "The paper is relatively comprehensive, and there are no major fundamental issues. However, two minor points require attention:\n1. **Paper Formatting:** The authors are requested to review the paper's formatting. For instance, on Page 21 of the supplementary materials, some figures/icons clearly exceed the paper's margins.\n2. **Evaluation Metrics:** The paper primarily uses the error between estimated Shapley values and the ground truth to evaluate performance. Readers are curious about how the proposed method performs under a broader set of evaluation metrics, such as Deletion/Insertion and mu-fidelity.", "questions": "Refer to the \"Weaknesses\" section above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761991998886}, {"id": "JacuuD3b2E", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11722/Reviewer_LoUX"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper introduces SIM-Shapley (Stochastic Iterative Momentum for Shapley values), a new method for approximating Shapley values, which are a key tool for feature attribution in explainable AI (XAI). The primary challenge with Shapley values is their exponential computational complexity, making them impractical for many real-world applications. The authors propose SIM-Shapley to address this by reformulating the Shapley value calculation as a stochastic optimization problem.", "review_text": "This paper introduces SIM-Shapley (Stochastic Iterative Momentum for Shapley values), a new method for approximating Shapley values, which are a key tool for feature attribution in explainable AI (XAI). The primary challenge with Shapley values is their exponential computational complexity, making them impractical for many real-world applications. The authors propose SIM-Shapley to address this by reformulating the Shapley value calculation as a stochastic optimization problem.", "strengths": "1.\tThe proposed method covers both local and global explanations.\n2.\tClear reformulation of KernelSHAP as a constrained stochastic optimization with a simple EMA update and a closed-form per-iteration solution. \n3.\tThe method stays model-agnostic.", "weaknesses": "1.\tThe authors claim that their method fundamentally re-conceptualizes SV computation. However, the core of their process is a series of mature stochastic optimization techniques.\n2.\tThe early stopping method (Eq.11) is heuristic and has no sensitivity analysis to the parameter epsilon.\n3.\tThe choice of the parameter xi in Eq. 12 is crude and lacks analysis.\n4.\tThe authors claim that the method can be used for Shapley Interactions, but provide no experiments to demonstrate this.", "questions": "1.\tLine 159: The constraint in equation 5 should preserve the efficiency property of SV.\n2.\tThe main convergence theorem targets the SIM-Shapley fixed point $\\beta^*$ that depends on $\\lambda$, not the unregularized KernelSHAP target $\\beta$. Hence, the theory guarantees a fast approach to $\\beta^*$rather than a fast approach to $\\beta$. Figure 1 suggests a small bias for small λ, but there is no explicit bound for $‖beta^*−\\beta‖$ or a rate as $\\lambda \\to 0$.\n3.\tThe proof of Theorem 1(Appendix B.1) seems to rely on a key assumption $mathrm{Var}(\\delta^{(j)})\\le \\mathrm{Var}(\\beta)$, which is not always right.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SIM-Shapley (Stochastic Iterative Momentum for Shapley values), a new method for approximating Shapley values, which are a key tool for feature attribution in explainable AI (XAI). The primary challenge with Shapley values is their exponential computational complexity, making them impractical for many real-world applications. The authors propose SIM-Shapley to address this by reformulating the Shapley value calculation as a stochastic optimization problem.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1.\tThe proposed method covers both local and global explanations.\n2.\tClear reformulation of KernelSHAP as a constrained stochastic optimization with a simple EMA update and a closed-form per-iteration solution. \n3.\tThe method stays model-agnostic.", "weaknesses": "1.\tThe authors claim that their method fundamentally re-conceptualizes SV computation. However, the core of their process is a series of mature stochastic optimization techniques.\n2.\tThe early stopping method (Eq.11) is heuristic and has no sensitivity analysis to the parameter epsilon.\n3.\tThe choice of the parameter xi in Eq. 12 is crude and lacks analysis.\n4.\tThe authors claim that the method can be used for Shapley Interactions, but provide no experiments to demonstrate this.", "questions": "1.\tLine 159: The constraint in equation 5 should preserve the efficiency property of SV.\n2.\tThe main convergence theorem targets the SIM-Shapley fixed point $\\beta^*$ that depends on $\\lambda$, not the unregularized KernelSHAP target $\\beta$. Hence, the theory guarantees a fast approach to $\\beta^*$rather than a fast approach to $\\beta$. Figure 1 suggests a small bias for small λ, but there is no explicit bound for $‖beta^*−\\beta‖$ or a rate as $\\lambda \\to 0$.\n3.\tThe proof of Theorem 1(Appendix B.1) seems to rely on a key assumption $mathrm{Var}(\\delta^{(j)})\\le \\mathrm{Var}(\\beta)$, which is not always right.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761811708433}, {"id": "n3al34L6ST", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11722/Reviewer_Jdo3"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper considers the problem of estimating Shapley values of a general value function $v$. They propose a linear regression-based method similar to KernelSHAP. Instead of solving the weighted and constrained least squares problem once, they iteratively solve a sequence of regression problems. When solving the next regression problem, they draw $m$ new samples and add a regularization term on the \"moving average\" solution of the regression problems so far.", "review_text": "This paper considers the problem of estimating Shapley values of a general value function $v$. They propose a linear regression-based method similar to KernelSHAP. Instead of solving the weighted and constrained least squares problem once, they iteratively solve a sequence of regression problems. When solving the next regression problem, they draw $m$ new samples and add a regularization term on the \"moving average\" solution of the regression problems so far.", "strengths": "N/A", "weaknesses": "Big:\n\n1. I think the biggest structural issue in this paper is how they discuss and compare to prior work. Across the discussions and experiments, they consider different subsets of the estimators FastSHAP, SimSHAP, LeverageSHAP, and KernelSHAP. The first issue is that they selectively compare to these algorithms e.g., they'll sometimes ignore LeverageSHAP or FastSHAP depending on the experiment. The bigger issue is that they don't consider other SOTA estimators which include PermutationSHAP, and (two recent methods) RegressionMSR and ProxySPEX.\n\n2. The paper ignores fundamental properties of estimators, or says things that are just incorrect about them:\n    * They claim prior estimators can't be used for general value functions $v$, but, except for methods like TreeSHAP (which isn't mentioned in the paper), this is just false. Basically all near-SOTA Shapley value estimators are agnostic to the value function. (I think they know this e.g., they present a \"variant\" of their estimator for global explanations in Appendix C but the only change is one line about how the value function $v$ is computed.) It is true that *implementations* of these algorithms are generally tied to specific value functions but, I would say a) the onus for modifying the standard implementation is on researchers if they want experiments on a new kind of value function, and b) the shap-iq library is extracting the value function logic to an arbitrary class that makes it easy to switch which value function is used.\n    * They treat KernelSHAP and LeverageSHAP as different algorithms. In reality, they're the same algorithm except a) the sampling distribution is different, and b) LeverageSHAP solves a constrained regression problem whereas KernelSHAP uses very large weights on the empty set and full sert. The fact that the experiments in the current paper have such a big gap in performance between these two methods indicates something is strange about the implementation.\n    * FastSHAP is fundamentally a different *kind* of estimator because it learns a neural network to predict the Shapley values with respect to *multiple* value functions. So if you're only estimating the Shapley value of a single value function, it doesn't really make sense to use FastSHAP. (To be fair, this is mentioned in passing in a runtime table in the experiments.)\n\n3. As for their algorithm, I'm confused bordering on concerned. Instead of solving one regression problem, they iteratively solve a sequence of algorithms. Intuitively, I don't see the value in distributing the total budget into a sequence of worse solutions than just solving the regression problem once (but hey, I could be wrong about this). The bigger concern I have is that, as written, Algorithm 1 is given a budget of $m$ (like the other estimators), but then makes $m$ independent samples for each of the $T$ iterations. Effectively, this algorithm is getting to see $T * m$ evaluations of the value functions, whereas the algorithms they compare to gets only $m$ evaluations. If this is how the experiments are implemented, they're deeply unfair because SimShapley is getting way more access to $v$ than the other algorithms.\n\n4. As mentioned earlier, their experiments are inconsistent in terms of which algorithms they compare to. And they present performance on different datasets in different ways. Pessimistically, this would indicate cherrypicking of experimental results. My **strong** suggestion is to standardize their experiments: Plot bias as a function of sample size in one big plot with one subplot per dataset. If you want, make a similar plot for time, and a similar one for \"global\" value functions (again, the only difference here is how $v$ is defined so, even if it means modifying some implementation code, you should compare to *every* Shapley value estimator that you mention). This would be a) more visually appealing, and b) much easier to see overall performance of each estimator in a fair way.\n\nHere are the recent (very good) Shapley value estimators you should compare to:\n\nProxySPEX: https://arxiv.org/abs/2505.17495\n\nRegressionMSR: https://arxiv.org/abs/2506.11849\n\nMinor:\n\n* There's already an algorithm called \"SimSHAP\" so calling this one \"SimShapley\" seems confusingly similar. Especially because this algorithm seems to be quite different from \"SimSHAP\".", "questions": "* Is there a typo in Equation 7b? I.e., should it be $\\delta^{(n+1)} = t \\beta^{(n)} + (1-t) \\delta^{(n)}$.\n\n* In Table 2, you run KernelSHAP with $m=64$ and SimShapley with $m=64$ and some $T$ (I assume greater than 1). How many evaluations does each algorithm get? Is it that KernelSHAP gets $64$ and SimShapley gets $64T$? What do you set $T$ to?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper considers the problem of estimating Shapley values of a general value function $v$. They propose a linear regression-based method similar to KernelSHAP. Instead of solving the weighted and constrained least squares problem once, they iteratively solve a sequence of regression problems. When solving the next regression problem, they draw $m$ new samples and add a regularization term on the \"moving average\" solution of the regression problems so far.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "N/A", "weaknesses": "Big:\n\n1. I think the biggest structural issue in this paper is how they discuss and compare to prior work. Across the discussions and experiments, they consider different subsets of the estimators FastSHAP, SimSHAP, LeverageSHAP, and KernelSHAP. The first issue is that they selectively compare to these algorithms e.g., they'll sometimes ignore LeverageSHAP or FastSHAP depending on the experiment. The bigger issue is that they don't consider other SOTA estimators which include PermutationSHAP, and (two recent methods) RegressionMSR and ProxySPEX.\n\n2. The paper ignores fundamental properties of estimators, or says things that are just incorrect about them:\n    * They claim prior estimators can't be used for general value functions $v$, but, except for methods like TreeSHAP (which isn't mentioned in the paper), this is just false. Basically all near-SOTA Shapley value estimators are agnostic to the value function. (I think they know this e.g., they present a \"variant\" of their estimator for global explanations in Appendix C but the only change is one line about how the value function $v$ is computed.) It is true that *implementations* of these algorithms are generally tied to specific value functions but, I would say a) the onus for modifying the standard implementation is on researchers if they want experiments on a new kind of value function, and b) the shap-iq library is extracting the value function logic to an arbitrary class that makes it easy to switch which value function is used.\n    * They treat KernelSHAP and LeverageSHAP as different algorithms. In reality, they're the same algorithm except a) the sampling distribution is different, and b) LeverageSHAP solves a constrained regression problem whereas KernelSHAP uses very large weights on the empty set and full sert. The fact that the experiments in the current paper have such a big gap in performance between these two methods indicates something is strange about the implementation.\n    * FastSHAP is fundamentally a different *kind* of estimator because it learns a neural network to predict the Shapley values with respect to *multiple* value functions. So if you're only estimating the Shapley value of a single value function, it doesn't really make sense to use FastSHAP. (To be fair, this is mentioned in passing in a runtime table in the experiments.)\n\n3. As for their algorithm, I'm confused bordering on concerned. Instead of solving one regression problem, they iteratively solve a sequence of algorithms. Intuitively, I don't see the value in distributing the total budget into a sequence of worse solutions than just solving the regression problem once (but hey, I could be wrong about this). The bigger concern I have is that, as written, Algorithm 1 is given a budget of $m$ (like the other estimators), but then makes $m$ independent samples for each of the $T$ iterations. Effectively, this algorithm is getting to see $T * m$ evaluations of the value functions, whereas the algorithms they compare to gets only $m$ evaluations. If this is how the experiments are implemented, they're deeply unfair because SimShapley is getting way more access to $v$ than the other algorithms.\n\n4. As mentioned earlier, their experiments are inconsistent in terms of which algorithms they compare to. And they present performance on different datasets in different ways. Pessimistically, this would indicate cherrypicking of experimental results. My **strong** suggestion is to standardize their experiments: Plot bias as a function of sample size in one big plot with one subplot per dataset. If you want, make a similar plot for time, and a similar one for \"global\" value functions (again, the only difference here is how $v$ is defined so, even if it means modifying some implementation code, you should compare to *every* Shapley value estimator that you mention). This would be a) more visually appealing, and b) much easier to see overall performance of each estimator in a fair way.\n\nHere are the recent (very good) Shapley value estimators you should compare to:\n\nProxySPEX: https://arxiv.org/abs/2505.17495\n\nRegressionMSR: https://arxiv.org/abs/2506.11849\n\nMinor:\n\n* There's already an algorithm called \"SimSHAP\" so calling this one \"SimShapley\" seems confusingly similar. Especially because this algorithm seems to be quite different from \"SimSHAP\".", "questions": "* Is there a typo in Equation 7b? I.e., should it be $\\delta^{(n+1)} = t \\beta^{(n)} + (1-t) \\delta^{(n)}$.\n\n* In Table 2, you run KernelSHAP with $m=64$ and SimShapley with $m=64$ and some $T$ (I assume greater than 1). How many evaluations does each algorithm get? Is it that KernelSHAP gets $64$ and SimShapley gets $64T$? What do you set $T$ to?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761751589611}, {"id": "UudT638quC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11722/Reviewer_ZozL"], "rating": 4, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 5, "summary": "The paper presents SIM-Shapley (SIM-SHAP) as a novel approximation method to estimate Shapely values. SIM-SHAP is quite related to many amortization and approximate methods relying on the weighted least squares representation of the Shapley value. SIM-SHAP extends the well known KernelSHAP approxmation method and Unbiased KernelSHAP (Covert and Lee, 2021) with an exponential moving average throughout the sampling and Shapley value computation procedure. The paper concludes by evaluating the approximation quality of SIM-SHAP compared to state-of-the-art methods. All in all the empirical evaluation shows that SIM-SHAP does not set the new state of the art but is comparable to it.\n\nThe following contains the **references** used throughout the review:\n- [1] https://inria.hal.science/hal-03414720v1/document\n- [2] https://arxiv.org/pdf/2506.11849\n- [3] https://ojs.aaai.org/index.php/AAAI/article/view/29225\n- [4] https://www.sciencedirect.com/science/article/abs/pii/S0305054808000804", "review_text": "The paper presents SIM-Shapley (SIM-SHAP) as a novel approximation method to estimate Shapely values. SIM-SHAP is quite related to many amortization and approximate methods relying on the weighted least squares representation of the Shapley value. SIM-SHAP extends the well known KernelSHAP approxmation method and Unbiased KernelSHAP (Covert and Lee, 2021) with an exponential moving average throughout the sampling and Shapley value computation procedure. The paper concludes by evaluating the approximation quality of SIM-SHAP compared to state-of-the-art methods. All in all the empirical evaluation shows that SIM-SHAP does not set the new state of the art but is comparable to it.\n\nThe following contains the **references** used throughout the review:\n- [1] https://inria.hal.science/hal-03414720v1/document\n- [2] https://arxiv.org/pdf/2506.11849\n- [3] https://ojs.aaai.org/index.php/AAAI/article/view/29225\n- [4] https://www.sciencedirect.com/science/article/abs/pii/S0305054808000804", "strengths": "- **Important Research:** Computation of Shapley Values is relevant for more and more application areas most prominently Explainable AI. Black box (game-agnostic) estimation methods are important with better methods proposed every year. Therein, this work contributes to an important research field. The most important aspect of the contribution is the dynamic nature of the estimation, and that it natively supports estimation of current estimation progress (which is often a positive argument brought in favor of [4]).\n---\n- **Well Written and Good Presentation:** The paper is well written and clear to follow. The contribution is clearly presented and well organized. The presentation good. \n---\n- **Noation:** The mathematical notation and presentation is good.\n---\n- **Good Baselines:** The paper compares itself already to the most important baseline methods. While the current SOTA [2] is missing, the current selection is already quite well chosen. Some improvements can be done.\n---\n- **Code:** I greatly appreciate the submission of the source code and the quality of the code. Thank you for including the demo notebooks they are interesting to go through.", "weaknesses": "- **Empirical Evaluation:** While the empirical evaluation is generally well done, it would be nice to also include some of the methods described in the below point in the evaluation such as [1] and [4]. The comparison with [4] is interesting since it is usually also very efficient but a bit more limited in terms of performance and thus serves as a good baseline how hard the problems are and how impressive the gains of improvements are. I also do not like that the empirical evaluation does not compare against *real ground truth* values at least in a couple of cases, but estimates them with KernelSHAP running longer. Real ground truths can be easily achieved by computing via brute force for datasets with a moderate number of features (16 features needs approximately $2^{16} \\approx 64k$ model calls). This should be doable for some settings.\n---\n- **Missing Related Work:** While the paper compares itself to modern methods to estimate the Shapley value empirically, the related work is not very well prepared in the current version of the manuscript and the majority of the related work is moved into the appendix. The whole SIM-SHAP method seems to me very related to [1] and potential follow-up works on this. I feel like this should be described in the manuscript. Some other impactful Shapley value estimation methods ([3] and [4]) are also missing. Most prominently, the RegressionMSR method proposed in [2] would need to be discussed in the work, since it presents the current state of the art in terms of Shapley value estimation. Since the paper was only released rather recently, I don't think an empirical evaluation is necessary (of course interesting) but delineating against is would definitely improve the work.\n---\n- **Doubts about LeverageSHAP usage:** I have my doubts about LeverageSHAP not being able to be used on classification tasks and the notion of it being imputer-dependent since LeverageSHAP is a black-box estimation method. While the *implementation* of LeverageSHAP may be limited to these tasks, it should be easily transferrable. Also see question 1.\n---\nFor me this work is **quite borderline** (in its current state sitting a bit on the reject side), but depending on the resolution (or partial resolution) of my concerns, I think **this paper could be accepted.**", "questions": "### Question 1\nYou write\n> SIM-Shapley uniquely provides both game and imputer agnosticism—capabilities absent in existing methods\n\nand \n> Leverage- SHAP (Musco & Witter, 2025) employs statistical leverage scores to improve sampling efficiency but is limited to mean-value imputation and specific task types\n\nHow is LeverageSHAP **not** imputer invariant? LeverageSHAP is basically KernelSHAP but a smarter coalition selection / sampling. The imputer used behind the value function does not influence the applicability of LeverageSHAP, no?\n\n---\n### Question 2\nHow does your method compare to [1]?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents SIM-Shapley (SIM-SHAP) as a novel approximation method to estimate Shapely values. SIM-SHAP is quite related to many amortization and approximate methods relying on the weighted least squares representation of the Shapley value. SIM-SHAP extends the well known KernelSHAP approxmation method and Unbiased KernelSHAP (Covert and Lee, 2021) with an exponential moving average throughout the sampling and Shapley value computation procedure. The paper concludes by evaluating the approximation quality of SIM-SHAP compared to state-of-the-art methods. All in all the empirical evaluation shows that SIM-SHAP does not set the new state of the art but is comparable to it.\n\nThe following contains the **references** used throughout the review:\n- [1] https://inria.hal.science/hal-03414720v1/document\n- [2] https://arxiv.org/pdf/2506.11849\n- [3] https://ojs.aaai.org/index.php/AAAI/article/view/29225\n- [4] https://www.sciencedirect.com/science/article/abs/pii/S0305054808000804", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "- **Important Research:** Computation of Shapley Values is relevant for more and more application areas most prominently Explainable AI. Black box (game-agnostic) estimation methods are important with better methods proposed every year. Therein, this work contributes to an important research field. The most important aspect of the contribution is the dynamic nature of the estimation, and that it natively supports estimation of current estimation progress (which is often a positive argument brought in favor of [4]).\n---\n- **Well Written and Good Presentation:** The paper is well written and clear to follow. The contribution is clearly presented and well organized. The presentation good. \n---\n- **Noation:** The mathematical notation and presentation is good.\n---\n- **Good Baselines:** The paper compares itself already to the most important baseline methods. While the current SOTA [2] is missing, the current selection is already quite well chosen. Some improvements can be done.\n---\n- **Code:** I greatly appreciate the submission of the source code and the quality of the code. Thank you for including the demo notebooks they are interesting to go through.", "weaknesses": "- **Empirical Evaluation:** While the empirical evaluation is generally well done, it would be nice to also include some of the methods described in the below point in the evaluation such as [1] and [4]. The comparison with [4] is interesting since it is usually also very efficient but a bit more limited in terms of performance and thus serves as a good baseline how hard the problems are and how impressive the gains of improvements are. I also do not like that the empirical evaluation does not compare against *real ground truth* values at least in a couple of cases, but estimates them with KernelSHAP running longer. Real ground truths can be easily achieved by computing via brute force for datasets with a moderate number of features (16 features needs approximately $2^{16} \\approx 64k$ model calls). This should be doable for some settings.\n---\n- **Missing Related Work:** While the paper compares itself to modern methods to estimate the Shapley value empirically, the related work is not very well prepared in the current version of the manuscript and the majority of the related work is moved into the appendix. The whole SIM-SHAP method seems to me very related to [1] and potential follow-up works on this. I feel like this should be described in the manuscript. Some other impactful Shapley value estimation methods ([3] and [4]) are also missing. Most prominently, the RegressionMSR method proposed in [2] would need to be discussed in the work, since it presents the current state of the art in terms of Shapley value estimation. Since the paper was only released rather recently, I don't think an empirical evaluation is necessary (of course interesting) but delineating against is would definitely improve the work.\n---\n- **Doubts about LeverageSHAP usage:** I have my doubts about LeverageSHAP not being able to be used on classification tasks and the notion of it being imputer-dependent since LeverageSHAP is a black-box estimation method. While the *implementation* of LeverageSHAP may be limited to these tasks, it should be easily transferrable. Also see question 1.\n---\nFor me this work is **quite borderline** (in its current state sitting a bit on the reject side), but depending on the resolution (or partial resolution) of my concerns, I think **this paper could be accepted.**", "questions": "### Question 1\nYou write\n> SIM-Shapley uniquely provides both game and imputer agnosticism—capabilities absent in existing methods\n\nand \n> Leverage- SHAP (Musco & Witter, 2025) employs statistical leverage scores to improve sampling efficiency but is limited to mean-value imputation and specific task types\n\nHow is LeverageSHAP **not** imputer invariant? LeverageSHAP is basically KernelSHAP but a smarter coalition selection / sampling. The imputer used behind the value function does not influence the applicability of LeverageSHAP, no?\n\n---\n### Question 2\nHow does your method compare to [1]?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761583977827}], "openreview_url": "https://openreview.net/forum?id=xg2ccsYpZY", "arxiv_id": "2505.08198", "paper_pdf": "papers/xg2ccsYpZY.pdf", "paper_pdf_sha256": "7cc6c06f4b2bff81d3a8db1a24d1def094c0e311299c11f673782a8f739f4c72", "paper_pdf_bytes": 760017, "paper_pdf_source": "openreview", "code_url": "https://github.com/nliulab/SIM-Shapley", "code_repository": "nliulab/SIM-Shapley", "code_commit": "afcb1456c28bcec078c5188e33a37ea6ed0f3b0e", "code_archive": "repos/xg2ccsYpZY.zip", "code_archive_sha256": "4b58356cc5e43ee89ca72b25cb97f050b60a664ab0f145d4ed694cc64e5f01fb", "code_archive_bytes": 3600452, "code_file_count": 31, "code_extensions": {".py": 27, ".ipynb": 4}, "github_disk_usage_kb": 4119, "github_languages": {"Python": 188239, "AGS Script": 46939}, "github_archived": false, "github_pushed_at": "2026-04-02T14:35:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sim-shapley-a-stable-and-computationally"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GYwH71ugtC", "year": 2025, "status": "rejected", "title": "Retrieval Augmented Time Series Forecasting", "authors": ["Sungwon Han", "Seungeon Lee", "Meeyoung Cha", "Sercan O Arik", "Jinsung Yoon"], "authorids": ["~Sungwon_Han1", "~Seungeon_Lee1", "~Meeyoung_Cha2", "~Sercan_O_Arik1", "~Jinsung_Yoon1"], "authors_source": "OpenReview API", "abstract": "Time series forecasting uses historical data to predict future trends, leveraging the relationships between past observations and available features. In this paper, we propose, RAFT, a retrieval-augmented time series forecasting method to provide sufficient inductive biases and complement the model's learning capacity. When forecasting the subsequent time frames, we directly retrieve historical data candidates from the training dataset with patterns most similar to the input, and utilize the future values of these candidates alongside the inputs to obtain predictions. This simple approach augments the model's capacity by externally providing information about past patterns via retrieval modules. Our empirical evaluations on eight benchmark datasets show that RAFT consistently outperforms contemporary baselines, an average win ratio of 86% for multivariate forecasting and 80% for univariate forecasting tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "XvOSaqJeMh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2137/Reviewer_CVZn"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes to incorporate a retrieval module for the time series forecasting task. It proposes to perform retrieval in observation space, selecting the `m` most similar lookback windows, and calculating a weighted average of their corresponding horizons, where the weights are a softmax of the similarity scores previously computed. Following this, several linear layers are used to project both the original lookback window and retrieved time series into the prediction. A \"multiple period\" extension is also proposed, which performs the same process on downsampled versions of the time series, which are incorporated into the final forecast.\n\nThe paper performs experiments on the long sequence forecasting setting and shows improved performance compared to several (older) baselines.", "review_text": "The paper proposes to incorporate a retrieval module for the time series forecasting task. It proposes to perform retrieval in observation space, selecting the `m` most similar lookback windows, and calculating a weighted average of their corresponding horizons, where the weights are a softmax of the similarity scores previously computed. Following this, several linear layers are used to project both the original lookback window and retrieved time series into the prediction. A \"multiple period\" extension is also proposed, which performs the same process on downsampled versions of the time series, which are incorporated into the final forecast.\n\nThe paper performs experiments on the long sequence forecasting setting and shows improved performance compared to several (older) baselines.", "strengths": "Retrieval is an interesting idea to explore in the context of time series forecasting. The paper shows that the proposed model improves over some baselines, and presents some analysis surrounding the proposed method.", "weaknesses": "From my reading, I am unclear on what is the set of datapoints the retrieval is being performed on. Is it on the entire time series, i.e. the whole training + validation set + the test set that is becoming available during rolling window evaluation? Or is there a limit to how much data is being searched over, and how is this set?\n\nRelated to this, is my concern regarding the motivation of the paper - the introduction argues that existing methods are memorizing the training set, and retrieval is a solution to help generalize by extracting information relevant to the query. However, it turns out that the proposed approach is still retrieving from the train set. I see no big difference between the proposed approach and a Transformer model which is also attending to the lookback window. Retrieval seems to only make sense in the zero-shot setting, where we are trying to make predictions on a time series from a completely new dataset, and performance can be improved by retrieving related time series from that dataset, and critically, this dataset wasn't in the training set, so that we can show that it is not just memorization. \n\nThe experiment design for ablating the effects of the different components of the proposed method, especially to isolate the improvements from the retrieval component can be improved. The current experiments simply take NLinear as \"without retrieval\", but the design space is much larger and this is an important set of experiments to better understand the effects of retrieval.", "questions": "1. Question from the weaknesses section regarding what is the set of datapoints for retrieval.\n2. The introduction states: ``This paper examines a critical open question in time-series forecasting: “do current models possess the necessary inductive biases and learning capacity to extract generalizable patterns from training data and achieve high accuracy?” '' -- what are the conclusions regarding this research question?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes to incorporate a retrieval module for the time series forecasting task. It proposes to perform retrieval in observation space, selecting the `m` most similar lookback windows, and calculating a weighted average of their corresponding horizons, where the weights are a softmax of the similarity scores previously computed. Following this, several linear layers are used to project both the original lookback window and retrieved time series into the prediction. A \"multiple period\" extension is also proposed, which performs the same process on downsampled versions of the time series, which are incorporated into the final forecast.\n\nThe paper performs experiments on the long sequence forecasting setting and shows improved performance compared to several (older) baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Retrieval is an interesting idea to explore in the context of time series forecasting. The paper shows that the proposed model improves over some baselines, and presents some analysis surrounding the proposed method.", "weaknesses": "From my reading, I am unclear on what is the set of datapoints the retrieval is being performed on. Is it on the entire time series, i.e. the whole training + validation set + the test set that is becoming available during rolling window evaluation? Or is there a limit to how much data is being searched over, and how is this set?\n\nRelated to this, is my concern regarding the motivation of the paper - the introduction argues that existing methods are memorizing the training set, and retrieval is a solution to help generalize by extracting information relevant to the query. However, it turns out that the proposed approach is still retrieving from the train set. I see no big difference between the proposed approach and a Transformer model which is also attending to the lookback window. Retrieval seems to only make sense in the zero-shot setting, where we are trying to make predictions on a time series from a completely new dataset, and performance can be improved by retrieving related time series from that dataset, and critically, this dataset wasn't in the training set, so that we can show that it is not just memorization. \n\nThe experiment design for ablating the effects of the different components of the proposed method, especially to isolate the improvements from the retrieval component can be improved. The current experiments simply take NLinear as \"without retrieval\", but the design space is much larger and this is an important set of experiments to better understand the effects of retrieval.", "questions": "1. Question from the weaknesses section regarding what is the set of datapoints for retrieval.\n2. The introduction states: ``This paper examines a critical open question in time-series forecasting: “do current models possess the necessary inductive biases and learning capacity to extract generalizable patterns from training data and achieve high accuracy?” '' -- what are the conclusions regarding this research question?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730963846312}, {"id": "aTPWsMiGR0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2137/Reviewer_qdCV"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes RAFT (Retrieval-Augmented Forecasting of Time-series), a novel method for time series forecasting that leverages a retrieval module to provide the model with relevant historical patterns. The key idea is to retrieve the most similar historical data to the current input and utilize the future values of these retrieved candidates to enhance predictions. This reduces the burden on the model to memorize all possible patterns, especially rare or complex ones.\nThe retrieval module operates by finding the most similar key patches to the input query from the entire training time series, and then retrieving the corresponding future value patches. An attention-like mechanism is used to weigh the retrieved value patches based on their similarity to the input. RAFT extends this idea to multiple time series generated by downsampling the original series at different periods, allowing it to capture patterns at various temporal resolutions.\nThe method is built upon a simple MLP architecture, demonstrating that a well-designed retrieval module can provide an effective inductive bias for time series forecasting. Extensive experiments on eight benchmark datasets show that RAFT consistently outperforms state-of-the-art baselines, achieving an average win ratio of 86% for multivariate forecasting and 80% for univariate forecasting tasks.\nFurther analyses using synthetic datasets reveal that RAFT is particularly beneficial when rare patterns repeat in the time series or when patterns are less temporally correlated. The retrieval module enables the model to directly leverage relevant historical patterns in such scenarios.\nOverall, the paper presents a novel perspective on enhancing time series forecasting models with retrieval-based methods, opening up new possibilities in this domain.", "review_text": "This paper proposes RAFT (Retrieval-Augmented Forecasting of Time-series), a novel method for time series forecasting that leverages a retrieval module to provide the model with relevant historical patterns. The key idea is to retrieve the most similar historical data to the current input and utilize the future values of these retrieved candidates to enhance predictions. This reduces the burden on the model to memorize all possible patterns, especially rare or complex ones.\nThe retrieval module operates by finding the most similar key patches to the input query from the entire training time series, and then retrieving the corresponding future value patches. An attention-like mechanism is used to weigh the retrieved value patches based on their similarity to the input. RAFT extends this idea to multiple time series generated by downsampling the original series at different periods, allowing it to capture patterns at various temporal resolutions.\nThe method is built upon a simple MLP architecture, demonstrating that a well-designed retrieval module can provide an effective inductive bias for time series forecasting. Extensive experiments on eight benchmark datasets show that RAFT consistently outperforms state-of-the-art baselines, achieving an average win ratio of 86% for multivariate forecasting and 80% for univariate forecasting tasks.\nFurther analyses using synthetic datasets reveal that RAFT is particularly beneficial when rare patterns repeat in the time series or when patterns are less temporally correlated. The retrieval module enables the model to directly leverage relevant historical patterns in such scenarios.\nOverall, the paper presents a novel perspective on enhancing time series forecasting models with retrieval-based methods, opening up new possibilities in this domain.", "strengths": "This paper presents several noteworthy strengths across various dimensions:\n\nOriginality: The proposed RAFT method offers a novel approach to time series forecasting by incorporating a retrieval module. While retrieval-augmented methods have been explored in other domains like natural language processing, their application to time series forecasting is innovative. By directly leveraging relevant historical patterns, RAFT introduces a new paradigm for handling complex and rare patterns in time series data.\nQuality: The paper is well-structured and thoroughly evaluates the proposed method. The authors provide a clear description of the retrieval module architecture and how it is integrated into the overall forecasting model. The experimental setup is comprehensive, considering both multivariate and univariate forecasting tasks across eight diverse benchmark datasets. The results convincingly demonstrate the superiority of RAFT over state-of-the-art baselines.\nClarity: The paper is well-written and easy to follow. The authors provide a clear motivation for their approach and explain the technical details of RAFT in a concise and understandable manner. The use of illustrative figures, such as the retrieval module architecture (Figure 2) and the overall RAFT architecture (Figure 3), enhances the clarity of the proposed method. The experimental results are presented in a readable format, making it easy to compare RAFT's performance against the baselines.\nSignificance: The paper makes a significant contribution to the field of time series forecasting. By demonstrating the effectiveness of retrieval-augmented methods, RAFT opens up new research directions and possibilities for improving forecasting models. The analyses using synthetic datasets provide valuable insights into the scenarios where retrieval is particularly beneficial, such as handling rare patterns or less temporally correlated data. These findings have important implications for real-world applications where such characteristics are common.\n\nMoreover, the paper's results challenge the current reliance on increasing model capacity to capture complex patterns. RAFT shows that a simpler MLP architecture, when augmented with a well-designed retrieval module, can outperform more sophisticated models. This highlights the potential of retrieval-based methods as a complementary approach to improving time series forecasting.\nIn summary, the paper's originality, quality, clarity, and significance make it a valuable contribution to the field, offering new insights and directions for future research in time series forecasting.", "weaknesses": "While the paper presents a novel and effective approach to time series forecasting, there are a few areas that could be improved or require further clarification:\n\nRetrieval module design: The current retrieval module uses a simple similarity measure (Pearson correlation) to find the most relevant historical patterns. However, the authors do not provide a thorough justification for this choice or explore alternative similarity measures. Time series data often exhibit complex, nonlinear, and nonstationary characteristics, which may not be well-captured by linear correlation. Exploring more sophisticated similarity measures or learning adaptive similarity functions could potentially improve the retrieval process and the overall forecasting performance.\nComputational efficiency: The paper does not provide a detailed analysis of the computational complexity and efficiency of the proposed method. The retrieval process involves comparing the input query with all key patches in the training data, which could be computationally expensive for large datasets. While the authors mention that the stride for the sliding window can be adjusted for computational efficiency (Section 3.2), they do not provide empirical results or discussions on the trade-off between computational cost and forecasting accuracy. A more comprehensive analysis of the method's scalability and efficiency would be valuable for practical applications.\nSensitivity to hyperparameters: The performance of RAFT may be sensitive to the choice of hyperparameters, such as the number of retrieved patches (m), the temperature (τ), and the set of periods (P) used for generating multiple time series. While the authors provide the chosen values of m for each experiment setting (Appendix B), they do not discuss how these values were determined or provide insights into the sensitivity of the results to these hyperparameters. A more systematic analysis of the impact of these hyperparameters on the forecasting performance would enhance the robustness and reproducibility of the proposed method.\nLimited ablation studies: The paper would benefit from more extensive ablation studies to better understand the individual contributions of the proposed components. For example, the authors could evaluate the performance of RAFT without the multi-period extension to assess the impact of capturing patterns at different temporal resolutions. Similarly, comparing the performance of RAFT with and without the attention-like weighting of the retrieved value patches could provide insights into the importance of this mechanism.\nEvaluation on more diverse datasets: While the paper evaluates RAFT on eight benchmark datasets, these datasets are primarily from the energy, traffic, and weather domains. To demonstrate the generalizability of the proposed method, it would be valuable to include datasets from a wider range of application domains, such as finance, healthcare, or social media. Moreover, the paper could benefit from evaluations on datasets with different characteristics, such as varying lengths, missing values, or irregularly sampled time series.\n\nAddressing these weaknesses would further strengthen the paper's contributions and provide a more comprehensive understanding of the proposed retrieval-augmented forecasting method. However, it is important to note that these weaknesses do not diminish the overall value and novelty of the work, and the authors have already made significant contributions to the field of time series forecasting.", "questions": "1. Choice of similarity measure: Can you provide more insights into the choice of Pearson correlation as the similarity measure in the retrieval module? Have you considered or experimented with other similarity measures, such as dynamic time warping (DTW), cross-correlation, or learned similarity functions? How do you think the choice of similarity measure affects the retrieval process and the overall forecasting performance?\n2. Computational efficiency: Can you provide more details on the computational complexity and efficiency of the proposed method, particularly the retrieval process? How does the computational cost scale with the size of the dataset and the length of the time series? Have you considered any techniques to improve the efficiency of the retrieval process, such as indexing or approximate nearest neighbor search?\n3. Hyperparameter sensitivity: How sensitive are the results to the choice of hyperparameters, such as the number of retrieved patches (m), the temperature (τ), and the set of periods (P)? Can you provide more details on how these hyperparameters were determined for each experiment setting? Have you considered using techniques like cross-validation or Bayesian optimization to tune these hyperparameters?\nAblation studies: Can you provide more ablation studies to investigate the individual contributions of the proposed components? For example, how does the performance of RAFT change when the multi-period extension is removed? How important is the attention-like weighting of the retrieved value patches compared to using a simple average or the most similar patch?\nEvaluation on diverse datasets: Have you considered evaluating RAFT on datasets from a wider range of application domains beyond energy, traffic, and weather? How do you expect the proposed method to perform on datasets with different characteristics, such as varying lengths, missing values, or irregularly sampled time series? Providing results on more diverse datasets could strengthen the claims of generalizability.\nHandling multiple retrieved patterns: In the current implementation, RAFT retrieves the top-m most similar patterns and aggregates them using an attention-like weighting scheme. Have you considered other approaches to handle multiple retrieved patterns, such as clustering similar patterns or using a more sophisticated aggregation method? How do you think these alternative approaches would impact the forecasting performance?\nComparison with other retrieval-based methods: While the paper compares RAFT with several state-of-the-art forecasting methods, it would be interesting to see a comparison with other retrieval-based methods, such as those mentioned in the related work. How does RAFT differ from these existing retrieval-based approaches, and how does it compare in terms of performance and efficiency?\nVisualization of retrieved patterns: Can you provide more visualizations of the retrieved patterns and their corresponding future values? It would be helpful to see examples of how the retrieved patterns contribute to the final forecasting results, particularly in cases where RAFT significantly outperforms the baselines. Such visualizations could provide additional insights into the effectiveness of the retrieval process.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes RAFT (Retrieval-Augmented Forecasting of Time-series), a novel method for time series forecasting that leverages a retrieval module to provide the model with relevant historical patterns. The key idea is to retrieve the most similar historical data to the current input and utilize the future values of these retrieved candidates to enhance predictions. This reduces the burden on the model to memorize all possible patterns, especially rare or complex ones.\nThe retrieval module operates by finding the most similar key patches to the input query from the entire training time series, and then retrieving the corresponding future value patches. An attention-like mechanism is used to weigh the retrieved value patches based on their similarity to the input. RAFT extends this idea to multiple time series generated by downsampling the original series at different periods, allowing it to capture patterns at various temporal resolutions.\nThe method is built upon a simple MLP architecture, demonstrating that a well-designed retrieval module can provide an effective inductive bias for time series forecasting. Extensive experiments on eight benchmark datasets show that RAFT consistently outperforms state-of-the-art baselines, achieving an average win ratio of 86% for multivariate forecasting and 80% for univariate forecasting tasks.\nFurther analyses using synthetic datasets reveal that RAFT is particularly beneficial when rare patterns repeat in the time series or when patterns are less temporally correlated. The retrieval module enables the model to directly leverage relevant historical patterns in such scenarios.\nOverall, the paper presents a novel perspective on enhancing time series forecasting models with retrieval-based methods, opening up new possibilities in this domain.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "This paper presents several noteworthy strengths across various dimensions:\n\nOriginality: The proposed RAFT method offers a novel approach to time series forecasting by incorporating a retrieval module. While retrieval-augmented methods have been explored in other domains like natural language processing, their application to time series forecasting is innovative. By directly leveraging relevant historical patterns, RAFT introduces a new paradigm for handling complex and rare patterns in time series data.\nQuality: The paper is well-structured and thoroughly evaluates the proposed method. The authors provide a clear description of the retrieval module architecture and how it is integrated into the overall forecasting model. The experimental setup is comprehensive, considering both multivariate and univariate forecasting tasks across eight diverse benchmark datasets. The results convincingly demonstrate the superiority of RAFT over state-of-the-art baselines.\nClarity: The paper is well-written and easy to follow. The authors provide a clear motivation for their approach and explain the technical details of RAFT in a concise and understandable manner. The use of illustrative figures, such as the retrieval module architecture (Figure 2) and the overall RAFT architecture (Figure 3), enhances the clarity of the proposed method. The experimental results are presented in a readable format, making it easy to compare RAFT's performance against the baselines.\nSignificance: The paper makes a significant contribution to the field of time series forecasting. By demonstrating the effectiveness of retrieval-augmented methods, RAFT opens up new research directions and possibilities for improving forecasting models. The analyses using synthetic datasets provide valuable insights into the scenarios where retrieval is particularly beneficial, such as handling rare patterns or less temporally correlated data. These findings have important implications for real-world applications where such characteristics are common.\n\nMoreover, the paper's results challenge the current reliance on increasing model capacity to capture complex patterns. RAFT shows that a simpler MLP architecture, when augmented with a well-designed retrieval module, can outperform more sophisticated models. This highlights the potential of retrieval-based methods as a complementary approach to improving time series forecasting.\nIn summary, the paper's originality, quality, clarity, and significance make it a valuable contribution to the field, offering new insights and directions for future research in time series forecasting.", "weaknesses": "While the paper presents a novel and effective approach to time series forecasting, there are a few areas that could be improved or require further clarification:\n\nRetrieval module design: The current retrieval module uses a simple similarity measure (Pearson correlation) to find the most relevant historical patterns. However, the authors do not provide a thorough justification for this choice or explore alternative similarity measures. Time series data often exhibit complex, nonlinear, and nonstationary characteristics, which may not be well-captured by linear correlation. Exploring more sophisticated similarity measures or learning adaptive similarity functions could potentially improve the retrieval process and the overall forecasting performance.\nComputational efficiency: The paper does not provide a detailed analysis of the computational complexity and efficiency of the proposed method. The retrieval process involves comparing the input query with all key patches in the training data, which could be computationally expensive for large datasets. While the authors mention that the stride for the sliding window can be adjusted for computational efficiency (Section 3.2), they do not provide empirical results or discussions on the trade-off between computational cost and forecasting accuracy. A more comprehensive analysis of the method's scalability and efficiency would be valuable for practical applications.\nSensitivity to hyperparameters: The performance of RAFT may be sensitive to the choice of hyperparameters, such as the number of retrieved patches (m), the temperature (τ), and the set of periods (P) used for generating multiple time series. While the authors provide the chosen values of m for each experiment setting (Appendix B), they do not discuss how these values were determined or provide insights into the sensitivity of the results to these hyperparameters. A more systematic analysis of the impact of these hyperparameters on the forecasting performance would enhance the robustness and reproducibility of the proposed method.\nLimited ablation studies: The paper would benefit from more extensive ablation studies to better understand the individual contributions of the proposed components. For example, the authors could evaluate the performance of RAFT without the multi-period extension to assess the impact of capturing patterns at different temporal resolutions. Similarly, comparing the performance of RAFT with and without the attention-like weighting of the retrieved value patches could provide insights into the importance of this mechanism.\nEvaluation on more diverse datasets: While the paper evaluates RAFT on eight benchmark datasets, these datasets are primarily from the energy, traffic, and weather domains. To demonstrate the generalizability of the proposed method, it would be valuable to include datasets from a wider range of application domains, such as finance, healthcare, or social media. Moreover, the paper could benefit from evaluations on datasets with different characteristics, such as varying lengths, missing values, or irregularly sampled time series.\n\nAddressing these weaknesses would further strengthen the paper's contributions and provide a more comprehensive understanding of the proposed retrieval-augmented forecasting method. However, it is important to note that these weaknesses do not diminish the overall value and novelty of the work, and the authors have already made significant contributions to the field of time series forecasting.", "questions": "1. Choice of similarity measure: Can you provide more insights into the choice of Pearson correlation as the similarity measure in the retrieval module? Have you considered or experimented with other similarity measures, such as dynamic time warping (DTW), cross-correlation, or learned similarity functions? How do you think the choice of similarity measure affects the retrieval process and the overall forecasting performance?\n2. Computational efficiency: Can you provide more details on the computational complexity and efficiency of the proposed method, particularly the retrieval process? How does the computational cost scale with the size of the dataset and the length of the time series? Have you considered any techniques to improve the efficiency of the retrieval process, such as indexing or approximate nearest neighbor search?\n3. Hyperparameter sensitivity: How sensitive are the results to the choice of hyperparameters, such as the number of retrieved patches (m), the temperature (τ), and the set of periods (P)? Can you provide more details on how these hyperparameters were determined for each experiment setting? Have you considered using techniques like cross-validation or Bayesian optimization to tune these hyperparameters?\nAblation studies: Can you provide more ablation studies to investigate the individual contributions of the proposed components? For example, how does the performance of RAFT change when the multi-period extension is removed? How important is the attention-like weighting of the retrieved value patches compared to using a simple average or the most similar patch?\nEvaluation on diverse datasets: Have you considered evaluating RAFT on datasets from a wider range of application domains beyond energy, traffic, and weather? How do you expect the proposed method to perform on datasets with different characteristics, such as varying lengths, missing values, or irregularly sampled time series? Providing results on more diverse datasets could strengthen the claims of generalizability.\nHandling multiple retrieved patterns: In the current implementation, RAFT retrieves the top-m most similar patterns and aggregates them using an attention-like weighting scheme. Have you considered other approaches to handle multiple retrieved patterns, such as clustering similar patterns or using a more sophisticated aggregation method? How do you think these alternative approaches would impact the forecasting performance?\nComparison with other retrieval-based methods: While the paper compares RAFT with several state-of-the-art forecasting methods, it would be interesting to see a comparison with other retrieval-based methods, such as those mentioned in the related work. How does RAFT differ from these existing retrieval-based approaches, and how does it compare in terms of performance and efficiency?\nVisualization of retrieved patterns: Can you provide more visualizations of the retrieved patterns and their corresponding future values? It would be helpful to see examples of how the retrieved patterns contribute to the final forecasting results, particularly in cases where RAFT significantly outperforms the baselines. Such visualizations could provide additional insights into the effectiveness of the retrieval process.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730701317010}, {"id": "cI52fFN8ze", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2137/Reviewer_Ccy2"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduced RAFT - a retrieval augmented time series forecasting model. The main idea is to use the a similarity function between the current time series context (query) and context and forecast horizon pairs (key/value) to retrieve similar time series patches from the training for forecasting. A similarity function between the query and key patches is used to compute attention weights for the value patches. These retrieved patches are then concatenated with the input and the forecast is generated with a linear model. Optionally, the model also uses downsampling of the input to different resolutions and using multiple downsampled retrieval results for forecasting. The authors compare their model with different baseline models on eight datasets. They also perform synthetic experiment to provide evidence when retrieval is useful. The authors's experiments suggest that retrieval is useful when keys are similar to the query, rate patterns are present, and the patterns are temporally less correlated.", "review_text": "The paper introduced RAFT - a retrieval augmented time series forecasting model. The main idea is to use the a similarity function between the current time series context (query) and context and forecast horizon pairs (key/value) to retrieve similar time series patches from the training for forecasting. A similarity function between the query and key patches is used to compute attention weights for the value patches. These retrieved patches are then concatenated with the input and the forecast is generated with a linear model. Optionally, the model also uses downsampling of the input to different resolutions and using multiple downsampled retrieval results for forecasting. The authors compare their model with different baseline models on eight datasets. They also perform synthetic experiment to provide evidence when retrieval is useful. The authors's experiments suggest that retrieval is useful when keys are similar to the query, rate patterns are present, and the patterns are temporally less correlated.", "strengths": "### Originality and Significance \nThe authors introduce a conceptually interesting model with a straightforward instantiation. As the proposed method is mainly a simplification of existing attention-based forecasting ideas, the significance in this work is mainly into the experimental evidence whether these simplified instantiations work well for time series forecasting. \n\n### Quality \nThe paper evaluates the model on eight datasets and several baselines from literature. Additionally, the authors discuss their results with additional synthetic experiments to demonstrate under which conditions the retrieval module in RAFT benefits brings a benefit to forecasting. However, I have concerns regarding the evaluation which I elaborate in the Weaknesses section. \n\n### Clarity \nThe paper is clearly written. The proposed idea and the experiments are easy to follow through the presentation of the proposed idea by the used notation and supporting illustrations.", "weaknesses": "The main two weaknesses in the paper are the presentation of the related work and the empirical evaluation. \n\n### Related work \nThe model uses a similarity function and attention weights to get a weighted average candidate value patches from the training dataset and an MLP network to forecast. This is conceptually similar to transformer variants for forecasting as the main difference is mainly in the interaction of learnable weights in the transformer architecture. Such a simplification might be effective, but this is not discussed in the related work. Moreover, there is related work on few-shot forecasting that presents an (arguably more complex) instantiation of the model proposed in this paper (Iwata and Kumagai, 2020; https://arxiv.org/abs/2009.14379). This work should be compared and discussed. In light of this prior work, the novelty of this work is limited. \n\nAnother concern I have with the related work is that recent work on pretrained/foundational time series model are not mentioned (Woo et al., ICML 2024: https://arxiv.org/abs/2402.02592; Das et al., ICML 2024: https://openreview.net/forum?id=jn2iTJas6h; Ansari et al., preprint 2024: https://arxiv.org/abs/2403.07815). I would argue that the proposed idea of using RAG for forecasting could benefit exactly this model class to forecast time series that are not in the pretrainig corpus. This should be discussed in the related work. \n\n### Evaluation\nThe main concern I have on the paper is on the evaluation in the Experiments section. \n\n1.The authors evaluate their method on only eight different datasets. While these datasets have been used widely, many more datasets became available, which allows for a more thorough evaluation. Several of these datasets (Woo et al., ICML 2024: https://arxiv.org/abs/2402.02592; Ansari et al., preprint 2024: https://arxiv.org/abs/2403.07815) from recent papers are publicly available. Running the evaluation on more datasets would give stronger evidence on the performance of the model. \n\n2. I also have concerns about the choice and setup of the baselines. The authors mention that they use the same experimental setup for each baseline, which includes number of training epochs. However, the number of training epochs are arguably an important hyperparameter for tuning a model. I would argue that the real-world performance is best compared by choosing the best possible setting for the baselines, rather than a uniform setting that might result into suboptimal performance of several baselines. Additionally, I think a strong baseline (PatchTST; Nie et al., ICLR 2023: https://arxiv.org/abs/2211.14730) is missing and should be included. When I compare the MSE/MAE of the models in this work with the models in the PatchTST paper, I find that the models in this work perform much weaker. This might suggest that the chosen setting leads to suboptimal performance of the baselines, which might affect the conclusion of this paper. For example: D-Linear has a 0.0764 MSE for the univariate setup in this work and a 0.056 MSE for ETTh1 in the PatchTST paper (Nie at al., ICLR 2023). This suggests that there is actually a better setting to run this baseline and this setting should be used for comparison. I noticed that both papers cite different sources of the datasets, but I checked briefly and at least the ETT datasets seem identical. There might be something different in the setup that I'm not aware off that also explains this difference. \n\nI consider this point critical and this needs to be addressed for me to consider to change my score. Specifically, I would ask the authors to revise the setup and use the settings for the baselines so they are comparable to the results in Nie at al. 2023. I would also like to ask the authors to include the PatchTST baseline. \n\n3. I also have concerns on the win matrix to compare the results. I would argue that in this setup it is not relevant how often RAFT outperforms other baselines, but rather how it compares to the strong baselines. Thus, the reported average win ratio of 86% is somewhat misleading and it would be more useful for the reader to report the win-rate to the next best model and the absolute/relative improvement in MSE/MAE when averaged over the datasets. I would also argue that in the context of the cited paper (Bahri et al., ICLR 2022), the win matrix is used over 65 datasets, while here only over 8 datasets with different forecast horizons. Hence, this introduces redundancy when aggregating the results in a win matrix. It is also not clear how ties are handled. Bahri et al., ICLR 2022 explicitly mention that ties are broken by a statistical test. How is this handled in this work? \n\nI would kindly ask the authors to report the win-rate to the next best model and the absolute/relative improvement in MSE/MAE when averaged over the datasets. I think this gives a more complete picture on performance of the model.", "questions": "1.  It is not clear how ties are handled in the statistical test. Bahri et al., ICLR 2022 explicitly mention that ties are broken by a statistical test. How is this handled in this work? \n\n2. Is the setup/datasets in this work noticeably different from the setup/datasets in the PatchTST paper? \n\n3. In several points of the paper the authors mention the inductive bias of RAFT and that it is more suited for forecasting. It is unclear to me what this inductive bias specifically means. In particular, there is one argument that existing models make i.i.d. assumptions and this is a limitation. How does RAFT overcome this limitation, especially with the empirical result that it is more effective when temporal correlation is lacking? I would kindly ask the authors what the specific inductive bias is that RAFT introduces and how it is a different inductive bias from existing models.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduced RAFT - a retrieval augmented time series forecasting model. The main idea is to use the a similarity function between the current time series context (query) and context and forecast horizon pairs (key/value) to retrieve similar time series patches from the training for forecasting. A similarity function between the query and key patches is used to compute attention weights for the value patches. These retrieved patches are then concatenated with the input and the forecast is generated with a linear model. Optionally, the model also uses downsampling of the input to different resolutions and using multiple downsampled retrieval results for forecasting. The authors compare their model with different baseline models on eight datasets. They also perform synthetic experiment to provide evidence when retrieval is useful. The authors's experiments suggest that retrieval is useful when keys are similar to the query, rate patterns are present, and the patterns are temporally less correlated.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "### Originality and Significance \nThe authors introduce a conceptually interesting model with a straightforward instantiation. As the proposed method is mainly a simplification of existing attention-based forecasting ideas, the significance in this work is mainly into the experimental evidence whether these simplified instantiations work well for time series forecasting. \n\n### Quality \nThe paper evaluates the model on eight datasets and several baselines from literature. Additionally, the authors discuss their results with additional synthetic experiments to demonstrate under which conditions the retrieval module in RAFT benefits brings a benefit to forecasting. However, I have concerns regarding the evaluation which I elaborate in the Weaknesses section. \n\n### Clarity \nThe paper is clearly written. The proposed idea and the experiments are easy to follow through the presentation of the proposed idea by the used notation and supporting illustrations.", "weaknesses": "The main two weaknesses in the paper are the presentation of the related work and the empirical evaluation. \n\n### Related work \nThe model uses a similarity function and attention weights to get a weighted average candidate value patches from the training dataset and an MLP network to forecast. This is conceptually similar to transformer variants for forecasting as the main difference is mainly in the interaction of learnable weights in the transformer architecture. Such a simplification might be effective, but this is not discussed in the related work. Moreover, there is related work on few-shot forecasting that presents an (arguably more complex) instantiation of the model proposed in this paper (Iwata and Kumagai, 2020; https://arxiv.org/abs/2009.14379). This work should be compared and discussed. In light of this prior work, the novelty of this work is limited. \n\nAnother concern I have with the related work is that recent work on pretrained/foundational time series model are not mentioned (Woo et al., ICML 2024: https://arxiv.org/abs/2402.02592; Das et al., ICML 2024: https://openreview.net/forum?id=jn2iTJas6h; Ansari et al., preprint 2024: https://arxiv.org/abs/2403.07815). I would argue that the proposed idea of using RAG for forecasting could benefit exactly this model class to forecast time series that are not in the pretrainig corpus. This should be discussed in the related work. \n\n### Evaluation\nThe main concern I have on the paper is on the evaluation in the Experiments section. \n\n1.The authors evaluate their method on only eight different datasets. While these datasets have been used widely, many more datasets became available, which allows for a more thorough evaluation. Several of these datasets (Woo et al., ICML 2024: https://arxiv.org/abs/2402.02592; Ansari et al., preprint 2024: https://arxiv.org/abs/2403.07815) from recent papers are publicly available. Running the evaluation on more datasets would give stronger evidence on the performance of the model. \n\n2. I also have concerns about the choice and setup of the baselines. The authors mention that they use the same experimental setup for each baseline, which includes number of training epochs. However, the number of training epochs are arguably an important hyperparameter for tuning a model. I would argue that the real-world performance is best compared by choosing the best possible setting for the baselines, rather than a uniform setting that might result into suboptimal performance of several baselines. Additionally, I think a strong baseline (PatchTST; Nie et al., ICLR 2023: https://arxiv.org/abs/2211.14730) is missing and should be included. When I compare the MSE/MAE of the models in this work with the models in the PatchTST paper, I find that the models in this work perform much weaker. This might suggest that the chosen setting leads to suboptimal performance of the baselines, which might affect the conclusion of this paper. For example: D-Linear has a 0.0764 MSE for the univariate setup in this work and a 0.056 MSE for ETTh1 in the PatchTST paper (Nie at al., ICLR 2023). This suggests that there is actually a better setting to run this baseline and this setting should be used for comparison. I noticed that both papers cite different sources of the datasets, but I checked briefly and at least the ETT datasets seem identical. There might be something different in the setup that I'm not aware off that also explains this difference. \n\nI consider this point critical and this needs to be addressed for me to consider to change my score. Specifically, I would ask the authors to revise the setup and use the settings for the baselines so they are comparable to the results in Nie at al. 2023. I would also like to ask the authors to include the PatchTST baseline. \n\n3. I also have concerns on the win matrix to compare the results. I would argue that in this setup it is not relevant how often RAFT outperforms other baselines, but rather how it compares to the strong baselines. Thus, the reported average win ratio of 86% is somewhat misleading and it would be more useful for the reader to report the win-rate to the next best model and the absolute/relative improvement in MSE/MAE when averaged over the datasets. I would also argue that in the context of the cited paper (Bahri et al., ICLR 2022), the win matrix is used over 65 datasets, while here only over 8 datasets with different forecast horizons. Hence, this introduces redundancy when aggregating the results in a win matrix. It is also not clear how ties are handled. Bahri et al., ICLR 2022 explicitly mention that ties are broken by a statistical test. How is this handled in this work? \n\nI would kindly ask the authors to report the win-rate to the next best model and the absolute/relative improvement in MSE/MAE when averaged over the datasets. I think this gives a more complete picture on performance of the model.", "questions": "1.  It is not clear how ties are handled in the statistical test. Bahri et al., ICLR 2022 explicitly mention that ties are broken by a statistical test. How is this handled in this work? \n\n2. Is the setup/datasets in this work noticeably different from the setup/datasets in the PatchTST paper? \n\n3. In several points of the paper the authors mention the inductive bias of RAFT and that it is more suited for forecasting. It is unclear to me what this inductive bias specifically means. In particular, there is one argument that existing models make i.i.d. assumptions and this is a limitation. How does RAFT overcome this limitation, especially with the empirical result that it is more effective when temporal correlation is lacking? I would kindly ask the authors what the specific inductive bias is that RAFT introduces and how it is a different inductive bias from existing models.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730299192426}], "openreview_url": "https://openreview.net/forum?id=GYwH71ugtC", "arxiv_id": "2411.08249", "paper_pdf": "papers/GYwH71ugtC.pdf", "paper_pdf_sha256": "2a1a9363dbe8d0ec13e633b7733f77976a1f3e5279681b5b2d2cfeab5d608a63", "paper_pdf_bytes": 1696089, "paper_pdf_source": "openreview", "code_url": "https://github.com/kutaytire/Retrieval-Augmented-Time-Series-Forecasting", "code_repository": "kutaytire/Retrieval-Augmented-Time-Series-Forecasting", "code_commit": "425e2af35797d0d32b63cfc554c7d97b3c54b390", "code_archive": "repos/GYwH71ugtC.zip", "code_archive_sha256": "f987205d1168765598ead579c0889258cc3fc1886efb3f06bddfc2d4e555bc31", "code_archive_bytes": 358038, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 284, "github_languages": {"Python": 109293}, "github_archived": false, "github_pushed_at": "2024-11-24T17:48:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/retrieval-augmented-time-series-forecasting"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9SwObx9Jdn", "year": 2024, "status": "rejected", "title": "Generation of Geodesics with Actor-Critic Reinforcement Learning to Predict Midpoints", "authors": ["Kazumi Kasaura"], "authorids": ["~Kazumi_Kasaura1"], "authors_source": "OpenReview API", "abstract": "Various tasks in the real world, such as path planning, can be reduced to the generation of geodesics on manifolds. For reinforcement learning to generate geodesics sequentially, we need to define rewards appropriately. To generate geodesics without any adjustment of rewards, we propose to use a modified version of sub-goal trees, called midpoint trees. While sub-goal trees consist of arbitrary intermediate points, midpoint trees consist of midpoints. In addition, we propose an actor-critic method to learn to predict midpoints and theoretically prove that, under mild assumptions, when the learning converges at the limit of infinite tree depth, the resulting policy generates exact midpoints.\nWe show experimentally that our proposed method outperforms existing methods in a certain path planning task.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "a2gVtoGXbG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3381/Reviewer_etxs"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a novel reinforcement learning framework, termed as 'midpoint tree', designed to recursively generate geodesics for path planning. The approach introduces an actor-critic learning method tailored to predict midpoint waypoints, facilitating the construction of paths in complex environments. The paper details both the theoretical underpinnings and the practical implications of the method, demonstrating its application to two distinct metrics, the Matsumoto metric and a car-like metric, and discussing its potential in fields such as image processing and physical systems modeling.", "review_text": "This paper presents a novel reinforcement learning framework, termed as 'midpoint tree', designed to recursively generate geodesics for path planning. The approach introduces an actor-critic learning method tailored to predict midpoint waypoints, facilitating the construction of paths in complex environments. The paper details both the theoretical underpinnings and the practical implications of the method, demonstrating its application to two distinct metrics, the Matsumoto metric and a car-like metric, and discussing its potential in fields such as image processing and physical systems modeling.", "strengths": "The paper presents a distinct approach to generating geodesics in reinforcement learning environments via a \"midpoint tree\" algorithm. The theoretical underpinnings are robust, complemented by a thorough experimental evaluation. The articulation is commendable, with the authors elucidating complex ideas succinctly. This work's originality and potential applicability are clear, indicating its prospective value in advancing research within reinforcement learning and robotics.", "weaknesses": "The paper lacks a broader range of examples to demonstrate the applicability of the method to more common robotic tasks like locomotion and manipulation planning. The experimental results, while encouraging, do not showcase a significant advantage over existing methods, raising questions about the practical benefits of the proposed approach. It requires certain assumptions that may not be present in typical robotic environments, such as the need for global coordinate systems and uniform sampling. The method might not be readily applicable to more complicated, dynamic environments.\n\nMore concretely:\n- The algorithm requires additional assumptions that may not be readily available or applicable in common robotic tasks, such as locomotion and manipulation planning. These assumptions include the need for global coordinate systems, obstacle-free environments, and environment-specific policy learning. The method's effectiveness is contingent on these conditions, which are not always present in more complex or dynamically changing real-world scenarios. Additionally, the challenge of generating globally optimal paths and dealing with the complexity of Finsler geodesics further limits its applicability to standard reinforcement learning tasks.\n\n- In the original wording, the paper mentions that the method \"only works well locally since we assume that manifolds have global coordinate systems and the continuous midpoint property may be satisfied only locally. For the generation of globally minimizing geodesics, we may have to divide manifolds, train policies for each local region and connect locally generated geodesics.\" It also states that \"the policy has to be learned for each environment. By modifying our method so that the actor and critic input information on environments, it may be possible to learn a policy applicable to different environments.\" These statements highlight the limitations regarding the need for specific geometric and topological assumptions that may not hold in typical RL tasks in robotics.\n\nA line of work on quasimetric distance for goal-conditioned RL seems related, which could provide important context and benchmarking. I'd be curious whether the proposed approach is related to them.\n- Tongzhou Wang et al., Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning, ICML 2023\n- Tongzhou Wang et al., On the Learning and Learnability of Quasimetrics, ICLR 2022", "questions": "- Can the authors provide additional examples where their method might be applicable, specifically within the realm of robotics tasks like locomotion and manipulation?\n- How does the proposed approach compare in terms of benefits and applicability to other realistic tasks, beyond what has been demonstrated in the paper?\n- Could the authors discuss the relationship and distinctions between their work and recent research on quasimetric learning for goal-conditioned RL?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel reinforcement learning framework, termed as 'midpoint tree', designed to recursively generate geodesics for path planning. The approach introduces an actor-critic learning method tailored to predict midpoint waypoints, facilitating the construction of paths in complex environments. The paper details both the theoretical underpinnings and the practical implications of the method, demonstrating its application to two distinct metrics, the Matsumoto metric and a car-like metric, and discussing its potential in fields such as image processing and physical systems modeling.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper presents a distinct approach to generating geodesics in reinforcement learning environments via a \"midpoint tree\" algorithm. The theoretical underpinnings are robust, complemented by a thorough experimental evaluation. The articulation is commendable, with the authors elucidating complex ideas succinctly. This work's originality and potential applicability are clear, indicating its prospective value in advancing research within reinforcement learning and robotics.", "weaknesses": "The paper lacks a broader range of examples to demonstrate the applicability of the method to more common robotic tasks like locomotion and manipulation planning. The experimental results, while encouraging, do not showcase a significant advantage over existing methods, raising questions about the practical benefits of the proposed approach. It requires certain assumptions that may not be present in typical robotic environments, such as the need for global coordinate systems and uniform sampling. The method might not be readily applicable to more complicated, dynamic environments.\n\nMore concretely:\n- The algorithm requires additional assumptions that may not be readily available or applicable in common robotic tasks, such as locomotion and manipulation planning. These assumptions include the need for global coordinate systems, obstacle-free environments, and environment-specific policy learning. The method's effectiveness is contingent on these conditions, which are not always present in more complex or dynamically changing real-world scenarios. Additionally, the challenge of generating globally optimal paths and dealing with the complexity of Finsler geodesics further limits its applicability to standard reinforcement learning tasks.\n\n- In the original wording, the paper mentions that the method \"only works well locally since we assume that manifolds have global coordinate systems and the continuous midpoint property may be satisfied only locally. For the generation of globally minimizing geodesics, we may have to divide manifolds, train policies for each local region and connect locally generated geodesics.\" It also states that \"the policy has to be learned for each environment. By modifying our method so that the actor and critic input information on environments, it may be possible to learn a policy applicable to different environments.\" These statements highlight the limitations regarding the need for specific geometric and topological assumptions that may not hold in typical RL tasks in robotics.\n\nA line of work on quasimetric distance for goal-conditioned RL seems related, which could provide important context and benchmarking. I'd be curious whether the proposed approach is related to them.\n- Tongzhou Wang et al., Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning, ICML 2023\n- Tongzhou Wang et al., On the Learning and Learnability of Quasimetrics, ICLR 2022", "questions": "- Can the authors provide additional examples where their method might be applicable, specifically within the realm of robotics tasks like locomotion and manipulation?\n- How does the proposed approach compare in terms of benefits and applicability to other realistic tasks, beyond what has been demonstrated in the paper?\n- Could the authors discuss the relationship and distinctions between their work and recent research on quasimetric learning for goal-conditioned RL?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699248533447}, {"id": "S3w3IKocGY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3381/Reviewer_ZvgN"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a modification of the sub-goal tree framework to use midpoints instead of arbitrary intermediate points and actor-critic instead of policy gradient for goal-conditioned reinforcement learning problems. With the two changes, the proposed method is able to generate equally divided waypoints and with better sample efficiency on deep trees. Theoretical proofs are given for the convergence of the proposed method. The proposed method shows comparable performance to baselines on several tasks with advantage of generating equally divided waypoints.", "review_text": "The paper proposes a modification of the sub-goal tree framework to use midpoints instead of arbitrary intermediate points and actor-critic instead of policy gradient for goal-conditioned reinforcement learning problems. With the two changes, the proposed method is able to generate equally divided waypoints and with better sample efficiency on deep trees. Theoretical proofs are given for the convergence of the proposed method. The proposed method shows comparable performance to baselines on several tasks with advantage of generating equally divided waypoints.", "strengths": "The paper is well-written and the method is well-motivated. The effectiveness of the proposed method is supported both theoretically and empirically. The generated waypoints with equal distances would be more useful than that of the previous method.", "weaknesses": "The novelty of the paper is not prominent compared to its base methods. \nThe experimental setting is a bit simplified. In section 6, the authors propose a penalty term to be added to deal with obstacles. Wondering how easy is it to generalize the proposed method to environments with obstacles.\nThe experiment results do not show clear performance improvements of the proposed method.", "questions": "Can we add some more explanation and justification on why the midpoint is not just a trivial extension of the existing method using arbitrary waypoints?\n\nCan we add more analysis on how the proposed method could be generalized to environments with obstacles?\n\nIn Figures 2 and 3, the proposed method does not show clear improvements compared to baselines. Is this expected? Can we add more explanations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a modification of the sub-goal tree framework to use midpoints instead of arbitrary intermediate points and actor-critic instead of policy gradient for goal-conditioned reinforcement learning problems. With the two changes, the proposed method is able to generate equally divided waypoints and with better sample efficiency on deep trees. Theoretical proofs are given for the convergence of the proposed method. The proposed method shows comparable performance to baselines on several tasks with advantage of generating equally divided waypoints.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper is well-written and the method is well-motivated. The effectiveness of the proposed method is supported both theoretically and empirically. The generated waypoints with equal distances would be more useful than that of the previous method.", "weaknesses": "The novelty of the paper is not prominent compared to its base methods. \nThe experimental setting is a bit simplified. In section 6, the authors propose a penalty term to be added to deal with obstacles. Wondering how easy is it to generalize the proposed method to environments with obstacles.\nThe experiment results do not show clear performance improvements of the proposed method.", "questions": "Can we add some more explanation and justification on why the midpoint is not just a trivial extension of the existing method using arbitrary waypoints?\n\nCan we add more analysis on how the proposed method could be generalized to environments with obstacles?\n\nIn Figures 2 and 3, the proposed method does not show clear improvements compared to baselines. Is this expected? Can we add more explanations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699041214388}, {"id": "mnVcO6btgi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3381/Reviewer_4ppY"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "summary": "This work focuses on path planning to generate geodesics in some manifold. It extends sub-goal tree framework (Jurgenson et al., 2020) to generate midpoints (equal distances to two given points), instead of any intermediate points. They train an actor to predict the midpoints, and a critic to predict the distance of two given points (s,g). It is also shown to converge to a unique optimal solution,  where the distance is given by some continuous approximation. The method is evaluated on two toy tasks to showcase its effectiveness over RL and previous planning approach.", "review_text": "This work focuses on path planning to generate geodesics in some manifold. It extends sub-goal tree framework (Jurgenson et al., 2020) to generate midpoints (equal distances to two given points), instead of any intermediate points. They train an actor to predict the midpoints, and a critic to predict the distance of two given points (s,g). It is also shown to converge to a unique optimal solution,  where the distance is given by some continuous approximation. The method is evaluated on two toy tasks to showcase its effectiveness over RL and previous planning approach.", "strengths": "The overall writing is rigorous, principled and looks solid work. But I am not sure of its significance.", "weaknesses": "Perhaps the motivation of this work can be better written. As the authors pointed out in their experiments, generating geodesic (path planning) can be simply tackled by RL by specifying a reward function related to the difference in distance. But it may have instability or other issue compared to path planning approaches. \n\nCould you give some explanation why Car-like task favors your approach, while Matsumoto task not? \n\nThe experiment scope is a bit narrow as only two toy tasks are evaluated. \n\nMinor: The description of methods in the experiments can be more complete – add a line of “ours” using Eq. 8 before “the following variants of our methods”.  The name “sequential RL” is a bit confusing as RL is sequential in nature. Perhaps “vanilla RL” or just “RL”, because your approach uses a non-conventional actor loss.", "questions": "I’m not familiar with path planning and differential manifold, so some of these comments are my educational guess.\n\n---- Post-rebuttal\n\nAfter reading the authors' response and other reviews, I think this work still requires more empirical evaluation on their approach. Thus, I lower my rating.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work focuses on path planning to generate geodesics in some manifold. It extends sub-goal tree framework (Jurgenson et al., 2020) to generate midpoints (equal distances to two given points), instead of any intermediate points. They train an actor to predict the midpoints, and a critic to predict the distance of two given points (s,g). It is also shown to converge to a unique optimal solution,  where the distance is given by some continuous approximation. The method is evaluated on two toy tasks to showcase its effectiveness over RL and previous planning approach.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The overall writing is rigorous, principled and looks solid work. But I am not sure of its significance.", "weaknesses": "Perhaps the motivation of this work can be better written. As the authors pointed out in their experiments, generating geodesic (path planning) can be simply tackled by RL by specifying a reward function related to the difference in distance. But it may have instability or other issue compared to path planning approaches. \n\nCould you give some explanation why Car-like task favors your approach, while Matsumoto task not? \n\nThe experiment scope is a bit narrow as only two toy tasks are evaluated. \n\nMinor: The description of methods in the experiments can be more complete – add a line of “ours” using Eq. 8 before “the following variants of our methods”.  The name “sequential RL” is a bit confusing as RL is sequential in nature. Perhaps “vanilla RL” or just “RL”, because your approach uses a non-conventional actor loss.", "questions": "I’m not familiar with path planning and differential manifold, so some of these comments are my educational guess.\n\n---- Post-rebuttal\n\nAfter reading the authors' response and other reviews, I think this work still requires more empirical evaluation on their approach. Thus, I lower my rating.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "code_of_conduct": "Yes"}, "tcdate": 1698273349070}, {"id": "Zvj7pLQUso", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3381/Reviewer_8vLp"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this paper, the authors study the problem of finding geodesics in general manifolds via reinforcement learning. The main idea is to divide the discovery task into smaller ones by predicting midpoints recursively. An actor-critic algorithm learns a policy to generate midpoints. Two empirical evaluations are provided to demonstrate the efficacy of the proposed algorithm.", "review_text": "In this paper, the authors study the problem of finding geodesics in general manifolds via reinforcement learning. The main idea is to divide the discovery task into smaller ones by predicting midpoints recursively. An actor-critic algorithm learns a policy to generate midpoints. Two empirical evaluations are provided to demonstrate the efficacy of the proposed algorithm.", "strengths": "1. Geodesics generation with reinforcement learning is a relatively under-explored research area. This work contributes by studying an actor-critic formulation and shows its effectiveness.\n\n2. A few design choices are explored, such as different variants of the actor loss. These results help illustrate some properties of the proposed algorithm.", "weaknesses": "1. The empirical evaluation environments are relatively artificial. I would expect some more practical tasks such as robotic motion planning to be more effective in demonstrating the significance of the contribution. \n\n2. The current baselines are all RL based. I think some classical motion planning algorithms should be included too, such as RRT (RRT*) and A* search. \n\n3. This is more of a clarification of the problem setting. It seems that the end goal of the learned policy is not necessarily finding the shortest path. The success criterion is stated as “all values of C(4) for two consecutive points are not greater than $\\epsilon$”, which does not imply that a path is the shortest. Is this correct? If so, this should be stated more clearly.", "questions": "1. Please provide some motivations for the definition of $C$ (Equation (4)). Also please explain what $df_x$ is in this definition.\n\n2. Why is Equation (5) hard to compute efficiently?\n\n3. How does one decide the depth parameter $D$ on Line 18 of Algorithm 1?\n\n4. In Equation (11), should the right-hand side be $d(x, y)$, the true distance rather than the local approximation? Either way, Equation (11) could use a more expanded explanation.\n\n5. In Proposition 2, what is $V_i$?\n\n6. In the Sequential Reinforcement Learning (Seq) baseline, why is the reward function (Equation (16)) scaled by $\\epsilon$? How does this decision affect the learning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors study the problem of finding geodesics in general manifolds via reinforcement learning. The main idea is to divide the discovery task into smaller ones by predicting midpoints recursively. An actor-critic algorithm learns a policy to generate midpoints. Two empirical evaluations are provided to demonstrate the efficacy of the proposed algorithm.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. Geodesics generation with reinforcement learning is a relatively under-explored research area. This work contributes by studying an actor-critic formulation and shows its effectiveness.\n\n2. A few design choices are explored, such as different variants of the actor loss. These results help illustrate some properties of the proposed algorithm.", "weaknesses": "1. The empirical evaluation environments are relatively artificial. I would expect some more practical tasks such as robotic motion planning to be more effective in demonstrating the significance of the contribution. \n\n2. The current baselines are all RL based. I think some classical motion planning algorithms should be included too, such as RRT (RRT*) and A* search. \n\n3. This is more of a clarification of the problem setting. It seems that the end goal of the learned policy is not necessarily finding the shortest path. The success criterion is stated as “all values of C(4) for two consecutive points are not greater than $\\epsilon$”, which does not imply that a path is the shortest. Is this correct? If so, this should be stated more clearly.", "questions": "1. Please provide some motivations for the definition of $C$ (Equation (4)). Also please explain what $df_x$ is in this definition.\n\n2. Why is Equation (5) hard to compute efficiently?\n\n3. How does one decide the depth parameter $D$ on Line 18 of Algorithm 1?\n\n4. In Equation (11), should the right-hand side be $d(x, y)$, the true distance rather than the local approximation? Either way, Equation (11) could use a more expanded explanation.\n\n5. In Proposition 2, what is $V_i$?\n\n6. In the Sequential Reinforcement Learning (Seq) baseline, why is the reward function (Equation (16)) scaled by $\\epsilon$? How does this decision affect the learning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697742831106}], "openreview_url": "https://openreview.net/forum?id=9SwObx9Jdn", "arxiv_id": "2407.01991", "paper_pdf": "papers/9SwObx9Jdn.pdf", "paper_pdf_sha256": "554508fbfe0ce295f7304355cf21d166a8718369080e27faccc162f6dff4588f", "paper_pdf_bytes": 757217, "paper_pdf_source": "openreview", "code_url": "https://github.com/omron-sinicx/midpoint_learning", "code_repository": "omron-sinicx/midpoint_learning", "code_commit": "6f8857ba0f9542dfb87c8941d5beeada55a39329", "code_archive": "repos/9SwObx9Jdn.zip", "code_archive_sha256": "b7db1d07fd8cd57bbc4de7cfa96e9bcfc718377c71716673613f0a9fee80bb46", "code_archive_bytes": 268713, "code_file_count": 79, "code_extensions": {".py": 78, ".sh": 1}, "github_disk_usage_kb": 225, "github_languages": {"Python": 338931, "Shell": 7148, "Dockerfile": 871}, "github_archived": false, "github_pushed_at": "2026-01-05T00:30:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generation-of-geodesics-with-actor-critic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "u6KhE9fapjX", "year": 2023, "status": "rejected", "title": "CAT: Collaborative Adversarial Training", "authors": ["xingbin liu", "Huafeng Kuang", "Xianming Lin", "GUANNAN JIANG", "YONGJIAN WU", "Xi Wang", "Rongrong Ji"], "authorids": ["~xingbin_liu1", "~Huafeng_Kuang1", "~Xianming_Lin1", "~GUANNAN_JIANG1", "~YONGJIAN_WU2", "~Xi_Wang12", "~Rongrong_Ji5"], "authors_source": "OpenReview API", "abstract": "Adversarial training can improve the robustness of neural networks. Previous adversarial training methods focus on a single training strategy and do not consider the collaboration between different training strategies. In this paper, we find different adversarial training methods have distinct robustness for sample instances. For example, an instance can be correctly classified by a model trained using standard adversarial training (AT) but not by a model trained using TRADES, and vice versa.  Based on this phenomenon, we propose a collaborative adversarial training framework to improve the robustness of neural networks. Specifically, we simultaneously use different adversarial training methods to train two robust models from scratch. We input the adversarial examples generated by each network to the peer network and use the logit of the peer network to guide the training of its network. Collaborative Adversarial Training (CAT) can improve both robustness and accuracy. Finally, Extensive experiments on CIFAR-10 and CIFAR-100 validated the effectiveness of our method.\nCAT achieved new state-of-the-art robustness without using any additional data on CIFAR-10 under the Auto-Attack benchmark.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "MSliCqtyK3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6149/Reviewer_MzNC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors made an observation that different adversarial defense strategies make different mistakes. Based on this observation, they proposed collaborative adversarial training, where simultaneously train two robust models. The objective of collaborative adversarial training is to minimize the symmetric KL-divergence between the logits of the first and second models. In the experiments, the authors showed improved robustness against AutoAttack on CIFAR-10 and CIFAR-100 datasets.", "review_text": "The authors proposed a collaborative adversarial training method that combines two defense methods. The approach is interesting and marginally novel. It improves robustness upon TRADES and ALP baselines. However, the paper contains many typos, the empirical comparison does not include more recent SOTA methods, and some citations (such as ensemble adversarial training) are missing. ", "strengths": "### Strengths\n\n- Proposed a framework for collaborative adversarial training, where two robust models are trained jointly.\n- Demonstrated promising experimental results on CIFAR-10 and CIFAR-100 benchmarks.\n\n### Weaknesses\n\n- The increased computational cost of training two robust models.\n- The lack of a theoretical analysis of the proposed method. It is not clear why the proposed objective will guide the training toward a more robust model compared to the single model, single attack, and single objective training.\n- The lack of more detailed experimental analysis and other baselines for comparison.\n\n### Questions\n- Can collaborative adversarial training be affected by catastrophic/robust overfitting?\n- Does collaborative adversarial training perform better than training one robust model with multiple attacks and/or multiple weighted objectives?\n- Can you include experiments with more than 2 defenses?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors made an observation that different adversarial defense strategies make different mistakes. Based on this observation, they proposed collaborative adversarial training, where simultaneously train two robust models. The objective of collaborative adversarial training is to minimize the symmetric KL-divergence between the logits of the first and second models. In the experiments, the authors showed improved robustness against AutoAttack on CIFAR-10 and CIFAR-100 datasets.", "strength_and_weaknesses": "### Strengths\n\n- Proposed a framework for collaborative adversarial training, where two robust models are trained jointly.\n- Demonstrated promising experimental results on CIFAR-10 and CIFAR-100 benchmarks.\n\n### Weaknesses\n\n- The increased computational cost of training two robust models.\n- The lack of a theoretical analysis of the proposed method. It is not clear why the proposed objective will guide the training toward a more robust model compared to the single model, single attack, and single objective training.\n- The lack of more detailed experimental analysis and other baselines for comparison.\n\n### Questions\n- Can collaborative adversarial training be affected by catastrophic/robust overfitting?\n- Does collaborative adversarial training perform better than training one robust model with multiple attacks and/or multiple weighted objectives?\n- Can you include experiments with more than 2 defenses?", "clarity,_quality,_novelty_and_reproducibility": "The paper is easy to understand and follow. The authors should proofread the paper to remove some of the typos (e.g. \"the TARDES-trained network). The idea of combining multiple defenses is not novel. There are several methods for training an ensemble of adversarially-trained models to improve robustness [1, 2], which were not cited nor compared with. The authors should relate the proposed method to ensemble methods for adversarial training. In comparison, the authors should consider using some of the more recent and stronger baselines for comparison. \n\n[1] Tramèr, F., Kurakin, A., Papernot, N., Goodfellow, I., Boneh, D., & McDaniel, P. (2018). Ensemble adversarial training: Attacks and defenses. In , International Conference on Learning Representations (pp. ). : .\n[2] Pang, T., Xu, K., Du, C., Chen, N., & Zhu, J. (2019). Improving adversarial robustness via promoting ensemble diversity. In K. Chaudhuri, & R. Salakhutdinov, Proceedings of the 36th International Conference on Machine Learning (pp. 4970–4979). Long Beach, California, USA: PMLR.", "summary_of_the_review": "The authors proposed a collaborative adversarial training method that combines two defense methods. The approach is interesting and marginally novel. It improves robustness upon TRADES and ALP baselines. However, the paper contains many typos, the empirical comparison does not include more recent SOTA methods, and some citations (such as ensemble adversarial training) are missing. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667360005853}, {"id": "YlNux05o_V", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6149/Reviewer_a6Jb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors observe a phenomenon that robust models adversarial trained by different methods (like AT and TRADES) have different reactions to the given input. Motivated by this, they proposed to perform collaboratively adversarial training to improve the robustness by training two neural networks from scratch.\n", "review_text": "Pls see the above comments.", "strengths": "Strength:\n\n1. The idea is straightforward, and the paper writing is easy to follow, which simply ensembles the knowledge from different adversarial training models.\n\n2. The experiments are extensive among a variety of datasets and networks, and the evaluation achieves state-of-the-art results.\n\n\nWeaknesses:\n\n1. The idea of the proposed method is quite lacks novelty. Similar ideas of collaboratively fusing knowledge have been proposed and studied for improving robustness [1][2][3].\n\n2. The paper does not theoretically analyse the benefits brought by fusing different adversarial training methods. The trivial experimental results do not clearly support their claims, the author seems to simply blend them together to get improvement without any further analysis.\n\n3. Is that necessary to train two different robust neural networks from scratch? Efficiency is the main challenge in adversarial training, and now the authors propose to scarify by doubling the time, making it worse. \n\n4. What if we train one neural network and fuse the knowledge by using different adversarial examples generated by different methods? This is a strong baseline that should be concluded for comparison.\n\n5. What is ALP? The abbreviation occurs suddenly in the experiment part without any reference or explanation.\n\n[1] Toward Learning Robust and Invariant Representations with Alignment Regularization and Data Augmentation\n\n[2] Improving adversarial robustness by learning shared information | Elsevier Enhanced Reader\n\n[3] Prior-Guided Adversarial Initialization for Fast Adversarial Training", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors observe a phenomenon that robust models adversarial trained by different methods (like AT and TRADES) have different reactions to the given input. Motivated by this, they proposed to perform collaboratively adversarial training to improve the robustness by training two neural networks from scratch.\n", "strength_and_weaknesses": "Strength:\n\n1. The idea is straightforward, and the paper writing is easy to follow, which simply ensembles the knowledge from different adversarial training models.\n\n2. The experiments are extensive among a variety of datasets and networks, and the evaluation achieves state-of-the-art results.\n\n\nWeaknesses:\n\n1. The idea of the proposed method is quite lacks novelty. Similar ideas of collaboratively fusing knowledge have been proposed and studied for improving robustness [1][2][3].\n\n2. The paper does not theoretically analyse the benefits brought by fusing different adversarial training methods. The trivial experimental results do not clearly support their claims, the author seems to simply blend them together to get improvement without any further analysis.\n\n3. Is that necessary to train two different robust neural networks from scratch? Efficiency is the main challenge in adversarial training, and now the authors propose to scarify by doubling the time, making it worse. \n\n4. What if we train one neural network and fuse the knowledge by using different adversarial examples generated by different methods? This is a strong baseline that should be concluded for comparison.\n\n5. What is ALP? The abbreviation occurs suddenly in the experiment part without any reference or explanation.\n\n[1] Toward Learning Robust and Invariant Representations with Alignment Regularization and Data Augmentation\n\n[2] Improving adversarial robustness by learning shared information | Elsevier Enhanced Reader\n\n[3] Prior-Guided Adversarial Initialization for Fast Adversarial Training", "clarity,_quality,_novelty_and_reproducibility": "Pls see the above comments.", "summary_of_the_review": "Pls see the above comments.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667084025896}, {"id": "kFZXZfc4Wy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6149/Reviewer_df3y"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "Summary.\n\nThis paper is dedicated to developing improved adversarial training methods. The authors are inspired by the observations of prediction discrepancy from different adversarially trained models. Specifically, they introduce a collaborative adversarial training framework (called CAT) to improve the robustness. CAT inputs the adversarial example generated by each network to the peer network and uses the peer networks' logit to guide its training. Experiments are conducted on CIFAR-10 and CIFAR-100.", "review_text": "Incremental ideas with insufficient studies.", "strengths": "Pros.\n\n1. The paper is well-written and easy to follow.\n2. Robustness is evaluated by diverse adversarial attackers beyond PGD.\n\n\n\nCons. \n\n1. The collaborative training technique can be regarded as a \"soft data augmentation\" since it will use the input sample and associated predictions from the peer networks. Therefore, (a) the authors need to compare with data augmentation approaches in adversarial training like \"Data Augmentation Can Improve Robustness\"; (b) it is hard to say that the comparison in this paper is fair since CAT use at least 2 times resource cost than other approaches.\n2. Only small-scale datasets are considered like CIFAR-10 and CIFAR-100. It is insufficient to support the effectiveness of the proposed methods. Larger datasets like ImageNet are required.\n3. Meanwhile, the network backbones for evaluation are also limited. More architectures like VGG and MobileNet are needed.\n4. The provided method is motivated by the prediction discrepancy between different robustified networks. However, it is unconvincing between several important baselines are missing. For example, the prediction intersection results for the same AT method with different random seeds or different random start, are required. Whether the prediction difference between TRADES and AT is larger than the results between two AT runs?\n5. More ablations are needed. For instance, how about involving more attackers (e.g., FAT, CW, auto attack) in collaborative training? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Summary.\n\nThis paper is dedicated to developing improved adversarial training methods. The authors are inspired by the observations of prediction discrepancy from different adversarially trained models. Specifically, they introduce a collaborative adversarial training framework (called CAT) to improve the robustness. CAT inputs the adversarial example generated by each network to the peer network and uses the peer networks' logit to guide its training. Experiments are conducted on CIFAR-10 and CIFAR-100.", "strength_and_weaknesses": "Pros.\n\n1. The paper is well-written and easy to follow.\n2. Robustness is evaluated by diverse adversarial attackers beyond PGD.\n\n\n\nCons. \n\n1. The collaborative training technique can be regarded as a \"soft data augmentation\" since it will use the input sample and associated predictions from the peer networks. Therefore, (a) the authors need to compare with data augmentation approaches in adversarial training like \"Data Augmentation Can Improve Robustness\"; (b) it is hard to say that the comparison in this paper is fair since CAT use at least 2 times resource cost than other approaches.\n2. Only small-scale datasets are considered like CIFAR-10 and CIFAR-100. It is insufficient to support the effectiveness of the proposed methods. Larger datasets like ImageNet are required.\n3. Meanwhile, the network backbones for evaluation are also limited. More architectures like VGG and MobileNet are needed.\n4. The provided method is motivated by the prediction discrepancy between different robustified networks. However, it is unconvincing between several important baselines are missing. For example, the prediction intersection results for the same AT method with different random seeds or different random start, are required. Whether the prediction difference between TRADES and AT is larger than the results between two AT runs?\n5. More ablations are needed. For instance, how about involving more attackers (e.g., FAT, CW, auto attack) in collaborative training? ", "clarity,_quality,_novelty_and_reproducibility": "The clarity is good.\n\nQuality and novelty are low.\n\nThe reproducibility is unknown.", "summary_of_the_review": "Incremental ideas with insufficient studies.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667009154963}, {"id": "dx9_BEzcIX", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6149/Reviewer_JNu7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a Collaborative Adversarial Training (CAT) framework, which is based on the observation that robust models trained with different training methods have different performance on different instances, despite the models having similar accuracies. This gives rise to the proposed collaborative approach: Given the setup of two networks being trained with two different training methods, an adversarial sample generated by one network is fed to the other network to obtain the corresponding logit, which is then utilized to guide the learning of the first network. The evaluation is performed on two benchmark datasets CIFAR10 and CIFAR100 against a number of baselines to test the effectiveness of the approach in improving model robustness. The authors also compare their method to various knowledge distillation techniques which work on a similar principle.", "review_text": "This work proposes a training regime to combine the benefits of different adversarial training algorithms, which is interesting and novel. However, there is little motivation on why this works, and how the trained models are different. Including some important ablations (mentioned above) can improve clarity on the impact of the proposed method. ", "strengths": "Strengths-\n\n- The proposed approach is intuitive and simple to implement.\n- The work shows improved robustness compared to the baseline models trained non-collaboratively.\n- The paper is clear and well-written. \n\nWeaknesses/ Questions -\n\n- The proposed method involves higher complexity when compared to baselines - 2x computational cost, 2x memory overhead, more hyperparamaters. \n- Could the authors elaborate further on why and how CAT improves performance? How are these adversarial training methods different in the first place - and why does combining them result in improvements? \n- Could the authors share the confusion matrix of (correct and incorrect) predictions from TRADES and PGD-AT (or ALP)? This would better motivate the approach.\n- The best results on Robustbench are actually obtained by applying Adversarial Weight perturbation (AWP). Could the proposed method TRADES-ALP also be integrated with AWP? It would be useful to compare gains of AWP and no-AWP runs. \n- Could the authors share an ablation of TRADES-TRADES, where the same method (TRADES) is used for training both models?\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a Collaborative Adversarial Training (CAT) framework, which is based on the observation that robust models trained with different training methods have different performance on different instances, despite the models having similar accuracies. This gives rise to the proposed collaborative approach: Given the setup of two networks being trained with two different training methods, an adversarial sample generated by one network is fed to the other network to obtain the corresponding logit, which is then utilized to guide the learning of the first network. The evaluation is performed on two benchmark datasets CIFAR10 and CIFAR100 against a number of baselines to test the effectiveness of the approach in improving model robustness. The authors also compare their method to various knowledge distillation techniques which work on a similar principle.", "strength_and_weaknesses": "Strengths-\n\n- The proposed approach is intuitive and simple to implement.\n- The work shows improved robustness compared to the baseline models trained non-collaboratively.\n- The paper is clear and well-written. \n\nWeaknesses/ Questions -\n\n- The proposed method involves higher complexity when compared to baselines - 2x computational cost, 2x memory overhead, more hyperparamaters. \n- Could the authors elaborate further on why and how CAT improves performance? How are these adversarial training methods different in the first place - and why does combining them result in improvements? \n- Could the authors share the confusion matrix of (correct and incorrect) predictions from TRADES and PGD-AT (or ALP)? This would better motivate the approach.\n- The best results on Robustbench are actually obtained by applying Adversarial Weight perturbation (AWP). Could the proposed method TRADES-ALP also be integrated with AWP? It would be useful to compare gains of AWP and no-AWP runs. \n- Could the authors share an ablation of TRADES-TRADES, where the same method (TRADES) is used for training both models?\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and well-written. Hyperparameters and other training settings have been discussed. The authors are encouraged to make the code available too. ", "summary_of_the_review": "This work proposes a training regime to combine the benefits of different adversarial training algorithms, which is interesting and novel. However, there is little motivation on why this works, and how the trained models are different. Including some important ablations (mentioned above) can improve clarity on the impact of the proposed method. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666987502718}], "openreview_url": "https://openreview.net/forum?id=u6KhE9fapjX", "arxiv_id": "2303.14922", "paper_pdf": "papers/u6KhE9fapjX.pdf", "paper_pdf_sha256": "b29f8f5c386abbef054c0014f4c797abb4d3f1fef594758b17a33be1e8b880bf", "paper_pdf_bytes": 324232, "paper_pdf_source": "openreview", "code_url": "https://github.com/liuxingbin/CAT", "code_repository": "liuxingbin/CAT", "code_commit": "2325fa16af644c1a6e83bf4761d4a8dece86763d", "code_archive": "repos/u6KhE9fapjX.zip", "code_archive_sha256": "9b1d3dab839509c451199e9a98032d8038a38a8333789c066b4f090662fdb949", "code_archive_bytes": 496510, "code_file_count": 12, "code_extensions": {".py": 10, ".sh": 2}, "github_disk_usage_kb": 486, "github_languages": {"Python": 53003, "Shell": 660}, "github_archived": false, "github_pushed_at": "2023-05-31T07:07:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cat-collaborative-adversarial-training"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "57T1ctyxtP", "year": 2022, "status": "rejected", "title": "Structured Stochastic Gradient MCMC", "authors": ["Antonios Alexos", "Alex James Boyd", "Stephan Mandt"], "authorids": ["~Antonios_Alexos1", "~Alex_James_Boyd1", "~Stephan_Mandt1"], "authors_source": "OpenReview API", "abstract": "Stochastic gradient Markov Chain Monte Carlo (SGMCMC) is considered the gold standard for Bayesian inference in large-scale models, such as Bayesian neural networks. Since practitioners face speed versus accuracy tradeoffs in these models, variational inference (VI) is often the preferable option. Unfortunately, VI makes strong assumptions on both the factorization and functional form of the posterior. In this work, we propose a new non-parametric variational approximation that makes no assumptions about the approximate posterior's functional form and allows practitioners to specify the exact dependencies the algorithm should respect or break. The approach relies on a new Langevin-type algorithm that operates on a modified energy function, where parts of the latent variables are averaged over samples from earlier iterations of the Markov chain. This way, statistical dependencies can be broken in a controlled way, allowing the chain to mix faster. This scheme can be further modified in a ``dropout'' manner, leading to even more scalability. By implementing the scheme on a ResNet-20 architecture, we obtain better predictive likelihoods and faster mixing time than full SGMCMC.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "O5jweuJO3iR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1138/Reviewer_NCgf"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a new hybrid method between MCMC and VI. The main idea of the paper is construction of a new energy function which allows to speed up the sampling significantly in comparison to the SGMCMC (stochastic gradient MCMC). An additional modification of the proposed algorithm is done by adopting a drop-out inspired approximation which allows for even better scalability. ", "review_text": "Overall I enjoyed reading the paper and I think the proposed method has merit, but I also do have some concerns.\n\n1)\tThe paper does not have proof of convergence for the proposed algorithm, a proof of convergence would have made the paper stronger. \n\n2)\tAlthough the paper deals with multivariate models and a potentially large number of parameters, the evaluation metrics are univariate. For example, the ESS in table 1. The authors should report the results for multivariate ESS, especially considering that on, for example, CIFAR-10 the results are only marginally better than SGHMC.  \n\n3)\tPlease note, that in the recent paper [1] it was shown that ESS and “other standard MCMC metrics which do not account for sample bias are not appropriate diagnostic tools for SGMCMC”. This paper [1] proposes to use the kernel Stein discrepancy metric for the assessment of SGMCMC methods. I would recommend the authors include it in their analysis. \n\n4)\tOn page 4 the authors assume that “This partitioning structure is assumed to be known a priori.” when talking about factorization of the parameters into mutually independent groups. While it can be naturally assumed in some models, is this a straightforward assumption of BNN, since the proposed algorithm is in particular motivated by application to BNN?\n\n5)\tOn page 4 the authors mention: “While it is unlikely to have the procedure initialize to a stationary state, we observe in practice that our scheme both tends to converge towards and remain in a stationary state. “ I would encourage the authors to directly refer to a figure or table which illustrates this claim. \n\n6)\tOn page 3 it states: “The gold standard for approximating the entire posterior distribution is by deploying Markov chain Monte Carlo (MCMC) algorithms. These methods work by producing an empirical distribution of samples through a random walk in parameter space. “  \nI would argue that this is not a correct statement as not all MCMC methods deploy random walk behavior (for example, not HMC or NUTS). \n\n7)\tIn the appendix on page 17 you mention “We observe that as we break dependencies we capture similar uncertainty intervals. ”. To me it looks like the confidence intervals actually change quite a bit, especially left/right of figure 4a and left/right of figures 4b and 4c, the confidence intervals become quite a bit more narrow.  Can this again be attributed to VI behavior? \n\n8)\tIn Figure 6 the average accuracy for the last model in the list is quite surprising. Do you have any intuition about that? \n\n9)\tPlease, check your references, both formatting, and whether some of the papers you cite have been published in the meantime. \n\n10) On page 3 the sentence \" The answer is affirmative and will be answered as follows\" probably can be re-written in a better way.  \n\n[1] Nemeth, Christopher, and Paul Fearnhead. \"Stochastic gradient Markov chain Monte Carlo.\" Journal of the American Statistical Association 116.533 (2021): 433-450.\n\nUPD: I increased my score for correctness to 3 after the authors reply and overall score to 5. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new hybrid method between MCMC and VI. The main idea of the paper is construction of a new energy function which allows to speed up the sampling significantly in comparison to the SGMCMC (stochastic gradient MCMC). An additional modification of the proposed algorithm is done by adopting a drop-out inspired approximation which allows for even better scalability. ", "main_review": "Overall I enjoyed reading the paper and I think the proposed method has merit, but I also do have some concerns.\n\n1)\tThe paper does not have proof of convergence for the proposed algorithm, a proof of convergence would have made the paper stronger. \n\n2)\tAlthough the paper deals with multivariate models and a potentially large number of parameters, the evaluation metrics are univariate. For example, the ESS in table 1. The authors should report the results for multivariate ESS, especially considering that on, for example, CIFAR-10 the results are only marginally better than SGHMC.  \n\n3)\tPlease note, that in the recent paper [1] it was shown that ESS and “other standard MCMC metrics which do not account for sample bias are not appropriate diagnostic tools for SGMCMC”. This paper [1] proposes to use the kernel Stein discrepancy metric for the assessment of SGMCMC methods. I would recommend the authors include it in their analysis. \n\n4)\tOn page 4 the authors assume that “This partitioning structure is assumed to be known a priori.” when talking about factorization of the parameters into mutually independent groups. While it can be naturally assumed in some models, is this a straightforward assumption of BNN, since the proposed algorithm is in particular motivated by application to BNN?\n\n5)\tOn page 4 the authors mention: “While it is unlikely to have the procedure initialize to a stationary state, we observe in practice that our scheme both tends to converge towards and remain in a stationary state. “ I would encourage the authors to directly refer to a figure or table which illustrates this claim. \n\n6)\tOn page 3 it states: “The gold standard for approximating the entire posterior distribution is by deploying Markov chain Monte Carlo (MCMC) algorithms. These methods work by producing an empirical distribution of samples through a random walk in parameter space. “  \nI would argue that this is not a correct statement as not all MCMC methods deploy random walk behavior (for example, not HMC or NUTS). \n\n7)\tIn the appendix on page 17 you mention “We observe that as we break dependencies we capture similar uncertainty intervals. ”. To me it looks like the confidence intervals actually change quite a bit, especially left/right of figure 4a and left/right of figures 4b and 4c, the confidence intervals become quite a bit more narrow.  Can this again be attributed to VI behavior? \n\n8)\tIn Figure 6 the average accuracy for the last model in the list is quite surprising. Do you have any intuition about that? \n\n9)\tPlease, check your references, both formatting, and whether some of the papers you cite have been published in the meantime. \n\n10) On page 3 the sentence \" The answer is affirmative and will be answered as follows\" probably can be re-written in a better way.  \n\n[1] Nemeth, Christopher, and Paul Fearnhead. \"Stochastic gradient Markov chain Monte Carlo.\" Journal of the American Statistical Association 116.533 (2021): 433-450.\n\nUPD: I increased my score for correctness to 3 after the authors reply and overall score to 5. ", "summary_of_the_review": "The paper takes a relevant direction in research trying to make inference more scalable for BNN (and other models) by combining MCMC and VI in a stochastic gradient MCMC framework. The paper, however, does not provide theoretical guarantees of convergence while empirical evaluation metrics were not chosen optimally (only univariate metrics which are also not optimal for SGMCMC methods as shownin [1]). Consequently, I would encourage the authors to address the critical points and resubmit the paper in the future. \n\n[1] Nemeth, Christopher, and Paul Fearnhead. \"Stochastic gradient Markov chain Monte Carlo.\" Journal of the American Statistical Association 116.533 (2021): 433-450.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635884119573}, {"id": "EnZJyYn8k4m", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1138/Reviewer_zQoE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper proposes a hybrid SGMCMC algorithm, a Langevin-type algorithm that operates on a modified energy function based on the variational inference (VI) approach. ", "review_text": "1. The proposed algorithm is not well motivated: the VI approach aims to obtain a point estimate, SGMCMC aims to draw a sequence of samples from the target posterior, but the proposed algorithm is to simulate a sequence of samples from a modified distribution.\nIt is unclear how well the modified distribution approximates the target distribution and how useful the pseudo-posterior samples are for statistical inference of the target distribution.   \n\n2. The comparison with the existing SGMCMC algorithm is not convincing. It seems that they are ``comparable in wall clock time'', while the baselines pSGLD and SGHMC might not be the state-of-the-art.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a hybrid SGMCMC algorithm, a Langevin-type algorithm that operates on a modified energy function based on the variational inference (VI) approach. ", "main_review": "1. The proposed algorithm is not well motivated: the VI approach aims to obtain a point estimate, SGMCMC aims to draw a sequence of samples from the target posterior, but the proposed algorithm is to simulate a sequence of samples from a modified distribution.\nIt is unclear how well the modified distribution approximates the target distribution and how useful the pseudo-posterior samples are for statistical inference of the target distribution.   \n\n2. The comparison with the existing SGMCMC algorithm is not convincing. It seems that they are ``comparable in wall clock time'', while the baselines pSGLD and SGHMC might not be the state-of-the-art.  ", "summary_of_the_review": "This paper proposes a hybrid SGMCMC algorithm, but the underlying theory is not fully developed and the performance of the algorithm is not fully explored. ", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635874948335}, {"id": "PQkks9wh3_q", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1138/Reviewer_qXUG"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work proposes using a structured variational approximation for stochastic gradient Markov Chain Monte Carlo. This allows someone to choose a factorization for the variational distribution (which factorization is best is unclear; several are studied). Analogously to coordinate ascent variational inference, the authors show that the best approximation is the Boltzmann energy function marginalized over the complements of every parameter group. \n\nHowever, this structured approximation is computationally expensive and requires the same number of evaluations of the approximation as there are parameter groups. This computational burden is alleviated by a dropout scheme, where instead of sampling from every parameter groups, parameters are masked using a dropout distribution, and the number of stochastic masks is a hyperparameter that controls regularization and fidelity to the structure imposed in the factorization. \n\nExperiments show that this is a viable way to impose structure on a variational distribution, and that mixing times are improved.", "review_text": "This is a very well-written, clear paper, with nice experiments to guide intuition. Congratulations! The effort shows. \n\nIf the paper \"attempt[s] to hybridize MCMC and VI\", why not compare to structured variational inference? Is the only difference between this approach and variational inference the additional noise in the gradient update?\n\nIf so, it could be stated in a sentence, which would help me be less confused. After all, Equation (4) is the same objective function, theorem 1 is roughly the same proof as coordinate ascent variational inference (where the parameter groups M are simply coordinates i), then these are more of an 'observation' than a theorem. But the key insight of using the coordinate ascent variational inference for SGMCMC is valuable and the main point of thh paper, and calling it a theorem may be helpful for readers. So it's up to the authors for deciding what's best - just wanted to point out that these connections and derivation analogs could be made more clear, which would make the method easier to understand and exposition more straightforward.\n\n- The most confusing sentence for me was belot Eq. 6: \"in practice, at timestep t...\". Maybe describe how \\hat q is composed of samples from previous timesteps. \n\n- small wording choices: \"completely joint\", \"well-approximates\", \"solid distributional assumptions\"\n\n- in equation 8, rho_i is not defined\n\nI also appreciate the authors including clean code as a supplementary zip file for reproducibility.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes using a structured variational approximation for stochastic gradient Markov Chain Monte Carlo. This allows someone to choose a factorization for the variational distribution (which factorization is best is unclear; several are studied). Analogously to coordinate ascent variational inference, the authors show that the best approximation is the Boltzmann energy function marginalized over the complements of every parameter group. \n\nHowever, this structured approximation is computationally expensive and requires the same number of evaluations of the approximation as there are parameter groups. This computational burden is alleviated by a dropout scheme, where instead of sampling from every parameter groups, parameters are masked using a dropout distribution, and the number of stochastic masks is a hyperparameter that controls regularization and fidelity to the structure imposed in the factorization. \n\nExperiments show that this is a viable way to impose structure on a variational distribution, and that mixing times are improved.", "main_review": "This is a very well-written, clear paper, with nice experiments to guide intuition. Congratulations! The effort shows. \n\nIf the paper \"attempt[s] to hybridize MCMC and VI\", why not compare to structured variational inference? Is the only difference between this approach and variational inference the additional noise in the gradient update?\n\nIf so, it could be stated in a sentence, which would help me be less confused. After all, Equation (4) is the same objective function, theorem 1 is roughly the same proof as coordinate ascent variational inference (where the parameter groups M are simply coordinates i), then these are more of an 'observation' than a theorem. But the key insight of using the coordinate ascent variational inference for SGMCMC is valuable and the main point of thh paper, and calling it a theorem may be helpful for readers. So it's up to the authors for deciding what's best - just wanted to point out that these connections and derivation analogs could be made more clear, which would make the method easier to understand and exposition more straightforward.\n\n- The most confusing sentence for me was belot Eq. 6: \"in practice, at timestep t...\". Maybe describe how \\hat q is composed of samples from previous timesteps. \n\n- small wording choices: \"completely joint\", \"well-approximates\", \"solid distributional assumptions\"\n\n- in equation 8, rho_i is not defined\n\nI also appreciate the authors including clean code as a supplementary zip file for reproducibility.\n", "summary_of_the_review": "This is a clearly written paper with a tightly scoped contribution. I advocate for acceptance, and looking forward to follow up work.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635865102226}, {"id": "Evlobf5AJvr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1138/Reviewer_AW5Q"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The author proposed a framework to incorporate the independence structure into the posterior inference for faster-mixing speed. To achieve that, the author designed two specific algorithms called S-SGMCMC and S_d-SGMCMC, respectively. Specifically, S-SGMCMC consisted of the following steps. First, the target random variables ($\\theta$) are gathered into mutually independent groups. Then, a modified energy function is derived by minimizing the KL divergence between the posterior $q$ and target $p(\\theta|D)$. The last step is to apply a standard SGMCMC method to draw samples from the resulting modified energy function. Further, the author also built a connection between dropout and the modified energy function, which results in a structure dropout SGMCMC (S_d-SGMCMC) with better scalability. \nThe author claimed the resulting algorithms achieved faster mixing speed and better classification accuracy when applied to real-world classification tasks. \n", "review_text": "## Strength\nPersonally, I find the proposed method interesting. The paper is written clearly and easy to follow, and the overall idea is easy to understand. I have checked the proofs of the theorem and they seem to be correct. To support the claims, the proposed methods were applied to real-world large data set to confirm their advantages compared to the standard full SGMCMC with some ablation studies, although I still think there is room for improvement. \n\n## Weakness and concerns\nAlthough the proposed method is easy to follow, I still find some parts unclear and elaborations are needed. First, the author claims that building the independent structure into the posterior can help improve the mixing speed (bottom of page 1, the first paragraph in the conclusion, etc.). However, I cannot see direct reasoning behind those claims, could you elaborate more on this? In section 6.1, I am not sure I understand the reasoning behind the better performances of S_d-SGMCMC. Can you elaborate more on \"regularizing the model\"?\nOn page 4, why $\\tilde{\\theta}^{(t,i)}$ is composed of samples from previous timesteps? I thought the author mentioned that a **single** sample $\\tilde{\\theta}^{(t,i)}$ from the **current** timestep is used for Monte Carlo approximation. I recommend putting more details in algorithm 1. For example, elaborating how $\\tilde{\\theta}^t_{\\neg i}$ is drawn from $\\hat{q}^t_{\\neg i}$. \nIn addition, why do appendix B and D use the same title? Can you merge them together?\n\nAnother concern is the theoretical soundness of proposed algorithms. For S-SGMCMC, the Monte Carlo estimation of the modified energy function requires previous samples. Does this break the Markov assumption of SGMCMC since it should only depend on the **current** samples? For example, when the sampler does not reach the stationary stage, $\\hat{q}^t$ still evolves with time, thus, the samples from the previous timesteps are not from $\\hat{q}^t$. \nApart from the stationary distribution of S-SGMCMC, I also wonder what is the stationary distribution (if exists) of the dropout version? How different is the stationary distribution of the dropout version compared to the optimal $q$? Any theoretical guarantees on the correctness of the dropout version? \n\nIn terms of the empirical evaluation, I wonder about the uncertainty quantification ability of the proposed algorithms, which is an important metric for SGMCMC methods, especially since the proposed methods also seem to underestimate the posterior variance. In addition, in the abstract, the author mentioned better predictive likelihoods. However, I can only find the accuracy metric in the experiment, which is different from the predictive likelihood. \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The author proposed a framework to incorporate the independence structure into the posterior inference for faster-mixing speed. To achieve that, the author designed two specific algorithms called S-SGMCMC and S_d-SGMCMC, respectively. Specifically, S-SGMCMC consisted of the following steps. First, the target random variables ($\\theta$) are gathered into mutually independent groups. Then, a modified energy function is derived by minimizing the KL divergence between the posterior $q$ and target $p(\\theta|D)$. The last step is to apply a standard SGMCMC method to draw samples from the resulting modified energy function. Further, the author also built a connection between dropout and the modified energy function, which results in a structure dropout SGMCMC (S_d-SGMCMC) with better scalability. \nThe author claimed the resulting algorithms achieved faster mixing speed and better classification accuracy when applied to real-world classification tasks. \n", "main_review": "## Strength\nPersonally, I find the proposed method interesting. The paper is written clearly and easy to follow, and the overall idea is easy to understand. I have checked the proofs of the theorem and they seem to be correct. To support the claims, the proposed methods were applied to real-world large data set to confirm their advantages compared to the standard full SGMCMC with some ablation studies, although I still think there is room for improvement. \n\n## Weakness and concerns\nAlthough the proposed method is easy to follow, I still find some parts unclear and elaborations are needed. First, the author claims that building the independent structure into the posterior can help improve the mixing speed (bottom of page 1, the first paragraph in the conclusion, etc.). However, I cannot see direct reasoning behind those claims, could you elaborate more on this? In section 6.1, I am not sure I understand the reasoning behind the better performances of S_d-SGMCMC. Can you elaborate more on \"regularizing the model\"?\nOn page 4, why $\\tilde{\\theta}^{(t,i)}$ is composed of samples from previous timesteps? I thought the author mentioned that a **single** sample $\\tilde{\\theta}^{(t,i)}$ from the **current** timestep is used for Monte Carlo approximation. I recommend putting more details in algorithm 1. For example, elaborating how $\\tilde{\\theta}^t_{\\neg i}$ is drawn from $\\hat{q}^t_{\\neg i}$. \nIn addition, why do appendix B and D use the same title? Can you merge them together?\n\nAnother concern is the theoretical soundness of proposed algorithms. For S-SGMCMC, the Monte Carlo estimation of the modified energy function requires previous samples. Does this break the Markov assumption of SGMCMC since it should only depend on the **current** samples? For example, when the sampler does not reach the stationary stage, $\\hat{q}^t$ still evolves with time, thus, the samples from the previous timesteps are not from $\\hat{q}^t$. \nApart from the stationary distribution of S-SGMCMC, I also wonder what is the stationary distribution (if exists) of the dropout version? How different is the stationary distribution of the dropout version compared to the optimal $q$? Any theoretical guarantees on the correctness of the dropout version? \n\nIn terms of the empirical evaluation, I wonder about the uncertainty quantification ability of the proposed algorithms, which is an important metric for SGMCMC methods, especially since the proposed methods also seem to underestimate the posterior variance. In addition, in the abstract, the author mentioned better predictive likelihoods. However, I can only find the accuracy metric in the experiment, which is different from the predictive likelihood. \n\n\n\n", "summary_of_the_review": "The paper presented an interesting approach to incorporate independent structures into posterior inference. However, there are still some ambiguities in motivation and theoretical soundness. To better support the claims, some metrics or experiments regarding uncertainty quantification should be added or discussed since the proposed methods also seem to underestimate the variance.  \n\n---\nI have read the author's responses. It address some of my concerns. However, for my last concern regarding the uncertainty (which is one of the reasons we consider MCMC over variational inference), the author did not mention it at all. Since this algorithm also underestimates the uncertainty due to minimizing the exclusive KL, I recommend the author to demonstrate that it can provide reasonable uncertainty quantification. Therefore, I will keep my original evaluations. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634898663609}], "openreview_url": "https://openreview.net/forum?id=57T1ctyxtP", "arxiv_id": "2107.09028", "paper_pdf": "papers/57T1ctyxtP.pdf", "paper_pdf_sha256": "f367cdaf6ed89f25b32cedc1a2ba7849ea7357d3f8554a6888292eaf778ab326", "paper_pdf_bytes": 871196, "paper_pdf_source": "openreview", "code_url": "https://github.com/mandt-lab/Structured-SG-MCMC", "code_repository": "mandt-lab/Structured-SG-MCMC", "code_commit": "c60e10c1eec51e9a1ff2c3f5e812c7eb3e7a5ae4", "code_archive": "repos/57T1ctyxtP.zip", "code_archive_sha256": "a318448525c031514261d1c8539f601bd0409c825d4d13c7fc12a0871900f246", "code_archive_bytes": 1799010, "code_file_count": 9, "code_extensions": {".py": 7, ".ipynb": 2}, "github_disk_usage_kb": 1798, "github_languages": {"Jupyter Notebook": 2406870, "Python": 142113}, "github_archived": false, "github_pushed_at": "2023-09-10T19:32:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/structured-stochastic-gradient-mcmc"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BygWRaVYwH", "year": 2020, "status": "rejected", "title": "Generalized Inner Loop Meta-Learning", "authors": ["Edward Grefenstette", "Brandon Amos", "Denis Yarats", "Phu Mon Htut", "Artem Molchanov", "Franziska Meier", "Douwe Kiela", "Kyunghyun Cho", "Soumith Chintala"], "authorids": ["egrefen@gmail.com", "brandon.amos.cs@gmail.com", "denisyarats@cs.nyu.edu", "pmh330@nyu.edu", "a.molchanov86@gmail.com", "fmeier@fb.com", "dkiela@fb.com", "kyunghyun.cho@nyu.edu", "soumith@gmail.com"], "authors_source": "OpenReview API", "abstract": "Many (but not all) approaches self-qualifying as \"meta-learning\" in deep learning and reinforcement learning fit a common pattern of approximating the solution to a nested optimization problem. In this paper, we give a formalization of this shared pattern, which we call GIMLI, prove its general requirements, and derive a general-purpose algorithm for implementing similar approaches. Based on this analysis and algorithm, we describe a library of our design, unnamedlib, which we share with the community to assist and enable future research into these kinds of meta-learning approaches. We end the paper by showcasing the practical applications of this framework and library through illustrative experiments and ablation studies which they facilitate.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ryeVrc4XcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper845/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe authors present a PyTorch based framework for performing second-order\nreverse mode autodiff for meta-learning.\n\nFirst, the authors present a formalization of a general prototypical\nmeta-learning setting.\nThey then provide an algorithm that solves this problem via gradient based\noptimization.\nFinally, perhaps the main contribution is a specific PyTorch implementation\nof said algorithm.\n\nThe type of meta-learning setting the authors consider is one where a gradient\nbased inner loop optimizer finds $\\theta^\\star$ by performing a finite number of steps.\nThe inner loop optimizer is parameterized through $\\varphi$ that consists of\ntwo parts $\\varphi^\\text{loss}$ and $\\varphi^\\text{opt}$.\nThe parameters $\\varphi^\\text{loss}$ are somehow part of the loss used for\ntraining in the inner loop. Example: Regularization paramter.\nThe parameters $\\varphi^\\text{opt}$ do not occur in the loss but in the\noptimizer step. Example: Learning rate.\n\nExample of an inner loop step:\n$\\theta^{k+1} := \\theta^k - \\alpha (\\nabla L(\\theta^k) + \\lambda \\nabla R(\\theta^k))$\nwhere $\\theta$ are the parameters of a neural network, $L$ is the training loss,\n$R$ is the regularizer, $\\alpha$ is the learning rate, $\\lambda$ is the regularization parameter.\nIn this example we would have $\\varphi = (\\alpha \\lambda)^T$.\n\nThe authos assume $\\theta^K$, the output of the inner loop after $K$ steps to\nbe differentiable wrt $\\varphi$.\nFurthermore, the meta-learning loss is assumed to be differentiable wrt $\\theta$\nso that a gradient of the meta-learning loss wrt to $\\varphi$ can be computed.\nThe authors also assume the meta-learning loss to be sufficient smooth in $\\varphi$\nsuch that a gradient based optimization can even be used for meta learning to\na local optimum.\n\nThe authors explicitly write down the reverse mode auto differentiation of\nthe inner loop and show how to, in that way, compute the gradient of the\nmeta-learning loss wrt to $\\varphi$.\n\nThe reverse (adjoint) mode auto differentiation of the above example inner loop step\nis the following step (iterated over in reverse down from $k = K$ to $k = 1$:\n$\\bar \\theta^k := (I - \\alpha (\\nabla^2 L(\\theta^k) + \\lambda \\nabla^2 R(\\theta^k)))^T \\bar \\theta^{k+1}$\n$\\bar \\alpha := \\bar \\alpha - (\\nabla l(\\theta^k) + \\lambda \\nabla R(\\theta^k))^T \\bar \\theta^{k+1}$\n$\\bar \\lambda := \\bar \\lambda - \\alpha \\nabla R(\\theta^k)^T \\bar \\theta^{k+1}$\nwhere $\\bar \\alpha$ accumulates the gradient of $\\theta$ wrt $\\alpha$ and\n$\\bar \\lambda$ accumulates the gradient of $\\theta$ wrt $\\lambda$.\n\nThe authors give some implementation details specific to some frameworks necessary\nfor implementing such \"gradient of an inner loop\".\n\nThe authors present experiments where they show how to meta-learn learning rates\nwith their framework.\nThey also how their framework can be used to quickly implement a MAML type\nmeta-learning optimizer ablation study comparing various combinations of\narchitecture, optimizer and inner loop steps etc..\n\nRecommendation:\nI propose to reject the paper.\nIn my eyes the only contribution is the implementation of a meta-learner in\nPyTorch based on well known methods.\nThe provided unifying formalization is theoretically inaccurate (see below)\nand to me come across as merely a motivation for their framework\n(but no value added compared to existing literature).\nThere is no new insight provided on the software engineering level either as\nfar as I can see.\n\nDetailed comments:\n- Page 1: ...provides tooling for analysing the provable requirements...\n\tseems like a complicated way of saying something that could be said simple\n\n- Page 2: Without loss of generality, ...\n\tYou are assuming a parametric model. Not sure what generality this phrase\n\trefers to.\n\n- Page 2: A formalization as $\\theta^\\star = argmin(\\theta, L(\\theta, \\varphi))$\n\tis inaccurate in the sense that it does not acknowledge the existing of\n\tmultiple optima.\n\tI would recommend not to use the $argmin$ operator here, since $argmin$ for\n\tsomething like a neural network would for example either return a global\n\toptimum (which no optimizer used in practice finds, and is not ment here)\n\tor would take on a set value with multiple local minima for example.\n\n\tIn the same context, the authors should mention the issues about uniqueness\n\tof optima (we are not even really finding optima when training neural networks),\n\timplicit functions / implicit differentiation\n\n\tIn the context of using stochastic optimizers one should also at least\n\tmention something about the differentiability of outputs of such optimizers\n\tand how they potentially depend on randomness of mini-batches\n\t(what if different randomness is used with the same or a perturbed\n\thyper parameter?)\n\n- Page 3: You mention the potential statefulness of the optimizer.\n\tWhy not explicitly carry it in the math notation?\n\tProbably things would get cluttered but saying it should be covered within\n\t$\\varphi$ does not seem reasonable to me.\n\n- Page 3: While this may seem like a fairly trivial formalization...\n\tYes, but also nesting this in an outer loop is fairly trivial in the sense\n\tthat it is a well known approach.\n\n- Page 4: there exist continuous hyperparam...\n\tIf they are not continuous then they should not even occur in this\n\tformalization so saying there exist... does not make much sense to me here\n\n\t$\\alpha \\subseteq \\varphi^\\text{opt}$ implies that $\\varphi^\\text{opt}$ is\n\ta set from notation although we are treating it as a vector everywhere else\n\n- Page 4: All of section 2.4 seems somewhat trivial to me, but I guess that is\n\thighly subjective.\n\n- Page 5: in the definition of stop operator perhaps use $:\\Leftrightarrow$\n\n- Page 5: Perhaps explicitly mention how your approach differs from a reverse\n\tmode differentiation of training or if it does not differ, say this.\n\n- Page 14: When talking about _S_GD (instead of just GD) perhaps mention\n\tsomething about non-existence of mini-batch randomness / being deterministic\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "Summary:\nThe authors present a PyTorch based framework for performing second-order\nreverse mode autodiff for meta-learning.\n\nFirst, the authors present a formalization of a general prototypical\nmeta-learning setting.\nThey then provide an algorithm that solves this problem via gradient based\noptimization.\nFinally, perhaps the main contribution is a specific PyTorch implementation\nof said algorithm.\n\nThe type of meta-learning setting the authors consider is one where a gradient\nbased inner loop optimizer finds $\\theta^\\star$ by performing a finite number of steps.\nThe inner loop optimizer is parameterized through $\\varphi$ that consists of\ntwo parts $\\varphi^\\text{loss}$ and $\\varphi^\\text{opt}$.\nThe parameters $\\varphi^\\text{loss}$ are somehow part of the loss used for\ntraining in the inner loop. Example: Regularization paramter.\nThe parameters $\\varphi^\\text{opt}$ do not occur in the loss but in the\noptimizer step. Example: Learning rate.\n\nExample of an inner loop step:\n$\\theta^{k+1} := \\theta^k - \\alpha (\\nabla L(\\theta^k) + \\lambda \\nabla R(\\theta^k))$\nwhere $\\theta$ are the parameters of a neural network, $L$ is the training loss,\n$R$ is the regularizer, $\\alpha$ is the learning rate, $\\lambda$ is the regularization parameter.\nIn this example we would have $\\varphi = (\\alpha \\lambda)^T$.\n\nThe authos assume $\\theta^K$, the output of the inner loop after $K$ steps to\nbe differentiable wrt $\\varphi$.\nFurthermore, the meta-learning loss is assumed to be differentiable wrt $\\theta$\nso that a gradient of the meta-learning loss wrt to $\\varphi$ can be computed.\nThe authors also assume the meta-learning loss to be sufficient smooth in $\\varphi$\nsuch that a gradient based optimization can even be used for meta learning to\na local optimum.\n\nThe authors explicitly write down the reverse mode auto differentiation of\nthe inner loop and show how to, in that way, compute the gradient of the\nmeta-learning loss wrt to $\\varphi$.\n\nThe reverse (adjoint) mode auto differentiation of the above example inner loop step\nis the following step (iterated over in reverse down from $k = K$ to $k = 1$:\n$\\bar \\theta^k := (I - \\alpha (\\nabla^2 L(\\theta^k) + \\lambda \\nabla^2 R(\\theta^k)))^T \\bar \\theta^{k+1}$\n$\\bar \\alpha := \\bar \\alpha - (\\nabla l(\\theta^k) + \\lambda \\nabla R(\\theta^k))^T \\bar \\theta^{k+1}$\n$\\bar \\lambda := \\bar \\lambda - \\alpha \\nabla R(\\theta^k)^T \\bar \\theta^{k+1}$\nwhere $\\bar \\alpha$ accumulates the gradient of $\\theta$ wrt $\\alpha$ and\n$\\bar \\lambda$ accumulates the gradient of $\\theta$ wrt $\\lambda$.\n\nThe authors give some implementation details specific to some frameworks necessary\nfor implementing such \"gradient of an inner loop\".\n\nThe authors present experiments where they show how to meta-learn learning rates\nwith their framework.\nThey also how their framework can be used to quickly implement a MAML type\nmeta-learning optimizer ablation study comparing various combinations of\narchitecture, optimizer and inner loop steps etc..\n\nRecommendation:\nI propose to reject the paper.\nIn my eyes the only contribution is the implementation of a meta-learner in\nPyTorch based on well known methods.\nThe provided unifying formalization is theoretically inaccurate (see below)\nand to me come across as merely a motivation for their framework\n(but no value added compared to existing literature).\nThere is no new insight provided on the software engineering level either as\nfar as I can see.\n\nDetailed comments:\n- Page 1: ...provides tooling for analysing the provable requirements...\n\tseems like a complicated way of saying something that could be said simple\n\n- Page 2: Without loss of generality, ...\n\tYou are assuming a parametric model. Not sure what generality this phrase\n\trefers to.\n\n- Page 2: A formalization as $\\theta^\\star = argmin(\\theta, L(\\theta, \\varphi))$\n\tis inaccurate in the sense that it does not acknowledge the existing of\n\tmultiple optima.\n\tI would recommend not to use the $argmin$ operator here, since $argmin$ for\n\tsomething like a neural network would for example either return a global\n\toptimum (which no optimizer used in practice finds, and is not ment here)\n\tor would take on a set value with multiple local minima for example.\n\n\tIn the same context, the authors should mention the issues about uniqueness\n\tof optima (we are not even really finding optima when training neural networks),\n\timplicit functions / implicit differentiation\n\n\tIn the context of using stochastic optimizers one should also at least\n\tmention something about the differentiability of outputs of such optimizers\n\tand how they potentially depend on randomness of mini-batches\n\t(what if different randomness is used with the same or a perturbed\n\thyper parameter?)\n\n- Page 3: You mention the potential statefulness of the optimizer.\n\tWhy not explicitly carry it in the math notation?\n\tProbably things would get cluttered but saying it should be covered within\n\t$\\varphi$ does not seem reasonable to me.\n\n- Page 3: While this may seem like a fairly trivial formalization...\n\tYes, but also nesting this in an outer loop is fairly trivial in the sense\n\tthat it is a well known approach.\n\n- Page 4: there exist continuous hyperparam...\n\tIf they are not continuous then they should not even occur in this\n\tformalization so saying there exist... does not make much sense to me here\n\n\t$\\alpha \\subseteq \\varphi^\\text{opt}$ implies that $\\varphi^\\text{opt}$ is\n\ta set from notation although we are treating it as a vector everywhere else\n\n- Page 4: All of section 2.4 seems somewhat trivial to me, but I guess that is\n\thighly subjective.\n\n- Page 5: in the definition of stop operator perhaps use $:\\Leftrightarrow$\n\n- Page 5: Perhaps explicitly mention how your approach differs from a reverse\n\tmode differentiation of training or if it does not differ, say this.\n\n- Page 14: When talking about _S_GD (instead of just GD) perhaps mention\n\tsomething about non-existence of mini-batch randomness / being deterministic\n"}, "tcdate": 1572190779639}, {"id": "r1lteyFy9B", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper845/AnonReviewer1"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work presented a general formulation of a wide class of existing meta-learning approaches, and proved the requirements that must be satisfied for such approaches to be possible.\n\nHalf of the work is focused on describing the unnamedlib library, which extends PyTorch to enable the easy\nand natural implementation of such meta-learning approaches.\n\nThe early sections are interesting, especially section 2, which gives some great insights to the existing inner loop pattern in meta-learning. However, from section 3, the paper has turned to examples and related works, where I was hoping the author would give more detailed analysis of the pattern. My concern is the authors have spent too much space on the unnamedlib library. So http://www.jmlr.org/mloss/ might be a more suitable place for publication.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory.", "review": "This work presented a general formulation of a wide class of existing meta-learning approaches, and proved the requirements that must be satisfied for such approaches to be possible.\n\nHalf of the work is focused on describing the unnamedlib library, which extends PyTorch to enable the easy\nand natural implementation of such meta-learning approaches.\n\nThe early sections are interesting, especially section 2, which gives some great insights to the existing inner loop pattern in meta-learning. However, from section 3, the paper has turned to examples and related works, where I was hoping the author would give more detailed analysis of the pattern. My concern is the authors have spent too much space on the unnamedlib library. So http://www.jmlr.org/mloss/ might be a more suitable place for publication."}, "tcdate": 1571946224812}, {"id": "BJgg7RijtS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper845/AnonReviewer3"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose the general formulation of recent meta-learning methods and propose a good library to use.\n\nPros:\n1. The general formulation of recent meta-learning methods is reasonable.\n2. The proposed library is easy to use.\n\nCons:\n\nThe paper lacks technical novelty. I understand the goal of this paper is to build a library. However, the paper only describes a general formulation for recent meta-learning methods (e.g., MAML) and implement the formulation. It is better to clarify and some key engineering challenges and do the corresponding experiments.\n\nIn addition, in the experiment parts, the authors only compare the results with MAML++. It will be more convincing if the authors can analyze other popular meta-learning methods (e.g.. Prototypical network [1], meta-LSTM [2]). \n\nAnother suggestion is that the authors can give some examples to connect current meta-learning models with the proposed general formulation. For example, the meaning of \\phi_i^opt, \\phi_i^loss in MAML, Prototype, Reptile, etc.\n\nIt is better to explain the meaning of different colors in Figure 3.\n\n[1] Snell, Jake, Kevin Swersky, and Richard Zemel. \"Prototypical networks for few-shot learning.\" Advances in Neural Information Processing Systems. 2017.\n[2] Ravi, Sachin, and Hugo Larochelle. \"Optimization as a model for few-shot learning.\" ICLR (2016).\n\n\n\nDecision after rebuttal: I have read the authors' responses. Like review 2, I also think the \"generalization\" is overclaimed, it only provides a general formulation. Thus, I finally decide to keep my score.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "title": "Official Blind Review #3", "review": "The authors propose the general formulation of recent meta-learning methods and propose a good library to use.\n\nPros:\n1. The general formulation of recent meta-learning methods is reasonable.\n2. The proposed library is easy to use.\n\nCons:\n\nThe paper lacks technical novelty. I understand the goal of this paper is to build a library. However, the paper only describes a general formulation for recent meta-learning methods (e.g., MAML) and implement the formulation. It is better to clarify and some key engineering challenges and do the corresponding experiments.\n\nIn addition, in the experiment parts, the authors only compare the results with MAML++. It will be more convincing if the authors can analyze other popular meta-learning methods (e.g.. Prototypical network [1], meta-LSTM [2]). \n\nAnother suggestion is that the authors can give some examples to connect current meta-learning models with the proposed general formulation. For example, the meaning of \\phi_i^opt, \\phi_i^loss in MAML, Prototype, Reptile, etc.\n\nIt is better to explain the meaning of different colors in Figure 3.\n\n[1] Snell, Jake, Kevin Swersky, and Richard Zemel. \"Prototypical networks for few-shot learning.\" Advances in Neural Information Processing Systems. 2017.\n[2] Ravi, Sachin, and Hugo Larochelle. \"Optimization as a model for few-shot learning.\" ICLR (2016).\n\n\n\nDecision after rebuttal: I have read the authors' responses. Like review 2, I also think the \"generalization\" is overclaimed, it only provides a general formulation. Thus, I finally decide to keep my score.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, "tcdate": 1571696151880}], "openreview_url": "https://openreview.net/forum?id=BygWRaVYwH", "arxiv_id": "1910.01727", "paper_pdf": "papers/BygWRaVYwH.pdf", "paper_pdf_sha256": "ab398d39418bf7d0666226ef670c09163533ca61e9f7f35b682001c2b2164acd", "paper_pdf_bytes": 2402275, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/higher", "code_repository": "facebookresearch/higher", "code_commit": "15a247ac06cac0d22601322677daff0dcfff062e", "code_archive": "repos/BygWRaVYwH.zip", "code_archive_sha256": "9342ae241bce7246b75df89a2807f6d1abc91d91644d3f642cc6dbc8e09afc53", "code_archive_bytes": 9067093, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 8888, "github_languages": {"Python": 132092}, "github_archived": true, "github_pushed_at": "2022-03-25T15:56:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalized-inner-loop-meta-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YQ5wH5C69t", "year": 2026, "status": "rejected", "title": "Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel Decoding", "authors": ["Runpeng Yu", "Xinyin Ma", "Xinchao Wang"], "authorids": ["~Runpeng_Yu2", "~Xinyin_Ma1", "~Xinchao_Wang1"], "authors_source": "OpenReview API", "abstract": "In this work, we present Dimple and Dimple+, two Discrete Diffusion Multimodal Large Language Models (dMLLMs). Dimple is initialized from a discrete diffusion Large Language Model (dLLM) without multimodal understanding ability, and learns such ability through a hybrid training paradigm that first applies autoregressive training and then switches to discrete diffusion training. Dimple+ is initialized from an autoregressive Multimodal Large Language Models, and acquires parallel decoding capability through pure discrete diffusion training. Both models achieve performance comparable to their autoregressive baselines, and Dimple+ establishes new state-of-the-art results among dMLLMs. To enhance inference efficiency, we propose Confident Decoding, which dynamically adjusts the number of tokens generated per iteration. Experiments show that it accelerates decoding by 2×–6× with only minor performance degradation. We also demonstrate that the Prefilling technique, previously used in autoregressive models, can be effectively applied to dMLLMs with bidirectional attention, achieving nearly lossless speedups of 1.7×–7×. Finally, we introduce the Structure Prior method, enabling fine-grained control over response format and reasoning structure, which is difficult to realize in autoregressive models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "km03nDurUa", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4970/Reviewer_h84d"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces Dimple and Dimple+, two Discrete Diffusion Multimodal Large Language Models (dMLLMs) that leverage discrete diffusion for parallel decoding in multimodal (vision-language) tasks. Dimple is initialized from a discrete diffusion LLM and learns multimodal understanding via a hybrid training paradigm (autoregressive pretraining, then discrete diffusion). Dimple+ is initialized from an autoregressive MLLM and acquires parallel decoding via pure discrete diffusion training. Both models achieve performance comparable to their autoregressive baselines, with Dimple+ establishing new state-of-the-art results among dMLLMs.\n\nThe authors propose several inference and training innovations:\n\n1.Confident Decoding: Dynamically adjusts the number of tokens generated per iteration, accelerating decoding by 2×–6× with minor performance degradation.\n\n2.Prefilling: Adapts a technique from AR models to dMLLMs, achieving nearly lossless speedups of 1.7×–7×.\nStructure Prior: Enables fine-grained control over response format and reasoning structure, which is difficult in AR models.\n\n3.Extensive experiments on standard multimodal benchmarks show that Dimple and Dimple+ match or surpass AR baselines and outperform other dMLLMs, with strong ablations and qualitative analyses.", "review_text": "This paper introduces Dimple and Dimple+, two Discrete Diffusion Multimodal Large Language Models (dMLLMs) that leverage discrete diffusion for parallel decoding in multimodal (vision-language) tasks. Dimple is initialized from a discrete diffusion LLM and learns multimodal understanding via a hybrid training paradigm (autoregressive pretraining, then discrete diffusion). Dimple+ is initialized from an autoregressive MLLM and acquires parallel decoding via pure discrete diffusion training. Both models achieve performance comparable to their autoregressive baselines, with Dimple+ establishing new state-of-the-art results among dMLLMs.\n\nThe authors propose several inference and training innovations:\n\n1.Confident Decoding: Dynamically adjusts the number of tokens generated per iteration, accelerating decoding by 2×–6× with minor performance degradation.\n\n2.Prefilling: Adapts a technique from AR models to dMLLMs, achieving nearly lossless speedups of 1.7×–7×.\nStructure Prior: Enables fine-grained control over response format and reasoning structure, which is difficult in AR models.\n\n3.Extensive experiments on standard multimodal benchmarks show that Dimple and Dimple+ match or surpass AR baselines and outperform other dMLLMs, with strong ablations and qualitative analyses.", "strengths": "The strengths of the paper are:\n1. Originality: Proposes a hybrid training paradigm (AR pretraining + diffusion) for dMLLMs, which is empirically effective. Introduces confident decoding and structure prior, enabling dynamic parallel decoding and fine-grained output control—capabilities not available in AR models. \n2. Clarity: Clear motivation, methodology, and results presentation. \n3. Significance: The proposed techniques (confident decoding, structure prior) are likely to inspire further research in efficient and controllable generation.", "weaknesses": "The weaknesses of the paper are:\n1. Novelty: The proposed ideas are engineering centric. The confident decoding with flexible steps is proposed in literature I believe. I don't weigh too high for the overall novelty of the paper, but is okay in some sense.\n2. Ablation on Structure Prior: While qualitative examples are provided, a more systematic quantitative evaluation of the structure prior’s impact would strengthen the claims.\n3. Lack of comparisons and experimental details.", "questions": "1. Will you release code, models, and scripts for reproducibility?\n2. I don't see supplementary materials of experimental details for reproducibility.\n3. In table1, are you going to compare the inference speed of Dimple+ to Qwen2.5L or its AR baselines? It is interesting to know how much the diffusion LLM is faster than AR model.\n4. Comparing Dimple and Dimple+, it is hard to draw the conclusion that initialization from AR training is better based on the experiment. The alignment and Instruction tuning phase of Dimple is done with much less training data and compute than Qwen2.5 VL. I believe the authors have used the 1.3M/0.8B data for this while Qwen2.5b-VL used many more, is it?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Dimple and Dimple+, two Discrete Diffusion Multimodal Large Language Models (dMLLMs) that leverage discrete diffusion for parallel decoding in multimodal (vision-language) tasks. Dimple is initialized from a discrete diffusion LLM and learns multimodal understanding via a hybrid training paradigm (autoregressive pretraining, then discrete diffusion). Dimple+ is initialized from an autoregressive MLLM and acquires parallel decoding via pure discrete diffusion training. Both models achieve performance comparable to their autoregressive baselines, with Dimple+ establishing new state-of-the-art results among dMLLMs.\n\nThe authors propose several inference and training innovations:\n\n1.Confident Decoding: Dynamically adjusts the number of tokens generated per iteration, accelerating decoding by 2×–6× with minor performance degradation.\n\n2.Prefilling: Adapts a technique from AR models to dMLLMs, achieving nearly lossless speedups of 1.7×–7×.\nStructure Prior: Enables fine-grained control over response format and reasoning structure, which is difficult in AR models.\n\n3.Extensive experiments on standard multimodal benchmarks show that Dimple and Dimple+ match or surpass AR baselines and outperform other dMLLMs, with strong ablations and qualitative analyses.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The strengths of the paper are:\n1. Originality: Proposes a hybrid training paradigm (AR pretraining + diffusion) for dMLLMs, which is empirically effective. Introduces confident decoding and structure prior, enabling dynamic parallel decoding and fine-grained output control—capabilities not available in AR models. \n2. Clarity: Clear motivation, methodology, and results presentation. \n3. Significance: The proposed techniques (confident decoding, structure prior) are likely to inspire further research in efficient and controllable generation.", "weaknesses": "The weaknesses of the paper are:\n1. Novelty: The proposed ideas are engineering centric. The confident decoding with flexible steps is proposed in literature I believe. I don't weigh too high for the overall novelty of the paper, but is okay in some sense.\n2. Ablation on Structure Prior: While qualitative examples are provided, a more systematic quantitative evaluation of the structure prior’s impact would strengthen the claims.\n3. Lack of comparisons and experimental details.", "questions": "1. Will you release code, models, and scripts for reproducibility?\n2. I don't see supplementary materials of experimental details for reproducibility.\n3. In table1, are you going to compare the inference speed of Dimple+ to Qwen2.5L or its AR baselines? It is interesting to know how much the diffusion LLM is faster than AR model.\n4. Comparing Dimple and Dimple+, it is hard to draw the conclusion that initialization from AR training is better based on the experiment. The alignment and Instruction tuning phase of Dimple is done with much less training data and compute than Qwen2.5 VL. I believe the authors have used the 1.3M/0.8B data for this while Qwen2.5b-VL used many more, is it?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761974538065}, {"id": "AWq5igfnlL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4970/Reviewer_6WKY"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces **Dimple and Dimple+**, two discrete diffusion multimodal LLMs (dMLLMs) that bring parallel decoding to vision-language models. Dimple follows a hybrid training pipeline, autoregressive (AR) alignment and instruction tuning, then discrete diffusion, while Dimple+ starts from an AR MLLM (Qwen2.5‑VL) and acquires parallel decoding via diffusion-only tuning. The authors propose Confident Decoding (adaptive multi-token updates per step), re-use Prefilling with bidirectional attention to cache prompt KV states, and demonstrate Structure Prior for fine‑grained output control. On 12 benchmarks, Dimple/Dimple+ generally match their AR baselines; Dimple+ attains state-of-the‑art results among dMLLMs versus LaViDA/LLaDA‑V with notably fewer training samples. Confident Decoding yields ~2x-6x fewer steps with minor loss, and Prefilling offers ~1.7x-7x speedups with small average drops. Qualitative cases show structured reasoning, JSON‑like formatting, and length control.", "review_text": "The paper introduces **Dimple and Dimple+**, two discrete diffusion multimodal LLMs (dMLLMs) that bring parallel decoding to vision-language models. Dimple follows a hybrid training pipeline, autoregressive (AR) alignment and instruction tuning, then discrete diffusion, while Dimple+ starts from an AR MLLM (Qwen2.5‑VL) and acquires parallel decoding via diffusion-only tuning. The authors propose Confident Decoding (adaptive multi-token updates per step), re-use Prefilling with bidirectional attention to cache prompt KV states, and demonstrate Structure Prior for fine‑grained output control. On 12 benchmarks, Dimple/Dimple+ generally match their AR baselines; Dimple+ attains state-of-the‑art results among dMLLMs versus LaViDA/LLaDA‑V with notably fewer training samples. Confident Decoding yields ~2x-6x fewer steps with minor loss, and Prefilling offers ~1.7x-7x speedups with small average drops. Qualitative cases show structured reasoning, JSON‑like formatting, and length control.", "strengths": "- Two pragmatic routes to dMLLMs: AR->diffusion (Dimple) and AR‑MLLM initialization->diffusion (Dimple+). Confident Decoding adaptively sets the number of updated tokens per step, unlike fixed‑K schedules in prior work. Structure Prior gives direct positional control, early answering and enforced formats, difficult for AR models.\n\n- Competitive accuracy vs. matched AR baselines on 12 benchmarks and clear SOTA among dMLLMs with far fewer training samples than LLaDA‑V. Ablations quantify Prefilling speedups ($\\sim$1.7x-7x) with small average performance drop ($\\sim$0.8\\%) and show Confident Decoding reduces steps by $\\sim$3x-6x at near‑baseline accuracy.\n\n- Training strategies, attention masks, and losses are explicit; examples on pp. 7-8 make the decoding behavior and controllability intuitive.\n\n- Encourages a practical path to parallel, controllable multimodal generation; speedups are meaningful for long‑prompt VLM inference.", "weaknesses": "- While Dimple+ is SOTA within discrete diffusion MLLMs, it lags a strong AR MLLM (Qwen2.5‑VL‑7B) on several tasks (e.g., ChartQA 74.7 vs. 87.3; OCRBench 699 vs. 783; Table 1), making the overall value proposition partly about parity plus speed/controllability, not raw accuracy.\n\n- The paper claims parity under the same budget, but AR vs. diffusion differ in token supervision density and FLOPs/step; matching only by iterations or tokens may not equate compute. A FLOPs‑normalized comparison would strengthen the claim.\n\n- Results are limited to ~7B scale; it is unclear how the approach scales to larger models or image‑dense prompts with much longer contexts. \n\n- Structure Prior is shown qualitatively (Tables 5–7) but lacks metrics such as exact‑format accuracy, early‑answer timing/quality trade‑offs, or robustness when priors conflict with content.\n\n- Confident Decoding hinges on the confidence threshold $\\gamma$; the main paper does not show sensitivity curves or cross‑task robustness (Table 4 reports one configuration).", "questions": "- How exactly is \"same training budget\" defined and matched (tokens, optimizer steps, wall‑clock, or FLOPs)? Please provide FLOPs‑level accounting for AR vs. diffusion phases and rerun a compute‑matched comparison.\n\n- How sensitive is Confident Decoding to $\\gamma$ across tasks and lengths? Could you report curves (accuracy vs. steps vs. $\\gamma$) and per‑task optima?\n\n- Have you tried larger backbones (e.g., ≥14B)? Do speedups and parity hold at scale, and does Prefilling remain near‑lossless with very long image sequences?\n\n- Can you quantify Structure Prior with exact-match format rates (e.g., strict JSON/LaTeX) and measure the earliness of correct answers vs. final sequence length? Any failure modes when priors are partially wrong or adversarial?\n\n- Beyond the listed vision–language tasks, how do Dimple/Dimple+ perform on purely textual reasoning/planning benchmarks where diffusion LMs have reportedly excelled?\n\n- For Dimple, can you show matched‑compute comparisons of (i) diffusion‑only vs. AR->diffusion, and (ii) varying the lengths of AA/AT/DT phases?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces **Dimple and Dimple+**, two discrete diffusion multimodal LLMs (dMLLMs) that bring parallel decoding to vision-language models. Dimple follows a hybrid training pipeline, autoregressive (AR) alignment and instruction tuning, then discrete diffusion, while Dimple+ starts from an AR MLLM (Qwen2.5‑VL) and acquires parallel decoding via diffusion-only tuning. The authors propose Confident Decoding (adaptive multi-token updates per step), re-use Prefilling with bidirectional attention to cache prompt KV states, and demonstrate Structure Prior for fine‑grained output control. On 12 benchmarks, Dimple/Dimple+ generally match their AR baselines; Dimple+ attains state-of-the‑art results among dMLLMs versus LaViDA/LLaDA‑V with notably fewer training samples. Confident Decoding yields ~2x-6x fewer steps with minor loss, and Prefilling offers ~1.7x-7x speedups with small average drops. Qualitative cases show structured reasoning, JSON‑like formatting, and length control.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Two pragmatic routes to dMLLMs: AR->diffusion (Dimple) and AR‑MLLM initialization->diffusion (Dimple+). Confident Decoding adaptively sets the number of updated tokens per step, unlike fixed‑K schedules in prior work. Structure Prior gives direct positional control, early answering and enforced formats, difficult for AR models.\n\n- Competitive accuracy vs. matched AR baselines on 12 benchmarks and clear SOTA among dMLLMs with far fewer training samples than LLaDA‑V. Ablations quantify Prefilling speedups ($\\sim$1.7x-7x) with small average performance drop ($\\sim$0.8\\%) and show Confident Decoding reduces steps by $\\sim$3x-6x at near‑baseline accuracy.\n\n- Training strategies, attention masks, and losses are explicit; examples on pp. 7-8 make the decoding behavior and controllability intuitive.\n\n- Encourages a practical path to parallel, controllable multimodal generation; speedups are meaningful for long‑prompt VLM inference.", "weaknesses": "- While Dimple+ is SOTA within discrete diffusion MLLMs, it lags a strong AR MLLM (Qwen2.5‑VL‑7B) on several tasks (e.g., ChartQA 74.7 vs. 87.3; OCRBench 699 vs. 783; Table 1), making the overall value proposition partly about parity plus speed/controllability, not raw accuracy.\n\n- The paper claims parity under the same budget, but AR vs. diffusion differ in token supervision density and FLOPs/step; matching only by iterations or tokens may not equate compute. A FLOPs‑normalized comparison would strengthen the claim.\n\n- Results are limited to ~7B scale; it is unclear how the approach scales to larger models or image‑dense prompts with much longer contexts. \n\n- Structure Prior is shown qualitatively (Tables 5–7) but lacks metrics such as exact‑format accuracy, early‑answer timing/quality trade‑offs, or robustness when priors conflict with content.\n\n- Confident Decoding hinges on the confidence threshold $\\gamma$; the main paper does not show sensitivity curves or cross‑task robustness (Table 4 reports one configuration).", "questions": "- How exactly is \"same training budget\" defined and matched (tokens, optimizer steps, wall‑clock, or FLOPs)? Please provide FLOPs‑level accounting for AR vs. diffusion phases and rerun a compute‑matched comparison.\n\n- How sensitive is Confident Decoding to $\\gamma$ across tasks and lengths? Could you report curves (accuracy vs. steps vs. $\\gamma$) and per‑task optima?\n\n- Have you tried larger backbones (e.g., ≥14B)? Do speedups and parity hold at scale, and does Prefilling remain near‑lossless with very long image sequences?\n\n- Can you quantify Structure Prior with exact-match format rates (e.g., strict JSON/LaTeX) and measure the earliness of correct answers vs. final sequence length? Any failure modes when priors are partially wrong or adversarial?\n\n- Beyond the listed vision–language tasks, how do Dimple/Dimple+ perform on purely textual reasoning/planning benchmarks where diffusion LMs have reportedly excelled?\n\n- For Dimple, can you show matched‑compute comparisons of (i) diffusion‑only vs. AR->diffusion, and (ii) varying the lengths of AA/AT/DT phases?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761719060496}, {"id": "wgkRE621RC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4970/Reviewer_3Bzn"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper explore the discrete diffusion for MLLMs, and present two diffusion MLLMs, termed Dimple and Dimple+. Besides, the authors also explore the parallel decoding to accelerate the inference of MLLMs.", "review_text": "This paper explore the discrete diffusion for MLLMs, and present two diffusion MLLMs, termed Dimple and Dimple+. Besides, the authors also explore the parallel decoding to accelerate the inference of MLLMs.", "strengths": "1. The authors explore the diffusion paradigm for MLLMs, and investigate this manner under the setting of pure LLMs and MLLMs. \n\n2. A confident decoding is proposed to accelerate the inference of MLLMs, which can achieve 2x-6x speedups.", "weaknesses": "1.  The experimental designs and results are hard to support the arguments. For instance, compared to the default QWen2.5-VL-7B, Dimple+ encounters obvious performance drops on multiple benchmarks.  Moreover, the results of Dimple-7B-AR baseline is also questionable. Compared to LLaVA-Next which uses a much weaker LLM, Dimple-AR-baseline perform worse on multiple benchmarks. These results are quite unreasonable. If the authors want to proof the merits of diffusion, they can make comparisons to open-sourced AR-MLLMs under the same experimental settings.\n\nBesides, some notations of Tab.1 are also incorrect. If Dimple-7B use QWen2.5-VL as the base MLLM, the training samples should not be 1.6B. It is 2.6T+1.6B. \n\n2. The actual contribution of Confident Decoding is unknown. From the paper, it is hard to recognize how many improvements and novelty the proposed Confident Decoding has compared to previous parallel decoding works. And how confident decoding implemented is also unclear. \n\nBesides, can the parallel decoding only be achieved via the diffusion paradigm, or it can also be done in AR-MLLMs? It is important to the readers to judge the significance of diffusion MLLMs compared to existing MLLM research. \n\n3. Following the first question, the comparison of DIMPLE to existing AR-MLLMs is not sufficient. More experiments should be conducted to show the merits of diffusion modeling.", "questions": "Q1 Can the parallel decoding only be achieved via the diffusion paradigm, or it can also be done in AR-MLLMs? It is important to the readers to judge the significance of diffusion MLLMs compared to existing MLLM research.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explore the discrete diffusion for MLLMs, and present two diffusion MLLMs, termed Dimple and Dimple+. Besides, the authors also explore the parallel decoding to accelerate the inference of MLLMs.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The authors explore the diffusion paradigm for MLLMs, and investigate this manner under the setting of pure LLMs and MLLMs. \n\n2. A confident decoding is proposed to accelerate the inference of MLLMs, which can achieve 2x-6x speedups.", "weaknesses": "1.  The experimental designs and results are hard to support the arguments. For instance, compared to the default QWen2.5-VL-7B, Dimple+ encounters obvious performance drops on multiple benchmarks.  Moreover, the results of Dimple-7B-AR baseline is also questionable. Compared to LLaVA-Next which uses a much weaker LLM, Dimple-AR-baseline perform worse on multiple benchmarks. These results are quite unreasonable. If the authors want to proof the merits of diffusion, they can make comparisons to open-sourced AR-MLLMs under the same experimental settings.\n\nBesides, some notations of Tab.1 are also incorrect. If Dimple-7B use QWen2.5-VL as the base MLLM, the training samples should not be 1.6B. It is 2.6T+1.6B. \n\n2. The actual contribution of Confident Decoding is unknown. From the paper, it is hard to recognize how many improvements and novelty the proposed Confident Decoding has compared to previous parallel decoding works. And how confident decoding implemented is also unclear. \n\nBesides, can the parallel decoding only be achieved via the diffusion paradigm, or it can also be done in AR-MLLMs? It is important to the readers to judge the significance of diffusion MLLMs compared to existing MLLM research. \n\n3. Following the first question, the comparison of DIMPLE to existing AR-MLLMs is not sufficient. More experiments should be conducted to show the merits of diffusion modeling.", "questions": "Q1 Can the parallel decoding only be achieved via the diffusion paradigm, or it can also be done in AR-MLLMs? It is important to the readers to judge the significance of diffusion MLLMs compared to existing MLLM research.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761277079321}, {"id": "8Rvj4aBbAi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4970/Reviewer_YGWg"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "This paper introduces Dimple and Dimple+, multimodal large language model with discrete diffusion decoding. \n\nThe motivation is to address the instability and inefficiency of pure diffusion models in multimodal tasks. \nAnd the paper validates two distinct strategies for converting AR models into diffusion-based multimodal models. (dimple and dimple+)\n\nIt also proposes a decoding strategy called “confident decoding” to reduce inference steps without obvious performance degradation.", "review_text": "This paper introduces Dimple and Dimple+, multimodal large language model with discrete diffusion decoding. \n\nThe motivation is to address the instability and inefficiency of pure diffusion models in multimodal tasks. \nAnd the paper validates two distinct strategies for converting AR models into diffusion-based multimodal models. (dimple and dimple+)\n\nIt also proposes a decoding strategy called “confident decoding” to reduce inference steps without obvious performance degradation.", "strengths": "1. The paper validates two practical strategies for converting AR models into diffusion-based multimodal models, providing clear guidance for the community. \n* Dimple: Starts from a diffusion-only LLM without multimodal capability. It first applies autoregressive training to align vision and language, then switches to diffusion-based instruction tuning. \n* Dimple+: Starts from a well-pretrained AR multimodal model (Qwen2.5-VL) and directly modifies its attention mask and loss function to enable diffusion training.\n2. It introduces Confident Decoding, a dynamic decoding method that improves inference efficiency without compromising output quality.\n3. From the code in supplementary material, the model is integrated into the Transformers library and will be fully open-sourced, providing a reproducible and accessible reference for the community.\n4. Through comparisons with AR baselines under identical training data, and with state-of-the-art diffusion models like LLaDA-V, Dimple+ demonstrates the effectiveness of diffusion training for multimodal tasks.", "weaknesses": "1. Although Dimple+ improves over its AR baseline, it still falls short compared to its initial AR models (Qwen2.5-VL) on several benchmarks. In particular, the performance gap on ChartQA is substantial, despite the training data includes similar samples (2 epoch results is much better than 1 epoch training), raising concerns about catastrophic forgetting and the scalability of diffusion models.\n\n2. The paper lacks analysis on the optimal number of decoding steps (k). While Confident Decoding dynamically adjusts k, it remains unclear whether increasing k beyond 8 could further improve performance, especially in complex tasks.\n\n3. There is no direct comparison between the inference speed of Dimple+ and standard AR models. Without this, it’s difficult to assess the practical efficiency gains of diffusion decoding.\n\n4. Despite the promising speedups and results reported for Confident Decoding, empirical testing with the open-sourced Dimple model shows noticeable degradation in long-context text generation. For example, when max_tokens is set to 64 or higher, even with step counts of 128 or 256, the model often fails to produce correct EOS tokens and instead generates excessive hallucinations. A controlled comparison with Dream-7B or other AR baseline would help clarify whether Dimple suffers from diffusion-induced degradation in textual capabilities.", "questions": "Overall, I think this is a strong paper that makes a meaningful contribution to the development of multimodal diffusion language models. However, I have a few questions that I hope the authors can clarify:\n1. **Conclusion and Fair Comparison Between Dimple and Dimple+** \nThe paper presents two training strategies—Dimple and Dimple+, but does not offer a direct comparison under the same training data. Could the authors provide results showing how the two models perform when trained on identical datasets? This would help clarify whether initializing from a pretrained AR multimodal model is consistently more effective than learning multimodal alignment from scratch.\n2. **Impact of Training Scale and Potential Forgetting**\nDimple+ inherits strong multimodal capabilities from Qwen2.5-VL, while Dimple builds them from new data. Given that Dimple+ is trained on limited data, is there a risk that diffusion tuning may gradually overwrite or degrade the original model’s capabilities? Could the authors comment on whether the performance gap between Dimple and Dimple+ might narrow or reverse with larger-scale training?\n3. **Trade-off Between Performance and Latency**\nThe supplementary material reports performance under a maximum decoding step (k=8), but does not explore how performance scales with larger k. Could the authors provide results showing whether increasing k beyond 8 leads to further improvements? Additionally, a comparison of inference latency between Dimple+ and standard AR models like Qwen2.5-VL would be valuable to understand the practical efficiency trade-offs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Dimple and Dimple+, multimodal large language model with discrete diffusion decoding. \n\nThe motivation is to address the instability and inefficiency of pure diffusion models in multimodal tasks. \nAnd the paper validates two distinct strategies for converting AR models into diffusion-based multimodal models. (dimple and dimple+)\n\nIt also proposes a decoding strategy called “confident decoding” to reduce inference steps without obvious performance degradation.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "1. The paper validates two practical strategies for converting AR models into diffusion-based multimodal models, providing clear guidance for the community. \n* Dimple: Starts from a diffusion-only LLM without multimodal capability. It first applies autoregressive training to align vision and language, then switches to diffusion-based instruction tuning. \n* Dimple+: Starts from a well-pretrained AR multimodal model (Qwen2.5-VL) and directly modifies its attention mask and loss function to enable diffusion training.\n2. It introduces Confident Decoding, a dynamic decoding method that improves inference efficiency without compromising output quality.\n3. From the code in supplementary material, the model is integrated into the Transformers library and will be fully open-sourced, providing a reproducible and accessible reference for the community.\n4. Through comparisons with AR baselines under identical training data, and with state-of-the-art diffusion models like LLaDA-V, Dimple+ demonstrates the effectiveness of diffusion training for multimodal tasks.", "weaknesses": "1. Although Dimple+ improves over its AR baseline, it still falls short compared to its initial AR models (Qwen2.5-VL) on several benchmarks. In particular, the performance gap on ChartQA is substantial, despite the training data includes similar samples (2 epoch results is much better than 1 epoch training), raising concerns about catastrophic forgetting and the scalability of diffusion models.\n\n2. The paper lacks analysis on the optimal number of decoding steps (k). While Confident Decoding dynamically adjusts k, it remains unclear whether increasing k beyond 8 could further improve performance, especially in complex tasks.\n\n3. There is no direct comparison between the inference speed of Dimple+ and standard AR models. Without this, it’s difficult to assess the practical efficiency gains of diffusion decoding.\n\n4. Despite the promising speedups and results reported for Confident Decoding, empirical testing with the open-sourced Dimple model shows noticeable degradation in long-context text generation. For example, when max_tokens is set to 64 or higher, even with step counts of 128 or 256, the model often fails to produce correct EOS tokens and instead generates excessive hallucinations. A controlled comparison with Dream-7B or other AR baseline would help clarify whether Dimple suffers from diffusion-induced degradation in textual capabilities.", "questions": "Overall, I think this is a strong paper that makes a meaningful contribution to the development of multimodal diffusion language models. However, I have a few questions that I hope the authors can clarify:\n1. **Conclusion and Fair Comparison Between Dimple and Dimple+** \nThe paper presents two training strategies—Dimple and Dimple+, but does not offer a direct comparison under the same training data. Could the authors provide results showing how the two models perform when trained on identical datasets? This would help clarify whether initializing from a pretrained AR multimodal model is consistently more effective than learning multimodal alignment from scratch.\n2. **Impact of Training Scale and Potential Forgetting**\nDimple+ inherits strong multimodal capabilities from Qwen2.5-VL, while Dimple builds them from new data. Given that Dimple+ is trained on limited data, is there a risk that diffusion tuning may gradually overwrite or degrade the original model’s capabilities? Could the authors comment on whether the performance gap between Dimple and Dimple+ might narrow or reverse with larger-scale training?\n3. **Trade-off Between Performance and Latency**\nThe supplementary material reports performance under a maximum decoding step (k=8), but does not explore how performance scales with larger k. Could the authors provide results showing whether increasing k beyond 8 leads to further improvements? Additionally, a comparison of inference latency between Dimple+ and standard AR models like Qwen2.5-VL would be valuable to understand the practical efficiency trade-offs.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761094234566}], "openreview_url": "https://openreview.net/forum?id=YQ5wH5C69t", "arxiv_id": "2505.16990", "paper_pdf": "papers/YQ5wH5C69t.pdf", "paper_pdf_sha256": "c8c968bea90e010653c90bc3b15b859fcfefb26d262e43c486b459d71b4b74ba", "paper_pdf_bytes": 978374, "paper_pdf_source": "openreview", "code_url": "https://github.com/yu-rp/Dimple", "code_repository": "yu-rp/Dimple", "code_commit": "e0ec36582a9ef9d9f205e65e2bc91b3726aab908", "code_archive": "repos/YQ5wH5C69t.zip", "code_archive_sha256": "1ffc3fc0b6dc3d4b634f979e6c5eb34e472b6d5274bc6160ec32528002c1b587", "code_archive_bytes": 4802750, "code_file_count": 480, "code_extensions": {".py": 463, ".ipynb": 14, ".sh": 3}, "github_disk_usage_kb": 4168, "github_languages": {"Python": 3754846, "Jupyter Notebook": 255855, "Shell": 4982}, "github_archived": false, "github_pushed_at": "2026-08-23T13:19:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dimple-discrete-diffusion-multimodal-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "baSU1eVLwS", "year": 2025, "status": "rejected", "title": "TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting", "authors": ["Peiyuan Liu", "Beiliang Wu", "Yifan Hu", "Naiqi Li", "Tao Dai", "Jigang Bao", "Shu-Tao Xia"], "authorids": ["~Peiyuan_Liu1", "~Beiliang_Wu1", "~Yifan_Hu8", "~Naiqi_Li1", "~Tao_Dai3", "~Jigang_Bao1", "~Shu-Tao_Xia1"], "authors_source": "OpenReview API", "abstract": "Non-stationarity poses significant challenges for multivariate time series forecasting due to the inherent short-term fluctuations and long-term trends that can lead to spurious regressions or obscure essential long-term relationships. Most existing methods either eliminate or retain non-stationarity without adequately addressing its distinct impacts on short-term and long-term modeling. Eliminating non-stationarity is essential for avoiding spurious regressions and capturing local dependencies in short-term modeling, while preserving it is crucial for revealing long-term cointegration across variates. In this paper, we propose TimeBridge, a novel framework designed to bridge the gap between non-stationarity and dependency modeling in long-term time series forecasting. By segmenting input series into smaller patches, TimeBridge applies Integrated Attention to mitigate short-term non-stationarity and capture stable dependencies within each variate, while Cointegrated Attention preserves non-stationarity to model long-term cointegration across variates. Extensive experiments show that TimeBridge consistently achieves state-of-the-art performance in both short-term and long-term forecasting. Additionally, TimeBridge demonstrates exceptional performance in financial forecasting on the CSI 500 and S\\&P 500 indices, further validating its robustness and effectiveness. The code is available in the supplementary material.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "tdxWLBk5qq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3735/Reviewer_5KbC"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "The paper introduces TimeBridge, a framework for multivariate time series forecasting that addresses the challenge of non-stationarity by differentiating between its impacts on short-term and long-term modeling. TimeBridge utilizes Integrated Attention to manage short-term fluctuations in batches, reducing spurious regressions and capturing local dependencies. Cointegrated Attention is introduced to preserve long-term non-stationarity across variates, enabling effective long-term dependency capture. Experiments on the CSI 500 and S&P 500 indices verify the short- and long-term forecasting performance. It is generally a method for handling nuanced non-stationarity effects in complex multivariate scenarios.", "review_text": "The paper introduces TimeBridge, a framework for multivariate time series forecasting that addresses the challenge of non-stationarity by differentiating between its impacts on short-term and long-term modeling. TimeBridge utilizes Integrated Attention to manage short-term fluctuations in batches, reducing spurious regressions and capturing local dependencies. Cointegrated Attention is introduced to preserve long-term non-stationarity across variates, enabling effective long-term dependency capture. Experiments on the CSI 500 and S&P 500 indices verify the short- and long-term forecasting performance. It is generally a method for handling nuanced non-stationarity effects in complex multivariate scenarios.", "strengths": "1. This paper is well written. The notations are clear.\n\n2. It provides up-to-date literature on learning techniques for capturing stationary/non-stationary and dependency in sequential data. It combines the treatments of non-stationarity and dependency modeling in one shoot and delivers convincing performances.\n\n3. The notion of cointegration of time series has been missing or forgotten in recent time series forecasting literature; this paper showed how cointegration could help reveal the non-stationary part when multiple time series evolve simultaneously. \nIn Operations Research, cointegration is a well-established technique, and it has been introduced to machine learning literature for long:\n\nMarco Cuturi, Alexandre D’Aspremont (ICML, 2013). https://proceedings.mlr.press/v28/cuturi13.html\n\nThe authors may refer to basic cointegration techniques to resolve the computational challenges or benchmark the extraction of stationary/non-stationary movements.\n\n4. The experiments are comprehensive and cover recent state-of-the-art competing methods. The results are convincing, as shown by a solid ablation study showing the contribution of the building blocks, e.g., Integrated Attention and Cointegrated Attention.", "weaknesses": "1. According to Figure 3, the building blocks of the proposed methods are streamlined with no conjugation. In some sense, this is brute force and remains room for improvement or further technical development, speaking of the systematic organic treatment of the non-stationarity and dependency modeling in long-term time-series forecasting.", "questions": "Q. As mentioned in Strength 3, it would be interesting to recap traditional techniques for cointegration in operations research and benchmark the stationary or non-stationary time series before discussing the impact of an attention-powered module.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces TimeBridge, a framework for multivariate time series forecasting that addresses the challenge of non-stationarity by differentiating between its impacts on short-term and long-term modeling. TimeBridge utilizes Integrated Attention to manage short-term fluctuations in batches, reducing spurious regressions and capturing local dependencies. Cointegrated Attention is introduced to preserve long-term non-stationarity across variates, enabling effective long-term dependency capture. Experiments on the CSI 500 and S&P 500 indices verify the short- and long-term forecasting performance. It is generally a method for handling nuanced non-stationarity effects in complex multivariate scenarios.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "1. This paper is well written. The notations are clear.\n\n2. It provides up-to-date literature on learning techniques for capturing stationary/non-stationary and dependency in sequential data. It combines the treatments of non-stationarity and dependency modeling in one shoot and delivers convincing performances.\n\n3. The notion of cointegration of time series has been missing or forgotten in recent time series forecasting literature; this paper showed how cointegration could help reveal the non-stationary part when multiple time series evolve simultaneously. \nIn Operations Research, cointegration is a well-established technique, and it has been introduced to machine learning literature for long:\n\nMarco Cuturi, Alexandre D’Aspremont (ICML, 2013). https://proceedings.mlr.press/v28/cuturi13.html\n\nThe authors may refer to basic cointegration techniques to resolve the computational challenges or benchmark the extraction of stationary/non-stationary movements.\n\n4. The experiments are comprehensive and cover recent state-of-the-art competing methods. The results are convincing, as shown by a solid ablation study showing the contribution of the building blocks, e.g., Integrated Attention and Cointegrated Attention.", "weaknesses": "1. According to Figure 3, the building blocks of the proposed methods are streamlined with no conjugation. In some sense, this is brute force and remains room for improvement or further technical development, speaking of the systematic organic treatment of the non-stationarity and dependency modeling in long-term time-series forecasting.", "questions": "Q. As mentioned in Strength 3, it would be interesting to recap traditional techniques for cointegration in operations research and benchmark the stationary or non-stationary time series before discussing the impact of an attention-powered module.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730734735787}, {"id": "1w2BFf8lsH", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3735/Reviewer_aC7S"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper presents TimeBridge for non-stationary time series forecasting. TimeBridge utilizes Integrated Attention to mitigate short-term non-stationarity and Cointegration Attention for modeling long-term non-stationarity. Experiment results on benchmarking time series demonstrate the effectiveness of TimeBridge.", "review_text": "The paper presents TimeBridge for non-stationary time series forecasting. TimeBridge utilizes Integrated Attention to mitigate short-term non-stationarity and Cointegration Attention for modeling long-term non-stationarity. Experiment results on benchmarking time series demonstrate the effectiveness of TimeBridge.", "strengths": "+ Non-stationarity is a major challenge in time series forecasting. The paper aims to address this important task\n+ Experiment results are encouraging, with a comprehensive ablation study.", "weaknesses": "- The idea of Integrated Attention and Cointegration Attention is not new. The model in general lacks novelty. \n\n- The improvement by TimeBridge is not as much as claimed in the paper (over 10%). On most of the datasets, TimeBridge achieves similar results or marginally better results (Table 1 and 2). In general, the popular benchmarking time series are relatively easier tasks. On finance applications, TimeBridge's performance is similar to TSmixer.", "questions": "What is TimeBridge's performance on bigger time series dataset, such as New York Taxi or Climate Data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents TimeBridge for non-stationary time series forecasting. TimeBridge utilizes Integrated Attention to mitigate short-term non-stationarity and Cointegration Attention for modeling long-term non-stationarity. Experiment results on benchmarking time series demonstrate the effectiveness of TimeBridge.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "+ Non-stationarity is a major challenge in time series forecasting. The paper aims to address this important task\n+ Experiment results are encouraging, with a comprehensive ablation study.", "weaknesses": "- The idea of Integrated Attention and Cointegration Attention is not new. The model in general lacks novelty. \n\n- The improvement by TimeBridge is not as much as claimed in the paper (over 10%). On most of the datasets, TimeBridge achieves similar results or marginally better results (Table 1 and 2). In general, the popular benchmarking time series are relatively easier tasks. On finance applications, TimeBridge's performance is similar to TSmixer.", "questions": "What is TimeBridge's performance on bigger time series dataset, such as New York Taxi or Climate Data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730709823153}, {"id": "aDZ2jahJPo", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission3735/Reviewer_nS1F"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces TimeBridge for predicting non-stationary time series, aiming to resolve the contradiction between short-term fluctuations and long-term trends. By dividing the input sequence into small segments and applying an integrated attention mechanism to reduce short-term instability while utilizing joint attention mechanisms to preserve non-stability to capture long-term co-integration relationships across variables, TimeBridge effectively captures stable dependencies without introducing pseudo-regression risks. Experimental results demonstrate that TimeBridge performs well on multiple tasks, particularly outperforming current state-of-the-art methods in financial time series prediction. Additionally, through a series of ablation studies, the importance of removing or preserving non-stability and the order of integrated and joint attention modules is further validated, revealing optimal model configurations under different dataset backgrounds. This work not only enhances the accuracy and robustness of time series prediction but also provides new perspectives and methods for handling non-stationary data in the future.", "review_text": "This paper introduces TimeBridge for predicting non-stationary time series, aiming to resolve the contradiction between short-term fluctuations and long-term trends. By dividing the input sequence into small segments and applying an integrated attention mechanism to reduce short-term instability while utilizing joint attention mechanisms to preserve non-stability to capture long-term co-integration relationships across variables, TimeBridge effectively captures stable dependencies without introducing pseudo-regression risks. Experimental results demonstrate that TimeBridge performs well on multiple tasks, particularly outperforming current state-of-the-art methods in financial time series prediction. Additionally, through a series of ablation studies, the importance of removing or preserving non-stability and the order of integrated and joint attention modules is further validated, revealing optimal model configurations under different dataset backgrounds. This work not only enhances the accuracy and robustness of time series prediction but also provides new perspectives and methods for handling non-stationary data in the future.", "strengths": "- Resolving the conflict between short-term fluctuations and long-term trends: TimeBridge can effectively capture stable dependencies by reducing short-term non-stability through dividing the input sequence into small fragments and applying integrated attention mechanisms, while utilizing joint attention mechanisms to preserve non-stability to capture cross-variable long-term co-integration relationships.\n\n- Improving the accuracy and robustness of time and sequence prediction: TimeBridge performs well on multiple tasks, particularly outperforming current state-of-the-art methods in financial time series prediction.\n\n- Providing new perspectives and methods for handling non-stationary data in the future: Through a series of ablation studies, this article further verifies the importance of removing or retaining non-stability, as well as the sequence of integrated and joint attention modules, and reveals the optimal model configuration under different datasets, providing new perspectives and methods for handling non-stationary data in the future.", "weaknesses": "- The paper lacks sufficient originality, as it primarily builds upon existing methods without presenting a clear, novel contribution to the field. This limits the manuscript's potential impact and reduces its value as an advancement in research.\n\n- The lack of experiments with varying input/output (I/O) ratios raises concerns about the fairness of the comparisons. Different I/O settings can affect model outcomes substantially, and without exploring these variations, the paper provides an incomplete evaluation of model performance.\n\n- The authors have only considered MSE and MAE as metrics for their long-term forecasting task, which is insufficient. This limited evaluation may obscure important aspects of model accuracy, particularly in cases where relative error size is a crucial factor for assessing model effectiveness.\n\n- The paper lacks a clear description of the hyperparameter tuning process for the comparative algorithms, which raises concerns about the comparisons. Without adequate tuning, it is difficult to ascertain if the reported performance differences are truly reflective of each model's capabilities.", "questions": "- The experimental settings are not reasonable. Why input length is set to 720, output length O is set to 96, 192, 336, 720 and not something more practical, like 1 month of data? \n\n- The paper does not account for the impact of varying input/output (I/O) ratios, which can significantly influence model performance. Fixing different input length for baselines without comprehensive experiments may lead to biased comparisons and limit the robustness of the results. The author needs to report the results of the baseline under the same experimental setup.\n\n- The paper lacks an evaluation of the algorithm's complexity, particularly with respect to its theoretical complexity, GPU resource cost, and runtime efficiency. Without this assessment, it is difficult to gauge the algorithm's practicality and scalability, especially in resource-constrained environments.\n\n- The absence of detailed hyperparameter optimization for the baseline algorithms undermines the fairness of the comparisons presented in the paper. Without this tuning, the results may not accurately represent the optimal performance of the comparative models, potentially leading to biased conclusions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces TimeBridge for predicting non-stationary time series, aiming to resolve the contradiction between short-term fluctuations and long-term trends. By dividing the input sequence into small segments and applying an integrated attention mechanism to reduce short-term instability while utilizing joint attention mechanisms to preserve non-stability to capture long-term co-integration relationships across variables, TimeBridge effectively captures stable dependencies without introducing pseudo-regression risks. Experimental results demonstrate that TimeBridge performs well on multiple tasks, particularly outperforming current state-of-the-art methods in financial time series prediction. Additionally, through a series of ablation studies, the importance of removing or preserving non-stability and the order of integrated and joint attention modules is further validated, revealing optimal model configurations under different dataset backgrounds. This work not only enhances the accuracy and robustness of time series prediction but also provides new perspectives and methods for handling non-stationary data in the future.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- Resolving the conflict between short-term fluctuations and long-term trends: TimeBridge can effectively capture stable dependencies by reducing short-term non-stability through dividing the input sequence into small fragments and applying integrated attention mechanisms, while utilizing joint attention mechanisms to preserve non-stability to capture cross-variable long-term co-integration relationships.\n\n- Improving the accuracy and robustness of time and sequence prediction: TimeBridge performs well on multiple tasks, particularly outperforming current state-of-the-art methods in financial time series prediction.\n\n- Providing new perspectives and methods for handling non-stationary data in the future: Through a series of ablation studies, this article further verifies the importance of removing or retaining non-stability, as well as the sequence of integrated and joint attention modules, and reveals the optimal model configuration under different datasets, providing new perspectives and methods for handling non-stationary data in the future.", "weaknesses": "- The paper lacks sufficient originality, as it primarily builds upon existing methods without presenting a clear, novel contribution to the field. This limits the manuscript's potential impact and reduces its value as an advancement in research.\n\n- The lack of experiments with varying input/output (I/O) ratios raises concerns about the fairness of the comparisons. Different I/O settings can affect model outcomes substantially, and without exploring these variations, the paper provides an incomplete evaluation of model performance.\n\n- The authors have only considered MSE and MAE as metrics for their long-term forecasting task, which is insufficient. This limited evaluation may obscure important aspects of model accuracy, particularly in cases where relative error size is a crucial factor for assessing model effectiveness.\n\n- The paper lacks a clear description of the hyperparameter tuning process for the comparative algorithms, which raises concerns about the comparisons. Without adequate tuning, it is difficult to ascertain if the reported performance differences are truly reflective of each model's capabilities.", "questions": "- The experimental settings are not reasonable. Why input length is set to 720, output length O is set to 96, 192, 336, 720 and not something more practical, like 1 month of data? \n\n- The paper does not account for the impact of varying input/output (I/O) ratios, which can significantly influence model performance. Fixing different input length for baselines without comprehensive experiments may lead to biased comparisons and limit the robustness of the results. The author needs to report the results of the baseline under the same experimental setup.\n\n- The paper lacks an evaluation of the algorithm's complexity, particularly with respect to its theoretical complexity, GPU resource cost, and runtime efficiency. Without this assessment, it is difficult to gauge the algorithm's practicality and scalability, especially in resource-constrained environments.\n\n- The absence of detailed hyperparameter optimization for the baseline algorithms undermines the fairness of the comparisons presented in the paper. Without this tuning, the results may not accurately represent the optimal performance of the comparative models, potentially leading to biased conclusions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730520235904}], "openreview_url": "https://openreview.net/forum?id=baSU1eVLwS", "arxiv_id": "2410.04442", "paper_pdf": "papers/baSU1eVLwS.pdf", "paper_pdf_sha256": "301de0175d6aa8efdf82709625cb618a7d0b266789cebbda0cdcd4bbde239f2f", "paper_pdf_bytes": 7923752, "paper_pdf_source": "openreview", "code_url": "https://github.com/Hank0626/TimeBridge", "code_repository": "Hank0626/TimeBridge", "code_commit": "0f9a83fbc3e1260c9ddd527c522dff0ce4b9554b", "code_archive": "repos/baSU1eVLwS.zip", "code_archive_sha256": "bb20739597b546e0291fe3dc6e5093e6e27c83975786a8dcfcab09d75cd214fe", "code_archive_bytes": 177652, "code_file_count": 16, "code_extensions": {".py": 15, ".sh": 1}, "github_disk_usage_kb": 291, "github_languages": {"Python": 87904, "Shell": 6642}, "github_archived": false, "github_pushed_at": "2025-05-16T04:33:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/timebridge-non-stationarity-matters-for-long"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tZ3JmSDbJM", "year": 2024, "status": "rejected", "title": "GRAPES: Learning to Sample Graphs for Scalable Graph Neural Networks", "authors": ["Taraneh Younesian", "Thiviyan Thanapalasingam", "Emile van Krieken", "Daniel Daza", "Peter Bloem"], "authorids": ["~Taraneh_Younesian2", "~Thiviyan_Thanapalasingam1", "~Emile_van_Krieken1", "~Daniel_Daza1", "~Peter_Bloem1"], "authors_source": "OpenReview API", "abstract": "Graph neural networks (GNNs) learn the representation of nodes in a graph by aggregating the neighborhood information in various ways. As these networks grow in depth, their receptive field grows exponentially due to the increase in neighborhood sizes, resulting in high memory costs. Graph sampling solves memory issues in GNNs by sampling a small ratio of the nodes in the graph. This way, GNNs can scale to much larger graphs. Most sampling methods focus on fixed sampling heuristics, which may not generalize to different structures or tasks. We introduce GRAPES, an adaptive graph sampling method that learns to identify sets of influential nodes for training a GNN classifier.  GRAPES uses a GFlowNet to learn node sampling probabilities given the classification objectives. We evaluate GRAPES across several small- and large-scale graph benchmarks and demonstrate its effectiveness in accuracy and scalability. In contrast to existing sampling methods, GRAPES maintains high accuracy even with small sample sizes and, therefore, can scale to very large graphs. Our code is publicly available at https://anonymous.4open.science/r/GRAPES.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "9kN3d76NlV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2529/Reviewer_uctj"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The work proposes a new classification model for large-scale graph data. It combines a learnable node-selection component and a graph neural network to construct a unified classification model. The node-selection component is devised with a GFlowNet, which is further parameterized by a graph neural network. Essentially it is an RL learning component that helps minimize the training loss. With this construction, it can greatly reduce the number of nodes in the classification component but without much performance drop.", "review_text": "The work proposes a new classification model for large-scale graph data. It combines a learnable node-selection component and a graph neural network to construct a unified classification model. The node-selection component is devised with a GFlowNet, which is further parameterized by a graph neural network. Essentially it is an RL learning component that helps minimize the training loss. With this construction, it can greatly reduce the number of nodes in the classification component but without much performance drop.", "strengths": "Compared with previous models such as AS-GCN, the sampling method of this model is more \"adaptive\". Although the model is much more complex than previous models, it does show some performance improvement.", "weaknesses": "1. The description of the algorithm is problematic. As far as I know, GFlowNet is a method proposed to sample for an energy-based model: it addresses a distribution approximation problem. It is a special case of an RL algorithm. In my view, the paper is a pure RL problem especially since the reward function is clearly defined. I think an RL formulation is straightforward from that. The formulation with GFlowNet is very misleading -- I spent hours before realizing that this is not a distribution approximation problem.  Actually, the reward scaling in 4.2 could be avoided within an RL formulation. \n\n2. The method is much more complex than previous methods because it has this extra learnable component. I don't know whether it is easy for others to apply such a model to a different application. To me, the simplicity of model tuning is more appealing than minor performance improvement: one may not see the improvement if the model cannot be well-tuned. \n\n3. The performance values of baseline methods reported in Table 1 are much lower than those reported in their original papers. I don't know how much I can trust the comparison. For example, Graph-SAINT has f1 scores, 0.511±0.001, 0.966±0.001,  and 0.653±0.003 on the Flickr, Reddit, and Yelp datasets. These numbers are much higher than the reported numbers in the submission.", "questions": "On the ogbn-products dataset, can you tune the number of samples (n) so that AS-GCN can also run on this dataset?\n\nCan you put data statistics in the experiment section? These numbers are important to the understanding of experiment results.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work proposes a new classification model for large-scale graph data. It combines a learnable node-selection component and a graph neural network to construct a unified classification model. The node-selection component is devised with a GFlowNet, which is further parameterized by a graph neural network. Essentially it is an RL learning component that helps minimize the training loss. With this construction, it can greatly reduce the number of nodes in the classification component but without much performance drop.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "Compared with previous models such as AS-GCN, the sampling method of this model is more \"adaptive\". Although the model is much more complex than previous models, it does show some performance improvement.", "weaknesses": "1. The description of the algorithm is problematic. As far as I know, GFlowNet is a method proposed to sample for an energy-based model: it addresses a distribution approximation problem. It is a special case of an RL algorithm. In my view, the paper is a pure RL problem especially since the reward function is clearly defined. I think an RL formulation is straightforward from that. The formulation with GFlowNet is very misleading -- I spent hours before realizing that this is not a distribution approximation problem.  Actually, the reward scaling in 4.2 could be avoided within an RL formulation. \n\n2. The method is much more complex than previous methods because it has this extra learnable component. I don't know whether it is easy for others to apply such a model to a different application. To me, the simplicity of model tuning is more appealing than minor performance improvement: one may not see the improvement if the model cannot be well-tuned. \n\n3. The performance values of baseline methods reported in Table 1 are much lower than those reported in their original papers. I don't know how much I can trust the comparison. For example, Graph-SAINT has f1 scores, 0.511±0.001, 0.966±0.001,  and 0.653±0.003 on the Flickr, Reddit, and Yelp datasets. These numbers are much higher than the reported numbers in the submission.", "questions": "On the ogbn-products dataset, can you tune the number of samples (n) so that AS-GCN can also run on this dataset?\n\nCan you put data statistics in the experiment section? These numbers are important to the understanding of experiment results.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698864578159}, {"id": "lYGKRKv3BU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2529/Reviewer_4YMr"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a novel mechanism for sampling subgraphs from a large graph during GNN training, with the goal of increasing scalability.\n\nThis sampling mechanism samples a fixed number k of nodes to add to the subgraph at each layer following a reinforcement learning policy parametrized by a GFlowNet and optimized using a trajectory balance loss. In addition to this framework, a key contribution of the paper is employing the loss of the GNN in the downstream task as the reward. Therefore, the model learns to sample nodes adaptively, in a manner that improves performance in the downstream task.\n\nThe efficacy of the architecture is verified via extensive numerical experiments on small, moderate-size, and large graph node classification tasks.", "review_text": "This paper introduces a novel mechanism for sampling subgraphs from a large graph during GNN training, with the goal of increasing scalability.\n\nThis sampling mechanism samples a fixed number k of nodes to add to the subgraph at each layer following a reinforcement learning policy parametrized by a GFlowNet and optimized using a trajectory balance loss. In addition to this framework, a key contribution of the paper is employing the loss of the GNN in the downstream task as the reward. Therefore, the model learns to sample nodes adaptively, in a manner that improves performance in the downstream task.\n\nThe efficacy of the architecture is verified via extensive numerical experiments on small, moderate-size, and large graph node classification tasks.", "strengths": "- The motivating ideas for the proposed sampling framework are very novel. They are a great example of integrating different concepts/techniques from modern deep learning to solve a relevant problem --- scalability of GNNs.\n- Although this is primarily an algorithm-based/application paper, the model, the algorithm, and the training mechanism are theoretically grounded, and the authors did a very good job at motivating and explaining the reasons behind their design choices.\n- The numerical results are extensive and convincing. I appreciate the inclusion of hypothesis tests for the rank of their method with respect to the baselines; the memory plots comparing GRAPES with GAS; the transferability plots of performance versus subgraph size; and the entropy plots. In particular, the transferability plots (Fig. 3) are very convincing in showing the superiority of GRAPES, as its performance is much more robust to reducing K. Further, the entropy plots are in direct agreement with the authors's claim that GRAPES is consistent in identifying important nodes.", "weaknesses": "- Some related work is missing, and perhaps also a comparison with other graph sampling baselines from the graph signal processing literature. Check, e.g., \"Efficient Sampling Set Selection for Bandlimited Graph Signals Using Graph Spectral Proxies\", by Anis and others, and papers therein (specifically, the works of Kovacevic and Moura; Chamon and Ribeiro; Segarra, Marques and Ribeiro; etc.). These papers are part of a subfield of graph signal processing---graph signal sampling---which studies how to sample graphs so as to maximize the preservation of their spectra. Since graph spectral information is typically very correlated with performance in graph machine learning tasks, I believe these are important references/comparisons to include.\n- The explanation of why the method is trained off-policy is not very clear for readers not familiar with reinforcement learning. There is a result which is only mentioned in passing---\"Importantly, GFlowNets [...] can learn from off-policy distributions without adjusting the objective\"---which is important in justifying the choice of off-policy training, and hence should be described in further detail (perhaps a short subsection) in the camera-ready. It would also be interesting to see empirical comparisons between training off-policy and using gradient estimation methods.\n- The numerical experiments only consider node classification tasks.\n- Other relevant line of related work is that on the \"transferability properties of GNNs\". See e.g. the work of Ruiz et al.", "questions": "- Have you analyzed the specific subgraphs that are sampled by GRAPES in different tasks? What are their characteristics (are they connected? do the sampled nodes have high centrality? etc.). GRAPES sounds like a nice tool for understanding which characteristics of a graph are most important in a given task.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel mechanism for sampling subgraphs from a large graph during GNN training, with the goal of increasing scalability.\n\nThis sampling mechanism samples a fixed number k of nodes to add to the subgraph at each layer following a reinforcement learning policy parametrized by a GFlowNet and optimized using a trajectory balance loss. In addition to this framework, a key contribution of the paper is employing the loss of the GNN in the downstream task as the reward. Therefore, the model learns to sample nodes adaptively, in a manner that improves performance in the downstream task.\n\nThe efficacy of the architecture is verified via extensive numerical experiments on small, moderate-size, and large graph node classification tasks.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "- The motivating ideas for the proposed sampling framework are very novel. They are a great example of integrating different concepts/techniques from modern deep learning to solve a relevant problem --- scalability of GNNs.\n- Although this is primarily an algorithm-based/application paper, the model, the algorithm, and the training mechanism are theoretically grounded, and the authors did a very good job at motivating and explaining the reasons behind their design choices.\n- The numerical results are extensive and convincing. I appreciate the inclusion of hypothesis tests for the rank of their method with respect to the baselines; the memory plots comparing GRAPES with GAS; the transferability plots of performance versus subgraph size; and the entropy plots. In particular, the transferability plots (Fig. 3) are very convincing in showing the superiority of GRAPES, as its performance is much more robust to reducing K. Further, the entropy plots are in direct agreement with the authors's claim that GRAPES is consistent in identifying important nodes.", "weaknesses": "- Some related work is missing, and perhaps also a comparison with other graph sampling baselines from the graph signal processing literature. Check, e.g., \"Efficient Sampling Set Selection for Bandlimited Graph Signals Using Graph Spectral Proxies\", by Anis and others, and papers therein (specifically, the works of Kovacevic and Moura; Chamon and Ribeiro; Segarra, Marques and Ribeiro; etc.). These papers are part of a subfield of graph signal processing---graph signal sampling---which studies how to sample graphs so as to maximize the preservation of their spectra. Since graph spectral information is typically very correlated with performance in graph machine learning tasks, I believe these are important references/comparisons to include.\n- The explanation of why the method is trained off-policy is not very clear for readers not familiar with reinforcement learning. There is a result which is only mentioned in passing---\"Importantly, GFlowNets [...] can learn from off-policy distributions without adjusting the objective\"---which is important in justifying the choice of off-policy training, and hence should be described in further detail (perhaps a short subsection) in the camera-ready. It would also be interesting to see empirical comparisons between training off-policy and using gradient estimation methods.\n- The numerical experiments only consider node classification tasks.\n- Other relevant line of related work is that on the \"transferability properties of GNNs\". See e.g. the work of Ruiz et al.", "questions": "- Have you analyzed the specific subgraphs that are sampled by GRAPES in different tasks? What are their characteristics (are they connected? do the sampled nodes have high centrality? etc.). GRAPES sounds like a nice tool for understanding which characteristics of a graph are most important in a given task.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698839317915}, {"id": "Owxw0TQufL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2529/Reviewer_iUex"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposes an adaptive sampling algorithm to learn an influential subgraph in each layer of a GNN classifier. Instead of fixed heuristics, the proposed method learns the preferences between nodes via the influence they have on the overall performance of the classifier. Designed based on the GFlowNet Architecture, it identifies a sequence of states that represent a sequence of samples in each layer of the GNN. They have shown an improvement in terms of F1 score and GPU memory utilization compared to other non-adaptive algorithms like FastGCN and LADIES, and even the adaptive AS-GCN method.", "review_text": "The paper proposes an adaptive sampling algorithm to learn an influential subgraph in each layer of a GNN classifier. Instead of fixed heuristics, the proposed method learns the preferences between nodes via the influence they have on the overall performance of the classifier. Designed based on the GFlowNet Architecture, it identifies a sequence of states that represent a sequence of samples in each layer of the GNN. They have shown an improvement in terms of F1 score and GPU memory utilization compared to other non-adaptive algorithms like FastGCN and LADIES, and even the adaptive AS-GCN method.", "strengths": "1. The method seems to have an improvement on F1 scores of GRAPES on most of the datasets compared to the other algorithms in the presented experimental setup. It has also proved to consume much less memory compared to GAS which has a different non-sampling strategy to reduce the scalability problem in large graphs. Although, it outperforms GRAPES in some datasets.\n\n1. Different types of results such as the F1 scores, GPU memory allocation, robustness and entropy are provided to demonstrate the effectiveness of the proposed algorithm.\n\n1. The paper, in general, is well-written and sufficient for the reader to understand the concepts involved.", "weaknesses": "1. **Experimental Setup**: In the presented setup, the proposed method outperforms the baselines. However, a few things about the setup are not clear:\n    1. It is not clear if the baselines were tuned on a validation set. Why was the batch size fixed to 256 for the main results table?\n    1. A related concern is the appearance of low F1-score compared to what is reported in other paper. Granted that this is in the transductive setting, I am not sure, if that should cause such decrease in performance. For instance, according to the GraphSAINT paper, it achieves 96.6% on Reddit in the inductive setting. In this paper, the result is much lower (80.50).\n    1. Comparison against GraphSAINT: Fixing 256 size and 256 samples does not seem fair for GraphSAINT. It uses a sample per minibatch instead of such a low number of samples. Also, GraphSAINT paper shows sample size of few thousands, while this paper uses sample size only up to $2^9$.  Also, the node sampler seems to have the lowest performance compared to other samplers of GraphSAINT, so other samplers should have been considered (edge, RW, Multidimensional RW).\n    1. Architectures beyond GCN should be considered. If the sampling approach improves over baselines for multiple architectures such as GAT, GIN, SAGE, then it would create a strong case for the proposed sampling approach. As of now, the central claim does not seem justified, \"GRAPES outperforms state-of-the-art sampling-based methods.\" What has been shown is that GRAPES outperforms other methods on a specific GCN architecture and under small sample sizes and number of samples.\n    1. Most of the baselines presented are relatively old.\n\n1. **Discussion on runtime** - The downside of being adaptive is that there is extra computation involved per batch. However, no discussion of training time has been presented. This would have helped with understanding execution time - F1 tradeoff.", "questions": "1. Were the baselines tuned on a validation set?\n1. Why are the baseline performance lower compared to what is seen in other papers? I would expect the transductive setting to improve the results compared to the inductive setting.\n1. How is the performance vs GraphSAINT with a higher number of samples and larger sample size?\n1. Have other architectures been considered (other than GCN)?\n1. Can you present a comparison of training times of the proposed approach vs the baselines?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an adaptive sampling algorithm to learn an influential subgraph in each layer of a GNN classifier. Instead of fixed heuristics, the proposed method learns the preferences between nodes via the influence they have on the overall performance of the classifier. Designed based on the GFlowNet Architecture, it identifies a sequence of states that represent a sequence of samples in each layer of the GNN. They have shown an improvement in terms of F1 score and GPU memory utilization compared to other non-adaptive algorithms like FastGCN and LADIES, and even the adaptive AS-GCN method.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The method seems to have an improvement on F1 scores of GRAPES on most of the datasets compared to the other algorithms in the presented experimental setup. It has also proved to consume much less memory compared to GAS which has a different non-sampling strategy to reduce the scalability problem in large graphs. Although, it outperforms GRAPES in some datasets.\n\n1. Different types of results such as the F1 scores, GPU memory allocation, robustness and entropy are provided to demonstrate the effectiveness of the proposed algorithm.\n\n1. The paper, in general, is well-written and sufficient for the reader to understand the concepts involved.", "weaknesses": "1. **Experimental Setup**: In the presented setup, the proposed method outperforms the baselines. However, a few things about the setup are not clear:\n    1. It is not clear if the baselines were tuned on a validation set. Why was the batch size fixed to 256 for the main results table?\n    1. A related concern is the appearance of low F1-score compared to what is reported in other paper. Granted that this is in the transductive setting, I am not sure, if that should cause such decrease in performance. For instance, according to the GraphSAINT paper, it achieves 96.6% on Reddit in the inductive setting. In this paper, the result is much lower (80.50).\n    1. Comparison against GraphSAINT: Fixing 256 size and 256 samples does not seem fair for GraphSAINT. It uses a sample per minibatch instead of such a low number of samples. Also, GraphSAINT paper shows sample size of few thousands, while this paper uses sample size only up to $2^9$.  Also, the node sampler seems to have the lowest performance compared to other samplers of GraphSAINT, so other samplers should have been considered (edge, RW, Multidimensional RW).\n    1. Architectures beyond GCN should be considered. If the sampling approach improves over baselines for multiple architectures such as GAT, GIN, SAGE, then it would create a strong case for the proposed sampling approach. As of now, the central claim does not seem justified, \"GRAPES outperforms state-of-the-art sampling-based methods.\" What has been shown is that GRAPES outperforms other methods on a specific GCN architecture and under small sample sizes and number of samples.\n    1. Most of the baselines presented are relatively old.\n\n1. **Discussion on runtime** - The downside of being adaptive is that there is extra computation involved per batch. However, no discussion of training time has been presented. This would have helped with understanding execution time - F1 tradeoff.", "questions": "1. Were the baselines tuned on a validation set?\n1. Why are the baseline performance lower compared to what is seen in other papers? I would expect the transductive setting to improve the results compared to the inductive setting.\n1. How is the performance vs GraphSAINT with a higher number of samples and larger sample size?\n1. Have other architectures been considered (other than GCN)?\n1. Can you present a comparison of training times of the proposed approach vs the baselines?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698819419532}, {"id": "1t7Dyqjoxf", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2529/Reviewer_J9DQ"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Paper is interested in sampling (smaller) subgraphs from (large) input graph when training Graph Neural Networks (GNNs), as a way of scaling of GNN training onto larger graphs. Unlike most earlier methods, where their sampling logic is fixed and non-trainable, the proposed method has a subgraph sampling function that is trainable. It samples subgraphs level-by-level, starting from an input batch. They sample next nodes from probability distribution over all nodes. It is conditioned on the nodes visited prior. They parameterize the distribution using GFLowNet [Bengio et al, 2021].", "review_text": "Paper is interested in sampling (smaller) subgraphs from (large) input graph when training Graph Neural Networks (GNNs), as a way of scaling of GNN training onto larger graphs. Unlike most earlier methods, where their sampling logic is fixed and non-trainable, the proposed method has a subgraph sampling function that is trainable. It samples subgraphs level-by-level, starting from an input batch. They sample next nodes from probability distribution over all nodes. It is conditioned on the nodes visited prior. They parameterize the distribution using GFLowNet [Bengio et al, 2021].", "strengths": "## Strengths\n**Problem space**: sampling subgraphs from large graphs when training GNNs, which can scale GNNs to very large graphs, is important, for both static graphs and dynamic graphs.\n\n**Novel Model definition**: they present \"sampling subgraphs\" as leaf-nodes of Finite State Machines (FSMs), and their FSMs look like trees (strong model assumptions*). This model definition is novel.\n\n**Clarity of writing**. The paper is concise and up-to-the-point. Algorithm 1 is ties the pieces well.\n\n\n## Summary\n\nThe novelty of construction appeals me to recommend this paper for acceptance. However, it has many small weaknesses. To do a better justice to the goodness of your work, if you have time to address all (or most) of my concerns and questions, within the main paper, I should be able to change my review.", "weaknesses": "Here, I point out major points that need revision. In addition, in the next **Questions** section, I ask for clarifications on more minor (but still important) issues.\n\n\n## The specialization of GFLowNet onto trees be explicit stated\n\nAbove Eq. 6, it says that $P_B(s_{l} | s_{l+1}) = 1$ -- this implies that the only way to get to $s_{l+1}$ is through $s_{l}$. This strong modeling assumption stems from \"s\" being the entire path of generated adjacencies $\\{A_0, \\dots, A_l \\}$. This produces a special case of \"finite state machines\" that specifically have states looking like trees.\n\n## Larger graphs?\n\nOnly one graph more than 1 million nodes whereas one central theme of the paper is about scale.\n\n\n## Application appeal\n\nI would wish that the paper considers applying their method to a problem space beyond graph sampling (or otherwise, show compelling use-cases for graph sampling). Specifically, could this method be used to *explain graphs*? E.g., in the integrated-gradient sense: the presence of which nodes or edges would cause a certain prediction.\n\n## Missing section on inference\n\nWhile the paper includes the information about training, it should also include information on how to do inference. Given a node at a (large) input test graph, are samples taken or the full graph around $n$?\n\n## Missing References on learnable sampling\n\nE.g.,\n\n* DSKReG; CIKM'2021\n* \"Performance-adaptive sampling strategy towards fast and accurate GNNs.\", KDD'2021\n* Submix; UAI'2023", "questions": "The following items are not clear. Please clarify them in the paper\n\nQ1:\nHow is the $GNN_F$ parameterized? Does it train a scalar for every node (i.e., lookup 1D embedding table) or is it a function of features? In my understanding, GNN_F models $P_F$ (correct me if I am wrong) i.e. should have support on the nodes\n\nQ2:\nIs the reward measured only on end states? (on the \"sum\" of sampled list of adjacency matrices) or on every intermediate state (e.g., sum of adjacencies at that point).\n\nQ3: **Runtime experiments** Would you report runtimes? E.g., on the largest dataset ogbn-products?\n\nQ4: **Repeat edges**. Can an edge be sampled twice? Does this have any impact on the GCN model?\n\nQ5: Is adjacency matrix $A^0$ same as $A_0$ (Algorithm 1)\n\nQ6: $Z(s_0)$ in text following Eq. 3 -- It is not clear whether scalar $log Z$ is modeled or if it is a constant and removed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Paper is interested in sampling (smaller) subgraphs from (large) input graph when training Graph Neural Networks (GNNs), as a way of scaling of GNN training onto larger graphs. Unlike most earlier methods, where their sampling logic is fixed and non-trainable, the proposed method has a subgraph sampling function that is trainable. It samples subgraphs level-by-level, starting from an input batch. They sample next nodes from probability distribution over all nodes. It is conditioned on the nodes visited prior. They parameterize the distribution using GFLowNet [Bengio et al, 2021].", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "## Strengths\n**Problem space**: sampling subgraphs from large graphs when training GNNs, which can scale GNNs to very large graphs, is important, for both static graphs and dynamic graphs.\n\n**Novel Model definition**: they present \"sampling subgraphs\" as leaf-nodes of Finite State Machines (FSMs), and their FSMs look like trees (strong model assumptions*). This model definition is novel.\n\n**Clarity of writing**. The paper is concise and up-to-the-point. Algorithm 1 is ties the pieces well.\n\n\n## Summary\n\nThe novelty of construction appeals me to recommend this paper for acceptance. However, it has many small weaknesses. To do a better justice to the goodness of your work, if you have time to address all (or most) of my concerns and questions, within the main paper, I should be able to change my review.", "weaknesses": "Here, I point out major points that need revision. In addition, in the next **Questions** section, I ask for clarifications on more minor (but still important) issues.\n\n\n## The specialization of GFLowNet onto trees be explicit stated\n\nAbove Eq. 6, it says that $P_B(s_{l} | s_{l+1}) = 1$ -- this implies that the only way to get to $s_{l+1}$ is through $s_{l}$. This strong modeling assumption stems from \"s\" being the entire path of generated adjacencies $\\{A_0, \\dots, A_l \\}$. This produces a special case of \"finite state machines\" that specifically have states looking like trees.\n\n## Larger graphs?\n\nOnly one graph more than 1 million nodes whereas one central theme of the paper is about scale.\n\n\n## Application appeal\n\nI would wish that the paper considers applying their method to a problem space beyond graph sampling (or otherwise, show compelling use-cases for graph sampling). Specifically, could this method be used to *explain graphs*? E.g., in the integrated-gradient sense: the presence of which nodes or edges would cause a certain prediction.\n\n## Missing section on inference\n\nWhile the paper includes the information about training, it should also include information on how to do inference. Given a node at a (large) input test graph, are samples taken or the full graph around $n$?\n\n## Missing References on learnable sampling\n\nE.g.,\n\n* DSKReG; CIKM'2021\n* \"Performance-adaptive sampling strategy towards fast and accurate GNNs.\", KDD'2021\n* Submix; UAI'2023", "questions": "The following items are not clear. Please clarify them in the paper\n\nQ1:\nHow is the $GNN_F$ parameterized? Does it train a scalar for every node (i.e., lookup 1D embedding table) or is it a function of features? In my understanding, GNN_F models $P_F$ (correct me if I am wrong) i.e. should have support on the nodes\n\nQ2:\nIs the reward measured only on end states? (on the \"sum\" of sampled list of adjacency matrices) or on every intermediate state (e.g., sum of adjacencies at that point).\n\nQ3: **Runtime experiments** Would you report runtimes? E.g., on the largest dataset ogbn-products?\n\nQ4: **Repeat edges**. Can an edge be sampled twice? Does this have any impact on the GCN model?\n\nQ5: Is adjacency matrix $A^0$ same as $A_0$ (Algorithm 1)\n\nQ6: $Z(s_0)$ in text following Eq. 3 -- It is not clear whether scalar $log Z$ is modeled or if it is a constant and removed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698800067598}], "openreview_url": "https://openreview.net/forum?id=tZ3JmSDbJM", "arxiv_id": "2310.03399", "paper_pdf": "papers/tZ3JmSDbJM.pdf", "paper_pdf_sha256": "0981b4c3da53e2fdb882056523e1c2da3f4ff279c577b2613e9c2194f674ac1e", "paper_pdf_bytes": 754802, "paper_pdf_source": "openreview", "code_url": "https://github.com/dfdazac/grapes", "code_repository": "dfdazac/grapes", "code_commit": "71ecebeaac896800aa4dd1d0f38c57ec222ef396", "code_archive": "repos/tZ3JmSDbJM.zip", "code_archive_sha256": "6f9294fe075e65e3ecf03d20216a47ae4cd820870912a80385c5cf730b9f5ca3", "code_archive_bytes": 79344, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 242, "github_languages": {"Python": 90863}, "github_archived": false, "github_pushed_at": "2025-04-01T12:36:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/grapes-learning-to-sample-graphs-for-scalable"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e9rdb24Yzqx", "year": 2023, "status": "rejected", "title": "Empirical analysis of representation learning and exploration in neural kernel bandits", "authors": ["Michal Lisicki", "Arash Afkanpour", "Graham W. Taylor"], "authorids": ["~Michal_Lisicki1", "~Arash_Afkanpour1", "~Graham_W._Taylor1"], "authors_source": "OpenReview API", "abstract": "Neural bandits have been shown to provide an efficient solution to practical sequential decision tasks that have nonlinear reward functions. The main contributor to that success is approximate Bayesian inference, which enables neural network (NN) training with uncertainty estimates. However, Bayesian NNs often suffer from a prohibitive computational overhead or operate on a subset of parameters. Alternatively, certain classes of infinite neural networks were shown to directly correspond to Gausian processes (GP) with neural kernels (NK). NK-GPs provide accurate uncertainty estimates and can be trained faster than most Bayesian NNs. We propose to guide common bandit policies with NK distributions and show that NK bandits achieve state-of-the-art performance on nonlinear structured data. Moreover, we propose a framework for measuring independently the ability of a bandit algorithm to learn representations and explore, and use it to analyze the impact of NK distributions w.r.t. those two aspects. We consider policies based on a GP and a Student's t-process (TP). Furthermore, we study practical considerations, such as training frequency and model partitioning. We believe our work will help better understand the impact of utilizing NKs in applied settings.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "f6N-90YSZl", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1146/Reviewer_st6H"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Established in deep learning literature that Neural Network (NN) is related to Neural Kernels (NK), especially between the optimization dynamic of infinitely wide NN has shown to be mostly captured by the NTK at initialization. Motivated by the correspondence, this paper proposed to guide bandit policy with NK-GP (or NK-TP to take account for varying noise level in feedback), which gives both reward predictions and uncertainty estimations at the same time. Their main algorithm and its variants are referred to as NK bandits.  \n\nEmpirically, this work shows NK bandits outperform other baselines on most of the tested UCI datasets.\n\nAs another claimed contribution, this paper designed a framework to empirically measure one bandit algorithm’s ability to learn representation and that to do exploration separately. \n\n", "review_text": "This work makes contribution by empirically comparing different bandit policies on nonlinear data and proposing a novel way to indepedently measure representation learn and exploration as well. But it is limited by not providing substantially improved method from existing work for the purpose of solving nonlinear bandits and not much insights on NK bandits that generally apply to applications were drawn from the experimental results.\n", "strengths": "**Strengths**:\n1. It gives a comprehensive side by side comparison between different bandit policies on nonlinear structure data: methods include NK bandits (multiple variants),  neural-linear (LiM2) bandits, neural bandits, linear TS / UCB and multitask GP.\n2. Proposes a way to measure one algorithm’s ability to learn representation and the ability to do exploration, by regulating the bandit environment with the tuple $(\\epsilon, \\delta)$, whose two components are controlling the learning complexity and the exploration urgency respectively.\n \n**Weaknesses**:\n1. Algorithmic novelty: it feels to me that the NK bandit policies proposed in this work are not substantially different from existing NK bandits, especially the $NeuralUCB_0$ by [Zhou et al.](http://arxiv.org/abs/1911.04462) in Appendix E, except for using separate kernels for different actions. There is some discussion in section 3.1 addressing this difference, but not sure if the separate kernel is a substantially improved design for general non-linear bandits or only fits for classification-converted bandits.\n2. Results interpretation: While the paper gives a thorough description of experiment results, the interpretation is less inconclusive. Also, the wheel dataset designed to separate representation learning and exploration is somehow simple in structure due to the low-dimensionality in context ( which I assume to be $2$ if directly following the setup in [Riquelme et al.](http://arxiv.org/abs/1802.09127)). So it’s questionable whether the results give insights of NK bandits that generally apply to applications.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "Established in deep learning literature that Neural Network (NN) is related to Neural Kernels (NK), especially between the optimization dynamic of infinitely wide NN has shown to be mostly captured by the NTK at initialization. Motivated by the correspondence, this paper proposed to guide bandit policy with NK-GP (or NK-TP to take account for varying noise level in feedback), which gives both reward predictions and uncertainty estimations at the same time. Their main algorithm and its variants are referred to as NK bandits.  \n\nEmpirically, this work shows NK bandits outperform other baselines on most of the tested UCI datasets.\n\nAs another claimed contribution, this paper designed a framework to empirically measure one bandit algorithm’s ability to learn representation and that to do exploration separately. \n\n", "strength_and_weaknesses": "**Strengths**:\n1. It gives a comprehensive side by side comparison between different bandit policies on nonlinear structure data: methods include NK bandits (multiple variants),  neural-linear (LiM2) bandits, neural bandits, linear TS / UCB and multitask GP.\n2. Proposes a way to measure one algorithm’s ability to learn representation and the ability to do exploration, by regulating the bandit environment with the tuple $(\\epsilon, \\delta)$, whose two components are controlling the learning complexity and the exploration urgency respectively.\n \n**Weaknesses**:\n1. Algorithmic novelty: it feels to me that the NK bandit policies proposed in this work are not substantially different from existing NK bandits, especially the $NeuralUCB_0$ by [Zhou et al.](http://arxiv.org/abs/1911.04462) in Appendix E, except for using separate kernels for different actions. There is some discussion in section 3.1 addressing this difference, but not sure if the separate kernel is a substantially improved design for general non-linear bandits or only fits for classification-converted bandits.\n2. Results interpretation: While the paper gives a thorough description of experiment results, the interpretation is less inconclusive. Also, the wheel dataset designed to separate representation learning and exploration is somehow simple in structure due to the low-dimensionality in context ( which I assume to be $2$ if directly following the setup in [Riquelme et al.](http://arxiv.org/abs/1802.09127)). So it’s questionable whether the results give insights of NK bandits that generally apply to applications.\n", "clarity,_quality,_novelty_and_reproducibility": "Evaluation/comments on novelty is referred to **strength/weakness**.\n\n**Clarity, Quality**: Though the rest of this paper is written clearly, the sections related to the experiment design (3 and 4) are somewhat difficult to follow for ones who are not familiar with the couple of papers cited in line: better to pull up some technical/mathematical details from the appendix.\n\n**Reproducibility**: Currently the code link is hidden, but I’d like to trust the authors that their results are reproducible.,\n", "summary_of_the_review": "This work makes contribution by empirically comparing different bandit policies on nonlinear data and proposing a novel way to indepedently measure representation learn and exploration as well. But it is limited by not providing substantially improved method from existing work for the purpose of solving nonlinear bandits and not much insights on NK bandits that generally apply to applications were drawn from the experimental results.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666892251214}, {"id": "z-hTmAOyw7", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1146/Reviewer_tHKv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper the authors propose an empirical analysis of neural kernel bandits, proposed as a UCB based alternative to Bayesian neural networks. Bayesian NNs have a high computational requirement as often an ensemble must be maintained in order to bootstrap a sample from the posterior, as in the case of Thompson sampling. The paper studies a version of kernel UCB where the kernel is realized by a linearized neural network in the limit where the width of the network goes to $\\infty$. The experiments compare against other neural network based bandit algorithms and show that there is a benefit for using NK bandits in terms of generalization ability and computational requirements. th authors also propose a practical setting where the ability of the algorithm to learn representations is studied and show a benefit for using NK bandits here.", "review_text": "- I believe that the studied algorithm is not super significant and misses the point of neural network function approximation.\n- I believe that there are scaling issues with the algorithm that cannot be easily resolved in spite of some proposed approaches. I think it requires a larger scale empirical study to assess whether these algorithms actually are statistically and computationally efficient.\n\nGiven these two facts, I propose a reject/weak-reject rating for the paper.", "strengths": "The paper consider NK bandits, or as studied in [1], termed as NeuralUCB_0 type algorithms.\nIn the highly cited paper [1], it is shown that when the neural network representation is allowed to change while training, and the changing representation is used to compute uncertainty estimates, the algorithm outperforms the fixed kernel representation based algorithms. This is in contrast to the results presented in this paper, which show the opposite phenomenon.\n\nAs the scale of the problem grows, the authors themselves remark in the paper that the complexity of computing the neural kernel is significant. However, the proposed solutions in the paper to remedy this issue are not satisfactory in my opinion.\n\nI also generally find it uncomfortable that the authors study an algorithm which does not use the ability of a neural network to learn a good representation for a problem, and rather use a fixed representation that an infinitely wide network would realize. While there are theorems showing equivalence in a highly overparameterized limit, this is not the typical limit in which practical multi-layer neural network architectures operate. As such the studied algorithm is inherently linear, and so I do not expect it to outperform nonlinear bandit algorithms as the problem scale grows.\n\n\n[1]: https://arxiv.org/pdf/1911.04462.pdf", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper the authors propose an empirical analysis of neural kernel bandits, proposed as a UCB based alternative to Bayesian neural networks. Bayesian NNs have a high computational requirement as often an ensemble must be maintained in order to bootstrap a sample from the posterior, as in the case of Thompson sampling. The paper studies a version of kernel UCB where the kernel is realized by a linearized neural network in the limit where the width of the network goes to $\\infty$. The experiments compare against other neural network based bandit algorithms and show that there is a benefit for using NK bandits in terms of generalization ability and computational requirements. th authors also propose a practical setting where the ability of the algorithm to learn representations is studied and show a benefit for using NK bandits here.", "strength_and_weaknesses": "The paper consider NK bandits, or as studied in [1], termed as NeuralUCB_0 type algorithms.\nIn the highly cited paper [1], it is shown that when the neural network representation is allowed to change while training, and the changing representation is used to compute uncertainty estimates, the algorithm outperforms the fixed kernel representation based algorithms. This is in contrast to the results presented in this paper, which show the opposite phenomenon.\n\nAs the scale of the problem grows, the authors themselves remark in the paper that the complexity of computing the neural kernel is significant. However, the proposed solutions in the paper to remedy this issue are not satisfactory in my opinion.\n\nI also generally find it uncomfortable that the authors study an algorithm which does not use the ability of a neural network to learn a good representation for a problem, and rather use a fixed representation that an infinitely wide network would realize. While there are theorems showing equivalence in a highly overparameterized limit, this is not the typical limit in which practical multi-layer neural network architectures operate. As such the studied algorithm is inherently linear, and so I do not expect it to outperform nonlinear bandit algorithms as the problem scale grows.\n\n\n[1]: https://arxiv.org/pdf/1911.04462.pdf", "clarity,_quality,_novelty_and_reproducibility": "I believe the clarity of the paper can be improved, by formally introducing and defining the neural kernel bandit algorithm more clearly and earlier in the paper. Otherwise it is largely free of typos, and reads OK.\n\nIn my opinion, the results are not significantly novel, although there is some value in carrying out an empirical study of NK bandits. However, I am not convinced that these results extend to larger scale datasets and problem settings, and I worry about the scaling issues in the considered algorithm. Both of these are not adequately addressed in the current paper, and I believe are the hard problems to study in practice.\n\nI have not checked the reproducibility of the experiments.", "summary_of_the_review": "- I believe that the studied algorithm is not super significant and misses the point of neural network function approximation.\n- I believe that there are scaling issues with the algorithm that cannot be easily resolved in spite of some proposed approaches. I think it requires a larger scale empirical study to assess whether these algorithms actually are statistically and computationally efficient.\n\nGiven these two facts, I propose a reject/weak-reject rating for the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666716540387}, {"id": "jwbuniJ85Dc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1146/Reviewer_r846"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper performs an empirical study of neural kernel bandit algorithms, to investigate the efficacy of these algorithms from different aspects such as representation learning and exploration ability. To achieve this, the paper has proposed a novel benchmark for contextual bandits which allows separately evaluating the efficacy of an algorithm in terms of representation learning and exploration.", "review_text": "I think the paper provides some important empirical insights regarding neural kernel bandits. I have an important concern which is the first one listed under \"Weaknesses\" above, which is regarding some discrepancies with the observations from recent works on neural bandits which explicitly use NNs for reward prediction. If this concern is addressed well, I'll be happy to increase my evaluation.", "strengths": "Strengths:\n- The empirical insights drawn from the empirical evaluations of this paper are interesting and can be of interest for the neural bandit community and the bandit community in general.\n- The proposed benchmark with the additional ability to evaluate the ability of an algorithm to learn representation is also interesting and can be potentially useful for the broader community.\n\nWeaknesses:\n- My biggest concern is regarding the comparison with the recent works on neural contextual bandits which explicitly train a neural network to predict the reward, including NeuralUCB and NeuralTS. **Firstly**, Figure 2 shows that NeuralUCB and NeuralTS are consistently outperformed by LinearUCB and LinearTS. This is surprising to me, because in the previous papers on NeuralUCB (Zhou et al., 2020) and NeuralTS (Zhang et al., 2020), they have both also performed experiments using the UCI datasets and had shown that both NeuralUCB and NeuralTS consistently outperform linear bandit algorithms. What's the reason for this discrepancy between your observations and theirs? **Secondly**, it is mentioned in the Introduction that \"...NKs have been shown to lack the full representational power of the corresponding NNs...\". The same statement has also been made by a few other papers such as NeuralUCB. As a result of this, the use of a neural network to predict the reward (i.e., using an NN as the term $\\mu_{a,t}$ on line 8 of Algorithm 1) is important for achieving good performances in neural bandits. Therefore, it is surprising that NeuralUCB and NeuralTS, which indeed use NNs for reward prediction, are significantly outperformed by the proposed algorithm which does not use NNs for reward prediction. Furthermore, in fact, the NeuralUCB paper also proposed another algorithm NeuralUCB0 which does not use NNs for reward prediction but instead simply treats NTK as a kernel in kernelized bandits. In this sense, NeuralUCB0 is similar to the proposed algorithm when NTKGP is used. But in their paper, they have shown that this NeuralUCB0 is consistently outperformed by NeuralUCB, which they have attributed to the fact that NeuralUCB0 does not use NNs for reward prediction.       \nThese two concerns here have broader implications, regarding whether it's useful/necessary to explicitly use an NN to predict the reward.\n- The technical contribution of the paper may be limited, since the proposed algorithm is a straightforward combination of existing methods.\n- Section 3.2: If I understand correctly, for the disjoint model, when estimating the posterior distribution for the reward of an arm, you only use the previous observations collected for this arm? So the observations from all other arms are not used? Isn't this a waste of data? Please clarify whether I misunderstood.\n- It's unclear to me which step of Algorithm 1 requires training a neural network?\n- (minor) Section 2, first paragraph, third line: \"inite-width\" should be \"infinite-width\".\n- (minor) Section 2.1, second last paragraph, first line: UCB is in fact not a \"stochastic\" policy.\n\nOther comments:\n- A concurrent work (paper [a] below) has also empirically evaluated neural kernel bandit algorithms in other real-world problems of autoML and reinforcement learning, and hence should be referenced.       \n[a] Sample-Then-Optimize batch neural Thompson sampling, NeurIPS 2022.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper performs an empirical study of neural kernel bandit algorithms, to investigate the efficacy of these algorithms from different aspects such as representation learning and exploration ability. To achieve this, the paper has proposed a novel benchmark for contextual bandits which allows separately evaluating the efficacy of an algorithm in terms of representation learning and exploration.", "strength_and_weaknesses": "Strengths:\n- The empirical insights drawn from the empirical evaluations of this paper are interesting and can be of interest for the neural bandit community and the bandit community in general.\n- The proposed benchmark with the additional ability to evaluate the ability of an algorithm to learn representation is also interesting and can be potentially useful for the broader community.\n\nWeaknesses:\n- My biggest concern is regarding the comparison with the recent works on neural contextual bandits which explicitly train a neural network to predict the reward, including NeuralUCB and NeuralTS. **Firstly**, Figure 2 shows that NeuralUCB and NeuralTS are consistently outperformed by LinearUCB and LinearTS. This is surprising to me, because in the previous papers on NeuralUCB (Zhou et al., 2020) and NeuralTS (Zhang et al., 2020), they have both also performed experiments using the UCI datasets and had shown that both NeuralUCB and NeuralTS consistently outperform linear bandit algorithms. What's the reason for this discrepancy between your observations and theirs? **Secondly**, it is mentioned in the Introduction that \"...NKs have been shown to lack the full representational power of the corresponding NNs...\". The same statement has also been made by a few other papers such as NeuralUCB. As a result of this, the use of a neural network to predict the reward (i.e., using an NN as the term $\\mu_{a,t}$ on line 8 of Algorithm 1) is important for achieving good performances in neural bandits. Therefore, it is surprising that NeuralUCB and NeuralTS, which indeed use NNs for reward prediction, are significantly outperformed by the proposed algorithm which does not use NNs for reward prediction. Furthermore, in fact, the NeuralUCB paper also proposed another algorithm NeuralUCB0 which does not use NNs for reward prediction but instead simply treats NTK as a kernel in kernelized bandits. In this sense, NeuralUCB0 is similar to the proposed algorithm when NTKGP is used. But in their paper, they have shown that this NeuralUCB0 is consistently outperformed by NeuralUCB, which they have attributed to the fact that NeuralUCB0 does not use NNs for reward prediction.       \nThese two concerns here have broader implications, regarding whether it's useful/necessary to explicitly use an NN to predict the reward.\n- The technical contribution of the paper may be limited, since the proposed algorithm is a straightforward combination of existing methods.\n- Section 3.2: If I understand correctly, for the disjoint model, when estimating the posterior distribution for the reward of an arm, you only use the previous observations collected for this arm? So the observations from all other arms are not used? Isn't this a waste of data? Please clarify whether I misunderstood.\n- It's unclear to me which step of Algorithm 1 requires training a neural network?\n- (minor) Section 2, first paragraph, third line: \"inite-width\" should be \"infinite-width\".\n- (minor) Section 2.1, second last paragraph, first line: UCB is in fact not a \"stochastic\" policy.\n\nOther comments:\n- A concurrent work (paper [a] below) has also empirically evaluated neural kernel bandit algorithms in other real-world problems of autoML and reinforcement learning, and hence should be referenced.       \n[a] Sample-Then-Optimize batch neural Thompson sampling, NeurIPS 2022.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is well written.\n\nQuality: The empirical evaluations are comprehensive and hence of high quality.\n\nNovelty: The technical novelty may be limited, since the proposed algorithm is a straightforward combination of existing methods.\n\nReproducibility: The code is submitted for reproducibility.", "summary_of_the_review": "I think the paper provides some important empirical insights regarding neural kernel bandits. I have an important concern which is the first one listed under \"Weaknesses\" above, which is regarding some discrepancies with the observations from recent works on neural bandits which explicitly use NNs for reward prediction. If this concern is addressed well, I'll be happy to increase my evaluation.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666620314147}, {"id": "JRxbMZiwMn", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1146/Reviewer_uxKz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents am empirical comparison between the performance of various neural kernels used in GPs for contextual bandit optimisation.  Comparisons are done using a modified wheel dataset, and the use of the student t-process to improve exploration in NK bandits is explored.", "review_text": "The paper is well written, but as this boils down to a comparison of kernels in GPs I feel that there is insufficient novelty here to warrant acceptance.", "strengths": "The background is interesting and the various NN kernel do have some potential for use in GP models for sequential model-based optimisation and bandits.\n\nMy main difficulty with this paper however is that, in the end, it appears to come down to a simple experimental comparison of various kernels, some of which happen to derive from various models or analysis of neural networks.  I am not convinced that this suffices to have a significant impact.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper presents am empirical comparison between the performance of various neural kernels used in GPs for contextual bandit optimisation.  Comparisons are done using a modified wheel dataset, and the use of the student t-process to improve exploration in NK bandits is explored.", "strength_and_weaknesses": "The background is interesting and the various NN kernel do have some potential for use in GP models for sequential model-based optimisation and bandits.\n\nMy main difficulty with this paper however is that, in the end, it appears to come down to a simple experimental comparison of various kernels, some of which happen to derive from various models or analysis of neural networks.  I am not convinced that this suffices to have a significant impact.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and the experiments appear to be reproducible.", "summary_of_the_review": "The paper is well written, but as this boils down to a comparison of kernels in GPs I feel that there is insufficient novelty here to warrant acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666581496494}], "openreview_url": "https://openreview.net/forum?id=e9rdb24Yzqx", "arxiv_id": "2111.03543", "paper_pdf": "papers/e9rdb24Yzqx.pdf", "paper_pdf_sha256": "6c16f7be104c6f8c6d00db1ab9dd01ef7785246e682be95232bfc67678ec7d8b", "paper_pdf_bytes": 614603, "paper_pdf_source": "openreview", "code_url": "https://github.com/VectorInstitute/NeuralKernelBandits", "code_repository": "VectorInstitute/NeuralKernelBandits", "code_commit": "6ddbfa3a410b50b84d9fb72c75803347da61b4cb", "code_archive": "repos/e9rdb24Yzqx.zip", "code_archive_sha256": "3c0bece48111776c6dc7eced07121c279a0d6eec96bdce20e17d3d9a23123d25", "code_archive_bytes": 409604, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 513, "github_languages": {}, "github_archived": false, "github_pushed_at": "2021-12-09T15:15:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/an-empirical-study-of-neural-kernel-bandits"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TD-5kgf13mH", "year": 2022, "status": "rejected", "title": "Sparse MoEs meet Efficient Ensembles", "authors": ["James Urquhart Allingham", "Florian Wenzel", "Zelda E Mariet", "Basil Mustafa", "Joan Puigcerver", "Neil Houlsby", "Ghassen Jerfel", "Vincent Fortuin", "Balaji Lakshminarayanan", "Jasper Snoek", "Dustin Tran", "Carlos Riquelme Ruiz", "Rodolphe Jenatton"], "authorids": ["~James_Urquhart_Allingham1", "~Florian_Wenzel1", "~Zelda_E_Mariet1", "~Basil_Mustafa1", "~Joan_Puigcerver1", "~Neil_Houlsby1", "~Ghassen_Jerfel1", "~Vincent_Fortuin1", "~Balaji_Lakshminarayanan1", "~Jasper_Snoek1", "~Dustin_Tran1", "~Carlos_Riquelme_Ruiz1", "~Rodolphe_Jenatton3"], "authors_source": "OpenReview API", "abstract": "Machine learning models based on the aggregated outputs of submodels, either at the activation or prediction levels, lead to strong performance. We study the interplay of two popular classes of such models: ensembles of neural networks and sparse mixture of experts (sparse MoEs). First, we show that these two approaches have complementary features whose combination is beneficial. Then, we present partitioned batch ensembles, an efficient ensemble of sparse MoEs that takes the best of both classes of models. Extensive experiments on fine-tuned vision transformers demonstrate the accuracy, log-likelihood, few-shot learning, robustness, and uncertainty calibration improvements of our approach over several challenging baselines. Partitioned batch ensembles not only scale to models with up to 2.7B parameters, but also provide larger performance gains for larger models. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "bfuB_Cnnwm6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper633/Reviewer_KpkQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors show empirically that Sparse MOEs and Ensembles have complementary features, and suggest that combining the two should lead to improved performance. Authors build on the Vision Transformer (ViT) for their experiments. To efficiently combine Sparse MOEs and Ensembles, the paper presents Partitioned Batch Ensembles (PBE), where the parameters of the self-attention layers are shared, and an ensemble of Sparse MOEs are used for the MLP layers of the Transformer blocks.\n\n", "review_text": "1. PBEs share the attention layers and use an ensemble of sparse MOEs for the MLP layers of Vision Transformers. Any reasons for this design choice, and not going the other way, i.e. sharing the MLPs and using Sparse MOEs for the self-attention layers ?\n2. Results in Table 2 suggest prediction level ensembling can be beneficial for uncertainty calibration. The Table reports the results for K = 2 and K = 4. Does this claim hold true for K > 4 as well ?\n3. The multi-head MoE presented in Section 3.2 stacks the K selected expert predictions. Are these K predictions averaged in order to compute the NLL and Error scores reported in Table 2 ?\n4. Fig 2 a: For every value of M, the log likelihood score increases till K=5, and drops at K=6. Any reasons behind this ?\n5. The idea of Batch Ensembling and Tiling isn't new, and has been used in previous works [A].\n\n[A] Yeming Wen, Dustin Tran, and Jimmy Ba. Batch ensemble: an alternative approach to efficient\nensemble and lifelong learning. In ICLR, 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors show empirically that Sparse MOEs and Ensembles have complementary features, and suggest that combining the two should lead to improved performance. Authors build on the Vision Transformer (ViT) for their experiments. To efficiently combine Sparse MOEs and Ensembles, the paper presents Partitioned Batch Ensembles (PBE), where the parameters of the self-attention layers are shared, and an ensemble of Sparse MOEs are used for the MLP layers of the Transformer blocks.\n\n", "main_review": "1. PBEs share the attention layers and use an ensemble of sparse MOEs for the MLP layers of Vision Transformers. Any reasons for this design choice, and not going the other way, i.e. sharing the MLPs and using Sparse MOEs for the self-attention layers ?\n2. Results in Table 2 suggest prediction level ensembling can be beneficial for uncertainty calibration. The Table reports the results for K = 2 and K = 4. Does this claim hold true for K > 4 as well ?\n3. The multi-head MoE presented in Section 3.2 stacks the K selected expert predictions. Are these K predictions averaged in order to compute the NLL and Error scores reported in Table 2 ?\n4. Fig 2 a: For every value of M, the log likelihood score increases till K=5, and drops at K=6. Any reasons behind this ?\n5. The idea of Batch Ensembling and Tiling isn't new, and has been used in previous works [A].\n\n[A] Yeming Wen, Dustin Tran, and Jimmy Ba. Batch ensemble: an alternative approach to efficient\nensemble and lifelong learning. In ICLR, 2019.\n", "summary_of_the_review": "I only have minor concerns with the paper, and I have stated those concerns in the Main Review. Other than that, the paper is novel and the experiments demonstrate the usefulness of the proposed Partitioned Batch Ensembles (PBE).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635965664345}, {"id": "yLMXq6Qoeyz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper633/Reviewer_cawf"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper investigates the benefits of combining (Sparse) Mixture of Experts (MoE) and ensembling. Sparse MoE’s employ conditional computation to reduce computational and environmental costs of DNNs while maintaining (or increasing) performance. On the other hand, ensembling models has been shown to achieve the highest robustness in the presence of dataset shift, e.g., higher accuracy and better estimation of uncertainty.\n\nThe present submission empirically demonstrates that Sparse MoE’s can be ensembled to attain the benefits of both techniques simultaneously, i.e., conditional computation (with its implications to scalability) together with robustness in the presence of dataset shift. The main components of the approach are: 1) Disjoint MoE’s as ensemble members. 2) Tiling of representations which enables all ensemble members to compute the output for a batch in a single forward pass.\n\nA number of experiments compare predictive performance vs computational cost for the proposed approach (pBE), vision Sparse MoE’s (V-MoE) and Vision Transformers (ViT). The results show that, under some conditions, pBE displays the known benefits of Sparse MoE’s and ensembling, which leads to performance gains w.r.t. ViT and V-MoE under metrics like accuracy, negative log-likelihood (NLL) and expected calibration error (ECE).", "review_text": "Strengths:\n\nThe paper provides confirmation that MoEs and ensembles have complementary (or additive) benefits. Moreover, the proposed approach (partitioned batch ensembles), which combines MoEs and ensembling, achieves the benefits of both techniques simultaneously. \n\nThe paper provides the first study of the robustness of vision sparse MoE’s (parallel submission).\n\nWeaknesses:\n\nThe proposed approach is incremental. In the words of the authors, the method “introduces two changes to V-MoEs: (a) the partitioning of the experts and (b) the tiling of the representations.”\n\nThe paper focuses on the comparison of specific architectures and training procedures. In this regard, conclusions one may be able to draw from the paper are too constrained. For example, can the benefit of “static” ensembles be achieved by “adaptive” ensembling? While this paper’s results suggest perhaps not, I don’t think this is conclusive. [The distinction between static and adaptive may be generalized in such a way as to enable finer control on aspect and degree of adaptation with implications to, e.g., the diversity and uncertainty calibration at the output of the ensemble.] Further, e.g., (Wen et al. 2020) claim that “diversity of initialization entirely determines the diversity of ensembling system.” The present paper notes ways to achieve diversity given a particular distinction between static vs adaptive ensembling, and training procedure. There are other dimensions (e.g., initialization, example balancing) which have direct relevance to diversity and thus, uncertainty calibration, which are not explored in this paper.\n\nSection 3.1 is an ablation study just like section 4.2 is. I find the discussion going back and forth to be detrimental to clarity and efficiency of communication.\n\nSection 3.2 compares a naive multi-head method -- without specifying how it is trained -- with a “standard” vision (sparse) MoE. I don’t find the results there clear enough to be compelling. For example, whether the multi-head variant is able to provide diverse predictions is more likely to do with the training procedure than with the architecture -- this is related, e.g., to the “load balancing” during training used in (Riquelme et al. 2021). Besides, the claims in section 3.2 are conflicting: we are told naive multi-head improves ECE but also that a different strategy is required for uncertainty calibration.\n\nThe experimental results (section 5) are simply stated. For example, most of the time (static) ensembling helps but when it doesn’t (e.g., out of distribution performance) the paper provides no discussion or insight.\n\nSome acronyms are used without introduction (the introduction appears after the first usage), e.g., OOD and NLL in page 4.\n\nIt’s not clear to me how the static ensembling portion of the proposed approach is “efficient” (section 4.1). For example, Batch Ensemble (Wen et al. 2020) generates the family of ensemble weights via element-wise multiplication of a shared matrix and “fast-weights.” In my understanding, in the case of the proposed approach any sharing of parameters is done within the ensemble members.\n\nHow are the dashed lines in figures 4 and 5 pareto frontiers? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper investigates the benefits of combining (Sparse) Mixture of Experts (MoE) and ensembling. Sparse MoE’s employ conditional computation to reduce computational and environmental costs of DNNs while maintaining (or increasing) performance. On the other hand, ensembling models has been shown to achieve the highest robustness in the presence of dataset shift, e.g., higher accuracy and better estimation of uncertainty.\n\nThe present submission empirically demonstrates that Sparse MoE’s can be ensembled to attain the benefits of both techniques simultaneously, i.e., conditional computation (with its implications to scalability) together with robustness in the presence of dataset shift. The main components of the approach are: 1) Disjoint MoE’s as ensemble members. 2) Tiling of representations which enables all ensemble members to compute the output for a batch in a single forward pass.\n\nA number of experiments compare predictive performance vs computational cost for the proposed approach (pBE), vision Sparse MoE’s (V-MoE) and Vision Transformers (ViT). The results show that, under some conditions, pBE displays the known benefits of Sparse MoE’s and ensembling, which leads to performance gains w.r.t. ViT and V-MoE under metrics like accuracy, negative log-likelihood (NLL) and expected calibration error (ECE).", "main_review": "Strengths:\n\nThe paper provides confirmation that MoEs and ensembles have complementary (or additive) benefits. Moreover, the proposed approach (partitioned batch ensembles), which combines MoEs and ensembling, achieves the benefits of both techniques simultaneously. \n\nThe paper provides the first study of the robustness of vision sparse MoE’s (parallel submission).\n\nWeaknesses:\n\nThe proposed approach is incremental. In the words of the authors, the method “introduces two changes to V-MoEs: (a) the partitioning of the experts and (b) the tiling of the representations.”\n\nThe paper focuses on the comparison of specific architectures and training procedures. In this regard, conclusions one may be able to draw from the paper are too constrained. For example, can the benefit of “static” ensembles be achieved by “adaptive” ensembling? While this paper’s results suggest perhaps not, I don’t think this is conclusive. [The distinction between static and adaptive may be generalized in such a way as to enable finer control on aspect and degree of adaptation with implications to, e.g., the diversity and uncertainty calibration at the output of the ensemble.] Further, e.g., (Wen et al. 2020) claim that “diversity of initialization entirely determines the diversity of ensembling system.” The present paper notes ways to achieve diversity given a particular distinction between static vs adaptive ensembling, and training procedure. There are other dimensions (e.g., initialization, example balancing) which have direct relevance to diversity and thus, uncertainty calibration, which are not explored in this paper.\n\nSection 3.1 is an ablation study just like section 4.2 is. I find the discussion going back and forth to be detrimental to clarity and efficiency of communication.\n\nSection 3.2 compares a naive multi-head method -- without specifying how it is trained -- with a “standard” vision (sparse) MoE. I don’t find the results there clear enough to be compelling. For example, whether the multi-head variant is able to provide diverse predictions is more likely to do with the training procedure than with the architecture -- this is related, e.g., to the “load balancing” during training used in (Riquelme et al. 2021). Besides, the claims in section 3.2 are conflicting: we are told naive multi-head improves ECE but also that a different strategy is required for uncertainty calibration.\n\nThe experimental results (section 5) are simply stated. For example, most of the time (static) ensembling helps but when it doesn’t (e.g., out of distribution performance) the paper provides no discussion or insight.\n\nSome acronyms are used without introduction (the introduction appears after the first usage), e.g., OOD and NLL in page 4.\n\nIt’s not clear to me how the static ensembling portion of the proposed approach is “efficient” (section 4.1). For example, Batch Ensemble (Wen et al. 2020) generates the family of ensemble weights via element-wise multiplication of a shared matrix and “fast-weights.” In my understanding, in the case of the proposed approach any sharing of parameters is done within the ensemble members.\n\nHow are the dashed lines in figures 4 and 5 pareto frontiers? ", "summary_of_the_review": "Overall I lean towards rejection. While the paper presents experimental results likely not found elsewhere, in my perception, that appears to be the sole reason for the paper. Specifically, the paper presents experimental results on the robustness of V-MoE’s (vision sparse MoE’s) and adds ensembling on top of the V-MoE model to increase robustness. However, no further insight or development (e.g., in architecture or optimization) is provided.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635885945538}, {"id": "ENNBBFEvMoC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper633/Reviewer_5srE"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper considers the combination of MoE and Ensembles for image recognition tasks to obtain both improvement of classification accuracy as well as stability of prediction. The proposed method is evaluated with extensive expereiments from large-scale image recognition to OOD and the results indicate better performance for the proposed method.", "review_text": "The paper is well written and the experiments are comprehensive. The main concerns about this paper are as follows.\n1. The paper lacks novelty. It just combines the MoE with Ensemble in a trivial manner. The tiling technique is quite simple and it is just the classical way of doing ensemble across models.\n2. The improvement is quite limited. In Table 3, I find although the proposed method achieves gains across different metrics e.g. NLL, Error, and ECE, the improvement in terms of accuracy is less than 0.5% which is quite small and cannot justify the effectiveness of the method. Furthermore, when compared to ViT and V-MoE in Figure 4,5,6, I can only find the curve for NLL instead of Error, which I think is more critical. \n3. The number of parameters should be reported and compared to ViT, since MoE is a parameter-consuming method and we cannot ignore this factor.\n4. The FLOPs cost by tiling and partitioning is another concern. More details about how to compute FLOPs for the proposed method and V-MoE should be clarified. For V-MoE, it is possible to first compute the coefficient for each expert and then synthesize the experts before extracting features. Therefore, it can save a large scale of computational costs. But due to tiling and partitioning, each branch should cause its own computation, which I think will cost a lot of extra computation and let me doubt the efficiency of the proposed method. I wonder whether the author computes the FLOPs with the optimal way.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper considers the combination of MoE and Ensembles for image recognition tasks to obtain both improvement of classification accuracy as well as stability of prediction. The proposed method is evaluated with extensive expereiments from large-scale image recognition to OOD and the results indicate better performance for the proposed method.", "main_review": "The paper is well written and the experiments are comprehensive. The main concerns about this paper are as follows.\n1. The paper lacks novelty. It just combines the MoE with Ensemble in a trivial manner. The tiling technique is quite simple and it is just the classical way of doing ensemble across models.\n2. The improvement is quite limited. In Table 3, I find although the proposed method achieves gains across different metrics e.g. NLL, Error, and ECE, the improvement in terms of accuracy is less than 0.5% which is quite small and cannot justify the effectiveness of the method. Furthermore, when compared to ViT and V-MoE in Figure 4,5,6, I can only find the curve for NLL instead of Error, which I think is more critical. \n3. The number of parameters should be reported and compared to ViT, since MoE is a parameter-consuming method and we cannot ignore this factor.\n4. The FLOPs cost by tiling and partitioning is another concern. More details about how to compute FLOPs for the proposed method and V-MoE should be clarified. For V-MoE, it is possible to first compute the coefficient for each expert and then synthesize the experts before extracting features. Therefore, it can save a large scale of computational costs. But due to tiling and partitioning, each branch should cause its own computation, which I think will cost a lot of extra computation and let me doubt the efficiency of the proposed method. I wonder whether the author computes the FLOPs with the optimal way.", "summary_of_the_review": "The paper still has a big space for further improvement. I recommend rejecting it this time.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634620105576}], "openreview_url": "https://openreview.net/forum?id=TD-5kgf13mH", "arxiv_id": "2110.03360", "paper_pdf": "papers/TD-5kgf13mH.pdf", "paper_pdf_sha256": "092a43edd66084b67086ec2e0cc0a2049f8e4c0bfa227e1cdd177bd0f09515d2", "paper_pdf_bytes": 1097965, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/vmoe", "code_repository": "google-research/vmoe", "code_commit": "6bb693bc81f949d654a9c5b1fc062374c92464cb", "code_archive": "repos/TD-5kgf13mH.zip", "code_archive_sha256": "02915bf096dec1d3d93d0033497b5ef0c394aabc5747f2e90eeed58ed70713b0", "code_archive_bytes": 1538795, "code_file_count": 102, "code_extensions": {".py": 98, ".ipynb": 3, ".sh": 1}, "github_disk_usage_kb": 1834, "github_languages": {"Jupyter Notebook": 1764344, "Python": 776887, "Shell": 2077}, "github_archived": false, "github_pushed_at": "2026-09-07T15:11:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sparse-moes-meet-efficient-ensembles"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Sva-fwURywB", "year": 2021, "status": "rejected", "title": "Efficiently Disentangle Causal Representations", "authors": ["Yuanpeng Li", "Joel Hestness", "Mohamed Elhoseiny", "Liang Zhao", "Kenneth Church"], "authorids": ["~Yuanpeng_Li2", "~Joel_Hestness2", "~Mohamed_Elhoseiny1", "~Liang_Zhao2", "~Kenneth_Church1"], "authors_source": "OpenReview API", "abstract": "In this paper, we propose a novel approach to efficiently learning disentangled representations with causal mechanisms, based on the difference of conditional probabilities in original and new distributions. We approximate the difference with model's generalization abilities so that it fits in standard machine learning framework and can be efficiently computed. In contrast to the state-of-the-art approach, which relies on learner's adaptation speed to new distribution, the proposed approach only requires evaluating the generalization ability of the model. We provide theoretical explanation for the advantage of the proposed method, and our experiments show that the proposed technique is 1.9-11.0x more sample efficient and 9.4-32.4x quicker than the previous method on various tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "jRXOnX_mi2L", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper265/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a novel approach for determining the causal direction between two random variables $A$ and $B$. The approach is based on the assumption that the conditional distribution $P_{A \\rightarrow B}(B \\mid B)$ does not change between the train and transfer distribution. As a result, a model that predicts the correct causal direction $A \\rightarrow B$ should generalize better from the train to the transfer distribution compared to a model predicting the wrong causal direction $B \\rightarrow A$. While previous work has proposed to use this insight to determine the causal direction by comparing the adaptation speed, this paper proposes to directly measure and compare the generalization performance. The results indicate that the proposed approach leads to the same performance in terms of causal relation prediction, but that it is more sample efficient and faster.\n\n\n\nStrengths:\n* The paper presents an interesting approach for determining the causal direction between two random variables. While the generally accepted wisdom is that a better causal model will lead to better generalization, their approach follows the reverse logic: if a causal model achieves better generalization, it is interpreted as the better (i.e. correct) causal model. The paper builds on this logic proposed by Bengio et al. (2020) and proposes a more efficient measure for the generalization performance of a model.\n* The results show a clear advantage of the proposed approach in terms of sample efficiency and speed.\n* The presented proofs seem correct and are presented in a clear way.\n\n\n\nWeaknesses:\n* The paper builds heavily on the previous approach presented by Bengio et al. (2020). While this is not a problem per se, the paper should be written in such a way that it can stand alone. In the current version, it is not possible to follow the experimental results without consulting the paper by Bengio et al. (2020). It would be very helpful to add a description of the data and models used, as well as the measures compared, e.g. what is $\\sigma(\\gamma)$?\n    * Consulting the paper by Bengio et al. (2020), it seems that the results in Figure 2 can only be interpreted as correct if the model also predicts the correct causal direction. I assume this is the case for the presented models, but it is not stated anywhere.\n* As far as I understand (Algorithm 2, line 3), one model is trained to optimize the loss for both causal directions simultaneously. That approach seems rather counterintuitive to me. I would imagine that this could lead to the model learning a conditional distribution that does not match either causal direction well, effectively rendering the proposed approach unworkable. Wouldn’t it be better to train one model for each causal direction separately and to compare the generalization performance between these models?\n* It would be interesting to see how the proposed approach compares to other methods for causal discovery.\n* The paper could be improved in terms of clarity:\n    * There are a few grammatical errors, especially the articles “the” and “a” are missing quite often.\n    * The structure could be reformatted to be more efficient. At the moment, the introduction spends a lot of time on a very general introduction of causality. I would recommend reducing this and instead, focus more on the proposed approach. Based on the current introduction, I could understand how the previous approach worked, but not the proposed one.\n    * The title should be more concrete. \n\n\n\nOther comments:\n* What happens if $S_G = 0$?\n* There is a mismatch in Figure 1.b): While the caption states that the x-axis shows the “computation time in seconds”, the x-axis in the plot is labeled as “Number of episodes”.\n* The paper could benefit from a clearer statement of all the assumptions that are necessary for the presented approach. For example, I would expect the approach to rely on noiseless dynamics and a fully observed setting without hidden confounders.\n* Regarding noiseless dynamics, the paper presents an experiment in the appendix showing how performance degrades when increasing the standard deviation of additive noise. For this experiment, it would be interesting to know the standard deviation of the underlying data, otherwise, it is unclear how the noise relates to that.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach, structure of paper could be improved", "review": "The paper presents a novel approach for determining the causal direction between two random variables $A$ and $B$. The approach is based on the assumption that the conditional distribution $P_{A \\rightarrow B}(B \\mid B)$ does not change between the train and transfer distribution. As a result, a model that predicts the correct causal direction $A \\rightarrow B$ should generalize better from the train to the transfer distribution compared to a model predicting the wrong causal direction $B \\rightarrow A$. While previous work has proposed to use this insight to determine the causal direction by comparing the adaptation speed, this paper proposes to directly measure and compare the generalization performance. The results indicate that the proposed approach leads to the same performance in terms of causal relation prediction, but that it is more sample efficient and faster.\n\n\n\nStrengths:\n* The paper presents an interesting approach for determining the causal direction between two random variables. While the generally accepted wisdom is that a better causal model will lead to better generalization, their approach follows the reverse logic: if a causal model achieves better generalization, it is interpreted as the better (i.e. correct) causal model. The paper builds on this logic proposed by Bengio et al. (2020) and proposes a more efficient measure for the generalization performance of a model.\n* The results show a clear advantage of the proposed approach in terms of sample efficiency and speed.\n* The presented proofs seem correct and are presented in a clear way.\n\n\n\nWeaknesses:\n* The paper builds heavily on the previous approach presented by Bengio et al. (2020). While this is not a problem per se, the paper should be written in such a way that it can stand alone. In the current version, it is not possible to follow the experimental results without consulting the paper by Bengio et al. (2020). It would be very helpful to add a description of the data and models used, as well as the measures compared, e.g. what is $\\sigma(\\gamma)$?\n    * Consulting the paper by Bengio et al. (2020), it seems that the results in Figure 2 can only be interpreted as correct if the model also predicts the correct causal direction. I assume this is the case for the presented models, but it is not stated anywhere.\n* As far as I understand (Algorithm 2, line 3), one model is trained to optimize the loss for both causal directions simultaneously. That approach seems rather counterintuitive to me. I would imagine that this could lead to the model learning a conditional distribution that does not match either causal direction well, effectively rendering the proposed approach unworkable. Wouldn’t it be better to train one model for each causal direction separately and to compare the generalization performance between these models?\n* It would be interesting to see how the proposed approach compares to other methods for causal discovery.\n* The paper could be improved in terms of clarity:\n    * There are a few grammatical errors, especially the articles “the” and “a” are missing quite often.\n    * The structure could be reformatted to be more efficient. At the moment, the introduction spends a lot of time on a very general introduction of causality. I would recommend reducing this and instead, focus more on the proposed approach. Based on the current introduction, I could understand how the previous approach worked, but not the proposed one.\n    * The title should be more concrete. \n\n\n\nOther comments:\n* What happens if $S_G = 0$?\n* There is a mismatch in Figure 1.b): While the caption states that the x-axis shows the “computation time in seconds”, the x-axis in the plot is labeled as “Number of episodes”.\n* The paper could benefit from a clearer statement of all the assumptions that are necessary for the presented approach. For example, I would expect the approach to rely on noiseless dynamics and a fully observed setting without hidden confounders.\n* Regarding noiseless dynamics, the paper presents an experiment in the appendix showing how performance degrades when increasing the standard deviation of additive noise. For this experiment, it would be interesting to know the standard deviation of the underlying data, otherwise, it is unclear how the noise relates to that.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604495974540}, {"id": "tqRjEngdEI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper265/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an efficient approach to learn disentanglement representation causal mechanism, based on a generalization loss between a training data set and transfer data sets. Empirical studies show it achieves better sample and computation efficiency than a previous work (Bengio et al., 2020). \n\nPros: \n- Improvement over the existing baseline is significant\n- Theoretical statements on the results, particularly on biased and unbiased estimators, offer good understanding for the approach. \n\nCons:\n-  Currently, I find the setting and results on real datasets rather weak. It would be great if authors could demonstrate such a causal direction identification problem in some real setting, with the distribution shift occurrence between different sets of data. \n- There are no comparison with other methods. It would be great if authors could show baseline methods used in (Mooij et al., 2016) perform, by considering the train alone and train+transfer data together. In (Mooij et al., 2016) datasets, representation learning are not needed; for other datasets when representation learning is needed, existing criterions of these baseline methods could replace the generalization loss. With these results, it would be better to judge the effectiveness of generalization loss. \n\n\nOther Thoughts:\n- Besides the marginal distribution shift, P(A|B) and P(B|A) may both change significantly, even one of them is the correct causal direction, as the underlying distribution may shift.  Can authors comment how their approach could handle such situations?\n- In representation learning, is the decoder never used?\n- since Section 2.4 is only a small part of paper, it may also be worth to also test the causal direction identification without the representation learning part, for example, in linear cases. \n- The statement \"this work in causal representation learning will be helpful for more advanced artificial intelligence\" is rather vague and pompous. \n\nRating: \nTo me the rating is borderline and I'm not yet convinced it is above the acceptance threshold, hence I left it as 5. Hopefully authors' rebuttal could address my concerns. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel idea but lack of empirical comparison, particularly for real datasets. ", "review": "The paper proposes an efficient approach to learn disentanglement representation causal mechanism, based on a generalization loss between a training data set and transfer data sets. Empirical studies show it achieves better sample and computation efficiency than a previous work (Bengio et al., 2020). \n\nPros: \n- Improvement over the existing baseline is significant\n- Theoretical statements on the results, particularly on biased and unbiased estimators, offer good understanding for the approach. \n\nCons:\n-  Currently, I find the setting and results on real datasets rather weak. It would be great if authors could demonstrate such a causal direction identification problem in some real setting, with the distribution shift occurrence between different sets of data. \n- There are no comparison with other methods. It would be great if authors could show baseline methods used in (Mooij et al., 2016) perform, by considering the train alone and train+transfer data together. In (Mooij et al., 2016) datasets, representation learning are not needed; for other datasets when representation learning is needed, existing criterions of these baseline methods could replace the generalization loss. With these results, it would be better to judge the effectiveness of generalization loss. \n\n\nOther Thoughts:\n- Besides the marginal distribution shift, P(A|B) and P(B|A) may both change significantly, even one of them is the correct causal direction, as the underlying distribution may shift.  Can authors comment how their approach could handle such situations?\n- In representation learning, is the decoder never used?\n- since Section 2.4 is only a small part of paper, it may also be worth to also test the causal direction identification without the representation learning part, for example, in linear cases. \n- The statement \"this work in causal representation learning will be helpful for more advanced artificial intelligence\" is rather vague and pompous. \n\nRating: \nTo me the rating is borderline and I'm not yet convinced it is above the acceptance threshold, hence I left it as 5. Hopefully authors' rebuttal could address my concerns. ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603925488803}, {"id": "bL9_GR3Uzj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper265/AnonReviewer3"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper describes an approach for learning a representation U(X,Y), V(X,Y) of data (X,Y) such that U and V are causally meaningful, and U causes V. The approach relies on observing data from two domains, P and Q, where P(V| U) = Q(V |U) (reflecting the causal structure). The approach is a modification of Bengio et al. The main new idea is that the objective function can be tweaked by replacing a KL divergence term with a term involving domain-shift induced generalization errors.\n\nThe paper covers an interesting subject, and the idea to directly use domain generalization error learn causal structure is exciting. However, the paper is clearly not yet baked. The writing is generally poor, and the key ideas have not been formalized. The main ideas of the paper are unclear, as are the validity of the core insights. In particular, there is a fundamental confusion between estimators and estimands.  The paper will require extensive revision and formalization before it's ready for public consumption.\n\nBelow I including some free form thoughts I had below. These give a flavor of my issues with the paper, and will hopefully provide some direction for the authors. However, I stress that these issues are only examples. The paper requires extensive revision.\n\n1. the phonemes/acoustics example just shows that structured learning may be beneficial, it doesn't rely on causality\n2. the explanation of equation 1 is unclear (or, possibly, wrong). Presumably, the actual aim is to compare the likelihoods of two distinct models corresponding to A->B and B<-A. The paper argues that the higher complexity model will have lower likelihood. But this is false in general; a very flexible model will simply memorize the training data.\n3. in general, the paper suffers from a confusion in the notation between population parameters and finite-sample estimators. The notation generally suggests the former, but the prose, appealing to sample-complexity handwaving, suggests the later. \n4. Is proposition 1 meant to be a theorem? The text doesn't reference any estimator, much less an 'unbiased' one.\n(update: I read the appendix, and the intended statement is simply, \"If P(A|B)=Q(A|B), but P(B|A)!=Q(B|A) then 0 = KL(P(A|B),Q(A|B)) < KL(P(B|A),Q(B|A))\"\n5. In section 2.3, \\script{L} has changed from denoting log-likelihood to denoting risk (incorrectly called loss in the prose)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea, clearly not yet baked", "review": "This paper describes an approach for learning a representation U(X,Y), V(X,Y) of data (X,Y) such that U and V are causally meaningful, and U causes V. The approach relies on observing data from two domains, P and Q, where P(V| U) = Q(V |U) (reflecting the causal structure). The approach is a modification of Bengio et al. The main new idea is that the objective function can be tweaked by replacing a KL divergence term with a term involving domain-shift induced generalization errors.\n\nThe paper covers an interesting subject, and the idea to directly use domain generalization error learn causal structure is exciting. However, the paper is clearly not yet baked. The writing is generally poor, and the key ideas have not been formalized. The main ideas of the paper are unclear, as are the validity of the core insights. In particular, there is a fundamental confusion between estimators and estimands.  The paper will require extensive revision and formalization before it's ready for public consumption.\n\nBelow I including some free form thoughts I had below. These give a flavor of my issues with the paper, and will hopefully provide some direction for the authors. However, I stress that these issues are only examples. The paper requires extensive revision.\n\n1. the phonemes/acoustics example just shows that structured learning may be beneficial, it doesn't rely on causality\n2. the explanation of equation 1 is unclear (or, possibly, wrong). Presumably, the actual aim is to compare the likelihoods of two distinct models corresponding to A->B and B<-A. The paper argues that the higher complexity model will have lower likelihood. But this is false in general; a very flexible model will simply memorize the training data.\n3. in general, the paper suffers from a confusion in the notation between population parameters and finite-sample estimators. The notation generally suggests the former, but the prose, appealing to sample-complexity handwaving, suggests the later. \n4. Is proposition 1 meant to be a theorem? The text doesn't reference any estimator, much less an 'unbiased' one.\n(update: I read the appendix, and the intended statement is simply, \"If P(A|B)=Q(A|B), but P(B|A)!=Q(B|A) then 0 = KL(P(A|B),Q(A|B)) < KL(P(B|A),Q(B|A))\"\n5. In section 2.3, \\script{L} has changed from denoting log-likelihood to denoting risk (incorrectly called loss in the prose)\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603843967428}], "openreview_url": "https://openreview.net/forum?id=Sva-fwURywB", "arxiv_id": "2201.01942", "paper_pdf": "papers/Sva-fwURywB.pdf", "paper_pdf_sha256": "bf67f56ed84f6d2d9af097ce4ef221895e78a6c6bc21de2fc264b3adf0bfc476", "paper_pdf_bytes": 518121, "paper_pdf_source": "openreview", "code_url": "https://github.com/yuanpeng16/EDCR", "code_repository": "yuanpeng16/EDCR", "code_commit": "781194191a82fbc85d43084af63e6a20c59ae670", "code_archive": "repos/Sva-fwURywB.zip", "code_archive_sha256": "e64bd48f6e961e3cfdecfcef5a7643c22c1d0b7ba4f82f3271226995d46e4813", "code_archive_bytes": 1298441, "code_file_count": 59, "code_extensions": {".py": 53, ".ipynb": 6}, "github_disk_usage_kb": 1256, "github_languages": {"Jupyter Notebook": 279672, "Python": 166392}, "github_archived": false, "github_pushed_at": "2023-11-24T16:13:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficiently-disentangle-causal-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SkglVlSFPS", "year": 2020, "status": "rejected", "title": "Uncertainty - sensitive learning and planning with ensembles", "authors": ["Piotr Miłoś", "Łukasz Kuciński", "Konrad Czechowski", "Piotr Kozakowski", "Maciej Klimek"], "authorids": ["pmilos@mimuw.edu.pl", "lukasz.kucinski@gmail.com", "konrad.czechowski@gmail.com", "p.kozakowski@mimuw.edu.pl", "maciej.klimek@gmail.com"], "authors_source": "OpenReview API", "abstract": "We propose a reinforcement learning framework for discrete environments in which an agent optimizes its behavior on two timescales. For the short one, it uses tree search methods to perform tactical decisions. The long strategic level is handled with an ensemble of value functions learned using $TD$-like backups. Combining these two techniques brings synergies. The planning module performs \\textit{what-if} analysis allowing to avoid short-term pitfalls and boost backups of the value function. Notably, our method performs well in environments with sparse rewards where standard $TD(1)$ backups fail. On the other hand, the value functions compensate for inherent short-sightedness of planning. Importantly, we use ensembles to measure the epistemic uncertainty of value functions. This serves two purposes: a) it stabilizes planning, b) it guides exploration. \n\nWe evaluate our methods on discrete environments with sparse rewards: the Deep sea chain environment, toy Montezuma's Revenge, and Sokoban. In all the cases, we obtain speed-up of learning and boost to the final performance.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "HJlP92j69H", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2236/AnonReviewer3"], "rating": "1: Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Uncertainty-Sensitive Learning and Planning with Ensembles\n=====================================================\n\nThis paper investigates the use of uncertainty-aware estimates in solving planning problems (RL with access to simulator).\nThe proposed algorithm combines a learned model-free value estimate with MCTS planning.\nAn ensemble of neural networks is used to model posterior uncertainty in the value estimate and drive efficient exploration.\n\n\nThere are several things to like about this paper:\n- The paper takes on several core issues in RL/planning research, most notably the synthesis of dealing with model-based and model-free uncertainty in RL.\n- The general flavour of the paper + algorithm seems to be reasonable. The proposal to use ensemble uncertainty estimates to drive model-based MCTS is interesting, natural, and I think it's a good one.\n- The proposed structure of the paper is quite nice, there is mostly a linear and logical progression of complexity in the experiments. This is nice to see clear benefits of the approach on the simplest possible settings and build up from there.\n- The effort to open source code + implementation details is laudable.\n\nHowever, there are several places where this paper falls short:\n- In general, the claims and results of the paper are far too vague to be fully understood and replicated. Take the main algorithm 1, it really seems like more of a \"sketch\" of a very general family of algorithms, rather than a specific description of a clear algorithm.\n- This vagueness is spread throughout the plots and figures as well... note that Figure 1 has no indication of how many steps have been evaluated, and Figure 2 has no indication for what value K > 0 was actually used. The clarity does not improve in Sections 3.2 and 3.3 where quite inconsistent performance metrics and presentations are presented.\n- Generally, the writing could be tightened quite a lot. In particular I would encourage you to think about whether each statement you make is clearly supported by some theorem, experiment or plot in your paper. For example, on page 3 \"We found this mechanism to be beneficial... see Section 3.3\" but then it's not clear exactly what statement shows that particular part of the mechanism was helpful, versus other issues associated with ensemble learning. There are more than a few typos... the on(e) in Osband... akin to ??... might be obtained by choosing from (the) ensemble...\n- It would be very helpful to clarify that the agent is given access to a simulator... so that this is not exactly the typical RL setting of sequential decision making. This should appear early in the paper.\n- The code that is released with the paper is also quite confusing, it is not structured with a clear README and includes many sections of dead/commented code. I was hoping the code might rescue some of the clarity, but I think that still needs work.\n\nOverall, I do think there is some interesting material here...\nIt's an important problem, and the core building blocks of combining model, value and uncertainty for better exploration is interesting.\nHowever, I just think the actual paper is not clear enough on the details.\nMy belief is that going through this paper very methodically and carefully to make sure that every single detail + claim is rigorously supported would help this paper immensely.\nFor that reason I have to say that I think it's a \"reject\" in its current form.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"rating": "1: Reject", "experience_assessment": "I have published in this field for several years.", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review": "Uncertainty-Sensitive Learning and Planning with Ensembles\n=====================================================\n\nThis paper investigates the use of uncertainty-aware estimates in solving planning problems (RL with access to simulator).\nThe proposed algorithm combines a learned model-free value estimate with MCTS planning.\nAn ensemble of neural networks is used to model posterior uncertainty in the value estimate and drive efficient exploration.\n\n\nThere are several things to like about this paper:\n- The paper takes on several core issues in RL/planning research, most notably the synthesis of dealing with model-based and model-free uncertainty in RL.\n- The general flavour of the paper + algorithm seems to be reasonable. The proposal to use ensemble uncertainty estimates to drive model-based MCTS is interesting, natural, and I think it's a good one.\n- The proposed structure of the paper is quite nice, there is mostly a linear and logical progression of complexity in the experiments. This is nice to see clear benefits of the approach on the simplest possible settings and build up from there.\n- The effort to open source code + implementation details is laudable.\n\nHowever, there are several places where this paper falls short:\n- In general, the claims and results of the paper are far too vague to be fully understood and replicated. Take the main algorithm 1, it really seems like more of a \"sketch\" of a very general family of algorithms, rather than a specific description of a clear algorithm.\n- This vagueness is spread throughout the plots and figures as well... note that Figure 1 has no indication of how many steps have been evaluated, and Figure 2 has no indication for what value K > 0 was actually used. The clarity does not improve in Sections 3.2 and 3.3 where quite inconsistent performance metrics and presentations are presented.\n- Generally, the writing could be tightened quite a lot. In particular I would encourage you to think about whether each statement you make is clearly supported by some theorem, experiment or plot in your paper. For example, on page 3 \"We found this mechanism to be beneficial... see Section 3.3\" but then it's not clear exactly what statement shows that particular part of the mechanism was helpful, versus other issues associated with ensemble learning. There are more than a few typos... the on(e) in Osband... akin to ??... might be obtained by choosing from (the) ensemble...\n- It would be very helpful to clarify that the agent is given access to a simulator... so that this is not exactly the typical RL setting of sequential decision making. This should appear early in the paper.\n- The code that is released with the paper is also quite confusing, it is not structured with a clear README and includes many sections of dead/commented code. I was hoping the code might rescue some of the clarity, but I think that still needs work.\n\nOverall, I do think there is some interesting material here...\nIt's an important problem, and the core building blocks of combining model, value and uncertainty for better exploration is interesting.\nHowever, I just think the actual paper is not clear enough on the details.\nMy belief is that going through this paper very methodically and carefully to make sure that every single detail + claim is rigorously supported would help this paper immensely.\nFor that reason I have to say that I think it's a \"reject\" in its current form."}, "tcdate": 1572875406656}, {"id": "H1lRg1qtcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2236/AnonReviewer4"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose to combine planning methods like MCTS with an ensemble of value functions to a) estimate the value of leaf nodes of the search tree and b) use the ensemble estimate of uncertainty to guide exploration during MCTS search. \nThe MCTS rollouts are also used as optimization targets for the value function.\n\nI believe this is a clear reject. On the one hand, the paper needs signficiantly more work on the writing and clarity. On the other hand I have several worries on the method and evaluation side. \n\nRegarding the presentation of the paper:\nOverall, the paper seems quite rushed. This is not a strong reason for rejection but should be improved in a future version. For example, punctuation and sentence structure is often wrong, the paper has only slighlty over 7 pages, a citation is undefined on p.7 and images and whitespace is formatted wrongly on occasion (e.g. top of page 6).\nMore importantly, on the content side, the experimental section is sufficiently clear and well written, however, the method description needs more detail and background information. The paper relies on several prior works which are referred to but not described (E.g. MCTS , the sampling mechanism by Osband et al. which they are using but not describing, the 'mask' from Osband et al which they are using but not describing).\nFurthermore, the algorithm itself is not described in sufficient detail:\n- How does the 'soft-penalization' work?\n- How exactly does the mechanism \"similar in fashing to\" Thomson sampling work?\n- Are you learning a model or do you have access to the true transition function?\n\nRegarding the method:\nI can't say anything definitive about the method as I'm not entirely clear how exactly it works. However, I have several worries that might need addressing:\n- It seems to me that the method relies on access to the _true_ transition and reward function and not on a learned model. This is a big difference to much of the prior work they compare against. This also makes the comparison against any pure model free method like PPO much less meaningful.\n- Similarly, manually avoiding dead-ends and loops is a very strong assumption \n- Also, being able to distinguish and use a fixed ratio of \"solved\" and \"unsolved\" episodes is a strong assumption. \n- The one main contribution seems to be a new way of how \\phi_a(x) is defined. Their particular choice needs a clearer motivation. Furthermore, if there is more contribution and differences to prior work, highlighting them more would help the reader understand the contribution. \n- As the work makes several strong assumptions regarding the environment and access to the model, significantly more work (e.g. ablation studies) is needed to clearly show which assumption and feature of the algorithm is important for performance (and ideally also why). For example (but that's just a first idea): To understand the impact of their choice of \\phi vs. their planning architecture, it would be be interesting to maybe train PPO using an exploration bonus based on \\phi. This would allow disentangling the contribution of: Access to the true model, \"discrete-environment-tricks\" like penalizing dead-ends, and exploration incentivication of \\phi. \n\nEdit:\nThank you for your response and the updated manuscript, which reads considerably better.\nI also agree with your point regarding the strength of assumption regarding \"solved\" and \"unsolved\" episodes.\n\nConsequently, I will raise my score to a \"weak reject\" to express that I think this is promising work.\n\nI do believe that ablation studies would add a lot to the paper as they would allow one to see which of the (many) added components help how much, for example between the selection function $\\phi$ and the various penalizations used. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "3: Weak Reject", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #4", "review": "The authors propose to combine planning methods like MCTS with an ensemble of value functions to a) estimate the value of leaf nodes of the search tree and b) use the ensemble estimate of uncertainty to guide exploration during MCTS search. \nThe MCTS rollouts are also used as optimization targets for the value function.\n\nI believe this is a clear reject. On the one hand, the paper needs signficiantly more work on the writing and clarity. On the other hand I have several worries on the method and evaluation side. \n\nRegarding the presentation of the paper:\nOverall, the paper seems quite rushed. This is not a strong reason for rejection but should be improved in a future version. For example, punctuation and sentence structure is often wrong, the paper has only slighlty over 7 pages, a citation is undefined on p.7 and images and whitespace is formatted wrongly on occasion (e.g. top of page 6).\nMore importantly, on the content side, the experimental section is sufficiently clear and well written, however, the method description needs more detail and background information. The paper relies on several prior works which are referred to but not described (E.g. MCTS , the sampling mechanism by Osband et al. which they are using but not describing, the 'mask' from Osband et al which they are using but not describing).\nFurthermore, the algorithm itself is not described in sufficient detail:\n- How does the 'soft-penalization' work?\n- How exactly does the mechanism \"similar in fashing to\" Thomson sampling work?\n- Are you learning a model or do you have access to the true transition function?\n\nRegarding the method:\nI can't say anything definitive about the method as I'm not entirely clear how exactly it works. However, I have several worries that might need addressing:\n- It seems to me that the method relies on access to the _true_ transition and reward function and not on a learned model. This is a big difference to much of the prior work they compare against. This also makes the comparison against any pure model free method like PPO much less meaningful.\n- Similarly, manually avoiding dead-ends and loops is a very strong assumption \n- Also, being able to distinguish and use a fixed ratio of \"solved\" and \"unsolved\" episodes is a strong assumption. \n- The one main contribution seems to be a new way of how \\phi_a(x) is defined. Their particular choice needs a clearer motivation. Furthermore, if there is more contribution and differences to prior work, highlighting them more would help the reader understand the contribution. \n- As the work makes several strong assumptions regarding the environment and access to the model, significantly more work (e.g. ablation studies) is needed to clearly show which assumption and feature of the algorithm is important for performance (and ideally also why). For example (but that's just a first idea): To understand the impact of their choice of \\phi vs. their planning architecture, it would be be interesting to maybe train PPO using an exploration bonus based on \\phi. This would allow disentangling the contribution of: Access to the true model, \"discrete-environment-tricks\" like penalizing dead-ends, and exploration incentivication of \\phi. \n\nEdit:\nThank you for your response and the updated manuscript, which reads considerably better.\nI also agree with your point regarding the strength of assumption regarding \"solved\" and \"unsolved\" episodes.\n\nConsequently, I will raise my score to a \"weak reject\" to express that I think this is promising work.\n\nI do believe that ablation studies would add a lot to the paper as they would allow one to see which of the (many) added components help how much, for example between the selection function $\\phi$ and the various penalizations used. ", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory."}, "tcdate": 1572605686009}, {"id": "HkgcGAHAFH", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper2236/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an approach blending model-based, model-free methods and utilizing risk-sensitivity information in ensembles as part of the value estimation and exploration process. The exploration is based on risk- sensitivity measures such as moments and relative majority vote. There is a lot of work currently trying to marry the model free with model based approaches for integrated planning and learning as the authors have mentioned in the related work section of the paper and also called out similar methods and techniques. The authors have provided evidence via experiments in three environments and shown good results of using this blended approach. Code is also provided for others to further carry out explorations in this research area.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "The paper proposes an approach blending model-based, model-free methods and utilizing risk-sensitivity information in ensembles as part of the value estimation and exploration process. The exploration is based on risk- sensitivity measures such as moments and relative majority vote. There is a lot of work currently trying to marry the model free with model based approaches for integrated planning and learning as the authors have mentioned in the related work section of the paper and also called out similar methods and techniques. The authors have provided evidence via experiments in three environments and shown good results of using this blended approach. Code is also provided for others to further carry out explorations in this research area.\n"}, "tcdate": 1571868177960}], "openreview_url": "https://openreview.net/forum?id=SkglVlSFPS", "arxiv_id": null, "paper_pdf": "papers/SkglVlSFPS.pdf", "paper_pdf_sha256": "27ed82277dff27e82e29165d57b6a1e4763e80bbf172d1e44eff01c704454099", "paper_pdf_bytes": 1406946, "paper_pdf_source": "openreview", "code_url": "https://github.com/learningandplanningICLR/learningandplanning", "code_repository": "learningandplanningICLR/learningandplanning", "code_commit": "10eac96831d9e61331c2879d33e12a117d259a2b", "code_archive": "repos/SkglVlSFPS.zip", "code_archive_sha256": "b81d7b2067cab901138e40f947b1a23f0c7e5aa9abfe3cbe4f287654457fd396", "code_archive_bytes": 8693125, "code_file_count": 291, "code_extensions": {".py": 284, ".ipynb": 4, ".cpp": 2, ".sh": 1}, "github_disk_usage_kb": 9515, "github_languages": {"Python": 355147, "Shell": 104}, "github_archived": false, "github_pushed_at": "2020-02-12T18:43:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/uncertainty-sensitive-learning-and-planning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FafH4GM4uc", "year": 2026, "status": "rejected", "title": "Universal Reasoner: A Single, Composable Plug-and-Play Reasoner for Frozen LLMs", "authors": ["Jaemin Kim", "Hangeol Chang", "Hyunmin Hwang", "Choonghan Kim", "Jong Chul Ye"], "authorids": ["~Jaemin_Kim2", "~Hangeol_Chang1", "~Hyunmin_Hwang1", "~Choonghan_Kim1", "~Jong_Chul_Ye1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) have demonstrated remarkable general capabilities, but enhancing skills such as reasoning often demands substantial computational resources and may compromise their generalization. While Parameter-Efficient Fine-Tuning (PEFT) methods offer a more resource-conscious alternative, they typically require retraining for each LLM backbone due to architectural dependencies. To address these challenges, here we propose Universal Reasoner (UniR) - a single, lightweight, composable, and plug-and-play reasoning module that can be used with larger frozen LLMs to endow it with specialized reasoning capabilities. Specifically, UniR decomposes the reward into a standalone reasoning module that is trained in a decoupled manner using predefined rewards, effectively translating trajectory-level signals into token-level guidance. Once trained, UniR can be combined with frozen LLMs at inference time by simply adding its output logits to those of the LLM backbones. This additive structure naturally enables modular composition: multiple UniR modules trained for different tasks can be jointly applied by summing their logits, enabling complex reasoning via composition. Experiments on mathematical reasoning and machine translation show that UniR surpasses existing baseline fine-tuning methods. Furthermore, UniR demonstrates weak-to-strong generalization: reasoning modules trained on smaller models effectively guide much larger LLMs. Beyond this, UniR generalizes across modalities such as in vision language models and domains such as medical reasoning. This makes UniR a cost-efficient, adaptable, and robust solution for enhancing reasoning in LLMs without compromising their core capabilities.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "2MswMcOGGu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5828/Reviewer_sxY5"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The paper introduces UniR, a lightweight reasoning module trained with verifiable rewards that can be plugged into frozen LLMs at inference by simply adding its logits to those of the backbone. The module is (i) architecture-agnostic (only tokenizer alignment required), (ii) composable (multiple UniRs can be summed for multi-task reasoning), and (iii) shown to transfer from small to much larger backbones without further tuning. Experiments on mathematical reasoning, machine translation, vision-language math, and medical risk prediction report gains over full-model GRPO fine-tuning while using far fewer trainable parameters and less memory.", "review_text": "The paper introduces UniR, a lightweight reasoning module trained with verifiable rewards that can be plugged into frozen LLMs at inference by simply adding its logits to those of the backbone. The module is (i) architecture-agnostic (only tokenizer alignment required), (ii) composable (multiple UniRs can be summed for multi-task reasoning), and (iii) shown to transfer from small to much larger backbones without further tuning. Experiments on mathematical reasoning, machine translation, vision-language math, and medical risk prediction report gains over full-model GRPO fine-tuning while using far fewer trainable parameters and less memory.", "strengths": "Originality\n- Proposes a clean, modular training objective mapping trajectory reward to token-level log-probabilities and uses additive logit fusion at inference which is a conceptually neat design.\n- Inference-time composition via logit addition is derived from a KL-regularized multi-objective RL objective (Eq. 6), giving the method theoretical grounding absent in prior adapter ensembles.\n\nEmpirical Quality\n- Consistent gains across 5 math benchmarks (up to +7.4 avg. over GRPO-full) and 2 translation directions (+1.8 BLEU) with ≤ 1.5 B trainable parameters.\n- Strong weak-to-strong transfer: a 1.5 B module trained with a 3 B backbone improves a frozen 14 B Qwen-2.5 by +2.5 avg. points (Fig. 2).\n- Composability demo: combining math + translation modules on German math→English task beats either specialist alone (Fig. 3).\n\nClarity & Reproducibility\n- Complete hyper-parameter tables (App. A.1), reward definitions, and prompts (App. D) provided.\n- Seeds and std-dev reported (10 runs, Tables 1–2).\n\nSignificance\n- Enables cheap, plug-and-play specialization of proprietary or edge-deployed models without gradients through the backbone.", "weaknesses": "- Tokenizer-alignment prerequisite limits practicality. The paper claims “architecture-agnostic” but requires shared tokenizer (sec. 1, line 6). Is there a way to extend the method with tokenizer-mismatch adaptation strategy (e.g., embedding-level mapping layer trained with UniR) or can we quantify performance drop under vocabulary misalignment.\n\n- The central identity (Eq. (4)), representing a trajectory reward as the sum of token log-probabilities of some policy, is assumed and only informally justified; Theorem 1 is informal and depends on strong assumptions (convergence of $\\pi_\\theta$ to $\\pi_{\\theta^*}$, uniqueness of optimum, and that π_r can exactly represent the reward) (Sec. 4.1 & 4.3). Can you provide a formal version of the theorem in the appendix with the details of the assumptions and proofs. The proof relies on the assumption that the guided policy has converged to the optimum. It would be beneficial to discuss the conditions under which this assumption holds. \n\n- The paper compares to GRPO variants and LoRA but does not compare to other recent inference-time guidance / token-level reward approaches (beyond brief GenARM mention).\n\n- Missing Related Work: The paper fails to cite or compare against a relevant line of work on plug-and-play modules for frozen transformers. Specifically, TART (\"TART: A Plug-and-Play Transformer for Zero-Shot Task Adaptation\", Bhatia et al., ICLR 2022) introduces a nearly identical architectural paradigm: a small, separately trained module whose outputs are combined with a frozen backbone model to achieve task adaptation. While TART's training objective and application (zero-shot classification) differ, the core concept of a \"plug-and-play reasoner for frozen LLMs\" is not entirely new.", "questions": "Please see the weakness section above for relevant questions. In addition, \n\n- Can you show failure cases where composed modules hurt performance? What are the symptoms of negative interference?\n- Code release plan -- will the code for this paper be released?\n- Regarding composability, how do you envision setting the combination weights $\\alpha_i$ in a real-world scenario without an expensive sweep for each new task combination? Have you considered methods for dynamically learning or setting these weights?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces UniR, a lightweight reasoning module trained with verifiable rewards that can be plugged into frozen LLMs at inference by simply adding its logits to those of the backbone. The module is (i) architecture-agnostic (only tokenizer alignment required), (ii) composable (multiple UniRs can be summed for multi-task reasoning), and (iii) shown to transfer from small to much larger backbones without further tuning. Experiments on mathematical reasoning, machine translation, vision-language math, and medical risk prediction report gains over full-model GRPO fine-tuning while using far fewer trainable parameters and less memory.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "Originality\n- Proposes a clean, modular training objective mapping trajectory reward to token-level log-probabilities and uses additive logit fusion at inference which is a conceptually neat design.\n- Inference-time composition via logit addition is derived from a KL-regularized multi-objective RL objective (Eq. 6), giving the method theoretical grounding absent in prior adapter ensembles.\n\nEmpirical Quality\n- Consistent gains across 5 math benchmarks (up to +7.4 avg. over GRPO-full) and 2 translation directions (+1.8 BLEU) with ≤ 1.5 B trainable parameters.\n- Strong weak-to-strong transfer: a 1.5 B module trained with a 3 B backbone improves a frozen 14 B Qwen-2.5 by +2.5 avg. points (Fig. 2).\n- Composability demo: combining math + translation modules on German math→English task beats either specialist alone (Fig. 3).\n\nClarity & Reproducibility\n- Complete hyper-parameter tables (App. A.1), reward definitions, and prompts (App. D) provided.\n- Seeds and std-dev reported (10 runs, Tables 1–2).\n\nSignificance\n- Enables cheap, plug-and-play specialization of proprietary or edge-deployed models without gradients through the backbone.", "weaknesses": "- Tokenizer-alignment prerequisite limits practicality. The paper claims “architecture-agnostic” but requires shared tokenizer (sec. 1, line 6). Is there a way to extend the method with tokenizer-mismatch adaptation strategy (e.g., embedding-level mapping layer trained with UniR) or can we quantify performance drop under vocabulary misalignment.\n\n- The central identity (Eq. (4)), representing a trajectory reward as the sum of token log-probabilities of some policy, is assumed and only informally justified; Theorem 1 is informal and depends on strong assumptions (convergence of $\\pi_\\theta$ to $\\pi_{\\theta^*}$, uniqueness of optimum, and that π_r can exactly represent the reward) (Sec. 4.1 & 4.3). Can you provide a formal version of the theorem in the appendix with the details of the assumptions and proofs. The proof relies on the assumption that the guided policy has converged to the optimum. It would be beneficial to discuss the conditions under which this assumption holds. \n\n- The paper compares to GRPO variants and LoRA but does not compare to other recent inference-time guidance / token-level reward approaches (beyond brief GenARM mention).\n\n- Missing Related Work: The paper fails to cite or compare against a relevant line of work on plug-and-play modules for frozen transformers. Specifically, TART (\"TART: A Plug-and-Play Transformer for Zero-Shot Task Adaptation\", Bhatia et al., ICLR 2022) introduces a nearly identical architectural paradigm: a small, separately trained module whose outputs are combined with a frozen backbone model to achieve task adaptation. While TART's training objective and application (zero-shot classification) differ, the core concept of a \"plug-and-play reasoner for frozen LLMs\" is not entirely new.", "questions": "Please see the weakness section above for relevant questions. In addition, \n\n- Can you show failure cases where composed modules hurt performance? What are the symptoms of negative interference?\n- Code release plan -- will the code for this paper be released?\n- Regarding composability, how do you envision setting the combination weights $\\alpha_i$ in a real-world scenario without an expensive sweep for each new task combination? Have you considered methods for dynamically learning or setting these weights?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1762143301792}, {"id": "6yVD5VFj0L", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5828/Reviewer_xWTC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "The paper proposes a single \"universal\" model that converts practically any LLM into a reasoning-enabled LLM by adding logits at inference time. This enables reasoning behavior without the cost of fine-tuning and without the reduction in quality that reasoning-fine-tuning often produces.", "review_text": "The paper proposes a single \"universal\" model that converts practically any LLM into a reasoning-enabled LLM by adding logits at inference time. This enables reasoning behavior without the cost of fine-tuning and without the reduction in quality that reasoning-fine-tuning often produces.", "strengths": "This model seems very useful both as a research artifact (it seems to tell us that there isn't all that much to reasoning behavior; not many weights are actually needed to represent it \"from scratch\") as well as to users of AI tech. Enabling a \"plug-and-play\" ecosystem of foundation models and reasoning models could have a high impact on the community.", "weaknesses": "Some of the baseline numbers are suspicious. For example, this paper reports a \"backbone only\" performance on MATH of 28.7 for llama3.2 3B, but huggingface here (https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) reports a 48.0 accuracy and that's consistent with what ArtificialAnalysis reports (https://artificialanalysis.ai/evaluations/math-500).  I'm also having trouble finding the other baseline numbers in prior work or online (GSM8K seems to mostly be 8-shot in evaluations of llama3 as opposed to the 0-shot evaluated in the paper).", "questions": "Are there sources that can corroborate that the baseline numbers in Table 1 and Table 2 are correct (i.e. what we should expect for these models)?\n\nCan you go into more detail about the architecture of the universal reasoner? The paper is vague on this point.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a single \"universal\" model that converts practically any LLM into a reasoning-enabled LLM by adding logits at inference time. This enables reasoning behavior without the cost of fine-tuning and without the reduction in quality that reasoning-fine-tuning often produces.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "This model seems very useful both as a research artifact (it seems to tell us that there isn't all that much to reasoning behavior; not many weights are actually needed to represent it \"from scratch\") as well as to users of AI tech. Enabling a \"plug-and-play\" ecosystem of foundation models and reasoning models could have a high impact on the community.", "weaknesses": "Some of the baseline numbers are suspicious. For example, this paper reports a \"backbone only\" performance on MATH of 28.7 for llama3.2 3B, but huggingface here (https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) reports a 48.0 accuracy and that's consistent with what ArtificialAnalysis reports (https://artificialanalysis.ai/evaluations/math-500).  I'm also having trouble finding the other baseline numbers in prior work or online (GSM8K seems to mostly be 8-shot in evaluations of llama3 as opposed to the 0-shot evaluated in the paper).", "questions": "Are there sources that can corroborate that the baseline numbers in Table 1 and Table 2 are correct (i.e. what we should expect for these models)?\n\nCan you go into more detail about the architecture of the universal reasoner? The paper is vague on this point.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762136738721}, {"id": "wcCq13szhB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5828/Reviewer_PpAD"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper studies LLM reasoning tasks and proposes UniR, a standalone reasoning module that reduces training costs while achieving generalization across models. UniR is composable and trained via GRPO. It provides token-level logit guidance to the backbone LLM. UniR demonstrates better performance than GRPO LoRA and GRPO Full with lower training costs. Additionally, UniR trained on a smaller backbone can improve task performance when combined with a larger backbone.", "review_text": "This paper studies LLM reasoning tasks and proposes UniR, a standalone reasoning module that reduces training costs while achieving generalization across models. UniR is composable and trained via GRPO. It provides token-level logit guidance to the backbone LLM. UniR demonstrates better performance than GRPO LoRA and GRPO Full with lower training costs. Additionally, UniR trained on a smaller backbone can improve task performance when combined with a larger backbone.", "strengths": "This paper has a clear motivation: reduce computational costs in post-training and improve the generalization of learned reasoning abilities. \n\nThe content is high quality and the writing is clear and easy to understand.\n\nThe experiments are extensive, covering various benchmarks, tasks, model sizes, and series. The proposed UniR shows better performance compared with baseline methods and improves performance on larger backbones that were not used to train UniR.\n\nWhen combined, different UniR modules show adjustable performance improvements on math and translation tasks.", "weaknesses": "One weakness is the lack of detailed discussion about time and resource costs. The paper should highlight how UniR differs from baseline methods to demonstrate its computational advantages. \n\nAnother point: the paper needs more analysis of the design choices for the standalone reasoning module. The authors use a smaller LLM as the reasoning module, which is a good initial step. However, have they experimented with using parts of a model instead of a full LLM? Could the reasoning module be further reduced in size?\n\nAdditionally, more analysis comparing outputs from baseline GRPO versus UniR would better illustrate UniR's influence. For example, what helpful adjustments does token-level logit modification provide compared to optimizing the model using only the final reward signal?", "questions": "Regarding the first contribution (Line 75–76): circumventing the need for expensive preference dataset creation is not a major contribution of this work. The tasks studied here do not require such preference dataset construction.\n\nIn Figure 3, what is the baseline performance—that is, only the base model without the reasoning module? Showing the baseline performance would help readers better understand this figure.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies LLM reasoning tasks and proposes UniR, a standalone reasoning module that reduces training costs while achieving generalization across models. UniR is composable and trained via GRPO. It provides token-level logit guidance to the backbone LLM. UniR demonstrates better performance than GRPO LoRA and GRPO Full with lower training costs. Additionally, UniR trained on a smaller backbone can improve task performance when combined with a larger backbone.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper has a clear motivation: reduce computational costs in post-training and improve the generalization of learned reasoning abilities. \n\nThe content is high quality and the writing is clear and easy to understand.\n\nThe experiments are extensive, covering various benchmarks, tasks, model sizes, and series. The proposed UniR shows better performance compared with baseline methods and improves performance on larger backbones that were not used to train UniR.\n\nWhen combined, different UniR modules show adjustable performance improvements on math and translation tasks.", "weaknesses": "One weakness is the lack of detailed discussion about time and resource costs. The paper should highlight how UniR differs from baseline methods to demonstrate its computational advantages. \n\nAnother point: the paper needs more analysis of the design choices for the standalone reasoning module. The authors use a smaller LLM as the reasoning module, which is a good initial step. However, have they experimented with using parts of a model instead of a full LLM? Could the reasoning module be further reduced in size?\n\nAdditionally, more analysis comparing outputs from baseline GRPO versus UniR would better illustrate UniR's influence. For example, what helpful adjustments does token-level logit modification provide compared to optimizing the model using only the final reward signal?", "questions": "Regarding the first contribution (Line 75–76): circumventing the need for expensive preference dataset creation is not a major contribution of this work. The tasks studied here do not require such preference dataset construction.\n\nIn Figure 3, what is the baseline performance—that is, only the base model without the reasoning module? Showing the baseline performance would help readers better understand this figure.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762012168177}, {"id": "786SGnH31g", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5828/Reviewer_8VRM"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper proposes a method to decouple reasoning from the base model by introducing a standalone, pluggable reasoning module that can be integrated with any model size or architecture by adding its output logits to those of the base model. The paper shows that it is possible to compose even two reasoning modules to combine task-specific abilities, achieving improved performance over both the base model and the GRPO-trained model (in math and translation tasks).", "review_text": "This paper proposes a method to decouple reasoning from the base model by introducing a standalone, pluggable reasoning module that can be integrated with any model size or architecture by adding its output logits to those of the base model. The paper shows that it is possible to compose even two reasoning modules to combine task-specific abilities, achieving improved performance over both the base model and the GRPO-trained model (in math and translation tasks).", "strengths": "1. A decoupled reasoning module that is model- and architecture-agnostic (assuming tokenizers can be aligned) with the base model, demonstrating that reasoning can be isolated and trained independently.\n\n2. The reasoning module can be seamlessly integrated by adding its output logits to those of the base model, and it appears to enhance performance even in much larger base models.\n\n3. A linear composition of reasoning modules is possible and effective, generalizing well even to unseen test sets (e.g., mortality prediction).\n\n4. The paper is well presented and demonstrates strong understanding and integration of relevant theoretical concepts.", "weaknesses": "1. How are the coefficients determined when combining reasoning modules?\n\n2. Is it possible to compose more than two modules, and if so, how does performance scale with additional compositions?", "questions": "1. Does the reasoning module require any SFT warm start? Prior studies have shown that directly starting with GRPO, especially on small models, often leads to suboptimal performance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to decouple reasoning from the base model by introducing a standalone, pluggable reasoning module that can be integrated with any model size or architecture by adding its output logits to those of the base model. The paper shows that it is possible to compose even two reasoning modules to combine task-specific abilities, achieving improved performance over both the base model and the GRPO-trained model (in math and translation tasks).", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. A decoupled reasoning module that is model- and architecture-agnostic (assuming tokenizers can be aligned) with the base model, demonstrating that reasoning can be isolated and trained independently.\n\n2. The reasoning module can be seamlessly integrated by adding its output logits to those of the base model, and it appears to enhance performance even in much larger base models.\n\n3. A linear composition of reasoning modules is possible and effective, generalizing well even to unseen test sets (e.g., mortality prediction).\n\n4. The paper is well presented and demonstrates strong understanding and integration of relevant theoretical concepts.", "weaknesses": "1. How are the coefficients determined when combining reasoning modules?\n\n2. Is it possible to compose more than two modules, and if so, how does performance scale with additional compositions?", "questions": "1. Does the reasoning module require any SFT warm start? Prior studies have shown that directly starting with GRPO, especially on small models, often leads to suboptimal performance.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761938440325}], "openreview_url": "https://openreview.net/forum?id=FafH4GM4uc", "arxiv_id": "2505.19075", "paper_pdf": "papers/FafH4GM4uc.pdf", "paper_pdf_sha256": "7cdf1137b6779da375c9f77bbfdab694abb343a6cfa6c997640e22a5c04b707a", "paper_pdf_bytes": 3057012, "paper_pdf_source": "openreview", "code_url": "https://github.com/hangeol/UniR", "code_repository": "hangeol/UniR", "code_commit": "c262e784d40c226621949daad6a72203fe5f3b83", "code_archive": "repos/FafH4GM4uc.zip", "code_archive_sha256": "69b4a977714a2edf9b394fbc1faba67124487ca9ea3c06d24188a6b37f5925df", "code_archive_bytes": 4324357, "code_file_count": 19, "code_extensions": {".py": 16, ".sh": 3}, "github_disk_usage_kb": 4200, "github_languages": {"Python": 395487, "Shell": 8498}, "github_archived": false, "github_pushed_at": "2025-11-26T12:37:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/universal-reasoner-a-single-composable-plug"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YUaQ6qrbRL", "year": 2025, "status": "rejected", "title": "Right Time to Learn: Promoting Generalization via Bio-inspired Spacing Effect in Knowledge Distillation", "authors": ["Guanglong Sun", "Hongwei Yan", "Liyuan Wang", "Qian Li", "Bo Lei", "Yi Zhong"], "authorids": ["~Guanglong_Sun1", "~Hongwei_Yan1", "~Liyuan_Wang1", "~Qian_Li3", "~Bo_Lei1", "~Yi_Zhong1"], "authors_source": "OpenReview API", "abstract": "Knowledge distillation (KD) is a powerful strategy for training deep neural networks (DNNs). While it was originally proposed to train a more compact “student” model from a large “teacher” model, many recent efforts have focused on adapting it as an effective way to promote generalization of the model itself, such as online KD and self KD. Here, we propose an easy-to-use and compatible strategy named Spaced KD to improve the effectiveness of both online KD and self KD, in which the student model distills knowledge from a teacher model trained with a space interval ahead. This strategy is inspired by a prominent theory named spacing effect in the field of biological learning and memory, positing that appropriate intervals between learning trials can significantly enhance learning performance. We provide an in-depth theoretical and empirical analysis showing that the benefits of the proposed spacing effect in KD stem from seeking a flat minima during stochastic gradient descent (SGD). We perform extensive experiments to demonstrate the effectiveness of our Spaced KD in improving the learning performance of DNNs (e.g., the additional performance gain is up to 2.31% and 3.34% on Tiny-ImageNet over online KD and self KD, respectively).", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "xiNZhzrhWg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2884/Reviewer_Ac8Y"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "In this work, the authors perform a scheme on the lines of online and self KD to promote generalization of the model. They propose SpacedKD which is inspired from spacing effect in biological learning which highlights that study sessions across time is critical for memory formation. They argue that this spacing effect can be generalized to a kind of ensemble learning at the temporal scale. They implement this spaced KD in both online and self KD setups. For online KD, they just train the teacher in advanced for a certain number of steps termed as spaced interval. Then they just freeze this teacher and distill this knowledge to the student using the same data. They further provide a theoritical justification of its better generalization than online KD by analyzing the Hessian Matrix of the loss function of student in both the scenarios by arguing that if the student model converges to a local minimizer at step t of SGD then expected value of trace of hessian matrix using Spaced KD is lesser.  Furthermore, they provide experimental results to verify their claim on the CIFAR-100, ImageNet-100 and Tiny-ImageNet datasets for both self and online KD where they show significant gains.", "review_text": "In this work, the authors perform a scheme on the lines of online and self KD to promote generalization of the model. They propose SpacedKD which is inspired from spacing effect in biological learning which highlights that study sessions across time is critical for memory formation. They argue that this spacing effect can be generalized to a kind of ensemble learning at the temporal scale. They implement this spaced KD in both online and self KD setups. For online KD, they just train the teacher in advanced for a certain number of steps termed as spaced interval. Then they just freeze this teacher and distill this knowledge to the student using the same data. They further provide a theoritical justification of its better generalization than online KD by analyzing the Hessian Matrix of the loss function of student in both the scenarios by arguing that if the student model converges to a local minimizer at step t of SGD then expected value of trace of hessian matrix using Spaced KD is lesser.  Furthermore, they provide experimental results to verify their claim on the CIFAR-100, ImageNet-100 and Tiny-ImageNet datasets for both self and online KD where they show significant gains.", "strengths": "The connection identified to the spacing effect in biological learning is quite interesting and provide intuitive backing for the method. The analysis of hessian of loss being lesser for spaced KD is also interesting and it being converging to a flatter local minima via SGD. \nThe results show prominent gains in performance which is again promising. Furthermore, the authors have also provided analysis of using different value of the interval various architectures along with the different architecture of teacher student pairs (inclduing different architectures for teacher). Also, the analysis in presence of various corruptions further backs the use case of their method.", "weaknesses": "Not exactly clear from the paper what is the exact difference between spaced and online KD setup. Also, why is the description of implementing spaced KD for self distillation not provided. I understand that it is pretty straightforward to adapt it but for completeness it should be there and there might be some smal changes that would be required into a naive adaptation that helps improving the result. \n\nRegarding results, even though the authors have mentioned in one of the figure captions the reason for not providing results on ViT architecture, I feel that it is not justified. To test its applicability it should be tested on state of the architectures and large scale datasets. The major results revolve around CIFAR-100, Tiny-ImageNet and ImageNet-100 architectures and only one result is provided in the ImageNet-1k (and that too with a ResNet-18 architecture) leading to speculations that it might not give this sort of gains on ImageNet-1k or other large scale settings. Also, I dont understand what the Appendix A.5 results want to convey.", "questions": "Please check the weakness section.  Appendix A.5 is confusing. Does it imply that distilling from a teacher trained s steps ahead is not useful? But then this is the method right that first you should update teacher for s steps and then using that distill to student for those s steps. Also it would have been very useful to provide a psuedo code comparison for online and spaced KD so as to what exactly is the differennce. Overall I feel the presentation of the paper could have been improved especially the emipirical section. \n\nAnyhow, right now I am going with the marginally below acceptance but I would be happy to reconsider my rating based on the justifications provided in the rebuttal.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors perform a scheme on the lines of online and self KD to promote generalization of the model. They propose SpacedKD which is inspired from spacing effect in biological learning which highlights that study sessions across time is critical for memory formation. They argue that this spacing effect can be generalized to a kind of ensemble learning at the temporal scale. They implement this spaced KD in both online and self KD setups. For online KD, they just train the teacher in advanced for a certain number of steps termed as spaced interval. Then they just freeze this teacher and distill this knowledge to the student using the same data. They further provide a theoritical justification of its better generalization than online KD by analyzing the Hessian Matrix of the loss function of student in both the scenarios by arguing that if the student model converges to a local minimizer at step t of SGD then expected value of trace of hessian matrix using Spaced KD is lesser.  Furthermore, they provide experimental results to verify their claim on the CIFAR-100, ImageNet-100 and Tiny-ImageNet datasets for both self and online KD where they show significant gains.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The connection identified to the spacing effect in biological learning is quite interesting and provide intuitive backing for the method. The analysis of hessian of loss being lesser for spaced KD is also interesting and it being converging to a flatter local minima via SGD. \nThe results show prominent gains in performance which is again promising. Furthermore, the authors have also provided analysis of using different value of the interval various architectures along with the different architecture of teacher student pairs (inclduing different architectures for teacher). Also, the analysis in presence of various corruptions further backs the use case of their method.", "weaknesses": "Not exactly clear from the paper what is the exact difference between spaced and online KD setup. Also, why is the description of implementing spaced KD for self distillation not provided. I understand that it is pretty straightforward to adapt it but for completeness it should be there and there might be some smal changes that would be required into a naive adaptation that helps improving the result. \n\nRegarding results, even though the authors have mentioned in one of the figure captions the reason for not providing results on ViT architecture, I feel that it is not justified. To test its applicability it should be tested on state of the architectures and large scale datasets. The major results revolve around CIFAR-100, Tiny-ImageNet and ImageNet-100 architectures and only one result is provided in the ImageNet-1k (and that too with a ResNet-18 architecture) leading to speculations that it might not give this sort of gains on ImageNet-1k or other large scale settings. Also, I dont understand what the Appendix A.5 results want to convey.", "questions": "Please check the weakness section.  Appendix A.5 is confusing. Does it imply that distilling from a teacher trained s steps ahead is not useful? But then this is the method right that first you should update teacher for s steps and then using that distill to student for those s steps. Also it would have been very useful to provide a psuedo code comparison for online and spaced KD so as to what exactly is the differennce. Overall I feel the presentation of the paper could have been improved especially the emipirical section. \n\nAnyhow, right now I am going with the marginally below acceptance but I would be happy to reconsider my rating based on the justifications provided in the rebuttal.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730758037356}, {"id": "Rh17CeEJYC", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2884/Reviewer_5J6i"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces a new method for online knowledge distillation that uses the teacher model trained ahead of the student at each epoch, resulting in a flatter loss landscape.", "review_text": "The paper introduces a new method for online knowledge distillation that uses the teacher model trained ahead of the student at each epoch, resulting in a flatter loss landscape.", "strengths": "The paper is well-written and easy to follow.\n\nThe experiments show the effective improvement of the proposed method over vanilla KD.", "weaknesses": "1) I did not get the connection between the proposed method and the spacing effect in biological learning mentioned in Section 3.2. The teacher has different weights from the student. Why does training for several additional epochs cause a spacing effect?\n\n2) The theoretical part is not robust. The proposed method assumes that the teacher trained ahead is close to the local minimizer $\\phi^*$, so why not use a well-trained model instead?\n\n3) There is a lack of comparison with other KD methods like [1,2]\n\n[1] Self-Distillation from the Last Mini-Batch for Consistency Regularization.\n\n[2] Self-Knowledge Distillation with Progressive Refinement of Targets.", "questions": "1) What are the numerical values of x and y in Figure 2a?\n\n2) As 0.5 epochs mean training on half of the dataset, would this cause a bias towards that half of the dataset? Is the training dataset shuffled?\n\n3) How can the proposed method be applied to self-distillation in [3], where the teacher is the deepest layer of models and the students are the shallow layers?\n\n4) What dataset was used for Figure 3, and which settings were used for the experiments?\n\n\n\n[3] Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a new method for online knowledge distillation that uses the teacher model trained ahead of the student at each epoch, resulting in a flatter loss landscape.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper is well-written and easy to follow.\n\nThe experiments show the effective improvement of the proposed method over vanilla KD.", "weaknesses": "1) I did not get the connection between the proposed method and the spacing effect in biological learning mentioned in Section 3.2. The teacher has different weights from the student. Why does training for several additional epochs cause a spacing effect?\n\n2) The theoretical part is not robust. The proposed method assumes that the teacher trained ahead is close to the local minimizer $\\phi^*$, so why not use a well-trained model instead?\n\n3) There is a lack of comparison with other KD methods like [1,2]\n\n[1] Self-Distillation from the Last Mini-Batch for Consistency Regularization.\n\n[2] Self-Knowledge Distillation with Progressive Refinement of Targets.", "questions": "1) What are the numerical values of x and y in Figure 2a?\n\n2) As 0.5 epochs mean training on half of the dataset, would this cause a bias towards that half of the dataset? Is the training dataset shuffled?\n\n3) How can the proposed method be applied to self-distillation in [3], where the teacher is the deepest layer of models and the students are the shallow layers?\n\n4) What dataset was used for Figure 3, and which settings were used for the experiments?\n\n\n\n[3] Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730661962591}, {"id": "iYwCukCYS7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2884/Reviewer_bFLi"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper studies a variant of Knowledge Distillation, designed with Space Interval learning, inspired by the spacing effects observed in biological learning. \n\nDesign: The Space Interval method is applied to Online KD, where the teacher model is trained a few intervals (s iterations) ahead of the student. The teacher’s weights are then frozen and used to train the student, who lags behind. The hypothesis is that, by updating the teacher less frequently, it provides more stable guidance to the student. \n\nTheoretical Analysis: The authors use the trace of the Hessian matrix as a measure of the loss landscape’s flatness, where a smaller Hessian trace indicates a flatter, and thus potentially more generalizable, landscape. They show that the loss landscape of the new paradigm has flatter Hessian trace ompared to the traditional Online KD, suggesting improved generalization. \n\nExperimental Results: The proposed method outperforms the traditional Online KD on CIFAR-100, Tiny-ImageNet, ImageNet-100, and ImageNet-1K datasets by approximately 1–2% on average with three ResNet variants. The authors also provide experiments demonstrating robustness improvements over online and self-KD counterparts, along with ablation studies on learning rate, batch size, and loss functions.", "review_text": "This paper studies a variant of Knowledge Distillation, designed with Space Interval learning, inspired by the spacing effects observed in biological learning. \n\nDesign: The Space Interval method is applied to Online KD, where the teacher model is trained a few intervals (s iterations) ahead of the student. The teacher’s weights are then frozen and used to train the student, who lags behind. The hypothesis is that, by updating the teacher less frequently, it provides more stable guidance to the student. \n\nTheoretical Analysis: The authors use the trace of the Hessian matrix as a measure of the loss landscape’s flatness, where a smaller Hessian trace indicates a flatter, and thus potentially more generalizable, landscape. They show that the loss landscape of the new paradigm has flatter Hessian trace ompared to the traditional Online KD, suggesting improved generalization. \n\nExperimental Results: The proposed method outperforms the traditional Online KD on CIFAR-100, Tiny-ImageNet, ImageNet-100, and ImageNet-1K datasets by approximately 1–2% on average with three ResNet variants. The authors also provide experiments demonstrating robustness improvements over online and self-KD counterparts, along with ablation studies on learning rate, batch size, and loss functions.", "strengths": "Overall, the paper is easy to follow. The idea seems simple to implement and appears interesting as inspired by biological learning processes. Empirically, it achieves a few percentage points of improvement over existing online methods, as demonstrated in the experimental sections.", "weaknesses": "The results are overall encouraging, there are few points to be clarified with the authors in term of bio-inspiration, the generality of the method, the settings used across experiments and the seemingly conflicts between theoretical and empirical results as in Questions section.", "questions": "1. While the bio-inspiration is interesting, the learning mechanism applied here differs from the spacing effects observed in biology. In biological learning, spacing effects enhance learning and memory retention of a subject when study sessions or practice trials are spread across intervals of time. However, in this application, the teacher is trained at intervals while the student learns continuously which seems the design is not well-aligned with the biological inspiration. Could the authors provide deeper details on this point to show the bio-inspiration is valid.\n\n2. In the online setting, most experiments use a teacher with the same architecture as the student. To demonstrate the generality of the method, could the authors conducting experiments where the teacher’s architecture is significantly larger than the student's?\n\n3. As shown in Fig. 3, the period of applying Space KD is critical. Are all experiments using the best setting or different experiments has different periods? Could the authors provide details on the optimal settings used across the experiments in the paper and explain how these settings were selected?\n\n4. Is there a conflict between the theoretical analysis and the empirical results? While the theoretical analysis is independent of the period during which Space KD is applied, the empirical results indicate that it is only useful in the last 10 epochs of training? Could the authors provides also the only online KD baseline's results in Fig. 3 for the comparision?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies a variant of Knowledge Distillation, designed with Space Interval learning, inspired by the spacing effects observed in biological learning. \n\nDesign: The Space Interval method is applied to Online KD, where the teacher model is trained a few intervals (s iterations) ahead of the student. The teacher’s weights are then frozen and used to train the student, who lags behind. The hypothesis is that, by updating the teacher less frequently, it provides more stable guidance to the student. \n\nTheoretical Analysis: The authors use the trace of the Hessian matrix as a measure of the loss landscape’s flatness, where a smaller Hessian trace indicates a flatter, and thus potentially more generalizable, landscape. They show that the loss landscape of the new paradigm has flatter Hessian trace ompared to the traditional Online KD, suggesting improved generalization. \n\nExperimental Results: The proposed method outperforms the traditional Online KD on CIFAR-100, Tiny-ImageNet, ImageNet-100, and ImageNet-1K datasets by approximately 1–2% on average with three ResNet variants. The authors also provide experiments demonstrating robustness improvements over online and self-KD counterparts, along with ablation studies on learning rate, batch size, and loss functions.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Overall, the paper is easy to follow. The idea seems simple to implement and appears interesting as inspired by biological learning processes. Empirically, it achieves a few percentage points of improvement over existing online methods, as demonstrated in the experimental sections.", "weaknesses": "The results are overall encouraging, there are few points to be clarified with the authors in term of bio-inspiration, the generality of the method, the settings used across experiments and the seemingly conflicts between theoretical and empirical results as in Questions section.", "questions": "1. While the bio-inspiration is interesting, the learning mechanism applied here differs from the spacing effects observed in biology. In biological learning, spacing effects enhance learning and memory retention of a subject when study sessions or practice trials are spread across intervals of time. However, in this application, the teacher is trained at intervals while the student learns continuously which seems the design is not well-aligned with the biological inspiration. Could the authors provide deeper details on this point to show the bio-inspiration is valid.\n\n2. In the online setting, most experiments use a teacher with the same architecture as the student. To demonstrate the generality of the method, could the authors conducting experiments where the teacher’s architecture is significantly larger than the student's?\n\n3. As shown in Fig. 3, the period of applying Space KD is critical. Are all experiments using the best setting or different experiments has different periods? Could the authors provide details on the optimal settings used across the experiments in the paper and explain how these settings were selected?\n\n4. Is there a conflict between the theoretical analysis and the empirical results? While the theoretical analysis is independent of the period during which Space KD is applied, the empirical results indicate that it is only useful in the last 10 epochs of training? Could the authors provides also the only online KD baseline's results in Fig. 3 for the comparision?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "n/a", "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730631547411}, {"id": "H643vmA4PD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2884/Reviewer_v8iU"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces Spaced Knowledge Distillation (Spaced KD), a bio-inspired approach to knowledge distillation that incorporates spacing intervals in the training process to enhance model generalization. Inspired by the \"spacing effect\" observed in biological learning, Spaced KD involves training a teacher model a few steps ahead of the student model and periodically transferring knowledge at intervals. This technique allows the student model to find flatter minima in the loss landscape, which leads to better generalization and improved robustness to noise. Extensive experiments demonstrate that Spaced KD yields superior performance across multiple datasets and architectures compared to traditional KD methods without additional training costs.", "review_text": "This paper introduces Spaced Knowledge Distillation (Spaced KD), a bio-inspired approach to knowledge distillation that incorporates spacing intervals in the training process to enhance model generalization. Inspired by the \"spacing effect\" observed in biological learning, Spaced KD involves training a teacher model a few steps ahead of the student model and periodically transferring knowledge at intervals. This technique allows the student model to find flatter minima in the loss landscape, which leads to better generalization and improved robustness to noise. Extensive experiments demonstrate that Spaced KD yields superior performance across multiple datasets and architectures compared to traditional KD methods without additional training costs.", "strengths": "1) Novelty: The paper introduces a fresh approach by integrating the bio-inspired “spacing effect” into knowledge distillation (KD), which is relatively unexplored in existing literature.\n\n2) Strong Theoretical Foundation: The authors provide a well-structured theoretical analysis linking Spaced KD to flat minima, supporting claims about enhanced generalization. The use of Hessian trace analysis to demonstrate why Spaced KD leads to better generalization adds rigor and robustness to the claims.\n\n3) Comprehensive Experiments: The experimental results are extensive, spanning multiple datasets (CIFAR-100, Tiny-ImageNet, ImageNet-100, and ImageNet-1K) and network architectures (ResNet variants, DeiT-Tiny, PiT-Tiny). These experiments show consistent improvements with Spaced KD, reinforcing the method’s general applicability.\n\n4) Practicality: Spaced KD is straightforward to implement, making it accessible for a broad range of applications without requiring additional training time or major architectural modifications. This ease of integration enhances its appeal for both academia and industry.\n\n5) Robustness and Generalization Gains: Spaced KD demonstrates superior robustness to noise and out-of-distribution data, an increasingly valued property in modern AI. This robustness adds significant value, especially for real-world applications where models encounter varied data distributions.", "weaknesses": "1) Incremental Novelty: While Spaced KD introduces a creative application of the spacing effect to knowledge distillation, it lacks a fundamentally new mechanism within the KD framework. This incremental contribution may limit its overall impact, as the method primarily builds on existing KD techniques without significantly advancing the underlying theory.\n\n2) Interval Sensitivity and Adaptability: The success of Spaced KD hinges on selecting an effective interval for knowledge transfer. However, the paper does not explore adaptive strategies for dynamically tuning this interval, which may constrain Spaced KD’s versatility across diverse tasks and datasets with varying training schedules.\n\n3) Applicability Across Model Sizes and Architectures: Spaced KD is demonstrated only with identical teacher and student models of the same size, without addressing model compression. It remains uncertain how this approach would perform if applied to different model sizes or architectures, such as a transformer teacher model paired with a CNN student. The potential to generalize Spaced KD for cross-architecture or cross-size distillation remains unexplored.", "questions": "1) Applicability Across Model Sizes and Architectures: How does Spaced KD perform when the teacher and student differ in size or architecture? Specifically, can it support scenarios where the teacher is larger or has a different architecture (e.g., a transformer teacher with a CNN student), or does it rely on identical model sizes?\n\n2) Comparison with Gradient-based Methods: How does Spaced KD differ from traditional gradient-based methods where gradients are updated after each batch? Given that Spaced KD introduces intervals between teacher and student updates, is this approach essentially a modified gradient update, or does it provide additional benefits that distinguish it from standard gradient optimization?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Spaced Knowledge Distillation (Spaced KD), a bio-inspired approach to knowledge distillation that incorporates spacing intervals in the training process to enhance model generalization. Inspired by the \"spacing effect\" observed in biological learning, Spaced KD involves training a teacher model a few steps ahead of the student model and periodically transferring knowledge at intervals. This technique allows the student model to find flatter minima in the loss landscape, which leads to better generalization and improved robustness to noise. Extensive experiments demonstrate that Spaced KD yields superior performance across multiple datasets and architectures compared to traditional KD methods without additional training costs.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1) Novelty: The paper introduces a fresh approach by integrating the bio-inspired “spacing effect” into knowledge distillation (KD), which is relatively unexplored in existing literature.\n\n2) Strong Theoretical Foundation: The authors provide a well-structured theoretical analysis linking Spaced KD to flat minima, supporting claims about enhanced generalization. The use of Hessian trace analysis to demonstrate why Spaced KD leads to better generalization adds rigor and robustness to the claims.\n\n3) Comprehensive Experiments: The experimental results are extensive, spanning multiple datasets (CIFAR-100, Tiny-ImageNet, ImageNet-100, and ImageNet-1K) and network architectures (ResNet variants, DeiT-Tiny, PiT-Tiny). These experiments show consistent improvements with Spaced KD, reinforcing the method’s general applicability.\n\n4) Practicality: Spaced KD is straightforward to implement, making it accessible for a broad range of applications without requiring additional training time or major architectural modifications. This ease of integration enhances its appeal for both academia and industry.\n\n5) Robustness and Generalization Gains: Spaced KD demonstrates superior robustness to noise and out-of-distribution data, an increasingly valued property in modern AI. This robustness adds significant value, especially for real-world applications where models encounter varied data distributions.", "weaknesses": "1) Incremental Novelty: While Spaced KD introduces a creative application of the spacing effect to knowledge distillation, it lacks a fundamentally new mechanism within the KD framework. This incremental contribution may limit its overall impact, as the method primarily builds on existing KD techniques without significantly advancing the underlying theory.\n\n2) Interval Sensitivity and Adaptability: The success of Spaced KD hinges on selecting an effective interval for knowledge transfer. However, the paper does not explore adaptive strategies for dynamically tuning this interval, which may constrain Spaced KD’s versatility across diverse tasks and datasets with varying training schedules.\n\n3) Applicability Across Model Sizes and Architectures: Spaced KD is demonstrated only with identical teacher and student models of the same size, without addressing model compression. It remains uncertain how this approach would perform if applied to different model sizes or architectures, such as a transformer teacher model paired with a CNN student. The potential to generalize Spaced KD for cross-architecture or cross-size distillation remains unexplored.", "questions": "1) Applicability Across Model Sizes and Architectures: How does Spaced KD perform when the teacher and student differ in size or architecture? Specifically, can it support scenarios where the teacher is larger or has a different architecture (e.g., a transformer teacher with a CNN student), or does it rely on identical model sizes?\n\n2) Comparison with Gradient-based Methods: How does Spaced KD differ from traditional gradient-based methods where gradients are updated after each batch? Given that Spaced KD introduces intervals between teacher and student updates, is this approach essentially a modified gradient update, or does it provide additional benefits that distinguish it from standard gradient optimization?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730248145814}], "openreview_url": "https://openreview.net/forum?id=YUaQ6qrbRL", "arxiv_id": "2502.06192", "paper_pdf": "papers/YUaQ6qrbRL.pdf", "paper_pdf_sha256": "eff8106c1b9a890e1c0f9c069827106292f52e907e5ad1f09d0f100b547fd316", "paper_pdf_bytes": 2573459, "paper_pdf_source": "openreview", "code_url": "https://github.com/SunGL001/Spaced-KD", "code_repository": "SunGL001/Spaced-KD", "code_commit": "deecb4da498f1bc1ded450d3129c86fc46def40d", "code_archive": "repos/YUaQ6qrbRL.zip", "code_archive_sha256": "5ec914b86a2a663ddeb1711971c141e5169ef92352d716d6dc1672fbae9b5d11", "code_archive_bytes": 287559, "code_file_count": 77, "code_extensions": {".py": 75, ".sh": 2}, "github_disk_usage_kb": 300, "github_languages": {"Python": 460121, "Shell": 1157}, "github_archived": false, "github_pushed_at": "2025-05-20T05:12:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/right-time-to-learn-promoting-generalization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3mdCet7vVv", "year": 2024, "status": "rejected", "title": "Maestro: Uncovering Low-Rank Structures via Trainable Decomposition", "authors": ["Samuel Horváth", "Stefanos Laskaridis", "Shashank Rajput", "Hongyi Wang"], "authorids": ["~Samuel_Horváth1", "~Stefanos_Laskaridis1", "~Shashank_Rajput1", "~Hongyi_Wang1"], "authors_source": "OpenReview API", "abstract": "Deep Neural Networks (DNNs) have been a large drivers and enablers for AI breakthroughs in recent years. These models have been getting larger in their attempt to become more accurate and tackle new upcoming use-cases, including AR/VR and intelligent assistants. However, the training process of such large models is a costly and time-consuming process, which typically yields a single model to fit all targets. To mitigate this, various techniques have been proposed in the literature, including pruning, sparsification, or quantization of the model weights and updates. While able to achieve high compression rates, they often incur computational overheads or accuracy penalties. Alternatively, factorization methods have been leveraged to incorporate low-rank compression in the training process. Such techniques (e.g., SVD) also frequently rely on the computationally expensive decomposition of layers and are potentially sub-optimal for non-linear models, such as DNNs. In this work, we take a further step in designing efficient low-rank models and propose MAESTRO, a framework for trainable low-rank layers. Instead of regularly applying a priori decompositions such as SVD, the low-rank structure is built into the training process through a generalized variant of Ordered Dropout. This method imposes an importance ordering via sampling on the decomposed DNN structure. Our theoretical analysis demonstrates that our method recovers the SVD decomposition of linear mapping on uniformly distributed data and PCA for linear autoencoders. We further apply our technique on DNNs and empirically illustrate that MAESTRO enables the extraction of lower footprint models that preserve model performance while allowing for graceful accuracy-latency tradeoffs for the deployment to devices of different capabilities.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "hi0mIdcLFJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4088/Reviewer_ZdCx"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes MAESTRO, which is a trainable low-rank approximation technique for deep neural networks. It proposes a progressive shrinking approach that decomposes the weights of each layer into low-rank components using an extended version of Ordered Dropout. This allows for efficient compression and trade-off between model size and accuracy. The method is evaluated on various models, datasets, and modalities, showing superior performance compared to other compression methods.", "review_text": "The paper proposes MAESTRO, which is a trainable low-rank approximation technique for deep neural networks. It proposes a progressive shrinking approach that decomposes the weights of each layer into low-rank components using an extended version of Ordered Dropout. This allows for efficient compression and trade-off between model size and accuracy. The method is evaluated on various models, datasets, and modalities, showing superior performance compared to other compression methods.", "strengths": "- The paper extends the Ordered Dropout technique to handle non-uniformity in the search space by allowing different ranks per layer. \n\n- It introduces a trainable aspect to the decomposition, which enables the model to reflect the data distribution. \n\n- It provides a latency-accuracy trade-off mechanism for deploying the network on constrained devices.", "weaknesses": "- The citation style seems not correct. It should include the author's names in place of numerical references.\n\n- Why the method named after \"Maestro\"? It is never introduced and seems weird to me.\n\n- The proposed technique appears as a logical improvement from Ordered Dropout. Its effectiveness, however, is primarily demonstrated through toy architectures and datasets, such as ResNet18 and Cifar10. For the method to gain practical and impactful validation, I recommend conducting additional experiments on more complex datasets like ImageNet to substantiate its superiority.\n\n- Building on the previous point, there are alternative methods that report better accuracy with more compact architectures. For instance, the OTOv2 framework:\n\nChen, Tianyi, et al. \"Only train once: A one-shot neural network training and pruning framework.\" Advances in Neural Information Processing Systems 34 (2021): 19637-19651.\n\nIt structurally prunes the model during training (hence still training efficient), and it achieves a 93.3% accuracy with only 0.55M parameters on Cifar10 using VGG16. This is in contrast to the 93.10% accuracy with 2.20M parameters reported by the proposed method. This comparison casts doubt on the practical utility and the advantages of the low-rank based method presented.", "questions": "See the weaknesses part above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes MAESTRO, which is a trainable low-rank approximation technique for deep neural networks. It proposes a progressive shrinking approach that decomposes the weights of each layer into low-rank components using an extended version of Ordered Dropout. This allows for efficient compression and trade-off between model size and accuracy. The method is evaluated on various models, datasets, and modalities, showing superior performance compared to other compression methods.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The paper extends the Ordered Dropout technique to handle non-uniformity in the search space by allowing different ranks per layer. \n\n- It introduces a trainable aspect to the decomposition, which enables the model to reflect the data distribution. \n\n- It provides a latency-accuracy trade-off mechanism for deploying the network on constrained devices.", "weaknesses": "- The citation style seems not correct. It should include the author's names in place of numerical references.\n\n- Why the method named after \"Maestro\"? It is never introduced and seems weird to me.\n\n- The proposed technique appears as a logical improvement from Ordered Dropout. Its effectiveness, however, is primarily demonstrated through toy architectures and datasets, such as ResNet18 and Cifar10. For the method to gain practical and impactful validation, I recommend conducting additional experiments on more complex datasets like ImageNet to substantiate its superiority.\n\n- Building on the previous point, there are alternative methods that report better accuracy with more compact architectures. For instance, the OTOv2 framework:\n\nChen, Tianyi, et al. \"Only train once: A one-shot neural network training and pruning framework.\" Advances in Neural Information Processing Systems 34 (2021): 19637-19651.\n\nIt structurally prunes the model during training (hence still training efficient), and it achieves a 93.3% accuracy with only 0.55M parameters on Cifar10 using VGG16. This is in contrast to the 93.10% accuracy with 2.20M parameters reported by the proposed method. This comparison casts doubt on the practical utility and the advantages of the low-rank based method presented.", "questions": "See the weaknesses part above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699077754729}, {"id": "ML8evsHc8c", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4088/Reviewer_Mad7"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a low-rank compression scheme for deep neural networks, which factorizes fully connected, convolutional, and attention layers in the form A=UV, and progressively reduces the rank of the U and V matrices. For convolutional layers the factorization is applied to the unrolled 2D matrix, while for attention layers it is applied to the Q, K, V matrices. They use ordered dropout and hierarchical group-lasso to facilitate the reduction of the rank of U/V matrices.", "review_text": "This paper proposes a low-rank compression scheme for deep neural networks, which factorizes fully connected, convolutional, and attention layers in the form A=UV, and progressively reduces the rank of the U and V matrices. For convolutional layers the factorization is applied to the unrolled 2D matrix, while for attention layers it is applied to the Q, K, V matrices. They use ordered dropout and hierarchical group-lasso to facilitate the reduction of the rank of U/V matrices.", "strengths": "Unlike unstructured pruning methods, low-rank compression can preserve the dense structure of matrices, which can extract more performance from GPUs. For the training of transformers on the Multi30k dataset shown in Table 3, the proposed method is able to reduce the number of parameters by more than half compared to the baseline (Pufferfish), while also reducing the perplexity.", "weaknesses": "Low-rank compression and Lasso have been around for a very long time, and the only novelty seems to be the use of ordered dropout. The improvement over existing methods is marginal for the experiments with CNNs. The proposed method is obviously very sensitive to the choice of the Lasso coefficient lambda, but there is no theory behind how it can be chosen effectively.", "questions": "How is the initial factorized mapping performed without SVD? How is the initial maximal rank r chosen?\n\nHow does the proposed method compare with other structured pruning methods?\n\nTypos\np.4 “multi-head attention (HMA)” > “multi-head attention (MHA)”\np.5 “we one could leverage” > “one could leverage”\np.5 “Singular Value Decomposition (SVD)” Why define this here when it has been repeatedly used in previous sections?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a low-rank compression scheme for deep neural networks, which factorizes fully connected, convolutional, and attention layers in the form A=UV, and progressively reduces the rank of the U and V matrices. For convolutional layers the factorization is applied to the unrolled 2D matrix, while for attention layers it is applied to the Q, K, V matrices. They use ordered dropout and hierarchical group-lasso to facilitate the reduction of the rank of U/V matrices.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "1 poor", "strengths": "Unlike unstructured pruning methods, low-rank compression can preserve the dense structure of matrices, which can extract more performance from GPUs. For the training of transformers on the Multi30k dataset shown in Table 3, the proposed method is able to reduce the number of parameters by more than half compared to the baseline (Pufferfish), while also reducing the perplexity.", "weaknesses": "Low-rank compression and Lasso have been around for a very long time, and the only novelty seems to be the use of ordered dropout. The improvement over existing methods is marginal for the experiments with CNNs. The proposed method is obviously very sensitive to the choice of the Lasso coefficient lambda, but there is no theory behind how it can be chosen effectively.", "questions": "How is the initial factorized mapping performed without SVD? How is the initial maximal rank r chosen?\n\nHow does the proposed method compare with other structured pruning methods?\n\nTypos\np.4 “multi-head attention (HMA)” > “multi-head attention (MHA)”\np.5 “we one could leverage” > “one could leverage”\np.5 “Singular Value Decomposition (SVD)” Why define this here when it has been repeatedly used in previous sections?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698820856397}, {"id": "hoNfoK0O6U", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4088/Reviewer_CAHp"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work mainly focuses on incorporating trainable low-rank layer decompositions in deep-learning models. The authors propose MAESTRO, which progressively finds the optimal rank of each layer during the training by imposing importance ordering via the existing Ordered Dropout technique. The redundant ranks are zeroed out by using the hierarchical group lasso term as the regularizer in the loss function. MAESTRO accounts for data distributions and the target function rather than applying SVD on pre-learned model weights.", "review_text": "This work mainly focuses on incorporating trainable low-rank layer decompositions in deep-learning models. The authors propose MAESTRO, which progressively finds the optimal rank of each layer during the training by imposing importance ordering via the existing Ordered Dropout technique. The redundant ranks are zeroed out by using the hierarchical group lasso term as the regularizer in the loss function. MAESTRO accounts for data distributions and the target function rather than applying SVD on pre-learned model weights.", "strengths": "The novelty of the work lies in applying the existing Ordered Dropout technique from Federated Learning (FjORD) to optimally order the heterogeneous ranks of various layers in DNNs based on importance criterion, which results in discovering layer-wise low-rank decompositions. In contrast to uniform dropout across the width in each layer ( FjORD), MAESTRO independently decomposes each layer to uncover optimal rank. The authors provide applications of MAESTRO to various layer types in CNNs, FC, and Transformers. \n\nThe paper is easy to understand and is well-structured. The experiments are comprehensive and justify theoretical insights.", "weaknesses": "1. The paper suffers from typos. The authors are encouraged to review and proofread the draft.\n- Page 1: …*find progressively*…\n- Page 2: …*novelly fuse*…\n- Page 3: ..*have been proposed*… (multiple instances)\n- Page 4: …*HMA*….\n- Page 5: ….*orthoghonal*….\n\n2. It is recommended that authors explore a better illustration for Figure 1. For instance, there is not much difference visually in Factorized mapping and Ordered Representation when printed in black/white. It might be helpful to provide a better illustration for the Ordered Dropout process (it is challenging to understand it with symbols without any reference in the figure caption. In current form, it is assumed that the readers will be familiar with OD). Since MAESTRO provides layer-wise decomposition and is generally applicable to various DNN layers, it might be useful to incorporate the various layer types of the DNN network (Sec 3.2) in Figure 1 as an overall summary of the proposed work and its applicability.", "questions": "Suggestions are provided in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work mainly focuses on incorporating trainable low-rank layer decompositions in deep-learning models. The authors propose MAESTRO, which progressively finds the optimal rank of each layer during the training by imposing importance ordering via the existing Ordered Dropout technique. The redundant ranks are zeroed out by using the hierarchical group lasso term as the regularizer in the loss function. MAESTRO accounts for data distributions and the target function rather than applying SVD on pre-learned model weights.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "The novelty of the work lies in applying the existing Ordered Dropout technique from Federated Learning (FjORD) to optimally order the heterogeneous ranks of various layers in DNNs based on importance criterion, which results in discovering layer-wise low-rank decompositions. In contrast to uniform dropout across the width in each layer ( FjORD), MAESTRO independently decomposes each layer to uncover optimal rank. The authors provide applications of MAESTRO to various layer types in CNNs, FC, and Transformers. \n\nThe paper is easy to understand and is well-structured. The experiments are comprehensive and justify theoretical insights.", "weaknesses": "1. The paper suffers from typos. The authors are encouraged to review and proofread the draft.\n- Page 1: …*find progressively*…\n- Page 2: …*novelly fuse*…\n- Page 3: ..*have been proposed*… (multiple instances)\n- Page 4: …*HMA*….\n- Page 5: ….*orthoghonal*….\n\n2. It is recommended that authors explore a better illustration for Figure 1. For instance, there is not much difference visually in Factorized mapping and Ordered Representation when printed in black/white. It might be helpful to provide a better illustration for the Ordered Dropout process (it is challenging to understand it with symbols without any reference in the figure caption. In current form, it is assumed that the readers will be familiar with OD). Since MAESTRO provides layer-wise decomposition and is generally applicable to various DNN layers, it might be useful to incorporate the various layer types of the DNN network (Sec 3.2) in Figure 1 as an overall summary of the proposed work and its applicability.", "questions": "Suggestions are provided in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698807229700}, {"id": "KPLklSOFhK", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4088/Reviewer_LrqS"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors introduce Maestro, a technique designed for efficient layer-wise low-rank factorization during training. This method incorporates an ordered drop strategy combined with group lasso regularization, encouraging the progressive adoption of lower-rank weights during training. The evaluation is conducted on CIFAR10, MNIST, and Multi-30k, comparing Maestro against various low-rank approaches and several pruning and quantization techniques. Furthermore, the paper offers multiple ablation studies and provides theoretical analysis for specific problems.", "review_text": "The authors introduce Maestro, a technique designed for efficient layer-wise low-rank factorization during training. This method incorporates an ordered drop strategy combined with group lasso regularization, encouraging the progressive adoption of lower-rank weights during training. The evaluation is conducted on CIFAR10, MNIST, and Multi-30k, comparing Maestro against various low-rank approaches and several pruning and quantization techniques. Furthermore, the paper offers multiple ablation studies and provides theoretical analysis for specific problems.", "strengths": "1. The paper is easy to follow. \n2. The theoretical properties are sound with the proposed method.\n3. The algorithm seems reasonable.", "weaknesses": "The algorithm seems reasonable to me. However, for the experiments, ImageNet results are missing. As an important benchmark, ImageNet is often used to compare performance between the compression-related tasks. For instance, Cutterfish presented their ResNet-50 results using the ImageNet dataset. To highlight effectiveness, it would be beneficial to include evaluations based on the ImageNet dataset. Additionally, tests on larger models would enhance the comprehensiveness of the study.\n\nIs the #GMACs the training cost? If not, please show the training cost.", "questions": "How does the proposed method perform on ViT and other larger models using the ImageNet dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce Maestro, a technique designed for efficient layer-wise low-rank factorization during training. This method incorporates an ordered drop strategy combined with group lasso regularization, encouraging the progressive adoption of lower-rank weights during training. The evaluation is conducted on CIFAR10, MNIST, and Multi-30k, comparing Maestro against various low-rank approaches and several pruning and quantization techniques. Furthermore, the paper offers multiple ablation studies and provides theoretical analysis for specific problems.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The paper is easy to follow. \n2. The theoretical properties are sound with the proposed method.\n3. The algorithm seems reasonable.", "weaknesses": "The algorithm seems reasonable to me. However, for the experiments, ImageNet results are missing. As an important benchmark, ImageNet is often used to compare performance between the compression-related tasks. For instance, Cutterfish presented their ResNet-50 results using the ImageNet dataset. To highlight effectiveness, it would be beneficial to include evaluations based on the ImageNet dataset. Additionally, tests on larger models would enhance the comprehensiveness of the study.\n\nIs the #GMACs the training cost? If not, please show the training cost.", "questions": "How does the proposed method perform on ViT and other larger models using the ImageNet dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698717511401}], "openreview_url": "https://openreview.net/forum?id=3mdCet7vVv", "arxiv_id": "2308.14929", "paper_pdf": "papers/3mdCet7vVv.pdf", "paper_pdf_sha256": "39eea620e9f5a71b233a1140d659e16b28b60ba6990b229df7447172dcd6d455", "paper_pdf_bytes": 2326242, "paper_pdf_source": "openreview", "code_url": "https://github.com/SamuelHorvath/Maestro-LoD", "code_repository": "SamuelHorvath/Maestro-LoD", "code_commit": "0003014a94f78a63e48d971b773b40b00277a207", "code_archive": "repos/3mdCet7vVv.zip", "code_archive_sha256": "ad1521c7df926e61f3e2b2f22c590ea4ce5471bf2ab485010b1f92ffa37aabf7", "code_archive_bytes": 262308, "code_file_count": 26, "code_extensions": {".py": 23, ".sh": 2, ".ipynb": 1}, "github_disk_usage_kb": 247, "github_languages": {"Jupyter Notebook": 304265, "Python": 135895, "Shell": 1019}, "github_archived": false, "github_pushed_at": "2024-06-14T17:21:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/maestro-uncovering-low-rank-structures-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rMbrVNxYuqZ", "year": 2023, "status": "rejected", "title": "EurNet: Efficient Multi-Range Relational Modeling of Spatial Multi-Relational Data", "authors": ["Minghao Xu", "Yuanfan Guo", "Yi Xu", "Jian Tang", "Xinlei Chen", "Yuandong Tian"], "authorids": ["~Minghao_Xu1", "~Yuanfan_Guo1", "~Yi_Xu2", "~Jian_Tang1", "~Xinlei_Chen1", "~Yuandong_Tian1"], "authors_source": "OpenReview API", "abstract": "Modeling spatial relationship in the data remains critical across many different tasks, such as image classification, semantic segmentation and protein structure understanding. Previous works often use a unified solution like relative positional encoding. However, there exists different kinds of spatial relations, including short-range, medium-range and long-range relations, and modeling them separately can better capture the focus of different tasks on the multi-range relations (e.g., short-range relations can be important in instance segmentation, while long-range relations should be upweighted for semantic segmentation). In this work, we introduce the EurNet for Efficient multi-range relational modeling. EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions. In the constructed graph, EurNet adopts a novel modeling layer, called gated relational message passing (GRMP), to propagate multi-relational information across the data. GRMP captures multiple relations within the data with little extra computational cost. We study EurNets in two important domains for image and protein structure modeling. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation verify the gains of EurNet over the previous SoTA FocalNet. On the EC and GO protein function prediction benchmarks, EurNet consistently surpasses the previous SoTA GearNet. Our results demonstrate the strength of EurNets on modeling spatial multi-relational data from various domains. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "K5HtuHTD2k2", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1169/Reviewer_nyUd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors introduce a novel layer, called Gated Relational Message Passing (GRMP), for efficient multi-relational modelling over graphs. Importantly, GRMP scales better than existing multi-relational models when increasing the number of considered relations.\n\nThe authors use GRMP as building block for Efficient multi-range relational graph neural networks (EurNet) that they apply to multi-scale modelling problems. Across image classification, object detection, image segmentation and protein function prediction tasks, the authors show that EurNet performs comparably or better than state-of-the-art methods.", "review_text": "This is an important contribution to multi-relational modelling and graph neural networks. The paper is clearly written, experiments are extensive and detailed supplementary info is provided.", "strengths": "Strengths:\n* Contributions are clear and important\n* Claims are well supported; experiments are extensive (compared with multiple methods, using different model capacities, and considered different datasets / tasks)\n* The paper is well written. Figures are clarifying.\n\nWeakness:\n* This is a minor point: Maybe the authors could more explicitly showcase the empirical advantages of GRMP speedup in an applied setting when comparing with graph-based multi-relational models. (this is somehow already discussed in the “Throughput analysis.” paragraph, but the authors could make it more evident)", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the authors introduce a novel layer, called Gated Relational Message Passing (GRMP), for efficient multi-relational modelling over graphs. Importantly, GRMP scales better than existing multi-relational models when increasing the number of considered relations.\n\nThe authors use GRMP as building block for Efficient multi-range relational graph neural networks (EurNet) that they apply to multi-scale modelling problems. Across image classification, object detection, image segmentation and protein function prediction tasks, the authors show that EurNet performs comparably or better than state-of-the-art methods.", "strength_and_weaknesses": "Strengths:\n* Contributions are clear and important\n* Claims are well supported; experiments are extensive (compared with multiple methods, using different model capacities, and considered different datasets / tasks)\n* The paper is well written. Figures are clarifying.\n\nWeakness:\n* This is a minor point: Maybe the authors could more explicitly showcase the empirical advantages of GRMP speedup in an applied setting when comparing with graph-based multi-relational models. (this is somehow already discussed in the “Throughput analysis.” paragraph, but the authors could make it more evident)", "clarity,_quality,_novelty_and_reproducibility": "* Clarity: clearly written\n* Novelty: Novel. New layer for multi-relational modelling over graphs at scale.\n* Reproducibility: Good. Hyperparameter choices are well justified. Criteria on the choice of baseline models are discussed. Authors provided detailed architectures of the proposed models. Source code in supplementary material.\n", "summary_of_the_review": "This is an important contribution to multi-relational modelling and graph neural networks. The paper is clearly written, experiments are extensive and detailed supplementary info is provided.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667353173901}, {"id": "3mpiGajMfm0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1169/Reviewer_S9ug"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a graph convolution module that learns representations of a node by aggregating the representations from its neighbour nodes in a graph. Compared with the standard one (RGConv from Schlichtkrull et al., 2018), the proposed module introduces factorised version of the kernel matrix and other adaptations to reduce the computational complexity. The paper then applied the proposed module to several tasks of computer vision as well as protein structure modeling.", "review_text": "Very comprehensive experiments and novelty is a bit weak.\n\nI have read the response of the authors. Additional experiments are appreciated.", "strengths": "Strength:\n\n1. The paper provides comprehensive experiments on various tasks of computer vision and protein structure modeling. In each of the tasks, the paper compares with strong baselines on large-scale datasets. Comprehensive ablation study is also provided to show the efficiency of the proposed module. The empirical study of the paper is strong and supports the claims of the paper well.\n\n2. The proposed module shows promising results when applied with different backbone architectures in various tasks, which demonstrates the usability of the method.\n\n3. The paper is well written and structured. Experimental details are elaborated.\n\nWeaknesses:\n\n1. Despite the comprehensive empirical study, I feel that the novelty of the proposed method might be a bit weak. The proposed method is an adaptation from a well studied method (RGConv) from the angle of efficiency. The technical depth might not be very strong. But I would not complain much as similarity might be a factor that the method can be applied in various tasks.\n\n2.  The construction of the graphs may require additional tuning such as the number of KNN. It is also a bit heuristic as it is constructed by some predefined rules (although it can be dynamic with the learning of the model). Moreover, although the paper claims \"spatial\", I feel that it does not apply specific configurations to deal with \"spatial\" and \"spatial\" is just one of the predefined relations.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduces a graph convolution module that learns representations of a node by aggregating the representations from its neighbour nodes in a graph. Compared with the standard one (RGConv from Schlichtkrull et al., 2018), the proposed module introduces factorised version of the kernel matrix and other adaptations to reduce the computational complexity. The paper then applied the proposed module to several tasks of computer vision as well as protein structure modeling.", "strength_and_weaknesses": "Strength:\n\n1. The paper provides comprehensive experiments on various tasks of computer vision and protein structure modeling. In each of the tasks, the paper compares with strong baselines on large-scale datasets. Comprehensive ablation study is also provided to show the efficiency of the proposed module. The empirical study of the paper is strong and supports the claims of the paper well.\n\n2. The proposed module shows promising results when applied with different backbone architectures in various tasks, which demonstrates the usability of the method.\n\n3. The paper is well written and structured. Experimental details are elaborated.\n\nWeaknesses:\n\n1. Despite the comprehensive empirical study, I feel that the novelty of the proposed method might be a bit weak. The proposed method is an adaptation from a well studied method (RGConv) from the angle of efficiency. The technical depth might not be very strong. But I would not complain much as similarity might be a factor that the method can be applied in various tasks.\n\n2.  The construction of the graphs may require additional tuning such as the number of KNN. It is also a bit heuristic as it is constructed by some predefined rules (although it can be dynamic with the learning of the model). Moreover, although the paper claims \"spatial\", I feel that it does not apply specific configurations to deal with \"spatial\" and \"spatial\" is just one of the predefined relations.", "clarity,_quality,_novelty_and_reproducibility": "I'm happy with the clarity, quality and reproducibility of the paper. I have minor concerns on novelty.", "summary_of_the_review": "Very comprehensive experiments and novelty is a bit weak.\n\nI have read the response of the authors. Additional experiments are appreciated.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666589027945}, {"id": "xhq6broc6c", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1169/Reviewer_WbJw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces the EurNet for Efficient multi-range relational modeling. It constructs the multi-relational graph to encode the short-medium and long-range relation embedding. It also introduces a gated relational message passing layer to achieve that. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show the effectiveness of proposed approaches. ", "review_text": "I appreciate the author present detailed rebuttal. However, the related works I present share the similar ideas with the author work. The authors also admit they ignore these related works. Please do cite these paper for new draft.\n\nAfter discussion with AC and other reviewers, I keep rating unchanged.", "strengths": "Strength \n1, The motivation is clear, and the writing is easy to follow. \n\n2, The experiments results are extensive. The proposed approach outperforms previous work FocalNet in many settings. \n\n3, The experiment results for protein function prediction are interesting. \n\nWeakness.\n1, The idea or motivation of using different range graph structure is not new before the vision transformer.  The authors do not cite or compare the closely related works: 1, Graph-Based Global Reasoning Networks CVPR-2019. 2, Dynamic Graph Message Passing Network CVPR-2020. 3, Dual Graph Convolutional Network for Semantic Segmentation, BMVC-2019.\n\n2, The core contribution in Equ.3 is not new or novel. It contains the different window sizes of MHSA-like operator and along with a channel attention. \n\t\n3, The core contribution in Equ.3 should be simplified for easier understanding.  For example, why not give a figure to better illustrate the core operation?\n\n4, What the advantages of combining different range for classification tasks?\nWhat are the advantages over CNN+ transformer-like models?\n\n5, The performance on ImageNet and COCO are not very competitive. In particular, compared with Convnext, the GFlops and Throughput increases while the performance improvements are within 0.2 over different baselines. \n\nMoreover, the proposed framework does not show any performance gains over con-current methods.\nCMT: Convolutional Neural Networks Meet Vision Transformers. CVPR-2022\nCMT: Convolutional Neural Networks Meet Vision Transformers. Arxiv-2022.02\n\n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces the EurNet for Efficient multi-range relational modeling. It constructs the multi-relational graph to encode the short-medium and long-range relation embedding. It also introduces a gated relational message passing layer to achieve that. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show the effectiveness of proposed approaches. ", "strength_and_weaknesses": "Strength \n1, The motivation is clear, and the writing is easy to follow. \n\n2, The experiments results are extensive. The proposed approach outperforms previous work FocalNet in many settings. \n\n3, The experiment results for protein function prediction are interesting. \n\nWeakness.\n1, The idea or motivation of using different range graph structure is not new before the vision transformer.  The authors do not cite or compare the closely related works: 1, Graph-Based Global Reasoning Networks CVPR-2019. 2, Dynamic Graph Message Passing Network CVPR-2020. 3, Dual Graph Convolutional Network for Semantic Segmentation, BMVC-2019.\n\n2, The core contribution in Equ.3 is not new or novel. It contains the different window sizes of MHSA-like operator and along with a channel attention. \n\t\n3, The core contribution in Equ.3 should be simplified for easier understanding.  For example, why not give a figure to better illustrate the core operation?\n\n4, What the advantages of combining different range for classification tasks?\nWhat are the advantages over CNN+ transformer-like models?\n\n5, The performance on ImageNet and COCO are not very competitive. In particular, compared with Convnext, the GFlops and Throughput increases while the performance improvements are within 0.2 over different baselines. \n\nMoreover, the proposed framework does not show any performance gains over con-current methods.\nCMT: Convolutional Neural Networks Meet Vision Transformers. CVPR-2022\nCMT: Convolutional Neural Networks Meet Vision Transformers. Arxiv-2022.02\n\n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity, Quality are good, while the novelty is not new.", "summary_of_the_review": "I appreciate the author present detailed rebuttal. However, the related works I present share the similar ideas with the author work. The authors also admit they ignore these related works. Please do cite these paper for new draft.\n\nAfter discussion with AC and other reviewers, I keep rating unchanged.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666515427490}, {"id": "NscG_xh8RM", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1169/Reviewer_aeee"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "1, Introduce a task called multi-range relational modeling.\n2, Their EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions.\n3, The proposed GRMP separately performs (1) relational message aggregation on each individual feature channel and (2) node-wise aggregation of different feature channels\n", "review_text": "The paper introduces EurNet to do multi-range relational modeling. The proposed GRMP separately models multi-relational graph.", "strengths": "Strength:\n1, Propose a novel modeling layer, called gated relational message passing (GRMP), to propagate multi-relational information across the data.\n2, Demonstrate EurNets in image and protein structure modeling cases.\nWeaknesses:\n1, Why your model is more \"efficient\" than other models?\n2, Why do we need multi-range relational modeling in Image classification?\n3, Do you have an inference time comparison between previous works?\n4, The method part is somehow hard to follow. Do you have any figures to demonstrate your method?\n5, Why relation-channel entangled aggregation is better?", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "1, Introduce a task called multi-range relational modeling.\n2, Their EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions.\n3, The proposed GRMP separately performs (1) relational message aggregation on each individual feature channel and (2) node-wise aggregation of different feature channels\n", "strength_and_weaknesses": "Strength:\n1, Propose a novel modeling layer, called gated relational message passing (GRMP), to propagate multi-relational information across the data.\n2, Demonstrate EurNets in image and protein structure modeling cases.\nWeaknesses:\n1, Why your model is more \"efficient\" than other models?\n2, Why do we need multi-range relational modeling in Image classification?\n3, Do you have an inference time comparison between previous works?\n4, The method part is somehow hard to follow. Do you have any figures to demonstrate your method?\n5, Why relation-channel entangled aggregation is better?", "clarity,_quality,_novelty_and_reproducibility": "Clarity, Quality, Novelty, and Reproducibility are OK.", "summary_of_the_review": "The paper introduces EurNet to do multi-range relational modeling. The proposed GRMP separately models multi-relational graph.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666498827964}], "openreview_url": "https://openreview.net/forum?id=rMbrVNxYuqZ", "arxiv_id": "2211.12941", "paper_pdf": "papers/rMbrVNxYuqZ.pdf", "paper_pdf_sha256": "da3a8b2b93f2150d82241922068f81c805ec84c31ce08588077a43e86f72eb01", "paper_pdf_bytes": 1729572, "paper_pdf_source": "openreview", "code_url": "https://github.com/hirl-team/EurNet-Image", "code_repository": "hirl-team/EurNet-Image", "code_commit": "7770123cb289b763646ca65fe4cdc30494ea26ee", "code_archive": "repos/rMbrVNxYuqZ.zip", "code_archive_sha256": "f95928dae53ec47199060050ed23fc089e88cd5d6e585a9554a4d7f615712c4b", "code_archive_bytes": 414575, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 527, "github_languages": {"Python": 99800}, "github_archived": false, "github_pushed_at": "2022-11-24T01:26:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/eurnet-efficient-multi-range-relational"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NP9T_pViXU", "year": 2022, "status": "rejected", "title": "VIMPAC: Video Pre-Training via Masked Token Prediction and Contrastive Learning", "authors": ["Hao Tan", "Jie Lei", "Thomas Wolf", "Mohit Bansal"], "authorids": ["~Hao_Tan1", "~Jie_Lei3", "~Thomas_Wolf1", "~Mohit_Bansal2"], "authors_source": "OpenReview API", "abstract": "Video understanding relies on perceiving the overall global content and modeling its internal connections (e.g., causality, movement, and spatio-temporal correspondence). To learn these interactions, we apply a mask-then-predict pre-training task on the discretized video tokens generated via VQ-VAE. Unlike language, where the text tokens are more independent, neighboring video tokens typically have strong correlations (e.g., consecutive video frames usually look very similar), and hence uniformly masking individual tokens will make the task too trivial to learn useful representations. To deal with this issue, we propose a block-wise masking strategy where we mask neighboring video tokens in both spatial and temporal domains. We also add an augmentation-free contrastive learning method to further capture the global content by predicting whether the video clips are sampled from the same video. We pre-train our model on uncurated videos and show that our pre-trained model can reach state-of-the-art results on several video understanding datasets (e.g., SSV2, Diving48). Lastly, we provide detailed analyses of the model scalability and pre-training method design. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ChwBji79v6n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4348/Reviewer_hrwa"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a few new techniques: 1) A new video modeling architecture that uses a pre-trained VQ-VAE to tokenize frames, followed by a transformer encoder that aggregates the features and produces the final action class label. 2) Pre-training such an architecture using self-supervision by a) Masked prediction: Authors mask out blocks of tokens and predict them using the context (akin to BERT) and b) contrastive learning: Authors use the representation for two clips from same video as a positive match and otherwise negative match. The final model is trained with a linear combination of the masked prediction and contrastive losses, and finally finetuned for downstream tasks. The model is pretrained on HowTo100M dataset, and finetuned on multiple downstream datasets, where it obtains gains on more temporal datasets like SS-v2.", "review_text": "## Strengths\n\n1. The proposed model is a fresh approach to video modeling and performs well for temporal reasoning tasks like SSv2. The proposed pre-training methods leverage ideas from BERT and shows it can be effective for computer vision tasks too (similar to BEIT). \n\n2. The ablations are interesting, eg table 1 which shows pre-training is imperative for a model like this and lends significant gains.\n\n## Weaknesses\n\n1. Missing/incorrect details\n- The paper in Table 1 compares their method to prior work and notes the pre-training dataset used. Authors report their method only uses \"HT\" for pretraining. However, the tokenizer is using a pre-trained DALL-E model, the training data for which should also be mentioned here? Given this is a paper about learning from unlabeled data, it is important authors are very explicit about *all* the data being used.\n\n2. Continuing from 1), it is unfortunate that no results in Table 1 are comparable to each other; authors use a different pre-training dataset (HT+DALL-E), model architecture (DALL-E + Transformer) and modalities (video only) which makes it impossible to perform any apples-to-apples comparisons. The standard setting in Feichtenhofer et al (2020) seems to be to perform all comparisons with unlabeled Kinetics dataset as pre-training. It would have been ideal if authors kept at least some of the axes of variation fixed. In absence of that, it is only fair to compare the proposed method to state of the art. In that regard it struggles on spatially heavy datasets, and on temporal datasets it performs better although is comparable to more recent SOTA (eg MVIT -- 68.7 without large scale pretraining or HowTo100M pretraining, though it does use K600 labels). \n\n3. The architecture seems quite constrained as it relies on tokenization which looses spatial information. As authors themselves point out, this is likely the reason it struggles on spatial datasets. However, this also means this limitation would limit the impact of the model as it might be very useful for spatial tasks (video detection/segmentation etc). Hence, even though it might be good at learning representations, it is limited to classification. Using a BEIT style framework (as authors also recommend for future work) would make this paper much stronger in that regard.\n\n4. Missing related work: Authors should cite MVIT and other recent transformer based video models. Also while it is not technically published work, I would encourage the authors to cite and compare with video swin transformer.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a few new techniques: 1) A new video modeling architecture that uses a pre-trained VQ-VAE to tokenize frames, followed by a transformer encoder that aggregates the features and produces the final action class label. 2) Pre-training such an architecture using self-supervision by a) Masked prediction: Authors mask out blocks of tokens and predict them using the context (akin to BERT) and b) contrastive learning: Authors use the representation for two clips from same video as a positive match and otherwise negative match. The final model is trained with a linear combination of the masked prediction and contrastive losses, and finally finetuned for downstream tasks. The model is pretrained on HowTo100M dataset, and finetuned on multiple downstream datasets, where it obtains gains on more temporal datasets like SS-v2.", "main_review": "## Strengths\n\n1. The proposed model is a fresh approach to video modeling and performs well for temporal reasoning tasks like SSv2. The proposed pre-training methods leverage ideas from BERT and shows it can be effective for computer vision tasks too (similar to BEIT). \n\n2. The ablations are interesting, eg table 1 which shows pre-training is imperative for a model like this and lends significant gains.\n\n## Weaknesses\n\n1. Missing/incorrect details\n- The paper in Table 1 compares their method to prior work and notes the pre-training dataset used. Authors report their method only uses \"HT\" for pretraining. However, the tokenizer is using a pre-trained DALL-E model, the training data for which should also be mentioned here? Given this is a paper about learning from unlabeled data, it is important authors are very explicit about *all* the data being used.\n\n2. Continuing from 1), it is unfortunate that no results in Table 1 are comparable to each other; authors use a different pre-training dataset (HT+DALL-E), model architecture (DALL-E + Transformer) and modalities (video only) which makes it impossible to perform any apples-to-apples comparisons. The standard setting in Feichtenhofer et al (2020) seems to be to perform all comparisons with unlabeled Kinetics dataset as pre-training. It would have been ideal if authors kept at least some of the axes of variation fixed. In absence of that, it is only fair to compare the proposed method to state of the art. In that regard it struggles on spatially heavy datasets, and on temporal datasets it performs better although is comparable to more recent SOTA (eg MVIT -- 68.7 without large scale pretraining or HowTo100M pretraining, though it does use K600 labels). \n\n3. The architecture seems quite constrained as it relies on tokenization which looses spatial information. As authors themselves point out, this is likely the reason it struggles on spatial datasets. However, this also means this limitation would limit the impact of the model as it might be very useful for spatial tasks (video detection/segmentation etc). Hence, even though it might be good at learning representations, it is limited to classification. Using a BEIT style framework (as authors also recommend for future work) would make this paper much stronger in that regard.\n\n4. Missing related work: Authors should cite MVIT and other recent transformer based video models. Also while it is not technically published work, I would encourage the authors to cite and compare with video swin transformer.", "summary_of_the_review": "The proposed model is interesting and in-line with recent work like BEIT etc, and obtains some decent results. However I find it hard to draw meaningful conclusions from the paper due to lack of proper apples-to-apples comparisons. Moreover, given the architecture is likely to have limited impact in the field (as I discuss in my weaknesses), I am borderline on this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635916832021}, {"id": "xzRQGYHcvXe", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4348/Reviewer_PC2U"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a new pre-training method named VIMPAC for video understanding, which is a combination of specially-designed masked token prediction objective and contrastive learning objective. VIMPAC is well-motivated and shows strong empirical results on several video understanding tasks.", "review_text": "## **Strengths**\n* VIMPAC achieves strong empirical results on temporally-heavy datasets.\n* Although the mask token prediction task has been proven to be an effective pretraining objective for image tasks, the block masking method designed by authors is well motivated.\n* Tokenizing all videos without spatial augmentation is a good way to reduce storage and io cost.\n\n## **Weaknesses**\n* This paper is utilizing a transformer as the main architecture, with a pretrained VQ-VAE as the tokenizer. I think it's great to have the baselines with the same arch for fair comparison. Bert-large is a relatively large model comparing to previous CNN models, and large models tend to work better in pretraining with big datasets.\n* The contrastive learning objective part doesn't make a lot of sense to me. It seems adding contrastive learning objective only provides marginal improvement, and might even hurt the performance on temporally-heavy datasets. Ablation studies on contrastive learning also lead to some counter-intuitive conclusions so I would expect more analysis and insights from this part.\n* I'm also interested if VIMPAC can work for tasks beyond action classification (might be hard due to the architecture design).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a new pre-training method named VIMPAC for video understanding, which is a combination of specially-designed masked token prediction objective and contrastive learning objective. VIMPAC is well-motivated and shows strong empirical results on several video understanding tasks.", "main_review": "## **Strengths**\n* VIMPAC achieves strong empirical results on temporally-heavy datasets.\n* Although the mask token prediction task has been proven to be an effective pretraining objective for image tasks, the block masking method designed by authors is well motivated.\n* Tokenizing all videos without spatial augmentation is a good way to reduce storage and io cost.\n\n## **Weaknesses**\n* This paper is utilizing a transformer as the main architecture, with a pretrained VQ-VAE as the tokenizer. I think it's great to have the baselines with the same arch for fair comparison. Bert-large is a relatively large model comparing to previous CNN models, and large models tend to work better in pretraining with big datasets.\n* The contrastive learning objective part doesn't make a lot of sense to me. It seems adding contrastive learning objective only provides marginal improvement, and might even hurt the performance on temporally-heavy datasets. Ablation studies on contrastive learning also lead to some counter-intuitive conclusions so I would expect more analysis and insights from this part.\n* I'm also interested if VIMPAC can work for tasks beyond action classification (might be hard due to the architecture design).", "summary_of_the_review": "With strengths and weaknesses I listed in previous section, I think this paper makes some reasonable contributions but is marginally below the iclr acceptance threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635877542285}, {"id": "9EzK2uXub_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4348/Reviewer_E8qx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new video pretraining method by combining masked token prediction and contrastive learning. Off-she-shelf VQ-VAE is used in this paper for discrete video tokens generation. In order to make the masked token prediction more effective, the authors proposed a block mask strategy where spatial-temporal neighboring video tokens are masked together. Experimental results are shown to validate its effectiveness.  ", "review_text": "Strength:\n+ the paper is well written, main ideas are clear enough, it is also easy to follow for readers.\n+ the proposed block masking scheme makes sense in terms of avoiding information leakage from nearby tokens and it is also validated to be quite better than i.i.d.\n+ extensive experiments are carried out\n\nWeakness:\n- novelty: \n（1）Combining masked token prediction (MP) and contrastive learning (CL) is quite straightforward. Both MP and CL are studied previously, simpling combining these techniques is kind of incremental. Yes, the proposed block masking strategy is a plus, but it will have problem by directly combining it to CL. In my opinion, global information loss will be caused by block masking, it is actually harmful to the CL task. If better combination mechanism of the two tasks can be invented, it will be a good work.\n  (2) Discrete video token generation is also off-the-shelf technique. VQ-VAE is not new.\n\n- performance and evaluation: \n  (1) the performances on \"spatially-heavy\" datasets, such as UCF101, K400 and HMDB51, are far from the state-of-the-arts, rather than comparable. Such results show that the proposed pretraining method is only good at \"local\" modeling, its global representation is not strong enough. This may not only be caused by the discretization of VQ-VAE, but also has something to do with the harmful blocking masking which can harm the CL.\n (2)  Now that VQ-VAE can be harmful to the spatial information, why it should still be leveraged. There is also no evaluation on this part. What if VQ-VAE is replaced by other feature encoder and we regress the masked contiguous video feature points? Why not using video frame patches as discrete input tokens? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new video pretraining method by combining masked token prediction and contrastive learning. Off-she-shelf VQ-VAE is used in this paper for discrete video tokens generation. In order to make the masked token prediction more effective, the authors proposed a block mask strategy where spatial-temporal neighboring video tokens are masked together. Experimental results are shown to validate its effectiveness.  ", "main_review": "Strength:\n+ the paper is well written, main ideas are clear enough, it is also easy to follow for readers.\n+ the proposed block masking scheme makes sense in terms of avoiding information leakage from nearby tokens and it is also validated to be quite better than i.i.d.\n+ extensive experiments are carried out\n\nWeakness:\n- novelty: \n（1）Combining masked token prediction (MP) and contrastive learning (CL) is quite straightforward. Both MP and CL are studied previously, simpling combining these techniques is kind of incremental. Yes, the proposed block masking strategy is a plus, but it will have problem by directly combining it to CL. In my opinion, global information loss will be caused by block masking, it is actually harmful to the CL task. If better combination mechanism of the two tasks can be invented, it will be a good work.\n  (2) Discrete video token generation is also off-the-shelf technique. VQ-VAE is not new.\n\n- performance and evaluation: \n  (1) the performances on \"spatially-heavy\" datasets, such as UCF101, K400 and HMDB51, are far from the state-of-the-arts, rather than comparable. Such results show that the proposed pretraining method is only good at \"local\" modeling, its global representation is not strong enough. This may not only be caused by the discretization of VQ-VAE, but also has something to do with the harmful blocking masking which can harm the CL.\n (2)  Now that VQ-VAE can be harmful to the spatial information, why it should still be leveraged. There is also no evaluation on this part. What if VQ-VAE is replaced by other feature encoder and we regress the masked contiguous video feature points? Why not using video frame patches as discrete input tokens? ", "summary_of_the_review": "Given the aforementioned weakness of this paper, I think the novelty of this paper is currently limited, and its performance is not good.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635737613919}], "openreview_url": "https://openreview.net/forum?id=NP9T_pViXU", "arxiv_id": "2106.11250", "paper_pdf": "papers/NP9T_pViXU.pdf", "paper_pdf_sha256": "f95f9908f9ddfd86c9da5eee5c2d7bb3ab6ad5b4ab809d5a6b2d17892bb80ee6", "paper_pdf_bytes": 966814, "paper_pdf_source": "openreview", "code_url": "https://github.com/airsplay/vimpac", "code_repository": "airsplay/vimpac", "code_commit": "0760d19eecfc477bd39c448c3eec6226e07e7e4f", "code_archive": "repos/NP9T_pViXU.zip", "code_archive_sha256": "6ce028aaee79bf74d28eb976631270d0fed997e6d4c4c7f38570ead3bac525c3", "code_archive_bytes": 993474, "code_file_count": 33, "code_extensions": {".py": 22, ".sh": 11}, "github_disk_usage_kb": 1870, "github_languages": {"Python": 176892, "Shell": 12444}, "github_archived": false, "github_pushed_at": "2022-06-03T00:13:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/vimpac-video-pre-training-via-masked-token"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_SKUm2AJpvN", "year": 2021, "status": "rejected", "title": "Decoupling Representation Learning from Reinforcement Learning", "authors": ["Adam Stooke", "Kimin Lee", "Pieter Abbeel", "Michael Laskin"], "authorids": ["~Adam_Stooke2", "~Kimin_Lee1", "~Pieter_Abbeel2", "mlaskin@berkeley.edu"], "authors_source": "OpenReview API", "abstract": "In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning.  To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss.  In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments.  Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others.  We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks.  Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation.  Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at \\url{hidden url}.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AYBhxVXFT3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper171/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an auxiliary task for learning better representations for reinforcement learning. The idea is interesting and is a very active area of research at the moment.\n\nMy main concern is that the paper is mostly an empirical evaluation, as the novel algorithm is mostly an extra contrastive loss. Further, it seems their method is almost exclusively meant for pixel-based environments (see point 1 below), so the authors should make this point more explicit. Given this empirical emphasis, I feel the authors could have performed a deeper exploration into understanding _why_ their proposed algorithm performs the way it does in the different environments considered. In particular, some very specific design decisions were made in evaluation (see, for instance, points 2, 5, 9, 11, 12, 13, 14, 15, 16 below). \nThe clarity of exposition could also use some improvement, as I detail below.\n\nMain questions/concerns:\n1. In Section 3, the authors write \"This task encourages the learned encoder to extract meaningful elements of the structure of the MDP from observations.\". This assumes some type of continuity in pixel space, relative to MDP dynamics, which is in general not true.\n2. In Figure 1, why doesn't the momentum encoder go through a residual predictor?\n3. It seems $\\theta$, $\\phi$, $\\bar{\\theta}$, and $\\bar{\\phi}$ are all updated independently, is this the case? What is the actual training regimen? Are the ATC and regular RL networks trained concurrently?\n4. It would help if you include a proper algorithm in the paper.\n5. Above equation (2), the authors say \"In our implementation, the positives from all other elements...\". It's not clear what \"positives from  all other elements\" means.\n6. In section 4.1 the authors say \"multiple seeds were run\", please specify how many.\n7. In section 4.2 the authors say they are capable of \"training the encoder online, fully detached from the RL agent\", but how is it fully detached if they share the conv layers?\n8. It's not clear what the difference between ATC and UL training is. In some experiments the authors use ATC, in others UL. Are they the same thing? For example, in Figure 3, which ones are ATC? Also in Figure 14 vs Figures 15 and 16?\n9. In Figure 3, why does one environment compare with pri and the other with 2x, but not both in both environments?\n10. In the **Atari** subsection on the comment of detached training, the authors point out subpar performance on Breakout and SpaceInvaders. On SpaceInvaders it's possible the screen changes could cause issues, but what do the authors think cause the subpar performance in Breakout?\n11. In point (iii) of section 4.3, what's the network architecture used for training the RL part?\n12. In section 4.3 the authrs say they \"drew expert demonstrations from partially-trained RL agents\". Were these all drawn from the same checkpoint?\n13. In the **DMControl** section, the VAE is trying to reproduce a frame $T$ steps in the future? What is the value of $T$ used? Did you try different values?\n14. Similar questin for the **Atari** subsection. Also for this section, does your VAE try to predict individual frames or stacked frames (as frame stacking is common in Atari experiments)?\n15. In Figure 8 top, are these after pre-training the encoder? If that is the case, regular RL would have used fewer frames in comparison, no? Where would RL be if left to train for longer?\n16. In section 4.5 please clarify what \"random shift augmentations\" are.\n17. In the **Encoder analysis** subsection, what do you mean by \"attention\"?\n18. In Figures 14, 15, and 16 it's not at all clear what we're supposed to be looking for, nor how they show that ATC/UL is focusing on the score/enemy and the others are not.\n\nMinor comments:\n1. At the bottom of page 2, the term \"POMDP\" has not been introduced yet.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review", "review": "This paper proposes an auxiliary task for learning better representations for reinforcement learning. The idea is interesting and is a very active area of research at the moment.\n\nMy main concern is that the paper is mostly an empirical evaluation, as the novel algorithm is mostly an extra contrastive loss. Further, it seems their method is almost exclusively meant for pixel-based environments (see point 1 below), so the authors should make this point more explicit. Given this empirical emphasis, I feel the authors could have performed a deeper exploration into understanding _why_ their proposed algorithm performs the way it does in the different environments considered. In particular, some very specific design decisions were made in evaluation (see, for instance, points 2, 5, 9, 11, 12, 13, 14, 15, 16 below). \nThe clarity of exposition could also use some improvement, as I detail below.\n\nMain questions/concerns:\n1. In Section 3, the authors write \"This task encourages the learned encoder to extract meaningful elements of the structure of the MDP from observations.\". This assumes some type of continuity in pixel space, relative to MDP dynamics, which is in general not true.\n2. In Figure 1, why doesn't the momentum encoder go through a residual predictor?\n3. It seems $\\theta$, $\\phi$, $\\bar{\\theta}$, and $\\bar{\\phi}$ are all updated independently, is this the case? What is the actual training regimen? Are the ATC and regular RL networks trained concurrently?\n4. It would help if you include a proper algorithm in the paper.\n5. Above equation (2), the authors say \"In our implementation, the positives from all other elements...\". It's not clear what \"positives from  all other elements\" means.\n6. In section 4.1 the authors say \"multiple seeds were run\", please specify how many.\n7. In section 4.2 the authors say they are capable of \"training the encoder online, fully detached from the RL agent\", but how is it fully detached if they share the conv layers?\n8. It's not clear what the difference between ATC and UL training is. In some experiments the authors use ATC, in others UL. Are they the same thing? For example, in Figure 3, which ones are ATC? Also in Figure 14 vs Figures 15 and 16?\n9. In Figure 3, why does one environment compare with pri and the other with 2x, but not both in both environments?\n10. In the **Atari** subsection on the comment of detached training, the authors point out subpar performance on Breakout and SpaceInvaders. On SpaceInvaders it's possible the screen changes could cause issues, but what do the authors think cause the subpar performance in Breakout?\n11. In point (iii) of section 4.3, what's the network architecture used for training the RL part?\n12. In section 4.3 the authrs say they \"drew expert demonstrations from partially-trained RL agents\". Were these all drawn from the same checkpoint?\n13. In the **DMControl** section, the VAE is trying to reproduce a frame $T$ steps in the future? What is the value of $T$ used? Did you try different values?\n14. Similar questin for the **Atari** subsection. Also for this section, does your VAE try to predict individual frames or stacked frames (as frame stacking is common in Atari experiments)?\n15. In Figure 8 top, are these after pre-training the encoder? If that is the case, regular RL would have used fewer frames in comparison, no? Where would RL be if left to train for longer?\n16. In section 4.5 please clarify what \"random shift augmentations\" are.\n17. In the **Encoder analysis** subsection, what do you mean by \"attention\"?\n18. In Figures 14, 15, and 16 it's not at all clear what we're supposed to be looking for, nor how they show that ATC/UL is focusing on the score/enemy and the others are not.\n\nMinor comments:\n1. At the bottom of page 2, the term \"POMDP\" has not been introduced yet.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603935005142}, {"id": "e6miFpBSNCB", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper171/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary of the paper:\nThe paper aims to define an unsupervised pretraining architecture that can be used to pretrain representations for a reinforcement learning agent. The proposed solution is called \"Augmented Temporal Contrast\" (ATC). It consists of an encoder-compressor-predictor architecture for an (augmented) observation that is trained in a small latent space by minimizing the InfoNCE loss between that prediction, and the encoding of some future state.\nThe representation used (the pretrained encoder) is then frozen, and a policy network is trained consuming these representations. The authors then evaluate their approach on several different tasks: a control domain (DMcontrol), Mazenagivation (DMLab) and Atari. They show that their approach - without finetuning the representations - achieves comparable results to end-to-end reinforcement learning on most of the different domains. They also show that adding their defined loss as a regularizer always helps learning. They finally show that their representations can generalize out-of-domain by running several multi-task experiments, while only having pretrained on one domain.\n\nCommentary on the goal of the paper:\n\nThe goal of the paper is extremely important: separating learning representations from learning policies would enable better transfer, possible more sample efficiency and lower variance in outcome. \n\nStrengths:\n- The authors propose a well-designed solution that combines existing approaches in a well thought-out way. \n- The paper is extremely well written\n- The paper has an extensive empirical section. \n- Results are generally very good.\n\nWeaknesses:\n- The goal of the paper is a bit vague. As I said above, I agree and understand the desire for the separation of representation learning and reinforcement learning. However, the paper would have been stronger if it had concentrated on a single benefit of this separation, and evaluated their approach on that. While they say that their goal is to investigate \"how to learn representations which are agnostic to rewards\", this is too general as well. (One example would be to say that the decoupling makes for better generalizes to new MDPs - but this is not the focus of the analysis, just an aspect. The paper is, unfortunately, less convincing for it.)\n- Along a similar line of thought, the results, while strong, are not as convincing as they could be, because the paper does not focus on the benefits of reward-agnostic representations learned by ATC. The results that consider the generalization advantage (the Multi-Task learning experiments) are weaker, without the paper offering an analysis as to why.\n\nIn total, I would argue that this is a well-written paper with interesting analysis that could be a lot stronger by narrowing the scope of the contained argument. I argue for rejection, because I can see an updated version of this paper to be a great paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A well-written paper with sound reasoning and mostly good results that should sharpen and narrow its focus", "review": "Summary of the paper:\nThe paper aims to define an unsupervised pretraining architecture that can be used to pretrain representations for a reinforcement learning agent. The proposed solution is called \"Augmented Temporal Contrast\" (ATC). It consists of an encoder-compressor-predictor architecture for an (augmented) observation that is trained in a small latent space by minimizing the InfoNCE loss between that prediction, and the encoding of some future state.\nThe representation used (the pretrained encoder) is then frozen, and a policy network is trained consuming these representations. The authors then evaluate their approach on several different tasks: a control domain (DMcontrol), Mazenagivation (DMLab) and Atari. They show that their approach - without finetuning the representations - achieves comparable results to end-to-end reinforcement learning on most of the different domains. They also show that adding their defined loss as a regularizer always helps learning. They finally show that their representations can generalize out-of-domain by running several multi-task experiments, while only having pretrained on one domain.\n\nCommentary on the goal of the paper:\n\nThe goal of the paper is extremely important: separating learning representations from learning policies would enable better transfer, possible more sample efficiency and lower variance in outcome. \n\nStrengths:\n- The authors propose a well-designed solution that combines existing approaches in a well thought-out way. \n- The paper is extremely well written\n- The paper has an extensive empirical section. \n- Results are generally very good.\n\nWeaknesses:\n- The goal of the paper is a bit vague. As I said above, I agree and understand the desire for the separation of representation learning and reinforcement learning. However, the paper would have been stronger if it had concentrated on a single benefit of this separation, and evaluated their approach on that. While they say that their goal is to investigate \"how to learn representations which are agnostic to rewards\", this is too general as well. (One example would be to say that the decoupling makes for better generalizes to new MDPs - but this is not the focus of the analysis, just an aspect. The paper is, unfortunately, less convincing for it.)\n- Along a similar line of thought, the results, while strong, are not as convincing as they could be, because the paper does not focus on the benefits of reward-agnostic representations learned by ATC. The results that consider the generalization advantage (the Multi-Task learning experiments) are weaker, without the paper offering an analysis as to why.\n\nIn total, I would argue that this is a well-written paper with interesting analysis that could be a lot stronger by narrowing the scope of the contained argument. I argue for rejection, because I can see an updated version of this paper to be a great paper.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603825750533}, {"id": "67WeylJ5OSx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper171/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Summary:**\nThis paper presents a new unsupervised learning method for learning latent representations for visual RL control domains. The method, Augmented Temporal Contrast (ATC), can be used alone to learn a representation to be combined with an RL algorithm, or as an auxiliary task in an end-to-end system. ATC matches or outperforms comparable end-to-end systems in several environments. The paper provides an extensive experimental study to support its claims.\n\n**Strengths:**\nThe paper is clearly written, and all of the main points are well articulated. ATC appears sufficiently novel, and is applicable to a wide variety of domains, and can be deployed in various configurations (e.g. auxiliary task, unsupervised pre-training, etc…). Included is a thorough experimental study that effectively demonstrates the performance of the method.\n\n**Weaknesses:**\nAlthough the experimental study seem thorough. I could not find the actual number of independent runs (seeds) for each domain listed anywhere. This information should be included so that the reader can better evaluate the variance of each method, and make more confident conclusions. \n\n**Recommendation:**\nOverall I vote to accept. The method presented in the paper is not revolutionary, but it appears to be novel and significant enough to be of interest to deep RL practitioners.\n\n**Questions:**\nHow many independent runs (seeds) were used in each of the domains? Can this information be included in the main text?\n\n**After Author Response and Discussion:**\nThanks to the authors for their responses. After reading the other reviews and the author responses, I am lowering my score to 5. I think that the number of independent runs used (especially on the smaller domains), and the way the results are presented with the min-max extent makes me less convinced of the results than I was in the initial review. Adding many more independent runs (seeds), especially on the smaller domains, would improve my confidence a lot. Overall I think the paper is of interest to the community, but the experiments and their analysis could be improved.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A new unsupervised learning method for learning latent representations for visual control domains.", "review": "**Summary:**\nThis paper presents a new unsupervised learning method for learning latent representations for visual RL control domains. The method, Augmented Temporal Contrast (ATC), can be used alone to learn a representation to be combined with an RL algorithm, or as an auxiliary task in an end-to-end system. ATC matches or outperforms comparable end-to-end systems in several environments. The paper provides an extensive experimental study to support its claims.\n\n**Strengths:**\nThe paper is clearly written, and all of the main points are well articulated. ATC appears sufficiently novel, and is applicable to a wide variety of domains, and can be deployed in various configurations (e.g. auxiliary task, unsupervised pre-training, etc…). Included is a thorough experimental study that effectively demonstrates the performance of the method.\n\n**Weaknesses:**\nAlthough the experimental study seem thorough. I could not find the actual number of independent runs (seeds) for each domain listed anywhere. This information should be included so that the reader can better evaluate the variance of each method, and make more confident conclusions. \n\n**Recommendation:**\nOverall I vote to accept. The method presented in the paper is not revolutionary, but it appears to be novel and significant enough to be of interest to deep RL practitioners.\n\n**Questions:**\nHow many independent runs (seeds) were used in each of the domains? Can this information be included in the main text?\n\n**After Author Response and Discussion:**\nThanks to the authors for their responses. After reading the other reviews and the author responses, I am lowering my score to 5. I think that the number of independent runs used (especially on the smaller domains), and the way the results are presented with the min-max extent makes me less convinced of the results than I was in the initial review. Adding many more independent runs (seeds), especially on the smaller domains, would improve my confidence a lot. Overall I think the paper is of interest to the community, but the experiments and their analysis could be improved.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603747984181}, {"id": "dYg8nU0sOFm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper171/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces an unsupervised task called Augmented Temporal Contrast which associates pairs of observations separated in time using a contrastive loss. The paper uses the task in several training regimes (in online RL, pretraining and multi-task RL).\n\nPros:\n- Well written and structured paper.\n- Interesting, general and simple to implement task.\n- Evaluated in several training regimes and on several environments.\n- Improves sample efficiency on most of the environments and setups, and improves over prior methods.\n- The attention maps in the paper and appendix are great and although they may be hand picked(?) examples, it highlights the issue with many approaches that can't model a goal rarely seen.\n\nCons:\n- You write you use multiple seeds but I don't see anywhere details on this? Consider adding it to the tables in appendix.\n- In Figure 3 one agent step is 4x environment frames? I suggest to make it clear in plot or in caption. In Figure 2, environment frames is used.\n\nComments/questions:\n- Wrt. 4.1, to what extent is the \"small\" replay for DMLab necessary over just using observations in batches/unrolls? Looking at Table 3 I can't see how large the replay is? 10k as in SAC or smaller?\n\nUpdate: Not all the experiments are particular thorough and the novelty less than expected.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good improvement over prior unsupervised RL tasks in many setups", "review": "The paper introduces an unsupervised task called Augmented Temporal Contrast which associates pairs of observations separated in time using a contrastive loss. The paper uses the task in several training regimes (in online RL, pretraining and multi-task RL).\n\nPros:\n- Well written and structured paper.\n- Interesting, general and simple to implement task.\n- Evaluated in several training regimes and on several environments.\n- Improves sample efficiency on most of the environments and setups, and improves over prior methods.\n- The attention maps in the paper and appendix are great and although they may be hand picked(?) examples, it highlights the issue with many approaches that can't model a goal rarely seen.\n\nCons:\n- You write you use multiple seeds but I don't see anywhere details on this? Consider adding it to the tables in appendix.\n- In Figure 3 one agent step is 4x environment frames? I suggest to make it clear in plot or in caption. In Figure 2, environment frames is used.\n\nComments/questions:\n- Wrt. 4.1, to what extent is the \"small\" replay for DMLab necessary over just using observations in batches/unrolls? Looking at Table 3 I can't see how large the replay is? 10k as in SAC or smaller?\n\nUpdate: Not all the experiments are particular thorough and the novelty less than expected.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1602929082392}], "openreview_url": "https://openreview.net/forum?id=_SKUm2AJpvN", "arxiv_id": "2009.08319", "paper_pdf": "papers/_SKUm2AJpvN.pdf", "paper_pdf_sha256": "3b7defc2675a66ae7a8698766e02295d20593f3ed0f6a3ad5a479be40e82bb2d", "paper_pdf_bytes": 1957948, "paper_pdf_source": "openreview", "code_url": "https://github.com/astooke/rlpyt", "code_repository": "astooke/rlpyt", "code_commit": "f04f23db1eb7b5915d88401fca67869968a07a37", "code_archive": "repos/_SKUm2AJpvN.zip", "code_archive_sha256": "9d6653ab186008726d8ed589c0656516f5d6b55d4e20dbf9bfa53e312a7e95cf", "code_archive_bytes": 618818, "code_file_count": 410, "code_extensions": {".py": 410}, "github_disk_usage_kb": 1272, "github_languages": {"Python": 1387616}, "github_archived": false, "github_pushed_at": "2021-01-04T20:48:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/decoupling-representation-learning-from"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Skeh1krtvH", "year": 2020, "status": "rejected", "title": "WaveFlow: A Compact Flow-based Model for Raw Audio", "authors": ["Wei Ping", "Kainan Peng", "Kexin Zhao", "Zhao Song"], "authorids": ["weiping.thu@gmail.com"], "authors_source": "OpenReview API", "abstract": "In this work, we present WaveFlow, a small-footprint generative flow for raw audio, which is trained with maximum likelihood without complicated density distillation and auxiliary losses as used in Parallel WaveNet.  It provides a unified view of flow-based models for raw audio, including autoregressive flow (e.g., WaveNet) and bipartite flow (e.g., WaveGlow) as special cases. We systematically study these likelihood-based generative models for raw waveforms in terms of test likelihood and speech fidelity. We demonstrate that WaveFlow can synthesize high-fidelity speech and obtain comparable likelihood as WaveNet, while only requiring a few sequential steps to generate very long waveforms.  In particular, our small-footprint WaveFlow has only 5.91M parameters and can generate 22.05kHz speech 15.39 times faster than real-time on a GPU without customized inference kernels.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "SJloae9gcS", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1484/AnonReviewer2"], "rating": "3: Weak Reject", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper re-organized the high dimensional 1-D raw waveform as 2-D matrix. This method simulated the autoregressive flow. Log-likelihood could be calculated in parallel. Autoregressive flow was only run on row dimension. The number of required parameters was desirable to synthesize high-fidelity speech with the speed faster than real time. Although this method could not achieve top one in ranking in every measurements, the resulting performance was still obtained with the best average results. \n\nIn general, this paper is clearly written, well organized and easy to follow. The authors carried out sufficient experiments and analyses, and proposed some rules of thumb to build a good model. On one hand, we may catch the contributions. But, on the other hand, the contributions were not clearly explained. The results were averaged but were not clearly explained.\n\nThe authors suggested to specify a bigger receptive field than the squeezed height. The property of getting better performance using deeper wavenet was \"not\" clearly explained and investigated. In the experiments, a small number of generative steps was considered. This is because short sequence based on autoregressive model was used. \n\nThis paper mentioned that using convolution queue could improve the synthesis speed. But, the synthesis speed has been fast enough because it is almost 15 times faster than real time. In practical applications, 100x faster is almost the same as 15x faster for humans. But, the task isn’t interacted with human. It is suggested to focuse on reducing the number of parameters or enhancing the log likelihood.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published one or two papers in this area.", "rating": "3: Weak Reject", "review_assessment:_thoroughness_in_paper_reading": "I made a quick assessment of this paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper re-organized the high dimensional 1-D raw waveform as 2-D matrix. This method simulated the autoregressive flow. Log-likelihood could be calculated in parallel. Autoregressive flow was only run on row dimension. The number of required parameters was desirable to synthesize high-fidelity speech with the speed faster than real time. Although this method could not achieve top one in ranking in every measurements, the resulting performance was still obtained with the best average results. \n\nIn general, this paper is clearly written, well organized and easy to follow. The authors carried out sufficient experiments and analyses, and proposed some rules of thumb to build a good model. On one hand, we may catch the contributions. But, on the other hand, the contributions were not clearly explained. The results were averaged but were not clearly explained.\n\nThe authors suggested to specify a bigger receptive field than the squeezed height. The property of getting better performance using deeper wavenet was \"not\" clearly explained and investigated. In the experiments, a small number of generative steps was considered. This is because short sequence based on autoregressive model was used. \n\nThis paper mentioned that using convolution queue could improve the synthesis speed. But, the synthesis speed has been fast enough because it is almost 15 times faster than real time. In practical applications, 100x faster is almost the same as 15x faster for humans. But, the task isn’t interacted with human. It is suggested to focuse on reducing the number of parameters or enhancing the log likelihood."}, "tcdate": 1572016323383}, {"id": "SkeuG_BAYB", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1484/AnonReviewer1"], "rating": "6: Weak Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This submission belongs to the field of text-to-speech synthesis. In particular it looks at a novel way of formulating a normalising flow using 2D rather than conventional 1D representation. Such reformulation enables to provide interpretations to several existing approaches as well as formulate a new one with quite interesting properties. This submission would benefit from a discussion of limitations of your approach. \n\nI believe there is a great deal of interest in the use of normalising flows in the text-to-speech area. I believe this submission could be a good contribution to the area. The test log-likelihoods look comparable to existing approaches with significantly worse inference times. The mean opinion scores (MOS) seem to approach one of the standard baselines with significantly worse inference times though at the expense of increasing the number of model parameters from 6M to 86M parameters whilst gaining only 0.2 in MOS. The submission would have benefited from discussion about model complexity/expressivity and it's impact on MOS for WaveFlow, WaveNet and other approaches. \n\nThe largest issues with this submission are:\n\n1) lack of proper technical description of your model in sections 1 and 2 making reading sections 1,2,3,etc in order awkward. It seems the order should be 3,4,(5),1,2,(5). \n2) complete omission of conditioning on text to be synthesised; anyone not familiar deeply with speech synthesis will wonder where does the text come in\n3) explicit statement of complexity for the operations involved using proper big-O notation; helps to avoid confusion about what do you mean by \"parallel\" (autoregressive WaveNet followed by parallel computation != parallel computation)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have published in this field for several years.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper thoroughly.", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "title": "Official Blind Review #1", "review_assessment:_checking_correctness_of_derivations_and_theory": "I carefully checked the derivations and theory.", "review": "This submission belongs to the field of text-to-speech synthesis. In particular it looks at a novel way of formulating a normalising flow using 2D rather than conventional 1D representation. Such reformulation enables to provide interpretations to several existing approaches as well as formulate a new one with quite interesting properties. This submission would benefit from a discussion of limitations of your approach. \n\nI believe there is a great deal of interest in the use of normalising flows in the text-to-speech area. I believe this submission could be a good contribution to the area. The test log-likelihoods look comparable to existing approaches with significantly worse inference times. The mean opinion scores (MOS) seem to approach one of the standard baselines with significantly worse inference times though at the expense of increasing the number of model parameters from 6M to 86M parameters whilst gaining only 0.2 in MOS. The submission would have benefited from discussion about model complexity/expressivity and it's impact on MOS for WaveFlow, WaveNet and other approaches. \n\nThe largest issues with this submission are:\n\n1) lack of proper technical description of your model in sections 1 and 2 making reading sections 1,2,3,etc in order awkward. It seems the order should be 3,4,(5),1,2,(5). \n2) complete omission of conditioning on text to be synthesised; anyone not familiar deeply with speech synthesis will wonder where does the text come in\n3) explicit statement of complexity for the operations involved using proper big-O notation; helps to avoid confusion about what do you mean by \"parallel\" (autoregressive WaveNet followed by parallel computation != parallel computation)\n"}, "tcdate": 1571866639996}, {"id": "rylJ0A96Yr", "reviewer_signature": ["ICLR.cc/2020/Conference/Paper1484/AnonReviewer3"], "rating": "8: Accept", "confidence": "", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n## Updated review\n\nI have read the rebuttal. The new version of the paper is definitely clearer, especially the contribution section and the experimental results. The new version addresses all my concerns, hence I am upgrading my rating to Accept.\n\n## Original review\n\nThis paper presents the WaveGlow model, a generative model for raw audio. The model is based on a 2D-matrix approach, which allows to generate the audio with a fixed amount of step. The model is shown to be a generalization of the two main approaches for raw audio generation, autoregressive flow and bipartite flow. The model is evaluated and compared with related work on an objective evaluation (Log-likelihood) and a subjective evaluation (MOS), and is shown to be a trade-off between memory footprint, generation speed and quality.\n\nI think this paper should be accepted, for the following reasons:\n- The theoretical framework presented is novel and significant, as it provides a unified view of the two main approaches for neural waveform generation.\n- The experiments are reasonably convincing, although they could be improved.\n\nDetailed comments:\n- In the subjective evaluation section (5.2), Table 5 is hard to decipher, especially given that there are three measurements to take into account, so it's not easy to see the benefit of the approach. Maybe the results should be organised differently, for instance grouping them according to one measurement could help, typically showing what speed and MOS each of the three models can achieve for a given model size. Maybe plotting speed vs MOS for the same model size could also be interesting. \n- In the same section, is the WaveNet model the original one, or the Parallel WaveNet ? if it's the original, why not include Parallel WaveNet in the table ?\n- Typo at the end of Section 1: \"We orgnize\" -> \"organize\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_checking_correctness_of_experiments": "I carefully checked the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #3", "review": "\n## Updated review\n\nI have read the rebuttal. The new version of the paper is definitely clearer, especially the contribution section and the experimental results. The new version addresses all my concerns, hence I am upgrading my rating to Accept.\n\n## Original review\n\nThis paper presents the WaveGlow model, a generative model for raw audio. The model is based on a 2D-matrix approach, which allows to generate the audio with a fixed amount of step. The model is shown to be a generalization of the two main approaches for raw audio generation, autoregressive flow and bipartite flow. The model is evaluated and compared with related work on an objective evaluation (Log-likelihood) and a subjective evaluation (MOS), and is shown to be a trade-off between memory footprint, generation speed and quality.\n\nI think this paper should be accepted, for the following reasons:\n- The theoretical framework presented is novel and significant, as it provides a unified view of the two main approaches for neural waveform generation.\n- The experiments are reasonably convincing, although they could be improved.\n\nDetailed comments:\n- In the subjective evaluation section (5.2), Table 5 is hard to decipher, especially given that there are three measurements to take into account, so it's not easy to see the benefit of the approach. Maybe the results should be organised differently, for instance grouping them according to one measurement could help, typically showing what speed and MOS each of the three models can achieve for a given model size. Maybe plotting speed vs MOS for the same model size could also be interesting. \n- In the same section, is the WaveNet model the original one, or the Parallel WaveNet ? if it's the original, why not include Parallel WaveNet in the table ?\n- Typo at the end of Section 1: \"We orgnize\" -> \"organize\"", "review_assessment:_checking_correctness_of_derivations_and_theory": "I did not assess the derivations or theory."}, "tcdate": 1571823302601}], "openreview_url": "https://openreview.net/forum?id=Skeh1krtvH", "arxiv_id": "1912.01219", "paper_pdf": "papers/Skeh1krtvH.pdf", "paper_pdf_sha256": "782282378863686b403d582a694a2992d6258dceb46adba6ad1425fdb4b6831b", "paper_pdf_bytes": 353151, "paper_pdf_source": "openreview", "code_url": "https://github.com/PaddlePaddle/Parakeet", "code_repository": "PaddlePaddle/Parakeet", "code_commit": "8705a2a8405e3c63f2174d69880d2b5525a6c9fd", "code_archive": "repos/Skeh1krtvH.zip", "code_archive_sha256": "6231ae9f0dcb71998bdcdb5cd4c411856d00d56afaafb8193989ad4fbb0fb22e", "code_archive_bytes": 5527625, "code_file_count": 242, "code_extensions": {".py": 196, ".sh": 44, ".ipynb": 2}, "github_disk_usage_kb": 9775, "github_languages": {"Python": 725526, "Shell": 22908}, "github_archived": true, "github_pushed_at": "2021-11-19T02:21:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/waveflow-a-compact-flow-based-model-for-raw-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UyiC52OPjn", "year": 2026, "status": "rejected", "title": "Characterizing Nonlinear Dynamics via Smooth Prototype Equivalences", "authors": ["Roy Friedman", "Noa Moriel", "Matthew Ricci", "Guy Pelc", "Yair Weiss", "Mor Nitzan"], "authorids": ["~Roy_Friedman2", "~Noa_Moriel1", "~Matthew_Ricci1", "~Guy_Pelc1", "~Yair_Weiss1", "~Mor_Nitzan1"], "authors_source": "OpenReview API", "abstract": "Characterizing dynamical systems given limited measurements is a common challenge throughout the physical and biological sciences. However, this task is challenging, especially due to transient variability in systems with equivalent long-term dynamics. We address this by introducing smooth prototype equivalences (SPE), a framework matches between sparse observations of phase space and prototypical behaviors using invertible neural networks. SPE enables classification by comparing the deformation loss of the observed sparse measurements to the prototype dynamics. Furthermore, our approach enables estimation of the invariant sets of the observed dynamics through the learned mapping from prototype space to data space. Our method outperforms existing techniques in the classification of oscillatory systems and can efficiently identify invariant structures like limit cycles and fixed points in an equation-free manner, even when only a small, noisy subset of the phase space is observed. Finally, we show how our method can be used for the detection of biological processes like the cell cycle trajectory from high-dimensional single-cell gene expression data.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "fU5c9F1lEu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16810/Reviewer_Pr4v"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes a Smooth Prototypical Equivalence (SPE) framework, a model designed to predict the invariant sets of unknown dynamical systems by pushforwarding known, simple prototypical vector fields. The model learns diffeomorphisms, parameterized by normalizing flows, that smoothly map between vector field data sampled from unknown systems and their prototypical counterparts. If a system is smoothly (or orbit) conjugate to a specific prototypical vector field, then the pushforward of the observations should match that prototypical vector field. This principle is encoded in the equivalence loss, which measures the discrepancy between the two. The prototypical vector field with the lowest equivalence loss is then selected as the canonical form of the data, and the pullback of its invariant set serves as an approximation of the invariant set of the data. The proposed method is benchmarked on several synthetic examples as well as a more complex single-cell gene expression dataset.", "review_text": "This paper proposes a Smooth Prototypical Equivalence (SPE) framework, a model designed to predict the invariant sets of unknown dynamical systems by pushforwarding known, simple prototypical vector fields. The model learns diffeomorphisms, parameterized by normalizing flows, that smoothly map between vector field data sampled from unknown systems and their prototypical counterparts. If a system is smoothly (or orbit) conjugate to a specific prototypical vector field, then the pushforward of the observations should match that prototypical vector field. This principle is encoded in the equivalence loss, which measures the discrepancy between the two. The prototypical vector field with the lowest equivalence loss is then selected as the canonical form of the data, and the pullback of its invariant set serves as an approximation of the invariant set of the data. The proposed method is benchmarked on several synthetic examples as well as a more complex single-cell gene expression dataset.", "strengths": "Matching the pushforward vector field with certain prototypes via smooth conjugacy is novel and conceptually sound. The paper is well-organized and easy to follow.", "weaknesses": "However, it is not entirely clear whether the smooth/orbit equivalence loss can be reliably applied in more general settings. When the equivalence loss is exactly zero, the pullback of the prototypical invariant set indeed corresponds to the true invariant set of the data dynamics. Yet, when the loss is nonzero, it is not guaranteed that a smaller equivalence loss implies a closer approximation of the true invariant set.\n\nIf the underlying invariant sets are hyperbolic, the equivalence loss might still be meaningful. The persistence theorem ensures that hyperbolic invariant sets cannot be destroyed by small perturbations. Thus, if the equivalence loss can be regarded as a perturbation metric, one might argue that a smaller loss increases the likelihood that the pullbacked invariant set remains approximately invariant. However, since the persistence theorem requires $C^1$-closeness, this justification is not entirely rigorous in the current formulation based on C0-closeness. It may be possible for the authors to derive a $C^1$ bound from $C^0$-closeness, though this is uncertain.\n\nMoreover, if the underlying invariant sets are non-hyperbolic, no such guarantee exists. Even a small perturbation (= a nonzero equivalence loss) can drastically alter the structure of the invariant sets. The benchmarked systems in the paper appear to be restricted to low-dimensional, hyperbolic cases (such as attracting limit cycles), where the theoretical assumptions implicitly hold. It remains unclear whether the proposed framework would perform reliably beyond these settings.", "questions": "The authors' method relies on the assumption that a smaller equivalence loss implies a closer correspondence between the true invariant set of the target system and the pullbacked one. However, this relationship is not theoretically justified in the paper. What theoretical guarantee does your equivalence loss provide regarding the recoverability or approximation quality of invariant sets? Specifically,\n\n- Can the equivalence loss be interpreted as a meaningful bound or metric (e.g., in the persistence theorem sense) on the deviation between true and estimated invariant sets?\n- If not, under what conditions (e.g., hyperbolicity, structural stability) can a smaller equivalence loss be expected to correspond to a more accurate recovery of the invariant structure?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a Smooth Prototypical Equivalence (SPE) framework, a model designed to predict the invariant sets of unknown dynamical systems by pushforwarding known, simple prototypical vector fields. The model learns diffeomorphisms, parameterized by normalizing flows, that smoothly map between vector field data sampled from unknown systems and their prototypical counterparts. If a system is smoothly (or orbit) conjugate to a specific prototypical vector field, then the pushforward of the observations should match that prototypical vector field. This principle is encoded in the equivalence loss, which measures the discrepancy between the two. The prototypical vector field with the lowest equivalence loss is then selected as the canonical form of the data, and the pullback of its invariant set serves as an approximation of the invariant set of the data. The proposed method is benchmarked on several synthetic examples as well as a more complex single-cell gene expression dataset.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "Matching the pushforward vector field with certain prototypes via smooth conjugacy is novel and conceptually sound. The paper is well-organized and easy to follow.", "weaknesses": "However, it is not entirely clear whether the smooth/orbit equivalence loss can be reliably applied in more general settings. When the equivalence loss is exactly zero, the pullback of the prototypical invariant set indeed corresponds to the true invariant set of the data dynamics. Yet, when the loss is nonzero, it is not guaranteed that a smaller equivalence loss implies a closer approximation of the true invariant set.\n\nIf the underlying invariant sets are hyperbolic, the equivalence loss might still be meaningful. The persistence theorem ensures that hyperbolic invariant sets cannot be destroyed by small perturbations. Thus, if the equivalence loss can be regarded as a perturbation metric, one might argue that a smaller loss increases the likelihood that the pullbacked invariant set remains approximately invariant. However, since the persistence theorem requires $C^1$-closeness, this justification is not entirely rigorous in the current formulation based on C0-closeness. It may be possible for the authors to derive a $C^1$ bound from $C^0$-closeness, though this is uncertain.\n\nMoreover, if the underlying invariant sets are non-hyperbolic, no such guarantee exists. Even a small perturbation (= a nonzero equivalence loss) can drastically alter the structure of the invariant sets. The benchmarked systems in the paper appear to be restricted to low-dimensional, hyperbolic cases (such as attracting limit cycles), where the theoretical assumptions implicitly hold. It remains unclear whether the proposed framework would perform reliably beyond these settings.", "questions": "The authors' method relies on the assumption that a smaller equivalence loss implies a closer correspondence between the true invariant set of the target system and the pullbacked one. However, this relationship is not theoretically justified in the paper. What theoretical guarantee does your equivalence loss provide regarding the recoverability or approximation quality of invariant sets? Specifically,\n\n- Can the equivalence loss be interpreted as a meaningful bound or metric (e.g., in the persistence theorem sense) on the deviation between true and estimated invariant sets?\n- If not, under what conditions (e.g., hyperbolicity, structural stability) can a smaller equivalence loss be expected to correspond to a more accurate recovery of the invariant structure?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761916142768}, {"id": "bGwKU3SXnb", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16810/Reviewer_1nY5"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes 'Smooth Prototype Equivalences (SPE)', a novel framework for characterizing the long-term behavior of nonlinear dynamical systems from sparse, noisy, and high-dimensional data, a common challenge in the physical and biological sciences.\nThe core idea is to learn a mapping that 'smoothly' deforms the data space into a known prototype space using Invertible Neural Networks (INNs), rather than directly modeling the complex, unknown dynamics. This mapping aligns the observed data with simple, well-understood 'prototype' dynamics (e.g., stable fixed points, simple limit cycles).\nThis approach provides two main functions:\n\tLocalization: It pinpoints the shape and location of hidden 'invariant sets' (e.g., limit cycles, fixed points) in the original data space via the learned inverse mapping (H^(-1)).\n\tClassification: It allows for classifying the system's dynamical regime by comparing the goodness-of-fit (lowest loss) across multiple prototypes.\nThe authors demonstrate that SPE is robust, outperforming existing techniques, especially in realistic, data-scarce, and noisy scenarios. Moreover, they successfully apply this method to high-dimensional biological data (single-cell gene expression data) to extract the complex 'cell cycle' trajectory.", "review_text": "This paper proposes 'Smooth Prototype Equivalences (SPE)', a novel framework for characterizing the long-term behavior of nonlinear dynamical systems from sparse, noisy, and high-dimensional data, a common challenge in the physical and biological sciences.\nThe core idea is to learn a mapping that 'smoothly' deforms the data space into a known prototype space using Invertible Neural Networks (INNs), rather than directly modeling the complex, unknown dynamics. This mapping aligns the observed data with simple, well-understood 'prototype' dynamics (e.g., stable fixed points, simple limit cycles).\nThis approach provides two main functions:\n\tLocalization: It pinpoints the shape and location of hidden 'invariant sets' (e.g., limit cycles, fixed points) in the original data space via the learned inverse mapping (H^(-1)).\n\tClassification: It allows for classifying the system's dynamical regime by comparing the goodness-of-fit (lowest loss) across multiple prototypes.\nThe authors demonstrate that SPE is robust, outperforming existing techniques, especially in realistic, data-scarce, and noisy scenarios. Moreover, they successfully apply this method to high-dimensional biological data (single-cell gene expression data) to extract the complex 'cell cycle' trajectory.", "strengths": "1. Data Efficiency and Robustness: This is the method's strongest point. As shown in Figure 3, it performs far more stably and accurately than other methods (SINDy, MLP), even with very few samples (N=25) and significant noise. This is critical for real-world scientific applications.\n2. Interpretability and Localization: Unlike simple 'black-box' classifiers (e.g., TWA), SPE provides the actual shape and location of the invariant set in the data space via H^(-1). This offers scientists deeper insights—not just \"what kind?\" but also \"where and in what shape?\"\n3. Scalability and Practicality: The method's effectiveness is demonstrated beyond 2D examples, scaling successfully to 6D (Figure 4) and 100D (Figure 5, scRNA-seq) high-dimensional data. The extraction of a biologically meaningful trajectory from 100D real-world cell cycle data is a particularly impressive result.\n4. Equation-Free Approach: It can identify the core structure of a dynamical system  without any knowledge of the system's governing equations.", "weaknesses": "1.\tDependence on Prototypes: The method relies on the user defining a 'dictionary' of prototypes in advance, based on what dynamics they expect to find. If the true dynamics are of a completely novel form or are not in the dictionary, the classification and localization may fail.\n2.\tLimitations on Complex Dynamics: As the authors note, this work primarily focuses on simple attractors like stable fixed points or limit cycles. Chaotic systems, characterized by 'strange attractors' with fractal structures, are difficult to map to simple prototypes using smooth equivalence.", "questions": "1.\tHow sensitive are the accuracy and stability of the results to the INN architecture (e.g., number of blocks, number of Fourier features)? Could you share any empirical guidelines for hyperparameter tuning?\n2.\tA question regarding the construction of the prototype dictionary: How would SPE handle a complex system that contains multiple different dynamical behaviors (e.g., a system with two fixed points and one limit cycle coexisting)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes 'Smooth Prototype Equivalences (SPE)', a novel framework for characterizing the long-term behavior of nonlinear dynamical systems from sparse, noisy, and high-dimensional data, a common challenge in the physical and biological sciences.\nThe core idea is to learn a mapping that 'smoothly' deforms the data space into a known prototype space using Invertible Neural Networks (INNs), rather than directly modeling the complex, unknown dynamics. This mapping aligns the observed data with simple, well-understood 'prototype' dynamics (e.g., stable fixed points, simple limit cycles).\nThis approach provides two main functions:\n\tLocalization: It pinpoints the shape and location of hidden 'invariant sets' (e.g., limit cycles, fixed points) in the original data space via the learned inverse mapping (H^(-1)).\n\tClassification: It allows for classifying the system's dynamical regime by comparing the goodness-of-fit (lowest loss) across multiple prototypes.\nThe authors demonstrate that SPE is robust, outperforming existing techniques, especially in realistic, data-scarce, and noisy scenarios. Moreover, they successfully apply this method to high-dimensional biological data (single-cell gene expression data) to extract the complex 'cell cycle' trajectory.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Data Efficiency and Robustness: This is the method's strongest point. As shown in Figure 3, it performs far more stably and accurately than other methods (SINDy, MLP), even with very few samples (N=25) and significant noise. This is critical for real-world scientific applications.\n2. Interpretability and Localization: Unlike simple 'black-box' classifiers (e.g., TWA), SPE provides the actual shape and location of the invariant set in the data space via H^(-1). This offers scientists deeper insights—not just \"what kind?\" but also \"where and in what shape?\"\n3. Scalability and Practicality: The method's effectiveness is demonstrated beyond 2D examples, scaling successfully to 6D (Figure 4) and 100D (Figure 5, scRNA-seq) high-dimensional data. The extraction of a biologically meaningful trajectory from 100D real-world cell cycle data is a particularly impressive result.\n4. Equation-Free Approach: It can identify the core structure of a dynamical system  without any knowledge of the system's governing equations.", "weaknesses": "1.\tDependence on Prototypes: The method relies on the user defining a 'dictionary' of prototypes in advance, based on what dynamics they expect to find. If the true dynamics are of a completely novel form or are not in the dictionary, the classification and localization may fail.\n2.\tLimitations on Complex Dynamics: As the authors note, this work primarily focuses on simple attractors like stable fixed points or limit cycles. Chaotic systems, characterized by 'strange attractors' with fractal structures, are difficult to map to simple prototypes using smooth equivalence.", "questions": "1.\tHow sensitive are the accuracy and stability of the results to the INN architecture (e.g., number of blocks, number of Fourier features)? Could you share any empirical guidelines for hyperparameter tuning?\n2.\tA question regarding the construction of the prototype dictionary: How would SPE handle a complex system that contains multiple different dynamical behaviors (e.g., a system with two fixed points and one limit cycle coexisting)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761900025074}, {"id": "EPCGQnM4DL", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16810/Reviewer_AQH7"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "This paper introduces a method called Smooth Prototype Equivalences (SPE) for characterizing dynamical systems from sparse, noisy observations without explicit governing equations. SPE uses invertible neural networks to learn smooth diffeomorphisms between observed data and predefined prototype dynamics (e.g., limit cycles or fixed points) to exploiting the systems sharing the same qualitative behavior. The method has two purposes: localizing invariant structures by mapping prototype attractors to data space via the learned inverse transformation, and classifying dynamical regimes by comparing equivalence losses across prototypes. Experiments show SPE outperforms existing baselines on sparse, noisy data and successfully extends to high-dimensional systems, including a 6D gene regulatory network and real 100-dimensional single-cell RNA velocity data where it recovers cell cycle gene expression patterns.", "review_text": "This paper introduces a method called Smooth Prototype Equivalences (SPE) for characterizing dynamical systems from sparse, noisy observations without explicit governing equations. SPE uses invertible neural networks to learn smooth diffeomorphisms between observed data and predefined prototype dynamics (e.g., limit cycles or fixed points) to exploiting the systems sharing the same qualitative behavior. The method has two purposes: localizing invariant structures by mapping prototype attractors to data space via the learned inverse transformation, and classifying dynamical regimes by comparing equivalence losses across prototypes. Experiments show SPE outperforms existing baselines on sparse, noisy data and successfully extends to high-dimensional systems, including a 6D gene regulatory network and real 100-dimensional single-cell RNA velocity data where it recovers cell cycle gene expression patterns.", "strengths": "(1) Technical contribution with Fourier Feature Coupling layers is a novel idea;\n(2) The use of prototypes (e.g., limit cycle, fixed point) provides clear physical and qualitative meaning to the learned mappings;\n(3) The method can infer invariant structures and classify system behaviors directly from sparse vector-field data without knowing governing equations;\n(4) The method effectively scales to high-dimensional systems and identifies the cell cycle from real single-cell RNA data, which shows practical value in biological science in a data-driven approach.", "weaknesses": "(1) Requires predefined prototypes based on domain knowledge in advance, which limits its applicability;\n(2) Training multiple invertible neural networks for different prototypes can be expensive, especially for high-dimensional data;\n(3) The method only guarantees local equivalence near observed data points.", "questions": "(1) Since SPE requires a set of prototype systems (e.g., limit cycles, fixed points), how should users select or construct these prototypes in practice? What happens if the true system exhibits a behavior not covered by the prototype?\n(2) Is the learned diffeomorphism $H_\\theta$ unique? Could different network initializations produce distinct but equally valid mappings?\n(3) Why Fourier features specifically? Have you tried other function approximators like neural ODEs or implicit layers for the coupling?\n(4) Given that Koopman operator methods are now also popular tools for data-driven dynamical analysis, how does SPE relate to them? Could these two frameworks complement each other? In your SPE equation (i.e., Eq. 1), the $\\partial_x H(x)\\cdot \\dot{x}$ is equivalent to $dH(x)/dt = \\mathcal{L}H(x)$ where $\\mathcal{L}$ is the Koopman generator. How would you think of it? Perhaps, considering the Koopman operator framework would help enhance the mathematical formulation of your method. Here is a list of papers you may have interest:\n\n  (a) https://link.springer.com/article/10.1007/s00332-015-9258-5\n\n  (b) https://www.aimsciences.org/article/doi/10.3934/jcd.2015005\n\n  (c) https://epubs.siam.org/doi/10.1137/21M1401243\n\n  (d) https://link.springer.com/article/10.1007/s11071-005-2824-x\n\n  (e) https://pubs.aip.org/aip/cha/article/27/10/103111/151485/Extended-dynamic-mode-decomposition-with\n\n  (f)  https://pubs.aip.org/aip/cha/article-abstract/35/10/103123/3368087/A-data-driven-framework-for-Koopman-semigroup?redirectedFrom=fulltext", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a method called Smooth Prototype Equivalences (SPE) for characterizing dynamical systems from sparse, noisy observations without explicit governing equations. SPE uses invertible neural networks to learn smooth diffeomorphisms between observed data and predefined prototype dynamics (e.g., limit cycles or fixed points) to exploiting the systems sharing the same qualitative behavior. The method has two purposes: localizing invariant structures by mapping prototype attractors to data space via the learned inverse transformation, and classifying dynamical regimes by comparing equivalence losses across prototypes. Experiments show SPE outperforms existing baselines on sparse, noisy data and successfully extends to high-dimensional systems, including a 6D gene regulatory network and real 100-dimensional single-cell RNA velocity data where it recovers cell cycle gene expression patterns.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "(1) Technical contribution with Fourier Feature Coupling layers is a novel idea;\n(2) The use of prototypes (e.g., limit cycle, fixed point) provides clear physical and qualitative meaning to the learned mappings;\n(3) The method can infer invariant structures and classify system behaviors directly from sparse vector-field data without knowing governing equations;\n(4) The method effectively scales to high-dimensional systems and identifies the cell cycle from real single-cell RNA data, which shows practical value in biological science in a data-driven approach.", "weaknesses": "(1) Requires predefined prototypes based on domain knowledge in advance, which limits its applicability;\n(2) Training multiple invertible neural networks for different prototypes can be expensive, especially for high-dimensional data;\n(3) The method only guarantees local equivalence near observed data points.", "questions": "(1) Since SPE requires a set of prototype systems (e.g., limit cycles, fixed points), how should users select or construct these prototypes in practice? What happens if the true system exhibits a behavior not covered by the prototype?\n(2) Is the learned diffeomorphism $H_\\theta$ unique? Could different network initializations produce distinct but equally valid mappings?\n(3) Why Fourier features specifically? Have you tried other function approximators like neural ODEs or implicit layers for the coupling?\n(4) Given that Koopman operator methods are now also popular tools for data-driven dynamical analysis, how does SPE relate to them? Could these two frameworks complement each other? In your SPE equation (i.e., Eq. 1), the $\\partial_x H(x)\\cdot \\dot{x}$ is equivalent to $dH(x)/dt = \\mathcal{L}H(x)$ where $\\mathcal{L}$ is the Koopman generator. How would you think of it? Perhaps, considering the Koopman operator framework would help enhance the mathematical formulation of your method. Here is a list of papers you may have interest:\n\n  (a) https://link.springer.com/article/10.1007/s00332-015-9258-5\n\n  (b) https://www.aimsciences.org/article/doi/10.3934/jcd.2015005\n\n  (c) https://epubs.siam.org/doi/10.1137/21M1401243\n\n  (d) https://link.springer.com/article/10.1007/s11071-005-2824-x\n\n  (e) https://pubs.aip.org/aip/cha/article/27/10/103111/151485/Extended-dynamic-mode-decomposition-with\n\n  (f)  https://pubs.aip.org/aip/cha/article-abstract/35/10/103123/3368087/A-data-driven-framework-for-Koopman-semigroup?redirectedFrom=fulltext", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1760862673926}], "openreview_url": "https://openreview.net/forum?id=UyiC52OPjn", "arxiv_id": "2503.10336", "paper_pdf": "papers/UyiC52OPjn.pdf", "paper_pdf_sha256": "4541fb38110d373647f468a7540ce5d1a8b3e0e702b58b4188ec48e910304b7a", "paper_pdf_bytes": 5491647, "paper_pdf_source": "openreview", "code_url": "https://github.com/nitzanlab/prototype-equivalences", "code_repository": "nitzanlab/prototype-equivalences", "code_commit": "5291a487ce4a3befe034b91d7386e6872c7509b9", "code_archive": "repos/UyiC52OPjn.zip", "code_archive_sha256": "5c4793bc67fd42f8440426352310e3e4aa1c31e1ed7cff46884778e00941b0b8", "code_archive_bytes": 1649493, "code_file_count": 13, "code_extensions": {".py": 11, ".ipynb": 2}, "github_disk_usage_kb": 4272, "github_languages": {"Jupyter Notebook": 296751, "Python": 125709}, "github_archived": false, "github_pushed_at": "2026-08-05T08:43:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/characterizing-nonlinear-dynamics-via-smooth"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S1OAqOtN5U", "year": 2025, "status": "rejected", "title": "Bayes Adaptive Monte Carlo Tree Search for Offline Model-based Reinforcement Learning", "authors": ["Jiayu Chen", "Wentse Chen", "Jeff Schneider"], "authorids": ["~Jiayu_Chen2", "~Wentse_Chen1", "~Jeff_Schneider1"], "authors_source": "OpenReview API", "abstract": "Offline reinforcement learning (RL) is a powerful approach for data-driven decision-making and control. Compared to model-free methods, offline model-based reinforcement learning (MBRL) explicitly learns world models from a static dataset and uses them as surrogate simulators, improving the data efficiency and enabling the learned policy to potentially generalize beyond the dataset support. However, there could be various MDPs that behave identically on the offline dataset and so dealing with the uncertainty about the true MDP can be challenging.  In this paper, we propose modeling offline MBRL as a Bayes Adaptive Markov Decision Process (BAMDP), which is a principled framework for addressing model uncertainty. We further introduce a novel Bayes Adaptive Monte-Carlo planning algorithm capable of solving BAMDPs in continuous state and action spaces with stochastic transitions. This planning process is based on Monte Carlo Tree Search and can be integrated into offline MBRL as a policy improvement operator in policy iteration. Our \"RL + Search\" framework follows in the footsteps of superhuman AIs like AlphaZero, improving on current offline MBRL methods by incorporating more computation input. The proposed algorithm significantly outperforms state-of-the-art model-based and model-free offline RL methods on twelve D4RL MuJoCo benchmark tasks and three target tracking tasks in a challenging, stochastic tokamak control simulator.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "lJ4lZjXUrw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8469/Reviewer_HEZk"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper propose to use a Bayesian adaptive method to train the model instead of end to end training approach used by MuZero. As indicate by the authors, this can modify the difficulty of model training in complex scenarios like continuous action space. Specifically, the method can be described in three main designs:1) it adapts DPW to handle continuous action and state spaces. 2) it use BA to refine the model during interactions and search process 3) It adapts MuZero’s supervised way to perform policy improvement. The experiment section shows methods using BA generally have a better performance than baselines, but the advantage of certain design seems not significant.", "review_text": "This paper propose to use a Bayesian adaptive method to train the model instead of end to end training approach used by MuZero. As indicate by the authors, this can modify the difficulty of model training in complex scenarios like continuous action space. Specifically, the method can be described in three main designs:1) it adapts DPW to handle continuous action and state spaces. 2) it use BA to refine the model during interactions and search process 3) It adapts MuZero’s supervised way to perform policy improvement. The experiment section shows methods using BA generally have a better performance than baselines, but the advantage of certain design seems not significant.", "strengths": "1. The authors describe the entire method in a clear way using pseudocode in Algorithms 1 and 2, and the insight of separating the model training from end-to-end process is explained in experiment section by comparing to sampled EfficientZero (sez). \n2. The experimental results indicate the effectiveness of the BA model training. Although the effectiveness of certain designs like MCTS search and supervised policy learning is not significant according to Table 1, I am glad to see the authors report the results honestly.", "weaknesses": "1. It seems the main improvement comes from BA model training according to Table 1. Thus, I wonder if there are other model-based RL methods using BA and deep ensembles to train the models, and why the authors did not include them in the baseline.\n2. In my opinion, the BA process adjust the belief of the models according to the sample results of current belief instead of the interaction with true environment, which may lead to accumulated error. For example, if the belief mistakenly concentrates on a wrong model and the sample results are likely to come from the wrong model, then future belief are likely to be more inclined to this wrong model. Could the authors provide some insights that why this not happen according to the experiment results?", "questions": "1. I am not sure about the necessity of adjusting the belief not only in the acting process but also in the searching process. Could the author explain the different influences of adjusting the model at the two stages?\n2. The authors indicate that sez does not work well due to the difficulty of end to end model training, but sez also use a different way of choosing actions nodes from DPW. Can the authors explain why they do not consider it as a main factor?\n3. I am curious about why the authors choose to update the policy in a trajectory way. MuZero just sample a series of transitions from the buffer, which can avoid the interaction in learned models. Could the authors explain the reason behind this design?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose to use a Bayesian adaptive method to train the model instead of end to end training approach used by MuZero. As indicate by the authors, this can modify the difficulty of model training in complex scenarios like continuous action space. Specifically, the method can be described in three main designs:1) it adapts DPW to handle continuous action and state spaces. 2) it use BA to refine the model during interactions and search process 3) It adapts MuZero’s supervised way to perform policy improvement. The experiment section shows methods using BA generally have a better performance than baselines, but the advantage of certain design seems not significant.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The authors describe the entire method in a clear way using pseudocode in Algorithms 1 and 2, and the insight of separating the model training from end-to-end process is explained in experiment section by comparing to sampled EfficientZero (sez). \n2. The experimental results indicate the effectiveness of the BA model training. Although the effectiveness of certain designs like MCTS search and supervised policy learning is not significant according to Table 1, I am glad to see the authors report the results honestly.", "weaknesses": "1. It seems the main improvement comes from BA model training according to Table 1. Thus, I wonder if there are other model-based RL methods using BA and deep ensembles to train the models, and why the authors did not include them in the baseline.\n2. In my opinion, the BA process adjust the belief of the models according to the sample results of current belief instead of the interaction with true environment, which may lead to accumulated error. For example, if the belief mistakenly concentrates on a wrong model and the sample results are likely to come from the wrong model, then future belief are likely to be more inclined to this wrong model. Could the authors provide some insights that why this not happen according to the experiment results?", "questions": "1. I am not sure about the necessity of adjusting the belief not only in the acting process but also in the searching process. Could the author explain the different influences of adjusting the model at the two stages?\n2. The authors indicate that sez does not work well due to the difficulty of end to end model training, but sez also use a different way of choosing actions nodes from DPW. Can the authors explain why they do not consider it as a main factor?\n3. I am curious about why the authors choose to update the policy in a trajectory way. MuZero just sample a series of transitions from the buffer, which can avoid the interaction in learned models. Could the authors explain the reason behind this design?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730785420127}, {"id": "wRIMeZnjSh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8469/Reviewer_khHT"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper addresses model uncertainty in offline MBRL by framing it as a BAMDP, allowing for adaptive belief updates over multiple potential world models. The authors propose a novel continuous MCTS method to solve BAMDPs in continuous and stochastic settings, integrating this planning approach as a policy improvement step in a policy iteration framework. The approach, which combines Bayesian RL with search, is shown to improve performance across a range of continuous control tasks.", "review_text": "This paper addresses model uncertainty in offline MBRL by framing it as a BAMDP, allowing for adaptive belief updates over multiple potential world models. The authors propose a novel continuous MCTS method to solve BAMDPs in continuous and stochastic settings, integrating this planning approach as a policy improvement step in a policy iteration framework. The approach, which combines Bayesian RL with search, is shown to improve performance across a range of continuous control tasks.", "strengths": "* Practical MCTS implementation for offline MBRL formulated as BAMDP: the paper introduces a practical implementation of the MCTS algorithm specifically adapted for continuous control tasks under the BAMDP framework. Though a straightforward approach, this integration is effective and brings continuous MCTS to offline MBRL with clear potential.\n\n* Well-executed ablative studies: the paper provides ablation studies that dissect the individual contributions of the algorithm’s components, demonstrating the benefits of the BAMDP formulation, MCTS planning, and the search-guided policy learning approach.", "weaknesses": "* Insufficient discussion of related work in Bayesian RL and offline MBRL: the related work section could better connect with existing research on modeling offline RL as a Bayes Adaptive MDP or epistemic POMDP. Ghosh et al. (2022) introduced a model-free approach to handling model uncertainty by training offline RL policies to adapt to belief changes, while Dorfman et al. (2021) framed offline meta-RL as a Bayesian RL problem, leveraging a belief-augmented MDP. Discussing these connections would position the current work more clearly within the broader landscape of Bayesian treatments of offline RL.\n\n* Lack of connection to prior policy learning with model-based search: using model-based search outcomes for policy learning within an actor-critic framework was introduced by Feinberg et al. (2018) and later applied to offline MBRL in Jeong et al. (2023). Although these methods do not use MCTS for rollouts, they share the concept of leveraging generated trajectories to improve action-value estimation for policy learning. Discussing these earlier works would clarify the paper’s contributions in relation to prior approaches.\n\n* Clarity issues in the experiment section: the structure of the experiment section could be improved to clarify the main research questions and key takeaways. At present, the content lacks focus on specific questions, making it challenging to follow the significance of each result. Organizing the discussion around clearly defined research questions could improve readability and focus.\n\n* Omission of recent offline MBRL baselines: the empirical evaluations rely on older methods (e.g., MOPO and MOReL), making it difficult to assess claims about state-of-the-art performance. Comparisons with more recent baselines, such as RAMBO (Rigter et al., 2022), MAPLE (Chen et al., 2021), and CBOP (Jeong et al., 2023), would strengthen the experimental claims and provide a more comprehensive evaluation of the proposed method’s effectiveness.\n\n__References:__\n\n- Ghosh et al. (2022): \"Offline RL policies should be trained to be adaptive.\"\n\n- Dorfman et al. (2021): \"Offline meta reinforcement learning – identifiability challenges and effective data collection strategies.\"\n\n- Feinberg et al. (2018): \"Model-Based Value Expansion for Efficient Model-Free Reinforcement Learning.\"\n\n- Jeong et al. (2023): \"Conservative Bayesian Model-Based Value Expansion for Offline Policy Optimization.\"\n\n- Rigter et al. (2022): \"Rambo-rl: Robust adversarial model-based offline reinforcement learning.\"\n\n- Chen et al. (2021): \"Offline model-based adaptable policy learning.\"", "questions": "* Belief updates during training: are the beliefs updated using imaginary rollouts sampled from the model ensemble? If so, could this introduce systematic biases? The paper mentions applying a reward penalty within a pessimistic MDP framework, but it would be useful to assess whether this is sufficient for accurate belief updates under the BAMDP setup.\n\n* Representation of returned policy in Algorithm 1: how is the policy $\\pi_{\\mathrm{ret}}$, returned by the SEARCH procedure, represented in practice? Further details would clarify its applicability to real-time policy learning.\n\n* Clarification on state representation: footnote 5 states that the policy takes the states consisting of both $s$ and $h$, but line 416 describes differently. Could the authors clarify this discrepancy? Additionally, MAPLE (Chen et al., 2021) employs an RNN-based policy network, which aligns well with the BAMDP framework. Using MAPLE as a baseline could enhance consistency with the proposed architecture.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses model uncertainty in offline MBRL by framing it as a BAMDP, allowing for adaptive belief updates over multiple potential world models. The authors propose a novel continuous MCTS method to solve BAMDPs in continuous and stochastic settings, integrating this planning approach as a policy improvement step in a policy iteration framework. The approach, which combines Bayesian RL with search, is shown to improve performance across a range of continuous control tasks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "* Practical MCTS implementation for offline MBRL formulated as BAMDP: the paper introduces a practical implementation of the MCTS algorithm specifically adapted for continuous control tasks under the BAMDP framework. Though a straightforward approach, this integration is effective and brings continuous MCTS to offline MBRL with clear potential.\n\n* Well-executed ablative studies: the paper provides ablation studies that dissect the individual contributions of the algorithm’s components, demonstrating the benefits of the BAMDP formulation, MCTS planning, and the search-guided policy learning approach.", "weaknesses": "* Insufficient discussion of related work in Bayesian RL and offline MBRL: the related work section could better connect with existing research on modeling offline RL as a Bayes Adaptive MDP or epistemic POMDP. Ghosh et al. (2022) introduced a model-free approach to handling model uncertainty by training offline RL policies to adapt to belief changes, while Dorfman et al. (2021) framed offline meta-RL as a Bayesian RL problem, leveraging a belief-augmented MDP. Discussing these connections would position the current work more clearly within the broader landscape of Bayesian treatments of offline RL.\n\n* Lack of connection to prior policy learning with model-based search: using model-based search outcomes for policy learning within an actor-critic framework was introduced by Feinberg et al. (2018) and later applied to offline MBRL in Jeong et al. (2023). Although these methods do not use MCTS for rollouts, they share the concept of leveraging generated trajectories to improve action-value estimation for policy learning. Discussing these earlier works would clarify the paper’s contributions in relation to prior approaches.\n\n* Clarity issues in the experiment section: the structure of the experiment section could be improved to clarify the main research questions and key takeaways. At present, the content lacks focus on specific questions, making it challenging to follow the significance of each result. Organizing the discussion around clearly defined research questions could improve readability and focus.\n\n* Omission of recent offline MBRL baselines: the empirical evaluations rely on older methods (e.g., MOPO and MOReL), making it difficult to assess claims about state-of-the-art performance. Comparisons with more recent baselines, such as RAMBO (Rigter et al., 2022), MAPLE (Chen et al., 2021), and CBOP (Jeong et al., 2023), would strengthen the experimental claims and provide a more comprehensive evaluation of the proposed method’s effectiveness.\n\n__References:__\n\n- Ghosh et al. (2022): \"Offline RL policies should be trained to be adaptive.\"\n\n- Dorfman et al. (2021): \"Offline meta reinforcement learning – identifiability challenges and effective data collection strategies.\"\n\n- Feinberg et al. (2018): \"Model-Based Value Expansion for Efficient Model-Free Reinforcement Learning.\"\n\n- Jeong et al. (2023): \"Conservative Bayesian Model-Based Value Expansion for Offline Policy Optimization.\"\n\n- Rigter et al. (2022): \"Rambo-rl: Robust adversarial model-based offline reinforcement learning.\"\n\n- Chen et al. (2021): \"Offline model-based adaptable policy learning.\"", "questions": "* Belief updates during training: are the beliefs updated using imaginary rollouts sampled from the model ensemble? If so, could this introduce systematic biases? The paper mentions applying a reward penalty within a pessimistic MDP framework, but it would be useful to assess whether this is sufficient for accurate belief updates under the BAMDP setup.\n\n* Representation of returned policy in Algorithm 1: how is the policy $\\pi_{\\mathrm{ret}}$, returned by the SEARCH procedure, represented in practice? Further details would clarify its applicability to real-time policy learning.\n\n* Clarification on state representation: footnote 5 states that the policy takes the states consisting of both $s$ and $h$, but line 416 describes differently. Could the authors clarify this discrepancy? Additionally, MAPLE (Chen et al., 2021) employs an RNN-based policy network, which aligns well with the BAMDP framework. Using MAPLE as a baseline could enhance consistency with the proposed architecture.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730693760245}, {"id": "uQ3O5qXugv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8469/Reviewer_YGvg"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper models offline model-based reinforcement learning as a Bayes Adaptive MDP, which enables the agent to leverage the model uncertainty during learning. By using MCTS with double progressive widening, the agent achieves good performance on the D4RL benchmark and three target-tracking tasks.", "review_text": "This paper models offline model-based reinforcement learning as a Bayes Adaptive MDP, which enables the agent to leverage the model uncertainty during learning. By using MCTS with double progressive widening, the agent achieves good performance on the D4RL benchmark and three target-tracking tasks.", "strengths": "The paper investigates a promising direction of leveraging the Bayesian RL framework combined with MCTS in offline RL. The paper is well-motivated and the performance on the D4RL benchmark outperforms baselines.", "weaknesses": "Experiments:\n- A more thorough analysis of the proposed method's performance is necessary. The observed performance improvement appears to result primarily from two factors: (1) the use of BAMDP, and (2) planning with MCTS. To verify this, it would be helpful to see an experiment where only BAMDP is applied, keeping other elements unchanged. While the BA-MBRL variant seems intended to isolate the effects of BAMDP, the paper does not clearly explain how this differs from other models like MOPO or MoReL. For instance, it’s unclear if the performance gap is solely due to BAMDP. Additionally, tracking the evolution of $b(\\theta)$ during updates would strengthen the claims about the role of BAMDP.\n- A comparison of runtime across methods would be valuable, especially to understand the impact of using BAMDP and MCTS on computational efficiency. Providing runtime metrics could help illustrate any trade-offs between performance gains and computational cost. (The paper only reports the runtime of Sampled EfficientZero.)\n- The inclusion of Figure 1 is somewhat ambiguous. The authors seem to suggest that the proposed method outperforms Sampled EfficientZero. However, a more detailed analysis is needed. For example, what factors contribute to the performance difference with BA-MCTS-SL? Is the advantage primarily due to using the Bayesian RL framework?\n-  In Line 472, the paper states, \"both Sampled EfficientZero and BA-MCTS-SL rely on supervised learning... Thus, purely mimicking the search result may be less sample-efficient than policy gradient methods.\" However, in Line 406, \"BA-MCTS-SL performs similarly to MA-MCTS, validating the effectiveness of both policy update mechanisms.\" This conclusion seems inconsistent. \n\nWriting: \n- Notation: what is hars'. It appears many times but I haven't found the definition. For example, in Algorithm1, the inputs of SIMULATE are defined as ($s, h$), $b(\\theta)$, $d$ (Line 225), while in Line 231, the inputs of SIMULATE are $(s', hars'), b', d-1$.\n- The update of $b(\\theta)$ in Eq 4. should be clarifed. \n- The figures should be added via PDF or SVG, otherwise, the figure can be blurry.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper models offline model-based reinforcement learning as a Bayes Adaptive MDP, which enables the agent to leverage the model uncertainty during learning. By using MCTS with double progressive widening, the agent achieves good performance on the D4RL benchmark and three target-tracking tasks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The paper investigates a promising direction of leveraging the Bayesian RL framework combined with MCTS in offline RL. The paper is well-motivated and the performance on the D4RL benchmark outperforms baselines.", "weaknesses": "Experiments:\n- A more thorough analysis of the proposed method's performance is necessary. The observed performance improvement appears to result primarily from two factors: (1) the use of BAMDP, and (2) planning with MCTS. To verify this, it would be helpful to see an experiment where only BAMDP is applied, keeping other elements unchanged. While the BA-MBRL variant seems intended to isolate the effects of BAMDP, the paper does not clearly explain how this differs from other models like MOPO or MoReL. For instance, it’s unclear if the performance gap is solely due to BAMDP. Additionally, tracking the evolution of $b(\\theta)$ during updates would strengthen the claims about the role of BAMDP.\n- A comparison of runtime across methods would be valuable, especially to understand the impact of using BAMDP and MCTS on computational efficiency. Providing runtime metrics could help illustrate any trade-offs between performance gains and computational cost. (The paper only reports the runtime of Sampled EfficientZero.)\n- The inclusion of Figure 1 is somewhat ambiguous. The authors seem to suggest that the proposed method outperforms Sampled EfficientZero. However, a more detailed analysis is needed. For example, what factors contribute to the performance difference with BA-MCTS-SL? Is the advantage primarily due to using the Bayesian RL framework?\n-  In Line 472, the paper states, \"both Sampled EfficientZero and BA-MCTS-SL rely on supervised learning... Thus, purely mimicking the search result may be less sample-efficient than policy gradient methods.\" However, in Line 406, \"BA-MCTS-SL performs similarly to MA-MCTS, validating the effectiveness of both policy update mechanisms.\" This conclusion seems inconsistent. \n\nWriting: \n- Notation: what is hars'. It appears many times but I haven't found the definition. For example, in Algorithm1, the inputs of SIMULATE are defined as ($s, h$), $b(\\theta)$, $d$ (Line 225), while in Line 231, the inputs of SIMULATE are $(s', hars'), b', d-1$.\n- The update of $b(\\theta)$ in Eq 4. should be clarifed. \n- The figures should be added via PDF or SVG, otherwise, the figure can be blurry.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730660835527}, {"id": "ERcBuu0Z95", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8469/Reviewer_grWz"], "rating": 5, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "Model-based offline reinforcement learning is a promising approach to improve the data efficiency and the generalization capability beyond the training dataset. To address the uncertainty issue when fitting the model with a static dataset, the conventional approach is to use an ensemble of world models, from which the uncertainty can be estimated. This work focuses on the adaptive ensemble mechanism where the belief over each model in the ensemble will be updated when the new samples are experienced by the algorithms. The proposed Bayes Adaptive Monte-Carlo planning algorithm has two key contributions: cast the offline MBRL as Bayes Adaptive MDP (BAMDP), based on which a Monte Carlo tree search method is used for planning. The experiments on D4RL MuJoCo and the three target tracking tasks show that the proposed method has better performance compared with state-of-the-art model-based and model-free offline RL methods.", "review_text": "Model-based offline reinforcement learning is a promising approach to improve the data efficiency and the generalization capability beyond the training dataset. To address the uncertainty issue when fitting the model with a static dataset, the conventional approach is to use an ensemble of world models, from which the uncertainty can be estimated. This work focuses on the adaptive ensemble mechanism where the belief over each model in the ensemble will be updated when the new samples are experienced by the algorithms. The proposed Bayes Adaptive Monte-Carlo planning algorithm has two key contributions: cast the offline MBRL as Bayes Adaptive MDP (BAMDP), based on which a Monte Carlo tree search method is used for planning. The experiments on D4RL MuJoCo and the three target tracking tasks show that the proposed method has better performance compared with state-of-the-art model-based and model-free offline RL methods.", "strengths": "- Experiments. This work considers both the D4RL MuJoCo and challenging stochastic target tracking tasks. The experiments also include an ablation study where each component of the proposed algorithms is verified, i.e., BA-MBRL, BA-MCTS, BA-MCTS-SL. The results are compared with many SOTA methods such as COMBO, MoReL, MOPO.\n- Problem Formulation. I find the adaptive ensemble mechanism is indeed a crucial problem in the ensemble-based method, and the proposed method using the BAMDP as a backbone to address this issue is novel. The algorithms are well-motivated.  I do believe that this paper has good contribution on the offline model based RL community.", "weaknesses": "- The comparison with related work is unclear. In particular, MuZero is a major  comparison object to motivate this study, while the statement in the paper is not clear. In Section 3, the paper claims that MuZero does not apply uncertainty estimation of the learned model, which has no direct relation with using latent models (i.e., estimating uncertainty in the latent space is possible). Please provide a more structured comparison of key related works to highlight the specific gaps.\n- Method section is confusing. Section 4.1 titled “ key role of deep ensemble,” while the contents are more about introducing BAMDP and reward design to incorporate the pessimistic-MDP. In Section 4.2, the newly introduced $(s,h)$ as a decision point seems out of nowhere, what is the connection with $(s,b)$ as introduced in BAMDP? The introduction of the algorithm in Section 4.2 is very hard to follow, and a paragraph title can help a lot. For instance, in the last paragraph where introducing the Bayes-optimal policy,  the author should directly introduce the design choice and then compare it with the previous methods. \n- Lack of comparison of the uncertainty evaluation. One of the key motivations of the proposed method is to use the samples to update the belief of ensembles. In the experiments, the related results on the benefits of such adaptation for ensemble-based methods are lacking (e.g., set the adaptive parameter to be zero).", "questions": "- what is the major purpose of having Fig.1? \n- What is the explanation of the lack of the convergence in Figure 2 for optimized method? Why in the last subfigure all the curves seem to struggle  to converge?\n- How to select $\\lambda$ in the reward and the number of the ensembles in practice?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Model-based offline reinforcement learning is a promising approach to improve the data efficiency and the generalization capability beyond the training dataset. To address the uncertainty issue when fitting the model with a static dataset, the conventional approach is to use an ensemble of world models, from which the uncertainty can be estimated. This work focuses on the adaptive ensemble mechanism where the belief over each model in the ensemble will be updated when the new samples are experienced by the algorithms. The proposed Bayes Adaptive Monte-Carlo planning algorithm has two key contributions: cast the offline MBRL as Bayes Adaptive MDP (BAMDP), based on which a Monte Carlo tree search method is used for planning. The experiments on D4RL MuJoCo and the three target tracking tasks show that the proposed method has better performance compared with state-of-the-art model-based and model-free offline RL methods.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "- Experiments. This work considers both the D4RL MuJoCo and challenging stochastic target tracking tasks. The experiments also include an ablation study where each component of the proposed algorithms is verified, i.e., BA-MBRL, BA-MCTS, BA-MCTS-SL. The results are compared with many SOTA methods such as COMBO, MoReL, MOPO.\n- Problem Formulation. I find the adaptive ensemble mechanism is indeed a crucial problem in the ensemble-based method, and the proposed method using the BAMDP as a backbone to address this issue is novel. The algorithms are well-motivated.  I do believe that this paper has good contribution on the offline model based RL community.", "weaknesses": "- The comparison with related work is unclear. In particular, MuZero is a major  comparison object to motivate this study, while the statement in the paper is not clear. In Section 3, the paper claims that MuZero does not apply uncertainty estimation of the learned model, which has no direct relation with using latent models (i.e., estimating uncertainty in the latent space is possible). Please provide a more structured comparison of key related works to highlight the specific gaps.\n- Method section is confusing. Section 4.1 titled “ key role of deep ensemble,” while the contents are more about introducing BAMDP and reward design to incorporate the pessimistic-MDP. In Section 4.2, the newly introduced $(s,h)$ as a decision point seems out of nowhere, what is the connection with $(s,b)$ as introduced in BAMDP? The introduction of the algorithm in Section 4.2 is very hard to follow, and a paragraph title can help a lot. For instance, in the last paragraph where introducing the Bayes-optimal policy,  the author should directly introduce the design choice and then compare it with the previous methods. \n- Lack of comparison of the uncertainty evaluation. One of the key motivations of the proposed method is to use the samples to update the belief of ensembles. In the experiments, the related results on the benefits of such adaptation for ensemble-based methods are lacking (e.g., set the adaptive parameter to be zero).", "questions": "- what is the major purpose of having Fig.1? \n- What is the explanation of the lack of the convergence in Figure 2 for optimized method? Why in the last subfigure all the curves seem to struggle  to converge?\n- How to select $\\lambda$ in the reward and the number of the ensembles in practice?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730618219602}], "openreview_url": "https://openreview.net/forum?id=S1OAqOtN5U", "arxiv_id": "2410.11234", "paper_pdf": "papers/S1OAqOtN5U.pdf", "paper_pdf_sha256": "12afde87eec63fec4b6a4eacc7d67bf41c246e020f4d8a5fd63098c16d439d36", "paper_pdf_bytes": 1755746, "paper_pdf_source": "openreview", "code_url": "https://github.com/LucasCJYSDL/Offline-RL-Kit", "code_repository": "LucasCJYSDL/Offline-RL-Kit", "code_commit": "0b15a36a7dad4ab9ed6b927c1368feca75670bca", "code_archive": "repos/S1OAqOtN5U.zip", "code_archive_sha256": "38a284549e567b6d69b1e32dcd833d9d156bd9afc96277356328ed9b49e724e8", "code_archive_bytes": 409140, "code_file_count": 56, "code_extensions": {".py": 51, ".cpp": 3, ".h": 2}, "github_disk_usage_kb": 305, "github_languages": {"Python": 225440, "C++": 23735, "Cython": 4004}, "github_archived": false, "github_pushed_at": "2025-03-19T22:18:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bayes-adaptive-monte-carlo-tree-search-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3b8CgMO5ix", "year": 2024, "status": "rejected", "title": "Model guidance via explanations turns image classifiers into segmentation models", "authors": ["Xiaoyan Yu", "Jannik Franzen", "Wojciech Samek", "Marina MC Höhne", "Dagmar Kainmueller"], "authorids": ["~Xiaoyan_Yu2", "~Jannik_Franzen1", "~Wojciech_Samek1", "~Marina_MC_Höhne1", "~Dagmar_Kainmueller2"], "authors_source": "OpenReview API", "abstract": "Heatmaps generated on inputs of image classification networks via explainable AI methods like Grad-CAM and LRP have been observed to resemble segmentations of input images in many cases. Consequently, heatmaps have also been leveraged for achieving weakly supervised segmentation with image-level supervision.\nOn the other hand, losses can be imposed on differentiable heatmaps, which has been shown to serve for (1) improving heatmaps to be more human-interpretable, (2) regularization of networks towards better generalization, (3) training diverse ensembles of networks, and (4) for explicitly ignoring confounding input features. Due to the latter use case, the paradigm of imposing losses on heatmaps is often referred to as \"Right for the right reasons\". \nWe unify these two lines of research by investigating semi-supervised segmentation as a novel use case for the Right for the Right Reasons paradigm. \nFirst, we show formal parallels between differentiable heatmap architectures and standard encoder-decoder architectures for image segmentation. \nSecond, we show that such differentiable heatmap architectures yield competitive results when trained with standard segmentation losses. \nThird, we show that such architectures allow for training with weak supervision in the form of image-level labels and small numbers of pixel-level labels, outperforming comparable encoder-decoder models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sJUYVrdZmY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3518/Reviewer_nMm3"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the weak supervision with image-level labels to achieve segmentation. It establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures commonly used for image segmentation.", "review_text": "This paper studies the weak supervision with image-level labels to achieve segmentation. It establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures commonly used for image segmentation.", "strengths": "The studied weak form supervision is interesting and helps understanding the learning of convolution neural networks.", "weaknesses": "The organization and presentation of this paper is poor, and some writing and language use is vague, the paper also lacks clear presentation of its contribution in the context of prior research work. for example, the LRP is used many times in the abstract, main text, image caption, section title, e.t.c, but without given concrete definition, which makes the paper quality poor.", "questions": "N.A.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the weak supervision with image-level labels to achieve segmentation. It establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures commonly used for image segmentation.", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "strengths": "The studied weak form supervision is interesting and helps understanding the learning of convolution neural networks.", "weaknesses": "The organization and presentation of this paper is poor, and some writing and language use is vague, the paper also lacks clear presentation of its contribution in the context of prior research work. for example, the LRP is used many times in the abstract, main text, image caption, section title, e.t.c, but without given concrete definition, which makes the paper quality poor.", "questions": "N.A.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699280993239}, {"id": "EeJ6erh1RD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3518/Reviewer_43DH"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures used for image segmentation. It conducts a comparative evaluation of these two approaches in terms of segmentation accuracy, finding that differentiable heatmap architectures, when trained with combined classification and segmentation loss, can achieve competitive segmentation performance. The authors also explore semi-supervised training with varying numbers of pixel-level labels, showing that differentiable heatmap architectures outperform standard U-Nets for segmentation in scenarios with few pixel-wise labels.", "review_text": "This paper establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures used for image segmentation. It conducts a comparative evaluation of these two approaches in terms of segmentation accuracy, finding that differentiable heatmap architectures, when trained with combined classification and segmentation loss, can achieve competitive segmentation performance. The authors also explore semi-supervised training with varying numbers of pixel-level labels, showing that differentiable heatmap architectures outperform standard U-Nets for segmentation in scenarios with few pixel-wise labels.", "strengths": "- This paper is well written and structured. I enjoyed reading this paper and find the idea quite interesting. The proposed unrolled LRP for benefits from having layer-wise guidance from the heatmaps with the skip and tied-in connections for accurate segmentation map prediction. However, it remains somewhat unclear whether this directly leads to improved segmentation performance, as indicated by the authors. It might be the case that the segmentation results are more closely tied to the quality of the heatmaps. Nevertheless, I feel that the idea is novel and is of sufficient interest to the research community.\n- The paper makes use of standard training objectives with the proposed unrolled LRP for semi-supervised segmentation, thereby making it directly usable with any task specific architectures. In general, I am in favour of simple and easy to plug-in methods that can complement already existing approaches.\n- The proposed method is well supported by experiment results. The empirical finding that concurrent training for classification and segmentation does not compromise classifier performance and holds up comparably to conventional segmentation architectures is quite intriguing. This finding suggests that the method is more generalizable to classification as well as segmentation.", "weaknesses": "- Continuining from one of my speculations I mentioned in strengths (point 1), the segmentation performance might be closely tied to the quality of heatmaps. This heatmap quality significantly based on amount of available data and the distribution of classes within a dataset. We already know that classification approaches tend to suffer in performance when there is imbalance in the number of samples per class, and I suspect that challenge may also extend to the segmentation performance. How robust is this method to such real-world scenarios with dataset and class imbalances?", "questions": "- Strengths (point 1) and weaknesses sections have an unanswered question for the authors to respond. I have listed a few more questions below.\n- Subsequently, bigger multi-label datasets with large number multiple instances per image can also create more uncertain regions the heatmaps. Is the method able to handle this?\n- How would the losses from orthogonal semi-supervised segmentation approaches affect the training with an unrolled LRP? Do you expect to see better performances?\n- There is no code currently available, will the authors make it available at some point?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper establishes formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures used for image segmentation. It conducts a comparative evaluation of these two approaches in terms of segmentation accuracy, finding that differentiable heatmap architectures, when trained with combined classification and segmentation loss, can achieve competitive segmentation performance. The authors also explore semi-supervised training with varying numbers of pixel-level labels, showing that differentiable heatmap architectures outperform standard U-Nets for segmentation in scenarios with few pixel-wise labels.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "- This paper is well written and structured. I enjoyed reading this paper and find the idea quite interesting. The proposed unrolled LRP for benefits from having layer-wise guidance from the heatmaps with the skip and tied-in connections for accurate segmentation map prediction. However, it remains somewhat unclear whether this directly leads to improved segmentation performance, as indicated by the authors. It might be the case that the segmentation results are more closely tied to the quality of the heatmaps. Nevertheless, I feel that the idea is novel and is of sufficient interest to the research community.\n- The paper makes use of standard training objectives with the proposed unrolled LRP for semi-supervised segmentation, thereby making it directly usable with any task specific architectures. In general, I am in favour of simple and easy to plug-in methods that can complement already existing approaches.\n- The proposed method is well supported by experiment results. The empirical finding that concurrent training for classification and segmentation does not compromise classifier performance and holds up comparably to conventional segmentation architectures is quite intriguing. This finding suggests that the method is more generalizable to classification as well as segmentation.", "weaknesses": "- Continuining from one of my speculations I mentioned in strengths (point 1), the segmentation performance might be closely tied to the quality of heatmaps. This heatmap quality significantly based on amount of available data and the distribution of classes within a dataset. We already know that classification approaches tend to suffer in performance when there is imbalance in the number of samples per class, and I suspect that challenge may also extend to the segmentation performance. How robust is this method to such real-world scenarios with dataset and class imbalances?", "questions": "- Strengths (point 1) and weaknesses sections have an unanswered question for the authors to respond. I have listed a few more questions below.\n- Subsequently, bigger multi-label datasets with large number multiple instances per image can also create more uncertain regions the heatmaps. Is the method able to handle this?\n- How would the losses from orthogonal semi-supervised segmentation approaches affect the training with an unrolled LRP? Do you expect to see better performances?\n- There is no code currently available, will the authors make it available at some point?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699232040419}, {"id": "UjXoLZvz3H", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3518/Reviewer_K7Hs"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper presents a method to turn image classifiers into segmentation models.", "review_text": "The paper presents a method to turn image classifiers into segmentation models.", "strengths": "It is a promising research direction to turn image classifiers directly into segmentation models, especially considering that there are many well-performed pre-trained (vision-language) classification models.", "weaknesses": "1. The paper lacks important comparisons with other weakly-supervised semantic segmentation methods that can also extract pseudo semantic masks from image-level labels.\n2. The authors claim they do not want to achieve the best semi-supervised performance, but the reported results are unacceptably too poor. And there seems not to be any ablations studies on the proposed method. It is strongly recommended to re-prepare the draft. I do not think this work has been well prepared for ICLR submission.", "questions": "Please refer to the above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method to turn image classifiers into segmentation models.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "It is a promising research direction to turn image classifiers directly into segmentation models, especially considering that there are many well-performed pre-trained (vision-language) classification models.", "weaknesses": "1. The paper lacks important comparisons with other weakly-supervised semantic segmentation methods that can also extract pseudo semantic masks from image-level labels.\n2. The authors claim they do not want to achieve the best semi-supervised performance, but the reported results are unacceptably too poor. And there seems not to be any ablations studies on the proposed method. It is strongly recommended to re-prepare the draft. I do not think this work has been well prepared for ICLR submission.", "questions": "Please refer to the above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698902842815}, {"id": "hcBLL4Og8c", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3518/Reviewer_a33j"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper optimizes heatmaps towards improved segmentation performance. The authors establish formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures for image segmentation. Their experimental results demonstrate that unrolled LRP trained with combined classification and segmentation loss, can achieve competitive segmentation performance across comparable U-Nets. And the architectures they showcased exhibit favourable outcomes in certain weakly supervised training scenarios.", "review_text": "The paper optimizes heatmaps towards improved segmentation performance. The authors establish formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures for image segmentation. Their experimental results demonstrate that unrolled LRP trained with combined classification and segmentation loss, can achieve competitive segmentation performance across comparable U-Nets. And the architectures they showcased exhibit favourable outcomes in certain weakly supervised training scenarios.", "strengths": "The paper establishes unrolled heatmap architectures as encoder-decoder-style convolutional architectures that can be trained for image segmentation.\n\nThe paper proposed the combined classification and segmentation loss and showed that differentiable heatmap architectures yield competitive results when trained with this loss.\n\nThe models proposed in this paper outperform comparable UNets in all supervision scenarios, with the performance margin increasing significantly as the level of pixel-level supervision decreases.", "weaknesses": "The paper did not conduct extensive datasets to verify the experimental results.\n\nThe contrast algorithm is restricted to U-Net and ignores relevant variants.\n\nThe paper specifically examines a limited number of cases comparing differentiable heatmap architectures with classical encoder-decoder architectures for image segmentation.\n\nThe language description is not clear, and there are some grammatical errors, such as case misuse.", "questions": "In Section 2.1 LRP BASICS of the paper, algorithm 1 should be presented with a clearer explanation.\n\nIn Section 2.2 UNROLLED LRP ARCHITECTURES FOR CONVOLUTIONAL CLASSIFIERS of the paper, what is the role of the 1x1 convolution mentioned in the \"Final classifier layers\" part?\n\nIn Section 2.3 LOSSES AND TRAINING of the paper, the different weight combinations of the loss function require more experimental support.\n\nIn Section 2.4 RELATION TO PREVIOUS FORMAL ANALYSES of the paper, “When trained with classification- and heatmap loss, the gradient of the classification loss backpropagates solely through the encoder, while the gradient of the heatmap loss backpropagates solely through the decoder. This can be leveraged for efficient training”, the phenomenon should be further validated through additional experiments or theoretical analysis in order to establish the credibility of this characteristic.\n\nIn Section 2.5 RELATION TO STANDARD ARCHITECTURES of the paper, unrolled heatmap architectures can unrolled heatmap architectures only be applied to U-Net or its related architectures? Can it be extended to a wider range of segmentation models? If not, please explain the reasons. If it can, please demonstrate its application and provide experimental comparisons with other architectures.\n\nIn Section 3 UNROLLED HEATMAP ARCHITECTURES FOR SEGMENTATION: RESULTS, please provide a more detailed description of the data selection and include validation on a wider range of datasets. Additionally, please include more comparative analysis regarding the improved U-Net models in the Quantitative Results part.\n\nIn Section 3 UNROLLED HEATMAP ARCHITECTURES FOR SEGMENTATION: RESULTS, “the ResNet18 UNet is outperformed by the ResNet50 UNet” ,the conclusion lacks experimental data support.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper optimizes heatmaps towards improved segmentation performance. The authors establish formal parallels between differentiable heatmap architectures and conventional encoder-decoder architectures for image segmentation. Their experimental results demonstrate that unrolled LRP trained with combined classification and segmentation loss, can achieve competitive segmentation performance across comparable U-Nets. And the architectures they showcased exhibit favourable outcomes in certain weakly supervised training scenarios.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "The paper establishes unrolled heatmap architectures as encoder-decoder-style convolutional architectures that can be trained for image segmentation.\n\nThe paper proposed the combined classification and segmentation loss and showed that differentiable heatmap architectures yield competitive results when trained with this loss.\n\nThe models proposed in this paper outperform comparable UNets in all supervision scenarios, with the performance margin increasing significantly as the level of pixel-level supervision decreases.", "weaknesses": "The paper did not conduct extensive datasets to verify the experimental results.\n\nThe contrast algorithm is restricted to U-Net and ignores relevant variants.\n\nThe paper specifically examines a limited number of cases comparing differentiable heatmap architectures with classical encoder-decoder architectures for image segmentation.\n\nThe language description is not clear, and there are some grammatical errors, such as case misuse.", "questions": "In Section 2.1 LRP BASICS of the paper, algorithm 1 should be presented with a clearer explanation.\n\nIn Section 2.2 UNROLLED LRP ARCHITECTURES FOR CONVOLUTIONAL CLASSIFIERS of the paper, what is the role of the 1x1 convolution mentioned in the \"Final classifier layers\" part?\n\nIn Section 2.3 LOSSES AND TRAINING of the paper, the different weight combinations of the loss function require more experimental support.\n\nIn Section 2.4 RELATION TO PREVIOUS FORMAL ANALYSES of the paper, “When trained with classification- and heatmap loss, the gradient of the classification loss backpropagates solely through the encoder, while the gradient of the heatmap loss backpropagates solely through the decoder. This can be leveraged for efficient training”, the phenomenon should be further validated through additional experiments or theoretical analysis in order to establish the credibility of this characteristic.\n\nIn Section 2.5 RELATION TO STANDARD ARCHITECTURES of the paper, unrolled heatmap architectures can unrolled heatmap architectures only be applied to U-Net or its related architectures? Can it be extended to a wider range of segmentation models? If not, please explain the reasons. If it can, please demonstrate its application and provide experimental comparisons with other architectures.\n\nIn Section 3 UNROLLED HEATMAP ARCHITECTURES FOR SEGMENTATION: RESULTS, please provide a more detailed description of the data selection and include validation on a wider range of datasets. Additionally, please include more comparative analysis regarding the improved U-Net models in the Quantitative Results part.\n\nIn Section 3 UNROLLED HEATMAP ARCHITECTURES FOR SEGMENTATION: RESULTS, “the ResNet18 UNet is outperformed by the ResNet50 UNet” ,the conclusion lacks experimental data support.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698641204620}], "openreview_url": "https://openreview.net/forum?id=3b8CgMO5ix", "arxiv_id": "2407.03009", "paper_pdf": "papers/3b8CgMO5ix.pdf", "paper_pdf_sha256": "f5c1193a783f1239784eab92b25fc01bd7a8ea14bae458855b71d0658f8d7466", "paper_pdf_bytes": 12292941, "paper_pdf_source": "openreview", "code_url": "https://github.com/Kainmueller-Lab/TW-autoencoder", "code_repository": "Kainmueller-Lab/TW-autoencoder", "code_commit": "01d3932a173a4504f4bfe9638864cb46da607e63", "code_archive": "repos/3b8CgMO5ix.zip", "code_archive_sha256": "d1c0ac05f0c7f125536f2090c1c1f0f6ac512ee51d994e1d7623a4fe9000636a", "code_archive_bytes": 75742, "code_file_count": 46, "code_extensions": {".py": 45, ".sh": 1}, "github_disk_usage_kb": 252, "github_languages": {"Python": 222916, "Shell": 953}, "github_archived": false, "github_pushed_at": "2024-07-18T11:09:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/model-guidance-via-explanations-turns-image"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "deit1AdsFU", "year": 2023, "status": "rejected", "title": "Learning To Invert: Simple Adaptive Attacks for Gradient Inversion in Federated Learning", "authors": ["Ruihan Wu", "Xiangyu Chen", "Chuan Guo", "Kilian Q Weinberger"], "authorids": ["~Ruihan_Wu1", "~Xiangyu_Chen1", "~Chuan_Guo1", "~Kilian_Q_Weinberger1"], "authors_source": "OpenReview API", "abstract": "Gradient inversion attack enables recovery of training samples from model updates in federated learning (FL) and constitutes a serious threat to data privacy. To mitigate this vulnerability, prior work proposed both principled defenses based on differential privacy, as well as heuristic defenses based on gradient compression as countermeasures. These defenses have so far been very effective, in particular those based on gradient compression that allow the model to maintain high accuracy while greatly reducing the attack's effectiveness. In this work, we argue that such findings do not accurately reflect the privacy risk in FL, and show that existing defenses can be broken by a simple adaptive attack that trains a model using auxiliary data to learn how to invert gradients on both vision and language tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "_MoS6jnFTI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3979/Reviewer_SYoc"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes to recover training data from gradient updates in federated learning. Specifically, it leverages an auxiliary dataset to obtain the gradients for these samples. A neural network model is then trained on them by mapping the gradient to the input. This paper also utilizes existing feature hashing method to reduce the dimensionality of the input gradient. The evaluation is conducted on one image classification dataset and one language task. The experimental results show the proposed method has better attack performance compared to existing techniques against different defense techniques.", "review_text": "This paper studies an important problem. However, the proposed method is based on unrealistic and impractical assumptions and settings. The technical novelty is also limited.\n", "strengths": "### Strength\n\n+ Important topic of gradient inversion attack\n+ Easy to follow\n\n\n### Weaknesses\n\n- Unrealistic assumption\n- Limited novelty\n- Impractical evaluation setup\n- Limited evaluation\n\n\n### Detailed comments\n\n* In the threat model, the paper assumes \"the FL protocol does not leverage secure aggregation\", which is an unrealistic assumption. The purpose of secure aggregation is to avoid revealing private information to other parties in a distributed setting such as federated learning. The paper discards the secure aggregation and builds an attack on top of such an unrealistic assumption. In addition, it assumes the gradient for each sample in the batch (actually it assumes the batch size is 1), which is impractical. It is not meaningful to study the problem in such a setting.\n\n* This paper leverages an auxiliary dataset to train a model for mapping the gradient to the input. Similar ideas have already been studied in many existing works. The dimensionality reduction issue is addressed by directly using an existing feature hashing technique. There is very limited technical contribution in this paper.\n\n* The paper uses the term \"auxiliary dataset\" to denote a dataset different from the training data by the subject model. However, in the evaluation, the paper directly uses the training data that the subject was trained on to construct their gradient inversion model, which is impractical.\n\n* The evaluation is conducted on both computer vision and natural language processing, which is good. However, only one dataset in each domain is evaluated. The model used on the computer vision task is LeNet, which is a very simple model structure. There is no evaluation on advanced and complex model structures.\n\n* Important details are missing in the paper. There is no data showing the performance of the subject model on corresponding tasks. The baseline GI-GIP uses ImageNet to train the generator in the original paper. It is unclear whether the authors directly use the trained generator from GI-GIP or retrain the generator on CIFAR-10. If it is the former case, the comparison is not considered fair.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes to recover training data from gradient updates in federated learning. Specifically, it leverages an auxiliary dataset to obtain the gradients for these samples. A neural network model is then trained on them by mapping the gradient to the input. This paper also utilizes existing feature hashing method to reduce the dimensionality of the input gradient. The evaluation is conducted on one image classification dataset and one language task. The experimental results show the proposed method has better attack performance compared to existing techniques against different defense techniques.", "strength_and_weaknesses": "### Strength\n\n+ Important topic of gradient inversion attack\n+ Easy to follow\n\n\n### Weaknesses\n\n- Unrealistic assumption\n- Limited novelty\n- Impractical evaluation setup\n- Limited evaluation\n\n\n### Detailed comments\n\n* In the threat model, the paper assumes \"the FL protocol does not leverage secure aggregation\", which is an unrealistic assumption. The purpose of secure aggregation is to avoid revealing private information to other parties in a distributed setting such as federated learning. The paper discards the secure aggregation and builds an attack on top of such an unrealistic assumption. In addition, it assumes the gradient for each sample in the batch (actually it assumes the batch size is 1), which is impractical. It is not meaningful to study the problem in such a setting.\n\n* This paper leverages an auxiliary dataset to train a model for mapping the gradient to the input. Similar ideas have already been studied in many existing works. The dimensionality reduction issue is addressed by directly using an existing feature hashing technique. There is very limited technical contribution in this paper.\n\n* The paper uses the term \"auxiliary dataset\" to denote a dataset different from the training data by the subject model. However, in the evaluation, the paper directly uses the training data that the subject was trained on to construct their gradient inversion model, which is impractical.\n\n* The evaluation is conducted on both computer vision and natural language processing, which is good. However, only one dataset in each domain is evaluated. The model used on the computer vision task is LeNet, which is a very simple model structure. There is no evaluation on advanced and complex model structures.\n\n* Important details are missing in the paper. There is no data showing the performance of the subject model on corresponding tasks. The baseline GI-GIP uses ImageNet to train the generator in the original paper. It is unclear whether the authors directly use the trained generator from GI-GIP or retrain the generator on CIFAR-10. If it is the former case, the comparison is not considered fair.", "clarity,_quality,_novelty_and_reproducibility": "### Clarity, Quality, Novelty\n\nPlease see detailed comments in Strength And Weaknesses.\n\n### Reproducibility\n\nThe submission includes the code. However the baselines are missing from the code. Without the aforementioned details regarding baselines, it is hard to reproduce the results.", "summary_of_the_review": "This paper studies an important problem. However, the proposed method is based on unrealistic and impractical assumptions and settings. The technical novelty is also limited.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667349721146}, {"id": "A_5tmJqWp_j", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3979/Reviewer_n8di"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper investigates potential privacy risk in federated learning. They found that existing privacy defenses in FL can be broken via a simple adaptive attack. In particular, the proposed learning-based approach (Learning to invert) aims to train an inversion model to reconstruct training samples from their gradient with the help from auxiliary dataset.  Experiments demonstrate the effectiveness of the proposed model. \n", "review_text": "Strength:\n1. The research problem is very important and this paper provides a simple but effective attacking method to conduct gradient inversion attacks in federated learning. \n2. This paper is easy to follow and clearly written.  \n3. The experimental results demonstrate the effectiveness of the proposed method. \n\n\nWeaknesses\n\nThere are some concerns regarding this paper.\n\n1. Auxiliary datasets are used to help learn the inversion model for privacy attacks in FL. In such cases, this task can be considered as a malicious server that wants to steal private information from some clients. The attacker task is weird to me. To steal information from one client, why not just a serve and a client pair, and attack this client directly? Such an attacking strategy is not hard to achieve. \n\n2. The batch size is set to 1. When the batch size increases, the proposed method can perform much worse. This may limit their applications to many real-world tasks, where many FL methods would consider using large batch sizes. Meanwhile, it's somehow unfair compared with other privacy attacks in FL.\n", "strengths": "Strength:\n1. The research problem is very important and this paper provides a simple but effective attacking method to conduct gradient inversion attacks in federated learning. \n2. This paper is easy to follow and clearly written.  \n3. The experimental results demonstrate the effectiveness of the proposed method. \n\n\nWeaknesses\n\nThere are some concerns regarding this paper.\n\n1. Auxiliary datasets are used to help learn the inversion model for privacy attacks in FL. In such cases, this task can be considered as a malicious server that wants to steal private information from some clients. The attacker task is weird to me. To steal information from one client, why not just a serve and a client pair, and attack this client directly? Such an attacking strategy is not hard to achieve. \n\n2. The batch size is set to 1. When the batch size increases, the proposed method can perform much worse. This may limit their applications to many real-world tasks, where many FL methods would consider using large batch sizes. Meanwhile, it's somehow unfair compared with other privacy attacks in FL.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper investigates potential privacy risk in federated learning. They found that existing privacy defenses in FL can be broken via a simple adaptive attack. In particular, the proposed learning-based approach (Learning to invert) aims to train an inversion model to reconstruct training samples from their gradient with the help from auxiliary dataset.  Experiments demonstrate the effectiveness of the proposed model. \n", "strength_and_weaknesses": "Strength:\n1. The research problem is very important and this paper provides a simple but effective attacking method to conduct gradient inversion attacks in federated learning. \n2. This paper is easy to follow and clearly written.  \n3. The experimental results demonstrate the effectiveness of the proposed method. \n\n\nWeaknesses\n\nThere are some concerns regarding this paper.\n\n1. Auxiliary datasets are used to help learn the inversion model for privacy attacks in FL. In such cases, this task can be considered as a malicious server that wants to steal private information from some clients. The attacker task is weird to me. To steal information from one client, why not just a serve and a client pair, and attack this client directly? Such an attacking strategy is not hard to achieve. \n\n2. The batch size is set to 1. When the batch size increases, the proposed method can perform much worse. This may limit their applications to many real-world tasks, where many FL methods would consider using large batch sizes. Meanwhile, it's somehow unfair compared with other privacy attacks in FL.\n", "clarity,_quality,_novelty_and_reproducibility": "N.A", "summary_of_the_review": "Strength:\n1. The research problem is very important and this paper provides a simple but effective attacking method to conduct gradient inversion attacks in federated learning. \n2. This paper is easy to follow and clearly written.  \n3. The experimental results demonstrate the effectiveness of the proposed method. \n\n\nWeaknesses\n\nThere are some concerns regarding this paper.\n\n1. Auxiliary datasets are used to help learn the inversion model for privacy attacks in FL. In such cases, this task can be considered as a malicious server that wants to steal private information from some clients. The attacker task is weird to me. To steal information from one client, why not just a serve and a client pair, and attack this client directly? Such an attacking strategy is not hard to achieve. \n\n2. The batch size is set to 1. When the batch size increases, the proposed method can perform much worse. This may limit their applications to many real-world tasks, where many FL methods would consider using large batch sizes. Meanwhile, it's somehow unfair compared with other privacy attacks in FL.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667114110654}, {"id": "xsiKcWraKJw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3979/Reviewer_6VqA"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper introduces a new gradient inversion method in Federated Learning. The proposed approach called \"Learning To Invert\" (LTI) directly attempts to learn the mapping between the gradient of a sample and the corresponding input sample using an auxiliary dataset.  A simple multi-layer perceptron (MLP) is used to learn this mapping. Dimensionality reduction is applied to the gradients to make the MLP practically feasible. The paper asserts that current defense mechanisms such as Sign Compression, Gradient Pruning, and Gaussian Perturbation cannot defend against the proposed attack, because the same transformations can be applied by the server before it learns to reconstruct. The method is evaluated on vision and language tasks (CIFAR10, WikiText).", "review_text": "The paper tackles an important problem in FL, but the assumptions involved are unrealistic, the novelty is limited, and the experiments are not comprehensive.", "strengths": "Strengths: \n1.\tEvaluation of the proposed method for both vision and language tasks. \n2.\tComparison of LTI with current SOTA gradient inversion attacks.\n3.\tAblation results based on the auxiliary dataset (both degree of overlap and size of the dataset).\n\nWeaknesses:\n\n1.    The attack works only in the presence of an auxiliary dataset (which has some overlap with the client distributions) at the server. Hence, the comparison with SOTA gradient inversion methods is not fair. For a fair comparison, other inversion methods should be updated to make use of the auxiliary dataset. When $\\beta=0$ (no overlap between the auxiliary and training sets), the MSE of the proposed method is the worst. Note that even for $\\beta=0.01$, there will be significant number of common samples between the two sets, which explains the superiority of the proposed method. Ideally, Tables 1 & 2 and Figures 2 & 4 should be reported for low values of $beta$ (< 0.1) or with an auxiliary dataset that is completely different (some other natural image and NLP dataset that is different from the training).\n\n2.  The main experimental setup is designed such that the training samples and samples in the auxiliary dataset are the SAME. Furthermore, the gradients are computed based on a single sample. So, effectively there is one-to-one mapping between a sample and its gradient, which any network can easily learn (all one needs is a indexing table!). So, the results in Table 1 and Figure 2 are unsurprising - in fact, it is somewhat underwhelming because exact reconstruction should be possible in this setting unless there is loss of some information in the dimensionality reduction step. \n\n3. There is no information about how the proposed approach will scale to higher fidelity data (say, images of 224 x 224 x 3 resolution). The MLP approach is likely to become practically infeasible for higher fidelity data.\n\n4. The other key aspect that is missing is the FL round in which the reconstruction is attempted. Several works in the literature have shown that reconstruction is easier in the first few rounds (when training from scratch), while it becomes harder in the later rounds. That is why multiple rounds of local training is typically carried out before the collaboration starts and this serves as a good defense against gradient inversion attacks. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper introduces a new gradient inversion method in Federated Learning. The proposed approach called \"Learning To Invert\" (LTI) directly attempts to learn the mapping between the gradient of a sample and the corresponding input sample using an auxiliary dataset.  A simple multi-layer perceptron (MLP) is used to learn this mapping. Dimensionality reduction is applied to the gradients to make the MLP practically feasible. The paper asserts that current defense mechanisms such as Sign Compression, Gradient Pruning, and Gaussian Perturbation cannot defend against the proposed attack, because the same transformations can be applied by the server before it learns to reconstruct. The method is evaluated on vision and language tasks (CIFAR10, WikiText).", "strength_and_weaknesses": "Strengths: \n1.\tEvaluation of the proposed method for both vision and language tasks. \n2.\tComparison of LTI with current SOTA gradient inversion attacks.\n3.\tAblation results based on the auxiliary dataset (both degree of overlap and size of the dataset).\n\nWeaknesses:\n\n1.    The attack works only in the presence of an auxiliary dataset (which has some overlap with the client distributions) at the server. Hence, the comparison with SOTA gradient inversion methods is not fair. For a fair comparison, other inversion methods should be updated to make use of the auxiliary dataset. When $\\beta=0$ (no overlap between the auxiliary and training sets), the MSE of the proposed method is the worst. Note that even for $\\beta=0.01$, there will be significant number of common samples between the two sets, which explains the superiority of the proposed method. Ideally, Tables 1 & 2 and Figures 2 & 4 should be reported for low values of $beta$ (< 0.1) or with an auxiliary dataset that is completely different (some other natural image and NLP dataset that is different from the training).\n\n2.  The main experimental setup is designed such that the training samples and samples in the auxiliary dataset are the SAME. Furthermore, the gradients are computed based on a single sample. So, effectively there is one-to-one mapping between a sample and its gradient, which any network can easily learn (all one needs is a indexing table!). So, the results in Table 1 and Figure 2 are unsurprising - in fact, it is somewhat underwhelming because exact reconstruction should be possible in this setting unless there is loss of some information in the dimensionality reduction step. \n\n3. There is no information about how the proposed approach will scale to higher fidelity data (say, images of 224 x 224 x 3 resolution). The MLP approach is likely to become practically infeasible for higher fidelity data.\n\n4. The other key aspect that is missing is the FL round in which the reconstruction is attempted. Several works in the literature have shown that reconstruction is easier in the first few rounds (when training from scratch), while it becomes harder in the later rounds. That is why multiple rounds of local training is typically carried out before the collaboration starts and this serves as a good defense against gradient inversion attacks. ", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written. However, the novelty is limited and the experiments are not comprehensive.There are no concerns about the reproducibility.", "summary_of_the_review": "The paper tackles an important problem in FL, but the assumptions involved are unrealistic, the novelty is limited, and the experiments are not comprehensive.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667046318772}, {"id": "pYW3jqF9f6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3979/Reviewer_8QEs"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a simple learning-based gradient inversion attack. The new method is trained using auxiliary data and can learn how to invert gradients on both vision and language tasks.  A new learning function is built to learn model parameters on the auxiliary data.  ", "review_text": "See the above. ", "strengths": "1. To learn the parameter theta, a large number of auxiliary data are required. This is also a limitation of the proposed method. A challenging problem would be reduce the number of the auxiliary data while keeping the performance. \n2. The learning function of eq.(2) is simple yet powerful. A new algorithm is expected to show the learning steps with data input/output and all the parameters used in the algorithm. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper presents a simple learning-based gradient inversion attack. The new method is trained using auxiliary data and can learn how to invert gradients on both vision and language tasks.  A new learning function is built to learn model parameters on the auxiliary data.  ", "strength_and_weaknesses": "1. To learn the parameter theta, a large number of auxiliary data are required. This is also a limitation of the proposed method. A challenging problem would be reduce the number of the auxiliary data while keeping the performance. \n2. The learning function of eq.(2) is simple yet powerful. A new algorithm is expected to show the learning steps with data input/output and all the parameters used in the algorithm. \n", "clarity,_quality,_novelty_and_reproducibility": "The codes and data are available. ", "summary_of_the_review": "See the above. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666671646625}, {"id": "r9-bf1GbltP", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3979/Reviewer_PX2s"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new gradient inversion attack against federated learning. By assuming that the server has access to some auxiliary data, the main idea of the paper is to learn a gradient inversion model from the auxiliary data. The paper shows that compared with previous optimization-based approaches such as Inverting Gradients (IG), Gradient Inversion with Generative Image Prior (GI-GIP), and TAG, the proposed learning to invert (LTI) method obtains better reconstruction accuracy on both an image task (CIFAR-10) and a language task (Wikitext), under both gradient perturbation and compression based defenses. ", "review_text": "The paper departs from existing optimization-based methods and proposes a simple learning-based approach to gradient inversion against federated learning. The approach is effective in the simple setting when the batch size is 1, and the server adopts gradient perturbation or gradient compression-based defenses. It is unclear if it can be applied to practical FL systems with larger batch sizes and strong defenses. Thus, the claim that existing defenses provide a false sense of security may not hold. ", "strengths": "Strengths\n\n1. The idea of learning an inversion model using auxiliary data seems a simple yet effective approach. \n2. Ablation studies show that even when the auxiliary data only consists of 500 images sampled from CIFAR-10 or contains 250 in-distribution samples, LTI still outperforms IG and GI-GIP, which is impressive. \n\nWeaknesses\n\n1. An important limitation of the proposed approach is that it only works for gradients computed from a single data sample. For a real FL system, even without using secure aggregation, a local update from a client is obtained by taking the gradient of a batch of data samples or through multiple gradient descent steps. Thus, the proposed approach is not sophisticated enough to be applied to real FL systems. \n2. Another limitation is that the proposed attack method is only tested against gradient perturbation and gradient compression, which were not originally designed for countering gradient inversion. It would be useful to understand how the proposed method performs against more recent defenses, such as Soteria [1], that target gradient inversion. \n\n[1] Jingwei Sun, Ang Li, Binghui Wang, Huanrui Yang, Hai Li, and Yiran Chen. Soteria: Provable defense against privacy leakage in federated learning from representation perspective. CVPR 2021.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new gradient inversion attack against federated learning. By assuming that the server has access to some auxiliary data, the main idea of the paper is to learn a gradient inversion model from the auxiliary data. The paper shows that compared with previous optimization-based approaches such as Inverting Gradients (IG), Gradient Inversion with Generative Image Prior (GI-GIP), and TAG, the proposed learning to invert (LTI) method obtains better reconstruction accuracy on both an image task (CIFAR-10) and a language task (Wikitext), under both gradient perturbation and compression based defenses. ", "strength_and_weaknesses": "Strengths\n\n1. The idea of learning an inversion model using auxiliary data seems a simple yet effective approach. \n2. Ablation studies show that even when the auxiliary data only consists of 500 images sampled from CIFAR-10 or contains 250 in-distribution samples, LTI still outperforms IG and GI-GIP, which is impressive. \n\nWeaknesses\n\n1. An important limitation of the proposed approach is that it only works for gradients computed from a single data sample. For a real FL system, even without using secure aggregation, a local update from a client is obtained by taking the gradient of a batch of data samples or through multiple gradient descent steps. Thus, the proposed approach is not sophisticated enough to be applied to real FL systems. \n2. Another limitation is that the proposed attack method is only tested against gradient perturbation and gradient compression, which were not originally designed for countering gradient inversion. It would be useful to understand how the proposed method performs against more recent defenses, such as Soteria [1], that target gradient inversion. \n\n[1] Jingwei Sun, Ang Li, Binghui Wang, Huanrui Yang, Hai Li, and Yiran Chen. Soteria: Provable defense against privacy leakage in federated learning from representation perspective. CVPR 2021.\n", "clarity,_quality,_novelty_and_reproducibility": "Multiple global models are generated during the FL training process.  An important detail that is unclear is which of them are used to train the gradient inversion module and which of them are used to evaluate the attacks and defenses. It is unlikely that the same model can be used for both purposes in practice, given the amount of time needed to train the gradient inversion module. Thus, the question is whether a module trained using early FL models can be effective for the inversion task against new FL models.  A related question is whether it is easier or harder to perform inversion attacks at the earlier stage of FL training. \n\nThe idea of training a gradient-inversion module using a small amount of auxiliary data seems novel.", "summary_of_the_review": "The paper departs from existing optimization-based methods and proposes a simple learning-based approach to gradient inversion against federated learning. The approach is effective in the simple setting when the batch size is 1, and the server adopts gradient perturbation or gradient compression-based defenses. It is unclear if it can be applied to practical FL systems with larger batch sizes and strong defenses. Thus, the claim that existing defenses provide a false sense of security may not hold. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666650372460}], "openreview_url": "https://openreview.net/forum?id=deit1AdsFU", "arxiv_id": "2210.10880", "paper_pdf": "papers/deit1AdsFU.pdf", "paper_pdf_sha256": "a15867ba5dafaf263010a24527d5a2eb716e3be31c8b322c251d1e1668f71843", "paper_pdf_bytes": 8404531, "paper_pdf_source": "openreview", "code_url": "https://github.com/wrh14/Learning_to_Invert", "code_repository": "wrh14/Learning_to_Invert", "code_commit": "a3464c58166967ce12c917d368e03e8129fb60a6", "code_archive": "repos/deit1AdsFU.zip", "code_archive_sha256": "f5ec677a330f0bfdf83db62d47c72f36476d0bd8723714671400df5c0ba8385a", "code_archive_bytes": 637585, "code_file_count": 68, "code_extensions": {".py": 66, ".ipynb": 2}, "github_disk_usage_kb": 528, "github_languages": {"Python": 606402, "Jupyter Notebook": 10915}, "github_archived": false, "github_pushed_at": "2024-06-15T03:55:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-to-invert-simple-adaptive-attacks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FFGDKzLasUa", "year": 2022, "status": "rejected", "title": "Stochastic Deep Networks with Linear Competing Units for Model-Agnostic Meta-Learning", "authors": ["Konstantinos Ι. Kalais", "Sotirios Chatzis"], "authorids": ["~Konstantinos_Ι._Kalais1", "~Sotirios_Chatzis1"], "authors_source": "OpenReview API", "abstract": "This work addresses meta-learning (ML) by considering deep networks with stochastic local winner-takes-all (LWTA) activations. This type of network units result in sparse representations from each model layer, as the units are organized into blocks where only one unit generates a non-zero output. The main operating principle of the introduced units lies on stochastic arguments, as the network performs posterior sampling over competing units to select the winner. Therefore, the proposed networks are explicitly designed to extract input data representations of sparse stochastic nature, as opposed to the currently standard deterministic representation paradigm. We posit that these modeling arguments, inspired from Bayesian statistics, allow for more robust modeling when uncertainty is high due to the limited availability of task-related training data; this is exactly the case with ML, which is the focus of this work. At training time, we rely on the reparameterization trick for Discrete distributions to perform reliable training via Monte-Carlo sampling. At inference time, we rely on Bayesian Model Averaging, which effectively averages over a number of sampled representations. As we experimentally show, our approach produces state-of-the-art predictive accuracy on standard few-shot image classification benchmarks; this is achieved without compromising computational efficiency.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AhB8PgRSyc1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper47/Reviewer_czsy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a stochastic local winner-takes-all (SLWTA) approach to learn data-driven (stochastic) sparsity. This is motivated by the limited availability of training data, especiially in few-shot meta learning scenarios. The authors propose a meta-learning algorithm for their SLWTA approach. As this is a probabilistic approach, one can draw from the approximate posterior at inference to obtained the posterior predictive distribution. The authors compare their method against MAML, FOMAML and Reptile on the Omniglot and Mini-Imagenet displaying better performance than the origiinal MAML algorithm.", "review_text": "[Strengths]\n\n* The idea to embed data-driven/learned sparsity in a meta-learning framework is really interesting. These methods have been shown to improve robustness, are open to continual learning and have been shown to enable learning independent mechanisms. It is a nice idea and this is backed up by competitive performance of the algorithm compared to original instantiations of MAML.\n\n* The paper is well written and clear. As the reviewer of this paper, I feel that I have sufficient information and understanding to implement it myself\n\n[Weaknesses]\n\n* I find the methodological advances proposed in the manuscript too incremental for publication at ICLR. The SLWTA method proposed within appears to be the method of Panousis et al. (2021) (cited in text) but adapted to the MAML paradigm.\n\n* The probabilistic model introduced by the authors bares many simiilarities with other data-driven sparsity models that are inline with Panousis et al. (2021). There is:\n1. https://arxiv.org/abs/1805.10896 - Adaptive Network Sparsification with Dependent Variational Beta-Bernoulli Dropout\n2. https://arxiv.org/abs/1912.02290 - Hierarchical Indian Buffet Neural Networks for Bayesian Continual Learning\nBoth who learn modularity with a similar ELBO as proposed by the authors. \n\n* I think there are too few competiting baselines. Why were more recent gradient-based ML algorithms not considered such as Bayesian MAML (https://arxiv.org/abs/1806.03836) which first framed MAML in a Bayesian perspective?\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a stochastic local winner-takes-all (SLWTA) approach to learn data-driven (stochastic) sparsity. This is motivated by the limited availability of training data, especiially in few-shot meta learning scenarios. The authors propose a meta-learning algorithm for their SLWTA approach. As this is a probabilistic approach, one can draw from the approximate posterior at inference to obtained the posterior predictive distribution. The authors compare their method against MAML, FOMAML and Reptile on the Omniglot and Mini-Imagenet displaying better performance than the origiinal MAML algorithm.", "main_review": "[Strengths]\n\n* The idea to embed data-driven/learned sparsity in a meta-learning framework is really interesting. These methods have been shown to improve robustness, are open to continual learning and have been shown to enable learning independent mechanisms. It is a nice idea and this is backed up by competitive performance of the algorithm compared to original instantiations of MAML.\n\n* The paper is well written and clear. As the reviewer of this paper, I feel that I have sufficient information and understanding to implement it myself\n\n[Weaknesses]\n\n* I find the methodological advances proposed in the manuscript too incremental for publication at ICLR. The SLWTA method proposed within appears to be the method of Panousis et al. (2021) (cited in text) but adapted to the MAML paradigm.\n\n* The probabilistic model introduced by the authors bares many simiilarities with other data-driven sparsity models that are inline with Panousis et al. (2021). There is:\n1. https://arxiv.org/abs/1805.10896 - Adaptive Network Sparsification with Dependent Variational Beta-Bernoulli Dropout\n2. https://arxiv.org/abs/1912.02290 - Hierarchical Indian Buffet Neural Networks for Bayesian Continual Learning\nBoth who learn modularity with a similar ELBO as proposed by the authors. \n\n* I think there are too few competiting baselines. Why were more recent gradient-based ML algorithms not considered such as Bayesian MAML (https://arxiv.org/abs/1806.03836) which first framed MAML in a Bayesian perspective?\n\n\n", "summary_of_the_review": "I like the idea to embed an stochastic LWTA approach into meta-learning. It makes sense and the motivations are well presented. However, the technical novelty of the paper is lacking and does not offer enough methodological advancement to be a strong candidate for acceptance. \n\n**** Post-rebuttal ****\nI have decided to increase my score from 3 to 5. My reasoning is in my response to the author rebuttal.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635977014305}, {"id": "ARBpPTI_aJL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper47/Reviewer_3Mcm"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes to use neural networks with so-called \"stochastic local winner-takes-all (LWTA)\" activations in the place of standard ReLU activations. These stochastic LWTA activations result in a model with sparse representations. The sparse gating is treated as a random variable. The authors also treat the weights of the neural network as random variables. Both sets of random variables are learnt using variational inference.", "review_text": "## Strengths\n\n* Strong empiracle results \n\n* Great set of ablation studies\n\n## Weaknesses\n\n### Lack of novelty\n\nThere already exist many probabilistic approaches to meta-learning which take uncertainty into account, which the authors have failed to acknowledge. Thus their claims of novelty in this regard are incorrect. For example, the statement \"these innovations constitute a radical departure from the currently prominent design paradigm in ML\" is far too strong a statement given the work of Finn et al. (2018), Nalisnick et al. (2021), Chen et al. (2020), and Gordon et al. (2019).\n\n### Somewhat narrow set of main experiments\n\nThe main set of results in the paper are based only on the Omniglot 20-way and Mini-imagenet 5-way tasks. While the results on these tasks are good, more tasks are required to paint an accurate picture of the method. Cifar100 few shot would be a good start, but even better would be a non-image classification task. I might suggest the ShapeNet View Reconstruction task as used in Gordon et al. (2019). \n\nFurthermore, at least some of the methods from Finn et al. (2018), Nalisnick et al. (2021), Chen et al. (2020), and Gordon et al. (2019) should be compared against. \n\nAn additional ablation study comparing the performance of Stochastic LWTA and standard LWTA would be informative. \n\nFor the effect of sample size ablation, plots with performance on the y-axis and B on the x-axis for B in {1,...,10} would be very interesting as they would hopefully show a \"knee-point\" at which it stops being worthwhile to increase B.\n\n### Lack of clarity\n\nOverall I found the presentation to be confusing in many places and occasionally incorrect. I will list the issues I spotted below. \n\n1. The authors frequently refer to \"stochastic arguments\", \"modelling arguments\", \"deep network arguments\", etc. However, the usage of the word \"argument\" in these contexts does not make sense to me and I am unsure exactly what is meant.\n\n2. The authors frequently use the word \"inference\" to refer to test time prediction. However, this word also refers to the probabilistic inference for random variables. Given that the authors' proposed method makes use of such inference this becomes somewhat confusing. I would recommend only using inference in the probabilistic sense. \n\n3. \\vec{y}_r is introduced in sec 2.2 but it is not mentioned until much later in sec 3.1 that \\vec{y} is the concatenation of \\vec{y}_r. \n\n4. It is not clear why it is a good idea to use a fixed competition function that always chooses the maximum activation. It would be great if the authors could provide some motivation/intuition for this choice. (As I mention below, I think that the Sparse-MoE approach with an input-dependent gating network makes more sense). \n\n5. It is not clear what the authors mean by \"We posit that this novel stochastic paradigm will give rise to more **potent** learned representations ...\" (emphasis my own). (This is another example of a claim that overstates the novelty of this work.)\n\n6. (Minor) I assume that the authors mean \"model-agnostic ML\" rather than \"meta-agnostic ML\" at the start of sec 3.1. \n\n7. (Minor) The authors introduce the symbols A_1 and A_2 to indicate the input and output dimensions of their linear units. Would I and O not be more clear? \n\n8. Paragraphs 2 and 3 are largely redundant as they mainly contain repeated information from sec 2.2.\n\n9. (Minor) In eqn 9, the notation [\\vec{y}]_{r, j} is used, but just above the notation y = [\\vec{y}_{r, j}]. I found this somewhat confusing. \n\n10. In the paragraph below eon 9, the authors first say that they \"postulate\" that the indicator variable is drawn from a Discrete distribution. Later in the paragraph they say that they \"stipulate\" this. Which one is it?\n\n11. Speaking of the \"Discrete\" probability distribution, would it not be more clear to refer to a \"Categorical\" probability distribution?  \n\n12. In the paragraph after eqn 10, the authors state that they impose a posterior density over the weights. However, it is more accurate to say that the prior density is impose and the posterior is inferred. \n\n13. The authors frequently refer to \"stochasticity\" as increasing the generalisation capacity. However, it should be noted that simply adding stochasticity does not help. It must be added in a principled and theoretically justified manner. \n\n14. \"This concludes the formulation of the proposed model-agnostic ML approach.\" – I suspect this line should be at the end of sec 3.3, rather than sec 3.1.\n\n15. The authors refer to independence among layers in the context of the mean-field approximation. I assume that this was meant to be independence among weights? \n\n16. Eqn 13, for the Gumbel-softmax does not show the purpose of the temperature \\tau.\n\n17. In algorithm 2, line 2, the authors suggest that the variational parameters {\\mu, \\sigma} should be sampled. I suspect that the authors mean \\theta or \\vec{w}, should be sampled? \n\n17. In algorithm 2, the authors should show the equation for the BMA. \n\n18. At the top of page 7, the authors refer to \"2 competing units per block\" and then \"2 neurons per block\". Are neurons the same as competing units? \n\n19. \"For the outer-loop, we use SGD with a linear annealed outer step size equal to 0.\" Should this actually be 0? \n\n20. What is the architecture that was used for the baseline methods? \n\n21. The authors claim a 10% improvement over the \"best alternatives\" however, I think it is a 10% improvement over the *worst* alternatives. \n\n22. In table 4, it surprises me that Reptile runs slower than MAML. I don't think this makes sense. Also in table 4, the best value in the \"inference\" column should be bolded. \n\n23. In table 5, does the parameter count for LWTA include both variational parameters \\mu and \\sigma^2? OR is it just \\theta? \n\n24. More motivation is needed for the statement \"It appears that convergence of our approach is even smoother; ...\". \n\n25. Figure 2b, should be replaced with Figure 2a but for Mini-imagenet. \n\n## Other comments\n\n1. There is a broad range of work involving so-called Sparse Mixture of Experts layers which are conceptually very similar to the stochastic LWTA activated layers. Sparse MoEs have the advantage that they are more computationally efficient. They also have, in my opinion, a more sensible input-dependent gating mechanism compared to simply choosing the most positive activation. These have not been mentioned at all in this work. See (Riquelme et al., 2021) and (Shazeer et al., 2017).\n\n2. I think that the statement \"there is an increasing body of evidence from Neuroscience that neurone with similar functions ...\" requires a citation. Perhaps (Lasner, 2009) is the source for this statement, but that is not clear.\n\n3. \"To ensure statistical significance ...\" – providing multiple runs (while greatly appreciated) does not make the results statistically significant. If the authors wish to claim statistical significance, they should perform the appropriate statistical tests.\n\n4. It is not clear to me that the authors' proposed method actually provides better results for the harder Mini-imagenet task. It is true that the raw gain of 10% Mini-imagenet is larger than the 2% for Omniglot. However, because Omniglot accuracy is nearly saturated by the baselines already, this 2% improvement may actually be even more impressive. \n\n## References\n\nCarlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, Neil Houlsby:\nScaling Vision with Sparse Mixture of Experts. CoRR abs/2106.05974 (2021)\n\nNoam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, Jeff Dean:\nOutrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. ICLR (Poster) 2017\n\nChelsea Finn, Kelvin Xu, Sergey Levine:\nProbabilistic Model-Agnostic Meta-Learning. NeurIPS 2018: 9537-9548\n\nEric T. Nalisnick, Jonathan Gordon, José Miguel Hernández-Lobato:\nPredictive Complexity Priors. AISTATS 2021: 694-702\n\nYutian Chen, Abram L. Friesen, Feryal Behbahani, Arnaud Doucet, David Budden, Matthew Hoffman, Nando de Freitas:\nModular Meta-Learning with Shrinkage. NeurIPS 2020\n\nJonathan Gordon, John Bronskill, Matthias Bauer, Sebastian Nowozin, Richard E. Turner:\nMeta-Learning Probabilistic Inference for Prediction. ICLR (Poster) 2019", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to use neural networks with so-called \"stochastic local winner-takes-all (LWTA)\" activations in the place of standard ReLU activations. These stochastic LWTA activations result in a model with sparse representations. The sparse gating is treated as a random variable. The authors also treat the weights of the neural network as random variables. Both sets of random variables are learnt using variational inference.", "main_review": "## Strengths\n\n* Strong empiracle results \n\n* Great set of ablation studies\n\n## Weaknesses\n\n### Lack of novelty\n\nThere already exist many probabilistic approaches to meta-learning which take uncertainty into account, which the authors have failed to acknowledge. Thus their claims of novelty in this regard are incorrect. For example, the statement \"these innovations constitute a radical departure from the currently prominent design paradigm in ML\" is far too strong a statement given the work of Finn et al. (2018), Nalisnick et al. (2021), Chen et al. (2020), and Gordon et al. (2019).\n\n### Somewhat narrow set of main experiments\n\nThe main set of results in the paper are based only on the Omniglot 20-way and Mini-imagenet 5-way tasks. While the results on these tasks are good, more tasks are required to paint an accurate picture of the method. Cifar100 few shot would be a good start, but even better would be a non-image classification task. I might suggest the ShapeNet View Reconstruction task as used in Gordon et al. (2019). \n\nFurthermore, at least some of the methods from Finn et al. (2018), Nalisnick et al. (2021), Chen et al. (2020), and Gordon et al. (2019) should be compared against. \n\nAn additional ablation study comparing the performance of Stochastic LWTA and standard LWTA would be informative. \n\nFor the effect of sample size ablation, plots with performance on the y-axis and B on the x-axis for B in {1,...,10} would be very interesting as they would hopefully show a \"knee-point\" at which it stops being worthwhile to increase B.\n\n### Lack of clarity\n\nOverall I found the presentation to be confusing in many places and occasionally incorrect. I will list the issues I spotted below. \n\n1. The authors frequently refer to \"stochastic arguments\", \"modelling arguments\", \"deep network arguments\", etc. However, the usage of the word \"argument\" in these contexts does not make sense to me and I am unsure exactly what is meant.\n\n2. The authors frequently use the word \"inference\" to refer to test time prediction. However, this word also refers to the probabilistic inference for random variables. Given that the authors' proposed method makes use of such inference this becomes somewhat confusing. I would recommend only using inference in the probabilistic sense. \n\n3. \\vec{y}_r is introduced in sec 2.2 but it is not mentioned until much later in sec 3.1 that \\vec{y} is the concatenation of \\vec{y}_r. \n\n4. It is not clear why it is a good idea to use a fixed competition function that always chooses the maximum activation. It would be great if the authors could provide some motivation/intuition for this choice. (As I mention below, I think that the Sparse-MoE approach with an input-dependent gating network makes more sense). \n\n5. It is not clear what the authors mean by \"We posit that this novel stochastic paradigm will give rise to more **potent** learned representations ...\" (emphasis my own). (This is another example of a claim that overstates the novelty of this work.)\n\n6. (Minor) I assume that the authors mean \"model-agnostic ML\" rather than \"meta-agnostic ML\" at the start of sec 3.1. \n\n7. (Minor) The authors introduce the symbols A_1 and A_2 to indicate the input and output dimensions of their linear units. Would I and O not be more clear? \n\n8. Paragraphs 2 and 3 are largely redundant as they mainly contain repeated information from sec 2.2.\n\n9. (Minor) In eqn 9, the notation [\\vec{y}]_{r, j} is used, but just above the notation y = [\\vec{y}_{r, j}]. I found this somewhat confusing. \n\n10. In the paragraph below eon 9, the authors first say that they \"postulate\" that the indicator variable is drawn from a Discrete distribution. Later in the paragraph they say that they \"stipulate\" this. Which one is it?\n\n11. Speaking of the \"Discrete\" probability distribution, would it not be more clear to refer to a \"Categorical\" probability distribution?  \n\n12. In the paragraph after eqn 10, the authors state that they impose a posterior density over the weights. However, it is more accurate to say that the prior density is impose and the posterior is inferred. \n\n13. The authors frequently refer to \"stochasticity\" as increasing the generalisation capacity. However, it should be noted that simply adding stochasticity does not help. It must be added in a principled and theoretically justified manner. \n\n14. \"This concludes the formulation of the proposed model-agnostic ML approach.\" – I suspect this line should be at the end of sec 3.3, rather than sec 3.1.\n\n15. The authors refer to independence among layers in the context of the mean-field approximation. I assume that this was meant to be independence among weights? \n\n16. Eqn 13, for the Gumbel-softmax does not show the purpose of the temperature \\tau.\n\n17. In algorithm 2, line 2, the authors suggest that the variational parameters {\\mu, \\sigma} should be sampled. I suspect that the authors mean \\theta or \\vec{w}, should be sampled? \n\n17. In algorithm 2, the authors should show the equation for the BMA. \n\n18. At the top of page 7, the authors refer to \"2 competing units per block\" and then \"2 neurons per block\". Are neurons the same as competing units? \n\n19. \"For the outer-loop, we use SGD with a linear annealed outer step size equal to 0.\" Should this actually be 0? \n\n20. What is the architecture that was used for the baseline methods? \n\n21. The authors claim a 10% improvement over the \"best alternatives\" however, I think it is a 10% improvement over the *worst* alternatives. \n\n22. In table 4, it surprises me that Reptile runs slower than MAML. I don't think this makes sense. Also in table 4, the best value in the \"inference\" column should be bolded. \n\n23. In table 5, does the parameter count for LWTA include both variational parameters \\mu and \\sigma^2? OR is it just \\theta? \n\n24. More motivation is needed for the statement \"It appears that convergence of our approach is even smoother; ...\". \n\n25. Figure 2b, should be replaced with Figure 2a but for Mini-imagenet. \n\n## Other comments\n\n1. There is a broad range of work involving so-called Sparse Mixture of Experts layers which are conceptually very similar to the stochastic LWTA activated layers. Sparse MoEs have the advantage that they are more computationally efficient. They also have, in my opinion, a more sensible input-dependent gating mechanism compared to simply choosing the most positive activation. These have not been mentioned at all in this work. See (Riquelme et al., 2021) and (Shazeer et al., 2017).\n\n2. I think that the statement \"there is an increasing body of evidence from Neuroscience that neurone with similar functions ...\" requires a citation. Perhaps (Lasner, 2009) is the source for this statement, but that is not clear.\n\n3. \"To ensure statistical significance ...\" – providing multiple runs (while greatly appreciated) does not make the results statistically significant. If the authors wish to claim statistical significance, they should perform the appropriate statistical tests.\n\n4. It is not clear to me that the authors' proposed method actually provides better results for the harder Mini-imagenet task. It is true that the raw gain of 10% Mini-imagenet is larger than the 2% for Omniglot. However, because Omniglot accuracy is nearly saturated by the baselines already, this 2% improvement may actually be even more impressive. \n\n## References\n\nCarlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, Neil Houlsby:\nScaling Vision with Sparse Mixture of Experts. CoRR abs/2106.05974 (2021)\n\nNoam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, Jeff Dean:\nOutrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. ICLR (Poster) 2017\n\nChelsea Finn, Kelvin Xu, Sergey Levine:\nProbabilistic Model-Agnostic Meta-Learning. NeurIPS 2018: 9537-9548\n\nEric T. Nalisnick, Jonathan Gordon, José Miguel Hernández-Lobato:\nPredictive Complexity Priors. AISTATS 2021: 694-702\n\nYutian Chen, Abram L. Friesen, Feryal Behbahani, Arnaud Doucet, David Budden, Matthew Hoffman, Nando de Freitas:\nModular Meta-Learning with Shrinkage. NeurIPS 2020\n\nJonathan Gordon, John Bronskill, Matthias Bauer, Sebastian Nowozin, Richard E. Turner:\nMeta-Learning Probabilistic Inference for Prediction. ICLR (Poster) 2019", "summary_of_the_review": "While this paper does have some strengths (strong empirical results, great set of ablation studies), it is let down by many issues of clarity, a lack of novelty, and a somewhat narrow set of main experiments. \n\n++++ Post-revision update ++++\n\nThe authors have addressed many of my concerns, including correcting some of my misunderstandings, improving the clarity of the text in many ways, and adding additional experimental results in the form of ablations, baselines, and a new task. With this in mind, I have increased my score from 3 to 5.\n\nI have not increased my score further as I still find some parts of the presentation slightly confusing (e.g. I don't think the word \"principals\" is much better than the word \"argument\") and more importantly, I still believe that having experiments only with image classification tasks is slightly too narrow.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635959759825}, {"id": "RcuOLck-V8O", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper47/Reviewer_rRh2"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors proposed to replace the classical activation units in MAML with stochastic local winner-takes-all (LWTA) units. The authors argued the biologically motivated LWTA units would lead to better performance in the few-shot setting where the support set in the target task is small, in which case the embedding representation of the target task would be noisy. The authors framed their proposal in the Bayesian setting, and proposed algorithms to estimate the distribution parameters.", "review_text": "The paper is interesting, but is not very clear. Here are a few examples that I was confused with when I first read the paper. I suggest the authors do a thorough revision to make the paper more approachable.\n\n1. Intuitively, why does LWTA units improve the performance with fewer data?\n2. What are the definitions of [\\xi]_{r, j} and [\\xi]_r? Because of this confusion, I couldn't really understand (9) or (10).\n3. What is a \"Discrete\" distribution? is it a \"multinoulli distribution\"?\n4. Why should the density of [\\xi]_r (again, I don't quite understand this notation) be proportional to the size of the activation?\n5. It is unclear to me why we need to sample W from a distribution, rather than obtaining its point estimate using ERM?\n6. The description of the training procedure on page 6 (especially the paragraph above (13) is very dense. I think I got a big picture of the proposal, but I don't think I can reproduce their methodology based on this description. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors proposed to replace the classical activation units in MAML with stochastic local winner-takes-all (LWTA) units. The authors argued the biologically motivated LWTA units would lead to better performance in the few-shot setting where the support set in the target task is small, in which case the embedding representation of the target task would be noisy. The authors framed their proposal in the Bayesian setting, and proposed algorithms to estimate the distribution parameters.", "main_review": "The paper is interesting, but is not very clear. Here are a few examples that I was confused with when I first read the paper. I suggest the authors do a thorough revision to make the paper more approachable.\n\n1. Intuitively, why does LWTA units improve the performance with fewer data?\n2. What are the definitions of [\\xi]_{r, j} and [\\xi]_r? Because of this confusion, I couldn't really understand (9) or (10).\n3. What is a \"Discrete\" distribution? is it a \"multinoulli distribution\"?\n4. Why should the density of [\\xi]_r (again, I don't quite understand this notation) be proportional to the size of the activation?\n5. It is unclear to me why we need to sample W from a distribution, rather than obtaining its point estimate using ERM?\n6. The description of the training procedure on page 6 (especially the paragraph above (13) is very dense. I think I got a big picture of the proposal, but I don't think I can reproduce their methodology based on this description. ", "summary_of_the_review": "I think the paper is interesting, but the presentation needs to be greatly improved. Many notations are used without introduction, and the description of the methodology is very dense that I am afraid that unless the reader is very familiar with the field, they could not reproduce the methods based on the description", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635873253544}, {"id": "Y6NBFDfOjqe", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper47/Reviewer_Nj3T"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed a novel meta-learning method by replacing standard nonlinearity functions in MAML with stochastic local winner-takes-all (LWTA) activations. The authors claim that such design results in sparse representations and benefit meta-learning. This method demonstrated superior performance over MAML and FOMAML over standard benchmarks such as Omniglot and Mini-ImageNet.", "review_text": "1. The novelty of the paper is limited. It applies stochastic LWTA to the standard MAML framework. Using Bayesian Model Averaging for its inferencing is also not new. Also, I do not find any advantage of this stochastic design which in fact induces large computation overhead.\n\n2. This paper lacks insight into how the proposed design is specifically related to meta-learning. Why does the \"sparse representation\" induced by the LWTA function help? \n \n3. The experiment results are strong. The authors provided a detailed experimental setting for reproducing the result. The improvement of the proposed model over MAML and FOMAML is significant. However, I would still suggest the authors compare their method to other recent state-of-the-art meta-learning methods.\n\n4. The authors provided an ablation study to empirically prove the stochastic LWTA is the underlying force for improving the model. However, I think the ablation study needs to be broken down into a deterministic LWTA function and a stochastic nature. It's not clear actually the model benefits from which part. The authors also discussed the computation overhead. These experiments are limited to small model settings without discussion of the role of batch size in the play.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a novel meta-learning method by replacing standard nonlinearity functions in MAML with stochastic local winner-takes-all (LWTA) activations. The authors claim that such design results in sparse representations and benefit meta-learning. This method demonstrated superior performance over MAML and FOMAML over standard benchmarks such as Omniglot and Mini-ImageNet.", "main_review": "1. The novelty of the paper is limited. It applies stochastic LWTA to the standard MAML framework. Using Bayesian Model Averaging for its inferencing is also not new. Also, I do not find any advantage of this stochastic design which in fact induces large computation overhead.\n\n2. This paper lacks insight into how the proposed design is specifically related to meta-learning. Why does the \"sparse representation\" induced by the LWTA function help? \n \n3. The experiment results are strong. The authors provided a detailed experimental setting for reproducing the result. The improvement of the proposed model over MAML and FOMAML is significant. However, I would still suggest the authors compare their method to other recent state-of-the-art meta-learning methods.\n\n4. The authors provided an ablation study to empirically prove the stochastic LWTA is the underlying force for improving the model. However, I think the ablation study needs to be broken down into a deterministic LWTA function and a stochastic nature. It's not clear actually the model benefits from which part. The authors also discussed the computation overhead. These experiments are limited to small model settings without discussion of the role of batch size in the play.\n", "summary_of_the_review": "This paper improves standard MAML by replacing the nonlinearity with stochastic LWTA. The idea has limited novelty and limited insight.\nThe pros are the experiments are well performed, and the results are significant. I appreciate the ablation study and the overhead complexity discussion.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635538806425}], "openreview_url": "https://openreview.net/forum?id=FFGDKzLasUa", "arxiv_id": "2208.01573", "paper_pdf": "papers/FFGDKzLasUa.pdf", "paper_pdf_sha256": "fc9fc13c23ca5b0cc33765ec221519988622468fd2d3069437fffd1c04c565ac", "paper_pdf_bytes": 1730465, "paper_pdf_source": "openreview", "code_url": "https://github.com/Kkalais/StochLWTA-ML", "code_repository": "Kkalais/StochLWTA-ML", "code_commit": "42d01bf70bf08612cd367e170cebe67fc9d1cfd1", "code_archive": "repos/FFGDKzLasUa.zip", "code_archive_sha256": "59a34a9bbe0eb248f2ff99e9cd2ee1e7ceba225c84a76977c755cdbe8452a03f", "code_archive_bytes": 2143291, "code_file_count": 26, "code_extensions": {".py": 23, ".ipynb": 3}, "github_disk_usage_kb": 2103, "github_languages": {"Jupyter Notebook": 574286, "Python": 186651}, "github_archived": false, "github_pushed_at": "2025-07-21T14:34:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/stochastic-deep-networks-with-linear-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1NRMmEUyXMu", "year": 2021, "status": "rejected", "title": "World Model as a Graph: Learning Latent Landmarks for Planning", "authors": ["Lunjun Zhang", "Ge Yang", "Bradly C Stadie"], "authorids": ["~Lunjun_Zhang1", "~Ge_Yang1", "~Bradly_C_Stadie1"], "authors_source": "OpenReview API", "abstract": "Planning, the ability to analyze the structure of a problem in the large and decompose it into interrelated subproblems, is a hallmark of human intelligence. While deep reinforcement learning (RL) has shown great promise for solving relatively straightforward control tasks, it remains an open problem how to best incorporate planning into existing deep RL paradigms to handle increasingly complex environments. One prominent framework, Model-Based RL, learns a world model and plans using step-by-step virtual rollouts. This type of world model quickly diverges from reality when the planning horizon increases, thus struggling at long-horizon planning. How can we learn world models that endow agents with the ability to do temporally extended reasoning? In this work, we propose to learn graph-structured world models composed of sparse, multi-step transitions. We devise a novel algorithm to learn latent landmarks that are scattered (in terms of reachability) across the goal space as the nodes on the graph. In this same graph, the edges are the reachability estimates distilled from Q-functions. On a variety of high-dimensional continuous control tasks ranging from robotic manipulation to navigation, we demonstrate that our method, named L^{3}P, significantly outperforms prior work, and is oftentimes the only method capable of leveraging both the robustness of model-free RL and generalization of graph-search algorithms. We believe our work is an important step towards scalable planning in reinforcement learning. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "wmP0mIFNJNY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2191/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper approaches long horizon planning by learning a sparse graphical representation. The proposed algorithm, L3P, proceeds by learning a latent space which enforces a distance measure, where this distance is learned to mimic the number of steps between states via a goal conditioned Q-function. A clustering algorithm is then used to represent this latent space through only a small, efficient set of latent landmarks. These landmarks are then connected if nearby via the distance, the estimate of which is refined via soft value iterations over the graph. L3P is demonstrated on a number of environments to be both data efficient and high performing compared to baselines.\n\nThe paper is clear and well written. The topic of graphical representations of state spaces to plan long horizons over via local predictions holds significant promise, as does sparsifying this representation. The latent space construction and sparsification procedures are reasonable, particularly tying the distance to the policy Q-function as in SoRB. A few notes on clarity:\n- The algorithm should be moved to the main body of the paper. (to reduce space, potentially remove a row of images in Fig. 4, stretch figure 1 to use the full width)\n- A video would be helpful of the full process, including the soft value iterations.\n- What algorithmic parameters (e.g., latent space dimensionality, network sizes, training data and number of states from which the latent space was constructed) were used?\n- How are L_rec and L_latent traded off?\n- The tasks, e.g., Fig 2 and 5, should be shown before the results in Fig. 3.\n- More intuition on the soft value iteration would be helpful. \n\nThe results on the shown tasks show clear benefits of L3P, particularly on the tasks where HER has most difficulty (likely the more long range tasks) L3P outperforms substantially. However, the results are somewhat lacking (1) on ablation, (2) demonstrating new environment generalization, and (3) showing significantly complex tasks.\n1) The authors show ablation for number of landmarks and d_max, but not for algorithmic changes like the soft value iteration. How much does this procedure help versus the use of latent marks to sparsify the space? It would also be interesting (though not fully necessary) to see comparison to Savinov 2018a (SPTM), which uses a learned distance predictor rather than the Q-function as in SoRB and L3P.\n2) It is not clear how this method would generalize to new environments, e.g., Section 5.5 in SoRB. This to me is key (particularly with generalization in the title, which I believe currently only refers to task length) and may be a challenge for L3P due to finding landmarks from limited coverage of a new scene.\n3) L3P should be shown on higher dimensional and longer horizon problems, as the maze is fairly short and simple , as is the box task (as can be seen by HER’s performance). These do not push the boundaries of the algorithm. Higher dimensional problems may particularly challenge the landmark learning (e.g., if the landmarks must be learned too in a higher dimensional latent space)\n\n\n__________\n\nAfter author response:\nI appreciate the author’s response. Previous topics:\n1) The author’s add ablations on the hard vs. soft min during the graph search, the additional results are informative, but not conclusive. Given that the overall performance is similar, the authors need to demonstrate the soft-min’s benefits for each experiment and over more training seeds. \n2) For generalization, the author’s confirm that this method is unable to generalize to new environments, though clearly it has other benefits in terms of data efficiency and robustness of solutions. I believe these are still important benefits, though it would be useful to discuss how generalization may be achieved.\n3) For harder experiments, the authors note that the baselines perform poorly there and thus these tasks were not considered. Though reasonable that the baselines are unable to perform in such cases, harder experiments would show the limit of the proposed algorithm. It would be useful to see for instance how well it scales with dimensionality, how quickly the success rate falls off.\n\nNew comments:\na) I believe the title change away from Generalization is an improvement, though the algorithm name \"WORLD MODEL AS A GRAPH\" seems to not capture the novel aspects of this work. This name I believe would be more readily applied to search on the replay buffer or semi-parametric topological memory.\nb) R2's point that much of the robustness may be a factor of choosing states further from the wall is an interesting one. It would be interesting to examine exactly *why* the method is robust.\nOverall, I believe the paper is interesting and proposes some novel ideas that have benefit, it requires more thorough analysis, and thus I am leaving my score unchanged.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Useful and efficient approach, but more ablations, generalizations, and environments should be examined.", "review": "This paper approaches long horizon planning by learning a sparse graphical representation. The proposed algorithm, L3P, proceeds by learning a latent space which enforces a distance measure, where this distance is learned to mimic the number of steps between states via a goal conditioned Q-function. A clustering algorithm is then used to represent this latent space through only a small, efficient set of latent landmarks. These landmarks are then connected if nearby via the distance, the estimate of which is refined via soft value iterations over the graph. L3P is demonstrated on a number of environments to be both data efficient and high performing compared to baselines.\n\nThe paper is clear and well written. The topic of graphical representations of state spaces to plan long horizons over via local predictions holds significant promise, as does sparsifying this representation. The latent space construction and sparsification procedures are reasonable, particularly tying the distance to the policy Q-function as in SoRB. A few notes on clarity:\n- The algorithm should be moved to the main body of the paper. (to reduce space, potentially remove a row of images in Fig. 4, stretch figure 1 to use the full width)\n- A video would be helpful of the full process, including the soft value iterations.\n- What algorithmic parameters (e.g., latent space dimensionality, network sizes, training data and number of states from which the latent space was constructed) were used?\n- How are L_rec and L_latent traded off?\n- The tasks, e.g., Fig 2 and 5, should be shown before the results in Fig. 3.\n- More intuition on the soft value iteration would be helpful. \n\nThe results on the shown tasks show clear benefits of L3P, particularly on the tasks where HER has most difficulty (likely the more long range tasks) L3P outperforms substantially. However, the results are somewhat lacking (1) on ablation, (2) demonstrating new environment generalization, and (3) showing significantly complex tasks.\n1) The authors show ablation for number of landmarks and d_max, but not for algorithmic changes like the soft value iteration. How much does this procedure help versus the use of latent marks to sparsify the space? It would also be interesting (though not fully necessary) to see comparison to Savinov 2018a (SPTM), which uses a learned distance predictor rather than the Q-function as in SoRB and L3P.\n2) It is not clear how this method would generalize to new environments, e.g., Section 5.5 in SoRB. This to me is key (particularly with generalization in the title, which I believe currently only refers to task length) and may be a challenge for L3P due to finding landmarks from limited coverage of a new scene.\n3) L3P should be shown on higher dimensional and longer horizon problems, as the maze is fairly short and simple , as is the box task (as can be seen by HER’s performance). These do not push the boundaries of the algorithm. Higher dimensional problems may particularly challenge the landmark learning (e.g., if the landmarks must be learned too in a higher dimensional latent space)\n\n\n__________\n\nAfter author response:\nI appreciate the author’s response. Previous topics:\n1) The author’s add ablations on the hard vs. soft min during the graph search, the additional results are informative, but not conclusive. Given that the overall performance is similar, the authors need to demonstrate the soft-min’s benefits for each experiment and over more training seeds. \n2) For generalization, the author’s confirm that this method is unable to generalize to new environments, though clearly it has other benefits in terms of data efficiency and robustness of solutions. I believe these are still important benefits, though it would be useful to discuss how generalization may be achieved.\n3) For harder experiments, the authors note that the baselines perform poorly there and thus these tasks were not considered. Though reasonable that the baselines are unable to perform in such cases, harder experiments would show the limit of the proposed algorithm. It would be useful to see for instance how well it scales with dimensionality, how quickly the success rate falls off.\n\nNew comments:\na) I believe the title change away from Generalization is an improvement, though the algorithm name \"WORLD MODEL AS A GRAPH\" seems to not capture the novel aspects of this work. This name I believe would be more readily applied to search on the replay buffer or semi-parametric topological memory.\nb) R2's point that much of the robustness may be a factor of choosing states further from the wall is an interesting one. It would be interesting to examine exactly *why* the method is robust.\nOverall, I believe the paper is interesting and proposes some novel ideas that have benefit, it requires more thorough analysis, and thus I am leaving my score unchanged.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604503726082}, {"id": "DJsCqwuDwTl", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2191/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Summary:\nThis paper presents a method for learning a sparse set of latent subgoal states during training. Using a goal-conditioned policy and the latent states, a simple planning algorithm that performs soft value iteration between the subgoal states is proposed to facilitate within-dataset generalization. The proposed method outperforms competing approaches on multiple simulated navigation and robotics tasks.\n\n##### Pros:\n- The results seem strong.\n- The method is straightforward and nice, and it seems to work more efficiently than previous methods.\n\n##### Cons:\n- Overall, the paper is incomplete. It’s very difficult to understand the whole system, how training proceeds, what the hyperparameters are, etc. There is very little detail about how this all works. The specific components of the latent landmark learning and the planning are mostly fleshed out (albeit with some missing details still) but the rest of the setup is mostly not described. How are the losses combined? Are there multiple learners or does the DDPG agent also do the latent space learning? How is training done? What are the hyperparameters used? \n- The generalization claims are a bit oversold as the generalization achieved is weak. In particular, the method requires that the test goal distribution have sufficient overlap with the train goal distribution for latent landmarks to be created near the desired test goals. \n- Are the experiment domains created here or existing work? If the former, more details are required. If the latter, citations are required.\n- Given the multiple changes with respect to prior work, additional ablations and experiments are needed to better understand why this method is performing well, and what contributions are important.\n\n#### Decision:\nOverall, I find it too hard to know exactly what is happening in this work. It would be very difficult, if not impossible, to reproduce this work from this paper alone. Without sufficient details about the system, it’s hard to evaluate, which means that I default to a reject, especially in conjunction with the other weaknesses listed above.\n\n#### Questions:\n1. What is the goal space G? Is it the state space? This should be defined.\n2. What exactly is \\Psi? Is it learned or is it given? This should be explained better.\n3. How is the autoencoder for the latent space learned within the overall system?\n4. The connection between D and Q is poorly explained. From the text, it reads as if you get Q from D. However, it makes more sense if the DDPG agent estimates Q and then you use eq(3) to compute D from Q. Which is it?\n5. Are both V and D really necessary? It seems like they encode the same thing, essentially. \n6. Why do you use soft value iteration? Is it a good choice? Is there a reason for this choice? Would other choices be better or worse? It seems likely that the need for the dist_max penalty is due to the use of the soft value iteration with a poorly set soft value iteration temperature. Is this true?\n7. Are the domains from Duan et al 2016 or some other existing work? Or did you create them from scratch (unlikely). Why are they labeled “-hard”? What is hard about them? These need to be cited and described. You have tons of space in the appendix to use freely. \n8. In Figure 3, is the x axis training timesteps? \n9. Figure 5 caption says the agent has learned landmarks that help avoid collision with the box. Where are these? Is there a figure? Any sort of information on these? Wouldn’t any landmark away from the box accomplish this?\n10. For Figure 6, why is choosing the maximum distance bound important? Why would the planning algorithm choose subgoals that are near the max distance bound given that the planner is finding the argmin subgoal? It seems likely that this hacky distance bound is unnecessary.\n11. The “planning algorithm” seems trivial. Am I missing something? Given the latent landmarks and the estimates for $d_{c \\rightarrow g}$, the planner just gets the next estimated subgoal and the distance to it, executes the goal-conditioned policy for that subgoal for that many steps, crosses out that subgoal, and then goes to the next subgoal. Is it non-trivial because $d_{s \\rightarrow c}$ gets updated after each step (not included in the pseudocode)?\n12. Is the GLS algorithm a contribution of this work or something from a previous work? Again, it's neither described sufficiently nor cited sufficiently.\n\n#### Comments:\n- The introduction uses unnecessary hyperbole and insufficient citations to make its points. It comes off as less compelling as a result.\n- Section 3 is called Background but really it contains the preliminaries and definitions of this work. It should not be called Background.\n- Should the variance vector of the centroids appear in equation (5)? It is mentioned once and then never discussed again. What values of it are learned? Is it important? \n- Similarly, the temperature for the soft value iteration is mentioned once and then never discussed again. What role does it play? Is it important? Would tuning this better remove the need for the dist_max penalty?\n- The notation $f_D(c_i)$ is unnecessarily complicated. These values correspond to states, no? So define them as $s_{c_i}$ or something similar to make the math and text more clear. As it is, equation (6) is unnecessarily cluttered.\n- I assume that the argmax in the subgoal computation at the end of section 4 should be an argmax over c? \n- There is a typo in the Figure 3 caption (PointmMaze). Further, the text in the subfigures is illegibly small.\n- Break lines 11 and 13 of algorithm 2 into two lines each.\n\n\n*****************\nAfter author response:\n\nI appreciate the updates you've made to the paper to better flesh it out, including the diagrams, pseudocode, and additional ablations. I've increased my score accordingly. Regarding generalization, I fully agree that it can be difficult to show. That said, when it's advertised in the title of the paper, I expect it to be clearly shown in the paper itself. It seems that the language has been greatly toned down in the updated version. However, without entirely re-reviewing the paper, I am unable to fully recommend acceptance.\n\nAside: I would have appreciated responses to my questions directly so that I don't need to dig through the rewritten paper to find the answers to my questions.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official review", "review": "#### Summary:\nThis paper presents a method for learning a sparse set of latent subgoal states during training. Using a goal-conditioned policy and the latent states, a simple planning algorithm that performs soft value iteration between the subgoal states is proposed to facilitate within-dataset generalization. The proposed method outperforms competing approaches on multiple simulated navigation and robotics tasks.\n\n##### Pros:\n- The results seem strong.\n- The method is straightforward and nice, and it seems to work more efficiently than previous methods.\n\n##### Cons:\n- Overall, the paper is incomplete. It’s very difficult to understand the whole system, how training proceeds, what the hyperparameters are, etc. There is very little detail about how this all works. The specific components of the latent landmark learning and the planning are mostly fleshed out (albeit with some missing details still) but the rest of the setup is mostly not described. How are the losses combined? Are there multiple learners or does the DDPG agent also do the latent space learning? How is training done? What are the hyperparameters used? \n- The generalization claims are a bit oversold as the generalization achieved is weak. In particular, the method requires that the test goal distribution have sufficient overlap with the train goal distribution for latent landmarks to be created near the desired test goals. \n- Are the experiment domains created here or existing work? If the former, more details are required. If the latter, citations are required.\n- Given the multiple changes with respect to prior work, additional ablations and experiments are needed to better understand why this method is performing well, and what contributions are important.\n\n#### Decision:\nOverall, I find it too hard to know exactly what is happening in this work. It would be very difficult, if not impossible, to reproduce this work from this paper alone. Without sufficient details about the system, it’s hard to evaluate, which means that I default to a reject, especially in conjunction with the other weaknesses listed above.\n\n#### Questions:\n1. What is the goal space G? Is it the state space? This should be defined.\n2. What exactly is \\Psi? Is it learned or is it given? This should be explained better.\n3. How is the autoencoder for the latent space learned within the overall system?\n4. The connection between D and Q is poorly explained. From the text, it reads as if you get Q from D. However, it makes more sense if the DDPG agent estimates Q and then you use eq(3) to compute D from Q. Which is it?\n5. Are both V and D really necessary? It seems like they encode the same thing, essentially. \n6. Why do you use soft value iteration? Is it a good choice? Is there a reason for this choice? Would other choices be better or worse? It seems likely that the need for the dist_max penalty is due to the use of the soft value iteration with a poorly set soft value iteration temperature. Is this true?\n7. Are the domains from Duan et al 2016 or some other existing work? Or did you create them from scratch (unlikely). Why are they labeled “-hard”? What is hard about them? These need to be cited and described. You have tons of space in the appendix to use freely. \n8. In Figure 3, is the x axis training timesteps? \n9. Figure 5 caption says the agent has learned landmarks that help avoid collision with the box. Where are these? Is there a figure? Any sort of information on these? Wouldn’t any landmark away from the box accomplish this?\n10. For Figure 6, why is choosing the maximum distance bound important? Why would the planning algorithm choose subgoals that are near the max distance bound given that the planner is finding the argmin subgoal? It seems likely that this hacky distance bound is unnecessary.\n11. The “planning algorithm” seems trivial. Am I missing something? Given the latent landmarks and the estimates for $d_{c \\rightarrow g}$, the planner just gets the next estimated subgoal and the distance to it, executes the goal-conditioned policy for that subgoal for that many steps, crosses out that subgoal, and then goes to the next subgoal. Is it non-trivial because $d_{s \\rightarrow c}$ gets updated after each step (not included in the pseudocode)?\n12. Is the GLS algorithm a contribution of this work or something from a previous work? Again, it's neither described sufficiently nor cited sufficiently.\n\n#### Comments:\n- The introduction uses unnecessary hyperbole and insufficient citations to make its points. It comes off as less compelling as a result.\n- Section 3 is called Background but really it contains the preliminaries and definitions of this work. It should not be called Background.\n- Should the variance vector of the centroids appear in equation (5)? It is mentioned once and then never discussed again. What values of it are learned? Is it important? \n- Similarly, the temperature for the soft value iteration is mentioned once and then never discussed again. What role does it play? Is it important? Would tuning this better remove the need for the dist_max penalty?\n- The notation $f_D(c_i)$ is unnecessarily complicated. These values correspond to states, no? So define them as $s_{c_i}$ or something similar to make the math and text more clear. As it is, equation (6) is unnecessarily cluttered.\n- I assume that the argmax in the subgoal computation at the end of section 4 should be an argmax over c? \n- There is a typo in the Figure 3 caption (PointmMaze). Further, the text in the subfigures is illegibly small.\n- Break lines 11 and 13 of algorithm 2 into two lines each.\n\n\n*****************\nAfter author response:\n\nI appreciate the updates you've made to the paper to better flesh it out, including the diagrams, pseudocode, and additional ablations. I've increased my score accordingly. Regarding generalization, I fully agree that it can be difficult to show. That said, when it's advertised in the title of the paper, I expect it to be clearly shown in the paper itself. It seems that the language has been greatly toned down in the updated version. However, without entirely re-reviewing the paper, I am unable to fully recommend acceptance.\n\nAside: I would have appreciated responses to my questions directly so that I don't need to dig through the rewritten paper to find the answers to my questions.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603923764792}, {"id": "fwPi_pZyptm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2191/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes an approach for automatically learning state abstraction on a RL problem, which can then be used for temporally extended planning using a search algorithm. The main contribution is introducing the concept of latent *landmarks*, a clustering of low dimensional state embeddings. Landmarks are defined on a latent space wherein the distance between two latents code is small if their corresponding high-dimensional states can reach each other in few environment steps. The paper proposes to cluster latent states, so that each cluster contains states that are easy to reach from each other; the landmarks correspond to the centers of these clusters. An algorithm for automatically learning this clustering is introduced in the paper.  With landmarks in hand, the paper proposes a soft-value iteration method to compute shortest path distances to the problem's goal. The approach is evaluated in a variety of domains, and compares favorably with recent state-of-the-art methods.\n\nIn general, I found the paper interesting and the key ideas intuitive and sensible. The paper is reasonably well-written, but there are some important points that are unclear (see below). The experimental section compares with very recent algorithms for planning over learned value functions (SORB and MSS), and the results on five continuous control task show substantial improvement over these methods.\n\nIn terms of clarification, I have a few questions for the authors:\n- With regards to the embedding used and $\\Psi$:\n  - The embedding operates over goals. Are these the goals of the problem? For example, in the AntMaze problem (Fig. 4), is $G = \\{ \\textit{red-square} \\} $? If so, doesn't this mean that  $\\forall s \\Psi(s) = \\textit{red-square}$, and thus $V(\\Psi(s), \\Psi(s')) = 0$ for all pairs of states?\n  - More reasonably, $V$ could be defined over states, so that $V(s_1, s_2)$ indicates the number of steps between $s_1$ and $s_2$. This can be achieved under the current formulation with $G = S$ and $\\Psi(s) = s$. Then $D(s_t, a, \\Psi(s_t))$ represents the number of steps between $s_{t+1}$ and $s_t$. Is this the case for most experiments? No details of the $\\Psi$ function used in the experiments was given, so I find this point somewhat confusing. \n  - If this is not the case, then I think it's important to explain where does this mapping function comes from, since it essentially provides some reward shaping for the problem. \n\n- Can you explain what is the motivation for using Algorithm 2? A common approach is to use Djikstra on the graph representation. Is there a reason why this can't be done (or is worse) in this case? \n\n- I don't fully understand Algorithm 1 to create the batch for Eq. (5). Can you explain this algorithm in more details? For example, the definition of $\\texttt{dist}$ is ambiguous; is it pairwise distance between all sampled goals? distance between $g_1$ and all the others? something else? In general, I'm having quite a hard time figuring out what this is doing.\n\nIn light of above, I think the clarity of the paper can be improved in many places. But, overall, I'm positive towards this work and I think it is a nice contribution, particularly since I'm not aware of other work creating explicit low-dimensional landmarks to be used for search. That being said, one exception that is missing from the literature review is [1], which creates a discrete latent representation of the environment and solves it using prioritized sweeping. \n\n[1] Corneil, Dane, Wulfram Gerstner, and Johanni Brea. \"Efficient Model-Based Deep Reinforcement Learning with Variational State Tabulation.\" International Conference on Machine Learning. 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Introduces new method for state abstraction on a latent space with compelling experiments. Writing unclear in some places.", "review": "This paper proposes an approach for automatically learning state abstraction on a RL problem, which can then be used for temporally extended planning using a search algorithm. The main contribution is introducing the concept of latent *landmarks*, a clustering of low dimensional state embeddings. Landmarks are defined on a latent space wherein the distance between two latents code is small if their corresponding high-dimensional states can reach each other in few environment steps. The paper proposes to cluster latent states, so that each cluster contains states that are easy to reach from each other; the landmarks correspond to the centers of these clusters. An algorithm for automatically learning this clustering is introduced in the paper.  With landmarks in hand, the paper proposes a soft-value iteration method to compute shortest path distances to the problem's goal. The approach is evaluated in a variety of domains, and compares favorably with recent state-of-the-art methods.\n\nIn general, I found the paper interesting and the key ideas intuitive and sensible. The paper is reasonably well-written, but there are some important points that are unclear (see below). The experimental section compares with very recent algorithms for planning over learned value functions (SORB and MSS), and the results on five continuous control task show substantial improvement over these methods.\n\nIn terms of clarification, I have a few questions for the authors:\n- With regards to the embedding used and $\\Psi$:\n  - The embedding operates over goals. Are these the goals of the problem? For example, in the AntMaze problem (Fig. 4), is $G = \\{ \\textit{red-square} \\} $? If so, doesn't this mean that  $\\forall s \\Psi(s) = \\textit{red-square}$, and thus $V(\\Psi(s), \\Psi(s')) = 0$ for all pairs of states?\n  - More reasonably, $V$ could be defined over states, so that $V(s_1, s_2)$ indicates the number of steps between $s_1$ and $s_2$. This can be achieved under the current formulation with $G = S$ and $\\Psi(s) = s$. Then $D(s_t, a, \\Psi(s_t))$ represents the number of steps between $s_{t+1}$ and $s_t$. Is this the case for most experiments? No details of the $\\Psi$ function used in the experiments was given, so I find this point somewhat confusing. \n  - If this is not the case, then I think it's important to explain where does this mapping function comes from, since it essentially provides some reward shaping for the problem. \n\n- Can you explain what is the motivation for using Algorithm 2? A common approach is to use Djikstra on the graph representation. Is there a reason why this can't be done (or is worse) in this case? \n\n- I don't fully understand Algorithm 1 to create the batch for Eq. (5). Can you explain this algorithm in more details? For example, the definition of $\\texttt{dist}$ is ambiguous; is it pairwise distance between all sampled goals? distance between $g_1$ and all the others? something else? In general, I'm having quite a hard time figuring out what this is doing.\n\nIn light of above, I think the clarity of the paper can be improved in many places. But, overall, I'm positive towards this work and I think it is a nice contribution, particularly since I'm not aware of other work creating explicit low-dimensional landmarks to be used for search. That being said, one exception that is missing from the literature review is [1], which creates a discrete latent representation of the environment and solves it using prioritized sweeping. \n\n[1] Corneil, Dane, Wulfram Gerstner, and Johanni Brea. \"Efficient Model-Based Deep Reinforcement Learning with Variational State Tabulation.\" International Conference on Machine Learning. 2018.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603898131856}, {"id": "_WeL21aim2-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2191/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper generalizes the previous graph-based planning algorithm for goal-conditioned RL algorithms by learning a latent metric space and building the graph by clustering. The experiments show that their approach outperforms both the HER and previous MSS/SORB baselines.\n\nStrong points:\n1. The paper achieves good performance on the challenging AntMaze and PickAndPlace environments.\n2. The clustering methods to find latent landmarks are novel in this setting.\n\nWeak points:\n1. The novelty of the paper is limited. Planning on the graph or learning a latent distance embedding for planning are all existing ideas. The combination is straightforward. The robustness of soft-min is well known. \n2. I am confused about why the authors' approach is better than MSS or SORB. It's not obvious that a distance-based clustering will perform significantly better than sampled landmarks (like farthest point sampling). I guess the reason is that clustering helps to avoid the landmarks near the wall (which often appears in the FP sampled methods). Those states are challenging as it has a higher possibility for the ant to collide with the wall. Can the author give some words about this? Moreover, the paper is not clear about the contributions of each component in the improvements. For example, no table or figure shows the benefits of the soft value iteration over the hard version. The author shall provide more ablation studies and explanations.\n3. Although the paper presents the overall idea well, there is still a lot of room for improvement in writing. For example, I suggest that the author replaces the hyphen with the comma in the abstract.\n\n~~~Based on the current presentation, the lack of ablation study, and the limited novelty, I think this paper is not good enough to be accepted. ~~~\n\nHere are some questions:\n1. Can we build graphs by sampling goal states and cluster them based on the metric space defined by the Q networks? What's the performance of this approach? Will it be worse? Why?\n2. Can a sample-based approach work in non-navigation environments? Motion planning algorithms can solve the high-level part of both pick&place or ant maze problem once the geometric model is known. If so, why do we need the expensive RL algorithm?\n\n\n---\n\nAfter author response：\n\nThe authors have improved the presentation and added the necessary ablation studies. I appreciate the authors' effort. I am glad to raise the score to reflect the changes. I hope this work will inspire future research on hierarchical planning algorithms.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A better graph-based planning algorithm for goal-conditioned RL, but the presentation is not very clear. ", "review": "This paper generalizes the previous graph-based planning algorithm for goal-conditioned RL algorithms by learning a latent metric space and building the graph by clustering. The experiments show that their approach outperforms both the HER and previous MSS/SORB baselines.\n\nStrong points:\n1. The paper achieves good performance on the challenging AntMaze and PickAndPlace environments.\n2. The clustering methods to find latent landmarks are novel in this setting.\n\nWeak points:\n1. The novelty of the paper is limited. Planning on the graph or learning a latent distance embedding for planning are all existing ideas. The combination is straightforward. The robustness of soft-min is well known. \n2. I am confused about why the authors' approach is better than MSS or SORB. It's not obvious that a distance-based clustering will perform significantly better than sampled landmarks (like farthest point sampling). I guess the reason is that clustering helps to avoid the landmarks near the wall (which often appears in the FP sampled methods). Those states are challenging as it has a higher possibility for the ant to collide with the wall. Can the author give some words about this? Moreover, the paper is not clear about the contributions of each component in the improvements. For example, no table or figure shows the benefits of the soft value iteration over the hard version. The author shall provide more ablation studies and explanations.\n3. Although the paper presents the overall idea well, there is still a lot of room for improvement in writing. For example, I suggest that the author replaces the hyphen with the comma in the abstract.\n\n~~~Based on the current presentation, the lack of ablation study, and the limited novelty, I think this paper is not good enough to be accepted. ~~~\n\nHere are some questions:\n1. Can we build graphs by sampling goal states and cluster them based on the metric space defined by the Q networks? What's the performance of this approach? Will it be worse? Why?\n2. Can a sample-based approach work in non-navigation environments? Motion planning algorithms can solve the high-level part of both pick&place or ant maze problem once the geometric model is known. If so, why do we need the expensive RL algorithm?\n\n\n---\n\nAfter author response：\n\nThe authors have improved the presentation and added the necessary ablation studies. I appreciate the authors' effort. I am glad to raise the score to reflect the changes. I hope this work will inspire future research on hierarchical planning algorithms.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603869806518}], "openreview_url": "https://openreview.net/forum?id=1NRMmEUyXMu", "arxiv_id": "2011.12491", "paper_pdf": "papers/1NRMmEUyXMu.pdf", "paper_pdf_sha256": "75a9eb65cc96c9c4bfa503ee1c414e68607d1a22c10add46d2c5201abc44ed17", "paper_pdf_bytes": 6599267, "paper_pdf_source": "openreview", "code_url": "https://github.com/LunjunZhang/world-model-as-a-graph", "code_repository": "LunjunZhang/world-model-as-a-graph", "code_commit": "b7d99b2efe15b9be162efe9a83cd99e9fdc88c33", "code_archive": "repos/1NRMmEUyXMu.zip", "code_archive_sha256": "f5de94aba1d2711cf3c4ffdafaa3f0dfeace888dc18d9255533c70a676448fd6", "code_archive_bytes": 1416445, "code_file_count": 69, "code_extensions": {".py": 64, ".sh": 5}, "github_disk_usage_kb": 1350, "github_languages": {"Python": 333076, "Shell": 4882}, "github_archived": false, "github_pushed_at": "2021-07-17T19:16:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/world-model-as-a-graph-learning-latent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MmYiosUoay", "year": 2026, "status": "rejected", "title": "Seeing is Understanding: Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs", "authors": ["Wei-Yao Wang", "Zhao Wang", "Helen Suzuki", "Yoshiyuki Kobayashi"], "authorids": ["~Wei-Yao_Wang1", "~Zhao_Wang6", "~Helen_Suzuki1", "~Yoshiyuki_Kobayashi2"], "authors_source": "OpenReview API", "abstract": "Recent Multimodal Large Language Models (MLLMs) have demonstrated significant progress in perceiving and reasoning over multimodal inquiries, ushering in a new research era for foundation models.\nHowever, vision-language misalignment in MLLMs has emerged as a critical challenge, where the textual responses generated by these models are not factually aligned with the given text-image inputs.\nExisting efforts to address vision-language misalignment have focused on developing specialized vision-language connectors or leveraging visual instruction tuning from diverse domains.\nIn this paper, we tackle this issue from a fundamental yet unexplored perspective by revisiting the core architecture of MLLMs.\nMost MLLMs are typically built on decoder-only LLMs consisting of a causal attention mechanism, which **limits the ability of the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text)**.\nTo address this problem a MLLM that unlocks causal attention into our proposed modality-mutual attention (MMA) to enable image tokens to attend to text tokens.\nThis simple yet effective design allows MMA to achieve state-of-the-art performance in 12 multimodal understanding benchmarks (**+5.5\\% on average across 2 LLMs backbones**) without introducing additional parameters.\nOur MMA design is intended to be generic, allowing for applications across various modalities, and scalable to accommodate diverse multimodal scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "yv7UtQz8We", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8141/Reviewer_VKts"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes Modality-Mutual Attention (MMA), a simple yet effective architecture for multimodal large language models (MLLMs). MMA unlocks the attention pathway from image tokens to text tokens during supervised fine-tuning, overcoming a key limitation of existing decoder-based models. MMA achieves a 5.5% average improvement across 12 benchmarks without adding parameters, demonstrates strong extensibility across model backbones, and provides transparent analysis and open-source results.", "review_text": "This paper proposes Modality-Mutual Attention (MMA), a simple yet effective architecture for multimodal large language models (MLLMs). MMA unlocks the attention pathway from image tokens to text tokens during supervised fine-tuning, overcoming a key limitation of existing decoder-based models. MMA achieves a 5.5% average improvement across 12 benchmarks without adding parameters, demonstrates strong extensibility across model backbones, and provides transparent analysis and open-source results.", "strengths": "- The approach of simply modifying the attention mask is both elegant and effective, directly tackling a core practical issue in VLMs. Its simplicity makes it highly exploratory and extensible, opening new avenues for deeper multimodal research.\n\n- The paper provides thoughtful analyses and visualizations to interpret the underlying attention mechanisms, supported by a relatively rich set of experimental evidence.", "weaknesses": "- While the paper presents a general masking formula for interleaved input, all experiments are conducted under the \"single-image + single-text\" setting. In benchmarks like CV-Bench and MMMU, samples with interleaved inputs are simplified to only the first image, leaving the robustness of MMA in multi-image, long-sequence, or multi-turn dialog scenarios unaddressed—these are currently key challenges for post-training VLMs.\n\n- MMA is effective only during supervised fine-tuning, and cannot be safely applied in the pretraining phase, which limits its broader impact and adaptability to foundational models. Existing experiments are performed only on small-scale models, and ablation studies show MMA's improvement in vision-knowledge mixed tasks is limited.\n\n- The explainability of the method is underdeveloped. The paper lacks a granular analysis of hallucination types and does not provide thorough theoretical derivations.", "questions": "- Results for Multi-image/Interleaved Sequences:\nCould the authors provide quantitative metrics and attention visualizations for MMA applied to truly multi-image inputs (e.g., interleaving 4–8 images)? As sequence length increases, is there evidence of attention dilution or gradient instability?\n\n- Relationship Between Mask Sparsity and Performance:\nIf the proportion of text tokens accessible to each image token is gradually increased (e.g., 10%, 30%, 50%, 100%), how does the performance curve behave? Is there a risk of “over-attention” where text tokens dominate and visual features become suppressed?\n\n- Scalability in Large-scale Training:\nAs multimodal training scales to tens of billions of samples, does MMA’s advantage diminish? Are there experiments or analyses on this phenomenon?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Modality-Mutual Attention (MMA), a simple yet effective architecture for multimodal large language models (MLLMs). MMA unlocks the attention pathway from image tokens to text tokens during supervised fine-tuning, overcoming a key limitation of existing decoder-based models. MMA achieves a 5.5% average improvement across 12 benchmarks without adding parameters, demonstrates strong extensibility across model backbones, and provides transparent analysis and open-source results.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The approach of simply modifying the attention mask is both elegant and effective, directly tackling a core practical issue in VLMs. Its simplicity makes it highly exploratory and extensible, opening new avenues for deeper multimodal research.\n\n- The paper provides thoughtful analyses and visualizations to interpret the underlying attention mechanisms, supported by a relatively rich set of experimental evidence.", "weaknesses": "- While the paper presents a general masking formula for interleaved input, all experiments are conducted under the \"single-image + single-text\" setting. In benchmarks like CV-Bench and MMMU, samples with interleaved inputs are simplified to only the first image, leaving the robustness of MMA in multi-image, long-sequence, or multi-turn dialog scenarios unaddressed—these are currently key challenges for post-training VLMs.\n\n- MMA is effective only during supervised fine-tuning, and cannot be safely applied in the pretraining phase, which limits its broader impact and adaptability to foundational models. Existing experiments are performed only on small-scale models, and ablation studies show MMA's improvement in vision-knowledge mixed tasks is limited.\n\n- The explainability of the method is underdeveloped. The paper lacks a granular analysis of hallucination types and does not provide thorough theoretical derivations.", "questions": "- Results for Multi-image/Interleaved Sequences:\nCould the authors provide quantitative metrics and attention visualizations for MMA applied to truly multi-image inputs (e.g., interleaving 4–8 images)? As sequence length increases, is there evidence of attention dilution or gradient instability?\n\n- Relationship Between Mask Sparsity and Performance:\nIf the proportion of text tokens accessible to each image token is gradually increased (e.g., 10%, 30%, 50%, 100%), how does the performance curve behave? Is there a risk of “over-attention” where text tokens dominate and visual features become suppressed?\n\n- Scalability in Large-scale Training:\nAs multimodal training scales to tens of billions of samples, does MMA’s advantage diminish? Are there experiments or analyses on this phenomenon?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761925583851}, {"id": "HpC2nMc2SQ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8141/Reviewer_xSEK"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper focuses on the core problem of \"visual-linguistic misalignment\" in multimodal large language models (MLLMs). It cuts into the causal attention mechanism at the bottom of the LLM architecture, and proposes a \"modal mutual attention (MMA) \" design. By modifying the attention mask, the image token can achieve cross-modal attention to the text token without adding additional parameters and computational costs. The research problem is accurately located (directly hitting the illusion pain points of current MLLMs), the method is concise and innovative (circumventing the limitations of traditional data/connector optimization), the experimental design is rigorous (covering 12 multimodal benchmarks + 2 model backbones), and the results are convincing (the average improvement is 5.5%, and the visual center task is significantly improved).", "review_text": "This paper focuses on the core problem of \"visual-linguistic misalignment\" in multimodal large language models (MLLMs). It cuts into the causal attention mechanism at the bottom of the LLM architecture, and proposes a \"modal mutual attention (MMA) \" design. By modifying the attention mask, the image token can achieve cross-modal attention to the text token without adding additional parameters and computational costs. The research problem is accurately located (directly hitting the illusion pain points of current MLLMs), the method is concise and innovative (circumventing the limitations of traditional data/connector optimization), the experimental design is rigorous (covering 12 multimodal benchmarks + 2 model backbones), and the results are convincing (the average improvement is 5.5%, and the visual center task is significantly improved).", "strengths": "1. Make clear the core limitation of traditional MLLMs - the \"decoder causal attention\" -based input order (image token before, text token after) results in \"forward modal (image) cannot obtain backward modal (text) information\", which in turn elicits the illusion of visually central tasks.\n2. The method design is simple and practical. First verify the necessity of \"cross-modal information flow\" through \"double-order training (DOT) \" (I & T and T & I two input sequential training), and then propose an MMA mechanism for the defect of \"doubling the training cost\" of DOT - only modify the attention mask (formula 5-6), allow the image token to focus on the text token, and still maintain the autoregressive characteristics when generating the response, taking into account \"validity\" and \"lightweight\". MMA does not increase parameters, does not increase computing costs (the amount of mask computation is consistent with the cause and effect of attention), and can be seamlessly integrated into existing MLLM training frameworks (only need to adjust the attention mask in the SFT stage), with a low threshold for landing and strong engineering practicality.\n3. Twelve benchmarks cover three types of tasks: \"General (MME, MMBench), Knowledge (MMMU, MathVista), Vision Center (POPE, CV-Bench) \", and two model backbones (Phi-3.5-Mini, LLaMA-3.2-3 B) verify generalization and avoid the chance of \"single model/benchmark\".", "weaknesses": "1. The paper mentions \"the use of standard causal attention in generating response TR\", but does not explicitly state \"whether modifying the attention mask at the input stage indirectly affects the autoregressive consistency of the generation process\" (e.g. whether there is modal confusion during generation). It is recommended to supplement the qualitative results of generative tasks (e.g. visual captioning) to verify that MMA does not compromise the autoregressive generation quality.\n2. The DOT performance in Table 2 is lower than the traditional causal attention on some benchmarks (such as MMB, SEED I) (such as the DOT of Phi-3.5-Mini is only 43.8 in MMB, lower than the traditional 64.9). The paper interprets it as \"different input order confusion model\", but does not rule out the influence of \"insufficient training steps\". It is recommended to supplement the DOT ablation experiment under \"longer training steps\" to verify whether the performance fluctuates due to insufficient training. If there is still no improvement, the specific mechanism of \"input order confusion\" needs to be further analyzed.", "questions": "The \"RR (Relation Reasoning) \" column value of \"AKI-4B\" in Table 9 is 0.65, which is significantly lower than that of other models (such as MMA is 67.8). It is speculated to be a clerical error (should be about 65.0). It needs to check the original data source and correct it.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on the core problem of \"visual-linguistic misalignment\" in multimodal large language models (MLLMs). It cuts into the causal attention mechanism at the bottom of the LLM architecture, and proposes a \"modal mutual attention (MMA) \" design. By modifying the attention mask, the image token can achieve cross-modal attention to the text token without adding additional parameters and computational costs. The research problem is accurately located (directly hitting the illusion pain points of current MLLMs), the method is concise and innovative (circumventing the limitations of traditional data/connector optimization), the experimental design is rigorous (covering 12 multimodal benchmarks + 2 model backbones), and the results are convincing (the average improvement is 5.5%, and the visual center task is significantly improved).", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. Make clear the core limitation of traditional MLLMs - the \"decoder causal attention\" -based input order (image token before, text token after) results in \"forward modal (image) cannot obtain backward modal (text) information\", which in turn elicits the illusion of visually central tasks.\n2. The method design is simple and practical. First verify the necessity of \"cross-modal information flow\" through \"double-order training (DOT) \" (I & T and T & I two input sequential training), and then propose an MMA mechanism for the defect of \"doubling the training cost\" of DOT - only modify the attention mask (formula 5-6), allow the image token to focus on the text token, and still maintain the autoregressive characteristics when generating the response, taking into account \"validity\" and \"lightweight\". MMA does not increase parameters, does not increase computing costs (the amount of mask computation is consistent with the cause and effect of attention), and can be seamlessly integrated into existing MLLM training frameworks (only need to adjust the attention mask in the SFT stage), with a low threshold for landing and strong engineering practicality.\n3. Twelve benchmarks cover three types of tasks: \"General (MME, MMBench), Knowledge (MMMU, MathVista), Vision Center (POPE, CV-Bench) \", and two model backbones (Phi-3.5-Mini, LLaMA-3.2-3 B) verify generalization and avoid the chance of \"single model/benchmark\".", "weaknesses": "1. The paper mentions \"the use of standard causal attention in generating response TR\", but does not explicitly state \"whether modifying the attention mask at the input stage indirectly affects the autoregressive consistency of the generation process\" (e.g. whether there is modal confusion during generation). It is recommended to supplement the qualitative results of generative tasks (e.g. visual captioning) to verify that MMA does not compromise the autoregressive generation quality.\n2. The DOT performance in Table 2 is lower than the traditional causal attention on some benchmarks (such as MMB, SEED I) (such as the DOT of Phi-3.5-Mini is only 43.8 in MMB, lower than the traditional 64.9). The paper interprets it as \"different input order confusion model\", but does not rule out the influence of \"insufficient training steps\". It is recommended to supplement the DOT ablation experiment under \"longer training steps\" to verify whether the performance fluctuates due to insufficient training. If there is still no improvement, the specific mechanism of \"input order confusion\" needs to be further analyzed.", "questions": "The \"RR (Relation Reasoning) \" column value of \"AKI-4B\" in Table 9 is 0.65, which is significantly lower than that of other models (such as MMA is 67.8). It is speculated to be a clerical error (should be about 65.0). It needs to check the original data source and correct it.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761820298171}, {"id": "4ydfIeTT8Q", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8141/Reviewer_Mhv4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper addresses the critical issue of vision-language misalignment in Multimodal Large Language Models (MLLMs), where generated textual responses are not factually grounded in the given image-text inputs, leading to object hallucinations. The proposed method, Modality-Mutual Attention (MMA), revisits the core architecture of MLLMs by fundamentally modifying the causal attention mechanism in decoder-only LLMs. It unlocks the attention mask to allow image tokens to actively attend and incorporate information from subsequent text tokens, thereby enabling dynamic, query-aware visual understanding. The authors also explore an intuitive Dual-Order Training approach but demonstrate that MMA is a more efficient and effective solution. A key advantage is that MMA achieves this without introducing any additional parameters or increasing computational costs. Extensive experiments show that the proposed method significantly outperforms strong baselines and state-of-the-art approaches across 12 diverse multimodal benchmarks, demonstrating improved performance particularly on vision-centric tasks.", "review_text": "This paper addresses the critical issue of vision-language misalignment in Multimodal Large Language Models (MLLMs), where generated textual responses are not factually grounded in the given image-text inputs, leading to object hallucinations. The proposed method, Modality-Mutual Attention (MMA), revisits the core architecture of MLLMs by fundamentally modifying the causal attention mechanism in decoder-only LLMs. It unlocks the attention mask to allow image tokens to actively attend and incorporate information from subsequent text tokens, thereby enabling dynamic, query-aware visual understanding. The authors also explore an intuitive Dual-Order Training approach but demonstrate that MMA is a more efficient and effective solution. A key advantage is that MMA achieves this without introducing any additional parameters or increasing computational costs. Extensive experiments show that the proposed method significantly outperforms strong baselines and state-of-the-art approaches across 12 diverse multimodal benchmarks, demonstrating improved performance particularly on vision-centric tasks.", "strengths": "1) Fundamental and Well-Motivated Problem: The paper identifies a fundamental architectural limitation of MLLMs as a root cause of hallucination.\n2) Efficient Solution: MMA requires no additional parameters, introduces no computational overhead, and can be seamlessly integrated into existing training pipelines, making it highly attractive for both research and deployment.\n3) Extensive and Convincing Empirical Validation: The method is thoroughly evaluated on 12 benchmarks using two LLM backbones, demonstrating consistent improvements, particularly on vision-centric tasks such as POPE and CV-Bench.", "weaknesses": "1) Limited Analysis of MMA's Inner Workings: A rigorous quantitative analysis of the cross-modal attention patterns is missing. For instance: How does the attention distribution from image tokens to text tokens evolve during training? Is there a consistent pattern (e.g., nouns, spatial relations) that image tokens learn to prioritize? This lack of in-depth analysis leaves the \"how\" of the performance gain somewhat as a black box.\n2) Lack of Novelty in Core Concept: The fundamental innovation is modest. As discussed, the use of non-causal, modality-aware attention masks is a well-established technique in multimodal learning. The paper does not adequately situate itself within this existing literature, creating an illusion of greater novelty than it possesses.\n3) Insufficient Exploration of Broader Applicability: The paper focuses on single image-text pairs. A key claim is that MMA is \"generic and scalable\", yet there is no empirical demonstration on interleaved multimodal data (image-text-image), multi-image inputs, or other modalities (e.g., audio).", "questions": "1) How does your work fundamentally differ from prior efforts that use non-causal or custom attention masks for multimodal modeling? Please clarify the specific novelty of the \"unlocking\" concept beyond its application to a new model class (decoder-only MLLMs).\n2) Have you identified any input conditions or task types where MMA fails to provide an improvement, or even degrades performance? For example, in scenarios with very verbose text or when the text contains substantial irrelevant information, could the additional cross-modal connections introduce noise? Discussing the limitations and failure modes would provide a more balanced view of the method's utility.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the critical issue of vision-language misalignment in Multimodal Large Language Models (MLLMs), where generated textual responses are not factually grounded in the given image-text inputs, leading to object hallucinations. The proposed method, Modality-Mutual Attention (MMA), revisits the core architecture of MLLMs by fundamentally modifying the causal attention mechanism in decoder-only LLMs. It unlocks the attention mask to allow image tokens to actively attend and incorporate information from subsequent text tokens, thereby enabling dynamic, query-aware visual understanding. The authors also explore an intuitive Dual-Order Training approach but demonstrate that MMA is a more efficient and effective solution. A key advantage is that MMA achieves this without introducing any additional parameters or increasing computational costs. Extensive experiments show that the proposed method significantly outperforms strong baselines and state-of-the-art approaches across 12 diverse multimodal benchmarks, demonstrating improved performance particularly on vision-centric tasks.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1) Fundamental and Well-Motivated Problem: The paper identifies a fundamental architectural limitation of MLLMs as a root cause of hallucination.\n2) Efficient Solution: MMA requires no additional parameters, introduces no computational overhead, and can be seamlessly integrated into existing training pipelines, making it highly attractive for both research and deployment.\n3) Extensive and Convincing Empirical Validation: The method is thoroughly evaluated on 12 benchmarks using two LLM backbones, demonstrating consistent improvements, particularly on vision-centric tasks such as POPE and CV-Bench.", "weaknesses": "1) Limited Analysis of MMA's Inner Workings: A rigorous quantitative analysis of the cross-modal attention patterns is missing. For instance: How does the attention distribution from image tokens to text tokens evolve during training? Is there a consistent pattern (e.g., nouns, spatial relations) that image tokens learn to prioritize? This lack of in-depth analysis leaves the \"how\" of the performance gain somewhat as a black box.\n2) Lack of Novelty in Core Concept: The fundamental innovation is modest. As discussed, the use of non-causal, modality-aware attention masks is a well-established technique in multimodal learning. The paper does not adequately situate itself within this existing literature, creating an illusion of greater novelty than it possesses.\n3) Insufficient Exploration of Broader Applicability: The paper focuses on single image-text pairs. A key claim is that MMA is \"generic and scalable\", yet there is no empirical demonstration on interleaved multimodal data (image-text-image), multi-image inputs, or other modalities (e.g., audio).", "questions": "1) How does your work fundamentally differ from prior efforts that use non-causal or custom attention masks for multimodal modeling? Please clarify the specific novelty of the \"unlocking\" concept beyond its application to a new model class (decoder-only MLLMs).\n2) Have you identified any input conditions or task types where MMA fails to provide an improvement, or even degrades performance? For example, in scenarios with very verbose text or when the text contains substantial irrelevant information, could the additional cross-modal connections introduce noise? Discussing the limitations and failure modes would provide a more balanced view of the method's utility.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761615767138}, {"id": "20tHSUNIsw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission8141/Reviewer_owr3"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper focuses on solving the limitations that the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text).  To address this problem, this paper proposes modality-mutual attention to enable image tokens to attend to text tokens. Experiments have shown that MMA achieves state-of-the-art performance in 12 multimodal understanding benchmarks.", "review_text": "This paper focuses on solving the limitations that the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text).  To address this problem, this paper proposes modality-mutual attention to enable image tokens to attend to text tokens. Experiments have shown that MMA achieves state-of-the-art performance in 12 multimodal understanding benchmarks.", "strengths": "1. This paper is well motivated.\n2. The proposed DOT method provides a plug-and-play advantage.\n3. The proposed MMA achieves significant improvement on multiple benchmarks.", "weaknesses": "1. The paper needs improvement in quality. For instance, clarify the significance of boldface and underline in Tab.1, which appears to differ from their meanings in Tab.2. There are repeated occurrences of underlines in Tab.2 with non-matching scores.\n2. As mentioned in the paper, DOT requires introducing doubled training overhead, as highlighted as a drawback.\n3. The novelty of the proposed method in this paper is limited.", "questions": "1.  Incorrect Improvements Result: In Tab.2, based on Phi-3.5-Mini-Instruct, MMA achieves 82.7 in POPE. The performance improvement should be (82.7-82.2)/82.2=0.6% instead of 1%.\n2. It would be more convincing to validate the proposed method based on other widely used MLLM models (e.g., Qwen2.5VL, LLaVA-OneVision, and InternVL3).\n3. In Tab.3, on nearly half of the benchmarks, AKI-4B's performance is worse than Qwen2-VL-2B's. Authors are requested to provide a reasonable explanation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on solving the limitations that the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text).  To address this problem, this paper proposes modality-mutual attention to enable image tokens to attend to text tokens. Experiments have shown that MMA achieves state-of-the-art performance in 12 multimodal understanding benchmarks.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This paper is well motivated.\n2. The proposed DOT method provides a plug-and-play advantage.\n3. The proposed MMA achieves significant improvement on multiple benchmarks.", "weaknesses": "1. The paper needs improvement in quality. For instance, clarify the significance of boldface and underline in Tab.1, which appears to differ from their meanings in Tab.2. There are repeated occurrences of underlines in Tab.2 with non-matching scores.\n2. As mentioned in the paper, DOT requires introducing doubled training overhead, as highlighted as a drawback.\n3. The novelty of the proposed method in this paper is limited.", "questions": "1.  Incorrect Improvements Result: In Tab.2, based on Phi-3.5-Mini-Instruct, MMA achieves 82.7 in POPE. The performance improvement should be (82.7-82.2)/82.2=0.6% instead of 1%.\n2. It would be more convincing to validate the proposed method based on other widely used MLLM models (e.g., Qwen2.5VL, LLaVA-OneVision, and InternVL3).\n3. In Tab.3, on nearly half of the benchmarks, AKI-4B's performance is worse than Qwen2-VL-2B's. Authors are requested to provide a reasonable explanation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761210914640}], "openreview_url": "https://openreview.net/forum?id=MmYiosUoay", "arxiv_id": "2503.02597", "paper_pdf": "papers/MmYiosUoay.pdf", "paper_pdf_sha256": "ea0d4fe2dbc08473e6f1fe100785ae95e241c0211fe74034a2f9e87c9bb55d3a", "paper_pdf_bytes": 6370324, "paper_pdf_source": "openreview", "code_url": "https://github.com/sony/aki", "code_repository": "sony/aki", "code_commit": "c7642317a4c9078efe87a5533360ff07f1fdc228", "code_archive": "repos/MmYiosUoay.zip", "code_archive_sha256": "79f33205054cd0bd6bb53cbdc07ee487d00df769d4c828e3a1749accd7113619", "code_archive_bytes": 4505644, "code_file_count": 41, "code_extensions": {".py": 37, ".sh": 4}, "github_disk_usage_kb": 4404, "github_languages": {"Python": 256569, "Shell": 1850}, "github_archived": false, "github_pushed_at": "2026-05-11T07:11:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/seeing-is-understanding-unlocking-causal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qUJsX3XMBH", "year": 2025, "status": "rejected", "title": "Rethinking Data Selection at Scale: Random Selection is Almost All You Need", "authors": ["Tingyu Xia", "Bowen Yu", "Kai Dang", "An Yang", "Yuan Wu", "Yuan Tian", "Yi Chang", "Junyang Lin"], "authorids": ["~Tingyu_Xia1", "~Bowen_Yu3", "~Kai_Dang1", "~An_Yang1", "~Yuan_Wu2", "~Yuan_Tian6", "~Yi_Chang4", "~Junyang_Lin1"], "authors_source": "OpenReview API", "abstract": "Supervised fine-tuning (SFT) is crucial for aligning Large Language Models (LLMs) with human instructions. The primary goal during SFT is to select a small yet representative subset of training data from the larger pool, such that fine-tuning with this subset achieves results comparable to or even exceeding those obtained using the entire dataset. However, most existing data selection techniques are designed for small-scale data pools, which fail to meet the demands of real-world SFT scenarios. In this paper, we replicated several self-scoring methods—those that do not rely on external model assistance—on two million-scale datasets, and found that nearly all methods struggled to significantly outperform random selection when dealing with such large-scale data pools. Moreover, our comparisons suggest that, during SFT, diversity in data selection is more critical than simply focusing on high-quality data. We also analyzed the limitations of several current approaches, explaining why they perform poorly on large-scale datasets and why they are unsuitable for such contexts. Finally, we found that filtering data by token length offers a stable and efficient method for improving results. This approach, particularly when training on long-text data, proves highly beneficial for relatively weaker base models, such as Llama3.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "Cc2sPQljz6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9357/Reviewer_2efs"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "In this paper authors proposed a data filtration technique for the SFT (supervised fine-tuning phase) of LLM training.\nRigorously testing existing SOTA methods that were proven to work on the small scale datasets, authors found that \nfiltering data by token length offers an efficient way for improving downstream results further. In addition to that authors demonstrated that SOTA methods fall short in quality when handling large-scale 1T datasets. Authors demonstrated that data diversity plays the most important role in the SFT phase, thus quality filtering is suboptimal.", "review_text": "In this paper authors proposed a data filtration technique for the SFT (supervised fine-tuning phase) of LLM training.\nRigorously testing existing SOTA methods that were proven to work on the small scale datasets, authors found that \nfiltering data by token length offers an efficient way for improving downstream results further. In addition to that authors demonstrated that SOTA methods fall short in quality when handling large-scale 1T datasets. Authors demonstrated that data diversity plays the most important role in the SFT phase, thus quality filtering is suboptimal.", "strengths": "Constructing a high-quality multi-corpora dataset that is useful for training LLMs is a hard and important task. Input data is crucial for making a high-quality reliable and safe LLM at scale. Authors explore the topic of improving the dataset construction process and suggest a simplified framework that does not require rigorous data cleaning at scale. Considering that developing and using filtration techniques at scale might present computational challenges, this paper makes a significant contribution to the field with their simplified proposal.", "weaknesses": "One of the challenges with assessing the quality of LLMs is the quality and generalizability of the downstream tasks. Majority of the times that amount of the overlap between training and downstream data can cause skewed assessments in the quality of data processing techniques. I could not find any statistics reported by the authors for the amount of the overlap present in their training corpuses and the downstream tasks. It is important to understand that before drawing any conclusions with regards to the data filtration techniques. \nIt might be the case that some of the filtration strategies are heavily biased towards the examples present in the downstream tasks, and not necessarily improve model's generazability qualities.", "questions": "I would like to see the stats on the overlap present in the tokens from your SFT datasets and downstream tasks. It would also be useful to know whether any deduplication method was applied on top of the SFT datasets to eliminate the effect of model's memorization.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper authors proposed a data filtration technique for the SFT (supervised fine-tuning phase) of LLM training.\nRigorously testing existing SOTA methods that were proven to work on the small scale datasets, authors found that \nfiltering data by token length offers an efficient way for improving downstream results further. In addition to that authors demonstrated that SOTA methods fall short in quality when handling large-scale 1T datasets. Authors demonstrated that data diversity plays the most important role in the SFT phase, thus quality filtering is suboptimal.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Constructing a high-quality multi-corpora dataset that is useful for training LLMs is a hard and important task. Input data is crucial for making a high-quality reliable and safe LLM at scale. Authors explore the topic of improving the dataset construction process and suggest a simplified framework that does not require rigorous data cleaning at scale. Considering that developing and using filtration techniques at scale might present computational challenges, this paper makes a significant contribution to the field with their simplified proposal.", "weaknesses": "One of the challenges with assessing the quality of LLMs is the quality and generalizability of the downstream tasks. Majority of the times that amount of the overlap between training and downstream data can cause skewed assessments in the quality of data processing techniques. I could not find any statistics reported by the authors for the amount of the overlap present in their training corpuses and the downstream tasks. It is important to understand that before drawing any conclusions with regards to the data filtration techniques. \nIt might be the case that some of the filtration strategies are heavily biased towards the examples present in the downstream tasks, and not necessarily improve model's generazability qualities.", "questions": "I would like to see the stats on the overlap present in the tokens from your SFT datasets and downstream tasks. It would also be useful to know whether any deduplication method was applied on top of the SFT datasets to eliminate the effect of model's memorization.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730747475931}, {"id": "4twc1auQh0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9357/Reviewer_yhuK"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper investigates the effectiveness of data selection methods for supervised fine-tuning (SFT) of Large Language Models (LLMs) at scale. Through extensive experimentation on million-scale datasets, the research reveals that most existing self-scoring data selection methods, which don't require external model assistance, fail to significantly outperform random selection when applied to large-scale data pools.", "review_text": "This paper investigates the effectiveness of data selection methods for supervised fine-tuning (SFT) of Large Language Models (LLMs) at scale. Through extensive experimentation on million-scale datasets, the research reveals that most existing self-scoring data selection methods, which don't require external model assistance, fail to significantly outperform random selection when applied to large-scale data pools.", "strengths": "1. This paper studies an interesting yet important task of selecting training data for supervised fine-tuning (SFT) in large language models.\n\n2. The experimental results reveal that various data selection methods struggle to significantly outperform random selection, a finding that will likely be beneficial to practitioners.", "weaknesses": "1. Some of the conclusions presented are already established in existing literature. For instance, paper [1] has previously demonstrated the importance of data diversity in SFT data selection with large-scale datasets, while paper [2] has shown the effectiveness of selecting longer answers for data selection. It appears the authors overlooked these references, which limits the novelty and contribution of this work.\n\n2. Although the paper identifies that certain data selection methods fail to outperform random selection, it lacks a thorough analysis—either theoretical or empirical—explaining why these approaches underperform. This absence of deeper exploration reduces the overall depth and insight provided by the study, which is required by top-tier ML conferences.\n\n3. The paper does not provide a clear definition of what is an \"extensive SFT dataset.\" In practical applications, it remains unclear how to determine whether a dataset qualifies as \"extensive,\" which could lead to ambiguity in interpreting the paper's findings and recommendations.\n\n\n[1] Bukharin, Alexander, and Tuo Zhao. \"Data diversity matters for robust instruction tuning.\" arXiv preprint arXiv:2311.14736 (2023).\n\n[2] Zhao, Hao, et al. \"Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning.\" Forty-first International Conference on Machine Learning.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the effectiveness of data selection methods for supervised fine-tuning (SFT) of Large Language Models (LLMs) at scale. Through extensive experimentation on million-scale datasets, the research reveals that most existing self-scoring data selection methods, which don't require external model assistance, fail to significantly outperform random selection when applied to large-scale data pools.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. This paper studies an interesting yet important task of selecting training data for supervised fine-tuning (SFT) in large language models.\n\n2. The experimental results reveal that various data selection methods struggle to significantly outperform random selection, a finding that will likely be beneficial to practitioners.", "weaknesses": "1. Some of the conclusions presented are already established in existing literature. For instance, paper [1] has previously demonstrated the importance of data diversity in SFT data selection with large-scale datasets, while paper [2] has shown the effectiveness of selecting longer answers for data selection. It appears the authors overlooked these references, which limits the novelty and contribution of this work.\n\n2. Although the paper identifies that certain data selection methods fail to outperform random selection, it lacks a thorough analysis—either theoretical or empirical—explaining why these approaches underperform. This absence of deeper exploration reduces the overall depth and insight provided by the study, which is required by top-tier ML conferences.\n\n3. The paper does not provide a clear definition of what is an \"extensive SFT dataset.\" In practical applications, it remains unclear how to determine whether a dataset qualifies as \"extensive,\" which could lead to ambiguity in interpreting the paper's findings and recommendations.\n\n\n[1] Bukharin, Alexander, and Tuo Zhao. \"Data diversity matters for robust instruction tuning.\" arXiv preprint arXiv:2311.14736 (2023).\n\n[2] Zhao, Hao, et al. \"Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning.\" Forty-first International Conference on Machine Learning.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730691879849}, {"id": "thiE3wsuYV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9357/Reviewer_HChF"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper studies the efficacy of various data filtering strategies when applied to large-scale SFT datasets (i.e.., OpenHermes2.5, WildChat). While these same strategies often outperform random selection on smaller-scale SFT datasets they are shown to be surprisingly  less impactful when applied to these larger-scale (state-of-the-art) open-source datasets. On these datasets, approaches that target diversity seem to be more effective than those that target quality. Finally, the authors test their own length-based filtering strategy, which outperforms the other methods when the base model is Llama3-8B.", "review_text": "This paper studies the efficacy of various data filtering strategies when applied to large-scale SFT datasets (i.e.., OpenHermes2.5, WildChat). While these same strategies often outperform random selection on smaller-scale SFT datasets they are shown to be surprisingly  less impactful when applied to these larger-scale (state-of-the-art) open-source datasets. On these datasets, approaches that target diversity seem to be more effective than those that target quality. Finally, the authors test their own length-based filtering strategy, which outperforms the other methods when the base model is Llama3-8B.", "strengths": "**Originality and Significance.** This paper performs a valuable and novel (to the best of my knowledge) empirical study to see whether existing SFT data selection methods remain effective when applied on larger SFT datasets. Because these larger SFT datasets are generally better than the smaller datasets these methods were previously tested on, it is important to see whether filtering works in this more state of the art regime. The main findings that (i) many of these methods no longer are as impactful; (ii) diversity is more important than quality-filtering are therefore quite significant and potentially highly-relevant to future SFT data curation efforts.\n\n**Quality.** This paper implements a good number of baselines and also is thorough about trying multiple base models and dataset sizes (post data selection). These variations are nice to test the robustness of the main takeaways.\n\n**Clarity.** Overall, most of the paper is easy to follow with the exception of the description of baseline methods (see W3).", "weaknesses": "**(W1) Isolating dataset size.** A central claim in the paper is that dataset size is what causes differences in the behavior of the baselines. However, there are also other differences between OH2.5/WildChat and older datasets besides data quantity; e.g., quality-based selection may be less impactful simply because these larger datasets are already the result of more careful curation and thus higher-quality compared to older ones. An experiment that would better isolate data quantity (and control for quality) would be to run the existing experiments on progressively smaller random subsets of OH2.5/WildChat (e.g., what happens when the starting pool is 10K/100K examples of OH2.5 v.s. 1M?)\n\n**(W2) Understanding run-to-run variance / Fairness of comparison to random.** The stated takeaway from the results is that *\"it is more efficient to randomly select training data...random selection reduces costs and yields superior training results\"* but I don't know if that's fully borne out by the results.\n- For the Qwen2 results (which achieve highest overall perforamnce), in the 50K regime, the best average results are achieved by non-random curation methods (Diverse and ZIP each score 1% higher than the max over the random runs in Tables 5 and 6 respectively). In the 10K regime, this is also almost true with the exception of 1/5 of the random runs.\n- In general, I'm not sure its fair to compare the max over 5 random runs to a single run of a selection method. if only 1 of 5 runs ends up doing better, then in practice you'd likely have to try multiple times when randomly sampling a subset that outperforms non-random selection, increasing costs. \n- Ideally, it would be nice to show the average and stdev across the 5 random runs and to also get a sense of run-to-run variation for the non-random methods (though I acknowledge this might be costly). \n\n**(W3) Clarity when describing existing methods.** I currently found Sec 3. a bit hard to follow. It may be helpful to define up front unified notation for variables that appear in multiple methods instead of using the differing notation from each of the original papers (e.g. whether a sample is given by $z$ or $x$ or $n$). Also, in some places, symbols are currently used without an explicit definition (e.g. $\\hat{\\Gamma}$ in LESS, $\\mathcal{C}$ in ZIP). \n\n**Miscellaneous Notes**\n- In captions for Tables 1 and 2, \"the suboptimal score\" is confusing terminology, I believe you meant \"second highest score\"? \n- Given that originally LESS is a targeted selection method, I'm not sure it's completely aligned with the spirit of the original method to set the target as a subset of the overall data pool? I suspect what might be leading to bad performance is that this may actually result in selecting a less diverse dataset (by prioritizing examples that represent the majority).", "questions": "(Q1) Could you share more details about Figure 1, given that the difference between small/large regimes is a key aspect of the paper? In particular, which datasets were used for the 10K-300K scale in Figure 1? Are these already random subsets of OH2.5/WildChat (basically what i suggested in W1) or are they some other datasets (e.g. Dolly, ShareGPT)?\n\n(Q2) What is the cost of the fine-tuning runs in your experiment setup? It would be helpful to know this relative to the costs of running the various data curation algorithms (when discussing them in 5.3).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the efficacy of various data filtering strategies when applied to large-scale SFT datasets (i.e.., OpenHermes2.5, WildChat). While these same strategies often outperform random selection on smaller-scale SFT datasets they are shown to be surprisingly  less impactful when applied to these larger-scale (state-of-the-art) open-source datasets. On these datasets, approaches that target diversity seem to be more effective than those that target quality. Finally, the authors test their own length-based filtering strategy, which outperforms the other methods when the base model is Llama3-8B.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "**Originality and Significance.** This paper performs a valuable and novel (to the best of my knowledge) empirical study to see whether existing SFT data selection methods remain effective when applied on larger SFT datasets. Because these larger SFT datasets are generally better than the smaller datasets these methods were previously tested on, it is important to see whether filtering works in this more state of the art regime. The main findings that (i) many of these methods no longer are as impactful; (ii) diversity is more important than quality-filtering are therefore quite significant and potentially highly-relevant to future SFT data curation efforts.\n\n**Quality.** This paper implements a good number of baselines and also is thorough about trying multiple base models and dataset sizes (post data selection). These variations are nice to test the robustness of the main takeaways.\n\n**Clarity.** Overall, most of the paper is easy to follow with the exception of the description of baseline methods (see W3).", "weaknesses": "**(W1) Isolating dataset size.** A central claim in the paper is that dataset size is what causes differences in the behavior of the baselines. However, there are also other differences between OH2.5/WildChat and older datasets besides data quantity; e.g., quality-based selection may be less impactful simply because these larger datasets are already the result of more careful curation and thus higher-quality compared to older ones. An experiment that would better isolate data quantity (and control for quality) would be to run the existing experiments on progressively smaller random subsets of OH2.5/WildChat (e.g., what happens when the starting pool is 10K/100K examples of OH2.5 v.s. 1M?)\n\n**(W2) Understanding run-to-run variance / Fairness of comparison to random.** The stated takeaway from the results is that *\"it is more efficient to randomly select training data...random selection reduces costs and yields superior training results\"* but I don't know if that's fully borne out by the results.\n- For the Qwen2 results (which achieve highest overall perforamnce), in the 50K regime, the best average results are achieved by non-random curation methods (Diverse and ZIP each score 1% higher than the max over the random runs in Tables 5 and 6 respectively). In the 10K regime, this is also almost true with the exception of 1/5 of the random runs.\n- In general, I'm not sure its fair to compare the max over 5 random runs to a single run of a selection method. if only 1 of 5 runs ends up doing better, then in practice you'd likely have to try multiple times when randomly sampling a subset that outperforms non-random selection, increasing costs. \n- Ideally, it would be nice to show the average and stdev across the 5 random runs and to also get a sense of run-to-run variation for the non-random methods (though I acknowledge this might be costly). \n\n**(W3) Clarity when describing existing methods.** I currently found Sec 3. a bit hard to follow. It may be helpful to define up front unified notation for variables that appear in multiple methods instead of using the differing notation from each of the original papers (e.g. whether a sample is given by $z$ or $x$ or $n$). Also, in some places, symbols are currently used without an explicit definition (e.g. $\\hat{\\Gamma}$ in LESS, $\\mathcal{C}$ in ZIP). \n\n**Miscellaneous Notes**\n- In captions for Tables 1 and 2, \"the suboptimal score\" is confusing terminology, I believe you meant \"second highest score\"? \n- Given that originally LESS is a targeted selection method, I'm not sure it's completely aligned with the spirit of the original method to set the target as a subset of the overall data pool? I suspect what might be leading to bad performance is that this may actually result in selecting a less diverse dataset (by prioritizing examples that represent the majority).", "questions": "(Q1) Could you share more details about Figure 1, given that the difference between small/large regimes is a key aspect of the paper? In particular, which datasets were used for the 10K-300K scale in Figure 1? Are these already random subsets of OH2.5/WildChat (basically what i suggested in W1) or are they some other datasets (e.g. Dolly, ShareGPT)?\n\n(Q2) What is the cost of the fine-tuning runs in your experiment setup? It would be helpful to know this relative to the costs of running the various data curation algorithms (when discussing them in 5.3).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730592161414}, {"id": "NG18D9Ims8", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9357/Reviewer_mJQN"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "In this paper, the authors study the effectiveness of data selection methods on large-scale datasets, which is largely unexplored. With experiments on large-scale datasets, the authors find that nearly existing data selection methods do not significantly outperform random selection on large-scale datasets. This paper also tries to explain the reason that leads to inferior performance of existing methods on a larger data pool. Ultimately, the authors propose a more effective data selection method based on token length. This paper offers researchers a new perspective on data selection in large language models.", "review_text": "In this paper, the authors study the effectiveness of data selection methods on large-scale datasets, which is largely unexplored. With experiments on large-scale datasets, the authors find that nearly existing data selection methods do not significantly outperform random selection on large-scale datasets. This paper also tries to explain the reason that leads to inferior performance of existing methods on a larger data pool. Ultimately, the authors propose a more effective data selection method based on token length. This paper offers researchers a new perspective on data selection in large language models.", "strengths": "The authors study the effectiveness of data selection methods on large-scale datasets, and find the inferior performance of existing methods when dealing with large-scale data pools. At the same time,  the authors try to analyze the reason that leads to inferior performance of existing methods on a larger data pool.", "weaknesses": "1. **Title Reflecting Content**: The title of this paper suggests that ”random\nselection is almost all you need”, while results show that the diversity-based\nselection method achieves superior performance with an average score of\n59.58 under Qwen2-7B in Table 1. The result does not indicate that random\nselection is a better method. In this paper, it seems to show that\ndiversity plays an important role in data selection for large-scale datasets. A more representative title improves clarity\nand helps set readers’ expectations.\n\n2. **Insufficient experiments**: This paper’s results show that existing data\nselection methods do not significantly outperform random selection on large-scale datasets. However, the authors only perform experiments on two models. Are there similar outcomes for different LLMs (e.g., different-sized models)?", "questions": "**Questions**: This paper proposes a data selection method by token\nlength to achieve optimal results. However, as shown in Table 3, does the Qwen2-7B model not get better results with this method compared with\nexisting methods? By the way, why is this selection method beneficial? In\nother words, what is the motivation for this approach? Moreover, what\ndoes it mean to ”reduce the uncertainty caused by randomness”?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors study the effectiveness of data selection methods on large-scale datasets, which is largely unexplored. With experiments on large-scale datasets, the authors find that nearly existing data selection methods do not significantly outperform random selection on large-scale datasets. This paper also tries to explain the reason that leads to inferior performance of existing methods on a larger data pool. Ultimately, the authors propose a more effective data selection method based on token length. This paper offers researchers a new perspective on data selection in large language models.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The authors study the effectiveness of data selection methods on large-scale datasets, and find the inferior performance of existing methods when dealing with large-scale data pools. At the same time,  the authors try to analyze the reason that leads to inferior performance of existing methods on a larger data pool.", "weaknesses": "1. **Title Reflecting Content**: The title of this paper suggests that ”random\nselection is almost all you need”, while results show that the diversity-based\nselection method achieves superior performance with an average score of\n59.58 under Qwen2-7B in Table 1. The result does not indicate that random\nselection is a better method. In this paper, it seems to show that\ndiversity plays an important role in data selection for large-scale datasets. A more representative title improves clarity\nand helps set readers’ expectations.\n\n2. **Insufficient experiments**: This paper’s results show that existing data\nselection methods do not significantly outperform random selection on large-scale datasets. However, the authors only perform experiments on two models. Are there similar outcomes for different LLMs (e.g., different-sized models)?", "questions": "**Questions**: This paper proposes a data selection method by token\nlength to achieve optimal results. However, as shown in Table 3, does the Qwen2-7B model not get better results with this method compared with\nexisting methods? By the way, why is this selection method beneficial? In\nother words, what is the motivation for this approach? Moreover, what\ndoes it mean to ”reduce the uncertainty caused by randomness”?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730356816555}, {"id": "zRVyZCD4V7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9357/Reviewer_MLme"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper investigates supervised fine-tuning (SFT) for large language models (LLMs), specifically focusing on selecting a representative subset of training data that performs as well as or better than using the full dataset. It reveals that existing data selection methods, particularly self-scoring techniques, struggle to outperform random selection on million-scale datasets and highlights the importance of diversity over mere data quality in large-scale scenarios. Additionally, the authors find that filtering data by token length enhances SFT, especially for weaker base models like Llama3, offering a stable and efficient improvement method.", "review_text": "This paper investigates supervised fine-tuning (SFT) for large language models (LLMs), specifically focusing on selecting a representative subset of training data that performs as well as or better than using the full dataset. It reveals that existing data selection methods, particularly self-scoring techniques, struggle to outperform random selection on million-scale datasets and highlights the importance of diversity over mere data quality in large-scale scenarios. Additionally, the authors find that filtering data by token length enhances SFT, especially for weaker base models like Llama3, offering a stable and efficient improvement method.", "strengths": "The finding that \"Most self-scoring data selection techniques do not significantly outperform random selection on large-scale datasets\" could contribute novel insights to the community.", "weaknesses": "- The motivation for this research is unclear. The authors claim that \"This discrepancy creates a gap between the present SFT data selection strategies and real-world applications,\" yet it remains unclear what these real-world applications are.\n\n- Finding 2, namely, \"Data quality-based selection methods are more effective than data diversity-based methods when dealing with a small-scale dataset from a single source,\" is interesting, but exploring why this discrepancy occurs would be more valuable.\n\n- Finding 3, that \"token length could be used for data filtering,\" does not fit well within the overall narrative. In general, while the paper presents three findings, they are neither well-motivated nor well-connected. As an exploratory paper, its contribution feels limited.\n\n- When using the term “significantly better,” please provide a p-value to substantiate the statistical significance of the results.\n\n- Clarification:\n\n  - The introduction should clarify what “scoring” means in this context and better motivate the need for scoring. Additionally, for the self-scoring approaches, the authors further categorise their techniques into two types: data quality-based methods and data diversity-based methods, but this distinction needs clearer explanation in the context.\n\n    Although further detail is provided in Section 3, the introduction is challenging to follow.\n\n  - In Figure 1, \"BBH benchmark\" requires a reference.\n\n  - Line 58: \"in practical applications\" needs clarification.\n\n- Typos:\n\n  - Line 39: missing a space before the citation.\n  - Table 1: some of the best numbers are not in bold.\n  - Line 324: missing a space before the bracket.", "questions": "- What real-world applications could benefit from the outcomes of this research?\n- What is the difference between data quality-based methods and data diversity-based approaches?\n- Why is utilising token length in SFT beneficial for weaker base language models, like Llama3-8B?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates supervised fine-tuning (SFT) for large language models (LLMs), specifically focusing on selecting a representative subset of training data that performs as well as or better than using the full dataset. It reveals that existing data selection methods, particularly self-scoring techniques, struggle to outperform random selection on million-scale datasets and highlights the importance of diversity over mere data quality in large-scale scenarios. Additionally, the authors find that filtering data by token length enhances SFT, especially for weaker base models like Llama3, offering a stable and efficient improvement method.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The finding that \"Most self-scoring data selection techniques do not significantly outperform random selection on large-scale datasets\" could contribute novel insights to the community.", "weaknesses": "- The motivation for this research is unclear. The authors claim that \"This discrepancy creates a gap between the present SFT data selection strategies and real-world applications,\" yet it remains unclear what these real-world applications are.\n\n- Finding 2, namely, \"Data quality-based selection methods are more effective than data diversity-based methods when dealing with a small-scale dataset from a single source,\" is interesting, but exploring why this discrepancy occurs would be more valuable.\n\n- Finding 3, that \"token length could be used for data filtering,\" does not fit well within the overall narrative. In general, while the paper presents three findings, they are neither well-motivated nor well-connected. As an exploratory paper, its contribution feels limited.\n\n- When using the term “significantly better,” please provide a p-value to substantiate the statistical significance of the results.\n\n- Clarification:\n\n  - The introduction should clarify what “scoring” means in this context and better motivate the need for scoring. Additionally, for the self-scoring approaches, the authors further categorise their techniques into two types: data quality-based methods and data diversity-based methods, but this distinction needs clearer explanation in the context.\n\n    Although further detail is provided in Section 3, the introduction is challenging to follow.\n\n  - In Figure 1, \"BBH benchmark\" requires a reference.\n\n  - Line 58: \"in practical applications\" needs clarification.\n\n- Typos:\n\n  - Line 39: missing a space before the citation.\n  - Table 1: some of the best numbers are not in bold.\n  - Line 324: missing a space before the bracket.", "questions": "- What real-world applications could benefit from the outcomes of this research?\n- What is the difference between data quality-based methods and data diversity-based approaches?\n- Why is utilising token length in SFT beneficial for weaker base language models, like Llama3-8B?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730160849608}], "openreview_url": "https://openreview.net/forum?id=qUJsX3XMBH", "arxiv_id": "2410.09335", "paper_pdf": "papers/qUJsX3XMBH.pdf", "paper_pdf_sha256": "65388a79d3f34feea031cce7c8ec4c34e132619e847fa6aa2cc6ef8330547d68", "paper_pdf_bytes": 418887, "paper_pdf_source": "openreview", "code_url": "https://github.com/xiatingyu/SFT-DataSelection-at-scale", "code_repository": "xiatingyu/SFT-DataSelection-at-scale", "code_commit": "5a635e42f84640f3c6b23a526fcfb1f251765f2b", "code_archive": "repos/qUJsX3XMBH.zip", "code_archive_sha256": "6e320586ae82a8f633e014d8d354d3bd5293a53650dc04c18c406c41e102c595", "code_archive_bytes": 375245, "code_file_count": 70, "code_extensions": {".py": 39, ".sh": 30, ".ipynb": 1}, "github_disk_usage_kb": 313, "github_languages": {"Jupyter Notebook": 285690, "Python": 227534, "Shell": 24823}, "github_archived": false, "github_pushed_at": "2025-02-09T03:05:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/rethinking-data-selection-at-scale-random"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9Z0yB8rmQ2", "year": 2024, "status": "rejected", "title": "Lyra: Orchestrating Dual Correction in Automated Theorem Proving", "authors": ["Chuanyang Zheng", "Haiming Wang", "Enze Xie", "Zhengying Liu", "Jiankai Sun", "Huajian Xin", "Jianhao Shen", "Zhenguo Li", "Yu Li"], "authorids": ["~Chuanyang_Zheng3", "~Haiming_Wang1", "~Enze_Xie1", "~Zhengying_Liu2", "~Jiankai_Sun6", "~Huajian_Xin1", "~Jianhao_Shen1", "~Zhenguo_Li1", "~Yu_Li1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) present an intriguing avenue for exploration in the field of formal theorem proving. Nevertheless, their full potential, particularly concerning the mitigation of hallucinations and refinement through prover error messages, remains an area that has yet to be thoroughly investigated. To enhance the effectiveness of LLMs in the field, we introduce the Lyra, a new framework that employs two distinct correction mechanisms: Tool Correction (TC) and Conjecture Correction (CC). To implement Tool Correction in the post-processing of formal proofs, we leverage prior knowledge to utilize predefined prover tools (e.g., Sledgehammer) for guiding the replacement of incorrect tools. Tool Correction significantly contributes to mitigating hallucinations, thereby improving the overall accuracy of the proof. In addition, we introduce Conjecture Correction, an error feedback mechanism designed to interact with prover to refine formal proof conjectures with prover error messages. Compared to the previous refinement framework, the proposed Conjecture Correction refines generation with instruction but does not collect paired (generation, error & refinement) prompts. Our method has achieved state-of-the-art (SOTA) performance on both miniF2F validation (48.0% → 55.3%) and test (45.5% → 51.2%). We also present 3 IMO problems solved by Lyra. We believe Tool Correction (post-process for hallucination mitigation) and Conjecture Correction (subgoal adjustment from interaction with environment) could provide a promising avenue for future research in this field.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "AwfBTIhIrF", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1096/Reviewer_Wc5t"], "rating": "6: marginally above the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes two methods for postprocessing and fixing the errors in the proof steps generated by LLM for theorem proving. The first method is Tool Correction, which tries a list of automation tactic such as sledgehammer, auto and arith on the fails steps. The second method is Conjecture Correction, which asks LLM to regenerate the proof step based on the error message or simply regenerate this step to start a new iteration. Experiments show that TC and CC could improve the performance significantly on the miniF2F valid (50.4%->55.3%) and test set( 42.6%->51.2%), and achieves the new state-of-the-art results.\n\n==================================\nPost-rebuttal:\nAfter reading authors' responses and other reviews, I upgraded my score to 6. \nAuthors' response largely addressed my questions. I still think the technical novelty of this paper is relatively weak. But the extensive experiments and ablation studies (including the new experiment results presented in the rebuttal) are very solid and could be helpful to the AITP community.", "review_text": "This paper proposes two methods for postprocessing and fixing the errors in the proof steps generated by LLM for theorem proving. The first method is Tool Correction, which tries a list of automation tactic such as sledgehammer, auto and arith on the fails steps. The second method is Conjecture Correction, which asks LLM to regenerate the proof step based on the error message or simply regenerate this step to start a new iteration. Experiments show that TC and CC could improve the performance significantly on the miniF2F valid (50.4%->55.3%) and test set( 42.6%->51.2%), and achieves the new state-of-the-art results.\n\n==================================\nPost-rebuttal:\nAfter reading authors' responses and other reviews, I upgraded my score to 6. \nAuthors' response largely addressed my questions. I still think the technical novelty of this paper is relatively weak. But the extensive experiments and ablation studies (including the new experiment results presented in the rebuttal) are very solid and could be helpful to the AITP community.", "strengths": "Both Tool Correction and Conjecture Correction are technically sound. Tool Correction implies the insights that LLM is better at generating the next conjecture to prove than closing the conjecture. Extensive search is helpful to close the conjecture.", "weaknesses": "The paper doesn't have much novelty in terms of the approach. It seems that the set of 11 tactics in TC have been proposed in DSP. The method of appending error message for self-debugging and generating multiple candidates have been commonly used for code generation.", "questions": "1 I think one good baseline would be replace Minerva and Codex in DSP with GPT4. So we still have GPT4 to generate the proof sketch (the intermediate conjectures) but use 11 tactics + sledgehammer to close the open goals. \n2 Could you calculate the number of wrong proof steps fixed by each tactic in TC?\n3 The proposed method adds a lot more computation. How much time would be token by calling GPT4, TC and CC? If we set a time limit for each question like 10 or 30 minutes, what would be the performance of TC/CC compared with the GPT4 baseline and DSP+GPT4 baseline?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes two methods for postprocessing and fixing the errors in the proof steps generated by LLM for theorem proving. The first method is Tool Correction, which tries a list of automation tactic such as sledgehammer, auto and arith on the fails steps. The second method is Conjecture Correction, which asks LLM to regenerate the proof step based on the error message or simply regenerate this step to start a new iteration. Experiments show that TC and CC could improve the performance significantly on the miniF2F valid (50.4%->55.3%) and test set( 42.6%->51.2%), and achieves the new state-of-the-art results.\n\n==================================\nPost-rebuttal:\nAfter reading authors' responses and other reviews, I upgraded my score to 6. \nAuthors' response largely addressed my questions. I still think the technical novelty of this paper is relatively weak. But the extensive experiments and ablation studies (including the new experiment results presented in the rebuttal) are very solid and could be helpful to the AITP community.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "Both Tool Correction and Conjecture Correction are technically sound. Tool Correction implies the insights that LLM is better at generating the next conjecture to prove than closing the conjecture. Extensive search is helpful to close the conjecture.", "weaknesses": "The paper doesn't have much novelty in terms of the approach. It seems that the set of 11 tactics in TC have been proposed in DSP. The method of appending error message for self-debugging and generating multiple candidates have been commonly used for code generation.", "questions": "1 I think one good baseline would be replace Minerva and Codex in DSP with GPT4. So we still have GPT4 to generate the proof sketch (the intermediate conjectures) but use 11 tactics + sledgehammer to close the open goals. \n2 Could you calculate the number of wrong proof steps fixed by each tactic in TC?\n3 The proposed method adds a lot more computation. How much time would be token by calling GPT4, TC and CC? If we set a time limit for each question like 10 or 30 minutes, what would be the performance of TC/CC compared with the GPT4 baseline and DSP+GPT4 baseline?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics concern.", "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699046111527}, {"id": "R3jZBvFWYw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1096/Reviewer_1Jjk"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents an LLM-augmented automated theorem proving framework called Lyra in a formal theorem proving environment. The distinguishing features of Lyra include Tool Correction (TC) and Conjecture Correction (CC), which respectively post-edit formal proofs emitted from LLMs given feedback from the proving environment. Good performance has been shown over the miniF2F dataset with 3 IMO problems being solved.", "review_text": "This paper presents an LLM-augmented automated theorem proving framework called Lyra in a formal theorem proving environment. The distinguishing features of Lyra include Tool Correction (TC) and Conjecture Correction (CC), which respectively post-edit formal proofs emitted from LLMs given feedback from the proving environment. Good performance has been shown over the miniF2F dataset with 3 IMO problems being solved.", "strengths": "The paper is well written with a clear illustration of its two contributions TC and CC. I especially appreciate the ablation study where Lyra downgrades to DSP without TC and CC. The performance gain also looks good.", "weaknesses": "Although I very much like the idea of CC, the innovation of TC appears slightly limited. At least to me, it does not involve much interaction with LLMs -- it is more like an exhaustive attempt on a set of heuristically chosen proof methods other than Sledgehammer.", "questions": "- page 1, 'However, they have not been able to post-process LLM generation or gradually refine previous generations.': The Baldur paper (https://arxiv.org/abs/2303.04910) has explored post-processing LLM-generated proofs. I would love to see a comparison between Baldur and Lyra if possible.\n- This paper is mostly based on the previous DSP paper, where LLMs are used to produce proof skeletons (i.e., unproved conjectures). In the tool correction part, it appears that LLMs are prompted to produce a full mechanised proof including the tactic 'by (simp add: div mult mod eq)', which is considered as LLM hallucination by the authors. Is that the case? If so, it might be a good idea to make the distinction clear as this may affect the ablation study. \n- Figure 3, the informal proof and the formal sketch are actually quite different. For example, the formal one does not cover continuity nor limit, which have been mentioned several times in the informal proofs. I was wondering if the authors could elaborate a bit on the discrepancy between the informal proof and the formal one.\n\nminor\n- page 3, 'conducted on LLLMs' -> 'conducted on LLMs'\n- page 5, 'As all formal proof begins with proof -': strictly speaking this is not quite true, as some Isabelle proofs start with 'proof (...)' or 'apply (...)', where ... can be some Isabelle tactics.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an LLM-augmented automated theorem proving framework called Lyra in a formal theorem proving environment. The distinguishing features of Lyra include Tool Correction (TC) and Conjecture Correction (CC), which respectively post-edit formal proofs emitted from LLMs given feedback from the proving environment. Good performance has been shown over the miniF2F dataset with 3 IMO problems being solved.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper is well written with a clear illustration of its two contributions TC and CC. I especially appreciate the ablation study where Lyra downgrades to DSP without TC and CC. The performance gain also looks good.", "weaknesses": "Although I very much like the idea of CC, the innovation of TC appears slightly limited. At least to me, it does not involve much interaction with LLMs -- it is more like an exhaustive attempt on a set of heuristically chosen proof methods other than Sledgehammer.", "questions": "- page 1, 'However, they have not been able to post-process LLM generation or gradually refine previous generations.': The Baldur paper (https://arxiv.org/abs/2303.04910) has explored post-processing LLM-generated proofs. I would love to see a comparison between Baldur and Lyra if possible.\n- This paper is mostly based on the previous DSP paper, where LLMs are used to produce proof skeletons (i.e., unproved conjectures). In the tool correction part, it appears that LLMs are prompted to produce a full mechanised proof including the tactic 'by (simp add: div mult mod eq)', which is considered as LLM hallucination by the authors. Is that the case? If so, it might be a good idea to make the distinction clear as this may affect the ablation study. \n- Figure 3, the informal proof and the formal sketch are actually quite different. For example, the formal one does not cover continuity nor limit, which have been mentioned several times in the informal proofs. I was wondering if the authors could elaborate a bit on the discrepancy between the informal proof and the formal one.\n\nminor\n- page 3, 'conducted on LLLMs' -> 'conducted on LLMs'\n- page 5, 'As all formal proof begins with proof -': strictly speaking this is not quite true, as some Isabelle proofs start with 'proof (...)' or 'apply (...)', where ... can be some Isabelle tactics.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698791868112}, {"id": "YEOpjgUdxZ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1096/Reviewer_NmsF"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a couple of approaches, namely Tool Correction (TC) and Conjecture Correction (CC), for improving the performance of LLM-guided autoformalization techniques based on the Draft-Sketch-Proof (DSP) paradigm. TC is a post-processing technique---once the LLM generates a formal proof sketch, the proof is fed to an Interactive Theorem Prover (ITP) and if the ITP returns an error, then TC uses simple heuristics to replace the tactics in the incorrect formal proof sketch one-by-one. CC, on the other hand, incorporates the error message from the ITP and generates a new prompt that includes the previous proof attempt and the error message. This interaction between the LLM and ITP is conductive up to 5 times. The proposed extensions to DSP are evaluated on the miniF2F benchmark and lead to improved proof success rates.", "review_text": "This paper presents a couple of approaches, namely Tool Correction (TC) and Conjecture Correction (CC), for improving the performance of LLM-guided autoformalization techniques based on the Draft-Sketch-Proof (DSP) paradigm. TC is a post-processing technique---once the LLM generates a formal proof sketch, the proof is fed to an Interactive Theorem Prover (ITP) and if the ITP returns an error, then TC uses simple heuristics to replace the tactics in the incorrect formal proof sketch one-by-one. CC, on the other hand, incorporates the error message from the ITP and generates a new prompt that includes the previous proof attempt and the error message. This interaction between the LLM and ITP is conductive up to 5 times. The proposed extensions to DSP are evaluated on the miniF2F benchmark and lead to improved proof success rates.", "strengths": "1. The techniques lead to an improvement over the state-of-the-art.\n2. I find it a little surprising that models like GPT-4 and Codex are able to understand and act based on the error messages from Isabelle. Unfortunately, the paper does not explore this surprising observation in depth.\n\n-----------------------------\nRating updated to 5 after discussion. Overall, I remain unconvinced by the technical contribution of TC but I find the empirical phenomenon of LLMs being able to interpret ITP error messages interesting\n\n-----------------------------\nRating further updated after much discussion about the general applicability of TC to 6. In particular, the authors presented evidence that TC helps improve performance on another dataset, namely, LISA.", "weaknesses": "1. I am not an expert in the area so it is possible that the empirical results might be surprising to experts in a way that I am unable to appreciate. However, I find that the techniques used to achieve the state-of-the-art results do not involve any significant technical or empirical insights.\n\n2. The heuristics used in TC seem too specific to the task and dataset. Are the heuristics used for TC transferable to other theorem proving tools and datasets? How do we know that these heuristics are not overfit to the miniF2f dataset? Do these heuristics represent some general insight into how mathematical theorems ought to be proved?\n\n3. CC simply incorporates the ITP error message into the prompt. On a technical level, this is an obvious idea that has been tried before for proof generation [1] and code generation [2,3]. Some of these approaches require fine-tuning the model, so the scientific question to be evaluated here is does incorporating error feedback without any fine-tuning help **in general** for autoformalization. Since models such as Codex and GPT-4 are not explicitly trained on the error messages from ITPs, one would not expect simply providing the error messages in the prompts to be generally helpful. While the empirical results here suggest that error messages help when the ITP is Isabelle, it remains to be evaluated if it is helpful with other ITPs. It would also be useful to analyze the nature of the error messages generated by Isabelle. Do the improvements depend on the quality of the error messages? One would expect so but this would be another useful aspect to empirically evaluate.\n\n[1] First, E., Rabe, M. N., Ringer, T., & Brun, Y. (2023). Baldur: whole-proof generation and repair with large language models. arXiv preprint arXiv:2303.04910.\n\n[2] Le, H., Wang, Y., Gotmare, A. D., Savarese, S., & Hoi, S. C. H. (2022). Coderl: Mastering code generation through pretrained models and deep reinforcement learning. Advances in Neural Information Processing Systems, 35, 21314-21328.\n\n[3] Wu, X., Cheriere, N., Zhang, C., & Narayanan, D. (2023). RustGen: An Augmentation Approach for Generating Compilable Rust Code with Large Language Models.", "questions": "1. Are the presented techniques overfit to miniF2F dataset and Isabelle? In particular, I am afraid this might be the case for TC.\n\n2. Would CC work with the error messages from a different ITP? How much does the availability of formal proofs and error messages for a particular ITP affect the effectiveness of CC?  How much does the quality of the error message affect CC?\n\n3. There are number of spelling errors in Algorithm 2.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a couple of approaches, namely Tool Correction (TC) and Conjecture Correction (CC), for improving the performance of LLM-guided autoformalization techniques based on the Draft-Sketch-Proof (DSP) paradigm. TC is a post-processing technique---once the LLM generates a formal proof sketch, the proof is fed to an Interactive Theorem Prover (ITP) and if the ITP returns an error, then TC uses simple heuristics to replace the tactics in the incorrect formal proof sketch one-by-one. CC, on the other hand, incorporates the error message from the ITP and generates a new prompt that includes the previous proof attempt and the error message. This interaction between the LLM and ITP is conductive up to 5 times. The proposed extensions to DSP are evaluated on the miniF2F benchmark and lead to improved proof success rates.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The techniques lead to an improvement over the state-of-the-art.\n2. I find it a little surprising that models like GPT-4 and Codex are able to understand and act based on the error messages from Isabelle. Unfortunately, the paper does not explore this surprising observation in depth.\n\n-----------------------------\nRating updated to 5 after discussion. Overall, I remain unconvinced by the technical contribution of TC but I find the empirical phenomenon of LLMs being able to interpret ITP error messages interesting\n\n-----------------------------\nRating further updated after much discussion about the general applicability of TC to 6. In particular, the authors presented evidence that TC helps improve performance on another dataset, namely, LISA.", "weaknesses": "1. I am not an expert in the area so it is possible that the empirical results might be surprising to experts in a way that I am unable to appreciate. However, I find that the techniques used to achieve the state-of-the-art results do not involve any significant technical or empirical insights.\n\n2. The heuristics used in TC seem too specific to the task and dataset. Are the heuristics used for TC transferable to other theorem proving tools and datasets? How do we know that these heuristics are not overfit to the miniF2f dataset? Do these heuristics represent some general insight into how mathematical theorems ought to be proved?\n\n3. CC simply incorporates the ITP error message into the prompt. On a technical level, this is an obvious idea that has been tried before for proof generation [1] and code generation [2,3]. Some of these approaches require fine-tuning the model, so the scientific question to be evaluated here is does incorporating error feedback without any fine-tuning help **in general** for autoformalization. Since models such as Codex and GPT-4 are not explicitly trained on the error messages from ITPs, one would not expect simply providing the error messages in the prompts to be generally helpful. While the empirical results here suggest that error messages help when the ITP is Isabelle, it remains to be evaluated if it is helpful with other ITPs. It would also be useful to analyze the nature of the error messages generated by Isabelle. Do the improvements depend on the quality of the error messages? One would expect so but this would be another useful aspect to empirically evaluate.\n\n[1] First, E., Rabe, M. N., Ringer, T., & Brun, Y. (2023). Baldur: whole-proof generation and repair with large language models. arXiv preprint arXiv:2303.04910.\n\n[2] Le, H., Wang, Y., Gotmare, A. D., Savarese, S., & Hoi, S. C. H. (2022). Coderl: Mastering code generation through pretrained models and deep reinforcement learning. Advances in Neural Information Processing Systems, 35, 21314-21328.\n\n[3] Wu, X., Cheriere, N., Zhang, C., & Narayanan, D. (2023). RustGen: An Augmentation Approach for Generating Compilable Rust Code with Large Language Models.", "questions": "1. Are the presented techniques overfit to miniF2F dataset and Isabelle? In particular, I am afraid this might be the case for TC.\n\n2. Would CC work with the error messages from a different ITP? How much does the availability of formal proofs and error messages for a particular ITP affect the effectiveness of CC?  How much does the quality of the error message affect CC?\n\n3. There are number of spelling errors in Algorithm 2.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698772919304}, {"id": "zIiCqDRTLG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1096/Reviewer_bwwN"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper presents Lyra, an automated system integrating large language models for autoformalisation and symbolic tools such as sledgehammer in proof assistants. The two core components, Tool Correction and Conjecture Correction, are critical for its achieving SoTA performance on the miniF2F benchmark.", "review_text": "This paper presents Lyra, an automated system integrating large language models for autoformalisation and symbolic tools such as sledgehammer in proof assistants. The two core components, Tool Correction and Conjecture Correction, are critical for its achieving SoTA performance on the miniF2F benchmark.", "strengths": "The proposed mechanism makes good intuitive sense. The performance improvement is impressive on the miniF2F benchmark. The ablation studies are well-presented and convincing.", "weaknesses": "There needs to be more contextualisation of the prior works. For the two major correction mechanisms, there have been direct prior works doing very similar or even identical things.\n- Tool correction: in the DSP work [1], the authors already used sledgehammer + heuristics to close conjectures made. The understanding is that the Lyra method first tries a LLM-generated tactic to close conjectures, and if it doesn't work, try sledgehammer. This is largely similar and should be noted.\n- In the Baldur work [2] from April 2023, the authors have proposed to use the proof assistant error message to repair the proofs. This is very similar to the conjecture correction with error messages and should be noted.\n\n[1] Albert Qiaochu Jiang, Sean Welleck, Jin Peng Zhou, Timothee Lacroix, Jiacheng Liu, Wenda Li, Mateja Jamnik, Guillaume Lample, and Yuhuai Wu. Draft, sketch, and prove: Guiding formal theorem provers with informal proofs. In The Eleventh International Conference on Learning Representations, 2023.\n\n[2] First, Emily, Markus N. Rabe, Talia Ringer, and Yuriy Brun. \"Baldur: whole-proof generation and repair with large language models.\" arXiv preprint arXiv:2303.04910 (2023).", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Lyra, an automated system integrating large language models for autoformalisation and symbolic tools such as sledgehammer in proof assistants. The two core components, Tool Correction and Conjecture Correction, are critical for its achieving SoTA performance on the miniF2F benchmark.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The proposed mechanism makes good intuitive sense. The performance improvement is impressive on the miniF2F benchmark. The ablation studies are well-presented and convincing.", "weaknesses": "There needs to be more contextualisation of the prior works. For the two major correction mechanisms, there have been direct prior works doing very similar or even identical things.\n- Tool correction: in the DSP work [1], the authors already used sledgehammer + heuristics to close conjectures made. The understanding is that the Lyra method first tries a LLM-generated tactic to close conjectures, and if it doesn't work, try sledgehammer. This is largely similar and should be noted.\n- In the Baldur work [2] from April 2023, the authors have proposed to use the proof assistant error message to repair the proofs. This is very similar to the conjecture correction with error messages and should be noted.\n\n[1] Albert Qiaochu Jiang, Sean Welleck, Jin Peng Zhou, Timothee Lacroix, Jiacheng Liu, Wenda Li, Mateja Jamnik, Guillaume Lample, and Yuhuai Wu. Draft, sketch, and prove: Guiding formal theorem provers with informal proofs. In The Eleventh International Conference on Learning Representations, 2023.\n\n[2] First, Emily, Markus N. Rabe, Talia Ringer, and Yuriy Brun. \"Baldur: whole-proof generation and repair with large language models.\" arXiv preprint arXiv:2303.04910 (2023).", "questions": "None.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1697905258318}], "openreview_url": "https://openreview.net/forum?id=9Z0yB8rmQ2", "arxiv_id": "2309.15806", "paper_pdf": "papers/9Z0yB8rmQ2.pdf", "paper_pdf_sha256": "4c9c2f59f6b73ce10222f6ea5dbf2fadd45a060352f84a5435485328a01da51f", "paper_pdf_bytes": 1024887, "paper_pdf_source": "openreview", "code_url": "https://github.com/chuanyang-Zheng/Lyra-theorem-prover", "code_repository": "chuanyang-Zheng/Lyra-theorem-prover", "code_commit": "4424ad734f1fb1ef9aaa13979cd273a54247acb7", "code_archive": "repos/9Z0yB8rmQ2.zip", "code_archive_sha256": "4463504a264ab7abfde1125b9259107b0b83ec3e150d030ba9214f3e5526aaa4", "code_archive_bytes": 286377, "code_file_count": 24, "code_extensions": {".py": 22, ".sh": 2}, "github_disk_usage_kb": 256, "github_languages": {"Python": 215424, "Isabelle": 154664, "Shell": 4382}, "github_archived": false, "github_pushed_at": "2024-07-02T06:22:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lyra-orchestrating-dual-correction-in"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zU2v47WF0Ku", "year": 2022, "status": "rejected", "title": "Implicit Bias of Linear Equivariant Networks", "authors": ["Hannah Lawrence", "Kristian Georgiev", "Andrew Dienes", "Bobak Kiani"], "authorids": ["~Hannah_Lawrence1", "~Kristian_Georgiev1", "adienes@mit.edu", "~Bobak_Kiani1"], "authors_source": "OpenReview API", "abstract": "Group equivariant convolutional neural networks (G-CNNs) are generalizations of convolutional neural networks (CNNs) which excel in a wide range of scientific and technical applications by explicitly encoding particular group symmetries, such as rotations and permutations, in their architectures. Although the success of G-CNNs is driven by the explicit symmetry bias of their convolutional architecture, a recent line of work has proposed that the implicit bias of training algorithms on a particular parameterization (or architecture) is key to understanding generalization for overparameterized neural nets. In this context, we show that $L$-layer full-width linear G-CNNs trained via gradient descent in a binary classification task converge to solutions with low-rank Fourier matrix coefficients, regularized by the $2/L$-Schatten matrix norm. Our work strictly generalizes previous analysis on the implicit bias of linear CNNs to linear G-CNNs over all finite groups, including the challenging setting of non-commutative symmetry groups (such as permutations). We validate our theorems via experiments on a variety of groups and empirically explore more realistic nonlinear networks, which locally capture similar regularization patterns. Finally, we provide intuitive interpretations of our Fourier-space implicit regularization results in real space via uncertainty principles.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tsWTCG-eUEJ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1553/Reviewer_Mfmn"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Because of the explicit inductive bias of G-CNN, other inductive biases have not been discussed much.\nIn this paper, the author focuses on and analyzes the non-explicit inductive bias of G-CNNs.\nTechnically, they present the non-explicit bias in terms of Fourier matrices by using a group Fourier transform, which depends on the group structure.\nThe results show that learning a linear G-CNN (with linearized β) by gradient descent implicitly biases the singular values of the Fourier matrix coefficients of β to be sparse.\nThis result has been experimentally confirmed in linear and some nonlinear situations.", "review_text": "Strength: Focusing on the non-explicit bias of G-CNN, the results of the analysis are satisfactory.\n\nWeaknesses: I do not see anything special about it.\n\nQuestions:\nFrom a theoretical point of view, we would like this kind of result to include the case where the group does not do group convolution as a case where the group is trivial. What does Theorem 1 imply when the group is trivial?\\\n\nIt would be great if this kind of result could provide some insight in analyzing over parametrized FNN. When we regress the G-equivariant function on an over parametrized FNN, can your results give any insight?\\\n\nThis result was for finite groups, but there are also convolutions for geometric groups, such as Lie conv. Please tell us about the part of this result where the finite group assumption works and give us some insight into the generalization to Lie groups.\n\nAfter the revision, the paper was improved, for example by generalization to the Lie group, but in the main the assessment did not change significantly, resulting in the following scores\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Because of the explicit inductive bias of G-CNN, other inductive biases have not been discussed much.\nIn this paper, the author focuses on and analyzes the non-explicit inductive bias of G-CNNs.\nTechnically, they present the non-explicit bias in terms of Fourier matrices by using a group Fourier transform, which depends on the group structure.\nThe results show that learning a linear G-CNN (with linearized β) by gradient descent implicitly biases the singular values of the Fourier matrix coefficients of β to be sparse.\nThis result has been experimentally confirmed in linear and some nonlinear situations.", "main_review": "Strength: Focusing on the non-explicit bias of G-CNN, the results of the analysis are satisfactory.\n\nWeaknesses: I do not see anything special about it.\n\nQuestions:\nFrom a theoretical point of view, we would like this kind of result to include the case where the group does not do group convolution as a case where the group is trivial. What does Theorem 1 imply when the group is trivial?\\\n\nIt would be great if this kind of result could provide some insight in analyzing over parametrized FNN. When we regress the G-equivariant function on an over parametrized FNN, can your results give any insight?\\\n\nThis result was for finite groups, but there are also convolutions for geometric groups, such as Lie conv. Please tell us about the part of this result where the finite group assumption works and give us some insight into the generalization to Lie groups.\n\nAfter the revision, the paper was improved, for example by generalization to the Lie group, but in the main the assessment did not change significantly, resulting in the following scores\n\n", "summary_of_the_review": "Basically, this paper is theoretically well done. The assumption of linearity is a strong one that is far from reality, but it is a reasonable assumption for the current theoretical analysis, and it is confirmed by experiments in the nonlinear case. Overall, I think this is a worthwhile paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "Not applicable", "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636080522965}, {"id": "Xevn5xmUQ7f", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1553/Reviewer_k7aV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies the induced bias of gradient descent on L-layer linear group-equivariant convolutional neural networks. Their main result is: gradient descent implicitly controls the 2/L-Schatten norm of the Fourier transform of the (linear) predictor. More precisely, they show that gradient descent trained on a L-layer group-equivariant CNN with exponential loss on a linearly separable dataset converges in direction to a first-order stationary point of an optimization problem minimizing the 2/L Schatten norm of the Fourier transform of the linear predictor subject to the margin being at least 1. \n", "review_text": "From a technical perspective, the results for abelian groups (as the authors note) follow from appropriately applying the result of Yun et al. (2020). (This is because convolutions in Fourier space can be expressed as pointwise multiplication for abelian groups.) The more interesting technical contribution is for non-abelian groups. In this case, convolutions corresponds to matrix multiplications of matrices which need-not be diagonal, and thus do not correspond to point wise multiplication. For this case, the authors directly analyze the stationary points of the optimization problem. \n\nAn appealing aspect of the paper is to highlight that the norm that is implicitly controlled by gradient descent on group equivariant networks takes a simple form, even for general (finite) groups. This technical point is made clearly in the paper, and the paper is generally well-written.  \n\nOne weakness of the paper is that the it does not explicitly connect the theoretical results with practical settings. First, since group-equivariant convolutional neural networks are not a standard architecture, it could be helpful if the authors added a discussion of what group structures have been used / are meaningful in practice, and provided insight (e.g. in the empirical section) into the implicit regularization of gradient descent in these special cases. \n\nFurthermore, the empirical results are quite limited. Based on Appendix D, the networks appear to be run on very small synthetic datasets (with 6-10 datapoints). For this reason, it is difficult to determine if the takeaways hold for more realistic classification tasks. It would be helpful if the authors ran experiments on standard binary classification datasets (e.g. MNIST or CIFAR). \n\nLastly, another weakness of this paper is that the results are restricted to single-channel linear networks with full-dimensional kernels. In light of previous work (i.e. Yun et al. (2020, Gunasekar et al. (2018)), the results are somewhat incremental. The paper mainly demonstrates that the 2/L-norm in Gunasekar et al. (2018) for L-layer linear CNNs can be replaced by the (2/L)-Schatten norm. While this norm (especially for non-abelian groups) does give rise to different sparsity structures, the analysis and the conceptual takeaways are somewhat similar those in Gunasekar et al. (2018). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the induced bias of gradient descent on L-layer linear group-equivariant convolutional neural networks. Their main result is: gradient descent implicitly controls the 2/L-Schatten norm of the Fourier transform of the (linear) predictor. More precisely, they show that gradient descent trained on a L-layer group-equivariant CNN with exponential loss on a linearly separable dataset converges in direction to a first-order stationary point of an optimization problem minimizing the 2/L Schatten norm of the Fourier transform of the linear predictor subject to the margin being at least 1. \n", "main_review": "From a technical perspective, the results for abelian groups (as the authors note) follow from appropriately applying the result of Yun et al. (2020). (This is because convolutions in Fourier space can be expressed as pointwise multiplication for abelian groups.) The more interesting technical contribution is for non-abelian groups. In this case, convolutions corresponds to matrix multiplications of matrices which need-not be diagonal, and thus do not correspond to point wise multiplication. For this case, the authors directly analyze the stationary points of the optimization problem. \n\nAn appealing aspect of the paper is to highlight that the norm that is implicitly controlled by gradient descent on group equivariant networks takes a simple form, even for general (finite) groups. This technical point is made clearly in the paper, and the paper is generally well-written.  \n\nOne weakness of the paper is that the it does not explicitly connect the theoretical results with practical settings. First, since group-equivariant convolutional neural networks are not a standard architecture, it could be helpful if the authors added a discussion of what group structures have been used / are meaningful in practice, and provided insight (e.g. in the empirical section) into the implicit regularization of gradient descent in these special cases. \n\nFurthermore, the empirical results are quite limited. Based on Appendix D, the networks appear to be run on very small synthetic datasets (with 6-10 datapoints). For this reason, it is difficult to determine if the takeaways hold for more realistic classification tasks. It would be helpful if the authors ran experiments on standard binary classification datasets (e.g. MNIST or CIFAR). \n\nLastly, another weakness of this paper is that the results are restricted to single-channel linear networks with full-dimensional kernels. In light of previous work (i.e. Yun et al. (2020, Gunasekar et al. (2018)), the results are somewhat incremental. The paper mainly demonstrates that the 2/L-norm in Gunasekar et al. (2018) for L-layer linear CNNs can be replaced by the (2/L)-Schatten norm. While this norm (especially for non-abelian groups) does give rise to different sparsity structures, the analysis and the conceptual takeaways are somewhat similar those in Gunasekar et al. (2018). \n", "summary_of_the_review": "I recommend weak rejection due to the incremental nature of the technical results in light of previous work, as well as the limitations of the empirical section. \n\n---- \nUpdate after author response: I appreciate the additional experiments on the MNIST dataset, and believe that this improves the empirical analysis provided in this paper. However, given the results in previous work (e.g. Gunasekar et al. '18, etc) for the trivial group, I still think that the results in this work are fairly restrictive since they only hold for single-channel linear networks with full-dimensional filters. Although previous work has shown that simple closed-form solutions may not exist in general, it would still be useful to provide insight into how these important parameters affect the implicit bias of gradient descent. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636011058364}, {"id": "7gKDTygP2n-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1553/Reviewer_nEDA"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper describes a theoretical analysis of the implicit bias in G-CNNs. The results show that for linear G-CNNs trained on linearly separable data using GD converge to sparse solutions in the Fourier domain (equivalently dense solutions in the real domain). The theory is theoretically confirmed and shows that the implicit bias also occurs to some extend for non-linear G-CNNs.", "review_text": "Strengths:\nAs far as I can assess, the paper is technically sound, novel and timely.\nThe paper is formal and precise.\nThe appendix is quite nice actually and is rather complete in mathematical background.\nThe paper is reproducible and has code provided.\n\nWeaknesses:\nMy main criticism is that the paper is very theoretical and feels a bit like a nice math exercise to prove something, however, without being clear why the problem is addressed in the first place (apart from the fact that no one did it yet). To me, as someone working with G-CNNs and having a good understanding of (irreducible) representation theory, but not in analysis, it is unclear what I should take home from the paper. \nI.e., why should we expect different results for non-Abelian compared to Abelian groups in the first place? \nHow should I interpret the bias towards dense solutions as a regularization effect? \nWhat are the implications?\n\nComments:\nIt would be helpful to have explicit statements about the main contributions/novelty of the paper.\n\nThe experiments in the figure make sense to me in the comparison G-CNN vs FC. Both map G-feature maps to scalars and the latter shows that the bias takes place due to architecture and not the data. However, I do not understand the comparison CNN to G-CNN. If the feature maps are functions on G, is a CNN not automatically a G-CNN? Then, what is the difference between the two? \n\nSome essential details are missing regarding the data. I cannot find it in the appendix either. The closest I get to a description is in caption of Figure 3: “trained … with six isotropic Gaussian data points”. What is an isotropic Gaussian data point? \n\nIn theorem 5.3 third line: is this a typo? (w.r.t. indexing to l should be l+1?, it says something like Fw_l-Fw_l \\geq \\lambda)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper describes a theoretical analysis of the implicit bias in G-CNNs. The results show that for linear G-CNNs trained on linearly separable data using GD converge to sparse solutions in the Fourier domain (equivalently dense solutions in the real domain). The theory is theoretically confirmed and shows that the implicit bias also occurs to some extend for non-linear G-CNNs.", "main_review": "Strengths:\nAs far as I can assess, the paper is technically sound, novel and timely.\nThe paper is formal and precise.\nThe appendix is quite nice actually and is rather complete in mathematical background.\nThe paper is reproducible and has code provided.\n\nWeaknesses:\nMy main criticism is that the paper is very theoretical and feels a bit like a nice math exercise to prove something, however, without being clear why the problem is addressed in the first place (apart from the fact that no one did it yet). To me, as someone working with G-CNNs and having a good understanding of (irreducible) representation theory, but not in analysis, it is unclear what I should take home from the paper. \nI.e., why should we expect different results for non-Abelian compared to Abelian groups in the first place? \nHow should I interpret the bias towards dense solutions as a regularization effect? \nWhat are the implications?\n\nComments:\nIt would be helpful to have explicit statements about the main contributions/novelty of the paper.\n\nThe experiments in the figure make sense to me in the comparison G-CNN vs FC. Both map G-feature maps to scalars and the latter shows that the bias takes place due to architecture and not the data. However, I do not understand the comparison CNN to G-CNN. If the feature maps are functions on G, is a CNN not automatically a G-CNN? Then, what is the difference between the two? \n\nSome essential details are missing regarding the data. I cannot find it in the appendix either. The closest I get to a description is in caption of Figure 3: “trained … with six isotropic Gaussian data points”. What is an isotropic Gaussian data point? \n\nIn theorem 5.3 third line: is this a typo? (w.r.t. indexing to l should be l+1?, it says something like Fw_l-Fw_l \\geq \\lambda)\n", "summary_of_the_review": "Although I am comfortable with group theory and basic harmonic analysis on groups through irreps, I find it hard to follow the derivations and make sense out of the statements. My impression is that the paper is too mathematical for ICLR audience, though I can really only speak for myself. The paper could improve a lot from more layman's explanations throughout the paper. For example, it already starts early in the paper with talk about Schatten norms, where to me it is not immediately obvious why these are considered.\n\nI understand that it is hard to please every reader, but I believe it is possible to convey the main messages better to a broader audience. Currently it seems overly tuned towards experts in analysis (of implicit biases).\n\nOverall, I do see value in publishing such work as the paper seems otherwise sound.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635891826682}, {"id": "etJIqhCK7Sw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1553/Reviewer_HWqt"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Group Equivariant CNNs (G-CNNs) generalize CNNs from the translation group to arbitrary groups. It has been shown that overparameterized/undetermined linear transformations and linear CNNs learned by gradient descent are implicitly biased, in that they find zero train loss solutions that minimize some norm on the predictor. This paper generalizes these results from linear CNNs to linear G-CNNs. The linear G-CNNs are constructed into a non-equivariant linear transformation, by doing L G-CNN layers, followed by a non-equivariant inner product with a G-feature. The authors show that the implicit bias is to minimize the p-norm of the singular values (the Schatten norm) of the Fourier transform of the resulting linear transformation.\nIn the case where G is Abelian, by a tensor factorization of the linear G-CNN, the authors can straightforwardly apply a result from prior work on the implicit bias of such tensor factorizations to prove their result.\nIn the case where G is not Abelian, the authors prove their results in three steps. (1) They use prior work on gradient descent on linear predictors that are polynomials of the parameters, which their linear G-CNN falls under, to show that the stationary points minimize norm the joint norm on all weights. (2) They define a linear regression optimization that minimizes the Schatten norm of the linear predictor. (3) They show that all optima of the polynomial linear predictor are also optima of the Schatten-normed linear regression. To do so, they find the subdifferential of the p-Schatten norm and show that the polynomial optimum is in the subdifferential.\n\nThe authors confirm their theory in toy experiments in the linear case and show that in the non-linear case, where the Schatten norm is computed by local linearization, that similar results may hold in the non-linear case.", "review_text": "The authors show for the models under consideration an interesting and novel result: the models are implicitly biased not by the real space norm of the linear predictor, but by the Schatten norm of the Fourier transform of the predictor. For the Abelian case, this required an incremental generalization of prior work, but for the non-Abelian case required an interesting proof method. The paper is very clearly written and was a pleasure to read.\n\nMy main problem with this work is that it considers a model class that I’ve never seen used (besides the lack of non-linearities): it uses group equivariant CNNs and then takes an inner product with a G-feature weight, resulting in a non-equivariant linear predictor. If the authors used invariant outputs, the linear invariant predictors would just be averaging, and the analysis wouldn’t work. As G-CNNs are chosen for their symmetry properties, studying them in a non-equivariant context makes little sense to me. Whether the implicit bias found in this paper generalizes to equivariant non-linear G-CNN networks has not been convincingly shown.\n\nFor me to recommend acceptance of this paper, the authors should make a more convincing point why the results for this non-equivariant model class are applicable to equivariant G-CNN networks people actually use (besides the fact that the analysis applies to linear networks).\n\nOther weaknesses:\n* The non-Abelian case states two assumptions (gradients converge in direction, iterates converge in direction to a classifier with positive margin), but these are not clearly defined, nor is it stated when these are satisfied. The paper would be improved if there would be some discussion of these assumptions.\n* The key point of the paper seems to be that Fourier Schatten norm, not spatial norm is the implicit bias of the learned networks. The paper would be stronger if it would give some more intuition about this point. Perhaps visualize some of the learned weights in both Fourier and spatial domain?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Group Equivariant CNNs (G-CNNs) generalize CNNs from the translation group to arbitrary groups. It has been shown that overparameterized/undetermined linear transformations and linear CNNs learned by gradient descent are implicitly biased, in that they find zero train loss solutions that minimize some norm on the predictor. This paper generalizes these results from linear CNNs to linear G-CNNs. The linear G-CNNs are constructed into a non-equivariant linear transformation, by doing L G-CNN layers, followed by a non-equivariant inner product with a G-feature. The authors show that the implicit bias is to minimize the p-norm of the singular values (the Schatten norm) of the Fourier transform of the resulting linear transformation.\nIn the case where G is Abelian, by a tensor factorization of the linear G-CNN, the authors can straightforwardly apply a result from prior work on the implicit bias of such tensor factorizations to prove their result.\nIn the case where G is not Abelian, the authors prove their results in three steps. (1) They use prior work on gradient descent on linear predictors that are polynomials of the parameters, which their linear G-CNN falls under, to show that the stationary points minimize norm the joint norm on all weights. (2) They define a linear regression optimization that minimizes the Schatten norm of the linear predictor. (3) They show that all optima of the polynomial linear predictor are also optima of the Schatten-normed linear regression. To do so, they find the subdifferential of the p-Schatten norm and show that the polynomial optimum is in the subdifferential.\n\nThe authors confirm their theory in toy experiments in the linear case and show that in the non-linear case, where the Schatten norm is computed by local linearization, that similar results may hold in the non-linear case.", "main_review": "The authors show for the models under consideration an interesting and novel result: the models are implicitly biased not by the real space norm of the linear predictor, but by the Schatten norm of the Fourier transform of the predictor. For the Abelian case, this required an incremental generalization of prior work, but for the non-Abelian case required an interesting proof method. The paper is very clearly written and was a pleasure to read.\n\nMy main problem with this work is that it considers a model class that I’ve never seen used (besides the lack of non-linearities): it uses group equivariant CNNs and then takes an inner product with a G-feature weight, resulting in a non-equivariant linear predictor. If the authors used invariant outputs, the linear invariant predictors would just be averaging, and the analysis wouldn’t work. As G-CNNs are chosen for their symmetry properties, studying them in a non-equivariant context makes little sense to me. Whether the implicit bias found in this paper generalizes to equivariant non-linear G-CNN networks has not been convincingly shown.\n\nFor me to recommend acceptance of this paper, the authors should make a more convincing point why the results for this non-equivariant model class are applicable to equivariant G-CNN networks people actually use (besides the fact that the analysis applies to linear networks).\n\nOther weaknesses:\n* The non-Abelian case states two assumptions (gradients converge in direction, iterates converge in direction to a classifier with positive margin), but these are not clearly defined, nor is it stated when these are satisfied. The paper would be improved if there would be some discussion of these assumptions.\n* The key point of the paper seems to be that Fourier Schatten norm, not spatial norm is the implicit bias of the learned networks. The paper would be stronger if it would give some more intuition about this point. Perhaps visualize some of the learned weights in both Fourier and spatial domain?", "summary_of_the_review": "While the authors use novel methods to derive an interesting implicit bias of the model class under consideration, this class, non-equivariant predictor based on equivariant neural network layers, is not what is generally used in practice. For the paper to be recommendable for acceptance, the authors should make a stronger argument why their results also apply to models actually used. If they do so convincingly, I'd increase my score.\n\n\n\nUpdated score to 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635682520116}], "openreview_url": "https://openreview.net/forum?id=zU2v47WF0Ku", "arxiv_id": "2110.06084", "paper_pdf": "papers/zU2v47WF0Ku.pdf", "paper_pdf_sha256": "01991cbb43389206a334071fd0632d3079a17d68f5e21417c045c2c7720bfdac", "paper_pdf_bytes": 4811337, "paper_pdf_source": "openreview", "code_url": "https://github.com/kristian-georgiev/implicit-bias-of-linear-equivariant-networks", "code_repository": "kristian-georgiev/implicit-bias-of-linear-equivariant-networks", "code_commit": "ac5a5536ea98f06b3b33870b73e700a09b736777", "code_archive": "repos/zU2v47WF0Ku.zip", "code_archive_sha256": "81c6720db7cbbee21ca7e46d4ac747780980a55481fdeaf21cadfd7e5b7e5ee2", "code_archive_bytes": 2393383, "code_file_count": 15, "code_extensions": {".ipynb": 10, ".py": 5}, "github_disk_usage_kb": 2153, "github_languages": {"Jupyter Notebook": 1370211, "Python": 39216, "Sage": 3859, "GAP": 1969}, "github_archived": false, "github_pushed_at": "2022-06-17T05:08:58Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/implicit-bias-of-linear-equivariant-networks-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cbtV7xGO9pS", "year": 2021, "status": "rejected", "title": "TEAC: Intergrating Trust Region and Max Entropy Actor Critic for Continuous Control", "authors": ["Hongyu Zang", "Xin Li", "Li Zhang", "Peiyao Zhao", "Mingzhong Wang"], "authorids": ["~Hongyu_Zang1", "~Xin_Li31", "~Li_Zhang18", "~Peiyao_Zhao1", "~Mingzhong_Wang1"], "authors_source": "OpenReview API", "abstract": "Trust region methods and maximum entropy methods are two state-of-the-art branches used in reinforcement learning (RL) for the benefits of stability and exploration in continuous environments, respectively. This paper proposes to integrate both branches in a unified framework, thus benefiting from both sides. We first transform the original RL objective to a constraint optimization problem and then proposes trust entropy actor-critic (TEAC), an off-policy algorithm to learn stable and sufficiently explored policies for continuous states and actions. TEAC trains the critic by minimizing the refined Bellman error and updates the actor by minimizing KL-divergence loss derived from the closed-form solution to the Lagrangian. \nWe prove that the policy evaluation and policy improvement in TEAC is guaranteed to converge.\nWe compare TEAC with 4 state-of-the-art solutions on 6 tasks in the MuJoCo environment. The results show that TEAC outperforms state-of-the-art solutions in terms of efficiency and effectiveness.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "HHNr4zcdQvc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper770/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes Trust Entropy Actor Critic (TEAC), a novel algorithm for reinforcement learning (RL) combining the idea of TRPO/PPO and max-entropy RL, together with the corresponding critic, actor and dual updates. The high level idea is that trust region methods ensure stability by constraining the KL divergence from the previous policy, while entropy regularization encourages exploration, and hence combining the two may achieve the best of both worlds and obtain a good trade-off between stability and exploration. To achieve this goal, the authors propose to augment the original trust-region subproblem in TRPO with an additional constraint on the lower bound of the policy entropy (together with two other trivial constraints corresponding to the validity of the policy in the MDP framework). Then by forming the Lagrangian function and setting the gradient to zero, the authors obtain both the critic (value) and actor (policy) updates with different choices of dual variables (corresponding to the two trivial constraints), together with the dual updates. Numerical experiments compared to some popular baseline RL algorithms are also reported to demonstrate the improvement of TEAC compared to the existing works. \n\nin general, the high level idea is interesting and reasonable, and some empirical improvement is also shown. However, there are several undefined terms and technical issues/mistakes that largely downgrade the quality and contribution of this paper. \n1. $Q^{\\pi}$ in (1) and (2) is not defined, and the meaning of $E_{\\rho_{\\pi}(s),\\pi(a|s)}$ is unclear. Also, according to standard literature like [1, Chapter 4], the expected reward $\\mathcal{J}(\\pi)$ should indeed be something like $\\frac{1}{1-\\gamma}E_{s\\sim \\rho_{\\pi}, a\\sim \\pi(\\cdot|s)}r(s,a)$, so the term inside the expectation should probably be the reward $r(s,a)$ instead of the Q function $Q^{\\pi}(s,a)$.\n2. The formulation in (2) is confusing. Firstly, it is unclear what the notation $E_{\\rho_{\\pi}(s)\\pi(a|s)}=1$ means, and what the difference is between $E_{\\rho_{\\pi}(s)\\pi(a|s)}$ and $E_{\\rho_{\\pi}(s),\\pi(a|s)}$. Similarly, it is unclear what $V$ is (as a function of $\\pi$ or a free variable), and what the notation $E_{\\rho_{\\pi}(s)\\pi(a|s)p(s’|s,a)}$ means. In general, all the terms appearing in the optimization problem should either be some constants or a function of $\\pi$, but this is not made clear by the authors. Also, it’s unclear why the authors switch the orders of $\\pi$ and $\\pi_{\\rm old}$ in the KL divergence compared to the original TRPO formulation, and some explanations should be provided. \n3. The formula (4) seems to be weird. Firstly, according to the right-hand side, it seems that the left-hand side should indeed be the (s,a)-entry of the gradient instead of the entire gradient. Secondly, it is unclear why $\\rho(s)$ disappears in the second equality. \n4. The authors argue that the dual variables $\\nu$ and $\\lambda$ can be taken arbitrarily as the policy is parametrized as a Gaussian via neural networks so that it is always a valid randomized policy (satisfying the third constraint in (2)), while the “value function” always satisfies the Bellman equation corresponding to the fourth constraint in (2). However, with the neural network parametrization $\\theta$, the optimization problem should also be solved w.r.t. $\\theta$ instead of $\\pi$, in which case (4) no longer makes sense (since it’s differentiating w.r.t. $\\pi$ instead of $\\theta$). Also, as mentioned above, the value function $V$ is undefined and hence the second claim does not make much sense. In addition, if the fourth constraint in (2) corresponds to the Bellman equation, then it is weird why no reward $r$ or discount factor $\\gamma$ is involved there. \n5. In Lemma 1, the “trust entropy Q-value” is undefined. \n6. In (10), why can $\\exp(-(\\alpha+\\beta+\\lambda)/(\\alpha+\\beta))$ serve as a normalization term? In particular, why does it hold that $\\int_a\\pi_{\\rm old}(a|s)^{\\alpha/(\\alpha+\\beta)}\\exp(Q^{\\pi}(s,a)/(\\alpha+\\beta))da$ always equal to $\\exp(-(\\alpha+\\beta+\\lambda)/(\\alpha+\\beta))$ for any state $s$?\n\nAlso, the experimental part has some limitations.\n1. The numerical improvement is not very convincing, as it seems that TEAC only improves over SAC in two out of six tasks and only has significant improvement on one of these two tasks. Some more thorough comparisons are needed to validate the empirical advantage of the proposed method. \n2. The actually implemented algorithm, Algorithm 1 contains several tricks unexplained in the main text. In particular, why do we need two target critic networks? And in the update (38), should $\\bar{\\phi}$ be $\\bar{\\phi}_i$? \n\nFinally, please find some miscellaneous suggestions/comments on additional references, technical issues fixing and typo fixing below.\n1. The work [2] seems to be closely related to the high level idea of this paper, and should better be compared with. \n2. In the abstract, “transforms” should be “transform”. In the first paragraph of the introduction, “learning process” should be “learning processes”. \n3. In the last paragraph of the introduction, “guaranteed to converge” is not very accurate. In fact, only the critic/policy evaluation is guaranteed to converge, and the authors show that policy improvement does hold (but may or may not lead to eventual convergence of the whole TEAC algorithm). \n4. In the first paragraph of Section 2, the definition of $\\rho_{\\pi}$ is not provided, although from the literature it probably means the discounted state-visitation distribution/measure. The authors should provide a clear definition for self-contained-ness, and the term “state of the trajectory distribution” does not seem to make much sense and should better be replaced with more standard terminologies like “discounted state-visitation distribution/measure”. Also, “qantified” should be “quantified”. \n5. In TEAC, the parameters $\\tau$ and $\\eta$ are always fixed. However, as the algorithm proceeds, it may not make much sense to keep a constant exploration power as required by the constant $\\eta>0$. Will a decreasing $\\eta$ be a better choice? Some discussions on this should better be provided. \n6. In (3), $\\rho(s)$ should be $\\rho_{\\pi}(s)$, and the dual variables $\\alpha$ and $\\beta$ should be required to be non-negative. Accordingly, the dual updates should probably better be projected onto the non-negative orthant. \n7. In the sentence before (4), “derivation” should be “derivative”. \n8. In (5), $Q$ should be $Q^{\\pi}$. \n9. In (20), the entropy terms lack right parentheses. \n\nHence in general, I think the paper is still not ready for publication and needs substantial fixing and improvement. \n\n[1] Agarwal, Alekh, Nan Jiang, and Sham M. Kakade. Reinforcement learning: Theory and algorithms. Technical Report, Department of Computer Science, University of Washington, 2019.\n\n[2] Pajarinen, Joni, Hong Linh Thai, Riad Akrour, Jan Peters, and Gerhard Neumann. \"Compatible natural gradient policy search.\" Machine Learning 108, no. 8-9 (2019): 1443-1466.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting idea with several technical issues", "review": "This paper proposes Trust Entropy Actor Critic (TEAC), a novel algorithm for reinforcement learning (RL) combining the idea of TRPO/PPO and max-entropy RL, together with the corresponding critic, actor and dual updates. The high level idea is that trust region methods ensure stability by constraining the KL divergence from the previous policy, while entropy regularization encourages exploration, and hence combining the two may achieve the best of both worlds and obtain a good trade-off between stability and exploration. To achieve this goal, the authors propose to augment the original trust-region subproblem in TRPO with an additional constraint on the lower bound of the policy entropy (together with two other trivial constraints corresponding to the validity of the policy in the MDP framework). Then by forming the Lagrangian function and setting the gradient to zero, the authors obtain both the critic (value) and actor (policy) updates with different choices of dual variables (corresponding to the two trivial constraints), together with the dual updates. Numerical experiments compared to some popular baseline RL algorithms are also reported to demonstrate the improvement of TEAC compared to the existing works. \n\nin general, the high level idea is interesting and reasonable, and some empirical improvement is also shown. However, there are several undefined terms and technical issues/mistakes that largely downgrade the quality and contribution of this paper. \n1. $Q^{\\pi}$ in (1) and (2) is not defined, and the meaning of $E_{\\rho_{\\pi}(s),\\pi(a|s)}$ is unclear. Also, according to standard literature like [1, Chapter 4], the expected reward $\\mathcal{J}(\\pi)$ should indeed be something like $\\frac{1}{1-\\gamma}E_{s\\sim \\rho_{\\pi}, a\\sim \\pi(\\cdot|s)}r(s,a)$, so the term inside the expectation should probably be the reward $r(s,a)$ instead of the Q function $Q^{\\pi}(s,a)$.\n2. The formulation in (2) is confusing. Firstly, it is unclear what the notation $E_{\\rho_{\\pi}(s)\\pi(a|s)}=1$ means, and what the difference is between $E_{\\rho_{\\pi}(s)\\pi(a|s)}$ and $E_{\\rho_{\\pi}(s),\\pi(a|s)}$. Similarly, it is unclear what $V$ is (as a function of $\\pi$ or a free variable), and what the notation $E_{\\rho_{\\pi}(s)\\pi(a|s)p(s’|s,a)}$ means. In general, all the terms appearing in the optimization problem should either be some constants or a function of $\\pi$, but this is not made clear by the authors. Also, it’s unclear why the authors switch the orders of $\\pi$ and $\\pi_{\\rm old}$ in the KL divergence compared to the original TRPO formulation, and some explanations should be provided. \n3. The formula (4) seems to be weird. Firstly, according to the right-hand side, it seems that the left-hand side should indeed be the (s,a)-entry of the gradient instead of the entire gradient. Secondly, it is unclear why $\\rho(s)$ disappears in the second equality. \n4. The authors argue that the dual variables $\\nu$ and $\\lambda$ can be taken arbitrarily as the policy is parametrized as a Gaussian via neural networks so that it is always a valid randomized policy (satisfying the third constraint in (2)), while the “value function” always satisfies the Bellman equation corresponding to the fourth constraint in (2). However, with the neural network parametrization $\\theta$, the optimization problem should also be solved w.r.t. $\\theta$ instead of $\\pi$, in which case (4) no longer makes sense (since it’s differentiating w.r.t. $\\pi$ instead of $\\theta$). Also, as mentioned above, the value function $V$ is undefined and hence the second claim does not make much sense. In addition, if the fourth constraint in (2) corresponds to the Bellman equation, then it is weird why no reward $r$ or discount factor $\\gamma$ is involved there. \n5. In Lemma 1, the “trust entropy Q-value” is undefined. \n6. In (10), why can $\\exp(-(\\alpha+\\beta+\\lambda)/(\\alpha+\\beta))$ serve as a normalization term? In particular, why does it hold that $\\int_a\\pi_{\\rm old}(a|s)^{\\alpha/(\\alpha+\\beta)}\\exp(Q^{\\pi}(s,a)/(\\alpha+\\beta))da$ always equal to $\\exp(-(\\alpha+\\beta+\\lambda)/(\\alpha+\\beta))$ for any state $s$?\n\nAlso, the experimental part has some limitations.\n1. The numerical improvement is not very convincing, as it seems that TEAC only improves over SAC in two out of six tasks and only has significant improvement on one of these two tasks. Some more thorough comparisons are needed to validate the empirical advantage of the proposed method. \n2. The actually implemented algorithm, Algorithm 1 contains several tricks unexplained in the main text. In particular, why do we need two target critic networks? And in the update (38), should $\\bar{\\phi}$ be $\\bar{\\phi}_i$? \n\nFinally, please find some miscellaneous suggestions/comments on additional references, technical issues fixing and typo fixing below.\n1. The work [2] seems to be closely related to the high level idea of this paper, and should better be compared with. \n2. In the abstract, “transforms” should be “transform”. In the first paragraph of the introduction, “learning process” should be “learning processes”. \n3. In the last paragraph of the introduction, “guaranteed to converge” is not very accurate. In fact, only the critic/policy evaluation is guaranteed to converge, and the authors show that policy improvement does hold (but may or may not lead to eventual convergence of the whole TEAC algorithm). \n4. In the first paragraph of Section 2, the definition of $\\rho_{\\pi}$ is not provided, although from the literature it probably means the discounted state-visitation distribution/measure. The authors should provide a clear definition for self-contained-ness, and the term “state of the trajectory distribution” does not seem to make much sense and should better be replaced with more standard terminologies like “discounted state-visitation distribution/measure”. Also, “qantified” should be “quantified”. \n5. In TEAC, the parameters $\\tau$ and $\\eta$ are always fixed. However, as the algorithm proceeds, it may not make much sense to keep a constant exploration power as required by the constant $\\eta>0$. Will a decreasing $\\eta$ be a better choice? Some discussions on this should better be provided. \n6. In (3), $\\rho(s)$ should be $\\rho_{\\pi}(s)$, and the dual variables $\\alpha$ and $\\beta$ should be required to be non-negative. Accordingly, the dual updates should probably better be projected onto the non-negative orthant. \n7. In the sentence before (4), “derivation” should be “derivative”. \n8. In (5), $Q$ should be $Q^{\\pi}$. \n9. In (20), the entropy terms lack right parentheses. \n\nHence in general, I think the paper is still not ready for publication and needs substantial fixing and improvement. \n\n[1] Agarwal, Alekh, Nan Jiang, and Sham M. Kakade. Reinforcement learning: Theory and algorithms. Technical Report, Department of Computer Science, University of Washington, 2019.\n\n[2] Pajarinen, Joni, Hong Linh Thai, Riad Akrour, Jan Peters, and Gerhard Neumann. \"Compatible natural gradient policy search.\" Machine Learning 108, no. 8-9 (2019): 1443-1466.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1605057589284}, {"id": "nVfPqhpmdx_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper770/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Previous work such as MOTO and GAC has developed approaches that combine trust regions for policy stability with maximum entropy to ensure adequate exploration.  This paper introduces an algorithm called TEAC in this spirit which adds several technical novelties to address limitations of previous work:\n - It uses a particular Gaussian policy parameterization for the actor to ensure that updated policies are easily representable in the same class while prior approaches used more complex solutions.\n - It uses a modified objective for the critic that incorporates the trust region and entropy terms\n - It uses a first (rather than second) order method to optimize the dual variables\n - It uses these to provide a proof of policy improvement.\nTEAC is evaluated on a number of MuJoCo tasks.\n\nI think the TEAC algorithm and theoretical analysis are a nice contribution and that ultimately this will be a nice paper.  However, I have some concerns with the positioning of the contribution, the rigor of the technical exposition, and the experiments.\n\n1) The introduction makes it seem like the contribution of the paper is proposing the idea of combining trust regions and maximum entropy, but as is discussed later this idea is already present in previous work (e.g MOTO and GAC).  The real contribution is discussed only in very compressed form in the last paragraph.  Having a fuller discussion in Section 4 after the technical exposition is reasonable, but the key ideas and problems the paper solves to make this approach work better than previous work should at least by clearly stated and some intuition or motivation provided in the introduction.\n\n2) In Equation (2), something about the notation for the =1 constraint doesn’t seem quite right.  The expectation isn’t being taken over anything.  This issue persists into 3.  The notation for the last MDP constraint seems a bit odd as well, although I supposes the intent is to have s’ drawn from the resulting distribution.\n\n3) Equation 5 still has \\lambda in it despite setting it to 0.\n\n4) Lemma 2: the old policy \\hat{\\pi} is being given, not defined right?  At least I don’t see a definition of it.\n\n5) The proof of Lemma 2 should point to A.3 not A.2.  I’m not sure why A.2 exists.  It doesn’t seem to be otherwise referenced and seems to be a repetition of material in the main text.\n\n6) The reason (27) is true should be explained rather than point generally to another paper without an specific explanation.  I also don’t understand why (28) is an application of (6) and there is an ellipsis here.  I assume what is going on is an argument about what happens as the Bellman operator is repeatedly applied as it approaches its limit, which we know from (10) is the final quantity.\n\n7) Given the close relationship, it seems odd that no comparison to MOTO is made.  GAC does not appear to provide one either, so even an indirect comparison does not appear to exist.  Furthermore GAC mysteriously disappears after Figure 1(b), which is problematic as it appears to be the most closely related work in terms of technique.\n\n8) The plots only show three trials of each algorithm, which limits the confidence of performance assessments based on them and is not what I would describe as “extensive” in the abstract and elsewhere.  While I agree based on the results shown that TEAC typically outperforms the non-SAC algorithms, the comparison with SAC is more mixed except for the strong performance in swimmer (although even here the strong performance of PPO seems similar to TEAC).  So I would tone down the unqualified claim that “TEAC outperforms the state-of-the-art solutions”.\n\nUpdates after author responses and discussion:\n(1) After the updates, most of the positioning issues have been improved\n(2) The technical exposition is improved, and the main argument I was bothered by is somewhat improved, though it still has a big jump near the end.\n(3) I'm still bothered that to demonstrate improvement the algorithm is tuned on a per-example basis but the baselines are not.  The results certainly show that the former is reasonable, but to make the comparison fair the latter needs to be done as well.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting algorithm and and analysis, but aspects of the paper need improvement", "review": "Previous work such as MOTO and GAC has developed approaches that combine trust regions for policy stability with maximum entropy to ensure adequate exploration.  This paper introduces an algorithm called TEAC in this spirit which adds several technical novelties to address limitations of previous work:\n - It uses a particular Gaussian policy parameterization for the actor to ensure that updated policies are easily representable in the same class while prior approaches used more complex solutions.\n - It uses a modified objective for the critic that incorporates the trust region and entropy terms\n - It uses a first (rather than second) order method to optimize the dual variables\n - It uses these to provide a proof of policy improvement.\nTEAC is evaluated on a number of MuJoCo tasks.\n\nI think the TEAC algorithm and theoretical analysis are a nice contribution and that ultimately this will be a nice paper.  However, I have some concerns with the positioning of the contribution, the rigor of the technical exposition, and the experiments.\n\n1) The introduction makes it seem like the contribution of the paper is proposing the idea of combining trust regions and maximum entropy, but as is discussed later this idea is already present in previous work (e.g MOTO and GAC).  The real contribution is discussed only in very compressed form in the last paragraph.  Having a fuller discussion in Section 4 after the technical exposition is reasonable, but the key ideas and problems the paper solves to make this approach work better than previous work should at least by clearly stated and some intuition or motivation provided in the introduction.\n\n2) In Equation (2), something about the notation for the =1 constraint doesn’t seem quite right.  The expectation isn’t being taken over anything.  This issue persists into 3.  The notation for the last MDP constraint seems a bit odd as well, although I supposes the intent is to have s’ drawn from the resulting distribution.\n\n3) Equation 5 still has \\lambda in it despite setting it to 0.\n\n4) Lemma 2: the old policy \\hat{\\pi} is being given, not defined right?  At least I don’t see a definition of it.\n\n5) The proof of Lemma 2 should point to A.3 not A.2.  I’m not sure why A.2 exists.  It doesn’t seem to be otherwise referenced and seems to be a repetition of material in the main text.\n\n6) The reason (27) is true should be explained rather than point generally to another paper without an specific explanation.  I also don’t understand why (28) is an application of (6) and there is an ellipsis here.  I assume what is going on is an argument about what happens as the Bellman operator is repeatedly applied as it approaches its limit, which we know from (10) is the final quantity.\n\n7) Given the close relationship, it seems odd that no comparison to MOTO is made.  GAC does not appear to provide one either, so even an indirect comparison does not appear to exist.  Furthermore GAC mysteriously disappears after Figure 1(b), which is problematic as it appears to be the most closely related work in terms of technique.\n\n8) The plots only show three trials of each algorithm, which limits the confidence of performance assessments based on them and is not what I would describe as “extensive” in the abstract and elsewhere.  While I agree based on the results shown that TEAC typically outperforms the non-SAC algorithms, the comparison with SAC is more mixed except for the strong performance in swimmer (although even here the strong performance of PPO seems similar to TEAC).  So I would tone down the unqualified claim that “TEAC outperforms the state-of-the-art solutions”.\n\nUpdates after author responses and discussion:\n(1) After the updates, most of the positioning issues have been improved\n(2) The technical exposition is improved, and the main argument I was bothered by is somewhat improved, though it still has a big jump near the end.\n(3) I'm still bothered that to demonstrate improvement the algorithm is tuned on a per-example basis but the baselines are not.  The results certainly show that the former is reasonable, but to make the comparison fair the latter needs to be done as well.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604159535948}, {"id": "xN-q_LKU8mh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper770/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose to add three more constraints to the well-known trust region policy optimization model. The first one of these constraints aims at keeping the entropy level higher than a given threshold. The authors then relax these constraints to obtain the Lagrangian function. An approach using primal-dual updates is proposed to obtain a solution.\n\nIn general, I find the paper difficult to follow and hence cannot properly assess its novelty. My comments and questions are as follows:\n\n- The function Q^\\pi in (1) is not defined at that stage.\n\n- How can we guarantee the feasibility of model (2) for different values of \\eta? Likewise, if we obtain a solution with tackling the Lagrangian function, then is that solution feasible for (2)?\n\n- How do you go from the first line to the second line in relation (4)?\n\n- Why \\lambda is in (5)?\n\n- What is Q-\\phi in (8)?\n\n- What does setting \\eta to dimension in Table 1 signify? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Hard to follow work proposing a trust region policy optimization model with additional constraints", "review": "The authors propose to add three more constraints to the well-known trust region policy optimization model. The first one of these constraints aims at keeping the entropy level higher than a given threshold. The authors then relax these constraints to obtain the Lagrangian function. An approach using primal-dual updates is proposed to obtain a solution.\n\nIn general, I find the paper difficult to follow and hence cannot properly assess its novelty. My comments and questions are as follows:\n\n- The function Q^\\pi in (1) is not defined at that stage.\n\n- How can we guarantee the feasibility of model (2) for different values of \\eta? Likewise, if we obtain a solution with tackling the Lagrangian function, then is that solution feasible for (2)?\n\n- How do you go from the first line to the second line in relation (4)?\n\n- Why \\lambda is in (5)?\n\n- What is Q-\\phi in (8)?\n\n- What does setting \\eta to dimension in Table 1 signify? ", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603896970746}, {"id": "zdjzn4DTdu", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper770/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "1: The reviewer's evaluation is an educated guess", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper addresses the problem of reinforcement learning in continuous spaces by formulating the problem as a constrained optimization problem. In this problem, the objective is maximizing the expected reward, and the constraints ensure that 1) the distance between the new and old policies is bounded, 2) the entropy is above a threshold, and 3) the assumptions of MDP hold. The paper then derives closed-form solutions for the Lagrangian, which are used for obtaining variable update rules. An empirical study examines the ideas against benchmark problems. \n\nThe paper has a good structure and style. The claims seem sound and the results look promising. However, since I'm not familiar with the topic, I cannot verify the details of those claims and results. I base my recommendation on a high-level idea of the paper. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Possibly a good paper", "review": "The paper addresses the problem of reinforcement learning in continuous spaces by formulating the problem as a constrained optimization problem. In this problem, the objective is maximizing the expected reward, and the constraints ensure that 1) the distance between the new and old policies is bounded, 2) the entropy is above a threshold, and 3) the assumptions of MDP hold. The paper then derives closed-form solutions for the Lagrangian, which are used for obtaining variable update rules. An empirical study examines the ideas against benchmark problems. \n\nThe paper has a good structure and style. The claims seem sound and the results look promising. However, since I'm not familiar with the topic, I cannot verify the details of those claims and results. I base my recommendation on a high-level idea of the paper. ", "rating": "7: Good paper, accept", "confidence": "1: The reviewer's evaluation is an educated guess"}, "tcdate": 1603877644023}], "openreview_url": "https://openreview.net/forum?id=cbtV7xGO9pS", "arxiv_id": null, "paper_pdf": "papers/cbtV7xGO9pS.pdf", "paper_pdf_sha256": "7b08de7608f3ab5794d9a32f5bf2aab60c35465d5b0a2c719d78bed6f1f5fe48", "paper_pdf_bytes": 3088734, "paper_pdf_source": "openreview", "code_url": "https://github.com/ICLR2021papersub/TEAC", "code_repository": "ICLR2021papersub/TEAC", "code_commit": "bad3488749963daf8cc71bc5dcb870e71b97abdd", "code_archive": "repos/cbtV7xGO9pS.zip", "code_archive_sha256": "dd32d7a66f767c10d0e1b463eda84c0ad1fab4931403ef59ceb66b4468fe5650", "code_archive_bytes": 724386, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 1368, "github_languages": {"Python": 212088}, "github_archived": false, "github_pushed_at": "2020-10-30T06:11:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/teac-intergrating-trust-region-and-max"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wyCnT4BUsT", "year": 2026, "status": "rejected", "title": "DeepCritic: Deliberate Critique with Large Language Models", "authors": ["Wenkai Yang", "Jingwen Chen", "Yankai Lin", "Ji-Rong Wen"], "authorids": ["~Wenkai_Yang1", "~Jingwen_Chen4", "~Yankai_Lin1", "~Ji-Rong_Wen1"], "authors_source": "OpenReview API", "abstract": "As Large Language Models (LLMs) are rapidly evolving, providing accurate feedback and scalable oversight on their outputs becomes an urgent and critical problem. Leveraging LLMs as critique models to achieve automated supervision is a promising solution. In this work, we focus on studying and enhancing the math critique ability of LLMs. Current LLM critics provide critiques that are too shallow and superficial on each step, leading to low judgment accuracy and struggling to offer sufficient feedback for the LLM generator to correct mistakes. To tackle this issue, we propose a novel and effective two-stage framework to develop LLM critics that are capable of deliberately critiquing on each reasoning step of math solutions. In the first stage, we carefully curate 4.5K long-form critiques as seed data for supervised fine-tuning. Each seed critique consists of deliberate step-wise critiques that includes multi-perspective verifications as well as in-depth critiques of initial critiques for each reasoning step. Then, we perform reinforcement learning on the fine-tuned model with either existing human-labeled data from PRM800K or our automatically annotated data obtained via Monte Carlo sampling-based correctness estimation, to further incentivize its critique ability. Our developed critique model built on Qwen2.5-7B-Instruct not only significantly outperforms existing LLM critics (including the same-sized DeepSeek-R1-Distill models and GPT-4o) on various error identification benchmarks, but also more effectively helps the LLM generator refine erroneous steps through more detailed feedback.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "W7MkPbRYU5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11560/Reviewer_Err4"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces DeepCritic, a two-stage framework for enhancing the critique capabilities of large language models (LLMs) in mathematical reasoning tasks. Existing LLM critics often produce shallow critiques, leading to low judgment accuracy and insufficient feedback for error correction. The authors curate 4.5K long-form critiques incorporating multi-perspective verification and meta-critiquing for supervised fine-tuning (SFT), followed by reinforcement learning (RL) using either human-annotated data (e.g., PRM800K) or automatically generated data via Monte Carlo sampling. Experiments on benchmarks like MR-GSM8K, PRM800K, and ProcessBench demonstrate that the resulting 7B-parameter model outperforms baselines including GPT-4o and DeepSeek-R1-Distill, while enabling better refinement of LLM generators through detailed feedback. Contributions include a novel data curation pipeline for deliberate critiquing and evidence of weak-to-strong supervision potential.", "review_text": "This paper introduces DeepCritic, a two-stage framework for enhancing the critique capabilities of large language models (LLMs) in mathematical reasoning tasks. Existing LLM critics often produce shallow critiques, leading to low judgment accuracy and insufficient feedback for error correction. The authors curate 4.5K long-form critiques incorporating multi-perspective verification and meta-critiquing for supervised fine-tuning (SFT), followed by reinforcement learning (RL) using either human-annotated data (e.g., PRM800K) or automatically generated data via Monte Carlo sampling. Experiments on benchmarks like MR-GSM8K, PRM800K, and ProcessBench demonstrate that the resulting 7B-parameter model outperforms baselines including GPT-4o and DeepSeek-R1-Distill, while enabling better refinement of LLM generators through detailed feedback. Contributions include a novel data curation pipeline for deliberate critiquing and evidence of weak-to-strong supervision potential.", "strengths": "The paper innovatively addresses the superficiality of LLM critiques by introducing a two-stage pipeline that combines iterative critique generation (initial + in-depth meta-critiquing) with RL, creatively adapting Monte Carlo sampling for automated RL data in math domains, which extends prior work on scalable oversight (e.g., Saunders et al., 2022) to deliberate reasoning without relying solely on human labels.\n\nThe methodology is rigorously evaluated across multiple benchmarks, showing substantial improvements (e.g., 20-point F1 gains post-SFT), with ablation studies validating key components like in-depth critiquing; the use of diverse RL data sources (human vs. auto-generated) provides robust evidence of the framework's effectiveness and generalizability.\n\nThe writing is clear, with detailed prompts, data statistics, and case studies illustrating deliberate critiquing; significantly, it advances LLM self-evolution by demonstrating how enhanced critics can supervise stronger generators (e.g., 7B critic refining 72B outputs), offering practical insights for superalignment and automated feedback in complex reasoning tasks.", "weaknesses": "While RL improves performance, the auto-generated data via Monte Carlo sampling discards certain solutions (e.g., fully correct/incorrect ones), potentially introducing biases toward medium-difficulty problems; this could be quantified with diversity metrics to ensure the data represents a wide range of math complexities.\n\nTest-time scaling results focus on majority voting and refinement, but lack comparisons with advanced baselines like outcome reward models (ORMs) or hybrid PRM-ORM setups, which might reveal limitations in handling very long reasoning chains or non-math domains.", "questions": "Given the emphasis on meta-critiquing in SFT data, could you elaborate on how often the model exhibits self-correction during inference (e.g., via quantitative analysis of generated critiques), and whether this transfers to out-of-distribution math problems like those in higher-level Olympiads? A response with additional metrics could strengthen claims of deliberate reasoning robustness.\n\nThe RL stage uses GRPO with a binary accuracy reward; how sensitive is performance to alternative reward designs, such as incorporating critique informativeness (e.g., via BLEU-like scores on feedback quality)? Experiments or ablation on this could address potential over-optimization toward judgment accuracy at the expense of feedback depth.\n\nIn weak-to-strong supervision experiments, the 7B critic refines 72B generators effectively, but what happens when the generator is even stronger (e.g., GPT-4o level) or in adversarial settings where solutions have subtle logical flaws? Providing case studies or extended results here could clarify the framework's limits in superalignment scenarios.\n\nTo better position your work in the literature, could you include a comparison with related critique enhancement methods such as CTRL, AlignRAG, Critique fine-tuning, one-shot CFT, and Critique-Guided Distillation? Specifically, discuss how your iterative meta-critiquing and RL pipeline differs in terms of data efficiency, critique depth, and applicability to math reasoning, potentially through qualitative or quantitative contrasts to highlight unique advantages or limitations?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces DeepCritic, a two-stage framework for enhancing the critique capabilities of large language models (LLMs) in mathematical reasoning tasks. Existing LLM critics often produce shallow critiques, leading to low judgment accuracy and insufficient feedback for error correction. The authors curate 4.5K long-form critiques incorporating multi-perspective verification and meta-critiquing for supervised fine-tuning (SFT), followed by reinforcement learning (RL) using either human-annotated data (e.g., PRM800K) or automatically generated data via Monte Carlo sampling. Experiments on benchmarks like MR-GSM8K, PRM800K, and ProcessBench demonstrate that the resulting 7B-parameter model outperforms baselines including GPT-4o and DeepSeek-R1-Distill, while enabling better refinement of LLM generators through detailed feedback. Contributions include a novel data curation pipeline for deliberate critiquing and evidence of weak-to-strong supervision potential.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper innovatively addresses the superficiality of LLM critiques by introducing a two-stage pipeline that combines iterative critique generation (initial + in-depth meta-critiquing) with RL, creatively adapting Monte Carlo sampling for automated RL data in math domains, which extends prior work on scalable oversight (e.g., Saunders et al., 2022) to deliberate reasoning without relying solely on human labels.\n\nThe methodology is rigorously evaluated across multiple benchmarks, showing substantial improvements (e.g., 20-point F1 gains post-SFT), with ablation studies validating key components like in-depth critiquing; the use of diverse RL data sources (human vs. auto-generated) provides robust evidence of the framework's effectiveness and generalizability.\n\nThe writing is clear, with detailed prompts, data statistics, and case studies illustrating deliberate critiquing; significantly, it advances LLM self-evolution by demonstrating how enhanced critics can supervise stronger generators (e.g., 7B critic refining 72B outputs), offering practical insights for superalignment and automated feedback in complex reasoning tasks.", "weaknesses": "While RL improves performance, the auto-generated data via Monte Carlo sampling discards certain solutions (e.g., fully correct/incorrect ones), potentially introducing biases toward medium-difficulty problems; this could be quantified with diversity metrics to ensure the data represents a wide range of math complexities.\n\nTest-time scaling results focus on majority voting and refinement, but lack comparisons with advanced baselines like outcome reward models (ORMs) or hybrid PRM-ORM setups, which might reveal limitations in handling very long reasoning chains or non-math domains.", "questions": "Given the emphasis on meta-critiquing in SFT data, could you elaborate on how often the model exhibits self-correction during inference (e.g., via quantitative analysis of generated critiques), and whether this transfers to out-of-distribution math problems like those in higher-level Olympiads? A response with additional metrics could strengthen claims of deliberate reasoning robustness.\n\nThe RL stage uses GRPO with a binary accuracy reward; how sensitive is performance to alternative reward designs, such as incorporating critique informativeness (e.g., via BLEU-like scores on feedback quality)? Experiments or ablation on this could address potential over-optimization toward judgment accuracy at the expense of feedback depth.\n\nIn weak-to-strong supervision experiments, the 7B critic refines 72B generators effectively, but what happens when the generator is even stronger (e.g., GPT-4o level) or in adversarial settings where solutions have subtle logical flaws? Providing case studies or extended results here could clarify the framework's limits in superalignment scenarios.\n\nTo better position your work in the literature, could you include a comparison with related critique enhancement methods such as CTRL, AlignRAG, Critique fine-tuning, one-shot CFT, and Critique-Guided Distillation? Specifically, discuss how your iterative meta-critiquing and RL pipeline differs in terms of data efficiency, critique depth, and applicability to math reasoning, potentially through qualitative or quantitative contrasts to highlight unique advantages or limitations?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761953047171}, {"id": "XNTnXKvQIF", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11560/Reviewer_PU76"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces DeepCritic, a framework to enhance the critique capabilities of LLMs, specifically for mathematical reasoning. The authors address the superficial nature of current LLM critiques by proposing following the same stages (SFT + RL) as other reasoning task. \nThe resulting 7B-parameter critic model outperforms larger models (like GPT-4o) and specialized math models on error identification benchmarks. It also scales well at test-time and effectively guides generators to refine their answers.", "review_text": "This paper introduces DeepCritic, a framework to enhance the critique capabilities of LLMs, specifically for mathematical reasoning. The authors address the superficial nature of current LLM critiques by proposing following the same stages (SFT + RL) as other reasoning task. \nThe resulting 7B-parameter critic model outperforms larger models (like GPT-4o) and specialized math models on error identification benchmarks. It also scales well at test-time and effectively guides generators to refine their answers.", "strengths": "1. Throughout experiments: the paper contains very throughout experiments to prove the idea. Although this is a well explored scope (SFT + RL), the full set of experiment is still nice.\n\n2. The paper demonstrates that RL with automatically constructed data (DeepCritic-7B-RL-Numina) also yields substantial gains, this confirms with finding from other papers, and proving that auto rating is valuable.", "weaknesses": "1. Lack of novelty: the framework adopted in this paper is well established in reasoning world, this work can be viewed as an application in the critique capability.\n\n2. Limited Domain: The paper focuses solely on mathematical reasoning. While a standard testbed, it's unclear if this deliberate critique approach generalizes well to more subjective or less structured domains (e.g., creative writing, complex instruction following).\n\n3. Dependency on Strong Teacher: The seed data generation relies heavily on a very capable model (Qwen2.5-72B-Instruct). The approach might be less viable if a significantly stronger teacher model isn't available for a given domain. As we know that distill is very effective for reasoning, the result is kind depending on this.", "questions": "1. When you use the LLM as critic, what prompt was used? Have you tried to improve the prompt to improve the performance?\n2. There are a lot mistakes in PRM800K (80% accuracy based on OpenAI), how does that affect the final result?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces DeepCritic, a framework to enhance the critique capabilities of LLMs, specifically for mathematical reasoning. The authors address the superficial nature of current LLM critiques by proposing following the same stages (SFT + RL) as other reasoning task. \nThe resulting 7B-parameter critic model outperforms larger models (like GPT-4o) and specialized math models on error identification benchmarks. It also scales well at test-time and effectively guides generators to refine their answers.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Throughout experiments: the paper contains very throughout experiments to prove the idea. Although this is a well explored scope (SFT + RL), the full set of experiment is still nice.\n\n2. The paper demonstrates that RL with automatically constructed data (DeepCritic-7B-RL-Numina) also yields substantial gains, this confirms with finding from other papers, and proving that auto rating is valuable.", "weaknesses": "1. Lack of novelty: the framework adopted in this paper is well established in reasoning world, this work can be viewed as an application in the critique capability.\n\n2. Limited Domain: The paper focuses solely on mathematical reasoning. While a standard testbed, it's unclear if this deliberate critique approach generalizes well to more subjective or less structured domains (e.g., creative writing, complex instruction following).\n\n3. Dependency on Strong Teacher: The seed data generation relies heavily on a very capable model (Qwen2.5-72B-Instruct). The approach might be less viable if a significantly stronger teacher model isn't available for a given domain. As we know that distill is very effective for reasoning, the result is kind depending on this.", "questions": "1. When you use the LLM as critic, what prompt was used? Have you tried to improve the prompt to improve the performance?\n2. There are a lot mistakes in PRM800K (80% accuracy based on OpenAI), how does that affect the final result?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761888908503}, {"id": "IgWiEKbLQf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission11560/Reviewer_y2i7"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper addresses a critical and timely issue: the superficial nature of critiques generated by Large Language Models (LLMs) when serving as evaluators, particularly in complex reasoning domains such as mathematics. The authors argue that existing LLM critics often produce shallow feedback that merely echoes the original solution's reasoning, leading to low error detection accuracy and insufficient guidance for correction.\n\nTo overcome this limitation, the paper introduces DeepCritic, a novel two-stage training framework designed to teach LLMs to perform \"deliberate critique.\"\n\nStage 1 (Supervised Fine-Tuning): The core innovation lies in its sophisticated data curation pipeline. First, a powerful teacher model (Qwen2.5-72B-Instruct) generates an \"initial critique\" for each step of a mathematical solution. Crucially, the teacher model then performs a deeper \"meta-critique\" on the initial critique itself. This involves multi-perspective verification and self-reflection to identify shortcomings in the initial assessment. Finally, these two critiques are synthesized into a single, high-quality, long-form \"deliberate critique.\" A curated dataset of 4.5K such examples is used to fine-tune the base model.\n\nStage 2 (Reinforcement Learning): Building upon the SFT model, RL is employed to further enhance its critique capabilities. The reward signals are sourced from two pathways: 1) existing human-annotated datasets (e.g., PRM800K), and 2) automatically generated step-level labels via Monte Carlo sampling, enabling scalable oversight.\n\nExperimental results demonstrate that the resulting DeepCritic-7B model significantly outperforms much larger and more capable baselines, including GPT-4o, on several mathematical error detection benchmarks. The paper further showcases the model's practical utility as both a verifier to improve solution accuracy and as a supervisor to guide other LLMs in refining their errors, even demonstrating a potential for weak-to-strong supervision.", "review_text": "This paper addresses a critical and timely issue: the superficial nature of critiques generated by Large Language Models (LLMs) when serving as evaluators, particularly in complex reasoning domains such as mathematics. The authors argue that existing LLM critics often produce shallow feedback that merely echoes the original solution's reasoning, leading to low error detection accuracy and insufficient guidance for correction.\n\nTo overcome this limitation, the paper introduces DeepCritic, a novel two-stage training framework designed to teach LLMs to perform \"deliberate critique.\"\n\nStage 1 (Supervised Fine-Tuning): The core innovation lies in its sophisticated data curation pipeline. First, a powerful teacher model (Qwen2.5-72B-Instruct) generates an \"initial critique\" for each step of a mathematical solution. Crucially, the teacher model then performs a deeper \"meta-critique\" on the initial critique itself. This involves multi-perspective verification and self-reflection to identify shortcomings in the initial assessment. Finally, these two critiques are synthesized into a single, high-quality, long-form \"deliberate critique.\" A curated dataset of 4.5K such examples is used to fine-tune the base model.\n\nStage 2 (Reinforcement Learning): Building upon the SFT model, RL is employed to further enhance its critique capabilities. The reward signals are sourced from two pathways: 1) existing human-annotated datasets (e.g., PRM800K), and 2) automatically generated step-level labels via Monte Carlo sampling, enabling scalable oversight.\n\nExperimental results demonstrate that the resulting DeepCritic-7B model significantly outperforms much larger and more capable baselines, including GPT-4o, on several mathematical error detection benchmarks. The paper further showcases the model's practical utility as both a verifier to improve solution accuracy and as a supervisor to guide other LLMs in refining their errors, even demonstrating a potential for weak-to-strong supervision.", "strengths": "- Importance and Timeliness of the Problem: The paper tackles a fundamental challenge at the forefront of LLM development. As the community shifts from outcome-based to process-based supervision, improving the quality of automated feedback (i.e., critique) is paramount for achieving scalable oversight and building more reliable and trustworthy LLMs. This work directly addresses a core bottleneck in this research direction.\n\n- Novel and Insightful Methodology: The paper's primary contribution is its ingenious \"meta-critique\" data generation strategy. This \"critique of the critique\" design cleverly pushes the model beyond simple right/wrong judgments, training it to perform multi-perspective verification, reflection, and self-correction. This is a far more profound approach than standard knowledge distillation, as it aims to instill a pattern of critical reasoning rather than just transferring knowledge.\n\n- Strong and Comprehensive Experimental Validation: The empirical results are compelling and well-executed.\n\t- The performance is impressive, with a 7B model outperforming top-tier models like GPT-4o on specialized tasks, clearly demonstrating the method's efficacy.\n\t- The evaluation is thorough, conducted across multiple standard benchmarks and supported by detailed ablation studies that justify the necessity of each component of the framework.\n\t- The paper effectively demonstrates the model's practical potential beyond benchmark scores, showcasing its value in applied scenarios like verified majority voting and solution refinement, with the latter hinting at the exciting prospect of weak-to-strong supervision.", "weaknesses": "- Scalability and Cost of the Data Generation Pipeline: The framework's main strength—its high-quality data—is also a potential weakness in terms of scalability. The data curation process is computationally intensive, requiring multiple long-sequence inference passes from a very large teacher model for each data point. This makes the cost of data generation exceedingly high, posing a significant challenge for scaling the dataset to millions of examples and potentially limiting its feasibility for large-scale industrial deployment.\n\n- Generalization to Other Domains: The method's success is demonstrated convincingly in the mathematical domain, which benefits from objective and verifiable ground truths. However, its generalizability to more subjective domains remains an open question. \n\n- High Dependency on the Teacher Model: The capabilities of the resulting DeepCritic model are inherently capped by the proficiency of the teacher model. Any systematic biases, knowledge gaps, or reasoning flaws present in the teacher are likely to be inherited, and possibly amplified, through the data generation process. The paper does not explore how the choice and capability of the teacher model impact the final outcome.\n\n- Ambiguity in Direct Application as an RL Reward Model: While RL is used in the second stage of training, the DeepCritic model itself does not produce a scalar reward suitable for direct use in standard online RL algorithms (e.g., PPO). Its output is a long-form text, making it challenging to efficiently integrate as a reward function in a closed-loop training setup. Its more immediate application appears to be as an offline data generator for preference tuning (e.g., DPO) or as a test-time refinement mechanism, while its role in online RL remains less clear.", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses a critical and timely issue: the superficial nature of critiques generated by Large Language Models (LLMs) when serving as evaluators, particularly in complex reasoning domains such as mathematics. The authors argue that existing LLM critics often produce shallow feedback that merely echoes the original solution's reasoning, leading to low error detection accuracy and insufficient guidance for correction.\n\nTo overcome this limitation, the paper introduces DeepCritic, a novel two-stage training framework designed to teach LLMs to perform \"deliberate critique.\"\n\nStage 1 (Supervised Fine-Tuning): The core innovation lies in its sophisticated data curation pipeline. First, a powerful teacher model (Qwen2.5-72B-Instruct) generates an \"initial critique\" for each step of a mathematical solution. Crucially, the teacher model then performs a deeper \"meta-critique\" on the initial critique itself. This involves multi-perspective verification and self-reflection to identify shortcomings in the initial assessment. Finally, these two critiques are synthesized into a single, high-quality, long-form \"deliberate critique.\" A curated dataset of 4.5K such examples is used to fine-tune the base model.\n\nStage 2 (Reinforcement Learning): Building upon the SFT model, RL is employed to further enhance its critique capabilities. The reward signals are sourced from two pathways: 1) existing human-annotated datasets (e.g., PRM800K), and 2) automatically generated step-level labels via Monte Carlo sampling, enabling scalable oversight.\n\nExperimental results demonstrate that the resulting DeepCritic-7B model significantly outperforms much larger and more capable baselines, including GPT-4o, on several mathematical error detection benchmarks. The paper further showcases the model's practical utility as both a verifier to improve solution accuracy and as a supervisor to guide other LLMs in refining their errors, even demonstrating a potential for weak-to-strong supervision.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Importance and Timeliness of the Problem: The paper tackles a fundamental challenge at the forefront of LLM development. As the community shifts from outcome-based to process-based supervision, improving the quality of automated feedback (i.e., critique) is paramount for achieving scalable oversight and building more reliable and trustworthy LLMs. This work directly addresses a core bottleneck in this research direction.\n\n- Novel and Insightful Methodology: The paper's primary contribution is its ingenious \"meta-critique\" data generation strategy. This \"critique of the critique\" design cleverly pushes the model beyond simple right/wrong judgments, training it to perform multi-perspective verification, reflection, and self-correction. This is a far more profound approach than standard knowledge distillation, as it aims to instill a pattern of critical reasoning rather than just transferring knowledge.\n\n- Strong and Comprehensive Experimental Validation: The empirical results are compelling and well-executed.\n\t- The performance is impressive, with a 7B model outperforming top-tier models like GPT-4o on specialized tasks, clearly demonstrating the method's efficacy.\n\t- The evaluation is thorough, conducted across multiple standard benchmarks and supported by detailed ablation studies that justify the necessity of each component of the framework.\n\t- The paper effectively demonstrates the model's practical potential beyond benchmark scores, showcasing its value in applied scenarios like verified majority voting and solution refinement, with the latter hinting at the exciting prospect of weak-to-strong supervision.", "weaknesses": "- Scalability and Cost of the Data Generation Pipeline: The framework's main strength—its high-quality data—is also a potential weakness in terms of scalability. The data curation process is computationally intensive, requiring multiple long-sequence inference passes from a very large teacher model for each data point. This makes the cost of data generation exceedingly high, posing a significant challenge for scaling the dataset to millions of examples and potentially limiting its feasibility for large-scale industrial deployment.\n\n- Generalization to Other Domains: The method's success is demonstrated convincingly in the mathematical domain, which benefits from objective and verifiable ground truths. However, its generalizability to more subjective domains remains an open question. \n\n- High Dependency on the Teacher Model: The capabilities of the resulting DeepCritic model are inherently capped by the proficiency of the teacher model. Any systematic biases, knowledge gaps, or reasoning flaws present in the teacher are likely to be inherited, and possibly amplified, through the data generation process. The paper does not explore how the choice and capability of the teacher model impact the final outcome.\n\n- Ambiguity in Direct Application as an RL Reward Model: While RL is used in the second stage of training, the DeepCritic model itself does not produce a scalar reward suitable for direct use in standard online RL algorithms (e.g., PPO). Its output is a long-form text, making it challenging to efficiently integrate as a reward function in a closed-loop training setup. Its more immediate application appears to be as an offline data generator for preference tuning (e.g., DPO) or as a test-time refinement mechanism, while its role in online RL remains less clear.", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761293458641}], "openreview_url": "https://openreview.net/forum?id=wyCnT4BUsT", "arxiv_id": "2505.00662", "paper_pdf": "papers/wyCnT4BUsT.pdf", "paper_pdf_sha256": "a5bc0f8fdd2ebe2b03684b968e67f72315e68ebc056c2403d248c3647d2372af", "paper_pdf_bytes": 1271889, "paper_pdf_source": "openreview", "code_url": "https://github.com/RUCBM/DeepCritic", "code_repository": "RUCBM/DeepCritic", "code_commit": "53eaf5e048187162451ff165e823c27d976d0e09", "code_archive": "repos/wyCnT4BUsT.zip", "code_archive_sha256": "d2db3f057f7b21ec1c4c7d9ce0c1e37a2d9af3590a4b845441a60dd01ce51a20", "code_archive_bytes": 4651850, "code_file_count": 24, "code_extensions": {".py": 19, ".sh": 5}, "github_disk_usage_kb": 4535, "github_languages": {"Python": 125044, "Shell": 3698}, "github_archived": false, "github_pushed_at": "2025-06-24T12:50:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deepcritic-deliberate-critique-with-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DDxLsxiZR8", "year": 2025, "status": "rejected", "title": "CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models", "authors": ["Xinle Cheng", "Zhuoming Chen", "Zhihao Jia"], "authorids": ["~Xinle_Cheng1", "~Zhuoming_Chen1", "~Zhihao_Jia2"], "authors_source": "OpenReview API", "abstract": "Diffusion models have transformed generative tasks, particularly in text-to-image synthesis, but their iterative denoising process is computationally intensive. We present a novel acceleration strategy that combines token-level pruning with cache mechanisms to address this challenge. By utilizing Noise Relative Magnitude, we identify significant token changes across iterations. Additionally, we incorporate spatial clustering and distributional balance to enhance token selection. Our experiments demonstrate 50\\%-60\\% reduction in computational cost while maintaining model performance, offering a substantial improvement in the efficiency of diffusion models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "XPrvha6Xkc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5910/Reviewer_3Lg7"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces CAT Pruning (Cluster-Aware Token Pruning), an acceleration technique for text-to-image diffusion models that aims to reduce computational cost by selectively updating tokens based on relative noise magnitude, spatial clustering, and balanced selection frequencies. By combining token pruning with caching, CAT Pruning demonstrates up to a 2x speedup and a 50-60% reduction in computation on two models (Stable Diffusion 3 and Pixart-Σ), two denoising steps (28 and 50), and two datasets (PartiPrompts and COCO2017), while maintaining CLIP score.", "review_text": "This paper introduces CAT Pruning (Cluster-Aware Token Pruning), an acceleration technique for text-to-image diffusion models that aims to reduce computational cost by selectively updating tokens based on relative noise magnitude, spatial clustering, and balanced selection frequencies. By combining token pruning with caching, CAT Pruning demonstrates up to a 2x speedup and a 50-60% reduction in computation on two models (Stable Diffusion 3 and Pixart-Σ), two denoising steps (28 and 50), and two datasets (PartiPrompts and COCO2017), while maintaining CLIP score.", "strengths": "1. The proposed method achieves notable speedup (up to 2x) while preserving CLIP Score.\n2. The proposed method can be applied to pre-trained models without additional training costs, making it a lightweight option for improving inference efficiency across different tasks.", "weaknesses": "1. The paper lacks a thorough discussion of prior token pruning work, making it difficult to assess the proposed method’s novelty and improvements over existing methods. A more detailed overview of token pruning techniques, their limitations, and how the proposed method addresses these would clarify its contributions.\n2. The study lacks comparisons with a diverse set of token pruning baselines and established training-free methods (e.g., caching), which limits a comprehensive view of the proposed method’s effectiveness relative to existing techniques.\n3. The qualitative differences shown in Figure 6 between clustering and non-clustering approaches are subtle, and there is no rigorous ablation study to quantitatively assess clustering’s impact on model performance. A thorough ablation study is needed to substantiate the claimed benefits of clustering.\n4. The paper relies solely on CLIP Score to assess image quality, which measures text-image alignment but not visual fidelity. Including metrics such as FID would provide a more complete evaluation of image quality and support claims of fidelity preservation.", "questions": "1. Could you clarify where the difference in inference speed arises between CAT Pruning and existing token pruning methods?\n2. What is t0 used in the experiments? \n3. Could you explain where the differences in speed and CLIP score arise between CAT Pruning and the AT-EDM baseline?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces CAT Pruning (Cluster-Aware Token Pruning), an acceleration technique for text-to-image diffusion models that aims to reduce computational cost by selectively updating tokens based on relative noise magnitude, spatial clustering, and balanced selection frequencies. By combining token pruning with caching, CAT Pruning demonstrates up to a 2x speedup and a 50-60% reduction in computation on two models (Stable Diffusion 3 and Pixart-Σ), two denoising steps (28 and 50), and two datasets (PartiPrompts and COCO2017), while maintaining CLIP score.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The proposed method achieves notable speedup (up to 2x) while preserving CLIP Score.\n2. The proposed method can be applied to pre-trained models without additional training costs, making it a lightweight option for improving inference efficiency across different tasks.", "weaknesses": "1. The paper lacks a thorough discussion of prior token pruning work, making it difficult to assess the proposed method’s novelty and improvements over existing methods. A more detailed overview of token pruning techniques, their limitations, and how the proposed method addresses these would clarify its contributions.\n2. The study lacks comparisons with a diverse set of token pruning baselines and established training-free methods (e.g., caching), which limits a comprehensive view of the proposed method’s effectiveness relative to existing techniques.\n3. The qualitative differences shown in Figure 6 between clustering and non-clustering approaches are subtle, and there is no rigorous ablation study to quantitatively assess clustering’s impact on model performance. A thorough ablation study is needed to substantiate the claimed benefits of clustering.\n4. The paper relies solely on CLIP Score to assess image quality, which measures text-image alignment but not visual fidelity. Including metrics such as FID would provide a more complete evaluation of image quality and support claims of fidelity preservation.", "questions": "1. Could you clarify where the difference in inference speed arises between CAT Pruning and existing token pruning methods?\n2. What is t0 used in the experiments? \n3. Could you explain where the differences in speed and CLIP score arise between CAT Pruning and the AT-EDM baseline?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730837632111}, {"id": "cVlC7NzjdW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5910/Reviewer_9Tq1"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "The paper presents CAT Pruning, a method to increase the computational efficiency of diffusion model sampling. CAT pruning targets the attention calculations of the DM and aims to eliminate irrelevant tokens. The method combines clustering results, noise magnitude and staleness of tokens to identify unimportant tokens. The results indicate no perceiptable loss of output information when pruning up to 70% of tokens which results in a speedup of up to 60%.", "review_text": "The paper presents CAT Pruning, a method to increase the computational efficiency of diffusion model sampling. CAT pruning targets the attention calculations of the DM and aims to eliminate irrelevant tokens. The method combines clustering results, noise magnitude and staleness of tokens to identify unimportant tokens. The results indicate no perceiptable loss of output information when pruning up to 70% of tokens which results in a speedup of up to 60%.", "strengths": "- **Important topic.** Computational efficenciy is one of the main limitations of Diffusion Models. Works addressing these issues are of high value to the field. \n- **Good performance** CAT Pruning appears to preseve the original performance of the model well, while yielding a decent speedup.", "weaknesses": "The paper has two overarching weaknesses that become apparent in a multitude of smaller issues.  \n\n**Presentation**\nKey aspects of the paper are presented badly, making it hard for readers to grasp the contributions made by CAT-Pruning\n- Underlying fundamentals of the method are not sufficiently explained. For example nowhere in the Abstract, Introduction or Conclusion to the authors elaborate that they prune tokens in the attention layers of the DM. In fact when asking multiple computer vision researchers what they assumed the method to be about, all of them assumed that tokens where pruned in the embeddings of the text prompt. \n- The authors use MACs as a key performance metric throughout the paper without ever explaining it\n- Similarly it is never explained between which samples the CLIP Score is calculated during evaluation \n- a third of Page 7 is just empty, but Page 8 is almost exclusively Figures\n- Tables 2 and 3 do not contain any bold numbers \n- The diffusion notations are somewhat disconnected from common conventions. For example, noise (estimates) are usually reffered to as $\\epsilon$, during generation diffusion steps $t$ should **decrease** since this is the reverse diffusion process. The authors also claim that Diffusion involves solving a reverse-time SDE, which is indeed one valid mathmatical foundation for diffusion. However, it specifically does not encompas the models actually used in the remainder of the paper, with SD3 being a rectified flow model that is specifically incompatibel with stochhastic algorithms. \n\n\n**Evaluation**\nFurther, there are problems with the validation of the poposed method. \n- The main claim is that the output image after pruning is perceptually similar to the image generated without pruning. However, that specific aspect is never empirically evaluated. The obvious choice here would have been to simply report LPIPS distance between the pruned und unpruned image. \n- Similarly the tradeoff between compute reduction and (un)pruning percentage $\\alpha$ is demonstrated with 4 examples but not ablated over empirically \n- In the same vain many of the design choices of the CAT algorithm are showcased with 1 or 2 qualitiative examples with limited to no empirical ablations. A more structured analysis on the importance/downstream influence of the different components in the selection algorithm would have been important \n- The empriical results in Tab. 2 and 3 do not contain any confidence intervals or standard deviations \n- There is no qualitative comparison with competing methods\n- The authors only compare against one other baseline, although other methods exist, including ∆-DiT [1], Faster Diffusion [2], or TGATE [3] or basic methods like KV caching\n- The most prominent choices for speeding up DM inference are, of course, distillation methods or consistency models. Consequently, it would be important to consider if CAT-Pruning still offers advantages when applied to distilled or consistency models.  \n\n[1] Pengtao Chen, et al.  ∆-DIT: A training-free acceleration method tailored for diffusion transformers. arXiv:2406.01125\n[2] Senmao Li et al. Faster diffusion: Rethinking the role of unet encoder in diffusion models.  arXiv:2312.09608, 2023.\n[3 Wentian Zhang, et al. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv:2404.02747, 2024.\n\n**Other**\n - Method appears to be limited to DiT architecture. At least no other architecture is considered", "questions": "- **Q1.** Is CAT-Pruning restricted to DiTs or does it also apply to other architectures like UNet DMs with attention?\n- **Q2.** You write that \"cached features must remain consistent across timesteps\" L 110 and that your methods combines token-level pruning with ache mechanisms L 64. How exactly is the cache optimization realized with CAT-Pruning? At no point in the paper do you mention which part of the method actually is responsible for the cache optimiztion. Is that achieved by the Frequency monitoring over timesteps?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents CAT Pruning, a method to increase the computational efficiency of diffusion model sampling. CAT pruning targets the attention calculations of the DM and aims to eliminate irrelevant tokens. The method combines clustering results, noise magnitude and staleness of tokens to identify unimportant tokens. The results indicate no perceiptable loss of output information when pruning up to 70% of tokens which results in a speedup of up to 60%.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "- **Important topic.** Computational efficenciy is one of the main limitations of Diffusion Models. Works addressing these issues are of high value to the field. \n- **Good performance** CAT Pruning appears to preseve the original performance of the model well, while yielding a decent speedup.", "weaknesses": "The paper has two overarching weaknesses that become apparent in a multitude of smaller issues.  \n\n**Presentation**\nKey aspects of the paper are presented badly, making it hard for readers to grasp the contributions made by CAT-Pruning\n- Underlying fundamentals of the method are not sufficiently explained. For example nowhere in the Abstract, Introduction or Conclusion to the authors elaborate that they prune tokens in the attention layers of the DM. In fact when asking multiple computer vision researchers what they assumed the method to be about, all of them assumed that tokens where pruned in the embeddings of the text prompt. \n- The authors use MACs as a key performance metric throughout the paper without ever explaining it\n- Similarly it is never explained between which samples the CLIP Score is calculated during evaluation \n- a third of Page 7 is just empty, but Page 8 is almost exclusively Figures\n- Tables 2 and 3 do not contain any bold numbers \n- The diffusion notations are somewhat disconnected from common conventions. For example, noise (estimates) are usually reffered to as $\\epsilon$, during generation diffusion steps $t$ should **decrease** since this is the reverse diffusion process. The authors also claim that Diffusion involves solving a reverse-time SDE, which is indeed one valid mathmatical foundation for diffusion. However, it specifically does not encompas the models actually used in the remainder of the paper, with SD3 being a rectified flow model that is specifically incompatibel with stochhastic algorithms. \n\n\n**Evaluation**\nFurther, there are problems with the validation of the poposed method. \n- The main claim is that the output image after pruning is perceptually similar to the image generated without pruning. However, that specific aspect is never empirically evaluated. The obvious choice here would have been to simply report LPIPS distance between the pruned und unpruned image. \n- Similarly the tradeoff between compute reduction and (un)pruning percentage $\\alpha$ is demonstrated with 4 examples but not ablated over empirically \n- In the same vain many of the design choices of the CAT algorithm are showcased with 1 or 2 qualitiative examples with limited to no empirical ablations. A more structured analysis on the importance/downstream influence of the different components in the selection algorithm would have been important \n- The empriical results in Tab. 2 and 3 do not contain any confidence intervals or standard deviations \n- There is no qualitative comparison with competing methods\n- The authors only compare against one other baseline, although other methods exist, including ∆-DiT [1], Faster Diffusion [2], or TGATE [3] or basic methods like KV caching\n- The most prominent choices for speeding up DM inference are, of course, distillation methods or consistency models. Consequently, it would be important to consider if CAT-Pruning still offers advantages when applied to distilled or consistency models.  \n\n[1] Pengtao Chen, et al.  ∆-DIT: A training-free acceleration method tailored for diffusion transformers. arXiv:2406.01125\n[2] Senmao Li et al. Faster diffusion: Rethinking the role of unet encoder in diffusion models.  arXiv:2312.09608, 2023.\n[3 Wentian Zhang, et al. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv:2404.02747, 2024.\n\n**Other**\n - Method appears to be limited to DiT architecture. At least no other architecture is considered", "questions": "- **Q1.** Is CAT-Pruning restricted to DiTs or does it also apply to other architectures like UNet DMs with attention?\n- **Q2.** You write that \"cached features must remain consistent across timesteps\" L 110 and that your methods combines token-level pruning with ache mechanisms L 64. How exactly is the cache optimization realized with CAT-Pruning? At no point in the paper do you mention which part of the method actually is responsible for the cache optimiztion. Is that achieved by the Frequency monitoring over timesteps?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730723378493}, {"id": "yJKOAx6rqI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5910/Reviewer_nXeT"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper (CAT Pruning) introduces a technique to increase the inference efficiency of diffusion models. In particular, the authors combine caching mechanisms with token-wise pruning which reduces the computational overhead during inference. The paper is overall well written and provides good quantitative results for latest text-to-image models (such as SD-3), however lacks some performance benchmarking against some baselines and a more holistic evaluation of generation beyond the CLIP-score.", "review_text": "The paper (CAT Pruning) introduces a technique to increase the inference efficiency of diffusion models. In particular, the authors combine caching mechanisms with token-wise pruning which reduces the computational overhead during inference. The paper is overall well written and provides good quantitative results for latest text-to-image models (such as SD-3), however lacks some performance benchmarking against some baselines and a more holistic evaluation of generation beyond the CLIP-score.", "strengths": "Below I state the strengths and weaknesses of the paper:\n\n- The paper is well-written and easy to follow — with a strong motivation. Moreover improving the efficiency during inference, is a practical problem which still has some open problems for the newer t2i models based on the transformer architecture.\n- The analysis on the token selection strategy based on relative noise magnitude (combined with spatial clustering and balancing) is comprehensive.", "weaknesses": "Weaknesses:\n\n- [Minor] The authors are suggested to provide a brief overview of the diffusion model architecture (e.g., SD-3 like architecture) to which their method is applicable.  This will improve readability by quite an extent.\n- [Major] The paper does not compare with some of the caching only baselines (e.g., TGATE which they cite). This would be important to understand the effectiveness of combining caching with pruning techniques. Although not directly comparable, I’d also suggest the authors to provide the distillation based baselines (less number of steps in the inference) in the paper, to provide a full picture of the effectiveness of the caching + pruning family of methods. \n- [Major] The authors provide the CLIP-score for generation quality — but it’d be more effective to show a more holistic evaluation of the method in terms of generation quality. For example, the authors can test on compositionality, long caption generation etc for their method in terms of generation. \n- [Minor]: The authors are suggested to provide more qualitative results comparing the generation across different methods improving efficiency.", "questions": "See Weakness. \n\nOverall, the paper introduces a technically solid method, but lacks some important comparisons on evaluations. I am happy to revisit my score during the rebuttal, if the authors respond to the Weaknesses adequately.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper (CAT Pruning) introduces a technique to increase the inference efficiency of diffusion models. In particular, the authors combine caching mechanisms with token-wise pruning which reduces the computational overhead during inference. The paper is overall well written and provides good quantitative results for latest text-to-image models (such as SD-3), however lacks some performance benchmarking against some baselines and a more holistic evaluation of generation beyond the CLIP-score.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Below I state the strengths and weaknesses of the paper:\n\n- The paper is well-written and easy to follow — with a strong motivation. Moreover improving the efficiency during inference, is a practical problem which still has some open problems for the newer t2i models based on the transformer architecture.\n- The analysis on the token selection strategy based on relative noise magnitude (combined with spatial clustering and balancing) is comprehensive.", "weaknesses": "Weaknesses:\n\n- [Minor] The authors are suggested to provide a brief overview of the diffusion model architecture (e.g., SD-3 like architecture) to which their method is applicable.  This will improve readability by quite an extent.\n- [Major] The paper does not compare with some of the caching only baselines (e.g., TGATE which they cite). This would be important to understand the effectiveness of combining caching with pruning techniques. Although not directly comparable, I’d also suggest the authors to provide the distillation based baselines (less number of steps in the inference) in the paper, to provide a full picture of the effectiveness of the caching + pruning family of methods. \n- [Major] The authors provide the CLIP-score for generation quality — but it’d be more effective to show a more holistic evaluation of the method in terms of generation quality. For example, the authors can test on compositionality, long caption generation etc for their method in terms of generation. \n- [Minor]: The authors are suggested to provide more qualitative results comparing the generation across different methods improving efficiency.", "questions": "See Weakness. \n\nOverall, the paper introduces a technically solid method, but lacks some important comparisons on evaluations. I am happy to revisit my score during the rebuttal, if the authors respond to the Weaknesses adequately.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730670951291}, {"id": "Lkwi0ff56T", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5910/Reviewer_upgQ"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes a new method for accelerating text-to-image diffusion models by selectively updating a subset of tokens during the denoising process. The authors introduce Cluster-Aware Token Pruning (CAT Pruning), which leverages the relative noise magnitude of tokens, their selection frequencies, and spatial clustering to achieve significant computational savings while maintaining the quality of generated images. They demonstrate that CAT Pruning can achieve up to a 50% reduction in computational costs with minimal impact on image quality, making diffusion models more efficient for generating high-resolution images. The paper also provides extensive experimental results on popular datasets and pretrained diffusion models, comparing CAT Pruning to existing methods and highlighting its superior performance.\n\nKey contributions of the paper include:\nProposing A Token Importance Ranking Procedure: The paper establishes a method for ranking token importance that considers not only noise magnitude but also selection frequencies across timesteps, ensuring consistent token selection.\nA Cluster-Aware Pruning Method: The authors propose a unique pruning method that integrates spatial clustering, leveraging positional encoding to maintain spatial coherence and detail preservation in generated images. This approach improves the quality of outputs compared to simple sequential token selection strategies.\nMaking the Case for Distributional Balance: The paper emphasizes the importance of distributional balance within clusters. This balance is achieved by considering both noise magnitude and selection frequencies when selecting tokens within each cluster. This contributes to a more nuanced pruning process that avoids over-emphasizing certain features at the expense of others.", "review_text": "This paper proposes a new method for accelerating text-to-image diffusion models by selectively updating a subset of tokens during the denoising process. The authors introduce Cluster-Aware Token Pruning (CAT Pruning), which leverages the relative noise magnitude of tokens, their selection frequencies, and spatial clustering to achieve significant computational savings while maintaining the quality of generated images. They demonstrate that CAT Pruning can achieve up to a 50% reduction in computational costs with minimal impact on image quality, making diffusion models more efficient for generating high-resolution images. The paper also provides extensive experimental results on popular datasets and pretrained diffusion models, comparing CAT Pruning to existing methods and highlighting its superior performance.\n\nKey contributions of the paper include:\nProposing A Token Importance Ranking Procedure: The paper establishes a method for ranking token importance that considers not only noise magnitude but also selection frequencies across timesteps, ensuring consistent token selection.\nA Cluster-Aware Pruning Method: The authors propose a unique pruning method that integrates spatial clustering, leveraging positional encoding to maintain spatial coherence and detail preservation in generated images. This approach improves the quality of outputs compared to simple sequential token selection strategies.\nMaking the Case for Distributional Balance: The paper emphasizes the importance of distributional balance within clusters. This balance is achieved by considering both noise magnitude and selection frequencies when selecting tokens within each cluster. This contributes to a more nuanced pruning process that avoids over-emphasizing certain features at the expense of others.", "strengths": "Originality: The paper presents a novel approach called CAT Pruning, which combines token-level pruning with caching techniques to accelerate text-to-image diffusion models. While previous works have explored caching and reuse mechanisms to reduce inference time, CAT Pruning focuses on optimizing at the intra-kernel level by reducing latency within individual kernel executions. The authors introduce the concept of “Relative Noise Magnitude” to identify significant token changes across denoising iterations. This concept is defined as the difference between the current predicted noise and the noise at step t0, which is defined as nt−nt0 and quantifies the relative change in noise They also incorporate spatial clustering and ensure distributional balance to enhance token selection, further improving efficiency and preserving model performance.\n\nQuality: The paper demonstrates a high level of quality through a reasonable amount of experimentation and analysis. The authors evaluate CAT Pruning on standard datasets like MS-COCO 2017 and PartiPrompts, using established pretrained diffusion models such as Stable Diffusion v3 and Pixart-Σ. They compare their method against relevant baselines, including the standard diffusion model output and AT-EDM, another token pruning technique. The results show significant reductions in computation costs (up to 50% reduction in MACs at 28 denoising steps and 60% at 50 denoising steps—although the authors should actually define the acronym MACs) while maintaining comparable or even superior CLIP scores. The authors provide visualizations of generated images at different sparsity levels, demonstrating the effectiveness of CAT Pruning in preserving image quality even with significant pruning. They also offer insights into the correlation between predicted noise and historical noise, justifying their token selection strategy.\n\nClarity: The paper is relatively well-written and structured, presenting the proposed method in a clear and concise manner. The authors provide a reasonable overview of the problem and related work, highlighting the limitations of existing approaches. They clearly define some key notation in Table 1 and describe the algorithm using illustrative examples and figures. The experimental setup is detailed, allowing for reproducibility and a clear understanding of the evaluation process. The results are presented in tables and visualized through figures, facilitating interpretation and analysis.\n\nSignificance: The paper addresses a crucial challenge in the field of text-to-image synthesis: the high computational cost of diffusion models. By significantly accelerating inference time without compromising image quality, CAT Pruning has the potential to make these powerful generative models more accessible for various applications. This work contributes to the growing body of research on optimizing diffusion models and could inspire further advancements in efficiency and scalability. The authors’ insights into token-level pruning and the exploitation of feature redundancy could benefit other generative tasks beyond text-to-image synthesis.", "weaknesses": "The paper has limited theoretical justification: The paper primarily relies on empirical observations and intuitions to justify the effectiveness of CAT Pruning. While the authors present Proposition 1 and provide a simplified proof in the appendix, a more rigorous theoretical analysis could strengthen the paper's contribution.  A deeper theoretical understanding of the relationship between relative noise magnitude, token staleness, and spatial clustering could lead to more informed design choices and potentially improved performance. For example, exploring the convergence properties of the algorithm or deriving bounds on the error introduced by pruning would provide valuable insights.\n\nThere is a lack of comparison with other pruning techniques, and other techniques in general: The paper compares CAT Pruning only with AT-EDM, another token pruning technique. However, a comprehensive comparison with a wider range of pruning methods for diffusion models, such as those leveraging model distillation, quantization, or low-rank factorization, would provide a more complete picture of the proposed method's strengths and limitations. This would allow for a more informed assessment of the relative performance and efficiency of CAT Pruning compared to other state-of-the-art techniques. In particular the paper seems to ignore the highly influential VQVAE and VQGAN based methods for leveraging compressed latent representations of data. These methods are extremely popular ways for reducing the computational load of diffusion models. See Gu et al., (2022) for example (among many other examples), i.e.\n\nGu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L. and Guo, B., 2022. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10696-10706).\n\n\"CAT Pruning\" focuses on optimizing token processing within a U-Net architecture. However, VQ-Diffusion demonstrates that utilizing a VQVAE to learn a compressed, discrete latent representation can lead to significant computational savings. By shifting the diffusion process to this lower-dimensional latent space, VQ-Diffusion achieves notable speed improvements.\nThe \"CAT Pruning\" paper does not acknowledge or compare its approach to this architectural shift towards latent space diffusion using VQVAEs, which constitutes a notable weakness. Similarly, the paper on \"High-Resolution Image Synthesis with Latent Diffusion Models\"  of\n\nRombach, R., Blattmann, A., Lorenz, D., Esser, P. and Ommer, B., 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10684-10695).\n\ndiscusses the popularity and advantages of VQGANs, particularly their ability to learn compressed latent representations that can be used for high-resolution image synthesis. VQGANs, in conjunction with autoregressive models, have emerged as powerful tools for high-resolution image synthesis. These models operate on a compressed, discretized latent space learned by a VQGAN, potentially offering significant computational advantages over pixel-based approaches. The \"High-Resolution Image Synthesis\" paper uses VQ-regularization as one method for training the autoencoder that produces the latent space for their LDMs. They discuss how their model \"can be interpreted as a VQGAN but with the quantization layer absorbed by the decoder.\"\n\nSince the proposed CAT Pruning technique emphasizes reducing computational costs as a primary goal, comparing this CAT token pruning method with VQVAE and VQGAN style approaches in a much more direct manner would be very helpful. These methods inherently operate in a compressed latent space, and understanding the relationships between these kinds of methods and the proposed method would be very helpful, comparing empirically would be strongly desired and it could demonstrate the relative efficiency of CAT Pruning much more clearly.\n\nIn summary, with respect to the issue of comparing the work here to these popular methods, the paper would be strengthened by:\n1.\tExplicitly at least acknowledging the popularity of VQVAE and VQGAN-based approaches for diffusion model acceleration.\n2.\tDiscussing the potential trade-offs between their token pruning within the U-Net architecture and the use of VQGANs and other approaches for latent space diffusion.\n3.\tIdeally, conducting more direct experiments to compare the performance, efficiency, and image quality of CAT Pruning against a VQGAN-based diffusion model.", "questions": "Questions: \n\nRegarding VQVAEs and Latent Space Diffusion: The \"Vector Quantized Diffusion Model for Text-to-Image Synthesis\" paper presents VQ-Diffusion, a method that uses a VQVAE to perform diffusion in a compressed latent space. This approach achieves significant speed improvements. \n\n* Given the potential efficiency gains of VQVAEs for diffusion, could the authors explain their rationale for focusing on token pruning within the U-Net architecture and not comparing with those methods, conceptually or emprically? \n* What are the perceived advantages and disadvantages of each approach?\n* Considering the importance of VQGANs in this domain, why weren't they included as a baseline for comparison in \"CAT Pruning\"?\nA key focus of \"CAT Pruning\" is computational efficiency. \n* Could the authors provide a more direct comparison of the efficiency gains of CAT Pruning against VQGAN-based diffusion models and LDMs? \nThis would involve metrics like inference time, memory usage, and the number of floating-point operations.\n\nSuggestions:\n\nWhile Proposition 1 is presented, a more comprehensive theoretical foundation for CAT Pruning would strengthen the paper. Pfdiff provides a very detailed analytical explanation for why their approach takes the decisions that it does. Brining the work here closer to that level of analytical analysis would be helpful. Some ideas for how that might be achieved include: \nDeriving bounds on the error introduced by pruning.\nExploring the relationship between noise magnitude, staleness, and clustering in more depth.\n\nAs a point of comparison \"Deep Cache\" emphasizes that the high-level features generated during the reverse diffusion process exhibit significant temporal consistency. This observation forms the basis for their caching mechanism, which avoids redundant computations. This work could benefit from explicitly acknowledging and discussing the role of temporal redundancy in the effectiveness of their method. Perhaps one could explain how the relative stability of certain tokens across timesteps (as captured by the \"staleness\" metric) might relate to the temporal consistency observed in \"Deep Cache.\" \"Deep Cache\" includes comparisons with various baselines, including pruning and distillation methods. \"CAT Pruning\" could benefit from a similarly more comprehensive evaluation. For example the work here could directly compare \"CAT Pruning\" with \"Deep Cache\" to assess their relative performance and efficiency.\n\nAT-EDM introduces a Denoising-Steps-Aware Pruning (DSAP) schedule that adjusts pruning ratios across different denoising timesteps. This schedule prunes fewer tokens in early steps when attention maps are less informative and more aggressively in later steps when redundancy is higher. CAT Pruning also acknowledges the varying importance of denoising steps and implements a prune-less schedule in early steps. Some kind of discussion and comparison of these different approaches and their motivations could be helpful.\n\nWhile the paper mentions the use of existing caching techniques, a more in-depth discussion of token recovery strategies, particularly in the context of subsequent convolutional layers, would strengthen the paper. Exploring alternative methods, such as the similarity-based copy technique proposed in AT-EDM, could further improve the effectiveness and generalizability of CAT Pruning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new method for accelerating text-to-image diffusion models by selectively updating a subset of tokens during the denoising process. The authors introduce Cluster-Aware Token Pruning (CAT Pruning), which leverages the relative noise magnitude of tokens, their selection frequencies, and spatial clustering to achieve significant computational savings while maintaining the quality of generated images. They demonstrate that CAT Pruning can achieve up to a 50% reduction in computational costs with minimal impact on image quality, making diffusion models more efficient for generating high-resolution images. The paper also provides extensive experimental results on popular datasets and pretrained diffusion models, comparing CAT Pruning to existing methods and highlighting its superior performance.\n\nKey contributions of the paper include:\nProposing A Token Importance Ranking Procedure: The paper establishes a method for ranking token importance that considers not only noise magnitude but also selection frequencies across timesteps, ensuring consistent token selection.\nA Cluster-Aware Pruning Method: The authors propose a unique pruning method that integrates spatial clustering, leveraging positional encoding to maintain spatial coherence and detail preservation in generated images. This approach improves the quality of outputs compared to simple sequential token selection strategies.\nMaking the Case for Distributional Balance: The paper emphasizes the importance of distributional balance within clusters. This balance is achieved by considering both noise magnitude and selection frequencies when selecting tokens within each cluster. This contributes to a more nuanced pruning process that avoids over-emphasizing certain features at the expense of others.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Originality: The paper presents a novel approach called CAT Pruning, which combines token-level pruning with caching techniques to accelerate text-to-image diffusion models. While previous works have explored caching and reuse mechanisms to reduce inference time, CAT Pruning focuses on optimizing at the intra-kernel level by reducing latency within individual kernel executions. The authors introduce the concept of “Relative Noise Magnitude” to identify significant token changes across denoising iterations. This concept is defined as the difference between the current predicted noise and the noise at step t0, which is defined as nt−nt0 and quantifies the relative change in noise They also incorporate spatial clustering and ensure distributional balance to enhance token selection, further improving efficiency and preserving model performance.\n\nQuality: The paper demonstrates a high level of quality through a reasonable amount of experimentation and analysis. The authors evaluate CAT Pruning on standard datasets like MS-COCO 2017 and PartiPrompts, using established pretrained diffusion models such as Stable Diffusion v3 and Pixart-Σ. They compare their method against relevant baselines, including the standard diffusion model output and AT-EDM, another token pruning technique. The results show significant reductions in computation costs (up to 50% reduction in MACs at 28 denoising steps and 60% at 50 denoising steps—although the authors should actually define the acronym MACs) while maintaining comparable or even superior CLIP scores. The authors provide visualizations of generated images at different sparsity levels, demonstrating the effectiveness of CAT Pruning in preserving image quality even with significant pruning. They also offer insights into the correlation between predicted noise and historical noise, justifying their token selection strategy.\n\nClarity: The paper is relatively well-written and structured, presenting the proposed method in a clear and concise manner. The authors provide a reasonable overview of the problem and related work, highlighting the limitations of existing approaches. They clearly define some key notation in Table 1 and describe the algorithm using illustrative examples and figures. The experimental setup is detailed, allowing for reproducibility and a clear understanding of the evaluation process. The results are presented in tables and visualized through figures, facilitating interpretation and analysis.\n\nSignificance: The paper addresses a crucial challenge in the field of text-to-image synthesis: the high computational cost of diffusion models. By significantly accelerating inference time without compromising image quality, CAT Pruning has the potential to make these powerful generative models more accessible for various applications. This work contributes to the growing body of research on optimizing diffusion models and could inspire further advancements in efficiency and scalability. The authors’ insights into token-level pruning and the exploitation of feature redundancy could benefit other generative tasks beyond text-to-image synthesis.", "weaknesses": "The paper has limited theoretical justification: The paper primarily relies on empirical observations and intuitions to justify the effectiveness of CAT Pruning. While the authors present Proposition 1 and provide a simplified proof in the appendix, a more rigorous theoretical analysis could strengthen the paper's contribution.  A deeper theoretical understanding of the relationship between relative noise magnitude, token staleness, and spatial clustering could lead to more informed design choices and potentially improved performance. For example, exploring the convergence properties of the algorithm or deriving bounds on the error introduced by pruning would provide valuable insights.\n\nThere is a lack of comparison with other pruning techniques, and other techniques in general: The paper compares CAT Pruning only with AT-EDM, another token pruning technique. However, a comprehensive comparison with a wider range of pruning methods for diffusion models, such as those leveraging model distillation, quantization, or low-rank factorization, would provide a more complete picture of the proposed method's strengths and limitations. This would allow for a more informed assessment of the relative performance and efficiency of CAT Pruning compared to other state-of-the-art techniques. In particular the paper seems to ignore the highly influential VQVAE and VQGAN based methods for leveraging compressed latent representations of data. These methods are extremely popular ways for reducing the computational load of diffusion models. See Gu et al., (2022) for example (among many other examples), i.e.\n\nGu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L. and Guo, B., 2022. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10696-10706).\n\n\"CAT Pruning\" focuses on optimizing token processing within a U-Net architecture. However, VQ-Diffusion demonstrates that utilizing a VQVAE to learn a compressed, discrete latent representation can lead to significant computational savings. By shifting the diffusion process to this lower-dimensional latent space, VQ-Diffusion achieves notable speed improvements.\nThe \"CAT Pruning\" paper does not acknowledge or compare its approach to this architectural shift towards latent space diffusion using VQVAEs, which constitutes a notable weakness. Similarly, the paper on \"High-Resolution Image Synthesis with Latent Diffusion Models\"  of\n\nRombach, R., Blattmann, A., Lorenz, D., Esser, P. and Ommer, B., 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10684-10695).\n\ndiscusses the popularity and advantages of VQGANs, particularly their ability to learn compressed latent representations that can be used for high-resolution image synthesis. VQGANs, in conjunction with autoregressive models, have emerged as powerful tools for high-resolution image synthesis. These models operate on a compressed, discretized latent space learned by a VQGAN, potentially offering significant computational advantages over pixel-based approaches. The \"High-Resolution Image Synthesis\" paper uses VQ-regularization as one method for training the autoencoder that produces the latent space for their LDMs. They discuss how their model \"can be interpreted as a VQGAN but with the quantization layer absorbed by the decoder.\"\n\nSince the proposed CAT Pruning technique emphasizes reducing computational costs as a primary goal, comparing this CAT token pruning method with VQVAE and VQGAN style approaches in a much more direct manner would be very helpful. These methods inherently operate in a compressed latent space, and understanding the relationships between these kinds of methods and the proposed method would be very helpful, comparing empirically would be strongly desired and it could demonstrate the relative efficiency of CAT Pruning much more clearly.\n\nIn summary, with respect to the issue of comparing the work here to these popular methods, the paper would be strengthened by:\n1.\tExplicitly at least acknowledging the popularity of VQVAE and VQGAN-based approaches for diffusion model acceleration.\n2.\tDiscussing the potential trade-offs between their token pruning within the U-Net architecture and the use of VQGANs and other approaches for latent space diffusion.\n3.\tIdeally, conducting more direct experiments to compare the performance, efficiency, and image quality of CAT Pruning against a VQGAN-based diffusion model.", "questions": "Questions: \n\nRegarding VQVAEs and Latent Space Diffusion: The \"Vector Quantized Diffusion Model for Text-to-Image Synthesis\" paper presents VQ-Diffusion, a method that uses a VQVAE to perform diffusion in a compressed latent space. This approach achieves significant speed improvements. \n\n* Given the potential efficiency gains of VQVAEs for diffusion, could the authors explain their rationale for focusing on token pruning within the U-Net architecture and not comparing with those methods, conceptually or emprically? \n* What are the perceived advantages and disadvantages of each approach?\n* Considering the importance of VQGANs in this domain, why weren't they included as a baseline for comparison in \"CAT Pruning\"?\nA key focus of \"CAT Pruning\" is computational efficiency. \n* Could the authors provide a more direct comparison of the efficiency gains of CAT Pruning against VQGAN-based diffusion models and LDMs? \nThis would involve metrics like inference time, memory usage, and the number of floating-point operations.\n\nSuggestions:\n\nWhile Proposition 1 is presented, a more comprehensive theoretical foundation for CAT Pruning would strengthen the paper. Pfdiff provides a very detailed analytical explanation for why their approach takes the decisions that it does. Brining the work here closer to that level of analytical analysis would be helpful. Some ideas for how that might be achieved include: \nDeriving bounds on the error introduced by pruning.\nExploring the relationship between noise magnitude, staleness, and clustering in more depth.\n\nAs a point of comparison \"Deep Cache\" emphasizes that the high-level features generated during the reverse diffusion process exhibit significant temporal consistency. This observation forms the basis for their caching mechanism, which avoids redundant computations. This work could benefit from explicitly acknowledging and discussing the role of temporal redundancy in the effectiveness of their method. Perhaps one could explain how the relative stability of certain tokens across timesteps (as captured by the \"staleness\" metric) might relate to the temporal consistency observed in \"Deep Cache.\" \"Deep Cache\" includes comparisons with various baselines, including pruning and distillation methods. \"CAT Pruning\" could benefit from a similarly more comprehensive evaluation. For example the work here could directly compare \"CAT Pruning\" with \"Deep Cache\" to assess their relative performance and efficiency.\n\nAT-EDM introduces a Denoising-Steps-Aware Pruning (DSAP) schedule that adjusts pruning ratios across different denoising timesteps. This schedule prunes fewer tokens in early steps when attention maps are less informative and more aggressively in later steps when redundancy is higher. CAT Pruning also acknowledges the varying importance of denoising steps and implements a prune-less schedule in early steps. Some kind of discussion and comparison of these different approaches and their motivations could be helpful.\n\nWhile the paper mentions the use of existing caching techniques, a more in-depth discussion of token recovery strategies, particularly in the context of subsequent convolutional layers, would strengthen the paper. Exploring alternative methods, such as the similarity-based copy technique proposed in AT-EDM, could further improve the effectiveness and generalizability of CAT Pruning.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730661709508}], "openreview_url": "https://openreview.net/forum?id=DDxLsxiZR8", "arxiv_id": "2502.00433", "paper_pdf": "papers/DDxLsxiZR8.pdf", "paper_pdf_sha256": "b0e71fb65d37783c4203afae0f90ff1b3cff73ec508469de303a702aeecb89d7", "paper_pdf_bytes": 3146516, "paper_pdf_source": "openreview", "code_url": "https://github.com/ada-cheng/CAT-Pruning", "code_repository": "ada-cheng/CAT-Pruning", "code_commit": "3b88889e4deea8f1033b509efe3fd5bab90a7b43", "code_archive": "repos/DDxLsxiZR8.zip", "code_archive_sha256": "09f45b3d3140b619917c40786910176793d7e5d301296f6d975da1f82b5fd511", "code_archive_bytes": 334769, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 324, "github_languages": {"Python": 207497}, "github_archived": false, "github_pushed_at": "2025-07-26T06:11:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cat-pruning-cluster-aware-token-pruning-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tAmfM1sORP", "year": 2024, "status": "rejected", "title": "Large Language Models can Learn Rules", "authors": ["Zhaocheng Zhu", "Yuan Xue", "Xinyun Chen", "Denny Zhou", "Jian Tang", "Dale Schuurmans", "Hanjun Dai"], "authorids": ["~Zhaocheng_Zhu1", "~Yuan_Xue5", "~Xinyun_Chen1", "~Denny_Zhou1", "~Jian_Tang1", "~Dale_Schuurmans1", "~Hanjun_Dai1"], "authors_source": "OpenReview API", "abstract": "When prompted with a few examples and intermediate steps, large language models (LLMs) have demonstrated impressive performance in various reasoning tasks. However, prompting methods that rely on implicit knowledge in an LLM often hallucinate incorrect answers when the implicit knowledge is wrong or inconsistent with the task. To tackle this problem, we present Hypotheses-to-Theories (HtT), a framework that learns a rule library for reasoning with LLMs. HtT contains two stages, an induction stage and a deduction stage. In the induction stage, an LLM is first asked to generate and verify rules over a set of training examples. Rules that appear and lead to correct answers sufficiently often are collected to form a rule library. In the deduction stage, the LLM is then prompted to employ the learned rule library to perform reasoning to answer test questions. Experiments on both numerical reasoning and relational reasoning problems show that HtT improves existing prompting methods, with an absolute gain of 11-27% in accuracy. The learned rules are also transferable to different models and to different forms of the same problem.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "tzGnP7ndG1", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4966/Reviewer_CJEP"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper discusses the challenges of large language models (LLMs) in complex reasoning tasks, addressing their tendencies for generating plausible but inaccurate outputs and decreased performance in unconventional knowledge scenarios. Then it proposes the Hypotheses-to-Theories (HtT) framework as a solution, inspired by the scientific method, incorporating rule induction and deduction stages to reduce hallucinations and improve reasoning accuracy. Empirical tests with GPT on numerical and relational reasoning datasets demonstrated significant performance improvements over baseline methods, showcasing the potential of HtT to enhance LLMs’ reasoning capabilities while mitigating existing challenges.", "review_text": "The paper discusses the challenges of large language models (LLMs) in complex reasoning tasks, addressing their tendencies for generating plausible but inaccurate outputs and decreased performance in unconventional knowledge scenarios. Then it proposes the Hypotheses-to-Theories (HtT) framework as a solution, inspired by the scientific method, incorporating rule induction and deduction stages to reduce hallucinations and improve reasoning accuracy. Empirical tests with GPT on numerical and relational reasoning datasets demonstrated significant performance improvements over baseline methods, showcasing the potential of HtT to enhance LLMs’ reasoning capabilities while mitigating existing challenges.", "strengths": "1.\tThe idea that uses LLM to act as a rule learner is novel, distinguishing itself from previous methodologies that typically employ alternative strategies to mitigate hallucination or rely on symbolic methods for rule acquisition.\n2.\tSeveral innovative tricks pertaining to prompts have been introduced, effectively addressing intricate implementation details and enhancing the method’s practicality.\n3.\tThe method's effectiveness has been validated through experiments.", "weaknesses": "1.\tAlthough the method presented in the article exhibits a certain degree of innovation, its articulation fails to meet the standards of ICLR, leaving many details unaddressed within the paper. This omission results in confusion among readers trying to grasp the intricacies of the proposed approach. \n\nFor instance, in the \"Induction from Deduction\" section, it is not specified how the rules are extracted — is it through regular expressions? \n\nIt is also unclear how the occurrence k and accuracy p are calculated based on the paper. These concepts are borrowed from the field of rule learning, yet the author does not elucidate how they are applied in the domain of natural language. This transition from rule learning to natural language processing necessitates a clear explanation, as the methodologies and challenges inherent to these domains can be vastly different. \n\nIn the appendix, considering Prompt 2, it is evident that a substantial number of ground rules are already present within the prompt. This raises a question: If we rely solely on the rules from the prompt, what level of performance can be achieved? \n\n2.\tThe citation is inconsistency. For instance, the first referenced NeurIPS paper does not include page numbers, while the second one does. The third citation is missing its source of publication, and the fourth one includes the conference name’s abbreviation, unlike the others. \n\nXinyun Chen, Chen Liang, Adams Wei Yu, Dawn Song, and Denny Zhou. Compositional generalization via neural-symbolic stack machines. In Advances in Neural Information Processing Systems, 2020.\n\nTom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.\n\nAntonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large language models for interpretable logical reasoning. 2023.\n\nAdam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 5418–5426, 2020.\n\n3.\tThe paper falls short of providing a comprehensive description regarding the spectrum of problems that the introduced approach is adept at solving. This leaves a ambiguity as to whether the proposed method can effectively tackle all varieties of hallucination issues, a matter that necessitates further elucidation for a complete understanding of the method’s capabilities and limitations.\n\n4.\tFurthermore, is the method capable of learning complex rules, such as first-order logic rules? If unification is required during reasoning, can LLMs still utilize these rules for inference?", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper discusses the challenges of large language models (LLMs) in complex reasoning tasks, addressing their tendencies for generating plausible but inaccurate outputs and decreased performance in unconventional knowledge scenarios. Then it proposes the Hypotheses-to-Theories (HtT) framework as a solution, inspired by the scientific method, incorporating rule induction and deduction stages to reduce hallucinations and improve reasoning accuracy. Empirical tests with GPT on numerical and relational reasoning datasets demonstrated significant performance improvements over baseline methods, showcasing the potential of HtT to enhance LLMs’ reasoning capabilities while mitigating existing challenges.", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1.\tThe idea that uses LLM to act as a rule learner is novel, distinguishing itself from previous methodologies that typically employ alternative strategies to mitigate hallucination or rely on symbolic methods for rule acquisition.\n2.\tSeveral innovative tricks pertaining to prompts have been introduced, effectively addressing intricate implementation details and enhancing the method’s practicality.\n3.\tThe method's effectiveness has been validated through experiments.", "weaknesses": "1.\tAlthough the method presented in the article exhibits a certain degree of innovation, its articulation fails to meet the standards of ICLR, leaving many details unaddressed within the paper. This omission results in confusion among readers trying to grasp the intricacies of the proposed approach. \n\nFor instance, in the \"Induction from Deduction\" section, it is not specified how the rules are extracted — is it through regular expressions? \n\nIt is also unclear how the occurrence k and accuracy p are calculated based on the paper. These concepts are borrowed from the field of rule learning, yet the author does not elucidate how they are applied in the domain of natural language. This transition from rule learning to natural language processing necessitates a clear explanation, as the methodologies and challenges inherent to these domains can be vastly different. \n\nIn the appendix, considering Prompt 2, it is evident that a substantial number of ground rules are already present within the prompt. This raises a question: If we rely solely on the rules from the prompt, what level of performance can be achieved? \n\n2.\tThe citation is inconsistency. For instance, the first referenced NeurIPS paper does not include page numbers, while the second one does. The third citation is missing its source of publication, and the fourth one includes the conference name’s abbreviation, unlike the others. \n\nXinyun Chen, Chen Liang, Adams Wei Yu, Dawn Song, and Denny Zhou. Compositional generalization via neural-symbolic stack machines. In Advances in Neural Information Processing Systems, 2020.\n\nTom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.\n\nAntonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large language models for interpretable logical reasoning. 2023.\n\nAdam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 5418–5426, 2020.\n\n3.\tThe paper falls short of providing a comprehensive description regarding the spectrum of problems that the introduced approach is adept at solving. This leaves a ambiguity as to whether the proposed method can effectively tackle all varieties of hallucination issues, a matter that necessitates further elucidation for a complete understanding of the method’s capabilities and limitations.\n\n4.\tFurthermore, is the method capable of learning complex rules, such as first-order logic rules? If unification is required during reasoning, can LLMs still utilize these rules for inference?", "questions": "See Weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698807760760}, {"id": "54LgCBp9dN", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4966/Reviewer_eHJ8"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The key idea of the paper is  that LLMs can learn rules from examples, and then these rules can be used to deduce answers for other queries to the LLMs. The paper proposes an induction and deduction step. In the induction step, the model infers rules from exemplars, some of which are thrown-out, based on coverage (how often is the rule used) and accuracy (how often is the rule correct) of the rules. In the deduction step, the induced rules are given as a knowledge base, and the model is expected to use rules to infer new answers. \n\nThe authors test there idea on simple synthetic datasets, by extending the Chain of Thoughts (CoT) and Least to Most (LtM) prompting methodology, showing consistent advantage of using the proposed method. The datasets consist of simple task of arithmetic in different bases and learning (simple) kinship relationship rules.", "review_text": "The key idea of the paper is  that LLMs can learn rules from examples, and then these rules can be used to deduce answers for other queries to the LLMs. The paper proposes an induction and deduction step. In the induction step, the model infers rules from exemplars, some of which are thrown-out, based on coverage (how often is the rule used) and accuracy (how often is the rule correct) of the rules. In the deduction step, the induced rules are given as a knowledge base, and the model is expected to use rules to infer new answers. \n\nThe authors test there idea on simple synthetic datasets, by extending the Chain of Thoughts (CoT) and Least to Most (LtM) prompting methodology, showing consistent advantage of using the proposed method. The datasets consist of simple task of arithmetic in different bases and learning (simple) kinship relationship rules.", "strengths": "- Easy and intuitive prompting method\n- Shows that LLMs can learn simple rules \n- Can be useful when rules that need to be learnt are simple true/false propositions", "weaknesses": "The general idea of the paper is nice, but the developed setup is too simplistic, and has not ben tested in any realistic setting. Specifically, I have the following concerns:\n \n- The setting is too simple, the learnt rules are just true/false propositions of the form \"A is B\"\n- The examples as shown in appendix are not very impressive, at least from a skim through, it seems the rules already exist explicitly in the prompt text. At this point how is this different from just knowledge retrieval as done in [1]. In fact the tasks presented in [1] seems much more nuanced than the one presented here. \n- The gains without XML tagging (an existing method in the prompting technical know-how) are marginal. Furthermore, (it seems to me) that the authors have not tested CoT and LtM with XML tagging, making it unclear how much of their gains are from tagging, and how much is from the extracted knowledge.\n\n[1] Trivedi et al. Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions. https://arxiv.org/pdf/2212.10509.pdf", "questions": "- On page 4, you explain that you add XML tags to the prompts, Do you add the XML tags to the prompts of the methods compared?\n- The examples that you show in appendix seem very simple rules, are you able to extract complex rules? --- beyond \"A is B\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The key idea of the paper is  that LLMs can learn rules from examples, and then these rules can be used to deduce answers for other queries to the LLMs. The paper proposes an induction and deduction step. In the induction step, the model infers rules from exemplars, some of which are thrown-out, based on coverage (how often is the rule used) and accuracy (how often is the rule correct) of the rules. In the deduction step, the induced rules are given as a knowledge base, and the model is expected to use rules to infer new answers. \n\nThe authors test there idea on simple synthetic datasets, by extending the Chain of Thoughts (CoT) and Least to Most (LtM) prompting methodology, showing consistent advantage of using the proposed method. The datasets consist of simple task of arithmetic in different bases and learning (simple) kinship relationship rules.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "- Easy and intuitive prompting method\n- Shows that LLMs can learn simple rules \n- Can be useful when rules that need to be learnt are simple true/false propositions", "weaknesses": "The general idea of the paper is nice, but the developed setup is too simplistic, and has not ben tested in any realistic setting. Specifically, I have the following concerns:\n \n- The setting is too simple, the learnt rules are just true/false propositions of the form \"A is B\"\n- The examples as shown in appendix are not very impressive, at least from a skim through, it seems the rules already exist explicitly in the prompt text. At this point how is this different from just knowledge retrieval as done in [1]. In fact the tasks presented in [1] seems much more nuanced than the one presented here. \n- The gains without XML tagging (an existing method in the prompting technical know-how) are marginal. Furthermore, (it seems to me) that the authors have not tested CoT and LtM with XML tagging, making it unclear how much of their gains are from tagging, and how much is from the extracted knowledge.\n\n[1] Trivedi et al. Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions. https://arxiv.org/pdf/2212.10509.pdf", "questions": "- On page 4, you explain that you add XML tags to the prompts, Do you add the XML tags to the prompts of the methods compared?\n- The examples that you show in appendix seem very simple rules, are you able to extract complex rules? --- beyond \"A is B\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698669508458}, {"id": "uxxzdN6qCy", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4966/Reviewer_r3qq"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper tries to improve LLM's reasoning ability, by inducing rules and applying the induced rules in deductive problems. Experimental results show that explicitly learning some rules and inject them into prompts can significantly benefit strong LLM models such as GPT-4, but not weaker ones like GPT-3.5.", "review_text": "This paper tries to improve LLM's reasoning ability, by inducing rules and applying the induced rules in deductive problems. Experimental results show that explicitly learning some rules and inject them into prompts can significantly benefit strong LLM models such as GPT-4, but not weaker ones like GPT-3.5.", "strengths": "1. Extensive experiments to verify the effectiveness of the proposed method.\n2. The presentation of this paper is articulate, and easy to read.\n3. The experimental results verified the effectiveness of this work.", "weaknesses": "1. Many details are missing, e.g., how does the induction stage work? How to use ground truth answers to verify the induced rules, does it require human annotators? What is the confidence of rules, how are they evaluated, and does LLM output a confidence score associated with the rules?\n2. This work is basically a technical report, some claims lack supportive facts or resources. For example, the authors say \"hallucination in LLMs resembles hypothesis generation in scientific discovery\", which is incorrect. In scientific discovery, hypotheses are generated by logical induction and abduction ([reference](https://plato.stanford.edu/entries/scientific-discovery/)), while the logic behind hallucination remains unknown.\n3. Rules in the form of natural language weaken the generalisation ability and usually cause ambiguity and may confuse people. This is exactly the reason why Gottfried Leibniz calls for mathematical logic. However, most of the time, humans, like LLMs, are using natural language for reasoning, so I don't think the method proposed by this paper is bad. It would be interesting to make a more comprehensive test for the learned rules using formal methods, for example, ask LLMs to abstract those numerical rules in Appendix C into higher-order forms, such as Peano axioms and see if LLMs can make use of those more advanced rules.", "questions": "Please see my above comments.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tries to improve LLM's reasoning ability, by inducing rules and applying the induced rules in deductive problems. Experimental results show that explicitly learning some rules and inject them into prompts can significantly benefit strong LLM models such as GPT-4, but not weaker ones like GPT-3.5.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Extensive experiments to verify the effectiveness of the proposed method.\n2. The presentation of this paper is articulate, and easy to read.\n3. The experimental results verified the effectiveness of this work.", "weaknesses": "1. Many details are missing, e.g., how does the induction stage work? How to use ground truth answers to verify the induced rules, does it require human annotators? What is the confidence of rules, how are they evaluated, and does LLM output a confidence score associated with the rules?\n2. This work is basically a technical report, some claims lack supportive facts or resources. For example, the authors say \"hallucination in LLMs resembles hypothesis generation in scientific discovery\", which is incorrect. In scientific discovery, hypotheses are generated by logical induction and abduction ([reference](https://plato.stanford.edu/entries/scientific-discovery/)), while the logic behind hallucination remains unknown.\n3. Rules in the form of natural language weaken the generalisation ability and usually cause ambiguity and may confuse people. This is exactly the reason why Gottfried Leibniz calls for mathematical logic. However, most of the time, humans, like LLMs, are using natural language for reasoning, so I don't think the method proposed by this paper is bad. It would be interesting to make a more comprehensive test for the learned rules using formal methods, for example, ask LLMs to abstract those numerical rules in Appendix C into higher-order forms, such as Peano axioms and see if LLMs can make use of those more advanced rules.", "questions": "Please see my above comments.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698652829118}, {"id": "7Fv7T3Yb4Q", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4966/Reviewer_Zpi4"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces the \"Hypotheses-to-Theories\" (HtT) framework, which is designed to equip LLMs with a rule library for conducting reasoning. HtT comprises two distinct stages: an induction stage and a deduction stage. In the induction stage, the LLM is initially tasked with generating and validating rules based on a set of training examples. Rules that frequently appear and lead to correct answers are aggregated to create a rule library. In the deduction stage, the LLM is then prompted to leverage this acquired rule library to engage in reasoning and respond to test questions. The authors have evaluated their approach on a numerical reasoning benchmark and a relational reasoning benchmark and argue that their approach can significantly enhance the performance of existing few-shot prompting methods.", "review_text": "This paper introduces the \"Hypotheses-to-Theories\" (HtT) framework, which is designed to equip LLMs with a rule library for conducting reasoning. HtT comprises two distinct stages: an induction stage and a deduction stage. In the induction stage, the LLM is initially tasked with generating and validating rules based on a set of training examples. Rules that frequently appear and lead to correct answers are aggregated to create a rule library. In the deduction stage, the LLM is then prompted to leverage this acquired rule library to engage in reasoning and respond to test questions. The authors have evaluated their approach on a numerical reasoning benchmark and a relational reasoning benchmark and argue that their approach can significantly enhance the performance of existing few-shot prompting methods.", "strengths": "The study and development of reasoning capabilities in LLMs is a very interesting and topical research area. LLMs have already demonstrated emerging capabilities across a wide range of reasoning tasks, primarily due to the evolution of sophisticated prompting methodologies. This paper provides further insights in this direction. Furthermore, the results of the experimental study seem to support the effectiveness of the proposed approach on existing benchmark datasets.", "weaknesses": "In spite of the pertinence of the addressed problem and the promising results from the experiments, I feel that paper also comes with \nsignificant weaknesses, which I will detail next. Specifically, I feel that the proposed Hypotheses-to-Theories (HtT) framework \nlacks comprehensive development within this submission. The authors describe general ideas but do not provide sufficient technical details \n  describing how their methods advance the state-of-the-art. The core technical contribution of the paper\nis succinctly described in just one and a half pages (pages 3 and 4), while the majority of the paper \nis devoted to the description of the experiments and the obtained results.  Unfortunately, the description of the approach\n appears somewhat lacking in depth and places an undue emphasis on \"tricks\" like XML tagging, detracting attention from fundamental principles and methodologies that can be adopted and further developed by other researchers.\n\nAs an example of the paper lacking important detail, I feel that the concept of a \"rule\" remains ambiguously defined in the paper. Specifically, the paper lacks an explicit description of the nature of \"rules\" that can be incorporated into the library and subsequently used in the deduction phase. The examples presented in Figure 1 proved to be somewhat confusing; for instance, the rule library in Figure 1 includes statements like \"3 + 4 = 7,\" which, in my view, represent concrete facts rather than rules. A rule typically constitutes a formalised, general statement that applies to a potentially infinite collection of objects (e.g., \"the successor of an even natural number is an odd number\" or \"all men are mortal\"). In this context, the assertion that the proposed approach can induce \"rules\" appears unjustified and potentially misleading.\n\nSimilarly, it remains unclear how these \"rules\" are to be applied. Rule application, in its essence, involves the process of using a general statement to derive new facts from existing information (e.g., given that 4 is an even natural number and the established rule that the successor of an even number is odd, we can deduce the new fact that 5 is an odd number). It remains unclear how a \"rule\" such as \"3 + 4 = 7\" would be employed in a deductive context to generate new insights in the aforementioned sense.", "questions": "I do not have specific questions. I believe that the paper should be substantially rewritten before it can be published at a top venue. In particular, the description of the core approach should be substantially expanded, the contributions to science should be emphasised and the new techniques developed should be made explicit so that they can be adopted and further developed by other researchers in the field.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the \"Hypotheses-to-Theories\" (HtT) framework, which is designed to equip LLMs with a rule library for conducting reasoning. HtT comprises two distinct stages: an induction stage and a deduction stage. In the induction stage, the LLM is initially tasked with generating and validating rules based on a set of training examples. Rules that frequently appear and lead to correct answers are aggregated to create a rule library. In the deduction stage, the LLM is then prompted to leverage this acquired rule library to engage in reasoning and respond to test questions. The authors have evaluated their approach on a numerical reasoning benchmark and a relational reasoning benchmark and argue that their approach can significantly enhance the performance of existing few-shot prompting methods.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The study and development of reasoning capabilities in LLMs is a very interesting and topical research area. LLMs have already demonstrated emerging capabilities across a wide range of reasoning tasks, primarily due to the evolution of sophisticated prompting methodologies. This paper provides further insights in this direction. Furthermore, the results of the experimental study seem to support the effectiveness of the proposed approach on existing benchmark datasets.", "weaknesses": "In spite of the pertinence of the addressed problem and the promising results from the experiments, I feel that paper also comes with \nsignificant weaknesses, which I will detail next. Specifically, I feel that the proposed Hypotheses-to-Theories (HtT) framework \nlacks comprehensive development within this submission. The authors describe general ideas but do not provide sufficient technical details \n  describing how their methods advance the state-of-the-art. The core technical contribution of the paper\nis succinctly described in just one and a half pages (pages 3 and 4), while the majority of the paper \nis devoted to the description of the experiments and the obtained results.  Unfortunately, the description of the approach\n appears somewhat lacking in depth and places an undue emphasis on \"tricks\" like XML tagging, detracting attention from fundamental principles and methodologies that can be adopted and further developed by other researchers.\n\nAs an example of the paper lacking important detail, I feel that the concept of a \"rule\" remains ambiguously defined in the paper. Specifically, the paper lacks an explicit description of the nature of \"rules\" that can be incorporated into the library and subsequently used in the deduction phase. The examples presented in Figure 1 proved to be somewhat confusing; for instance, the rule library in Figure 1 includes statements like \"3 + 4 = 7,\" which, in my view, represent concrete facts rather than rules. A rule typically constitutes a formalised, general statement that applies to a potentially infinite collection of objects (e.g., \"the successor of an even natural number is an odd number\" or \"all men are mortal\"). In this context, the assertion that the proposed approach can induce \"rules\" appears unjustified and potentially misleading.\n\nSimilarly, it remains unclear how these \"rules\" are to be applied. Rule application, in its essence, involves the process of using a general statement to derive new facts from existing information (e.g., given that 4 is an even natural number and the established rule that the successor of an even number is odd, we can deduce the new fact that 5 is an odd number). It remains unclear how a \"rule\" such as \"3 + 4 = 7\" would be employed in a deductive context to generate new insights in the aforementioned sense.", "questions": "I do not have specific questions. I believe that the paper should be substantially rewritten before it can be published at a top venue. In particular, the description of the core approach should be substantially expanded, the contributions to science should be emphasised and the new techniques developed should be made explicit so that they can be adopted and further developed by other researchers in the field.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697194853791}], "openreview_url": "https://openreview.net/forum?id=tAmfM1sORP", "arxiv_id": "2310.07064", "paper_pdf": "papers/tAmfM1sORP.pdf", "paper_pdf_sha256": "cbcab388152cd96c31c8f6ea2b79f1966bef41583f08ee6ff5b33cdb222345b9", "paper_pdf_bytes": 423040, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-deepmind/llms_can_learn_rules", "code_repository": "google-deepmind/llms_can_learn_rules", "code_commit": "5c2c523690720a314932dcd8f77c321a3bd9c1f8", "code_archive": "repos/tAmfM1sORP.zip", "code_archive_sha256": "3eeb7d588e9298e9c126088986ea527b9ad2ba7c383a91a8a6ff16c204bbc08c", "code_archive_bytes": 265211, "code_file_count": 7, "code_extensions": {".py": 6, ".sh": 1}, "github_disk_usage_kb": 262, "github_languages": {"Python": 30326, "Shell": 1770}, "github_archived": false, "github_pushed_at": "2024-12-06T01:54:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/large-language-models-can-learn-rules"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7PURWDjJCf3", "year": 2023, "status": "rejected", "title": "Slimmable Networks for Contrastive Self-supervised Learning", "authors": ["Shuai Zhao", "Xiaohan Wang", "Linchao Zhu", "Yi Yang"], "authorids": ["~Shuai_Zhao1", "~Xiaohan_Wang2", "~Linchao_Zhu1", "~Yi_Yang22"], "authors_source": "OpenReview API", "abstract": "Self-supervised learning makes great progress in large model pre-training but suffers in training small models. Previous solutions to this problem mainly rely on knowledge distillation and indeed have a two-stage learning procedure: first train a large teacher model, then distill it to improve the generalization ability of small ones. In this work, we present a new one-stage solution to obtain pre-trained small models without extra teachers: slimmable networks for contrastive self-supervised learning (SlimCLR). A slimmable network contains a full network and several weight-sharing sub-networks. We can pre-train for only one time and obtain various networks including small ones with low computation costs. However, in self-supervised cases, the interference between weight-sharing networks leads to severe performance degradation. One evidence of the interference is gradient imbalance: a small proportion of parameters produces dominant gradients during backpropagation, and the main parameters may not be fully optimized. The divergence in gradient directions of various networks may also cause interference between networks. To overcome these problems, we make the main parameters produce dominant gradients and provide consistent guidance for sub-networks via three techniques: slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Besides, a switchable linear probe layer is applied during linear evaluation to avoid the interference of weight-sharing linear layers. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "t5d7dWxrim", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper130/Reviewer_yYC4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposed a slimmable contrastive self-supervised learning framework for building small models with SimCLR. It applied the well-known slimmable network and solved the gradient imbalance problem in the training by slow start training, online distillation, and loss reweighting, etc.\n", "review_text": "The paper studied an interesting topic and solved it by applying an existing approach with some minor improvements on the training loss. The evaluation is not sufficient to justify the motivation and contributions.\n", "strengths": "1. The proposed approach achieved balanced performance between multiple architecture sizes and outperformed the other small SSL models.\n2. The paper is well written and easy to follow.\n\nWeakness\n1. The novelty is limited as this paper simply applies the slimmable architecture with some minor improvements on the training loss.\n2. Given the paper target on training a small SSL model, it would be interesting to compare the performance with the small distilled model in terms of accuracy, model size, FLOPs. The current evaluation is not sufficient to justify the motivation.\n3. The evaluation can be improved by evaluating with other downstream datasets and tasks.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposed a slimmable contrastive self-supervised learning framework for building small models with SimCLR. It applied the well-known slimmable network and solved the gradient imbalance problem in the training by slow start training, online distillation, and loss reweighting, etc.\n", "strength_and_weaknesses": "1. The proposed approach achieved balanced performance between multiple architecture sizes and outperformed the other small SSL models.\n2. The paper is well written and easy to follow.\n\nWeakness\n1. The novelty is limited as this paper simply applies the slimmable architecture with some minor improvements on the training loss.\n2. Given the paper target on training a small SSL model, it would be interesting to compare the performance with the small distilled model in terms of accuracy, model size, FLOPs. The current evaluation is not sufficient to justify the motivation.\n3. The evaluation can be improved by evaluating with other downstream datasets and tasks.\n", "clarity,_quality,_novelty_and_reproducibility": "1. The paper is well written and easy to read.\n\n2. The novelty is limited since the paper mainly applied the well-studied slimmable network.\n\n3. The proposed approach is reproducible.\n", "summary_of_the_review": "The paper studied an interesting topic and solved it by applying an existing approach with some minor improvements on the training loss. The evaluation is not sufficient to justify the motivation and contributions.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667460640877}, {"id": "k4atcdIRCkI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper130/Reviewer_enLj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Motivated by the performance gap between self-supervised learning and its counter supervised model, this work provides a one-stage self-supervised small model pretraining protocol. The slimmable network idea is used to get the representations of one augmentation $x_1$. Weights are shared between the set of slimmable encoders (the small one shared with the larger one). All the outputs from slimmable encoders are used as anchors in InfoNCE. The sum of loss from different anchors is calculated eventually. However, this naive implementation will cause gradient imbalance. The work proposed updating the full encoder first, knowledge distillation from the full encoder to the sub-network, and loss reweighting to mitigate the imbalance issue.  ", "review_text": "Overall this is an interesting work even though the idea of slammable is not new and no theory is provided. I hope the author could address my questions in the rebuttal. I’m wondering whether a similar range of performance drops will happen on R-18 like on R-50. ", "strengths": "An interesting work and clear motivation. \n\nHowever, I am not convinced about the gradient imbalance in learning slimmable networks. The measurement used in the paper seems not sufficient to me. Imagine that for each iteration the network weights are updated in a cyclic way, i.e., weights at (1,2,3) are updated in the first iteration, then (4,5,6), (7,8,9), (10,11,12) .... In this way the gradient norm ratio will be low as well, but the network can still be trained reasonably. Another question is since the paper mentioned Res18 in the motivation, is there a comparison of R-18 compared with R-18 distilled by R-50 teacher? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "Motivated by the performance gap between self-supervised learning and its counter supervised model, this work provides a one-stage self-supervised small model pretraining protocol. The slimmable network idea is used to get the representations of one augmentation $x_1$. Weights are shared between the set of slimmable encoders (the small one shared with the larger one). All the outputs from slimmable encoders are used as anchors in InfoNCE. The sum of loss from different anchors is calculated eventually. However, this naive implementation will cause gradient imbalance. The work proposed updating the full encoder first, knowledge distillation from the full encoder to the sub-network, and loss reweighting to mitigate the imbalance issue.  ", "strength_and_weaknesses": "An interesting work and clear motivation. \n\nHowever, I am not convinced about the gradient imbalance in learning slimmable networks. The measurement used in the paper seems not sufficient to me. Imagine that for each iteration the network weights are updated in a cyclic way, i.e., weights at (1,2,3) are updated in the first iteration, then (4,5,6), (7,8,9), (10,11,12) .... In this way the gradient norm ratio will be low as well, but the network can still be trained reasonably. Another question is since the paper mentioned Res18 in the motivation, is there a comparison of R-18 compared with R-18 distilled by R-50 teacher? ", "clarity,_quality,_novelty_and_reproducibility": "Clear, and well-written. But the novelty seems limited to me, with no theory, and the hypothesis seems not convincing. Reproducibility should be ok.", "summary_of_the_review": "Overall this is an interesting work even though the idea of slammable is not new and no theory is provided. I hope the author could address my questions in the rebuttal. I’m wondering whether a similar range of performance drops will happen on R-18 like on R-50. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666667895509}, {"id": "rq9Y9b-h7_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper130/Reviewer_1XPr"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies slimmable networks under the contrastive self-supervised learning (SlimCLR) setting. It proposes some strategies to solve the interference between weight-sharing sub-networks during training, including slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes, and a switchable linear probe layer. SlimCLR is evaluated on ImageNet based on ResNet-50 using MoCo v2 and v3 frameworks.", "review_text": "Please refer to the above.", "strengths": "### Strength\n\nThis paper is a good study of the effect of slimmable networks for contrastive self-supervised learning. The empirical analysis has some contribution to the community.\n\n### Weakness\n\n**1: The technical novelty is limited and some details are confusing.**\n\nSlimmable networks are a special case of widely studied one-shot NAS (e.g., [R2, R3, R4, R5, R6]), which only considers the width dimension (see discussion in OFA [R2]). There are many techniques to deal with interference among subnetworks. Specifically,\n* “Slow start” belongs to progressive training in one-shot NAS. For example, OFA proposes a “Progressive Shrinking” strategy, which starts with training the largest sub-network and then progressively fine-tunes the network to support smaller sub-networks by gradually adding them into the sampling space.\n\n * “Online distillation” was originally proposed in US-Nets [Yu et al., 2019b]. Apart from the inplace distillation, it also proposes the sandwich rule.\n\n* “loss reweights” aims to assign larger weights for sub-networks with large widths. However, it violates the training objective of slimmable networks. The objective is to make each supported sub-network maintain the same level of accuracy as independently training a network with the same architectural configuration, rather than only training an accurate large “supernet”. This is evidenced in Table 2 (e), where adding loss reweighting makes R-50(0.25) perform worse, so what’s the meaning there?\n\n**2:  Another concern is what are the fundamental differences between self-supervised and supervised training for slimmable networks?** This is not clear to me as all the training techniques used are common practices in supervised training.\n\n**3: What’s the relationship between unsupervised NAS (e.g., [R4, R5]), including the contrastive self-supervised one (e.g., [R6])?** \n\n4: In Page 8, the authors study 4 possible cases of loss reweighting and show the results in Table 2e. However, I find case (3) archives the best performance for most widths but the paper uses case (1) by default in Eq. (5). I disagree with the author's explanation that “To ensure the performance of the smallest network, we adopt the reweighting manner (1) in practice” as all sub-networks with different widths should be equally important. Otherwise, what’s the meaning of slimmable networks there? \n\t\t\t\t\n**5: The paper lacks mathematical modeling for the gradient divergence issue which leads to the optimization difficulty claimed by the authors.** I think there are only four possible widths and it is not difficult to analyze the gradient magnitude and directions using SGD with maths formulations. Also, some theoretical analysis on convergence is expected, even assuming a linear neural network is fine [R1].\n\n**6: The experiments are far from enough to justify the effectiveness of the proposed method.**\n\n* 6.1:  The results are merely based on the ResNet-50 backbone. However, I would like to see more ResNet backbones such as R-101 and R-152. More importantly, experiments on Vision Transformers, such as ViT-B in MoCo v3, must be included in the experiments.  \n\n* 6.2:  The paper only evaluates the representation quality using linear probing. However, it must evaluate transfer learning performance which is the standard practice in self-supervised learning (e.g., in MoCo v3). For example, experiments on dataset transfer and downstream tasks such as dense segmentation and detection on COCO and ADE20k are needed.  \n\n* 6.3: How about training the whole network (width 1.0) first then using network pruning (e.g., [R7]) to obtain small networks (width 0.25, 0.5, 0.75)? As this strategy can avoid the interference issue during training. \n\n* 6.4: It lacks comparisons with methods dealing with sub-network interference, such as switchable BN [Yu et al., 2019], sandwich rules [Yu et al., 2019b] and many others.\n\n**7:  The discussions and references in related work are far from enough.** There are few discussions with single-shot NAS and unsupervised NAS methods. In addition, as I point out in the technical novelty part, the differences and advantages with the related work must be discussed. \n\n**8:  Writing also needs to be improved.** \n\n* 8.1: What is the definition of the “main parameters” in the introduction? \n\n* 8.2: In Sec. 3.2, “..., where $L$ is the loss function”. It should be defined in Eq. (1) where it first appears.\n\n* 8.3: Many grammar issues. I only point out a few. “Slimmable neworks” in Sec. 2; “server performance degradation” in Sec. 3.2. \n\n9: In Sec. 3.2, authors argue that the two ratios in Figure 3 should be large enough. “In Figure 3f, ..., are larger than 1.0 by a clear margin”. It does provide a clear concept of how large is good enough. In my opinion, it also depends on the network architectures and self-supervised learning frameworks. So Figure 3 may not be statistically significant. \n\n\n**References:**\n\n[R1]: “On the optimization of Deep Networks: Implicit Acceleration by Overparameterization”, ICML 2018\n\n[R2]: “ONCE FOR ALL: TRAIN ONE NETWORK AND SPECIALIZE IT FOR EFFICIENT DEPLOYMENT”, ICLR 2020\n\n[R3]: “BigNAS: Scaling Up Neural Architecture Search with Big Single-Stage Models”, ECCV 2020\n\n[R4]: “Are Labels Necessary for Neural Architecture Search?”, ECCV 2020\n\n[R5]: “Does Unsupervised Architecture Representation Learning Help Neural Architecture Search?”, NeurIPS 2020\n\n[R6]: “Contrastive Self-supervised Neural Architecture Search”, Arxiv 2021\n\n[R7]: “Resrep: Lossless cnn pruning via decoupling remembering and forgetting”, CVPR 2022", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper studies slimmable networks under the contrastive self-supervised learning (SlimCLR) setting. It proposes some strategies to solve the interference between weight-sharing sub-networks during training, including slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes, and a switchable linear probe layer. SlimCLR is evaluated on ImageNet based on ResNet-50 using MoCo v2 and v3 frameworks.", "strength_and_weaknesses": "### Strength\n\nThis paper is a good study of the effect of slimmable networks for contrastive self-supervised learning. The empirical analysis has some contribution to the community.\n\n### Weakness\n\n**1: The technical novelty is limited and some details are confusing.**\n\nSlimmable networks are a special case of widely studied one-shot NAS (e.g., [R2, R3, R4, R5, R6]), which only considers the width dimension (see discussion in OFA [R2]). There are many techniques to deal with interference among subnetworks. Specifically,\n* “Slow start” belongs to progressive training in one-shot NAS. For example, OFA proposes a “Progressive Shrinking” strategy, which starts with training the largest sub-network and then progressively fine-tunes the network to support smaller sub-networks by gradually adding them into the sampling space.\n\n * “Online distillation” was originally proposed in US-Nets [Yu et al., 2019b]. Apart from the inplace distillation, it also proposes the sandwich rule.\n\n* “loss reweights” aims to assign larger weights for sub-networks with large widths. However, it violates the training objective of slimmable networks. The objective is to make each supported sub-network maintain the same level of accuracy as independently training a network with the same architectural configuration, rather than only training an accurate large “supernet”. This is evidenced in Table 2 (e), where adding loss reweighting makes R-50(0.25) perform worse, so what’s the meaning there?\n\n**2:  Another concern is what are the fundamental differences between self-supervised and supervised training for slimmable networks?** This is not clear to me as all the training techniques used are common practices in supervised training.\n\n**3: What’s the relationship between unsupervised NAS (e.g., [R4, R5]), including the contrastive self-supervised one (e.g., [R6])?** \n\n4: In Page 8, the authors study 4 possible cases of loss reweighting and show the results in Table 2e. However, I find case (3) archives the best performance for most widths but the paper uses case (1) by default in Eq. (5). I disagree with the author's explanation that “To ensure the performance of the smallest network, we adopt the reweighting manner (1) in practice” as all sub-networks with different widths should be equally important. Otherwise, what’s the meaning of slimmable networks there? \n\t\t\t\t\n**5: The paper lacks mathematical modeling for the gradient divergence issue which leads to the optimization difficulty claimed by the authors.** I think there are only four possible widths and it is not difficult to analyze the gradient magnitude and directions using SGD with maths formulations. Also, some theoretical analysis on convergence is expected, even assuming a linear neural network is fine [R1].\n\n**6: The experiments are far from enough to justify the effectiveness of the proposed method.**\n\n* 6.1:  The results are merely based on the ResNet-50 backbone. However, I would like to see more ResNet backbones such as R-101 and R-152. More importantly, experiments on Vision Transformers, such as ViT-B in MoCo v3, must be included in the experiments.  \n\n* 6.2:  The paper only evaluates the representation quality using linear probing. However, it must evaluate transfer learning performance which is the standard practice in self-supervised learning (e.g., in MoCo v3). For example, experiments on dataset transfer and downstream tasks such as dense segmentation and detection on COCO and ADE20k are needed.  \n\n* 6.3: How about training the whole network (width 1.0) first then using network pruning (e.g., [R7]) to obtain small networks (width 0.25, 0.5, 0.75)? As this strategy can avoid the interference issue during training. \n\n* 6.4: It lacks comparisons with methods dealing with sub-network interference, such as switchable BN [Yu et al., 2019], sandwich rules [Yu et al., 2019b] and many others.\n\n**7:  The discussions and references in related work are far from enough.** There are few discussions with single-shot NAS and unsupervised NAS methods. In addition, as I point out in the technical novelty part, the differences and advantages with the related work must be discussed. \n\n**8:  Writing also needs to be improved.** \n\n* 8.1: What is the definition of the “main parameters” in the introduction? \n\n* 8.2: In Sec. 3.2, “..., where $L$ is the loss function”. It should be defined in Eq. (1) where it first appears.\n\n* 8.3: Many grammar issues. I only point out a few. “Slimmable neworks” in Sec. 2; “server performance degradation” in Sec. 3.2. \n\n9: In Sec. 3.2, authors argue that the two ratios in Figure 3 should be large enough. “In Figure 3f, ..., are larger than 1.0 by a clear margin”. It does provide a clear concept of how large is good enough. In my opinion, it also depends on the network architectures and self-supervised learning frameworks. So Figure 3 may not be statistically significant. \n\n\n**References:**\n\n[R1]: “On the optimization of Deep Networks: Implicit Acceleration by Overparameterization”, ICML 2018\n\n[R2]: “ONCE FOR ALL: TRAIN ONE NETWORK AND SPECIALIZE IT FOR EFFICIENT DEPLOYMENT”, ICLR 2020\n\n[R3]: “BigNAS: Scaling Up Neural Architecture Search with Big Single-Stage Models”, ECCV 2020\n\n[R4]: “Are Labels Necessary for Neural Architecture Search?”, ECCV 2020\n\n[R5]: “Does Unsupervised Architecture Representation Learning Help Neural Architecture Search?”, NeurIPS 2020\n\n[R6]: “Contrastive Self-supervised Neural Architecture Search”, Arxiv 2021\n\n[R7]: “Resrep: Lossless cnn pruning via decoupling remembering and forgetting”, CVPR 2022", "clarity,_quality,_novelty_and_reproducibility": "From the technical part, the paper lacks novelty and I don’t see techniques customized to train slimmable networks in the self-supervised setting. From the experimental perspective, it lacks essential empirical studies. Overall, the paper is clearly below the acceptance threshold.", "summary_of_the_review": "Please refer to the above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666500920715}, {"id": "cO0twSNKy3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper130/Reviewer_5mjb"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to learn slimmable neural networks with contrastive self-supervised learning without labels. To further improve the performance of sub-networks, this paper suggests using (i) slow-start training for sub-networks, (ii) online distillation, and (iii) loss reweighting. This increases the gradient norm of the main parameters, leading to performance improvements in the main network. This paper demonstrates that the main and sub-networks outperform existing two-stage distillation approaches.\n", "review_text": "While learning slimmable neural networks is interesting, I feel the lack of methodological novelty and the weak motivation about gradient imbalance. In addition, this paper is hard to follow with confused notations. Hence I vote for rejection.\n", "strengths": "Strengths\n- I think the main strength of this paper is that the proposed method outperforms other distillation approaches.\n\nWeaknesses\n- I feel the lack of methodological novelty. The idea of training slimmable neural networks is not new, and online distillation was already used in [1]. I also think other techniques are just engineering, and they have often been utilized in other literature: e.g., loss reweighting is a common strategy for multi-task learning.\n- Although this paper obtains some gains from the proposed techniques, the gains are marginal, and most of the gains come from using a large number of training epochs or MoCo-v3, not the techniques.\n- The training cost is linearly increasing with respect to the number of sub-networks. This could be critical since self-supervised learning is often time-consuming.\n- The motivation, the gradient imbalance, is really critical? I think the motivation is somewhat weak: the gradient imbalance can occur in any neural network layer because there exist dominant and minor neurons in any layer (e.g., think about singular value decomposition of weight matrix). Training sub-networks can be considered as determining the order of neuron importance. Note that the result in Fig (3a) is due to the random order of the neurons.\n  - Could you provide the distribution of singular values of the full weight matrix instead of gradient norms? I think it would be better to understand how training slimmable networks affect the weight matrix.\n- This paper is hard to follow: there are many confusing notations and descriptions.\n  - It would be better to use colors in Figure 1. In addition, it would be better to include the results of the proposed method in Figure 1 for readers.\n  - In general, notations should be defined before using them. For example, there is no description of ξ when using the parameter.\n  - Suggest to use $\\Theta_{w_i}=\\\\{\\theta_{w_1},\\ldots,\\theta_{w_i\\}\\\\}$ instead of writing $\\Theta_{w_j}\\subset\\Theta_{w_i}$ if $w_j<w_i$.\n  - In Eq (1), why $\\xi_1$ instead of $\\xi_i$? There is no explanation for this.\n  - In the self-supervised learning literature, the first MLP is often referred to as projection and the second MLP as prediction. I recommend following the conventional terminologies. For example, SlimCLR-MoCo-v2 should use projection instead of prediction. The current usage causes some confusion.\n  - What is $\\theta_{1.0\\setminus0.25}$?\n  - What are the main and minor parameters? And why are they main and minor?\n\n[1] Yu & Huang, Universally Slimmable Networks and Improved Training Techniques, 2019\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper aims to learn slimmable neural networks with contrastive self-supervised learning without labels. To further improve the performance of sub-networks, this paper suggests using (i) slow-start training for sub-networks, (ii) online distillation, and (iii) loss reweighting. This increases the gradient norm of the main parameters, leading to performance improvements in the main network. This paper demonstrates that the main and sub-networks outperform existing two-stage distillation approaches.\n", "strength_and_weaknesses": "Strengths\n- I think the main strength of this paper is that the proposed method outperforms other distillation approaches.\n\nWeaknesses\n- I feel the lack of methodological novelty. The idea of training slimmable neural networks is not new, and online distillation was already used in [1]. I also think other techniques are just engineering, and they have often been utilized in other literature: e.g., loss reweighting is a common strategy for multi-task learning.\n- Although this paper obtains some gains from the proposed techniques, the gains are marginal, and most of the gains come from using a large number of training epochs or MoCo-v3, not the techniques.\n- The training cost is linearly increasing with respect to the number of sub-networks. This could be critical since self-supervised learning is often time-consuming.\n- The motivation, the gradient imbalance, is really critical? I think the motivation is somewhat weak: the gradient imbalance can occur in any neural network layer because there exist dominant and minor neurons in any layer (e.g., think about singular value decomposition of weight matrix). Training sub-networks can be considered as determining the order of neuron importance. Note that the result in Fig (3a) is due to the random order of the neurons.\n  - Could you provide the distribution of singular values of the full weight matrix instead of gradient norms? I think it would be better to understand how training slimmable networks affect the weight matrix.\n- This paper is hard to follow: there are many confusing notations and descriptions.\n  - It would be better to use colors in Figure 1. In addition, it would be better to include the results of the proposed method in Figure 1 for readers.\n  - In general, notations should be defined before using them. For example, there is no description of ξ when using the parameter.\n  - Suggest to use $\\Theta_{w_i}=\\\\{\\theta_{w_1},\\ldots,\\theta_{w_i\\}\\\\}$ instead of writing $\\Theta_{w_j}\\subset\\Theta_{w_i}$ if $w_j<w_i$.\n  - In Eq (1), why $\\xi_1$ instead of $\\xi_i$? There is no explanation for this.\n  - In the self-supervised learning literature, the first MLP is often referred to as projection and the second MLP as prediction. I recommend following the conventional terminologies. For example, SlimCLR-MoCo-v2 should use projection instead of prediction. The current usage causes some confusion.\n  - What is $\\theta_{1.0\\setminus0.25}$?\n  - What are the main and minor parameters? And why are they main and minor?\n\n[1] Yu & Huang, Universally Slimmable Networks and Improved Training Techniques, 2019\n", "clarity,_quality,_novelty_and_reproducibility": "The detailed comments are described in the previous section. In summary,\n- Clarity :: This proposed idea is clear, but its description is hard to follow in general.\n- Quality :: The empirical results are somewhat strong, but it is not clear where the performance gains come from (proposed techniques or MoCo-v3).\n- Novelty :: This paper lacks methodological novelty.\n- Reproducibility :: This paper describes the implementation details well.\n", "summary_of_the_review": "While learning slimmable neural networks is interesting, I feel the lack of methodological novelty and the weak motivation about gradient imbalance. In addition, this paper is hard to follow with confused notations. Hence I vote for rejection.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666355537954}], "openreview_url": "https://openreview.net/forum?id=7PURWDjJCf3", "arxiv_id": "2209.15525", "paper_pdf": "papers/7PURWDjJCf3.pdf", "paper_pdf_sha256": "768457d3d4c79d1c6cb0934192c3d524f8ad73eea3bf0d7fb95e7291351d8085", "paper_pdf_bytes": 2204918, "paper_pdf_source": "openreview", "code_url": "https://github.com/mzhaoshuai/SlimCLR", "code_repository": "mzhaoshuai/SlimCLR", "code_commit": "d975c2ab3fb1aaa0c3dbe2520a09ed449970d5cd", "code_archive": "repos/7PURWDjJCf3.zip", "code_archive_sha256": "966816b32815780d66e617f915c1af091753185541fb30c529b979f78f2a4696", "code_archive_bytes": 629787, "code_file_count": 155, "code_extensions": {".py": 146, ".sh": 9}, "github_disk_usage_kb": 582, "github_languages": {"Python": 657129, "Shell": 20640}, "github_archived": false, "github_pushed_at": "2025-11-18T04:07:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/slimmable-networks-for-contrastive-self"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Qu_XudmGajz", "year": 2022, "status": "rejected", "title": "Structured Uncertainty in the Observation Space of Variational Autoencoders", "authors": ["James Langley", "Miguel Monteiro", "Charles Jones", "Nick Pawlowski", "Ben Glocker"], "authorids": ["~James_Langley1", "~Miguel_Monteiro1", "~Charles_Jones4", "~Nick_Pawlowski2", "~Ben_Glocker1"], "authors_source": "OpenReview API", "abstract": "Variational autoencoders (VAEs) are a popular class of deep generative models with many variants and a wide range of applications. Improvements upon the standard VAE mostly focus on the modelling of the posterior distribution over the latent space and the properties of the neural network decoder. In contrast, improving the model for the observational distribution is rarely considered and typically defaults to a pixel-wise independent categorical or normal distribution. In image synthesis, sampling from such distributions produces spatially-incoherent results with uncorrelated pixel noise, resulting in only the sample mean being somewhat useful as an output prediction. In this paper, we aim to stay true to VAE theory by improving the samples from the observational distribution. We propose an alternative model for the observation space, encoding spatial dependencies via a low-rank parameterization. We demonstrate that this new observational distribution has the ability to capture relevant covariance between pixels, resulting in spatially-coherent samples. In contrast to pixel-wise independent distributions, our samples seem to contain semantically meaningful variations from the mean allowing the prediction of multiple plausible outputs with a single forward pass.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Dq9wbfEFeWD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2681/Reviewer_2FaS"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to include some structure in the observation model of the standard VAE, typically implemented with a normal distribution with diagonal covariance matrix. The authors follow the work in [Monteiro 2020] which include low rank structure structure in the covariance matrix for the observation model in another generative framework.\n", "review_text": "Strengths\n- The authors provide an adequate introduction and summary of the literature, and highlight how their proposed method could be plugged into existing models.\n\nWeaknesses:\n- The approach is not radically novel and the paper could improve further by providing all details to allow the reader to fully reproduce their results. This could be included in an extended supplementary section.\n- Another part that could greatly improve the paper is some analysis of the scale of the variables. \n- The paper would greatly benefit of analysis showing how the approach extended into a more complex VAE framework further improve any downstream task. \n- How should the reader interpret the scales of the Lagrangian multiplier and the slack variable in Eq (2)? In section 4.1 these values are reported for (2) datasets. Adding additional analysis/explanation for these variables would benefit the paper and help persuade any reader about the robustness of  the results presented.\n- Figure 2 would also benefit from further explanation to better understand how these results support the claim of the authors.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to include some structure in the observation model of the standard VAE, typically implemented with a normal distribution with diagonal covariance matrix. The authors follow the work in [Monteiro 2020] which include low rank structure structure in the covariance matrix for the observation model in another generative framework.\n", "main_review": "Strengths\n- The authors provide an adequate introduction and summary of the literature, and highlight how their proposed method could be plugged into existing models.\n\nWeaknesses:\n- The approach is not radically novel and the paper could improve further by providing all details to allow the reader to fully reproduce their results. This could be included in an extended supplementary section.\n- Another part that could greatly improve the paper is some analysis of the scale of the variables. \n- The paper would greatly benefit of analysis showing how the approach extended into a more complex VAE framework further improve any downstream task. \n- How should the reader interpret the scales of the Lagrangian multiplier and the slack variable in Eq (2)? In section 4.1 these values are reported for (2) datasets. Adding additional analysis/explanation for these variables would benefit the paper and help persuade any reader about the robustness of  the results presented.\n- Figure 2 would also benefit from further explanation to better understand how these results support the claim of the authors.", "summary_of_the_review": "\nThe paper provides an extension of the work in [Monteiro et al. 2020] to the case of VAEs. The paper provides adequate references but it would greatly benefit for more downstream experiments and detailed explanation as suggested in the main review.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636386635925}, {"id": "wU3wSDRlVij", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2681/Reviewer_cypB"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors aim at improving the canonical VAE model by replacing the standard iid Gaussian likelihood with a multivariate Gaussian with (low-rank + diagonal) covariance.\n\nIn applications to CelebA and a brain MRI dataset from UK Biobank, the authors compare the proposed structured-observation-space VAE with a canonical VAE, showing that the samples from their model have lower Fréchet Inception Distance (FID) scores than both means and samples from the canonical VAE. The authors then evaluate the expressiveness of the representations learned in the observation space by visually evaluating interpolations in the observation space, and images obtained by rescaling the principal components of the observational covariance matrix. Finally, the authors show how their model can be used for interactive editing—i.e., editing a small number of pixels and using the conditional distribution to infer coherent edits in the remainder of the image.", "review_text": "The problem the authors tackle in this paper and the proposed solution are interesting. The submission is technically sound and the paper is well written.\n\nHowever, I have some major concerns about the originality of the work and feel that further comparison with other methods would be needed to show improvements over the state-of-the-art methods. Specifically:\n1) this work is closely related to Monteiro et al. (2020), who used a similar model in a segmentation setting. Is the model proposed here the same considered in Monteiro et al? If not, the authors should highlight methodological differences. If yes, the authors should highlight the conceptual contributions of their work.\n2) Dorta et al. (2018) considered a closely-related model, but making a different modeling choice for the covariance matrix. The authors should compare their results with those obtained with the model from Dorta et al (e.g., comparison of FID scores of generated samples).\n\nOther questions / concerns:\n- I find it hard to judge whether the means from the structured-observation-space VAE look visually more realistic than the mean from the canonical VAE. Can the author comment on this? Also, does not \"fixing the covariance diagonal to a small positive scalar\" artifactual reduce independent noise in the samples? I wonder if a visual comparison with canonical VAE is fair in this setting.\n- Are there other potential use cases for the observational model beyond pixel editing?\n- I think another way to improve upon the canonical VAE likelihood is to consider an iid gaussian likelihood on perceptual features [1]. The authors should consider this method in the related method section and potentially include this technique in the comparison.\n\n[1] Hou X, Shen L, Sun K, Qiu G. Deep feature consistent variational autoencoder. In2017 IEEE Winter Conference on Applications of Computer Vision (WACV) 2017 Mar 24 (pp. 1133-1141). IEEE.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors aim at improving the canonical VAE model by replacing the standard iid Gaussian likelihood with a multivariate Gaussian with (low-rank + diagonal) covariance.\n\nIn applications to CelebA and a brain MRI dataset from UK Biobank, the authors compare the proposed structured-observation-space VAE with a canonical VAE, showing that the samples from their model have lower Fréchet Inception Distance (FID) scores than both means and samples from the canonical VAE. The authors then evaluate the expressiveness of the representations learned in the observation space by visually evaluating interpolations in the observation space, and images obtained by rescaling the principal components of the observational covariance matrix. Finally, the authors show how their model can be used for interactive editing—i.e., editing a small number of pixels and using the conditional distribution to infer coherent edits in the remainder of the image.", "main_review": "The problem the authors tackle in this paper and the proposed solution are interesting. The submission is technically sound and the paper is well written.\n\nHowever, I have some major concerns about the originality of the work and feel that further comparison with other methods would be needed to show improvements over the state-of-the-art methods. Specifically:\n1) this work is closely related to Monteiro et al. (2020), who used a similar model in a segmentation setting. Is the model proposed here the same considered in Monteiro et al? If not, the authors should highlight methodological differences. If yes, the authors should highlight the conceptual contributions of their work.\n2) Dorta et al. (2018) considered a closely-related model, but making a different modeling choice for the covariance matrix. The authors should compare their results with those obtained with the model from Dorta et al (e.g., comparison of FID scores of generated samples).\n\nOther questions / concerns:\n- I find it hard to judge whether the means from the structured-observation-space VAE look visually more realistic than the mean from the canonical VAE. Can the author comment on this? Also, does not \"fixing the covariance diagonal to a small positive scalar\" artifactual reduce independent noise in the samples? I wonder if a visual comparison with canonical VAE is fair in this setting.\n- Are there other potential use cases for the observational model beyond pixel editing?\n- I think another way to improve upon the canonical VAE likelihood is to consider an iid gaussian likelihood on perceptual features [1]. The authors should consider this method in the related method section and potentially include this technique in the comparison.\n\n[1] Hou X, Shen L, Sun K, Qiu G. Deep feature consistent variational autoencoder. In2017 IEEE Winter Conference on Applications of Computer Vision (WACV) 2017 Mar 24 (pp. 1133-1141). IEEE.", "summary_of_the_review": "The paper is sound, well written and tackles an interesting problem. However, more work is needed to demonstrate improvement upon state of the art methods (specially, Dorta et al, 2020, and potentially Hou et al, 2017).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636369953839}, {"id": "qf01PjnvbSo", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2681/Reviewer_DqfM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In the standard Variational Autoencoder framework the statistics of the decoder output are assumed to be pixel-independent Gaussian which can lead to problems when sampling from the model when covariances are missing. To overcome these limitations, the authors propose to use the network architecture proposed by Monteiro at al. 2020 in the decoder of a variational autoencoder. In the approacah of \nMonteiro at al. 2020, the output distribution is modeled as low-rank multivariate normal. Qualitative results on the CelebA dataset demonstrate that samples from the proposed modified variational autoencoder capture covariances between different parts of the image.", "review_text": "More quantitative results are needed to demonstrate possible advantages of the proposed method. Overall the only contrubution of this paper is to transfer the method of Monteiro et al. into the variational autoencder framework. Therefore this paper should clearly demonstrate the benefits of the proposed method over the established practice of using the predicted mean output vector of the decoder.\nFig. 4 a) and b) compare the proposed method with the established practice of using the predicted mean output vector of the decoder. However in Fig. b) the samples from the proposed method shows some unfavourable color shift. Thus the mean vector creates the more realistic image and would still be the preferable option.\n\nMore quantitative results are needed:\n\n- \"Structured uncertainty\": Is the uncertainty calibrated? Please provide experiments on synthetic data, for example sampled from a multivariate distribution with known mean and variance to demonstrate that the framework actually models the uncertainty of the data distribution.\n- Use of constraints in Eq. 1: Is convergence of the model affected? Are the Lagrangian multupliers maximized during optimization? Does this affect the numerical stabiilty during training?\n- How were the dataset-specific slack variables in Eq, 1 chosen? \n\nReferences\n\n[Monteiro at al. 2020] Monteiro, Miguel, et al. \"Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty, Stochastic Segmentation Networks: Modelling Spatially Correlated Aleatoric Uncertainty, Neurips 2020\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In the standard Variational Autoencoder framework the statistics of the decoder output are assumed to be pixel-independent Gaussian which can lead to problems when sampling from the model when covariances are missing. To overcome these limitations, the authors propose to use the network architecture proposed by Monteiro at al. 2020 in the decoder of a variational autoencoder. In the approacah of \nMonteiro at al. 2020, the output distribution is modeled as low-rank multivariate normal. Qualitative results on the CelebA dataset demonstrate that samples from the proposed modified variational autoencoder capture covariances between different parts of the image.", "main_review": "More quantitative results are needed to demonstrate possible advantages of the proposed method. Overall the only contrubution of this paper is to transfer the method of Monteiro et al. into the variational autoencder framework. Therefore this paper should clearly demonstrate the benefits of the proposed method over the established practice of using the predicted mean output vector of the decoder.\nFig. 4 a) and b) compare the proposed method with the established practice of using the predicted mean output vector of the decoder. However in Fig. b) the samples from the proposed method shows some unfavourable color shift. Thus the mean vector creates the more realistic image and would still be the preferable option.\n\nMore quantitative results are needed:\n\n- \"Structured uncertainty\": Is the uncertainty calibrated? Please provide experiments on synthetic data, for example sampled from a multivariate distribution with known mean and variance to demonstrate that the framework actually models the uncertainty of the data distribution.\n- Use of constraints in Eq. 1: Is convergence of the model affected? Are the Lagrangian multupliers maximized during optimization? Does this affect the numerical stabiilty during training?\n- How were the dataset-specific slack variables in Eq, 1 chosen? \n\nReferences\n\n[Monteiro at al. 2020] Monteiro, Miguel, et al. \"Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty, Stochastic Segmentation Networks: Modelling Spatially Correlated Aleatoric Uncertainty, Neurips 2020\n", "summary_of_the_review": "The contribution of this paper is limited, it merely takes an existing method to capture covariances of model outputs and incoroporates it into the decoder of a variational autoencoder. More quantitative results are needed to demonstrate possible advantages of the proposed method and calibration of the estimated covariances. Further details on parameter optimization and convergence are needed.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636316393629}, {"id": "XW5XCEEyZyM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2681/Reviewer_uSoj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes to use a low-rank plus diagonal covariance matrix, rather than the usual diagonal ones, in the decoder of Gaussian VAEs. The authors empirically show the advantages of using more expressive covariance matrices in the decoder.", "review_text": "This paper is well-written, easy to follow, and well motivated. I agree with the authors that having more flexible decoder distributions, and not just more flexible neural networks (mapping to parameters of simple distributions) is an understudied area in VAEs that deserves work. The paper is very simple and relies mostly on empirical results, but I do think that the authors properly show improvements and added benefits over using diagonal covariance matrices.\n\nOne concern I have though is that it seems like adding this structure to the covariance matrix come with its share of training instabilities. The authors need to add regularizers (eq. 2), and some of the parameters used look incredibly specific (e.g. $\\xi_H=-504750$ and $\\xi_H=-199250$ for different datasets). How were the values of $\\xi_H$ chosen? How robust are the results to changes in this hyperparameter? I think a careful answer to these questions should be presented, at least in the appendix to strengthen the paper.\n\nAlso, while I think the experiments do suggest improvements over diagonal covariance matrices, I think there are relevant additional experiments that are left out. Firstly, as the authors correctly point out, they are not the first to use a structured covariance matrix in the decoder of a Gaussian VAE [1], although this previous work uses sparse covariance matrices rather than low-rank plus diagonal ones. The absence of comparisons against this other structure significantly weakens the paper in my opinion: the baseline should not be diagonal covariance, it should be sparse covariances based on pixel neighbouring structure (as in [1]); and the reader should be convinced that low-rank plus diagonal is a more sensible choice than sparse. In other words, the authors claim \"([1]) limits its ability to capture long-range spatial dependencies\", and they should empirically verify that this actually hurts performance. Some more minor experimental things that I believe are also left out are: The authors should report \"Our VAE (means)\" in Table 1 as well. I imagine these results would be worse that the standard VAE's given for example the grey backgrounds in CELEBA, but I think this is actually interesting! It means the decoder is actually being used as a distribution in the generative model. I also believe that log-likelihoods (or ELBOs) should be reported in Table 1.\n\nI will consider increasing my score if my main issues (hyperparameter instability, and comparisons against [1]) are adequately addressed during the rebuttal period.\n\nMinor things:\n\n-Citation links seem to be formatted in a different way than other papers I reviewed (they appear in blue in other papers).\n\n-The notation S = H x W seems a bit weird: why collapse only height and width, but not channels C into a single variable? That is, S could be defined as S = H x W x C and notation would be simplified.\n\n-Above eq.3, $t \\in [0..1]$ should be $t \\in [0,1]$ I believe.\n\n-The slerp function in eq.3 should be explicitly written down, at least in the appendix.\n\n[1] Structured uncertainty prediction networks, Dorta et al. 2018\n\n========================================================================================================\n\nUPDATE 1 AFTER REBITTAL\n\n========================================================================================================\n\nI have read the author's rebuttal, and while I understand that adding the comparisons against Dorta et al. is a significant amount of work, I am not willing to increase my score until definitive comparisons have been done, even if agree with the authors that the method of Dorta should not add large-scale noise, as ELBO/FID comparisons could remain interesting.\n\n========================================================================================================\n\nUPDATE 2 AFTER REBITTAL\n\n========================================================================================================\n\nI have read the updated author's rebuttal with comparisons against Dorta et al.'s method. I decided to increase my score, and still want to encourage the authors to include ELBO comparisons in an updated version of the manuscript.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to use a low-rank plus diagonal covariance matrix, rather than the usual diagonal ones, in the decoder of Gaussian VAEs. The authors empirically show the advantages of using more expressive covariance matrices in the decoder.", "main_review": "This paper is well-written, easy to follow, and well motivated. I agree with the authors that having more flexible decoder distributions, and not just more flexible neural networks (mapping to parameters of simple distributions) is an understudied area in VAEs that deserves work. The paper is very simple and relies mostly on empirical results, but I do think that the authors properly show improvements and added benefits over using diagonal covariance matrices.\n\nOne concern I have though is that it seems like adding this structure to the covariance matrix come with its share of training instabilities. The authors need to add regularizers (eq. 2), and some of the parameters used look incredibly specific (e.g. $\\xi_H=-504750$ and $\\xi_H=-199250$ for different datasets). How were the values of $\\xi_H$ chosen? How robust are the results to changes in this hyperparameter? I think a careful answer to these questions should be presented, at least in the appendix to strengthen the paper.\n\nAlso, while I think the experiments do suggest improvements over diagonal covariance matrices, I think there are relevant additional experiments that are left out. Firstly, as the authors correctly point out, they are not the first to use a structured covariance matrix in the decoder of a Gaussian VAE [1], although this previous work uses sparse covariance matrices rather than low-rank plus diagonal ones. The absence of comparisons against this other structure significantly weakens the paper in my opinion: the baseline should not be diagonal covariance, it should be sparse covariances based on pixel neighbouring structure (as in [1]); and the reader should be convinced that low-rank plus diagonal is a more sensible choice than sparse. In other words, the authors claim \"([1]) limits its ability to capture long-range spatial dependencies\", and they should empirically verify that this actually hurts performance. Some more minor experimental things that I believe are also left out are: The authors should report \"Our VAE (means)\" in Table 1 as well. I imagine these results would be worse that the standard VAE's given for example the grey backgrounds in CELEBA, but I think this is actually interesting! It means the decoder is actually being used as a distribution in the generative model. I also believe that log-likelihoods (or ELBOs) should be reported in Table 1.\n\nI will consider increasing my score if my main issues (hyperparameter instability, and comparisons against [1]) are adequately addressed during the rebuttal period.\n\nMinor things:\n\n-Citation links seem to be formatted in a different way than other papers I reviewed (they appear in blue in other papers).\n\n-The notation S = H x W seems a bit weird: why collapse only height and width, but not channels C into a single variable? That is, S could be defined as S = H x W x C and notation would be simplified.\n\n-Above eq.3, $t \\in [0..1]$ should be $t \\in [0,1]$ I believe.\n\n-The slerp function in eq.3 should be explicitly written down, at least in the appendix.\n\n[1] Structured uncertainty prediction networks, Dorta et al. 2018\n\n========================================================================================================\n\nUPDATE 1 AFTER REBITTAL\n\n========================================================================================================\n\nI have read the author's rebuttal, and while I understand that adding the comparisons against Dorta et al. is a significant amount of work, I am not willing to increase my score until definitive comparisons have been done, even if agree with the authors that the method of Dorta should not add large-scale noise, as ELBO/FID comparisons could remain interesting.\n\n========================================================================================================\n\nUPDATE 2 AFTER REBITTAL\n\n========================================================================================================\n\nI have read the updated author's rebuttal with comparisons against Dorta et al.'s method. I decided to increase my score, and still want to encourage the authors to include ELBO comparisons in an updated version of the manuscript.", "summary_of_the_review": "While I think this paper proposes a sensible idea and shows improvements over using diagonal covariances in Gaussian VAE decoders, I have concerns about the stability of the proposed method; and I believe that the authors should compare against a stronger baseline than the one they use for the experiments to be fully convincing.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635863029403}], "openreview_url": "https://openreview.net/forum?id=Qu_XudmGajz", "arxiv_id": "2205.12533", "paper_pdf": "papers/Qu_XudmGajz.pdf", "paper_pdf_sha256": "baad970b9fe313794529d8fc7729e3b6b34752b6af86a988c1caa959bf8a5af1", "paper_pdf_bytes": 47036636, "paper_pdf_source": "openreview", "code_url": "https://github.com/biomedia-mira/sos-vae", "code_repository": "biomedia-mira/sos-vae", "code_commit": "cc02afb2551dee5b472741dce3d1e82e0996c83a", "code_archive": "repos/Qu_XudmGajz.zip", "code_archive_sha256": "12e59ab95dd7d30e72f07eb9a745e0b6c973405eb6c5d6fe0b3749a7207ce0c7", "code_archive_bytes": 2262927, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 2204, "github_languages": {"Python": 113931}, "github_archived": false, "github_pushed_at": "2022-10-28T22:11:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/structured-uncertainty-in-the-observation-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "lJgbDxGhJ4r", "year": 2021, "status": "rejected", "title": "OpenCoS: Contrastive Semi-supervised Learning for Handling Open-set Unlabeled Data", "authors": ["Jongjin Park", "Sukmin Yun", "Jongheon Jeong", "Jinwoo Shin"], "authorids": ["~Jongjin_Park1", "~Sukmin_Yun1", "~Jongheon_Jeong1", "~Jinwoo_Shin1"], "authors_source": "OpenReview API", "abstract": "Modern semi-supervised learning methods conventionally assume both labeled and unlabeled data have the same class distribution. However, unlabeled data may include out-of-class samples in practice; those that cannot have one-hot encoded labels from a closed-set of classes in label data, i.e., unlabeled data is an open-set. In this paper, we introduce OpenCoS, a method for handling this realistic semi-supervised learning scenario based on a recent framework of contrastive learning. One of our key findings is that out-of-class samples in the unlabeled dataset can be identified effectively via (unsupervised) contrastive learning. OpenCoS utilizes this information to overcome the failure modes in the existing state-of-the-art semi-supervised methods, e.g., ReMixMatch or FixMatch. In particular, we propose to assign soft-labels for out-of-class samples using the representation learned from contrastive learning. Our extensive experimental results show the effectiveness of OpenCoS, fixing the state-of-the-art semi-supervised methods to be suitable for diverse scenarios involving open-set unlabeled data.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "wRe855ial8", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2540/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\nSummary:\n\nThe paper proposes a new approach for open set semi-supervised learning, where there are unlabeled data from classes not in the labeled data. The paper uses a contrastive representation learning paradigm to learn a feature encoder and a similarity measurement. Then the paper filters outlier samples by the similarity measurement and further utilizes outlier samples with soft labels. The separate BN layers address the distribution shift between in-class and out-class data.\n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for accepting. I like the idea of learning representation in an unsupervised way for both labeled and unlabeled data. My major concern is about how to select a reasonable threshold and why the threshold works well (see cons below). Hopefully the authors can address my concern in the rebuttal period. \n\n########################################################################## \n\nPros:\n\nThe paper addresses a very interesting and practical problem in semi-supervised learning, where the unlabeled samples may include out-class samples. This problem promotes the practical use of semi-supervised learning in real world applications.\n\nThe paper employs contrastive representation learning to learn the encoder and similarity measurement, which preserves the similarity between features of a sample transformed by different transformations. The learned encoder can encode the semantic information into the feature for both labeled and unlabeled data in an unsupervised way. \n\nThe paper filters the out-class samples by the learned similarity measure based on a threshold, which can better filter out out-class samples.\n\nThe paper employs different batch normalization layers for the in-class and out-class samples, which avoids the influence of the distribution shift.\n\nThe paper is well-written and the claims are clearly clarified.\n\nExperimental results on three semi-supervised learning benchmarks: CIFAR-10, CIFAR-100 and ImageNet, show that the proposed method can detect open-class samples and achieves higher accuracy the previous semi-supervised learning methods. Qualitative results show that the method is not very sensitive to hyper-parameters.\n\n##########################################################################\n\nCons:\n\nThe performance gain of the proposed method is only over closed-set semi-supervised learning methods. The paper does not compare with the state-of-the-art open set semi-supervised learning method (Guo et al., 2020).\n\nThe threshold hyper-parameter is the key to filtering out-class samples. The paper decides the threshold by mean minus 2 standard deviation. Could the authors explain why choosing this value? Is there any insight in this threshold or why does this threshold fitful for any dataset?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for the paper", "review": "##########################################################################\nSummary:\n\nThe paper proposes a new approach for open set semi-supervised learning, where there are unlabeled data from classes not in the labeled data. The paper uses a contrastive representation learning paradigm to learn a feature encoder and a similarity measurement. Then the paper filters outlier samples by the similarity measurement and further utilizes outlier samples with soft labels. The separate BN layers address the distribution shift between in-class and out-class data.\n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for accepting. I like the idea of learning representation in an unsupervised way for both labeled and unlabeled data. My major concern is about how to select a reasonable threshold and why the threshold works well (see cons below). Hopefully the authors can address my concern in the rebuttal period. \n\n########################################################################## \n\nPros:\n\nThe paper addresses a very interesting and practical problem in semi-supervised learning, where the unlabeled samples may include out-class samples. This problem promotes the practical use of semi-supervised learning in real world applications.\n\nThe paper employs contrastive representation learning to learn the encoder and similarity measurement, which preserves the similarity between features of a sample transformed by different transformations. The learned encoder can encode the semantic information into the feature for both labeled and unlabeled data in an unsupervised way. \n\nThe paper filters the out-class samples by the learned similarity measure based on a threshold, which can better filter out out-class samples.\n\nThe paper employs different batch normalization layers for the in-class and out-class samples, which avoids the influence of the distribution shift.\n\nThe paper is well-written and the claims are clearly clarified.\n\nExperimental results on three semi-supervised learning benchmarks: CIFAR-10, CIFAR-100 and ImageNet, show that the proposed method can detect open-class samples and achieves higher accuracy the previous semi-supervised learning methods. Qualitative results show that the method is not very sensitive to hyper-parameters.\n\n##########################################################################\n\nCons:\n\nThe performance gain of the proposed method is only over closed-set semi-supervised learning methods. The paper does not compare with the state-of-the-art open set semi-supervised learning method (Guo et al., 2020).\n\nThe threshold hyper-parameter is the key to filtering out-class samples. The paper decides the threshold by mean minus 2 standard deviation. Could the authors explain why choosing this value? Is there any insight in this threshold or why does this threshold fitful for any dataset?\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603966542073}, {"id": "taEABg8678a", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2540/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper considers the problem of semi-supervised learning, where the unlabeled data may include out-of-class samples. To address this task, the paper proposes a method consisting of three steps: (1) detecting out-of-class samples in the unlabeled set, (2) assigning soft-labels to the detected out-of-class samples using class-conditional likelihoods from labeled data, and (3) using auxiliary batch normalization layers  to help mitigate the class distribution mismatch problem. Experiments are conducted on CIFAR-10, CIFAR-100, ImageNet datasets. Results show improvements over competing methods.\n\nQuality: This paper is well written and well organized. I find this paper easy to follow.\n\nClarify: The preliminaries section clearly describes the setting of semi-supervised learning concerned in this paper and also clearly describes the contrastive representation learning. I like the way that the authors include a preliminaries section before the proposed method section. This allows me to understand which parts of the method are built upon existing approaches and helps me better identify the contributions and new things of this paper.\n\nOriginality: While I think the idea of this paper makes sense, I think there are several parts that are highly similar from a recent paper published in ECCV 2020 [a]. While the setting considered in [a] (for metric learning problems) is a bit different from that concerned in this submission (for image classification tasks), the high-level idea (learning representations that can be used to describe unlabeled images with labels different from those in the training set) is very similar. For example, the idea of Equation 5 and Equation 6 in this submission is almost exactly the same as that in Equation 2 of [a]. In addition, the idea of Equation 9 of this submission is very similar to that of Equation 3 of [a]. However, the authors did not acknowledge the similarity with [a] in the submission. This will make readers feel like the idea of Equation 5 and Equation 6 is original in this paper.\n\n[a] Chen et al. Learning to learn in a semi-supervised fashion. In ECCV, 2020. https://arxiv.org/pdf/2008.11203.pdf\n\nSignificance: Given that some parts of this paper are highly similar to [a], the significance of this paper is downplayed.\n\nRequest for author response: I would like to see how the authors compare their paper with [a] in terms of problem setting, idea of class-wise similarity and representation learning. In particular, I would like to know the pros and cons of this paper compared to [a]. It would be great if the authors can talk about whether the semantics-oriented similarity representation in [a] could be used (and how to use it) to help improve the performance of the proposed method in the setting concerned by the paper.\n\nRating: Given that there are many parts similar to [a] and they are not acknowledged in the paper, I can only rate 4 for this submission at this point. I will reevaluate this paper after seeing the reviews from the other reviews as well as the author response.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of OpenCoS", "review": "This paper considers the problem of semi-supervised learning, where the unlabeled data may include out-of-class samples. To address this task, the paper proposes a method consisting of three steps: (1) detecting out-of-class samples in the unlabeled set, (2) assigning soft-labels to the detected out-of-class samples using class-conditional likelihoods from labeled data, and (3) using auxiliary batch normalization layers  to help mitigate the class distribution mismatch problem. Experiments are conducted on CIFAR-10, CIFAR-100, ImageNet datasets. Results show improvements over competing methods.\n\nQuality: This paper is well written and well organized. I find this paper easy to follow.\n\nClarify: The preliminaries section clearly describes the setting of semi-supervised learning concerned in this paper and also clearly describes the contrastive representation learning. I like the way that the authors include a preliminaries section before the proposed method section. This allows me to understand which parts of the method are built upon existing approaches and helps me better identify the contributions and new things of this paper.\n\nOriginality: While I think the idea of this paper makes sense, I think there are several parts that are highly similar from a recent paper published in ECCV 2020 [a]. While the setting considered in [a] (for metric learning problems) is a bit different from that concerned in this submission (for image classification tasks), the high-level idea (learning representations that can be used to describe unlabeled images with labels different from those in the training set) is very similar. For example, the idea of Equation 5 and Equation 6 in this submission is almost exactly the same as that in Equation 2 of [a]. In addition, the idea of Equation 9 of this submission is very similar to that of Equation 3 of [a]. However, the authors did not acknowledge the similarity with [a] in the submission. This will make readers feel like the idea of Equation 5 and Equation 6 is original in this paper.\n\n[a] Chen et al. Learning to learn in a semi-supervised fashion. In ECCV, 2020. https://arxiv.org/pdf/2008.11203.pdf\n\nSignificance: Given that some parts of this paper are highly similar to [a], the significance of this paper is downplayed.\n\nRequest for author response: I would like to see how the authors compare their paper with [a] in terms of problem setting, idea of class-wise similarity and representation learning. In particular, I would like to know the pros and cons of this paper compared to [a]. It would be great if the authors can talk about whether the semantics-oriented similarity representation in [a] could be used (and how to use it) to help improve the performance of the proposed method in the setting concerned by the paper.\n\nRating: Given that there are many parts similar to [a] and they are not acknowledged in the paper, I can only rate 4 for this submission at this point. I will reevaluate this paper after seeing the reviews from the other reviews as well as the author response.", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603871205852}, {"id": "LKP5-OFrf-B", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2540/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors proposed to address the task of semi-supervised learning (SSL) by contrastive learning techniques, and the proposed techniques can be applied to handle open-set unlabeled data (i.e., the label spaces between label and unlabeled data are partially disjoint). I found the paper clearly written and easy to follow. The review comments are discussed below.\n\nThe proposed learning scheme is extended from SimCLR by Chen et al. More precisely, the authors perform unsupervised learning using D_l (w/o observing labels) and D_u using SimCLR techniques. In order to handle out-of-class samples in D_u, the authors present the idea of learning class-wise prototypical representation based on the above contrastive features. Detection of such samples is performed by a simple cosine similarity comparison between each instance in D_u and the prototypes of D_l. The main contribution lies in the last stage, i.e., SSL with auxiliary loss, which trains classifiers for recognizing in-class samples while assigning soft label scores for out-of-sample ones. The use of such soft labels allows the training of such unlabeled and out-of-class samples, which would be the major novelty of this work. (I do not feel that the use of batch norm would be viewed as technical contributions.)\n\nOverall, I feel that most of the technical components come from existing works (e.g., SimCLR from Chen et al. 2020, auxiliary batch norm from Xie et al. 2020.). The use of soft-label assignments has also been proposed by existing works, which makes the overall technical contributions to be marginal. The lack of verification on the soft label assignment would be the concern as well. For example, is \"cat\" assigned with 0.1*leopard + 0.2*lion + 0.7*tiger? Most importantly, are out-of-class samples assigned with uniformly soft labels (i.e., 1/C)? Out-of-class samples might still exhibit similarity with selected in-class categories, and thus forcing their soft labels to be a uniform distribution might not seem to be practical (if that's the case). Existing vision and learning models have been considering class-similarity based representation, etc. techniques for handling zero-shot or open-dataset learning problems (e.g., Xian et al. PAMI'18 and Scheirer et al. PAMI'12). Based on the above remarks, I feel that the paper is yet above the ICLR standard for acceptance.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Solid work but marginal novelty/technical contributions", "review": "The authors proposed to address the task of semi-supervised learning (SSL) by contrastive learning techniques, and the proposed techniques can be applied to handle open-set unlabeled data (i.e., the label spaces between label and unlabeled data are partially disjoint). I found the paper clearly written and easy to follow. The review comments are discussed below.\n\nThe proposed learning scheme is extended from SimCLR by Chen et al. More precisely, the authors perform unsupervised learning using D_l (w/o observing labels) and D_u using SimCLR techniques. In order to handle out-of-class samples in D_u, the authors present the idea of learning class-wise prototypical representation based on the above contrastive features. Detection of such samples is performed by a simple cosine similarity comparison between each instance in D_u and the prototypes of D_l. The main contribution lies in the last stage, i.e., SSL with auxiliary loss, which trains classifiers for recognizing in-class samples while assigning soft label scores for out-of-sample ones. The use of such soft labels allows the training of such unlabeled and out-of-class samples, which would be the major novelty of this work. (I do not feel that the use of batch norm would be viewed as technical contributions.)\n\nOverall, I feel that most of the technical components come from existing works (e.g., SimCLR from Chen et al. 2020, auxiliary batch norm from Xie et al. 2020.). The use of soft-label assignments has also been proposed by existing works, which makes the overall technical contributions to be marginal. The lack of verification on the soft label assignment would be the concern as well. For example, is \"cat\" assigned with 0.1*leopard + 0.2*lion + 0.7*tiger? Most importantly, are out-of-class samples assigned with uniformly soft labels (i.e., 1/C)? Out-of-class samples might still exhibit similarity with selected in-class categories, and thus forcing their soft labels to be a uniform distribution might not seem to be practical (if that's the case). Existing vision and learning models have been considering class-similarity based representation, etc. techniques for handling zero-shot or open-dataset learning problems (e.g., Xian et al. PAMI'18 and Scheirer et al. PAMI'12). Based on the above remarks, I feel that the paper is yet above the ICLR standard for acceptance.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603804423985}, {"id": "otBlKaAUQw", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2540/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Quality: \nThe quality of the paper is below average. The loss function that integrates out-of-class samples is counter intuitive and seems to be chosen based on improved empirical evidence.\n\nClarity:\nThe paper reads well.\n\nOriginality/Significance:\nThe originality of the approach is limited. Main ideas are borrowed from SimCLR and applied to SSL. \n\nDetailed Comments:\n\nThe paper uses contrastive learning idea proposed in SimCLR (Chen et al. 2020) to detect out-of-class samples and treat them in a different way than in-class unlabeled samples during semi-supervised learning. It also explores the idea of auxiliary batch normalization (from Xie et al. 2020) in the open-set SSL setting but the results of the ablation study suggest the level of improvement achieved by this normalization is negligible and the most of the improvement comes from more accurate detection of out-of-class samples through using the projection header function introduced in SimCLR paper. Although results across multiple benchmark datasets report significant improvement over other SSL techniques these improvements could be artificial as other techniques have no way of handling out-of-class samples. Integrating these other out-of-class detection techniques in ReMixMatch and comparing results with the proposed technique would offer a more compelling argument. Overall, the paper proposes a nice practical idea but for publication in ICLR one would like to see more theoretical insight along with empirical evidence. \n\nClass conditional likelihoods has been shown to be not very useful in detecting out-of-distribution samples in cross-entropy loss learning. What aspect of contrastive learning makes it more useful for open-set classification?\n\nCross-entropy term in equation 8 does not make much sense. It is simply computing the loss with respect to an incorrect label. If a sample belongs to an unknown class it is not clear why this would help SSL. If the purpose here is to capture shared characteristics of the samples then SimCLR trained with both labeled and unlabeled data already takes care of it. The authors try to justify this by considering that some seen classes would be similar to the unseen one but for a fine-grained classification task such samples may still hurt predictive performance. If there are no similar classes among labeled classes the authors argue that this loss will have a uniform affect for all classes. Again, this is not a compelling argument. No matter how dissimilar the unseen sample to labeled classes are it would be more similar to some of the classes than some others. Due to the normalization affect the weight distributions will significantly deviate from a uniform distribution. \n\nBaseline techniques are all SSL techniques. These are guaranteed to perform worse than the proposed technique because they have no way of handling out-of-class samples. What about other baselines? From Table 3 one can see that the main contribution comes from the detection of out-of-class samples. The authors compare the detection performance of their technique against other standard outlier detection techniques and show that the proposed detection outperforms all of them to achieve the highest AUC in the \"isolated\" detection task. However, it is still not clear why the projection header g achieves something that softmax probabilities cannot do. In other words why do class-conditional probabilities obtained from g are useful for out-of-class detection but the probabilities that one would obtain from the softmax layer of an architecture (such resnet) trained by cross-entropy loss is not much useful for the same task. Interesting  that not much insight has been provided in the SimCLR paper either. Along the same lines, as a contrastive loss function, triplet loss also seems to do well in open-world settings. \n\nMinor:\n\nPlease correct the following:\n\npage1 Compared to prior approaches approaches have that bypassed ...", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper uses contrastive learning idea proposed in SimCLR (Chen et al. 2020) to detect out-of-class samples and treat them in a different way than in-class unlabeled samples during semi-supervised learning.", "review": "Quality: \nThe quality of the paper is below average. The loss function that integrates out-of-class samples is counter intuitive and seems to be chosen based on improved empirical evidence.\n\nClarity:\nThe paper reads well.\n\nOriginality/Significance:\nThe originality of the approach is limited. Main ideas are borrowed from SimCLR and applied to SSL. \n\nDetailed Comments:\n\nThe paper uses contrastive learning idea proposed in SimCLR (Chen et al. 2020) to detect out-of-class samples and treat them in a different way than in-class unlabeled samples during semi-supervised learning. It also explores the idea of auxiliary batch normalization (from Xie et al. 2020) in the open-set SSL setting but the results of the ablation study suggest the level of improvement achieved by this normalization is negligible and the most of the improvement comes from more accurate detection of out-of-class samples through using the projection header function introduced in SimCLR paper. Although results across multiple benchmark datasets report significant improvement over other SSL techniques these improvements could be artificial as other techniques have no way of handling out-of-class samples. Integrating these other out-of-class detection techniques in ReMixMatch and comparing results with the proposed technique would offer a more compelling argument. Overall, the paper proposes a nice practical idea but for publication in ICLR one would like to see more theoretical insight along with empirical evidence. \n\nClass conditional likelihoods has been shown to be not very useful in detecting out-of-distribution samples in cross-entropy loss learning. What aspect of contrastive learning makes it more useful for open-set classification?\n\nCross-entropy term in equation 8 does not make much sense. It is simply computing the loss with respect to an incorrect label. If a sample belongs to an unknown class it is not clear why this would help SSL. If the purpose here is to capture shared characteristics of the samples then SimCLR trained with both labeled and unlabeled data already takes care of it. The authors try to justify this by considering that some seen classes would be similar to the unseen one but for a fine-grained classification task such samples may still hurt predictive performance. If there are no similar classes among labeled classes the authors argue that this loss will have a uniform affect for all classes. Again, this is not a compelling argument. No matter how dissimilar the unseen sample to labeled classes are it would be more similar to some of the classes than some others. Due to the normalization affect the weight distributions will significantly deviate from a uniform distribution. \n\nBaseline techniques are all SSL techniques. These are guaranteed to perform worse than the proposed technique because they have no way of handling out-of-class samples. What about other baselines? From Table 3 one can see that the main contribution comes from the detection of out-of-class samples. The authors compare the detection performance of their technique against other standard outlier detection techniques and show that the proposed detection outperforms all of them to achieve the highest AUC in the \"isolated\" detection task. However, it is still not clear why the projection header g achieves something that softmax probabilities cannot do. In other words why do class-conditional probabilities obtained from g are useful for out-of-class detection but the probabilities that one would obtain from the softmax layer of an architecture (such resnet) trained by cross-entropy loss is not much useful for the same task. Interesting  that not much insight has been provided in the SimCLR paper either. Along the same lines, as a contrastive loss function, triplet loss also seems to do well in open-world settings. \n\nMinor:\n\nPlease correct the following:\n\npage1 Compared to prior approaches approaches have that bypassed ...", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603397031225}], "openreview_url": "https://openreview.net/forum?id=lJgbDxGhJ4r", "arxiv_id": "2107.08943", "paper_pdf": "papers/lJgbDxGhJ4r.pdf", "paper_pdf_sha256": "a4fc42ef14d533800188e561a531b632211d36ab862dd782ef0b7f02cf2148c6", "paper_pdf_bytes": 1422343, "paper_pdf_source": "openreview", "code_url": "https://github.com/alinlab/OpenCoS", "code_repository": "alinlab/OpenCoS", "code_commit": "59003724045f82cf1ca54b2d509da7abc5aefe96", "code_archive": "repos/lJgbDxGhJ4r.zip", "code_archive_sha256": "6b8950d86121ad9c904ac288f5d6972dc2312275058e69a3bfc8c63a58a6a078", "code_archive_bytes": 1502669, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 1425, "github_languages": {"Python": 281937}, "github_archived": false, "github_pushed_at": "2022-06-16T08:04:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/opencos-contrastive-semi-supervised-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "6rm4ZC1nnP", "year": 2026, "status": "rejected", "title": "LDC-MTL: Balancing Multi-Task Learning through Scalable Loss Discrepancy Control", "authors": ["Peiyao Xiao", "Chaosheng Dong", "Shaofeng Zou", "Kaiyi Ji"], "authorids": ["~Peiyao_Xiao1", "~Chaosheng_Dong1", "~Shaofeng_Zou1", "~Kaiyi_Ji1"], "authors_source": "OpenReview API", "abstract": "Multi-task learning (MTL) has been widely adopted for its ability to simultaneously learn multiple tasks. While existing gradient manipulation methods often yield more balanced solutions than simple scalarization-based approaches, they typically incur a significant computational overhead of $\\mathcal{O}(K)$ in both time and memory, where $K$ is the number of tasks. In this paper, we propose LDC-MTL, a simple and scalable loss discrepancy control approach for MTL, formulated from a bilevel optimization perspective. Our method incorporates two key components: (i) a bilevel formulation for fine-grained loss discrepancy control, and (ii) a scalable first-order bilevel algorithm that requires only $\\mathcal{O}(1)$ time and memory. Theoretically, we prove that LDC-MTL guarantees convergence not only to a stationary point of the bilevel problem with loss discrepancy control but also to an $\\epsilon$-accurate Pareto stationary point for all $K$ loss functions under mild conditions. Extensive experiments on diverse multi-task datasets demonstrate the superior performance of LDC-MTL in both accuracy and efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sWUJMPyu22", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9715/Reviewer_xsvW"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper proposes LDC-MTL, a scalable loss discrepancy control method for multi-task learning (MTL). It formulates MTL as a bilevel optimization problem, where the lower level minimizes a weighted sum of task losses and the upper level minimizes the discrepancies among those losses. A first-order, single-loop approximation is introduced, leading to an $\\mathcal{O}(1)$ time and memory algorithm that scales with the number of tasks. Theoretical analysis establishes convergence both to a stationary point of the bilevel problem and to an $\\epsilon$-accurate Pareto-stationary point. Experiments on CelebA ($40$ tasks), QM9 ($11$ tasks), NYU-v$2$ ($3$ tasks), and Cityscapes ($2$ tasks) show competitive or superior $\\Delta m\\\\%$ compared to scalarization and gradient-manipulation baselines, with runtime close to linear scalarization (LS).", "review_text": "The paper proposes LDC-MTL, a scalable loss discrepancy control method for multi-task learning (MTL). It formulates MTL as a bilevel optimization problem, where the lower level minimizes a weighted sum of task losses and the upper level minimizes the discrepancies among those losses. A first-order, single-loop approximation is introduced, leading to an $\\mathcal{O}(1)$ time and memory algorithm that scales with the number of tasks. Theoretical analysis establishes convergence both to a stationary point of the bilevel problem and to an $\\epsilon$-accurate Pareto-stationary point. Experiments on CelebA ($40$ tasks), QM9 ($11$ tasks), NYU-v$2$ ($3$ tasks), and Cityscapes ($2$ tasks) show competitive or superior $\\Delta m\\\\%$ compared to scalarization and gradient-manipulation baselines, with runtime close to linear scalarization (LS).", "strengths": "The paper's main strengths lie in its clear bilevel formulation for controlling task-wise loss discrepancies, which provides a principled alternative to existing loss balancing methods. The proposed approach is both scalable and lightweight, featuring a single-loop update that avoids the typical $\\mathcal{O}(K)$ gradient storage overhead found in many multi-task algorithms. Theoretical analysis further reinforces the work by providing convergence guarantees that ensure Pareto stationarity under mild assumptions. Empirically, the paper presents comprehensive experiments across four benchmark datasets and compares against competitive baselines, demonstrating consistent improvements. Moreover, the additional analyses on gradient conflict reduction and loss variance offer valuable insights into the behavior and interpretability of the proposed method.", "weaknesses": "1. The claimed $O(1)$ efficiency critically depends on the empirical observation that $||\\nabla_W g(W^t,z^t_N)||$ remains small. While this is illustrated for the Cityscapes dataset $(K = 2)$, providing similar empirical evidence for other datasets, particularly CelebA $(K = 40)$, would strengthen the generality of this claim.\n\n2. The experimental analyses on loss discrepancy and gradient conflict (Table $3$, Figures $6$ and $7$) are conducted only against linear scalarization. For a fairer assessment, comparisons should also include the strongest loss balancing baselines, such as GO4Align, which achieves top performance in the main experiments.", "questions": "1. Could you provide a performance comparison between Algorithm $1$ and Algorithm $2$? Additionally, what do you think of a hybrid strategy that runs Algorithm $2$ until $||\\nabla_W g(W^t, z^t_N)||$ drops below a threshold relative to $||\\nabla_W g(W^t, x^t)||$, and then switches to Algorithm $1$ for the remainder of training?\n\n2. Could you include a comparison with ConsMTL $[1]$, a recent state-of-the-art gradient-manipulation method that formulates MTL as a bi-level optimization over shared and task-specific parameters? This would strengthen the claim that LDC-MTL consistently outperforms gradient-based baselines.\n\n3. In Appendix A-$7$, it is stated that the scatter plots in Figures $5 (a)$ and $5 (b)$ follow the same setting as Figure $8$ in $[2]$. Does this imply that the axis labels in Figures $5 (a–b)$ might have been mislabeled? Additionally, could you include FAMO and GO4Align (each trained with distinct random seeds) in this Pareto front analysis for a more comprehensive comparison?\n\n$[1]$ Qin, Xiaohan, Xiaoxing Wang, and Junchi Yan. \"Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters.\" Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.\n\n$[2]$ Xin, Derrick, et al. \"Do current multi-task optimization methods in deep learning even help?.\" Advances in neural information processing systems 35 (2022): 13597-13609.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes LDC-MTL, a scalable loss discrepancy control method for multi-task learning (MTL). It formulates MTL as a bilevel optimization problem, where the lower level minimizes a weighted sum of task losses and the upper level minimizes the discrepancies among those losses. A first-order, single-loop approximation is introduced, leading to an $\\mathcal{O}(1)$ time and memory algorithm that scales with the number of tasks. Theoretical analysis establishes convergence both to a stationary point of the bilevel problem and to an $\\epsilon$-accurate Pareto-stationary point. Experiments on CelebA ($40$ tasks), QM9 ($11$ tasks), NYU-v$2$ ($3$ tasks), and Cityscapes ($2$ tasks) show competitive or superior $\\Delta m\\\\%$ compared to scalarization and gradient-manipulation baselines, with runtime close to linear scalarization (LS).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper's main strengths lie in its clear bilevel formulation for controlling task-wise loss discrepancies, which provides a principled alternative to existing loss balancing methods. The proposed approach is both scalable and lightweight, featuring a single-loop update that avoids the typical $\\mathcal{O}(K)$ gradient storage overhead found in many multi-task algorithms. Theoretical analysis further reinforces the work by providing convergence guarantees that ensure Pareto stationarity under mild assumptions. Empirically, the paper presents comprehensive experiments across four benchmark datasets and compares against competitive baselines, demonstrating consistent improvements. Moreover, the additional analyses on gradient conflict reduction and loss variance offer valuable insights into the behavior and interpretability of the proposed method.", "weaknesses": "1. The claimed $O(1)$ efficiency critically depends on the empirical observation that $||\\nabla_W g(W^t,z^t_N)||$ remains small. While this is illustrated for the Cityscapes dataset $(K = 2)$, providing similar empirical evidence for other datasets, particularly CelebA $(K = 40)$, would strengthen the generality of this claim.\n\n2. The experimental analyses on loss discrepancy and gradient conflict (Table $3$, Figures $6$ and $7$) are conducted only against linear scalarization. For a fairer assessment, comparisons should also include the strongest loss balancing baselines, such as GO4Align, which achieves top performance in the main experiments.", "questions": "1. Could you provide a performance comparison between Algorithm $1$ and Algorithm $2$? Additionally, what do you think of a hybrid strategy that runs Algorithm $2$ until $||\\nabla_W g(W^t, z^t_N)||$ drops below a threshold relative to $||\\nabla_W g(W^t, x^t)||$, and then switches to Algorithm $1$ for the remainder of training?\n\n2. Could you include a comparison with ConsMTL $[1]$, a recent state-of-the-art gradient-manipulation method that formulates MTL as a bi-level optimization over shared and task-specific parameters? This would strengthen the claim that LDC-MTL consistently outperforms gradient-based baselines.\n\n3. In Appendix A-$7$, it is stated that the scatter plots in Figures $5 (a)$ and $5 (b)$ follow the same setting as Figure $8$ in $[2]$. Does this imply that the axis labels in Figures $5 (a–b)$ might have been mislabeled? Additionally, could you include FAMO and GO4Align (each trained with distinct random seeds) in this Pareto front analysis for a more comprehensive comparison?\n\n$[1]$ Qin, Xiaohan, Xiaoxing Wang, and Junchi Yan. \"Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters.\" Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.\n\n$[2]$ Xin, Derrick, et al. \"Do current multi-task optimization methods in deep learning even help?.\" Advances in neural information processing systems 35 (2022): 13597-13609.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761956971248}, {"id": "wRTJuBTpUd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9715/Reviewer_shRo"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This work proposed a first-order bi-level optimization approach to solve the multi-task learning problem. The proposed method mainly contains two components: (i) a bilevel formulation for fine-grained loss discrepancy control, and (ii) a scalable first-order bilevel algorithm that requires only constant order time and memory.", "review_text": "This work proposed a first-order bi-level optimization approach to solve the multi-task learning problem. The proposed method mainly contains two components: (i) a bilevel formulation for fine-grained loss discrepancy control, and (ii) a scalable first-order bilevel algorithm that requires only constant order time and memory.", "strengths": "1.  The proposed method avoids solving the complex bi-level structure in BLO. It also achieves good performance in MTL. \n2. The paper structure is clear and easy to follow.\n3. A theoretical proof to support the effectiveness of the proposed method.", "weaknesses": "1. Lacking $\\mathcal{O}(1)$ baselines such as \"Smooth Tchebycheff Scalarization for Multi-Objective Optimization, ICML 2024.\"\n\n2. Using bi-level optimization to solve the MTL problem is not new. The \"related work\" part does not mention those works, such as \"Multi-objective meta-learning, AIJ\" and \"A first-order multi-gradient algorithm for multi-objective bi-level optimization, ECAI 2024\".\n\n3. The proposed approach is quite like a penalty-based bi-level optimization approach, such as \"On Penalty-based Bilevel Gradient Descent Method\". Though the author cites some papers  (Kwon et al., 2023; Yang et al., 2023), but does not discuss the differences and how similar these approaches are. This is important for evaluating the theoretical contributions of this paper. In my opinion, the theoretical result does not show an advantage.", "questions": "See the weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposed a first-order bi-level optimization approach to solve the multi-task learning problem. The proposed method mainly contains two components: (i) a bilevel formulation for fine-grained loss discrepancy control, and (ii) a scalable first-order bilevel algorithm that requires only constant order time and memory.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.  The proposed method avoids solving the complex bi-level structure in BLO. It also achieves good performance in MTL. \n2. The paper structure is clear and easy to follow.\n3. A theoretical proof to support the effectiveness of the proposed method.", "weaknesses": "1. Lacking $\\mathcal{O}(1)$ baselines such as \"Smooth Tchebycheff Scalarization for Multi-Objective Optimization, ICML 2024.\"\n\n2. Using bi-level optimization to solve the MTL problem is not new. The \"related work\" part does not mention those works, such as \"Multi-objective meta-learning, AIJ\" and \"A first-order multi-gradient algorithm for multi-objective bi-level optimization, ECAI 2024\".\n\n3. The proposed approach is quite like a penalty-based bi-level optimization approach, such as \"On Penalty-based Bilevel Gradient Descent Method\". Though the author cites some papers  (Kwon et al., 2023; Yang et al., 2023), but does not discuss the differences and how similar these approaches are. This is important for evaluating the theoretical contributions of this paper. In my opinion, the theoretical result does not show an advantage.", "questions": "See the weaknesses part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761880228412}, {"id": "S02cEXIufN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9715/Reviewer_7Dov"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces LDC-MTL, a scalable multi-task learning method that tackles the persistent issue of loss discrepancy across task objectives. Motivated by the difficulty of balancing tasks whose losses operate on different scales or progress at different speeds, LDC-MTL formulates the learning challenge as a bilevel optimization problem: the lower-level optimizes a weighted sum of task losses, and the upper-level adjusts the weights to minimize their discrepancies. The authors propose a single-loop, first-order algorithm that claims $\\mathcal{O}(1)$ memory and time overhead per iteration, compared to $\\mathcal{O}(K)$ for previous gradient manipulation techniques. Experiments across standard MTL benchmarks demonstrate the efficiency and competitive accuracy of LDC-MTL.", "review_text": "This paper introduces LDC-MTL, a scalable multi-task learning method that tackles the persistent issue of loss discrepancy across task objectives. Motivated by the difficulty of balancing tasks whose losses operate on different scales or progress at different speeds, LDC-MTL formulates the learning challenge as a bilevel optimization problem: the lower-level optimizes a weighted sum of task losses, and the upper-level adjusts the weights to minimize their discrepancies. The authors propose a single-loop, first-order algorithm that claims $\\mathcal{O}(1)$ memory and time overhead per iteration, compared to $\\mathcal{O}(K)$ for previous gradient manipulation techniques. Experiments across standard MTL benchmarks demonstrate the efficiency and competitive accuracy of LDC-MTL.", "strengths": "1. The bilevel formulation provides a clear and principled avenue for directly controlling loss discrepancies while aiming for balanced performance across tasks (Section 4).\n\n2. The authors conducted extensive experiments to verify the effectiveness of the proposed LDC-MTL.\n\n3. The authors provide a thorough theoretical analysis, showing that the algorithm achieves $\\epsilon$-Pareto stationarity under standard Lipschitz and PL conditions.", "weaknesses": "1. The authors claim that LDC-MTL achieves $\\mathcal{O}(1)$ memory and time overhead per iteration. However, this property has already been stated in prior work such as FAMO. In fact, all loss-based MTL methods (e.g., UW, DWA, and more recently GO4Align) inherently possess this characteristic, as they aggregate task losses directly rather than computing gradients for each task separately—thus requiring only a single backward pass per iteration. Therefore, this advantage cannot be regarded as a distinctive contribution of the proposed method.\n\n2. The idea of employing bilevel optimization for multi-task learning is not novel. Recent works, including GO4Align and ConsMTL [2], have already adopted similar formulations, making the conceptual contribution of this paper relatively incremental.\n\n3. In Figure 1, the authors present the performance of LDC-MTL and other methods on a toy example. However, since there is no dominating solution along the Pareto front, convergence to the same Pareto stationary point cannot be considered a meaningful advantage in the absence of any prior preference among tasks.\n\n4. The proposed method involves several hyperparameters, such as the learning rate $\\alpha$ and the penalty coefficient $\\lambda$. The authors determine these through grid search, which contradicts the claimed advantage of computational efficiency, as it implies multiple runs are needed to find suitable values. More thorough ablation studies and analyses are necessary to demonstrate the robustness of the method and to provide practical guidance for hyperparameter selection.\n\n5. Confusing supplementary material. The authors appear to provide code in the supplementary material, yet upon inspection, there seems to be no actual implementation corresponding to the proposed LDC-MTL method. This raises concerns about reproducibility and transparency; the authors should ensure that the provided materials are complete and directly related to the presented work.\n\n6. Missing references. The paper overlooks several key related works [1,2,3], some of which report superior empirical performance (though this does not affect my evaluation of the paper). In the MTL field, there are no universally accepted quantitative metrics for superiority—measures such as $\\Delta m\\\\%$ and MR are only proxy indicators, and it is well known that $\\Delta m\\\\%$ tends to favor baselines with lower initial performance. Therefore, I do not insist that the authors include these methods in their experiments, but they should at least acknowledge and differentiate them properly in the Related Work section.\n\n[1] Independent component alignment for multi-task learning. CVPR 2023\n\n[2] Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters. CVPR 2025\n\n[5] Rep-MTL: Unleashing the Power of Representation-level Task Saliency for Multi-Task Learning. ICCV 2025", "questions": "Refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LDC-MTL, a scalable multi-task learning method that tackles the persistent issue of loss discrepancy across task objectives. Motivated by the difficulty of balancing tasks whose losses operate on different scales or progress at different speeds, LDC-MTL formulates the learning challenge as a bilevel optimization problem: the lower-level optimizes a weighted sum of task losses, and the upper-level adjusts the weights to minimize their discrepancies. The authors propose a single-loop, first-order algorithm that claims $\\mathcal{O}(1)$ memory and time overhead per iteration, compared to $\\mathcal{O}(K)$ for previous gradient manipulation techniques. Experiments across standard MTL benchmarks demonstrate the efficiency and competitive accuracy of LDC-MTL.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The bilevel formulation provides a clear and principled avenue for directly controlling loss discrepancies while aiming for balanced performance across tasks (Section 4).\n\n2. The authors conducted extensive experiments to verify the effectiveness of the proposed LDC-MTL.\n\n3. The authors provide a thorough theoretical analysis, showing that the algorithm achieves $\\epsilon$-Pareto stationarity under standard Lipschitz and PL conditions.", "weaknesses": "1. The authors claim that LDC-MTL achieves $\\mathcal{O}(1)$ memory and time overhead per iteration. However, this property has already been stated in prior work such as FAMO. In fact, all loss-based MTL methods (e.g., UW, DWA, and more recently GO4Align) inherently possess this characteristic, as they aggregate task losses directly rather than computing gradients for each task separately—thus requiring only a single backward pass per iteration. Therefore, this advantage cannot be regarded as a distinctive contribution of the proposed method.\n\n2. The idea of employing bilevel optimization for multi-task learning is not novel. Recent works, including GO4Align and ConsMTL [2], have already adopted similar formulations, making the conceptual contribution of this paper relatively incremental.\n\n3. In Figure 1, the authors present the performance of LDC-MTL and other methods on a toy example. However, since there is no dominating solution along the Pareto front, convergence to the same Pareto stationary point cannot be considered a meaningful advantage in the absence of any prior preference among tasks.\n\n4. The proposed method involves several hyperparameters, such as the learning rate $\\alpha$ and the penalty coefficient $\\lambda$. The authors determine these through grid search, which contradicts the claimed advantage of computational efficiency, as it implies multiple runs are needed to find suitable values. More thorough ablation studies and analyses are necessary to demonstrate the robustness of the method and to provide practical guidance for hyperparameter selection.\n\n5. Confusing supplementary material. The authors appear to provide code in the supplementary material, yet upon inspection, there seems to be no actual implementation corresponding to the proposed LDC-MTL method. This raises concerns about reproducibility and transparency; the authors should ensure that the provided materials are complete and directly related to the presented work.\n\n6. Missing references. The paper overlooks several key related works [1,2,3], some of which report superior empirical performance (though this does not affect my evaluation of the paper). In the MTL field, there are no universally accepted quantitative metrics for superiority—measures such as $\\Delta m\\\\%$ and MR are only proxy indicators, and it is well known that $\\Delta m\\\\%$ tends to favor baselines with lower initial performance. Therefore, I do not insist that the authors include these methods in their experiments, but they should at least acknowledge and differentiate them properly in the Related Work section.\n\n[1] Independent component alignment for multi-task learning. CVPR 2023\n\n[2] Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters. CVPR 2025\n\n[5] Rep-MTL: Unleashing the Power of Representation-level Task Saliency for Multi-Task Learning. ICCV 2025", "questions": "Refer to weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761828089030}, {"id": "TYKJ8EoPl6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission9715/Reviewer_PdRf"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper investigates MTL from the perspective of loss balancing. Specifically, it proposes a bi-level loss balancing paradigm to dynamically adjust the task weights and promote the balanced optimization. Besides, it also proposes a first-order alternative for efficiency purposes. Extensive experiments have demonstrated the proposed method's ability to achieve competitive performance across multiple mainstream MTL datasets.", "review_text": "This paper investigates MTL from the perspective of loss balancing. Specifically, it proposes a bi-level loss balancing paradigm to dynamically adjust the task weights and promote the balanced optimization. Besides, it also proposes a first-order alternative for efficiency purposes. Extensive experiments have demonstrated the proposed method's ability to achieve competitive performance across multiple mainstream MTL datasets.", "strengths": "1. Employing a bi-level paradigm for MTL becomes popular recently, thus this paper is timely.\n2. Developing efficient MTL algorithms is crucial in the era of large foundation models, and this paper addresses a timely and important issue.\n3. Extensive evaluation demonstrates excellent performance across mainstream MTL datasets.", "weaknesses": "1. It remains unclear how Eqn. (1) facilitates balanced task learning speeds, especially given that $\\tau$ is fixed at 1 in most experimental settings.\n2. The paper lacks sufficient details regarding the router model, which plays a critical role in the proposed framework.\n3. The proposed method adopts a different training schedule (e.g., learning rate, batch size) compared to prior MTL approaches [1][2]; please provide a discussion on GPU memory usage under this setting.\n4. It is recommended to report another widely adopted MTL metric, Mean Rank (MR), for a more comprehensive evaluation [1][2].\n5. Please specify which version of LDC-MTL was used for evaluation.\n\nReference:\n\n[1] Fair Resource Allocation in Multi-Task Learning. ICML 2024.\n\n[2] FAMO: fast adaptive multitask optimization. NeurIPS 2023.", "questions": "Please refer to the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates MTL from the perspective of loss balancing. Specifically, it proposes a bi-level loss balancing paradigm to dynamically adjust the task weights and promote the balanced optimization. Besides, it also proposes a first-order alternative for efficiency purposes. Extensive experiments have demonstrated the proposed method's ability to achieve competitive performance across multiple mainstream MTL datasets.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. Employing a bi-level paradigm for MTL becomes popular recently, thus this paper is timely.\n2. Developing efficient MTL algorithms is crucial in the era of large foundation models, and this paper addresses a timely and important issue.\n3. Extensive evaluation demonstrates excellent performance across mainstream MTL datasets.", "weaknesses": "1. It remains unclear how Eqn. (1) facilitates balanced task learning speeds, especially given that $\\tau$ is fixed at 1 in most experimental settings.\n2. The paper lacks sufficient details regarding the router model, which plays a critical role in the proposed framework.\n3. The proposed method adopts a different training schedule (e.g., learning rate, batch size) compared to prior MTL approaches [1][2]; please provide a discussion on GPU memory usage under this setting.\n4. It is recommended to report another widely adopted MTL metric, Mean Rank (MR), for a more comprehensive evaluation [1][2].\n5. Please specify which version of LDC-MTL was used for evaluation.\n\nReference:\n\n[1] Fair Resource Allocation in Multi-Task Learning. ICML 2024.\n\n[2] FAMO: fast adaptive multitask optimization. NeurIPS 2023.", "questions": "Please refer to the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761550502503}], "openreview_url": "https://openreview.net/forum?id=6rm4ZC1nnP", "arxiv_id": "2502.08585", "paper_pdf": "papers/6rm4ZC1nnP.pdf", "paper_pdf_sha256": "a1cb15bb99381f7a685fd643cfc7b43ccbb1c3393936e8e0fa4cf9b760c69fba", "paper_pdf_bytes": 1512519, "paper_pdf_source": "openreview", "code_url": "https://github.com/OptMN-Lab/LDC-MTL", "code_repository": "OptMN-Lab/LDC-MTL", "code_commit": "7c3946518689bf05be196c7bd51947b2cf885c9e", "code_archive": "repos/6rm4ZC1nnP.zip", "code_archive_sha256": "b659cca14ea0b2b5787299f6f00a17213201b26367d66fba2a84d7eb352b6b30", "code_archive_bytes": 3543370, "code_file_count": 24, "code_extensions": {".py": 20, ".sh": 4}, "github_disk_usage_kb": 4871, "github_languages": {"Python": 190554, "Shell": 806}, "github_archived": false, "github_pushed_at": "2025-05-15T16:15:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/scalable-bilevel-loss-balancing-for-multi"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ogmzNfeRl7", "year": 2025, "status": "rejected", "title": "Correlations Are Ruining Your Gradient Descent", "authors": ["Nasir Ahmad"], "authorids": ["~Nasir_Ahmad1"], "authors_source": "OpenReview API", "abstract": "Herein the topics of (natural) gradient descent, data decorrelation, and approximate methods for backpropagation are brought into a common discussion. Natural gradient descent illuminates how gradient vectors, pointing at directions of steepest descent, can be improved by considering the local curvature of loss landscapes. We extend this perspective and show that to fully solve the problem illuminated by natural gradients in neural networks, one must recognise that correlations in the data at any linear transformation, including node responses at every layer of a neural network, cause a non-orthonormal relationship between the model's parameters. To solve this requires a method for decorrelating inputs at each individual layer of a neural network. We describe a range of methods which have been proposed for decorrelation and whitening of node output, and expand on these to provide a novel method specifically useful for distributed computing and computational neuroscience. Implementing decorrelation within multi-layer neural networks, we can show that not only is training via backpropagation sped up significantly but also existing approximations of backpropagation, which have failed catastrophically in the past, benefit significantly in their accuracy and convergence speed. This has the potential to provide a route forward for approximate gradient descent methods which have previously been discarded, training approaches for analogue and neuromorphic hardware, and potentially insights as to the efficacy and utility of decorrelation processes in the brain.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "E9n0SY1gdW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1090/Reviewer_Ggrt"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "Starting from natural gradient descent, the authors show that correlations in data can cause non-orthonormal relationship between the model's parameters. To mitigate this, they propose a decorrelation mechanism that uses only local information. They demonstrate that this mechanism improves BP accuracy, but also notably improves results for alternate training methods such as feedback alignment, and node perturbation.", "review_text": "Starting from natural gradient descent, the authors show that correlations in data can cause non-orthonormal relationship between the model's parameters. To mitigate this, they propose a decorrelation mechanism that uses only local information. They demonstrate that this mechanism improves BP accuracy, but also notably improves results for alternate training methods such as feedback alignment, and node perturbation.", "strengths": "- The focus on the input correlation part of natural gradient descent is novel, as is the proposed decorrelation mechanism.\n- The observation that this improves performance for BP alternatives is very significant, since these methods have the potential to be more efficient for training\n- The narrative exposition was easy to follow, and the concepts are explained very clearly\n- The (almost) empirical evaluation supports the conclusions of the paper\n- connections to biology is interesting", "weaknesses": "- The discussion of methods from computational neuroscience is very cursory, and would benefit from more details such as forms of specific learning rules\n- The recurrence aspect of the decorrelation rule could be discussed in more depth in the main text.\n- It's not clear if the proposed decorrelation method only has a recurrent formulation, or can be efficiently implemented without recurrence as well.", "questions": "- What is the computational cost of the proposed decorrelation method?\n-  What are the implications of having a recurrent formulation for computational efficiency and implementation?\n- There are second order methods which seem to have gotten some traction such as shampoo [1, 3] and K-FAC [2]. The statement in the first paragraph against 2nd order methods needs more nuance.\n\n[1] Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned stochastic tensor optimization. In International Conference on Machine Learning, pp. 1842–1850. PMLR, 2018.\n[2] Yi Ren and Donald Goldfarb. Tensor normal training for deep learning models. Advances in Neural Information Processing Systems, 34:26040–26052, 2021.\n[3] Vyas, N., Morwani, D., Zhao, R., Shapira, I., Brandfonbrener, D., Janson, L., and Kakade, S. (2024). SOAP: Improving and Stabilizing Shampoo using Adam. Preprint at arXiv, https://doi.org/10.48550/arXiv.2409.11321 https://doi.org/10.48550/arXiv.2409.11321.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Starting from natural gradient descent, the authors show that correlations in data can cause non-orthonormal relationship between the model's parameters. To mitigate this, they propose a decorrelation mechanism that uses only local information. They demonstrate that this mechanism improves BP accuracy, but also notably improves results for alternate training methods such as feedback alignment, and node perturbation.", "soundness": 4, "presentation": 3, "contribution": 4, "strengths": "- The focus on the input correlation part of natural gradient descent is novel, as is the proposed decorrelation mechanism.\n- The observation that this improves performance for BP alternatives is very significant, since these methods have the potential to be more efficient for training\n- The narrative exposition was easy to follow, and the concepts are explained very clearly\n- The (almost) empirical evaluation supports the conclusions of the paper\n- connections to biology is interesting", "weaknesses": "- The discussion of methods from computational neuroscience is very cursory, and would benefit from more details such as forms of specific learning rules\n- The recurrence aspect of the decorrelation rule could be discussed in more depth in the main text.\n- It's not clear if the proposed decorrelation method only has a recurrent formulation, or can be efficiently implemented without recurrence as well.", "questions": "- What is the computational cost of the proposed decorrelation method?\n-  What are the implications of having a recurrent formulation for computational efficiency and implementation?\n- There are second order methods which seem to have gotten some traction such as shampoo [1, 3] and K-FAC [2]. The statement in the first paragraph against 2nd order methods needs more nuance.\n\n[1] Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned stochastic tensor optimization. In International Conference on Machine Learning, pp. 1842–1850. PMLR, 2018.\n[2] Yi Ren and Donald Goldfarb. Tensor normal training for deep learning models. Advances in Neural Information Processing Systems, 34:26040–26052, 2021.\n[3] Vyas, N., Morwani, D., Zhao, R., Shapira, I., Brandfonbrener, D., Janson, L., and Kakade, S. (2024). SOAP: Improving and Stabilizing Shampoo using Adam. Preprint at arXiv, https://doi.org/10.48550/arXiv.2409.11321 https://doi.org/10.48550/arXiv.2409.11321.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730757755066}, {"id": "eU8SUz5ii5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1090/Reviewer_qqzv"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "In this work, the authors address the challenge of data correlations impacting neural activity and the learning process within neural networks. They take a comprehensive approach, beginning with an analysis that connects data correlations to the non-Euclidean geometry of the parameter space. Specifically, they demonstrate that correlations in input data lead to longer paths through parameter space during learning. This observation is rooted in the theory of Natural Gradient Descent, as originally developed by Amari in the late 90's, where parameter space is considered curved and defined locally by the inverse Fisher information matrix. The authors provide evidence that input correlations influence these longer paths.\nNext, they propose an incremental learning rule capable of decorrelating inputs within each neural network layer. This rule facilitates a linear transformation to reduce input correlations progressively. The authors demonstrate that employing such a decorrelation mechanism accelerates learning and potentially improves generalization across several learning paradigms, including standard backpropagation, Node Perturbation, and Feedback Alignment. The concluding part of the paper discusses the implications of removing correlations in biological and artificial settings.", "review_text": "In this work, the authors address the challenge of data correlations impacting neural activity and the learning process within neural networks. They take a comprehensive approach, beginning with an analysis that connects data correlations to the non-Euclidean geometry of the parameter space. Specifically, they demonstrate that correlations in input data lead to longer paths through parameter space during learning. This observation is rooted in the theory of Natural Gradient Descent, as originally developed by Amari in the late 90's, where parameter space is considered curved and defined locally by the inverse Fisher information matrix. The authors provide evidence that input correlations influence these longer paths.\nNext, they propose an incremental learning rule capable of decorrelating inputs within each neural network layer. This rule facilitates a linear transformation to reduce input correlations progressively. The authors demonstrate that employing such a decorrelation mechanism accelerates learning and potentially improves generalization across several learning paradigms, including standard backpropagation, Node Perturbation, and Feedback Alignment. The concluding part of the paper discusses the implications of removing correlations in biological and artificial settings.", "strengths": "• The initial sections of the paper provide an insightful overview of Natural Gradient Descent and offer a compelling perspective on how Amari’s theoretical framework can be contextualized within deep learning. While typical applications of Natural Gradients focus on Bayesian inference, where the Fisher information naturally defines the metric, the authors present an innovative interpretation that broadens its relevance.\n\n• The simulations included in the study convincingly demonstrate that implementing decorrelation at each layer significantly enhances training efficiency, particularly within the Feedback Alignment framework.\n\n• The paper offers a comprehensive overview of the impact of data correlations on training dynamics, which serves to enrich the reader’s understanding of this often underexplored aspect of deep learning optimization.", "weaknesses": "The primary issue with this paper is its limited coherence and lack of substantial innovation. The manuscript presents small incremental contributions across several topics and attempts to integrate them into a unified framework. This approach results in a paper that straddles the line between a review and an opinion piece, which does not align well with the expectations for an ICLR submission. Below, I outline the specific areas where the paper falls short in terms of novelty:\n1. **Review of Natural Gradient Descent**: The first section provides a summary of Natural Gradient Descent, highlighting that input correlations can be inferred from the derivation of the metric tensor. However, this observation is neither surprising nor particularly useful, as it merely reiterates that input correlations are intrinsic to Natural Gradient theory. Moreover, the concept of learning dynamics in curved parameter spaces is not applied or built upon in subsequent sections, limiting the relevance of this part of the paper to the overall narrative.\n\n2. **Derivation of a Local Learning Rule**: The second section presents a derivation for a local learning rule intended to decorrelate inputs to neural network layers. While the authors point out the novelty of their incremental rule (line 340), this rule closely resembles existing methods such as recursive least squares (e.g., Sussillo and Abbott, 2009), which also stem from the goal of decorrelating neural activity. The resemblance is even clearer when studying the derivation of the proposed decorrelation learning rule in Appendix C. Notably, the proposed rule introduces a novel scalar rescaling that maintains the norm. However, this addition is minor, and the overall innovation is minimal.\n\n3. **Numerical Experiments on Layer-wise Decorrelation**: The third section includes numerical results showing the advantages of input decorrelation for training deep networks. Prior work has already demonstrated that input decorrelation accelerates learning in frameworks such as backpropagation and Node Perturbation. Consequently, the main novelty here is the improved performance observed in the Feedback Alignment framework. However, the paper does not provide sufficient explanation or analysis to elucidate why decorrelation significantly enhances Feedback Alignment, nor does the performance match standard backpropagation. This omission limits the impact of this result, rendering it another incremental contribution.\n\nCollectively, these points suggest that the novelty presented in this work does not meet the standards expected for ICLR.\n\nFinally, the discussion section of the paper, while extensive, lacks coherence and leans heavily on speculation. For instance, the portion discussing biological plausibility is unclear in its message. Although the authors correctly note that decorrelation is observed in various neural circuits, the connection between these observations and the mechanisms or theories presented in the paper remains ambiguous. Additionally, the claim that decorrelation improves generalization (line 479) is inadequately supported by the data. The training curves in Figure 4 do not appear to reach convergence, as indicated by the non-saturating test accuracy. This raises doubts about the robustness of the claimed generalization improvements.", "questions": "To better align this work with the standards and expectations of an ICLR paper, the authors need to sharpen their message and clearly articulate the novel contributions of each section. Specifically, the authors may consider addressing the following questions:\n1. How does the consideration of Natural Gradient Descent and the dynamics in curved parameter space contribute to the understanding or derivation of the incremental learning rule? Can this rule be directly derived from the principles of the Natural Gradient theorem?\n2. What differentiates the proposed learning rule from existing frameworks for decorrelating neural activity, such as Recursive Least Squares? If there are differences, what are the specific theoretical or practical advantages of the proposed rule?\n3. Could the authors provide more extensive numerical analyses to demonstrate whether the benefits of their approach extend beyond faster learning to improved generalization? While Natural Gradient theory primarily explains accelerated convergence, do the authors observe any substantial gains in generalization performance?\n\nLastly, I want to note that I enjoyed reading the paper and its broad review, which connects the concept of natural gradients to deep network training. However, this discussion currently feels more suited to an opinion piece.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors address the challenge of data correlations impacting neural activity and the learning process within neural networks. They take a comprehensive approach, beginning with an analysis that connects data correlations to the non-Euclidean geometry of the parameter space. Specifically, they demonstrate that correlations in input data lead to longer paths through parameter space during learning. This observation is rooted in the theory of Natural Gradient Descent, as originally developed by Amari in the late 90's, where parameter space is considered curved and defined locally by the inverse Fisher information matrix. The authors provide evidence that input correlations influence these longer paths.\nNext, they propose an incremental learning rule capable of decorrelating inputs within each neural network layer. This rule facilitates a linear transformation to reduce input correlations progressively. The authors demonstrate that employing such a decorrelation mechanism accelerates learning and potentially improves generalization across several learning paradigms, including standard backpropagation, Node Perturbation, and Feedback Alignment. The concluding part of the paper discusses the implications of removing correlations in biological and artificial settings.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "• The initial sections of the paper provide an insightful overview of Natural Gradient Descent and offer a compelling perspective on how Amari’s theoretical framework can be contextualized within deep learning. While typical applications of Natural Gradients focus on Bayesian inference, where the Fisher information naturally defines the metric, the authors present an innovative interpretation that broadens its relevance.\n\n• The simulations included in the study convincingly demonstrate that implementing decorrelation at each layer significantly enhances training efficiency, particularly within the Feedback Alignment framework.\n\n• The paper offers a comprehensive overview of the impact of data correlations on training dynamics, which serves to enrich the reader’s understanding of this often underexplored aspect of deep learning optimization.", "weaknesses": "The primary issue with this paper is its limited coherence and lack of substantial innovation. The manuscript presents small incremental contributions across several topics and attempts to integrate them into a unified framework. This approach results in a paper that straddles the line between a review and an opinion piece, which does not align well with the expectations for an ICLR submission. Below, I outline the specific areas where the paper falls short in terms of novelty:\n1. **Review of Natural Gradient Descent**: The first section provides a summary of Natural Gradient Descent, highlighting that input correlations can be inferred from the derivation of the metric tensor. However, this observation is neither surprising nor particularly useful, as it merely reiterates that input correlations are intrinsic to Natural Gradient theory. Moreover, the concept of learning dynamics in curved parameter spaces is not applied or built upon in subsequent sections, limiting the relevance of this part of the paper to the overall narrative.\n\n2. **Derivation of a Local Learning Rule**: The second section presents a derivation for a local learning rule intended to decorrelate inputs to neural network layers. While the authors point out the novelty of their incremental rule (line 340), this rule closely resembles existing methods such as recursive least squares (e.g., Sussillo and Abbott, 2009), which also stem from the goal of decorrelating neural activity. The resemblance is even clearer when studying the derivation of the proposed decorrelation learning rule in Appendix C. Notably, the proposed rule introduces a novel scalar rescaling that maintains the norm. However, this addition is minor, and the overall innovation is minimal.\n\n3. **Numerical Experiments on Layer-wise Decorrelation**: The third section includes numerical results showing the advantages of input decorrelation for training deep networks. Prior work has already demonstrated that input decorrelation accelerates learning in frameworks such as backpropagation and Node Perturbation. Consequently, the main novelty here is the improved performance observed in the Feedback Alignment framework. However, the paper does not provide sufficient explanation or analysis to elucidate why decorrelation significantly enhances Feedback Alignment, nor does the performance match standard backpropagation. This omission limits the impact of this result, rendering it another incremental contribution.\n\nCollectively, these points suggest that the novelty presented in this work does not meet the standards expected for ICLR.\n\nFinally, the discussion section of the paper, while extensive, lacks coherence and leans heavily on speculation. For instance, the portion discussing biological plausibility is unclear in its message. Although the authors correctly note that decorrelation is observed in various neural circuits, the connection between these observations and the mechanisms or theories presented in the paper remains ambiguous. Additionally, the claim that decorrelation improves generalization (line 479) is inadequately supported by the data. The training curves in Figure 4 do not appear to reach convergence, as indicated by the non-saturating test accuracy. This raises doubts about the robustness of the claimed generalization improvements.", "questions": "To better align this work with the standards and expectations of an ICLR paper, the authors need to sharpen their message and clearly articulate the novel contributions of each section. Specifically, the authors may consider addressing the following questions:\n1. How does the consideration of Natural Gradient Descent and the dynamics in curved parameter space contribute to the understanding or derivation of the incremental learning rule? Can this rule be directly derived from the principles of the Natural Gradient theorem?\n2. What differentiates the proposed learning rule from existing frameworks for decorrelating neural activity, such as Recursive Least Squares? If there are differences, what are the specific theoretical or practical advantages of the proposed rule?\n3. Could the authors provide more extensive numerical analyses to demonstrate whether the benefits of their approach extend beyond faster learning to improved generalization? While Natural Gradient theory primarily explains accelerated convergence, do the authors observe any substantial gains in generalization performance?\n\nLastly, I want to note that I enjoyed reading the paper and its broad review, which connects the concept of natural gradients to deep network training. However, this discussion currently feels more suited to an opinion piece.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730649237649}, {"id": "FtWI0EQABc", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1090/Reviewer_HgPq"], "rating": 5, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This submission didactically draws together approximate gradient methods, natural gradient descent, and decorrelation methods, to ultimately improve approximate gradient methods. More specifically the paper shows how natural gradients defined with respect to changes in the loss function highlight the importance of data correlation in determining the alignment between GD and natural gradient updates. With this insight in hand, approximate gradient descent methods are shown to be improved by decorrelating the network activations.", "review_text": "This submission didactically draws together approximate gradient methods, natural gradient descent, and decorrelation methods, to ultimately improve approximate gradient methods. More specifically the paper shows how natural gradients defined with respect to changes in the loss function highlight the importance of data correlation in determining the alignment between GD and natural gradient updates. With this insight in hand, approximate gradient descent methods are shown to be improved by decorrelating the network activations.", "strengths": "I believe this is an original synthesis of concepts resulting in an original and significant modification to approximate gradient methods. \nThis paper is extremely well written, and generally of high clarity and quality throughout.", "weaknesses": "- It is not clear that the alignment to natural gradients is what is underlying the improved performance. \n- No assessment of how successful the decorrelation updates are in aligning regular updates to the natural updates. How big is the contribution of gradient correlations? \n- While the improvements in FA and NP are striking, the experimental results are not finely tuned. It is possible with GD to obtain better results than those shown, and it is unknown if FA and NP will match this improvement. \n- To verify decorrelated FA scales, harder experiments would strengthen the paper (e.g those in Bartunov 2018. Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures for ANNs.)\n- Given biological relevance, showing application and relevance to recurrent networks would improve the paper.", "questions": "- typo optimization 153\n- Assumption on 182 that NGD independent layers, how serious is this?\n- 163, should the expectations on G not be over y too?  \n- 251 skewed is being used non-technically? \n- 260 type orthogonal \n- 264 citation missing, gradient correlations\n- What is the connection between natural gradients and the approximate gradient methods. Why in particular should the approximate methods benefit from natural gradients?\n- The discussion of the orthonormal basis of parameters (page 5) induced by data correlations does not make sense to me. Are the parameters already not orthornormal?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This submission didactically draws together approximate gradient methods, natural gradient descent, and decorrelation methods, to ultimately improve approximate gradient methods. More specifically the paper shows how natural gradients defined with respect to changes in the loss function highlight the importance of data correlation in determining the alignment between GD and natural gradient updates. With this insight in hand, approximate gradient descent methods are shown to be improved by decorrelating the network activations.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "I believe this is an original synthesis of concepts resulting in an original and significant modification to approximate gradient methods. \nThis paper is extremely well written, and generally of high clarity and quality throughout.", "weaknesses": "- It is not clear that the alignment to natural gradients is what is underlying the improved performance. \n- No assessment of how successful the decorrelation updates are in aligning regular updates to the natural updates. How big is the contribution of gradient correlations? \n- While the improvements in FA and NP are striking, the experimental results are not finely tuned. It is possible with GD to obtain better results than those shown, and it is unknown if FA and NP will match this improvement. \n- To verify decorrelated FA scales, harder experiments would strengthen the paper (e.g those in Bartunov 2018. Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures for ANNs.)\n- Given biological relevance, showing application and relevance to recurrent networks would improve the paper.", "questions": "- typo optimization 153\n- Assumption on 182 that NGD independent layers, how serious is this?\n- 163, should the expectations on G not be over y too?  \n- 251 skewed is being used non-technically? \n- 260 type orthogonal \n- 264 citation missing, gradient correlations\n- What is the connection between natural gradients and the approximate gradient methods. Why in particular should the approximate methods benefit from natural gradients?\n- The discussion of the orthonormal basis of parameters (page 5) induced by data correlations does not make sense to me. Are the parameters already not orthornormal?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730594956098}], "openreview_url": "https://openreview.net/forum?id=ogmzNfeRl7", "arxiv_id": "2407.10780", "paper_pdf": "papers/ogmzNfeRl7.pdf", "paper_pdf_sha256": "613e8814a56b2b3850c33e230833b73d32bae9ce91cf3e7c0e94c2af0e3cc46f", "paper_pdf_bytes": 1715785, "paper_pdf_source": "openreview", "code_url": "https://github.com/nasiryahm/CorrelationsRuinGD", "code_repository": "nasiryahm/CorrelationsRuinGD", "code_commit": "00cf47050bbd20e6a153e10bd86e1651524f9779", "code_archive": "repos/ogmzNfeRl7.zip", "code_archive_sha256": "880c79deed5c0065c46f1d09d3963ac9dda1ef9dcae8b87c9eccff103ba647ac", "code_archive_bytes": 80833, "code_file_count": 10, "code_extensions": {".py": 9, ".ipynb": 1}, "github_disk_usage_kb": 329, "github_languages": {"Jupyter Notebook": 98905, "Python": 53694}, "github_archived": false, "github_pushed_at": "2025-06-25T09:21:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/correlations-are-ruining-your-gradient"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Q8cVivO5k5", "year": 2024, "status": "rejected", "title": "Large-Batch, Iteration-Efficient Neural Bayesian Design Optimization", "authors": ["Navid Ansari", "Hans-peter Seidel", "Vahid Babaei"], "authorids": ["~Navid_Ansari1", "~Hans-peter_Seidel1", "~Vahid_Babaei1"], "authors_source": "OpenReview API", "abstract": "Bayesian optimization (BO) provides a powerful framework for optimizing black-box, expensive-to-evaluate functions. It is therefore an attractive tool for engineering design problems, typically involving multiple objectives. Thanks to the rapid advances in fabrication and measurement methods as well as parallel computing infrastructure, querying many design problems can be heavily parallelized. This class of problems challenges BO with an unprecedented setup where it has to deal with very large batches, shifting its focus from sample efficiency to iteration efficiency. We present a novel Bayesian optimization framework specifically tailored to address these limitations. Our key contribution is a highly scalable, sample-based acquisition function that performs a non-dominated sorting of not only the objectives but also their associated uncertainty. We show that our acquisition function in combination with different Bayesian neural network surrogates is effective in data-intensive environments with a minimal number of iterations. We demonstrate the superiority of our method by comparing it with state-of-the-art multi-objective optimizations. We perform our evaluation on two real-world problems - airfoil design and 3D printing -showcasing the applicability and efficiency of our approach. Our code is available at: https://github.com/an-on-ym-ous/lbn_mobo", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "BDsvy4Rsna", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5310/Reviewer_acfL"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents a new method to perform multi-objective BO with the large-batch setting. It proposes to use a Bayesian Neural Network (BNN) created by Deep Ensembles as a surrogate model. It also proposes an NSGA-II based acquisition function that is claimed to be able to scale to large-batch setting better than current acquisition functions. The main idea for the proposed acquisition function (2MD acquisition function) is to simultaneously maximize the predicted objectives and the associated uncertainties, both are given by the BNN surrogate model.\n\nThe method (LBN-MOBO) is evaluated on synthetic functions (1 in the main paper and 4 in the appendix), and 2 real-world problems.", "review_text": "This paper presents a new method to perform multi-objective BO with the large-batch setting. It proposes to use a Bayesian Neural Network (BNN) created by Deep Ensembles as a surrogate model. It also proposes an NSGA-II based acquisition function that is claimed to be able to scale to large-batch setting better than current acquisition functions. The main idea for the proposed acquisition function (2MD acquisition function) is to simultaneously maximize the predicted objectives and the associated uncertainties, both are given by the BNN surrogate model.\n\nThe method (LBN-MOBO) is evaluated on synthetic functions (1 in the main paper and 4 in the appendix), and 2 real-world problems.", "strengths": "- The paper tackles an important problem which is performing multi-objective BO with the large-batch setting.\n- The paper proposes an acquisition function for applying large-batch when performing BO while the current acquisition functions (qEHVI, qParEGO, qNEHVI) struggle, in terms of computation time. The concept of the proposed acquisition function is intuitive: it seems to further encourage explorative behavior, because it also maximizes the uncertainties in the surrogate model.", "weaknesses": "- Some technical details are not described clearly, making it sometimes hard to catch the main idea of the paper. For example, the formal problem statement is not described, the proposed method makes use of only epistemic uncertainty but the concept of epistemic uncertainty is not explained in the Background. The organization of the paper is sometimes a bit confused, for example, the overall process of BO should not be placed in method section.\n- The use of BNN as the surrogate model to enhance the performance is surely promising, however, BNN has many problems. For example, the tuning of its hyperparameters could be another optimization problem or the uncertainty provided by the BNNs could be inaccurate. However, these problems are not discussed in the paper. Furthermore, I don't understand why the proposed method only requires the epistemic uncertainty. There are no motivation, explanation, or insights about this choice and why it work.\n- The proposed acquisition function of optimizing both the prediction and the uncertainty and the usage of NSGA-II to optimize this acquisition function seems to be not too novel for me. The idea is very similar to UCB. There are no deep analysis regarding this proposed acquisition function and why it will work well.\n- The experimental evaluation is very limited. It doesn't compare against other baselines in the main paper. In Section 5.1, it is not convincing to choose the surrogate model (inference method) by using only 1 synthetic experimental result. \n- Related works should mention other types of surrogate models apart from GP and BNN, such as TPE and RF. And also, it is worth mentioning why BNN is preferred over these models.\n- Section 4.1 only covers the modification for Deep Ensembles method. It is not clear how to apply the modification for other inference methods (SGHMC, HMC, DKL, IBNN), so as to compare in Figure 3. \n\nMinor:\n- Authors should use \\citep{} and \\citet{} separately when citing the references.", "questions": "Apart from my comments in the Weaknesses section, the authors can answer the following questions:\n- The concept of the 2MD acquisition function is quite similar to Upper Confidence Bound with a specific exploration factor. It seems that in UCB, both the prediction and the uncertainty are incorporated to compute the acquisition function, while 2MD use the two values as separate objective to optimize. Can the authors point out some differences between the UCB and 2MD?\n- In Figure 1, why there is no surrogate SGHMC, HCM, Deep Ensembles paired with qNEHVI and qParEGO. How many function evaluations in total for this experiment?\n- Is batch size b > 1000 a normal batch size in real-world problem? There seems to be no reference to any applications using such large batch.\n- The two real-world problems use b=15,000 for airfoil problem and b=20,000 for printer problem, on a total of 10 iterations. With such a large number of function evaluations (150,000 and 200,000), can LBN-MOBO outperform Evolutionary Computation methods, e.g., MOEA/D, NSGA-II? These two EC methods are quite powerful for solving multi-objective optimization problems.\n\n[1] B. Paria, K. Kandasamy, and B. Póczos. A flexible framework for multi-objective Bayesian optimization using random scalarizations. In Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, volume 115, 2020\n\n[2] Daulton, Samuel, David Eriksson, Maximilian Balandat, and Eytan Bakshy. \"Multi-objective bayesian optimization over high-dimensional search spaces.\" In Uncertainty in Artificial Intelligence, pp. 507-517. PMLR, 2022.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a new method to perform multi-objective BO with the large-batch setting. It proposes to use a Bayesian Neural Network (BNN) created by Deep Ensembles as a surrogate model. It also proposes an NSGA-II based acquisition function that is claimed to be able to scale to large-batch setting better than current acquisition functions. The main idea for the proposed acquisition function (2MD acquisition function) is to simultaneously maximize the predicted objectives and the associated uncertainties, both are given by the BNN surrogate model.\n\nThe method (LBN-MOBO) is evaluated on synthetic functions (1 in the main paper and 4 in the appendix), and 2 real-world problems.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The paper tackles an important problem which is performing multi-objective BO with the large-batch setting.\n- The paper proposes an acquisition function for applying large-batch when performing BO while the current acquisition functions (qEHVI, qParEGO, qNEHVI) struggle, in terms of computation time. The concept of the proposed acquisition function is intuitive: it seems to further encourage explorative behavior, because it also maximizes the uncertainties in the surrogate model.", "weaknesses": "- Some technical details are not described clearly, making it sometimes hard to catch the main idea of the paper. For example, the formal problem statement is not described, the proposed method makes use of only epistemic uncertainty but the concept of epistemic uncertainty is not explained in the Background. The organization of the paper is sometimes a bit confused, for example, the overall process of BO should not be placed in method section.\n- The use of BNN as the surrogate model to enhance the performance is surely promising, however, BNN has many problems. For example, the tuning of its hyperparameters could be another optimization problem or the uncertainty provided by the BNNs could be inaccurate. However, these problems are not discussed in the paper. Furthermore, I don't understand why the proposed method only requires the epistemic uncertainty. There are no motivation, explanation, or insights about this choice and why it work.\n- The proposed acquisition function of optimizing both the prediction and the uncertainty and the usage of NSGA-II to optimize this acquisition function seems to be not too novel for me. The idea is very similar to UCB. There are no deep analysis regarding this proposed acquisition function and why it will work well.\n- The experimental evaluation is very limited. It doesn't compare against other baselines in the main paper. In Section 5.1, it is not convincing to choose the surrogate model (inference method) by using only 1 synthetic experimental result. \n- Related works should mention other types of surrogate models apart from GP and BNN, such as TPE and RF. And also, it is worth mentioning why BNN is preferred over these models.\n- Section 4.1 only covers the modification for Deep Ensembles method. It is not clear how to apply the modification for other inference methods (SGHMC, HMC, DKL, IBNN), so as to compare in Figure 3. \n\nMinor:\n- Authors should use \\citep{} and \\citet{} separately when citing the references.", "questions": "Apart from my comments in the Weaknesses section, the authors can answer the following questions:\n- The concept of the 2MD acquisition function is quite similar to Upper Confidence Bound with a specific exploration factor. It seems that in UCB, both the prediction and the uncertainty are incorporated to compute the acquisition function, while 2MD use the two values as separate objective to optimize. Can the authors point out some differences between the UCB and 2MD?\n- In Figure 1, why there is no surrogate SGHMC, HCM, Deep Ensembles paired with qNEHVI and qParEGO. How many function evaluations in total for this experiment?\n- Is batch size b > 1000 a normal batch size in real-world problem? There seems to be no reference to any applications using such large batch.\n- The two real-world problems use b=15,000 for airfoil problem and b=20,000 for printer problem, on a total of 10 iterations. With such a large number of function evaluations (150,000 and 200,000), can LBN-MOBO outperform Evolutionary Computation methods, e.g., MOEA/D, NSGA-II? These two EC methods are quite powerful for solving multi-objective optimization problems.\n\n[1] B. Paria, K. Kandasamy, and B. Póczos. A flexible framework for multi-objective Bayesian optimization using random scalarizations. In Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, volume 115, 2020\n\n[2] Daulton, Samuel, David Eriksson, Maximilian Balandat, and Eytan Bakshy. \"Multi-objective bayesian optimization over high-dimensional search spaces.\" In Uncertainty in Artificial Intelligence, pp. 507-517. PMLR, 2022.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698833519351}, {"id": "967WZP44Qj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5310/Reviewer_s2Pz"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper considers the problem of designing Bayesian optimization algorithms for the setting of large batches of evaluations in order to optimize a black-box function. An acquisition function is constructed as multiobjective optimization over multiple predictive mean and uncertainty functions modeled by a deep ensemble. Experiments are performed on two real-world benchmarks.", "review_text": "The paper considers the problem of designing Bayesian optimization algorithms for the setting of large batches of evaluations in order to optimize a black-box function. An acquisition function is constructed as multiobjective optimization over multiple predictive mean and uncertainty functions modeled by a deep ensemble. Experiments are performed on two real-world benchmarks.", "strengths": "- The paper considers an important problem relevant to real-world applications in engineering design.\n\n- I especially like the real world evaluation on two interesting benchmarks: airfoil design and 3D printing. It would be an interesting contribution to the BO community if they are released in the open-source code. \n\n- The idea is simple and works well on the benchmarks.", "weaknesses": "- Although I like the simplicity of the approach, the reasoning behind choosing this instantiation of multiobjective optimization is not entirely clear. Please considering some more analysis about the principles behind the proposed acquisition function. \n\n- Some relevant related work that can be useful to discuss in the paper:\n\t- A very similar idea utilizing multiobjective acquisition function with predicted mean and variance as objectives. \n\n\t[1] Gupta, S., Shilton, A., Rana, S., & Venkatesh, S. (2018, March). Exploiting strategy-space diversity for batch Bayesian optimization. In International conference on artificial intelligence and statistics (pp. 538-547). PMLR.\n\t- There has been a bunch of work on making thompson sampling work for large batch sizes in both continuous and combinatorial design spaces. \n\n\t[2] Hernández-Lobato, J. M., Requeima, J., Pyzer-Knapp, E. O., & Aspuru-Guzik, A. (2017, July). Parallel and distributed Thompson sampling for large-scale accelerated exploration of chemical space. In International conference on machine learning (pp. 1470-1479). PMLR.\n\n\t[3] Deshwal, A., Belakaria, S., & Doppa, J. R. (2021, May). Mercer features for efficient combinatorial Bayesian optimization. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 8, pp. 7210-7218).\n\n\t[4] Vakili, S., Moss, H., Artemev, A., Dutordoir, V., & Picheny, V. (2021). Scalable Thompson sampling using sparse Gaussian process models. Advances in neural information processing systems, 34, 5631-5643.\n\n- Probably a nit, but I think calling deep ensembles as a bayesian neural network is not entirely correct.", "questions": "Please see weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers the problem of designing Bayesian optimization algorithms for the setting of large batches of evaluations in order to optimize a black-box function. An acquisition function is constructed as multiobjective optimization over multiple predictive mean and uncertainty functions modeled by a deep ensemble. Experiments are performed on two real-world benchmarks.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The paper considers an important problem relevant to real-world applications in engineering design.\n\n- I especially like the real world evaluation on two interesting benchmarks: airfoil design and 3D printing. It would be an interesting contribution to the BO community if they are released in the open-source code. \n\n- The idea is simple and works well on the benchmarks.", "weaknesses": "- Although I like the simplicity of the approach, the reasoning behind choosing this instantiation of multiobjective optimization is not entirely clear. Please considering some more analysis about the principles behind the proposed acquisition function. \n\n- Some relevant related work that can be useful to discuss in the paper:\n\t- A very similar idea utilizing multiobjective acquisition function with predicted mean and variance as objectives. \n\n\t[1] Gupta, S., Shilton, A., Rana, S., & Venkatesh, S. (2018, March). Exploiting strategy-space diversity for batch Bayesian optimization. In International conference on artificial intelligence and statistics (pp. 538-547). PMLR.\n\t- There has been a bunch of work on making thompson sampling work for large batch sizes in both continuous and combinatorial design spaces. \n\n\t[2] Hernández-Lobato, J. M., Requeima, J., Pyzer-Knapp, E. O., & Aspuru-Guzik, A. (2017, July). Parallel and distributed Thompson sampling for large-scale accelerated exploration of chemical space. In International conference on machine learning (pp. 1470-1479). PMLR.\n\n\t[3] Deshwal, A., Belakaria, S., & Doppa, J. R. (2021, May). Mercer features for efficient combinatorial Bayesian optimization. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 8, pp. 7210-7218).\n\n\t[4] Vakili, S., Moss, H., Artemev, A., Dutordoir, V., & Picheny, V. (2021). Scalable Thompson sampling using sparse Gaussian process models. Advances in neural information processing systems, 34, 5631-5643.\n\n- Probably a nit, but I think calling deep ensembles as a bayesian neural network is not entirely correct.", "questions": "Please see weaknesses section above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698738230902}, {"id": "X5bLGGrm80", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5310/Reviewer_9JiQ"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper considers a setting in which BO is applied to solve black-box optimization problems where there are multiple objectives and a query is expensive, but the batch size can be extremely large. To address this challenge, the authors propose an acquisition function that is more scalable and takes into account the uncertainty of candidates. Empirically, this BO algorithm is applied to solve two realistic black-box optimization problems in this setting.", "review_text": "The paper considers a setting in which BO is applied to solve black-box optimization problems where there are multiple objectives and a query is expensive, but the batch size can be extremely large. To address this challenge, the authors propose an acquisition function that is more scalable and takes into account the uncertainty of candidates. Empirically, this BO algorithm is applied to solve two realistic black-box optimization problems in this setting.", "strengths": "1.\tThe paper considers a novel black-box optimization setting where there are multiple objectives and the batch size can be very large. The authors empirically observe that contemporary multi-objective batch acquisition functions do not scale well with respect to the batch size.\n2.\tTo solve this issue, the authors propose a modified version of Deep Ensembles to approximate BNN and an acquisition function to maximize both predicted objectives and the uncertainty measure.", "weaknesses": "1.\tContribution is not enough. The innovation of the paper can be summarized into a new predictive model with a minor modification on the original Deep Ensembles model and a new multi-objective batch acquisition function. For the predictive model, the reason for modifying the uncertainty measurement part from aleatory noise to epistemic noise is unclear. Also, the benefit from this change is not verified in the paper. Moreover, the novelty of 2$\\textit{M}$D acquisition compared to the other acquisition functions is not clear either, except for being more scalable.\n2.\tThe empirical results presented are somewhat unconvincing. The reason why the authors choose deep ensembles as a surrogate to test in the two subsequent realistic tasks is that its runtime is shorter and achieves higher hypervolume. However, since this appears in only one experiment, its generalized performances to other tasks are not necessarily better than other surrogates. \n3.\tMore benchmark models should be considered in the two realistic tasks. In these two tasks, only two models are considered, i.e., modified deep ensemble + 2$\\textit{M}$D and dropout + 2$\\textit{M}$D. Therefore, whether the modified deep ensemble + 2MD indeed performs well enough is not clear. It would be great if the authors can also consider more models as benchmarks.\n4.\tLimited theory is developed for this new BO method.", "questions": "1.\tHow is the “time” defined in Figures 1 and 3? \n2.\tHow well does the modified deep ensemble quantify uncertainty, compared to the original deep ensemble?\n3.\tHow does 2MD work? It looks like a key ingredient of the new acquisition is the NSGA-II. However, this is not explained in the main body of the paper. Also, how to implement the acquisition function, one of the most important parts in this paper, is not clearly explained. The only relevant statement is the last three lines in page 6. \n4.\tIn reality, the magnitude or the range of $F$ is usually unknown. How do the authors suggest to balance the tradeoff between the output of the predictive model and the uncertainty?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers a setting in which BO is applied to solve black-box optimization problems where there are multiple objectives and a query is expensive, but the batch size can be extremely large. To address this challenge, the authors propose an acquisition function that is more scalable and takes into account the uncertainty of candidates. Empirically, this BO algorithm is applied to solve two realistic black-box optimization problems in this setting.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1.\tThe paper considers a novel black-box optimization setting where there are multiple objectives and the batch size can be very large. The authors empirically observe that contemporary multi-objective batch acquisition functions do not scale well with respect to the batch size.\n2.\tTo solve this issue, the authors propose a modified version of Deep Ensembles to approximate BNN and an acquisition function to maximize both predicted objectives and the uncertainty measure.", "weaknesses": "1.\tContribution is not enough. The innovation of the paper can be summarized into a new predictive model with a minor modification on the original Deep Ensembles model and a new multi-objective batch acquisition function. For the predictive model, the reason for modifying the uncertainty measurement part from aleatory noise to epistemic noise is unclear. Also, the benefit from this change is not verified in the paper. Moreover, the novelty of 2$\\textit{M}$D acquisition compared to the other acquisition functions is not clear either, except for being more scalable.\n2.\tThe empirical results presented are somewhat unconvincing. The reason why the authors choose deep ensembles as a surrogate to test in the two subsequent realistic tasks is that its runtime is shorter and achieves higher hypervolume. However, since this appears in only one experiment, its generalized performances to other tasks are not necessarily better than other surrogates. \n3.\tMore benchmark models should be considered in the two realistic tasks. In these two tasks, only two models are considered, i.e., modified deep ensemble + 2$\\textit{M}$D and dropout + 2$\\textit{M}$D. Therefore, whether the modified deep ensemble + 2MD indeed performs well enough is not clear. It would be great if the authors can also consider more models as benchmarks.\n4.\tLimited theory is developed for this new BO method.", "questions": "1.\tHow is the “time” defined in Figures 1 and 3? \n2.\tHow well does the modified deep ensemble quantify uncertainty, compared to the original deep ensemble?\n3.\tHow does 2MD work? It looks like a key ingredient of the new acquisition is the NSGA-II. However, this is not explained in the main body of the paper. Also, how to implement the acquisition function, one of the most important parts in this paper, is not clearly explained. The only relevant statement is the last three lines in page 6. \n4.\tIn reality, the magnitude or the range of $F$ is usually unknown. How do the authors suggest to balance the tradeoff between the output of the predictive model and the uncertainty?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA.", "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698696183292}, {"id": "FZjlsGgcYG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5310/Reviewer_2FsD"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose a Multi objective optimisation algorithm with a focus on large batch sizes (up to 1000s of points) and few iterations (as low as 10).\n\nGiven an objective functions $f:\\mathcal{X}\\to \\mathbb{R}^M$ where $\\mathcal{X}\\subset \\mathbb{R}^d$,  the method is a Bayesian model based algorithm, they propose to use Bayesian neural networks as the surrogate model to predict $\\underline{y}=\\hat{f}(x)$ as well as the epistemic uncertainty $Var(\\hat{f(x)})$, these models which can scale to large dataset sizes much more effectively than the more traditional Gaussian processes.\n\nIn order to determine a new batch of points to be evaluated, the authors propose to concatenate the model predictions and uncertainties $[\\hat{f(x)}, Var(\\hat{f}(x))] \\in \\mathbb{R}^{2M}$, which can then be fed into NSGA-II, a popular evolutionary aglgorithm, which can find the set of pareto optimal points $x_1,...,x_B\\in \\mathcal{X}$ that form the pareto front in the augmented output space $\\mathbb{R}^{2M}$. In other words, these are point that are predicted to have high value and/or high uncertainty.\n\nThe authors perform experiments with a range of off-the-shelf Bayesian neural network methods and determine Deep Ensembles to be the best candidate surrogate model.", "review_text": "The authors propose a Multi objective optimisation algorithm with a focus on large batch sizes (up to 1000s of points) and few iterations (as low as 10).\n\nGiven an objective functions $f:\\mathcal{X}\\to \\mathbb{R}^M$ where $\\mathcal{X}\\subset \\mathbb{R}^d$,  the method is a Bayesian model based algorithm, they propose to use Bayesian neural networks as the surrogate model to predict $\\underline{y}=\\hat{f}(x)$ as well as the epistemic uncertainty $Var(\\hat{f(x)})$, these models which can scale to large dataset sizes much more effectively than the more traditional Gaussian processes.\n\nIn order to determine a new batch of points to be evaluated, the authors propose to concatenate the model predictions and uncertainties $[\\hat{f(x)}, Var(\\hat{f}(x))] \\in \\mathbb{R}^{2M}$, which can then be fed into NSGA-II, a popular evolutionary aglgorithm, which can find the set of pareto optimal points $x_1,...,x_B\\in \\mathcal{X}$ that form the pareto front in the augmented output space $\\mathbb{R}^{2M}$. In other words, these are point that are predicted to have high value and/or high uncertainty.\n\nThe authors perform experiments with a range of off-the-shelf Bayesian neural network methods and determine Deep Ensembles to be the best candidate surrogate model.", "strengths": "- Simplicty, elegance.\n  - Bayesian neural networks have become a work horse surrogate model in the Bayesian Optimization community in recent years\n  - NSGA-II is a very popular well established mainstream algorithm in the multi objective community\n  - concatenating predictions and uncertainties to be fed into NSGA-II seems a very reasonable good idea\n  - altogether the method avoids introducing any sophisticated new engineering, and instead opts to intelligently combine established components from the community with some well justified tweaks.\n\n- clearly written, I enjoyed the exposition of related work.\n\n- Section 5.4, running algorithm without using uncertainties I felt was a very nice experiment and cleawrly demonstrated their benefit.", "weaknesses": "I only have minor comments\n\n- I see the authors discuss this in Appendix E but MOO is an very large field and I would be very surprised if batch construction by finding the pareto front of concatenated predictions and uncertainties has not been considered before, (it _seems_ so obvious!), \n- upon first reading, I felt the title was somewhat cluttered.", "questions": "- presumably for small use cases, optimising 2 simple objectives over 2 dimensions batchsize 2, i.e. the ideal use case for any GP-BO, the proposed method would suffer, is there a crossover from where more simple GP-BO methods fail and LBN-MOBO would be best?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a Multi objective optimisation algorithm with a focus on large batch sizes (up to 1000s of points) and few iterations (as low as 10).\n\nGiven an objective functions $f:\\mathcal{X}\\to \\mathbb{R}^M$ where $\\mathcal{X}\\subset \\mathbb{R}^d$,  the method is a Bayesian model based algorithm, they propose to use Bayesian neural networks as the surrogate model to predict $\\underline{y}=\\hat{f}(x)$ as well as the epistemic uncertainty $Var(\\hat{f(x)})$, these models which can scale to large dataset sizes much more effectively than the more traditional Gaussian processes.\n\nIn order to determine a new batch of points to be evaluated, the authors propose to concatenate the model predictions and uncertainties $[\\hat{f(x)}, Var(\\hat{f}(x))] \\in \\mathbb{R}^{2M}$, which can then be fed into NSGA-II, a popular evolutionary aglgorithm, which can find the set of pareto optimal points $x_1,...,x_B\\in \\mathcal{X}$ that form the pareto front in the augmented output space $\\mathbb{R}^{2M}$. In other words, these are point that are predicted to have high value and/or high uncertainty.\n\nThe authors perform experiments with a range of off-the-shelf Bayesian neural network methods and determine Deep Ensembles to be the best candidate surrogate model.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- Simplicty, elegance.\n  - Bayesian neural networks have become a work horse surrogate model in the Bayesian Optimization community in recent years\n  - NSGA-II is a very popular well established mainstream algorithm in the multi objective community\n  - concatenating predictions and uncertainties to be fed into NSGA-II seems a very reasonable good idea\n  - altogether the method avoids introducing any sophisticated new engineering, and instead opts to intelligently combine established components from the community with some well justified tweaks.\n\n- clearly written, I enjoyed the exposition of related work.\n\n- Section 5.4, running algorithm without using uncertainties I felt was a very nice experiment and cleawrly demonstrated their benefit.", "weaknesses": "I only have minor comments\n\n- I see the authors discuss this in Appendix E but MOO is an very large field and I would be very surprised if batch construction by finding the pareto front of concatenated predictions and uncertainties has not been considered before, (it _seems_ so obvious!), \n- upon first reading, I felt the title was somewhat cluttered.", "questions": "- presumably for small use cases, optimising 2 simple objectives over 2 dimensions batchsize 2, i.e. the ideal use case for any GP-BO, the proposed method would suffer, is there a crossover from where more simple GP-BO methods fail and LBN-MOBO would be best?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698692409761}], "openreview_url": "https://openreview.net/forum?id=Q8cVivO5k5", "arxiv_id": "2306.01095", "paper_pdf": "papers/Q8cVivO5k5.pdf", "paper_pdf_sha256": "f97e8cccb755588400aeb1d9819a025066426fca479a6f88ac461925b2467efe", "paper_pdf_bytes": 7788836, "paper_pdf_source": "openreview", "code_url": "https://github.com/AnsariNavid/lbn_mobo", "code_repository": "AnsariNavid/lbn_mobo", "code_commit": "da980750ecea1ed386d9ff54c16f6ae9c279e905", "code_archive": "repos/Q8cVivO5k5.zip", "code_archive_sha256": "04c4f3012720256a5464e793d4d5bcd545baf628f5cc8624e6bf4be97931e9cd", "code_archive_bytes": 211249, "code_file_count": 67, "code_extensions": {".py": 46, ".sh": 19, ".ipynb": 1, ".m": 1}, "github_disk_usage_kb": 264, "github_languages": {"Jupyter Notebook": 314060, "Python": 108599, "Shell": 4986, "MATLAB": 2605}, "github_archived": false, "github_pushed_at": "2023-05-22T14:44:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/large-batch-neural-multi-objective-bayesian"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "NHfSJAWhKTw", "year": 2023, "status": "rejected", "title": "A Closer Look at Self-supervised Lightweight Vision Transformers", "authors": ["Shaoru Wang", "Jin Gao", "Zeming Li", "Weiming Hu"], "authorids": ["~Shaoru_Wang1", "~Jin_Gao1", "~Zeming_Li2", "~Weiming_Hu1"], "authors_source": "OpenReview API", "abstract": "Self-supervised learning on large-scale Vision Transformers (ViTs) as pre-training methods has achieved promising downstream performance. Yet, how much these pre-training paradigms promote lightweight ViTs' performance is considerably less studied. In this work, we mainly develop and benchmark self-supervised pre-training methods, e.g., contrastive-learning-based MoCo-v3, masked-image-modeling-based MAE on image classification tasks, and some downstream dense prediction tasks. We surprisingly find that if proper pre-training is adopted, even vanilla lightweight ViTs show comparable performance on ImageNet to previous SOTA networks with delicate architecture design. We also point out some defects of such pre-training, \\eg, failing to benefit from large-scale pre-training data and showing inferior performance on data-insufficient downstream tasks. Furthermore, we analyze and clearly show the effect of such pre-training by analyzing the properties of the layer representation and attention maps for related models. Finally, based on the above analyses, a distillation strategy during pre-training is developed, which leads to further downstream performance improvement for MAE-based pre-training.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "1IE5PzEDKj", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1538/Reviewer_PEwP"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This is an empirical paper focusing on exploring the task of self-supervised learning on lightweight vision transformers (ViTs). In this paper, the authors observed several discoveries, such as: the tiny model fails to benefit from large-scale pre-training data and shows inferior performance on data-insufficient downstream tasks. They further proposed a distillation strategy during pretraining to improve the representation ability of compact ViTs. Experiments are conducted on ImageNet pre-training and multiple downstream tasks and datasets.", "review_text": "Overall, this is an empirical paper with some trivial observations which are not well supported by the experiments and some are even not new. Also, the organization and writing can be improved significantly in this paper. Thus, I tend to reject it.", "strengths": "### Strengths:\n\n   - It is interesting to explore suitable methods for lightweight vision transformers or other efficient models in the self-supervised learning manner. \n\n   - This paper provided extensive experiments on ImageNet pre-training and multiple downstream datasets and tasks, such as classification, object detection, and segmentation.\n\n\n### Weaknesses:\n\n   - Though this is an empirical paper, the novelty and originality in it are fairly limited, as well as the significance and contribution which are also not strong. The observations that tiny models fail to benefit from large-scale pre-training data, and rely more on the downstream dataset scale are a little bit straightforward and not surprising. The use of knowledge distillation for self-supervised learning on lightweight models also has been proposed for a long time, e.g., on low-bit efficient models [1] and mobile-level models [2].\n\n[1] Shen, Z., Liu, Z., Qin, J., Huang, L., Cheng, K. T., & Savvides, M. (2021). S2-bnn: Bridging the gap between self-supervised real and 1-bit neural networks via guided distribution calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2165-2174).\n\n[2] Fang, Zhiyuan, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. \"SEED: Self-supervised Distillation For Visual Representation.\" In International Conference on Learning Representations. 2021.\n\n   - Some statements in this paper are not well supported, such as “lower layers of the pre-trained models matter more than higher ones if sufficient downstream data is provided, while higher layers matter in data-insufficient downstream tasks.” I think a better and fair comparison design is crucial and also necessary for this argument. The current experiments for this part are not rigorous to prove it.\n\n   - The writing and organization of this paper can also be improved. For instance, it’s not clear to me why Table 1 is located in the early part of the paper. I did not get much information from it and do not know what the insight of this table is.\n\n   - Overall, this paper seems a little bit incremental without providing new conclusions or discoveries over previous literature.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This is an empirical paper focusing on exploring the task of self-supervised learning on lightweight vision transformers (ViTs). In this paper, the authors observed several discoveries, such as: the tiny model fails to benefit from large-scale pre-training data and shows inferior performance on data-insufficient downstream tasks. They further proposed a distillation strategy during pretraining to improve the representation ability of compact ViTs. Experiments are conducted on ImageNet pre-training and multiple downstream tasks and datasets.", "strength_and_weaknesses": "### Strengths:\n\n   - It is interesting to explore suitable methods for lightweight vision transformers or other efficient models in the self-supervised learning manner. \n\n   - This paper provided extensive experiments on ImageNet pre-training and multiple downstream datasets and tasks, such as classification, object detection, and segmentation.\n\n\n### Weaknesses:\n\n   - Though this is an empirical paper, the novelty and originality in it are fairly limited, as well as the significance and contribution which are also not strong. The observations that tiny models fail to benefit from large-scale pre-training data, and rely more on the downstream dataset scale are a little bit straightforward and not surprising. The use of knowledge distillation for self-supervised learning on lightweight models also has been proposed for a long time, e.g., on low-bit efficient models [1] and mobile-level models [2].\n\n[1] Shen, Z., Liu, Z., Qin, J., Huang, L., Cheng, K. T., & Savvides, M. (2021). S2-bnn: Bridging the gap between self-supervised real and 1-bit neural networks via guided distribution calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2165-2174).\n\n[2] Fang, Zhiyuan, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. \"SEED: Self-supervised Distillation For Visual Representation.\" In International Conference on Learning Representations. 2021.\n\n   - Some statements in this paper are not well supported, such as “lower layers of the pre-trained models matter more than higher ones if sufficient downstream data is provided, while higher layers matter in data-insufficient downstream tasks.” I think a better and fair comparison design is crucial and also necessary for this argument. The current experiments for this part are not rigorous to prove it.\n\n   - The writing and organization of this paper can also be improved. For instance, it’s not clear to me why Table 1 is located in the early part of the paper. I did not get much information from it and do not know what the insight of this table is.\n\n   - Overall, this paper seems a little bit incremental without providing new conclusions or discoveries over previous literature.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The clarity of this paper is qualified. Novelty and originality are somewhat limited. Since this paper did not introduce any new and concrete approach, reproducibility is not applicable.", "summary_of_the_review": "Overall, this is an empirical paper with some trivial observations which are not well supported by the experiments and some are even not new. Also, the organization and writing can be improved significantly in this paper. Thus, I tend to reject it.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666777130103}, {"id": "QgSFBaSbw6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1538/Reviewer_UM5C"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to analyze how the lightweight ViTs (e.g., ViT-tiny and DeiT-tiny) perform when using self-supervised pretraining methods. By conducting a variety of experiments, the authors show that: 1) MIM-based methods, like MAE, helps more than contrastive learning based methods, like MOCOv3; 2) Low-level layers matter more than high-level layers when sufficient data for finetuning is available; 3) KD methods help improve the representative ability of high-level layers.\n\nPrevious works mostly focus on large-scale models, which often contain more than 80M parameters. In contrast, these paper aims to reveal how lightweight ViTs behave on both ImageNet and some downstream tasks, which makes this paper much more different than previous works on self-supervised methods.", "review_text": "The intention of this paper is interesting. There are no relevant papers reporting similar conclusions. Though the novelty of this paper is not significant enough, regarding the thorough experiments that have been done, I give a score of 5 at this moment. I would like to lift the rating if the authors can more clearly explain the significance of the contributions.", "strengths": "Strength\n- Analyzing the performance of lightweight ViTs in self-supervised learning manners is important in that in most cases it is not appropriate to use large-scale models but only tiny-sized ones. This paper clearly shows how lightweight ViTs behave under different settings and evaluates the pretrained models on downstream tasks.\n- Authors show that low-level layers are more important than high-level layers after pretraining. This is an important signal for the development of SSL methods in future.\n- Distilling the knowledge from a large model to lightweight ones is an interesting topic. This paper provides an effective way to do so and analyze how to do distillation helps more for the lifting the model performance.\n\nWeaknesses\n- Only two SSL methods are selected for presentation. One is MOCOv3 and the other one is MAE. It would be better if two methods are selected for explanation for each type of SSL method. The conclusion would be more convincing.\n- From Table 3, we see that the pre-training benefits little from large-scale data.  With only 10% of the ImageNet 1k data are provided, the classification performance after finetuning is already good. I suppose there are some explanations on the reasons but obviously there is not.\n- The analysis is interesting. I am looking forward to taking some message about how to design better network architectures or how to develop more advanced self-supervised learning methods.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper aims to analyze how the lightweight ViTs (e.g., ViT-tiny and DeiT-tiny) perform when using self-supervised pretraining methods. By conducting a variety of experiments, the authors show that: 1) MIM-based methods, like MAE, helps more than contrastive learning based methods, like MOCOv3; 2) Low-level layers matter more than high-level layers when sufficient data for finetuning is available; 3) KD methods help improve the representative ability of high-level layers.\n\nPrevious works mostly focus on large-scale models, which often contain more than 80M parameters. In contrast, these paper aims to reveal how lightweight ViTs behave on both ImageNet and some downstream tasks, which makes this paper much more different than previous works on self-supervised methods.", "strength_and_weaknesses": "Strength\n- Analyzing the performance of lightweight ViTs in self-supervised learning manners is important in that in most cases it is not appropriate to use large-scale models but only tiny-sized ones. This paper clearly shows how lightweight ViTs behave under different settings and evaluates the pretrained models on downstream tasks.\n- Authors show that low-level layers are more important than high-level layers after pretraining. This is an important signal for the development of SSL methods in future.\n- Distilling the knowledge from a large model to lightweight ones is an interesting topic. This paper provides an effective way to do so and analyze how to do distillation helps more for the lifting the model performance.\n\nWeaknesses\n- Only two SSL methods are selected for presentation. One is MOCOv3 and the other one is MAE. It would be better if two methods are selected for explanation for each type of SSL method. The conclusion would be more convincing.\n- From Table 3, we see that the pre-training benefits little from large-scale data.  With only 10% of the ImageNet 1k data are provided, the classification performance after finetuning is already good. I suppose there are some explanations on the reasons but obviously there is not.\n- The analysis is interesting. I am looking forward to taking some message about how to design better network architectures or how to develop more advanced self-supervised learning methods.", "clarity,_quality,_novelty_and_reproducibility": "This paper is well written and easy to understand. This reveals some secrets behind SSL on visual recognition using lightweight ViT models. The novelty is ok but not significant enough for publication.", "summary_of_the_review": "The intention of this paper is interesting. There are no relevant papers reporting similar conclusions. Though the novelty of this paper is not significant enough, regarding the thorough experiments that have been done, I give a score of 5 at this moment. I would like to lift the rating if the authors can more clearly explain the significance of the contributions.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666492664909}, {"id": "ub_hCVK5X-6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1538/Reviewer_nNkX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper presents an analysis of two self-supervised training methods for lightweight vision transformers. The methods are MAE (a masking method) and MoCoV3 (a contrastive method). It shows that although MAE appears stronger when training (w/o labels) and evaluating on ImageNet, the roles reverse on downstream tasks. It analyzes the differences of these two networks and compares them a supervised method (DeitT), by looking at layer-by-layer similarity metrics, as well as analysis of the locality vs. globality of the attention maps. Finally it proposes to use distillation from a large-scale MAE model to improve the tiny MAE. It does this with significant improvements to downstream tasks.", "review_text": "The papers presents a lot of analysis around lightweight ViTs, and shows the efficacy of student/teacher distillation for mask-based self-supervised pre-training. The paper presents few novelties in terms of technique and method, but the quality and novelty of the analysis and strength of final distilled results are solid contributions. The topic is of broad interest, so many could benefit from this work. Balancing the lack of novelties with the potential benefits it is close to borderline for me. However, I do lean in favor of acceptance.", "strengths": "Strengths:\n- This is a pretty in depth analysis of the difference between a MAE- and a MoCo-based self-supervised ViT.\n- The distillation of the MAE-Tiny from MAE-Base leads to significant improvements over the MAE-Tiny without teacher guidance. The final results compared to supervised pretraining is strong and a significant achievement.\n\nWeaknesses:\n- The paper mostly presents analysis and new applications of old techniques. It is pretty light on novelties though.\n- Not all analysis is tied back to experimental results, which means it's not clear if we can draw actionable conclusions from it. If the analyses had led to more actionable changes that resulted in improvements, the case would have been stronger.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents an analysis of two self-supervised training methods for lightweight vision transformers. The methods are MAE (a masking method) and MoCoV3 (a contrastive method). It shows that although MAE appears stronger when training (w/o labels) and evaluating on ImageNet, the roles reverse on downstream tasks. It analyzes the differences of these two networks and compares them a supervised method (DeitT), by looking at layer-by-layer similarity metrics, as well as analysis of the locality vs. globality of the attention maps. Finally it proposes to use distillation from a large-scale MAE model to improve the tiny MAE. It does this with significant improvements to downstream tasks.", "strength_and_weaknesses": "Strengths:\n- This is a pretty in depth analysis of the difference between a MAE- and a MoCo-based self-supervised ViT.\n- The distillation of the MAE-Tiny from MAE-Base leads to significant improvements over the MAE-Tiny without teacher guidance. The final results compared to supervised pretraining is strong and a significant achievement.\n\nWeaknesses:\n- The paper mostly presents analysis and new applications of old techniques. It is pretty light on novelties though.\n- Not all analysis is tied back to experimental results, which means it's not clear if we can draw actionable conclusions from it. If the analyses had led to more actionable changes that resulted in improvements, the case would have been stronger.", "clarity,_quality,_novelty_and_reproducibility": "Clarity\nThe paper is fairly clear, although there are times I wish ablations had been kept separate. For instance, \"we adopt the layer-wise attention-based distillation\" and then two paragraphs later \"For simplicity, we only apply distillation on the attention maps of the last layer\". This causes some confusion and it's easier for the reader if you present one main method, and leave ablations to separate sections.\n\nQuality\nThe analysis and results are sound and of good quality. The experimental evaluation is varied and relevant.\n\nNovelty\nAs for novelties, this paper mostly presents analysis and the final distilled MAE uses standard methods. The novelty is in applying these to more lightweight ViTs.\n\nReproducibility\nAs for reproducibility, it is unclear if source code is forthcoming. \n\nSuggestions:\n- Adding a training data column to tables 4 and 6  would help to make it clear that DeiT used labels and the other methods did not. Having to go through the text to realize this takes time.\n- Why not include DeiT in table 6 as well? The distilled MAE actually outperforms the supervised method on several benchmarks. This is significant and should be highlighted.", "summary_of_the_review": "The papers presents a lot of analysis around lightweight ViTs, and shows the efficacy of student/teacher distillation for mask-based self-supervised pre-training. The paper presents few novelties in terms of technique and method, but the quality and novelty of the analysis and strength of final distilled results are solid contributions. The topic is of broad interest, so many could benefit from this work. Balancing the lack of novelties with the potential benefits it is close to borderline for me. However, I do lean in favor of acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666392161713}], "openreview_url": "https://openreview.net/forum?id=NHfSJAWhKTw", "arxiv_id": "2205.14443", "paper_pdf": "papers/NHfSJAWhKTw.pdf", "paper_pdf_sha256": "677cdd497dadddf393f8d5f6e86987d727cf0b78100afedafe921868f58f4856", "paper_pdf_bytes": 1274963, "paper_pdf_source": "openreview", "code_url": "https://github.com/wangsr126/MAE-Lite", "code_repository": "wangsr126/MAE-Lite", "code_commit": "a5d1a74865b661c0e3a91e3ab1d5331ad1deac9c", "code_archive": "repos/NHfSJAWhKTw.zip", "code_archive_sha256": "66bb39272f229210ea715058df81c31d6795938d61cd96020aa8f8a6127c7910", "code_archive_bytes": 753174, "code_file_count": 238, "code_extensions": {".py": 229, ".sh": 8, ".ipynb": 1}, "github_disk_usage_kb": 611, "github_languages": {"Python": 1158942, "Shell": 9347, "Jupyter Notebook": 2581, "Makefile": 198}, "github_archived": false, "github_pushed_at": "2025-03-02T15:39:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-closer-look-at-self-supervised-lightweight"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZzwfldvDLpC", "year": 2022, "status": "rejected", "title": "Let Your Heart Speak in its Mother Tongue: Multilingual Captioning of Cardiac Signals", "authors": ["Dani Kiyasseh", "Tingting Zhu", "David A. Clifton"], "authorids": ["~Dani_Kiyasseh1", "~Tingting_Zhu1", "~David_A._Clifton1"], "authors_source": "OpenReview API", "abstract": "Cardiac signals convey a significant amount of information about the health status of a patient. Upon recording these signals, cardiologists are expected to manually generate an accompanying report to share with physicians and patients. Generating these reports, however, can be time-consuming and error-prone, while also exhibiting a high degree of intra- and inter-physician variability. To address this, we design a neural, multilingual, cardiac signal captioning framework. In the process, we propose a discriminative multilingual representation learning method, RTLP, which randomly replaces tokens with those from a different language and tasks a network with identifying the language of all tokens. We show that RTLP performs on par with state-of-the-art pre-training methods such as MLM and MARGE, while generating more clinically accurate reports than MLM. We also show that, with RTLP, multilingual fine-tuning can be preferable to its monolingual counterpart, a phenomenon we refer to as the \\textit{blessing of multilinguality}.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ekRSXGXbYwm", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper339/Reviewer_KZQE"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors of the paper have introduced a multilingual cardiac signal captioning framework. The authors have proposed a neural framework that generates captions for the cardiac signals in multiple languages simultaneously. The authors add an auxiliary task of identifying the language of some of the tokens to improve the performance of the decoder. The proposed framework achieves on par performance with state of the art pre-training methods. ", "review_text": "The strengths of the paper: \n1. The paper is well presented and the problem introduced in the paper is interesting as well as important. \n2. The authors have tried multiple pre-training methods to train their decoder and use a relevant auxiliary task to supplement the performance of the proposed model.\n3. The reports generated by the proposed framework seems to be quite coherent and similar to the gold standard captions.\n\nThe weaknesses of the paper: \n1. The authors have mentioned that though there has been work around EEG captioning (Biswal et al 2019, 2020), the previous papers did not explore the multilingual captioning of the cardiac signals. But in the case of the proposed framework the gold-label captions for languages other than English are trained on captions generated via Google Translate API. So the ceiling of the decoder model is as good as the Translate API itself. If that is the case, wouldn't it be better to just generate the English caption and then translate them via the API. The cited paper (Conneau et al) uses such a strategy to generate for different languages to augment the dataset and generate new translations which the model might not have seen. \n2. I do not understand the use case of generating all languages at once. The more suitable framework would be where the report is generated as per the requirement in which the source language is provided as an input to the decoder (<source_lg such as en, fr etc.><START>). \n3. If MLM performs as good or better, in most cases, for generating the captions. What would justify the use-case of RTLP? It is an interesting approach but I do not see an added advantage of the proposed model.\n4. The application is highly interesting but almost all parts of the proposed framework/models have been explored in a similar setting except for the multilingual aspect, which I am having a hard time understanding the use case of. In the proposed problem formulation, I would try to learn a really good cardiac signal captioning 'en' model and then either use the API itself or train a translator model. This would reduce the model parameters and make the whole framework more efficient.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors of the paper have introduced a multilingual cardiac signal captioning framework. The authors have proposed a neural framework that generates captions for the cardiac signals in multiple languages simultaneously. The authors add an auxiliary task of identifying the language of some of the tokens to improve the performance of the decoder. The proposed framework achieves on par performance with state of the art pre-training methods. ", "main_review": "The strengths of the paper: \n1. The paper is well presented and the problem introduced in the paper is interesting as well as important. \n2. The authors have tried multiple pre-training methods to train their decoder and use a relevant auxiliary task to supplement the performance of the proposed model.\n3. The reports generated by the proposed framework seems to be quite coherent and similar to the gold standard captions.\n\nThe weaknesses of the paper: \n1. The authors have mentioned that though there has been work around EEG captioning (Biswal et al 2019, 2020), the previous papers did not explore the multilingual captioning of the cardiac signals. But in the case of the proposed framework the gold-label captions for languages other than English are trained on captions generated via Google Translate API. So the ceiling of the decoder model is as good as the Translate API itself. If that is the case, wouldn't it be better to just generate the English caption and then translate them via the API. The cited paper (Conneau et al) uses such a strategy to generate for different languages to augment the dataset and generate new translations which the model might not have seen. \n2. I do not understand the use case of generating all languages at once. The more suitable framework would be where the report is generated as per the requirement in which the source language is provided as an input to the decoder (<source_lg such as en, fr etc.><START>). \n3. If MLM performs as good or better, in most cases, for generating the captions. What would justify the use-case of RTLP? It is an interesting approach but I do not see an added advantage of the proposed model.\n4. The application is highly interesting but almost all parts of the proposed framework/models have been explored in a similar setting except for the multilingual aspect, which I am having a hard time understanding the use case of. In the proposed problem formulation, I would try to learn a really good cardiac signal captioning 'en' model and then either use the API itself or train a translator model. This would reduce the model parameters and make the whole framework more efficient.", "summary_of_the_review": "The paper proposed a solution to an interesting problem but as mentioned in the weakness section above I could not comprehend clearly the use-case of the multilingual decoder. But I would be highly interested in discussing with the authors and other reviewers during the rebuttal period if I might be missing the aspect of using multilingual data for the framework. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636321202342}, {"id": "No8x_RPCDoD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper339/Reviewer_VuBG"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The goal of this paper is to develop an approach for generating multilingual ECG reports. The authors propose a new multilingual pretraining method in which tokens are randomly replaced with those from a different language, and the model must learn to identify the language of all tokens. The method – RTLP – performs similarly to MLM, ELECTRA, and MARGE according to BLEU-1, METEOR, and ROUGE-L metrics on generation of ECG reports, and the authors qualitatively assess that the generated reports are clinically accurate. The paper compares monolingual and multilingual versions of the models and find that RTLP benefits from multilingual training.", "review_text": "Strengths: This paper is very well written. Multilingual text generation is an interesting task, and more multilingual work is needed in the ML for health space. \n\nWeaknesses: I have several concerns about the clinical utility of this task as well as the evaluation approach.  \n-\tFirst of all, I think clarification is needed to describe the utility of the task setup. Why is the task framed as generation of the ECG report rather than framing the task as multi-label classification or slot-filling, especially given the known faithfulness issues with text generation? There are some existing approaches for automatic ECG interpretation. How does this work fit into the existing approaches? A portion of the ECG reports from the PTB-XL dataset are actually automatically generated (See Data Acquisition under https://physionet.org/content/ptb-xl/1.0.1/). Do you filter out those notes during evaluation? How does your method compare to those automatically generated reports?\n-\tA major claim in the paper is that RTLP generates more clinically accurate reports than MLM, yet the only analysis in the paper related to this is a qualitative analysis of a single report. A more systematic analysis of the quality of generation would be useful to support the claim made in the appendix. Can you ask clinicians to evaluate the utility of the generated reports or evaluate clinical utility by using the generated reports to predict conditions identifiable from the ECG? I think that it’s fine that the RTLP method performs comparable to existing methods, but I am not sure from the current paper what the utility of using RTLP is. \n-\tMore generally, I think that this paper is trying to do two things at once – present new methods for multilingual pretraining while also developing a method of ECG captioning. If the emphasis is on the former, then I would expect to see evaluation against other multilingual pretraining setups such as the Unicoder (Huang 2019a). If the core contribution is the latter, then clinical utility of the method as well as comparison to baselines for ECG captioning (or similar methods) is especially important. \n-\tI’m a bit confused as to why the diversity of the generated reports is emphasized during evaluation. While I agree that the generated reports should be faithful to the associated ECG, diversity may not actually be necessary metric to aim for in a medical context. For instance, if many of the reports are normal, you would want similar reports for each normal ECG (i.e. low diversity). \n-\tMy understanding is that reports are generated in other languages using Google Translate. While this makes sense to generate multilingual reports for training, it seems a bit strange to then evaluate your model performance on these silver-standard noisy reports. Do you have a held out set of gold standard reports in different languages for evaluation (other than German)? \n\nOther Comments:\n-\tWhy do you only consider ECG segments with one label assigned to them? I would expect that the associated reports would be significantly easier than including all reports. \n-\tYou might consider changing the terminology from “cardiac arrythmia” categories to something broader since hypertrophy (one of the categories) is not technically a cardiac arrythmia (although it can be detected via ECG & it does predispose you to them)\n-\tI think it’d be helpful to include an example of some of the tokens that are sampled during pretraining using your semantically similar strategy for selecting target tokens. How well does this work in languages that have very different syntactic structures compared to the source language?\n-\tDo you pretrain the cardiac signal representation learning model on the entire dataset or just the training set? If the entire set, how well does this generalize to setting where you don’t have the associated labels?\n-\tWhat kind of tokenization is used in the model? Which Spacy tokenizer?\n-\tIt’d be helpful to reference the appendix when describing the setup in section 3/5 so that the reader knows that more detailed architecture information is there.\n-\tI’d be interested to know if other multilingual pretraining setups also struggle with Greek. \n-\tIt’d be helpful to show the original ECG report with punctuation + make the ECG larger so that they are easier to read\n-\tWhy do you think RTLP benefits from fine-tuning on multiple languages, but MARGE does not?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The goal of this paper is to develop an approach for generating multilingual ECG reports. The authors propose a new multilingual pretraining method in which tokens are randomly replaced with those from a different language, and the model must learn to identify the language of all tokens. The method – RTLP – performs similarly to MLM, ELECTRA, and MARGE according to BLEU-1, METEOR, and ROUGE-L metrics on generation of ECG reports, and the authors qualitatively assess that the generated reports are clinically accurate. The paper compares monolingual and multilingual versions of the models and find that RTLP benefits from multilingual training.", "main_review": "Strengths: This paper is very well written. Multilingual text generation is an interesting task, and more multilingual work is needed in the ML for health space. \n\nWeaknesses: I have several concerns about the clinical utility of this task as well as the evaluation approach.  \n-\tFirst of all, I think clarification is needed to describe the utility of the task setup. Why is the task framed as generation of the ECG report rather than framing the task as multi-label classification or slot-filling, especially given the known faithfulness issues with text generation? There are some existing approaches for automatic ECG interpretation. How does this work fit into the existing approaches? A portion of the ECG reports from the PTB-XL dataset are actually automatically generated (See Data Acquisition under https://physionet.org/content/ptb-xl/1.0.1/). Do you filter out those notes during evaluation? How does your method compare to those automatically generated reports?\n-\tA major claim in the paper is that RTLP generates more clinically accurate reports than MLM, yet the only analysis in the paper related to this is a qualitative analysis of a single report. A more systematic analysis of the quality of generation would be useful to support the claim made in the appendix. Can you ask clinicians to evaluate the utility of the generated reports or evaluate clinical utility by using the generated reports to predict conditions identifiable from the ECG? I think that it’s fine that the RTLP method performs comparable to existing methods, but I am not sure from the current paper what the utility of using RTLP is. \n-\tMore generally, I think that this paper is trying to do two things at once – present new methods for multilingual pretraining while also developing a method of ECG captioning. If the emphasis is on the former, then I would expect to see evaluation against other multilingual pretraining setups such as the Unicoder (Huang 2019a). If the core contribution is the latter, then clinical utility of the method as well as comparison to baselines for ECG captioning (or similar methods) is especially important. \n-\tI’m a bit confused as to why the diversity of the generated reports is emphasized during evaluation. While I agree that the generated reports should be faithful to the associated ECG, diversity may not actually be necessary metric to aim for in a medical context. For instance, if many of the reports are normal, you would want similar reports for each normal ECG (i.e. low diversity). \n-\tMy understanding is that reports are generated in other languages using Google Translate. While this makes sense to generate multilingual reports for training, it seems a bit strange to then evaluate your model performance on these silver-standard noisy reports. Do you have a held out set of gold standard reports in different languages for evaluation (other than German)? \n\nOther Comments:\n-\tWhy do you only consider ECG segments with one label assigned to them? I would expect that the associated reports would be significantly easier than including all reports. \n-\tYou might consider changing the terminology from “cardiac arrythmia” categories to something broader since hypertrophy (one of the categories) is not technically a cardiac arrythmia (although it can be detected via ECG & it does predispose you to them)\n-\tI think it’d be helpful to include an example of some of the tokens that are sampled during pretraining using your semantically similar strategy for selecting target tokens. How well does this work in languages that have very different syntactic structures compared to the source language?\n-\tDo you pretrain the cardiac signal representation learning model on the entire dataset or just the training set? If the entire set, how well does this generalize to setting where you don’t have the associated labels?\n-\tWhat kind of tokenization is used in the model? Which Spacy tokenizer?\n-\tIt’d be helpful to reference the appendix when describing the setup in section 3/5 so that the reader knows that more detailed architecture information is there.\n-\tI’d be interested to know if other multilingual pretraining setups also struggle with Greek. \n-\tIt’d be helpful to show the original ECG report with punctuation + make the ECG larger so that they are easier to read\n-\tWhy do you think RTLP benefits from fine-tuning on multiple languages, but MARGE does not?\n\n", "summary_of_the_review": "Overall, the combined unclear clinical utility, lack of support for a major claim in the paper (that RTLP generates more clinically accurate reports), and concerns about the evaluation strategies lead me to recommend this paper for rejection.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636067327582}, {"id": "R5SqYYmumRF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper339/Reviewer_HBxE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to build a multilingual cardiac signal captioning system to generate ECG reports, which describe the clinical findings in the input of electrocardiogram (ECG) signals. In particular, the proposed system can generate desirable and fluent reports in multiple languages, i.e., German, Greek, English, Spanish, French, Italian, and Portuguese. The experiments on a public dataset verify the effectiveness of the proposed approach, which performs on par with state-of-the-art language pre-training methods.", "review_text": "Strengths:\n1. This paper is clearly written. The paper is easy to follow and understand.\n2. The targeted problems, i.e., cardiac signal captioning and multilingual captioning, are novel and important in both artificial intelligence and clinical medicine.\n3. The proposed multilingual cardiac signal captioning system is well-motivated, novel, and interesting.\n4. The experiments and analysis are extensive and solid.\n\nWeaknesses:\n1. The presentation can be further improved. In the Abstract and Introduction, can you give more explanations about \"Generating these reports, however, can be time-consuming and error-prone, while also exhibiting a high degree of intra- and inter-physician variability\"? For example, why generating these reports can be error-prone? What problems will be brought by the high degree of intra- and inter-physician variability (can you give some examples in the Introduction)? \n\n2. Although the targeted problems are important and novel, after reading this paper, I am still confused about how your proposed method is related to this cardiac signal captioning. I think the proposed approach can be used for conventional image captioning as well. In other words, have you solved the challenges and problems that are unique to the multilingual cardiac signal captioning?\n\n3. The experiment should be improved. Firstly, the evaluation metrics used in this paper, e.g.,  BLEU and ROUGE, are all general metrics for text generation tasks. So, it is unclear why your proposed approach can bring improvements. In what aspects can the proposed method improve the performance of the multilingual cardiac signal captioning? Secondly, in Table 1, why is your proposed method lower than the baselines in some settings? Can you give more explanations? Thirdly, the Google Translate model is not specifically designed for biomedical texts, so you can give more analysis of the Google Translate model. \n\n4. The paper is written in an optimistic tone that leads the reader to assume the proposed approach is rather good. However, I am more interested in knowing if the approach brings errors? And what type of errors does it bring? And why?\n\n5. The related work is insufficient. It is suggested to add more discussions about the report generation for other types of medical signals, e.g., chest X-rays, which have been widely explored in existing papers [1][2][3][4][5][6][7].\n\nMissing References:\n\n[1] On the Automatic Generation of Medical Imaging Reports. In ACL, 2018\n\n[2] Show, Describe and Conclude: On Exploiting the Structure Information of Chest X-ray Reports. In ACL, 2019.\n\n[3] When Radiology Report Generation Meets Knowledge Graph. In AAAI, 2020.\n\n[4] Exploring and Distilling Posterior and Prior Knowledge for Radiology Report Generation. In CVPR, 2021.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to build a multilingual cardiac signal captioning system to generate ECG reports, which describe the clinical findings in the input of electrocardiogram (ECG) signals. In particular, the proposed system can generate desirable and fluent reports in multiple languages, i.e., German, Greek, English, Spanish, French, Italian, and Portuguese. The experiments on a public dataset verify the effectiveness of the proposed approach, which performs on par with state-of-the-art language pre-training methods.", "main_review": "Strengths:\n1. This paper is clearly written. The paper is easy to follow and understand.\n2. The targeted problems, i.e., cardiac signal captioning and multilingual captioning, are novel and important in both artificial intelligence and clinical medicine.\n3. The proposed multilingual cardiac signal captioning system is well-motivated, novel, and interesting.\n4. The experiments and analysis are extensive and solid.\n\nWeaknesses:\n1. The presentation can be further improved. In the Abstract and Introduction, can you give more explanations about \"Generating these reports, however, can be time-consuming and error-prone, while also exhibiting a high degree of intra- and inter-physician variability\"? For example, why generating these reports can be error-prone? What problems will be brought by the high degree of intra- and inter-physician variability (can you give some examples in the Introduction)? \n\n2. Although the targeted problems are important and novel, after reading this paper, I am still confused about how your proposed method is related to this cardiac signal captioning. I think the proposed approach can be used for conventional image captioning as well. In other words, have you solved the challenges and problems that are unique to the multilingual cardiac signal captioning?\n\n3. The experiment should be improved. Firstly, the evaluation metrics used in this paper, e.g.,  BLEU and ROUGE, are all general metrics for text generation tasks. So, it is unclear why your proposed approach can bring improvements. In what aspects can the proposed method improve the performance of the multilingual cardiac signal captioning? Secondly, in Table 1, why is your proposed method lower than the baselines in some settings? Can you give more explanations? Thirdly, the Google Translate model is not specifically designed for biomedical texts, so you can give more analysis of the Google Translate model. \n\n4. The paper is written in an optimistic tone that leads the reader to assume the proposed approach is rather good. However, I am more interested in knowing if the approach brings errors? And what type of errors does it bring? And why?\n\n5. The related work is insufficient. It is suggested to add more discussions about the report generation for other types of medical signals, e.g., chest X-rays, which have been widely explored in existing papers [1][2][3][4][5][6][7].\n\nMissing References:\n\n[1] On the Automatic Generation of Medical Imaging Reports. In ACL, 2018\n\n[2] Show, Describe and Conclude: On Exploiting the Structure Information of Chest X-ray Reports. In ACL, 2019.\n\n[3] When Radiology Report Generation Meets Knowledge Graph. In AAAI, 2020.\n\n[4] Exploring and Distilling Posterior and Prior Knowledge for Radiology Report Generation. In CVPR, 2021.", "summary_of_the_review": "The paper is well-written and the motivation sounds reasonable. The targeted problems, i.e., cardiac signal captioning and multilingual captioning, are novel and important. The presented approach is novel and interesting. Thus, I tend to accept this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635860707310}, {"id": "MPEaJFq8orr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper339/Reviewer_VEwE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a new report generation system for cardiac signals. It applied replaced token language prediction (RTLP) settings to improve the report generation performance. Experiments show that the proposed RTLP framework can achieve comparable performance with SOTA models. Extensive analyses were conducted to examine the diversity and the performance in multilanguage settings. ", "review_text": "This paper proposed a novel multilingual pretraining language model settings. The idea is very interesting and the experiment setting is solid.   This paper conducted many result analyses which are helpful to understand the model's performance. \n\n1. Using BLUE and ROUGE as the metrics is not enough in the evaluation of clinical report generation. The precision of content is critical in clinical reports. High BLUE/ROUGE does not represent the report is well written. In my opinion,  the performance of clinical text generation should be validated based on its content, although it is very hard and requires extensive manual work. \n\n2. The gold-standard reports for non-English languages came from Google translation. This setting may lead to evaluation bias. It would be better to conduct a sensitivity analysis by using other translation tools.   \n\n3. It is not clear which network structure was used to represent the cardiac signals? Convolutional NN? Transformor? Other time series models?\n\n4. It is not clear how to build the categorical distribution when selecting the semantic-similar target language tokens (page 4).\n\n5. The \"blessing of multilinguality\"  comes from the low performance of RTLP in a single language. Based on this low baseline, the improvement from multilanguage has limited value. (RTLP + multilanguage can not beat MARGE).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a new report generation system for cardiac signals. It applied replaced token language prediction (RTLP) settings to improve the report generation performance. Experiments show that the proposed RTLP framework can achieve comparable performance with SOTA models. Extensive analyses were conducted to examine the diversity and the performance in multilanguage settings. ", "main_review": "This paper proposed a novel multilingual pretraining language model settings. The idea is very interesting and the experiment setting is solid.   This paper conducted many result analyses which are helpful to understand the model's performance. \n\n1. Using BLUE and ROUGE as the metrics is not enough in the evaluation of clinical report generation. The precision of content is critical in clinical reports. High BLUE/ROUGE does not represent the report is well written. In my opinion,  the performance of clinical text generation should be validated based on its content, although it is very hard and requires extensive manual work. \n\n2. The gold-standard reports for non-English languages came from Google translation. This setting may lead to evaluation bias. It would be better to conduct a sensitivity analysis by using other translation tools.   \n\n3. It is not clear which network structure was used to represent the cardiac signals? Convolutional NN? Transformor? Other time series models?\n\n4. It is not clear how to build the categorical distribution when selecting the semantic-similar target language tokens (page 4).\n\n5. The \"blessing of multilinguality\"  comes from the low performance of RTLP in a single language. Based on this low baseline, the improvement from multilanguage has limited value. (RTLP + multilanguage can not beat MARGE).", "summary_of_the_review": "The proposed model is novel and interesting. While the evaluation metrics in clinical text generation are not persuasive. And some of the important details were missing in the current draft. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635782934743}], "openreview_url": "https://openreview.net/forum?id=ZzwfldvDLpC", "arxiv_id": "2103.11011", "paper_pdf": "papers/ZzwfldvDLpC.pdf", "paper_pdf_sha256": "eabcf26638e62b32986478cf48f85ea5299213d4c7a74a3bc14133df4bbdcc72", "paper_pdf_bytes": 4227129, "paper_pdf_source": "openreview", "code_url": "https://github.com/danikiyasseh/RTLP", "code_repository": "danikiyasseh/RTLP", "code_commit": "e52f3e2488ba2d5eb6be0fa011daac002e736c16", "code_archive": "repos/ZzwfldvDLpC.zip", "code_archive_sha256": "5610078a8f05b39912a7cbbe2438d94353f900eb4f5f74920cf88e149e0f1a78", "code_archive_bytes": 1567052, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 2341, "github_languages": {"Python": 429536, "HTML": 1}, "github_archived": false, "github_pushed_at": "2022-08-25T16:43:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/let-your-heart-speak-in-its-mother-tongue"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "j5d9qacxdZa", "year": 2021, "status": "rejected", "title": "Energy-Based Models for Continual Learning", "authors": ["Shuang Li", "Yilun Du", "Gido Martijn van de Ven", "Antonio Torralba", "Igor Mordatch"], "authorids": ["~Shuang_Li5", "~Yilun_Du1", "~Gido_Martijn_van_de_Ven1", "~Antonio_Torralba1", "~Igor_Mordatch4"], "authors_source": "OpenReview API", "abstract": "We motivate Energy-Based Models (EBMs) as a promising model class for continual learning problems. Instead of tackling continual learning via the use of external memory, growing models, or regularization, EBMs have a natural way to support a dynamically-growing number of tasks and classes and less interference with old tasks. We show that EBMs are adaptable to a more general continual learning setting where the data distribution changes without the notion of explicitly delineated tasks. We also find that EBMs outperform the baseline methods by a large margin on several continual learning benchmarks. These observations point towards EBMs as a class of models naturally inclined towards the continual learning regime.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "g3L9ZPZ84I1", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1451/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper explores the usage of EBMs in continual learning for classification. Although the application of EBMs in continual learning is novel, the general idea is a special case of the usage of EBMs for structured prediction, which has been widely studied. For instance, multi-class classification can be considered as a special version of multi-label classification, which has been studied in Belanger and McCallum (2016) and a set of follow-up works. The main difference here is that multi-class classification is a simpler problem, and all possible classes can be enumerated in O(N), but in multi-label classification, more complicated inference such as gradient-descent based approaches must be used.\nThe contrastive training can be seen as a special case of margin-based training (Belanger and McCallum, 2016; Rooshenas et al. 2019), where the margin is infinity.\nI believe the works in using EBMs for structured prediction must be cited here as they are closely related.\n\nThe authors also explored the effect of ML training on interference with past data and showed that using single sample ML approximation can significantly alleviate the catastrophic forgetting problem.  \nI believe that this is an interesting observation. \n\nTypo: \"current bath\" in Section 5.1.4\n\nBelanger and McCallum (2016), Structured Prediction Energy Networks.\nGygli et al. (2017), Deep Value Networks Learn to Evaluate and Iteratively Refine Structured Outputs.\nRooshenas et al. (2019), Search-Guided, Lightly-supervised Training of Structured Prediction Energy Networks", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Maximum likelihood training results in catastrophic forgetting", "review": "This paper explores the usage of EBMs in continual learning for classification. Although the application of EBMs in continual learning is novel, the general idea is a special case of the usage of EBMs for structured prediction, which has been widely studied. For instance, multi-class classification can be considered as a special version of multi-label classification, which has been studied in Belanger and McCallum (2016) and a set of follow-up works. The main difference here is that multi-class classification is a simpler problem, and all possible classes can be enumerated in O(N), but in multi-label classification, more complicated inference such as gradient-descent based approaches must be used.\nThe contrastive training can be seen as a special case of margin-based training (Belanger and McCallum, 2016; Rooshenas et al. 2019), where the margin is infinity.\nI believe the works in using EBMs for structured prediction must be cited here as they are closely related.\n\nThe authors also explored the effect of ML training on interference with past data and showed that using single sample ML approximation can significantly alleviate the catastrophic forgetting problem.  \nI believe that this is an interesting observation. \n\nTypo: \"current bath\" in Section 5.1.4\n\nBelanger and McCallum (2016), Structured Prediction Energy Networks.\nGygli et al. (2017), Deep Value Networks Learn to Evaluate and Iteratively Refine Structured Outputs.\nRooshenas et al. (2019), Search-Guided, Lightly-supervised Training of Structured Prediction Energy Networks", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604257666728}, {"id": "3dStbOjNRc2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1451/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "==========================\n\nbefore revision\n\n==========================\n\nReview: Motivated by the effectiveness and naturalness of energy-based models, this paper proposes to use energy-based learning framework for continual learning. Empirical studies are performed to validate the proposed strategy on several continual learning benchmarks.\n\nStrength: \n+ This paper applies EBMs to the task of continual learning, which is interesting and relevant to ICLR conference. \n+ The paper is well written and easy to follow. \n+ The paper is technically sound, since the formulation of the EBMs are well derived by other prior works.  \n\n\nConcerns: \n+ The contribution of the paper is insufficient for publication. The energy-based learning framework for discriminative purpose has been developed for a long time, even though recently researchers in the field machine learning are enthusiastic about developing energy-based models for data generation. \n+ The underlying theory of the proposed method is developed by other papers, the only contribution of this paper is to apply the EBM to the continual learning, which is quite straightforward. \n+ Missing key reference about EBM for discriminative learning. The core of this paper is mainly based on the finding of the transition between discriminative EBM and generative EBM, which is originally presented in reference [a]. The current paper misses to discuss and cite this paper. \n+ missing relevant reference about generative EBMs in related work. Even though this paper is not directly related to EBMs for data generation, but it DID discuss the development of it in its paper. The current related work about EBMs for generative purpose is incomplete in the sense that it skipped some pioneering works and important application with EBMs. For examples,  [1] is the first paper to use ConvNet-parameterized EBMs with Langevin for image generation.  Training EBMs with assisting networks can be found in [2] and [3].  Also, writing a section of comprehensive related works about energy-based learning is not necessary but encouraging.  \n\nreferences\n+ [1] A Theory of Generative ConvNet (ICML 2016)\n+ [2] Cooperative learning of descriptor and generator networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI 2018).\n+ [3] Divergence triangle for joint training of generator model, energy-based model, and inference model. (CVPR 2019)   \n\n===================================\n\nAfter a revision\n\n===================================\nThank you for your efforts to revise the paper. The revised parts about related work look good to me. I agree on that citing all those EBM application papers is not necessary. But doing so can provide a comprehensive and complete development of  the DeepNet-EBM. Again, this is not required and it will not affect the rating.   \n\nI also acknowledge the existing contributions in the current paper and admit that such a direction is promising, but I still feel that the current paper doesn't fully explore this area with more solid experiments. Thus, the whole contribution is quite marginal. By taking into account all these concerns, I will change my rating from 4 to 5.    \n  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Insufficient novelty + relevant reference missing", "review": "==========================\n\nbefore revision\n\n==========================\n\nReview: Motivated by the effectiveness and naturalness of energy-based models, this paper proposes to use energy-based learning framework for continual learning. Empirical studies are performed to validate the proposed strategy on several continual learning benchmarks.\n\nStrength: \n+ This paper applies EBMs to the task of continual learning, which is interesting and relevant to ICLR conference. \n+ The paper is well written and easy to follow. \n+ The paper is technically sound, since the formulation of the EBMs are well derived by other prior works.  \n\n\nConcerns: \n+ The contribution of the paper is insufficient for publication. The energy-based learning framework for discriminative purpose has been developed for a long time, even though recently researchers in the field machine learning are enthusiastic about developing energy-based models for data generation. \n+ The underlying theory of the proposed method is developed by other papers, the only contribution of this paper is to apply the EBM to the continual learning, which is quite straightforward. \n+ Missing key reference about EBM for discriminative learning. The core of this paper is mainly based on the finding of the transition between discriminative EBM and generative EBM, which is originally presented in reference [a]. The current paper misses to discuss and cite this paper. \n+ missing relevant reference about generative EBMs in related work. Even though this paper is not directly related to EBMs for data generation, but it DID discuss the development of it in its paper. The current related work about EBMs for generative purpose is incomplete in the sense that it skipped some pioneering works and important application with EBMs. For examples,  [1] is the first paper to use ConvNet-parameterized EBMs with Langevin for image generation.  Training EBMs with assisting networks can be found in [2] and [3].  Also, writing a section of comprehensive related works about energy-based learning is not necessary but encouraging.  \n\nreferences\n+ [1] A Theory of Generative ConvNet (ICML 2016)\n+ [2] Cooperative learning of descriptor and generator networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI 2018).\n+ [3] Divergence triangle for joint training of generator model, energy-based model, and inference model. (CVPR 2019)   \n\n===================================\n\nAfter a revision\n\n===================================\nThank you for your efforts to revise the paper. The revised parts about related work look good to me. I agree on that citing all those EBM application papers is not necessary. But doing so can provide a comprehensive and complete development of  the DeepNet-EBM. Again, this is not required and it will not affect the rating.   \n\nI also acknowledge the existing contributions in the current paper and admit that such a direction is promising, but I still feel that the current paper doesn't fully explore this area with more solid experiments. Thus, the whole contribution is quite marginal. By taking into account all these concerns, I will change my rating from 4 to 5.    \n  \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603956640447}, {"id": "qZ74kfSktaL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1451/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This work shows that energy-based models (EBMs)  are a promising model class for continual learning problems. According to the experiments, EBMs outperform the baseline methods by several continual learning benchmarks.\n \n+ves: \n1. It is interesting to see that energy-based models are introduced for the classification continual learning problems. In the paper, the authors show that EBMs achieve a significant improvement on four standard CL benchmarks. It is really surprising that the improvements are so large. \n \n2. Some analysis, for example,  “energy landscape” and “Interference with past data.” seems interesting. It shows benefit of energy-based model for continual learning. Both of them indicate that y EBMs suffer less from catastrophic forgetting.\n \n \nConcerns\n1. It seems like the gradient computation of equation (5) is not corresponding to equation (4). \n\n2. It is nice to see the improvement of EBMS. In this work, a difference architecture is usually in EBMS. The architecture is different from the baseline. It is nice that EMBS has a more flexible architecture to score the input x and output y. However, the improvement of your work is from architecture or from the training. It is better to do a clear claim. If the benefit is from the architecture, do you treat it as a black box? Or it is from the training objective?  It is possible to show some learned structure in your formulation. The architecture of EBM seems large. \n\n\nQuestions during the rebuttal period: \nPlease address and clarify the cons above:\n \n1. The experiment setting detail is unclear. The training details of the proposed method and baseline are unclear.  And the baseline detail is a little confusing to me. In section 5.1.2, it says “ All the baselines and EBMs are based on the same model architecture”. It seems the architecture of EBMs is different.\n\n2. Some related work on energy-based model: multiple label classification[1] , sequence labeling [2] and machine translation [3]\n[1] Structured Prediction Energy Networks, ICML 2016\n[2] Benchmarking Approximate Inference Methods for Neural Structured Prediction. NAACL 2019\n[3] ENGINE: Energy-Based Inference Networks for Non-Autoregressive Machine Translation\nACL 2020\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review 3", "review": "Summary: This work shows that energy-based models (EBMs)  are a promising model class for continual learning problems. According to the experiments, EBMs outperform the baseline methods by several continual learning benchmarks.\n \n+ves: \n1. It is interesting to see that energy-based models are introduced for the classification continual learning problems. In the paper, the authors show that EBMs achieve a significant improvement on four standard CL benchmarks. It is really surprising that the improvements are so large. \n \n2. Some analysis, for example,  “energy landscape” and “Interference with past data.” seems interesting. It shows benefit of energy-based model for continual learning. Both of them indicate that y EBMs suffer less from catastrophic forgetting.\n \n \nConcerns\n1. It seems like the gradient computation of equation (5) is not corresponding to equation (4). \n\n2. It is nice to see the improvement of EBMS. In this work, a difference architecture is usually in EBMS. The architecture is different from the baseline. It is nice that EMBS has a more flexible architecture to score the input x and output y. However, the improvement of your work is from architecture or from the training. It is better to do a clear claim. If the benefit is from the architecture, do you treat it as a black box? Or it is from the training objective?  It is possible to show some learned structure in your formulation. The architecture of EBM seems large. \n\n\nQuestions during the rebuttal period: \nPlease address and clarify the cons above:\n \n1. The experiment setting detail is unclear. The training details of the proposed method and baseline are unclear.  And the baseline detail is a little confusing to me. In section 5.1.2, it says “ All the baselines and EBMs are based on the same model architecture”. It seems the architecture of EBMs is different.\n\n2. Some related work on energy-based model: multiple label classification[1] , sequence labeling [2] and machine translation [3]\n[1] Structured Prediction Energy Networks, ICML 2016\n[2] Benchmarking Approximate Inference Methods for Neural Structured Prediction. NAACL 2019\n[3] ENGINE: Energy-Based Inference Networks for Non-Autoregressive Machine Translation\nACL 2020\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603830342478}, {"id": "QzeltLz-bzH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1451/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The writing of the paper needs improvements. I am still unsure how the paper learns each new task. You talked about batches but never talked about how each new task is learned specifically. It creates doubts in my mind. E.g., does each batch contain some examples from old tasks? Is this training like for multi-task learning?\n\nYou wrote “In the continual learning setting, we assume classes in Y_B are uniformly distributed in every new batch.” Is Y_B fixed for each task or each batch? \n\nYou need a negative class in each batch. This means that your algorithm cannot work in the scenario where a task has one class only. Most of existing techniques can handle this case although they use 2 or more classes in a task in their experiments. I think this is a serious limitation. It is only suitable for Task-IL, which is an easier problem to solve. \n\nI think the claim that existing techniques need to fix the number of classes beforehand is not correct. I know hat most of them fix the number in their code or experiments, but I don’t see why they cannot use a large number or dynamically add new class heads when needed, say, using cross-entropy as the loss function. \n\nThe experiments are not well described, and baselines are old. Many newer baselines should be included, e.g., \n\nLearning a Unified Classifier Incrementally via Rebalancing. CVPR 2019. \nOvercoming catastrophic forgetting for continual learning via model adaptation. ICLR, 2019. \nRandom path selection for continual learning. NeurIPS 2019\nContinuous learning of context-dependent processing in neural networks. Nature Machine Intelligence, 2019\nLarge scale incremental learning. CVPR 2019\n\nFor each dataset, you used one setting for tasks only, e.g., CIFAR100, 10 tasks. More than one setting should be tried to show the generality of the approach. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposes an energy-based model for continual learning, which seems to be new. However, I have several concerns about the paper, which are described in the detailed comments. ", "review": "The writing of the paper needs improvements. I am still unsure how the paper learns each new task. You talked about batches but never talked about how each new task is learned specifically. It creates doubts in my mind. E.g., does each batch contain some examples from old tasks? Is this training like for multi-task learning?\n\nYou wrote “In the continual learning setting, we assume classes in Y_B are uniformly distributed in every new batch.” Is Y_B fixed for each task or each batch? \n\nYou need a negative class in each batch. This means that your algorithm cannot work in the scenario where a task has one class only. Most of existing techniques can handle this case although they use 2 or more classes in a task in their experiments. I think this is a serious limitation. It is only suitable for Task-IL, which is an easier problem to solve. \n\nI think the claim that existing techniques need to fix the number of classes beforehand is not correct. I know hat most of them fix the number in their code or experiments, but I don’t see why they cannot use a large number or dynamically add new class heads when needed, say, using cross-entropy as the loss function. \n\nThe experiments are not well described, and baselines are old. Many newer baselines should be included, e.g., \n\nLearning a Unified Classifier Incrementally via Rebalancing. CVPR 2019. \nOvercoming catastrophic forgetting for continual learning via model adaptation. ICLR, 2019. \nRandom path selection for continual learning. NeurIPS 2019\nContinuous learning of context-dependent processing in neural networks. Nature Machine Intelligence, 2019\nLarge scale incremental learning. CVPR 2019\n\nFor each dataset, you used one setting for tasks only, e.g., CIFAR100, 10 tasks. More than one setting should be tried to show the generality of the approach. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603729260902}], "openreview_url": "https://openreview.net/forum?id=j5d9qacxdZa", "arxiv_id": "2011.12216", "paper_pdf": "papers/j5d9qacxdZa.pdf", "paper_pdf_sha256": "d82cdf180c473dde5a7fcb2607cc9f5ea4dde8d4e4c3ab761c541718ae4fdea6", "paper_pdf_bytes": 4948780, "paper_pdf_source": "openreview", "code_url": "https://github.com/ShuangLI59/ebm-continual-learning", "code_repository": "ShuangLI59/ebm-continual-learning", "code_commit": "0450d69ac01625c1d227356e5367e002aaae65a4", "code_archive": "repos/j5d9qacxdZa.zip", "code_archive_sha256": "f12f2ed64e2c5a7d6e808a266336af2c0dcff2959bb426be6e70a50b426e013e", "code_archive_bytes": 1536377, "code_file_count": 41, "code_extensions": {".py": 25, ".sh": 16}, "github_disk_usage_kb": 1484, "github_languages": {"Python": 232622, "Shell": 3426}, "github_archived": false, "github_pushed_at": "2022-09-13T02:08:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/energy-based-models-for-continual-learning-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FnwU7ogRzv", "year": 2026, "status": "rejected", "title": "CRAKEN: Cybersecurity LLM Agent with Knowledge-Based Execution", "authors": ["Minghao Shao", "Haoran Xi", "NANDA RANI", "Meet Udeshi", "Venkata Sai Charan Putrevu", "Kimberly Milner", "Brendan Dolan-Gavitt", "Sandeep K. Shukla", "Prashanth Krishnamurthy", "Farshad Khorrami", "Ramesh Karri", "Muhammad Shafique"], "authorids": ["~Minghao_Shao3", "~Haoran_Xi1", "~NANDA_RANI1", "~Meet_Udeshi1", "~Venkata_Sai_Charan_Putrevu1", "~Kimberly_Milner1", "~Brendan_Dolan-Gavitt1", "~Sandeep_K._Shukla1", "~Prashanth_Krishnamurthy1", "~Farshad_Khorrami1", "~Ramesh_Karri1", "~Muhammad_Shafique1"], "authors_source": "OpenReview API", "abstract": "Large Language Model (LLM) agents can automate cybersecurity tasks and can adapt to the evolving cybersecurity landscape without re-engineering. While LLM agents have demonstrated cybersecurity capabilities on Capture-The-Flag (CTF) competitions, they have two key limitations: accessing latest cybersecurity expertise beyond training data, and integrating new knowledge into complex task planning. Knowledge-based approaches that incorporate technical understanding into the task-solving automation can tackle these limitations. We present CRAKEN, a knowledge-based LLM agent framework that improves cybersecurity capability through three core mechanisms: contextual decomposition of task-critical information, iterative self-reflected knowledge retrieval, and solution injection that transforms insights into adaptive attack strategies. CRAKEN combines advanced retrieval algorithms with prompt injection-based integration. Comprehensive evaluations with different configurations show CRAKEN's effectiveness in multi-stage vulnerability detection and exploitation compared to previous approaches. Our extensible architecture establishes new methodologies for embedding new security knowledge into LLM-driven cybersecurity agentic systems. CRAKEN obtained an accuracy of 22% on NYU CTF Bench with our collected simple CTF write-up dataset and shows a 15.7% of the solution distribution difference, indicating the effectiveness of integrating knowledge into automated CTF solving for challenges that requires knowledge from specific domains.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "toxfqJAP0S", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13539/Reviewer_tBDn"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces a system for solving cybersecurity Capture-The-Flag (CTFs) that uses language model agents + retrieval of domain knowledge (CTF writeups). On the NYU CTF benchmark, the proposed system improves over the baseline.", "review_text": "This paper introduces a system for solving cybersecurity Capture-The-Flag (CTFs) that uses language model agents + retrieval of domain knowledge (CTF writeups). On the NYU CTF benchmark, the proposed system improves over the baseline.", "strengths": "Cybersecurity LM agents is an exciting area, and using prior domain knowledge is a sensible approach.\nThe new system (CRAKEN) does improve over prior work.", "weaknesses": "The system is rather complex and it is hard to tell which components are most helpful (Table 2 might have athe information but it combines a bunch of variations like model, which is orthogonal to the contributions of the paper). It would be clearer to make clear the three axes of variation: models, scaffolds (RAG or not), and information available to the agent.\nOnly the NYU CTF dataset is used.  What about Cybench, XBOW, Intercode, CTF-Dojo? The paper would be empirically stronger if it showed the applicability of the method across multiple datasets.\nI know this space moves fast, but it would be interesting to see how the latest models (GPT-5, Claude 4.5, etc.) work.", "questions": "How do you ensure there is no train-test contamination, especially since you're adding new sources of information (the CTF writeups)?\nHow were the prompts tuned for each of the different models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a system for solving cybersecurity Capture-The-Flag (CTFs) that uses language model agents + retrieval of domain knowledge (CTF writeups). On the NYU CTF benchmark, the proposed system improves over the baseline.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "Cybersecurity LM agents is an exciting area, and using prior domain knowledge is a sensible approach.\nThe new system (CRAKEN) does improve over prior work.", "weaknesses": "The system is rather complex and it is hard to tell which components are most helpful (Table 2 might have athe information but it combines a bunch of variations like model, which is orthogonal to the contributions of the paper). It would be clearer to make clear the three axes of variation: models, scaffolds (RAG or not), and information available to the agent.\nOnly the NYU CTF dataset is used.  What about Cybench, XBOW, Intercode, CTF-Dojo? The paper would be empirically stronger if it showed the applicability of the method across multiple datasets.\nI know this space moves fast, but it would be interesting to see how the latest models (GPT-5, Claude 4.5, etc.) work.", "questions": "How do you ensure there is no train-test contamination, especially since you're adding new sources of information (the CTF writeups)?\nHow were the prompts tuned for each of the different models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762129436319}, {"id": "SQLC2WcibJ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13539/Reviewer_49us"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper presents CRAKEN, an LLM agent built to complete the NYU CTF Benchmark using retrieval from different knowledge databases. CRAKEN is able to achieve a 3% improvement on NYU CTF Bench.", "review_text": "This paper presents CRAKEN, an LLM agent built to complete the NYU CTF Benchmark using retrieval from different knowledge databases. CRAKEN is able to achieve a 3% improvement on NYU CTF Bench.", "strengths": "The work does a good job presenting the complex CRAKEN system, and presents a thorough understanding of previous agents built for CTFs and CTF benchmarks themselves. The authors are thorough in their evaluations. They present results for different configurations of CRAKEN, and analyze failure modes and the performance of the graph RAG system.", "weaknesses": "Ultimately, this work adds RAG capabilities to a previously presented agentic framework from Xu we al. in order to achieve a minimal increase in performance on the NYU CTF benchmark. The planner-executor based framework is not novel, nor is adding RAG to agents to improve performance. Additionally, despite compiling and presenting a complex system for fetching information relevant to the CTF task at hand, the system only performs 3% more in total on a singular benchmark. This is not nearly a significant enough improvement nor wide enough breadth of evaluations to claim that the agentic framework is state-of-the-art or a worthwhile contribution to ICLR. In fact, under some configurations of the agent performance drops below their baseline comparison of D-CIPHER. This is not to say that negative results should be punished, rather that there is not sufficient breadth of evidence here to show that this system consistently improves performance across the board. Finally, simply evaluating on one CTF benchmark is not enough. CTFs alone are not a reasonable metric for evaluating the cybersecurity capabilities of an agent, and therefore cannot be the sole evaluation metric for a new agent framework. This would be strengthened significantly had the authors presented results for a) other CTF benchmarks, such as Cybench of XBOW's validation set, and b) other, more realistic cybersecurity evaluations, like MHBench, BountyBench, or CVEBench. Currently, the lack of generalizability means that this work is not as strong as it could be.", "questions": "1. In table 2, some of the CRAKEN configurations lead to worse performance than the D-CIPHER baseline. Do you have any intuitions as to why an increase in knowledge access results in worse performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents CRAKEN, an LLM agent built to complete the NYU CTF Benchmark using retrieval from different knowledge databases. CRAKEN is able to achieve a 3% improvement on NYU CTF Bench.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "The work does a good job presenting the complex CRAKEN system, and presents a thorough understanding of previous agents built for CTFs and CTF benchmarks themselves. The authors are thorough in their evaluations. They present results for different configurations of CRAKEN, and analyze failure modes and the performance of the graph RAG system.", "weaknesses": "Ultimately, this work adds RAG capabilities to a previously presented agentic framework from Xu we al. in order to achieve a minimal increase in performance on the NYU CTF benchmark. The planner-executor based framework is not novel, nor is adding RAG to agents to improve performance. Additionally, despite compiling and presenting a complex system for fetching information relevant to the CTF task at hand, the system only performs 3% more in total on a singular benchmark. This is not nearly a significant enough improvement nor wide enough breadth of evaluations to claim that the agentic framework is state-of-the-art or a worthwhile contribution to ICLR. In fact, under some configurations of the agent performance drops below their baseline comparison of D-CIPHER. This is not to say that negative results should be punished, rather that there is not sufficient breadth of evidence here to show that this system consistently improves performance across the board. Finally, simply evaluating on one CTF benchmark is not enough. CTFs alone are not a reasonable metric for evaluating the cybersecurity capabilities of an agent, and therefore cannot be the sole evaluation metric for a new agent framework. This would be strengthened significantly had the authors presented results for a) other CTF benchmarks, such as Cybench of XBOW's validation set, and b) other, more realistic cybersecurity evaluations, like MHBench, BountyBench, or CVEBench. Currently, the lack of generalizability means that this work is not as strong as it could be.", "questions": "1. In table 2, some of the CRAKEN configurations lead to worse performance than the D-CIPHER baseline. Do you have any intuitions as to why an increase in knowledge access results in worse performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761958538618}, {"id": "nvKZOu5OHn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13539/Reviewer_QLh2"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a knowledge-based LLM agent framework, CRAKEN, which integrates RAG techniques with LLM agents for cybersecurity tasks. Given a Capture-The-Flag (CTF) challenge, the planner-executor based system first decomposes the challenge into multiple sub-tasks and assigns each to an executor. During sub-task execution, each executor is enhanced with knowledge hints retrieved through a Self-RAG framework that employs a hybrid strategy combining traditional vector-similarity search and Graph-RAG methods. Leveraging the retrieved knowledge, the executors solve the challenge. CRAKEN is evaluated on the NYU CTF Bench dataset, which contains 200 CTF challenges from 2017 to 2023, achieving a 3% performance improvement over the state-of-the-art method, D-CIPHER. Additionally, the experiment on different knowledge databases used in the retrieval process demonstrate that step-by-step operational knowledge extracted from CTF write-ups provides superior guidance compared to general or mixed datasets.", "review_text": "This paper proposes a knowledge-based LLM agent framework, CRAKEN, which integrates RAG techniques with LLM agents for cybersecurity tasks. Given a Capture-The-Flag (CTF) challenge, the planner-executor based system first decomposes the challenge into multiple sub-tasks and assigns each to an executor. During sub-task execution, each executor is enhanced with knowledge hints retrieved through a Self-RAG framework that employs a hybrid strategy combining traditional vector-similarity search and Graph-RAG methods. Leveraging the retrieved knowledge, the executors solve the challenge. CRAKEN is evaluated on the NYU CTF Bench dataset, which contains 200 CTF challenges from 2017 to 2023, achieving a 3% performance improvement over the state-of-the-art method, D-CIPHER. Additionally, the experiment on different knowledge databases used in the retrieval process demonstrate that step-by-step operational knowledge extracted from CTF write-ups provides superior guidance compared to general or mixed datasets.", "strengths": "1. The overall quality of the writing is sufficient.\n2. Comprehensive evaluation with different LLMs and knowledge databases. The paper provides a comprehensive evaluation of CRAKEN on four powerful close-source LLMs, including Claude 3.5 Sonnet, Claude 3.7 Sonnet, GPT 4o and GPT 4.1, and a open-source LLM, namely DeepSeek V3. Moreover, the finding derived from the experiments, showing that step-by-step operational CTF write-ups are the most effective knowledge source for the RAG process, is valuable.\n3. Clear breakdown of time overhead and cost. In the evaluation result, the paper clearly clarify the average cost per solved CTF and the execution latency of either all attempts and successful cases. The low cost of the framework shows its practicality.", "weaknesses": "## The novelty is poor.\nThe evaluation of integrating RAG with cybersecurity LLM agents on CTF tasks is valuable. However, the work is built upon the existing D-CIPHER framework [1] and directly adopts previously proposed RAG components, namely Self-RAG [2] and Graph-RAG [3]. Although the paper claims to employ an optimized version of Self-RAG, the differences between it and the original implementation described in [2] are not clearly explained. Overall, since the proposed methodology appears to be mainly a combination of existing approaches, the work represents an incremental study with poor novelty.\n\n## Unclear technical details\n- The decomposition strategies for task-critical information and the rationale behind their design are not clearly described. Although the contextual decomposition process appears to be a crucial step, its implementation is not thoroughly explained in Section 3 (CRAKEN Architecture). Furthermore, no assessment of the decomposition process’s soundness is provided.\n- The criteria for baseline selection are unclear. As shown in Table 1, there are at least seven existing works on LLM agents for cybersecurity. However, the evaluation only considers D-CIPHER and EnIGMA as baselines, without explaining why these two were chosen over the others. This lack of justification raises concerns about potential bias in the evaluation, as the comparison is limited to only two baselines.\n- The retry mechanism discussed in Section 5 (RESULTS – Retrieval Process Analysis) is neither detailed in Section 3 (CRAKEN ARCHITECTURE – Retrieval Process) nor illustrated in Algorithm 1. Additionally, as shown in Figure 4, it is unclear why 33.7% of cases in the Retry process reach Success directly, bypassing the Grade Answer step. Further clarification on the design and operation of this retry mechanism is required.\n\n## Insufficient experiments on post-cutoff benchmark\nThe experiment on the flag-leakage-free cut-off benchmark does not provide strong evidence of CRAKEN's performance persistence on unfamiliar tasks. In the results, there is a notable gap between the 13.64% solve rate on this benchmark and the 21% solve rate on the NYU CTF Bench. While this discrepancy may be partly explained by the difference in dataset sizes (approximately 1:10), the results of D-CIPHER on these 22 post-cutoff cases should also be included. Having this comparison would help demonstrate whether the performance gap between D-CIPHER and CRAKEN remains consistent, thereby supporting the claim that CRAKEN’s superior performance persists even on unfamiliar tasks.\n\n## Difficulty in evaluating the reproducibility\nThe source code and datasets are not publicly available at the time of review; only the prompts are open-source, which makes it difficult to assess reproducibility.\n\n[1] Meet Udeshi and Minghao Shao and Haoran Xi and Nanda Rani and Kimberly Milner and Venkata Sai Charan Putrevu and Brendan Dolan-Gavitt and Sandeep Kumar Shukla and Prashanth Krishnamurthy and Farshad Khorrami and Ramesh Karri and Muhammad Shafique. D-CIPHER: Dynamic Collaborative Intelligent Multi-Agent System with Planner and Heterogeneous Executors for Offensive Security. https://arxiv.org/abs/2502.10931 \\\n[2] Akari Asai and Zeqiu Wu and Yizhong Wang and Avirup Sil and Hannaneh Hajishirzi. Self-{RAG}: Learning to Retrieve, Generate, and Critique through Self-Reflection. In International Conference on Learning Representations, 2023. \\\n[3] Yuntong Hu, Zhihan Lei, Zheng Zhang, Bo Pan, Chen Ling, and Liang Zhao. Grag: Graph retrieval-augmented generation. arXiv preprint arXiv:2405.16506, 2024.", "questions": "For detailed major concerns, please see the Weaknesses\n\n1. Please justify the novelty of CRAKEN, and clarify the differences between the optimized Self-RAG used and the original implementation\n2. Please explain the design of the contextual decomposition process and include an evaluation on its soundness in the experiment.\n3. Please justify the reason of only choosing D-CIPHER and EnIGMA as baselines over other existing works listed in Table 1.\n4. Please address the confusion in Figure 4 or clarify the Retry process.\n5. Please provide the evaluation result of D-CIPHER on the post-cutoff dataset.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a knowledge-based LLM agent framework, CRAKEN, which integrates RAG techniques with LLM agents for cybersecurity tasks. Given a Capture-The-Flag (CTF) challenge, the planner-executor based system first decomposes the challenge into multiple sub-tasks and assigns each to an executor. During sub-task execution, each executor is enhanced with knowledge hints retrieved through a Self-RAG framework that employs a hybrid strategy combining traditional vector-similarity search and Graph-RAG methods. Leveraging the retrieved knowledge, the executors solve the challenge. CRAKEN is evaluated on the NYU CTF Bench dataset, which contains 200 CTF challenges from 2017 to 2023, achieving a 3% performance improvement over the state-of-the-art method, D-CIPHER. Additionally, the experiment on different knowledge databases used in the retrieval process demonstrate that step-by-step operational knowledge extracted from CTF write-ups provides superior guidance compared to general or mixed datasets.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The overall quality of the writing is sufficient.\n2. Comprehensive evaluation with different LLMs and knowledge databases. The paper provides a comprehensive evaluation of CRAKEN on four powerful close-source LLMs, including Claude 3.5 Sonnet, Claude 3.7 Sonnet, GPT 4o and GPT 4.1, and a open-source LLM, namely DeepSeek V3. Moreover, the finding derived from the experiments, showing that step-by-step operational CTF write-ups are the most effective knowledge source for the RAG process, is valuable.\n3. Clear breakdown of time overhead and cost. In the evaluation result, the paper clearly clarify the average cost per solved CTF and the execution latency of either all attempts and successful cases. The low cost of the framework shows its practicality.", "weaknesses": "## The novelty is poor.\nThe evaluation of integrating RAG with cybersecurity LLM agents on CTF tasks is valuable. However, the work is built upon the existing D-CIPHER framework [1] and directly adopts previously proposed RAG components, namely Self-RAG [2] and Graph-RAG [3]. Although the paper claims to employ an optimized version of Self-RAG, the differences between it and the original implementation described in [2] are not clearly explained. Overall, since the proposed methodology appears to be mainly a combination of existing approaches, the work represents an incremental study with poor novelty.\n\n## Unclear technical details\n- The decomposition strategies for task-critical information and the rationale behind their design are not clearly described. Although the contextual decomposition process appears to be a crucial step, its implementation is not thoroughly explained in Section 3 (CRAKEN Architecture). Furthermore, no assessment of the decomposition process’s soundness is provided.\n- The criteria for baseline selection are unclear. As shown in Table 1, there are at least seven existing works on LLM agents for cybersecurity. However, the evaluation only considers D-CIPHER and EnIGMA as baselines, without explaining why these two were chosen over the others. This lack of justification raises concerns about potential bias in the evaluation, as the comparison is limited to only two baselines.\n- The retry mechanism discussed in Section 5 (RESULTS – Retrieval Process Analysis) is neither detailed in Section 3 (CRAKEN ARCHITECTURE – Retrieval Process) nor illustrated in Algorithm 1. Additionally, as shown in Figure 4, it is unclear why 33.7% of cases in the Retry process reach Success directly, bypassing the Grade Answer step. Further clarification on the design and operation of this retry mechanism is required.\n\n## Insufficient experiments on post-cutoff benchmark\nThe experiment on the flag-leakage-free cut-off benchmark does not provide strong evidence of CRAKEN's performance persistence on unfamiliar tasks. In the results, there is a notable gap between the 13.64% solve rate on this benchmark and the 21% solve rate on the NYU CTF Bench. While this discrepancy may be partly explained by the difference in dataset sizes (approximately 1:10), the results of D-CIPHER on these 22 post-cutoff cases should also be included. Having this comparison would help demonstrate whether the performance gap between D-CIPHER and CRAKEN remains consistent, thereby supporting the claim that CRAKEN’s superior performance persists even on unfamiliar tasks.\n\n## Difficulty in evaluating the reproducibility\nThe source code and datasets are not publicly available at the time of review; only the prompts are open-source, which makes it difficult to assess reproducibility.\n\n[1] Meet Udeshi and Minghao Shao and Haoran Xi and Nanda Rani and Kimberly Milner and Venkata Sai Charan Putrevu and Brendan Dolan-Gavitt and Sandeep Kumar Shukla and Prashanth Krishnamurthy and Farshad Khorrami and Ramesh Karri and Muhammad Shafique. D-CIPHER: Dynamic Collaborative Intelligent Multi-Agent System with Planner and Heterogeneous Executors for Offensive Security. https://arxiv.org/abs/2502.10931 \\\n[2] Akari Asai and Zeqiu Wu and Yizhong Wang and Avirup Sil and Hannaneh Hajishirzi. Self-{RAG}: Learning to Retrieve, Generate, and Critique through Self-Reflection. In International Conference on Learning Representations, 2023. \\\n[3] Yuntong Hu, Zhihan Lei, Zheng Zhang, Bo Pan, Chen Ling, and Liang Zhao. Grag: Graph retrieval-augmented generation. arXiv preprint arXiv:2405.16506, 2024.", "questions": "For detailed major concerns, please see the Weaknesses\n\n1. Please justify the novelty of CRAKEN, and clarify the differences between the optimized Self-RAG used and the original implementation\n2. Please explain the design of the contextual decomposition process and include an evaluation on its soundness in the experiment.\n3. Please justify the reason of only choosing D-CIPHER and EnIGMA as baselines over other existing works listed in Table 1.\n4. Please address the confusion in Figure 4 or clarify the Retry process.\n5. Please provide the evaluation result of D-CIPHER on the post-cutoff dataset.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761922564091}, {"id": "KUn05QsWIu", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13539/Reviewer_US46"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper presents CRAKEN, a framework for improving LLM performance on cybersecurity tasks, especially CTF challenges. The authors point out two key problems with current LLM agents: they can't access security knowledge beyond their training data, and they struggle to integrate new knowledge into multi-stage planning and execution. The paper tackles this through knowledge-based mechanisms that combine structured and unstructured information.\nMain contributions are: a knowledge-enhanced agent architecture with task decomposition, retrieval and injection capabilities; a self-evaluating Self-RAG retrieval process for better output accuracy; Graph-RAG to model entities and relationships using knowledge graphs; a knowledge base with CTF writeups, attack payloads, and code snippets; and empirical evaluation on CTF benchmarks.", "review_text": "This paper presents CRAKEN, a framework for improving LLM performance on cybersecurity tasks, especially CTF challenges. The authors point out two key problems with current LLM agents: they can't access security knowledge beyond their training data, and they struggle to integrate new knowledge into multi-stage planning and execution. The paper tackles this through knowledge-based mechanisms that combine structured and unstructured information.\nMain contributions are: a knowledge-enhanced agent architecture with task decomposition, retrieval and injection capabilities; a self-evaluating Self-RAG retrieval process for better output accuracy; Graph-RAG to model entities and relationships using knowledge graphs; a knowledge base with CTF writeups, attack payloads, and code snippets; and empirical evaluation on CTF benchmarks.", "strengths": "[+] Comprehensive result analysis including failure type statistics (Figure 5) and retrieval process distribution (Figure 4)\n\n[+] Clear architecture (Figure 1) and module definitions", "weaknesses": "[-] Incremental research. Core methods borrowed from existing work (Self-RAG, Graph-RAG)\n\n[-] Limited performance improvements\n\n[-] Inconsistent experimental configurations with insufficient control of variables\n\n[-] Lack of module implementation details", "questions": "Q1: Grading module implementation\nSection 3 describes three grading components (Relevance Grader, Hallucination Grader, and Solved Grader) in the retrieval pipeline, but I couldn't find implementation details. Are these prompt-based LLM calls with binary classification instructions, fine-tuned classifiers on labeled cybersecurity data, or rule-based heuristics?\n\nQ2: Graph-RAG\nHow are the queries for Graph-RAG derived? It is not clear how the initial queries are setup, and how do they evolve if retrieval fails. Do they come from the responses from the command executions?\n\nQ3: Hybrid retrieval fusion\nFigure 2 shows the hybrid Graph-RAG approach combining structured graph search and vector retrieval, but the fusion mechanism is unclear. Are results from both retrievers simply concatenated or weighted by relevance scores? How do you handle conflicts when graph-based and vector-based results contradict each other?\n\nQ4: Knowledge graph construction\nSection 4 mentions the knowledge graph uses semantic triplets like \"appeared_in\" and \"mentions\" shown in Figure 2, but critical details are missing. What's the complete relation taxonomy? Is there a formal ontology guiding triplet extraction? How are entities (vulnerabilities, exploits, techniques) identified and canonicalized from unstructured CTF writeups? Is the graph construction fully automated via LLM extraction or does it involve manual curation?\n\nQ5: Inconsistent experimental setup\nTable 2 shows several inconsistencies. Why does CRAKEN w/ Graph-RAG only report Claude 3.5 Sonnet results while other configurations test multiple LLMs? This makes it hard to isolate Graph-RAG's contribution from model selection effects. Also, GPT-4.1 performs worse with CRAKEN (11.5%) than D-CIPHER (13.5%), which seems to contradict the framework's premise - what explains this regression? Finally, the paper evaluates DeepSeek V3 but doesn't analyze why this open-source model underperforms so dramatically (2-3% vs 18-22% for commercial models). Do you plan to include more open-source models in future work to verify the system's generalizability?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents CRAKEN, a framework for improving LLM performance on cybersecurity tasks, especially CTF challenges. The authors point out two key problems with current LLM agents: they can't access security knowledge beyond their training data, and they struggle to integrate new knowledge into multi-stage planning and execution. The paper tackles this through knowledge-based mechanisms that combine structured and unstructured information.\nMain contributions are: a knowledge-enhanced agent architecture with task decomposition, retrieval and injection capabilities; a self-evaluating Self-RAG retrieval process for better output accuracy; Graph-RAG to model entities and relationships using knowledge graphs; a knowledge base with CTF writeups, attack payloads, and code snippets; and empirical evaluation on CTF benchmarks.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "[+] Comprehensive result analysis including failure type statistics (Figure 5) and retrieval process distribution (Figure 4)\n\n[+] Clear architecture (Figure 1) and module definitions", "weaknesses": "[-] Incremental research. Core methods borrowed from existing work (Self-RAG, Graph-RAG)\n\n[-] Limited performance improvements\n\n[-] Inconsistent experimental configurations with insufficient control of variables\n\n[-] Lack of module implementation details", "questions": "Q1: Grading module implementation\nSection 3 describes three grading components (Relevance Grader, Hallucination Grader, and Solved Grader) in the retrieval pipeline, but I couldn't find implementation details. Are these prompt-based LLM calls with binary classification instructions, fine-tuned classifiers on labeled cybersecurity data, or rule-based heuristics?\n\nQ2: Graph-RAG\nHow are the queries for Graph-RAG derived? It is not clear how the initial queries are setup, and how do they evolve if retrieval fails. Do they come from the responses from the command executions?\n\nQ3: Hybrid retrieval fusion\nFigure 2 shows the hybrid Graph-RAG approach combining structured graph search and vector retrieval, but the fusion mechanism is unclear. Are results from both retrievers simply concatenated or weighted by relevance scores? How do you handle conflicts when graph-based and vector-based results contradict each other?\n\nQ4: Knowledge graph construction\nSection 4 mentions the knowledge graph uses semantic triplets like \"appeared_in\" and \"mentions\" shown in Figure 2, but critical details are missing. What's the complete relation taxonomy? Is there a formal ontology guiding triplet extraction? How are entities (vulnerabilities, exploits, techniques) identified and canonicalized from unstructured CTF writeups? Is the graph construction fully automated via LLM extraction or does it involve manual curation?\n\nQ5: Inconsistent experimental setup\nTable 2 shows several inconsistencies. Why does CRAKEN w/ Graph-RAG only report Claude 3.5 Sonnet results while other configurations test multiple LLMs? This makes it hard to isolate Graph-RAG's contribution from model selection effects. Also, GPT-4.1 performs worse with CRAKEN (11.5%) than D-CIPHER (13.5%), which seems to contradict the framework's premise - what explains this regression? Finally, the paper evaluates DeepSeek V3 but doesn't analyze why this open-source model underperforms so dramatically (2-3% vs 18-22% for commercial models). Do you plan to include more open-source models in future work to verify the system's generalizability?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed.", "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761523792483}], "openreview_url": "https://openreview.net/forum?id=FnwU7ogRzv", "arxiv_id": "2505.17107", "paper_pdf": "papers/FnwU7ogRzv.pdf", "paper_pdf_sha256": "b9d83bb69e045abc6e4e42de67e57f736f97e56e54958164c761427d1dbee26d", "paper_pdf_bytes": 874852, "paper_pdf_source": "openreview", "code_url": "https://github.com/NYU-LLM-CTF/nyuctf_agents_craken", "code_repository": "NYU-LLM-CTF/nyuctf_agents_craken", "code_commit": "748bc8986bae2e02eb69dbebce07006a8b325a73", "code_archive": "repos/FnwU7ogRzv.zip", "code_archive_sha256": "e08a3112d197e51e8294a3ae805dad3198236a528c517e95087c00a4f0495604", "code_archive_bytes": 54704, "code_file_count": 19, "code_extensions": {".py": 18, ".sh": 1}, "github_disk_usage_kb": 5291, "github_languages": {"Python": 170827, "Shell": 58}, "github_archived": false, "github_pushed_at": "2025-07-13T23:23:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/craken-cybersecurity-llm-agent-with-knowledge"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2bn7gayfz9", "year": 2025, "status": "rejected", "title": "CTBench: A Library and Benchmark for Certified Training", "authors": ["Yuhao Mao", "Stefan Balauca", "Martin Vechev"], "authorids": ["~Yuhao_Mao1", "~Stefan_Balauca1", "~Martin_Vechev1"], "authors_source": "OpenReview API", "abstract": "Training certifiably robust neural networks is an important but challenging task. While many algorithms for (deterministic) certified training have been proposed, they are often evaluated on different training schedules, certification methods, and systematically under-tuned hyperparameters, making it difficult to compare their performance. To address this challenge, we introduce CTBench, a unified library and a high-quality benchmark for certified training that evaluates all algorithms under fair settings and systematically tuned hyperparameters. We show that (1) almost all algorithms in CTBench surpass the corresponding reported performance in literature in the magnitude of algorithmic improvements, thus establishing new state-of-the-art, and (2) the claimed advantage of recent algorithms drops significantly when we enhance the outdated baselines with a fair training schedule, a fair certification method and well-tuned hyperparameters. Based on CTBench, we provide insights into the current state of certified training and suggest future research directions. We are confident that CTBench will serve as a benchmark and testbed for future research in certified training.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "PHqrCIdJeX", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6522/Reviewer_xBxH"], "rating": 5, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "The paper presents CTBENCH, a standardized library and benchmark designed to fairly evaluate certified training algorithms for neural networks, addressing the inconsistency in previous evaluations due to varied training schedules, certification methods, and under-optimized hyperparameters. By testing all algorithms under consistent conditions with tuned hyperparameters, CTBENCH reveals that most certified training methods perform better than previously reported, setting new benchmarks. Through CTBENCH, authors uncover several interesting properties of models training with certified methods.", "review_text": "The paper presents CTBENCH, a standardized library and benchmark designed to fairly evaluate certified training algorithms for neural networks, addressing the inconsistency in previous evaluations due to varied training schedules, certification methods, and under-optimized hyperparameters. By testing all algorithms under consistent conditions with tuned hyperparameters, CTBENCH reveals that most certified training methods perform better than previously reported, setting new benchmarks. Through CTBENCH, authors uncover several interesting properties of models training with certified methods.", "strengths": "1. The paper proposes a new benchmark for certified robustness methods for image classifiers.\n2. Authors implement several prominent certified robustness methods in a unified framework, thereby standardizing the implementations to facilitate future research.\n3. Furthermore, authors correct implementation mistakes and perform systematic hyperparameter tuning to fully realize the potential of all methods.\n4. Authors present several interesting findings regarding the properties of ceritified robustness methods, for example, models trained using distinct methods have a high overlap in the examples they succeed and fail on, uncovering a sample-specific inherent difficulty level that can be leveraged to improve training. And, these methods can boost OOD generalization for specific corruptions, and hurt generalization for others.", "weaknesses": "1. Authors incorrectly state that the benchmark from Li et al. is not up to date as \"it reports 89% and 51% best certified accuracy for MNIST epsilon = 0.3 and CIFAR-10 epsilon = 2/255 in its evaluation, respectively, while recent methods have achieved more than 93% and 62%\". However, at the time of this review, the numbers on Li et al.'s leaderboard (https://sokcertifiedrobustness.github.io/leaderboard/) are even higher than 93% and 62%, they are 94.02% and 68.2%. Furthermore, the leaderboard toppers are defenses from 2019/2021. It appears that the authors might have pulled their numbers from a stale source.\n2. In order to be an improvement over the existing benchmark (of Li et.al.), one important requirement is comprable or improved comprehensiveness. Based on the results in the paper, the proposed benchmark is significantly less comprehensive than Li et. al. on two important directions: (i) number of defenses evaluated, (ii) number of diverse models used during evaluation. While I understand that the proposed work can be made more comprehensive by running more experiments, this is not the case currently and so is worth poining out.\n3. Furthermore, as stated in the limitations section, the propsoed benchmark only focuses on deterministic certified robustness in the L_infinity space. Whereas, Li et. al.'s benchmark uses both determinisitc and probabilistic certified methods, and covers all the popularly used norms in literature (i.e., L_1, L_2, L_infinity). Thereby further hurting the comprehensiveness of the proposed benchmark.\n4. Some of the findings presented in this paper are expected and already established by prior works (see Questions).\n5. The main contribution of the paper is a unified code-based (and benchmark) for promiment certified robustness methods. Even though authors uncover several interesting findings while reproducing and tuning sota methods, the nature of the contributions of this paper are heavily empirical (not enough technical novelty). As such, this paper is much better suited for venues like TMLR that put emphasis on contributions of such nature.", "questions": "1. It is already well established by previous works that robustness training increases local smoothness. What is unique about the findings presented in this paper?\n2. It is also previously established that adversarially robust trianing methods tend to have higher sample complexity, and therefore are more likely to overfit (less regularization). Other than the choice of metric, what is unique about the findings in Section 5.4.?\n3. Is there an explanation for why the model performs worse for certain corruptions? How will these results be affected if we use different L_p norms? For example, I would expect a model trained to be robust in the L_2 space to be better resistant to Gaussian noise and less resistant to salt and pepper noise.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents CTBENCH, a standardized library and benchmark designed to fairly evaluate certified training algorithms for neural networks, addressing the inconsistency in previous evaluations due to varied training schedules, certification methods, and under-optimized hyperparameters. By testing all algorithms under consistent conditions with tuned hyperparameters, CTBENCH reveals that most certified training methods perform better than previously reported, setting new benchmarks. Through CTBENCH, authors uncover several interesting properties of models training with certified methods.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "1. The paper proposes a new benchmark for certified robustness methods for image classifiers.\n2. Authors implement several prominent certified robustness methods in a unified framework, thereby standardizing the implementations to facilitate future research.\n3. Furthermore, authors correct implementation mistakes and perform systematic hyperparameter tuning to fully realize the potential of all methods.\n4. Authors present several interesting findings regarding the properties of ceritified robustness methods, for example, models trained using distinct methods have a high overlap in the examples they succeed and fail on, uncovering a sample-specific inherent difficulty level that can be leveraged to improve training. And, these methods can boost OOD generalization for specific corruptions, and hurt generalization for others.", "weaknesses": "1. Authors incorrectly state that the benchmark from Li et al. is not up to date as \"it reports 89% and 51% best certified accuracy for MNIST epsilon = 0.3 and CIFAR-10 epsilon = 2/255 in its evaluation, respectively, while recent methods have achieved more than 93% and 62%\". However, at the time of this review, the numbers on Li et al.'s leaderboard (https://sokcertifiedrobustness.github.io/leaderboard/) are even higher than 93% and 62%, they are 94.02% and 68.2%. Furthermore, the leaderboard toppers are defenses from 2019/2021. It appears that the authors might have pulled their numbers from a stale source.\n2. In order to be an improvement over the existing benchmark (of Li et.al.), one important requirement is comprable or improved comprehensiveness. Based on the results in the paper, the proposed benchmark is significantly less comprehensive than Li et. al. on two important directions: (i) number of defenses evaluated, (ii) number of diverse models used during evaluation. While I understand that the proposed work can be made more comprehensive by running more experiments, this is not the case currently and so is worth poining out.\n3. Furthermore, as stated in the limitations section, the propsoed benchmark only focuses on deterministic certified robustness in the L_infinity space. Whereas, Li et. al.'s benchmark uses both determinisitc and probabilistic certified methods, and covers all the popularly used norms in literature (i.e., L_1, L_2, L_infinity). Thereby further hurting the comprehensiveness of the proposed benchmark.\n4. Some of the findings presented in this paper are expected and already established by prior works (see Questions).\n5. The main contribution of the paper is a unified code-based (and benchmark) for promiment certified robustness methods. Even though authors uncover several interesting findings while reproducing and tuning sota methods, the nature of the contributions of this paper are heavily empirical (not enough technical novelty). As such, this paper is much better suited for venues like TMLR that put emphasis on contributions of such nature.", "questions": "1. It is already well established by previous works that robustness training increases local smoothness. What is unique about the findings presented in this paper?\n2. It is also previously established that adversarially robust trianing methods tend to have higher sample complexity, and therefore are more likely to overfit (less regularization). Other than the choice of metric, what is unique about the findings in Section 5.4.?\n3. Is there an explanation for why the model performs worse for certain corruptions? How will these results be affected if we use different L_p norms? For example, I would expect a model trained to be robust in the L_2 space to be better resistant to Gaussian noise and less resistant to salt and pepper noise.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731199412945}, {"id": "CQvJhhdmy4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6522/Reviewer_Z4Eb"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "The paper introduces a benchmark for certified robust training. The goal is to standardize the hyperparameters, training schedules (& other training configurations) between competing methods in certified training. The purported advantages of newer methods are lower when older baselines were given equivalent optimization and testing conditions. The work covers several popular approaches like PGD, IBP, and CROWN-IBP.", "review_text": "The paper introduces a benchmark for certified robust training. The goal is to standardize the hyperparameters, training schedules (& other training configurations) between competing methods in certified training. The purported advantages of newer methods are lower when older baselines were given equivalent optimization and testing conditions. The work covers several popular approaches like PGD, IBP, and CROWN-IBP.", "strengths": "-- The paper is well-written and typeset well\n\n-- Tackles an important problem in the field: the inconsistent evaluation of different certified training methods. I think the field needed this kind of paper. \n\n-- It's not only a benchmark paper but provides some analysis into certified model behavior in loss fragmentation (showing certified models reduce fragmentation compared to adversarial training), have shared mistake patterns, model utilization metrics, and generalization performance (showing certified training provides benefits for certain types of corruptions).", "weaknesses": "-- The novelty of the paper is limited since it's just focused on benchmarking existing methods. Certified robustness is a relatively new field and the field needs methods as much as unifying benchmarks. I do believe the lack of novelty is mitigated to an extent by the analysis provided in Section 5.\n\n-- I wonder about the sustainability of the benchmark since there are other leaderboards for adversarial training (e.g. RobustBench). Others may want to submit their work to an existing leaderboard rather than standardize to adopt your settings.\n\n-- I'm a bit confused about the purpose of the fragmentation experiments. Robust models lead to fewer flipped neurons in the presence of noise, but why should we care? This is after all expected given they are more robust in general to input noise. I believe these experiments may be valuable but the authors should articulate why.", "questions": "Some questions I had while reading:\n\n-- Why do methods like TAPS and MTL-IBP achieve better accuracy while deactivating more neurons?\n\n-- Is there a theoretical framework to explain the relationship between neuron deactivation and robustness? \n\n-- Is there a way to understand and leverage the shared mistakes patterns to improve certified training? Or is it natural that mistakes would overlap (similar to how mistakes overlap in natural training)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a benchmark for certified robust training. The goal is to standardize the hyperparameters, training schedules (& other training configurations) between competing methods in certified training. The purported advantages of newer methods are lower when older baselines were given equivalent optimization and testing conditions. The work covers several popular approaches like PGD, IBP, and CROWN-IBP.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "-- The paper is well-written and typeset well\n\n-- Tackles an important problem in the field: the inconsistent evaluation of different certified training methods. I think the field needed this kind of paper. \n\n-- It's not only a benchmark paper but provides some analysis into certified model behavior in loss fragmentation (showing certified models reduce fragmentation compared to adversarial training), have shared mistake patterns, model utilization metrics, and generalization performance (showing certified training provides benefits for certain types of corruptions).", "weaknesses": "-- The novelty of the paper is limited since it's just focused on benchmarking existing methods. Certified robustness is a relatively new field and the field needs methods as much as unifying benchmarks. I do believe the lack of novelty is mitigated to an extent by the analysis provided in Section 5.\n\n-- I wonder about the sustainability of the benchmark since there are other leaderboards for adversarial training (e.g. RobustBench). Others may want to submit their work to an existing leaderboard rather than standardize to adopt your settings.\n\n-- I'm a bit confused about the purpose of the fragmentation experiments. Robust models lead to fewer flipped neurons in the presence of noise, but why should we care? This is after all expected given they are more robust in general to input noise. I believe these experiments may be valuable but the authors should articulate why.", "questions": "Some questions I had while reading:\n\n-- Why do methods like TAPS and MTL-IBP achieve better accuracy while deactivating more neurons?\n\n-- Is there a theoretical framework to explain the relationship between neuron deactivation and robustness? \n\n-- Is there a way to understand and leverage the shared mistakes patterns to improve certified training? Or is it natural that mistakes would overlap (similar to how mistakes overlap in natural training)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730671730968}, {"id": "O9d1QZmAnq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6522/Reviewer_t9Mr"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a library for benchmarking certified training methods under unified settings. It uses the best practices for certified training from (Shi et al., 2021), such as CNN7 architecture with batch normalization, IBP initialization, warm-up schedule and warm-up regularizers. To improve generalization, it uses L1 regularization and stochastic weight averaging (Izmailov et al., 2018). From the implementation perspective, the authors propose to use full batch statistics to address problems with batch normalization when gradient accumulation or PGD attack is performed. The paper claims that the improvements of recent methods in certified training drop significantly compared to older IBP training method under the same settings with proper hyperparameter tuning. Further, the authors analyze different aspects of the training methods: regularization strength, model utilization, loss fragmentation, OOD generalization and shared mistakes.", "review_text": "This paper proposes a library for benchmarking certified training methods under unified settings. It uses the best practices for certified training from (Shi et al., 2021), such as CNN7 architecture with batch normalization, IBP initialization, warm-up schedule and warm-up regularizers. To improve generalization, it uses L1 regularization and stochastic weight averaging (Izmailov et al., 2018). From the implementation perspective, the authors propose to use full batch statistics to address problems with batch normalization when gradient accumulation or PGD attack is performed. The paper claims that the improvements of recent methods in certified training drop significantly compared to older IBP training method under the same settings with proper hyperparameter tuning. Further, the authors analyze different aspects of the training methods: regularization strength, model utilization, loss fragmentation, OOD generalization and shared mistakes.", "strengths": "- The paper raises an important question of fairly assessing the algorithmic improvements of recent certified training methods compared to older IBP-based training. Since the evaluation depends on many factors and components, the paper proposes to fix some of them to the best-known ones and to properly tune the rest. \n- The writing is clear (except for the presentation of Table 1), the code for benchmarking, and the weights of pre-trained models are provided. \n- The analysis of training methods leads to interesting conclusions. Particularly, the relationship between propagation tightness and certified accuracy at larger epsilon, i.e. the absence of correlation, is surprising.", "weaknesses": "I believe the **experiments are insufficient** to support the main claims of the paper. Particularly:\n\n1. **Accuracy-robustness tradeoffs are not considered**. Improvements in robustness can be due to decreased natural accuracy, and vice versa [a, b, c]. For example, in Table 1 for CIFAR-10 at 2/255 the implementations of following methods choose a different point at accuracy-robustness tradeoff curve compared to the one in literature, getting higher robustness at the cost of reduced accuracy: CROWN-IBP, SABR, STAPS, MTL-IBP, making claims about technical improvements unsupported. In this regard, the baselines such as ACERT [a], and ACE [b] are missing. Accuracy-robustness tradeoff curves and metrics such as ART-score [a] can be used to capture the improvements in the tradeoff.\n2. **Error bars are missing**.  The presented improvements over the results in the literature could be statistically insignificant. For example, the experimental results for CIFAR-10 at 8/255 in paper by Shi et al. (2021) show standard deviation of $\\pm0.3$ for certified accuracy and of $\\pm0.4-0.7$ for natural accuracy, which makes improvements in both accuracy and robustness in Table 1 for SABR and TAPS within the error of standard deviation. \n3. **Training costs are not considered**. Different methods require different amount of computational costs for training, which could be an important factor to consider in benchmarking.\n4. **Certification costs are not considered**. Since some certified training methods allow computing tight certified bounds using efficient \"online\" certification methods, such as IBP (Gowal et al., 2018, Mao et al., 2024), the IBP-based certified accuracy or IBP-based certified radius [a] could also be compared. The cost of test-time verification might be an important factor in choosing a training method.\n\nSince this is a paper proposing a benchmark, it **lacks original** contributions. In terms of evaluation setting, most of the components were already used consistently in previous works.\n\nSmaller comments:\n- The main results in Table 1 are hard to parse and analyze due to large amount of numbers to compare. Accuracy-robustness plots could help with improving clarity.\n- Due to shared mistakes, the paper claims that \"_... there could be an intrinsic difficulty score for each input_\". The certified radius of robustness of each point, described in [a, d], could serve as such score. The average certified radius and/or the histogram of radii [d] can be compared in the benchmark. The adaptive training methods can be discussed in this regard.\n\n[a] Nurlanov, Z., Schmidt, F.R., Bernard, F. (2024). Adaptive Certified Training: Towards Better Accuracy-Robustness Tradeoffs. In: Bifet, A., et al. Machine Learning and Knowledge Discovery in Databases. Research Track and Demo Track. ECML PKDD 2024. Lecture Notes in Computer Science(), vol 14948. Springer, Cham. https://doi.org/10.1007/978-3-031-70371-3_8\n\n[b] Müller, M. N., Balunović, M., & Vechev, M. (2021). Certify or predict: Boosting certified robustness with compositional architectures. In International Conference on Learning Representations (ICLR 2021).\n\n[c] Tsipras, D., Santurkar, S., Engstrom, L., Turner, A., & Madry, A (2019). Robustness May Be at Odds with Accuracy. In International Conference on Learning Representations (ICLR 2019).\n\n[d] Bosman, A. W., Hoos, H. H., & van Rijn, J. N. (2023). A preliminary study of critical robustness distributions in neural network verification. In Proceedings of the 6th workshop on formal methods for ML-enabled autonomous systems.", "questions": "The main concerns about the experiments are raised in the weaknesses section. If these can be addressed, I would be happy to change my opinion.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a library for benchmarking certified training methods under unified settings. It uses the best practices for certified training from (Shi et al., 2021), such as CNN7 architecture with batch normalization, IBP initialization, warm-up schedule and warm-up regularizers. To improve generalization, it uses L1 regularization and stochastic weight averaging (Izmailov et al., 2018). From the implementation perspective, the authors propose to use full batch statistics to address problems with batch normalization when gradient accumulation or PGD attack is performed. The paper claims that the improvements of recent methods in certified training drop significantly compared to older IBP training method under the same settings with proper hyperparameter tuning. Further, the authors analyze different aspects of the training methods: regularization strength, model utilization, loss fragmentation, OOD generalization and shared mistakes.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper raises an important question of fairly assessing the algorithmic improvements of recent certified training methods compared to older IBP-based training. Since the evaluation depends on many factors and components, the paper proposes to fix some of them to the best-known ones and to properly tune the rest. \n- The writing is clear (except for the presentation of Table 1), the code for benchmarking, and the weights of pre-trained models are provided. \n- The analysis of training methods leads to interesting conclusions. Particularly, the relationship between propagation tightness and certified accuracy at larger epsilon, i.e. the absence of correlation, is surprising.", "weaknesses": "I believe the **experiments are insufficient** to support the main claims of the paper. Particularly:\n\n1. **Accuracy-robustness tradeoffs are not considered**. Improvements in robustness can be due to decreased natural accuracy, and vice versa [a, b, c]. For example, in Table 1 for CIFAR-10 at 2/255 the implementations of following methods choose a different point at accuracy-robustness tradeoff curve compared to the one in literature, getting higher robustness at the cost of reduced accuracy: CROWN-IBP, SABR, STAPS, MTL-IBP, making claims about technical improvements unsupported. In this regard, the baselines such as ACERT [a], and ACE [b] are missing. Accuracy-robustness tradeoff curves and metrics such as ART-score [a] can be used to capture the improvements in the tradeoff.\n2. **Error bars are missing**.  The presented improvements over the results in the literature could be statistically insignificant. For example, the experimental results for CIFAR-10 at 8/255 in paper by Shi et al. (2021) show standard deviation of $\\pm0.3$ for certified accuracy and of $\\pm0.4-0.7$ for natural accuracy, which makes improvements in both accuracy and robustness in Table 1 for SABR and TAPS within the error of standard deviation. \n3. **Training costs are not considered**. Different methods require different amount of computational costs for training, which could be an important factor to consider in benchmarking.\n4. **Certification costs are not considered**. Since some certified training methods allow computing tight certified bounds using efficient \"online\" certification methods, such as IBP (Gowal et al., 2018, Mao et al., 2024), the IBP-based certified accuracy or IBP-based certified radius [a] could also be compared. The cost of test-time verification might be an important factor in choosing a training method.\n\nSince this is a paper proposing a benchmark, it **lacks original** contributions. In terms of evaluation setting, most of the components were already used consistently in previous works.\n\nSmaller comments:\n- The main results in Table 1 are hard to parse and analyze due to large amount of numbers to compare. Accuracy-robustness plots could help with improving clarity.\n- Due to shared mistakes, the paper claims that \"_... there could be an intrinsic difficulty score for each input_\". The certified radius of robustness of each point, described in [a, d], could serve as such score. The average certified radius and/or the histogram of radii [d] can be compared in the benchmark. The adaptive training methods can be discussed in this regard.\n\n[a] Nurlanov, Z., Schmidt, F.R., Bernard, F. (2024). Adaptive Certified Training: Towards Better Accuracy-Robustness Tradeoffs. In: Bifet, A., et al. Machine Learning and Knowledge Discovery in Databases. Research Track and Demo Track. ECML PKDD 2024. Lecture Notes in Computer Science(), vol 14948. Springer, Cham. https://doi.org/10.1007/978-3-031-70371-3_8\n\n[b] Müller, M. N., Balunović, M., & Vechev, M. (2021). Certify or predict: Boosting certified robustness with compositional architectures. In International Conference on Learning Representations (ICLR 2021).\n\n[c] Tsipras, D., Santurkar, S., Engstrom, L., Turner, A., & Madry, A (2019). Robustness May Be at Odds with Accuracy. In International Conference on Learning Representations (ICLR 2019).\n\n[d] Bosman, A. W., Hoos, H. H., & van Rijn, J. N. (2023). A preliminary study of critical robustness distributions in neural network verification. In Proceedings of the 6th workshop on formal methods for ML-enabled autonomous systems.", "questions": "The main concerns about the experiments are raised in the weaknesses section. If these can be addressed, I would be happy to change my opinion.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730473397523}, {"id": "N9Gvc2DOVg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6522/Reviewer_s9m2"], "rating": 5, "soundness": 4, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "The The paper introduces CTBENCH, a unified library and benchmark for evaluating certified training methods for neural networks. It addresses the challenges in comparing existing certified training algorithms by standardizing training schedules, certification methods, and hyperparameter tuning. The authors demonstrate that most algorithms in CTBENCH surpass previously reported results, revealing that much of the perceived advantage of newer methods diminishes when outdated baselines are properly tuned. The benchmark provides insights into certified training methods, encouraging future research by offering a consistent framework and re-establishing state-of-the-art performance.", "review_text": "The The paper introduces CTBENCH, a unified library and benchmark for evaluating certified training methods for neural networks. It addresses the challenges in comparing existing certified training algorithms by standardizing training schedules, certification methods, and hyperparameter tuning. The authors demonstrate that most algorithms in CTBENCH surpass previously reported results, revealing that much of the perceived advantage of newer methods diminishes when outdated baselines are properly tuned. The benchmark provides insights into certified training methods, encouraging future research by offering a consistent framework and re-establishing state-of-the-art performance.", "strengths": "1. The paper is well presented, well written, with clear goals and objectives.\n\n2- While this is for a non-expert is not obvious, but the amount of experiments and computation required in this paper is beyond impressive.\n\n3- The insights of the paper are particularly helpful. I personally did not expect that current SOTA methods are under performing. However, it was not that surprising that the improvements over IBP for larger epsilons are not that big.\n\n4- Paper sheds light on a relatively good problem.", "weaknesses": "1. The paper focuses solely on deterministic certified training, overlooking advancements in randomized certified robustness. I believe the paper should have cited works like Cohen et al., Matthias et al. (l1 certification with differential privacy -- early works from 2019), Greg Yang (\"All Shapes and Sizes\" paper), among many others.\n\n2. The paper only considers infinity ball, neglecting other perturbation sets. While this is generally okay, some insights with a few experiments in other perturbation sets might be helpful. It is not clear whether the proposed tricks in the library as part of the unified certified training would work for other perturbation sets (e.g., L2). If they do not, it raises the question of whether we would need a separate library for each perturbation set. The next steps are unclear if that is the case.\n\n3. Some conclusions on the impact of tuning and modifications, while valid, lack formal decomposition, making it difficult to quantify individual contributions. No clarity on the contribution of each individual component (batch norm, etc) towards the final performance. A small systematic study will be very helpful.\n\n4. The evaluation is based on a single model architecture (CNN7); the paper should demonstrate that the library and recommendations hold across different architectures.\n\n\nGeneral comment: Interest in certified models has significantly declined over the past two years. At ECCV, for example, there were notably fewer submissions and accepted papers on adversarial attacks, even though this topic was previously very popular in vision conferences. One reason for this decline could be the uncertainty around where such certifications can be practically deployed, especially given the massive scale of current models, which are thousands of times larger than the CNNs discussed here. Furthermore, as models shift towards generative architectures, it’s unclear who will find this domain relevant. While the paper makes valuable contributions, this direction feels somewhat outdated by about two years and the question of the benefit for it is very unclear and vague, at least to me. I would love to hear the authors take on this.\n\nMinor Comments:\n1. Cite \"is NP-complete\" line 321.\n2.  Is not typical robust accuracy (adv acc) for PGD around 48% on 8/255 CIFAR10? Or is because you use CNN7.\n3. adversarial accuracy is not well defined in line 135. You need to say that it is empirical and serves as an upper bound to the robust accuracy.\n4. certified accuracy defined in 133 is not correct. It should be the portion of *correctly* classified samples that are certifiably robust.", "questions": "See above; I would love to hear the authors comments on each of the weakness above along with a response to the general comment.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The The paper introduces CTBENCH, a unified library and benchmark for evaluating certified training methods for neural networks. It addresses the challenges in comparing existing certified training algorithms by standardizing training schedules, certification methods, and hyperparameter tuning. The authors demonstrate that most algorithms in CTBENCH surpass previously reported results, revealing that much of the perceived advantage of newer methods diminishes when outdated baselines are properly tuned. The benchmark provides insights into certified training methods, encouraging future research by offering a consistent framework and re-establishing state-of-the-art performance.", "soundness": 4, "presentation": 4, "contribution": 2, "strengths": "1. The paper is well presented, well written, with clear goals and objectives.\n\n2- While this is for a non-expert is not obvious, but the amount of experiments and computation required in this paper is beyond impressive.\n\n3- The insights of the paper are particularly helpful. I personally did not expect that current SOTA methods are under performing. However, it was not that surprising that the improvements over IBP for larger epsilons are not that big.\n\n4- Paper sheds light on a relatively good problem.", "weaknesses": "1. The paper focuses solely on deterministic certified training, overlooking advancements in randomized certified robustness. I believe the paper should have cited works like Cohen et al., Matthias et al. (l1 certification with differential privacy -- early works from 2019), Greg Yang (\"All Shapes and Sizes\" paper), among many others.\n\n2. The paper only considers infinity ball, neglecting other perturbation sets. While this is generally okay, some insights with a few experiments in other perturbation sets might be helpful. It is not clear whether the proposed tricks in the library as part of the unified certified training would work for other perturbation sets (e.g., L2). If they do not, it raises the question of whether we would need a separate library for each perturbation set. The next steps are unclear if that is the case.\n\n3. Some conclusions on the impact of tuning and modifications, while valid, lack formal decomposition, making it difficult to quantify individual contributions. No clarity on the contribution of each individual component (batch norm, etc) towards the final performance. A small systematic study will be very helpful.\n\n4. The evaluation is based on a single model architecture (CNN7); the paper should demonstrate that the library and recommendations hold across different architectures.\n\n\nGeneral comment: Interest in certified models has significantly declined over the past two years. At ECCV, for example, there were notably fewer submissions and accepted papers on adversarial attacks, even though this topic was previously very popular in vision conferences. One reason for this decline could be the uncertainty around where such certifications can be practically deployed, especially given the massive scale of current models, which are thousands of times larger than the CNNs discussed here. Furthermore, as models shift towards generative architectures, it’s unclear who will find this domain relevant. While the paper makes valuable contributions, this direction feels somewhat outdated by about two years and the question of the benefit for it is very unclear and vague, at least to me. I would love to hear the authors take on this.\n\nMinor Comments:\n1. Cite \"is NP-complete\" line 321.\n2.  Is not typical robust accuracy (adv acc) for PGD around 48% on 8/255 CIFAR10? Or is because you use CNN7.\n3. adversarial accuracy is not well defined in line 135. You need to say that it is empirical and serves as an upper bound to the robust accuracy.\n4. certified accuracy defined in 133 is not correct. It should be the portion of *correctly* classified samples that are certifiably robust.", "questions": "See above; I would love to hear the authors comments on each of the weakness above along with a response to the general comment.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729595913260}], "openreview_url": "https://openreview.net/forum?id=2bn7gayfz9", "arxiv_id": "2406.04848", "paper_pdf": "papers/2bn7gayfz9.pdf", "paper_pdf_sha256": "f74b6a53ba22f84d39e82c30946768df6f2e691a42e7a2632fd250123dc0b8e7", "paper_pdf_bytes": 422592, "paper_pdf_source": "openreview", "code_url": "https://github.com/eth-sri/CTBench", "code_repository": "eth-sri/CTBench", "code_commit": "0f18162cba85c54d7fc28af62b2c5926efd3695d", "code_archive": "repos/2bn7gayfz9.zip", "code_archive_sha256": "ad1d962e99a18871f250d4d0fb4bb19106535b7c6348e0b8e1a676106214be26", "code_archive_bytes": 214501, "code_file_count": 93, "code_extensions": {".sh": 65, ".py": 28}, "github_disk_usage_kb": 335, "github_languages": {"Python": 578214, "Shell": 58890}, "github_archived": false, "github_pushed_at": "2026-04-25T17:55:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ctbench-a-library-and-benchmark-for-certified"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hJEMTDOwKx", "year": 2024, "status": "rejected", "title": "Language Models as Semantic Indexers", "authors": ["Bowen Jin", "Hansi Zeng", "Guoyin Wang", "Xiusi Chen", "Tianxin Wei", "Ruirui Li", "Zhengyang Wang", "Zheng Li", "Yang Li", "Hanqing Lu", "Suhang Wang", "Jiawei Han", "Xianfeng Tang"], "authorids": ["~Bowen_Jin1", "~Hansi_Zeng1", "~Guoyin_Wang1", "~Xiusi_Chen1", "~Tianxin_Wei1", "~Ruirui_Li3", "~Zhengyang_Wang1", "~Zheng_Li9", "~Yang_Li80", "~Hanqing_Lu3", "~Suhang_Wang1", "~Jiawei_Han1", "~Xianfeng_Tang1"], "authors_source": "OpenReview API", "abstract": "Semantic identifier (ID) is an important concept in information retrieval that aims to preserve the semantics of objects such as documents and items inside their IDs. Previous studies typically adopt a two-stage pipeline to learn semantic IDs by first procuring embeddings using off-the-shelf text encoders and then deriving IDs based on the embeddings. However, each step introduces potential information loss and there is usually an inherent mismatch between the distribution of embeddings within the latent space produced by text encoders and the anticipated distribution required for semantic indexing. Nevertheless, it is non-trivial to design a method that can learn the document’s semantic representations and its hierarchical structure simultaneously, given that semantic IDs are discrete and sequentially structured, and the semantic supervision is deficient. In this paper, we introduce LMINDEXER, a self-supervised framework to learn semantic IDs with a generative language model. We tackle the challenge of sequential discrete ID by introducing a semantic indexer capable of generating neural sequential discrete representations with progressive training and contrastive learning. In response to the semantic supervision deficiency, we propose to train the model with a self-supervised document reconstruction objective. The learned semantic indexer can facilitate various downstream tasks, such as recommendation and retrieval. We conduct experiments on three tasks including recommendation, product search, and document retrieval on five datasets from various domains, where LMINDEXER outperforms competitive baselines significantly and consistently. Code is available at https://anonymous.4open.science/r/ICLR24-submit-B2E7/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "YFGCEvEzDb", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4028/Reviewer_4DWK"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper formulates the problem of learning semantic IDs by simultaneously capturing the document's semantic representations and its hierarchical structure. It introduces an innovative self-supervised approach designed to acquire semantic IDs directly from the input document using a generative language model. Experimental results on five datasets from various domains demonstrate that the proposed method consistently outperforms competitive baselines by a significant margin.", "review_text": "This paper formulates the problem of learning semantic IDs by simultaneously capturing the document's semantic representations and its hierarchical structure. It introduces an innovative self-supervised approach designed to acquire semantic IDs directly from the input document using a generative language model. Experimental results on five datasets from various domains demonstrate that the proposed method consistently outperforms competitive baselines by a significant margin.", "strengths": "1. This paper presents the \"LMINDEXER\" approach as a solution to the challenges inherent in generating semantic IDs from textual data. The approach is carefully crafted to capture both the semantic representations and hierarchical structure of documents simultaneously.\n\n2. The paper demonstrates the effectiveness of the LMINDEXER approach through empirical evidence gathered from experiments on three distinct downstream tasks, utilizing data from diverse domains.\n\n3. The paper exhibits a well-organized structure and offers an easily digestible reading experience.", "weaknesses": "1. While this paper presents an approach termed LMINDEXER, it's important to note that the novelty of the method is somewhat limited. Additionally, the paper lacks a comprehensive discussion of related work, including notable prior efforts that have explored the use of encoders for text encoding and decoders for reconstruction in the context of information retrieval. Several works, such as [1], [2], and [3], have examined similar techniques and deserve acknowledgment for their contributions to the field.\n\n[1] Less is More: Pretrain a Strong Siamese Encoder for Dense Text Retrieval Using a Weak Decoder (EMNLP 2021)\n\n[2] A Contrastive Pre-training Approach to Learn Discriminative Autoencoder for Dense Retrieval (CIKM 2022)\n\n[3] RetroMAE: Pre-Training Retrieval-oriented Language Models Via Masked Auto-Encoder (EMNLP 2022)\n\n\n2. In the context of document retrieval, it would be beneficial to broaden the comparison to include more generative retrieval baselines. For instance, evaluating how the LMINDEXER approach compares to FM-index-based SEAL or other recent generative retrieval methods would provide a more comprehensive understanding of its performance. Focusing solely on comparisons with DSI, which may be considered a relatively weaker baseline, might not offer a complete picture of the method's capabilities.\n\n3. While the proposed self-supervised approach has the potential to address the \"new document problem\" by automatically acquiring semantic IDs, it would be valuable to see experimental results that explicitly address this issue. Incorporating experiments that involve adding new documents to the model and assessing its adaptability and performance in such scenarios would provide a more robust evaluation.\n\n4. An important consideration when using automatically generated semantic IDs is the possibility of duplication. It is crucial to include a discussion or explanation of how the LMINDEXER approach handles or mitigates this potential issue. A detailed exploration of the approach's robustness in preventing or addressing duplication would enhance the paper's completeness and practicality.", "questions": "1. The authors should consider expanding their related work section to include the most recent developments in generative Information Retrieval (IR) techniques. Additionally, they should include a comparative analysis of these recent works alongside the proposed LMINDEXER method in the experimental section. This would provide a more comprehensive overview of how LMINDEXER stacks up against the state-of-the-art in generative IR.\n\n2. It is essential for the authors to address the issue of handling new documents within the LMIndexer framework. The paper should discuss how this framework manages the addition of new documents, what mechanisms or strategies are employed, and the performance of the LMIndexer approach in comparison to baseline methods when confronted with this \"new document\" scenario. This analysis would help assess the adaptability and robustness of the approach.\n\n3. To ensure that distinct semantic IDs are assigned to different documents, the paper should provide detailed explanations and discussions regarding the mechanisms and safeguards in place within the LMIndexer framework. Experimental results or case studies showcasing how the system maintains distinct semantic IDs for various documents would add substantial value to the paper, reinforcing its practicality and effectiveness in preventing semantic ID duplication.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper formulates the problem of learning semantic IDs by simultaneously capturing the document's semantic representations and its hierarchical structure. It introduces an innovative self-supervised approach designed to acquire semantic IDs directly from the input document using a generative language model. Experimental results on five datasets from various domains demonstrate that the proposed method consistently outperforms competitive baselines by a significant margin.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. This paper presents the \"LMINDEXER\" approach as a solution to the challenges inherent in generating semantic IDs from textual data. The approach is carefully crafted to capture both the semantic representations and hierarchical structure of documents simultaneously.\n\n2. The paper demonstrates the effectiveness of the LMINDEXER approach through empirical evidence gathered from experiments on three distinct downstream tasks, utilizing data from diverse domains.\n\n3. The paper exhibits a well-organized structure and offers an easily digestible reading experience.", "weaknesses": "1. While this paper presents an approach termed LMINDEXER, it's important to note that the novelty of the method is somewhat limited. Additionally, the paper lacks a comprehensive discussion of related work, including notable prior efforts that have explored the use of encoders for text encoding and decoders for reconstruction in the context of information retrieval. Several works, such as [1], [2], and [3], have examined similar techniques and deserve acknowledgment for their contributions to the field.\n\n[1] Less is More: Pretrain a Strong Siamese Encoder for Dense Text Retrieval Using a Weak Decoder (EMNLP 2021)\n\n[2] A Contrastive Pre-training Approach to Learn Discriminative Autoencoder for Dense Retrieval (CIKM 2022)\n\n[3] RetroMAE: Pre-Training Retrieval-oriented Language Models Via Masked Auto-Encoder (EMNLP 2022)\n\n\n2. In the context of document retrieval, it would be beneficial to broaden the comparison to include more generative retrieval baselines. For instance, evaluating how the LMINDEXER approach compares to FM-index-based SEAL or other recent generative retrieval methods would provide a more comprehensive understanding of its performance. Focusing solely on comparisons with DSI, which may be considered a relatively weaker baseline, might not offer a complete picture of the method's capabilities.\n\n3. While the proposed self-supervised approach has the potential to address the \"new document problem\" by automatically acquiring semantic IDs, it would be valuable to see experimental results that explicitly address this issue. Incorporating experiments that involve adding new documents to the model and assessing its adaptability and performance in such scenarios would provide a more robust evaluation.\n\n4. An important consideration when using automatically generated semantic IDs is the possibility of duplication. It is crucial to include a discussion or explanation of how the LMINDEXER approach handles or mitigates this potential issue. A detailed exploration of the approach's robustness in preventing or addressing duplication would enhance the paper's completeness and practicality.", "questions": "1. The authors should consider expanding their related work section to include the most recent developments in generative Information Retrieval (IR) techniques. Additionally, they should include a comparative analysis of these recent works alongside the proposed LMINDEXER method in the experimental section. This would provide a more comprehensive overview of how LMINDEXER stacks up against the state-of-the-art in generative IR.\n\n2. It is essential for the authors to address the issue of handling new documents within the LMIndexer framework. The paper should discuss how this framework manages the addition of new documents, what mechanisms or strategies are employed, and the performance of the LMIndexer approach in comparison to baseline methods when confronted with this \"new document\" scenario. This analysis would help assess the adaptability and robustness of the approach.\n\n3. To ensure that distinct semantic IDs are assigned to different documents, the paper should provide detailed explanations and discussions regarding the mechanisms and safeguards in place within the LMIndexer framework. Experimental results or case studies showcasing how the system maintains distinct semantic IDs for various documents would add substantial value to the paper, reinforcing its practicality and effectiveness in preventing semantic ID duplication.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698824512477}, {"id": "qBhsyYlNqD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4028/Reviewer_dU5G"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose LMIndexer, a self-supervised method to learn the semantic IDs of documents using sequence-to-sequence models. A semantic ID of a document is a sequence of integers that are indexes/row numbers of a codebook embedding matrix. Three loss functions are designed and used to train the sequence-to-sequence model: a reconstruction loss, a contrastive loss, and a commitment loss. The proposed method is evaluated on three downstream tasks: sequential recommendation, product search, and document retrieval. The results show good improvement over some SOTA methods.", "review_text": "The authors propose LMIndexer, a self-supervised method to learn the semantic IDs of documents using sequence-to-sequence models. A semantic ID of a document is a sequence of integers that are indexes/row numbers of a codebook embedding matrix. Three loss functions are designed and used to train the sequence-to-sequence model: a reconstruction loss, a contrastive loss, and a commitment loss. The proposed method is evaluated on three downstream tasks: sequential recommendation, product search, and document retrieval. The results show good improvement over some SOTA methods.", "strengths": "-\tThe proposed method formulates the semantic ID learning problem as a sequence-to-sequence learning method, which is novel according to related work discussed in the paper.\n-\tTechnical challenges are described clearly.\n-\tSOTA techniques are used in the proposed framework.\n-\tThe experimental results show that the proposed method outperforms some SOTA methods in the three downstream tasks.", "weaknesses": "-\tThe paper says that the proposed method \"learns the document’s discrete semantic embeddings and its hierarchical structure simultaneously\". But it is not clear what the authors mean by the hierarchical structure of a document, how the proposed method is guaranteed to learn such a structure, and whether the proposed method actually learns such a structure. \n\n-\tThe size of the semantic ID (T) is set to less than or equal to 3 in the experiments, which is surprisingly small. Figure 5 shows the performance increases with T. Why not trying bigger T values? Would an ID with length 3 be informative enough?\n\n-   How should the codebook size be determined? It seems that only three sizes are tried in the experiments. How does the size of the codebook affect the performance?\t\n\n-  The performance metric used in ID quantitative study, AMI (in Table 1), is not defined or explained. \n\n-   Not clear what the word clouds picture (Figure 3) is trying to show? The text explaining it is confusing. What do you mean by “two semantic ID prefixes”? Are the two prefixes from the same generated ID or two different IDs? \n\n-   It would be better if authors provided other performance metrics such as latency for the retrieval in comparison with baselines.\n\n-  The authors mentioned the model could be fine-tuned on downstream tasks such as retrieval or recommendation tasks. Fine-tuning would change the model weights and then the previously generated semantic IDs would be changed as well, which may affect the ground-truth ID used in fine-tuning. Does it cause any problem?", "questions": "Please see the questions in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose LMIndexer, a self-supervised method to learn the semantic IDs of documents using sequence-to-sequence models. A semantic ID of a document is a sequence of integers that are indexes/row numbers of a codebook embedding matrix. Three loss functions are designed and used to train the sequence-to-sequence model: a reconstruction loss, a contrastive loss, and a commitment loss. The proposed method is evaluated on three downstream tasks: sequential recommendation, product search, and document retrieval. The results show good improvement over some SOTA methods.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "-\tThe proposed method formulates the semantic ID learning problem as a sequence-to-sequence learning method, which is novel according to related work discussed in the paper.\n-\tTechnical challenges are described clearly.\n-\tSOTA techniques are used in the proposed framework.\n-\tThe experimental results show that the proposed method outperforms some SOTA methods in the three downstream tasks.", "weaknesses": "-\tThe paper says that the proposed method \"learns the document’s discrete semantic embeddings and its hierarchical structure simultaneously\". But it is not clear what the authors mean by the hierarchical structure of a document, how the proposed method is guaranteed to learn such a structure, and whether the proposed method actually learns such a structure. \n\n-\tThe size of the semantic ID (T) is set to less than or equal to 3 in the experiments, which is surprisingly small. Figure 5 shows the performance increases with T. Why not trying bigger T values? Would an ID with length 3 be informative enough?\n\n-   How should the codebook size be determined? It seems that only three sizes are tried in the experiments. How does the size of the codebook affect the performance?\t\n\n-  The performance metric used in ID quantitative study, AMI (in Table 1), is not defined or explained. \n\n-   Not clear what the word clouds picture (Figure 3) is trying to show? The text explaining it is confusing. What do you mean by “two semantic ID prefixes”? Are the two prefixes from the same generated ID or two different IDs? \n\n-   It would be better if authors provided other performance metrics such as latency for the retrieval in comparison with baselines.\n\n-  The authors mentioned the model could be fine-tuned on downstream tasks such as retrieval or recommendation tasks. Fine-tuning would change the model weights and then the previously generated semantic IDs would be changed as well, which may affect the ground-truth ID used in fine-tuning. Does it cause any problem?", "questions": "Please see the questions in the above section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698818037028}, {"id": "ufgPDJGXSv", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4028/Reviewer_HnqK"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Discrete semantic IDs are useful in information retrieval tasks. They are often learned by performing some sort of hierarchical clustering over off-the-shelf item representations which is not ideal as they may not be aligned with the downstream task. This paper presents an approach to learn discrete semantic IDs of items in a self-supervised manner. The proposed approach uses a transformer decoder architecture which encodes the item description and decodes it into semantic IDs which is further coupled with a small transformer model that consumes these semantic IDs, item description and tries to perform MLM task. The paper also describes and suggests ways to circumvent the challenges encountered while learning these semantic IDs.", "review_text": "Discrete semantic IDs are useful in information retrieval tasks. They are often learned by performing some sort of hierarchical clustering over off-the-shelf item representations which is not ideal as they may not be aligned with the downstream task. This paper presents an approach to learn discrete semantic IDs of items in a self-supervised manner. The proposed approach uses a transformer decoder architecture which encodes the item description and decodes it into semantic IDs which is further coupled with a small transformer model that consumes these semantic IDs, item description and tries to perform MLM task. The paper also describes and suggests ways to circumvent the challenges encountered while learning these semantic IDs.", "strengths": "- The paper is in general well-written and easy to follow\n- The approach is novel for learning semantic IDs in information retrieval, and the optimization challenges in such learning problems is highlighted", "weaknesses": "My major concerns with the paper are the weak evaluation and baselines, and overall the training seems to need a lot of bells and whistles to succeed.\n- DPR dual encoder is not a strong baseline; DPR is almost 10% behind SOTA dual-encoder approaches on standard benchmarks\n- Baselines in section 4.2 are weak since they are using an off-the-shelf text encoder and hence have no knowledge about the task; a very simple baseline that could be tried here is to train a dual-encoder model on this corpus and then cluster the embeddings", "questions": "- The learned semantic ID lengths seem to be very small (1-3), is this because of training instability when scaling to larger ID lengths? Do the other generative baselines also use the same ID lengths?  \n- Why not use the full MS-Marco dataset instead of the 1M sampled one?\n- How does the performance gets affected when using a more powerful reconstructor? perhaps an ablation on the number of layers in the reconstructor might be helpful here\n- Is the contrastive loss $\\mathcal{L}_{\\text{contrastive}}$ computed over all documents?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Discrete semantic IDs are useful in information retrieval tasks. They are often learned by performing some sort of hierarchical clustering over off-the-shelf item representations which is not ideal as they may not be aligned with the downstream task. This paper presents an approach to learn discrete semantic IDs of items in a self-supervised manner. The proposed approach uses a transformer decoder architecture which encodes the item description and decodes it into semantic IDs which is further coupled with a small transformer model that consumes these semantic IDs, item description and tries to perform MLM task. The paper also describes and suggests ways to circumvent the challenges encountered while learning these semantic IDs.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is in general well-written and easy to follow\n- The approach is novel for learning semantic IDs in information retrieval, and the optimization challenges in such learning problems is highlighted", "weaknesses": "My major concerns with the paper are the weak evaluation and baselines, and overall the training seems to need a lot of bells and whistles to succeed.\n- DPR dual encoder is not a strong baseline; DPR is almost 10% behind SOTA dual-encoder approaches on standard benchmarks\n- Baselines in section 4.2 are weak since they are using an off-the-shelf text encoder and hence have no knowledge about the task; a very simple baseline that could be tried here is to train a dual-encoder model on this corpus and then cluster the embeddings", "questions": "- The learned semantic ID lengths seem to be very small (1-3), is this because of training instability when scaling to larger ID lengths? Do the other generative baselines also use the same ID lengths?  \n- Why not use the full MS-Marco dataset instead of the 1M sampled one?\n- How does the performance gets affected when using a more powerful reconstructor? perhaps an ablation on the number of layers in the reconstructor might be helpful here\n- Is the contrastive loss $\\mathcal{L}_{\\text{contrastive}}$ computed over all documents?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698790056529}, {"id": "OsI288env2", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4028/Reviewer_2Nr6"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Semantic id is an interesting and promising topic.\nThis task reform counting id to the identifier with semantic, which will faciliate many intelligence application including e-commerce.\nThis work presents a self-supervised framework to learn semantic IDs with a generative language model.\nBy progressive learning and contrastive learning, the authors achieve sequential discrete semantic indexier.\nThe paper is well-motivated and well-written.\n\nMajor Concerns:\n1. We have not seen a full section to anlayze the semantic IDs by demonstrating all sorts of key examples.\nAs a rising domain, many readers may wonder what the semantic IDs exactly look like and how they benefit the downstream tasks.\n2. In your experiments, could you provide some human annotator evaluation for the semantic of your generated IDs?\nThis is a key experiment, though we have seen enough assessment.", "review_text": "Semantic id is an interesting and promising topic.\nThis task reform counting id to the identifier with semantic, which will faciliate many intelligence application including e-commerce.\nThis work presents a self-supervised framework to learn semantic IDs with a generative language model.\nBy progressive learning and contrastive learning, the authors achieve sequential discrete semantic indexier.\nThe paper is well-motivated and well-written.\n\nMajor Concerns:\n1. We have not seen a full section to anlayze the semantic IDs by demonstrating all sorts of key examples.\nAs a rising domain, many readers may wonder what the semantic IDs exactly look like and how they benefit the downstream tasks.\n2. In your experiments, could you provide some human annotator evaluation for the semantic of your generated IDs?\nThis is a key experiment, though we have seen enough assessment.", "strengths": "Semantic id is an interesting and promising topic.\nThis task reform counting id to the identifier with semantic, which will faciliate many intelligence application including e-commerce.\nThis work presents a self-supervised framework to learn semantic IDs with a generative language model.\nBy progressive learning and contrastive learning, the authors achieve sequential discrete semantic indexier.\nThe paper is well-motivated and well-written.", "weaknesses": "Experimental analysis is not sufficient to support the contribution.", "questions": "Major Concerns:\n1. We have not seen a full section to anlayze the semantic IDs by demonstrating all sorts of key examples.\nAs a rising domain, many readers may wonder what the semantic IDs exactly look like and how they benefit the downstream tasks.\n2. In your experiments, could you provide some human annotator evaluation for the semantic of your generated IDs?\nThis is a key experiment, though we have seen enough assessment.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Semantic id is an interesting and promising topic.\nThis task reform counting id to the identifier with semantic, which will faciliate many intelligence application including e-commerce.\nThis work presents a self-supervised framework to learn semantic IDs with a generative language model.\nBy progressive learning and contrastive learning, the authors achieve sequential discrete semantic indexier.\nThe paper is well-motivated and well-written.\n\nMajor Concerns:\n1. We have not seen a full section to anlayze the semantic IDs by demonstrating all sorts of key examples.\nAs a rising domain, many readers may wonder what the semantic IDs exactly look like and how they benefit the downstream tasks.\n2. In your experiments, could you provide some human annotator evaluation for the semantic of your generated IDs?\nThis is a key experiment, though we have seen enough assessment.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "Semantic id is an interesting and promising topic.\nThis task reform counting id to the identifier with semantic, which will faciliate many intelligence application including e-commerce.\nThis work presents a self-supervised framework to learn semantic IDs with a generative language model.\nBy progressive learning and contrastive learning, the authors achieve sequential discrete semantic indexier.\nThe paper is well-motivated and well-written.", "weaknesses": "Experimental analysis is not sufficient to support the contribution.", "questions": "Major Concerns:\n1. We have not seen a full section to anlayze the semantic IDs by demonstrating all sorts of key examples.\nAs a rising domain, many readers may wonder what the semantic IDs exactly look like and how they benefit the downstream tasks.\n2. In your experiments, could you provide some human annotator evaluation for the semantic of your generated IDs?\nThis is a key experiment, though we have seen enough assessment.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698640152871}], "openreview_url": "https://openreview.net/forum?id=hJEMTDOwKx", "arxiv_id": "2310.07815", "paper_pdf": "papers/hJEMTDOwKx.pdf", "paper_pdf_sha256": "05ff3e71cb4c675b5bfbeef76d4e1b5b04bff322c831d71e98419cc129199f1a", "paper_pdf_bytes": 1254820, "paper_pdf_source": "openreview", "code_url": "https://github.com/PeterGriffinJin/LMIndexer", "code_repository": "PeterGriffinJin/LMIndexer", "code_commit": "e6dc4cc1ddf1f1615f643771c9047063b8e49528", "code_archive": "repos/hJEMTDOwKx.zip", "code_archive_sha256": "8ca1c376945ef5828c23a88ea6cc60771394f74e0a08f2932d6bc2426cc4cb63", "code_archive_bytes": 316165, "code_file_count": 97, "code_extensions": {".py": 51, ".sh": 45, ".ipynb": 1}, "github_disk_usage_kb": 267, "github_languages": {"Python": 630707, "Shell": 83923, "Jupyter Notebook": 19503}, "github_archived": false, "github_pushed_at": "2024-05-02T12:48:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/language-models-as-semantic-indexers"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hUr6K4D9f7P", "year": 2022, "status": "rejected", "title": "Adversarial Weight Perturbation Improves Generalization in Graph Neural Networks", "authors": ["Yihan Wu", "Aleksandar Bojchevski", "Heng Huang"], "authorids": ["~Yihan_Wu1", "~Aleksandar_Bojchevski1", "~Heng_Huang1"], "authors_source": "OpenReview API", "abstract": "There is growing theoretical and empirical evidence that flatter local minima tend to improve generalization. An efficient and effective technique for finding such minima is Adversarial Weight Perturbation (AWP). The main idea is to minimize the loss w.r.t. a bounded worst-case perturbation of the model parameters by (approximately) solving an associated min-max problem. Intuitively, we favor local minima with a small loss in a neighborhood around them. The benefits of AWP, and more generally the connections between flatness and generalization, have been extensively studied for i.i.d. data such as images. In this paper we initiate the first study of this phenomenon for graph data. Along the way, we identify a vanishing-gradient issue with all existing formulations of AWP and we propose Weighted Truncated AWP (WT-AWP) to alleviate this issue. We show that regularizing graph neural networks with WT-AWP consistently improves both natural and robust generalization across many different graph learning tasks and models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gkHybJcv4NX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1934/Reviewer_DWF4"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper designed and tested WT-AWP, a new adversarial weight perturbation approach, on graph neural networks. They demonstrated that by locating flat local minima, our WT-AWP can improve the regularization of GNNs. They carried out comprehensive tests to verify the method. WT-AWP reliably increases GNN performance on a wide range of graph learning tasks, including node classification, graph defense, and graph classification.", "review_text": "This paper focuses on extending AWP to GNN. This paper analyzes the vanishing-gradient issue existing in AWP and gives a more detailed theoretical proof. The experimental part of this article is comprehensive, and the experimental results also verify the advantages of the proposed method.\n\nHowever, there still exist some insufficiency and confusing places.\n\n1, What is the term h in equation(3). The statements said h is a monotonously increasing function. Where is it from?\n\n2, The paper utilized comprehensive experiments to show the efficiency when facing attacks, but no analysis and theoretical proof for this advantage. I think the source of the advantage in handling attacks is worth clarification. Especially, in Table 2, some experimental results under attacks are even better than the natural setting.\n\n3, Figure1 and Figure3 are blurred. I can not distinguish face color and border color. It is recommended to use vector graphics.\n\n4, The sentence on Page 5 footnotes \"Perturbing only the second layer instead performs similarly.\" is confusing for me.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper designed and tested WT-AWP, a new adversarial weight perturbation approach, on graph neural networks. They demonstrated that by locating flat local minima, our WT-AWP can improve the regularization of GNNs. They carried out comprehensive tests to verify the method. WT-AWP reliably increases GNN performance on a wide range of graph learning tasks, including node classification, graph defense, and graph classification.", "main_review": "This paper focuses on extending AWP to GNN. This paper analyzes the vanishing-gradient issue existing in AWP and gives a more detailed theoretical proof. The experimental part of this article is comprehensive, and the experimental results also verify the advantages of the proposed method.\n\nHowever, there still exist some insufficiency and confusing places.\n\n1, What is the term h in equation(3). The statements said h is a monotonously increasing function. Where is it from?\n\n2, The paper utilized comprehensive experiments to show the efficiency when facing attacks, but no analysis and theoretical proof for this advantage. I think the source of the advantage in handling attacks is worth clarification. Especially, in Table 2, some experimental results under attacks are even better than the natural setting.\n\n3, Figure1 and Figure3 are blurred. I can not distinguish face color and border color. It is recommended to use vector graphics.\n\n4, The sentence on Page 5 footnotes \"Perturbing only the second layer instead performs similarly.\" is confusing for me.\n", "summary_of_the_review": "The paper is well-organized and has comprehensive experiments to defense its method. However, some statements and equations are not clear enough. The illustrations and figures are not meticulous. I hope the authors can further polish the paper in detail.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635875052517}, {"id": "RI2I8L-IUr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1934/Reviewer_2WEe"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose a variant of adversarial weight perturbations / sharpness aware minimization for graph (convolutional) neural networks for node and graph classification. In particular, they make two adjustments: “truncating”, i.e., limiting the weight perturbation to specific layers, and weighting the sharpness aware loss with the regular loss during training.", "review_text": "Strengths:\n- The paper is generally clearly structured and written; the toy example visualizations are nice and the algorithm helps understanding.\n- An interesting problem/question is addressed: how to effectively use AWP/sharpness-aware minimization for graph neural networks.\n- It is also nice to show that the optimum does not change with AWP training.\n- Experiments for clean and robust accuracy with several models are conducted. Besides mean, standard deviation is reported which is nice to judge the improvement.\n- The approach improves across the board, even though improvement is small at times.\n- Analysis in terms of flatness (gradients and visualization) and ablation studies are provided.\n\nWeaknesses:\n- In the introduction, the authors state the i.i.d assumption as an important criteria for existing flatness approaches. However, it is not discussed when this becomes relevant. As far as I am aware, the i.i.id assumption does not play an important role in related discussion/papers. Could the authors comment on that? As I understand the paper, the main point is to see how these approaches work in non i.i.d settings such as node classification, but the paper does not really theoretically discuss this issue.\n- One paper that could be included in related work as it is quite related:\n\n[a] https://arxiv.org/abs/1609.04836\n\n- Also a discussion of scale-invariance and the criticism of [b] is missing (this should be handled as in Stutz et al., but it would be interesting at least to say why this generalizes to the GCNs considered)\n\n[b] https://arxiv.org/pdf/1703.04933.pdf \n\n- Notation-wise, the main sections could be improved by making explicit that a per-layer neighborhood B is used from the beginning. Currently, this is made clear in Eq. (4) while Eq. (1) to (3) suggest that all parameters are appended.\n- Also, are the GCN baselines equipped with biases in addition to the weights? If so, are biases and weights treated as separate layers (as in Stutz et al.)?\n- In terms of contributions, the theoretical contributions (mainly Thm 1) and methodological contributions seem a bit limited. Although I have not seen Thm 1 in other papers (which makes it refreshing to actually see it), I found that the result is very intuitive and the proof is also quite straight-forward. The techniques employed to improve AWP also seem very specific to graph problems where larger \\rho are used than for vision problems. Thus, I see the contributions mostly in verifying this approach on graph data.\n- Regarding the vanishing gradient problem: I am having difficulties understanding why \\rho has to be chosen as large as done throughout the toy example and the experiments. Stutz et al. And Wu et al. Consider very small \\rho of 0.5% (=0.005 or lower). Obviously these are much deeper networks and not graph neural networks. I am wondering if the authors could give more details on why a large \\rho is needed. Is it because of the 2-layer structure or the architecture differences?\n- The two proposed approaches do actually not improve performance on the toy example. While I understand that it is meant for illustration purposes, I believe it is badly chosen. I see that without truncation and weighting the accuracy is very bad, but shouldn’t you show that you can improve over the baseline of 98% and not be stuck at 95%? Did you try optimizing hyper-parameters, or is it a problem where TW-AWP just does not help (which would be interesting!)?\n- Regarding 5.2, I am not entirely convinced that the gradient norm is the best indicator of flatness exactly because of the scale-invariance argumentation of Stutz et al.: Can the authors comment on that? I guess that the used GCNs do not use batch normalization (I have not seen BN used for graph neural networks before), but scale-invariance is a problem as described in [b]. Also, Fig. 4 (c) and (d) has some outliers that are not really explained.\n- Table 2 is hard to read and very small.\n- Some ablation that I find missing: ablation regarding layers. Which layer to skip? Obviously, the networks are two-layer networks, but I would find it very interesting whether it is always the last layer to skip (also in deeper networks) or always the first layer to perturb (also in deeper networks).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a variant of adversarial weight perturbations / sharpness aware minimization for graph (convolutional) neural networks for node and graph classification. In particular, they make two adjustments: “truncating”, i.e., limiting the weight perturbation to specific layers, and weighting the sharpness aware loss with the regular loss during training.", "main_review": "Strengths:\n- The paper is generally clearly structured and written; the toy example visualizations are nice and the algorithm helps understanding.\n- An interesting problem/question is addressed: how to effectively use AWP/sharpness-aware minimization for graph neural networks.\n- It is also nice to show that the optimum does not change with AWP training.\n- Experiments for clean and robust accuracy with several models are conducted. Besides mean, standard deviation is reported which is nice to judge the improvement.\n- The approach improves across the board, even though improvement is small at times.\n- Analysis in terms of flatness (gradients and visualization) and ablation studies are provided.\n\nWeaknesses:\n- In the introduction, the authors state the i.i.d assumption as an important criteria for existing flatness approaches. However, it is not discussed when this becomes relevant. As far as I am aware, the i.i.id assumption does not play an important role in related discussion/papers. Could the authors comment on that? As I understand the paper, the main point is to see how these approaches work in non i.i.d settings such as node classification, but the paper does not really theoretically discuss this issue.\n- One paper that could be included in related work as it is quite related:\n\n[a] https://arxiv.org/abs/1609.04836\n\n- Also a discussion of scale-invariance and the criticism of [b] is missing (this should be handled as in Stutz et al., but it would be interesting at least to say why this generalizes to the GCNs considered)\n\n[b] https://arxiv.org/pdf/1703.04933.pdf \n\n- Notation-wise, the main sections could be improved by making explicit that a per-layer neighborhood B is used from the beginning. Currently, this is made clear in Eq. (4) while Eq. (1) to (3) suggest that all parameters are appended.\n- Also, are the GCN baselines equipped with biases in addition to the weights? If so, are biases and weights treated as separate layers (as in Stutz et al.)?\n- In terms of contributions, the theoretical contributions (mainly Thm 1) and methodological contributions seem a bit limited. Although I have not seen Thm 1 in other papers (which makes it refreshing to actually see it), I found that the result is very intuitive and the proof is also quite straight-forward. The techniques employed to improve AWP also seem very specific to graph problems where larger \\rho are used than for vision problems. Thus, I see the contributions mostly in verifying this approach on graph data.\n- Regarding the vanishing gradient problem: I am having difficulties understanding why \\rho has to be chosen as large as done throughout the toy example and the experiments. Stutz et al. And Wu et al. Consider very small \\rho of 0.5% (=0.005 or lower). Obviously these are much deeper networks and not graph neural networks. I am wondering if the authors could give more details on why a large \\rho is needed. Is it because of the 2-layer structure or the architecture differences?\n- The two proposed approaches do actually not improve performance on the toy example. While I understand that it is meant for illustration purposes, I believe it is badly chosen. I see that without truncation and weighting the accuracy is very bad, but shouldn’t you show that you can improve over the baseline of 98% and not be stuck at 95%? Did you try optimizing hyper-parameters, or is it a problem where TW-AWP just does not help (which would be interesting!)?\n- Regarding 5.2, I am not entirely convinced that the gradient norm is the best indicator of flatness exactly because of the scale-invariance argumentation of Stutz et al.: Can the authors comment on that? I guess that the used GCNs do not use batch normalization (I have not seen BN used for graph neural networks before), but scale-invariance is a problem as described in [b]. Also, Fig. 4 (c) and (d) has some outliers that are not really explained.\n- Table 2 is hard to read and very small.\n- Some ablation that I find missing: ablation regarding layers. Which layer to skip? Obviously, the networks are two-layer networks, but I would find it very interesting whether it is always the last layer to skip (also in deeper networks) or always the first layer to perturb (also in deeper networks).", "summary_of_the_review": "I appreciate this paper in terms of applying AWP to graph neural networks and showing how it needs to be adapted to work well. Methodological contributions are small, however, and improvements vary across datasets and cases. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635856751181}, {"id": "RWgiUhPcUSR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1934/Reviewer_MYVq"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors extend the line of work related to adversarial weight perturbations: first showing that a vanishing gradient issue in standard AWP can hinder training. To remedy this, some natural tweaks are applied to AWP. The authors then focus on using AWP to train graph neural networks, demonstrating a minor boost in both clean and robust accuracy. \n", "review_text": "**Strengths:**\nThe line of work concerning adversarial weight perturbations is interesting and significant, as it is one of a few locations where the nascent theory of deep networks can provide easy tweaks to training that improve generalization. Identifying a problem with the current approach (even on MLPs) is valuable and significant, and ample evidence is presented to suggest the vanishing gradient problem is actually what is occurring. The proposed solutions are simple and easy to implement. The experimentation is thorough: while the improvements provided by AWP are very minor (less than a percent in some instances), it is yet another easy and cheap trick to ever-so-slightly boost the performance of a network. \n\n**Weaknesses:**\nFirst, while it is an important insight to notice the vanishing gradient, the proposed remedies (weight truncation and weighted AWP) are natural and not particularly novel. These are likely the first things that one would try to mitigate the observed problems with AWP and don't represent a great stride forward in the field. Similarly, it is not particularly surprising that the benefits of AWP as evidenced in MLPs carry over to GNNs: not enough motivation or evidence is presented to make this seem surprising or unexpected. Third, Theorem 1 is trivial. If the definition of Ltrain is to be the one that only performs a single first-order step, then of course the gradient evaluated at a minimum is going to be zero: the interesting question here is how the true AWP loss (eq2) relates to the standard training loss. This theorem doesn't add add anything to the story. \n\n**Questions:**\n- One of my complaints is about the novelty of contribution re: GNNs vs MLPs. Did I miss this, or is there simply a much more pronounced effect of the gradient-vanishing phenomenon in GNNS than MLPs? \n- How much do the gradient norms (with respect to the weights) actually change when training under AWP vs WT-AWP vs standard training?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors extend the line of work related to adversarial weight perturbations: first showing that a vanishing gradient issue in standard AWP can hinder training. To remedy this, some natural tweaks are applied to AWP. The authors then focus on using AWP to train graph neural networks, demonstrating a minor boost in both clean and robust accuracy. \n", "main_review": "**Strengths:**\nThe line of work concerning adversarial weight perturbations is interesting and significant, as it is one of a few locations where the nascent theory of deep networks can provide easy tweaks to training that improve generalization. Identifying a problem with the current approach (even on MLPs) is valuable and significant, and ample evidence is presented to suggest the vanishing gradient problem is actually what is occurring. The proposed solutions are simple and easy to implement. The experimentation is thorough: while the improvements provided by AWP are very minor (less than a percent in some instances), it is yet another easy and cheap trick to ever-so-slightly boost the performance of a network. \n\n**Weaknesses:**\nFirst, while it is an important insight to notice the vanishing gradient, the proposed remedies (weight truncation and weighted AWP) are natural and not particularly novel. These are likely the first things that one would try to mitigate the observed problems with AWP and don't represent a great stride forward in the field. Similarly, it is not particularly surprising that the benefits of AWP as evidenced in MLPs carry over to GNNs: not enough motivation or evidence is presented to make this seem surprising or unexpected. Third, Theorem 1 is trivial. If the definition of Ltrain is to be the one that only performs a single first-order step, then of course the gradient evaluated at a minimum is going to be zero: the interesting question here is how the true AWP loss (eq2) relates to the standard training loss. This theorem doesn't add add anything to the story. \n\n**Questions:**\n- One of my complaints is about the novelty of contribution re: GNNs vs MLPs. Did I miss this, or is there simply a much more pronounced effect of the gradient-vanishing phenomenon in GNNS than MLPs? \n- How much do the gradient norms (with respect to the weights) actually change when training under AWP vs WT-AWP vs standard training?", "summary_of_the_review": "This is an interesting line of work, and the pointing out of (and subsequent fixing of) the gradient-vanishing phenomenon in AWP is a valuable contribution. Past this, none of the results are strikingly novel or groundbreaking. I think this is a borderline paper, tending towards rejection, but could be convinced to boost my score slightly if I've misunderstood something. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635444781367}, {"id": "-YzVlROXSg_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1934/Reviewer_gQzb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work studies how AWP improves the GNN, mainly at the generalization aspect. The authors first derive several theoretical results that can be easily derived from the existing works on CNN, e.g., Wu et al., 2020a and Foret et a;., 2021 in section 3.. In section 4, the authors show that if they directly apply large $\\rho$ as suggested by (3), they will encounter the gradient vanishing phenomenon during the training. Based on this phenomenon, the authors propose two ways to solve this. One is truncated AWP, the other is weighted AWP. Then, the authors conduct simple experiments using Vanilla GCN, GCN+AWP, GCN+TAWP, GCN+WAWP, where Vanilla perform the best whereas GCN+TAWP, GCN+WAWP have smoother decision boundary. And they decide to combine T+W with AWP (to be frank, I do not know why they do not add experiment result of WT-AWP here.) In section 5, the authors conduct a lot of experiments like clean accuracy, robust accuracy, etc, which follow the existing works.\n\nThe contribution as I can see is that the authors are the first ones to conduct extensive experiments using AWP on graph dataset. However, I do not see other contributions listed by the authors, which I will explain later in the main review.", "review_text": "Strengths: I am satisfied with the numerical experiments in section 5, although they largely follows the routes of existing work. \n\nWeaknesses: I will point the weaknesses by sections.\n\nSection 3: i) The theoretical results are very incremental based on existing works e.g., Wu et al., 2020a, which can be easily derived or directly used. ii) the AWP algorithm listed here is the same as Wu et al., 2020a (only add a letter A).\n\nSection 4: i) After I read the paper thoroughly twice, I do not see how you assign $\\theta^{(awp)}$ and $\\theta^{(normal)}$, which is very important. If you have that in your paper, first let me know where I can find it, second move that to section 4 and explain it. It is unacceptable not having it in section 4. ii) Do not have theory or even intuition about T-AWP and W-AWP. In my perspective, how you justify your algorithms is very important, which is much important than the so called \"theoretical\" results in Section 3. You can even save the space in section 3 for you to defend your algorithms in section 4. I believe what I point out here should be your novelty for this paper. At least, please, add some good intuitions for your algorithms (I will add points if you do that.) iii) Please add the WT-AWG result in Figure 3.\n\nSummary: Good: experiments; Weak: i) too incremental and lack of novelty; ii) do not provides theoretical/intuitive explanation of the proposed algorithms.\n\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work studies how AWP improves the GNN, mainly at the generalization aspect. The authors first derive several theoretical results that can be easily derived from the existing works on CNN, e.g., Wu et al., 2020a and Foret et a;., 2021 in section 3.. In section 4, the authors show that if they directly apply large $\\rho$ as suggested by (3), they will encounter the gradient vanishing phenomenon during the training. Based on this phenomenon, the authors propose two ways to solve this. One is truncated AWP, the other is weighted AWP. Then, the authors conduct simple experiments using Vanilla GCN, GCN+AWP, GCN+TAWP, GCN+WAWP, where Vanilla perform the best whereas GCN+TAWP, GCN+WAWP have smoother decision boundary. And they decide to combine T+W with AWP (to be frank, I do not know why they do not add experiment result of WT-AWP here.) In section 5, the authors conduct a lot of experiments like clean accuracy, robust accuracy, etc, which follow the existing works.\n\nThe contribution as I can see is that the authors are the first ones to conduct extensive experiments using AWP on graph dataset. However, I do not see other contributions listed by the authors, which I will explain later in the main review.", "main_review": "Strengths: I am satisfied with the numerical experiments in section 5, although they largely follows the routes of existing work. \n\nWeaknesses: I will point the weaknesses by sections.\n\nSection 3: i) The theoretical results are very incremental based on existing works e.g., Wu et al., 2020a, which can be easily derived or directly used. ii) the AWP algorithm listed here is the same as Wu et al., 2020a (only add a letter A).\n\nSection 4: i) After I read the paper thoroughly twice, I do not see how you assign $\\theta^{(awp)}$ and $\\theta^{(normal)}$, which is very important. If you have that in your paper, first let me know where I can find it, second move that to section 4 and explain it. It is unacceptable not having it in section 4. ii) Do not have theory or even intuition about T-AWP and W-AWP. In my perspective, how you justify your algorithms is very important, which is much important than the so called \"theoretical\" results in Section 3. You can even save the space in section 3 for you to defend your algorithms in section 4. I believe what I point out here should be your novelty for this paper. At least, please, add some good intuitions for your algorithms (I will add points if you do that.) iii) Please add the WT-AWG result in Figure 3.\n\nSummary: Good: experiments; Weak: i) too incremental and lack of novelty; ii) do not provides theoretical/intuitive explanation of the proposed algorithms.\n\n\n\n\n\n", "summary_of_the_review": "This is work i) is quite incremental compared to the existing works, ii) does not explain/justify/describe the algorithms well. I do not think this paper is good enough for ICLR.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635200402910}], "openreview_url": "https://openreview.net/forum?id=hUr6K4D9f7P", "arxiv_id": "2212.04983", "paper_pdf": "papers/hUr6K4D9f7P.pdf", "paper_pdf_sha256": "c541437eda108e345a55204f7db67e35e187f3ef76edbf29fdc940ce104139f3", "paper_pdf_bytes": 1108766, "paper_pdf_source": "openreview", "code_url": "https://github.com/yihwu/WT-AWP", "code_repository": "yihwu/WT-AWP", "code_commit": "aa0dae0c64521d885fed6dce32c5a35a3a292f3e", "code_archive": "repos/hUr6K4D9f7P.zip", "code_archive_sha256": "aa802b68a1dfa8e2defeb045525eb97e555f6f6f3e6769c0a98bfec8c0f01402", "code_archive_bytes": 2581455, "code_file_count": 59, "code_extensions": {".py": 56, ".ipynb": 3}, "github_disk_usage_kb": 2349, "github_languages": {"Python": 2912518, "Jupyter Notebook": 341294}, "github_archived": false, "github_pushed_at": "2022-11-22T07:28:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-weight-perturbation-improves-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zbEupOtJFF", "year": 2021, "status": "rejected", "title": "On interaction between augmentations and corruptions in natural corruption robustness", "authors": ["Eric Mintun", "Alexander Kirillov", "Saining Xie"], "authorids": ["~Eric_Mintun1", "~Alexander_Kirillov1", "~Saining_Xie2"], "authors_source": "OpenReview API", "abstract": "Invariance to a broad array of image corruptions, such as warping, noise, or color shifts, is an important aspect of building robust models in computer vision. Recently, several new data augmentations have been proposed that significantly improve performance on ImageNet-C, a benchmark of such corruptions. However, there is still a lack of basic understanding on the relationship between data augmentations and test-time corruptions. To this end, we develop a feature space for image transforms, and then use a new measure in this space between augmentations and corruptions called the Minimal Sample Distance to demonstrate there is a strong correlation between similarity and performance. We then investigate recent data augmentations and observe a significant degradation in corruption robustness when the test-time corruptions are sampled to be perceptually dissimilar from ImageNet-C in this feature space. Our results suggest that test error can be improved by training on perceptually similar augmentations, and data augmentations may risk overfitting to the existing benchmark.  We hope our results and tools will allow for more robust progress towards improving robustness to image corruptions.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "qIGKxk2Q14W", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper704/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces a metric to quantify the perceptual similarity between different kind of corruptions and uses it to show that training on a corruptions induces robustness to other corruptions which are perceptually similar. Analyzing the current data augmentation based methods for robustness using this lens, the authors hypothesis that we are currently overfitting to the existing robustness benchmark (ImageNet-C) and proposes a new benchmark with a new set of corruptions. \n\nI think the author's observations that training on some corruptions helps the network to be robust to similar corruptions in test time is quite intuitive. I appreciate that the paper tries to quantify this and performs an extensive empirical study. The insight from this study and the new proposed benchmark will be useful to further advance the research in this area. \n\nI have a some questions/clarifications:\n1. What are the augmentations that were used while training the network used to extract features for the metric? Does the metric will probably depend on the architecture of the network used as well. Have you verified that the metric is robust to different architectures and the same conclusion holds? I am not sure if this is in the supplementary material somewhere and I missed it. \n2. Do we find any non-intuitive pairs of corruptions which are similar? It occurs to me that most geometric based corruptions are similar and noise based corruptions are similar etc, but is there any pair of corruptions across these groups that the metric identifies as similar?\n\n\nUpdate after rebuttal: I appreciate the author response. I will maintain my original score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Review", "review": "The paper introduces a metric to quantify the perceptual similarity between different kind of corruptions and uses it to show that training on a corruptions induces robustness to other corruptions which are perceptually similar. Analyzing the current data augmentation based methods for robustness using this lens, the authors hypothesis that we are currently overfitting to the existing robustness benchmark (ImageNet-C) and proposes a new benchmark with a new set of corruptions. \n\nI think the author's observations that training on some corruptions helps the network to be robust to similar corruptions in test time is quite intuitive. I appreciate that the paper tries to quantify this and performs an extensive empirical study. The insight from this study and the new proposed benchmark will be useful to further advance the research in this area. \n\nI have a some questions/clarifications:\n1. What are the augmentations that were used while training the network used to extract features for the metric? Does the metric will probably depend on the architecture of the network used as well. Have you verified that the metric is robust to different architectures and the same conclusion holds? I am not sure if this is in the supplementary material somewhere and I missed it. \n2. Do we find any non-intuitive pairs of corruptions which are similar? It occurs to me that most geometric based corruptions are similar and noise based corruptions are similar etc, but is there any pair of corruptions across these groups that the metric identifies as similar?\n\n\nUpdate after rebuttal: I appreciate the author response. I will maintain my original score.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604027615623}, {"id": "GECfZ-aPmhQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper704/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper introduces the Minimal Sample Distance (MSD): a measure of the minimal distance, in a trained network representation space, between samples modified with an augmentation and the average of all samples modified by a corruption. It uses this metric to claim that there exists a high correlation between the corruption error and the MSD of a given augmentation. This way, it claims that focusing on benchmarks like ImageNet-C may lead to overfitting to the corruptions present in that benchmark. \n\nOne problem is that this correlation isn’t true for all augmentations. Only 4 are highlighted in the main text. As the paper describes, a few corruptions have spearman coefficient of less than 0.5. Particularly notable is brightness which, despite having very low spearman coefficient, is used to show that Patch Gaussian has “overfit” to the noise corruptions in the ImageNet-C benchmark. This is especially worrying since in their original paper, Patch Gaussian shows improvement in non-noise as well, which couldn’t have come from overfitting. Additionally, why was AutoAugment, but not RandAugment tested?\n\nAnother concern is that it may not make sense to compare single augmentations (such as Patch Gaussian) with augmentation policies, such as AutoAugment, RandAugment, and AugMix. It’s possible that individual augmentations in these policies “overfit” to the corruptions as well, but that this isn’t shown in MSD due to the use of many corruptions. In which case, using PatchGaussian in the RandAugment search space (for instance) would resolve any issues. \n\nThe paper argues that AutoAugment \"overfits\" while AugMix doesn't, but that's only because AugMix explicitly removed any augmentations in ImageNet-C from its search space, so it would make sense to repeat this experiment with the augmentations present in AugMix in order to confirm that it doesn't indeed \"overfit\".\n\nThe paper then suggests that one solution would be to use MSD to sample dissimilar corruptions to test on. However, given that it seems like there’s no evidence of overfitting for augmentation policies that encourage a diversity of augmentations, such as AugMix (which is in line with Yin et al 2019). Then I’m not sure what the benefit of expanding the robustness benchmark is. If current methods have indeed overfit, why won’t we also overfit to the new benchmark’s corruptions? \n\nOverall, the paper presents interesting results and discussion. \n\nUpdate after rebuttal: I appreciate the authors' response and clarifications. I maintain my original score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting analysis, but some strange comparisons", "review": "The paper introduces the Minimal Sample Distance (MSD): a measure of the minimal distance, in a trained network representation space, between samples modified with an augmentation and the average of all samples modified by a corruption. It uses this metric to claim that there exists a high correlation between the corruption error and the MSD of a given augmentation. This way, it claims that focusing on benchmarks like ImageNet-C may lead to overfitting to the corruptions present in that benchmark. \n\nOne problem is that this correlation isn’t true for all augmentations. Only 4 are highlighted in the main text. As the paper describes, a few corruptions have spearman coefficient of less than 0.5. Particularly notable is brightness which, despite having very low spearman coefficient, is used to show that Patch Gaussian has “overfit” to the noise corruptions in the ImageNet-C benchmark. This is especially worrying since in their original paper, Patch Gaussian shows improvement in non-noise as well, which couldn’t have come from overfitting. Additionally, why was AutoAugment, but not RandAugment tested?\n\nAnother concern is that it may not make sense to compare single augmentations (such as Patch Gaussian) with augmentation policies, such as AutoAugment, RandAugment, and AugMix. It’s possible that individual augmentations in these policies “overfit” to the corruptions as well, but that this isn’t shown in MSD due to the use of many corruptions. In which case, using PatchGaussian in the RandAugment search space (for instance) would resolve any issues. \n\nThe paper argues that AutoAugment \"overfits\" while AugMix doesn't, but that's only because AugMix explicitly removed any augmentations in ImageNet-C from its search space, so it would make sense to repeat this experiment with the augmentations present in AugMix in order to confirm that it doesn't indeed \"overfit\".\n\nThe paper then suggests that one solution would be to use MSD to sample dissimilar corruptions to test on. However, given that it seems like there’s no evidence of overfitting for augmentation policies that encourage a diversity of augmentations, such as AugMix (which is in line with Yin et al 2019). Then I’m not sure what the benefit of expanding the robustness benchmark is. If current methods have indeed overfit, why won’t we also overfit to the new benchmark’s corruptions? \n\nOverall, the paper presents interesting results and discussion. \n\nUpdate after rebuttal: I appreciate the authors' response and clarifications. I maintain my original score.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603920497507}, {"id": "2OMZK3oJOhh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper704/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary \nThe paper studies the importance of similarity between augmentations and corruptions for improving performance on those corruptions. To measure the distance between the augmentation and corruption distributions, the paper proposes a new metric, Minimal Sample Distance (MSD), which is the perceptual similarity between an average corruption and the closest augmentation from a finite set of samples sampled from the augmented data distribution. It is shown that MSD overcomes the drawbacks of distributional distance measures like Maximum Mean Discrepancy (MMD). A new benchmark, called ImageNet-C-bar, made up of corruptions that are perceptually dissimilar to ImageNet-C, is introduced. Using standard evaluation, it is empirically shown that several recent augmentation schemes show degraded performance on the new dataset, suggesting that they generate augmentations only perceptually similar to ImageNet-C and thus are prone to overfitting. \n\n+ves\n+ Although the notion of the relation between augmentations and test-time corruptions has somewhat already been empirically observed and stated in many previous works, the paper tries to correlate this relation statistically. To my knowledge, this is the first such work.\n+ Through comprehensive evaluations, the paper shows that the proposition of computing MSD rather than MMD correlates well with the relation between augmentations and corruptions observed in reality.\n+ A new benchmark, ImageNet-C-bar is proposed, which shows a useful result to the community that recent SOTA augmentation methods have degraded performance on the new dataset because they generate augmentations close to ImageNet-C corruptions. \n\nConcerns\n- The paper says (pg 2) - “we empirically show an intuitive yet surprisingly overlooked finding: Augmentation-corruption perceptual similarity is a strong predictor of corruption error”. However, this notion of the relation between augmentations and test-time corruptions does not seem very surprising. It’s perhaps well-known that DNNs will generalize well only when test distributions are fairly similar to training distributions. Hence, the importance of this observation does seem to be overemphasized, although this is certainly of use.  \n\n- It's been recently shown that removing texture biases by introducing stylized transformations also improves robustness to common-corruptions (Geirhos et al, ICLR 2019). However, stylized transformations don’t look close to any Imagenet-C corruptions. If MSD of stylized transforms is large (which intuitively seems so), then it will mean that MSD is not a reliable metric in such a case. Was this studied? Would MSD be a reliable metric even in such cases? It would be good to understand this, to get a more well-rounded picture of the proposed metric. \n\n- Paper introduces MSD as a distance metric. However, distance metrics should be symmetric in nature. It is not quite evident from just Eq-1 if MSD is symmetric. \n\n- In addition, adding high severity corruption as training augmentation leads to better performance on the same low severity corruption but vice-versa is not true in general. This suggests that measures should perhaps ideally be asymmetric. Clarifying the notion of “distance metric” in this work may be important to make the work mathematically correct.\n\n- Standard choices for measuring perceptual distances are VGG-16 or 19 networks pre-trained on ImageNet. However, the paper chooses to use WRN-40-10 trained on CIFAR-10. This seems to deviate from standard settings. The paper should explain the rationale behind their choice. It will be great to show an ablation on how this choice of feature extractor (VGG-19, Robust VGG etc) affects the MSD. Ideally, the metric should be robust i.e. shouldn’t be sensitive to the choice of feature extractor.\n\n- The practical utility of ImageNet-C-bar seems limited and unclear. The only use that I can think of is using it to identify overfitting on ImageNet-C. I would be happy to understand what I am missing here.\n\nPOST-REBUTTAL:\n\nI thank the authors for the response and the revisions to the paper. I appreciate the authors' efforts towards the rebuttal. I am however left with some concerns which did not have a convincing resolution:\n\n* Regarding the comment on how the proposed analysis would look for stylized transforms, the authors say in the response that \"...we don’t expect that perceptual similarity is the only cause of improved corruption error, only that perceptual similarity is particularly salient for predicting generalization to dissimilar corruptions...\". The work seems to be one-sided in this regard. If stylized transforms don't look perceptually similar to ImageNet-C corruptions but provide robustness, this counters the proposed hypothesis. It then becomes important to say where the proposed analysis is meaningful and where it is not. This seems to be lacking at this time in the work.\n\n* Regarding the robustness of MSD to the choice of feature extractor (as also asked by R1), the revised paper includes results on VGG as feature extractor (thanks to the authors for this), but uses a model that is finetuned for CIFAR-10. In general, perceptual similarity is studied directly taking VGG pre-trained on ImageNet - without finetuning on the target dataset. This leaves this question open, and makes one wonder if the latter features did not support the analysis.\n\n* The utility of Imagenet-C-bar as an additional benchmark to check the goodness of performance on Imagenet-C seems a bit convoluted. Would we need a Imagenet-C-bar-bar to check the goodness for corruptions that may be beyond perceptual similarity (such as stylized transforms)? This is not very convincing.\n\nOverall, I am still on the fence on this work (and retain my original decision at this time). I think the paper does present an interesting insight, but I am not very convinced it has been studied and explored comprehensively enough. I would have ideally preferred to give a borderline (neither positive nor negative) decision, and will not be disappointed if the work is accepted, considering the interesting insights it offers. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review of AnonReviewer2", "review": "Summary \nThe paper studies the importance of similarity between augmentations and corruptions for improving performance on those corruptions. To measure the distance between the augmentation and corruption distributions, the paper proposes a new metric, Minimal Sample Distance (MSD), which is the perceptual similarity between an average corruption and the closest augmentation from a finite set of samples sampled from the augmented data distribution. It is shown that MSD overcomes the drawbacks of distributional distance measures like Maximum Mean Discrepancy (MMD). A new benchmark, called ImageNet-C-bar, made up of corruptions that are perceptually dissimilar to ImageNet-C, is introduced. Using standard evaluation, it is empirically shown that several recent augmentation schemes show degraded performance on the new dataset, suggesting that they generate augmentations only perceptually similar to ImageNet-C and thus are prone to overfitting. \n\n+ves\n+ Although the notion of the relation between augmentations and test-time corruptions has somewhat already been empirically observed and stated in many previous works, the paper tries to correlate this relation statistically. To my knowledge, this is the first such work.\n+ Through comprehensive evaluations, the paper shows that the proposition of computing MSD rather than MMD correlates well with the relation between augmentations and corruptions observed in reality.\n+ A new benchmark, ImageNet-C-bar is proposed, which shows a useful result to the community that recent SOTA augmentation methods have degraded performance on the new dataset because they generate augmentations close to ImageNet-C corruptions. \n\nConcerns\n- The paper says (pg 2) - “we empirically show an intuitive yet surprisingly overlooked finding: Augmentation-corruption perceptual similarity is a strong predictor of corruption error”. However, this notion of the relation between augmentations and test-time corruptions does not seem very surprising. It’s perhaps well-known that DNNs will generalize well only when test distributions are fairly similar to training distributions. Hence, the importance of this observation does seem to be overemphasized, although this is certainly of use.  \n\n- It's been recently shown that removing texture biases by introducing stylized transformations also improves robustness to common-corruptions (Geirhos et al, ICLR 2019). However, stylized transformations don’t look close to any Imagenet-C corruptions. If MSD of stylized transforms is large (which intuitively seems so), then it will mean that MSD is not a reliable metric in such a case. Was this studied? Would MSD be a reliable metric even in such cases? It would be good to understand this, to get a more well-rounded picture of the proposed metric. \n\n- Paper introduces MSD as a distance metric. However, distance metrics should be symmetric in nature. It is not quite evident from just Eq-1 if MSD is symmetric. \n\n- In addition, adding high severity corruption as training augmentation leads to better performance on the same low severity corruption but vice-versa is not true in general. This suggests that measures should perhaps ideally be asymmetric. Clarifying the notion of “distance metric” in this work may be important to make the work mathematically correct.\n\n- Standard choices for measuring perceptual distances are VGG-16 or 19 networks pre-trained on ImageNet. However, the paper chooses to use WRN-40-10 trained on CIFAR-10. This seems to deviate from standard settings. The paper should explain the rationale behind their choice. It will be great to show an ablation on how this choice of feature extractor (VGG-19, Robust VGG etc) affects the MSD. Ideally, the metric should be robust i.e. shouldn’t be sensitive to the choice of feature extractor.\n\n- The practical utility of ImageNet-C-bar seems limited and unclear. The only use that I can think of is using it to identify overfitting on ImageNet-C. I would be happy to understand what I am missing here.\n\nPOST-REBUTTAL:\n\nI thank the authors for the response and the revisions to the paper. I appreciate the authors' efforts towards the rebuttal. I am however left with some concerns which did not have a convincing resolution:\n\n* Regarding the comment on how the proposed analysis would look for stylized transforms, the authors say in the response that \"...we don’t expect that perceptual similarity is the only cause of improved corruption error, only that perceptual similarity is particularly salient for predicting generalization to dissimilar corruptions...\". The work seems to be one-sided in this regard. If stylized transforms don't look perceptually similar to ImageNet-C corruptions but provide robustness, this counters the proposed hypothesis. It then becomes important to say where the proposed analysis is meaningful and where it is not. This seems to be lacking at this time in the work.\n\n* Regarding the robustness of MSD to the choice of feature extractor (as also asked by R1), the revised paper includes results on VGG as feature extractor (thanks to the authors for this), but uses a model that is finetuned for CIFAR-10. In general, perceptual similarity is studied directly taking VGG pre-trained on ImageNet - without finetuning on the target dataset. This leaves this question open, and makes one wonder if the latter features did not support the analysis.\n\n* The utility of Imagenet-C-bar as an additional benchmark to check the goodness of performance on Imagenet-C seems a bit convoluted. Would we need a Imagenet-C-bar-bar to check the goodness for corruptions that may be beyond perceptual similarity (such as stylized transforms)? This is not very convincing.\n\nOverall, I am still on the fence on this work (and retain my original decision at this time). I think the paper does present an interesting insight, but I am not very convinced it has been studied and explored comprehensively enough. I would have ideally preferred to give a borderline (neither positive nor negative) decision, and will not be disappointed if the work is accepted, considering the interesting insights it offers. ", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603899059217}, {"id": "u926VQgaoci", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper704/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes ImageNet-\\bar{C} which uses a smaller number of carefully chosen corruptions, compared to ImageNet-C. The authors try to argue that previous work is overfitting to ImageNet-C. They claim \"overfitting indeed occurs.\" Additionally, they propose \"Minimum Sample Distance,\" showing that they can predict generalization performance using feature embedding distances.\n\nI don't think they marshaled substantial, far-reaching evidence that \"overfitting indeed occurs.\" They rendered numerous corruptions from Huxtable, 2006; Gladman, 2016 (which I very much appreciate) and adversarially chose the worst corruptions. As a result, the best models performed worse by a few percent compared to ImageNet-C.\n_A drop is inevitable and expected given the adversarial selection. If the drop was consistently more than, say, 20%, then there might be strong evidence of overfitting._ Since the degradation for some techniques is small, this strikes me as evidence that by-and-large the community isn't overfitting. (Note it was evident to all that Patch Gaussian and ANT were specialized to noise corruptions.) \"Is Robustness Robust?\" It seems like the answer is \"yes\" but the authors are trying to argue that the answer is \"no.\"\n\nThis paper is fairly similar to _Increasing the Coverage and Balance of Robustness Benchmarks by Using Non-Overlapping Corruptions_ (submission #1101). If these papers have very different ratings, then there's a problem with this review process.\n\n\"AugMix, which introduces a broad array of both geometric and color augmentations\"\nNo it doesn't. It removes several color augmentations from AutoAugment and doesn't introduce any augmentation primitive beyond elementwise convex combinations.\n\n\"On other corruptions, the increase in error is even worse than the mean would suggest, and even broad augmentations like AugMix....\"\nAutoAugment is more broad; AugMix is narrower.\n\n\"Second, AutoAugment is the only tested augmentation scheme that was designed before the release of ImageNet-C\"\nAugMix uses a proper subset of AutoAugment's augmentations. It's not as though AugMix added in distortions to fit ImageNet-C; it removes augmentations because AA fits some of ImageNet-C's corruptions directly. This observation makes their overfitting case hard to maintain. \n\nAlso, if AutoAugment's full augmentation list is fair game for ImageNet-\\bar{C}, then the authors should train AugMix with the full set of augmentations and make that comparison; the generalization discrepancy would likely be even smaller were the augmentation sets made equal.\n\nFrom the \"Corruption robustness as a secondary learning task\" section:\n\"To mitigate overfitting, standard machine learning practice would dictate a training/validation/test set split; it is only the size and breadth of modern vision datasets that has allowed this to be neglected in certain cases recently.\"\n\"having a held-out test set that is not used during model development seems necessary.\"\n\"ImageNet-C has only 15 corruption types\"\n\"ImageNet-C could serve as a validation set and ImageNet-\\bar{C} could serve as a held-out test set\"\nEssentially, since the community lacks a validation set, ImageNet-C should become the validation set, and ImageNet-\\bar{C} should become the test set.\nThis section might be negligent or worse for not acknowledging the already existent ImageNet-C validation set. ImageNet-C provides a validation set with about half the corruptions of ImageNet-\\bar{C}. There are 19 available ImageNet-C corruptions, so the community already has a validation set.\n\nThe authors should cite or compare to other works that use feature distances to predict generalization. An example is \"The Frechet Distance of training and test distribution predicts the generalization gap.\"\n\nUpdate: \"Noisy Student and Assemble-ResNet use without removing overlapping augmentations, yet they test on ImageNet-C.\" This is a bad practice and I should hope the authors of this paper only have experiments where there is a train-test mismatch (otherwise we're not testing robustness to distribution shift).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "\"Is Robustness Robust?\" It would appear so.", "review": "This paper proposes ImageNet-\\bar{C} which uses a smaller number of carefully chosen corruptions, compared to ImageNet-C. The authors try to argue that previous work is overfitting to ImageNet-C. They claim \"overfitting indeed occurs.\" Additionally, they propose \"Minimum Sample Distance,\" showing that they can predict generalization performance using feature embedding distances.\n\nI don't think they marshaled substantial, far-reaching evidence that \"overfitting indeed occurs.\" They rendered numerous corruptions from Huxtable, 2006; Gladman, 2016 (which I very much appreciate) and adversarially chose the worst corruptions. As a result, the best models performed worse by a few percent compared to ImageNet-C.\n_A drop is inevitable and expected given the adversarial selection. If the drop was consistently more than, say, 20%, then there might be strong evidence of overfitting._ Since the degradation for some techniques is small, this strikes me as evidence that by-and-large the community isn't overfitting. (Note it was evident to all that Patch Gaussian and ANT were specialized to noise corruptions.) \"Is Robustness Robust?\" It seems like the answer is \"yes\" but the authors are trying to argue that the answer is \"no.\"\n\nThis paper is fairly similar to _Increasing the Coverage and Balance of Robustness Benchmarks by Using Non-Overlapping Corruptions_ (submission #1101). If these papers have very different ratings, then there's a problem with this review process.\n\n\"AugMix, which introduces a broad array of both geometric and color augmentations\"\nNo it doesn't. It removes several color augmentations from AutoAugment and doesn't introduce any augmentation primitive beyond elementwise convex combinations.\n\n\"On other corruptions, the increase in error is even worse than the mean would suggest, and even broad augmentations like AugMix....\"\nAutoAugment is more broad; AugMix is narrower.\n\n\"Second, AutoAugment is the only tested augmentation scheme that was designed before the release of ImageNet-C\"\nAugMix uses a proper subset of AutoAugment's augmentations. It's not as though AugMix added in distortions to fit ImageNet-C; it removes augmentations because AA fits some of ImageNet-C's corruptions directly. This observation makes their overfitting case hard to maintain. \n\nAlso, if AutoAugment's full augmentation list is fair game for ImageNet-\\bar{C}, then the authors should train AugMix with the full set of augmentations and make that comparison; the generalization discrepancy would likely be even smaller were the augmentation sets made equal.\n\nFrom the \"Corruption robustness as a secondary learning task\" section:\n\"To mitigate overfitting, standard machine learning practice would dictate a training/validation/test set split; it is only the size and breadth of modern vision datasets that has allowed this to be neglected in certain cases recently.\"\n\"having a held-out test set that is not used during model development seems necessary.\"\n\"ImageNet-C has only 15 corruption types\"\n\"ImageNet-C could serve as a validation set and ImageNet-\\bar{C} could serve as a held-out test set\"\nEssentially, since the community lacks a validation set, ImageNet-C should become the validation set, and ImageNet-\\bar{C} should become the test set.\nThis section might be negligent or worse for not acknowledging the already existent ImageNet-C validation set. ImageNet-C provides a validation set with about half the corruptions of ImageNet-\\bar{C}. There are 19 available ImageNet-C corruptions, so the community already has a validation set.\n\nThe authors should cite or compare to other works that use feature distances to predict generalization. An example is \"The Frechet Distance of training and test distribution predicts the generalization gap.\"\n\nUpdate: \"Noisy Student and Assemble-ResNet use without removing overlapping augmentations, yet they test on ImageNet-C.\" This is a bad practice and I should hope the authors of this paper only have experiments where there is a train-test mismatch (otherwise we're not testing robustness to distribution shift).", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603861308912}], "openreview_url": "https://openreview.net/forum?id=zbEupOtJFF", "arxiv_id": "2102.11273", "paper_pdf": "papers/zbEupOtJFF.pdf", "paper_pdf_sha256": "c9a3fe7b9361c8a05078ba00cc2271f420f6adb682c789ef46803f3865ee4a6e", "paper_pdf_bytes": 6672450, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/augmentation-corruption", "code_repository": "facebookresearch/augmentation-corruption", "code_commit": "ba4d5a5e9132fe98fcd4be1ac90abedbbb188794", "code_archive": "repos/zbEupOtJFF.zip", "code_archive_sha256": "f536df67ed4b1ed909f05970784c4d9b43a83f28810bb2e7e8b032a72ca7092d", "code_archive_bytes": 1586667, "code_file_count": 74, "code_extensions": {".py": 55, ".sh": 18, ".ipynb": 1}, "github_disk_usage_kb": 1510, "github_languages": {"Python": 347384, "Shell": 48867, "Jupyter Notebook": 17469}, "github_archived": true, "github_pushed_at": "2022-11-06T23:27:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/on-interaction-between-augmentations-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Tj5rHP8yrQ", "year": 2026, "status": "rejected", "title": "CompeteSMoE - Statistically Guaranteed Mixture of Experts Training via Competition", "authors": ["Nam V. Nguyen", "Huy Nguyen", "Quang Pham", "Van Nguyen", "Savitha Ramasamy", "Nhat Ho"], "authorids": ["~Nam_V._Nguyen1", "~Huy_Nguyen5", "~Quang_Pham1", "~Van_Nguyen3", "~Savitha_Ramasamy1", "~Nhat_Ho1"], "authors_source": "OpenReview API", "abstract": "Sparse mixture of experts (SMoE) offers an appealing solution to scale up the model complexity beyond the mean of increasing the network's depth or width. However, we argue that effective SMoE training remains challenging because of the suboptimal routing process where experts that perform computation do not directly contribute to the routing process. In this work, we propose competition, a novel mechanism to route tokens to experts with the highest neural response. Theoretically, we show that the competition mechanism enjoys a better sample efficiency than the traditional softmax routing. Furthermore, we develop CompeteSMoE, a simple yet effective algorithm to train large language models by deploying a router to learn the competition policy, thus enjoying strong performances at a low training overhead. Our extensive empirical evaluations on both the visual instruction tuning and language pre-training tasks demonstrate the efficacy, robustness, and scalability of CompeteSMoE compared to state-of-the-art SMoE strategies. We will publish the implementation upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "2w5N07bqd6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15638/Reviewer_6ffV"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces a novel competition-based training mechanism for SMoE models that achieves faster convergence than traditional routing. Building on this idea, the authors develop CompeteSMoE, a method that leverages competition to enhance SMoE training. To support their claims, the paper provides a theoretical convergence analysis of the proposed approach. Finally, experiments are conducted on popular vision and text benchmarks, complemented by an additional analysis of routing behavior.", "review_text": "This paper introduces a novel competition-based training mechanism for SMoE models that achieves faster convergence than traditional routing. Building on this idea, the authors develop CompeteSMoE, a method that leverages competition to enhance SMoE training. To support their claims, the paper provides a theoretical convergence analysis of the proposed approach. Finally, experiments are conducted on popular vision and text benchmarks, complemented by an additional analysis of routing behavior.", "strengths": "Despite addressing the well-researched subfield of routing in SMoE, the authors successfully identify an original and interesting perspective on this topic. The paper is well-written and easy to follow.  A significant strength is that the authors ground their methodological claims in theoretical analysis. Furthermore, the proposed method is evaluated across both images and text, which strongly supports the paper's conclusions.", "weaknesses": "The claim in the abstract regarding a scalable method is not substantiated in the main text, as the experiments were conducted on a very limited scale. To support the scalability claim, additional experiments involving larger datasets and more diverse conditions would be necessary.", "questions": "- I am not convinced by ECR metric (Section 5.2.2 b). Intuitively, I would like router network to have a possibility to adapt to updated experts till the end of the training process. Why the authors claim that it is actually good, that the rate decaying faster?\n\n- Section 4.2 does not appear to be directly relevant to the main objectives of the study.\n\n- Upcycling was used as a means to bypass the costly pre-training phase. However, I believe that conducting smaller-scale experiments involving pre-training would likely yield more reliable insights.\n\nMinor comment:\n- No units in Table 4 for Train / Infer", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel competition-based training mechanism for SMoE models that achieves faster convergence than traditional routing. Building on this idea, the authors develop CompeteSMoE, a method that leverages competition to enhance SMoE training. To support their claims, the paper provides a theoretical convergence analysis of the proposed approach. Finally, experiments are conducted on popular vision and text benchmarks, complemented by an additional analysis of routing behavior.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Despite addressing the well-researched subfield of routing in SMoE, the authors successfully identify an original and interesting perspective on this topic. The paper is well-written and easy to follow.  A significant strength is that the authors ground their methodological claims in theoretical analysis. Furthermore, the proposed method is evaluated across both images and text, which strongly supports the paper's conclusions.", "weaknesses": "The claim in the abstract regarding a scalable method is not substantiated in the main text, as the experiments were conducted on a very limited scale. To support the scalability claim, additional experiments involving larger datasets and more diverse conditions would be necessary.", "questions": "- I am not convinced by ECR metric (Section 5.2.2 b). Intuitively, I would like router network to have a possibility to adapt to updated experts till the end of the training process. Why the authors claim that it is actually good, that the rate decaying faster?\n\n- Section 4.2 does not appear to be directly relevant to the main objectives of the study.\n\n- Upcycling was used as a means to bypass the costly pre-training phase. However, I believe that conducting smaller-scale experiments involving pre-training would likely yield more reliable insights.\n\nMinor comment:\n- No units in Table 4 for Train / Infer", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761998627392}, {"id": "f296YQntdi", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15638/Reviewer_WUdu"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces CompeteSMoE, a new approach to training the routing module in Sparse Mixture of Experts (SMoE) models. The method builds on a competition mechanism inspired by the winner-take-all principle. Periodically during training, all experts are activated for each token, and routing decisions are made based on the magnitude of their neural responses. These competition-based assignments serve as supervision signals for the router, which is trained to approximate the winner-take-all routing behavior. The authors propose a practical algorithm with scheduled router training and demonstrate promising empirical performance on both visual instruction tuning (VIT) and language pretraining tasks, outperforming several existing SMoE variants.\n\nThe breadth of analysis is satisfactory, though the language pretraining section would benefit from additional experiments that show whether the efficacy of CompeteSMoE holds up to scale, as well as the most popular modern architectural choices (more in Weaknesses  and Questions)", "review_text": "The paper introduces CompeteSMoE, a new approach to training the routing module in Sparse Mixture of Experts (SMoE) models. The method builds on a competition mechanism inspired by the winner-take-all principle. Periodically during training, all experts are activated for each token, and routing decisions are made based on the magnitude of their neural responses. These competition-based assignments serve as supervision signals for the router, which is trained to approximate the winner-take-all routing behavior. The authors propose a practical algorithm with scheduled router training and demonstrate promising empirical performance on both visual instruction tuning (VIT) and language pretraining tasks, outperforming several existing SMoE variants.\n\nThe breadth of analysis is satisfactory, though the language pretraining section would benefit from additional experiments that show whether the efficacy of CompeteSMoE holds up to scale, as well as the most popular modern architectural choices (more in Weaknesses  and Questions)", "strengths": "- The paper has a clear motivation and a relatively straightforward implementation of the proposed idea.  \n- The presentation is clear and easy to follow.  \n- The formulation of periodical winner-takes-all for SMoE routing decision, alongside having the router emulate those deisions appears novel, to the best of my knowledge.  \n- Experiments cover both vision and language domains.  \n- The paper includes thorough ablations analyzing the individual contributions of the competition mechanism and the diversity loss.  \n- The experiment that rotates routing choices to test whether the model benefits from routing decisions or behaves randomly is an interesting addition.  \n- The results are generally positive, though the effect sizes on benchmarks are small.  \n- The analysis of wall-clock time and peak memory during training shows minimal overhead for the proposed method.", "weaknesses": "- The language pretraining setup uses 13B tokens for a 1B-parameter model. According to the Chinchilla rule of thumb (around 20 tokens per parameter), this corresponds to a moderate undertraining regime, while many modern models are heavily overtrained. See e.g. [Qwen3 Technical Report](https://arxiv.org/pdf/2505.09388)\n- The chosen configuration of `topk = 8` out of 24 experts is not the most typical. Recent trends favor sparser settings where only a few experts are active out of a large total. An ablation over different sparsity levels would make the empirical evaluation more complete.  \n- The observed effect sizes on benchmarks are small. The improvements are more consistent in the vision domain, but those experiments rely on upcycled models, which complicates the interpretation relative to the more straightforward from-scratch language pretraining setting.  \n- Limited ablation of hyperparameters: while ad-hoc guidelines for α, β, γ, and Amax are provided, the robustness of results to these choices is not extensively analyzed in the main text.  \n- Minor issues with typos and unclear formulations (e.g., “200” instead of “200K” around line 1545, based on the supplementary code).", "questions": "- To strengthen the paper on the language pretraining front:  \n  1. It would be helpful to investigate how the efficiency of CompeteSMoE changes with training horizon, model size, and Mixture-of-Experts architecture configurations (total number of experts and number of selected experts).  \n  2. The effect size on benchmarks is small, and given the scale of experiments it's unclear how much of the variation is noise. It would be very informative to see evaluation benchmarks, training loss and validation loss trajectories throughout the entire training run, ideally for different configurations from suggestion 1.  \n- Why was the diversity loss applied during pretraining from scratch, since the model was not upcycled?  \n- Could the authors clarify the meaning of the following passage (lines 129–134)? It is somewhat unclear how joint learning of the task loss and competition policy alleviates the sample inefficiency issue or why this makes training feasible on limited hardware.\n- It would be very valuable to understand the impact of frequency of competition on model quality. If memory is an issue, the authors can consider training a smaller model and thus being able to vary the competition probability up.\n- Teaching the router to pick magnitude-maximising experts biases the model toward bigger activations. It would be very interesting to see the evolution of mean magnitude of tokens in the residual stream after each MoE layer, and how it evolves over time, and how all that changes from baseline SMoE to CompeteSMoE. The reason to care about this is that large activations can be difficult to quantize, which comes up in low-precision training.\n- The paper would benefit from an analysis of the evolution of routing scores through training", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces CompeteSMoE, a new approach to training the routing module in Sparse Mixture of Experts (SMoE) models. The method builds on a competition mechanism inspired by the winner-take-all principle. Periodically during training, all experts are activated for each token, and routing decisions are made based on the magnitude of their neural responses. These competition-based assignments serve as supervision signals for the router, which is trained to approximate the winner-take-all routing behavior. The authors propose a practical algorithm with scheduled router training and demonstrate promising empirical performance on both visual instruction tuning (VIT) and language pretraining tasks, outperforming several existing SMoE variants.\n\nThe breadth of analysis is satisfactory, though the language pretraining section would benefit from additional experiments that show whether the efficacy of CompeteSMoE holds up to scale, as well as the most popular modern architectural choices (more in Weaknesses  and Questions)", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The paper has a clear motivation and a relatively straightforward implementation of the proposed idea.  \n- The presentation is clear and easy to follow.  \n- The formulation of periodical winner-takes-all for SMoE routing decision, alongside having the router emulate those deisions appears novel, to the best of my knowledge.  \n- Experiments cover both vision and language domains.  \n- The paper includes thorough ablations analyzing the individual contributions of the competition mechanism and the diversity loss.  \n- The experiment that rotates routing choices to test whether the model benefits from routing decisions or behaves randomly is an interesting addition.  \n- The results are generally positive, though the effect sizes on benchmarks are small.  \n- The analysis of wall-clock time and peak memory during training shows minimal overhead for the proposed method.", "weaknesses": "- The language pretraining setup uses 13B tokens for a 1B-parameter model. According to the Chinchilla rule of thumb (around 20 tokens per parameter), this corresponds to a moderate undertraining regime, while many modern models are heavily overtrained. See e.g. [Qwen3 Technical Report](https://arxiv.org/pdf/2505.09388)\n- The chosen configuration of `topk = 8` out of 24 experts is not the most typical. Recent trends favor sparser settings where only a few experts are active out of a large total. An ablation over different sparsity levels would make the empirical evaluation more complete.  \n- The observed effect sizes on benchmarks are small. The improvements are more consistent in the vision domain, but those experiments rely on upcycled models, which complicates the interpretation relative to the more straightforward from-scratch language pretraining setting.  \n- Limited ablation of hyperparameters: while ad-hoc guidelines for α, β, γ, and Amax are provided, the robustness of results to these choices is not extensively analyzed in the main text.  \n- Minor issues with typos and unclear formulations (e.g., “200” instead of “200K” around line 1545, based on the supplementary code).", "questions": "- To strengthen the paper on the language pretraining front:  \n  1. It would be helpful to investigate how the efficiency of CompeteSMoE changes with training horizon, model size, and Mixture-of-Experts architecture configurations (total number of experts and number of selected experts).  \n  2. The effect size on benchmarks is small, and given the scale of experiments it's unclear how much of the variation is noise. It would be very informative to see evaluation benchmarks, training loss and validation loss trajectories throughout the entire training run, ideally for different configurations from suggestion 1.  \n- Why was the diversity loss applied during pretraining from scratch, since the model was not upcycled?  \n- Could the authors clarify the meaning of the following passage (lines 129–134)? It is somewhat unclear how joint learning of the task loss and competition policy alleviates the sample inefficiency issue or why this makes training feasible on limited hardware.\n- It would be very valuable to understand the impact of frequency of competition on model quality. If memory is an issue, the authors can consider training a smaller model and thus being able to vary the competition probability up.\n- Teaching the router to pick magnitude-maximising experts biases the model toward bigger activations. It would be very interesting to see the evolution of mean magnitude of tokens in the residual stream after each MoE layer, and how it evolves over time, and how all that changes from baseline SMoE to CompeteSMoE. The reason to care about this is that large activations can be difficult to quantize, which comes up in low-precision training.\n- The paper would benefit from an analysis of the evolution of routing scores through training", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761693274740}, {"id": "epZ3VcF9MX", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15638/Reviewer_EjcQ"], "rating": 8, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 2, "summary": "The authors propose CompeteSMoE, a SMoE training algorithm based on a competition mechanism, aimed at addressing the suboptimal issue of separation between expert computation and routing decisions in the routing process of traditional SMoE. Inspired by the Winner-take-all principle, its core idea is to activate all experts and select the Top-K experts based on their neural responses. Theoretically, it is proven that this mechanism has better sample efficiency and convergence rate than traditional softmax routing. Experiments are conducted on a wide range of benchmarks, demonstrating the effectiveness of the scheme.", "review_text": "The authors propose CompeteSMoE, a SMoE training algorithm based on a competition mechanism, aimed at addressing the suboptimal issue of separation between expert computation and routing decisions in the routing process of traditional SMoE. Inspired by the Winner-take-all principle, its core idea is to activate all experts and select the Top-K experts based on their neural responses. Theoretically, it is proven that this mechanism has better sample efficiency and convergence rate than traditional softmax routing. Experiments are conducted on a wide range of benchmarks, demonstrating the effectiveness of the scheme.", "strengths": "1 The benchmarks covered in the experiments are relatively comprehensive, including both vision and language models of different types, providing strong support for the generalizability of the scheme.\n\n2 The authors provide detailed hyperparameter configurations, making the reproducibility of the method convincing.\n\n3 It provides statistical guarantees for the competition mechanism in SMoE. Through the convergence analysis of Gaussian MoE models, the convergence rate of the Total Variation distance for density estimation is derived, and a lower bound is established based on Voronoi loss, filling the gap in the theoretical analysis of the competition mechanism.", "weaknesses": "1 Regarding the hyperparameters ω and A used in this manuscript, what is their hyperparameter sensitivity for LLMs of different sizes?\n\n2 How does the change in the total number of experts and the number of activated experts affect the method proposed in this paper?", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose CompeteSMoE, a SMoE training algorithm based on a competition mechanism, aimed at addressing the suboptimal issue of separation between expert computation and routing decisions in the routing process of traditional SMoE. Inspired by the Winner-take-all principle, its core idea is to activate all experts and select the Top-K experts based on their neural responses. Theoretically, it is proven that this mechanism has better sample efficiency and convergence rate than traditional softmax routing. Experiments are conducted on a wide range of benchmarks, demonstrating the effectiveness of the scheme.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1 The benchmarks covered in the experiments are relatively comprehensive, including both vision and language models of different types, providing strong support for the generalizability of the scheme.\n\n2 The authors provide detailed hyperparameter configurations, making the reproducibility of the method convincing.\n\n3 It provides statistical guarantees for the competition mechanism in SMoE. Through the convergence analysis of Gaussian MoE models, the convergence rate of the Total Variation distance for density estimation is derived, and a lower bound is established based on Voronoi loss, filling the gap in the theoretical analysis of the competition mechanism.", "weaknesses": "1 Regarding the hyperparameters ω and A used in this manuscript, what is their hyperparameter sensitivity for LLMs of different sizes?\n\n2 How does the change in the total number of experts and the number of activated experts affect the method proposed in this paper?", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1760948497300}], "openreview_url": "https://openreview.net/forum?id=Tj5rHP8yrQ", "arxiv_id": "2505.13380", "paper_pdf": "papers/Tj5rHP8yrQ.pdf", "paper_pdf_sha256": "2129ac6aa9e0482889ed7975d6b3b3c41e47f09f7c3b7b1f387fb05193c493d9", "paper_pdf_bytes": 832107, "paper_pdf_source": "openreview", "code_url": "https://github.com/Fsoft-AIC/CompeteSMoE", "code_repository": "Fsoft-AIC/CompeteSMoE", "code_commit": "ab48bb62aa7edb855328e375b9e089e9b922c9d7", "code_archive": "repos/Tj5rHP8yrQ.zip", "code_archive_sha256": "ff551a9b49cb66865030ec64c3f857b3aed36e3733cf604f13e094b3b6983101", "code_archive_bytes": 6254766, "code_file_count": 449, "code_extensions": {".py": 429, ".sh": 11, ".ipynb": 9}, "github_disk_usage_kb": 5686, "github_languages": {"Python": 3123102, "Jupyter Notebook": 151475, "Shell": 11757}, "github_archived": false, "github_pushed_at": "2025-08-23T08:02:14Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/competesmoe-statistically-guaranteed-mixture"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GYk0thSY1M", "year": 2025, "status": "rejected", "title": "Recurrent Context Compression: Efficiently Expanding the Context Window of LLM", "authors": ["ChensenHuang", "Guibo Zhu", "Xuepeng Wang", "Dong Yi", "Yifei Luo", "Haoran Chen", "Guojing Ge", "Jinqiao Wang"], "authorids": ["~ChensenHuang1", "~Guibo_Zhu1", "~Xuepeng_Wang1", "~Dong_Yi2", "~Yifei_Luo2", "~Haoran_Chen1", "~Guojing_Ge2", "~Jinqiao_Wang1"], "authors_source": "OpenReview API", "abstract": "To extend the context length of Transformer-based large language models (LLMs) and improve comprehension capabilities, researchers often encounter constraints stemming from finite computational resources and bounded memory capacities. This work proposes a novel approach, termed Recurrent Context Compression (RCC), designed to efficiently expand the context window length of LLMs. Furthermore, we delve into the prevalent issue of degraded model performance when both instructional prompts and contextual information undergo compression for downstream tasks. To address this challenge, we propose a novel instruction reconstruction methodology aimed at mitigating the detrimental effects of this compression process. The effectiveness of our proposed approach was validated across multiple tasks while achieving an impressive context compression rate of at least 32x. On text reconstruction task, we maintain a BLEU-4 score close to 0.95. On passkey retrieval task, we achieve nearly 100% accuracy involving an extensive sequence length of 1 million tokens. On long-text question-answering task, we obtain comparable performance with the non-compressed LLM in F1 and Rouge scores. Our method also demonstrated competitive performance in long-text question-answering tasks compared to non-compressed methods, while significantly saving storage resources.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NBwjE2lSgi", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2386/Reviewer_Jqs9"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces Recurrent Context Compression (RCC), a technique for long-context language modeling. The input is segmented into smaller chunks, the encoder encodes multiple chunks at a time, and the hidden state of the last token in each segment at a layer is fed into the decoder. On the decoder side, the hidden states from the first layer are prepended to the decoder input tokens and the hidden states from the encoder are connected to the hidden states in the decoder at each layer.\nThe paper also introduces reconstruction tasks to help the decoder model to enhance memory and retrieval ability.\nThe method is evaluated on the task reconstruction task, passkey retrieval, and LongBench.", "review_text": "The paper introduces Recurrent Context Compression (RCC), a technique for long-context language modeling. The input is segmented into smaller chunks, the encoder encodes multiple chunks at a time, and the hidden state of the last token in each segment at a layer is fed into the decoder. On the decoder side, the hidden states from the first layer are prepended to the decoder input tokens and the hidden states from the encoder are connected to the hidden states in the decoder at each layer.\nThe paper also introduces reconstruction tasks to help the decoder model to enhance memory and retrieval ability.\nThe method is evaluated on the task reconstruction task, passkey retrieval, and LongBench.", "strengths": "- The method is well-motivated—building long-context language models is a challenge and useful task. The method is designed to tackle the training and inference challenges and achieves better memory usage compared to some previous work.\n- The paper includes experiments using different settings of RCC, such as different encoder and decoder models, and different ways of using the instructions during training.\n- RCC achieves strong PassKey results at 1M input tokens in supervised settings, which suggests the architecture may be capable of passing such synthetic stress tests.", "weaknesses": "- In RCC, the question and the instruction are first reconstructed from the encoder hidden states. However, this means that these tokens are still inputted to the decoder, and the time and space complexity is worse than simply putting these tokens in the decoder inputs in the first place. \n- The evaluation of downstream applications is limited to only the QA subsets from LongBench, and the comparison is limited to only Pythia with and without fine-tuning (Table 2). Similarly, the passkey experiments (Table 1) are limited in comparisons with previous related long-context language modeling works, such as AutoCompressor (Chevalier et al., 2023), LLoCO (Tan et al., 2024), Unlimiformer (Bertsch et al., 2023), and StreamingLLM (Xiao et al., 2024). Due to the lack of comparisons, it is unclear how RCC compares to or improves upon previous methods.\n- Overall, there are still some issues with the evaluation settings (detailed below in the Questions section), but the work could really benefit from testing RCC on more downstream applications as well as on out-of-domain tasks.", "questions": "- What are the results of RCC on more out-of-domain tasks other than QA tasks? I would be interested in seeing the results of LongBench in other categories like summarization and few-shot learning. \n- How do you evaluate LongBench at 0-2k? Do you only use the test samples where the input length is in that range or do you do truncation? \n- Similarly, what is the performance on the passkey retrieval tasks without specifically fine-tuning on the task? In previous works (Fu et al., 2024, Dubey et al., 2024), the synthetic tasks are often evaluated in a zero-shot setting without doing any fine-tuning as a stress test, it would reveal more about the method if it were only trained on the Pile and test on the synthetic tasks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Recurrent Context Compression (RCC), a technique for long-context language modeling. The input is segmented into smaller chunks, the encoder encodes multiple chunks at a time, and the hidden state of the last token in each segment at a layer is fed into the decoder. On the decoder side, the hidden states from the first layer are prepended to the decoder input tokens and the hidden states from the encoder are connected to the hidden states in the decoder at each layer.\nThe paper also introduces reconstruction tasks to help the decoder model to enhance memory and retrieval ability.\nThe method is evaluated on the task reconstruction task, passkey retrieval, and LongBench.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The method is well-motivated—building long-context language models is a challenge and useful task. The method is designed to tackle the training and inference challenges and achieves better memory usage compared to some previous work.\n- The paper includes experiments using different settings of RCC, such as different encoder and decoder models, and different ways of using the instructions during training.\n- RCC achieves strong PassKey results at 1M input tokens in supervised settings, which suggests the architecture may be capable of passing such synthetic stress tests.", "weaknesses": "- In RCC, the question and the instruction are first reconstructed from the encoder hidden states. However, this means that these tokens are still inputted to the decoder, and the time and space complexity is worse than simply putting these tokens in the decoder inputs in the first place. \n- The evaluation of downstream applications is limited to only the QA subsets from LongBench, and the comparison is limited to only Pythia with and without fine-tuning (Table 2). Similarly, the passkey experiments (Table 1) are limited in comparisons with previous related long-context language modeling works, such as AutoCompressor (Chevalier et al., 2023), LLoCO (Tan et al., 2024), Unlimiformer (Bertsch et al., 2023), and StreamingLLM (Xiao et al., 2024). Due to the lack of comparisons, it is unclear how RCC compares to or improves upon previous methods.\n- Overall, there are still some issues with the evaluation settings (detailed below in the Questions section), but the work could really benefit from testing RCC on more downstream applications as well as on out-of-domain tasks.", "questions": "- What are the results of RCC on more out-of-domain tasks other than QA tasks? I would be interested in seeing the results of LongBench in other categories like summarization and few-shot learning. \n- How do you evaluate LongBench at 0-2k? Do you only use the test samples where the input length is in that range or do you do truncation? \n- Similarly, what is the performance on the passkey retrieval tasks without specifically fine-tuning on the task? In previous works (Fu et al., 2024, Dubey et al., 2024), the synthetic tasks are often evaluated in a zero-shot setting without doing any fine-tuning as a stress test, it would reveal more about the method if it were only trained on the Pile and test on the synthetic tasks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730686317003}, {"id": "cvkywzsG8P", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2386/Reviewer_bbjH"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper studies expanding the context window of the language model by introducing a decoder that chunks the input context into fix size pieces and compressing each. They also consider the effect of prompt and how to compress prompts such that degredation in different forms of prompting is minimized. They evaluate their method on a set of benchmarks and show effectiveness up to 1M length context.", "review_text": "The paper studies expanding the context window of the language model by introducing a decoder that chunks the input context into fix size pieces and compressing each. They also consider the effect of prompt and how to compress prompts such that degredation in different forms of prompting is minimized. They evaluate their method on a set of benchmarks and show effectiveness up to 1M length context.", "strengths": "- The paper targets a very important problem. Effectively increasing the context of language models is an active area of research and has important practical implications\n - The work is original. Although I have some hesitations about the appraoch, introducing a decoder the way has been presented is novel to the best of my knowledge.", "weaknesses": "- The approach proposed basically doubles the number of parameters required for the model. This has both training and inference implications.\n- As shown in Table 2, the method results in 2 point degradation for the context length that is within the pretraining setting. This is significant and further evaluations and studies is needed. One of the main challenges of long context extension is preserving short context benchmarks.", "questions": "- For the decoder part, have the authors considered a model that uses a bi-directional attention mask?\n- Have the authors compared the method to models such as T5 with iso parameters (decoder + encoder)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies expanding the context window of the language model by introducing a decoder that chunks the input context into fix size pieces and compressing each. They also consider the effect of prompt and how to compress prompts such that degredation in different forms of prompting is minimized. They evaluate their method on a set of benchmarks and show effectiveness up to 1M length context.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "- The paper targets a very important problem. Effectively increasing the context of language models is an active area of research and has important practical implications\n - The work is original. Although I have some hesitations about the appraoch, introducing a decoder the way has been presented is novel to the best of my knowledge.", "weaknesses": "- The approach proposed basically doubles the number of parameters required for the model. This has both training and inference implications.\n- As shown in Table 2, the method results in 2 point degradation for the context length that is within the pretraining setting. This is significant and further evaluations and studies is needed. One of the main challenges of long context extension is preserving short context benchmarks.", "questions": "- For the decoder part, have the authors considered a model that uses a bi-directional attention mask?\n- Have the authors compared the method to models such as T5 with iso parameters (decoder + encoder)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730677602518}, {"id": "W2c5e4ikAu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2386/Reviewer_dDMd"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The study proposes a new KV cache compression method, called Recurrent Context Compression (RCC). The proposed framework consists of a RCC encoder and a RCC decoder. The RCC encoder compresses the input prompt segment-by-segment into a much shorter vector of context. The compressed prompt is then under-gone under a MLP, and fed to the RCC decoder afterward. The framework requires fine-tuning. The proposed method claims significant performance gain on long-context tasks.", "review_text": "The study proposes a new KV cache compression method, called Recurrent Context Compression (RCC). The proposed framework consists of a RCC encoder and a RCC decoder. The RCC encoder compresses the input prompt segment-by-segment into a much shorter vector of context. The compressed prompt is then under-gone under a MLP, and fed to the RCC decoder afterward. The framework requires fine-tuning. The proposed method claims significant performance gain on long-context tasks.", "strengths": "- The model claims substantial performance improvement on long-context task.\n- The idea is straight-forward and easy to understand.", "weaknesses": "1. The experiment lacks comparison with other KV cache compression methods & baselines. For example, how is the proposed method compared with other methods on LongBench-E Document tasks or NIAH [1] [2] [3] [4]? \n2. What is the memory usage & time efficiency gain in comparison with other methods given the proposed method require fine-tuning to perform prompt compression?\n3. The architecture used in the experiment, i.e. pythia-1.4b & mamba-1.4b, is small. I also saw the study did experiment on llama-2-7b in section 4.3. Can the study provide more results on more architectures such as llama3 family or mistral family for the encoder & decoder?\n4. If we vary the length of segment to compress in the encoder, I'm wondering how influence can this factor be? Can we do an ablation study to test this?\n5. Even though the observation on context-instruction confusion due to prompt compression seems interesting, is there any experimental result to support this claim?\n\n6. Nitpicking: the result presentation part is somehow unclear:\n- In Table 1: what  does three numbers in each cell present? For example, \"96/98/94\"\n- In Table 2: is each cell the average result among 4 subtasks from LongBench-E with the specific context length from the column's header?\n\nWhile the proposed method is interesting, I feel like the experiment is incomplete to show the effectiveness of the method, so I recommend rejection for further improvement for now, but I am willing to reconsider and increase the score a bit if the authors can address my upper concerns.\n\n[1] MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention. Jiang et al., https://arxiv.org/pdf/2407.02490 \n\n[2] Razorattention: Efficient kv cache compression through retrieval heads, 2024. Tang et al., https: //arxiv.org/abs/2407.15891\n\n[3] InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory, Xiao et al., https://arxiv.org/pdf/2402.04617\n\n[4] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache, Liu et al., https://arxiv.org/pdf/2402.02750", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The study proposes a new KV cache compression method, called Recurrent Context Compression (RCC). The proposed framework consists of a RCC encoder and a RCC decoder. The RCC encoder compresses the input prompt segment-by-segment into a much shorter vector of context. The compressed prompt is then under-gone under a MLP, and fed to the RCC decoder afterward. The framework requires fine-tuning. The proposed method claims significant performance gain on long-context tasks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- The model claims substantial performance improvement on long-context task.\n- The idea is straight-forward and easy to understand.", "weaknesses": "1. The experiment lacks comparison with other KV cache compression methods & baselines. For example, how is the proposed method compared with other methods on LongBench-E Document tasks or NIAH [1] [2] [3] [4]? \n2. What is the memory usage & time efficiency gain in comparison with other methods given the proposed method require fine-tuning to perform prompt compression?\n3. The architecture used in the experiment, i.e. pythia-1.4b & mamba-1.4b, is small. I also saw the study did experiment on llama-2-7b in section 4.3. Can the study provide more results on more architectures such as llama3 family or mistral family for the encoder & decoder?\n4. If we vary the length of segment to compress in the encoder, I'm wondering how influence can this factor be? Can we do an ablation study to test this?\n5. Even though the observation on context-instruction confusion due to prompt compression seems interesting, is there any experimental result to support this claim?\n\n6. Nitpicking: the result presentation part is somehow unclear:\n- In Table 1: what  does three numbers in each cell present? For example, \"96/98/94\"\n- In Table 2: is each cell the average result among 4 subtasks from LongBench-E with the specific context length from the column's header?\n\nWhile the proposed method is interesting, I feel like the experiment is incomplete to show the effectiveness of the method, so I recommend rejection for further improvement for now, but I am willing to reconsider and increase the score a bit if the authors can address my upper concerns.\n\n[1] MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention. Jiang et al., https://arxiv.org/pdf/2407.02490 \n\n[2] Razorattention: Efficient kv cache compression through retrieval heads, 2024. Tang et al., https: //arxiv.org/abs/2407.15891\n\n[3] InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory, Xiao et al., https://arxiv.org/pdf/2402.04617\n\n[4] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache, Liu et al., https://arxiv.org/pdf/2402.02750", "questions": "Please see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730442072548}, {"id": "bzAOtRBLsl", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2386/Reviewer_t2N7"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes Recurrent Context Compression, or RCC for short, that is designed to extend the context length of large language models. RCC can be trained with both text reconstruction and text continuation objectives, and a special instruction reconstruction process is used to mitigate the issue of performance degradation after compression. RCC is tested across multiple tasks and achieves comparable performance to un-compressed methods.", "review_text": "This paper proposes Recurrent Context Compression, or RCC for short, that is designed to extend the context length of large language models. RCC can be trained with both text reconstruction and text continuation objectives, and a special instruction reconstruction process is used to mitigate the issue of performance degradation after compression. RCC is tested across multiple tasks and achieves comparable performance to un-compressed methods.", "strengths": "S1. RCC achieves a high compression ratio compared to its baseline method ICAE and is proven to preserve most of the context information across experiments of text reconstruction and downstream tasks.\n\nS2. The training paradigm of RCC is carefully designed. For example, it proposes an instruction-aware reconstruction task to mitigate the performance degradation issue. Experiments show that the instruction reconstruction is taking effect in the ability of LLMs to follow the instructions. It also proposes a short-to-long training paradigm to save the computational overhead during training.", "weaknesses": "Some ideas proposed in this paper are not new:\n\nW1: The paper improves the final-layer compression solution of ICAE (Ge et al., 2024) and AutoCompressor (Chevalier et al., 2023) by incorporating the hidden representation of each layer. However, similar ideas are used by prior work such as Compressive Transformers (Rae et al., 2020) and Dodo (Qin et al., 2024). The solution adopted by RCC is to split and re-use the last token representation of each segment at each layer, while this idea has already been applied by prior work. Compressive Transformers compress each segment with a pooling operator and Dodo dynamically selects the split points instead of employing a fixed-length segmentation, both working at layer level. \n\nW2: Gradual training on from short to long sequences was deployed in many prior works, though for different purposes. E.g., the training sequences of Longformer are gradually prolonged from the limit of BERT to much longer.\n\nSome experiments in this paper are inconclusive:\n\nW3: The paper claims that RCC outperforms ICAE in the task of text reconstruction. However, text reconstruction is a relatively easy task for LLMs and a BLEU score of 0.95 isn't good enough. E.g., in the Dodo paper, LLaMA can achieve a BLEU of 0.98 with a compression ratio of 20x. \n\nW4: Many experiments only have infini-transformers as the baseline. Comparing to a single baseline is not sufficient to support that RCC is superior to prior methods. Some related methods, such as Recurrent Memory Transformer (RMT), should be tested.\n\nOther comments:\n\nW5: RCC can compress the context after the pre-filling. However, the major computational overhead of the transformer decoding is not the attention but the non-quadratic terms according to the estimation of this blogpost. This effect can be stronger for larger models. It would be great if the authors could show the wallclock time difference between LLM decodings with and without RCC.\n\nReferences:\n- Bulatov, Aydar, Yury Kuratov, and Mikhail Burtsev. \"Recurrent memory transformer.\" Advances in Neural Information Processing Systems 35 (2022): 11079-11091.\n- Qin, Guanghui, et al. \"Dodo: Dynamic Contextual Compression for Decoder-only LMs.\" Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.\n- Beltagy, Iz, Matthew E. Peters, and Arman Cohan. \"Longformer: The long-document transformer.\" arXiv preprint arXiv:2004.05150 (2020).", "questions": "Q1. RCC is trained to re-construct the instruction to prevent the model from forgetting the instruction. However, an alternative approach is to prepend/append the instruction after the compressed representation. As shown in Table 2, RCC-Ins-Human achieves better performance than compressing the instruction. Can you discuss the advantage of compressing the instructions?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Recurrent Context Compression, or RCC for short, that is designed to extend the context length of large language models. RCC can be trained with both text reconstruction and text continuation objectives, and a special instruction reconstruction process is used to mitigate the issue of performance degradation after compression. RCC is tested across multiple tasks and achieves comparable performance to un-compressed methods.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "S1. RCC achieves a high compression ratio compared to its baseline method ICAE and is proven to preserve most of the context information across experiments of text reconstruction and downstream tasks.\n\nS2. The training paradigm of RCC is carefully designed. For example, it proposes an instruction-aware reconstruction task to mitigate the performance degradation issue. Experiments show that the instruction reconstruction is taking effect in the ability of LLMs to follow the instructions. It also proposes a short-to-long training paradigm to save the computational overhead during training.", "weaknesses": "Some ideas proposed in this paper are not new:\n\nW1: The paper improves the final-layer compression solution of ICAE (Ge et al., 2024) and AutoCompressor (Chevalier et al., 2023) by incorporating the hidden representation of each layer. However, similar ideas are used by prior work such as Compressive Transformers (Rae et al., 2020) and Dodo (Qin et al., 2024). The solution adopted by RCC is to split and re-use the last token representation of each segment at each layer, while this idea has already been applied by prior work. Compressive Transformers compress each segment with a pooling operator and Dodo dynamically selects the split points instead of employing a fixed-length segmentation, both working at layer level. \n\nW2: Gradual training on from short to long sequences was deployed in many prior works, though for different purposes. E.g., the training sequences of Longformer are gradually prolonged from the limit of BERT to much longer.\n\nSome experiments in this paper are inconclusive:\n\nW3: The paper claims that RCC outperforms ICAE in the task of text reconstruction. However, text reconstruction is a relatively easy task for LLMs and a BLEU score of 0.95 isn't good enough. E.g., in the Dodo paper, LLaMA can achieve a BLEU of 0.98 with a compression ratio of 20x. \n\nW4: Many experiments only have infini-transformers as the baseline. Comparing to a single baseline is not sufficient to support that RCC is superior to prior methods. Some related methods, such as Recurrent Memory Transformer (RMT), should be tested.\n\nOther comments:\n\nW5: RCC can compress the context after the pre-filling. However, the major computational overhead of the transformer decoding is not the attention but the non-quadratic terms according to the estimation of this blogpost. This effect can be stronger for larger models. It would be great if the authors could show the wallclock time difference between LLM decodings with and without RCC.\n\nReferences:\n- Bulatov, Aydar, Yury Kuratov, and Mikhail Burtsev. \"Recurrent memory transformer.\" Advances in Neural Information Processing Systems 35 (2022): 11079-11091.\n- Qin, Guanghui, et al. \"Dodo: Dynamic Contextual Compression for Decoder-only LMs.\" Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.\n- Beltagy, Iz, Matthew E. Peters, and Arman Cohan. \"Longformer: The long-document transformer.\" arXiv preprint arXiv:2004.05150 (2020).", "questions": "Q1. RCC is trained to re-construct the instruction to prevent the model from forgetting the instruction. However, an alternative approach is to prepend/append the instruction after the compressed representation. As shown in Table 2, RCC-Ins-Human achieves better performance than compressing the instruction. Can you discuss the advantage of compressing the instructions?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729750246358}], "openreview_url": "https://openreview.net/forum?id=GYk0thSY1M", "arxiv_id": "2406.06110", "paper_pdf": "papers/GYk0thSY1M.pdf", "paper_pdf_sha256": "b628086ad454173cf7779ed38cd7383a4eae84e8bc94934859d635547f9439d1", "paper_pdf_bytes": 587121, "paper_pdf_source": "openreview", "code_url": "https://github.com/WUHU-G/RCC_Transformer", "code_repository": "WUHU-G/RCC_Transformer", "code_commit": "b9c5486c49708cd284dbf1229fe6bc054d424453", "code_archive": "repos/GYk0thSY1M.zip", "code_archive_sha256": "d95d1c0e10e74490ebfe2dfb319de02fcf96746a3c7a4cc88b37dbbedff2a15e", "code_archive_bytes": 274055, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 339, "github_languages": {"Python": 40685}, "github_archived": false, "github_pushed_at": "2024-06-13T02:27:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/recurrent-context-compression-efficiently"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "U2ZIgcrg7Z", "year": 2024, "status": "rejected", "title": "ZOOPFL: EXPLORING BLACK-BOX FOUNDATION MODELS FOR PERSONALIZED FEDERATED LEARNING", "authors": ["Wang Lu", "Hao Yu", "Jindong Wang", "Damien Teney", "Haohan Wang", "Yiqiang Chen", "Qiang Yang", "Xing Xie", "Xiangyang Ji"], "authorids": ["~Wang_Lu2", "~Hao_Yu8", "~Jindong_Wang1", "~Damien_Teney1", "~Haohan_Wang1", "~Yiqiang_Chen1", "~Qiang_Yang1", "~Xing_Xie3", "~Xiangyang_Ji1"], "authors_source": "OpenReview API", "abstract": "When personalized federated learning (FL) meets large foundation models, new challenges arise from various limitations in resources. In addition to typical limitations such as data, computation, and communication costs, access to the models is also often limited. This paper endeavors to solve both the challenges of limited resources and personalization. i.e., distribution shifts between clients. To do so, we propose a method named ZOOPFL that uses Zeroth-Order Optimization for Personalized Federated Learning. ZOOPFL avoids direct interference with the foundation models and instead learns to adapt its inputs through zeroth-order optimization. In addition, we employ simple yet effective linear projections to remap its predictions for personalization. To reduce the computation costs and enhance personalization, we propose input surgery to incorporate an auto-encoder with low-dimensional and client-specific embeddings. We provide theoretical support for ZOOPFL to analyze its convergence. Extensive empirical experiments on computer vision and natural language processing tasks using popular foundation models demonstrate its effectiveness for FL on black-box foundation models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "RSKdKdiypw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7431/Reviewer_MDQp"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper presents a method for personalized federated learning (FL) while relying on the existence of foundation models at clients. The main idea is to train some additional components (auto-encoder and semantic re-mapping) that are applied before or after the foundation model. Zeroth-order optimization has been applied due to the assumption that the foundation model cannot be accessed for purposes other than inference. Experimental results confirm the advantage of the proposed method compared to some baselines.", "review_text": "The paper presents a method for personalized federated learning (FL) while relying on the existence of foundation models at clients. The main idea is to train some additional components (auto-encoder and semantic re-mapping) that are applied before or after the foundation model. Zeroth-order optimization has been applied due to the assumption that the foundation model cannot be accessed for purposes other than inference. Experimental results confirm the advantage of the proposed method compared to some baselines.", "strengths": "- The consideration of foundation models in FL is an important research direction.", "weaknesses": "- The paper assumes that foundation models are located at FL clients, but the clients cannot perform back-propagation on these models. It is not clear in what practical scenario such an assumption would hold. It is worth noting that most large language models (LLMs) nowadays are hosted in the cloud. Obviously, transmitting data to the cloud, even for inference, violates the privacy promise provided by FL. It seems that the authors of this paper try to overcome this privacy violation by assuming that the foundation model is hosted on each client. However, this has several issues. First, many types of LLMs are not feasible to run on mobile devices, which means that the proposed approach may only be possible in the case of cross-silo FL but not cross-device FL. Second, and more importantly, if the foundation model is hosted at the client, it is unclear why gradients cannot be computed, since each client has full access to its model in this case. \n- Overall, the proposed approach is a combination of several known techniques, including zeroth-order optimization, so the novelty seems limited. \n- The method requires additional components to be added to an existing foundation model, which appears to be a patch instead of a long-term solution. These additional components will cause additional computational overhead, which has not been studied in the paper.", "questions": "My questions are related to the weaknesses mentioned above, which are summarized as follows:\n- In what practical scenario would a FL client host a foundation model, but does not have full access to it?\n- What are the key technical challenges and novel solution in this work?\n- What is the additional computational overhead of the additional components (auto-encoder and semantic re-mapping) in the proposed method, when the full combined model is used for inference? It would be helpful to measure and compare the inference time with and without these additional components on a real device.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a method for personalized federated learning (FL) while relying on the existence of foundation models at clients. The main idea is to train some additional components (auto-encoder and semantic re-mapping) that are applied before or after the foundation model. Zeroth-order optimization has been applied due to the assumption that the foundation model cannot be accessed for purposes other than inference. Experimental results confirm the advantage of the proposed method compared to some baselines.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The consideration of foundation models in FL is an important research direction.", "weaknesses": "- The paper assumes that foundation models are located at FL clients, but the clients cannot perform back-propagation on these models. It is not clear in what practical scenario such an assumption would hold. It is worth noting that most large language models (LLMs) nowadays are hosted in the cloud. Obviously, transmitting data to the cloud, even for inference, violates the privacy promise provided by FL. It seems that the authors of this paper try to overcome this privacy violation by assuming that the foundation model is hosted on each client. However, this has several issues. First, many types of LLMs are not feasible to run on mobile devices, which means that the proposed approach may only be possible in the case of cross-silo FL but not cross-device FL. Second, and more importantly, if the foundation model is hosted at the client, it is unclear why gradients cannot be computed, since each client has full access to its model in this case. \n- Overall, the proposed approach is a combination of several known techniques, including zeroth-order optimization, so the novelty seems limited. \n- The method requires additional components to be added to an existing foundation model, which appears to be a patch instead of a long-term solution. These additional components will cause additional computational overhead, which has not been studied in the paper.", "questions": "My questions are related to the weaknesses mentioned above, which are summarized as follows:\n- In what practical scenario would a FL client host a foundation model, but does not have full access to it?\n- What are the key technical challenges and novel solution in this work?\n- What is the additional computational overhead of the additional components (auto-encoder and semantic re-mapping) in the proposed method, when the full combined model is used for inference? It would be helpful to measure and compare the inference time with and without these additional components on a real device.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699233190023}, {"id": "Vsb4g6PFuN", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7431/Reviewer_1Crz"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors designed a method that treats foundation models as black boxes. The idea is to use zeoth-order optimisation to  make sure that fine-tuning can be done efficiently on-device (e.g., to train through federated learning)", "review_text": "The authors designed a method that treats foundation models as black boxes. The idea is to use zeoth-order optimisation to  make sure that fine-tuning can be done efficiently on-device (e.g., to train through federated learning)", "strengths": "- Being able to fine-tune Foundation Models (FM) in a privacy-preserving way is an important problem\n- Overall this paper helps us understand the scenario of incorporating FM in FL settings. \n- The use of zeroth-order optimization, input surgery and semantic re-mapping are interesting contributions here", "weaknesses": "Some areas to improve:\n\n- While the idea of using FM as black box is interesting, there might be some privacy implications. It is unclear if the input to the FM reveals any information about the private input that is used both for training and during inference. I assume that the FM --being a black box and too big to be hosted on-device-- is run externally). As a result, this method might limit the ability of FL to offer privacy-preserving training. \nIn other words,  If we assume that the \"black box\" in figure 2.b runs externally, what are the privacy implications wrt to its input and output crossing the device boundary. If it runs on-device then what are the assumptions wrt to its size and the fact that it is a black box. \n\n- The authors assume that the FM is a black box. With more and more FM being open-sourced, it would be great if the authors can further motivate their approach and what might be the main advantages of incorporating a black box. \n\n- The evaluation is mostly done on rather simple benchmarks. I was wondering if the proposed approach (to train just parts of the model) would carry enough capacity to tackle larger tasks. Maybe some discussion or even evaluation on a more complex task would be great. \n\n- The paper might benefit from some understanding of the memory footprint and computation complexity of this method. Overall, the main target of this method is to make FM training possible with FL (on-device). As a result, we should have a good understanding on the memory/computation overhead.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors designed a method that treats foundation models as black boxes. The idea is to use zeoth-order optimisation to  make sure that fine-tuning can be done efficiently on-device (e.g., to train through federated learning)", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Being able to fine-tune Foundation Models (FM) in a privacy-preserving way is an important problem\n- Overall this paper helps us understand the scenario of incorporating FM in FL settings. \n- The use of zeroth-order optimization, input surgery and semantic re-mapping are interesting contributions here", "weaknesses": "Some areas to improve:\n\n- While the idea of using FM as black box is interesting, there might be some privacy implications. It is unclear if the input to the FM reveals any information about the private input that is used both for training and during inference. I assume that the FM --being a black box and too big to be hosted on-device-- is run externally). As a result, this method might limit the ability of FL to offer privacy-preserving training. \nIn other words,  If we assume that the \"black box\" in figure 2.b runs externally, what are the privacy implications wrt to its input and output crossing the device boundary. If it runs on-device then what are the assumptions wrt to its size and the fact that it is a black box. \n\n- The authors assume that the FM is a black box. With more and more FM being open-sourced, it would be great if the authors can further motivate their approach and what might be the main advantages of incorporating a black box. \n\n- The evaluation is mostly done on rather simple benchmarks. I was wondering if the proposed approach (to train just parts of the model) would carry enough capacity to tackle larger tasks. Maybe some discussion or even evaluation on a more complex task would be great. \n\n- The paper might benefit from some understanding of the memory footprint and computation complexity of this method. Overall, the main target of this method is to make FM training possible with FL (on-device). As a result, we should have a good understanding on the memory/computation overhead.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698772952418}, {"id": "9beNi98vOi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7431/Reviewer_hxrH"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposed ZOOPFL, a zeroth-order optimization system for the black-box local model under a federated learning setup. Instead of directly fine-tune the black-box foundational model, ZOOPFL learns input surgery and semantic re-mapping for black-box large foundation models in federated learning. ZOOPFL aims to adapt inputs to models and project outputs to meaningful semantic space. The experiment shows that ZOOPFL performs better than the ZS setup in both NLP and CV benchmark datasets.", "review_text": "This paper proposed ZOOPFL, a zeroth-order optimization system for the black-box local model under a federated learning setup. Instead of directly fine-tune the black-box foundational model, ZOOPFL learns input surgery and semantic re-mapping for black-box large foundation models in federated learning. ZOOPFL aims to adapt inputs to models and project outputs to meaningful semantic space. The experiment shows that ZOOPFL performs better than the ZS setup in both NLP and CV benchmark datasets.", "strengths": "1. In the current foundational model era, the black-box foundational model is becoming popular. It is important to propose some ideas to efficiently personalize the foundational model without direct interference with it. Compared to other existing works related to foundational model with FL, the proposed ZOOPFL is the first to achieve federated learning with large black-box models, which is very relative to the current challenges.\n\n2. The paper is well-written and clearly structured. The author selects two different data modalities to validate the soundness of the proposed ZooPFL.", "weaknesses": "1. The idea seems very similar to the soft-prompt training [1], which is also working on the input surgery without directly inference the foundational model. What is the benefit of the auto-encoder pre-training in your paper?\n\n2. What are the benefits of personalization? In the experiment part, it mainly focused on the overall accuracy boost compared to ZS, which does not reflect anything regarding to personalization.\n\n3. I suggest the paper should be more clear about the only baseline ZS. I checked several times in the paper, and I could not understand what ZS stands for and why it is a suitable baseline for ZooPFL.\n\n\n[1]. Wang, Zifeng et al. “Learning to Prompt for Continual Learning.” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 139-149.", "questions": "1. What does ZS stand for in the paper? Does it stand for zero-shot training?\n\n2. I am curious why the author selected the personalized FL as a topic to discuss. Even without step 3, this paper still makes its point regarding how to efficiently use the black-box model under FL setup.\n\n3. I am not very clear why ZooPFL needs Semantic re-mapping. Could the author elaborate more on this?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed ZOOPFL, a zeroth-order optimization system for the black-box local model under a federated learning setup. Instead of directly fine-tune the black-box foundational model, ZOOPFL learns input surgery and semantic re-mapping for black-box large foundation models in federated learning. ZOOPFL aims to adapt inputs to models and project outputs to meaningful semantic space. The experiment shows that ZOOPFL performs better than the ZS setup in both NLP and CV benchmark datasets.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. In the current foundational model era, the black-box foundational model is becoming popular. It is important to propose some ideas to efficiently personalize the foundational model without direct interference with it. Compared to other existing works related to foundational model with FL, the proposed ZOOPFL is the first to achieve federated learning with large black-box models, which is very relative to the current challenges.\n\n2. The paper is well-written and clearly structured. The author selects two different data modalities to validate the soundness of the proposed ZooPFL.", "weaknesses": "1. The idea seems very similar to the soft-prompt training [1], which is also working on the input surgery without directly inference the foundational model. What is the benefit of the auto-encoder pre-training in your paper?\n\n2. What are the benefits of personalization? In the experiment part, it mainly focused on the overall accuracy boost compared to ZS, which does not reflect anything regarding to personalization.\n\n3. I suggest the paper should be more clear about the only baseline ZS. I checked several times in the paper, and I could not understand what ZS stands for and why it is a suitable baseline for ZooPFL.\n\n\n[1]. Wang, Zifeng et al. “Learning to Prompt for Continual Learning.” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 139-149.", "questions": "1. What does ZS stand for in the paper? Does it stand for zero-shot training?\n\n2. I am curious why the author selected the personalized FL as a topic to discuss. Even without step 3, this paper still makes its point regarding how to efficiently use the black-box model under FL setup.\n\n3. I am not very clear why ZooPFL needs Semantic re-mapping. Could the author elaborate more on this?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698707536510}, {"id": "Nnhw0DRMWw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7431/Reviewer_RdGi"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper addresses challenges in personalized federated learning with large foundation models and limited resources, including data, computation, and model access. The proposed method, ZOOPFL (Zeroth-Order Optimization for Personalized Federated Learning), adapts inputs through zeroth-order optimization and uses linear projections for personalization. Input surgery is introduced to reduce computation costs and enhance personalization.", "review_text": "This paper addresses challenges in personalized federated learning with large foundation models and limited resources, including data, computation, and model access. The proposed method, ZOOPFL (Zeroth-Order Optimization for Personalized Federated Learning), adapts inputs through zeroth-order optimization and uses linear projections for personalization. Input surgery is introduced to reduce computation costs and enhance personalization.", "strengths": "- The experiments are comprehensive, covering multiple datasets in both computer vision (CV) and natural language processing (NLP) applications.\n- The paper's focus on federated learning settings that address both data privacy and model privacy is intriguing.", "weaknesses": "My main concerns are the validity and privacy risks of this FL setting. \n- First, the black-box FL setting lacks practicality. The paper assumes the existence of large foundation models on clients in the form of encrypted assets, and it does not require the uploading of transformed inputs. However, this does not align with the most common scenarios in machine learning model services, such as the access of various black-box large language models like ChatGPT. In practical scenarios, local data needs to be uploaded to the model service provider.  \n- Second, the motivation for deploying zeroth-order optimization methods based on the local encrypted black-box model setup is not well-motivated. This setting implies that it is entirely possible to train an white-box emulator [1] as a proxy for the black-box model and directly perform first-order optimization based on the white-box emulator. However, the authors do not provide relevant discussions and experimental comparisons.\n- In terms of model privacy, the privacy leakage of a black-box model is closely related to the number of queries [2], but the authors do not provide theoretical or empirical studies on this. \n- The experimental section lacks ablation experiments with varying levels of noise added on the transformed data and visualizations of transformed data.\n\n[1] Xiao, Guangxuan, Ji Lin, and Song Han. \"Offsite-tuning: Transfer learning without full model.\" arXiv preprint arXiv:2302.04870 (2023).  \n[2] Tsai, Yun-Yun, Pin-Yu Chen, and Tsung-Yi Ho. \"Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources.\" International Conference on Machine Learning. PMLR, 2020.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses challenges in personalized federated learning with large foundation models and limited resources, including data, computation, and model access. The proposed method, ZOOPFL (Zeroth-Order Optimization for Personalized Federated Learning), adapts inputs through zeroth-order optimization and uses linear projections for personalization. Input surgery is introduced to reduce computation costs and enhance personalization.", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "strengths": "- The experiments are comprehensive, covering multiple datasets in both computer vision (CV) and natural language processing (NLP) applications.\n- The paper's focus on federated learning settings that address both data privacy and model privacy is intriguing.", "weaknesses": "My main concerns are the validity and privacy risks of this FL setting. \n- First, the black-box FL setting lacks practicality. The paper assumes the existence of large foundation models on clients in the form of encrypted assets, and it does not require the uploading of transformed inputs. However, this does not align with the most common scenarios in machine learning model services, such as the access of various black-box large language models like ChatGPT. In practical scenarios, local data needs to be uploaded to the model service provider.  \n- Second, the motivation for deploying zeroth-order optimization methods based on the local encrypted black-box model setup is not well-motivated. This setting implies that it is entirely possible to train an white-box emulator [1] as a proxy for the black-box model and directly perform first-order optimization based on the white-box emulator. However, the authors do not provide relevant discussions and experimental comparisons.\n- In terms of model privacy, the privacy leakage of a black-box model is closely related to the number of queries [2], but the authors do not provide theoretical or empirical studies on this. \n- The experimental section lacks ablation experiments with varying levels of noise added on the transformed data and visualizations of transformed data.\n\n[1] Xiao, Guangxuan, Ji Lin, and Song Han. \"Offsite-tuning: Transfer learning without full model.\" arXiv preprint arXiv:2302.04870 (2023).  \n[2] Tsai, Yun-Yun, Pin-Yu Chen, and Tsung-Yi Ho. \"Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources.\" International Conference on Machine Learning. PMLR, 2020.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698701788892}], "openreview_url": "https://openreview.net/forum?id=U2ZIgcrg7Z", "arxiv_id": "2310.05143", "paper_pdf": "papers/U2ZIgcrg7Z.pdf", "paper_pdf_sha256": "a51156d337e77165e8ec1552603dd54bedafe5d76c58c3bb6895b3feb3cb4225", "paper_pdf_bytes": 2140291, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/PersonalizedFL", "code_repository": "microsoft/PersonalizedFL", "code_commit": "441670edb5744d23af46867cce9a5fbec0f8ead3", "code_archive": "repos/U2ZIgcrg7Z.zip", "code_archive_sha256": "521bb8e68c430baf3e428468640c02aba5f11ce74194fbed0c9acd9ae583d203", "code_archive_bytes": 264542, "code_file_count": 22, "code_extensions": {".py": 21, ".sh": 1}, "github_disk_usage_kb": 279, "github_languages": {"Python": 75730, "Shell": 1210}, "github_archived": false, "github_pushed_at": "2023-10-04T01:05:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/zoopfl-exploring-black-box-foundation-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Gb2Rndy5595", "year": 2023, "status": "rejected", "title": "Context Autoencoder for Self-Supervised Representation Learning", "authors": ["Xiaokang Chen", "Mingyu Ding", "Xiaodi Wang", "Ying Xin", "Shentong Mo", "Yunhao Wang", "Shumin Han", "Ping Luo", "Gang Zeng", "Jingdong Wang"], "authorids": ["~Xiaokang_Chen1", "~Mingyu_Ding1", "~Xiaodi_Wang2", "~Ying_Xin1", "~Shentong_Mo1", "~Yunhao_Wang1", "~Shumin_Han1", "~Ping_Luo2", "~Gang_Zeng1", "~Jingdong_Wang1"], "authors_source": "OpenReview API", "abstract": "We present a novel masked image modeling (MIM) approach, context autoencoder (CAE), for self-supervised representation pretraining. The goal is to pretrain an encoder by solving the pretext task: estimate the masked patches from the visible patches in an image. Our approach first feeds the visible patches into the encoder, extracting the representations. Then, we make predictions from visible patches to masked patches in the encoded representation space. We introduce an alignment constraint, encouraging that the representations for masked patches, predicted from the encoded representations of visible patches, are aligned with the masked patch presentations computed from the encoder. In other words, the predicted representations are expected to lie in the encoded representation space, which empirically shows the benefit to representation learning. Last, the predicted masked patch representations are mapped to the targets of the pretext task through a decoder.\nOne additional characteristic is that our approach encourages the separation of the representation learning part (encoder), and the pretext task completion part that will be replaced by the downstream task part. In contrast, previous MIM methods (e.g., BEiT and MAE) couple the two parts, potentially limiting the representation learning quality. We demonstrate the effectiveness of our CAE through superior transfer performance in downstream tasks: semantic segmentation, and object detection and instance segmentation.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "hzPOF35_MB0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper846/Reviewer_KNoM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposed a novel contextual autoencoder framework for SSL. The proposed CAE approach introduced a cross-attention module to learn the latent space knowledge between the visible and masked patches. The paper had a comprehensive analysis between the proposed approach and the existing SSL approaches which gave some insight to the reader to understand the core contributions.  \n", "review_text": "The paper is an improvement work over BEIT and MAE with some minor novelties. The proposed CAE algorithm achieved the SOTA performance on various backbone and downstream tasks. The paper can be improved by clarifying the technical details and adding ablation studies.\n", "strengths": "Strength\n1. The paper is well written and easy to follow. The comprehensive analysis between the proposed approach and the existing SSL works add a valuable insight into its contributions.\n2. The evaluation confirms the proposed approach can outperform the SOTA SSL approach on various backbone and downstream tasks.\n\nWeakness\n1. The novelty of this paper is limited. The proposed CAE is an improvement work over MAE and BEIT by further exploiting the knowledge between the masked and visible patches in the latent space. There is no sufficient evidence that the proposed cross-attention module can further improve representation performance learned by CAE.\n2. The proposed CAE approach used a random block-wise masking approach. It is necessary to add an ablation study about the performance between various masking approaches, e.g. random patch masking. Moreover, the masking ratio is set to 0.5 by default. What would the performance be for other masking ratios? An ablation study would help clarify that.\n3. Some technical details are missed in the paper. E.g. how to decode \\bar{Y}_m from \\bar{Z}_m in figure 7a is not clear. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposed a novel contextual autoencoder framework for SSL. The proposed CAE approach introduced a cross-attention module to learn the latent space knowledge between the visible and masked patches. The paper had a comprehensive analysis between the proposed approach and the existing SSL approaches which gave some insight to the reader to understand the core contributions.  \n", "strength_and_weaknesses": "Strength\n1. The paper is well written and easy to follow. The comprehensive analysis between the proposed approach and the existing SSL works add a valuable insight into its contributions.\n2. The evaluation confirms the proposed approach can outperform the SOTA SSL approach on various backbone and downstream tasks.\n\nWeakness\n1. The novelty of this paper is limited. The proposed CAE is an improvement work over MAE and BEIT by further exploiting the knowledge between the masked and visible patches in the latent space. There is no sufficient evidence that the proposed cross-attention module can further improve representation performance learned by CAE.\n2. The proposed CAE approach used a random block-wise masking approach. It is necessary to add an ablation study about the performance between various masking approaches, e.g. random patch masking. Moreover, the masking ratio is set to 0.5 by default. What would the performance be for other masking ratios? An ablation study would help clarify that.\n3. Some technical details are missed in the paper. E.g. how to decode \\bar{Y}_m from \\bar{Z}_m in figure 7a is not clear. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written with most of the details clearly presented.\n\nThe paper’s quality is good in general but missed some technical details.\n\nThe paper is an improvement of BEIT and MAE. The novelty is limited as the comments above.\n", "summary_of_the_review": "The paper is an improvement work over BEIT and MAE with some minor novelties. The proposed CAE algorithm achieved the SOTA performance on various backbone and downstream tasks. The paper can be improved by clarifying the technical details and adding ablation studies.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667286684500}, {"id": "AWlwMmTGlVY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper846/Reviewer_TXRx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents context autoencoder (CAE), which follows BEiT and uses an additional latent contextual regressor to make predictions for the masked patches with cross attention over the visible patches. By doing that we can encourage the masked patches' representation predicted from visible patches are aligned with the one from the encoder and then benefits the learning and downstream transferring. Experiments on downstream tasks also show that it can achieve better performance compared with the baseline BEiT and MAE.", "review_text": "I think this paper shows some interesting findings and good performance, although the technical contribution is somewhat limited from my perspective. Thus my rating is 6.", "strengths": "The paper is well-written. The motivation of this paper is clear and easy to follow.\n\nThe proposed attentive probing is interesting and well-motivated, however, it seems that even with this metric the CAE trained with 800 epochs using ViT-B cannot surpass the contrastive learning-based model MoCo-v3 according to Table 1.\n\nPerformance wise, the downstream task results on COCO look superior, especially when using ViT-B, but when moving to ViT-L the gain seems diminished (from ~2 points to ~0.5 points compared with MAE on 1600ep).\n\nThe technical part seems to be somewhat incremental, how important/critical the alignment is? Seems like this issue only happens when making predictions in the representation space. For example, MAE without the alignment loss will not show meaningless output like Figure 3. The role of the cross-attention based latent contextual regressor is also not clear, one could also use the output from the MAE decoder's first layer to compute the alignment loss.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents context autoencoder (CAE), which follows BEiT and uses an additional latent contextual regressor to make predictions for the masked patches with cross attention over the visible patches. By doing that we can encourage the masked patches' representation predicted from visible patches are aligned with the one from the encoder and then benefits the learning and downstream transferring. Experiments on downstream tasks also show that it can achieve better performance compared with the baseline BEiT and MAE.", "strength_and_weaknesses": "The paper is well-written. The motivation of this paper is clear and easy to follow.\n\nThe proposed attentive probing is interesting and well-motivated, however, it seems that even with this metric the CAE trained with 800 epochs using ViT-B cannot surpass the contrastive learning-based model MoCo-v3 according to Table 1.\n\nPerformance wise, the downstream task results on COCO look superior, especially when using ViT-B, but when moving to ViT-L the gain seems diminished (from ~2 points to ~0.5 points compared with MAE on 1600ep).\n\nThe technical part seems to be somewhat incremental, how important/critical the alignment is? Seems like this issue only happens when making predictions in the representation space. For example, MAE without the alignment loss will not show meaningless output like Figure 3. The role of the cross-attention based latent contextual regressor is also not clear, one could also use the output from the MAE decoder's first layer to compute the alignment loss.", "clarity,_quality,_novelty_and_reproducibility": "In the paragraph of `Relation to BEiT and MAE.`, the authors mentioned that `In MAE (He et al., 2022), the so-called\ndecoder may play a partial role for representation learning as the representations of the visible\npatches are also updated in the MAE decoder`, which is somewhat inaccurate and makes me confused as the \"updated representation\" for visible patches is neither supervised with the reconstruction loss nor used for downstream tasks. Whether using extra layers to process representations for visible patches should not be a key difference and the cross attention used in this paper also has MLP for the key/value (which are the visible patches feature).\n\nFigure 3 is somewhat hard to follow without referring to the caption, might be better to add some illustration about how the output is generated.\n\nMisc: I believe the citation of LARS is incorrect.", "summary_of_the_review": "I think this paper shows some interesting findings and good performance, although the technical contribution is somewhat limited from my perspective. Thus my rating is 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666578584566}, {"id": "hTgjKfhn-E", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper846/Reviewer_b4Df"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed a method for self-supervised representation learning based on masked image modeling. Unlike previous approaches, the proposed method decouple representation learning part and pretext task completion part where the learning signal comes from the reconstruction in the encoded representation space rather than in image space. Alignment constraint is introduced encouraging the predicted representations to be lied in the encoded one. The experiments are conducted on three different downstream tasks; semantic segmentation, object detection, and instance segmentation, surpassing the previous approaches.", "review_text": "As the weaknesses outweigh the strengths, I lean towards 5: marginally below the acceptance threshold at this point.", "strengths": "** Strengths\n- Comparisons to the concurrent methods with analysis and discussion\n- Extensive experimental results achieving state-of-the-art performances on various downstream tasks\n- Clear implementation details\n- Using pretrained tokenizer for masking strategy is interesting\n\n** Weaknesses\n- The idea of reconstructing just the features of the masked patches is also used in iBOT\n- The reliance on a pre-trained tokenizer could be cumbersome. While being with 250M images as extra data, there are not many insights and improvements compared to previous works, e.g. competitive performances compared to iBOT and MAE even they don't require using such pre-trained tokenizers\n- ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors proposed a method for self-supervised representation learning based on masked image modeling. Unlike previous approaches, the proposed method decouple representation learning part and pretext task completion part where the learning signal comes from the reconstruction in the encoded representation space rather than in image space. Alignment constraint is introduced encouraging the predicted representations to be lied in the encoded one. The experiments are conducted on three different downstream tasks; semantic segmentation, object detection, and instance segmentation, surpassing the previous approaches.", "strength_and_weaknesses": "** Strengths\n- Comparisons to the concurrent methods with analysis and discussion\n- Extensive experimental results achieving state-of-the-art performances on various downstream tasks\n- Clear implementation details\n- Using pretrained tokenizer for masking strategy is interesting\n\n** Weaknesses\n- The idea of reconstructing just the features of the masked patches is also used in iBOT\n- The reliance on a pre-trained tokenizer could be cumbersome. While being with 250M images as extra data, there are not many insights and improvements compared to previous works, e.g. competitive performances compared to iBOT and MAE even they don't require using such pre-trained tokenizers\n- ", "clarity,_quality,_novelty_and_reproducibility": "Please see Strength And Weaknesses section above.", "summary_of_the_review": "As the weaknesses outweigh the strengths, I lean towards 5: marginally below the acceptance threshold at this point.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666502014635}, {"id": "nFwISSBSTF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper846/Reviewer_cbX1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to decouple the encoder and decoder space of MAE, thus improving the representation quality. Similar to contrastive learning, a Latent contextual regressor is further used to align the predicted masked patch with the encoded masked patch. Experiments on downstream object detection and semantic segmentation show the effectiveness of the proposed method.", "review_text": "I am inclined to accept the paper if the author can dispel my concerns well.", "strengths": "Strength:  \nGood results on downstream transferring learning, especially for semantic segmentation.  \nClear idea, decoupling the space of encoder and decoder explicitly may benefit the learning of representation.  \n\nWeaknesses:  \n(1) In fact, the “Context” in the Title does not correspond well with the interpretation in the main manuscript. MAE, BEiT and other Masked Image Modeling approaches can also be understood as including context information. I am not sure what specific meaning the context in the paper refers to.  \n(2) Several details and designs are not clear:  \n-- Why is the masking rate set to 50%? (Section 3.2) Why not 75% or other values?  \n-- How to balance the two loss functions? (Eq (1))  \n-- How to choose the number of layers in the Latent Contextual Regressor and Decoder? The number of layers may be related to the performance.    \n\n(3) Questions about the experiments:  \n-- In fact, the finetuning results in Table 1 are only slightly higher than MAE (0.3%).  \n-- How about removing the Decoder? That’s only containing the latent contextual regressor. In addition, in Table 2, why not use the finetuning or linear probe metrics?  \n-- In Table 3, according to the MAE paper, the detection performance of MAE with ViT-B is 50.3%, not 48.4%. In addition, MAE+ViTDet [1] has already achieved higher results (ViT-B: 51.6% and ViT-L: 56.7%).  \n[1] Yanghao Li, et al. Exploring Plain Vision Transformer Backbones for Object Detection. ECCV2022.  \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposes to decouple the encoder and decoder space of MAE, thus improving the representation quality. Similar to contrastive learning, a Latent contextual regressor is further used to align the predicted masked patch with the encoded masked patch. Experiments on downstream object detection and semantic segmentation show the effectiveness of the proposed method.", "strength_and_weaknesses": "Strength:  \nGood results on downstream transferring learning, especially for semantic segmentation.  \nClear idea, decoupling the space of encoder and decoder explicitly may benefit the learning of representation.  \n\nWeaknesses:  \n(1) In fact, the “Context” in the Title does not correspond well with the interpretation in the main manuscript. MAE, BEiT and other Masked Image Modeling approaches can also be understood as including context information. I am not sure what specific meaning the context in the paper refers to.  \n(2) Several details and designs are not clear:  \n-- Why is the masking rate set to 50%? (Section 3.2) Why not 75% or other values?  \n-- How to balance the two loss functions? (Eq (1))  \n-- How to choose the number of layers in the Latent Contextual Regressor and Decoder? The number of layers may be related to the performance.    \n\n(3) Questions about the experiments:  \n-- In fact, the finetuning results in Table 1 are only slightly higher than MAE (0.3%).  \n-- How about removing the Decoder? That’s only containing the latent contextual regressor. In addition, in Table 2, why not use the finetuning or linear probe metrics?  \n-- In Table 3, according to the MAE paper, the detection performance of MAE with ViT-B is 50.3%, not 48.4%. In addition, MAE+ViTDet [1] has already achieved higher results (ViT-B: 51.6% and ViT-L: 56.7%).  \n[1] Yanghao Li, et al. Exploring Plain Vision Transformer Backbones for Object Detection. ECCV2022.  \n", "clarity,_quality,_novelty_and_reproducibility": "Overall, the proposed method may be of some inspiration to the community. However, some details and experiments need to be further investigated.  \nLack of limitations (e.g. training costs analysis).  \n", "summary_of_the_review": "I am inclined to accept the paper if the author can dispel my concerns well.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666351471675}], "openreview_url": "https://openreview.net/forum?id=Gb2Rndy5595", "arxiv_id": "2202.03026", "paper_pdf": "papers/Gb2Rndy5595.pdf", "paper_pdf_sha256": "4a8595a46e8939fb558f5ea4172977215b515901ea72201e30d8c4c1dbd6798a", "paper_pdf_bytes": 12318393, "paper_pdf_source": "openreview", "code_url": "https://github.com/Atten4Vis/CAE", "code_repository": "Atten4Vis/CAE", "code_commit": "a7fd1628176358e3e76b0b042b0f6e9f3cd7c76c", "code_archive": "repos/Gb2Rndy5595.zip", "code_archive_sha256": "a6976061e708b93dc0342f85a13ad873737c513d26ee862459164dd3fb9f7a79", "code_archive_bytes": 756589, "code_file_count": 236, "code_extensions": {".py": 223, ".sh": 13}, "github_disk_usage_kb": 628, "github_languages": {"Python": 1231842, "Shell": 23917}, "github_archived": false, "github_pushed_at": "2023-11-28T06:45:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/context-autoencoder-for-self-supervised"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UFYYol-bRq", "year": 2022, "status": "rejected", "title": "ANCER: Anisotropic Certification  via Sample-wise Volume Maximization", "authors": ["Francisco Eiras", "Motasem Alfarra", "Philip Torr", "M. Pawan Kumar", "Puneet K. Dokania", "Bernard Ghanem", "Adel Bibi"], "authorids": ["~Francisco_Eiras1", "~Motasem_Alfarra1", "~Philip_Torr1", "~M._Pawan_Kumar1", "~Puneet_K._Dokania1", "~Bernard_Ghanem1", "~Adel_Bibi1"], "authors_source": "OpenReview API", "abstract": "Randomized smoothing has recently emerged as an effective tool that enables certification of deep neural network classifiers at scale. All prior art on randomized smoothing has focused on isotropic $\\ell_p$ certification, which has the advantage of yielding certificates that can be easily compared among isotropic methods via $\\ell_p$-norm radius. However, isotropic certification limits the region that can be certified around an input to worst-case adversaries, i.e., it cannot reason about other \"close\", potentially large, constant prediction safe regions. To alleviate this issue, (i) we theoretically extend the isotropic randomized smoothing $\\ell_1$ and $\\ell_2$ certificates to their generalized anisotropic counterparts following a simplified analysis. Moreover, (ii) we propose evaluation metrics allowing for the comparison of general certificates - a certificate is superior to another if it certifies a superset region - with the quantification of each certificate through the volume of the certified region. We introduce ANCER, a framework for obtaining anisotropic certificates for a given test set sample via volume maximization. We achieve it by generalizing memory-based certification of data-dependent classifiers. Our empirical results demonstrate that ANCER achieves state-of-the-art $\\ell_1$ and $\\ell_2$ certified accuracy on CIFAR-10 and ImageNet, while certifying larger regions in terms of volume, highlighting the benefits of moving away from isotropic analysis.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "PDQLTOMKK8n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper441/Reviewer_xapX"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes the anisotropic version of randomized smoothing. Evaluation metrics based on the volume of the certified region are proposed, allowing comparisons with the certified regions provided from isotropic randomized smoothing. Experimental results show the usefulness of introducing anisotropic randomized smoothing as it certifies larger regions.", "review_text": "Strength\n\nThis paper is fairly interesting and insightful as it sufficiently demonstrates the advantage of introducing anisotropic robustness certificates. Although section 4.2 and 4.3, the main derivations of the non-isometric certification, are directly based on Salman et al. 2019(a), it is useful and allows less restrictive certification geometry. The discussion and adoption of Alfarra et al. 2020 in maximizing the volume through proxy radius are also interesting. Section 7.3 is especially helpful as it delves into the root keys of the better performance over isotropic DD.\n\n\nWeakness\n\nThe paper mentioned all prior works considered smoothing with isotropic distributions and hence certified isotropic ell_p-ball regions, but [*] admits non-isotropic certified radius bound via first-order certification. It will be helpful if the authors can include comparisons theoretically and empirically with this prior work, allowing a more fair evaluation of the significance of the paper.\n\nMohapatra, Jeet, Ching-Yun Ko, Tsui-Wei Weng, Pin-Yu Chen, Sijia Liu, and Luca Daniel. \"Higher-Order Certification for Randomized Smoothing.\" Advances in Neural Information Processing Systems 33 (2020).\n\n\nMinor error: There is a redundant 'the' in the first sentence of section 7.3.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes the anisotropic version of randomized smoothing. Evaluation metrics based on the volume of the certified region are proposed, allowing comparisons with the certified regions provided from isotropic randomized smoothing. Experimental results show the usefulness of introducing anisotropic randomized smoothing as it certifies larger regions.", "main_review": "Strength\n\nThis paper is fairly interesting and insightful as it sufficiently demonstrates the advantage of introducing anisotropic robustness certificates. Although section 4.2 and 4.3, the main derivations of the non-isometric certification, are directly based on Salman et al. 2019(a), it is useful and allows less restrictive certification geometry. The discussion and adoption of Alfarra et al. 2020 in maximizing the volume through proxy radius are also interesting. Section 7.3 is especially helpful as it delves into the root keys of the better performance over isotropic DD.\n\n\nWeakness\n\nThe paper mentioned all prior works considered smoothing with isotropic distributions and hence certified isotropic ell_p-ball regions, but [*] admits non-isotropic certified radius bound via first-order certification. It will be helpful if the authors can include comparisons theoretically and empirically with this prior work, allowing a more fair evaluation of the significance of the paper.\n\nMohapatra, Jeet, Ching-Yun Ko, Tsui-Wei Weng, Pin-Yu Chen, Sijia Liu, and Luca Daniel. \"Higher-Order Certification for Randomized Smoothing.\" Advances in Neural Information Processing Systems 33 (2020).\n\n\nMinor error: There is a redundant 'the' in the first sentence of section 7.3.", "summary_of_the_review": "The paper contributes adequately to the community and provides enough details regarding the problem formulation and empirical performance-cost tradeoff. Including a more complete comparison with the literature will make it a stronger submission.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635866547157}, {"id": "7YCKLcuu1Dl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper441/Reviewer_xsin"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors extend the standard (isotropic) l_1/l_2 randomized smoothing certificate to anisotropic smoothing distributions (and anisotropic regions). They propose to maximize the volume of the certified region for each sample independently, using a memory procedure to preserve soundness.", "review_text": "The paper is well written, polished, and easy to follow. The anistropic part of the proposed approach is well-motivated, however the sample-wise part has several issues (see below) making it unsuitable to use in practice. Moreover, some of the theoretical results are not novel (see below).\n\nComments:\n- The derivation of the certificate in term of the Lipschitz constant under the dual norm (Proposition 1, Theorem 1) is insightful. However, this result is already known in the literature in greater generality and should be properly cited. Specifically, Theorem 1 in [1] derives the local (\\alpha,\\beta)-Lipschitz constant over an open set. The special case of a scalar function is also discussed (see Eq. 3 in [1]) and it coincides with the result in Proposition 1 (if we are in the special case where the open set is R). The discussion of how to use such a Lipschitz constant for certification is also discussed in [1] (see section E.2).\n- The authors already discuss a know issue of data-dependent classifiers which when not tackled can lead to certificates that are not sound. To address the issue they adapt the memory-based procedure introduced in Alfarra et al. While this procedure does make the certificate sound it has other problems.\n  - First, by introducing memory the proposed anisotropic certificate (or in fact any method to compute a certified region) becomes completely irrelevant. Namely, for sample i construct any region R_i and define that the prediction in that region is some arbitrary C_i. Then run the post-processing certification step in Algorithm 1 (or similarly the procedure from Alfarra et al.). Since the obtained regions will not intersect with any certified regions in the memory M the procedure is still sound. This is true regardless of how the region R_i was constructed.\n  - A second, and more important, issue is that the memory makes the certificate dependent on the order of the incoming test samples. This provides a new avenue for attack, i.e. the adversary can optimize the order of the test samples to decrease the utility of the final obtained smoothed classifier.\n  - Finally, the success of this memory approach also somewhat depends on the \"sparsity\" of the test samples. Namely, by using a small test set since the samples are in a high-dimensional space the distance between them tends to be bigger than the (proxy) radii of the certified regions. However, in a real-world application we are likely to have many more test samples which would increase the number of intersections when running Algorithm 1.\n  - Note, all of these issue also apply to Alfarra et al. and are simply inherited in this approach.\n- Another issue with the proposed approach is the optimization procedure described in section C. The optimization suffers from issues such as: inconsistent estimation due to clamping and not using confidence bounds, sensitivity to initialization, high gradient variance, etc. These and other issues have been described in great detail in section C.2 in [2], focusing on Alfarra et al., although, given the similarity, the same issues are inherited in this approach.\n- Relatedly, in [2] the authors show that input-dependent randomized smoothing suffers from the curse of dimensionality. A discussion of this issue and how it applies to the proposed certificate should be included.\n- The derivations for the special cases of certifying ellipsoids and generalized cross polytopes are useful. Note that similar results are derived in [3] (see Theorem 5 for a weighted l_2 metric equivalent to a diagonal \\Sigma, and also Theorem 3) albeit in a slightly different context. A discussion on how these two relate should be included.  \n- The discussion of how to evaluate the anisotropic certificates and the formal definition of a \"superior certificate\" is appreciated. \n- The experiments are executed well.\n\nQuestions:\n- Assuming that estimate of the Lipschitz constant L is tight is the certificate provided by Theorem 1 also tight? For the special case of l_2 certification with \"\\Sigma = \\sigma^2 I\" tightness is shown in previous work. Does it hold in general?\n- Are there any benefits to training with anisotropic noise?\n\nReferences:\n1. Jordan and Dimakis. \"Exactly Computing the Local Lipschitz Constant of ReLU networks\"\n2. Sukenik et al. \"Intriguing Properties of Input-dependent Randomized Smoothing\"\n3. Yeom and Fredrikson. \"Individual Fairness Revisited: Transferring Techniques from Adversarial Robustness\"", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors extend the standard (isotropic) l_1/l_2 randomized smoothing certificate to anisotropic smoothing distributions (and anisotropic regions). They propose to maximize the volume of the certified region for each sample independently, using a memory procedure to preserve soundness.", "main_review": "The paper is well written, polished, and easy to follow. The anistropic part of the proposed approach is well-motivated, however the sample-wise part has several issues (see below) making it unsuitable to use in practice. Moreover, some of the theoretical results are not novel (see below).\n\nComments:\n- The derivation of the certificate in term of the Lipschitz constant under the dual norm (Proposition 1, Theorem 1) is insightful. However, this result is already known in the literature in greater generality and should be properly cited. Specifically, Theorem 1 in [1] derives the local (\\alpha,\\beta)-Lipschitz constant over an open set. The special case of a scalar function is also discussed (see Eq. 3 in [1]) and it coincides with the result in Proposition 1 (if we are in the special case where the open set is R). The discussion of how to use such a Lipschitz constant for certification is also discussed in [1] (see section E.2).\n- The authors already discuss a know issue of data-dependent classifiers which when not tackled can lead to certificates that are not sound. To address the issue they adapt the memory-based procedure introduced in Alfarra et al. While this procedure does make the certificate sound it has other problems.\n  - First, by introducing memory the proposed anisotropic certificate (or in fact any method to compute a certified region) becomes completely irrelevant. Namely, for sample i construct any region R_i and define that the prediction in that region is some arbitrary C_i. Then run the post-processing certification step in Algorithm 1 (or similarly the procedure from Alfarra et al.). Since the obtained regions will not intersect with any certified regions in the memory M the procedure is still sound. This is true regardless of how the region R_i was constructed.\n  - A second, and more important, issue is that the memory makes the certificate dependent on the order of the incoming test samples. This provides a new avenue for attack, i.e. the adversary can optimize the order of the test samples to decrease the utility of the final obtained smoothed classifier.\n  - Finally, the success of this memory approach also somewhat depends on the \"sparsity\" of the test samples. Namely, by using a small test set since the samples are in a high-dimensional space the distance between them tends to be bigger than the (proxy) radii of the certified regions. However, in a real-world application we are likely to have many more test samples which would increase the number of intersections when running Algorithm 1.\n  - Note, all of these issue also apply to Alfarra et al. and are simply inherited in this approach.\n- Another issue with the proposed approach is the optimization procedure described in section C. The optimization suffers from issues such as: inconsistent estimation due to clamping and not using confidence bounds, sensitivity to initialization, high gradient variance, etc. These and other issues have been described in great detail in section C.2 in [2], focusing on Alfarra et al., although, given the similarity, the same issues are inherited in this approach.\n- Relatedly, in [2] the authors show that input-dependent randomized smoothing suffers from the curse of dimensionality. A discussion of this issue and how it applies to the proposed certificate should be included.\n- The derivations for the special cases of certifying ellipsoids and generalized cross polytopes are useful. Note that similar results are derived in [3] (see Theorem 5 for a weighted l_2 metric equivalent to a diagonal \\Sigma, and also Theorem 3) albeit in a slightly different context. A discussion on how these two relate should be included.  \n- The discussion of how to evaluate the anisotropic certificates and the formal definition of a \"superior certificate\" is appreciated. \n- The experiments are executed well.\n\nQuestions:\n- Assuming that estimate of the Lipschitz constant L is tight is the certificate provided by Theorem 1 also tight? For the special case of l_2 certification with \"\\Sigma = \\sigma^2 I\" tightness is shown in previous work. Does it hold in general?\n- Are there any benefits to training with anisotropic noise?\n\nReferences:\n1. Jordan and Dimakis. \"Exactly Computing the Local Lipschitz Constant of ReLU networks\"\n2. Sukenik et al. \"Intriguing Properties of Input-dependent Randomized Smoothing\"\n3. Yeom and Fredrikson. \"Individual Fairness Revisited: Transferring Techniques from Adversarial Robustness\"", "summary_of_the_review": "The paper is well written, polished, and easy to follow. The anistropic part of the proposed approach is well-motivated, however the sample-wise part has several issues (see main review) making it unsuitable to use in practice. Moreover, some of the theoretical results are not novel (see main review).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635863941014}, {"id": "d8HMqSY6qfR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper441/Reviewer_fXE9"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors discuss the extension of $\\ell_{p}$-randomized smoothing to anisotropic counterparts.\nIn particular, they consider the extension of $\\ell_{2}$-certificates from (hyper)spheres to (hyper)ellipsoids\nby sampling anisotropic rather than isotropic Gaussian noise, as well as the extension of $\\ell_{1}$-certificates to cross-polytopes by sampling scaled Uniform noise rather than unscaled noise.\nFurther, the authors discuss how these extended certificates can be compared to their base-counter parts to establish superiority (inclusion).\nThe introduced certification algorithm, ANCER, utilizes the idea of data-dependent randomized smoothing to find the anisotropic shape with maximal certification volume.\nIn experimental evaluation on Cifar-10 and Imagenet the authors show that the obtained certificates permit higher isotropic certificating radii than other methods in the per-sample optimization setting.", "review_text": "The paper is well written and easy to follow.\nAt the high-level the idea of anisotropic certificates is mathematical simple, yet interesting and, as shown in the evaluation, effective.\n\nThe mathematical ideas introduced in sections 4 and 5 mostly follow directly from prior work (as is acknowledged in footnote 2).\nANCER, presented in sections 6 and C, also seems straight forward and correct. However, I have concerns about the data dependent nature of the algorithm (see below).\n\nThe evaluation section is quite thorough and obtains SOTA results.\nHowever, what is completely missing is a discussion of the inference/prediction procedure.\n\nBased on this my main concerns are:\n- The usefulness of data-dependence approach (see below) and whether comparison with non-data-dependent approaches is fair.\n- The novelty of the approach: Many of the contributions are direct consequences or applications of other approaches.\n  While normally, in the light of good results, I would not mind this, in the case of this publication I am unsure, due to my first concern, how good the results really are.\n\nFurthermore, I am wondering whether ACR gains can be realized in the anisotropic setting without per-datapoint optimization (e.g. finding a anisotropic smoothing parameters offline on the training set, and subsequently using them in certification)?\n\nData Dependence:\nData-dependent randomized smoothing (DDRS) [1] has been shown to be unsound in its original form and was then patched by the addition of memory-based certification in the latest draft of [1] as well as in this paper.\nWhile the memory-based approach seems to fix the unsoundness of DDRS when applied to perturbed versions of previously certified samples, I have a few questions regarding its suitability as a practical certified defense:\n- Neither this paper nor [1] discuss inference/prediction (e.g., PREDICT in [2]). Can you elaborate on how a prediction procedure would look like, that (i) is consistent with the result of the memory-based certification and (ii) is efficient (i.e. does not need to compute the certification radius at inference time)?\n- Can you clarify whether the order in which inputs are presented can influence the output of the model? If so, would the model have to include a history of all samples, predictions, and certified radii previously seen? Consequently, would any obtained guarantees only be valid for a model with the exact same history, hence preventing parallel application?\n- Can you discuss the possibility of an attacker deliberately presenting inputs (e.g. in small pockets of one class in the decision landscape) that influence the model’s predictions on future inputs?\n\nDepending on these questions, data-dependent and non-data-dependent approaches (e.g. [2]) seem to target very different settings. If this is the case, I believe a direct comparison and the claim to be SOTA to be unfair.  To avoid confusion in the field, this difference in setting and all its implications should be made abundantly clear.\n\nFurther, while I am very open to discuss these points and, depending on the answers, raise my score, I am unsure whether this is the right place to discuss these issues as they mostly concern the key contributions of [1], which has not yet been published in a peer reviewed format, potentially due to these very questions.\n\n[1] Data Dependent Randomized Smoothing, Alfarra et al.; arXiv 2020/2021\n\n[2] Certified Adversarial Robustness via Randomized Smoothing, Cohen et al.; ICML 2019\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors discuss the extension of $\\ell_{p}$-randomized smoothing to anisotropic counterparts.\nIn particular, they consider the extension of $\\ell_{2}$-certificates from (hyper)spheres to (hyper)ellipsoids\nby sampling anisotropic rather than isotropic Gaussian noise, as well as the extension of $\\ell_{1}$-certificates to cross-polytopes by sampling scaled Uniform noise rather than unscaled noise.\nFurther, the authors discuss how these extended certificates can be compared to their base-counter parts to establish superiority (inclusion).\nThe introduced certification algorithm, ANCER, utilizes the idea of data-dependent randomized smoothing to find the anisotropic shape with maximal certification volume.\nIn experimental evaluation on Cifar-10 and Imagenet the authors show that the obtained certificates permit higher isotropic certificating radii than other methods in the per-sample optimization setting.", "main_review": "The paper is well written and easy to follow.\nAt the high-level the idea of anisotropic certificates is mathematical simple, yet interesting and, as shown in the evaluation, effective.\n\nThe mathematical ideas introduced in sections 4 and 5 mostly follow directly from prior work (as is acknowledged in footnote 2).\nANCER, presented in sections 6 and C, also seems straight forward and correct. However, I have concerns about the data dependent nature of the algorithm (see below).\n\nThe evaluation section is quite thorough and obtains SOTA results.\nHowever, what is completely missing is a discussion of the inference/prediction procedure.\n\nBased on this my main concerns are:\n- The usefulness of data-dependence approach (see below) and whether comparison with non-data-dependent approaches is fair.\n- The novelty of the approach: Many of the contributions are direct consequences or applications of other approaches.\n  While normally, in the light of good results, I would not mind this, in the case of this publication I am unsure, due to my first concern, how good the results really are.\n\nFurthermore, I am wondering whether ACR gains can be realized in the anisotropic setting without per-datapoint optimization (e.g. finding a anisotropic smoothing parameters offline on the training set, and subsequently using them in certification)?\n\nData Dependence:\nData-dependent randomized smoothing (DDRS) [1] has been shown to be unsound in its original form and was then patched by the addition of memory-based certification in the latest draft of [1] as well as in this paper.\nWhile the memory-based approach seems to fix the unsoundness of DDRS when applied to perturbed versions of previously certified samples, I have a few questions regarding its suitability as a practical certified defense:\n- Neither this paper nor [1] discuss inference/prediction (e.g., PREDICT in [2]). Can you elaborate on how a prediction procedure would look like, that (i) is consistent with the result of the memory-based certification and (ii) is efficient (i.e. does not need to compute the certification radius at inference time)?\n- Can you clarify whether the order in which inputs are presented can influence the output of the model? If so, would the model have to include a history of all samples, predictions, and certified radii previously seen? Consequently, would any obtained guarantees only be valid for a model with the exact same history, hence preventing parallel application?\n- Can you discuss the possibility of an attacker deliberately presenting inputs (e.g. in small pockets of one class in the decision landscape) that influence the model’s predictions on future inputs?\n\nDepending on these questions, data-dependent and non-data-dependent approaches (e.g. [2]) seem to target very different settings. If this is the case, I believe a direct comparison and the claim to be SOTA to be unfair.  To avoid confusion in the field, this difference in setting and all its implications should be made abundantly clear.\n\nFurther, while I am very open to discuss these points and, depending on the answers, raise my score, I am unsure whether this is the right place to discuss these issues as they mostly concern the key contributions of [1], which has not yet been published in a peer reviewed format, potentially due to these very questions.\n\n[1] Data Dependent Randomized Smoothing, Alfarra et al.; arXiv 2020/2021\n\n[2] Certified Adversarial Robustness via Randomized Smoothing, Cohen et al.; ICML 2019\n\n", "summary_of_the_review": "The paper presents anisotropic robustness certificates, which seems to be technically correct and conceptually interesting.\nWhile not stellar I believe the novelty and motivation are sufficient, given the results.\nHowever, my reception of the results, and thereby the overall paper hinges on the data-dependent optimization procedure considered in the paper, of which I am currently not convinced.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635812105883}, {"id": "IY8-Z0T0hGh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper441/Reviewer_Fomp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors provide a technique for a data-dependent randomized smoothing that provides anisotropic certificates of robustness. This may be viewed as an extension of the work of Alfarra et al, where the certificates provided are axis-aligned cross-polytopes or ellipsoids; however surprisingly the proposed approach provides even tighter bounds than Alfarra's in the isotropic setting.", "review_text": "**Strengths:**\n\nRandomized smoothing, being the most scalable of robustness certification techniques, is an important research direction, and this work significantly advances the state of the art, even in the isotropic setting. The proposed technique is clearly explained and the analysis to justify this approach is simpler than that of Cohen or Yang. While the problem setting is slightly different than prior arts (anisotropic vs isotropic), a natural and fair quantitative comparison approach is described. The most significant and surprising result of this work: that these isotropic certificates outperform even Alfarra's approach is given ample discussion, and empirically-supported justification. The experiments are thorough and fairly compare against prior work.\n\n**Weaknesses:**\n\nA primary pitfall of this work is that anisotropic certificates are not particularly well-motivated. The community-accepted threat model for adversarial attackers is typically isotropic, and probing the shape of the decision boundary is of dubious utility. If this were a primary goal, more efforts would be paid to considering anisotropic certificates that are not axis-aligned (i.e. Sigma and Lambda are not diagonal). Another downside to this work is that it relies heavily upon the approach of Alfarra. Importantly, this also means AnCer inherits the main inelegance of Alfarra's approach: the memory-based classifier. This raises serious concerns in that the certificates provided are dependent upon the order in which the model is queried. While this is perhaps borderline not-kosher, it has not been a concern in practice in prior works, but it would be nice to see some evidence of this claim here.\n\n**Questions:**\n\n- How does Ancer perform for non axis-aligned anisotropic regions? Searching over orthogonal matrices (for the rotation) in addition to the diagonal anisotropic components could drastically increase the number of optimization variables as well as complicate the optimization landscape. How does this affect performance in terms of both runtime and reported certificates?\n- Are there examples for which the memory bank of the memory-based classifier is leveraged? In particular, how does the minimum distance between a pair of cross-class test data points compare to the bounds provided? \n- What insights does AnCer tell us about the shape of the decision boundary? ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors provide a technique for a data-dependent randomized smoothing that provides anisotropic certificates of robustness. This may be viewed as an extension of the work of Alfarra et al, where the certificates provided are axis-aligned cross-polytopes or ellipsoids; however surprisingly the proposed approach provides even tighter bounds than Alfarra's in the isotropic setting.", "main_review": "**Strengths:**\n\nRandomized smoothing, being the most scalable of robustness certification techniques, is an important research direction, and this work significantly advances the state of the art, even in the isotropic setting. The proposed technique is clearly explained and the analysis to justify this approach is simpler than that of Cohen or Yang. While the problem setting is slightly different than prior arts (anisotropic vs isotropic), a natural and fair quantitative comparison approach is described. The most significant and surprising result of this work: that these isotropic certificates outperform even Alfarra's approach is given ample discussion, and empirically-supported justification. The experiments are thorough and fairly compare against prior work.\n\n**Weaknesses:**\n\nA primary pitfall of this work is that anisotropic certificates are not particularly well-motivated. The community-accepted threat model for adversarial attackers is typically isotropic, and probing the shape of the decision boundary is of dubious utility. If this were a primary goal, more efforts would be paid to considering anisotropic certificates that are not axis-aligned (i.e. Sigma and Lambda are not diagonal). Another downside to this work is that it relies heavily upon the approach of Alfarra. Importantly, this also means AnCer inherits the main inelegance of Alfarra's approach: the memory-based classifier. This raises serious concerns in that the certificates provided are dependent upon the order in which the model is queried. While this is perhaps borderline not-kosher, it has not been a concern in practice in prior works, but it would be nice to see some evidence of this claim here.\n\n**Questions:**\n\n- How does Ancer perform for non axis-aligned anisotropic regions? Searching over orthogonal matrices (for the rotation) in addition to the diagonal anisotropic components could drastically increase the number of optimization variables as well as complicate the optimization landscape. How does this affect performance in terms of both runtime and reported certificates?\n- Are there examples for which the memory bank of the memory-based classifier is leveraged? In particular, how does the minimum distance between a pair of cross-class test data points compare to the bounds provided? \n- What insights does AnCer tell us about the shape of the decision boundary? ", "summary_of_the_review": "This approach advances the state of the art in an important line of work. A slightly nonstandard problem setting is considered, but even when comparing to the standard problem, AnCer improves upon prior works. The analysis is clean, the results surprising, and the experiments are thorough. I have no qualms recommending acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635518865052}], "openreview_url": "https://openreview.net/forum?id=UFYYol-bRq", "arxiv_id": "2107.04570", "paper_pdf": "papers/UFYYol-bRq.pdf", "paper_pdf_sha256": "be7e9b329ebbd8c439ef1f81e3e4f0137de709fa6817e36a2a855dc472af25d4", "paper_pdf_bytes": 5273045, "paper_pdf_source": "openreview", "code_url": "https://github.com/MotasemAlfarra/ANCER", "code_repository": "MotasemAlfarra/ANCER", "code_commit": "98f869e924c482ac246c2245ae3b247e08c7e04b", "code_archive": "repos/UFYYol-bRq.zip", "code_archive_sha256": "5c6a45368fd08960b3750c07b16eccb551b730fda282cf253222b049e909a1fb", "code_archive_bytes": 270942, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 2392, "github_languages": {"Python": 32651}, "github_archived": false, "github_pushed_at": "2022-09-09T08:45:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ancer-anisotropic-certification-via-sample"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Q1aiM7sCi1", "year": 2021, "status": "rejected", "title": "Fuzzy c-Means Clustering for Persistence Diagrams", "authors": ["Thomas Davies", "Jack Aspinall", "Bryan Wilder", "Long Tran-Thanh"], "authorids": ["~Thomas_Davies1", "jack.aspinall@materials.ox.ac.uk", "~Bryan_Wilder1", "long.tran-thanh@warwick.ac.uk"], "authors_source": "OpenReview API", "abstract": "Persistence diagrams concisely represent the topology of a point cloud whilst having strong theoretical guarantees. Most current approaches to integrating topological information into machine learning implicitly map persistence diagrams to a Hilbert space, resulting in deformation of the underlying metric structure whilst also generally requiring prior knowledge about the true topology of the space. In this paper we give an algorithm for Fuzzy c-Means (FCM) clustering directly on the space of persistence diagrams, enabling unsupervised learning that automatically captures the topological structure of data, with no prior knowledge or additional processing of persistence diagrams. We prove the same convergence guarantees as traditional FCM clustering: every convergent subsequence of iterates tends to a local minimum or saddle point. We end by presenting experiments where our fuzzy topological clustering algorithm allows for unsupervised top-$k$ candidate selection in settings where (i) the properties of persistence diagrams make them the natural choice over geometric equivalents, and (ii) the probabilistic membership values let us rank candidates in settings where verifying candidate suitability is expensive: lattice structure classification in materials science and pre-trained model selection in machine learning.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "6wTpzgLuFL2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper890/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a new clustering algorithm for persistence diagrams. They use fuzzy c means clustering. The motivation for fuzzy clustering is that it allows each datum (persistence diagram) to have weighted (soft) membership in different clusters. The partial membership value is the ratio of the distance to that cluster center to sum of all memberships to other clusters. Empirical results on synthetic and real data shows that clustering of persistence diagrams outperforms others that depends on geometry. A convergence theorem is provided, based on previous fuzzy k-means convergence proof.\n\nThis paper is well written. The clustering of persistence diagrams is an important practical problem and the proposed algorithm does seem to work.\n\nHowever, overall the benefit of the new soft-membership clustering is not clear. The authors claim that (Lacombe et al. NeurIPS 2018) is hard clustering. But indeed (Lacombe et al. 18) relaxed the diagram to a continuous function. So it is unclear why it cannot have some of the benefits of the proposed method. Also the following paper needs to be discussed.\n\nPersistence Bag-of-Words for Topological Data Analysis - IJCAI'19\n\nEmpirically, some important tables should be included in the main paper. It is difficult to find the results in the supplemental material. The comparison to hard membership is hard to find. Also maybe Lacombe et al should be compared with? \n\nFurthermore, the real world data experiments is very specially chosen to show that when geometric transformation is involved, topology-based method is better than geometry-dependent baselines. But this is kind of obvious. And there are definitely better adaptations of these baselines to be more robust to these geometric transformations (e.g., using ICP algorithm to compute distance for WBC). I suggest the authors look for better dataset (ideally datasets used by supervised methods) and show that PH based clustering algorithm is useful in these datasets.\n\nI think the bottomline to me is that the key ideas are really not quite surprising. \n\n------------\n\nI have read the authors' response. I do not think the answers addressed my concerns.\n\nOn the positive side, I do think the paper is written well and the idea is clear. It is true that a soft-clustering of PDs on the true Wasserstein distance has not been done. The paper also provided proof of the convergence of the algorithm.\n\nMy main concerns are the following:\n\nFirst and most importantly, the method is not particularly surprising. It basically combines the most classic soft-clustering algorithm and the Frechet mean computation algorithm of Turner et al. Theoretically, the proof is extending the convergence result of the soft-clustering algorithm to the Wasserstein distance of PDs. I would not say the proof is trivial. But it is not that surprising, as we already knew that the Frechet mean of PDs is computable. I would be much more excited if the theoretical result is about the optimality rather than just convergence.\n\nEmpirically, the paper did not provide comparison with (Latombe et al. 18). Just saying that they did not do it exactly in the PD space is not good enough. A lot of practically powerful methods (persistence image, various kernels, etc) are approximations/relaxations outside the PD space. These approximations/relaxations can bring computational advantage, and sometimes better learning efficiency. Therefore, we need to know how this method is compared with (Latombe et al. 18) in efficiency and clustering performance. (Latombe et al. 18) can naturally have both hard- and soft-clustering versions. A thorough comparison with the different versions can show how important it is to stick with the PD space rather than the relaxation. My guess is that in practice sticking with PD space is not that important, or maybe even worse due to bad local optima of the Frechet mean. But I would be very happy to be proven wrong.\n\nAnother issue is the limited experiments. The material data does seem to be a good fit. But the authors could use some of the classic topology-friendly data (shape, dynamic data, graph) from existing supervised methods. Any labeled binary/multiclass data can be used to evaluate clustering. An even more ambitious goal is to prove the usefulness of clustering in the supervised task. For example, the authors can show that a bag-of-words approach (using the clustering result) can improve classification performance.\n\nOverall, I feel that the methodology is not very exciting to me, and the experiments are insufficient. If the main argument is the algorithm computes on the PD space and the proof of convergence, this paper may better fit a theoretical conference.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Not Quite Exciting", "review": "The paper proposes a new clustering algorithm for persistence diagrams. They use fuzzy c means clustering. The motivation for fuzzy clustering is that it allows each datum (persistence diagram) to have weighted (soft) membership in different clusters. The partial membership value is the ratio of the distance to that cluster center to sum of all memberships to other clusters. Empirical results on synthetic and real data shows that clustering of persistence diagrams outperforms others that depends on geometry. A convergence theorem is provided, based on previous fuzzy k-means convergence proof.\n\nThis paper is well written. The clustering of persistence diagrams is an important practical problem and the proposed algorithm does seem to work.\n\nHowever, overall the benefit of the new soft-membership clustering is not clear. The authors claim that (Lacombe et al. NeurIPS 2018) is hard clustering. But indeed (Lacombe et al. 18) relaxed the diagram to a continuous function. So it is unclear why it cannot have some of the benefits of the proposed method. Also the following paper needs to be discussed.\n\nPersistence Bag-of-Words for Topological Data Analysis - IJCAI'19\n\nEmpirically, some important tables should be included in the main paper. It is difficult to find the results in the supplemental material. The comparison to hard membership is hard to find. Also maybe Lacombe et al should be compared with? \n\nFurthermore, the real world data experiments is very specially chosen to show that when geometric transformation is involved, topology-based method is better than geometry-dependent baselines. But this is kind of obvious. And there are definitely better adaptations of these baselines to be more robust to these geometric transformations (e.g., using ICP algorithm to compute distance for WBC). I suggest the authors look for better dataset (ideally datasets used by supervised methods) and show that PH based clustering algorithm is useful in these datasets.\n\nI think the bottomline to me is that the key ideas are really not quite surprising. \n\n------------\n\nI have read the authors' response. I do not think the answers addressed my concerns.\n\nOn the positive side, I do think the paper is written well and the idea is clear. It is true that a soft-clustering of PDs on the true Wasserstein distance has not been done. The paper also provided proof of the convergence of the algorithm.\n\nMy main concerns are the following:\n\nFirst and most importantly, the method is not particularly surprising. It basically combines the most classic soft-clustering algorithm and the Frechet mean computation algorithm of Turner et al. Theoretically, the proof is extending the convergence result of the soft-clustering algorithm to the Wasserstein distance of PDs. I would not say the proof is trivial. But it is not that surprising, as we already knew that the Frechet mean of PDs is computable. I would be much more excited if the theoretical result is about the optimality rather than just convergence.\n\nEmpirically, the paper did not provide comparison with (Latombe et al. 18). Just saying that they did not do it exactly in the PD space is not good enough. A lot of practically powerful methods (persistence image, various kernels, etc) are approximations/relaxations outside the PD space. These approximations/relaxations can bring computational advantage, and sometimes better learning efficiency. Therefore, we need to know how this method is compared with (Latombe et al. 18) in efficiency and clustering performance. (Latombe et al. 18) can naturally have both hard- and soft-clustering versions. A thorough comparison with the different versions can show how important it is to stick with the PD space rather than the relaxation. My guess is that in practice sticking with PD space is not that important, or maybe even worse due to bad local optima of the Frechet mean. But I would be very happy to be proven wrong.\n\nAnother issue is the limited experiments. The material data does seem to be a good fit. But the authors could use some of the classic topology-friendly data (shape, dynamic data, graph) from existing supervised methods. Any labeled binary/multiclass data can be used to evaluate clustering. An even more ambitious goal is to prove the usefulness of clustering in the supervised task. For example, the authors can show that a bag-of-words approach (using the clustering result) can improve classification performance.\n\nOverall, I feel that the methodology is not very exciting to me, and the experiments are insufficient. If the main argument is the algorithm computes on the PD space and the proof of convergence, this paper may better fit a theoretical conference.\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604034433996}, {"id": "WEdl0dkAmP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper890/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims to apply Fuzzy c-Means (FCM) clustering to persistence diagrams and prove convergent subsequence of iterates tends to a local minimum or saddle point. The motivation of the work is direct and clear. This paper addresses the problem of persistence diagram clustering via using  weighted Frechet mean . \n\nThis is a incremental work though replacing Euclidean distance in FCM by Wasserstein distance. The contributions in this work are not quite promising. For instance, the weighted Frechet mean in Sec3.2 is the well-known Wasserstein Barycenter whose behavior is well studied in optimal transport works. Thus Theorem2 can be not be considered as the contribution of this work. \n\nAnother drawback is that experiment in this work is very weak. There are only three dataset tested in this work. It's not quite convincing. It would be promising if the proposed work valid in other shape datasets such as SHREC2010 or SHREC2014. These datasets were frequently used for testing algorithms of topological data analysis .\n\nI cannot recognize the merits of this work compared with previous papers, such as: Large Scale computation of Means and Clusters for Persistence Diagrams using Optimal Transport.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This is a incremental work to cluster persistence diagrams", "review": "This paper aims to apply Fuzzy c-Means (FCM) clustering to persistence diagrams and prove convergent subsequence of iterates tends to a local minimum or saddle point. The motivation of the work is direct and clear. This paper addresses the problem of persistence diagram clustering via using  weighted Frechet mean . \n\nThis is a incremental work though replacing Euclidean distance in FCM by Wasserstein distance. The contributions in this work are not quite promising. For instance, the weighted Frechet mean in Sec3.2 is the well-known Wasserstein Barycenter whose behavior is well studied in optimal transport works. Thus Theorem2 can be not be considered as the contribution of this work. \n\nAnother drawback is that experiment in this work is very weak. There are only three dataset tested in this work. It's not quite convincing. It would be promising if the proposed work valid in other shape datasets such as SHREC2010 or SHREC2014. These datasets were frequently used for testing algorithms of topological data analysis .\n\nI cannot recognize the merits of this work compared with previous papers, such as: Large Scale computation of Means and Clusters for Persistence Diagrams using Optimal Transport.\n\n", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603952852016}, {"id": "CWomj3cJi_2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper890/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Synopsis of the paper\n\nThis paper develops a novel algorithm for performing fuzzy clustering\n(i.e. non-hard assignment of points to cluster centres) on persistence\ndiagrams, i.e. topological data descriptors. This is a highly relevant\ncontribution because persistence diagrams 'live' in a space that makes\nmetric analyses somewhat cumbersome. By contrast, the proposed method,\neven though this is not strictly highlighted in the paper, can be used\nas a principled way to obtain 'representatives' of a data set.\n\nThe critical algorithmic insight of the paper lies in developing a new\nway to calculate Fréchet means; this makes it possible to adapt FCM to\nthis domain.\n\nA set of experiments demonstrates the utility of the proposed approach.\n\n# Summary of the review\n\nThis is a well-written paper with a highly relevant contribution for\nthe TDA community. I am excited to see such a clustering algorithm\nfinally emerge for persistence diagrams, and I envision that this paper\nwill be a very useful contribution to the field.\n\nThat being said, there are some issues in the current write-up that\nprevent me from fully endorsing this work for now, namely:\n\n1. Presentation: the paper will be somewhat confusing for non-experts in\n   TDA. While this is to be expected to some extent, there are several\n   places in which the paper could be improved to provide some more\n   intuition, making it possible that even non-experts can appreciate\n   the contribution.\n\n2. The experiments appear to be somewhat preliminary. The experiment\n   with synthetic data, for example, only comprises few samples that\n   are clustered; a large portion of this section is spent on discussing\n   an application in materials science instead (which is of course\n   important, but I feel that it is hard to both appreciate the\n   application domain and the algorithm at the same time). In addition,\n   some details about the empirical behaviour of the method are not\n   discussed.\n\nIf these two points were to be rectified in a revision of the paper,\nit would help the contribution to shine more. Please see below for more\ndetailed comments.\n\n# Detailed comments\n\n- In terms of exposition, I would suggest to cite the Vietoris--Rips\n  complex construction (and also refer to the complex by this\n  alternative name; I think VR is much more common that just 'Rips')\n\n- Would it not be easier to show a Vietoris--Rips filtration instead of\n  the Čech one? I would suggest updating Figure 1 to account for this.\n  To provide a more intuitive view to TDA, the caption of the figure\n  could also be extended to describe the creation of cycles in the\n  point cloud, for example. This is the first figure that readers will\n  see, so any updates are well worth the effort, in my opinion. \n\n- Footnote 2 needs some clarification: is the paper suggesting to\n  compute the full VR complex (until the full simplex of $n$ vertices\n  has been reached)? If not, there *will* be multiple points at\n  infinity.\n\n- Would it not be possible to sidestep the issue of infinite points\n  entirely by using, say, extended persistence? It is my understanding\n  that the method does not 'care' about the way the persistence diagrams\n  are calculated, right? Hence, there are no structural assumptions\n  being made here.\n\n- The Fréchet mean is not necessary unique. Does this pose any problems\n  for the algorithm? I would assume that the calculated clustering might\n  also not be unique (or one of multiple equivalent solutions), but\n  I lack the intuition here. This should be briefly discussed.\n\n- Before equation 2, it should be 'Fréchet mean $\\widehat{\\mathbb{D}}$',\n  i.e. the variable indicating the mean should be used here. \n\n- How is the convergence behaviour of the algorithm? How many\n  iterations does it usually take until clusters start to stabilise? \n\n- The section on 'Computing the Fréchet mean' could be improved in terms\n  of the flow. Would it be possible to provide an overview algorithm as\n  well?\n\n- In terms of the limitations of the method, it is my understanding that\n  it heavily depends on the Wasserstein distance. Is this correct? I am\n  asking because this distance is known to be computationally\n  challenging to compute, so I was wondering while reading the paper\n  whether a similar algorithm could be derived for *kernels* between\n  persistence diagrams. While the experiments discuss runtime already in\n  the supplements, I am not convinced about the overall scalability of\n  the algorithm. (this is not a fault of the paper, I merely want the\n  limitations to be discussed in more details)\n\n- As already mentioned above, I would suggest updating the discussion of\n  the lattice structure data if possible. I feel that this is an\n  interesting topic, but I would rather see more experiments in the\n  paper and a shortened description of the background (it could be put\n  into the supplemental section).\n\n- As an additional suggestion for improving the experiments, I would\n  suggest running the synthetic test data experiments with more\n  diagrams. This *can* be an excellent introductory experiment to\n  showcase the capabilities of the algorithm, but at present, it falls\n  slightly short of that.\n\n- To me, the 'decision boundaries' section could be extended. This is\n  a really exciting application; the ability to find representatives \n  of diagrams opens up all kinds of new avenues! Is it possible to link\n  this more to previous results, i.e. Ramamurthy et al.?\n\n- As for the discussion on generalisation capabilities, I would suggest\n  citing prior work (Rieck et al., 'Neural Persistence: A Complexity\n  Measure for Deep Neural Networks Using Algebraic Topology'), which\n  mentions relationships between topology-based measures and\n  generalisation capabilities.\n\nAll in all, I am convinced that this has the potential to be a strong\naddition to the TDA community!\n\n# Style & clarity\n\nThe paper is well-written; there are a few sentences that I failed to\nparse correctly, though:\n\n- 'invariance to the basis symmetries': should this be 'basic\n  symmetries' instead? Moreover, why is there an ellipsis (...) after\n  'physics'? Should this be '[...]' to indicate that parts of the\n  quotation were left out?\n\n- As a matter of personal style preference, I would prefer to say 'The\n  algorithm by Turner et al.' rather than \"Turner et al.'s algorithm\".\n  The latter strikes me as somewhat confusing.\n\n- I would suggest to use no contractions in formal writing, hence\n  \"cannot\" instead of \"can't\" etc.; this is a minor point, but it since\n  the remainder of the paper is written so neatly, I cannot help but\n  point out ways to improve it even more.\n\n- 'so does not' --> 'so it does not'\n\n# Update after rebuttal & discussions\n\nI thank the authors for their thorough rebuttal. While the technical details are acknowledged and addressed for the most part, the experimental setup could still be improved. R1 mentioned that the work by Lacombe et al. might also be applicable as a comparison partner. Investing in a more thorough scenario would strengthen the paper by a lot.\n\n# Further update after discussions\n\nThe primary subject of our discussions concerned the experimental setup. While I still see this paper favourably, it would be strengthened by a more in-depth comparison with the work by Lacombe et al. The core of the paper would be more convincing if the utility of the fuzzy clustering could be highlighted better in a set of scenarios that are more comparable with existing TDA literature.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good contribution that finally enables clustering of topological descriptors", "review": "# Synopsis of the paper\n\nThis paper develops a novel algorithm for performing fuzzy clustering\n(i.e. non-hard assignment of points to cluster centres) on persistence\ndiagrams, i.e. topological data descriptors. This is a highly relevant\ncontribution because persistence diagrams 'live' in a space that makes\nmetric analyses somewhat cumbersome. By contrast, the proposed method,\neven though this is not strictly highlighted in the paper, can be used\nas a principled way to obtain 'representatives' of a data set.\n\nThe critical algorithmic insight of the paper lies in developing a new\nway to calculate Fréchet means; this makes it possible to adapt FCM to\nthis domain.\n\nA set of experiments demonstrates the utility of the proposed approach.\n\n# Summary of the review\n\nThis is a well-written paper with a highly relevant contribution for\nthe TDA community. I am excited to see such a clustering algorithm\nfinally emerge for persistence diagrams, and I envision that this paper\nwill be a very useful contribution to the field.\n\nThat being said, there are some issues in the current write-up that\nprevent me from fully endorsing this work for now, namely:\n\n1. Presentation: the paper will be somewhat confusing for non-experts in\n   TDA. While this is to be expected to some extent, there are several\n   places in which the paper could be improved to provide some more\n   intuition, making it possible that even non-experts can appreciate\n   the contribution.\n\n2. The experiments appear to be somewhat preliminary. The experiment\n   with synthetic data, for example, only comprises few samples that\n   are clustered; a large portion of this section is spent on discussing\n   an application in materials science instead (which is of course\n   important, but I feel that it is hard to both appreciate the\n   application domain and the algorithm at the same time). In addition,\n   some details about the empirical behaviour of the method are not\n   discussed.\n\nIf these two points were to be rectified in a revision of the paper,\nit would help the contribution to shine more. Please see below for more\ndetailed comments.\n\n# Detailed comments\n\n- In terms of exposition, I would suggest to cite the Vietoris--Rips\n  complex construction (and also refer to the complex by this\n  alternative name; I think VR is much more common that just 'Rips')\n\n- Would it not be easier to show a Vietoris--Rips filtration instead of\n  the Čech one? I would suggest updating Figure 1 to account for this.\n  To provide a more intuitive view to TDA, the caption of the figure\n  could also be extended to describe the creation of cycles in the\n  point cloud, for example. This is the first figure that readers will\n  see, so any updates are well worth the effort, in my opinion. \n\n- Footnote 2 needs some clarification: is the paper suggesting to\n  compute the full VR complex (until the full simplex of $n$ vertices\n  has been reached)? If not, there *will* be multiple points at\n  infinity.\n\n- Would it not be possible to sidestep the issue of infinite points\n  entirely by using, say, extended persistence? It is my understanding\n  that the method does not 'care' about the way the persistence diagrams\n  are calculated, right? Hence, there are no structural assumptions\n  being made here.\n\n- The Fréchet mean is not necessary unique. Does this pose any problems\n  for the algorithm? I would assume that the calculated clustering might\n  also not be unique (or one of multiple equivalent solutions), but\n  I lack the intuition here. This should be briefly discussed.\n\n- Before equation 2, it should be 'Fréchet mean $\\widehat{\\mathbb{D}}$',\n  i.e. the variable indicating the mean should be used here. \n\n- How is the convergence behaviour of the algorithm? How many\n  iterations does it usually take until clusters start to stabilise? \n\n- The section on 'Computing the Fréchet mean' could be improved in terms\n  of the flow. Would it be possible to provide an overview algorithm as\n  well?\n\n- In terms of the limitations of the method, it is my understanding that\n  it heavily depends on the Wasserstein distance. Is this correct? I am\n  asking because this distance is known to be computationally\n  challenging to compute, so I was wondering while reading the paper\n  whether a similar algorithm could be derived for *kernels* between\n  persistence diagrams. While the experiments discuss runtime already in\n  the supplements, I am not convinced about the overall scalability of\n  the algorithm. (this is not a fault of the paper, I merely want the\n  limitations to be discussed in more details)\n\n- As already mentioned above, I would suggest updating the discussion of\n  the lattice structure data if possible. I feel that this is an\n  interesting topic, but I would rather see more experiments in the\n  paper and a shortened description of the background (it could be put\n  into the supplemental section).\n\n- As an additional suggestion for improving the experiments, I would\n  suggest running the synthetic test data experiments with more\n  diagrams. This *can* be an excellent introductory experiment to\n  showcase the capabilities of the algorithm, but at present, it falls\n  slightly short of that.\n\n- To me, the 'decision boundaries' section could be extended. This is\n  a really exciting application; the ability to find representatives \n  of diagrams opens up all kinds of new avenues! Is it possible to link\n  this more to previous results, i.e. Ramamurthy et al.?\n\n- As for the discussion on generalisation capabilities, I would suggest\n  citing prior work (Rieck et al., 'Neural Persistence: A Complexity\n  Measure for Deep Neural Networks Using Algebraic Topology'), which\n  mentions relationships between topology-based measures and\n  generalisation capabilities.\n\nAll in all, I am convinced that this has the potential to be a strong\naddition to the TDA community!\n\n# Style & clarity\n\nThe paper is well-written; there are a few sentences that I failed to\nparse correctly, though:\n\n- 'invariance to the basis symmetries': should this be 'basic\n  symmetries' instead? Moreover, why is there an ellipsis (...) after\n  'physics'? Should this be '[...]' to indicate that parts of the\n  quotation were left out?\n\n- As a matter of personal style preference, I would prefer to say 'The\n  algorithm by Turner et al.' rather than \"Turner et al.'s algorithm\".\n  The latter strikes me as somewhat confusing.\n\n- I would suggest to use no contractions in formal writing, hence\n  \"cannot\" instead of \"can't\" etc.; this is a minor point, but it since\n  the remainder of the paper is written so neatly, I cannot help but\n  point out ways to improve it even more.\n\n- 'so does not' --> 'so it does not'\n\n# Update after rebuttal & discussions\n\nI thank the authors for their thorough rebuttal. While the technical details are acknowledged and addressed for the most part, the experimental setup could still be improved. R1 mentioned that the work by Lacombe et al. might also be applicable as a comparison partner. Investing in a more thorough scenario would strengthen the paper by a lot.\n\n# Further update after discussions\n\nThe primary subject of our discussions concerned the experimental setup. While I still see this paper favourably, it would be strengthened by a more in-depth comparison with the work by Lacombe et al. The core of the paper would be more convincing if the utility of the fuzzy clustering could be highlighted better in a set of scenarios that are more comparable with existing TDA literature.", "rating": "6: Marginally above acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603882123991}, {"id": "6iofO9Cr1e", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper890/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a novel algorithm for the fuzzy clustering of persistence diagrams. To determine cluster centroids, the Wasserstein-2 distance is used to minimize the weighted Fr\\’echet mean between a potential cluster center and all PDs considered for clustering. The authors proof convergence of the clustering algorithm and conduct experiments on 1) synthetic data, 2) lattice structures, and 3) decision boundaries of neural networks. In the last experiment is was shown that models whose PDs cluster close to the PDs of a given task lead to higher classification performance than random classifiers, demonstrating the merit of PDs as a useful tool for model selection. \n\nThe paper is very well written and clearly presented. Even though I did not check all proofs in the appendix, it seems like a mathematically sound approach that leads to good clustering results. The contributions and the relevance of the topic are apparent; however, I think the experimental section is weaker than it could be. While I like the experiments on the synthetic data set and the decision boundaries, I think the weakest part of the paper are the experiments on the lattice structures. Respective paragraphs are packed with technical details which are not crucial to understand the problem at hand. The results are very qualitative and outsourced to the appendix. Furthermore, Figure 3 is interesting to look at and to get a better understand of the structural differences of the data; however, it would be nice to have visual representations of the results of 4.2. For example, would it be possible to choose $c=4$ and find the four different categories that you show in Figure 3 (by selecting 10-20 structures from each category and run your algorithm on them)? You could visualize a representative of each category (like in the current figure) and their respective PDs (just plotted on top of each other for each category). In a third panel, it would be interesting to see a representation of the centroids and all PDs using Multidimensional Scaling based on the W2 distance. I would also recommend using a more quantitative way of evaluating cluster quality (e.g. [1]).\nA second weakness I see in the experiments is that in the introduction, you mention previous work on fuzzy discrete distribution clustering (by de Carvalho et al.). Still, you do not compare your results to their method (does adding the diagonal make such a big difference?).  \n\nMinor comments and questions:\n\n•\tRelated work: When you mention the correlation of topological complexity with generalization ability, you miss the work by Rieck et al. [2].\n\n•\tI am a fan of citing software libraries when they provide how to in their website. Hence I would recommend citing Ripser [3] (see https://github.com/Ripser/ripser for BibTeX entry) and potentially other Software you used in the main paper.\n\n•\tYou say “[…] the vectorization […] required by Lacombe et al.’s algorithm makes it unsuitable for integration into our work”. Could you elaborate on this? Can you not even choose this as a comparison partner?\n\n•\tWhat do you mean when you say other persistence-based learning strategies required prior knowledge of a ‘correct’ target topology which can’t plausibly be known? The work by Moor et al. for example, extracts persistence directly from the input, which is “trivial” to know since you can compute it. \n\n•\tTopological Preliminaries: Why do you introduce the \\v{C}ech complex when you use Vietoris-Rips for your experiments? \n\n•\tOne feat of your approach is that you add tuples on the diagonal to make the cardinality of all compared PDs match. I wonder how susceptible your approach is to increasingly different cardinality (that has to be filled). Maybe you can conduct a small ablation study on your synthetic data set and evaluate how “good” your mean PDs are when cardinalities are increasingly different.\n \n[1]: RJGB Campello, ER Hruschka. A fuzzy extension of the silhouette width criterion for cluster analysis. Fuzzy Sets and Systems, 2006 – Elsevier\n\n[2]: Bastian Rieck, Matteo Togninalli, Christian Bock, Michael Moor, Max Horn, Thomas Gumbsch, Karsten Borgwardt. Neural Persistence: A Complexity Measure for Deep Neural Networks Using Algebraic Topology. ICLR 2019. \n\n[3]: Ulrich Bauer. Ripser: efficient computation of Vietoris-Rips persistence barcodes. Preprint.\n\n======================\nUpdate: Thank you for your rebuttal, I still think this is a promising paper and lean towards acceptance. However, after the discussion and the update of the manuscript, my score will remain the same.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Sound approach with weaknesses in the experiments", "review": "The authors propose a novel algorithm for the fuzzy clustering of persistence diagrams. To determine cluster centroids, the Wasserstein-2 distance is used to minimize the weighted Fr\\’echet mean between a potential cluster center and all PDs considered for clustering. The authors proof convergence of the clustering algorithm and conduct experiments on 1) synthetic data, 2) lattice structures, and 3) decision boundaries of neural networks. In the last experiment is was shown that models whose PDs cluster close to the PDs of a given task lead to higher classification performance than random classifiers, demonstrating the merit of PDs as a useful tool for model selection. \n\nThe paper is very well written and clearly presented. Even though I did not check all proofs in the appendix, it seems like a mathematically sound approach that leads to good clustering results. The contributions and the relevance of the topic are apparent; however, I think the experimental section is weaker than it could be. While I like the experiments on the synthetic data set and the decision boundaries, I think the weakest part of the paper are the experiments on the lattice structures. Respective paragraphs are packed with technical details which are not crucial to understand the problem at hand. The results are very qualitative and outsourced to the appendix. Furthermore, Figure 3 is interesting to look at and to get a better understand of the structural differences of the data; however, it would be nice to have visual representations of the results of 4.2. For example, would it be possible to choose $c=4$ and find the four different categories that you show in Figure 3 (by selecting 10-20 structures from each category and run your algorithm on them)? You could visualize a representative of each category (like in the current figure) and their respective PDs (just plotted on top of each other for each category). In a third panel, it would be interesting to see a representation of the centroids and all PDs using Multidimensional Scaling based on the W2 distance. I would also recommend using a more quantitative way of evaluating cluster quality (e.g. [1]).\nA second weakness I see in the experiments is that in the introduction, you mention previous work on fuzzy discrete distribution clustering (by de Carvalho et al.). Still, you do not compare your results to their method (does adding the diagonal make such a big difference?).  \n\nMinor comments and questions:\n\n•\tRelated work: When you mention the correlation of topological complexity with generalization ability, you miss the work by Rieck et al. [2].\n\n•\tI am a fan of citing software libraries when they provide how to in their website. Hence I would recommend citing Ripser [3] (see https://github.com/Ripser/ripser for BibTeX entry) and potentially other Software you used in the main paper.\n\n•\tYou say “[…] the vectorization […] required by Lacombe et al.’s algorithm makes it unsuitable for integration into our work”. Could you elaborate on this? Can you not even choose this as a comparison partner?\n\n•\tWhat do you mean when you say other persistence-based learning strategies required prior knowledge of a ‘correct’ target topology which can’t plausibly be known? The work by Moor et al. for example, extracts persistence directly from the input, which is “trivial” to know since you can compute it. \n\n•\tTopological Preliminaries: Why do you introduce the \\v{C}ech complex when you use Vietoris-Rips for your experiments? \n\n•\tOne feat of your approach is that you add tuples on the diagonal to make the cardinality of all compared PDs match. I wonder how susceptible your approach is to increasingly different cardinality (that has to be filled). Maybe you can conduct a small ablation study on your synthetic data set and evaluate how “good” your mean PDs are when cardinalities are increasingly different.\n \n[1]: RJGB Campello, ER Hruschka. A fuzzy extension of the silhouette width criterion for cluster analysis. Fuzzy Sets and Systems, 2006 – Elsevier\n\n[2]: Bastian Rieck, Matteo Togninalli, Christian Bock, Michael Moor, Max Horn, Thomas Gumbsch, Karsten Borgwardt. Neural Persistence: A Complexity Measure for Deep Neural Networks Using Algebraic Topology. ICLR 2019. \n\n[3]: Ulrich Bauer. Ripser: efficient computation of Vietoris-Rips persistence barcodes. Preprint.\n\n======================\nUpdate: Thank you for your rebuttal, I still think this is a promising paper and lean towards acceptance. However, after the discussion and the update of the manuscript, my score will remain the same.  \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603792338911}], "openreview_url": "https://openreview.net/forum?id=Q1aiM7sCi1", "arxiv_id": "2006.02796", "paper_pdf": "papers/Q1aiM7sCi1.pdf", "paper_pdf_sha256": "c91cae8d7bf52f22210719cd3a4a51a0415206c53211909163e45557211892b3", "paper_pdf_bytes": 1655816, "paper_pdf_source": "openreview", "code_url": "https://github.com/tomogwen/fpdcluster", "code_repository": "tomogwen/fpdcluster", "code_commit": "8bcd7c01e6120893a545aaaac1e73d793b6210f0", "code_archive": "repos/Q1aiM7sCi1.zip", "code_archive_sha256": "28e21f59c6dc8f3325aab16023e01ec5da4744b6049cffdda912cb48f27e2b18", "code_archive_bytes": 1023682, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 1534, "github_languages": {"D2": 311660, "Python": 27322}, "github_archived": false, "github_pushed_at": "2023-10-14T12:00:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fuzzy-c-means-clustering-for-persistence"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mAutPdnHIN", "year": 2026, "status": "rejected", "title": "ASyMOB: Algebraic Symbolic Mathematical Operations Benchmark", "authors": ["Michael Shalyt", "Rotem Elimelech", "Ido Kaminer"], "authorids": ["~Michael_Shalyt1", "~Rotem_Elimelech1", "~Ido_Kaminer1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) are increasingly applied to symbolic mathematics, yet existing evaluations often conflate pattern memorization with genuine reasoning. To address this gap, we present **ASyMOB**, a high-resolution dataset of **35,368** validated symbolic math problems spanning integration, limits, differential equations, series, and hypergeometrics. Unlike prior benchmarks, **ASyMOB** systematically perturbs each seed problem using symbolic, numeric, and equivalence-preserving transformations, enabling a fine-grained assessment of generalization and robustness.\nOur evaluation reveals three key findings: (1) most models’ performance collapses under minor perturbations, while frontier systems exhibit substantial robustness, suggesting an emerging *\"phase transition\"* from memorization to generalization; (2) integrated code tools stabilize performance, particularly for weaker models; and (3) we identify examples where Computer Algebra Systems (CAS) fail while LLMs succeed, as well as problems solved only via a hybrid LLM-CAS approach, highlighting a promising integration frontier.\n**ASyMOB** serves as a principled diagnostic tool for measuring and accelerating progress toward building verifiable, trustworthy AI for scientific discovery.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "qHEH2wurku", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22638/Reviewer_ScbR"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This work introduces ASyMOB, a 35,368 problem benchmark for symbolic mathematics (integration, limits, differential equations, series, etc.) built to test symbolic reasoning by systematically perturbing seed problems and then validating answers via equivalence-aware \nsymbolic and numeric checks. For every seed problem, they obtain variants of controlled difficulty through three methods: 1) symbolic perturbations, which perturb N symbols; numeric variants, which replace all or exactly one symbol with N-digit integers; and equivalence variants, which insert identities equal to 1 (trigonometric, hyperbolic, logarithmic, complex-exponential, and series). They further note that the dataset can be re-generated before assessing a new LLM, making their dataset more resilient against benchmark hacking or memorization. Finally, they find that 1) models' performance drops even under small perturbations, with frontier models being more robust, 2) incorporating code tools stabilizes performance, especially for weaker models, and 3) they observe instances where CAS fails but LLMs succeed and problems that can only be solved by combining both.", "review_text": "This work introduces ASyMOB, a 35,368 problem benchmark for symbolic mathematics (integration, limits, differential equations, series, etc.) built to test symbolic reasoning by systematically perturbing seed problems and then validating answers via equivalence-aware \nsymbolic and numeric checks. For every seed problem, they obtain variants of controlled difficulty through three methods: 1) symbolic perturbations, which perturb N symbols; numeric variants, which replace all or exactly one symbol with N-digit integers; and equivalence variants, which insert identities equal to 1 (trigonometric, hyperbolic, logarithmic, complex-exponential, and series). They further note that the dataset can be re-generated before assessing a new LLM, making their dataset more resilient against benchmark hacking or memorization. Finally, they find that 1) models' performance drops even under small perturbations, with frontier models being more robust, 2) incorporating code tools stabilizes performance, especially for weaker models, and 3) they observe instances where CAS fails but LLMs succeed and problems that can only be solved by combining both.", "strengths": "The work clearly explains how problems are built and expanded into symbolic, numeric, and equivalence variants, with worked examples for each. \n* ASyMOB fills gaps in existing literature dataset, targeting symbolic manipulation (integration, limits, DEs, series, hypergeometrics) rather than text-to-math. It offers controlled difficulty via systematic perturbations and broad university-level problem coverage that previous benchmarks lack. \n* Dataset instances are created with random transforms, and the dataset can be re-generated before evaluation. This design reduces leakage/memorization risk compared to static test sets. \n* The work illustrates that models’ performance degrades sharply under small perturbations, while frontier models are more robust. Tool use helps weaker models, and hybrid LLM + CAS solves cases where either alone fails.", "weaknesses": "* The work documents qualitative examples where CAS fails but LLMs succeed, and a case solvable only by an LLM + CAS hybrid (Figure 6). Further, it argues that symbolics hurt CAS more than LLMs. What’s missing is a dataset-level percentage/table partitioning successes into LLM-only, CAS-only, and hybrid categories across perturbations. Adding this would substantively strengthen the claim.\n* Some of the perturbations appear to be somewhat contrived. This may not necessarily be a bad thing, but it could mean that the benchmark is testing perturbations that would not reasonably appear in the real world. This is related to my first question below.", "questions": "* Does performance on ASyMOB correlate with performance on other real-world tasks? E.g., performance on math contests that occurred after model releases.\n* Would training/fine-tuning on ASyMOB perturbations lead to stronger models?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces ASyMOB, a 35,368 problem benchmark for symbolic mathematics (integration, limits, differential equations, series, etc.) built to test symbolic reasoning by systematically perturbing seed problems and then validating answers via equivalence-aware \nsymbolic and numeric checks. For every seed problem, they obtain variants of controlled difficulty through three methods: 1) symbolic perturbations, which perturb N symbols; numeric variants, which replace all or exactly one symbol with N-digit integers; and equivalence variants, which insert identities equal to 1 (trigonometric, hyperbolic, logarithmic, complex-exponential, and series). They further note that the dataset can be re-generated before assessing a new LLM, making their dataset more resilient against benchmark hacking or memorization. Finally, they find that 1) models' performance drops even under small perturbations, with frontier models being more robust, 2) incorporating code tools stabilizes performance, especially for weaker models, and 3) they observe instances where CAS fails but LLMs succeed and problems that can only be solved by combining both.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The work clearly explains how problems are built and expanded into symbolic, numeric, and equivalence variants, with worked examples for each. \n* ASyMOB fills gaps in existing literature dataset, targeting symbolic manipulation (integration, limits, DEs, series, hypergeometrics) rather than text-to-math. It offers controlled difficulty via systematic perturbations and broad university-level problem coverage that previous benchmarks lack. \n* Dataset instances are created with random transforms, and the dataset can be re-generated before evaluation. This design reduces leakage/memorization risk compared to static test sets. \n* The work illustrates that models’ performance degrades sharply under small perturbations, while frontier models are more robust. Tool use helps weaker models, and hybrid LLM + CAS solves cases where either alone fails.", "weaknesses": "* The work documents qualitative examples where CAS fails but LLMs succeed, and a case solvable only by an LLM + CAS hybrid (Figure 6). Further, it argues that symbolics hurt CAS more than LLMs. What’s missing is a dataset-level percentage/table partitioning successes into LLM-only, CAS-only, and hybrid categories across perturbations. Adding this would substantively strengthen the claim.\n* Some of the perturbations appear to be somewhat contrived. This may not necessarily be a bad thing, but it could mean that the benchmark is testing perturbations that would not reasonably appear in the real world. This is related to my first question below.", "questions": "* Does performance on ASyMOB correlate with performance on other real-world tasks? E.g., performance on math contests that occurred after model releases.\n* Would training/fine-tuning on ASyMOB perturbations lead to stronger models?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761994169801}, {"id": "lkbg58tJwT", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22638/Reviewer_eDb2"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper presents ASyMOB (Algebraic Symbolic Mathematical Operations Benchmark), a large-scale benchmark for evaluating large language models (LLMs) on symbolic mathematical reasoning. ASyMOB contains 35,368 verified problems generated from 100 seed questions through systematic symbolic, numeric, and equivalence-preserving perturbations, covering integration, limits, differential equations, series, and hypergeometric functions. These perturbations maintain mathematical equivalence while testing robustness to structural and numerical variation. The authors evaluate multiple open- and closed-weight LLMs (e.g., GPT-4o, Gemini 2.5, DeepSeek-R1) and find substantial performance degradation under minor perturbations (average accuracy drops from 74.6% to 46.8%).  However, frontier models show notably higher robustness, suggesting an emerging transition from memorization to true symbolic generalization. Additionally, ASyMOB reveals that LLMs sometimes outperform traditional CAS systems, and that hybrid LLM + CAS approaches can solve problems unsolvable by either alone.", "review_text": "This paper presents ASyMOB (Algebraic Symbolic Mathematical Operations Benchmark), a large-scale benchmark for evaluating large language models (LLMs) on symbolic mathematical reasoning. ASyMOB contains 35,368 verified problems generated from 100 seed questions through systematic symbolic, numeric, and equivalence-preserving perturbations, covering integration, limits, differential equations, series, and hypergeometric functions. These perturbations maintain mathematical equivalence while testing robustness to structural and numerical variation. The authors evaluate multiple open- and closed-weight LLMs (e.g., GPT-4o, Gemini 2.5, DeepSeek-R1) and find substantial performance degradation under minor perturbations (average accuracy drops from 74.6% to 46.8%).  However, frontier models show notably higher robustness, suggesting an emerging transition from memorization to true symbolic generalization. Additionally, ASyMOB reveals that LLMs sometimes outperform traditional CAS systems, and that hybrid LLM + CAS approaches can solve problems unsolvable by either alone.", "strengths": "- ASyMOB isolates symbolic mathematical reasoning from linguistic understanding, providing a clean test of algebraic manipulation skills.\n- The symbolic, numeric, and equivalence perturbations enable fine-grained evaluation of robustness and generalization.\n- Dual symbolic–numeric verification ensures reliability, and the findings reveal meaningful trends such as a \"phase transition\" toward genuine reasoning in frontier LLMs.", "weaknesses": "- The scope is somehow limited. The benchmark focuses narrowly on algebraic operations, omitting other mathematical reasoning domains, such as geometry or proofs.\n- Some generated variants may be mathematically artificial and not representative of real-world symbolic problems.\n- Several key conclusions, such as the role of code integration and hybrid tool use in improving LLM reasoning, have already been explored in prior work on tool-augmented or agentic LLMs, making the contributions more incremental than novel.", "questions": "1. How would the proposed perturbation framework generalize to other mathematical domains, such as geometry, proofs, or word problems that involve both symbolic and linguistic reasoning?\n2. The paper interprets the robustness of frontier models as evidence of a \"phase transition\" from memorization to genuine symbolic reasoning. However, how do the authors rule out the possibility that these models simply memorize or interpolate over a much larger region of symbolic patterns, rather than performing true reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents ASyMOB (Algebraic Symbolic Mathematical Operations Benchmark), a large-scale benchmark for evaluating large language models (LLMs) on symbolic mathematical reasoning. ASyMOB contains 35,368 verified problems generated from 100 seed questions through systematic symbolic, numeric, and equivalence-preserving perturbations, covering integration, limits, differential equations, series, and hypergeometric functions. These perturbations maintain mathematical equivalence while testing robustness to structural and numerical variation. The authors evaluate multiple open- and closed-weight LLMs (e.g., GPT-4o, Gemini 2.5, DeepSeek-R1) and find substantial performance degradation under minor perturbations (average accuracy drops from 74.6% to 46.8%).  However, frontier models show notably higher robustness, suggesting an emerging transition from memorization to true symbolic generalization. Additionally, ASyMOB reveals that LLMs sometimes outperform traditional CAS systems, and that hybrid LLM + CAS approaches can solve problems unsolvable by either alone.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- ASyMOB isolates symbolic mathematical reasoning from linguistic understanding, providing a clean test of algebraic manipulation skills.\n- The symbolic, numeric, and equivalence perturbations enable fine-grained evaluation of robustness and generalization.\n- Dual symbolic–numeric verification ensures reliability, and the findings reveal meaningful trends such as a \"phase transition\" toward genuine reasoning in frontier LLMs.", "weaknesses": "- The scope is somehow limited. The benchmark focuses narrowly on algebraic operations, omitting other mathematical reasoning domains, such as geometry or proofs.\n- Some generated variants may be mathematically artificial and not representative of real-world symbolic problems.\n- Several key conclusions, such as the role of code integration and hybrid tool use in improving LLM reasoning, have already been explored in prior work on tool-augmented or agentic LLMs, making the contributions more incremental than novel.", "questions": "1. How would the proposed perturbation framework generalize to other mathematical domains, such as geometry, proofs, or word problems that involve both symbolic and linguistic reasoning?\n2. The paper interprets the robustness of frontier models as evidence of a \"phase transition\" from memorization to genuine symbolic reasoning. However, how do the authors rule out the possibility that these models simply memorize or interpolate over a much larger region of symbolic patterns, rather than performing true reasoning?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761988371273}, {"id": "wWllWJcIld", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission22638/Reviewer_P3Tt"], "rating": 4, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper proposes a novel benchmark for LLM mathematical reasoning, namely ASyMOB. This benchmark perturbs existing problems using symbolic, numeric, and equivalence-preserving transformations to ensure robust evaluation. Several key insights are identified to guide the future development of mathematical LLM.", "review_text": "This paper proposes a novel benchmark for LLM mathematical reasoning, namely ASyMOB. This benchmark perturbs existing problems using symbolic, numeric, and equivalence-preserving transformations to ensure robust evaluation. Several key insights are identified to guide the future development of mathematical LLM.", "strengths": "- The benchmark is reasonable, as the symbolic and numeric versions can fully evaluate the ability of LLMs to address mathematical reasoning.\n- The provided examples are well-motivated, as identifying cases where both LLMs and symbolic systems do not perform well can help guide further research directions.", "weaknesses": "- The novelty of this paper requires further clarification. As noted at the end of this paper, GSM-Symbolic has conducted similar research and reached comparable conclusions. Therefore, it is important for the authors to clearly articulate the unique contribution and positioning of this work within the field, especially given the prior work, i.e., GSM-Symbolic. The authors should carefully clarify the difference between their benchmark and existing work. Additionally, some closely related studies are missing from the discussion, such as [1], which also employs neuro-symbolic methods for data generation. Including and discussing such relevant literature would further strengthen the paper. \n- The conclusions and findings presented in this paper are trivial. For example, the integration of code tools has already been discussed in [2], and this is not a significant discovery or a new insight within the domain of LLM reasoning. In fact, numerous studies on agentic workflows [3] have previously explored approaches that enable LLMs to utilize a variety of tools.\n- The presentation of this paper has significant room for improvement. For example, the paper ends with a lot of blank space. Some technical details are missing, such as the details of equivalence-preserving transformations, which are not clearly presented.\n\n[1] Zenan Li, Zhi Zhou, Yuan Yao, Xian Zhang, Yu-Feng Li, Chun Cao, Fan Yang, Xiaoxing Ma. Neuro-Symbolic Data Generation for Math Reasoning. NeurIPS 2024.\n\n[2] Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, Graham Neubig. PAL: Program-aided Language Models. ICML 2023.\n\n[3] Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, Chenglin Wu. AFlow: Automating Agentic Workflow Generation. ICLR 2025.", "questions": "Please refer to the weaknesses section. Additionally,\n1. The list of equivalence-preserving transformations provided in the appendix appears to be too limited. It is unclear whether this small set of transformations is sufficient to comprehensively evaluate LLM capabilities, or whether it merely identifies a few isolated cases that might temporarily \"hack\" or exploit LLM behavior.\n2. The evaluation costs associated with this benchmark should be further clarified. As a benchmark, the evaluation process should be both easy and affordable for researchers to carry out. It should also be clear whether the benchmark is sufficiently inexpensive to allow for comprehensive evaluation, or whether results obtained from evaluating only a subset of the benchmark are truly representative of overall performance.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel benchmark for LLM mathematical reasoning, namely ASyMOB. This benchmark perturbs existing problems using symbolic, numeric, and equivalence-preserving transformations to ensure robust evaluation. Several key insights are identified to guide the future development of mathematical LLM.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "- The benchmark is reasonable, as the symbolic and numeric versions can fully evaluate the ability of LLMs to address mathematical reasoning.\n- The provided examples are well-motivated, as identifying cases where both LLMs and symbolic systems do not perform well can help guide further research directions.", "weaknesses": "- The novelty of this paper requires further clarification. As noted at the end of this paper, GSM-Symbolic has conducted similar research and reached comparable conclusions. Therefore, it is important for the authors to clearly articulate the unique contribution and positioning of this work within the field, especially given the prior work, i.e., GSM-Symbolic. The authors should carefully clarify the difference between their benchmark and existing work. Additionally, some closely related studies are missing from the discussion, such as [1], which also employs neuro-symbolic methods for data generation. Including and discussing such relevant literature would further strengthen the paper. \n- The conclusions and findings presented in this paper are trivial. For example, the integration of code tools has already been discussed in [2], and this is not a significant discovery or a new insight within the domain of LLM reasoning. In fact, numerous studies on agentic workflows [3] have previously explored approaches that enable LLMs to utilize a variety of tools.\n- The presentation of this paper has significant room for improvement. For example, the paper ends with a lot of blank space. Some technical details are missing, such as the details of equivalence-preserving transformations, which are not clearly presented.\n\n[1] Zenan Li, Zhi Zhou, Yuan Yao, Xian Zhang, Yu-Feng Li, Chun Cao, Fan Yang, Xiaoxing Ma. Neuro-Symbolic Data Generation for Math Reasoning. NeurIPS 2024.\n\n[2] Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, Graham Neubig. PAL: Program-aided Language Models. ICML 2023.\n\n[3] Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, Chenglin Wu. AFlow: Automating Agentic Workflow Generation. ICLR 2025.", "questions": "Please refer to the weaknesses section. Additionally,\n1. The list of equivalence-preserving transformations provided in the appendix appears to be too limited. It is unclear whether this small set of transformations is sufficient to comprehensively evaluate LLM capabilities, or whether it merely identifies a few isolated cases that might temporarily \"hack\" or exploit LLM behavior.\n2. The evaluation costs associated with this benchmark should be further clarified. As a benchmark, the evaluation process should be both easy and affordable for researchers to carry out. It should also be clear whether the benchmark is sufficiently inexpensive to allow for comprehensive evaluation, or whether results obtained from evaluating only a subset of the benchmark are truly representative of overall performance.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761817299841}], "openreview_url": "https://openreview.net/forum?id=mAutPdnHIN", "arxiv_id": "2505.23851", "paper_pdf": "papers/mAutPdnHIN.pdf", "paper_pdf_sha256": "7b02a17294644008602dd70a7d99e4fbc8aeebd30ec0bc30a6f01d455e56915b", "paper_pdf_bytes": 899298, "paper_pdf_source": "openreview", "code_url": "https://github.com/RamanujanMachine/ASyMOB", "code_repository": "RamanujanMachine/ASyMOB", "code_commit": "172712d0a12eed39a47cd1df82f690121c57520d", "code_archive": "repos/mAutPdnHIN.zip", "code_archive_sha256": "166001cc68ff36168a2231af4f327a372b7af0235df1a7b8f571c1093f5e3337", "code_archive_bytes": 1439949, "code_file_count": 17, "code_extensions": {".py": 16, ".sql": 1}, "github_disk_usage_kb": 5714, "github_languages": {"Python": 96250}, "github_archived": false, "github_pushed_at": "2026-06-08T23:49:01Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/asymob-algebraic-symbolic-mathematical"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0fwJMANq9P", "year": 2025, "status": "rejected", "title": "Efficient Heuristics Generation for Solving Combinatorial Optimization Problems Using Large Language Models", "authors": ["Xuan Wu", "Di Wang", "Zhiguang Cao", "Chunguo Wu", "Lijie Wen", "Chunyan Miao", "Yubin Xiao", "You Zhou"], "authorids": ["~Xuan_Wu7", "~Di_Wang26", "~Zhiguang_Cao1", "~Chunguo_Wu1", "~Lijie_Wen1", "~Chunyan_Miao1", "~Yubin_Xiao1", "~You_Zhou5"], "authors_source": "OpenReview API", "abstract": "Recent studies exploited Large Language Models (LLMs) to autonomously generate heuristics for solving Combinatorial Optimization Problems (COPs), by prompting LLMs to first provide search directions and then derive heuristics accordingly. However, the absence of task-specific knowledge in prompts often leads LLMs to provide unspecific search directions, obstructing the derivation of well-performing heuristics. Moreover, evaluating the derived heuristics remains resource-intensive, especially for those semantically equivalent ones, often requiring unnecessary resource expenditure. To enable LLMs to provide specific search directions, we propose the Hercules algorithm, which leverages our designed Core Abstraction Prompting (CAP) method to abstract the core components from elite heuristics and incorporate them as prior knowledge in prompts. We theoretically prove the effectiveness of CAP in reducing unspecificity and provide empirical results in this work. To reduce the required computing resources for evaluating the derived heuristics, we propose few-shot Performance Prediction Prompting (PPP), a first-of-its-kind method for the Heuristic Generation (HG) task. PPP leverages LLMs to predict the fitness values of newly derived heuristics by analyzing their semantic similarity to previously evaluated ones. We further develop two tailored mechanisms for PPP to enhance predictive accuracy and determine unreliable predictions, respectively. The use of PPP makes Hercules more resource-efficient and we name this variant Hercules-P. Extensive experiments across various HG tasks, COPs, and LLMs demonstrate that Hercules outperforms the state-of-the-art LLM-based HG algorithms, while Hercules-P excels at minimizing computing resources. In addition, we illustrate the effectiveness of CAP, PPP, and the other proposed mechanisms by conducting relevant ablation studies.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "kK1xaevCZP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12808/Reviewer_jJ2Y"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper presents Hercules, an LLM-based algorithm for generating heuristics for combinatorial problems. The paper seems to extend the framework in Ye et al., 2024 with a more advanced direction generation (based on identifying core components in heuristics) as well as an LLM-based fitness calculation. The experiments show gains over the baselines.", "review_text": "The paper presents Hercules, an LLM-based algorithm for generating heuristics for combinatorial problems. The paper seems to extend the framework in Ye et al., 2024 with a more advanced direction generation (based on identifying core components in heuristics) as well as an LLM-based fitness calculation. The experiments show gains over the baselines.", "strengths": "Strengths:\n- The topic is interesting and of recent interest\n- The approach (CAP and PPP) seems novel.\n- The experiments show significant gains over the baselines in deriving penalty heuristics for guided local search, as well as more moderate gains on constructive heuristics for TSP, heuristic measures for ant colony optimization, and reshaping of attention scores in neural combinatorial optimization,", "weaknesses": "Weaknesses:\n- I found the claim about information gain to be quite confusing. \n\t- First, a lot of information is missing: why the number of core components corresponds to the number of heuristics (can we not have multiple core components per heuristic or the same core component in multiple heuristics)? why do we assume that the set of all possible directions can be partitioned into mutually exclusive subsets that correspond to components (can we not have the same direction for multiple core components)?\n\t- Second, it is really not clear why the information gain means we get better heuristics (as indicated in lines 284-285)? If the components generated are of low-quality the directions may be of lower quality as well.\n\n- Experimental evaluation:\n\t- It is not clear what is being reported under gain: the definition is based on \"the performance of ...\" but it is not clear how performance is measured.\n\n- Writing: the writing could improve as a lot of information is not clearly presented. For example there are no clear definitions for a range of terms like parent heuristics, elite heuristics, etc.\n\n- The paper does not provide significant insight into the impact of the proposed techniques (CAP and PPP) beyond the experimental results. For example, it would be interesting to show an analysis of the correlation between predicted fitness values and quality of heuristics.", "questions": "I would appreciate the authors response and clarification on the points listed under \"weaknesses\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents Hercules, an LLM-based algorithm for generating heuristics for combinatorial problems. The paper seems to extend the framework in Ye et al., 2024 with a more advanced direction generation (based on identifying core components in heuristics) as well as an LLM-based fitness calculation. The experiments show gains over the baselines.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "Strengths:\n- The topic is interesting and of recent interest\n- The approach (CAP and PPP) seems novel.\n- The experiments show significant gains over the baselines in deriving penalty heuristics for guided local search, as well as more moderate gains on constructive heuristics for TSP, heuristic measures for ant colony optimization, and reshaping of attention scores in neural combinatorial optimization,", "weaknesses": "Weaknesses:\n- I found the claim about information gain to be quite confusing. \n\t- First, a lot of information is missing: why the number of core components corresponds to the number of heuristics (can we not have multiple core components per heuristic or the same core component in multiple heuristics)? why do we assume that the set of all possible directions can be partitioned into mutually exclusive subsets that correspond to components (can we not have the same direction for multiple core components)?\n\t- Second, it is really not clear why the information gain means we get better heuristics (as indicated in lines 284-285)? If the components generated are of low-quality the directions may be of lower quality as well.\n\n- Experimental evaluation:\n\t- It is not clear what is being reported under gain: the definition is based on \"the performance of ...\" but it is not clear how performance is measured.\n\n- Writing: the writing could improve as a lot of information is not clearly presented. For example there are no clear definitions for a range of terms like parent heuristics, elite heuristics, etc.\n\n- The paper does not provide significant insight into the impact of the proposed techniques (CAP and PPP) beyond the experimental results. For example, it would be interesting to show an analysis of the correlation between predicted fitness values and quality of heuristics.", "questions": "I would appreciate the authors response and clarification on the points listed under \"weaknesses\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730685424040}, {"id": "scFxa8difI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12808/Reviewer_KGdN"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper studies the generation of heuristics for combinatorial optimization problems using LLMs. \n\nThe work continues similar work in this space that tries to mimic evolutionary computation (crossover, mutation) via LLMs. The result is an LLM-infused metaheuristic algorithm. \n\nUnlike the previous work, the paper claims 1) to address introducing more problem specificity into the prompts and 2) to speed up the process by using LLMs to predict the performance of generated accuracies to skip over evaluating them fully. \n\nOverall, I enjoyed reading this paper, and I appreciated the work that went into building an end-to-end pipeline with several components.", "review_text": "This paper studies the generation of heuristics for combinatorial optimization problems using LLMs. \n\nThe work continues similar work in this space that tries to mimic evolutionary computation (crossover, mutation) via LLMs. The result is an LLM-infused metaheuristic algorithm. \n\nUnlike the previous work, the paper claims 1) to address introducing more problem specificity into the prompts and 2) to speed up the process by using LLMs to predict the performance of generated accuracies to skip over evaluating them fully. \n\nOverall, I enjoyed reading this paper, and I appreciated the work that went into building an end-to-end pipeline with several components.", "strengths": "Integrating LLMs with heuristic solving is an exciting combination. \nThe paper implements an end-to-end pipeline that starts with a seed query that then mimics evolutionary computing via LLMs, yielding heuristics that can be embedded in the Local Search Meta-Heuristics for different combinatorial problems. \nThe connection with Information Gain is an excellent addition\nFrom a practical perspective, the paper considers several details into account such as reducing costs via LLM predictors.", "weaknesses": "As rightly noted in the paper, the idea of mimicking evolutionary computation via LLMs is not new. In fact, most (all?) crossover and mutation operators are from Ye et. al. 2024. On the one hand, the experiments and the ablation study show that the proposed modifications might offer some benefit in the results, and on the other hand, they can be regarded as incremental, and it is not clear what's the main takeaway. \n\nRegarding the presentation, I found it difficult/confusing that many moving parts are introduced as large components with several acronyms Hercules, CAP, PPP, EXEMPLAR, Cons --but after all, the provided pseudocode shows the overall algorithm, so I am not sure what these abstractions add to the presentation. Also, the paper claims our \"propriety\" CAP algorithm  --what does that mean? \n\nThe idea of adding more specificity to the prompts seems reasonable at a high level, but the paper overindexes too much into the example in Figure 1. The information gain analysis is interesting (and is borrowed from a Hu et. al. 2024) but at the end what happens is we select top-k core components. And that's also not uniform, we do that only some number of iterations (denoted by \\lambda in the paper), all of which remain as more hyper-parameters to deal with.\n\nThe experiments cover TSP, CVPR, Binpacking, Multi-Knapsacks. Importantly, the starting seed function seems critical to the approach. The method generates heuristics but the overall approach to solve these problems are meta-heuristics.  (please correct me if I understand this correctly). For TSP, we use guided local search. For BinPacking and Knapsacks we use Ant-Colony Optimization. One might argue that the settings of the outer meta-heuristics and their performance are crucial to the overall results and not just the heuristics (generated by LLMs here.) The experiments do not discuss or study any of this. \n\nAdditionally, all comparisons are with other LLM-based heuristics generations. Note that this is quite a costly approach (hence some effort with performance predictors to save time etc.). According to the tables in the appendix, we are consuming many many minutes upto 5 hours. Then, it is not clear to me how to fairly evaluate these results. How does the same GLS and ACO without the advanced heuristics found by LLM but with standard heuristics perform given the same amount of time? (Btw, does this time include LLM queries or only running the heuristics after the LLM generates them against the instances?) \n\nThis might not be surprising that the choice of LLM quite affects the results (Table 1; LLama vs GPT-4o). But then it makes one wonder how much of the value comes from the many moving components proposed here vs. plain and simple, the underlying LLM.", "questions": "Could you provide details on the outer meta-heuristics (GLS, ACO etc.)? How much of the results are due to the LLM integration with CAP, PPP etc. vs. the meta-heuristics leading the search into good solutions. \nIt would be interesting to know the comparison between default GLS, ACO or even other baselines for TSP, BinPacking, MKP to position the results in this paper. As is, it is hard to evaluate the significance", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the generation of heuristics for combinatorial optimization problems using LLMs. \n\nThe work continues similar work in this space that tries to mimic evolutionary computation (crossover, mutation) via LLMs. The result is an LLM-infused metaheuristic algorithm. \n\nUnlike the previous work, the paper claims 1) to address introducing more problem specificity into the prompts and 2) to speed up the process by using LLMs to predict the performance of generated accuracies to skip over evaluating them fully. \n\nOverall, I enjoyed reading this paper, and I appreciated the work that went into building an end-to-end pipeline with several components.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "Integrating LLMs with heuristic solving is an exciting combination. \nThe paper implements an end-to-end pipeline that starts with a seed query that then mimics evolutionary computing via LLMs, yielding heuristics that can be embedded in the Local Search Meta-Heuristics for different combinatorial problems. \nThe connection with Information Gain is an excellent addition\nFrom a practical perspective, the paper considers several details into account such as reducing costs via LLM predictors.", "weaknesses": "As rightly noted in the paper, the idea of mimicking evolutionary computation via LLMs is not new. In fact, most (all?) crossover and mutation operators are from Ye et. al. 2024. On the one hand, the experiments and the ablation study show that the proposed modifications might offer some benefit in the results, and on the other hand, they can be regarded as incremental, and it is not clear what's the main takeaway. \n\nRegarding the presentation, I found it difficult/confusing that many moving parts are introduced as large components with several acronyms Hercules, CAP, PPP, EXEMPLAR, Cons --but after all, the provided pseudocode shows the overall algorithm, so I am not sure what these abstractions add to the presentation. Also, the paper claims our \"propriety\" CAP algorithm  --what does that mean? \n\nThe idea of adding more specificity to the prompts seems reasonable at a high level, but the paper overindexes too much into the example in Figure 1. The information gain analysis is interesting (and is borrowed from a Hu et. al. 2024) but at the end what happens is we select top-k core components. And that's also not uniform, we do that only some number of iterations (denoted by \\lambda in the paper), all of which remain as more hyper-parameters to deal with.\n\nThe experiments cover TSP, CVPR, Binpacking, Multi-Knapsacks. Importantly, the starting seed function seems critical to the approach. The method generates heuristics but the overall approach to solve these problems are meta-heuristics.  (please correct me if I understand this correctly). For TSP, we use guided local search. For BinPacking and Knapsacks we use Ant-Colony Optimization. One might argue that the settings of the outer meta-heuristics and their performance are crucial to the overall results and not just the heuristics (generated by LLMs here.) The experiments do not discuss or study any of this. \n\nAdditionally, all comparisons are with other LLM-based heuristics generations. Note that this is quite a costly approach (hence some effort with performance predictors to save time etc.). According to the tables in the appendix, we are consuming many many minutes upto 5 hours. Then, it is not clear to me how to fairly evaluate these results. How does the same GLS and ACO without the advanced heuristics found by LLM but with standard heuristics perform given the same amount of time? (Btw, does this time include LLM queries or only running the heuristics after the LLM generates them against the instances?) \n\nThis might not be surprising that the choice of LLM quite affects the results (Table 1; LLama vs GPT-4o). But then it makes one wonder how much of the value comes from the many moving components proposed here vs. plain and simple, the underlying LLM.", "questions": "Could you provide details on the outer meta-heuristics (GLS, ACO etc.)? How much of the results are due to the LLM integration with CAP, PPP etc. vs. the meta-heuristics leading the search into good solutions. \nIt would be interesting to know the comparison between default GLS, ACO or even other baselines for TSP, BinPacking, MKP to position the results in this paper. As is, it is hard to evaluate the significance", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730680576517}, {"id": "uT18DbPALb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12808/Reviewer_QjjZ"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper proposes a framework to use LLMs to generate heuristics for solving\noptimization problems. The authors describe their framework and evaluate it\nempirically, comparing to other approaches in the literature.", "review_text": "The paper proposes a framework to use LLMs to generate heuristics for solving\noptimization problems. The authors describe their framework and evaluate it\nempirically, comparing to other approaches in the literature.", "strengths": "The proposed framework is interesting and seems to work well in practice.", "weaknesses": "The choice of KGLS as seed heuristics should be justified as it was not designed\nfor the general TSP. Why not LKH? This should also be considered in the\nempirical evaluation; in particular to answer the question of whether KGLS is a\nreasonable heuristic to start with in this case (improving over a weak heuristic\nis easier than improving over a strong heuristic).\n\nFigure 5 has no axis labels.", "questions": "The differences are small in some cases and it would be great if the authors could\nprovide error bounds or confidence intervals for the empirical results.\n\nWhy is KGLS is reasonable base heuristic?\n\nUpdate after responses: Thank you for your responses. I have updated my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a framework to use LLMs to generate heuristics for solving\noptimization problems. The authors describe their framework and evaluate it\nempirically, comparing to other approaches in the literature.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The proposed framework is interesting and seems to work well in practice.", "weaknesses": "The choice of KGLS as seed heuristics should be justified as it was not designed\nfor the general TSP. Why not LKH? This should also be considered in the\nempirical evaluation; in particular to answer the question of whether KGLS is a\nreasonable heuristic to start with in this case (improving over a weak heuristic\nis easier than improving over a strong heuristic).\n\nFigure 5 has no axis labels.", "questions": "The differences are small in some cases and it would be great if the authors could\nprovide error bounds or confidence intervals for the empirical results.\n\nWhy is KGLS is reasonable base heuristic?\n\nUpdate after responses: Thank you for your responses. I have updated my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730484503148}, {"id": "6mkWdBMzCJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12808/Reviewer_f4Hd"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper explores the application of LLM in autonomously generating heuristics to solve COPs and proposes a novel algorithm named Hercules to address the two main challenges of existing approaches.\n\nHercules utilizes the Core Abstraction Prompting (CAP) method to abstract core components from elite heuristics and incorporate them as prior knowledge in prompts, thereby reducing the specificity of search directions. This paper further introduces Hercules-P, an efficient variant of Hercules that integrates CAP with the novel Performance Prediction Prompting (PPP) method. PPP leverages LLMs to predict the fitness values of newly derived heuristics based on their semantic similarity to previously evaluated ones, significantly reducing the required computing resources.", "review_text": "This paper explores the application of LLM in autonomously generating heuristics to solve COPs and proposes a novel algorithm named Hercules to address the two main challenges of existing approaches.\n\nHercules utilizes the Core Abstraction Prompting (CAP) method to abstract core components from elite heuristics and incorporate them as prior knowledge in prompts, thereby reducing the specificity of search directions. This paper further introduces Hercules-P, an efficient variant of Hercules that integrates CAP with the novel Performance Prediction Prompting (PPP) method. PPP leverages LLMs to predict the fitness values of newly derived heuristics based on their semantic similarity to previously evaluated ones, significantly reducing the required computing resources.", "strengths": "1. The paper is laudable for its well-structured and logical presentation, providing a comprehensive understanding of the research topic.\n\n2. The article is praiseworthy for its extensive experimental data and significant findings. The authors have selected a number of baselines for comparative experiments on different benchmarks.\n\n3. The supplement provided in this article is adequate. It explains in detail for the reader what is not expanded in detail in the paper, including specific experimental data, hyperparameter settings, Critical Difference Analysis, etc.", "weaknesses": "1. In the literature of TSP and CVRP, it is known that those conventional heuristic algorithms, such as LKH [1] and EAX [2], exhibit robust performance. It appears, however, that this submission does not address LKH and EAX, nor does it provide a comparative analysis of the proposed algorithm against these established methods.\n\n1. In lines 85-100 of the Introduction section, the authors describe two challenges to LLM-based HG methods, mentioning in the second challenge that these methods introduce numerous linear operations and conditional branches that make the GPU less efficient for these algorithms. In lines 113-128, the authors claim to have proposed Hercules-P in order to better address the second challenge, but I don't seem to have read in the manuscript how Hercules-P reduces linear operations and conditional branches, making GPUs more efficient in processing these algorithms. May I ask if the authors have solved this challenge? If not, these representations are inappropriate.\n\n2. Does the appearance of CVRP in Section 4.4 stand for Capacitated Vehicle Routing Problem? The authors do not explain what CVRP stands for in the body of the manuscript, and the only explanation appears in the code comments in the Appendix section (line 1337). This cannot be very clear to the reader when it comes to understanding the manuscript.\n\n3. In Section 2.3, the authors mention two challenges for NCO solvers: improving generalisation capabilities and large-scale COPs performance. In Table 5, for LEHD, the performance improvement of either Hercules or Hercules-p gradually decreases as the problem size of TSP or VCRP increases. Does this mean that Hercules also fails to address the challenges faced by NCO solvers? Can Hercules still provide performance gains when the problem size is larger? Further discussion is requested from the authors.\n\n\n## References\n[1] Keld Helsgaun. General k-opt submoves for the Lin-Kernighan TSP heuristic. Mathematical Programming Computation 1(2-3): 119-163 (2009)\n\n[2] Yuichi Nagata, Shigenobu Kobayashi. A Powerful Genetic Algorithm Using Edge Assembly Crossover for the Traveling Salesman Problem. INFORMS Journal on Computing 25(2): 346-363 (2013)", "questions": "Please reply to my comments in \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the application of LLM in autonomously generating heuristics to solve COPs and proposes a novel algorithm named Hercules to address the two main challenges of existing approaches.\n\nHercules utilizes the Core Abstraction Prompting (CAP) method to abstract core components from elite heuristics and incorporate them as prior knowledge in prompts, thereby reducing the specificity of search directions. This paper further introduces Hercules-P, an efficient variant of Hercules that integrates CAP with the novel Performance Prediction Prompting (PPP) method. PPP leverages LLMs to predict the fitness values of newly derived heuristics based on their semantic similarity to previously evaluated ones, significantly reducing the required computing resources.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper is laudable for its well-structured and logical presentation, providing a comprehensive understanding of the research topic.\n\n2. The article is praiseworthy for its extensive experimental data and significant findings. The authors have selected a number of baselines for comparative experiments on different benchmarks.\n\n3. The supplement provided in this article is adequate. It explains in detail for the reader what is not expanded in detail in the paper, including specific experimental data, hyperparameter settings, Critical Difference Analysis, etc.", "weaknesses": "1. In the literature of TSP and CVRP, it is known that those conventional heuristic algorithms, such as LKH [1] and EAX [2], exhibit robust performance. It appears, however, that this submission does not address LKH and EAX, nor does it provide a comparative analysis of the proposed algorithm against these established methods.\n\n1. In lines 85-100 of the Introduction section, the authors describe two challenges to LLM-based HG methods, mentioning in the second challenge that these methods introduce numerous linear operations and conditional branches that make the GPU less efficient for these algorithms. In lines 113-128, the authors claim to have proposed Hercules-P in order to better address the second challenge, but I don't seem to have read in the manuscript how Hercules-P reduces linear operations and conditional branches, making GPUs more efficient in processing these algorithms. May I ask if the authors have solved this challenge? If not, these representations are inappropriate.\n\n2. Does the appearance of CVRP in Section 4.4 stand for Capacitated Vehicle Routing Problem? The authors do not explain what CVRP stands for in the body of the manuscript, and the only explanation appears in the code comments in the Appendix section (line 1337). This cannot be very clear to the reader when it comes to understanding the manuscript.\n\n3. In Section 2.3, the authors mention two challenges for NCO solvers: improving generalisation capabilities and large-scale COPs performance. In Table 5, for LEHD, the performance improvement of either Hercules or Hercules-p gradually decreases as the problem size of TSP or VCRP increases. Does this mean that Hercules also fails to address the challenges faced by NCO solvers? Can Hercules still provide performance gains when the problem size is larger? Further discussion is requested from the authors.\n\n\n## References\n[1] Keld Helsgaun. General k-opt submoves for the Lin-Kernighan TSP heuristic. Mathematical Programming Computation 1(2-3): 119-163 (2009)\n\n[2] Yuichi Nagata, Shigenobu Kobayashi. A Powerful Genetic Algorithm Using Edge Assembly Crossover for the Traveling Salesman Problem. INFORMS Journal on Computing 25(2): 346-363 (2013)", "questions": "Please reply to my comments in \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "This submission does not have ethics concerns.", "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730102082668}], "openreview_url": "https://openreview.net/forum?id=0fwJMANq9P", "arxiv_id": "2505.12627", "paper_pdf": "papers/0fwJMANq9P.pdf", "paper_pdf_sha256": "3e4d0c5fb444d440dde3c0e1e1c341f5225801123bc032eefed33ffe68a628bf", "paper_pdf_bytes": 668719, "paper_pdf_source": "openreview", "code_url": "https://github.com/wuuu110/Hercules", "code_repository": "wuuu110/Hercules", "code_commit": "bc532df7325373c564e1926916c8be6bbb645216", "code_archive": "repos/0fwJMANq9P.zip", "code_archive_sha256": "5bf3ec189239dda797d2257ef7c958deff95c6afd4e9dab7d9e365332db6b612", "code_archive_bytes": 200195, "code_file_count": 51, "code_extensions": {".py": 51}, "github_disk_usage_kb": 340, "github_languages": {"Python": 361661}, "github_archived": false, "github_pushed_at": "2025-05-20T05:59:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-heuristics-generation-for-solving"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "w8eCnnq57m", "year": 2024, "status": "rejected", "title": "LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition", "authors": ["Chengsong Huang", "Qian Liu", "Bill Yuchen Lin", "Chao Du", "Tianyu Pang", "Min Lin"], "authorids": ["~Chengsong_Huang1", "~Qian_Liu2", "~Bill_Yuchen_Lin1", "~Chao_Du1", "~Tianyu_Pang1", "~Min_Lin1"], "authors_source": "OpenReview API", "abstract": "Low-rank adaptations (LoRA) are often employed to fine-tune large language models (LLMs) for new tasks. This paper investigates LoRA composability for cross-task generalization and introduces LoraHub, a simple framework devised for the purposive assembly of LoRA modules trained on diverse given tasks, with the objective of achieving adaptable performance on unseen tasks. With just a few examples from a new task, LoraHub can fluidly combine multiple LoRA modules, eliminating the need for human expertise and assumptions.  Notably, the composition requires neither additional model parameters nor gradients.  Our empirical results, derived from the Big-Bench Hard benchmark, suggest that LoraHub can effectively mimic the performance of in-context learning in few-shot scenarios, excluding the necessity of in-context examples alongside each inference input.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "WMWqjIwSXA", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission97/Reviewer_pXQC"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a method to improve the generalizability of LLMs to new tasks. The proposed method can be summarized as below:\n1. Train several models on a set of downstream tasks (one model for each task). This is done in a parameter-efficient way using the LORA method. This results in a set of LORA models. \n2. Randomly choose a subset of the LORA models.\n3. Learn a set of weights to combine these LORA models using a few (for eg, 5) samples from a new task not seen in step 1. The weights are learnt using gradient-free optimization. \n4. Use the learnt weights to generate a combination of the LORA models and combine it with the base LLMs for inference. \n\nThe result is a method that can be used to improve generalizability when the number of samples available for a new task is very low.", "review_text": "This paper proposes a method to improve the generalizability of LLMs to new tasks. The proposed method can be summarized as below:\n1. Train several models on a set of downstream tasks (one model for each task). This is done in a parameter-efficient way using the LORA method. This results in a set of LORA models. \n2. Randomly choose a subset of the LORA models.\n3. Learn a set of weights to combine these LORA models using a few (for eg, 5) samples from a new task not seen in step 1. The weights are learnt using gradient-free optimization. \n4. Use the learnt weights to generate a combination of the LORA models and combine it with the base LLMs for inference. \n\nThe result is a method that can be used to improve generalizability when the number of samples available for a new task is very low.", "strengths": "- The paper presents an interesting way to fine-tune LLMs on tasks where the number of training examples might be very low. \n- The proposed method uses gradient-free optimization to minimize resource requirements, since the number of parameters being learnt is very low.\n-The method outperforms zero-shot deployment of the base LLM.", "weaknesses": "The main weakness of the paper is its performance compared to in-context learning (ICL), as highlighted in Table. 1. The authors acknowledge this in the paper but justify by saying that their method uses fewer tokens in their fine-tuning process compared to in-context learning. However, I feel that ICL is a very straight-forward and easy way to improve generalizability and that the problem that the authors are addressing is a minor one. Further, the performance of the proposed method also weakens the paper. From a practitioner's perspective I feel that the proposed method will be less appealing to just using ICL.", "questions": "- Can the authors provide more insight into the impact of the reduced number of tokens required in the proposed method? What does that mean to a practitioner in the end? Especially considering the fact that the proposed method first needs the training of several LORA modules, the burden on a practitioner may actually be higher.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to improve the generalizability of LLMs to new tasks. The proposed method can be summarized as below:\n1. Train several models on a set of downstream tasks (one model for each task). This is done in a parameter-efficient way using the LORA method. This results in a set of LORA models. \n2. Randomly choose a subset of the LORA models.\n3. Learn a set of weights to combine these LORA models using a few (for eg, 5) samples from a new task not seen in step 1. The weights are learnt using gradient-free optimization. \n4. Use the learnt weights to generate a combination of the LORA models and combine it with the base LLMs for inference. \n\nThe result is a method that can be used to improve generalizability when the number of samples available for a new task is very low.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper presents an interesting way to fine-tune LLMs on tasks where the number of training examples might be very low. \n- The proposed method uses gradient-free optimization to minimize resource requirements, since the number of parameters being learnt is very low.\n-The method outperforms zero-shot deployment of the base LLM.", "weaknesses": "The main weakness of the paper is its performance compared to in-context learning (ICL), as highlighted in Table. 1. The authors acknowledge this in the paper but justify by saying that their method uses fewer tokens in their fine-tuning process compared to in-context learning. However, I feel that ICL is a very straight-forward and easy way to improve generalizability and that the problem that the authors are addressing is a minor one. Further, the performance of the proposed method also weakens the paper. From a practitioner's perspective I feel that the proposed method will be less appealing to just using ICL.", "questions": "- Can the authors provide more insight into the impact of the reduced number of tokens required in the proposed method? What does that mean to a practitioner in the end? Especially considering the fact that the proposed method first needs the training of several LORA modules, the burden on a practitioner may actually be higher.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699049999736}, {"id": "bdHzkKHaM0", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission97/Reviewer_ZnG8"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper tackles the task of employing `LoRA` (parameter-efficient low rank adapters) for few-shot learning of unseen tasks:: When only few samples are available for a new task, then training a new `LoRA` module on it may not work well. Instead, the paper considers the setting where several `LoRA` modules are available, pretrained on a set of upstream tasks $T_i$. Then, given a set of few samples for a new task, the proposed `LoRAHub` aims to learn the ideal \"meta-weights\" $w_i$ such that the agglomerated `LoRA` module $\\sum_i w_i LoRA_{T_i }$ yields the best performance for the new few-shot task (Note that this required that all upstream `LoRA` modules have the same rank).\n\nTo optimize the weights $w$, `LoraHub` employs a gradient-free method based on combinatorial optimization, `Shiwa`, and uses it to minimize the loss on the given few-shot samples, with an additional regularization term on the weights $w$.\n\nThe proposed method is then evaluated using a `Flan-T5` backbone on the `Big-Bench Hard` benchmark. The upstream tasks are the 200 tasks originally used to instruct `Flan-T5`; however, in practice, a random subset of 20 tasks is used in each run, such that we only have 20 weights $w_i$ to tune in each run. Overall, `LoraHub` performs almost on-par with in-context learning method, with the advantage that it requires shorter input prompts hence fewer tokens to process (essentially identical to zero-shot learning).", "review_text": "The paper tackles the task of employing `LoRA` (parameter-efficient low rank adapters) for few-shot learning of unseen tasks:: When only few samples are available for a new task, then training a new `LoRA` module on it may not work well. Instead, the paper considers the setting where several `LoRA` modules are available, pretrained on a set of upstream tasks $T_i$. Then, given a set of few samples for a new task, the proposed `LoRAHub` aims to learn the ideal \"meta-weights\" $w_i$ such that the agglomerated `LoRA` module $\\sum_i w_i LoRA_{T_i }$ yields the best performance for the new few-shot task (Note that this required that all upstream `LoRA` modules have the same rank).\n\nTo optimize the weights $w$, `LoraHub` employs a gradient-free method based on combinatorial optimization, `Shiwa`, and uses it to minimize the loss on the given few-shot samples, with an additional regularization term on the weights $w$.\n\nThe proposed method is then evaluated using a `Flan-T5` backbone on the `Big-Bench Hard` benchmark. The upstream tasks are the 200 tasks originally used to instruct `Flan-T5`; however, in practice, a random subset of 20 tasks is used in each run, such that we only have 20 weights $w_i$ to tune in each run. Overall, `LoraHub` performs almost on-par with in-context learning method, with the advantage that it requires shorter input prompts hence fewer tokens to process (essentially identical to zero-shot learning).", "strengths": "- **Interesting idea and motivation**: I find the core idea of the paper very interesting with potential applications to fields such as multi-task learning or continual learning. While there are similar concurrent idea mixing mixture-of-experts with LoRA, I found the results of the paper and focus on few-shot multi-task novel and insightful.\n\n- **Clear writing**: Overall the paper is clearly written, easy to read and understand. \n\n- **Detailled experimental analysis**: I found the experimental analysis (Section 5 in the paper) quite interesting and raises interesting properties as well as limitations of the proposed method.", "weaknesses": "- **Computational cost of optimization**: Unlike In-Context Learning, `LoraHub` does not need to process additional tokens hence a reduced inference cost. However, it also adds the cost to optimize the combination weights $w$ on the input few-shot samples, in particular when many upstream tasks are available. It would be interesting to discuss the trade-off between these two costs, e.g. say we have some few-shot samples but only want to solve the associated task once, it might be more practical to use in-context learning rather than the optimization pipeline of `LoraHub` ?\n\n- **Optimization of $w$ for many tasks and robustness of `LoRAHub`:** It's not clear to me how the optimization methods scale to a higher number of upstream LoRA modules either in terms of cost (see previous point) or performance: In **Figure 4**, we see that increasing the number of LoRA module does not always improve performance but strongly affects the variance of the outputs. This suggests that the optimization procedure is either noisy and/or does not converge well. As a consequence, it means that the number of candidate upstream LoRA modules, $N$, must be carefully selected (and the optimal $N$ even seems to be task dependent from **Figure 4**) which introduces an additional hyperparameter. This can be an important limitation for real-life applications.", "questions": "- On the topic of **optimizing $w$ for many tasks**, I am wondering if authors have considered alternative techniques which might be more robust to optimization noise (beyond the prefiltering strategy mentioned in Section 7): e.g. a hierarchical approach (optimize $w$ for multiple random subsets of 20 LoRA candidates, then learn $w'$ for these agglomerated modules) or a curriculum like approach (gradually drop some of the candidates when optimizing $w$ if they are consistently given very low weights) ?\n\n- **Question/Suggestion on Table 1**:  it is not clear to me whether Table 1 reports results averaged on 5 random seeds for *all methods* or only for `LoraHub` and whether the different random seeds impact the choice of query few-shot samples, or only optimization(e.g. initialization, and library of Lora modules). Maybe a more complete evaluation metric would be to report (avg) and (best) for all methods (or even some form of significance test) to understand how robust the other methods are to the random seed. \n\n- **Figure 3 and variance**: I think **Figure 3** would be much more convincing with error bars or some notion of variance. This figure's aim is to illustrate that \"LoRA with few samples does not work as well as LoraHub's few-shot learning\" , however the results are only available for 3 tasks, for which the assumption only holds until 20 few-shot samples;therefore it's not clear how the insight generalizes to more few-shot settings.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tackles the task of employing `LoRA` (parameter-efficient low rank adapters) for few-shot learning of unseen tasks:: When only few samples are available for a new task, then training a new `LoRA` module on it may not work well. Instead, the paper considers the setting where several `LoRA` modules are available, pretrained on a set of upstream tasks $T_i$. Then, given a set of few samples for a new task, the proposed `LoRAHub` aims to learn the ideal \"meta-weights\" $w_i$ such that the agglomerated `LoRA` module $\\sum_i w_i LoRA_{T_i }$ yields the best performance for the new few-shot task (Note that this required that all upstream `LoRA` modules have the same rank).\n\nTo optimize the weights $w$, `LoraHub` employs a gradient-free method based on combinatorial optimization, `Shiwa`, and uses it to minimize the loss on the given few-shot samples, with an additional regularization term on the weights $w$.\n\nThe proposed method is then evaluated using a `Flan-T5` backbone on the `Big-Bench Hard` benchmark. The upstream tasks are the 200 tasks originally used to instruct `Flan-T5`; however, in practice, a random subset of 20 tasks is used in each run, such that we only have 20 weights $w_i$ to tune in each run. Overall, `LoraHub` performs almost on-par with in-context learning method, with the advantage that it requires shorter input prompts hence fewer tokens to process (essentially identical to zero-shot learning).", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- **Interesting idea and motivation**: I find the core idea of the paper very interesting with potential applications to fields such as multi-task learning or continual learning. While there are similar concurrent idea mixing mixture-of-experts with LoRA, I found the results of the paper and focus on few-shot multi-task novel and insightful.\n\n- **Clear writing**: Overall the paper is clearly written, easy to read and understand. \n\n- **Detailled experimental analysis**: I found the experimental analysis (Section 5 in the paper) quite interesting and raises interesting properties as well as limitations of the proposed method.", "weaknesses": "- **Computational cost of optimization**: Unlike In-Context Learning, `LoraHub` does not need to process additional tokens hence a reduced inference cost. However, it also adds the cost to optimize the combination weights $w$ on the input few-shot samples, in particular when many upstream tasks are available. It would be interesting to discuss the trade-off between these two costs, e.g. say we have some few-shot samples but only want to solve the associated task once, it might be more practical to use in-context learning rather than the optimization pipeline of `LoraHub` ?\n\n- **Optimization of $w$ for many tasks and robustness of `LoRAHub`:** It's not clear to me how the optimization methods scale to a higher number of upstream LoRA modules either in terms of cost (see previous point) or performance: In **Figure 4**, we see that increasing the number of LoRA module does not always improve performance but strongly affects the variance of the outputs. This suggests that the optimization procedure is either noisy and/or does not converge well. As a consequence, it means that the number of candidate upstream LoRA modules, $N$, must be carefully selected (and the optimal $N$ even seems to be task dependent from **Figure 4**) which introduces an additional hyperparameter. This can be an important limitation for real-life applications.", "questions": "- On the topic of **optimizing $w$ for many tasks**, I am wondering if authors have considered alternative techniques which might be more robust to optimization noise (beyond the prefiltering strategy mentioned in Section 7): e.g. a hierarchical approach (optimize $w$ for multiple random subsets of 20 LoRA candidates, then learn $w'$ for these agglomerated modules) or a curriculum like approach (gradually drop some of the candidates when optimizing $w$ if they are consistently given very low weights) ?\n\n- **Question/Suggestion on Table 1**:  it is not clear to me whether Table 1 reports results averaged on 5 random seeds for *all methods* or only for `LoraHub` and whether the different random seeds impact the choice of query few-shot samples, or only optimization(e.g. initialization, and library of Lora modules). Maybe a more complete evaluation metric would be to report (avg) and (best) for all methods (or even some form of significance test) to understand how robust the other methods are to the random seed. \n\n- **Figure 3 and variance**: I think **Figure 3** would be much more convincing with error bars or some notion of variance. This figure's aim is to illustrate that \"LoRA with few samples does not work as well as LoraHub's few-shot learning\" , however the results are only available for 3 tasks, for which the assumption only holds until 20 few-shot samples;therefore it's not clear how the insight generalizes to more few-shot settings.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698767503462}, {"id": "ZRcRwEWGGe", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission97/Reviewer_8M9N"], "rating": "5: marginally below the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper introduces LoRAHub, a method designed to enhance performance on unseen tasks by re-utilizing trained LoRA (Low Rank Approximation) parameters across different tasks. LoRAHub operates by weighting each of these parameters, learned through a minimal set of examples on unseen tasks, thereby outperforming zero-shot baselines while achieving comparable results to in-context learning.", "review_text": "The paper introduces LoRAHub, a method designed to enhance performance on unseen tasks by re-utilizing trained LoRA (Low Rank Approximation) parameters across different tasks. LoRAHub operates by weighting each of these parameters, learned through a minimal set of examples on unseen tasks, thereby outperforming zero-shot baselines while achieving comparable results to in-context learning.", "strengths": "1. The paper presents a novel approach to leveraging previously learned LoRA parameters to improve performance on unseen tasks.\n\n2. It demonstrates competitive performance compared to in-context learning while outperforming zero-shot baselines, showcasing the potential of the method.\n\n3. It would be beneficial if the authors could release the fine-tuned LoRA weight to the community.", "weaknesses": "1. The choice of using FLAN-T5-Large as the base model is questionable as a model pre-trained on unsupervised text might have been more relevant for fine-tuning / LoRA fine-tuning on upstream tasks from the FLAN collection.\n\n2. The paper lacks clarity in explaining the rationale behind maintaining the same rank for the composed LoRA module and could benefit from exploring higher rank matrices when composing.\n\n3. The selection of 20 LoRAs for unseen tasks seems arbitrary and might limit the method’s performance. An iterative procedure or justification for this selection would have been beneficial.\n\n4. The absence of certain baselines, such as the average performance for BBH's top 5 upstream tasks, leaves gaps in the evaluation.\n\n5. The baseline corresponding to the retrieval of a trained LoRAs, when given a handful of examples from unseen tasks, is missing. For example, see https://arxiv.org/abs/2302.03202", "questions": "1. Can you elaborate on the interpretation of negative coefficients for the LoRA weights?\n\n2. Why was the decision made to maintain the same rank for the composed LoRA module? Have higher rank matrices been explored and if so, what were the findings?\n\n3. Is the selection of 20 LoRAs for unseen tasks fixed or is there an iterative procedure to this selection? How does this choice impact the method’s performance on unseen tasks?\n\n4. I believe the strength of the method lies in cases where there are a handful of examples from unseen tasks. If it's otherwise, that it is beneficial to use a larger number of examples, it would make sense to compare against methods like IA3 [https://arxiv.org/abs/2205.05638], which fine-tunes efficiently on few-shot examples from unseen tasks.\n\n5. Why was FLAN-T5-Large chosen as the base model over a model pre-trained on unsupervised text? Wouldn't it be strange to fine-tune on FLAN tasks using LoRA on an already FLAN multitask-trained backbone model?\n\n6. How does the absolute value of LoRA weight not exceeding 1.5 relate to the method's performance in section 4.2, and is there a particular significance to this threshold?\n\n7. It would be beneficial to include parameter efficient fine-tuning and traditional fine-tuning performance in Table 1, especially with the setup of a limited number of examples.\n\n8. Is In-Context Learning (ICL) performed on the same base Language Model (LLM), or is it conducted using larger decoder only LLMs? What are the implications of this choice on the comparison of results?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces LoRAHub, a method designed to enhance performance on unseen tasks by re-utilizing trained LoRA (Low Rank Approximation) parameters across different tasks. LoRAHub operates by weighting each of these parameters, learned through a minimal set of examples on unseen tasks, thereby outperforming zero-shot baselines while achieving comparable results to in-context learning.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The paper presents a novel approach to leveraging previously learned LoRA parameters to improve performance on unseen tasks.\n\n2. It demonstrates competitive performance compared to in-context learning while outperforming zero-shot baselines, showcasing the potential of the method.\n\n3. It would be beneficial if the authors could release the fine-tuned LoRA weight to the community.", "weaknesses": "1. The choice of using FLAN-T5-Large as the base model is questionable as a model pre-trained on unsupervised text might have been more relevant for fine-tuning / LoRA fine-tuning on upstream tasks from the FLAN collection.\n\n2. The paper lacks clarity in explaining the rationale behind maintaining the same rank for the composed LoRA module and could benefit from exploring higher rank matrices when composing.\n\n3. The selection of 20 LoRAs for unseen tasks seems arbitrary and might limit the method’s performance. An iterative procedure or justification for this selection would have been beneficial.\n\n4. The absence of certain baselines, such as the average performance for BBH's top 5 upstream tasks, leaves gaps in the evaluation.\n\n5. The baseline corresponding to the retrieval of a trained LoRAs, when given a handful of examples from unseen tasks, is missing. For example, see https://arxiv.org/abs/2302.03202", "questions": "1. Can you elaborate on the interpretation of negative coefficients for the LoRA weights?\n\n2. Why was the decision made to maintain the same rank for the composed LoRA module? Have higher rank matrices been explored and if so, what were the findings?\n\n3. Is the selection of 20 LoRAs for unseen tasks fixed or is there an iterative procedure to this selection? How does this choice impact the method’s performance on unseen tasks?\n\n4. I believe the strength of the method lies in cases where there are a handful of examples from unseen tasks. If it's otherwise, that it is beneficial to use a larger number of examples, it would make sense to compare against methods like IA3 [https://arxiv.org/abs/2205.05638], which fine-tunes efficiently on few-shot examples from unseen tasks.\n\n5. Why was FLAN-T5-Large chosen as the base model over a model pre-trained on unsupervised text? Wouldn't it be strange to fine-tune on FLAN tasks using LoRA on an already FLAN multitask-trained backbone model?\n\n6. How does the absolute value of LoRA weight not exceeding 1.5 relate to the method's performance in section 4.2, and is there a particular significance to this threshold?\n\n7. It would be beneficial to include parameter efficient fine-tuning and traditional fine-tuning performance in Table 1, especially with the setup of a limited number of examples.\n\n8. Is In-Context Learning (ICL) performed on the same base Language Model (LLM), or is it conducted using larger decoder only LLMs? What are the implications of this choice on the comparison of results?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698687649180}], "openreview_url": "https://openreview.net/forum?id=w8eCnnq57m", "arxiv_id": "2307.13269", "paper_pdf": "papers/w8eCnnq57m.pdf", "paper_pdf_sha256": "4ce8c9e6faf96627c129463d6db335743eb6ab6445e202e583d86b32c45a6a26", "paper_pdf_bytes": 557225, "paper_pdf_source": "openreview", "code_url": "https://github.com/sail-sg/lorahub", "code_repository": "sail-sg/lorahub", "code_commit": "df73afe5f38d9ff0fd1cd43774be51c79c581cc3", "code_archive": "repos/w8eCnnq57m.zip", "code_archive_sha256": "c4d5490c774ebc9ea13f7ca51137a75fa12b71836a3d7306b640aea101ba58b3", "code_archive_bytes": 268071, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 281, "github_languages": {"Python": 110046, "Shell": 1501}, "github_archived": false, "github_pushed_at": "2024-07-22T00:38:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lorahub-efficient-cross-task-generalization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fUX3bszZSOw", "year": 2023, "status": "rejected", "title": "Do You Remember? Overcoming Catastrophic Forgetting for Fake Audio Detection", "authors": ["XiaoHui Zhang", "Jiangyan Yi", "Chenglong Wang", "Chu Yuan Zhang", "Jianhua Tao"], "authorids": ["~XiaoHui_Zhang4", "~Jiangyan_Yi1", "~Chenglong_Wang5", "~Chu_Yuan_Zhang1", "~Jianhua_Tao2"], "authors_source": "OpenReview API", "abstract": "Current fake audio detection algorithms achieve promising performances on most datasets. However, their performance may be significantly degraded when dealing with audio of a different dataset. The orthogonal weight modification to overcome catastrophic forgetting does not consider the similarity of some audio, including fake audio obtained by the same algorithm and genuine audio, on different datasets. To overcome this limitation, we propose a continual learning algorithm for fake audio detection to overcome catastrophic forgetting, called Regularized Adaptive Weight Modification (RAWM). Specifically, when fine-tuning a detection network, our approach adaptively computes the direction of weight modification according to the ratio of genuine utterances and fake utterances. The adaptive modification direction ensures the network can detect fake audio on the new dataset while preserving its knowledge of previous model, thus mitigating catastrophic forgetting. In addition, orthogonal weight modification of fake audios in the new dataset will skew the distribution of inferences on audio in the previous dataset with similar acoustic characteristics, so we introduce a regularization constraint to force the network to remember this distribution. We evaluate our approach across multiple datasets and obtain a significant performance improvement on cross-dataset experiments.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "O0V3eRIJez", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4338/Reviewer_a37q"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of continual learning in the context of fake audio detection. One of the main challenges of continual learning is catastrophic forgetting. The paper proposes a new algorithm called Regularized Adaptive Weight Modification (RAWM). The motivation is genuine audios are more similar than fake audios in different datasets. The proposed approach can adaptively modify the direction of weights according to the ratio of genuine audio and fake audio of each batch in the process of fine-tuning. The paper evaluated the proposed approach across multiple fake audio detection datasets and showed promising results in the continual learning setting. \n", "review_text": "The paper proposes a novel method (RAWM) for continual learning in the context of fake audio detection. Experiments demonstrate that the proposed method obtains strong performance compared to the prior methods. The paper can be made stronger if the proposed method can be extended to more applications or to general recognition tasks.\n\n", "strengths": "Strength \n\nThe paper proposes a novel method called Regularized Adaptive Weight Modification (RAWM) to address the problem of catastrophic forgetting in the context of fake audio detection. The proposed method builds on prior work orthogonal weight modification (OWM).\n\nThe approach is well-motivated (genuine audios are more similar than fake audios in different datasets. The proposed method obtains a better tradeoff between learning from a new dataset while not forgetting the past knowledge.)\n\nThe paper provides a satisfactory literature review.\n\nThe paper evaluates the proposed approach across multiple fake audio detection datasets and shows promising results in the continual learning setting. \n\nWeaknesses\n\nThe proposed method seems to be general for continual learning. However, experiments only demonstrate its value for the task of fake audio detection. Can authors comment on whether the proposed method can benefit any other tasks other than fake audio detection? The paper can be made stronger if the proposed method can be extended to more applications or to general recognition tasks (e.g., image/video/audio recognition).\n\nDoes the proposed method also work for multiclass classification tasks? If so, what are the necessary changes?\n\nPresentation of section 3 can be improved due to lack of context. It will be useful to include a paragraph to describe the problem setting (e.g., notation of model layers, notation for each dataset, notation for gradient etc.). A lot of the context is only introduced in section 4 (“We consider a feed-forward network consisting of L + 1 layers..”).\n\n\nMinor comments:\n\nIn the second equation of EQ1,  the numerator is P_l(i-1,j) \\bar{X}(l-1). Should \\bar{X}(l-1) be indexed with i,j ?\n\nSection 3.2, “y_o and y_o are the old and new ground truth” (typos?).\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper considers the problem of continual learning in the context of fake audio detection. One of the main challenges of continual learning is catastrophic forgetting. The paper proposes a new algorithm called Regularized Adaptive Weight Modification (RAWM). The motivation is genuine audios are more similar than fake audios in different datasets. The proposed approach can adaptively modify the direction of weights according to the ratio of genuine audio and fake audio of each batch in the process of fine-tuning. The paper evaluated the proposed approach across multiple fake audio detection datasets and showed promising results in the continual learning setting. \n", "strength_and_weaknesses": "Strength \n\nThe paper proposes a novel method called Regularized Adaptive Weight Modification (RAWM) to address the problem of catastrophic forgetting in the context of fake audio detection. The proposed method builds on prior work orthogonal weight modification (OWM).\n\nThe approach is well-motivated (genuine audios are more similar than fake audios in different datasets. The proposed method obtains a better tradeoff between learning from a new dataset while not forgetting the past knowledge.)\n\nThe paper provides a satisfactory literature review.\n\nThe paper evaluates the proposed approach across multiple fake audio detection datasets and shows promising results in the continual learning setting. \n\nWeaknesses\n\nThe proposed method seems to be general for continual learning. However, experiments only demonstrate its value for the task of fake audio detection. Can authors comment on whether the proposed method can benefit any other tasks other than fake audio detection? The paper can be made stronger if the proposed method can be extended to more applications or to general recognition tasks (e.g., image/video/audio recognition).\n\nDoes the proposed method also work for multiclass classification tasks? If so, what are the necessary changes?\n\nPresentation of section 3 can be improved due to lack of context. It will be useful to include a paragraph to describe the problem setting (e.g., notation of model layers, notation for each dataset, notation for gradient etc.). A lot of the context is only introduced in section 4 (“We consider a feed-forward network consisting of L + 1 layers..”).\n\n\nMinor comments:\n\nIn the second equation of EQ1,  the numerator is P_l(i-1,j) \\bar{X}(l-1). Should \\bar{X}(l-1) be indexed with i,j ?\n\nSection 3.2, “y_o and y_o are the old and new ground truth” (typos?).\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: Overall the paper is clear. However, some of the presentation can be made more clear (see Weaknesses) and there are also a few typos in the paper (see Minor comments).\n\nQuality: The proposed method is well-motivated and experiments demonstrate the effectiveness of the proposed method. However, the proposed method is limited to the task of fake audio detection. \n\nNovelty: The proposed method is incremental to OWM, but provides additional novel technical insights for the problem of fake audio detection.\n\nReproducibility: The paper seems reproducible. \n", "summary_of_the_review": "The paper proposes a novel method (RAWM) for continual learning in the context of fake audio detection. Experiments demonstrate that the proposed method obtains strong performance compared to the prior methods. The paper can be made stronger if the proposed method can be extended to more applications or to general recognition tasks.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667287681110}, {"id": "iYWlaz9Cd5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4338/Reviewer_d2it"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors address the problem of \"catastrophic forgetting\" in the context of fake audio detection. When a network trained on one dataset (D1) and is fine tuned on another dataset (D2), the fine-tuned network loses its original performance on the D1. This is referred to as catastrophic forgetting for the considered task of fake audio detection. There are several approaches addressing this problem. One of the recent approaches is using the orthogonal weight modification (OWM) technique. This is based on the viewpoint that modifying the weights in a direction orthogonal to the Weights subspace corresponding to D1 would ensure that the modified network's performance on D1 would not deteriorate. Thus, the idea while fine-tuning on D2 is to modify the weight update equation by modifying the weights in a direction orthogonal to that corresponding to weights subspace representing D1. Such a weight modification has a regularization effect on the network to retain its performance on D1. This however presents a significant drawback: The Performance of this regularized network on D2 suffers compared to without regularization. \n\nIn this paper the authors address the above two issues in OWM by proposing Regularized Adaptive Weight Modification (RAWM) where the above two problems are addressed respectively by a nice observation that blindly updating the weights in an orthogonal direction to the subspace corresponding to D1 might be a bad idea as there could be several genuine audios in D1 which are similar to D2 and for those samples - the regular update might not harm network performance on D2 while improving its performance on D2. This is accounted in the weight update by modifying the direction of update to be based on the number of genuine vs fake examples in each batch and using this to adapt the orthogonal weight update. In the degenrate case of all fake examples this proposed approach would boil down to the OWM method.   \n\nOverall the paper is well written with clear problem definition, experimental results and useful comparisons. ", "review_text": "Overall the proposed solution of adapting the weight modification by accounting for the genuine and fake samples in each batch seems interesting and shows promising results. The methodology is clearly explained with detailed insightful evaluation. ", "strengths": "Strengths:\n1. Paper is well written and the solution proposed addresses the problem quite well. \n2. The experimental validation is sufficiently elaborate and shows good improvement over the baselines considered. \n\nWeakness:\n1. There are a few parts of the paper which are not particularly well written and makes it a difficult read. For instance:\n a) In Section 4.2 the notations and equations 9, 10 are not in sync.\n b) Symbols and notations are not clearly explained on several occasions. \n c) Out of nowhere in Section 5.3, authors start talking about a \"Model-4\n2. How does the proposed method compare to training on all datasets ?\n3. It is not clear why the assumption that genuine audio across different datasets is more likely to be similar than fake audios from different datasets is necessary ? Can the authors not think about exploiting fake audio similarity as well in their technique ?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors address the problem of \"catastrophic forgetting\" in the context of fake audio detection. When a network trained on one dataset (D1) and is fine tuned on another dataset (D2), the fine-tuned network loses its original performance on the D1. This is referred to as catastrophic forgetting for the considered task of fake audio detection. There are several approaches addressing this problem. One of the recent approaches is using the orthogonal weight modification (OWM) technique. This is based on the viewpoint that modifying the weights in a direction orthogonal to the Weights subspace corresponding to D1 would ensure that the modified network's performance on D1 would not deteriorate. Thus, the idea while fine-tuning on D2 is to modify the weight update equation by modifying the weights in a direction orthogonal to that corresponding to weights subspace representing D1. Such a weight modification has a regularization effect on the network to retain its performance on D1. This however presents a significant drawback: The Performance of this regularized network on D2 suffers compared to without regularization. \n\nIn this paper the authors address the above two issues in OWM by proposing Regularized Adaptive Weight Modification (RAWM) where the above two problems are addressed respectively by a nice observation that blindly updating the weights in an orthogonal direction to the subspace corresponding to D1 might be a bad idea as there could be several genuine audios in D1 which are similar to D2 and for those samples - the regular update might not harm network performance on D2 while improving its performance on D2. This is accounted in the weight update by modifying the direction of update to be based on the number of genuine vs fake examples in each batch and using this to adapt the orthogonal weight update. In the degenrate case of all fake examples this proposed approach would boil down to the OWM method.   \n\nOverall the paper is well written with clear problem definition, experimental results and useful comparisons. ", "strength_and_weaknesses": "Strengths:\n1. Paper is well written and the solution proposed addresses the problem quite well. \n2. The experimental validation is sufficiently elaborate and shows good improvement over the baselines considered. \n\nWeakness:\n1. There are a few parts of the paper which are not particularly well written and makes it a difficult read. For instance:\n a) In Section 4.2 the notations and equations 9, 10 are not in sync.\n b) Symbols and notations are not clearly explained on several occasions. \n c) Out of nowhere in Section 5.3, authors start talking about a \"Model-4\n2. How does the proposed method compare to training on all datasets ?\n3. It is not clear why the assumption that genuine audio across different datasets is more likely to be similar than fake audios from different datasets is necessary ? Can the authors not think about exploiting fake audio similarity as well in their technique ?", "clarity,_quality,_novelty_and_reproducibility": "1. Clear Enough\n2. Good Quality\n3. I am not entirely sure this work is reproducible. ", "summary_of_the_review": "Overall the proposed solution of adapting the weight modification by accounting for the genuine and fake samples in each batch seems interesting and shows promising results. The methodology is clearly explained with detailed insightful evaluation. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666833453185}, {"id": "EbzvcSEfuA", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4338/Reviewer_jUDi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a new method for continual learning which leverages structure found in the particular task of fake audio detection. Specifically, the proposed method augments continual learning algorithms by leveraging structural differences found in fake audio detection between two individual distribution shifts latent in the source and target datasets: source fake -> target fake, and source real -> target real. Compared to other continual learning algorithms, the proposed algorithm leads to large improvements in performance on both source and target datasets.", "review_text": "Overall, this paper presents some interesting ideas and compelling results, but the current version struggles w/ (1) motivating the setting of interest (continual learning for fake audio detection) including lack of comparison to methods which can access the source data, and (2) clarity of explanation.", "strengths": "The primary strength of this paper is that of the results: the proposed algorithm clearly outperforms other comparable continual learning algorithms for standard fake audio detection tasks. The primary weakness is a lack of motivation for the problem formulation explored here. Specifically, this paper explores continual learning for fake audio detection, for which there are two key issues:\n\n(1) I would argue that the biggest practical challenge in fake audio detection is not _adapting_ to new domains w/ many examples (as is explored in continual learning) but rather improving _zero/few-shot generalization_ to new domains. The threat model in fake audio detection should emphasize the persistent development of new methods for generating fake audio. Instead, the exploration of continual learning suggests a perpetual cat-and-mouse game where detection models must be manually adapted to new domains.\n\n(2) The motivation for the _continual_ learning setup (where source data is inaccessible) is unclear. The introduction says in one sentence that “in some practical situations, it is almost impossible to obtain the old data”, but does not suggest any particular situations where that would be the case. Moreover, all of the experiments _contrive_ such situations (since the source dataset is always accessible), and there are no comparisons to “oracle” performance w/ access to source data.\n\nAdditionally, this paper struggles w/ clarity (see next section).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new method for continual learning which leverages structure found in the particular task of fake audio detection. Specifically, the proposed method augments continual learning algorithms by leveraging structural differences found in fake audio detection between two individual distribution shifts latent in the source and target datasets: source fake -> target fake, and source real -> target real. Compared to other continual learning algorithms, the proposed algorithm leads to large improvements in performance on both source and target datasets.", "strength_and_weaknesses": "The primary strength of this paper is that of the results: the proposed algorithm clearly outperforms other comparable continual learning algorithms for standard fake audio detection tasks. The primary weakness is a lack of motivation for the problem formulation explored here. Specifically, this paper explores continual learning for fake audio detection, for which there are two key issues:\n\n(1) I would argue that the biggest practical challenge in fake audio detection is not _adapting_ to new domains w/ many examples (as is explored in continual learning) but rather improving _zero/few-shot generalization_ to new domains. The threat model in fake audio detection should emphasize the persistent development of new methods for generating fake audio. Instead, the exploration of continual learning suggests a perpetual cat-and-mouse game where detection models must be manually adapted to new domains.\n\n(2) The motivation for the _continual_ learning setup (where source data is inaccessible) is unclear. The introduction says in one sentence that “in some practical situations, it is almost impossible to obtain the old data”, but does not suggest any particular situations where that would be the case. Moreover, all of the experiments _contrive_ such situations (since the source dataset is always accessible), and there are no comparisons to “oracle” performance w/ access to source data.\n\nAdditionally, this paper struggles w/ clarity (see next section).", "clarity,_quality,_novelty_and_reproducibility": "**Clarity**\n\nOverall, I found this paper to be quite unclear and required several passes to understand. A primary issue is definition and consistency of terminology. For example, this paper primarily concerns “fake audio detection” and “continual learning” but neither are defined anywhere. Also, this paper fluctuates between using “source / target” (in Figure 1 / Experiments) and “old / new” (throughout) naming conventions for adaptation. Important notations like T_1, T_2, and T_3 are introduced in the header of Table 1 instead of Section 5.1 where the reader might expect to see them.\n\n\nA key question I was left with after reading this paper: in Table 6, why do the majority of the continual learning algorithms outperform simple fine-tuning on the target dataset? This may be an understood phenomenon in the continual learning research community but I find it incredibly counterintuitive as an outsider. Can the authors clarify the “free lunch” aspect of continual learning, especially in this context of fake audio detection?\n\nAnother comment is that the introduction could do more to clarify the specific assumptions about fake audio detection that motivate the need for tailored continual learning algorithms. The third paragraph of the intro attempts to do this in one sentence (starting with “Because the…”), but this sentence is exceptionally difficult to parse and entangled w/ the proposed method (instead of stated in a method-agnostic fashion). A clearer upfront explanation of this would have greatly improved my ability to understand the rest of the paper on the first pass.\n\nThere are also numerous places throughout the paper containing misleading or ambiguous statements. For example, the first paragraph of section 5.1 suggests that the training subset of the source dataset will be used, but the training subsets of the target datasets will not be (which is wrong in the context of continual learning).\n\n** Quality **: Accepting the sole focus on continual learning, the experiments are reasonably well-designed. However, I would still have very much liked to see a comparison to “oracle” methods w/ access to the source dataset - how much performance do we lose by making the assumption that source datasets are inaccessible?\n\n** Novelty **: The proposed method is adequately interesting and novel - a stronger version of this paper might explore the proposed method for other tasks (besides fake audio detection) with similar structure.\n\n** Reproducibility **: It would be an extraordinary challenge to reproduce this algorithm from the notation / information in this paper alone - would the authors be able to release code?\n", "summary_of_the_review": "Overall, this paper presents some interesting ideas and compelling results, but the current version struggles w/ (1) motivating the setting of interest (continual learning for fake audio detection) including lack of comparison to methods which can access the source data, and (2) clarity of explanation.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666647679806}, {"id": "YZUmPft_Rkz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4338/Reviewer_VFqH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a continual learning algorithm, called Regularized Adaptive Weight Modification (RAWM) to overcome catastrophic forgetting for fake audio detection. RAWM is based on the previously published orthogonal weight modification (OWM). Because OWM does not consider the similarity of some audio, including fake audio obtained by the same algorithm and real audio, on different datasets. To solve this limitation, adaptive modification direction and a regularization constraint are proposed. Experimental results show a good performance improvement compared to baselines.", "review_text": "This paper is based on the previously published OWM paper, and two modifications (adaptive modification direction and a regularization constraint) are proposed. As pointed out earlier, the reason to apply Q in adaptive modification direction is unclear, and the novelty of using a teacher model for regularization constraint is somewhat limited.", "strengths": "Strength:\nThe proposed continual learning algorithm is designed for the application of fake audio detection. The authors observe some problems when using conventional continual learning algorithms and propose two methods (adaptive modification direction and a regularization constraint) to overcome them. \n\nWeaknesses:\nTo avoid using training data from the previous tasks, the proposed regularization constraint is based on a teacher model, which is not a very novel idea. On the other hand, the way to apply an adaptive modification direction is not very clear to me. In page 4, the authors claim that “In the training process on the new dataset, they should be trained without any modification.” However, the authors propose a new projector Q which is orthogonal to the projector P from OWM. My question is: why not just use the original gradient from SGD, ∆W^BP in this case? \nOther comments:\n1. Is P a square matrix? Otherwise, how to compute Q using e.q. (4)?\n\n2. In the original paper of ‘OWM’ (Continual Learning of Context-dependent Processing in Neural Networks), the authors claim that ” Recursive Least Square (RLS) algorithm which can be used to train feedforward and recurrent neural networks to achieve fast convergence’ However in this paper, the proposed fake audio detection model is based on Wav2vec (which includes convolutional encoder) and a CNN classifier. I’m not sure how OWM performs on convolutional layers. My concern is that I think OWM assumes the calculation between input and model weights is through matrix multiplication instead of convolutional operation.\n\n3. In Fig. 1. of the original paper of ‘OWM’, they claim that OWM can reach the position inside the overlapping subspace between two tasks. However, in Fig. 1. of your paper, you mention that OWM can not reach the region of dataset 2. Please explain why there is such a difference.\n\n4. In Table 6, why even Fine-tune cannot obtain the best performance on the new task compared to other methods?\n\n5. It would be great to also report the ‘experience-replay-based’ method in the experiment.\n\nTypo:\n1. Page 1, fae audio detection -> fake audio detection \n\n2. Page 3, yo and yo are the old and new ground truth -> yo and yn are the old and new ground truth\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a continual learning algorithm, called Regularized Adaptive Weight Modification (RAWM) to overcome catastrophic forgetting for fake audio detection. RAWM is based on the previously published orthogonal weight modification (OWM). Because OWM does not consider the similarity of some audio, including fake audio obtained by the same algorithm and real audio, on different datasets. To solve this limitation, adaptive modification direction and a regularization constraint are proposed. Experimental results show a good performance improvement compared to baselines.", "strength_and_weaknesses": "Strength:\nThe proposed continual learning algorithm is designed for the application of fake audio detection. The authors observe some problems when using conventional continual learning algorithms and propose two methods (adaptive modification direction and a regularization constraint) to overcome them. \n\nWeaknesses:\nTo avoid using training data from the previous tasks, the proposed regularization constraint is based on a teacher model, which is not a very novel idea. On the other hand, the way to apply an adaptive modification direction is not very clear to me. In page 4, the authors claim that “In the training process on the new dataset, they should be trained without any modification.” However, the authors propose a new projector Q which is orthogonal to the projector P from OWM. My question is: why not just use the original gradient from SGD, ∆W^BP in this case? \nOther comments:\n1. Is P a square matrix? Otherwise, how to compute Q using e.q. (4)?\n\n2. In the original paper of ‘OWM’ (Continual Learning of Context-dependent Processing in Neural Networks), the authors claim that ” Recursive Least Square (RLS) algorithm which can be used to train feedforward and recurrent neural networks to achieve fast convergence’ However in this paper, the proposed fake audio detection model is based on Wav2vec (which includes convolutional encoder) and a CNN classifier. I’m not sure how OWM performs on convolutional layers. My concern is that I think OWM assumes the calculation between input and model weights is through matrix multiplication instead of convolutional operation.\n\n3. In Fig. 1. of the original paper of ‘OWM’, they claim that OWM can reach the position inside the overlapping subspace between two tasks. However, in Fig. 1. of your paper, you mention that OWM can not reach the region of dataset 2. Please explain why there is such a difference.\n\n4. In Table 6, why even Fine-tune cannot obtain the best performance on the new task compared to other methods?\n\n5. It would be great to also report the ‘experience-replay-based’ method in the experiment.\n\nTypo:\n1. Page 1, fae audio detection -> fake audio detection \n\n2. Page 3, yo and yo are the old and new ground truth -> yo and yn are the old and new ground truth\n", "clarity,_quality,_novelty_and_reproducibility": "This paper presents the problem quite clearly, however, the reason for using projector Q instead of normal gradient need to be explained. ", "summary_of_the_review": "This paper is based on the previously published OWM paper, and two modifications (adaptive modification direction and a regularization constraint) are proposed. As pointed out earlier, the reason to apply Q in adaptive modification direction is unclear, and the novelty of using a teacher model for regularization constraint is somewhat limited.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666401049258}], "openreview_url": "https://openreview.net/forum?id=fUX3bszZSOw", "arxiv_id": "2308.03300", "paper_pdf": "papers/fUX3bszZSOw.pdf", "paper_pdf_sha256": "4e3e72b1c89b0085d90f2aae4cf92eacd32aee980d6e69703c3961fd01e091b5", "paper_pdf_bytes": 1633379, "paper_pdf_source": "openreview", "code_url": "https://github.com/Cecile-hi/Regularized-Adaptive-Weight-Modification", "code_repository": "Cecile-hi/Regularized-Adaptive-Weight-Modification", "code_commit": "bb663ee083d8d9f8309701f2255b87916eff55c6", "code_archive": "repos/fUX3bszZSOw.zip", "code_archive_sha256": "06be15f23719a8053871703285a433ab3dbed68ee8d4c151d00b8a840d1ff4a3", "code_archive_bytes": 714128, "code_file_count": 156, "code_extensions": {".py": 154, ".sh": 2}, "github_disk_usage_kb": 647, "github_languages": {"Python": 1307444, "Shell": 363}, "github_archived": false, "github_pushed_at": "2024-09-26T08:14:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/do-you-remember-overcoming-catastrophic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Yn4CPz_LRKO", "year": 2022, "status": "rejected", "title": "Conditional GANs with Auxiliary Discriminative Classifier", "authors": ["Liang Hou", "Qi Cao", "Huawei Shen", "Xueqi Cheng"], "authorids": ["~Liang_Hou1", "~Qi_Cao1", "~Huawei_Shen1", "~Xueqi_Cheng1"], "authors_source": "OpenReview API", "abstract": "Conditional generative models aim to learn the underlying joint distribution of data and labels, and thus realize conditional generation. Among them, auxiliary classifier generative adversarial networks (AC-GAN) have been widely used, but suffer from the problem of low intra-class diversity on generated samples. In this paper, we point out that the fundamental reason is that the classifier of AC-GAN is generator-agnostic, and therefore cannot provide informative guidance to the generator to approximate the target distribution, resulting in minimization of conditional entropy that decreases the intra-class diversity. Motivated by this observation, we propose a novel conditional GAN with auxiliary \\textit{discriminative} classifier (ADC-GAN) to resolve the problem of AC-GAN. Specifically, the proposed auxiliary \\textit{discriminative} classifier becomes generator-aware by recognizing the labels of the real data and the generated data \\textit{discriminatively}. Our theoretical analysis reveals that the generator can faithfully replicate the target distribution even without the original discriminator, making the proposed ADC-GAN robust to the hyper-parameter and stable during the training process. Extensive experimental results on synthetic and real-world datasets demonstrate the superiority of ADC-GAN on conditional generative modeling compared to competing methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "LbdeIQnr2Df", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1198/Reviewer_mZT7"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper is about improving conditional GANs. To be specifically, it also aims to resolve the bias issue of ACGAN by proposing a discriminative classifier. The discriminative classifier is a hybrid model of discriminator and classifier, where it has to not tell real or fake, but also the class.  Preliminary analysis of the proposed method are provided. Experiments are conducted on the standard benchmarks. ", "review_text": "Strength:\n\n* The proposed discriminative classifiers seems interesting to resolve the biased issue of ACGAN. \n\nWeakness:\n\n* The key results seems to be Thm 2, which is based on Prop. 2. However, the proof shows pm(x, y, l) = pm(x, y, 1) + pm(x, y, 0) = 1/2 p(x, y) + 1/2q(x, y). at the very beginning. How do you get the second equation?  I guess it may not be a fatal error, but I can't tell the correctness at the moment. \n\n* The main criticism of AC-GAN is being generator-agnostic, which I think it's not fully appropriate. A common practical implementation of AC-GAN is also using generated data to train classifier, which I think it's a straightforward idea. Under this, whether it can simply resolve the generator-agnostic issue? However, by just doing so, the performance seems not as competitive as the reported numbers of the proposed method.  Could you comment on it? \n\n* The reported numbers looks not quite consistent with other works to me.  \n\n  - For CIFAR10 results, the reported FID are all below 7, however, it's not the case for most of the existing works. For example, all the models reported here https://github.com/POSTECH-CVLab/PyTorch-StudioGAN#cifar10-3x32x32 are all with FID > 7. \n\n - For CIFAR100, the numbers reported in TAC-GAN seems better than the 11.37 reported in the paper for the TAC-GAN and 7.98 for the proposed method. \n\n- For ImageNet, again the reported BigGAN is worse than it should be.  For example, see Table 2 in https://arxiv.org/pdf/2111.01118v1.pdf which provides a nice comparison. \n\n* Some descriptions are not accurate. For example, there are multiple sentences about \"GANs are notoriously unstable to train\", which is true back to 2014. However, there are already \"tons\" of papers working on it to resolve the issues. The authors should at least cite those and make a fair description.  Secondly, the below Eq (6), the authors mention they proposed method can be \"unbiasedly optimize\", which is not true under alternative and minibatch setting. Again, there are lots of works (check all the works on GAN optimization) discussing this issue. The authors should remove this incorrect claim. \n\n* Proposition 1 seems trivial. Also, Theorem 1 is quite similar to the analysis in TAC-GAN. I would highly suggest removing Proposition 1,  which you don't have to pretend to be a theoretical paper. Also, cite TAC-GAN for the analysis. We should not copy or redo the analysis from the predecessor. \n\n* I have the concern of the analysis. Most of the results rely on assuming sth components are optimal. However, in reality, they are not hold in reality, and there is no convergence analysis provided.  Could the authors comment on it?  \n\n* The biased issue of ACGAN is known.  For example, \n\n  - AC-GAN Learns a Biased Distribution, 2017\n - Unbiased Auxiliary Classifier GANs with MINE, 2020\n\n There are many more. The authors should provide a better overview for the progress of this direction.\n\n\nQuestion:\n\n* In Table 1, PD-GAN is optimizing JS(P||Q) while the proposed ADC-GAN is optimizing KL(P||Q).  To me, there should not much difference. Any insights why the proposed ADC-GAN, which optimizes an asymmeric loss, should be better? \n\nSuggestion:\n\n* There are some very recent works in NeurIPS and also highly relayed, \n  - Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training, NeurIPS 2021\n  - A Unified View of cGANs with and without Classifiers, NeurIPS 2021\n  Although they are posted online after ICLR deadline, I would strongly encourage the authors comment on the similarities and differences between the proposed work and these two, because they will be on public for a when the paper decision of ICLR is out anyway. Note that I won't judge the paper decision based on this, but I think it's great to have for the community. It's fine even though the ideas are overlapping. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper is about improving conditional GANs. To be specifically, it also aims to resolve the bias issue of ACGAN by proposing a discriminative classifier. The discriminative classifier is a hybrid model of discriminator and classifier, where it has to not tell real or fake, but also the class.  Preliminary analysis of the proposed method are provided. Experiments are conducted on the standard benchmarks. ", "main_review": "Strength:\n\n* The proposed discriminative classifiers seems interesting to resolve the biased issue of ACGAN. \n\nWeakness:\n\n* The key results seems to be Thm 2, which is based on Prop. 2. However, the proof shows pm(x, y, l) = pm(x, y, 1) + pm(x, y, 0) = 1/2 p(x, y) + 1/2q(x, y). at the very beginning. How do you get the second equation?  I guess it may not be a fatal error, but I can't tell the correctness at the moment. \n\n* The main criticism of AC-GAN is being generator-agnostic, which I think it's not fully appropriate. A common practical implementation of AC-GAN is also using generated data to train classifier, which I think it's a straightforward idea. Under this, whether it can simply resolve the generator-agnostic issue? However, by just doing so, the performance seems not as competitive as the reported numbers of the proposed method.  Could you comment on it? \n\n* The reported numbers looks not quite consistent with other works to me.  \n\n  - For CIFAR10 results, the reported FID are all below 7, however, it's not the case for most of the existing works. For example, all the models reported here https://github.com/POSTECH-CVLab/PyTorch-StudioGAN#cifar10-3x32x32 are all with FID > 7. \n\n - For CIFAR100, the numbers reported in TAC-GAN seems better than the 11.37 reported in the paper for the TAC-GAN and 7.98 for the proposed method. \n\n- For ImageNet, again the reported BigGAN is worse than it should be.  For example, see Table 2 in https://arxiv.org/pdf/2111.01118v1.pdf which provides a nice comparison. \n\n* Some descriptions are not accurate. For example, there are multiple sentences about \"GANs are notoriously unstable to train\", which is true back to 2014. However, there are already \"tons\" of papers working on it to resolve the issues. The authors should at least cite those and make a fair description.  Secondly, the below Eq (6), the authors mention they proposed method can be \"unbiasedly optimize\", which is not true under alternative and minibatch setting. Again, there are lots of works (check all the works on GAN optimization) discussing this issue. The authors should remove this incorrect claim. \n\n* Proposition 1 seems trivial. Also, Theorem 1 is quite similar to the analysis in TAC-GAN. I would highly suggest removing Proposition 1,  which you don't have to pretend to be a theoretical paper. Also, cite TAC-GAN for the analysis. We should not copy or redo the analysis from the predecessor. \n\n* I have the concern of the analysis. Most of the results rely on assuming sth components are optimal. However, in reality, they are not hold in reality, and there is no convergence analysis provided.  Could the authors comment on it?  \n\n* The biased issue of ACGAN is known.  For example, \n\n  - AC-GAN Learns a Biased Distribution, 2017\n - Unbiased Auxiliary Classifier GANs with MINE, 2020\n\n There are many more. The authors should provide a better overview for the progress of this direction.\n\n\nQuestion:\n\n* In Table 1, PD-GAN is optimizing JS(P||Q) while the proposed ADC-GAN is optimizing KL(P||Q).  To me, there should not much difference. Any insights why the proposed ADC-GAN, which optimizes an asymmeric loss, should be better? \n\nSuggestion:\n\n* There are some very recent works in NeurIPS and also highly relayed, \n  - Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training, NeurIPS 2021\n  - A Unified View of cGANs with and without Classifiers, NeurIPS 2021\n  Although they are posted online after ICLR deadline, I would strongly encourage the authors comment on the similarities and differences between the proposed work and these two, because they will be on public for a when the paper decision of ICLR is out anyway. Note that I won't judge the paper decision based on this, but I think it's great to have for the community. It's fine even though the ideas are overlapping. ", "summary_of_the_review": "The main concerns are first on the analysis, which I couldn't tell the correctness at the moment. Second issue is the empirical results, which seems not consistent with other works. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635894327146}, {"id": "sPLQC5gW_2w", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1198/Reviewer_ebJs"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new conditional GAN model that employs a discriminative classifier that predicts in the joint space of label and real/fake domain. The theoretical analysis shows the proposed ADC-GAN can minimize the reverse KL between joint $Q_{X,Y}$ and $P_{X,Y}$.", "review_text": "## Strengths\n1. The paper presents interesting analysis of AC-GAN, TAC-GAN, PD-GAN. Especially the Theorem 3 reveals potential drawbacks of TAC-GAN.\n2. The experimental results on synthetic and real datasets demonstrate the superiority of ADC-GAN on conditional generative modeling tasks.\n\n## Weaknesses\n1. I don't fully agree with some claims made by the authors:\n   1. Page 5, footnote 2, it might be true that term (a) is \"ignored\" or set to zero, but it is not sufficient to say this inductive bias is a mistake.\n2. Equation 8, this is not the original form of TAC-GAN. The original TAC-GAN is built upon AC-GAN, so term (c) in Equation 3 in the TAC-GAN paper is missing. I notice that Equation 8 is consistent with the actual implementation of TAC-GAN, but I guess it is good to state this clearly in the paper.\n3. Implementations of ADC-GAN and TAC-GAN are the same? As I checked the provided code in supplementary, I think the proposed ADC-GAN is very similar to TAC-GAN (as defined in Equation 8): In fact, if spectral norm (SN) and bias are not used in the linear classification layer, they are exactly equivalent. This is because the weight of $C_d$ is just $C$ and $C_{mi}$ stacked together. I would consider this as an implementation difference. Note that Theorem 2 and 3 are different, I doubt the superior performance of ADC-GAN might come from the difference in SN or a different choice of hyperparameters. In such case, it would be helpful if the author could provide code for MoG experiment, which is cleaner, simpler, and no SN applied (if the code is borrowed from TAC-GAN). Please correct me if I am wrong, and I'm happy to amend my score accordingly.\n4. It would be helpful if the author could provide results of ADC-GAN on ImageNet at 256 resolution.\n5. Table 2, why not use the reported numbers in TAC-GAN paper? I checked Table 1 in TAC-GAN paper, and their FID on CIFAR100 is 7.22 which is lower than the reported 7.98, any explanation?\n6. It is nice to see ADC-GAN worked even without GAN loss, I am curious how the model performs (w/o GAN loss) on challenging datasets such as ImageNet? It is also surprising that in Supplementary Table 4, Hinge loss does worse than no GAN loss: if Theorem 2 holds, the solution set of ADC-GAN is a subset of (unconditional) GAN (say with hinge loss), so adding GAN loss wouldn't affect ADC-GAN training. Can the author explain this?\n7. Comparing Theorem 2 with a plain cGAN which minimizes $JS(Q_{X,Y}||P_{X,Y})$, does the reverse KL tend to cause mode collapse? (in theory it seems that JS is better than reverse KL? a typical yet imprecise point to be made is that KL causes mode averaging and reverse KL causes mode collapse. [1])\n\n[1] Zhao, Miaoyun, et al. \"Bridging Maximum Likelihood and Adversarial Learning via α-Divergence.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 04. 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new conditional GAN model that employs a discriminative classifier that predicts in the joint space of label and real/fake domain. The theoretical analysis shows the proposed ADC-GAN can minimize the reverse KL between joint $Q_{X,Y}$ and $P_{X,Y}$.", "main_review": "## Strengths\n1. The paper presents interesting analysis of AC-GAN, TAC-GAN, PD-GAN. Especially the Theorem 3 reveals potential drawbacks of TAC-GAN.\n2. The experimental results on synthetic and real datasets demonstrate the superiority of ADC-GAN on conditional generative modeling tasks.\n\n## Weaknesses\n1. I don't fully agree with some claims made by the authors:\n   1. Page 5, footnote 2, it might be true that term (a) is \"ignored\" or set to zero, but it is not sufficient to say this inductive bias is a mistake.\n2. Equation 8, this is not the original form of TAC-GAN. The original TAC-GAN is built upon AC-GAN, so term (c) in Equation 3 in the TAC-GAN paper is missing. I notice that Equation 8 is consistent with the actual implementation of TAC-GAN, but I guess it is good to state this clearly in the paper.\n3. Implementations of ADC-GAN and TAC-GAN are the same? As I checked the provided code in supplementary, I think the proposed ADC-GAN is very similar to TAC-GAN (as defined in Equation 8): In fact, if spectral norm (SN) and bias are not used in the linear classification layer, they are exactly equivalent. This is because the weight of $C_d$ is just $C$ and $C_{mi}$ stacked together. I would consider this as an implementation difference. Note that Theorem 2 and 3 are different, I doubt the superior performance of ADC-GAN might come from the difference in SN or a different choice of hyperparameters. In such case, it would be helpful if the author could provide code for MoG experiment, which is cleaner, simpler, and no SN applied (if the code is borrowed from TAC-GAN). Please correct me if I am wrong, and I'm happy to amend my score accordingly.\n4. It would be helpful if the author could provide results of ADC-GAN on ImageNet at 256 resolution.\n5. Table 2, why not use the reported numbers in TAC-GAN paper? I checked Table 1 in TAC-GAN paper, and their FID on CIFAR100 is 7.22 which is lower than the reported 7.98, any explanation?\n6. It is nice to see ADC-GAN worked even without GAN loss, I am curious how the model performs (w/o GAN loss) on challenging datasets such as ImageNet? It is also surprising that in Supplementary Table 4, Hinge loss does worse than no GAN loss: if Theorem 2 holds, the solution set of ADC-GAN is a subset of (unconditional) GAN (say with hinge loss), so adding GAN loss wouldn't affect ADC-GAN training. Can the author explain this?\n7. Comparing Theorem 2 with a plain cGAN which minimizes $JS(Q_{X,Y}||P_{X,Y})$, does the reverse KL tend to cause mode collapse? (in theory it seems that JS is better than reverse KL? a typical yet imprecise point to be made is that KL causes mode averaging and reverse KL causes mode collapse. [1])\n\n[1] Zhao, Miaoyun, et al. \"Bridging Maximum Likelihood and Adversarial Learning via α-Divergence.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 04. 2020.", "summary_of_the_review": "I think the paper presents an interesting analysis of TAC-GAN, AC-GAN, and other cGAN methods. My concern is that the proposed method has the same actual implementation as existing method (TAC-GAN). The paper also lacks results on high-resolution generation. I am willing to raise my score if my concerns are resolved.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635829443753}, {"id": "F9V6r9VE3YK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1198/Reviewer_DPgR"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes the Auxiliary Discriminative Classifier GAN (ADC-GAN) to eliminate a contractionary objective and conditional entropy in ACGAN generator training. Specifically, the authors mathematically demonstrate that training ACGAN without a discriminative label classifier causes minimizing an undesirable divergence (KL(q(x)||p(x))) which conflicts with the joint distribution matching (KL(q(x,y)||p(x,y))). Also, they insist that the lack of intra-class diversity of ACGAN results from the absence of generator guidance for training the discriminator. To resolve all these issues, they devise a new classifier, the auxiliary discriminative classifier and deploy the new classifier directly on the ACGAN framework. Experiments demonstrate that ADC-GAN can successfully learn the joint distribution whose conditional marginals have non-negligible support overlap using MoG dataset. In addition, they show the effectiveness of ADCGAN compared to ACGAN, projection discriminator, and TAC-GAN on four benchmark datasets (CIFAR10, CIFAR100, Tiny-ImageNet, and ImageNet) using IS, FID, iFID metrics. ", "review_text": "Strengths:\n\n(+) The paper exactly points out the primitive problem of ACGAN from the optimization perspective. Since ACGAN is widely adopted in the machine learning area, analyzing problems of ACGAN is necessary and valuable.\n\n(+) The proposed auxiliary discriminative classifier is reasonable, and easy to implement. Also, ADC-GAN does not require much computational burden.\n\n(+) Section 4.2 is very interesting and the explanation of why projection discriminator fails to approximate the joint distribution in Figure 2 is reasonable.\n\nWeaknesses:\n\n(-) It seems that Theorem 1 has already been covered in TAC-GAN paper (paragraphs below eq.4 of the TAC-GAN paper [R1]). Although mathematical formulations are different from each other, the arguments of Theorem 1 and the paragraphs seem to be very similar. I think it is essential to clarify differences between two arguments.\n\n(-) It seems that all experiments were conducted once. It would be better to conduct experiments several times since GANs have been known to have a large performance variance. \n\n(-) The contribution that ADC-GAN can generate diverse images compared to ACGAN and TAC-GAN is not fully demonstrated. Although FID has been a widely used metric to measure fidelity and diversity of generated images, I think It is not enough. I recommend the authors to utilize the improved precision and recall [R2], classification accuracy score [R3], or density and coverage [R4] to quantify the ability of generating diverse images of ADC-GAN.\n\n(-) In section 5.1, the authors conducted the distribution learning experiment using one-dimensional conditional gaussians whose supports are overlapped. I accept that ADC-GAN can learn the joint distribution which consists of the one-dimensional conditional gaussians better than PD-GAN, AC-GAN, and TAC-GAN. However, what about a joint distribution which consists of conditional gaussians with disjoint supports? Can ADC-GAN learn the joint distribution better than other cGANs?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes the Auxiliary Discriminative Classifier GAN (ADC-GAN) to eliminate a contractionary objective and conditional entropy in ACGAN generator training. Specifically, the authors mathematically demonstrate that training ACGAN without a discriminative label classifier causes minimizing an undesirable divergence (KL(q(x)||p(x))) which conflicts with the joint distribution matching (KL(q(x,y)||p(x,y))). Also, they insist that the lack of intra-class diversity of ACGAN results from the absence of generator guidance for training the discriminator. To resolve all these issues, they devise a new classifier, the auxiliary discriminative classifier and deploy the new classifier directly on the ACGAN framework. Experiments demonstrate that ADC-GAN can successfully learn the joint distribution whose conditional marginals have non-negligible support overlap using MoG dataset. In addition, they show the effectiveness of ADCGAN compared to ACGAN, projection discriminator, and TAC-GAN on four benchmark datasets (CIFAR10, CIFAR100, Tiny-ImageNet, and ImageNet) using IS, FID, iFID metrics. ", "main_review": "Strengths:\n\n(+) The paper exactly points out the primitive problem of ACGAN from the optimization perspective. Since ACGAN is widely adopted in the machine learning area, analyzing problems of ACGAN is necessary and valuable.\n\n(+) The proposed auxiliary discriminative classifier is reasonable, and easy to implement. Also, ADC-GAN does not require much computational burden.\n\n(+) Section 4.2 is very interesting and the explanation of why projection discriminator fails to approximate the joint distribution in Figure 2 is reasonable.\n\nWeaknesses:\n\n(-) It seems that Theorem 1 has already been covered in TAC-GAN paper (paragraphs below eq.4 of the TAC-GAN paper [R1]). Although mathematical formulations are different from each other, the arguments of Theorem 1 and the paragraphs seem to be very similar. I think it is essential to clarify differences between two arguments.\n\n(-) It seems that all experiments were conducted once. It would be better to conduct experiments several times since GANs have been known to have a large performance variance. \n\n(-) The contribution that ADC-GAN can generate diverse images compared to ACGAN and TAC-GAN is not fully demonstrated. Although FID has been a widely used metric to measure fidelity and diversity of generated images, I think It is not enough. I recommend the authors to utilize the improved precision and recall [R2], classification accuracy score [R3], or density and coverage [R4] to quantify the ability of generating diverse images of ADC-GAN.\n\n(-) In section 5.1, the authors conducted the distribution learning experiment using one-dimensional conditional gaussians whose supports are overlapped. I accept that ADC-GAN can learn the joint distribution which consists of the one-dimensional conditional gaussians better than PD-GAN, AC-GAN, and TAC-GAN. However, what about a joint distribution which consists of conditional gaussians with disjoint supports? Can ADC-GAN learn the joint distribution better than other cGANs?", "summary_of_the_review": "The authors propose a new type of ACGAN named ADC-GAN to address an improper optimization process of ACGAN. They apply adversarial training not only for the discriminator but also for the auxiliary classifier to eliminate a contradictory divergence and conditional entropy in ACGAN training. In the experimental results, they prove the effectiveness of ADC-GAN using synthetic datasets and various benchmark datasets. However, I think Theorem 1 has already been addressed in TAC-GAN paper and experimental results do not fully demonstrate the effectiveness of the proposed method. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635811317491}, {"id": "Glk94RfuJ5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1198/Reviewer_uPwH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "- This paper aims to solve the low intra-class diversity on generated images of AC-GAN, a classifier-based cGAN. \n- As far as I know, this is an important issue that limits classifier-based cGANs (the counterpart is the projection-based cGAN, i.e, PD-GAN).\n- The authors point out that the reason is that the classifier of AC-GAN is generator-agnostic and minimization of conditional entropy decreases the intra-class diversity. \n- The authors propose ADC-GAN (auxiliary discriminative classifier) to solve this problem, and theoretical analysis is also presented.\n", "review_text": "Strengths:\n\n- The problem of low intra-class diversity of classifier-based cGANs is important. \n    - Reasons: The projection-based cGAN, PD-GAN, does not suffer from the low intra-class diversity problem, but it converges more slowly than classifier-based cGANs (see Omni-GAN, arXiv:2011.13074).  The classifier-based cGANs converge faster but suffer from low intra-class diversity. Therefore, it is of great value to improve classifier-based cGANs so that we can completely abandon the PD-GAN converging slowly in practice. \n- This paper provides a theoretical perspective for analyzing the loss function of different cGANs. \n- The proposed ADC-GAN is very simple to implement without additional computational overhead.\n\nWeaknesses:\n\n- Please detail in the paper how the FID in Table 2 is calculated (for example, how many generated images are used, whether the training set or the validation set is used, and whether the inception model is from PyTorch or tensorflow). In addition, I also recommend including IS in Table 2.\n- In Equ. 5, the notations of $C_d(y,1|x)$ and $C_d(y,0|x)$ are a bit confusing. After checking the code in the supplementary material, I understood the meaning of the equation. In fact, $C_d(y,1|x)$ and $C_d(y,0|x)$ are implemented using a fully connected layer with output dimensions of num_classes * 2. I suggest that the author use much clearer notation to make it easier to understand. The author can refer to the notation of equation 8, which is clear. \n- In Figure 4e, I am surprised that as a classifier-based cGAN, ADC-GAN did not suffer from mode collapse. The author also does not seem to apply weight decay for the discriminator, as Omni-GAN does. I am not sure if the author has used other regularization techniques to stabilize the training. As far as I know, if do not add regularizations such as weight decay, other classifier-based cGANs will collapse earlier, such as AC-GAN, Multi-hinge GAN, and Omni-GAN. It would be better if the author could explain this phenomenon. \n- I know that there is an improved version of AC-GAN (ImAC-GAN), as discussed in section 3.3 of the Omni-GAN paper. ImAC-GAN has a clear performance gain compared to AC-GAN. It would be better if the author could discuss the relationship between ImAC-GAN and ADC-GAN. \n- In Figure 4 (c) and (f), the results of T-SNE visualization are not very convincing. In my opinion, the author should not use the discriminator to extract feature representations, because the PD-GAN discriminator is less supervised than ADC-GAN. I suggest that the author use a pre-trained classification model to extract features for a fair comparison.\n- In the caption of Figure 4, the author says that the T-SNE uses training data, but in the last paragraph of section 5.3, the author says that the T-SNE uses the validation data. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "- This paper aims to solve the low intra-class diversity on generated images of AC-GAN, a classifier-based cGAN. \n- As far as I know, this is an important issue that limits classifier-based cGANs (the counterpart is the projection-based cGAN, i.e, PD-GAN).\n- The authors point out that the reason is that the classifier of AC-GAN is generator-agnostic and minimization of conditional entropy decreases the intra-class diversity. \n- The authors propose ADC-GAN (auxiliary discriminative classifier) to solve this problem, and theoretical analysis is also presented.\n", "main_review": "Strengths:\n\n- The problem of low intra-class diversity of classifier-based cGANs is important. \n    - Reasons: The projection-based cGAN, PD-GAN, does not suffer from the low intra-class diversity problem, but it converges more slowly than classifier-based cGANs (see Omni-GAN, arXiv:2011.13074).  The classifier-based cGANs converge faster but suffer from low intra-class diversity. Therefore, it is of great value to improve classifier-based cGANs so that we can completely abandon the PD-GAN converging slowly in practice. \n- This paper provides a theoretical perspective for analyzing the loss function of different cGANs. \n- The proposed ADC-GAN is very simple to implement without additional computational overhead.\n\nWeaknesses:\n\n- Please detail in the paper how the FID in Table 2 is calculated (for example, how many generated images are used, whether the training set or the validation set is used, and whether the inception model is from PyTorch or tensorflow). In addition, I also recommend including IS in Table 2.\n- In Equ. 5, the notations of $C_d(y,1|x)$ and $C_d(y,0|x)$ are a bit confusing. After checking the code in the supplementary material, I understood the meaning of the equation. In fact, $C_d(y,1|x)$ and $C_d(y,0|x)$ are implemented using a fully connected layer with output dimensions of num_classes * 2. I suggest that the author use much clearer notation to make it easier to understand. The author can refer to the notation of equation 8, which is clear. \n- In Figure 4e, I am surprised that as a classifier-based cGAN, ADC-GAN did not suffer from mode collapse. The author also does not seem to apply weight decay for the discriminator, as Omni-GAN does. I am not sure if the author has used other regularization techniques to stabilize the training. As far as I know, if do not add regularizations such as weight decay, other classifier-based cGANs will collapse earlier, such as AC-GAN, Multi-hinge GAN, and Omni-GAN. It would be better if the author could explain this phenomenon. \n- I know that there is an improved version of AC-GAN (ImAC-GAN), as discussed in section 3.3 of the Omni-GAN paper. ImAC-GAN has a clear performance gain compared to AC-GAN. It would be better if the author could discuss the relationship between ImAC-GAN and ADC-GAN. \n- In Figure 4 (c) and (f), the results of T-SNE visualization are not very convincing. In my opinion, the author should not use the discriminator to extract feature representations, because the PD-GAN discriminator is less supervised than ADC-GAN. I suggest that the author use a pre-trained classification model to extract features for a fair comparison.\n- In the caption of Figure 4, the author says that the T-SNE uses training data, but in the last paragraph of section 5.3, the author says that the T-SNE uses the validation data. \n", "summary_of_the_review": "The author proposes a simple but seemingly promising classifier-based cGAN. I hope the author can answer my questions above in detail. I will improve the score based on the author's answer. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635761075351}], "openreview_url": "https://openreview.net/forum?id=Yn4CPz_LRKO", "arxiv_id": "2107.10060", "paper_pdf": "papers/Yn4CPz_LRKO.pdf", "paper_pdf_sha256": "6082ae521f88fe5a135412119572edf86d9fa40aad32849c5cc27253d2f8db6c", "paper_pdf_bytes": 1694234, "paper_pdf_source": "openreview", "code_url": "https://github.com/liang-hou/adcgan", "code_repository": "liang-hou/adcgan", "code_commit": "29ecfa74dff78286e5966035a2ec8c6e96e6b4e1", "code_archive": "repos/Yn4CPz_LRKO.zip", "code_archive_sha256": "e55345ca538a7bd1d6e712da66a5db6798ff9092116aaafdc7a96f020a368203", "code_archive_bytes": 2634450, "code_file_count": 48, "code_extensions": {".sh": 25, ".py": 23}, "github_disk_usage_kb": 2613, "github_languages": {"Python": 276530, "Shell": 14184}, "github_archived": false, "github_pushed_at": "2023-06-11T17:18:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cgans-with-auxiliary-discriminative"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VbCVU10R7K", "year": 2021, "status": "rejected", "title": "Offline policy selection under Uncertainty", "authors": ["Mengjiao Yang", "Bo Dai", "Ofir Nachum", "George Tucker", "Dale Schuurmans"], "authorids": ["~Mengjiao_Yang1", "~Bo_Dai1", "~Ofir_Nachum1", "~George_Tucker1", "~Dale_Schuurmans1"], "authors_source": "OpenReview API", "abstract": "The presence of uncertainty in policy evaluation  significantly complicates the process of policy ranking and selection in real-world settings. We formally consider offline policy selection as learning preferences over a set of policy prospects given a fixed experience dataset. While one can select or rank policies based on point estimates of their policy values or high-confidence intervals, access to the full distribution over one's belief of the policy value enables more flexible selection algorithms under a wider range of downstream evaluation metrics. We propose BayesDICE for estimating this belief distribution in terms of posteriors of distribution correction ratios derived from stochastic constraints (as opposed to explicit likelihood, which is not available). Empirically, BayesDICE is highly competitive to existing state-of-the-art approaches in confidence interval estimation. More importantly, we show how the belief distribution estimated by BayesDICE may be used to rank policies with respect to any arbitrary downstream policy selection metric, and we empirically demonstrate that this selection procedure significantly outperforms existing approaches, such as ranking policies according to mean or high-confidence lower bound value estimates.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "7E7iM8RDv4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2249/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a method BayesDICE to estimate posteriors over candidate policy values, which can be used for downstream policy selection. Specifically, the authors estimate the posteriors over the correction ratios for state-action pairs, which optimize a combined metric of a chance constraint from collected data and KL from the prior. Computationally, the authors demonstrate the advantages of their approach by having better performances in both coverage and power for policy evaluation and better downstream ranking with respect to different metrics for policy selection.\n\nI think the paper is well written and has made a good investigation of their approach. It would be better if the authors can talk about how different choices of prior influence the results. Besides, for policy selection, when we have a large pool of candidates, how would this be adapted to multiple testing or how would the number of candidates influence the results. Finally, solving eq.14 needs estimation of some expectation over the state action visitation, while the data itself is not independent, how would this influence the results?\n\nsome minor comments\n- section 3.1, the first displayed equation: should it be summing from t=0 to infinity?\n- typo, appendix B, the first sentence, `\"consider exploitin\" -> exploiting ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "good paper that brings new perspective for offline policy selection", "review": "This paper proposes a method BayesDICE to estimate posteriors over candidate policy values, which can be used for downstream policy selection. Specifically, the authors estimate the posteriors over the correction ratios for state-action pairs, which optimize a combined metric of a chance constraint from collected data and KL from the prior. Computationally, the authors demonstrate the advantages of their approach by having better performances in both coverage and power for policy evaluation and better downstream ranking with respect to different metrics for policy selection.\n\nI think the paper is well written and has made a good investigation of their approach. It would be better if the authors can talk about how different choices of prior influence the results. Besides, for policy selection, when we have a large pool of candidates, how would this be adapted to multiple testing or how would the number of candidates influence the results. Finally, solving eq.14 needs estimation of some expectation over the state action visitation, while the data itself is not independent, how would this influence the results?\n\nsome minor comments\n- section 3.1, the first displayed equation: should it be summing from t=0 to infinity?\n- typo, appendix B, the first sentence, `\"consider exploitin\" -> exploiting ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603760713060}, {"id": "9Bmv9QjvWh", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2249/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. Significance:\n- This paper studies off-policy model selection problem, which we aim to select a subset of policies based on different evaluation metric, such as top-k precision, top-k accuracy, etc. I am less convinced by the motivation of off-policy model selection, since if we want to find some policy to deploy, does off-policy learning address the problem? It would be great if the authors could give a more practical and clear motivation on this, especially compared with learning. \n\n2. Quality & Clarity:\n- This paper is clearly written and easy to read, though I feel some details in the Appendix should be move into the main paper, especially some setup for the experiments, like the number of overall policies to be evaluated, what is the f in the experiments?\n- The method is based on building posterior distribution of stationary distribution ratio, which tried to build a loss function of the Bellman residual on the average state-action visitation measure, and a KL regularization with the prior distribution. This idea is not new (see DualDice) however utilize them to learn the distribution of density ratio is novel. \n- I love the vision that when we have different evaluation metrics, we may need the whole distribution of the policy value to perform effective selection, and point estimate w. confidence interval may not be suffice in this setting. Regarding to this observation, I can see the importance/significance of this work.\n- The method does make sense to me, and the empirical evaluation is well done (though some problems listed below). Important metrics are examined, like coverage and length of interval. Also, I like the experiments that the authors examine different metrics, and show how point estimate w. confidence estimation gives sub-optimal performance, this makes the motivation much clearer, about why we need the whole posterior distribution.\n\n3. Questions:\n- Regarding to the scale of the method, do we need really to go though all permutations in Alg 1? For some structured score S, can we just use sorting? Also, i see the experiments, there is only 5 policy to select, I am concerned about the scalability of the method. Could you justify more this?\n- Just curious, any explanation about why it fails in some domain, say Taxi in Fig 2? It could be better to discuss more about when the method works, when it fails? Taxi in Fig 2 may be a good example to diagnose this?\n- In experiments, for baseline methods , the authors are using per-step IPS +/- conf level. However, I feel it is kind of unfair for baseline methods, since basically the BayesDICE are using marginalized IPS, is it possible to change the baseline estimate to marginalized IPS for a fair comparison?\n- Can you comment how you address the stochasticity in the reward?\n- Some typo, say Eq4.\n\nOverall, I like the problem it raises, that there exists some metric when a point estimate of policy value +/- conf level may not suffice to give optimal performance. The authors did a pretty good job in addressing this problem, though some problems remain and needs improvement. I am willing to improve my score if my questions are addressed. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper studies off-policy model selection under various evaluation metrics. Basically a posterior distribution of the stationary density ratio is estimated, then utilize marginalized IPS to estimate the distribution of policy value.", "review": "1. Significance:\n- This paper studies off-policy model selection problem, which we aim to select a subset of policies based on different evaluation metric, such as top-k precision, top-k accuracy, etc. I am less convinced by the motivation of off-policy model selection, since if we want to find some policy to deploy, does off-policy learning address the problem? It would be great if the authors could give a more practical and clear motivation on this, especially compared with learning. \n\n2. Quality & Clarity:\n- This paper is clearly written and easy to read, though I feel some details in the Appendix should be move into the main paper, especially some setup for the experiments, like the number of overall policies to be evaluated, what is the f in the experiments?\n- The method is based on building posterior distribution of stationary distribution ratio, which tried to build a loss function of the Bellman residual on the average state-action visitation measure, and a KL regularization with the prior distribution. This idea is not new (see DualDice) however utilize them to learn the distribution of density ratio is novel. \n- I love the vision that when we have different evaluation metrics, we may need the whole distribution of the policy value to perform effective selection, and point estimate w. confidence interval may not be suffice in this setting. Regarding to this observation, I can see the importance/significance of this work.\n- The method does make sense to me, and the empirical evaluation is well done (though some problems listed below). Important metrics are examined, like coverage and length of interval. Also, I like the experiments that the authors examine different metrics, and show how point estimate w. confidence estimation gives sub-optimal performance, this makes the motivation much clearer, about why we need the whole posterior distribution.\n\n3. Questions:\n- Regarding to the scale of the method, do we need really to go though all permutations in Alg 1? For some structured score S, can we just use sorting? Also, i see the experiments, there is only 5 policy to select, I am concerned about the scalability of the method. Could you justify more this?\n- Just curious, any explanation about why it fails in some domain, say Taxi in Fig 2? It could be better to discuss more about when the method works, when it fails? Taxi in Fig 2 may be a good example to diagnose this?\n- In experiments, for baseline methods , the authors are using per-step IPS +/- conf level. However, I feel it is kind of unfair for baseline methods, since basically the BayesDICE are using marginalized IPS, is it possible to change the baseline estimate to marginalized IPS for a fair comparison?\n- Can you comment how you address the stochasticity in the reward?\n- Some typo, say Eq4.\n\nOverall, I like the problem it raises, that there exists some metric when a point estimate of policy value +/- conf level may not suffice to give optimal performance. The authors did a pretty good job in addressing this problem, though some problems remain and needs improvement. I am willing to improve my score if my questions are addressed. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603541445021}, {"id": "0UoDJeBJvxC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2249/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "## Review\n\nGiven as set of pre-specified policies, this paper proposes a Bayesian method to estimate the posterior distribution of their average values, by estimating posterior distributions of their discounted stationary distribution ratios. These posterior distributions are used for off-policy evaluation in various ways.\n\n\n \n## Positives\n\n+ The idea focusing on estimating nonlinear functionals of multiple policy values is appealing.\n\n\n \n## Major concerns\n\n+ The results in Figure 2 are a bit surprising. We expect that methods based on concentration inequalities like Bernstein or student-t to be somewhat conservative, but the results suggest that their confidence intervals are extremely wide. For example, in the Bandit case, even after 200 samples, the interval log-width would suggest that Bernstein's confidence intervals is more than 7x the confidence intervals suggested by BayesDICE. What explains these results?\n\n+ Again on Figure 2, if \"Bernstein\" and \"Student t\" are unbiased methods, then having a very wide confidence interval should translate into over-coverage. However, they seem to be *under*-covering the true value. Are these methods somehow heavily biased? If not, what explains the under-coverage?\n\n+ The paper proposes a method for evaluating non-linear functionals of policy values, such as ranking scores over their values. However, it seems to me that in order to evaluate such nonlinear functions one would require knowledge about the *joint* distribution of values over all policies of interest. In the notation of the paper, one would require knowledge of $q(\\bar{\\rho}_1, ..., \\bar{\\rho}_N)$. However, it is not clear from the method description in Section 3.2 how one is able to estimate this joint distribution. Instead, it seems to me that all we get is $q(\\bar{\\rho}_i)$ for each policy $i$ -- that is, their marginal distributions. If that is the correct interpretation of what's going on in Section 3.2, then that raises the question of whether these distributions are independent. \n\n\n\n \n## Minor concerns\n\n+ In appendix C.1, I did not understand the description of the \"bandit\" environment. Are rewards binary?\n\n+ Several symbols are not formally defined. E.g., on page 4, (lowercase) r(s,a) is not defined.\n\n+ In Algorithm 1, what is the role of quantity L*? Why do we need it as a stopping rule?\n\n+ Some notes on exposition.\n\n  - The authors take some time to reveal what is their estimand --- the discounted stationary distribution ratios. As a reader, I would have benefited from having that explained much earlier, even before the conversation about ranking evaluation.\n\n  - Section 3.2: the authors could have dedicated some more space developing the intuition for their method (e.g. an abridged version of Nachum and Dai 2020), even if that meant relegating some of the mathematical details to the appendix. As it stands, the section makes the paper incomprehensible as a standalone piece of research.\n\n\n## Typos\n\n+ The indices on the sum in the definition of \"stationary visitation\" (p.4) are wrong. On the next line, the last conditioning should have been s[i+1] ~ T(.|s[i],a[i]) instead of s[i+1] ~ T(.|s[t],a[t]).\n+ Philip Thomas' \"High-confidence off- policy evaluation\" citation shows up twice in the bibliography.\n+ indentical --> identical (Pg. 5)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": " ", "review": "## Review\n\nGiven as set of pre-specified policies, this paper proposes a Bayesian method to estimate the posterior distribution of their average values, by estimating posterior distributions of their discounted stationary distribution ratios. These posterior distributions are used for off-policy evaluation in various ways.\n\n\n \n## Positives\n\n+ The idea focusing on estimating nonlinear functionals of multiple policy values is appealing.\n\n\n \n## Major concerns\n\n+ The results in Figure 2 are a bit surprising. We expect that methods based on concentration inequalities like Bernstein or student-t to be somewhat conservative, but the results suggest that their confidence intervals are extremely wide. For example, in the Bandit case, even after 200 samples, the interval log-width would suggest that Bernstein's confidence intervals is more than 7x the confidence intervals suggested by BayesDICE. What explains these results?\n\n+ Again on Figure 2, if \"Bernstein\" and \"Student t\" are unbiased methods, then having a very wide confidence interval should translate into over-coverage. However, they seem to be *under*-covering the true value. Are these methods somehow heavily biased? If not, what explains the under-coverage?\n\n+ The paper proposes a method for evaluating non-linear functionals of policy values, such as ranking scores over their values. However, it seems to me that in order to evaluate such nonlinear functions one would require knowledge about the *joint* distribution of values over all policies of interest. In the notation of the paper, one would require knowledge of $q(\\bar{\\rho}_1, ..., \\bar{\\rho}_N)$. However, it is not clear from the method description in Section 3.2 how one is able to estimate this joint distribution. Instead, it seems to me that all we get is $q(\\bar{\\rho}_i)$ for each policy $i$ -- that is, their marginal distributions. If that is the correct interpretation of what's going on in Section 3.2, then that raises the question of whether these distributions are independent. \n\n\n\n \n## Minor concerns\n\n+ In appendix C.1, I did not understand the description of the \"bandit\" environment. Are rewards binary?\n\n+ Several symbols are not formally defined. E.g., on page 4, (lowercase) r(s,a) is not defined.\n\n+ In Algorithm 1, what is the role of quantity L*? Why do we need it as a stopping rule?\n\n+ Some notes on exposition.\n\n  - The authors take some time to reveal what is their estimand --- the discounted stationary distribution ratios. As a reader, I would have benefited from having that explained much earlier, even before the conversation about ranking evaluation.\n\n  - Section 3.2: the authors could have dedicated some more space developing the intuition for their method (e.g. an abridged version of Nachum and Dai 2020), even if that meant relegating some of the mathematical details to the appendix. As it stands, the section makes the paper incomprehensible as a standalone piece of research.\n\n\n## Typos\n\n+ The indices on the sum in the definition of \"stationary visitation\" (p.4) are wrong. On the next line, the last conditioning should have been s[i+1] ~ T(.|s[i],a[i]) instead of s[i+1] ~ T(.|s[t],a[t]).\n+ Philip Thomas' \"High-confidence off- policy evaluation\" citation shows up twice in the bibliography.\n+ indentical --> identical (Pg. 5)", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603499270800}], "openreview_url": "https://openreview.net/forum?id=VbCVU10R7K", "arxiv_id": "2012.06919", "paper_pdf": "papers/VbCVU10R7K.pdf", "paper_pdf_sha256": "9e5ff9eba10b12dc490ce3399498f90b99c00ec05c8a07b5088c432ba98a1cfd", "paper_pdf_bytes": 5387397, "paper_pdf_source": "openreview", "code_url": "https://github.com/google-research/dice_rl", "code_repository": "google-research/dice_rl", "code_commit": "5ee67f7f2145d295bdc4759b62e5c6193af31acf", "code_archive": "repos/VbCVU10R7K.zip", "code_archive_sha256": "cc05b5fa6d97faaf799830bbfde82cd9de5afdaa9d595e8ff33781f76b60acb5", "code_archive_bytes": 1415580, "code_file_count": 98, "code_extensions": {".py": 96, ".sh": 2}, "github_disk_usage_kb": 1569, "github_languages": {"Python": 582029, "Shell": 1366}, "github_archived": false, "github_pushed_at": "2026-07-30T00:11:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/offline-policy-selection-under-uncertainty-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "KuhCUX2oIt", "year": 2026, "status": "rejected", "title": "LEAD: Large Foundation Model for EEG-Based Alzheimer’s Disease Detection", "authors": ["Yihe Wang", "Nan Huang", "Nadia Mammone", "Marco Cecchi", "Xiang Zhang"], "authorids": ["~Yihe_Wang2", "~Nan_Huang2", "~Nadia_Mammone1", "~Marco_Cecchi1", "~Xiang_Zhang10"], "authors_source": "OpenReview API", "abstract": "Electroencephalography (EEG) provides a non-invasive, highly accessible, and cost-effective approach for detecting Alzheimer’s disease (AD). However, existing methods, whether based on handcrafted feature engineering or standard deep learning, face two major challenges: 1) the lack of large-scale EEG-AD datasets for robust representation learning, and 2) the absence of a dedicated deep learning pipeline for subject-level detection, which is more clinically meaningful than the commonly used sample-level detection. To address these gaps, we have curated the world's largest EEG-AD corpus to date, comprising 2,255 subjects. Leveraging this unique data corpus, we propose LEAD, the first large-scale foundation model for EEG analysis in dementia. Our approach provides an innovative framework for subject-level AD detection, including: 1) a comprehensive preprocessing pipeline such as artifact removal, resampling, and filtering, and a newly proposed multi-scale segmentation strategy, 2) a subject-regularized spatio–temporal transformer trained with a novel subject-level cross-entropy loss and an indices group-shuffling algorithm, and 3) AD-guided contrastive pre-training. We pre-train on 12 datasets (3 AD-related and 9 non-AD) and fine-tune/test on 4 AD datasets. Compared with 10 baselines, LEAD consistently obtains superior subject-level detection performance under the challenging subject-independent cross-validation protocol. On the benchmark ADFTD dataset, our model achieves an impressive subject-level Sensitivity of 90.91\\% under the leave-one-subject-out (LOSO) setting. These results strongly validate the effectiveness of our method for real-world EEG-based AD detection. Source code: \\url{https://anonymous.4open.science/r/LEAD-3B51}", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "beF47swFf1", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1141/Reviewer_fDC2"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces LEAD, a large-scale model for EEG-based Alzheimer’s Disease detection that is pre-trained on 3 AD datasets and 9 non-AD datasets by employing both sample- and subject-level contrastive learning, followed by fine-tuning on 60% of subjects from 4 AD datasets with joint sample- and subject-level cross-entropy losses, while the remaining 40% are reserved for validation and evaluation. Their main contribution is integrating a multi-scale segmentation strategy, a subject-regularized spatio–temporal\ntransformer with both sample- and subject-level cross-entropy losses, and an indices group-shuffling algorithm to reduce differences between sample-level performance  and subject-level detection. The authors fine-tuned and tested the model on four datasets,  mainly report sensitivity and F1-scores and report performance improvements in comparison to 10 baseline results. In addition, the authors conducted an ablation study to evaluate the effectiveness of each domain-relevant pre-training dataset and an ablation study on the ADFTD dataset to assess the contribution of each module.", "review_text": "The paper introduces LEAD, a large-scale model for EEG-based Alzheimer’s Disease detection that is pre-trained on 3 AD datasets and 9 non-AD datasets by employing both sample- and subject-level contrastive learning, followed by fine-tuning on 60% of subjects from 4 AD datasets with joint sample- and subject-level cross-entropy losses, while the remaining 40% are reserved for validation and evaluation. Their main contribution is integrating a multi-scale segmentation strategy, a subject-regularized spatio–temporal\ntransformer with both sample- and subject-level cross-entropy losses, and an indices group-shuffling algorithm to reduce differences between sample-level performance  and subject-level detection. The authors fine-tuned and tested the model on four datasets,  mainly report sensitivity and F1-scores and report performance improvements in comparison to 10 baseline results. In addition, the authors conducted an ablation study to evaluate the effectiveness of each domain-relevant pre-training dataset and an ablation study on the ADFTD dataset to assess the contribution of each module.", "strengths": "* Performance comparison among different groups on the ADFTD dataset \n* The paper provides an ablation study of modules, and an ablation study of pre-training datasets. \n* binary and multi-class results are provided\n* Each evaluation is repeated with five random seeds reporting the mean and standard deviation\n* creating larger, open-source datasets from heterogeneous clinical populations  is important to address  the lack of large-scale EEG datasets and is essential to investigate cross-institutional generalisability.", "weaknesses": "* In addition, Leave-One-Subject-Out Analysis is provided, however it is limited to one dataset (ADFTD) and only for the LEAD model.\n* The information in the repository does not match that in the document. For example repository: 9AD and 7 non-AD, 5 AD for finetuning; Paper: 7AD, 9 non-AD, 4 AD for finetuning.\n* The authors compare the performance of the LEAD model to 10 baselines. Two models are foundation models, however the other models are not designed for AD detection. \n* A direct comparison of the results to previously published state-of-the-art results on the four datasets is not provided, nor is any information given about the typical range of classification performance for AD detection.\n* For a direct comparison, please provide the sensitivity for 95% specificity, and also other performance metrics, such as accuracy, balanced accuracy, specificity, PR curves, ROC curves, AUCs, etc.\ndetailed analysis is focused on the ADFTD dataset, no detailed analysis for the other three evaluation sets\n* Overall, multi-class classification performances (Sensitivity and F1-score) of LEAD are in the range of 60-77%, and for binary classification in the range of 60-93%. On the benchmark ADFTD dataset, the  model achieves a subject-level sensitivity of 90.91% under the leave-one-subject-out (LOSO) setting.  However, it is unclear how this compares to previous state-of-the-art results.\n* The baselines used in this work are mostly designed for other EEG tasks: \n  - EEGNet: BCI\n  - TST: multivariate time series datasets (heartbeat, Facedetection, etc; no EEG)\n  - EEGInception: Motor Imagery BCI\n  - EEGConformer: BCI motor imagery and emotion recognition paradigms\n  - BIOT: pre-trained on 6 datasets, fine-tuned on abnormal EEG detection\n  - Medformer: included AD detection, multi-class performance on ADFTD: 97.50% F1 score\n  - MedGNN:  medical time series classification: sample-based F1score ADFD dataset: 98.30%\n  - LaBraM: pre-trained foundation model evaluated on various EEG detection tasks, such as including abnormal detection, event type  classification, emotion recognition, and gait prediction\n  - CBraMod: pre-trained foundation model for various  BCI tasks\n\t\n* A statistical evaluation is missing.\n* It remains unclear what clinical improvements this approach provides, especially compared to previously introduced approaches.", "questions": "* Why are the ADFTD, CNBPM, Cognision, CAUEEG chosen as final dataset for fine-tuning and evaluation and not AD-Auditory, BrainLat and P-ADIC?\n* What are the criteria for selecting the baseline? Why did you not use DNN designed for AD detection, such as SteadyNet, MNet or ADformer?\n* Are the baselines trained using the subject-level cross-entropy loss and indices group-shuffling?\n* To what extent is the imbalance of the dataset taken into account in relation to the loss function?\n* How does the performance of LEAD compare to other models on datasets that have not been pre-trained on multiple datasets?\n* Combining multiple EEG datasets, there are several key factors that have to be considered, including data heterogeneity and acquisition variability, institutional biases (distribution mismatch), demographic disparities (dataset imbalance, physical characteristics), data governance and handling. What kind of methodological and technical solutions (e.g. data harmonisation) are applied to address these challenges?  \n\nSuggestions: \n\n* Include a discussion of the challenges faced when combining multiple EEG datasets and how they were addressed in this work.\n* For a more robust comparison, please include baselines that have been applied for AD detection, such as SteadyNet, ADformer or MNet, as well as a detailed comparison to previous published results on these four datasets.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces LEAD, a large-scale model for EEG-based Alzheimer’s Disease detection that is pre-trained on 3 AD datasets and 9 non-AD datasets by employing both sample- and subject-level contrastive learning, followed by fine-tuning on 60% of subjects from 4 AD datasets with joint sample- and subject-level cross-entropy losses, while the remaining 40% are reserved for validation and evaluation. Their main contribution is integrating a multi-scale segmentation strategy, a subject-regularized spatio–temporal\ntransformer with both sample- and subject-level cross-entropy losses, and an indices group-shuffling algorithm to reduce differences between sample-level performance  and subject-level detection. The authors fine-tuned and tested the model on four datasets,  mainly report sensitivity and F1-scores and report performance improvements in comparison to 10 baseline results. In addition, the authors conducted an ablation study to evaluate the effectiveness of each domain-relevant pre-training dataset and an ablation study on the ADFTD dataset to assess the contribution of each module.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "* Performance comparison among different groups on the ADFTD dataset \n* The paper provides an ablation study of modules, and an ablation study of pre-training datasets. \n* binary and multi-class results are provided\n* Each evaluation is repeated with five random seeds reporting the mean and standard deviation\n* creating larger, open-source datasets from heterogeneous clinical populations  is important to address  the lack of large-scale EEG datasets and is essential to investigate cross-institutional generalisability.", "weaknesses": "* In addition, Leave-One-Subject-Out Analysis is provided, however it is limited to one dataset (ADFTD) and only for the LEAD model.\n* The information in the repository does not match that in the document. For example repository: 9AD and 7 non-AD, 5 AD for finetuning; Paper: 7AD, 9 non-AD, 4 AD for finetuning.\n* The authors compare the performance of the LEAD model to 10 baselines. Two models are foundation models, however the other models are not designed for AD detection. \n* A direct comparison of the results to previously published state-of-the-art results on the four datasets is not provided, nor is any information given about the typical range of classification performance for AD detection.\n* For a direct comparison, please provide the sensitivity for 95% specificity, and also other performance metrics, such as accuracy, balanced accuracy, specificity, PR curves, ROC curves, AUCs, etc.\ndetailed analysis is focused on the ADFTD dataset, no detailed analysis for the other three evaluation sets\n* Overall, multi-class classification performances (Sensitivity and F1-score) of LEAD are in the range of 60-77%, and for binary classification in the range of 60-93%. On the benchmark ADFTD dataset, the  model achieves a subject-level sensitivity of 90.91% under the leave-one-subject-out (LOSO) setting.  However, it is unclear how this compares to previous state-of-the-art results.\n* The baselines used in this work are mostly designed for other EEG tasks: \n  - EEGNet: BCI\n  - TST: multivariate time series datasets (heartbeat, Facedetection, etc; no EEG)\n  - EEGInception: Motor Imagery BCI\n  - EEGConformer: BCI motor imagery and emotion recognition paradigms\n  - BIOT: pre-trained on 6 datasets, fine-tuned on abnormal EEG detection\n  - Medformer: included AD detection, multi-class performance on ADFTD: 97.50% F1 score\n  - MedGNN:  medical time series classification: sample-based F1score ADFD dataset: 98.30%\n  - LaBraM: pre-trained foundation model evaluated on various EEG detection tasks, such as including abnormal detection, event type  classification, emotion recognition, and gait prediction\n  - CBraMod: pre-trained foundation model for various  BCI tasks\n\t\n* A statistical evaluation is missing.\n* It remains unclear what clinical improvements this approach provides, especially compared to previously introduced approaches.", "questions": "* Why are the ADFTD, CNBPM, Cognision, CAUEEG chosen as final dataset for fine-tuning and evaluation and not AD-Auditory, BrainLat and P-ADIC?\n* What are the criteria for selecting the baseline? Why did you not use DNN designed for AD detection, such as SteadyNet, MNet or ADformer?\n* Are the baselines trained using the subject-level cross-entropy loss and indices group-shuffling?\n* To what extent is the imbalance of the dataset taken into account in relation to the loss function?\n* How does the performance of LEAD compare to other models on datasets that have not been pre-trained on multiple datasets?\n* Combining multiple EEG datasets, there are several key factors that have to be considered, including data heterogeneity and acquisition variability, institutional biases (distribution mismatch), demographic disparities (dataset imbalance, physical characteristics), data governance and handling. What kind of methodological and technical solutions (e.g. data harmonisation) are applied to address these challenges?  \n\nSuggestions: \n\n* Include a discussion of the challenges faced when combining multiple EEG datasets and how they were addressed in this work.\n* For a more robust comparison, please include baselines that have been applied for AD detection, such as SteadyNet, ADformer or MNet, as well as a detailed comparison to previous published results on these four datasets.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762009168710}, {"id": "NWX5P60WkN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1141/Reviewer_thwC"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces a foundation model trained to process EEG data from pathological individuals, with a focus on Alzheimer’s disease detection downstream tasks at the subject level. A transformer architecture with separate temporal and spatial branches is first pretrained using contrastive objectives at the sample and subject levels. The model is then finetuned to perform Alzheimer’s disease-related classification tasks such as pathology type or stage classification. For this, a corpus of public and private datasets (2255 subjects, 442h of recording) was curated and used to pretrain and evaluate performance of the model. Results suggest that in most cases the proposed model outperforms other pipelines based on handcrafted features, fully supervised architectures, or “generalist” EEG foundation models.", "review_text": "The paper introduces a foundation model trained to process EEG data from pathological individuals, with a focus on Alzheimer’s disease detection downstream tasks at the subject level. A transformer architecture with separate temporal and spatial branches is first pretrained using contrastive objectives at the sample and subject levels. The model is then finetuned to perform Alzheimer’s disease-related classification tasks such as pathology type or stage classification. For this, a corpus of public and private datasets (2255 subjects, 442h of recording) was curated and used to pretrain and evaluate performance of the model. Results suggest that in most cases the proposed model outperforms other pipelines based on handcrafted features, fully supervised architectures, or “generalist” EEG foundation models.", "strengths": "Originality\n* The paper presents the first EEG foundation model pretrained specifically on pathological data with a focus on Alzheimer’s disease. While there are multiple papers in the literature focused on building EEG foundation models, I am not aware of any that are specifically focused on this subtype of data.\n* Similarly, the curation and unified preprocessing of a large corpus of AD-specific datasets appears novel.\n\nQuality\n* The proposed method is compared to several relevant baselines (handcrafted features approach, supervised models, and two other “generalist” EEG foundation models), on 4 Alzheimer’s disease datasets from different sources. These comparisons suggest that the proposed model almost always outperforms baselines, which supports the claim of foundation model training.\n* The provided ablation studies suggest adding pretraining steps yield the largest performance increase, which supports the use of a pretraining-finetuning pipeline. The ablation study on the pretraining corpus is also informative and supports the use of the curated corpus.\n\nClarity\n* The paper is overall well written and easy to follow, and technical decisions are well supported.\n\nSignificance\n* Based on the reported strong results on the 4 downstream datasets, this paper is likely to be relevant for the EEG and Alzheimer’s disease prediction community. More generally, this can provide a strong point of comparison for evaluating “generalist” EEG foundation models on pathological EEG detection tasks.", "weaknesses": "1. While the baseline approaches are diverse, comparisons with architectures that have been specifically designed and/or tested on AD data in the past appear to be missing. Appendix A lists a few of them, which I believe are not considered in the results of Tables 3 and 7. Most notably, the last model to be listed, ADformer, presents many similarities to the proposed work both in terms of architecture and evaluation protocol. Similarly, it would be interesting to compare performance to existing EEG foundation models that have also tested their models on similar tasks, for instance [1]. See Q1.\n\n[1] Yuan, Zhizhang, et al. \"Brainwave: A brain signal foundation model for clinical applications.\" arXiv preprint arXiv:2402.10251 (2024).\n\n2. The reason for focusing on subject-level evaluation is well explained in the manuscript. However, it seems to me that the proposed approach relies on the same strategy as previous work to yield subject-level predictions, i.e. majority voting from sample-level predictions. Along with the limited performance increase yielded by the subject cross-entropy loss (Table 5), this weakens the listed contribution of building a novel subject-level detection pipeline.\n\n3. The contrastive learning strategy at the core of the paper is taken from a previous paper (Wang et al., 2023), which limits the novelty of this work on the algorithm side. Of note, this reference is only mentioned in the appendix, whereas it seems central enough that it should be cited in the main text.\n\n4. The preprocessing pipeline relies on heavy preprocessing steps, i.e. ICA and rejection of artifactual components, and requires adapting electrode montages to a common 19-channel montage. Given existing foundation models have shown promising results with limited or no preprocessing steps applied (e.g. LaBraM and CBraMod), with flexible montage support, this limits the flexibility of the proposed model in comparison (see Q2). Additionally, filtering the signals below 45 Hz may get rid of relevant information (see Q3).", "questions": "1. How does LEAD compare to existing deep learning approaches specifically designed for AD-related classification tasks?\n2. Have the authors evaluated the importance of the artifact correction step?\n3. Cutting off information in higher frequency bands (above 45 Hz) may be limiting as information in the gamma band has been reported as relevant for AD classification, e.g. in Wang et al. (2017). Why was this choice made?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a foundation model trained to process EEG data from pathological individuals, with a focus on Alzheimer’s disease detection downstream tasks at the subject level. A transformer architecture with separate temporal and spatial branches is first pretrained using contrastive objectives at the sample and subject levels. The model is then finetuned to perform Alzheimer’s disease-related classification tasks such as pathology type or stage classification. For this, a corpus of public and private datasets (2255 subjects, 442h of recording) was curated and used to pretrain and evaluate performance of the model. Results suggest that in most cases the proposed model outperforms other pipelines based on handcrafted features, fully supervised architectures, or “generalist” EEG foundation models.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Originality\n* The paper presents the first EEG foundation model pretrained specifically on pathological data with a focus on Alzheimer’s disease. While there are multiple papers in the literature focused on building EEG foundation models, I am not aware of any that are specifically focused on this subtype of data.\n* Similarly, the curation and unified preprocessing of a large corpus of AD-specific datasets appears novel.\n\nQuality\n* The proposed method is compared to several relevant baselines (handcrafted features approach, supervised models, and two other “generalist” EEG foundation models), on 4 Alzheimer’s disease datasets from different sources. These comparisons suggest that the proposed model almost always outperforms baselines, which supports the claim of foundation model training.\n* The provided ablation studies suggest adding pretraining steps yield the largest performance increase, which supports the use of a pretraining-finetuning pipeline. The ablation study on the pretraining corpus is also informative and supports the use of the curated corpus.\n\nClarity\n* The paper is overall well written and easy to follow, and technical decisions are well supported.\n\nSignificance\n* Based on the reported strong results on the 4 downstream datasets, this paper is likely to be relevant for the EEG and Alzheimer’s disease prediction community. More generally, this can provide a strong point of comparison for evaluating “generalist” EEG foundation models on pathological EEG detection tasks.", "weaknesses": "1. While the baseline approaches are diverse, comparisons with architectures that have been specifically designed and/or tested on AD data in the past appear to be missing. Appendix A lists a few of them, which I believe are not considered in the results of Tables 3 and 7. Most notably, the last model to be listed, ADformer, presents many similarities to the proposed work both in terms of architecture and evaluation protocol. Similarly, it would be interesting to compare performance to existing EEG foundation models that have also tested their models on similar tasks, for instance [1]. See Q1.\n\n[1] Yuan, Zhizhang, et al. \"Brainwave: A brain signal foundation model for clinical applications.\" arXiv preprint arXiv:2402.10251 (2024).\n\n2. The reason for focusing on subject-level evaluation is well explained in the manuscript. However, it seems to me that the proposed approach relies on the same strategy as previous work to yield subject-level predictions, i.e. majority voting from sample-level predictions. Along with the limited performance increase yielded by the subject cross-entropy loss (Table 5), this weakens the listed contribution of building a novel subject-level detection pipeline.\n\n3. The contrastive learning strategy at the core of the paper is taken from a previous paper (Wang et al., 2023), which limits the novelty of this work on the algorithm side. Of note, this reference is only mentioned in the appendix, whereas it seems central enough that it should be cited in the main text.\n\n4. The preprocessing pipeline relies on heavy preprocessing steps, i.e. ICA and rejection of artifactual components, and requires adapting electrode montages to a common 19-channel montage. Given existing foundation models have shown promising results with limited or no preprocessing steps applied (e.g. LaBraM and CBraMod), with flexible montage support, this limits the flexibility of the proposed model in comparison (see Q2). Additionally, filtering the signals below 45 Hz may get rid of relevant information (see Q3).", "questions": "1. How does LEAD compare to existing deep learning approaches specifically designed for AD-related classification tasks?\n2. Have the authors evaluated the importance of the artifact correction step?\n3. Cutting off information in higher frequency bands (above 45 Hz) may be limiting as information in the gamma band has been reported as relevant for AD classification, e.g. in Wang et al. (2017). Why was this choice made?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761987111031}, {"id": "3HTsSsJnCN", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1141/Reviewer_Dwx4"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes LEAD, a large-scale, subject-regularized spatio-temporal transformer for EEG-based Alzheimer’s disease detection. It achieves state-of-the-art subject-level performance across four AD datasets.", "review_text": "This paper proposes LEAD, a large-scale, subject-regularized spatio-temporal transformer for EEG-based Alzheimer’s disease detection. It achieves state-of-the-art subject-level performance across four AD datasets.", "strengths": "S1. The authors curate the largest unified EEG–AD dataset to date, comprising 2,255 subjects from heterogeneous sources, significantly expanding the available data scale for clinical modeling.\n\nS2. They introduce a subject-regularized spatio-temporal transformer with group shuffling and a subject-level loss to directly optimize subject-level consistency.\n\nS3. The proposed domain-relevant contrastive pretraining effectively enhances representation learning and downstream AD detection performance.", "weaknesses": "W1. Signal frequency and channel unification may cause information loss. To handle heterogeneous datasets, all recordings were resampled to 128 Hz and mapped to 19 channels. Although this ensures format consistency, it inevitably sacrifices high-frequency components and temporal resolution.\n\nW2. The COGNISION dataset (7 channels) was spatially interpolated to 19 channels, which achieves formal consistency but weakens the neurophysiological grounding. In addition, Figure 5 and the channel-importance visualization section do not specify their derivation or correspondence to AD-related regions, making the interpretation resemble a simple visual description rather than a physiologically meaningful analysis.\n\nW3. The model processes EEG in the time domain, no frequency-specific representations or analyses are provided. It remains unclear whether the learned patterns align with known AD-related spectral alterations.\n\nW4. The contrastive learning component shows limited methodological innovation. The subject-level contrast appears to simply cluster samples from the same subject rather than capture disease-specific dynamics, and the sample-level contrast largely reuses existing SimCLR/MoCo augmentations. \n\nW5. A few citation styles appear incorrect at sentence endings. The authors may consider using the \\citep{} format.", "questions": "Q1. Given the variability in diagnostic criteria and MMSE thresholds for AD/MCI/HC among institutions, how did the authors ensure consistent labeling across datasets such as AD-AUDITORY (https://openneuro.org/datasets/ds005048/versions/1.0.0) and ADFTD (https://openneuro.org/datasets/ds004504/versions/1.0.8)? Were any normalization or re-annotation steps applied?\n\nQ2. What is the rationale for fixing the configuration at 128 Hz and 19 channels? Was this choice made for engineering convenience, or was it supported by empirical analysis showing optimal performance?\n\nQ3. Since LEAD’s pretraining remains disease-oriented rather than task-agnostic, in what sense should it be regarded as a foundation model rather than a cross-disease mixed pretraining framework?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes LEAD, a large-scale, subject-regularized spatio-temporal transformer for EEG-based Alzheimer’s disease detection. It achieves state-of-the-art subject-level performance across four AD datasets.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "S1. The authors curate the largest unified EEG–AD dataset to date, comprising 2,255 subjects from heterogeneous sources, significantly expanding the available data scale for clinical modeling.\n\nS2. They introduce a subject-regularized spatio-temporal transformer with group shuffling and a subject-level loss to directly optimize subject-level consistency.\n\nS3. The proposed domain-relevant contrastive pretraining effectively enhances representation learning and downstream AD detection performance.", "weaknesses": "W1. Signal frequency and channel unification may cause information loss. To handle heterogeneous datasets, all recordings were resampled to 128 Hz and mapped to 19 channels. Although this ensures format consistency, it inevitably sacrifices high-frequency components and temporal resolution.\n\nW2. The COGNISION dataset (7 channels) was spatially interpolated to 19 channels, which achieves formal consistency but weakens the neurophysiological grounding. In addition, Figure 5 and the channel-importance visualization section do not specify their derivation or correspondence to AD-related regions, making the interpretation resemble a simple visual description rather than a physiologically meaningful analysis.\n\nW3. The model processes EEG in the time domain, no frequency-specific representations or analyses are provided. It remains unclear whether the learned patterns align with known AD-related spectral alterations.\n\nW4. The contrastive learning component shows limited methodological innovation. The subject-level contrast appears to simply cluster samples from the same subject rather than capture disease-specific dynamics, and the sample-level contrast largely reuses existing SimCLR/MoCo augmentations. \n\nW5. A few citation styles appear incorrect at sentence endings. The authors may consider using the \\citep{} format.", "questions": "Q1. Given the variability in diagnostic criteria and MMSE thresholds for AD/MCI/HC among institutions, how did the authors ensure consistent labeling across datasets such as AD-AUDITORY (https://openneuro.org/datasets/ds005048/versions/1.0.0) and ADFTD (https://openneuro.org/datasets/ds004504/versions/1.0.8)? Were any normalization or re-annotation steps applied?\n\nQ2. What is the rationale for fixing the configuration at 128 Hz and 19 channels? Was this choice made for engineering convenience, or was it supported by empirical analysis showing optimal performance?\n\nQ3. Since LEAD’s pretraining remains disease-oriented rather than task-agnostic, in what sense should it be regarded as a foundation model rather than a cross-disease mixed pretraining framework?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761986293496}, {"id": "vwqNm8JzNY", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission1141/Reviewer_qMSz"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces LEAD, a large-scale foundation model for EEG-based Alzheimer's Disease (AD) detection. The authors address two primary challenges in the field: the lack of large-scale EEG-AD datasets and the common focus on sample-level detection rather than the more clinically meaningful subject-level detection.", "review_text": "This paper introduces LEAD, a large-scale foundation model for EEG-based Alzheimer's Disease (AD) detection. The authors address two primary challenges in the field: the lack of large-scale EEG-AD datasets and the common focus on sample-level detection rather than the more clinically meaningful subject-level detection.", "strengths": "1. The proposed methodological additions are directly target the subject-level problem.\n2. The paper convincingly shows that its domain-relevant pre-training strategy (using ~730 hours of data) can outperform foundation models pre-trained on much larger, more general EEG corpora (LaBraM at ~2,000 hours, CBraMod at ~9,000 hours). This highlights the importance of data selection for pre-training.", "weaknesses": "1. The paper frames itself as the 'first foundation model for EEG-based AD detection'. This premise is arguably a contradiction in terms. Foundation models (such as the baselines LaBraM and CBraMod ) are intended to learn general, robust, and task-agnostic representations from massive, broad data. Specific tasks, like AD detection, are meant to be downstream applications solved by fine-tuning these general models. The authors' approach, curating a large but highly specific pre-training corpus of 'domain-relevant' neurological disorders, is a form of specialized transfer learning. It does not constitute the creation of a new 'foundation model' in the accepted sense. It is, rather, a highly effective domain-specific model, and this framing overstates the generality of its contribution.\n\n2. The core technical contribution beyond the (non-trivial) data engineering appears to be incremental. The proposed model is an adaptation of a spatio-temporal transformer, and the key novel components (subject-level loss, index shuffling) are application-specific engineering solutions. \n\n3. More concerningly, the value of the architectural changes is directly contradicted by the paper's own ablation study.\n* The \"Vanilla Backbone\" alone achieves a strong subject-level F1 score of 73.11%. The full LEAD pipeline, with all modules, achieves 75.77%, a marginal improvement of only ~2.66%.\n* Adding \"Half-Overlapping\" segmentation causes performance to drop from 73.11% to 72.12%. Later, adding \"Multi-Scale Segmentation\" also causes a performance drop from 75.49% to 74.96%, which suggests the architectural components are, at best, offering a negligible and inconsistent benefit. The vast majority of the gain  comes from \"+Sample/Subject-Level Pre-training\", reinforcing that the contribution is in the pre-training data strategy, not the LEAD architecture itself.\n\n4.  For a tool aimed at Alzheimer's Disease, the most vital clinical task is early detection, which requires distinguishing MCI from HC and AD. The model's performance on this task is weak.\n* On the multi-class datasets that include MCI, the subject-level F1 scores are low (~64.3% and ~59.6%, respectively).\n* On the largest dataset, CAUEEG (1,379 subjects), the LEAD model's subject-level F1 score (59.61%) is statistically indistinct from, several baselines, such as LaBraM (58.92%).\n\n5. The abstract states the model achieves a \"Sensitivity of 90.91%\" on the LOSO benchmark , but Table 9 and Appendix F.2 refer to this 90.91% value as the \"Subject-level Accuracy\". Could the authors please clarify which metric this value represents?", "questions": "Please check above. I will reconsider my assessment after checking authors' response.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces LEAD, a large-scale foundation model for EEG-based Alzheimer's Disease (AD) detection. The authors address two primary challenges in the field: the lack of large-scale EEG-AD datasets and the common focus on sample-level detection rather than the more clinically meaningful subject-level detection.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The proposed methodological additions are directly target the subject-level problem.\n2. The paper convincingly shows that its domain-relevant pre-training strategy (using ~730 hours of data) can outperform foundation models pre-trained on much larger, more general EEG corpora (LaBraM at ~2,000 hours, CBraMod at ~9,000 hours). This highlights the importance of data selection for pre-training.", "weaknesses": "1. The paper frames itself as the 'first foundation model for EEG-based AD detection'. This premise is arguably a contradiction in terms. Foundation models (such as the baselines LaBraM and CBraMod ) are intended to learn general, robust, and task-agnostic representations from massive, broad data. Specific tasks, like AD detection, are meant to be downstream applications solved by fine-tuning these general models. The authors' approach, curating a large but highly specific pre-training corpus of 'domain-relevant' neurological disorders, is a form of specialized transfer learning. It does not constitute the creation of a new 'foundation model' in the accepted sense. It is, rather, a highly effective domain-specific model, and this framing overstates the generality of its contribution.\n\n2. The core technical contribution beyond the (non-trivial) data engineering appears to be incremental. The proposed model is an adaptation of a spatio-temporal transformer, and the key novel components (subject-level loss, index shuffling) are application-specific engineering solutions. \n\n3. More concerningly, the value of the architectural changes is directly contradicted by the paper's own ablation study.\n* The \"Vanilla Backbone\" alone achieves a strong subject-level F1 score of 73.11%. The full LEAD pipeline, with all modules, achieves 75.77%, a marginal improvement of only ~2.66%.\n* Adding \"Half-Overlapping\" segmentation causes performance to drop from 73.11% to 72.12%. Later, adding \"Multi-Scale Segmentation\" also causes a performance drop from 75.49% to 74.96%, which suggests the architectural components are, at best, offering a negligible and inconsistent benefit. The vast majority of the gain  comes from \"+Sample/Subject-Level Pre-training\", reinforcing that the contribution is in the pre-training data strategy, not the LEAD architecture itself.\n\n4.  For a tool aimed at Alzheimer's Disease, the most vital clinical task is early detection, which requires distinguishing MCI from HC and AD. The model's performance on this task is weak.\n* On the multi-class datasets that include MCI, the subject-level F1 scores are low (~64.3% and ~59.6%, respectively).\n* On the largest dataset, CAUEEG (1,379 subjects), the LEAD model's subject-level F1 score (59.61%) is statistically indistinct from, several baselines, such as LaBraM (58.92%).\n\n5. The abstract states the model achieves a \"Sensitivity of 90.91%\" on the LOSO benchmark , but Table 9 and Appendix F.2 refer to this 90.91% value as the \"Subject-level Accuracy\". Could the authors please clarify which metric this value represents?", "questions": "Please check above. I will reconsider my assessment after checking authors' response.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761686977185}], "openreview_url": "https://openreview.net/forum?id=KuhCUX2oIt", "arxiv_id": "2502.01678", "paper_pdf": "papers/KuhCUX2oIt.pdf", "paper_pdf_sha256": "28d845a6a3c2bf6aaad13d787b5a3f87b6d8778abed3b42124a598a427286cdf", "paper_pdf_bytes": 2148190, "paper_pdf_source": "openreview", "code_url": "https://github.com/DL4mHealth/LEAD", "code_repository": "DL4mHealth/LEAD", "code_commit": "ec35aadb1bc068fcacdcd8d036964514ca2a708f", "code_archive": "repos/KuhCUX2oIt.zip", "code_archive_sha256": "1b5c8b68a7b4f426f3e86a2fe6af37791322193010525516be45fd384bdf1a08", "code_archive_bytes": 3205704, "code_file_count": 114, "code_extensions": {".py": 68, ".sh": 26, ".ipynb": 20}, "github_disk_usage_kb": 5765, "github_languages": {"Jupyter Notebook": 22251065, "Python": 451390, "Shell": 52879}, "github_archived": false, "github_pushed_at": "2026-04-01T07:40:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/lead-large-foundation-model-for-eeg-based"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3Wuvqc4xoy", "year": 2025, "status": "rejected", "title": "Learning Efficient Representations of Neutrino Telescope Events", "authors": ["Felix J. Yu", "Nicholas Kamp", "Carlos A. Argüelles"], "authorids": ["~Felix_J._Yu1", "~Nicholas_Kamp1", "~Carlos_A._Argüelles1"], "authors_source": "OpenReview API", "abstract": "Neutrino telescopes detect rare interactions of particles produced in some of the most extreme environments in the Universe. This is accomplished by instrumenting a cubic-kilometer volume of naturally occurring transparent medium with light sensors. Given their substantial size and the high frequency of background interactions, these telescopes amass an enormous quantity of large variance, high-dimensional data. These attributes create substantial challenges for analyzing and reconstructing interactions, particularly when utilizing machine learning (ML) techniques. In this paper, we present a novel approach, called om2vec, that employs transformer-based variational autoencoders to efficiently represent neutrino telescope events by learning compact and descriptive latent representations. We demonstrate that these latent representations offer enhanced flexibility and improved computational efficiency, thereby facilitating downstream tasks in data analysis.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "jsMRINbUO2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5289/Reviewer_mVDH"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This article presents an approach to learning representations of neutrino events by leveraging a transformer-based variational autoencoder. The model is trained to capture the photon arrival time distribution, and the learned representations are evaluated using the Jensen-Shannon divergence to assess reconstruction quality. Furthermore, the authors explore the applicability of these representations in a downstream task – angular reconstruction.", "review_text": "This article presents an approach to learning representations of neutrino events by leveraging a transformer-based variational autoencoder. The model is trained to capture the photon arrival time distribution, and the learned representations are evaluated using the Jensen-Shannon divergence to assess reconstruction quality. Furthermore, the authors explore the applicability of these representations in a downstream task – angular reconstruction.", "strengths": "The application of machine learning techniques in scientific research is a vital and rapidly evolving field. We are delighted to see submissions in this area and encourage researchers to share their relevant work.", "weaknesses": "This article requires significant improvements in its writing and technical accuracy. Numerous technical details are either unclear, incorrect, or require further clarification (see Questions for specific concerns). As it stands, the article's technical clarity is compromised, which may lead to confusion and misinterpretation. A thorough revision is necessary to ensure the article's technical details are accurate, clear, and concise.", "questions": "* In Fig. 2, the “autoencoder” outputs some probabilities through the softmax activation. This is a confusing design. How is the reconstruction loss applied in this case? \n* In section 4.2, the training methodology for the three models and the utilization of om2vec are unclear. Can you provide a more detailed explanation of the training process and how om2vec is incorporated?\n* Are there any additional physics features that could be included in the time series data, beyond the current single feature of photon hits?\n* In lines 179-180, the authors wrote “We opted for a learnable memory embedding for the transformer decoder layers, ensuring that the decoder portion of the architecture remains entirely independent of the encoder”. Please elaborate on the memory embedding block about its design.\n* The model and training details in Table 1 are incomplete and unclear. Can you provide a more comprehensive description of the model architecture, including the number of encoder and decoder layers used?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article presents an approach to learning representations of neutrino events by leveraging a transformer-based variational autoencoder. The model is trained to capture the photon arrival time distribution, and the learned representations are evaluated using the Jensen-Shannon divergence to assess reconstruction quality. Furthermore, the authors explore the applicability of these representations in a downstream task – angular reconstruction.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "The application of machine learning techniques in scientific research is a vital and rapidly evolving field. We are delighted to see submissions in this area and encourage researchers to share their relevant work.", "weaknesses": "This article requires significant improvements in its writing and technical accuracy. Numerous technical details are either unclear, incorrect, or require further clarification (see Questions for specific concerns). As it stands, the article's technical clarity is compromised, which may lead to confusion and misinterpretation. A thorough revision is necessary to ensure the article's technical details are accurate, clear, and concise.", "questions": "* In Fig. 2, the “autoencoder” outputs some probabilities through the softmax activation. This is a confusing design. How is the reconstruction loss applied in this case? \n* In section 4.2, the training methodology for the three models and the utilization of om2vec are unclear. Can you provide a more detailed explanation of the training process and how om2vec is incorporated?\n* Are there any additional physics features that could be included in the time series data, beyond the current single feature of photon hits?\n* In lines 179-180, the authors wrote “We opted for a learnable memory embedding for the transformer decoder layers, ensuring that the decoder portion of the architecture remains entirely independent of the encoder”. Please elaborate on the memory embedding block about its design.\n* The model and training details in Table 1 are incomplete and unclear. Can you provide a more comprehensive description of the model architecture, including the number of encoder and decoder layers used?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730711128930}, {"id": "TEvLddXVWv", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5289/Reviewer_nfb2"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 4, "confidence": 4, "summary": "This paper presents om2vec, a novel approach leveraging transformer-based variational autoencoders (VAEs) to create compact, descriptive latent representations of photon arrival time distributions (PATDs) from neutrino telescope events. The proposed model is designed to handle the high-dimensional, variable-length data typical of neutrino observatories like IceCube. om2vec aims to outperform conventional approaches, such as asymmetric Gaussian mixture models (AGMMs), by improving reconstruction accuracy, runtime efficiency, and reliability while being less dependent on hyperparameters. The paper details the architecture, training, and testing with simulated datasets, comparing the method’s performance with traditional AGMMs and exploring its utility for downstream tasks like angular reconstruction.", "review_text": "This paper presents om2vec, a novel approach leveraging transformer-based variational autoencoders (VAEs) to create compact, descriptive latent representations of photon arrival time distributions (PATDs) from neutrino telescope events. The proposed model is designed to handle the high-dimensional, variable-length data typical of neutrino observatories like IceCube. om2vec aims to outperform conventional approaches, such as asymmetric Gaussian mixture models (AGMMs), by improving reconstruction accuracy, runtime efficiency, and reliability while being less dependent on hyperparameters. The paper details the architecture, training, and testing with simulated datasets, comparing the method’s performance with traditional AGMMs and exploring its utility for downstream tasks like angular reconstruction.", "strengths": "- Originality: Applying transformer-based VAEs to neutrino event data is novel and demonstrates a creative extension of ML techniques to physical sciences.\n- Quality: Comprehensive evaluation of the model against AGMMs, showing significant improvements in reconstruction accuracy, computational efficiency, and robustness.\n- Clarity: The architectural details, data processing steps, and experimental methods are described with clarity, making the paper accessible to readers familiar with ML and neutrino physics.\n- Significance: The ability to improve data processing and enable better downstream analyses has substantial implications for neutrino research and potentially for other high-dimensional physics datasets.", "weaknesses": "- Generalizability: While the results are promising, it would be helpful to see a more extensive discussion on how the method might generalize across different types of neutrino observatories or non-simulated real-world data.\n- Comparison Baseline: Although om2vec is compared with AGMMs, additional comparisons with other potential ML approaches (e.g., deep CNNs or LSTMs) for PATD representation might strengthen the case for its use.\n- Hyperparameter Sensitivity: While the model claims reduced dependence on hyperparameters, an exploration of performance variability with different encoder/decoder block configurations or latent dimension sizes would provide deeper insights into its stability.", "questions": "1. How does the model’s performance vary with different encoder/decoder block architectures or deeper networks?\n2. Can the approach be adapted or extended to handle data from other types of particle physics experiments with different signal characteristics?\n3. Have real-world data tests been considered, and if so, what were the challenges and results?\n4. Is there potential for this method to contribute to real-time data processing in neutrino observatories under field conditions?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents om2vec, a novel approach leveraging transformer-based variational autoencoders (VAEs) to create compact, descriptive latent representations of photon arrival time distributions (PATDs) from neutrino telescope events. The proposed model is designed to handle the high-dimensional, variable-length data typical of neutrino observatories like IceCube. om2vec aims to outperform conventional approaches, such as asymmetric Gaussian mixture models (AGMMs), by improving reconstruction accuracy, runtime efficiency, and reliability while being less dependent on hyperparameters. The paper details the architecture, training, and testing with simulated datasets, comparing the method’s performance with traditional AGMMs and exploring its utility for downstream tasks like angular reconstruction.", "soundness": 4, "presentation": 4, "contribution": 4, "strengths": "- Originality: Applying transformer-based VAEs to neutrino event data is novel and demonstrates a creative extension of ML techniques to physical sciences.\n- Quality: Comprehensive evaluation of the model against AGMMs, showing significant improvements in reconstruction accuracy, computational efficiency, and robustness.\n- Clarity: The architectural details, data processing steps, and experimental methods are described with clarity, making the paper accessible to readers familiar with ML and neutrino physics.\n- Significance: The ability to improve data processing and enable better downstream analyses has substantial implications for neutrino research and potentially for other high-dimensional physics datasets.", "weaknesses": "- Generalizability: While the results are promising, it would be helpful to see a more extensive discussion on how the method might generalize across different types of neutrino observatories or non-simulated real-world data.\n- Comparison Baseline: Although om2vec is compared with AGMMs, additional comparisons with other potential ML approaches (e.g., deep CNNs or LSTMs) for PATD representation might strengthen the case for its use.\n- Hyperparameter Sensitivity: While the model claims reduced dependence on hyperparameters, an exploration of performance variability with different encoder/decoder block configurations or latent dimension sizes would provide deeper insights into its stability.", "questions": "1. How does the model’s performance vary with different encoder/decoder block architectures or deeper networks?\n2. Can the approach be adapted or extended to handle data from other types of particle physics experiments with different signal characteristics?\n3. Have real-world data tests been considered, and if so, what were the challenges and results?\n4. Is there potential for this method to contribute to real-time data processing in neutrino observatories under field conditions?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730691745659}, {"id": "SqqE2Se1X0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5289/Reviewer_sTJ4"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This develops a variational autoencoder to create a generative model for data produced by neutrino telescopes. The architecture is based on transformers, and results in a flexible representation and improved computation.", "review_text": "This develops a variational autoencoder to create a generative model for data produced by neutrino telescopes. The architecture is based on transformers, and results in a flexible representation and improved computation.", "strengths": "The application is certainly interesting and compelling. I also like the rationale of the work. There's a clear scientific motivation for these problems.", "weaknesses": "Several aspects. First, this is an ML focused conference so I would have appreciated greater details on the encoder and decoder without having to dig through the source code. Why transformers as opposed to a simpler architecture? Is there some kind transformation of the features that would allow for an MLP. Even if not, I would appreciate these as baselines as opposed to a traditional statistical model when comparing performance.\n\nAlso having worked with these a lot, I'm willing to bet that there was a substantial amount of tweaking required for learning rate and architecture parameters. If not, I'm certain performance can be improved dramatically by taking these steps. Another example, the runtime isn't really compelling to me. This is a feed-forward network, clearly it's going to be quicker than the alternatives. Should be supplementary, which would make more space for the fitting details I discussed.\n\nOverall, this seems written for a scientific audience rather than an ML audience. I very much appreciate the application and clear motivation so I hope it's resubmitted. It just seems like some of the details we find interesting were glossed over and need to be improved for this to be accepted.", "questions": "Not at the moment, will see other reviewers' comments.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This develops a variational autoencoder to create a generative model for data produced by neutrino telescopes. The architecture is based on transformers, and results in a flexible representation and improved computation.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The application is certainly interesting and compelling. I also like the rationale of the work. There's a clear scientific motivation for these problems.", "weaknesses": "Several aspects. First, this is an ML focused conference so I would have appreciated greater details on the encoder and decoder without having to dig through the source code. Why transformers as opposed to a simpler architecture? Is there some kind transformation of the features that would allow for an MLP. Even if not, I would appreciate these as baselines as opposed to a traditional statistical model when comparing performance.\n\nAlso having worked with these a lot, I'm willing to bet that there was a substantial amount of tweaking required for learning rate and architecture parameters. If not, I'm certain performance can be improved dramatically by taking these steps. Another example, the runtime isn't really compelling to me. This is a feed-forward network, clearly it's going to be quicker than the alternatives. Should be supplementary, which would make more space for the fitting details I discussed.\n\nOverall, this seems written for a scientific audience rather than an ML audience. I very much appreciate the application and clear motivation so I hope it's resubmitted. It just seems like some of the details we find interesting were glossed over and need to be improved for this to be accepted.", "questions": "Not at the moment, will see other reviewers' comments.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730580740999}, {"id": "mwqALXvIRj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5289/Reviewer_nXo1"], "rating": 1, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "The paper titled \"Learning Efficient Representations of Neutrino Telescope Events\" introduces a novel approach called om2vec, which utilizes transformer-based variational autoencoders (VAEs) to effectively represent neutrino telescope events. The study addresses the challenges posed by high-dimensional, variable-length Photon arrival time distributions (PATDs) recorded by optical modules in neutrino telescopes, particularly focusing on the IceCube Neutrino Observatory.", "review_text": "The paper titled \"Learning Efficient Representations of Neutrino Telescope Events\" introduces a novel approach called om2vec, which utilizes transformer-based variational autoencoders (VAEs) to effectively represent neutrino telescope events. The study addresses the challenges posed by high-dimensional, variable-length Photon arrival time distributions (PATDs) recorded by optical modules in neutrino telescopes, particularly focusing on the IceCube Neutrino Observatory.", "strengths": "- The use of a transformer-based variational autoencoder (VAE), called om2vec, represents an innovative approach for neutrino event data analysis, which has traditionally relied on more conventional statistical methods or simple summary statistics. \n- The paper pushes the boundaries of machine learning applications within high-energy physics, specifically neutrino detection. \n- By applying a VAE with transformer components to a unique scientific data source, the paper contributes to bridging techniques between disciplines, such as physics, machine learning, and data science. This could encourage further cross-disciplinary research and adaptation of machine learning models to complex scientific problems.", "weaknesses": "- The paper lacks a clear structure and does not adequately address related work. If this is indeed the first study applying deep learning techniques to the domain of neutrino telescopes, it is essential to include a dedicated **Related Works** section to provide context for this research.\n\n- The figures in the paper are oversized. I recommend the authors resize them to a more standard dimension to enhance the overall presentation quality. The current size does not meet the standards expected for conference presentations.\n\n- There are several typographical errors throughout the paper (e.g., lines 127, 484, etc.), which detract from its readability and should be addressed to improve clarity.\n\n- The objective function is unclear, and the problem is not well-defined. The paper jumps directly to the results, with only a brief discussion of the classical $KL$ divergence. A significant improvement is needed in presenting a comprehensive **Proposed Methods** section that clearly defines the final objective function, rather than merely referring to it in the **Results section** (lines 228 to 230).\n\n- Some statements in the paper are ambiguous or inaccurate. For example, the assertion in lines 223 to 232 that \"the re-parameterization trick is utilized to construct the latent representation $z$, a vector of user-defined length referred to as the latent dimension. This technique guarantees that the latent space remains continuous and that similar representations within this space reconstruct to similar PATDs\" is misleading and not entirely accurate.  However, the reparameterization trick separates the randomness of sampling (handled by $\\epsilon$) from the parameters $\\mu$ and $\\sigma$, which allows to compute gradients with respect to these parameters. I recommend that the authors deepen their understanding of this concept from this paper [1].\n\nI would be willing to consider increasing my rating, but only if these issues are adequately addressed. As it stands, the current version of the paper is not ready for publication.\n\n**Refrences:**\n\n[1] Kingma, Diederik P., and Max Welling. \"An introduction to variational autoencoders.\" Foundations and Trends® in Machine Learning 12.4 (2019): 307-392", "questions": "The paper is somewhat limited as it presents results solely based on training and testing with simulated events, which may not accurately reflect real-world measurement data. Given that the approach uses a VAE-based transformer, it may perform better with simulated data that follows known distributions. Do you have access to any existing real-world datasets? If so, I would appreciate your feedback on this aspect.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper titled \"Learning Efficient Representations of Neutrino Telescope Events\" introduces a novel approach called om2vec, which utilizes transformer-based variational autoencoders (VAEs) to effectively represent neutrino telescope events. The study addresses the challenges posed by high-dimensional, variable-length Photon arrival time distributions (PATDs) recorded by optical modules in neutrino telescopes, particularly focusing on the IceCube Neutrino Observatory.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "- The use of a transformer-based variational autoencoder (VAE), called om2vec, represents an innovative approach for neutrino event data analysis, which has traditionally relied on more conventional statistical methods or simple summary statistics. \n- The paper pushes the boundaries of machine learning applications within high-energy physics, specifically neutrino detection. \n- By applying a VAE with transformer components to a unique scientific data source, the paper contributes to bridging techniques between disciplines, such as physics, machine learning, and data science. This could encourage further cross-disciplinary research and adaptation of machine learning models to complex scientific problems.", "weaknesses": "- The paper lacks a clear structure and does not adequately address related work. If this is indeed the first study applying deep learning techniques to the domain of neutrino telescopes, it is essential to include a dedicated **Related Works** section to provide context for this research.\n\n- The figures in the paper are oversized. I recommend the authors resize them to a more standard dimension to enhance the overall presentation quality. The current size does not meet the standards expected for conference presentations.\n\n- There are several typographical errors throughout the paper (e.g., lines 127, 484, etc.), which detract from its readability and should be addressed to improve clarity.\n\n- The objective function is unclear, and the problem is not well-defined. The paper jumps directly to the results, with only a brief discussion of the classical $KL$ divergence. A significant improvement is needed in presenting a comprehensive **Proposed Methods** section that clearly defines the final objective function, rather than merely referring to it in the **Results section** (lines 228 to 230).\n\n- Some statements in the paper are ambiguous or inaccurate. For example, the assertion in lines 223 to 232 that \"the re-parameterization trick is utilized to construct the latent representation $z$, a vector of user-defined length referred to as the latent dimension. This technique guarantees that the latent space remains continuous and that similar representations within this space reconstruct to similar PATDs\" is misleading and not entirely accurate.  However, the reparameterization trick separates the randomness of sampling (handled by $\\epsilon$) from the parameters $\\mu$ and $\\sigma$, which allows to compute gradients with respect to these parameters. I recommend that the authors deepen their understanding of this concept from this paper [1].\n\nI would be willing to consider increasing my rating, but only if these issues are adequately addressed. As it stands, the current version of the paper is not ready for publication.\n\n**Refrences:**\n\n[1] Kingma, Diederik P., and Max Welling. \"An introduction to variational autoencoders.\" Foundations and Trends® in Machine Learning 12.4 (2019): 307-392", "questions": "The paper is somewhat limited as it presents results solely based on training and testing with simulated events, which may not accurately reflect real-world measurement data. Given that the approach uses a VAE-based transformer, it may perform better with simulated data that follows known distributions. Do you have access to any existing real-world datasets? If so, I would appreciate your feedback on this aspect.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730580610851}], "openreview_url": "https://openreview.net/forum?id=3Wuvqc4xoy", "arxiv_id": "2410.13148", "paper_pdf": "papers/3Wuvqc4xoy.pdf", "paper_pdf_sha256": "b4a75aab91c7fe43593e28061037dd529d35043e8951855b5cdac92af7dfb66e", "paper_pdf_bytes": 3815067, "paper_pdf_source": "openreview", "code_url": "https://github.com/felixyu7/om2vec", "code_repository": "felixyu7/om2vec", "code_commit": "60394e1dc5293126e32e4dab8510eca1f4fcbd12", "code_archive": "repos/3Wuvqc4xoy.zip", "code_archive_sha256": "0f36d3f6efe9befbc2d6b29ea75b5b6e6eeb9c482c1125084d1c1d8da82316a8", "code_archive_bytes": 14083, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 344, "github_languages": {"Python": 31260}, "github_archived": false, "github_pushed_at": "2025-07-07T19:26:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-efficient-representations-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UTGv8CayNt", "year": 2024, "status": "rejected", "title": "Chain-of-Thought Predictive Control", "authors": ["Zhiwei Jia", "Vineet Thumuluri", "Fangchen Liu", "Linghao Chen", "Zhiao Huang", "Hao Su"], "authorids": ["~Zhiwei_Jia1", "~Vineet_Thumuluri1", "~Fangchen_Liu2", "~Linghao_Chen2", "~Zhiao_Huang1", "~Hao_Su1"], "authors_source": "OpenReview API", "abstract": "We study generalizable policy learning from demonstrations for complex low-level control tasks (e.g., contact-rich object manipulations). We propose a novel hierarchical imitation learning method that utilizes scalable, albeit sub-optimal, demonstrations. Firstly, we propose an observation space-agnostic approach that efficiently discovers the multi-step subgoal decomposition (sequences of key observations) of the demos in an unsupervised manner. By grouping temporarily close and functionally similar actions into subskill-level segments, the discovered breakpoints (the segment boundaries) constitute a chain of planning steps (i.e., the chain-of-thought) to complete the task. Next, we propose a Transformer-based design that effectively learns to predict the chain-of-thought (CoT) as the high-level guidance for low-level action. We couple action and CoT predictions via prompt tokens and a hybrid masking strategy, which enable dynamically updated CoT guidance at test time and improve feature representation of the trajectory for generalizable policy learning. Our method, named Chain-of-Thought Predictive Control (CoTPC), consistently surpasses existing strong baselines on a wide range of challenging low-level manipulation tasks with scalable yet sub-optimal demos.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "vG9dBVqa5G", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9232/Reviewer_2mSo"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work proposes CoTPC, a Transformer-based architecture that performs hierarchical planning.\nAn important part of this method is the unsupervised discovery of subgoals, which assume that temporally close and similar actions belong to the same subskills.\nThe overall architecture uses learned subgoal embeddings and uses the goals discovered by the unsupervised algorithm to train these learned embeddings, as an auxiliary loss.\nThe remainder of the architecture fits into the family of the Behaviour transformer, with some design differences that impact the performance, and utilize the CoT learned embeddings.\nExperiments on the Moving Maze, Franka Kitchen and ManiSkill2 environments show the effectiveness of CoTPC over baselines.\nAn ablation study is also presented to showcase the importance of different design choices in the architecture.", "review_text": "This work proposes CoTPC, a Transformer-based architecture that performs hierarchical planning.\nAn important part of this method is the unsupervised discovery of subgoals, which assume that temporally close and similar actions belong to the same subskills.\nThe overall architecture uses learned subgoal embeddings and uses the goals discovered by the unsupervised algorithm to train these learned embeddings, as an auxiliary loss.\nThe remainder of the architecture fits into the family of the Behaviour transformer, with some design differences that impact the performance, and utilize the CoT learned embeddings.\nExperiments on the Moving Maze, Franka Kitchen and ManiSkill2 environments show the effectiveness of CoTPC over baselines.\nAn ablation study is also presented to showcase the importance of different design choices in the architecture.", "strengths": "- The dataset chosen for experiments are relevant, and the results are quite convincing, with the CoT model clearly performing better than the baselines. I thought the exposition of the experiments section clear and easy to follow. The choice made on experiments were clear and, to my knowledge, the choice of baselines are fair.\n\n- Although the architecture is limited in novelty, the method as a whole is novel and the specific design decisions are novel. I particularly liked the use of learned embeddings that are trained using an auxiliary loss with targets generated using an unsupervised subgoal discovery process.\n\n- The proposed work is clearly motivated and properly positioned in the literature.", "weaknesses": "- The writing of section 4 needs improvements. Specifically, I found Section 4.2.2 quite hard to parse and easy to get lost. I would encourage the authors to re-write this section and redo Figure 1 so that things are clearer. \n    - Specifically, I'm confused by what the inputs of $g_\\text{CoT}$. \n    - What is the CoT predictor? It appears once in the last paragraph of section 4.2.2. \n    - Why aren't actions used as inputs to $g^x(.)$ functions?\n    - What exactly are the contents of $\\{ \\mathbf{S}^\\text{CoT}_{...}\\}$? Do they change with time? What is the (...) subscript? \n    - There seems to be $T$ CoT features, but that seems confusing as these are suppose to represent subgoals and are trained using the auxiliary $L_\\text{CoT}$. I thought PELT was minimizing the number of goals. How do you ensure alignment between the learned tokens and the output of PELT?\n\n- I found the ablation study to have limited value. I think it should aim to provide the reader with more intuition on what exactly is learned by CoT embeddings. This could possibly be shown on the maze environment, and would show clearly the discovery of the subgoals. I see Figure 5 in the appendix and it is a step in the right direction, but in my opinion, it would be be easier and more informative to show in the maze environment.\n\n- As a minor point, I would ask the authors to express the limitations of the approach. For instance, is it always possible to assume that actions that are similar or close temporally belong to the same subskill?", "questions": "- It seems like BeT should have been explained in section 3 as the method seems to be heavily based on it. I leave that up to the authors to decide, but it could potentially ease the exposition. An alternative could be to present it in the appendix, similar to the Rt-1 paragraph.\n\n- In the last paragraph of section 5.4 named \"Shared tokens for CoT and action predictions\", I think the different variants should be shown as a three part diagram two help the reader understand the differences in inputs. As of right now, I am not entirely sure what the differences are. \n\n- I would be curious to see what the effect of setting the component of the auxiliary loss term to 0, but keeping learnable prompt tokens. I think this is basically equivalent to BC but with added learnable tokens , maintain the same capacity as the CoTPC architecture.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes CoTPC, a Transformer-based architecture that performs hierarchical planning.\nAn important part of this method is the unsupervised discovery of subgoals, which assume that temporally close and similar actions belong to the same subskills.\nThe overall architecture uses learned subgoal embeddings and uses the goals discovered by the unsupervised algorithm to train these learned embeddings, as an auxiliary loss.\nThe remainder of the architecture fits into the family of the Behaviour transformer, with some design differences that impact the performance, and utilize the CoT learned embeddings.\nExperiments on the Moving Maze, Franka Kitchen and ManiSkill2 environments show the effectiveness of CoTPC over baselines.\nAn ablation study is also presented to showcase the importance of different design choices in the architecture.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- The dataset chosen for experiments are relevant, and the results are quite convincing, with the CoT model clearly performing better than the baselines. I thought the exposition of the experiments section clear and easy to follow. The choice made on experiments were clear and, to my knowledge, the choice of baselines are fair.\n\n- Although the architecture is limited in novelty, the method as a whole is novel and the specific design decisions are novel. I particularly liked the use of learned embeddings that are trained using an auxiliary loss with targets generated using an unsupervised subgoal discovery process.\n\n- The proposed work is clearly motivated and properly positioned in the literature.", "weaknesses": "- The writing of section 4 needs improvements. Specifically, I found Section 4.2.2 quite hard to parse and easy to get lost. I would encourage the authors to re-write this section and redo Figure 1 so that things are clearer. \n    - Specifically, I'm confused by what the inputs of $g_\\text{CoT}$. \n    - What is the CoT predictor? It appears once in the last paragraph of section 4.2.2. \n    - Why aren't actions used as inputs to $g^x(.)$ functions?\n    - What exactly are the contents of $\\{ \\mathbf{S}^\\text{CoT}_{...}\\}$? Do they change with time? What is the (...) subscript? \n    - There seems to be $T$ CoT features, but that seems confusing as these are suppose to represent subgoals and are trained using the auxiliary $L_\\text{CoT}$. I thought PELT was minimizing the number of goals. How do you ensure alignment between the learned tokens and the output of PELT?\n\n- I found the ablation study to have limited value. I think it should aim to provide the reader with more intuition on what exactly is learned by CoT embeddings. This could possibly be shown on the maze environment, and would show clearly the discovery of the subgoals. I see Figure 5 in the appendix and it is a step in the right direction, but in my opinion, it would be be easier and more informative to show in the maze environment.\n\n- As a minor point, I would ask the authors to express the limitations of the approach. For instance, is it always possible to assume that actions that are similar or close temporally belong to the same subskill?", "questions": "- It seems like BeT should have been explained in section 3 as the method seems to be heavily based on it. I leave that up to the authors to decide, but it could potentially ease the exposition. An alternative could be to present it in the appendix, similar to the Rt-1 paragraph.\n\n- In the last paragraph of section 5.4 named \"Shared tokens for CoT and action predictions\", I think the different variants should be shown as a three part diagram two help the reader understand the differences in inputs. As of right now, I am not entirely sure what the differences are. \n\n- I would be curious to see what the effect of setting the component of the auxiliary loss term to 0, but keeping learnable prompt tokens. I think this is basically equivalent to BC but with added learnable tokens , maintain the same capacity as the CoTPC architecture.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698892316945}, {"id": "EgEsiXumdz", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9232/Reviewer_iFzy"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents CoTPC, a behavior cloning method that predicts simultaneously multiple future sub-goals (Chain-of-thoughts), as well as low-level actions.  It also presents a method for discovering brakpoints and a chain of planning steps. It's evaluated on state-based tasks across various settings, from 2D moving maze, to franka kitchen, then to several tasks in maniskill2. The experiments show CoTPC outperforms other baselines as well as other ablation choices. It also show some preliminary results on 2 real world tasks.", "review_text": "This paper presents CoTPC, a behavior cloning method that predicts simultaneously multiple future sub-goals (Chain-of-thoughts), as well as low-level actions.  It also presents a method for discovering brakpoints and a chain of planning steps. It's evaluated on state-based tasks across various settings, from 2D moving maze, to franka kitchen, then to several tasks in maniskill2. The experiments show CoTPC outperforms other baselines as well as other ablation choices. It also show some preliminary results on 2 real world tasks.", "strengths": "I like the general direction this paper is pursuing. Addressing suboptimality in demonstrations by finding shared hierarchical patterns and key states makes a lot of sense. Predicting a sequence of subgoals simultaneously, as opposed to auto-regressively one-by-one, is also reasonable in terms of better guiding low-level actions prediction.\nThe paper has a set of extensive experiments, as well as some preliminary study using realistic visual inputs and real-world experiments.\nIn addition, i was a reviewer reviewing this paper during its previous round of submission, back then one of my major concern is lacking of a automated machanisim for extracting key states from the demo. This has been addressed to some extent in this version.", "weaknesses": "* I am still not fully convinced by using the term 'Chain-of-Thought'...\n*  Real world evaluation is a bit too simple", "questions": "I have no further questions since I reviewed this before.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents CoTPC, a behavior cloning method that predicts simultaneously multiple future sub-goals (Chain-of-thoughts), as well as low-level actions.  It also presents a method for discovering brakpoints and a chain of planning steps. It's evaluated on state-based tasks across various settings, from 2D moving maze, to franka kitchen, then to several tasks in maniskill2. The experiments show CoTPC outperforms other baselines as well as other ablation choices. It also show some preliminary results on 2 real world tasks.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "I like the general direction this paper is pursuing. Addressing suboptimality in demonstrations by finding shared hierarchical patterns and key states makes a lot of sense. Predicting a sequence of subgoals simultaneously, as opposed to auto-regressively one-by-one, is also reasonable in terms of better guiding low-level actions prediction.\nThe paper has a set of extensive experiments, as well as some preliminary study using realistic visual inputs and real-world experiments.\nIn addition, i was a reviewer reviewing this paper during its previous round of submission, back then one of my major concern is lacking of a automated machanisim for extracting key states from the demo. This has been addressed to some extent in this version.", "weaknesses": "* I am still not fully convinced by using the term 'Chain-of-Thought'...\n*  Real world evaluation is a bit too simple", "questions": "I have no further questions since I reviewed this before.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698698553061}, {"id": "NPBsWRP7HY", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9232/Reviewer_HDpH"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors present Chain-of-Thought (CoT) Predictive control, a transformer based approach to policy learning, trained via sequence modeling. The transformer is augmented with learnable CoT prompt tokens that guide low-level action learning. In addition, the transformer is trained to predict the next and last high-level prompt, further encouraging abstractions that capture higher level semantic information. The high-level prompts are discovered in an unsupervised manner, as changepoints in time, discovered with Pruned Exact Linear Time methods, using cosine similarity as a cost metric. The model is trained on suboptimal demos and surpasses other transformer based methods on held-out tasks.", "review_text": "The authors present Chain-of-Thought (CoT) Predictive control, a transformer based approach to policy learning, trained via sequence modeling. The transformer is augmented with learnable CoT prompt tokens that guide low-level action learning. In addition, the transformer is trained to predict the next and last high-level prompt, further encouraging abstractions that capture higher level semantic information. The high-level prompts are discovered in an unsupervised manner, as changepoints in time, discovered with Pruned Exact Linear Time methods, using cosine similarity as a cost metric. The model is trained on suboptimal demos and surpasses other transformer based methods on held-out tasks.", "strengths": "The paper has several strengths:\n\n\n1) Reasonably well written\n2) Simplistic and effective approach that outperforms similar methods\n3) Nice to see their approach applied to complex dynamic settings, and generalizing favorably", "weaknesses": "My main concern with the method is their motivation for how they perform sub-task decomposition. Cosine similarity metric seems heuristical, not well motivated, and anecdotal. Whilst the results are promising, it is unclear whether more principled decompositions  would lead to better results: e.g. obtained via bottleneck options [1] or gaussian processes [2]. The paper would benefit from a greater discussion/comparison on this front. It is unclear to me whether their decomposition approach would favor different tasks, with distinct action-space statistics.  \n\nIn addition, there are a couple of presentation limitations:\n\n1) Citations are not in the correct ICLR format (surname and year)\n2) Results are lacking confidence intervals (how many runs/model seeds)\n\n[1] - Salter, Sasha, et al. \"Mo2: Model-based offline options.\" Conference on Lifelong Learning Agents. PMLR, 2022.\n\n[2] - Saatçi, Yunus, Ryan D. Turner, and Carl E. Rasmussen. \"Gaussian process change point models.\" Proceedings of the 27th International Conference on Machine Learning (ICML-10). 2010.", "questions": "1) How well does this approach to sub-task decomposition scale to larger action-spaces?\n2) How sensitive is their approach to the beta parameter that controls the number of detected changepoints?\n3) Fig 2 - Can the author's comment on what the action groupings correspond to intuitively for these examples?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present Chain-of-Thought (CoT) Predictive control, a transformer based approach to policy learning, trained via sequence modeling. The transformer is augmented with learnable CoT prompt tokens that guide low-level action learning. In addition, the transformer is trained to predict the next and last high-level prompt, further encouraging abstractions that capture higher level semantic information. The high-level prompts are discovered in an unsupervised manner, as changepoints in time, discovered with Pruned Exact Linear Time methods, using cosine similarity as a cost metric. The model is trained on suboptimal demos and surpasses other transformer based methods on held-out tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper has several strengths:\n\n\n1) Reasonably well written\n2) Simplistic and effective approach that outperforms similar methods\n3) Nice to see their approach applied to complex dynamic settings, and generalizing favorably", "weaknesses": "My main concern with the method is their motivation for how they perform sub-task decomposition. Cosine similarity metric seems heuristical, not well motivated, and anecdotal. Whilst the results are promising, it is unclear whether more principled decompositions  would lead to better results: e.g. obtained via bottleneck options [1] or gaussian processes [2]. The paper would benefit from a greater discussion/comparison on this front. It is unclear to me whether their decomposition approach would favor different tasks, with distinct action-space statistics.  \n\nIn addition, there are a couple of presentation limitations:\n\n1) Citations are not in the correct ICLR format (surname and year)\n2) Results are lacking confidence intervals (how many runs/model seeds)\n\n[1] - Salter, Sasha, et al. \"Mo2: Model-based offline options.\" Conference on Lifelong Learning Agents. PMLR, 2022.\n\n[2] - Saatçi, Yunus, Ryan D. Turner, and Carl E. Rasmussen. \"Gaussian process change point models.\" Proceedings of the 27th International Conference on Machine Learning (ICML-10). 2010.", "questions": "1) How well does this approach to sub-task decomposition scale to larger action-spaces?\n2) How sensitive is their approach to the beta parameter that controls the number of detected changepoints?\n3) Fig 2 - Can the author's comment on what the action groupings correspond to intuitively for these examples?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698530815253}, {"id": "R0VwFwzJIM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9232/Reviewer_QpAT"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a policy learning from the demonstration method. The authors propose a novel hierarchical imitation learning that utilizes scalable demonstrations. The demonstration is decomposed into a sequence of key observations, and then CoT is leveraged to generalize policy learning.", "review_text": "This paper proposes a policy learning from the demonstration method. The authors propose a novel hierarchical imitation learning that utilizes scalable demonstrations. The demonstration is decomposed into a sequence of key observations, and then CoT is leveraged to generalize policy learning.", "strengths": "1. I think the author's environment looks cool and real. This is very important in the policy learning domain and in the robotic domain. I see the supplementary video and I believe that this environment provides a good test environment for policy learning methods. The authors are encouraged by the reviewer to continue their study in this domain, even though this paper may be rejected.\n\n2. The general direction of learning policy with demonstrations is good, and the hierarchical RL formulation for this task is also sound and important, although those two ideas are not very novel.\n\n3. The numerical results are good and the improvement looks significant.\n\n4. I like the supplementary video as well, although it can be improved (see below).", "weaknesses": "# 1. About the novelty\n\n(1) I think CoT is not very novel in this context and I don't think the proposed approach can be regarded as CoT. The proposed approach refers to a demonstration split methodology (or a subgoal discovery mechanism in hierarchical RL). Since no language is involved, I don't think this method can be related to CoT.\n\n(2) This paper is not positioned well in the context of hierarchical RL + demonstration, so the novelty is not well-stated. The authors should mention, discuss, or compare the following works [1-5]. Disclaim: I am not an author of any of these works. The idea of learning subgoals by similarity or diversity is not novel.\n\n(3) Using transformers in policy learning is not novel as well, and I don't think the major goal of this paper is related to architecture design. If the goal is to claim the novelty of CoT control, the authors should try other architectures as well. Since nowadays transformers are very common architectures, I don't think this can be claimed as a major novelty.\n\n(4) It's unclear why the authors use a single model to learn CoT and learn the action. The authors should try to compare to architectures like [6].\n\n# 2. About the experiments\n\n(1) The authors should discuss their results and summarize the conclusions in the main text. I'm confused about the main result in Tab. 1, and this is because the results are not discussed (the authors only say results in Tab.1 without further explanations). It seems that BC, DT, and BeT are not hierarchical RL methods. So it is very unfair. The authors should not compare to methods that do not use subgoals. Instead, the authors should mainly compare to hierarchical RL methods that leverage demonstrations. The authors should search for [1-5] as long as their follow-up works to get the most related hierarchical RL methods to compare to. \n\n(2) The variance should be shown for each method, and the learning curves are required.\n\n(3) The supplementary videos should be combined with a presentation video, which can reveal the comparisons between the proposed approach and previous works.\n\n# 3. About the writing\n\n(1) The function of the model is not discussed and section 4.2 looks confusing as a result. I think the function and signature of the model should be discussed prior to the model details.\n\n(2) The details of the CoT algorithm are unclear. Particularly, these two sentences look very confusing to me.\n\n\"therefore, propose to group contiguous actions into segments, using a similarity-based heuristic to find these subskills.\" How does the group algorithm work? How to discretize continuous actions?\n\n\"We then utilize the Pruned Exact Linear Time (PELT) method [38] with cosine similarity as the cost metric to generate the changepoints in a per-trajectory manner.\" How does this work? I'm not familiar with the PELT method and this should be discussed in detail.\n\n(3) The authors assume the readers know the decision transformer and detection transformer in advance, which is not very good.\n\n[1] Jiang, Yiding, et al. \"Learning Options via Compression.\" Advances in Neural Information Processing Systems 35 (2022): 21184-21199.\n\n[2] Eysenbach, Benjamin, et al. \"Diversity is all you need: Learning skills without a reward function.\" arXiv preprint arXiv:1802.06070 (2018).\n\n[3] Konidaris, George, et al. \"Robot learning from demonstration by constructing skill trees.\" The International Journal of Robotics Research 31.3 (2012): 360-375.\n\n[4] Pickett, Marc, and Andrew G. Barto. \"Policyblocks: An algorithm for creating useful macro-actions in reinforcement learning.\" ICML. Vol. 19. 2002.\n\n[5] Kipf, Thomas, et al. \"Compile: Compositional imitation learning and execution.\" International Conference on Machine Learning. PMLR, 2019.\n\n[6] Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning", "questions": "1. How does the number of CoT affect the policy learning results?\n\n2. If the transformer's weights are tuned, why the CoT can accurately serve as a target signal?\n\n3. How efficient is the transformer architecture?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a policy learning from the demonstration method. The authors propose a novel hierarchical imitation learning that utilizes scalable demonstrations. The demonstration is decomposed into a sequence of key observations, and then CoT is leveraged to generalize policy learning.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. I think the author's environment looks cool and real. This is very important in the policy learning domain and in the robotic domain. I see the supplementary video and I believe that this environment provides a good test environment for policy learning methods. The authors are encouraged by the reviewer to continue their study in this domain, even though this paper may be rejected.\n\n2. The general direction of learning policy with demonstrations is good, and the hierarchical RL formulation for this task is also sound and important, although those two ideas are not very novel.\n\n3. The numerical results are good and the improvement looks significant.\n\n4. I like the supplementary video as well, although it can be improved (see below).", "weaknesses": "# 1. About the novelty\n\n(1) I think CoT is not very novel in this context and I don't think the proposed approach can be regarded as CoT. The proposed approach refers to a demonstration split methodology (or a subgoal discovery mechanism in hierarchical RL). Since no language is involved, I don't think this method can be related to CoT.\n\n(2) This paper is not positioned well in the context of hierarchical RL + demonstration, so the novelty is not well-stated. The authors should mention, discuss, or compare the following works [1-5]. Disclaim: I am not an author of any of these works. The idea of learning subgoals by similarity or diversity is not novel.\n\n(3) Using transformers in policy learning is not novel as well, and I don't think the major goal of this paper is related to architecture design. If the goal is to claim the novelty of CoT control, the authors should try other architectures as well. Since nowadays transformers are very common architectures, I don't think this can be claimed as a major novelty.\n\n(4) It's unclear why the authors use a single model to learn CoT and learn the action. The authors should try to compare to architectures like [6].\n\n# 2. About the experiments\n\n(1) The authors should discuss their results and summarize the conclusions in the main text. I'm confused about the main result in Tab. 1, and this is because the results are not discussed (the authors only say results in Tab.1 without further explanations). It seems that BC, DT, and BeT are not hierarchical RL methods. So it is very unfair. The authors should not compare to methods that do not use subgoals. Instead, the authors should mainly compare to hierarchical RL methods that leverage demonstrations. The authors should search for [1-5] as long as their follow-up works to get the most related hierarchical RL methods to compare to. \n\n(2) The variance should be shown for each method, and the learning curves are required.\n\n(3) The supplementary videos should be combined with a presentation video, which can reveal the comparisons between the proposed approach and previous works.\n\n# 3. About the writing\n\n(1) The function of the model is not discussed and section 4.2 looks confusing as a result. I think the function and signature of the model should be discussed prior to the model details.\n\n(2) The details of the CoT algorithm are unclear. Particularly, these two sentences look very confusing to me.\n\n\"therefore, propose to group contiguous actions into segments, using a similarity-based heuristic to find these subskills.\" How does the group algorithm work? How to discretize continuous actions?\n\n\"We then utilize the Pruned Exact Linear Time (PELT) method [38] with cosine similarity as the cost metric to generate the changepoints in a per-trajectory manner.\" How does this work? I'm not familiar with the PELT method and this should be discussed in detail.\n\n(3) The authors assume the readers know the decision transformer and detection transformer in advance, which is not very good.\n\n[1] Jiang, Yiding, et al. \"Learning Options via Compression.\" Advances in Neural Information Processing Systems 35 (2022): 21184-21199.\n\n[2] Eysenbach, Benjamin, et al. \"Diversity is all you need: Learning skills without a reward function.\" arXiv preprint arXiv:1802.06070 (2018).\n\n[3] Konidaris, George, et al. \"Robot learning from demonstration by constructing skill trees.\" The International Journal of Robotics Research 31.3 (2012): 360-375.\n\n[4] Pickett, Marc, and Andrew G. Barto. \"Policyblocks: An algorithm for creating useful macro-actions in reinforcement learning.\" ICML. Vol. 19. 2002.\n\n[5] Kipf, Thomas, et al. \"Compile: Compositional imitation learning and execution.\" International Conference on Machine Learning. PMLR, 2019.\n\n[6] Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning", "questions": "1. How does the number of CoT affect the policy learning results?\n\n2. If the transformer's weights are tuned, why the CoT can accurately serve as a target signal?\n\n3. How efficient is the transformer architecture?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697247129368}], "openreview_url": "https://openreview.net/forum?id=UTGv8CayNt", "arxiv_id": "2304.00776", "paper_pdf": "papers/UTGv8CayNt.pdf", "paper_pdf_sha256": "a8356f018ce2525dd1ac4f5d8becf5410d6861b0a751cfd9439fe05d08a85fef", "paper_pdf_bytes": 2263379, "paper_pdf_source": "openreview", "code_url": "https://github.com/SeanJia/CoTPC", "code_repository": "SeanJia/CoTPC", "code_commit": "1c971be6cd5bdbfd59e0ad5607be0d5c149b9b5b", "code_archive": "repos/UTGv8CayNt.zip", "code_archive_sha256": "fb46e2bfca37887a5954b748d773d02cce036b29201d5fe3f048218dbf28f824", "code_archive_bytes": 215039, "code_file_count": 13, "code_extensions": {".py": 10, ".sh": 3}, "github_disk_usage_kb": 297, "github_languages": {"Python": 67575, "Shell": 1818}, "github_archived": false, "github_pushed_at": "2023-05-01T00:43:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/chain-of-thought-predictive-control"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "alaQzRbCY9w", "year": 2022, "status": "rejected", "title": "Bolstering Stochastic Gradient Descent with Model Building", "authors": ["Ilker Birbil", "Özgür Martin", "Gönenc Onay", "Figen Öztoprak"], "authorids": ["~Ilker_Birbil1", "~Özgür_Martin1", "~Gönenc_Onay1", "~Figen_Öztoprak1"], "authors_source": "OpenReview API", "abstract": "Stochastic gradient descent method and its variants constitute the core optimization algorithms that achieve good convergence rates for solving machine learning problems. These rates are obtained especially when these algorithms are fine-tuned for the application at hand. Although this tuning process can require large computational costs, recent work has shown that these costs can be reduced by line search methods that iteratively adjust the stepsize. We propose an alternative approach to stochastic line search by using a new algorithm based on forward step model building. This model building step incorporates a second-order information that allows adjusting not only the stepsize but also the search direction. Noting that deep learning model parameters come in groups (layers of tensors), our method builds its model and calculates a new step for each parameter group.  This novel diagonalization approach makes the selected step lengths adaptive. We provide convergence rate analysis, and experimentally show that the proposed algorithm achieves faster convergence and better generalization in most problems. Moreover, our experiments show that the proposed method is quite robust as it converges for a wide range of initial stepsizes.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "2cinxn3g7pr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2842/Reviewer_xM1f"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a model building approach that replaces the one-step backtracking in stochastic line search methods. The authors provided some convergence guarantees for their proposed method and evaluated their approach on some image classification tasks.", "review_text": "1- The majority of the manuscript is well-written and easy to understand. However, some parts require further explanation and clarification that the reviewer explains in the following comments.\n\n2- The literature part needs to be strengthened. \n\n3- The point that is mentioned in Page 2 \"It also incorporates the most recent curvature information from the current point. This is in contrast with the stochastic quasi-Newton methods which use the information from the previous steps\" is $\\textbf{not}$ novel! The studies [1] and [2] (to name a few) introduced this concept earlier.\n\n4- Does the inequality (3) guarantee a decrease in the main objective function or just the stochastic objective function? If just the stochastic objective function, how can we trust the step length?\n\n5- In Page 3 (line 3 in the second paragraph) the authors mentioned group $p$ which is not introduced before. The text needs to be more clear and readable.\n\n6- One of the main goals of this paper is related to reducing the computational costs of tuning. However, in the proposed algorithms 1 and 2, there are many hyper-parameters that still requires tuning. The authors need to exactly mention how they claim the idea of reducing tuning attempts.\n\n7- The theoretical results are quite basic and more stronger results are expected. Theorem 1 does not guarantee any convergence, it just shows that the norm of gradient (in expectation) is bounded above by some terms that are NOT diminishing as the algorithm progresses. If $\\alpha_k = \\mathcal{O}(L^{-2})$ (which is super tiny), the proposed method converges to the neighborhood of stationary point. Even with diminishing step sizes, the proved convergence result does not guarantee any convergence to a stationary point (just the neighborhood around a stationary point). All in all, the theoretical results definitely require modification.\n\n8- Do the authors consider the extra cost per each iteration in the plots with $x$-axis as $\\textbf{Epochs}$? Also, instead of averaging over 5 runs, it is better to show the error band. The mini-batch size is not reported for the experiments, it is very important hyper-parameter as well.\n\n9- That is good that the authors provided some results regarding the robustness of their method.\n\n10- In the abstract, it is mentioned that $\\textbf{This novel\ndiagonalization approach makes the selected step lengths adaptive.}$ The proposed method is not adaptive in terms of step size! In both algorithms 1 and 2, there is not any adaptive rule for the step size.\n\n\n[1] Gao, W., & Goldfarb, D. (2018). Block BFGS methods. SIAM Journal on Optimization, 28(2), 1205-1231.\n\n[2] Berahas, A. S., Jahani, M., Richtárik, P., & Takáč, M. (2020). Quasi-newton methods for deep learning: Forget the past, just sample.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper presents a model building approach that replaces the one-step backtracking in stochastic line search methods. The authors provided some convergence guarantees for their proposed method and evaluated their approach on some image classification tasks.", "main_review": "1- The majority of the manuscript is well-written and easy to understand. However, some parts require further explanation and clarification that the reviewer explains in the following comments.\n\n2- The literature part needs to be strengthened. \n\n3- The point that is mentioned in Page 2 \"It also incorporates the most recent curvature information from the current point. This is in contrast with the stochastic quasi-Newton methods which use the information from the previous steps\" is $\\textbf{not}$ novel! The studies [1] and [2] (to name a few) introduced this concept earlier.\n\n4- Does the inequality (3) guarantee a decrease in the main objective function or just the stochastic objective function? If just the stochastic objective function, how can we trust the step length?\n\n5- In Page 3 (line 3 in the second paragraph) the authors mentioned group $p$ which is not introduced before. The text needs to be more clear and readable.\n\n6- One of the main goals of this paper is related to reducing the computational costs of tuning. However, in the proposed algorithms 1 and 2, there are many hyper-parameters that still requires tuning. The authors need to exactly mention how they claim the idea of reducing tuning attempts.\n\n7- The theoretical results are quite basic and more stronger results are expected. Theorem 1 does not guarantee any convergence, it just shows that the norm of gradient (in expectation) is bounded above by some terms that are NOT diminishing as the algorithm progresses. If $\\alpha_k = \\mathcal{O}(L^{-2})$ (which is super tiny), the proposed method converges to the neighborhood of stationary point. Even with diminishing step sizes, the proved convergence result does not guarantee any convergence to a stationary point (just the neighborhood around a stationary point). All in all, the theoretical results definitely require modification.\n\n8- Do the authors consider the extra cost per each iteration in the plots with $x$-axis as $\\textbf{Epochs}$? Also, instead of averaging over 5 runs, it is better to show the error band. The mini-batch size is not reported for the experiments, it is very important hyper-parameter as well.\n\n9- That is good that the authors provided some results regarding the robustness of their method.\n\n10- In the abstract, it is mentioned that $\\textbf{This novel\ndiagonalization approach makes the selected step lengths adaptive.}$ The proposed method is not adaptive in terms of step size! In both algorithms 1 and 2, there is not any adaptive rule for the step size.\n\n\n[1] Gao, W., & Goldfarb, D. (2018). Block BFGS methods. SIAM Journal on Optimization, 28(2), 1205-1231.\n\n[2] Berahas, A. S., Jahani, M., Richtárik, P., & Takáč, M. (2020). Quasi-newton methods for deep learning: Forget the past, just sample.", "summary_of_the_review": "The theoretical results needs to be more strengthened. The contribution is ok but not good enough.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635894299084}, {"id": "jBN0St3PyAx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2842/Reviewer_hEnG"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a stochastic version of the model building algorithm (SMB)  by taking the subsequent iterate to be the minimizer of a quadratic model build from two iterates (the current and the one based on conventional SGD). Empirical results suggest that the proposed SMB has some great generalization performance.", "review_text": "Strengths  \nThere is some novelty in the proposed algorithm, and the paper has made some efforts to extend the model building approach to the stochastic setting. \n\nWeakness\n\n1. From theoretical side, by enforcing a relatively strong independence batch assumption, the convergence analysis is not a surprise given the reference Wang 2017. However, the experiments shows that SMB  consistently outperforms SMBi, it seems to me that dependence is the key for getting better performance.  Hence, it is unsatisfactory if the convergence property of SMB is not demystified and completely ignored.\n\n2. Since the paper has targeted its proposed algorithm a special case of quasi-newton method (e.g. (6)), it would be unsatisfactory if comparison with SQN is not made. \n\n3. The experiment conclusion is based on a single run, it is not convicing to me that the reported advantage of SMB is as reliable as described. I would recommend spending more time effort in the experiments before submission.\n\n4 Figure 3 shows that SLS outperforms SGD and ADAM while SGD and ADAM outperform SMB and SMBi. The last figure shows that SMB with Auto-scheduled stepsize performs slightly better than SLS. What is the most important factor in getting good performance? Would the benefits mainly come from the Auto-scheduled stepsize? Can we equip SGD or ADAM with similar scheduled stepsize? This appears to be very promising. \n\n\nSome minor issues.\n1. \\mathbb{E}[F(x; \\xi )] is defined at the beginning but never used.\n\n2. It is confusing that $n$ is used to define both the number of parameter blocks and the dimension of X.\n\n3. What is  $f_{k,p}$ defined in page 3?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents a stochastic version of the model building algorithm (SMB)  by taking the subsequent iterate to be the minimizer of a quadratic model build from two iterates (the current and the one based on conventional SGD). Empirical results suggest that the proposed SMB has some great generalization performance.", "main_review": "Strengths  \nThere is some novelty in the proposed algorithm, and the paper has made some efforts to extend the model building approach to the stochastic setting. \n\nWeakness\n\n1. From theoretical side, by enforcing a relatively strong independence batch assumption, the convergence analysis is not a surprise given the reference Wang 2017. However, the experiments shows that SMB  consistently outperforms SMBi, it seems to me that dependence is the key for getting better performance.  Hence, it is unsatisfactory if the convergence property of SMB is not demystified and completely ignored.\n\n2. Since the paper has targeted its proposed algorithm a special case of quasi-newton method (e.g. (6)), it would be unsatisfactory if comparison with SQN is not made. \n\n3. The experiment conclusion is based on a single run, it is not convicing to me that the reported advantage of SMB is as reliable as described. I would recommend spending more time effort in the experiments before submission.\n\n4 Figure 3 shows that SLS outperforms SGD and ADAM while SGD and ADAM outperform SMB and SMBi. The last figure shows that SMB with Auto-scheduled stepsize performs slightly better than SLS. What is the most important factor in getting good performance? Would the benefits mainly come from the Auto-scheduled stepsize? Can we equip SGD or ADAM with similar scheduled stepsize? This appears to be very promising. \n\n\nSome minor issues.\n1. \\mathbb{E}[F(x; \\xi )] is defined at the beginning but never used.\n\n2. It is confusing that $n$ is used to define both the number of parameter blocks and the dimension of X.\n\n3. What is  $f_{k,p}$ defined in page 3?\n\n", "summary_of_the_review": "While this paper makes some interesting contribution of a new stochastic algorithm for deep learning, it appears to be far from a complete work. I think the paper needs to be substantially improved in both theory and empirical study.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635873310549}, {"id": "aJS9na2o5rR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2842/Reviewer_8Xqy"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "\nThis paper proposes an alternative to stochastic line search which is based on forwarding step model building which corrects the direction of move and its magnitude at the same time. In its proposed algorithm it first checks if the given step size satisfies the stochastic line search. If yes then just use the step size and do the SGD update. Otherwise, it builds linear models around two points and combines these two models and minimizes this new model and this becomes the new iterate value. \n\n", "review_text": "1- The paper is an incremental work and is mainly a stochastic variant of Öztoprak et al 2018. This paper doesn't offer anything technical since all the theoretical results are just slight modifications of already existing results. On the other hand, the empirical results also don't show a big advancement. \n\n2-  Generally tuning both direction and step size is not a new approach. For example, the diagonal variant of AdaGrad has a similar approach. Also in the proposed algorithm in each step, it can require a matrix-vector multiplication and it cannot be relaxed. Parameter partitioning is a way to make the matrix block-diagonal and to reduce the matrix-vector multiplication. However, in the paper, it is not clarified why partitioning is useful. \n\n3- The paper claims that unlike line search there is no backtracking step. However, to set \\eta their proposed algorithm needs backtracking. \n\n4- It hasn’t been mentioned in the paper how many times each experiment ran? It would be more reliable if each experiment runs several times and the average of them is plotted. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "\nThis paper proposes an alternative to stochastic line search which is based on forwarding step model building which corrects the direction of move and its magnitude at the same time. In its proposed algorithm it first checks if the given step size satisfies the stochastic line search. If yes then just use the step size and do the SGD update. Otherwise, it builds linear models around two points and combines these two models and minimizes this new model and this becomes the new iterate value. \n\n", "main_review": "1- The paper is an incremental work and is mainly a stochastic variant of Öztoprak et al 2018. This paper doesn't offer anything technical since all the theoretical results are just slight modifications of already existing results. On the other hand, the empirical results also don't show a big advancement. \n\n2-  Generally tuning both direction and step size is not a new approach. For example, the diagonal variant of AdaGrad has a similar approach. Also in the proposed algorithm in each step, it can require a matrix-vector multiplication and it cannot be relaxed. Parameter partitioning is a way to make the matrix block-diagonal and to reduce the matrix-vector multiplication. However, in the paper, it is not clarified why partitioning is useful. \n\n3- The paper claims that unlike line search there is no backtracking step. However, to set \\eta their proposed algorithm needs backtracking. \n\n4- It hasn’t been mentioned in the paper how many times each experiment ran? It would be more reliable if each experiment runs several times and the average of them is plotted. \n", "summary_of_the_review": "The paper is an incremental work and is mainly a stochastic variant of Öztoprak et al 2018. This paper doesn't offer anything technical since all the theoretical results are just slight modifications of already existing results. On the other hand, the empirical results also don't show a big advancement. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635828675105}, {"id": "c8DXxlwh7ZD", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2842/Reviewer_4z74"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed a method called stochastic model building (SMB) that uses a combination of existing techniques to get faster convergence in stochastic non-convex optimization. In particular, they use a stochastic adaptation of the model-building globalization strategy from Oztoprak and Birbil (2018), in which the deterministic Armijo condition check is also computed using stochastic gradients. By re-writing their update as a preconditioned SGD step, they are able to bound the spectrum of the preconditioner. This allows them to obtain convergence results for smooth and non-convex objectives by directly invoking the analysis of Wang et al. (2017). Experiments demonstrate that with additional heuristics, their proposed method can outperform state-of-the-art optimizers on common deep learning benchmarks.", "review_text": "**Strengths**\nThe proposed method SMB/SMBi has several advantages. First of all, unlike general preconditioned stochastic gradient methods, the preconditioner here is a blockwise diagonal matrix where each block corresponds to a parameter group in the deep neural network. In addition, because the preconditioned update is only expressed so to facilitate theoretical analysis, the actual implementation itself does not involve computing or storing matrices at all --- only vector inner products are involved in each parameter group update. This is particularly suitable for implementation in popular deep learning libraries like PyTorch where parameters are stored in groups. \n\nSMB also does not use backtracking as is typical in line search with the Armijo condition. When the condition is failed for the stochastic gradient direction, SMB simply computes a new direction based on a stochastic \"quasi-Newton\" update, and take a step along this direction instead. This lets SMB avoid potentially doing many additional forward passes which can be expensive. \n\nFinally, the empirical results show that by using additional simple heuristics to adjust the step size based on how often the Armijo condition has failed in the previous epoch, SMB can achieve state-of-the-art performance on CIFAR10 with a ResNet34 architecture, both in terms of the training loss and test accuracy. \n\n\n**Weaknesses**\nThe main concerns I have with this paper are:\n1. Although convergence results are given in the non-convex setting, the analysis is rather trivial. The crux of the proofs for the theorems is to first re-write the SMB/SMBi update as a preconditioned stochastic gradient step $x_{k+1} = x_k - \\alpha_k H_k g_k$, and this part comes directly from the work of Oztoprak and Birbil (2018) except here the deterministic gradients are replaced with their stochastic counterparts. Then the proof boils down to bounding the extreme eigenvalues of each of the diagonal blocks in $H_k$. This part of the analysis is also nearly identical to what's in Oztoprak and Birbil (2018). Finally, with the extreme eigenvalues at hand, the non-convex results given by Wang et al. (2017) can be directly invoked to obtain Theorem 1 and 2 in this paper. The theoretical analysis therefore is simple as previous works have done all the heavy-lifting,  but at the same time I consider the contribution to be neither novel nor significant.\n2. Because the algorithm is directly adapted from that of Oztoprak and Birbil (2018), the preconditioner $H_k$ would be correlated with the stochastic gradient. This would not fit into the analysis of Wang et al. (2017) as their theory requires $H_k$ to be independent of $\\nabla f(x_k,\\xi_k)$ conditioned on the past. The algorithm in Wang et al. (2017) constructs $H_k$  based on a truncated history of the past stochastic gradients and therefore is independent of the randomness in iteration $k$. To circumvent this issue, the authors here suggest to analyze a different version of SMB, which uses an independent minibatch to compute $H_k$, and the variant is named SMBi. Although this lets the analysis go through, it inevitably introduces a *gap between theory and practice*, since the experiments demonstrate that SMB almost always perform better than SMBi. This observation makes sense as although we are using stochastic quantities to construct the model, the model can be seen as a locally approximation to the minibatch objective, while using independent batches to compute the model does not seem to make a lot of sense. \n3. Figure 1: Although it is true that in this particular case (seed, rather), \"$s_k$  has a smaller length compared to the trial step $s_k^t$, the step we would have taken with $s_k^t$ is usually scaled with a step size instead of the unit step size taken by $s_k$ in SMB. More importantly, the direction obtained in $s_k^t$, the stochastic gradient step depends on which feature vector $\\xi_k$ is sampled, and so it shouldn't always be the case that $s_k^t$ \"lies along a direction decreasing the function value\" while $s_k$ does not, unless this can be proven otherwise. For these reasons I find this illustration slightly misleading. \n4. Robustness to step size: assuming the experiments in Figure 6 are using constant step sizes, it is unclear to me what the point of robustness to step size means given that SMB/SMBi only works well in general when step size auto-scheduling is required. If the authors indeed recommend the step size heuristic in Section 4 to be adopted in practice, then I see little value in this robustness experiment, but I could be interpreting the experiment incorrectly. Instead, I would recommend a robustness experiment on the other hyperparameters such as the percentage threshold mentioned in Section 4 and the step size multiplication factor. The last issue I have with Figure 6 is it seems like for MNIST-MLP, different step size can lead to an order of magnitude difference in the final training loss. Does that imply SMB isn't very robust to the step size on this dataset/model combination?\n5. It would be great to see a replication of the auto-scheduling experiment on CIFAR100-ResNet34 as well, especially to see whether with the auto-scheduling, the \"converge fast but generalize worse\" observation in Figure 3 (c,d) can be mitigated.\n\n**Minor concerns / typos**\n- In the objective (Equations (1) and (2)), the dimensionality of the problem is $n$; however, $n$ is later also used to denote the number of parameter groups at the top of page 3. The latter implies the overall dimensionality of the problem should be $\\sum_{j=1}^n p_j$ where $p_j$ is the number of parameters in group $j$. \n- It's unclear to me why $f_k=f(x_k,\\xi_k)$ is not expressed as the sum of individual function values of a mini-batch, as you do for the stochastic gradients $g_k$. I suppose it should be and this is just to avoid cluttering, or are the $f_k$'s computed in some other way?\n- Top of page 5: when concluding that $\\mathbb{E}[H_kg_k] = H_k\\nabla f$, one should be more explicit that the expectation is taken over $\\xi_k$, the first minibatch sampled at that iteration, excluding the randomness in $H_k$, since $H_k$ is also a random quantity it's just that its randomness comes from a separate, independent minibatch. \n- Top of page 6: \"For given $T$, $R$ be a random variable\" -> \"For *a* given $T$, *let* $R$ be a random variable\"\n- Proof of Theorem 1 (page 11, section A.1):\n\t- First line: \"can be expressed a special\" -> \"can be expressed *as* a special\"\n\t- Please use a different notation for $\\sigma=(y_k^\\top s_k^t + 2\\delta)^2 - \\|s_k^t\\|^2 \\|y_k\\|^2$ as $\\sigma$ is already used in Assumption $\\mathbb{E}_{\\xi_k}[\\|g(x_k,\\xi_k)-\\nabla f(x_k)\\|^2] \\leq \\sigma^2$.\n- Proof of Theorem 1 (page 12, section A.1):\n\t- Second equality in upper bounding $\\lambda_{\\max}$: the last $+y_{k,p}^\\top g_{k,p}$ should be a minus instead. (typo)\n- It would be helpful to further explain the motivation behind the *stochastic model building* part of the algorithm, rather than just citing away Oztoprak and Birbil (2018). In particular, a justification of why their globalization strategy should also work well in the stochastic setting would significantly strengthen the paper.\n\n---\n**References mentioned**:\nOztoprak and Birbil (2018): An alternative globalization strategy for unconstrained optimization\nWang et al. (2017): Stochastic Quasi-Newton Methods for Nonconvex Stochastic Optimization \nNocedal and Wright (2006): Numerical Optimization\nVaswani et al. (2019) : Painless Stochastic Gradient: Interpolation, Line-Search, and Convergence Rates\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors proposed a method called stochastic model building (SMB) that uses a combination of existing techniques to get faster convergence in stochastic non-convex optimization. In particular, they use a stochastic adaptation of the model-building globalization strategy from Oztoprak and Birbil (2018), in which the deterministic Armijo condition check is also computed using stochastic gradients. By re-writing their update as a preconditioned SGD step, they are able to bound the spectrum of the preconditioner. This allows them to obtain convergence results for smooth and non-convex objectives by directly invoking the analysis of Wang et al. (2017). Experiments demonstrate that with additional heuristics, their proposed method can outperform state-of-the-art optimizers on common deep learning benchmarks.", "main_review": "**Strengths**\nThe proposed method SMB/SMBi has several advantages. First of all, unlike general preconditioned stochastic gradient methods, the preconditioner here is a blockwise diagonal matrix where each block corresponds to a parameter group in the deep neural network. In addition, because the preconditioned update is only expressed so to facilitate theoretical analysis, the actual implementation itself does not involve computing or storing matrices at all --- only vector inner products are involved in each parameter group update. This is particularly suitable for implementation in popular deep learning libraries like PyTorch where parameters are stored in groups. \n\nSMB also does not use backtracking as is typical in line search with the Armijo condition. When the condition is failed for the stochastic gradient direction, SMB simply computes a new direction based on a stochastic \"quasi-Newton\" update, and take a step along this direction instead. This lets SMB avoid potentially doing many additional forward passes which can be expensive. \n\nFinally, the empirical results show that by using additional simple heuristics to adjust the step size based on how often the Armijo condition has failed in the previous epoch, SMB can achieve state-of-the-art performance on CIFAR10 with a ResNet34 architecture, both in terms of the training loss and test accuracy. \n\n\n**Weaknesses**\nThe main concerns I have with this paper are:\n1. Although convergence results are given in the non-convex setting, the analysis is rather trivial. The crux of the proofs for the theorems is to first re-write the SMB/SMBi update as a preconditioned stochastic gradient step $x_{k+1} = x_k - \\alpha_k H_k g_k$, and this part comes directly from the work of Oztoprak and Birbil (2018) except here the deterministic gradients are replaced with their stochastic counterparts. Then the proof boils down to bounding the extreme eigenvalues of each of the diagonal blocks in $H_k$. This part of the analysis is also nearly identical to what's in Oztoprak and Birbil (2018). Finally, with the extreme eigenvalues at hand, the non-convex results given by Wang et al. (2017) can be directly invoked to obtain Theorem 1 and 2 in this paper. The theoretical analysis therefore is simple as previous works have done all the heavy-lifting,  but at the same time I consider the contribution to be neither novel nor significant.\n2. Because the algorithm is directly adapted from that of Oztoprak and Birbil (2018), the preconditioner $H_k$ would be correlated with the stochastic gradient. This would not fit into the analysis of Wang et al. (2017) as their theory requires $H_k$ to be independent of $\\nabla f(x_k,\\xi_k)$ conditioned on the past. The algorithm in Wang et al. (2017) constructs $H_k$  based on a truncated history of the past stochastic gradients and therefore is independent of the randomness in iteration $k$. To circumvent this issue, the authors here suggest to analyze a different version of SMB, which uses an independent minibatch to compute $H_k$, and the variant is named SMBi. Although this lets the analysis go through, it inevitably introduces a *gap between theory and practice*, since the experiments demonstrate that SMB almost always perform better than SMBi. This observation makes sense as although we are using stochastic quantities to construct the model, the model can be seen as a locally approximation to the minibatch objective, while using independent batches to compute the model does not seem to make a lot of sense. \n3. Figure 1: Although it is true that in this particular case (seed, rather), \"$s_k$  has a smaller length compared to the trial step $s_k^t$, the step we would have taken with $s_k^t$ is usually scaled with a step size instead of the unit step size taken by $s_k$ in SMB. More importantly, the direction obtained in $s_k^t$, the stochastic gradient step depends on which feature vector $\\xi_k$ is sampled, and so it shouldn't always be the case that $s_k^t$ \"lies along a direction decreasing the function value\" while $s_k$ does not, unless this can be proven otherwise. For these reasons I find this illustration slightly misleading. \n4. Robustness to step size: assuming the experiments in Figure 6 are using constant step sizes, it is unclear to me what the point of robustness to step size means given that SMB/SMBi only works well in general when step size auto-scheduling is required. If the authors indeed recommend the step size heuristic in Section 4 to be adopted in practice, then I see little value in this robustness experiment, but I could be interpreting the experiment incorrectly. Instead, I would recommend a robustness experiment on the other hyperparameters such as the percentage threshold mentioned in Section 4 and the step size multiplication factor. The last issue I have with Figure 6 is it seems like for MNIST-MLP, different step size can lead to an order of magnitude difference in the final training loss. Does that imply SMB isn't very robust to the step size on this dataset/model combination?\n5. It would be great to see a replication of the auto-scheduling experiment on CIFAR100-ResNet34 as well, especially to see whether with the auto-scheduling, the \"converge fast but generalize worse\" observation in Figure 3 (c,d) can be mitigated.\n\n**Minor concerns / typos**\n- In the objective (Equations (1) and (2)), the dimensionality of the problem is $n$; however, $n$ is later also used to denote the number of parameter groups at the top of page 3. The latter implies the overall dimensionality of the problem should be $\\sum_{j=1}^n p_j$ where $p_j$ is the number of parameters in group $j$. \n- It's unclear to me why $f_k=f(x_k,\\xi_k)$ is not expressed as the sum of individual function values of a mini-batch, as you do for the stochastic gradients $g_k$. I suppose it should be and this is just to avoid cluttering, or are the $f_k$'s computed in some other way?\n- Top of page 5: when concluding that $\\mathbb{E}[H_kg_k] = H_k\\nabla f$, one should be more explicit that the expectation is taken over $\\xi_k$, the first minibatch sampled at that iteration, excluding the randomness in $H_k$, since $H_k$ is also a random quantity it's just that its randomness comes from a separate, independent minibatch. \n- Top of page 6: \"For given $T$, $R$ be a random variable\" -> \"For *a* given $T$, *let* $R$ be a random variable\"\n- Proof of Theorem 1 (page 11, section A.1):\n\t- First line: \"can be expressed a special\" -> \"can be expressed *as* a special\"\n\t- Please use a different notation for $\\sigma=(y_k^\\top s_k^t + 2\\delta)^2 - \\|s_k^t\\|^2 \\|y_k\\|^2$ as $\\sigma$ is already used in Assumption $\\mathbb{E}_{\\xi_k}[\\|g(x_k,\\xi_k)-\\nabla f(x_k)\\|^2] \\leq \\sigma^2$.\n- Proof of Theorem 1 (page 12, section A.1):\n\t- Second equality in upper bounding $\\lambda_{\\max}$: the last $+y_{k,p}^\\top g_{k,p}$ should be a minus instead. (typo)\n- It would be helpful to further explain the motivation behind the *stochastic model building* part of the algorithm, rather than just citing away Oztoprak and Birbil (2018). In particular, a justification of why their globalization strategy should also work well in the stochastic setting would significantly strengthen the paper.\n\n---\n**References mentioned**:\nOztoprak and Birbil (2018): An alternative globalization strategy for unconstrained optimization\nWang et al. (2017): Stochastic Quasi-Newton Methods for Nonconvex Stochastic Optimization \nNocedal and Wright (2006): Numerical Optimization\nVaswani et al. (2019) : Painless Stochastic Gradient: Interpolation, Line-Search, and Convergence Rates\n", "summary_of_the_review": "Although the paper proposes an interesting, easy-to-implement algorithm for non-convex stochastic optimization, I am hesitant to recommend for acceptance at this point for its lack of significantly novel contribution and some confusing aspects of the experiments. I am willing to increase my score if my concerns are well addressed.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635448668417}], "openreview_url": "https://openreview.net/forum?id=alaQzRbCY9w", "arxiv_id": "2111.07058", "paper_pdf": "papers/alaQzRbCY9w.pdf", "paper_pdf_sha256": "c8640f5e63b98f80e48a730ebf87250d83fe93a6f5750bca93e54fd843963406", "paper_pdf_bytes": 1577973, "paper_pdf_source": "openreview", "code_url": "https://github.com/sibirbil/SMB", "code_repository": "sibirbil/SMB", "code_commit": "17fb8ba4f440a5e36e66d701385ecdef6cb05723", "code_archive": "repos/alaQzRbCY9w.zip", "code_archive_sha256": "fe6636bb802a0fefd225b034c8c5aaf15cc298ffaed48f9d5f501c79b7797f27", "code_archive_bytes": 86928, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 2671, "github_languages": {"Python": 48703}, "github_archived": false, "github_pushed_at": "2023-02-15T16:44:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bolstering-stochastic-gradient-descent-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XEw5Onu69uu", "year": 2021, "status": "rejected", "title": "Self-Labeling of Fully Mediating Representations by Graph Alignment", "authors": ["Martijn Oldenhof", "Adam Arany", "Yves Moreau", "Jaak Simm"], "authorids": ["~Martijn_Oldenhof1", "~Adam_Arany1", "~Yves_Moreau2", "~Jaak_Simm1"], "authors_source": "OpenReview API", "abstract": "To be able to predict a molecular graph structure ($W$) given a 2D image of a chemical compound ($U$)  is a challenging problem in machine learning. We are interested to learn $f: U \\rightarrow W$ where we have a fully mediating representation $V$ such that $f$ factors into $U \\rightarrow V \\rightarrow W$. However, observing V requires detailed and expensive labels. We propose \\textbf{graph aligning} approach that generates rich or detailed labels given normal labels $W$. In this paper we investigate the scenario of domain adaptation from the source domain where we have access to the expensive labels $V$ to the target domain where only normal labels W are available. Focusing on the problem of predicting chemical compound graphs from 2D images the fully mediating layer is represented using the planar embedding of the chemical graph structure we are predicting. The use of a fully mediating layer implies some assumptions on the mechanism of the underlying process. However if the assumptions are correct it should allow the machine learning model to be more interpretable, generalize better and be more  data efficient at training time.\nThe empirical results show that, using only 4000 data points, we obtain up to 4x improvement of performance after domain adaptation to target domain compared to pretrained model only on the source domain. After domain adaptation, the model is even able to detect atom types that were never seen in the original source domain. Finally, on the Maybridge data set the proposed self-labeling approach reached higher performance than the current state of the art.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "QYCxs3VLiPd", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1739/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n\nThe paper describes a method to convert 2D molecular images to molecular graph structures, with applications in extracting raw chemical structures from journal articles and other publications. The model has two components: a semantic segmentation network that first predicts the location of the atoms and bonds in the image, and a series of classification networks that classifies each segment. The paper proposes some domain adaptation techniques to reduce the amount of expensive ‘pixel-wise’ labels required for training the segmentation network\n\n##########################################################################\n\nReasons for score: \n\nOverall, I currently vote for rejection. I have some questions about the current evaluation setup that I hope the authors could clarify  \n\n##########################################################################\n\nStrengths:\n\n*The proposed iterative strong labeling and graph alignment framework seems to improve the performance of the pre-trained model  \n\nWeaknesses:\n\n*The model evaluation seems to show some mixed results. On the Mayfield dataset, there is an improvement over some baseline models, but on the Indigo dataset, the proposed model seems to be perform significantly worse than the model described in the citing reference 30. (~40% vs ~80%) \n\n*I found that some parts of the paper were difficult for me to understand (see below)\n\n##########################################################################\n\nQuestions and other comments:\n\n*Paper clarity:\n**I think there should be more information on how the underlying Chemgrapher model converts 2D molecular images to the molecular graph structures. How does the model actually construct the graph structure? Also, there is a lot of analysis on the image segmentation part of the model (eg figure 5, 6). Does the image classification part of the Chemgrapher model play any significant role? \n**Information about the Indigo and Maybridge dataset could be provided in a more accessible way. Eg the total number of examples in each dataset\n**Some additional information about how the original model was pre-trained would be useful\n\n*I think there needs to be more justification for why this proposed approach [option 1] of mapping the 2D molecule image to the molecular graph structure via an intermediate representation that explicitly identifies all the atoms and bonds in the image is preferred over the alternative approach [option 2] of directly mapping the 2D molecule image to the molecular graph structure (eg using by outputting a smiles text representation that can be converted to the molecular graph). The cost of option 1 is that it requires the very expensive pixel-wise labels that describes the locations of atoms and bonds in the 2D molecule image to train the segmentation model, thus motivating the domain adaptation part of this work. In terms of pure performance, it’s not clear to me that option 1 is superior, for example, [ref 30] which directly predicts the molecular smiles from the 2D molecular image [option 2] can attain ~80% in the indigo dataset, while this proposed approach seems to attain only ~40%\n\n*Figures 2 and 3 show various performance metrics over multiple iterations of the re-training. How did you decide which iteration to stop the re-training? Also, how computationally expensive is it to perform the iterative re-training procedure?\n\n*In Figure 3b, the performance at iteration 0 is ~72%, which I’m understanding to be the vanilla performance of the Chemgrapher model on the Maybridge dataset? But Table 1 shows that the performance of the Chemgrapher model on Maybridge is 83.3%\n\n*What is the reasoning for allowing a max of 2 node substitutions or 1 edge substitution for the ‘correcting graph alignment’ case?\n\n*In the future, it would be interesting to see how this proposed method compares in this recently published benchmark: https://github.com/Kohulan/OCSR_Review. NB: out of scope for this ICLR submission since it was published after the paper submission deadline\n\n\nRef [30]: Joshua Staker, Kyle Marshall, Robert Abel, and Carolyn M. McQuaw. Molecular Structure\nExtraction from Documents Using Deep Learning. J. Chem. Inf. Model., 59(3):1017–1029,\nMarch 2019. ISSN 1549-9596. doi: 10.1021/acs.jcim.8b00669. URL https://doi.org/10.1021/acs.jcim.8b00669.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Recommendation to reject", "review": "##########################################################################\n\nSummary:\n\nThe paper describes a method to convert 2D molecular images to molecular graph structures, with applications in extracting raw chemical structures from journal articles and other publications. The model has two components: a semantic segmentation network that first predicts the location of the atoms and bonds in the image, and a series of classification networks that classifies each segment. The paper proposes some domain adaptation techniques to reduce the amount of expensive ‘pixel-wise’ labels required for training the segmentation network\n\n##########################################################################\n\nReasons for score: \n\nOverall, I currently vote for rejection. I have some questions about the current evaluation setup that I hope the authors could clarify  \n\n##########################################################################\n\nStrengths:\n\n*The proposed iterative strong labeling and graph alignment framework seems to improve the performance of the pre-trained model  \n\nWeaknesses:\n\n*The model evaluation seems to show some mixed results. On the Mayfield dataset, there is an improvement over some baseline models, but on the Indigo dataset, the proposed model seems to be perform significantly worse than the model described in the citing reference 30. (~40% vs ~80%) \n\n*I found that some parts of the paper were difficult for me to understand (see below)\n\n##########################################################################\n\nQuestions and other comments:\n\n*Paper clarity:\n**I think there should be more information on how the underlying Chemgrapher model converts 2D molecular images to the molecular graph structures. How does the model actually construct the graph structure? Also, there is a lot of analysis on the image segmentation part of the model (eg figure 5, 6). Does the image classification part of the Chemgrapher model play any significant role? \n**Information about the Indigo and Maybridge dataset could be provided in a more accessible way. Eg the total number of examples in each dataset\n**Some additional information about how the original model was pre-trained would be useful\n\n*I think there needs to be more justification for why this proposed approach [option 1] of mapping the 2D molecule image to the molecular graph structure via an intermediate representation that explicitly identifies all the atoms and bonds in the image is preferred over the alternative approach [option 2] of directly mapping the 2D molecule image to the molecular graph structure (eg using by outputting a smiles text representation that can be converted to the molecular graph). The cost of option 1 is that it requires the very expensive pixel-wise labels that describes the locations of atoms and bonds in the 2D molecule image to train the segmentation model, thus motivating the domain adaptation part of this work. In terms of pure performance, it’s not clear to me that option 1 is superior, for example, [ref 30] which directly predicts the molecular smiles from the 2D molecular image [option 2] can attain ~80% in the indigo dataset, while this proposed approach seems to attain only ~40%\n\n*Figures 2 and 3 show various performance metrics over multiple iterations of the re-training. How did you decide which iteration to stop the re-training? Also, how computationally expensive is it to perform the iterative re-training procedure?\n\n*In Figure 3b, the performance at iteration 0 is ~72%, which I’m understanding to be the vanilla performance of the Chemgrapher model on the Maybridge dataset? But Table 1 shows that the performance of the Chemgrapher model on Maybridge is 83.3%\n\n*What is the reasoning for allowing a max of 2 node substitutions or 1 edge substitution for the ‘correcting graph alignment’ case?\n\n*In the future, it would be interesting to see how this proposed method compares in this recently published benchmark: https://github.com/Kohulan/OCSR_Review. NB: out of scope for this ICLR submission since it was published after the paper submission deadline\n\n\nRef [30]: Joshua Staker, Kyle Marshall, Robert Abel, and Carolyn M. McQuaw. Molecular Structure\nExtraction from Documents Using Deep Learning. J. Chem. Inf. Model., 59(3):1017–1029,\nMarch 2019. ISSN 1549-9596. doi: 10.1021/acs.jcim.8b00669. URL https://doi.org/10.1021/acs.jcim.8b00669.\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604034966169}, {"id": "XKrLwTGbHxg", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1739/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a domain adaption technique for self-labeling for strong/expensive planer graph labels given normal labels. For the application of molecular graphs, a graph alignment method on the planer graph level is proposed to find an isomorphism with a minimal edit distance of the predicted strong labels. The results show the proposed method can gradually correct the strong labels and improve the prediction performance with interpretable explanations. \n\nHowever, there are several concerns about the paper:\n1. The correction requires close edit distances between wrong labels and correct labels, and thus requires a relatively accurate U->V function. A plot regarding the correct percentage related to the accuracy of the initial U->V function will be interesting.\n2. How are strong labels picked? The authors mention the strong labels are \"a selection of these 4,000 datapoints\". But ablation study on the selection, size, and quality(edit distance to other graphs) of those datapoints should be investigated. \n3. The cost for the optimization problem argmin|e| is not discussed. The normal complexity for a single datapoint is N^d where N is the number of distance 1 editing and d is the distance allowed. A faster searching algorithm with domain knowledge is expected.\n4. For the other methods mentioned in table 1, are strong labels required? The percentage of strong labels should be reported for a fair comparison. The performance with different initial correct percentage will be interesting to investigate. \n5. The generalization of the model is not discussed. The adaption only works when the error of the prediction strong labels is smaller than k edit distance. However, the prediction depends on the distribution of the strong labels, the quality of the pre-trained models, the size of the graph, etc.\n\nOther questions:\n1. The structure of the paper can be improved. For example, the background of chemical structure recognition and the settings could be introduced at the beginning of the paper. \n2. In the experiments part, why the percentage correct decrease with iterations? \n3. In the graph alignment part, why to introduce sub-graph isomorphism?\n4. What is the relationship between function U->V, V->W, the segmentation network, the classification network?\n5. The segmentation network uses 134K  images. How is this data related to the training data?\n\nOverall, the idea of gradually correcting strong labels using graph alignment is interesting, but more discussions and results are required to make the paper stronger.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Self-Labeling of Fully Mediating Representations by Graph Alignment", "review": "The authors propose a domain adaption technique for self-labeling for strong/expensive planer graph labels given normal labels. For the application of molecular graphs, a graph alignment method on the planer graph level is proposed to find an isomorphism with a minimal edit distance of the predicted strong labels. The results show the proposed method can gradually correct the strong labels and improve the prediction performance with interpretable explanations. \n\nHowever, there are several concerns about the paper:\n1. The correction requires close edit distances between wrong labels and correct labels, and thus requires a relatively accurate U->V function. A plot regarding the correct percentage related to the accuracy of the initial U->V function will be interesting.\n2. How are strong labels picked? The authors mention the strong labels are \"a selection of these 4,000 datapoints\". But ablation study on the selection, size, and quality(edit distance to other graphs) of those datapoints should be investigated. \n3. The cost for the optimization problem argmin|e| is not discussed. The normal complexity for a single datapoint is N^d where N is the number of distance 1 editing and d is the distance allowed. A faster searching algorithm with domain knowledge is expected.\n4. For the other methods mentioned in table 1, are strong labels required? The percentage of strong labels should be reported for a fair comparison. The performance with different initial correct percentage will be interesting to investigate. \n5. The generalization of the model is not discussed. The adaption only works when the error of the prediction strong labels is smaller than k edit distance. However, the prediction depends on the distribution of the strong labels, the quality of the pre-trained models, the size of the graph, etc.\n\nOther questions:\n1. The structure of the paper can be improved. For example, the background of chemical structure recognition and the settings could be introduced at the beginning of the paper. \n2. In the experiments part, why the percentage correct decrease with iterations? \n3. In the graph alignment part, why to introduce sub-graph isomorphism?\n4. What is the relationship between function U->V, V->W, the segmentation network, the classification network?\n5. The segmentation network uses 134K  images. How is this data related to the training data?\n\nOverall, the idea of gradually correcting strong labels using graph alignment is interesting, but more discussions and results are required to make the paper stronger.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603976957120}, {"id": "seOkrlNmZn-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1739/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\nSummary:\nThis article presents a methodology to generate the complete molecular graph (including connectivity between atoms and functional groups) from a 2D image by using an intermediate representation (called fully mediating representation, V). This problem is of high relevance and is likely to have a considerable impact in the chemistry community. The particular advance of this work is to include such intermediate representation based on a graph alignment approach that generates “strong” labels. \n##########################################################################\nReasons for score: \nOverall, I vote for not accepting. As mentioned before, this problem is of high relevance and is likely to have a considerable impact in the chemistry community, I see myself using it, in particular if it is paired with SMILES. Nevertheless, my major concern is about the clarity of the paper, but I think that beyond clarity, the lack of the full or at least partial code should be available from the beginning for reviewers. This would make the evaluation of the article much easier. \n\n##########################################################################Pros: \n \n1. The paper addresses an interesting problem that, if solved, every chemist would like to make use of it in everyday life. \n \n2. The proposed method uses and intermediate step in the learning process to encode a higher complexity as well as information. This provides flexibility that is reflected in a marginal gain in error prediction compared to other methods but it needs a lower number of data points. \n \n##########################################################################\nCons: \nAbstract: The abstract should be better structured. It should state first what the framework is or the context in which this work is set. The proceed to the specifics and the technical part. \n\nA specific motivation for the use of the mediating representation is missing. \n\nBe specific, “relatively low number of data points” doesn't provide any information. \n\nThe authors repeatedly use “strong labels are expensive”, just for the sake of completeness, would be good to be explicit in this fact instead of assuming that the reader will infer what the meaning is. \n\nThe authors state: “In order to measure empirically the performance of our method of self-labeling fully mediating representations we start with a pre-trained model and perform two steps.” What type of model or trained on what? It should be specific. \n\nThe reference “Slot attention” should be described more in detail and the advantages of the authors’ method over the Hungarian algorithm should be very clear.\n\nI think a clear sentence is missing to describe in detail what the authors mean by strong and weak labels. In this regard, the section “Weak Supervision” is not straightforward to read. It should be rewritten or rearranged to make a better and clear reading. \n\n“We also assume the map E(v) which gives all allowed graph edits for the graph v” it is not clear.\n\nWhat is the origin on the such a different performance on the Indigo and Maybridge datasets.\n\nFig. 5 and 6 nicely summarise the good performance of the model, but in order to understand better the method and its limitations or type of graphs that struggle with, it would be good to present out-layers where the system doesn't work.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Report on \"Self-Labeling of Fully Mediating Representations by Graph Alignment\"", "review": "##########################################################################\nSummary:\nThis article presents a methodology to generate the complete molecular graph (including connectivity between atoms and functional groups) from a 2D image by using an intermediate representation (called fully mediating representation, V). This problem is of high relevance and is likely to have a considerable impact in the chemistry community. The particular advance of this work is to include such intermediate representation based on a graph alignment approach that generates “strong” labels. \n##########################################################################\nReasons for score: \nOverall, I vote for not accepting. As mentioned before, this problem is of high relevance and is likely to have a considerable impact in the chemistry community, I see myself using it, in particular if it is paired with SMILES. Nevertheless, my major concern is about the clarity of the paper, but I think that beyond clarity, the lack of the full or at least partial code should be available from the beginning for reviewers. This would make the evaluation of the article much easier. \n\n##########################################################################Pros: \n \n1. The paper addresses an interesting problem that, if solved, every chemist would like to make use of it in everyday life. \n \n2. The proposed method uses and intermediate step in the learning process to encode a higher complexity as well as information. This provides flexibility that is reflected in a marginal gain in error prediction compared to other methods but it needs a lower number of data points. \n \n##########################################################################\nCons: \nAbstract: The abstract should be better structured. It should state first what the framework is or the context in which this work is set. The proceed to the specifics and the technical part. \n\nA specific motivation for the use of the mediating representation is missing. \n\nBe specific, “relatively low number of data points” doesn't provide any information. \n\nThe authors repeatedly use “strong labels are expensive”, just for the sake of completeness, would be good to be explicit in this fact instead of assuming that the reader will infer what the meaning is. \n\nThe authors state: “In order to measure empirically the performance of our method of self-labeling fully mediating representations we start with a pre-trained model and perform two steps.” What type of model or trained on what? It should be specific. \n\nThe reference “Slot attention” should be described more in detail and the advantages of the authors’ method over the Hungarian algorithm should be very clear.\n\nI think a clear sentence is missing to describe in detail what the authors mean by strong and weak labels. In this regard, the section “Weak Supervision” is not straightforward to read. It should be rewritten or rearranged to make a better and clear reading. \n\n“We also assume the map E(v) which gives all allowed graph edits for the graph v” it is not clear.\n\nWhat is the origin on the such a different performance on the Indigo and Maybridge datasets.\n\nFig. 5 and 6 nicely summarise the good performance of the model, but in order to understand better the method and its limitations or type of graphs that struggle with, it would be good to present out-layers where the system doesn't work.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603926613237}, {"id": "oip8fx2MU05", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1739/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Clarity\nThe paper is well written. The relevant work, introduction and main text allows the user to understand the problem easily.\n\nOriginality\nPaper uses intermediate layers (called mediating layers) as a concept to solve the problem of mapping  an input (that can represent a graph on a 2D image) to the graph structure. The base system for segmentation and classification is derived from previous work. The paper introduces graph aligning the mediating layer to the ground truth labels and propose an algorithm for the process. The originality is in the proposal of graph alignment as an intermediate step to generate additional training data while allowing different edit corrections.\n\nQuality\nThe experimentation showing the effectiveness of using a mediating layer compared to the base seems adequate for the 2 datasets. However there doesn't seem to be a good comparison to previous work. Moreover the previous work had different sampling (number of training and test samples) and is hard to understand if Table 1 is useful i.e. the comparison of numbers aren't useful. Having a clear fair comparison would have suggested the mediating layer's use more strongly.\n\nSignificance\nThe work claims to improve previous performance by a few percent points but as noted above the comparison may not be proper.\n\nQuestions\n1. Since the values of V are not really known (both during training and testing), why are the labels termed strong labels?\ni.e. is the correctness of the planar graph verified at any stage?\n\n2. There is a statement that the mediating layer can make the model more interpretable. But there is no further discussion in the paper on how this is true.\n\n3. It is not clear how the system is able to correctly classify never observed atoms and bonds in test examples. Is this valid only during training and alignment? Otherwise how does it learn to classify them?\n\n4. Above the key contribution section on page 2, there is mention that after domain adaptation, the performance is checked on the same test data. Why is the use of doing this? And why would we expect better results than not using domain adaptation?\n\n5. How is set of V initialized? Is it an empty set in the beginning? In algo 1, input seems to mention there are already m strong labels before the start, does that need some correction? Also, perhaps one line needs update of S\nS <- appendStrongLabels(T, (u,v));\n\n6. Do iteration 0 always correspond to the Oldenhof model or were they different models, not clear from the text other than mention that it was pre-trained?  \n\n\n\n\nGoal of the paper is to learn a function that maps an input (that can represent a graph on a 2D image) to the graph structure. Paper has experiments showing importance of mediating layer but doesn't make similar comparisions to previous work making it harder to understand if the mediating layer is really useful or not.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Goal of the paper is to learn a function that maps an input (that can represent a graph on a 2D image) to the graph structure. Paper has experiments showing importance of mediating layer but doesn't make similar comparisions to previous work making it harder to understand if the mediating layer is really useful or not.", "review": "Clarity\nThe paper is well written. The relevant work, introduction and main text allows the user to understand the problem easily.\n\nOriginality\nPaper uses intermediate layers (called mediating layers) as a concept to solve the problem of mapping  an input (that can represent a graph on a 2D image) to the graph structure. The base system for segmentation and classification is derived from previous work. The paper introduces graph aligning the mediating layer to the ground truth labels and propose an algorithm for the process. The originality is in the proposal of graph alignment as an intermediate step to generate additional training data while allowing different edit corrections.\n\nQuality\nThe experimentation showing the effectiveness of using a mediating layer compared to the base seems adequate for the 2 datasets. However there doesn't seem to be a good comparison to previous work. Moreover the previous work had different sampling (number of training and test samples) and is hard to understand if Table 1 is useful i.e. the comparison of numbers aren't useful. Having a clear fair comparison would have suggested the mediating layer's use more strongly.\n\nSignificance\nThe work claims to improve previous performance by a few percent points but as noted above the comparison may not be proper.\n\nQuestions\n1. Since the values of V are not really known (both during training and testing), why are the labels termed strong labels?\ni.e. is the correctness of the planar graph verified at any stage?\n\n2. There is a statement that the mediating layer can make the model more interpretable. But there is no further discussion in the paper on how this is true.\n\n3. It is not clear how the system is able to correctly classify never observed atoms and bonds in test examples. Is this valid only during training and alignment? Otherwise how does it learn to classify them?\n\n4. Above the key contribution section on page 2, there is mention that after domain adaptation, the performance is checked on the same test data. Why is the use of doing this? And why would we expect better results than not using domain adaptation?\n\n5. How is set of V initialized? Is it an empty set in the beginning? In algo 1, input seems to mention there are already m strong labels before the start, does that need some correction? Also, perhaps one line needs update of S\nS <- appendStrongLabels(T, (u,v));\n\n6. Do iteration 0 always correspond to the Oldenhof model or were they different models, not clear from the text other than mention that it was pre-trained?  \n\n\n\n\nGoal of the paper is to learn a function that maps an input (that can represent a graph on a 2D image) to the graph structure. Paper has experiments showing importance of mediating layer but doesn't make similar comparisions to previous work making it harder to understand if the mediating layer is really useful or not.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603786073493}], "openreview_url": "https://openreview.net/forum?id=XEw5Onu69uu", "arxiv_id": "2103.14133", "paper_pdf": "papers/XEw5Onu69uu.pdf", "paper_pdf_sha256": "11b887ee3ea1e57b05d84ba267dfda746ee2aaede94e6d5aa21e41f344f241ee", "paper_pdf_bytes": 1208307, "paper_pdf_source": "openreview", "code_url": "https://github.com/biolearning-stadius/chemgrapher-self-rich-labeling", "code_repository": "biolearning-stadius/chemgrapher-self-rich-labeling", "code_commit": "7018a7d2ea0e9288dc94cdce5fb7024d4c14cd55", "code_archive": "repos/XEw5Onu69uu.zip", "code_archive_sha256": "a5fd427ca5577951a8b26772e7ccb79c1c0f3e311811b6a5a21b412f8fa5f1f1", "code_archive_bytes": 1553802, "code_file_count": 23, "code_extensions": {".py": 22, ".ipynb": 1}, "github_disk_usage_kb": 1635, "github_languages": {"Jupyter Notebook": 185146, "Python": 122012}, "github_archived": false, "github_pushed_at": "2021-09-02T09:39:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/self-labeling-of-fully-mediating-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qiOIqQ9CwA", "year": 2026, "status": "rejected", "title": "Optimal Stepsize for Diffusion Sampling", "authors": ["Jianning Pei", "Han Hu", "Shuyang Gu"], "authorids": ["~Jianning_Pei1", "~Han_Hu1", "~Shuyang_Gu1"], "authors_source": "OpenReview API", "abstract": "Diffusion models achieve remarkable generation quality but suffer from computational intensive sampling due to suboptimal step discretization. While existing works focus on optimizing denoising directions, we address the principled design of stepsize schedules. This paper proposes Optimal Stepsize Distillation, a dynamic programming framework that extracts theoretically optimal schedules by distilling knowledge from reference trajectories. By reformulating stepsize optimization as recursive error minimization, our method guarantees global discretization bounds through optimal substructure exploitation. Crucially, the distilled schedules demonstrate strong robustness across architectures, ODE solvers, and noise schedules. Experiments show 10x accelerated text-to-image generation while preserving 99.4% performance on GenEval.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "7RPXzfT2UB", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24253/Reviewer_3qCY"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a dynamic programming idea for obtaining the optimal stepsize in diffusion sampling process via recursive approximation. Empirically, it demonstrates the method across multiple architectures and solvers. The idea leads to upto 10x accelerated Text to image generation without loss in quality.", "review_text": "This paper proposes a dynamic programming idea for obtaining the optimal stepsize in diffusion sampling process via recursive approximation. Empirically, it demonstrates the method across multiple architectures and solvers. The idea leads to upto 10x accelerated Text to image generation without loss in quality.", "strengths": "- The idea of distilling a compact non-uniform schedule is pragmatic and useful.\n- Empirical gains are promising and seem to have been tested on models including  Masked Autoregressive generation model and OpenSora.", "weaknesses": "It seems like this paper has been formatted poorly. It is a bit difficult to read. Missing details:\n- L391 - The sentence is cutoff \n- L393 - Visualization results for Open-Sora are promised but not available in the Appendix it looks like.\n- Figure 5 does not seem to be referenced anywhere.\n\n\nOther Weaknesses:\n- Authors claim that the scheduler generalizes across ODE solvers but there is no ablation study to show this claim.\n- PSNR evaluation is insufficient. Metric like FID needs to be also evaluated.\n- Authors should also compare how much does the teacher trajectory generation + DP search cost compared to the sampling cost.\nThis factor may vary for various tasks and applications. Authors should consider discussing the costs of their method compared to plain sampling.  \n\nOverall, the motivation of this paper is good but the benefit is not clearly shown through experiments. Evaluations are not complete (missing metrics). Claims need to be supported with experiments and more evaluation metrics. This is why i am rejecting the paper.", "questions": "- Authors cite [Align Your Steps](https://arxiv.org/pdf/2404.14507) paper but do not compare with it. AYS does compare itself to deterministic solvers in Table-1. Can the authors comment why they chose to not compare?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a dynamic programming idea for obtaining the optimal stepsize in diffusion sampling process via recursive approximation. Empirically, it demonstrates the method across multiple architectures and solvers. The idea leads to upto 10x accelerated Text to image generation without loss in quality.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The idea of distilling a compact non-uniform schedule is pragmatic and useful.\n- Empirical gains are promising and seem to have been tested on models including  Masked Autoregressive generation model and OpenSora.", "weaknesses": "It seems like this paper has been formatted poorly. It is a bit difficult to read. Missing details:\n- L391 - The sentence is cutoff \n- L393 - Visualization results for Open-Sora are promised but not available in the Appendix it looks like.\n- Figure 5 does not seem to be referenced anywhere.\n\n\nOther Weaknesses:\n- Authors claim that the scheduler generalizes across ODE solvers but there is no ablation study to show this claim.\n- PSNR evaluation is insufficient. Metric like FID needs to be also evaluated.\n- Authors should also compare how much does the teacher trajectory generation + DP search cost compared to the sampling cost.\nThis factor may vary for various tasks and applications. Authors should consider discussing the costs of their method compared to plain sampling.  \n\nOverall, the motivation of this paper is good but the benefit is not clearly shown through experiments. Evaluations are not complete (missing metrics). Claims need to be supported with experiments and more evaluation metrics. This is why i am rejecting the paper.", "questions": "- Authors cite [Align Your Steps](https://arxiv.org/pdf/2404.14507) paper but do not compare with it. AYS does compare itself to deterministic solvers in Table-1. Can the authors comment why they chose to not compare?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761986015429}, {"id": "4mTrEZRgl0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24253/Reviewer_D3on"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 2, "summary": "This paper addresses the efficiency bottleneck of diffusion sampling by focusing on stepsize optimization, a relatively underexplored aspect compared to denoising direction refinement. The authors propose Optimal Stepsize Distillation (OSS), a dynamic programming (DP) framework that derives theoretically optimal stepsize schedules by minimizing global discretization error between a high-step “teacher” trajectory and a low-step “student” trajectory.", "review_text": "This paper addresses the efficiency bottleneck of diffusion sampling by focusing on stepsize optimization, a relatively underexplored aspect compared to denoising direction refinement. The authors propose Optimal Stepsize Distillation (OSS), a dynamic programming (DP) framework that derives theoretically optimal stepsize schedules by minimizing global discretization error between a high-step “teacher” trajectory and a low-step “student” trajectory.", "strengths": "1. The paper formalizes stepsize optimization as a dynamic programming problem, which is theoretically elegant and clear.\n\n2. The experimental design is comprehensive and solid, which includes both image tasks and video tasks.", "weaknesses": "No major concern, just suggestions or minor questions.\n\n1. Evaluation metrics: The paper mainly reports PSNR and FID, without including perceptual or semantic alignment metrics (e.g., CLIP-Score, ImageReward, or human preference), which would reflect generation quality in other views.\n\n2. Can the OSS framework be integrated jointly with solver optimizations such as DPM-Solver++ or UniPC in a unified search space, and if so, would further improvements emerge?\n\n\nAs I don't fully understand the theoretical analysis in this paper, I would rate 8 with a low confidence score.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the efficiency bottleneck of diffusion sampling by focusing on stepsize optimization, a relatively underexplored aspect compared to denoising direction refinement. The authors propose Optimal Stepsize Distillation (OSS), a dynamic programming (DP) framework that derives theoretically optimal stepsize schedules by minimizing global discretization error between a high-step “teacher” trajectory and a low-step “student” trajectory.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The paper formalizes stepsize optimization as a dynamic programming problem, which is theoretically elegant and clear.\n\n2. The experimental design is comprehensive and solid, which includes both image tasks and video tasks.", "weaknesses": "No major concern, just suggestions or minor questions.\n\n1. Evaluation metrics: The paper mainly reports PSNR and FID, without including perceptual or semantic alignment metrics (e.g., CLIP-Score, ImageReward, or human preference), which would reflect generation quality in other views.\n\n2. Can the OSS framework be integrated jointly with solver optimizations such as DPM-Solver++ or UniPC in a unified search space, and if so, would further improvements emerge?\n\n\nAs I don't fully understand the theoretical analysis in this paper, I would rate 8 with a low confidence score.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761984727204}, {"id": "zR7ceU6pkV", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission24253/Reviewer_YzEh"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes a dynamic programming-based framework, Optimal Stepsize Distillation (OSS), to optimize the step sizes (“stepsize schedule”) used in diffusion model sampling. Instead of relying on heuristic or hand-crafted stepsize choices, they recast the problem as a knowledge distillation task: a low-step (student) sampler aims to match the output of a high-step (teacher) sampler, minimizing global discretization error. The approach leverages the recursive structure of the sampling process and offers theoretically optimal solution via dynamic programming. OSS is shown to be architecture-agnostic and robust across different solvers, noise schedules, and tasks, enabling up to 10x acceleration in diffusion-based image and video generation with negligible loss in performance, as evidenced on the GenEval benchmark and other datasets. The paper provides extensive empirical results, detailed derivations, and justifies the approach with theoretical guarantees.", "review_text": "The paper proposes a dynamic programming-based framework, Optimal Stepsize Distillation (OSS), to optimize the step sizes (“stepsize schedule”) used in diffusion model sampling. Instead of relying on heuristic or hand-crafted stepsize choices, they recast the problem as a knowledge distillation task: a low-step (student) sampler aims to match the output of a high-step (teacher) sampler, minimizing global discretization error. The approach leverages the recursive structure of the sampling process and offers theoretically optimal solution via dynamic programming. OSS is shown to be architecture-agnostic and robust across different solvers, noise schedules, and tasks, enabling up to 10x acceleration in diffusion-based image and video generation with negligible loss in performance, as evidenced on the GenEval benchmark and other datasets. The paper provides extensive empirical results, detailed derivations, and justifies the approach with theoretical guarantees.", "strengths": "1. The recasting of stepsize optimization as a dynamic programming problem is well-motivated and rigorously constructed (see Section 3.3 and Lemma 3.1), with detailed equations and a clear algorithmic description (Algorithm 1 on pg. 14). The approach exploits the optimal substructure property, which is proven and illustrated.\n2. The framework is general—applying to both first- and high-order ODE solvers (see Algorithm 2 in Appendix B, and main Section 3.3.3), and decouples stepsize schedule from direction choice.\n3. Experiments are broad and robust, including ablation on noise schedules (Table 2, pg. 7), teacher stepsize granularity (Table 3, pg. 7), and evaluation across multiple datasets (Table 5, pg. 9). Figure 4 (pg. 6) illustrates how amplitude misalignment can be corrected by their calibration method—demonstrating engagement with practical sampling artifacts.", "weaknesses": "1. While the paper mentions several prior stepsize/schedule optimization works (e.g., GITS, DM, LD3), it does not clearly distinguish how OSS fundamentally surpasses or aligns with recent works such as DDSS [Watson et al., 2022] [1] and “Adaptive Time-Stepping Schedules for Diffusion Models” [Yuzhu et al., 2024] [2], both of which are missing from the reference and comparative experiments, and are strongly related in optimizing time steps or sampling schedules for improved efficiency. For instance, Section 2 does not articulate the improvement over gradient-based or convergence-bound approaches in these missing papers.\n2. While the proof of optimal substructure (Appendix A) is mathematically sound under certain assumptions, the practical optimality is limited by (a) the granularity of teacher steps (as the student schedule is a subset of teacher steps), and (b) the potential mismatch between $L_2$ error and perceptual or application-specific metrics (see ablation Table 6, Appendix E). The claim of “global discretization bounds” should be qualified with these approximations and the fact that, in practice, optimality is with respect to the chosen metric (e.g., PSNR or MSE), not necessarily downstream sample quality.\n3. The core algorithm minimizes $L_2$ distance between student and teacher trajectories (main Eq. 9, Alg. 1), but Table 6 (Appendix E) shows this choice can misalign with perceptual image metrics (e.g., Inception-V3 or feature distance). The methodology may lead to over-smoothing or suboptimal perceptual outcomes, as seen by the drop in PSNR after amplitude calibration (Section 4.2.1). This tradeoff between MSE/PSNR and perceptual detail is not thoroughly discussed or ablated.\n4. While Figure 4 and Section 3.4/4.2.1 demonstrate the need for amplitude calibration, this fix is relatively ad hoc (a simple affine per-step transformation) and lacks a thorough theoretical or empirical analysis of its impact, especially on more challenging data distributions or with colored noise. There is no ablation on its robustness, generalizability, or risk of overfitting/failure modes.\n5. While dynamic programming is, in theory, efficient, OSS appears to run a nontrivial search over a potentially large $N \\times M$ table (Algorithm 1, Figure 3). There is no detailed discussion or timing comparison of search overhead versus gains in sampling time, nor a complexity analysis, especially for high-order solvers or for real-world large models.\n\n[1] Watson D, Chan W, Ho J, et al. Learning fast samplers for diffusion models by differentiating through sample quality[C]//International Conference on Learning Representations. 2021.\n\n[2] Chen Y, He F, Fu S, et al. Adaptive time-stepping schedules for diffusion models[C]//The 40th Conference on Uncertainty in Artificial Intelligence. 2024.", "questions": "1. How does OSS perform relative to gradient-based sampler optimization methods such as DDSS [Watson et al., 2022], both in terms of sampling quality (FID, IS) and search/compute overhead? Is there an empirical or conceptual advantage in practice?\n2. Could the authors elaborate on the practical runtime/complexity tradeoff of the DP-based schedule search, especially for very deep teacher schedules or high-resolution images? Is the search cost amortized or negligible compared to model sampling cost?\n3. Are there plans or methodology to extend OSS to adaptive (test-time) schedule computation, avoiding pre-computed mean sequence or per-sample search, to further increase practicality?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a dynamic programming-based framework, Optimal Stepsize Distillation (OSS), to optimize the step sizes (“stepsize schedule”) used in diffusion model sampling. Instead of relying on heuristic or hand-crafted stepsize choices, they recast the problem as a knowledge distillation task: a low-step (student) sampler aims to match the output of a high-step (teacher) sampler, minimizing global discretization error. The approach leverages the recursive structure of the sampling process and offers theoretically optimal solution via dynamic programming. OSS is shown to be architecture-agnostic and robust across different solvers, noise schedules, and tasks, enabling up to 10x acceleration in diffusion-based image and video generation with negligible loss in performance, as evidenced on the GenEval benchmark and other datasets. The paper provides extensive empirical results, detailed derivations, and justifies the approach with theoretical guarantees.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The recasting of stepsize optimization as a dynamic programming problem is well-motivated and rigorously constructed (see Section 3.3 and Lemma 3.1), with detailed equations and a clear algorithmic description (Algorithm 1 on pg. 14). The approach exploits the optimal substructure property, which is proven and illustrated.\n2. The framework is general—applying to both first- and high-order ODE solvers (see Algorithm 2 in Appendix B, and main Section 3.3.3), and decouples stepsize schedule from direction choice.\n3. Experiments are broad and robust, including ablation on noise schedules (Table 2, pg. 7), teacher stepsize granularity (Table 3, pg. 7), and evaluation across multiple datasets (Table 5, pg. 9). Figure 4 (pg. 6) illustrates how amplitude misalignment can be corrected by their calibration method—demonstrating engagement with practical sampling artifacts.", "weaknesses": "1. While the paper mentions several prior stepsize/schedule optimization works (e.g., GITS, DM, LD3), it does not clearly distinguish how OSS fundamentally surpasses or aligns with recent works such as DDSS [Watson et al., 2022] [1] and “Adaptive Time-Stepping Schedules for Diffusion Models” [Yuzhu et al., 2024] [2], both of which are missing from the reference and comparative experiments, and are strongly related in optimizing time steps or sampling schedules for improved efficiency. For instance, Section 2 does not articulate the improvement over gradient-based or convergence-bound approaches in these missing papers.\n2. While the proof of optimal substructure (Appendix A) is mathematically sound under certain assumptions, the practical optimality is limited by (a) the granularity of teacher steps (as the student schedule is a subset of teacher steps), and (b) the potential mismatch between $L_2$ error and perceptual or application-specific metrics (see ablation Table 6, Appendix E). The claim of “global discretization bounds” should be qualified with these approximations and the fact that, in practice, optimality is with respect to the chosen metric (e.g., PSNR or MSE), not necessarily downstream sample quality.\n3. The core algorithm minimizes $L_2$ distance between student and teacher trajectories (main Eq. 9, Alg. 1), but Table 6 (Appendix E) shows this choice can misalign with perceptual image metrics (e.g., Inception-V3 or feature distance). The methodology may lead to over-smoothing or suboptimal perceptual outcomes, as seen by the drop in PSNR after amplitude calibration (Section 4.2.1). This tradeoff between MSE/PSNR and perceptual detail is not thoroughly discussed or ablated.\n4. While Figure 4 and Section 3.4/4.2.1 demonstrate the need for amplitude calibration, this fix is relatively ad hoc (a simple affine per-step transformation) and lacks a thorough theoretical or empirical analysis of its impact, especially on more challenging data distributions or with colored noise. There is no ablation on its robustness, generalizability, or risk of overfitting/failure modes.\n5. While dynamic programming is, in theory, efficient, OSS appears to run a nontrivial search over a potentially large $N \\times M$ table (Algorithm 1, Figure 3). There is no detailed discussion or timing comparison of search overhead versus gains in sampling time, nor a complexity analysis, especially for high-order solvers or for real-world large models.\n\n[1] Watson D, Chan W, Ho J, et al. Learning fast samplers for diffusion models by differentiating through sample quality[C]//International Conference on Learning Representations. 2021.\n\n[2] Chen Y, He F, Fu S, et al. Adaptive time-stepping schedules for diffusion models[C]//The 40th Conference on Uncertainty in Artificial Intelligence. 2024.", "questions": "1. How does OSS perform relative to gradient-based sampler optimization methods such as DDSS [Watson et al., 2022], both in terms of sampling quality (FID, IS) and search/compute overhead? Is there an empirical or conceptual advantage in practice?\n2. Could the authors elaborate on the practical runtime/complexity tradeoff of the DP-based schedule search, especially for very deep teacher schedules or high-resolution images? Is the search cost amortized or negligible compared to model sampling cost?\n3. Are there plans or methodology to extend OSS to adaptive (test-time) schedule computation, avoiding pre-computed mean sequence or per-sample search, to further increase practicality?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761298739692}], "openreview_url": "https://openreview.net/forum?id=qiOIqQ9CwA", "arxiv_id": "2503.21774", "paper_pdf": "papers/qiOIqQ9CwA.pdf", "paper_pdf_sha256": "ade4e5a2ed53aff1db1fe45f258548b09292d60c941a17e739fde0381e5ee9dd", "paper_pdf_bytes": 43910505, "paper_pdf_source": "openreview", "code_url": "https://github.com/bebebe666/OptimalSteps", "code_repository": "bebebe666/OptimalSteps", "code_commit": "ee350436c86e29088c0cf550300a2e6a5911a7ee", "code_archive": "repos/qiOIqQ9CwA.zip", "code_archive_sha256": "8a6e85af2d138a34cf05c7a4bc3fa53b1b4ac03a3c89140fd74e90305c7cfedc", "code_archive_bytes": 6127121, "code_file_count": 15, "code_extensions": {".py": 11, ".sh": 4}, "github_disk_usage_kb": 5993, "github_languages": {"Python": 17789, "Shell": 3320}, "github_archived": false, "github_pushed_at": "2025-04-13T05:54:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/optimal-stepsize-for-diffusion-sampling"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PIHPmNNp7w", "year": 2025, "status": "rejected", "title": "Retrieval-Augmented Decision Transformer: External Memory for In-context RL", "authors": ["Thomas Schmied", "Fabian Paischer", "Vihang Prakash Patil", "Markus Hofmarcher", "Razvan Pascanu", "Sepp Hochreiter"], "authorids": ["~Thomas_Schmied1", "~Fabian_Paischer1", "~Vihang_Prakash_Patil1", "~Markus_Hofmarcher1", "~Razvan_Pascanu1", "~Sepp_Hochreiter1"], "authors_source": "OpenReview API", "abstract": "In-context learning (ICL) is the ability of a model to learn a new task by observing a few exemplars in its context. While prevalent in NLP, this capability has recently also been observed in Reinforcement Learning (RL) settings. Prior in-context RL methods, however, require entire episodes in the agent's context. Given that complex environments typically lead to long episodes with sparse rewards, these methods are constrained to simple environments with short episodes. To address these challenges, we introduce Retrieval-Augmented Decision Transformer (RA-DT). RA-DT employs an external memory mechanism to store past experiences from which it retrieves only sub-trajectories relevant for the current situation. The retrieval component in RA-DT does not require training and can be entirely domain-agnostic. We evaluate the capabilities of RA-DT on grid-world environments, robotics simulations, and procedurally-generated video games. On grid worlds, RA-DT outperforms baselines, while using only a fraction of their context length. Furthermore, we illuminate the limitations of current in-context RL methods on complex environments and discuss future directions. To facilitate future research, we release datasets for four of the considered environments.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "betoaZFMy4", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9530/Reviewer_6hs5"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper tackles the challenge of in-context learning in complex RL environments, where long episodes and sparse rewards make traditional approaches difficult to apply effectively. To address this, the authors introduce the Retrieval-Augmented Decision Transformer (RA-DT), which utilizes an external memory to store past experiences. This memory enables the selective retrieval of relevant sub-trajectories based on the current context, allowing the model to make informed decisions without relying on entire episodes. Through extensive experiments, the authors demonstrate the effectiveness of RA-DT and show its superior performance over baseline methods. Additionally, they release datasets for multiple environments to support future research in in-context RL.", "review_text": "This paper tackles the challenge of in-context learning in complex RL environments, where long episodes and sparse rewards make traditional approaches difficult to apply effectively. To address this, the authors introduce the Retrieval-Augmented Decision Transformer (RA-DT), which utilizes an external memory to store past experiences. This memory enables the selective retrieval of relevant sub-trajectories based on the current context, allowing the model to make informed decisions without relying on entire episodes. Through extensive experiments, the authors demonstrate the effectiveness of RA-DT and show its superior performance over baseline methods. Additionally, they release datasets for multiple environments to support future research in in-context RL.", "strengths": "1. The proposed RA-DT leverages an external memory mechanism to store and retrieve relevant sub-trajectories, effectively tackling the challenge of managing context in environments with long episodes and sparse rewards.\n2. The authors conducted extensive experiments to showcase the method’s effectiveness and to highlight the contribution of each module.\n3. The authors released datasets for multiple environments, offering valuable resources to support future research in in-context RL.", "weaknesses": "1. Although the authors explain the rationale for using each module in RA-DT and demonstrate their effectiveness through experiments, the proposed RA-DT appears to be a combination of various existing methods [1, 2, 3]. What specific innovations do each of these methods introduce compared to the original approaches?\n2. The processes of searching forsimilar experiences and reweighting retrieved experiences introduce additional computational overhead, which is neither discussed in detail nor evaluated in the experiments. Could you discuss the computational overhead introduced by the search and reweighting processes, and how this compares to the baseline methods?\n\n[1] History compression via language models in reinforcement learning.\n[2] Retrieval-augmented multimodal language modeling.\n[3] Generative agents: Interactive simulacra of human behavior.\n\nI look forward to the authors' response and am open to further discussion.", "questions": "1. Could you provide information on the runtime and computational resources required for the proposed method compared to the baseline methods? (The same as the second point in weaknesses)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the challenge of in-context learning in complex RL environments, where long episodes and sparse rewards make traditional approaches difficult to apply effectively. To address this, the authors introduce the Retrieval-Augmented Decision Transformer (RA-DT), which utilizes an external memory to store past experiences. This memory enables the selective retrieval of relevant sub-trajectories based on the current context, allowing the model to make informed decisions without relying on entire episodes. Through extensive experiments, the authors demonstrate the effectiveness of RA-DT and show its superior performance over baseline methods. Additionally, they release datasets for multiple environments to support future research in in-context RL.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The proposed RA-DT leverages an external memory mechanism to store and retrieve relevant sub-trajectories, effectively tackling the challenge of managing context in environments with long episodes and sparse rewards.\n2. The authors conducted extensive experiments to showcase the method’s effectiveness and to highlight the contribution of each module.\n3. The authors released datasets for multiple environments, offering valuable resources to support future research in in-context RL.", "weaknesses": "1. Although the authors explain the rationale for using each module in RA-DT and demonstrate their effectiveness through experiments, the proposed RA-DT appears to be a combination of various existing methods [1, 2, 3]. What specific innovations do each of these methods introduce compared to the original approaches?\n2. The processes of searching forsimilar experiences and reweighting retrieved experiences introduce additional computational overhead, which is neither discussed in detail nor evaluated in the experiments. Could you discuss the computational overhead introduced by the search and reweighting processes, and how this compares to the baseline methods?\n\n[1] History compression via language models in reinforcement learning.\n[2] Retrieval-augmented multimodal language modeling.\n[3] Generative agents: Interactive simulacra of human behavior.\n\nI look forward to the authors' response and am open to further discussion.", "questions": "1. Could you provide information on the runtime and computational resources required for the proposed method compared to the baseline methods? (The same as the second point in weaknesses)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730533060108}, {"id": "DSpy3MdPu0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9530/Reviewer_FDLc"], "rating": 5, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The work introduces a variant of transformer-based offline RL by modifying the Decision Transformer architecture by an external memory component and associated retrieval mechanism. The retrieval-augmented decision transformers (RA-DT) retrieval mechanism enables retrieval of relevant sub-trajectories such that the in-context learning behavior of the DT is improved. The approach is evaluated on a variety of RL problems, though strong benefits of the proposed methods can only be shown on grid-world environments.", "review_text": "The work introduces a variant of transformer-based offline RL by modifying the Decision Transformer architecture by an external memory component and associated retrieval mechanism. The retrieval-augmented decision transformers (RA-DT) retrieval mechanism enables retrieval of relevant sub-trajectories such that the in-context learning behavior of the DT is improved. The approach is evaluated on a variety of RL problems, though strong benefits of the proposed methods can only be shown on grid-world environments.", "strengths": "The retrieval-augmentation idea is compelling. It particularly shines in its ability of cutting down the context length that is required for in-context reinforcement learning.\nFurther, the idea of utilizing pre-trained language models for the domain-agnostic embedding is very interesting and could be useful in this type of in-context RL. \nLastly, the work evaluates the approach on a broad variety of problems including more toy-like problems to larger and more complex ones.", "weaknesses": "While the work provides a broad evaluation of the proposed method, the experiments highlight the shortcomings of the method. Often the proposed method does not achieve better in-context learning abilities, particularly on problems that are not grid-worlds. Due to these shortcomings the authors provide a longer discussion section on the potential shortcomings of offline in-context RL methods, though this discussion does not provide explanations or better understanding why the proposed method did not work. In short, this section leaves many open questions about RA-DT .\nSimilarly, the domain agnostic embedding idea is put in the spotlight as a big benefit. However, experiments with this embedding seem to only have been conducted in the simplest setting where RA-DT works and there it does not seem to be a generally good choice. Further investigation is needed to properly understand if the domain agnostic approach actually  is a good choice or if it was simply random chance that it worked for some of the simplest cases.\nAs one central aspect of this work is concerned with reducing the context size needed for action prediction, I would have expected (at least a short) discussion on the work of Melo, ICML 22 (\"Transformers are Meta-Reinforcement Learners\") as this approach uses a \"working memory\" of only the five most recent transitions, when learning to reinforcement learn with a transformer. While this work might not be in the exact ICL learning regime as the work under review, I believe that they are related enough that this work warrants further comparison. Particularly of interest should be that in Melos work longer horizons become detrimental for learning and the learned agent struggles to generalize to completely unseen settings. A followup work to Melos (Shala et al. 2024 (\"Hierarchical Transformers are Efficient Meta-Reinforcement Learners\")) aims at improving the ICL abilities of this approach by introducing a \"cross-episodic\" memory where not only the most recent transitions are part of the working memory, but snippets of the most recent episodes as well. While this increases the overall used context size again, it does not require full episodes in the context and shows superior ICL on completely new test environments.", "questions": "* How does RA-DT + domain-agnostic compare to the baselines and RA-DT in the complex environments?\n* Why can the works on using Transformers for Meta-RL use much shorter context-sizes?\n* Which of the shortcomings of ICRL does RA-DT suffer from (the most) and which of the provided experiments show this?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work introduces a variant of transformer-based offline RL by modifying the Decision Transformer architecture by an external memory component and associated retrieval mechanism. The retrieval-augmented decision transformers (RA-DT) retrieval mechanism enables retrieval of relevant sub-trajectories such that the in-context learning behavior of the DT is improved. The approach is evaluated on a variety of RL problems, though strong benefits of the proposed methods can only be shown on grid-world environments.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "The retrieval-augmentation idea is compelling. It particularly shines in its ability of cutting down the context length that is required for in-context reinforcement learning.\nFurther, the idea of utilizing pre-trained language models for the domain-agnostic embedding is very interesting and could be useful in this type of in-context RL. \nLastly, the work evaluates the approach on a broad variety of problems including more toy-like problems to larger and more complex ones.", "weaknesses": "While the work provides a broad evaluation of the proposed method, the experiments highlight the shortcomings of the method. Often the proposed method does not achieve better in-context learning abilities, particularly on problems that are not grid-worlds. Due to these shortcomings the authors provide a longer discussion section on the potential shortcomings of offline in-context RL methods, though this discussion does not provide explanations or better understanding why the proposed method did not work. In short, this section leaves many open questions about RA-DT .\nSimilarly, the domain agnostic embedding idea is put in the spotlight as a big benefit. However, experiments with this embedding seem to only have been conducted in the simplest setting where RA-DT works and there it does not seem to be a generally good choice. Further investigation is needed to properly understand if the domain agnostic approach actually  is a good choice or if it was simply random chance that it worked for some of the simplest cases.\nAs one central aspect of this work is concerned with reducing the context size needed for action prediction, I would have expected (at least a short) discussion on the work of Melo, ICML 22 (\"Transformers are Meta-Reinforcement Learners\") as this approach uses a \"working memory\" of only the five most recent transitions, when learning to reinforcement learn with a transformer. While this work might not be in the exact ICL learning regime as the work under review, I believe that they are related enough that this work warrants further comparison. Particularly of interest should be that in Melos work longer horizons become detrimental for learning and the learned agent struggles to generalize to completely unseen settings. A followup work to Melos (Shala et al. 2024 (\"Hierarchical Transformers are Efficient Meta-Reinforcement Learners\")) aims at improving the ICL abilities of this approach by introducing a \"cross-episodic\" memory where not only the most recent transitions are part of the working memory, but snippets of the most recent episodes as well. While this increases the overall used context size again, it does not require full episodes in the context and shows superior ICL on completely new test environments.", "questions": "* How does RA-DT + domain-agnostic compare to the baselines and RA-DT in the complex environments?\n* Why can the works on using Transformers for Meta-RL use much shorter context-sizes?\n* Which of the shortcomings of ICRL does RA-DT suffer from (the most) and which of the provided experiments show this?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730274754803}, {"id": "pXwVRydPZJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9530/Reviewer_Yuoy"], "rating": 1, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper introduces Retrieval-Augmented Decision Transformer (RA-DT), a method that employs an external memory mechanism to store and retrieve relevant past experiences, enabling efficient in-context reinforcement learning in environments with long episodes and sparse rewards.", "review_text": "This paper introduces Retrieval-Augmented Decision Transformer (RA-DT), a method that employs an external memory mechanism to store and retrieve relevant past experiences, enabling efficient in-context reinforcement learning in environments with long episodes and sparse rewards.", "strengths": "-Model agnostic Retrieval-Augmentation\n\n-Smaller context length and boost in Performance", "weaknesses": "1) Novelty: The novelty of this work seems limited as it primarily leverages existing techniques like retrieval augmentation and Decision Transformers.\n\n2) Qaulity Data Availability and Relevance: If quality relevant data is not available is in storage then it is not very helpful.", "questions": "1) Technical Challenges: Could you please elaborate on the specific technical challenges (that authors solve) in this work?  (Not including what prior works do). If Retrieval Augmentation and decision transformer already exist then what is the contribution?\n\n2) Large Context Length and Retrieval Augmentation: How does your approach handle longer trajectories, and what are the potential limitations of using large context lengths in retrieval augmentation?\n\n3) Domain-Agnostic Model Performance: In Figure 3b, why does the RA-DT + domain-agnostic model exhibit a performance decline compared to the RA-DT model?\n\n4) In Figure 4, does running for more episodes ensure other approaches converge? Can you present those results to ensure baselines converge but require more samples?\n\n5) Computational Efficiency: How does the computational cost of your approach compare to the baseline methods, especially in terms of inference time?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Retrieval-Augmented Decision Transformer (RA-DT), a method that employs an external memory mechanism to store and retrieve relevant past experiences, enabling efficient in-context reinforcement learning in environments with long episodes and sparse rewards.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "-Model agnostic Retrieval-Augmentation\n\n-Smaller context length and boost in Performance", "weaknesses": "1) Novelty: The novelty of this work seems limited as it primarily leverages existing techniques like retrieval augmentation and Decision Transformers.\n\n2) Qaulity Data Availability and Relevance: If quality relevant data is not available is in storage then it is not very helpful.", "questions": "1) Technical Challenges: Could you please elaborate on the specific technical challenges (that authors solve) in this work?  (Not including what prior works do). If Retrieval Augmentation and decision transformer already exist then what is the contribution?\n\n2) Large Context Length and Retrieval Augmentation: How does your approach handle longer trajectories, and what are the potential limitations of using large context lengths in retrieval augmentation?\n\n3) Domain-Agnostic Model Performance: In Figure 3b, why does the RA-DT + domain-agnostic model exhibit a performance decline compared to the RA-DT model?\n\n4) In Figure 4, does running for more episodes ensure other approaches converge? Can you present those results to ensure baselines converge but require more samples?\n\n5) Computational Efficiency: How does the computational cost of your approach compare to the baseline methods, especially in terms of inference time?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729976181512}], "openreview_url": "https://openreview.net/forum?id=PIHPmNNp7w", "arxiv_id": "2410.07071", "paper_pdf": "papers/PIHPmNNp7w.pdf", "paper_pdf_sha256": "1871449b61b208290f9be73ca01277721718faa459a8e1d3656a58d345234951", "paper_pdf_bytes": 11811785, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-jku/RA-DT", "code_repository": "ml-jku/RA-DT", "code_commit": "40adec5cc4a8f3aeef1e84c5a203eb55ebd9d481", "code_archive": "repos/PIHPmNNp7w.zip", "code_archive_sha256": "e0b5ab37a38cbba26b08aaaa8b1019a34a49501f0f2bd6e2fe9ec95e7bb84ac1", "code_archive_bytes": 384060, "code_file_count": 75, "code_extensions": {".py": 75}, "github_disk_usage_kb": 349, "github_languages": {"Python": 849387}, "github_archived": false, "github_pushed_at": "2024-10-27T16:10:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/retrieval-augmented-decision-transformer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GY1fKFXG5i", "year": 2024, "status": "rejected", "title": "Non-Vacuous Generalization Bounds for Large Language Models", "authors": ["Sanae Lotfi", "Marc Anton Finzi", "Yilun Kuang", "Tim G. J. Rudner", "Micah Goldblum", "Andrew Gordon Wilson"], "authorids": ["~Sanae_Lotfi1", "~Marc_Anton_Finzi1", "~Yilun_Kuang1", "~Tim_G._J._Rudner2", "~Micah_Goldblum1", "~Andrew_Gordon_Wilson1"], "authors_source": "OpenReview API", "abstract": "Modern language models can contain billions of parameters, raising the question of whether they can generalize beyond the training data or simply regurgitate their training corpora. We provide the first non-vacuous generalization bounds for pretrained large language models (LLMs), indicating that language models are capable of discovering regularities that generalize to unseen data. In particular, we derive a compression bound that is valid for the unbounded log-likelihood loss using prediction smoothing, and we extend the bound to handle subsampling, accelerating bound computation on massive datasets. To achieve the extreme level of compression required for non-vacuous generalization bounds, we devise SubLoRA, a low-dimensional non-linear parameterization. Using this approach, we find that larger models have better generalization bounds and are more compressible than smaller models.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "s9ciWh5QAI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5752/Reviewer_dFjh"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Low Rank Adaption where parameter updates $\\Delta W$ is taken to be a product of two lower rank matrices that can be learnt; is combined with subspace training that uses  a projection to/from lower dimension subspace and arithmetic coding; to propose a combination of the two suitable for LLMs. Under i.i.d. assumption, risk generalization bounds are derived for LLMs on token prediction task. These bounds (that depend on empirical risk) are computed for several methods, with the bounds, and the empirical performance is shown to be best for the proposed method.", "review_text": "Low Rank Adaption where parameter updates $\\Delta W$ is taken to be a product of two lower rank matrices that can be learnt; is combined with subspace training that uses  a projection to/from lower dimension subspace and arithmetic coding; to propose a combination of the two suitable for LLMs. Under i.i.d. assumption, risk generalization bounds are derived for LLMs on token prediction task. These bounds (that depend on empirical risk) are computed for several methods, with the bounds, and the empirical performance is shown to be best for the proposed method.", "strengths": "- Compression bounds obtained with a neat smoothing trick\n- Code provided and also one that does not require industrial grade resources (but see Questions)\n- Very nicely written and presented\n- Improves upon previously reported bounds\n- Context of research contributions and related work reviewed very well\n- A bound is given where the empirical risk can be computed over a small subsample of the dataset\n- Bounds computed for some  methods, bounds best for SubLoRA", "weaknesses": "**Task and Error**\n- I would suggest better clarity about what the task is (token prediction) early on in the paper and also in the abstract and methodology section.\n- Also perhaps more clarity about errors. When it is said top-1 error, do you mean the worst sequence, or the worst token? \n\n**I.I.D. Assumption**\n\t- i.i.d assumption is a very strong assumption\n\t- but then a workaround is found which considers sequences not tokens, but then it's unclear how the parameter L is chosen", "questions": "**Artifact**\n- It seems the provided code is running the glue suite, and the code involves LoRA, but not subLoRA? And the other folder seems to be from nanoGPT?\n\t- Glue suite seems to be unrelated to the task in paper: token prediction?\n- At first because of the copyright line in the provided file, I was going to make a comment how it seems double blindness of the review is compromised as it seemed the submission was from the Hugging Face Team, but then I realized it is a minor modification of the same file from the Hugging Face Repository\n- Can you please check if the provided code is what you intended to submit? And let me know if I am missing something? And also some rough system specifications on which it can be run?\n\n**Comparison with Other Methods**\n- when comparing LoRA and Subspace (e.g. in figure 1) we see a comparison of train error, but not test error. Is there a reason for that?\n\n**I.I.D Assumption**\n- Can you please comment on why, despite the i.i.d. assumption, these results are still significant?\n\n**Possible Minor Typos / Formatting / Clarity** \n- page 6. 'payed'\n- abbreviation NLL used without definition\n- references need to be reviewed for formatting\n- page 4, second last line, shouldn't it be \"u: = flatten( ...)\" instead of \"LoRA(u) := flatten ( ... )\"\n- page 6, last paragraph says we use \"several\" values for r, giving the impression they are more than two, but the appendix mentions two values {1,4}\n- page 14, second last paragraph: \"Choosing an overall ... \". Please recheck this for clarity.\n\n\nIn summary the only major reservations I have are about the provided code and the i.i.d assumption, to the best of my knowledge. I apologize if I overlooked something important. Please feel free to correct me for any errors I may have made while reviewing; and to address these concerns. Thank you.\n\nMy vote is to for strong accept, conditional on addressing concerns regarding the prototype code.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Low Rank Adaption where parameter updates $\\Delta W$ is taken to be a product of two lower rank matrices that can be learnt; is combined with subspace training that uses  a projection to/from lower dimension subspace and arithmetic coding; to propose a combination of the two suitable for LLMs. Under i.i.d. assumption, risk generalization bounds are derived for LLMs on token prediction task. These bounds (that depend on empirical risk) are computed for several methods, with the bounds, and the empirical performance is shown to be best for the proposed method.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Compression bounds obtained with a neat smoothing trick\n- Code provided and also one that does not require industrial grade resources (but see Questions)\n- Very nicely written and presented\n- Improves upon previously reported bounds\n- Context of research contributions and related work reviewed very well\n- A bound is given where the empirical risk can be computed over a small subsample of the dataset\n- Bounds computed for some  methods, bounds best for SubLoRA", "weaknesses": "**Task and Error**\n- I would suggest better clarity about what the task is (token prediction) early on in the paper and also in the abstract and methodology section.\n- Also perhaps more clarity about errors. When it is said top-1 error, do you mean the worst sequence, or the worst token? \n\n**I.I.D. Assumption**\n\t- i.i.d assumption is a very strong assumption\n\t- but then a workaround is found which considers sequences not tokens, but then it's unclear how the parameter L is chosen", "questions": "**Artifact**\n- It seems the provided code is running the glue suite, and the code involves LoRA, but not subLoRA? And the other folder seems to be from nanoGPT?\n\t- Glue suite seems to be unrelated to the task in paper: token prediction?\n- At first because of the copyright line in the provided file, I was going to make a comment how it seems double blindness of the review is compromised as it seemed the submission was from the Hugging Face Team, but then I realized it is a minor modification of the same file from the Hugging Face Repository\n- Can you please check if the provided code is what you intended to submit? And let me know if I am missing something? And also some rough system specifications on which it can be run?\n\n**Comparison with Other Methods**\n- when comparing LoRA and Subspace (e.g. in figure 1) we see a comparison of train error, but not test error. Is there a reason for that?\n\n**I.I.D Assumption**\n- Can you please comment on why, despite the i.i.d. assumption, these results are still significant?\n\n**Possible Minor Typos / Formatting / Clarity** \n- page 6. 'payed'\n- abbreviation NLL used without definition\n- references need to be reviewed for formatting\n- page 4, second last line, shouldn't it be \"u: = flatten( ...)\" instead of \"LoRA(u) := flatten ( ... )\"\n- page 6, last paragraph says we use \"several\" values for r, giving the impression they are more than two, but the appendix mentions two values {1,4}\n- page 14, second last paragraph: \"Choosing an overall ... \". Please recheck this for clarity.\n\n\nIn summary the only major reservations I have are about the provided code and the i.i.d assumption, to the best of my knowledge. I apologize if I overlooked something important. Please feel free to correct me for any errors I may have made while reviewing; and to address these concerns. Thank you.\n\nMy vote is to for strong accept, conditional on addressing concerns regarding the prototype code.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698991771903}, {"id": "iAVimBweyo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5752/Reviewer_9VJi"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper aims to compute non-vacuous generalization bounds that apply to LLM pre-training. The authors employ a compression-based PAC-Bayes approach to achieve this goal. To obtain good compression they employ both LoRA and subspace training. Further, they apply a trick to bound the prediction probability of tokens so that the NLL becomes bounded, making it amenable to standard PAC-Bayes analysis. Finally, their experimental results show that their approach outperforms state-of-the-art generalization bounds.", "review_text": "The paper aims to compute non-vacuous generalization bounds that apply to LLM pre-training. The authors employ a compression-based PAC-Bayes approach to achieve this goal. To obtain good compression they employ both LoRA and subspace training. Further, they apply a trick to bound the prediction probability of tokens so that the NLL becomes bounded, making it amenable to standard PAC-Bayes analysis. Finally, their experimental results show that their approach outperforms state-of-the-art generalization bounds.", "strengths": "- The authors compute non-vacuous generalization bounds for LLM which can be challenging. \n- The paper is well-structured and the language is good. \n- The experimental results seem to outperform the current state of the art.", "weaknesses": "- While the paper is well-written for the most part, some parts are confusing. \n- The technical contribution is moderate on the conceptual part is fair, however, engineering a working non-vacuous bound can be challenging.", "questions": "- In Equation (2) it seems that LoRA is applied to $Pw$, it is not clear to me how that can be implemented, in particular, how the weights can have the forms $Pw$ and $UV$ simultaneously. It might, however, be just a typo and the equation should be $P \\cdot LoRA(w)$. \n- In section 4.4, it is not clear why we can assume that $\\hat{R}_{\\sigma_{i}}(h)$ are independent. In general, taking a random sample of a sequence does not make such a random sub-sample independent.\n\nA minor typo:\n$Q_1, Q_2 \\sim \\mathcal{N}(0,1)^{\\sqrt{D}\\times d}$ ----> $Q_1, Q_2 \\sim \\mathcal{N}(0,1)^{\\sqrt{D}\\times \\sqrt{d}}$", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to compute non-vacuous generalization bounds that apply to LLM pre-training. The authors employ a compression-based PAC-Bayes approach to achieve this goal. To obtain good compression they employ both LoRA and subspace training. Further, they apply a trick to bound the prediction probability of tokens so that the NLL becomes bounded, making it amenable to standard PAC-Bayes analysis. Finally, their experimental results show that their approach outperforms state-of-the-art generalization bounds.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The authors compute non-vacuous generalization bounds for LLM which can be challenging. \n- The paper is well-structured and the language is good. \n- The experimental results seem to outperform the current state of the art.", "weaknesses": "- While the paper is well-written for the most part, some parts are confusing. \n- The technical contribution is moderate on the conceptual part is fair, however, engineering a working non-vacuous bound can be challenging.", "questions": "- In Equation (2) it seems that LoRA is applied to $Pw$, it is not clear to me how that can be implemented, in particular, how the weights can have the forms $Pw$ and $UV$ simultaneously. It might, however, be just a typo and the equation should be $P \\cdot LoRA(w)$. \n- In section 4.4, it is not clear why we can assume that $\\hat{R}_{\\sigma_{i}}(h)$ are independent. In general, taking a random sample of a sequence does not make such a random sub-sample independent.\n\nA minor typo:\n$Q_1, Q_2 \\sim \\mathcal{N}(0,1)^{\\sqrt{D}\\times d}$ ----> $Q_1, Q_2 \\sim \\mathcal{N}(0,1)^{\\sqrt{D}\\times \\sqrt{d}}$", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698840173309}, {"id": "NYQRjGgVmh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5752/Reviewer_TUxG"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors aim to establish non-vacuous generalization bounds for large language models (LLMs) by employing a compression-based approach. The challenges they address include (1) dealing with the non-iid nature of tokens, (2) handling unbounded loss, and (3) managing extremely large model parameters. To tackle these issues, the authors propose specific solutions: for (1), leveraging entire sequences to treat tokens as iid; for (2), introducing a smoothed model to manage unbounded loss; and for (3), developing SubLoRA, a novel technique that combines LoRA and subspace training during the training process. Notably, the authors demonstrate the non-vacuous nature of the derived bound for GPT-2.\n\nIn consideration of the fact that this represents the first non-vacuous generalization bound in the context of Large Language Models (LLMs), I have assigned a moderately favorable score. It's important to note that I have not conducted an exhaustive investigation to confirm whether this is indeed the inaugural non-vacuous bound, and I am relying on the authors' assertion in this regard.\nMy reasons for not awarding a higher score can be attributed to the following factors:\n\n1. I hold the perspective that the existence of vacuous bounds in LLMs during the pretraining phase might not be of paramount significance.\n2. The authors primarily synthesize existing techniques rather than introducing fundamentally novel methods. It should be acknowledged that this does not necessarily translate to incremental progress since discovering these techniques is not a straightforward endeavor. However, a higher rating is withheld due to the absence of groundbreaking or distinctly tailored contributions to LLMs.", "review_text": "In this paper, the authors aim to establish non-vacuous generalization bounds for large language models (LLMs) by employing a compression-based approach. The challenges they address include (1) dealing with the non-iid nature of tokens, (2) handling unbounded loss, and (3) managing extremely large model parameters. To tackle these issues, the authors propose specific solutions: for (1), leveraging entire sequences to treat tokens as iid; for (2), introducing a smoothed model to manage unbounded loss; and for (3), developing SubLoRA, a novel technique that combines LoRA and subspace training during the training process. Notably, the authors demonstrate the non-vacuous nature of the derived bound for GPT-2.\n\nIn consideration of the fact that this represents the first non-vacuous generalization bound in the context of Large Language Models (LLMs), I have assigned a moderately favorable score. It's important to note that I have not conducted an exhaustive investigation to confirm whether this is indeed the inaugural non-vacuous bound, and I am relying on the authors' assertion in this regard.\nMy reasons for not awarding a higher score can be attributed to the following factors:\n\n1. I hold the perspective that the existence of vacuous bounds in LLMs during the pretraining phase might not be of paramount significance.\n2. The authors primarily synthesize existing techniques rather than introducing fundamentally novel methods. It should be acknowledged that this does not necessarily translate to incremental progress since discovering these techniques is not a straightforward endeavor. However, a higher rating is withheld due to the absence of groundbreaking or distinctly tailored contributions to LLMs.", "strengths": "1. The authors pioneer the introduction of a non-vacuous generalization bound for LLMs in the pretraining phase.\n2. The SubLoRA method, which amalgamates LoRA and subspace training, enhances model compressibility.\n3. The paper highlights and addresses various challenges in deriving generalization bounds for LLMs.\n4. Experimental verification of the proposed bounds adds credibility to the research.", "weaknesses": "1. While the authors emphasize the importance of non-vacuous bounds during the pretraining phase, a more detailed justification of its significance *within the LLM context* would enhance the paper's impact. At least I am not sure that for LLM, generalization in the pretraining phase is such important. \n2. Given the point 1, I wish to see some novel techniques. However, it seems that the techniques in this paper are not very novel. The methods presented in this paper largely combine existing techniques (e.g., SubLoRA). \n3. The paper claims that tokens exhibit non-iid behavior but resolves this issue by considering entire sequences, which might be considered a somewhat coarse approach.\n\nThe three points stop me from giving a higher score.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors aim to establish non-vacuous generalization bounds for large language models (LLMs) by employing a compression-based approach. The challenges they address include (1) dealing with the non-iid nature of tokens, (2) handling unbounded loss, and (3) managing extremely large model parameters. To tackle these issues, the authors propose specific solutions: for (1), leveraging entire sequences to treat tokens as iid; for (2), introducing a smoothed model to manage unbounded loss; and for (3), developing SubLoRA, a novel technique that combines LoRA and subspace training during the training process. Notably, the authors demonstrate the non-vacuous nature of the derived bound for GPT-2.\n\nIn consideration of the fact that this represents the first non-vacuous generalization bound in the context of Large Language Models (LLMs), I have assigned a moderately favorable score. It's important to note that I have not conducted an exhaustive investigation to confirm whether this is indeed the inaugural non-vacuous bound, and I am relying on the authors' assertion in this regard.\nMy reasons for not awarding a higher score can be attributed to the following factors:\n\n1. I hold the perspective that the existence of vacuous bounds in LLMs during the pretraining phase might not be of paramount significance.\n2. The authors primarily synthesize existing techniques rather than introducing fundamentally novel methods. It should be acknowledged that this does not necessarily translate to incremental progress since discovering these techniques is not a straightforward endeavor. However, a higher rating is withheld due to the absence of groundbreaking or distinctly tailored contributions to LLMs.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The authors pioneer the introduction of a non-vacuous generalization bound for LLMs in the pretraining phase.\n2. The SubLoRA method, which amalgamates LoRA and subspace training, enhances model compressibility.\n3. The paper highlights and addresses various challenges in deriving generalization bounds for LLMs.\n4. Experimental verification of the proposed bounds adds credibility to the research.", "weaknesses": "1. While the authors emphasize the importance of non-vacuous bounds during the pretraining phase, a more detailed justification of its significance *within the LLM context* would enhance the paper's impact. At least I am not sure that for LLM, generalization in the pretraining phase is such important. \n2. Given the point 1, I wish to see some novel techniques. However, it seems that the techniques in this paper are not very novel. The methods presented in this paper largely combine existing techniques (e.g., SubLoRA). \n3. The paper claims that tokens exhibit non-iid behavior but resolves this issue by considering entire sequences, which might be considered a somewhat coarse approach.\n\nThe three points stop me from giving a higher score.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698756129341}, {"id": "CeKoiHxrtn", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5752/Reviewer_DsFt"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper derives a compression bound that is valid for the unbounded log-likelihood loss using\nprediction smoothing, and we extend the bound to handle subsampling, accelerating bound computation on massive datasets. Using this approach, we find that larger models have better generalization bounds and are more compressible than\nsmaller models.", "review_text": "This paper derives a compression bound that is valid for the unbounded log-likelihood loss using\nprediction smoothing, and we extend the bound to handle subsampling, accelerating bound computation on massive datasets. Using this approach, we find that larger models have better generalization bounds and are more compressible than\nsmaller models.", "strengths": "1. This paper provides the first non-vacuous generalization bounds for LLM\npertaining by using extreme levels of model compression. The bounds suggest that compression\nbounds present new possibilities for understanding how and why language models generalize.\n2. The experiments verify their theoretical results.", "weaknesses": "None", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper derives a compression bound that is valid for the unbounded log-likelihood loss using\nprediction smoothing, and we extend the bound to handle subsampling, accelerating bound computation on massive datasets. Using this approach, we find that larger models have better generalization bounds and are more compressible than\nsmaller models.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. This paper provides the first non-vacuous generalization bounds for LLM\npertaining by using extreme levels of model compression. The bounds suggest that compression\nbounds present new possibilities for understanding how and why language models generalize.\n2. The experiments verify their theoretical results.", "weaknesses": "None", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "8: accept, good paper", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698582789717}, {"id": "waweZBUyak", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5752/Reviewer_GeY3"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors derive generalization bounds for auto-regressive\nlanguage models pre-training combining four ingredients:\n\n1.  a non-uniform hypothesis prior in the PAC-bound\n2.  the intrinsic dimension to bound the complexity of the neural model\n3. a form of label smoothing (prediction-smoothing) for bounding the\nnegative log-likelihood\n4. subsampling to reduce the cost\nof empirical risk estimation. \n\nThey also propose to combine LoRA with\nsubspace fine-tuning for finding the intrinsic dimension more\nefficiently. The empirical part demonstrates their generalization bounds by\nperforming experiments on GPT-2-style models with less than 200M parameters.", "review_text": "The authors derive generalization bounds for auto-regressive\nlanguage models pre-training combining four ingredients:\n\n1.  a non-uniform hypothesis prior in the PAC-bound\n2.  the intrinsic dimension to bound the complexity of the neural model\n3. a form of label smoothing (prediction-smoothing) for bounding the\nnegative log-likelihood\n4. subsampling to reduce the cost\nof empirical risk estimation. \n\nThey also propose to combine LoRA with\nsubspace fine-tuning for finding the intrinsic dimension more\nefficiently. The empirical part demonstrates their generalization bounds by\nperforming experiments on GPT-2-style models with less than 200M parameters.", "strengths": "1. A simple and elegant proposal, prediction smoothing, to accomodate the unbounded NLL loss in deriving generalization bounds.\n\n2. A novel proposal that combines LoRA with subspace training for pre-training on a lower dimensional subspace.\n\n3. An empirical verification of the effect that text structure has on generalization bounds.", "weaknesses": "1. The paper seems to lack a thorough comparison to previous theoretical work that\n  highlights the novel theoretical contributions. For example, the cited\n  **Aghajanyan et. al (2020)** proves a generalization bound for\n  classifiers that have been obtained by fine-tuning a pre-trained\n  language model. Their bound already relies on intrinsic dimension\n  (Ingredient 2) to reduce the hypothesis space and what seems to me a version of Ingredient 1 for compression based on **Arora et al**: *Stronger Generalization Bounds for Deep Nets via a Compression Approach*. Ingredients (1, 2) and a version of Ingredient (4) seem to be present in **Lofti et. al**.\n  \n* Exposition could be improved if the main generalization results\n  were stated as a Theorem with a discussion of the proof ingredients.\n  \n* Experiments are carried out on small model sizes (from the Appendix it seems\n  < 200M). It is then unclear if these findings would generalize to\n  large LMs, e.g. to the > 10B scale. The bounds for NLL seem\n  to improve in the tested scales, but it is unclear how these are related\n  to the generalization abilities in the few-shot or instruction following\n  capabilities that large LMs exhibit.", "questions": "My initial rating inclines towards rejection because I have some\nconcerns regarding:\n\n1. the novelty of the bounds and the theoretical arguments wrt to previous work\n2. the empirical part seems limited to small model sizes\n3. the benefits of SubLoRA in downstream applications are not clear.\n\nI am leaving some questions that would help me to improve\nmy assessment and in case increase the initial rating.\n\n**Questions**:\n\n1. Could you highlight the novel contributions and comparison to the work discussed in Weaknesses 1?\n\n2. What was the biggest model you trained and using how many tokens?\n\n3. In Table 1 is it the case that Subspace only is enough to achieve non-vacuous generalization bounds?\n\n 4. What is the tradeoff between SubLoRA and standard pre-training?\n \n5. If a pre-trained model exhibits few-shot capabilities, does its counterpart that was pre-trained with SubLoRA exhibits the same\n    abilities?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors derive generalization bounds for auto-regressive\nlanguage models pre-training combining four ingredients:\n\n1.  a non-uniform hypothesis prior in the PAC-bound\n2.  the intrinsic dimension to bound the complexity of the neural model\n3. a form of label smoothing (prediction-smoothing) for bounding the\nnegative log-likelihood\n4. subsampling to reduce the cost\nof empirical risk estimation. \n\nThey also propose to combine LoRA with\nsubspace fine-tuning for finding the intrinsic dimension more\nefficiently. The empirical part demonstrates their generalization bounds by\nperforming experiments on GPT-2-style models with less than 200M parameters.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. A simple and elegant proposal, prediction smoothing, to accomodate the unbounded NLL loss in deriving generalization bounds.\n\n2. A novel proposal that combines LoRA with subspace training for pre-training on a lower dimensional subspace.\n\n3. An empirical verification of the effect that text structure has on generalization bounds.", "weaknesses": "1. The paper seems to lack a thorough comparison to previous theoretical work that\n  highlights the novel theoretical contributions. For example, the cited\n  **Aghajanyan et. al (2020)** proves a generalization bound for\n  classifiers that have been obtained by fine-tuning a pre-trained\n  language model. Their bound already relies on intrinsic dimension\n  (Ingredient 2) to reduce the hypothesis space and what seems to me a version of Ingredient 1 for compression based on **Arora et al**: *Stronger Generalization Bounds for Deep Nets via a Compression Approach*. Ingredients (1, 2) and a version of Ingredient (4) seem to be present in **Lofti et. al**.\n  \n* Exposition could be improved if the main generalization results\n  were stated as a Theorem with a discussion of the proof ingredients.\n  \n* Experiments are carried out on small model sizes (from the Appendix it seems\n  < 200M). It is then unclear if these findings would generalize to\n  large LMs, e.g. to the > 10B scale. The bounds for NLL seem\n  to improve in the tested scales, but it is unclear how these are related\n  to the generalization abilities in the few-shot or instruction following\n  capabilities that large LMs exhibit.", "questions": "My initial rating inclines towards rejection because I have some\nconcerns regarding:\n\n1. the novelty of the bounds and the theoretical arguments wrt to previous work\n2. the empirical part seems limited to small model sizes\n3. the benefits of SubLoRA in downstream applications are not clear.\n\nI am leaving some questions that would help me to improve\nmy assessment and in case increase the initial rating.\n\n**Questions**:\n\n1. Could you highlight the novel contributions and comparison to the work discussed in Weaknesses 1?\n\n2. What was the biggest model you trained and using how many tokens?\n\n3. In Table 1 is it the case that Subspace only is enough to achieve non-vacuous generalization bounds?\n\n 4. What is the tradeoff between SubLoRA and standard pre-training?\n \n5. If a pre-trained model exhibits few-shot capabilities, does its counterpart that was pre-trained with SubLoRA exhibits the same\n    abilities?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698404441677}], "openreview_url": "https://openreview.net/forum?id=GY1fKFXG5i", "arxiv_id": "2312.17173", "paper_pdf": "papers/GY1fKFXG5i.pdf", "paper_pdf_sha256": "04d76647540f51fc646fc5a187ee916012d89d1963d8f7f1dcb99c89fc3c20e7", "paper_pdf_bytes": 362216, "paper_pdf_source": "openreview", "code_url": "https://github.com/Sanaelotfi/sublora-bounds-for-llms", "code_repository": "Sanaelotfi/sublora-bounds-for-llms", "code_commit": "c606ea664e54fe60cfd870167bf6d1c1183fedf1", "code_archive": "repos/GY1fKFXG5i.zip", "code_archive_sha256": "9e3f91b35553c7183df48379e9e9122da9c8283abe392b5447ee5eeca52c964f", "code_archive_bytes": 310471, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 300, "github_languages": {"Python": 197077}, "github_archived": false, "github_pushed_at": "2024-06-04T01:38:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/non-vacuous-generalization-bounds-for-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "cwiFbXPW4G0", "year": 2023, "status": "rejected", "title": "Learning for Edge-Weighted Online Bipartite Matching with Robustness Guarantees", "authors": ["Pengfei Li", "Jianyi Yang", "Shaolei Ren"], "authorids": ["~Pengfei_Li2", "~Jianyi_Yang1", "~Shaolei_Ren1"], "authors_source": "OpenReview API", "abstract": "Many real-world problems, such as online ad display, can be formulated as online bipartite matching. The crucial challenge lies in the nature of sequentially-revealed online item information, based on which we make irreversible matching decisions at each step. While numerous expert online algorithms have been proposed with bounded worst-case competitive ratios, they may not offer satisfactory performance in average cases. On the other hand, reinforcement learning (RL) has been applied to improve the average performance, but they lack robustness and can perform arbitrarily badly. In this paper, we propose a novel RL-based approach to edge-weighted online bipartite matching with robustness guarantees (LOMAR), achieving both good average-case and good worst-case performance. The key novelty of LOMAR is a new online switching operation which, based on a judiciously-designed condition to hedge against future uncertainties, decides whether to follow the expert's decision or the RL decision for each online item arrival. We prove that for any $\\rho \\in [0,1]$, LOMAR is $\\rho$-competitive against any given expert online algorithm. To improve the average performance, we  train the RL policy by explicitly considering the online switching operation. Finally, we run empirical experiments to demonstrate the advantages of LOMAR compared to existing baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "LYch1q2K2Gw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1962/Reviewer_XizC"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a RL-based approach for online weighted bipartite matching. A key novelty of the paper is to augment the decisions of the RL agent with a classic online algorithm to obtain robustness guarantees. Upon the arrival of an online vertex, the algorithm queries both a base online algorithm (called expert) and the RL agent. It follows the action of the RL agent unless doing so has the possibility that the algorithm incurs significant loss with respect to the expert. ", "review_text": "The paper proposes a framework to use an expert algorithm to guide the decisions of an RL agent for online bipartite matching. I like the problem direction but find the current paper version a bit lacking with respect to details of (i) RL training procedure, (ii) Difficulty in naively adapting standard RL techniques, (iii) thorough empirical evaluation.", "strengths": "Strengths:\n+ Unlike most work on learning augmented algorithms that aim to augment online algorithms with some predictions / advice from ML models; the paper takes the complementary view of an online algorithm used as a safety railing to guide a ML model.\n\nWeaknesses:\n- The model feels too brittle - in the general case when the edge weights are unknown - the algorithm essentially only follows the decisions of the expert and can not use the RL agent at all. Maybe a restriction to a more specific problem would yield better insights.\n\n- It’s not clear what the focus of the paper is. Section 4 is concerned with how to use the RL agent’s decisions while maintaining robustness against a fixed expert. It would be better to restructure this as a meta-algorithm that takes in two algorithms A (expert) and B (RL agent): The meta algorithm switches between the actions of A and B and guarantees to maintain a given robustness wrt A while maximizing the number of times it uses actions of B. (Note that this section is completely independent of RL). Indeed, viewed in this light, the meta-algorithm and its analysis is almost trivial. [ Also worth referring to “combining” algorithms for paging and metrical task systems (e.g. Fiat et al)]. \n\n- Section 5 then deals with the challenges of training the RL agent when used along with the switching algorithm above. I would much rather see more time and space allotted to this section and clarify the training process in more details.\n\n- The empirical section includes preliminary experiments and does not demonstrate strong positive results. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a RL-based approach for online weighted bipartite matching. A key novelty of the paper is to augment the decisions of the RL agent with a classic online algorithm to obtain robustness guarantees. Upon the arrival of an online vertex, the algorithm queries both a base online algorithm (called expert) and the RL agent. It follows the action of the RL agent unless doing so has the possibility that the algorithm incurs significant loss with respect to the expert. ", "strength_and_weaknesses": "Strengths:\n+ Unlike most work on learning augmented algorithms that aim to augment online algorithms with some predictions / advice from ML models; the paper takes the complementary view of an online algorithm used as a safety railing to guide a ML model.\n\nWeaknesses:\n- The model feels too brittle - in the general case when the edge weights are unknown - the algorithm essentially only follows the decisions of the expert and can not use the RL agent at all. Maybe a restriction to a more specific problem would yield better insights.\n\n- It’s not clear what the focus of the paper is. Section 4 is concerned with how to use the RL agent’s decisions while maintaining robustness against a fixed expert. It would be better to restructure this as a meta-algorithm that takes in two algorithms A (expert) and B (RL agent): The meta algorithm switches between the actions of A and B and guarantees to maintain a given robustness wrt A while maximizing the number of times it uses actions of B. (Note that this section is completely independent of RL). Indeed, viewed in this light, the meta-algorithm and its analysis is almost trivial. [ Also worth referring to “combining” algorithms for paging and metrical task systems (e.g. Fiat et al)]. \n\n- Section 5 then deals with the challenges of training the RL agent when used along with the switching algorithm above. I would much rather see more time and space allotted to this section and clarify the training process in more details.\n\n- The empirical section includes preliminary experiments and does not demonstrate strong positive results. \n", "clarity,_quality,_novelty_and_reproducibility": "I found the paper a bit hard to read and follow - as I mentioned earlier in the review, in my view, the paper focused on the wrong aspects of the work. The switching framework itself is not particularly novel within the online algorithms and is indeed the first thing one would think of when faced with such a question. ", "summary_of_the_review": "The paper proposes a framework to use an expert algorithm to guide the decisions of an RL agent for online bipartite matching. I like the problem direction but find the current paper version a bit lacking with respect to details of (i) RL training procedure, (ii) Difficulty in naively adapting standard RL techniques, (iii) thorough empirical evaluation.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666820654645}, {"id": "n_g3jr2qtkc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1962/Reviewer_D8g2"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies the problem of online weighted b-matching problem augmented with an RL-based expert algorithm. The goal of the RL algorithm is to leverage the expert such that when the RL algorithm's decision is bad while the expert algorithm's decision is good, the RL algorithm achieves robustness by using the expert algorithm's choice, while also having high enough average performance across the board, even when the expert is bad. They study this problem both in the free disposal and no-free disposal setting (where you can't throw away already matched edges) and show empirically competitive results that improves over the RL-only based algorithm.", "review_text": "Overall, I am on the border since I would like the empirical section to be strengthened. In particular, the paper introduces complex ideas and would like to see when this would significantly perform better than hand-tuned optimal online algorithms and when it would perform worse. This kind of extensive analysis will be useful to future research as well as a practitioner who would like to try these ideas in the wild. ", "strengths": "Strengths\n\n+ Robustifying RL algorithms are practically important. Online Matching problems are also practically important with many economic consequences. Thus, I find combing Robust RL area with online matching an important problem and understanding the limits, foundations and novel algorithms is a key contribution. This paper makes a step and adds to this direction.\n\n+ The paper is well-written and overall, the ideas are clearly explained. The online switching based algorithm is intuitive and practical. \n\n+ The paper considers an exhaustive space of both non-free disposal and free disposal settings which are both plently available in practical settings such as ads and recommender systems.\n\n\nWeakness\n\n- The primary weakness I find is in experiments. In particular, the studied problem is actually theoretically solvable using worst-case online algorithms when the inputs are stochastic. I wonder if the authors can compare against these optimal algorithms in the experimental section. In particular, when the inputs are i.i.d., these worst case algortihms can achieve competitive ratio > 0.7 in the worst-case and sometimes much higher on a random instance. Given that in the experiments, the competitive ratio obtained is of the order of 0.7-0.8, I think this comparison would make the paper really strong.\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper studies the problem of online weighted b-matching problem augmented with an RL-based expert algorithm. The goal of the RL algorithm is to leverage the expert such that when the RL algorithm's decision is bad while the expert algorithm's decision is good, the RL algorithm achieves robustness by using the expert algorithm's choice, while also having high enough average performance across the board, even when the expert is bad. They study this problem both in the free disposal and no-free disposal setting (where you can't throw away already matched edges) and show empirically competitive results that improves over the RL-only based algorithm.", "strength_and_weaknesses": "Strengths\n\n+ Robustifying RL algorithms are practically important. Online Matching problems are also practically important with many economic consequences. Thus, I find combing Robust RL area with online matching an important problem and understanding the limits, foundations and novel algorithms is a key contribution. This paper makes a step and adds to this direction.\n\n+ The paper is well-written and overall, the ideas are clearly explained. The online switching based algorithm is intuitive and practical. \n\n+ The paper considers an exhaustive space of both non-free disposal and free disposal settings which are both plently available in practical settings such as ads and recommender systems.\n\n\nWeakness\n\n- The primary weakness I find is in experiments. In particular, the studied problem is actually theoretically solvable using worst-case online algorithms when the inputs are stochastic. I wonder if the authors can compare against these optimal algorithms in the experimental section. In particular, when the inputs are i.i.d., these worst case algortihms can achieve competitive ratio > 0.7 in the worst-case and sometimes much higher on a random instance. Given that in the experiments, the competitive ratio obtained is of the order of 0.7-0.8, I think this comparison would make the paper really strong.\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity\n\n+ The paper is clear and well-written. The literature survey and placement is done fairly. \n\n\nQuality\n\n+ The considered research is high-quality and solving an important problem. Overall the direction is significant.\n\nNovelty\n\n+ The considered robust RL algorithm for Online matching is very novel. These combine two different areas of research that don't often intersect. The paper is novel in combining disparate ideas together.\n\nReproduciblity\n\n+ As far as I can tell, the paper contains sufficient details to reproduce the experiments", "summary_of_the_review": "Overall, I am on the border since I would like the empirical section to be strengthened. In particular, the paper introduces complex ideas and would like to see when this would significantly perform better than hand-tuned optimal online algorithms and when it would perform worse. This kind of extensive analysis will be useful to future research as well as a practitioner who would like to try these ideas in the wild. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666757207276}, {"id": "niHP5CamCg", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1962/Reviewer_7xYF"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a reinforcement-learning approach (that the authors call LOMAR) for edge-weighted online bipartite matching which is claimed to have good average but also worst case performance. The approach can handle both the no-free-disposal and the free-disposal setting with the second one allowing for dropping some edges from the matching to include some other ones. The main body of the paper concentrates and no-free-disposal and free disposal is deferred to the appendix.\n\nThe presented algorithm (LOMAR) is quite technical and the main novelty is involving a \"robustness constraints\" which helps decide on whether to follow the expert or the RL decision in each step. The algorithm is then analysed in terms of competitive ratio against an expert online algorithm. Finally, the RL training is incorporated within LOMAR by considering a switching operation based on the aforementioned constraint, and experiments are presented.", "review_text": "Overall this is a reasonably good paper, but I feel that more work needs to be done with respect to placing the results within the context of the literature.", "strengths": "Strengths\n- Important problem that is at the core of many real-life problems.\n- Generally well written.\n- The combination of RL and robustification is quite interesting the way it is done here\n\nWeaknesses\n- It is not clear how LOMAR compares to other robust algorithms for the problem stemming from the \"learning augmented algorithms\" literature. Although these algorithms are robust as well, it seems conceivable that LOMAR has an advantage by incorporating the robustification into the learning. It would be nice to see at least experimental comparison to these results and maybe a more extensive discussion on differences and similarities (some of these papers are already cited).\n- The literature review and some citations are quite off and give the impression that the authors just searched for some keywords and wrote a couple of sentences on some papers that seemed relevant. As an example, Boyar et al., 2016 is used (i) to support that worst case performance is not necessarily good average-case performance (page 1), (ii) as an example for ML-augmented algorithms (page 2) and (iii) to support the definition of \"strict\" competitive ratio. This paper is a survey on \"advice complexity\" and although it is not wrong to cite it in the context of this paper, advice complexity (in general) concerns itself with receiving *accurate* advice and investigating what the least possible amount (and appropriate encoding) of this advice is in order to guarantee optimality. This comes in contrast to ML-augmented algorithms where the advice may be suboptimal but still needs to be incorporated in order to provide improved results in case it is adequate. Also for the definition of competitive ratio I would go to the source and cite one of the papers that introduced the concept, or perhaps a book on online algorithms. Also, I am not familiar with the Gupta and Roughgarden paper but I find it strange that it is cited as an ML-augmented algorithms paper -- since it predates the area.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper presents a reinforcement-learning approach (that the authors call LOMAR) for edge-weighted online bipartite matching which is claimed to have good average but also worst case performance. The approach can handle both the no-free-disposal and the free-disposal setting with the second one allowing for dropping some edges from the matching to include some other ones. The main body of the paper concentrates and no-free-disposal and free disposal is deferred to the appendix.\n\nThe presented algorithm (LOMAR) is quite technical and the main novelty is involving a \"robustness constraints\" which helps decide on whether to follow the expert or the RL decision in each step. The algorithm is then analysed in terms of competitive ratio against an expert online algorithm. Finally, the RL training is incorporated within LOMAR by considering a switching operation based on the aforementioned constraint, and experiments are presented.", "strength_and_weaknesses": "Strengths\n- Important problem that is at the core of many real-life problems.\n- Generally well written.\n- The combination of RL and robustification is quite interesting the way it is done here\n\nWeaknesses\n- It is not clear how LOMAR compares to other robust algorithms for the problem stemming from the \"learning augmented algorithms\" literature. Although these algorithms are robust as well, it seems conceivable that LOMAR has an advantage by incorporating the robustification into the learning. It would be nice to see at least experimental comparison to these results and maybe a more extensive discussion on differences and similarities (some of these papers are already cited).\n- The literature review and some citations are quite off and give the impression that the authors just searched for some keywords and wrote a couple of sentences on some papers that seemed relevant. As an example, Boyar et al., 2016 is used (i) to support that worst case performance is not necessarily good average-case performance (page 1), (ii) as an example for ML-augmented algorithms (page 2) and (iii) to support the definition of \"strict\" competitive ratio. This paper is a survey on \"advice complexity\" and although it is not wrong to cite it in the context of this paper, advice complexity (in general) concerns itself with receiving *accurate* advice and investigating what the least possible amount (and appropriate encoding) of this advice is in order to guarantee optimality. This comes in contrast to ML-augmented algorithms where the advice may be suboptimal but still needs to be incorporated in order to provide improved results in case it is adequate. Also for the definition of competitive ratio I would go to the source and cite one of the papers that introduced the concept, or perhaps a book on online algorithms. Also, I am not familiar with the Gupta and Roughgarden paper but I find it strange that it is cited as an ML-augmented algorithms paper -- since it predates the area.", "clarity,_quality,_novelty_and_reproducibility": "The paper is of good quality, clarity and as far as I can say also novel.", "summary_of_the_review": "Overall this is a reasonably good paper, but I feel that more work needs to be done with respect to placing the results within the context of the literature.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethics concerns.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666627925066}], "openreview_url": "https://openreview.net/forum?id=cwiFbXPW4G0", "arxiv_id": "2306.00172", "paper_pdf": "papers/cwiFbXPW4G0.pdf", "paper_pdf_sha256": "c22fb89087ec769907de6de169c2f567782c149838560417b3eec7e181246aea", "paper_pdf_bytes": 479149, "paper_pdf_source": "openreview", "code_url": "https://github.com/Ren-Research/LOMAR", "code_repository": "Ren-Research/LOMAR", "code_commit": "7710662a31b55fdafe1c542da1842d8ed9f2e7a9", "code_archive": "repos/cwiFbXPW4G0.zip", "code_archive_sha256": "705514101b0a81274b01b361c9136fb3914c41542aeb17c326c845caa01a2874", "code_archive_bytes": 704788, "code_file_count": 40, "code_extensions": {".py": 40}, "github_disk_usage_kb": 682, "github_languages": {"Python": 323617}, "github_archived": false, "github_pushed_at": "2023-08-09T02:45:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-for-edge-weighted-online-bipartite"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8QE3pwEVc8P", "year": 2022, "status": "rejected", "title": "Zero-Cost Operation Scoring in Differentiable Architecture Search", "authors": ["Lichuan Xiang", "Łukasz Dudziak", "Mohamed S Abdelfattah", "Thomas Chun Pong Chau", "Nicholas Donald Lane", "Hongkai Wen"], "authorids": ["~Lichuan_Xiang1", "~Łukasz_Dudziak1", "~Mohamed_S_Abdelfattah1", "~Thomas_Chun_Pong_Chau1", "~Nicholas_Donald_Lane1", "~Hongkai_Wen1"], "authors_source": "OpenReview API", "abstract": "Differentiable neural architecture search (NAS) has attracted significant attention in recent years due to its ability to quickly discover promising architectures of deep neural networks even in very large search spaces. Despite its success, many differentiable NAS methods lack robustness and may degenerate to trivial architectures with excessive parameter-free operations such as skip connections thus leading to inferior performance. In fact, selecting operations based on the magnitude of architectural parameters was recently proven to be fundamentally wrong, showcasing the need to rethink how operation scoring and selection occurs in differentiable NAS. To this end, we formalize and analyze a fundamental component of differentiable NAS: local \"operation scoring\" that occurs at each choice of operation.\nWhen comparing existing operation scoring functions, we find that existing methods can be viewed as inexact proxies for accuracy.\nWe also find that existing methods perform poorly when analyzed empirically on NAS benchmarks. From this perspective, we introduce new training-free proxies to the context of differentiable NAS, and show that we can significantly speed up the search process while improving accuracy on multiple search spaces. We take inspiration from zero-cost proxies that were recently studied in the context of sample-based NAS but shown to degrade significantly for larger search spaces like DARTS. Our novel \"perturbation-based zero-cost operation scoring\" (Zero-Cost-PT) improves searching time and accuracy compared to the best available differentiable architecture search for many search space sizes, including very large ones. Specifically, we are able improve accuracy compared to the best current method (DARTS-PT) on the DARTS CNN search space while being over 40x faster (total searching time 25 minutes on a single GPU). Our code is available at: https://github.com/avail-upon-acceptance.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ahf0brbPsC", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2943/Reviewer_podk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Differentiable neural architecture search (NAS), and more recently, perturbation-based operation selection in differentiable NAS, is a popular recent area of study in NAS. In parallel, zero-cost proxies are also gaining in popularity in NAS. In this paper, the authors combine the insights from perturbation-based operation selection and zero-cost proxies, to create a new method that sees substantial time speedups. Specifically, they formalize \"operation scoring\", and use zero-cost proxies to score operations with the perturbation paradigm. This leads them to a new NAS method (Zero-Cost-PT) that outperforms existing methods. They evaluate their proposed algorithm on the DARTS search space (and its subsets) and NAS-Bench-201, showing that it can achieve up to 40x speedups.", "review_text": "## Strengths\n- This is a great combination of two popular areas: zero-cost proxies, and differentiable NAS via perturbation-based operation scoring. It is a natural idea to combine these two things, but also not trivial to do it well.\n- These ideas can have a high impact.\n- It seems that they give new insights for the celebrated “Rethinking Architecture Selection in Differentiable NAS” paper, including more proof of why DARTS-PT does well, and a refutation of the claim that disc-acc does well.\n- The appendix is pretty good and has a lot of information for reproducibility (but for code, see weaknesses).\n- Sections 3.1 and 3.2 are nicely done and make sense, and Fig 2 in particular is very interesting.\n\n## Weaknesses\n- Over-reliance on NAS-Bench-201. Almost all of Sections 3 and 4 focus on NAS-Bench-201. But this is the smallest search space, and in particular, it is known that zero-cost proxies do well on NAS-Bench-201 and bad on larger search spaces (the authors mention this as well). Although the authors also test their full algorithm on the DARTS search space in Section 5, it would strengthen the paper to do the analysis of Sections 3 and 4 on at least one other search space. NAS-Bench-1shot1 is the best option because it is size 360k and can run one-shot algorithms. TransNAS-Bench-101 can also run one-shot, and it is smaller but would add diversity. Finally, it might be possible to run some of the experiments using the 60k trained architectures from NAS-Bench-301).\n- Better ablation between the search and the validation stage. Sections 3 and 4 motivate the search part of the algorithm, but adding in the validation stage makes it less clear what leads to strong performance, especially on DARTS where the search step was never run in isolation. I think it makes it less apples-to-apples comparisons, e.g. because DARTS, DARTS-PT etc do not have the final validation stage. The authors did have an ablation, but it is only on NAS-Bench-201 and does not cover search only (\"validation only\" is basically NASWOT). In order to have a better comparison between ZC-PT and DARTS-PT, the authors could run ZC-PT once with no validation stage. Or, run ZC-PT and DARTS-PT both V times and do the validation stage for both of them. For Tables 2, 4, 5. As a result, I think the claim in the intro that \"Our novel Zero-Cost-PT improves searching time and accuracy compared to the best available differentiable architecture search for many search space sizes, including very large ones.\" is misleading. \n- The authors claim in Section 7 that they released their code in the supplementary material, but unfortunately there was no supplementary material submitted. If the authors provide the code during the rebuttal period, I will feel more positive about the paper (for example, anonymous github).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Differentiable neural architecture search (NAS), and more recently, perturbation-based operation selection in differentiable NAS, is a popular recent area of study in NAS. In parallel, zero-cost proxies are also gaining in popularity in NAS. In this paper, the authors combine the insights from perturbation-based operation selection and zero-cost proxies, to create a new method that sees substantial time speedups. Specifically, they formalize \"operation scoring\", and use zero-cost proxies to score operations with the perturbation paradigm. This leads them to a new NAS method (Zero-Cost-PT) that outperforms existing methods. They evaluate their proposed algorithm on the DARTS search space (and its subsets) and NAS-Bench-201, showing that it can achieve up to 40x speedups.", "main_review": "## Strengths\n- This is a great combination of two popular areas: zero-cost proxies, and differentiable NAS via perturbation-based operation scoring. It is a natural idea to combine these two things, but also not trivial to do it well.\n- These ideas can have a high impact.\n- It seems that they give new insights for the celebrated “Rethinking Architecture Selection in Differentiable NAS” paper, including more proof of why DARTS-PT does well, and a refutation of the claim that disc-acc does well.\n- The appendix is pretty good and has a lot of information for reproducibility (but for code, see weaknesses).\n- Sections 3.1 and 3.2 are nicely done and make sense, and Fig 2 in particular is very interesting.\n\n## Weaknesses\n- Over-reliance on NAS-Bench-201. Almost all of Sections 3 and 4 focus on NAS-Bench-201. But this is the smallest search space, and in particular, it is known that zero-cost proxies do well on NAS-Bench-201 and bad on larger search spaces (the authors mention this as well). Although the authors also test their full algorithm on the DARTS search space in Section 5, it would strengthen the paper to do the analysis of Sections 3 and 4 on at least one other search space. NAS-Bench-1shot1 is the best option because it is size 360k and can run one-shot algorithms. TransNAS-Bench-101 can also run one-shot, and it is smaller but would add diversity. Finally, it might be possible to run some of the experiments using the 60k trained architectures from NAS-Bench-301).\n- Better ablation between the search and the validation stage. Sections 3 and 4 motivate the search part of the algorithm, but adding in the validation stage makes it less clear what leads to strong performance, especially on DARTS where the search step was never run in isolation. I think it makes it less apples-to-apples comparisons, e.g. because DARTS, DARTS-PT etc do not have the final validation stage. The authors did have an ablation, but it is only on NAS-Bench-201 and does not cover search only (\"validation only\" is basically NASWOT). In order to have a better comparison between ZC-PT and DARTS-PT, the authors could run ZC-PT once with no validation stage. Or, run ZC-PT and DARTS-PT both V times and do the validation stage for both of them. For Tables 2, 4, 5. As a result, I think the claim in the intro that \"Our novel Zero-Cost-PT improves searching time and accuracy compared to the best available differentiable architecture search for many search space sizes, including very large ones.\" is misleading. \n- The authors claim in Section 7 that they released their code in the supplementary material, but unfortunately there was no supplementary material submitted. If the authors provide the code during the rebuttal period, I will feel more positive about the paper (for example, anonymous github).\n", "summary_of_the_review": "This paper combines two popular techniques in a nice way. I think the ideas in this paper can be pretty impactful. I do have three weaknesses about reliance on NAS-Bench-201, a fairer experimental setup, and code release, so I will give a weak accept.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None.", "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635896218574}, {"id": "HAI0vJnx_hM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2943/Reviewer_G1Xa"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors introduce new training-free proxies to the context of differentiable NAS, which can speed up the search process while improving accuracy. Further, the authors propose, evaluate and compare perturbation-based zero-cost operation scoring (Zero-Cost-PT) for differentiable NAS. Extensive experiments empirically show the effectiveness of Zero-Cost-PT in six search spaces and 3 datasets. ", "review_text": "Pros:\nThis paper introduces the lightweight operation scoring based on zero-cost proxies, which empirically outperform existing operation scoring functions. The experiments show that the perturbation is more effective than discretization and propose Zero-Cost-PT. Extensive experiments empirically show that the proposed method outperforms the best available differentiable architecture search in terms of searching time and accuracy. \n\nCons:\nThe zero-cost proxies method has been used in the previous work. This work only borrows it into differentiable architecture search, which is only a combination. There exist some typos and it is a little difficult to understand Figure 1. Some annotations can be moved to the caption. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors introduce new training-free proxies to the context of differentiable NAS, which can speed up the search process while improving accuracy. Further, the authors propose, evaluate and compare perturbation-based zero-cost operation scoring (Zero-Cost-PT) for differentiable NAS. Extensive experiments empirically show the effectiveness of Zero-Cost-PT in six search spaces and 3 datasets. ", "main_review": "Pros:\nThis paper introduces the lightweight operation scoring based on zero-cost proxies, which empirically outperform existing operation scoring functions. The experiments show that the perturbation is more effective than discretization and propose Zero-Cost-PT. Extensive experiments empirically show that the proposed method outperforms the best available differentiable architecture search in terms of searching time and accuracy. \n\nCons:\nThe zero-cost proxies method has been used in the previous work. This work only borrows it into differentiable architecture search, which is only a combination. There exist some typos and it is a little difficult to understand Figure 1. Some annotations can be moved to the caption. \n", "summary_of_the_review": "Generally speaking, I think it is incremental work. Although there is a lack of theoretical contributions, a large number of experiments provide empirical contributions. It is difficult for me to put forward some new opinions about these large numbers of experiments. I give the above acceptance threshold, but reject is OK.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635860152530}, {"id": "L-z75e1BmW", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2943/Reviewer_q4ge"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes to use the zero-cost proxy in the operation selection in differentiable neural architecture search. Specifically, the paper formalizes the selection procedure into perturbation and discretization. By introducing the zero-cost proxy, they first find the proxy could yield a better ranking compared to both Darts and Darts-PT in the Nas-Bench 201. They also experiment with different proxies proposed by previous work and find they all perform better than the Darts-PT. In the Darts-CNN and S1-S4 experiments, they select one of the best-performed proxies and find it could yield better accuracy than other baselines.", "review_text": "Pros:\n1. The paper is well-written and easy to follow. However, analyze part is slightly too long however the paper is in general in good shape.\n2. The experiment results look promising and interesting. It could achieve better accuracy across several benchmarks and some spaces, which could be useful to the community.\n\nCons:\n1. The paper's novelty is quite limited. It simply combines the differentiable neural architecture search with the one-shot one. Although it has been shown and addressed across the paper the zero-cost proxies are quite useful, however, it lacks insight on how it is correlated with the final performance ranking. \n2. At the same time, all the zero-cost proxies are proposed by other works and there is limited understanding of why some specific proxies are clearly better than others in the differentiable setting but not in the one-shot setting. \n3. Although the paper formalizes the problem into perturbation and discretion, I find only discretization is used in the final experiments, where the formalized framework seems not that important.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to use the zero-cost proxy in the operation selection in differentiable neural architecture search. Specifically, the paper formalizes the selection procedure into perturbation and discretization. By introducing the zero-cost proxy, they first find the proxy could yield a better ranking compared to both Darts and Darts-PT in the Nas-Bench 201. They also experiment with different proxies proposed by previous work and find they all perform better than the Darts-PT. In the Darts-CNN and S1-S4 experiments, they select one of the best-performed proxies and find it could yield better accuracy than other baselines.", "main_review": "Pros:\n1. The paper is well-written and easy to follow. However, analyze part is slightly too long however the paper is in general in good shape.\n2. The experiment results look promising and interesting. It could achieve better accuracy across several benchmarks and some spaces, which could be useful to the community.\n\nCons:\n1. The paper's novelty is quite limited. It simply combines the differentiable neural architecture search with the one-shot one. Although it has been shown and addressed across the paper the zero-cost proxies are quite useful, however, it lacks insight on how it is correlated with the final performance ranking. \n2. At the same time, all the zero-cost proxies are proposed by other works and there is limited understanding of why some specific proxies are clearly better than others in the differentiable setting but not in the one-shot setting. \n3. Although the paper formalizes the problem into perturbation and discretion, I find only discretization is used in the final experiments, where the formalized framework seems not that important.", "summary_of_the_review": "Although the paper has shown us the zero-cost proxy has great power in operation search in the differentiable neural architecture search, the paper stills lack insight on analyzing how it works. Therefore, I rate it as a borderline paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635599416996}, {"id": "68Wq3y_vxPe", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2943/Reviewer_86cZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work combines training-free proxies and perturbation-based subnetwork evaluation to achieve efficient NAS. The author also conducted comprehensive study of training-free proxies under different training iterations.", "review_text": "Strength:\nThis work studied different training-free proxies and leveraged perturbation-based methods for NAS.\n\nWeakness:\nMy main concern is the novelty of this work. Specifically,\n1. This work did not propose any new proxies for evaluating networks performance. If I understand correctly, the proxies listed in Table 2 are all from previous works. The authors did not claim any novelty in these training-free proxies either. Although the authors mentioned \"opportunity to use a new class of training-free proxies\" in Sec. 3.1, there is no concrete proposal in this work.\n2. The operation sorting method (Eq. 5) is also similar to the perturbation method proposed by [1]. The main difference is that the authors adopt this perturbation at supernet's initialization.\n\n[1] Rethinking architecture selection in differentiable NAS.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work combines training-free proxies and perturbation-based subnetwork evaluation to achieve efficient NAS. The author also conducted comprehensive study of training-free proxies under different training iterations.", "main_review": "Strength:\nThis work studied different training-free proxies and leveraged perturbation-based methods for NAS.\n\nWeakness:\nMy main concern is the novelty of this work. Specifically,\n1. This work did not propose any new proxies for evaluating networks performance. If I understand correctly, the proxies listed in Table 2 are all from previous works. The authors did not claim any novelty in these training-free proxies either. Although the authors mentioned \"opportunity to use a new class of training-free proxies\" in Sec. 3.1, there is no concrete proposal in this work.\n2. The operation sorting method (Eq. 5) is also similar to the perturbation method proposed by [1]. The main difference is that the authors adopt this perturbation at supernet's initialization.\n\n[1] Rethinking architecture selection in differentiable NAS.", "summary_of_the_review": "In addition to the novelty issue mentioned above, some results and description of this work also makes me a bit confusing. For example, when demonstrating Figure 2, the author did not mention which concrete training-free proxy they used for $S$ in Eq. 5. Also, the purpose of \"retrain for 5 epochs\" is unclear to me. Why does the author want to conduct this experiment given the \"zero-cost\" motivation in this work?", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634426533066}], "openreview_url": "https://openreview.net/forum?id=8QE3pwEVc8P", "arxiv_id": "2106.06799", "paper_pdf": "papers/8QE3pwEVc8P.pdf", "paper_pdf_sha256": "aa5dd9bcb28ebb7ec261ff973d60e9a01ecf2012c7287464d33f77e5417e77d0", "paper_pdf_bytes": 1426415, "paper_pdf_source": "openreview", "code_url": "https://github.com/visionbasicagent/zerocost_operation_score", "code_repository": "visionbasicagent/zerocost_operation_score", "code_commit": "55fc52b29a8d1be086937a7e8045d155ec0cf63f", "code_archive": "repos/8QE3pwEVc8P.zip", "code_archive_sha256": "16e9163ccc8694ced2039ff210c27c2528c06074bd659e816b9498c1f2c8480c", "code_archive_bytes": 4029109, "code_file_count": 230, "code_extensions": {".py": 123, ".sh": 94, ".ipynb": 13}, "github_disk_usage_kb": 3010, "github_languages": {"Jupyter Notebook": 15717226, "Python": 877775, "Shell": 166619}, "github_archived": false, "github_pushed_at": "2022-12-01T19:51:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/zero-cost-proxies-meet-differentiable"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fycxGdpCCmW", "year": 2021, "status": "rejected", "title": "Hybrid Discriminative-Generative Training via Contrastive Learning", "authors": ["Hao Liu", "Pieter Abbeel"], "authorids": ["~Hao_Liu1", "~Pieter_Abbeel2"], "authors_source": "OpenReview API", "abstract": "Contrastive learning and supervised learning have both seen significant progress and success. However, thus far they have largely been treated as two separate objectives, brought together only by having a shared neural network. In this paper we show that through the perspective of hybrid discriminative-generative training of energy-based models we can make a direct connection between contrastive learning and supervised learning. Beyond presenting this unified view, we show our specific choice of approximation of the energy-based loss significantly improves energy-based models and contrastive learning based methods in confidence-calibration, out-of-distribution detection, adversarial robustness, generative modeling, and image classification tasks. In addition to significantly improved performance, our method also gets rid of SGLD training and does not suffer from training instability. Our evaluations also demonstrate that our method performs better than or on par with state-of-the-art hand-tailored methods in each task. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "OD5vdPJauN_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3002/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n- Paper proposes Hybrid Discriminative Generative training of Energy based models (HDGE) which combines supervised and generative modeling by using a contrastive approximation of the energy based loss\n- Approach shows this is better than baselines on various tasks like confidence calibration, OOD detection, robustness and classification accuracy\n\n\nClarity\n- Overall well written paper. Figures and tables are informative and supplement the flow.\n- Formatting error in figure 4 in appendix\n\n\nNovelty\n- Paper proposes a simple but unified view of contrastive training, generative and discriminative modeling - a nice, novel contribution with empirically strong results\n- Gets rid of computationally expensive SGLD by using contrastive approximation which was a key limitation of prior energy based modeling work like JEM\n\n\nSignificance\n- Results are compelling across a wide range of tasks over existing (EBM) baselines including calibration, robustness, OOD detection, generative modeling and classification accuracy\n\n\nQuestions/clarifications/comments\n- “, Grathwohl et al. (2019) show the alternative class-conditional EBM p(y|x) leads to significant improvement in generative modeling while retain compelling classification accuracy” -> Not sure the JEM model is class conditional\n\n- How alpha = 0.5 (the weighting chosen)? The details are not presented.\n\n- Error bars are missing for the classification accuracy experiments in Table 1 which makes it hard to verify improvements especially wrt supervised contrastive loss method\n\n- Detail on how classification accuracy is computed when using generative term in HDGE is missing? Is it a linear classifier on top of learned representations?\n\n- “Prior work show that fitting a density model on the data and consider examples with low likelihood to be OOD is effective” -> Not completely true see https://arxiv.org/abs/1810.09136\n\n- Please share exact details on how p(x) for OOD score is calculated\n- Error bars again missing in Table 2\n\n- “We find HDGE performs beyond the performance of a strong baseline classifier” - this is a strong statement as only for CIFAR10/Celeb A the gains of HDGE are clear\n\n- Why was the Winkens et al, 2020 contrastive baseline not used here to compare in Table 2 - https://arxiv.org/abs/2007.05566?\n- “HDGE is conceptual simple to implement, scalable, and powerful.” -> conceptually. Also scalability is a somewhat strong claim as the main datasets used here are CIFAR variants.\n\n- Was HDGE + JEM experiments also performed for OOD detection?\n\n- Legend in figure 4 should be “HDGE” not “HDSE”?\n\nOverall good effort with seemingly good improvements over prior efforts on hybrid EBMs over a number of tasks. Main concern is lack of error bars which makes it hard to validate claims in certain cases.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "a good submission empirically improving on top of prior hybrid EBM works", "review": "Summary\n- Paper proposes Hybrid Discriminative Generative training of Energy based models (HDGE) which combines supervised and generative modeling by using a contrastive approximation of the energy based loss\n- Approach shows this is better than baselines on various tasks like confidence calibration, OOD detection, robustness and classification accuracy\n\n\nClarity\n- Overall well written paper. Figures and tables are informative and supplement the flow.\n- Formatting error in figure 4 in appendix\n\n\nNovelty\n- Paper proposes a simple but unified view of contrastive training, generative and discriminative modeling - a nice, novel contribution with empirically strong results\n- Gets rid of computationally expensive SGLD by using contrastive approximation which was a key limitation of prior energy based modeling work like JEM\n\n\nSignificance\n- Results are compelling across a wide range of tasks over existing (EBM) baselines including calibration, robustness, OOD detection, generative modeling and classification accuracy\n\n\nQuestions/clarifications/comments\n- “, Grathwohl et al. (2019) show the alternative class-conditional EBM p(y|x) leads to significant improvement in generative modeling while retain compelling classification accuracy” -> Not sure the JEM model is class conditional\n\n- How alpha = 0.5 (the weighting chosen)? The details are not presented.\n\n- Error bars are missing for the classification accuracy experiments in Table 1 which makes it hard to verify improvements especially wrt supervised contrastive loss method\n\n- Detail on how classification accuracy is computed when using generative term in HDGE is missing? Is it a linear classifier on top of learned representations?\n\n- “Prior work show that fitting a density model on the data and consider examples with low likelihood to be OOD is effective” -> Not completely true see https://arxiv.org/abs/1810.09136\n\n- Please share exact details on how p(x) for OOD score is calculated\n- Error bars again missing in Table 2\n\n- “We find HDGE performs beyond the performance of a strong baseline classifier” - this is a strong statement as only for CIFAR10/Celeb A the gains of HDGE are clear\n\n- Why was the Winkens et al, 2020 contrastive baseline not used here to compare in Table 2 - https://arxiv.org/abs/2007.05566?\n- “HDGE is conceptual simple to implement, scalable, and powerful.” -> conceptually. Also scalability is a somewhat strong claim as the main datasets used here are CIFAR variants.\n\n- Was HDGE + JEM experiments also performed for OOD detection?\n\n- Legend in figure 4 should be “HDGE” not “HDSE”?\n\nOverall good effort with seemingly good improvements over prior efforts on hybrid EBMs over a number of tasks. Main concern is lack of error bars which makes it hard to validate claims in certain cases.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604361269934}, {"id": "5NwFaKQ5y-Y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3002/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes HDGE - a simple method to improve over JEM. JEM is optimized using a combination of two terms:\n$\\log p(y|x) + \\log p(x)$\n\nThe first term is optimized using the standard cross-entropy loss, while the second term is optimized using SGLD. Running SGLD chains in each iteration can cause instability. In HDGE, instead of optimizing $\\log p(x)$, an approximation to the conditional density $\\log p(x|y)$ is optimized. The idea is to approximate the normalization constant $Z(\\theta)$ with an empirical averaging of energy functions over a large memory bank. This yields a simple objective to optimize. The benefit of using such an approximation is that this eliminates the need for running SGLD, thereby improving the stability of training.\n\nThe idea itself is simple and intuitive. Experiments show that HDGE consistently outperform / perform on-par with JEM on image classification, OOD detection and calibration.\n\nShould we call this a contrastive objective? Im not super convinced if the objective of $\\log p(x|y)$ can be called a contrastive objective. Because in contrastive losses, we always focus on pairs of samples, i.e., we contrast the representation of one sample to another, while the objective in this paper takes the form similar to cross-entropy loss instead. Should the loss be called something instead?\n\nI would like the authors to have a discussion on the training stability of HDGE compared to JEM. It looks like HDGE would be more stable since we don't need to run SGLD, but this message should be made more clear as this is the most important improvement over JEM.\n\nThe performance improvement in table 1 is marginal. So, it is important to perform multiple runs and report mean and standard deviations to understand the statistical significance of the results.\n\nCan HDGE be used for generative modeling? i.e., how do you sample from p(x)? Experiments in appendix show that HDGE can be used in combination with JEM for generative modeling, but this again requires running SGLD. Can HDGE be used in isolation for generative modeling tasks?\n\nHow does the performance compare with other SOTA metods for OOD detection and calibation? Some comparisons that could be done include (but not limited to): Ren et al., \"Likelihood Ratios for Out-of-Distribution Detection\", Padhy et al., \"Revisiting One-vs-All Classifiers for Predictive Uncertainty and Out-of-Distribution Detection in Neural Networks\", Morningstar et al. \"Density of States Estimation for Out-of-Distribution Detection\", etc. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple yet effective method for improvement over JEM", "review": "The paper proposes HDGE - a simple method to improve over JEM. JEM is optimized using a combination of two terms:\n$\\log p(y|x) + \\log p(x)$\n\nThe first term is optimized using the standard cross-entropy loss, while the second term is optimized using SGLD. Running SGLD chains in each iteration can cause instability. In HDGE, instead of optimizing $\\log p(x)$, an approximation to the conditional density $\\log p(x|y)$ is optimized. The idea is to approximate the normalization constant $Z(\\theta)$ with an empirical averaging of energy functions over a large memory bank. This yields a simple objective to optimize. The benefit of using such an approximation is that this eliminates the need for running SGLD, thereby improving the stability of training.\n\nThe idea itself is simple and intuitive. Experiments show that HDGE consistently outperform / perform on-par with JEM on image classification, OOD detection and calibration.\n\nShould we call this a contrastive objective? Im not super convinced if the objective of $\\log p(x|y)$ can be called a contrastive objective. Because in contrastive losses, we always focus on pairs of samples, i.e., we contrast the representation of one sample to another, while the objective in this paper takes the form similar to cross-entropy loss instead. Should the loss be called something instead?\n\nI would like the authors to have a discussion on the training stability of HDGE compared to JEM. It looks like HDGE would be more stable since we don't need to run SGLD, but this message should be made more clear as this is the most important improvement over JEM.\n\nThe performance improvement in table 1 is marginal. So, it is important to perform multiple runs and report mean and standard deviations to understand the statistical significance of the results.\n\nCan HDGE be used for generative modeling? i.e., how do you sample from p(x)? Experiments in appendix show that HDGE can be used in combination with JEM for generative modeling, but this again requires running SGLD. Can HDGE be used in isolation for generative modeling tasks?\n\nHow does the performance compare with other SOTA metods for OOD detection and calibation? Some comparisons that could be done include (but not limited to): Ren et al., \"Likelihood Ratios for Out-of-Distribution Detection\", Padhy et al., \"Revisiting One-vs-All Classifiers for Predictive Uncertainty and Out-of-Distribution Detection in Neural Networks\", Morningstar et al. \"Density of States Estimation for Out-of-Distribution Detection\", etc. \n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603985642665}, {"id": "5EPTqE3Y8u_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3002/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1, Summary of contribution:\nThis paper gives a unified view of contrastive learning and supervised learning through energy-based model (EBM) training. Specifically, it claims that the contrastive loss can approximate the log-likelihood of the energy-based model, and shows that the proposed objective works on par with or better than its counterparts. \n\n2, Strengths and Weaknesses:\nThe proposed method of training energy-based models achieves good performance on various tasks (e.g. OOD) without SGLD which usually requires a lot of tricks to work well on high-dimensional data.  This is the strength of this paper, and the community might be able to leverage the idea like this to extend the application of EBM to other domains in the future. \nMeanwhile,  as the authors admit themselves,  the approximations are not thoroughly justified. Whether the research is appropriately bridging the gap between EBM and contrastive learning is questionable. This is the weakness of the paper. \n\n3, Recommendation:\nMarginally below the acceptance threshold. I believe that their claimed connection to EBM is overstated.\n\n4, Reasons for Recommendation:\nThere are several factors that make me question whether the model learned by their proposed method can be legitimately interpreted as a variant of EBM. To name a few, \n(1) As the authors write themselves, the approximation in eq (13) is crude. While the summands in the denominator are sampled from probability distributions (data distribution), when in fact it is supposed to be integral with respect to Lebesgue measure. \n(2) The algorithm seems to be introducing some mysterious trick that makes the model ignore the gradient with respect to negative samples and normalization factor. For some unstated reason, the algorithm is also normalizing the logits.\n\nThe sheer results look promising, and it seems that the method is succeeding in capturing some aspects of the data distribution. However, they do not necessarily justify the claimed connection between HDGE and EBM. It could be that HDGE is just succeeding in learning the support of the dataset.  \nIn order for this paper to reach the standard of ICLR, I believe that the authors must revise the strength of their work and redesign their experiments as such.  \n\nI believe that the paper can be improved if, within the scope of revision, the author can provide more solid justifications for approximations and tricks used in the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "1, Summary of contribution:\nThis paper gives a unified view of contrastive learning and supervised learning through energy-based model (EBM) training. Specifically, it claims that the contrastive loss can approximate the log-likelihood of the energy-based model, and shows that the proposed objective works on par with or better than its counterparts. \n\n2, Strengths and Weaknesses:\nThe proposed method of training energy-based models achieves good performance on various tasks (e.g. OOD) without SGLD which usually requires a lot of tricks to work well on high-dimensional data.  This is the strength of this paper, and the community might be able to leverage the idea like this to extend the application of EBM to other domains in the future. \nMeanwhile,  as the authors admit themselves,  the approximations are not thoroughly justified. Whether the research is appropriately bridging the gap between EBM and contrastive learning is questionable. This is the weakness of the paper. \n\n3, Recommendation:\nMarginally below the acceptance threshold. I believe that their claimed connection to EBM is overstated.\n\n4, Reasons for Recommendation:\nThere are several factors that make me question whether the model learned by their proposed method can be legitimately interpreted as a variant of EBM. To name a few, \n(1) As the authors write themselves, the approximation in eq (13) is crude. While the summands in the denominator are sampled from probability distributions (data distribution), when in fact it is supposed to be integral with respect to Lebesgue measure. \n(2) The algorithm seems to be introducing some mysterious trick that makes the model ignore the gradient with respect to negative samples and normalization factor. For some unstated reason, the algorithm is also normalizing the logits.\n\nThe sheer results look promising, and it seems that the method is succeeding in capturing some aspects of the data distribution. However, they do not necessarily justify the claimed connection between HDGE and EBM. It could be that HDGE is just succeeding in learning the support of the dataset.  \nIn order for this paper to reach the standard of ICLR, I believe that the authors must revise the strength of their work and redesign their experiments as such.  \n\nI believe that the paper can be improved if, within the scope of revision, the author can provide more solid justifications for approximations and tricks used in the paper.", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603891883593}, {"id": "MUfUK2TS8rU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3002/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "===============Update after rebuttal period================\nThe connection between the contrastive learning objective and discriminative learning is made via \"resemblance\". And the author claims the \"resemblance\" as a theoretical contribution, which the first reason I vote for a clear rejection. This issue has not been addressed by the authors. The second reason for my rejection of the paper is the paper requires an effort to make it self-contained, especially for the experimental section. I remain my score of clear rejection.\n\n=======================================================\nThis paper connected contrastive-learning and supervised learning from the perspective of the energy-based models. Then, the authors combine both objectives and evaluate the presented method on various datasets and tasks.\n\nStrengths: The paper attempts to connect supervised and contrastive learning. I like the attempt. But unfortunately, I don't think it is valid. See explanations as follows.\n\nWeakness:\n1. I feel the claim in the paper is too strong. The approximation from equation 12 to 13 is very crude. Specifically, the approximation states that the infinite integral (for the normalization constant) can be replaced by a finite sum, which is generally not true. \n2. Even if we assume the above approximation is fine, the connection with contrastive learning is very unclear. Precisely, the approximation is for modeling p(x|y), yet the contrastive learning is modeling p(x_1|x_2) with  x_1 and x_2 being the outcomes from correlated data. The authors do not discuss or compare between p(x|y) and p(x_1|x_2), and hence it makes the connection very vague. \n3. The resulting objective (eq. 15) is a combination of the discriminative and generative modeling, which has already been studied. \n4. On page 4, \"the representation captures important information between similar data points, and therefore might improve performance on downstream tasks.\" This sentence is super vague, and I can't understand what the \"important information\" is and why if we capture \"this important information\", we \"may improve performance on downstream tasks.\" The author should spend time polishing the presentation.\n5. The main complaint of the presentation is the overclaim for the experimental section. I understand the contents are too much, and hence the author must move some experimental sections into Appendix. The author claims that the proposed method is performed on adversarial modeling and generative modeling, while these two sections only appear in the Appendix. In the last few lines of page 5, the author seems to rush the remaining experimental sections into the Appendix and asks the reviewer/reader to read themselves. The author should spend time arranging the contents and make sure the paper is self-contained. \n\n==================================\nSummary of the reasons why I vote for rejection:\n1. The main contribution of the paper by connecting supervised learning and contrastive learning is overclaimed. The approximation of the intractable normalization term is not appropriate. The connection with contrastive learning is not solid.\n2. The paper doesn't seem to be ready for submission. The content is not organized well and some ambiguous wordings should be avoided.\n\n\n[1] Representation Learning with Contrastive Predictive Coding by Oord et al.\n[2] On Variational Bounds of Mutual Information by Poole et al.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "===============Update after rebuttal period================\nThe connection between the contrastive learning objective and discriminative learning is made via \"resemblance\". And the author claims the \"resemblance\" as a theoretical contribution, which the first reason I vote for a clear rejection. This issue has not been addressed by the authors. The second reason for my rejection of the paper is the paper requires an effort to make it self-contained, especially for the experimental section. I remain my score of clear rejection.\n\n=======================================================\nThis paper connected contrastive-learning and supervised learning from the perspective of the energy-based models. Then, the authors combine both objectives and evaluate the presented method on various datasets and tasks.\n\nStrengths: The paper attempts to connect supervised and contrastive learning. I like the attempt. But unfortunately, I don't think it is valid. See explanations as follows.\n\nWeakness:\n1. I feel the claim in the paper is too strong. The approximation from equation 12 to 13 is very crude. Specifically, the approximation states that the infinite integral (for the normalization constant) can be replaced by a finite sum, which is generally not true. \n2. Even if we assume the above approximation is fine, the connection with contrastive learning is very unclear. Precisely, the approximation is for modeling p(x|y), yet the contrastive learning is modeling p(x_1|x_2) with  x_1 and x_2 being the outcomes from correlated data. The authors do not discuss or compare between p(x|y) and p(x_1|x_2), and hence it makes the connection very vague. \n3. The resulting objective (eq. 15) is a combination of the discriminative and generative modeling, which has already been studied. \n4. On page 4, \"the representation captures important information between similar data points, and therefore might improve performance on downstream tasks.\" This sentence is super vague, and I can't understand what the \"important information\" is and why if we capture \"this important information\", we \"may improve performance on downstream tasks.\" The author should spend time polishing the presentation.\n5. The main complaint of the presentation is the overclaim for the experimental section. I understand the contents are too much, and hence the author must move some experimental sections into Appendix. The author claims that the proposed method is performed on adversarial modeling and generative modeling, while these two sections only appear in the Appendix. In the last few lines of page 5, the author seems to rush the remaining experimental sections into the Appendix and asks the reviewer/reader to read themselves. The author should spend time arranging the contents and make sure the paper is self-contained. \n\n==================================\nSummary of the reasons why I vote for rejection:\n1. The main contribution of the paper by connecting supervised learning and contrastive learning is overclaimed. The approximation of the intractable normalization term is not appropriate. The connection with contrastive learning is not solid.\n2. The paper doesn't seem to be ready for submission. The content is not organized well and some ambiguous wordings should be avoided.\n\n\n[1] Representation Learning with Contrastive Predictive Coding by Oord et al.\n[2] On Variational Bounds of Mutual Information by Poole et al.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603861846780}], "openreview_url": "https://openreview.net/forum?id=fycxGdpCCmW", "arxiv_id": "2007.09070", "paper_pdf": "papers/fycxGdpCCmW.pdf", "paper_pdf_sha256": "b1f7de6d844c0fe1fa671c3b1531b6904291739e485a7eeb3a86caf75fd844c0", "paper_pdf_bytes": 716991, "paper_pdf_source": "openreview", "code_url": "https://github.com/haoliuhl/hybrid-discriminative-generative", "code_repository": "haoliuhl/hybrid-discriminative-generative", "code_commit": "9f12b3e8f53dcc20bdc47359ffba1894748f1c8f", "code_archive": "repos/fycxGdpCCmW.zip", "code_archive_sha256": "bba232cd59a576a245bdac2300dd0e375e25980095a77ae97100f4aed9d28cf1", "code_archive_bytes": 1483099, "code_file_count": 16, "code_extensions": {".py": 11, ".js": 5}, "github_disk_usage_kb": 1648, "github_languages": {"Python": 153412, "JavaScript": 28156, "HTML": 23946, "CSS": 8270}, "github_archived": false, "github_pushed_at": "2023-05-01T20:42:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hybrid-discriminative-generative-training-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4eYSSSDle6", "year": 2026, "status": "rejected", "title": "PRL: Prompts from Reinforcement Learning", "authors": ["Paweł Batorski", "Adrian Kosmala", "Paul Swoboda"], "authorids": ["~Paweł_Batorski2", "~Adrian_Kosmala1", "~Paul_Swoboda1"], "authors_source": "OpenReview API", "abstract": "Effective prompt engineering remains a central challenge in fully harnessing the\ncapabilities of LLMs. While well-designed prompts can dramatically enhance\nperformance, crafting them typically demands expert intuition and a nuanced understanding of the task. Moreover, the most impactful prompts often hinge on\nsubtle semantic cues, ones that may elude human perception but are crucial for\nguiding LLM behavior. In this paper, we introduce PRL (Prompts from Reinforcement Learning), a novel RL-based approach for automatic prompt generation.\nUnlike previous methods, PRL can produce novel few-shot examples that were not\nseen during training. Our approach achieves state-of-the-art performance across a\nrange of benchmarks, including text classification, simplification, summarization,\nand reasoning. On the classification task, it surpasses prior methods by 2.58%\nover APE and 1.00% over EvoPrompt. Additionally, it improves the average\nROUGE scores on the summarization task by 4.32 over APE and by 2.12 over\nEvoPrompt and the SARI score on simplification by 6.93 over APE and by 6.01\nover EvoPrompt. On the GSM8K mathematical reasoning benchmark, PRL further\nimproves accuracy by 2.72% over APE and by 4.53% over EvoPrompt. We will\nmake our implementation publicly available upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Qh7iphTJje", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4095/Reviewer_tiGz"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces PRL (Prompts from Reinforcement Learning), a reinforcement learning-based\napproach to automatically generating and optimizing prompts for large language models (LLMs). PRL uniquely\nenables the synthesis of novel few-shot examples not seen during training and integrates explicit reasoning\nsteps before prompt output. This method achieves excellent empirical performance in tasks such as text\nclassification. The research scheme includes carefully designed reward shaping, prompt selection, and\ndetailed ablation experiments.", "review_text": "This paper introduces PRL (Prompts from Reinforcement Learning), a reinforcement learning-based\napproach to automatically generating and optimizing prompts for large language models (LLMs). PRL uniquely\nenables the synthesis of novel few-shot examples not seen during training and integrates explicit reasoning\nsteps before prompt output. This method achieves excellent empirical performance in tasks such as text\nclassification. The research scheme includes carefully designed reward shaping, prompt selection, and\ndetailed ablation experiments.", "strengths": "1. PRL devises a clear RL-based prompt optimization loop. Compared to other methods, PRL can create few-\nshot prompt examples not limited to the original training data\n\n2. PRL is evaluated across varied tasks, and multiple ablation studies dissect the contribution of prompt\nselection, few-shot examples, and explicit reasoning. Additionally, its effectiveness is verified on models of\ndifferent architectures and sizes.", "weaknesses": "1. Compared to other methods, PRL requires more computational resources for training and has insufficient\ngeneralization ability, which means that PRL needs to be retrained for different tasks, greatly limiting its\nusability.\n\n2. There is insufficient discussion on the instability, scalability, and generalization of reinforcement learning:\n- Can training on a single task generalize to other tasks? Furthermore, how does the performance and\ngeneralization of simultaneous multi-task learning compare to that of single-task learning?\n- Is model training sensitive to the introduced reward function, does reward manipulation exist, and\nwhat is the interaction between format specification rewards and task correctness rewards?\n- Is it effective to use a different generator model for training than the one used to evaluation model?\nFurthermore, can the same generator model be directly used for different evaluation models after\ntraining?\n\n3. One of the paper's cores is that few-shot prompting behavior \"spontaneously emerges\" from the RL setup,\nyet there is little to no formal analysis or justification. A more rigorous explanation (e.g., does the reward\nlandscape incentivize synthesis, or is it an artifact of the prompt generator's architecture?) is sorely missing.\n\n4. The compared methods are limited to those before 2023, lacking comparison results with new methods\nfrom 2024-2025.", "questions": "see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PRL (Prompts from Reinforcement Learning), a reinforcement learning-based\napproach to automatically generating and optimizing prompts for large language models (LLMs). PRL uniquely\nenables the synthesis of novel few-shot examples not seen during training and integrates explicit reasoning\nsteps before prompt output. This method achieves excellent empirical performance in tasks such as text\nclassification. The research scheme includes carefully designed reward shaping, prompt selection, and\ndetailed ablation experiments.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. PRL devises a clear RL-based prompt optimization loop. Compared to other methods, PRL can create few-\nshot prompt examples not limited to the original training data\n\n2. PRL is evaluated across varied tasks, and multiple ablation studies dissect the contribution of prompt\nselection, few-shot examples, and explicit reasoning. Additionally, its effectiveness is verified on models of\ndifferent architectures and sizes.", "weaknesses": "1. Compared to other methods, PRL requires more computational resources for training and has insufficient\ngeneralization ability, which means that PRL needs to be retrained for different tasks, greatly limiting its\nusability.\n\n2. There is insufficient discussion on the instability, scalability, and generalization of reinforcement learning:\n- Can training on a single task generalize to other tasks? Furthermore, how does the performance and\ngeneralization of simultaneous multi-task learning compare to that of single-task learning?\n- Is model training sensitive to the introduced reward function, does reward manipulation exist, and\nwhat is the interaction between format specification rewards and task correctness rewards?\n- Is it effective to use a different generator model for training than the one used to evaluation model?\nFurthermore, can the same generator model be directly used for different evaluation models after\ntraining?\n\n3. One of the paper's cores is that few-shot prompting behavior \"spontaneously emerges\" from the RL setup,\nyet there is little to no formal analysis or justification. A more rigorous explanation (e.g., does the reward\nlandscape incentivize synthesis, or is it an artifact of the prompt generator's architecture?) is sorely missing.\n\n4. The compared methods are limited to those before 2023, lacking comparison results with new methods\nfrom 2024-2025.", "questions": "see weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761966551636}, {"id": "jPX9FI034N", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4095/Reviewer_tMpA"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces PRL, a RL framework that trains a prompt-generation model to automatically produce effective task prompts. It can be viewed as another R1-style (GRPO RL training) approach applied to the domain of automatic prompt generation.", "review_text": "The paper introduces PRL, a RL framework that trains a prompt-generation model to automatically produce effective task prompts. It can be viewed as another R1-style (GRPO RL training) approach applied to the domain of automatic prompt generation.", "strengths": "1. Framing prompt optimization as a reinforcement learning problem with a frozen evaluator is clear and straightforward.\n2. The method demonstrates consistent empirical improvements across multiple benchmark tasks.", "weaknesses": "1. The paper motivates prompt optimization as an underexplored and crucial problem, but recent work (e.g., instruction tuning, preference alignment, RLHF) has substantially reduced the marginal importance of prompt optimization for strong LLMs. The paper would benefit from a deeper discussion or quantitative evidence that prompt optimization still provides meaningful gains for modern aligned models. \n2. Most baseline methods cited for automatic prompt generation date back to 2022–2023. It would strengthen the paper to incorporate or discuss more recent developments in this area.\n3. The proposed method requires RL training for each dataset, which introduces significant computational overhead compared to evolutionary or heuristic prompt optimizers. The paper lacks a systematic analysis of the resulting latency and compute cost versus accuracy trade-off, making it difficult to assess practical efficiency.\n4. The motivation for training a separate prompt generator model is somewhat unconvincing. The paper does not compare against stronger or larger generator models (e.g., Qwen2.5-72B, GPT) or expert-crafted prompts, which could potentially achieve comparable results without RL training.\n5. Since the generator must be retrained for every dataset, the method’s scalability and generality are limited. The paper does not explore whether a single generator can generalize across multiple datasets or related domains.\n6. The chosen benchmarks (e.g., GSM8K, SAMSum, SST, AGNews) are somewhat outdated and relatively simple. Evaluating on more challenging or recent datasets would better demonstrate PRL’s robustness and contemporary relevance.\n7. The ablation on model size lacks a clear causal interpretation. Performance improvements when scaling from 7B to 32B evaluators could largely stem from the inherent capability gain of the larger models rather than from PRL’s prompt optimization.", "questions": "Could you include detailed statistics for the training, validation, and test datasets used in your experiments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces PRL, a RL framework that trains a prompt-generation model to automatically produce effective task prompts. It can be viewed as another R1-style (GRPO RL training) approach applied to the domain of automatic prompt generation.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Framing prompt optimization as a reinforcement learning problem with a frozen evaluator is clear and straightforward.\n2. The method demonstrates consistent empirical improvements across multiple benchmark tasks.", "weaknesses": "1. The paper motivates prompt optimization as an underexplored and crucial problem, but recent work (e.g., instruction tuning, preference alignment, RLHF) has substantially reduced the marginal importance of prompt optimization for strong LLMs. The paper would benefit from a deeper discussion or quantitative evidence that prompt optimization still provides meaningful gains for modern aligned models. \n2. Most baseline methods cited for automatic prompt generation date back to 2022–2023. It would strengthen the paper to incorporate or discuss more recent developments in this area.\n3. The proposed method requires RL training for each dataset, which introduces significant computational overhead compared to evolutionary or heuristic prompt optimizers. The paper lacks a systematic analysis of the resulting latency and compute cost versus accuracy trade-off, making it difficult to assess practical efficiency.\n4. The motivation for training a separate prompt generator model is somewhat unconvincing. The paper does not compare against stronger or larger generator models (e.g., Qwen2.5-72B, GPT) or expert-crafted prompts, which could potentially achieve comparable results without RL training.\n5. Since the generator must be retrained for every dataset, the method’s scalability and generality are limited. The paper does not explore whether a single generator can generalize across multiple datasets or related domains.\n6. The chosen benchmarks (e.g., GSM8K, SAMSum, SST, AGNews) are somewhat outdated and relatively simple. Evaluating on more challenging or recent datasets would better demonstrate PRL’s robustness and contemporary relevance.\n7. The ablation on model size lacks a clear causal interpretation. Performance improvements when scaling from 7B to 32B evaluators could largely stem from the inherent capability gain of the larger models rather than from PRL’s prompt optimization.", "questions": "Could you include detailed statistics for the training, validation, and test datasets used in your experiments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761588537698}, {"id": "IggJfC1GrK", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4095/Reviewer_vKbp"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "Effective prompt engineering remains a **central challenge** in fully harnessing the capabilities of Large Language Models (LLMs). While precise input prompts can guide LLMs to perform complex tasks, the most impactful prompts often rely on *subtle semantic cues* that may *elude human perception*.\n\nTo address this, the paper introduces **PRL (Prompts from Reinforcement Learning)**, a novel RL-based approach for automatic prompt generation.\n\n**PRL's key innovations include:**\n\n1.  **Novel Few-shot Example Synthesis:** Unlike previous methods like Automatic Prompt Optimization (APO), which are restricted to selecting few-shot examples from training data, PRL is capable of generating and selecting **novel few-shot examples** that were *not seen during training*.\n2.  **Explicit Reasoning Integration:** PRL incorporates a *reasoning phase* prior to prompt generation, where the prompt generator first produces a **rationale** (enclosed within `<think>` tags) to guide the final output.\n3.  **RL Optimization Cycle:** PRL trains a **Prompt Generator ($\\pi_{\\text{generator}}$)** (a trainable language model) to refine base prompts and generate a corresponding reasoning trace. This generated prompt is evaluated by a **frozen Evaluation Model ($\\pi_{\\text{eval}}$)** (an LLM used only for inference), which calculates rewards based on formatting and task performance. The generator is optimized using the *Group Relative Policy Optimization (GRPO)* update rule.\n4.  **Prompt Selection Strategy:** The method uses a prompt selection strategy to mitigate training instability and noisy feedback, regularly testing generated prompts on the validation set and keeping the best overall one.\n\n**Main Contributions and Results:**\n\nPRL achieves **state-of-the-art performance** across a range of benchmarks: text classification, summarization, simplification, and GSM8K mathematical reasoning.\n\n*   On the **classification task**, PRL surpasses APE by **2.58%** and EvoPrompt by **1.00%** in mean accuracy.\n*   On **summarization**, it improves average ROUGE scores by **4.32** over APE and **2.12** over EvoPrompt.\n*   On **simplification**, it improves the SARI score by **6.93** over APE and **6.01** over EvoPrompt.\n*   On the **GSM8K mathematical reasoning** benchmark, PRL improves accuracy by **2.72%** over APE and **4.53%** over EvoPrompt.\n*   The research further suggests that **RL-based optimization naturally leads to the emergence of few-shot prompting behavior**.", "review_text": "Effective prompt engineering remains a **central challenge** in fully harnessing the capabilities of Large Language Models (LLMs). While precise input prompts can guide LLMs to perform complex tasks, the most impactful prompts often rely on *subtle semantic cues* that may *elude human perception*.\n\nTo address this, the paper introduces **PRL (Prompts from Reinforcement Learning)**, a novel RL-based approach for automatic prompt generation.\n\n**PRL's key innovations include:**\n\n1.  **Novel Few-shot Example Synthesis:** Unlike previous methods like Automatic Prompt Optimization (APO), which are restricted to selecting few-shot examples from training data, PRL is capable of generating and selecting **novel few-shot examples** that were *not seen during training*.\n2.  **Explicit Reasoning Integration:** PRL incorporates a *reasoning phase* prior to prompt generation, where the prompt generator first produces a **rationale** (enclosed within `<think>` tags) to guide the final output.\n3.  **RL Optimization Cycle:** PRL trains a **Prompt Generator ($\\pi_{\\text{generator}}$)** (a trainable language model) to refine base prompts and generate a corresponding reasoning trace. This generated prompt is evaluated by a **frozen Evaluation Model ($\\pi_{\\text{eval}}$)** (an LLM used only for inference), which calculates rewards based on formatting and task performance. The generator is optimized using the *Group Relative Policy Optimization (GRPO)* update rule.\n4.  **Prompt Selection Strategy:** The method uses a prompt selection strategy to mitigate training instability and noisy feedback, regularly testing generated prompts on the validation set and keeping the best overall one.\n\n**Main Contributions and Results:**\n\nPRL achieves **state-of-the-art performance** across a range of benchmarks: text classification, summarization, simplification, and GSM8K mathematical reasoning.\n\n*   On the **classification task**, PRL surpasses APE by **2.58%** and EvoPrompt by **1.00%** in mean accuracy.\n*   On **summarization**, it improves average ROUGE scores by **4.32** over APE and **2.12** over EvoPrompt.\n*   On **simplification**, it improves the SARI score by **6.93** over APE and **6.01** over EvoPrompt.\n*   On the **GSM8K mathematical reasoning** benchmark, PRL improves accuracy by **2.72%** over APE and **4.53%** over EvoPrompt.\n*   The research further suggests that **RL-based optimization naturally leads to the emergence of few-shot prompting behavior**.", "strengths": "**Originality:**\n\n*   PRL is highlighted as the **first RL-based prompt optimization method** capable of *generating and selecting novel, task-specific few-shot examples*. This is a crucial distinction, as it moves beyond the constraint of only using few-shot examples already present in the training data, a limitation faced by methods like APO.\n*   The observation that few-shot examples emerge **spontaneously** during the RL training process, without explicit encouragement, is a unique and insightful finding regarding how RL shapes LLM prompting behavior.\n\n**Quality & Significance:**\n\n*   The method demonstrates **superior performance** consistently across all four evaluated task types (classification, summarization, simplification, and reasoning), validating its robust generalizability and effectiveness.\n*   **Ablation studies confirm the value of core components:** The explicit reasoning phase proved critical, leading to a *substantial drop in accuracy* (from 75.05 to 60.12) on the SUBJ dataset when omitted. Furthermore, the Prompt Selection strategy was shown to *improve final performance and enhance training efficiency*, helping to manage the high variance inherent in RL training.\n*   PRL is shown to be effective even when applied to **larger, more powerful LLMs** (e.g., Qwen2-32B-instruct), demonstrating that even these models remain *vulnerable to prompt variation* and can benefit significantly from PRL's tailored prompts.\n\n**Clarity:**\n\n*   The architecture, including the roles of the Prompt Generator and the frozen Evaluation Model, and the overall RL training scheme (Figure 2) are clearly described.\n*   The reward function is systematically broken down into components: **formatting rewards** ($r_{token}$, $r_{structure}$) for the Generator, and **task performance rewards** ($r_{format}$, $r_{alignment}$) for the Evaluation Model, providing clarity on how the model’s behavior is guided.", "weaknesses": "1.  **Significant Computational Overhead:** The paper explicitly lists as a limitation that the improved performance is obtained at the cost of a **significantly greater computational expense** than related, comparatively simpler work. The experimental setup involved training over 48 hours using two NVIDIA A100 GPUs (40 GB each). The paper lacks concrete discussion or suggested methods for mitigating or quantifying this increased computational burden, which limits its practical accessibility.\n2.  **Task-Specific Retraining Requirement:** Currently, the prompt generator must be **retrained for each new task**. The authors acknowledge that developing a *universal prompt generator* is a \"desideratum\" (an ideal goal), indicating that the current method’s efficiency is limited when facing a wide variety of tasks or zero-shot scenarios.\n3.  **Insufficient Detail on Few-shot Synthesis:** PRL's ability to **autonomously synthesize relevant few-shot examples** not present in the training set is a major contribution. However, the paper does not delve into *how* the prompt generator, guided by RL, manages to *create* these task-aligned, non-redundant examples—specifically, the internal reasoning or constraints utilized by $\\pi_{\\text{generator}}$ in its thought process ($\\langle \\text{think} \\rangle$ tag) to achieve this synthesis.\n4.  **Sensitivity of Baselines to Evaluation Model Choice:** The authors note that when reproducing EvoPrompt results, the relative effectiveness of its DE and GA variants was **sensitive to the choice of the underlying language model** (Qwen2.5-7B-Instruct). Although the authors ensured a fair comparison by using the same Evaluation Model across all baselines, this inherent sensitivity suggests that the observed superiority of PRL might be conditional on the chosen model, necessitating more explicit acknowledgement of this limitation in interpreting the main results.", "questions": "1.  **Computational Efficiency and Trade-offs:** Given that PRL requires a *significantly greater computational expense*, could the authors provide a more detailed analysis of the performance gains versus the resource cost? Are there specific tuning levers (e.g., Prompt Selection frequency $t$, or the number of sampled prompts $n$) that could be adjusted to reduce training time substantially while maintaining competitive performance against baselines?\n2.  **Mechanics of Task-Dependent Few-Shot Emergence:** Few-shot examples were critical for classification tasks (improving accuracy significantly, e.g., SUBJ from 66.75 to 77.95), but PRL **consistently opted *not* to include few-shot examples** for the summarization task. What drives this striking, task-dependent behavior? Which specific components of the comprehensive reward function $R$ (e.g., $r_{alignment}$ or $r_{structure}$) are responsible for prompting $\\pi_{\\text{generator}}$ to spontaneously generate few-shot examples for classification but omit them for generation tasks?\n3.  **Path to a Universal Prompt Generator:** The paper identifies the need to *retrain the generator for each new task* as a limitation, noting that a *universal prompt generator* is a \"desideratum\". What are the initial conceptual steps or future research directions the authors are considering to enable the Prompt Generator to transfer or generalize prompting knowledge across different tasks without requiring full retraining?\n4.  **Baseline Robustness Across Evaluation Models:** All SOTA claims are established by evaluating baselines on the Qwen2.5-7B-Instruct model. Given the noted sensitivity of EvoPrompt variants to the base model choice, how would a *complete* set of benchmark results (including APE, EvoPrompt, and APO where applicable) compare if a distinct architecture, such as LLaMA 3.1-8B-Instruct, was used as the Evaluation Model for *all* methods, similar to the setup used for the portability study?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Effective prompt engineering remains a **central challenge** in fully harnessing the capabilities of Large Language Models (LLMs). While precise input prompts can guide LLMs to perform complex tasks, the most impactful prompts often rely on *subtle semantic cues* that may *elude human perception*.\n\nTo address this, the paper introduces **PRL (Prompts from Reinforcement Learning)**, a novel RL-based approach for automatic prompt generation.\n\n**PRL's key innovations include:**\n\n1.  **Novel Few-shot Example Synthesis:** Unlike previous methods like Automatic Prompt Optimization (APO), which are restricted to selecting few-shot examples from training data, PRL is capable of generating and selecting **novel few-shot examples** that were *not seen during training*.\n2.  **Explicit Reasoning Integration:** PRL incorporates a *reasoning phase* prior to prompt generation, where the prompt generator first produces a **rationale** (enclosed within `<think>` tags) to guide the final output.\n3.  **RL Optimization Cycle:** PRL trains a **Prompt Generator ($\\pi_{\\text{generator}}$)** (a trainable language model) to refine base prompts and generate a corresponding reasoning trace. This generated prompt is evaluated by a **frozen Evaluation Model ($\\pi_{\\text{eval}}$)** (an LLM used only for inference), which calculates rewards based on formatting and task performance. The generator is optimized using the *Group Relative Policy Optimization (GRPO)* update rule.\n4.  **Prompt Selection Strategy:** The method uses a prompt selection strategy to mitigate training instability and noisy feedback, regularly testing generated prompts on the validation set and keeping the best overall one.\n\n**Main Contributions and Results:**\n\nPRL achieves **state-of-the-art performance** across a range of benchmarks: text classification, summarization, simplification, and GSM8K mathematical reasoning.\n\n*   On the **classification task**, PRL surpasses APE by **2.58%** and EvoPrompt by **1.00%** in mean accuracy.\n*   On **summarization**, it improves average ROUGE scores by **4.32** over APE and **2.12** over EvoPrompt.\n*   On **simplification**, it improves the SARI score by **6.93** over APE and **6.01** over EvoPrompt.\n*   On the **GSM8K mathematical reasoning** benchmark, PRL improves accuracy by **2.72%** over APE and **4.53%** over EvoPrompt.\n*   The research further suggests that **RL-based optimization naturally leads to the emergence of few-shot prompting behavior**.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "**Originality:**\n\n*   PRL is highlighted as the **first RL-based prompt optimization method** capable of *generating and selecting novel, task-specific few-shot examples*. This is a crucial distinction, as it moves beyond the constraint of only using few-shot examples already present in the training data, a limitation faced by methods like APO.\n*   The observation that few-shot examples emerge **spontaneously** during the RL training process, without explicit encouragement, is a unique and insightful finding regarding how RL shapes LLM prompting behavior.\n\n**Quality & Significance:**\n\n*   The method demonstrates **superior performance** consistently across all four evaluated task types (classification, summarization, simplification, and reasoning), validating its robust generalizability and effectiveness.\n*   **Ablation studies confirm the value of core components:** The explicit reasoning phase proved critical, leading to a *substantial drop in accuracy* (from 75.05 to 60.12) on the SUBJ dataset when omitted. Furthermore, the Prompt Selection strategy was shown to *improve final performance and enhance training efficiency*, helping to manage the high variance inherent in RL training.\n*   PRL is shown to be effective even when applied to **larger, more powerful LLMs** (e.g., Qwen2-32B-instruct), demonstrating that even these models remain *vulnerable to prompt variation* and can benefit significantly from PRL's tailored prompts.\n\n**Clarity:**\n\n*   The architecture, including the roles of the Prompt Generator and the frozen Evaluation Model, and the overall RL training scheme (Figure 2) are clearly described.\n*   The reward function is systematically broken down into components: **formatting rewards** ($r_{token}$, $r_{structure}$) for the Generator, and **task performance rewards** ($r_{format}$, $r_{alignment}$) for the Evaluation Model, providing clarity on how the model’s behavior is guided.", "weaknesses": "1.  **Significant Computational Overhead:** The paper explicitly lists as a limitation that the improved performance is obtained at the cost of a **significantly greater computational expense** than related, comparatively simpler work. The experimental setup involved training over 48 hours using two NVIDIA A100 GPUs (40 GB each). The paper lacks concrete discussion or suggested methods for mitigating or quantifying this increased computational burden, which limits its practical accessibility.\n2.  **Task-Specific Retraining Requirement:** Currently, the prompt generator must be **retrained for each new task**. The authors acknowledge that developing a *universal prompt generator* is a \"desideratum\" (an ideal goal), indicating that the current method’s efficiency is limited when facing a wide variety of tasks or zero-shot scenarios.\n3.  **Insufficient Detail on Few-shot Synthesis:** PRL's ability to **autonomously synthesize relevant few-shot examples** not present in the training set is a major contribution. However, the paper does not delve into *how* the prompt generator, guided by RL, manages to *create* these task-aligned, non-redundant examples—specifically, the internal reasoning or constraints utilized by $\\pi_{\\text{generator}}$ in its thought process ($\\langle \\text{think} \\rangle$ tag) to achieve this synthesis.\n4.  **Sensitivity of Baselines to Evaluation Model Choice:** The authors note that when reproducing EvoPrompt results, the relative effectiveness of its DE and GA variants was **sensitive to the choice of the underlying language model** (Qwen2.5-7B-Instruct). Although the authors ensured a fair comparison by using the same Evaluation Model across all baselines, this inherent sensitivity suggests that the observed superiority of PRL might be conditional on the chosen model, necessitating more explicit acknowledgement of this limitation in interpreting the main results.", "questions": "1.  **Computational Efficiency and Trade-offs:** Given that PRL requires a *significantly greater computational expense*, could the authors provide a more detailed analysis of the performance gains versus the resource cost? Are there specific tuning levers (e.g., Prompt Selection frequency $t$, or the number of sampled prompts $n$) that could be adjusted to reduce training time substantially while maintaining competitive performance against baselines?\n2.  **Mechanics of Task-Dependent Few-Shot Emergence:** Few-shot examples were critical for classification tasks (improving accuracy significantly, e.g., SUBJ from 66.75 to 77.95), but PRL **consistently opted *not* to include few-shot examples** for the summarization task. What drives this striking, task-dependent behavior? Which specific components of the comprehensive reward function $R$ (e.g., $r_{alignment}$ or $r_{structure}$) are responsible for prompting $\\pi_{\\text{generator}}$ to spontaneously generate few-shot examples for classification but omit them for generation tasks?\n3.  **Path to a Universal Prompt Generator:** The paper identifies the need to *retrain the generator for each new task* as a limitation, noting that a *universal prompt generator* is a \"desideratum\". What are the initial conceptual steps or future research directions the authors are considering to enable the Prompt Generator to transfer or generalize prompting knowledge across different tasks without requiring full retraining?\n4.  **Baseline Robustness Across Evaluation Models:** All SOTA claims are established by evaluating baselines on the Qwen2.5-7B-Instruct model. Given the noted sensitivity of EvoPrompt variants to the base model choice, how would a *complete* set of benchmark results (including APE, EvoPrompt, and APO where applicable) compare if a distinct architecture, such as LLaMA 3.1-8B-Instruct, was used as the Evaluation Model for *all* methods, similar to the setup used for the portability study?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761566384808}, {"id": "0goXLAloyz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4095/Reviewer_rKJG"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents PRL (Prompts from Reinforcement Learning), a method for automatic prompt generation using reinforcement learning. The approach formulates prompt construction as a learning problem, enabling the model to generate few-shot examples that were not seen during training. The authors evaluate PRL on several tasks, including text classification, text simplification, summarization, and the GSM8K reasoning benchmark. Experimental results show consistent improvements over prior prompt optimization methods such as APE and EvoPrompt, with reported gains in accuracy, ROUGE, SARI, and reasoning performance. The paper claims that PRL offers a general and effective framework for enhancing LLM performance through learned prompts", "review_text": "This paper presents PRL (Prompts from Reinforcement Learning), a method for automatic prompt generation using reinforcement learning. The approach formulates prompt construction as a learning problem, enabling the model to generate few-shot examples that were not seen during training. The authors evaluate PRL on several tasks, including text classification, text simplification, summarization, and the GSM8K reasoning benchmark. Experimental results show consistent improvements over prior prompt optimization methods such as APE and EvoPrompt, with reported gains in accuracy, ROUGE, SARI, and reasoning performance. The paper claims that PRL offers a general and effective framework for enhancing LLM performance through learned prompts", "strengths": "1.\tThe method is clear and intuitive. The paper is well-written and easy to follow.\n\n2.\tOverall, the experimental results effectively support the authors’ claims.", "weaknesses": "1.\tCompared to other approaches discussed in the paper, this method appears to be more computationally intensive.\n\n2.\tThe core of the method lies in the prompt generator, but the authors only use Qwen2-7B as the base model. It would be more convincing to evaluate PRL on models of different sizes and from different model families.\n\n3.\tThe generalization ability of PRL is not thoroughly studied. As different models exhibit varying levels of sensitivity and preference toward prompts, PRL may need to train separate prompt generators for different evaluation models. More tests about this could be further included.\n\n4.\tAdditional ablation studies could strengthen the paper. For instance, the influence of the “thinking process” in the prompt generator is not clearly analyzed. How much would the results change if this component were removed? If the performance remains similar, the additional computation might be unnecessary.", "questions": "1.\tThe format of the citation at line 053 seems incorrect.\n\n2.\tThe prompt generator in PRL requires a base prompt. How sensitive is the method to the quality of this base prompt? Would iterative refinement improve performance? Additionally, what would happen if prompts were generated entirely from scratch?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents PRL (Prompts from Reinforcement Learning), a method for automatic prompt generation using reinforcement learning. The approach formulates prompt construction as a learning problem, enabling the model to generate few-shot examples that were not seen during training. The authors evaluate PRL on several tasks, including text classification, text simplification, summarization, and the GSM8K reasoning benchmark. Experimental results show consistent improvements over prior prompt optimization methods such as APE and EvoPrompt, with reported gains in accuracy, ROUGE, SARI, and reasoning performance. The paper claims that PRL offers a general and effective framework for enhancing LLM performance through learned prompts", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.\tThe method is clear and intuitive. The paper is well-written and easy to follow.\n\n2.\tOverall, the experimental results effectively support the authors’ claims.", "weaknesses": "1.\tCompared to other approaches discussed in the paper, this method appears to be more computationally intensive.\n\n2.\tThe core of the method lies in the prompt generator, but the authors only use Qwen2-7B as the base model. It would be more convincing to evaluate PRL on models of different sizes and from different model families.\n\n3.\tThe generalization ability of PRL is not thoroughly studied. As different models exhibit varying levels of sensitivity and preference toward prompts, PRL may need to train separate prompt generators for different evaluation models. More tests about this could be further included.\n\n4.\tAdditional ablation studies could strengthen the paper. For instance, the influence of the “thinking process” in the prompt generator is not clearly analyzed. How much would the results change if this component were removed? If the performance remains similar, the additional computation might be unnecessary.", "questions": "1.\tThe format of the citation at line 053 seems incorrect.\n\n2.\tThe prompt generator in PRL requires a base prompt. How sensitive is the method to the quality of this base prompt? Would iterative refinement improve performance? Additionally, what would happen if prompts were generated entirely from scratch?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761550022129}], "openreview_url": "https://openreview.net/forum?id=4eYSSSDle6", "arxiv_id": "2505.14412", "paper_pdf": "papers/4eYSSSDle6.pdf", "paper_pdf_sha256": "f1566a76074a2d37315d2af3a19028f167133afec792b8cdef1ff017cabaac61", "paper_pdf_bytes": 380024, "paper_pdf_source": "openreview", "code_url": "https://github.com/Batorskq/PRL-Prompts-from-Reinforcement-Learning", "code_repository": "Batorskq/PRL-Prompts-from-Reinforcement-Learning", "code_commit": "f9797cb987f5fd9f2ce5aa7b9e1e15b12f6ea328", "code_archive": "repos/4eYSSSDle6.zip", "code_archive_sha256": "b1c37c1a30cb5a61d90d3ade1bb1af8207ed3e00720d5aacbd20ea2ff0bbf95f", "code_archive_bytes": 7940967, "code_file_count": 276, "code_extensions": {".py": 267, ".sh": 9}, "github_disk_usage_kb": 6953, "github_languages": {"Python": 1790258, "Shell": 11501}, "github_archived": false, "github_pushed_at": "2025-06-05T11:48:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/prl-prompts-from-reinforcement-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "O6W9SJRZRA", "year": 2025, "status": "rejected", "title": "Right on Time: Revising Time Series Models by Constraining their Explanations", "authors": ["Maurice Kraus", "David Steinmann", "Antonia Wüst", "Andre Kokozinski", "Kristian Kersting"], "authorids": ["~Maurice_Kraus1", "~David_Steinmann1", "~Antonia_Wüst1", "~Andre_Kokozinski1", "~Kristian_Kersting1"], "authors_source": "OpenReview API", "abstract": "The reliability of deep time series models is often compromised by their tendency to rely on confounding factors, which may lead to incorrect outputs. Our newly recorded, naturally confounded dataset named P2S from a real mechanical production line emphasizes this. To avoid “Clever-Hans” moments in time series, i.e., to mitigate confounders, we introduce the method Right on Time (RioT). RioT enables, for the first time interactions with model explanations across both the time and frequency domain. Feedback on explanations in both domains is then used to constrain the model, steering it away from the annotated confounding factors. The dual-domain interaction strategy is crucial for effectively addressing confounders in time series datasets. We empirically demonstrate that RioT can effectively guide models away from the wrong reasons in P2S as well as popular time series classification and forecasting datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "uNbSRXRCUj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10443/Reviewer_EVUz"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes a right for the right reason method (RRR) for time series data, which they refer to as Right On Time RioT. Similar to the RRR method, the method involves first extracting explanations for the model's predictions. They use IG for explanations, and then, using human feedback, they add a loss to penalize the model for looking at the incorrect locations. Since they do this for time series, they propose adding two losses, one for the time domain and one for the frequency domain. RioT reduces the model's reliance on confounding factors during training. The paper introduces a new dataset called PRODUCTION PRESS SENSOR DATA (P2S), which captures sensor readings from an industrial high-speed press in the sheet metal working industry. This dataset is characterized by naturally occurring confounders that can lead to inaccurate predictions when used for training models to detect production faults. P2S provides explicitly annotated confounders, P2S can be used for testing and comparing methods to mitigate the impact of confounders on real-world data.\n\n## RioT\n\n1- *Explain* The paper proposes using IG to extract model explanations for classification problems. The paper uses a modified version of IG with the input magnitude and not the input sign to compute the IG attribution equation 1; for forecasting, the paper uses the average over the forecasting window as an explanation, as shown in equation 2. To get explanations in the frequency domain, the paper transforms the explanations from the time domain to the frequency domain using the Fourier transformation.\n\n2- *Obtain* The paper assumes the user provides annotations of cofounders, resulting in a mask {0,1} mask where 1 represents the presence of a confounder.\n\n3-*Revise* An additional loss term is added to the objective that penalizes the model for looking at the wrong location. This is done by multiplying the explanation produced by the IG with the annotation mask. They propose adding a loss term for both frequency and time domains.\n\n## Experiments\n\n# Datasets\n\n- The paper introduces P2S, which is annotated with confounders. It is a classification dataset. In this dataset, the paper refers to the run speed as a confounder, and experts highlight regions in the time series where there is a difference in the time series due to the difference in the run speed, not in the label itself.\n\n- The paper evaluates UCR/UEA datasets to test whether RioT can help mitigate confounders by adding spatial or frequency shortcuts to the datasets. These confounders result in false correlations between patterns and class labels or forecasting signals within the training data, which do not appear in the validation or test data. An annotation mask is created to simulate human feedback based on the confounder's region or frequency.\n\n# Models\n\n- For classification, they use RioT on FCN and OFA models.\n\n- For regression, they evaluate adding RioT to the TiDE, PatchTST, and NBEATS models.\n\n# Results\n\n- For classification: The paper showed that adding RioT helps improve the accuracy when the data has spurious correlations present in the training dataset but not in the test data across 4 datasets in both spatial and frequency domains.\n\n- For forecasting: The paper showed that adding RioT for confounding datasets reduces MSE in some cases, giving MSE even better the the unconfounded datasets.\n\n- On P2S: They showed that even with partial human feedback, RioT could help the model focus on the correct regions in the input data, resulting in better overall accuracy.\n\n- Multiple confounders: The paper investigates when confounders exist in both frequency and spatial domains; it shows that using the aggregate loss functions with both frequency and spatial feedback gives the best results in this case.\n\n- Human Feedback: The paper shows that even with a small amount of feedback, 5% of the original dataset significantly boosts accuracy. The paper also shows RioT is robust to noisy feedback. They showed that with even 50% noisy feedback, using RioT can still result in some performance gains.", "review_text": "The paper proposes a right for the right reason method (RRR) for time series data, which they refer to as Right On Time RioT. Similar to the RRR method, the method involves first extracting explanations for the model's predictions. They use IG for explanations, and then, using human feedback, they add a loss to penalize the model for looking at the incorrect locations. Since they do this for time series, they propose adding two losses, one for the time domain and one for the frequency domain. RioT reduces the model's reliance on confounding factors during training. The paper introduces a new dataset called PRODUCTION PRESS SENSOR DATA (P2S), which captures sensor readings from an industrial high-speed press in the sheet metal working industry. This dataset is characterized by naturally occurring confounders that can lead to inaccurate predictions when used for training models to detect production faults. P2S provides explicitly annotated confounders, P2S can be used for testing and comparing methods to mitigate the impact of confounders on real-world data.\n\n## RioT\n\n1- *Explain* The paper proposes using IG to extract model explanations for classification problems. The paper uses a modified version of IG with the input magnitude and not the input sign to compute the IG attribution equation 1; for forecasting, the paper uses the average over the forecasting window as an explanation, as shown in equation 2. To get explanations in the frequency domain, the paper transforms the explanations from the time domain to the frequency domain using the Fourier transformation.\n\n2- *Obtain* The paper assumes the user provides annotations of cofounders, resulting in a mask {0,1} mask where 1 represents the presence of a confounder.\n\n3-*Revise* An additional loss term is added to the objective that penalizes the model for looking at the wrong location. This is done by multiplying the explanation produced by the IG with the annotation mask. They propose adding a loss term for both frequency and time domains.\n\n## Experiments\n\n# Datasets\n\n- The paper introduces P2S, which is annotated with confounders. It is a classification dataset. In this dataset, the paper refers to the run speed as a confounder, and experts highlight regions in the time series where there is a difference in the time series due to the difference in the run speed, not in the label itself.\n\n- The paper evaluates UCR/UEA datasets to test whether RioT can help mitigate confounders by adding spatial or frequency shortcuts to the datasets. These confounders result in false correlations between patterns and class labels or forecasting signals within the training data, which do not appear in the validation or test data. An annotation mask is created to simulate human feedback based on the confounder's region or frequency.\n\n# Models\n\n- For classification, they use RioT on FCN and OFA models.\n\n- For regression, they evaluate adding RioT to the TiDE, PatchTST, and NBEATS models.\n\n# Results\n\n- For classification: The paper showed that adding RioT helps improve the accuracy when the data has spurious correlations present in the training dataset but not in the test data across 4 datasets in both spatial and frequency domains.\n\n- For forecasting: The paper showed that adding RioT for confounding datasets reduces MSE in some cases, giving MSE even better the the unconfounded datasets.\n\n- On P2S: They showed that even with partial human feedback, RioT could help the model focus on the correct regions in the input data, resulting in better overall accuracy.\n\n- Multiple confounders: The paper investigates when confounders exist in both frequency and spatial domains; it shows that using the aggregate loss functions with both frequency and spatial feedback gives the best results in this case.\n\n- Human Feedback: The paper shows that even with a small amount of feedback, 5% of the original dataset significantly boosts accuracy. The paper also shows RioT is robust to noisy feedback. They showed that with even 50% noisy feedback, using RioT can still result in some performance gains.", "strengths": "## Originality\n\n- The application of RRR methods to time series data is novel.\n- Introducing two RRR loss terms for spatial and time domains is novel.\n- The P2S dataset is a new contribution, featuring confounder annotations that can be utilized for future method evaluations.\n- The unique approach to adding confounders to the UCR/UEA datasets is original; this dataset can serve a similar purpose to DomainBed in image analysis (https://github.com/facebookresearch/DomainBed) for future model evaluations.\n\n## Quality\n\n- The application of RRR in this context is well-justified.\n- The paper presents robust empirical results.\n- It explores various scenarios where feedback was either incomplete or incorrect, demonstrating that RioT remains effective in such cases.\n- The paper is well-written and easy to understand.\n\n## Significance\n\n- With the introduction of the new P2S dataset and the proposed setup for UCR/UEA, this work is quite significant. It provides valuable resources that the machine learning community can use and build upon.", "weaknesses": "## Novelty:\n-  RioT itself is not novel, it is just applying RRR to time series data.\n## Clarity:\n- Some minor clarity issues; please see the question section.", "questions": "- What is $\\bar{x}$ in equation 1? if it's the reference point, please clarify and discuss how you choose the reference point for time series data.\n\n- In equation 2, how is $e'_i(x)$ calculated?\n\n- In line 249, why is the second-order derivative required? IG traditionally uses first-order derivatives.\n\n- In Figure 4, it is difficult to see the attribution colors.\n\n- In the experiment \"Removing Multiple Confounders at Once,\" are the masks (that mimic human feedback) given in both spatial and frequency domains? i.e., different masks for different domains? If this is not the case, I am not sure why applying RRR in both frequency and time domains will necessarily help. Since one is the transformation of the other. What is your intuition as to why this might be the case?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a right for the right reason method (RRR) for time series data, which they refer to as Right On Time RioT. Similar to the RRR method, the method involves first extracting explanations for the model's predictions. They use IG for explanations, and then, using human feedback, they add a loss to penalize the model for looking at the incorrect locations. Since they do this for time series, they propose adding two losses, one for the time domain and one for the frequency domain. RioT reduces the model's reliance on confounding factors during training. The paper introduces a new dataset called PRODUCTION PRESS SENSOR DATA (P2S), which captures sensor readings from an industrial high-speed press in the sheet metal working industry. This dataset is characterized by naturally occurring confounders that can lead to inaccurate predictions when used for training models to detect production faults. P2S provides explicitly annotated confounders, P2S can be used for testing and comparing methods to mitigate the impact of confounders on real-world data.\n\n## RioT\n\n1- *Explain* The paper proposes using IG to extract model explanations for classification problems. The paper uses a modified version of IG with the input magnitude and not the input sign to compute the IG attribution equation 1; for forecasting, the paper uses the average over the forecasting window as an explanation, as shown in equation 2. To get explanations in the frequency domain, the paper transforms the explanations from the time domain to the frequency domain using the Fourier transformation.\n\n2- *Obtain* The paper assumes the user provides annotations of cofounders, resulting in a mask {0,1} mask where 1 represents the presence of a confounder.\n\n3-*Revise* An additional loss term is added to the objective that penalizes the model for looking at the wrong location. This is done by multiplying the explanation produced by the IG with the annotation mask. They propose adding a loss term for both frequency and time domains.\n\n## Experiments\n\n# Datasets\n\n- The paper introduces P2S, which is annotated with confounders. It is a classification dataset. In this dataset, the paper refers to the run speed as a confounder, and experts highlight regions in the time series where there is a difference in the time series due to the difference in the run speed, not in the label itself.\n\n- The paper evaluates UCR/UEA datasets to test whether RioT can help mitigate confounders by adding spatial or frequency shortcuts to the datasets. These confounders result in false correlations between patterns and class labels or forecasting signals within the training data, which do not appear in the validation or test data. An annotation mask is created to simulate human feedback based on the confounder's region or frequency.\n\n# Models\n\n- For classification, they use RioT on FCN and OFA models.\n\n- For regression, they evaluate adding RioT to the TiDE, PatchTST, and NBEATS models.\n\n# Results\n\n- For classification: The paper showed that adding RioT helps improve the accuracy when the data has spurious correlations present in the training dataset but not in the test data across 4 datasets in both spatial and frequency domains.\n\n- For forecasting: The paper showed that adding RioT for confounding datasets reduces MSE in some cases, giving MSE even better the the unconfounded datasets.\n\n- On P2S: They showed that even with partial human feedback, RioT could help the model focus on the correct regions in the input data, resulting in better overall accuracy.\n\n- Multiple confounders: The paper investigates when confounders exist in both frequency and spatial domains; it shows that using the aggregate loss functions with both frequency and spatial feedback gives the best results in this case.\n\n- Human Feedback: The paper shows that even with a small amount of feedback, 5% of the original dataset significantly boosts accuracy. The paper also shows RioT is robust to noisy feedback. They showed that with even 50% noisy feedback, using RioT can still result in some performance gains.", "soundness": 4, "presentation": 3, "contribution": 3, "strengths": "## Originality\n\n- The application of RRR methods to time series data is novel.\n- Introducing two RRR loss terms for spatial and time domains is novel.\n- The P2S dataset is a new contribution, featuring confounder annotations that can be utilized for future method evaluations.\n- The unique approach to adding confounders to the UCR/UEA datasets is original; this dataset can serve a similar purpose to DomainBed in image analysis (https://github.com/facebookresearch/DomainBed) for future model evaluations.\n\n## Quality\n\n- The application of RRR in this context is well-justified.\n- The paper presents robust empirical results.\n- It explores various scenarios where feedback was either incomplete or incorrect, demonstrating that RioT remains effective in such cases.\n- The paper is well-written and easy to understand.\n\n## Significance\n\n- With the introduction of the new P2S dataset and the proposed setup for UCR/UEA, this work is quite significant. It provides valuable resources that the machine learning community can use and build upon.", "weaknesses": "## Novelty:\n-  RioT itself is not novel, it is just applying RRR to time series data.\n## Clarity:\n- Some minor clarity issues; please see the question section.", "questions": "- What is $\\bar{x}$ in equation 1? if it's the reference point, please clarify and discuss how you choose the reference point for time series data.\n\n- In equation 2, how is $e'_i(x)$ calculated?\n\n- In line 249, why is the second-order derivative required? IG traditionally uses first-order derivatives.\n\n- In Figure 4, it is difficult to see the attribution colors.\n\n- In the experiment \"Removing Multiple Confounders at Once,\" are the masks (that mimic human feedback) given in both spatial and frequency domains? i.e., different masks for different domains? If this is not the case, I am not sure why applying RRR in both frequency and time domains will necessarily help. Since one is the transformation of the other. What is your intuition as to why this might be the case?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730650156921}, {"id": "L3mjWLHg3p", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10443/Reviewer_aQ3G"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper titled introduces a novel method named RioT for mitigating confounding factors in time series models. The authors emphasize the importance of addressing 'Clever-Hans' moments in time series models where spurious correlations may compromise model performance. The key contributions are:\n1. Introduction of a new real-world dataset P2S containing annotated confounders from industrial sensor data.\n2. Proposal of RioT, a feedback-based technique that operates across both time and frequency domains to mitigate confounders.\n3. Empirical evidence showcasing that RioT improves the generalization of time series models across multiple datasets and scenarios.", "review_text": "The paper titled introduces a novel method named RioT for mitigating confounding factors in time series models. The authors emphasize the importance of addressing 'Clever-Hans' moments in time series models where spurious correlations may compromise model performance. The key contributions are:\n1. Introduction of a new real-world dataset P2S containing annotated confounders from industrial sensor data.\n2. Proposal of RioT, a feedback-based technique that operates across both time and frequency domains to mitigate confounders.\n3. Empirical evidence showcasing that RioT improves the generalization of time series models across multiple datasets and scenarios.", "strengths": "1. High practical relevance: The focus on confounders in industrial sensor data highlights the method's real-world applicability.\n2. Novel dual-domain feedback: Operating across both time and frequency domains is an innovative extension to existing XIL paradigms.\n3. Empirical rigor: The experiments span across multiple datasets, models, and configurations, adding robustness to the findings.", "weaknesses": "1. Dependency on human feedback: The method requires expert annotations, which may be costly and challenging to acquire at scale.\n2. Training cost: Incorporating feedback through gradient-based explanations adds computational overhead.\n3. Limited exploration of feedback noise impact: Although robustness tests are performed, more exploration of noisy annotations could be useful to assess real-world viability.", "questions": "1. In Figure 1, it is unclear how annotations on the pretraining data are handled. Can the authors clarify if the feedback on pretraining is incorporated into final model updates?\n2. Could you elaborate on any instances within which incorporating frequency domain feedback negatively affected performance? It would be particularly insightful to explore scenarios where the frequency domain annotations might conflict with those in the time domain.\n3. Given the increased training cost, what strategies can practitioners use to balance the feedback quality and computational efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper titled introduces a novel method named RioT for mitigating confounding factors in time series models. The authors emphasize the importance of addressing 'Clever-Hans' moments in time series models where spurious correlations may compromise model performance. The key contributions are:\n1. Introduction of a new real-world dataset P2S containing annotated confounders from industrial sensor data.\n2. Proposal of RioT, a feedback-based technique that operates across both time and frequency domains to mitigate confounders.\n3. Empirical evidence showcasing that RioT improves the generalization of time series models across multiple datasets and scenarios.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. High practical relevance: The focus on confounders in industrial sensor data highlights the method's real-world applicability.\n2. Novel dual-domain feedback: Operating across both time and frequency domains is an innovative extension to existing XIL paradigms.\n3. Empirical rigor: The experiments span across multiple datasets, models, and configurations, adding robustness to the findings.", "weaknesses": "1. Dependency on human feedback: The method requires expert annotations, which may be costly and challenging to acquire at scale.\n2. Training cost: Incorporating feedback through gradient-based explanations adds computational overhead.\n3. Limited exploration of feedback noise impact: Although robustness tests are performed, more exploration of noisy annotations could be useful to assess real-world viability.", "questions": "1. In Figure 1, it is unclear how annotations on the pretraining data are handled. Can the authors clarify if the feedback on pretraining is incorporated into final model updates?\n2. Could you elaborate on any instances within which incorporating frequency domain feedback negatively affected performance? It would be particularly insightful to explore scenarios where the frequency domain annotations might conflict with those in the time domain.\n3. Given the increased training cost, what strategies can practitioners use to balance the feedback quality and computational efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730512273547}, {"id": "571h3lGPzU", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10443/Reviewer_4ai3"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces RioT, a method to mitigate confounders in time series analysis. The authors adapt established explanatory interactive learning (XIL) methods for time series analysis. XIL includes human-in-the-loop approaches that provide feedback on model explanations, particularly penalising model decisions based on confounding factors such as artifacts and noise. Evaluations across various architectures and downstream applications indicate that integrating human feedback into the training process may help reduce confounding factors in time series analysis.", "review_text": "The paper introduces RioT, a method to mitigate confounders in time series analysis. The authors adapt established explanatory interactive learning (XIL) methods for time series analysis. XIL includes human-in-the-loop approaches that provide feedback on model explanations, particularly penalising model decisions based on confounding factors such as artifacts and noise. Evaluations across various architectures and downstream applications indicate that integrating human feedback into the training process may help reduce confounding factors in time series analysis.", "strengths": "1. The paper is well structured.\n\n2. The paper is clearly written.\n\n3. The authors conduct various experiments including different model architectures and downstream applications to evaluate their method.\n\n4. The authors provide their code to support reproducibility.", "weaknesses": "1. Table 1 to Table 4 as well as the statement \"we evaluate on [...] popular datasets [...] with 70% / 30% train / test splits\" (ll. 268-269) indicate that the authors did not use a validation set, i.e. tuned their models on the test set. The authors should follow good practice and adhere to established procedures, e.g. as introduced by [1] for forecasting (train/val/test split of 60/20/20 for ETTm1, and 70/10/20 for Energy and Weather). \n\n2. The proposed approach requires human feedback for each downstream application, which is expensive as domain knowledge is required. Furthermore, the authors claim that \"RioT is well-equipped to manage the practical challenges associated with human feedback\" (ll 475-476). However, such human-in-the-loop approaches are vulnerable to adversarial attacks, i.e. intentionally poisoned feedback, which the authors do not address in their work. \n\n3. The approach focuses on uni-variate time series analysis. Since the authors do not explain how to extend their method to multi-variate time series, its applicability to real-world scenario remains limited. Additionally, the authors further constrain their approach by excluding \"datasets with time series of different length or missing values\" (l. 752), focusing on \"datasets with less than 10 classes\" (l. 769), and omitting \"all datasets with less than 1000 training samples\" (ll. 771-772) from their analysis. These constraints, especially the first and last, rule out applications in domains with limited data availability, e.g. medicine.\n\n4. The authors use integrated gradients [2] to provide explanations $e(x)$, assuming that labels are available, e.g. for classification tasks, and of good quality. For forecasting tasks, pseudo-labels are created by computing an explanation for each time step $w$ in the forcasting horizon $W$ and creating a single representative explanation using global average pooling over all $W$ explanations, which requires further investigations to rule out points of failure. \n\n5. The methodology section requires clarification, as the authors introduce variables, including $\\bar{x}$, $\\tilde{x}$, and $\\alpha$ in Equation (1), without providing definitions.\n\n6. The reproducibility is not fully supported, as the authors did not describe how they tuned the models for the experiments. \n\n[1] Zhou et al. \"Informer: Beyond efficient transformer for long sequence time-series forecasting.\" AAAI (2021).\n\n[2] Sundararajan et al. \"Axiomatic attribution for deep networks.\" ICML (2017).", "questions": "1. How to provide feedback for medical domains and domains which might contain non-trivial confounders? Does the approach translate well across domains such that feedback might not be required for specific applications?\n\n2. How to evaluate whether the human feedback is of good quality? Is there a way to quantitatively measure feedback quality other than training the model with the respective human feedback and eventually evaluating the downstream performance?\n\n3. How is the human-in-the-loop supposed to identify the confounders in the real and imaginary part within the frequency domain? Could the authors further elaborate on this and provide visual examples for clarity?\n\n4. The model needs to be trained to derive the explanations using integrated gradients. How many iterations of training, i.e. explanations $e(x)$, and successive feedback, i.e. $a(x)$, are required to achieve a sufficiently good model? How does the downstream performance evolve as the number of iterations increase?\n\n5. Could overfitting on confounders be mitigated by carefully selecting the validation set? As the authors assume that \"confounders [...] are absent in [the] test set\" (ll. 309-310), why not construct a validation set with non-confounded samples and use an early stopping criterion on the validation metric to prevent overfitting? This experimental setup would demonstrate how the proposed RioT performs in comparison to cheap early stopping techniques, that do not require human feedback.\n\n6. The authors investigate fully supervised settings. Have they considered to analyse whether pre-trained models, i.e. models that are trained using self-supervised methods, are less prone to overfitting on confounders?\n\n7. Do the authors \"report average and standard deviation over 5 runs\" (l. 305) across different seeds set during fine-tuning?\n\n8. The authors \"add spatial (sp) or frequency (freq) shortcuts to the datasets from the UCR and Darts repositories\" (l. 308). For visualisation purposes, could they also provide examples of representative data samples before and after adding these shortcuts?\n\n9. Could the authors add a qualitative analysis of the learned latent space to visualise how confounders may influence the model prediction?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces RioT, a method to mitigate confounders in time series analysis. The authors adapt established explanatory interactive learning (XIL) methods for time series analysis. XIL includes human-in-the-loop approaches that provide feedback on model explanations, particularly penalising model decisions based on confounding factors such as artifacts and noise. Evaluations across various architectures and downstream applications indicate that integrating human feedback into the training process may help reduce confounding factors in time series analysis.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper is well structured.\n\n2. The paper is clearly written.\n\n3. The authors conduct various experiments including different model architectures and downstream applications to evaluate their method.\n\n4. The authors provide their code to support reproducibility.", "weaknesses": "1. Table 1 to Table 4 as well as the statement \"we evaluate on [...] popular datasets [...] with 70% / 30% train / test splits\" (ll. 268-269) indicate that the authors did not use a validation set, i.e. tuned their models on the test set. The authors should follow good practice and adhere to established procedures, e.g. as introduced by [1] for forecasting (train/val/test split of 60/20/20 for ETTm1, and 70/10/20 for Energy and Weather). \n\n2. The proposed approach requires human feedback for each downstream application, which is expensive as domain knowledge is required. Furthermore, the authors claim that \"RioT is well-equipped to manage the practical challenges associated with human feedback\" (ll 475-476). However, such human-in-the-loop approaches are vulnerable to adversarial attacks, i.e. intentionally poisoned feedback, which the authors do not address in their work. \n\n3. The approach focuses on uni-variate time series analysis. Since the authors do not explain how to extend their method to multi-variate time series, its applicability to real-world scenario remains limited. Additionally, the authors further constrain their approach by excluding \"datasets with time series of different length or missing values\" (l. 752), focusing on \"datasets with less than 10 classes\" (l. 769), and omitting \"all datasets with less than 1000 training samples\" (ll. 771-772) from their analysis. These constraints, especially the first and last, rule out applications in domains with limited data availability, e.g. medicine.\n\n4. The authors use integrated gradients [2] to provide explanations $e(x)$, assuming that labels are available, e.g. for classification tasks, and of good quality. For forecasting tasks, pseudo-labels are created by computing an explanation for each time step $w$ in the forcasting horizon $W$ and creating a single representative explanation using global average pooling over all $W$ explanations, which requires further investigations to rule out points of failure. \n\n5. The methodology section requires clarification, as the authors introduce variables, including $\\bar{x}$, $\\tilde{x}$, and $\\alpha$ in Equation (1), without providing definitions.\n\n6. The reproducibility is not fully supported, as the authors did not describe how they tuned the models for the experiments. \n\n[1] Zhou et al. \"Informer: Beyond efficient transformer for long sequence time-series forecasting.\" AAAI (2021).\n\n[2] Sundararajan et al. \"Axiomatic attribution for deep networks.\" ICML (2017).", "questions": "1. How to provide feedback for medical domains and domains which might contain non-trivial confounders? Does the approach translate well across domains such that feedback might not be required for specific applications?\n\n2. How to evaluate whether the human feedback is of good quality? Is there a way to quantitatively measure feedback quality other than training the model with the respective human feedback and eventually evaluating the downstream performance?\n\n3. How is the human-in-the-loop supposed to identify the confounders in the real and imaginary part within the frequency domain? Could the authors further elaborate on this and provide visual examples for clarity?\n\n4. The model needs to be trained to derive the explanations using integrated gradients. How many iterations of training, i.e. explanations $e(x)$, and successive feedback, i.e. $a(x)$, are required to achieve a sufficiently good model? How does the downstream performance evolve as the number of iterations increase?\n\n5. Could overfitting on confounders be mitigated by carefully selecting the validation set? As the authors assume that \"confounders [...] are absent in [the] test set\" (ll. 309-310), why not construct a validation set with non-confounded samples and use an early stopping criterion on the validation metric to prevent overfitting? This experimental setup would demonstrate how the proposed RioT performs in comparison to cheap early stopping techniques, that do not require human feedback.\n\n6. The authors investigate fully supervised settings. Have they considered to analyse whether pre-trained models, i.e. models that are trained using self-supervised methods, are less prone to overfitting on confounders?\n\n7. Do the authors \"report average and standard deviation over 5 runs\" (l. 305) across different seeds set during fine-tuning?\n\n8. The authors \"add spatial (sp) or frequency (freq) shortcuts to the datasets from the UCR and Darts repositories\" (l. 308). For visualisation purposes, could they also provide examples of representative data samples before and after adding these shortcuts?\n\n9. Could the authors add a qualitative analysis of the learned latent space to visualise how confounders may influence the model prediction?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730199981746}, {"id": "ac95XuyWTU", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10443/Reviewer_acPg"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper proposed a method named Right on Time (RioT), which can enable human interactions with the time series model explanations in both time and frequency domains. The authors also created a confounded dataset named P2S from a real scenario and conducted detailed experiments on both synthetic datasets and this real-world datasets for evaluation.", "review_text": "This paper proposed a method named Right on Time (RioT), which can enable human interactions with the time series model explanations in both time and frequency domains. The authors also created a confounded dataset named P2S from a real scenario and conducted detailed experiments on both synthetic datasets and this real-world datasets for evaluation.", "strengths": "* The idea of human-in-the-loop for removing confounding factors in time series modelling is interesting.\n* The published confounded dataset will be beneficial to the research community.", "weaknesses": "* The proposed method seems to be not very practical in real world applications. Firstly, the proposed method relies on human annotations for the whole datasets, which will require huge labelling costs. Secondly, for many real-world applications, the confounders are often unknown, so that human labelling may not be trivial, especially in frequency domain. \n* The proposed method assumes that all the confounders could be lablled by human experts. First, if all the condounders are known, it will be very convenient to apply causal learning based methods to avoid these confounding effects. Second, the proposed method penalizes the confounding factors by a penality term, which can only reduce their effects but cannot completely remove their effects. Directly removing the confounding factors from the input seems to be more effective than the proposed penality term in my opinion.\n* The frequency domain explanation seems to be a trivial Fourier transformation of the time domain explanation according the descriptions in line 186/187. This means that the two types of explanations are essentially the same thing with different format, which may not be the optimal cases in my opinion. I am not sure if it would be better to directly learn frequency domain explanation independently, i.e., tranforming the input data into frequency domain and then applying IG for obtaining explanations.\n* The presentation should be significantly improved. Here are a few examples:\n    * line 34: in practice Geirhos et al. (2020) -> in practice (Geirhos et al., 2020)\n    * line 124: casual analysis -> causal analysis\n    * line 149: Given is a dataset - > Given a dataset", "questions": "For the major questions, please refer to my comments above.\n\nHere are some additionl questions:\n1. For Table 1, will the accuracy be the same as the unconfounded case if you directly remove all the confounders from the synthetic dataset? If so, a penality term will be unnecessary.\n2. In line 200, you mentioned \"apply the same annotation mask to many samples\". How did you do that automatically?\n3. Would it be possible to apply other unconfounding/causal learning/robust learning methods on the P2S datasest? It would be helpful to compare with other types of methods in addition to comparing different variants of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a method named Right on Time (RioT), which can enable human interactions with the time series model explanations in both time and frequency domains. The authors also created a confounded dataset named P2S from a real scenario and conducted detailed experiments on both synthetic datasets and this real-world datasets for evaluation.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "* The idea of human-in-the-loop for removing confounding factors in time series modelling is interesting.\n* The published confounded dataset will be beneficial to the research community.", "weaknesses": "* The proposed method seems to be not very practical in real world applications. Firstly, the proposed method relies on human annotations for the whole datasets, which will require huge labelling costs. Secondly, for many real-world applications, the confounders are often unknown, so that human labelling may not be trivial, especially in frequency domain. \n* The proposed method assumes that all the confounders could be lablled by human experts. First, if all the condounders are known, it will be very convenient to apply causal learning based methods to avoid these confounding effects. Second, the proposed method penalizes the confounding factors by a penality term, which can only reduce their effects but cannot completely remove their effects. Directly removing the confounding factors from the input seems to be more effective than the proposed penality term in my opinion.\n* The frequency domain explanation seems to be a trivial Fourier transformation of the time domain explanation according the descriptions in line 186/187. This means that the two types of explanations are essentially the same thing with different format, which may not be the optimal cases in my opinion. I am not sure if it would be better to directly learn frequency domain explanation independently, i.e., tranforming the input data into frequency domain and then applying IG for obtaining explanations.\n* The presentation should be significantly improved. Here are a few examples:\n    * line 34: in practice Geirhos et al. (2020) -> in practice (Geirhos et al., 2020)\n    * line 124: casual analysis -> causal analysis\n    * line 149: Given is a dataset - > Given a dataset", "questions": "For the major questions, please refer to my comments above.\n\nHere are some additionl questions:\n1. For Table 1, will the accuracy be the same as the unconfounded case if you directly remove all the confounders from the synthetic dataset? If so, a penality term will be unnecessary.\n2. In line 200, you mentioned \"apply the same annotation mask to many samples\". How did you do that automatically?\n3. Would it be possible to apply other unconfounding/causal learning/robust learning methods on the P2S datasest? It would be helpful to compare with other types of methods in addition to comparing different variants of the proposed method.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729565040136}], "openreview_url": "https://openreview.net/forum?id=O6W9SJRZRA", "arxiv_id": "2402.12921", "paper_pdf": "papers/O6W9SJRZRA.pdf", "paper_pdf_sha256": "6bfb009dd3dc87b73bf435d1ed7fb5263e4730490588c417ad8ed4a8df0e50f4", "paper_pdf_bytes": 2753655, "paper_pdf_source": "openreview", "code_url": "https://github.com/ml-research/RioT", "code_repository": "ml-research/RioT", "code_commit": "0480158d53ec9b2218eec89d7b3f3f8f02277b09", "code_archive": "repos/O6W9SJRZRA.zip", "code_archive_sha256": "27d523b40012ccb9d5e532118e91723b54300dce2c644ac1c73196264aec2547", "code_archive_bytes": 391126, "code_file_count": 86, "code_extensions": {".py": 84, ".ipynb": 2}, "github_disk_usage_kb": 354, "github_languages": {"Jupyter Notebook": 373269, "Python": 291429, "Dockerfile": 811}, "github_archived": false, "github_pushed_at": "2024-06-18T15:51:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/right-on-time-revising-time-series-models-by"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VmqTuFMk68", "year": 2024, "status": "rejected", "title": "Trainable Transformer in Transformer", "authors": ["Abhishek Panigrahi", "Sadhika Malladi", "Mengzhou Xia", "Sanjeev Arora"], "authorids": ["~Abhishek_Panigrahi1", "~Sadhika_Malladi2", "~Mengzhou_Xia1", "~Sanjeev_Arora1"], "authors_source": "OpenReview API", "abstract": "Recent works attribute the capability of in-context learning (ICL) in large pre-trained language models to implicitly simulating and fine-tuning an internal model (e.g., linear or 2-layer MLP) during inference. However, such constructions require large memory overhead, which makes simulation of more sophisticated internal models intractable. In this work, we propose a new efficient construction, Transformer in Transformer (in short, TINT), that allows a transformer to simulate and fine-tune more complex models during inference (e.g., pre-trained language models). In particular, we introduce innovative approximation techniques that allow a TINT model with less than 2 billion parameters to simulate and fine-tune a 125 million parameter transformer model within a single forward pass. TINT accommodates many common transformer variants and its design ideas also improve the efficiency of past instantiations of simple models inside transformers. We conduct end-to-end experiments to validate the internal fine-tuning procedure of TINT on various language modeling and downstream tasks. For example, even with a limited one-step budget, we observe TINT for a OPT-125M model improves performance by 4 − 16% absolute on average compared to OPT-125M. These findings suggest that large pre-trained language models are capable of performing intricate subroutines. To facilitate further work, a modular and extensible codebase for TINT is included.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "g7WNmCg1Ca", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4281/Reviewer_zqc2"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work proposes TinT, a parameter-efficient construction to allow transformers to simulate forward and backward passes. Theoretically, the paper shows that TinT uses fewer parameters than prior construction based on the vanilla transformer architectures.  Empirically, this paper conducts experiments on language tasks to verify that TinT archives similar in-context learning performance compared to dynamic evaluation (which performs one GD step model update during inference) and larger models of a comparable parameter size.", "review_text": "This work proposes TinT, a parameter-efficient construction to allow transformers to simulate forward and backward passes. Theoretically, the paper shows that TinT uses fewer parameters than prior construction based on the vanilla transformer architectures.  Empirically, this paper conducts experiments on language tasks to verify that TinT archives similar in-context learning performance compared to dynamic evaluation (which performs one GD step model update during inference) and larger models of a comparable parameter size.", "strengths": "1. New construction for simulating a backward pass within a model is provided, using fewer model parameters than prior constructions. \n2. Experiments on more realistic language tasks are provided. \n3. The idea of TinT might be useful in designing new language models for other applications.", "weaknesses": "1. Writtings could be improved in some places. For two examples, \n* In definition 2.1, what are the \"relevant\" auxiliary model weights? The current definition is a bit difficult for me to interpret. \n* In definition 2.3, are $p_t$'s referring to positional embedding? Could you explain why there aren't positional embeddings in definition 2.10.\n\n2. Theorem 2.5 shows linear attention could be approximated by softmax attention. Can softmax attention also be approximated by linear attention? If not, I feel Theorem 2.5 alone does not suffice to justify the claim that \"Thus, we often use linear attention in TINT\". Let me know if I have misunderstood anything. In addition, is the claimed parameter saving based on linear attention or self-attention? \n\n3. Definition 2.8 uses finite difference to approximate gradient. I am wondering if we can do this from end to end. That is, can we simulate a backward pass by doing finite-difference and two forward-pass? What's the disadvantage of doing so? \n\n4. This work provides experiments on language tasks, while prior works provide experiments on simulated tasks (e.g., Akyurek et al 2022 did ICL for linear regression). So the empirical results are not directly comparable with prior works. \n\n5. I feel an important prior work [1] is missed. Specifically, [1] also did approximation theory for ICL using transformers. How would the required number of parameters in the construction in this work compare to theirs? \n\n\n[1] Bai, Yu, Fan Chen, Huan Wang, Caiming Xiong, and Song Mei. \"Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection.\" NeurIPS 2023", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes TinT, a parameter-efficient construction to allow transformers to simulate forward and backward passes. Theoretically, the paper shows that TinT uses fewer parameters than prior construction based on the vanilla transformer architectures.  Empirically, this paper conducts experiments on language tasks to verify that TinT archives similar in-context learning performance compared to dynamic evaluation (which performs one GD step model update during inference) and larger models of a comparable parameter size.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. New construction for simulating a backward pass within a model is provided, using fewer model parameters than prior constructions. \n2. Experiments on more realistic language tasks are provided. \n3. The idea of TinT might be useful in designing new language models for other applications.", "weaknesses": "1. Writtings could be improved in some places. For two examples, \n* In definition 2.1, what are the \"relevant\" auxiliary model weights? The current definition is a bit difficult for me to interpret. \n* In definition 2.3, are $p_t$'s referring to positional embedding? Could you explain why there aren't positional embeddings in definition 2.10.\n\n2. Theorem 2.5 shows linear attention could be approximated by softmax attention. Can softmax attention also be approximated by linear attention? If not, I feel Theorem 2.5 alone does not suffice to justify the claim that \"Thus, we often use linear attention in TINT\". Let me know if I have misunderstood anything. In addition, is the claimed parameter saving based on linear attention or self-attention? \n\n3. Definition 2.8 uses finite difference to approximate gradient. I am wondering if we can do this from end to end. That is, can we simulate a backward pass by doing finite-difference and two forward-pass? What's the disadvantage of doing so? \n\n4. This work provides experiments on language tasks, while prior works provide experiments on simulated tasks (e.g., Akyurek et al 2022 did ICL for linear regression). So the empirical results are not directly comparable with prior works. \n\n5. I feel an important prior work [1] is missed. Specifically, [1] also did approximation theory for ICL using transformers. How would the required number of parameters in the construction in this work compare to theirs? \n\n\n[1] Bai, Yu, Fan Chen, Huan Wang, Caiming Xiong, and Song Mei. \"Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection.\" NeurIPS 2023", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699641888524}, {"id": "GTFlrjlCOe", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4281/Reviewer_k6Mj"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces an innovative and parameter-efficient transformer construction, TINT, capable of simulating complex models, such as pre-trained transformers, and applying fine-tuning in-context via a single forward pass.", "review_text": "This paper introduces an innovative and parameter-efficient transformer construction, TINT, capable of simulating complex models, such as pre-trained transformers, and applying fine-tuning in-context via a single forward pass.", "strengths": "1. The paper details a novel transformer construction that enables forward and backward operations, as well as parameter updates, within a single inference pass, and without the need for weight adjustments in the TINT model. This method surpasses previous approaches in terms of parameter efficiency and the complexity of models it can simulate.\n2. The TINT architecture's efficiency is corroborated through real-data evaluations.", "weaknesses": "1. The TINT model requires access to auxiliary model weights, which it uses in prefix embeddings. This dependency differs from some previous works where the transformer independently and implicitly learns the auxiliary model's weights (and architecture) from the provided in-context dataset.\n2. Despite its parameter efficiency, the TINT model's reliance on prefix embeddings to access auxiliary weights may lead to longer input sequence, which could potentially reduce computational and memory efficiency during inference due to the transformer's sequence length dependency.", "questions": "1. The capacity of the TINT structure in mimicking complex models like pre-trained transformers is well-documented, but is it equally adaptable to simpler auxiliary models, such as 2-layer MLPs?\n2. Is the TINT's structure inherently dependent on the configuration of the auxiliary models it simulates, for instance in terms of number of layers, activation functions, or attention mechanisms?\n3. The details of the Backward Module depicted in Figure 1 are unclear. Can you provide more insight into it, particularly regarding the connections between the Backward Module's inputs/outputs across layers and their relationship to the Forward Module's output?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an innovative and parameter-efficient transformer construction, TINT, capable of simulating complex models, such as pre-trained transformers, and applying fine-tuning in-context via a single forward pass.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper details a novel transformer construction that enables forward and backward operations, as well as parameter updates, within a single inference pass, and without the need for weight adjustments in the TINT model. This method surpasses previous approaches in terms of parameter efficiency and the complexity of models it can simulate.\n2. The TINT architecture's efficiency is corroborated through real-data evaluations.", "weaknesses": "1. The TINT model requires access to auxiliary model weights, which it uses in prefix embeddings. This dependency differs from some previous works where the transformer independently and implicitly learns the auxiliary model's weights (and architecture) from the provided in-context dataset.\n2. Despite its parameter efficiency, the TINT model's reliance on prefix embeddings to access auxiliary weights may lead to longer input sequence, which could potentially reduce computational and memory efficiency during inference due to the transformer's sequence length dependency.", "questions": "1. The capacity of the TINT structure in mimicking complex models like pre-trained transformers is well-documented, but is it equally adaptable to simpler auxiliary models, such as 2-layer MLPs?\n2. Is the TINT's structure inherently dependent on the configuration of the auxiliary models it simulates, for instance in terms of number of layers, activation functions, or attention mechanisms?\n3. The details of the Backward Module depicted in Figure 1 are unclear. Can you provide more insight into it, particularly regarding the connections between the Backward Module's inputs/outputs across layers and their relationship to the Forward Module's output?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699308552259}, {"id": "cghExXkARV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4281/Reviewer_KkDE"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a new construction, Transformers in Transformer (TinT), that allows a transformer to simulate and fine-tune transformer models during inference. The authors introduce innovative approximation techniques that allow TinT with less than 2 Billion parameters to simulate and fine-tune a 125 million-parameter transformer model. The authors conduct experiments to validate the internal fine-tuning procedure of TinT on various tasks.", "review_text": "This paper proposes a new construction, Transformers in Transformer (TinT), that allows a transformer to simulate and fine-tune transformer models during inference. The authors introduce innovative approximation techniques that allow TinT with less than 2 Billion parameters to simulate and fine-tune a 125 million-parameter transformer model. The authors conduct experiments to validate the internal fine-tuning procedure of TinT on various tasks.", "strengths": "1. Using a transformer to perform in-context-learning to fine-tune a transformer sounds like a quite fancy idea. \n2. The authors performed experiments and demonstrated the effectiveness of TinT.", "weaknesses": "The writing of this paper is not completely clear, which makes it hard to understand the exact architecture of TinT. More specifically: \n1. In Figure 1, it is not clear what are the dimensions of V_k, e_i, \\partial y_j. Are there multiple back-propagation and gradient update steps in the TinT forward pass? \n2. What is the meaning of notation $\\partial$? There are many of them in Sections 2.5 and 2.6, but I don't understand the equations that contain this symbol. \n3. What are the parameters of the auxiliary transformer, and what are the trainable parameters of the TinT? \n4. I cannot see from the description of Section 2.7 PARAMETER SHARING IN THE TINT how the parameters of TinT are shared. \n5. Since the architecture is unclear, I don't see where the $5 \\times$, $H_{sim} \\times$, $4 \\times$ savings in Section 2.2 come from.", "questions": "1. Could the authors write down more concretely what is the architecture of the TinT? \n2. Could the authors explain their experimental setup more clearly?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new construction, Transformers in Transformer (TinT), that allows a transformer to simulate and fine-tune transformer models during inference. The authors introduce innovative approximation techniques that allow TinT with less than 2 Billion parameters to simulate and fine-tune a 125 million-parameter transformer model. The authors conduct experiments to validate the internal fine-tuning procedure of TinT on various tasks.", "soundness": "3 good", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1. Using a transformer to perform in-context-learning to fine-tune a transformer sounds like a quite fancy idea. \n2. The authors performed experiments and demonstrated the effectiveness of TinT.", "weaknesses": "The writing of this paper is not completely clear, which makes it hard to understand the exact architecture of TinT. More specifically: \n1. In Figure 1, it is not clear what are the dimensions of V_k, e_i, \\partial y_j. Are there multiple back-propagation and gradient update steps in the TinT forward pass? \n2. What is the meaning of notation $\\partial$? There are many of them in Sections 2.5 and 2.6, but I don't understand the equations that contain this symbol. \n3. What are the parameters of the auxiliary transformer, and what are the trainable parameters of the TinT? \n4. I cannot see from the description of Section 2.7 PARAMETER SHARING IN THE TINT how the parameters of TinT are shared. \n5. Since the architecture is unclear, I don't see where the $5 \\times$, $H_{sim} \\times$, $4 \\times$ savings in Section 2.2 come from.", "questions": "1. Could the authors write down more concretely what is the architecture of the TinT? \n2. Could the authors explain their experimental setup more clearly?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698889983735}, {"id": "jPrM5Nn0kL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4281/Reviewer_LDni"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper provides a construction for a large transformer as a simulator, called TINT, such that it can simulate (on the auxiliary model / the smaller model) the forward pass, do back propagation to update the parameters, and then simulate another forward pass to output the final results. To improve the parameter efficiency, they provided the construction to approximate the layer norm, self-attention matrix, and their back-propagation process. Prior work focused on doing a forward pass one step of gradient descent on linear models or some much easier models, but they showed that actually a larger Transformer can simulate the forward and backward pass for another smaller Transformer. They showed that a GPT2 model can serve as a simulator to train a 125M OPT model in the forward  pass (of larger model) and showed that the perplexity can be improved by 0.3-0.7 on average.\n\nThis idea is very creative and worth investigating in the future. The theory result seems good and solid, but I still have some questions about the motivation and the theory part (see below). It is possible for me to change my score based on author's response and other reviewers' opinion.", "review_text": "This paper provides a construction for a large transformer as a simulator, called TINT, such that it can simulate (on the auxiliary model / the smaller model) the forward pass, do back propagation to update the parameters, and then simulate another forward pass to output the final results. To improve the parameter efficiency, they provided the construction to approximate the layer norm, self-attention matrix, and their back-propagation process. Prior work focused on doing a forward pass one step of gradient descent on linear models or some much easier models, but they showed that actually a larger Transformer can simulate the forward and backward pass for another smaller Transformer. They showed that a GPT2 model can serve as a simulator to train a 125M OPT model in the forward  pass (of larger model) and showed that the perplexity can be improved by 0.3-0.7 on average.\n\nThis idea is very creative and worth investigating in the future. The theory result seems good and solid, but I still have some questions about the motivation and the theory part (see below). It is possible for me to change my score based on author's response and other reviewers' opinion.", "strengths": "1. The idea of doing forward ad backward pass of a Transformer in the single forward pass of another larger Transformer is creative. \n\n2. They spent lots of efforts on improving the parameter efficiency of the larger Transformer (the simulator) by approximation the derivatives  of softmax-attention layer and layer norm, which is very good. The theory looks solid and correct, and the way to approximate linear Transformers using non-linear Transformers (especially the soft-max Transformers) is a topic that may be of separate interest of some community.\n\n3. The experiments look good and there is a good improvement on the performance when trained this OPT model on larger GPT model.", "weaknesses": "1. The theory part, although the proof seems correct, it is better to have a theorem to include all results of approximating each part of a forward/backward pass of fine-tune a Transformer. For example, this should look like 'there is a transformer with XXX layers and XXX heads, and XXX dimensions such that, when you have an auxiliary model with XXX layers/heads/dimension and a sequence of tokens, it can approximation the objective function (something you want to approximate in the forward pass of the simulator) with an error of at most epsilon'. I think adding a result like this in your main result section (section 2) after the basic setup is essential.\n\nThe most important issue in the theorem above is that how many parameters you need in the simulator Transformers to approximate the in-context learn the fine-tuning objective of the auxiliary Transformer (with XX parameters). The proportion of the number of parameters (parameter size of big TF / param size of auxiliary one) matters since you claim that your construction is more parameter efficient than the construction in previous works [1,2]. \n\nBased on the proportion of parameter size, you can then determine that in the experiments, in order to train an OPT model with 125M params in a single forward pass of the simulator, what size model do you need to use as a simulator.\n\n2. The structure of the paper, I believe, is not optimal, since you put lots of details of definitions and the results of approximating each part of TF in the main text, which I believe can be postponed to the appendix. This makes the paper harder to follow. I believe the authors may need to include more motivations and high-level description in the main text, as well as the the 'global theorem' I suggested in the first point. Then, the concrete way to approximating the back-propagation of soft-max and some details can be deferred in the appendix. I think the 'key component' section is especially hard to follow and I think maybe it will be slightly better to illustrate the structure of simulators in more detail in this subsection, instead of providing some formal definitions (like 2.3,2.6,2.7) in the main text (these are well-known definitions after all).\n\n3. Why do you want to implement/approximate the forward/backward pass of Transformers in-context? \n\nMy understanding is that, you want to do something like 'in-context fine-tuning' for Transformers, because by doing this in-context, you do not need to 'do actual parameter updates' (in classical fine-tuning procedure, you need to literally 'update the parameters'). Then, the natural question is: which is the better and more efficient way for fine-tuning ----- in-context fine-tuning or direct fine-tuning? \n\nIn the paper, the authors did not show **whether it is more computationally efficient for in-context fine-tuning than direct fine-tuning**. By in-context fine-tuning, you need to do a forward pass in a much larger model; and by direct fine-tuning, you need to compute the gradient and update the parameters. Which one need a smaller number of computation (in terms of FLOPS or other metrics)? \n\n4. Another question for the motivation: why do you use Transformers as the simulators? Are there any obvious reasons that make Transformer better than other architecture of Neural Networks here? Since what you want to do is to compute the forward/backward pass of the auxiliary Transformer in a single forward pass of the simulator model, which is a very complex task, why not considering other types of models as the outer simulator? I know that in some tasks that Transformers or attention models showed a better in-context learnability than CNN/RNN, but the tasks they considered are very much different from yours, so I am not sure whether this reason is sufficient enough.\n\n[1] Ekin Akyurek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. arXiv preprint arXiv:2211.15661, 2022.\n[2] Johannes von Oswald, Eyvind Niklasson, Ettore Randazzo, Jo˜ao Sacramento, Alexander Mordvintsev, Andrey Zhmoginov, and Max Vladymyrov. Transformers learn in-context by gradient descent, 2022.", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper provides a construction for a large transformer as a simulator, called TINT, such that it can simulate (on the auxiliary model / the smaller model) the forward pass, do back propagation to update the parameters, and then simulate another forward pass to output the final results. To improve the parameter efficiency, they provided the construction to approximate the layer norm, self-attention matrix, and their back-propagation process. Prior work focused on doing a forward pass one step of gradient descent on linear models or some much easier models, but they showed that actually a larger Transformer can simulate the forward and backward pass for another smaller Transformer. They showed that a GPT2 model can serve as a simulator to train a 125M OPT model in the forward  pass (of larger model) and showed that the perplexity can be improved by 0.3-0.7 on average.\n\nThis idea is very creative and worth investigating in the future. The theory result seems good and solid, but I still have some questions about the motivation and the theory part (see below). It is possible for me to change my score based on author's response and other reviewers' opinion.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The idea of doing forward ad backward pass of a Transformer in the single forward pass of another larger Transformer is creative. \n\n2. They spent lots of efforts on improving the parameter efficiency of the larger Transformer (the simulator) by approximation the derivatives  of softmax-attention layer and layer norm, which is very good. The theory looks solid and correct, and the way to approximate linear Transformers using non-linear Transformers (especially the soft-max Transformers) is a topic that may be of separate interest of some community.\n\n3. The experiments look good and there is a good improvement on the performance when trained this OPT model on larger GPT model.", "weaknesses": "1. The theory part, although the proof seems correct, it is better to have a theorem to include all results of approximating each part of a forward/backward pass of fine-tune a Transformer. For example, this should look like 'there is a transformer with XXX layers and XXX heads, and XXX dimensions such that, when you have an auxiliary model with XXX layers/heads/dimension and a sequence of tokens, it can approximation the objective function (something you want to approximate in the forward pass of the simulator) with an error of at most epsilon'. I think adding a result like this in your main result section (section 2) after the basic setup is essential.\n\nThe most important issue in the theorem above is that how many parameters you need in the simulator Transformers to approximate the in-context learn the fine-tuning objective of the auxiliary Transformer (with XX parameters). The proportion of the number of parameters (parameter size of big TF / param size of auxiliary one) matters since you claim that your construction is more parameter efficient than the construction in previous works [1,2]. \n\nBased on the proportion of parameter size, you can then determine that in the experiments, in order to train an OPT model with 125M params in a single forward pass of the simulator, what size model do you need to use as a simulator.\n\n2. The structure of the paper, I believe, is not optimal, since you put lots of details of definitions and the results of approximating each part of TF in the main text, which I believe can be postponed to the appendix. This makes the paper harder to follow. I believe the authors may need to include more motivations and high-level description in the main text, as well as the the 'global theorem' I suggested in the first point. Then, the concrete way to approximating the back-propagation of soft-max and some details can be deferred in the appendix. I think the 'key component' section is especially hard to follow and I think maybe it will be slightly better to illustrate the structure of simulators in more detail in this subsection, instead of providing some formal definitions (like 2.3,2.6,2.7) in the main text (these are well-known definitions after all).\n\n3. Why do you want to implement/approximate the forward/backward pass of Transformers in-context? \n\nMy understanding is that, you want to do something like 'in-context fine-tuning' for Transformers, because by doing this in-context, you do not need to 'do actual parameter updates' (in classical fine-tuning procedure, you need to literally 'update the parameters'). Then, the natural question is: which is the better and more efficient way for fine-tuning ----- in-context fine-tuning or direct fine-tuning? \n\nIn the paper, the authors did not show **whether it is more computationally efficient for in-context fine-tuning than direct fine-tuning**. By in-context fine-tuning, you need to do a forward pass in a much larger model; and by direct fine-tuning, you need to compute the gradient and update the parameters. Which one need a smaller number of computation (in terms of FLOPS or other metrics)? \n\n4. Another question for the motivation: why do you use Transformers as the simulators? Are there any obvious reasons that make Transformer better than other architecture of Neural Networks here? Since what you want to do is to compute the forward/backward pass of the auxiliary Transformer in a single forward pass of the simulator model, which is a very complex task, why not considering other types of models as the outer simulator? I know that in some tasks that Transformers or attention models showed a better in-context learnability than CNN/RNN, but the tasks they considered are very much different from yours, so I am not sure whether this reason is sufficient enough.\n\n[1] Ekin Akyurek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. arXiv preprint arXiv:2211.15661, 2022.\n[2] Johannes von Oswald, Eyvind Niklasson, Ettore Randazzo, Jo˜ao Sacramento, Alexander Mordvintsev, Andrey Zhmoginov, and Max Vladymyrov. Transformers learn in-context by gradient descent, 2022.", "questions": "/", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698511308366}], "openreview_url": "https://openreview.net/forum?id=VmqTuFMk68", "arxiv_id": "2307.01189", "paper_pdf": "papers/VmqTuFMk68.pdf", "paper_pdf_sha256": "98ec028b07033358362f20d42a8140dfdfa00362f3773c28f8de52652c3832c1", "paper_pdf_bytes": 3692509, "paper_pdf_source": "openreview", "code_url": "https://github.com/abhishekpanigrahi1996/transformer_in_transformer", "code_repository": "abhishekpanigrahi1996/transformer_in_transformer", "code_commit": "1722081db59c1a9459941ec9e2c2f039fba9f331", "code_archive": "repos/VmqTuFMk68.zip", "code_archive_sha256": "5121d3b818a7748a5094d969bb0fc6da3d10596fd5a03d42b0ce550559403ea6", "code_archive_bytes": 321987, "code_file_count": 40, "code_extensions": {".py": 39, ".sh": 1}, "github_disk_usage_kb": 315, "github_languages": {"Python": 500283, "Shell": 3505}, "github_archived": false, "github_pushed_at": "2023-10-11T17:55:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/trainable-transformer-in-transformer"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QEmn_Hvh7j8", "year": 2023, "status": "rejected", "title": "Private GANs, Revisited", "authors": ["Alex Bie", "Gautam Kamath", "Guojun Zhang"], "authorids": ["~Alex_Bie1", "~Gautam_Kamath1", "~Guojun_Zhang1"], "authors_source": "OpenReview API", "abstract": "We show that with improved training, the standard approach for differentially private GANs -- updating the discriminator with noisy gradients -- achieves or competes with state-of-the-art results for private image synthesis. Existing instantiations of this approach neglect to consider how adding noise only to discriminator updates disrupts the careful balance between generator and discriminator necessary for successful GAN training. We show that a simple fix restores parity: taking more discriminator steps between generator steps. Finally, with the goal of restoring parity between generator and discriminator, we experiment with further modifications to improve discriminator training and see further improvements. For MNIST at $\\eps=10$, our private GANs improve the record FID from 48.4 to 13.0, as well as downstream classifier accuracy from 83.2\\% to 95.0\\%.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ulppKDiNDo", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper621/Reviewer_wVny"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a simple strategy to improve DP-GANs by using more update steps for discriminator, adopting a step scheduler, and taking a larger batch size. The authors provide empirical findings to justify such improvement by linking generation quality with the discriminator accuracy though the training process. Experiments on MNIST and FashionMNIST show very promising results in terms of generation quality.", "review_text": "This paper has interesting empirical findings and promising results, while at current stage the proposed solution is not mature enough to meet the standard of ICLR. However, along its direction, this work is potential to have an impact in the future, probably with an elaborated method or some theoretical support.", "strengths": "Strength:\n\n1. The proposed techniques are very effective to achieve much better FID on MNIST and FashionMNIST.\n\n2. The empirical explanation of balancing the discriminator and maximizing generator steps taken when discriminator accuracy is high, is interesting and inspirational.\n\nWeaknesses:\n\n1. Overall, I think the proposed method is currently limited to an engineering technique, which might not be mature enough. However, I do believe it is highly potential to develop a principled method with high impact along the current direction. For example, I would recommend further elaborate on the step scheduler to be more adaptive and include more ablation studies. The authors might be able to draw some inspirations from StyleGAN2-ADA for such paradigm.\n\n2. The improvement of utility accuracy is not as significant as FID. It would also be better to include an experiment on some larger datasets such as CelebA.\n\n3. The writing and the organization could be improved.\n- Although fancy, the title is not very informative.\n- Figure 1 and figure 2 share the same purpose.\n- Section 3.3 is duplicate to the previous content.\n- Section 5.1 and 5.2 are not well aligned with the theme of Section 5 and the results are not well linked to the discriminator accuracy.\n- In Algorithm 1, some symbols are used without definition.\n- In Table 1, \"(This work)\" is ambiguous.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a simple strategy to improve DP-GANs by using more update steps for discriminator, adopting a step scheduler, and taking a larger batch size. The authors provide empirical findings to justify such improvement by linking generation quality with the discriminator accuracy though the training process. Experiments on MNIST and FashionMNIST show very promising results in terms of generation quality.", "strength_and_weaknesses": "Strength:\n\n1. The proposed techniques are very effective to achieve much better FID on MNIST and FashionMNIST.\n\n2. The empirical explanation of balancing the discriminator and maximizing generator steps taken when discriminator accuracy is high, is interesting and inspirational.\n\nWeaknesses:\n\n1. Overall, I think the proposed method is currently limited to an engineering technique, which might not be mature enough. However, I do believe it is highly potential to develop a principled method with high impact along the current direction. For example, I would recommend further elaborate on the step scheduler to be more adaptive and include more ablation studies. The authors might be able to draw some inspirations from StyleGAN2-ADA for such paradigm.\n\n2. The improvement of utility accuracy is not as significant as FID. It would also be better to include an experiment on some larger datasets such as CelebA.\n\n3. The writing and the organization could be improved.\n- Although fancy, the title is not very informative.\n- Figure 1 and figure 2 share the same purpose.\n- Section 3.3 is duplicate to the previous content.\n- Section 5.1 and 5.2 are not well aligned with the theme of Section 5 and the results are not well linked to the discriminator accuracy.\n- In Algorithm 1, some symbols are used without definition.\n- In Table 1, \"(This work)\" is ambiguous.", "clarity,_quality,_novelty_and_reproducibility": "The clarity and quality are good in general with some issues to be improved. \nThis paper is novel in terms of empirical findings, while the proposed method is currently limited to an engineering technique.\nThe reproducibility is good with sufficient implementation details provided.", "summary_of_the_review": "This paper has interesting empirical findings and promising results, while at current stage the proposed solution is not mature enough to meet the standard of ICLR. However, along its direction, this work is potential to have an impact in the future, probably with an elaborated method or some theoretical support.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667118114228}, {"id": "DfGIsGAbPO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper621/Reviewer_dASH"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides two empirical findings on how to train differential private gans: larger batch size, and more discriminator steps. Experimental results show superior performance against existing baselines.", "review_text": "Overall, this paper is a practical guide for training private gans. However the findings are empirical and rather trivial. In addition, some details are missing.", "strengths": "Strength\n\n1. In the literature of differential privacy gans, it seems like the utilization of more recent architectures and training tricks (e.g. StyleGAN2,3, …) for higher quality samples are mostly unexplored. This paper provides a good point that these “tricks” could largely boost the performance of private gans.\n\nWeakness\n\n1. From my perspective, this work is with limited novelty. Since larger batch size and take more discriminator steps has been studied widely in normal gan papers. It seems like natural attempts to try these tricks on private gans.\n2. Several missing details. For example, how does (10, 10^-5)-DP calculated from B=128,sigma=1,C=1,T=450000 (in section 3.1)? Please provide the formal equation on calculating this. In addition, noise scales, n_D, and batch size also affect privacy, in section 5.1, how can you targeting the same \\epsilon with different values of these?", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper provides two empirical findings on how to train differential private gans: larger batch size, and more discriminator steps. Experimental results show superior performance against existing baselines.", "strength_and_weaknesses": "Strength\n\n1. In the literature of differential privacy gans, it seems like the utilization of more recent architectures and training tricks (e.g. StyleGAN2,3, …) for higher quality samples are mostly unexplored. This paper provides a good point that these “tricks” could largely boost the performance of private gans.\n\nWeakness\n\n1. From my perspective, this work is with limited novelty. Since larger batch size and take more discriminator steps has been studied widely in normal gan papers. It seems like natural attempts to try these tricks on private gans.\n2. Several missing details. For example, how does (10, 10^-5)-DP calculated from B=128,sigma=1,C=1,T=450000 (in section 3.1)? Please provide the formal equation on calculating this. In addition, noise scales, n_D, and batch size also affect privacy, in section 5.1, how can you targeting the same \\epsilon with different values of these?", "clarity,_quality,_novelty_and_reproducibility": "Some details are missing, code is not provided.", "summary_of_the_review": "Overall, this paper is a practical guide for training private gans. However the findings are empirical and rather trivial. In addition, some details are missing.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666677126188}, {"id": "IKSvXe-aRgK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper621/Reviewer_hSom"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors revisit the DPGAN paper and state that the added noise to discriminator training disrupts the balance between the generator and discriminator training. They show that by tuning the number of update steps taking by the discriminator for every generator update step, specifically, taking more steps significantly improves the results. They show that for MNIST \\at epsilon = 10 their private GAN FID goes from 48.4  to 13.0 and the downstream accuracy of the classifier goes up from 83.2% to 95.0%. ", "review_text": "The paper present a very slight modification, a focus on a parameter change in the DPGAN method. It propose to increase the discriminator update frequency when noise is introduced to the discriminator (as in the DPGAN training case). The paper discuss the motivation to their method at length. Finally, the authors show inconclusive experiments, where the method works for some regime of the noise level. ", "strengths": "### Strength\nThe paper provides a simple yet effective method to improve the quality of the generated data of DPGAN. The experiments show significant improvement in the tested benchmarks. The authors discuss in length about whey more steps are needed in the case of DPGAN. In addition, the authors discuss the role of the batch size in DPGAN settings and the additional benefits of scheduling the discriminator frequency update (start low then increase). \n\n### Weaknesses\nThe novelty of the work is marginal. The method work best at \\epsilon = 10, but on high privacy/ low \\epsilon regime sometime the results are even worse.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors revisit the DPGAN paper and state that the added noise to discriminator training disrupts the balance between the generator and discriminator training. They show that by tuning the number of update steps taking by the discriminator for every generator update step, specifically, taking more steps significantly improves the results. They show that for MNIST \\at epsilon = 10 their private GAN FID goes from 48.4  to 13.0 and the downstream accuracy of the classifier goes up from 83.2% to 95.0%. ", "strength_and_weaknesses": "### Strength\nThe paper provides a simple yet effective method to improve the quality of the generated data of DPGAN. The experiments show significant improvement in the tested benchmarks. The authors discuss in length about whey more steps are needed in the case of DPGAN. In addition, the authors discuss the role of the batch size in DPGAN settings and the additional benefits of scheduling the discriminator frequency update (start low then increase). \n\n### Weaknesses\nThe novelty of the work is marginal. The method work best at \\epsilon = 10, but on high privacy/ low \\epsilon regime sometime the results are even worse.", "clarity,_quality,_novelty_and_reproducibility": "The paper is written in a clear and easy to follow fashion, the work might be important to the community although the novelty of the work is not substantial. The work could be easily be reproduce.", "summary_of_the_review": "The paper present a very slight modification, a focus on a parameter change in the DPGAN method. It propose to increase the discriminator update frequency when noise is introduced to the discriminator (as in the DPGAN training case). The paper discuss the motivation to their method at length. Finally, the authors show inconclusive experiments, where the method works for some regime of the noise level. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "no ethics concerns", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666190265325}], "openreview_url": "https://openreview.net/forum?id=QEmn_Hvh7j8", "arxiv_id": "2302.02936", "paper_pdf": "papers/QEmn_Hvh7j8.pdf", "paper_pdf_sha256": "ea524c63ea455d0332e874d3b67c64a46d65b941288c9ee806dd1e5f1e177396", "paper_pdf_bytes": 3233598, "paper_pdf_source": "openreview", "code_url": "https://github.com/alexbie98/dpgan-revisit", "code_repository": "alexbie98/dpgan-revisit", "code_commit": "2233d9bc7f787e19750fc83c3a5944b43e25a9df", "code_archive": "repos/QEmn_Hvh7j8.zip", "code_archive_sha256": "3b3a4d5f98f2bb52c32b2b3accc0380ba68ccc652ae1997de502dd68da3b9f46", "code_archive_bytes": 489139, "code_file_count": 29, "code_extensions": {".py": 29}, "github_disk_usage_kb": 710, "github_languages": {"Python": 117435}, "github_archived": false, "github_pushed_at": "2023-10-05T14:00:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/private-gans-revisited"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "2PSrjVtj6gU", "year": 2022, "status": "rejected", "title": "Graph Attention Multi-layer Perceptron", "authors": ["Wentao Zhang", "Ziqi Yin", "Zeang Sheng", "Yang Li", "Wen Ouyang", "Xiaosen Li", "Yangyu Tao", "Zhi Yang", "Bin CUI"], "authorids": ["~Wentao_Zhang1", "~Ziqi_Yin1", "~Zeang_Sheng1", "~Yang_Li36", "~Wen_Ouyang1", "~Xiaosen_Li1", "~Yangyu_Tao2", "~Zhi_Yang4", "~Bin_CUI2"], "authors_source": "OpenReview API", "abstract": "Recently, graph neural networks (GNNs) have achieved a stride of success in many graph-based applications. However, most GNNs suffer from a critical issue: representation learned is constructed based on a fixed k-hop neighborhood and insensitive to individual needs for each node, which greatly hampers the performance of GNNs. To satisfy the unique needs of each node, we propose a new architecture -- Graph Attention Multi-Layer Perceptron (GAMLP). This architecture combines multi-scale knowledge and learns to capture the underlying correlations between different scales of knowledge with two novel attention mechanisms: Recursive attention and Jumping Knowledge (JK) attention. Instead of using node feature only, the knowledge within node labels is also exploited to reinforce the performance of GAMLP. Extensive experiments on 12 real-world datasets demonstrate that GAMLP achieves state-of-the-art performance while enjoying high scalability and efficiency.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ffj2E3RIJL1", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper610/Reviewer_Z4vM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposed the Graph Attention Multi-Layer Perceptron (GAMLP) model that combines features extracted from two streams: 1) node-adaptive features that are obtained via \"Graph-wise / layer-wise propagation\" (Section 2.2), 2) features obtained from label propagation. The author also proposed two ways for performing the node-adaptive attention: JK attention, and Recursive attention (Section 3.3). Extensive experiments are conducted on 9 transductive datasets and 3 inductive datasets. The results show that the proposed GAMLP(JK) and GAMLP(R) can outperform the baselines.\n", "review_text": "- Overall Novelty\n\nCombining layer-wise propagation and label propagation is a good idea. However, it is not clear what's new in GAMLP except ensembling two types of models. Fusing these two streams do not look very challenging and it is foreseeable that the model can achieve competitive performance as GBP (based on layer-wise propagation) or UniMP (based on label propagation). The proposed \"JK attention\" and \"Recursive attention\" are also not very novel and are largely borrowed from previous works like JK-Net.\n\n- Technical Quality + Writing\n\nThe technical quality of the paper is good. The author summarized the current landscape of scalable GNNs and categorized them as 1) sampling-based, 2) graph-wise propagation and 3) layer-wise propagation. The author also provided the reference to prior works such as label propagation and JK-Net. Also, the author clearly stated the general algorithm in Figure 2 and the detailed design in Section 3. However, the author can still improve the paper by providing insights on some technical designs, e.g.\n\n    1. Why choose the `cos(...)` formula in designing the \"Last Residual Connection\" module?\n    2. What's the fundamental differences between JK attention and Recursive attention? Why would you propose two variants?\n\n- Experiments\n\nThe author conducted extensive experiments in 9 transductive datasets and 3 inductive datasets. However, I find that the baseline selection is not consistent across the experiments. For example, GBP is not compared in Table 2. UniMP is not compared in Table 1.\n\nIn addition, the trend of validation and test accuracy reported in Table 5 is not consistent (models with higher validation accuracy tend to have lower test accuracy). The author needs to clarify the inconsistency.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposed the Graph Attention Multi-Layer Perceptron (GAMLP) model that combines features extracted from two streams: 1) node-adaptive features that are obtained via \"Graph-wise / layer-wise propagation\" (Section 2.2), 2) features obtained from label propagation. The author also proposed two ways for performing the node-adaptive attention: JK attention, and Recursive attention (Section 3.3). Extensive experiments are conducted on 9 transductive datasets and 3 inductive datasets. The results show that the proposed GAMLP(JK) and GAMLP(R) can outperform the baselines.\n", "main_review": "- Overall Novelty\n\nCombining layer-wise propagation and label propagation is a good idea. However, it is not clear what's new in GAMLP except ensembling two types of models. Fusing these two streams do not look very challenging and it is foreseeable that the model can achieve competitive performance as GBP (based on layer-wise propagation) or UniMP (based on label propagation). The proposed \"JK attention\" and \"Recursive attention\" are also not very novel and are largely borrowed from previous works like JK-Net.\n\n- Technical Quality + Writing\n\nThe technical quality of the paper is good. The author summarized the current landscape of scalable GNNs and categorized them as 1) sampling-based, 2) graph-wise propagation and 3) layer-wise propagation. The author also provided the reference to prior works such as label propagation and JK-Net. Also, the author clearly stated the general algorithm in Figure 2 and the detailed design in Section 3. However, the author can still improve the paper by providing insights on some technical designs, e.g.\n\n    1. Why choose the `cos(...)` formula in designing the \"Last Residual Connection\" module?\n    2. What's the fundamental differences between JK attention and Recursive attention? Why would you propose two variants?\n\n- Experiments\n\nThe author conducted extensive experiments in 9 transductive datasets and 3 inductive datasets. However, I find that the baseline selection is not consistent across the experiments. For example, GBP is not compared in Table 2. UniMP is not compared in Table 1.\n\nIn addition, the trend of validation and test accuracy reported in Table 5 is not consistent (models with higher validation accuracy tend to have lower test accuracy). The author needs to clarify the inconsistency.", "summary_of_the_review": "The general idea of GAMLP is neat and reasonable. However, I'm hesitant to accept the paper because I haven't seen a contribution that is clearly novel. In addition, there are some inconsistencies in the experiments.\n\n——— Post Rebuttal ———\nThe author has addressed my concerns regarding the experiments and the motivation of proposing two attention mechanisms (recursive attention v.s JK attention). Given the nice results and scalability of GAMLP, I’ve increased my score to 6.  However, I’m still concerned about the novelty of the paper because it proposed several ideas but none of them is really novel. The node-wise feature propagation idea looks like a direct extension of the previous layer-wise / graph-wise propagation mechanisms. The major difference is that the node-wise propagation will use the propagated features from multiple depths. Thus, I haven’t increased my score to accept due to lack of novelty. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635894204988}, {"id": "mIeM2aWXISv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper610/Reviewer_eUi3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes the scalable GAMLP architecture that can learn node-adaptive features and can utilize label propagation to improve the model performance. Comprehensive experiments are provided to demonstrate the effectiveness of GAMLP, discuss the importance of each component in GAMLP, and compare its efficiency against some classic and SOTA baseline models.", "review_text": "Strengths:\n1. The writing of this paper is very clear, the organization is easy to follow and the main ideas are delivered clearly.\n2. I appreciate the comprehensive experiments. \n3. The model performance is impressive. \n\n===\n\nWeakness：\n1. I feel there are some similarities between this paper and the work of SLE [1]. According to my understanding, both papers try to improve the SIGN model [2] by introducing a node-wise attention mechanism and utilizing label propagation to further enhance the performance. I am wondering, what is the key difference between GAMLP and SLE, and what are the advantages of GAMLP compared to SLE?\n2. I am also interested in the performance comparison between GAMLP and SLE.\n3. For the ablation study (section 4.3) Q2 (2), the authors compared the attention mechanism with different reference vectors. However, I am also interested to see how can the proposed attention mechanism affect performance?  For example, what would happen if each Wk in eq(3) is a learnable scaler? or what would happen if directly concatenate [X^(0)...X^(k)]?\n\n===\n\n[1] Sun, Chuxiong, and Guoshi Wu. \"Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training.\" arXiv preprint arXiv:2104.09376 (2021).\n[2] Frasca, Fabrizio, et al. \"Sign: Scalable inception graph neural networks.\" arXiv preprint arXiv:2004.11198 (2020).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes the scalable GAMLP architecture that can learn node-adaptive features and can utilize label propagation to improve the model performance. Comprehensive experiments are provided to demonstrate the effectiveness of GAMLP, discuss the importance of each component in GAMLP, and compare its efficiency against some classic and SOTA baseline models.", "main_review": "Strengths:\n1. The writing of this paper is very clear, the organization is easy to follow and the main ideas are delivered clearly.\n2. I appreciate the comprehensive experiments. \n3. The model performance is impressive. \n\n===\n\nWeakness：\n1. I feel there are some similarities between this paper and the work of SLE [1]. According to my understanding, both papers try to improve the SIGN model [2] by introducing a node-wise attention mechanism and utilizing label propagation to further enhance the performance. I am wondering, what is the key difference between GAMLP and SLE, and what are the advantages of GAMLP compared to SLE?\n2. I am also interested in the performance comparison between GAMLP and SLE.\n3. For the ablation study (section 4.3) Q2 (2), the authors compared the attention mechanism with different reference vectors. However, I am also interested to see how can the proposed attention mechanism affect performance?  For example, what would happen if each Wk in eq(3) is a learnable scaler? or what would happen if directly concatenate [X^(0)...X^(k)]?\n\n===\n\n[1] Sun, Chuxiong, and Guoshi Wu. \"Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training.\" arXiv preprint arXiv:2104.09376 (2021).\n[2] Frasca, Fabrizio, et al. \"Sign: Scalable inception graph neural networks.\" arXiv preprint arXiv:2004.11198 (2020).", "summary_of_the_review": "My major concern is the novelty. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635893539883}, {"id": "L0Sjj8yAXpF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper610/Reviewer_Cwzz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes an efficient method for inductive/transductive node labeling based on fast feature propagation, label propagation, and attention and yields good experimental results.", "review_text": "The paper introduces an efficient way to perform node labeling tasks on inductive and transductive setting. It achieves good results on the widely used benchmarks. Here are the questions / concerns:\n\n1- In some parts it is difficult to follow the paper. And at some points the information is missing. e.g., definition of T in eq(2) and s in eq (7). eq (7) and (8) seems to be identical. The dimensions are not explicitly mentioned which could have made the paper more clear.\n\n2- In terms of contribution, the paper seems to have incremental contributions with respect to previous work but not a significant one. Using a weighted average over the features from different layers seems very similar to attention graph pooling and also feature propagation is not different from previous work. Also using label propagation has been previously explored.\n\n3- The other potential limitation of the proposed model is that it is tailored to supervised setting and won't be able to perform in contrastive setting with linear evaluation protocol.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes an efficient method for inductive/transductive node labeling based on fast feature propagation, label propagation, and attention and yields good experimental results.", "main_review": "The paper introduces an efficient way to perform node labeling tasks on inductive and transductive setting. It achieves good results on the widely used benchmarks. Here are the questions / concerns:\n\n1- In some parts it is difficult to follow the paper. And at some points the information is missing. e.g., definition of T in eq(2) and s in eq (7). eq (7) and (8) seems to be identical. The dimensions are not explicitly mentioned which could have made the paper more clear.\n\n2- In terms of contribution, the paper seems to have incremental contributions with respect to previous work but not a significant one. Using a weighted average over the features from different layers seems very similar to attention graph pooling and also feature propagation is not different from previous work. Also using label propagation has been previously explored.\n\n3- The other potential limitation of the proposed model is that it is tailored to supervised setting and won't be able to perform in contrastive setting with linear evaluation protocol.", "summary_of_the_review": "The paper introduces some incremental contributions with empirical support. But at the end of day, the contributions are slight variations of previous work. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635712435249}, {"id": "WDrOa_gIMjF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper610/Reviewer_XqaP"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a simplified graph attention model that operates in two stages: first label propagation of the labels and features and then learning MLP on obtained predictions. The experiments show improvement on 12 graph classification datasets. ", "review_text": "The paper proposes an end-to-end model that first propagates separately node features and labels, combines them with attention weights, aggregates embeddings and trains MLP model. For node-adaptive attention mechanism the authors propose two schemes: Recursive attention and JK attention. Experiments clearly show state-of-the-art performance. \n\nDespite its advantages, I see the following limitations of this work: \n1) Limited novelty. All the blocks used in the model were presented in the past and showed superior performance. Hence it's expected to see superior performance when combining those. \n\n2. Figure 1 used to discuss limitations of the existing approaches is not plotted for the proposed approach. Does GAMLP solves the aforementioned limitations? \n\n3.Cora/citeseer/pubmed datasets are quite small to motivate such an approach. While there are OGB datasets, all of these data are homophily-based. It would be good to compare the model on heterophily-based data [1] (where LP may fail) and on data with heterogeneous node features [2] (where graph-agnostic MLP and GBM can lead to better results). It's also interesting to compare with solely MLP baseline and GBDT baseline [2]\n\n[1] New Benchmarks for Learning on Non-Homophilous Graphs https://arxiv.org/abs/2104.01404\n\n[2] Boost then Convolve: Gradient Boosting Meets Graph Neural Networks https://arxiv.org/abs/2101.08543 \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a simplified graph attention model that operates in two stages: first label propagation of the labels and features and then learning MLP on obtained predictions. The experiments show improvement on 12 graph classification datasets. ", "main_review": "The paper proposes an end-to-end model that first propagates separately node features and labels, combines them with attention weights, aggregates embeddings and trains MLP model. For node-adaptive attention mechanism the authors propose two schemes: Recursive attention and JK attention. Experiments clearly show state-of-the-art performance. \n\nDespite its advantages, I see the following limitations of this work: \n1) Limited novelty. All the blocks used in the model were presented in the past and showed superior performance. Hence it's expected to see superior performance when combining those. \n\n2. Figure 1 used to discuss limitations of the existing approaches is not plotted for the proposed approach. Does GAMLP solves the aforementioned limitations? \n\n3.Cora/citeseer/pubmed datasets are quite small to motivate such an approach. While there are OGB datasets, all of these data are homophily-based. It would be good to compare the model on heterophily-based data [1] (where LP may fail) and on data with heterogeneous node features [2] (where graph-agnostic MLP and GBM can lead to better results). It's also interesting to compare with solely MLP baseline and GBDT baseline [2]\n\n[1] New Benchmarks for Learning on Non-Homophilous Graphs https://arxiv.org/abs/2104.01404\n\n[2] Boost then Convolve: Gradient Boosting Meets Graph Neural Networks https://arxiv.org/abs/2101.08543 \n\n", "summary_of_the_review": "The paper is well-written and attains good performance metrics; however, contributions are limited and are composed of previously studied ideas. To improve the paper I would seek a good application of such model besides the standard node classification (e.g. such applications that require simplicity of ML models). ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635528201416}], "openreview_url": "https://openreview.net/forum?id=2PSrjVtj6gU", "arxiv_id": "2206.04355", "paper_pdf": "papers/2PSrjVtj6gU.pdf", "paper_pdf_sha256": "f0060f3ef66ee90a5ddf5e29fd6c06411e50d4e370c739f7482748e6af33faa0", "paper_pdf_bytes": 700295, "paper_pdf_source": "openreview", "code_url": "https://github.com/PKU-DAIR/GAMLP", "code_repository": "PKU-DAIR/GAMLP", "code_commit": "83dc5d4a414c1829ed4a793410d3bd6037ee89ee", "code_archive": "repos/2PSrjVtj6gU.zip", "code_archive_sha256": "cc28d0919c20311ff05a335c267ee9523e5b8a83df9ea94be48f684d4f5e30be", "code_archive_bytes": 1183856, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 3027, "github_languages": {"Python": 59499, "Shell": 1028}, "github_archived": false, "github_pushed_at": "2022-06-25T09:04:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/graph-attention-multi-layer-perceptron-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nCY83KxoehA", "year": 2021, "status": "rejected", "title": "Automated Concatenation of Embeddings for Structured Prediction", "authors": ["Xinyu Wang", "Yong Jiang", "Nguyen Bach", "Tao Wang", "Zhongqiang Huang", "Fei Huang", "Kewei Tu"], "authorids": ["~Xinyu_Wang3", "~Yong_Jiang1", "~Nguyen_Bach1", "~Tao_Wang4", "~Zhongqiang_Huang1", "~Fei_Huang2", "~Kewei_Tu1"], "authors_source": "OpenReview API", "abstract": "Pretrained contextualized embeddings are powerful word representations for structured prediction tasks. Recent work found that better word representations can be obtained by concatenating different types of embeddings. However, the selection of embeddings to form the best concatenated representation usually varies depending on the task and the collection of candidate embeddings, and the ever-increasing number of embedding types makes it a more difficult problem. In this paper, we propose Automated Concatenation of Embeddings (ACE) to automate the process of finding better concatenations of embeddings for structured prediction tasks, based on a formulation inspired by recent progress on neural architecture search. Specifically, a controller alternately samples a concatenation of embeddings, according to its current belief of the effectiveness of individual embedding types in consideration for a task, and updates the belief based on a reward. We follow strategies in reinforcement learning to optimize the parameters of the controller and compute the reward based on the accuracy of a task model, which is fed with the sampled concatenation as input and trained on a task dataset. Empirical results on 6 tasks and 21 datasets show that our approach outperforms strong baselines and achieves state-of-the-art performance with fine-tuned embeddings in the vast majority of evaluations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "aycUbSu26Y", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1096/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Updates after discussion/revision period:\n\nI think the revisions have improved the paper, but I'm not willing to increase my score or to fight for the paper. Overall, I think the paper represents a minor contribution, with its rigorous experimentation and some of its ideas, and that others may benefit from reading it, but I don't know that it is at the level of a typical ICLR publication. \n\n\n--------\n\nThis paper describes an approach to choosing a subset of several options for (optionally contextualized) word embeddings to use in NLP tasks. Ideas are drawn from neural architecture search (NAS) and RL. The basic idea is to maximize dev set accuracy by searching over the space of embedding sets to use. There are some tweaks, including avoiding retraining-from-scratch for each set by keeping a single generalized model with all embeddings in it (where subsets can be chosen by setting some matrices to zero). Experiments are done on many NLP tasks (tagging, chunking, NER, parsing, etc.), leading to state-of-the-art results on nearly all test sets. \n\nThis paper represents an impressive number of experiments, considering several types of embeddings and many NLP tasks/datasets. It is well-written on the whole, though there are a few things I had confusion or concern about (details below). I lean positive on this paper, as it has an interesting algorithm that is more practical than prior work in NAS and has some promising results. However, I also have a few high-level concerns, described below:\n\n1. I'm not sure if I'm thoroughly convinced of the empirical superiority of ACE. The primary baselines are All (using all embeddings always) and Random (random search over subsets of embeddings). Random is a little better than All on average, and ACE is a little better than Random on average. The average difference of 0.5 in Table 1 between ACE and Random is largely due to the 8 aspect extraction (AE) datasets, for which the differences are sometimes sizable. However, across the 17 other results in the table (tagging, NER, chunking, and parsing), none differ by more than 0.4, and the average difference between Random and ACE on those other 17 numbers is 0.17 (computed by me). Possibly statistically significant, especially because there are so many different datasets, but less impressive. If one were to deploy a method like this in practice, one would likely start by trying random search because it's so simple and doesn't require the slightly specialized learning framework and reward function in ACE.  \n\nRelatedly, I am concerned that the All baseline is not strong enough. Another natural baseline would be to start with All but then add a_l parameters as gates on the input embeddings, just as they are present in ACE. (The a_l parameters could be normalized to be between 0 and 1 by passing them each through a sigmoid before being multiplied with embedding vectors.) By having a single parameter to weight each embedding type in this way, the new version of All could switch on or off entire embedding types without adding many more parameters, which would make it more similar to the other methods. This would let us see the results of this stronger version of All (which, I would argue, is more likely to be used in practice than the current version of All). \n\n2. My second concern is about the following sentence in Sec. 4.2: \"If the pretrained contextualized embeddings are not available for a particular language, we use the pretrained contextualized embeddings for English instead.\" I find this to be a rather surprising decision, as it could add a great deal of noise for non-English languages. This could be especially problematic for the All baseline which doesn't have an easy way to switch off a noise type of embedding. It would be nice to know for which tasks/datasets this English embedding replacement was done in practice. I looked at Appendix A.4, but I wasn't able to determine from that section which embedding types were missing for which datasets. \n\n3. CoNLL 2003 does not contain gold chunk labels. It contains automatic chunk labels (as can be confirmed by checking the original paper). CoNLL 2003 should not be used for chunking experiments. Unfortunately this mistake has been repeated in many papers. Please remove all CoNLL 2003 chunking experiments. \n\nSome additional (less major) questions are below:\n\nIn Appendix B.2, why does ACE work better than retraining? I wouldn't have expected this to happen. \n\nIn Sec. 3.1, it's odd that the BiLSTM-CRF and BiLSTM-Biaffine functions only take in V as the only argument, where V is a function solely of the input x, not of the output y. Why is it not a function of y as well?\n\nIn Sec. 3.2, I find the phrasing \"concatenation with the mask\" to be a bit confusing. I don't think the mask is being concatenated; I think it's being multiplied elementwise with the embeddings. \n\nRight above Eq. (7), there is the text \"Taking m = 1\" -- what is m?\n\nIn Sec. 4.2: What exactly is meant by \"character embeddings\"? There are many ways to embed words using characters. How are the character embeddings composed to form a word embedding?\n\nHow was the random search done? I see the sentence \"For Random, we use the same training settings as our approach\", which makes me assume there were 30 steps, but I think this should be made more explicit. Since the random approach and ACE are different algorithms with different hyperparameters, it's not clear to me what is meant by using \"the same training settings\". \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "reasonably interesting, though some concerns remain", "review": "Updates after discussion/revision period:\n\nI think the revisions have improved the paper, but I'm not willing to increase my score or to fight for the paper. Overall, I think the paper represents a minor contribution, with its rigorous experimentation and some of its ideas, and that others may benefit from reading it, but I don't know that it is at the level of a typical ICLR publication. \n\n\n--------\n\nThis paper describes an approach to choosing a subset of several options for (optionally contextualized) word embeddings to use in NLP tasks. Ideas are drawn from neural architecture search (NAS) and RL. The basic idea is to maximize dev set accuracy by searching over the space of embedding sets to use. There are some tweaks, including avoiding retraining-from-scratch for each set by keeping a single generalized model with all embeddings in it (where subsets can be chosen by setting some matrices to zero). Experiments are done on many NLP tasks (tagging, chunking, NER, parsing, etc.), leading to state-of-the-art results on nearly all test sets. \n\nThis paper represents an impressive number of experiments, considering several types of embeddings and many NLP tasks/datasets. It is well-written on the whole, though there are a few things I had confusion or concern about (details below). I lean positive on this paper, as it has an interesting algorithm that is more practical than prior work in NAS and has some promising results. However, I also have a few high-level concerns, described below:\n\n1. I'm not sure if I'm thoroughly convinced of the empirical superiority of ACE. The primary baselines are All (using all embeddings always) and Random (random search over subsets of embeddings). Random is a little better than All on average, and ACE is a little better than Random on average. The average difference of 0.5 in Table 1 between ACE and Random is largely due to the 8 aspect extraction (AE) datasets, for which the differences are sometimes sizable. However, across the 17 other results in the table (tagging, NER, chunking, and parsing), none differ by more than 0.4, and the average difference between Random and ACE on those other 17 numbers is 0.17 (computed by me). Possibly statistically significant, especially because there are so many different datasets, but less impressive. If one were to deploy a method like this in practice, one would likely start by trying random search because it's so simple and doesn't require the slightly specialized learning framework and reward function in ACE.  \n\nRelatedly, I am concerned that the All baseline is not strong enough. Another natural baseline would be to start with All but then add a_l parameters as gates on the input embeddings, just as they are present in ACE. (The a_l parameters could be normalized to be between 0 and 1 by passing them each through a sigmoid before being multiplied with embedding vectors.) By having a single parameter to weight each embedding type in this way, the new version of All could switch on or off entire embedding types without adding many more parameters, which would make it more similar to the other methods. This would let us see the results of this stronger version of All (which, I would argue, is more likely to be used in practice than the current version of All). \n\n2. My second concern is about the following sentence in Sec. 4.2: \"If the pretrained contextualized embeddings are not available for a particular language, we use the pretrained contextualized embeddings for English instead.\" I find this to be a rather surprising decision, as it could add a great deal of noise for non-English languages. This could be especially problematic for the All baseline which doesn't have an easy way to switch off a noise type of embedding. It would be nice to know for which tasks/datasets this English embedding replacement was done in practice. I looked at Appendix A.4, but I wasn't able to determine from that section which embedding types were missing for which datasets. \n\n3. CoNLL 2003 does not contain gold chunk labels. It contains automatic chunk labels (as can be confirmed by checking the original paper). CoNLL 2003 should not be used for chunking experiments. Unfortunately this mistake has been repeated in many papers. Please remove all CoNLL 2003 chunking experiments. \n\nSome additional (less major) questions are below:\n\nIn Appendix B.2, why does ACE work better than retraining? I wouldn't have expected this to happen. \n\nIn Sec. 3.1, it's odd that the BiLSTM-CRF and BiLSTM-Biaffine functions only take in V as the only argument, where V is a function solely of the input x, not of the output y. Why is it not a function of y as well?\n\nIn Sec. 3.2, I find the phrasing \"concatenation with the mask\" to be a bit confusing. I don't think the mask is being concatenated; I think it's being multiplied elementwise with the embeddings. \n\nRight above Eq. (7), there is the text \"Taking m = 1\" -- what is m?\n\nIn Sec. 4.2: What exactly is meant by \"character embeddings\"? There are many ways to embed words using characters. How are the character embeddings composed to form a word embedding?\n\nHow was the random search done? I see the sentence \"For Random, we use the same training settings as our approach\", which makes me assume there were 30 steps, but I think this should be made more explicit. Since the random approach and ACE are different algorithms with different hyperparameters, it's not clear to me what is meant by using \"the same training settings\". \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603998662355}, {"id": "nQ-wfd0Xx7a", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1096/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n \nThis paper proposes to automate the concatenation of word embeddings (obtained using different strategies) to produce powerful word representations for a given downstream task. To this end, the paper develops an approach based on Neural Architecture Search, wherein the search space is comprised of embedding candidates obtained using different concatenations. Using an accuracy-based reward function, it is showed that ACE can determine more effective concatenations. ACE is evaluated using  extensive experiments with different tasks and datasets, and it outperforms the two baselines (to different degrees) -- random search and concatenating all embeddings with no subselection.\n\n##########################################################################\n\nPositives:\n- The idea of using NAS to construct concatenated embeddings is interesting and the formulation is clearly developed in the paper.\n- The proposed approach is generic and can support different types of structured outputs (sequences, graphs etc.)\n- The search process is computationally efficient and can be even run on a single GPU.\n- A simple modification (based on a discount factor) is proposed to the reward function design that leads to non-trivial performance improvements.\n- Strong experiment design: The proposed approach is evaluated on a large suite of datasets and tasks, and in many cases \n\nConcerns:\n- While the overall idea is interesting, the design choices made in the paper are not fully justified. Since ACE already pretrains the task model for each of the embeddings independently to begin with, why not adopt a \"boosting\" style approach instead of the naive \"ALL\" baseline. It is not surprising that even random search (known to be a strong baseline) consistently outperforms \"ALL\". The key challenge in concatenating disparate emebddings is that they can predict with varying degrees of confidence in different parts of the data and sequential inclusion of embeddings could be effective. In my opinion, the baselines chosen for concatenation are weak.\n- Why is \"accuracy\" the best choice for reward design? There could be two different embeddings that could produce the same accuracy with varying levels of confidence (or empirical calibration). Unlike conventional ensemble learners, each contextualized representation is not a weak learner and hence it will be critical to take into account confidence estimates.\n\nOverall, though the paper is experimentally strong, the design choices and the baselines need to be better justified.\n##########################################################################\n\nQuestions during rebuttal period: \n\nPlease respond to the questions under concerns.\n\n##########################################################################\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "AUTOMATED CONCATENATION OF EMBEDDINGS FOR STRUCTURED PREDICTION", "review": "Summary:\n \nThis paper proposes to automate the concatenation of word embeddings (obtained using different strategies) to produce powerful word representations for a given downstream task. To this end, the paper develops an approach based on Neural Architecture Search, wherein the search space is comprised of embedding candidates obtained using different concatenations. Using an accuracy-based reward function, it is showed that ACE can determine more effective concatenations. ACE is evaluated using  extensive experiments with different tasks and datasets, and it outperforms the two baselines (to different degrees) -- random search and concatenating all embeddings with no subselection.\n\n##########################################################################\n\nPositives:\n- The idea of using NAS to construct concatenated embeddings is interesting and the formulation is clearly developed in the paper.\n- The proposed approach is generic and can support different types of structured outputs (sequences, graphs etc.)\n- The search process is computationally efficient and can be even run on a single GPU.\n- A simple modification (based on a discount factor) is proposed to the reward function design that leads to non-trivial performance improvements.\n- Strong experiment design: The proposed approach is evaluated on a large suite of datasets and tasks, and in many cases \n\nConcerns:\n- While the overall idea is interesting, the design choices made in the paper are not fully justified. Since ACE already pretrains the task model for each of the embeddings independently to begin with, why not adopt a \"boosting\" style approach instead of the naive \"ALL\" baseline. It is not surprising that even random search (known to be a strong baseline) consistently outperforms \"ALL\". The key challenge in concatenating disparate emebddings is that they can predict with varying degrees of confidence in different parts of the data and sequential inclusion of embeddings could be effective. In my opinion, the baselines chosen for concatenation are weak.\n- Why is \"accuracy\" the best choice for reward design? There could be two different embeddings that could produce the same accuracy with varying levels of confidence (or empirical calibration). Unlike conventional ensemble learners, each contextualized representation is not a weak learner and hence it will be critical to take into account confidence estimates.\n\nOverall, though the paper is experimentally strong, the design choices and the baselines need to be better justified.\n##########################################################################\n\nQuestions during rebuttal period: \n\nPlease respond to the questions under concerns.\n\n##########################################################################\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603897880078}, {"id": "QzpGk_O_TP4", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1096/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper explores a way of learning how to automatically construct a concatenated set of embeddings for structured prediction tasks in NLP. The paper's model takes up to L embeddings concatenated together and feeds them into standard models (BiLSTM-CRFs or the BiLSTM-Biaffine technique of Dozat and Manning) to tackle problems like POS tagging, NER, dependency parsing, and more.  Search over embedding concaneations is expressed as a search over binary masks of length L.  The controller for this search is parameterized by an independent Bernoulli for each mask position.  The paper's approach learns the controller parameters with policy gradient, where the reward function is (a modified version of) the accuracy on the development set for the given task. This modified reward uses all samples throughout training to effectively get a more fine-grained baseline for the current timestep based on prior samples.  Notably, the paper uses embeddings that are already fine-tuned for each task, as fine-tuning the concatenated embeddings is hard due to divergent step sizes and steep computational requirements.\n\nResults show gains over randomly searching the space of binary masks. The overall model outperforms XLM-R in a range of multilingual settings.\n\nThis paper has some nice empirical results and the simplicity of its approach is attractive. But there are two shortcomings of the paper I will discuss.\n\nMOTIVATION/COMPARISONS\n\nThe authors motivate their technique by drawing parallels to neural architecture search. But I actually think what the authors are doing more closely resembles ensembling, system combination, or model stacking, e.g.:\nhttps://www.aclweb.org/anthology/N09-2064.pdf\nhttps://www.aclweb.org/anthology/N18-1201.pdf\n\nWhen you take large Transformer models (I have to imagine that the Transformers are contributing more to the performance than GloVe and other static word embeddings -- and Table 11 supports this somewhat) and staple a BiLSTM-CRF on top of them, most of the computation is happening in the (fixed) large Transformer. Most NAS methods I'm familiar with re-learn fundamental aspects of the architecture (e.g., the Evolved Transformer), while fixing most of the architecture and re-learning a last layer or two is more suggestive of system combination or model stacking.\n\nMy main question is: did the authors try comparing to an ensemble or post-hoc combination of the predictions according to different models?  Computationally this would be cheaper than what the authors did. It's also much faster to search over 2^L-1 possibilities when checking each possibility just requires decoding the dev set rather than training the BiLSTM-CRF -- actually, this can be done very efficiently if each model's logits are cached.\n\nThere are more sophisticated variants of this like in the papers I linked above where each model has its own weights or additional inputs are used. Intellectually, I think these approaches are related, and they should be discussed and compared to.\n\nRESULTS\n\nAs for the results, Table 1's gains are small -- they are consistent over random search, but I don't find them all that convincing.  There are too many embeddings here for ALL to work well -- my guess would be that a smaller set would yield better performance.\n\nTables 2-4 show improvements over existing baselines, XLM-R, and XLNet. This performance is commendable. However, again, I don't know how this compares to ensembling across a few leading approaches (like mBERT and XLM-R for the cross-lingual tasks).\n\nCONCLUSION\n\nIn the end, I'm not sure how readily this approach will be picked up by others. Because the embeddings aren't themselves fine-tuned as part of the ensemble, it really feels more like a fine-tuned ensemble of existing models rather than true NAS. And the overhead of this approach is significant: it requires running many training runs over large collections of existing pre-trained models to get a small improvement over the current state-of-the-art. This is a possibly useful datapoint to have in the literature, but it feels like the technique isn't quite right to lead to more work in this area.\n\nMINOR:\n\n\"use BiLSTM-Biaffine model (Dozat & Manning, 2017) for graph-structured outputs\"\n\nThis is a very particular structure, namely a directed minimum spanning tree (MST), though projective trees are also possible using the Eisner algorithm. The paper should specify that it's these, and not arbitrary graphs that are being produced here.\n\n-------------------\n\nUPDATE AFTER RESPONSE\n\nThanks for the response and the additional experiments. The comparison between ACE and these other techniques is nice to see, although I'll note that both SWAF and voting shouldn't make totally independent predictions in tasks like NER, but should at least respect constraints in the label space (not sure if there were applied or not).\n\nIn the end, my opinion of this paper largely comes down to the practicality of this technique and its likelihood to be adopted more generally. This results in a large, complex model, and while I am now convinced that the authors have a better ensembling/combination technique than some others, I think it still falls short of a real \"neural architecture search\" contribution or a really exciting result.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Some nice results, but I'm not fully convinced by the technique", "review": "This paper explores a way of learning how to automatically construct a concatenated set of embeddings for structured prediction tasks in NLP. The paper's model takes up to L embeddings concatenated together and feeds them into standard models (BiLSTM-CRFs or the BiLSTM-Biaffine technique of Dozat and Manning) to tackle problems like POS tagging, NER, dependency parsing, and more.  Search over embedding concaneations is expressed as a search over binary masks of length L.  The controller for this search is parameterized by an independent Bernoulli for each mask position.  The paper's approach learns the controller parameters with policy gradient, where the reward function is (a modified version of) the accuracy on the development set for the given task. This modified reward uses all samples throughout training to effectively get a more fine-grained baseline for the current timestep based on prior samples.  Notably, the paper uses embeddings that are already fine-tuned for each task, as fine-tuning the concatenated embeddings is hard due to divergent step sizes and steep computational requirements.\n\nResults show gains over randomly searching the space of binary masks. The overall model outperforms XLM-R in a range of multilingual settings.\n\nThis paper has some nice empirical results and the simplicity of its approach is attractive. But there are two shortcomings of the paper I will discuss.\n\nMOTIVATION/COMPARISONS\n\nThe authors motivate their technique by drawing parallels to neural architecture search. But I actually think what the authors are doing more closely resembles ensembling, system combination, or model stacking, e.g.:\nhttps://www.aclweb.org/anthology/N09-2064.pdf\nhttps://www.aclweb.org/anthology/N18-1201.pdf\n\nWhen you take large Transformer models (I have to imagine that the Transformers are contributing more to the performance than GloVe and other static word embeddings -- and Table 11 supports this somewhat) and staple a BiLSTM-CRF on top of them, most of the computation is happening in the (fixed) large Transformer. Most NAS methods I'm familiar with re-learn fundamental aspects of the architecture (e.g., the Evolved Transformer), while fixing most of the architecture and re-learning a last layer or two is more suggestive of system combination or model stacking.\n\nMy main question is: did the authors try comparing to an ensemble or post-hoc combination of the predictions according to different models?  Computationally this would be cheaper than what the authors did. It's also much faster to search over 2^L-1 possibilities when checking each possibility just requires decoding the dev set rather than training the BiLSTM-CRF -- actually, this can be done very efficiently if each model's logits are cached.\n\nThere are more sophisticated variants of this like in the papers I linked above where each model has its own weights or additional inputs are used. Intellectually, I think these approaches are related, and they should be discussed and compared to.\n\nRESULTS\n\nAs for the results, Table 1's gains are small -- they are consistent over random search, but I don't find them all that convincing.  There are too many embeddings here for ALL to work well -- my guess would be that a smaller set would yield better performance.\n\nTables 2-4 show improvements over existing baselines, XLM-R, and XLNet. This performance is commendable. However, again, I don't know how this compares to ensembling across a few leading approaches (like mBERT and XLM-R for the cross-lingual tasks).\n\nCONCLUSION\n\nIn the end, I'm not sure how readily this approach will be picked up by others. Because the embeddings aren't themselves fine-tuned as part of the ensemble, it really feels more like a fine-tuned ensemble of existing models rather than true NAS. And the overhead of this approach is significant: it requires running many training runs over large collections of existing pre-trained models to get a small improvement over the current state-of-the-art. This is a possibly useful datapoint to have in the literature, but it feels like the technique isn't quite right to lead to more work in this area.\n\nMINOR:\n\n\"use BiLSTM-Biaffine model (Dozat & Manning, 2017) for graph-structured outputs\"\n\nThis is a very particular structure, namely a directed minimum spanning tree (MST), though projective trees are also possible using the Eisner algorithm. The paper should specify that it's these, and not arbitrary graphs that are being produced here.\n\n-------------------\n\nUPDATE AFTER RESPONSE\n\nThanks for the response and the additional experiments. The comparison between ACE and these other techniques is nice to see, although I'll note that both SWAF and voting shouldn't make totally independent predictions in tasks like NER, but should at least respect constraints in the label space (not sure if there were applied or not).\n\nIn the end, my opinion of this paper largely comes down to the practicality of this technique and its likelihood to be adopted more generally. This results in a large, complex model, and while I am now convinced that the authors have a better ensembling/combination technique than some others, I think it still falls short of a real \"neural architecture search\" contribution or a really exciting result.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603862410132}, {"id": "NyAPJqvXtyH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1096/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper introduced an interesting application of reinforcement learning in the selection of concatenation of contextual/non-contextual word embeddings.  It is clever to limit the search space on the selection of embedding sources rather than search the whole network structure, as the current strategy is much easier in the training step. The author(s) conducted many experiments that compared many other models (including SOTA models, and ablation study). Those results are pretty good and impressive.\n\nThe main concern is the necessity of using the concatenation of those contextual embeddings. Calculating different contextual embeddings will cost many computing resources and affect the model's speed. Instead of using reinforcement learning to learn the concatenation of different embeddings, why not use the ensemble model to aggregate the results from different contextual embedding-based models? For example, we can use three embeddings BERT, ELMO, Glove to build three separate models and then aggerate their predicted results, the computing speed/resource may be similar to the learned embedding concatenation model, but is it possible that the ensemble model outperforms the ACE model? More experiments of the comparison with ensemble models should be conducted to prove the necessity of ACE. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting, good results, but too heavy and complex", "review": "This paper introduced an interesting application of reinforcement learning in the selection of concatenation of contextual/non-contextual word embeddings.  It is clever to limit the search space on the selection of embedding sources rather than search the whole network structure, as the current strategy is much easier in the training step. The author(s) conducted many experiments that compared many other models (including SOTA models, and ablation study). Those results are pretty good and impressive.\n\nThe main concern is the necessity of using the concatenation of those contextual embeddings. Calculating different contextual embeddings will cost many computing resources and affect the model's speed. Instead of using reinforcement learning to learn the concatenation of different embeddings, why not use the ensemble model to aggregate the results from different contextual embedding-based models? For example, we can use three embeddings BERT, ELMO, Glove to build three separate models and then aggerate their predicted results, the computing speed/resource may be similar to the learned embedding concatenation model, but is it possible that the ensemble model outperforms the ACE model? More experiments of the comparison with ensemble models should be conducted to prove the necessity of ACE. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603076554967}], "openreview_url": "https://openreview.net/forum?id=nCY83KxoehA", "arxiv_id": "2010.05006", "paper_pdf": "papers/nCY83KxoehA.pdf", "paper_pdf_sha256": "a3c69d41fa029be962e56d93214e2fc09181deabb5f0be33fc1c7bfb04ffd848", "paper_pdf_bytes": 401053, "paper_pdf_source": "openreview", "code_url": "https://github.com/Alibaba-NLP/ACE", "code_repository": "Alibaba-NLP/ACE", "code_commit": "cf50440b5d4ab44f74feb4214733224265428db4", "code_archive": "repos/nCY83KxoehA.zip", "code_archive_sha256": "91055b0b5bac3f17a3e0649024ed0f846fe80507facaa1a22a9bff8164a4e169", "code_archive_bytes": 1456656, "code_file_count": 123, "code_extensions": {".py": 123}, "github_disk_usage_kb": 1758, "github_languages": {"Python": 1615035, "Perl": 12743}, "github_archived": false, "github_pushed_at": "2022-12-02T09:13:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/automated-concatenation-of-embeddings-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "AEgyitdRWf", "year": 2026, "status": "rejected", "title": "ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering", "authors": ["Zexi Liu", "Jingyi Chai", "Xinyu Zhu", "Shuo Tang", "Rui Ye", "Weiyu Ma", "Bo Zhang", "LEI BAI", "Siheng Chen"], "authorids": ["~Zexi_Liu1", "~Jingyi_Chai1", "~Xinyu_Zhu5", "~Shuo_Tang2", "~Rui_Ye1", "~Weiyu_Ma1", "~Bo_Zhang17", "~LEI_BAI1", "~Siheng_Chen1"], "authors_source": "OpenReview API", "abstract": "The emergence of large language model (LLM)-based agents has significantly advanced the development of autonomous machine learning (ML) engineering. However, the dominant prompt-based paradigm exhibits limitations: smaller models lack the capacity to learn from execution trajectories for generalization, while large proprietary models incur high computational overhead, restricting accessibility and scalability.\nFocusing on this, for the first time, we explore the paradigm of learning-based agentic ML, where an LLM agent learns through interactive experimentation on ML tasks using online reinforcement learning (RL). To realize this, we propose a novel agentic ML training framework with three key components:\n(1) exploration-enriched fine-tuning, which enables LLM agents to generate diverse actions for enhanced RL exploration;\n(2) step-wise RL, which enables training on a single action step, accelerating experience collection and improving training efficiency;\n(3) an agentic ML-specific reward module, which unifies varied ML feedback signals into consistent rewards for RL optimization. \nLeveraging this framework, we train ML-Agent, driven by a 7B-sized Qwen-2.5 LLM for autonomous ML.\nDespite training on only 9 ML tasks, our 7B-sized ML-Agent achieves comparable performance to agents using much larger proprietary LLMs (e.g., GPT-5) but at significantly lower computational cost, demonstrating strong performance and cross-task generalization.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "DcPmPVVv7w", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4134/Reviewer_D4Wx"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes an online reinforcement learning agent training framework applied to agentic machine learning tasks. The proposed training framework involves three stages: exploration-enriched fine-tuning, step-wise RL, and rewards that are specific to agentic ML tasks. Experimentally, the authors show that with limited training, their 7B agent can match the performance of frontier models such as GPT-5.", "review_text": "This paper proposes an online reinforcement learning agent training framework applied to agentic machine learning tasks. The proposed training framework involves three stages: exploration-enriched fine-tuning, step-wise RL, and rewards that are specific to agentic ML tasks. Experimentally, the authors show that with limited training, their 7B agent can match the performance of frontier models such as GPT-5.", "strengths": "- This paper provides strong evidence that RL can be used successfully to improve smaller (7B) models to the point of enabling complex agentic behavior on ML-specific tasks, which I view as a strong contribution. \n- The proposed technical solutions are quite reasonable. In particular, the authors note that ML agents can suffer from a lack of exploration, and provide an SFT technique that specifically targets this problem. The proposed exploration-enriched fine-tuning technique is both novel and well-executed. \n- The results on the tasks that were evaluated suggest that the method works and is competitive with frontier models. This is a strength, however, I have discussed other concerns about the evaluation details in the weaknesses section.", "weaknesses": "- The authors train on only 9 tasks from subsets of MLAgentBench and MLE-Bench. How were these tasks chosen? Details about the rationale behind the training task selection seem missing. \n- During evaluation, the authors evaluated on a small number of held-out tasks from MLE-Bench. It is unclear how this evaluation subset was selected from the larger set of tasks that were not used during training, even within MLE-Bench. A more convincing evaluation might involve evaluating on entire held-out benchmarks. \n- Training requires generating trajectories using GPT-4o-mini (or some other expert model) which might fundamentally limit the complexity of tasks on which it can be used, to tasks that are already solvable by frontier models.", "questions": "- Related to the first weakness point, can the authors provide the specific criteria used to select the 9 training tasks? \n- MLE-bench contains significantly more than 10 tasks. How were the 10 held-out evaluation tasks selected from the full set of unused tasks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an online reinforcement learning agent training framework applied to agentic machine learning tasks. The proposed training framework involves three stages: exploration-enriched fine-tuning, step-wise RL, and rewards that are specific to agentic ML tasks. Experimentally, the authors show that with limited training, their 7B agent can match the performance of frontier models such as GPT-5.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- This paper provides strong evidence that RL can be used successfully to improve smaller (7B) models to the point of enabling complex agentic behavior on ML-specific tasks, which I view as a strong contribution. \n- The proposed technical solutions are quite reasonable. In particular, the authors note that ML agents can suffer from a lack of exploration, and provide an SFT technique that specifically targets this problem. The proposed exploration-enriched fine-tuning technique is both novel and well-executed. \n- The results on the tasks that were evaluated suggest that the method works and is competitive with frontier models. This is a strength, however, I have discussed other concerns about the evaluation details in the weaknesses section.", "weaknesses": "- The authors train on only 9 tasks from subsets of MLAgentBench and MLE-Bench. How were these tasks chosen? Details about the rationale behind the training task selection seem missing. \n- During evaluation, the authors evaluated on a small number of held-out tasks from MLE-Bench. It is unclear how this evaluation subset was selected from the larger set of tasks that were not used during training, even within MLE-Bench. A more convincing evaluation might involve evaluating on entire held-out benchmarks. \n- Training requires generating trajectories using GPT-4o-mini (or some other expert model) which might fundamentally limit the complexity of tasks on which it can be used, to tasks that are already solvable by frontier models.", "questions": "- Related to the first weakness point, can the authors provide the specific criteria used to select the 9 training tasks? \n- MLE-bench contains significantly more than 10 tasks. How were the 10 held-out evaluation tasks selected from the full set of unused tasks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762112796698}, {"id": "sEwf0xXhNS", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4134/Reviewer_i2zT"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces RL to train ML Engineering agents to learn from past experiences. They address three problems in training an ML Agent: first, small ML agents lack exploration; they address this by distilling from a larger model. Second, to overcome the slow feedback loop of typical ML experiments, the authors introduce a step-wise reinforcement learning (RL) paradigm. Instead of learning from entire trajectories, they reformulate the problem to learn from single action steps sampled from the pre-collected expert states. Finally, this RL process is guided by a carefully designed agentic ML-specific reward module. Experiments show comparable performance of a trained 7B model to a prompted 600B model. Although the experimental results are promising, the paper lacks discussion on highly probable reward hacking (due to reward design) and distribution shift problems (due to step-wise RL).\n\nOverall, the paper introduces a novel learning-based paradigm that enables smaller models to achieve strong performance, addresses practical challenges. However, the approach has fundamental limitations in true exploration due to dependence on expert-generated state distributions, limited task diversity in training, incomplete cost analysis that excludes training overhead, and potential generalization concerns to tasks outside the benchmark distribution.", "review_text": "The paper introduces RL to train ML Engineering agents to learn from past experiences. They address three problems in training an ML Agent: first, small ML agents lack exploration; they address this by distilling from a larger model. Second, to overcome the slow feedback loop of typical ML experiments, the authors introduce a step-wise reinforcement learning (RL) paradigm. Instead of learning from entire trajectories, they reformulate the problem to learn from single action steps sampled from the pre-collected expert states. Finally, this RL process is guided by a carefully designed agentic ML-specific reward module. Experiments show comparable performance of a trained 7B model to a prompted 600B model. Although the experimental results are promising, the paper lacks discussion on highly probable reward hacking (due to reward design) and distribution shift problems (due to step-wise RL).\n\nOverall, the paper introduces a novel learning-based paradigm that enables smaller models to achieve strong performance, addresses practical challenges. However, the approach has fundamental limitations in true exploration due to dependence on expert-generated state distributions, limited task diversity in training, incomplete cost analysis that excludes training overhead, and potential generalization concerns to tasks outside the benchmark distribution.", "strengths": "The paper's core strength is developing an autonomous ML agent that learns from experience rather than a prompt-engineered heuristic. The proposed step‑wise RL objective makes online training easier in ML settings with slow experiments. Empirical results show competitive performance of a 7B backbone model compared to Larger 600B size backbone models without training.\nThe paper systematically validates each component's contribution, showing that all three technical components are necessary. Despite training on only 9 tasks, the model shows meaningful performance gains on 10 held-out tasks.", "weaknesses": "1. Different execution times: The paper, although it proposes RL fine-tuning for MLE agents, does not consider a practical problem in training agents specifically for MLE tasks. Different actions from the same state could have different execution times, and if optimized in a vanilla fashion, would lead to more time-consuming yet optimal solutions being explored less. Consider that given a neural network, the agent adds extra layers in one action, opposed to changing the learning rate in a parallel action (people generally perform parallel rollouts for efficiency), then the second type of action would lead to faster rewards and hence be more frequent during training. \n2. Reward function is coarse: Mapping all the error cases to -1 and all the valid or corner edits to 0 would treat syntax issues, dependency issues alike, and all the types of out-of-memory issues (model loading OOM, memory leak OOM, and large batch size OOM) the same, although they convey different information and levels of mistakes and understanding. Moreover, the reward considers only the task accuracy and ignores compute, memory, cost, etc., important parameters. Although the reward is simple, it is prone to reward hacking. Another instance of reward hacking could be that the agent learns to find good seeds to make the performance increase, rather than actually learning to improve the architecture or training algorithm. \n3. Trajectory Collection is offline with respect to the training policy: The paper states that they use the collected trajectories and at each step, ask the smaller agent to propose an action on a state sampled from the larger agent. This might lead to the problems of distribution shift when using the smaller agent to generate the trajectories from scratch. This might also bias the learning towards expert behavior and limit exploration. This is more akin to a mix of off-policy behavior cloning and on-policy action generation. The agent cannot learn to recover from or explore states outside the expert trajectory distribution, which severely constrains the \"learning through experimentation\" claim. This raises the question, how would this approach extend to domains where high-quality expert trajectories aren't available?\n4. Incomplete Cost Analysis: Figure 2 only compares the inference costs per trajectory but excludes the substantial training and inference costs required for data collection using GPT-4o-mini, supervised fine-tuning, and RL training. For a fair comparison, these costs should be amortized over expected usage.\n5. Narrow Task Distribution and Generalization: Training on mainly regression and classification, and then testing on similar tasks raises the question of true cross-task generalization (except one generation task in testing). The paper lacks evidence of generalization to: (a) fundamentally different ML tasks, e.g., if trained on supervised learning, can it handle RL tasks?, (b) different data modalities not seen in training.\n6. Evaluation Metric Limitations: The Performance gain delta metric depends heavily on the initial script quality, which may vary across tasks. A task with a poor initial script will show larger gains for the same absolute improvement. The paper would benefit from also reporting: (a) absolute performance metrics, (b) comparison against human expert solutions, and (c) success rate in reaching specific performance thresholds.\n7. Script editing using a different model: Note that the training script is not edited by the model being trained. The actual editing is done by a different model (according to Table 4 and prompts shown), yet there is no discussion on which model or agent scaffold was used for this action. Moreover, if the policy being trained is not making edits, why is it penalized for errors induced by the editing model?", "questions": "1. How do you prevent the agent from 'reward hacking' by discovering shortcuts, such as finding optimal random seeds, rather than learning generalizable ML engineering improvements?\n2. Since the trajectories are collected offline, how does this affect the test time performance of the model when it has to generate an action on its own trajectories that it has never seen during training time? What is the performance of your model when starting from an incorrect or suboptimal state compared to a pre-trained prompt-based method?\n3. How do you account for more complex solutions that are less explored during RL due to the high time of execution, leading to fewer occurrences?\n4. Which model is used to edit the script?\n5. What is the average trajectory length at testing time? What is the average execution time of the solution generated by the trained model?\n6. What is the total computational cost, including expert trajectory collection, fine-tuning, and RL training? How many trajectories would need to be run at inference time to amortize this cost compared to a strong prompt-based method directly?\n7. How sensitive is the approach to expert trajectory quality? What happens if you use trajectories from a weaker model?\n8. Can you provide evidence of generalization to qualitatively different ML tasks? For example, if trained on supervised learning tasks, can it handle reinforcement learning or unsupervised learning tasks?\n9. How do you select the baseline script for each task? The reward and evaluation both depend on this selection.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces RL to train ML Engineering agents to learn from past experiences. They address three problems in training an ML Agent: first, small ML agents lack exploration; they address this by distilling from a larger model. Second, to overcome the slow feedback loop of typical ML experiments, the authors introduce a step-wise reinforcement learning (RL) paradigm. Instead of learning from entire trajectories, they reformulate the problem to learn from single action steps sampled from the pre-collected expert states. Finally, this RL process is guided by a carefully designed agentic ML-specific reward module. Experiments show comparable performance of a trained 7B model to a prompted 600B model. Although the experimental results are promising, the paper lacks discussion on highly probable reward hacking (due to reward design) and distribution shift problems (due to step-wise RL).\n\nOverall, the paper introduces a novel learning-based paradigm that enables smaller models to achieve strong performance, addresses practical challenges. However, the approach has fundamental limitations in true exploration due to dependence on expert-generated state distributions, limited task diversity in training, incomplete cost analysis that excludes training overhead, and potential generalization concerns to tasks outside the benchmark distribution.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "The paper's core strength is developing an autonomous ML agent that learns from experience rather than a prompt-engineered heuristic. The proposed step‑wise RL objective makes online training easier in ML settings with slow experiments. Empirical results show competitive performance of a 7B backbone model compared to Larger 600B size backbone models without training.\nThe paper systematically validates each component's contribution, showing that all three technical components are necessary. Despite training on only 9 tasks, the model shows meaningful performance gains on 10 held-out tasks.", "weaknesses": "1. Different execution times: The paper, although it proposes RL fine-tuning for MLE agents, does not consider a practical problem in training agents specifically for MLE tasks. Different actions from the same state could have different execution times, and if optimized in a vanilla fashion, would lead to more time-consuming yet optimal solutions being explored less. Consider that given a neural network, the agent adds extra layers in one action, opposed to changing the learning rate in a parallel action (people generally perform parallel rollouts for efficiency), then the second type of action would lead to faster rewards and hence be more frequent during training. \n2. Reward function is coarse: Mapping all the error cases to -1 and all the valid or corner edits to 0 would treat syntax issues, dependency issues alike, and all the types of out-of-memory issues (model loading OOM, memory leak OOM, and large batch size OOM) the same, although they convey different information and levels of mistakes and understanding. Moreover, the reward considers only the task accuracy and ignores compute, memory, cost, etc., important parameters. Although the reward is simple, it is prone to reward hacking. Another instance of reward hacking could be that the agent learns to find good seeds to make the performance increase, rather than actually learning to improve the architecture or training algorithm. \n3. Trajectory Collection is offline with respect to the training policy: The paper states that they use the collected trajectories and at each step, ask the smaller agent to propose an action on a state sampled from the larger agent. This might lead to the problems of distribution shift when using the smaller agent to generate the trajectories from scratch. This might also bias the learning towards expert behavior and limit exploration. This is more akin to a mix of off-policy behavior cloning and on-policy action generation. The agent cannot learn to recover from or explore states outside the expert trajectory distribution, which severely constrains the \"learning through experimentation\" claim. This raises the question, how would this approach extend to domains where high-quality expert trajectories aren't available?\n4. Incomplete Cost Analysis: Figure 2 only compares the inference costs per trajectory but excludes the substantial training and inference costs required for data collection using GPT-4o-mini, supervised fine-tuning, and RL training. For a fair comparison, these costs should be amortized over expected usage.\n5. Narrow Task Distribution and Generalization: Training on mainly regression and classification, and then testing on similar tasks raises the question of true cross-task generalization (except one generation task in testing). The paper lacks evidence of generalization to: (a) fundamentally different ML tasks, e.g., if trained on supervised learning, can it handle RL tasks?, (b) different data modalities not seen in training.\n6. Evaluation Metric Limitations: The Performance gain delta metric depends heavily on the initial script quality, which may vary across tasks. A task with a poor initial script will show larger gains for the same absolute improvement. The paper would benefit from also reporting: (a) absolute performance metrics, (b) comparison against human expert solutions, and (c) success rate in reaching specific performance thresholds.\n7. Script editing using a different model: Note that the training script is not edited by the model being trained. The actual editing is done by a different model (according to Table 4 and prompts shown), yet there is no discussion on which model or agent scaffold was used for this action. Moreover, if the policy being trained is not making edits, why is it penalized for errors induced by the editing model?", "questions": "1. How do you prevent the agent from 'reward hacking' by discovering shortcuts, such as finding optimal random seeds, rather than learning generalizable ML engineering improvements?\n2. Since the trajectories are collected offline, how does this affect the test time performance of the model when it has to generate an action on its own trajectories that it has never seen during training time? What is the performance of your model when starting from an incorrect or suboptimal state compared to a pre-trained prompt-based method?\n3. How do you account for more complex solutions that are less explored during RL due to the high time of execution, leading to fewer occurrences?\n4. Which model is used to edit the script?\n5. What is the average trajectory length at testing time? What is the average execution time of the solution generated by the trained model?\n6. What is the total computational cost, including expert trajectory collection, fine-tuning, and RL training? How many trajectories would need to be run at inference time to amortize this cost compared to a strong prompt-based method directly?\n7. How sensitive is the approach to expert trajectory quality? What happens if you use trajectories from a weaker model?\n8. Can you provide evidence of generalization to qualitatively different ML tasks? For example, if trained on supervised learning tasks, can it handle reinforcement learning or unsupervised learning tasks?\n9. How do you select the baseline script for each task? The reward and evaluation both depend on this selection.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761988460775}, {"id": "bd39MuhE9d", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4134/Reviewer_6UHw"], "rating": 2, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper introduces ML Agent, a method for fine-tuning LLMs using one-step RL, a lightweight alternative to traditional multi-step RL algorithms like PPO or DPO.\nInstead of simulating full dialogue trajectories or optimizing cumulative rewards, ML Agent applies a single-step policy improvement based on feedback signals (e.g., preference or quality scores) for individual responses.\nThe paper shows comparable results of post-training on a small-sized model compared to proprietary LLMs such as GPT-5.", "review_text": "The paper introduces ML Agent, a method for fine-tuning LLMs using one-step RL, a lightweight alternative to traditional multi-step RL algorithms like PPO or DPO.\nInstead of simulating full dialogue trajectories or optimizing cumulative rewards, ML Agent applies a single-step policy improvement based on feedback signals (e.g., preference or quality scores) for individual responses.\nThe paper shows comparable results of post-training on a small-sized model compared to proprietary LLMs such as GPT-5.", "strengths": "The new RL update method, if proven efficient and comparable or better to SOTA techniques, could be a good tool for LLM post-training when tool calls (such as training a ML model) are expensive.", "weaknesses": "- I am having a hard time seeing the novelty of this approach. The lack of comparison with a broader range of papers in the literature (Reinforcement Learning for Machine Learning Engineering Agents, MLE-Dojo, MLGym, AgentGym...) makes it hard to really see how much  is brought by this paper.\n- I don't see a theoretical justification for the One-step RL method, and the empirical validation is itself a bit weak. I think the authors should show that this is a sound thing to do by showing:\n  - Better sample efficiency vs PPO, or\n  - Equal or better final alignment compared to a well-tuned PPO model.\nThe statistical rigor (significance etc) is also limited.", "questions": "- Could the authors clarify how their method relates to recent frameworks like Reinforcement Learning for Machine Learning Engineering Agents, MLE-Dojo, MLGym, or AgentGym? The literature on agents / tools and ML scientists has exploded these past months. In particular, does ML Agent address any limitations or gaps identified in those works?\n- The paper presents one-step RL as novel, but similar formulations (reward-weighted log-likelihood updates) have been studied extensively. What, specifically, is new here, the algorithm, the training pipeline, something else?\n- Does the one-step approach introduce bias relative to multi-step RL methods such as PPO?\n- Why do the authors not include comparisons against PPO-based RLHF, or at least controlled reimplementations using similar datasets and reward models? One could use a simple problem with a small model if compute availability is an issue.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces ML Agent, a method for fine-tuning LLMs using one-step RL, a lightweight alternative to traditional multi-step RL algorithms like PPO or DPO.\nInstead of simulating full dialogue trajectories or optimizing cumulative rewards, ML Agent applies a single-step policy improvement based on feedback signals (e.g., preference or quality scores) for individual responses.\nThe paper shows comparable results of post-training on a small-sized model compared to proprietary LLMs such as GPT-5.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "The new RL update method, if proven efficient and comparable or better to SOTA techniques, could be a good tool for LLM post-training when tool calls (such as training a ML model) are expensive.", "weaknesses": "- I am having a hard time seeing the novelty of this approach. The lack of comparison with a broader range of papers in the literature (Reinforcement Learning for Machine Learning Engineering Agents, MLE-Dojo, MLGym, AgentGym...) makes it hard to really see how much  is brought by this paper.\n- I don't see a theoretical justification for the One-step RL method, and the empirical validation is itself a bit weak. I think the authors should show that this is a sound thing to do by showing:\n  - Better sample efficiency vs PPO, or\n  - Equal or better final alignment compared to a well-tuned PPO model.\nThe statistical rigor (significance etc) is also limited.", "questions": "- Could the authors clarify how their method relates to recent frameworks like Reinforcement Learning for Machine Learning Engineering Agents, MLE-Dojo, MLGym, or AgentGym? The literature on agents / tools and ML scientists has exploded these past months. In particular, does ML Agent address any limitations or gaps identified in those works?\n- The paper presents one-step RL as novel, but similar formulations (reward-weighted log-likelihood updates) have been studied extensively. What, specifically, is new here, the algorithm, the training pipeline, something else?\n- Does the one-step approach introduce bias relative to multi-step RL methods such as PPO?\n- Why do the authors not include comparisons against PPO-based RLHF, or at least controlled reimplementations using similar datasets and reward models? One could use a simple problem with a small model if compute availability is an issue.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761850208829}, {"id": "0gFcYGwJ2I", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission4134/Reviewer_ixvU"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors design a training framework that enables LLMs to learn from environment interactions efficiently, leveraging exploration-enriched fine-tuning, step-wise RL, and a ML-specific reward module. The authors train ML-Agent, a 7B-parameter model based on Qwen-2.5, which outperformed most baselines, and showed strong generalization result.", "review_text": "The authors design a training framework that enables LLMs to learn from environment interactions efficiently, leveraging exploration-enriched fine-tuning, step-wise RL, and a ML-specific reward module. The authors train ML-Agent, a 7B-parameter model based on Qwen-2.5, which outperformed most baselines, and showed strong generalization result.", "strengths": "1. Modular framework design. The three proposed components (exploration-enriched fine-tuning, step-wise RL, and reward unification) are intuitive, complementary, and grounded in practical RL challenges for ML agents.\n2. Strong empirical results. The 7B- model trained with this framework rivals GPT-5-driven agents, and shows great generalization.", "weaknesses": "1. Sparse ablation and analysis. The contribution exploration-enriched fine-tuning is not clearly isolated in ablation studies.\n2. Lack of novelty. The three components of the proposed framework, exploration-enriched fine-tuning, step-wise RL, and the agentic ML-specific reward module, appear to be additive rather than integrated or co-designed.", "questions": "1. Could you report quantitative ablations isolating each component, e.g., without exploration-enriched fine-tuning, without step-wise RL, and without reward normalization?\n2. Does exploration-enriched fine-tuning empirically increase coverage of the action space?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors design a training framework that enables LLMs to learn from environment interactions efficiently, leveraging exploration-enriched fine-tuning, step-wise RL, and a ML-specific reward module. The authors train ML-Agent, a 7B-parameter model based on Qwen-2.5, which outperformed most baselines, and showed strong generalization result.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Modular framework design. The three proposed components (exploration-enriched fine-tuning, step-wise RL, and reward unification) are intuitive, complementary, and grounded in practical RL challenges for ML agents.\n2. Strong empirical results. The 7B- model trained with this framework rivals GPT-5-driven agents, and shows great generalization.", "weaknesses": "1. Sparse ablation and analysis. The contribution exploration-enriched fine-tuning is not clearly isolated in ablation studies.\n2. Lack of novelty. The three components of the proposed framework, exploration-enriched fine-tuning, step-wise RL, and the agentic ML-specific reward module, appear to be additive rather than integrated or co-designed.", "questions": "1. Could you report quantitative ablations isolating each component, e.g., without exploration-enriched fine-tuning, without step-wise RL, and without reward normalization?\n2. Does exploration-enriched fine-tuning empirically increase coverage of the action space?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No.", "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761578726504}], "openreview_url": "https://openreview.net/forum?id=AEgyitdRWf", "arxiv_id": "2505.23723", "paper_pdf": "papers/AEgyitdRWf.pdf", "paper_pdf_sha256": "860c4bcd5f977fd96dee280d1b269e40187a1e98d91ba4605c4080df351f5137", "paper_pdf_bytes": 2392506, "paper_pdf_source": "openreview", "code_url": "https://github.com/MASWorks/ML-Agent", "code_repository": "MASWorks/ML-Agent", "code_commit": "15932e7525deb99d59f7416bbe8c75077cff3690", "code_archive": "repos/AEgyitdRWf.zip", "code_archive_sha256": "c0207c1ddb4cf5870bdda0284dfbfcac5ec13fa7a44cc5c4e0ed04a15d175414", "code_archive_bytes": 7311896, "code_file_count": 156, "code_extensions": {".py": 150, ".sh": 6}, "github_disk_usage_kb": 7394, "github_languages": {"Python": 1089531, "Shell": 2413, "Makefile": 391}, "github_archived": false, "github_pushed_at": "2025-06-21T18:17:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ml-agent-reinforcing-llm-agents-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "f4mQ2SU5tp", "year": 2025, "status": "rejected", "title": "IntLoRA: Integral Low-rank Adaptation of Quantized Diffusion Models", "authors": ["Hang Guo", "Yawei Li", "Tao Dai", "Shu-Tao Xia", "Luca Benini"], "authorids": ["~Hang_Guo3", "~Yawei_Li1", "~Tao_Dai3", "~Shu-Tao_Xia1", "~Luca_Benini2"], "authors_source": "OpenReview API", "abstract": "Fine-tuning large-scale text-to-image diffusion models for various downstream tasks has yielded impressive results. However, the heavy computational burdens of tuning large models prevent personal customization. Recent advances have attempted to employ parameter-efficient fine-tuning (PEFT) techniques to adapt the floating-point (FP) or quantized pre-trained weights. Nonetheless, the adaptation parameters in existing works are still restricted to FP arithmetic, hindering hardware-friendly acceleration. In this work, we propose IntLoRA, to further push the efficiency limits by using integer type (INT) low-rank parameters to adapt the quantized diffusion models. By working in the integer arithmetic, our IntLoRA offers three key advantages: (i) for fine-tuning, the pre-trained weights are quantized, reducing memory usage; (ii) for storage, both pre-trained and low-rank weights are in INT which consumes less disk space; (iii) for inference, IntLoRA weights can be naturally merged into quantized pre-trained weights through efficient integer multiplication or bit-shifting, eliminating additional post-training quantization. Extensive experiments demonstrate that IntLoRA can achieve performance on par with or even superior to the vanilla LoRA, accompanied by significant efficiency improvements.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "pkppS9DHDB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2140/Reviewer_wQ3r"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors propose IntLoRA, which employes INT low-rank parameters to adapt the quantized diffusion models, to address the arithmetic inconsistency between the quantized pre-trained weights and adaptation weights. All the process will operate directly on INT arithmetic in training and merged weights at inference stage will also be on INT format. Apparently, it will speed up both the training and inference stage.", "review_text": "The authors propose IntLoRA, which employes INT low-rank parameters to adapt the quantized diffusion models, to address the arithmetic inconsistency between the quantized pre-trained weights and adaptation weights. All the process will operate directly on INT arithmetic in training and merged weights at inference stage will also be on INT format. Apparently, it will speed up both the training and inference stage.", "strengths": "1) The presentation of challenges and contributions are clear to me. \n2) The presentation of methods and experiments are clear to me.", "weaknesses": "1) Recent works, e.g., EfficientDM and LoftQ, have worked out quantized diffusion models using LoRA. Different from this paper, EfficientDM focus on quantized merged weights at inference stage. Can you compare with it? BTW, I am doubting the importance of the speed-up for the training stage as the pretrained weights is fixed, only small adapters are updated.  Can you show how much speed-up as compared to full-precision? or at least theorotically show the improvement?\n\n2) From my understanding, the proposed methods also can be used in NLP tasks. As the baselines have results on NLP tasks in their papers, why not verify the proposed methods on NLP tasks? Could you show some experimental results on NLP tasks?\n\n3) For diffusion models, seems it have standard/common verification on some tasks, please check EfficientDM paper. As all baselines are implemented by authors, I just doubt the reliablity because of hyper-parameter tuning. Maybe it is better compare with SOTA results in paper.", "questions": "Have asked in Weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose IntLoRA, which employes INT low-rank parameters to adapt the quantized diffusion models, to address the arithmetic inconsistency between the quantized pre-trained weights and adaptation weights. All the process will operate directly on INT arithmetic in training and merged weights at inference stage will also be on INT format. Apparently, it will speed up both the training and inference stage.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1) The presentation of challenges and contributions are clear to me. \n2) The presentation of methods and experiments are clear to me.", "weaknesses": "1) Recent works, e.g., EfficientDM and LoftQ, have worked out quantized diffusion models using LoRA. Different from this paper, EfficientDM focus on quantized merged weights at inference stage. Can you compare with it? BTW, I am doubting the importance of the speed-up for the training stage as the pretrained weights is fixed, only small adapters are updated.  Can you show how much speed-up as compared to full-precision? or at least theorotically show the improvement?\n\n2) From my understanding, the proposed methods also can be used in NLP tasks. As the baselines have results on NLP tasks in their papers, why not verify the proposed methods on NLP tasks? Could you show some experimental results on NLP tasks?\n\n3) For diffusion models, seems it have standard/common verification on some tasks, please check EfficientDM paper. As all baselines are implemented by authors, I just doubt the reliablity because of hyper-parameter tuning. Maybe it is better compare with SOTA results in paper.", "questions": "Have asked in Weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730777943783}, {"id": "t4rbf946kj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2140/Reviewer_qTL5"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper introduces IntLoRA, a method that quantizes the LoRA branch to integers for parameter-efficient fine-tuning (PEFT) in diffusion models. Traditional approaches typically use 16-bit precision for LoRA branches and focus on quantizing only the base model during fine-tuning. However, for inference, these methods often require re-quantization when the LoRA branches are fused with the base model. Unlike prior methods, such as QLoRA, IntLoRA fine-tunes the LoRA branches in a quantization-aware manner. To maintain adapter performance, it incorporates techniques like Adaptation-Quantization Separation (AQS), Multiplicative Low-rank Adaptation (MLA), and Variance Matching Control (VMC), and supports both linear and log2 quantization. Extensive experiments across various downstream tasks—including subject-driven generation, controllable generation, and style customization—demonstrate the effectiveness of IntLoRA in both visual quality and model size reduction.", "review_text": "This paper introduces IntLoRA, a method that quantizes the LoRA branch to integers for parameter-efficient fine-tuning (PEFT) in diffusion models. Traditional approaches typically use 16-bit precision for LoRA branches and focus on quantizing only the base model during fine-tuning. However, for inference, these methods often require re-quantization when the LoRA branches are fused with the base model. Unlike prior methods, such as QLoRA, IntLoRA fine-tunes the LoRA branches in a quantization-aware manner. To maintain adapter performance, it incorporates techniques like Adaptation-Quantization Separation (AQS), Multiplicative Low-rank Adaptation (MLA), and Variance Matching Control (VMC), and supports both linear and log2 quantization. Extensive experiments across various downstream tasks—including subject-driven generation, controllable generation, and style customization—demonstrate the effectiveness of IntLoRA in both visual quality and model size reduction.", "strengths": "* The paper presents a strong motivation, addressing a key limitation in current quantization methods: the need for re-quantization after applying adapters. This limitation has prevented the wide adoption of quantization methods in diffusion models.\n* The concepts of Adaptation-Quantization Separation (AQS), Multiplicative Low-rank Adaptation (MLA), and Variance Matching Control (VMC) are innovative and well-supported by mathematical analysis. Their effectiveness is further validated through comprehensive experimental results.\n* The authors' support for log2 quantization, a challenging but highly hardware-efficient setting that further enhances the practical utility of IntLoRA.", "weaknesses": "* The paper exclusively focuses on weight-only quantization, where activations remain at 16-bit precision. Since diffusion models are primarily compute-bound, quantizing only the weights does not yield speedups during inference. Additionally, in a weight-only quantization setting, users can choose not to fuse adapters, thereby avoiding re-quantization. As a result, the primary benefit of IntLoRA is just a reduction in adapter size compared to conventional PEFT methods—a relatively minor contribution, given that LoRA adapters are already small.\n* he proposed PEFT method does not reduce training memory usage compared to existing PEFT methods (e.g., QLoRA). Furthermore, quantization-aware training of adapters may slow down training and even increase memory usage. The paper should discuss the training speed and memory usage.\n* Clarification on storage requirements for $\\mathbf{R}$ is necessary. Storing $\\mathbf{R}$ after training the adapters would require significant storage, which may limit the method's practicality. If $\\mathbf{R}$ is not stored, the authors should explain how it is obtained during inference. Also, note that $\\mathbf{R}$ should derive from $\\mathbf{W}_{\\text{round}}$ rather than $\\mathbf{W}$, as only quantized weights are accessible during PEFT.\n* Writing:\n  * Using \"FP\" to refer to original LoRA training is ambiguous, as there are low-precision formats like FP8 or FP4. Consider using \"16-bit\" instead. Note also that QLoRA uses NF4 for weight representation, which is not integer-based. Therefore, Lines 190–202 need to be rewritten.\n  * Line 156: The dimension of $\\mathbf{B}$ should be $d \\times C_{\\text{in}}$.\n  * Figure 2: The caption lacks detail, and the figure is difficult to interpret.\n  * Line 229: Refer to Section 5.3 on how to determine $\\mathbf{R}$.\n  * Line 265: Need to clarify $\\rho$ is the correlation coefficient.\n  * Table 4: Correct \"FP addtion\" to \"FP addition\".\n  * Figure 7: Add axis labels for clarity.", "questions": "* Line 252: Could you provide any insights or justification for the assumption that $\\mathbf{A}\\mathbf{B}$ is orders of magnitude smaller than $\\mathbf{R}$?\n* Line 260: What is the purpose of a zero-mean adaptation term, and why is it better?\n* Table 4: Why does the storage reduction exceed $8\\times$ (e.g., $9.8/1.2 > 8$ and $0.34/0.04 > 8$)? If the model is stored in 32 bits, the reduction should be slightly less than $8\\times$ due to the additional overhead from zero points and scaling factors. If stored in 16 bits (the preferred format), the reduction should not exceed $4\\times$.\n* Figure 6: Could you clarify the difference between VMC $\\sigma_{\\mathbf{R}}$ and the appropriate $\\sigma_{\\mathbf{R}}$? Additionally, how can one observe the large quantization error from the second image in the second row of the figure?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces IntLoRA, a method that quantizes the LoRA branch to integers for parameter-efficient fine-tuning (PEFT) in diffusion models. Traditional approaches typically use 16-bit precision for LoRA branches and focus on quantizing only the base model during fine-tuning. However, for inference, these methods often require re-quantization when the LoRA branches are fused with the base model. Unlike prior methods, such as QLoRA, IntLoRA fine-tunes the LoRA branches in a quantization-aware manner. To maintain adapter performance, it incorporates techniques like Adaptation-Quantization Separation (AQS), Multiplicative Low-rank Adaptation (MLA), and Variance Matching Control (VMC), and supports both linear and log2 quantization. Extensive experiments across various downstream tasks—including subject-driven generation, controllable generation, and style customization—demonstrate the effectiveness of IntLoRA in both visual quality and model size reduction.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "* The paper presents a strong motivation, addressing a key limitation in current quantization methods: the need for re-quantization after applying adapters. This limitation has prevented the wide adoption of quantization methods in diffusion models.\n* The concepts of Adaptation-Quantization Separation (AQS), Multiplicative Low-rank Adaptation (MLA), and Variance Matching Control (VMC) are innovative and well-supported by mathematical analysis. Their effectiveness is further validated through comprehensive experimental results.\n* The authors' support for log2 quantization, a challenging but highly hardware-efficient setting that further enhances the practical utility of IntLoRA.", "weaknesses": "* The paper exclusively focuses on weight-only quantization, where activations remain at 16-bit precision. Since diffusion models are primarily compute-bound, quantizing only the weights does not yield speedups during inference. Additionally, in a weight-only quantization setting, users can choose not to fuse adapters, thereby avoiding re-quantization. As a result, the primary benefit of IntLoRA is just a reduction in adapter size compared to conventional PEFT methods—a relatively minor contribution, given that LoRA adapters are already small.\n* he proposed PEFT method does not reduce training memory usage compared to existing PEFT methods (e.g., QLoRA). Furthermore, quantization-aware training of adapters may slow down training and even increase memory usage. The paper should discuss the training speed and memory usage.\n* Clarification on storage requirements for $\\mathbf{R}$ is necessary. Storing $\\mathbf{R}$ after training the adapters would require significant storage, which may limit the method's practicality. If $\\mathbf{R}$ is not stored, the authors should explain how it is obtained during inference. Also, note that $\\mathbf{R}$ should derive from $\\mathbf{W}_{\\text{round}}$ rather than $\\mathbf{W}$, as only quantized weights are accessible during PEFT.\n* Writing:\n  * Using \"FP\" to refer to original LoRA training is ambiguous, as there are low-precision formats like FP8 or FP4. Consider using \"16-bit\" instead. Note also that QLoRA uses NF4 for weight representation, which is not integer-based. Therefore, Lines 190–202 need to be rewritten.\n  * Line 156: The dimension of $\\mathbf{B}$ should be $d \\times C_{\\text{in}}$.\n  * Figure 2: The caption lacks detail, and the figure is difficult to interpret.\n  * Line 229: Refer to Section 5.3 on how to determine $\\mathbf{R}$.\n  * Line 265: Need to clarify $\\rho$ is the correlation coefficient.\n  * Table 4: Correct \"FP addtion\" to \"FP addition\".\n  * Figure 7: Add axis labels for clarity.", "questions": "* Line 252: Could you provide any insights or justification for the assumption that $\\mathbf{A}\\mathbf{B}$ is orders of magnitude smaller than $\\mathbf{R}$?\n* Line 260: What is the purpose of a zero-mean adaptation term, and why is it better?\n* Table 4: Why does the storage reduction exceed $8\\times$ (e.g., $9.8/1.2 > 8$ and $0.34/0.04 > 8$)? If the model is stored in 32 bits, the reduction should be slightly less than $8\\times$ due to the additional overhead from zero points and scaling factors. If stored in 16 bits (the preferred format), the reduction should not exceed $4\\times$.\n* Figure 6: Could you clarify the difference between VMC $\\sigma_{\\mathbf{R}}$ and the appropriate $\\sigma_{\\mathbf{R}}$? Additionally, how can one observe the large quantization error from the second image in the second row of the figure?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730604566933}, {"id": "LJhLQ8h5lr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2140/Reviewer_SB93"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces IntLoRA, a pioneering method for adapting quantized diffusion models to optimize both training and inference efficiencies. It tackles the arithmetic inconsistency between low-rank parameters and pre-trained model weights by employing integer-based low-rank parameters. During training, the method applies quantization to the pre-trained model weights, and post-training, integrates these integer low-rank parameters seamlessly for efficient inference. To facilitate quantization-aware training, the authors propose the Adaptation-Quantization Separation (AQS), which allows for the coexistence of zero-initialized gradients and a quantization-friendly distribution. Additionally, a Variance Matching Control (VMC) mechanism is developed to finely tune the channel-aware variance, ensuring precise adaptation control.  Extensive experimental results indicate that IntLoRA not only competes with but often surpasses traditional methods in performance, significantly reducing memory usage and computational demands.", "review_text": "The paper introduces IntLoRA, a pioneering method for adapting quantized diffusion models to optimize both training and inference efficiencies. It tackles the arithmetic inconsistency between low-rank parameters and pre-trained model weights by employing integer-based low-rank parameters. During training, the method applies quantization to the pre-trained model weights, and post-training, integrates these integer low-rank parameters seamlessly for efficient inference. To facilitate quantization-aware training, the authors propose the Adaptation-Quantization Separation (AQS), which allows for the coexistence of zero-initialized gradients and a quantization-friendly distribution. Additionally, a Variance Matching Control (VMC) mechanism is developed to finely tune the channel-aware variance, ensuring precise adaptation control.  Extensive experimental results indicate that IntLoRA not only competes with but often surpasses traditional methods in performance, significantly reducing memory usage and computational demands.", "strengths": "The paper demonstrates a strong and clear motivation by addressing a critical challenge in the domain of machine learning models—specifically, the adaptation of quantized diffusion models using integer low-rank parameters. This innovative approach effectively bridges the gap between maintaining precision and enhancing computational efficiency. Furthermore, the paper successfully resolves the prevalent issue of arithmetic inconsistency between pre-trained weights and adaptation parameters.", "weaknesses": "1.\tThe clarity of some explanations could be improved, particularly concerning why the all-zero distribution is not quantization-friendly. The text should elucidate why such a distribution necessitates a distinct quantizer design at the beginning of tuning. Providing a detailed example or analytical explanation on gradients would greatly enhance understanding, highlighting the specific challenges and potential impacts on the model’s performance during quantization.\n\n2.\tIn the description of the VMC, the authors introduce a scalar $\\alpha$ used as an exponent of $r$ to finely adjust the search for the optimal auxiliary matrix. However, the method for conducting this search remains unclear. It would be beneficial for the authors to specify the objective function used in this search process. Clarifying how $\\alpha$ and $r$ are optimized, including the criteria and algorithms employed, would provide a more comprehensive understanding of the method’s effectiveness and implementation.", "questions": "1. The proposed method, IntLoRA, demonstrates substantial potential to enhance the performance of quantized models beyond simple adaptation. Integrating IntLoRA’s innovative techniques, such as integer low-rank parameters and network-wide reconstruction calibration, not only maintains but also significantly boosts the efficacy of models operating under quantization constraints.\n\n2. The authors introduce the use of an auxiliary matrix in the AQS. It would be helpful to clarify whether the introduction of this auxiliary matrix introduces additional quantization difficulty like outliers to the pre-trained model weights.\n\n3. The paper overlooks an important baseline in its experimental setup or discussion. The study “EfficientDM” by He et al., 2023, which similarly employs low-rank parameters for fine-tuning pre-trained models, is notably absent. Like IntLoRA, EfficientDM also allows for the merging of low-rank parameters into pre-trained model weights post-training. Including this baseline could enhance the comparative analysis, providing a clearer benchmark of IntLoRA’s performance against existing methodologies with similar strategies.\n\n4. The authors merge integer low-rank parameters with quantized pre-trained model weights after training, a process that might lead to value overflow. It would be beneficial for the paper to address how this potential issue is managed. Specifically, detailing the mechanisms or techniques employed to prevent or mitigate overflow—such as scaling factors, normalization methods, or safeguards within the merging algorithm—would provide a clearer understanding of the reliability of the proposed method in handling data integrity during integration.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces IntLoRA, a pioneering method for adapting quantized diffusion models to optimize both training and inference efficiencies. It tackles the arithmetic inconsistency between low-rank parameters and pre-trained model weights by employing integer-based low-rank parameters. During training, the method applies quantization to the pre-trained model weights, and post-training, integrates these integer low-rank parameters seamlessly for efficient inference. To facilitate quantization-aware training, the authors propose the Adaptation-Quantization Separation (AQS), which allows for the coexistence of zero-initialized gradients and a quantization-friendly distribution. Additionally, a Variance Matching Control (VMC) mechanism is developed to finely tune the channel-aware variance, ensuring precise adaptation control.  Extensive experimental results indicate that IntLoRA not only competes with but often surpasses traditional methods in performance, significantly reducing memory usage and computational demands.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper demonstrates a strong and clear motivation by addressing a critical challenge in the domain of machine learning models—specifically, the adaptation of quantized diffusion models using integer low-rank parameters. This innovative approach effectively bridges the gap between maintaining precision and enhancing computational efficiency. Furthermore, the paper successfully resolves the prevalent issue of arithmetic inconsistency between pre-trained weights and adaptation parameters.", "weaknesses": "1.\tThe clarity of some explanations could be improved, particularly concerning why the all-zero distribution is not quantization-friendly. The text should elucidate why such a distribution necessitates a distinct quantizer design at the beginning of tuning. Providing a detailed example or analytical explanation on gradients would greatly enhance understanding, highlighting the specific challenges and potential impacts on the model’s performance during quantization.\n\n2.\tIn the description of the VMC, the authors introduce a scalar $\\alpha$ used as an exponent of $r$ to finely adjust the search for the optimal auxiliary matrix. However, the method for conducting this search remains unclear. It would be beneficial for the authors to specify the objective function used in this search process. Clarifying how $\\alpha$ and $r$ are optimized, including the criteria and algorithms employed, would provide a more comprehensive understanding of the method’s effectiveness and implementation.", "questions": "1. The proposed method, IntLoRA, demonstrates substantial potential to enhance the performance of quantized models beyond simple adaptation. Integrating IntLoRA’s innovative techniques, such as integer low-rank parameters and network-wide reconstruction calibration, not only maintains but also significantly boosts the efficacy of models operating under quantization constraints.\n\n2. The authors introduce the use of an auxiliary matrix in the AQS. It would be helpful to clarify whether the introduction of this auxiliary matrix introduces additional quantization difficulty like outliers to the pre-trained model weights.\n\n3. The paper overlooks an important baseline in its experimental setup or discussion. The study “EfficientDM” by He et al., 2023, which similarly employs low-rank parameters for fine-tuning pre-trained models, is notably absent. Like IntLoRA, EfficientDM also allows for the merging of low-rank parameters into pre-trained model weights post-training. Including this baseline could enhance the comparative analysis, providing a clearer benchmark of IntLoRA’s performance against existing methodologies with similar strategies.\n\n4. The authors merge integer low-rank parameters with quantized pre-trained model weights after training, a process that might lead to value overflow. It would be beneficial for the paper to address how this potential issue is managed. Specifically, detailing the mechanisms or techniques employed to prevent or mitigate overflow—such as scaling factors, normalization methods, or safeguards within the merging algorithm—would provide a clearer understanding of the reliability of the proposed method in handling data integrity during integration.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730429947036}, {"id": "CKtZ2ScOWf", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2140/Reviewer_c58Y"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces IntLoRA, which uses integral low-rank parameters to adapt quantized diffusion models. First, Adaptation-quantization separation (AQS) is proposed to overcome the difficulty of vanilla zero-initialized LoRA parameters is not quantization-friendly. Then, Multiplicative low-rank adaptation is proposed to enable the fusion of quantized weights and LoRA parameters. Variance matching control is further proposed to make the distribution suitable for log2-quantization. Experimental results show that IntLoRA improves the efficiency of adaptation while maintaining performance.", "review_text": "The paper introduces IntLoRA, which uses integral low-rank parameters to adapt quantized diffusion models. First, Adaptation-quantization separation (AQS) is proposed to overcome the difficulty of vanilla zero-initialized LoRA parameters is not quantization-friendly. Then, Multiplicative low-rank adaptation is proposed to enable the fusion of quantized weights and LoRA parameters. Variance matching control is further proposed to make the distribution suitable for log2-quantization. Experimental results show that IntLoRA improves the efficiency of adaptation while maintaining performance.", "strengths": "1.\tReducing storage costs is significant for adaptation tasks.\n2.\tThe idea of multiplicative low-rank adaptation is novel.", "weaknesses": "1.\tThe training costs are not reported in the paper.\n2.\tIn section 5.2, the authors mentioned that “IntLoRA is the only one that achieves INT type adaptation weights”, which is not true since EfficientDM also uses quantized LoRA weights.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces IntLoRA, which uses integral low-rank parameters to adapt quantized diffusion models. First, Adaptation-quantization separation (AQS) is proposed to overcome the difficulty of vanilla zero-initialized LoRA parameters is not quantization-friendly. Then, Multiplicative low-rank adaptation is proposed to enable the fusion of quantized weights and LoRA parameters. Variance matching control is further proposed to make the distribution suitable for log2-quantization. Experimental results show that IntLoRA improves the efficiency of adaptation while maintaining performance.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1.\tReducing storage costs is significant for adaptation tasks.\n2.\tThe idea of multiplicative low-rank adaptation is novel.", "weaknesses": "1.\tThe training costs are not reported in the paper.\n2.\tIn section 5.2, the authors mentioned that “IntLoRA is the only one that achieves INT type adaptation weights”, which is not true since EfficientDM also uses quantized LoRA weights.", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730356025035}], "openreview_url": "https://openreview.net/forum?id=f4mQ2SU5tp", "arxiv_id": "2410.21759", "paper_pdf": "papers/f4mQ2SU5tp.pdf", "paper_pdf_sha256": "492452adf5049baefb33b9651cbcacfebb16a8a6dfe5b92d6b8ab6c43ff447f0", "paper_pdf_bytes": 4691337, "paper_pdf_source": "openreview", "code_url": "https://github.com/csguoh/IntLoRA", "code_repository": "csguoh/IntLoRA", "code_commit": "65a8257a4311e0feab9b33477c9748d13d8ce17b", "code_archive": "repos/f4mQ2SU5tp.zip", "code_archive_sha256": "10ce454be8b0b948d631cf377c4b144f9adeca2d32df503b960660c75ac62578", "code_archive_bytes": 367037, "code_file_count": 9, "code_extensions": {".py": 8, ".sh": 1}, "github_disk_usage_kb": 358, "github_languages": {"Python": 99330, "Shell": 8177}, "github_archived": false, "github_pushed_at": "2024-11-25T04:23:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/intlora-integral-low-rank-adaptation-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "17ZbByq95E", "year": 2024, "status": "rejected", "title": "Memory-Efficient Backpropagation through Large Linear Layers", "authors": ["Daniel Bershatsky", "Aleksandr Mikhalev", "Aleksandr Katrutsa", "Julia Gusak", "Daniil Merkulov", "Ivan Oseledets"], "authorids": ["~Daniel_Bershatsky1", "~Aleksandr_Mikhalev1", "~Aleksandr_Katrutsa1", "~Julia_Gusak1", "~Daniil_Merkulov1", "~Ivan_Oseledets1"], "authors_source": "OpenReview API", "abstract": "In modern neural networks like Transformers, linear layers require significant memory to store activations during backward pass. This study proposes a memory reduction approach to perform backpropagation through linear layers. Since the gradients of linear layers are computed by matrix multiplications, we consider methods for randomized matrix multiplications and demonstrate that they require less memory with a moderate decrease of the test accuracy. Also, we investigate the variance of the gradient estimate induced by the randomized matrix multiplication. We compare this variance with the variance coming from gradient estimation based on the batch of samples. We demonstrate the benefits of the proposed method on the fine-tuning of the pretrained RoBERTa model on GLUE tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "G87jRpwEnS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5200/Reviewer_aWYY"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The presented paper proposes a stochastic estimator for the gradient of a linear transform. The proposed estimator is unbiased and relies on a randomized algorithm for matrix multiplication. In practice, this results in the introduction of a randomized linear layer where instead of storing the full batch of inputs during the forward pass, a random projection of the input tensor onto a lower dimensional space is stored for the backward pass. This results in memory savings together with a tradeoff on the variance of the resulting gradient estimate. The authors derive a theoretical analysis of the additional variance introduced by the randomized matrix multiplication and conduct experiments on fine-tuning tasks for NLP applications.\n\nThe presented method shares strong similarities with previous work on randomized matrix multiplication in the context of deep learning [1].\n\n[1] EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression", "review_text": "The presented paper proposes a stochastic estimator for the gradient of a linear transform. The proposed estimator is unbiased and relies on a randomized algorithm for matrix multiplication. In practice, this results in the introduction of a randomized linear layer where instead of storing the full batch of inputs during the forward pass, a random projection of the input tensor onto a lower dimensional space is stored for the backward pass. This results in memory savings together with a tradeoff on the variance of the resulting gradient estimate. The authors derive a theoretical analysis of the additional variance introduced by the randomized matrix multiplication and conduct experiments on fine-tuning tasks for NLP applications.\n\nThe presented method shares strong similarities with previous work on randomized matrix multiplication in the context of deep learning [1].\n\n[1] EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression", "strengths": "1) The authors provide a theoretical analysis of the additional variance induced by the randomized matrix multiplication procedure.\n2) The authors reports practical savings observed in the fine-tuning tasks with respect to their implementation.", "weaknesses": "1) The paper lacks comparison to other works aiming at reducing memory requirements during training. For example, both [2] and [3] compress activations with a quantization procedure. [1] jointly uses randomized matrix multiplications and quantization to compress the stored activations. The lack of comparison with these methods makes it difficult to evaluate the relevance of the contribution against these alternatives. These papers aren't mentioned in the related work section.\n2) The EXACT method [1] stores a random projection of the input where the compression is applied to feature channels. In the presented work, the compression is applied to the batch dimension. More generally, it would be interesting to know the practical importance of the dimension chosen for compression in the presented method.\n3) In a fully-connected layer, the weights themselves have a substantial memory footprint, which explains why retaining only 10% of the input dimensionality only yields 25% memory savings. For convolutional architectures, however, the weight space is much smaller, meaning that the relative savings would be much higher. It would be interesting to know how the proposed method translates in this setting.\n4) Given that transformers are now widely adopted in both NLP and computer vision, it would also be interesting to know how the proposed methods perform for a ViT on CIFAR10 and ImageNet, which would enrich the experimental section with minimal coding efforts.\n5) [Minor comment] If the RMM forward block sends $W, b, S^{\\top}X$ then the FC forward block should send $W, b, X$ instead of $W, b$.\n\n[2] ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training\n[3] GACT: Activation Compressed Training for Generic Network Architectures", "questions": "1) Could the author provides comparison with other works based on quantization of the activations?\n2) Could the author provide experiments on convolutional architectures or on computer vision benchmarks with a transformer-based architecture?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The presented paper proposes a stochastic estimator for the gradient of a linear transform. The proposed estimator is unbiased and relies on a randomized algorithm for matrix multiplication. In practice, this results in the introduction of a randomized linear layer where instead of storing the full batch of inputs during the forward pass, a random projection of the input tensor onto a lower dimensional space is stored for the backward pass. This results in memory savings together with a tradeoff on the variance of the resulting gradient estimate. The authors derive a theoretical analysis of the additional variance introduced by the randomized matrix multiplication and conduct experiments on fine-tuning tasks for NLP applications.\n\nThe presented method shares strong similarities with previous work on randomized matrix multiplication in the context of deep learning [1].\n\n[1] EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1) The authors provide a theoretical analysis of the additional variance induced by the randomized matrix multiplication procedure.\n2) The authors reports practical savings observed in the fine-tuning tasks with respect to their implementation.", "weaknesses": "1) The paper lacks comparison to other works aiming at reducing memory requirements during training. For example, both [2] and [3] compress activations with a quantization procedure. [1] jointly uses randomized matrix multiplications and quantization to compress the stored activations. The lack of comparison with these methods makes it difficult to evaluate the relevance of the contribution against these alternatives. These papers aren't mentioned in the related work section.\n2) The EXACT method [1] stores a random projection of the input where the compression is applied to feature channels. In the presented work, the compression is applied to the batch dimension. More generally, it would be interesting to know the practical importance of the dimension chosen for compression in the presented method.\n3) In a fully-connected layer, the weights themselves have a substantial memory footprint, which explains why retaining only 10% of the input dimensionality only yields 25% memory savings. For convolutional architectures, however, the weight space is much smaller, meaning that the relative savings would be much higher. It would be interesting to know how the proposed method translates in this setting.\n4) Given that transformers are now widely adopted in both NLP and computer vision, it would also be interesting to know how the proposed methods perform for a ViT on CIFAR10 and ImageNet, which would enrich the experimental section with minimal coding efforts.\n5) [Minor comment] If the RMM forward block sends $W, b, S^{\\top}X$ then the FC forward block should send $W, b, X$ instead of $W, b$.\n\n[2] ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training\n[3] GACT: Activation Compressed Training for Generic Network Architectures", "questions": "1) Could the author provides comparison with other works based on quantization of the activations?\n2) Could the author provide experiments on convolutional architectures or on computer vision benchmarks with a transformer-based architecture?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699043767696}, {"id": "C91enBLy7V", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5200/Reviewer_fccg"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work presents a memory-efficient method for computing the needed gradients in backpropagation training through large linear layers.\nThe method uses the randomized matrix multiplication approach to achieve this.\nIt randomly creates a matrix S that projects the signal X into a smaller dimension space.\nThe reduction is controlled by $\\rho$, which takes up a value between 0 and 1.\nThe only required condition for S is that its autocorrelation equals the identity matrix I.\n\nThis work further analyzed the gradient variance induced by the randomized computation.\nIt also provides a detailed explanation of how the new approach reduces the memory requirement for backpropagation.\nThe work provides simulation results to support the main claim.", "review_text": "This work presents a memory-efficient method for computing the needed gradients in backpropagation training through large linear layers.\nThe method uses the randomized matrix multiplication approach to achieve this.\nIt randomly creates a matrix S that projects the signal X into a smaller dimension space.\nThe reduction is controlled by $\\rho$, which takes up a value between 0 and 1.\nThe only required condition for S is that its autocorrelation equals the identity matrix I.\n\nThis work further analyzed the gradient variance induced by the randomized computation.\nIt also provides a detailed explanation of how the new approach reduces the memory requirement for backpropagation.\nThe work provides simulation results to support the main claim.", "strengths": "The paper is well organized and the literature review provides enough context needed to understand this work.\nThe topic is an interesting one that is worth pursuing, especially in the era of ever-growing large models that rely on transformer modules.", "weaknesses": "One of the suggested claims (under section 2.3) is not well-supported.\nThis work seems to suggest that the difference between the exact gradient and the randomized approximation is a form of noise injection.\nIt went on to state that it can play the role of a regularizer.\nYou will need to model the effect of the randomized error as such to make this claim. \nI do not think this paper contains such an explanation. \n\nThe paper states that the only requirement for the random matrix S is that its autocorrelation be equal to I.\nBut the definition of the random Gaussian random matrix suggests that the autocorrelation is  $\\frac{1}{B_{proj}} I$.\nI think you should align the two.\nYou can treat the constant as a scale on the learning rate.", "questions": "1). How will the sample size of $S$ affect the performance of this method?\nI think the algorithm statement suggests that you only picked one sample of S.\nIf we compute the gradient over multiple samples of $S$ and take the average, will this improve the performance or not? \nThis follows from equation (4), which relies on the expectation of the randomized gradients over $S$\nIt will be interesting to see the trade-off between the additional workload and the \n \n2). How does the depth of the network affect the variance and error induced by the randomized gradients?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a memory-efficient method for computing the needed gradients in backpropagation training through large linear layers.\nThe method uses the randomized matrix multiplication approach to achieve this.\nIt randomly creates a matrix S that projects the signal X into a smaller dimension space.\nThe reduction is controlled by $\\rho$, which takes up a value between 0 and 1.\nThe only required condition for S is that its autocorrelation equals the identity matrix I.\n\nThis work further analyzed the gradient variance induced by the randomized computation.\nIt also provides a detailed explanation of how the new approach reduces the memory requirement for backpropagation.\nThe work provides simulation results to support the main claim.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is well organized and the literature review provides enough context needed to understand this work.\nThe topic is an interesting one that is worth pursuing, especially in the era of ever-growing large models that rely on transformer modules.", "weaknesses": "One of the suggested claims (under section 2.3) is not well-supported.\nThis work seems to suggest that the difference between the exact gradient and the randomized approximation is a form of noise injection.\nIt went on to state that it can play the role of a regularizer.\nYou will need to model the effect of the randomized error as such to make this claim. \nI do not think this paper contains such an explanation. \n\nThe paper states that the only requirement for the random matrix S is that its autocorrelation be equal to I.\nBut the definition of the random Gaussian random matrix suggests that the autocorrelation is  $\\frac{1}{B_{proj}} I$.\nI think you should align the two.\nYou can treat the constant as a scale on the learning rate.", "questions": "1). How will the sample size of $S$ affect the performance of this method?\nI think the algorithm statement suggests that you only picked one sample of S.\nIf we compute the gradient over multiple samples of $S$ and take the average, will this improve the performance or not? \nThis follows from equation (4), which relies on the expectation of the randomized gradients over $S$\nIt will be interesting to see the trade-off between the additional workload and the \n \n2). How does the depth of the network affect the variance and error induced by the randomized gradients?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698736272647}, {"id": "IHxLQYUTHt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5200/Reviewer_k3XT"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This article proposes a way to reduce the memory cost of the backpropagation of linear layers, with Randomized Matrix Multiplication.", "review_text": "This article proposes a way to reduce the memory cost of the backpropagation of linear layers, with Randomized Matrix Multiplication.", "strengths": "This article is relatively well-written and easy to follow.\nThe RMM algorithm is simple and well-motivated, lowering immediately the memory cost of backward operations on linear layers in DNN.\nThe authors bring a novel analysis of variance in both SGD and RMM.", "weaknesses": "**Novelty** The novelty of the method is very limited. The main idea is already present in Adelman et al. (2021), even though their goal was not to save the memory during the stage. The authors do not bring many contributions over Adelman, the theoretical analysis of the variance being unclear in particular. \n\n**Variance** $D_{SGD}$ and $D_{RMM}$  are not defined clearly (or in equations (10) and (11), is $D_{RMM}=D$?). In the Appendix, the equation (21) $D^2_{SGD}(X,Y)= \\frac 1 {B−1} D^2 _Z(X,Y)$ is not clear.\n\nMore generally, it is quite unclear what should be the conclusion of the analysis of the variances of SGD and RMM in Sections 2.3 and 3.3. The theoretical analysis gives an upper bound of the variance of RMM, but we observe that in practice the variance of RMM is much higher than the one of SGD, meaning that RMM brings a lot of noise to the training, which seems to hinder training. What is the conclusion of these sections? I also would at least have appreciated the theoretical upper bound on Figure 3.\n\n**Computation cost** The computation requirements should be moved from the appendix to the main paper and discussed there, considering their importance. In particular, the comparison between the computation costs $O(BN^2)$ and $O(ρBN(B+ N))$ is important, as it showcases that the method may be slower, an important drawback that is confirmed in Table 3. For such a high computation time cost, other methods that compress the activation vector $X$ will be favored.\n\n**Pretrained network** Experiments being done on a pre-trained network do not allow a clear measurement of the degradation due to RMM during training.\n\nSection 3.4 brings nothing new compared to Table 2.", "questions": "\"In Adelman et al. (2021) the construction of S requires the knowledge of the norms of the rows of Y\". I may be mistaken but I do not understand why this is the case. Bernoulli-CRS is unbiased whatever the values of $p_{j}$ are. Therefore it can be applied to RMM with no knowledge of $Y$ for any distribution $p_j$.\n\nThe results of Section 3.3 seem surprising. Why is the SGD variance increasing during training, while the network should converge?\n\nWhy say that \"Subsampled Orthonormal with Random Signs (SORS)\" is not considered if there is a Comparison of Randomized MatMuls in Section 3.5?\nConsidering the cost of computation of RMM, why should we choose Gaussian RMM rather than DCT or DFT, which the authors note have a much smaller computation cost?\n\n\"Moreover, we empirically found that our randomization is faster if ρ ≤ 0.1.\" Doesn't Table 3 show that even with $\\rho=0.1$ the randomization is still longer than no RMM?\n\nVarious remarks and errors:\n* Section 2.2: \"this should be?\", \"target really accurate\", \"it may seem\"\n* Section 3 \"We rewrite implementation\". PRNG stage $G$ is only used here. \" compression in 5–10 times\" \"study influence of randomized\" \"However, overfitting point\"\n*", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This article proposes a way to reduce the memory cost of the backpropagation of linear layers, with Randomized Matrix Multiplication.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "This article is relatively well-written and easy to follow.\nThe RMM algorithm is simple and well-motivated, lowering immediately the memory cost of backward operations on linear layers in DNN.\nThe authors bring a novel analysis of variance in both SGD and RMM.", "weaknesses": "**Novelty** The novelty of the method is very limited. The main idea is already present in Adelman et al. (2021), even though their goal was not to save the memory during the stage. The authors do not bring many contributions over Adelman, the theoretical analysis of the variance being unclear in particular. \n\n**Variance** $D_{SGD}$ and $D_{RMM}$  are not defined clearly (or in equations (10) and (11), is $D_{RMM}=D$?). In the Appendix, the equation (21) $D^2_{SGD}(X,Y)= \\frac 1 {B−1} D^2 _Z(X,Y)$ is not clear.\n\nMore generally, it is quite unclear what should be the conclusion of the analysis of the variances of SGD and RMM in Sections 2.3 and 3.3. The theoretical analysis gives an upper bound of the variance of RMM, but we observe that in practice the variance of RMM is much higher than the one of SGD, meaning that RMM brings a lot of noise to the training, which seems to hinder training. What is the conclusion of these sections? I also would at least have appreciated the theoretical upper bound on Figure 3.\n\n**Computation cost** The computation requirements should be moved from the appendix to the main paper and discussed there, considering their importance. In particular, the comparison between the computation costs $O(BN^2)$ and $O(ρBN(B+ N))$ is important, as it showcases that the method may be slower, an important drawback that is confirmed in Table 3. For such a high computation time cost, other methods that compress the activation vector $X$ will be favored.\n\n**Pretrained network** Experiments being done on a pre-trained network do not allow a clear measurement of the degradation due to RMM during training.\n\nSection 3.4 brings nothing new compared to Table 2.", "questions": "\"In Adelman et al. (2021) the construction of S requires the knowledge of the norms of the rows of Y\". I may be mistaken but I do not understand why this is the case. Bernoulli-CRS is unbiased whatever the values of $p_{j}$ are. Therefore it can be applied to RMM with no knowledge of $Y$ for any distribution $p_j$.\n\nThe results of Section 3.3 seem surprising. Why is the SGD variance increasing during training, while the network should converge?\n\nWhy say that \"Subsampled Orthonormal with Random Signs (SORS)\" is not considered if there is a Comparison of Randomized MatMuls in Section 3.5?\nConsidering the cost of computation of RMM, why should we choose Gaussian RMM rather than DCT or DFT, which the authors note have a much smaller computation cost?\n\n\"Moreover, we empirically found that our randomization is faster if ρ ≤ 0.1.\" Doesn't Table 3 show that even with $\\rho=0.1$ the randomization is still longer than no RMM?\n\nVarious remarks and errors:\n* Section 2.2: \"this should be?\", \"target really accurate\", \"it may seem\"\n* Section 3 \"We rewrite implementation\". PRNG stage $G$ is only used here. \" compression in 5–10 times\" \"study influence of randomized\" \"However, overfitting point\"\n*", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698692018632}, {"id": "ivigjXjyWf", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5200/Reviewer_JVog"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The proposed RMM utilizes random projection to reduce the memory required to store input activations of the linear layer while still achieving comparable accuracy.", "review_text": "The proposed RMM utilizes random projection to reduce the memory required to store input activations of the linear layer while still achieving comparable accuracy.", "strengths": "This paper is \n\n1. well-written, \n\n2. provides justification for the RMM method through a comparison of gradient variance with SGD, and \n\n3. demonstrates the usefulness of the algorithm in RoBERTa.", "weaknesses": "1. Lack of experiemntal results\n\n: The paper claims the efficacy of the method for large linear layers, but only presents experimental results from the GLUE benchmark of RoBERTa. A performance comparison with larger models, such as LLaMA-2-7B with over 7B parameters, seems necessary.\n\n2. Lack of comparison between previous works\n\n: In fact, there have been many attempts in the past to save memory through activation compression. Notably, GACT (2022) is capable of compressing all input activations, including those of the linear layer, to an average of 4-bit in BERT-Large. A comparison between such existing methods and RMM seems necessary.\n\n3. Lack of novelty\n\n: As the author pointed out, random projection has long been a widely used method, and applying it to activation compression doesn't necessarily mean it lacks novelty. However, there have already been cases using random projection under the same objective of activation compression. EXACT (2022) compresses activations using random projection, which seems strikingly similar to the proposed RMM.", "questions": "1. Lack of experiemntal results\n\n: Please provide experimental results on larger models like LLaMA-2-7B. This is crucial in bolstering the claims made in the paper.\n\n2. Comparision with previous works\n\n: A quantitative experimental comparison with recently proposed activation compression algorithms like GACT (2022), EXACT (2022), AAL (2023), and DropIT (2023) seems necessary. If the review period is limited, please at least provide a comparison with GACT (2022), which is the-state-of-art activation compression technique,  in terms of 1) accuracy and 2) memory saving through experimental data. For the remaining algorithms, a qualitative analysis can be included in the related work section.\n\n3. Lack of novelty\n\n: A detailed comparison with EXACT, which also uses random projection, seems necessary. To offer some advice, 1) Unlike EXACT, RMM appears to compress the batch dimension. Demonstrating mathematically that this difference allows RMM to further reduce gradient variance, or 2) showing through actual experimental results that RMM offers a significant accuracy improvement over EXACT, would likely address concerns regarding novelty.\n\nHowever, one concern is that, based on the comparison results in the Appendix for GCN2, EXACT seems to achieve higher accuracy than RMM. Given that both algorithms adopt very similar approaches, the accuracy of RMM must necessarily be higher than that of the existing EXACT.\n\n* GACT: Activation Compressed Training for Generic Network Architectures (2022)\n* EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression  (2022)\n* DropIT: Dropping Intermediate Tensors for Memory-Efficient DNN Training (2023)\n* Learning with Auxiliary Activation for Memory-Efficient Training  (2023)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The proposed RMM utilizes random projection to reduce the memory required to store input activations of the linear layer while still achieving comparable accuracy.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "This paper is \n\n1. well-written, \n\n2. provides justification for the RMM method through a comparison of gradient variance with SGD, and \n\n3. demonstrates the usefulness of the algorithm in RoBERTa.", "weaknesses": "1. Lack of experiemntal results\n\n: The paper claims the efficacy of the method for large linear layers, but only presents experimental results from the GLUE benchmark of RoBERTa. A performance comparison with larger models, such as LLaMA-2-7B with over 7B parameters, seems necessary.\n\n2. Lack of comparison between previous works\n\n: In fact, there have been many attempts in the past to save memory through activation compression. Notably, GACT (2022) is capable of compressing all input activations, including those of the linear layer, to an average of 4-bit in BERT-Large. A comparison between such existing methods and RMM seems necessary.\n\n3. Lack of novelty\n\n: As the author pointed out, random projection has long been a widely used method, and applying it to activation compression doesn't necessarily mean it lacks novelty. However, there have already been cases using random projection under the same objective of activation compression. EXACT (2022) compresses activations using random projection, which seems strikingly similar to the proposed RMM.", "questions": "1. Lack of experiemntal results\n\n: Please provide experimental results on larger models like LLaMA-2-7B. This is crucial in bolstering the claims made in the paper.\n\n2. Comparision with previous works\n\n: A quantitative experimental comparison with recently proposed activation compression algorithms like GACT (2022), EXACT (2022), AAL (2023), and DropIT (2023) seems necessary. If the review period is limited, please at least provide a comparison with GACT (2022), which is the-state-of-art activation compression technique,  in terms of 1) accuracy and 2) memory saving through experimental data. For the remaining algorithms, a qualitative analysis can be included in the related work section.\n\n3. Lack of novelty\n\n: A detailed comparison with EXACT, which also uses random projection, seems necessary. To offer some advice, 1) Unlike EXACT, RMM appears to compress the batch dimension. Demonstrating mathematically that this difference allows RMM to further reduce gradient variance, or 2) showing through actual experimental results that RMM offers a significant accuracy improvement over EXACT, would likely address concerns regarding novelty.\n\nHowever, one concern is that, based on the comparison results in the Appendix for GCN2, EXACT seems to achieve higher accuracy than RMM. Given that both algorithms adopt very similar approaches, the accuracy of RMM must necessarily be higher than that of the existing EXACT.\n\n* GACT: Activation Compressed Training for Generic Network Architectures (2022)\n* EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression  (2022)\n* DropIT: Dropping Intermediate Tensors for Memory-Efficient DNN Training (2023)\n* Learning with Auxiliary Activation for Memory-Efficient Training  (2023)", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698655367049}], "openreview_url": "https://openreview.net/forum?id=17ZbByq95E", "arxiv_id": "2201.13195", "paper_pdf": "papers/17ZbByq95E.pdf", "paper_pdf_sha256": "74f1ef669098cb715ce131858a8bbe5267936d9ea86d47a41cbcd8b4fac2f10a", "paper_pdf_bytes": 474137, "paper_pdf_source": "openreview", "code_url": "https://github.com/skolai/fewbit", "code_repository": "skolai/fewbit", "code_commit": "940706bbbc40b11ade61d42253f80dc960d35793", "code_archive": "repos/17ZbByq95E.zip", "code_archive_sha256": "330303eaa470a820abfde80e7eb67e3d37feaa9b41e71eef065f95da6ef2d0e8", "code_archive_bytes": 310980, "code_file_count": 53, "code_extensions": {".py": 33, ".cc": 5, ".h": 5, ".sh": 4, ".ipynb": 4, ".cu": 2}, "github_disk_usage_kb": 321, "github_languages": {"Python": 141112, "C++": 36054, "Cuda": 30862, "Shell": 5429, "CMake": 3968, "Dockerfile": 1542}, "github_archived": false, "github_pushed_at": "2023-07-26T04:42:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/memory-efficient-backpropagation-through"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "zHSaBQtj-l", "year": 2023, "status": "rejected", "title": "Differentiable Rendering with Reparameterized Volume Sampling", "authors": ["Kirill Struminsky", "Oleg Desheulin"], "authorids": ["~Kirill_Struminsky1", "~Oleg_Desheulin2"], "authors_source": "OpenReview API", "abstract": "We propose an alternative rendering algorithm for neural radiance fields based on importance sampling. In view synthesis, a neural radiance field approximates underlying density and radiance fields based on a sparse set of views of a scene. To generate a pixel of a novel view, it marches a ray through the pixel and computes a weighted sum of radiance emitted from a dense set of ray points. This rendering algorithm is fully differentiable and facilitates gradient-based optimization of the fields. However, in practice, only a tiny opaque portion of the ray contributes most of the radiance to the sum. Therefore, we can avoid computing radiance in the rest part. In this work, we use importance sampling to pick non-transparent points on the ray. Specifically, we generate samples according to the probability distribution induced by the density field. Our main contribution is the reparameterization of the sampling algorithm. It allows end-to-end learning with gradient descent as in the original rendering algorithm. With our approach, we can optimize a neural radiance field with just a few radiance field evaluations per ray. As a result, we alleviate the costs associated with the color component of the neural radiance field.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "daUTibIM2k", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5331/Reviewer_bm7A"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes an importance sampling scheme for the radiance evaluation of NeRF-style volumetric rendering. Importance sampling replaces the coarse network in NeRF with direct sampling of the high resolution density field. This promises a simpler algorithm without the need for a proxy loss on the coarse network, and improved quality/performance, as radiance samples are concentrated in regions of high opacity, while transparent regions are skipped.\n\nThe paper describes the relevant background on volumetric rendering and outlines how the piecewise constant ray integral can be recast, first as a monte-carlo estimate in ray space, and then as as an integral over the normalized opacity along the ray.\n\nThe paper describes in detail how the approximate opacity is computed with a piecewise approximation, and how samples from the distribution are drawn via inverse transform sampling. A complication of the paper is that in order to construct an approximation of the cumulative density function, the technique does require a relatively fine (256) sampling of the density field, so any performance improvement of importance sampling would be gained solely by saving radiance evaluations.\n\nThere is a insightful synthetic experiment that compares the proposed integration scheme to uniform sampling in ray space (with ablations for stratified sampling).\n\nThe paper then presents comparisons to NeRF and a voxel-based technique (DVGO). It appears from these experiments that both baselines (full NeRF with coarse model, and an unguided NeRF without hierarchical sampling) perform better than the proposed technique.\n\n", "review_text": "I'm leaning against accepting this paper. The empirical evaluation is unfortunately not very impressive. The technical content is solid, but only addresses a narrow aspect of the overall algorithm, and does not include enough new ideas to put the paper over the bar in light of the mixed experimental results.\n", "strengths": "The paper is well written and well motivated. The description of the technical contribution is clear, and Figures 1-3 are good visualizations of the technique. \n\nThe outlined idea holds some promise, given that NeRFs coarse-then-fine scheme, despite its effectiveness, is not thoroughly analyzed in the original paper.\n\nA drawback of the proposed technique is the need to finely sample the density field to create the approximation for inverse transform sampling. This seems to preclude major performance benefits that would otherwise be associated with importance sampling.\n\nThe experiments reported in Section 5.2 suggest that the proposed importance sampling scheme just doesn't work very well in its current implementation. The reconstruction quality is consistently lower than NeRF's, even if the coarse network of NeRF is disabled. (Could the authors clarify how many radiance evaluations the NeRF models in Table 1 and 2 perform?)\n\nThe comparison to DVGO also fails to show significant performance or quality improvements.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes an importance sampling scheme for the radiance evaluation of NeRF-style volumetric rendering. Importance sampling replaces the coarse network in NeRF with direct sampling of the high resolution density field. This promises a simpler algorithm without the need for a proxy loss on the coarse network, and improved quality/performance, as radiance samples are concentrated in regions of high opacity, while transparent regions are skipped.\n\nThe paper describes the relevant background on volumetric rendering and outlines how the piecewise constant ray integral can be recast, first as a monte-carlo estimate in ray space, and then as as an integral over the normalized opacity along the ray.\n\nThe paper describes in detail how the approximate opacity is computed with a piecewise approximation, and how samples from the distribution are drawn via inverse transform sampling. A complication of the paper is that in order to construct an approximation of the cumulative density function, the technique does require a relatively fine (256) sampling of the density field, so any performance improvement of importance sampling would be gained solely by saving radiance evaluations.\n\nThere is a insightful synthetic experiment that compares the proposed integration scheme to uniform sampling in ray space (with ablations for stratified sampling).\n\nThe paper then presents comparisons to NeRF and a voxel-based technique (DVGO). It appears from these experiments that both baselines (full NeRF with coarse model, and an unguided NeRF without hierarchical sampling) perform better than the proposed technique.\n\n", "strength_and_weaknesses": "The paper is well written and well motivated. The description of the technical contribution is clear, and Figures 1-3 are good visualizations of the technique. \n\nThe outlined idea holds some promise, given that NeRFs coarse-then-fine scheme, despite its effectiveness, is not thoroughly analyzed in the original paper.\n\nA drawback of the proposed technique is the need to finely sample the density field to create the approximation for inverse transform sampling. This seems to preclude major performance benefits that would otherwise be associated with importance sampling.\n\nThe experiments reported in Section 5.2 suggest that the proposed importance sampling scheme just doesn't work very well in its current implementation. The reconstruction quality is consistently lower than NeRF's, even if the coarse network of NeRF is disabled. (Could the authors clarify how many radiance evaluations the NeRF models in Table 1 and 2 perform?)\n\nThe comparison to DVGO also fails to show significant performance or quality improvements.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written. There are some typos (\"improtance\" in the Table 2 section) and some grammatical errors, but these don't hurt understanding of the paper. The contribution is not very novel, given that NeRF does point out the need for some form of importance sampling, and implements an effective ad-hoc importance sampling-style scheme.", "summary_of_the_review": "I'm leaning against accepting this paper. The empirical evaluation is unfortunately not very impressive. The technical content is solid, but only addresses a narrow aspect of the overall algorithm, and does not include enough new ideas to put the paper over the bar in light of the mixed experimental results.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666668576902}, {"id": "f0vl_YPMTk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5331/Reviewer_yzB7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims at avoiding computing radiance in less contributive parts by reparameterizing the sampling algorithm. This can help decrease the number of evaluations to MLPs. This sounds reasonable. But the experiments do not meet the authors’ expectations. First, decreasing the MLP evaluations should indeed improve efficiency a lot. But the results show only minor improvements on NeRF baseline. This reminds us of if the proposed approach itself slows down the efficiency. Second, DVGO is an explicit volume-based approach which is extremely efficient. I have to say the author have chosen a wrong baseline. Because the efficiency of explicit volume-based approach is not severely slow down by the great number of sampling points. Third, a big number of sampling points is vital to rendering quality. So, the rendering quality of this paper degrades. Fourth, the experiments cannot support the claim. Please refer to the weakness for more details.", "review_text": "The unimpressive experiment results mostly lead to a negative opinion to this paper. \nThe claimed contribution on efficiency cannot be supported by the experiment. Some constraints (time used on reaching the same psnr, or evaluating the produced psnr using the same training time) have to be set for a fair comparison. And the rendering quality is also not good.\n", "strengths": "Strength:\n1) The mathematical analysis looks great.\n\nWeakness:\n1)\tThe proposed method can lead to improvements on efficiency but degrades on quality. But the experiments presented in the paper are hard to support the claimed contribution. It would be fair if make experiments of evaluating two methods on the time used of reaching the same PSNR. Or vice versa, evaluating the produced PNSR with the same training time. \n2)\tThe improved efficiency is very slight compared with baseline approach. Normally, a smaller number of evaluations to MLPs should lead to great efficiency improvements. But in this paper, the efficiency improvement is little. The author should analyze the reason deeply, is this caused by the extra time used on solving Eq. 14 and 10?\n3)\tThe reasons of piece wise constant has worse quality than piece wise linear are not analyzed.\n4)\tThe reasons of different performance on three experiment setups (Fig. 3) are not analyzed.\n5)\t“Notably, in our experiments, the model trained with 64 samples achieved PSNR 34.33 even with 16 during evaluation stage.” I cannot find the 34.33 in paper.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper aims at avoiding computing radiance in less contributive parts by reparameterizing the sampling algorithm. This can help decrease the number of evaluations to MLPs. This sounds reasonable. But the experiments do not meet the authors’ expectations. First, decreasing the MLP evaluations should indeed improve efficiency a lot. But the results show only minor improvements on NeRF baseline. This reminds us of if the proposed approach itself slows down the efficiency. Second, DVGO is an explicit volume-based approach which is extremely efficient. I have to say the author have chosen a wrong baseline. Because the efficiency of explicit volume-based approach is not severely slow down by the great number of sampling points. Third, a big number of sampling points is vital to rendering quality. So, the rendering quality of this paper degrades. Fourth, the experiments cannot support the claim. Please refer to the weakness for more details.", "strength_and_weaknesses": "Strength:\n1) The mathematical analysis looks great.\n\nWeakness:\n1)\tThe proposed method can lead to improvements on efficiency but degrades on quality. But the experiments presented in the paper are hard to support the claimed contribution. It would be fair if make experiments of evaluating two methods on the time used of reaching the same PSNR. Or vice versa, evaluating the produced PNSR with the same training time. \n2)\tThe improved efficiency is very slight compared with baseline approach. Normally, a smaller number of evaluations to MLPs should lead to great efficiency improvements. But in this paper, the efficiency improvement is little. The author should analyze the reason deeply, is this caused by the extra time used on solving Eq. 14 and 10?\n3)\tThe reasons of piece wise constant has worse quality than piece wise linear are not analyzed.\n4)\tThe reasons of different performance on three experiment setups (Fig. 3) are not analyzed.\n5)\t“Notably, in our experiments, the model trained with 64 samples achieved PSNR 34.33 even with 16 during evaluation stage.” I cannot find the 34.33 in paper.\n", "clarity,_quality,_novelty_and_reproducibility": "The originality is good. Clarity is fair. Quality is fair.", "summary_of_the_review": "The unimpressive experiment results mostly lead to a negative opinion to this paper. \nThe claimed contribution on efficiency cannot be supported by the experiment. Some constraints (time used on reaching the same psnr, or evaluating the produced psnr using the same training time) have to be set for a fair comparison. And the rendering quality is also not good.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666625620478}, {"id": "CUm5beNa-8F", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5331/Reviewer_XnSd"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Volume rendering is a critically important algorithm in NeRF (neural radiance field) methods, which connects 3D real points with 2D image pixels. This paper aims at exploring an alternative rendering algorithm mainly focusing on the importance sampling for integration. A reparameterized Monte Carlo estimation method is proposed to enable end-to-end optimization, where fewer samples are needed for integration to save cost.", "review_text": "The main concerns lie in experimental evaluation and comparisons. For its current state, no evident advantages over previous works are shown.", "strengths": "**Strength**  \nThe topic of designing better sampling methods for volume rendering is interesting and valuable. This paper proposes a reparameterization method to enable end-to-end differentiable optimization, which is reasonable and has been evaluated in NeRF-provided benchmarks.\n\n**Weaknesses**    \n**The biggest concerns come from the experimental evaluation and comparisons.**  \n* The proposed method shows a worse rendering quality while the speedup over the original NeRF and DVGO is not so significant. \n* Related works, e.g. NeRF-ID, DO-NeRF, AutoInt, TermiNeRF, should also be compared in experiments with regard to both rendering quality and speed. \n* Some highly related works are missing, including but not limited to [1, 2]. Please have a careful check.\n* Experimental evaluation is not so comprehensive. Real scenes, e.g. LLFF ones, need to be included.\n\n**Minor**:   \n* The citation style does not follow the ICLR-required author-year ones.\n\n[1] Fang J, Xie L, Wang X, et al. Neusample: Neural sample field for efficient view synthesis[J]. arXiv preprint arXiv:2111.15552, 2021.   \n[2] Kurz A, Neff T, Lv Z, et al. AdaNeRF: Adaptive Sampling for Real-time Rendering of Neural Radiance Fields[J]. arXiv preprint arXiv:2207.10312, 2022.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "Volume rendering is a critically important algorithm in NeRF (neural radiance field) methods, which connects 3D real points with 2D image pixels. This paper aims at exploring an alternative rendering algorithm mainly focusing on the importance sampling for integration. A reparameterized Monte Carlo estimation method is proposed to enable end-to-end optimization, where fewer samples are needed for integration to save cost.", "strength_and_weaknesses": "**Strength**  \nThe topic of designing better sampling methods for volume rendering is interesting and valuable. This paper proposes a reparameterization method to enable end-to-end differentiable optimization, which is reasonable and has been evaluated in NeRF-provided benchmarks.\n\n**Weaknesses**    \n**The biggest concerns come from the experimental evaluation and comparisons.**  \n* The proposed method shows a worse rendering quality while the speedup over the original NeRF and DVGO is not so significant. \n* Related works, e.g. NeRF-ID, DO-NeRF, AutoInt, TermiNeRF, should also be compared in experiments with regard to both rendering quality and speed. \n* Some highly related works are missing, including but not limited to [1, 2]. Please have a careful check.\n* Experimental evaluation is not so comprehensive. Real scenes, e.g. LLFF ones, need to be included.\n\n**Minor**:   \n* The citation style does not follow the ICLR-required author-year ones.\n\n[1] Fang J, Xie L, Wang X, et al. Neusample: Neural sample field for efficient view synthesis[J]. arXiv preprint arXiv:2111.15552, 2021.   \n[2] Kurz A, Neff T, Lv Z, et al. AdaNeRF: Adaptive Sampling for Real-time Rendering of Neural Radiance Fields[J]. arXiv preprint arXiv:2207.10312, 2022.\n", "clarity,_quality,_novelty_and_reproducibility": "* **Clarity**: Most contents are clear and easy to follow.\n* **Quality**: The proposed method has not yet been evaluated and studied on substantial benchmarks and settings (refer to weaknesses).\n* **Novelty**: The proposed method is somewhat new but has some points in common with previous ones, e.g. density estimation.\n* **Reproducibility**: Most implementation details are provided and a script is provided. It is likely to reproduce \nthe method easily.", "summary_of_the_review": "The main concerns lie in experimental evaluation and comparisons. For its current state, no evident advantages over previous works are shown.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666624757611}, {"id": "sa8JZbmj0Mx", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5331/Reviewer_D4iJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors proposed a rendering method for a neural radiance field. When marching on each light ray, the authors chose the sampling points using the idea of importance sampling; the samples are drawn in accordance with the contributions to the final rendered color. The method was designed so as to reduce the number of the radiance queries, which (they expected) would bring the computational advantages over the previous methods. They incorporated this sampling strategy into two baseline methods, NeRF and DIGO, and compared the performance with the original methods. ", "review_text": "As mentioned in \"Strength and Weaknesses\", I appreciate the logical explanation for the importance sampling for a neural radiance field. However, I think the authors' contribution is insufficient in the current form; they have not established the effectiveness of their proposal: computational advantages over the previous methods. Therefore, I recommend rejection for this submission.", "strengths": "Strengths\n\n- The idea of importance sampling is explained well in a logical manner. As far as I read, the theory and the related derivations are correct. Moreover, this idea was successfully implemented as a trainable module that can be incorporated into the existing neural rendering methods.\n\nWeaknesses\n\n- The authors failed to establish the effectiveness (computational advantages over the previous methods) of their proposal.  As shown in Tables 2 and 3, the authors' method achieved only a small amount (up to 25 percent) of reduction in computation time  compared to the baselines (NeRF and DVGO) at the cost of quality decline. The authors should notice the fact that the number of sampling points can be changed also for the original methods (NeRF and DVGO) so as to find a desirable trade-off between the computational cost and rendering quality. However, the authors used a fixed number of samples for the baselines,  and thus, I cannot see the advantage of the authors' proposal over the baseline methods. Moreover, the authors did not compare their proposed method against other speed-up methods for neural radiance fields. For example, DONeRF achieved a significant improvement in quality-speed trade-off: 15-78 times fewer computation and comparable rendering quality than/to the baseline NeRF. KiloNeRF achieved more than 2000 times speed up while maintaining the rendering quality. Compared to these works, the contribution of the present submission seems quite limited. \n\n- Insufficient consideration/evaluation for the method's design. The proposed method was designed to reduce the number of \"radiance queries\", but it still needs many \"density queries\" to obtain accurate opacity along the light ray. Moreover, the re-parameterization for the importance sampling introduced additional computational cost. Therefore, it seems that the authors' method cannot save the computation so much as the other speed-up methods. The authors should evaluate computational cost (e.g. in FLOP per pixel) involved in each of these steps (density queries, re-parameterization, and radiance queries) to see the problem clearer. Moreover, it seems that the authors sticked to the idea of using a single radiance field, rejecting the use of auxiliary/multiple networks. However, the use of auxiliary/multiple networks has practical advantages in terms of the speed-quality trade-off.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors proposed a rendering method for a neural radiance field. When marching on each light ray, the authors chose the sampling points using the idea of importance sampling; the samples are drawn in accordance with the contributions to the final rendered color. The method was designed so as to reduce the number of the radiance queries, which (they expected) would bring the computational advantages over the previous methods. They incorporated this sampling strategy into two baseline methods, NeRF and DIGO, and compared the performance with the original methods. ", "strength_and_weaknesses": "Strengths\n\n- The idea of importance sampling is explained well in a logical manner. As far as I read, the theory and the related derivations are correct. Moreover, this idea was successfully implemented as a trainable module that can be incorporated into the existing neural rendering methods.\n\nWeaknesses\n\n- The authors failed to establish the effectiveness (computational advantages over the previous methods) of their proposal.  As shown in Tables 2 and 3, the authors' method achieved only a small amount (up to 25 percent) of reduction in computation time  compared to the baselines (NeRF and DVGO) at the cost of quality decline. The authors should notice the fact that the number of sampling points can be changed also for the original methods (NeRF and DVGO) so as to find a desirable trade-off between the computational cost and rendering quality. However, the authors used a fixed number of samples for the baselines,  and thus, I cannot see the advantage of the authors' proposal over the baseline methods. Moreover, the authors did not compare their proposed method against other speed-up methods for neural radiance fields. For example, DONeRF achieved a significant improvement in quality-speed trade-off: 15-78 times fewer computation and comparable rendering quality than/to the baseline NeRF. KiloNeRF achieved more than 2000 times speed up while maintaining the rendering quality. Compared to these works, the contribution of the present submission seems quite limited. \n\n- Insufficient consideration/evaluation for the method's design. The proposed method was designed to reduce the number of \"radiance queries\", but it still needs many \"density queries\" to obtain accurate opacity along the light ray. Moreover, the re-parameterization for the importance sampling introduced additional computational cost. Therefore, it seems that the authors' method cannot save the computation so much as the other speed-up methods. The authors should evaluate computational cost (e.g. in FLOP per pixel) involved in each of these steps (density queries, re-parameterization, and radiance queries) to see the problem clearer. Moreover, it seems that the authors sticked to the idea of using a single radiance field, rejecting the use of auxiliary/multiple networks. However, the use of auxiliary/multiple networks has practical advantages in terms of the speed-quality trade-off.", "clarity,_quality,_novelty_and_reproducibility": "I found no significant problems in these respects. The paper itself is well written and easy to follow. The proposed idea is clearly presented. I think the authors made a reasonable effort to include sufficient information for reproduction.\n\nThe author' idea, the use of accurate importance sampling for a neural radiance field, is somewhat new in the sense that  I do not know the same work that has been done/published before. However, as I mentioned in \"Strength and Weaknesses\", I do not think that this idea (in the current form) is reasonable and useful for advancing the technology in this field.", "summary_of_the_review": "As mentioned in \"Strength and Weaknesses\", I appreciate the logical explanation for the importance sampling for a neural radiance field. However, I think the authors' contribution is insufficient in the current form; they have not established the effectiveness of their proposal: computational advantages over the previous methods. Therefore, I recommend rejection for this submission.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No specific concerns.", "recommendation": "3: reject, not good enough"}, "tcdate": 1666061378227}], "openreview_url": "https://openreview.net/forum?id=zHSaBQtj-l", "arxiv_id": "2302.10970", "paper_pdf": "papers/zHSaBQtj-l.pdf", "paper_pdf_sha256": "d297a60270bccf9ea8cfdc0f012983c1d59f68dcb49de45bc14c247569dcf531", "paper_pdf_bytes": 4336943, "paper_pdf_source": "openreview", "code_url": "https://github.com/GreatDrake/reparameterized-volume-sampling", "code_repository": "GreatDrake/reparameterized-volume-sampling", "code_commit": "1fc6979f465e0e556786eee36648c7231c4006a9", "code_archive": "repos/zHSaBQtj-l.zip", "code_archive_sha256": "ead7086e7d32d43e88d4251c3bb2c08321eb3b1f3a18d2840208909b03c444f9", "code_archive_bytes": 703419, "code_file_count": 9, "code_extensions": {".py": 8, ".ipynb": 1}, "github_disk_usage_kb": 713, "github_languages": {"Jupyter Notebook": 603423, "Python": 77328}, "github_archived": false, "github_pushed_at": "2024-04-21T13:44:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/differentiable-rendering-with-reparameterized"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8nXkyH2_s6", "year": 2021, "status": "rejected", "title": "Neural networks behave as hash encoders: An empirical study", "authors": ["Fengxiang He", "Shiye Lei", "Jianmin Ji", "Dacheng Tao"], "authorids": ["~Fengxiang_He1", "leishiye@gmail.com", "jianmin@ustc.edu.cn", "~Dacheng_Tao1"], "authors_source": "OpenReview API", "abstract": "The input space of a neural network with ReLU-like activations is partitioned into multiple linear regions, each corresponding to a specific activation pattern of the included ReLU-like activations. We demonstrate that this partition exhibits the following encoding properties across a variety of deep learning models: (1) {\\it determinism}: almost every linear region contains at most one training example. We can therefore represent almost every training example by a unique activation pattern, which is parameterized by a {\\it neural code}; and (2) {\\it categorization}: according to the neural code, simple algorithms, such as $K$-Means, $K$-NN, and logistic regression, can achieve fairly good performance on both training and test data. These encoding properties surprisingly suggest that {\\it normal neural networks well-trained for classification behave as hash encoders without any extra efforts.} In addition, the encoding properties exhibit variability in different scenarios. {Further experiments demonstrate that {\\it model size}, {\\it training time}, {\\it training sample size}, {\\it regularization}, and {\\it label noise} contribute in shaping the encoding properties, while the impacts of the first three are dominant.} We then define an {\\it activation hash phase chart} to represent the space expanded by {model size}, training time, training sample size, and the encoding properties, which is divided into three canonical regions: {\\it under-expressive regime}, {\\it critically-expressive regime}, and {\\it sufficiently-expressive regime}.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "bIOWQHLFaJN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1829/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This draft proposes to use the relu activation pattern of the neurons in the neural network as the hash code for the input. Essentially the input features are bucketized into small piecewise linear regions. The authors show empirically that the proposed hash code has small collision and high accuracy given certain conditions including\n1. The features are around the sample manifold \n2. the training time is long enough\n3. the network is wide enough\n4. the training sample size is large enough\nThe authors also found empirically the effect of regularization is relatively small on the encoding properties.\n\nI feel this is an interesting thought but I am not sure if this is the first work on it. Some other possible issues:\n1. Figure 4(a) somehow shows the redundancy is the smallest at epoch 0, then it goes high after 1 epoch and decreases slowly as the number epochs grows. Can authors provide some explanation on this? Does it suggest the random initialization of the neural network gives a good hash code in terms of the redundancy metric? (of course the accuracy will be bad)\n2. The authors used K-means as another benchmark to compare. To me k-means is an unsupervised clustering algorithm. How do you get the accuracy from k-means? How do you match the cluster id to the labels?\n3. My biggest complaint is that only the MNIST data set is investigated in the experiment. MNIST is too easy to show any conclusive results. You may need to work on other data sets such as image net, cifar 100, or NLP related data set to draw a convincing conclusion.\n4. All the conclusions are purely empirical. Can authors provide some explanation or intuition on why the redundancy ratio decreases as the training time grows? Is this related to the type of optimizer being used? Why can a larger sample size also help reduce the redundancy ratio?\n\n\nOverall I think this draft has some really good ideas but the empirical result is not quite conclusive due to the lack of extensive experimentation.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "MNIST alone is not enough to draw any decisive conclusions", "review": "This draft proposes to use the relu activation pattern of the neurons in the neural network as the hash code for the input. Essentially the input features are bucketized into small piecewise linear regions. The authors show empirically that the proposed hash code has small collision and high accuracy given certain conditions including\n1. The features are around the sample manifold \n2. the training time is long enough\n3. the network is wide enough\n4. the training sample size is large enough\nThe authors also found empirically the effect of regularization is relatively small on the encoding properties.\n\nI feel this is an interesting thought but I am not sure if this is the first work on it. Some other possible issues:\n1. Figure 4(a) somehow shows the redundancy is the smallest at epoch 0, then it goes high after 1 epoch and decreases slowly as the number epochs grows. Can authors provide some explanation on this? Does it suggest the random initialization of the neural network gives a good hash code in terms of the redundancy metric? (of course the accuracy will be bad)\n2. The authors used K-means as another benchmark to compare. To me k-means is an unsupervised clustering algorithm. How do you get the accuracy from k-means? How do you match the cluster id to the labels?\n3. My biggest complaint is that only the MNIST data set is investigated in the experiment. MNIST is too easy to show any conclusive results. You may need to work on other data sets such as image net, cifar 100, or NLP related data set to draw a convincing conclusion.\n4. All the conclusions are purely empirical. Can authors provide some explanation or intuition on why the redundancy ratio decreases as the training time grows? Is this related to the type of optimizer being used? Why can a larger sample size also help reduce the redundancy ratio?\n\n\nOverall I think this draft has some really good ideas but the empirical result is not quite conclusive due to the lack of extensive experimentation.\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603857311211}, {"id": "W30ojOnalj", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1829/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Update after rebuttal:** I thank the authors for their detailed responses and the additional experiments. The responses addressed most of my concerns. I noticed that I had the wrong notion of redundancy ratio in my mind (I'm glad the authors now give a more formal definition of this concept as I think this would trip up many other readers). I'm also glad that the authors have clarified the difference between their results and those reported in Hanin and Rolnick (ICML, 2019). Given these, I'm happy to increase my score to a weak accept (a weak accept, because I'm still not quite sure about the significance of the results reported in this paper). \n\n--------------------------------------------------------------\nThis paper reports some observations about properties of linear regions in deep ReLU networks. Unfortunately, I have several major issues with the paper. \n\n(1) First of all, I am a bit confused about the motivation behind this work. The motivation is initially couched in terms of hash codes, so one gets the impression that the authors are going to propose a new hash coding scheme using neural networks. But this is clearly not the case. The proposed coding scheme is practically useless as a hash code because of the enormous dimensionality of the codes (I estimate this to be on the order of millions even for the toy MNIST case studied in the paper, but it would actually be very helpful for the reader if the authors explicitly mentioned the dimensionality of the proposed hash codes--at least the order of magnitude--). This is also why the authors are stuck with the toy MNIST dataset throughout the paper, because the proposed scheme is completely impractical for any reasonably large dataset and model.\n\n(2) This begs the question: what exactly is the significance of the observation that linear regions satisfy some properties of hash codes, if they’re not going to be used as hash codes? It’s not meaningful to just point out that something satisfies some properties of hash codes. One can point to a million different things that satisfy some properties of hash codes. What exactly is the significance of linear regions having these properties?\n\n(3) This brings me to the related works section. This section is written very shallowly, the authors do not do a good job of situating their contributions in the context of prior works. You cannot just say: “Other advances in linear region counting include …” (p. 3) and then cite a bunch of references. You have to tell us what each of these papers did and how what you’re doing in this paper differs from these earlier works and makes a meaningful and novel contribution to the prior literature.\n\n(4) The concept of redundancy ratio seems to play a central role in the paper, but it is not defined formally, just a verbal (potentially ambiguous) definition is given. Please define this concept formally to avoid any ambiguities. I’m also not sure this concept is the right one for quantifying the goodness of a code: consider a case where each point in a dataset shares its linear region with exactly one other point in the dataset vs. a case where all points belong to the same linear region. It seems that the redundancy ratio will be 100% in both cases, however intuitively the code in the first case should be much better than in the second case (for example for retrieval).\n\n(5) Relatedly, Figure 3c suggests that the redundancy ratio is close to zero for an untrained random network. Then, by the authors’ own definition, the encoding is actually very good before training. Please consider what this means (also see point 7 below).\n\n(6) Unfortunately, I don’t think classification results for MNIST only are very meaningful. Almost anything will get above 99% accuracy on MNIST. Moreover, no effort is made by the authors to understand what drives good test accuracy in these experiments. A very straightforward explanation is that accurate classification is primarily driven by higher layers, so one actually doesn’t need most of the dimensions in the hash code for good classification performance (similar “cache” models using high layer features have been proposed before: e.g. Dubey et al., CVPR 2019; Orhan, NeurIPS 2018; Khandelwal et al., ICLR 2020).\n \n(7) Most importantly, one of the main phenomena observed in this paper (the change of redundancy ratio over training) has already been reported in Hanin and Rolnick (ICML, 2019): they note that the number of linear regions in a deep ReLU net first decreases and then increases during training (please read their section 3 carefully). This would easily explain the trajectory of the redundancy ratio observed in Figure 3c (and in Figure 4a) in this paper. Moreover, the concept of diameter is also rigorously defined (and diameters of linear regions studied) in Hanin and Rolnick (ICML, 2019), but this is not acknowledged at all by the authors. This is a pretty serious omission. \n\n(8) Typos: should be: “Mapping induced by a relu network” (p. 1), “it is worth noting” (p. 2)\n“Geometric properties of linear regions” (p.3). “Another diameters of linear regions”? (p. 5)\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "official review", "review": "**Update after rebuttal:** I thank the authors for their detailed responses and the additional experiments. The responses addressed most of my concerns. I noticed that I had the wrong notion of redundancy ratio in my mind (I'm glad the authors now give a more formal definition of this concept as I think this would trip up many other readers). I'm also glad that the authors have clarified the difference between their results and those reported in Hanin and Rolnick (ICML, 2019). Given these, I'm happy to increase my score to a weak accept (a weak accept, because I'm still not quite sure about the significance of the results reported in this paper). \n\n--------------------------------------------------------------\nThis paper reports some observations about properties of linear regions in deep ReLU networks. Unfortunately, I have several major issues with the paper. \n\n(1) First of all, I am a bit confused about the motivation behind this work. The motivation is initially couched in terms of hash codes, so one gets the impression that the authors are going to propose a new hash coding scheme using neural networks. But this is clearly not the case. The proposed coding scheme is practically useless as a hash code because of the enormous dimensionality of the codes (I estimate this to be on the order of millions even for the toy MNIST case studied in the paper, but it would actually be very helpful for the reader if the authors explicitly mentioned the dimensionality of the proposed hash codes--at least the order of magnitude--). This is also why the authors are stuck with the toy MNIST dataset throughout the paper, because the proposed scheme is completely impractical for any reasonably large dataset and model.\n\n(2) This begs the question: what exactly is the significance of the observation that linear regions satisfy some properties of hash codes, if they’re not going to be used as hash codes? It’s not meaningful to just point out that something satisfies some properties of hash codes. One can point to a million different things that satisfy some properties of hash codes. What exactly is the significance of linear regions having these properties?\n\n(3) This brings me to the related works section. This section is written very shallowly, the authors do not do a good job of situating their contributions in the context of prior works. You cannot just say: “Other advances in linear region counting include …” (p. 3) and then cite a bunch of references. You have to tell us what each of these papers did and how what you’re doing in this paper differs from these earlier works and makes a meaningful and novel contribution to the prior literature.\n\n(4) The concept of redundancy ratio seems to play a central role in the paper, but it is not defined formally, just a verbal (potentially ambiguous) definition is given. Please define this concept formally to avoid any ambiguities. I’m also not sure this concept is the right one for quantifying the goodness of a code: consider a case where each point in a dataset shares its linear region with exactly one other point in the dataset vs. a case where all points belong to the same linear region. It seems that the redundancy ratio will be 100% in both cases, however intuitively the code in the first case should be much better than in the second case (for example for retrieval).\n\n(5) Relatedly, Figure 3c suggests that the redundancy ratio is close to zero for an untrained random network. Then, by the authors’ own definition, the encoding is actually very good before training. Please consider what this means (also see point 7 below).\n\n(6) Unfortunately, I don’t think classification results for MNIST only are very meaningful. Almost anything will get above 99% accuracy on MNIST. Moreover, no effort is made by the authors to understand what drives good test accuracy in these experiments. A very straightforward explanation is that accurate classification is primarily driven by higher layers, so one actually doesn’t need most of the dimensions in the hash code for good classification performance (similar “cache” models using high layer features have been proposed before: e.g. Dubey et al., CVPR 2019; Orhan, NeurIPS 2018; Khandelwal et al., ICLR 2020).\n \n(7) Most importantly, one of the main phenomena observed in this paper (the change of redundancy ratio over training) has already been reported in Hanin and Rolnick (ICML, 2019): they note that the number of linear regions in a deep ReLU net first decreases and then increases during training (please read their section 3 carefully). This would easily explain the trajectory of the redundancy ratio observed in Figure 3c (and in Figure 4a) in this paper. Moreover, the concept of diameter is also rigorously defined (and diameters of linear regions studied) in Hanin and Rolnick (ICML, 2019), but this is not acknowledged at all by the authors. This is a pretty serious omission. \n\n(8) Typos: should be: “Mapping induced by a relu network” (p. 1), “it is worth noting” (p. 2)\n“Geometric properties of linear regions” (p.3). “Another diameters of linear regions”? (p. 5)\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603821300591}, {"id": "UXdEFurD4yE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1829/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Update after authors' response**\nI am very happy to see the additional results on CIFAR-10, and the layer-wise ablations and other control experiments. To me, these results have shed a lot of light onto how the observed hashing effect can be explained. These explanations mostly confirm intuitions. Nonetheless I think it's worth reporting the empirical verification of these intuitions that tie in with earlier results reported in other papers. To me, the most interesting aspect of the work is how the hashing properties change over training (and across layers). The paper could still be improved by experiments on e.g. CIFAR-100, ImageNet and non-vision tasks, as well as more mathematically sophisticated definitions of some of the measures (e.g. average stochastic activation diameter). I personally think the results are now sufficient (and sufficiently backed up) for a publication and most of the criticism raised by the other reviewers has been addressed sufficiently (for me). I would now rate the paper as a 6.5 - but to facilitate the reviewer's discussion I will take a clear stance and have thus raised my score to 7.\n---\n\n**Summary**\nThe paper investigates activation patterns in ReLU networks, which are known to be piecewise linear. The main finding is that trained MNIST classifiers produce unique activation patterns for most points from the data-distribution (but not for random data). To be precise, an activation pattern is a binary matrix where each entry corresponds to one ReLU unit, and is 1 if the unit has non-zero activation and 0 otherwise. This means that trained MNIST classifiers can be viewed as “hashing” each data-point into a unique activation pattern. Importantly, simple clustering/classification algorithms (K-means, K-nearest-neighbors, and logistic regression) on these hashed patterns lead to good classification accuracy, implying that the hashed patterns follow some underlying geometric regularity (which is not trivially expected from an arbitrary hashing function).\n\n---------------------------------------------------------------------\n**Main contributions, Novelty, Impact**\n1) Systematic study of the “hashing” behavior of MNIST classifiers. To the best of my knowledge this particular type of analysis is novel. The findings are interesting, though perhaps not totally surprising. The impact of the results is currently limited by 2 important factors: (i) results are shown on MNIST only, and (ii) it is currently unclear whether the hashing effect is mostly explained by early layer activations, or not. To address (ii) layer-wise ablations would be crucial (see improvements for more details).\n\n2) Various ablations and control experiments to identify the impact of certain hyper-parameters and architectural variations. The ablations and control experiments are interesting to further characterize the observed hashing effect. Two important control experiments are currently missing: (i) results for untrained networks, and (ii) results when fitting and random labels (to determine whether the observed effects are mainly input-data-driven, or whether the hashed landscape is mainly shaped by label information).\n\n3) Definition of an “Activation Hash Phase Chart” with three qualitatively distinct reasons. The main idea behind this is interesting, but the paper never shows such a diagram, and (importantly!) how the chart is meant to be used (hyper-parameter tuning, or rather as a diagnostic tool, …?). It is thus currently unclear whether such a chart could have wider impact in the community.\n\n---------------------------------------------------------------------\n**Score and reasons for score**\nThe main observation in the paper is interesting and supported by a number of experiments. Unfortunately the generality of the findings is currently fairly limited since evaluation happened for MNIST classification only, and some control experiments are missing. I think the ingredients for a strong and interesting paper are there, but it does not quite come together yet. I therefore currently suggest a major revision of the work, which means resubmission at another conference. I personally want to strongly encourage the authors to spend a bit more time and work on the paper to turn it into the strongest version possible - in which case I think the paper could be quite impactful.  I am of course happy to reconsider my final verdict based on the other reviews and the authors' response.\n \nPerhaps the most important unaddressed point is how the observed hashing property can be reconciled with the good generalization performance: if each test-datapoint is mapped to a unique pattern, how does the classifier generalize well? I suspect that the hashing property is mostly a property of early layers that gets washed out towards the output of the classifier (but it might also be the case that the hashing property holds essentially until the final layer). I think the hashing property itself is not very surprising, but what’s surprising is that the “goodness-of-hash” changes quite a bit over training - not only for training points but also test data, but not random data. Shedding more light on this would make the paper much stronger in my opinion.\n\n---------------------------------------------------------------------\n**Strengths**\n * Observed hashing effect and hypothesis stated fits quite well with the wider literature (which is also well cited in the paper).\n * The experiments that are conducted are thorough (multiple repetitions per run, covering a broad range of hyperparameters).\n * To me Fig 3c is the most convincing result that the hashing effect is (at least partly) an caused by training and not simply a generic feature neural networks with a certain width.\n---------------------------------------------------------------------\n**Weaknesses**\n * Results are on MNIST only. Historically it’s often been the case that strong results on MNIST would not carry over to more complex data. Additionally, at least some core parts of the analysis does not require training networks (but could even be performed e.g. with pre-trained classifiers on ImageNet) - there is thus no severe computational bottleneck, which is often the case when going beyond MNIST.\n * The “Average stochastic activation diameter” is a quite crude measure and results must thus be taken with a (large) grain of salt. It would be good to perform some control experiments and sanity checks to make sure that the measure behaves as expected, particularly in high-dimensional spaces.\n * The current paper reports the hashing effect and starts relating it to what’s known in the literature, and has some experiments that try to understand the underlying *causes* for the hashing effect. However, while some factors are found to have an influence on the strength of the effect, some control experiments are still missing (training on random labels, results on untrained networks, and an analysis of how the results change when starting to leave out more and more of the early layers).\n\n---------------------------------------------------------------------\n**Correctness**\nOverall the methodology, results, and conclusions seem mostly fine (I’m currently not very convinced by the “stochastic activation diameter” and would not read too much into the corresponding results). Additionally some claims are not entirely supported (in fullest generality), based on the results shown, see comments for more on this.\n\n---------------------------------------------------------------------\n**Clarity**\nThe main idea is well presented and related literature is nicely cited. However, some of the writing is quite redundant (some parts of the intro appear as literal copies later in the text). Most importantly the writing in some parts of the manuscript seems quite rushed with quite a few typos and some sentences/passages that could be rephrased for more fluent reading.\n\n---------------------------------------------------------------------\n**Improvements (that would make me raise my score) / major issues (that need to be addressed)**\n1) Experiments on more complex datasets.\n\n2) One question that is currently unresolved is: is the hashing effect mostly attributed to early layer activations? Ultimately, a high-accuracy classifier will “lump together” all datapoints of a certain class when looking at the network output only. The question is whether this really happens at the very last layer or already earlier in the network. Similarly, when considering the input to the network (the raw data) the hashing effect holds since each data-point is unique. It is conceivable that the first layer activations only marginally transform the data in which case it would be somewhat trivially expected to see the hashing effect (when considering all activations simultaneously). However that might not explain e.g. the K-NN results.\nI think it would be very insightful to compute the redundancy ratio layer-wise and/or when leaving out more and more of the early layer activations (i.e. more and more rows of the activation pattern matrix). Additionally it would be great to see how this evolves over time, i.e. is the hashing effect initially mostly localized in early layers and does it gradually shape deeper activations over training? This would also shed some light on the very important issue of how a network that maps each (test-) data-point to a unique pattern generalize well?\n\n3) Another unresolved question is whether it’s mostly the structure of the input-data or the labels driving the organization of the hashed space? The random data experiments answers this partially. Additionally it would be interesting to see what happens when (i) training with random data, (ii) training with random labels - is the hashing effect still there, does the K-NN classification still work?\n\n4) Clarify: Does Fig 3c and 4a show results for untrained networks? I.e. is the redundancy ratio near 0 for training, test and random data in an untrained network? I would not be entirely surprised by that (a “reservoir effect”) but if that’s the case that should be commented/discussed in the paper, and improvement 3) mentioned above would become even more important. If the figures do not show results for untrained networks then please run the corresponding experiments and add them to the figures and Table 1.\n\n5) Clarify: Random data (Fig 3c). Was the network trained on random data, or do the dotted lines show networks trained on unaltered data, evaluated with random data?\n\n6) Clarify: Random data (Fig 3). Was the non-random data normalized or not (i.e. is the additional “unit-ball” noise small or large compared to the data). Ideally show some examples of the random data in the appendix.\n\n7) P3: “It is worth noting that the volume of boundaries between linear regions is zero” - is this still true for non-ReLU nonlinearities (e.g. sigmoids)? If not what are the consequences (can you still easily make the claims on P1: “This linear region partition can be extended to the neural networks containing smooth activations”)? Otherwise please rephrase the claims to refer to ReLU networks only.\n\n8) I disagree that model capacity is well measured by layer width. Please use the term ‘model-size’ instead of ‘model-capacity’ throughout the text. Model capacity is a more complex concept that is influenced by regularizers and other architectural properties (also note that the term capacity has e.g. a well-defined meaning in information theory, and when applied to neural networks it does not simply correspond to layer-width).\n\n9) Sec 5.4: I disagree that regularization “has very little impact” (as mentioned in the abstract and intro). Looking at the redundancy ratio for weight decay (unfortunately only shown in the appendix) one can clearly see a significant and systematic impact of the regularizer towards higher redundancy ratios (as theoretically expected) for some networks (I guess the impact is stronger for larger networks, unfortunately Fig 8 in the appendix does not allow to precisely answer which networks are which).\n\n---------------------------------------------------------------------\n**Minor comments**\nA) Formally define what “well-trained” means. The term is used quite often and it is unclear whether it simply means converged, or whether it refers to the trained classifier having to have a certain performance.\n\nB) There is quite an extensive body of literature (mainly 90s and early 2000s) on “reservoir effects” in randomly initialized, untrained networks (e.g. echo state networks and liquid state machines, however the latter use recurrent random nets). Perhaps it’s worth checking that literature for similar results.\n\nC) Remark 1: is really only the *training* distribution meant, i.e. without the *test* data, or is it the unaltered data generating distribution (i.e. without unit-ball noise)?\n\nD) Is the red histogram in Fig 3a and 3b the same (i.e. does Fig 3b use the network trained with 500 epochs)?\n\nE) P2 - Sufficiently-expressive regime: “This regime involves almost all common scenarios in the current practice of deep learning”. This is a bit of a strong claim which is not fully supported by the experiments - please tone it down a bit. It is for instance unclear whether the effect holds for non-classification tasks, and variational methods with strong entropy-based regularizers, or Dropout, ...\n\nF) P2- The Rosenblatt 1961 citation is not entirely accurate, MLP today typically only loosely refers to the original Perceptron (stacked into multiple-layers), most notably the latter is not trained via gradient backpropagation. I think it’s fine to use the term MLP without citation, or point out that MLP refers to a multi-layer feedforward network (trained via backprop).\n\nG) First paragraph in Sec. 4 is very redundant with the first two bullet points on P2 (parts of the text are literally copied). This is not a good writing style.\n\nH) P4 - first bullet point: “Generally, a larger redundancy ratio corresponds a worse encoding property.”. This is a quite hand-wavy statement - “worse” with respect to what? One could argue that for instance for good generalization high redundancy could be good.\n\nI) Fig 3: “10 epochs (red) and 500 epochs (blue),” does not match the figure legend where red and blue are swapped.\n\nJ) Fig 3: Panel b says “Rondom” data.\n\nK) Should the x-axis in Fig 3c be 10^x where x is what’s currently shown on the axis? (Similar to how 4a is labelled?)\n\nL) Some typos\nP2: It is worths noting\nP2:  By contrast, our the partition in activation hash phase chart characerizes goodnessof-hash.\nP3: For the brevity\nP3: activation statue\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting empirical observation that intuitively makes sense - but currently lacking some important controls.", "review": "**Update after authors' response**\nI am very happy to see the additional results on CIFAR-10, and the layer-wise ablations and other control experiments. To me, these results have shed a lot of light onto how the observed hashing effect can be explained. These explanations mostly confirm intuitions. Nonetheless I think it's worth reporting the empirical verification of these intuitions that tie in with earlier results reported in other papers. To me, the most interesting aspect of the work is how the hashing properties change over training (and across layers). The paper could still be improved by experiments on e.g. CIFAR-100, ImageNet and non-vision tasks, as well as more mathematically sophisticated definitions of some of the measures (e.g. average stochastic activation diameter). I personally think the results are now sufficient (and sufficiently backed up) for a publication and most of the criticism raised by the other reviewers has been addressed sufficiently (for me). I would now rate the paper as a 6.5 - but to facilitate the reviewer's discussion I will take a clear stance and have thus raised my score to 7.\n---\n\n**Summary**\nThe paper investigates activation patterns in ReLU networks, which are known to be piecewise linear. The main finding is that trained MNIST classifiers produce unique activation patterns for most points from the data-distribution (but not for random data). To be precise, an activation pattern is a binary matrix where each entry corresponds to one ReLU unit, and is 1 if the unit has non-zero activation and 0 otherwise. This means that trained MNIST classifiers can be viewed as “hashing” each data-point into a unique activation pattern. Importantly, simple clustering/classification algorithms (K-means, K-nearest-neighbors, and logistic regression) on these hashed patterns lead to good classification accuracy, implying that the hashed patterns follow some underlying geometric regularity (which is not trivially expected from an arbitrary hashing function).\n\n---------------------------------------------------------------------\n**Main contributions, Novelty, Impact**\n1) Systematic study of the “hashing” behavior of MNIST classifiers. To the best of my knowledge this particular type of analysis is novel. The findings are interesting, though perhaps not totally surprising. The impact of the results is currently limited by 2 important factors: (i) results are shown on MNIST only, and (ii) it is currently unclear whether the hashing effect is mostly explained by early layer activations, or not. To address (ii) layer-wise ablations would be crucial (see improvements for more details).\n\n2) Various ablations and control experiments to identify the impact of certain hyper-parameters and architectural variations. The ablations and control experiments are interesting to further characterize the observed hashing effect. Two important control experiments are currently missing: (i) results for untrained networks, and (ii) results when fitting and random labels (to determine whether the observed effects are mainly input-data-driven, or whether the hashed landscape is mainly shaped by label information).\n\n3) Definition of an “Activation Hash Phase Chart” with three qualitatively distinct reasons. The main idea behind this is interesting, but the paper never shows such a diagram, and (importantly!) how the chart is meant to be used (hyper-parameter tuning, or rather as a diagnostic tool, …?). It is thus currently unclear whether such a chart could have wider impact in the community.\n\n---------------------------------------------------------------------\n**Score and reasons for score**\nThe main observation in the paper is interesting and supported by a number of experiments. Unfortunately the generality of the findings is currently fairly limited since evaluation happened for MNIST classification only, and some control experiments are missing. I think the ingredients for a strong and interesting paper are there, but it does not quite come together yet. I therefore currently suggest a major revision of the work, which means resubmission at another conference. I personally want to strongly encourage the authors to spend a bit more time and work on the paper to turn it into the strongest version possible - in which case I think the paper could be quite impactful.  I am of course happy to reconsider my final verdict based on the other reviews and the authors' response.\n \nPerhaps the most important unaddressed point is how the observed hashing property can be reconciled with the good generalization performance: if each test-datapoint is mapped to a unique pattern, how does the classifier generalize well? I suspect that the hashing property is mostly a property of early layers that gets washed out towards the output of the classifier (but it might also be the case that the hashing property holds essentially until the final layer). I think the hashing property itself is not very surprising, but what’s surprising is that the “goodness-of-hash” changes quite a bit over training - not only for training points but also test data, but not random data. Shedding more light on this would make the paper much stronger in my opinion.\n\n---------------------------------------------------------------------\n**Strengths**\n * Observed hashing effect and hypothesis stated fits quite well with the wider literature (which is also well cited in the paper).\n * The experiments that are conducted are thorough (multiple repetitions per run, covering a broad range of hyperparameters).\n * To me Fig 3c is the most convincing result that the hashing effect is (at least partly) an caused by training and not simply a generic feature neural networks with a certain width.\n---------------------------------------------------------------------\n**Weaknesses**\n * Results are on MNIST only. Historically it’s often been the case that strong results on MNIST would not carry over to more complex data. Additionally, at least some core parts of the analysis does not require training networks (but could even be performed e.g. with pre-trained classifiers on ImageNet) - there is thus no severe computational bottleneck, which is often the case when going beyond MNIST.\n * The “Average stochastic activation diameter” is a quite crude measure and results must thus be taken with a (large) grain of salt. It would be good to perform some control experiments and sanity checks to make sure that the measure behaves as expected, particularly in high-dimensional spaces.\n * The current paper reports the hashing effect and starts relating it to what’s known in the literature, and has some experiments that try to understand the underlying *causes* for the hashing effect. However, while some factors are found to have an influence on the strength of the effect, some control experiments are still missing (training on random labels, results on untrained networks, and an analysis of how the results change when starting to leave out more and more of the early layers).\n\n---------------------------------------------------------------------\n**Correctness**\nOverall the methodology, results, and conclusions seem mostly fine (I’m currently not very convinced by the “stochastic activation diameter” and would not read too much into the corresponding results). Additionally some claims are not entirely supported (in fullest generality), based on the results shown, see comments for more on this.\n\n---------------------------------------------------------------------\n**Clarity**\nThe main idea is well presented and related literature is nicely cited. However, some of the writing is quite redundant (some parts of the intro appear as literal copies later in the text). Most importantly the writing in some parts of the manuscript seems quite rushed with quite a few typos and some sentences/passages that could be rephrased for more fluent reading.\n\n---------------------------------------------------------------------\n**Improvements (that would make me raise my score) / major issues (that need to be addressed)**\n1) Experiments on more complex datasets.\n\n2) One question that is currently unresolved is: is the hashing effect mostly attributed to early layer activations? Ultimately, a high-accuracy classifier will “lump together” all datapoints of a certain class when looking at the network output only. The question is whether this really happens at the very last layer or already earlier in the network. Similarly, when considering the input to the network (the raw data) the hashing effect holds since each data-point is unique. It is conceivable that the first layer activations only marginally transform the data in which case it would be somewhat trivially expected to see the hashing effect (when considering all activations simultaneously). However that might not explain e.g. the K-NN results.\nI think it would be very insightful to compute the redundancy ratio layer-wise and/or when leaving out more and more of the early layer activations (i.e. more and more rows of the activation pattern matrix). Additionally it would be great to see how this evolves over time, i.e. is the hashing effect initially mostly localized in early layers and does it gradually shape deeper activations over training? This would also shed some light on the very important issue of how a network that maps each (test-) data-point to a unique pattern generalize well?\n\n3) Another unresolved question is whether it’s mostly the structure of the input-data or the labels driving the organization of the hashed space? The random data experiments answers this partially. Additionally it would be interesting to see what happens when (i) training with random data, (ii) training with random labels - is the hashing effect still there, does the K-NN classification still work?\n\n4) Clarify: Does Fig 3c and 4a show results for untrained networks? I.e. is the redundancy ratio near 0 for training, test and random data in an untrained network? I would not be entirely surprised by that (a “reservoir effect”) but if that’s the case that should be commented/discussed in the paper, and improvement 3) mentioned above would become even more important. If the figures do not show results for untrained networks then please run the corresponding experiments and add them to the figures and Table 1.\n\n5) Clarify: Random data (Fig 3c). Was the network trained on random data, or do the dotted lines show networks trained on unaltered data, evaluated with random data?\n\n6) Clarify: Random data (Fig 3). Was the non-random data normalized or not (i.e. is the additional “unit-ball” noise small or large compared to the data). Ideally show some examples of the random data in the appendix.\n\n7) P3: “It is worth noting that the volume of boundaries between linear regions is zero” - is this still true for non-ReLU nonlinearities (e.g. sigmoids)? If not what are the consequences (can you still easily make the claims on P1: “This linear region partition can be extended to the neural networks containing smooth activations”)? Otherwise please rephrase the claims to refer to ReLU networks only.\n\n8) I disagree that model capacity is well measured by layer width. Please use the term ‘model-size’ instead of ‘model-capacity’ throughout the text. Model capacity is a more complex concept that is influenced by regularizers and other architectural properties (also note that the term capacity has e.g. a well-defined meaning in information theory, and when applied to neural networks it does not simply correspond to layer-width).\n\n9) Sec 5.4: I disagree that regularization “has very little impact” (as mentioned in the abstract and intro). Looking at the redundancy ratio for weight decay (unfortunately only shown in the appendix) one can clearly see a significant and systematic impact of the regularizer towards higher redundancy ratios (as theoretically expected) for some networks (I guess the impact is stronger for larger networks, unfortunately Fig 8 in the appendix does not allow to precisely answer which networks are which).\n\n---------------------------------------------------------------------\n**Minor comments**\nA) Formally define what “well-trained” means. The term is used quite often and it is unclear whether it simply means converged, or whether it refers to the trained classifier having to have a certain performance.\n\nB) There is quite an extensive body of literature (mainly 90s and early 2000s) on “reservoir effects” in randomly initialized, untrained networks (e.g. echo state networks and liquid state machines, however the latter use recurrent random nets). Perhaps it’s worth checking that literature for similar results.\n\nC) Remark 1: is really only the *training* distribution meant, i.e. without the *test* data, or is it the unaltered data generating distribution (i.e. without unit-ball noise)?\n\nD) Is the red histogram in Fig 3a and 3b the same (i.e. does Fig 3b use the network trained with 500 epochs)?\n\nE) P2 - Sufficiently-expressive regime: “This regime involves almost all common scenarios in the current practice of deep learning”. This is a bit of a strong claim which is not fully supported by the experiments - please tone it down a bit. It is for instance unclear whether the effect holds for non-classification tasks, and variational methods with strong entropy-based regularizers, or Dropout, ...\n\nF) P2- The Rosenblatt 1961 citation is not entirely accurate, MLP today typically only loosely refers to the original Perceptron (stacked into multiple-layers), most notably the latter is not trained via gradient backpropagation. I think it’s fine to use the term MLP without citation, or point out that MLP refers to a multi-layer feedforward network (trained via backprop).\n\nG) First paragraph in Sec. 4 is very redundant with the first two bullet points on P2 (parts of the text are literally copied). This is not a good writing style.\n\nH) P4 - first bullet point: “Generally, a larger redundancy ratio corresponds a worse encoding property.”. This is a quite hand-wavy statement - “worse” with respect to what? One could argue that for instance for good generalization high redundancy could be good.\n\nI) Fig 3: “10 epochs (red) and 500 epochs (blue),” does not match the figure legend where red and blue are swapped.\n\nJ) Fig 3: Panel b says “Rondom” data.\n\nK) Should the x-axis in Fig 3c be 10^x where x is what’s currently shown on the axis? (Similar to how 4a is labelled?)\n\nL) Some typos\nP2: It is worths noting\nP2:  By contrast, our the partition in activation hash phase chart characerizes goodnessof-hash.\nP3: For the brevity\nP3: activation statue\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603644964188}, {"id": "SDqOQqU29av", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1829/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the problem of linear partition in the linear spaces of neural networks with ReLU-like activations is studied. It demonstrates that such a partition exhibits two properties, including determinism and categorization, across a variety of deep learning models \n\n————————————————————————————————————————————\n\nThis paper presents an interesting work, however, there are a few issues/comments with the work:\n\n1.This paper mentions that “simple classification and clustering algorithms, such as K-NN, logistic regression, and K-means can achieve fairly good training accuracy and test accuracy on the neural code space”,  About “achieve fairly good training accuracy and test accuracy”, maybe some comparisons with existing methods will strength this statement.\n\n2.The paper concludes that model capacity, training time, and sample size play important roles in shaping the encoding properties, while several popular regularizers have little impact on the encoding properties.  As for the conclusion of the regularization technique, it is better to try other methods, such as drop-out, to verify gradient clipping and weight decay not only on MLP but also on RNN, before reaching such a conclusion. Moreover, only for MLP on MNIST, it seems unconvincing.  It will be better to try out some existing/well-known NN models on more datasets.
\n3.I was confused about the sentence below Section 5.4 “we trained 345 MLPs on the MNIST dataset” and the last sentence in the caption of Figure6 “Totally, 115 models are involved in one scatter. “. If possible, could you please explain that?\n\n4.In Figure 2 (c) and (d), “Diameter“ on the horizontal axis correspondings to the “Layer width” in Figure 2 (a) and (b)\n\n5.What is the meaning of “Frequency” on the vertical axis?\n\n6.In Figure 6, the third and fourth subfigures mean that for some of MLPs with depths between 3 and 100, “ without BN” would be better than  “with BN”, BN sometimes harms the performance of these models? is that so?\n\n7.It needs some space between the second and the third paragraphs below Figure 2.\n\n8.In deep models, depth generally is more important than width. Why only analyze the model capacity by width?\n\n9.For the definitions of three regions “\tUnder-expressive regime”, “Critically-expressive regime”, and “Sufficiently-expressive regime”, may not need to be repeated on page 8, as similar definitions are explained on page 2. Also, is that possible to quantify/formalize them? If so, it would be more interesting and useful.\n\n10.The format of references is inconsistent, for example, references “Jingdong Wang, Ting Zhang, Nicu Sebe, Heng Tao Shen, et al.. 2017” and “Li Yuan, Eng Hock Francis Tay, Ping Li, and Jiashi Feng. 2019”.\n\n==========================================================================================\n\n\"In addition, I am not sure the description would be enough to reproduce and no code seems to provide.\" In the beginning, I did not find the codes related to this paper, but later, the author(s) uploaded codes. Thanks.  For the codes,  if setting random seed program-wide in codes maybe it will be more helpful for reproducing. \n\n==========================================================================================\n\n————————————————————————————————————————————\n\nOverall, I think this is an interesting piece of work that may be of interest to people to explore/understand (deep) neural networks, but I think the results/conclusions require more careful/extended analysis.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Hash encoders of Neural networks", "review": "In this paper, the problem of linear partition in the linear spaces of neural networks with ReLU-like activations is studied. It demonstrates that such a partition exhibits two properties, including determinism and categorization, across a variety of deep learning models \n\n————————————————————————————————————————————\n\nThis paper presents an interesting work, however, there are a few issues/comments with the work:\n\n1.This paper mentions that “simple classification and clustering algorithms, such as K-NN, logistic regression, and K-means can achieve fairly good training accuracy and test accuracy on the neural code space”,  About “achieve fairly good training accuracy and test accuracy”, maybe some comparisons with existing methods will strength this statement.\n\n2.The paper concludes that model capacity, training time, and sample size play important roles in shaping the encoding properties, while several popular regularizers have little impact on the encoding properties.  As for the conclusion of the regularization technique, it is better to try other methods, such as drop-out, to verify gradient clipping and weight decay not only on MLP but also on RNN, before reaching such a conclusion. Moreover, only for MLP on MNIST, it seems unconvincing.  It will be better to try out some existing/well-known NN models on more datasets.
\n3.I was confused about the sentence below Section 5.4 “we trained 345 MLPs on the MNIST dataset” and the last sentence in the caption of Figure6 “Totally, 115 models are involved in one scatter. “. If possible, could you please explain that?\n\n4.In Figure 2 (c) and (d), “Diameter“ on the horizontal axis correspondings to the “Layer width” in Figure 2 (a) and (b)\n\n5.What is the meaning of “Frequency” on the vertical axis?\n\n6.In Figure 6, the third and fourth subfigures mean that for some of MLPs with depths between 3 and 100, “ without BN” would be better than  “with BN”, BN sometimes harms the performance of these models? is that so?\n\n7.It needs some space between the second and the third paragraphs below Figure 2.\n\n8.In deep models, depth generally is more important than width. Why only analyze the model capacity by width?\n\n9.For the definitions of three regions “\tUnder-expressive regime”, “Critically-expressive regime”, and “Sufficiently-expressive regime”, may not need to be repeated on page 8, as similar definitions are explained on page 2. Also, is that possible to quantify/formalize them? If so, it would be more interesting and useful.\n\n10.The format of references is inconsistent, for example, references “Jingdong Wang, Ting Zhang, Nicu Sebe, Heng Tao Shen, et al.. 2017” and “Li Yuan, Eng Hock Francis Tay, Ping Li, and Jiashi Feng. 2019”.\n\n==========================================================================================\n\n\"In addition, I am not sure the description would be enough to reproduce and no code seems to provide.\" In the beginning, I did not find the codes related to this paper, but later, the author(s) uploaded codes. Thanks.  For the codes,  if setting random seed program-wide in codes maybe it will be more helpful for reproducing. \n\n==========================================================================================\n\n————————————————————————————————————————————\n\nOverall, I think this is an interesting piece of work that may be of interest to people to explore/understand (deep) neural networks, but I think the results/conclusions require more careful/extended analysis.", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1603601711102}], "openreview_url": "https://openreview.net/forum?id=8nXkyH2_s6", "arxiv_id": "2101.05490", "paper_pdf": "papers/8nXkyH2_s6.pdf", "paper_pdf_sha256": "e07a0c4bf55da59616a0fcf40bc32c09a0b9cb779bb1670aeaa8ba51badf72cb", "paper_pdf_bytes": 1089361, "paper_pdf_source": "openreview", "code_url": "https://github.com/LeavesLei/activation-code", "code_repository": "LeavesLei/activation-code", "code_commit": "6be962f5e6c081d0bbefd6a30f31b13728da7752", "code_archive": "repos/8nXkyH2_s6.zip", "code_archive_sha256": "a4bc5085e3e930bca96833b326175109900c7f7e546e7aebd615e73d37bdf373", "code_archive_bytes": 810980, "code_file_count": 136, "code_extensions": {".sh": 69, ".py": 67}, "github_disk_usage_kb": 1795, "github_languages": {"Python": 323069, "Shell": 23651}, "github_archived": false, "github_pushed_at": "2021-05-28T08:55:34Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/neural-networks-behave-as-hash-encoders-an-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Xq4NdAodoA", "year": 2026, "status": "rejected", "title": "Generative Modeling with Bayesian Sample Inference", "authors": ["Marten Lienen", "Marcel Kollovieh", "Stephan Günnemann"], "authorids": ["~Marten_Lienen1", "~Marcel_Kollovieh1", "~Stephan_Günnemann1"], "authors_source": "OpenReview API", "abstract": "We derive a novel generative model from iterative Gaussian posterior inference. By treating the generated sample as an unknown variable, we can formulate the sampling process in the language of Bayesian probability. Our model uses a sequence of prediction and posterior update steps to iteratively narrow down the unknown sample starting from a broad initial belief. In addition to a rigorous theoretical analysis, we establish a connection between our model and diffusion models and show that it includes Bayesian Flow Networks (BFNs) as a special case. In our experiments, we demonstrate that our model improves sample quality on ImageNet32 over both BFNs and the closely related Variational Diffusion Models, while achieving equivalent log-likelihoods on ImageNet32 and CIFAR10.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "5UR5Mzi81Z", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13057/Reviewer_yESK"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper introduces Bayesian Sample Inference (BSI), a generative framework that reformulates sample generation as iterative Gaussian posterior updates. The proposed model uses a sequence of prediction and posterior update steps to iteratively narrow down the unknown sample starting from a broad initial belief. The paper include an extensive theoretical analysis along with empirical evaluation on the ImageNet 32x32 and CIFAR-10 datasets.", "review_text": "The paper introduces Bayesian Sample Inference (BSI), a generative framework that reformulates sample generation as iterative Gaussian posterior updates. The proposed model uses a sequence of prediction and posterior update steps to iteratively narrow down the unknown sample starting from a broad initial belief. The paper include an extensive theoretical analysis along with empirical evaluation on the ImageNet 32x32 and CIFAR-10 datasets.", "strengths": "The paper provides a novel perspective on generative modeling that frames sample generation as iterative Bayesian inference.\n• The proposed BSI framework is quite general, and the paper shows that BSI includes BFN (“Bayesian Flow Networks”) as a special case.\n• The includes a succinct and easy to understand comparison between the BSI and BFN / VDM frameworks.", "weaknesses": "• The proposed BSI only leads to minor improvements over the BFN and VDM models. In fact, the loglikelihood BPD metric is nearly identical compared to the BFN and VDM models on both the ImageNet 32x32 and CIFAR-10 datasets.\n• The paper only considers a small set of baselines – BFN and VDM – on ImageNet 32x32 and CIFAR-10. The paper should also compare to state of the art models such as “Improved Denoising Diffusion Probabilistic Models, arXiv 2025”, “Neural Diffusion Models, arXiv 2024”.\n• The paper only considers the low-resolution ImageNet 32x32 and CIFAR-10 datasets. The paper should provide results on the higher resolution ImageNet 64x64 dataset, as provided in the VDM (“Variational Diffusion Models”) paper.\n• The paper does not provide any analysis of the number of steps required for convergence by the proposed BSI model compared to the BFN and VDM models. In Figure 2, the proposed BSI model requires more steps compared to the BFN and VDM models to converge.\n• The difference between the proposed BSI model compared to the BFN model is minor. The effective difference is only in the choice of the non-deterministic hyper-prior $p(\\mu_0)$ in BSI.", "questions": "• The advantage of the proposed BSI framework over the BFN/ VDM frameworks should be described in more detail.\n• The choice of baselines should be motivated in more detail.\n• The paper should include more results on higher resolution image datasets, e.g., ImageNet 64 x 64.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Bayesian Sample Inference (BSI), a generative framework that reformulates sample generation as iterative Gaussian posterior updates. The proposed model uses a sequence of prediction and posterior update steps to iteratively narrow down the unknown sample starting from a broad initial belief. The paper include an extensive theoretical analysis along with empirical evaluation on the ImageNet 32x32 and CIFAR-10 datasets.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper provides a novel perspective on generative modeling that frames sample generation as iterative Bayesian inference.\n• The proposed BSI framework is quite general, and the paper shows that BSI includes BFN (“Bayesian Flow Networks”) as a special case.\n• The includes a succinct and easy to understand comparison between the BSI and BFN / VDM frameworks.", "weaknesses": "• The proposed BSI only leads to minor improvements over the BFN and VDM models. In fact, the loglikelihood BPD metric is nearly identical compared to the BFN and VDM models on both the ImageNet 32x32 and CIFAR-10 datasets.\n• The paper only considers a small set of baselines – BFN and VDM – on ImageNet 32x32 and CIFAR-10. The paper should also compare to state of the art models such as “Improved Denoising Diffusion Probabilistic Models, arXiv 2025”, “Neural Diffusion Models, arXiv 2024”.\n• The paper only considers the low-resolution ImageNet 32x32 and CIFAR-10 datasets. The paper should provide results on the higher resolution ImageNet 64x64 dataset, as provided in the VDM (“Variational Diffusion Models”) paper.\n• The paper does not provide any analysis of the number of steps required for convergence by the proposed BSI model compared to the BFN and VDM models. In Figure 2, the proposed BSI model requires more steps compared to the BFN and VDM models to converge.\n• The difference between the proposed BSI model compared to the BFN model is minor. The effective difference is only in the choice of the non-deterministic hyper-prior $p(\\mu_0)$ in BSI.", "questions": "• The advantage of the proposed BSI framework over the BFN/ VDM frameworks should be described in more detail.\n• The choice of baselines should be motivated in more detail.\n• The paper should include more results on higher resolution image datasets, e.g., ImageNet 64 x 64.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762122258251}, {"id": "Ol38HkXF3R", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13057/Reviewer_sXog"], "rating": 2, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes Bayesian Sample Inference (BSI), a generative modeling framework that treats generation as iterative posterior inference on an unknown sample using a sequence of prediction and Bayesian updates . The authors derive an ELBO in both finite-step and continuous limits, enabling likelihood-based training with a variance-reduced estimator via log-uniform importance sampling. They analyze relationships to Variational Diffusion Models (VDM) and Bayesian Flow Networks (BFN), showing BFN as a special case and connecting BSI to diffusion via a (generally) non-Markov forward process . Empirically, BSI matches likelihoods and improves FID on ImageNet32 vs. VDM/BFN using the same DiT backbone, with additional CIFAR-10 likelihood comparisons using a shared U-Net .", "review_text": "The paper proposes Bayesian Sample Inference (BSI), a generative modeling framework that treats generation as iterative posterior inference on an unknown sample using a sequence of prediction and Bayesian updates . The authors derive an ELBO in both finite-step and continuous limits, enabling likelihood-based training with a variance-reduced estimator via log-uniform importance sampling. They analyze relationships to Variational Diffusion Models (VDM) and Bayesian Flow Networks (BFN), showing BFN as a special case and connecting BSI to diffusion via a (generally) non-Markov forward process . Empirically, BSI matches likelihoods and improves FID on ImageNet32 vs. VDM/BFN using the same DiT backbone, with additional CIFAR-10 likelihood comparisons using a shared U-Net .", "strengths": "- Clear probabilistic formulation with closed-form update and a principled ELBO; continuous-limit bound is neat and intuitive.\n\n- Concrete connection to VDM/BFN; proof that BFN is a limit case and discussion of Markov vs. non-Markov forward processes add theoretical clarity.\n\n- Variance reduction via log-uniform sampling is motivated analytically and validated empirically.\n\n- Competitive or better results on ImageNet32 (lower FID at matched likelihoods) and matched BPD on CIFAR-10 with controlled architectures/seeds.", "weaknesses": "- Empirical scope is narrow: small-resolution datasets (ImageNet32, CIFAR-10); no high-res, text-conditioned, or scaling-law study to substantiate broader quality claims.\n\n- The theoretical comparison to VDM emphasizes simplicity of the BSI/BFN update, but practical stability/accuracy trade-offs vs. VDM’s log-space parameterization are not quantified.", "questions": "- The proposed Bayesian perspective vs. VDM/BFN: what concrete theoretical advantages does BSI deliver (e.g., tighter/cleaner ELBO, better conditioning, robustness) beyond interpretability? Please make the comparison formal.\n\n- It would be important to support the argument that BSI has higher generation quality than VDM and BFN if there are larger scales such as higher resolutions(e.g., ImageNet-256 and ImageNet-512) or experiments on scaling laws. Please provide results generated with larger resolution or text conditions, and performance trends as the model/data/number of steps increases.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Bayesian Sample Inference (BSI), a generative modeling framework that treats generation as iterative posterior inference on an unknown sample using a sequence of prediction and Bayesian updates . The authors derive an ELBO in both finite-step and continuous limits, enabling likelihood-based training with a variance-reduced estimator via log-uniform importance sampling. They analyze relationships to Variational Diffusion Models (VDM) and Bayesian Flow Networks (BFN), showing BFN as a special case and connecting BSI to diffusion via a (generally) non-Markov forward process . Empirically, BSI matches likelihoods and improves FID on ImageNet32 vs. VDM/BFN using the same DiT backbone, with additional CIFAR-10 likelihood comparisons using a shared U-Net .", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- Clear probabilistic formulation with closed-form update and a principled ELBO; continuous-limit bound is neat and intuitive.\n\n- Concrete connection to VDM/BFN; proof that BFN is a limit case and discussion of Markov vs. non-Markov forward processes add theoretical clarity.\n\n- Variance reduction via log-uniform sampling is motivated analytically and validated empirically.\n\n- Competitive or better results on ImageNet32 (lower FID at matched likelihoods) and matched BPD on CIFAR-10 with controlled architectures/seeds.", "weaknesses": "- Empirical scope is narrow: small-resolution datasets (ImageNet32, CIFAR-10); no high-res, text-conditioned, or scaling-law study to substantiate broader quality claims.\n\n- The theoretical comparison to VDM emphasizes simplicity of the BSI/BFN update, but practical stability/accuracy trade-offs vs. VDM’s log-space parameterization are not quantified.", "questions": "- The proposed Bayesian perspective vs. VDM/BFN: what concrete theoretical advantages does BSI deliver (e.g., tighter/cleaner ELBO, better conditioning, robustness) beyond interpretability? Please make the comparison formal.\n\n- It would be important to support the argument that BSI has higher generation quality than VDM and BFN if there are larger scales such as higher resolutions(e.g., ImageNet-256 and ImageNet-512) or experiments on scaling laws. Please provide results generated with larger resolution or text conditions, and performance trends as the model/data/number of steps increases.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761992133081}, {"id": "SxguA7usn2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13057/Reviewer_9kRj"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces Bayesian Sample Inference (BSI), a novel generative model that frames the generation process as iterative Bayesian inference. The core idea is to treat a data sample as an unknown variable and to iteratively refine a belief about it, starting from a broad prior and updating it based on noisy measurements of model predictions. The authors derive an Evidence Lower Bound (ELBO) for both finite and infinite-step versions of this process, enabling likelihood-based training. Through this framework, they show that Bayesian Flow Networks (BFNs) are a special case of BSI, providing a new perspective on them. Experiments on ImageNet32 and CIFAR10 demonstrate that BSI achieves sample quality (FID) superior to both Variational Diffusion Models (VDM) and BFNs, while maintaining equivalent log-likelihoods (BPD). The significance of this work lies in its introduction of a conceptually simple and elegant Bayesian perspective on generative modeling that unifies existing models and proves to be empirically effective.", "review_text": "This paper introduces Bayesian Sample Inference (BSI), a novel generative model that frames the generation process as iterative Bayesian inference. The core idea is to treat a data sample as an unknown variable and to iteratively refine a belief about it, starting from a broad prior and updating it based on noisy measurements of model predictions. The authors derive an Evidence Lower Bound (ELBO) for both finite and infinite-step versions of this process, enabling likelihood-based training. Through this framework, they show that Bayesian Flow Networks (BFNs) are a special case of BSI, providing a new perspective on them. Experiments on ImageNet32 and CIFAR10 demonstrate that BSI achieves sample quality (FID) superior to both Variational Diffusion Models (VDM) and BFNs, while maintaining equivalent log-likelihoods (BPD). The significance of this work lies in its introduction of a conceptually simple and elegant Bayesian perspective on generative modeling that unifies existing models and proves to be empirically effective.", "strengths": "*   **Novelty:** The primary strength of the paper is its novel framing of generative modeling. Viewing sample generation as the process of inferring a fixed but unknown variable via iterative posterior updates is a conceptually clean and powerful idea. This Bayesian perspective is a welcome contribution, offering a different way to think about the gradual refinement process that characterizes modern generative models.\n\n*   **Theoretical Soundness:** The paper provides a rigorous theoretical foundation for BSI. The derivation of the ELBO (Theorems 3.1 and 3.2) is solid and connects the model to the established practice of variational inference. A key contribution is showing that BFNs emerge as a special case of BSI when the initial prior is deterministic (infinite precision). This not only unifies BFNs within a more general framework but also simplifies and clarifies their connection to diffusion models. The analysis of the \"noising\" process (Section A.2) and how BSI's non-Markovian forward process differs from BFN's Markovian one is insightful.\n\n*   **Strong Empirical Results:** The experimental validation, though limited in scope, is compelling within its domain. The authors compare BSI against two highly relevant and strong baselines: VDM and BFN. On the ImageNet32 benchmark, BSI achieves a statistically significant improvement in sample quality (FID of 8.9 vs. 9.9 for VDM and 11.0 for BFN) while matching their performance on log-likelihood. This is a strong result that demonstrates a clear advantage of the proposed formulation.", "weaknesses": "Despite its elegant formulation and promising results I still have doubts on the methodology and experimental validation, which raise several critical questions that weaken its overall contribution. The claims of novelty and superiority feel overstated upon closer inspection of the underlying algorithm and the lack of crucial ablation studies.\n\n1.  **Conflation of Re-framing with Algorithmic Novelty:** The core training algorithm (Algorithm 2) is structurally almost identical to standard denoising diffusion models. It samples a noise level (`λ`), creates a noisy input (`μ_λ`), and trains a network `f_θ` to predict the original sample `x` from `μ_λ`. While the Bayesian derivation is new, the final algorithm is not. This raises the question of whether BSI is a fundamentally new *class* of models or a well-motivated re-derivation of a specific noise schedule and prior for existing models.\n\n2.  **Lack of Ablation Studies:** The paper's central claim is that BSI's performance advantage over BFN comes from its non-deterministic prior (`p(μ₀)` with `λ₀⁻¹ > 0`). This is a critical claim that is not substantiated with sufficient evidence. The paper only compares the two extremes: BSI (`λ₀⁻¹ = 100`) and BFN (`λ₀⁻¹ = 0`). A proper scientific validation would require an ablation study showing how FID and BPD vary across a range of `λ₀⁻¹` values to demonstrate a clear trend and justify the chosen hyperparameter. The same applies to other key design choices, such as the `p(λ)` proposal distribution.\n\n3.  **Questionable Experimental Comparison:** The main empirical comparison is between BSI, BFN, and VDM using a DiT-L-2 backbone. Figure 2 clearly shows that the input distributions for these models at the highest noise levels are qualitatively different. BSI and BFN start from (near) pure noise, whereas VDM's noisiest input mean (`Np(0.08x, 1)`) still contains significant structural information about the image. The DiT architecture, with its global self-attention mechanism, is particularly well-suited to reconstructing images from unstructured noise. This setup may be unfairly biased towards BSI/BFN and confounding the architectural strengths with the merits of the generative formulation.\n\n4.  **Disconnect Between Training Objective and Sampling:** The model is trained to optimize an \"infinite-step\" ELBO, which is a continuous integral over all precision levels `λ`. However, sampling is always performed in a finite, discrete number of steps. The paper does not provide a theoretical justification for why optimizing the continuous objective is ideal for the discrete sampling task. This is particularly relevant for the low-step sampling regime, which is crucial for practical efficiency but not explored in the paper.", "questions": "*   **Question 1:** The final training objective (optimizing `L_M`) amounts to a noise-conditioned denoising task, structurally similar to other diffusion models. Beyond the novel derivation, what is the fundamental algorithmic difference between BSI and a VDM/BFN where the noise schedule is designed to match BSI's effective `λ` distribution and a non-zero initial noise level is used?\n\n*   **Question 2:** The choice of a log-uniform proposal distribution for importance sampling is motivated by an approximation where `f_θ(μ, λ) ≈ μ`. How does the validity of this approximation vary across precision levels `λ`, and have you investigated the sensitivity of training stability and final performance to this specific choice of `p(λ)`? Could a poorly chosen `p(λ)` mask or artificially inflate the model's true performance?\n\n*   **Question 3:** The comparison with VDM is positioned as a key result. However, Figure 2 shows that the nature of the noisiest inputs for VDM and BSI are qualitatively different. Could the superior FID of BSI be an artifact of the DiT architecture being better suited to BSI's noise distribution and prior, rather than an inherent advantage of the BSI framework itself? How would the models compare with an architecture less biased towards global attention, like a pure U-Net on the same task?\n\n*   **Question 4:** The central claim for BSI's improved FID over BFN is the use of a non-deterministic initial belief (`λ₀ < ∞`). However, this is only validated by comparing the two extremes. The paper lacks a proper ablation study on the effect of the initial noise variance `λ₀⁻¹`. How does the FID score change as `λ₀⁻¹` is varied between 0 (BFN) and 100 (BSI)?\n\n*   **Question 5:** Training is performed by optimizing the infinite-step ELBO (`L_M`), which involves a continuous integral over precision `λ`. Sampling, however, is a discrete, finite-step process. What is the theoretical justification for this being the optimal training strategy for finite-step sampling, and how does this approach affect performance in the very low-step sampling regime (e.g., < 20 steps)?\n\n*   **Question 6:** The training objective, which minimizes the squared error between the true sample `x` and the model's prediction `f_θ(μ_λ, λ)`, bears a strong resemblance to the denoising score-matching objective. Can the BSI objective be formally interpreted as a form of score matching? Specifically, how does the BSI update rule relate to the updates in score-based generative modeling, and could this connection provide further insights into the model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Bayesian Sample Inference (BSI), a novel generative model that frames the generation process as iterative Bayesian inference. The core idea is to treat a data sample as an unknown variable and to iteratively refine a belief about it, starting from a broad prior and updating it based on noisy measurements of model predictions. The authors derive an Evidence Lower Bound (ELBO) for both finite and infinite-step versions of this process, enabling likelihood-based training. Through this framework, they show that Bayesian Flow Networks (BFNs) are a special case of BSI, providing a new perspective on them. Experiments on ImageNet32 and CIFAR10 demonstrate that BSI achieves sample quality (FID) superior to both Variational Diffusion Models (VDM) and BFNs, while maintaining equivalent log-likelihoods (BPD). The significance of this work lies in its introduction of a conceptually simple and elegant Bayesian perspective on generative modeling that unifies existing models and proves to be empirically effective.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "*   **Novelty:** The primary strength of the paper is its novel framing of generative modeling. Viewing sample generation as the process of inferring a fixed but unknown variable via iterative posterior updates is a conceptually clean and powerful idea. This Bayesian perspective is a welcome contribution, offering a different way to think about the gradual refinement process that characterizes modern generative models.\n\n*   **Theoretical Soundness:** The paper provides a rigorous theoretical foundation for BSI. The derivation of the ELBO (Theorems 3.1 and 3.2) is solid and connects the model to the established practice of variational inference. A key contribution is showing that BFNs emerge as a special case of BSI when the initial prior is deterministic (infinite precision). This not only unifies BFNs within a more general framework but also simplifies and clarifies their connection to diffusion models. The analysis of the \"noising\" process (Section A.2) and how BSI's non-Markovian forward process differs from BFN's Markovian one is insightful.\n\n*   **Strong Empirical Results:** The experimental validation, though limited in scope, is compelling within its domain. The authors compare BSI against two highly relevant and strong baselines: VDM and BFN. On the ImageNet32 benchmark, BSI achieves a statistically significant improvement in sample quality (FID of 8.9 vs. 9.9 for VDM and 11.0 for BFN) while matching their performance on log-likelihood. This is a strong result that demonstrates a clear advantage of the proposed formulation.", "weaknesses": "Despite its elegant formulation and promising results I still have doubts on the methodology and experimental validation, which raise several critical questions that weaken its overall contribution. The claims of novelty and superiority feel overstated upon closer inspection of the underlying algorithm and the lack of crucial ablation studies.\n\n1.  **Conflation of Re-framing with Algorithmic Novelty:** The core training algorithm (Algorithm 2) is structurally almost identical to standard denoising diffusion models. It samples a noise level (`λ`), creates a noisy input (`μ_λ`), and trains a network `f_θ` to predict the original sample `x` from `μ_λ`. While the Bayesian derivation is new, the final algorithm is not. This raises the question of whether BSI is a fundamentally new *class* of models or a well-motivated re-derivation of a specific noise schedule and prior for existing models.\n\n2.  **Lack of Ablation Studies:** The paper's central claim is that BSI's performance advantage over BFN comes from its non-deterministic prior (`p(μ₀)` with `λ₀⁻¹ > 0`). This is a critical claim that is not substantiated with sufficient evidence. The paper only compares the two extremes: BSI (`λ₀⁻¹ = 100`) and BFN (`λ₀⁻¹ = 0`). A proper scientific validation would require an ablation study showing how FID and BPD vary across a range of `λ₀⁻¹` values to demonstrate a clear trend and justify the chosen hyperparameter. The same applies to other key design choices, such as the `p(λ)` proposal distribution.\n\n3.  **Questionable Experimental Comparison:** The main empirical comparison is between BSI, BFN, and VDM using a DiT-L-2 backbone. Figure 2 clearly shows that the input distributions for these models at the highest noise levels are qualitatively different. BSI and BFN start from (near) pure noise, whereas VDM's noisiest input mean (`Np(0.08x, 1)`) still contains significant structural information about the image. The DiT architecture, with its global self-attention mechanism, is particularly well-suited to reconstructing images from unstructured noise. This setup may be unfairly biased towards BSI/BFN and confounding the architectural strengths with the merits of the generative formulation.\n\n4.  **Disconnect Between Training Objective and Sampling:** The model is trained to optimize an \"infinite-step\" ELBO, which is a continuous integral over all precision levels `λ`. However, sampling is always performed in a finite, discrete number of steps. The paper does not provide a theoretical justification for why optimizing the continuous objective is ideal for the discrete sampling task. This is particularly relevant for the low-step sampling regime, which is crucial for practical efficiency but not explored in the paper.", "questions": "*   **Question 1:** The final training objective (optimizing `L_M`) amounts to a noise-conditioned denoising task, structurally similar to other diffusion models. Beyond the novel derivation, what is the fundamental algorithmic difference between BSI and a VDM/BFN where the noise schedule is designed to match BSI's effective `λ` distribution and a non-zero initial noise level is used?\n\n*   **Question 2:** The choice of a log-uniform proposal distribution for importance sampling is motivated by an approximation where `f_θ(μ, λ) ≈ μ`. How does the validity of this approximation vary across precision levels `λ`, and have you investigated the sensitivity of training stability and final performance to this specific choice of `p(λ)`? Could a poorly chosen `p(λ)` mask or artificially inflate the model's true performance?\n\n*   **Question 3:** The comparison with VDM is positioned as a key result. However, Figure 2 shows that the nature of the noisiest inputs for VDM and BSI are qualitatively different. Could the superior FID of BSI be an artifact of the DiT architecture being better suited to BSI's noise distribution and prior, rather than an inherent advantage of the BSI framework itself? How would the models compare with an architecture less biased towards global attention, like a pure U-Net on the same task?\n\n*   **Question 4:** The central claim for BSI's improved FID over BFN is the use of a non-deterministic initial belief (`λ₀ < ∞`). However, this is only validated by comparing the two extremes. The paper lacks a proper ablation study on the effect of the initial noise variance `λ₀⁻¹`. How does the FID score change as `λ₀⁻¹` is varied between 0 (BFN) and 100 (BSI)?\n\n*   **Question 5:** Training is performed by optimizing the infinite-step ELBO (`L_M`), which involves a continuous integral over precision `λ`. Sampling, however, is a discrete, finite-step process. What is the theoretical justification for this being the optimal training strategy for finite-step sampling, and how does this approach affect performance in the very low-step sampling regime (e.g., < 20 steps)?\n\n*   **Question 6:** The training objective, which minimizes the squared error between the true sample `x` and the model's prediction `f_θ(μ_λ, λ)`, bears a strong resemblance to the denoising score-matching objective. Can the BSI objective be formally interpreted as a form of score matching? Specifically, how does the BSI update rule relate to the updates in score-based generative modeling, and could this connection provide further insights into the model?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761386316682}, {"id": "RhHULvOUsY", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13057/Reviewer_WnXZ"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "The authors present a generative algorithm that can be framed as iterative Bayesian inference. They derive an ELBO objective for the model, and show in a series of numerical experiments that they perform competitively with related models such as Bayesian Flow Networks, or Variational Diffusion models, on ImageNet and CIFAR datasets. The appendix carefully discusses the differences with other frameworks, notably diffusion models.", "review_text": "The authors present a generative algorithm that can be framed as iterative Bayesian inference. They derive an ELBO objective for the model, and show in a series of numerical experiments that they perform competitively with related models such as Bayesian Flow Networks, or Variational Diffusion models, on ImageNet and CIFAR datasets. The appendix carefully discusses the differences with other frameworks, notably diffusion models.", "strengths": "The paper is well written, and technical statements are satisfyingly accompanied with explanatory discussion. The interpretation of the generative model in terms of iterative Bayesian inference is interesting, and the numerical experiments seem convincing, although I have limited expertise in evaluating the latter.", "weaknesses": "My main concern revolves around the novelty of the theoretical generative framework, and how it relates to existing works. The authors already provide a detailed discussion in Appendix A.2 on the differences with standard diffusion models, highlighting how the noising process in their framework is non-Markovian. However, the closely related stochastic interpolant models (see e.g. Albergo et al, Stochastic Interpolants: A Unifying Framework for Flows and Diffusions, 2023) is to the best of my reading not discussed at all. The noising process in the latter seem to the best of my understanding very closely related to that considered by the authors. Notably, the denoising process (1) that involves the posterior denoiser of $x$ given the noisy observation $y$ resembles e.g. (4.3) in the provided reference, notably in the continuous limit of l.170. It is not impossible I have misunderstood some key contributions of the submission, and I am happy to raise my score after clarifications of the authors. However, I cannot be in favor of acceptance unless a detailed discussion is included to clarify differences to this line of works.", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present a generative algorithm that can be framed as iterative Bayesian inference. They derive an ELBO objective for the model, and show in a series of numerical experiments that they perform competitively with related models such as Bayesian Flow Networks, or Variational Diffusion models, on ImageNet and CIFAR datasets. The appendix carefully discusses the differences with other frameworks, notably diffusion models.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "The paper is well written, and technical statements are satisfyingly accompanied with explanatory discussion. The interpretation of the generative model in terms of iterative Bayesian inference is interesting, and the numerical experiments seem convincing, although I have limited expertise in evaluating the latter.", "weaknesses": "My main concern revolves around the novelty of the theoretical generative framework, and how it relates to existing works. The authors already provide a detailed discussion in Appendix A.2 on the differences with standard diffusion models, highlighting how the noising process in their framework is non-Markovian. However, the closely related stochastic interpolant models (see e.g. Albergo et al, Stochastic Interpolants: A Unifying Framework for Flows and Diffusions, 2023) is to the best of my reading not discussed at all. The noising process in the latter seem to the best of my understanding very closely related to that considered by the authors. Notably, the denoising process (1) that involves the posterior denoiser of $x$ given the noisy observation $y$ resembles e.g. (4.3) in the provided reference, notably in the continuous limit of l.170. It is not impossible I have misunderstood some key contributions of the submission, and I am happy to raise my score after clarifications of the authors. However, I cannot be in favor of acceptance unless a detailed discussion is included to clarify differences to this line of works.", "questions": "Please see the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760952488924}], "openreview_url": "https://openreview.net/forum?id=Xq4NdAodoA", "arxiv_id": "2502.07580", "paper_pdf": "papers/Xq4NdAodoA.pdf", "paper_pdf_sha256": "45bc0fe5d84ca629ba40629a7afd658a34a9e54b0740dedb4aeee71a1b489ce8", "paper_pdf_bytes": 1576494, "paper_pdf_source": "openreview", "code_url": "https://github.com/martenlienen/bsi", "code_repository": "martenlienen/bsi", "code_commit": "5b8a6acf17f2c7106d6fa31a474738e3b412c9c1", "code_archive": "repos/Xq4NdAodoA.zip", "code_archive_sha256": "00094761be0e38c281c9f17c770f2ff23a01f14944dbb71a63401fb7d9083827", "code_archive_bytes": 7537638, "code_file_count": 49, "code_extensions": {".py": 48, ".ipynb": 1}, "github_disk_usage_kb": 7457, "github_languages": {"Python": 197117, "Jupyter Notebook": 179113}, "github_archived": false, "github_pushed_at": "2026-08-14T05:54:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generative-modeling-with-bayesian-sample"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e3odKmatZr", "year": 2025, "status": "rejected", "title": "Critique-out-Loud Reward Models", "authors": ["Zachary Ankner", "Mansheej Paul", "Brandon Cui", "Jonathan Daniel Chang", "Prithviraj Ammanabrolu"], "authorids": ["~Zachary_Ankner1", "~Mansheej_Paul1", "~Brandon_Cui1", "~Jonathan_Daniel_Chang1", "~Prithviraj_Ammanabrolu1"], "authors_source": "OpenReview API", "abstract": "Traditionally, reward models used for reinforcement learning from human feedback (RLHF) are trained to directly predict preference scores without leveraging the generation capabilities of the underlying large language model (LLM). This limits the capabilities of reward models as they must reason implicitly about the quality of a response, i.e., preference modeling must be performed in a single forward pass through the model. To enable reward models to reason explicitly about the quality of a response, we introduce Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response that is then used to predict a scalar reward for the quality of the response. We demonstrate the success of CLoud reward models for both Llama-3-8B and 70B base models: compared to classic reward models CLoud reward models improve pairwise preference classification accuracy on RewardBench by 4.65 and 5.84 percentage points for the 8B and 70B base models respectively. Furthermore, CLoud reward models lead to a Pareto improvement for win rate on ArenaHard when used as the scoring model for Best-of-N. Finally, we explore how to exploit the dynamic inference compute capabilities of CLoud reward models by performing self-consistency decoding for reward prediction.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "CWVS7iz6JB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9196/Reviewer_T6ho"], "rating": 3, "soundness": 2, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper proposes a reward modeling approach that combines next-token-prediction loss with binary rating loss, which improves the performance of the Llama3-8B and Llama3-70B language models on RewardBench and BoN. The authors conducted extensive ablation studies to analyze their proposed method.", "review_text": "This paper proposes a reward modeling approach that combines next-token-prediction loss with binary rating loss, which improves the performance of the Llama3-8B and Llama3-70B language models on RewardBench and BoN. The authors conducted extensive ablation studies to analyze their proposed method.", "strengths": "- Well written, the technical approach is very clear, and Figure 1 and Figure 2 provide a clear understanding of the CLoud reward model training process.\n- Based on the experimental results published by the authors, the CLoud method seems to effectively improve the model's performance on RewardBench.\n- The author's ablation studies are very comprehensive.", "weaknesses": "- One of the most critical applications of the Reward Model is RLHF, yet the authors have not verified the improvement of CLoud RM over traditional RMs in RLHF. I believe this would greatly limit the widespread application of this method.\n- It is not very clear what significant issues in Reward Modeling CLoud has addressed, such as reward overoptimization. From the experimental results, it appears to only enhance the reward model's ability to distinguish between binary responses.\n- The experimental setup seems inconsistent. Based on my understanding from page 5, the authors' preference dataset and BoN dataset are derived from sampling Llama3-8B-Instruct. However, is it appropriate to use data generated by an 8B model for experiments with a 70B model? What is the rationale behind this choice? The authors should present this in the paper.\n- The above point leads to my skepticism about the author's claim that \"8B CLoud reward model even outperforms the 70B classic reward model.\"\n- The experiment on self-consistency is interesting, but the author's research is not thorough enough. If they could investigate how the quality of critique generation affects scoring accuracy, and combine this with methods like COT, TOT, etc., for further exploration, it would be more impressive.\n\n**Minors**\n\n- I'm not particularly fond of the title \"Results\" for Section 3; terms like \"Experiment\" might be more appropriate.\n- RQ2 in Section 3 seems a bit odd because there hasn't been any prior mention of on-policy/off-policy concepts, which makes it difficult to understand.", "questions": "- How much more overhead does CLoud RM incur during inference compared to traditional RM? Given that CLoud RM needs to generate an entire sequence before scoring, it is foreseeable that it would be more time-consuming than traditional RM. My concern about the inference cost grew even more after discovering that the author used the self-consistency method for multiple sampling averages.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a reward modeling approach that combines next-token-prediction loss with binary rating loss, which improves the performance of the Llama3-8B and Llama3-70B language models on RewardBench and BoN. The authors conducted extensive ablation studies to analyze their proposed method.", "soundness": 2, "presentation": 4, "contribution": 2, "strengths": "- Well written, the technical approach is very clear, and Figure 1 and Figure 2 provide a clear understanding of the CLoud reward model training process.\n- Based on the experimental results published by the authors, the CLoud method seems to effectively improve the model's performance on RewardBench.\n- The author's ablation studies are very comprehensive.", "weaknesses": "- One of the most critical applications of the Reward Model is RLHF, yet the authors have not verified the improvement of CLoud RM over traditional RMs in RLHF. I believe this would greatly limit the widespread application of this method.\n- It is not very clear what significant issues in Reward Modeling CLoud has addressed, such as reward overoptimization. From the experimental results, it appears to only enhance the reward model's ability to distinguish between binary responses.\n- The experimental setup seems inconsistent. Based on my understanding from page 5, the authors' preference dataset and BoN dataset are derived from sampling Llama3-8B-Instruct. However, is it appropriate to use data generated by an 8B model for experiments with a 70B model? What is the rationale behind this choice? The authors should present this in the paper.\n- The above point leads to my skepticism about the author's claim that \"8B CLoud reward model even outperforms the 70B classic reward model.\"\n- The experiment on self-consistency is interesting, but the author's research is not thorough enough. If they could investigate how the quality of critique generation affects scoring accuracy, and combine this with methods like COT, TOT, etc., for further exploration, it would be more impressive.\n\n**Minors**\n\n- I'm not particularly fond of the title \"Results\" for Section 3; terms like \"Experiment\" might be more appropriate.\n- RQ2 in Section 3 seems a bit odd because there hasn't been any prior mention of on-policy/off-policy concepts, which makes it difficult to understand.", "questions": "- How much more overhead does CLoud RM incur during inference compared to traditional RM? Given that CLoud RM needs to generate an entire sequence before scoring, it is foreseeable that it would be more time-consuming than traditional RM. My concern about the inference cost grew even more after discovering that the author used the self-consistency method for multiple sampling averages.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730729394711}, {"id": "YDQQwoIwMF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9196/Reviewer_bVCe"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces “Critique-out-Loud” (CLoud) reward model, which aims to enhance reinforcement learning from human feedback (RLHF) by generating criticisms before giving rewards. This approach improves the performance and policy quality of preference classification, especially in LLM-based preference models, by unifying reward modelling with chain-of-thought reasoning.", "review_text": "This paper introduces “Critique-out-Loud” (CLoud) reward model, which aims to enhance reinforcement learning from human feedback (RLHF) by generating criticisms before giving rewards. This approach improves the performance and policy quality of preference classification, especially in LLM-based preference models, by unifying reward modelling with chain-of-thought reasoning.", "strengths": "An innovative approach is proposed that combines critique generation with reward prediction to address the limitations of traditional reward models that lack explicit reasoning.\nEmpirical results show significant performance improvements on both preference modelling and BoN compared to traditional models.\nProvides additional tests in strategy training and self-consistent decoding, demonstrating potential application scenarios.", "weaknesses": "Criticism generation methods, while innovative, may introduce bias if the criticisms reflect some fixed bias of LLMs rather than true preferences.\nThe process of generating criticisms can be viewed as a sort of distillation from the larger model, and the comparisons provided in the paper are slightly unfair compared to scoring directly using the reward model. More ablation should be provided.\nCritic generation and inference practices lack rational supervision. For GenRM, further discussion should be proposed on the construction of the critique process.", "questions": "The approach in the paper can be viewed as incremental compared to a general reward model, so performance increases do not fully substantiate its value. More experiments are needed to prove that Critique-out-Loud works universally.\n1. How does the performance of this type of operation compare to simply llm-judge to produce preference labeling?\n2. Do authors try to output the reward scores before generating the corresponding critiques, and whether this format also improve the performance of the reward model？\n3. How generalizable is the approach given that the existing training data is largely constructed on full domain samples? If this is the paradigm rather than a gain from the data itself, we should see an increase in scoring capabilities on OOD data when trained only on domain-specific data (e.g., trained on math data and tested on safety data)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces “Critique-out-Loud” (CLoud) reward model, which aims to enhance reinforcement learning from human feedback (RLHF) by generating criticisms before giving rewards. This approach improves the performance and policy quality of preference classification, especially in LLM-based preference models, by unifying reward modelling with chain-of-thought reasoning.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "An innovative approach is proposed that combines critique generation with reward prediction to address the limitations of traditional reward models that lack explicit reasoning.\nEmpirical results show significant performance improvements on both preference modelling and BoN compared to traditional models.\nProvides additional tests in strategy training and self-consistent decoding, demonstrating potential application scenarios.", "weaknesses": "Criticism generation methods, while innovative, may introduce bias if the criticisms reflect some fixed bias of LLMs rather than true preferences.\nThe process of generating criticisms can be viewed as a sort of distillation from the larger model, and the comparisons provided in the paper are slightly unfair compared to scoring directly using the reward model. More ablation should be provided.\nCritic generation and inference practices lack rational supervision. For GenRM, further discussion should be proposed on the construction of the critique process.", "questions": "The approach in the paper can be viewed as incremental compared to a general reward model, so performance increases do not fully substantiate its value. More experiments are needed to prove that Critique-out-Loud works universally.\n1. How does the performance of this type of operation compare to simply llm-judge to produce preference labeling?\n2. Do authors try to output the reward scores before generating the corresponding critiques, and whether this format also improve the performance of the reward model？\n3. How generalizable is the approach given that the existing training data is largely constructed on full domain samples? If this is the paradigm rather than a gain from the data itself, we should see an increase in scoring capabilities on OOD data when trained only on domain-specific data (e.g., trained on math data and tested on safety data)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730719891011}, {"id": "0TpUZEu2j9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9196/Reviewer_sPGi"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The authors propose Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response, which is then used to predict a scalar reward for the quality of the response. Their effectiveness is demonstrated across different models and benchmarks.The motivation behind this paper is interesting and contributes to the ongoing advancements in RLHF fine-tuning for current LLMs.\nThe writing is clear and easy to understand.", "review_text": "The authors propose Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response, which is then used to predict a scalar reward for the quality of the response. Their effectiveness is demonstrated across different models and benchmarks.The motivation behind this paper is interesting and contributes to the ongoing advancements in RLHF fine-tuning for current LLMs.\nThe writing is clear and easy to understand.", "strengths": "1. The motivation behind this paper is interesting and contributes to the ongoing advancements in RLHF fine-tuning for current LLMs.\n\n2. The writing is clear and easy to understand.", "weaknesses": "1. The authors constructed their dataset based on UltraFeedback and UltraInstruct, requiring Llama-3.1-405B-Instruct to generate critiques for chosen and non-chosen responses. This seems to introduce additional annotation costs and complexities, which is a drawback of the method. How effective would this pipeline be in a pure self-critique setting?\n\n2. The authors test the method's effectiveness in the safety domain, where critique is currently widely used in LLM safety alignment scenarios. This is intriguing. However, the authors' safety evaluation of the method on RewardBench raises concerns, as my understanding is that RewardBench aggregates existing datasets with inconsistent safety classifications across different safety datasets. Consequently, the detailed assessments in safety and reasoning may lack credibility. I recommend that the authors evaluate the method on specific datasets for different categories, such as reasoning on GSM8K [5] and safety on BeaverTails [6], which are designed for specific LLM categories.\n\n3. In validating the effectiveness of CLoud RM, apart from testing on RewardBench, the authors conducted only BoN experiments. What is the impact of Cloud RM on RLHF? This is a critical aspect that I would like to address, especially since Cloud RM adds the additional annotation cost of critiques. The authors claim that the scalar reward is more accurate but have not conducted any RLHF experiments [3].\n\n4. This method inevitably increases the complexity and cost of the pipeline. I am particularly interested in novel uses of critique beyond merely improving the accuracy of scalar rewards. Since it incorporates additional information, it is not surprising that the scalar reward's accuracy improves compared to a pure reward model. However, its influence on RLHF could yield other applications, such as directly aligning models using critiques across multiple dimensions, such as Aligner [4], which trains based on correct data (a type of critique) to enable offline alignment. It would be interesting to explore whether learning from richer critique feedback is feasible [2]. Simply using a critique head and a scoring head is not particularly innovative, as discussed in relation to help steer.\n\n5. Some intriguing ablation studies have not been conducted. For example, allowing the model to output a critique before producing a scalar reward could highlight the differences between training using the Cloud RM approach versus traditional methods. (I am unaware if such an experiment has been done; please correct me if I’m mistaken.)\n\n[1] Self-critiquing models for assisting human evaluators\n\n[2] Training Language Models with Language Feedback\n\n[3] OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework\n\n[4] Aligner: Efficient Alignment by Learning to Correct\n\n[5] https://huggingface.co/datasets/openai/gsm8k\n\n[6] https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF\n\nOverall, the motivation behind this paper is compelling. I welcome further discussions with the authors during the rebuttal period.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response, which is then used to predict a scalar reward for the quality of the response. Their effectiveness is demonstrated across different models and benchmarks.The motivation behind this paper is interesting and contributes to the ongoing advancements in RLHF fine-tuning for current LLMs.\nThe writing is clear and easy to understand.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The motivation behind this paper is interesting and contributes to the ongoing advancements in RLHF fine-tuning for current LLMs.\n\n2. The writing is clear and easy to understand.", "weaknesses": "1. The authors constructed their dataset based on UltraFeedback and UltraInstruct, requiring Llama-3.1-405B-Instruct to generate critiques for chosen and non-chosen responses. This seems to introduce additional annotation costs and complexities, which is a drawback of the method. How effective would this pipeline be in a pure self-critique setting?\n\n2. The authors test the method's effectiveness in the safety domain, where critique is currently widely used in LLM safety alignment scenarios. This is intriguing. However, the authors' safety evaluation of the method on RewardBench raises concerns, as my understanding is that RewardBench aggregates existing datasets with inconsistent safety classifications across different safety datasets. Consequently, the detailed assessments in safety and reasoning may lack credibility. I recommend that the authors evaluate the method on specific datasets for different categories, such as reasoning on GSM8K [5] and safety on BeaverTails [6], which are designed for specific LLM categories.\n\n3. In validating the effectiveness of CLoud RM, apart from testing on RewardBench, the authors conducted only BoN experiments. What is the impact of Cloud RM on RLHF? This is a critical aspect that I would like to address, especially since Cloud RM adds the additional annotation cost of critiques. The authors claim that the scalar reward is more accurate but have not conducted any RLHF experiments [3].\n\n4. This method inevitably increases the complexity and cost of the pipeline. I am particularly interested in novel uses of critique beyond merely improving the accuracy of scalar rewards. Since it incorporates additional information, it is not surprising that the scalar reward's accuracy improves compared to a pure reward model. However, its influence on RLHF could yield other applications, such as directly aligning models using critiques across multiple dimensions, such as Aligner [4], which trains based on correct data (a type of critique) to enable offline alignment. It would be interesting to explore whether learning from richer critique feedback is feasible [2]. Simply using a critique head and a scoring head is not particularly innovative, as discussed in relation to help steer.\n\n5. Some intriguing ablation studies have not been conducted. For example, allowing the model to output a critique before producing a scalar reward could highlight the differences between training using the Cloud RM approach versus traditional methods. (I am unaware if such an experiment has been done; please correct me if I’m mistaken.)\n\n[1] Self-critiquing models for assisting human evaluators\n\n[2] Training Language Models with Language Feedback\n\n[3] OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework\n\n[4] Aligner: Efficient Alignment by Learning to Correct\n\n[5] https://huggingface.co/datasets/openai/gsm8k\n\n[6] https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF\n\nOverall, the motivation behind this paper is compelling. I welcome further discussions with the authors during the rebuttal period.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730538366332}, {"id": "F4oLL9INdE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9196/Reviewer_pirR"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 4, "confidence": 5, "summary": "The paper proposes an interesting idea of joining training the LM head and RM head for reward model so that it first generates CoT-type natural-langauge-based critique, then let the reward model generate new scalar score. Experiments show that the performance greatly improves compared with normal reward training pipelines.", "review_text": "The paper proposes an interesting idea of joining training the LM head and RM head for reward model so that it first generates CoT-type natural-langauge-based critique, then let the reward model generate new scalar score. Experiments show that the performance greatly improves compared with normal reward training pipelines.", "strengths": "The idea of training an LM head for critique appears neat and novel. The experiments are sound and look at both the performance on reward bench and arena hard for downstream performance of a trained model.", "weaknesses": "I don't see major weakness from the paper and I like the proposed idea a lot. But I would appreciate if the authors could help clarify some of the questions I have below. Please refer to question section.", "questions": "1. The authors use Llama 405B to generate oracle critique data for SFT, and replace the oracle data with fine-tuned model's self-genereated critique. Usually when you distill from larger models, there will be direct gain in model performance. How do we know that if the performance gain is mainly from distilling responses from large models, or from the methodology itself? If the critique SFT data is mostly self-generated (not on the fine-tuned one), would we still expect such high gain in downstream performance? This will be helpful for the case where we want to improve the largest model available with only binary human feedback.\n\n2. The generation of SFT data is using llama 3.1 405B but the training model is llama 3 instead of llama 3.1, what's the reason behind this? Would the gain of the same methodology appear weaker when it comes to stronger base RM?\n\n3. Does the gain mainly come from the diverse critiques generated, or from more clear instructions of what the RM should look like? For example, if we provide a very comprehensive fixed system prompt asking the response to focus on safety, helpfulness, comprehensiveness, (e.g. by modifying the prompt in Figure 9) and provide a combination of system prompt + user prompt + response to RM, would it achieve similar performance as user prompt + response + critique?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an interesting idea of joining training the LM head and RM head for reward model so that it first generates CoT-type natural-langauge-based critique, then let the reward model generate new scalar score. Experiments show that the performance greatly improves compared with normal reward training pipelines.", "soundness": 4, "presentation": 4, "contribution": 4, "strengths": "The idea of training an LM head for critique appears neat and novel. The experiments are sound and look at both the performance on reward bench and arena hard for downstream performance of a trained model.", "weaknesses": "I don't see major weakness from the paper and I like the proposed idea a lot. But I would appreciate if the authors could help clarify some of the questions I have below. Please refer to question section.", "questions": "1. The authors use Llama 405B to generate oracle critique data for SFT, and replace the oracle data with fine-tuned model's self-genereated critique. Usually when you distill from larger models, there will be direct gain in model performance. How do we know that if the performance gain is mainly from distilling responses from large models, or from the methodology itself? If the critique SFT data is mostly self-generated (not on the fine-tuned one), would we still expect such high gain in downstream performance? This will be helpful for the case where we want to improve the largest model available with only binary human feedback.\n\n2. The generation of SFT data is using llama 3.1 405B but the training model is llama 3 instead of llama 3.1, what's the reason behind this? Would the gain of the same methodology appear weaker when it comes to stronger base RM?\n\n3. Does the gain mainly come from the diverse critiques generated, or from more clear instructions of what the RM should look like? For example, if we provide a very comprehensive fixed system prompt asking the response to focus on safety, helpfulness, comprehensiveness, (e.g. by modifying the prompt in Figure 9) and provide a combination of system prompt + user prompt + response to RM, would it achieve similar performance as user prompt + response + critique?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730485908639}], "openreview_url": "https://openreview.net/forum?id=e3odKmatZr", "arxiv_id": "2408.11791", "paper_pdf": "papers/e3odKmatZr.pdf", "paper_pdf_sha256": "7a128b93c833ac7cfc77cf2cbba0d1770f1c289609ec2b549cfc84625fe53bc3", "paper_pdf_bytes": 1763330, "paper_pdf_source": "openreview", "code_url": "https://github.com/zankner/CLoud", "code_repository": "zankner/CLoud", "code_commit": "fac417e2c0f45fab9083bf9b77066de490e8288c", "code_archive": "repos/e3odKmatZr.zip", "code_archive_sha256": "da1b3911374b086d2d2cb64f352ce2ca1fa5847fb6d598cfb3d7e113429a8e24", "code_archive_bytes": 378121, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 362, "github_languages": {"Python": 95335}, "github_archived": false, "github_pushed_at": "2024-10-18T19:38:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/critique-out-loud-reward-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "01LMSeReNvY", "year": 2023, "status": "rejected", "title": "PromptBoosting: Black-Box Text Classification with Ten Forward Passes", "authors": ["Bairu Hou", "Joe O'Connor", "Jacob Andreas", "Shiyu Chang", "Yang Zhang"], "authorids": ["~Bairu_Hou2", "~Joe_O'Connor1", "~Jacob_Andreas1", "~Shiyu_Chang2", "~Yang_Zhang3"], "authors_source": "OpenReview API", "abstract": "We describe PromptBoosting, a query-efficient procedure for building a text classifier from a neural language model (LM) without access to the LM’s parameters, gradients, or hidden representations. This form of “black-box” classifier training has become increasingly important as the cost of training and inference in large-scale LMs grows. But existing black-box LM classifier learning approaches are themselves computationally inefficient, typically specializing LMs to the target task by searching in a large space of (discrete or continuous) prompts using zeroth-order optimization methods. Instead of directly optimizing in prompt space, PromptBoosting obtains a small pool of prompts via a gradient-free approach and then constructs a large pool of weak learners by pairing these prompts with different elements of the LM’s output distribution. These weak learners are then ensembled using the AdaBoost algorithm. The entire learning process requires only a small number of forward passes and no backward pass. Experiments show that PromptBoosting achieves state-of-the-art performance in multiple black-box few-shot classification tasks, and matches or outperforms full fine-tuning in both few-shot and standard learning paradigms, while training 10x faster than existing black-box methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Frm57yAMh-", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5684/Reviewer_QXXp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThis paper proposes a prompting-based method for building a text classifier from language models (LMs), which is called PromptBoosting. The idea of the proposed method is to first obtain a small pool of prompts via gradient-free approach, and then contract a large pool of weak learners by pairing the prompts with different elements of the LM’s output distribution. Then the adaboost algorithm is applied to ensemble the weak learners. Empirical studies claim that PromptBoosting can achieve state-of-the-art performance in different few-shot classification tasks while being 10x faster than existing black-box methods.", "review_text": "The paper applies the adaboost algorithm to the existing prompting methods to get the better text classifier with pretrained LM. While the experimental results look promising, the lack of novelty and comprehensiveness of experiments are two concerns.", "strengths": "Strength:\n\n1. The method is very intuitive in applying the boosting idea to the prompt-based methods such as Gao et al. (2021). Also, the paper is well written. So in general, this paper is easy to understand.\n2. Learning the verbalizer has a closed form solution, which makes it very efficient.\n\nWeaknesses:\n\n1. The novelty of the paper seems slim. It seems like a direct application of adaboost on the existing prompting method Gao et al. (2021).\n2. The evaluation is not convincing. \n    2.1. First of all, according to the implementation, this method can only be applied to text classification tasks, making it a bit narrow. \n    2.2. Besides, while there a quite a few text classification tasks, the authors only select 8 without giving the rational why they are selected. In contrast, Gao et at. (2021) evaluates their method on 16 tasks, mostly from (Super)GLUE. I suggest  the author at least extend the evaluation on more representative tasks to show the effectiveness is indeed universal.\n    2.3. Finally, there should be some simple baselines to compare with, e.g., prompt ensemble with majority vote. \u0010But they are never shown up in the experiments.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\nThis paper proposes a prompting-based method for building a text classifier from language models (LMs), which is called PromptBoosting. The idea of the proposed method is to first obtain a small pool of prompts via gradient-free approach, and then contract a large pool of weak learners by pairing the prompts with different elements of the LM’s output distribution. Then the adaboost algorithm is applied to ensemble the weak learners. Empirical studies claim that PromptBoosting can achieve state-of-the-art performance in different few-shot classification tasks while being 10x faster than existing black-box methods.", "strength_and_weaknesses": "Strength:\n\n1. The method is very intuitive in applying the boosting idea to the prompt-based methods such as Gao et al. (2021). Also, the paper is well written. So in general, this paper is easy to understand.\n2. Learning the verbalizer has a closed form solution, which makes it very efficient.\n\nWeaknesses:\n\n1. The novelty of the paper seems slim. It seems like a direct application of adaboost on the existing prompting method Gao et al. (2021).\n2. The evaluation is not convincing. \n    2.1. First of all, according to the implementation, this method can only be applied to text classification tasks, making it a bit narrow. \n    2.2. Besides, while there a quite a few text classification tasks, the authors only select 8 without giving the rational why they are selected. In contrast, Gao et at. (2021) evaluates their method on 16 tasks, mostly from (Super)GLUE. I suggest  the author at least extend the evaluation on more representative tasks to show the effectiveness is indeed universal.\n    2.3. Finally, there should be some simple baselines to compare with, e.g., prompt ensemble with majority vote. \u0010But they are never shown up in the experiments.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity of the paper is good that ideas are clearly conveyed.\n\nThe novelty is slim, as it sounds like a direct application of adaboost on the prompting methods.\n\nI did not check the code so I cannot judge the reproducibility.", "summary_of_the_review": "The paper applies the adaboost algorithm to the existing prompting methods to get the better text classifier with pretrained LM. While the experimental results look promising, the lack of novelty and comprehensiveness of experiments are two concerns.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667182450733}, {"id": "56lovpgLsrY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5684/Reviewer_jDoU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed PromptBoosting, which works on black box setting for text classification with LLM and prompt. The idea is to build a set of weak learners and each of them is associated with a prompt (raw text) and the final model is trained with adaboost over these weak learners. The resulting weak learners achieve state-of-the-art performance in multiple black-box few-shot text classification tasks, and matches or even outperforms full fine-tuning in both few-shot and standard learning.", "review_text": "The paper is well written and easy to read. The idea is simple and makes sense. There are lots of experiment provided in the paper showing good performance of the proposed method for text classification.  However I am mostly concerned about the limit of the scope of the work (limited to text classification only).", "strengths": "Strength:\n\n1: Black box setting for LLM with prompt is an interesting direction and would have lots of applications, especially for the cases the LLM is too large.\n\n2: Paper is well written and the proposed idea makes sense. \n\n3: Comparing with many methods and datasets, and showing better results than other methods.\n\nWeakness:\n\n1: The scope of the work seems limited. This work seems can only be used for text classification (due to the Verbalizer step which needs the end task to be classification task only). Therefore a large amount of use cases for LLM with few shot learning settings such as summarization, etc cannot be directly applied with the proposed method.\n\n2: Missing some ablation studies. Such as: the prompts sets are generated by optimization-free method proposed by Gao et al. (2021). How about using other prompt sets, maybe just original training examples as a simple baseline? And the adaboost is used for ensemble weak learners, how about use other ensemble methods?  \n\n3: Some of the experimental results need further explanation.\n 3.1 Is the training time (wall time) including prompt set generating time? Or just the time for training weak learners? Could you provide time for each step of the proposed method?\n\n3.2 In Table 2, DART and LM-BFF need lots of backward and forward passes of the model, why the training time is close to the proposed method? Does it mean the forward and backward time is not the bottleneck? \n\n3.3. And also finetuning results are almost always worse than the proposed method (and DART and LM-BFF), which is beyond my expectation. It is better to have some explanation for that.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed PromptBoosting, which works on black box setting for text classification with LLM and prompt. The idea is to build a set of weak learners and each of them is associated with a prompt (raw text) and the final model is trained with adaboost over these weak learners. The resulting weak learners achieve state-of-the-art performance in multiple black-box few-shot text classification tasks, and matches or even outperforms full fine-tuning in both few-shot and standard learning.", "strength_and_weaknesses": "Strength:\n\n1: Black box setting for LLM with prompt is an interesting direction and would have lots of applications, especially for the cases the LLM is too large.\n\n2: Paper is well written and the proposed idea makes sense. \n\n3: Comparing with many methods and datasets, and showing better results than other methods.\n\nWeakness:\n\n1: The scope of the work seems limited. This work seems can only be used for text classification (due to the Verbalizer step which needs the end task to be classification task only). Therefore a large amount of use cases for LLM with few shot learning settings such as summarization, etc cannot be directly applied with the proposed method.\n\n2: Missing some ablation studies. Such as: the prompts sets are generated by optimization-free method proposed by Gao et al. (2021). How about using other prompt sets, maybe just original training examples as a simple baseline? And the adaboost is used for ensemble weak learners, how about use other ensemble methods?  \n\n3: Some of the experimental results need further explanation.\n 3.1 Is the training time (wall time) including prompt set generating time? Or just the time for training weak learners? Could you provide time for each step of the proposed method?\n\n3.2 In Table 2, DART and LM-BFF need lots of backward and forward passes of the model, why the training time is close to the proposed method? Does it mean the forward and backward time is not the bottleneck? \n\n3.3. And also finetuning results are almost always worse than the proposed method (and DART and LM-BFF), which is beyond my expectation. It is better to have some explanation for that.", "clarity,_quality,_novelty_and_reproducibility": "1: The writing is good and easy to read.\n\n2: The work seems novel to me. \n\n3: Pre-released code for the proposed method is included in the supplemental material.", "summary_of_the_review": "The paper is well written and easy to read. The idea is simple and makes sense. There are lots of experiment provided in the paper showing good performance of the proposed method for text classification.  However I am mostly concerned about the limit of the scope of the work (limited to text classification only).", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667073204153}, {"id": "m3NRpazS3w", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5684/Reviewer_QZLQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes PROMPTBOOSTING, a novel black-box prompt learning approach that does not rely on searching for an optimal prompt and which can thus drastically improve the computational efficiency over the existing method. Specifically, the proposed approach obtains a small pool of prompts via a gradient-free approach and then constructs a large pool of weak learners by pairing these prompts with different elements of the LM’s output distribution. These weak learners are then ensembled using the AdaBoost algorithm. The learning process requires only a small number of forward passes per batch and no backward pass. Experiments show that PROMPTBOOSTING achieves state-of-the-art performance in multiple black-box few-shot classification tasks and matches or outperforms full fine-tuning in both few-shot and standard learning paradigms while training 10x faster than existing black-box methods. Overall, this paper is well-written and well-motivated. ", "review_text": "Overall, I think this paper is well-written and well-motivated. Some small issues are the lack of detailed analysis of different prompt templates. I think black-box tuning is a promising solution for foundation models, and this work can help inspire future work.", "strengths": "Strength：\n\n1.This paper proposes a novel approach that does not rely on searching for an optimal prompt and which can thus drastically improve the computational efficiency over the existing method. In my opinion, the proposed approach is well-written and well-motivated.\n\n2.Comprehensive experimental results show that the proposed approach is efficient and effective.\n\nWeaknesses:\n\nFrom my point of view, I think the prompt set may have a big influence on the performance. As the paper mentioned, \"the aforementioned approach can be replaced with any other optimization-free prompt generation methods,\" so can you provide the performance with other prompt generation methods?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes PROMPTBOOSTING, a novel black-box prompt learning approach that does not rely on searching for an optimal prompt and which can thus drastically improve the computational efficiency over the existing method. Specifically, the proposed approach obtains a small pool of prompts via a gradient-free approach and then constructs a large pool of weak learners by pairing these prompts with different elements of the LM’s output distribution. These weak learners are then ensembled using the AdaBoost algorithm. The learning process requires only a small number of forward passes per batch and no backward pass. Experiments show that PROMPTBOOSTING achieves state-of-the-art performance in multiple black-box few-shot classification tasks and matches or outperforms full fine-tuning in both few-shot and standard learning paradigms while training 10x faster than existing black-box methods. Overall, this paper is well-written and well-motivated. ", "strength_and_weaknesses": "Strength：\n\n1.This paper proposes a novel approach that does not rely on searching for an optimal prompt and which can thus drastically improve the computational efficiency over the existing method. In my opinion, the proposed approach is well-written and well-motivated.\n\n2.Comprehensive experimental results show that the proposed approach is efficient and effective.\n\nWeaknesses:\n\nFrom my point of view, I think the prompt set may have a big influence on the performance. As the paper mentioned, \"the aforementioned approach can be replaced with any other optimization-free prompt generation methods,\" so can you provide the performance with other prompt generation methods?\n", "clarity,_quality,_novelty_and_reproducibility": "See weaknesses. Besides, the code is without running scripts. How to run the code?", "summary_of_the_review": "Overall, I think this paper is well-written and well-motivated. Some small issues are the lack of detailed analysis of different prompt templates. I think black-box tuning is a promising solution for foundation models, and this work can help inspire future work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666706688540}, {"id": "bdQz9_J_OPa", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5684/Reviewer_cHte"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper deals with the problem of prompt-style finetuning when gradient information is not available. The solution is using adaboost to quickly learn prompt-style learners. Experimental results on standard benchmark showed competitive results over previous black-box methods.", "review_text": "Overall I like the research problem and the method proposed. Not very difficult or sophisticated solution but sounds working. I hope authors can add further studies to validate the root cause. Disclaimer is that I am not particularly familiar with fine-tuning under black-box methods. I do study black-box methods under other setups but not sure if this paper covers all recent works, and if recent works covered all possible black-box methods.", "strengths": "- Strength\n 1. Clearly written, well motivated\n 2. Competitive results\n 3. Reasonable methods\n\n-Weakness\n1. I do believe the method will work but somehow I don't see a further discussion on the best scheme when boosting is used.\n2. not so many black-box method is deployed in this direction, and there is no thorough discussion of the potential applicability of previous black-box methods. That is, I believe zero-order method has a broader literature but this paper didn't connect with that well.", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper deals with the problem of prompt-style finetuning when gradient information is not available. The solution is using adaboost to quickly learn prompt-style learners. Experimental results on standard benchmark showed competitive results over previous black-box methods.", "strength_and_weaknesses": "- Strength\n 1. Clearly written, well motivated\n 2. Competitive results\n 3. Reasonable methods\n\n-Weakness\n1. I do believe the method will work but somehow I don't see a further discussion on the best scheme when boosting is used.\n2. not so many black-box method is deployed in this direction, and there is no thorough discussion of the potential applicability of previous black-box methods. That is, I believe zero-order method has a broader literature but this paper didn't connect with that well.", "clarity,_quality,_novelty_and_reproducibility": "1. Very clearly written paper. Well motivated examples.\n2. Reproducibility could be a problem as it's about wall clock time. And somehow the process of base prompt generation process is not clear to me.\n3. Novelty wise I think it's an interesting way to learn the model under zero-gradient case, but further studies can be done to investigate more.\n\nFollowings are my main questions:\n\n1. I'd like to see how boosting helps. Is it possible for you to build the prompt base using gradients and then apply it with adaboost? In sum, I do think the proposed base prompt generation is not the best (due to gradient limitation) so supposedly a better prompt + boosting will go better. \n\n2. Can you also provide number for prompt ensemble methods? Want to see the difference. How good/bad compared to vanilla fine-tuning.\n\n3. other baseline methods on SST-2 MR  performs well but failed significantly on the rest. That's strange to me and it requires an explanation to this.\n\n4. I am wondering the transferrability of RoBERTa to GPT-3. It's known that prompt-base methods works much better on GPT-3 compared to GPT-2. RoBERTA is a much smaller model compared to GPT-3. Not sure if this will make the experiments in the paper much worthless. \n\n", "summary_of_the_review": "Overall I like the research problem and the method proposed. Not very difficult or sophisticated solution but sounds working. I hope authors can add further studies to validate the root cause. Disclaimer is that I am not particularly familiar with fine-tuning under black-box methods. I do study black-box methods under other setups but not sure if this paper covers all recent works, and if recent works covered all possible black-box methods.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666566715397}], "openreview_url": "https://openreview.net/forum?id=01LMSeReNvY", "arxiv_id": "2212.09257", "paper_pdf": "papers/01LMSeReNvY.pdf", "paper_pdf_sha256": "a7492816417762dbfd110f87eb6b6cbf4b16e9d7c1806461fa8e949b8ed458be", "paper_pdf_bytes": 700916, "paper_pdf_source": "openreview", "code_url": "https://github.com/UCSB-NLP-Chang/PromptBoosting", "code_repository": "UCSB-NLP-Chang/PromptBoosting", "code_commit": "6843af7da67a27f21f1de76950a7f0c0f3fb7785", "code_archive": "repos/01LMSeReNvY.zip", "code_archive_sha256": "df0b24f71dc2ce1c47abf0ac5eb5f215ae73dbdd41bb8600db3ebb40c36f3a11", "code_archive_bytes": 1088188, "code_file_count": 23, "code_extensions": {".py": 23}, "github_disk_usage_kb": 747, "github_languages": {"Python": 193104}, "github_archived": false, "github_pushed_at": "2023-09-05T19:08:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/promptboosting-black-box-text-classification"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FOR2VqgJXb", "year": 2021, "status": "rejected", "title": "Evaluating representations by the complexity of learning low-loss predictors", "authors": ["William F Whitney", "Min Jae Song", "David Brandfonbrener", "Jaan Altosaar", "Kyunghyun Cho"], "authorids": ["~William_F_Whitney1", "~Min_Jae_Song1", "~David_Brandfonbrener1", "~Jaan_Altosaar1", "~Kyunghyun_Cho1"], "authors_source": "OpenReview API", "abstract": "We consider the problem of evaluating representations of data for use in solving a downstream task. We propose to measure the quality of a representation by the complexity of learning a predictor on top of the representation that achieves low loss on a task of interest. To this end, we introduce two measures: surplus description length (SDL) and $\\varepsilon$ sample complexity ($\\varepsilon$SC). To compare our methods to prior work, we also present a framework based on plotting the validation loss versus dataset size (the \"loss-data\" curve). Existing measures, such as mutual information and minimum description length, correspond to slices and integrals along the data-axis of the loss-data curve, while ours correspond to slices and integrals along the loss-axis. This analysis shows that prior methods measure properties of an evaluation dataset of a specified size, whereas our methods measure properties of a predictor with a specified loss. We conclude with experiments on real data to compare the behavior of these methods over datasets of varying size.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "RSRyFi7TPoH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2491/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors address the issue of evaluating the quality of representations based on performance of a classifier for a downstream task. The premise is that conventional metrics for the downstream classifier such as Validation Accuracy, MDL or MI are flawed due to dependence on the size of the dataset (VA, MDL) or because they ignore statistical/computational complexity of the classifier (MI).\nThe alternatives proposed in the paper SDL/\\epsilon-SC, which attempt to measure the complexity of learning the classifier to \\epsilon-tolerance.\n\nThe core premise is reasonable. The metric/evaluation of the representation should not depend on the amount of data available for the downstream task. However, it seems like the proposed alternative replaces an arbitrary dataset size, with an arbitrary loss threshold (small values of \\epsilon may lead no measurement (infinity for both SDL and \\epsilon-SC), while larger values of \\epsilon may also lead to arbitrariness. The authors do say that the they want an evaluation measure which depends on the data distribution, but seem to not pay sufficient attention to the \"distribution\" part (measurements are noisy, and we should understand what is noise, and what is a significant improvement, possibly by computing confidence intervals). The authors specifically mention a downside of MI is that it ignores computational complexity of the classifiers, but that seems to be a valid complaint of their approach as well (presumably more (computationally) complex classifiers can achieve a loss threshold), though their approach does address the issue with statistical complexity. \nThere are other issues - there are typically many tasks that we might want to use a common representation for. The arbitrariness of epsilon makes combining evaluations across different tasks difficult. Further, we typically have a fixed amount of data for each downstream task - it would seem that the approach proposed here would require an arbitrary amount of data for each task, which may not be feasible.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for \"Evaluating representations by the complexity of learning low-loss predictors\"", "review": "The authors address the issue of evaluating the quality of representations based on performance of a classifier for a downstream task. The premise is that conventional metrics for the downstream classifier such as Validation Accuracy, MDL or MI are flawed due to dependence on the size of the dataset (VA, MDL) or because they ignore statistical/computational complexity of the classifier (MI).\nThe alternatives proposed in the paper SDL/\\epsilon-SC, which attempt to measure the complexity of learning the classifier to \\epsilon-tolerance.\n\nThe core premise is reasonable. The metric/evaluation of the representation should not depend on the amount of data available for the downstream task. However, it seems like the proposed alternative replaces an arbitrary dataset size, with an arbitrary loss threshold (small values of \\epsilon may lead no measurement (infinity for both SDL and \\epsilon-SC), while larger values of \\epsilon may also lead to arbitrariness. The authors do say that the they want an evaluation measure which depends on the data distribution, but seem to not pay sufficient attention to the \"distribution\" part (measurements are noisy, and we should understand what is noise, and what is a significant improvement, possibly by computing confidence intervals). The authors specifically mention a downside of MI is that it ignores computational complexity of the classifiers, but that seems to be a valid complaint of their approach as well (presumably more (computationally) complex classifiers can achieve a loss threshold), though their approach does address the issue with statistical complexity. \nThere are other issues - there are typically many tasks that we might want to use a common representation for. The arbitrariness of epsilon makes combining evaluations across different tasks difficult. Further, we typically have a fixed amount of data for each downstream task - it would seem that the approach proposed here would require an arbitrary amount of data for each task, which may not be feasible.\n\n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604260353222}, {"id": "fZlgfBFBzPm", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2491/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "*Summary of the paper:\nThe paper discusses new measures to evaluating the quality of representations.\nThey present two new measures namely SDL and $\\epsilon$-sample complexity, describe their benefits and apply experimental study, comparing these measures to standard baselines.\n\n* Comments\nThe suggested measures are simple modifications of existing baselines (MDL,MI).\nThe authors emphasize that their measures are independent of dataset size $n$, but this only applies when $n$ is large enough and when $\\epsilon$ is a good estimate of the conditional entropy.\n\nThe authors also mention two issues regarding existing measures (MDL,MI)\n- They are insensitive to statistical complexity\n- They are insensitive to computational complexity\nwhile these are true for the theoretical measures, I don't see what is the benefit of the current measures in this context. This is since old and new measures are leaning on an approximate  computation of $L(A,i)$.\n\n\n*Summary of review\nI don't see any clear benefit of the suggested measures.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "*Summary of the paper:\nThe paper discusses new measures to evaluating the quality of representations.\nThey present two new measures namely SDL and $\\epsilon$-sample complexity, describe their benefits and apply experimental study, comparing these measures to standard baselines.\n\n* Comments\nThe suggested measures are simple modifications of existing baselines (MDL,MI).\nThe authors emphasize that their measures are independent of dataset size $n$, but this only applies when $n$ is large enough and when $\\epsilon$ is a good estimate of the conditional entropy.\n\nThe authors also mention two issues regarding existing measures (MDL,MI)\n- They are insensitive to statistical complexity\n- They are insensitive to computational complexity\nwhile these are true for the theoretical measures, I don't see what is the benefit of the current measures in this context. This is since old and new measures are leaning on an approximate  computation of $L(A,i)$.\n\n\n*Summary of review\nI don't see any clear benefit of the suggested measures.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604047606110}, {"id": "AQCkrWrrtc8", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2491/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The submission addresses the problem of representation evaluation from the perspective of efficient learning of downstream predictors. Leveraging the introduced loss-data curve framework, the paper studies and demonstrates the limitations of the existing methods in terms of their implicit dependency on evaluation dataset size. Motivated by practicality and interpretability of the measures for choosing the best representations, the paper introduces two novel methods, $\\epsilon$ sample complexity ($\\epsilon$SC) and surplus description length (SDL), which are well-motivated and supported both theoretically and empirically. The paper also delivers efficient implementation.\n\nThe paper has excellent motivation and discussion about how existing methods are inconsistent with representation quality, computational complexity and are not robust when evaluation dataset size changes—thus deeming them inapt to answer practical questions. For example, how hard is it to learn a predictor on the given representation? How many samples with given the representation are needed to achieve a specified performance? Is it even possible to reach a specified performance given this representation?\nThe proposed methods are designed to handle these questions and are robust to the evaluation dataset size secured against precipitate decisions on which representation is best. The paper considers the best representation to be the one that allows the most efficient learning of a downstream predictor.\n\nAlthough I do not immediately see weak points, yet the following things should be taken into account when applying the proposed methods — the assumption of monotonically improving predictors, the dependency of the proposed and existing methods for representation evaluation on the probe design, and potential disagreement between $\\epsilon$SC and SDL (as seen in Figure 3 for $\\epsilon=0.5$).\nAlthough very plausible, the assumption might not hold when there is a sufficient shift between training and validation data. There might be situations when representations may overfit to training data and not reflect properties of validation set points, potentially causing problems in both proposed methods. While dependency on the predictors is addressed in the paper as a separate problem and a direction for future work, it would help if the authors discussed the cases when there is a disagreement between $\\epsilon$SC and SDL.\n\nOverall, this submission seems to be a solid contribution and should be accepted as it rigorously addresses a problem of representation evaluation, which is of broad interest to the ICLR community.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good discussion and motivation, well grounded methods, accept", "review": "The submission addresses the problem of representation evaluation from the perspective of efficient learning of downstream predictors. Leveraging the introduced loss-data curve framework, the paper studies and demonstrates the limitations of the existing methods in terms of their implicit dependency on evaluation dataset size. Motivated by practicality and interpretability of the measures for choosing the best representations, the paper introduces two novel methods, $\\epsilon$ sample complexity ($\\epsilon$SC) and surplus description length (SDL), which are well-motivated and supported both theoretically and empirically. The paper also delivers efficient implementation.\n\nThe paper has excellent motivation and discussion about how existing methods are inconsistent with representation quality, computational complexity and are not robust when evaluation dataset size changes—thus deeming them inapt to answer practical questions. For example, how hard is it to learn a predictor on the given representation? How many samples with given the representation are needed to achieve a specified performance? Is it even possible to reach a specified performance given this representation?\nThe proposed methods are designed to handle these questions and are robust to the evaluation dataset size secured against precipitate decisions on which representation is best. The paper considers the best representation to be the one that allows the most efficient learning of a downstream predictor.\n\nAlthough I do not immediately see weak points, yet the following things should be taken into account when applying the proposed methods — the assumption of monotonically improving predictors, the dependency of the proposed and existing methods for representation evaluation on the probe design, and potential disagreement between $\\epsilon$SC and SDL (as seen in Figure 3 for $\\epsilon=0.5$).\nAlthough very plausible, the assumption might not hold when there is a sufficient shift between training and validation data. There might be situations when representations may overfit to training data and not reflect properties of validation set points, potentially causing problems in both proposed methods. While dependency on the predictors is addressed in the paper as a separate problem and a direction for future work, it would help if the authors discussed the cases when there is a disagreement between $\\epsilon$SC and SDL.\n\nOverall, this submission seems to be a solid contribution and should be accepted as it rigorously addresses a problem of representation evaluation, which is of broad interest to the ICLR community.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603884144400}], "openreview_url": "https://openreview.net/forum?id=FOR2VqgJXb", "arxiv_id": "2009.07368", "paper_pdf": "papers/FOR2VqgJXb.pdf", "paper_pdf_sha256": "d19be31d60dc849c00836981f1d287e0048530d717db528739d0a9f8518ed9e4", "paper_pdf_bytes": 354076, "paper_pdf_source": "openreview", "code_url": "https://github.com/willwhitney/reprieve", "code_repository": "willwhitney/reprieve", "code_commit": "180a82092c973fca572fad35d4c21f0075764f4a", "code_archive": "repos/FOR2VqgJXb.zip", "code_archive_sha256": "78d929b83bd50dee5b37d5b728d4e52708e901e3f56579a24192ce47c6258298", "code_archive_bytes": 389534, "code_file_count": 16, "code_extensions": {".py": 15, ".ipynb": 1}, "github_disk_usage_kb": 1869, "github_languages": {"Python": 46532}, "github_archived": false, "github_pushed_at": "2021-11-21T15:30:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evaluating-representations-by-the-complexity"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vnlsxFbWSB", "year": 2026, "status": "rejected", "title": "Bench-CoE: A Framework for Collaboration of Experts from Benchmark", "authors": ["Yuanshuai Wang", "Xingjian Zhang", "Jinkun Zhao", "Siwei Wen", "Peilin Feng", "Shuhao Liao", "Lei Huang", "wenjun wu"], "authorids": ["~Yuanshuai_Wang1", "~Xingjian_Zhang2", "~Jinkun_Zhao1", "~Siwei_Wen5", "~Peilin_Feng1", "~Shuhao_Liao2", "~Lei_Huang1", "~wenjun_wu3"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) are key technologies that drive intelligent systems to handle multiple tasks. To meet the demands of various tasks, an increasing number of LLMs-driven experts with diverse capabilities have been developed, spreading from language to visual understanding and generalization, accompanied by corresponding benchmarks to evaluate their performance. This paper proposes the Bench-CoE framework, which enables Collaboration of Experts (CoE) by effectively leveraging benchmark evaluations to achieve optimal performance across various tasks. Bench-CoE consists of a set of specialized expert models, a router for assigning tasks to corresponding experts, and a benchmark dataset for training the router. Based on this framework, we first formulate Query-Level Bench-CoE that is an abstraction of existing CoE methods exploiting the benchmark dataset. We further propose Subject-Level Bench-CoE, a new method that effectively addresses the potential issues of Query-Level Bench-CoE in poor generalization and labeling costs during training the router. Experiments show that the Query-Level Bench-CoE excels in in-distribution tasks, while the Subject-Level Bench-CoE demonstrates stronger out-of-distribution generalization. Our proposed Bench-CoE achieves efficient expert collaboration with minimal training label costs, improving adaptability in multi-task and cross-domain scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "yjwaDYCwvd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16362/Reviewer_FHpQ"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes Bench-CoE, a framework for collaborating multiple expert LLMs by leveraging benchmark evaluation results for routing. The authors formalize 1) Query-Level Bench-CoE, which routes queries based on which expert performs best on each individual query, and 2) Subject-Level Bench-CoE, which classifies queries into subjects and routes to experts based on subject-level benchmark performance. Experiments on multimodal tasks and language tasks show that Query-Level excels on in-distribution data while Subject-Level achieves better cross-domain generalization.", "review_text": "This paper proposes Bench-CoE, a framework for collaborating multiple expert LLMs by leveraging benchmark evaluation results for routing. The authors formalize 1) Query-Level Bench-CoE, which routes queries based on which expert performs best on each individual query, and 2) Subject-Level Bench-CoE, which classifies queries into subjects and routes to experts based on subject-level benchmark performance. Experiments on multimodal tasks and language tasks show that Query-Level excels on in-distribution data while Subject-Level achieves better cross-domain generalization.", "strengths": "- The framework is well-defined and its difference from the existing methods are clearly stated.\n- The proposed methods are evaluated under both in-domain and out-of-domain scenarios which makes the evaluation comprehensive.\n- Experiments show performance improvement over the existing baselines.", "weaknesses": "There are existing works that propose to route experts according to the queries or topics which are not included in the paper as baselines. The proposed methods, therefore, IMO, are not very novel and bear limited impact to the research community.", "questions": "How do the proposed methods perform when compared to the \"LoRA Soups\" line of works, which seek to merge LoRA weights instead of routing inputs to them [1]?\n\n[1] Prabhakar, Akshara, et al. \"Lora soups: Merging loras for practical skill composition tasks.\" Proceedings of the 31st International Conference on Computational Linguistics: Industry Track. 2025.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Bench-CoE, a framework for collaborating multiple expert LLMs by leveraging benchmark evaluation results for routing. The authors formalize 1) Query-Level Bench-CoE, which routes queries based on which expert performs best on each individual query, and 2) Subject-Level Bench-CoE, which classifies queries into subjects and routes to experts based on subject-level benchmark performance. Experiments on multimodal tasks and language tasks show that Query-Level excels on in-distribution data while Subject-Level achieves better cross-domain generalization.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The framework is well-defined and its difference from the existing methods are clearly stated.\n- The proposed methods are evaluated under both in-domain and out-of-domain scenarios which makes the evaluation comprehensive.\n- Experiments show performance improvement over the existing baselines.", "weaknesses": "There are existing works that propose to route experts according to the queries or topics which are not included in the paper as baselines. The proposed methods, therefore, IMO, are not very novel and bear limited impact to the research community.", "questions": "How do the proposed methods perform when compared to the \"LoRA Soups\" line of works, which seek to merge LoRA weights instead of routing inputs to them [1]?\n\n[1] Prabhakar, Akshara, et al. \"Lora soups: Merging loras for practical skill composition tasks.\" Proceedings of the 31st International Conference on Computational Linguistics: Industry Track. 2025.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762966169311}, {"id": "A3VdyZgcSa", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16362/Reviewer_UVbT"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes Bench-CoE, a new framework that trains a router to select expert models for a specific query. The paper proposes 2 methods: (1) query based router which directly routes the query to an expert model (2) subject based which first maps the query to a subject, then finds the expert that excels most at this subject to answer the question. The router is trained based on models' results on benchmark. The paper then conducts experiments and shows that this framework performs better than single expert models.", "review_text": "This paper proposes Bench-CoE, a new framework that trains a router to select expert models for a specific query. The paper proposes 2 methods: (1) query based router which directly routes the query to an expert model (2) subject based which first maps the query to a subject, then finds the expert that excels most at this subject to answer the question. The router is trained based on models' results on benchmark. The paper then conducts experiments and shows that this framework performs better than single expert models.", "strengths": "- Proposes a new framework that trains a router to select an expert for a given query \n- Sees some performance gain over single experts.\n- The paper writing is generally clear and easy to follow", "weaknesses": "- Experiments are not solid enough and lack some essential baselines. The paper only compares their method of combining multiple experts with just a single model. It is not surprising that this would perform better than a single model. The paper should conduct more detailed analysis with simple baseline methods such as majority vote, or other router training methods / other expert selection methods. \n- The selected experts are outdated and not essentially the state-of-the-art models, it would be useful to incorporate the current state-of-the-art models / combine strong models with weak models and conduct how the performance would change.\n- In each experiment setting, how many expert models there should be and what the expert models should be are heuristically chosen. The paper lacks essential ablation studies on these design choices.\n- The design of the subject level router is not convincing enough. Especially in OOD cases because the performance is intrinsically bounded by the subject labels.", "questions": "- For out of distribution cases (for example, when the subject labels from the training set and the test set are completely different), would the subject level router still help? It would be useful to discuss this.\n- How does the method perform compared with other baseline methods as mentioned in the weakness section?\n- How does the expert number / expert choices affect performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Bench-CoE, a new framework that trains a router to select expert models for a specific query. The paper proposes 2 methods: (1) query based router which directly routes the query to an expert model (2) subject based which first maps the query to a subject, then finds the expert that excels most at this subject to answer the question. The router is trained based on models' results on benchmark. The paper then conducts experiments and shows that this framework performs better than single expert models.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- Proposes a new framework that trains a router to select an expert for a given query \n- Sees some performance gain over single experts.\n- The paper writing is generally clear and easy to follow", "weaknesses": "- Experiments are not solid enough and lack some essential baselines. The paper only compares their method of combining multiple experts with just a single model. It is not surprising that this would perform better than a single model. The paper should conduct more detailed analysis with simple baseline methods such as majority vote, or other router training methods / other expert selection methods. \n- The selected experts are outdated and not essentially the state-of-the-art models, it would be useful to incorporate the current state-of-the-art models / combine strong models with weak models and conduct how the performance would change.\n- In each experiment setting, how many expert models there should be and what the expert models should be are heuristically chosen. The paper lacks essential ablation studies on these design choices.\n- The design of the subject level router is not convincing enough. Especially in OOD cases because the performance is intrinsically bounded by the subject labels.", "questions": "- For out of distribution cases (for example, when the subject labels from the training set and the test set are completely different), would the subject level router still help? It would be useful to discuss this.\n- How does the method perform compared with other baseline methods as mentioned in the weakness section?\n- How does the expert number / expert choices affect performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761989559262}, {"id": "avn6wTBQAI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16362/Reviewer_yC4P"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This work follows the line of Collaboration of Experts and proposes 2 abstractions that route experts, Query-level and subject-level Bench-CoT and show interesting results that Query-level excels in in-distribution tasks and Subject-level works better in OOD tasks", "review_text": "This work follows the line of Collaboration of Experts and proposes 2 abstractions that route experts, Query-level and subject-level Bench-CoT and show interesting results that Query-level excels in in-distribution tasks and Subject-level works better in OOD tasks", "strengths": "1. The methodology seems to be novel, simple, and effective, and potentially efficient.\n2. The related work section is well-written and helps with the understanding of the scope of this work", "weaknesses": "1. There is usually a suite of benchmarks people in this domain test to showcase that other benchmark performances do not drop; in this work, they do MMStar, and there are more OOD/Cross-Bench datasets like MME, etc., as well. It would be more convincing to show that the performances using this method are on par or do not drop much.\n2. Usually, it is great to show a model of different sizes for ablations to show the Subject-level and query-level different advantages. Doesn't need many, but good to have.\n3. L160, not sure where WGM, LGM come from\n4. The citation format can use some care.", "questions": "1. Subject-level and query-level has different advantage, so what is the recommendation for practitioner", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work follows the line of Collaboration of Experts and proposes 2 abstractions that route experts, Query-level and subject-level Bench-CoT and show interesting results that Query-level excels in in-distribution tasks and Subject-level works better in OOD tasks", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The methodology seems to be novel, simple, and effective, and potentially efficient.\n2. The related work section is well-written and helps with the understanding of the scope of this work", "weaknesses": "1. There is usually a suite of benchmarks people in this domain test to showcase that other benchmark performances do not drop; in this work, they do MMStar, and there are more OOD/Cross-Bench datasets like MME, etc., as well. It would be more convincing to show that the performances using this method are on par or do not drop much.\n2. Usually, it is great to show a model of different sizes for ablations to show the Subject-level and query-level different advantages. Doesn't need many, but good to have.\n3. L160, not sure where WGM, LGM come from\n4. The citation format can use some care.", "questions": "1. Subject-level and query-level has different advantage, so what is the recommendation for practitioner", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761939006484}, {"id": "o4XWSTrjRd", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16362/Reviewer_jzbB"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper builds a framework to route queries to different LLM-based experts to achieve high performance and generalize across queries from new tasks. It proposes to use subject level meta-data of benchmark evaluations to train a router. Assuming the subject-expert mapping (i.e., best expert for each subject) is available through benchmark evaluation, the router is trained to predict the subject of the query, to then route the query to the corresponding best expert.", "review_text": "The paper builds a framework to route queries to different LLM-based experts to achieve high performance and generalize across queries from new tasks. It proposes to use subject level meta-data of benchmark evaluations to train a router. Assuming the subject-expert mapping (i.e., best expert for each subject) is available through benchmark evaluation, the router is trained to predict the subject of the query, to then route the query to the corresponding best expert.", "strengths": "- The problem is highly relevant with respect to efficiency, reuse, and collaborative development. \n- The paper is easy to follow \n- It provides ablations showing that subject level routing generalizes better compared to existing approaches that route at query level without utilizing subject level meta-data", "weaknesses": "- It assumes that benchmarks have clearly separated subject level meta-data in them, which might not always be true. Categorizing a benchmark into a set of distinct subjects/expertise is a challenging problem on its own.\n- There are cases where even though the subject remains the same, there might be different difficulty associated with them which won’t be captured if routing is learnt at subject level. For example, GSM8k vs AIME benchmark fall under math subject, where as there might multiple experts associated with this subject and they have different performances across the datasets in a given subject. \n- Naive evaluation doesn’t make sense. You can’t have the same training and test dataset. \n- For other evaluations, please provide non-zero shot baselines like best expert in the pool or best expert per query when evaluated with every expert to get a sense of benefit of the approach. For these baselines, see (https://arxiv.org/pdf/2402.05859)", "questions": "Please see weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper builds a framework to route queries to different LLM-based experts to achieve high performance and generalize across queries from new tasks. It proposes to use subject level meta-data of benchmark evaluations to train a router. Assuming the subject-expert mapping (i.e., best expert for each subject) is available through benchmark evaluation, the router is trained to predict the subject of the query, to then route the query to the corresponding best expert.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The problem is highly relevant with respect to efficiency, reuse, and collaborative development. \n- The paper is easy to follow \n- It provides ablations showing that subject level routing generalizes better compared to existing approaches that route at query level without utilizing subject level meta-data", "weaknesses": "- It assumes that benchmarks have clearly separated subject level meta-data in them, which might not always be true. Categorizing a benchmark into a set of distinct subjects/expertise is a challenging problem on its own.\n- There are cases where even though the subject remains the same, there might be different difficulty associated with them which won’t be captured if routing is learnt at subject level. For example, GSM8k vs AIME benchmark fall under math subject, where as there might multiple experts associated with this subject and they have different performances across the datasets in a given subject. \n- Naive evaluation doesn’t make sense. You can’t have the same training and test dataset. \n- For other evaluations, please provide non-zero shot baselines like best expert in the pool or best expert per query when evaluated with every expert to get a sense of benefit of the approach. For these baselines, see (https://arxiv.org/pdf/2402.05859)", "questions": "Please see weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761765552597}, {"id": "AuaYAetibj", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission16362/Reviewer_JVyp"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes the Bench-CoE framework to enable effective collaboration among LLMs by leveraging benchmark evaluation results. The authors validate Bench-CoE through experiments on multimodal (MMMU, MMStar) and NLP (MMLU-Pro, BigBench-Hard) tasks, demonstrating that Subject-Level Bench-CoE outperforms individual experts and Query-Level methods.", "review_text": "This paper proposes the Bench-CoE framework to enable effective collaboration among LLMs by leveraging benchmark evaluation results. The authors validate Bench-CoE through experiments on multimodal (MMMU, MMStar) and NLP (MMLU-Pro, BigBench-Hard) tasks, demonstrating that Subject-Level Bench-CoE outperforms individual experts and Query-Level methods.", "strengths": "- The core insight of using benchmark evaluations as \"free labels\" for router training effectively solves two key pain points of existing CoE methods.\n- The subject-expert mapping mechanism allows dynamic updates (e.g., integrating new experts or updating leaderboard rankings) without retraining the router", "weaknesses": "- While the paper contrasts Bench-CoE with Mixture of Experts (MoE) and traditional CoE methods, it overlooks recent works that also leverage benchmarks for model selection or routing.\n- The paper mentions using BERT and TinyLLaVA as classifiers but provides no details on training details.\n- The experiments are weak, it only compare with single methods.\n- The citation format is not correct.", "questions": "- What is the contribution of this work compared to recently proposed routing methods?\n- What if multiple experts perform equally well on a subject (e.g., two models with <1% accuracy difference on MMMU’s Math subject)? How does the framework handle ambiguity in mapping?\n- How to choose the subject effectively? What if I want to add new subject?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the Bench-CoE framework to enable effective collaboration among LLMs by leveraging benchmark evaluation results. The authors validate Bench-CoE through experiments on multimodal (MMMU, MMStar) and NLP (MMLU-Pro, BigBench-Hard) tasks, demonstrating that Subject-Level Bench-CoE outperforms individual experts and Query-Level methods.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The core insight of using benchmark evaluations as \"free labels\" for router training effectively solves two key pain points of existing CoE methods.\n- The subject-expert mapping mechanism allows dynamic updates (e.g., integrating new experts or updating leaderboard rankings) without retraining the router", "weaknesses": "- While the paper contrasts Bench-CoE with Mixture of Experts (MoE) and traditional CoE methods, it overlooks recent works that also leverage benchmarks for model selection or routing.\n- The paper mentions using BERT and TinyLLaVA as classifiers but provides no details on training details.\n- The experiments are weak, it only compare with single methods.\n- The citation format is not correct.", "questions": "- What is the contribution of this work compared to recently proposed routing methods?\n- What if multiple experts perform equally well on a subject (e.g., two models with <1% accuracy difference on MMMU’s Math subject)? How does the framework handle ambiguity in mapping?\n- How to choose the subject effectively? What if I want to add new subject?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761720112220}], "openreview_url": "https://openreview.net/forum?id=vnlsxFbWSB", "arxiv_id": "2412.04167", "paper_pdf": "papers/vnlsxFbWSB.pdf", "paper_pdf_sha256": "6f392a8c90df25218c53f6852fb37d0ced17942cf7d9b0f9f2c6b2b1887eb19b", "paper_pdf_bytes": 1047702, "paper_pdf_source": "openreview", "code_url": "https://github.com/ZhangXJ199/Bench-CoE", "code_repository": "ZhangXJ199/Bench-CoE", "code_commit": "fab309d6c397e92d60907ec345a8989043afa04c", "code_archive": "repos/vnlsxFbWSB.zip", "code_archive_sha256": "e28ddcbfbb200506ef93fa136458dc675b2b54a97e0d069b0fb483806dc6bab4", "code_archive_bytes": 7402808, "code_file_count": 127, "code_extensions": {".py": 125, ".sh": 2}, "github_disk_usage_kb": 7522, "github_languages": {"Python": 1060126, "Shell": 1497}, "github_archived": false, "github_pushed_at": "2025-04-27T05:44:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/bench-coe-a-framework-for-collaboration-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "j9wBgcxa7N", "year": 2025, "status": "rejected", "title": "MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning", "authors": ["Justin Chen", "Archiki Prasad", "Swarnadeep Saha", "Elias Stengel-Eskin", "Mohit Bansal"], "authorids": ["~Justin_Chen1", "~Archiki_Prasad1", "~Swarnadeep_Saha2", "~Elias_Stengel-Eskin1", "~Mohit_Bansal2"], "authors_source": "OpenReview API", "abstract": "Large Language Models' (LLM) reasoning can be improved using test-time aggregation strategies, i.e., generating multiple samples for each problem and aggregating over them to find a better answer. While these improve performance, they often reach a saturation point beyond which additional samples provide no return. Refinement offers an alternative by using model-generated feedback to improve answer quality. However, refinement introduces three key challenges: (1) Excessive refinement: Uniformly refining all instances can cause over-correction and reduce overall performance. (2) Inability to localize and address errors: LLMs have a limited ability to self-correct and struggle to identify and correct their own mistakes in a targeted way. (3) Insufficient refinement: Deciding how many iterations of refinement are needed is non-trivial, and stopping too soon could leave errors unaddressed. To tackle these issues, we propose MAgICoRe, a framework for Multi-Agent Iteration for Coarse-to-fine Refinement. MAgICoRe aims to avoid excessive refinement by categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving hard ones with fine-grained and iterative multi-agent refinement. To enable more granular error localization, we incorporate external step-wise reward model (RM) scores. To ensure effective refinement, we employ a multi-agent loop with three agents: the Solver, the Reviewer (which generates targeted feedback based on step-wise RM scores) and the Refiner (which incorporates feedback and generates new solutions). To ensure sufficient refinement, we re-evaluate updated solutions, iteratively initiating further rounds of multi-agent refinement. We evaluate MAgICoRe on Llama-3-8B and GPT-3.5 and show its effectiveness across five math reasoning datasets, with consistent gains for all datasets and models. Even one iteration of MAgICoRe beats Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% while using less than 50% of the samples. Unlike iterative refinement with baseline methods, MAgICoRe continues to improve with more iterations. Finally, our ablations highlight the importance of MAgICoRe's use of RMs and multi-agent communication.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "dEtrK8IxBg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11411/Reviewer_ywAv"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces MAGICORE, an inference framework for LLM reasoning that leverages a multi-agent approach with three agents: the Solver, the Reviewer, and the Refiner. The Solver generates multiple solutions, which are then scored by the Reviewer using a Process Reward Model. The Refiner subsequently refines solutions based on their quality and confidence scores. MAGICORE demonstrates superior performance over several aggregation methods.", "review_text": "This paper introduces MAGICORE, an inference framework for LLM reasoning that leverages a multi-agent approach with three agents: the Solver, the Reviewer, and the Refiner. The Solver generates multiple solutions, which are then scored by the Reviewer using a Process Reward Model. The Refiner subsequently refines solutions based on their quality and confidence scores. MAGICORE demonstrates superior performance over several aggregation methods.", "strengths": "1. The writing is clear and easy to follow.\n2. The proposed method is efficient and straightforward, achieving better performance than self-aggregation and self-refinement methods.", "weaknesses": "1. The paper lacks a comparison with methods that also use a PRM. While baselines are based on self-aggregation and self-refinement, it remains unclear how MAGICORE would perform compared to methods like self-consistency with PRM. This omission may make the comparison in Figure 1 less conclusive.\n\n2. The method assumes that LLMs can improve through fine-grained stepwise refinement. Adding preliminary experiments or a discussion section to validate this assumption could enhance the paper’s robustness.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces MAGICORE, an inference framework for LLM reasoning that leverages a multi-agent approach with three agents: the Solver, the Reviewer, and the Refiner. The Solver generates multiple solutions, which are then scored by the Reviewer using a Process Reward Model. The Refiner subsequently refines solutions based on their quality and confidence scores. MAGICORE demonstrates superior performance over several aggregation methods.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The writing is clear and easy to follow.\n2. The proposed method is efficient and straightforward, achieving better performance than self-aggregation and self-refinement methods.", "weaknesses": "1. The paper lacks a comparison with methods that also use a PRM. While baselines are based on self-aggregation and self-refinement, it remains unclear how MAGICORE would perform compared to methods like self-consistency with PRM. This omission may make the comparison in Figure 1 less conclusive.\n\n2. The method assumes that LLMs can improve through fine-grained stepwise refinement. Adding preliminary experiments or a discussion section to validate this assumption could enhance the paper’s robustness.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730724971663}, {"id": "zkvfaR3zc7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11411/Reviewer_k9ff"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "MAGICORE is a multi-agent framework designed to improve large language model (LLM) reasoning by selectively refining answers based on difficulty, leveraging external reward models (RMs) for targeted feedback and iterative refinement. This approach addresses key issues in traditional refinement methods, such as over-correction and insufficient error localization, achieving superior performance across five math reasoning datasets compared to baseline methods. Notably, MAGICORE shows continued improvement with more iterations, unlike other methods, highlighting the importance of selective and iterative refinement using RMs.", "review_text": "MAGICORE is a multi-agent framework designed to improve large language model (LLM) reasoning by selectively refining answers based on difficulty, leveraging external reward models (RMs) for targeted feedback and iterative refinement. This approach addresses key issues in traditional refinement methods, such as over-correction and insufficient error localization, achieving superior performance across five math reasoning datasets compared to baseline methods. Notably, MAGICORE shows continued improvement with more iterations, unlike other methods, highlighting the importance of selective and iterative refinement using RMs.", "strengths": "1. This paper proposes multi-agent framework, named MAGICORE, effectively categorizes problems into \"easy\" and \"hard\" cases, applying coarse-grained aggregation to simpler problems and fine-grained, multi-agent refinement to challenging ones. This selective refinement approach minimizes over-correction and ensures computational resources are allocated efficiently, resulting in higher performance without excessive sampling.\n\n2. By incorporating step-wise Reward Model (RM) scores, MAGICORE significantly improves error localization, allowing for more accurate, step-by-step feedback. This targeted feedback mechanism enables the framework to address specific mistakes precisely, enhancing the overall quality of model outputs.\n\n3. This paper conducts comprehensive experiments to verify the effectiveness of the proposed framework MAGICORE.", "weaknesses": "1.  MAGICORE relies heavily on external Reward Models to assess difficulty and provide targeted feedback, which could introduce dependency on the quality of these RMs. However, the better reward model is not easy to obtain. If the RMs are not well-tuned or suited to the specific dataset, the refinement process might misidentify errors, potentially impacting overall performance.\n\n2. While MAGICORE shows strong results on math reasoning tasks, it is unclear how well it would perform on other types of reasoning tasks, which may have different error patterns. More evaluations should be conducted across a broader range of tasks to verify the generalizability of this framework beyond math.\n\n3. Although MAGICORE improves with additional iterations, the optimal stopping point remains unclear, despite the authors conducting experiments with 1 to 5 iterations for validation.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "MAGICORE is a multi-agent framework designed to improve large language model (LLM) reasoning by selectively refining answers based on difficulty, leveraging external reward models (RMs) for targeted feedback and iterative refinement. This approach addresses key issues in traditional refinement methods, such as over-correction and insufficient error localization, achieving superior performance across five math reasoning datasets compared to baseline methods. Notably, MAGICORE shows continued improvement with more iterations, unlike other methods, highlighting the importance of selective and iterative refinement using RMs.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper proposes multi-agent framework, named MAGICORE, effectively categorizes problems into \"easy\" and \"hard\" cases, applying coarse-grained aggregation to simpler problems and fine-grained, multi-agent refinement to challenging ones. This selective refinement approach minimizes over-correction and ensures computational resources are allocated efficiently, resulting in higher performance without excessive sampling.\n\n2. By incorporating step-wise Reward Model (RM) scores, MAGICORE significantly improves error localization, allowing for more accurate, step-by-step feedback. This targeted feedback mechanism enables the framework to address specific mistakes precisely, enhancing the overall quality of model outputs.\n\n3. This paper conducts comprehensive experiments to verify the effectiveness of the proposed framework MAGICORE.", "weaknesses": "1.  MAGICORE relies heavily on external Reward Models to assess difficulty and provide targeted feedback, which could introduce dependency on the quality of these RMs. However, the better reward model is not easy to obtain. If the RMs are not well-tuned or suited to the specific dataset, the refinement process might misidentify errors, potentially impacting overall performance.\n\n2. While MAGICORE shows strong results on math reasoning tasks, it is unclear how well it would perform on other types of reasoning tasks, which may have different error patterns. More evaluations should be conducted across a broader range of tasks to verify the generalizability of this framework beyond math.\n\n3. Although MAGICORE improves with additional iterations, the optimal stopping point remains unclear, despite the authors conducting experiments with 1 to 5 iterations for validation.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730722679851}, {"id": "xWtZzY5mqZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11411/Reviewer_eYto"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces MAGICORE, a multi-agent, iterative coarse-to-fine refinement framework designed to improve reasoning performance in LLMs. The system uses three agents—Solver, Reviewer, and Refiner—who collaborate iteratively, guided by external RMs that provide both global and step-wise feedback. The framework is evaluated on five math reasoning datasets, showing consistent gains across models and datasets, outperforming existing self-refinement and aggregation-based baselines.", "review_text": "The paper introduces MAGICORE, a multi-agent, iterative coarse-to-fine refinement framework designed to improve reasoning performance in LLMs. The system uses three agents—Solver, Reviewer, and Refiner—who collaborate iteratively, guided by external RMs that provide both global and step-wise feedback. The framework is evaluated on five math reasoning datasets, showing consistent gains across models and datasets, outperforming existing self-refinement and aggregation-based baselines.", "strengths": "MAGICORE combines a multi-agent system with a coarse-to-fine approach, allowing efficient resource allocation that avoids the pitfalls of excessive or insufficient refinement.\n\nThe integration of external reward models (ORM and PRM) for both global and step-wise scoring enables precise error identification and correction.\n\nMAGICORE outperforms higher-budget baselines with fewer samples, making it a promising framework for cost-effective large-scale reasoning.", "weaknesses": "The three technical designs to address the three major challenges seem isolated and independent, making the overall framework not a “unified” solution.\n\nEach of the three technical designs has been fully utilized by existing research works, e.g., least-to-most prompting, multi-agent discussion, and iterative refinement. This paper does not propose any fundamentally new method, limiting its technical contributions.\n\nThe selected baselines for comparison are insufficient. Several self-correction, self-refinement, tree searching, or multi-agent reasoning methods are not included.\n\nThe experiments are only conducted on math reasoning datasets, leaving several other domains unexplored, e.g., program coding, commonsense reasoning, and symbolic reasoning. These are also essential domains in LLM reasoning.\n\nOnly two LLMs are considered. Several up-to-date and more powerful models, e.g., GPT-4 and llama3 70B/405B, should also be investigated.\n\nThe paper seems too dense to read. The presentation necessitates much improvement.", "questions": "Is the MAGICORE framework suitable for other types of reasoning? What modifications would be necessary for other domains? For example, in terms of condition evaluation and the refinement process?\n\nHow does MAGICORE handle misleading questions where the Solver consistently arrives at the same incorrect answer across multiple solutions? Would MAGICORE classify such cases as 'easy,' given that only one of the ORM or PRM needs to be 'fooled' during classification?\n\nThe refinement process relies on PRM-generated step-wise scores to guide targeted feedback. How sensitive is this process to the accuracy of PRM's step-wise evaluations? If PRM incorrectly assigns high scores to flawed steps, could this lead to ineffective or even harmful refinements? How does MAGICORE mitigate such risks in error localization?\n\nThe reviewer and refiner roles are distinct, with the Reviewer generating feedback based on step-wise scores. What specific mechanisms ensure that the Reviewer’s feedback is both relevant and actionable for the Refiner? Are there cases where ambiguous or overly general feedback could hinder the refinement process?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces MAGICORE, a multi-agent, iterative coarse-to-fine refinement framework designed to improve reasoning performance in LLMs. The system uses three agents—Solver, Reviewer, and Refiner—who collaborate iteratively, guided by external RMs that provide both global and step-wise feedback. The framework is evaluated on five math reasoning datasets, showing consistent gains across models and datasets, outperforming existing self-refinement and aggregation-based baselines.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "MAGICORE combines a multi-agent system with a coarse-to-fine approach, allowing efficient resource allocation that avoids the pitfalls of excessive or insufficient refinement.\n\nThe integration of external reward models (ORM and PRM) for both global and step-wise scoring enables precise error identification and correction.\n\nMAGICORE outperforms higher-budget baselines with fewer samples, making it a promising framework for cost-effective large-scale reasoning.", "weaknesses": "The three technical designs to address the three major challenges seem isolated and independent, making the overall framework not a “unified” solution.\n\nEach of the three technical designs has been fully utilized by existing research works, e.g., least-to-most prompting, multi-agent discussion, and iterative refinement. This paper does not propose any fundamentally new method, limiting its technical contributions.\n\nThe selected baselines for comparison are insufficient. Several self-correction, self-refinement, tree searching, or multi-agent reasoning methods are not included.\n\nThe experiments are only conducted on math reasoning datasets, leaving several other domains unexplored, e.g., program coding, commonsense reasoning, and symbolic reasoning. These are also essential domains in LLM reasoning.\n\nOnly two LLMs are considered. Several up-to-date and more powerful models, e.g., GPT-4 and llama3 70B/405B, should also be investigated.\n\nThe paper seems too dense to read. The presentation necessitates much improvement.", "questions": "Is the MAGICORE framework suitable for other types of reasoning? What modifications would be necessary for other domains? For example, in terms of condition evaluation and the refinement process?\n\nHow does MAGICORE handle misleading questions where the Solver consistently arrives at the same incorrect answer across multiple solutions? Would MAGICORE classify such cases as 'easy,' given that only one of the ORM or PRM needs to be 'fooled' during classification?\n\nThe refinement process relies on PRM-generated step-wise scores to guide targeted feedback. How sensitive is this process to the accuracy of PRM's step-wise evaluations? If PRM incorrectly assigns high scores to flawed steps, could this lead to ineffective or even harmful refinements? How does MAGICORE mitigate such risks in error localization?\n\nThe reviewer and refiner roles are distinct, with the Reviewer generating feedback based on step-wise scores. What specific mechanisms ensure that the Reviewer’s feedback is both relevant and actionable for the Refiner? Are there cases where ambiguous or overly general feedback could hinder the refinement process?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730642032598}, {"id": "Ec0c5yIcut", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11411/Reviewer_cViA"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper presents MAgICoRe (Multi-Agent, Iterative, Coarse-to-Fine Refinement), a framework designed to enhance the answer quality of LLMs. The framework addresses three major challenges in refining LLM outputs: avoiding excessive refinement (which can lead to over-correction), targeting error localization, and ensuring sufficient refinement. It proposes to use three agents (solver, reviewer, and refiner), and two reward models (PRM for local scores and ORM for global scores) to enhance the base model’s performance. The key idea behind the proposed framework is categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving hard ones with fine-grained and iterative multi-agent refinement. The authors evaluate MAgICoRe on 5 datasets and 2 models and show that it obtains accuracy higher than weighted self-consistency.", "review_text": "This paper presents MAgICoRe (Multi-Agent, Iterative, Coarse-to-Fine Refinement), a framework designed to enhance the answer quality of LLMs. The framework addresses three major challenges in refining LLM outputs: avoiding excessive refinement (which can lead to over-correction), targeting error localization, and ensuring sufficient refinement. It proposes to use three agents (solver, reviewer, and refiner), and two reward models (PRM for local scores and ORM for global scores) to enhance the base model’s performance. The key idea behind the proposed framework is categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving hard ones with fine-grained and iterative multi-agent refinement. The authors evaluate MAgICoRe on 5 datasets and 2 models and show that it obtains accuracy higher than weighted self-consistency.", "strengths": "The idea of incorporating RM models and selective refinement with LLMs is very interesting. \nThe results are very promising, and this method seems to be better than the baselines.\nThe authors open source their code.", "weaknesses": "Writing\nW1: The writing of Section 2 could be clearer. Some definitions of the conditions (in Appendix B) are very important to understand the method. Please move it up to the main context. This is especially true when you have a complex Figure 2. If there is no space, please move related works to the appendix.\n\nNovelty\nW2: While the reviewer appreciated the proposed framework, the classification of this method is not very clear. If we borrow the classification as used by [1], do the authors consider this work as an intrinsic or extrinsic refinement? If intrinsic, then you should not use separate RM models (instead you should finetune Llama3-8b to be your RM). If extrinsic, then this method seems less accurate than, say, RAG-assisted refinement.\n\nW3: The framework is a combination of multi-agent and RM for selective refinement.  The novelty of the entire system may be hard to justify, as individual parts of the framework do not seem very new. For example, multi-agent is a very well-studied topic now, and in terms of selective refinement, [2] uses implicit confidence of LLM to selectively refine the response. [3] is less similar, but it also uses confidence to select which LLM to use for prediction. In the context of LLM, confidence is similar to the explicit RM (although the reviewer agrees that RM is better than confidence). However, as the authors admit in the paper, the idea of using RM is also not novel (l153).\n\nEvaluation\nW4: SVAMP is a dataset with 700 training and 300 testing samples. The paper says that the authors have evaluated the 1000 samples of SVAMP (l304), and the results (for example 78.1) do seem like that. This is problematic though. Please report the result on the test set only. \n\nW5: Please specify the subtasks in MMLU as used. The reviewer failed to add up the numbers to 974 for the standard math-related questions (l308).\n\nW6: Please add GPT4 for a subset of the test datasets.\n\nW7: The cost of the proposed framework seems very high compared to decoding-only frameworks (majority voting or vanilla self-consistency). The method runs models multiple times, whereas decoding-only frameworks mainly just decode and do not run the main model. Can the authors please provide the inference delay or cost for the comparison?\n\nReference\n[1] Huang, Jie, et al. \"Large language models cannot self-correct reasoning yet.\" arXiv preprint arXiv:2310.01798 (2023).\n[2] Li, Loka, et al. \"Confidence matters: Revisiting intrinsic self-correction capabilities of large language models.\" arXiv preprint arXiv:2402.12563 (2024).\n[3] Nie, Lunyiu, et al. \"Online Cascade Learning for Efficient Inference over Streams.\" Forty-first International Conference on Machine Learning.", "questions": "Please see the weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents MAgICoRe (Multi-Agent, Iterative, Coarse-to-Fine Refinement), a framework designed to enhance the answer quality of LLMs. The framework addresses three major challenges in refining LLM outputs: avoiding excessive refinement (which can lead to over-correction), targeting error localization, and ensuring sufficient refinement. It proposes to use three agents (solver, reviewer, and refiner), and two reward models (PRM for local scores and ORM for global scores) to enhance the base model’s performance. The key idea behind the proposed framework is categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving hard ones with fine-grained and iterative multi-agent refinement. The authors evaluate MAgICoRe on 5 datasets and 2 models and show that it obtains accuracy higher than weighted self-consistency.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The idea of incorporating RM models and selective refinement with LLMs is very interesting. \nThe results are very promising, and this method seems to be better than the baselines.\nThe authors open source their code.", "weaknesses": "Writing\nW1: The writing of Section 2 could be clearer. Some definitions of the conditions (in Appendix B) are very important to understand the method. Please move it up to the main context. This is especially true when you have a complex Figure 2. If there is no space, please move related works to the appendix.\n\nNovelty\nW2: While the reviewer appreciated the proposed framework, the classification of this method is not very clear. If we borrow the classification as used by [1], do the authors consider this work as an intrinsic or extrinsic refinement? If intrinsic, then you should not use separate RM models (instead you should finetune Llama3-8b to be your RM). If extrinsic, then this method seems less accurate than, say, RAG-assisted refinement.\n\nW3: The framework is a combination of multi-agent and RM for selective refinement.  The novelty of the entire system may be hard to justify, as individual parts of the framework do not seem very new. For example, multi-agent is a very well-studied topic now, and in terms of selective refinement, [2] uses implicit confidence of LLM to selectively refine the response. [3] is less similar, but it also uses confidence to select which LLM to use for prediction. In the context of LLM, confidence is similar to the explicit RM (although the reviewer agrees that RM is better than confidence). However, as the authors admit in the paper, the idea of using RM is also not novel (l153).\n\nEvaluation\nW4: SVAMP is a dataset with 700 training and 300 testing samples. The paper says that the authors have evaluated the 1000 samples of SVAMP (l304), and the results (for example 78.1) do seem like that. This is problematic though. Please report the result on the test set only. \n\nW5: Please specify the subtasks in MMLU as used. The reviewer failed to add up the numbers to 974 for the standard math-related questions (l308).\n\nW6: Please add GPT4 for a subset of the test datasets.\n\nW7: The cost of the proposed framework seems very high compared to decoding-only frameworks (majority voting or vanilla self-consistency). The method runs models multiple times, whereas decoding-only frameworks mainly just decode and do not run the main model. Can the authors please provide the inference delay or cost for the comparison?\n\nReference\n[1] Huang, Jie, et al. \"Large language models cannot self-correct reasoning yet.\" arXiv preprint arXiv:2310.01798 (2023).\n[2] Li, Loka, et al. \"Confidence matters: Revisiting intrinsic self-correction capabilities of large language models.\" arXiv preprint arXiv:2402.12563 (2024).\n[3] Nie, Lunyiu, et al. \"Online Cascade Learning for Efficient Inference over Streams.\" Forty-first International Conference on Machine Learning.", "questions": "Please see the weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730601276593}, {"id": "GDdVHXoyhG", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11411/Reviewer_u3Va"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "The authors focus on achieving reliable problem-solving with LLMs through refinement. They identify three main issues in existing refinement methods: excessive refinement, inability to localize and address errors, and insufficient refinement. To address these challenges, they propose MAGICORE, an adaptive framework designed to enhance both performance and efficiency in multi-step reasoning with LLMs by intelligently applying test-time aggregation and targeted refinement. Experimental results show that MAGICORE outperforms leading alternative approaches by a significant margin.", "review_text": "The authors focus on achieving reliable problem-solving with LLMs through refinement. They identify three main issues in existing refinement methods: excessive refinement, inability to localize and address errors, and insufficient refinement. To address these challenges, they propose MAGICORE, an adaptive framework designed to enhance both performance and efficiency in multi-step reasoning with LLMs by intelligently applying test-time aggregation and targeted refinement. Experimental results show that MAGICORE outperforms leading alternative approaches by a significant margin.", "strengths": "The paper is easy to follow and includes figures explaining the intuitive concept behind the proposed framework.\n\nThe authors propose a framework aiming to simultaneously address all issues in the existing refinement approaches. In particular, the authors try to design a complex process to avoid or alleviate the problems related to each issue. \n\nThe experimental results presented in the paper demonstrate the state-of-the-art performance of the proposed framework.", "weaknesses": "1. The major issue of this paper is the lack of research focus and key insights. At first glance, the authors open a big question aiming to address all issues in the existing refinement approaches. However, this question is too big to hold for both the researcher and the reader because understanding and gaining insights into one issue to capture solid motivations may be challenging enough. For instance, in the introduction, even Figure 1 contains too much content to understand, and it is hard for me to figure out which major problem (based on your research experience) should be the most critical. Interestingly, in this submission, the authors emphasize the importance of identifying difficult problems and allocating more resources to them (i.e., 'over-correction'). However, in this paper, they present everything uniformly, making each part overly complex in both conceptual explanation and writing.\n\n2. The consequence of presenting everything together is that, to me, the paper's main contribution to the community becomes difficult to identify. It is difficult to discern the in-depth insights the authors gained from related work that motivated their approach design. Specifically, can the authors claim that the proposed framework addresses all three issues present in existing approaches? If not, which issue does the framework address more effectively, and which one remains less resolved? If the framework can address one issue well, why and what makes this framework achieve better performance in this issue? \n\n3. Besides, I may not fully capture the relation or connections between each module of this framework. By saying connection, I mean that why the next module is necessary after the previous one finishes. For example, after identifying hard problems based on the RM scores, why do we need to perform a complex iteration multi-agent process? Is it good enough for me to use an existing complex approach to address these hard problems? I believe this misunderstanding stems from the complexity of the framework and the authors' vague problem definition. In most moments during the paper reading, before I fully appreciate and explore a module in depth, the author has already started defining a new problem and proposing a more complex solution.\n\n4. More importantly, with so many equally important agents involved, hallucinations could become a significant issue: if any agent produces an error, it may impact other modules. Worse, based on my experience, the issues caused by hallucinations tend to accumulate over iterations rather than being resolved. However, the authors do not discuss this obvious core drawback of LLMs. For example, some existing work suggests that their framework relies on advanced LLMs and performs poorly with weaker LLMs. The authors of this paper fail to provide such a necessary discussion. \n\n5. Additionally, while reading the paper, I noticed that the authors provide only brief explanations without engaging in in-depth discussions before quickly moving on. As a result, I am left wondering what findings or insights, beyond these design elements, might better illuminate the corresponding issues. This is because engaging in in-depth, specific discussions rather than quickly presenting design elements should be a core focus within the community.\n\n\n6. In aiming to address all three issues by incorporating these modules into the framework, the current experiments are insufficient to demonstrate its effectiveness or provide a deeper understanding. For instance, we might ask about the 'cost' of this framework. By 'cost,' I refer to the number of interactions required and the number or length of prompts needed to solve the problem. Second, a more specific ablation study may be necessary to examine the three issues in existing approaches, identifying which issue is most severe and to what extent each part of the framework addresses it. Besides, if 1-2 issues are ignored, I can remove or simplify the corresponding modules from the framework. Then, how about the performance? Third, please include more state-of-the-art approaches in the baseline, as Self-Refine may be outdated given the recent advancements in the field (refinement).\n\n\n7. Last but not least, I believe this paper does not align with my preferences, as the authors seem to combine multiple novel ideas into a single framework, which can be represented as A + B + C... Due to this, in addition to lacking experience in addressing a major, well-defined problem step-by-step, I feel there are some overclaims in the paper.  --- all three issues mentioned in the paper can be addressed (again without in-depth insights).", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors focus on achieving reliable problem-solving with LLMs through refinement. They identify three main issues in existing refinement methods: excessive refinement, inability to localize and address errors, and insufficient refinement. To address these challenges, they propose MAGICORE, an adaptive framework designed to enhance both performance and efficiency in multi-step reasoning with LLMs by intelligently applying test-time aggregation and targeted refinement. Experimental results show that MAGICORE outperforms leading alternative approaches by a significant margin.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper is easy to follow and includes figures explaining the intuitive concept behind the proposed framework.\n\nThe authors propose a framework aiming to simultaneously address all issues in the existing refinement approaches. In particular, the authors try to design a complex process to avoid or alleviate the problems related to each issue. \n\nThe experimental results presented in the paper demonstrate the state-of-the-art performance of the proposed framework.", "weaknesses": "1. The major issue of this paper is the lack of research focus and key insights. At first glance, the authors open a big question aiming to address all issues in the existing refinement approaches. However, this question is too big to hold for both the researcher and the reader because understanding and gaining insights into one issue to capture solid motivations may be challenging enough. For instance, in the introduction, even Figure 1 contains too much content to understand, and it is hard for me to figure out which major problem (based on your research experience) should be the most critical. Interestingly, in this submission, the authors emphasize the importance of identifying difficult problems and allocating more resources to them (i.e., 'over-correction'). However, in this paper, they present everything uniformly, making each part overly complex in both conceptual explanation and writing.\n\n2. The consequence of presenting everything together is that, to me, the paper's main contribution to the community becomes difficult to identify. It is difficult to discern the in-depth insights the authors gained from related work that motivated their approach design. Specifically, can the authors claim that the proposed framework addresses all three issues present in existing approaches? If not, which issue does the framework address more effectively, and which one remains less resolved? If the framework can address one issue well, why and what makes this framework achieve better performance in this issue? \n\n3. Besides, I may not fully capture the relation or connections between each module of this framework. By saying connection, I mean that why the next module is necessary after the previous one finishes. For example, after identifying hard problems based on the RM scores, why do we need to perform a complex iteration multi-agent process? Is it good enough for me to use an existing complex approach to address these hard problems? I believe this misunderstanding stems from the complexity of the framework and the authors' vague problem definition. In most moments during the paper reading, before I fully appreciate and explore a module in depth, the author has already started defining a new problem and proposing a more complex solution.\n\n4. More importantly, with so many equally important agents involved, hallucinations could become a significant issue: if any agent produces an error, it may impact other modules. Worse, based on my experience, the issues caused by hallucinations tend to accumulate over iterations rather than being resolved. However, the authors do not discuss this obvious core drawback of LLMs. For example, some existing work suggests that their framework relies on advanced LLMs and performs poorly with weaker LLMs. The authors of this paper fail to provide such a necessary discussion. \n\n5. Additionally, while reading the paper, I noticed that the authors provide only brief explanations without engaging in in-depth discussions before quickly moving on. As a result, I am left wondering what findings or insights, beyond these design elements, might better illuminate the corresponding issues. This is because engaging in in-depth, specific discussions rather than quickly presenting design elements should be a core focus within the community.\n\n\n6. In aiming to address all three issues by incorporating these modules into the framework, the current experiments are insufficient to demonstrate its effectiveness or provide a deeper understanding. For instance, we might ask about the 'cost' of this framework. By 'cost,' I refer to the number of interactions required and the number or length of prompts needed to solve the problem. Second, a more specific ablation study may be necessary to examine the three issues in existing approaches, identifying which issue is most severe and to what extent each part of the framework addresses it. Besides, if 1-2 issues are ignored, I can remove or simplify the corresponding modules from the framework. Then, how about the performance? Third, please include more state-of-the-art approaches in the baseline, as Self-Refine may be outdated given the recent advancements in the field (refinement).\n\n\n7. Last but not least, I believe this paper does not align with my preferences, as the authors seem to combine multiple novel ideas into a single framework, which can be represented as A + B + C... Due to this, in addition to lacking experience in addressing a major, well-defined problem step-by-step, I feel there are some overclaims in the paper.  --- all three issues mentioned in the paper can be addressed (again without in-depth insights).", "questions": "See weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730513214664}], "openreview_url": "https://openreview.net/forum?id=j9wBgcxa7N", "arxiv_id": "2409.12147", "paper_pdf": "papers/j9wBgcxa7N.pdf", "paper_pdf_sha256": "9845342614e4bfeb7d8cd67c48a4fea6f5b086f81f5505a26cf94c973efb14bf", "paper_pdf_bytes": 929378, "paper_pdf_source": "openreview", "code_url": "https://github.com/dinobby/MAgICoRE", "code_repository": "dinobby/MAgICoRE", "code_commit": "1388006b309663da69b142102d819d3be4d34890", "code_archive": "repos/j9wBgcxa7N.zip", "code_archive_sha256": "de87d379bd0bb1ca31f895558bf5c3450f82737d25a5de7242f137690d3b266e", "code_archive_bytes": 306128, "code_file_count": 8, "code_extensions": {".py": 8}, "github_disk_usage_kb": 365, "github_languages": {"Python": 48065}, "github_archived": false, "github_pushed_at": "2024-09-19T06:17:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/magicore-multi-agent-iterative-coarse-to-fine"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0aEUd9UtiA", "year": 2024, "status": "rejected", "title": "DiffCPS: Diffusion Model based Constrained Policy Search for Offline Reinforcement Learning", "authors": ["Longxiang He", "Linrui Zhang", "Junbo Tan", "Xueqian Wang"], "authorids": ["~Longxiang_He2", "~Linrui_Zhang1", "~Junbo_Tan1", "~Xueqian_Wang1"], "authors_source": "OpenReview API", "abstract": "Constrained policy search (CPS) is a fundamental problem in offline reinforcement learning, which is generally solved by advantage weighted regression (AWR). However, previous methods may still encounter out-of-distribution actions due to the limited expressivity of Gaussian-based policies.  On the other hand, directly applying the state-of-the-art models with distribution expression capabilities (i.e., diffusion models) in the AWR framework is insufficient since AWR requires exact policy probability densities, which is intractable in diffusion models. In this paper, we propose a novel approach called $\\textbf{Diffusion Model based Constrained Policy Search (DiffCPS)}$, which tackles the diffusion-based constrained policy search without resorting to AWR.  The theoretical analysis reveals our key insights by leveraging the action distribution of the diffusion model to eliminate the policy distribution constraint in the CPS and then utilizing the Evidence Lower Bound (ELBO) of diffusion-based policy to approximate the KL constraint. Consequently, DiffCPS admits the high expressivity of diffusion models while circumventing the cumbersome density calculation brought by AWR. Extensive experimental results based on the D4RL benchmark demonstrate the efficacy of our approach. We empirically show that DiffCPS achieves better or at least competitive performance compared to traditional AWR-based baselines as well as recent diffusion-based offline RL methods. Code will be made publicly available upon acceptance.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "CrIclP2Knk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1330/Reviewer_UbgT"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper studies a constrained policy search in offline reinforcement learning. To increase the expressivity of Gaussian-based policies, the authors propose to use diffusion model to represent policy. The authors formulate a diffusion model based constrained policy optimization problem, and propose a constrained policy search algorithm. The authors also provide experiments to show the performance of this method.", "review_text": "The paper studies a constrained policy search in offline reinforcement learning. To increase the expressivity of Gaussian-based policies, the authors propose to use diffusion model to represent policy. The authors formulate a diffusion model based constrained policy optimization problem, and propose a constrained policy search algorithm. The authors also provide experiments to show the performance of this method.", "strengths": "- The paper is well organized, and the key idea is delivered. \n\n- The authors provide an example to show the expressivity limitation in standard advantage regression methods, which justifies the necessity of introducing diffusion models. \n\n- The authors use a popular diffusion model: DDPM to represent policy, and present a new constrained policy search method, which is intuitively simple and easy to implement. \n\n- Experimental results demonstrate comprable performance compared with state-of-the-art methods.", "weaknesses": "- The importance or motivation of propositions, theorems, and corollaries is not well explained. Most results can be more directly obtained using simple calculations that are known in diffusion model. \n\n- The use of diffusion model as policy in constrained policy search has been studied in offline RL. Due to the similarity, it is important to distinguish them in an explicit way. \n\n- The authors claim strong duality for Equation (12) according to the duality in the convex optimization. However, the constrained policy search is a non-convex problem. It is not justified if the strong duality still holds.  \n\n- The main result is empirical. It is useful if the authors could provide performance analyses, which can strengthen the method with solid theoretical guarantees.", "questions": "- A large paragraph of this paper introduces known results. Can the authors highlight more new developments compared with existing methods?\n\n- It is not clear to me the strong duality of Equation (12). Can the authors justify it? \n\n- How does the problem (18)-(19) can solve the original problem (12)? \n\n- Are there other diffusion models useful for constrained policy search? It is useful if the authors could discuss the generalization of this approach. \n\n- Training diffusion model can be inefficient. What are computational times for the methods in experiments?\n\n- The examples in experiments are created in simulated environments. Any realistic offline dataset you can use to show performance of your algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper studies a constrained policy search in offline reinforcement learning. To increase the expressivity of Gaussian-based policies, the authors propose to use diffusion model to represent policy. The authors formulate a diffusion model based constrained policy optimization problem, and propose a constrained policy search algorithm. The authors also provide experiments to show the performance of this method.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The paper is well organized, and the key idea is delivered. \n\n- The authors provide an example to show the expressivity limitation in standard advantage regression methods, which justifies the necessity of introducing diffusion models. \n\n- The authors use a popular diffusion model: DDPM to represent policy, and present a new constrained policy search method, which is intuitively simple and easy to implement. \n\n- Experimental results demonstrate comprable performance compared with state-of-the-art methods.", "weaknesses": "- The importance or motivation of propositions, theorems, and corollaries is not well explained. Most results can be more directly obtained using simple calculations that are known in diffusion model. \n\n- The use of diffusion model as policy in constrained policy search has been studied in offline RL. Due to the similarity, it is important to distinguish them in an explicit way. \n\n- The authors claim strong duality for Equation (12) according to the duality in the convex optimization. However, the constrained policy search is a non-convex problem. It is not justified if the strong duality still holds.  \n\n- The main result is empirical. It is useful if the authors could provide performance analyses, which can strengthen the method with solid theoretical guarantees.", "questions": "- A large paragraph of this paper introduces known results. Can the authors highlight more new developments compared with existing methods?\n\n- It is not clear to me the strong duality of Equation (12). Can the authors justify it? \n\n- How does the problem (18)-(19) can solve the original problem (12)? \n\n- Are there other diffusion models useful for constrained policy search? It is useful if the authors could discuss the generalization of this approach. \n\n- Training diffusion model can be inefficient. What are computational times for the methods in experiments?\n\n- The examples in experiments are created in simulated environments. Any realistic offline dataset you can use to show performance of your algorithm?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698990785326}, {"id": "xdiBNtB5EM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1330/Reviewer_hHF8"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors tackle a KL-constrained offline RL problem, where a RL policy is trained but additional constraints are put on the distribution shift between the original policy and the trained policy. The authors point out the weakness in using a unimodal Gaussian policy and propose to use a diffusion process as a policy parametrization. The authors show some desirable properties of using diffusion as a policy parametrization, and propose a primal-dual iteration algorithm to solve KL-constrained offline RL. The authors compare against baselines in D4RL and present ablation studies of the algorithm.", "review_text": "The authors tackle a KL-constrained offline RL problem, where a RL policy is trained but additional constraints are put on the distribution shift between the original policy and the trained policy. The authors point out the weakness in using a unimodal Gaussian policy and propose to use a diffusion process as a policy parametrization. The authors show some desirable properties of using diffusion as a policy parametrization, and propose a primal-dual iteration algorithm to solve KL-constrained offline RL. The authors compare against baselines in D4RL and present ablation studies of the algorithm.", "strengths": "1. Overall, this is a good paper with strong theoretical and empirical results. \n2. The visualization for the problem of having unimodal policies is very intuitive, and motivates a richer class of policies that can model more multimodal distributions in offline RL. \n3. The ablation studies of the algorithm is extensive and the authors have done good research into best practices in training diffusion models.", "weaknesses": "1. The proposed baselines are entirely numerical and quantitative; while convincing, it would have been nice to see some qualitative behavior of DiffCPS compared to other baselines as well in order to strengthen the authors' claims. For instance, trajectory stitching is a popular example in offline RL where unimodal policies might possibly fail if done naively. In simple offline RL tasks such as the X data collection task in Diffuser [1], does DiffCPS succeed?\n\n2. One interpretation of the authors' work is that when we use a generative model of the training data from the behavior policy, this has the effect of automatically constraining the distribution shift between the learned policy and the behavior policy. The authors are missing some relevant work in this direction in the context of model-based offline RL. For instance, [2] motivates a very similar objective with DiffCPS, where the cross-entropy term is penalized rather than constrained (with the difference that the distribution shift is on the state-action occupation measures rather than the action distribution in the model-based setting).\n\n3. On a related note, offline RL also has model-free and model-based approaches, and the author's approach is model-free. The title might give the impression that this method is model-based, though I am aware that the authors' intention was to say it's based on a diffusion generative model. Maybe Diffusion based might be a better title?\n\n[1] Janner et al., \"Planning with Diffusion for Flexible Behavior Synthesis\", ICML 2022\n\n[2] Suh et al., \"Fighting Uncertainty with Gradients: Offline Reinforcement Learning with Diffusion Score Matching\", CoRL 2023", "questions": "1. Have the authors considered using Augmented Lagrangian?\n2. In Theorem 3.1, what is $d_{\\pi_b(s)}$? Should this be the occupation measure of the behavior policy, which the authors previously denote using $\\rho$? or the initial distribution of the behavior policy?\n3. How specific are the authors' claims to diffusion models? For instance, if we had trained a denoising autoencoder model (assuming they can be trained well), would the theorems still hold?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors tackle a KL-constrained offline RL problem, where a RL policy is trained but additional constraints are put on the distribution shift between the original policy and the trained policy. The authors point out the weakness in using a unimodal Gaussian policy and propose to use a diffusion process as a policy parametrization. The authors show some desirable properties of using diffusion as a policy parametrization, and propose a primal-dual iteration algorithm to solve KL-constrained offline RL. The authors compare against baselines in D4RL and present ablation studies of the algorithm.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. Overall, this is a good paper with strong theoretical and empirical results. \n2. The visualization for the problem of having unimodal policies is very intuitive, and motivates a richer class of policies that can model more multimodal distributions in offline RL. \n3. The ablation studies of the algorithm is extensive and the authors have done good research into best practices in training diffusion models.", "weaknesses": "1. The proposed baselines are entirely numerical and quantitative; while convincing, it would have been nice to see some qualitative behavior of DiffCPS compared to other baselines as well in order to strengthen the authors' claims. For instance, trajectory stitching is a popular example in offline RL where unimodal policies might possibly fail if done naively. In simple offline RL tasks such as the X data collection task in Diffuser [1], does DiffCPS succeed?\n\n2. One interpretation of the authors' work is that when we use a generative model of the training data from the behavior policy, this has the effect of automatically constraining the distribution shift between the learned policy and the behavior policy. The authors are missing some relevant work in this direction in the context of model-based offline RL. For instance, [2] motivates a very similar objective with DiffCPS, where the cross-entropy term is penalized rather than constrained (with the difference that the distribution shift is on the state-action occupation measures rather than the action distribution in the model-based setting).\n\n3. On a related note, offline RL also has model-free and model-based approaches, and the author's approach is model-free. The title might give the impression that this method is model-based, though I am aware that the authors' intention was to say it's based on a diffusion generative model. Maybe Diffusion based might be a better title?\n\n[1] Janner et al., \"Planning with Diffusion for Flexible Behavior Synthesis\", ICML 2022\n\n[2] Suh et al., \"Fighting Uncertainty with Gradients: Offline Reinforcement Learning with Diffusion Score Matching\", CoRL 2023", "questions": "1. Have the authors considered using Augmented Lagrangian?\n2. In Theorem 3.1, what is $d_{\\pi_b(s)}$? Should this be the occupation measure of the behavior policy, which the authors previously denote using $\\rho$? or the initial distribution of the behavior policy?\n3. How specific are the authors' claims to diffusion models? For instance, if we had trained a denoising autoencoder model (assuming they can be trained well), would the theorems still hold?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698638100011}, {"id": "0RCgOz2UEc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1330/Reviewer_4x5N"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors propose a new algorithm called DiffCPS for solving offline RL problems. They claim that DiffCPS can learn a diffusion-based policy avoiding the difficult density calculation brought by traditional AWR framework. They present theoretical justification for their approach and perform an empirical study to validate their proposed algorithms.", "review_text": "The authors propose a new algorithm called DiffCPS for solving offline RL problems. They claim that DiffCPS can learn a diffusion-based policy avoiding the difficult density calculation brought by traditional AWR framework. They present theoretical justification for their approach and perform an empirical study to validate their proposed algorithms.", "strengths": "* The numerical experiments are well designed, with a detailed section of ablation study. Also, the toy example is intuitive and interesting.", "weaknesses": "* Corollary 3.1.1 and Theorem 3.2 are flawed. In the proof, the authors claim $J(\\mu)$ is an affine function of $\\mu$, which is **false**. In fact, $J(\\mu)$ is **not** convex w.r.t. $\\mu$ and the duality results can not hold. Consequently, the theoretical analysis presented in this paper is compromised, significantly undermining its contributions.\n\n* The paper lacks clarity. Multiple mathematical objects are introduced without clear definitions. For example, $Q_\\phi(s,a)$, $\\epsilon_\\theta(x_i,i)$, $f_\\phi(y\\mid x_i)$, $H(\\cdot,\\cdot)$, etc. . If the reader is not familiar with the literature on RL and diffusion models, she/he will definitely get confused by the undefined notations. Also, I think the authors should provide more explanations about the exact methods used to update $\\mu$ and $\\lambda$. It would be appreciated if the authors could give a complete and precise description of DiffCPSS. (If there are problems with page limits the authors can put it in the appendix.)\n\n* The empirical performance of the proposed algorithm exhibits only marginal improvement when compared to the baselines.", "questions": "* The authors deploy diffusion-based policies in place of the traditional Gaussian policies to handle the multi-modal problem. I wonder whether there are choices other than diffusion models and naive Gaussians. For example, flow models can express a rich distribution class and have explicit densities. Is it possible to fit the flow models into the framework of AWR? \n\n* In the toy example, the authors claim that \"SfBC incorrectly models the dataset as a circle instead of a noisy circle\". I checked the image and did not find such a difference. Can the authors give an explanation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a new algorithm called DiffCPS for solving offline RL problems. They claim that DiffCPS can learn a diffusion-based policy avoiding the difficult density calculation brought by traditional AWR framework. They present theoretical justification for their approach and perform an empirical study to validate their proposed algorithms.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "* The numerical experiments are well designed, with a detailed section of ablation study. Also, the toy example is intuitive and interesting.", "weaknesses": "* Corollary 3.1.1 and Theorem 3.2 are flawed. In the proof, the authors claim $J(\\mu)$ is an affine function of $\\mu$, which is **false**. In fact, $J(\\mu)$ is **not** convex w.r.t. $\\mu$ and the duality results can not hold. Consequently, the theoretical analysis presented in this paper is compromised, significantly undermining its contributions.\n\n* The paper lacks clarity. Multiple mathematical objects are introduced without clear definitions. For example, $Q_\\phi(s,a)$, $\\epsilon_\\theta(x_i,i)$, $f_\\phi(y\\mid x_i)$, $H(\\cdot,\\cdot)$, etc. . If the reader is not familiar with the literature on RL and diffusion models, she/he will definitely get confused by the undefined notations. Also, I think the authors should provide more explanations about the exact methods used to update $\\mu$ and $\\lambda$. It would be appreciated if the authors could give a complete and precise description of DiffCPSS. (If there are problems with page limits the authors can put it in the appendix.)\n\n* The empirical performance of the proposed algorithm exhibits only marginal improvement when compared to the baselines.", "questions": "* The authors deploy diffusion-based policies in place of the traditional Gaussian policies to handle the multi-modal problem. I wonder whether there are choices other than diffusion models and naive Gaussians. For example, flow models can express a rich distribution class and have explicit densities. Is it possible to fit the flow models into the framework of AWR? \n\n* In the toy example, the authors claim that \"SfBC incorrectly models the dataset as a circle instead of a noisy circle\". I checked the image and did not find such a difference. Can the authors give an explanation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1697549814468}], "openreview_url": "https://openreview.net/forum?id=0aEUd9UtiA", "arxiv_id": "2310.05333", "paper_pdf": "papers/0aEUd9UtiA.pdf", "paper_pdf_sha256": "22c566f0eb1b3e96ce15bf9a5f87e9c6540f3de10d854215ef528987d829f16a", "paper_pdf_bytes": 2207264, "paper_pdf_source": "openreview", "code_url": "https://github.com/felix-thu/DiffCPS", "code_repository": "felix-thu/DiffCPS", "code_commit": "2d8de2878bdc2bb2f59bbb9f7ce929897cf7cb15", "code_archive": "repos/0aEUd9UtiA.zip", "code_archive_sha256": "3f7a18f629a0c2ad024c463a3080b848785bffd3208d275cdfd4a5f3d19db200", "code_archive_bytes": 332031, "code_file_count": 11, "code_extensions": {".py": 11}, "github_disk_usage_kb": 329, "github_languages": {"Python": 89657}, "github_archived": false, "github_pushed_at": "2024-09-09T16:28:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/diffcps-diffusion-model-based-constrained"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YjKqWExiy6s", "year": 2023, "status": "rejected", "title": "Eliminating Catastrophic Overfitting Via Abnormal Adversarial Examples Regularization", "authors": ["Runqi Lin", "Chaojian Yu", "Tongliang Liu"], "authorids": ["~Runqi_Lin1", "~Chaojian_Yu1", "~Tongliang_Liu1"], "authors_source": "OpenReview API", "abstract": "Single-step adversarial training (SSAT) is shown to be able to defend against iterative-step adversarial attacks to achieve both efficiency and robustness. However, SSAT suffers from catastrophic overfitting (CO) with strong adversaries, showing that the classifier decision boundaries are highly distorted and robust accuracy against iterative-step adversarial attacks suddenly drops from peak to nearly 0% in a few epochs. In this work, we find that some adversarial examples generated on the network trained by SSAT exhibit anomalous behaviour, that is, although the training data is generated by the inner maximization process, the loss of some adversarial examples decreases instead, which we called abnormal adversarial examples. Furthermore, network optimization on these abnormal adversarial examples will further accelerate the model decision boundaries distortion, and correspondingly, the number of abnormal adversarial examples will sharply increase with CO. These observations motivate us to prevent CO by hindering the generation of abnormal adversarial examples. Specifically, we design a novel method, Abnormal Adversarial Examples Regularization (AAER), which explicitly regularizes the number and logits variation of abnormal adversarial examples to hinder the model from generating abnormal adversarial examples. Extensive experiments demonstrate that our method can prevent CO and further boost adversarial robustness with strong adversaries.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "f-9ZUSYt24", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper656/Reviewer_DBF4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new training objective for single-step adversarial training to reduce catastrophic overfitting (CO). In particular, the paper establishes a correlation between CO and the presence of \"abnormal adversarial examples,\" i.e., instances where taking a single step of FGSM actually decreases the loss rather than increasing it. To this end, the authors propose a regularizer called AAER meant to penalize abnormal adversarial examples, and show that it often improves the accuracy of SSAT.", "review_text": "In summary, the work proposes a new training objective for regularizing single-step adversarial training. Some of the motivation is a bit unclear and there is a natural baseline that the work does not explore. Overall, I think the paper as is is slightly below the bar for publication, but I am willing to raise my score.", "strengths": "Strengths:\n\n- The observation that CO co-occurs with abnormal adversarial examples is both novel and interesting.\n- The attack seems efficient (e.g., not much slower than N-FGSM) while providing better adversarial accuracy.\n- The authors conduct an extensive evaluation of other methods and baselines.\n\nWeaknesses:\n\n- Given the link between CO and abnormal adversarial examples, a natural baseline would be to simply ignore any abnormal examples in the loss and train only on examples where the loss actually increased. Does this not work? If so, why?\n- The method does seem to improve on N-FGSM in terms of robust accuracy, but this often comes at the expense of clean accuracy, so it's unclear if the proposed method truly dominates N-FGSM. There is also some misleading bolding in the Table (i.e., for CIFAR-100 16/255 it doesn't look like the confidence intervals of the two bolded entries actually overlap) that should be fixed.\n- I am confused about the motivation behind Eq. 7---it seems like it is encouraging the network to classify the original and \"normal adversarial\" examples very differently (judging by the negative term)? Please correct me if I am wrong here, but it seems very counterintuitive.\n- Finally, there are many non-standard additions to the training routine (I'm thinking mainly of the linearly increasing $\\alpha$, as the other terms were ablated quite well). Does normal FGSM with linearly increasing alpha still suffer from CO?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new training objective for single-step adversarial training to reduce catastrophic overfitting (CO). In particular, the paper establishes a correlation between CO and the presence of \"abnormal adversarial examples,\" i.e., instances where taking a single step of FGSM actually decreases the loss rather than increasing it. To this end, the authors propose a regularizer called AAER meant to penalize abnormal adversarial examples, and show that it often improves the accuracy of SSAT.", "strength_and_weaknesses": "Strengths:\n\n- The observation that CO co-occurs with abnormal adversarial examples is both novel and interesting.\n- The attack seems efficient (e.g., not much slower than N-FGSM) while providing better adversarial accuracy.\n- The authors conduct an extensive evaluation of other methods and baselines.\n\nWeaknesses:\n\n- Given the link between CO and abnormal adversarial examples, a natural baseline would be to simply ignore any abnormal examples in the loss and train only on examples where the loss actually increased. Does this not work? If so, why?\n- The method does seem to improve on N-FGSM in terms of robust accuracy, but this often comes at the expense of clean accuracy, so it's unclear if the proposed method truly dominates N-FGSM. There is also some misleading bolding in the Table (i.e., for CIFAR-100 16/255 it doesn't look like the confidence intervals of the two bolded entries actually overlap) that should be fixed.\n- I am confused about the motivation behind Eq. 7---it seems like it is encouraging the network to classify the original and \"normal adversarial\" examples very differently (judging by the negative term)? Please correct me if I am wrong here, but it seems very counterintuitive.\n- Finally, there are many non-standard additions to the training routine (I'm thinking mainly of the linearly increasing $\\alpha$, as the other terms were ablated quite well). Does normal FGSM with linearly increasing alpha still suffer from CO?", "clarity,_quality,_novelty_and_reproducibility": "The work is certainly novel and original. The quality is reasonable, bar the weaknesses above. The clarity could be improved---there are some grammatical errors throughout, and there are some unclear sentences, e.g.:\n\n- Furthermore, the strength of AAER depends on the interaction of two parts learning objective, which can reflect the degree of classifier distortion more comprehensively and flexibly.\n- Based on the above analysis, we design a novel regularization term, AAER, which aims to suppress the anomalous training samples by the (i) number and the (ii) logits variation, ultimately achieving the purpose of preventing CO, which is shown in the following formula\n...\n\n", "summary_of_the_review": "In summary, the work proposes a new training objective for regularizing single-step adversarial training. Some of the motivation is a bit unclear and there is a natural baseline that the work does not explore. Overall, I think the paper as is is slightly below the bar for publication, but I am willing to raise my score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666668882397}, {"id": "A0oR-A5VZ_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper656/Reviewer_qRtw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThis paper studies the robust overfitting of one-step adversarial training and finds that robust overfitting is related to the increase of abnormal adversarial examples generated for training. Abnormal adversarial examples have even lower losses after adding the adversarial perturbations, which can be a reason for causing robust overfitting in one-step adversarial training. Based on this observation, this paper proposes a novel regularization term to reduce the number of abnormal adversarial examples. By comparing the proposed methods with other baselines on CIFAR10 and CIFAR100 datasets, the evaluation results show that the proposed method achieves better robustness against the PGD attack.\n", "review_text": "There are several points/places lacking clarity in the paper  (as discussed in the weakness). Further clarification and details may be necessary.\n\nReproducibility: The experiment settings section does not include all experiment details and no code is provided.\n", "strengths": "Strength: \n+ This paper points out the existence of abnormal adversarial examples when robust overfitting happens and speculates that robust overfitting in single-step adversarial training is caused by abnormal adversarial examples. \n\n+ This paper proposed a novel regularization term to reduce the number of abnormal adversarial examples in training.\n\n+ The evaluation results show that the proposed method outperforms other baseline methods.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\nThis paper studies the robust overfitting of one-step adversarial training and finds that robust overfitting is related to the increase of abnormal adversarial examples generated for training. Abnormal adversarial examples have even lower losses after adding the adversarial perturbations, which can be a reason for causing robust overfitting in one-step adversarial training. Based on this observation, this paper proposes a novel regularization term to reduce the number of abnormal adversarial examples. By comparing the proposed methods with other baselines on CIFAR10 and CIFAR100 datasets, the evaluation results show that the proposed method achieves better robustness against the PGD attack.\n", "strength_and_weaknesses": "Strength: \n+ This paper points out the existence of abnormal adversarial examples when robust overfitting happens and speculates that robust overfitting in single-step adversarial training is caused by abnormal adversarial examples. \n\n+ This paper proposed a novel regularization term to reduce the number of abnormal adversarial examples in training.\n\n+ The evaluation results show that the proposed method outperforms other baseline methods.\n", "clarity,_quality,_novelty_and_reproducibility": "Weakness:\n\n+ There are several places that I am not quite clear about in writing and may need further clarification:\n1. In equation 5, is $\\eta$ the random noise in the allowed perturbation ball? Why should we need an $\\eta$ here? \n2. In Line 7 to line 11 of algorithm 1, $\\eta$ is calculated for each iteration. Why does Algorithm 1 need this? What is the role of this component in the algorithm? What is $\\eta_t$ and $\\delta_t$ here.\n3. In Line 15, there is a multiplier $t/T$, what is this for?\n\n+ I don’t find some important experiment details: What is the number of training epochs and learning rate scheduler for the proposed training method? Which checkpoint are you using for evaluation? What’s the robustness difference between the best checkpoint and the last checkpoint? It will be nice to draw a figure to show how robustness changes for the proposed algorithm (a figure like Figure.1). \n\n+ The robustness reported in Table.1 is evaluated with the PGD attack, which is always found an overestimated robustness issue. Compared to the PGD attack, AutoAttack is usually more reliable to evaluate the robustness. Although the appendix compares the AA robustness with the vanilla PGD-based adversarial training method, it will be more convincible to compare the robustness of the proposed method with other SOTA baselines against AutoAttack.\n", "summary_of_the_review": "There are several points/places lacking clarity in the paper  (as discussed in the weakness). Further clarification and details may be necessary.\n\nReproducibility: The experiment settings section does not include all experiment details and no code is provided.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666635377487}, {"id": "zWGwZY2t6_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper656/Reviewer_gJHh"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Recently, FGSM-based adversarial training faced an emergency problem of catastrophic overfitting. To mitigate this, the authors first observed that the catastrophic overfitting is related to some abnormal examples, whose loss decreases under the FGSM attack. Based on this observation, in this paper they proposed a new abnormal adversarial example regularization (AAER), which considers two key components: one is about the number of the abnormal examples, while the other is about the logits variation. The experiments on CIFAR-10/100 show that the proposed AAER is effectiveness.\n  ", "review_text": "This paper finds some interesting observations, but the proposed technique does not appear to be strongly correlated with these observations. Moreover, the experiments are not very sufficient. ", "strengths": "**Strength**:\n1. The observation of the abnormal adversarial examples is interesting. \n2. The experiments show their effectiveness on two widely-used CIFAR datasets.\n\n**Weaknesses**:\n1. In section 3, the authors analyzed the number and logit variation of normal and abnormal examples, respectively. Section 3.1 shows that the abnormal adversarial examples dominate the catastrophic overfitting since the number has a significant increase when catastrophic overfitting exists. In section 3.2, the authors found that the logit variations during the inner process will increase after the catastrophic overfitting. Both of these two observations are new and interesting. And the authors claimed that their proposed AAER is inspired by them. However, it's very difficult for me to understand how Eq.7 is derived from these two observations? \n1.1 From the second observation of logit variations, whether it is a normal sample or an abnormal sample, the change of logical distance shows an increasing trend, and the trend of distance change is also the same. Therefore, why does Eq.7 reduce the gap between these two type of adversarial examples? \n1.2 At the same time, Eq.7 is not very rigorous, because the $x_k$ in the second term may contain abnormal examples. \n1.3 The authors wanted to reduce the number of the abnormal examples. So they introduced a term $\\frac{n}{m}$ in Eq.8. But the authors did not analyze the mechanism of this design.\n\n2. The authors conducted experiments on CIFAR-10 and CIFAR-100. I don't think the experiments are enough. As we know, FGSM-based adversarial training is one of the most effective training schemes that can also train the model on a large-scale dataset. So it's better to evaluate the proposed method and analyze the properties of abnormal adversarial examples on a large dataset, like Tiny-ImageNet or ImageNet. \n\n3. Since the proposed AAER is to reduce the impact of abnormal adversarial examples. However, there were no corresponding experiments on this part in the experiment. For example, show the number curve after using AAER (as shown in Figure 3 (a)). Also, as training progresses, it is better to show the loss curve.\n\n4. On page 7, the authors said that they used the unclipped perturbation to achieve better robustness. However, this is not a commonly used technique. This may cause the pixel value of the adversarial sample to exceed the actual value. The authors need to clarify the reasons for using this technique.\n\n5. On page 8, what's the parameter $\\beta$? In previous section, I didn't find this hyper-parameter. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "Recently, FGSM-based adversarial training faced an emergency problem of catastrophic overfitting. To mitigate this, the authors first observed that the catastrophic overfitting is related to some abnormal examples, whose loss decreases under the FGSM attack. Based on this observation, in this paper they proposed a new abnormal adversarial example regularization (AAER), which considers two key components: one is about the number of the abnormal examples, while the other is about the logits variation. The experiments on CIFAR-10/100 show that the proposed AAER is effectiveness.\n  ", "strength_and_weaknesses": "**Strength**:\n1. The observation of the abnormal adversarial examples is interesting. \n2. The experiments show their effectiveness on two widely-used CIFAR datasets.\n\n**Weaknesses**:\n1. In section 3, the authors analyzed the number and logit variation of normal and abnormal examples, respectively. Section 3.1 shows that the abnormal adversarial examples dominate the catastrophic overfitting since the number has a significant increase when catastrophic overfitting exists. In section 3.2, the authors found that the logit variations during the inner process will increase after the catastrophic overfitting. Both of these two observations are new and interesting. And the authors claimed that their proposed AAER is inspired by them. However, it's very difficult for me to understand how Eq.7 is derived from these two observations? \n1.1 From the second observation of logit variations, whether it is a normal sample or an abnormal sample, the change of logical distance shows an increasing trend, and the trend of distance change is also the same. Therefore, why does Eq.7 reduce the gap between these two type of adversarial examples? \n1.2 At the same time, Eq.7 is not very rigorous, because the $x_k$ in the second term may contain abnormal examples. \n1.3 The authors wanted to reduce the number of the abnormal examples. So they introduced a term $\\frac{n}{m}$ in Eq.8. But the authors did not analyze the mechanism of this design.\n\n2. The authors conducted experiments on CIFAR-10 and CIFAR-100. I don't think the experiments are enough. As we know, FGSM-based adversarial training is one of the most effective training schemes that can also train the model on a large-scale dataset. So it's better to evaluate the proposed method and analyze the properties of abnormal adversarial examples on a large dataset, like Tiny-ImageNet or ImageNet. \n\n3. Since the proposed AAER is to reduce the impact of abnormal adversarial examples. However, there were no corresponding experiments on this part in the experiment. For example, show the number curve after using AAER (as shown in Figure 3 (a)). Also, as training progresses, it is better to show the loss curve.\n\n4. On page 7, the authors said that they used the unclipped perturbation to achieve better robustness. However, this is not a commonly used technique. This may cause the pixel value of the adversarial sample to exceed the actual value. The authors need to clarify the reasons for using this technique.\n\n5. On page 8, what's the parameter $\\beta$? In previous section, I didn't find this hyper-parameter. ", "clarity,_quality,_novelty_and_reproducibility": "The observations are new and I think the proposed method is easy to reproduce. But the experiments are not very sufficient to verify the proposed regularization technique.", "summary_of_the_review": "This paper finds some interesting observations, but the proposed technique does not appear to be strongly correlated with these observations. Moreover, the experiments are not very sufficient. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666597771584}, {"id": "M6m51BpRra", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper656/Reviewer_5TBo"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes Abnormal Adversarial Examples Regularization (AAER), a regularizer to address the known catastrophic overfitting issue observed in fast adversarial training (AT). AAER is based on suppressing the output difference of abnormal adversarial examples (irregular adversarial samples such that their loss is smaller after adding perturbation) and their number, as well as maximizing the difference of model output between unperturbed and perturbed samples. ", "review_text": "1. Figure 3 is a very nice motivating example of the challenge and the proposed solution. However, it is unclear whether the emergence of abnormal examples is unique to the fast AT setting or not. To complete the analysis, I suggest running the same analysis on standard AT. If standard AT does not suffer from this issue but fast AT does, then this is a more convincing analysis.\n\n2. How critical is the second term in eq. (7)? In Line 15 of Algorithm 1, I noticed the authors use a zero threshold $\\max${$\\cdot$,0}. Does this suggest the second term in eq. (7), which contributes to negative numbers fo AAER, may not be important? Can the authors run an ablation study of the experiments with and without the second term?\n\n3. In most cases, the AAER_RUC method is the best regularizer. However, it's unclear to me what \"unclipped perturbation\" means in this context. Do the authors mean allowing the perturbed samples to go beyond the pixel values, or the authors did not clip the norm of the perturbation? Both cases make little sense to me, and the result needs more justification.\n\n4. Can AAER show complementary performance gain together with existing methods (e.g., gradient alignment) to mitigate catastrophic forgetting? For instance, any benefit of combining AAER with any of the existing methods in Table 1?\n", "strengths": "Strength:\n1. The proposed method is simple to implement and has low overhead\n\nWeakness:\n1. The uniqueness of Fig. 3 to fast AT needs more justification\n2. The role of the second term in eq. (7) is unclear\n3. What is \"unclipped perturbation\"? Is it a rational regularizer? Why it gives better performance?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes Abnormal Adversarial Examples Regularization (AAER), a regularizer to address the known catastrophic overfitting issue observed in fast adversarial training (AT). AAER is based on suppressing the output difference of abnormal adversarial examples (irregular adversarial samples such that their loss is smaller after adding perturbation) and their number, as well as maximizing the difference of model output between unperturbed and perturbed samples. ", "strength_and_weaknesses": "Strength:\n1. The proposed method is simple to implement and has low overhead\n\nWeakness:\n1. The uniqueness of Fig. 3 to fast AT needs more justification\n2. The role of the second term in eq. (7) is unclear\n3. What is \"unclipped perturbation\"? Is it a rational regularizer? Why it gives better performance?", "clarity,_quality,_novelty_and_reproducibility": "Although the proposed method is novel and addresses an important challenge in fast AT, the current presentation of this submission has several issues that need to be addressed. I list my main concerns and comments as follows.", "summary_of_the_review": "1. Figure 3 is a very nice motivating example of the challenge and the proposed solution. However, it is unclear whether the emergence of abnormal examples is unique to the fast AT setting or not. To complete the analysis, I suggest running the same analysis on standard AT. If standard AT does not suffer from this issue but fast AT does, then this is a more convincing analysis.\n\n2. How critical is the second term in eq. (7)? In Line 15 of Algorithm 1, I noticed the authors use a zero threshold $\\max${$\\cdot$,0}. Does this suggest the second term in eq. (7), which contributes to negative numbers fo AAER, may not be important? Can the authors run an ablation study of the experiments with and without the second term?\n\n3. In most cases, the AAER_RUC method is the best regularizer. However, it's unclear to me what \"unclipped perturbation\" means in this context. Do the authors mean allowing the perturbed samples to go beyond the pixel values, or the authors did not clip the norm of the perturbation? Both cases make little sense to me, and the result needs more justification.\n\n4. Can AAER show complementary performance gain together with existing methods (e.g., gradient alignment) to mitigate catastrophic forgetting? For instance, any benefit of combining AAER with any of the existing methods in Table 1?\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666324152874}], "openreview_url": "https://openreview.net/forum?id=YjKqWExiy6s", "arxiv_id": "2404.08154", "paper_pdf": "papers/YjKqWExiy6s.pdf", "paper_pdf_sha256": "75560f0a0429313afd4fdb7a50f72d05a1d31dbc36334b364aa475a022705a17", "paper_pdf_bytes": 2198026, "paper_pdf_source": "openreview", "code_url": "https://github.com/tmllab/2023_NeurIPS_AAER", "code_repository": "tmllab/2023_NeurIPS_AAER", "code_commit": "5ad2591c1f9be9e797420f713aa45a3e288da269", "code_archive": "repos/YjKqWExiy6s.zip", "code_archive_sha256": "91cd60175e8c2823362921813b3f1faa6fcdd70e39e4102b1cd25c4e56074800", "code_archive_bytes": 773838, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 767, "github_languages": {"Python": 112075}, "github_archived": false, "github_pushed_at": "2025-02-15T07:05:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/eliminating-catastrophic-overfitting-via-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9Cwxjd6nRh", "year": 2022, "status": "rejected", "title": "High Fidelity Visualization of What Your Self-Supervised Representation Knows About", "authors": ["Florian Bordes", "Randall Balestriero", "Pascal Vincent"], "authorids": ["~Florian_Bordes1", "~Randall_Balestriero1", "~Pascal_Vincent1"], "authors_source": "OpenReview API", "abstract": "Discovering what is learned by neural networks remains a challenge. In self-supervised learning, classification is the most common task used to evaluate how good a representation is. However, relying only on such downstream task can limit our understanding of how much information is contained in the representation of a given input. In this work, we study how to visualize representations learned with self-supervised models. We investigate a simple gradient descend based method to match a target representation and show the limitations of such techniques. We overcome these limitations by developing a representation-conditioned diffusion model (RCDM) that is able to generate high-quality inputs that share commonalities with a given representation. We further demonstrate how our model's generation quality is on par with state-of-the-art generative models and how the representation conditioning brings new avenues to analyze and improve self-supervised models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "h8KZkBLY1vQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1720/Reviewer_MCTj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a representation analysis technique to particularly understand what information is contained in self-supervised models. The authors propose a visualization technique based on conditional diffusion models that seeks to synthesize realistic-looking inputs whose representation matches that of a target image. Finding commonalities among these synthesized images likely provides clues about the target representation. ", "review_text": "Strengths: \n\n1. The question is meaningful, and the proposed technique is interesting. \n2. Visualizations generated using the proposed technique look promising in addressing the question laid out in the study.\n3. The paper is very well-written, and the analyses are well-motivated.\n\nCritiques: \n1. Some conclusions are stated as though they are specific to SSL but it might be good to know whether or not they are applicable for supervised training as well. For instance, the authors claim that `mapping to the same SSL representation as a natural image is not sufficient for producing a similar realistic-looking image’. My intuition is that that should be true for the supervised case as well due to the very nature of gradient-based matching on a deep layer representation when starting from random noise images. This result is not very surprising and has been described before. \n\n2. Imposing the 'naturalness’ constraint always arises out of a need for human understandability but it is not the most faithful account of what information is contained in the representation. I am a little concerned by the use of the phrase 'high fidelity’ in this setting to describe the visualization/model inversion process.  Perhaps it would be useful to briefly discuss the effect of different constraints on the input space in terms of fidelity – one with fewer control knobs (regularizations, naturalness prior etc.) is going to rank better on the fidelity scale..?  And the perceptual quality of image generation is not likely a good reflection of its faithfulness to the representation. \n\n\n3. Given that the goal was to characterize (through visualization) what is contained in a self-supervised representation, I think the results are not particularly intriguing and seem inadequate in their present form. Additional work and more rigorous analyses, including exploring more models, producing stronger results by analyzing more images etc., seem needed to strengthen the paper. \n\n4. The proposed method for analyzing representations is also highly subjective (not unlike other interpretability approaches I concede) but still there are no recommendations or thorough analysis about how to find the key factors driving a representation. e.g., how many images should be sampled from the RCDM to find the commonalities? Further, the conclusions that are drawn using this conditional image syntheses approach could just as well have been drawn by analyzing the representations directly (for instance, by seeing how well the scale of images can be decoded from the representation). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a representation analysis technique to particularly understand what information is contained in self-supervised models. The authors propose a visualization technique based on conditional diffusion models that seeks to synthesize realistic-looking inputs whose representation matches that of a target image. Finding commonalities among these synthesized images likely provides clues about the target representation. ", "main_review": "Strengths: \n\n1. The question is meaningful, and the proposed technique is interesting. \n2. Visualizations generated using the proposed technique look promising in addressing the question laid out in the study.\n3. The paper is very well-written, and the analyses are well-motivated.\n\nCritiques: \n1. Some conclusions are stated as though they are specific to SSL but it might be good to know whether or not they are applicable for supervised training as well. For instance, the authors claim that `mapping to the same SSL representation as a natural image is not sufficient for producing a similar realistic-looking image’. My intuition is that that should be true for the supervised case as well due to the very nature of gradient-based matching on a deep layer representation when starting from random noise images. This result is not very surprising and has been described before. \n\n2. Imposing the 'naturalness’ constraint always arises out of a need for human understandability but it is not the most faithful account of what information is contained in the representation. I am a little concerned by the use of the phrase 'high fidelity’ in this setting to describe the visualization/model inversion process.  Perhaps it would be useful to briefly discuss the effect of different constraints on the input space in terms of fidelity – one with fewer control knobs (regularizations, naturalness prior etc.) is going to rank better on the fidelity scale..?  And the perceptual quality of image generation is not likely a good reflection of its faithfulness to the representation. \n\n\n3. Given that the goal was to characterize (through visualization) what is contained in a self-supervised representation, I think the results are not particularly intriguing and seem inadequate in their present form. Additional work and more rigorous analyses, including exploring more models, producing stronger results by analyzing more images etc., seem needed to strengthen the paper. \n\n4. The proposed method for analyzing representations is also highly subjective (not unlike other interpretability approaches I concede) but still there are no recommendations or thorough analysis about how to find the key factors driving a representation. e.g., how many images should be sampled from the RCDM to find the commonalities? Further, the conclusions that are drawn using this conditional image syntheses approach could just as well have been drawn by analyzing the representations directly (for instance, by seeing how well the scale of images can be decoded from the representation). \n", "summary_of_the_review": "I enjoyed reading the paper. However, I'm uncertain about how much value this paper adds in efforts towards understanding the nature of self-supervised representations and the results are not very interesting. As a result, I think the paper falls slightly below the acceptance threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635992640926}, {"id": "qAfb1xqone6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1720/Reviewer_rQaS"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a technique for visualizing the representations of fixed pretrained self-supervised neural networks. More specifically, by using a conditional denoising diffusion probabilistic model (DDPM, coined _RCDM_ here), a learned neural network representation $h$ can be mapped (back) to image space (i.e., the space of \"natural images\") and, by drawing multiple samples conditioned on the same $h$, the invariances learned by the (self-supervised) network can be visualized. Experiments with different pre-trained models show that the method is indeed able to synthesize high quality natural images corresponding to the learned representations.", "review_text": "The main idea of the paper, i.e. explaining the representations (and their invariances) learned by deep neural networks via visualizations through natural images is intuitive and very reasonable. By demonstrating that gradient-based methods do not suffice (Sec. 2) to reconstruct _natural_ images, the authors show the need for a probabilistic approach and, by opting for a conditional diffusion model, provide a method capable of generating high-quality images corresponding to the representations, as demonstrated in the experimental section. The main drawback of this paper, however, is that this idea is not novel and the paper fails to discuss prior work which uses different implementations for the conditional generative model, e.g. GANs [1], autoregressive models [2] or normalizing flows [3]. Authors should clearly situate their work within this existing literature and discuss what advantages/differences an approach based on DDPMs offers. For example, is it possible to use a deterministic approximation of the diffusion process (such as DDIM, see [4]) to extract, analogously to [3], an explicit representation of the learned invariances? Further, I do not quite understand the focus on SSL representations, as the proposed techniques can easily be applied to any type of representation, independent of the training paradigm. \n\nMoreover, a quantitative assessment of the correspondences between representations and synthesized images is missing. For example, can the synthesized representations be fed back into the (SSL) model and does this then yield the same representation $h$?\n\n_Mixed comments:_\n\n- I think the paper would benefit from focusing on __explaining__ (SSL) representations and thus dropping the part on synthesis (mainly Appendix B), as it mainly relies on methods that already have been developed and discussed in the context of pure synthesis (ADM).\n- On the contrary, Appendix E and especially Fig. 20 should be moved to the main paper, because they provide interesting analysis of different SSL models/representations.\n- 1st paragraph of 4.2.: I do not quite understand the motivation here: In the case the density of $p(x|h)$ approaches a delta-distribution, the task becomes more easy and in the limit, could be implemented through a regression model.\n- Following on from this: the paper should demonstrate the need for a probabilistic approach by comparing, for example, with a simple regression baseline $x = f_{\\theta}(h)$. A more advanced baseline would be the IC-GAN discussed in Sec. 4.2 (without neighborhood sampling).\n- _\"Diffusion models do not usually explicitly express an energy function\"_ (Footnote 4) The usual training mechanism is to train the reweighted epsilon-parameterization from [5] which corresponds to a denoised score matching objective. This score corresponds to the derivative of an energy which can be plugged into MCMC sampling, see [6].\n- The figure on the first page should show several examples of each model to emphasize the visualization capability of the invariances.\n- The interpolation in Fig. 4c should show several \"interpolation-paths\", demonstrating the stochasticity of the proposed approach (DDPM)\n- Does the proposed conditioning mechanism via conditional batch norm provide significant advantages over other methods (e.g. addition to the timestep-embedding, AdaIN, cross-attention, ...)?\n\n__References:__\n- [1]: Shocher, A., Gandelsman, Y., Mosseri, I., Yarom, M., Irani, M., Freeman, W.T., Dekel, T.: Semantic Pyramid for Image Generation. \n- [2]: Nash, C., Kushman, N., Williams, C.K.: Inverting Supervised Representations with Autoregressive Neural Density Models.\n- [3]: Rombach, R., Esser, P., Ommer, B.: Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with INNs\n- [4]: Song, J., Meng, C., Ermon, S.: Denoising Diffusion Implicit Models\n- [5]: Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models\n- [6]: Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a technique for visualizing the representations of fixed pretrained self-supervised neural networks. More specifically, by using a conditional denoising diffusion probabilistic model (DDPM, coined _RCDM_ here), a learned neural network representation $h$ can be mapped (back) to image space (i.e., the space of \"natural images\") and, by drawing multiple samples conditioned on the same $h$, the invariances learned by the (self-supervised) network can be visualized. Experiments with different pre-trained models show that the method is indeed able to synthesize high quality natural images corresponding to the learned representations.", "main_review": "The main idea of the paper, i.e. explaining the representations (and their invariances) learned by deep neural networks via visualizations through natural images is intuitive and very reasonable. By demonstrating that gradient-based methods do not suffice (Sec. 2) to reconstruct _natural_ images, the authors show the need for a probabilistic approach and, by opting for a conditional diffusion model, provide a method capable of generating high-quality images corresponding to the representations, as demonstrated in the experimental section. The main drawback of this paper, however, is that this idea is not novel and the paper fails to discuss prior work which uses different implementations for the conditional generative model, e.g. GANs [1], autoregressive models [2] or normalizing flows [3]. Authors should clearly situate their work within this existing literature and discuss what advantages/differences an approach based on DDPMs offers. For example, is it possible to use a deterministic approximation of the diffusion process (such as DDIM, see [4]) to extract, analogously to [3], an explicit representation of the learned invariances? Further, I do not quite understand the focus on SSL representations, as the proposed techniques can easily be applied to any type of representation, independent of the training paradigm. \n\nMoreover, a quantitative assessment of the correspondences between representations and synthesized images is missing. For example, can the synthesized representations be fed back into the (SSL) model and does this then yield the same representation $h$?\n\n_Mixed comments:_\n\n- I think the paper would benefit from focusing on __explaining__ (SSL) representations and thus dropping the part on synthesis (mainly Appendix B), as it mainly relies on methods that already have been developed and discussed in the context of pure synthesis (ADM).\n- On the contrary, Appendix E and especially Fig. 20 should be moved to the main paper, because they provide interesting analysis of different SSL models/representations.\n- 1st paragraph of 4.2.: I do not quite understand the motivation here: In the case the density of $p(x|h)$ approaches a delta-distribution, the task becomes more easy and in the limit, could be implemented through a regression model.\n- Following on from this: the paper should demonstrate the need for a probabilistic approach by comparing, for example, with a simple regression baseline $x = f_{\\theta}(h)$. A more advanced baseline would be the IC-GAN discussed in Sec. 4.2 (without neighborhood sampling).\n- _\"Diffusion models do not usually explicitly express an energy function\"_ (Footnote 4) The usual training mechanism is to train the reweighted epsilon-parameterization from [5] which corresponds to a denoised score matching objective. This score corresponds to the derivative of an energy which can be plugged into MCMC sampling, see [6].\n- The figure on the first page should show several examples of each model to emphasize the visualization capability of the invariances.\n- The interpolation in Fig. 4c should show several \"interpolation-paths\", demonstrating the stochasticity of the proposed approach (DDPM)\n- Does the proposed conditioning mechanism via conditional batch norm provide significant advantages over other methods (e.g. addition to the timestep-embedding, AdaIN, cross-attention, ...)?\n\n__References:__\n- [1]: Shocher, A., Gandelsman, Y., Mosseri, I., Yarom, M., Irani, M., Freeman, W.T., Dekel, T.: Semantic Pyramid for Image Generation. \n- [2]: Nash, C., Kushman, N., Williams, C.K.: Inverting Supervised Representations with Autoregressive Neural Density Models.\n- [3]: Rombach, R., Esser, P., Ommer, B.: Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with INNs\n- [4]: Song, J., Meng, C., Ermon, S.: Denoising Diffusion Implicit Models\n- [5]: Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models\n- [6]: Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations\n", "summary_of_the_review": "I think that the present paper deals with a very important problem, namely the \"understanding\" of learned neural representations. The proposed approach to do this by synthesizing \"natural\" images generated by a generative model conditioned on these learned representations (here a diffusion model) is intuitive and reasonable. The generated samples are of high quality and seem to be in agreement with the representations. \n\nUnfortunately, this approach is not novel (only the use of a diffusion model instead of {GANS, ARMs, flows} is new here) and the present work would, in my opinion, benefit greatly from being placed in the context of this existing work. In addition, there are some specific questions (see above) that I think are not yet answered satisfactorily. \n\n++++++++++++++++++++++++++\n\nScore raised to 6 after the rebuttal\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635855705138}, {"id": "BQhP1rmp6dK", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1720/Reviewer_YLR1"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work introduces develops a method for sampling different natural images that a pretrained encoder maps to the same or similar representation. To do this, the authors propose a modification to the diffusion model of Dhariwal & Nichol for sampling conditioned on a given encoded representation $h=f(x)$ that is well suited to conditioning on high dimensional vectors (i.e., representations). This new method is used to qualitatively compare encoders trained via a number of methods (supervised, DINO, SimCLR, Barlow Twins, VICReg etc.).\n\n", "review_text": "\nThis paper considers a very worthy problem, that of analyzing encoders via visualizations, and clearly develops a high quality generative modeling approach with which to do this. I think this paper may be on its way to being a vey nice paper, however I have concerns with it’s current state. \n\nCurrent strengths include:\n\n- high quality of generative models used (the generated samples are high fidelity). \n- plenty of qualitative evaluation - the appendix contains many pages of additional visualizations. \n- the method is able to visually discern some differences between supervised and self-supervised trained encoders. The differences between  self-supervised methods seems a little harder to extract, besides noting that some (e.g., DINO) exhibit better invariance than others (e.g., SimCLR).\n\n\nMy two broad concerns (i.e., \"weaknesses\"), which I discuss in more detail momentarily, are:\n\n- A lack of *quantitative results*. All results presented are qualitative and visual in nature (with the exception of a sanity check that the generative models are producing photorealistic samples). \n- A lack of demonstration of useful applications for the proposed method. \n\n**Quantitative results:**\n\nWhile a large part of generative modeling is focused on qualitative production of photorealistic samples, in order to properly compare you method and analyze pertained encoders I think it is critical to have more quantitative evaluation. \n\nI am sure the authors can come up with many better ideas for additional quantitative evaluations than I can, but I will also attempt to offer some actionable feedback:\n\n- The main premise of the work is that the samples $x’$ from your diffusion model are such that $f(x’)$ are very close to the original representation h. It seems almost a must that this is quantitatively confirmed. \n- You argue verbally that for a given representation $h$, IC-GAN generates samples which map to a neighborhood of $h$, whereas your method generates samples that map very closely to $h$. While this may very well be true, it would be much more convincing if you quantitatively showed this. \n\n**Applications of method:**\n\nAt present, the work mainly attempts to visualize some simple properties that one would expect a self-supervised encoder to have - e.g., a more instance-level invariance then supervised encoders, and invariance to the data augmentations used training training. Unfortunately I don't find this to be sufficiently insightful right now. In order to truly demonstrate the value of this method, it would be very valuable to take more steps towards devising practically useful tools that are enabled by this method. Some ideas of the kind of thing I am imagining are:\n\n- Could your visualization method be used to determine a priori which out of a group of encoders will perform best at a given downstream task. If possible, this would be very cool since the trial and error evaluation of pretrained encoders is annoyingly heuristic. E.g. specifically, could you develop some metric based on your visualization method that correlates well with (e.g.,) segmentation performance. \n- Suppose you have two encoders $f$ and $f’$, and you know that f performs much better on some downstream task $T$. Could you use the diffusion model for $f$ to *generate new training samples* with which to finetune the encoder f’ itself (not just the linear probe) so that the performance of $f’$ on task $T$ improves. \n\n\nTo be clear, I am not asking you to try any of these ideas specifically. Instead I am trying to illustrate the type of thing I envisage when I ask for extensions that analyze encoders in more detail. This is a promise that was made in the abstract — “new avenues to analyze and improve self-supervised models” — but which was underdelivered on in my eyes. \n\n---\n**Miscellaneous comments and questions:**\n\n- The difference in FID and IS for your method compared to ADM seems very large considering your method is a modification of theirs. Do you know what accounts for this? Is it better model training? Is it the fact that your method is instance conditioned, & therefore requires less variation in samples? Something else?\n- Some experiments seem orthogonal to the idea of analyzing pretrained encoders. E.g. the value of Fig 4c) is unclear to me (4 a) and b) are good) and seem to be more of the form of a “sanity check” that the generative model is working sensibly. \n- Please fix your bibliography. Many citations are missing their publication venues (e.g. Barlow Twins and at least three of Yang Song’s papers), and some citation simply list: 1) authors names, 2) paper’s name, 3) year — i.e. they don’t even have an arXiv number, let alone a publication venue (e.g. SimCLR and VICReg papers). \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work introduces develops a method for sampling different natural images that a pretrained encoder maps to the same or similar representation. To do this, the authors propose a modification to the diffusion model of Dhariwal & Nichol for sampling conditioned on a given encoded representation $h=f(x)$ that is well suited to conditioning on high dimensional vectors (i.e., representations). This new method is used to qualitatively compare encoders trained via a number of methods (supervised, DINO, SimCLR, Barlow Twins, VICReg etc.).\n\n", "main_review": "\nThis paper considers a very worthy problem, that of analyzing encoders via visualizations, and clearly develops a high quality generative modeling approach with which to do this. I think this paper may be on its way to being a vey nice paper, however I have concerns with it’s current state. \n\nCurrent strengths include:\n\n- high quality of generative models used (the generated samples are high fidelity). \n- plenty of qualitative evaluation - the appendix contains many pages of additional visualizations. \n- the method is able to visually discern some differences between supervised and self-supervised trained encoders. The differences between  self-supervised methods seems a little harder to extract, besides noting that some (e.g., DINO) exhibit better invariance than others (e.g., SimCLR).\n\n\nMy two broad concerns (i.e., \"weaknesses\"), which I discuss in more detail momentarily, are:\n\n- A lack of *quantitative results*. All results presented are qualitative and visual in nature (with the exception of a sanity check that the generative models are producing photorealistic samples). \n- A lack of demonstration of useful applications for the proposed method. \n\n**Quantitative results:**\n\nWhile a large part of generative modeling is focused on qualitative production of photorealistic samples, in order to properly compare you method and analyze pertained encoders I think it is critical to have more quantitative evaluation. \n\nI am sure the authors can come up with many better ideas for additional quantitative evaluations than I can, but I will also attempt to offer some actionable feedback:\n\n- The main premise of the work is that the samples $x’$ from your diffusion model are such that $f(x’)$ are very close to the original representation h. It seems almost a must that this is quantitatively confirmed. \n- You argue verbally that for a given representation $h$, IC-GAN generates samples which map to a neighborhood of $h$, whereas your method generates samples that map very closely to $h$. While this may very well be true, it would be much more convincing if you quantitatively showed this. \n\n**Applications of method:**\n\nAt present, the work mainly attempts to visualize some simple properties that one would expect a self-supervised encoder to have - e.g., a more instance-level invariance then supervised encoders, and invariance to the data augmentations used training training. Unfortunately I don't find this to be sufficiently insightful right now. In order to truly demonstrate the value of this method, it would be very valuable to take more steps towards devising practically useful tools that are enabled by this method. Some ideas of the kind of thing I am imagining are:\n\n- Could your visualization method be used to determine a priori which out of a group of encoders will perform best at a given downstream task. If possible, this would be very cool since the trial and error evaluation of pretrained encoders is annoyingly heuristic. E.g. specifically, could you develop some metric based on your visualization method that correlates well with (e.g.,) segmentation performance. \n- Suppose you have two encoders $f$ and $f’$, and you know that f performs much better on some downstream task $T$. Could you use the diffusion model for $f$ to *generate new training samples* with which to finetune the encoder f’ itself (not just the linear probe) so that the performance of $f’$ on task $T$ improves. \n\n\nTo be clear, I am not asking you to try any of these ideas specifically. Instead I am trying to illustrate the type of thing I envisage when I ask for extensions that analyze encoders in more detail. This is a promise that was made in the abstract — “new avenues to analyze and improve self-supervised models” — but which was underdelivered on in my eyes. \n\n---\n**Miscellaneous comments and questions:**\n\n- The difference in FID and IS for your method compared to ADM seems very large considering your method is a modification of theirs. Do you know what accounts for this? Is it better model training? Is it the fact that your method is instance conditioned, & therefore requires less variation in samples? Something else?\n- Some experiments seem orthogonal to the idea of analyzing pretrained encoders. E.g. the value of Fig 4c) is unclear to me (4 a) and b) are good) and seem to be more of the form of a “sanity check” that the generative model is working sensibly. \n- Please fix your bibliography. Many citations are missing their publication venues (e.g. Barlow Twins and at least three of Yang Song’s papers), and some citation simply list: 1) authors names, 2) paper’s name, 3) year — i.e. they don’t even have an arXiv number, let alone a publication venue (e.g. SimCLR and VICReg papers). \n\n", "summary_of_the_review": "\nThe reasons I am concerned about these two main objections are: \n- quantitative results are almost a must in order to properly evaluate methodological progress, and  check that the method is behaving as it should, and to properly compare models/methods etc.,\n- The methodological contribution of the representation conditioned diffusion model is relatively small (it sufficed to simply describe the changes one needs to make to ADM in words). This is not a problem in itself. But I do think it is therefore important to demonstrate the method's usefulness in analyzing and understanding pretrained models. \n\nConsequently, I am currently not in favor of acceptance. However, to recognize the generally high quality of the work thus far I am opting for a weak reject. I am unlikely to raise the score to an accept without significant updates to the work, or strong arguments in favor from other reviewers. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635801457745}, {"id": "Wx7CGOt496", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1720/Reviewer_mFTx"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a conditional diffusion model that can be used to visualize representations learned by SSL or supervised models.\n\nThe proposed Representation-conditioned Diffusion Model (RCDM) can generate images that are both close in the representation space to a given image, and looking realistic.\n\nThe paper also provides empirical results comparing conditionally generated samples from models trained on features from the backbone or projector of a SSL model, as well as supervised features. These experiments provide insight to understanding SSL, such as showing that projector is the key of SSL being invariance to transformations.\n", "review_text": "**Strength**: This work provides a visualization tool for understanding NN representations. The proposed RCDM is not technically challenging (it is adapted from ADM in the prior work), but it works well for probing the information in NN representations without sacrificing generation quality.\n\nThe experiments cover different SSL methods and the comparison with supervised learning, showing results on 1) in/out of distribution generation, 2) interpolation, 3) super-resolution, 4) unconditioned generation, 5) algebraic manipulation. \n\nOther aspects:\n  - The writing is clear and easy to follow.\n  - The relation and comparison with prior work is discussed adequately.\n  - The code will be released.\n\n**Questions**:\n  - Figure 1: it would be better to provide more context for the relative distance: e.g. is the difference between 0.4% and 3.3% significant?  Should a 3% distance considered small?\n  - Figure 6: projector head vs backbone: are features from the projector able to get as close to the original feature as features from the backbone? If not, how much further are the projector features compared to the backbone features?\n\nMinor comments:\n  - The writing for Figure 7 is repetitive, i.e. the last paragraph on page 8 and the caption for Figure 7.\n  - last sentence in the first paragraph: \"...or discriminating instances.\"\n  - the paragraph below equation (2): should the dimension be $K$ rather than $S$?\n  - Sec 4.1, second line in the second paragraph: model \"dependent\"\n  - Last sentence of the caption of Figure 4: should be column 2 to 6 (rather than 7)\n  - Appendix section C: it should link to Figure 15 rather than Figure 21.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a conditional diffusion model that can be used to visualize representations learned by SSL or supervised models.\n\nThe proposed Representation-conditioned Diffusion Model (RCDM) can generate images that are both close in the representation space to a given image, and looking realistic.\n\nThe paper also provides empirical results comparing conditionally generated samples from models trained on features from the backbone or projector of a SSL model, as well as supervised features. These experiments provide insight to understanding SSL, such as showing that projector is the key of SSL being invariance to transformations.\n", "main_review": "**Strength**: This work provides a visualization tool for understanding NN representations. The proposed RCDM is not technically challenging (it is adapted from ADM in the prior work), but it works well for probing the information in NN representations without sacrificing generation quality.\n\nThe experiments cover different SSL methods and the comparison with supervised learning, showing results on 1) in/out of distribution generation, 2) interpolation, 3) super-resolution, 4) unconditioned generation, 5) algebraic manipulation. \n\nOther aspects:\n  - The writing is clear and easy to follow.\n  - The relation and comparison with prior work is discussed adequately.\n  - The code will be released.\n\n**Questions**:\n  - Figure 1: it would be better to provide more context for the relative distance: e.g. is the difference between 0.4% and 3.3% significant?  Should a 3% distance considered small?\n  - Figure 6: projector head vs backbone: are features from the projector able to get as close to the original feature as features from the backbone? If not, how much further are the projector features compared to the backbone features?\n\nMinor comments:\n  - The writing for Figure 7 is repetitive, i.e. the last paragraph on page 8 and the caption for Figure 7.\n  - last sentence in the first paragraph: \"...or discriminating instances.\"\n  - the paragraph below equation (2): should the dimension be $K$ rather than $S$?\n  - Sec 4.1, second line in the second paragraph: model \"dependent\"\n  - Last sentence of the caption of Figure 4: should be column 2 to 6 (rather than 7)\n  - Appendix section C: it should link to Figure 15 rather than Figure 21.\n\n", "summary_of_the_review": "The method in this paper is unsurprising but gives good results.\nI find the experiments thorough and insightful, and would recommend an accept.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635642730247}], "openreview_url": "https://openreview.net/forum?id=9Cwxjd6nRh", "arxiv_id": "2112.09164", "paper_pdf": "papers/9Cwxjd6nRh.pdf", "paper_pdf_sha256": "e2819039495fa338837ee48e0f35ca1bb67dbdb02f8d45d0cca8106991b64b60", "paper_pdf_bytes": 44596143, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/RCDM", "code_repository": "facebookresearch/RCDM", "code_commit": "71daaf10a73bb2012864f0827c68d209fc92b0a5", "code_archive": "repos/9Cwxjd6nRh.zip", "code_archive_sha256": "4778ee2854148947cb035b8eb56a6e92c898e8db70648271145fb82e02e8b9c2", "code_archive_bytes": 3425002, "code_file_count": 28, "code_extensions": {".py": 28}, "github_disk_usage_kb": 3345, "github_languages": {"Python": 255010}, "github_archived": true, "github_pushed_at": "2023-05-03T20:06:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/high-fidelity-visualization-of-what-your-self-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ohdw3t-8VCY", "year": 2021, "status": "rejected", "title": "CTRLsum: Towards Generic Controllable Text Summarization", "authors": ["Junxian He", "Wojciech Maciej Kryscinski", "Bryan McCann", "Nazneen Rajani", "Caiming Xiong"], "authorids": ["~Junxian_He1", "~Wojciech_Maciej_Kryscinski1", "~Bryan_McCann1", "~Nazneen_Rajani1", "~Caiming_Xiong1"], "authors_source": "OpenReview API", "abstract": "Current summarization systems yield generic summaries that are disconnected from users' preferences and expectations. To address this limitation, we present CTRLsum, a novel framework for controllable summarization. Our approach enables users to control multiple aspects of generated summaries by interacting with the summarization system through textual input in the form of a set of keywords or descriptive prompts. Using a single unified model, CTRLsum is able to achieve a broad scope of summary manipulation at inference time without requiring additional human annotations or pre-defining a set of control aspects during training. We quantitatively demonstrate the effectiveness of our approach on three domains of summarization datasets and five control aspects: 1) entity-centric and 2) length-controllable summarization, 3) contribution summarization on scientific papers, 4) invention purpose summarization on patent filings, and 5) question-guided summarization on news articles in a reading comprehension setting. Moreover, when used in a standard, uncontrolled summarization setting, CTRLsum achieves state-of-the-art results on the CNN/DailyMail dataset.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "seX3Hn_4roa", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2271/AnonReviewer5"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a two-stage summarization system where a document is provided along with (optionally) keywords or a prompt. This supplemental information helps to guide the summarization and possibly make it more user-specific. The keywords and prompt can also be guessed automatically by a BERT-base model, which seems to improve automatic metrics on CNN/daily mail.\n\nStrengths:\n* Moving beyond the conventional 'document/summary' framework of existing summarization approaches is a strength of this paper. This paper studies a few different ways that the summaries can be controlled: through prompts, entities, or one-sentence summaries of summaries (contribution and purpose summarization). These seem novel at least to this reviewer and could be helpful for future work.\n* When using oracle guidance, performance increases on several different datasets for slightly different forms of summarization (CNN/DM, arxiv, bigpatent).\n* The idea of using a two-stage approach (with a BERT-Base extractor to guess keywords to guide the summary) seems novel to this reviewer, and it seems to enable this approach to perform well even in an unconditional setting.\n\nWeaknesses:\n* The main weakness to this reviewer is that the evaluation might not be sufficiently convincing to test the key hypothesis: that these keywords/prompts can enable users to get summaries that are closer to their intent (like Figure 1). To this reviewer, this necessitates a human evaluation. Though testing factual correctness in Table 3 seems like a good start to this reviewer, measuring overall summarization quality (both conditional and unconditional on user intent) through a human evaluation seems necessary.\n* (minor) one possible reason why the two-stage approach might perform better on unconditional summarization is because there are more parameters when ensembling BERT-Base and BART. Possibly doing something multitask within a single BART model might be cleaner and could clearly test whether the gains come from more parameters/computation, or the keyword approach.\n\n\nOverall, to this reviewer, this paper seems like it would be strong if it had human evaluations of summarization quality. I would be willing to raise my score if those were provided.\n\n----\n\nUpdate: thanks for the additional human evaluation results! These help and the results on excluding unimportant entities seem strong to this reviewer. Perhaps it might be more helpful for the annotators themselves to try to interact with the summarizer in some way, but that's a more minor point.\n\nAnyways, I bumped up my score from 5->7.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "review": "This paper proposes a two-stage summarization system where a document is provided along with (optionally) keywords or a prompt. This supplemental information helps to guide the summarization and possibly make it more user-specific. The keywords and prompt can also be guessed automatically by a BERT-base model, which seems to improve automatic metrics on CNN/daily mail.\n\nStrengths:\n* Moving beyond the conventional 'document/summary' framework of existing summarization approaches is a strength of this paper. This paper studies a few different ways that the summaries can be controlled: through prompts, entities, or one-sentence summaries of summaries (contribution and purpose summarization). These seem novel at least to this reviewer and could be helpful for future work.\n* When using oracle guidance, performance increases on several different datasets for slightly different forms of summarization (CNN/DM, arxiv, bigpatent).\n* The idea of using a two-stage approach (with a BERT-Base extractor to guess keywords to guide the summary) seems novel to this reviewer, and it seems to enable this approach to perform well even in an unconditional setting.\n\nWeaknesses:\n* The main weakness to this reviewer is that the evaluation might not be sufficiently convincing to test the key hypothesis: that these keywords/prompts can enable users to get summaries that are closer to their intent (like Figure 1). To this reviewer, this necessitates a human evaluation. Though testing factual correctness in Table 3 seems like a good start to this reviewer, measuring overall summarization quality (both conditional and unconditional on user intent) through a human evaluation seems necessary.\n* (minor) one possible reason why the two-stage approach might perform better on unconditional summarization is because there are more parameters when ensembling BERT-Base and BART. Possibly doing something multitask within a single BART model might be cleaner and could clearly test whether the gains come from more parameters/computation, or the keyword approach.\n\n\nOverall, to this reviewer, this paper seems like it would be strong if it had human evaluations of summarization quality. I would be willing to raise my score if those were provided.\n\n----\n\nUpdate: thanks for the additional human evaluation results! These help and the results on excluding unimportant entities seem strong to this reviewer. Perhaps it might be more helpful for the annotators themselves to try to interact with the summarizer in some way, but that's a more minor point.\n\nAnyways, I bumped up my score from 5->7.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604696768570}, {"id": "gg7-miWPeLQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2271/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "# Summary:\n\nBuilds/extends on Controllable Abstractive Summarization (Fan et al) using keywords and other prompts. There’s two phases, and both phases are independent:\n\n1. Extract keywords, z,  using a BERT classifier/sequence-tagger trained to predict keywords\n2. Fine-tune BART (Lewis et al) to learn p(summary | document, z).\n\nOne can use automatic keywords using (1) and get uncontrolled generation, for which they present SOTA results on some summarization tasks. Two datasets are collected: (a) intro->contributions from arxiv papers; (b) patent->one-sentence summary, which are used to measure performance of prompts specific to those tasks.\n\n# Pros:\n1. Improves state-of-the-art results on some summarization benchmarks.\n2. Provides BERTScore results in addition to ROUGE.\n3. Results provided across multiple summarization datasets.\n4. Interesting new datasets for measuring document+prompt->summary performance\n5. Interesting zero-shot/transfer results from summarization to Question-answering.\n\n# Cons: \n1. Contribution in methods over Fan et al + BART (Lewis et al) is minimal. Results in Fan et al were weak because the underlying model was much weaker than BART, so it is unclear how much results here improved from simply using BART and adding control tokens from Fan et al, which would be a useful baseline to have that is omitted.\n2. Since the focus of the paper is on controlling generation, more results on this would be informative. It is unclear how well control works when not using oracle words or automatically extracted keywords, i.e. user-controlled. An MTurk experiment evaluating how well control works would be useful in assessing this.\n3. Comparing BART/PEGASUS to BART+BERT-based model is a little unfair since BERT is another large model in the system, i.e. the total amount of compute and number of parameters is much greater. A more fair comparison would be to compare using a smaller BART or PEGASUS model such that the model sizes are comparable. It is unclear whether the proposed system would do better than BART/PEGASUS scaled to the same amount of total parameters.\n4. For the prompt tests, how well does BART/PEGASUS do if the decoder is prompted? This baseline would be useful to have. That is, it’s unclear how the p(y | x, z) improves over using p(y | x) by simply prompting the decoder.\n\n# Clarifications/questions:\n1. For \"CONTRIBUTION AND PURPOSE SUMMARIZATION\", what is the fine-tuning process? Are the models fine-tuned on (paper/patent, keywords)->abstract task before testing on intro->contribution generation? \n2. How are entities randomly selected in the example decodes in the Appendix?\n3. Are any of the prompts used in training or is it zero-shot?\n4. What is zero-shot state-of-the-art on the QA tasks? Please add to Table 5. GPT-3 zero-shot results would also be informative.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting techniques for some summarization tasks, but unclear contribution over Fan et al.", "review": "# Summary:\n\nBuilds/extends on Controllable Abstractive Summarization (Fan et al) using keywords and other prompts. There’s two phases, and both phases are independent:\n\n1. Extract keywords, z,  using a BERT classifier/sequence-tagger trained to predict keywords\n2. Fine-tune BART (Lewis et al) to learn p(summary | document, z).\n\nOne can use automatic keywords using (1) and get uncontrolled generation, for which they present SOTA results on some summarization tasks. Two datasets are collected: (a) intro->contributions from arxiv papers; (b) patent->one-sentence summary, which are used to measure performance of prompts specific to those tasks.\n\n# Pros:\n1. Improves state-of-the-art results on some summarization benchmarks.\n2. Provides BERTScore results in addition to ROUGE.\n3. Results provided across multiple summarization datasets.\n4. Interesting new datasets for measuring document+prompt->summary performance\n5. Interesting zero-shot/transfer results from summarization to Question-answering.\n\n# Cons: \n1. Contribution in methods over Fan et al + BART (Lewis et al) is minimal. Results in Fan et al were weak because the underlying model was much weaker than BART, so it is unclear how much results here improved from simply using BART and adding control tokens from Fan et al, which would be a useful baseline to have that is omitted.\n2. Since the focus of the paper is on controlling generation, more results on this would be informative. It is unclear how well control works when not using oracle words or automatically extracted keywords, i.e. user-controlled. An MTurk experiment evaluating how well control works would be useful in assessing this.\n3. Comparing BART/PEGASUS to BART+BERT-based model is a little unfair since BERT is another large model in the system, i.e. the total amount of compute and number of parameters is much greater. A more fair comparison would be to compare using a smaller BART or PEGASUS model such that the model sizes are comparable. It is unclear whether the proposed system would do better than BART/PEGASUS scaled to the same amount of total parameters.\n4. For the prompt tests, how well does BART/PEGASUS do if the decoder is prompted? This baseline would be useful to have. That is, it’s unclear how the p(y | x, z) improves over using p(y | x) by simply prompting the decoder.\n\n# Clarifications/questions:\n1. For \"CONTRIBUTION AND PURPOSE SUMMARIZATION\", what is the fine-tuning process? Are the models fine-tuned on (paper/patent, keywords)->abstract task before testing on intro->contribution generation? \n2. How are entities randomly selected in the example decodes in the Appendix?\n3. Are any of the prompts used in training or is it zero-shot?\n4. What is zero-shot state-of-the-art on the QA tasks? Please add to Table 5. GPT-3 zero-shot results would also be informative.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603920334670}, {"id": "Qo366HRkgq_", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2271/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose an abstractive document summarization model that can\ngenerate summaries that target a specific set of keywords or prompts. This is\nin contrast to generic summarization models that learn to summarize a document\nbut are difficult to control or direct. The authors propose a straightforward\nway of obtaining keywords from an article similar in spirit to Gerhmann et al.\n2018. Alternatively, \"ground truth\" keywords can be found using a reference\nsummary. In either case, the keywords are prepended to the input document and\na BART model is fine-tuned to generate summaries using both the document and\nkeyword content.\n\nThe authors go on to show how a model trained in such a way, which they refer\nto as CTRLsum, can generate entity-focused summaries, by using an entity name\nas the keyword prefix. Additionally, providing differing numbers of keywords\ncan be used to control the length of the generated summary.  The authors also\nshow that CTRLsum can respond sensibly to prompts, i.e. instead of providing\nkeywords, a question or initial phrase is provided.  Useful summarization\nbehavior can be achieved including zero shot question answering, or\nenumeration of a research paper's contributions or the purpose of an\ninvention.\n\nWhile the paper feels largely like an extension of Keskar et al. 2019, the\nevaluation of the proposed methods is very thorough on a variety of settings\nand domains. I especially enjoyed the break out of entity targeted summaries\nbased on whether the entity occurred in the lead and/or reference summary. The\nuse of prompts to obtain question answering and more focused\ncontributions/purpose summarization was also very interesting.\n\nI would be happy to see this paper accepted to ICLR. This paper offers a\nsimple method of obtaining a variety of focused or targeted summarization\nbehaviors from a BART summarization model. In general, I would like to see\nmore work like this exploring methods of controlling pretrained language\nmodels.  The evaluation of the correctness of the generated utterances\nsuggests that this method provides fairly reliable control.\n\nThere several areas where the paper could improve. The explanation of how\nlength control is achieved was not very clear. It would help to have examples\nlike those shown in the appendix present in the section introducing length\ncontrol. \n\nComparisons are to the standard BART model or to Fan et al. (2018) which\nsimilarly prepend important control information to the input. It would be\ninteresting to see an evaluation that compared CTRLsum to BART with a\nconstrained decoding method, such as dynamic beam allocation [1].\n\nAdditionally, the authors should say more about the differences in entity\ncontrol of their method and Fan et al. 2018, which seem on their face to be\nsimilar.\n\nDid the authors experiment with pairs of entities as entity controls? It would\nbe especially interesting to see whether the model preserves the correct\nrelationship between entities, especially for entities that didn't occur in\nthe same sentences in the original document, e.g. one important entity and one\nunimportant entity.\n\n[1] Matt Post and David Vilar. Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation. ACL. 2018.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A simple but effective method of focusing abstractive summarization models.", "review": "The authors propose an abstractive document summarization model that can\ngenerate summaries that target a specific set of keywords or prompts. This is\nin contrast to generic summarization models that learn to summarize a document\nbut are difficult to control or direct. The authors propose a straightforward\nway of obtaining keywords from an article similar in spirit to Gerhmann et al.\n2018. Alternatively, \"ground truth\" keywords can be found using a reference\nsummary. In either case, the keywords are prepended to the input document and\na BART model is fine-tuned to generate summaries using both the document and\nkeyword content.\n\nThe authors go on to show how a model trained in such a way, which they refer\nto as CTRLsum, can generate entity-focused summaries, by using an entity name\nas the keyword prefix. Additionally, providing differing numbers of keywords\ncan be used to control the length of the generated summary.  The authors also\nshow that CTRLsum can respond sensibly to prompts, i.e. instead of providing\nkeywords, a question or initial phrase is provided.  Useful summarization\nbehavior can be achieved including zero shot question answering, or\nenumeration of a research paper's contributions or the purpose of an\ninvention.\n\nWhile the paper feels largely like an extension of Keskar et al. 2019, the\nevaluation of the proposed methods is very thorough on a variety of settings\nand domains. I especially enjoyed the break out of entity targeted summaries\nbased on whether the entity occurred in the lead and/or reference summary. The\nuse of prompts to obtain question answering and more focused\ncontributions/purpose summarization was also very interesting.\n\nI would be happy to see this paper accepted to ICLR. This paper offers a\nsimple method of obtaining a variety of focused or targeted summarization\nbehaviors from a BART summarization model. In general, I would like to see\nmore work like this exploring methods of controlling pretrained language\nmodels.  The evaluation of the correctness of the generated utterances\nsuggests that this method provides fairly reliable control.\n\nThere several areas where the paper could improve. The explanation of how\nlength control is achieved was not very clear. It would help to have examples\nlike those shown in the appendix present in the section introducing length\ncontrol. \n\nComparisons are to the standard BART model or to Fan et al. (2018) which\nsimilarly prepend important control information to the input. It would be\ninteresting to see an evaluation that compared CTRLsum to BART with a\nconstrained decoding method, such as dynamic beam allocation [1].\n\nAdditionally, the authors should say more about the differences in entity\ncontrol of their method and Fan et al. 2018, which seem on their face to be\nsimilar.\n\nDid the authors experiment with pairs of entities as entity controls? It would\nbe especially interesting to see whether the model preserves the correct\nrelationship between entities, especially for entities that didn't occur in\nthe same sentences in the original document, e.g. one important entity and one\nunimportant entity.\n\n[1] Matt Post and David Vilar. Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation. ACL. 2018.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603855368568}, {"id": "5Mc64UkeQFv", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2271/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Paper Summary:\n* This paper proposes a framework for controllable summarization, CTRLsum.  It is different from standard summarization models that CTRLsum uses a set of keywords extracted from the source text automatically or descriptive prompts to control the summary.  Experiments with three domains of summarization datasets and five control aspects.\n\nStrengthes:\n* The authors investigated the effectiveness of the proposed model through extensive experiments.\n\nWeaknesses:\n* The proposed method that uses keywords as an additional input text is almost the same as CIT (Saito et al., 2020), and the scores of CTRLsum on the CNNDM dataset reported in Table 7 does not outperformed those of CIT.  Also, it is not novel to use descriptive prompts to control natural language generation.\n   * Saito et al.: Abstractive Summarization with Combination of Pre-trained Sequence-to-Sequence and Saliency Models. CoRR abs/2003.13028 (2020)\n* I think that the author's claim, \"keywords and prompts are complementary\", is not evaluated fully.\n\nQuestions:\n\n* With respect to contribution summarization, did you evaluate CTRLsum(keyword without prompt) and CTRLsum(prompt without keywords)?  The control tokens \"the main contributions of this paper are : ( 1 )\" is far from the keywords used during training, and so I think that the keywords are not effective for contribution summarization.  In fact, BART that uses prompt worked well for contribution summarization.\n\n* Did you evaluate the ablation tests with respect to the special token \"|\" and keyword dropout?\n\n* Can CTRLsum control the generation with multiple aspects (length and entity control, length and QA control, etc.) simultaneously?  The length of  summaries generated by CTRLsum is strongly dependent the number of keywords, and so I think it is difficult to simultaneously control multiple aspects including length control.\n\nUpdate:\nThank you for the answers to my questions and additional experiments. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "Paper Summary:\n* This paper proposes a framework for controllable summarization, CTRLsum.  It is different from standard summarization models that CTRLsum uses a set of keywords extracted from the source text automatically or descriptive prompts to control the summary.  Experiments with three domains of summarization datasets and five control aspects.\n\nStrengthes:\n* The authors investigated the effectiveness of the proposed model through extensive experiments.\n\nWeaknesses:\n* The proposed method that uses keywords as an additional input text is almost the same as CIT (Saito et al., 2020), and the scores of CTRLsum on the CNNDM dataset reported in Table 7 does not outperformed those of CIT.  Also, it is not novel to use descriptive prompts to control natural language generation.\n   * Saito et al.: Abstractive Summarization with Combination of Pre-trained Sequence-to-Sequence and Saliency Models. CoRR abs/2003.13028 (2020)\n* I think that the author's claim, \"keywords and prompts are complementary\", is not evaluated fully.\n\nQuestions:\n\n* With respect to contribution summarization, did you evaluate CTRLsum(keyword without prompt) and CTRLsum(prompt without keywords)?  The control tokens \"the main contributions of this paper are : ( 1 )\" is far from the keywords used during training, and so I think that the keywords are not effective for contribution summarization.  In fact, BART that uses prompt worked well for contribution summarization.\n\n* Did you evaluate the ablation tests with respect to the special token \"|\" and keyword dropout?\n\n* Can CTRLsum control the generation with multiple aspects (length and entity control, length and QA control, etc.) simultaneously?  The length of  summaries generated by CTRLsum is strongly dependent the number of keywords, and so I think it is difficult to simultaneously control multiple aspects including length control.\n\nUpdate:\nThank you for the answers to my questions and additional experiments. ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603847595467}], "openreview_url": "https://openreview.net/forum?id=ohdw3t-8VCY", "arxiv_id": "2012.04281", "paper_pdf": "papers/ohdw3t-8VCY.pdf", "paper_pdf_sha256": "7df38c89ff91288cef0ac6eb33396b248fb88620209b215048792980dddb3d44", "paper_pdf_bytes": 387961, "paper_pdf_source": "openreview", "code_url": "https://github.com/salesforce/ctrl-sum", "code_repository": "salesforce/ctrl-sum", "code_commit": "01c1a6cf0e286346321829ea88d30c85fbede0ca", "code_archive": "repos/ohdw3t-8VCY.zip", "code_archive_sha256": "32126739d7e095fbd94c9be1414ee068b5334bfc5025cecf6029354994e8d363", "code_archive_bytes": 1831724, "code_file_count": 23, "code_extensions": {".py": 15, ".sh": 8}, "github_disk_usage_kb": 1877, "github_languages": {"Python": 174890, "Shell": 15348}, "github_archived": true, "github_pushed_at": "2025-05-01T17:29:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/ctrlsum-towards-generic-controllable-text-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YBgjDBYPzz", "year": 2026, "status": "rejected", "title": "Top-Theta Attention: Sparsifying Transformers by Compensated Thresholding", "authors": ["Konstantin Berestizshevsky", "Renzo Andri", "Lukas Cavigelli"], "authorids": ["~Konstantin_Berestizshevsky1", "~Renzo_Andri1", "~Lukas_Cavigelli1"], "authors_source": "OpenReview API", "abstract": "We present Top-Theta (Top-θ) Attention, a training-free method for sparsifying transformer attention during inference. Our key insight is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row. This approach enables content-based sparsity without retraining, and it remains robust across data domains. We further introduce compensation techniques to preserve accuracy under aggressive sparsification, establishing attention thresholding as a practical and principled alternative to top-k attention. We provide extensive evaluation on natural language processing tasks, showing that Top-θ achieves 3-10x reduction in $V$-cache usage and up to 10x fewer attention elements during inference while degrading no more than 1% in accuracy.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "mh6ql1EQWF", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7914/Reviewer_PArU"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper proposes Top-Theta Attention, a training-free sparsification method for transformer attention at inference. Key idea: instead of doing per-row top-k search, the authors pre-calibrate a static per-head, per-row threshold so that, on average, each attention row keeps ≈k important entries. At inference, attention scores are just compared to this threshold — a pure elementwise op — so no row-wise dependency, and it works with tiled / distributed kernels. To offset the loss from dropping entries, they add two compensations: Softmax Denominator Compensation (SDC) to better approximate post-softmax sparsity, and V-Mean Compensation (VMC) to add back the average contribution of discarded tokens. On LLaMA-2, LLaMA-3 and LLaMA-3.1 (7B–70B), across ARC-C/E, HellaSwag, HumanEval, and LongBench, they report 3–10× fewer V-rows / attention elements with ≤1% accuracy loss, sometimes even small gains, and the thresholds calibrated on ARC-C transfer to code and long-context tasks, suggesting the thresholds are model-not-data specific. Calibration needs only a few hundred samples and is one-time per model.", "review_text": "The paper proposes Top-Theta Attention, a training-free sparsification method for transformer attention at inference. Key idea: instead of doing per-row top-k search, the authors pre-calibrate a static per-head, per-row threshold so that, on average, each attention row keeps ≈k important entries. At inference, attention scores are just compared to this threshold — a pure elementwise op — so no row-wise dependency, and it works with tiled / distributed kernels. To offset the loss from dropping entries, they add two compensations: Softmax Denominator Compensation (SDC) to better approximate post-softmax sparsity, and V-Mean Compensation (VMC) to add back the average contribution of discarded tokens. On LLaMA-2, LLaMA-3 and LLaMA-3.1 (7B–70B), across ARC-C/E, HellaSwag, HumanEval, and LongBench, they report 3–10× fewer V-rows / attention elements with ≤1% accuracy loss, sometimes even small gains, and the thresholds calibrated on ARC-C transfer to code and long-context tasks, suggesting the thresholds are model-not-data specific. Calibration needs only a few hundred samples and is one-time per model.", "strengths": "1. **Training-free & model-centric.** Calibrate once per model, a few hundred samples, then reuse across domains (ARC-C → HumanEval/LongBench) — this is much cheaper than retrain/fine-tune-based sparsity.\n2. **Tile- and kernel-friendly formulation.** Pure elementwise thresholding; no per-row top-k that breaks tiling. This is exactly where existing top-k attention hurts.\n3. **Strong empirical coverage.** LLaMA-2, LLaMA-3, LLaMA-3.1; 7B→70B; prefill (QA) and decoding (HumanEval, LongBench); GQA case analyzed.\n4. **Storage overhead negligible**. Thresholds are tiny (~12MB for LLaMA-3-70B).", "weaknesses": "1. **No wall-clock / kernel-level evaluation.** The main selling point is “better for tiled / distributed kernels than top-k,” but the paper does not show an actual implementation or runtime vs. FlashAttention+Top-k baselines on GPU.\n2. **Calibration cost vs. model/library changes not fully discussed.** If KV layout or rotary settings change (common in serving stacks), do we need to recalibrate?", "questions": "1. **On SDC choice.** You present three SDC estimators (offline, exp-threshold, exact). For an actual GPU kernel, which one do you expect to be used, and what’s the real compute/memory overhead relative to plain Top-j? Please quantify for LLaMA-3-8B decode (one token).\n2. **GQA union strategy.** You mention two alternatives (discard low-support V-rows; or union across heads). Did you test either? If not, what made you decide on per-head selections only?\n3. **Per-row vs parametric thresholds.** You mention fitting a parametric function to rows to reduce storage. Did you actually try it, and what is the accuracy drop vs. full per-row table?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Top-Theta Attention, a training-free sparsification method for transformer attention at inference. Key idea: instead of doing per-row top-k search, the authors pre-calibrate a static per-head, per-row threshold so that, on average, each attention row keeps ≈k important entries. At inference, attention scores are just compared to this threshold — a pure elementwise op — so no row-wise dependency, and it works with tiled / distributed kernels. To offset the loss from dropping entries, they add two compensations: Softmax Denominator Compensation (SDC) to better approximate post-softmax sparsity, and V-Mean Compensation (VMC) to add back the average contribution of discarded tokens. On LLaMA-2, LLaMA-3 and LLaMA-3.1 (7B–70B), across ARC-C/E, HellaSwag, HumanEval, and LongBench, they report 3–10× fewer V-rows / attention elements with ≤1% accuracy loss, sometimes even small gains, and the thresholds calibrated on ARC-C transfer to code and long-context tasks, suggesting the thresholds are model-not-data specific. Calibration needs only a few hundred samples and is one-time per model.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "1. **Training-free & model-centric.** Calibrate once per model, a few hundred samples, then reuse across domains (ARC-C → HumanEval/LongBench) — this is much cheaper than retrain/fine-tune-based sparsity.\n2. **Tile- and kernel-friendly formulation.** Pure elementwise thresholding; no per-row top-k that breaks tiling. This is exactly where existing top-k attention hurts.\n3. **Strong empirical coverage.** LLaMA-2, LLaMA-3, LLaMA-3.1; 7B→70B; prefill (QA) and decoding (HumanEval, LongBench); GQA case analyzed.\n4. **Storage overhead negligible**. Thresholds are tiny (~12MB for LLaMA-3-70B).", "weaknesses": "1. **No wall-clock / kernel-level evaluation.** The main selling point is “better for tiled / distributed kernels than top-k,” but the paper does not show an actual implementation or runtime vs. FlashAttention+Top-k baselines on GPU.\n2. **Calibration cost vs. model/library changes not fully discussed.** If KV layout or rotary settings change (common in serving stacks), do we need to recalibrate?", "questions": "1. **On SDC choice.** You present three SDC estimators (offline, exp-threshold, exact). For an actual GPU kernel, which one do you expect to be used, and what’s the real compute/memory overhead relative to plain Top-j? Please quantify for LLaMA-3-8B decode (one token).\n2. **GQA union strategy.** You mention two alternatives (discard low-support V-rows; or union across heads). Did you test either? If not, what made you decide on per-head selections only?\n3. **Per-row vs parametric thresholds.** You mention fitting a parametric function to rows to reduce storage. Did you actually try it, and what is the accuracy drop vs. full per-row table?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762091863034}, {"id": "zddF8yEnwY", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7914/Reviewer_WyYY"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 5, "summary": "This paper proposes Top-Theta Attention, a training-free method for sparsifying transformer attention during inference. The key idea is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row.", "review_text": "This paper proposes Top-Theta Attention, a training-free method for sparsifying transformer attention during inference. The key idea is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row.", "strengths": "+ The proposed method does not require training.\n+ The proposed method shows effectiveness on the state-of-the-art open-source LLM models.", "weaknesses": "- Limited novelty. Sparse attention has been extensively studied, and many state-of-the-art methods already exist. The proposed Top-Theta or threshold-based sparsification appears to be a minor variation of the well-known top-k attention, which has been applied both in standard transformer architectures and in large language models. The contribution, therefore, seems incremental rather than fundamentally new.\n\n- Unclear motivation. The paper does not clearly explain why a content-based attention mechanism is necessary or how it addresses specific limitations of existing methods. The proposed approach essentially performs pruning by comparing attention scores against a threshold, but it remains unclear why this thresholding scheme is conceptually or practically superior.\n\n- Unjustified design choices and missing ablations. Several components of the proposed method are introduced without sufficient justification or explanation. Moreover, the paper lacks ablation studies to demonstrate the contribution of each component or to support the design decisions made.\n\n- Insufficient experimental validation. The experimental section lacks comprehensive comparisons with existing sparse-attention methods and does not include evaluations on modern large-scale LLMs. As a result, it is challenging to evaluate the generality and practical effectiveness of the proposed technique.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Top-Theta Attention, a training-free method for sparsifying transformer attention during inference. The key idea is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "+ The proposed method does not require training.\n+ The proposed method shows effectiveness on the state-of-the-art open-source LLM models.", "weaknesses": "- Limited novelty. Sparse attention has been extensively studied, and many state-of-the-art methods already exist. The proposed Top-Theta or threshold-based sparsification appears to be a minor variation of the well-known top-k attention, which has been applied both in standard transformer architectures and in large language models. The contribution, therefore, seems incremental rather than fundamentally new.\n\n- Unclear motivation. The paper does not clearly explain why a content-based attention mechanism is necessary or how it addresses specific limitations of existing methods. The proposed approach essentially performs pruning by comparing attention scores against a threshold, but it remains unclear why this thresholding scheme is conceptually or practically superior.\n\n- Unjustified design choices and missing ablations. Several components of the proposed method are introduced without sufficient justification or explanation. Moreover, the paper lacks ablation studies to demonstrate the contribution of each component or to support the design decisions made.\n\n- Insufficient experimental validation. The experimental section lacks comprehensive comparisons with existing sparse-attention methods and does not include evaluations on modern large-scale LLMs. As a result, it is challenging to evaluate the generality and practical effectiveness of the proposed technique.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761885364762}, {"id": "1lX8jGgcbc", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7914/Reviewer_zNgo"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a novel method called Top-$\\theta$ attention, which can approximate top-$k$ attention computation without using sorting algorithms or other approximation variants. This approach demonstrates a 3- to 10-fold reduction in V-cache usage and up to 10 times fewer attention elements during inference, while degrading accuracy by no more than 1%.", "review_text": "This paper proposes a novel method called Top-$\\theta$ attention, which can approximate top-$k$ attention computation without using sorting algorithms or other approximation variants. This approach demonstrates a 3- to 10-fold reduction in V-cache usage and up to 10 times fewer attention elements during inference, while degrading accuracy by no more than 1%.", "strengths": "* **High Efficiency**: Reduces **V-cache usage by 3–10×** and attention elements by **up to 10×** with <1% accuracy loss.\n\n* **Robust to Domain Shift**: Thresholds are **model-intrinsic** — calibrated once and work across tasks and datasets.\n\n* **Better Than Top-\\(k\\)**: Replaces expensive top-\\(k\\) search with **constant-time thresholding**, removing row-wise dependencies.", "weaknesses": "* This paper lacks a comprehensive review of the field of sparse attention, which has already been popular to study for a long time.\n* This paper lacks a comprehensive experimental comparison with current popular methods; it is not clear how this method outperforms other sparsity-based attention approximation algorithms.\n* The authors present a new method with promising performance in this work, but after checking the whole paper, it is still not clear what this paper wants to solve. If they just provide a new method that approximates attention efficiently without losing too much accuracy, they should do a comprehensive experimental comparison with current popular methods; otherwise, the community cannot tell why this work is better than others and why people should use this method rather than others.\n* This paper lacks a comparison of the efficiency among the top-$\\theta$ attention and other methods, especially when it involves additional inference; the additional computational cost should be considered, too.", "questions": "* Please see Weaknesses\n* Line 124-125 \"We conjecture that the top-k search algorithms....\": Could you explain why this conjecture holds?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel method called Top-$\\theta$ attention, which can approximate top-$k$ attention computation without using sorting algorithms or other approximation variants. This approach demonstrates a 3- to 10-fold reduction in V-cache usage and up to 10 times fewer attention elements during inference, while degrading accuracy by no more than 1%.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* **High Efficiency**: Reduces **V-cache usage by 3–10×** and attention elements by **up to 10×** with <1% accuracy loss.\n\n* **Robust to Domain Shift**: Thresholds are **model-intrinsic** — calibrated once and work across tasks and datasets.\n\n* **Better Than Top-\\(k\\)**: Replaces expensive top-\\(k\\) search with **constant-time thresholding**, removing row-wise dependencies.", "weaknesses": "* This paper lacks a comprehensive review of the field of sparse attention, which has already been popular to study for a long time.\n* This paper lacks a comprehensive experimental comparison with current popular methods; it is not clear how this method outperforms other sparsity-based attention approximation algorithms.\n* The authors present a new method with promising performance in this work, but after checking the whole paper, it is still not clear what this paper wants to solve. If they just provide a new method that approximates attention efficiently without losing too much accuracy, they should do a comprehensive experimental comparison with current popular methods; otherwise, the community cannot tell why this work is better than others and why people should use this method rather than others.\n* This paper lacks a comparison of the efficiency among the top-$\\theta$ attention and other methods, especially when it involves additional inference; the additional computational cost should be considered, too.", "questions": "* Please see Weaknesses\n* Line 124-125 \"We conjecture that the top-k search algorithms....\": Could you explain why this conjecture holds?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761243855913}], "openreview_url": "https://openreview.net/forum?id=YBgjDBYPzz", "arxiv_id": "2502.08363", "paper_pdf": "papers/YBgjDBYPzz.pdf", "paper_pdf_sha256": "26fdc80d8cfeb1fc700da2f9aaffb44a0c5a4f84628ac1acf7ea523970f00ebf", "paper_pdf_bytes": 4913761, "paper_pdf_source": "openreview", "code_url": "https://github.com/huawei-csl/top-theta-attention", "code_repository": "huawei-csl/top-theta-attention", "code_commit": "fa3ab6889137994cd393162ff8236993f3391053", "code_archive": "repos/YBgjDBYPzz.zip", "code_archive_sha256": "0debb1e23d3363604f7f959602e361dffd9739123b68c86f148e94f70266926d", "code_archive_bytes": 5863761, "code_file_count": 25, "code_extensions": {".py": 17, ".ipynb": 6, ".sh": 2}, "github_disk_usage_kb": 7596, "github_languages": {"Jupyter Notebook": 6079920, "Python": 351171, "Shell": 3266}, "github_archived": false, "github_pushed_at": "2026-06-04T11:37:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/top-theta-attention-sparsifying-transformers"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OANUpvmnuf", "year": 2025, "status": "rejected", "title": "Choices are More Important than Efforts: LLM Enables Efficient Multi-Agent Exploration", "authors": ["Yun Qu", "Boyuan Wang", "Yuhang Jiang", "Jianzhun Shao", "Yixiu Mao", "Chang Liu", "Cheems Wang", "Xiangyang Ji"], "authorids": ["~Yun_Qu2", "~Boyuan_Wang1", "~Yuhang_Jiang3", "~Jianzhun_Shao1", "~Yixiu_Mao2", "~Chang_Liu9", "~Cheems_Wang1", "~Xiangyang_Ji1"], "authors_source": "OpenReview API", "abstract": "With expansive state-action spaces, efficient multi-agent exploration remains a longstanding challenge in reinforcement learning.\nAlthough pursuing novelty, diversity, or uncertainty attracts increasing attention, redundant efforts brought by exploration without proper guidance choices poses a practical issue for the community.\nThis paper introduces a systematic approach, termed LEMAE, choosing to channel informative task-relevant guidance from a knowledgeable Large Language Model (LLM) for Efficient Multi-Agent Exploration. \nSpecifically, we ground linguistic knowledge from LLM into symbolic key states, that are critical for task fulfillment, in a discriminative manner at low LLM inference costs. \nTo unleash the power of key states, \nwe design Subspace-based Hindsight Intrinsic Reward (SHIR) to guide agents toward key states by increasing reward density.  Additionally, we build the Key State Memory Tree (KSMT) to track transitions between key states in a specific task for organized exploration. Benefiting from diminishing redundant explorations, LEMAE outperforms existing SOTA approaches on the challenging benchmarks (e.g., SMAC and MPE) by a large margin, achieving a 10x acceleration in certain scenarios.\nOur code is available at https://anonymous.4open.science/r/LEMAE.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "QmFrRBeL4a", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1833/Reviewer_Kmbo"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper study utilizing LLMs to improve exploration for multi-agent RL setting. The key motivation is that LLMs, pre-trained on general knowledge, can extract task-relevant guidance for improving exploration. The proposed method,  LEMAE, identifies key states by prompting LLMs, and encourages the agents to move forward key states using a memory tree. Experiment results on SMAC and MPE show that LEMAE outperforms various RL methods in terms of improvement speed and final performance.", "review_text": "This paper study utilizing LLMs to improve exploration for multi-agent RL setting. The key motivation is that LLMs, pre-trained on general knowledge, can extract task-relevant guidance for improving exploration. The proposed method,  LEMAE, identifies key states by prompting LLMs, and encourages the agents to move forward key states using a memory tree. Experiment results on SMAC and MPE show that LEMAE outperforms various RL methods in terms of improvement speed and final performance.", "strengths": "Developing an effective way to leverage the linguistic knowledge in LLMs for RL is a promising approach. This paper makes a step forward by proposing the idea of 'discriminating key state' and tree search. I believe this could be a valuable supplement to RL community with the integration of large foundation models. \n\nIn addition, I list some other pros as follow:\n\n1. Overall the paper is well-organized and well-established. It is easy to follow the motivation and method design. \n2. The proposed method seems a general method that can be utilized for wide RL problems, e.g., single agent setting. It would be interesting to see the results on single agent RL tasks.\n3. The experiment is comprehensive, including wide ranges of baselines and evaluation tasks.", "weaknesses": "It seems that LEMAE is quite limited to specific tasks, whose state is simple to describe. A key point of LEMAE is to address the symbolic state representation issue for LLMs. However, for tasks with complex state representation, such as robotics control, it is hard to explain a specific state value. LEMAE is hard to be applied in such domain, which is an important sub-filed of RL tasks. \n\nBesides, LEMAE is proposed for multi-agent RL problem, but it lacks a consideration for solve MARL training, e.g., cooperation between partners. Taking this into consideration would make the method more reliable and efficient.\n\nThere are also some more suggestions on the writing:\n1. I feel that the introduction of the method in Section 1 is a simple re-writing of the abstract. \n2. I do not find the importance of putting \"CHOICES ARE MORE IMPORTANT THAN EFFORTS\" in the title. I think it is not the main highlight of this paper. Maybe it is better to consider from the perspective of LLMs.", "questions": "Refer to weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper study utilizing LLMs to improve exploration for multi-agent RL setting. The key motivation is that LLMs, pre-trained on general knowledge, can extract task-relevant guidance for improving exploration. The proposed method,  LEMAE, identifies key states by prompting LLMs, and encourages the agents to move forward key states using a memory tree. Experiment results on SMAC and MPE show that LEMAE outperforms various RL methods in terms of improvement speed and final performance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Developing an effective way to leverage the linguistic knowledge in LLMs for RL is a promising approach. This paper makes a step forward by proposing the idea of 'discriminating key state' and tree search. I believe this could be a valuable supplement to RL community with the integration of large foundation models. \n\nIn addition, I list some other pros as follow:\n\n1. Overall the paper is well-organized and well-established. It is easy to follow the motivation and method design. \n2. The proposed method seems a general method that can be utilized for wide RL problems, e.g., single agent setting. It would be interesting to see the results on single agent RL tasks.\n3. The experiment is comprehensive, including wide ranges of baselines and evaluation tasks.", "weaknesses": "It seems that LEMAE is quite limited to specific tasks, whose state is simple to describe. A key point of LEMAE is to address the symbolic state representation issue for LLMs. However, for tasks with complex state representation, such as robotics control, it is hard to explain a specific state value. LEMAE is hard to be applied in such domain, which is an important sub-filed of RL tasks. \n\nBesides, LEMAE is proposed for multi-agent RL problem, but it lacks a consideration for solve MARL training, e.g., cooperation between partners. Taking this into consideration would make the method more reliable and efficient.\n\nThere are also some more suggestions on the writing:\n1. I feel that the introduction of the method in Section 1 is a simple re-writing of the abstract. \n2. I do not find the importance of putting \"CHOICES ARE MORE IMPORTANT THAN EFFORTS\" in the title. I think it is not the main highlight of this paper. Maybe it is better to consider from the perspective of LLMs.", "questions": "Refer to weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731254519198}, {"id": "yMkCWoVDGu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1833/Reviewer_kuvr"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes to improve exploration in multi-agent environments by generating key states with the aid of LLMs and then use hindsight experience replay to generate auxiliary rewards. Specifically, task descriptions as well as task-relevant knowledge is converted to text and then fed to an LLM that is tasked to generate a function (e.g., written in Python) that takes in a state and returns whether this is a goal state. During a typical online reinforcement learning procedure, each experienced state is examined by the key state checking function, and if it is masked as a key state, hindsight experience replay will be used to add auxiliary rewards to prior states in the same trajectory/episode. The proposed algorithm was tested on the Multiple-Particle Environment and the StarCraft Multi-Agent Challenge.", "review_text": "This paper proposes to improve exploration in multi-agent environments by generating key states with the aid of LLMs and then use hindsight experience replay to generate auxiliary rewards. Specifically, task descriptions as well as task-relevant knowledge is converted to text and then fed to an LLM that is tasked to generate a function (e.g., written in Python) that takes in a state and returns whether this is a goal state. During a typical online reinforcement learning procedure, each experienced state is examined by the key state checking function, and if it is masked as a key state, hindsight experience replay will be used to add auxiliary rewards to prior states in the same trajectory/episode. The proposed algorithm was tested on the Multiple-Particle Environment and the StarCraft Multi-Agent Challenge.", "strengths": "Hindsight experience replay is a very effective auxiliary reward generation methods that have shown successes in many RL environments. However, its main limitation is the difficulty to correctly “identify” goal states. This paper proposes to use LLM to generate such “key states”.", "weaknesses": "- Connection with other LLM-based reward shaping algorithms. While the paper discusses many LLM-based reward generation algorithms in their related works, better and more comprehensive comparison with these methods are needed in the experiment sections to better demonstrate the effectiveness of the proposed algorithm (e.g., Xu et al., 2023 and Liu et al., 2023 cited in the paper). Specifically, as mentioned in the paper, specific prompts for description the task/state space/action space are used for each environment independently. Whether the same set of prompts are used in relevant baselines? If the prompt is good enough for the LLM to generate key state description functions, it seems natural that with similar prompts the LLM can also directly generate good auxiliary rewards and/or reward functions, which are done by many existing LLM-based reward generation algorithms. It would be very helpful to clearly describe the settings and make sure that the relevant baselines are also exposed to the same prompts. \n\n- There is only one LLM-based reward generation baseline adopted. Are there particular reasons for only including this LLM-based baseline algorithm? Since the proposed method leverages knowledge from LLMs, it would be more natural to include more LLM-based baselines.\n\n- Generalizability of the LLM-based key-state detector to other environments. The adopted environments have discrete and small state/action spaces, which makes it easier to create key-state detectors. Also, the environments are relatively simple to describe with natural language. I wonder how this method performs in environments with more complex state space/action space/transition dynamics, for example in robotics tasks?\n\n- Minor: relation with Go-Explore [1]. This key states memory tree seems to be related to go-explore. Could the authors discuss their similarities and differences?\n\n\n[1] Ecoffet, Adrien, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. \"Go-explore: a new approach for hard-exploration problems.\" arXiv preprint arXiv:1901.10995 (2019).", "questions": "- How does the proposed method compare with other LLM-based reward shaping algorithms?\n- Could the authors clarify whether the same prompts were used across all LLM-based methods?\n- How well can the proposed algorithm scale to environments with complex state space/action space/transition dynamics?\n- It seems that the proposed method can be applied to both single-agent and multi-agent environments. Is there a particular reason to use it primarily in multi-agent environments?\n- How does the key states memory tree compare to the memory structure used in Go-Explore?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to improve exploration in multi-agent environments by generating key states with the aid of LLMs and then use hindsight experience replay to generate auxiliary rewards. Specifically, task descriptions as well as task-relevant knowledge is converted to text and then fed to an LLM that is tasked to generate a function (e.g., written in Python) that takes in a state and returns whether this is a goal state. During a typical online reinforcement learning procedure, each experienced state is examined by the key state checking function, and if it is masked as a key state, hindsight experience replay will be used to add auxiliary rewards to prior states in the same trajectory/episode. The proposed algorithm was tested on the Multiple-Particle Environment and the StarCraft Multi-Agent Challenge.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Hindsight experience replay is a very effective auxiliary reward generation methods that have shown successes in many RL environments. However, its main limitation is the difficulty to correctly “identify” goal states. This paper proposes to use LLM to generate such “key states”.", "weaknesses": "- Connection with other LLM-based reward shaping algorithms. While the paper discusses many LLM-based reward generation algorithms in their related works, better and more comprehensive comparison with these methods are needed in the experiment sections to better demonstrate the effectiveness of the proposed algorithm (e.g., Xu et al., 2023 and Liu et al., 2023 cited in the paper). Specifically, as mentioned in the paper, specific prompts for description the task/state space/action space are used for each environment independently. Whether the same set of prompts are used in relevant baselines? If the prompt is good enough for the LLM to generate key state description functions, it seems natural that with similar prompts the LLM can also directly generate good auxiliary rewards and/or reward functions, which are done by many existing LLM-based reward generation algorithms. It would be very helpful to clearly describe the settings and make sure that the relevant baselines are also exposed to the same prompts. \n\n- There is only one LLM-based reward generation baseline adopted. Are there particular reasons for only including this LLM-based baseline algorithm? Since the proposed method leverages knowledge from LLMs, it would be more natural to include more LLM-based baselines.\n\n- Generalizability of the LLM-based key-state detector to other environments. The adopted environments have discrete and small state/action spaces, which makes it easier to create key-state detectors. Also, the environments are relatively simple to describe with natural language. I wonder how this method performs in environments with more complex state space/action space/transition dynamics, for example in robotics tasks?\n\n- Minor: relation with Go-Explore [1]. This key states memory tree seems to be related to go-explore. Could the authors discuss their similarities and differences?\n\n\n[1] Ecoffet, Adrien, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. \"Go-explore: a new approach for hard-exploration problems.\" arXiv preprint arXiv:1901.10995 (2019).", "questions": "- How does the proposed method compare with other LLM-based reward shaping algorithms?\n- Could the authors clarify whether the same prompts were used across all LLM-based methods?\n- How well can the proposed algorithm scale to environments with complex state space/action space/transition dynamics?\n- It seems that the proposed method can be applied to both single-agent and multi-agent environments. Is there a particular reason to use it primarily in multi-agent environments?\n- How does the key states memory tree compare to the memory structure used in Go-Explore?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730731742907}, {"id": "52yegRV8Nq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1833/Reviewer_AW9u"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper proposes to channel informative task-relevant guidance from a knowledgeable LLM for efficient Multi-Agent Exploration. It uses LLM to localize key states and use hindsight reward for exploration (also use memory tree to orginaze the key states). The results show that the proposed methods achieve better results than most \"traditional\" deep MARL methods.", "review_text": "This paper proposes to channel informative task-relevant guidance from a knowledgeable LLM for efficient Multi-Agent Exploration. It uses LLM to localize key states and use hindsight reward for exploration (also use memory tree to orginaze the key states). The results show that the proposed methods achieve better results than most \"traditional\" deep MARL methods.", "strengths": "1. important topic and novel methods: Efficient multi-agent exploration is important for MAS, the proposed methods leverage LLM for this goal, which is novel as the reviewer knows.  \n\n2. memory tree to orginaze the key states for guided exploration (instead of greedy redundant exploration) is interesting and novel.  \n\n3. The paper is clear and easy to understand.\n\n4. The results show that the proposed methods achieve better results than most baseline methods.", "weaknesses": "1. Proposition 4.1 is one of hte most important motivation for task-relevant information based exploration. But the setting is the one-dimensional asymmetric random walk problem. What is means for a more realistic scenarios (e.g., MPE, SMAC)?  Could the authors discuss how the insights from Proposition 4.1 might generalize or relate to more complex environments (higher-dimensional state spaces with multiple agents) like MPE and SMAC?\n\n2. The key states localization is the most important for the proposed methods. But current description (i..e, Section 4.2)  is not clear about the method detaills. (besides discrimination by LLM rather than generation; m discriminator functions). Could the authors discuss the process of generating the m discriminator functions (or the detailed prompts), how the number m is determined, or how the discriminator functions are applied to states in practice. And how to make sure the LLM-based discriminators can work well as the designed intention?\n\n3. The baselines compared seems to be general MARL methods. It is better to compare MA-exploration methods and methods that also combine LLM with MARL (e.g., [1,2,3]).  Could the authors explain why these specific baselines were not included, and could you either add comparisons to these methods or discuss how their approach differs conceptually from these related works. This would help contextualize your contribution within the most relevant recent literature.\n\n[1] Learning When to Explore in Multi-Agent Reinforcement Learning. TCYB 2023.  (which learns to when to explore)\n[2] Controlling Large Language Model-based Agents for Large-Scale Decision-Making: An Actor-Critic Approach. (which uses two LLMs for exploration-exploitation tradeoff)", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to channel informative task-relevant guidance from a knowledgeable LLM for efficient Multi-Agent Exploration. It uses LLM to localize key states and use hindsight reward for exploration (also use memory tree to orginaze the key states). The results show that the proposed methods achieve better results than most \"traditional\" deep MARL methods.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. important topic and novel methods: Efficient multi-agent exploration is important for MAS, the proposed methods leverage LLM for this goal, which is novel as the reviewer knows.  \n\n2. memory tree to orginaze the key states for guided exploration (instead of greedy redundant exploration) is interesting and novel.  \n\n3. The paper is clear and easy to understand.\n\n4. The results show that the proposed methods achieve better results than most baseline methods.", "weaknesses": "1. Proposition 4.1 is one of hte most important motivation for task-relevant information based exploration. But the setting is the one-dimensional asymmetric random walk problem. What is means for a more realistic scenarios (e.g., MPE, SMAC)?  Could the authors discuss how the insights from Proposition 4.1 might generalize or relate to more complex environments (higher-dimensional state spaces with multiple agents) like MPE and SMAC?\n\n2. The key states localization is the most important for the proposed methods. But current description (i..e, Section 4.2)  is not clear about the method detaills. (besides discrimination by LLM rather than generation; m discriminator functions). Could the authors discuss the process of generating the m discriminator functions (or the detailed prompts), how the number m is determined, or how the discriminator functions are applied to states in practice. And how to make sure the LLM-based discriminators can work well as the designed intention?\n\n3. The baselines compared seems to be general MARL methods. It is better to compare MA-exploration methods and methods that also combine LLM with MARL (e.g., [1,2,3]).  Could the authors explain why these specific baselines were not included, and could you either add comparisons to these methods or discuss how their approach differs conceptually from these related works. This would help contextualize your contribution within the most relevant recent literature.\n\n[1] Learning When to Explore in Multi-Agent Reinforcement Learning. TCYB 2023.  (which learns to when to explore)\n[2] Controlling Large Language Model-based Agents for Large-Scale Decision-Making: An Actor-Critic Approach. (which uses two LLMs for exploration-exploitation tradeoff)", "questions": "see weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730084458076}, {"id": "ToLdQjGYJj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission1833/Reviewer_5ctA"], "rating": 1, "soundness": 1, "presentation": 1, "contribution": 1, "confidence": 5, "summary": "This paper presents an approach for using LLMs to identify key states for multi-agent reinforcement learning agents to more efficiently explore in procedurally generated environments. Given a trajectory, the LLM generates key states, which are then organized in a key states memory tree, and used in a hindsight experience replay-like approach.", "review_text": "This paper presents an approach for using LLMs to identify key states for multi-agent reinforcement learning agents to more efficiently explore in procedurally generated environments. Given a trajectory, the LLM generates key states, which are then organized in a key states memory tree, and used in a hindsight experience replay-like approach.", "strengths": "- I believe the paper provides sufficient detail in the appendix to reproduce it.\n- The direction of the paper, addressing hard exploration in multi-agent systems, is an important one.", "weaknesses": "## Major\nA) The main results of the paper center around The StarCraft Multi-Agent Challenge (SMACv1) by Samvelyan et al. (2019). However, it is well established by now that SMACv1 is saturated and too simple. The authors should use the 2022 released SMACv2 [1] for their experiments. From the SMACv2 paper: “after years of sustained improvement on SMAC, algorithms now achieve near-perfect performance. In this work, we conduct new analysis demonstrating that SMAC lacks the stochasticity and partial observability to require complex *closed-loop* policies. In particular, we show that an *open-loop* policy conditioned only on the timestep can achieve non-trivial win rates for many SMAC scenarios.”\n\nB) Using language observations or LLMs to guide exploration of RL agents for hard exploration problems has been extensively studied in the single-agent setup. The authors have not covered important very related methods (e.g. [2-4]) in this space. In particular I would expect Motif [3] and OMNI [4] to be tried out as a baseline in the multi-agent RL (MARL) setting. In fact, I would go as far and say that the proposed method is not MARL-specific at all, so the authors really need to show that it provides benefits over existing exploration methods in the single agent setting.\n\nC) Related to B, it is even unclear to me whether explicit key states are necessary for exploration in the MARL environments used here. I would like to see a comparison to MARL with general intrinsic reward mechanisms such as E3B [5] to answer the question whether key states significantly improve exploration over generic intrinsic reward methods.\n\nD) The authors claim that “KSMT helps avoid redundant exploration throughout the state space, particularly beneficial in more complicated real-world scenarios” despite that no experiments on such real-world problems are provided. Either tone it down or provide additional empirical evidence in more-real world environments. I would be very interested in the latter, in particular seeing experiments on environments that do not have a symbolic observation space.\n\n[1] Ellis, B., Moalla, S., Samvelyan, M., Sun, M., Mahajan, A., Foerster, J. N., & Whiteson, S. (2022). SMACv2: An Improved Benchmark for Cooperative Multi-Agent Reinforcement Learning. arXiv. https://doi.org/10.48550/arXiv.2212.07489\n\n[2] Mu, J., Zhong, V., Raileanu, R., Jiang, M., Goodman, N., Rocktäschel, T., & Grefenstette, E. (2022). Improving Intrinsic Exploration with Language Abstractions. arXiv. https://doi.org/10.48550/arXiv.2202.08938\n\n[3] Klissarov, M., D’Oro, P., Sodhani, S., Raileanu, R., Bacon, P.-L., Vincent, P., … Henaff, M. (2023). Motif: Intrinsic Motivation from Artificial Intelligence Feedback. arXiv. https://doi.org/10.48550/arXiv.2310.00166\n\n[4] Zhang, J., Lehman, J., Stanley, K., & Clune, J. (2023). OMNI: Open-endedness via Models of human Notions of Interestingness. ICLR 2024.\n\n[5] Henaff, M., Raileanu, R., Jiang, M., & Rocktäschel, T. (2023). Exploration via Elliptical Episodic Bonuses. arXiv. https://doi.org/10.48550/arXiv.2210.05805\n\n\n## Minor\n- LLMs instead of LLM (e.g. in the title); also \"Efforts\" -> \"Effort\" in the title\n- Be more precise with what you mean by “redundant exploration”\n- Use more formal language (e.g. not “brand-new” in line 089)\n- L160: “to clarity” -> “to clarify”\n- Figure 2 is somewhat unclear to me, in particular the LLM-mapping from Prompt to Response. I particularly find it hard to map what’s formalized in 4.2 (e.g. L209 to L212) to what’s shown in Figure 2. Also, what feedback flows directly back from response to prompt?", "questions": "- L190: What do you mean by “the initial policy is asymmetric”?\n- L247: What’s the rationale for using Manhattan Distance instead of, for example, cosine similarity?\n\nWhat the authors would have to demonstrate to see an improved rating from me:\n- Redo experiments on SMACv2 (A)\n- Add baselines Motif and OMNI, requiring them to be evaluated on MARL (B)\n- Add E3B baseline, again integrating that into the MARL setup, or, alternatively (since the proposed method is not MARL-specific) run the proposed method on single-agent procedurally generated environments. (C) \n- Demonstrate that the proposed method also works for non-symbolic observations (D)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an approach for using LLMs to identify key states for multi-agent reinforcement learning agents to more efficiently explore in procedurally generated environments. Given a trajectory, the LLM generates key states, which are then organized in a key states memory tree, and used in a hindsight experience replay-like approach.", "soundness": 1, "presentation": 1, "contribution": 1, "strengths": "- I believe the paper provides sufficient detail in the appendix to reproduce it.\n- The direction of the paper, addressing hard exploration in multi-agent systems, is an important one.", "weaknesses": "## Major\nA) The main results of the paper center around The StarCraft Multi-Agent Challenge (SMACv1) by Samvelyan et al. (2019). However, it is well established by now that SMACv1 is saturated and too simple. The authors should use the 2022 released SMACv2 [1] for their experiments. From the SMACv2 paper: “after years of sustained improvement on SMAC, algorithms now achieve near-perfect performance. In this work, we conduct new analysis demonstrating that SMAC lacks the stochasticity and partial observability to require complex *closed-loop* policies. In particular, we show that an *open-loop* policy conditioned only on the timestep can achieve non-trivial win rates for many SMAC scenarios.”\n\nB) Using language observations or LLMs to guide exploration of RL agents for hard exploration problems has been extensively studied in the single-agent setup. The authors have not covered important very related methods (e.g. [2-4]) in this space. In particular I would expect Motif [3] and OMNI [4] to be tried out as a baseline in the multi-agent RL (MARL) setting. In fact, I would go as far and say that the proposed method is not MARL-specific at all, so the authors really need to show that it provides benefits over existing exploration methods in the single agent setting.\n\nC) Related to B, it is even unclear to me whether explicit key states are necessary for exploration in the MARL environments used here. I would like to see a comparison to MARL with general intrinsic reward mechanisms such as E3B [5] to answer the question whether key states significantly improve exploration over generic intrinsic reward methods.\n\nD) The authors claim that “KSMT helps avoid redundant exploration throughout the state space, particularly beneficial in more complicated real-world scenarios” despite that no experiments on such real-world problems are provided. Either tone it down or provide additional empirical evidence in more-real world environments. I would be very interested in the latter, in particular seeing experiments on environments that do not have a symbolic observation space.\n\n[1] Ellis, B., Moalla, S., Samvelyan, M., Sun, M., Mahajan, A., Foerster, J. N., & Whiteson, S. (2022). SMACv2: An Improved Benchmark for Cooperative Multi-Agent Reinforcement Learning. arXiv. https://doi.org/10.48550/arXiv.2212.07489\n\n[2] Mu, J., Zhong, V., Raileanu, R., Jiang, M., Goodman, N., Rocktäschel, T., & Grefenstette, E. (2022). Improving Intrinsic Exploration with Language Abstractions. arXiv. https://doi.org/10.48550/arXiv.2202.08938\n\n[3] Klissarov, M., D’Oro, P., Sodhani, S., Raileanu, R., Bacon, P.-L., Vincent, P., … Henaff, M. (2023). Motif: Intrinsic Motivation from Artificial Intelligence Feedback. arXiv. https://doi.org/10.48550/arXiv.2310.00166\n\n[4] Zhang, J., Lehman, J., Stanley, K., & Clune, J. (2023). OMNI: Open-endedness via Models of human Notions of Interestingness. ICLR 2024.\n\n[5] Henaff, M., Raileanu, R., Jiang, M., & Rocktäschel, T. (2023). Exploration via Elliptical Episodic Bonuses. arXiv. https://doi.org/10.48550/arXiv.2210.05805\n\n\n## Minor\n- LLMs instead of LLM (e.g. in the title); also \"Efforts\" -> \"Effort\" in the title\n- Be more precise with what you mean by “redundant exploration”\n- Use more formal language (e.g. not “brand-new” in line 089)\n- L160: “to clarity” -> “to clarify”\n- Figure 2 is somewhat unclear to me, in particular the LLM-mapping from Prompt to Response. I particularly find it hard to map what’s formalized in 4.2 (e.g. L209 to L212) to what’s shown in Figure 2. Also, what feedback flows directly back from response to prompt?", "questions": "- L190: What do you mean by “the initial policy is asymmetric”?\n- L247: What’s the rationale for using Manhattan Distance instead of, for example, cosine similarity?\n\nWhat the authors would have to demonstrate to see an improved rating from me:\n- Redo experiments on SMACv2 (A)\n- Add baselines Motif and OMNI, requiring them to be evaluated on MARL (B)\n- Add E3B baseline, again integrating that into the MARL setup, or, alternatively (since the proposed method is not MARL-specific) run the proposed method on single-agent procedurally generated environments. (C) \n- Demonstrate that the proposed method also works for non-symbolic observations (D)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1729253818357}], "openreview_url": "https://openreview.net/forum?id=OANUpvmnuf", "arxiv_id": "2410.02511", "paper_pdf": "papers/OANUpvmnuf.pdf", "paper_pdf_sha256": "cc3a22bb34cd614ab808aa296535ef3e270687752fd0c153436cad4d978bea5b", "paper_pdf_bytes": 6752884, "paper_pdf_source": "openreview", "code_url": "https://github.com/hijkzzz/pymarl2", "code_repository": "hijkzzz/pymarl2", "code_commit": "8ccac7c5aa134422a2e3009be735d23cdf8ce2f8", "code_archive": "repos/OANUpvmnuf.zip", "code_archive_sha256": "240e13e2e93cc3360a458d7589b872e6143e2424c3160892b3bcb2288ac8057a", "code_archive_bytes": 350311, "code_file_count": 81, "code_extensions": {".py": 76, ".sh": 5}, "github_disk_usage_kb": 373, "github_languages": {"Python": 374490, "Shell": 3103}, "github_archived": false, "github_pushed_at": "2024-05-18T03:22:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/choices-are-more-important-than-efforts-llm"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "q38SZkUmUh", "year": 2024, "status": "rejected", "title": "FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation", "authors": ["Tu Vu", "Mohit Iyyer", "Xuezhi Wang", "Noah Constant", "Jerry Wei", "Jason Wei", "Chris Tar", "Yun-Hsuan Sung", "Denny Zhou", "Quoc V Le", "Thang Luong"], "authorids": ["~Tu_Vu1", "~Mohit_Iyyer1", "~Xuezhi_Wang3", "~Noah_Constant1", "~Jerry_Wei1", "~Jason_Wei1", "ctar@google.com", "~Yun-Hsuan_Sung1", "~Denny_Zhou1", "~Quoc_V_Le1", "~Thang_Luong1"], "authors_source": "OpenReview API", "abstract": "Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world.  In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt.  Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers.  Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future research, we will release FreshQA after blind review and commit to updating it at regular intervals.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "hIZx0cmBWo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8494/Reviewer_BRXn"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses the challenge of large language models (LLMs) not being updated with current information, leading to inaccuracies in their responses. The authors introduce \"FreshQA\", a dynamic QA benchmark designed to test the factuality of LLM-generated answers, especially for questions requiring up-to-date knowledge or debunking false premises. Through extensive human evaluations, they highlight the limitations of current LLMs in addressing fast-changing and false-premise questions. To combat this, they propose FreshPrompt, an in-context learning method that augments LLMs with up-to-date information from search engines, significantly enhancing their factuality.", "review_text": "The paper addresses the challenge of large language models (LLMs) not being updated with current information, leading to inaccuracies in their responses. The authors introduce \"FreshQA\", a dynamic QA benchmark designed to test the factuality of LLM-generated answers, especially for questions requiring up-to-date knowledge or debunking false premises. Through extensive human evaluations, they highlight the limitations of current LLMs in addressing fast-changing and false-premise questions. To combat this, they propose FreshPrompt, an in-context learning method that augments LLMs with up-to-date information from search engines, significantly enhancing their factuality.", "strengths": "1. Tackles a paramount limitation of LLMs – their reliance on outdated or erroneous knowledge.\n\n2. Unveils a dynamic benchmark, FreshQA, capable of evolving over time, which stands as a potent tool for continuous evaluations.\n\n3. Implements a thorough evaluation procedure to gauge both the accuracy and potential hallucination in LLM responses.", "weaknesses": "1. The evolving nature of FreshQA could pose challenges for researchers aiming for consistent benchmarks over varied time frames.\n\n2. The FreshQA dataset bears similarities with RealTimeQA and TimeQA, which somewhat dilutes the novelty of this work, although it remains a complementary addition.\n\n3. The proposed method isn't entirely groundbreaking, given precedents like internet-augmented LLM [1] and REPLUG [2].\n[1] Internet-augmented language models through few-shot prompting for open-domain question answering\n[2] Replug: Retrieval-augmented black-box language models.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of large language models (LLMs) not being updated with current information, leading to inaccuracies in their responses. The authors introduce \"FreshQA\", a dynamic QA benchmark designed to test the factuality of LLM-generated answers, especially for questions requiring up-to-date knowledge or debunking false premises. Through extensive human evaluations, they highlight the limitations of current LLMs in addressing fast-changing and false-premise questions. To combat this, they propose FreshPrompt, an in-context learning method that augments LLMs with up-to-date information from search engines, significantly enhancing their factuality.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Tackles a paramount limitation of LLMs – their reliance on outdated or erroneous knowledge.\n\n2. Unveils a dynamic benchmark, FreshQA, capable of evolving over time, which stands as a potent tool for continuous evaluations.\n\n3. Implements a thorough evaluation procedure to gauge both the accuracy and potential hallucination in LLM responses.", "weaknesses": "1. The evolving nature of FreshQA could pose challenges for researchers aiming for consistent benchmarks over varied time frames.\n\n2. The FreshQA dataset bears similarities with RealTimeQA and TimeQA, which somewhat dilutes the novelty of this work, although it remains a complementary addition.\n\n3. The proposed method isn't entirely groundbreaking, given precedents like internet-augmented LLM [1] and REPLUG [2].\n[1] Internet-augmented language models through few-shot prompting for open-domain question answering\n[2] Replug: Retrieval-augmented black-box language models.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698883849746}, {"id": "6J2BbTDQaV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8494/Reviewer_4BtD"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper analyzes how Large Language Models (LLMs) perform when queried with questions involving fast-changing knowledge or questions requiring the debunking of false premises. To this end, they introduce a new high-quality Q&A dataset, FreshQA, comprising 600 questions. They evaluate the performance of multiple LLMs under two regimes: STRICT, in which an answer is correct if and only if the reasoning and the answer are correct, and RELAXED, in which it is sufficient for the answer to the question to be correct. This sheds light on whether the models can arrive at the correct answer while hallucinating some facts about the world. To improve performance, the authors devise a method to incorporate search results from Google into the context of the LLM. The conclusion is that LLMs struggle with fast-changing and false premise questions (although GPT-4 seems to perform quite well on these types of questions if the knowledge is before its cut-off date), and retrieval-augmented generation improves the performance of LLMs.", "review_text": "The paper analyzes how Large Language Models (LLMs) perform when queried with questions involving fast-changing knowledge or questions requiring the debunking of false premises. To this end, they introduce a new high-quality Q&A dataset, FreshQA, comprising 600 questions. They evaluate the performance of multiple LLMs under two regimes: STRICT, in which an answer is correct if and only if the reasoning and the answer are correct, and RELAXED, in which it is sufficient for the answer to the question to be correct. This sheds light on whether the models can arrive at the correct answer while hallucinating some facts about the world. To improve performance, the authors devise a method to incorporate search results from Google into the context of the LLM. The conclusion is that LLMs struggle with fast-changing and false premise questions (although GPT-4 seems to perform quite well on these types of questions if the knowledge is before its cut-off date), and retrieval-augmented generation improves the performance of LLMs.", "strengths": "- Well-written and easy to read\n- Provides a high-quality dataset of 600 questions, along with useful question taxonomy (required knowledge dates, is multi-hop, has false premise, etc.)", "weaknesses": "- Dataset and paper lacks any type of automatic evaluation metric, everything is based on human eval (major)\n- Findings are not particularly novel or unexpected (minor)\n- Methodology for retrieval seem to be tailored to Google Search (as the authors acknowledge, minor)", "questions": "Thanks to the authors for the thoroughly executed paper. Although I think that the results are not particularly unexpected or novel (e.g. LLMs struggle with multi-hop questions as noted in Self-Ask, or questions that involve fresh, novel and changing knowledge, and that retrieval from a search engine helps), I think the main scientific contribution of this paper might be considered the high-quality dataset of questions.\n\nThis leads us to the main weakness of the paper. If I understand correctly, the evaluation in this paper is performed exclusively by human annotation. While I appreciate this substantial evaluation effort, it raises some concerns, especially given that one of the authors' goals is to provide this dataset to the community to facilitate further research in this field.\n\nHow many annotators were used to perform each system's evaluation in this paper? Inter-annotator agreement between two evaluators on a subset of 100 questions is reported in the appendix, and it is shown that the agreement is high. Do the authors expect everyone in the research community using the dataset to follow the same evaluation protocol to assess their own systems?\n\nThese questions leave me wondering whether the authors could have created an automatic evaluation metric that might make the dataset more broadly useful to the community and less subject to annotator variance.\n\nFor example, by carefully crafting the questions with multiple-choice answers (with carefully selected hard negative options that could elicit some of the capabilities the authors are examining, e.g., integrating options with incorrect reasoning but correct answers). Multiple-choice is a strategy already used in well-known QA benchmarks such as MMLU or the suite of the original 11 T0 held-out tasks and lends itself nicely to computing accuracy. Can the authors elaborate on why they didn't choose this option?\n\nOverall, I appreciate the effort put into this paper, but I am a bit unsure about the broad usability of such dataset in the current state. I hope the authors can help me clarify these doubts and, if so, I will be more than willing to increase my score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper analyzes how Large Language Models (LLMs) perform when queried with questions involving fast-changing knowledge or questions requiring the debunking of false premises. To this end, they introduce a new high-quality Q&A dataset, FreshQA, comprising 600 questions. They evaluate the performance of multiple LLMs under two regimes: STRICT, in which an answer is correct if and only if the reasoning and the answer are correct, and RELAXED, in which it is sufficient for the answer to the question to be correct. This sheds light on whether the models can arrive at the correct answer while hallucinating some facts about the world. To improve performance, the authors devise a method to incorporate search results from Google into the context of the LLM. The conclusion is that LLMs struggle with fast-changing and false premise questions (although GPT-4 seems to perform quite well on these types of questions if the knowledge is before its cut-off date), and retrieval-augmented generation improves the performance of LLMs.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Well-written and easy to read\n- Provides a high-quality dataset of 600 questions, along with useful question taxonomy (required knowledge dates, is multi-hop, has false premise, etc.)", "weaknesses": "- Dataset and paper lacks any type of automatic evaluation metric, everything is based on human eval (major)\n- Findings are not particularly novel or unexpected (minor)\n- Methodology for retrieval seem to be tailored to Google Search (as the authors acknowledge, minor)", "questions": "Thanks to the authors for the thoroughly executed paper. Although I think that the results are not particularly unexpected or novel (e.g. LLMs struggle with multi-hop questions as noted in Self-Ask, or questions that involve fresh, novel and changing knowledge, and that retrieval from a search engine helps), I think the main scientific contribution of this paper might be considered the high-quality dataset of questions.\n\nThis leads us to the main weakness of the paper. If I understand correctly, the evaluation in this paper is performed exclusively by human annotation. While I appreciate this substantial evaluation effort, it raises some concerns, especially given that one of the authors' goals is to provide this dataset to the community to facilitate further research in this field.\n\nHow many annotators were used to perform each system's evaluation in this paper? Inter-annotator agreement between two evaluators on a subset of 100 questions is reported in the appendix, and it is shown that the agreement is high. Do the authors expect everyone in the research community using the dataset to follow the same evaluation protocol to assess their own systems?\n\nThese questions leave me wondering whether the authors could have created an automatic evaluation metric that might make the dataset more broadly useful to the community and less subject to annotator variance.\n\nFor example, by carefully crafting the questions with multiple-choice answers (with carefully selected hard negative options that could elicit some of the capabilities the authors are examining, e.g., integrating options with incorrect reasoning but correct answers). Multiple-choice is a strategy already used in well-known QA benchmarks such as MMLU or the suite of the original 11 T0 held-out tasks and lends itself nicely to computing accuracy. Can the authors elaborate on why they didn't choose this option?\n\nOverall, I appreciate the effort put into this paper, but I am a bit unsure about the broad usability of such dataset in the current state. I hope the authors can help me clarify these doubts and, if so, I will be more than willing to increase my score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698611130708}, {"id": "J9WdTnLN2u", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8494/Reviewer_NG2g"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this work the authors conduct detailed study over the factuality and hallucination of text generated by Large Language Models (LLMs) and propose a question answering benchmark FRESHQA which mainly focuses on questions requiring fast-changing world knowledge. Moreover, the authors also present FRESHPROMPT, a few-shot prompting method aiming to boost the performance of LLMs in acquiring up-to-date knowledge retrieved from search engines. Experiments on several LLMs show FRESHPROMPT outperforms other search engine-augmented prompting methods.", "review_text": "In this work the authors conduct detailed study over the factuality and hallucination of text generated by Large Language Models (LLMs) and propose a question answering benchmark FRESHQA which mainly focuses on questions requiring fast-changing world knowledge. Moreover, the authors also present FRESHPROMPT, a few-shot prompting method aiming to boost the performance of LLMs in acquiring up-to-date knowledge retrieved from search engines. Experiments on several LLMs show FRESHPROMPT outperforms other search engine-augmented prompting methods.", "strengths": "1.\tA novel QA benchmark with 600 questions divided into four main categories: never-changing questions, slow-changing questions, fast-changing questions and false-premise questions is proposed for testing factuality of LLMs. Both questions are generated by nlp researchers and online freelancers and cover a wide range of topics. The authors also commit to update the dataset to get up-to-date knowledge. This benchmark is likely to benefit LLM research a lot.\n2.\tThe paper is well written with clear motivations. A QA benchmark for LLMs is first presented and analysis of different models’ score follows. Then a search engine-based prompting method FRESHPROMPT is proposed to alleviate the problem of factuality, with \n3.\tGood empirical results. FRESHPROMPT outperforms PERPLEXITY.AI and SELF-ASK on GPT-3.5 and GPT-4. The authors also conduct detailed ablation studies to analyze the results from different perspective.", "weaknesses": "1.\tThe authors conduct experiments on T5, PaLM and GPT series LLMs and show the influence of parameter size on benchmark score. However, I think more experiments on different famous LLMs like LLaMA, Falcon, etc are needed as benchmark baselines. \n2.\tFor better visualization, the best results in Table 1 need to be displayed in bold.", "questions": "na", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work the authors conduct detailed study over the factuality and hallucination of text generated by Large Language Models (LLMs) and propose a question answering benchmark FRESHQA which mainly focuses on questions requiring fast-changing world knowledge. Moreover, the authors also present FRESHPROMPT, a few-shot prompting method aiming to boost the performance of LLMs in acquiring up-to-date knowledge retrieved from search engines. Experiments on several LLMs show FRESHPROMPT outperforms other search engine-augmented prompting methods.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1.\tA novel QA benchmark with 600 questions divided into four main categories: never-changing questions, slow-changing questions, fast-changing questions and false-premise questions is proposed for testing factuality of LLMs. Both questions are generated by nlp researchers and online freelancers and cover a wide range of topics. The authors also commit to update the dataset to get up-to-date knowledge. This benchmark is likely to benefit LLM research a lot.\n2.\tThe paper is well written with clear motivations. A QA benchmark for LLMs is first presented and analysis of different models’ score follows. Then a search engine-based prompting method FRESHPROMPT is proposed to alleviate the problem of factuality, with \n3.\tGood empirical results. FRESHPROMPT outperforms PERPLEXITY.AI and SELF-ASK on GPT-3.5 and GPT-4. The authors also conduct detailed ablation studies to analyze the results from different perspective.", "weaknesses": "1.\tThe authors conduct experiments on T5, PaLM and GPT series LLMs and show the influence of parameter size on benchmark score. However, I think more experiments on different famous LLMs like LLaMA, Falcon, etc are needed as benchmark baselines. \n2.\tFor better visualization, the best results in Table 1 need to be displayed in bold.", "questions": "na", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698457033210}], "openreview_url": "https://openreview.net/forum?id=q38SZkUmUh", "arxiv_id": "2310.03214", "paper_pdf": "papers/q38SZkUmUh.pdf", "paper_pdf_sha256": "65c3af0884306b6f5d750991b09bb4f6c1b6f6846917abf32af4009193544643", "paper_pdf_bytes": 2172992, "paper_pdf_source": "openreview", "code_url": "https://github.com/freshllms/freshqa", "code_repository": "freshllms/freshqa", "code_commit": "7d2d3683991916f3633e480548a6aa5c9a62e3db", "code_archive": "repos/q38SZkUmUh.zip", "code_archive_sha256": "7eb5ace9d7d9af0a39c1b3e4cf3e90c7f87bee1c39c1b1e61543e54c7ec03bee", "code_archive_bytes": 39840, "code_file_count": 3, "code_extensions": {".ipynb": 3}, "github_disk_usage_kb": 343, "github_languages": {"Jupyter Notebook": 146459}, "github_archived": false, "github_pushed_at": "2026-05-01T22:43:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/freshllms-refreshing-large-language-models"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qhplAU1BOZW", "year": 2023, "status": "rejected", "title": "Lottery Aware Sparsity Hunting: Enabling Federated Learning on Resource-Limited Edge", "authors": ["Sara Babakniya", "Souvik Kundu", "Saurav Prakash", "Yue Niu", "Salman Avestimehr"], "authorids": ["~Sara_Babakniya1", "~Souvik_Kundu2", "~Saurav_Prakash1", "~Yue_Niu1", "~Salman_Avestimehr1"], "authors_source": "OpenReview API", "abstract": "Limited computation and communication capabilities of clients pose significant challenges in federated learning (FL) over resource-limited edge nodes. A potential solution to this problem is to deploy off-the-shelf sparse learning algorithms that train a binary sparse mask on each client with the expectation of training a consistent sparse server mask yielding sparse weight tensors. However, as we investigate in this paper, such naive deployments result in a significant drop in accuracy compared to FL with dense models, especially for clients with limited resource budgets. In particular, our investigations reveal a serious lack of consensus among the trained sparsity masks on clients, which prevents convergence for the server mask and potentially leads to a substantial drop in model performance. Based on such key observations, we propose federated lottery aware sparsity hunting (FLASH), a unified sparse learning framework to make the server win a lottery in terms of yielding a sparse sub-model, able to maintain classification performance under highly resource-limited client settings. Moreover, to support FL on different devices requiring different parameter density, we leverage our findings to present hetero-FLASH, where clients can have different target sparsity budgets based on their device resource limits. Experimental evaluations with multiple models on various datasets (both IID and non-IID) show superiority of our models in closing the gap with unpruned baseline while yielding up to ∼10.1% improved accuracy with ∼10.26x fewer communication costs, compared to existing alternatives, at similar hyperparameter settings.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ziKgGWlwmgN", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper826/Reviewer_womR"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper tries to improve the sparse network training efficiency in a Federated learning framework in two folds: 1) bridging the gap between the sparse network and its dense counterpart (e.g., FedAvg) and 2) saving the communication cost between clients and the server.   To this end, the proposed method performs a two-stage sparse training, including the mask-sensitive evaluation (for mask initialization) and networking training, and investigates two different mask learning strategies -- 1) fixed masks and 2) jointly training over masks and weights. Experiments on three public datasets were provided in terms of both IID and non-IID settings.   ", "review_text": "Overall, the paper provides a technically sound method to solve several practical challenges in incorporating sparse network training into a Federated learning framework. However, please refer to my concerns on methodology, experiment, and network depth in the above sections.  ", "strengths": "Pros:\n- The proposed hetero-FLASH method is interesting and well adapted to the Federated learning setting over diverse edge resources. While the sampling and fusion strategy is straightforward, the proposed solution poses a good baseline for exploring dynamic budgets in a Federated learning framework. \n- It is technically sound to bridge the gap between sparse network training and its dense counterpart by developing a two-stage method to evaluate mask sensitivity in advance. \n- Two mask learning strategies, namely SPDST and JMWST, were designed and developed on three datasets under two different settings. \n\nCons:\n- While several observations were discussed and provided with empirical evidence, it somewhat lacks in-depth or theoretical analysis regarding the reasons behind these observations. Also, all the observations were provided on the same model architecture (ResNet18), which might lead to model bias in the conclusions. \n- The proposed stage 1 lies in the key contribution of this work. Thus, some ablated models like SPDST w/o state 1, and JMWST w/o state 1 are expected to be provided in the experiment. Plus, it is unclear to me how the proposed method encourages mask convergence -- by individually using SPDST or JMWST, or by dynamically switching between these two methods?\n- The experiment results are less convincing due to the lack of baselines (e.g., Huang et al. (2022) and Bibikar et al. (2021)) and ablated models. Although mask convergence has been well discussed, training convergence and parameter analysis on mask initialization are expected. It also remains unclear if the proposed method can be applied in a deeper model, such as ResNet34, ResNet50, and ResNet101", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper tries to improve the sparse network training efficiency in a Federated learning framework in two folds: 1) bridging the gap between the sparse network and its dense counterpart (e.g., FedAvg) and 2) saving the communication cost between clients and the server.   To this end, the proposed method performs a two-stage sparse training, including the mask-sensitive evaluation (for mask initialization) and networking training, and investigates two different mask learning strategies -- 1) fixed masks and 2) jointly training over masks and weights. Experiments on three public datasets were provided in terms of both IID and non-IID settings.   ", "strength_and_weaknesses": "Pros:\n- The proposed hetero-FLASH method is interesting and well adapted to the Federated learning setting over diverse edge resources. While the sampling and fusion strategy is straightforward, the proposed solution poses a good baseline for exploring dynamic budgets in a Federated learning framework. \n- It is technically sound to bridge the gap between sparse network training and its dense counterpart by developing a two-stage method to evaluate mask sensitivity in advance. \n- Two mask learning strategies, namely SPDST and JMWST, were designed and developed on three datasets under two different settings. \n\nCons:\n- While several observations were discussed and provided with empirical evidence, it somewhat lacks in-depth or theoretical analysis regarding the reasons behind these observations. Also, all the observations were provided on the same model architecture (ResNet18), which might lead to model bias in the conclusions. \n- The proposed stage 1 lies in the key contribution of this work. Thus, some ablated models like SPDST w/o state 1, and JMWST w/o state 1 are expected to be provided in the experiment. Plus, it is unclear to me how the proposed method encourages mask convergence -- by individually using SPDST or JMWST, or by dynamically switching between these two methods?\n- The experiment results are less convincing due to the lack of baselines (e.g., Huang et al. (2022) and Bibikar et al. (2021)) and ablated models. Although mask convergence has been well discussed, training convergence and parameter analysis on mask initialization are expected. It also remains unclear if the proposed method can be applied in a deeper model, such as ResNet34, ResNet50, and ResNet101", "clarity,_quality,_novelty_and_reproducibility": "*Clarity*: The methodology in the paper is somewhat hard to follow due to 1) the missing preliminary knowledge for specifying the Federated learning setting, 2) the lack of an overall framework or formal formulation of the proposed method, and 3) the mixup between algorithm lines and descriptions. \n\n*Novelty*: The proposed method seems novel to me owing to its two-stage sparse mask training method and the practical heterogeneous device budget setting. However, it remains unclear to me what is the key contribution of this work to address the observed limitations. The main technical concern is whether the proposed method can be used in very deep neural networks (like over 100 layers and the recent transformer-based architecture). ", "summary_of_the_review": "Overall, the paper provides a technically sound method to solve several practical challenges in incorporating sparse network training into a Federated learning framework. However, please refer to my concerns on methodology, experiment, and network depth in the above sections.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667185492843}, {"id": "nILSMC57Ju", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper826/Reviewer_vopV"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper targets sparse training in the federated learning domain. The paper starts with multiple observations (on small models) about how disagreement of clients on the sparsity mask pattern could significantly reduce the model accuracy, compared to a centralized training. It also shows that this disagreement on the sparsity mask pattern diverges as the training progresses. Based on these observations, the authors propose a two-stage sparse federated learning method with two main purposes; (1) decoupling the identification of initial sparse mask from training and (2) collaboratively learning sparse mask and model weights. Finally, the paper extends the method for a heterogeneous environment (which I think may not be aligned with the main message of the paper) where each client has a different resource budget for which different sparsity mask is required.", "review_text": "I think the paper is well-motivated with clear description of the observations. The authors went deep into a better understanding of the root cause of low model accuracy in a federated learning setting. While I agree with the authors that this disagreement between sparsity mask patterns *could* be one of the reasons for low accuracy, but I am not convinced that this is the sole reason for such performance degradation. In addition, as mentioned in the questions, it would be great if the authors could provide additional results for higher density ratio to better understand whether such problem exists in a more realistic setting.", "strengths": "**Strength**\n\n- The paper starts off with multiple observations about the meager accuracy of models in a sparse federated learning setting and then proposes multiple solutions to mitigate the challenges.\n\n- The results across few models seem to be promising, both in terms of final model accuracy and reduced communication. \n\n**Weaknesses**\n\n- While I liked the structure of the paper, it is not clear how bringing the heterogeneity of a system is aligned with the message of the paper. \n\n- While sparsity is an interesting paradigm to reduce various bottlenecks in models, I am not sure if that low density rate (where the most benefit of the paper emerges) at the cost of accuracy degradation would be an acceptable solution even for federated learning. As such, this is not clear if the proposed solution is for a realistic problem setting.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper targets sparse training in the federated learning domain. The paper starts with multiple observations (on small models) about how disagreement of clients on the sparsity mask pattern could significantly reduce the model accuracy, compared to a centralized training. It also shows that this disagreement on the sparsity mask pattern diverges as the training progresses. Based on these observations, the authors propose a two-stage sparse federated learning method with two main purposes; (1) decoupling the identification of initial sparse mask from training and (2) collaboratively learning sparse mask and model weights. Finally, the paper extends the method for a heterogeneous environment (which I think may not be aligned with the main message of the paper) where each client has a different resource budget for which different sparsity mask is required.", "strength_and_weaknesses": "**Strength**\n\n- The paper starts off with multiple observations about the meager accuracy of models in a sparse federated learning setting and then proposes multiple solutions to mitigate the challenges.\n\n- The results across few models seem to be promising, both in terms of final model accuracy and reduced communication. \n\n**Weaknesses**\n\n- While I liked the structure of the paper, it is not clear how bringing the heterogeneity of a system is aligned with the message of the paper. \n\n- While sparsity is an interesting paradigm to reduce various bottlenecks in models, I am not sure if that low density rate (where the most benefit of the paper emerges) at the cost of accuracy degradation would be an acceptable solution even for federated learning. As such, this is not clear if the proposed solution is for a realistic problem setting.", "clarity,_quality,_novelty_and_reproducibility": "(1) Is the model performance sensitive to the number of clients? How the model accuracy and SM changes as the number of clients vary?\n\n(2) Can you also show SM mismatch (similar to Figure 3) for modest density ratio? If I understand correctly you are claiming that such problem only exists for low density ratio?\n\n(3) Similarly, is there any breaking point for the density ratio in which the SM mismatch ratio emerges? What I am trying to understand is to better understand the core reason behind this SM mismatch ratio. Is this just a byproduct of higher probability of mismatch as the number of zeros increase?\n\n(4) (not looking for results) Do you have any intuition how your method can be applied to other forms of sparsity? For example, as you may be familiar, structure sparsity (e.g. N:M, Pixelated Butterfly) methods are more hardware-friendly and efficient for both inference and training. Do you have any intuition how this SM mismatch affects these forms of sparsity?", "summary_of_the_review": "I think the paper is well-motivated with clear description of the observations. The authors went deep into a better understanding of the root cause of low model accuracy in a federated learning setting. While I agree with the authors that this disagreement between sparsity mask patterns *could* be one of the reasons for low accuracy, but I am not convinced that this is the sole reason for such performance degradation. In addition, as mentioned in the questions, it would be great if the authors could provide additional results for higher density ratio to better understand whether such problem exists in a more realistic setting.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666636160443}, {"id": "LKUvZWoTBnu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper826/Reviewer_cFmd"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studies sparse training in a federated learning setting. The authors show that a naïve implementation has significant accuracy degradation, and they proposed a method called federated lottery-aware sparsity hunting. Experiments on ResNet-18 on MNIST, EMNIST, and CIFAR-10 show that the proposed method improves accuracy and reduces communication costs.", "review_text": "The setting of this work is unclear to me. The applicability of the proposed method also seems narrow. ", "strengths": "## strength\nThe authors identify an interesting problem: the accuracy drops for sparse training in the federated learning setting.\n\n## weakness\n1. The paper is hard to follow. I don't understand why FL cannot achieve converged sparse masks. The only difference between centralized and federated learning is data distribution, but you can always have a consensus model on the server by averaging the gradients from workers, right? Or if you use local SGD and average the models, you can fix the sparse mask on each worker between synchronizations, right? \n\n2. The proposed algorithm is designed based on one particular pruning method (Kundu et al. 2021). Does it apply to other sparse training methods (e.g. RigL)? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies sparse training in a federated learning setting. The authors show that a naïve implementation has significant accuracy degradation, and they proposed a method called federated lottery-aware sparsity hunting. Experiments on ResNet-18 on MNIST, EMNIST, and CIFAR-10 show that the proposed method improves accuracy and reduces communication costs.", "strength_and_weaknesses": "## strength\nThe authors identify an interesting problem: the accuracy drops for sparse training in the federated learning setting.\n\n## weakness\n1. The paper is hard to follow. I don't understand why FL cannot achieve converged sparse masks. The only difference between centralized and federated learning is data distribution, but you can always have a consensus model on the server by averaging the gradients from workers, right? Or if you use local SGD and average the models, you can fix the sparse mask on each worker between synchronizations, right? \n\n2. The proposed algorithm is designed based on one particular pruning method (Kundu et al. 2021). Does it apply to other sparse training methods (e.g. RigL)? ", "clarity,_quality,_novelty_and_reproducibility": "I think the motivation and idea would be better explained if the authors could provide a rigorous description of the baseline sparse federated learning algorithm in Section3. ", "summary_of_the_review": "The setting of this work is unclear to me. The applicability of the proposed method also seems narrow. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666565366422}], "openreview_url": "https://openreview.net/forum?id=qhplAU1BOZW", "arxiv_id": "2208.13092", "paper_pdf": "papers/qhplAU1BOZW.pdf", "paper_pdf_sha256": "88534b715314720914f09d213eea012d2e13dde7fbf7f5c6beff0521bf7d95b9", "paper_pdf_bytes": 16715668, "paper_pdf_source": "openreview", "code_url": "https://github.com/SaraBabakN/flash_fl", "code_repository": "SaraBabakN/flash_fl", "code_commit": "fb883133c48490e7d4e3f21cce4212a27f618a22", "code_archive": "repos/qhplAU1BOZW.zip", "code_archive_sha256": "976ce2eb95014d02e09f63ed5d9f3c4a4b3a58971bc6ffb89ddc176e701b4e7a", "code_archive_bytes": 219002, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 824, "github_languages": {"Python": 66195}, "github_archived": false, "github_pushed_at": "2023-11-04T07:06:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/federated-sparse-training-lottery-aware-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "u2JeVfXIQa", "year": 2022, "status": "rejected", "title": "Adaptive Cross-Layer Attention for Image Restoration", "authors": ["Yancheng Wang", "Yingzhen Yang", "Chong Chen", "Ning Xu"], "authorids": ["~Yancheng_Wang2", "~Yingzhen_Yang1", "chongchen@kuaishou.com", "~Ning_Xu3"], "authors_source": "OpenReview API", "abstract": "Non-local attention module has been proven to be crucial for image restoration. Conventional non-local attention processes features of each layer separately, so it risks missing correlation between features among different layers. To address this problem, we propose Cross-Layer Attention (CLA) module in this paper. Instead of finding correlated key pixels within the same layer, each query pixel is allowed to attend to key pixels at previous layers of the network. In order to mitigate the expensive computational cost of such hierarchical attention design, only a small fixed number of keys can be selected for each query from a previous layer. We further propose a variant of CLA termed Adaptive Cross-Layer Attention (ACLA). In ACLA, the number of keys to be aggregated for each query is dynamically selected. A neural architecture search method is used to find the insert positions of ACLA modules to render a compact neural network with compelling performance. Extensive experiments on image restoration tasks including single image super-resolution, image denoising, image demosaicing, and image compression artifacts reduction validate the effectiveness and efficiency of ACLA.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Y3ytzQl_nkX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3241/Reviewer_g1nJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work presents cross-layer attention (CLA) modules to find informative keys across different CNN layers for each query feature. Furthermore, an adaptive cross-layer attention (ACLA) is also formulated to dynamically select keys from different CNN layers by using a NAS method. After that, the authors embedded the presented CLA or ACLA modules into EDSR to formulate a deep model for image restoration. Experimental results on several image restoration tasks show that the developed deep model outperforms state-of-the-art methods.", "review_text": "Strengths:\n(1)\tThis work presents cross-layer attention (CLA) modules and adaptive cross-layer attention (ACLA) to model non-local pixel correlations by considering feature correlations among different layers.\n(2)\tExperimental results show that ACLA works well for different image restoration tasks.\nWeaknesses:\n1.\tIn Tables 1, 2, & 3, the authors are suggested to add an experiment to replace CLA with the classical non-local block.\n2.\tIt is unclear why finding the correlated key pixels from previous CNN layers is capable to enhance the image restoration performance.\n3.\tCompared to CLA, the ACLA dynamically selects key pixels by using a NAS. However, according to Table 1, ACLA only slightly improves CLA in terms of super-resolution accuracy. It tends to degrade the effectiveness of the NAS of ACLA. Please discuss it.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work presents cross-layer attention (CLA) modules to find informative keys across different CNN layers for each query feature. Furthermore, an adaptive cross-layer attention (ACLA) is also formulated to dynamically select keys from different CNN layers by using a NAS method. After that, the authors embedded the presented CLA or ACLA modules into EDSR to formulate a deep model for image restoration. Experimental results on several image restoration tasks show that the developed deep model outperforms state-of-the-art methods.", "main_review": "Strengths:\n(1)\tThis work presents cross-layer attention (CLA) modules and adaptive cross-layer attention (ACLA) to model non-local pixel correlations by considering feature correlations among different layers.\n(2)\tExperimental results show that ACLA works well for different image restoration tasks.\nWeaknesses:\n1.\tIn Tables 1, 2, & 3, the authors are suggested to add an experiment to replace CLA with the classical non-local block.\n2.\tIt is unclear why finding the correlated key pixels from previous CNN layers is capable to enhance the image restoration performance.\n3.\tCompared to CLA, the ACLA dynamically selects key pixels by using a NAS. However, according to Table 1, ACLA only slightly improves CLA in terms of super-resolution accuracy. It tends to degrade the effectiveness of the NAS of ACLA. Please discuss it.\n", "summary_of_the_review": "Please refer to the weaknesses of the main review.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1639219093806}, {"id": "iZ_KJqWcGY6", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3241/Reviewer_6N93"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a novel Cross Layer Attention (CLA) module for image restoration tasks. The CLA module doesn’t look for correlated key pixels within the same layer as other algorithms do; instead query pixels attend to key pixels at previous layers of the network. To reduce computational complexity deformable convolution is used to reduce the number of sampled keys.  To further reduce computational complexity, the paper proposes a modified CLA called Adaptive Cross Layer Attention (ACLA). In ACLA, the number of collected keys for each query is dynamically selected through a gating mechanism.  Further, a neural architecture search method was used to find the proper positions to insert the ACLA modules in the backbone network.  The paper provides comprehensive experiments on image restoration tasks to validate the effectiveness and efficiency of ACLA. Conducted experiments include single image super resolution, image denoising, image demosaicing and image compression artifact reduction.  Contributions are the CLA and ACLA modules designed to efficiently perform attention across layers in a neural network.", "review_text": "Attention is a powerful mechanism and has found utility in numerous tasks in both high and low level computer vision as effectively argued in the paper.  Therefore paper is addressing an interesting topic relevant to the ICLR community.\n\nNote however, the idea of exploiting correlations across layers in convolutional networks isn’t new, for example, the papers\n\n•\tChang et al., “EPSNet: Efficient Panoptic Segmentation Network with Cross-layer Attention Fusion,” ACCV 2020 also applies non-local attention at different layers, in a hierarchical network architecture used for image segmentation.\n\n•\tMei et al., “Image Super-Resolution with Cross-Scale Non-Local Attention and Exhaustive Self-Exemplars Mining,” CVPR 2020 from the authors of the PANet (Mei et al., 2020) referenced and compared to in the paper; has similar motivations for image restoration tasks.\n\nHowever, the approach taken in this paper appears to be novel through its use of deformable convolution, gating, and neural architecture search to reduce the computational complexity.  \n\nThe paper is clearly described.  It would be helpful if the paper commented on the availability of source code in order for others to replicate the results.  The approach is fairly generic so the method may find applicability in many tasks beyond image restoration.\n\nExperimental results cover a wide range of image restoration tasks to prove the effectiveness and efficiency of the proposed module.  Although the method often produces the best results, the incremental improvement is arguably small; for example in Table 4 the PSNR increases by 0.2 dB compared to EDSR.  However, the visual results are compelling particularly in the appendix.  It seems somewhat surprising to this reviewer that the PSNR improvement is so small yet visually the results appear to be more pronounced (for example, in Figure 5).  \n\nStrengths: \n•\tWhile cross-layer attention isn’t a new idea, the approach taken in this paper has novelty and provides a fairly generic approach that could be used with different backbones.\n•\tThe method carefully addresses the issue of computational complexity to produce a practical working solution.\n\n•\tComprehensive experiments including quantitate analysis and ablation study shows how the proposed algorithm performs efficiently.\n\n•\tThe paper is clearly written and includes useful figures, e.g. Figure 2.\n\nWeaknesses.\n•\tThe primary disadvantage is the small incremental benefit for some of the steps of the method.  For example, the ablation study in Table 6 shows only a 0.07 dB improvement on Set5 for ACLA compared to CLA.  Given that ACLA requires neural architecture search, dynamic key selection and more it may be a lot of work for limited benefit to squeeze out less than 0.1 dB improvement.\n\nSmall corrections and suggestions:\n\n•\tThe first sentence in the introduction describes degradation as “irreversible”.  But the idea of the image restoration is to restore the degradation.  If the image is restored, isn’t the degradation reversed?  Perhaps this qualifying statement could be removed or clarified. \n\n•\tPage 1 argues that most CNN-based image restoration methods tend to produce smooth results, potentially due to only referring to keys in the same layer in their attention modules.  This seems highly speculative – it would be better to show this is the case or possibly remove this.  It’s well known there is a perception-distoration tradeoff in image restoration; e.g. Blau et al., “The Perception-Distortion Tradeoff,” CVPR 2018 and blurring often results in minimizing distortion using standard losses and can be addressed through generative methods that hallucinate detail at the expense of PSNR.\n\n•\tPage 4, please change “To find for keys” to “To search for keys”\n\n•\tPage 5, please change “ALDA” to “ACLA”\n\n•\tPage 5, please change “top op CLA” to “top of CLA”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a novel Cross Layer Attention (CLA) module for image restoration tasks. The CLA module doesn’t look for correlated key pixels within the same layer as other algorithms do; instead query pixels attend to key pixels at previous layers of the network. To reduce computational complexity deformable convolution is used to reduce the number of sampled keys.  To further reduce computational complexity, the paper proposes a modified CLA called Adaptive Cross Layer Attention (ACLA). In ACLA, the number of collected keys for each query is dynamically selected through a gating mechanism.  Further, a neural architecture search method was used to find the proper positions to insert the ACLA modules in the backbone network.  The paper provides comprehensive experiments on image restoration tasks to validate the effectiveness and efficiency of ACLA. Conducted experiments include single image super resolution, image denoising, image demosaicing and image compression artifact reduction.  Contributions are the CLA and ACLA modules designed to efficiently perform attention across layers in a neural network.", "main_review": "Attention is a powerful mechanism and has found utility in numerous tasks in both high and low level computer vision as effectively argued in the paper.  Therefore paper is addressing an interesting topic relevant to the ICLR community.\n\nNote however, the idea of exploiting correlations across layers in convolutional networks isn’t new, for example, the papers\n\n•\tChang et al., “EPSNet: Efficient Panoptic Segmentation Network with Cross-layer Attention Fusion,” ACCV 2020 also applies non-local attention at different layers, in a hierarchical network architecture used for image segmentation.\n\n•\tMei et al., “Image Super-Resolution with Cross-Scale Non-Local Attention and Exhaustive Self-Exemplars Mining,” CVPR 2020 from the authors of the PANet (Mei et al., 2020) referenced and compared to in the paper; has similar motivations for image restoration tasks.\n\nHowever, the approach taken in this paper appears to be novel through its use of deformable convolution, gating, and neural architecture search to reduce the computational complexity.  \n\nThe paper is clearly described.  It would be helpful if the paper commented on the availability of source code in order for others to replicate the results.  The approach is fairly generic so the method may find applicability in many tasks beyond image restoration.\n\nExperimental results cover a wide range of image restoration tasks to prove the effectiveness and efficiency of the proposed module.  Although the method often produces the best results, the incremental improvement is arguably small; for example in Table 4 the PSNR increases by 0.2 dB compared to EDSR.  However, the visual results are compelling particularly in the appendix.  It seems somewhat surprising to this reviewer that the PSNR improvement is so small yet visually the results appear to be more pronounced (for example, in Figure 5).  \n\nStrengths: \n•\tWhile cross-layer attention isn’t a new idea, the approach taken in this paper has novelty and provides a fairly generic approach that could be used with different backbones.\n•\tThe method carefully addresses the issue of computational complexity to produce a practical working solution.\n\n•\tComprehensive experiments including quantitate analysis and ablation study shows how the proposed algorithm performs efficiently.\n\n•\tThe paper is clearly written and includes useful figures, e.g. Figure 2.\n\nWeaknesses.\n•\tThe primary disadvantage is the small incremental benefit for some of the steps of the method.  For example, the ablation study in Table 6 shows only a 0.07 dB improvement on Set5 for ACLA compared to CLA.  Given that ACLA requires neural architecture search, dynamic key selection and more it may be a lot of work for limited benefit to squeeze out less than 0.1 dB improvement.\n\nSmall corrections and suggestions:\n\n•\tThe first sentence in the introduction describes degradation as “irreversible”.  But the idea of the image restoration is to restore the degradation.  If the image is restored, isn’t the degradation reversed?  Perhaps this qualifying statement could be removed or clarified. \n\n•\tPage 1 argues that most CNN-based image restoration methods tend to produce smooth results, potentially due to only referring to keys in the same layer in their attention modules.  This seems highly speculative – it would be better to show this is the case or possibly remove this.  It’s well known there is a perception-distoration tradeoff in image restoration; e.g. Blau et al., “The Perception-Distortion Tradeoff,” CVPR 2018 and blurring often results in minimizing distortion using standard losses and can be addressed through generative methods that hallucinate detail at the expense of PSNR.\n\n•\tPage 4, please change “To find for keys” to “To search for keys”\n\n•\tPage 5, please change “ALDA” to “ACLA”\n\n•\tPage 5, please change “top op CLA” to “top of CLA”\n", "summary_of_the_review": "Overall, this paper address an important problem of performing attention across different layers in a neural network and proposes a novel approach that carefully handles the computational complexity.  The method is demonstrated to work effectively in several image restoration methods.  Numerically the results are only slightly better than competitors, but it does reflect an advance.  This reviewer is somewhere between marginally above and accept for this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635687869464}, {"id": "h2bPEmrAb8n", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3241/Reviewer_HQYo"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a cross-layer attention module for the image restoration tasks. Unlike previous conventional non-local attention approaches that find correlated keys within the same layer, the proposed method selects keys across the layers. In order to prevent expensive computational costs, the authors propose adaptive cross-layer attention modules. The proposed method is validated on various image restoration tasks including SR, denoising, demosaicing, and compression.\n", "review_text": "Strength\n\n- The proposed method seems quite new, and properly designed.\n\n- The proposed method is applied to various image restoration tasks to show the generality power of the method.\n\nWeakness\n\n- Isn't just nonlocal attention good enough? What is the exact problem of nonlocal attention?  An experimental analysis is needed to determine the specific benefits of using cross-layer nonlocal attention. In particular, what are the advantages of using it for image restoration tasks?\n\n- It would be better to mention differences from LAM more specifically.\n\n- In Table1, why is there no CLA in RCAN? \n\n- In Table1, compared to RDN or HAN, the performance is not very good. How different is the inference time? If the speed is not very fast, I think that the meaning of the proposed method is not great.\n\n- In Table2, and Table3, similar to SR, there is not much difference in performance compared to the existing method, and the baseline model is already quite good. It is interpreted that the effect of the proposed method is not large.\n\n- Why is the order of HAN and RDN in x4 different from x2 and x3?\n\n- In Table 5, the difference in performance according to L is not very large, but on the contrary, FLOP and Pram increase significantly. It seems that the effect of the CLA module is not so great.\n\n- In other ablation studies, various experiments are performed, but there is little difference in performance between each, making meaningful analysis difficult.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a cross-layer attention module for the image restoration tasks. Unlike previous conventional non-local attention approaches that find correlated keys within the same layer, the proposed method selects keys across the layers. In order to prevent expensive computational costs, the authors propose adaptive cross-layer attention modules. The proposed method is validated on various image restoration tasks including SR, denoising, demosaicing, and compression.\n", "main_review": "Strength\n\n- The proposed method seems quite new, and properly designed.\n\n- The proposed method is applied to various image restoration tasks to show the generality power of the method.\n\nWeakness\n\n- Isn't just nonlocal attention good enough? What is the exact problem of nonlocal attention?  An experimental analysis is needed to determine the specific benefits of using cross-layer nonlocal attention. In particular, what are the advantages of using it for image restoration tasks?\n\n- It would be better to mention differences from LAM more specifically.\n\n- In Table1, why is there no CLA in RCAN? \n\n- In Table1, compared to RDN or HAN, the performance is not very good. How different is the inference time? If the speed is not very fast, I think that the meaning of the proposed method is not great.\n\n- In Table2, and Table3, similar to SR, there is not much difference in performance compared to the existing method, and the baseline model is already quite good. It is interpreted that the effect of the proposed method is not large.\n\n- Why is the order of HAN and RDN in x4 different from x2 and x3?\n\n- In Table 5, the difference in performance according to L is not very large, but on the contrary, FLOP and Pram increase significantly. It seems that the effect of the CLA module is not so great.\n\n- In other ablation studies, various experiments are performed, but there is little difference in performance between each, making meaningful analysis difficult.\n", "summary_of_the_review": "- This paper is technically sound and quite new. However, the performance difference according to the proposed method is not significant in most experiments.\n\n- In other words, it is difficult to say that the value of the proposed method is that high. Even technically, I think that difference with the existing methods is not significant. (Of course, the proposed method have different points compared to existing methods)\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635325412062}, {"id": "A9JWCOQRavY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3241/Reviewer_Xgo2"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a Cross-Layer Attention(CLA) module in order to capture the correlations. Between features among different layers. Besides, an Adaptive Cross-Layer Attention (ACLA) is proposed to reduce the computational cost of the Cross-Layer Attention(CLA) module. Lastly,a neural architecture search method is used to find the insert positions of ACLA modules to further improve performance. ", "review_text": "Extensive experiments are conducted to validate the proposed method both in quality and quantity.\n\n(1)\tThe effective of layer attention is a commonly used method in high-level and low-level computer vision task. There are various layer attention module was proposed in the past three years. So, the main contribution of the paper is lack of novelty. More importantly, the findings of the paper has been general knowledge in computer vision community. The authors may need to focus on more important and valuable problems in image restoration field.\n\n(2)\tThe claim of “However, LAM neglects the difference of spatial positions of features.” is not strictly correct. The authors have used the CSAM to exploit the spatial positions of features in their paper. The authors may also need to compare the proposed to LAM. \n\n(3)\tThe claim of “both IPT and SwinIR requires large-scale datasets for good performance.” is not exactly correct. SwinIR only use a small training set to train their network. \n\n(4)\tMore importantly, in the experiments, there are only the proposed CLA and ACLA results ,which can’t present the comparative experiment with other attention mechanisms. The author may need explain the difference between the proposed CLA and other attention mechanisms in computer vision field.\n\n(5)\tThe equation (13) should be further expained clearly.\n\n(6)\tIn the experiments ,there are lack of params such as single image super-resolution、image denoising and image compression artifacts reduction. Also,the RCAN+CLA may neglect by the authors. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a Cross-Layer Attention(CLA) module in order to capture the correlations. Between features among different layers. Besides, an Adaptive Cross-Layer Attention (ACLA) is proposed to reduce the computational cost of the Cross-Layer Attention(CLA) module. Lastly,a neural architecture search method is used to find the insert positions of ACLA modules to further improve performance. ", "main_review": "Extensive experiments are conducted to validate the proposed method both in quality and quantity.\n\n(1)\tThe effective of layer attention is a commonly used method in high-level and low-level computer vision task. There are various layer attention module was proposed in the past three years. So, the main contribution of the paper is lack of novelty. More importantly, the findings of the paper has been general knowledge in computer vision community. The authors may need to focus on more important and valuable problems in image restoration field.\n\n(2)\tThe claim of “However, LAM neglects the difference of spatial positions of features.” is not strictly correct. The authors have used the CSAM to exploit the spatial positions of features in their paper. The authors may also need to compare the proposed to LAM. \n\n(3)\tThe claim of “both IPT and SwinIR requires large-scale datasets for good performance.” is not exactly correct. SwinIR only use a small training set to train their network. \n\n(4)\tMore importantly, in the experiments, there are only the proposed CLA and ACLA results ,which can’t present the comparative experiment with other attention mechanisms. The author may need explain the difference between the proposed CLA and other attention mechanisms in computer vision field.\n\n(5)\tThe equation (13) should be further expained clearly.\n\n(6)\tIn the experiments ,there are lack of params such as single image super-resolution、image denoising and image compression artifacts reduction. Also,the RCAN+CLA may neglect by the authors. \n\n", "summary_of_the_review": "Please see the paper weakness. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634811066488}], "openreview_url": "https://openreview.net/forum?id=u2JeVfXIQa", "arxiv_id": "2203.03619", "paper_pdf": "papers/u2JeVfXIQa.pdf", "paper_pdf_sha256": "f0a2f5a94fab68e20d1b7273dd910d759ca4cf17af372ffa9be9d2d07658ded8", "paper_pdf_bytes": 1851316, "paper_pdf_source": "openreview", "code_url": "https://github.com/Statistical-Deep-Learning/ACLA-IKS", "code_repository": "Statistical-Deep-Learning/ACLA-IKS", "code_commit": "aaefb77dcc44ba81d06c13b3d12db2b641c15273", "code_archive": "repos/u2JeVfXIQa.zip", "code_archive_sha256": "3978e80dada063057fd8071aaf83a086103c758d428e01b6cda787227e2a7847", "code_archive_bytes": 3785892, "code_file_count": 160, "code_extensions": {".py": 127, ".sh": 18, ".ipynb": 6, ".h": 3, ".cpp": 2, ".m": 2, ".cu": 1, ".cuh": 1}, "github_disk_usage_kb": 3434, "github_languages": {"Python": 490756, "Cuda": 62010, "Jupyter Notebook": 52308, "MATLAB": 17294, "Shell": 15601, "C++": 6172}, "github_archived": false, "github_pushed_at": "2025-11-06T23:50:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adaptive-cross-layer-attention-for-image-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "nPVlVsBTiJ", "year": 2021, "status": "rejected", "title": "Adversarial Boot Camp: label free certified robustness in one epoch", "authors": ["Ryan Campbell", "Chris Finlay", "Adam M Oberman"], "authorids": ["~Ryan_Campbell2", "~Chris_Finlay1", "~Adam_M_Oberman1"], "authors_source": "OpenReview API", "abstract": "Machine learning models are vulnerable to adversarial attacks.  One approach to addressing this vulnerability is certification, which focuses on models that are guaranteed to be robust for a given perturbation size.  A drawback of recent certified models is that they are stochastic: they require multiple computationally expensive model evaluations with random noise added to a given image. In our work, we present a deterministic certification approach which results in a certifiably robust model. This approach is based on an equivalence between training with a particular regularized loss, and the expected values of Gaussian averages. We achieve certified models on ImageNet-1k by retraining a model with this loss for one epoch without the use of label information.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "6I11oU9l_gU", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper667/AnonReviewer4"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to use a deterministic classifier to replace the sampling\nprocess in randomized smoothing based certifiably robust models.  The goal of\ntraining a deterministic robust classifier to avoid the high cost of randomized\nsmoothing is a right direction to look at. \n\nThe reason that we can get certified robustness with Gaussian smoothing is that\nthe smoothed classifier becomes Lipschitz (see Salman et al.). Unfortunately,\nit is a stochastic classifier and typically it is impossible to access the\nsmoothed classifier directly, and that's why in randomized smoothing (e.g.,\nCohen's PREDICT procedure), sampling is necessary.\n\nThe approach in this work is to train a smoothed classifier that essentially\nreturns the same prediction as the mean of the originally stochastic\nclassifier. The authors pointed out the connection between Gaussian smoothed\nclassifier and gradient regularization, so they use gradient regularization to\nobtain the desired deterministic classifier.\n\nUnfortunately, it seems to me that the proposed approach is not sound. With\ngradient regularization, we try to make the learned classifier to be smooth,\nhowever there is no guarantee that such a learned classifier will indeed be\nsmooth and produce the same outcome as the original Gaussian smoothed\nclassifier, so we cannot use the outcome of this classifier to replace Cohen's\nSampleUnderNoise procedure.  More precisely, optimizing the proposed loss\nfunction (3) does not guarantee the learned classifier f^smooth to be Lipschitz,\nwhere the original Gaussian smoothed classifier is guaranteed to be Lipschitz.\nAlthough the loss attempts to do so with gradient regularization, there is no\nguarantee here. So the entire procedure is not certified anymore.\n\nOn the positive side, the proposed method may work as a good empirical defense,\nsince the smoothed classifier can be learned quickly and can be more robust\nthan the original classifier. This may be advantageous for certain applications\nwhere adversarial training is too slow or we don't want to retrain the original\nclassifier.\n\nBecause of the fundamental problem mentioned above, I cannot recommend\nacceptance of this paper. I am willing to discuss with the authors further in\ncase I misunderstand some parts of the paper.\n\n---\n### After rebuttal:\n\nAfter reading the rebuttal, I feel my main concern is still not addressed by the response. The authors agree that they may provide a different kind of guarantee for the certificates as in (Cohen et al., Salman et al.). An empirical comparison between the output of the proposed model and the mean from sampling is not sufficient. We hope the authors can improve on this point and provide formal robustness guarantees like those in (Cohen et al., Salman et al.).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Problems with the proposed approach", "review": "This paper proposes to use a deterministic classifier to replace the sampling\nprocess in randomized smoothing based certifiably robust models.  The goal of\ntraining a deterministic robust classifier to avoid the high cost of randomized\nsmoothing is a right direction to look at. \n\nThe reason that we can get certified robustness with Gaussian smoothing is that\nthe smoothed classifier becomes Lipschitz (see Salman et al.). Unfortunately,\nit is a stochastic classifier and typically it is impossible to access the\nsmoothed classifier directly, and that's why in randomized smoothing (e.g.,\nCohen's PREDICT procedure), sampling is necessary.\n\nThe approach in this work is to train a smoothed classifier that essentially\nreturns the same prediction as the mean of the originally stochastic\nclassifier. The authors pointed out the connection between Gaussian smoothed\nclassifier and gradient regularization, so they use gradient regularization to\nobtain the desired deterministic classifier.\n\nUnfortunately, it seems to me that the proposed approach is not sound. With\ngradient regularization, we try to make the learned classifier to be smooth,\nhowever there is no guarantee that such a learned classifier will indeed be\nsmooth and produce the same outcome as the original Gaussian smoothed\nclassifier, so we cannot use the outcome of this classifier to replace Cohen's\nSampleUnderNoise procedure.  More precisely, optimizing the proposed loss\nfunction (3) does not guarantee the learned classifier f^smooth to be Lipschitz,\nwhere the original Gaussian smoothed classifier is guaranteed to be Lipschitz.\nAlthough the loss attempts to do so with gradient regularization, there is no\nguarantee here. So the entire procedure is not certified anymore.\n\nOn the positive side, the proposed method may work as a good empirical defense,\nsince the smoothed classifier can be learned quickly and can be more robust\nthan the original classifier. This may be advantageous for certain applications\nwhere adversarial training is too slow or we don't want to retrain the original\nclassifier.\n\nBecause of the fundamental problem mentioned above, I cannot recommend\nacceptance of this paper. I am willing to discuss with the authors further in\ncase I misunderstand some parts of the paper.\n\n---\n### After rebuttal:\n\nAfter reading the rebuttal, I feel my main concern is still not addressed by the response. The authors agree that they may provide a different kind of guarantee for the certificates as in (Cohen et al., Salman et al.). An empirical comparison between the output of the proposed model and the mean from sampling is not sufficient. We hope the authors can improve on this point and provide formal robustness guarantees like those in (Cohen et al., Salman et al.).\n", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603874471727}, {"id": "xYMg204ey0J", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper667/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Randomized smoothing is the major way to certify the robustness of large scale networks, however, it requires sampling from Gaussian distribution many times, which is not fast enough for real-time inference. This paper uses a regularized loss to get deterministic Gaussian averaged results. This paper points out an interesting direction for certifying robustness, the method is simple and effective.\n\nStrength:\n\n1. The paper is clearly written. \n\n2. It is very interesting to see a method that can certify the robustness without the computational intensive randomized smoothing.\n\n3. The method is simple and effective, does not require much computation resources, which improves the inference speed by around 10 times (Table 2).\n\n4. The speed for randomized smoothing certification is a major concern for the community. This paper address this problem. If this paper is really effective, it can have a broad impact on the research community for robustness certification.\n\nWeakness and Questions:\n\n1. What if the attacker directly attacks the objective function for certification (e.g. equation 2) ? Given that the certification is deterministic, is it possible to fool the certification method?\n\n2. In Figure 2, it seems the deterministic under performs some baselines.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper propose a deterministic method for certified robustness under adversarial attack. Different from the prior random smoothing approach, this paper uses deterministic inference and achieves times of speed up. The idea is novel and interesting.", "review": "Randomized smoothing is the major way to certify the robustness of large scale networks, however, it requires sampling from Gaussian distribution many times, which is not fast enough for real-time inference. This paper uses a regularized loss to get deterministic Gaussian averaged results. This paper points out an interesting direction for certifying robustness, the method is simple and effective.\n\nStrength:\n\n1. The paper is clearly written. \n\n2. It is very interesting to see a method that can certify the robustness without the computational intensive randomized smoothing.\n\n3. The method is simple and effective, does not require much computation resources, which improves the inference speed by around 10 times (Table 2).\n\n4. The speed for randomized smoothing certification is a major concern for the community. This paper address this problem. If this paper is really effective, it can have a broad impact on the research community for robustness certification.\n\nWeakness and Questions:\n\n1. What if the attacker directly attacks the objective function for certification (e.g. equation 2) ? Given that the certification is deterministic, is it possible to fool the certification method?\n\n2. In Figure 2, it seems the deterministic under performs some baselines.", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603812304539}, {"id": "kas_iz80FKJ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper667/AnonReviewer1"], "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper claims that a (computationally intractable) randomized smoothing of any classifier can be distilled into the (deterministic) classifier itself via fine-tuning it with gradient penalty. This is motivated by a theoretical result that Gaussian smoothing of a classifier is equivalent to solving a certain heat equation, which can be approximated by a regularized loss training. Experimental results use the resulting deterministic classifier to compute the certified radius compared to (stochastic) smoothed classifiers, arguing its efficiency and higher certified radius of the proposed method. \n\nThe relation between randomized smoothing and PDE seems to be an interesting direction to explore. My biggest concern is that, however, whether the proposed deterministic smooth classifier is indeed *certifiably* robust in practice. Apart from its theoretical motivation, I could not find any statistical guarantee that the deterministic smooth classifier is provably close to its stochastic counterpart for every input x, which is essential to claim the certifiable robustness of the proposed model. Otherwise, how one could prove that the deterministic smooth classifier is indeed robust to adversarial examples? - although randomized smoothing may require much more computation time, it at least provides such a statistical yet practical guarantee, namely as the CERTIFY algorithm [1].\n\nAnother possible way to go is to show that the *empirical* robustness of the deterministic smooth classifier is non-trivial. In this context, actually, the message of the paper can be \"very\" surprising (and very unlikely, at the same time): one can robustify any classifier by a single pass of fine-tuning with gradient penalty, even without adversarial training. The paper indeed provide a related result, i.e., empirical test accuracy on adversarial attacks, but the current evaluation is too weak to support the claim: considering a bunch of defense papers that are broken after published [2, 3], it is too hard for me to believe the results as is even it could bear from a specific configuration of PGD and DDN. I would recommend the paper to follow the standard guidelines suggested by [4]. The paper could explore more possible configurations of PGD attacks, or other types of attacks (e.g., gradient-free attacks [5], or black-box attacks), just to name a few.\n\n[1] Cohen et al., Certified Adversarial Robustness via Randomized Smoothing, ICML 2019.\n\n[2] Athalye et al., Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples, ICML 2018.\n\n[3] Tramer et al., On Adaptive Attacks to Adversarial Example Defenses, 2020.\n\n[4] Carlini et al., On Evaluating Adversarial Robustness, 2019.\n\n[5] Uesato et al., Adversarial Risk and the Dangers of Evaluating Against Weak Attacks, ICML 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "The paper claims that a (computationally intractable) randomized smoothing of any classifier can be distilled into the (deterministic) classifier itself via fine-tuning it with gradient penalty. This is motivated by a theoretical result that Gaussian smoothing of a classifier is equivalent to solving a certain heat equation, which can be approximated by a regularized loss training. Experimental results use the resulting deterministic classifier to compute the certified radius compared to (stochastic) smoothed classifiers, arguing its efficiency and higher certified radius of the proposed method. \n\nThe relation between randomized smoothing and PDE seems to be an interesting direction to explore. My biggest concern is that, however, whether the proposed deterministic smooth classifier is indeed *certifiably* robust in practice. Apart from its theoretical motivation, I could not find any statistical guarantee that the deterministic smooth classifier is provably close to its stochastic counterpart for every input x, which is essential to claim the certifiable robustness of the proposed model. Otherwise, how one could prove that the deterministic smooth classifier is indeed robust to adversarial examples? - although randomized smoothing may require much more computation time, it at least provides such a statistical yet practical guarantee, namely as the CERTIFY algorithm [1].\n\nAnother possible way to go is to show that the *empirical* robustness of the deterministic smooth classifier is non-trivial. In this context, actually, the message of the paper can be \"very\" surprising (and very unlikely, at the same time): one can robustify any classifier by a single pass of fine-tuning with gradient penalty, even without adversarial training. The paper indeed provide a related result, i.e., empirical test accuracy on adversarial attacks, but the current evaluation is too weak to support the claim: considering a bunch of defense papers that are broken after published [2, 3], it is too hard for me to believe the results as is even it could bear from a specific configuration of PGD and DDN. I would recommend the paper to follow the standard guidelines suggested by [4]. The paper could explore more possible configurations of PGD attacks, or other types of attacks (e.g., gradient-free attacks [5], or black-box attacks), just to name a few.\n\n[1] Cohen et al., Certified Adversarial Robustness via Randomized Smoothing, ICML 2019.\n\n[2] Athalye et al., Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples, ICML 2018.\n\n[3] Tramer et al., On Adaptive Attacks to Adversarial Example Defenses, 2020.\n\n[4] Carlini et al., On Evaluating Adversarial Robustness, 2019.\n\n[5] Uesato et al., Adversarial Risk and the Dangers of Evaluating Against Weak Attacks, ICML 2018.", "rating": "3: Clear rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603773941408}, {"id": "PYL8Z03dmmy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper667/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes to finetune a pretrained network with a gradient norm regularizer to mimic the Gaussian noise augmentation during training. I have two main concerns about this paper:\n\n1. As claimed by the authors, they are motivated to solve the stochasticity of previous certified defenses like random smoothing. However,  there are mainstream certified training methods that are deterministic during inference (with a single query) [1][2][3][4], but are not discussed or empirically compared in this paper.\n\n2. The proposed regularizer function in Eq.(3) is actually equivalent to L-2 adversarial training, which can be simply proved by a dual norm trick [5]. This makes the results in Table 4 seem counter-intuitive, where a from-scratch L-2 adversarially trained model performs worse than a model finetuned by a similar mechanism. I suggest the authors better explain these both formally and empirically.\n\nMinor:\nIn Sec.4.2, the authors claim that they use the pretrained ImageNet model from Madry GitHub. But as far as I know, there are only pretrained models on MNIST and CIFAR-10 in their repository, could the authors provide an official link to this baseline? \n\n\n\nReference:\n\n[1] Wong et al. Provable defenses against adversarial examples via the convex outer adversarial polytope. ICML 2018\n\n[2] Wong et al. Scaling provable adversarial defenses. NeurIPS 2018\n\n[3] Dvijotham et al. Training verified learners with learned verifiers. arXiv 2018\n\n[4] Dvijotham et al. A dual approach to scalable verification of deep networks. UAI 2018\n\n[5] Simon-Gabriel et al. First-order adversarial vulnerability of neural networks and input dimension. ICML 2019", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Unconvinced results and insufficient discussion on related work", "review": "This paper proposes to finetune a pretrained network with a gradient norm regularizer to mimic the Gaussian noise augmentation during training. I have two main concerns about this paper:\n\n1. As claimed by the authors, they are motivated to solve the stochasticity of previous certified defenses like random smoothing. However,  there are mainstream certified training methods that are deterministic during inference (with a single query) [1][2][3][4], but are not discussed or empirically compared in this paper.\n\n2. The proposed regularizer function in Eq.(3) is actually equivalent to L-2 adversarial training, which can be simply proved by a dual norm trick [5]. This makes the results in Table 4 seem counter-intuitive, where a from-scratch L-2 adversarially trained model performs worse than a model finetuned by a similar mechanism. I suggest the authors better explain these both formally and empirically.\n\nMinor:\nIn Sec.4.2, the authors claim that they use the pretrained ImageNet model from Madry GitHub. But as far as I know, there are only pretrained models on MNIST and CIFAR-10 in their repository, could the authors provide an official link to this baseline? \n\n\n\nReference:\n\n[1] Wong et al. Provable defenses against adversarial examples via the convex outer adversarial polytope. ICML 2018\n\n[2] Wong et al. Scaling provable adversarial defenses. NeurIPS 2018\n\n[3] Dvijotham et al. Training verified learners with learned verifiers. arXiv 2018\n\n[4] Dvijotham et al. A dual approach to scalable verification of deep networks. UAI 2018\n\n[5] Simon-Gabriel et al. First-order adversarial vulnerability of neural networks and input dimension. ICML 2019", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603635898411}], "openreview_url": "https://openreview.net/forum?id=nPVlVsBTiJ", "arxiv_id": "2010.02508", "paper_pdf": "papers/nPVlVsBTiJ.pdf", "paper_pdf_sha256": "f9ca0805c0c46466ecec5c73facc7a969fec08078581600b7d57376d44843cf8", "paper_pdf_bytes": 785534, "paper_pdf_source": "openreview", "code_url": "https://github.com/ryancampbell514/HeatSmoothing", "code_repository": "ryancampbell514/HeatSmoothing", "code_commit": "a67b3d857ea6bb37c34c6ce591f7626052278035", "code_archive": "repos/nPVlVsBTiJ.zip", "code_archive_sha256": "f03dd8d49c3585c0ceef82f4cddec9c186a02444c496b4e312a3d494c5aeb38c", "code_archive_bytes": 562492, "code_file_count": 52, "code_extensions": {".py": 47, ".sh": 3, ".ipynb": 2}, "github_disk_usage_kb": 1988, "github_languages": {"Python": 286446, "Jupyter Notebook": 114568, "Shell": 4949}, "github_archived": false, "github_pushed_at": "2020-10-08T17:32:43Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-boot-camp-label-free-certified-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RYwtJyOP3k", "year": 2026, "status": "rejected", "title": "Probably Approximately Correct Labels", "authors": ["Emmanuel Candes", "Andrew Ilyas", "Tijana Zrnic"], "authorids": ["~Emmanuel_Candes1", "~Andrew_Ilyas1", "~Tijana_Zrnic1"], "authors_source": "OpenReview API", "abstract": "Obtaining high-quality labeled datasets is often costly, requiring either\nhuman annotation or expensive experiments. \nIn theory, powerful pre-trained AI models provide an opportunity to \nautomatically label datasets and save costs. \nUnfortunately, these models provide no guarantees on their accuracy, \nmaking wholesale replacement of manual labeling impractical.\nIn this work, we propose a method for leveraging pre-trained AI models to curate \ncost-effective and high-quality datasets.\nIn particular, our approach results in\n*probably approximately correct labels*: with high probability, the overall\nlabeling error is small. \nOur method is nonasymptotically valid under minimal assumptions on the dataset or\nthe AI model being studied, and thus enables rigorous yet efficient dataset\ncuration using modern AI models. We demonstrate the benefits of the methodology\nthrough text annotation with large language models, image labeling with\npre-trained vision models, and protein folding analysis with AlphaFold.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "AHyOxVNHWl", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17773/Reviewer_FoeC"], "rating": 4, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 3, "summary": "This paper introduces Probably Approximately Correct (PAC) Labeling, aiming to provide provable guarantees for AI-assisted labeling.\nIt proposes to select an uncertainty threshold such that all samples with model confidence above the threshold are labeled automatically, while others are sent to experts.\nThe guarantee states that, with high probability, the global labeling error is ≤ ε.\nExtensions include a multi-model PAC Router and a practical multicalibration procedure to improve uncertainty reliability.", "review_text": "This paper introduces Probably Approximately Correct (PAC) Labeling, aiming to provide provable guarantees for AI-assisted labeling.\nIt proposes to select an uncertainty threshold such that all samples with model confidence above the threshold are labeled automatically, while others are sent to experts.\nThe guarantee states that, with high probability, the global labeling error is ≤ ε.\nExtensions include a multi-model PAC Router and a practical multicalibration procedure to improve uncertainty reliability.", "strengths": "1. Relevant and well-motivated problem formulation. Applying PAC reasoning to automatic labeling is timely and practically meaningful.  \n   It formalizes a common heuristic (confidence-based filtering) into a statistically grounded process.  \n\n2. Simplicity and generality. The proposed approach is model-agnostic and directly applicable across diverse labeling pipelines.  \n\n3. Comprehensive empirical coverage. Experiments span multiple modalities and show consistent compliance with PAC bounds while reducing expert effort.  \n\n4. Clear presentation and awareness of practical issues.   The inclusion of multicalibration demonstrates attention to real-world miscalibration.", "weaknesses": "1. Limited overall contribution despite novel framing.\nWhile the paper introduces new formulations (e.g., PAC Labeling and PAC Router) that are conceptually fresh, the underlying theoretical substance remains limited.\nThe analysis primarily builds on existing mean-upper-bound PAC results without introducing new bounds, assumptions, or insights into the nature of uncertainty in labeling.\nConsequently, although the problem setting is well-motivated, the contribution lies more in repackaging and integration than in theoretical advancement.\n\n\n2. Disconnect between calibration and PAC validity.\nMulticalibration empirically adjusts uncertainty scores but may violate the monotonicity assumption required for PAC reasoning.  \n   The paper provides no formal analysis showing that PAC guarantees still hold after calibration.\n\n3. Lack of intuitive and competitive comparisons.\nAlthough experiments are broad, they do not clearly show cost–error trade-offs (e.g., comparing methods under equal error rates).  \n   Baselines are limited to simple heuristics; stronger competitors like conformal labeling or active selection are missing.  \n   As a result, the empirical advantage remains qualitative rather than quantitative.", "questions": "1. What concrete theoretical or methodological innovations are introduced beyond adapting existing PAC mean-upper-bound results?  \n\n2. Does multicalibration theoretically preserve the assumptions required for PAC guarantees, or is it purely an empirical fix?  \n\n3. Have the authors evaluated cost–error trade-offs, e.g., comparing expert costs at equal error levels across methods?  \n\n4. How does the framework behave under correlated errors or severe miscalibration beyond what multicalibration can correct?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Probably Approximately Correct (PAC) Labeling, aiming to provide provable guarantees for AI-assisted labeling.\nIt proposes to select an uncertainty threshold such that all samples with model confidence above the threshold are labeled automatically, while others are sent to experts.\nThe guarantee states that, with high probability, the global labeling error is ≤ ε.\nExtensions include a multi-model PAC Router and a practical multicalibration procedure to improve uncertainty reliability.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "1. Relevant and well-motivated problem formulation. Applying PAC reasoning to automatic labeling is timely and practically meaningful.  \n   It formalizes a common heuristic (confidence-based filtering) into a statistically grounded process.  \n\n2. Simplicity and generality. The proposed approach is model-agnostic and directly applicable across diverse labeling pipelines.  \n\n3. Comprehensive empirical coverage. Experiments span multiple modalities and show consistent compliance with PAC bounds while reducing expert effort.  \n\n4. Clear presentation and awareness of practical issues.   The inclusion of multicalibration demonstrates attention to real-world miscalibration.", "weaknesses": "1. Limited overall contribution despite novel framing.\nWhile the paper introduces new formulations (e.g., PAC Labeling and PAC Router) that are conceptually fresh, the underlying theoretical substance remains limited.\nThe analysis primarily builds on existing mean-upper-bound PAC results without introducing new bounds, assumptions, or insights into the nature of uncertainty in labeling.\nConsequently, although the problem setting is well-motivated, the contribution lies more in repackaging and integration than in theoretical advancement.\n\n\n2. Disconnect between calibration and PAC validity.\nMulticalibration empirically adjusts uncertainty scores but may violate the monotonicity assumption required for PAC reasoning.  \n   The paper provides no formal analysis showing that PAC guarantees still hold after calibration.\n\n3. Lack of intuitive and competitive comparisons.\nAlthough experiments are broad, they do not clearly show cost–error trade-offs (e.g., comparing methods under equal error rates).  \n   Baselines are limited to simple heuristics; stronger competitors like conformal labeling or active selection are missing.  \n   As a result, the empirical advantage remains qualitative rather than quantitative.", "questions": "1. What concrete theoretical or methodological innovations are introduced beyond adapting existing PAC mean-upper-bound results?  \n\n2. Does multicalibration theoretically preserve the assumptions required for PAC guarantees, or is it purely an empirical fix?  \n\n3. Have the authors evaluated cost–error trade-offs, e.g., comparing expert costs at equal error levels across methods?  \n\n4. How does the framework behave under correlated errors or severe miscalibration beyond what multicalibration can correct?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761926100716}, {"id": "WFxDptlYSh", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17773/Reviewer_R54A"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper proposes a method to create high-quality labels by leveraging cheap predictions from AI models alongside costly expert labels. The method has statistical guarantee: with high probability, the final labeled dataset's error will not exceed a user-specified threshold. For data points labeled by AI models, only those with uncertainty exceeds a threshold are sent for expert labeling, minimizing overall cost. The paper also outlines a more complex PAC Router for optimally selecting between multiple AI models and presents experiments demonstrating budget savings.", "review_text": "The paper proposes a method to create high-quality labels by leveraging cheap predictions from AI models alongside costly expert labels. The method has statistical guarantee: with high probability, the final labeled dataset's error will not exceed a user-specified threshold. For data points labeled by AI models, only those with uncertainty exceeds a threshold are sent for expert labeling, minimizing overall cost. The paper also outlines a more complex PAC Router for optimally selecting between multiple AI models and presents experiments demonstrating budget savings.", "strengths": "- The paper considers a very practical setup with AI models generating labels for all data and then experts annotating the most valuable subset. This is helpful in generating labeled training data in low resource domains.\n- The proposed method has nice statistical guarantee that with high probability, the final labeled dataset's error will not exceed a user-specified threshold.\n- The effectiveness of the method in terms of cost saving is demonstrated empirically.", "weaknesses": "- In practice, the cost of high quality AI models is not neglectable, especially given the existence of test time scaling. Wonder how does this affect the proposed approach? Also wonder if allocating the whole expert budget to AI models test time scaling achieves better results?\n\n- The considered baselines are too naive. A natural baseline is to use active learning. For example, first initialize the active learning model to be the model trained on labels generated by AI models, then select data points for experts to label using a standard active learning process. Also, a better baseline than the second baseline is to get more labels from AI models (with different prompts or AI models) until the cost matches the proposed method, and then aggregate the labels (e.g. with existing methods from the crowd sourcing literature).", "questions": "- how does the method compare to active learning or aggregated labels from many AI models (with different prompts and models).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method to create high-quality labels by leveraging cheap predictions from AI models alongside costly expert labels. The method has statistical guarantee: with high probability, the final labeled dataset's error will not exceed a user-specified threshold. For data points labeled by AI models, only those with uncertainty exceeds a threshold are sent for expert labeling, minimizing overall cost. The paper also outlines a more complex PAC Router for optimally selecting between multiple AI models and presents experiments demonstrating budget savings.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper considers a very practical setup with AI models generating labels for all data and then experts annotating the most valuable subset. This is helpful in generating labeled training data in low resource domains.\n- The proposed method has nice statistical guarantee that with high probability, the final labeled dataset's error will not exceed a user-specified threshold.\n- The effectiveness of the method in terms of cost saving is demonstrated empirically.", "weaknesses": "- In practice, the cost of high quality AI models is not neglectable, especially given the existence of test time scaling. Wonder how does this affect the proposed approach? Also wonder if allocating the whole expert budget to AI models test time scaling achieves better results?\n\n- The considered baselines are too naive. A natural baseline is to use active learning. For example, first initialize the active learning model to be the model trained on labels generated by AI models, then select data points for experts to label using a standard active learning process. Also, a better baseline than the second baseline is to get more labels from AI models (with different prompts or AI models) until the cost matches the proposed method, and then aggregate the labels (e.g. with existing methods from the crowd sourcing literature).", "questions": "- how does the method compare to active learning or aggregated labels from many AI models (with different prompts and models).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761710864279}, {"id": "NgmT1m9Ij2", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17773/Reviewer_rfSa"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "This paper addresses the high cost of building high-quality labeled datasets. The authors claim that while many labeling methods exist, they have a critical limitation of providing no guarantees on accuracy. Therefore, the paper tackles the problem of how to guarantee the error rate, proposing a methodology called \"PAC labeling.\" The method first uses AI labeling, classifies labels based on confidence, and then decides whether to use human labeling based on a certain threshold. The authors state this is done under minimal statistical assumptions. As a result, they achieve high performance compared to an \"AI only\" baseline and also show better performance than a fixed-threshold baseline.", "review_text": "This paper addresses the high cost of building high-quality labeled datasets. The authors claim that while many labeling methods exist, they have a critical limitation of providing no guarantees on accuracy. Therefore, the paper tackles the problem of how to guarantee the error rate, proposing a methodology called \"PAC labeling.\" The method first uses AI labeling, classifies labels based on confidence, and then decides whether to use human labeling based on a certain threshold. The authors state this is done under minimal statistical assumptions. As a result, they achieve high performance compared to an \"AI only\" baseline and also show better performance than a fixed-threshold baseline.", "strengths": "This paper addresses an important problem and is well-timed with the current trend of synthetic data labeling, particularly in its focus on providing statistical guarantees.", "weaknesses": "The biggest problem is that the baseline comparison feels contrived, and the paper's contribution seems overstated. The authors claim that existing methods fail to provide mathematical guarantees, but such methodologies clearly exist. For example, Conformal Prediction (CP), which the authors themselves mention, deals with a very similar problem. The CP methodology quantifies prediction uncertainty into a prediction set with guaranteed coverage; it statistically guarantees the probability that the true label is within this set. Can this not be used to solve the same problem? The same applies to papers like \"Learning to Defer...\". This methodology also solves the same problem. Why did the authors not compare against these methods and evaluate their contributions toe-to-toe?", "questions": "See the Weakness section. Why were direct competitors that also provide statistical guarantees such as Conformal Prediction and Learning to Defer, omitted from the experimental comparison, and why was their contribution not evaluated \"toe-to-toe\"?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the high cost of building high-quality labeled datasets. The authors claim that while many labeling methods exist, they have a critical limitation of providing no guarantees on accuracy. Therefore, the paper tackles the problem of how to guarantee the error rate, proposing a methodology called \"PAC labeling.\" The method first uses AI labeling, classifies labels based on confidence, and then decides whether to use human labeling based on a certain threshold. The authors state this is done under minimal statistical assumptions. As a result, they achieve high performance compared to an \"AI only\" baseline and also show better performance than a fixed-threshold baseline.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "This paper addresses an important problem and is well-timed with the current trend of synthetic data labeling, particularly in its focus on providing statistical guarantees.", "weaknesses": "The biggest problem is that the baseline comparison feels contrived, and the paper's contribution seems overstated. The authors claim that existing methods fail to provide mathematical guarantees, but such methodologies clearly exist. For example, Conformal Prediction (CP), which the authors themselves mention, deals with a very similar problem. The CP methodology quantifies prediction uncertainty into a prediction set with guaranteed coverage; it statistically guarantees the probability that the true label is within this set. Can this not be used to solve the same problem? The same applies to papers like \"Learning to Defer...\". This methodology also solves the same problem. Why did the authors not compare against these methods and evaluate their contributions toe-to-toe?", "questions": "See the Weakness section. Why were direct competitors that also provide statistical guarantees such as Conformal Prediction and Learning to Defer, omitted from the experimental comparison, and why was their contribution not evaluated \"toe-to-toe\"?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1761702171049}, {"id": "asj2AEEOU3", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission17773/Reviewer_sfRp"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper tackles the problem that building high-quality labeled datasets is costly, and while leveraging AI model predictions can reduce costs, the lack of accuracy guarantees makes full replacement difficult. It proposes a mathematical framework to label datasets using AI models, mathematically bounding the error below epsilon and the confidence above 1-a. Uncertainty scores play a central role in this process, where expert labels are collected only for data exceeding an uncertainty threshold u^. To ensure precise uncertainty measurement, calibration correcting uncertainty per data cluster is performed. The authors experimentally verify the broad effectiveness of the PAC approach across diverse domains including natural language, images, and proteins.", "review_text": "This paper tackles the problem that building high-quality labeled datasets is costly, and while leveraging AI model predictions can reduce costs, the lack of accuracy guarantees makes full replacement difficult. It proposes a mathematical framework to label datasets using AI models, mathematically bounding the error below epsilon and the confidence above 1-a. Uncertainty scores play a central role in this process, where expert labels are collected only for data exceeding an uncertainty threshold u^. To ensure precise uncertainty measurement, calibration correcting uncertainty per data cluster is performed. The authors experimentally verify the broad effectiveness of the PAC approach across diverse domains including natural language, images, and proteins.", "strengths": "- Provides a formal mathematical framework to set an appropriate uncertainty threshold between expert labels and model-generated labels, with a rigorously defined iid variable within this framework.  \n- Clearly points out the critical importance of precise uncertainty measurement to guarantee the quality of AI-generated labels, enhancing uncertainty estimation with calibration techniques.  \n- Demonstrates through experiments on various data modalities—natural language, image, and protein structure—that the PAC method is a broadly applicable algorithm across multiple modalities and models.", "weaknesses": "- The uncertainty threshold is set universally across the dataset, but actual model inferences often show variations in uncertainty by class, due to reasons such as insufficient training samples for specific classes or confusing classes. Has the approach of using class-specific thresholds for more efficient labeling been considered?  \n- How is the value of m determined? This seems like a very critical hyperparameter for the method. If this value is heuristically determined, it would be difficult to consider the algorithm as robustly functioning. If it is described somewhere in the paper that I missed, it would be helpful to be informed.\n- From the experimental results, it is hard to say that PAC brings a meaningful gain. In Figure 1, the PAC results lie between naive thresholding and AI-only interpolation, suggesting the PAC method mainly adjusts the trade-off between error and budget, rather than achieving gains beyond existing trade-off lines. If the authors’ intent is to enable trade-off adjustment using PAC, it makes sense, but this can partially be achieved just by constant threshold adjustment, limiting the contribution.  \n- There is no conclusion section, and therefore no mention of limitations and future work. Usually, limitations and future work part is considered as an important part of a paper, so this omission is a clear weakness.", "questions": "- As model overparameterization increases, over-confidence in model predictions worsens. Are there any countermeasures proposed for this?  \n- Depending on the reader’s background, some may find the format of protein folding labels unfamiliar. Would it improve the completeness of the paper to briefly explain the label format of the various domains you used in the experiment section?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper tackles the problem that building high-quality labeled datasets is costly, and while leveraging AI model predictions can reduce costs, the lack of accuracy guarantees makes full replacement difficult. It proposes a mathematical framework to label datasets using AI models, mathematically bounding the error below epsilon and the confidence above 1-a. Uncertainty scores play a central role in this process, where expert labels are collected only for data exceeding an uncertainty threshold u^. To ensure precise uncertainty measurement, calibration correcting uncertainty per data cluster is performed. The authors experimentally verify the broad effectiveness of the PAC approach across diverse domains including natural language, images, and proteins.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- Provides a formal mathematical framework to set an appropriate uncertainty threshold between expert labels and model-generated labels, with a rigorously defined iid variable within this framework.  \n- Clearly points out the critical importance of precise uncertainty measurement to guarantee the quality of AI-generated labels, enhancing uncertainty estimation with calibration techniques.  \n- Demonstrates through experiments on various data modalities—natural language, image, and protein structure—that the PAC method is a broadly applicable algorithm across multiple modalities and models.", "weaknesses": "- The uncertainty threshold is set universally across the dataset, but actual model inferences often show variations in uncertainty by class, due to reasons such as insufficient training samples for specific classes or confusing classes. Has the approach of using class-specific thresholds for more efficient labeling been considered?  \n- How is the value of m determined? This seems like a very critical hyperparameter for the method. If this value is heuristically determined, it would be difficult to consider the algorithm as robustly functioning. If it is described somewhere in the paper that I missed, it would be helpful to be informed.\n- From the experimental results, it is hard to say that PAC brings a meaningful gain. In Figure 1, the PAC results lie between naive thresholding and AI-only interpolation, suggesting the PAC method mainly adjusts the trade-off between error and budget, rather than achieving gains beyond existing trade-off lines. If the authors’ intent is to enable trade-off adjustment using PAC, it makes sense, but this can partially be achieved just by constant threshold adjustment, limiting the contribution.  \n- There is no conclusion section, and therefore no mention of limitations and future work. Usually, limitations and future work part is considered as an important part of a paper, so this omission is a clear weakness.", "questions": "- As model overparameterization increases, over-confidence in model predictions worsens. Are there any countermeasures proposed for this?  \n- Depending on the reader’s background, some may find the format of protein folding labels unfamiliar. Would it improve the completeness of the paper to briefly explain the label format of the various domains you used in the experiment section?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761277845105}], "openreview_url": "https://openreview.net/forum?id=RYwtJyOP3k", "arxiv_id": "2506.10908", "paper_pdf": "papers/RYwtJyOP3k.pdf", "paper_pdf_sha256": "f0e07a02481cec4b5b435452eb445f152afc99067218518f8d89203bffc1fbb0", "paper_pdf_bytes": 1030000, "paper_pdf_source": "openreview", "code_url": "https://github.com/tijana-zrnic/pac-labels", "code_repository": "tijana-zrnic/pac-labels", "code_commit": "b415b58756b14b384529ac9cf146bd5d4c8139aa", "code_archive": "repos/RYwtJyOP3k.zip", "code_archive_sha256": "38020d18668956f268bd2520397262c99c7ed7858bec0e9a31160cbcbf512dbc", "code_archive_bytes": 12267016, "code_file_count": 6, "code_extensions": {".ipynb": 3, ".py": 3}, "github_disk_usage_kb": 7615, "github_languages": {"Jupyter Notebook": 28007, "Python": 20683}, "github_archived": false, "github_pushed_at": "2025-06-13T06:30:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/probably-approximately-correct-labels"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "leSbzBtofH", "year": 2025, "status": "rejected", "title": "AutoAdvExBench: Benchmarking Autonomous Exploitation of Adversarial Example Defenses", "authors": ["Nicholas Carlini", "Edoardo Debenedetti", "Javier Rando", "Milad Nasr", "Florian Tramèr"], "authorids": ["~Nicholas_Carlini1", "~Edoardo_Debenedetti1", "~Javier_Rando2", "~Milad_Nasr2", "~Florian_Tramèr1"], "authors_source": "OpenReview API", "abstract": "We introduce AutoAdvExBench, a benchmark to evaluate if large language models (LLMs)\ncan autonomously exploit defenses to adversarial examples.\nWe believe our benchmark will be valuable to several distinct audiences. \nFirst, it measures if models can match the abilities of expert adversarial machine learning researchers.\nSecond, it serves as a challenging evaluation for reasoning capabilities that\ncan measure LLMs' ability to understand and interact with sophisticated codebases. \nAnd third, \nsince many adversarial examples defenses have been broken in the past,\nthis benchmark allows for evaluating the ability of LLMs to reproduce\nprior research results automatically.\nWe then benchmark the ability of current LLMs to solve this benchmark,\nand find most are unable to succeed.\nOur strongest agent, with a human-guided prompt,\nis only able to successfully generate adversarial examples on 6 of the 51 defenses in our benchmark.\nThis benchmark is publicly accessible at redacted for review.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "DgSMYSftzt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_G63v"], "rating": 8, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "This paper proposes a new benchmark to test LLM capabilities: whether they can generate adversarial attacks to proposed adversarial defenses. The authors crawled arXiv and filtered the papers to find adversarial defense methods with easily reproducible code. Current state-of-the-art LLMs were tested on this new benchmark, and do not perform well.", "review_text": "This paper proposes a new benchmark to test LLM capabilities: whether they can generate adversarial attacks to proposed adversarial defenses. The authors crawled arXiv and filtered the papers to find adversarial defense methods with easily reproducible code. Current state-of-the-art LLMs were tested on this new benchmark, and do not perform well.", "strengths": "The authors propose a novel benchmark to test the capabilities of LLMs. The proposed task is a real-world research/security setting, where the \"correct\" answer may not even be known, and yet there is a quantitative measurement that can be extracted to evaluate the abilities of the LLM. As a result, this provides a benchmark which may still be useful even if a model has surpassed human level performance in this domain. The authors are upfront about the many limitations of the benchmark.", "weaknesses": "- The benchmark is evaluated in a limited setting. Even though to be successful, the model must be proficient in several different domains, the scope of the task is fairly limited.\n- As mentioned, benchmark contamination is a potential issue, especially when considering the use of this benchmark well into the future.\n- The benchmark does not appear to provide a meaningful continuous measure for current LLMs ability to generate novel attacks. Instead, it seems limited to whether they can successfully implement a known attack, as all of their current limited success on the benchmark is due to benchmark contamination.", "questions": "- As the models are currently not producing novel adversarial attacks, what does this benchmark measure that is not covered by existing benchmarks? Is the utility of this benchmark primarily for much more capable LLMs?\n- How often do you expect to update the benchmark? If it is every time a new attack to a defense is published, then would this make it difficult to compare models trained at different points in time? If not, then might the benchmark provide an illusion of progress, even if it is just due of benchmark contamination?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new benchmark to test LLM capabilities: whether they can generate adversarial attacks to proposed adversarial defenses. The authors crawled arXiv and filtered the papers to find adversarial defense methods with easily reproducible code. Current state-of-the-art LLMs were tested on this new benchmark, and do not perform well.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "The authors propose a novel benchmark to test the capabilities of LLMs. The proposed task is a real-world research/security setting, where the \"correct\" answer may not even be known, and yet there is a quantitative measurement that can be extracted to evaluate the abilities of the LLM. As a result, this provides a benchmark which may still be useful even if a model has surpassed human level performance in this domain. The authors are upfront about the many limitations of the benchmark.", "weaknesses": "- The benchmark is evaluated in a limited setting. Even though to be successful, the model must be proficient in several different domains, the scope of the task is fairly limited.\n- As mentioned, benchmark contamination is a potential issue, especially when considering the use of this benchmark well into the future.\n- The benchmark does not appear to provide a meaningful continuous measure for current LLMs ability to generate novel attacks. Instead, it seems limited to whether they can successfully implement a known attack, as all of their current limited success on the benchmark is due to benchmark contamination.", "questions": "- As the models are currently not producing novel adversarial attacks, what does this benchmark measure that is not covered by existing benchmarks? Is the utility of this benchmark primarily for much more capable LLMs?\n- How often do you expect to update the benchmark? If it is every time a new attack to a defense is published, then would this make it difficult to compare models trained at different points in time? If not, then might the benchmark provide an illusion of progress, even if it is just due of benchmark contamination?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731055271240}, {"id": "r7jvdbcLIP", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_uSMe"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "This paper proposes a benchmark, AutoAdvExBench, to measure large language models' (LLMs) ability to exploit other AI systems. To be specific, AutoAdvExBench evaluates LLMs' ability to construct adversarial examples that bypass the corresponding defense methods. Experimental results demonstrate that AutoAdvExBench is challenging with a low attack success rate: only 6 of the 51 defenses are successfully attacked by their strongest agent.", "review_text": "This paper proposes a benchmark, AutoAdvExBench, to measure large language models' (LLMs) ability to exploit other AI systems. To be specific, AutoAdvExBench evaluates LLMs' ability to construct adversarial examples that bypass the corresponding defense methods. Experimental results demonstrate that AutoAdvExBench is challenging with a low attack success rate: only 6 of the 51 defenses are successfully attacked by their strongest agent.", "strengths": "+ Measuring AI's ability to exploit other AI systems is necessary and meaningful for preparing for future safety risks. Automatic AI exploitation becomes feasible due to the automated nature of the AI agent system and could result in catastrophic risks. This paper provides a realistic implementation for such a speculative threat model. Therefore, it will be a good proxy to monitor the progress of AI and prepare for the possible safety risks. \n\n+ The construction process of AutoAdvExBench is solid: this paper collects reproducible and diverse defense papers with manual checks to compose their benchmark, which will provide a strong basis for measuring AI exploitation ability.", "weaknesses": "- **The evaluation metrics lack comprehensiveness.** The paper only reports robust accuracy and the number of successful attacks as metrics, neglecting more detailed analyses of the various capabilities agents need to overcome the benchmark. Areas like code comprehension, code completion, long-context understanding, and the novelty of proposed ideas are overlooked. For example, while the paper finds that only a quarter of defenses can be made differentiable (line 480)—a necessary step before designing new attacks—it’s unclear which capabilities agents lack that lead to these limitations. Simply reporting post-attack robust accuracy does not reveal which specific capabilities are bottlenecks, nor does it offer insights into where current models fall short.\n\n- **Data contamination is a significant concern** that the authors have not adequately addressed. They do not provide empirical results to demonstrate the extent of data contamination in their benchmark. Although they state (line 262) that allowing agents access to paper data does not significantly improve success rates, this does not definitively rule out contamination. Furthermore, as the benchmark does not use the most advanced agent frameworks, it’s conceivable that more sophisticated agents could leverage memorized information to generate attack codes if contamination were present, raising questions about the benchmark’s validity and challenge level.\n\n- **The captions of each table and figure lack essential details, making them difficult to follow.** For instance, in Figure 2, it’s unclear how the caption’s statement, \"each line plots the number of defenses that reduce the robust accuracy to a given level,\" corresponds to the curves in the figure. Similarly, in Table 3, the meanings of terms such as \"forward pass,\" \"differentiable,\" and \"FGSM attack\" and the respective numbers are not fully explained. This also seems inconsistent with the analysis of FGSM in the main text (line 486), where an accuracy rate of 84% is mentioned.\n\n- **The benchmark does not leverage the latest agent frameworks, such as multi-agent systems with specialized roles, which could potentially address the challenges posed by the benchmark.** The coding capabilities of the agents used in the paper seem quite limited (see Table 3), suggesting that the benchmark may not be challenging enough for cutting-edge agent models.", "questions": "- **The analysis of diversity among selected defense papers is insufficient.** Diversity is a crucial metric for this benchmark, yet the distinctions between each defense paper are not clearly highlighted (lines 224–230). It would be helpful if the authors could highlight the unique aspects of each defense paper and delineate their differences.\n\nI would like to raise my score if the authors could address my concerns and questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a benchmark, AutoAdvExBench, to measure large language models' (LLMs) ability to exploit other AI systems. To be specific, AutoAdvExBench evaluates LLMs' ability to construct adversarial examples that bypass the corresponding defense methods. Experimental results demonstrate that AutoAdvExBench is challenging with a low attack success rate: only 6 of the 51 defenses are successfully attacked by their strongest agent.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "+ Measuring AI's ability to exploit other AI systems is necessary and meaningful for preparing for future safety risks. Automatic AI exploitation becomes feasible due to the automated nature of the AI agent system and could result in catastrophic risks. This paper provides a realistic implementation for such a speculative threat model. Therefore, it will be a good proxy to monitor the progress of AI and prepare for the possible safety risks. \n\n+ The construction process of AutoAdvExBench is solid: this paper collects reproducible and diverse defense papers with manual checks to compose their benchmark, which will provide a strong basis for measuring AI exploitation ability.", "weaknesses": "- **The evaluation metrics lack comprehensiveness.** The paper only reports robust accuracy and the number of successful attacks as metrics, neglecting more detailed analyses of the various capabilities agents need to overcome the benchmark. Areas like code comprehension, code completion, long-context understanding, and the novelty of proposed ideas are overlooked. For example, while the paper finds that only a quarter of defenses can be made differentiable (line 480)—a necessary step before designing new attacks—it’s unclear which capabilities agents lack that lead to these limitations. Simply reporting post-attack robust accuracy does not reveal which specific capabilities are bottlenecks, nor does it offer insights into where current models fall short.\n\n- **Data contamination is a significant concern** that the authors have not adequately addressed. They do not provide empirical results to demonstrate the extent of data contamination in their benchmark. Although they state (line 262) that allowing agents access to paper data does not significantly improve success rates, this does not definitively rule out contamination. Furthermore, as the benchmark does not use the most advanced agent frameworks, it’s conceivable that more sophisticated agents could leverage memorized information to generate attack codes if contamination were present, raising questions about the benchmark’s validity and challenge level.\n\n- **The captions of each table and figure lack essential details, making them difficult to follow.** For instance, in Figure 2, it’s unclear how the caption’s statement, \"each line plots the number of defenses that reduce the robust accuracy to a given level,\" corresponds to the curves in the figure. Similarly, in Table 3, the meanings of terms such as \"forward pass,\" \"differentiable,\" and \"FGSM attack\" and the respective numbers are not fully explained. This also seems inconsistent with the analysis of FGSM in the main text (line 486), where an accuracy rate of 84% is mentioned.\n\n- **The benchmark does not leverage the latest agent frameworks, such as multi-agent systems with specialized roles, which could potentially address the challenges posed by the benchmark.** The coding capabilities of the agents used in the paper seem quite limited (see Table 3), suggesting that the benchmark may not be challenging enough for cutting-edge agent models.", "questions": "- **The analysis of diversity among selected defense papers is insufficient.** Diversity is a crucial metric for this benchmark, yet the distinctions between each defense paper are not clearly highlighted (lines 224–230). It would be helpful if the authors could highlight the unique aspects of each defense paper and delineate their differences.\n\nI would like to raise my score if the authors could address my concerns and questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730718428133}, {"id": "k8e3jnNf1U", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_qJjZ"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes an agentic benchmark for LLMs where the task is to automatically break various adversarial robustness defenses. It includes implementations and PDFs for a large number of adversarial robustness defenses, some of which have published vulnerabilities. The goal of the LLM agent is to output 1000 perturbed images under a standard \\ell_{\\infty} bound that break the defenses.", "review_text": "This paper proposes an agentic benchmark for LLMs where the task is to automatically break various adversarial robustness defenses. It includes implementations and PDFs for a large number of adversarial robustness defenses, some of which have published vulnerabilities. The goal of the LLM agent is to output 1000 perturbed images under a standard \\ell_{\\infty} bound that break the defenses.", "strengths": "- This is a clever idea for an agentic benchmark. The task is complex but easy to evaluate, and it provides a way to measure how useful LLM agents could be for stress-testing defenses proposed by the ML community, which I think is an interesting future use case of AI agents.\n- The work is timely. Multiple new agentic benchmarks have been proposed recently, including SWE-bench and MLE-bench. This paper continues that line of benchmarking work, but with an emphasis on automated stress-testing for adversarial training defenses.\n- The writing is clear.\n- The benchmark seems well-designed. The task is easy to understand. Useful data was curated to enable the agents to perform the task (paper code and PDFs).\n- The baseline evaluations show that the task is tractable.", "weaknesses": "For defenses with published vulnerabilities, it would be good to include a check for whether the model is aware of these vulnerabilities or discovers them from scratch. I realize this isn't relevant to current models, since they aren't very good yet, but it could be an interesting thing to check in future models. I see your response in lines 261-262. This is just a comment that the paper would be stronger with proactive measures to address this.", "questions": "No questions for now.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an agentic benchmark for LLMs where the task is to automatically break various adversarial robustness defenses. It includes implementations and PDFs for a large number of adversarial robustness defenses, some of which have published vulnerabilities. The goal of the LLM agent is to output 1000 perturbed images under a standard \\ell_{\\infty} bound that break the defenses.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- This is a clever idea for an agentic benchmark. The task is complex but easy to evaluate, and it provides a way to measure how useful LLM agents could be for stress-testing defenses proposed by the ML community, which I think is an interesting future use case of AI agents.\n- The work is timely. Multiple new agentic benchmarks have been proposed recently, including SWE-bench and MLE-bench. This paper continues that line of benchmarking work, but with an emphasis on automated stress-testing for adversarial training defenses.\n- The writing is clear.\n- The benchmark seems well-designed. The task is easy to understand. Useful data was curated to enable the agents to perform the task (paper code and PDFs).\n- The baseline evaluations show that the task is tractable.", "weaknesses": "For defenses with published vulnerabilities, it would be good to include a check for whether the model is aware of these vulnerabilities or discovers them from scratch. I realize this isn't relevant to current models, since they aren't very good yet, but it could be an interesting thing to check in future models. I see your response in lines 261-262. This is just a comment that the paper would be stronger with proactive measures to address this.", "questions": "No questions for now.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730686670960}, {"id": "wOw3fAzbOq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_Xffd"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper introduces AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously break defenses to adversarial examples, given access to defense (a) the paper describing the defense, (b) the source code of the defense, (c) a correct forward pass implementation of the defense, (d) a perturbation bound, and (e) 1,000 images that should be attacked. The authors attempt both an end-to-end approach and a four-step, human-guided approach. Based on the low success rates, the authors conclude that current language models cannot autonomously break most adversarial example defenses.", "review_text": "This paper introduces AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously break defenses to adversarial examples, given access to defense (a) the paper describing the defense, (b) the source code of the defense, (c) a correct forward pass implementation of the defense, (d) a perturbation bound, and (e) 1,000 images that should be attacked. The authors attempt both an end-to-end approach and a four-step, human-guided approach. Based on the low success rates, the authors conclude that current language models cannot autonomously break most adversarial example defenses.", "strengths": "-  An interesting attempt to test LLMs on a well-established security problem, i.e., defenses against adversarial examples. The authors have provided strong motivations for proposing such a new benchmark.\n\n- The paper is written well, clearly stating the contributions beyond the literature and the significance. \n\n- The fact that current LLMs largely fail in solving such a well-defined problem is somewhat surprising and so calls for future work.", "weaknesses": "Presentation:\n\n1. The reviewer appreciates that the authors have put much effort into describing the limitations, even as early as in Section 3.3. However, several points in Section 3.3 Limitations and Section 3.1 Motivation are actually from the same perspective but do not well connect. Specifically, as the reviewer understands, although the benchmark is “difficult” and “security-relevant”, ”Adversarial examples attacks are not representative of common security exploits”. Although the benchmark is “Messy”, its “Research code is not representative of production code”. It may help if the authors could first present the perspectives, and then talk about the motivations/advantages and limitations of the benchmark in each perspective.\n\n2. Throughout the paper, only Figure 2 was referred to in the main text, making the more summarized information in tables/figures not helpful for understanding the paper. The authors should explicitly reference relevant tables and figures when discussing results in the main text. This would help readers connect the discussion to the supporting data more easily.\n\n3. The term “successfully attacked” is not defined before use. For example, it appears in the second to last sentence of the Abstract and the caption of Table 3. In addition, it is not clear what the numbers in Table 3 mean. \n\n\nExperiment:\n\n1. Although it is OK to show current LLMs are not good at solving the new benchmark, there is no exploration of potential ways to improve them. For example, It is interesting that current LLMs still struggle with simple operations, e.g., “models were only able to implement a differentiable forward pass in 23% of cases”. However, the authors have not attempted to solve this and have not even discussed potential solutions. They could include ideas for targeted training and improved prompting strategies.\n\n2. The authors state that “We impose no time restriction, on the number of unsuccessful attempts an adversary makes, on the runtime of the algorithm, or on the cost of the attack. However, we strongly encourage reporting these numbers so that future work will be able to draw comparisons between methods that are exceptionally expensive to run, and methods that are cheaper.” The reviewer does not get why they have not reported these numbers in this paper (which seems not to require too much additional effort).", "questions": "Please see the above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously break defenses to adversarial examples, given access to defense (a) the paper describing the defense, (b) the source code of the defense, (c) a correct forward pass implementation of the defense, (d) a perturbation bound, and (e) 1,000 images that should be attacked. The authors attempt both an end-to-end approach and a four-step, human-guided approach. Based on the low success rates, the authors conclude that current language models cannot autonomously break most adversarial example defenses.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "-  An interesting attempt to test LLMs on a well-established security problem, i.e., defenses against adversarial examples. The authors have provided strong motivations for proposing such a new benchmark.\n\n- The paper is written well, clearly stating the contributions beyond the literature and the significance. \n\n- The fact that current LLMs largely fail in solving such a well-defined problem is somewhat surprising and so calls for future work.", "weaknesses": "Presentation:\n\n1. The reviewer appreciates that the authors have put much effort into describing the limitations, even as early as in Section 3.3. However, several points in Section 3.3 Limitations and Section 3.1 Motivation are actually from the same perspective but do not well connect. Specifically, as the reviewer understands, although the benchmark is “difficult” and “security-relevant”, ”Adversarial examples attacks are not representative of common security exploits”. Although the benchmark is “Messy”, its “Research code is not representative of production code”. It may help if the authors could first present the perspectives, and then talk about the motivations/advantages and limitations of the benchmark in each perspective.\n\n2. Throughout the paper, only Figure 2 was referred to in the main text, making the more summarized information in tables/figures not helpful for understanding the paper. The authors should explicitly reference relevant tables and figures when discussing results in the main text. This would help readers connect the discussion to the supporting data more easily.\n\n3. The term “successfully attacked” is not defined before use. For example, it appears in the second to last sentence of the Abstract and the caption of Table 3. In addition, it is not clear what the numbers in Table 3 mean. \n\n\nExperiment:\n\n1. Although it is OK to show current LLMs are not good at solving the new benchmark, there is no exploration of potential ways to improve them. For example, It is interesting that current LLMs still struggle with simple operations, e.g., “models were only able to implement a differentiable forward pass in 23% of cases”. However, the authors have not attempted to solve this and have not even discussed potential solutions. They could include ideas for targeted training and improved prompting strategies.\n\n2. The authors state that “We impose no time restriction, on the number of unsuccessful attempts an adversary makes, on the runtime of the algorithm, or on the cost of the attack. However, we strongly encourage reporting these numbers so that future work will be able to draw comparisons between methods that are exceptionally expensive to run, and methods that are cheaper.” The reviewer does not get why they have not reported these numbers in this paper (which seems not to require too much additional effort).", "questions": "Please see the above weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730291475302}, {"id": "ywbqDBFvSF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_enyd"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper introduces AutoAdvExBench, a benchmark designed to evaluate the ability of large language models (LLMs) to attack adversarial example defenses. The benchmark includes 51 defenses from 37 papers, and the LLMs are provided with the research paper and the corresponding code for the forward pass of the defense.\nThe authors test state-of-the-art LLMs like Claude 3.5-Sonnet and GPT-4o in various scenarios, finding that current models struggle significantly. Zero-shot attempts were entirely unsuccessful, and even with iterative debugging, only a few defenses were successfully attacked. \nThe paper highlights the challenges LLMs face in automating security tasks, emphasizing their limitations in understanding and exploiting such defenses.", "review_text": "The paper introduces AutoAdvExBench, a benchmark designed to evaluate the ability of large language models (LLMs) to attack adversarial example defenses. The benchmark includes 51 defenses from 37 papers, and the LLMs are provided with the research paper and the corresponding code for the forward pass of the defense.\nThe authors test state-of-the-art LLMs like Claude 3.5-Sonnet and GPT-4o in various scenarios, finding that current models struggle significantly. Zero-shot attempts were entirely unsuccessful, and even with iterative debugging, only a few defenses were successfully attacked. \nThe paper highlights the challenges LLMs face in automating security tasks, emphasizing their limitations in understanding and exploiting such defenses.", "strengths": "1. The authors invested significant effort in filtering relevant papers and collecting reproducible code, resulting in a comprehensive and valuable dataset of 51 adversarial defenses from 37 papers. This rigorous collection process enhances the benchmark's credibility and relevance to the research community.\n\n2. The benchmark setup involves working with \"messy\" research codebases, which closely reflects real-world challenges in adversarial defense scenarios. This realistic approach adds significant value, as it pushes current LLMs to handle complex, imperfect code environments typical in real applications.\n\n3. The paper employs definitive evaluation methods that do not rely on black-box metrics, such as LLM-as-a-judge approaches. Instead, it uses measurable, transparent performance metrics, ensuring that the results are both interpretable and replicable.", "weaknesses": "1. The paper lacks detailed examples of interactions with the LLMs, such as specific prompts used or a complete end-to-end example of at least one of the defenses and its corresponding developed attack stages. Including visual aids or detailed walk-throughs of successful or failed attacks could greatly enhance readability and engagement. The absence of such details makes the content dense leading to a less engaging reading experience.\n\n2. The current experimental settings are overly general for the LLMs, so it's unsurprising to observe such low success rates. From my experience, even for simpler coding tasks, I find that even the best LLMs require significantly more detail—such as an initial code draft, context explanations, and multiple rounds of conversation and reflection, as in a chat scenario—to achieve the desired results. For example, would a 6-turn interaction (interactive human-AI chat session) with these models (like GitHub Copilot settings) yield much better results? \n\n3. The paper could have explored more powerful methods, such as ICL few-shot examples, fine-tuning, or retrieval-augmented generation (RAG), to better evaluate LLMs' capabilities in solving the benchmark. These approaches could have provided valuable insights into the strengths and limitations of LLMs when leveraging different training and interaction methodologies. The current experimental settings are necessary but insufficient.\n\n4. This is a coding task, yet the paper's background lacks information on state-of-the-art LLMs specifically trained for coding. You've tested with GPT-4o and Claude 3.5 Sonnet, but are there other LLMs worth experimenting with? \n\n5. In Section 5.1 (End-to-End Evaluation), in the sentence \"Unsurprisingly, we find that current defenses fail completely at this task ...\" do you mean \"current models\"?", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["Yes, Privacy, security and safety", "Yes, Potentially harmful insights, methodologies and applications"], "all_content": {"summary": "The paper introduces AutoAdvExBench, a benchmark designed to evaluate the ability of large language models (LLMs) to attack adversarial example defenses. The benchmark includes 51 defenses from 37 papers, and the LLMs are provided with the research paper and the corresponding code for the forward pass of the defense.\nThe authors test state-of-the-art LLMs like Claude 3.5-Sonnet and GPT-4o in various scenarios, finding that current models struggle significantly. Zero-shot attempts were entirely unsuccessful, and even with iterative debugging, only a few defenses were successfully attacked. \nThe paper highlights the challenges LLMs face in automating security tasks, emphasizing their limitations in understanding and exploiting such defenses.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. The authors invested significant effort in filtering relevant papers and collecting reproducible code, resulting in a comprehensive and valuable dataset of 51 adversarial defenses from 37 papers. This rigorous collection process enhances the benchmark's credibility and relevance to the research community.\n\n2. The benchmark setup involves working with \"messy\" research codebases, which closely reflects real-world challenges in adversarial defense scenarios. This realistic approach adds significant value, as it pushes current LLMs to handle complex, imperfect code environments typical in real applications.\n\n3. The paper employs definitive evaluation methods that do not rely on black-box metrics, such as LLM-as-a-judge approaches. Instead, it uses measurable, transparent performance metrics, ensuring that the results are both interpretable and replicable.", "weaknesses": "1. The paper lacks detailed examples of interactions with the LLMs, such as specific prompts used or a complete end-to-end example of at least one of the defenses and its corresponding developed attack stages. Including visual aids or detailed walk-throughs of successful or failed attacks could greatly enhance readability and engagement. The absence of such details makes the content dense leading to a less engaging reading experience.\n\n2. The current experimental settings are overly general for the LLMs, so it's unsurprising to observe such low success rates. From my experience, even for simpler coding tasks, I find that even the best LLMs require significantly more detail—such as an initial code draft, context explanations, and multiple rounds of conversation and reflection, as in a chat scenario—to achieve the desired results. For example, would a 6-turn interaction (interactive human-AI chat session) with these models (like GitHub Copilot settings) yield much better results? \n\n3. The paper could have explored more powerful methods, such as ICL few-shot examples, fine-tuning, or retrieval-augmented generation (RAG), to better evaluate LLMs' capabilities in solving the benchmark. These approaches could have provided valuable insights into the strengths and limitations of LLMs when leveraging different training and interaction methodologies. The current experimental settings are necessary but insufficient.\n\n4. This is a coding task, yet the paper's background lacks information on state-of-the-art LLMs specifically trained for coding. You've tested with GPT-4o and Claude 3.5 Sonnet, but are there other LLMs worth experimenting with? \n\n5. In Section 5.1 (End-to-End Evaluation), in the sentence \"Unsurprisingly, we find that current defenses fail completely at this task ...\" do you mean \"current models\"?", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["Yes, Privacy, security and safety", "Yes, Potentially harmful insights, methodologies and applications"], "details_of_ethics_concerns": "The authors have effectively adhered to the ethical concerns.", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730156806883}, {"id": "bDQkQGauPq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission14023/Reviewer_gDHc"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper presented AutoAdvExBench, which evaluated if LLM can automatically exploit defenses to adversarial examples. Contributions include:\n- evaluates if models can match the abilities of expert adversarial machine learning researchers\n- measure LLMs’ ability to understand and interact with sophisticated codebases\n- evaluates the ability of LLMs to reproduce prior research results automatically\n\nThe results show: the strongest agent, with a human-guided prompt, is only able to successfully generate adversarial examples on 6 of the 51 defenses in the benchmark.", "review_text": "The paper presented AutoAdvExBench, which evaluated if LLM can automatically exploit defenses to adversarial examples. Contributions include:\n- evaluates if models can match the abilities of expert adversarial machine learning researchers\n- measure LLMs’ ability to understand and interact with sophisticated codebases\n- evaluates the ability of LLMs to reproduce prior research results automatically\n\nThe results show: the strongest agent, with a human-guided prompt, is only able to successfully generate adversarial examples on 6 of the 51 defenses in the benchmark.", "strengths": "- It's an interesting benchmark to evaluate LLMs' capacity to autonomously exploit defenses to adversarial examples.\n- The defense implementations are collected comprehensively and rigorously.\n- The limitations are fairly presented.", "weaknesses": "The presentation can be improved:\n- It takes me a long time to find out that adversarial examples are for image classifiers. LLMs also have so-called adversarial attacks- it would be better to make it clear of the specific task in the abstract as well as early in the introduction to make it clear.\n- The introduction didn't actually specify what it is \"autonomously generate exploits on adversarial example defenses\", before jumping into what the benchmark includes and what is its impact.\n- The abstract focuses on the potential impact of the benchmark a lot, Yet it shall better inform what is the benchmark is really about-- the first sentence didn't convey it clearly.\n\nAs a benchmark paper, more LLMs should be considered and benchmarked. However, we only see the results for Claude 3.5 and GPT-4o.\n\nIf that is because other LLMs are so weak for the task, I personally think a good benchmark for current sota LLMs should be able to distinguish their capacity. If most of them fail for the benchmark, then it may not be a good choice at this time.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presented AutoAdvExBench, which evaluated if LLM can automatically exploit defenses to adversarial examples. Contributions include:\n- evaluates if models can match the abilities of expert adversarial machine learning researchers\n- measure LLMs’ ability to understand and interact with sophisticated codebases\n- evaluates the ability of LLMs to reproduce prior research results automatically\n\nThe results show: the strongest agent, with a human-guided prompt, is only able to successfully generate adversarial examples on 6 of the 51 defenses in the benchmark.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- It's an interesting benchmark to evaluate LLMs' capacity to autonomously exploit defenses to adversarial examples.\n- The defense implementations are collected comprehensively and rigorously.\n- The limitations are fairly presented.", "weaknesses": "The presentation can be improved:\n- It takes me a long time to find out that adversarial examples are for image classifiers. LLMs also have so-called adversarial attacks- it would be better to make it clear of the specific task in the abstract as well as early in the introduction to make it clear.\n- The introduction didn't actually specify what it is \"autonomously generate exploits on adversarial example defenses\", before jumping into what the benchmark includes and what is its impact.\n- The abstract focuses on the potential impact of the benchmark a lot, Yet it shall better inform what is the benchmark is really about-- the first sentence didn't convey it clearly.\n\nAs a benchmark paper, more LLMs should be considered and benchmarked. However, we only see the results for Claude 3.5 and GPT-4o.\n\nIf that is because other LLMs are so weak for the task, I personally think a good benchmark for current sota LLMs should be able to distinguish their capacity. If most of them fail for the benchmark, then it may not be a good choice at this time.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729997641073}], "openreview_url": "https://openreview.net/forum?id=leSbzBtofH", "arxiv_id": "2503.01811", "paper_pdf": "papers/leSbzBtofH.pdf", "paper_pdf_sha256": "3dbbb97e83361880472f876f10dd59a6ada4d4c4f1e0068eab4276bd9ec1dea3", "paper_pdf_bytes": 250543, "paper_pdf_source": "openreview", "code_url": "https://github.com/ethz-spylab/autoadvexbench", "code_repository": "ethz-spylab/autoadvexbench", "code_commit": "8da21de3aa3442b339b519a0f5c82b54905efbf1", "code_archive": "repos/leSbzBtofH.zip", "code_archive_sha256": "79ec0284af77407ac1cdfda75e7e82e5b0743baafc953763d99a853a743c6e06", "code_archive_bytes": 397012, "code_file_count": 32, "code_extensions": {".py": 31, ".js": 1}, "github_disk_usage_kb": 374, "github_languages": {"Python": 93972, "HTML": 21117}, "github_archived": false, "github_pushed_at": "2025-05-21T18:36:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/2503-01811"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XJiN1VkgA0", "year": 2024, "status": "rejected", "title": "Generating with Confidence: Uncertainty Quantification for Black-box Large Language Models", "authors": ["Zhen Lin", "Shubhendu Trivedi", "Jimeng Sun"], "authorids": ["~Zhen_Lin2", "~Shubhendu_Trivedi2", "~Jimeng_Sun3"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) specializing in natural language generation (NLG) have recently started exhibiting promising capabilities across a variety of domains. However, gauging the trustworthiness of responses generated by LLMs remains an open challenge, with limited research on uncertainty quantification (UQ) for NLG. Furthermore, existing literature typically assumes white-box access to language models, which is becoming unrealistic either due to the closed-source nature of the latest LLMs or computational constraints. In this work, we investigate UQ in NLG for *black-box* LLMs. We first differentiate *uncertainty* vs *confidence*: the former refers to the “dispersion” of the potential predictions for a fixed input, and the latter refers to the confidence on a particular prediction/generation. We then propose and compare several confidence/uncertainty metrics, applying them to *selective* NLG where unreliable results could either be ignored or yielded for further assessment. Experiments were carried out with several popular LLMs on question-answering datasets (for evaluation purposes). Results reveal that a simple metric for the semantic dispersion can be a reliable predictor of the quality of LLM responses, providing valuable insights for practitioners on uncertainty management when adopting LLMs.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "dJgUfjK6Y4", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8527/Reviewer_1J6D"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposed several metrics to estimate the model output uncertainty for a given question without access to the logic, i.e., close models (openAI models). The proposed metric uses three types of similarity methods to estimate multiple model generations for a question, then uses the pair-wise similarity score to estimate the model’s uncertainty on a given question with different metrics like the eigenvalue of the graph laplacian. They later evaluate the metrics compared with baselines and white-box uncertainty measures. The results showed improvement.", "review_text": "The paper proposed several metrics to estimate the model output uncertainty for a given question without access to the logic, i.e., close models (openAI models). The proposed metric uses three types of similarity methods to estimate multiple model generations for a question, then uses the pair-wise similarity score to estimate the model’s uncertainty on a given question with different metrics like the eigenvalue of the graph laplacian. They later evaluate the metrics compared with baselines and white-box uncertainty measures. The results showed improvement.", "strengths": "1. The paper proposed an interesting idea of estimating black-box generation output without access to the output logits.\n2. The motivation is strong, and the paper is timely as the wide usage of close source models estimating model uncertainty without logits could help downstream tasks avoid using uncertain model output.\n3. The proposed method is intuitive and reasonable to be applied to this situation.", "weaknesses": "1. Some more analysis should be conducted, e.g., the impact of similarity functions and intuitions of why one similarity function is more preferred than the other.\n2. Some quantitative examples would be great, such as what kind of questions are evaluated as uncertain as the proposed metric vs the baselines, etc.\n3. The presentation and writing of the paper can be more intuitive. Tables 1, 2, and 3 are very hard to read; an alternative is to include the important numbers and put the rest into the appendix.", "questions": "1. Is there any reason for the definition of the degree matrix?\n2. Have you tried other similarity measures, such as cosine similarity, using existing word embeddings?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed several metrics to estimate the model output uncertainty for a given question without access to the logic, i.e., close models (openAI models). The proposed metric uses three types of similarity methods to estimate multiple model generations for a question, then uses the pair-wise similarity score to estimate the model’s uncertainty on a given question with different metrics like the eigenvalue of the graph laplacian. They later evaluate the metrics compared with baselines and white-box uncertainty measures. The results showed improvement.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The paper proposed an interesting idea of estimating black-box generation output without access to the output logits.\n2. The motivation is strong, and the paper is timely as the wide usage of close source models estimating model uncertainty without logits could help downstream tasks avoid using uncertain model output.\n3. The proposed method is intuitive and reasonable to be applied to this situation.", "weaknesses": "1. Some more analysis should be conducted, e.g., the impact of similarity functions and intuitions of why one similarity function is more preferred than the other.\n2. Some quantitative examples would be great, such as what kind of questions are evaluated as uncertain as the proposed metric vs the baselines, etc.\n3. The presentation and writing of the paper can be more intuitive. Tables 1, 2, and 3 are very hard to read; an alternative is to include the important numbers and put the rest into the appendix.", "questions": "1. Is there any reason for the definition of the degree matrix?\n2. Have you tried other similarity measures, such as cosine similarity, using existing word embeddings?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699035350846}, {"id": "Aexoj64XZI", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8527/Reviewer_LJF2"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes a new method for uncertainty quantification for NLG when the model used is a black-box LLMs (no access to logits). The framework used consists of three steps (1) response samples generation, (2) pairwise similarity calculation between responses, and (3) uncertainty and confidence score estimation. Using this framework, the authors compare different estimation methods using number of semantic sets, sum of eigenvalues, degree matrix, and eccentricity. For the pairwise similarity, they compare the standard Jaccard similarity as well as model-based method using predicted probabilities from a natural language inference (NLI) model. Experiment results show that the proposed uncertainty and confidence score estimation outperform the baselines, and sometimes better than white-box methods.", "review_text": "The paper proposes a new method for uncertainty quantification for NLG when the model used is a black-box LLMs (no access to logits). The framework used consists of three steps (1) response samples generation, (2) pairwise similarity calculation between responses, and (3) uncertainty and confidence score estimation. Using this framework, the authors compare different estimation methods using number of semantic sets, sum of eigenvalues, degree matrix, and eccentricity. For the pairwise similarity, they compare the standard Jaccard similarity as well as model-based method using predicted probabilities from a natural language inference (NLI) model. Experiment results show that the proposed uncertainty and confidence score estimation outperform the baselines, and sometimes better than white-box methods.", "strengths": "- A novel approach to estimate uncertainty and model confidence.\n- Thorough experiments on multiple QA datasets and public LLMs.", "weaknesses": "- The paper addresses an interesting question that relevant to most of the current LLM works. However, I feel the work is not fully complete, with many results but lack of analysis and insights on what to be the main findings of the paper. There are also some missing details, e.g. what DeBERTa model used for NLI? Was it trained using NLI data?\n- There is a discussion regarding uncertainty vs. model confidence and why it matters, but along the way the differentiation is not clear - Sec 3.2 does not even differentiate the two in details, where I expect that the discussion should be there because of the subsection title.\n- There are many variations/ablations regarding the methods used, i.e. similarity metrics and also uncertainty/confidence measurement but little analysis on which perform best, why, and what is the recommendation for future work if they want to quantify uncertainty / model confidence.\n- I am also a bit concerned regarding evaluation with GPT model. Although the authors perform human verification on a subset, but the subset is pretty small (33 samples per dataset). I think many of these datasets are factual, thus human evaluation might not be as tricky as say creative writing task, so having more human verification on the results would make the claim stronger. There isn't also explanation about inter-annotator agreement on the verification. Though, I appreciate that the authors acknowledge the limitation of using GPT models for evaluation.", "questions": "- What is the reasoning for a_{NLI, contra}? If you use a standard NLI model which has three classes, Eq 4 (right) would take into account the entailment and neutral class (since it is 1-p_{contra}, is that correct?\n- Unclear why uncertainty is used for expected accuracy and model confidence for individual accuracy.\n\nSuggestions for paper:\n- Provide examples on when the uncertainty and confidence estimation aligns/doesn't align with the prediction\n- More depth analysis on why a particular method (similarity metric/quantification method) outperform the other methods used in the experiment", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new method for uncertainty quantification for NLG when the model used is a black-box LLMs (no access to logits). The framework used consists of three steps (1) response samples generation, (2) pairwise similarity calculation between responses, and (3) uncertainty and confidence score estimation. Using this framework, the authors compare different estimation methods using number of semantic sets, sum of eigenvalues, degree matrix, and eccentricity. For the pairwise similarity, they compare the standard Jaccard similarity as well as model-based method using predicted probabilities from a natural language inference (NLI) model. Experiment results show that the proposed uncertainty and confidence score estimation outperform the baselines, and sometimes better than white-box methods.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- A novel approach to estimate uncertainty and model confidence.\n- Thorough experiments on multiple QA datasets and public LLMs.", "weaknesses": "- The paper addresses an interesting question that relevant to most of the current LLM works. However, I feel the work is not fully complete, with many results but lack of analysis and insights on what to be the main findings of the paper. There are also some missing details, e.g. what DeBERTa model used for NLI? Was it trained using NLI data?\n- There is a discussion regarding uncertainty vs. model confidence and why it matters, but along the way the differentiation is not clear - Sec 3.2 does not even differentiate the two in details, where I expect that the discussion should be there because of the subsection title.\n- There are many variations/ablations regarding the methods used, i.e. similarity metrics and also uncertainty/confidence measurement but little analysis on which perform best, why, and what is the recommendation for future work if they want to quantify uncertainty / model confidence.\n- I am also a bit concerned regarding evaluation with GPT model. Although the authors perform human verification on a subset, but the subset is pretty small (33 samples per dataset). I think many of these datasets are factual, thus human evaluation might not be as tricky as say creative writing task, so having more human verification on the results would make the claim stronger. There isn't also explanation about inter-annotator agreement on the verification. Though, I appreciate that the authors acknowledge the limitation of using GPT models for evaluation.", "questions": "- What is the reasoning for a_{NLI, contra}? If you use a standard NLI model which has three classes, Eq 4 (right) would take into account the entailment and neutral class (since it is 1-p_{contra}, is that correct?\n- Unclear why uncertainty is used for expected accuracy and model confidence for individual accuracy.\n\nSuggestions for paper:\n- Provide examples on when the uncertainty and confidence estimation aligns/doesn't align with the prediction\n- More depth analysis on why a particular method (similarity metric/quantification method) outperform the other methods used in the experiment", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698832996249}, {"id": "23cEu7rUCR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8527/Reviewer_YpC9"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents an innovative approach for generating confidence and uncertainty scores within the context of natural language generation (NLG), while working under the realistic constraint of no white-box access to the underlying model. In this scenario, where only the generated sequences are observable and the language model (LLM) can be queried repeatedly for confidence and uncertainty assessments, the proposed method leverages pair-wise similarity or entailment probabilities between generated samples to construct a weight graph. Spectral clustering is employed on that graph to compute confidence and uncertainty scores. Intriguingly, the paper's results suggest that this method can outperform traditional white-box techniques in certain instances, highlighting its potential for enhancing NLG performance.", "review_text": "This paper presents an innovative approach for generating confidence and uncertainty scores within the context of natural language generation (NLG), while working under the realistic constraint of no white-box access to the underlying model. In this scenario, where only the generated sequences are observable and the language model (LLM) can be queried repeatedly for confidence and uncertainty assessments, the proposed method leverages pair-wise similarity or entailment probabilities between generated samples to construct a weight graph. Spectral clustering is employed on that graph to compute confidence and uncertainty scores. Intriguingly, the paper's results suggest that this method can outperform traditional white-box techniques in certain instances, highlighting its potential for enhancing NLG performance.", "strengths": "- Explores an important problem: with the advent of LLMs it is necessary to address how to compute confidence and uncertainty estimates which was easier to obtain in traditional deep learning setting. \n- Paper is well written and easy to follow \n- The paper proposes an interesting idea to use spectral clustering to produce uncertainty and confidence measures.\n- Evaluates performance on different QA datasets and different LLMs as well.", "weaknesses": "- It might be helpful to measure expected or adaptive calibration error to show this method outputs confidence scores which are better in comparison to other methods. \n- Relevant work: https://arxiv.org/pdf/2306.13063.pdf, might be helpful to include comparison between their method and your proposed method.", "questions": "- As pointed out in the weakness, for confidence scores measuring ECE or ACE might be helpful in showcasing the effectiveness of the approach. \n- In cases where we have access to model embeddings, I think same method of clustering could be applied where we use similarity between embeddings instead of entailment prob/Jaccard similarity. Would this method give better confidence or uncertainty estimates in those settings as well?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents an innovative approach for generating confidence and uncertainty scores within the context of natural language generation (NLG), while working under the realistic constraint of no white-box access to the underlying model. In this scenario, where only the generated sequences are observable and the language model (LLM) can be queried repeatedly for confidence and uncertainty assessments, the proposed method leverages pair-wise similarity or entailment probabilities between generated samples to construct a weight graph. Spectral clustering is employed on that graph to compute confidence and uncertainty scores. Intriguingly, the paper's results suggest that this method can outperform traditional white-box techniques in certain instances, highlighting its potential for enhancing NLG performance.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Explores an important problem: with the advent of LLMs it is necessary to address how to compute confidence and uncertainty estimates which was easier to obtain in traditional deep learning setting. \n- Paper is well written and easy to follow \n- The paper proposes an interesting idea to use spectral clustering to produce uncertainty and confidence measures.\n- Evaluates performance on different QA datasets and different LLMs as well.", "weaknesses": "- It might be helpful to measure expected or adaptive calibration error to show this method outputs confidence scores which are better in comparison to other methods. \n- Relevant work: https://arxiv.org/pdf/2306.13063.pdf, might be helpful to include comparison between their method and your proposed method.", "questions": "- As pointed out in the weakness, for confidence scores measuring ECE or ACE might be helpful in showcasing the effectiveness of the approach. \n- In cases where we have access to model embeddings, I think same method of clustering could be applied where we use similarity between embeddings instead of entailment prob/Jaccard similarity. Would this method give better confidence or uncertainty estimates in those settings as well?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698802146556}, {"id": "O66TS4u7OJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8527/Reviewer_sn9f"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper explores how to detect whether text generated from an LLM is accurate or not, without having access to the models probabilities (so called \"black box\" inference). It does so based on decoding multiple outputs (token sequences) for a single input via sampling and then exploring various measures for how much these multiple outputs agree with each other. \nIt reports experimental results across a range of different models on the task of question answering (QA). QA is chosen to make it easier to evaluate the accuracy of model outputs. \nRelevant literature from other ML fields which have made greater progress than LLMs/NLP in quantifying model uncertainty are referenced.", "review_text": "The paper explores how to detect whether text generated from an LLM is accurate or not, without having access to the models probabilities (so called \"black box\" inference). It does so based on decoding multiple outputs (token sequences) for a single input via sampling and then exploring various measures for how much these multiple outputs agree with each other. \nIt reports experimental results across a range of different models on the task of question answering (QA). QA is chosen to make it easier to evaluate the accuracy of model outputs. \nRelevant literature from other ML fields which have made greater progress than LLMs/NLP in quantifying model uncertainty are referenced.", "strengths": "On the whole the paper is well written. The topic of uncertainty quantification of LLMs is well motivated. The paper builds of some very recent work in this field, and reports good empirical results across several models which indicate the proposed methods for measuring black-box uncertainty are promising. \nRepeating the experiments and reporting mean + stdDev is helpful for gaining confidence in the reported numbers. \nIt's interesting to see another example of how poorly white-box probabilities reflect accuracy, per tables 1,2,3.", "weaknesses": "The paper mentions several times that uncertainty is not the same as confidence. Confidence scores are defined based on the token probabilities, which are not available in black-box inference settings. The paper however isn't very clear in how uncertainty and confidence are not the same thing, with multiple sentences referring to \"uncertainty/confidence\" as though they are the same, and section 4.2. not written in a way that helps clarify the working definitions of how these differ either. This is frustrating for the reader. \n\nThe main limitation is only using question answering to test the models here. There are a spectrum of NLG tasks, and I agree with the authors that caring about uncertainty is less likely for completely open generation. However there are tasks like data-to-text (ie knowledge grounded) NLG, where the outputs need to be trusted, and it remains an open issue whether these methods reported here are transferable to tasks other than QA. The QA task is also interesting in that it is extracting answers from the LLMs weights, rather than using the LLM as a general tool for e.g. the data-to-text task, where the info is given as structural inputs.", "questions": "* In the introduction it's presented that LLMs are gaining attention -> it's crucial to quantify their uncertainty.  After reading the paper, I still have to ask why? It seems _crucial_ to improve their accuracy. Is the only rational for improving their uncertainty so that their accuracy can be improved, e.g. by blocking uncertain outputs? Or is there other reasons as well?\n* How is the issue of model calibration different?\n* Do you think these results transfer to other tasks such as data-to-text? Or do you think there may be different results beyond QA tasks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper explores how to detect whether text generated from an LLM is accurate or not, without having access to the models probabilities (so called \"black box\" inference). It does so based on decoding multiple outputs (token sequences) for a single input via sampling and then exploring various measures for how much these multiple outputs agree with each other. \nIt reports experimental results across a range of different models on the task of question answering (QA). QA is chosen to make it easier to evaluate the accuracy of model outputs. \nRelevant literature from other ML fields which have made greater progress than LLMs/NLP in quantifying model uncertainty are referenced.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "On the whole the paper is well written. The topic of uncertainty quantification of LLMs is well motivated. The paper builds of some very recent work in this field, and reports good empirical results across several models which indicate the proposed methods for measuring black-box uncertainty are promising. \nRepeating the experiments and reporting mean + stdDev is helpful for gaining confidence in the reported numbers. \nIt's interesting to see another example of how poorly white-box probabilities reflect accuracy, per tables 1,2,3.", "weaknesses": "The paper mentions several times that uncertainty is not the same as confidence. Confidence scores are defined based on the token probabilities, which are not available in black-box inference settings. The paper however isn't very clear in how uncertainty and confidence are not the same thing, with multiple sentences referring to \"uncertainty/confidence\" as though they are the same, and section 4.2. not written in a way that helps clarify the working definitions of how these differ either. This is frustrating for the reader. \n\nThe main limitation is only using question answering to test the models here. There are a spectrum of NLG tasks, and I agree with the authors that caring about uncertainty is less likely for completely open generation. However there are tasks like data-to-text (ie knowledge grounded) NLG, where the outputs need to be trusted, and it remains an open issue whether these methods reported here are transferable to tasks other than QA. The QA task is also interesting in that it is extracting answers from the LLMs weights, rather than using the LLM as a general tool for e.g. the data-to-text task, where the info is given as structural inputs.", "questions": "* In the introduction it's presented that LLMs are gaining attention -> it's crucial to quantify their uncertainty.  After reading the paper, I still have to ask why? It seems _crucial_ to improve their accuracy. Is the only rational for improving their uncertainty so that their accuracy can be improved, e.g. by blocking uncertain outputs? Or is there other reasons as well?\n* How is the issue of model calibration different?\n* Do you think these results transfer to other tasks such as data-to-text? Or do you think there may be different results beyond QA tasks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698177659827}], "openreview_url": "https://openreview.net/forum?id=XJiN1VkgA0", "arxiv_id": "2305.19187", "paper_pdf": "papers/XJiN1VkgA0.pdf", "paper_pdf_sha256": "6fe887b68b1544c5ac4a5f576b66d4939c7ec6e27360566492527c8d9a52fcb0", "paper_pdf_bytes": 711403, "paper_pdf_source": "openreview", "code_url": "https://github.com/zlin7/UQ-NLG", "code_repository": "zlin7/UQ-NLG", "code_commit": "ecaea91741a6c1076e9069b56d32a9929eea8ff9", "code_archive": "repos/XJiN1VkgA0.zip", "code_archive_sha256": "b8914f75d45f8be4a7828b14affe9f54ec8a93cf61ec67c5f864569714605a56", "code_archive_bytes": 346125, "code_file_count": 18, "code_extensions": {".py": 16, ".ipynb": 2}, "github_disk_usage_kb": 345, "github_languages": {"Jupyter Notebook": 262628, "Python": 97829}, "github_archived": false, "github_pushed_at": "2024-06-30T21:12:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generating-with-confidence-uncertainty"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pcBJT4bgbpH", "year": 2023, "status": "rejected", "title": "Attention Flows for General Transformers", "authors": ["Niklas Metzger", "Christopher Hahn", "Julian Siber", "Frederik Schmitt", "Bernd Finkbeiner"], "authorids": ["~Niklas_Metzger1", "~Christopher_Hahn1", "~Julian_Siber1", "~Frederik_Schmitt1", "~Bernd_Finkbeiner1"], "authors_source": "OpenReview API", "abstract": "In this paper, we study the computation of how much an input token in a Transformer model influences its prediction. We formalize a method to construct a flow network out of the attention values of encoder-only Transformer models and extend it to general Transformer architectures, including an auto-regressive decoder. We show that running a maxflow algorithm on the flow network construction yields Shapley values, which determine a player's impact in cooperative game theory. By interpreting the input tokens in the flow network as players, we can compute their influence on the total attention flow leading to the decoder's decision. Additionally, we provide a library that computes and visualizes the attention flow of arbitrary Transformer models. We show the usefulness of our implementation on various models trained on natural language processing and reasoning tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "bH59bXG0lA", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6356/Reviewer_wsJr"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work extends the attention flow method proposed in Abnar and Zuidema 2020 to encoder-decoder and decoder-only transformers. The major contribution is based on the observation that later predicted words have more incoming edges than earlier words, such that to ensure positional independence, this work proposes a method to normalize maxflow values. In addition, this work draws connection between maxflow attention and Shapley values by defining payoffs as the sum of maxflows and showing the equivalence under this definition. Experiments on several tasks show that this method is able to gain insights into token importance for a prediction task.", "review_text": "My major concerns are: 1. the change to the original attention flow paper is incremental; and 2. a correlation study with feature importance found by other methods (such as Shapley values) is missing and it's not clear if the found maxflow values mean anything. Therefore, I'm leaning towards rejecting this paper.", "strengths": "Strengths:\n1. Extends attention flow to encoder-decoder and decoder-only transformers.\n2. Empirical analysis of token importance on several tasks.\n\nWeaknesses:\n1. The change to Abnar and Zuidema 2020 seems incremental.\n2. The connection to Shapley values is drawn by defining value function to be based on maxflows, so the equivalence is not surprising at all. A more interesting connection would be to use the actual log probability of the target token as payoff values and see if there's a correlation.\n3. Related to 2, it is not clear if attention maxflows reflect actual token importance. It would be more convincing if the maxflow values can be compared against feature importance obtained using other methods (such as Shapley values using log probability of target token as payoffs, or simply gradient-based saliency maps). For example, does the first-token bias found by maxflow hold for Shapley values?\n4. The analysis on individual attention heads doesn't make that much sense to me, since the same head id across different layers does not mean they have anything in common.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This work extends the attention flow method proposed in Abnar and Zuidema 2020 to encoder-decoder and decoder-only transformers. The major contribution is based on the observation that later predicted words have more incoming edges than earlier words, such that to ensure positional independence, this work proposes a method to normalize maxflow values. In addition, this work draws connection between maxflow attention and Shapley values by defining payoffs as the sum of maxflows and showing the equivalence under this definition. Experiments on several tasks show that this method is able to gain insights into token importance for a prediction task.", "strength_and_weaknesses": "Strengths:\n1. Extends attention flow to encoder-decoder and decoder-only transformers.\n2. Empirical analysis of token importance on several tasks.\n\nWeaknesses:\n1. The change to Abnar and Zuidema 2020 seems incremental.\n2. The connection to Shapley values is drawn by defining value function to be based on maxflows, so the equivalence is not surprising at all. A more interesting connection would be to use the actual log probability of the target token as payoff values and see if there's a correlation.\n3. Related to 2, it is not clear if attention maxflows reflect actual token importance. It would be more convincing if the maxflow values can be compared against feature importance obtained using other methods (such as Shapley values using log probability of target token as payoffs, or simply gradient-based saliency maps). For example, does the first-token bias found by maxflow hold for Shapley values?\n4. The analysis on individual attention heads doesn't make that much sense to me, since the same head id across different layers does not mean they have anything in common.", "clarity,_quality,_novelty_and_reproducibility": "This paper is original and well-written. However, I think the connection to Shapley values in its current form is over-complicating things and it's not necessary for the understanding of the paper.", "summary_of_the_review": "My major concerns are: 1. the change to the original attention flow paper is incremental; and 2. a correlation study with feature importance found by other methods (such as Shapley values) is missing and it's not clear if the found maxflow values mean anything. Therefore, I'm leaning towards rejecting this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666890641104}, {"id": "xnfKeF5kIg9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6356/Reviewer_4MwZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper constructs an attention flow network out of encoder (& decoder) Transformers, and release a library to visualize the attention flow of Transformers.", "review_text": "Overall, the paper feels like an extension to the original Attention Flows paper and not like a standalone paper. The experiments are limited, and the knowledge uncovered by the visualizations and methods are not impactful.\n\nI did not read the Appendices thoroughly.", "strengths": "**Strengths**\n\n- (+) Authors explain that Attention Flow can be connected between an encoder and decoder of a Transformer.\n\n**Weaknesses**\n\n- (- -) The paper’s analysis is very limited. <5 sentences are analyzed in the main paper to show that their method works. The insights they uncover from the shapley values about the role that individual heads play in transformer attention are not novel.\n- (- -) The visualization library seems limited and not thoroughly introduced even though it was a central part of the abstract. The paper’s figures simply show tokens next to shapley values as colored blocks. The background color of the tokens overemphasizes small differences in the shapley value when all values are similar and is more deceiving than helpful. Additionally, the fact that attention flow primarily attends to punctuation and special tokens is not meaningful to understanding the data domain.\n- (-) Across the board, I am unable to understand why the experimental results are particularly meaningful. The experimental section feels more like showing that you can get numbers at all using this method.\n\n **Other Comments**\n\n- Statement “In practice, however, heads are biased towards keeping their respective tasks” [across layers] is not defended, and it has been my experience that in Transformers this is entirely not the case. In Transformers that have no weight sharing, there is no mechanism that encourages a head to perform the same association function across layers.\n- Figs 4 and 5 are not self contained. I cannot see the input tokens in the plot itself, making the results difficult to interpret.\n- Figure 2’s explanation on the top of page 4, “encoder (top) and a decoder (bottom)” is inconsistent with the design of the figure itself.\n- Unclear what “positional independence” (sec 1 page 2) of computed maxflow values means — positional encoding is embedded into the token representation itself and attention learns to attend to each token+position? I could not find where this was further explained in the paper", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper constructs an attention flow network out of encoder (& decoder) Transformers, and release a library to visualize the attention flow of Transformers.", "strength_and_weaknesses": "**Strengths**\n\n- (+) Authors explain that Attention Flow can be connected between an encoder and decoder of a Transformer.\n\n**Weaknesses**\n\n- (- -) The paper’s analysis is very limited. <5 sentences are analyzed in the main paper to show that their method works. The insights they uncover from the shapley values about the role that individual heads play in transformer attention are not novel.\n- (- -) The visualization library seems limited and not thoroughly introduced even though it was a central part of the abstract. The paper’s figures simply show tokens next to shapley values as colored blocks. The background color of the tokens overemphasizes small differences in the shapley value when all values are similar and is more deceiving than helpful. Additionally, the fact that attention flow primarily attends to punctuation and special tokens is not meaningful to understanding the data domain.\n- (-) Across the board, I am unable to understand why the experimental results are particularly meaningful. The experimental section feels more like showing that you can get numbers at all using this method.\n\n **Other Comments**\n\n- Statement “In practice, however, heads are biased towards keeping their respective tasks” [across layers] is not defended, and it has been my experience that in Transformers this is entirely not the case. In Transformers that have no weight sharing, there is no mechanism that encourages a head to perform the same association function across layers.\n- Figs 4 and 5 are not self contained. I cannot see the input tokens in the plot itself, making the results difficult to interpret.\n- Figure 2’s explanation on the top of page 4, “encoder (top) and a decoder (bottom)” is inconsistent with the design of the figure itself.\n- Unclear what “positional independence” (sec 1 page 2) of computed maxflow values means — positional encoding is embedded into the token representation itself and attention learns to attend to each token+position? I could not find where this was further explained in the paper", "clarity,_quality,_novelty_and_reproducibility": "I finished the paper unsure of the impact of its contribution. The paper is not clear on the motivation for their method, and I gather no new insights from their analysis on the attention of Transformers.", "summary_of_the_review": "Overall, the paper feels like an extension to the original Attention Flows paper and not like a standalone paper. The experiments are limited, and the knowledge uncovered by the visualizations and methods are not impactful.\n\nI did not read the Appendices thoroughly.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666611840295}, {"id": "gn61_AxpSR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6356/Reviewer_DmXz"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the author provides a comprehensive formulation of the attention flows for different architectures of transformers (encoder, decoder, encoder-decoder). In order to account for the auto-regressive decoder structure, the author adjusted the attention flow to ensure the positional independence of the computed maxflow values. Following the previous work, the author further shows that the maximum attention flow corresponds to the Shapley value under the three architectures of transformers. ", "review_text": "Explaining the decision from large language models has always been a focus since the popularity of these models. This paper improves on top of similar existing works but providing a comprehensive formulation on the attention flow that covers decoder and encode-decoder too. It is of great interest to this community to see progress that leads to better explainability of the model. On the other hand, this paper's main contribution is not significant enough for ICLR since the major works (both the attention flow and the equivalence to Shapley value) exist before and it did not contribute significantly outside the existing framework. The performance evaluation is also tricky since only a few examples are shown but lacks of more experiments to demonstrate the strength of attention flow method. I would like to see some improvement on the matter. ", "strengths": "The strength of this paper lies in the following aspects:\n\n1. Although the attention flow formulation for transformer encoder has been proposed before, this paper extends this to cover the rest of two architectures so that this method can be applied to auto-regressive structure. The model formulation is solid with minor typos. \n\n2. The connections between attention flow values and the Shapley values has been shown before. This work, however, is able to explore the assumption on the positional independence so that the proof is simplified and is able to avoid some issues behind the underlying assumptions of the previous proof. \n\nThe weakness of this paper lies in the following aspects:\n\n1. Lack of objective and reliable metric to decide if the attention flow is relevant or not. It is stated in the paper also that the attention flow value is not supposed to be interpreted as causal factors to the prediction or translation. It is merely an associational factor. This together with lack of large scale studies and peer-reviews in terms of the performance of the method limits its application in real life decision making. \n\n2. The proof that the attention flow is Shapley value relies on the effectiveness of the positional independence assumption. A solid verification and analysis is necessary to convince the reader that the positional independence is guaranteed in experiments. It is very challenging in general to check on this assumption even if the weight is normalized for tokens output later.\n\n3. The proof that the attention flow is Shapley value under the positional independence assumption needs to be shown in detail instead of explained vaguely. The contribution of this paper is its mathematical formulation of the attention flow. It is necessary to show this critical proposition. Hope that this part can be added in appendix later on.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the author provides a comprehensive formulation of the attention flows for different architectures of transformers (encoder, decoder, encoder-decoder). In order to account for the auto-regressive decoder structure, the author adjusted the attention flow to ensure the positional independence of the computed maxflow values. Following the previous work, the author further shows that the maximum attention flow corresponds to the Shapley value under the three architectures of transformers. ", "strength_and_weaknesses": "The strength of this paper lies in the following aspects:\n\n1. Although the attention flow formulation for transformer encoder has been proposed before, this paper extends this to cover the rest of two architectures so that this method can be applied to auto-regressive structure. The model formulation is solid with minor typos. \n\n2. The connections between attention flow values and the Shapley values has been shown before. This work, however, is able to explore the assumption on the positional independence so that the proof is simplified and is able to avoid some issues behind the underlying assumptions of the previous proof. \n\nThe weakness of this paper lies in the following aspects:\n\n1. Lack of objective and reliable metric to decide if the attention flow is relevant or not. It is stated in the paper also that the attention flow value is not supposed to be interpreted as causal factors to the prediction or translation. It is merely an associational factor. This together with lack of large scale studies and peer-reviews in terms of the performance of the method limits its application in real life decision making. \n\n2. The proof that the attention flow is Shapley value relies on the effectiveness of the positional independence assumption. A solid verification and analysis is necessary to convince the reader that the positional independence is guaranteed in experiments. It is very challenging in general to check on this assumption even if the weight is normalized for tokens output later.\n\n3. The proof that the attention flow is Shapley value under the positional independence assumption needs to be shown in detail instead of explained vaguely. The contribution of this paper is its mathematical formulation of the attention flow. It is necessary to show this critical proposition. Hope that this part can be added in appendix later on.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written, with clear motivation and overall structure. The code provided is also workable. ", "summary_of_the_review": "Explaining the decision from large language models has always been a focus since the popularity of these models. This paper improves on top of similar existing works but providing a comprehensive formulation on the attention flow that covers decoder and encode-decoder too. It is of great interest to this community to see progress that leads to better explainability of the model. On the other hand, this paper's main contribution is not significant enough for ICLR since the major works (both the attention flow and the equivalence to Shapley value) exist before and it did not contribute significantly outside the existing framework. The performance evaluation is also tricky since only a few examples are shown but lacks of more experiments to demonstrate the strength of attention flow method. I would like to see some improvement on the matter. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666476909056}, {"id": "l3FClHE8xgf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6356/Reviewer_uJJx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a method to quantify the impact of each input token on the outcome of a Transformer model. The algorithm constructs a flow network whose edges contain attention values, and solves the max flow problem for each vertex corresponding to an input token. The idea is applied to encoder only, decoder only, and encoder-decoder models.\n\nThe paper makes some connections with game theoretical concepts to justify the meaningfulness of the approach and the computed scores.\n\nFinally, the paper shows some experiments where the token-importance values are computed, and offers an explanation regarding the underlying logic or intuition behind each example.", "review_text": "This paper extends [1, 2] to decoder only and encoder-decoder architectures, and provides some experiments where the method is applied. The contribution of the paper seems modest, and maybe some of the mathematical formalization could be replaced with more examples or applications.\n\n[1] = Abnar and Zuidema, Quantifying Attention Flow in Transformers.\n[2] = Ethayarajh and Jurafsky, Attention Flows are Shapley Value Explanations.", "strengths": "The overall approach makes intuitive sense (note it isn't new), but the outputs are numbers that are still hard to interpret or use, as there are many moving parts. After going over the paper, I ended up with the impression of having read about a technical tool that isn't solving any specific problem. Given the amount of follow-up work on [1], it seems the ML community probably has found applications for it, though.\n\nI appreciate the authors' efforts and honesty at the end of the paper to list limitations of the work. One example is the fact that Transformer models also apply layers other than attention, like MLPs. I'm not sure if the proposed method accounts for token transformations via those; maybe the fact that they affect the subsequent attention scores is enough? i.e. if an MLP layer zeroes-out a token, the attention scores used to solve the max-flow problem (matrix A in the paper) will propagate the zeros, right? Otherwise results would be misleading.\n\nIn terms of experiments, as a proposal, I think a simple vision example should be much more informative. Each input token in VIT corresponds to a patch, and for image classification, the authors could show the input image superimposed with the corresponding attention score of each patch. Ideally, the attention centers around the object that is to be classified. It would also be interesting to see how solving a max-flow problem with hundreds of input tokens (like in VIT) scales in terms of runtime.\n\nThe authors offer some time estimates in 4.1 (text completion), but given the complexity and engineering required to compute attention flows, I was expecting a more in-depth analysis of how expensive it is to compute these numbers for real-world models. Grouping inputs or skipping layers, as mentioned in Section 2.5 is an interesting approach, but no evidence is provided on that direction suggesting it could work.\n\nIt's hard to conclude much from some of the experiments (like Figure 4 or 5a).\n\n[1] = Abnar and Zuidema, Quantifying Attention Flow in Transformers.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a method to quantify the impact of each input token on the outcome of a Transformer model. The algorithm constructs a flow network whose edges contain attention values, and solves the max flow problem for each vertex corresponding to an input token. The idea is applied to encoder only, decoder only, and encoder-decoder models.\n\nThe paper makes some connections with game theoretical concepts to justify the meaningfulness of the approach and the computed scores.\n\nFinally, the paper shows some experiments where the token-importance values are computed, and offers an explanation regarding the underlying logic or intuition behind each example.", "strength_and_weaknesses": "The overall approach makes intuitive sense (note it isn't new), but the outputs are numbers that are still hard to interpret or use, as there are many moving parts. After going over the paper, I ended up with the impression of having read about a technical tool that isn't solving any specific problem. Given the amount of follow-up work on [1], it seems the ML community probably has found applications for it, though.\n\nI appreciate the authors' efforts and honesty at the end of the paper to list limitations of the work. One example is the fact that Transformer models also apply layers other than attention, like MLPs. I'm not sure if the proposed method accounts for token transformations via those; maybe the fact that they affect the subsequent attention scores is enough? i.e. if an MLP layer zeroes-out a token, the attention scores used to solve the max-flow problem (matrix A in the paper) will propagate the zeros, right? Otherwise results would be misleading.\n\nIn terms of experiments, as a proposal, I think a simple vision example should be much more informative. Each input token in VIT corresponds to a patch, and for image classification, the authors could show the input image superimposed with the corresponding attention score of each patch. Ideally, the attention centers around the object that is to be classified. It would also be interesting to see how solving a max-flow problem with hundreds of input tokens (like in VIT) scales in terms of runtime.\n\nThe authors offer some time estimates in 4.1 (text completion), but given the complexity and engineering required to compute attention flows, I was expecting a more in-depth analysis of how expensive it is to compute these numbers for real-world models. Grouping inputs or skipping layers, as mentioned in Section 2.5 is an interesting approach, but no evidence is provided on that direction suggesting it could work.\n\nIt's hard to conclude much from some of the experiments (like Figure 4 or 5a).\n\n[1] = Abnar and Zuidema, Quantifying Attention Flow in Transformers.", "clarity,_quality,_novelty_and_reproducibility": "The paper makes an effort to formalize the problem and make connections with game theoretical concepts. I'm not sure these connections are worth the consequent over-technification of the paper, even though it is claimed to be one of the central contributions. Maybe more examples or experiments showing how these scores behave in practice or could be feed as inputs to something else could be more useful.\n\nIn general, the paper reads well and smoothly.\n\nThe authors will open source the code eventually, and thus results should be reproducible.\n\nIn terms of novelty, the paper ends the introduction with a related-work paragraph. This work mainly builds on top of two previous works [1, 2]. [1] introduces attention flows (mainly for Transformer encoders), and [2] shows attention flows are Shapley values for a specific formulation of a cooperative game. My understanding is that the contribution of this paper is to extend the ideas in [1, 2] to decoder-only and encoder-decoder models, for example by handling temporal dependence in autoregressive models (e.g. positional independence).\n\nProviding the code will be useful for the ML community, to explore and analyze trained models.\n\n[1] = Abnar and Zuidema, Quantifying Attention Flow in Transformers.\n[2] = Ethayarajh and Jurafsky, Attention Flows are Shapley Value Explanations.", "summary_of_the_review": "This paper extends [1, 2] to decoder only and encoder-decoder architectures, and provides some experiments where the method is applied. The contribution of the paper seems modest, and maybe some of the mathematical formalization could be replaced with more examples or applications.\n\n[1] = Abnar and Zuidema, Quantifying Attention Flow in Transformers.\n[2] = Ethayarajh and Jurafsky, Attention Flows are Shapley Value Explanations.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666268791676}], "openreview_url": "https://openreview.net/forum?id=pcBJT4bgbpH", "arxiv_id": "2205.15389", "paper_pdf": "papers/pcBJT4bgbpH.pdf", "paper_pdf_sha256": "8e3695108d43694082a958ca8e0c65057752efe072b63ed8bf478366b2821839", "paper_pdf_bytes": 8711270, "paper_pdf_source": "openreview", "code_url": "https://github.com/reactive-systems/ml2", "code_repository": "reactive-systems/ml2", "code_commit": "33d9696c94de6d27aa836ae8118118a7277ff35c", "code_archive": "repos/pcBJT4bgbpH.zip", "code_archive_sha256": "0c7557ed1e26ad1408b973c8a1f0e8f05314334a3d659f408e52f6f83f174a4a", "code_archive_bytes": 556930, "code_file_count": 436, "code_extensions": {".py": 411, ".ipynb": 23, ".sh": 2}, "github_disk_usage_kb": 833, "github_languages": {"Python": 1368065, "Jupyter Notebook": 116581, "Dockerfile": 15364, "Shell": 2114}, "github_archived": false, "github_pushed_at": "2025-04-01T14:13:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/attention-flows-for-general-transformers"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "TQ75Md-FqQp", "year": 2022, "status": "rejected", "title": "Efficient and Modular Implicit Differentiation", "authors": ["Mathieu Blondel", "Quentin Berthet", "marco cuturi", "Roy Frostig", "Stephan Hoyer", "Felipe Llinares-López", "Fabian Pedregosa", "Jean-Philippe Vert"], "authorids": ["~Mathieu_Blondel1", "~Quentin_Berthet2", "~marco_cuturi2", "~Roy_Frostig1", "~Stephan_Hoyer1", "~Felipe_Llinares-López1", "~Fabian_Pedregosa1", "~Jean-Philippe_Vert1"], "authors_source": "OpenReview API", "abstract": "Automatic differentiation (autodiff) has revolutionized machine learning.  It allows expressing complex computations by composing elementary ones in creative ways and removes the tedious burden of computing their derivatives by hand. More recently, differentiation of optimization problem solutions has attracted a great deal of research, with applications as a layer in a neural network, and in bi-level optimization, including hyper-parameter optimization. However, the formulae for these derivatives often involves a tedious manual derivation and implementation. In this paper, we propose a unified, efficient and modular approach for implicit differentiation of optimization problems. In our approach, the user defines directly in Python a function $F$ capturing the optimality conditions of the problem to be differentiated. Once this is done, we leverage autodiff of $F$ to automatically differentiate the optimization problem. This way, our approach combines the benefits of implicit differentiation and autodiff.  We show that seemingly simple principles allow to recover all recently proposed implicit differentiation methods and create new ones easily. We describe in details a JAX implementation of our framework and demonstrate the ease of differentiating through optimization problems thanks to it on four diverse tasks: hyperparameter optimization of multiclass SVMs, dataset distillation, task-driven dictionary learning and sensitivity analysis of molecular dynamics.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "IwKMNAl6nic", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4362/Reviewer_Q3Lr"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "10: strong accept, should be highlighted at the conference", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a modular and efficient framework along with its JAX implementation for the implicit differentiation of optimization problems. The user defines the function F capturing the optimality conditions of the problem to be differentiated; then the framework combines implicit differentiation and autodiff of F to automatically differentiate the optimization problem. The proposed framework is labeled as efficient, since it doesn’t have to unroll the computational graph like in autodiff, and modular since it doesn’t require case-by-case mathematical derivation like in implicit differentiation.\nThe authors show that existing implicit differentiation methods can be instantiated in their framework. They provide and empirically validate new bounds on the Jacobian error when the optimization problem is only solved approximately.\nThe authors implemented four illustrative applications of their framework ( Hyperparameter Optimization Of Multiclass SVM; Dataset Distillation; Task-Driven Dictionary Learning; Sensitivity Analysis Of Molecular Dynamics).\nCode and implementation in JAX are provided along with the paper.\n", "review_text": "The paper proposes a modular and efficient framework along with its JAX implementation for the implicit differentiation of optimization problems. Firstly, the user defines the function F capturing the optimality conditions of the problem to be differentiated. Then the method frames the differentiation problem as a resolution of the linear system of equations (eq. 2) and combines implicit differentiation and autodiff of F to automatically differentiate the optimization problem.\nThe main novelty of the paper resides in the efficiency and modularity of the framework proposed to solve bi-level optimization problems. This framework allows to abstract away low-level details and significantly lowers the barrier to use implicit differentiation. \nBi-level optimization problems are ubiquitous in ML (hyperparameter optimization, meta-learning, NAS) and the framework proposed allows to address these problems efficiently.\n\nThe paper is well written and touches upon very relevant topics to the community. The framework, its implementation, and the theoretical insights of the Jacobian error are novel and useful contributions. Overall, the proposed method can positively impact the community since abstracting away low-level details of implicit differentiation allows opening new research directions. Therefore I strongly encourage accepting the paper (modulo minor comments below) and I think it should be highlighted at the conference.\n\n\nCOMMENTS:\n\n1. Figure 1 proposes to the reader an illustrative example of the framework. Although the mathematical notation in the paper is well defined in a specific paragraph, the interpretation of the code functions/notation is left to the reader. For example, when you define “X_tr, y_tr = load_data()” the meaning of X_tr is left to the reader (might be X_transpose as well as X_train). Furthermore, an inexperienced reader with the numpy/jax notation might misunderstand the example proposed (functions jnp.eye and jnp.linalg.solve are not defined). I suggest you add line comments or a paragraph with the code notation.\n2. In section 2.1 General Principle, the inner working of the method is proposed; however, it is difficult for me to untangle the novelties presented with respect to what is already known in the literature. Is the procedure of differentiating a root as a linear system part of the novelty? Or the main novelty resides in how to solve the linear system? Please clarify your contribution in this section.\n3. The paper doesn’t contain an explicit definition of “Implicit Differentiation”. I think it would be beneficial to the reader to have a brief explicit definition of Implicit Differentiation.\n\n\nMINOR TYPOS/GRAMMAR CORRECTION:\n- “exiting implicit differentiation” -> “existing implicit differentiation”\n- “we hope that this paper” -> “we hope this paper”\n- “The derivation and implementation in these works is always case-by-case” -> “The derivation and implementation in these works are always case-by-case”\n- “Often times” -> “Oftentimes”\n- “this learned data set achieves small loss” -> “this learned data set achieves a small loss”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a modular and efficient framework along with its JAX implementation for the implicit differentiation of optimization problems. The user defines the function F capturing the optimality conditions of the problem to be differentiated; then the framework combines implicit differentiation and autodiff of F to automatically differentiate the optimization problem. The proposed framework is labeled as efficient, since it doesn’t have to unroll the computational graph like in autodiff, and modular since it doesn’t require case-by-case mathematical derivation like in implicit differentiation.\nThe authors show that existing implicit differentiation methods can be instantiated in their framework. They provide and empirically validate new bounds on the Jacobian error when the optimization problem is only solved approximately.\nThe authors implemented four illustrative applications of their framework ( Hyperparameter Optimization Of Multiclass SVM; Dataset Distillation; Task-Driven Dictionary Learning; Sensitivity Analysis Of Molecular Dynamics).\nCode and implementation in JAX are provided along with the paper.\n", "main_review": "The paper proposes a modular and efficient framework along with its JAX implementation for the implicit differentiation of optimization problems. Firstly, the user defines the function F capturing the optimality conditions of the problem to be differentiated. Then the method frames the differentiation problem as a resolution of the linear system of equations (eq. 2) and combines implicit differentiation and autodiff of F to automatically differentiate the optimization problem.\nThe main novelty of the paper resides in the efficiency and modularity of the framework proposed to solve bi-level optimization problems. This framework allows to abstract away low-level details and significantly lowers the barrier to use implicit differentiation. \nBi-level optimization problems are ubiquitous in ML (hyperparameter optimization, meta-learning, NAS) and the framework proposed allows to address these problems efficiently.\n\nThe paper is well written and touches upon very relevant topics to the community. The framework, its implementation, and the theoretical insights of the Jacobian error are novel and useful contributions. Overall, the proposed method can positively impact the community since abstracting away low-level details of implicit differentiation allows opening new research directions. Therefore I strongly encourage accepting the paper (modulo minor comments below) and I think it should be highlighted at the conference.\n\n\nCOMMENTS:\n\n1. Figure 1 proposes to the reader an illustrative example of the framework. Although the mathematical notation in the paper is well defined in a specific paragraph, the interpretation of the code functions/notation is left to the reader. For example, when you define “X_tr, y_tr = load_data()” the meaning of X_tr is left to the reader (might be X_transpose as well as X_train). Furthermore, an inexperienced reader with the numpy/jax notation might misunderstand the example proposed (functions jnp.eye and jnp.linalg.solve are not defined). I suggest you add line comments or a paragraph with the code notation.\n2. In section 2.1 General Principle, the inner working of the method is proposed; however, it is difficult for me to untangle the novelties presented with respect to what is already known in the literature. Is the procedure of differentiating a root as a linear system part of the novelty? Or the main novelty resides in how to solve the linear system? Please clarify your contribution in this section.\n3. The paper doesn’t contain an explicit definition of “Implicit Differentiation”. I think it would be beneficial to the reader to have a brief explicit definition of Implicit Differentiation.\n\n\nMINOR TYPOS/GRAMMAR CORRECTION:\n- “exiting implicit differentiation” -> “existing implicit differentiation”\n- “we hope that this paper” -> “we hope this paper”\n- “The derivation and implementation in these works is always case-by-case” -> “The derivation and implementation in these works are always case-by-case”\n- “Often times” -> “Oftentimes”\n- “this learned data set achieves small loss” -> “this learned data set achieves a small loss”\n", "summary_of_the_review": "The paper proposes a modular and efficient framework along with its JAX implementation for the implicit differentiation of optimization problems. This framework allows to abstract away low-level details and significantly lowers the barrier to use implicit differentiation. Bi-level optimization problems are ubiquitous in ML (hyperparameter optimization, meta-learning, NAS) and a procedure to automatically and efficiently tackle them is much needed by the community. I believe that this paper should be accepted and highlighted at the conference.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "10: strong accept, should be highlighted at the conference", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635921200297}, {"id": "TWJeqkh0P3W", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4362/Reviewer_h9WU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper introduces a Jax package for implicitly differentiating various numerical solvers. Concretely, the authors develop a systemic methodology for producing gradients for a variety of optimization problems. Then, the authors prove that the Jacobian solution to the approximate numerical solution produces close enough gradients. Finally, the authors show the power of their framework on four test tasks.", "review_text": "## Strengths\n\n* The proposed package is a nice unification of the core implicit differentiation procedures. I believe that this construction has the potential to speed up development and application of implicit differentiation methods.\n* The theorem is a good, nice-to-know theorem for the literature.\n* The test tasks are nice and varied and showcase the different ways the package could be applied.\n\n## Weaknesses\n\n* The paper can effectively be split into the \"code\" portion, the \"theorem\" portion, and the \"examples\" portion. None of the portions, by themselves, meet the threshold for ICLR. In particular, the code is a package of preexisting work, and I am unaware of any similar types of packages being published to major conferences (see for example, the baseline TorchDyn library [1]). The theorem also has a straightforward, routine proof, so not much theoretical insight is gained here. Finally, the examples are smaller scale and are more toy than previous work such as [2, 3].\n* The core contribution, the software, is not presented in a digestible manner. For example, the only code in the main paper is given in Figure 1, and that code is confusing. In particular, the functions $f$ and $F$ are not good names for variables, the jax.grad call operates on two arguments so should have argnums=0, and the choice to write out the functions in pure jax code clutters up the entire block. In addition, I would hope that Section 2.2 could be rewritten with more code examples. Currently, this section feels like a rehash of preexisting methods, and code would be helpful in showing how one can directly implement these in practice.\n\n## References\n\n[1] https://arxiv.org/abs/2009.09346\n\n[2] https://arxiv.org/abs/1703.00443\n\n[3] https://arxiv.org/abs/1909.01377", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a Jax package for implicitly differentiating various numerical solvers. Concretely, the authors develop a systemic methodology for producing gradients for a variety of optimization problems. Then, the authors prove that the Jacobian solution to the approximate numerical solution produces close enough gradients. Finally, the authors show the power of their framework on four test tasks.", "main_review": "## Strengths\n\n* The proposed package is a nice unification of the core implicit differentiation procedures. I believe that this construction has the potential to speed up development and application of implicit differentiation methods.\n* The theorem is a good, nice-to-know theorem for the literature.\n* The test tasks are nice and varied and showcase the different ways the package could be applied.\n\n## Weaknesses\n\n* The paper can effectively be split into the \"code\" portion, the \"theorem\" portion, and the \"examples\" portion. None of the portions, by themselves, meet the threshold for ICLR. In particular, the code is a package of preexisting work, and I am unaware of any similar types of packages being published to major conferences (see for example, the baseline TorchDyn library [1]). The theorem also has a straightforward, routine proof, so not much theoretical insight is gained here. Finally, the examples are smaller scale and are more toy than previous work such as [2, 3].\n* The core contribution, the software, is not presented in a digestible manner. For example, the only code in the main paper is given in Figure 1, and that code is confusing. In particular, the functions $f$ and $F$ are not good names for variables, the jax.grad call operates on two arguments so should have argnums=0, and the choice to write out the functions in pure jax code clutters up the entire block. In addition, I would hope that Section 2.2 could be rewritten with more code examples. Currently, this section feels like a rehash of preexisting methods, and code would be helpful in showing how one can directly implement these in practice.\n\n## References\n\n[1] https://arxiv.org/abs/2009.09346\n\n[2] https://arxiv.org/abs/1703.00443\n\n[3] https://arxiv.org/abs/1909.01377", "summary_of_the_review": "I enjoyed the paper and think that the package can contribute much to popularize implicit differentiation tools. However, I feel that the paper is not sufficiently novel on any front to warrant publication. Furthermore, I would encourage the authors to present the main contribution, the package, in a more digestible manner.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635834295126}, {"id": "wQTyRy-gPH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4362/Reviewer_b1rc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides a unified tool for combining the implicit differentiation technique and the  automatic differentiation method widely used in existing deep learning packages such as PyTorch and TensorFlow. The proposed implementation is easy to use for numeric optimization such as bilevel optimization, meta-learning and hyperparameter optimization because it covers many existing schemes such as fixed point, KKT point, projected method. In the experiments, the authors illustrate how their tool can be useful to simply the implementation. ", "review_text": "My detailed comments are given as below.\n\nStrength: \n\n1. The presentation of this paper is good and most parts are easy to follow. The motivation of this paper is clear, i.e., to provide a unified and easy-to-use implementation for implicit differentiation by leveraging the tool from automatic differentiation. The tool developed here seems to be useful and cover many existing designs of interest. \n\n2. Since implicit differentiation is useful and has critical use in practical applications now, e.g., meta-learning and automated machine learning, I feel an easy-to-use implementation can further promote the developments of more advanced algorithms. I feel this is good. \n\nWeakness:\n\n1. One of my concerns is the novelty of this paper. Since all things are not new, e.g., implicit differentiation, automatic differentiation, matrix-vector product computation in deep learning package, I feel this paper does not contribute a new algorithm design, but an easy-to-use package for many practical algorithms.\n\n2. The analysis is not new as well, since there are already many works on studying the iteration complexity of the response Jacobian and the hypergradient, which I find are missing in this paper. I list some of them and highly encourage the authors to add them (and some related works therein) and provide a short discussion. \n\n1) K. Ji, J. Yang, and Y. Liang. Bilevel optimization: Convergence analysis and enhanced design. In Proc.\nInternational Conference on Machine Learning (ICML), 2021.\n\n2) Grazzi, R., Franceschi, L., Pontil, M., and Salzo, S. On the iteration complexity of hypergradient computation. In Proc. International Conference on Machine Learning (ICML), 2020.\n\n3. Experiments are conducted over only small datasets and models, e.g., synthetic data and no DNN involved. I am wondering whether this package is generalizable enough for such more practical settings. Therefore, I suggest the authors can provide some experiments on larger datasets and models. \n \n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper provides a unified tool for combining the implicit differentiation technique and the  automatic differentiation method widely used in existing deep learning packages such as PyTorch and TensorFlow. The proposed implementation is easy to use for numeric optimization such as bilevel optimization, meta-learning and hyperparameter optimization because it covers many existing schemes such as fixed point, KKT point, projected method. In the experiments, the authors illustrate how their tool can be useful to simply the implementation. ", "main_review": "My detailed comments are given as below.\n\nStrength: \n\n1. The presentation of this paper is good and most parts are easy to follow. The motivation of this paper is clear, i.e., to provide a unified and easy-to-use implementation for implicit differentiation by leveraging the tool from automatic differentiation. The tool developed here seems to be useful and cover many existing designs of interest. \n\n2. Since implicit differentiation is useful and has critical use in practical applications now, e.g., meta-learning and automated machine learning, I feel an easy-to-use implementation can further promote the developments of more advanced algorithms. I feel this is good. \n\nWeakness:\n\n1. One of my concerns is the novelty of this paper. Since all things are not new, e.g., implicit differentiation, automatic differentiation, matrix-vector product computation in deep learning package, I feel this paper does not contribute a new algorithm design, but an easy-to-use package for many practical algorithms.\n\n2. The analysis is not new as well, since there are already many works on studying the iteration complexity of the response Jacobian and the hypergradient, which I find are missing in this paper. I list some of them and highly encourage the authors to add them (and some related works therein) and provide a short discussion. \n\n1) K. Ji, J. Yang, and Y. Liang. Bilevel optimization: Convergence analysis and enhanced design. In Proc.\nInternational Conference on Machine Learning (ICML), 2021.\n\n2) Grazzi, R., Franceschi, L., Pontil, M., and Salzo, S. On the iteration complexity of hypergradient computation. In Proc. International Conference on Machine Learning (ICML), 2020.\n\n3. Experiments are conducted over only small datasets and models, e.g., synthetic data and no DNN involved. I am wondering whether this package is generalizable enough for such more practical settings. Therefore, I suggest the authors can provide some experiments on larger datasets and models. \n \n\n\n\n\n", "summary_of_the_review": "Overall, I feel this work provides a useful tool for implementing the implicit differentiation in various scenarios. Although I feel the novelty of this work is not that high, I am still slightly positive about it. I am open to increase my score if the authors can address my concerns and add missing related works. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635613809873}], "openreview_url": "https://openreview.net/forum?id=TQ75Md-FqQp", "arxiv_id": "2105.15183", "paper_pdf": "papers/TQ75Md-FqQp.pdf", "paper_pdf_sha256": "c80fc1805f72a0ce8f7321dc71ace7191bda8ab64b2d513f37abfdf1e27a52f5", "paper_pdf_bytes": 2202952, "paper_pdf_source": "openreview", "code_url": "https://github.com/google/jaxopt", "code_repository": "google/jaxopt", "code_commit": "176b13830bc552236b32a4c3c4654f196c8b4cd6", "code_archive": "repos/TQ75Md-FqQp.zip", "code_archive_sha256": "97c9a4807ae55bc66f2027544a5f2d6e295aa18a7b5d0e49c9d7edf12986feb2", "code_archive_bytes": 2247572, "code_file_count": 126, "code_extensions": {".py": 118, ".ipynb": 8}, "github_disk_usage_kb": 3535, "github_languages": {"Python": 922179}, "github_archived": false, "github_pushed_at": "2026-09-07T21:43:54Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-and-modular-implicit"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "5wmNjjvGOXh", "year": 2021, "status": "rejected", "title": "Selfish Sparse RNN Training", "authors": ["Shiwei Liu", "Decebal Constantin Mocanu", "Yulong Pei", "Mykola Pechenizkiy"], "authorids": ["~Shiwei_Liu2", "~Decebal_Constantin_Mocanu1", "~Yulong_Pei1", "~Mykola_Pechenizkiy1"], "authors_source": "OpenReview API", "abstract": "Sparse neural networks have been widely applied to reduce the necessary resource requirements to train and deploy over-parameterized deep neural networks. For inference acceleration, methods that induce sparsity from a pre-trained dense network (dense-to-sparse) work effectively. Recently, dynamic sparse training (DST) has been proposed to train sparse neural networks without pre-training a large and dense network (sparse-to-sparse), so that the training process can also be accelerated. However, previous sparse-to-sparse methods mainly focus on Multilayer Perceptron Networks (MLPs) and Convolutional Neural Networks (CNNs), failing to match the performance of dense-to-sparse methods in Recurrent Neural Networks (RNNs) setting. In this paper, we propose an approach to train sparse RNNs with a fixed parameter count in one single run, without compromising performance. During training, we allow RNN layers to have a non-uniform redistribution across cell weights for a better regularization. Further, we introduce SNT-ASGD, a variant of the averaged stochastic gradient optimizer, which significantly improves the performance of all sparse training methods for RNNs. Using these strategies, we achieve state-of-the-art sparse training results, even better than dense model results, with various types of RNNs on Penn TreeBank and Wikitext-2 datasets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "-hg1W59XWZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1613/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper claims that the previous sparse training methods mainly focus on MLP and CNN, and fail to perform very well in RNNs. Hence, the authors proposed an approach to train sparse RNNs with a fixed FLOPs budget.\nThe proposed technique is based on defining a mask matrix $M$ and refining it during training. It is initialized randomly to have the desired sparsity level $S$. After each training epoch, a fraction $p$ of the weights with the smallest magnitude is removed, i.e., those locations are zeroed in the mask M.\nNext, the same amount of parameters are randomly added to M again. \nMoreover, a variant of the averaged stochastic gradient optimizer (SNT-ASGD) is developed for the training of sparse RNN to account for the effect of weight masks during training.\nThey showed that in practice, the requirements for efficient sparse training of RNNs are different than CNN and MLP.\n\nStrengths:\nBy adding some refinements and tweaks to the existing techniques (masking for sparse training and adapting the NT-ASGD), the authors were able to achieve good performance to train sparse RNNs. The paper has a rather extensive set of simulations and experimental setups to analyze the best setup which yields good sparse training, e.g., comparing uniform vs ER distribution for masks, sensitivity to hyperparameters, ... Moreover, they have considered a fairly diverse set of RNN architectures to evaluate their method.\n\nWeaknesses and questions:\nCompared to the existing methods, the technical novelty of the paper is minor. It can be seen as some tweaks and improvements to the existing ones (although I admit that those changes are essential for the method to work for RNN.). \nWhat is special about the method that makes it specific to RNN? In other words, is it possible to use the same method for sparse training of MLP and CNN?\nA minor issue with the paper is the FLOPS analysis the authors used. Effectively, they use the sparsity of the parameters as a measure of FLOPS, not the actual FLOPS that might depend on the sparsity structure, HW, or software implementation. It would be a good idea to directly mention and use total sparsity, instead of FLOPS which can mislead the readers.\n\nSome parts of the method are not clear enough, e.g., \n1. In the paper, it is stated that \"magnitude weight removal\" is applied to non-RNN layers. Do the authors mean that for the parameters of RNN, this step is skipped?\n2. In \"cell weight redistribution\", it is suggested that the \"magnitude weight removal\" is applied to the whole set of RNN parameters $\\{\\theta_1, \\ldots, \\theta_t\\}$. However, in \"random weight growth\", it is mentioned that the same number of weights is grown immediately after weight removal, i.e., $R$ and $P$ have the same number of 1's. So, does it mean that the number of 1's in mask $M_i$ for each weight $\\theta_i$ ($1\\leq i \\leq t$) remains fixed S during training?\n3. Another aspect of training that is unclear for me is the parameters that are updated. Is $\\theta$ updated during training or only $\\theta_s$ is updated? As a result, if a weight is removed in one epoch and its value at the time of removal was $\\alpha$, and later regrown at another epoch, is its initial value set to 0 or started from its previous value before \"weight removal\", i.e. $\\alpha$?\n4. Did the authors add any regularizer (e.g., $\\ell_1$) to the training loss to improve sparsity in their experiments?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "minor technical novelities that lead to improved performance over state of the art", "review": "The paper claims that the previous sparse training methods mainly focus on MLP and CNN, and fail to perform very well in RNNs. Hence, the authors proposed an approach to train sparse RNNs with a fixed FLOPs budget.\nThe proposed technique is based on defining a mask matrix $M$ and refining it during training. It is initialized randomly to have the desired sparsity level $S$. After each training epoch, a fraction $p$ of the weights with the smallest magnitude is removed, i.e., those locations are zeroed in the mask M.\nNext, the same amount of parameters are randomly added to M again. \nMoreover, a variant of the averaged stochastic gradient optimizer (SNT-ASGD) is developed for the training of sparse RNN to account for the effect of weight masks during training.\nThey showed that in practice, the requirements for efficient sparse training of RNNs are different than CNN and MLP.\n\nStrengths:\nBy adding some refinements and tweaks to the existing techniques (masking for sparse training and adapting the NT-ASGD), the authors were able to achieve good performance to train sparse RNNs. The paper has a rather extensive set of simulations and experimental setups to analyze the best setup which yields good sparse training, e.g., comparing uniform vs ER distribution for masks, sensitivity to hyperparameters, ... Moreover, they have considered a fairly diverse set of RNN architectures to evaluate their method.\n\nWeaknesses and questions:\nCompared to the existing methods, the technical novelty of the paper is minor. It can be seen as some tweaks and improvements to the existing ones (although I admit that those changes are essential for the method to work for RNN.). \nWhat is special about the method that makes it specific to RNN? In other words, is it possible to use the same method for sparse training of MLP and CNN?\nA minor issue with the paper is the FLOPS analysis the authors used. Effectively, they use the sparsity of the parameters as a measure of FLOPS, not the actual FLOPS that might depend on the sparsity structure, HW, or software implementation. It would be a good idea to directly mention and use total sparsity, instead of FLOPS which can mislead the readers.\n\nSome parts of the method are not clear enough, e.g., \n1. In the paper, it is stated that \"magnitude weight removal\" is applied to non-RNN layers. Do the authors mean that for the parameters of RNN, this step is skipped?\n2. In \"cell weight redistribution\", it is suggested that the \"magnitude weight removal\" is applied to the whole set of RNN parameters $\\{\\theta_1, \\ldots, \\theta_t\\}$. However, in \"random weight growth\", it is mentioned that the same number of weights is grown immediately after weight removal, i.e., $R$ and $P$ have the same number of 1's. So, does it mean that the number of 1's in mask $M_i$ for each weight $\\theta_i$ ($1\\leq i \\leq t$) remains fixed S during training?\n3. Another aspect of training that is unclear for me is the parameters that are updated. Is $\\theta$ updated during training or only $\\theta_s$ is updated? As a result, if a weight is removed in one epoch and its value at the time of removal was $\\alpha$, and later regrown at another epoch, is its initial value set to 0 or started from its previous value before \"weight removal\", i.e. $\\alpha$?\n4. Did the authors add any regularizer (e.g., $\\ell_1$) to the training loss to improve sparsity in their experiments?\n", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603842101129}, {"id": "lhMqduuW5XY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1613/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors propose an approach to train sparse recurrent models, and a sparse variant of the NT-ASGD. The proposed method mixes some interesting novel methodologies and achieves interesting empirical results on Penn Treebank and WikiText-2 language modeling tasks. \nIn general, the paper is well written and interesting, but in section 3 many explanations about the rationale behind some architectural choices of the selfish-RNN methodology are only partially explained, and sometimes they are just related to empirical results (e.g. in the cell weight redistribution). To me, a more theoretical explanation would significantly improve the manuscript readability.\nIn section 4 many different approaches were considered. But there are a few points that are not clear. The authors report the results of a “small\" dense network, but no information about this model is reported in the text.\nReading the results reported in table 5 of the appendix, I found it interesting that the performance of the DSR improves significantly by using SNT-ASGD instead of Adam  (it outperforms the Selfish-RNN). This table shows how much the optimizer influences model performance.  Even if the ablation study reported in appendix A highlights the benefits of the SNT-ASGD, the results reported in table 5 show that the impact of this component is even more important than the selfish-RNN. Honestly, I think that is fairer to compare all the methods using the same optimization algorithm, therefore my suggestion is to move this table in the main paper and extend the analysis of these results. \nReading the manuscript it is not clear how the hyper-parameters considered in the experimental campaigns have been chosen. By reading the first part of section 4.1 seems like parameters like the removing rate or the number of epochs are set without performing any validation on them. Even in appendix D, hyper-parameters (e.g. the learning rate, or the batch size) used to test the RHM are just listed. The authors should insert a more extensive explanation about how the hyper-parameters various models/approaches considered in the comparison have been validated. To perform a fair comparison the hyper-parameters of each model should be chosen according to its performance on the validation set.\nIn this regard, it is important also to highlight how the hyper-parameters are chosen because some SOTA models achieved better results. For instance on the Penn Treebank dataset in “On The State Of The Art Of Evaluation In Neural Language Models”, Melis et al. report perplexities on the test set of 59.7.\nexploiting better the research space. The reported results in the paper (and in Appendix L) show the benefits of using this approach, but honestly, to me, it is not clear if it helps in exploring the state space. In general, it is not clear what is the reason why the model benefits from using the random growth approach. Moreover, in “Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware” the gradient guided growth strategy outperforms the other sparse training approaches considered in the paper, even in the RNN case. Therefore a more extended evaluation/discussion of this point is required.\nAnother recently proposed approach that uses sparsity in recurrent models is defined in “Intrinsically Sparse Long Short-Term Memory Networks” by Liu et al. the author should compare this approach with the selfish-LSTM.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting paper that proposes an approach to train sparse recurrent models, and a sparse variant of the NT-ASGD.", "review": "In this paper, the authors propose an approach to train sparse recurrent models, and a sparse variant of the NT-ASGD. The proposed method mixes some interesting novel methodologies and achieves interesting empirical results on Penn Treebank and WikiText-2 language modeling tasks. \nIn general, the paper is well written and interesting, but in section 3 many explanations about the rationale behind some architectural choices of the selfish-RNN methodology are only partially explained, and sometimes they are just related to empirical results (e.g. in the cell weight redistribution). To me, a more theoretical explanation would significantly improve the manuscript readability.\nIn section 4 many different approaches were considered. But there are a few points that are not clear. The authors report the results of a “small\" dense network, but no information about this model is reported in the text.\nReading the results reported in table 5 of the appendix, I found it interesting that the performance of the DSR improves significantly by using SNT-ASGD instead of Adam  (it outperforms the Selfish-RNN). This table shows how much the optimizer influences model performance.  Even if the ablation study reported in appendix A highlights the benefits of the SNT-ASGD, the results reported in table 5 show that the impact of this component is even more important than the selfish-RNN. Honestly, I think that is fairer to compare all the methods using the same optimization algorithm, therefore my suggestion is to move this table in the main paper and extend the analysis of these results. \nReading the manuscript it is not clear how the hyper-parameters considered in the experimental campaigns have been chosen. By reading the first part of section 4.1 seems like parameters like the removing rate or the number of epochs are set without performing any validation on them. Even in appendix D, hyper-parameters (e.g. the learning rate, or the batch size) used to test the RHM are just listed. The authors should insert a more extensive explanation about how the hyper-parameters various models/approaches considered in the comparison have been validated. To perform a fair comparison the hyper-parameters of each model should be chosen according to its performance on the validation set.\nIn this regard, it is important also to highlight how the hyper-parameters are chosen because some SOTA models achieved better results. For instance on the Penn Treebank dataset in “On The State Of The Art Of Evaluation In Neural Language Models”, Melis et al. report perplexities on the test set of 59.7.\nexploiting better the research space. The reported results in the paper (and in Appendix L) show the benefits of using this approach, but honestly, to me, it is not clear if it helps in exploring the state space. In general, it is not clear what is the reason why the model benefits from using the random growth approach. Moreover, in “Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware” the gradient guided growth strategy outperforms the other sparse training approaches considered in the paper, even in the RNN case. Therefore a more extended evaluation/discussion of this point is required.\nAnother recently proposed approach that uses sparsity in recurrent models is defined in “Intrinsically Sparse Long Short-Term Memory Networks” by Liu et al. the author should compare this approach with the selfish-LSTM.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603734269512}, {"id": "x4_uParF5zR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1613/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: \nThe authors improve sparse for recurrent neural networks by developing a greedy redistribution rule for gates and adapting the ASGD optimizer for sparse networks. The work provides good results and a rich analysis of their and related methods.\n\nStrong points:\n- very rigorous experimental setup and analysis\n- Solid evidence for many new insights into some sparse training phenomena. The work provides broadens our understanding of sparse training.\n\nWeak points:\n- Some might complain that RNNs are outdated. I see this only as a minor weak point. Indeed, RNNs are not much used anymore, but many of the insights the paper provides are quite universal.\n- The fixed FLOPS only seems to be a by-product of the algorithm and particular network structure but not necessarily an algorithmic contribution. This makes the paper a bit confusing.\n\nRecommendation (short):\nThis is a very solid paper with exemplary experimentation and analysis. It provides many unique insights that are very valuable for anyone who wants to work in the field of sparse training. I recommend accepting this paper.\n\nRecommendation (long):\nI think this paper is one of these papers, which is a very solid all-around. The authors invested quite a bit of time in creating rigorous experimental setups that test hypotheses. In particular, I like the graph analysis of sparse connective between networks. Findings of different initialization schemes and performance of other sparse training methods are precious and make the overall literature on sparse training robust. I can see that this paper may seem a bit boring and less impactful to some reviewers, but good science like this is not about being exciting but about providing rigorous results for a small problem. This paper does exactly that. I think any good conference should encourage good science by accepting papers like this one.\n\nComments for authors:\nSolid work. Here some additional comments and questions.\n- Please feed your paper through a grammar/spellchecker. There are multiple errors which make the paper hard to read in some sections\n- It is not entirely clear why ASGD is needed for good performance. Can you elaborate, please?\n- Do you have any idea how does ER initialization relates to eigenvalues of recurrent matrices? If you can make a connection here, it would be a quite insightful addition to the paper since the top eigenvalue of the recurrent matrix determines the overall long-term behavior of the recurrent matrix and is known to influence behavior.\n- I would drop the fixed FLOPS contribution and focus on the other parts of the paper. You have more than enough contributions, and the space is better devoted to making the other contributions as clear as possible.\n- The cell weight redistribution algorithm description is unclear. A weight cannot have \"more parameters, I think you mean to say gate-neurons with large magnitude weights gain more parameters over time.\n- The sparse topology algorithm: Is the correlation between weights computed overall test set outputs between two networks/weights?\n- Figure 3, unclear. What does Figure 3 (left) show exactly? It is unclear what random initialization means: different sparsity patterns, different weight values, or both? What does the seed do here? Does it affect sparsity pattern, data order, weight values, etc.?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Very solid paper on sparse training of RNNs which presents insights valuable for important questions on sparse training.", "review": "Summary: \nThe authors improve sparse for recurrent neural networks by developing a greedy redistribution rule for gates and adapting the ASGD optimizer for sparse networks. The work provides good results and a rich analysis of their and related methods.\n\nStrong points:\n- very rigorous experimental setup and analysis\n- Solid evidence for many new insights into some sparse training phenomena. The work provides broadens our understanding of sparse training.\n\nWeak points:\n- Some might complain that RNNs are outdated. I see this only as a minor weak point. Indeed, RNNs are not much used anymore, but many of the insights the paper provides are quite universal.\n- The fixed FLOPS only seems to be a by-product of the algorithm and particular network structure but not necessarily an algorithmic contribution. This makes the paper a bit confusing.\n\nRecommendation (short):\nThis is a very solid paper with exemplary experimentation and analysis. It provides many unique insights that are very valuable for anyone who wants to work in the field of sparse training. I recommend accepting this paper.\n\nRecommendation (long):\nI think this paper is one of these papers, which is a very solid all-around. The authors invested quite a bit of time in creating rigorous experimental setups that test hypotheses. In particular, I like the graph analysis of sparse connective between networks. Findings of different initialization schemes and performance of other sparse training methods are precious and make the overall literature on sparse training robust. I can see that this paper may seem a bit boring and less impactful to some reviewers, but good science like this is not about being exciting but about providing rigorous results for a small problem. This paper does exactly that. I think any good conference should encourage good science by accepting papers like this one.\n\nComments for authors:\nSolid work. Here some additional comments and questions.\n- Please feed your paper through a grammar/spellchecker. There are multiple errors which make the paper hard to read in some sections\n- It is not entirely clear why ASGD is needed for good performance. Can you elaborate, please?\n- Do you have any idea how does ER initialization relates to eigenvalues of recurrent matrices? If you can make a connection here, it would be a quite insightful addition to the paper since the top eigenvalue of the recurrent matrix determines the overall long-term behavior of the recurrent matrix and is known to influence behavior.\n- I would drop the fixed FLOPS contribution and focus on the other parts of the paper. You have more than enough contributions, and the space is better devoted to making the other contributions as clear as possible.\n- The cell weight redistribution algorithm description is unclear. A weight cannot have \"more parameters, I think you mean to say gate-neurons with large magnitude weights gain more parameters over time.\n- The sparse topology algorithm: Is the correlation between weights computed overall test set outputs between two networks/weights?\n- Figure 3, unclear. What does Figure 3 (left) show exactly? It is unclear what random initialization means: different sparsity patterns, different weight values, or both? What does the seed do here? Does it affect sparsity pattern, data order, weight values, etc.?\n", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603495717939}, {"id": "-nMGFcluav", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1613/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In this paper, the authors studied the possibility of sparsity exploration in Recurrent Neural Networks (RNNs) training. The main contributions include two parts: (1) Selfish-RNN training algorithm in Section 3.1 (2) SNT-ASGD optimizer in Section 3.2. The key idea of the Selfish-RNN training algorithm is a non-uniform redistribution across cell weights for better regularization. The authors mentioned previous sparse training techniques mainly focus on Multilayer Perceptron Networks (MLPs) and Convolutional Neural Networks (CNNs) rather than RNNs. This claim seems to be doubtful because one-time SVD + fine-tuning usually works very well for most RNN training applications in the industry.\n\nOverall, this paper is carefully written and provides some interesting empirical results. However, due to the lack of some important information, it is hard to evaluate the contribution of this paper.\n\nHere are some of my questions.\n\nSNT-ASGD needs to save the weights w_i,t from iteration Ti to iteration K, will that cost additional memory?\n\nThe authors mentioned that they picked Adam optimizer for SET, DSR, SNFS, and RigL. Is Adam the best optimizer to build a strong baseline? I suspect Adam may not be the best optimizer for each of them.\n\nThe authors need to give more information on the hyper-parameters like the learning rate. The selection of hyper-parameters usually significantly affects the convergence/generalization performance of an RNN model. For example, the way of learning rate decay has a big impact on the performance of training Penn TreeBank dataset.\n\nCan the authors report the training epochs and wall-clock time (e.g. in Table 2)? The sparsity typically makes modern hardware like GPUs perform poorly. That may be a concern. That’s the reason why researchers are studying structure sparsity. For future work, an analysis of computation (flops) to communication (memory access frequency) ratio seems to be necessary.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "lack of some important information", "review": "In this paper, the authors studied the possibility of sparsity exploration in Recurrent Neural Networks (RNNs) training. The main contributions include two parts: (1) Selfish-RNN training algorithm in Section 3.1 (2) SNT-ASGD optimizer in Section 3.2. The key idea of the Selfish-RNN training algorithm is a non-uniform redistribution across cell weights for better regularization. The authors mentioned previous sparse training techniques mainly focus on Multilayer Perceptron Networks (MLPs) and Convolutional Neural Networks (CNNs) rather than RNNs. This claim seems to be doubtful because one-time SVD + fine-tuning usually works very well for most RNN training applications in the industry.\n\nOverall, this paper is carefully written and provides some interesting empirical results. However, due to the lack of some important information, it is hard to evaluate the contribution of this paper.\n\nHere are some of my questions.\n\nSNT-ASGD needs to save the weights w_i,t from iteration Ti to iteration K, will that cost additional memory?\n\nThe authors mentioned that they picked Adam optimizer for SET, DSR, SNFS, and RigL. Is Adam the best optimizer to build a strong baseline? I suspect Adam may not be the best optimizer for each of them.\n\nThe authors need to give more information on the hyper-parameters like the learning rate. The selection of hyper-parameters usually significantly affects the convergence/generalization performance of an RNN model. For example, the way of learning rate decay has a big impact on the performance of training Penn TreeBank dataset.\n\nCan the authors report the training epochs and wall-clock time (e.g. in Table 2)? The sparsity typically makes modern hardware like GPUs perform poorly. That may be a concern. That’s the reason why researchers are studying structure sparsity. For future work, an analysis of computation (flops) to communication (memory access frequency) ratio seems to be necessary.", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1602665126639}], "openreview_url": "https://openreview.net/forum?id=5wmNjjvGOXh", "arxiv_id": "2101.09048", "paper_pdf": "papers/5wmNjjvGOXh.pdf", "paper_pdf_sha256": "8d88788cb8f686585bc754a7b77b04ccd1adf48f5b65a4e949d3d0417c3ad671", "paper_pdf_bytes": 1853573, "paper_pdf_source": "openreview", "code_url": "https://github.com/Shiweiliuiiiiiii/Selfish-RNN", "code_repository": "Shiweiliuiiiiiii/Selfish-RNN", "code_commit": "06b03db04bfe93f20f8fc703e9a78a83cea777a7", "code_archive": "repos/5wmNjjvGOXh.zip", "code_archive_sha256": "0504ce22b9e498a9509623020f599bba977ede7efdc5b9201b8fde7f52123e55", "code_archive_bytes": 2046934, "code_file_count": 13, "code_extensions": {".py": 13}, "github_disk_usage_kb": 2101, "github_languages": {"Python": 104692}, "github_archived": false, "github_pushed_at": "2021-10-08T03:52:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/selfish-sparse-rnn-training-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uwoA5iyTC6", "year": 2026, "status": "rejected", "title": "Time Series Forecasting as Reasoning: A Slow-Thinking Approach with Reinforced LLMs", "authors": ["Yucong Luo", "Yitong Zhou", "Mingyue Cheng", "Jiahao Wang", "Daoyu Wang"], "authorids": ["~Yucong_Luo1", "~Yitong_Zhou2", "~Mingyue_Cheng1", "~Jiahao_Wang25", "~Daoyu_Wang1"], "authors_source": "OpenReview API", "abstract": "To advance time series forecasting (TSF), various methods have been proposed to improve prediction accuracy, evolving from statistical techniques to data-driven deep learning architectures. Despite their effectiveness, most existing methods still adhere to a fast thinking paradigm - relying on pattern recognition and trend prediction as their core modeling philosophy, lacking an explicit \"thinking process\" that incorporates intermediate time series reasoning. Meanwhile, emerging slow-thinking LLMs (e.g., ChatGPT-o1) have shown remarkable multi-step reasoning capabilities, offering an alternative way to overcome these issues. However, prompt engineering alone presents several limitations—including high computational cost, privacy risks, and limited capacity for in-depth domain-specific time series reasoning. To address these limitations, a more promising approach is to train LLMs to develop slow-thinking capabilities and acquire strong time series reasoning skills. To acquire such slow thinking reasoning capabilities, we propose Time-R1, a two-stage reinforcement fine-tuning framework designed to enhance multi-step reasoning ability of LLMs for time series forecasting. Specifically, the first stage conducts supervised fine-tuning for warmup adaptation, while the second stage employs reinforcement learning to improve the model's generalization ability. Particularly, we introduce GRIP (group-based relative importance for policy optimization), which utilizes non-uniform sampling along with a fine-grained multi-objective reward specifically designed for time series forecasting to further encourage and optimize the model's exploration of effective reasoning paths. Experiments demonstrate that Time-R1 significantly improves forecast performance across diverse datasets. Source code is available https://anonymous.4open.science/r/Time-R1-NeurIPS-2025/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "aGqVfQCxSg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12694/Reviewer_Sj3u"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper proposes Time-R1, a two-stage reinforcement fine-tuning (RFT) framework designed to enhance LLMs' reasoning capabilities for time series forecasting (TSF). The approach consists of: (1) a warmup supervised fine-tuning (SFT) stage using chain-of-thought (CoT) trajectories generated by DeepSeek-R1, and (2) a reinforcement learning stage employing GRIP, which uses non-uniform sampling and adaptive trajectory weighting with a multi-objective reward function. Experiments on nine datasets demonstrate improvements over baseline methods including traditional deep learning models and LLM-based approaches. The authors argue that slow-thinking reasoning provides better temporal understanding than fast-thinking pattern matching.", "review_text": "This paper proposes Time-R1, a two-stage reinforcement fine-tuning (RFT) framework designed to enhance LLMs' reasoning capabilities for time series forecasting (TSF). The approach consists of: (1) a warmup supervised fine-tuning (SFT) stage using chain-of-thought (CoT) trajectories generated by DeepSeek-R1, and (2) a reinforcement learning stage employing GRIP, which uses non-uniform sampling and adaptive trajectory weighting with a multi-objective reward function. Experiments on nine datasets demonstrate improvements over baseline methods including traditional deep learning models and LLM-based approaches. The authors argue that slow-thinking reasoning provides better temporal understanding than fast-thinking pattern matching.", "strengths": "(1)  The distinction between \"fast-thinking\" (pattern matching) and \"slow-thinking\" (explicit reasoning) paradigms is intuitive and well-articulated. The motivation that time series contain complex temporal logic warranting structured reasoning is compelling.\n(2) The combination of SFT (for format stabilization and basic reasoning) followed by RL (for generalization and reasoning refinement) is a principled approach that addresses practical training challenges. The ablation study (Table 3, Figure 4b) effectively demonstrates the necessity of both stages.", "weaknesses": "(1) Using LLMs to generate synthetic CoT trajectories for fine-tuning is not novel (e.g., similar to approaches in reasoning LLMs). The key distinction from existing work is unclear.\n(2) While the adaptive weighting is presented as novel, it is essentially a variant of importance sampling with softmax normalization. The core algorithmic novelty is limited. The comparison with GRPO (Figure 4a) shows only marginal improvements.\n(3) The paper claims \"slow thinking\" but this mirrors existing reasoning-based LLM approaches (o1, DeepSeek-R1). What is fundamentally new beyond applying them to TSF?\n(4) SFT data is generated using DeepSeek-R1, which itself is a slow-thinking model. This creates a conceptual circularity: improving reasoning for TSF via a model that already has slow-thinking capabilities.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Time-R1, a two-stage reinforcement fine-tuning (RFT) framework designed to enhance LLMs' reasoning capabilities for time series forecasting (TSF). The approach consists of: (1) a warmup supervised fine-tuning (SFT) stage using chain-of-thought (CoT) trajectories generated by DeepSeek-R1, and (2) a reinforcement learning stage employing GRIP, which uses non-uniform sampling and adaptive trajectory weighting with a multi-objective reward function. Experiments on nine datasets demonstrate improvements over baseline methods including traditional deep learning models and LLM-based approaches. The authors argue that slow-thinking reasoning provides better temporal understanding than fast-thinking pattern matching.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "(1)  The distinction between \"fast-thinking\" (pattern matching) and \"slow-thinking\" (explicit reasoning) paradigms is intuitive and well-articulated. The motivation that time series contain complex temporal logic warranting structured reasoning is compelling.\n(2) The combination of SFT (for format stabilization and basic reasoning) followed by RL (for generalization and reasoning refinement) is a principled approach that addresses practical training challenges. The ablation study (Table 3, Figure 4b) effectively demonstrates the necessity of both stages.", "weaknesses": "(1) Using LLMs to generate synthetic CoT trajectories for fine-tuning is not novel (e.g., similar to approaches in reasoning LLMs). The key distinction from existing work is unclear.\n(2) While the adaptive weighting is presented as novel, it is essentially a variant of importance sampling with softmax normalization. The core algorithmic novelty is limited. The comparison with GRPO (Figure 4a) shows only marginal improvements.\n(3) The paper claims \"slow thinking\" but this mirrors existing reasoning-based LLM approaches (o1, DeepSeek-R1). What is fundamentally new beyond applying them to TSF?\n(4) SFT data is generated using DeepSeek-R1, which itself is a slow-thinking model. This creates a conceptual circularity: improving reasoning for TSF via a model that already has slow-thinking capabilities.", "questions": "See Weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762073154734}, {"id": "JHgHzv2hUe", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12694/Reviewer_hii8"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper proposes Time-R1, a framework that recasts time series forecasting as a reasoning task for an LLM. Moving away from traditional \"fast-thinking\" models that directly map historical data to future values, Time-R1 adopts a \"slow-thinking\" paradigm. In this approach, the LLM first generates an explicit, step-by-step reasoning process about the time series' properties (trends, seasonality, etc.) before outputting the final numerical forecast.\n\nThe framework employs a two-stage training process. First, SFT is used to adapt the LLM to the task format and instill basic reasoning patterns, using synthetic reasoning trajectories generated by a more powerful \"teacher\" model. Second, RL is applied to refine and generalize this reasoning ability. A key part of the RL stage is a novel optimization algorithm called GRIP, which uses non-uniform sampling to focus on high-reward reasoning paths, guided by a fine-grained, multi-objective reward function. Experiments show that a model trained on a single dataset (ETTh1) can generalize effectively to eight other unseen datasets, outperforming traditional TSF models.", "review_text": "This paper proposes Time-R1, a framework that recasts time series forecasting as a reasoning task for an LLM. Moving away from traditional \"fast-thinking\" models that directly map historical data to future values, Time-R1 adopts a \"slow-thinking\" paradigm. In this approach, the LLM first generates an explicit, step-by-step reasoning process about the time series' properties (trends, seasonality, etc.) before outputting the final numerical forecast.\n\nThe framework employs a two-stage training process. First, SFT is used to adapt the LLM to the task format and instill basic reasoning patterns, using synthetic reasoning trajectories generated by a more powerful \"teacher\" model. Second, RL is applied to refine and generalize this reasoning ability. A key part of the RL stage is a novel optimization algorithm called GRIP, which uses non-uniform sampling to focus on high-reward reasoning paths, guided by a fine-grained, multi-objective reward function. Experiments show that a model trained on a single dataset (ETTh1) can generalize effectively to eight other unseen datasets, outperforming traditional TSF models.", "strengths": "1. Framing TSF as a \"slow-thinking\" reasoning task is a compelling and innovative idea that could open up new avenues for building more intelligent forecasting systems.\n2. The model's ability to train on one dataset and perform well on eight others is a significant strength, showcasing that it learns transferable reasoning skills rather than just dataset-specific patterns.\n3. The generation of an explicit reasoning chain before the forecast provides a degree of interpretability that is absent in most traditional \"black-box\" TSF models.\n4. The paper includes comprehensive ablations that effectively demonstrate the importance of each part of the complex pipeline, from SFT and RL to the individual reward components.", "weaknesses": "1. The entire framework is exceptionally complex, involving synthetic data generation with a powerful teacher model, SFT, and a sophisticated RL pipeline with a custom optimizer. The reported inference speed (~3 samples/s) and training requirements make it impractical for many real-world applications. The trade-off between accuracy and efficiency is severe. \n2. The SFT stage relies on reasoning paths generated by another LLM that is prompted with the ground-truth answer. This could lead to the model learning to generate \"post-hoc justifications\" rather than engaging in genuine, from-scratch reasoning. The quality of the final model is heavily dependent on the quality of the teacher model.\n3. The multi-objective reward function is a complex combination of several terms, and the paper provides little justification for the specific formulation and weighting of these components. This makes the design feel ad-hoc and potentially difficult to transfer to other tasks.\n4. While presented as a key contribution, the GRIP algorithm offers a modest improvement over existing methods like GRPO. The core ideas of focusing on high-reward samples are not entirely new, and the ablation in Figure 4a confirms the improvement is not dramatic.", "questions": "I don't think all TSF tasks require slow thinking. Is it possible to clearly define when we need slow thinking and when fast thinking is sufficient to obtain good results? Even for slow thinking, what specific real-world applications do you envision for this \"slow-thinking\" paradigm, where the benefits in accuracy and interpretability would justify the extreme costs in latency and compute?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Time-R1, a framework that recasts time series forecasting as a reasoning task for an LLM. Moving away from traditional \"fast-thinking\" models that directly map historical data to future values, Time-R1 adopts a \"slow-thinking\" paradigm. In this approach, the LLM first generates an explicit, step-by-step reasoning process about the time series' properties (trends, seasonality, etc.) before outputting the final numerical forecast.\n\nThe framework employs a two-stage training process. First, SFT is used to adapt the LLM to the task format and instill basic reasoning patterns, using synthetic reasoning trajectories generated by a more powerful \"teacher\" model. Second, RL is applied to refine and generalize this reasoning ability. A key part of the RL stage is a novel optimization algorithm called GRIP, which uses non-uniform sampling to focus on high-reward reasoning paths, guided by a fine-grained, multi-objective reward function. Experiments show that a model trained on a single dataset (ETTh1) can generalize effectively to eight other unseen datasets, outperforming traditional TSF models.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Framing TSF as a \"slow-thinking\" reasoning task is a compelling and innovative idea that could open up new avenues for building more intelligent forecasting systems.\n2. The model's ability to train on one dataset and perform well on eight others is a significant strength, showcasing that it learns transferable reasoning skills rather than just dataset-specific patterns.\n3. The generation of an explicit reasoning chain before the forecast provides a degree of interpretability that is absent in most traditional \"black-box\" TSF models.\n4. The paper includes comprehensive ablations that effectively demonstrate the importance of each part of the complex pipeline, from SFT and RL to the individual reward components.", "weaknesses": "1. The entire framework is exceptionally complex, involving synthetic data generation with a powerful teacher model, SFT, and a sophisticated RL pipeline with a custom optimizer. The reported inference speed (~3 samples/s) and training requirements make it impractical for many real-world applications. The trade-off between accuracy and efficiency is severe. \n2. The SFT stage relies on reasoning paths generated by another LLM that is prompted with the ground-truth answer. This could lead to the model learning to generate \"post-hoc justifications\" rather than engaging in genuine, from-scratch reasoning. The quality of the final model is heavily dependent on the quality of the teacher model.\n3. The multi-objective reward function is a complex combination of several terms, and the paper provides little justification for the specific formulation and weighting of these components. This makes the design feel ad-hoc and potentially difficult to transfer to other tasks.\n4. While presented as a key contribution, the GRIP algorithm offers a modest improvement over existing methods like GRPO. The core ideas of focusing on high-reward samples are not entirely new, and the ablation in Figure 4a confirms the improvement is not dramatic.", "questions": "I don't think all TSF tasks require slow thinking. Is it possible to clearly define when we need slow thinking and when fast thinking is sufficient to obtain good results? Even for slow thinking, what specific real-world applications do you envision for this \"slow-thinking\" paradigm, where the benefits in accuracy and interpretability would justify the extreme costs in latency and compute?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761987878426}, {"id": "Nmbs3tzeDI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12694/Reviewer_hSn1"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper propose Time-R1, a novel time series forecasting framework that trains large language model with slow-thinking reasoning capabilities via a two-stage reinforcement fine-tuning pipeline. First, supervised fine-tuning uses synthetic chain-of-thought trajectories to teach the model temporal analysis and output formatting. Second, reinforcement learning with a fine-grained multi-objective reward function enhances generalization. A key innovation is GRIP, a non-uniform sampling and adaptive weighting strategy for optimizing reasoning paths. Experiments on nine diverse datasets show Time-R1 outperforms traditional deep learning and LLM-based baselines in MSE and MAE, improving temporal coherence and out-of-distribution generalization. By integrating explicit reasoning into TSF, Time-R1 addresses the \"fast thinking\" limitation of existing methods, offering both interpretability and state-of-the-art performance.", "review_text": "This paper propose Time-R1, a novel time series forecasting framework that trains large language model with slow-thinking reasoning capabilities via a two-stage reinforcement fine-tuning pipeline. First, supervised fine-tuning uses synthetic chain-of-thought trajectories to teach the model temporal analysis and output formatting. Second, reinforcement learning with a fine-grained multi-objective reward function enhances generalization. A key innovation is GRIP, a non-uniform sampling and adaptive weighting strategy for optimizing reasoning paths. Experiments on nine diverse datasets show Time-R1 outperforms traditional deep learning and LLM-based baselines in MSE and MAE, improving temporal coherence and out-of-distribution generalization. By integrating explicit reasoning into TSF, Time-R1 addresses the \"fast thinking\" limitation of existing methods, offering both interpretability and state-of-the-art performance.", "strengths": "1. Introduces slow-thinking reasoning for TSF, replacing direct pattern mapping with explicit step-by-step temporal inference, enhancing interpretability and logical consistency.\n\n2. Designing Two-Stage RFT Framework, SFT warmup ensures proper formatting and basic reasoning, while RL with tailored rewards optimizes for TSF-specific goals.\n\n3. Non-uniform sampling and adaptive weighting balance exploration/exploitation, reducing computational cost while amplifying gradient signals from high-quality reasoning paths.\n\n4. Very strong generalization, which trained on one dataset (ETTh1) but achieves superior performance across nine diverse domains, outperforming domain-specific baselines without task-specific fine-tuning.", "weaknesses": "1. There is significant discrepancy between the MSE losses of each reported model and the error values in their respective original papers, For instance, the MSE of the Exchange dataset is extremely small, while the errors of ETTh1 and ETTm1 are notably large. Additionally, there is a lack of necessary setup details for comparing the baseline models, and different model configurations may lead to unfair comparisons. \n\n2. Using text modality for input and output of time series may not  inefficient. For example, a numerical value with 4 decimal places like 16.3864 may require 2–4 tokens to represent, especially for larger number of variates and  time series lengths. \n\n3. It is not very reasonable to achieve better performance on other time series datasets by only training on the ETTh1 dataset, as the inherent characteristics and distributions of different data may vary significantly such as stock or foreign exchange time series. Please explain the underlying principle with experiments or theoretical support. \n\n4. There are minor writing errors. For example, in Appendix A.3: \"offering computational efficiency but suffering from off-policy issues (?).\"", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose Time-R1, a novel time series forecasting framework that trains large language model with slow-thinking reasoning capabilities via a two-stage reinforcement fine-tuning pipeline. First, supervised fine-tuning uses synthetic chain-of-thought trajectories to teach the model temporal analysis and output formatting. Second, reinforcement learning with a fine-grained multi-objective reward function enhances generalization. A key innovation is GRIP, a non-uniform sampling and adaptive weighting strategy for optimizing reasoning paths. Experiments on nine diverse datasets show Time-R1 outperforms traditional deep learning and LLM-based baselines in MSE and MAE, improving temporal coherence and out-of-distribution generalization. By integrating explicit reasoning into TSF, Time-R1 addresses the \"fast thinking\" limitation of existing methods, offering both interpretability and state-of-the-art performance.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. Introduces slow-thinking reasoning for TSF, replacing direct pattern mapping with explicit step-by-step temporal inference, enhancing interpretability and logical consistency.\n\n2. Designing Two-Stage RFT Framework, SFT warmup ensures proper formatting and basic reasoning, while RL with tailored rewards optimizes for TSF-specific goals.\n\n3. Non-uniform sampling and adaptive weighting balance exploration/exploitation, reducing computational cost while amplifying gradient signals from high-quality reasoning paths.\n\n4. Very strong generalization, which trained on one dataset (ETTh1) but achieves superior performance across nine diverse domains, outperforming domain-specific baselines without task-specific fine-tuning.", "weaknesses": "1. There is significant discrepancy between the MSE losses of each reported model and the error values in their respective original papers, For instance, the MSE of the Exchange dataset is extremely small, while the errors of ETTh1 and ETTm1 are notably large. Additionally, there is a lack of necessary setup details for comparing the baseline models, and different model configurations may lead to unfair comparisons. \n\n2. Using text modality for input and output of time series may not  inefficient. For example, a numerical value with 4 decimal places like 16.3864 may require 2–4 tokens to represent, especially for larger number of variates and  time series lengths. \n\n3. It is not very reasonable to achieve better performance on other time series datasets by only training on the ETTh1 dataset, as the inherent characteristics and distributions of different data may vary significantly such as stock or foreign exchange time series. Please explain the underlying principle with experiments or theoretical support. \n\n4. There are minor writing errors. For example, in Appendix A.3: \"offering computational efficiency but suffering from off-policy issues (?).\"", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761639411011}, {"id": "KcvipNwXz7", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12694/Reviewer_nXNz"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces Time-R1, a two-stage reinforcement fine-tuning framework designed to enhance large language models (LLMs) for time series forecasting (TSF) by developing explicit multi-step reasoning capabilities.", "review_text": "The paper introduces Time-R1, a two-stage reinforcement fine-tuning framework designed to enhance large language models (LLMs) for time series forecasting (TSF) by developing explicit multi-step reasoning capabilities.", "strengths": "- Several LLM-based models are used as benchmarks. Competitive forecasting benchmarks like PatchTST are also included.\n- Presentation and results in Table 3 are clear and well-structured.\n- The purposed approach appears to outperform most baselines across datasets.\n- The figures are visually effective and contribute to understanding the proposed framework.\n- The paper’s proposed framework for fine-tuning dataset construction is a valuable direction.", "weaknesses": "Several claims in the paper motivating use of reasoning for time series models are either vague or overstated:\n   - For example, the statement “These models map history to future directly without detecting regime changes or performing step-by-step inference, resembling fast (not deliberate) thinking for time series” lacks precision. The notion of “regime changes” is not well defined. Modeling covariates could, for instance, already capture temporal dynamics associated with external events.\n   - Lines 62–63: The phrase “time series often reflect more complex temporal logic, which should not merely be ‘fitted’—they should be understood and reasoned” is conceptually unclear. What is meant by “complex temporal logic” (e.g., domain knowledge, text-based covariates)? The authors should better justify why “reasoning” about such logic is necessary or beneficial compared to existing data-driven approaches.\n   - Line 84: The assertion that the model is “fine-tuned for memorization” is a strong and potentially misleading claim. Forecasting fine-tuning does not inherently imply memorization, and the distinction between memorization and generalization is an active topic of research that requires supporting evidence.\n\nMethodology contribution concerns: \n- Reasoning evaluation (Eq. 1): The evaluation procedure for reasoning remains unclear. If the metric is based solely on achieving the lowest MSE, it does not actually assess whether the model engages in reasoning, it merely rewards numerical accuracy.\n\n- Reward design (Eqs. 3 and 4): The proposed reward functions appear conceptually similar to traditional loss formulations such as MSE and trend–seasonality decomposition. This raises questions about how they differ from conventional forecasting objectives. The authors critique prior approaches for lacking explicit reasoning, yet the reward relies on classical statistical constructs. The paper should clarify how this formulation advances the goal of reasoning-based forecasting. Moreover, this seems inconsistent with statements in the introduction (Lines 51–52): “Although effective in benchmarks, their underlying logic is largely based on pattern recognition (Cheng et al., 2025a) and trend prediction, lacking an explicit reasoning process.”\n\n- Model architecture clarity: The underlying architecture on which Time-R1 is built is not clearly described. Rather than presenting Time-R1 as a standalone model, the contributions of the reasoning-oriented fine-tuning and reinforcement learning framework, would be more compelling if shown to consistently improve performance across multiple LLM backbones.\n\nBenchmark concerns:\n- The paper lacks comparisons with non-LLM-based Time Series Foundation Models (e.g., TimesFM, Moirai, Moment) and classical statistical baselines (e.g., ARIMA, ETS). Including such comparisons is crucial to validate its claimed advantages and motivate the use of LLMs for forecasting.", "questions": "Other:\n- Lines 261-269: formulation of the GRIP Objective could be more clearly articulated to outline the provided equations.\n- The purposed multi-trajectory approach in lines 292–300 is interesting. It could be interesting to explore ensembling of top-k sample trajectories. Such approach could reduce forecast variance an important consideration for forecasting in practice.\n- In addition to the reward based on returned prediction sequence length, including a reward function based on prediction variance from distributional loss functions could help further improve model performance.\n- Normalization is very often used in time series forecasting models. It could be interesting to show an ablation study with and without normalization to see how robust this approach is. \n- Some acronym and functions are not defined:\n   - Function ‘g’ (Line 107) not defined.\n   - SFT (Lines 095, 151) not defined.\n   - Extrema (Eq. 5) and Eq. 6 unclear. Unclear how extrema are determined.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces Time-R1, a two-stage reinforcement fine-tuning framework designed to enhance large language models (LLMs) for time series forecasting (TSF) by developing explicit multi-step reasoning capabilities.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Several LLM-based models are used as benchmarks. Competitive forecasting benchmarks like PatchTST are also included.\n- Presentation and results in Table 3 are clear and well-structured.\n- The purposed approach appears to outperform most baselines across datasets.\n- The figures are visually effective and contribute to understanding the proposed framework.\n- The paper’s proposed framework for fine-tuning dataset construction is a valuable direction.", "weaknesses": "Several claims in the paper motivating use of reasoning for time series models are either vague or overstated:\n   - For example, the statement “These models map history to future directly without detecting regime changes or performing step-by-step inference, resembling fast (not deliberate) thinking for time series” lacks precision. The notion of “regime changes” is not well defined. Modeling covariates could, for instance, already capture temporal dynamics associated with external events.\n   - Lines 62–63: The phrase “time series often reflect more complex temporal logic, which should not merely be ‘fitted’—they should be understood and reasoned” is conceptually unclear. What is meant by “complex temporal logic” (e.g., domain knowledge, text-based covariates)? The authors should better justify why “reasoning” about such logic is necessary or beneficial compared to existing data-driven approaches.\n   - Line 84: The assertion that the model is “fine-tuned for memorization” is a strong and potentially misleading claim. Forecasting fine-tuning does not inherently imply memorization, and the distinction between memorization and generalization is an active topic of research that requires supporting evidence.\n\nMethodology contribution concerns: \n- Reasoning evaluation (Eq. 1): The evaluation procedure for reasoning remains unclear. If the metric is based solely on achieving the lowest MSE, it does not actually assess whether the model engages in reasoning, it merely rewards numerical accuracy.\n\n- Reward design (Eqs. 3 and 4): The proposed reward functions appear conceptually similar to traditional loss formulations such as MSE and trend–seasonality decomposition. This raises questions about how they differ from conventional forecasting objectives. The authors critique prior approaches for lacking explicit reasoning, yet the reward relies on classical statistical constructs. The paper should clarify how this formulation advances the goal of reasoning-based forecasting. Moreover, this seems inconsistent with statements in the introduction (Lines 51–52): “Although effective in benchmarks, their underlying logic is largely based on pattern recognition (Cheng et al., 2025a) and trend prediction, lacking an explicit reasoning process.”\n\n- Model architecture clarity: The underlying architecture on which Time-R1 is built is not clearly described. Rather than presenting Time-R1 as a standalone model, the contributions of the reasoning-oriented fine-tuning and reinforcement learning framework, would be more compelling if shown to consistently improve performance across multiple LLM backbones.\n\nBenchmark concerns:\n- The paper lacks comparisons with non-LLM-based Time Series Foundation Models (e.g., TimesFM, Moirai, Moment) and classical statistical baselines (e.g., ARIMA, ETS). Including such comparisons is crucial to validate its claimed advantages and motivate the use of LLMs for forecasting.", "questions": "Other:\n- Lines 261-269: formulation of the GRIP Objective could be more clearly articulated to outline the provided equations.\n- The purposed multi-trajectory approach in lines 292–300 is interesting. It could be interesting to explore ensembling of top-k sample trajectories. Such approach could reduce forecast variance an important consideration for forecasting in practice.\n- In addition to the reward based on returned prediction sequence length, including a reward function based on prediction variance from distributional loss functions could help further improve model performance.\n- Normalization is very often used in time series forecasting models. It could be interesting to show an ablation study with and without normalization to see how robust this approach is. \n- Some acronym and functions are not defined:\n   - Function ‘g’ (Line 107) not defined.\n   - SFT (Lines 095, 151) not defined.\n   - Extrema (Eq. 5) and Eq. 6 unclear. Unclear how extrema are determined.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760743868652}], "openreview_url": "https://openreview.net/forum?id=uwoA5iyTC6", "arxiv_id": "2506.10630", "paper_pdf": "papers/uwoA5iyTC6.pdf", "paper_pdf_sha256": "e71338135a187cf3ecf4d2440013466d2798916226d2f56fec55e3c53ad8c1ad", "paper_pdf_bytes": 6664272, "paper_pdf_source": "openreview", "code_url": "https://github.com/ustc-time-series/Time-R1", "code_repository": "ustc-time-series/Time-R1", "code_commit": "2cd666eaddac37734b45103d654b6d2d22bf23ed", "code_archive": "repos/uwoA5iyTC6.zip", "code_archive_sha256": "b16dc8a9161bbb16427a6e4d162446b69d6f827a345bba5fd99e1b53a634b135", "code_archive_bytes": 7496202, "code_file_count": 259, "code_extensions": {".py": 203, ".sh": 54, ".ipynb": 2}, "github_disk_usage_kb": 7771, "github_languages": {"Python": 1502870, "Shell": 23013}, "github_archived": false, "github_pushed_at": "2026-04-14T08:32:23Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/time-series-forecasting-as-reasoning-a-slow"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pVL4bYKOGM", "year": 2025, "status": "rejected", "title": "Conformal prediction for causal effects of continuous treatments", "authors": ["Maresa Schröder", "Dennis Frauen", "Jonas Schweisthal", "Konstantin Hess", "Valentyn Melnychuk", "Stefan Feuerriegel"], "authorids": ["~Maresa_Schröder1", "~Dennis_Frauen1", "~Jonas_Schweisthal1", "~Konstantin_Hess1", "~Valentyn_Melnychuk1", "~Stefan_Feuerriegel1"], "authors_source": "OpenReview API", "abstract": "Uncertainty quantification of causal effects is crucial for safety-critical applications such as personalized medicine. A powerful approach for this is conformal prediction, which has several practical benefits due to model-agnostic finite-sample guarantees. Yet, existing methods for conformal prediction of causal effects are limited to binary/discrete treatments and make highly restrictive assumptions such as known propensity scores. In this work, we provide a novel conformal prediction method for potential outcomes of continuous treatments. We account for the additional uncertainty introduced through propensity estimation so that our conformal prediction intervals are valid even if the propensity score is unknown. Our contributions are three-fold: (1) We derive finite-sample prediction intervals for potential outcomes of continuous treatments. (2) We provide an algorithm for calculating the derived intervals. (3) We demonstrate the effectiveness of the conformal prediction intervals in experiments on synthetic and medical datasets. To the best of our knowledge, we are the first to propose conformal prediction for continuous treatments when the propensity score is unknown and must be estimated from data.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "QgCvBhMdHS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4984/Reviewer_qNm2"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper proposes a conformal prediction method for continuous treatments, addressing the challenge of unknown propensity scores. It derives finite-sample prediction intervals, provides an algorithm for calculating them, and demonstrates effectiveness on synthetic and real datasets. This is the first method for continuous treatments with unknown propensity scores.", "review_text": "This paper proposes a conformal prediction method for continuous treatments, addressing the challenge of unknown propensity scores. It derives finite-sample prediction intervals, provides an algorithm for calculating them, and demonstrates effectiveness on synthetic and real datasets. This is the first method for continuous treatments with unknown propensity scores.", "strengths": "Clear writing & solid theoretical results", "weaknesses": "The reviewer is not an expert in conformal prediction for causal inference but would assume there are working targeting the studies setting: uncertainty quantification of causal effects for continuous treatment.\n\nEquation 14 seems to use the kernel function to approximate the indicator function. Does this imply some smoothness assumptions?\n\nThe optimal parameter $\\sigma^*$ represents a trade-off between the uncertainty in the prediction and the uncertainty in the interval construction. It is similar to the bandwidth approximation of the indicator function using a kernel function. Are there any empirical values ​​here, such as \\sigma = cn^{-2/5} to  trade-off between the prediction and interval.\n\nThe method is demonstrated on specific datasets and scenarios. While the authors claim that their method scales to high-dimensional settings, further validation on a broader range of datasets and treatment types would strengthen the generalizability of the findings.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a conformal prediction method for continuous treatments, addressing the challenge of unknown propensity scores. It derives finite-sample prediction intervals, provides an algorithm for calculating them, and demonstrates effectiveness on synthetic and real datasets. This is the first method for continuous treatments with unknown propensity scores.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Clear writing & solid theoretical results", "weaknesses": "The reviewer is not an expert in conformal prediction for causal inference but would assume there are working targeting the studies setting: uncertainty quantification of causal effects for continuous treatment.\n\nEquation 14 seems to use the kernel function to approximate the indicator function. Does this imply some smoothness assumptions?\n\nThe optimal parameter $\\sigma^*$ represents a trade-off between the uncertainty in the prediction and the uncertainty in the interval construction. It is similar to the bandwidth approximation of the indicator function using a kernel function. Are there any empirical values ​​here, such as \\sigma = cn^{-2/5} to  trade-off between the prediction and interval.\n\nThe method is demonstrated on specific datasets and scenarios. While the authors claim that their method scales to high-dimensional settings, further validation on a broader range of datasets and treatment types would strengthen the generalizability of the findings.", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730711828406}, {"id": "qUIQk25KQI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4984/Reviewer_xiCH"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "In the paper, the authors propose a Conformal-prediction (CP) based method to obtain the interval estimation of the potential outcomes under continuous treatments. \nTo achieve this, the authors follows (Gibbs et al., 2023; Romano et al., 2019) and reformulated split CP as an augmented quantile regression.\nThe authors derive finite-sample prediction intervals for cases where the propensity score function is known and unknown.\nMoreover, they provide finite-sample theoretical guarantees.", "review_text": "In the paper, the authors propose a Conformal-prediction (CP) based method to obtain the interval estimation of the potential outcomes under continuous treatments. \nTo achieve this, the authors follows (Gibbs et al., 2023; Romano et al., 2019) and reformulated split CP as an augmented quantile regression.\nThe authors derive finite-sample prediction intervals for cases where the propensity score function is known and unknown.\nMoreover, they provide finite-sample theoretical guarantees.", "strengths": "1. They study a very important question in the causal inference and provide novel methods. \n2. They provide the finite-sample theorem for their methods. \n3. They clearly and accessibly present the research question and methods, using figures and concrete examples.", "weaknesses": "The authors do not have a good evaluation of their method in terms of numerical studies. \n\nFirstly, the authors only focus on the empirical coverage, and claim that it should be the higher the better. \nIn my opinion, it seems not very reasonable. \nIf we simply choose the [-inf, inf] as the interval, it is always has the 100% empirical coverage. \nA more balanced evaluation should consider both the interval length and empirical coverage. Additionally, the empirical coverage should closely match the nominal coverage for optimal results.\n\n\nWhen checking the numerical results, I do believe the method in the implementation is not very stable under some cases. \nFor example, in Fig 4 ($\\alpha=0.2$, 5), with only 50 repetitions, I can see many outliers in the boxplot, indicating there may be some issues there. Similar issue can be found in Fig 5 ($\\alpha=0.1$, 1) and  ($\\alpha=0.2$, 5). \nAnyway, I know such unstableness may be inevitable for a complicated method, I think the authors should have some discussions on this issue.", "questions": "1. Major question is in **Weakness** part \n2. I think the real data in the paper can not justify the superiority of authors' method well as I can see at range [0.6, 1], it yields not very reasonable prediction (blood pressure can be less than 0). \nAlso, the interval estimate is too wide to be meaningful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In the paper, the authors propose a Conformal-prediction (CP) based method to obtain the interval estimation of the potential outcomes under continuous treatments. \nTo achieve this, the authors follows (Gibbs et al., 2023; Romano et al., 2019) and reformulated split CP as an augmented quantile regression.\nThe authors derive finite-sample prediction intervals for cases where the propensity score function is known and unknown.\nMoreover, they provide finite-sample theoretical guarantees.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. They study a very important question in the causal inference and provide novel methods. \n2. They provide the finite-sample theorem for their methods. \n3. They clearly and accessibly present the research question and methods, using figures and concrete examples.", "weaknesses": "The authors do not have a good evaluation of their method in terms of numerical studies. \n\nFirstly, the authors only focus on the empirical coverage, and claim that it should be the higher the better. \nIn my opinion, it seems not very reasonable. \nIf we simply choose the [-inf, inf] as the interval, it is always has the 100% empirical coverage. \nA more balanced evaluation should consider both the interval length and empirical coverage. Additionally, the empirical coverage should closely match the nominal coverage for optimal results.\n\n\nWhen checking the numerical results, I do believe the method in the implementation is not very stable under some cases. \nFor example, in Fig 4 ($\\alpha=0.2$, 5), with only 50 repetitions, I can see many outliers in the boxplot, indicating there may be some issues there. Similar issue can be found in Fig 5 ($\\alpha=0.1$, 1) and  ($\\alpha=0.2$, 5). \nAnyway, I know such unstableness may be inevitable for a complicated method, I think the authors should have some discussions on this issue.", "questions": "1. Major question is in **Weakness** part \n2. I think the real data in the paper can not justify the superiority of authors' method well as I can see at range [0.6, 1], it yields not very reasonable prediction (blood pressure can be less than 0). \nAlso, the interval estimate is too wide to be meaningful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730620078190}, {"id": "jDnymZVsey", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4984/Reviewer_wfcX"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "Conformal prediction for causal estimands is useful because it provides finite-sample coverage guarantees. Predicting potential outcomes is challenging due to the potentially unknown distribution shift between the training set and the test set, which is subject to intervention. This work extends conformal prediction under domain shift to address the challenges that are introduced when the treatment variable is continuous, and the causal estimand is a dose response.", "review_text": "Conformal prediction for causal estimands is useful because it provides finite-sample coverage guarantees. Predicting potential outcomes is challenging due to the potentially unknown distribution shift between the training set and the test set, which is subject to intervention. This work extends conformal prediction under domain shift to address the challenges that are introduced when the treatment variable is continuous, and the causal estimand is a dose response.", "strengths": "The problem of conformal prediction of dose responses is clearly significant. The authors proposed an original formulation that uses Gaussian smoothing over the treatments and finds an optimal bandwidth for each problem instance. It is also important to consider soft interventions (by the authors' definition) in addition to hard interventions, since it is common in medicine and other settings to study incremental changes.", "weaknesses": "* I do not understand some core elements of the motivation. What is the intention behind the separation of challenge (a) and challenge (b)? Estimating the propensity score is more difficult with continuous treatments, but it is still a challenge with discrete treatments. By my understanding, the main challenge for bringing conformal prediction to continuous treatments is that you rarely see data points with the same treatment value, so you have to do some sort of smoothing. The motivating challenges do not mention this (unless I am misreading one of them).\n * I also found the roles of soft and hard interventions confusing. Under \"Scenarios\" in Section 4, the authors explain that soft interventions are relevant when the propensity is known, while hard interventions are reserved for when the propensity is unknown. This storyline is convenient for the mathematical derivations. However, I am not convinced: even if the propensity function is unknown, it can still be of interest to estimate the effect of increasing the treatment by a small amount on top of the existing policy. These cases have been studied before:\n   * Many papers by Chernozhukov et al. that mention \"average causal derivatives\"\n   * Marmarelis et al. \"Policy Learning for Localized Interventions from Observational Data\" in AISTATS 2024.\n * When the propensity / policy is unknown, the validity of the conformal prediction relies on a constant $M$ that bounds the error of the estimated propensity. This resembles a sensitivity analysis. It appears difficult to know which $M$ to use.\n * The results from the medical dataset do not seem convincing. It is good to see that intervals are wider in treatment regions that are less common in the data. However, the fact that intervals are wider than MC-dropout is not necessarily a good thing.\n\nMinor notes.\n * \"Quantile regression might yield non-unique solutions, so we later restrict the analysis to solvers invariant to the data ordering\". Please expand on this sentence. How does ordering invariance help?\n * The algorithm text is small and quite dense.\n * An interpretation of the optimized $\\sigma$ is provided in the supplement, but further discussion would be helpful. Please clarify its role in the optimization problem.", "questions": "* Consider building semi-synthetic experiments for stronger results.\n * How can $M$ of Assumption 1 be calibrated in practice, for tightness?\n * Do you assume some kind of continuity in the dose response?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Conformal prediction for causal estimands is useful because it provides finite-sample coverage guarantees. Predicting potential outcomes is challenging due to the potentially unknown distribution shift between the training set and the test set, which is subject to intervention. This work extends conformal prediction under domain shift to address the challenges that are introduced when the treatment variable is continuous, and the causal estimand is a dose response.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "The problem of conformal prediction of dose responses is clearly significant. The authors proposed an original formulation that uses Gaussian smoothing over the treatments and finds an optimal bandwidth for each problem instance. It is also important to consider soft interventions (by the authors' definition) in addition to hard interventions, since it is common in medicine and other settings to study incremental changes.", "weaknesses": "* I do not understand some core elements of the motivation. What is the intention behind the separation of challenge (a) and challenge (b)? Estimating the propensity score is more difficult with continuous treatments, but it is still a challenge with discrete treatments. By my understanding, the main challenge for bringing conformal prediction to continuous treatments is that you rarely see data points with the same treatment value, so you have to do some sort of smoothing. The motivating challenges do not mention this (unless I am misreading one of them).\n * I also found the roles of soft and hard interventions confusing. Under \"Scenarios\" in Section 4, the authors explain that soft interventions are relevant when the propensity is known, while hard interventions are reserved for when the propensity is unknown. This storyline is convenient for the mathematical derivations. However, I am not convinced: even if the propensity function is unknown, it can still be of interest to estimate the effect of increasing the treatment by a small amount on top of the existing policy. These cases have been studied before:\n   * Many papers by Chernozhukov et al. that mention \"average causal derivatives\"\n   * Marmarelis et al. \"Policy Learning for Localized Interventions from Observational Data\" in AISTATS 2024.\n * When the propensity / policy is unknown, the validity of the conformal prediction relies on a constant $M$ that bounds the error of the estimated propensity. This resembles a sensitivity analysis. It appears difficult to know which $M$ to use.\n * The results from the medical dataset do not seem convincing. It is good to see that intervals are wider in treatment regions that are less common in the data. However, the fact that intervals are wider than MC-dropout is not necessarily a good thing.\n\nMinor notes.\n * \"Quantile regression might yield non-unique solutions, so we later restrict the analysis to solvers invariant to the data ordering\". Please expand on this sentence. How does ordering invariance help?\n * The algorithm text is small and quite dense.\n * An interpretation of the optimized $\\sigma$ is provided in the supplement, but further discussion would be helpful. Please clarify its role in the optimization problem.", "questions": "* Consider building semi-synthetic experiments for stronger results.\n * How can $M$ of Assumption 1 be calibrated in practice, for tightness?\n * Do you assume some kind of continuity in the dose response?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730327402977}, {"id": "XGVv0NRSzM", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission4984/Reviewer_nAyf"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper applies the framework introduced by Gibbs et al. (2023) to produce prediction intervals, with a finite-sample coverage guarantee, for potential outcomes of continuous treatments. Gibbs et al. (2023) can provide certain types of conditional coverage guarantees by reformulating the coverage as coverage over a class of covariate shifts. Gibbs et al. (2023) already showed that given an appropriate choice of $\\mathcal{F}$, that their prediction set achieves coverage under covariate shift, so long that the covariate shift can be represented as, $P_X$ titled by $f$, and $f \\in \\mathcal{F}$.\n\nThis work postulates the causal effect of continuous treatments as soft and hard interventions.  A soft intervention can be seen as a value shift in the generalized propensity score. A hard intervention can be seen as shifting the generalized propensity score to a Dirac delta function to the desired treatment value. In the paper, they define an appropriate function class for the different interventions and deal with their specific covariate shifts. For the hard intervention setting, they also ensure that the function class can deal with an estimated generalized propensity score with a bounded estimation error of the propensity function. \n\nThey provide some experiments on synthetic datasets (from their own) and medical ICU datasets, benchmarking their prediciton intervals against an MC dropout and Ensemble approach.\n\n[1] I. Gibbs, J. J. Cherian, and E. J. Candès, “Conformal Prediction With Conditional Guarantees,” Dec. 20, 2023, arXiv: arXiv:2305.12616. doi: 10.48550/arXiv.2305.12616.", "review_text": "The paper applies the framework introduced by Gibbs et al. (2023) to produce prediction intervals, with a finite-sample coverage guarantee, for potential outcomes of continuous treatments. Gibbs et al. (2023) can provide certain types of conditional coverage guarantees by reformulating the coverage as coverage over a class of covariate shifts. Gibbs et al. (2023) already showed that given an appropriate choice of $\\mathcal{F}$, that their prediction set achieves coverage under covariate shift, so long that the covariate shift can be represented as, $P_X$ titled by $f$, and $f \\in \\mathcal{F}$.\n\nThis work postulates the causal effect of continuous treatments as soft and hard interventions.  A soft intervention can be seen as a value shift in the generalized propensity score. A hard intervention can be seen as shifting the generalized propensity score to a Dirac delta function to the desired treatment value. In the paper, they define an appropriate function class for the different interventions and deal with their specific covariate shifts. For the hard intervention setting, they also ensure that the function class can deal with an estimated generalized propensity score with a bounded estimation error of the propensity function. \n\nThey provide some experiments on synthetic datasets (from their own) and medical ICU datasets, benchmarking their prediciton intervals against an MC dropout and Ensemble approach.\n\n[1] I. Gibbs, J. J. Cherian, and E. J. Candès, “Conformal Prediction With Conditional Guarantees,” Dec. 20, 2023, arXiv: arXiv:2305.12616. doi: 10.48550/arXiv.2305.12616.", "strengths": "- **Clear Motivation:** The paper's objective is well-placed within causal machine learning, particularly in high-stakes areas like medicine, where uncertainty quantification is critical for decision-making.\n- **Novel Contribution:** The authors propose a novel method for deriving conformal prediction intervals in the presence of continuous treatments and unknown propensity scores. This is a significant contribution, as most prior work focuses on binary or discrete treatments.", "weaknesses": "- **Terminology Issues:** Terms such as \"finite-sample predictive intervals\" should be revised to \"finite-sample validity guarantees\" to align with established terminology in the field. Additionally, (line 090) “strong finite-sample mathematical guarantees” should be “finite-sample validity guarantees”. Similarly, the frequent use of the term \"exact\" in relation to conditional coverage is misleading, as exact conditional coverage is known to be unattainable. I also assume you do not provide exact coverage in the context of exact and conditional coverage.\n- **Relation to Lei and Candes (2021):** In Figure 2, and the appendix, the paper claims that the work of Lei and Candes can not deal with an unknown propensity, this is not true it defines and proves a coverage gap based on the error in the estimation of the likelihood ratio’s (= propensity in the case of treatment effect) in weighted CP.\n- **Relation to Gibbs et al. (2023): T**he paper mentions the paper of Gibbs et al. (2023) in that it reformulates the split CP as an augmented quantile regression and the use of Lemma 1. But throughout the work, several notions from this work are taken, such as the tilting of the propensity function (lines 210-215), Theorem 1 (lines 262-276) and Theorem 2 (lines 352-371) which is heavily inspired by this work as it can even be seen in the borrowed notation of Section 4 of Gibbs et al. (2023), Algorithm 1 similarly with Algorithm 1 of Gibbs et al. (2023).\n- **Under-coverage in Empirical Results**: The experimental results show some under-coverage, particularly for specific significance levels (e.g., α = 0.1). This variability raises concerns about the method's robustness, and the number of runs (10) in some experiments is too low to draw reliable conclusions. More runs would improve the statistical robustness of the findings.\n- **Computational Complexity**: While the paper introduces an algorithm, the appendix discusses the method's computational efficiency. It points out that complexity can be high due to the optimization solver. This should be included in the limitations. Additionally, you should include your experiments' fitting and inference times in the paper, as this could indicate if computational complexity is an issue in practice.\n- **Too simple simulation experiments: T**he proposed synthetic experiments are way too simplistic, with only one discrete confounder, continuous treatment, and continuous outcome. This is not in line with other treatment effect literature; we cannot adequately evaluate the method with these experiments.\n- **Baseline**: The choice of baseline methods, particularly MC dropout, may not be ideal for comparison. Since these approaches only quantify epistemic uncertainty, which is probably minimal due to the simple experiment settings, consider a benchmark approach that quantifies both aleatoric and epistemic uncertainty, or at least aleatoric such as the Gaussian process or maximum likelihood estimator. The MLE can be easily combined with MC dropout to get both aleatoric and epistemic to get a reasonable benchmark. Also, you did not evaluate vanilla CP without considering covariate shift as a benchmark.", "questions": "- What is your added contribution to the work of Gibbs et al. (2023), which you leverage in this work?\n- Can you explain the variability in empirical coverage across different significance levels, particularly for α = 0.1? Do you plan to increase the number of runs or experiments to solidify your claims?\n- Could you indicate how long it takes to run these experiments?\n- I think there is a typo on line 476 that should be Table 1, right?\n- In your synthetic experiments, why are the hard interventions related to x and not just hard interventions?\n- Related to Assumption 1 (line 295-299), does this assumption mean that the estimation error needs to be the same for each X given an intervention $a$, otherwise should it not be $c_a(x)$ instead of $c_a$.\n- Related to Theorem 2, what is the meaning of resulting $\\sigma$ and $c_a$ of the optimization problem.\n- Related to Lemma 4 (lines 712-736):\n    - I think you need to clarify that this is only relevant for absolute residual non-conformity score.\n    - Next, I think the proof is complete. You do not prove the connection between the Bonferonni corrected significance level and the targeted significance level.\n    - In my opinion, the proof for this lemma is even elementary since it is just an application of the Bonferonni correction.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper applies the framework introduced by Gibbs et al. (2023) to produce prediction intervals, with a finite-sample coverage guarantee, for potential outcomes of continuous treatments. Gibbs et al. (2023) can provide certain types of conditional coverage guarantees by reformulating the coverage as coverage over a class of covariate shifts. Gibbs et al. (2023) already showed that given an appropriate choice of $\\mathcal{F}$, that their prediction set achieves coverage under covariate shift, so long that the covariate shift can be represented as, $P_X$ titled by $f$, and $f \\in \\mathcal{F}$.\n\nThis work postulates the causal effect of continuous treatments as soft and hard interventions.  A soft intervention can be seen as a value shift in the generalized propensity score. A hard intervention can be seen as shifting the generalized propensity score to a Dirac delta function to the desired treatment value. In the paper, they define an appropriate function class for the different interventions and deal with their specific covariate shifts. For the hard intervention setting, they also ensure that the function class can deal with an estimated generalized propensity score with a bounded estimation error of the propensity function. \n\nThey provide some experiments on synthetic datasets (from their own) and medical ICU datasets, benchmarking their prediciton intervals against an MC dropout and Ensemble approach.\n\n[1] I. Gibbs, J. J. Cherian, and E. J. Candès, “Conformal Prediction With Conditional Guarantees,” Dec. 20, 2023, arXiv: arXiv:2305.12616. doi: 10.48550/arXiv.2305.12616.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- **Clear Motivation:** The paper's objective is well-placed within causal machine learning, particularly in high-stakes areas like medicine, where uncertainty quantification is critical for decision-making.\n- **Novel Contribution:** The authors propose a novel method for deriving conformal prediction intervals in the presence of continuous treatments and unknown propensity scores. This is a significant contribution, as most prior work focuses on binary or discrete treatments.", "weaknesses": "- **Terminology Issues:** Terms such as \"finite-sample predictive intervals\" should be revised to \"finite-sample validity guarantees\" to align with established terminology in the field. Additionally, (line 090) “strong finite-sample mathematical guarantees” should be “finite-sample validity guarantees”. Similarly, the frequent use of the term \"exact\" in relation to conditional coverage is misleading, as exact conditional coverage is known to be unattainable. I also assume you do not provide exact coverage in the context of exact and conditional coverage.\n- **Relation to Lei and Candes (2021):** In Figure 2, and the appendix, the paper claims that the work of Lei and Candes can not deal with an unknown propensity, this is not true it defines and proves a coverage gap based on the error in the estimation of the likelihood ratio’s (= propensity in the case of treatment effect) in weighted CP.\n- **Relation to Gibbs et al. (2023): T**he paper mentions the paper of Gibbs et al. (2023) in that it reformulates the split CP as an augmented quantile regression and the use of Lemma 1. But throughout the work, several notions from this work are taken, such as the tilting of the propensity function (lines 210-215), Theorem 1 (lines 262-276) and Theorem 2 (lines 352-371) which is heavily inspired by this work as it can even be seen in the borrowed notation of Section 4 of Gibbs et al. (2023), Algorithm 1 similarly with Algorithm 1 of Gibbs et al. (2023).\n- **Under-coverage in Empirical Results**: The experimental results show some under-coverage, particularly for specific significance levels (e.g., α = 0.1). This variability raises concerns about the method's robustness, and the number of runs (10) in some experiments is too low to draw reliable conclusions. More runs would improve the statistical robustness of the findings.\n- **Computational Complexity**: While the paper introduces an algorithm, the appendix discusses the method's computational efficiency. It points out that complexity can be high due to the optimization solver. This should be included in the limitations. Additionally, you should include your experiments' fitting and inference times in the paper, as this could indicate if computational complexity is an issue in practice.\n- **Too simple simulation experiments: T**he proposed synthetic experiments are way too simplistic, with only one discrete confounder, continuous treatment, and continuous outcome. This is not in line with other treatment effect literature; we cannot adequately evaluate the method with these experiments.\n- **Baseline**: The choice of baseline methods, particularly MC dropout, may not be ideal for comparison. Since these approaches only quantify epistemic uncertainty, which is probably minimal due to the simple experiment settings, consider a benchmark approach that quantifies both aleatoric and epistemic uncertainty, or at least aleatoric such as the Gaussian process or maximum likelihood estimator. The MLE can be easily combined with MC dropout to get both aleatoric and epistemic to get a reasonable benchmark. Also, you did not evaluate vanilla CP without considering covariate shift as a benchmark.", "questions": "- What is your added contribution to the work of Gibbs et al. (2023), which you leverage in this work?\n- Can you explain the variability in empirical coverage across different significance levels, particularly for α = 0.1? Do you plan to increase the number of runs or experiments to solidify your claims?\n- Could you indicate how long it takes to run these experiments?\n- I think there is a typo on line 476 that should be Table 1, right?\n- In your synthetic experiments, why are the hard interventions related to x and not just hard interventions?\n- Related to Assumption 1 (line 295-299), does this assumption mean that the estimation error needs to be the same for each X given an intervention $a$, otherwise should it not be $c_a(x)$ instead of $c_a$.\n- Related to Theorem 2, what is the meaning of resulting $\\sigma$ and $c_a$ of the optimization problem.\n- Related to Lemma 4 (lines 712-736):\n    - I think you need to clarify that this is only relevant for absolute residual non-conformity score.\n    - Next, I think the proof is complete. You do not prove the connection between the Bonferonni corrected significance level and the targeted significance level.\n    - In my opinion, the proof for this lemma is even elementary since it is just an application of the Bonferonni correction.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729690406999}], "openreview_url": "https://openreview.net/forum?id=pVL4bYKOGM", "arxiv_id": "2407.03094", "paper_pdf": "papers/pVL4bYKOGM.pdf", "paper_pdf_sha256": "6e0fe044a34f2471fc6cc544cf6f01e0748186b0e043027539398867ee64b184", "paper_pdf_bytes": 1848233, "paper_pdf_source": "openreview", "code_url": "https://github.com/m-schroder/ContinuousCausalCP", "code_repository": "m-schroder/ContinuousCausalCP", "code_commit": "d557555d6bb38239809b2c4d846c2768e9a279b7", "code_archive": "repos/pVL4bYKOGM.zip", "code_archive_sha256": "8e999ad3baa99b7926d1c2268094e4f43dd6cbd3c351846b9fd04d06103de02a", "code_archive_bytes": 96126, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 374, "github_languages": {"Python": 58734}, "github_archived": false, "github_pushed_at": "2024-07-04T11:28:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/conformal-prediction-for-causal-effects-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "npf3gREtf7", "year": 2024, "status": "rejected", "title": "Which Examples to Annotate for In-Context Learning? Towards Effective and Efficient Selection", "authors": ["Costas Mavromatis", "Balasubramaniam Srinivasan", "Zhengyuan Shen", "Jiani Zhang", "Huzefa Rangwala", "Christos Faloutsos", "George Karypis"], "authorids": ["~Costas_Mavromatis1", "~Balasubramaniam_Srinivasan1", "~Zhengyuan_Shen1", "~Jiani_Zhang2", "~Huzefa_Rangwala2", "~Christos_Faloutsos1", "~George_Karypis1"], "authors_source": "OpenReview API", "abstract": "Large Language Models (LLMs) can adapt to new tasks via in-context learning (ICL). ICL is efficient as it does not require any parameter updates to the trained LLM, but only few annotated examples as input for the LLM. In this work, we investigate an active learning approach for ICL, where there is a limited budget for annotating examples. We propose a model-adaptive optimization-free algorithm, termed AdaICL, which identifies examples that the model is uncertain about, and performs semantic diversity-based example selection. Diversity-based sampling improves overall effectiveness, while uncertainty sampling improves budget efficiency and helps the LLM learn new information. Moreover, AdaICL poses its sampling strategy as a Maximum Coverage problem, that dynamically adapts based on the model’s feedback and can be approximately solved via greedy algorithms. Extensive experiments on nine datasets and seven LLMs show that AdaICL improves performance by 4.4% accuracy points over SOTA (7.7% relative improvement), is up to 3× more budget-efficient than performing annotations uniformly at random, while it outperforms SOTA with 2× fewer ICL examples.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zvwf1hEq2p", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8269/Reviewer_hvtd"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes an active learning approach for ICL, which combines diversity-based sampling and uncertainty-based sampling. It introduces three versions of the proposed framework, including ADAICL-BASE, ADAICL, and ADAICL+. The base version performs k-means clustering over the identified hard examples, while ADAICL quantifies whether each example can help the model learn new information, and the plus version further equips a reweighting schema for the MAXCOVER problem to ensure dense regions with hard examples are preferred. Experiments study nine NLP datasets and GSM8K, with 1.3B to 65B LLMs across several model families.", "review_text": "This paper proposes an active learning approach for ICL, which combines diversity-based sampling and uncertainty-based sampling. It introduces three versions of the proposed framework, including ADAICL-BASE, ADAICL, and ADAICL+. The base version performs k-means clustering over the identified hard examples, while ADAICL quantifies whether each example can help the model learn new information, and the plus version further equips a reweighting schema for the MAXCOVER problem to ensure dense regions with hard examples are preferred. Experiments study nine NLP datasets and GSM8K, with 1.3B to 65B LLMs across several model families.", "strengths": "1. The paper is presented in a coherent manner and easy to follow.\n2. The logical progression connecting the various methodological variants is well-articulated. \n3. The improvements over the baseline is good.", "weaknesses": "1. Active learning for NLP is a well-studied area. This paper lacks the illustration of why AL for ICL is challenging or the key difference compared to AL for fine-tuning based NLP. Otherwise, why do not directly apply multiple sophisticated query policies proposed in AL to the ICL example selection problem?\n\n2. Although not for ICL, combining diversity and uncertainty for data selection have been studied in previous literature:\\\nEntropy-Based Active Learning for Object Detection With Progressive Diversity Constraint;\\\nCold-start data selection for few-shot language model fine-tuning: A prompt-based uncertainty propagation approach;\\\nACTUNE: Uncertainty-Based Active Self-Training for Active Fine-Tuning of Pretrained Language Models.\n\n3. The baseline methods are limited. Although the related work discusses a batch of work for active learning in ICL/NLP, only a few simple baselines are compared in experiments. How does ADAICL compare to other AL methods? Meanwhile, I am also wondering about the performance of zero-shot GPT-4 or GPT-3.5-turbo.\n\n4. The method relies on the estimation of model uncertainty, which is only suited for the LLMs with moderate scales, For those most recent LLMs with hundreds billions parameters, usually we do not have a way to obtain its uncertainty.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an active learning approach for ICL, which combines diversity-based sampling and uncertainty-based sampling. It introduces three versions of the proposed framework, including ADAICL-BASE, ADAICL, and ADAICL+. The base version performs k-means clustering over the identified hard examples, while ADAICL quantifies whether each example can help the model learn new information, and the plus version further equips a reweighting schema for the MAXCOVER problem to ensure dense regions with hard examples are preferred. Experiments study nine NLP datasets and GSM8K, with 1.3B to 65B LLMs across several model families.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "1. The paper is presented in a coherent manner and easy to follow.\n2. The logical progression connecting the various methodological variants is well-articulated. \n3. The improvements over the baseline is good.", "weaknesses": "1. Active learning for NLP is a well-studied area. This paper lacks the illustration of why AL for ICL is challenging or the key difference compared to AL for fine-tuning based NLP. Otherwise, why do not directly apply multiple sophisticated query policies proposed in AL to the ICL example selection problem?\n\n2. Although not for ICL, combining diversity and uncertainty for data selection have been studied in previous literature:\\\nEntropy-Based Active Learning for Object Detection With Progressive Diversity Constraint;\\\nCold-start data selection for few-shot language model fine-tuning: A prompt-based uncertainty propagation approach;\\\nACTUNE: Uncertainty-Based Active Self-Training for Active Fine-Tuning of Pretrained Language Models.\n\n3. The baseline methods are limited. Although the related work discusses a batch of work for active learning in ICL/NLP, only a few simple baselines are compared in experiments. How does ADAICL compare to other AL methods? Meanwhile, I am also wondering about the performance of zero-shot GPT-4 or GPT-3.5-turbo.\n\n4. The method relies on the estimation of model uncertainty, which is only suited for the LLMs with moderate scales, For those most recent LLMs with hundreds billions parameters, usually we do not have a way to obtain its uncertainty.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699164287686}, {"id": "OGz8eKE16C", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8269/Reviewer_m5zU"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose an example selection method for in-context learning via active learning techniques. Given an unlabeled set $\\mathcal{U}$ of examples, the authors first choose the set of hard examples $\\mathcal{U}_h$ based on model's confidence score $u_i$. Then choose examples up to a given budget $B$ from centroids of each k-mean cluster, which is termed as $\\texttt{AdaIcl-base}$. To improve the method, the Maximum Coverage problem is applied. The goal is then to choose $B$ most representative example that cover the most semantic space by building a global graph based on the semantic embedding space ($\\texttt{AdaIcl}$). To further improve the method, hard examples are selected from denser regions (instead of outliers) by implementing the re-weighting schema ($\\texttt{AdaIcl+}$). The results show effective improvement over nine datasets and seven LLM models.", "review_text": "The authors propose an example selection method for in-context learning via active learning techniques. Given an unlabeled set $\\mathcal{U}$ of examples, the authors first choose the set of hard examples $\\mathcal{U}_h$ based on model's confidence score $u_i$. Then choose examples up to a given budget $B$ from centroids of each k-mean cluster, which is termed as $\\texttt{AdaIcl-base}$. To improve the method, the Maximum Coverage problem is applied. The goal is then to choose $B$ most representative example that cover the most semantic space by building a global graph based on the semantic embedding space ($\\texttt{AdaIcl}$). To further improve the method, hard examples are selected from denser regions (instead of outliers) by implementing the re-weighting schema ($\\texttt{AdaIcl+}$). The results show effective improvement over nine datasets and seven LLM models.", "strengths": "+ The authors provide an intuitive approach which does make sense with additional efficiency.\n+ The paper is well structured.\n+ Results over many models and datasets showing great performance of variants of $\\texttt{AdaICL}$.", "weaknesses": "- The construction of the graph is highly dependent on other off-the-shelf encoders, which might not be representative of the target model.\n- The method requires multiple prompts to get LLM feedbacks (probability scores) for each example, which is expensive.\n- Quite outdated models used (even given the ICLR submission date), so it is hard to verify if the method is applicable to more up-to-date LLMs.", "questions": "- How are you performing k-mean clustering for $\\texttt{AdaIcl-base}$? \n\n- What embedder are you using for choosing top-k examples?\n\n- Have you compared to any similarity based baselines?\n\n- Have you compared to any retriever-based baselines?\n\n- Question about the practical setting. If we need to annotate the selected examples on demand based on the query, then why not annotate directly the query? \n\n- Why are you choosing top-$N_{\\theta}$ examples based on probability scores? Would uncertain examples mean bottom-$N_{\\theta}$? \n\n- RQ4. is provided but not addressed in the main paper nor referred to Appendix?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose an example selection method for in-context learning via active learning techniques. Given an unlabeled set $\\mathcal{U}$ of examples, the authors first choose the set of hard examples $\\mathcal{U}_h$ based on model's confidence score $u_i$. Then choose examples up to a given budget $B$ from centroids of each k-mean cluster, which is termed as $\\texttt{AdaIcl-base}$. To improve the method, the Maximum Coverage problem is applied. The goal is then to choose $B$ most representative example that cover the most semantic space by building a global graph based on the semantic embedding space ($\\texttt{AdaIcl}$). To further improve the method, hard examples are selected from denser regions (instead of outliers) by implementing the re-weighting schema ($\\texttt{AdaIcl+}$). The results show effective improvement over nine datasets and seven LLM models.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "+ The authors provide an intuitive approach which does make sense with additional efficiency.\n+ The paper is well structured.\n+ Results over many models and datasets showing great performance of variants of $\\texttt{AdaICL}$.", "weaknesses": "- The construction of the graph is highly dependent on other off-the-shelf encoders, which might not be representative of the target model.\n- The method requires multiple prompts to get LLM feedbacks (probability scores) for each example, which is expensive.\n- Quite outdated models used (even given the ICLR submission date), so it is hard to verify if the method is applicable to more up-to-date LLMs.", "questions": "- How are you performing k-mean clustering for $\\texttt{AdaIcl-base}$? \n\n- What embedder are you using for choosing top-k examples?\n\n- Have you compared to any similarity based baselines?\n\n- Have you compared to any retriever-based baselines?\n\n- Question about the practical setting. If we need to annotate the selected examples on demand based on the query, then why not annotate directly the query? \n\n- Why are you choosing top-$N_{\\theta}$ examples based on probability scores? Would uncertain examples mean bottom-$N_{\\theta}$? \n\n- RQ4. is provided but not addressed in the main paper nor referred to Appendix?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698801967405}, {"id": "FIfVd2Rw0K", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8269/Reviewer_6M3N"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a method to select samples to present for ICL based on set coverage in an embedding space. The proposed method, AdaICL, uses a greedy approximation for the MaxCover problem to select sets that cover as many hard problems as possible. AdaICL outperforms baselines on a wide variety of tasks and does not appear to be too expensive to run.\n\nI think this paper is closer to a 7 than 6 but unfortunately 7 is not an option on the rating scale.", "review_text": "This paper presents a method to select samples to present for ICL based on set coverage in an embedding space. The proposed method, AdaICL, uses a greedy approximation for the MaxCover problem to select sets that cover as many hard problems as possible. AdaICL outperforms baselines on a wide variety of tasks and does not appear to be too expensive to run.\n\nI think this paper is closer to a 7 than 6 but unfortunately 7 is not an option on the rating scale.", "strengths": "- Good empirical performance on a wide variety of LLMs and datasets.\n- AdaICL appears to outperform kMeans based retrievers such as Votek and AdaICL base\n- AdaICL appears to be more robust to sample presentation order than the baselines on some tasks", "weaknesses": "- Why is the semantic similiarity space determined by a 3rd party embedding model such as SBERT? Shouldn't this be from the LLM itself, such as from an embedding layer?\n- How does overall performance depend on $m$ in $G_m$?\n- My understanding is that AdaICL queries a LLM many times to get confidence scores *before* feeding the final prompt in to get $y_{test}$. How do the various baselines and AdaICL compare when limited to the same LLM query budget? For example, random selection requires 0 queries, which could be far cheaper than running AdaICL.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a method to select samples to present for ICL based on set coverage in an embedding space. The proposed method, AdaICL, uses a greedy approximation for the MaxCover problem to select sets that cover as many hard problems as possible. AdaICL outperforms baselines on a wide variety of tasks and does not appear to be too expensive to run.\n\nI think this paper is closer to a 7 than 6 but unfortunately 7 is not an option on the rating scale.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- Good empirical performance on a wide variety of LLMs and datasets.\n- AdaICL appears to outperform kMeans based retrievers such as Votek and AdaICL base\n- AdaICL appears to be more robust to sample presentation order than the baselines on some tasks", "weaknesses": "- Why is the semantic similiarity space determined by a 3rd party embedding model such as SBERT? Shouldn't this be from the LLM itself, such as from an embedding layer?\n- How does overall performance depend on $m$ in $G_m$?\n- My understanding is that AdaICL queries a LLM many times to get confidence scores *before* feeding the final prompt in to get $y_{test}$. How do the various baselines and AdaICL compare when limited to the same LLM query budget? For example, random selection requires 0 queries, which could be far cheaper than running AdaICL.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698009053317}, {"id": "q41Ksq1xSU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission8269/Reviewer_LBWT"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper explores in-context learning (ICL) for Large Language Models (LLMs) where the goal is to adapt the model to new tasks with minimal annotation. It introduces ADAICL, an active learning approach that efficiently selects examples for annotation within a limited budget. ADAICL identifies uncertain examples for the model and uses semantic diversity-based selection. This approach, treated as a maximum coverage problem, dynamically adapts based on the model's feedback. Experiments across datasets and LLMs demonstrate ADAICL's superiority, improving accuracy over state-of-the-art works and being up to 3 times more budget-efficient than random annotation. It also outperformed existing methods with half the number of annotated examples.", "review_text": "This paper explores in-context learning (ICL) for Large Language Models (LLMs) where the goal is to adapt the model to new tasks with minimal annotation. It introduces ADAICL, an active learning approach that efficiently selects examples for annotation within a limited budget. ADAICL identifies uncertain examples for the model and uses semantic diversity-based selection. This approach, treated as a maximum coverage problem, dynamically adapts based on the model's feedback. Experiments across datasets and LLMs demonstrate ADAICL's superiority, improving accuracy over state-of-the-art works and being up to 3 times more budget-efficient than random annotation. It also outperformed existing methods with half the number of annotated examples.", "strengths": "- The motivation is clear. It is significant to study how to determine which unlabeled instances should be labeled. This is important to reduce the cost of annotations in in-context learning. \n- Experimental results are overall great. On a series of tasks, the proposed method can achieve the best performance.", "weaknesses": "- The contributions are somewhat overclaimed.\n- Technical contributions are not sufficient. \n- The writing also should be polished. For the current form, there are a series of unclear justifications.\n\nMore details about the above weaknesses can be checked below.", "questions": "- It is possible to meet outliers if the method overemphasizes the selection of diverse data. However, in the current form, it is not clear how to address the issue. \n- Does the $k$-NN retriever equal the similar retriever in the Vote-$k$ paper?\n- This paper claims that \"However, these approaches do not consider which examples help the LLM learn new information and may waste resources for annotating examples whose answers are already within the model’s knowledge.\" I am somewhat confused about this claim. The method Vote-$k$ also uses the feedback of LLMs. Could the paper give more details about this?\n- Does the uncertain with respect to one example equal that the LLM can learn it accurately?\n- The paper argues that previous work assumes a high-resource setting, where a large set of ICL examples is already annotated. Could the paper provide some detailed examples for better understanding?\n- For Section 4.1, could the paper provide more details about how to obtain the probability with respect to the label or demonstrations?\n- Compared with previous work such as Vote-$k$, the method proposed by this work is more complex. It introduces a series of hyper-parameters. How to balance them in practice? Also, is there a time advantage of the proposed method over baselines?\n- What is the definition of \"egonet\" in this paper?\n- For Figure 5, could the paper supplement the comparison between all methods not just the best baseline?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores in-context learning (ICL) for Large Language Models (LLMs) where the goal is to adapt the model to new tasks with minimal annotation. It introduces ADAICL, an active learning approach that efficiently selects examples for annotation within a limited budget. ADAICL identifies uncertain examples for the model and uses semantic diversity-based selection. This approach, treated as a maximum coverage problem, dynamically adapts based on the model's feedback. Experiments across datasets and LLMs demonstrate ADAICL's superiority, improving accuracy over state-of-the-art works and being up to 3 times more budget-efficient than random annotation. It also outperformed existing methods with half the number of annotated examples.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "- The motivation is clear. It is significant to study how to determine which unlabeled instances should be labeled. This is important to reduce the cost of annotations in in-context learning. \n- Experimental results are overall great. On a series of tasks, the proposed method can achieve the best performance.", "weaknesses": "- The contributions are somewhat overclaimed.\n- Technical contributions are not sufficient. \n- The writing also should be polished. For the current form, there are a series of unclear justifications.\n\nMore details about the above weaknesses can be checked below.", "questions": "- It is possible to meet outliers if the method overemphasizes the selection of diverse data. However, in the current form, it is not clear how to address the issue. \n- Does the $k$-NN retriever equal the similar retriever in the Vote-$k$ paper?\n- This paper claims that \"However, these approaches do not consider which examples help the LLM learn new information and may waste resources for annotating examples whose answers are already within the model’s knowledge.\" I am somewhat confused about this claim. The method Vote-$k$ also uses the feedback of LLMs. Could the paper give more details about this?\n- Does the uncertain with respect to one example equal that the LLM can learn it accurately?\n- The paper argues that previous work assumes a high-resource setting, where a large set of ICL examples is already annotated. Could the paper provide some detailed examples for better understanding?\n- For Section 4.1, could the paper provide more details about how to obtain the probability with respect to the label or demonstrations?\n- Compared with previous work such as Vote-$k$, the method proposed by this work is more complex. It introduces a series of hyper-parameters. How to balance them in practice? Also, is there a time advantage of the proposed method over baselines?\n- What is the definition of \"egonet\" in this paper?\n- For Figure 5, could the paper supplement the comparison between all methods not just the best baseline?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697340465195}], "openreview_url": "https://openreview.net/forum?id=npf3gREtf7", "arxiv_id": "2310.20046", "paper_pdf": "papers/npf3gREtf7.pdf", "paper_pdf_sha256": "3a2c953209a509c4d3882001a9bf84bb744ad7e4246ababaa33a0eeda22bdeb6", "paper_pdf_bytes": 1364854, "paper_pdf_source": "openreview", "code_url": "https://github.com/amazon-science/adaptive-in-context-learning", "code_repository": "amazon-science/adaptive-in-context-learning", "code_commit": "d0ea1c7d333517de592b4c1bf3862ab2ddfa027a", "code_archive": "repos/npf3gREtf7.zip", "code_archive_sha256": "2dbbfd33c298268ae85ac003d2ebd7988b899983ea75beb21986e42e30d990ed", "code_archive_bytes": 374045, "code_file_count": 13, "code_extensions": {".py": 12, ".sh": 1}, "github_disk_usage_kb": 354, "github_languages": {"Python": 234643, "Shell": 690}, "github_archived": true, "github_pushed_at": "2023-10-30T20:14:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/which-examples-to-annotate-for-in-context"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MtGmCCPJD-", "year": 2023, "status": "rejected", "title": "Repository-Level Prompt Generation for Large Language Models of Code", "authors": ["Disha Shrivastava", "Hugo Larochelle", "Daniel Tarlow"], "authorids": ["~Disha_Shrivastava1", "~Hugo_Larochelle1", "~Daniel_Tarlow1"], "authors_source": "OpenReview API", "abstract": "With the success of large language models (LLMs) of code and their use as code assistants (e.g.\\ Codex used in GitHub Copilot, techniques for introducing domain-specific knowledge in the prompt design process become important. In this work, we propose a framework called Repo-Level Prompt Generator that learns to generate example-specific prompts using prompt proposals. The prompt proposals take context from the entire repository, thereby incorporating both the structure of the repository and the context from other relevant files (e.g.\\ imports, parent class files). Our technique doesn't require any access to the weights of the LLM, making it applicable in cases where we only have black-box access to the LLM. We conduct experiments on the task of single-line code-autocompletion using code repositories taken from Google Code archives. We demonstrate that an oracle constructed from our prompt proposals gives a remarkably high relative improvement of 36\\% over Codex, showing the quality of these proposals. Further, we show that when we train a model to select the best prompt proposal, we can achieve significant performance gains over Codex and other baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "bmtUB-Rhu_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4604/Reviewer_oQyE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a method to make prompt proposals from repository-level context and add concatenate the prompt proposals with normal context to LLM in order to feed more related information to achieve better generations. The paper explores different sources of repository level prompts, as well as prompt context types. Experiment results show that additional repository level prompt help improve the performance significantly.\n", "review_text": "Overall, I think this is a decent paper with novel method proposed and well-performed experiments. I have some concerns on the paper, and would want to see more experiments and results to make the paper more complete.", "strengths": "Strength:\n- The paper is well-written. The method is well explained with examples. Experiment results are extensive with analysis.\n- The paper explores several variants on the prompt sources and prompt types, and made a nice figure to demonstrate which prompt source works the best, though only one is used in experiment at a time.\n\n\nWeakness:\n- Taking the whole file from the whole import file seems unnecessarily large context. Rather, if the import statement only imports one class or one function from the module, it makes sense to only consider the function implementation, rather than the whole file where the function resides. This especially applies to 7) Method Names and Bodies prompt type.\n- The paper only considers one prompt proposal, which is a large limitation of the paper. Why don't the authors take top-k prompt proposals and put into context instead of just one? Besides, I'm curious whether there is a better choice of prompt proposal, for example, intuitively the function names, docstrings, and variable names of the imported function could help inform the model the functionality of the imported function.\n- It seems that, RLPG-H is more like classification, while RLPG-R is more like retrieval. Would a lexical search method like BM25 could retrieve relevant prompts already compared? That way no training for the PPC is needed. And I would expect it works reasonably well.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a method to make prompt proposals from repository-level context and add concatenate the prompt proposals with normal context to LLM in order to feed more related information to achieve better generations. The paper explores different sources of repository level prompts, as well as prompt context types. Experiment results show that additional repository level prompt help improve the performance significantly.\n", "strength_and_weaknesses": "Strength:\n- The paper is well-written. The method is well explained with examples. Experiment results are extensive with analysis.\n- The paper explores several variants on the prompt sources and prompt types, and made a nice figure to demonstrate which prompt source works the best, though only one is used in experiment at a time.\n\n\nWeakness:\n- Taking the whole file from the whole import file seems unnecessarily large context. Rather, if the import statement only imports one class or one function from the module, it makes sense to only consider the function implementation, rather than the whole file where the function resides. This especially applies to 7) Method Names and Bodies prompt type.\n- The paper only considers one prompt proposal, which is a large limitation of the paper. Why don't the authors take top-k prompt proposals and put into context instead of just one? Besides, I'm curious whether there is a better choice of prompt proposal, for example, intuitively the function names, docstrings, and variable names of the imported function could help inform the model the functionality of the imported function.\n- It seems that, RLPG-H is more like classification, while RLPG-R is more like retrieval. Would a lexical search method like BM25 could retrieve relevant prompts already compared? That way no training for the PPC is needed. And I would expect it works reasonably well.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is well-writtern and is of good quality. The proposed method is somewhat novel. And I expect that the results should be easy to reproduce.\n", "summary_of_the_review": "Overall, I think this is a decent paper with novel method proposed and well-performed experiments. I have some concerns on the paper, and would want to see more experiments and results to make the paper more complete.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666851224074}, {"id": "-ksgVN5wz7R", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4604/Reviewer_xrrm"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper is part of the stream of papers that try and generate code given a prompt. This paper tries to assess that if additional context is given as part of the prompt, whether the performance can improve. In particular, they predefine different types of contexts and then train a classifier to predict which context should be prepended to the prompt. To get the training data, they given different types of contexts to Codex and if a particular context is able to make the completion match the ground truth completion, they label that context as 1. They then train a multi-label classifier and use its predictions at inference time to decide what context to add on to the prompt.\n\nThey compare against several baselines - no additional context, oracle context (using the ground truth correct contexts) which is the upper bound, random context, nearest neighbors in representation space from the random contexts, lines using the closest identifier throughout the repo, and context to the right of the line to be completed.\n\nThe strongest baseline is the right context baseline against which they show a 1-4% relative improvement and a 14-16% relative improvement over no additional context (Table 1)", "review_text": "While I believe the ideas in the paper are interesting and novel, the fact that we do not know the pretraining data which could easily overlap with the test data + the deduping with respect to the the known part of pretraining data being imperfect (based on repo name match instead of file content match or suffix array deduping), the results are not reliable.", "strengths": "Strengths:\n\n(1) The idea of generating training labels for the correct contexts and training a classifier to predict what the correct contexts should be is novel and interesting\n\n(2) They investigate multiple possible baselines \n\nWeaknesses:\n\n(1) Overlap with pretraining data -- Github is not the only source of training data for Codex (https://openai.com/blog/openai-codex/ -- \"OpenAI Codex is a descendant of GPT-3; its training data contains both natural language and billions of lines of source code from publicly available sources, including code in public GitHub repositories\"). Thus it is very possible that Google Code -- which they use for their test data -- is part of the training data for Codex. Furthermore, deduping based on just repo name match is highly imperfect. A much better way would be file level deduping or suffix array based deduping. Unless there's a better understanding of what the pretraining data consists of, the results might be invalid. The authors could consider using a model like CodeGen for which the pretraining dataset is known. Alternatively they could use code published after June 2021 (the training data cutoff for the Codex davinci-002 model), dedupe it against the code data available upto June 2021 and use that as their test data.\n\n(2) The improvements over both right context as well as identifier usage baselines are modest.\n\n(3) They don't explore how different prompt proposals could be combined with each other", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper is part of the stream of papers that try and generate code given a prompt. This paper tries to assess that if additional context is given as part of the prompt, whether the performance can improve. In particular, they predefine different types of contexts and then train a classifier to predict which context should be prepended to the prompt. To get the training data, they given different types of contexts to Codex and if a particular context is able to make the completion match the ground truth completion, they label that context as 1. They then train a multi-label classifier and use its predictions at inference time to decide what context to add on to the prompt.\n\nThey compare against several baselines - no additional context, oracle context (using the ground truth correct contexts) which is the upper bound, random context, nearest neighbors in representation space from the random contexts, lines using the closest identifier throughout the repo, and context to the right of the line to be completed.\n\nThe strongest baseline is the right context baseline against which they show a 1-4% relative improvement and a 14-16% relative improvement over no additional context (Table 1)", "strength_and_weaknesses": "Strengths:\n\n(1) The idea of generating training labels for the correct contexts and training a classifier to predict what the correct contexts should be is novel and interesting\n\n(2) They investigate multiple possible baselines \n\nWeaknesses:\n\n(1) Overlap with pretraining data -- Github is not the only source of training data for Codex (https://openai.com/blog/openai-codex/ -- \"OpenAI Codex is a descendant of GPT-3; its training data contains both natural language and billions of lines of source code from publicly available sources, including code in public GitHub repositories\"). Thus it is very possible that Google Code -- which they use for their test data -- is part of the training data for Codex. Furthermore, deduping based on just repo name match is highly imperfect. A much better way would be file level deduping or suffix array based deduping. Unless there's a better understanding of what the pretraining data consists of, the results might be invalid. The authors could consider using a model like CodeGen for which the pretraining dataset is known. Alternatively they could use code published after June 2021 (the training data cutoff for the Codex davinci-002 model), dedupe it against the code data available upto June 2021 and use that as their test data.\n\n(2) The improvements over both right context as well as identifier usage baselines are modest.\n\n(3) They don't explore how different prompt proposals could be combined with each other", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\nIt's not clear how they are selecting the prompt proposal at inference time. Are they taking the argmax over the probabilities for the different prompt proposals?\n\nQuality:\nError bars are not given. It's unclear how much overlap the test set has with the pretraining data.\n", "summary_of_the_review": "While I believe the ideas in the paper are interesting and novel, the fact that we do not know the pretraining data which could easily overlap with the test data + the deduping with respect to the the known part of pretraining data being imperfect (based on repo name match instead of file content match or suffix array deduping), the results are not reliable.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666772319204}, {"id": "ddhYN3Isbv", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4604/Reviewer_VAEL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a repository-level prompt generation (RLPG) approach for language models of source code. Without having access to the weight of LLMs, RLPG can improve the performance of LM in code completion task. Through training the prompt proposal classifier, the prompt composer can generate high quality prompts for LM, increasing the accuracy of code completion.", "review_text": "I appreciate that the proposed technique can generate prompts in a black box manner. Different from previous work about line-level code completion, this paper considers the completion from the cursor to the end of the line. It involves a wide range of prompt sources and prompt context types. The proposed approach is well ablated and extensive experiments are conducted. RLPG is also efficient for both training and inference.\n\nThe proposed work learns to generate example-specific prompts using prompt proposals, which are taken from the structure of the repository and the context from other relevant files (e.g., imports, parent class files). The authors may also consider other code information such as API. See, for example:\n\nLyu et al., Embedding API dependency graph for neural code generation. Empirical Software Engineering, 26(4):61, 2021. \n\nFor RLPG-H, the used code context is two lines before and after the hole. Are two lines (or four) enough? Maybe the target hole is related to some statements far away from the target (e.g. some global variable definition).\n\nThe Identifier Usage (Random) also achieves promising results, why?\n\nFigure 1 is not easy to understand.\n\nRecently, the following work also applies prompt learning to pre-trained code models: Wang et al., No More Fine-Tuning? An Experimental Evaluation of Prompt Tuning in Code Intelligence, https://arxiv.org/abs/2207.11680.\n\nThe proposed technique is more about software engineering/programming language, the novel technical contribution to neural networks/language models is a bit limited. \n", "strengths": "Strengths:\n\n- A novel way of generating prompts.\n\n- Extensive experiments.\n\n- RLPG is efficient for both training and inference.\n\nWeaknesses:\n\n- The paper is better for a software engineering/programming language conference.\n\n- Only the structure of the repository and the context from other relevant files are considered in the prompt proposals.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a repository-level prompt generation (RLPG) approach for language models of source code. Without having access to the weight of LLMs, RLPG can improve the performance of LM in code completion task. Through training the prompt proposal classifier, the prompt composer can generate high quality prompts for LM, increasing the accuracy of code completion.", "strength_and_weaknesses": "Strengths:\n\n- A novel way of generating prompts.\n\n- Extensive experiments.\n\n- RLPG is efficient for both training and inference.\n\nWeaknesses:\n\n- The paper is better for a software engineering/programming language conference.\n\n- Only the structure of the repository and the context from other relevant files are considered in the prompt proposals.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is generally well written. The implementation of the model is not available for replication study. The proposed work appears to be novel.", "summary_of_the_review": "I appreciate that the proposed technique can generate prompts in a black box manner. Different from previous work about line-level code completion, this paper considers the completion from the cursor to the end of the line. It involves a wide range of prompt sources and prompt context types. The proposed approach is well ablated and extensive experiments are conducted. RLPG is also efficient for both training and inference.\n\nThe proposed work learns to generate example-specific prompts using prompt proposals, which are taken from the structure of the repository and the context from other relevant files (e.g., imports, parent class files). The authors may also consider other code information such as API. See, for example:\n\nLyu et al., Embedding API dependency graph for neural code generation. Empirical Software Engineering, 26(4):61, 2021. \n\nFor RLPG-H, the used code context is two lines before and after the hole. Are two lines (or four) enough? Maybe the target hole is related to some statements far away from the target (e.g. some global variable definition).\n\nThe Identifier Usage (Random) also achieves promising results, why?\n\nFigure 1 is not easy to understand.\n\nRecently, the following work also applies prompt learning to pre-trained code models: Wang et al., No More Fine-Tuning? An Experimental Evaluation of Prompt Tuning in Code Intelligence, https://arxiv.org/abs/2207.11680.\n\nThe proposed technique is more about software engineering/programming language, the novel technical contribution to neural networks/language models is a bit limited. \n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666601925028}, {"id": "hhWb5DvQ_i_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4604/Reviewer_9is8"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors propose prompt generator based on the whole code repository for the code completion tasks.\nAuthors show a significant improvement over baseline Codex results by using their prompt generator.", "review_text": "I think that the paper presents a novel approach for improving results of tasks such as code completion by generating prompts from whole repository using a trained framework. This is interesting work and should be accepted to ICLR.", "strengths": "Strengths:\n- Addresses an important and novel area of good prompt generation for LLM tasks\n- Approach does not require access to the weights of LLM\n- Suggests and implements prompt generator that uses information from the whole repository. The prompt generator also uses repository level prompt \"proposals\" (rules/suggestions).\n- Using generated prompts significantly improves Codex results\n\nWeaknesses:\n\n- It is not clear whether prompt proposals are really per-repository. It seems that they will be pretty universal and should not vary much. So the value of repository-specific prompt proposals is not really established or proven. This is a minor issue though.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "Authors propose prompt generator based on the whole code repository for the code completion tasks.\nAuthors show a significant improvement over baseline Codex results by using their prompt generator.", "strength_and_weaknesses": "Strengths:\n- Addresses an important and novel area of good prompt generation for LLM tasks\n- Approach does not require access to the weights of LLM\n- Suggests and implements prompt generator that uses information from the whole repository. The prompt generator also uses repository level prompt \"proposals\" (rules/suggestions).\n- Using generated prompts significantly improves Codex results\n\nWeaknesses:\n\n- It is not clear whether prompt proposals are really per-repository. It seems that they will be pretty universal and should not vary much. So the value of repository-specific prompt proposals is not really established or proven. This is a minor issue though.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written. \nI think that the paper presents a novel approach for improving results of tasks such as code completion by generating prompts from whole repository using a trained framework.\nThe results of the paper should be reproducible assuming that the dataset used by authors to train the PPC is available.\n\n", "summary_of_the_review": "I think that the paper presents a novel approach for improving results of tasks such as code completion by generating prompts from whole repository using a trained framework. This is interesting work and should be accepted to ICLR.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666581802322}], "openreview_url": "https://openreview.net/forum?id=MtGmCCPJD-", "arxiv_id": "2206.12839", "paper_pdf": "papers/MtGmCCPJD-.pdf", "paper_pdf_sha256": "2d88ae7c3f1764c4ba10c08def0acb0edc46217ab8b3b3dc2fc55ac169f42238", "paper_pdf_bytes": 509561, "paper_pdf_source": "openreview", "code_url": "https://github.com/shrivastavadisha/repo_level_prompt_generation", "code_repository": "shrivastavadisha/repo_level_prompt_generation", "code_commit": "3af5f3424740448d8e325b3726e61944f6eec8b6", "code_archive": "repos/MtGmCCPJD-.zip", "code_archive_sha256": "d220dde993f61f5712b042b3f1077b29abdb317a01728c84e79a48276920b50d", "code_archive_bytes": 459465, "code_file_count": 21, "code_extensions": {".py": 21}, "github_disk_usage_kb": 865, "github_languages": {"Python": 150598}, "github_archived": false, "github_pushed_at": "2023-04-22T15:23:11Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/repository-level-prompt-generation-for-large"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "G-7GlfTneYg", "year": 2022, "status": "rejected", "title": "VoiceFixer: Toward General Speech Restoration with Neural Vocoder", "authors": ["Haohe Liu", "Qiuqiang Kong", "Qiao Tian", "Yan Zhao", "DeLiang Wang", "Chuanzeng Huang", "Yuxuan Wang"], "authorids": ["~Haohe_Liu1", "~Qiuqiang_Kong1", "~Qiao_Tian1", "~Yan_Zhao6", "~DeLiang_Wang1", "~Chuanzeng_Huang1", "~Yuxuan_Wang1"], "authors_source": "OpenReview API", "abstract": "Speech restoration aims to remove distortions in speech signals. Prior methods mainly focus on single-task speech restoration (SSR), such as speech denoising or speech declipping. However, SSR systems only focus on one task and do not address the general speech restoration problem. In addition, previous SSR systems show limited performance in some speech restoration tasks such as speech super-resolution. To overcome those limitations, we propose a general speech restoration (GSR) task that attempts to remove multiple distortions \u0002simultaneously. Furthermore, we propose VoiceFixer, a generative framework to address the GSR task. VoiceFixer consists of an analysis stage and a synthesis stage to mimic the speech analysis and comprehension of the human auditory system. We employ a ResUNet to model the analysis stage and a neural vocoder to model the synthesis stage. We evaluate VoiceFixer with additive noise, room \u0002reverberation, low-resolution, and clipping distortions. Our baseline GSR model achieves a 0.499 higher mean opinion score (MOS) than the speech denoising SSR model. VoiceFixer further surpasses the GSR baseline model on the MOS score by 0.256. Moreover, we observe that VoiceFixer generalizes well to severely degraded real speech recordings, indicating its potential in restoring old movies and historical speeches.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "pt1A8nc4bFQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3103/Reviewer_Hsxt"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes an approach called VoiceFixer which is aimed at restoring degraded speech signals. The paper considers a variety of speech degradations - additive noise, reverberations, clipping and limited bandwidth. The paper describes a two stage approach in which the first stage aims to produce restored mel-spectrogram and then a vocoder is used to synthesize the  speech from the restored mel-spectrogram. Experiments are done using the VCTK dataset and experiments are done using single distortions as well as combinations of all 4 distortions. ", "review_text": "1. The paper claims that most speech restoration works are driven by a single task. I do not completely agree with this statement. There are plenty of works which have explored speech denoising and dereverberations together, conventional signal processing based methods as well as neural network based methods [R1, R3, R4, R5]. Some works such as HiFi-GAN also consider other types of distortions such as equalization distortions. \n\n\n2. Overall, I find very little to no novelty in the paper with respect to the proposed method. Two stage approaches like the one proposed in the paper have been used before. In fact, the prior work [R2, R6] takes more or less the exact same approach for speech enhancement -- reconstruct acoustic features and then use vocoder. \n\n\n3. Moreover, the paper does not propose any novel method for each stage. The first stage (analysis stage) uses a pre-existing architecture with well-known L1 based reconstruction loss. Similarly, the second stage also uses a prior vocoder (TFGAN) without any modifications. \n\n\n4. On the empirical front also, the paper leaves out quite a few desirable and expected experiments. Experiments are done using only the VCTK dataset. VCTK and VCTK-demand are both relatively small datasets and do not provide the best picture of the performance. Moreover, one would expect comparisons with latest state of the art approaches for denoising (enhancement) and dereverberations. There are quite a lot of works in these areas and it is expected to compare with reasonable state of the art methods. \n\n\n5. It would have been interesting if the paper showed results using different methods for mel-restoration (they could use a variety of prior works here) and different vocoders in stage 2. Considering that the two stage approach and the methods used in each stage have been done before, this would have helped establish the significance of different approaches for analysis and the synthesis stages. The significance of each stage are not properly decoupled as well. Can we keep the same vocoder but improve the overall system somehow by improving analysis stage ? Similarly, the other way round as well. \n\n\n6. It's not clear to me if Denoise-UNet, Dereverb-UNet etc are single stage or 2 stage approaches. Results tables mention them as a one-stage model. How are restored speeches obtained ? Are they directly predicting clean speech or similar to the two stage predict restored mel and then uses vocoder for synthesis. In the latter case, these are also two stage approaches ??\n\n\n7. The paper claims to be solving general speech restoration. This could have been a focus of the paper. Showing how training with all distortions helps in improving restoration in case of single as well as multiple distortions. However, this is not emphasized and clear through experiments and discussions. \n\n\n8. While such a two stage approach makes sense, the connection between human hearing and the proposed approach is rather weak. I am not sure if it is a good idea to make that connection here. \n\n\n9. Except for minor language issues here and there, the paper is clearly written. The authors have also provided demos and codes through anonymous github links. \n\n\n10.  What happens if the two stages are trained jointly -- in an end to end fashion ? Is it better or worse than training separately ? \n\n\n11. Some analysis w.r.t level of distortions (e.g SNR level for additive noise, RT60 for reverberations, amount of clipping etc.) would have helped better understand the performance. \n\n\nR1: Joint Dereverberation And Noise Reduction Using Beamforming\nAnd A Single-channel Speech Enhancement Scheme, Cauchi et. al., 2014\n\n\nR2: Speaker Independence Of Neural Vocoders And Their Effect On Parametric Resynthesis Speech Enhancement, Maiti and Mandel, 2019. \n\nR3: Denoising-and-dereverberation Hierarchical Neural Vocoder For Robust Waveform Generation, Ai et. al., 2020\n\nR4: A Unified Convolutional Beamformer For Simultaneous Denoising And Dereverberation, Nakatani et. al., 2019. \n\nR5: HiFi-GAN: High-Fidelity Denoising and Dereverberation Based on Speech Deep Features in Adversarial Networks, Su et. al., 2019\n\nR6: Speech Enhancement Using Speech Synthesis Techniques, Maiti, 2021\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes an approach called VoiceFixer which is aimed at restoring degraded speech signals. The paper considers a variety of speech degradations - additive noise, reverberations, clipping and limited bandwidth. The paper describes a two stage approach in which the first stage aims to produce restored mel-spectrogram and then a vocoder is used to synthesize the  speech from the restored mel-spectrogram. Experiments are done using the VCTK dataset and experiments are done using single distortions as well as combinations of all 4 distortions. ", "main_review": "1. The paper claims that most speech restoration works are driven by a single task. I do not completely agree with this statement. There are plenty of works which have explored speech denoising and dereverberations together, conventional signal processing based methods as well as neural network based methods [R1, R3, R4, R5]. Some works such as HiFi-GAN also consider other types of distortions such as equalization distortions. \n\n\n2. Overall, I find very little to no novelty in the paper with respect to the proposed method. Two stage approaches like the one proposed in the paper have been used before. In fact, the prior work [R2, R6] takes more or less the exact same approach for speech enhancement -- reconstruct acoustic features and then use vocoder. \n\n\n3. Moreover, the paper does not propose any novel method for each stage. The first stage (analysis stage) uses a pre-existing architecture with well-known L1 based reconstruction loss. Similarly, the second stage also uses a prior vocoder (TFGAN) without any modifications. \n\n\n4. On the empirical front also, the paper leaves out quite a few desirable and expected experiments. Experiments are done using only the VCTK dataset. VCTK and VCTK-demand are both relatively small datasets and do not provide the best picture of the performance. Moreover, one would expect comparisons with latest state of the art approaches for denoising (enhancement) and dereverberations. There are quite a lot of works in these areas and it is expected to compare with reasonable state of the art methods. \n\n\n5. It would have been interesting if the paper showed results using different methods for mel-restoration (they could use a variety of prior works here) and different vocoders in stage 2. Considering that the two stage approach and the methods used in each stage have been done before, this would have helped establish the significance of different approaches for analysis and the synthesis stages. The significance of each stage are not properly decoupled as well. Can we keep the same vocoder but improve the overall system somehow by improving analysis stage ? Similarly, the other way round as well. \n\n\n6. It's not clear to me if Denoise-UNet, Dereverb-UNet etc are single stage or 2 stage approaches. Results tables mention them as a one-stage model. How are restored speeches obtained ? Are they directly predicting clean speech or similar to the two stage predict restored mel and then uses vocoder for synthesis. In the latter case, these are also two stage approaches ??\n\n\n7. The paper claims to be solving general speech restoration. This could have been a focus of the paper. Showing how training with all distortions helps in improving restoration in case of single as well as multiple distortions. However, this is not emphasized and clear through experiments and discussions. \n\n\n8. While such a two stage approach makes sense, the connection between human hearing and the proposed approach is rather weak. I am not sure if it is a good idea to make that connection here. \n\n\n9. Except for minor language issues here and there, the paper is clearly written. The authors have also provided demos and codes through anonymous github links. \n\n\n10.  What happens if the two stages are trained jointly -- in an end to end fashion ? Is it better or worse than training separately ? \n\n\n11. Some analysis w.r.t level of distortions (e.g SNR level for additive noise, RT60 for reverberations, amount of clipping etc.) would have helped better understand the performance. \n\n\nR1: Joint Dereverberation And Noise Reduction Using Beamforming\nAnd A Single-channel Speech Enhancement Scheme, Cauchi et. al., 2014\n\n\nR2: Speaker Independence Of Neural Vocoders And Their Effect On Parametric Resynthesis Speech Enhancement, Maiti and Mandel, 2019. \n\nR3: Denoising-and-dereverberation Hierarchical Neural Vocoder For Robust Waveform Generation, Ai et. al., 2020\n\nR4: A Unified Convolutional Beamformer For Simultaneous Denoising And Dereverberation, Nakatani et. al., 2019. \n\nR5: HiFi-GAN: High-Fidelity Denoising and Dereverberation Based on Speech Deep Features in Adversarial Networks, Su et. al., 2019\n\nR6: Speech Enhancement Using Speech Synthesis Techniques, Maiti, 2021\n", "summary_of_the_review": "The novelty of the paper in terms of the methods is not much. The analysis and synthesis approach as well as the method for each stage are not particularly novel. Several prior works have used such methods. The experimental results and analyses are also not strong and does not provide interesting insights. Overall, the paper lacks strength on both fronts in the current form. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1636181257005}, {"id": "ebHXdJfea0d", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3103/Reviewer_UVJ3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a general speech restoration (GSR) task that tries to remove multiple distortions in a single model. In addition, it also presents a generative framework called VoiceFixer consisting of analysis and synthesis stages to address the general speech restoration task. In VoiceFixer, the authors employ a ResNet for modeling the analysis stage and a TFGAN-based neural vocoder for synthesis stage. They report that their baseline GSR and VoiceFixer surpass the single speech restoration (SSR) models with more improved results by the latter. Their idea was well described and the experiments are systematical and extensive. The results are consistent and clear. \nThe contribution of this paper is to incorporate a variety of speech restoration tasks including speech denoising, super-resolution, dereverberation, declipping, etc. in a single unified task called GSR. Another is the proposal of a well-performing generative speech restoration framework called VoiceFixer.\n", "review_text": "In this paper, the authors propose a single model-based general speech restoration (GSR) task for removing multiple distortions simultaneously. They also propose a two-stage-based generative framework called VoiceFixer to address the general speech restoration task. To develop and evaluate these models, they performed extensive experiments in the areas of speech restoration such as denoising, dereverberation, super-resolution, and declipping.\nTheir proposed single model-based GSR approach employing ResUNet provided improved results over the well-known existing single-task speech restoration techniques. VoiceFixer even surpasses their single model-based GSR approach. It is meaningful and impressive that the two proposed single-model-based approaches outperform the well-known existing single-task speech restoration techniques in almost every speech restoration tasks described in the paper.\nThe authors report that VoiceFixer outperforms the GSR approach in the experimental results. However, it is unclear which makes VoiceFixer superior to the GSR approach. The author needs to clarify this issue.\nA weak point may be that VoiceFixer restores speech as a synthesis form by employing a neural vocoder which inevitably entails some degree of distortion when compared with real speech. Therefore, when the raw speech has relatively small amount of distortion, the restored speech is still vocoder-decoded speech which may be inferior to the uncorrupted clean speech. In this case, the merit of VoiceFixer over the IFFT-based approach may be decreased. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a general speech restoration (GSR) task that tries to remove multiple distortions in a single model. In addition, it also presents a generative framework called VoiceFixer consisting of analysis and synthesis stages to address the general speech restoration task. In VoiceFixer, the authors employ a ResNet for modeling the analysis stage and a TFGAN-based neural vocoder for synthesis stage. They report that their baseline GSR and VoiceFixer surpass the single speech restoration (SSR) models with more improved results by the latter. Their idea was well described and the experiments are systematical and extensive. The results are consistent and clear. \nThe contribution of this paper is to incorporate a variety of speech restoration tasks including speech denoising, super-resolution, dereverberation, declipping, etc. in a single unified task called GSR. Another is the proposal of a well-performing generative speech restoration framework called VoiceFixer.\n", "main_review": "In this paper, the authors propose a single model-based general speech restoration (GSR) task for removing multiple distortions simultaneously. They also propose a two-stage-based generative framework called VoiceFixer to address the general speech restoration task. To develop and evaluate these models, they performed extensive experiments in the areas of speech restoration such as denoising, dereverberation, super-resolution, and declipping.\nTheir proposed single model-based GSR approach employing ResUNet provided improved results over the well-known existing single-task speech restoration techniques. VoiceFixer even surpasses their single model-based GSR approach. It is meaningful and impressive that the two proposed single-model-based approaches outperform the well-known existing single-task speech restoration techniques in almost every speech restoration tasks described in the paper.\nThe authors report that VoiceFixer outperforms the GSR approach in the experimental results. However, it is unclear which makes VoiceFixer superior to the GSR approach. The author needs to clarify this issue.\nA weak point may be that VoiceFixer restores speech as a synthesis form by employing a neural vocoder which inevitably entails some degree of distortion when compared with real speech. Therefore, when the raw speech has relatively small amount of distortion, the restored speech is still vocoder-decoded speech which may be inferior to the uncorrupted clean speech. In this case, the merit of VoiceFixer over the IFFT-based approach may be decreased. \n", "summary_of_the_review": "The reviewed paper proposed two new approaches for speech restoration. The first approach is a single-model-based generative speech restoration technique that can cope with a variety of speech restoration tasks simultaneously. The other is a generative speech restoration frame called VoiceFixer which provides enhanced performance over the well-known existing techniques as well as the proposed GSR approach. \nIn spite of some unclear and weak points mentioned above, the reviewer thinks the two proposed approaches highly and believes that they could impact positively on the related society.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635849335434}, {"id": "d4v-r4tpQiW", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3103/Reviewer_Q5gX"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper introduces a unified view of several speech restoration problems including denoising, decliping, dereverberation and audio super-resolution. In order to solve the problem of the general speech restoration task, the authors propose a U-Net architecture which is trained on all of these tasks simultaneously during training time. The authors conduct extensive experiments for the general speech restoration task as well as the individual tasks where they compare the proposed models with more specialized models bounded to each distortion. The experimental results show that the proposed VoiceFixer combination of the model and the analysis-synthesis procedure are capable of effectively removing the speech distortions and in some cases outperform previous approaches in the literature.", "review_text": "First of all, I would like to mention that this is a well written paper which discusses an interesting idea for addressing the problem of speech distortions in a more general way. The paper also details many experimental design choices that researchers would need to replicate the results, which is a strong plus. It is also important that the authors have conducted subjective evaluation experiments to show how their algorithm performs wrt the opinion of various humans. Overall, I believe that the paper has great potential if some concerns are addressed. Please see below my concerns and questions to the authors in descending order of importance (0 is the most important):\n\n0. The main problem of this paper stems from the limited originality of the idea behind the general speech restoration method proposed. Specifically, the generative setup where a generative network tries to capture the distribution of clean speech and generate relevant frames has been recently proposed in [1] and shown great potential for speech enhancement. On the other hand, breaking the optimization process in two distinct steps for operating on a feature space where the speech restoration could be easier to obtain has been proposed in [2] for speech and universal audio separation tasks. Although the combination of the two ideas is interesting, the paper does not sufficiently discuss how this approach works better (or not) compared to the end-to-end optimization where discriminative training would be used with the exact same computational power used.\n\n1. The authors need to also show the extra computational load that the VoiceFixer method induces compared to the gain that it delivers. Specifically, I have the following suggestions: The authors propose a complex architecture that might severely enlarge the actual memory footprint of a forward pass of the network (or backward) mainly because of the intermediate activations tensors. Moreover, the  authors should report the number of trainable parameters, the number of FLOPs as well as the actual memory requirements, time consumed during training on GPU / inference on CPU. Most importantly, this comparison could also be used to adress the concern above that would show that VoiceFixer helps even against the same computational load.\n\n2. The training algorithm as described in the Appendix (Algorithm 1) could be utilized using any time-domain state of the art network. For example, I do not see the reason of why not directly using a state-of-the-art speech denoising network [3, 4, 5] and train it for all the 4 tasks using Algorithm 1 and simply put a few extra upsampling layers for the super-resolution audio task. Also, I would expect that the model comparisons, especially for tasks such as denoising, which have open-sourced benchmark datasets, to be conducted using datasets like the DNS-Challenge [6] (but this is more optional compared to the main point of direct comparison with current state-of-the-art models).\n\n3. Although the authors provided a few sample spectrograms in the Appendix, it is really hard to understand from an image how the actual speech sounds. Since the MOS have been reported and generative setups have been employed it is really important to include a few random reconstructed speech files to assess the performance and let the reviewers actually hear the results of the method and compare.\n\n4. The authors have identified that because of the use of generative setups it might be the case that point-wise metrics such as SISNR do not capture how realistic a generated speech file sounds. Although this is true in general, this decline in performance is often caused by time-domain shifts that can be easily fixed with the appropriate padding (e.g. the metric wrt log MSE on the spectrogram domain does not display a similar behavior, probably because the computation window is wider).\n\n5. How the authors explain that for the super-resolution task (e.g. Figure 6e), the mean MOS for a few models are better than the high-resolution target itself and also worse than the unprocessed file? This is extremely odd and is not at all consistent with Figure 6b. \n\n**Minor**\n- There are probably more typos but here are a few that I have found:\n\n    - we design a similar metrics → we design similar metrics \n    - The experiment code can conduct evaluations → The code can be used to conduct the evaluations\n    - Sections  4.1 and 4.4, include “:” after each bold word\n\n- Last Paragraph of Section D3 Appendix:\n\n    - all these examples prove the effectiveness → all these examples showcase the effectiveness\n    - which proves the advantages → which displays the advantages\n    - Last but not least, despite the abnormal harmonic structure in the low-frequency part in Figure 12g, our proposed model can still repair it into a normal distribution, → the phrasing repairing something into a normal distribution is both wrong and misleading.\n\n- What is the reason where some vectors are denoted with bold and others not? (E.g. equation 29)\n\n[1] Adam Polyak, Lior Wolf, Yossi Adi, Ori Kabeli, and Yaniv Taigman. High fidelity speech regeneration with application to speech enhancement. In Proceedings of the IEEE Conference on Acoustics, Speech, and Signal Processing, pp. 7143–7147, 2021.\n\n[2] Efthymios Tzinis, Shrikant Venkataramani, Zhepei Wang, Cem Subakan, Paris Smaragdis. “Two-step sound source separation: Training on learned latent targets.” In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2020 May 4 (pp. 31-35).\n\n[3] Umut Isik, Ritwik Giri, Neerad Phansalkar, Jean-Marc Valin, Karim Helwani, and Arvindh Krishnaswamy, “Poconet: Better speech enhancement with frequency-positional embeddings, semi-supervised conversational data, and biased loss,” in Proc. Interspeech, 2020, pp. 2487–2491.\n\n[4] Ashutosh Pandey and DeLiang Wang, “Dense cnn with selfattention for time-domain speech enhancement,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1270–1279, 2021.\n\n[5] Défossez, A., Synnaeve, G., Adi, Y. (2020) Real Time Speech Enhancement in the Waveform Domain. Proc. Interspeech 2020, pp. 3291-3295.\n\n[6] Chandan KA Reddy et al., “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” in Proc. Interspeech, 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a unified view of several speech restoration problems including denoising, decliping, dereverberation and audio super-resolution. In order to solve the problem of the general speech restoration task, the authors propose a U-Net architecture which is trained on all of these tasks simultaneously during training time. The authors conduct extensive experiments for the general speech restoration task as well as the individual tasks where they compare the proposed models with more specialized models bounded to each distortion. The experimental results show that the proposed VoiceFixer combination of the model and the analysis-synthesis procedure are capable of effectively removing the speech distortions and in some cases outperform previous approaches in the literature.", "main_review": "First of all, I would like to mention that this is a well written paper which discusses an interesting idea for addressing the problem of speech distortions in a more general way. The paper also details many experimental design choices that researchers would need to replicate the results, which is a strong plus. It is also important that the authors have conducted subjective evaluation experiments to show how their algorithm performs wrt the opinion of various humans. Overall, I believe that the paper has great potential if some concerns are addressed. Please see below my concerns and questions to the authors in descending order of importance (0 is the most important):\n\n0. The main problem of this paper stems from the limited originality of the idea behind the general speech restoration method proposed. Specifically, the generative setup where a generative network tries to capture the distribution of clean speech and generate relevant frames has been recently proposed in [1] and shown great potential for speech enhancement. On the other hand, breaking the optimization process in two distinct steps for operating on a feature space where the speech restoration could be easier to obtain has been proposed in [2] for speech and universal audio separation tasks. Although the combination of the two ideas is interesting, the paper does not sufficiently discuss how this approach works better (or not) compared to the end-to-end optimization where discriminative training would be used with the exact same computational power used.\n\n1. The authors need to also show the extra computational load that the VoiceFixer method induces compared to the gain that it delivers. Specifically, I have the following suggestions: The authors propose a complex architecture that might severely enlarge the actual memory footprint of a forward pass of the network (or backward) mainly because of the intermediate activations tensors. Moreover, the  authors should report the number of trainable parameters, the number of FLOPs as well as the actual memory requirements, time consumed during training on GPU / inference on CPU. Most importantly, this comparison could also be used to adress the concern above that would show that VoiceFixer helps even against the same computational load.\n\n2. The training algorithm as described in the Appendix (Algorithm 1) could be utilized using any time-domain state of the art network. For example, I do not see the reason of why not directly using a state-of-the-art speech denoising network [3, 4, 5] and train it for all the 4 tasks using Algorithm 1 and simply put a few extra upsampling layers for the super-resolution audio task. Also, I would expect that the model comparisons, especially for tasks such as denoising, which have open-sourced benchmark datasets, to be conducted using datasets like the DNS-Challenge [6] (but this is more optional compared to the main point of direct comparison with current state-of-the-art models).\n\n3. Although the authors provided a few sample spectrograms in the Appendix, it is really hard to understand from an image how the actual speech sounds. Since the MOS have been reported and generative setups have been employed it is really important to include a few random reconstructed speech files to assess the performance and let the reviewers actually hear the results of the method and compare.\n\n4. The authors have identified that because of the use of generative setups it might be the case that point-wise metrics such as SISNR do not capture how realistic a generated speech file sounds. Although this is true in general, this decline in performance is often caused by time-domain shifts that can be easily fixed with the appropriate padding (e.g. the metric wrt log MSE on the spectrogram domain does not display a similar behavior, probably because the computation window is wider).\n\n5. How the authors explain that for the super-resolution task (e.g. Figure 6e), the mean MOS for a few models are better than the high-resolution target itself and also worse than the unprocessed file? This is extremely odd and is not at all consistent with Figure 6b. \n\n**Minor**\n- There are probably more typos but here are a few that I have found:\n\n    - we design a similar metrics → we design similar metrics \n    - The experiment code can conduct evaluations → The code can be used to conduct the evaluations\n    - Sections  4.1 and 4.4, include “:” after each bold word\n\n- Last Paragraph of Section D3 Appendix:\n\n    - all these examples prove the effectiveness → all these examples showcase the effectiveness\n    - which proves the advantages → which displays the advantages\n    - Last but not least, despite the abnormal harmonic structure in the low-frequency part in Figure 12g, our proposed model can still repair it into a normal distribution, → the phrasing repairing something into a normal distribution is both wrong and misleading.\n\n- What is the reason where some vectors are denoted with bold and others not? (E.g. equation 29)\n\n[1] Adam Polyak, Lior Wolf, Yossi Adi, Ori Kabeli, and Yaniv Taigman. High fidelity speech regeneration with application to speech enhancement. In Proceedings of the IEEE Conference on Acoustics, Speech, and Signal Processing, pp. 7143–7147, 2021.\n\n[2] Efthymios Tzinis, Shrikant Venkataramani, Zhepei Wang, Cem Subakan, Paris Smaragdis. “Two-step sound source separation: Training on learned latent targets.” In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2020 May 4 (pp. 31-35).\n\n[3] Umut Isik, Ritwik Giri, Neerad Phansalkar, Jean-Marc Valin, Karim Helwani, and Arvindh Krishnaswamy, “Poconet: Better speech enhancement with frequency-positional embeddings, semi-supervised conversational data, and biased loss,” in Proc. Interspeech, 2020, pp. 2487–2491.\n\n[4] Ashutosh Pandey and DeLiang Wang, “Dense cnn with selfattention for time-domain speech enhancement,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1270–1279, 2021.\n\n[5] Défossez, A., Synnaeve, G., Adi, Y. (2020) Real Time Speech Enhancement in the Waveform Domain. Proc. Interspeech 2020, pp. 3291-3295.\n\n[6] Chandan KA Reddy et al., “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” in Proc. Interspeech, 2020.", "summary_of_the_review": "Overall, this is a well written paper which discusses an interesting idea for addressing the problem of removing speech distortions in a more general way with strong results. The paper discusses many experimental design choices that researchers would need to replicate the results, which is a strong plus. It is also important that the authors have conducted subjective evaluation experiments to show how their algorithm performs. I think that the paper could have a great potential in the field, however, in its current form, the paper has several limitations which downplay its true potential. Specifically, showing how the proposed method is different from ones presented in the literature as well as why it is more important than simply employing more computational power. Moreoever, there are a few misconceptions in the experimental section that authors need to address before the paper is ready for a publication.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635750119146}, {"id": "FBwKXUzE9q", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3103/Reviewer_AG4n"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a single system to deal with the speech enhancement tasks of denoising, dereverb, bandwidth extension (BWE) and declipping. The system is a two-stage system composed of an analysis module producing mel-band masks and a synthesis module using a vocoder. Both modules reuse existing architectures. Results comparing the proposed system to author-derived counterparts of it and to some existing systems for specific tasks show improvement for the proposed system in some cases, while achieving similar performance as existing systems in other cases. Data and code for reproducibility are provided.", "review_text": "Strengths & weaknesses:\n- [$+$] Considering the more general speech enhancement task rather than isolated tasks is an existing effort that deserves more attention.\n- [$-$] Considering more general speech enhancement tasks is not novel, as claimed in the paper.\n- [$+$] Great effort in the literature review for single-task models.\n- [$-$] Misses a substantial amount of prior work for the multiple-task models (see below). Also in result evaluation of, for instance, the denoising task.\n- [$-$] Reusing existing architectures and training procedures. Lack of novelty from a machine learning perspective.\n- [$-$] No assessment of the effectiveness of the system components, specially the different losses used at different stages.\n- [$+$] Great effort in considering separate evaluations for the different tasks.\n- [$-$] Results fail at showing improvement in some tasks. Some baselines for the denoising task are not appropriately chosen.\n- [$-$] No details on subjective evaluation.\n- [$+$] Great effort on reproducibility, both on the data and code sides.\n- [$-$] Positioning of the paper should be changed.\n- [$-$] Writing and presentation should be improved.\n\nDetailed major issues:\n1. The paper positions itself as \"propose[ing] a general speech restoration (GSR) task\" (example sentence from the Abstract). However, the task has been proposed for quite some time now; the only difference is that perhaps the 4 specific distortions considered in the paper have not been considered together by previous work. Some examples of previous systems performing more than one \"speech restoration\" task are [1], considering clipping, gaps, codecs, [2], considering BWE, clipping, gaps and whispering, [3], considering denoising, dereverb and equalization, and [4-9], considering denoising and dereverberation. To my knowledge, one of the earliest references to the general task of considering multiple speech distortions dates back to 2019 [2].\n1. In my opinion, the paper contains many unsupported claims. One example is in Sec 1: \"To address the mismatch problem... restoration (GSR)\" (mismatch problem is not characterized, formally defined or empirically proven; GSR is not shown to solve the mismatch problem, actually no mismatch data is considered). Another example is in Sec 3.2: \"Through the two-stage processing... human perception of speech\" (no empirical result; unclear motivation). Another example is in Sec 3.2.2: \"The energy loss and phase loss... metallic sounds\" (this is not demonstrated by any result or ablation; also authors do not provide any intuitive motivation). Another example is in Sec 4.1: \"we design a similar metrics... discrepancies on the spectrograms\" (it not shown anywhere that this new metric really helps in comparing generative vs non-generative models, actually results in Table 7 show this is not the case). \n1. In a number of places, the paper tries to establish a relation between a two-stage strategy of analysis/synthesis and \"biological mechanisms\". I personally think this is far-fetched, misleading and unnecessary. Biological mechanisms are poorly introduced, it is hard to see how model design derives from such biological mechanisms, no experiment is performed to validate such relation, etc.\n1. Using a two-stage model, which to my understanding is presented as part of the novelty of the paper, is a common strategy among speech enhancement systems that tackle more than one distortion (see refs. [4-9]). Usually, those are already composed of an analysis plus a synthesis step using a vocoder. Even other papers on denoising employ these two stages (see for example [10]).\n1. Architectures for the analysis and synthesis modules are known and the authors do not perform any substantial modification (or at least they do not mention them nor evaluate their worth). \n1. Most losses used for training are quite known and some of them are standard. The system uses a large combination of those losses and, besides not providing specific motivation in most of the cases, no results are reported assessing the need of such losses or the impact that they have in final performance. It thus results into a complicated training procedure which may be even unnecessary, and provides no learnings for the speech community.\n1. No details are given regarding subjective tests (note that authors sometimes confuse objective with subjective tests, see below). How many listeners? Were they experts? Were they native speakers? Which procedure? Which stimuli? How many of them? Was there some randomization? Specifying the procedure followed in subjective testing is an important point that, if wrong or not specified, by itself alone justifies rejection for a paper in a prestigious venue like ICLR.\n1. Distortions are considered to follow a specific order. However, I have serious doubts of whether this is a realistic consideration. Moreover, if the system really aims at generalization, shouldn't we consider all possible orderings even if they are artificial? (thinking here of human generated signals like podcasts for instance).\n1. I don't think it makes much sense to compare SSR models with GSR ones in the GSR task (as done in Table 1 for instance). It is good that authors compare SSR/GSR models in the SSR tasks, but doing it on GSR is unfair.\n1. In my opinion, most of the results in SSR tasks do not show a significant improvement in terms of MOS. Plots (a,d,e,f) in Fig 6 show that existing systems or simpler counterparts of the proposed system achieve comparable or even better MOS. Only in the tasks of super-resolution and dereverberation the proposed system seems to have a winning edge. This is not bad per se, but puts into perspective the value of the approach.\n1. Competitors for the task of denoising are poorly chosen. There are many systems performing much better than the ones reported as baselines in Table 7, both on the supervised and generative side. Some examples are [11-15].\n1. Some parts of the writing contain typos or could be better phrased.\n\nMinor comments:\n- Abstract: \"However, SSR systems...\" --> Almost same sentence as before.\n- Page 1, paragraph 1: \"due to speech\" --> since\n- I found Sec 2 unnecessary. It could have been moved to the appendix (and this would have left some space to move more interesting parts of the appendix, like some results to the main paper).\n- The used vocoder seems to be deterministic (not generative), as it does not have a noise source. It is just an adversarial network, but not generative.\n- I personally find that the AISHELL data set contains a lot of reverb and is of rather low quality. I wonder why can this be useful to train the model to do dereverberation when AISHELL is used as clean data.\n- Sec 4.4, GSR paragraph: There seems to be a confusion between objective metrics (PESQ, LSD, etc) and subjective metrics (MOS). The authors swap the terms objective and subjective.\n- Some references are incomplete (missing page numbers, journal/conference name, book publisher, etc).\n\nReferences:\n1. https://ieeexplore.ieee.org/document/9414721\n1. http://dx.doi.org/10.21437/Interspeech.2019-2688\n1. https://pixl.cs.princeton.edu/pubs/Su_2021_HSS/Su-HiFi-GAN-2-WASPAA-2021.pdf\n1. https://arxiv.org/pdf/2006.00687v1.pdf\n1. http://arxiv.org/abs/2102.00429\n1. https://arxiv.org/abs/2011.05038\n1. https://arxiv.org/abs/2011.03955\n1. https://gfx.cs.princeton.edu/pubs/Su_2019_PM/Su_2019_enhancement.pdf\n1. https://arxiv.org/abs/2006.05694\n1. https://arxiv.org/pdf/2004.04001.pdf\n1. https://arxiv.org/pdf/2010.11860v1.pdf\n1. https://arxiv.org/abs/2104.03538\n1. https://www.isca-speech.org/archive/pdfs/interspeech_2021/hsieh21_interspeech.pdf\n1. https://arxiv.org/pdf/2006.12847.pdf\n1. https://www.isca-speech.org/archive/pdfs/interspeech_2021/kim21h_interspeech.pdf", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a single system to deal with the speech enhancement tasks of denoising, dereverb, bandwidth extension (BWE) and declipping. The system is a two-stage system composed of an analysis module producing mel-band masks and a synthesis module using a vocoder. Both modules reuse existing architectures. Results comparing the proposed system to author-derived counterparts of it and to some existing systems for specific tasks show improvement for the proposed system in some cases, while achieving similar performance as existing systems in other cases. Data and code for reproducibility are provided.", "main_review": "Strengths & weaknesses:\n- [$+$] Considering the more general speech enhancement task rather than isolated tasks is an existing effort that deserves more attention.\n- [$-$] Considering more general speech enhancement tasks is not novel, as claimed in the paper.\n- [$+$] Great effort in the literature review for single-task models.\n- [$-$] Misses a substantial amount of prior work for the multiple-task models (see below). Also in result evaluation of, for instance, the denoising task.\n- [$-$] Reusing existing architectures and training procedures. Lack of novelty from a machine learning perspective.\n- [$-$] No assessment of the effectiveness of the system components, specially the different losses used at different stages.\n- [$+$] Great effort in considering separate evaluations for the different tasks.\n- [$-$] Results fail at showing improvement in some tasks. Some baselines for the denoising task are not appropriately chosen.\n- [$-$] No details on subjective evaluation.\n- [$+$] Great effort on reproducibility, both on the data and code sides.\n- [$-$] Positioning of the paper should be changed.\n- [$-$] Writing and presentation should be improved.\n\nDetailed major issues:\n1. The paper positions itself as \"propose[ing] a general speech restoration (GSR) task\" (example sentence from the Abstract). However, the task has been proposed for quite some time now; the only difference is that perhaps the 4 specific distortions considered in the paper have not been considered together by previous work. Some examples of previous systems performing more than one \"speech restoration\" task are [1], considering clipping, gaps, codecs, [2], considering BWE, clipping, gaps and whispering, [3], considering denoising, dereverb and equalization, and [4-9], considering denoising and dereverberation. To my knowledge, one of the earliest references to the general task of considering multiple speech distortions dates back to 2019 [2].\n1. In my opinion, the paper contains many unsupported claims. One example is in Sec 1: \"To address the mismatch problem... restoration (GSR)\" (mismatch problem is not characterized, formally defined or empirically proven; GSR is not shown to solve the mismatch problem, actually no mismatch data is considered). Another example is in Sec 3.2: \"Through the two-stage processing... human perception of speech\" (no empirical result; unclear motivation). Another example is in Sec 3.2.2: \"The energy loss and phase loss... metallic sounds\" (this is not demonstrated by any result or ablation; also authors do not provide any intuitive motivation). Another example is in Sec 4.1: \"we design a similar metrics... discrepancies on the spectrograms\" (it not shown anywhere that this new metric really helps in comparing generative vs non-generative models, actually results in Table 7 show this is not the case). \n1. In a number of places, the paper tries to establish a relation between a two-stage strategy of analysis/synthesis and \"biological mechanisms\". I personally think this is far-fetched, misleading and unnecessary. Biological mechanisms are poorly introduced, it is hard to see how model design derives from such biological mechanisms, no experiment is performed to validate such relation, etc.\n1. Using a two-stage model, which to my understanding is presented as part of the novelty of the paper, is a common strategy among speech enhancement systems that tackle more than one distortion (see refs. [4-9]). Usually, those are already composed of an analysis plus a synthesis step using a vocoder. Even other papers on denoising employ these two stages (see for example [10]).\n1. Architectures for the analysis and synthesis modules are known and the authors do not perform any substantial modification (or at least they do not mention them nor evaluate their worth). \n1. Most losses used for training are quite known and some of them are standard. The system uses a large combination of those losses and, besides not providing specific motivation in most of the cases, no results are reported assessing the need of such losses or the impact that they have in final performance. It thus results into a complicated training procedure which may be even unnecessary, and provides no learnings for the speech community.\n1. No details are given regarding subjective tests (note that authors sometimes confuse objective with subjective tests, see below). How many listeners? Were they experts? Were they native speakers? Which procedure? Which stimuli? How many of them? Was there some randomization? Specifying the procedure followed in subjective testing is an important point that, if wrong or not specified, by itself alone justifies rejection for a paper in a prestigious venue like ICLR.\n1. Distortions are considered to follow a specific order. However, I have serious doubts of whether this is a realistic consideration. Moreover, if the system really aims at generalization, shouldn't we consider all possible orderings even if they are artificial? (thinking here of human generated signals like podcasts for instance).\n1. I don't think it makes much sense to compare SSR models with GSR ones in the GSR task (as done in Table 1 for instance). It is good that authors compare SSR/GSR models in the SSR tasks, but doing it on GSR is unfair.\n1. In my opinion, most of the results in SSR tasks do not show a significant improvement in terms of MOS. Plots (a,d,e,f) in Fig 6 show that existing systems or simpler counterparts of the proposed system achieve comparable or even better MOS. Only in the tasks of super-resolution and dereverberation the proposed system seems to have a winning edge. This is not bad per se, but puts into perspective the value of the approach.\n1. Competitors for the task of denoising are poorly chosen. There are many systems performing much better than the ones reported as baselines in Table 7, both on the supervised and generative side. Some examples are [11-15].\n1. Some parts of the writing contain typos or could be better phrased.\n\nMinor comments:\n- Abstract: \"However, SSR systems...\" --> Almost same sentence as before.\n- Page 1, paragraph 1: \"due to speech\" --> since\n- I found Sec 2 unnecessary. It could have been moved to the appendix (and this would have left some space to move more interesting parts of the appendix, like some results to the main paper).\n- The used vocoder seems to be deterministic (not generative), as it does not have a noise source. It is just an adversarial network, but not generative.\n- I personally find that the AISHELL data set contains a lot of reverb and is of rather low quality. I wonder why can this be useful to train the model to do dereverberation when AISHELL is used as clean data.\n- Sec 4.4, GSR paragraph: There seems to be a confusion between objective metrics (PESQ, LSD, etc) and subjective metrics (MOS). The authors swap the terms objective and subjective.\n- Some references are incomplete (missing page numbers, journal/conference name, book publisher, etc).\n\nReferences:\n1. https://ieeexplore.ieee.org/document/9414721\n1. http://dx.doi.org/10.21437/Interspeech.2019-2688\n1. https://pixl.cs.princeton.edu/pubs/Su_2021_HSS/Su-HiFi-GAN-2-WASPAA-2021.pdf\n1. https://arxiv.org/pdf/2006.00687v1.pdf\n1. http://arxiv.org/abs/2102.00429\n1. https://arxiv.org/abs/2011.05038\n1. https://arxiv.org/abs/2011.03955\n1. https://gfx.cs.princeton.edu/pubs/Su_2019_PM/Su_2019_enhancement.pdf\n1. https://arxiv.org/abs/2006.05694\n1. https://arxiv.org/pdf/2004.04001.pdf\n1. https://arxiv.org/pdf/2010.11860v1.pdf\n1. https://arxiv.org/abs/2104.03538\n1. https://www.isca-speech.org/archive/pdfs/interspeech_2021/hsieh21_interspeech.pdf\n1. https://arxiv.org/pdf/2006.12847.pdf\n1. https://www.isca-speech.org/archive/pdfs/interspeech_2021/kim21h_interspeech.pdf", "summary_of_the_review": "I think that the paper is not at the level of ICLR standards. Although it could be considered that it achieves interesting results in an existing but rather unexplored task, it contains almost no learnings for the ICLR audience as there seem to be no novel machine learning parts. In this sense, a more specialized venue could be perhaps more appropriate (e.g., InterSpeech, ICASSP, WASPAA). However, in my opinion, both results and writing/presentation also have serious problems that should be solved before resubmitting to such venues. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635332406049}], "openreview_url": "https://openreview.net/forum?id=G-7GlfTneYg", "arxiv_id": "2109.13731", "paper_pdf": "papers/G-7GlfTneYg.pdf", "paper_pdf_sha256": "4b934e4e19c82adad798f4c54fec4282c8584972e63bd3c18134883a867721a1", "paper_pdf_bytes": 18109096, "paper_pdf_source": "openreview", "code_url": "https://github.com/haoheliu/voicefixer", "code_repository": "haoheliu/voicefixer", "code_commit": "aae2253c85f97a87b844b6832384de249c23ab37", "code_archive": "repos/G-7GlfTneYg.zip", "code_archive_sha256": "2d2b90c23a7f767ea796ed71f057066a5309b3c4ed4b120a41d9015a4da048be", "code_archive_bytes": 2290754, "code_file_count": 32, "code_extensions": {".py": 31, ".sh": 1}, "github_disk_usage_kb": 3946, "github_languages": {"Python": 166428, "Dockerfile": 1335, "Shell": 52, "Batchfile": 33}, "github_archived": false, "github_pushed_at": "2025-02-17T14:13:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/voicefixer-toward-general-speech-restoration"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jYkO_0z2TAr", "year": 2021, "status": "rejected", "title": "Zero-Shot Learning with Common Sense Knowledge Graphs", "authors": ["Nihal Nayak", "Stephen Bach"], "authorids": ["~Nihal_Nayak1", "~Stephen_Bach1"], "authors_source": "OpenReview API", "abstract": "Zero-shot learning relies on semantic class representations such as hand-engineered attributes or learned embeddings to predict classes without any labeled examples. We propose to learn class representations from common sense knowledge graphs. Common sense knowledge graphs are an untapped source of explicit high-level knowledge that requires little human effort to apply to a range of tasks. To capture the knowledge in the graph, we introduce ZSL-KG, a general-purpose framework with a novel transformer graph convolutional network (TrGCN) to generate class representations. Our proposed TrGCN architecture computes non-linear combinations of the node neighbourhood and leads to significant improvements on zero-shot learning tasks. We report new state-of-the-art accuracies on six zero-shot benchmark datasets in object classification, intent classification, and fine-grained entity typing tasks. ZSL-KG outperforms the specialized state-of-the-art method for each task by an average 1.7 accuracy points and outperforms the general-purpose method with the best average accuracy by 5.3 points. Our ablation study on ZSL-KG with alternate graph neural networks shows that our transformer-based aggregator adds up to 2.8 accuracy points improvement on these tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "fAaVjc-86D2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper547/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The submission proposed to leverage a commonsense knowledge graph and an attention GNN based model to aggregate the node features on the graph for the problem of zero-shot learning. \n\nThe main concern is the technical contribution is limited:\n\n- The authors mention \"Existing methods for zero-shot object classification such as GCNZ and DGP cannot be\nadapted to common sense knowledge graphs as they do not scale to large graphs or require a directed\nacyclic graph such as WordNet.\" Can the authors elaborate on why GCNZ and DGP can not scale to large knowledge graphs? To my understanding, GCNZ and DGP use WordNet knowledge graphs which are large-scale. And [a] also proposed to leverage GNN structure and knowledge graph, can the authors explain more comparisons with respect to this paper?\n\n- The proposed GAT model is similar to the previous works, i.e., GCNZ, and DGP. To me, the model GNN and attention (Transformer in the paper) model have already been discussed in these two works and the only difference seems to be the knowledge graph, where the previous paper uses the WordNet while the current submission uses the ConceptNet. (I think WordNet is a more straightforward knowledge graph format, can the authors explain why extra efforts on making a ConceptNet is better?) Also, [e] this paper has already proposed to leverage ConcepNet and GNN structures for ZSL, can the authors explain the difference with it please?\n\n- The authors claim the structure of Transformer while actually, it is the structure of dot product attention. The appliance of the attention of ZSL has been a lot, such as [a,b,c,d]. So I think the technical contribution is limited and it is better to also include the discussions with these works.\n\nBased on the previous three points, I believe the technical contribution is limited for this submission.\n\nThe experiments do not support the claims:\n\n- The authors claim the model can scale to large size while in the experiment section,  the results on ImageNet is not reported. The most related two works GCNZ and DPN both reported their results on ImageNet, thus i believe the performance result on this dataset is necessary. \n\n- I am expected to see performance on the generalized zero-shot learning since this setting is more practical and includes the prediction on both seen and unseen classes. The most recent papers including the previous works listed in the experiment table have the results on GZSL, so do the GCNZ and DPN. Thus, I think the results on GZSL are also necessary to support the claims in this submission.\n\n\nWriting:\nEven the basic format is not well organized, such as the tables are out of columns and spelling errors.\n\nRelated works:\nThe current related works section is a little bit precise, and I would recommend the authors to include the related works mentioned in my previous reviews and discuss the difference and how this submission is different from previous works. Since there is plenty of previous literature that is somewhat related to the current submission, a more detailed explanation of the relationships to previous works is recommended.\n\n[a] Attribute Propagation Network for Graph Zero-shot Learning, AAAI 2021\n\n[b] Attentive Region Embedding Network for ZSL, CVPR19.\n\n[c] Semantic-Guided Multi-Attention Localization for ZSL, NIPS19.\n\n[d] Attribute Attention for Semantic Disambiguation in ZSL, ICCV19.\n\n[e] TGG: Transferable Graph Generation for Zero-shot and Few-shot Learning", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Limited technical contribution and experiment results not convincing", "review": "The submission proposed to leverage a commonsense knowledge graph and an attention GNN based model to aggregate the node features on the graph for the problem of zero-shot learning. \n\nThe main concern is the technical contribution is limited:\n\n- The authors mention \"Existing methods for zero-shot object classification such as GCNZ and DGP cannot be\nadapted to common sense knowledge graphs as they do not scale to large graphs or require a directed\nacyclic graph such as WordNet.\" Can the authors elaborate on why GCNZ and DGP can not scale to large knowledge graphs? To my understanding, GCNZ and DGP use WordNet knowledge graphs which are large-scale. And [a] also proposed to leverage GNN structure and knowledge graph, can the authors explain more comparisons with respect to this paper?\n\n- The proposed GAT model is similar to the previous works, i.e., GCNZ, and DGP. To me, the model GNN and attention (Transformer in the paper) model have already been discussed in these two works and the only difference seems to be the knowledge graph, where the previous paper uses the WordNet while the current submission uses the ConceptNet. (I think WordNet is a more straightforward knowledge graph format, can the authors explain why extra efforts on making a ConceptNet is better?) Also, [e] this paper has already proposed to leverage ConcepNet and GNN structures for ZSL, can the authors explain the difference with it please?\n\n- The authors claim the structure of Transformer while actually, it is the structure of dot product attention. The appliance of the attention of ZSL has been a lot, such as [a,b,c,d]. So I think the technical contribution is limited and it is better to also include the discussions with these works.\n\nBased on the previous three points, I believe the technical contribution is limited for this submission.\n\nThe experiments do not support the claims:\n\n- The authors claim the model can scale to large size while in the experiment section,  the results on ImageNet is not reported. The most related two works GCNZ and DPN both reported their results on ImageNet, thus i believe the performance result on this dataset is necessary. \n\n- I am expected to see performance on the generalized zero-shot learning since this setting is more practical and includes the prediction on both seen and unseen classes. The most recent papers including the previous works listed in the experiment table have the results on GZSL, so do the GCNZ and DPN. Thus, I think the results on GZSL are also necessary to support the claims in this submission.\n\n\nWriting:\nEven the basic format is not well organized, such as the tables are out of columns and spelling errors.\n\nRelated works:\nThe current related works section is a little bit precise, and I would recommend the authors to include the related works mentioned in my previous reviews and discuss the difference and how this submission is different from previous works. Since there is plenty of previous literature that is somewhat related to the current submission, a more detailed explanation of the relationships to previous works is recommended.\n\n[a] Attribute Propagation Network for Graph Zero-shot Learning, AAAI 2021\n\n[b] Attentive Region Embedding Network for ZSL, CVPR19.\n\n[c] Semantic-Guided Multi-Attention Localization for ZSL, NIPS19.\n\n[d] Attribute Attention for Semantic Disambiguation in ZSL, ICCV19.\n\n[e] TGG: Transferable Graph Generation for Zero-shot and Few-shot Learning", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604467575640}, {"id": "XC3jVK0gU3Z", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper547/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes to tackle the zero-shot learning problem by learning class representation from commonsense knowledge graphs. To capture knowledge in graphs, the paper proposes to perform Transformer-based aggregation. \n\nPros:\n- The motivation of using commonsense knowledge graphs to enhance zero-short learning is interesting (although it has been explored in the previous work). \n- The paper is in general clearly written and easy to follow.\n\nCons:\n- In term of methodology, the novelty of paper is very limited---it uses Transformer to aggregate knowledge over graphs. This is pretty straightforward, and the previous work (e.g., (Kampffmeyer et al.)) has tried Graph Attention Networks (GAT)-based models for zero-shot learning. Transformer and GAT are very similar, particularly in the setup of this work (e.g., without sequence/position embedding and the graphs are not fully connected). The differences claimed in this paper between GAT-based and Transformer-based models are rather marginal.\n\n- The empirical comparison to GAT-based models in the experiments is not clear enough. For example, Table 4 shows that the proposed model is better than ZSL-KG-GAT, but detailed analyses were not provided to make this clearer to help understand why that is the case (and hence better understand the contributions of this paper). For example, whether the performance differences are related to the use of layer normalization, the two-layer feed-forward nets in equation (4), or other reasons? Previous work suggested DGP is better than GAT-based models (Kampffmeyer et al.) and in this current submission GAT and transformer-based models are better. More discussions will be helpful. \n\n- It will be more helpful if the paper describes more details about the models in comparison, e.g., details about ZSL-KG-GAT such as the setup of multi-heads.\n\nMore comments:\n- The title of the paper may be made more specific, particularly given much work has been done by using commonsense/knowledge graphs for zero-short learning. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novelty of the proposed model is limited and more analyses are needed in experiments", "review": "The paper proposes to tackle the zero-shot learning problem by learning class representation from commonsense knowledge graphs. To capture knowledge in graphs, the paper proposes to perform Transformer-based aggregation. \n\nPros:\n- The motivation of using commonsense knowledge graphs to enhance zero-short learning is interesting (although it has been explored in the previous work). \n- The paper is in general clearly written and easy to follow.\n\nCons:\n- In term of methodology, the novelty of paper is very limited---it uses Transformer to aggregate knowledge over graphs. This is pretty straightforward, and the previous work (e.g., (Kampffmeyer et al.)) has tried Graph Attention Networks (GAT)-based models for zero-shot learning. Transformer and GAT are very similar, particularly in the setup of this work (e.g., without sequence/position embedding and the graphs are not fully connected). The differences claimed in this paper between GAT-based and Transformer-based models are rather marginal.\n\n- The empirical comparison to GAT-based models in the experiments is not clear enough. For example, Table 4 shows that the proposed model is better than ZSL-KG-GAT, but detailed analyses were not provided to make this clearer to help understand why that is the case (and hence better understand the contributions of this paper). For example, whether the performance differences are related to the use of layer normalization, the two-layer feed-forward nets in equation (4), or other reasons? Previous work suggested DGP is better than GAT-based models (Kampffmeyer et al.) and in this current submission GAT and transformer-based models are better. More discussions will be helpful. \n\n- It will be more helpful if the paper describes more details about the models in comparison, e.g., details about ZSL-KG-GAT such as the setup of multi-heads.\n\nMore comments:\n- The title of the paper may be made more specific, particularly given much work has been done by using commonsense/knowledge graphs for zero-short learning. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603949342188}, {"id": "XHtjbdrWFor", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper547/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n\nThis paper tackles zero-shot learning by leveraging the large-scale knowledge graph i.e., ConceptNet to propagate the knowledge learned from seen classes to unseen classes. The authors propose a novel propagation rule that aggregates node embeddings by the self-attention technique. It is infeasible to run GCN on such large-scale knowledge graph. Therefore they reduce the knolwedge graph by adopting a neighborhood sampling strategy based on random walks. The method is evaluated on multiple zero-shot learning tasks including object classification, intent classification and fine-grained entity typing. The SOTA results are achieved.\n\nPros\n\n-This paper is well motivated. The idea of leveraging a large-scale common sense knowledge graph that contains richer semantic relationship is reasonable and novel. I agree that previous GCN methods for zero-shot learning are limited by the wordnet. \n\n-The self-attention or transformer-based aggregator is novel as well. Applying such non-linear aggregation makes graph capture complex relationships between nodes. Therefore, I believe the technical contribution is significant. \n\n-The evaluation is conducted on multiple zero-shot learning tasks and the results show that the proposed approach generalizes  well to different tasks. \n\n-The paper is written well and easy to follow in general, although some important details and justification of the results are missing (I will points out the issues below).  \n\nCons\n\n-I am particularly not convinced by the object classification results (Table 1). This paper only reports results on two small datasets i.e., AWA2 and aPY, in the zero-shot learning setting, while the GCN baselines i.e., GCNZ, SGCN and DGP, are all evaluated on the large-scale ImageNet in both zero-shot and generalized zero-shot learning settings. I strongly suggest the authors to compare with GCNZ, SGCN and DGP on the ImageNet.\n\n-The results on the BBN dataset (Table 3) are not impressive. The proposed approach performs much worse than DZET \nin two metrics i.e., Loose Mic. and Loose Mac. The authors fail to give some justification for these bad results.  \n\n-The neighborhood sampling technique is rarely evaluated. I believe it is interesting to ablate different sampling strategies e.g., uniformly ramdom sampling, importance sampling (Chen et al., 2018a) and random walks. Implementation details for sampling are not provided either which makes it hard to reproduce the results. \n\n-Writing can be further improved in the following perspectives. A discussion on the difference between the common sense knowledge graph and the standard knowledge like WordNet adopted by previous ZSL works should be included. The justification for the bad results on the BBN dataset is missing. There are not enough implementation details in order to reproduce the results e.g., what are the text features used for the intent classification? How many nodes are selected in the graph sampling step?  \n\nMinor questions\n\n-In the last paragraph of Section 4.1, I do not get why other GCN approaches are referred as general-purpose methods? \n\nJustification of the rating\n\n-Overall, I think this paper has significant technical contributions. However, due to the issues I have pointed out above, my initial score is only 6 (marginally above the acceptance threshold.). If the authors can address my concerns, in particular, providing better results on the ImageNet dataset, I will be very happy to upgrade my score.  \n\n----------------------------------------------------\nPost-rebuttal\n\nThanks for the new ZSL and GZSL results on ImageNet. The results are convincing to me. My other concerns are properly addressed too. I read the concerns from R4 and R5. To my knowledge, using common sense knowledge graph and the GCN with self-attention are novel in the zero-shot learning literature. I decide to increase my score to an accept.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "Summary\n\nThis paper tackles zero-shot learning by leveraging the large-scale knowledge graph i.e., ConceptNet to propagate the knowledge learned from seen classes to unseen classes. The authors propose a novel propagation rule that aggregates node embeddings by the self-attention technique. It is infeasible to run GCN on such large-scale knowledge graph. Therefore they reduce the knolwedge graph by adopting a neighborhood sampling strategy based on random walks. The method is evaluated on multiple zero-shot learning tasks including object classification, intent classification and fine-grained entity typing. The SOTA results are achieved.\n\nPros\n\n-This paper is well motivated. The idea of leveraging a large-scale common sense knowledge graph that contains richer semantic relationship is reasonable and novel. I agree that previous GCN methods for zero-shot learning are limited by the wordnet. \n\n-The self-attention or transformer-based aggregator is novel as well. Applying such non-linear aggregation makes graph capture complex relationships between nodes. Therefore, I believe the technical contribution is significant. \n\n-The evaluation is conducted on multiple zero-shot learning tasks and the results show that the proposed approach generalizes  well to different tasks. \n\n-The paper is written well and easy to follow in general, although some important details and justification of the results are missing (I will points out the issues below).  \n\nCons\n\n-I am particularly not convinced by the object classification results (Table 1). This paper only reports results on two small datasets i.e., AWA2 and aPY, in the zero-shot learning setting, while the GCN baselines i.e., GCNZ, SGCN and DGP, are all evaluated on the large-scale ImageNet in both zero-shot and generalized zero-shot learning settings. I strongly suggest the authors to compare with GCNZ, SGCN and DGP on the ImageNet.\n\n-The results on the BBN dataset (Table 3) are not impressive. The proposed approach performs much worse than DZET \nin two metrics i.e., Loose Mic. and Loose Mac. The authors fail to give some justification for these bad results.  \n\n-The neighborhood sampling technique is rarely evaluated. I believe it is interesting to ablate different sampling strategies e.g., uniformly ramdom sampling, importance sampling (Chen et al., 2018a) and random walks. Implementation details for sampling are not provided either which makes it hard to reproduce the results. \n\n-Writing can be further improved in the following perspectives. A discussion on the difference between the common sense knowledge graph and the standard knowledge like WordNet adopted by previous ZSL works should be included. The justification for the bad results on the BBN dataset is missing. There are not enough implementation details in order to reproduce the results e.g., what are the text features used for the intent classification? How many nodes are selected in the graph sampling step?  \n\nMinor questions\n\n-In the last paragraph of Section 4.1, I do not get why other GCN approaches are referred as general-purpose methods? \n\nJustification of the rating\n\n-Overall, I think this paper has significant technical contributions. However, due to the issues I have pointed out above, my initial score is only 6 (marginally above the acceptance threshold.). If the authors can address my concerns, in particular, providing better results on the ImageNet dataset, I will be very happy to upgrade my score.  \n\n----------------------------------------------------\nPost-rebuttal\n\nThanks for the new ZSL and GZSL results on ImageNet. The results are convincing to me. My other concerns are properly addressed too. I read the concerns from R4 and R5. To my knowledge, using common sense knowledge graph and the GCN with self-attention are novel in the zero-shot learning literature. I decide to increase my score to an accept.", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603893625085}], "openreview_url": "https://openreview.net/forum?id=jYkO_0z2TAr", "arxiv_id": "2006.10713", "paper_pdf": "papers/jYkO_0z2TAr.pdf", "paper_pdf_sha256": "4e858af206ee3c02b4724cc720e1600cfdbad41c109938ba1b32dafd77d88ae4", "paper_pdf_bytes": 490474, "paper_pdf_source": "openreview", "code_url": "https://github.com/BatsResearch/nayak-tmlr22-code", "code_repository": "BatsResearch/nayak-tmlr22-code", "code_commit": "fd86c1d2dfb2dd540fd744ce8c49bf58719abadf", "code_archive": "repos/jYkO_0z2TAr.zip", "code_archive_sha256": "3cabbebf859e82a84ee72c4ba043d96edfca5aacfbc7698b95174201b9e6b70a", "code_archive_bytes": 2280638, "code_file_count": 55, "code_extensions": {".py": 55}, "github_disk_usage_kb": 2212, "github_languages": {"Python": 302147}, "github_archived": false, "github_pushed_at": "2022-07-25T11:24:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/zero-shot-learning-with-common-sense"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3DZeEUTwhq", "year": 2026, "status": "rejected", "title": "Are You Getting What You Pay For? Auditing Model Substitution in LLM APIs", "authors": ["Will Cai", "Tianneng Shi", "Xuandong Zhao", "Dawn Song"], "authorids": ["~Will_Cai1", "~Tianneng_Shi1", "~Xuandong_Zhao1", "~Dawn_Song1"], "authors_source": "OpenReview API", "abstract": "Commercial Large Language Model (LLM) APIs create a fundamental trust problem: users pay for specific models but have no guarantee that providers deliver them faithfully. Providers may covertly substitute cheaper alternatives (e.g., quantized versions, smaller models) to reduce costs while maintaining advertised pricing. We formalize this model substitution problem and systematically evaluate detection methods under realistic adversarial conditions. Our empirical analysis reveals that software-only methods are fundamentally unreliable: statistical tests on text outputs are query-intensive and fail against subtle substitutions, while methods using log probabilities are defeated by inherent inference nondeterminism in production environments. We argue that this verification gap can be more effectively closed with hardware-level security. We propose and evaluate the use of Trusted Execution Environments (TEEs) as one practical and robust solution. Our findings demonstrate that TEEs can provide provable cryptographic guarantees of model integrity with only a modest performance overhead, offering a clear and actionable path to ensure users get what they pay for.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Y0j8msNnpC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7360/Reviewer_eP5M"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper studies the problem of auditing model substitution in LLM APIs. When users use an API provider to generate text from an LLM, they may want to check that they are being served the correct model that they are paying for. API providers may have an incentive to secretly use a different model, e.g., a smaller model to save on compute costs.\n\nThis paper runs experiments to claim that existing software-based auditing methods are unreliable in adversarial scenarios, such as quantization and randomized substitution. Therefore, this paper proposes trusted execution environments (TEEs) as a solution, where hardware enclaves provide cryptographic attestation reports of correct model and code execution.", "review_text": "This paper studies the problem of auditing model substitution in LLM APIs. When users use an API provider to generate text from an LLM, they may want to check that they are being served the correct model that they are paying for. API providers may have an incentive to secretly use a different model, e.g., a smaller model to save on compute costs.\n\nThis paper runs experiments to claim that existing software-based auditing methods are unreliable in adversarial scenarios, such as quantization and randomized substitution. Therefore, this paper proposes trusted execution environments (TEEs) as a solution, where hardware enclaves provide cryptographic attestation reports of correct model and code execution.", "strengths": "The problem of auditing model substitution is well-motivated and important given the widespread usage of LLM APIs today. The introduction clearly and succinctly states the problem setting. The proposed use of TEEs for auditing model substitution appears to be a novel idea.", "weaknesses": "I don’t think that the paper has substantial enough contributions for what is typically expected in a publication. The majority of the paper consists of simple evaluations of existing methods on adversarial settings, such as subtle differences due to quantization. Note that several of these existing methods were originally designed for problems other than auditing model substitution. For example, [Sun et al. (2025)](https://arxiv.org/abs/2502.12150) studies the problem of predicting which model generated a particular text, e.g., ChatGPT vs Claude.\n\n[Gao et al. (2024)](https://arxiv.org/abs/2410.20247) studies the same problem of auditing model substitution as the submitted paper, and proposes and evaluates principled, rigorous statistical tests (MMD) for doing so. Gao et al. (2024) showed strong results in controlled, simulated settings, and also ran audits on real-world LLM APIs. I am not convinced by the submitted paper’s claim that MMD is an unreliable and impractical audit method.\n\nThe submitted paper runs a small experiment claiming that MMD achieves low power for detecting quantization of small models and randomized substitution. However, the paper does not report how many completions are sampled for each prompt when running MMD, which is a crucial factor impacting the power, as found in Gao et al. (2024). I suspect that given more samples, MMD would achieve higher power.\n\nAnother recent arXiv preprint that proposes an audit for model substitution is [Zhu et al. (2025)](https://arxiv.org/abs/2506.06975), although I do not expect a detailed experimental comparison since it is a preprint and recent (June 2025).\n\nThe proposed TEE method does not seem practical for deployment in real-world APIs, given the complexity of modern LLM inference infrastructures. Larger models are split across many GPUs, and complex routing strategies are employed due to prompt caching, mixture of experts, etc. In addition, the proposed TEE method requires the open-sourcing of inference code, which would likely be unacceptable for both proprietary model providers and providers of open-source models. Such code would reveal information about model architecture, inference optimizations and tricks, etc.\n\nNote that TEEs require API providers to make significant implementation changes, whereas audits such as MMD can be run by users on any black-box LLM API, without needing the API provider to agree to it. So TEEs are not a comparable replacement for black-box audits.", "questions": "### Main questions\n\n1. Line 141: shouldn’t this be $p \\\\to 0$ for a low substitution rate? Then it should be $p$ upward on line 143\\.  \n2. Line 246: doesn’t the prompt have to appear in the MMD equation? It only makes sense to compare completions that came from the same prompt.  \n   * It is confusing to use $x$ to denote completions here, as $x$ was used for prompts and $y$ for completions earlier in the paper.  \n   * The model distribution was only defined as the conditional distribution $P(y \\\\mid x)$ earlier in the paper, so it is not clear what exactly drawing $x \\\\sim P$ means here.  \n3. Figure 1 (left): why are there two bars for original and quantized? Comparing the original and quantized models is one test, so why isn’t there just one value for the power?  \n4. Line 253: how many completions are sampled for each prompt in the test? Gao et al. (2024) found that the power is sometimes low when few samples are used, but high when more samples are used.  \n5. Line 269: where does Gao et al. (2024) state that *“MMD-based auditing is only effective under strictly controlled local inference environments, limiting its practicality”*? Gao et al. (2024) run audits on real-world API providers, so I’m not sure where this is coming from.  \n6. Line 317: the size of the drop in accuracy for higher temperatures seems significantly higher than the drop from quantization. The drop from higher temperature is around 5–10 percentage points, whereas the difference from quantization is usually around 1–2 percentage points.  \n7. Can’t the effect of higher temperature be mitigated by taking the most common answer across queries, instead of averaging? The relative probabilities between answers should not change in higher temperatures.  \n8. Line 350 states that both $k \\= 20$ and $k \\= 100$ is used, but only one agreement rate is reported. Which value of $k$ is used?  \n9. In section 4.1.5, how different are the greedy decoding outputs when they do not match? Are they completely different, or just off by one or two tokens?  \n   * What is the agreement rate between providers, i.e., comparing Together vs. OpenRouter instead of Together vs. local baseline?  \n   * Are the greedy outputs deterministic if you query the provider with the same prompt multiple times, or do they vary?  \n10. Line 410: this definition of LSH is contradictory, how can “similar activations yield similar hashes” while also “small changes cause detectable differences”? LSH is designed so that similar items receive the same hash with probability, so the second part seems to be incorrect.\n\n### Minor questions\n\n1. Line 246: why does MMD have the squared superscript? It does not appear with the square elsewhere in the paper.  \n2. Line 252: what is $T$? I assume that $L$ is the completion length, but this is also not explicitly stated.  \n3. Line 252 states that Llama 70B is used, but Figure 1 (left) shows only Llama 8B and Mistral 7B.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of auditing model substitution in LLM APIs. When users use an API provider to generate text from an LLM, they may want to check that they are being served the correct model that they are paying for. API providers may have an incentive to secretly use a different model, e.g., a smaller model to save on compute costs.\n\nThis paper runs experiments to claim that existing software-based auditing methods are unreliable in adversarial scenarios, such as quantization and randomized substitution. Therefore, this paper proposes trusted execution environments (TEEs) as a solution, where hardware enclaves provide cryptographic attestation reports of correct model and code execution.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "The problem of auditing model substitution is well-motivated and important given the widespread usage of LLM APIs today. The introduction clearly and succinctly states the problem setting. The proposed use of TEEs for auditing model substitution appears to be a novel idea.", "weaknesses": "I don’t think that the paper has substantial enough contributions for what is typically expected in a publication. The majority of the paper consists of simple evaluations of existing methods on adversarial settings, such as subtle differences due to quantization. Note that several of these existing methods were originally designed for problems other than auditing model substitution. For example, [Sun et al. (2025)](https://arxiv.org/abs/2502.12150) studies the problem of predicting which model generated a particular text, e.g., ChatGPT vs Claude.\n\n[Gao et al. (2024)](https://arxiv.org/abs/2410.20247) studies the same problem of auditing model substitution as the submitted paper, and proposes and evaluates principled, rigorous statistical tests (MMD) for doing so. Gao et al. (2024) showed strong results in controlled, simulated settings, and also ran audits on real-world LLM APIs. I am not convinced by the submitted paper’s claim that MMD is an unreliable and impractical audit method.\n\nThe submitted paper runs a small experiment claiming that MMD achieves low power for detecting quantization of small models and randomized substitution. However, the paper does not report how many completions are sampled for each prompt when running MMD, which is a crucial factor impacting the power, as found in Gao et al. (2024). I suspect that given more samples, MMD would achieve higher power.\n\nAnother recent arXiv preprint that proposes an audit for model substitution is [Zhu et al. (2025)](https://arxiv.org/abs/2506.06975), although I do not expect a detailed experimental comparison since it is a preprint and recent (June 2025).\n\nThe proposed TEE method does not seem practical for deployment in real-world APIs, given the complexity of modern LLM inference infrastructures. Larger models are split across many GPUs, and complex routing strategies are employed due to prompt caching, mixture of experts, etc. In addition, the proposed TEE method requires the open-sourcing of inference code, which would likely be unacceptable for both proprietary model providers and providers of open-source models. Such code would reveal information about model architecture, inference optimizations and tricks, etc.\n\nNote that TEEs require API providers to make significant implementation changes, whereas audits such as MMD can be run by users on any black-box LLM API, without needing the API provider to agree to it. So TEEs are not a comparable replacement for black-box audits.", "questions": "### Main questions\n\n1. Line 141: shouldn’t this be $p \\\\to 0$ for a low substitution rate? Then it should be $p$ upward on line 143\\.  \n2. Line 246: doesn’t the prompt have to appear in the MMD equation? It only makes sense to compare completions that came from the same prompt.  \n   * It is confusing to use $x$ to denote completions here, as $x$ was used for prompts and $y$ for completions earlier in the paper.  \n   * The model distribution was only defined as the conditional distribution $P(y \\\\mid x)$ earlier in the paper, so it is not clear what exactly drawing $x \\\\sim P$ means here.  \n3. Figure 1 (left): why are there two bars for original and quantized? Comparing the original and quantized models is one test, so why isn’t there just one value for the power?  \n4. Line 253: how many completions are sampled for each prompt in the test? Gao et al. (2024) found that the power is sometimes low when few samples are used, but high when more samples are used.  \n5. Line 269: where does Gao et al. (2024) state that *“MMD-based auditing is only effective under strictly controlled local inference environments, limiting its practicality”*? Gao et al. (2024) run audits on real-world API providers, so I’m not sure where this is coming from.  \n6. Line 317: the size of the drop in accuracy for higher temperatures seems significantly higher than the drop from quantization. The drop from higher temperature is around 5–10 percentage points, whereas the difference from quantization is usually around 1–2 percentage points.  \n7. Can’t the effect of higher temperature be mitigated by taking the most common answer across queries, instead of averaging? The relative probabilities between answers should not change in higher temperatures.  \n8. Line 350 states that both $k \\= 20$ and $k \\= 100$ is used, but only one agreement rate is reported. Which value of $k$ is used?  \n9. In section 4.1.5, how different are the greedy decoding outputs when they do not match? Are they completely different, or just off by one or two tokens?  \n   * What is the agreement rate between providers, i.e., comparing Together vs. OpenRouter instead of Together vs. local baseline?  \n   * Are the greedy outputs deterministic if you query the provider with the same prompt multiple times, or do they vary?  \n10. Line 410: this definition of LSH is contradictory, how can “similar activations yield similar hashes” while also “small changes cause detectable differences”? LSH is designed so that similar items receive the same hash with probability, so the second part seems to be incorrect.\n\n### Minor questions\n\n1. Line 246: why does MMD have the squared superscript? It does not appear with the square elsewhere in the paper.  \n2. Line 252: what is $T$? I assume that $L$ is the completion length, but this is also not explicitly stated.  \n3. Line 252 states that Llama 70B is used, but Figure 1 (left) shows only Llama 8B and Mistral 7B.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761712626113}, {"id": "02fQESGQpg", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7360/Reviewer_4MTF"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper addresses the critical trust problem in commercial LLM APIs where users pay for specific models but cannot verify if providers are faithfully serving them. The authors formalize the model substitution detection problem, systematically evaluate existing software-based detection methods under adversarial conditions, and propose Trusted Execution Environments (TEEs) as a robust hardware-based solution. Through comprehensive empirical analysis across multiple detection techniques (text classifiers, identity prompting, statistical tests, log probability verification), they demonstrate that software-only approaches are fundamentally unreliable due to inference nondeterminism and subtle substitution strategies, while TEEs provide cryptographic guarantees with modest performance overhead (2.88-16.18%).", "review_text": "This paper addresses the critical trust problem in commercial LLM APIs where users pay for specific models but cannot verify if providers are faithfully serving them. The authors formalize the model substitution detection problem, systematically evaluate existing software-based detection methods under adversarial conditions, and propose Trusted Execution Environments (TEEs) as a robust hardware-based solution. Through comprehensive empirical analysis across multiple detection techniques (text classifiers, identity prompting, statistical tests, log probability verification), they demonstrate that software-only approaches are fundamentally unreliable due to inference nondeterminism and subtle substitution strategies, while TEEs provide cryptographic guarantees with modest performance overhead (2.88-16.18%).", "strengths": "This paper makes several important contributions to addressing the critical trust problem in commercial LLM APIs. Most notably, it tackles a timely and economically significant issue where providers have strong incentives to substitute cheaper models while maintaining premium pricing, creating a fundamental trust gap in the current ecosystem. The authors provide a comprehensive and systematic evaluation of detection methods that goes well beyond previous work by considering realistic adversarial scenarios including quantization substitution, randomized model mixing, and benchmark evasion attacks. Their empirical analysis reveals crucial practical insights, particularly the surprising finding that production-level inference nondeterminism—arising from factors like batch variance and heterogeneous backends—can defeat seemingly robust detection methods such as log probability verification, even when these methods work well in controlled laboratory settings.\n\nThe paper's technical approach demonstrates strong rigor through its formal problem definition using a hypothesis testing framework (H₀: Pactual = Pspec vs H₁: Pactual ≠ Pspec) that properly grounds the empirical evaluation. The experimental methodology is sound, employing appropriate statistical tests like Maximum Mean Discrepancy with permutation testing, and the results are clearly presented through effective visualizations that illustrate key findings such as the relationship between substitution probability and detection power. Perhaps most importantly, the paper goes beyond merely identifying problems to propose and validate a practical solution through Trusted Execution Environments (TEEs). The authors demonstrate that TEEs can provide cryptographic guarantees of model integrity with surprisingly modest performance overhead—only 2.88% throughput reduction under high concurrency with 64 concurrent requests—making this a genuinely deployable solution for the industry. This combination of thorough problem analysis, systematic evaluation under adversarial conditions, practical insights about real-world deployment challenges, and an actionable solution with empirical validation makes this work a valuable contribution that directly addresses pressing concerns in the rapidly evolving LLM API ecosystem.", "weaknesses": "1. Limited TEE Evaluation: Only evaluates one model (Llama-3-8B) on one GPU (H100). Lacks testing on larger models (70B+) where memory constraints and multi-GPU setups could significantly impact feasibility and overhead. Since TEEs are proposed as the primary solution, this limited evaluation scope severely constrains confidence in the solution's general applicability.\n\n2. Limited Real-World Validation: All experiments use controlled settings with known models and substitutions. The paper lacks evaluation with actual commercial APIs or evidence of real-world substitution detection. Without validation against production systems or case studies of actual substitution attempts, the practical applicability of both detection methods and the TEE solution remains uncertain.\n\n3. Narrow Attack Scenarios: Focuses primarily on basic quantization and randomized substitution. Doesn't explore sophisticated adversarial strategies like adaptive substitution based on query patterns, gradual model degradation, or hybrid evasion techniques. This limits understanding of how robust the detection methods and TEE solution would be against determined adversaries with economic incentives to evade detection.", "questions": "1. How would the TEE solution scale to larger models (175B+ parameters) that require multi-GPU setups? What are the technical challenges and expected overheads?\n\n2. Have you considered hybrid approaches combining multiple detection methods (e.g., periodic TEE attestation with continuous statistical monitoring)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the critical trust problem in commercial LLM APIs where users pay for specific models but cannot verify if providers are faithfully serving them. The authors formalize the model substitution detection problem, systematically evaluate existing software-based detection methods under adversarial conditions, and propose Trusted Execution Environments (TEEs) as a robust hardware-based solution. Through comprehensive empirical analysis across multiple detection techniques (text classifiers, identity prompting, statistical tests, log probability verification), they demonstrate that software-only approaches are fundamentally unreliable due to inference nondeterminism and subtle substitution strategies, while TEEs provide cryptographic guarantees with modest performance overhead (2.88-16.18%).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "This paper makes several important contributions to addressing the critical trust problem in commercial LLM APIs. Most notably, it tackles a timely and economically significant issue where providers have strong incentives to substitute cheaper models while maintaining premium pricing, creating a fundamental trust gap in the current ecosystem. The authors provide a comprehensive and systematic evaluation of detection methods that goes well beyond previous work by considering realistic adversarial scenarios including quantization substitution, randomized model mixing, and benchmark evasion attacks. Their empirical analysis reveals crucial practical insights, particularly the surprising finding that production-level inference nondeterminism—arising from factors like batch variance and heterogeneous backends—can defeat seemingly robust detection methods such as log probability verification, even when these methods work well in controlled laboratory settings.\n\nThe paper's technical approach demonstrates strong rigor through its formal problem definition using a hypothesis testing framework (H₀: Pactual = Pspec vs H₁: Pactual ≠ Pspec) that properly grounds the empirical evaluation. The experimental methodology is sound, employing appropriate statistical tests like Maximum Mean Discrepancy with permutation testing, and the results are clearly presented through effective visualizations that illustrate key findings such as the relationship between substitution probability and detection power. Perhaps most importantly, the paper goes beyond merely identifying problems to propose and validate a practical solution through Trusted Execution Environments (TEEs). The authors demonstrate that TEEs can provide cryptographic guarantees of model integrity with surprisingly modest performance overhead—only 2.88% throughput reduction under high concurrency with 64 concurrent requests—making this a genuinely deployable solution for the industry. This combination of thorough problem analysis, systematic evaluation under adversarial conditions, practical insights about real-world deployment challenges, and an actionable solution with empirical validation makes this work a valuable contribution that directly addresses pressing concerns in the rapidly evolving LLM API ecosystem.", "weaknesses": "1. Limited TEE Evaluation: Only evaluates one model (Llama-3-8B) on one GPU (H100). Lacks testing on larger models (70B+) where memory constraints and multi-GPU setups could significantly impact feasibility and overhead. Since TEEs are proposed as the primary solution, this limited evaluation scope severely constrains confidence in the solution's general applicability.\n\n2. Limited Real-World Validation: All experiments use controlled settings with known models and substitutions. The paper lacks evaluation with actual commercial APIs or evidence of real-world substitution detection. Without validation against production systems or case studies of actual substitution attempts, the practical applicability of both detection methods and the TEE solution remains uncertain.\n\n3. Narrow Attack Scenarios: Focuses primarily on basic quantization and randomized substitution. Doesn't explore sophisticated adversarial strategies like adaptive substitution based on query patterns, gradual model degradation, or hybrid evasion techniques. This limits understanding of how robust the detection methods and TEE solution would be against determined adversaries with economic incentives to evade detection.", "questions": "1. How would the TEE solution scale to larger models (175B+ parameters) that require multi-GPU setups? What are the technical challenges and expected overheads?\n\n2. Have you considered hybrid approaches combining multiple detection methods (e.g., periodic TEE attestation with continuous statistical monitoring)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761610859718}, {"id": "mLVTzA3QX5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7360/Reviewer_gV7d"], "rating": 2, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper studies the problem of auditing model substitutions in LLM APIs. It formalizes the hypothesis test and evaluate a series of techniques, including text-only tests, log-probability-based tests, activation-based tests. They find these methods all fail under at least one type of adversarial substitution attack, and they claim that trusted execution environments provide the only robust guarantee for solving the problem.", "review_text": "This paper studies the problem of auditing model substitutions in LLM APIs. It formalizes the hypothesis test and evaluate a series of techniques, including text-only tests, log-probability-based tests, activation-based tests. They find these methods all fail under at least one type of adversarial substitution attack, and they claim that trusted execution environments provide the only robust guarantee for solving the problem.", "strengths": "1. Important problem and clear motivation: The problem they study is timely and important, given the widespread use of blackbox LLM APIs. \n2. Comprehensive empirical experiments: the paper systematically tests several verification strategies with different types of substitution attacks. The negative results are informative for future work.", "weaknesses": "### Major weaknesses\n\n1. Definition of the auditing goal is too strong: the null hypothesis $H_0$ requires equality for all  (end of Section 3.1), which seems unattainable in realistic settings; the paper itself later also attributes failures of some methods to \"production-level inference nondeterminism.\" Relaxed formulation of the hypotheses would be more coherent. However, with relaxed formulation, some substitution methods may no longer make sense, as discussed in the next point.\n2. Quantization attack framing: the paper shows failure of most methods with quantization attacks. However, it is unclear from the text whether existing providers claim the dtype of their models, which makes \"quantization substitution\" potentially indistinguishable from allowed, valid, and legal backend optimization. Moreover, because benchmark performance deltas under quantization are small (as shown in Table 3), escaping detection may be both easy and practically operationally benign. More clarification on the motivation for this substitution method is needed. For example, if replacing model A with B improves provider-side cost-effectiveness but does not hurt user experience in any empirically observed scenario, do the authors still aim to detect this substitution? Also, it would be beneficial to consider the \"severity\" of constitution by how much they impact model performance, and then consider evaluation metrics for detection methods that take this \"severity\" into account.\n3. TEE section lack details and depth: the paper claims that TEE is \"the only currently viable mechanism\" and lists this as one of their main contribution. However, TEE is only discussed at the very end of the paper with a short subsection. What exactly is attested with TEE? Does this include model weights and inference hyperparameters? How is the trust in this attestation verified by end-user? When I read up to the end of section one, I thought TEE as an auditing method would be a significant part of the main text.\n4. Novelty is limited: the paper feels more like a position or survey paper. Much of section 4 implements known detecting techniques. If I understand it correctly, the main new contribution (as claimed by the authors) is the argument for using the existing TEE as a auditing method and the overhead measurement. This is valuable engineering evaluation but seems methodologically thin for ICLR.\n\n### Minor issues\n1. Adversary model is too narrow: the randomized substitution attack mixes models uniformly randomly, but strategic provider would very likely condition their substitution decisions on certain features of the request. The paper does not evaluate adaptive substitution. which is arguably the more realistic case.\n2. Benchmark-based detection seems brittle: it seems that the adversary can simply return fixed outputs for all known benchmark queries.", "questions": "## Questions\nMany questions are already discussed in the weaknesses section. Some selected/additional questions are:\n1. Would you consider relaxing the null hypothesis? If so, how? How does this relaxation change your conclusions?\n2. How do you propose distinguishing \"benign\" quantization (operationally equivalent for users) from substitutions that meaningfully violate a service agreement?\n3. Is the production-level deployment of TEE practical? What are some potential blockers? How can a user verify that every request was served by the TEE?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of auditing model substitutions in LLM APIs. It formalizes the hypothesis test and evaluate a series of techniques, including text-only tests, log-probability-based tests, activation-based tests. They find these methods all fail under at least one type of adversarial substitution attack, and they claim that trusted execution environments provide the only robust guarantee for solving the problem.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. Important problem and clear motivation: The problem they study is timely and important, given the widespread use of blackbox LLM APIs. \n2. Comprehensive empirical experiments: the paper systematically tests several verification strategies with different types of substitution attacks. The negative results are informative for future work.", "weaknesses": "### Major weaknesses\n\n1. Definition of the auditing goal is too strong: the null hypothesis $H_0$ requires equality for all  (end of Section 3.1), which seems unattainable in realistic settings; the paper itself later also attributes failures of some methods to \"production-level inference nondeterminism.\" Relaxed formulation of the hypotheses would be more coherent. However, with relaxed formulation, some substitution methods may no longer make sense, as discussed in the next point.\n2. Quantization attack framing: the paper shows failure of most methods with quantization attacks. However, it is unclear from the text whether existing providers claim the dtype of their models, which makes \"quantization substitution\" potentially indistinguishable from allowed, valid, and legal backend optimization. Moreover, because benchmark performance deltas under quantization are small (as shown in Table 3), escaping detection may be both easy and practically operationally benign. More clarification on the motivation for this substitution method is needed. For example, if replacing model A with B improves provider-side cost-effectiveness but does not hurt user experience in any empirically observed scenario, do the authors still aim to detect this substitution? Also, it would be beneficial to consider the \"severity\" of constitution by how much they impact model performance, and then consider evaluation metrics for detection methods that take this \"severity\" into account.\n3. TEE section lack details and depth: the paper claims that TEE is \"the only currently viable mechanism\" and lists this as one of their main contribution. However, TEE is only discussed at the very end of the paper with a short subsection. What exactly is attested with TEE? Does this include model weights and inference hyperparameters? How is the trust in this attestation verified by end-user? When I read up to the end of section one, I thought TEE as an auditing method would be a significant part of the main text.\n4. Novelty is limited: the paper feels more like a position or survey paper. Much of section 4 implements known detecting techniques. If I understand it correctly, the main new contribution (as claimed by the authors) is the argument for using the existing TEE as a auditing method and the overhead measurement. This is valuable engineering evaluation but seems methodologically thin for ICLR.\n\n### Minor issues\n1. Adversary model is too narrow: the randomized substitution attack mixes models uniformly randomly, but strategic provider would very likely condition their substitution decisions on certain features of the request. The paper does not evaluate adaptive substitution. which is arguably the more realistic case.\n2. Benchmark-based detection seems brittle: it seems that the adversary can simply return fixed outputs for all known benchmark queries.", "questions": "## Questions\nMany questions are already discussed in the weaknesses section. Some selected/additional questions are:\n1. Would you consider relaxing the null hypothesis? If so, how? How does this relaxation change your conclusions?\n2. How do you propose distinguishing \"benign\" quantization (operationally equivalent for users) from substitutions that meaningfully violate a service agreement?\n3. Is the production-level deployment of TEE practical? What are some potential blockers? How can a user verify that every request was served by the TEE?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761540420964}, {"id": "8pgQZKpj4M", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission7360/Reviewer_pbz9"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper addresses the critical issue of \\emph{model substitution in commercial LLM APIs}---cases where service providers covertly replace an advertised model with a cheaper or quantized variant. The authors (1) formalize the model substitution problem and threat model, (2) evaluate existing software-only auditing methods such as text classifiers, MMD-based statistical tests, benchmark probing, and log-probability comparisons, and (3) propose \\emph{Trusted Execution Environments (TEEs)} as a cryptographically secure and deployable solution. Through extensive experiments, they demonstrate that software-level approaches are unreliable in realistic, non-deterministic production settings, whereas TEEs provide provable model integrity guarantees with modest overhead (≈9–16\\% latency). The paper concludes that TEEs are currently the most practical solution to ensure users “get what they pay for.”", "review_text": "This paper addresses the critical issue of \\emph{model substitution in commercial LLM APIs}---cases where service providers covertly replace an advertised model with a cheaper or quantized variant. The authors (1) formalize the model substitution problem and threat model, (2) evaluate existing software-only auditing methods such as text classifiers, MMD-based statistical tests, benchmark probing, and log-probability comparisons, and (3) propose \\emph{Trusted Execution Environments (TEEs)} as a cryptographically secure and deployable solution. Through extensive experiments, they demonstrate that software-level approaches are unreliable in realistic, non-deterministic production settings, whereas TEEs provide provable model integrity guarantees with modest overhead (≈9–16\\% latency). The paper concludes that TEEs are currently the most practical solution to ensure users “get what they pay for.”", "strengths": "- **Timely and relevant:** The paper targets a real trust gap in today’s LLM API ecosystem: users are billed for “Model A” but may actually get a quantized or smaller substitute.\n- **Thorough evaluation:** The paper tests multiple auditing approaches under realistic adversarial settings:\n  - text classifiers fail to distinguish full-precision vs. quantized variants (≈50% accuracy, i.e. random);\n  - MMD tests lose power when the provider routes only part of traffic to the substitute;\n  - benchmark-based auditing is confounded by hidden decoding parameters and caching;\n  - greedy decoding and logprob comparison are broken by inference nondeterminism across stacks and batching.\n- **Clear negative result:** It convincingly argues that software-only auditing is fundamentally brittle in production.\n- **Actionable answer:** TEEs are proposed as a concrete, deployable alternative. The paper reports only ~9–16% first-token latency overhead and small throughput impact, which supports feasibility.\n- **Responsible positioning:** The authors discuss implications for users, providers, and policymakers, not just measurements.", "weaknesses": "- **Novelty is mostly integrative:** The paper does not introduce a new detection algorithm or new attestation primitive. The contribution is primarily empirical and architectural.\n- **Limited attack surface coverage:** The evaluation focuses on quantization, routing/mixing, and prompt caching. Other realistic substitutions (light fine-tuning on domain data, pruning, speculative decoding with a small draft model, etc.) are mentioned but not deeply quantified.\n- **Security depth of TEEs:** The paper treats TEEs as “the answer,” but does not analyze:\n  - side-channel leakage,\n  - malicious firmware / compromised root of trust,\n  - dishonest attestation endpoints in multi-tenant datacenters.\n  These are important if we’re going to claim TEEs “solve” the problem.\n- **Scalability and ops questions:** The TEE measurements are on a single GPU. It’s unclear how this extends to multi-GPU inference, pipeline parallelism, or high-throughput commercial routing.\n- **Reproducibility / accessibility:** Verifying the TEE claims requires specialized hardware (confidential H100 / Blackwell stacks). That limits community validation.", "questions": "1. How does your auditing approach behave under *partial* substitution (e.g. 10–20% of requests silently routed to a cheaper model) when the attacker also slightly perturbs temperature or top-p to inject noise?\n2. Can the proposed TEE-based workflow support multi-GPU inference or model sharding? If different GPU enclaves each serve a shard, how is attestation composed?\n3. Could lightweight cryptographic proofs (e.g. ZKPs on selective layers, or TopLoc-style activation hashes) be combined with TEEs to reduce trust in the hardware vendor?\n4. For open-source models that customers can run locally: do you see a practical “self-audit mode” that does not require TEEs?\n5. Could the authors clarify the quantization settings used (e.g., INT8 vs FP8 schemes, calibration methods) for Table 2 and Figure 1?  \n   Were the same prompts and decoding parameters kept constant across quantized and full-precision variants?\n\n6. In Figure 1, how sensitive are the MMD results to kernel choice and sample size?  \n   Did you observe any instability in power estimates under different prompt distributions?\n\n7. For Table 4, how many runs were averaged to compute the reported latency and throughput overheads, and were warm-up tokens or network effects included in those measurements?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the critical issue of \\emph{model substitution in commercial LLM APIs}---cases where service providers covertly replace an advertised model with a cheaper or quantized variant. The authors (1) formalize the model substitution problem and threat model, (2) evaluate existing software-only auditing methods such as text classifiers, MMD-based statistical tests, benchmark probing, and log-probability comparisons, and (3) propose \\emph{Trusted Execution Environments (TEEs)} as a cryptographically secure and deployable solution. Through extensive experiments, they demonstrate that software-level approaches are unreliable in realistic, non-deterministic production settings, whereas TEEs provide provable model integrity guarantees with modest overhead (≈9–16\\% latency). The paper concludes that TEEs are currently the most practical solution to ensure users “get what they pay for.”", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- **Timely and relevant:** The paper targets a real trust gap in today’s LLM API ecosystem: users are billed for “Model A” but may actually get a quantized or smaller substitute.\n- **Thorough evaluation:** The paper tests multiple auditing approaches under realistic adversarial settings:\n  - text classifiers fail to distinguish full-precision vs. quantized variants (≈50% accuracy, i.e. random);\n  - MMD tests lose power when the provider routes only part of traffic to the substitute;\n  - benchmark-based auditing is confounded by hidden decoding parameters and caching;\n  - greedy decoding and logprob comparison are broken by inference nondeterminism across stacks and batching.\n- **Clear negative result:** It convincingly argues that software-only auditing is fundamentally brittle in production.\n- **Actionable answer:** TEEs are proposed as a concrete, deployable alternative. The paper reports only ~9–16% first-token latency overhead and small throughput impact, which supports feasibility.\n- **Responsible positioning:** The authors discuss implications for users, providers, and policymakers, not just measurements.", "weaknesses": "- **Novelty is mostly integrative:** The paper does not introduce a new detection algorithm or new attestation primitive. The contribution is primarily empirical and architectural.\n- **Limited attack surface coverage:** The evaluation focuses on quantization, routing/mixing, and prompt caching. Other realistic substitutions (light fine-tuning on domain data, pruning, speculative decoding with a small draft model, etc.) are mentioned but not deeply quantified.\n- **Security depth of TEEs:** The paper treats TEEs as “the answer,” but does not analyze:\n  - side-channel leakage,\n  - malicious firmware / compromised root of trust,\n  - dishonest attestation endpoints in multi-tenant datacenters.\n  These are important if we’re going to claim TEEs “solve” the problem.\n- **Scalability and ops questions:** The TEE measurements are on a single GPU. It’s unclear how this extends to multi-GPU inference, pipeline parallelism, or high-throughput commercial routing.\n- **Reproducibility / accessibility:** Verifying the TEE claims requires specialized hardware (confidential H100 / Blackwell stacks). That limits community validation.", "questions": "1. How does your auditing approach behave under *partial* substitution (e.g. 10–20% of requests silently routed to a cheaper model) when the attacker also slightly perturbs temperature or top-p to inject noise?\n2. Can the proposed TEE-based workflow support multi-GPU inference or model sharding? If different GPU enclaves each serve a shard, how is attestation composed?\n3. Could lightweight cryptographic proofs (e.g. ZKPs on selective layers, or TopLoc-style activation hashes) be combined with TEEs to reduce trust in the hardware vendor?\n4. For open-source models that customers can run locally: do you see a practical “self-audit mode” that does not require TEEs?\n5. Could the authors clarify the quantization settings used (e.g., INT8 vs FP8 schemes, calibration methods) for Table 2 and Figure 1?  \n   Were the same prompts and decoding parameters kept constant across quantized and full-precision variants?\n\n6. In Figure 1, how sensitive are the MMD results to kernel choice and sample size?  \n   Did you observe any instability in power estimates under different prompt distributions?\n\n7. For Table 4, how many runs were averaged to compute the reported latency and throughput overheads, and were warm-up tokens or network effects included in those measurements?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761333221371}], "openreview_url": "https://openreview.net/forum?id=3DZeEUTwhq", "arxiv_id": "2504.04715", "paper_pdf": "papers/3DZeEUTwhq.pdf", "paper_pdf_sha256": "d1d2e34ab04395637bdf86f1e4148b3dcc2003f87913e613caa9b313cc7810ea", "paper_pdf_bytes": 688098, "paper_pdf_source": "openreview", "code_url": "https://github.com/sunblaze-ucb/llm-api-audit", "code_repository": "sunblaze-ucb/llm-api-audit", "code_commit": "21ffeef65c4c92689e4db5a6a8ec2a779d2bc049", "code_archive": "repos/3DZeEUTwhq.zip", "code_archive_sha256": "ec41d15b8c3142121872c13b845e74f8d6377f973d8ca42663456af21a085add", "code_archive_bytes": 11433031, "code_file_count": 409, "code_extensions": {".py": 393, ".sh": 11, ".ipynb": 4, ".cpp": 1}, "github_disk_usage_kb": 7772, "github_languages": {"Python": 1873356, "Jupyter Notebook": 85561, "Shell": 11893, "C++": 6748}, "github_archived": false, "github_pushed_at": "2025-04-10T06:22:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/are-you-getting-what-you-pay-for-auditing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uLAAVg0ymc", "year": 2025, "status": "rejected", "title": "Do we need rebalancing strategies? A theoretical and empirical study around SMOTE and its variants", "authors": ["Abdoulaye SAKHO", "Emmanuel Malherbe", "Erwan Scornet"], "authorids": ["~Abdoulaye_SAKHO1", "~Emmanuel_Malherbe3", "~Erwan_Scornet1"], "authors_source": "OpenReview API", "abstract": "Synthetic Minority Oversampling Technique (SMOTE) is a common rebalancing strategy for handling imbalanced tabular data sets. However, few works analyze SMOTE theoretically. In this paper, we prove that SMOTE  (with default parameter) tends to copy the original minority samples asymptotically. We also prove that SMOTE exhibits boundary artifacts, thus justifying existing SMOTE variants. Then we introduce two new SMOTE-related strategies, and compare them with state-of-the-art rebalancing procedures. Surprisingly, for most data sets, we observe that applying no rebalancing strategy is competitive in terms of predictive performances, with tuned random forests, logistic regression or LightGBM. For highly imbalanced data sets, our new methods, named CV-SMOTE and Multivariate Gaussian SMOTE, are competitive. Besides, our analysis sheds some lights on the behavior of common rebalancing strategies, when used in conjunction with random forests.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "hsHVuhcKeI", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6639/Reviewer_aPEe"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper makes a theretical analysis of the well-known SMOTE method for imbalance classification, and proposes two simplest variants.", "review_text": "This paper makes a theretical analysis of the well-known SMOTE method for imbalance classification, and proposes two simplest variants.", "strengths": "This paper involves both theoretical analysis and method development", "weaknesses": "1) The theretical analysis is based on the assumption that $n$ is sufficiently large such that $K/n$ tends to zero, but this assumption is totally impractical in the setting of imbalanced learning where minority class has few examples.\n2) The theretical analysis provides little information with respect to imbalanced learning, and it simply discusses the SMOTE method in the general learning setting. The result does not help understand the essence of imbalanced learning, and does not motivate how to solve the problem.\n3) The connection between theoretical analysis and the method proposal is not well established. \n4) The two proposed methods are not attractive. CV SMOTE is simply a hyperparameter tuning based on cross-validation, which is usually effective and commonly used technique as an engineering approach, without any theoretical or methodological contribution. MGS requires that there are at least $d+1$ minority-class examples, which may be impractical for high-dimensional data. In addition, even there are more than $d+1$ minority-class examples available, it has to assume that the covariance matrix is diagonal, which is also not the case in practice and thus over-generalized samples would be generated, which may do harm to the performance of imbalanced learning.  \n5) The experimental results are not significant.", "questions": "Please refer to the weaknesses 1, 2, 3 and 4.\n\nMore specifically, \n1) Could you provide bounds or approximations that hold for small $n$, instead of huge $n$?\n2) Could you analyze how SMOTE affects the decision boundary or classification margins in imbalanced settings? \n3) Could you explain the reasons for designing these two variants based on the theoretical analysis in your paper and your responses to the questions 1) and 2)?\n4) Have you considered any theoretical guarantees for CV SMOTE? Have you explored regularization techniques for MGS to handle high-dimensional data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper makes a theretical analysis of the well-known SMOTE method for imbalance classification, and proposes two simplest variants.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "This paper involves both theoretical analysis and method development", "weaknesses": "1) The theretical analysis is based on the assumption that $n$ is sufficiently large such that $K/n$ tends to zero, but this assumption is totally impractical in the setting of imbalanced learning where minority class has few examples.\n2) The theretical analysis provides little information with respect to imbalanced learning, and it simply discusses the SMOTE method in the general learning setting. The result does not help understand the essence of imbalanced learning, and does not motivate how to solve the problem.\n3) The connection between theoretical analysis and the method proposal is not well established. \n4) The two proposed methods are not attractive. CV SMOTE is simply a hyperparameter tuning based on cross-validation, which is usually effective and commonly used technique as an engineering approach, without any theoretical or methodological contribution. MGS requires that there are at least $d+1$ minority-class examples, which may be impractical for high-dimensional data. In addition, even there are more than $d+1$ minority-class examples available, it has to assume that the covariance matrix is diagonal, which is also not the case in practice and thus over-generalized samples would be generated, which may do harm to the performance of imbalanced learning.  \n5) The experimental results are not significant.", "questions": "Please refer to the weaknesses 1, 2, 3 and 4.\n\nMore specifically, \n1) Could you provide bounds or approximations that hold for small $n$, instead of huge $n$?\n2) Could you analyze how SMOTE affects the decision boundary or classification margins in imbalanced settings? \n3) Could you explain the reasons for designing these two variants based on the theoretical analysis in your paper and your responses to the questions 1) and 2)?\n4) Have you considered any theoretical guarantees for CV SMOTE? Have you explored regularization techniques for MGS to handle high-dimensional data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730609838692}, {"id": "6q0JCoD8NY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6639/Reviewer_rG2v"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents a comprehensive study of synthetic rebalancing strategies from theoretical and computational perspectives, further proposing two promising approaches specifically designed for highly imbalanced datasets. The findings reveal that for most mildly imbalanced datasets, applying no rebalancing strategy can be competitive, while rebalancing methods show clear benefits in scenarios with high imbalance.", "review_text": "This paper presents a comprehensive study of synthetic rebalancing strategies from theoretical and computational perspectives, further proposing two promising approaches specifically designed for highly imbalanced datasets. The findings reveal that for most mildly imbalanced datasets, applying no rebalancing strategy can be competitive, while rebalancing methods show clear benefits in scenarios with high imbalance.", "strengths": "1. The paper provides an in-depth theoretical analysis of SMOTE, offering valuable insights into its behavior and mechanics, which enhances the understanding of its effectiveness in handling imbalanced datasets.\n2. The paper is well-written, with a clear articulation of its objectives and contributions. In particular, the Introduction and Related Works sections effectively outline the development of solutions for imbalanced datasets, providing a strong foundation for the study's motivation and purpose.", "weaknesses": "1. It would be incorrect and confusing about the definition of imbalanced ratio. For example, in line 147, the authors define the imbalance ratio as the proportion of minority samples to the total samples (minority samples/total samples). However, in line 400, the authors mention that \"applying no strategy is the best, probably highlighting that the imbalance ratio is not high enough\". This wording suggest low-imbalanced datasets while not high imbalance ratio in your definition indicates highly imbalanced datasets.\n2. The novelty of the paper appears limited in several aspects. a. While the paper offers a thorough theoretical examination of SMOTE, the insights may not extend significantly beyond those of earlier works. I may not have the complete picture, however, and remain open to further clarification from the authors on their theoretical contributions. b. CV-SMOTE addresses hyperparameter selection via cross validation, while MGS introduces Gaussian sampling. Could the authors provide any unique aspects of their implementation or theoretical justifications that set the approaches apart from existing methods? c. The observation that applying no rebalancing strategy performs competitively on mildly imbalanced datasets is consistent with existing knowledge. Could the authors discuss how their empirical results add to or refine existing knowledge?\n3. The study focuses solely on binary classification and tabular data. It would be beneficial to discuss how the findings might extend to multiclass problems to improve the paper's generalizability.", "questions": "My opinions and questions are outlined in the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a comprehensive study of synthetic rebalancing strategies from theoretical and computational perspectives, further proposing two promising approaches specifically designed for highly imbalanced datasets. The findings reveal that for most mildly imbalanced datasets, applying no rebalancing strategy can be competitive, while rebalancing methods show clear benefits in scenarios with high imbalance.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The paper provides an in-depth theoretical analysis of SMOTE, offering valuable insights into its behavior and mechanics, which enhances the understanding of its effectiveness in handling imbalanced datasets.\n2. The paper is well-written, with a clear articulation of its objectives and contributions. In particular, the Introduction and Related Works sections effectively outline the development of solutions for imbalanced datasets, providing a strong foundation for the study's motivation and purpose.", "weaknesses": "1. It would be incorrect and confusing about the definition of imbalanced ratio. For example, in line 147, the authors define the imbalance ratio as the proportion of minority samples to the total samples (minority samples/total samples). However, in line 400, the authors mention that \"applying no strategy is the best, probably highlighting that the imbalance ratio is not high enough\". This wording suggest low-imbalanced datasets while not high imbalance ratio in your definition indicates highly imbalanced datasets.\n2. The novelty of the paper appears limited in several aspects. a. While the paper offers a thorough theoretical examination of SMOTE, the insights may not extend significantly beyond those of earlier works. I may not have the complete picture, however, and remain open to further clarification from the authors on their theoretical contributions. b. CV-SMOTE addresses hyperparameter selection via cross validation, while MGS introduces Gaussian sampling. Could the authors provide any unique aspects of their implementation or theoretical justifications that set the approaches apart from existing methods? c. The observation that applying no rebalancing strategy performs competitively on mildly imbalanced datasets is consistent with existing knowledge. Could the authors discuss how their empirical results add to or refine existing knowledge?\n3. The study focuses solely on binary classification and tabular data. It would be beneficial to discuss how the findings might extend to multiclass problems to improve the paper's generalizability.", "questions": "My opinions and questions are outlined in the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730535033793}, {"id": "XC8U07THP3", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6639/Reviewer_XTgD"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper analyzes SMOTE theoretically. Its contributions include:\n\n(1) It proves that, without tuning the hyperparameter K (usually set to 5), SMOTE asymptotically copies the original minority samples, therefore lacking the intrinsic variability required in any synthetic generative procedure. It provides numerical illustrations of this limitation.\n\n(2) It proves that SMOTE density vanishes near the boundary of the support of the minority distribution, therefore justifying the introduction of SMOTE variants such as BorderLine SMOTE.\n\n(3) It introduces two SMOTE alternatives, CVSMOTE and Multivariate Gaussian SMOTE (MGS). It evaluates these two new strategies and state-of-the-art rebalancing strategies on several real-world data sets using random forests/logistic regression/LightGBM. These experiments show that applying no strategy is competitive for most data sets. For the remaining data sets, the proposed strategies are among the best strategies in terms of predictive performance. The analysis of experiment results also provides some explanations about the good behavior of RUS, due to an implicit regularization in presence of random forests classifiers.", "review_text": "This paper analyzes SMOTE theoretically. Its contributions include:\n\n(1) It proves that, without tuning the hyperparameter K (usually set to 5), SMOTE asymptotically copies the original minority samples, therefore lacking the intrinsic variability required in any synthetic generative procedure. It provides numerical illustrations of this limitation.\n\n(2) It proves that SMOTE density vanishes near the boundary of the support of the minority distribution, therefore justifying the introduction of SMOTE variants such as BorderLine SMOTE.\n\n(3) It introduces two SMOTE alternatives, CVSMOTE and Multivariate Gaussian SMOTE (MGS). It evaluates these two new strategies and state-of-the-art rebalancing strategies on several real-world data sets using random forests/logistic regression/LightGBM. These experiments show that applying no strategy is competitive for most data sets. For the remaining data sets, the proposed strategies are among the best strategies in terms of predictive performance. The analysis of experiment results also provides some explanations about the good behavior of RUS, due to an implicit regularization in presence of random forests classifiers.", "strengths": "1. This paper theoretically analyzes Synthetic Minority Oversampling Technique (SMOTE), which is a widely adopted rebalancing strategy for handling imbalanced tabular data sets and has numerous variants. The theoretical analysis is comprehensive.\n2. The paper proposes two SMOTE alternatives: CVSMOTE and Multivariate Gaussian SMOTE (MGS), thus making minor technical innovations.\n3. Through extensive experiments on various datasets, the paper demonstrates that applying no rebalancing strategy can be competitive for some data sets. This finding is insightful.", "weaknesses": "Please refer to questions.", "questions": "1. It is recommended to clearly explain the implications of the metric \\(\\bar{C}(Z, X)/\\bar{C}(\\tilde{X}, X)\\), such as what it signifies when the value approaches 1, 0, or exceeds 1, as well as the impact of the corresponding synthetic samples on the classification model performance, which currently lacks discussion.\n\n2. Regarding Theorem 3.2, the author's claim that \"This highlights a good behavior of the default setting of SMOTE (K = 5), as it can create more data points, different from the original sample, and distributed as the original sample\" needs further justification. The reasoning for \"create more data points\" is unclear, and with finite \\(n\\) in practice, the conclusion \"distributed as the original sample\" may require additional support.\n\n3. In the statement, “Choosing \\(K\\) that increases with \\(n\\) leads to larger characteristic distances: SMOTE observations are more distant from their central points. Corollary 3.6 leads us to choose \\(K\\) such that \\(K/n\\) does not tend too fast to zero, so that SMOTE observations are not too close to the original minority samples,” it's hard to understand “Choosing \\(K\\) that increases with \\(n\\)” and “tend too fast to zero.” These informal expressions can be confusing. It is suggested to explain the practical meaning of larger characteristic distances, such as diversity, to aid understanding.\n\n4. In Section 3.2, THEORETICAL RESULTS ON SMOTE, could explicit ranges for \\(K\\) values be provided? Is there a trade-off between \"regenerating the distribution of the minority class,\" avoiding \"boundary bias,\" and achieving \"more diversity\"?\n\n5. The term “boundary artifacts” appears only in the abstract and conclusion but not elsewhere in the paper. Furthermore, no reference is found in the original BorderLine SMOTE paper. Additionally, the author claims to “prove that SMOTE exhibits boundary artifacts, therefore justifying the introduction of SMOTE variants such as BorderLine SMOTE (Section 3),” so it is recommended to provide relevant citations and explain how BorderLine SMOTE addresses this issue.\n\n6. Concerns regarding Section 3.3, NUMERICAL ILLUSTRATIONS:\n   - (1) In the first result analysis, the author states, “Note that, for the other asymptotics in \\(K\\), the diversity of SMOTE observations increases with \\(n\\), meaning \\(\\bar{C}(Z, X)\\) gets closer to \\(\\bar{C}(\\tilde{X}, X)\\). This behavior in terms of average distance is ideal since \\(\\tilde{X}\\) is drawn from the same theoretical distribution as \\(X\\). On the contrary, \\(K = 5\\) maintains a lower average distance, indicating a lack of diversity in generated points.” Based on previous descriptions, higher diversity is considered better, so \\(K = 5\\) with low diversity should be undesirable, which conflicts with the statement “This highlights a good behavior of the default setting of SMOTE (K = 5).”\n   - (2) How should “this diversity is asymptotically more important for xxx” be understood?\n   - (3) How should “By construction, SMOTE data points are close to central points, which may explain why the quantity of interest in Figure 1 is smaller than 1” be interpreted? Furthermore, in the second result analysis, some values of the metric \\(\\bar{C}(Z, X)/\\bar{C}(\\tilde{X}, X)\\) exceed 1; what implications would this have?\n\n7. The conclusion that “applying no strategy is competitive for most data sets” based solely on “the imbalance ratio is not high enough or the learning task not difficult enough” is not rigorous enough. A more detailed analysis of dataset characteristics would help identify when it’s appropriate to apply no strategy versus when to use rebalancing strategies.\n\n8. It is recommended to compare some newer SMOTE variants, as some of the rebalancing strategies used in the experiments are somewhat outdated.\n\n9. More precise and standardized descriptions would improve readability. For example, phrases like “guarantees asymptotically” and “smoothed out” require clarification. Consistent terminology should be used instead of mixing “SMOTE data points,” “SMOTE samples,” and “SMOTE observations.” Additionally, the title of Table 2 could specify the experimental module for clarity.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper analyzes SMOTE theoretically. Its contributions include:\n\n(1) It proves that, without tuning the hyperparameter K (usually set to 5), SMOTE asymptotically copies the original minority samples, therefore lacking the intrinsic variability required in any synthetic generative procedure. It provides numerical illustrations of this limitation.\n\n(2) It proves that SMOTE density vanishes near the boundary of the support of the minority distribution, therefore justifying the introduction of SMOTE variants such as BorderLine SMOTE.\n\n(3) It introduces two SMOTE alternatives, CVSMOTE and Multivariate Gaussian SMOTE (MGS). It evaluates these two new strategies and state-of-the-art rebalancing strategies on several real-world data sets using random forests/logistic regression/LightGBM. These experiments show that applying no strategy is competitive for most data sets. For the remaining data sets, the proposed strategies are among the best strategies in terms of predictive performance. The analysis of experiment results also provides some explanations about the good behavior of RUS, due to an implicit regularization in presence of random forests classifiers.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. This paper theoretically analyzes Synthetic Minority Oversampling Technique (SMOTE), which is a widely adopted rebalancing strategy for handling imbalanced tabular data sets and has numerous variants. The theoretical analysis is comprehensive.\n2. The paper proposes two SMOTE alternatives: CVSMOTE and Multivariate Gaussian SMOTE (MGS), thus making minor technical innovations.\n3. Through extensive experiments on various datasets, the paper demonstrates that applying no rebalancing strategy can be competitive for some data sets. This finding is insightful.", "weaknesses": "Please refer to questions.", "questions": "1. It is recommended to clearly explain the implications of the metric \\(\\bar{C}(Z, X)/\\bar{C}(\\tilde{X}, X)\\), such as what it signifies when the value approaches 1, 0, or exceeds 1, as well as the impact of the corresponding synthetic samples on the classification model performance, which currently lacks discussion.\n\n2. Regarding Theorem 3.2, the author's claim that \"This highlights a good behavior of the default setting of SMOTE (K = 5), as it can create more data points, different from the original sample, and distributed as the original sample\" needs further justification. The reasoning for \"create more data points\" is unclear, and with finite \\(n\\) in practice, the conclusion \"distributed as the original sample\" may require additional support.\n\n3. In the statement, “Choosing \\(K\\) that increases with \\(n\\) leads to larger characteristic distances: SMOTE observations are more distant from their central points. Corollary 3.6 leads us to choose \\(K\\) such that \\(K/n\\) does not tend too fast to zero, so that SMOTE observations are not too close to the original minority samples,” it's hard to understand “Choosing \\(K\\) that increases with \\(n\\)” and “tend too fast to zero.” These informal expressions can be confusing. It is suggested to explain the practical meaning of larger characteristic distances, such as diversity, to aid understanding.\n\n4. In Section 3.2, THEORETICAL RESULTS ON SMOTE, could explicit ranges for \\(K\\) values be provided? Is there a trade-off between \"regenerating the distribution of the minority class,\" avoiding \"boundary bias,\" and achieving \"more diversity\"?\n\n5. The term “boundary artifacts” appears only in the abstract and conclusion but not elsewhere in the paper. Furthermore, no reference is found in the original BorderLine SMOTE paper. Additionally, the author claims to “prove that SMOTE exhibits boundary artifacts, therefore justifying the introduction of SMOTE variants such as BorderLine SMOTE (Section 3),” so it is recommended to provide relevant citations and explain how BorderLine SMOTE addresses this issue.\n\n6. Concerns regarding Section 3.3, NUMERICAL ILLUSTRATIONS:\n   - (1) In the first result analysis, the author states, “Note that, for the other asymptotics in \\(K\\), the diversity of SMOTE observations increases with \\(n\\), meaning \\(\\bar{C}(Z, X)\\) gets closer to \\(\\bar{C}(\\tilde{X}, X)\\). This behavior in terms of average distance is ideal since \\(\\tilde{X}\\) is drawn from the same theoretical distribution as \\(X\\). On the contrary, \\(K = 5\\) maintains a lower average distance, indicating a lack of diversity in generated points.” Based on previous descriptions, higher diversity is considered better, so \\(K = 5\\) with low diversity should be undesirable, which conflicts with the statement “This highlights a good behavior of the default setting of SMOTE (K = 5).”\n   - (2) How should “this diversity is asymptotically more important for xxx” be understood?\n   - (3) How should “By construction, SMOTE data points are close to central points, which may explain why the quantity of interest in Figure 1 is smaller than 1” be interpreted? Furthermore, in the second result analysis, some values of the metric \\(\\bar{C}(Z, X)/\\bar{C}(\\tilde{X}, X)\\) exceed 1; what implications would this have?\n\n7. The conclusion that “applying no strategy is competitive for most data sets” based solely on “the imbalance ratio is not high enough or the learning task not difficult enough” is not rigorous enough. A more detailed analysis of dataset characteristics would help identify when it’s appropriate to apply no strategy versus when to use rebalancing strategies.\n\n8. It is recommended to compare some newer SMOTE variants, as some of the rebalancing strategies used in the experiments are somewhat outdated.\n\n9. More precise and standardized descriptions would improve readability. For example, phrases like “guarantees asymptotically” and “smoothed out” require clarification. Consistent terminology should be used instead of mixing “SMOTE data points,” “SMOTE samples,” and “SMOTE observations.” Additionally, the title of Table 2 could specify the experimental module for clarity.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730510033071}, {"id": "HDSSKFThrh", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6639/Reviewer_7ocJ"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper investigates the effectiveness of rebalancing strategies for handling imbalanced datasets in binary classification, specifically focusing on the Synthetic Minority Oversampling Technique (SMOTE) and its variants. The authors provide a theoretical analysis demonstrating that default SMOTE tends to asymptotically copy original minority samples and exhibits boundary artifacts. They introduce two new strategies, CV-SMOTE and Multivariate Gaussian SMOTE (MGS), and compare their performance against traditional methods like Random Under Sampling (RUS), Random Over Sampling (ROS), and Class Weighting (CW). The empirical results indicate that for many datasets, applying no rebalancing strategy is competitive, while the proposed methods show promise for highly imbalanced datasets.", "review_text": "The paper investigates the effectiveness of rebalancing strategies for handling imbalanced datasets in binary classification, specifically focusing on the Synthetic Minority Oversampling Technique (SMOTE) and its variants. The authors provide a theoretical analysis demonstrating that default SMOTE tends to asymptotically copy original minority samples and exhibits boundary artifacts. They introduce two new strategies, CV-SMOTE and Multivariate Gaussian SMOTE (MGS), and compare their performance against traditional methods like Random Under Sampling (RUS), Random Over Sampling (ROS), and Class Weighting (CW). The empirical results indicate that for many datasets, applying no rebalancing strategy is competitive, while the proposed methods show promise for highly imbalanced datasets.", "strengths": "1. The paper is well-organized, with clear sections that guide the reader through the theoretical analysis, empirical evaluation, and results.\n2. The paper provides a solid theoretical foundation for understanding the behavior of SMOTE, including its limitations and boundary artifacts.\n3. The authors conduct a thorough empirical evaluation across multiple datasets.", "weaknesses": "1. The diversity in terms of data characteristics (e.g., feature types, distributions) could be expanded.\n2. More recent references and baselines should be included.", "questions": "What criteria should practitioners use to select the hyperparameter grid for CV-SMOTE? Are there specific guidelines or best practices that can be derived from the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates the effectiveness of rebalancing strategies for handling imbalanced datasets in binary classification, specifically focusing on the Synthetic Minority Oversampling Technique (SMOTE) and its variants. The authors provide a theoretical analysis demonstrating that default SMOTE tends to asymptotically copy original minority samples and exhibits boundary artifacts. They introduce two new strategies, CV-SMOTE and Multivariate Gaussian SMOTE (MGS), and compare their performance against traditional methods like Random Under Sampling (RUS), Random Over Sampling (ROS), and Class Weighting (CW). The empirical results indicate that for many datasets, applying no rebalancing strategy is competitive, while the proposed methods show promise for highly imbalanced datasets.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "1. The paper is well-organized, with clear sections that guide the reader through the theoretical analysis, empirical evaluation, and results.\n2. The paper provides a solid theoretical foundation for understanding the behavior of SMOTE, including its limitations and boundary artifacts.\n3. The authors conduct a thorough empirical evaluation across multiple datasets.", "weaknesses": "1. The diversity in terms of data characteristics (e.g., feature types, distributions) could be expanded.\n2. More recent references and baselines should be included.", "questions": "What criteria should practitioners use to select the hyperparameter grid for CV-SMOTE? Are there specific guidelines or best practices that can be derived from the experiments?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730211748742}], "openreview_url": "https://openreview.net/forum?id=uLAAVg0ymc", "arxiv_id": "2402.03819", "paper_pdf": "papers/uLAAVg0ymc.pdf", "paper_pdf_sha256": "534af05494452613ed9d4eb4c133140dc887364d31ed35afa35e5b8740f70947", "paper_pdf_bytes": 530463, "paper_pdf_source": "openreview", "code_url": "https://github.com/artefactory/smote_strategies_study", "code_repository": "artefactory/smote_strategies_study", "code_commit": "476f902a7aafbea99432d0ac5de95509e2ff275a", "code_archive": "repos/uLAAVg0ymc.zip", "code_archive_sha256": "9fc8edb25052a82b46e9453ef84830192bd0043d71d931030e4e9e502614727c", "code_archive_bytes": 143969, "code_file_count": 10, "code_extensions": {".ipynb": 6, ".py": 4}, "github_disk_usage_kb": 375, "github_languages": {"Jupyter Notebook": 254510, "Python": 98462}, "github_archived": false, "github_pushed_at": "2025-09-02T11:37:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/theoretical-and-experimental-study-of-smote"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "K7KQkiHanD", "year": 2024, "status": "rejected", "title": "One-for-All: Generalized LoRA for Parameter-Efficient Fine-tuning", "authors": ["Arnav Chavan", "Zhuang Liu", "Deepak Gupta", "Eric Xing", "Zhiqiang Shen"], "authorids": ["~Arnav_Chavan1", "~Zhuang_Liu1", "~Deepak_Gupta2", "~Eric_Xing1", "~Zhiqiang_Shen1"], "authors_source": "OpenReview API", "abstract": "We present Generalized LoRA (GLoRA), a flexible approach for universal parameter-efficient fine-tuning tasks. Enhancing Low-Rank Adaptation (LoRA), GLoRA employs a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations, providing more flexibility and capability across diverse tasks and datasets. Moreover, GLoRA facilitates efficient parameter adaptation by employing a scalable, modular, layer-wise structure search that learns individual adapter of each layer. \nOriginating from a unified mathematical formulation, GLoRA exhibits strong transfer learning, few-shot learning and domain generalization abilities, as it adapts to new tasks through not only weights but also additional dimensions like activations. Comprehensive experiments demonstrate that GLoRA outperforms all previous methods in natural, specialized, and structured benchmarks in the vision field, achieving superior accuracy with fewer parameters and computations. Our models on LLaMA-1 and 2 also show considerable enhancements compared to the original LoRA in the language domain. Furthermore, our structural re-parameterization design ensures that GLoRA incurs no extra inference cost, rendering it a practical solution for resource-limited applications.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "FlH3fzupTg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2326/Reviewer_jzkZ"], "rating": "5: marginally below the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper discusses an enhancement to Low-Rank Adaptation (LoRA), which is call GLoR and can be a flexible approach to optimize model inference results. Experiments on llama and vit shows that GLoRA can improve over original LoRA consistently.", "review_text": "The paper discusses an enhancement to Low-Rank Adaptation (LoRA), which is call GLoR and can be a flexible approach to optimize model inference results. Experiments on llama and vit shows that GLoRA can improve over original LoRA consistently.", "strengths": "This paper is clearly presented and well organized. The authors also provide a detailed discussion of related works and variants.\n\nGLoRA referneces the inspirations from RepVGG, which introduces fusable parameters during training to improve model capacity. This can generally bring improvements without extra inference cost as shown in the experiments. \n\nGLoRA offers a unifed framework that includes multiple fine-tuning paradigms and provides a more generalized prompt mdule design per layer. The final scheme is searched via evolutional algorithms, providng better capacity and flexibility.", "weaknesses": "The authors claim that GLoRA can be \"seamlessly integrate into the base network\", but it seems such design is for linear layer only. But there are many other type of operators like conv / normalization layers. How can GLoRA be combined with those layers?\n\nThe evolutional search (Sec 2.4) is crucial for GLoRA as it decides which layer and scheme to use during fine-tuning. However, the details of the search and final chosen paradigms are not clearly discussed in the main paper. \n\nAs the abstract emphasiss the llama experiments. The table 2 is not solid neough to support the authors' claim. For example, the mean and variance is not included; the number of learnable paramters is missing; the lora baseline for llama-v2 is not reported. Seems like the experiments are rushed and may not be solid.", "questions": "No extra inference cost is the novelty from RepVGG or LoRA, which should be claimed as the contribution of GLoRA\n\nThe main experiments are based on ViT but abstract emphasis for llama. Please make the claim conssitent.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper discusses an enhancement to Low-Rank Adaptation (LoRA), which is call GLoR and can be a flexible approach to optimize model inference results. Experiments on llama and vit shows that GLoRA can improve over original LoRA consistently.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "This paper is clearly presented and well organized. The authors also provide a detailed discussion of related works and variants.\n\nGLoRA referneces the inspirations from RepVGG, which introduces fusable parameters during training to improve model capacity. This can generally bring improvements without extra inference cost as shown in the experiments. \n\nGLoRA offers a unifed framework that includes multiple fine-tuning paradigms and provides a more generalized prompt mdule design per layer. The final scheme is searched via evolutional algorithms, providng better capacity and flexibility.", "weaknesses": "The authors claim that GLoRA can be \"seamlessly integrate into the base network\", but it seems such design is for linear layer only. But there are many other type of operators like conv / normalization layers. How can GLoRA be combined with those layers?\n\nThe evolutional search (Sec 2.4) is crucial for GLoRA as it decides which layer and scheme to use during fine-tuning. However, the details of the search and final chosen paradigms are not clearly discussed in the main paper. \n\nAs the abstract emphasiss the llama experiments. The table 2 is not solid neough to support the authors' claim. For example, the mean and variance is not included; the number of learnable paramters is missing; the lora baseline for llama-v2 is not reported. Seems like the experiments are rushed and may not be solid.", "questions": "No extra inference cost is the novelty from RepVGG or LoRA, which should be claimed as the contribution of GLoRA\n\nThe main experiments are based on ViT but abstract emphasis for llama. Please make the claim conssitent.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699265522529}, {"id": "iRswZ7XKEg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2326/Reviewer_1LBP"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents a new PEFT module named Generalized LoRA (GLoRA), which can be applied to different tasks. The authors claim that GLoRA is more general than existing PEFT modules since it can facilitate efficient parameter adaptation by employing a more scalable structure search. Moreover, the authors conduct multiple experiments including few-shot learning and domain generalization tasks to demonstrate the effectiveness of GLoRA.", "review_text": "This paper presents a new PEFT module named Generalized LoRA (GLoRA), which can be applied to different tasks. The authors claim that GLoRA is more general than existing PEFT modules since it can facilitate efficient parameter adaptation by employing a more scalable structure search. Moreover, the authors conduct multiple experiments including few-shot learning and domain generalization tasks to demonstrate the effectiveness of GLoRA.", "strengths": "1. GLoRA has a re-parameterization design. It is more similar to LORA than Adapter. It makes GLoRA more flexible since it does not need to change the structure of the original backbone. And it incurs no extra inference cost.\n2. GLoRA integrates multiple methods and can perform similar effects as most of the existing PEFT modules.\n3. The authors conduct multiple experiments to demonstrate the generality and effectiveness of GLoRA.", "weaknesses": "1. It seems that GLoRA is not general, since it has an evolutionary search procedure to obtain the suitable components. The idea is similar to Neural Prompt Search [1]. GLoRA is not a fixed design as existing modules, which might limit its practicality.\n2. GLoRA has a large search space, which might yield huge time costs. However, the authors have not mentioned the actual training time and memory cost of GLoRA, which is very important for PEFT modules.\n3. The authors introduce multiple PEFT modules including AdaptFormer, LoRA, etc. So could GLoRA simulate all of these modules? As far as I know, these modules are applied on different layers (LoRA in multi-head self-attention layers, while AdaptFormer in MLP layers). And which layer is GLoRA applied in practice?\n\n[1] Neural Prompt Search. In https://arxiv.org/abs/2206.04673.", "questions": "See \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a new PEFT module named Generalized LoRA (GLoRA), which can be applied to different tasks. The authors claim that GLoRA is more general than existing PEFT modules since it can facilitate efficient parameter adaptation by employing a more scalable structure search. Moreover, the authors conduct multiple experiments including few-shot learning and domain generalization tasks to demonstrate the effectiveness of GLoRA.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. GLoRA has a re-parameterization design. It is more similar to LORA than Adapter. It makes GLoRA more flexible since it does not need to change the structure of the original backbone. And it incurs no extra inference cost.\n2. GLoRA integrates multiple methods and can perform similar effects as most of the existing PEFT modules.\n3. The authors conduct multiple experiments to demonstrate the generality and effectiveness of GLoRA.", "weaknesses": "1. It seems that GLoRA is not general, since it has an evolutionary search procedure to obtain the suitable components. The idea is similar to Neural Prompt Search [1]. GLoRA is not a fixed design as existing modules, which might limit its practicality.\n2. GLoRA has a large search space, which might yield huge time costs. However, the authors have not mentioned the actual training time and memory cost of GLoRA, which is very important for PEFT modules.\n3. The authors introduce multiple PEFT modules including AdaptFormer, LoRA, etc. So could GLoRA simulate all of these modules? As far as I know, these modules are applied on different layers (LoRA in multi-head self-attention layers, while AdaptFormer in MLP layers). And which layer is GLoRA applied in practice?\n\n[1] Neural Prompt Search. In https://arxiv.org/abs/2206.04673.", "questions": "See \"Weaknesses\".", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699029036130}, {"id": "KBx0xPiP5V", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2326/Reviewer_8AV4"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper presents Generalized LoRA (GLoRA), an efficient framework for fine-tuning machine learning models. Building on Low-Rank Adaptation (LoRA), GLoRA introduces an advanced prompt module that not only refines pre-trained model weights but also modulates intermediate activations. Uniquely, this prompt module operates individually across each model layer, ensuring versatility across various tasks and datasets.\n\nGLoRA employs a cohesive mathematical strategy to adapt to new tasks, modifying both weights and activations. This methodology positions it strongly for transfer learning, few-shot learning, and domain generalization.\n\nThe authors substantiate GLoRA's efficacy through experiments on diverse datasets and tasks, encompassing downstream fine-tuning, few-shot learning, domain generalization, and recent popular LLMs. The results demonstrate GLoRA's superior performance over prior techniques in these areas. Remarkably, despite its heightened capability, GLoRA demands fewer parameters and computations without incurring additional inference costs, akin to LoRA, making it an optimal choice for resource-constrained applications.", "review_text": "The paper presents Generalized LoRA (GLoRA), an efficient framework for fine-tuning machine learning models. Building on Low-Rank Adaptation (LoRA), GLoRA introduces an advanced prompt module that not only refines pre-trained model weights but also modulates intermediate activations. Uniquely, this prompt module operates individually across each model layer, ensuring versatility across various tasks and datasets.\n\nGLoRA employs a cohesive mathematical strategy to adapt to new tasks, modifying both weights and activations. This methodology positions it strongly for transfer learning, few-shot learning, and domain generalization.\n\nThe authors substantiate GLoRA's efficacy through experiments on diverse datasets and tasks, encompassing downstream fine-tuning, few-shot learning, domain generalization, and recent popular LLMs. The results demonstrate GLoRA's superior performance over prior techniques in these areas. Remarkably, despite its heightened capability, GLoRA demands fewer parameters and computations without incurring additional inference costs, akin to LoRA, making it an optimal choice for resource-constrained applications.", "strengths": "GLoRA effectively consolidates previous parameter-efficient fine-tuning methods within Equation 10. Importantly, all adjustable support tensors are linear, which makes structural re-parameterization readily accessible.\n\nThe paper highlights GLoRA's commendable capacity to generalize across diverse tasks, an invaluable quality in machine learning and a frequently challenging facet of model development.", "weaknesses": "Structural re-parameterization requires storing the full set of weights (including bias) for every individual downstream task. This means that as the number of these tasks increases, the storage needs can become prohibitively large. Although this approach might improve inference performance, the substantial storage overhead can be a major impediment for real-world deployment, especially when multiple tuned-models are needed to handle different downstream tasks.\n\nThe clarity of the paper is occasionally compromised by abrupt topic transitions, such as the unexpected introduction of \"GLoRA with Higher Capacity\" in section 2.6, without prior elucidation of terms like H_i and H_ini. A more coherent and gradual introduction of these concepts would enhance readability.\n\nThe authors touch on memory and training time costs but fail to provide concrete figures to substantiate their claims. Offering detailed, quantitative data on these costs would provide readers with a clearer picture of GLoRA's practical ramifications.", "questions": "Please refer to 'weakness' part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents Generalized LoRA (GLoRA), an efficient framework for fine-tuning machine learning models. Building on Low-Rank Adaptation (LoRA), GLoRA introduces an advanced prompt module that not only refines pre-trained model weights but also modulates intermediate activations. Uniquely, this prompt module operates individually across each model layer, ensuring versatility across various tasks and datasets.\n\nGLoRA employs a cohesive mathematical strategy to adapt to new tasks, modifying both weights and activations. This methodology positions it strongly for transfer learning, few-shot learning, and domain generalization.\n\nThe authors substantiate GLoRA's efficacy through experiments on diverse datasets and tasks, encompassing downstream fine-tuning, few-shot learning, domain generalization, and recent popular LLMs. The results demonstrate GLoRA's superior performance over prior techniques in these areas. Remarkably, despite its heightened capability, GLoRA demands fewer parameters and computations without incurring additional inference costs, akin to LoRA, making it an optimal choice for resource-constrained applications.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "GLoRA effectively consolidates previous parameter-efficient fine-tuning methods within Equation 10. Importantly, all adjustable support tensors are linear, which makes structural re-parameterization readily accessible.\n\nThe paper highlights GLoRA's commendable capacity to generalize across diverse tasks, an invaluable quality in machine learning and a frequently challenging facet of model development.", "weaknesses": "Structural re-parameterization requires storing the full set of weights (including bias) for every individual downstream task. This means that as the number of these tasks increases, the storage needs can become prohibitively large. Although this approach might improve inference performance, the substantial storage overhead can be a major impediment for real-world deployment, especially when multiple tuned-models are needed to handle different downstream tasks.\n\nThe clarity of the paper is occasionally compromised by abrupt topic transitions, such as the unexpected introduction of \"GLoRA with Higher Capacity\" in section 2.6, without prior elucidation of terms like H_i and H_ini. A more coherent and gradual introduction of these concepts would enhance readability.\n\nThe authors touch on memory and training time costs but fail to provide concrete figures to substantiate their claims. Offering detailed, quantitative data on these costs would provide readers with a clearer picture of GLoRA's practical ramifications.", "questions": "Please refer to 'weakness' part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698733556714}, {"id": "moqu3UINPk", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2326/Reviewer_QDBh"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a general form of low-rank adaptation for vision and language large models, based on a unified formula which the authors claim to encompass serveral previous parameter efficient finetuning methods such as VPT and LoRA. As for training networks with GLoRA, the authors exploit an evolutionary strategy to search for the best subnet after training the supernet. Extensive experiments on vision and language benchmarks show the effectiveness of the propose method.", "review_text": "The paper proposes a general form of low-rank adaptation for vision and language large models, based on a unified formula which the authors claim to encompass serveral previous parameter efficient finetuning methods such as VPT and LoRA. As for training networks with GLoRA, the authors exploit an evolutionary strategy to search for the best subnet after training the supernet. Extensive experiments on vision and language benchmarks show the effectiveness of the propose method.", "strengths": "1. The paper rethinks serveral previous PEFT methods and unifies them with a general form, contributing a novel perspective.\n2. To obtain the task-specific GLoRA network, the authors first train the supernet and then search for the best subnet.\n3. Extensive experiments are conducted on both vision and language benchmarks, and also in few-shot learning and domain generalization, showing the effectiveness GLoRA.", "weaknesses": "1. The presentation of the paper could be improved, especially when comparing with previous PEFT methods. The authors could draw figures or list tables to show how existing methods can be integrated into GLoRA framework, e.g. what are the specifications of the A/B/C/D/E support tensors in Eq. (10).\n2. I wonder how much training time (supernet training and subnet searching) does GLoRA cost, such that to compare with existing methods more clearly from the perspective of training efficiency.", "questions": "1. There exist some typos: 1) \"PETL\"(maybe PEFT?) at the end of page 3 (first line of Sec. Limitations); 2) 4th line of page 4: \"wieght\" -> weight.\n2. How about the performance if we do not add the weight/bias scaling term: W_0 x A and D x b_0 in Eq. (10) ? Or else, which of the five tensors are really necessary in terms of efficiency and efficacy ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a general form of low-rank adaptation for vision and language large models, based on a unified formula which the authors claim to encompass serveral previous parameter efficient finetuning methods such as VPT and LoRA. As for training networks with GLoRA, the authors exploit an evolutionary strategy to search for the best subnet after training the supernet. Extensive experiments on vision and language benchmarks show the effectiveness of the propose method.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "1. The paper rethinks serveral previous PEFT methods and unifies them with a general form, contributing a novel perspective.\n2. To obtain the task-specific GLoRA network, the authors first train the supernet and then search for the best subnet.\n3. Extensive experiments are conducted on both vision and language benchmarks, and also in few-shot learning and domain generalization, showing the effectiveness GLoRA.", "weaknesses": "1. The presentation of the paper could be improved, especially when comparing with previous PEFT methods. The authors could draw figures or list tables to show how existing methods can be integrated into GLoRA framework, e.g. what are the specifications of the A/B/C/D/E support tensors in Eq. (10).\n2. I wonder how much training time (supernet training and subnet searching) does GLoRA cost, such that to compare with existing methods more clearly from the perspective of training efficiency.", "questions": "1. There exist some typos: 1) \"PETL\"(maybe PEFT?) at the end of page 3 (first line of Sec. Limitations); 2) 4th line of page 4: \"wieght\" -> weight.\n2. How about the performance if we do not add the weight/bias scaling term: W_0 x A and D x b_0 in Eq. (10) ? Or else, which of the five tensors are really necessary in terms of efficiency and efficacy ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698466751000}], "openreview_url": "https://openreview.net/forum?id=K7KQkiHanD", "arxiv_id": "2306.07967", "paper_pdf": "papers/K7KQkiHanD.pdf", "paper_pdf_sha256": "fe6fd2a0e19a44a4b39b4fb2ade6c68b50e2e87f99bf25b6d0ab7f357ddc586e", "paper_pdf_bytes": 786191, "paper_pdf_source": "openreview", "code_url": "https://github.com/Arnav0400/ViT-Slim", "code_repository": "Arnav0400/ViT-Slim", "code_commit": "390467ddf529d747168709ce2f007a85a0399cca", "code_archive": "repos/K7KQkiHanD.zip", "code_archive_sha256": "046de9581cfd6cd723aeec9fd014f3bc55f6f862c9970e3585046ec4e695998c", "code_archive_bytes": 352907, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 367, "github_languages": {"Python": 171073}, "github_archived": false, "github_pushed_at": "2025-08-24T17:10:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/one-for-all-generalized-lora-for-parameter"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ESR6hysKDsW", "year": 2023, "status": "rejected", "title": "Class-Incremental Learning with Repetition", "authors": ["Hamed Hemati", "Andrea Cossu", "Antonio Carta", "Julio Hurtado", "Lorenzo Pellegrini", "Davide Bacciu", "Vincenzo Lomonaco", "Damian Borth"], "authorids": ["~Hamed_Hemati1", "~Andrea_Cossu1", "~Antonio_Carta1", "~Julio_Hurtado1", "~Lorenzo_Pellegrini1", "~Davide_Bacciu1", "~Vincenzo_Lomonaco1", "~Damian_Borth1"], "authors_source": "OpenReview API", "abstract": "Real-world data streams naturally include the repetition of previous concepts. From a Continual Learning (CL) perspective, repetition is a property of the environment and, unlike replay, cannot be controlled by the user. Nowadays, Class-Incremental scenarios represent the leading test-bed for assessing and comparing CL strategies. This family of scenarios is very easy to use, but it never allows revisiting previously seen classes, thus completely disregarding the role of repetition. We focus on the family of Class-Incremental with Repetition (CIR) scenarios, where repetition is embedded in the definition of the stream. We propose two stochastic scenario generators that produce a wide range of CIR scenarios starting from a single dataset and a few control parameters. We conduct the first comprehensive evaluation of repetition in CL by studying the behavior of existing CL strategies under different CIR scenarios. We then present a novel replay strategy that exploits repetition and counteracts the natural imbalance present in the stream. On both CIFAR100 and TinyImageNet, our strategy outperforms other replay approaches, which are not designed for environments with repetition.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "EmqkcFDinBy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2894/Reviewer_rWm4"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes a study on a different flavor of continual learning scenarios than previously considered in the literature. Precisely, one in which data separation is not strict among concepts, and later tasks can revisit one of the previously introduced classes. The authors name this \"class incremental with repetitions\" (CIL).\n\nThe authors show that in these scenarios storing and replaying data is less important (which is not surprising), unless it corrects for imbalanced exposures to the different classes. Therefore the authors propose a sampling method which balances the class representation during training.\n\nExperiments on the standard computer vision tasks show comparisons between domain incremental, class incremental and class incremental with repetitions.", "review_text": "The CL community would benefit from an analysis of a more natural stream of tasks than the extreme cases considered so far. Due to the clear exposure, interesting analysis and experiments I suggest the paper should be accepted.", "strengths": "Strengths:\n - novelty of the analysis\n - relevant CL algorithms analysed under the different scenarios\n - well written paper\n\nWeaknesses:\n - the results are not necessarily surprising; e.g. it is known that balancing classes in training batches benefits performance / optimization", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a study on a different flavor of continual learning scenarios than previously considered in the literature. Precisely, one in which data separation is not strict among concepts, and later tasks can revisit one of the previously introduced classes. The authors name this \"class incremental with repetitions\" (CIL).\n\nThe authors show that in these scenarios storing and replaying data is less important (which is not surprising), unless it corrects for imbalanced exposures to the different classes. Therefore the authors propose a sampling method which balances the class representation during training.\n\nExperiments on the standard computer vision tasks show comparisons between domain incremental, class incremental and class incremental with repetitions.", "strength_and_weaknesses": "Strengths:\n - novelty of the analysis\n - relevant CL algorithms analysed under the different scenarios\n - well written paper\n\nWeaknesses:\n - the results are not necessarily surprising; e.g. it is known that balancing classes in training batches benefits performance / optimization", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and clear about the details, therefore I consider it's easily reproducible.", "summary_of_the_review": "The CL community would benefit from an analysis of a more natural stream of tasks than the extreme cases considered so far. Due to the clear exposure, interesting analysis and experiments I suggest the paper should be accepted.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667161463758}, {"id": "LPxpGV8u8CC", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2894/Reviewer_MpJD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper focuses on class-incremental with repetition (CIR), which ranges from class- and domain-incremental learning. For building streams and managing repetition over time, the authors propose two CIR generators (i.e., slot-based and sampling-based), which are from two aspects to generate data streams. Moreover, the authors propose a frequency-aware (FA) storage policy tailored to CIR scenarios. Further, the authors perform the first comprehensive evaluation on CIFAR 100 and TinyImageNet to compare other replay approaches in the CIR scenario, and then show the effectiveness of FA.\n", "review_text": "This is not a high-quality paper without a clear expression and lack of solid theoretical analyses, which could not stimulate other research works. So, I recommend rejection.", "strengths": "Strengths\n1. The idea of CIR is absorbing and reasonable. Because for realistic data streams, it is natural to occurrences previously seen classes in new coming samples. Further, they designed a paradigm to build a stream with repetition via the proposed generators. \nWeaknesses\n\n1. The novelty of this paper is insufficient. In my view, this paper conducted a sampling procedure via two generators. As for frequency-aware replay, its improvement is dealing with unbalanced data. What is neglected in this paper, however, is that there have been numerous works on adaptative store policy in continual learning.\n\n2. The writing of this paper is not clear enough to follow. Firstly, the explanation of the discrepancy between CIR and traditional incremental learning (i.e., class- and domain-incremental) is not clear. Figure 1 replaced with actual objects may give a more intuitive understanding of the authors' motivation. Moreover, it is necessary to use a coherent algorithm to introduce the overall procedure, however, absent in the main text. Finally, the conclusion part is tedious and should be refined.\n\n3. The lack of formulas in this paper makes comprehending have more barriers. For example, the generators build streams part lacks the complementary mathematical explanation. \n\n4. The authors should provide comprehensive theoretical analyses of their method, in my view, which is necessary for solid work.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper focuses on class-incremental with repetition (CIR), which ranges from class- and domain-incremental learning. For building streams and managing repetition over time, the authors propose two CIR generators (i.e., slot-based and sampling-based), which are from two aspects to generate data streams. Moreover, the authors propose a frequency-aware (FA) storage policy tailored to CIR scenarios. Further, the authors perform the first comprehensive evaluation on CIFAR 100 and TinyImageNet to compare other replay approaches in the CIR scenario, and then show the effectiveness of FA.\n", "strength_and_weaknesses": "Strengths\n1. The idea of CIR is absorbing and reasonable. Because for realistic data streams, it is natural to occurrences previously seen classes in new coming samples. Further, they designed a paradigm to build a stream with repetition via the proposed generators. \nWeaknesses\n\n1. The novelty of this paper is insufficient. In my view, this paper conducted a sampling procedure via two generators. As for frequency-aware replay, its improvement is dealing with unbalanced data. What is neglected in this paper, however, is that there have been numerous works on adaptative store policy in continual learning.\n\n2. The writing of this paper is not clear enough to follow. Firstly, the explanation of the discrepancy between CIR and traditional incremental learning (i.e., class- and domain-incremental) is not clear. Figure 1 replaced with actual objects may give a more intuitive understanding of the authors' motivation. Moreover, it is necessary to use a coherent algorithm to introduce the overall procedure, however, absent in the main text. Finally, the conclusion part is tedious and should be refined.\n\n3. The lack of formulas in this paper makes comprehending have more barriers. For example, the generators build streams part lacks the complementary mathematical explanation. \n\n4. The authors should provide comprehensive theoretical analyses of their method, in my view, which is necessary for solid work.\n", "clarity,_quality,_novelty_and_reproducibility": "Overall, this paper is not high-quality work.\nClarity: This paper is not easy to follow. The figures in this paper don't help others understand the authors' motivation. For example, Figure 1 does not make sense in intuitively showing the difference between CIR and CI (or DI). Besides, the authors don't provide a visual explanation of the slot-based generator. The tedious conclusion part makes it difficult to grip the main contributions of this paper.\n\nNovelty: This paper is not rich in novelty, and lacks relevant theoretical analyses. The main contribution of this paper, in my view, is to build a CIR data stream. Recently, analogous works with frequency-aware replay are not rare.\n\nReproducibility: The authors promise to provide open-source code and the algorithm.\n", "summary_of_the_review": "This is not a high-quality paper without a clear expression and lack of solid theoretical analyses, which could not stimulate other research works. So, I recommend rejection.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666369043838}, {"id": "OmuUC31DAjz", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2894/Reviewer_U14g"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "For the possible repetition in Continual Learning, the paper makes contributes in the following three areas: (1) The author proposes two stochastic scenario generators that produce a wide range of CIR scenarios starting from a single dataset and a few control parameters. (2) The author conducts a comprehensive evaluation of repetition in CL by studying the behavior of existing CL strategies under different CIR scenarios. (3) The author presents a novel replay strategy that exploits repetition and counteracts the natural imbalance present in the stream.", "review_text": "Class-Incremental with Repetition (CIR) scenarios are important in realistic scenarios but have not been sufficiently studied so far. This paper proposes two practical CIR generators and performs a comprehensive series of comparative experiments of existing CL methods, which provides the CL area with a referenceable basis for further research. Therefore, it is recommended to accept this paper if the questions about the weaknesses can be answered.", "strengths": "Strengths: The paper proposes two practical CIR generators and performs a comprehensive series of comparative experiments of existing CL methods, which provides the CL area with a referenceable basis for further research. \n\nWeaknesses: The analysis of the FA storage policy is not sufficient. In the paper, the FA method is only proposed based on ER. Can the FA method be applied to other CL methods? How does the ER-FA method perform when the dataset is not unbalanced? Does FA cause a decrease in class accuracy for frequent classes? The author may want to respond to the weakness the reviewer raised.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "For the possible repetition in Continual Learning, the paper makes contributes in the following three areas: (1) The author proposes two stochastic scenario generators that produce a wide range of CIR scenarios starting from a single dataset and a few control parameters. (2) The author conducts a comprehensive evaluation of repetition in CL by studying the behavior of existing CL strategies under different CIR scenarios. (3) The author presents a novel replay strategy that exploits repetition and counteracts the natural imbalance present in the stream.", "strength_and_weaknesses": "Strengths: The paper proposes two practical CIR generators and performs a comprehensive series of comparative experiments of existing CL methods, which provides the CL area with a referenceable basis for further research. \n\nWeaknesses: The analysis of the FA storage policy is not sufficient. In the paper, the FA method is only proposed based on ER. Can the FA method be applied to other CL methods? How does the ER-FA method perform when the dataset is not unbalanced? Does FA cause a decrease in class accuracy for frequent classes? The author may want to respond to the weakness the reviewer raised.\n", "clarity,_quality,_novelty_and_reproducibility": "In terms of algorithms, the paper describes the steps of the algorithms in detail so that the algorithms can be clearly understood.\n\nIn terms of experimental results, the paper provides multiple experiments in various settings and uses high-quality graphs to present the experimental results.\n\nIn terms of novelty, the paper proposes two novel CLR generators and an improved FA method on ER method.\n\nIn terms of reproducibility, the author promises to provide an open-source implementation of our generators and algorithms with the scripts needed to reproduce the results reported in the paper. \n\nIn summary, the paper is well above average in terms of clarity, quality, novelty, and reproducibility. \n", "summary_of_the_review": "Class-Incremental with Repetition (CIR) scenarios are important in realistic scenarios but have not been sufficiently studied so far. This paper proposes two practical CIR generators and performs a comprehensive series of comparative experiments of existing CL methods, which provides the CL area with a referenceable basis for further research. Therefore, it is recommended to accept this paper if the questions about the weaknesses can be answered.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666253060865}], "openreview_url": "https://openreview.net/forum?id=ESR6hysKDsW", "arxiv_id": "2301.11396", "paper_pdf": "papers/ESR6hysKDsW.pdf", "paper_pdf_sha256": "1e207e1d3488917781efede8cde237c154479521c6254024c9734223b9d14d53", "paper_pdf_bytes": 1545202, "paper_pdf_source": "openreview", "code_url": "https://github.com/HamedHemati/CIR", "code_repository": "HamedHemati/CIR", "code_commit": "ca3305aba37e08ed1067f0da4847dba247c1569d", "code_archive": "repos/ESR6hysKDsW.zip", "code_archive_sha256": "8f99cff9675234bee1bd98559a4c68b9e043863f6d4decee5920f01f2c496e04", "code_archive_bytes": 829247, "code_file_count": 47, "code_extensions": {".py": 33, ".ipynb": 7, ".sh": 7}, "github_disk_usage_kb": 877, "github_languages": {"Jupyter Notebook": 901350, "Python": 133802, "Shell": 12069}, "github_archived": true, "github_pushed_at": "2024-02-15T13:11:42Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/class-incremental-learning-with-repetition"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tJCwZBHm-jW", "year": 2022, "status": "rejected", "title": "Image2Point: 3D Point-Cloud Understanding with 2D Image Pretrained Models", "authors": ["Chenfeng Xu", "Shijia Yang", "Bohan Zhai", "Bichen Wu", "Xiangyu Yue", "Wei Zhan", "Peter Vajda", "Kurt Keutzer", "Masayoshi Tomizuka"], "authorids": ["~Chenfeng_Xu1", "~Shijia_Yang1", "~Bohan_Zhai1", "~Bichen_Wu1", "~Xiangyu_Yue1", "~Wei_Zhan2", "~Peter_Vajda1", "~Kurt_Keutzer3", "~Masayoshi_Tomizuka2"], "authors_source": "OpenReview API", "abstract": "3D point-clouds and 2D images are different visual representations of the physical world. While human vision can understand both representations, computer vision models designed for 2D image and 3D point-cloud understanding are quite different.\nOur paper explores the potential for transferring between these two representations by empirically investigating the feasibility of the transfer, the benefits of the transfer, and shedding light on why the transfer works.\nWe discovered that we can indeed use the same architecture and pretrained weights of a neural net model to understand both images and point-clouds. Specifically, we can transfer the pretrained image model to a point-cloud model by \\textit{inflating} 2D convolutional filters to 3D and then \\textbf{f}inetuning the \\textbf{i}mage-\\textbf{p}retrained models (FIP). \nWe discover that, surprisingly, models with minimal finetuning efforts --- only on input, output, and optionally batch normalization layers, can achieve competitive performance on 3D point-cloud classification, beating a wide range of point-cloud models that adopt task-specific architectures and use a variety of tricks. When finetuning the whole model, the performance further improves significantly. Meanwhile, we also find that FIP improves data efficiency, achieving up to 10.0 points top-1 accuracy gain on few-shot classification. It also speeds up the training of point-cloud models by up to 11.1x to reach a target accuracy.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "H5DHmJb_F9c", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4157/Reviewer_aajt"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduced \"inflation\" (2D CNN kernel to 3D one by repeating on the 3rd axis) to transform a 2D image pre-train backbone into a 3D version so that 3D point cloud task can be benefit from 2D image pretraining. Detailed experiments are performed to address this idea. Results show minimal finetuning efforts can achieve competitive performance on 3D tasks. The authors are kind of \"surprised\" by these results.", "review_text": "\nStrengths: \n1. Idea is good and helpful for the 3D point cloud field.\n2. The experiments are designed to test different advantage of this idea. The results support the arguments. \n\nWeaknesses / questions: \n1. Since such idea has already used in other 2D-3D domain, the novelty is limited. Such paper needs extra emphasis on discussing why it works and ablation studies. \n  - In discussion section, the author provides some experiments, but the result seems quite random (figure 4). The \"shape representation\" are better transferred seems make sense, but why? The author tried to explain the finding but seems not quite convincing.\n  - When designing inflation, is there a difference on which axis to inflate. x, y or z axis. This is different to video since 3 axis is symmetric. \n2. The biggest concern is the parameter sizes and network structure.\n  - In 4.1, the author did not compare the parameter size. I doubt the fairness of the comparison with the baseline since it is possible that the performance gain is fully from the increasing parameters. \n  - In 4.2, there is a possibility that the ResNet structure is not good structure to train on point cloud making it quite bad on scratch training. For pointnet++, the training method is quite strange. \n3. The writing for this paper is interesting... not formal enough. The author used 10 \"surprising(ly)\" in this paper...", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduced \"inflation\" (2D CNN kernel to 3D one by repeating on the 3rd axis) to transform a 2D image pre-train backbone into a 3D version so that 3D point cloud task can be benefit from 2D image pretraining. Detailed experiments are performed to address this idea. Results show minimal finetuning efforts can achieve competitive performance on 3D tasks. The authors are kind of \"surprised\" by these results.", "main_review": "\nStrengths: \n1. Idea is good and helpful for the 3D point cloud field.\n2. The experiments are designed to test different advantage of this idea. The results support the arguments. \n\nWeaknesses / questions: \n1. Since such idea has already used in other 2D-3D domain, the novelty is limited. Such paper needs extra emphasis on discussing why it works and ablation studies. \n  - In discussion section, the author provides some experiments, but the result seems quite random (figure 4). The \"shape representation\" are better transferred seems make sense, but why? The author tried to explain the finding but seems not quite convincing.\n  - When designing inflation, is there a difference on which axis to inflate. x, y or z axis. This is different to video since 3 axis is symmetric. \n2. The biggest concern is the parameter sizes and network structure.\n  - In 4.1, the author did not compare the parameter size. I doubt the fairness of the comparison with the baseline since it is possible that the performance gain is fully from the increasing parameters. \n  - In 4.2, there is a possibility that the ResNet structure is not good structure to train on point cloud making it quite bad on scratch training. For pointnet++, the training method is quite strange. \n3. The writing for this paper is interesting... not formal enough. The author used 10 \"surprising(ly)\" in this paper...", "summary_of_the_review": "This is an interesting idea. This paper provided detailed experiments to test different advantage of this idea. However, the discussion on why it work seems not quite convincing. Also, part of the experiments missed some details. Some conclusions may not correct.\nWe expect the author provided more convincing discussion and more comprehensive comparison detail.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636360496802}, {"id": "qSxlT6S-p13", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4157/Reviewer_DFK1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed a pipeline for transferring convolutional network weights that are pre-trained on 2D images to 3D convolution networks. The proposed approach is to inflate the 2D kernels to 3D kernels similar to video models. The experiments are conducted on 2D image datasets such as ImageNet, and 3D datasets such as ModelNet.", "review_text": "Strength:\n\n- This paper investigates a new problem of transferring 2D pre-training to 3D.\n- The presentation is clear and easy to follow.\n- The experiments are extensive. The performance gain using the 2D pre-training is notable and non-trivial, showing the proposed weight transferring technique has the ability to convert useful learned information on 2D to 3D.\n\nWeakness:\n\n- It's good that the authors showed some visualizations to help understand the proposed framework. However, as the authors pointed out, the visualization cannot show why the transfer works or what information can be transferred.\n- The last sentence on page 8 is questionable: the conclusion of \"shape representations are better transferred from image to point-cloud\" cannot be inferred from overall dataset performances. Also, what does \"shape representations\" mean? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed a pipeline for transferring convolutional network weights that are pre-trained on 2D images to 3D convolution networks. The proposed approach is to inflate the 2D kernels to 3D kernels similar to video models. The experiments are conducted on 2D image datasets such as ImageNet, and 3D datasets such as ModelNet.", "main_review": "Strength:\n\n- This paper investigates a new problem of transferring 2D pre-training to 3D.\n- The presentation is clear and easy to follow.\n- The experiments are extensive. The performance gain using the 2D pre-training is notable and non-trivial, showing the proposed weight transferring technique has the ability to convert useful learned information on 2D to 3D.\n\nWeakness:\n\n- It's good that the authors showed some visualizations to help understand the proposed framework. However, as the authors pointed out, the visualization cannot show why the transfer works or what information can be transferred.\n- The last sentence on page 8 is questionable: the conclusion of \"shape representations are better transferred from image to point-cloud\" cannot be inferred from overall dataset performances. Also, what does \"shape representations\" mean? \n", "summary_of_the_review": "This paper proposed a solution to a new problem of transferring 2D weights to 3D. The solution works well on many dataset pairs. Though some claims made in the paper are questionable, this paper deserves to be published at the conference.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635910982866}, {"id": "z1SBMPf8Gw_", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4157/Reviewer_KdMv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes to transfer the 2D image features for 3D point cloud understanding tasks. Specifically, the proposed method first inflates the 2D convolution filters from a pretrained model to 3D and then optimizes only input, output and optionally batch normalization layer.  Detailed experiments show that the proposed finetuned-image-pretrained models can improve the performance and data efficiency.\n", "review_text": "Strengths\n- The paper proposes a model to connect 2D and 3D representations in transfer learning.\n- Detailed experiments show that the proposed model improves the performance and data efficiency compared to training from scratch.\n\nWeaknesses\n- Will larger dataset leads to better performance? A lot of comparisons between results from datasets of different size are missing in Table 1. \n- The paper shows that using 2D pretrained features improve performance. However, results of state-of-the-art task specific methods should also be listed. In addition, it would be better to compare with different self-supervised learning method.\nFor example, for shape classification on ModelNet40:\nFoldingnet: Point cloud auto-encoder via deep grid deformation. CVPR 2018\nPointcontrast: Unsupervised pretraining for 3d point cloud understanding. ECCV 2020\n-  Explanation why this kind of inflating 2D filter to 3D is reasonable. There is a huge domain gap between 2D dataset and 3D dataset. It's still unclear why this kind of transfer is reasonable. It would be better to add theoretical analysis besides experiments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes to transfer the 2D image features for 3D point cloud understanding tasks. Specifically, the proposed method first inflates the 2D convolution filters from a pretrained model to 3D and then optimizes only input, output and optionally batch normalization layer.  Detailed experiments show that the proposed finetuned-image-pretrained models can improve the performance and data efficiency.\n", "main_review": "Strengths\n- The paper proposes a model to connect 2D and 3D representations in transfer learning.\n- Detailed experiments show that the proposed model improves the performance and data efficiency compared to training from scratch.\n\nWeaknesses\n- Will larger dataset leads to better performance? A lot of comparisons between results from datasets of different size are missing in Table 1. \n- The paper shows that using 2D pretrained features improve performance. However, results of state-of-the-art task specific methods should also be listed. In addition, it would be better to compare with different self-supervised learning method.\nFor example, for shape classification on ModelNet40:\nFoldingnet: Point cloud auto-encoder via deep grid deformation. CVPR 2018\nPointcontrast: Unsupervised pretraining for 3d point cloud understanding. ECCV 2020\n-  Explanation why this kind of inflating 2D filter to 3D is reasonable. There is a huge domain gap between 2D dataset and 3D dataset. It's still unclear why this kind of transfer is reasonable. It would be better to add theoretical analysis besides experiments.", "summary_of_the_review": "Although the experiments in the paper show that the proposed model can improve the performance and data efficiency compared to training from scratch, the overall experiments results are not strong enough. In addition, it's still unclear why this kind of 2D to 3D transfer is reasonable. I am inclined to reject.\n\nUpdate: After reading the feedbacks from the author, I raise my score to 6.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635897052128}, {"id": "uglJT9PfIGh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4157/Reviewer_2iHv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper described an experiment of transfer learning between image and point cloud data. Despite of apparent dissimilarity between the two domain, the author was able to use structure and weights from pertained models in image domain to get reasonable performance in point cloud tasks like classification and segmentation. The author also made comparison of using pertained weights versus training from scratch, the result shows training pertained weights improves data efficiency. The proposed method is also proven to be useful when the training data is limited by training under few-shot learning setting. \n", "review_text": "Pros: The idea of transfer learning between image and point cloud data seems to be novel and surprising.\n\nCons: \n1. The idea of inflating model weight from 2D to 3D is not new[1].  \n2. The performance on point cloud segmentation on SemanticKitti dataset, while is reasonably well, still has gap to the state-of-the-art(This one is minor)\n3. The effectiveness of pretrained weight would be more convincing if the author could do further ablation study to initialize new model with expanding subsets of layers and exploring the effect on model performance. It will indicate which part of the pertained model is really useful in this transferred learning setting.\n\n=====\nUpdate after rebuttal: After reading the author's response and the new experimental results amended by the author, I raise my score to 6.\n\n[1] Carreira, Joao, and Andrew Zisserman. \"Quo vadis, action recognition? a new model and the kinetics dataset.\" proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper described an experiment of transfer learning between image and point cloud data. Despite of apparent dissimilarity between the two domain, the author was able to use structure and weights from pertained models in image domain to get reasonable performance in point cloud tasks like classification and segmentation. The author also made comparison of using pertained weights versus training from scratch, the result shows training pertained weights improves data efficiency. The proposed method is also proven to be useful when the training data is limited by training under few-shot learning setting. \n", "main_review": "Pros: The idea of transfer learning between image and point cloud data seems to be novel and surprising.\n\nCons: \n1. The idea of inflating model weight from 2D to 3D is not new[1].  \n2. The performance on point cloud segmentation on SemanticKitti dataset, while is reasonably well, still has gap to the state-of-the-art(This one is minor)\n3. The effectiveness of pretrained weight would be more convincing if the author could do further ablation study to initialize new model with expanding subsets of layers and exploring the effect on model performance. It will indicate which part of the pertained model is really useful in this transferred learning setting.\n\n=====\nUpdate after rebuttal: After reading the author's response and the new experimental results amended by the author, I raise my score to 6.\n\n[1] Carreira, Joao, and Andrew Zisserman. \"Quo vadis, action recognition? a new model and the kinetics dataset.\" proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.", "summary_of_the_review": "Overall the paper has interesting experimental result on transfer learning between image and point cloud data. The authors did a good job explaining the idea, concepts, procedures and experiments, showing the effectiveness of the transferred model on performance. data efficiency and usefulness on few-shot setting. The experiment itself could be further extended though.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635895962148}, {"id": "1zCA49f-mw7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4157/Reviewer_SBBb"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposed to perform the 3D point cloud understanding based on the pretrained 2D image models. Simple inflation technical, same as the work in I3D, are used to convert the 2D convnet to 3D convnet.\n\nExtensive experiments performed on the point cloud classification and segmentation tasks demonstrate the effectiveness of the work.\n", "review_text": "Strength:\n\nThe paper is simple yet very interesting, which simple inflate the 2D convolution kernels to 3D kernels to learn the representation of the 3D point clouds. 2D images and 3D point cloud are of great differences. Thus, it seems to impractical to simply performed the inflation and make it work. However, the authors proposed one simple idea and demonstrate that the idea does work.  I like the simple idea and would like to recommend the acceptance.\n\n\nWeakness:\n1. Some detailed information is missing. Therefore, I am not able to follow the detailed process. In Table 1, the authors list the performances of pointnet++ and the performances. I am curious how to performance the inflation in point net++\n2. The authors perform the experiments on the classification and segmeantion. What about the performances on detection task.\n\n\n\n\n=========\nPost-rebuttal\n\nI have read the authors rebuttal as well as the other fellow reviewers' comments and the corresponding discussions. \n\nI agree with the other fellow reviewers' comment, 2D inflating to 3D has been studied in video action recognition. However, to me, inflating 2D CNN of image for 3D pointcloud is new. Therefore, I do think the paper is of novelty. \nAlso I agree with the other reviewers' comments that the discussion of why the proposed method works does not been clearly addressed. \n\n\nConsidering this, I am lowering my rating from 8 to 6.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposed to perform the 3D point cloud understanding based on the pretrained 2D image models. Simple inflation technical, same as the work in I3D, are used to convert the 2D convnet to 3D convnet.\n\nExtensive experiments performed on the point cloud classification and segmentation tasks demonstrate the effectiveness of the work.\n", "main_review": "Strength:\n\nThe paper is simple yet very interesting, which simple inflate the 2D convolution kernels to 3D kernels to learn the representation of the 3D point clouds. 2D images and 3D point cloud are of great differences. Thus, it seems to impractical to simply performed the inflation and make it work. However, the authors proposed one simple idea and demonstrate that the idea does work.  I like the simple idea and would like to recommend the acceptance.\n\n\nWeakness:\n1. Some detailed information is missing. Therefore, I am not able to follow the detailed process. In Table 1, the authors list the performances of pointnet++ and the performances. I am curious how to performance the inflation in point net++\n2. The authors perform the experiments on the classification and segmeantion. What about the performances on detection task.\n\n\n\n\n=========\nPost-rebuttal\n\nI have read the authors rebuttal as well as the other fellow reviewers' comments and the corresponding discussions. \n\nI agree with the other fellow reviewers' comment, 2D inflating to 3D has been studied in video action recognition. However, to me, inflating 2D CNN of image for 3D pointcloud is new. Therefore, I do think the paper is of novelty. \nAlso I agree with the other reviewers' comments that the discussion of why the proposed method works does not been clearly addressed. \n\n\nConsidering this, I am lowering my rating from 8 to 6.\n", "summary_of_the_review": "It seems to the reviewer that this is the first paper to use 2D image pertained models for 3D point cloud understanding. Although the idea is simple, the work is interesting and inspiring. Therefore, I am recommending the acceptance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635781546412}], "openreview_url": "https://openreview.net/forum?id=tJCwZBHm-jW", "arxiv_id": "2106.04180", "paper_pdf": "papers/tJCwZBHm-jW.pdf", "paper_pdf_sha256": "03bd1b1923a6b2f9b09a42124957ccb97ee999331e20e2d49d59ad4f01879307", "paper_pdf_bytes": 8133728, "paper_pdf_source": "openreview", "code_url": "https://github.com/chenfengxu714/image2point", "code_repository": "chenfengxu714/image2point", "code_commit": "6efba318bcd4316dbe91e3a0daec6d8c9769cc8f", "code_archive": "repos/tJCwZBHm-jW.zip", "code_archive_sha256": "eb71a80727a67584ccc039dfcf8fbe20bfe0e3281b7c4648a3697fb9e3487b27", "code_archive_bytes": 647661, "code_file_count": 27, "code_extensions": {".py": 27}, "github_disk_usage_kb": 4153, "github_languages": {"Python": 136558}, "github_archived": false, "github_pushed_at": "2022-11-16T23:56:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/image2point-3d-point-cloud-understanding-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Ns8v4jHGyAV", "year": 2021, "status": "rejected", "title": "Matrix Shuffle-Exchange Networks for Hard 2D Tasks", "authors": ["Emīls Ozoliņš", "Karlis Freivalds", "Agris Šostaks"], "authorids": ["~Emīls_Ozoliņš1", "~Karlis_Freivalds1", "agris.sostaks@lumii.lv"], "authors_source": "OpenReview API", "abstract": "Convolutional neural networks have become the main tools for processing two-dimensional data. They work well for images, yet convolutions have a limited receptive field that prevents its applications to more complex 2D tasks. We propose a new neural model, called Matrix Shuffle-Exchange network, that can efficiently exploit long-range dependencies in 2D data and has comparable speed to a convolutional neural network. It is derived from  Neural Shuffle-Exchange network and has $\\mathcal{O}( \\log{n})$ layers and $\\mathcal{O}( n^2 \\log{n})$ total time and space complexity for processing a $n \\times n$ data matrix. We show that the Matrix Shuffle-Exchange network is well-suited for algorithmic and logical reasoning tasks on matrices and dense graphs, exceeding convolutional and graph neural network baselines. Its distinct advantage is the capability of retaining full long-range dependency modelling when generalizing to larger instances -- much larger than could be processed with models equipped with a dense attention mechanism.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ZBXCK-41L_6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2949/AnonReviewer5"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work proposes the Neural Shuffle-Exchange Network to capture both local and global dependencies for 2D data. The idea extends the 1D Neural Shuffle-Exchange Network to its 2D application. The proposed method first converts 2D data to 1D following the Z-order, then apply several Quaternary Switch and Quaternary Shuffle layers, and finally convert the data back to 2D space. The experimental results show that the proposed method can obtain better performance with reasonable computational cost. \n\nStrengths:\n+ The proposed method is very interesting. It studies how to capture long-term dependencies for 2D data in an efficient way. \n+ It uses the Z-order curve to flatten 2D data into 1D. It can preserve locality information for features. \n+ The experimental results show the proposed method can obtain better performance.\n\nWeaknesses:\n-  This work extends the Neural Shuffle-Exchange Network 1D to 2D. Compared with the 1D NSE, its technical contribution is incremental, and the novelty is limited.  \n-  Even though the Z-order curve can locality information, it is not clear whether the spatial information in 2D data can be preserved. It should be discussed. \n-  The attention mechanism can learn non-local dependencies effectively. It should be compared in the experimental studies. In addition, it is true that the vanilla version of attention has O(n^4) complexity but there are several improved variants with competitive performance but lower computational cost.\n- The proposed method is also applied to graph data and is compared with GIN. However, for graph data, I would suggest conducting experiments on benchmark datasets.\n\n=====Update after rebuttal=====\n\nI have read the authors' rebuttal. I still believe the novelty is limited, and hence I keep my score unchanged. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #5", "review": "This work proposes the Neural Shuffle-Exchange Network to capture both local and global dependencies for 2D data. The idea extends the 1D Neural Shuffle-Exchange Network to its 2D application. The proposed method first converts 2D data to 1D following the Z-order, then apply several Quaternary Switch and Quaternary Shuffle layers, and finally convert the data back to 2D space. The experimental results show that the proposed method can obtain better performance with reasonable computational cost. \n\nStrengths:\n+ The proposed method is very interesting. It studies how to capture long-term dependencies for 2D data in an efficient way. \n+ It uses the Z-order curve to flatten 2D data into 1D. It can preserve locality information for features. \n+ The experimental results show the proposed method can obtain better performance.\n\nWeaknesses:\n-  This work extends the Neural Shuffle-Exchange Network 1D to 2D. Compared with the 1D NSE, its technical contribution is incremental, and the novelty is limited.  \n-  Even though the Z-order curve can locality information, it is not clear whether the spatial information in 2D data can be preserved. It should be discussed. \n-  The attention mechanism can learn non-local dependencies effectively. It should be compared in the experimental studies. In addition, it is true that the vanilla version of attention has O(n^4) complexity but there are several improved variants with competitive performance but lower computational cost.\n- The proposed method is also applied to graph data and is compared with GIN. However, for graph data, I would suggest conducting experiments on benchmark datasets.\n\n=====Update after rebuttal=====\n\nI have read the authors' rebuttal. I still believe the novelty is limited, and hence I keep my score unchanged. ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604795300489}, {"id": "lPoA46QwraO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2949/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper adapts the recently introduced Neural Shuffle-Exchange (NSE) network for 2D inputs. This is done by introducing two changes; flattening 2D input using a z-order curve, and using 4-to-4 switches rather than 2-to-2. Experiments on synthetic data show that the proposed model can outperform baselines.\n\n### Strengths\n\n1. The extension of NSE to 2D domains is an interesting research direction.\n2. The use of the Z-order curve to map from 2D to 1D such that locality is approximately preserved is interesting.\n3. The model performs well for the tasks explored here.\n\n### Weaknesses\n\n1. The overall contributions are limited. While flattening 2D data using a z-order curve is interesting, it is a well known technique for spatial indexing. The use of z-order curves is also not entirely unprecedented in the field. For example, [1, 2] also use z-order curves in the context of 2D data, and [3] in the context of 3D data.\n2. The paper states that naive flattening of 2D data is expensive since it results in long sequences that are very slow to process with a NSE. The paper then proposes a technique which flattens 2D into an equally long sequence. This is very confusing. I'm assuming that reported speed improvements are due to the use of 4-to-4 rather than 2-to-2 switch, is this correct? Wouldn't this technique also be applicable to the original NSE?\n3. The experiments are limited, consisting exclusively of synthetic data and a single baseline for each task. While an experiment involving images (CIFAR-10) is presented in the Appendix, the description is brief, and the model performs no better than a feed-forward network. This is concerning as CV seems like a natural domain for exploration, especially given overlapping motivation  between this work and recent literature exploring transformers in the vision domain [4, 5]. As such, while the results are promising, they are not enough to convince me of that Matrix-SE is likely to be useful for practical  problems of interest.\n4. The presentation of technical and experimental details could be improved.\n   1. I'm confused about how weights are shared. In particular, while the figure make it clear that weights are shared among Shuffle layers in the same block (similarly for Inverse Shuffle layers), I'm unsure if they are shared between units, although I suspect they are not.\n   2. The original NSE paper pads \"the input sequence to that length by placing the sequence at a random position and adding zeros on both ends.\" Is a similar technique used in this work? If not, and weights are not shared between units (Point 4.1), I'm confused about how the model can generalize to larger matrices.\n   3. The paper does not contain information about how hyper-parameters were tuned, or information about stability of results. How robust is the model to different hyper-parameter settings? Are reported metrics averaged over multiple runs?\n   4. It should be clarified in the paper that Z-Order curves are only approximately locality preserving and contain large discontinuous jumps.\n   5. In section 5.2, the paper states that \"Graph Isomorphism Network struggles to solve these tasks.\" However, on the hardest task, GIN outperforms Matrix-SE on triangle-finding on the largest graphs.\n\n### Recommendation\n\nI recommend rejection. While I believe the extension of NSE to 2D domains is a worthwhile pursuit, I do not believe this paper represents significant progress is this regard.\n\n### Minor Issues\n\n1. Section 5.1: \"All tasks, except matrix squaring, **has** simple O(n2) algorithms.\" => \"All tasks, except matrix squaring, **have** simple O(n2) algorithms.\"\n2. Section 5.2: \"a common **practise**\" =>  \"a common **practice**\"\n3. Inconsistent spacing is used around parenthesis.\n\n### References\n\n1. Zhang, Jianjin, et al. \"Z-order recurrent neural networks for video prediction.\" 2019 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2019.\n2. Kumar Jayaraman, Pradeep, et al. \"Quadtree convolutional neural networks.\" *Proceedings of the European Conference on Computer Vision (ECCV)*. 2018.\n3. Corcoran, Thomas, et al. \"A spatial mapping algorithm with applications in deep learning-based structure classification.\" arXiv preprint arXiv:1802.02532 (2018).\n4. Parmar, Niki, et al. \"Image transformer.\" *arXiv preprint arXiv:1802.05751* (2018).\n5. Child, Rewon, et al. \"Generating long sequences with sparse transformers.\" *arXiv preprint arXiv:1904.10509* (2019).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting direction, but more work needed", "review": "This paper adapts the recently introduced Neural Shuffle-Exchange (NSE) network for 2D inputs. This is done by introducing two changes; flattening 2D input using a z-order curve, and using 4-to-4 switches rather than 2-to-2. Experiments on synthetic data show that the proposed model can outperform baselines.\n\n### Strengths\n\n1. The extension of NSE to 2D domains is an interesting research direction.\n2. The use of the Z-order curve to map from 2D to 1D such that locality is approximately preserved is interesting.\n3. The model performs well for the tasks explored here.\n\n### Weaknesses\n\n1. The overall contributions are limited. While flattening 2D data using a z-order curve is interesting, it is a well known technique for spatial indexing. The use of z-order curves is also not entirely unprecedented in the field. For example, [1, 2] also use z-order curves in the context of 2D data, and [3] in the context of 3D data.\n2. The paper states that naive flattening of 2D data is expensive since it results in long sequences that are very slow to process with a NSE. The paper then proposes a technique which flattens 2D into an equally long sequence. This is very confusing. I'm assuming that reported speed improvements are due to the use of 4-to-4 rather than 2-to-2 switch, is this correct? Wouldn't this technique also be applicable to the original NSE?\n3. The experiments are limited, consisting exclusively of synthetic data and a single baseline for each task. While an experiment involving images (CIFAR-10) is presented in the Appendix, the description is brief, and the model performs no better than a feed-forward network. This is concerning as CV seems like a natural domain for exploration, especially given overlapping motivation  between this work and recent literature exploring transformers in the vision domain [4, 5]. As such, while the results are promising, they are not enough to convince me of that Matrix-SE is likely to be useful for practical  problems of interest.\n4. The presentation of technical and experimental details could be improved.\n   1. I'm confused about how weights are shared. In particular, while the figure make it clear that weights are shared among Shuffle layers in the same block (similarly for Inverse Shuffle layers), I'm unsure if they are shared between units, although I suspect they are not.\n   2. The original NSE paper pads \"the input sequence to that length by placing the sequence at a random position and adding zeros on both ends.\" Is a similar technique used in this work? If not, and weights are not shared between units (Point 4.1), I'm confused about how the model can generalize to larger matrices.\n   3. The paper does not contain information about how hyper-parameters were tuned, or information about stability of results. How robust is the model to different hyper-parameter settings? Are reported metrics averaged over multiple runs?\n   4. It should be clarified in the paper that Z-Order curves are only approximately locality preserving and contain large discontinuous jumps.\n   5. In section 5.2, the paper states that \"Graph Isomorphism Network struggles to solve these tasks.\" However, on the hardest task, GIN outperforms Matrix-SE on triangle-finding on the largest graphs.\n\n### Recommendation\n\nI recommend rejection. While I believe the extension of NSE to 2D domains is a worthwhile pursuit, I do not believe this paper represents significant progress is this regard.\n\n### Minor Issues\n\n1. Section 5.1: \"All tasks, except matrix squaring, **has** simple O(n2) algorithms.\" => \"All tasks, except matrix squaring, **have** simple O(n2) algorithms.\"\n2. Section 5.2: \"a common **practise**\" =>  \"a common **practice**\"\n3. Inconsistent spacing is used around parenthesis.\n\n### References\n\n1. Zhang, Jianjin, et al. \"Z-order recurrent neural networks for video prediction.\" 2019 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2019.\n2. Kumar Jayaraman, Pradeep, et al. \"Quadtree convolutional neural networks.\" *Proceedings of the European Conference on Computer Vision (ECCV)*. 2018.\n3. Corcoran, Thomas, et al. \"A spatial mapping algorithm with applications in deep learning-based structure classification.\" arXiv preprint arXiv:1802.02532 (2018).\n4. Parmar, Niki, et al. \"Image transformer.\" *arXiv preprint arXiv:1802.05751* (2018).\n5. Child, Rewon, et al. \"Generating long sequences with sparse transformers.\" *arXiv preprint arXiv:1904.10509* (2019).", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603906915734}, {"id": "8SOKJGKQ1lE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2949/AnonReviewer1"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: The paper proposes a network architecture called Matrix Shuffle-Exchange (Matrix-SE) that can learn many logical reasoning tasks on 2D data and graph. It has complexity O(n^2 log n) for 2D input of size n x n, which is much smaller than the complexity of naive attention applied to 2D data (O(n^4)). The proposed architecture is an adaptation of the Neural Shuffle-Exchange network architecture (Freivalds et al., 2019), moving from 1D to 2D data. This adaptation is done by using a Z-order iteration of the 2D input, then performing radix-4 shuffle and radix-4 exchange, instead of radix-2. This model is shown to be able to solve several hard tasks on 2D data, such as inferring algorithms on binary matrices (transpose, rotation, bitwise XOR, matrix squaring), graph operations (component labeling, triangle finding, transitivity), and solving Sudoku puzzles. The experiments show impressive results and the model's ability to generalize to test inputs of larger sizes than those in the training set.\n\n======\n\nStrengths:\n1. Wide variety of algorithmic and logical reasoning tasks. I enjoyed reading the experiment section, and the tasks are fun and creative.\n2. Impressive generalization on larger sizes. On those tasks mentioned above, the Matrix-SE model generalizes well to 2D arrays that have larger sizes than those in the training set, out-performing baselines (ResNet, Graph Isomorphism Network).\n3. Simple model design. The generalization from Neural Shuffle-Exchange network to Matrix-SE is natural and straightforward.\n4. The paper is well-written and easy to read.\n\nWeaknesses:\n1. Lack of theoretical characterization of the model. More concretely, what kind of operations can be represented by Matrix-SE? The theoretical motivation seems to be from the classic result that Benes networks can represent any permutation. However, it's not clear how expressive the proposed model is.\n2. More realistic tasks. Do the logical reasoning tasks in more realistic scenario, such as modeling social networks represented as graphs?\n\n======\n\nOverall, I vote for accepting. The proposed model is simple, can perform logical reasoning tasks on 2D data, and generalizes well beyond sizes that are in the training set.\n\n======\n\nAdditional feedback and questions:\n- Do all the matrices in section 5.1 have binary values?\n- How is accuracy defined in Table 1 and 2. Does the output matrix have to match the label exactly? Or is it accuracy per element?\n- In Section 5.4, why is Residual-SE so much slower (9x) than Matrix-SE? I think it should only be 2x slower, because the depth of Residual-SE is twice that of the Matrix-SE.\n\n==== After rebuttal: Thank you for clarifying details. I maintain my rating. \nShowing that the inductive bias from Z-order curve and shuffle-exchange allows generalization to larger sizes is very interesting. Overall I vote for accepting.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Simple model that performs well on logical reasoning tasks on 2D data, and generalizes to larger sizes than trained.", "review": "Summary: The paper proposes a network architecture called Matrix Shuffle-Exchange (Matrix-SE) that can learn many logical reasoning tasks on 2D data and graph. It has complexity O(n^2 log n) for 2D input of size n x n, which is much smaller than the complexity of naive attention applied to 2D data (O(n^4)). The proposed architecture is an adaptation of the Neural Shuffle-Exchange network architecture (Freivalds et al., 2019), moving from 1D to 2D data. This adaptation is done by using a Z-order iteration of the 2D input, then performing radix-4 shuffle and radix-4 exchange, instead of radix-2. This model is shown to be able to solve several hard tasks on 2D data, such as inferring algorithms on binary matrices (transpose, rotation, bitwise XOR, matrix squaring), graph operations (component labeling, triangle finding, transitivity), and solving Sudoku puzzles. The experiments show impressive results and the model's ability to generalize to test inputs of larger sizes than those in the training set.\n\n======\n\nStrengths:\n1. Wide variety of algorithmic and logical reasoning tasks. I enjoyed reading the experiment section, and the tasks are fun and creative.\n2. Impressive generalization on larger sizes. On those tasks mentioned above, the Matrix-SE model generalizes well to 2D arrays that have larger sizes than those in the training set, out-performing baselines (ResNet, Graph Isomorphism Network).\n3. Simple model design. The generalization from Neural Shuffle-Exchange network to Matrix-SE is natural and straightforward.\n4. The paper is well-written and easy to read.\n\nWeaknesses:\n1. Lack of theoretical characterization of the model. More concretely, what kind of operations can be represented by Matrix-SE? The theoretical motivation seems to be from the classic result that Benes networks can represent any permutation. However, it's not clear how expressive the proposed model is.\n2. More realistic tasks. Do the logical reasoning tasks in more realistic scenario, such as modeling social networks represented as graphs?\n\n======\n\nOverall, I vote for accepting. The proposed model is simple, can perform logical reasoning tasks on 2D data, and generalizes well beyond sizes that are in the training set.\n\n======\n\nAdditional feedback and questions:\n- Do all the matrices in section 5.1 have binary values?\n- How is accuracy defined in Table 1 and 2. Does the output matrix have to match the label exactly? Or is it accuracy per element?\n- In Section 5.4, why is Residual-SE so much slower (9x) than Matrix-SE? I think it should only be 2x slower, because the depth of Residual-SE is twice that of the Matrix-SE.\n\n==== After rebuttal: Thank you for clarifying details. I maintain my rating. \nShowing that the inductive bias from Z-order curve and shuffle-exchange allows generalization to larger sizes is very interesting. Overall I vote for accepting.\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603886732978}], "openreview_url": "https://openreview.net/forum?id=Ns8v4jHGyAV", "arxiv_id": "2006.15892", "paper_pdf": "papers/Ns8v4jHGyAV.pdf", "paper_pdf_sha256": "9136d045c232fdb281cbe7a75fb2c890965df6d7e516c48cc8284355f5ecd69f", "paper_pdf_bytes": 530449, "paper_pdf_source": "openreview", "code_url": "https://github.com/LUMII-Syslab/Matrix-SE", "code_repository": "LUMII-Syslab/Matrix-SE", "code_commit": "f398589d23d973836b78cf7dd5cf0872bff42f9c", "code_archive": "repos/Ns8v4jHGyAV.zip", "code_archive_sha256": "bc0c75dd22d020a8a06a52ae8e85e45b7bb7507701f01342f15218377210b614", "code_archive_bytes": 469630, "code_file_count": 26, "code_extensions": {".py": 26}, "github_disk_usage_kb": 2526, "github_languages": {"Python": 120235}, "github_archived": false, "github_pushed_at": "2021-02-03T09:21:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/switchblade-a-neural-network-for-hard-2d"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DdHrylM8Tr", "year": 2026, "status": "rejected", "title": "AgentAlign: Navigating Safety Alignment in the Shift from Informative to Agentic Large Language Models", "authors": ["Jinchuan Zhang", "Lu Yin", "Yan Zhou", "Songlin Hu"], "authorids": ["~Jinchuan_Zhang1", "~Lu_Yin7", "~Yan_Zhou8", "~Songlin_Hu2"], "authors_source": "OpenReview API", "abstract": "The emergence of agentic capabilities in large language models fundamentally transforms their risk profile from passive information providers to autonomous action executors, introducing unprecedented safety challenges that existing alignment methods fail to address. Current approaches lack systematic frameworks for understanding and modeling the behavioral patterns underlying malicious agentic\nactivities, leading to brittle safety measures that collapse when confronted with multi-step harmful requests. We introduce AgentAlign, a novel behavioral modeling framework for agentic alignment that systematically captures malicious activity patterns through abstract behavior chains – structured representations of action sequences that characterize how harmful objectives are pursued across diverse tool-use scenarios. By instantiating these behavioral abstractions within comprehensive simulated environments, our framework enables scalable generation of authentic, executable training scenarios that preserve complex multi-step dynamics while avoiding real-world risks. Extensive evaluation across three model families demonstrates substantial safety improvements (35.8% to 79.5% enhancement in refusal rates) while maintaining or improving utility on benign tasks, significantly outperforming existing prompting-based defenses.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "I82abifj91", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5011/Reviewer_BDDN"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "The authors introduce AgentAlign, a framework that models malicious agent behaviors through \"abstract behavior chains\" - structured representations of multi-step harmful action sequences. These chains are instantiated in simulated environments to generate training data that balances safety and utility. Experiments across three model families show substantial improvements while maintaining performance on benign tasks.", "review_text": "The authors introduce AgentAlign, a framework that models malicious agent behaviors through \"abstract behavior chains\" - structured representations of multi-step harmful action sequences. These chains are instantiated in simulated environments to generate training data that balances safety and utility. Experiments across three model families show substantial improvements while maintaining performance on benign tasks.", "strengths": "1. Relevant problem: The safety gap between conversational and agentic LLMs is a real concern worth investigating.\n2. Systematic data generation: The abstract behavior chain framework provides a structured approach to generating multi-step harmful scenarios, which is important in AI safety research.", "weaknesses": "1. The method is missing critical comparisons. This is fundamentally a data generation method, yet there are no comparisons to: \n- Existing safety datasets: GuardSet-X [1], ToolAlign (only briefly mentioned in the related work), and other multi-step safety datasets. How does training on your dataset compare to training on these?\n- Guardrail systems: Why not compare against ShieldAgent [2], LlamaGuard, or other input filtering approaches? These operate at inference time without requiring model retraining. The paper doesn't justify why fine-tuning is necessary when you could simply filter inputs with an existing safety classifier.\n- Other data generation approaches: What about simple augmentation of existing red-teaming datasets? Or using LLMs to generate harmful agent scenarios with different prompting strategies?\n\n2. The evaluated models (GPT-4o, Qwen-2.5) are already outdated. More recent models should be evaluated.\n\n3. Transferability issue not addressed. If the agent is equipped with new sets of tools (or APIs), will the model still show a good refusal rate?\n\n[1] Kang, Mintong, et al. \"Guardset-x: Massive multi-domain safety policy-grounded guardrail dataset.\" arXiv preprint arXiv:2506.19054 (2025).\n[2]", "questions": "1. Why not compare to LlamaGuard or ShieldAgent as input filters? This seems like the most obvious baseline.\n2. How does performance compare when training on other existing safety datasets? \n3. What happens with completely different tool ecosystems? If I deploy your trained model with an entirely new set of APIs, does the safety transfer?\n4. Can you show this works on current frontier models? The models tested are outdated.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors introduce AgentAlign, a framework that models malicious agent behaviors through \"abstract behavior chains\" - structured representations of multi-step harmful action sequences. These chains are instantiated in simulated environments to generate training data that balances safety and utility. Experiments across three model families show substantial improvements while maintaining performance on benign tasks.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Relevant problem: The safety gap between conversational and agentic LLMs is a real concern worth investigating.\n2. Systematic data generation: The abstract behavior chain framework provides a structured approach to generating multi-step harmful scenarios, which is important in AI safety research.", "weaknesses": "1. The method is missing critical comparisons. This is fundamentally a data generation method, yet there are no comparisons to: \n- Existing safety datasets: GuardSet-X [1], ToolAlign (only briefly mentioned in the related work), and other multi-step safety datasets. How does training on your dataset compare to training on these?\n- Guardrail systems: Why not compare against ShieldAgent [2], LlamaGuard, or other input filtering approaches? These operate at inference time without requiring model retraining. The paper doesn't justify why fine-tuning is necessary when you could simply filter inputs with an existing safety classifier.\n- Other data generation approaches: What about simple augmentation of existing red-teaming datasets? Or using LLMs to generate harmful agent scenarios with different prompting strategies?\n\n2. The evaluated models (GPT-4o, Qwen-2.5) are already outdated. More recent models should be evaluated.\n\n3. Transferability issue not addressed. If the agent is equipped with new sets of tools (or APIs), will the model still show a good refusal rate?\n\n[1] Kang, Mintong, et al. \"Guardset-x: Massive multi-domain safety policy-grounded guardrail dataset.\" arXiv preprint arXiv:2506.19054 (2025).\n[2]", "questions": "1. Why not compare to LlamaGuard or ShieldAgent as input filters? This seems like the most obvious baseline.\n2. How does performance compare when training on other existing safety datasets? \n3. What happens with completely different tool ecosystems? If I deploy your trained model with an entirely new set of APIs, does the safety transfer?\n4. Can you show this works on current frontier models? The models tested are outdated.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761978338020}, {"id": "0pAM0pa35B", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5011/Reviewer_eCxY"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 4, "confidence": 3, "summary": "This paper proposes AgentAlign, a framework for improving the safety alignment of agentic large language models (LLMs) that can execute multi-step actions and use external tools. The key idea is to represent potential harmful or benign behaviors as Abstract Behavior Chains. The authors construct a simulation environment with 86 functional APIs across 9 tool categories to safely generate and validate a large-scale dataset (~18K samples) of harmful and benign agentic instructions. AgentAlign demonstrates significant improvements on the AgentHarm and ToolSword benchmarks. Specifically, the method substantially increases the refusal rate on harmful tasks while maintaining utility on benign tasks.", "review_text": "This paper proposes AgentAlign, a framework for improving the safety alignment of agentic large language models (LLMs) that can execute multi-step actions and use external tools. The key idea is to represent potential harmful or benign behaviors as Abstract Behavior Chains. The authors construct a simulation environment with 86 functional APIs across 9 tool categories to safely generate and validate a large-scale dataset (~18K samples) of harmful and benign agentic instructions. AgentAlign demonstrates significant improvements on the AgentHarm and ToolSword benchmarks. Specifically, the method substantially increases the refusal rate on harmful tasks while maintaining utility on benign tasks.", "strengths": "* Clear motivation and presentation: The paper provides a well-motivated discussion of the emerging safety challenges in agentic LLMs, supported by concrete examples and quantitative evidence. The writing is clear and easy to follow.\n* Originality: The idea of modeling safety through Abstract Behavior Chains is novel and insightful, as it captures multi-step harmful behaviors at the behavioral logic level rather than relying on surface text filters.\n* The proposed simulation environment and accompanying dataset are strong contributions, enabling safe and systematic synthesis of agentic tasks for alignment training. The semantic and execution validation framework is particularly rigorous and enhances data reliability.\n* The experiments are comprehensive, covering multiple open-source models and benchmarks (AgentHarm, ToolSword).", "weaknesses": "* While the paper is strong overall, it would benefit from a more comprehensive discussion of related work on plug-and-use safety guardrails for agents, such as GuardAgent, Conseca, and Agrail, to better position AgentAlign within this growing research space.\n* The training setup is not clearly described in the main text; readers may find it difficult to understand how the proposed dataset and objectives are applied during fine-tuning. Including a concise summary of the training process (currently only in the appendix) would significantly improve clarity.\n* Similarly, the simulation environment is central to the paper’s contribution, but its implementation details and accessibility are limited. Open-sourcing or providing more technical documentation on the environment and dataset would enhance reproducibility and impact.\n* On the empirical side, there is a slight drop in benign task performance for some models after applying AgentAlign, and results on Ministral and Qwen remain below Claude-3.5-Haiku. Moreover, it would strengthen the work to include comparisons against other representative guardrail systems such as Llama-Guard 3.", "questions": "Most of my questions overlap with the weaknesses mentioned above. In addition, I have one question regarding the data generation process:\n* Could the authors clarify why Claude-3.5-Haiku was chosen to generate refusal responses for harmful instructions, while Mistral-Large was used to generate benign trajectories?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes AgentAlign, a framework for improving the safety alignment of agentic large language models (LLMs) that can execute multi-step actions and use external tools. The key idea is to represent potential harmful or benign behaviors as Abstract Behavior Chains. The authors construct a simulation environment with 86 functional APIs across 9 tool categories to safely generate and validate a large-scale dataset (~18K samples) of harmful and benign agentic instructions. AgentAlign demonstrates significant improvements on the AgentHarm and ToolSword benchmarks. Specifically, the method substantially increases the refusal rate on harmful tasks while maintaining utility on benign tasks.", "soundness": 2, "presentation": 3, "contribution": 4, "strengths": "* Clear motivation and presentation: The paper provides a well-motivated discussion of the emerging safety challenges in agentic LLMs, supported by concrete examples and quantitative evidence. The writing is clear and easy to follow.\n* Originality: The idea of modeling safety through Abstract Behavior Chains is novel and insightful, as it captures multi-step harmful behaviors at the behavioral logic level rather than relying on surface text filters.\n* The proposed simulation environment and accompanying dataset are strong contributions, enabling safe and systematic synthesis of agentic tasks for alignment training. The semantic and execution validation framework is particularly rigorous and enhances data reliability.\n* The experiments are comprehensive, covering multiple open-source models and benchmarks (AgentHarm, ToolSword).", "weaknesses": "* While the paper is strong overall, it would benefit from a more comprehensive discussion of related work on plug-and-use safety guardrails for agents, such as GuardAgent, Conseca, and Agrail, to better position AgentAlign within this growing research space.\n* The training setup is not clearly described in the main text; readers may find it difficult to understand how the proposed dataset and objectives are applied during fine-tuning. Including a concise summary of the training process (currently only in the appendix) would significantly improve clarity.\n* Similarly, the simulation environment is central to the paper’s contribution, but its implementation details and accessibility are limited. Open-sourcing or providing more technical documentation on the environment and dataset would enhance reproducibility and impact.\n* On the empirical side, there is a slight drop in benign task performance for some models after applying AgentAlign, and results on Ministral and Qwen remain below Claude-3.5-Haiku. Moreover, it would strengthen the work to include comparisons against other representative guardrail systems such as Llama-Guard 3.", "questions": "Most of my questions overlap with the weaknesses mentioned above. In addition, I have one question regarding the data generation process:\n* Could the authors clarify why Claude-3.5-Haiku was chosen to generate refusal responses for harmful instructions, while Mistral-Large was used to generate benign trajectories?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761965623973}, {"id": "kXXbhMN6ba", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission5011/Reviewer_fXJL"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This work introduces AgentAlign, a new training framework to improve agent safety in which the attackers instruct the agents to execute malicious tasks. They show positive results where AgentAlign outperformed basic prompting defense baselines such as CoT, ReAct, and refusal prompts.", "review_text": "This work introduces AgentAlign, a new training framework to improve agent safety in which the attackers instruct the agents to execute malicious tasks. They show positive results where AgentAlign outperformed basic prompting defense baselines such as CoT, ReAct, and refusal prompts.", "strengths": "- Solid synthetic data generation pipeline with human validations.\n- Great diagrams and plots that help explain things clearly.\n- Showed great performance when compared to prompting baselines.", "weaknesses": "- Only compared to weak prompting baselines. For example, I'd appreciate it if you could at least add one common guardrail baseline, such as Llama Guard 4? This is a common defense method and will be helpful to include to see whether it complements AgentAlign or it's already very effective enough on the benchmarks evaluated. \n- No adversarial pressure studied. In reality, attackers will not give up after one try, and will likely apply the existing, common jailbreaking approach to the original harmful instructions. It'd be great to see a section that studies how robust this training approach is to for example, automated red teaming, multi-turn or decomposition attack, prefill attack, etc.\n- There is no limitation section. This work studies the single-turn attack for the agents, but it's unclear whether this works for multi-turn attack, etc. some papers you can cite in multi-turn or decomposition attacks include: (1) Monitoring Decomposition Attacks in LLMs with\nLightweight Sequential Monitors and (2) Breach By A Thousand Leaks: Unsafe Information Leakage in `Safe' AI Responses\n- No error bars were provided in any of the stats reported.", "questions": "Questions are basically the weakness i mentioned:\n- Could you include llama guard 4 as a baseline?\n- Could you test this approach on a few common jailbreaking techniques mentioned in Weakness 2?\n- Could you add a limitation section?\n- Could you at least report the standard errors for table 2? because the baseline and AgentAlign are quite close.\n\nHappy to raise my score if all of these are addressed, and AgentAlign still outperforms the new baseline and still holds the positive results under the adversarial pressure.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces AgentAlign, a new training framework to improve agent safety in which the attackers instruct the agents to execute malicious tasks. They show positive results where AgentAlign outperformed basic prompting defense baselines such as CoT, ReAct, and refusal prompts.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Solid synthetic data generation pipeline with human validations.\n- Great diagrams and plots that help explain things clearly.\n- Showed great performance when compared to prompting baselines.", "weaknesses": "- Only compared to weak prompting baselines. For example, I'd appreciate it if you could at least add one common guardrail baseline, such as Llama Guard 4? This is a common defense method and will be helpful to include to see whether it complements AgentAlign or it's already very effective enough on the benchmarks evaluated. \n- No adversarial pressure studied. In reality, attackers will not give up after one try, and will likely apply the existing, common jailbreaking approach to the original harmful instructions. It'd be great to see a section that studies how robust this training approach is to for example, automated red teaming, multi-turn or decomposition attack, prefill attack, etc.\n- There is no limitation section. This work studies the single-turn attack for the agents, but it's unclear whether this works for multi-turn attack, etc. some papers you can cite in multi-turn or decomposition attacks include: (1) Monitoring Decomposition Attacks in LLMs with\nLightweight Sequential Monitors and (2) Breach By A Thousand Leaks: Unsafe Information Leakage in `Safe' AI Responses\n- No error bars were provided in any of the stats reported.", "questions": "Questions are basically the weakness i mentioned:\n- Could you include llama guard 4 as a baseline?\n- Could you test this approach on a few common jailbreaking techniques mentioned in Weakness 2?\n- Could you add a limitation section?\n- Could you at least report the standard errors for table 2? because the baseline and AgentAlign are quite close.\n\nHappy to raise my score if all of these are addressed, and AgentAlign still outperforms the new baseline and still holds the positive results under the adversarial pressure.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760919616669}], "openreview_url": "https://openreview.net/forum?id=DdHrylM8Tr", "arxiv_id": "2505.23020", "paper_pdf": "papers/DdHrylM8Tr.pdf", "paper_pdf_sha256": "25f041c7e9edc5ae49d76dbc7c1318b3b3a4ed4484996d5ffe15b374702b1cb5", "paper_pdf_bytes": 601537, "paper_pdf_source": "openreview", "code_url": "https://github.com/jc-ryan/AgentAlign", "code_repository": "jc-ryan/AgentAlign", "code_commit": "efb728f7211163a39eacdeab77b4d4e0618a5bc7", "code_archive": "repos/DdHrylM8Tr.zip", "code_archive_sha256": "c25f2b40b8789c2f7c666427112b43ee0e6164f1b4bbed7f6c1e325f9698cc0f", "code_archive_bytes": 15115598, "code_file_count": 24, "code_extensions": {".py": 24}, "github_disk_usage_kb": 7934, "github_languages": {"Python": 289689}, "github_archived": false, "github_pushed_at": "2025-06-02T04:42:57Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/agentalign-navigating-safety-alignment-in-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WkpqUVcSTy", "year": 2025, "status": "rejected", "title": "SlowFast-LLaVA: A strong training-free baseline for video large language models", "authors": ["Mingze Xu", "Mingfei Gao", "Zhe Gan", "Hong-You Chen", "Zhengfeng Lai", "Haiming Gang", "Kai Kang", "Afshin Dehghan"], "authorids": ["~Mingze_Xu2", "~Mingfei_Gao1", "~Zhe_Gan1", "~Hong-You_Chen1", "~Zhengfeng_Lai1", "~Haiming_Gang1", "~Kai_Kang2", "~Afshin_Dehghan5"], "authors_source": "OpenReview API", "abstract": "We propose SlowFast-LLaVA (or SF-LLaVA for short), a training-free video large language model (LLM) that can jointly capture the detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is realized by using a two-stream SlowFast design of inputs for Video LLMs to aggregate features from sampled video frames in an effective way. Specifically, the Slow pathway extracts features at a low frame rate while keeping as many spatial details as possible (e.g., with 24x24 tokens), and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride (e.g., downsampling 6x) to focus on the motion cues. As a result, this design allows us to adequately capture both spatial and temporal features that are beneficial for understanding details in the video. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks. On some benchmarks, it achieves comparable or even better performance compared to state-of-the-art Video LLMs that are fine-tuned on video datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "urrSWeqwUd", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission610/Reviewer_Hn2a"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a training free method that converts an image MLLM into a video MLLM. This was achieved by simply concatenating features or frames in previous works. This paper proposes to concatenate features in a slow fast manner that combines a few high resolution feature maps with a large number of low resolution feature maps, which enabled LLM to see both local details and temporal context. The method SF-LLAVA achieves improved results on video MLLM benchmarks.", "review_text": "This paper proposes a training free method that converts an image MLLM into a video MLLM. This was achieved by simply concatenating features or frames in previous works. This paper proposes to concatenate features in a slow fast manner that combines a few high resolution feature maps with a large number of low resolution feature maps, which enabled LLM to see both local details and temporal context. The method SF-LLAVA achieves improved results on video MLLM benchmarks.", "strengths": "1. The proposed method is simple, intuitive, and efficient. It involves no training but obtains a video MLLM for free.\n2. Ablations are performed extensively on multiple datasets and benchmarks, focusing on the optimal selection of the slow and fast resolutions.", "weaknesses": "1. Limited novelty compared with existing methods such as IG-VLM, or PLLaVA. It has been shown that simply concatenating frames (or frame features) gives a good MLLM without or with only a little training. On top of these existing works, it is straightforward to see or expect that a cleverer selection of feature such as adopting the SlowFast idea can work a little bit better.\n2. Limited performance gain on top of the baselines such as IG-VLM, as shown in Table 2 (e.g. +1.1% on NextQA), as well as in the ablation, line 413, where adding the fast path improves marginally on top of the 12 frame slow only model.\n3. It is less well justified to keep “training free” as a requirement. It is true that training free is a nice property but existing training-based methods such as pllava and VILA involve only a lightweight fine-tuning stage that initializes mostly from image (LLaVA) model and learns only a projector or arguably simply concept selection/merging layers.\n4. Limited comparison with strong training based methods, such as VILA, and more comprehensive benchmarks such as MVBench or VideoMME. The benchmarks that are currently studied are either spatial biased (i.e. a 1-frame model can do quite well) or temporal-biased (EgoSchema that is based on high level semantics), rather than those that require both spatial and temporal reasoning.\n5. It may be interesting to see the difference in the order of input tokens. The current model looks like a bag of words model with the LLM doing most of the heavy lifting. Does the order of the tokens really matter? Does the model really learn motion? Do the datasets require/measure motion? This may give more insights to the community, if the goal is not to push the best results on video benchmarks.", "questions": "1. Since SF-LLAVA does not require training, is it possible to compute features offline to save some memory and thus use a lot more tokens than those methods that compute vision features during inference? In this case, is it possible to show that SF-LLAVA can work as a general add-on to any model that improves inference for any model? e.g. trained on 4 frames but can infer on many more frames with SF-Model?\n2. Does SF-LLaVA achieve good results on zero-shot generalization cases, compared with video trained methods such as pllava and vila? e.g. these methods are trained on specific datasets and therefore work well on those datasets that are similar to the training distribution, while SF-LLAVA works out of box and better on some datasets that are more different than the training sets of these existing papers.\n3. Following the discussion on weakness #3, is it possible to apply SF-LLaVA on larger scale models where video training is less feasible such as a 70B/90B llama3 model? This may be a real advantage of training-free vs. training-based methods.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a training free method that converts an image MLLM into a video MLLM. This was achieved by simply concatenating features or frames in previous works. This paper proposes to concatenate features in a slow fast manner that combines a few high resolution feature maps with a large number of low resolution feature maps, which enabled LLM to see both local details and temporal context. The method SF-LLAVA achieves improved results on video MLLM benchmarks.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The proposed method is simple, intuitive, and efficient. It involves no training but obtains a video MLLM for free.\n2. Ablations are performed extensively on multiple datasets and benchmarks, focusing on the optimal selection of the slow and fast resolutions.", "weaknesses": "1. Limited novelty compared with existing methods such as IG-VLM, or PLLaVA. It has been shown that simply concatenating frames (or frame features) gives a good MLLM without or with only a little training. On top of these existing works, it is straightforward to see or expect that a cleverer selection of feature such as adopting the SlowFast idea can work a little bit better.\n2. Limited performance gain on top of the baselines such as IG-VLM, as shown in Table 2 (e.g. +1.1% on NextQA), as well as in the ablation, line 413, where adding the fast path improves marginally on top of the 12 frame slow only model.\n3. It is less well justified to keep “training free” as a requirement. It is true that training free is a nice property but existing training-based methods such as pllava and VILA involve only a lightweight fine-tuning stage that initializes mostly from image (LLaVA) model and learns only a projector or arguably simply concept selection/merging layers.\n4. Limited comparison with strong training based methods, such as VILA, and more comprehensive benchmarks such as MVBench or VideoMME. The benchmarks that are currently studied are either spatial biased (i.e. a 1-frame model can do quite well) or temporal-biased (EgoSchema that is based on high level semantics), rather than those that require both spatial and temporal reasoning.\n5. It may be interesting to see the difference in the order of input tokens. The current model looks like a bag of words model with the LLM doing most of the heavy lifting. Does the order of the tokens really matter? Does the model really learn motion? Do the datasets require/measure motion? This may give more insights to the community, if the goal is not to push the best results on video benchmarks.", "questions": "1. Since SF-LLAVA does not require training, is it possible to compute features offline to save some memory and thus use a lot more tokens than those methods that compute vision features during inference? In this case, is it possible to show that SF-LLAVA can work as a general add-on to any model that improves inference for any model? e.g. trained on 4 frames but can infer on many more frames with SF-Model?\n2. Does SF-LLaVA achieve good results on zero-shot generalization cases, compared with video trained methods such as pllava and vila? e.g. these methods are trained on specific datasets and therefore work well on those datasets that are similar to the training distribution, while SF-LLAVA works out of box and better on some datasets that are more different than the training sets of these existing papers.\n3. Following the discussion on weakness #3, is it possible to apply SF-LLaVA on larger scale models where video training is less feasible such as a 70B/90B llama3 model? This may be a real advantage of training-free vs. training-based methods.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730693418975}, {"id": "VmGClkT364", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission610/Reviewer_jACj"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper proposes SlowFast-LLaVA (SF-LLaVA), a training-free video large language model (LLM) that can jointly capture detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is achieved by using a two-stream SlowFast design of inputs, where the Slow pathway extracts features at a low frame rate while keeping spatial detail, and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride to focus on motion cues. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks, and achieves comparable or even better performance compared to state-of-the-art video LLMs that are fine-tuned on video datasets.", "review_text": "The paper proposes SlowFast-LLaVA (SF-LLaVA), a training-free video large language model (LLM) that can jointly capture detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is achieved by using a two-stream SlowFast design of inputs, where the Slow pathway extracts features at a low frame rate while keeping spatial detail, and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride to focus on motion cues. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks, and achieves comparable or even better performance compared to state-of-the-art video LLMs that are fine-tuned on video datasets.", "strengths": "1. the proposed method to build video-llm is simple, novel, and the slowfast design to aggregate spatial and temportal information is reasonable.\n2.the paper is in good writing for understanding, and ablation experiments to provide detailed results for the influential factors of the provided method.", "weaknesses": "1. video-mllms could not only do recgnition and captioning of a video, but also exploite the knowledge of LLM to do further recognition reseasoning. In order to comprehensively evaluate the proposed training-free method to build a video-mllm. \n\na. I suggest authors provide evaluations not only on captioning benchmarks, such as MSVD, MSR-VTT and so on, but also on benchmarks specifically for video-mllm, such as video-mme, mvbench, videovista, MLVU and so on, which could provide more comprehensive video-llm performance comparation, the evaluation resutls could provide important new insights about the model's performance.\n\n2. the two-stream method is only applied on llava-next image-llm, if authors could give proof that this idea could be generalized to other image-llm models, the two-stream token aggregation design is more persuasive and thought-provoking for futher video-llm research. Here, I list some recent  image-llm models with structure different from llava, blip2(with q-former structure), internvl(with pixel shuffle structure), Ovis(Ovis: Structural Embedding Alignment for\n Multimodal Large Language Model, with visual embedding table), Aria（An Open Multimodal Native Mixture-of-Experts Model， with MoE structure）and suggest authors give test or theoretical analysis on these models.\n\n3.Althogh training-free is a good feature for the method, a training-based model could gain a more powerful performance, just as  LLaVA-NeXT-Video-DPO is better than SF-LLaVA(Both built on LLaVA-NeXT-Image ). \n\na. I suggest authors discuss specific use cases or scenarios where a training-free approach might be preferable despite potentially lower performance. \n\nb. I also suggest authors elaborate on potential trade-offs between training-free and fine-tuned approaches in terms of factors like deployment flexibility, computational requirements, or ability to handle novel domains.", "questions": "questions are listed above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes SlowFast-LLaVA (SF-LLaVA), a training-free video large language model (LLM) that can jointly capture detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is achieved by using a two-stream SlowFast design of inputs, where the Slow pathway extracts features at a low frame rate while keeping spatial detail, and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride to focus on motion cues. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks, and achieves comparable or even better performance compared to state-of-the-art video LLMs that are fine-tuned on video datasets.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. the proposed method to build video-llm is simple, novel, and the slowfast design to aggregate spatial and temportal information is reasonable.\n2.the paper is in good writing for understanding, and ablation experiments to provide detailed results for the influential factors of the provided method.", "weaknesses": "1. video-mllms could not only do recgnition and captioning of a video, but also exploite the knowledge of LLM to do further recognition reseasoning. In order to comprehensively evaluate the proposed training-free method to build a video-mllm. \n\na. I suggest authors provide evaluations not only on captioning benchmarks, such as MSVD, MSR-VTT and so on, but also on benchmarks specifically for video-mllm, such as video-mme, mvbench, videovista, MLVU and so on, which could provide more comprehensive video-llm performance comparation, the evaluation resutls could provide important new insights about the model's performance.\n\n2. the two-stream method is only applied on llava-next image-llm, if authors could give proof that this idea could be generalized to other image-llm models, the two-stream token aggregation design is more persuasive and thought-provoking for futher video-llm research. Here, I list some recent  image-llm models with structure different from llava, blip2(with q-former structure), internvl(with pixel shuffle structure), Ovis(Ovis: Structural Embedding Alignment for\n Multimodal Large Language Model, with visual embedding table), Aria（An Open Multimodal Native Mixture-of-Experts Model， with MoE structure）and suggest authors give test or theoretical analysis on these models.\n\n3.Althogh training-free is a good feature for the method, a training-based model could gain a more powerful performance, just as  LLaVA-NeXT-Video-DPO is better than SF-LLaVA(Both built on LLaVA-NeXT-Image ). \n\na. I suggest authors discuss specific use cases or scenarios where a training-free approach might be preferable despite potentially lower performance. \n\nb. I also suggest authors elaborate on potential trade-offs between training-free and fine-tuned approaches in terms of factors like deployment flexibility, computational requirements, or ability to handle novel domains.", "questions": "questions are listed above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730556188477}, {"id": "wl7xR8VXQV", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission610/Reviewer_yZMr"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper presents SF-LLaVA, a training-free video LLM that requires no additional fine-tuning to work for various video tasks.  Technically, the authors propose a SlowFast design of inputs for Video LLMs to capture both detailed spatial semantics and long-range temporal context, which is a simple and effective approach. The experiments on 3 video tasks with 8 benchmarks demonstrate the effectiveness and superiority of SF-LLaVA compared to training-free and SFT methods. Comprehensive ablation studies provide valuable insights for future research.", "review_text": "This paper presents SF-LLaVA, a training-free video LLM that requires no additional fine-tuning to work for various video tasks.  Technically, the authors propose a SlowFast design of inputs for Video LLMs to capture both detailed spatial semantics and long-range temporal context, which is a simple and effective approach. The experiments on 3 video tasks with 8 benchmarks demonstrate the effectiveness and superiority of SF-LLaVA compared to training-free and SFT methods. Comprehensive ablation studies provide valuable insights for future research.", "strengths": "+ The motivation for enabling LLMs to understand video inputs through a training-free approach is reasonable. \n\n+ The paper is easy to read and well-structured.\n\n+ Experiments on a diverse set of 8 video benchmarks demonstrate the effectiveness of SF-LLaVA.", "weaknesses": "- Experiments were conducted only on LLaVA-NeXT (despite including 7B and 34B models), with no additional results from other open-source LLMs provided.\n\n\n- The quantitative results in Figure 5 lack the results without the SlowFast design.", "questions": "-  Can the proposed method work for particularly long video understanding, such as LVBench[1]?\n\n-  The paper indicates that it lacks the capability to detect the precise start and end times. Can fine-grained frame-level textual descriptions be introduced to enhance the spatio-temporal understanding of videos?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents SF-LLaVA, a training-free video LLM that requires no additional fine-tuning to work for various video tasks.  Technically, the authors propose a SlowFast design of inputs for Video LLMs to capture both detailed spatial semantics and long-range temporal context, which is a simple and effective approach. The experiments on 3 video tasks with 8 benchmarks demonstrate the effectiveness and superiority of SF-LLaVA compared to training-free and SFT methods. Comprehensive ablation studies provide valuable insights for future research.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "+ The motivation for enabling LLMs to understand video inputs through a training-free approach is reasonable. \n\n+ The paper is easy to read and well-structured.\n\n+ Experiments on a diverse set of 8 video benchmarks demonstrate the effectiveness of SF-LLaVA.", "weaknesses": "- Experiments were conducted only on LLaVA-NeXT (despite including 7B and 34B models), with no additional results from other open-source LLMs provided.\n\n\n- The quantitative results in Figure 5 lack the results without the SlowFast design.", "questions": "-  Can the proposed method work for particularly long video understanding, such as LVBench[1]?\n\n-  The paper indicates that it lacks the capability to detect the precise start and end times. Can fine-grained frame-level textual descriptions be introduced to enhance the spatio-temporal understanding of videos?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730383558483}, {"id": "8u0LkaE7Hg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission610/Reviewer_zed3"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "Inspired by the idea of two stream networks in action recognition, this paper introduces a really simple sampling scheme of the frames and spatial pooling for training free evaluation of image LLM on videos. They show consistent improvements over the baseline on many tasks.", "review_text": "Inspired by the idea of two stream networks in action recognition, this paper introduces a really simple sampling scheme of the frames and spatial pooling for training free evaluation of image LLM on videos. They show consistent improvements over the baseline on many tasks.", "strengths": "The idea is extremely simple (which is a good thing) and battle tested in action recognition.\n\nThe evaluations and ablations seem comprehensive and fair.", "weaknesses": "The paper suffers from excessive repetition. Given the simplicity of the core idea, it seems unnecessarily stretched to fill 10 pages and could be more concise. Additionally, table captions, such as in Table 1, lack added value; they simply reiterate the table content without providing further insights. Similarly, the experiment descriptions merely repeat the accuracy values from the tables, without offering additional intuition or interpretation.\n\nFurthermore, the prompt aspect could be explored in greater depth, as it represents an additional axis to consider in zero-shot adaptation of vision-based large language models (LLMs) for videos.", "questions": "It seems that the prompt aspect in zero-shot adaptation could add substantial value if further explored. Have you considered additional prompt engineering or variations in the prompts to evaluate their impact on the model’s performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Inspired by the idea of two stream networks in action recognition, this paper introduces a really simple sampling scheme of the frames and spatial pooling for training free evaluation of image LLM on videos. They show consistent improvements over the baseline on many tasks.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The idea is extremely simple (which is a good thing) and battle tested in action recognition.\n\nThe evaluations and ablations seem comprehensive and fair.", "weaknesses": "The paper suffers from excessive repetition. Given the simplicity of the core idea, it seems unnecessarily stretched to fill 10 pages and could be more concise. Additionally, table captions, such as in Table 1, lack added value; they simply reiterate the table content without providing further insights. Similarly, the experiment descriptions merely repeat the accuracy values from the tables, without offering additional intuition or interpretation.\n\nFurthermore, the prompt aspect could be explored in greater depth, as it represents an additional axis to consider in zero-shot adaptation of vision-based large language models (LLMs) for videos.", "questions": "It seems that the prompt aspect in zero-shot adaptation could add substantial value if further explored. Have you considered additional prompt engineering or variations in the prompts to evaluate their impact on the model’s performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729952698907}], "openreview_url": "https://openreview.net/forum?id=WkpqUVcSTy", "arxiv_id": "2407.15841", "paper_pdf": "papers/WkpqUVcSTy.pdf", "paper_pdf_sha256": "aab8f27afbd00e073502aa56244af6190f70832ebf3507cd5c363979e1296cc8", "paper_pdf_bytes": 2435222, "paper_pdf_source": "openreview", "code_url": "https://github.com/apple/ml-slowfast-llava", "code_repository": "apple/ml-slowfast-llava", "code_commit": "bea0f73c106b91404ef403353e184b278fcc64a7", "code_archive": "repos/WkpqUVcSTy.zip", "code_archive_sha256": "58950ff192384e32750ad4f3698b8a232189e8020ae4ef3227b9ec21e8c678c4", "code_archive_bytes": 422991, "code_file_count": 52, "code_extensions": {".py": 38, ".sh": 14}, "github_disk_usage_kb": 384, "github_languages": {"Python": 165263, "Shell": 35913}, "github_archived": false, "github_pushed_at": "2024-09-16T21:44:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/slowfast-llava-a-strong-training-free"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XqLcFMMwNb", "year": 2024, "status": "rejected", "title": "MM-LDM: Multi-Modal Latent Diffusion Model for Sounding Video Generation", "authors": ["Mingzhen Sun", "Weining Wang", "Yanyuan Qiao", "Longteng Guo", "Jiahui Sun", "Xinxin Zhu", "Jing Liu"], "authorids": ["~Mingzhen_Sun1", "~Weining_Wang3", "~Yanyuan_Qiao1", "~Longteng_Guo1", "~Jiahui_Sun2", "~Xinxin_Zhu1", "~Jing_Liu1"], "authors_source": "OpenReview API", "abstract": "Sounding video generation (SVG) is a challenging audio-video joint generation task that requires both single-modal realism and cross-modal consistency.\nPrevious diffusion-based methods tackled SVG within the original signal space, resulting in a huge computation burden.\nIn this paper, we introduce a novel multi-modal latent diffusion model (MM-LDM), which establishes a perceptual latent space that is perceptually equivalent to the original audio-video signal space but drastically reduces computational complexity.\nWe unify the representation of audio and video signals and construct a shared high-level semantic feature space to bridge the information gap between audio and video modalities.\nFurthermore, we present a novel cross-modal sampling guidance that extends our generative models to audio-to-video and video-to-audio conditional generation tasks. \nWe obtain the new state-of-the-art results with significant quality and efficiency gains.\nIn particular, our method achieves an overall improvement in all evaluation metrics and a faster training and sampling speed.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "NdzM5l44H3", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission281/Reviewer_1VF7"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The present paper proposes a framework based on the latent diffusion model to address the challenge of audio-visual joint generation. In comparison to the baseline (MM-Diffusion), the generation scheme proposed in this article, which operates in the latent space, offers significantly enhanced accuracy and reduces the computational burden. Moreover, the application of contrastive loss is also more judicious than in previous methods.\nHowever, the author's writing exhibits some instances of ambiguity, which may suggest a lack of thorough understanding of the latent diffusion model. While the article's motivation and approach are commendable, I suggest that the author consider further revision and refinement of the manuscript before submitting it to the next conference.", "review_text": "The present paper proposes a framework based on the latent diffusion model to address the challenge of audio-visual joint generation. In comparison to the baseline (MM-Diffusion), the generation scheme proposed in this article, which operates in the latent space, offers significantly enhanced accuracy and reduces the computational burden. Moreover, the application of contrastive loss is also more judicious than in previous methods.\nHowever, the author's writing exhibits some instances of ambiguity, which may suggest a lack of thorough understanding of the latent diffusion model. While the article's motivation and approach are commendable, I suggest that the author consider further revision and refinement of the manuscript before submitting it to the next conference.", "strengths": "1. The application of diffusion in latent space has been demonstrated in the fields of image and video generation, thereby warranting its extension to multi-modal generation. This approach to model generation is highly relevant in the current context of machine learning and artificial intelligence, where multi-modal data is increasingly prevalent. By leveraging the power of diffusion in latent space, multi-modal models can be developed that can generate diverse outputs across various modalities. This approach offers significant benefits in terms of model robustness, scalability, and generalization, making it an attractive choice for businesses and academic researchers alike.\n\n2. The present study's results exhibit substantial enhancements from both qualitative and quantitative perspectives. The research outcomes demonstrate a significant improvement in both the quality and quantity of the study's output.\n\n3. The proposed module in this article has been validated through extensive ablation experiments. The results of these experiments indicate the module's efficacy in addressing the intended objectives.", "weaknesses": "1. I think the author's understanding of latent diffusion is not deep enough, and there are many unprofessional and unscientific descriptions in the writing process. For details, see Questions 1, 2 and 4.\n\n2. The sign of equation (7) is confusing. According to the paper, (n_a^t,n_v^t) are predicted noise features. But obviously this variable is not a predicted value, but a variable that satisfies the N(0,1) distribution.\n\n3. The Implementation Details only give the training details of the multi-modal autoencoder, but not the training details of the diffusion model. Or is it that the model in this paper does not need to first train an autoencoder and then perform diffusion training, like [1]?\n\n[1] Blattmann, Andreas, et al. \"Align your latents: High-resolution video synthesis with latent diffusion models.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.", "questions": "1. \"we can leverage pre-trained image diffusion models to be our signal decoders\". How is this step implemented? Are the \"signal decoders\" used here the diffusion unet in the image diffusion model? I don't understand how image diffusion model can act as decoder in autoencoder.\n\n2. What is the relationship between the content drawn in Figure 3(b) and T? Why does the multi-modal autoencoder perform a denoising process?\n\n3. In contrastive loss, how are positive and negative samples constructed, and how to define \"matched pairs\" during the implementation process?\n\n4. \"we utilize the ϵ-prediction to optimize our signal decoder, which involves the noise mean square error loss\". What is the relationship between the process of training autoencoder and the method of noise prediction?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The present paper proposes a framework based on the latent diffusion model to address the challenge of audio-visual joint generation. In comparison to the baseline (MM-Diffusion), the generation scheme proposed in this article, which operates in the latent space, offers significantly enhanced accuracy and reduces the computational burden. Moreover, the application of contrastive loss is also more judicious than in previous methods.\nHowever, the author's writing exhibits some instances of ambiguity, which may suggest a lack of thorough understanding of the latent diffusion model. While the article's motivation and approach are commendable, I suggest that the author consider further revision and refinement of the manuscript before submitting it to the next conference.", "soundness": "2 fair", "presentation": "1 poor", "contribution": "2 fair", "strengths": "1. The application of diffusion in latent space has been demonstrated in the fields of image and video generation, thereby warranting its extension to multi-modal generation. This approach to model generation is highly relevant in the current context of machine learning and artificial intelligence, where multi-modal data is increasingly prevalent. By leveraging the power of diffusion in latent space, multi-modal models can be developed that can generate diverse outputs across various modalities. This approach offers significant benefits in terms of model robustness, scalability, and generalization, making it an attractive choice for businesses and academic researchers alike.\n\n2. The present study's results exhibit substantial enhancements from both qualitative and quantitative perspectives. The research outcomes demonstrate a significant improvement in both the quality and quantity of the study's output.\n\n3. The proposed module in this article has been validated through extensive ablation experiments. The results of these experiments indicate the module's efficacy in addressing the intended objectives.", "weaknesses": "1. I think the author's understanding of latent diffusion is not deep enough, and there are many unprofessional and unscientific descriptions in the writing process. For details, see Questions 1, 2 and 4.\n\n2. The sign of equation (7) is confusing. According to the paper, (n_a^t,n_v^t) are predicted noise features. But obviously this variable is not a predicted value, but a variable that satisfies the N(0,1) distribution.\n\n3. The Implementation Details only give the training details of the multi-modal autoencoder, but not the training details of the diffusion model. Or is it that the model in this paper does not need to first train an autoencoder and then perform diffusion training, like [1]?\n\n[1] Blattmann, Andreas, et al. \"Align your latents: High-resolution video synthesis with latent diffusion models.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.", "questions": "1. \"we can leverage pre-trained image diffusion models to be our signal decoders\". How is this step implemented? Are the \"signal decoders\" used here the diffusion unet in the image diffusion model? I don't understand how image diffusion model can act as decoder in autoencoder.\n\n2. What is the relationship between the content drawn in Figure 3(b) and T? Why does the multi-modal autoencoder perform a denoising process?\n\n3. In contrastive loss, how are positive and negative samples constructed, and how to define \"matched pairs\" during the implementation process?\n\n4. \"we utilize the ϵ-prediction to optimize our signal decoder, which involves the noise mean square error loss\". What is the relationship between the process of training autoencoder and the method of noise prediction?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698823005402}, {"id": "snL6VDTBbE", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission281/Reviewer_ApFF"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "In this paper, the author studies the problem of sounding video generation (SVG) and proposes a novel multi-modality video generation model in latent space named MM-LDM. The idea of incorporating the latent diffusion model and SVG task is interesting and the overall result is promising.", "review_text": "In this paper, the author studies the problem of sounding video generation (SVG) and proposes a novel multi-modality video generation model in latent space named MM-LDM. The idea of incorporating the latent diffusion model and SVG task is interesting and the overall result is promising.", "strengths": "1. The idea of modeling audio and video in latent space for sounding video generation is interesting and promising.\n2. The writing is good and the results further demonstrate the effectiveness of the proposed method.", "weaknesses": "1. Similar ideas to the conditional generation section have been proposed in many papers which seems too weak to list as a technical contribution in the paper. I would like the author to claim this point as a \"bonus\" of the proposed model in the paper.\n2. The visual quality of MM-Diffusion results in Fig. 4 seems quite different from their original paper even considering the result has been super-resolved by the SR model. Is there any explanation for that? The visual quality of the results seems to be in low resolution or processed by some simple SR methods like nearest-neighbor.", "questions": "Please refer to the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the author studies the problem of sounding video generation (SVG) and proposes a novel multi-modality video generation model in latent space named MM-LDM. The idea of incorporating the latent diffusion model and SVG task is interesting and the overall result is promising.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The idea of modeling audio and video in latent space for sounding video generation is interesting and promising.\n2. The writing is good and the results further demonstrate the effectiveness of the proposed method.", "weaknesses": "1. Similar ideas to the conditional generation section have been proposed in many papers which seems too weak to list as a technical contribution in the paper. I would like the author to claim this point as a \"bonus\" of the proposed model in the paper.\n2. The visual quality of MM-Diffusion results in Fig. 4 seems quite different from their original paper even considering the result has been super-resolved by the SR model. Is there any explanation for that? The visual quality of the results seems to be in low resolution or processed by some simple SR methods like nearest-neighbor.", "questions": "Please refer to the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698723854259}, {"id": "VaP6TRNQ63", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission281/Reviewer_9ERg"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper introduces the Multi-Modal Latent Diffusion Model (MM-LDM) for Sounding Video Generation (SVG). The main contributions of this paper are two-fold:\n1. MM-LDM establishes audio and video latent spaces for SVG, which significantly reduces computational complexity.\n2. MM-LDM proposes a multi-modal autoencoder to compress video and audio signals from pixel space to a semantically shared latent space.\nThe proposed method achieves state-of-the-art results, demonstrating its effectiveness.", "review_text": "This paper introduces the Multi-Modal Latent Diffusion Model (MM-LDM) for Sounding Video Generation (SVG). The main contributions of this paper are two-fold:\n1. MM-LDM establishes audio and video latent spaces for SVG, which significantly reduces computational complexity.\n2. MM-LDM proposes a multi-modal autoencoder to compress video and audio signals from pixel space to a semantically shared latent space.\nThe proposed method achieves state-of-the-art results, demonstrating its effectiveness.", "strengths": "1. The authors attempt to solve a novel and valuable problem and design a reasonable framework for this purpose.\n2. The multimodal VAE designed by the author is interesting, establishing semantic latent spaces for audio and video modalities. Further, the authors use a shared multimodal decoder introduced in cross-modal alignment, which can inspire future multimodal generation.\n3. The experimental results are promising in metrics, demonstrating the effectiveness of the proposed method. In particular, the MM-LDM achieves state-of-the-art results.", "weaknesses": "1. The designing of a multimodal VAE is innovative, but it may not be as effective as that of two separate VAEs. The authors should compare their multimodal VAE with most direct audio and video VAEs, which would better demonstrate the effectiveness of multimodal VAE.\n2. I have viewed the generated results provided by the author, and only some results from the AIST++ dataset are available (MM-LDM: Multi-Modal Latent Diffusion Model for Sounding Video Generation (anonymouss765.github.io). However, the movements of the video characters are not natural, and due to the similarity of the generated audio, it is not clear whether the two modalities are in temporal alignment. Although the author has demonstrated significant success in metrics for their proposed method, it is necessary to supplement a human evaluation with MM-Diffusion to enhance credibility.\n3. The paper only does experiments on small datasets, and there may be serious overfitting problems for diffusion-based methods. The author should discuss the generalization of their method on larger datasets.", "questions": "1. In Table 2, parts “Multi-Modal Generative Models on Audio-to-Video Generation” and “Multi-Modal Generative Models on Video-to-Audio Generation”, the results of MM-Diffusion don’t take the other modality as a condition input.\n2. There is a typo in the heading of Table 3. “Latent Average Poolong” should be “Latent Average Pooling”.\n3. On page 5, in the section “Signal Decoding”, the authors state that they initialize their signal decoder with parameters of a pre-trained image diffusion model to reduce training time and enhance the quality of reconstruction. It is confusing to initialize the decoder with a diffusion model as they are for different objectives. I wonder if this initialization has positive benefits.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces the Multi-Modal Latent Diffusion Model (MM-LDM) for Sounding Video Generation (SVG). The main contributions of this paper are two-fold:\n1. MM-LDM establishes audio and video latent spaces for SVG, which significantly reduces computational complexity.\n2. MM-LDM proposes a multi-modal autoencoder to compress video and audio signals from pixel space to a semantically shared latent space.\nThe proposed method achieves state-of-the-art results, demonstrating its effectiveness.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The authors attempt to solve a novel and valuable problem and design a reasonable framework for this purpose.\n2. The multimodal VAE designed by the author is interesting, establishing semantic latent spaces for audio and video modalities. Further, the authors use a shared multimodal decoder introduced in cross-modal alignment, which can inspire future multimodal generation.\n3. The experimental results are promising in metrics, demonstrating the effectiveness of the proposed method. In particular, the MM-LDM achieves state-of-the-art results.", "weaknesses": "1. The designing of a multimodal VAE is innovative, but it may not be as effective as that of two separate VAEs. The authors should compare their multimodal VAE with most direct audio and video VAEs, which would better demonstrate the effectiveness of multimodal VAE.\n2. I have viewed the generated results provided by the author, and only some results from the AIST++ dataset are available (MM-LDM: Multi-Modal Latent Diffusion Model for Sounding Video Generation (anonymouss765.github.io). However, the movements of the video characters are not natural, and due to the similarity of the generated audio, it is not clear whether the two modalities are in temporal alignment. Although the author has demonstrated significant success in metrics for their proposed method, it is necessary to supplement a human evaluation with MM-Diffusion to enhance credibility.\n3. The paper only does experiments on small datasets, and there may be serious overfitting problems for diffusion-based methods. The author should discuss the generalization of their method on larger datasets.", "questions": "1. In Table 2, parts “Multi-Modal Generative Models on Audio-to-Video Generation” and “Multi-Modal Generative Models on Video-to-Audio Generation”, the results of MM-Diffusion don’t take the other modality as a condition input.\n2. There is a typo in the heading of Table 3. “Latent Average Poolong” should be “Latent Average Pooling”.\n3. On page 5, in the section “Signal Decoding”, the authors state that they initialize their signal decoder with parameters of a pre-trained image diffusion model to reduce training time and enhance the quality of reconstruction. It is confusing to initialize the decoder with a diffusion model as they are for different objectives. I wonder if this initialization has positive benefits.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698649154166}, {"id": "REEjLHrA48", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission281/Reviewer_n6co"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a multi-modal latent diffusion model named SVG for audio and video generation. Both audio and video signals are into latent spaces and then learn joint semantic features via classification and contrastive loss. The resulting semantic features can be used as conditional signals to improve audio-to-video and video-to-audio generation. The experiments were conducted on two sounding video datasets and the quantitative results are better than the baselines.", "review_text": "The paper proposes a multi-modal latent diffusion model named SVG for audio and video generation. Both audio and video signals are into latent spaces and then learn joint semantic features via classification and contrastive loss. The resulting semantic features can be used as conditional signals to improve audio-to-video and video-to-audio generation. The experiments were conducted on two sounding video datasets and the quantitative results are better than the baselines.", "strengths": "1) Audio-visual cross-modal generation is a challenging task. The authors proposed a promising approach to bridge the gap between audio and video efficiently.\n2) The quality of the generated samples looks good to me.\n3) The paper is easy to understand.", "weaknesses": "1) The paper is more like a straightforward follow-up on MM-Diffusion. The major difference is changing the diffusion targets from the raw signal domains to the latent spaces. However, based on the literature, it is almost trivial or obvious that transferring to latent space could achieve better results and improve training efficiency in any diffusion generation. From this perspective, the technical novelty of this paper is limited.\n2) The datasets used in the paper are pretty limited. The AIST and Landscape are small-size datasets. The proposed method could overfit the dataset, and indeed, diffusion is really good at overfitting. While the authors mentioned that they have not yet extended the method to open-domain sounding video datasets, I believe that is actually the critical research problem required to solve.\n3) While the quality of the generated samples on the webpage is nice for 1 second, I believe it is quite limited and probably hard to generalize for a longer time.\n4) In addition to quantitative metrics, I believe it is quite important to have a subjective evaluation of the generated samples, especially for audio.", "questions": "1) Have you tried doing applications on audio-visual continuation? For example, the model is conditioned on the first 1s audio, and then you can generate the next 1s audio and visual together. Based on your approach, it seems like these applications are also feasible.\n2) There exist so many losses and learnable embeddings within the audio-visual autoencoder. How do you search those weights of different losses? While there are ablation studies, it is only in one direction. I am interested in how you eventually could find the best combinations of all terms.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a multi-modal latent diffusion model named SVG for audio and video generation. Both audio and video signals are into latent spaces and then learn joint semantic features via classification and contrastive loss. The resulting semantic features can be used as conditional signals to improve audio-to-video and video-to-audio generation. The experiments were conducted on two sounding video datasets and the quantitative results are better than the baselines.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1) Audio-visual cross-modal generation is a challenging task. The authors proposed a promising approach to bridge the gap between audio and video efficiently.\n2) The quality of the generated samples looks good to me.\n3) The paper is easy to understand.", "weaknesses": "1) The paper is more like a straightforward follow-up on MM-Diffusion. The major difference is changing the diffusion targets from the raw signal domains to the latent spaces. However, based on the literature, it is almost trivial or obvious that transferring to latent space could achieve better results and improve training efficiency in any diffusion generation. From this perspective, the technical novelty of this paper is limited.\n2) The datasets used in the paper are pretty limited. The AIST and Landscape are small-size datasets. The proposed method could overfit the dataset, and indeed, diffusion is really good at overfitting. While the authors mentioned that they have not yet extended the method to open-domain sounding video datasets, I believe that is actually the critical research problem required to solve.\n3) While the quality of the generated samples on the webpage is nice for 1 second, I believe it is quite limited and probably hard to generalize for a longer time.\n4) In addition to quantitative metrics, I believe it is quite important to have a subjective evaluation of the generated samples, especially for audio.", "questions": "1) Have you tried doing applications on audio-visual continuation? For example, the model is conditioned on the first 1s audio, and then you can generate the next 1s audio and visual together. Based on your approach, it seems like these applications are also feasible.\n2) There exist so many losses and learnable embeddings within the audio-visual autoencoder. How do you search those weights of different losses? While there are ablation studies, it is only in one direction. I am interested in how you eventually could find the best combinations of all terms.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698645421138}], "openreview_url": "https://openreview.net/forum?id=XqLcFMMwNb", "arxiv_id": "2410.01594", "paper_pdf": "papers/XqLcFMMwNb.pdf", "paper_pdf_sha256": "2839738e9086a2bd55bf678ca29282ed4eef5f72ab62d796a7dd725431bba470", "paper_pdf_bytes": 3860841, "paper_pdf_source": "openreview", "code_url": "https://github.com/mzsun01/MM-LDM", "code_repository": "mzsun01/MM-LDM", "code_commit": "5db569aef6d3f637022a2dc6b7947cbc016f01c4", "code_archive": "repos/XqLcFMMwNb.zip", "code_archive_sha256": "7352b51b13776acf37ab580ccde85174c8434b87a219786a402aba4e20ab32e9", "code_archive_bytes": 271400, "code_file_count": 43, "code_extensions": {".py": 42, ".sh": 1}, "github_disk_usage_kb": 371, "github_languages": {"Python": 560854, "Shell": 166}, "github_archived": false, "github_pushed_at": "2024-04-12T07:34:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/mm-ldm-multi-modal-latent-diffusion-model-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "HZJje06x6IO", "year": 2023, "status": "rejected", "title": "Global Context Vision Transformers", "authors": ["Ali Hatamizadeh", "Hongxu Yin", "Jan Kautz", "Pavlo Molchanov"], "authorids": ["~Ali_Hatamizadeh1", "~Hongxu_Yin2", "~Jan_Kautz1", "~Pavlo_Molchanov1"], "authors_source": "OpenReview API", "abstract": "We propose global context vision transformer (GC ViT), a novel architecture that enhances parameter and compute utilization for computer vision tasks. The core of the novel model are  global context self-attention modules, joint with standard local self-attention, to effectively yet efficiently model both long and short-range spatial interactions, as an alternative to complex operations such as an attention masks or local windows shifting. While the local self-attention modules are responsible for modeling short-range information, the global query tokens are shared across all global self-attention modules to interact with local key and values. In addition, we address the lack of inductive bias in ViTs and improve the modeling of inter-channel dependencies by proposing a novel downsampler which leverages a parameter-efficient fused inverted residual block. The proposed GC ViT achieves new state-of-the-art performance across image classification, object detection and semantic segmentation tasks. On ImageNet-1K dataset for classification, the tiny, small and base variants of GC ViT with 28M, 51M and 90M parameters achieve 83.4%, 83.9% and 84.4% Top-1 accuracy, respectively, surpassing comparably-sized prior art such as CNN-based ConvNeXt and ViT-based Swin Transformer. Pre-trained GC ViT backbones in downstream tasks of object detection, instance segmentation, and semantic segmentation on MS COCO and ADE20K datasets outperform prior work consistently, sometimes by large margins.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "1OtJVS6_8I", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1237/Reviewer_YBSt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors propose a novel module to combine the local and global context. The key of this method is to introduce global query tokens into the local context module. The advantage of this method is that it can merge the context information from short and long-term ranges. The authors do experiments for image classification, detection, and segmentation tasks to show the effectiveness of the method.", "review_text": "Please follow the previous parts.", "strengths": "Strength\n1. The experiments show this method is effective and can be generalized to many different tasks.\n2. The problem is essential for network architecture design and the method is straightforward.\n3. The visualization results show the effectiveness of this method.\n\nWeakness\n1. The combination of local and global context is a long-standing problem and has been explored by many methods. The method for merging local and global contexts lacks novelty.\n2. The improvement between the proposed method and other SOTA methods is minor.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors propose a novel module to combine the local and global context. The key of this method is to introduce global query tokens into the local context module. The advantage of this method is that it can merge the context information from short and long-term ranges. The authors do experiments for image classification, detection, and segmentation tasks to show the effectiveness of the method.", "strength_and_weaknesses": "Strength\n1. The experiments show this method is effective and can be generalized to many different tasks.\n2. The problem is essential for network architecture design and the method is straightforward.\n3. The visualization results show the effectiveness of this method.\n\nWeakness\n1. The combination of local and global context is a long-standing problem and has been explored by many methods. The method for merging local and global contexts lacks novelty.\n2. The improvement between the proposed method and other SOTA methods is minor.", "clarity,_quality,_novelty_and_reproducibility": "This paper is easy to follow. The novelty of this paper is somehow limited. I think researchers can reproduce this paper easily.", "summary_of_the_review": "Please follow the previous parts.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666864541672}, {"id": "RD-aVJwq0HU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1237/Reviewer_t1Sn"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper proposed a new hirerchical transformers architecture. The paper revisit the attention mechanism and the downsampling layers.\nThe claim of the paper are the following: \n\n• We introduce a compute and parameter-optimized hierarchical ViT with reparametrization of the\ndesign space (e.g., embedding dimension, number of heads, MLP ratio).\n\n• We design an efficient CNN-like token generator that encodes spatial features at different resolutions\nfor global query representations.\n\n• We propose global query tokens that can effectively capture contextual information in an efficient\nmanner and model both local and global interactions.\n\n• We introduce a parameter-efficient downsampling module with modified Fused MB-Conv blocks\nthat not only integrates inductive bias but also enables the modeling of inter-channel dependencies.\n\n• We demonstrate new SOTA benchmarks for : (1) ImageNet classification with Pareto fronts on\nImageNet-1K for model size and FLOPs (see Fig. 1), and (2) downstream tasks such as detection,\ninstance segmentation and semantic segmentation on MS COCO and ADE20K, respectively.\n", "review_text": "The idea of the paper is interesting but unfortunately the contribution does not seem to be justified experimentally. The paper seems to be incremental. There is a lack of discussion and comparison with very similar approaches in the literature.\n\n\n====== Post rebuttal ====\n\nThe rebuttal raised a very important concern: \n\nThe authors report maximum accuracy during training by evaluating at each epoch in order to maximise performance on the ImageNet-1k validation set (there is no separate test set for ImageNet). Although the authors argue that it is a common practice, there is no evidence in the mentioned papers that this is the approach used. Furthermore I consider this approach to be detrimental to the community as it puts scientific understanding behind the optimal performance.\n\nOn the 3 logs provided the impact of this practice is important on the log with the bigger model +0.2% and impact the conclusion. The authors no longer have the other logs of their experiments, especially those of the ablation, so it is impossible to know the impact of the method. There is no multi-seed evaluation that shows that 0.1% on ImageNet-v2 is significant although it is claimed in the rebuttal.\n\nIn the current state the experiments are mainly designed to optimise the performance of the GC ViT architecture at the expense of an understanding of the impact of each component of the proposed method.\n\nSo I will lower my score.", "strengths": "Strength\n- The paper is well written and easy to follow\n- There is evaluation on different tasks\n\nWeaknesses:\n- Missing comparaison and discussion:\n\na) The global query tokens claim: The proposed global query token behaviour seem to be quite similar to the attention pattern mechanism in the BigBird paper[1] and also the local global attention in the paper EdgeViT[2]. It is necessary to discuss the contribution of the proposed method to these two papers. Because at present the contribution of the paper on this aspect is not clear.\n\n[1] Zaheer et al., Big Bird: Transformers for Longer Sequences\n[2] Pan et al., EdgeViTs: Competing Light-weight CNNs on Mobile Devices with Vision Transformers\n\nb) The compute and parameter-optimized hierarchical ViT and SOTA claim: in Figure 1 and Table 1: Indeed, the EdgeViT paper exploit similar ideas (local-global attention, Hierachical architecture) and seem to have better FLOPs-accuracy and Parameter accuracy trade-off (EdgeViT-S: 11.1M params, 1.9G FLOPs, 81.0 top-1 on ImageNet vs GC ViT-XXT 12M params, 2.1G FLOPs, 79.8 top-1 on ImageNet)  It is therefore necessary to add this architecture to the comparison. Moreover, the paper mention EfficientNetV2[1] because GC ViT exploit some similar blocks and seem also better/similar for the FLOPs-accuracy and parameter accuracy trade-off (EfficientNetV2-M 54M params, 24G flops, 85.1 top-1 on ImageNet  vs GC ViT-L  201M params, 32.6 G FLOPs, 84.6 top-1 ImageNet). So it's important to add also EfficinetNet-v2 in the comparaison.\n\nGiven these elements it is probably best to down tone the state-of-the art claims in the paper. Because it is not clear which state of the art it is.\n\n[1] Tan et al., EfficientNetV2: Smaller Models and Faster Training\n\nb) The efficient CNN-like token generator: Many papers have proposed convolution blocks for token generation like LeViT[1] or XCiT[2].\nThere is no discussion or comparison of performance with the previous approach. It is therefore impossible to evaluate the contribution of the paper on this point.\n\n[1] Graham et al., LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference\n[2] El-Nouby et al., XCiT: Cross-Covariance Image Transformers\n\n- Overfitting evaluation: For the ImageNet dataset there is no separate test and validation set so it's important to evaluate the level of overfitting by doing evaluation of the models on ImageNet-v2[1].\n\n[1] Recht et al., are we done with ImageNet?\n\n- Missing metrics: There is only params and FLOPs in the different table. It's important to measure other trade-off like Latency and memory consumption. Indeed, the proposed architecture use MBconv which are known to be good on trade-off FLOPs accuracy and parameter accuracy.\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposed a new hirerchical transformers architecture. The paper revisit the attention mechanism and the downsampling layers.\nThe claim of the paper are the following: \n\n• We introduce a compute and parameter-optimized hierarchical ViT with reparametrization of the\ndesign space (e.g., embedding dimension, number of heads, MLP ratio).\n\n• We design an efficient CNN-like token generator that encodes spatial features at different resolutions\nfor global query representations.\n\n• We propose global query tokens that can effectively capture contextual information in an efficient\nmanner and model both local and global interactions.\n\n• We introduce a parameter-efficient downsampling module with modified Fused MB-Conv blocks\nthat not only integrates inductive bias but also enables the modeling of inter-channel dependencies.\n\n• We demonstrate new SOTA benchmarks for : (1) ImageNet classification with Pareto fronts on\nImageNet-1K for model size and FLOPs (see Fig. 1), and (2) downstream tasks such as detection,\ninstance segmentation and semantic segmentation on MS COCO and ADE20K, respectively.\n", "strength_and_weaknesses": "Strength\n- The paper is well written and easy to follow\n- There is evaluation on different tasks\n\nWeaknesses:\n- Missing comparaison and discussion:\n\na) The global query tokens claim: The proposed global query token behaviour seem to be quite similar to the attention pattern mechanism in the BigBird paper[1] and also the local global attention in the paper EdgeViT[2]. It is necessary to discuss the contribution of the proposed method to these two papers. Because at present the contribution of the paper on this aspect is not clear.\n\n[1] Zaheer et al., Big Bird: Transformers for Longer Sequences\n[2] Pan et al., EdgeViTs: Competing Light-weight CNNs on Mobile Devices with Vision Transformers\n\nb) The compute and parameter-optimized hierarchical ViT and SOTA claim: in Figure 1 and Table 1: Indeed, the EdgeViT paper exploit similar ideas (local-global attention, Hierachical architecture) and seem to have better FLOPs-accuracy and Parameter accuracy trade-off (EdgeViT-S: 11.1M params, 1.9G FLOPs, 81.0 top-1 on ImageNet vs GC ViT-XXT 12M params, 2.1G FLOPs, 79.8 top-1 on ImageNet)  It is therefore necessary to add this architecture to the comparison. Moreover, the paper mention EfficientNetV2[1] because GC ViT exploit some similar blocks and seem also better/similar for the FLOPs-accuracy and parameter accuracy trade-off (EfficientNetV2-M 54M params, 24G flops, 85.1 top-1 on ImageNet  vs GC ViT-L  201M params, 32.6 G FLOPs, 84.6 top-1 ImageNet). So it's important to add also EfficinetNet-v2 in the comparaison.\n\nGiven these elements it is probably best to down tone the state-of-the art claims in the paper. Because it is not clear which state of the art it is.\n\n[1] Tan et al., EfficientNetV2: Smaller Models and Faster Training\n\nb) The efficient CNN-like token generator: Many papers have proposed convolution blocks for token generation like LeViT[1] or XCiT[2].\nThere is no discussion or comparison of performance with the previous approach. It is therefore impossible to evaluate the contribution of the paper on this point.\n\n[1] Graham et al., LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference\n[2] El-Nouby et al., XCiT: Cross-Covariance Image Transformers\n\n- Overfitting evaluation: For the ImageNet dataset there is no separate test and validation set so it's important to evaluate the level of overfitting by doing evaluation of the models on ImageNet-v2[1].\n\n[1] Recht et al., are we done with ImageNet?\n\n- Missing metrics: There is only params and FLOPs in the different table. It's important to measure other trade-off like Latency and memory consumption. Indeed, the proposed architecture use MBconv which are known to be good on trade-off FLOPs accuracy and parameter accuracy.\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and easy to follow. Unfortunately, the claims did not seem to be supported and the novelty seems very marginal. The paper seems to be reproducible", "summary_of_the_review": "The idea of the paper is interesting but unfortunately the contribution does not seem to be justified experimentally. The paper seems to be incremental. There is a lack of discussion and comparison with very similar approaches in the literature.\n\n\n====== Post rebuttal ====\n\nThe rebuttal raised a very important concern: \n\nThe authors report maximum accuracy during training by evaluating at each epoch in order to maximise performance on the ImageNet-1k validation set (there is no separate test set for ImageNet). Although the authors argue that it is a common practice, there is no evidence in the mentioned papers that this is the approach used. Furthermore I consider this approach to be detrimental to the community as it puts scientific understanding behind the optimal performance.\n\nOn the 3 logs provided the impact of this practice is important on the log with the bigger model +0.2% and impact the conclusion. The authors no longer have the other logs of their experiments, especially those of the ablation, so it is impossible to know the impact of the method. There is no multi-seed evaluation that shows that 0.1% on ImageNet-v2 is significant although it is claimed in the rebuttal.\n\nIn the current state the experiments are mainly designed to optimise the performance of the GC ViT architecture at the expense of an understanding of the impact of each component of the proposed method.\n\nSo I will lower my score.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject"}, "tcdate": 1666848948699}, {"id": "QixU8ZHmRkj", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1237/Reviewer_Fuf9"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduced GC ViT that can efficiently capture global context by utilizing global query tokens and interact with local regions. \nThis paper demonstrate through experiments that this simple modification on traditional ViTs could get SOTA results for image classification on ImageNet-1K dataset and downstream tasks.", "review_text": "Overall this paper proposed a simple and yet effective module that could add reasonable improvement to current ViT. This work could be useful to the community once the code is released for verification/comparison.  ", "strengths": "Strength:\n1. the idea is simple and straightforward, from the experiment it seems the improvement is reasonable.\n\nWeakness:\n1. The comparison of GC-VIT to other models clearly shows the quantitative improvements in Tab.1. In addition, the ablation study shows the proposed module could be quite plug-able into existing models. It will be more appealing to see how GC module improves existing models by simply plug them into them. This will inevitably add #param  and FLOPS, but it gives more intuitive evidence on the effectiveness. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper introduced GC ViT that can efficiently capture global context by utilizing global query tokens and interact with local regions. \nThis paper demonstrate through experiments that this simple modification on traditional ViTs could get SOTA results for image classification on ImageNet-1K dataset and downstream tasks.", "strength_and_weaknesses": "Strength:\n1. the idea is simple and straightforward, from the experiment it seems the improvement is reasonable.\n\nWeakness:\n1. The comparison of GC-VIT to other models clearly shows the quantitative improvements in Tab.1. In addition, the ablation study shows the proposed module could be quite plug-able into existing models. It will be more appealing to see how GC module improves existing models by simply plug them into them. This will inevitably add #param  and FLOPS, but it gives more intuitive evidence on the effectiveness. ", "clarity,_quality,_novelty_and_reproducibility": "There are no problems with Clarify, Quality and Novelty. \nHowever, since the paper is based on plain empirical/experimental findings. The released code will be critical in verifying the reproducibility of the paper. ", "summary_of_the_review": "Overall this paper proposed a simple and yet effective module that could add reasonable improvement to current ViT. This work could be useful to the community once the code is released for verification/comparison.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666727221804}, {"id": "fRHrsYK3D_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1237/Reviewer_V9pP"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a novel global context vision transformer built with a simple combination of global context self-attention modules and local self-attention modules to model\nboth long-range and short-range interactions. Besides, the authors propose to model the inter-channel dependencies with an inverted residual block.\nThe proposed GC ViT shows encouraging results across image classification, object detection, instance segmentation, and semantic segmentation tasks.\n", "review_text": "Designing new versatile architecture is always very important for the whole computer vision community. The overall quality is OK but the results are relatively WEAK to support the claim that the proposed method really outperforms the previous SoTA methods such as Swin-Transformer.\nPlease carefully address the above-listed weaknesses. I will increase the ratings if the authors could well address these concerns.", "strengths": "> Strengths\n\n✅ The presented idea is very simple and easy to follow.\n\n✅ The overall writing is of high quality, such as Figure 2, Figure 3, Figure 4, Figure 5, and Figure 6 is all of GOOD quality and help the readers to understand the key\nidea easily.\n\n✅ The authors provide rich details and conduct rich experiments to verify the effectiveness of the proposed approach.\n\n> Weaknesses\n\n❎ The experimental results are relatively WEAK and the authors are encouraged to verify the advantages over Swin-T based on the very recent object detection and segmentation frameworks\nsuch as DINO-DETR[1], H-DETR[2], and Mask2Former[3]. For example, Mask2Former+Swin-B (pre-trained on ImageNet1K) achieves PQ=55.1, AP=46.7, and mIoU=52.4 on COCO panoptic segmentation, COCO instance\nsegmentation, and ADE20K semantic segmentation tasks respectively. I would like to expect the authors to present stronger results and release the related code upon acceptance.\n\n[1] https://github.com/IDEA-Research/DINO\n\n[2] https://github.com/HDETR/H-Deformable-DETR \n\n[3] https://github.com/facebookresearch/Mask2Former \n\n❎ According to Table 1 & Table 2, the proposed GC Vit backbones require more parameters and GFLOPs than the previous approaches such as CSWin and ConvNeXt. For example, GC ViT-B requires 1018 GFLOPs vs. ConvNeXt-B requires 964 GFLOPs and GC ViT-B only gains 0.2 measured by AP^box.\n\n❎ The authors are encouraged to verify the advantages of the proposed backbone under larger model scales. Specifically, the authors should construct a GC ViT-L and GC ViT-H instead of only comparing the models under moderate scales as scaling up is one of the most important aspects of an important backbone design.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes a novel global context vision transformer built with a simple combination of global context self-attention modules and local self-attention modules to model\nboth long-range and short-range interactions. Besides, the authors propose to model the inter-channel dependencies with an inverted residual block.\nThe proposed GC ViT shows encouraging results across image classification, object detection, instance segmentation, and semantic segmentation tasks.\n", "strength_and_weaknesses": "> Strengths\n\n✅ The presented idea is very simple and easy to follow.\n\n✅ The overall writing is of high quality, such as Figure 2, Figure 3, Figure 4, Figure 5, and Figure 6 is all of GOOD quality and help the readers to understand the key\nidea easily.\n\n✅ The authors provide rich details and conduct rich experiments to verify the effectiveness of the proposed approach.\n\n> Weaknesses\n\n❎ The experimental results are relatively WEAK and the authors are encouraged to verify the advantages over Swin-T based on the very recent object detection and segmentation frameworks\nsuch as DINO-DETR[1], H-DETR[2], and Mask2Former[3]. For example, Mask2Former+Swin-B (pre-trained on ImageNet1K) achieves PQ=55.1, AP=46.7, and mIoU=52.4 on COCO panoptic segmentation, COCO instance\nsegmentation, and ADE20K semantic segmentation tasks respectively. I would like to expect the authors to present stronger results and release the related code upon acceptance.\n\n[1] https://github.com/IDEA-Research/DINO\n\n[2] https://github.com/HDETR/H-Deformable-DETR \n\n[3] https://github.com/facebookresearch/Mask2Former \n\n❎ According to Table 1 & Table 2, the proposed GC Vit backbones require more parameters and GFLOPs than the previous approaches such as CSWin and ConvNeXt. For example, GC ViT-B requires 1018 GFLOPs vs. ConvNeXt-B requires 964 GFLOPs and GC ViT-B only gains 0.2 measured by AP^box.\n\n❎ The authors are encouraged to verify the advantages of the proposed backbone under larger model scales. Specifically, the authors should construct a GC ViT-L and GC ViT-H instead of only comparing the models under moderate scales as scaling up is one of the most important aspects of an important backbone design.", "clarity,_quality,_novelty_and_reproducibility": "> Clarity\n\nGood.\n\n> Quality\n\nGood.\n\n> Novelty\n\nGood.\n\n> Reproducibility\n\nNo code is available.", "summary_of_the_review": "Designing new versatile architecture is always very important for the whole computer vision community. The overall quality is OK but the results are relatively WEAK to support the claim that the proposed method really outperforms the previous SoTA methods such as Swin-Transformer.\nPlease carefully address the above-listed weaknesses. I will increase the ratings if the authors could well address these concerns.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No further concerns.", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666243101206}], "openreview_url": "https://openreview.net/forum?id=HZJje06x6IO", "arxiv_id": "2206.09959", "paper_pdf": "papers/HZJje06x6IO.pdf", "paper_pdf_sha256": "0b7b5944d09a2db51ae20680ee0b5b3206a78a56f7241bf25ca1ac7288bd3d03", "paper_pdf_bytes": 892774, "paper_pdf_source": "openreview", "code_url": "https://github.com/NVlabs/GCVit", "code_repository": "NVlabs/GCVit", "code_commit": "8e1941f2099eeb99ee9e8472a5ddbd6b94a8958f", "code_archive": "repos/HZJje06x6IO.zip", "code_archive_sha256": "82adf96e1c063295ebc79d12403e52ac0d85f027928269464b3682a3852bef2d", "code_archive_bytes": 636848, "code_file_count": 29, "code_extensions": {".py": 27, ".sh": 2}, "github_disk_usage_kb": 879, "github_languages": {"Python": 198144, "Shell": 541}, "github_archived": false, "github_pushed_at": "2023-12-22T13:04:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/global-context-vision-transformers"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "in1ynkrXyMH", "year": 2022, "status": "rejected", "title": "Introspective Learning : A Two-Stage approach for Inference in Neural Networks", "authors": ["Mohit Prabhushankar", "Ghassan AlRegib"], "authorids": ["~Mohit_Prabhushankar1", "~Ghassan_AlRegib1"], "authors_source": "OpenReview API", "abstract": "In this paper, we advocate for two stages in a neural network's decision making process. The first is the existing feed-forward inference framework where patterns in given data are sensed and associated with previously learned patterns. The second stage is a slower reflection stage where we ask the network to reflect on its feed-forward decision by considering and evaluating all available choices. Together, we term the two stages as introspective learning. We use gradients of trained neural networks as a measurement of this reflection. We perceptually visualize the explanations from both stages to provide a visual grounding to introspection. For the application of recognition, we show that an introspective network is $4\\%$ more robust and $42\\%$ less prone to calibration errors when generalizing to noisy data. We also illustrate the value of introspective networks in downstream tasks that require generalizability and calibration including active learning and out-of-distribution detection. Finally, we ground the proposed machine introspection to human introspection in the application of image quality assessment.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "ZJ08YDHdcht", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3981/Reviewer_yYCU"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present a novel 2-stage technique to improve the classification accuracy of feedforward ANNs.  In particular, after the feedforward pass, a so-called introspective stage occurs, the goal of which is to ascertain why the particular class label was provided rather than a different label.  This stage, which tends to improve the accuracy of the class predictions, is modular and can be added onto networks under varying task conditions.", "review_text": "Vanilla feedforward networks are single-stage, filtering the inputs as they pass forward through the network and predicting the class of the input in the final layer. Motivated by the human tendency to not only gather sensory data, but to reflect on those data in the face of uncertainty to arrive at a classification, the authors introduce a second, introspective or reflective, stage. \n\nThis stage is defined as the change in the parameters in the network if a different label were assigned to the network's input, and can be measured as a gradient with respect to the network weights of a loss function.  After showing how the calculation can be simplified in space and time, they then demonstrate how introspection can improve the generalizability of classifications to distributions that are shifted from the original training sets and can improve the calibration of the network, the confidence of the classification minus the accuracy.  They finally demonstrate how the approach can be used in a number of different applications.\n\nOverall, the introspective approach is, to the reviewer's knowledge, novel and potentially of interest as a general method for improving the performance of feed-forward networks. The authors provide the reader with intuition (e.g., Fig 2), rigor, and empirical data (Fig 3/4).  One potential issue that arises in this empirical data is that introspection does not improve the accuracy of ResNet-18 on CIFAR-10C, as they state, although the ECE decreases systematically.  \n\nUnfortunately, it is difficult to judge by eye whether the differences in Fig 4 are statistically significant.  The authors should include error bars or some other visual cue that indicate the uncertainty of the plotted values.  Additionally, without a statistically significant result, it is difficult to convince the reader that the authors' approach will work in a meaningful manner empirically in cases of interest.  \n\nIf possible, I would suggest the authors find such an example, or, if they cannot, to perhaps indicate future work that might improve the technique in such a way that it could in fact do so. Beyond this, the paper is written and organized well and clearly.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors present a novel 2-stage technique to improve the classification accuracy of feedforward ANNs.  In particular, after the feedforward pass, a so-called introspective stage occurs, the goal of which is to ascertain why the particular class label was provided rather than a different label.  This stage, which tends to improve the accuracy of the class predictions, is modular and can be added onto networks under varying task conditions.", "main_review": "Vanilla feedforward networks are single-stage, filtering the inputs as they pass forward through the network and predicting the class of the input in the final layer. Motivated by the human tendency to not only gather sensory data, but to reflect on those data in the face of uncertainty to arrive at a classification, the authors introduce a second, introspective or reflective, stage. \n\nThis stage is defined as the change in the parameters in the network if a different label were assigned to the network's input, and can be measured as a gradient with respect to the network weights of a loss function.  After showing how the calculation can be simplified in space and time, they then demonstrate how introspection can improve the generalizability of classifications to distributions that are shifted from the original training sets and can improve the calibration of the network, the confidence of the classification minus the accuracy.  They finally demonstrate how the approach can be used in a number of different applications.\n\nOverall, the introspective approach is, to the reviewer's knowledge, novel and potentially of interest as a general method for improving the performance of feed-forward networks. The authors provide the reader with intuition (e.g., Fig 2), rigor, and empirical data (Fig 3/4).  One potential issue that arises in this empirical data is that introspection does not improve the accuracy of ResNet-18 on CIFAR-10C, as they state, although the ECE decreases systematically.  \n\nUnfortunately, it is difficult to judge by eye whether the differences in Fig 4 are statistically significant.  The authors should include error bars or some other visual cue that indicate the uncertainty of the plotted values.  Additionally, without a statistically significant result, it is difficult to convince the reader that the authors' approach will work in a meaningful manner empirically in cases of interest.  \n\nIf possible, I would suggest the authors find such an example, or, if they cannot, to perhaps indicate future work that might improve the technique in such a way that it could in fact do so. Beyond this, the paper is written and organized well and clearly.", "summary_of_the_review": "The authors present a new method, introspection, that can be added on top of feedforward networks to improve accuracy and generalization of the networks.  While they have written a well-structured paper that provides both intuition and rigor, the empirical results are somewhat lacking in that there are no statistically significant results that support the claims of accuracy improvements that introspection can obtain.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636368584123}, {"id": "KY2lzbPs8dA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3981/Reviewer_DxqE"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose a modification to standard neural networks used for object classification tasks to incorporate what they call “introspective learning”. This consists on training a multi-layer perceptron (MLP) on the neural network introspective features. These are obtained by calculating the gradients over the last layer weights of the model on a loss function corresponding to posing an introspective question: why is the correct label A instead of B? The authors apply this procedure to several neural networks of the ResNet model family and show that the introspective-networks have better generalization to distributional shifts and smaller calibration errors on datasets with the same classes and image sizes as CIFAR-10. Finally, they show further improvements in multiple applications.", "review_text": "# Strengths\n* An interesting modification to standard neural networks for object recognition with improvements in generalization and  calibration\n* Extensive experiments showing the usefulness of the method in multiple applications\n\n# Weaknesses\n* Confusing explanation of the proposed method lacking a visual representation of the full model architecture, its components, and what the introspective features represent\n* Model is evaluated only on CIFAR-10 related datasets which are very low resolution images and few possible classes\n* According to the authors, the proposed method does not scale well for larger datasets. This makes its applicability very limited as CIFAR-10 is only a toy dataset\n* Some controls seem rather arbitrary and a better choice of parameters for the alternative baselines could have been done\n\n# Detailed review\n## Role of introspection for object recognition\nObject recognition in humans is an extremely fast visual computation, taking place within ~200ms of stimulus presentation, and it is unlikely to depend on introspection. It is not clear the motivation of why including introspective learning in artificial neural networks would improve their performance. Furthermore, standard neural networks are typically optimized to classify an object as belonging to one class and not belonging to the others (all classes are used in the cost function).  In that sense, it could be argued that they also are trained to answer the questions illustrated in Figure 1.\n\n## Confusing explanation of the proposed methodology\nUnfortunately, the authors could have done a better job at explaining their method. In my opinion, some of the mathematical details could be moved to the appendix, and in their place, the authors should focus on clarifying the training and inference process of the introspective networks. A visual representation of the full model architecture, its components, what the introspective features represent, and how the model is trained would be extremely helpful to the reader. Because of this it was not clear where the computational complexity of the introspective networks come from. The MLP used is very standard and the introspective features are obtained from the model’s last layer gradients.\n\n## Poor choice of controls\nSince adding the 3 layers to the standard model made it considerably worse, the choice of using 3 additional layers on the control for the same number of parameters was a poor choice by the authors. The MLP in the introspective network has a higher number of features but only a single layer. A better comparison would be to make the last layer of the standard model with more features, keeping only 1 layer, or keeping the same penultimate layer and adding a new much wider layer to compensate the smaller number of parameters. \n\n## Minor comments\n* Fonts in several figures are very small. Please improve their readability\n* In Figure 4b and Figure 6, the x-axis should be corruption severity and go from 1 to 5\n* Figure 2 does not add much and could go to supplementary\n* Figure 1 is very high-level and in its place, the authors should provide a more concrete visualization of their method\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose a modification to standard neural networks used for object classification tasks to incorporate what they call “introspective learning”. This consists on training a multi-layer perceptron (MLP) on the neural network introspective features. These are obtained by calculating the gradients over the last layer weights of the model on a loss function corresponding to posing an introspective question: why is the correct label A instead of B? The authors apply this procedure to several neural networks of the ResNet model family and show that the introspective-networks have better generalization to distributional shifts and smaller calibration errors on datasets with the same classes and image sizes as CIFAR-10. Finally, they show further improvements in multiple applications.", "main_review": "# Strengths\n* An interesting modification to standard neural networks for object recognition with improvements in generalization and  calibration\n* Extensive experiments showing the usefulness of the method in multiple applications\n\n# Weaknesses\n* Confusing explanation of the proposed method lacking a visual representation of the full model architecture, its components, and what the introspective features represent\n* Model is evaluated only on CIFAR-10 related datasets which are very low resolution images and few possible classes\n* According to the authors, the proposed method does not scale well for larger datasets. This makes its applicability very limited as CIFAR-10 is only a toy dataset\n* Some controls seem rather arbitrary and a better choice of parameters for the alternative baselines could have been done\n\n# Detailed review\n## Role of introspection for object recognition\nObject recognition in humans is an extremely fast visual computation, taking place within ~200ms of stimulus presentation, and it is unlikely to depend on introspection. It is not clear the motivation of why including introspective learning in artificial neural networks would improve their performance. Furthermore, standard neural networks are typically optimized to classify an object as belonging to one class and not belonging to the others (all classes are used in the cost function).  In that sense, it could be argued that they also are trained to answer the questions illustrated in Figure 1.\n\n## Confusing explanation of the proposed methodology\nUnfortunately, the authors could have done a better job at explaining their method. In my opinion, some of the mathematical details could be moved to the appendix, and in their place, the authors should focus on clarifying the training and inference process of the introspective networks. A visual representation of the full model architecture, its components, what the introspective features represent, and how the model is trained would be extremely helpful to the reader. Because of this it was not clear where the computational complexity of the introspective networks come from. The MLP used is very standard and the introspective features are obtained from the model’s last layer gradients.\n\n## Poor choice of controls\nSince adding the 3 layers to the standard model made it considerably worse, the choice of using 3 additional layers on the control for the same number of parameters was a poor choice by the authors. The MLP in the introspective network has a higher number of features but only a single layer. A better comparison would be to make the last layer of the standard model with more features, keeping only 1 layer, or keeping the same penultimate layer and adding a new much wider layer to compensate the smaller number of parameters. \n\n## Minor comments\n* Fonts in several figures are very small. Please improve their readability\n* In Figure 4b and Figure 6, the x-axis should be corruption severity and go from 1 to 5\n* Figure 2 does not add much and could go to supplementary\n* Figure 1 is very high-level and in its place, the authors should provide a more concrete visualization of their method\n", "summary_of_the_review": "The authors propose a very interesting modification to neural networks for object recognition and do a convincing job in showing that it improves generalization and calibration. However, the proposed method is only tested in CIFAR-10 and its use in larger networks and datasets is questionable. Furthermore, the explanation of the method is unnecessarily confusing and better controls could have been chosen to make a more convincing case.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635971209264}, {"id": "lYTaN7SHpqk", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3981/Reviewer_Q4tj"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes introspective networks, a technique utilizing the gradients of a trained base model to make more robust predictions when faced with distribution shift.", "review_text": "The paper proposes an interesting and novel approach for improving robustness and calibration under distribution shift.  Results on CIFAR10-Corrupted are very promising, and the authors further show the method is able to improve in active learning and OOD detection experiments when inputs are corrupted with noise.\n\nLemma 1: assumptions should be stated precisely in the lemma statement itself. As it stands, the lemma statement makes it seem like the statement holds exactly, which seems very unreasonable at first glance.\nRegarding the reasonableness of the assumptions, a  \"well trained\" network might have high confidence predictions on input data it was trained on, but not necessarily on test data. Empirically, it has been observed that test data (particularly OOD test data) tends to have much higher entropy in predictions. Regarding theorem 1, can you test how well this approximation holds? In CIFAR10, there are only 10 classes, so it's reasonable to actually compute and store the full gradient features (especially since only the last layer is used). It would be good to see how a classifier performs using the full set of introspective features instead of the gradient wrt to the all ones vector. I would also highly recommend experiments to see if introspection (and this approximation in particular) scale well to more complex tasks with more classes like CIFAR100 or Imagenet, as I believe the current set of experiments to be quite limited.\n\n**Questions and Suggested Ablations:**\n\nCan you elaborate on why we should expect the introspective features to provide benefits when the new classifier is trained on the exact same training data as the base network? It seems very plausible to me that the gradients on training points the base\nnetwork has already overfitted heavily are a very different distribution than the gradients on previously unseen points, so I wouldn't necessarily have expected the introspective classifier to generalize well. I'm also curious what would happen if the introspection network is trained on a separate validation set.\n\nHow does this compare to directly taking the activations of the network instead of gradients? The final activations should be very closely related to the gradients of the last layer, differing by multiplicative terms corresponding to derivatives of the loss wrt to the output logits. This experiment would help to examine whether the structure of the gradient information from each class over possibly some benefits of just retraining model components on top of a pretrained model features.\n\nDid you examine introspective features from other layers?\n\nIn the OOD detection, are the same adversarial images used for the feedforward and introspective network, or is the adversarial image for the introspective method specifically targeted at the extra network using introspective features?\n\nSome missing related work:\n\nGradients as features: [2] is a crucial piece of missing related work, which also considers using gradients of a network as features, and for example show some similar results in recovering the original classifier performance on the trianing set. It would also be good to discuss and reference [1], which similarly uses gradients of base models as features, but applies it towards finetuning for transfer learning.\n \nConditional/predictive normalized maximum likelihood: Conditional/predictive normalized maxmimum likelihood [3, 4] considers retraining models for each label of the test input and combining these models' predictions to output a final prediction, essentially introspecting for each possible label to get predictions. While this work differs by using introspection to provide features for a separate classifier, it would be good to mention these other works as applying similar ideas of utilizing information on how model's change for specific labels of the test input in order improve predictions in the face of distribution shift.\n\n**Misc comments:**\n\nThere's a bit of a weird jump in the assumptions in the paper: lemma 1 assumes cross-entropy loss, but in sec 3, we're assuming network is trained with MSE for the derivation. \n\nIn addition to showing ECE for CIFAR10-C, I would also recommend including uncertainty-aware metrics like NLL and Brier score in the appendix, as these are proper scoring rules.\n\n\n**Citations:**\n\n[1] Zinkevich, Martin A. et al. “Holographic Feature Representations of Deep Networks.” UAI (2017).\n\n[2] Mu, Fangzhou, Yingyu Liang, and Yin Li. \"Gradients as features for deep representation learning.\" ICLR (2020)\n\n[3] Zhou, Aurick, and Sergey Levine. \"Amortized Conditional Normalized Maximum Likelihood: Reliable Out of Distribution Uncertainty Estimation.\" International Conference on Machine Learning (2021)\n\n[4] Bibas, Koby, Yaniv Fogel, and Meir Feder. \"Deep pnml: Predictive normalized maximum likelihood for deep neural networks.\" arXiv preprint arXiv:1904.12286 (2019).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes introspective networks, a technique utilizing the gradients of a trained base model to make more robust predictions when faced with distribution shift.", "main_review": "The paper proposes an interesting and novel approach for improving robustness and calibration under distribution shift.  Results on CIFAR10-Corrupted are very promising, and the authors further show the method is able to improve in active learning and OOD detection experiments when inputs are corrupted with noise.\n\nLemma 1: assumptions should be stated precisely in the lemma statement itself. As it stands, the lemma statement makes it seem like the statement holds exactly, which seems very unreasonable at first glance.\nRegarding the reasonableness of the assumptions, a  \"well trained\" network might have high confidence predictions on input data it was trained on, but not necessarily on test data. Empirically, it has been observed that test data (particularly OOD test data) tends to have much higher entropy in predictions. Regarding theorem 1, can you test how well this approximation holds? In CIFAR10, there are only 10 classes, so it's reasonable to actually compute and store the full gradient features (especially since only the last layer is used). It would be good to see how a classifier performs using the full set of introspective features instead of the gradient wrt to the all ones vector. I would also highly recommend experiments to see if introspection (and this approximation in particular) scale well to more complex tasks with more classes like CIFAR100 or Imagenet, as I believe the current set of experiments to be quite limited.\n\n**Questions and Suggested Ablations:**\n\nCan you elaborate on why we should expect the introspective features to provide benefits when the new classifier is trained on the exact same training data as the base network? It seems very plausible to me that the gradients on training points the base\nnetwork has already overfitted heavily are a very different distribution than the gradients on previously unseen points, so I wouldn't necessarily have expected the introspective classifier to generalize well. I'm also curious what would happen if the introspection network is trained on a separate validation set.\n\nHow does this compare to directly taking the activations of the network instead of gradients? The final activations should be very closely related to the gradients of the last layer, differing by multiplicative terms corresponding to derivatives of the loss wrt to the output logits. This experiment would help to examine whether the structure of the gradient information from each class over possibly some benefits of just retraining model components on top of a pretrained model features.\n\nDid you examine introspective features from other layers?\n\nIn the OOD detection, are the same adversarial images used for the feedforward and introspective network, or is the adversarial image for the introspective method specifically targeted at the extra network using introspective features?\n\nSome missing related work:\n\nGradients as features: [2] is a crucial piece of missing related work, which also considers using gradients of a network as features, and for example show some similar results in recovering the original classifier performance on the trianing set. It would also be good to discuss and reference [1], which similarly uses gradients of base models as features, but applies it towards finetuning for transfer learning.\n \nConditional/predictive normalized maximum likelihood: Conditional/predictive normalized maxmimum likelihood [3, 4] considers retraining models for each label of the test input and combining these models' predictions to output a final prediction, essentially introspecting for each possible label to get predictions. While this work differs by using introspection to provide features for a separate classifier, it would be good to mention these other works as applying similar ideas of utilizing information on how model's change for specific labels of the test input in order improve predictions in the face of distribution shift.\n\n**Misc comments:**\n\nThere's a bit of a weird jump in the assumptions in the paper: lemma 1 assumes cross-entropy loss, but in sec 3, we're assuming network is trained with MSE for the derivation. \n\nIn addition to showing ECE for CIFAR10-C, I would also recommend including uncertainty-aware metrics like NLL and Brier score in the appendix, as these are proper scoring rules.\n\n\n**Citations:**\n\n[1] Zinkevich, Martin A. et al. “Holographic Feature Representations of Deep Networks.” UAI (2017).\n\n[2] Mu, Fangzhou, Yingyu Liang, and Yin Li. \"Gradients as features for deep representation learning.\" ICLR (2020)\n\n[3] Zhou, Aurick, and Sergey Levine. \"Amortized Conditional Normalized Maximum Likelihood: Reliable Out of Distribution Uncertainty Estimation.\" International Conference on Machine Learning (2021)\n\n[4] Bibas, Koby, Yaniv Fogel, and Meir Feder. \"Deep pnml: Predictive normalized maximum likelihood for deep neural networks.\" arXiv preprint arXiv:1904.12286 (2019).\n", "summary_of_the_review": "Overall, the paper presents an interesting approach with promising results for improving robustness, both in terms of accuracy and uncertainty estimation. I believe with a more clear presentation, more thorough discussion of related work, more extensive experimental evaluations, and ablations/evaluations of specific assumptions and design choices, this will be a valuable contribution.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635832394468}, {"id": "pvT6MT4OOya", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3981/Reviewer_8FMs"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors present a new inference pipeline for neural networks $f$ trained to solve classification problems. They augment the features processed by $f$ with the loss gradients with respect to the weights of the last layer. Their pipeline requires two steps: the first one coincides with a standard evaluation of $f(x)$, while gradients are evaluated and processed in the second step to derive the final prediction. Experiments show slight improvements in classification accuracy and cuts of calibration errors. ", "review_text": "The reading of the paper does not flow properly. The methods are hard to understand and are not technically sound. Many core elements of the math discourse are introduced abruptly, such as $L$ and $\\mathcal{H}$. The definition of introspection is vague and not formal. The gap between the intuition about introspection presented in the introduction and the practical realization described in the methods sections remains wide and leaves the readers without a proper foundation to build their understanding. For instance, the intuition of introspection the authors give in the introduction and in Figure 1 calls for a reflection stage applied only to some promising alternative classes, while the proposed method applies the reflection stage to every single class.\nMoreover, I think that some less relevant sections should be shortened and others more relevant should be deepened. For instance, given the nature of the scientific contribution, this paper does not require in my opinion such an extensive Application section, while spending more words on making clear the methods in the middle sections and the general idea in the Introduction could provide more value to the reader.\n\nYet, to me, the idea of enriching neural networks features with gradient information through a two-stage inference process seems novel and sounds interesting. Results suggest that this research direction can be relevant, so I encourage the authors to pursue this work and focus on communicating it properly.\n\nIn the following, I list some comments about specific parts of the paper:\n\nINTRODUCTION\n- I believe that philosophical references can be of great value to the ML and AI community. In this sense I appreciate the reference to Locke, I wish the authors have expanded more this point.\n- \"For the application of recognition... ...EQUATION (1)\" This part is not clear. I do not understand why the authors use the projection concept in this context. $L$ seems a tag to indicate the final layer but then becomes a natural number ($L-1$). The first equation about $y_{feat}$ has a reference to $y$ that does not appear in the equation.\n\nINTROSPECTIVE FEATURES\n- The authors use the term explanations without defining what is an explanation for them and justify that definition. It is a textual string? A heat map on an image? Why we can consider heat maps as *explanations*? \n- The definition of introspection is very vague. \"...is the measurement of change induced in the network parameters...\", explicit this measurement, state the quantities. The reader here is expecting a formal definition but this definition is not formal.\n\nINTROSPECTIVE FEATURE VISUALIZATION\n- \"...Grad-CAM indicates that the pink and round body, and straight beak are the reasons for the decision. [...] the network highlights the neck of the spoonbill to indicate that since S-shaped neck is not observed, x cannot be a flamingo. ...\" I believe that here authors are overinterpreting the heat-maps. How can we say that the heat-map is caused by shapes rather than patterns? The clean high-level concepts of neck, beak, straight, S-shaped are what the network sees or what the authors see?\n\nINTROSPECTIVE FEATURE EXTRACTION\n- The assumption that a forward and backward pass through f is O(1) does not sound familiar to me. There is a dependence with respect to the size of the input. \n- Probably this subsection could be shorter, and at least part of the computational cost consideration could go in the appendix.\n\nINTROSPECTIVE NETWORK\n- H appears suddenly and everything is built on it. This is one of the more obscure parts of the paper.\n- The assumption of a mean squared error loss is not trivial, we usually do not train classifiers with the mean squared error loss but with cross-entropy. \n- \"having networks that introspect consecutively, by creating a trade-off\" is not clear.\n-  let the reader grasp how the reflection stage can fix the first inference of the sensing stage, spend some words on this case or maybe provide an example.\n\nEXPERIMENTS\n- \"Given a data distribution $x \\in X$\" why distribution?\n- \"10 separate bins based on their prediction confidence\". Considering the outputs of a neural network as confidences is problematic, this point should be expanded. DNNs are known to have unreliable uncertainty estimates (MacKay, 1995, Szegedy et al., 2014, Nguyen et al., 2015)", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors present a new inference pipeline for neural networks $f$ trained to solve classification problems. They augment the features processed by $f$ with the loss gradients with respect to the weights of the last layer. Their pipeline requires two steps: the first one coincides with a standard evaluation of $f(x)$, while gradients are evaluated and processed in the second step to derive the final prediction. Experiments show slight improvements in classification accuracy and cuts of calibration errors. ", "main_review": "The reading of the paper does not flow properly. The methods are hard to understand and are not technically sound. Many core elements of the math discourse are introduced abruptly, such as $L$ and $\\mathcal{H}$. The definition of introspection is vague and not formal. The gap between the intuition about introspection presented in the introduction and the practical realization described in the methods sections remains wide and leaves the readers without a proper foundation to build their understanding. For instance, the intuition of introspection the authors give in the introduction and in Figure 1 calls for a reflection stage applied only to some promising alternative classes, while the proposed method applies the reflection stage to every single class.\nMoreover, I think that some less relevant sections should be shortened and others more relevant should be deepened. For instance, given the nature of the scientific contribution, this paper does not require in my opinion such an extensive Application section, while spending more words on making clear the methods in the middle sections and the general idea in the Introduction could provide more value to the reader.\n\nYet, to me, the idea of enriching neural networks features with gradient information through a two-stage inference process seems novel and sounds interesting. Results suggest that this research direction can be relevant, so I encourage the authors to pursue this work and focus on communicating it properly.\n\nIn the following, I list some comments about specific parts of the paper:\n\nINTRODUCTION\n- I believe that philosophical references can be of great value to the ML and AI community. In this sense I appreciate the reference to Locke, I wish the authors have expanded more this point.\n- \"For the application of recognition... ...EQUATION (1)\" This part is not clear. I do not understand why the authors use the projection concept in this context. $L$ seems a tag to indicate the final layer but then becomes a natural number ($L-1$). The first equation about $y_{feat}$ has a reference to $y$ that does not appear in the equation.\n\nINTROSPECTIVE FEATURES\n- The authors use the term explanations without defining what is an explanation for them and justify that definition. It is a textual string? A heat map on an image? Why we can consider heat maps as *explanations*? \n- The definition of introspection is very vague. \"...is the measurement of change induced in the network parameters...\", explicit this measurement, state the quantities. The reader here is expecting a formal definition but this definition is not formal.\n\nINTROSPECTIVE FEATURE VISUALIZATION\n- \"...Grad-CAM indicates that the pink and round body, and straight beak are the reasons for the decision. [...] the network highlights the neck of the spoonbill to indicate that since S-shaped neck is not observed, x cannot be a flamingo. ...\" I believe that here authors are overinterpreting the heat-maps. How can we say that the heat-map is caused by shapes rather than patterns? The clean high-level concepts of neck, beak, straight, S-shaped are what the network sees or what the authors see?\n\nINTROSPECTIVE FEATURE EXTRACTION\n- The assumption that a forward and backward pass through f is O(1) does not sound familiar to me. There is a dependence with respect to the size of the input. \n- Probably this subsection could be shorter, and at least part of the computational cost consideration could go in the appendix.\n\nINTROSPECTIVE NETWORK\n- H appears suddenly and everything is built on it. This is one of the more obscure parts of the paper.\n- The assumption of a mean squared error loss is not trivial, we usually do not train classifiers with the mean squared error loss but with cross-entropy. \n- \"having networks that introspect consecutively, by creating a trade-off\" is not clear.\n-  let the reader grasp how the reflection stage can fix the first inference of the sensing stage, spend some words on this case or maybe provide an example.\n\nEXPERIMENTS\n- \"Given a data distribution $x \\in X$\" why distribution?\n- \"10 separate bins based on their prediction confidence\". Considering the outputs of a neural network as confidences is problematic, this point should be expanded. DNNs are known to have unreliable uncertainty estimates (MacKay, 1995, Szegedy et al., 2014, Nguyen et al., 2015)", "summary_of_the_review": "The core parts of the paper are not clear and do not allow a proper comprehension and evaluation of the scientific contribution of this work. \nThe overall idea seems valuable and experimental results appear promising. Yet, with the current manuscript, I can only judge the potential of this work rather than the work itself, so at this stage my recommendation is a 3: reject, not good enough.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635776559904}], "openreview_url": "https://openreview.net/forum?id=in1ynkrXyMH", "arxiv_id": "2209.08425", "paper_pdf": "papers/in1ynkrXyMH.pdf", "paper_pdf_sha256": "d8bbd42572fa5fa1d3142db44be8a2ca2e130ad0b512f3aba86bd218fe0d69d2", "paper_pdf_bytes": 6856615, "paper_pdf_source": "openreview", "code_url": "https://github.com/olivesgatech/Introspective-Learning", "code_repository": "olivesgatech/Introspective-Learning", "code_commit": "fd9a358f45d7b417d1687d3bcdcce18facba2b2a", "code_archive": "repos/in1ynkrXyMH.zip", "code_archive_sha256": "e2b0417600deb54e46f3d5fcb6b1f10cd27197ea91f6e2ec521ce7c869e9f238", "code_archive_bytes": 4259454, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 4165, "github_languages": {"Python": 24392}, "github_archived": false, "github_pushed_at": "2022-09-26T13:40:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/introspective-learning-a-two-stage-approach-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "GVNGAaY2Dr1", "year": 2021, "status": "rejected", "title": "Multi-Agent Collaboration via Reward Attribution Decomposition", "authors": ["Tianjun Zhang", "Huazhe Xu", "Xiaolong Wang", "Yi Wu", "Kurt Keutzer", "Joseph E. Gonzalez", "Yuandong Tian"], "authorids": ["~Tianjun_Zhang1", "~Huazhe_Xu1", "~Xiaolong_Wang3", "~Yi_Wu1", "~Kurt_Keutzer1", "~Joseph_E._Gonzalez1", "~Yuandong_Tian1"], "authors_source": "OpenReview API", "abstract": "Recent advances in multi-agent reinforcement learning (MARL) have achieved super-human performance in games like Quake 3 and Dota 2. Unfortunately, these techniques require orders-of-magnitude more training rounds than humans and don't generalize to new agent configurations even on the same game. In this work, we propose Collaborative Q-learning (CollaQ) that achieves state-of-the-art performance in the StarCraft multi-agent challenge and supports ad hoc team play. We first formulate multi-agent collaboration as a joint optimization on reward assignment and show that each agent has an approximately optimal policy that decomposes into two parts: one part that only relies on the agent's own state, and the other part that is related to states of nearby agents. Following this novel finding, CollaQ decomposes the Q-function of each agent into a self term and an interactive term, with a Multi-Agent Reward Attribution (MARA) loss that regularizes the training. CollaQ is evaluated on various StarCraft maps and shows that it outperforms existing state-of-the-art techniques  (i.e., QMIX, QTRAN, and VDN) by improving the win rate by 40% with the same number of samples. In the more challenging ad hoc team play setting (i.e., reweight/add/remove units without re-training or finetuning), CollaQ outperforms previous SoTA by over 30%. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rabHEzOzQQL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2134/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Good theoretical analysis and compliant experiment performance\n\n1. The Limitation of Theorem 1: the authors have said that ``the optimal gap of r_i heavily depends on the size of s_i^{local}. But in experiments (including experiments in appendix), the authors only discussed the claimed optimal setting (“using the observation o_i of agent i covers $s_i^{local}$”). More experiments on other size of s_i^{local} could be added to better prove the conclusion. (Whether the best choice can not or hard to be proven mathematically.)\n2. Some expression problems which may cause confusions: 1) Too many kinds of rewards including perceived reward, local reward, external reward and etc. The definitions of them are not very clear.  For example,  the perceived reward is really confused; 2) A brief algorithm flow chat and pseudo codes are needed for better understanding of how the algorithm works.\n3. As this work is eventually an MARL work in solving ad hoc team setting games by decomposing reward. Some explicit comparisons (May be in form or experiments or brief analysis) should be added with some MARL methods(SSD: Social Influnce as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning; PBRS: Reward shaping for knowledge-based multi-objective multi-agent reinforcement learning). Only the credit assignment problem in RL is discussed, the authors need to discuss more on some other related works like social dilemma in MARL or reward shaping?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good theoretical analysis and compliant experiment performance", "review": "Good theoretical analysis and compliant experiment performance\n\n1. The Limitation of Theorem 1: the authors have said that ``the optimal gap of r_i heavily depends on the size of s_i^{local}. But in experiments (including experiments in appendix), the authors only discussed the claimed optimal setting (“using the observation o_i of agent i covers $s_i^{local}$”). More experiments on other size of s_i^{local} could be added to better prove the conclusion. (Whether the best choice can not or hard to be proven mathematically.)\n2. Some expression problems which may cause confusions: 1) Too many kinds of rewards including perceived reward, local reward, external reward and etc. The definitions of them are not very clear.  For example,  the perceived reward is really confused; 2) A brief algorithm flow chat and pseudo codes are needed for better understanding of how the algorithm works.\n3. As this work is eventually an MARL work in solving ad hoc team setting games by decomposing reward. Some explicit comparisons (May be in form or experiments or brief analysis) should be added with some MARL methods(SSD: Social Influnce as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning; PBRS: Reward shaping for knowledge-based multi-objective multi-agent reinforcement learning). Only the credit assignment problem in RL is discussed, the authors need to discuss more on some other related works like social dilemma in MARL or reward shaping?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604041980692}, {"id": "UTzNZNeS31D", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2134/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper focuses on reward attribution in multiagent reinforcement learning, proposing a new algorithm, essentially by splitting value functions to what agents can achieve individually and learning separately how different individual rewards interact. This is an old idea, but the paper essentially applies it to the state-of-the-art deep learning machinery, producing impressive results on the hardest games state-of-the-art algorithms can manage. \n\nAs is the case with many similar areas, the work emphasises translation of an idea to the deelp learning setting, with the usual caveats, i.e. that there is not a major new methodological or conceptual insight, and the main advance is in terms of scaling up to real-world scenarios rather than having a better explicit algorithm to manage reward decomposition at runtime in an online learning setting. In other words, we learn more about how to solve challenging games rather than about the key underlying AI problem. \n\nThe paper does not offer a huge amount of novelty, but rather presents a solid deep RL engineering approach to solving the wider problem in a specific setting. Nonetheless, the technical material is well-developed, the presentation is overall of a high quality, and the experimental results extensive (and impressive). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel method for reward attribution achieves impressive results in state-of-the-art benchmark domains", "review": "The paper focuses on reward attribution in multiagent reinforcement learning, proposing a new algorithm, essentially by splitting value functions to what agents can achieve individually and learning separately how different individual rewards interact. This is an old idea, but the paper essentially applies it to the state-of-the-art deep learning machinery, producing impressive results on the hardest games state-of-the-art algorithms can manage. \n\nAs is the case with many similar areas, the work emphasises translation of an idea to the deelp learning setting, with the usual caveats, i.e. that there is not a major new methodological or conceptual insight, and the main advance is in terms of scaling up to real-world scenarios rather than having a better explicit algorithm to manage reward decomposition at runtime in an online learning setting. In other words, we learn more about how to solve challenging games rather than about the key underlying AI problem. \n\nThe paper does not offer a huge amount of novelty, but rather presents a solid deep RL engineering approach to solving the wider problem in a specific setting. Nonetheless, the technical material is well-developed, the presentation is overall of a high quality, and the experimental results extensive (and impressive). ", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603911216994}, {"id": "LC6KEYKa8An", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2134/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "To address the ad hoc team play, the authors propose a residual term of Q function, which additionally considers the states of nearby agents. A novel MARA loss is introduced to the residual term as a regularization to achieve the reward assignment implicitly. The proposed CollaQ could be easily built on QMIX and trained end-to-end. CollaQ outperforms other baselines on various tasks with the ad hoc team play setting. \n\nThe paper is very clear and well-structured. To the best of my knowledge, the MARA regularization is novel enough. The Ad-hoc MARL is an important problem in real-world applications but has not been fully studied. The interactive term with regularization is a practical and promising method to solve this problem and could be followed by other researchers.\n\nHowever, I still have some concerns:\n\nFirst, the theoretical analysis of reward assignment is not close to the implementation of CollaQ. There is no real assignment mechanism. A MARA loss is derived from the theoretical analysis to achieve the reward assignment implicitly, but the MARA loss could be more straightforwardly interpreted as that the Q value should be equal to the individual value when the agent cannot observe other agents. From this perspective, the complex reward assignment is not necessary. Moreover, CollaQ is built on QMIX. However, the individual value function in QMIX does not estimate a real expected return, and the value has no meaning. Is the theoretical analysis of reward assignment still valid in QMIX?\n\nI do not find any experiments to support the claim that \"agents using CollaQ would first learn to solve the problem pretending no other agents are around using Qalone then try to learn interaction with local agents through Qcollab.\" I think it is over-claimed and should be removed. Splitting the end-to-end learning process into two learning stages might harm the learning. \n\nThe visualizations in Fig 3 are helpful to understand how CollaQ works, but they are special cases. Statistical results are more convincing to verify how CollaQ influences the decision.\n\nAt the test time of StarCraft, are the IDs shuffled at each timestep or only at the first timestep of an episode?\n\n----------Update after author response----------\n\nI thank the authors for the detailed response. Most of my concerns have been addressed, and I decide to keep my score.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Reviews", "review": "To address the ad hoc team play, the authors propose a residual term of Q function, which additionally considers the states of nearby agents. A novel MARA loss is introduced to the residual term as a regularization to achieve the reward assignment implicitly. The proposed CollaQ could be easily built on QMIX and trained end-to-end. CollaQ outperforms other baselines on various tasks with the ad hoc team play setting. \n\nThe paper is very clear and well-structured. To the best of my knowledge, the MARA regularization is novel enough. The Ad-hoc MARL is an important problem in real-world applications but has not been fully studied. The interactive term with regularization is a practical and promising method to solve this problem and could be followed by other researchers.\n\nHowever, I still have some concerns:\n\nFirst, the theoretical analysis of reward assignment is not close to the implementation of CollaQ. There is no real assignment mechanism. A MARA loss is derived from the theoretical analysis to achieve the reward assignment implicitly, but the MARA loss could be more straightforwardly interpreted as that the Q value should be equal to the individual value when the agent cannot observe other agents. From this perspective, the complex reward assignment is not necessary. Moreover, CollaQ is built on QMIX. However, the individual value function in QMIX does not estimate a real expected return, and the value has no meaning. Is the theoretical analysis of reward assignment still valid in QMIX?\n\nI do not find any experiments to support the claim that \"agents using CollaQ would first learn to solve the problem pretending no other agents are around using Qalone then try to learn interaction with local agents through Qcollab.\" I think it is over-claimed and should be removed. Splitting the end-to-end learning process into two learning stages might harm the learning. \n\nThe visualizations in Fig 3 are helpful to understand how CollaQ works, but they are special cases. Statistical results are more convincing to verify how CollaQ influences the decision.\n\nAt the test time of StarCraft, are the IDs shuffled at each timestep or only at the first timestep of an episode?\n\n----------Update after author response----------\n\nI thank the authors for the detailed response. Most of my concerns have been addressed, and I decide to keep my score.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603261952433}, {"id": "xk6cb4IbgjT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2134/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies a team of agents that collaborate to maximize a global objective, where all agents receive the same reward value as a function of their state and actions. The approach suggested here is to assign each agent a virtual (\"perceived\") reward function such that all the virtual rewards sum up to the actual reward. Then the problem from the point of view of agent i can be solved with local value functions that depend on the local state of agent i and its perceived reward.  It is shown that such a decoupled policy exists that approximates the optimal policy if the state structure decomposes \"well enough\" into local states. Experiments demonstrate the advantages of this approach on a resource collection game on the StarCraft Multi-Agent Challange, with dramatic improvements over existing methods for the scenarios that were tested. \n\nThe paper is mostly easy to follow, and is well-written and well motivated. The idea of this paper is nice and intuitive. Being less familiar with the literature, It's hard for me to judge how novel this idea is. It seems like there must exist multi-agent works that exploit the local structure of the interaction to reduce the dimensionality. Hence, as a start, the contribution of this paper can be enhanced by extending the literature review to include works of that kind, even if under different contexts. If indeed this idea is unique enough (and had a major impact on the design of the algorithm), then this is a significant contribution. Then, the other main issue is with improving the rigor of the presentation and the math. While it's clear what the results are showing, some guessing is needed to fill in some gaps. \n\nThe reward and state structure: I guess that the \"external rewards\" referred to above (1) and the \"same reward\" that all agents share are the same thing. Please clarify and unify the definitions. Then it's not clear what's going on with the dimensions of the local states. The local states appear in the second paragraph of Section 2 and are never properly defined, and also the local state spaces S_i are not well defined. Are the cardinalities of S_i and A_i the same for all i? otherwise, the constraint in (1) isn't clear. Additionally, it's not clear how can one measure distances of states that live in different state spaces (of different agents) as is often being done in the proofs (e.g., the proof of Theorem 1, the definition of D, and so on). On that note, the idea of \"nearby agents\" should be more rigorously defined. Ignoring these gaps, it looks like Theorem 1 basically shows that the better the problem decomposes into local environments, the easier it becomes to solve it distributedly. It's not clear if the math provides any added value on top of this important yet simple observation. \n\nThe connection between theory and practice: Assigning perceived rewards to simplify the MARL problem is an elegant and appealing idea. For this reason, it's important to carefully discuss to what extent this idea actually influenced the algorithm that was tested in practice. Subsection 2.3 raises some concerns in this regard since the algorithm resorts to \"end-to-end learning of Q_i\", in what seems like a total bypass of the idea of the perceived rewards. Looking at equation (3) or (4), one can get the impression that the perceived rewards are just an interpretation of what happens \"inside\" when training the Q-values after splitting them as in (4). By itself, (4) makes a lot of sense and is very natural, so it's not clear if it isn't' easier to come up directly with (4) without knowing anything about perceived rewards. This raises the question: why isn't it possible to bypass the idea of the perceived rewards and motivate the paper based on (4), which is closer to the practical algorithm? answering this question is crucial to claim a significant contribution since otherwise the are two loosely related parts in this paper. \n\nPrecise statements: The statements of the mathematical results often lack some definitions. The statement of theorem 1 doesn't define R_{max} (but Lemma 3 does). The statement of Lemma 3 doesn't define the distance between states (but Lemma 2 does). Please make the statements more standalone and well defined. On the same note, make sure that important definitions don't randomly appear in the paper, sometimes too late (e.g., local states). \n\nExperiments: The experimental results are overall nice and promising.  My only question here is why the resource collection scenario only compares to IQL and not to QTRAN/VDN/QMIX like the StarCraft scenario? \n\nVague sentences and typos:\n\n\"each agent i is acted independently on its own state\" - acting? based on its own state? \n\n(s_1,...,s_N) (bottom of page 2) - needs to be s_K\n\n\"so that\" in Theorem 1 - such that \n\n\"We found that using the observation o_i of agent i covers s_i^{local} works sufficiently well\" - not clear. \n\n\"since 1990s\" - since the 1990s. \n\n\"We sometimes also replace... in Eq.7 by its target to further stabilize training\" (Page 12)- what does sometimes mean? how can one reproduce this?\n\n\"doesn't\" - does not \n\n\"Applying Lemma 1 and notice that all other rewards does not change\" - do not change \n\n\"Define the remote agents s_i^{remote}...\" isn't that a set of states and not of agents? please rephrase. \n\n\"the more distant between relevant rewards istes from remote agents\" - the larger the distance?  \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice idea, but needs some serious clarification and polishing ", "review": "The paper studies a team of agents that collaborate to maximize a global objective, where all agents receive the same reward value as a function of their state and actions. The approach suggested here is to assign each agent a virtual (\"perceived\") reward function such that all the virtual rewards sum up to the actual reward. Then the problem from the point of view of agent i can be solved with local value functions that depend on the local state of agent i and its perceived reward.  It is shown that such a decoupled policy exists that approximates the optimal policy if the state structure decomposes \"well enough\" into local states. Experiments demonstrate the advantages of this approach on a resource collection game on the StarCraft Multi-Agent Challange, with dramatic improvements over existing methods for the scenarios that were tested. \n\nThe paper is mostly easy to follow, and is well-written and well motivated. The idea of this paper is nice and intuitive. Being less familiar with the literature, It's hard for me to judge how novel this idea is. It seems like there must exist multi-agent works that exploit the local structure of the interaction to reduce the dimensionality. Hence, as a start, the contribution of this paper can be enhanced by extending the literature review to include works of that kind, even if under different contexts. If indeed this idea is unique enough (and had a major impact on the design of the algorithm), then this is a significant contribution. Then, the other main issue is with improving the rigor of the presentation and the math. While it's clear what the results are showing, some guessing is needed to fill in some gaps. \n\nThe reward and state structure: I guess that the \"external rewards\" referred to above (1) and the \"same reward\" that all agents share are the same thing. Please clarify and unify the definitions. Then it's not clear what's going on with the dimensions of the local states. The local states appear in the second paragraph of Section 2 and are never properly defined, and also the local state spaces S_i are not well defined. Are the cardinalities of S_i and A_i the same for all i? otherwise, the constraint in (1) isn't clear. Additionally, it's not clear how can one measure distances of states that live in different state spaces (of different agents) as is often being done in the proofs (e.g., the proof of Theorem 1, the definition of D, and so on). On that note, the idea of \"nearby agents\" should be more rigorously defined. Ignoring these gaps, it looks like Theorem 1 basically shows that the better the problem decomposes into local environments, the easier it becomes to solve it distributedly. It's not clear if the math provides any added value on top of this important yet simple observation. \n\nThe connection between theory and practice: Assigning perceived rewards to simplify the MARL problem is an elegant and appealing idea. For this reason, it's important to carefully discuss to what extent this idea actually influenced the algorithm that was tested in practice. Subsection 2.3 raises some concerns in this regard since the algorithm resorts to \"end-to-end learning of Q_i\", in what seems like a total bypass of the idea of the perceived rewards. Looking at equation (3) or (4), one can get the impression that the perceived rewards are just an interpretation of what happens \"inside\" when training the Q-values after splitting them as in (4). By itself, (4) makes a lot of sense and is very natural, so it's not clear if it isn't' easier to come up directly with (4) without knowing anything about perceived rewards. This raises the question: why isn't it possible to bypass the idea of the perceived rewards and motivate the paper based on (4), which is closer to the practical algorithm? answering this question is crucial to claim a significant contribution since otherwise the are two loosely related parts in this paper. \n\nPrecise statements: The statements of the mathematical results often lack some definitions. The statement of theorem 1 doesn't define R_{max} (but Lemma 3 does). The statement of Lemma 3 doesn't define the distance between states (but Lemma 2 does). Please make the statements more standalone and well defined. On the same note, make sure that important definitions don't randomly appear in the paper, sometimes too late (e.g., local states). \n\nExperiments: The experimental results are overall nice and promising.  My only question here is why the resource collection scenario only compares to IQL and not to QTRAN/VDN/QMIX like the StarCraft scenario? \n\nVague sentences and typos:\n\n\"each agent i is acted independently on its own state\" - acting? based on its own state? \n\n(s_1,...,s_N) (bottom of page 2) - needs to be s_K\n\n\"so that\" in Theorem 1 - such that \n\n\"We found that using the observation o_i of agent i covers s_i^{local} works sufficiently well\" - not clear. \n\n\"since 1990s\" - since the 1990s. \n\n\"We sometimes also replace... in Eq.7 by its target to further stabilize training\" (Page 12)- what does sometimes mean? how can one reproduce this?\n\n\"doesn't\" - does not \n\n\"Applying Lemma 1 and notice that all other rewards does not change\" - do not change \n\n\"Define the remote agents s_i^{remote}...\" isn't that a set of states and not of agents? please rephrase. \n\n\"the more distant between relevant rewards istes from remote agents\" - the larger the distance?  \n\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1602920356244}], "openreview_url": "https://openreview.net/forum?id=GVNGAaY2Dr1", "arxiv_id": "2010.08531", "paper_pdf": "papers/GVNGAaY2Dr1.pdf", "paper_pdf_sha256": "af3be00a2a42879548e1f83bb21506e9282fd1eacc565b8ecdd0de404fbe59a8", "paper_pdf_bytes": 3480548, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/CollaQ", "code_repository": "facebookresearch/CollaQ", "code_commit": "ac43314fcf955a21fd21cf644bea864bf4b5013a", "code_archive": "repos/GVNGAaY2Dr1.zip", "code_archive_sha256": "a93407365f39234746ece9400acaa03da19de4ddb6dd41122ac88cc52cfa6158", "code_archive_bytes": 470533, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 2578, "github_languages": {"Python": 124283}, "github_archived": true, "github_pushed_at": "2023-08-14T21:56:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-agent-collaboration-via-reward-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "VD7GaNY1tJ", "year": 2026, "status": "rejected", "title": "Robustly Improving LLM Fairness in Realistic Settings via Interpretability", "authors": ["Adam Karvonen", "Samuel Marks"], "authorids": ["~Adam_Karvonen1", "~Samuel_Marks1"], "authors_source": "OpenReview API", "abstract": "Large language models (LLMs) are increasingly deployed in high-stakes hiring applications, making decisions that directly impact people's careers and livelihoods. While prior studies suggest simple anti-bias prompts can eliminate demographic biases in controlled evaluations, we find these mitigations fail when realistic contextual details are introduced. We address these failures through internal bias mitigation: by identifying and neutralizing sensitive attribute directions within model activations, we achieve robust bias reduction across all tested scenarios.\nAcross leading commercial (GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash) and open-source models (Gemma-2 27B, Gemma-3, Mistral-24B), we find that adding realistic context such as company names, culture descriptions from public careers pages, and selective hiring constraints (e.g.,``only accept candidates in the top 10\\%\") induces significant racial and gender biases (up to 12\\% differences in interview rates). When these biases emerge, they consistently favor Black over White candidates and female over male candidates across all tested models and scenarios. Moreover, models can infer demographics and become biased from subtle cues like college affiliations, with these biases remaining invisible even when inspecting the model's chain-of-thought reasoning.\nTo address these limitations, our internal bias mitigation identifies race and gender-correlated directions and applies affine concept editing at inference time. Despite using directions from a simple synthetic dataset, the intervention generalizes robustly, consistently reducing bias to very low levels (typically under 1\\%, always below 2.5\\%) while largely maintaining model performance.\nOur findings suggest that practitioners deploying LLMs for hiring should adopt more realistic evaluation methodologies and consider internal mitigation strategies for equitable outcomes.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "pRTBJkNgex", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20181/Reviewer_aZFN"], "rating": 2, "soundness": 1, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "The authors study the fairness of LLMs in the context of screening candidate resumes. The paper concludes that, at first glance, there is little (but significant) bias from all the open source and commercial models examined, and this bias can be successfully mitigated using simple anti-bias prompts. However, when the tasks is enriched by providing additional description about the hiring company culture (from the company website) or requesting selective criteria (e.g., only top 10%) suddenly all models exhibit stronger bias towards preferring female over male and black over white candidates. Furthermore, in this scenario the authors claim that simple anti-bias prompt are no longer effective. Instead the authors tried an existing technique (affine concept editing - ACE  from Marshall et al., 2025) for concept editing based on identifying directions in activation space correlated with demographics and at inference time shift activations along those directions to a neutral midpoint. This approach resulted in a more effective bias mitigation than prompting while minimally impacting model performance (measured as MMLU).", "review_text": "The authors study the fairness of LLMs in the context of screening candidate resumes. The paper concludes that, at first glance, there is little (but significant) bias from all the open source and commercial models examined, and this bias can be successfully mitigated using simple anti-bias prompts. However, when the tasks is enriched by providing additional description about the hiring company culture (from the company website) or requesting selective criteria (e.g., only top 10%) suddenly all models exhibit stronger bias towards preferring female over male and black over white candidates. Furthermore, in this scenario the authors claim that simple anti-bias prompt are no longer effective. Instead the authors tried an existing technique (affine concept editing - ACE  from Marshall et al., 2025) for concept editing based on identifying directions in activation space correlated with demographics and at inference time shift activations along those directions to a neutral midpoint. This approach resulted in a more effective bias mitigation than prompting while minimally impacting model performance (measured as MMLU).", "strengths": "- The call for action to use more realistic and challenging evaluation for bias is very important. \n- The overall raising awareness about bias in LLMs, especially when employed for important decision such as hiring is also very important. \n- The authors examine a good set of models both commercial and open sourced.\n- The proposed intervention seemed effective on an individual demographic axes.", "weaknesses": "The paper mentions that biases favor Black over White and female over male in their setting — this is counter to many fairness concerns (which typically focus on disadvantage to historically marginalized groups). When such strong claims are presented the evaluation protocol needs to be extremely solid, clearly explained and results need to be thoroughly analyzed. While this should be true always, in this specific context, it is even more important. Unfortunately I find the paper lacking in all those aspects.\n\nProtocol for the bias evaluation\n- Lack of control for confounders. From the paper it is unclear how the exact task / evaluation was carried out. Specifically, there is the risk that the model is choosing a candidate not because of race or gender but because the resume could have been objectively better. The authors argue that this is not the case because bias is reduced after the concept edit of those demographics, however, this remains an indirect proof and other factors may play a role when one starts to modify the model’s activations. In an extreme example, if I make the decision completely random bias will go away, yet it does not mean that the model was using race/gender to make the decision. Please understand this is just an extreme example to drive my point home. A standard fairness protocol would hold all other features constant and vary only the sensitive attribute. In my opinion the correct way to assess this would be to take the same resume and create two copies, modifying those cues that might induce the model to believe the come from different gender and race candidate and then measure if the model still have a preference. It is not fully clear whether this “counterfactual” style control was achieved, and therefore, why this bias is being measured.\n- One of the main finding is that biases in “realistic” contexts favor Black over White and female over male. The authors proposed some hypothesis about why this might emerge but the protocol does not deeply explore why the bias flips/arise in this direction. With such strong claims one needs a strong protocol (as per my point above) and a deeper analysis of the results to understand what is happening. \n\nProtocol for the mitigation\n- The authors correctly emphasize the importance of realistic tasks. However, the test they conduct analyze independently per axis (either gender OR race) while in practice those axis appear in conjunction. This is important for two aspects: on the one hand intersectional biases tend to be even stronger than single axes biases, on the other hand multi-concept editing is known to be particularly challenging. So there is no evidence that the proposed mitigation, in real tasks scenarios, will actually work. Given the emphasis posed on the realistic tasks this is a rather strong unrealistic simplification. \n- Prompt results are aggregated across all prompts tested. Why? The most interesting results is for each model the result from the best prompt. In fact from the results in the appendix it seems that for every model there exists one prompt that reach 0 bias in both race and gender. \n- The authors focus on Chain of Thought (CoT) but there is no example of few shots which might be more effective. Providing a specific resume where cues injected to imply different gender/race and instruct the model that those two resumes should be valued equally (this is just an example) might actually be more effective than CoT.\n- Some results are difficult to compare. For example Table 5 shows that some prompts seem effective for the acceptance rate task in the “realistic scenario” with additional context from Meta. However the comparable results using the proposed mitigation are not presented.\n- [Minor] While I believe that few shot might work better than CoT, the CoT evaluation seems limited: manual keyword search + GPT-4o automated review of reasoning. The exact methodology for evaluating faithfulness is not fully described (what keywords, what thresholds, how many samples). This part is not clear and rigorous and may benefit from more formal probing.\n\n\nOverstatement / not fair presentation of the results / weak arguments\n- In some cases the concept editing mitigation seems to be overcorrecting. To the extent that it ends up inducing a pro-stereotypical bias of a comparable, or at times even higher magnitude. E.g., Figure 3 (a) Gemma-3 12B and Mistral,  Figure 5 Gemma-2 27B, Figure 7 Gemma-3 12B. Yet these weaknesses are not commented.\n- Some claims are overstated. For example “biases consistently favor Black over White candidates and female over male candidates across all tested models and scenarios” is strong, yet there are example in the Appendix (e.g., Figure 5 and 6) where the bias is in the opposite direction. It would help to tone the statement down given those results. \n- Some arguments seem counter intuitive and weak. For example the authors state that “Internal interventions have intuitive advantages over external methods such as prompting. Real-world hiring contexts are inherently complex and multifaceted, involving countless variations of job descriptions, domains, prompts, and candidate information. Ensuring consistently unbiased responses across every possible input scenario via prompt engineering alone may be unrealistic.”  Surely changing the prompt is much more intuitive and easier to do than performing concept editing. Additionally, a “ complex and multifaceted” task surely requires multi-concept editing which, as mentioned above, the current empirical evaluation do not present (and multi concept editing is known to be challenging).
\n\nWriting\n- Sampling / dataset construction: The setup uses synthetic or canned “candidate resumes” and “job descriptions” plus added context. But the paper gives limited detail  about how the candidate/resume set was constructed (number of candidates, how demographic attributes were randomized, how subtle cues like “College affiliation” were inserted). While Appendix provides some, but not all, this info the main paper could better summarize it. \n- Some of the writing can be controversial and lack adequate citation and support. For example “To signal White candidates, we used the predominantly White institutions (PWIs) Georgetown University and Emory University.” According to who?\n- The paper report standard deviation but it is unclear out of how many runs and what was different in each of them.\n- The exact way the intervention for concept mitigation was constructed is not clear. The authors state “Crucially, our intervention (constructed from explicit name-based demo-graphic signals) successfully mitigates these implicit biases.” But how the synthetic dataset was used is missing.\n- Table 1 shows the MMLU when intervening (I suppose) on race (it’s actually not fully clear). But the equivalent table when intervening on gender is not presented. \n- [Minor] Many key information are in the appendix. For example the exact prompts used appear there but are not referenced from the main paper. In section 3.2 when talking about prompting you should add the reference to the Appendix where the actual prompts are listed.\n\n\nSuggestions
\t\n- You have some hypothesis that adding “only accept candidates in the top 10%” etc… causes the bias to appear but there is no ablation that shows which of these contextual information induce the bias. This would be a great addition to the work.\n- In addition to MMLU it would be great to add Perplexity and some measure of fluency.\n- As mentioned above I don’t believe CoT would be particularly useful. However, an interesting experiments could be to test CoT before and after concept editing and analyze the differences in the CoT. \n- The full work assumes that demographic fairness equates to simply balancing acceptance rates by race/gender. This is a perhaps understandable simplification yet this aspect should be discussed in the limitation section (and maybe even mentioned in the introduction).", "questions": "Unfortunately all the points above would need to be addressed and I don't believe this is doable within a rebuttal. Even assuming that the protocol is correct and it is simply a matter of better explanation that would completely change the manuscript and would require a new review. However, I remain open to the fact that I might completely miss understood the work and would be happy to discuss it with the authors. \n\nDuring the rebuttal please don’t worry about all the points tagged as “Minor” nor the Suggestions. Those are for you to consider in order to make the work stronger.", "flag_for_ethics_review": ["Yes, Discrimination / bias / fairness concerns"], "all_content": {"summary": "The authors study the fairness of LLMs in the context of screening candidate resumes. The paper concludes that, at first glance, there is little (but significant) bias from all the open source and commercial models examined, and this bias can be successfully mitigated using simple anti-bias prompts. However, when the tasks is enriched by providing additional description about the hiring company culture (from the company website) or requesting selective criteria (e.g., only top 10%) suddenly all models exhibit stronger bias towards preferring female over male and black over white candidates. Furthermore, in this scenario the authors claim that simple anti-bias prompt are no longer effective. Instead the authors tried an existing technique (affine concept editing - ACE  from Marshall et al., 2025) for concept editing based on identifying directions in activation space correlated with demographics and at inference time shift activations along those directions to a neutral midpoint. This approach resulted in a more effective bias mitigation than prompting while minimally impacting model performance (measured as MMLU).", "soundness": 1, "presentation": 1, "contribution": 1, "strengths": "- The call for action to use more realistic and challenging evaluation for bias is very important. \n- The overall raising awareness about bias in LLMs, especially when employed for important decision such as hiring is also very important. \n- The authors examine a good set of models both commercial and open sourced.\n- The proposed intervention seemed effective on an individual demographic axes.", "weaknesses": "The paper mentions that biases favor Black over White and female over male in their setting — this is counter to many fairness concerns (which typically focus on disadvantage to historically marginalized groups). When such strong claims are presented the evaluation protocol needs to be extremely solid, clearly explained and results need to be thoroughly analyzed. While this should be true always, in this specific context, it is even more important. Unfortunately I find the paper lacking in all those aspects.\n\nProtocol for the bias evaluation\n- Lack of control for confounders. From the paper it is unclear how the exact task / evaluation was carried out. Specifically, there is the risk that the model is choosing a candidate not because of race or gender but because the resume could have been objectively better. The authors argue that this is not the case because bias is reduced after the concept edit of those demographics, however, this remains an indirect proof and other factors may play a role when one starts to modify the model’s activations. In an extreme example, if I make the decision completely random bias will go away, yet it does not mean that the model was using race/gender to make the decision. Please understand this is just an extreme example to drive my point home. A standard fairness protocol would hold all other features constant and vary only the sensitive attribute. In my opinion the correct way to assess this would be to take the same resume and create two copies, modifying those cues that might induce the model to believe the come from different gender and race candidate and then measure if the model still have a preference. It is not fully clear whether this “counterfactual” style control was achieved, and therefore, why this bias is being measured.\n- One of the main finding is that biases in “realistic” contexts favor Black over White and female over male. The authors proposed some hypothesis about why this might emerge but the protocol does not deeply explore why the bias flips/arise in this direction. With such strong claims one needs a strong protocol (as per my point above) and a deeper analysis of the results to understand what is happening. \n\nProtocol for the mitigation\n- The authors correctly emphasize the importance of realistic tasks. However, the test they conduct analyze independently per axis (either gender OR race) while in practice those axis appear in conjunction. This is important for two aspects: on the one hand intersectional biases tend to be even stronger than single axes biases, on the other hand multi-concept editing is known to be particularly challenging. So there is no evidence that the proposed mitigation, in real tasks scenarios, will actually work. Given the emphasis posed on the realistic tasks this is a rather strong unrealistic simplification. \n- Prompt results are aggregated across all prompts tested. Why? The most interesting results is for each model the result from the best prompt. In fact from the results in the appendix it seems that for every model there exists one prompt that reach 0 bias in both race and gender. \n- The authors focus on Chain of Thought (CoT) but there is no example of few shots which might be more effective. Providing a specific resume where cues injected to imply different gender/race and instruct the model that those two resumes should be valued equally (this is just an example) might actually be more effective than CoT.\n- Some results are difficult to compare. For example Table 5 shows that some prompts seem effective for the acceptance rate task in the “realistic scenario” with additional context from Meta. However the comparable results using the proposed mitigation are not presented.\n- [Minor] While I believe that few shot might work better than CoT, the CoT evaluation seems limited: manual keyword search + GPT-4o automated review of reasoning. The exact methodology for evaluating faithfulness is not fully described (what keywords, what thresholds, how many samples). This part is not clear and rigorous and may benefit from more formal probing.\n\n\nOverstatement / not fair presentation of the results / weak arguments\n- In some cases the concept editing mitigation seems to be overcorrecting. To the extent that it ends up inducing a pro-stereotypical bias of a comparable, or at times even higher magnitude. E.g., Figure 3 (a) Gemma-3 12B and Mistral,  Figure 5 Gemma-2 27B, Figure 7 Gemma-3 12B. Yet these weaknesses are not commented.\n- Some claims are overstated. For example “biases consistently favor Black over White candidates and female over male candidates across all tested models and scenarios” is strong, yet there are example in the Appendix (e.g., Figure 5 and 6) where the bias is in the opposite direction. It would help to tone the statement down given those results. \n- Some arguments seem counter intuitive and weak. For example the authors state that “Internal interventions have intuitive advantages over external methods such as prompting. Real-world hiring contexts are inherently complex and multifaceted, involving countless variations of job descriptions, domains, prompts, and candidate information. Ensuring consistently unbiased responses across every possible input scenario via prompt engineering alone may be unrealistic.”  Surely changing the prompt is much more intuitive and easier to do than performing concept editing. Additionally, a “ complex and multifaceted” task surely requires multi-concept editing which, as mentioned above, the current empirical evaluation do not present (and multi concept editing is known to be challenging).
\n\nWriting\n- Sampling / dataset construction: The setup uses synthetic or canned “candidate resumes” and “job descriptions” plus added context. But the paper gives limited detail  about how the candidate/resume set was constructed (number of candidates, how demographic attributes were randomized, how subtle cues like “College affiliation” were inserted). While Appendix provides some, but not all, this info the main paper could better summarize it. \n- Some of the writing can be controversial and lack adequate citation and support. For example “To signal White candidates, we used the predominantly White institutions (PWIs) Georgetown University and Emory University.” According to who?\n- The paper report standard deviation but it is unclear out of how many runs and what was different in each of them.\n- The exact way the intervention for concept mitigation was constructed is not clear. The authors state “Crucially, our intervention (constructed from explicit name-based demo-graphic signals) successfully mitigates these implicit biases.” But how the synthetic dataset was used is missing.\n- Table 1 shows the MMLU when intervening (I suppose) on race (it’s actually not fully clear). But the equivalent table when intervening on gender is not presented. \n- [Minor] Many key information are in the appendix. For example the exact prompts used appear there but are not referenced from the main paper. In section 3.2 when talking about prompting you should add the reference to the Appendix where the actual prompts are listed.\n\n\nSuggestions
\t\n- You have some hypothesis that adding “only accept candidates in the top 10%” etc… causes the bias to appear but there is no ablation that shows which of these contextual information induce the bias. This would be a great addition to the work.\n- In addition to MMLU it would be great to add Perplexity and some measure of fluency.\n- As mentioned above I don’t believe CoT would be particularly useful. However, an interesting experiments could be to test CoT before and after concept editing and analyze the differences in the CoT. \n- The full work assumes that demographic fairness equates to simply balancing acceptance rates by race/gender. This is a perhaps understandable simplification yet this aspect should be discussed in the limitation section (and maybe even mentioned in the introduction).", "questions": "Unfortunately all the points above would need to be addressed and I don't believe this is doable within a rebuttal. Even assuming that the protocol is correct and it is simply a matter of better explanation that would completely change the manuscript and would require a new review. However, I remain open to the fact that I might completely miss understood the work and would be happy to discuss it with the authors. \n\nDuring the rebuttal please don’t worry about all the points tagged as “Minor” nor the Suggestions. Those are for you to consider in order to make the work stronger.", "flag_for_ethics_review": ["Yes, Discrimination / bias / fairness concerns"], "details_of_ethics_concerns": "The paper concludes that biases favor Black over White and female over male in their setting — this is counter to many fairness concerns (which typically focus on disadvantage to historically marginalized groups). However such strong conclusions are not supported by a strong and clear protocol evaluation. \n\nI am not sure this is \"just\" a reason to reject or if Ethics review needs to be carried out.", "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761912823779}, {"id": "dIz9eHzFaR", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20181/Reviewer_TqmZ"], "rating": 2, "soundness": 1, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper investigates biases in LLMs in the context of hiring decisions. The authors first show that prompt-based debiasing techniques can improve model fairness only superficially, while they reduce bias in simple prompt settings, the bias reemerges (by up to 12%) once additional contextual details are introduced into the same task. To address this, the authors propose a representation-level intervention that edits demographic concepts in the model’s internal representations at inference time. This approach aims to directly suppress biased latent associations without retraining. The authors demonstrate that this intervention reduces bias in 4x models while maintaining minimal degradation in general model performance (MMLU).", "review_text": "This paper investigates biases in LLMs in the context of hiring decisions. The authors first show that prompt-based debiasing techniques can improve model fairness only superficially, while they reduce bias in simple prompt settings, the bias reemerges (by up to 12%) once additional contextual details are introduced into the same task. To address this, the authors propose a representation-level intervention that edits demographic concepts in the model’s internal representations at inference time. This approach aims to directly suppress biased latent associations without retraining. The authors demonstrate that this intervention reduces bias in 4x models while maintaining minimal degradation in general model performance (MMLU).", "strengths": "1. Open-source contribution: The authors release their codebase, data and method, which supports transparency and allows for reproducibility and future extensions.\n\n2. Problem relevance: Bias in LLM-based hiring systems is an important and timely issue. The focus on robustness of debiasing under varying context complexity is conceptually interesting.", "weaknesses": "1. Overstatement of realism: The paper overclaims the “real-world” nature of its simulated hiring settings. Adding elements like company names, cultural descriptions, or hiring constraints (e.g., “hire the top 10% of candidates”) adds contextual richness, but it does not necessarily make the task realistic. The authors provide no evidence (e.g., comparisons to real hiring data or expert validation) to support that these additions meaningfully increase realism.\n\n2. Prompt debiasing claims conflict with prior work: The finding that prompt-based debiasing is effective contrasts with several prior works that show the opposite (e.g., Rethinking Prompt-based Debiasing in LLMs, Yang et al.; Social Bias Evaluation for LLMs Requires Prompt Variations, Hida et al.). The paper should reconcile these discrepancies or more carefully contextualize its findings in light of this literature.\n\n3. Missing prompt details: The authors mention using four debiasing prompts but fail to disclose their content. This omission significantly limits reproducibility and makes it impossible to assess whether the prompts are reasonable or comparable to prior work.\n\n4. Insufficient baselines: The paper only tests vanilla prompt-based debiasing, ignoring stronger and better-established prompt-debasing baselines such as iterative self-debiasing (Gallegos et al., Self-Debiasing LLMs), implication and self-refinement prompts (Furniturewala et al., Thinking Fair and Slow). Without such comparisons, the conclusion that prompt-based methods are inadequate under context variation is not sufficiently supported.\n\n5. Loose use of interpretability: The paper claims interpretability benefits from its latent editing method but provides no concrete interpretability evidence. E.g., analysis of sparsity, or semantic coherence of modified directions. If interpretability is a claimed contribution, this must be empirically demonstrated.\n\n6. Choice of solution requires justification: the proposed method resembles existing inference-time steering approaches such as FairSteer (Li et al.), which also modulate demographic activation directions but allow for dynamic control of bias levels. The authors should explain why their approach is preferable or complementary to these alternatives.\n\n7. Limited Fig 1 results: Figure 1 only reports results on three models from the Gemma family and Mistral. It is unclear why common benchmarks such as GPT, Claude, or Gemini are omitted. Broader model coverage would strengthen generality claims.\n\n8. Single dataset limitation: Experiments rely on a single dataset, limiting claims of robustness and generalization. Bias evaluation across multiple hiring datasets or tasks would make conclusions more credible.", "questions": "1. The paper reports “anti-stereotyping” effects; that is, cases where Black or female candidates are preferred over white or male candidates. What is the occupation distribution in these results? Are the occupations predominantly counter-stereotypical (e.g., women in male-dominated fields)? This finding requires clearer interpretation, ideally supported by a breakdown of results by occupation category.\n\n2. What do the authors mean by this: “whitening” demographic directions?\n\n3. The authors state that “RL-trained reasoning models may exhibit more faithful CoT, but we found no evidence of this.” However, reasoning models are not evaluated in the paper. This claim lacks empirical grounding and should be removed or should be further supported.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates biases in LLMs in the context of hiring decisions. The authors first show that prompt-based debiasing techniques can improve model fairness only superficially, while they reduce bias in simple prompt settings, the bias reemerges (by up to 12%) once additional contextual details are introduced into the same task. To address this, the authors propose a representation-level intervention that edits demographic concepts in the model’s internal representations at inference time. This approach aims to directly suppress biased latent associations without retraining. The authors demonstrate that this intervention reduces bias in 4x models while maintaining minimal degradation in general model performance (MMLU).", "soundness": 1, "presentation": 2, "contribution": 2, "strengths": "1. Open-source contribution: The authors release their codebase, data and method, which supports transparency and allows for reproducibility and future extensions.\n\n2. Problem relevance: Bias in LLM-based hiring systems is an important and timely issue. The focus on robustness of debiasing under varying context complexity is conceptually interesting.", "weaknesses": "1. Overstatement of realism: The paper overclaims the “real-world” nature of its simulated hiring settings. Adding elements like company names, cultural descriptions, or hiring constraints (e.g., “hire the top 10% of candidates”) adds contextual richness, but it does not necessarily make the task realistic. The authors provide no evidence (e.g., comparisons to real hiring data or expert validation) to support that these additions meaningfully increase realism.\n\n2. Prompt debiasing claims conflict with prior work: The finding that prompt-based debiasing is effective contrasts with several prior works that show the opposite (e.g., Rethinking Prompt-based Debiasing in LLMs, Yang et al.; Social Bias Evaluation for LLMs Requires Prompt Variations, Hida et al.). The paper should reconcile these discrepancies or more carefully contextualize its findings in light of this literature.\n\n3. Missing prompt details: The authors mention using four debiasing prompts but fail to disclose their content. This omission significantly limits reproducibility and makes it impossible to assess whether the prompts are reasonable or comparable to prior work.\n\n4. Insufficient baselines: The paper only tests vanilla prompt-based debiasing, ignoring stronger and better-established prompt-debasing baselines such as iterative self-debiasing (Gallegos et al., Self-Debiasing LLMs), implication and self-refinement prompts (Furniturewala et al., Thinking Fair and Slow). Without such comparisons, the conclusion that prompt-based methods are inadequate under context variation is not sufficiently supported.\n\n5. Loose use of interpretability: The paper claims interpretability benefits from its latent editing method but provides no concrete interpretability evidence. E.g., analysis of sparsity, or semantic coherence of modified directions. If interpretability is a claimed contribution, this must be empirically demonstrated.\n\n6. Choice of solution requires justification: the proposed method resembles existing inference-time steering approaches such as FairSteer (Li et al.), which also modulate demographic activation directions but allow for dynamic control of bias levels. The authors should explain why their approach is preferable or complementary to these alternatives.\n\n7. Limited Fig 1 results: Figure 1 only reports results on three models from the Gemma family and Mistral. It is unclear why common benchmarks such as GPT, Claude, or Gemini are omitted. Broader model coverage would strengthen generality claims.\n\n8. Single dataset limitation: Experiments rely on a single dataset, limiting claims of robustness and generalization. Bias evaluation across multiple hiring datasets or tasks would make conclusions more credible.", "questions": "1. The paper reports “anti-stereotyping” effects; that is, cases where Black or female candidates are preferred over white or male candidates. What is the occupation distribution in these results? Are the occupations predominantly counter-stereotypical (e.g., women in male-dominated fields)? This finding requires clearer interpretation, ideally supported by a breakdown of results by occupation category.\n\n2. What do the authors mean by this: “whitening” demographic directions?\n\n3. The authors state that “RL-trained reasoning models may exhibit more faithful CoT, but we found no evidence of this.” However, reasoning models are not evaluated in the paper. This claim lacks empirical grounding and should be removed or should be further supported.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761883206675}, {"id": "aJz54DwbFI", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission20181/Reviewer_Y857"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper demonstrates that simple prompt-based anti-bias mitigations for LLMs, which work well in controlled evaluations, fail when realistic contextual details are introduced to hiring scenarios. The authors show that adding company-specific information, culture descriptions, and selective hiring constraints induces significant racial and gender biases (up to 12% differences in interview rates) across leading commercial and open-source models. To address this, they propose internal bias mitigation using affine concept editing on model activations, which robustly reduces bias to under 2.5% while maintaining model performance.", "review_text": "This paper demonstrates that simple prompt-based anti-bias mitigations for LLMs, which work well in controlled evaluations, fail when realistic contextual details are introduced to hiring scenarios. The authors show that adding company-specific information, culture descriptions, and selective hiring constraints induces significant racial and gender biases (up to 12% differences in interview rates) across leading commercial and open-source models. To address this, they propose internal bias mitigation using affine concept editing on model activations, which robustly reduces bias to under 2.5% while maintaining model performance.", "strengths": "- Important real-world problem: The paper addresses a critical issue as LLMs are increasingly deployed in high-stakes hiring applications with direct impact on people's livelihoods.\n\n- Strong empirical findings: The demonstration that prompt-based mitigations become brittle under realistic conditions is well-documented across multiple models and scenarios.\n\n- Robust internal intervention: The proposed affine concept editing approach shows consistent effectiveness across different contexts.\n\n- Comprehensive evaluation: The experiments cover multiple commercial and open-source models, various contextual conditions, and include both explicit and implicit demographic indicators.", "weaknesses": "- Dataset quality issues: The authors acknowledge in Appendix E that 22% of resumes contained unintended demographic indicators, though they claim minimal impact on results.\n\n- Mechanistic clarity and design choices. Directions are estimated from synthetic data and applied at all layers/tokens. The paper doesn’t ablate which layers matter, how many directions per attribute are needed, or compare ACE against other linear methods (e.g., LEACE, NOP, DAS) in this setting. Gemma-3 sensitivity indicates model-specific fragility that deserves deeper analysis.", "questions": "- How were company culture blurbs sourced, token-length-matched, and cleaned (e.g., removal of DEI phrases)? Do results persist under length-matched neutral placebos or shuffled company/job pairings?\n\n- ACE application details: which layers carry most of the effect, and does restricting ACE to those layers recover similar mitigation with less capability impact (especially on Gemma-3)?\n\n- How stable are the identified bias directions across different random seeds and prompt orderings?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper demonstrates that simple prompt-based anti-bias mitigations for LLMs, which work well in controlled evaluations, fail when realistic contextual details are introduced to hiring scenarios. The authors show that adding company-specific information, culture descriptions, and selective hiring constraints induces significant racial and gender biases (up to 12% differences in interview rates) across leading commercial and open-source models. To address this, they propose internal bias mitigation using affine concept editing on model activations, which robustly reduces bias to under 2.5% while maintaining model performance.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- Important real-world problem: The paper addresses a critical issue as LLMs are increasingly deployed in high-stakes hiring applications with direct impact on people's livelihoods.\n\n- Strong empirical findings: The demonstration that prompt-based mitigations become brittle under realistic conditions is well-documented across multiple models and scenarios.\n\n- Robust internal intervention: The proposed affine concept editing approach shows consistent effectiveness across different contexts.\n\n- Comprehensive evaluation: The experiments cover multiple commercial and open-source models, various contextual conditions, and include both explicit and implicit demographic indicators.", "weaknesses": "- Dataset quality issues: The authors acknowledge in Appendix E that 22% of resumes contained unintended demographic indicators, though they claim minimal impact on results.\n\n- Mechanistic clarity and design choices. Directions are estimated from synthetic data and applied at all layers/tokens. The paper doesn’t ablate which layers matter, how many directions per attribute are needed, or compare ACE against other linear methods (e.g., LEACE, NOP, DAS) in this setting. Gemma-3 sensitivity indicates model-specific fragility that deserves deeper analysis.", "questions": "- How were company culture blurbs sourced, token-length-matched, and cleaned (e.g., removal of DEI phrases)? Do results persist under length-matched neutral placebos or shuffled company/job pairings?\n\n- ACE application details: which layers carry most of the effect, and does restricting ACE to those layers recover similar mitigation with less capability impact (especially on Gemma-3)?\n\n- How stable are the identified bias directions across different random seeds and prompt orderings?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760682436399}], "openreview_url": "https://openreview.net/forum?id=VD7GaNY1tJ", "arxiv_id": "2506.10922", "paper_pdf": "papers/VD7GaNY1tJ.pdf", "paper_pdf_sha256": "108fc54bdbe9240959f5c11a2bc8c4bcfadc28a3504aefc740780cc8bf177985", "paper_pdf_bytes": 481133, "paper_pdf_source": "openreview", "code_url": "https://github.com/adamkarvonen/llm_bias", "code_repository": "adamkarvonen/llm_bias", "code_commit": "11ce50baccd7fdf4f8eca6936a5e8682a5df4c1b", "code_archive": "repos/VD7GaNY1tJ.zip", "code_archive_sha256": "2a1229891c5e75001cc47e715f36fe7af4a910f797d563810c49376e6cce0add", "code_archive_bytes": 8546158, "code_file_count": 35, "code_extensions": {".py": 27, ".ipynb": 8}, "github_disk_usage_kb": 8359, "github_languages": {"Python": 277378, "Jupyter Notebook": 217687}, "github_archived": false, "github_pushed_at": "2025-10-03T15:14:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/robustly-improving-llm-fairness-in-realistic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DQfHkEcUqV", "year": 2025, "status": "rejected", "title": "Learning Extrapolative Sequence Transformations from Markov Chains", "authors": ["Sophia Hager", "Aleem Khan", "Andrew Wang", "Nicholas Andrews"], "authorids": ["~Sophia_Hager1", "~Aleem_Khan1", "~Andrew_Wang3", "~Nicholas_Andrews2"], "authors_source": "OpenReview API", "abstract": "Most successful applications of deep learning involve similar training and test conditions. However, for some generative tasks, samples should improve desirable properties beyond previously known values, which requires the ability to generate novel hypotheses that extrapolate beyond training data. While large language models have been successfully extended to a variety of sequence modeling problems, greedy autoregressive sampling can struggle to explore the solution space sufficiently to extrapolate, especially when the properties of interest are global to the sequence. On the other hand, sequence-level sampling methods such as Markov chain Monte Carlo (MCMC) offer theoretical guarantees about capturing the distribution of interest, but suffer from the curse of dimensionality in discrete structured spaces. We propose a new approach that bridges the gap between MCMC and autoregressive sampling, which may be viewed as off-policy reinforcement learning. Our approach uses selected states from Markov chains as a source of training data for an autoregressive inference network, which is then able to generate novel sequences at test time that extrapolate along the sequence-level properties of interest. The proposed approach is validated on three problems: protein sequence design, text sentiment control, and text anonymization. We find that the learned inference network confers many of the same (and sometimes better) generalization benefits compared to the slow sampling process, but with the additional benefit of high sample efficiency.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "bxYoyJzbbF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11835/Reviewer_iCrL"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a method for learning sample-efficient extrapolative sequence transformations by first using Markov Chain Monte Carlo (MCMC) sampling to explore the solution space, then training a separate inference network on selected states from these chains. The authors demonstrate their approach on three tasks: protein sequence design, text sentiment control, and text anonymization. The key innovation is sub-sampling informative state transitions from MCMC chains to train a more efficient model that can achieve similar or better performance with significantly fewer inference steps.", "review_text": "This paper proposes a method for learning sample-efficient extrapolative sequence transformations by first using Markov Chain Monte Carlo (MCMC) sampling to explore the solution space, then training a separate inference network on selected states from these chains. The authors demonstrate their approach on three tasks: protein sequence design, text sentiment control, and text anonymization. The key innovation is sub-sampling informative state transitions from MCMC chains to train a more efficient model that can achieve similar or better performance with significantly fewer inference steps.", "strengths": "- Tackles an important problem (extrapolative generation) with a novel approach\n- Impressive empirical results, especially on protein engineering\n- Thorough ablation studies examining different components\n- Clear practical benefits in terms of sample efficiency", "weaknesses": "1. The validation methodology for protein engineering is fundamentally flawed. Without a separate validation set, there's no reliable way to assess generalization ability. While the authors acknowledge this and attempt some mitigation strategies, they don't address the core issue of potential overfitting to the test set through architecture and hyperparameter choices.\n\n2. The theoretical justification for extrapolation capabilities is weak. Although they discuss how their approach captures dependencies between hidden states, they fail to provide a compelling explanation for why this enables better extrapolation. The paper lacks formal analysis of extrapolation capabilities or theoretical bounds on performance.\n\n3. The baseline comparisons are inadequate, particularly in protein engineering where they only compare against two baselines from a single paper. This omits numerous recent protein design methods and alternative approaches. While sentiment and anonymization tasks include more baselines, they still miss obvious comparisons with state-of-the-art methods.\n\n4. Several key ablation studies are missing. There's no exploration of different MCMC sampling strategies, limited analysis of energy function designs, and no investigation of how performance scales with sequence length. These missing analyses make it harder to understand which components are crucial for success.\n\n5. The energy function design choices feel arbitrary and aren't well justified. The weighting between different terms lacks thorough explanation, and there's no meaningful discussion of how to design effective energy functions for new tasks, limiting the method's applicability to other domains.\n\n6. Memory consumption issues are acknowledged but not adequately addressed. The authors note their models use significantly more memory than alternatives but offer no solutions. The analysis lacks discussion of how usage scales with sequence length or batch size, and fails to analyze memory-computation tradeoffs.\n\n7. Reproducibility and practical concerns are significant. Many implementation details are buried in appendices, hyperparameter sensitivity isn't thoroughly analyzed, and there's limited discussion of failure cases or scalability to longer sequences. These omissions make it difficult for others to build upon their work or apply it to real-world problems.", "questions": "- Why not include more baselines, especially recent work in protein engineering?\n- How sensitive is the method to the choice of energy function?\n- Have you explored using curriculum learning approaches for training?\n- How does performance scale with sequence length?\n- What are the key factors limiting memory efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method for learning sample-efficient extrapolative sequence transformations by first using Markov Chain Monte Carlo (MCMC) sampling to explore the solution space, then training a separate inference network on selected states from these chains. The authors demonstrate their approach on three tasks: protein sequence design, text sentiment control, and text anonymization. The key innovation is sub-sampling informative state transitions from MCMC chains to train a more efficient model that can achieve similar or better performance with significantly fewer inference steps.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- Tackles an important problem (extrapolative generation) with a novel approach\n- Impressive empirical results, especially on protein engineering\n- Thorough ablation studies examining different components\n- Clear practical benefits in terms of sample efficiency", "weaknesses": "1. The validation methodology for protein engineering is fundamentally flawed. Without a separate validation set, there's no reliable way to assess generalization ability. While the authors acknowledge this and attempt some mitigation strategies, they don't address the core issue of potential overfitting to the test set through architecture and hyperparameter choices.\n\n2. The theoretical justification for extrapolation capabilities is weak. Although they discuss how their approach captures dependencies between hidden states, they fail to provide a compelling explanation for why this enables better extrapolation. The paper lacks formal analysis of extrapolation capabilities or theoretical bounds on performance.\n\n3. The baseline comparisons are inadequate, particularly in protein engineering where they only compare against two baselines from a single paper. This omits numerous recent protein design methods and alternative approaches. While sentiment and anonymization tasks include more baselines, they still miss obvious comparisons with state-of-the-art methods.\n\n4. Several key ablation studies are missing. There's no exploration of different MCMC sampling strategies, limited analysis of energy function designs, and no investigation of how performance scales with sequence length. These missing analyses make it harder to understand which components are crucial for success.\n\n5. The energy function design choices feel arbitrary and aren't well justified. The weighting between different terms lacks thorough explanation, and there's no meaningful discussion of how to design effective energy functions for new tasks, limiting the method's applicability to other domains.\n\n6. Memory consumption issues are acknowledged but not adequately addressed. The authors note their models use significantly more memory than alternatives but offer no solutions. The analysis lacks discussion of how usage scales with sequence length or batch size, and fails to analyze memory-computation tradeoffs.\n\n7. Reproducibility and practical concerns are significant. Many implementation details are buried in appendices, hyperparameter sensitivity isn't thoroughly analyzed, and there's limited discussion of failure cases or scalability to longer sequences. These omissions make it difficult for others to build upon their work or apply it to real-world problems.", "questions": "- Why not include more baselines, especially recent work in protein engineering?\n- How sensitive is the method to the choice of energy function?\n- Have you explored using curriculum learning approaches for training?\n- How does performance scale with sequence length?\n- What are the key factors limiting memory efficiency?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730729101397}, {"id": "VrsDDUF3Rx", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11835/Reviewer_dU2W"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 3, "confidence": 4, "summary": "This paper studies the problem of intractable inference in sequence models. It is proposed to sample MCMC chains using an infilling LM as a proposal, to create subchains using various procedures (e.g., selecting the points where the target energy decreases), and to train an autoregressive model on these subchains. This model can then be used to sample modes of the target distribution more efficiently. This method is evaluated on a protein design task and two language generation (editing) tasks and shows somewhat promising results.", "review_text": "This paper studies the problem of intractable inference in sequence models. It is proposed to sample MCMC chains using an infilling LM as a proposal, to create subchains using various procedures (e.g., selecting the points where the target energy decreases), and to train an autoregressive model on these subchains. This model can then be used to sample modes of the target distribution more efficiently. This method is evaluated on a protein design task and two language generation (editing) tasks and shows somewhat promising results.", "strengths": "- The procedure of amortising subchains of MCMC chains by a non-Markovian sequence model is interesting and could be applicable to structured sampling problems well beyond the protein and language tasks studied here. \n  - For example, chain-of-thought reasoning, among other intractable inference problems in language, has been interpreted as latent variable inference and addressed using MCMC ([Phan et al.,  NeurIPS'23](https://arxiv.org/abs/2312.02179), [Lew et al.](https://arxiv.org/abs/2306.03081)), amortisation ([Hu et al., ICLR'24](https://arxiv.org/abs/2310.04363)), hybrid methods ([Zhao et al., ICML'24](https://arxiv.org/abs/2404.17546)), and distillation into tractable models ([Zhang et al., 2024](https://arxiv.org/abs/2406.13892)).\n- The inclusion of both LM and biological sequence design tasks is a good way to show the generality (although toy experiments on something synthetic and *very* simple, where the target is well-understood, would also be helpful).\n- Results generally comparable with prior work in each task with (possibly -- see questions below) higher efficiency of inference.", "weaknesses": "Overall, both the writing and the experiments need substantial improvement. It is quite difficult to understand the algorithm and the experiment setup, nor are the results particularly strong.\n\n- The writing in the first two pages is imprecise and does not set up the problem well.\n  - The first sentence of the abstract does not make sense to me. (Doesn't \"desired outputs\" already presuppose we are talking about generative modelling -- so what is the meaning of saying generative models are \"appealing\"? \"Sequence-level\" as opposed to what?) \n    - In general, the abstract does not explain well the problem setting and does not prime the reader to expect what is to come. The reader is left wondering: Are we talking about generative modelling from data or sampling given a density? Are we training a new model or constraining an existing one?\n  - The introduction does not set up the problem to be solved in a clear way. \n    - The first 2.5 paragraphs or so of the introduction seem to be saying approximately \"generalisation is important\". \n    - However, the **main objects** involved in the problem setting are not introduced. Thus it is not clear what is being talked about when \"original\" and \"sampled\" sequences are mentioned, what the EBM has to do with infilling, etc.\n    - Second paragraph: Lu et al. and Holtzman et al. are inaccurately described. The latter studies strategies to clip the tail of the next-token distributions, while Lu et al. studies MCTS as proposed by reference [59] in that paper, which seems to have no connection to such clipping.\n    - In the last paragraph of the intro, $q_\\theta$ appears, but it is never defined.\n- Section 2: Please define all symbols when they are first used. As far as I could infer:\n  - $\\cal X$: space of sequences\n  - $\\cal Y$: set of \"properties\"\n  - $o:{\\cal X}\\to{\\cal Y}$: oracle assigning a property to every sequence\n  - ${\\cal X}^{\\rm train}$: training set\n  - ${\\cal Y}^{\\rm train}$: unclear; is it $\\{o(x):x\\in{\\cal X}^{\\rm train}\\}$? That is, we wish to generate sequences whose properties, as computed by an oracle, were not observed in the training set?\n- Section 3:\n  - Is the score giving the density or the log-density? Contradiction between \"probability proportional to sequence-level score $s(x)$\" (L144) and the equation in L149.\n  - The following does not make sense to me (LL156-158): \"masked language modeling objectives implicitly define EBM with the masked-infilling objective serving as a proposal distribution\".\n    - How can a [training] **objective** define a distribution? Do you mean that the trained model defines a distribution? Same in LL162-163: the self-supervised pre-training process seems to have nothing to do with the use of the trained infilling model (so how can we say that MH sampling is based on the self-supervised training?).\n    - An EBM is defined by an energy function, not by a proposal distribution used to perform MH on it. Do you mean that the infilling model approximates a collapsed Gibbs sampler for a certain EBM? Or do you mean that the infilling model is an effective proposal for MH sampling of EBMs that happened to be trained on similar data?\n    - Even in the former case, things are somewhat more subtle than this. The conditional distributions given by infilling may not be compatible with any joint distribution, see, for example, [Henningen and Kim, ACL'23](https://arxiv.org/abs/2305.15501) and [Wang and Cho, NeuralGen@NAACL'19](https://arxiv.org/abs/1902.04094).\n    - In the end, I am left confused about the setting: the infilling model defines a EBM (L157) and is also a proposal distribution that is used to propose transitions (L160), but the proposals are accepted/rejected using the MH criterion on a EBM (L161), which seems to require an explicit energy model to compute the acceptance probability.\n  - A little diagram would help to understand the different ways of forming trajectories for training $q_\\theta$. (It is also confusing that $q$ (proposal) and $q_\\theta$ (amortised sampler) are denoted by the same letter.)\n- Experiments:\n  - Missing details:\n    - In each experiment, please specify exactly, in an equation, what the target energy model is. I found some notes on this in Appendix D, but they left me with more questions. In general, I would suggest to move some experiment details to the appendix, but state as directly as possible what sampling problem is being solved.\n    - It is not stated which version of training data for $q_\\theta$ is used. It can be inferred from the ablation study in Section 5, but would be good to make explicit. \n  - I am not convinced by evaluations in 4.1:\n    - The fluency is not reported for ICE.\n    - Extrapolation metric: why should we expect that \"make this negative\" would lead necessarily to 1-star and not to 2-star reviews? It would be more informative to compare the full distributions of the base model and with editing, for each editing method.\n      - How to understand that the second and third columns in Table 2 sum to more than 1, if they are the proportions of reviews with (in the example of negative sentiment) 2 and 1 stars, respectively?\n      - In fact, generating **too many** 1-star reviews, together with the high fluency, could actually indicate mode collapse. There is no diversity metric to show that this is not happening.\n  - The comment on diversity also applies to Section 4.2. Why was ICE not considered in this problem?\n  - How do the methods compare in execution wall time (both for training and drawing a single sample)?", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of intractable inference in sequence models. It is proposed to sample MCMC chains using an infilling LM as a proposal, to create subchains using various procedures (e.g., selecting the points where the target energy decreases), and to train an autoregressive model on these subchains. This model can then be used to sample modes of the target distribution more efficiently. This method is evaluated on a protein design task and two language generation (editing) tasks and shows somewhat promising results.", "soundness": 2, "presentation": 1, "contribution": 3, "strengths": "- The procedure of amortising subchains of MCMC chains by a non-Markovian sequence model is interesting and could be applicable to structured sampling problems well beyond the protein and language tasks studied here. \n  - For example, chain-of-thought reasoning, among other intractable inference problems in language, has been interpreted as latent variable inference and addressed using MCMC ([Phan et al.,  NeurIPS'23](https://arxiv.org/abs/2312.02179), [Lew et al.](https://arxiv.org/abs/2306.03081)), amortisation ([Hu et al., ICLR'24](https://arxiv.org/abs/2310.04363)), hybrid methods ([Zhao et al., ICML'24](https://arxiv.org/abs/2404.17546)), and distillation into tractable models ([Zhang et al., 2024](https://arxiv.org/abs/2406.13892)).\n- The inclusion of both LM and biological sequence design tasks is a good way to show the generality (although toy experiments on something synthetic and *very* simple, where the target is well-understood, would also be helpful).\n- Results generally comparable with prior work in each task with (possibly -- see questions below) higher efficiency of inference.", "weaknesses": "Overall, both the writing and the experiments need substantial improvement. It is quite difficult to understand the algorithm and the experiment setup, nor are the results particularly strong.\n\n- The writing in the first two pages is imprecise and does not set up the problem well.\n  - The first sentence of the abstract does not make sense to me. (Doesn't \"desired outputs\" already presuppose we are talking about generative modelling -- so what is the meaning of saying generative models are \"appealing\"? \"Sequence-level\" as opposed to what?) \n    - In general, the abstract does not explain well the problem setting and does not prime the reader to expect what is to come. The reader is left wondering: Are we talking about generative modelling from data or sampling given a density? Are we training a new model or constraining an existing one?\n  - The introduction does not set up the problem to be solved in a clear way. \n    - The first 2.5 paragraphs or so of the introduction seem to be saying approximately \"generalisation is important\". \n    - However, the **main objects** involved in the problem setting are not introduced. Thus it is not clear what is being talked about when \"original\" and \"sampled\" sequences are mentioned, what the EBM has to do with infilling, etc.\n    - Second paragraph: Lu et al. and Holtzman et al. are inaccurately described. The latter studies strategies to clip the tail of the next-token distributions, while Lu et al. studies MCTS as proposed by reference [59] in that paper, which seems to have no connection to such clipping.\n    - In the last paragraph of the intro, $q_\\theta$ appears, but it is never defined.\n- Section 2: Please define all symbols when they are first used. As far as I could infer:\n  - $\\cal X$: space of sequences\n  - $\\cal Y$: set of \"properties\"\n  - $o:{\\cal X}\\to{\\cal Y}$: oracle assigning a property to every sequence\n  - ${\\cal X}^{\\rm train}$: training set\n  - ${\\cal Y}^{\\rm train}$: unclear; is it $\\{o(x):x\\in{\\cal X}^{\\rm train}\\}$? That is, we wish to generate sequences whose properties, as computed by an oracle, were not observed in the training set?\n- Section 3:\n  - Is the score giving the density or the log-density? Contradiction between \"probability proportional to sequence-level score $s(x)$\" (L144) and the equation in L149.\n  - The following does not make sense to me (LL156-158): \"masked language modeling objectives implicitly define EBM with the masked-infilling objective serving as a proposal distribution\".\n    - How can a [training] **objective** define a distribution? Do you mean that the trained model defines a distribution? Same in LL162-163: the self-supervised pre-training process seems to have nothing to do with the use of the trained infilling model (so how can we say that MH sampling is based on the self-supervised training?).\n    - An EBM is defined by an energy function, not by a proposal distribution used to perform MH on it. Do you mean that the infilling model approximates a collapsed Gibbs sampler for a certain EBM? Or do you mean that the infilling model is an effective proposal for MH sampling of EBMs that happened to be trained on similar data?\n    - Even in the former case, things are somewhat more subtle than this. The conditional distributions given by infilling may not be compatible with any joint distribution, see, for example, [Henningen and Kim, ACL'23](https://arxiv.org/abs/2305.15501) and [Wang and Cho, NeuralGen@NAACL'19](https://arxiv.org/abs/1902.04094).\n    - In the end, I am left confused about the setting: the infilling model defines a EBM (L157) and is also a proposal distribution that is used to propose transitions (L160), but the proposals are accepted/rejected using the MH criterion on a EBM (L161), which seems to require an explicit energy model to compute the acceptance probability.\n  - A little diagram would help to understand the different ways of forming trajectories for training $q_\\theta$. (It is also confusing that $q$ (proposal) and $q_\\theta$ (amortised sampler) are denoted by the same letter.)\n- Experiments:\n  - Missing details:\n    - In each experiment, please specify exactly, in an equation, what the target energy model is. I found some notes on this in Appendix D, but they left me with more questions. In general, I would suggest to move some experiment details to the appendix, but state as directly as possible what sampling problem is being solved.\n    - It is not stated which version of training data for $q_\\theta$ is used. It can be inferred from the ablation study in Section 5, but would be good to make explicit. \n  - I am not convinced by evaluations in 4.1:\n    - The fluency is not reported for ICE.\n    - Extrapolation metric: why should we expect that \"make this negative\" would lead necessarily to 1-star and not to 2-star reviews? It would be more informative to compare the full distributions of the base model and with editing, for each editing method.\n      - How to understand that the second and third columns in Table 2 sum to more than 1, if they are the proportions of reviews with (in the example of negative sentiment) 2 and 1 stars, respectively?\n      - In fact, generating **too many** 1-star reviews, together with the high fluency, could actually indicate mode collapse. There is no diversity metric to show that this is not happening.\n  - The comment on diversity also applies to Section 4.2. Why was ICE not considered in this problem?\n  - How do the methods compare in execution wall time (both for training and drawing a single sample)?", "questions": "Please see above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730713847814}, {"id": "EvoB76pBJJ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11835/Reviewer_8bYp"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper proposes a method for enhancing sequence generation in models that need to extrapolate beyond training data that aims to extend beyond Padmakumar et al. (2023). It aims to be comparable with sampling-based methods such as MCMC in performance while being computationally efficient as existing works. To this end, it introduces an inference network trained on selected states from Markov chains. This approach is tested on protein sequence design, text sentiment control, and text anonymization tasks.", "review_text": "The paper proposes a method for enhancing sequence generation in models that need to extrapolate beyond training data that aims to extend beyond Padmakumar et al. (2023). It aims to be comparable with sampling-based methods such as MCMC in performance while being computationally efficient as existing works. To this end, it introduces an inference network trained on selected states from Markov chains. This approach is tested on protein sequence design, text sentiment control, and text anonymization tasks.", "strengths": "Overall, despite the fact that the paper is not well-written and easily understood, I admit that the paper conveys the problem background, the motivation, as well as their method relatively clearly. From my understanding, the problem of conditional sequence generation with extrapolated target properties is undoubtedly important, and I encourage the authors to conduct in-depth study of the problem. In addition, the goal of trying to find a method with (1) good performance, such as sampling-based methods, and (2) efficiency, such as existing iteration-based methods, is well-motivated and acute.", "weaknesses": "In my opinion, there are several major concerns that prevent current paper from being accepted. Specifically,\n\n- In the first place, instead of trying to improve on the existing ICE strategy, the authors should analyze whether and why current autoregressive (AR) models are unable to extrapolate, given their belief that iterative transformations are an alternative. Current LLMs have undoubtedly demonstrated their capacities to generalize beyond training data, especially in data-abundant areas. As a result, various conditional generation methods have proven effective, even when the conditions are highly user-crafted and unlikely seen in the training set. To well motivate the paper, the authors should explicitly analyze, (better) both theoretically and empirically, that current conditional generation methods such as prompts-based techniques or classifier guidance techniques, are insufficient for extrapolation. \n- Since the authors, unlike ICE, are training an additional model to directly \"predict\" minor transformations that improve a target property, it becomes unclear to me why the original AR models equipped with additional data cannot do this. I do not see the incentive to train an additional model (as a surrogate) for this. Specifically, if you consider each transformed \"sequence\" as a \"token\" (that actually is a sequence) of an \"AR\" model and train on this \"token sequence\", I believe the model is, by nature, designed to generate sequences with better target property iteratively. In this case, what is the reason to have two models in the first place?\n- Following the second point, I think the current comparison is unfair because the proposed method requires a large amount of preference data pairs to train. However, this data needs to be generated via standard MCMC-based methods and is unlikely to be generated computationally efficiently. As a result, the fact that the proposed method outperforms existing methods can be a natural fact that it is just leveraging more information, which again makes me confused about whether some alternatives can leverage these data more effectively. A more fair comparison should be made. For instance, when comparing efficiency, data generation & training time should be presented; when comparing performance, you should also show that existing methods with the dataset you generate are still inferior to your method.\n- Finally but minorly, I think the presence of tables is far from clear. For instance, in Table 1, $q_\\theta$ has lower $-1$ and $-2.5$ scores, and I am not sure whether this can be referred to as \"outperform\". There might be a misunderstanding, but the authors should make it clear what we are expected to read from each table.", "questions": "No additional questions beyond the weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method for enhancing sequence generation in models that need to extrapolate beyond training data that aims to extend beyond Padmakumar et al. (2023). It aims to be comparable with sampling-based methods such as MCMC in performance while being computationally efficient as existing works. To this end, it introduces an inference network trained on selected states from Markov chains. This approach is tested on protein sequence design, text sentiment control, and text anonymization tasks.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "Overall, despite the fact that the paper is not well-written and easily understood, I admit that the paper conveys the problem background, the motivation, as well as their method relatively clearly. From my understanding, the problem of conditional sequence generation with extrapolated target properties is undoubtedly important, and I encourage the authors to conduct in-depth study of the problem. In addition, the goal of trying to find a method with (1) good performance, such as sampling-based methods, and (2) efficiency, such as existing iteration-based methods, is well-motivated and acute.", "weaknesses": "In my opinion, there are several major concerns that prevent current paper from being accepted. Specifically,\n\n- In the first place, instead of trying to improve on the existing ICE strategy, the authors should analyze whether and why current autoregressive (AR) models are unable to extrapolate, given their belief that iterative transformations are an alternative. Current LLMs have undoubtedly demonstrated their capacities to generalize beyond training data, especially in data-abundant areas. As a result, various conditional generation methods have proven effective, even when the conditions are highly user-crafted and unlikely seen in the training set. To well motivate the paper, the authors should explicitly analyze, (better) both theoretically and empirically, that current conditional generation methods such as prompts-based techniques or classifier guidance techniques, are insufficient for extrapolation. \n- Since the authors, unlike ICE, are training an additional model to directly \"predict\" minor transformations that improve a target property, it becomes unclear to me why the original AR models equipped with additional data cannot do this. I do not see the incentive to train an additional model (as a surrogate) for this. Specifically, if you consider each transformed \"sequence\" as a \"token\" (that actually is a sequence) of an \"AR\" model and train on this \"token sequence\", I believe the model is, by nature, designed to generate sequences with better target property iteratively. In this case, what is the reason to have two models in the first place?\n- Following the second point, I think the current comparison is unfair because the proposed method requires a large amount of preference data pairs to train. However, this data needs to be generated via standard MCMC-based methods and is unlikely to be generated computationally efficiently. As a result, the fact that the proposed method outperforms existing methods can be a natural fact that it is just leveraging more information, which again makes me confused about whether some alternatives can leverage these data more effectively. A more fair comparison should be made. For instance, when comparing efficiency, data generation & training time should be presented; when comparing performance, you should also show that existing methods with the dataset you generate are still inferior to your method.\n- Finally but minorly, I think the presence of tables is far from clear. For instance, in Table 1, $q_\\theta$ has lower $-1$ and $-2.5$ scores, and I am not sure whether this can be referred to as \"outperform\". There might be a misunderstanding, but the authors should make it clear what we are expected to read from each table.", "questions": "No additional questions beyond the weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730543358055}, {"id": "81YrDHiwB2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission11835/Reviewer_jQ93"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper addresses the challenge of **efficient extrapolation in sequence modeling**. The problem setup is as follows: Given a scoring oracle (or a function approximating it) that evaluates sequences—such as protein stability or sentiment of review comments—and training data limited to a specific score range, $\\mathcal{I}$, the objective is to generate sequences that achieve scores outside this range. This extrapolation aims to transform sequences beyond the initial constraints imposed by the training dataset.\n\nTo me, the proposed solution can be framed as a reinforcement learning (RL) approach with the following key steps:\n\n- Sample sequences from an initial model (policy).\n- Randomly mask parts of each sequence.\n- Predict and fill masked sections using a masked language model (MLM).\n- Evaluate the transformed sequence via the scoring function.\n- Retain transformations that yield improved scores.\n\nAdditional heuristics are employed to enhance training efficiency, such as optimizing inference by selectively dropping intermediate reasoning steps when appropriate.", "review_text": "This paper addresses the challenge of **efficient extrapolation in sequence modeling**. The problem setup is as follows: Given a scoring oracle (or a function approximating it) that evaluates sequences—such as protein stability or sentiment of review comments—and training data limited to a specific score range, $\\mathcal{I}$, the objective is to generate sequences that achieve scores outside this range. This extrapolation aims to transform sequences beyond the initial constraints imposed by the training dataset.\n\nTo me, the proposed solution can be framed as a reinforcement learning (RL) approach with the following key steps:\n\n- Sample sequences from an initial model (policy).\n- Randomly mask parts of each sequence.\n- Predict and fill masked sections using a masked language model (MLM).\n- Evaluate the transformed sequence via the scoring function.\n- Retain transformations that yield improved scores.\n\nAdditional heuristics are employed to enhance training efficiency, such as optimizing inference by selectively dropping intermediate reasoning steps when appropriate.", "strengths": "- The experimental results on sentiment extrapolation and stable protein sequences indicate improvements on the proposed metrics, supporting the method’s effectiveness in generating high-scoring sequences beyond the training data’s score range while maintaining the chosen reference quality measures.\n\n- The work provides insights into the limitations of MCMC, and demonstrates how a structured RL approach can offer improvements in extrapolative tasks.", "weaknesses": "1. **Lack of context**:\n   Although some works on RL for controlled generation are discussed in the appendix, presenting the approach as an RL-based sequence generation technique in the main text would better contextualize the work. The core steps—sampling, masking, scoring, and retaining high-reward samples—closely resemble an RL framework and have been studied in the context of language modeling. For instance I can think of the following:\n\n   - The MLM and scoring function could be interpreted as critics, similar to those used in RL formulations for text generation [1,2]. I think both works have similar criteria of retaining only if the rollout is correct (0-1 rewards), which is similar to your formalization.\n   - The approach of appending scored sequences mirrors reward conditioning or goal-conditioned policy training in RL [3,4].\n\n   It would be easier to understand, compare, and extend this work if it is framed within an established framework. \n\n2. **Comparisons with RL Baselines**:\n   Given the alignment with RL methodologies, comparisons to baseline RL approaches would strengthen the empirical analysis. For example, on tasks like protein engineering, including a simple reward-conditioning baseline that takes ddG values as input or similar methods referenced in the appendix would provide a more complete evaluation. For instance, the protein sequence experiments are only compared with Padmakumar et al. (2023). Comparing with simple/basic RL methods would also work also as an ablation, justifying the more complex steps of the proposed method.\n\n3. **Clarity and Editorial Suggestions**:\n   Certain sections of the paper could benefit from additional clarity. For instance, the sentence on line 365 is incomplete, and some overly long sentences could be simplified for readability. Introducing mathematical notation earlier (e.g., at the end of page 2) would also improve consistency and clarify the problem setup throughout the paper.", "questions": "To improve the interpretability of the evaluation, it would be beneficial to include example (text) sequences alongside quantitative metrics, such as fluency or Sentence-BERT similarity scores. While these metrics offer quantitative insights, they are difficult to interpret without sample outputs to contextualize them. For instance, it is vert vague for me what the impact of a 0.13% improvement in fluency or a BERT similarity of 0.1 would be. It would be clearer if  these metrics are accompanied by representative examples, e.g. showcasing semantic similarity, etc. I wonder how the actual output texts compare to baselines.\n\n**References**\n\n1. Zelikman et al., *STaR: Bootstrapping Reasoning With Reasoning*\n2. Zelikman et al., *Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking*\n3. Lynch and Semanet, *Language Conditioned Imitation Learning over Unstructured Data*\n4. Shypula, Madaan, et al., *Learning Performance-Improving Code Edits*", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the challenge of **efficient extrapolation in sequence modeling**. The problem setup is as follows: Given a scoring oracle (or a function approximating it) that evaluates sequences—such as protein stability or sentiment of review comments—and training data limited to a specific score range, $\\mathcal{I}$, the objective is to generate sequences that achieve scores outside this range. This extrapolation aims to transform sequences beyond the initial constraints imposed by the training dataset.\n\nTo me, the proposed solution can be framed as a reinforcement learning (RL) approach with the following key steps:\n\n- Sample sequences from an initial model (policy).\n- Randomly mask parts of each sequence.\n- Predict and fill masked sections using a masked language model (MLM).\n- Evaluate the transformed sequence via the scoring function.\n- Retain transformations that yield improved scores.\n\nAdditional heuristics are employed to enhance training efficiency, such as optimizing inference by selectively dropping intermediate reasoning steps when appropriate.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The experimental results on sentiment extrapolation and stable protein sequences indicate improvements on the proposed metrics, supporting the method’s effectiveness in generating high-scoring sequences beyond the training data’s score range while maintaining the chosen reference quality measures.\n\n- The work provides insights into the limitations of MCMC, and demonstrates how a structured RL approach can offer improvements in extrapolative tasks.", "weaknesses": "1. **Lack of context**:\n   Although some works on RL for controlled generation are discussed in the appendix, presenting the approach as an RL-based sequence generation technique in the main text would better contextualize the work. The core steps—sampling, masking, scoring, and retaining high-reward samples—closely resemble an RL framework and have been studied in the context of language modeling. For instance I can think of the following:\n\n   - The MLM and scoring function could be interpreted as critics, similar to those used in RL formulations for text generation [1,2]. I think both works have similar criteria of retaining only if the rollout is correct (0-1 rewards), which is similar to your formalization.\n   - The approach of appending scored sequences mirrors reward conditioning or goal-conditioned policy training in RL [3,4].\n\n   It would be easier to understand, compare, and extend this work if it is framed within an established framework. \n\n2. **Comparisons with RL Baselines**:\n   Given the alignment with RL methodologies, comparisons to baseline RL approaches would strengthen the empirical analysis. For example, on tasks like protein engineering, including a simple reward-conditioning baseline that takes ddG values as input or similar methods referenced in the appendix would provide a more complete evaluation. For instance, the protein sequence experiments are only compared with Padmakumar et al. (2023). Comparing with simple/basic RL methods would also work also as an ablation, justifying the more complex steps of the proposed method.\n\n3. **Clarity and Editorial Suggestions**:\n   Certain sections of the paper could benefit from additional clarity. For instance, the sentence on line 365 is incomplete, and some overly long sentences could be simplified for readability. Introducing mathematical notation earlier (e.g., at the end of page 2) would also improve consistency and clarify the problem setup throughout the paper.", "questions": "To improve the interpretability of the evaluation, it would be beneficial to include example (text) sequences alongside quantitative metrics, such as fluency or Sentence-BERT similarity scores. While these metrics offer quantitative insights, they are difficult to interpret without sample outputs to contextualize them. For instance, it is vert vague for me what the impact of a 0.13% improvement in fluency or a BERT similarity of 0.1 would be. It would be clearer if  these metrics are accompanied by representative examples, e.g. showcasing semantic similarity, etc. I wonder how the actual output texts compare to baselines.\n\n**References**\n\n1. Zelikman et al., *STaR: Bootstrapping Reasoning With Reasoning*\n2. Zelikman et al., *Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking*\n3. Lynch and Semanet, *Language Conditioned Imitation Learning over Unstructured Data*\n4. Shypula, Madaan, et al., *Learning Performance-Improving Code Edits*", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730390111748}], "openreview_url": "https://openreview.net/forum?id=DQfHkEcUqV", "arxiv_id": "2505.20251", "paper_pdf": "papers/DQfHkEcUqV.pdf", "paper_pdf_sha256": "35d14ebf9a3beec549c00b78a9ac00f3651b1b269837fb3bc091747148d218ba", "paper_pdf_bytes": 354935, "paper_pdf_source": "openreview", "code_url": "https://github.com/sophia-hager/learning-MCMC-extrapolation", "code_repository": "sophia-hager/learning-MCMC-extrapolation", "code_commit": "1692bbd894c0a9b226888006f2b1e67e95c10229", "code_archive": "repos/DQfHkEcUqV.zip", "code_archive_sha256": "9177f4d2078fda4565068800ead9be9264b524d1fbc3e1d899c13ffe45f20216", "code_archive_bytes": 416653, "code_file_count": 43, "code_extensions": {".py": 35, ".sh": 8}, "github_disk_usage_kb": 388, "github_languages": {"Python": 221029, "Shell": 2777}, "github_archived": false, "github_pushed_at": "2025-08-07T23:26:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-extrapolative-sequence"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RxhOEngX8s", "year": 2024, "status": "rejected", "title": "Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection", "authors": ["Charles Guille-Escuret", "Pierre-Andre Noel", "Ioannis Mitliagkas", "David Vazquez", "Joao Monteiro"], "authorids": ["~Charles_Guille-Escuret1", "~Pierre-Andre_Noel1", "~Ioannis_Mitliagkas1", "~David_Vazquez1", "~Joao_Monteiro1"], "authors_source": "OpenReview API", "abstract": "Deployed machine learning systems can be improved using methods detecting out-of-distribution (OOD) inputs. Existing research mainly focuses on one type of distribution shift: detecting samples from novel classes, absent from the training set. However, real-world systems encounter a broad variety of anomalous inputs, and the OOD literature neglects this diversity. This work categorizes five distinct types of distribution shifts and critically evaluates the performance of recent OOD detection methods on each of them. We publicly release our benchmark under the name BROAD (Benchmarking Resilience Over Anomaly Diversity). We find that while these methods excel in detecting novel classes, their performances are inconsistent across other types of distribution shifts. In other words, they can only reliably detect unexpected inputs that they have been specifically designed to expect. As a first step toward broad OOD detection, we learn a Gaussian mixture generative model for existing detection scores, enabling an ensemble detection approach that is more consistent and comprehensive for broad OOD detection, with improved performances over existing methods. Our code to download BROAD and reproduce our experiments will be released upon publication.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "BWD3y6yps6", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6028/Reviewer_WvA2"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper reviews Out of Distribution (OOD) Detection, in the sense of samples seen in the real world that were not covered in the training data. Two approaches are \n1) to create robust systems, designed to not degrade on OOD data, and \n2) to flag samples uncharacteristic of the training data.\nDistribution shift detection appears more practical, but can be fooled, by the multiple ways OOD can occur for such diverse reasons (in images): as novel classes, adversarial attacks, synthetics, corruptions multiple labels. \n\nThe paper introduces \n - a new OOD benchmark with 12 datasets representing the various OOD reasons\n - benchmarking of a variety of existing methods published in the last decade,\n - And demonstration of a Gaussian Mixture Model of an ensemble of existing methods with significant gains over existing methods.", "review_text": "This paper reviews Out of Distribution (OOD) Detection, in the sense of samples seen in the real world that were not covered in the training data. Two approaches are \n1) to create robust systems, designed to not degrade on OOD data, and \n2) to flag samples uncharacteristic of the training data.\nDistribution shift detection appears more practical, but can be fooled, by the multiple ways OOD can occur for such diverse reasons (in images): as novel classes, adversarial attacks, synthetics, corruptions multiple labels. \n\nThe paper introduces \n - a new OOD benchmark with 12 datasets representing the various OOD reasons\n - benchmarking of a variety of existing methods published in the last decade,\n - And demonstration of a Gaussian Mixture Model of an ensemble of existing methods with significant gains over existing methods.", "strengths": "The paper offers a comprehensive view of why image recognition models may fail on images that have not been seen during training, and by a comprehensive set of tests demonstrates the relative value of existing detection methods when applied to tasks they were intended for, and for other OOD tasks that are related, but not explicitly targeted by the existing method.   Of note is the interesting comment on how to build OOD detectors with generative models, by use of a function they designate as \"h(x)\" that so to speak \"sees inside\" the generative network, offering an extended feature set for detection.  \n\nThe paper introduces a OOD classifier that is a combination of existing methods by use of a Gaussian mixture model that has better coverage, and achieves an AUC score on average superior to existing methods.", "weaknesses": "There isn't sufficient detail in the paper to re-construct the Gaussian mixture model (GMMs) proposed by the authors. GMMs are conventionally used to estimate density functions for oddly-shaped distributions, e.g. with multiple modes.  It is intuitive, in fact not unexpected, that creating an ensemble of detectors has better  performance on average than any individual detector, so the novelty of this finding is limited. however the results from the paper are not reproducible from the paper's contents. Given the scores how is the GMM density learnt? In what sense is this an ensemble? How does this generate a classification and thus an AUC? \n\nOne gets the sense that this work would find a better audience in a more engineering-oriented conference where testing comparisons of performance were the primary interest, and algorithmic aspects were not.", "questions": "If the construction and output of the GMM classifier is actually revealed in the paper and this review overlooked it, please explain.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper reviews Out of Distribution (OOD) Detection, in the sense of samples seen in the real world that were not covered in the training data. Two approaches are \n1) to create robust systems, designed to not degrade on OOD data, and \n2) to flag samples uncharacteristic of the training data.\nDistribution shift detection appears more practical, but can be fooled, by the multiple ways OOD can occur for such diverse reasons (in images): as novel classes, adversarial attacks, synthetics, corruptions multiple labels. \n\nThe paper introduces \n - a new OOD benchmark with 12 datasets representing the various OOD reasons\n - benchmarking of a variety of existing methods published in the last decade,\n - And demonstration of a Gaussian Mixture Model of an ensemble of existing methods with significant gains over existing methods.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "The paper offers a comprehensive view of why image recognition models may fail on images that have not been seen during training, and by a comprehensive set of tests demonstrates the relative value of existing detection methods when applied to tasks they were intended for, and for other OOD tasks that are related, but not explicitly targeted by the existing method.   Of note is the interesting comment on how to build OOD detectors with generative models, by use of a function they designate as \"h(x)\" that so to speak \"sees inside\" the generative network, offering an extended feature set for detection.  \n\nThe paper introduces a OOD classifier that is a combination of existing methods by use of a Gaussian mixture model that has better coverage, and achieves an AUC score on average superior to existing methods.", "weaknesses": "There isn't sufficient detail in the paper to re-construct the Gaussian mixture model (GMMs) proposed by the authors. GMMs are conventionally used to estimate density functions for oddly-shaped distributions, e.g. with multiple modes.  It is intuitive, in fact not unexpected, that creating an ensemble of detectors has better  performance on average than any individual detector, so the novelty of this finding is limited. however the results from the paper are not reproducible from the paper's contents. Given the scores how is the GMM density learnt? In what sense is this an ensemble? How does this generate a classification and thus an AUC? \n\nOne gets the sense that this work would find a better audience in a more engineering-oriented conference where testing comparisons of performance were the primary interest, and algorithmic aspects were not.", "questions": "If the construction and output of the GMM classifier is actually revealed in the paper and this review overlooked it, please explain.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699033119670}, {"id": "RcyPD9hAzB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6028/Reviewer_xuax"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper propose a new visual OOD detection benchmark consisting of 5 types of distributional shifts (1) novel classes (2) adversarial perturbations (3) synthetic images (4) corruptions (5) images with multiple objects. The paper further evaluates the performance of various OOD detection methods (that do not require training/fine-tuning) and observe that the performance is inconsistent across different types of distribution shifts. Lastly, the authors propose to ensemble various scores with Gaussian mixture models, which demonstrates better performance.", "review_text": "The paper propose a new visual OOD detection benchmark consisting of 5 types of distributional shifts (1) novel classes (2) adversarial perturbations (3) synthetic images (4) corruptions (5) images with multiple objects. The paper further evaluates the performance of various OOD detection methods (that do not require training/fine-tuning) and observe that the performance is inconsistent across different types of distribution shifts. Lastly, the authors propose to ensemble various scores with Gaussian mixture models, which demonstrates better performance.", "strengths": "- The overall organization of the paper is clear and easy to follow.\n- The proposed ensembling method is straightforward and demonstrate good performance.", "weaknesses": "- The major weakness of the paper is that most OOD detection scores considered in the paper are proposed to only handle novel classes. Expecting such OOD scores to detect adversarial perturbations and corruptions may be **out-of-scope** and unrealistic.\n\nIn particular, recent work [1] has demonstrated that when OOD samples are not involved during training (the setting considered in this work), it can be **theoretically impossible** to expect common OOD detection methods to work. Despite that detecting adversarial perturbations and corruptions are interesting tasks, directly utilizing post-hoc OOD detection scores is ill-justified and the failure is expected.\n\n- In the multi-label scenario, it seems more reasonable to use object detection models instead of classification models. The failure of OOD detection based on classification models is expected. It would be more interesting to see the performance of OOD detection given bounding boxes.", "questions": "Method:\n- Can authors justify in theory or principle why OOD detection methods are suitable for detecting adversarial perturbations and corruptions?\n\nExperiments: \n- Can authors provide further OOD detection results with object detection models for the multi-label case?\n\n[1] Fang et al., Is Out-of-Distribution Detection Learnable?, NeurIPS 2022", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper propose a new visual OOD detection benchmark consisting of 5 types of distributional shifts (1) novel classes (2) adversarial perturbations (3) synthetic images (4) corruptions (5) images with multiple objects. The paper further evaluates the performance of various OOD detection methods (that do not require training/fine-tuning) and observe that the performance is inconsistent across different types of distribution shifts. Lastly, the authors propose to ensemble various scores with Gaussian mixture models, which demonstrates better performance.", "soundness": "1 poor", "presentation": "3 good", "contribution": "1 poor", "strengths": "- The overall organization of the paper is clear and easy to follow.\n- The proposed ensembling method is straightforward and demonstrate good performance.", "weaknesses": "- The major weakness of the paper is that most OOD detection scores considered in the paper are proposed to only handle novel classes. Expecting such OOD scores to detect adversarial perturbations and corruptions may be **out-of-scope** and unrealistic.\n\nIn particular, recent work [1] has demonstrated that when OOD samples are not involved during training (the setting considered in this work), it can be **theoretically impossible** to expect common OOD detection methods to work. Despite that detecting adversarial perturbations and corruptions are interesting tasks, directly utilizing post-hoc OOD detection scores is ill-justified and the failure is expected.\n\n- In the multi-label scenario, it seems more reasonable to use object detection models instead of classification models. The failure of OOD detection based on classification models is expected. It would be more interesting to see the performance of OOD detection given bounding boxes.", "questions": "Method:\n- Can authors justify in theory or principle why OOD detection methods are suitable for detecting adversarial perturbations and corruptions?\n\nExperiments: \n- Can authors provide further OOD detection results with object detection models for the multi-label case?\n\n[1] Fang et al., Is Out-of-Distribution Detection Learnable?, NeurIPS 2022", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698813274474}, {"id": "vyIj3bcxfG", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6028/Reviewer_GVGN"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper propose a OOD benchmark comprising five different types of distribution shift. The result shows that the performance of OOD detection methods are not consistent over different types of distribution shift. The paper propose a method to ensemble different OOD detection methods to achieve consistent performence over different types of distribution shift.", "review_text": "This paper propose a OOD benchmark comprising five different types of distribution shift. The result shows that the performance of OOD detection methods are not consistent over different types of distribution shift. The paper propose a method to ensemble different OOD detection methods to achieve consistent performence over different types of distribution shift.", "strengths": "1). The assessment of a broder OOD detection capabilities is interesting and probablity important for future OOD detection method development.\n\n2). Extensive experiments have been done to benchmark the recent OOD detection methods.\n\n3). Overall, the paper is clear and well-written.", "weaknesses": "My concerns are mainly about the proposed method.\n\n1). The ensemble of OOD detection methods seems ad-hoc for this benchmark by evaluating and picking some of the methods that perform relatively well on the benchmark.\n\n2). The proposed method of fitting GMM over scores from different OOD detection methods does not make sense to me. For example, in Sec.3, it says \"this approach is adept at identifying atypical realizations of the underlying scores, even in situations where the marginal likelihood of each score is high, but their joint likelihood is low.\" It would be weird if most in-distribution samples can not achieve high likelihood at each score while out-of-distribution samples can, since all these methods aim at measuring if the sample is in-distribution. In other words, what is the advantage of fitting GMM over taking average of different scores?\n\n3). The time complexity or the scalability of the proposed method is still of concern. Though the results shows that the time complexity of the proposed ENS-F is acceptable at 25% of the time for a normal inference. However, when comparing to the methods it ensembles (e.g. MSP takes 1% additional time), the time complexity is extremely high.", "questions": "1). The proposed method uses a validation set to fit GMM, does it affect the performance? With the use of additional data, is it fair to compare the proposed method with other baseline methods?\n\n2). Does the proposed ensemble method outperform simple ensemble methods such as taking the average of the scores or the largest score? If so, why would it be?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper propose a OOD benchmark comprising five different types of distribution shift. The result shows that the performance of OOD detection methods are not consistent over different types of distribution shift. The paper propose a method to ensemble different OOD detection methods to achieve consistent performence over different types of distribution shift.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "1). The assessment of a broder OOD detection capabilities is interesting and probablity important for future OOD detection method development.\n\n2). Extensive experiments have been done to benchmark the recent OOD detection methods.\n\n3). Overall, the paper is clear and well-written.", "weaknesses": "My concerns are mainly about the proposed method.\n\n1). The ensemble of OOD detection methods seems ad-hoc for this benchmark by evaluating and picking some of the methods that perform relatively well on the benchmark.\n\n2). The proposed method of fitting GMM over scores from different OOD detection methods does not make sense to me. For example, in Sec.3, it says \"this approach is adept at identifying atypical realizations of the underlying scores, even in situations where the marginal likelihood of each score is high, but their joint likelihood is low.\" It would be weird if most in-distribution samples can not achieve high likelihood at each score while out-of-distribution samples can, since all these methods aim at measuring if the sample is in-distribution. In other words, what is the advantage of fitting GMM over taking average of different scores?\n\n3). The time complexity or the scalability of the proposed method is still of concern. Though the results shows that the time complexity of the proposed ENS-F is acceptable at 25% of the time for a normal inference. However, when comparing to the methods it ensembles (e.g. MSP takes 1% additional time), the time complexity is extremely high.", "questions": "1). The proposed method uses a validation set to fit GMM, does it affect the performance? With the use of additional data, is it fair to compare the proposed method with other baseline methods?\n\n2). Does the proposed ensemble method outperform simple ensemble methods such as taking the average of the scores or the largest score? If so, why would it be?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698571150791}, {"id": "XxnYlCmWlx", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6028/Reviewer_Jh7c"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses the problem of OOD detection in deployed machine learning systems. The authors analyze existing OOD detectors, identifying a common limitation in their adaptability to diverse distribution shifts. To address this, they propose a new benchmark named BROAD, designed to evaluate OOD detectors across a wide spectrum of distribution shifts. This benchmark examines various scenarios, including novel classes, adversarial perturbations, synthetic images, corruptions, and multi-class inputs. Additionally, the authors introduce an approach that leverages an ensemble of reliable OOD detectors combined with a GMM.", "review_text": "The paper addresses the problem of OOD detection in deployed machine learning systems. The authors analyze existing OOD detectors, identifying a common limitation in their adaptability to diverse distribution shifts. To address this, they propose a new benchmark named BROAD, designed to evaluate OOD detectors across a wide spectrum of distribution shifts. This benchmark examines various scenarios, including novel classes, adversarial perturbations, synthetic images, corruptions, and multi-class inputs. Additionally, the authors introduce an approach that leverages an ensemble of reliable OOD detectors combined with a GMM.", "strengths": "The paper explains the existing limitations of current OOD detectors, providing a clear and compelling rationale for advancement in this domain. The paper's main strength lies in the introduction of the BROAD benchmark, which establishes a robust evaluation framework. This benchmark not only rigorously assesses OOD detector performance but also offers a critical insight into their ability to extend beyond the scope of novel classes. Through meticulous experimentation and analysis, the authors provide compelling evidence of OOD detector capabilities and potential areas of improvement. This well-constructed evaluation framework significantly contributes to the depth and reliability of the research findings.", "weaknesses": "The paper introduces an ensemble method for broad OOD detection; however, there are notable weaknesses. Its efficiency and inability to be scaled up raise questions about its practicality in real-world applications. Moreover, the paper lacks a clear roadmap or forward-looking guidance for the broader OOD detection community on how to effectively approach the challenges posed by the BROAD benchmark. While the intent to introduce a method is appreciated, the proposed method is clearly not the ideal solution to this intricate problem, as it does not offer new ideas or new directions to the OOD community. A more comprehensive analysis and discussion on potential future directions and steps for advancing OOD detection would greatly enhance the paper's overall impact and utility to the research community. This would provide valuable insights for researchers looking to build upon this work and make meaningful strides in the field of OOD detection.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the problem of OOD detection in deployed machine learning systems. The authors analyze existing OOD detectors, identifying a common limitation in their adaptability to diverse distribution shifts. To address this, they propose a new benchmark named BROAD, designed to evaluate OOD detectors across a wide spectrum of distribution shifts. This benchmark examines various scenarios, including novel classes, adversarial perturbations, synthetic images, corruptions, and multi-class inputs. Additionally, the authors introduce an approach that leverages an ensemble of reliable OOD detectors combined with a GMM.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper explains the existing limitations of current OOD detectors, providing a clear and compelling rationale for advancement in this domain. The paper's main strength lies in the introduction of the BROAD benchmark, which establishes a robust evaluation framework. This benchmark not only rigorously assesses OOD detector performance but also offers a critical insight into their ability to extend beyond the scope of novel classes. Through meticulous experimentation and analysis, the authors provide compelling evidence of OOD detector capabilities and potential areas of improvement. This well-constructed evaluation framework significantly contributes to the depth and reliability of the research findings.", "weaknesses": "The paper introduces an ensemble method for broad OOD detection; however, there are notable weaknesses. Its efficiency and inability to be scaled up raise questions about its practicality in real-world applications. Moreover, the paper lacks a clear roadmap or forward-looking guidance for the broader OOD detection community on how to effectively approach the challenges posed by the BROAD benchmark. While the intent to introduce a method is appreciated, the proposed method is clearly not the ideal solution to this intricate problem, as it does not offer new ideas or new directions to the OOD community. A more comprehensive analysis and discussion on potential future directions and steps for advancing OOD detection would greatly enhance the paper's overall impact and utility to the research community. This would provide valuable insights for researchers looking to build upon this work and make meaningful strides in the field of OOD detection.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697538871890}], "openreview_url": "https://openreview.net/forum?id=RxhOEngX8s", "arxiv_id": "2308.11480", "paper_pdf": "papers/RxhOEngX8s.pdf", "paper_pdf_sha256": "0d1b319ec90c1738e8473cef1b3bf4c7f17f021c3203106db75ea2b769c19338", "paper_pdf_bytes": 2040858, "paper_pdf_source": "openreview", "code_url": "https://github.com/ServiceNow/broad-openood", "code_repository": "ServiceNow/broad-openood", "code_commit": "7a9414801544c1a4b759e453873442be9600fbe1", "code_archive": "repos/RxhOEngX8s.zip", "code_archive_sha256": "fabf981968cbe5eac8c6fabe51b49999f2959f63732864d5ac83962bd7c00857", "code_archive_bytes": 561816, "code_file_count": 427, "code_extensions": {".sh": 238, ".py": 186, ".ipynb": 3}, "github_disk_usage_kb": 372, "github_languages": {"Python": 647435, "Jupyter Notebook": 276787, "Shell": 158806}, "github_archived": false, "github_pushed_at": "2023-08-22T11:04:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/expecting-the-unexpected-towards-broad-out-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "luajgSjRlew", "year": 2023, "status": "rejected", "title": "Social and environmental impact of recent developments in machine learning on biology and chemistry research", "authors": ["Daniel Probst"], "authorids": ["~Daniel_Probst2"], "authors_source": "OpenReview API", "abstract": "Potential societal and environmental effects such as the rapidly increasing resource use and the associated environmental impact, reproducibility issues, and exclusivity, the privatization of ML research leading to a public research brain-drain, a narrowing of the research effort caused by a focus on deep learning, and the introduction of biases through a lack of sociodemographic diversity in data and personnel caused by recent developments in machine learning are a current topic of discussion and scientific publications. However, these discussions and publications focus mainly on computer science-adjacent fields, including computer vision and natural language processing or basic ML research. Using bibliometric analysis of the complete and full-text analysis of the open-access literature, we show that the same observations can be made for applied machine learning in chemistry and biology. These developments can potentially affect basic and applied research, such as drug discovery and development, beyond the known issue of biased data sets.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tXuo3ALmsDI", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1682/Reviewer_dCqg"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nThe authors go over trends in the research and use of machine learning\nand in particular deep learning in the applied sciences, via analysis\nof the open-access literature (bibliometrics and/or full-text analyses),\nand highlight areas for concern. The findings are somewhat similar to\nprior results of the analyses of recent trends in machine learning R&D \nand its related fields.\n", "review_text": "The paper reports on trends in ml and how it has been developing in\nthat past decade, specially in biology and chemistry.  Several of\nthese trends are certainly concerning (the de-democratizatin of ml\nwork).\n\nHowever, I don't think that ICLR is the venue for this publication,\njust as literature reviews are highly useful, but not relevant in this\nconference. The paper does not have a technical contribution.\n", "strengths": "Strengths:\n\n-- A variety of useful insights and takeaways into how  ml (in particular large-scale deep ml) is impacting other\n   scientific fields, in particular biology and chemistry, and several\n   trends are estimated from analysis of published research  (societal/environmental/energy impact).  \n\nWeaknesses:\n\n-- No or very little technical contributions.  The ICLR venue may not to be\n   relevant for this report.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "\nThe authors go over trends in the research and use of machine learning\nand in particular deep learning in the applied sciences, via analysis\nof the open-access literature (bibliometrics and/or full-text analyses),\nand highlight areas for concern. The findings are somewhat similar to\nprior results of the analyses of recent trends in machine learning R&D \nand its related fields.\n", "strength_and_weaknesses": "Strengths:\n\n-- A variety of useful insights and takeaways into how  ml (in particular large-scale deep ml) is impacting other\n   scientific fields, in particular biology and chemistry, and several\n   trends are estimated from analysis of published research  (societal/environmental/energy impact).  \n\nWeaknesses:\n\n-- No or very little technical contributions.  The ICLR venue may not to be\n   relevant for this report.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity is good, although the introduction abruptly dives into the\nhistory of hardware development.  The paper uses existing techniques\nbut the novelty mainly lies in applying the analyses to research in \nbiology and chemistry.\n\n\n", "summary_of_the_review": "The paper reports on trends in ml and how it has been developing in\nthat past decade, specially in biology and chemistry.  Several of\nthese trends are certainly concerning (the de-democratizatin of ml\nwork).\n\nHowever, I don't think that ICLR is the venue for this publication,\njust as literature reviews are highly useful, but not relevant in this\nconference. The paper does not have a technical contribution.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666674695971}, {"id": "r-GWKER-L2", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1682/Reviewer_6kvg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper draws attention to the societal and environmental impacts of machine learning research in chemistry and biology. The paper specifically makes a contribution in the following components such as: environmental considerations, citation inequality, academic brain-drain, and scientific considerations:\n\nThe study makes the following key assertions:\n\n* During the past two decades, there has been an  unparalleled shift of professors from university to industry. \n* chemistry saw an increase in transitions of machine learning practitioners from academia to industry, while overall transitions in biology remain slightly skewed towards industry \n* Technology companies drive machine learning research. \n* More citations in chemistry and biology are generated from Pharm, tech, and “other” companies compared to academia. \n* In terms of environmental impact, the data was limited and as a result the  quantitative analysis of the environmental impact of  machine learning research in biology and chemistry was limited. ", "review_text": "\nThe paper presents the critical interfaces of social and environmental impact of machine learning research specifically in chemistry and biology.  The findings reported in the paper are applicable to other fields as well.  This work provided a basis for discussion as key fields such as drug development, genetics, and cancer research can be affected as well. Overall, this overview provided essential knowledge for the advancement of machine learning research in biology and chemistry. \n\n\n", "strengths": "Strength: results reported were consistently compared with those in literature. \n\nWeakness 1: The sentences are just too long. Need to be broken down. For example, this sentence is too long : “Potential societal and environmental effects such as the rapidly increasing resource use and the associated environmental impact, reproducibility issues, and exclusivity, the privatization of ML research leading to a public research brain-drain, a narrowing of the research effort caused by a focus on deep learning, and the introduction of biases through a lack of sociodemographic diversity in data and personnel caused by recent developments in machine learning are a current topic of discussion and scientific publications.\n\nWeakness 2: A simple figure highlighting all the steps/methods that were followed in generating the paper findings is missing.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper draws attention to the societal and environmental impacts of machine learning research in chemistry and biology. The paper specifically makes a contribution in the following components such as: environmental considerations, citation inequality, academic brain-drain, and scientific considerations:\n\nThe study makes the following key assertions:\n\n* During the past two decades, there has been an  unparalleled shift of professors from university to industry. \n* chemistry saw an increase in transitions of machine learning practitioners from academia to industry, while overall transitions in biology remain slightly skewed towards industry \n* Technology companies drive machine learning research. \n* More citations in chemistry and biology are generated from Pharm, tech, and “other” companies compared to academia. \n* In terms of environmental impact, the data was limited and as a result the  quantitative analysis of the environmental impact of  machine learning research in biology and chemistry was limited. ", "strength_and_weaknesses": "Strength: results reported were consistently compared with those in literature. \n\nWeakness 1: The sentences are just too long. Need to be broken down. For example, this sentence is too long : “Potential societal and environmental effects such as the rapidly increasing resource use and the associated environmental impact, reproducibility issues, and exclusivity, the privatization of ML research leading to a public research brain-drain, a narrowing of the research effort caused by a focus on deep learning, and the introduction of biases through a lack of sociodemographic diversity in data and personnel caused by recent developments in machine learning are a current topic of discussion and scientific publications.\n\nWeakness 2: A simple figure highlighting all the steps/methods that were followed in generating the paper findings is missing.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper presents novel knowledge as it specifically draws attention to the impact of machine learning research in biology and chemistry. The findings are also general in  that they apply to other fields as well. \nThe results reported in the study may be reproduced as the authors have shared  data  and code used in this study. ", "summary_of_the_review": "\nThe paper presents the critical interfaces of social and environmental impact of machine learning research specifically in chemistry and biology.  The findings reported in the paper are applicable to other fields as well.  This work provided a basis for discussion as key fields such as drug development, genetics, and cancer research can be affected as well. Overall, this overview provided essential knowledge for the advancement of machine learning research in biology and chemistry. \n\n\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "None.", "recommendation": "8: accept, good paper"}, "tcdate": 1666601589379}, {"id": "r8Pqrh_nPBB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1682/Reviewer_RDG6"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research.", "review_text": "Interesting paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research.\n\nIt falls within the remit of the conference, as the call for papers includes work on \"societal considerations of representation learning \nincluding fairness, safety, privacy, and interpretability, and explainability\"\n\nEven if interesting and certainly relevant, my main qualm with this study is that it lacks coherent structure. A lot of ideas and information\nare thrown in the pot, but they are barely organized at all, jumping from a topic to the next, almost without the respite of a paragraph ending.\nThis is compounded by a rather unirthodox referencing system.\n\nI definitely like the content, but it seems to me that it would require a different venue in which these ideas could breath, be it a jounal position paper or a book chapter.\n", "strengths": "Strengths:\nLots of information and very interesting topic\nWeaknesses:\nMinimally coherent structure. At some points it almost seems like a stream of thoughts dump.\nMixed topics\nUnorthodox information source referencing", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "Paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research.", "strength_and_weaknesses": "Strengths:\nLots of information and very interesting topic\nWeaknesses:\nMinimally coherent structure. At some points it almost seems like a stream of thoughts dump.\nMixed topics\nUnorthodox information source referencing", "clarity,_quality,_novelty_and_reproducibility": "Not very clear for the reasons stated above.\nDefinitely novel from the point of vew of the research and certainly reproducible.", "summary_of_the_review": "Interesting paper studying social and environmental impact of recent developments in ML (with a focus on DL) on biology and chemistry research.\n\nIt falls within the remit of the conference, as the call for papers includes work on \"societal considerations of representation learning \nincluding fairness, safety, privacy, and interpretability, and explainability\"\n\nEven if interesting and certainly relevant, my main qualm with this study is that it lacks coherent structure. A lot of ideas and information\nare thrown in the pot, but they are barely organized at all, jumping from a topic to the next, almost without the respite of a paragraph ending.\nThis is compounded by a rather unirthodox referencing system.\n\nI definitely like the content, but it seems to me that it would require a different venue in which these ideas could breath, be it a jounal position paper or a book chapter.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666370936066}, {"id": "-WqqhTE8Er", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1682/Reviewer_s4UR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "\nIn this paper, the authors perform several quantitive studies for understanding the social and environmental impact when applying machine learning methods for biology and chemistry research. Some of authors observations include a potential inequity increase in applied machine learning research in biology and chemistry, and big companies tend to be actively involved in machine learning based research.\n\n", "review_text": "\nThis paper is not proposing some new tasks or introducing some new methods. Although the topic and the authors' findings are interesting, it seems this paper is not suitable for a conference like ICLR.", "strengths": "\nThe topic is of importance and in general interesting.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "\nIn this paper, the authors perform several quantitive studies for understanding the social and environmental impact when applying machine learning methods for biology and chemistry research. Some of authors observations include a potential inequity increase in applied machine learning research in biology and chemistry, and big companies tend to be actively involved in machine learning based research.\n\n", "strength_and_weaknesses": "\nThe topic is of importance and in general interesting.", "clarity,_quality,_novelty_and_reproducibility": "\nAll experiments seem to be reproducible.", "summary_of_the_review": "\nThis paper is not proposing some new tasks or introducing some new methods. Although the topic and the authors' findings are interesting, it seems this paper is not suitable for a conference like ICLR.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666126746798}], "openreview_url": "https://openreview.net/forum?id=luajgSjRlew", "arxiv_id": "2210.00356", "paper_pdf": "papers/luajgSjRlew.pdf", "paper_pdf_sha256": "dcc1ee623895a8498778699700f9e1b72419c3090dd29c22687b51dc1038b4bd", "paper_pdf_bytes": 423657, "paper_pdf_source": "openreview", "code_url": "https://github.com/daenuprobst/anon_aichem", "code_repository": "daenuprobst/anon_aichem", "code_commit": "1fed4455d480029851d7311b30808ddefc50654d", "code_archive": "repos/luajgSjRlew.zip", "code_archive_sha256": "583381e7daaa78a14c6dbf0d3f50a698cfc14ee750e40908f7dc3520a62080d6", "code_archive_bytes": 989642, "code_file_count": 12, "code_extensions": {".py": 8, ".ipynb": 4}, "github_disk_usage_kb": 977, "github_languages": {"Jupyter Notebook": 1948436, "Python": 42528}, "github_archived": false, "github_pushed_at": "2022-09-29T11:43:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/social-and-environmental-impact-of-recent"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UTTrevGchy", "year": 2022, "status": "rejected", "title": "Learning Diverse Options via InfoMax Termination Critic", "authors": ["Yuji Kanagawa", "Tomoyuki Kaneko"], "authorids": ["~Yuji_Kanagawa1", "~Tomoyuki_Kaneko1"], "authors_source": "OpenReview API", "abstract": "We consider the problem of autonomously learning reusable temporally extended actions, or options, in reinforcement learning. While options can speed up transfer learning by serving as reusable building blocks, learning reusable options for unknown task distribution remains challenging. Motivated by the recent success of mutual information (MI) based skill learning, we hypothesize that more diverse options are more reusable. To this end, we propose a method for learning termination conditions of options by maximizing MI between options and corresponding state transitions. We derive a scalable approximation of this MI maximization via gradient ascent, yielding the InfoMax Termination Critic (IMTC) algorithm. Our experiments demonstrate that IMTC significantly improves the diversity of learned options without rewards, combined with an intrinsic option learning method. Moreover, we test the reusability of learned options by transferring options into various tasks, confirming that IMTC helps quick adaptation, especially in complex domains where an agent needs to manipulate objects.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "X4FfHup_yfx", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1337/Reviewer_SWkr"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents an approach for learning diverse temporally extended and reusable options. It is based on the assumption that learning diverse options are generally useful for downstream tasks. Thus, the approach aims to discover options without making any assumptions about the downstream tasks. The main idea behind the approach is to discover options by maximizing the mutual information between the options and the corresponding state transitions and demonstrates the options discovered by this approach. Also, the paper demonstrates that these options are useful for faster learning in a downstream task. \n", "review_text": "The paper presents an extension of the termination critic idea, which is called the InfoMax Termination Critic. Specifically, the approach is different from Termination Critic as it involves computing the entropy over the final states as a function of the option and the start state (termination critic only computes the entropy of the final states as a function of the option).\n\nThe overall approach is clear and well-presented.\n\nI do have a number of questions related to this approach and would like some clarifications from the authors: \n\n1. The Fig 2 presented shows qualitatively how the terminations for options look for IMTC and Termination Critic. Would it be possible to demonstrate a similar visualization on the standard 4 room gridworld, identical to the one presented in Termination Critic’s paper? It would be much easier to make a comparison with the published result as opposed to reimplementing termination critic and demonstrating the differences on a new but smaller gridworld.\n\n2. If I understand the approach correctly, the terminations for the options are discovered through IMTC but the policy for those options are discovered through Variational Intrinsic Control (VIC). Is my understanding correct? \nThis seems like a roundabout way of discovering options that are diverse. Why doesn’t the approach learn option-policies where the rewards for those options are obtained through the discovered termination functions?  \nAlso, this makes it unclear if the performance gains are due to the discovered terminations or due to VIC. Perhaps adding an ablation to address this is necessary to make it clear.\n\n3. Would it be possible to make Fig 3 more prominent? It is hard to understand what each of the option-policies are looking like. Also, how are the options being discovered for OC + VIC? I thought that OC requires extrinsic rewards (from tasks) to discover options and in the experiments considered in this paper, there are no extrinsic rewards (i.e., reward-free RL). \n\n4. Because the approach is very much related to Termination Critic, I think it is important to include Termination Critic as a baseline in the large scale experiments (the plots in Fig 3 and Fig 5). Without this result, it is difficult to understand how the proposed work is better than its prior work.\n\n5. What are the error bars in Fig 5? \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents an approach for learning diverse temporally extended and reusable options. It is based on the assumption that learning diverse options are generally useful for downstream tasks. Thus, the approach aims to discover options without making any assumptions about the downstream tasks. The main idea behind the approach is to discover options by maximizing the mutual information between the options and the corresponding state transitions and demonstrates the options discovered by this approach. Also, the paper demonstrates that these options are useful for faster learning in a downstream task. \n", "main_review": "The paper presents an extension of the termination critic idea, which is called the InfoMax Termination Critic. Specifically, the approach is different from Termination Critic as it involves computing the entropy over the final states as a function of the option and the start state (termination critic only computes the entropy of the final states as a function of the option).\n\nThe overall approach is clear and well-presented.\n\nI do have a number of questions related to this approach and would like some clarifications from the authors: \n\n1. The Fig 2 presented shows qualitatively how the terminations for options look for IMTC and Termination Critic. Would it be possible to demonstrate a similar visualization on the standard 4 room gridworld, identical to the one presented in Termination Critic’s paper? It would be much easier to make a comparison with the published result as opposed to reimplementing termination critic and demonstrating the differences on a new but smaller gridworld.\n\n2. If I understand the approach correctly, the terminations for the options are discovered through IMTC but the policy for those options are discovered through Variational Intrinsic Control (VIC). Is my understanding correct? \nThis seems like a roundabout way of discovering options that are diverse. Why doesn’t the approach learn option-policies where the rewards for those options are obtained through the discovered termination functions?  \nAlso, this makes it unclear if the performance gains are due to the discovered terminations or due to VIC. Perhaps adding an ablation to address this is necessary to make it clear.\n\n3. Would it be possible to make Fig 3 more prominent? It is hard to understand what each of the option-policies are looking like. Also, how are the options being discovered for OC + VIC? I thought that OC requires extrinsic rewards (from tasks) to discover options and in the experiments considered in this paper, there are no extrinsic rewards (i.e., reward-free RL). \n\n4. Because the approach is very much related to Termination Critic, I think it is important to include Termination Critic as a baseline in the large scale experiments (the plots in Fig 3 and Fig 5). Without this result, it is difficult to understand how the proposed work is better than its prior work.\n\n5. What are the error bars in Fig 5? \n", "summary_of_the_review": "Overall, I think the paper addresses an important problem in RL: discovering diverse options when there are no external tasks. The paper presents an adaptation of Termination Critic. \nHowever, some of the design choices made by the approach does not seem to be reasonable. Specifically, why use Variational Intrinsic Control in combination with the proposed approach when there is a direct way of learning option-policies from the discovered termination functions. Also, using VIC in this way makes it unclear as to whether the performance gains are due to the proposed approach or due to VIC.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636296741586}, {"id": "9-z2u4j5Riv", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1337/Reviewer_XKs3"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposed an algorithm, termed infomax Termination Critic (IMTC) to learn diversified options in RL. This algorithm learns termination conditions of options by maximizing mutual information between options and corresponding state transitions. The experiments demonstrate the IMTC algorithm learns diversified options and can be reused in various tasks.", "review_text": "Strength: \n1.\tthe method is well motivated and derived \n2.\tEmpirical study in various domains show diversity of the learned options and its usefulness in transfer settings. \n\nWeakness:\n1.\tThe method is incremental and seems trivial. The method is directly derived from the termination gradient theorem from Harutyunyan et al. (2019) with reward signal constructed based on mutual information between terminal state and option. \n\n2.\tFailed to mention and compare some relevant references, E.g., reference [1] also considered learned diversified options using pseudo-reward constructed based on divergence between action distribution of different options. \n\n[1] Kamat & Precup, Diversity-Enriched Option-Critic, Neurips2020\n\n[2] Ramesh et al., Successor Options: an option discovery framework for reinforcement learning, IJCAI2019\n\nI would like to see some discussions of these references and experimental comparisons.\n\n3. This work is mostly based on empirical study, the paper could be strengthened if some theoretical analysis, such as sample/computation complexity, convergence etc. can be provided. \n\nOther comments:\nIn the abstract, “our experiments demonstrate … without rewards”,  you really mean “without environmental/extrinsic rewards”.\nProposition 2 seems simply a trivial extension of proposition 1. \nIn the experiment (5.1), why the policy over option $\\mu(o|x)$ is fixed to be $1/|O|$, instead of learned?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed an algorithm, termed infomax Termination Critic (IMTC) to learn diversified options in RL. This algorithm learns termination conditions of options by maximizing mutual information between options and corresponding state transitions. The experiments demonstrate the IMTC algorithm learns diversified options and can be reused in various tasks.", "main_review": "Strength: \n1.\tthe method is well motivated and derived \n2.\tEmpirical study in various domains show diversity of the learned options and its usefulness in transfer settings. \n\nWeakness:\n1.\tThe method is incremental and seems trivial. The method is directly derived from the termination gradient theorem from Harutyunyan et al. (2019) with reward signal constructed based on mutual information between terminal state and option. \n\n2.\tFailed to mention and compare some relevant references, E.g., reference [1] also considered learned diversified options using pseudo-reward constructed based on divergence between action distribution of different options. \n\n[1] Kamat & Precup, Diversity-Enriched Option-Critic, Neurips2020\n\n[2] Ramesh et al., Successor Options: an option discovery framework for reinforcement learning, IJCAI2019\n\nI would like to see some discussions of these references and experimental comparisons.\n\n3. This work is mostly based on empirical study, the paper could be strengthened if some theoretical analysis, such as sample/computation complexity, convergence etc. can be provided. \n\nOther comments:\nIn the abstract, “our experiments demonstrate … without rewards”,  you really mean “without environmental/extrinsic rewards”.\nProposition 2 seems simply a trivial extension of proposition 1. \nIn the experiment (5.1), why the policy over option $\\mu(o|x)$ is fixed to be $1/|O|$, instead of learned?\n", "summary_of_the_review": "In all, this paper studied an important problem in reinforcement, learning diversified options that can be reused in various tasks. Their empirical study is able to justify the diversify of the learned options and shows superior performance in transfer learning settings comparing to some baseline.  However, there are still some key references missing and should be compared against. Also, the method seems trivial extension of many existing methods. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "I don't have ethics concerns of this paper.", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635912056274}, {"id": "4xekqPTiiQf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1337/Reviewer_rEv8"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose an objective for learning a diverse set of options in which the goal is to maximize the entropy of terminating states from any initial state while at the same time minimize the entropy of terminating states given a specific option. \nIntuitively, this means learning options that tend to be deterministic, while at the same time are able to reach diverse terminating states.\nThe authors then evaluate against other option learning frameworks in a series of diverse environments.", "review_text": "Overall I enjoyed the paper, it is well organized and well written, and it touches in a topic that I personally find very interesting which is options learning.\n\nStrengths:\n- Well written paper and easy to follow.\n- Straight forward objective with a nice intuitive explanation. Should be easy to re-implement for the reader.\n- Evaluation done in a set of diverse tasks.\n\nWeakness:\n- Maybe a concrete definition of what's meant by diversity in this paper could help. The objective encourages the options to learn different terminating states, but not different trajectories. So, based on the objective two options that \"move right\" for 38 steps, and one moves \"up\" on step 39 and the other moves \"down\" on step 39 would be diverse, and would maximize the objective.\nIf this is not a behavior that does not happen in practice, it would be nice to have a discussion around it, since nothing strictly prohibits this from happening. \n\n- Where the number of options pre-defined? If so, how did you pick that number?\n\n- On eq. 9, what is the assumption on the reward function to take the maximum over 0 as optimistic. What happens if all rewards are positive in a task? Taking the max over 0 would probably not help with understimation.\n\n- Looking at the results, it's interesting to see that \"reach\" tasks with large action spaces is where the proposed method tends to underperform. Does the complexity of the task or the action space play a role in how difficult it is to learn options? I would like to see some move in-depth analysis on the results.\n\n- Following on the results, I would suggest avoiding making non-specific statements like: \"IMTC outperforms other methods with a large margin in complex PointBilliard, AntBilliard, and AntPush\". In that case, my next question is \"what is a large margin?\" How significant are these values? How many times did you re-run the experiments to make that claim?\nI would like to see some more specific numbers around these results.\n\n\nNitpick:\n- At the beginning of the paper, there is quite a bit of emphasis on \"we hypothesis diverse options are reusable\". This is a generally well-accepted hypothesis, but the way its written it makes it sound that no one thought of it before.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose an objective for learning a diverse set of options in which the goal is to maximize the entropy of terminating states from any initial state while at the same time minimize the entropy of terminating states given a specific option. \nIntuitively, this means learning options that tend to be deterministic, while at the same time are able to reach diverse terminating states.\nThe authors then evaluate against other option learning frameworks in a series of diverse environments.", "main_review": "Overall I enjoyed the paper, it is well organized and well written, and it touches in a topic that I personally find very interesting which is options learning.\n\nStrengths:\n- Well written paper and easy to follow.\n- Straight forward objective with a nice intuitive explanation. Should be easy to re-implement for the reader.\n- Evaluation done in a set of diverse tasks.\n\nWeakness:\n- Maybe a concrete definition of what's meant by diversity in this paper could help. The objective encourages the options to learn different terminating states, but not different trajectories. So, based on the objective two options that \"move right\" for 38 steps, and one moves \"up\" on step 39 and the other moves \"down\" on step 39 would be diverse, and would maximize the objective.\nIf this is not a behavior that does not happen in practice, it would be nice to have a discussion around it, since nothing strictly prohibits this from happening. \n\n- Where the number of options pre-defined? If so, how did you pick that number?\n\n- On eq. 9, what is the assumption on the reward function to take the maximum over 0 as optimistic. What happens if all rewards are positive in a task? Taking the max over 0 would probably not help with understimation.\n\n- Looking at the results, it's interesting to see that \"reach\" tasks with large action spaces is where the proposed method tends to underperform. Does the complexity of the task or the action space play a role in how difficult it is to learn options? I would like to see some move in-depth analysis on the results.\n\n- Following on the results, I would suggest avoiding making non-specific statements like: \"IMTC outperforms other methods with a large margin in complex PointBilliard, AntBilliard, and AntPush\". In that case, my next question is \"what is a large margin?\" How significant are these values? How many times did you re-run the experiments to make that claim?\nI would like to see some more specific numbers around these results.\n\n\nNitpick:\n- At the beginning of the paper, there is quite a bit of emphasis on \"we hypothesis diverse options are reusable\". This is a generally well-accepted hypothesis, but the way its written it makes it sound that no one thought of it before.\n\n\n", "summary_of_the_review": "Well written paper with an interesting objective that makes intuitive sense.\nThe results look promising, but the discussion and analysis could have more depth; there are questions left unanswered on this paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635691374115}, {"id": "4OmkJxFFVJw", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1337/Reviewer_YG8c"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose an HRL algorithm that uses the VIC objective to discover the options, i.e., the termination condition of the options is trained to maximize the mutual information between the set of options and their terminating states. These options are first trained without access to a reward function in an unsupervised manner. Later on there are experiments that show how an RL agent can re-use these options in downstream tasks. ", "review_text": "The general idea of using the termination gradient in VIC is novel and interesting. That said, I feel that the method which was implemented has some gap from the theory, and that the method that is eventually implemented is too similar to VIC. Since the VIC paper was published there were quite a few papers extending it, like DIAYN and RVIC to give two examples. Thus, the novelty of this work is more limited, as it is only comparing to VIC as a baseline. I also feel that I don't learn enough from the experimental section and the algorithm description about the actual difference proposed here. \n\nMore concretely, the implementation of the option model is not clear to me and is not well justified. The theoretical derivation suggests using a model that predicts the probability of seeing s_f after executing option o from state s. However in practice, this model is not used, and it is replaced with the discriminator as being done in VIC, which is a different kind of a model. Furthermore, using this model instead of the option model makes the methods much more similar empirically to VIC and much less novel in my opinion. I think this should be clarified and discussed in much more detail, e.g., what is exactly the difference from VIC.  \n\nSection 4.2 makes some decisions that are not very clear to me and I would have like to see some experiments supporting them. I don't see it as a major part of the paper though and it might have been better to put this section in the supplementary. \n\nThere seems to be a mistake in the definition of the \"undiscounted occupancy measure\". First of all, it is unclear which occupancy measure is it; is it the average reward occupancy? is it the finite horizon? being undiscounted doesn't tell much. Secondly, on what is the expectation being taken over? if it is on trajectories executed by the policy, then shouldn't the state occupancy be infinity for all non transient states? the average reward and finite horizon state occupancies do not suffer from this issue", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose an HRL algorithm that uses the VIC objective to discover the options, i.e., the termination condition of the options is trained to maximize the mutual information between the set of options and their terminating states. These options are first trained without access to a reward function in an unsupervised manner. Later on there are experiments that show how an RL agent can re-use these options in downstream tasks. ", "main_review": "The general idea of using the termination gradient in VIC is novel and interesting. That said, I feel that the method which was implemented has some gap from the theory, and that the method that is eventually implemented is too similar to VIC. Since the VIC paper was published there were quite a few papers extending it, like DIAYN and RVIC to give two examples. Thus, the novelty of this work is more limited, as it is only comparing to VIC as a baseline. I also feel that I don't learn enough from the experimental section and the algorithm description about the actual difference proposed here. \n\nMore concretely, the implementation of the option model is not clear to me and is not well justified. The theoretical derivation suggests using a model that predicts the probability of seeing s_f after executing option o from state s. However in practice, this model is not used, and it is replaced with the discriminator as being done in VIC, which is a different kind of a model. Furthermore, using this model instead of the option model makes the methods much more similar empirically to VIC and much less novel in my opinion. I think this should be clarified and discussed in much more detail, e.g., what is exactly the difference from VIC.  \n\nSection 4.2 makes some decisions that are not very clear to me and I would have like to see some experiments supporting them. I don't see it as a major part of the paper though and it might have been better to put this section in the supplementary. \n\nThere seems to be a mistake in the definition of the \"undiscounted occupancy measure\". First of all, it is unclear which occupancy measure is it; is it the average reward occupancy? is it the finite horizon? being undiscounted doesn't tell much. Secondly, on what is the expectation being taken over? if it is on trajectories executed by the policy, then shouldn't the state occupancy be infinity for all non transient states? the average reward and finite horizon state occupancies do not suffer from this issue", "summary_of_the_review": "Nice idea, but the connection between the theory and practice is somewhat missing, resulting in an algorithm that might be too similar to VIC, and the differences are not explained clearly enough nor supported by the experiments. In addition, including more baselines would have made the empirical contributions stronger. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634815871650}], "openreview_url": "https://openreview.net/forum?id=UTTrevGchy", "arxiv_id": "2010.02756", "paper_pdf": "papers/UTTrevGchy.pdf", "paper_pdf_sha256": "fa886e840c90f0a4921bbced03dbf0cdd122b5d0f8b317d4ee38dcfd82fd0221", "paper_pdf_bytes": 7825823, "paper_pdf_source": "openreview", "code_url": "https://github.com/kngwyu/infomax-option-critic", "code_repository": "kngwyu/infomax-option-critic", "code_commit": "9d907c041c1d0280db9b23eb2fdf9e0033e33bf3", "code_archive": "repos/UTTrevGchy.zip", "code_archive_sha256": "dd5ed1420d3d685e407d63b957414d00e8da56dba9d9e49fd7bbb6a9bd0f099d", "code_archive_bytes": 4359739, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 4256, "github_languages": {"Python": 94250}, "github_archived": false, "github_pushed_at": "2020-10-07T04:19:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/diverse-exploration-via-infomax-options-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Jf24xdaAwF9", "year": 2021, "status": "rejected", "title": "Self-Activating Neural Ensembles for Continual Reinforcement Learning", "authors": ["Sam Powers", "Abhinav Gupta"], "authorids": ["~Sam_Powers1", "~Abhinav_Gupta1"], "authors_source": "OpenReview API", "abstract": "The ability for an agent to continuously learn new skills without catastrophically forgetting existing knowledge is of critical importance for the development of generally intelligent agents. Most methods devised to address this problem depend heavily on well-defined task boundaries which simplify the problem considerably. Our task-agnostic method, Self-Activating Neural Ensembles (SANE), uses a hierarchical modular architecture designed to avoid catastrophic forgetting without making any such assumptions. At each timestep a path through the SANE tree is activated; during training only activated nodes are updated, ensuring that unused nodes do not undergo catastrophic forgetting. Additionally, new nodes are created as needed, allowing the system to leverage and retain old skills while growing and learning new ones. We demonstrate our approach on MNIST and a set of grid world environments, demonstrating that SANE does not undergo catastrophic forgetting where existing methods do.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "MinW53KvF_e", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2033/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work addresses multi-task learning where task boundaries are unknown. The approach is to construct a dynamic decision tree with nodes made up of small networks. Nodes are merged and promoted in the tree based on learned error bounds on value function estimates. Inference through the tree works by selecting nodes with the highest value prediction. It is an interesting approach for modular learning, even within the same environment.\n\nThe paper is clearly written. The approach is clearly explained. The empirical evaluation is thorough, with good control of the baselines.\n\nNode promotion is done so that a new task can take a new path through the tree, avoiding catastrophic forgetting. This is done by comparing the value function estimates of a child node to its parents. A large enough difference in value function estimate triggers a node promotion. An important assumption here is that different tasks differ substantially in long term returns. This may not be the case in general. Tasks can differ in various ways apart from the total reward. Sufficient discussion is not provided on this assumption and how it affects the results.\n\nAt the end of page 5, it is claimed that sequential MNIST is “significantly more realistic”. Why is this the case? \n\nFigure 4 is a bit hard to understand. What is the “reward” here? What is the action space for this task?\n\nIn table 1, why is the SANE accuracy for digit “0” not 100%? All other methods are 100% and this is the only class so far unless I’m misinterpreting the table.\n\nWhile SANE manages to avoid catastrophic forgetting compared to the baselines, the classification accuracy for each digit does settle back down to less than 50% in most cases. Why is this the case?\n\nHow effective is the strategy of detecting task change using value function bounds in the mini world environment? The rewards do not seem to be very different. I would be interested in seeing some examples of transitions that trigger a node promotion.\n\nThe results in this environment are not so convincing. Is it perhaps because of the values being similar across tasks?\n\nSome more discussion of how this work differs from other approaches that combine deep learning and decision tree could be included.\n\nIn section 2.3 “Training the Critics”, it is stated that the critic is trying to minimize the value v(s). Why is this done?\n\nFuture work could include using distribution RL for detecting outliers and node promotion.\n\nThis is an exciting direction of work and a difficult problem without known task boundaries. The approach is sound with decent empirical evaluations. The paper could benefit from some clarity in the explanation of the results and the assumptions made in the method. Some more in-depth analysis of the trees contracted by this method and their dynamic behavior through learning would shed some more light on some of the results.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An interesting approach for a difficult multi-task problem", "review": "This work addresses multi-task learning where task boundaries are unknown. The approach is to construct a dynamic decision tree with nodes made up of small networks. Nodes are merged and promoted in the tree based on learned error bounds on value function estimates. Inference through the tree works by selecting nodes with the highest value prediction. It is an interesting approach for modular learning, even within the same environment.\n\nThe paper is clearly written. The approach is clearly explained. The empirical evaluation is thorough, with good control of the baselines.\n\nNode promotion is done so that a new task can take a new path through the tree, avoiding catastrophic forgetting. This is done by comparing the value function estimates of a child node to its parents. A large enough difference in value function estimate triggers a node promotion. An important assumption here is that different tasks differ substantially in long term returns. This may not be the case in general. Tasks can differ in various ways apart from the total reward. Sufficient discussion is not provided on this assumption and how it affects the results.\n\nAt the end of page 5, it is claimed that sequential MNIST is “significantly more realistic”. Why is this the case? \n\nFigure 4 is a bit hard to understand. What is the “reward” here? What is the action space for this task?\n\nIn table 1, why is the SANE accuracy for digit “0” not 100%? All other methods are 100% and this is the only class so far unless I’m misinterpreting the table.\n\nWhile SANE manages to avoid catastrophic forgetting compared to the baselines, the classification accuracy for each digit does settle back down to less than 50% in most cases. Why is this the case?\n\nHow effective is the strategy of detecting task change using value function bounds in the mini world environment? The rewards do not seem to be very different. I would be interested in seeing some examples of transitions that trigger a node promotion.\n\nThe results in this environment are not so convincing. Is it perhaps because of the values being similar across tasks?\n\nSome more discussion of how this work differs from other approaches that combine deep learning and decision tree could be included.\n\nIn section 2.3 “Training the Critics”, it is stated that the critic is trying to minimize the value v(s). Why is this done?\n\nFuture work could include using distribution RL for detecting outliers and node promotion.\n\nThis is an exciting direction of work and a difficult problem without known task boundaries. The approach is sound with decent empirical evaluations. The paper could benefit from some clarity in the explanation of the results and the assumptions made in the method. Some more in-depth analysis of the trees contracted by this method and their dynamic behavior through learning would shed some more light on some of the results.\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603902502489}, {"id": "X5wEinoJ5aa", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2033/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes SANE -- an architecture and a training algorithm for continual learning. The SANE model consists of a tree where each node can act as an RL agent and where nodes act according to the dispatching mechanism based on their reward prediction. This allows to activate and update only those agents that are specialized in the current task and thus may prevent catastrophic forgetting caused by updating the whole model.\n\nNovelty:\nDifferent parts of the method did have appearance in the prior literature. For example, the general idea to organize the model into a hierarhical structure with a certain kind of dispatching can be found in hierarchical softmax [1] with a more modern variant [2]. More recently, [3] proposed a similar model with a very close motivation to address catastrophic forgetting. In my opinion, authors should make a more extensive literature review and more clearly justify novelty of the method.\n\nClarity:\nAfter reading the paper several times, I did not get a clear enough picture of the proposed model operates. Perhaps, a more formal algorithmic description would help. \n\nSignificance:\nThe empirical evaluation only consists of relatively simple experiments. MNIST, for example, is hardly representative of a real-world continual learning task. Authors do not compare SANE to string baselines that are specifically designed to prevent catastrophic forgetting, even to EWC which is cited. Even though, EWC does rely on task boundaries information, it would at the very least define the gap that exists to the methods that are agnoistic to this kind of priveledged information. \nThe lack of a thorough experimental study makes it difficult to argue for high significance of this work to the community.\n\nOther comments:\n I think what authors mean by \"discounted reward\" and $r_{disc}$ is usually denoted as \"discounted return\", this should be clarified.\n\nReferences:\n[1] A scalable hierarchical distributed language model. Andriy Mnih and Geoffrey Hinton. 2008.\n[2] Self-organized Hierarchical Softmax. Yikang Shen, Shawn Tan, Chrisopher Pal, Aaron Courville. 2017\n[3] Gated Linear Networks. Joel Veness, Tor Lattimore, David Budden, Avishkar Bhoopchand, Christopher Mattern, Agnieszka Grabska-Barwinska, Eren Sezener, Jianan Wang, Peter Toth, Simon Schmitt, Marcus Hutter. 2019", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The paper proposes SANE -- an architecture and a training algorithm for continual learning. The SANE model consists of a tree where each node can act as an RL agent and where nodes act according to the dispatching mechanism based on their reward prediction. This allows to activate and update only those agents that are specialized in the current task and thus may prevent catastrophic forgetting caused by updating the whole model.\n\nNovelty:\nDifferent parts of the method did have appearance in the prior literature. For example, the general idea to organize the model into a hierarhical structure with a certain kind of dispatching can be found in hierarchical softmax [1] with a more modern variant [2]. More recently, [3] proposed a similar model with a very close motivation to address catastrophic forgetting. In my opinion, authors should make a more extensive literature review and more clearly justify novelty of the method.\n\nClarity:\nAfter reading the paper several times, I did not get a clear enough picture of the proposed model operates. Perhaps, a more formal algorithmic description would help. \n\nSignificance:\nThe empirical evaluation only consists of relatively simple experiments. MNIST, for example, is hardly representative of a real-world continual learning task. Authors do not compare SANE to string baselines that are specifically designed to prevent catastrophic forgetting, even to EWC which is cited. Even though, EWC does rely on task boundaries information, it would at the very least define the gap that exists to the methods that are agnoistic to this kind of priveledged information. \nThe lack of a thorough experimental study makes it difficult to argue for high significance of this work to the community.\n\nOther comments:\n I think what authors mean by \"discounted reward\" and $r_{disc}$ is usually denoted as \"discounted return\", this should be clarified.\n\nReferences:\n[1] A scalable hierarchical distributed language model. Andriy Mnih and Geoffrey Hinton. 2008.\n[2] Self-organized Hierarchical Softmax. Yikang Shen, Shawn Tan, Chrisopher Pal, Aaron Courville. 2017\n[3] Gated Linear Networks. Joel Veness, Tor Lattimore, David Budden, Avishkar Bhoopchand, Christopher Mattern, Agnieszka Grabska-Barwinska, Eren Sezener, Jianan Wang, Peter Toth, Simon Schmitt, Marcus Hutter. 2019", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603878693583}, {"id": "XTd9W7h-RaH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2033/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Edit\n\nUpdated score from 4 to 5\n\n## Summary\n\n* The paper proposes a task agnostic CL method called SANE. It uses a hierarchical, modular tree-like architecture where nodes are neural networks (policy + critic + replay buffer). \n* The architecture is quite interesting. Alongside regular parameter update operations, it also supports updates to the tree-structure.\n* Each state (that a learning agent encounters) activates a new path in the tree (thus reducing negative interference) while still enabling knowledge sharing.\n* Experiments on MNIST and Mini-Grid to show the usefulness of SANE over the baselines.\n\n## Objective\n\n* Addressing the problem of continual learning in the context of reinforcement learning.\n\n## Strong Points\n\n* The proposed architecture is quite intriguing. \n* It has several practical desiderata like:\n  * Updates on the tree-structure. More nodes can be added (on the fly) while keeping the overall complexity (of training/inference) in check.\n  * Works in the task-agnostic setting, which is more general and practical than the task-aware setting.\n* The writing (while missing some key details, see \"Areas for Improvement\") generally flows well and conveys high-level ideas.\n* It is quite interesting (and motivating) that SANE retains knowledge of the previous tasks as it continues training on the new tasks. I think of the tree-structure as a way of inducing a prior for the learning agent. The tree-update operations are along the direction of learning a useful inductive bias (without hardcoding the structure). SANE could be a good step in the direction of learning priors/inductive biases from data alone.\n\n## Areas for Improvement\n\n* While the paper describes the high-level idea well, it is very vague in several design choices/implementation details. The lack of detail makes it hard to evaluate the usefulness of the approach. Some examples:\n  * The paragraph on merging nodes is very vague. What does it mean to find the closest nodes in the policy space? Do you compute the L2 norm between the predictions of the policies of all possible inputs? How are policies combined by doing a \"weighted sum of usage count\"? What does \"retraining once\" mean. How do you decide when to stop training the merged node.\n\n* Setup: The paper mentions that they are operating in the continual RL setup. The MNIST task is a one-step RL problem (or a one step decision-making problem). The MiniGrid setup is more exciting, but the paper considers only three tasks (difficult to evaluate the approach's utility when the number of tasks increases). It does not show convincing results on even those (more at a later point). \n\n* Unconvincing results: \n\nLet's start with the MiniGrid results. The paper mentions that \"The Empty8x8 graph is the most insightful: CLEAR has almost perfect recall of this environment while 2-Room is training, but catastrophically forgets it during the training of Unlock.\". Let us see the performance of CLEAR on Unlock -- Figure 5 (2nd row, 3rd column, MiniGrid Unlock). We see that around 1.5M steps, the CLEAR baseline reaches about 90% performance and starts to saturate. SANE is performing approximately 20% at this point, and even after training for 2.25M steps, SANE reaches around 80% success. My argument is, CLEAR starts overfitting after 1.5M steps while SANE is not overfitting even after 2.25M steps. Now, look at Figure 5 (2nd Row, 1st Column MiniGrid Empty). Around 1.5M steps, CLEAR is performing close to the SANE model.  I could argue that there is no need to train CLEAR (and other baselines) for additional 750K steps just because the SANE model needs to be updated. Updating for so many extra steps is also causing the CLEAR baseline to overfit to the Unlock task. Please note that I am not criticizing the decision to train SANE for 2.25M steps. I am criticizing the choice of training CLEAR much longer. \n\nRegarding the MNIST results, I have the same worry. We can see that the baselines reach close to 100% very quickly by zooming into the plots, while the SANE model takes much longer. Repeating myself,  I am not criticizing the decision to train SANE much longer. I am criticizing the choice of training the other baselines much longer, thus forcing them to overfit to the current task.\n\n* Choice of baselines: Apart from the criticism of the experiment results, I found the baselines' choice to be quite weak as well. While the authors rightly point out that task-agnostic learning is a much harder setup, they consider very few baselines (actually only one baseline, since Impala is not for CL). Regarding CLEAR and CLEARx40, I do not think CLEARx40 is a fair baseline. Using much more parameters makes the model more likely to overfit, given that the mode is always trained for a fixed number of steps, irrespective of when it starts to overfit. Some more reasonable baselines would have been:\n\n@InProceedings{Aljundi_2019_CVPR,\nauthor = {Aljundi, Rahaf and Kelchtermans, Klaas and Tuytelaars, Tinne},\ntitle = {Task-Free Continual Learning},\nbooktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},\nmonth = {June},\nyear = {2019}\n}\n\n@incollection{NIPS2019_9357,\ntitle = {Online Continual Learning with Maximal Interfered Retrieval},\nauthor = {Aljundi, Rahaf and Belilovsky, Eugene and Tuytelaars, Tinne and Charlin, Laurent and Caccia, Massimo and Lin, Min and Page-Caccia, Lucas},\nbooktitle = {Advances in Neural Information Processing Systems 32},\neditor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\\textquotesingle Alch\\'{e}-Buc and E. Fox and R. Garnett},\npages = {11849--11860},\nyear = {2019},\npublisher = {Curran Associates, Inc.},\nurl = {http://papers.nips.cc/paper/9357-online-continual-learning-with-maximal-interfered-retrieval.pdf}\n}\n\n\n@inproceedings{\nLee2020A,\ntitle={A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning},\nauthor={Soochan Lee and Junsoo Ha and Dongsu Zhang and Gunhee Kim},\nbooktitle={International Conference on Learning Representations},\nyear={2020},\nurl={https://openreview.net/forum?id=SJxSOJStPr}\n}\n\n* Limited related work: Apart from some articles mentioned in the previous point, some recent works look at learning ensembles of \"self-activating\" policies.\n\nhttps://icml.cc/Conferences/2020/ScheduleMultitrack?event=6293\n\n@inproceedings{\nGoyal2020Reinforcement,\ntitle={Reinforcement Learning with Competitive  Ensembles of Information-Constrained Primitives},\nauthor={Anirudh Goyal and Shagun Sodhani and Jonathan Binas and Xue Bin Peng and Sergey Levine and Yoshua Bengio},\nbooktitle={International Conference on Learning Representations},\nyear={2020},\nurl={https://openreview.net/forum?id=ryxgJTEYDr}\n}\n\n\n\n\n* Additional Results: It is difficult to understand what the nodes are learning. The appendix has one example, which is not sufficient to understand what the nodes are learning. It is also not clear why would the learning model try to activate different paths. This relates to the problem of learning options in RL and the commonly observed problem where only one option is active during most of the training. At the minimum, it will be useful to see some results/plots describing how the tree evolves during training and how the frequency of use of different nodes changes.\n\n* Ablations: The paper introduces several hyper-parameters but does not consider any sort of ablations with them. This makes it difficult to understand what hyper-parameter values are important and which are set arbitrarily.\n\n## Some other questions\n\nPlease note that these are some general questions, and I did not consider any of these design choices as \"wrong\" or \"against the paper.\"\n\n* Why did the paper use L2 norm and not KL penalty (section 2.3)\n\n* The paper mentions that they are operating in a task-agnostic setting. While that is true, they are still training/evaluating the model in the more limited sequential (one-task-at-a-time) setting. Is there a reason to not work in the more general any-task-at-any-time setting?\n\n* Could the authors also include a discussion of the training time for the baselines vs. SANE. My understanding is, training SANE takes a lot more time/steps.\n\n* Could the authors add some description about how rewards are computed for the MNIST setup.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Idea, limited execution", "review": "Edit\n\nUpdated score from 4 to 5\n\n## Summary\n\n* The paper proposes a task agnostic CL method called SANE. It uses a hierarchical, modular tree-like architecture where nodes are neural networks (policy + critic + replay buffer). \n* The architecture is quite interesting. Alongside regular parameter update operations, it also supports updates to the tree-structure.\n* Each state (that a learning agent encounters) activates a new path in the tree (thus reducing negative interference) while still enabling knowledge sharing.\n* Experiments on MNIST and Mini-Grid to show the usefulness of SANE over the baselines.\n\n## Objective\n\n* Addressing the problem of continual learning in the context of reinforcement learning.\n\n## Strong Points\n\n* The proposed architecture is quite intriguing. \n* It has several practical desiderata like:\n  * Updates on the tree-structure. More nodes can be added (on the fly) while keeping the overall complexity (of training/inference) in check.\n  * Works in the task-agnostic setting, which is more general and practical than the task-aware setting.\n* The writing (while missing some key details, see \"Areas for Improvement\") generally flows well and conveys high-level ideas.\n* It is quite interesting (and motivating) that SANE retains knowledge of the previous tasks as it continues training on the new tasks. I think of the tree-structure as a way of inducing a prior for the learning agent. The tree-update operations are along the direction of learning a useful inductive bias (without hardcoding the structure). SANE could be a good step in the direction of learning priors/inductive biases from data alone.\n\n## Areas for Improvement\n\n* While the paper describes the high-level idea well, it is very vague in several design choices/implementation details. The lack of detail makes it hard to evaluate the usefulness of the approach. Some examples:\n  * The paragraph on merging nodes is very vague. What does it mean to find the closest nodes in the policy space? Do you compute the L2 norm between the predictions of the policies of all possible inputs? How are policies combined by doing a \"weighted sum of usage count\"? What does \"retraining once\" mean. How do you decide when to stop training the merged node.\n\n* Setup: The paper mentions that they are operating in the continual RL setup. The MNIST task is a one-step RL problem (or a one step decision-making problem). The MiniGrid setup is more exciting, but the paper considers only three tasks (difficult to evaluate the approach's utility when the number of tasks increases). It does not show convincing results on even those (more at a later point). \n\n* Unconvincing results: \n\nLet's start with the MiniGrid results. The paper mentions that \"The Empty8x8 graph is the most insightful: CLEAR has almost perfect recall of this environment while 2-Room is training, but catastrophically forgets it during the training of Unlock.\". Let us see the performance of CLEAR on Unlock -- Figure 5 (2nd row, 3rd column, MiniGrid Unlock). We see that around 1.5M steps, the CLEAR baseline reaches about 90% performance and starts to saturate. SANE is performing approximately 20% at this point, and even after training for 2.25M steps, SANE reaches around 80% success. My argument is, CLEAR starts overfitting after 1.5M steps while SANE is not overfitting even after 2.25M steps. Now, look at Figure 5 (2nd Row, 1st Column MiniGrid Empty). Around 1.5M steps, CLEAR is performing close to the SANE model.  I could argue that there is no need to train CLEAR (and other baselines) for additional 750K steps just because the SANE model needs to be updated. Updating for so many extra steps is also causing the CLEAR baseline to overfit to the Unlock task. Please note that I am not criticizing the decision to train SANE for 2.25M steps. I am criticizing the choice of training CLEAR much longer. \n\nRegarding the MNIST results, I have the same worry. We can see that the baselines reach close to 100% very quickly by zooming into the plots, while the SANE model takes much longer. Repeating myself,  I am not criticizing the decision to train SANE much longer. I am criticizing the choice of training the other baselines much longer, thus forcing them to overfit to the current task.\n\n* Choice of baselines: Apart from the criticism of the experiment results, I found the baselines' choice to be quite weak as well. While the authors rightly point out that task-agnostic learning is a much harder setup, they consider very few baselines (actually only one baseline, since Impala is not for CL). Regarding CLEAR and CLEARx40, I do not think CLEARx40 is a fair baseline. Using much more parameters makes the model more likely to overfit, given that the mode is always trained for a fixed number of steps, irrespective of when it starts to overfit. Some more reasonable baselines would have been:\n\n@InProceedings{Aljundi_2019_CVPR,\nauthor = {Aljundi, Rahaf and Kelchtermans, Klaas and Tuytelaars, Tinne},\ntitle = {Task-Free Continual Learning},\nbooktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},\nmonth = {June},\nyear = {2019}\n}\n\n@incollection{NIPS2019_9357,\ntitle = {Online Continual Learning with Maximal Interfered Retrieval},\nauthor = {Aljundi, Rahaf and Belilovsky, Eugene and Tuytelaars, Tinne and Charlin, Laurent and Caccia, Massimo and Lin, Min and Page-Caccia, Lucas},\nbooktitle = {Advances in Neural Information Processing Systems 32},\neditor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\\textquotesingle Alch\\'{e}-Buc and E. Fox and R. Garnett},\npages = {11849--11860},\nyear = {2019},\npublisher = {Curran Associates, Inc.},\nurl = {http://papers.nips.cc/paper/9357-online-continual-learning-with-maximal-interfered-retrieval.pdf}\n}\n\n\n@inproceedings{\nLee2020A,\ntitle={A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning},\nauthor={Soochan Lee and Junsoo Ha and Dongsu Zhang and Gunhee Kim},\nbooktitle={International Conference on Learning Representations},\nyear={2020},\nurl={https://openreview.net/forum?id=SJxSOJStPr}\n}\n\n* Limited related work: Apart from some articles mentioned in the previous point, some recent works look at learning ensembles of \"self-activating\" policies.\n\nhttps://icml.cc/Conferences/2020/ScheduleMultitrack?event=6293\n\n@inproceedings{\nGoyal2020Reinforcement,\ntitle={Reinforcement Learning with Competitive  Ensembles of Information-Constrained Primitives},\nauthor={Anirudh Goyal and Shagun Sodhani and Jonathan Binas and Xue Bin Peng and Sergey Levine and Yoshua Bengio},\nbooktitle={International Conference on Learning Representations},\nyear={2020},\nurl={https://openreview.net/forum?id=ryxgJTEYDr}\n}\n\n\n\n\n* Additional Results: It is difficult to understand what the nodes are learning. The appendix has one example, which is not sufficient to understand what the nodes are learning. It is also not clear why would the learning model try to activate different paths. This relates to the problem of learning options in RL and the commonly observed problem where only one option is active during most of the training. At the minimum, it will be useful to see some results/plots describing how the tree evolves during training and how the frequency of use of different nodes changes.\n\n* Ablations: The paper introduces several hyper-parameters but does not consider any sort of ablations with them. This makes it difficult to understand what hyper-parameter values are important and which are set arbitrarily.\n\n## Some other questions\n\nPlease note that these are some general questions, and I did not consider any of these design choices as \"wrong\" or \"against the paper.\"\n\n* Why did the paper use L2 norm and not KL penalty (section 2.3)\n\n* The paper mentions that they are operating in a task-agnostic setting. While that is true, they are still training/evaluating the model in the more limited sequential (one-task-at-a-time) setting. Is there a reason to not work in the more general any-task-at-any-time setting?\n\n* Could the authors also include a discussion of the training time for the baselines vs. SANE. My understanding is, training SANE takes a lot more time/steps.\n\n* Could the authors add some description about how rewards are computed for the MNIST setup.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603833458630}, {"id": "TSEW0lnhBvl", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2033/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Post-rebuttal update:\n\nI thank the authors for their responses to my queries and for updating the paper. I think the paper has improved, the main reason being that the replay-ratio parameter sweeps for CLEAR in the MNIST experiments show that SANE does improve the stability-plasticity tradeoff, in particular the runs with a 0.97 replay ratio (only shown in the rebuttal pdf but I think should make it into the appendix of the paper), which show that the memory of CLEAR cannot be improved to match SANE's without severely affecting its ability to learn new tasks. The only other set of experiments, the sequence of three minigrid tasks, however, do not show an overall improvement over CLEAR, and so overall the experimental evaluation is still weak. It is hard to judge how good the method is without extending the range of baselines (I appreciate the addition of the and task settings. As a result, I am only increasing my score from a 4 to a 5, but I do think SANE is an interesting method and I encourage the authors to keep working at it to prove its effectiveness.\n\n---------------------------\nOriginal review:\n\nThis paper presents SANE (Self-Activating Neural Ensembles) a new dynamic architecture method for mitigating catastrophic forgetting in neural networks in the context of reinforcement learning that is task-agnostic. The architecture consists of a tree where each node comprises its own value critic, policy and replay buffer. Action selection at each time step begins by starting at the root of the three and activating the child that predicts the most optimistic value, and repeating this procedure from the activated child until a leaf node is reached, whose policy is used to act with. Training has two components: (i) the parameters in activated nodes are updated using REINFORCE, and (ii) structural updates are made to the tree whereby children that are sufficiently different to their parents are promoted up the tree and given their own children, and children that are similar to each other are merged in order to limit growth of the tree. The method is empirically compared to two baselines, IMPALA and IMPALA with CLEAR (an experience-replay based continual learning method), on two sequences of tasks, sequential MNIST and a sequence of three egocentric gridworld tasks (MiniGrid). On sequential MNIST, SANE displays less forgetting but less adaptivity than the baselines, and on the gridworld tasks the results are mixed.\n\nSANE is an interesting and inventive method but I do not think it is ready for publication yet because (i) the experiments do not sufficiently support the claim that SANE improves catastrophic forgetting over the (limited set of) baselines, and (ii) SANE has a lot of components and hyperparameters but it is not clear which aspects are important for it to work. I believe the paper could be much improved by (i) demonstrating that the baselines have been adequately tuned (and perhaps increasing the set of methods compared to, e.g. with another dynamic architecture method), (ii) including a number of ablation studies or other experiments that better elucidate the inner workings of SANE and justify architectural choices.\n\nPositives:\n\t•\tSANE incorporates a number of interesting mechanisms that might intuitively be useful for mitigating catastrophic forgetting. For example, by only updating activated nodes, modularity is preserved during training in a way that can protect forgetting in other parts of the network. Also, the structural updates provide an interesting way of expanding and contracting the architecture as necessary under fixed resource constraints - an important challenge for continual learning.
\n\t•\tThe problem that SANE tackles, i.e. task-agnostic continual learning, is an important one for which there are not a large number of approaches in the literature.
\n\t•\tAn effort is made to equalise the number of parameters used by the baseline agents and SANE in the experiments for a fair comparison, and experiments are repeated with several random seeds.
\n\nConcerns:\n\t•\tSANE is claimed to improve on the baselines of IMPALA and CLEAR on the axis of mitigating catastrophic forgetting, but the experiments do not demonstrate this clearly.
\n\t◦\tIn the sequential MNIST experiments, SANE retains some performance on earlier tasks as training progress but the performance of both IMPALA and CLEAR seems to drop almost instantaneously after the task is switched. On the other hand, both IMPALA and CLEAR learn the new task extremely quickly compared to SANE, which usually does not even reach the same maximum level of performance as the baselines. This suggests that SANE shifts the tradeoff of adaptivity vs. remembering towards remembering, but doesn’t show that the tradeoff is improved vs the baselines. In the CLEAR paper, there is an important hyperparameter that mediates this tradeoff, which is the proportion of replay experiences vs. new experiences used for training - this parameter should be tuned here to be able to show that CLEAR is indeed inferior to SANE on this task. (Judging from the code it seems that a 50/50 split is used, is this correct?)
\n\t◦\tIn the sequence of three MiniGrid tasks, even though the performance of the CLEAR agent on the first task eventually drops below that of the SANE agent, the SANE agent initially deteriorates a lot faster. For the second task there seems to be no significant difference between agents. Additionally, the SANE agent is much slower to learn on the third task, never reaching the performance of the CLEAR agent, again demonstrating that the SANE agent is trading off adaptivity for slightly better memory. It’s not evident that SANE is improving the tradeoff vs. CLEAR without tuning the latter’s hyperparameters.
\n\t•\tIt is difficult to discern which elements of SANE are important for its performance in order to justify some of the architectural choices. For example:
\n\t◦\tThe critics are trained to estimate both the value and the variance of the value, and, in the inference stage, the node with the highest upper bound of estimated reward is chosen. How important is this UCB-like policy for the performance of the agent, and the value of the hyperparameter alpha that scales the standard deviation term? Would SANE work just as well if just the value estimate was used to choose the next node? Could this lead to a wrong solution if applied in a highly stochastic environment?
\n\t◦\tThe policies of the nodes in SANE use REINFORCE but IMPALA and CLEAR use VMPO; is there a reason for this difference?
\n\t◦\tIn the structural updates, for node promotion the value functions of a child and its parent are compared, but for node merging, the policies are compared - is there a rationale behind these choices? 
\n\t◦\tIn the MiniGrid experiments, SANE is run with either 10 or 15 top level nodes, resulting in a relatively big difference in performance for the 1st and 3rd tasks, giving some indication of the sensitivity of this hyperparameter. Why were 20 top-level nodes chosen for the MNIST experiments and was the performance sensitive to this choice? How about other hyperparams such as number of child nodes and the number of times a node has to have been used before it can be promoted?
\n\nOther comments:\n\t•\tThe number of baselines used for comparison is limited, especially given that one of them (IMPALA) is not designed for continual learning. It would at least be good to cite other task-agnostic methods such as [1,2].
\n\t•\tIt is slightly odd to frame sequential MNIST as an RL problem, when it is a supervised learning task.
\n\t•\tA visualisation is given in the appendix showing an example of a node that seems to represent a policy that always moves the agent forward one step in the gridworld. Presumably this is shown to indicate the specialisation of modules in the tree; it would be very interesting to see more evidence of specialisation/diversity across modules since this is one of the motivating factors for the architecture and also to see how the tree structurally adapts at task switches - are there more promotions when the data distribution changes? 
\n\n[1] Aljundi, Rahaf, Klaas Kelchtermans, and Tinne Tuytelaars. \"Task-free continual learning.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n[2] Zeno, Chen, et al. \"Task agnostic continual learning using online variational bayes.\" arXiv preprint arXiv:1803.10123 (2018).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea but lacking in empirical validation and justification of architectural choices", "review": "Post-rebuttal update:\n\nI thank the authors for their responses to my queries and for updating the paper. I think the paper has improved, the main reason being that the replay-ratio parameter sweeps for CLEAR in the MNIST experiments show that SANE does improve the stability-plasticity tradeoff, in particular the runs with a 0.97 replay ratio (only shown in the rebuttal pdf but I think should make it into the appendix of the paper), which show that the memory of CLEAR cannot be improved to match SANE's without severely affecting its ability to learn new tasks. The only other set of experiments, the sequence of three minigrid tasks, however, do not show an overall improvement over CLEAR, and so overall the experimental evaluation is still weak. It is hard to judge how good the method is without extending the range of baselines (I appreciate the addition of the and task settings. As a result, I am only increasing my score from a 4 to a 5, but I do think SANE is an interesting method and I encourage the authors to keep working at it to prove its effectiveness.\n\n---------------------------\nOriginal review:\n\nThis paper presents SANE (Self-Activating Neural Ensembles) a new dynamic architecture method for mitigating catastrophic forgetting in neural networks in the context of reinforcement learning that is task-agnostic. The architecture consists of a tree where each node comprises its own value critic, policy and replay buffer. Action selection at each time step begins by starting at the root of the three and activating the child that predicts the most optimistic value, and repeating this procedure from the activated child until a leaf node is reached, whose policy is used to act with. Training has two components: (i) the parameters in activated nodes are updated using REINFORCE, and (ii) structural updates are made to the tree whereby children that are sufficiently different to their parents are promoted up the tree and given their own children, and children that are similar to each other are merged in order to limit growth of the tree. The method is empirically compared to two baselines, IMPALA and IMPALA with CLEAR (an experience-replay based continual learning method), on two sequences of tasks, sequential MNIST and a sequence of three egocentric gridworld tasks (MiniGrid). On sequential MNIST, SANE displays less forgetting but less adaptivity than the baselines, and on the gridworld tasks the results are mixed.\n\nSANE is an interesting and inventive method but I do not think it is ready for publication yet because (i) the experiments do not sufficiently support the claim that SANE improves catastrophic forgetting over the (limited set of) baselines, and (ii) SANE has a lot of components and hyperparameters but it is not clear which aspects are important for it to work. I believe the paper could be much improved by (i) demonstrating that the baselines have been adequately tuned (and perhaps increasing the set of methods compared to, e.g. with another dynamic architecture method), (ii) including a number of ablation studies or other experiments that better elucidate the inner workings of SANE and justify architectural choices.\n\nPositives:\n\t•\tSANE incorporates a number of interesting mechanisms that might intuitively be useful for mitigating catastrophic forgetting. For example, by only updating activated nodes, modularity is preserved during training in a way that can protect forgetting in other parts of the network. Also, the structural updates provide an interesting way of expanding and contracting the architecture as necessary under fixed resource constraints - an important challenge for continual learning.
\n\t•\tThe problem that SANE tackles, i.e. task-agnostic continual learning, is an important one for which there are not a large number of approaches in the literature.
\n\t•\tAn effort is made to equalise the number of parameters used by the baseline agents and SANE in the experiments for a fair comparison, and experiments are repeated with several random seeds.
\n\nConcerns:\n\t•\tSANE is claimed to improve on the baselines of IMPALA and CLEAR on the axis of mitigating catastrophic forgetting, but the experiments do not demonstrate this clearly.
\n\t◦\tIn the sequential MNIST experiments, SANE retains some performance on earlier tasks as training progress but the performance of both IMPALA and CLEAR seems to drop almost instantaneously after the task is switched. On the other hand, both IMPALA and CLEAR learn the new task extremely quickly compared to SANE, which usually does not even reach the same maximum level of performance as the baselines. This suggests that SANE shifts the tradeoff of adaptivity vs. remembering towards remembering, but doesn’t show that the tradeoff is improved vs the baselines. In the CLEAR paper, there is an important hyperparameter that mediates this tradeoff, which is the proportion of replay experiences vs. new experiences used for training - this parameter should be tuned here to be able to show that CLEAR is indeed inferior to SANE on this task. (Judging from the code it seems that a 50/50 split is used, is this correct?)
\n\t◦\tIn the sequence of three MiniGrid tasks, even though the performance of the CLEAR agent on the first task eventually drops below that of the SANE agent, the SANE agent initially deteriorates a lot faster. For the second task there seems to be no significant difference between agents. Additionally, the SANE agent is much slower to learn on the third task, never reaching the performance of the CLEAR agent, again demonstrating that the SANE agent is trading off adaptivity for slightly better memory. It’s not evident that SANE is improving the tradeoff vs. CLEAR without tuning the latter’s hyperparameters.
\n\t•\tIt is difficult to discern which elements of SANE are important for its performance in order to justify some of the architectural choices. For example:
\n\t◦\tThe critics are trained to estimate both the value and the variance of the value, and, in the inference stage, the node with the highest upper bound of estimated reward is chosen. How important is this UCB-like policy for the performance of the agent, and the value of the hyperparameter alpha that scales the standard deviation term? Would SANE work just as well if just the value estimate was used to choose the next node? Could this lead to a wrong solution if applied in a highly stochastic environment?
\n\t◦\tThe policies of the nodes in SANE use REINFORCE but IMPALA and CLEAR use VMPO; is there a reason for this difference?
\n\t◦\tIn the structural updates, for node promotion the value functions of a child and its parent are compared, but for node merging, the policies are compared - is there a rationale behind these choices? 
\n\t◦\tIn the MiniGrid experiments, SANE is run with either 10 or 15 top level nodes, resulting in a relatively big difference in performance for the 1st and 3rd tasks, giving some indication of the sensitivity of this hyperparameter. Why were 20 top-level nodes chosen for the MNIST experiments and was the performance sensitive to this choice? How about other hyperparams such as number of child nodes and the number of times a node has to have been used before it can be promoted?
\n\nOther comments:\n\t•\tThe number of baselines used for comparison is limited, especially given that one of them (IMPALA) is not designed for continual learning. It would at least be good to cite other task-agnostic methods such as [1,2].
\n\t•\tIt is slightly odd to frame sequential MNIST as an RL problem, when it is a supervised learning task.
\n\t•\tA visualisation is given in the appendix showing an example of a node that seems to represent a policy that always moves the agent forward one step in the gridworld. Presumably this is shown to indicate the specialisation of modules in the tree; it would be very interesting to see more evidence of specialisation/diversity across modules since this is one of the motivating factors for the architecture and also to see how the tree structurally adapts at task switches - are there more promotions when the data distribution changes? 
\n\n[1] Aljundi, Rahaf, Klaas Kelchtermans, and Tinne Tuytelaars. \"Task-free continual learning.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n[2] Zeno, Chen, et al. \"Task agnostic continual learning using online variational bayes.\" arXiv preprint arXiv:1803.10123 (2018).", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603702664749}], "openreview_url": "https://openreview.net/forum?id=Jf24xdaAwF9", "arxiv_id": "2301.00141", "paper_pdf": "papers/Jf24xdaAwF9.pdf", "paper_pdf_sha256": "0eee8234ca66b80d4b4cbb8a116b278961d35b81e7d05e13adcd1baee7aac1d3", "paper_pdf_bytes": 2871749, "paper_pdf_source": "openreview", "code_url": "https://github.com/AGI-Labs/continual_rl", "code_repository": "AGI-Labs/continual_rl", "code_commit": "f2754bb282757829765beb4703f24b87efa13ff9", "code_archive": "repos/Jf24xdaAwF9.zip", "code_archive_sha256": "efbc4a642120339684e9f830a7a2108670d7a6ed80ab446c4a3718ba15a19373", "code_archive_bytes": 677532, "code_file_count": 88, "code_extensions": {".py": 88}, "github_disk_usage_kb": 2633, "github_languages": {"Python": 480548}, "github_archived": false, "github_pushed_at": "2023-07-06T14:04:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/self-activating-neural-ensembles-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Hd1Wciuyka", "year": 2026, "status": "rejected", "title": "Jodi: Unification of Visual Generation and Understanding via Joint Modeling", "authors": ["Yifeng Xu", "Zhenliang He", "Meina Kan", "Shiguang Shan", "Xilin Chen"], "authorids": ["~Yifeng_Xu1", "~Zhenliang_He2", "~Meina_Kan1", "~Shiguang_Shan2", "~Xilin_Chen1"], "authors_source": "OpenReview API", "abstract": "Visual generation and understanding are two deeply interconnected aspects of human intelligence, yet they have been traditionally treated as separate tasks in machine learning. In this paper, we propose Jodi, a diffusion framework that unifies visual generation and understanding by jointly modeling the image domain and multiple label domains. Specifically, Jodi is built upon a linear diffusion transformer along with a Role-Switch mechanism, which enables it to perform three particular types of tasks: (1) joint generation, where the model simultaneously generates images and multiple labels; (2) controllable generation, where images are generated conditioned on any combination of labels; and (3) image perception, where multiple labels can be predicted at once from a given image. Furthermore, we present the Joint-1.6M dataset, which contains 200K high-quality images collected from public sources, automatic labels for 7 visual domains, and LLM-generated captions. Extensive experiments demonstrate that our Jodi excels in generation tasks and performs competitively in understanding tasks. Besides, Jodi exhibits strong extensibility to new visual domains. Codes, data, and model weights will be publicly available.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "yZsoqUSigf", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12035/Reviewer_yWSM"], "rating": 6, "soundness": 4, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "This paper introduces Jodi, a diffusion-based model that tries to unify image generation and image understanding by jointly modeling images and multiple label modalities. With the proposed “Role-Switch” mechanism, the same model can generate images with labels, generate images conditioned on any combination of labels, or predict labels from an input image. It uses a linear diffusion transformer and domain-invariant positional embeddings to keep computation manageable and maintain consistency across domains. The authors also provide a new dataset covering 8 visual domains. Experiments show good performance on both generation and perception tasks and strong scalability. Overall, the idea is intuitive and nicely executed, and the results suggest that a unified approach like this is becoming useful.", "review_text": "This paper introduces Jodi, a diffusion-based model that tries to unify image generation and image understanding by jointly modeling images and multiple label modalities. With the proposed “Role-Switch” mechanism, the same model can generate images with labels, generate images conditioned on any combination of labels, or predict labels from an input image. It uses a linear diffusion transformer and domain-invariant positional embeddings to keep computation manageable and maintain consistency across domains. The authors also provide a new dataset covering 8 visual domains. Experiments show good performance on both generation and perception tasks and strong scalability. Overall, the idea is intuitive and nicely executed, and the results suggest that a unified approach like this is becoming useful.", "strengths": "* The role switch mechanism:  by randomly switching each domain between being generated, used as conditioning, or ignored, the model learns all the key distributions for both image generation and perception at once -- giving one network the flexibility to do many things.\n\n* Paper is well motivated, and generally well written and easy to follow.\n\n* The proposed method uses shared positional embeddings across domains (+ small role tags) so that the model knows which pixels align spatially, making it easier to keep different visual modalities consistent with each other.\n\n* The proposed Joint generation could be very valuable in many downstream applications, such as artistic manipulation or asset usage in the traditional frameworks. \n  \n* Experimental results validate the proposed method convincingly.", "weaknesses": "* “Extensibility to more domains” is claimed, but it’s unclear how much effort is needed for an entirely new modality.\n\n* Baselines for multimodal generation/understanding might not share the same supervision or label availability. \nThis needs to be clarified in the paper. \n\n* Randomly switching roles during training might make optimization harder or convergence slower, especially as domains increase.\n\n* Even with linear attention, jointly modeling 8+ domains could still be memory-intensive and slow for large resolutions.\n\n* The “multiple label domains” are pseudo labels (depth, normals, etc.), so errors/noise in these domains may propagate through the unified model.", "questions": "* How well does the method scale to more detailed label domains (e.g., full-class/instance segmentation or dense keypoints) without redesign?\n\n* What are the failure cases on real or diverse datasets, especially with human subjects? Although the limitations are discussed, the failure cases are not provided (neither in the main paper nor in the supplementary).\n\n* Can the authors provide per-domain trade-offs -- do any tasks degrade notably compared to specialist methods?\n The results currently reported appear not to be on fair ground comparisons.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Jodi, a diffusion-based model that tries to unify image generation and image understanding by jointly modeling images and multiple label modalities. With the proposed “Role-Switch” mechanism, the same model can generate images with labels, generate images conditioned on any combination of labels, or predict labels from an input image. It uses a linear diffusion transformer and domain-invariant positional embeddings to keep computation manageable and maintain consistency across domains. The authors also provide a new dataset covering 8 visual domains. Experiments show good performance on both generation and perception tasks and strong scalability. Overall, the idea is intuitive and nicely executed, and the results suggest that a unified approach like this is becoming useful.", "soundness": 4, "presentation": 3, "contribution": 4, "strengths": "* The role switch mechanism:  by randomly switching each domain between being generated, used as conditioning, or ignored, the model learns all the key distributions for both image generation and perception at once -- giving one network the flexibility to do many things.\n\n* Paper is well motivated, and generally well written and easy to follow.\n\n* The proposed method uses shared positional embeddings across domains (+ small role tags) so that the model knows which pixels align spatially, making it easier to keep different visual modalities consistent with each other.\n\n* The proposed Joint generation could be very valuable in many downstream applications, such as artistic manipulation or asset usage in the traditional frameworks. \n  \n* Experimental results validate the proposed method convincingly.", "weaknesses": "* “Extensibility to more domains” is claimed, but it’s unclear how much effort is needed for an entirely new modality.\n\n* Baselines for multimodal generation/understanding might not share the same supervision or label availability. \nThis needs to be clarified in the paper. \n\n* Randomly switching roles during training might make optimization harder or convergence slower, especially as domains increase.\n\n* Even with linear attention, jointly modeling 8+ domains could still be memory-intensive and slow for large resolutions.\n\n* The “multiple label domains” are pseudo labels (depth, normals, etc.), so errors/noise in these domains may propagate through the unified model.", "questions": "* How well does the method scale to more detailed label domains (e.g., full-class/instance segmentation or dense keypoints) without redesign?\n\n* What are the failure cases on real or diverse datasets, especially with human subjects? Although the limitations are discussed, the failure cases are not provided (neither in the main paper nor in the supplementary).\n\n* Can the authors provide per-domain trade-offs -- do any tasks degrade notably compared to specialist methods?\n The results currently reported appear not to be on fair ground comparisons.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1762071716931}, {"id": "NAAgMEolbZ", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12035/Reviewer_LQZ6"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents Jodi, a unified diffusion framework designed to jointly model the distribution of images ($x$) and multiple label domains ($y^n$), such as depth, normal, and segmentation. The goal is to unify visual generation and understanding, which are typically treated as separate tasks. The core technical contribution is a \"Role-Switch\" mechanism, where each domain is randomly assigned as a generation target [G], a condition [C], or ignored [X] during training. This principled approach, based on joint probability modeling, allows the single model to perform three distinct tasks: joint generation ($p(x, y, ...)$), controllable generation ($p(x|y, ...)$), and multi-label image perception ($p(y, ...|x)$). The framework is built on an efficient linear diffusion transformer to handle the computational load of many domains and is trained on a newly curated \"Joint-1.6M\" dataset.", "review_text": "This paper presents Jodi, a unified diffusion framework designed to jointly model the distribution of images ($x$) and multiple label domains ($y^n$), such as depth, normal, and segmentation. The goal is to unify visual generation and understanding, which are typically treated as separate tasks. The core technical contribution is a \"Role-Switch\" mechanism, where each domain is randomly assigned as a generation target [G], a condition [C], or ignored [X] during training. This principled approach, based on joint probability modeling, allows the single model to perform three distinct tasks: joint generation ($p(x, y, ...)$), controllable generation ($p(x|y, ...)$), and multi-label image perception ($p(y, ...|x)$). The framework is built on an efficient linear diffusion transformer to handle the computational load of many domains and is trained on a newly curated \"Joint-1.6M\" dataset.", "strengths": "1. Principled and Elegant Framework: The core idea of unifying $p(x|y)$ and $p(y|x)$ by modeling the joint distribution $p(x, y)$ is statistically elegant. The \"Role-Switch\" mechanism is a clever and direct implementation of this principle, forcing the model to learn a wide range of conditional and marginal distributions within one architecture.\n2. The authors made smart architectural choices. The use of a linear diffusion transformer (Sana) correctly identifies and solves the $\\mathcal{O}(M^2)$ computational bottleneck of multi-domain ($M$) attention. The \"masked linear attention\" and \"domain-invariant positional embeddings\" are solid supporting contributions that are well-motivated and essential for the framework to function.\n3. The paper introduces the \"Joint-1.6M\" dataset (200K images + 7 predicted labels) and aggregates 90K images with ground-truth labels. This is a valuable contribution to the community that will enable future research in joint visual modeling.", "weaknesses": "The model's \"unification\" comes at a significant cost to specialist performance. Though perform better than omni-models on edge detection and normal estimation tasks, Jodi performs noticeably worse than SOTA specialist models on albedo estimation and depth estimation. For example, in depth estimation (Table 2), Jodi achieves 10.1 AbsRel on NYUv2, while the specialist Lotus-D achieves 5.1. In albedo estimation (Table 4), Jodi gets 15.5 PSNR, while the specialist RGB2X gets 20.6. The model excels at generality but does not \"excel in... understanding tasks\" as claimed.", "questions": "1. The perception results (especially Tables 2, 4) are lower than some SOTA unified models and specialist models significantly. Is this performance gap an unavoidable cost of the \"jack of all trades\" unification, or do the authors see a clear path for Jodi to actually surpass some specialist models in understanding tasks?\n2. What was the sampling strategy for the [G], [C], and [X] roles? Were they sampled uniformly at random for all domains? Or was a curriculum used (e.g., more [C] roles early in training) to stabilize the learning of this complex joint distribution?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Jodi, a unified diffusion framework designed to jointly model the distribution of images ($x$) and multiple label domains ($y^n$), such as depth, normal, and segmentation. The goal is to unify visual generation and understanding, which are typically treated as separate tasks. The core technical contribution is a \"Role-Switch\" mechanism, where each domain is randomly assigned as a generation target [G], a condition [C], or ignored [X] during training. This principled approach, based on joint probability modeling, allows the single model to perform three distinct tasks: joint generation ($p(x, y, ...)$), controllable generation ($p(x|y, ...)$), and multi-label image perception ($p(y, ...|x)$). The framework is built on an efficient linear diffusion transformer to handle the computational load of many domains and is trained on a newly curated \"Joint-1.6M\" dataset.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. Principled and Elegant Framework: The core idea of unifying $p(x|y)$ and $p(y|x)$ by modeling the joint distribution $p(x, y)$ is statistically elegant. The \"Role-Switch\" mechanism is a clever and direct implementation of this principle, forcing the model to learn a wide range of conditional and marginal distributions within one architecture.\n2. The authors made smart architectural choices. The use of a linear diffusion transformer (Sana) correctly identifies and solves the $\\mathcal{O}(M^2)$ computational bottleneck of multi-domain ($M$) attention. The \"masked linear attention\" and \"domain-invariant positional embeddings\" are solid supporting contributions that are well-motivated and essential for the framework to function.\n3. The paper introduces the \"Joint-1.6M\" dataset (200K images + 7 predicted labels) and aggregates 90K images with ground-truth labels. This is a valuable contribution to the community that will enable future research in joint visual modeling.", "weaknesses": "The model's \"unification\" comes at a significant cost to specialist performance. Though perform better than omni-models on edge detection and normal estimation tasks, Jodi performs noticeably worse than SOTA specialist models on albedo estimation and depth estimation. For example, in depth estimation (Table 2), Jodi achieves 10.1 AbsRel on NYUv2, while the specialist Lotus-D achieves 5.1. In albedo estimation (Table 4), Jodi gets 15.5 PSNR, while the specialist RGB2X gets 20.6. The model excels at generality but does not \"excel in... understanding tasks\" as claimed.", "questions": "1. The perception results (especially Tables 2, 4) are lower than some SOTA unified models and specialist models significantly. Is this performance gap an unavoidable cost of the \"jack of all trades\" unification, or do the authors see a clear path for Jodi to actually surpass some specialist models in understanding tasks?\n2. What was the sampling strategy for the [G], [C], and [X] roles? Were they sampled uniformly at random for all domains? Or was a curriculum used (e.g., more [C] roles early in training) to stabilize the learning of this complex joint distribution?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762000367026}, {"id": "ZTbAWyWHnC", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12035/Reviewer_qawp"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper proposes Jodi, a unified diffusion model that can jointly generate images and their corresponding conditions (depth, mask, etc). To achieve this, Jodi considers a unified input format (role switching) and utilizes domain-invariant positional embeddings to optimize flow-based DiT model. Extensive experiments on various benchmarks demonstrate their methods efficacy in unified generation.", "review_text": "The paper proposes Jodi, a unified diffusion model that can jointly generate images and their corresponding conditions (depth, mask, etc). To achieve this, Jodi considers a unified input format (role switching) and utilizes domain-invariant positional embeddings to optimize flow-based DiT model. Extensive experiments on various benchmarks demonstrate their methods efficacy in unified generation.", "strengths": "I like this paper due to the following strengths:\n\n(1) Flexible Control: The framework naturally supports complex, multi-modal conditioning, offering unparalleled flexibility for creative applications.\n\n(2) Enhanced Generalization: By forcing the model to simultaneously learn the generative process  and the analytical structure, JODI is likely to develop a richer, more robust latent representation of visual concepts.", "weaknesses": "(1) Lack of T2I evaluation. As demonstrated in Fig. 11, the generated image is the bridge between text and other condisions. Therefore, I would like to see some comparison with T2I methods on GenEval or other T2I benchmarks.\n\n(2) The proposed method achieves condition to image by modeling the joint distribution of various conditions. However, some generation tasks (e.g., depth to image [a]) could be evaluated in a more precise manner. The authors should compare, or at least, discuss with such related works. \n\n(3) Minor: In Tab. 1-openpose, controlnet seems to achieve better result than yours in terms of LPIPS.\n\n\n[a] 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. In ICLR'25.\n[b] 3DIS-FLUX: simple and efficient multi-instance generation with DiT rendering. In Arxiv.", "questions": "Please refer to \"weaknesses\"", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Jodi, a unified diffusion model that can jointly generate images and their corresponding conditions (depth, mask, etc). To achieve this, Jodi considers a unified input format (role switching) and utilizes domain-invariant positional embeddings to optimize flow-based DiT model. Extensive experiments on various benchmarks demonstrate their methods efficacy in unified generation.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "I like this paper due to the following strengths:\n\n(1) Flexible Control: The framework naturally supports complex, multi-modal conditioning, offering unparalleled flexibility for creative applications.\n\n(2) Enhanced Generalization: By forcing the model to simultaneously learn the generative process  and the analytical structure, JODI is likely to develop a richer, more robust latent representation of visual concepts.", "weaknesses": "(1) Lack of T2I evaluation. As demonstrated in Fig. 11, the generated image is the bridge between text and other condisions. Therefore, I would like to see some comparison with T2I methods on GenEval or other T2I benchmarks.\n\n(2) The proposed method achieves condition to image by modeling the joint distribution of various conditions. However, some generation tasks (e.g., depth to image [a]) could be evaluated in a more precise manner. The authors should compare, or at least, discuss with such related works. \n\n(3) Minor: In Tab. 1-openpose, controlnet seems to achieve better result than yours in terms of LPIPS.\n\n\n[a] 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. In ICLR'25.\n[b] 3DIS-FLUX: simple and efficient multi-instance generation with DiT rendering. In Arxiv.", "questions": "Please refer to \"weaknesses\"", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761477106717}, {"id": "Yk9mfPHrdz", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12035/Reviewer_Qnvb"], "rating": 2, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper proposes a joint modeling framework to train an all-in-one generative model that can do joint generation of images and caption, conditional image generation, and label prediction.", "review_text": "This paper proposes a joint modeling framework to train an all-in-one generative model that can do joint generation of images and caption, conditional image generation, and label prediction.", "strengths": "- I like the probabilistic perspective on joint modeling p(x, y) = p(x | y) p(y) = p(y | x) p(x). Although simple, it provides an intuitive motivation for why you would want to pursue such a method.\n- Visual results clearly show that your method works.\n- The paper is well-written and easy to follow.", "weaknesses": "see questions", "questions": "- Why would somebody want to use this instead of just chaining a few off-the-shelf models to do the target tasks? I get that there's some elegance to having it all in one model, but the pretrained models are already so powerful, that I'm skeptical that somebody would use this approach in practice. Furthermore, the results show that the specialist models often beat Jodi (although Jodi is better than the other joint models).\n- What is the methodological novelty?\n- Have you considered more modern fast attention mechanisms? [1] Or why not only train a few conditions at a time to get around the quadratic scaling?\n\n[1] Guo, Han, et al. \"Log-linear attention.\" arXiv preprint arXiv:2506.04761 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a joint modeling framework to train an all-in-one generative model that can do joint generation of images and caption, conditional image generation, and label prediction.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- I like the probabilistic perspective on joint modeling p(x, y) = p(x | y) p(y) = p(y | x) p(x). Although simple, it provides an intuitive motivation for why you would want to pursue such a method.\n- Visual results clearly show that your method works.\n- The paper is well-written and easy to follow.", "weaknesses": "see questions", "questions": "- Why would somebody want to use this instead of just chaining a few off-the-shelf models to do the target tasks? I get that there's some elegance to having it all in one model, but the pretrained models are already so powerful, that I'm skeptical that somebody would use this approach in practice. Furthermore, the results show that the specialist models often beat Jodi (although Jodi is better than the other joint models).\n- What is the methodological novelty?\n- Have you considered more modern fast attention mechanisms? [1] Or why not only train a few conditions at a time to get around the quadratic scaling?\n\n[1] Guo, Han, et al. \"Log-linear attention.\" arXiv preprint arXiv:2506.04761 (2025).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761213686106}], "openreview_url": "https://openreview.net/forum?id=Hd1Wciuyka", "arxiv_id": "2505.19084", "paper_pdf": "papers/Hd1Wciuyka.pdf", "paper_pdf_sha256": "7e113f1eb5e658bf8bbc3e34d4e2aaccd272d507edfd06b5760315be4b56ed16", "paper_pdf_bytes": 33196323, "paper_pdf_source": "openreview", "code_url": "https://github.com/VIPL-GENUN/Jodi", "code_repository": "VIPL-GENUN/Jodi", "code_commit": "c21bdcb8817284980f68321fa7375c5376c37d6d", "code_archive": "repos/Hd1Wciuyka.zip", "code_archive_sha256": "0ff0a1794f60f41eaadf7593950e0402bb0586fd822b1df416006c934d568acc", "code_archive_bytes": 8646852, "code_file_count": 72, "code_extensions": {".py": 70, ".sh": 2}, "github_disk_usage_kb": 8435, "github_languages": {"Python": 500890, "Shell": 870}, "github_archived": false, "github_pushed_at": "2026-03-06T02:46:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/jodi-unification-of-visual-generation-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EBBeSbmAyh", "year": 2025, "status": "rejected", "title": "Towards Constraint-aware Learning for Resource Allocation in NFV-enabled Networks", "authors": ["Tianfu Wang", "Long Yang", "Chao Wang", "Chuan Qin", "Liwei Deng", "Li Shen", "Hui Xiong"], "authorids": ["~Tianfu_Wang4", "~Long_Yang4", "~Chao_Wang14", "~Chuan_Qin1", "~Liwei_Deng2", "~Li_Shen1", "~Hui_Xiong1"], "authors_source": "OpenReview API", "abstract": "Virtual Network Embedding (VNE) is a challenging combinatorial optimization problem that refers to resource allocation associated with hard and multifaceted constraints in network function virtualization (NFV). Existing works for VNE struggle to handle such complex constraints, leading to compromised system performance and stability. In this paper, we propose a \\textbf{CON}straint-\\textbf{A}ware \\textbf{L}earning framework for VNE, named \\textbf{CONAL}, to achieve efficient constraint management. Concretely, we formulate the VNE problem as a constrained Markov decision process with violation tolerance. This modeling approach aims to improve both resource utilization and solution feasibility by precisely evaluating solution quality and the degree of constraint violation. We also propose a reachability-guided optimization with an adaptive reachability budget method that dynamically assigns budget values. This method achieves persistent zero violation to guarantee the feasibility of VNE solutions and more stable policy optimization by handling instances without any feasible solution. Furthermore, we propose a constraint-aware graph representation method to efficiently learn cross-graph relations and constrained path connectivity in VNE. Finally, extensive experimental results demonstrate the superiority of our proposed method over state-of-the-art baselines. Our code is available at \\href{https://anonymous.4open.science/r/iclr25-conal}{https://anonymous.4open.science/r/iclr25-conal}.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "cscLEq7Bq1", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5553/Reviewer_S9pS"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper proposes a new framework called constraint-Aware Learning (CONAL) to address the Virtual Network Embedding (VNE) problem in network virtualization. Specifically, the paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances.", "review_text": "The paper proposes a new framework called constraint-Aware Learning (CONAL) to address the Virtual Network Embedding (VNE) problem in network virtualization. Specifically, the paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances.", "strengths": "The paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances.", "weaknesses": "1. The paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances. However, when dealing with unsolvable instances, no policy can satisfy the constraints, the Lagrange multiplier λ may tend to infinity, leading to numerical instability during training. Instability may affect the policy's performance on solvable instances. Provide empirical evidence of the behavior of the Lagrange multiplier λ during training. Specifically, plot the variation of λ over training iterations or time to illustrate how it evolves, especially in the presence of unsolvable instances.\n2. The augmentation methods used in the path-bandwidth contrast module (physical link addition ϕA  and virtual link addition ϕB) lack sufficient theoretical and empirical justification. The choice of augmentation ratio ϵ significantly affects model performance, but the paper does not provide detailed analysis or guidelines for selecting these parameters. Provide theoretical explanations for how the augmentation methods contribute to improved bandwidth awareness. \n3. The integration of virtual and physical networks into a heterogeneous graph with numerous cross-graph links can lead to information redundancy and noise. Noise from irrelevant links can hinder the model's ability to learn meaningful representations.\n4. The experiments are mainly conducted on simulated environments and limited network topologies (e.g., GEANT and BRAIN). This may not adequately demonstrate the model's performance.\n5. Given the rapid development of NFV, some relevant literatures are missing to be discussed, e.g., NFVdeep: Adaptive Online Service Function Chain Deployment with Deep Reinforcement Learning, iwqos’19; Adaptive VNF Scaling and Flow Routing with Proactive Demand Prediction, infocom’18; FlexNFV: Flexible Network Service Chaining with Dynamic Scaling, network’19; Joint Optimization of Chain Placement and Request Scheduling for Network Function Virtualization, icdcs’17, etc.", "questions": "Overall，when infeasible instances exist in the Virtual Network Embedding (VNE) problem (i.e., there are no embedding solutions that satisfy all constraints), the optimization method employed in the paper causes the Lagrange multipliers (λ) to grow unbounded during training. The unbounded growth of λ leads to overflows or underflows in numerical calculation. The paper does not provide a robust method for detecting infeasible instances, nor does it implement any controls or mitigations for the growth of λ. This oversight means that, when faced with infeasible instances, the model may fail to operate correctly.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a new framework called constraint-Aware Learning (CONAL) to address the Virtual Network Embedding (VNE) problem in network virtualization. Specifically, the paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances.", "weaknesses": "1. The paper models the VNE problem as a violation-tolerant CMDP and introduces an adaptive reachability budget (ARB) to handle unsolvable instances. However, when dealing with unsolvable instances, no policy can satisfy the constraints, the Lagrange multiplier λ may tend to infinity, leading to numerical instability during training. Instability may affect the policy's performance on solvable instances. Provide empirical evidence of the behavior of the Lagrange multiplier λ during training. Specifically, plot the variation of λ over training iterations or time to illustrate how it evolves, especially in the presence of unsolvable instances.\n2. The augmentation methods used in the path-bandwidth contrast module (physical link addition ϕA  and virtual link addition ϕB) lack sufficient theoretical and empirical justification. The choice of augmentation ratio ϵ significantly affects model performance, but the paper does not provide detailed analysis or guidelines for selecting these parameters. Provide theoretical explanations for how the augmentation methods contribute to improved bandwidth awareness. \n3. The integration of virtual and physical networks into a heterogeneous graph with numerous cross-graph links can lead to information redundancy and noise. Noise from irrelevant links can hinder the model's ability to learn meaningful representations.\n4. The experiments are mainly conducted on simulated environments and limited network topologies (e.g., GEANT and BRAIN). This may not adequately demonstrate the model's performance.\n5. Given the rapid development of NFV, some relevant literatures are missing to be discussed, e.g., NFVdeep: Adaptive Online Service Function Chain Deployment with Deep Reinforcement Learning, iwqos’19; Adaptive VNF Scaling and Flow Routing with Proactive Demand Prediction, infocom’18; FlexNFV: Flexible Network Service Chaining with Dynamic Scaling, network’19; Joint Optimization of Chain Placement and Request Scheduling for Network Function Virtualization, icdcs’17, etc.", "questions": "Overall，when infeasible instances exist in the Virtual Network Embedding (VNE) problem (i.e., there are no embedding solutions that satisfy all constraints), the optimization method employed in the paper causes the Lagrange multipliers (λ) to grow unbounded during training. The unbounded growth of λ leads to overflows or underflows in numerical calculation. The paper does not provide a robust method for detecting infeasible instances, nor does it implement any controls or mitigations for the growth of λ. This oversight means that, when faced with infeasible instances, the model may fail to operate correctly.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730649803676}, {"id": "01iTTBe2IO", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5553/Reviewer_9yrG"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 3, "summary": "The paper tackles the VNE problem within NFV networks. Recognizing the limitations of existing solutions in handling intricate constraints and unsolvable instances, the authors propose a framework called Constraint-Aware Learning, formulates the VNE problem as a violation-tolerant constrained Markov Decision Process and introduces a reachability-guided optimization with adaptive reachability budgets. Additionally, the framework incorporates a constraint-aware graph representation method to capture cross-graph interactions and bandwidth-constrained path connectivity.", "review_text": "The paper tackles the VNE problem within NFV networks. Recognizing the limitations of existing solutions in handling intricate constraints and unsolvable instances, the authors propose a framework called Constraint-Aware Learning, formulates the VNE problem as a violation-tolerant constrained Markov Decision Process and introduces a reachability-guided optimization with adaptive reachability budgets. Additionally, the framework incorporates a constraint-aware graph representation method to capture cross-graph interactions and bandwidth-constrained path connectivity.", "strengths": "The paper is well-written and tackles a significant problem, offering practical implications for real-world network systems and potential applicability to other optimization challenges. The performance is benchmarked against several state-of-the-art baselines. Experiments are conducted across a wide range of network scenarios", "weaknesses": "The framework seems assumes a static PN setting. This assumption may not hold in highly dynamic network environments, such as mobile edge computing. \n\nThe focus is mainly on computing and bandwidth constraints. Other important factors, such as latency, reliability, and energy efficiency, are not addressed.", "questions": "1. How does the proposed method perform in dynamic network environments where the physical network topology and resource availabilities could change over time? \n\n2. Can you provide a more detailed analysis of the computational complexity of the proposed method, especially in comparison to baseline methods? \n\n3. Could you elaborate on the rationale behind using contrastive learning in the constraint-aware graph representation module?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper tackles the VNE problem within NFV networks. Recognizing the limitations of existing solutions in handling intricate constraints and unsolvable instances, the authors propose a framework called Constraint-Aware Learning, formulates the VNE problem as a violation-tolerant constrained Markov Decision Process and introduces a reachability-guided optimization with adaptive reachability budgets. Additionally, the framework incorporates a constraint-aware graph representation method to capture cross-graph interactions and bandwidth-constrained path connectivity.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "The paper is well-written and tackles a significant problem, offering practical implications for real-world network systems and potential applicability to other optimization challenges. The performance is benchmarked against several state-of-the-art baselines. Experiments are conducted across a wide range of network scenarios", "weaknesses": "The framework seems assumes a static PN setting. This assumption may not hold in highly dynamic network environments, such as mobile edge computing. \n\nThe focus is mainly on computing and bandwidth constraints. Other important factors, such as latency, reliability, and energy efficiency, are not addressed.", "questions": "1. How does the proposed method perform in dynamic network environments where the physical network topology and resource availabilities could change over time? \n\n2. Can you provide a more detailed analysis of the computational complexity of the proposed method, especially in comparison to baseline methods? \n\n3. Could you elaborate on the rationale behind using contrastive learning in the constraint-aware graph representation module?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730594448281}, {"id": "8sujUx1toR", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5553/Reviewer_iEV6"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper proposes a solution based on a Constrained Markov Decision Process for resource allocation in NFV-enabled networks. The problem of resource allocation in those networks (called Virtual Network Embedding in the related literature) is well-known in the research community and several solutions for it have already been proposed. Anyway, the proposed solution is sufficiently original and shows to achieve good performance results if compared with the primary baselines already existing in the literature.\nHowever, the paper is weak in terms of in-depth technical insights about how to efficiently implement the proposed solution, of insufficient experimental evaluation and validation, and of potential impact in the field (see the following parts of this review form).", "review_text": "The paper proposes a solution based on a Constrained Markov Decision Process for resource allocation in NFV-enabled networks. The problem of resource allocation in those networks (called Virtual Network Embedding in the related literature) is well-known in the research community and several solutions for it have already been proposed. Anyway, the proposed solution is sufficiently original and shows to achieve good performance results if compared with the primary baselines already existing in the literature.\nHowever, the paper is weak in terms of in-depth technical insights about how to efficiently implement the proposed solution, of insufficient experimental evaluation and validation, and of potential impact in the field (see the following parts of this review form).", "strengths": "- The addressed topic is interesting and relevant, even if already well investigated in the related literature\n- The proposed problem formulation and the deriving algorithmic solution are technically sound and do not exhibit big technical flaws\n- The reported performance results are interesting and show that the proposed solution can outperform several related baselines in the existing literature\n- The paper is generally well organized and well written", "weaknesses": "- The VNE problem has been investigated several times in the related literature. To be impactful, there is the need that novel solutions in the field do not propose only an algorithmic solution but also the design and implementation of a prototype integrated into real cloud/edge deployment environments. Otherwise, the level of technical originality and relevance could only be limited, given the status of maturity of the research field\n- The paper does not include in-depth technical insights about how to exactly achieve an effective and efficient design/implementation of the proposed solution into a real prototype. No lessons learned from the experience of real deployment and evaluation in in-the-field deployment scenarios\n- No systems engineering considerations and lessons learned about how to optimally configure and deploy the proposed solution\n- The reported performance results are obtained by adopting simulation assumptions that are not realistic for many real deployment environments. I can understand that other papers in the literature have adopted a similar approach, but this is too simplistic. At least the validity of the used assumptions should be better justified and motivated in the paper. In addition, why not using real traces from real deployment environments, in particular for request demands?\n- Even if the paper is generally well organized and well written, a few writing inaccuracies are still present in the manuscript and call for some minor revision work in order to improve the paper presentation style. Only to mention one example: \"Addtional\" in page 24.", "questions": "Please see the previous parts of this review form, in particular the weaknesses part above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a solution based on a Constrained Markov Decision Process for resource allocation in NFV-enabled networks. The problem of resource allocation in those networks (called Virtual Network Embedding in the related literature) is well-known in the research community and several solutions for it have already been proposed. Anyway, the proposed solution is sufficiently original and shows to achieve good performance results if compared with the primary baselines already existing in the literature.\nHowever, the paper is weak in terms of in-depth technical insights about how to efficiently implement the proposed solution, of insufficient experimental evaluation and validation, and of potential impact in the field (see the following parts of this review form).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The addressed topic is interesting and relevant, even if already well investigated in the related literature\n- The proposed problem formulation and the deriving algorithmic solution are technically sound and do not exhibit big technical flaws\n- The reported performance results are interesting and show that the proposed solution can outperform several related baselines in the existing literature\n- The paper is generally well organized and well written", "weaknesses": "- The VNE problem has been investigated several times in the related literature. To be impactful, there is the need that novel solutions in the field do not propose only an algorithmic solution but also the design and implementation of a prototype integrated into real cloud/edge deployment environments. Otherwise, the level of technical originality and relevance could only be limited, given the status of maturity of the research field\n- The paper does not include in-depth technical insights about how to exactly achieve an effective and efficient design/implementation of the proposed solution into a real prototype. No lessons learned from the experience of real deployment and evaluation in in-the-field deployment scenarios\n- No systems engineering considerations and lessons learned about how to optimally configure and deploy the proposed solution\n- The reported performance results are obtained by adopting simulation assumptions that are not realistic for many real deployment environments. I can understand that other papers in the literature have adopted a similar approach, but this is too simplistic. At least the validity of the used assumptions should be better justified and motivated in the paper. In addition, why not using real traces from real deployment environments, in particular for request demands?\n- Even if the paper is generally well organized and well written, a few writing inaccuracies are still present in the manuscript and call for some minor revision work in order to improve the paper presentation style. Only to mention one example: \"Addtional\" in page 24.", "questions": "Please see the previous parts of this review form, in particular the weaknesses part above.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A.", "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730360965838}, {"id": "EJAhHIj0bQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5553/Reviewer_z4vG"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 5, "summary": "This paper presents a Constraint-aware Network Abstraction Layer (CONAL) tailored for Virtual Network Embedding (VNE) to advance constraint management and improve training robustness, key factors for optimizing network system performance and reliability. By framing VNE as a violation-tolerant Constrained Markov Decision Process (CMDP), the authors aim to enhance solution quality and feasibility, ensuring complete solutions that accurately assess solution quality. The paper introduces a reachability-guided objective, paired with an adaptive feasibility budget method, to guarantee ongoing constraint satisfaction while reducing policy conservativeness and stabilizing policy optimization even with unsolvable instances. To address the complexity of VNE constraints, a constraint-aware graph representation is proposed, featuring a heterogeneous modeling module to capture cross-graph relationships and a path-bandwidth contrast module for heightened sensitivity to bandwidth constraints.", "review_text": "This paper presents a Constraint-aware Network Abstraction Layer (CONAL) tailored for Virtual Network Embedding (VNE) to advance constraint management and improve training robustness, key factors for optimizing network system performance and reliability. By framing VNE as a violation-tolerant Constrained Markov Decision Process (CMDP), the authors aim to enhance solution quality and feasibility, ensuring complete solutions that accurately assess solution quality. The paper introduces a reachability-guided objective, paired with an adaptive feasibility budget method, to guarantee ongoing constraint satisfaction while reducing policy conservativeness and stabilizing policy optimization even with unsolvable instances. To address the complexity of VNE constraints, a constraint-aware graph representation is proposed, featuring a heterogeneous modeling module to capture cross-graph relationships and a path-bandwidth contrast module for heightened sensitivity to bandwidth constraints.", "strengths": "The authors propose a violation-tolerant Constrained Markov Decision Process (CMDP) modeling approach, which effectively evaluates solution quality and constraint violation levels, thereby enhancing solution feasibility and resource utilization efficiency.", "weaknesses": "- There has been some work examining about RL and NFV [1,2] and various approaches have been devised for solving the constraint violations therein, and it is not clear where this paper excels in relation to them.\n  [1] Gu L, Zeng D, Li W, et al. Intelligent VNF orchestration and flow scheduling via model-assisted deep reinforcement learning[J]. IEEE Journal on Selected Areas in Communications, 2019, 38(2): 279-291.\n  [2] Zeng Y, Qu Z, Guo S, et al. SafeDRL: Dynamic Microservice Provisioning With Reliability and Latency Guarantees in Edge Environments[J]. IEEE Transactions on Computers, 2023.\n- Why applying DRL to cope with VNE is unclear, where existing heuristics do not always face high time overhead and it is not clear for what specific problems face what specific limited performance and why?\n- The fact that real-world system validation is not based on real-world system implementations but is still based on simulation should not be blown out of proportion.\n- The author claims that reinforcement learning learns effective strategies from unlabeled datasets, however, reinforcement learning actually learns strategies through interaction with the environment.\n- Is constraint violation really acceptable for VNE ? Is it reasonable that constraint violations are allowed in the designed solution?", "questions": "Compare their work with more existing studies on solving constraint violations in applying RL to NFV （not just the papers mentioned above）, and compare it with them in experiments and related works.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a Constraint-aware Network Abstraction Layer (CONAL) tailored for Virtual Network Embedding (VNE) to advance constraint management and improve training robustness, key factors for optimizing network system performance and reliability. By framing VNE as a violation-tolerant Constrained Markov Decision Process (CMDP), the authors aim to enhance solution quality and feasibility, ensuring complete solutions that accurately assess solution quality. The paper introduces a reachability-guided objective, paired with an adaptive feasibility budget method, to guarantee ongoing constraint satisfaction while reducing policy conservativeness and stabilizing policy optimization even with unsolvable instances. To address the complexity of VNE constraints, a constraint-aware graph representation is proposed, featuring a heterogeneous modeling module to capture cross-graph relationships and a path-bandwidth contrast module for heightened sensitivity to bandwidth constraints.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The authors propose a violation-tolerant Constrained Markov Decision Process (CMDP) modeling approach, which effectively evaluates solution quality and constraint violation levels, thereby enhancing solution feasibility and resource utilization efficiency.", "weaknesses": "- There has been some work examining about RL and NFV [1,2] and various approaches have been devised for solving the constraint violations therein, and it is not clear where this paper excels in relation to them.\n  [1] Gu L, Zeng D, Li W, et al. Intelligent VNF orchestration and flow scheduling via model-assisted deep reinforcement learning[J]. IEEE Journal on Selected Areas in Communications, 2019, 38(2): 279-291.\n  [2] Zeng Y, Qu Z, Guo S, et al. SafeDRL: Dynamic Microservice Provisioning With Reliability and Latency Guarantees in Edge Environments[J]. IEEE Transactions on Computers, 2023.\n- Why applying DRL to cope with VNE is unclear, where existing heuristics do not always face high time overhead and it is not clear for what specific problems face what specific limited performance and why?\n- The fact that real-world system validation is not based on real-world system implementations but is still based on simulation should not be blown out of proportion.\n- The author claims that reinforcement learning learns effective strategies from unlabeled datasets, however, reinforcement learning actually learns strategies through interaction with the environment.\n- Is constraint violation really acceptable for VNE ? Is it reasonable that constraint violations are allowed in the designed solution?", "questions": "Compare their work with more existing studies on solving constraint violations in applying RL to NFV （not just the papers mentioned above）, and compare it with them in experiments and related works.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730114361855}], "openreview_url": "https://openreview.net/forum?id=EBBeSbmAyh", "arxiv_id": "2410.22999", "paper_pdf": "papers/EBBeSbmAyh.pdf", "paper_pdf_sha256": "1fd5a143bda9078bf1b4442865b055aff1e04de08fe1c354dca8382bbb389656", "paper_pdf_bytes": 1182935, "paper_pdf_source": "openreview", "code_url": "https://github.com/GeminiLight/conal-vne", "code_repository": "GeminiLight/conal-vne", "code_commit": "768b322f314d4d33f9ceddeb2a27bf031ba5ee10", "code_archive": "repos/EBBeSbmAyh.zip", "code_archive_sha256": "f3ecb0c7c22d5fa24e82db96834d40bc827cd4fb8dcb4ecd047766945181edaf", "code_archive_bytes": 435026, "code_file_count": 98, "code_extensions": {".py": 91, ".sh": 7}, "github_disk_usage_kb": 390, "github_languages": {"Python": 729337, "Shell": 28174}, "github_archived": false, "github_pushed_at": "2024-10-11T07:01:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-constraint-aware-learning-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xsts7MRLey", "year": 2024, "status": "rejected", "title": "DEEP UNSUPERVISED DOMAIN ADAPTATION FOR TIME SERIES CLASSIFICATION: A BENCHMARK", "authors": ["Hassan Ismail Fawaz", "Ganesh Del Grosso", "Tanguy Kerdoncuff", "Aurelie Boisbunon", "Illyyne Saffar"], "authorids": ["~Hassan_Ismail_Fawaz1", "~Ganesh_Del_Grosso1", "~Tanguy_Kerdoncuff1", "~Aurelie_Boisbunon1", "~Illyyne_Saffar1"], "authors_source": "OpenReview API", "abstract": "Unsupervised Domain Adaptation (UDA) aims to harness labeled source data to train models for unlabeled target data. Despite extensive research in domains like computer vision and natural language processing, UDA remains underexplored for time series data, which has widespread real-world applications ranging from medicine and manufacturing to earth observation and human activity recognition. Our paper addresses this gap by introducing a comprehensive benchmark for evaluating UDA techniques for time series classification, with a focus on deep learning methods. We provide seven new benchmark datasets covering various domain shifts and temporal dynamics, facilitating fair and standardized UDA method assessments with state of the art neural network backbones (e.g. Inception) for time series data. This benchmark offers insights into the strengths and limitations of the evaluated approaches while preserving the unsupervised nature of domain adaptation, making it directly applicable to practical problems. Our paper serves as a vital resource for researchers and practitioners, advancing domain adaptation solutions for time series data and fostering innovation in this critical field.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "uWnBkqUE5l", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1464/Reviewer_KuTn"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper presents a thorough benchmarking study on time-series unsupervised domain adaptation, primarily focusing on deep learning techniques. It examines the impact of model backbones and hyperparameter tuning approaches. Furthermore, the authors evaluate various existing unsupervised domain adaptation methods across multiple domains, including seven new benchmark datasets.", "review_text": "This paper presents a thorough benchmarking study on time-series unsupervised domain adaptation, primarily focusing on deep learning techniques. It examines the impact of model backbones and hyperparameter tuning approaches. Furthermore, the authors evaluate various existing unsupervised domain adaptation methods across multiple domains, including seven new benchmark datasets.", "strengths": "1. The study delivers a detailed benchmark on unsupervised domain adaptation for time-series data, delving into the effect of domain adaptation algorithms, model backbones, and hyperparameter tuning strategies.\n2. The paper evaluates a range of unsupervised domain adaptation methods on datasets from diverse domains, including seven newly introduced datasets and existing benchmarks.", "weaknesses": "1. The discussion on the effect of model backbone in the paper is limited primarily to the Inception model. A broader examination involving diverse backbone models is crucial to substantiate the claim that \"backbones do not have a significant impact\".\n2. More discussions on different types of unsupervised domain adaptation methods would be beneficial. Specifically, it would be informative to explore under what specific conditions certain domain adaptation approaches may outperform others.\n3. Additional discussions regarding the choice of model backbones is helpful too. For example, I am curious if Inception is the best model backbone over all domains, or we need different model backbones for different time-series domains and data characteristics.\n4. Figure 2 (b) and 5 should be merged since they lead to similar findings.", "questions": "See Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a thorough benchmarking study on time-series unsupervised domain adaptation, primarily focusing on deep learning techniques. It examines the impact of model backbones and hyperparameter tuning approaches. Furthermore, the authors evaluate various existing unsupervised domain adaptation methods across multiple domains, including seven new benchmark datasets.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The study delivers a detailed benchmark on unsupervised domain adaptation for time-series data, delving into the effect of domain adaptation algorithms, model backbones, and hyperparameter tuning strategies.\n2. The paper evaluates a range of unsupervised domain adaptation methods on datasets from diverse domains, including seven newly introduced datasets and existing benchmarks.", "weaknesses": "1. The discussion on the effect of model backbone in the paper is limited primarily to the Inception model. A broader examination involving diverse backbone models is crucial to substantiate the claim that \"backbones do not have a significant impact\".\n2. More discussions on different types of unsupervised domain adaptation methods would be beneficial. Specifically, it would be informative to explore under what specific conditions certain domain adaptation approaches may outperform others.\n3. Additional discussions regarding the choice of model backbones is helpful too. For example, I am curious if Inception is the best model backbone over all domains, or we need different model backbones for different time-series domains and data characteristics.\n4. Figure 2 (b) and 5 should be merged since they lead to similar findings.", "questions": "See Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698810719114}, {"id": "lqVdnkGAgt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1464/Reviewer_hmFL"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "In this paper, the authors present benchmark research on deep unsupervised domain adaptation (UDA) for time series classification (TSC). Specifically, seven new datasets are introduced for this TSC UDA task, and experiments of several existing TSC UDA baselines are tested on these datasets.", "review_text": "In this paper, the authors present benchmark research on deep unsupervised domain adaptation (UDA) for time series classification (TSC). Specifically, seven new datasets are introduced for this TSC UDA task, and experiments of several existing TSC UDA baselines are tested on these datasets.", "strengths": "Strength:\n\n1.\tThe paper introduces 7 new datasets for TSC UDA task.\n\n2.\tThe paper conduct experiments on several existing UDA baselines on the new datasets.\n\n3.\tThe paper has potential to be a benchmark for the following TSC UDA research.", "weaknesses": "Weakness:\n\n1.\tThe major concern of the work is on the technical novelty. All the datasets and baselines (including hyper-parameter tuning methods) are from existing literatures, and there is no novel technical contribution proposed. \n\n2.\tFor UDA TSC, some important related works are missing, for instance (to name a few), unsupervised video domain adaptation [ref1], transfer gaussian process [ref2], and time-series domain adaptation [ref3]. The authors may need to present a more comprehensive related work section to discuss more related works. \n\n3.\tMore analyses on the new datasets are expected, for instance, the domain discrepancy analyses (both marginal and conditional can be involved) on different domain pairs. From the experiments results of the source only baseline, the domain discrepancy differs considerably among different domain pairs, see, Table 5 domain 0 and domain 3, Table 9 domain 9 and domain 18. \n\n4.\tMore analyses on the comparison results are also expected. For instance, analyzing why some baselines achieve positive transfer on some tasks but negative transfer on others, e.g., see OTDA/VRADA in table 9. It would be more interesting to see constructive insights or conclusions that can benefit the community. \n\n[ref1] Video Unsupervised Domain Adaptation with Deep Learning\n\n[ref2] Adaptive Transfer Kernel Learning for Transfer Gaussian Process Regression\n\n[ref3] Time Series Domain Adaptation via Sparse Associative Structure Alignment", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors present benchmark research on deep unsupervised domain adaptation (UDA) for time series classification (TSC). Specifically, seven new datasets are introduced for this TSC UDA task, and experiments of several existing TSC UDA baselines are tested on these datasets.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "Strength:\n\n1.\tThe paper introduces 7 new datasets for TSC UDA task.\n\n2.\tThe paper conduct experiments on several existing UDA baselines on the new datasets.\n\n3.\tThe paper has potential to be a benchmark for the following TSC UDA research.", "weaknesses": "Weakness:\n\n1.\tThe major concern of the work is on the technical novelty. All the datasets and baselines (including hyper-parameter tuning methods) are from existing literatures, and there is no novel technical contribution proposed. \n\n2.\tFor UDA TSC, some important related works are missing, for instance (to name a few), unsupervised video domain adaptation [ref1], transfer gaussian process [ref2], and time-series domain adaptation [ref3]. The authors may need to present a more comprehensive related work section to discuss more related works. \n\n3.\tMore analyses on the new datasets are expected, for instance, the domain discrepancy analyses (both marginal and conditional can be involved) on different domain pairs. From the experiments results of the source only baseline, the domain discrepancy differs considerably among different domain pairs, see, Table 5 domain 0 and domain 3, Table 9 domain 9 and domain 18. \n\n4.\tMore analyses on the comparison results are also expected. For instance, analyzing why some baselines achieve positive transfer on some tasks but negative transfer on others, e.g., see OTDA/VRADA in table 9. It would be more interesting to see constructive insights or conclusions that can benefit the community. \n\n[ref1] Video Unsupervised Domain Adaptation with Deep Learning\n\n[ref2] Adaptive Transfer Kernel Learning for Transfer Gaussian Process Regression\n\n[ref3] Time Series Domain Adaptation via Sparse Associative Structure Alignment", "questions": "Please refer to the weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698635943999}, {"id": "Z91NkkQr3P", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1464/Reviewer_wzDS"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work explores the use of unsupervised domain adaptation (UDA) for time series classification (TSC), with a particular focus on deep learning methods. In UDA, which has been extensively explored in vision and natural language applications, two domains of data exist: a labelled source domain and an unlabelled target domain that has some form of shift in the time series data (e.g. differences in data used for training and data used during deployment). The objective is to leverage the labelled source data to make predicts in the target domain.\n\nIn addition to five existing datasets, this work proposes the use of seven new datasets for UDA TSC (taken from existing sources). This collection of datasets serves as a benchmarking evaluation tool for assessing the efficacy of different UDA TSC deep learning approaches, notably with different algorithms, hyperparameter optimisation approaches, and model backbones. Consistent experimentation is used to compare the performance of these different approaches and make observations of which elements contribute the most to performance.", "review_text": "This work explores the use of unsupervised domain adaptation (UDA) for time series classification (TSC), with a particular focus on deep learning methods. In UDA, which has been extensively explored in vision and natural language applications, two domains of data exist: a labelled source domain and an unlabelled target domain that has some form of shift in the time series data (e.g. differences in data used for training and data used during deployment). The objective is to leverage the labelled source data to make predicts in the target domain.\n\nIn addition to five existing datasets, this work proposes the use of seven new datasets for UDA TSC (taken from existing sources). This collection of datasets serves as a benchmarking evaluation tool for assessing the efficacy of different UDA TSC deep learning approaches, notably with different algorithms, hyperparameter optimisation approaches, and model backbones. Consistent experimentation is used to compare the performance of these different approaches and make observations of which elements contribute the most to performance.", "strengths": "**Originality**  \nO1. The main novelty of the work lies in the proposal of additional datasets and a consistent, fair framework for evaluation the UDA TSC methods. This also extends to the insights that can be drawn from this evaluation.  \nO2. There is also some originality in the deep UDA methods that are used, notably using consistent a consistent backbone (InceptionTime) across different approaches.  \n\n**Quality**  \nQ1. The experimental setup is well-structured and makes steps to ensure fairness across all algorithms (e.g. limiting GPU time for training/hyperopt).  \nQ2. Results analysis provides some comparisons between the choice of different classifiers, hyperopt methods, and backbones.  \n\n**Clarity**  \nC1. Clear descriptions of all the different elements of the experiments are given (models, hyperopt methods, datasets, and pipelines).  \nC2. Figures are communicative and support conclusions drawn from the work.\n\n**Significance**  \nS1. This work could serve as a stable baseline for further developing UDA TSC deep learning approaches, helping to progress the area of research.  \nS2. Insights into the performance of different methods (e.g. InceptionRain seemingly being the strongest method) is useful for establishing the current SOTA and assessing the relative performance.", "weaknesses": "**Presentation of Results**  \nP1. While Figure 1 compares model performance within hyperopt methods (a, b, c), it does not provide an overall comparison of all models with all hyperopt methods. As such, it is difficult to determine which complete approach (model + tuning approach) actually has the best performance. An additional critical difference diagram comparing the (5?) top methods for each tuning approach would make this much clearer.  \nP2. While the figures provide some information, and the Appendix gives a full set of results for each dataset, it remains difficult to assess the margins between the approaches. A summary table of average accuracy across all datasets for each experimental configuration would be beneficial in conveying this information.  \nP3. Further variations of Figure 4 for other models/datasets would be useful to see if the revealed trend is consistent.  \nP4. I think the violin plots in Figure 7 are a strong way of communicating the results, and potentially should be moved to the main body if possible. As mentioned above, combining the some selection of the top methods for each tuning approach into a single plot would further aid comparison.  \n\n**Significance**  \nS1. I believe this work has the most potential if the evaluation is released to allow further development of methods. I appreciate the source code is planned to be released upon acceptance, but potentially taking this a step further and allowing for easy reproducibility/extensibility would improve the impact of the work and help progress the research area.", "questions": "1. To what extent is there dataset imbalance in the datasets? Additional results, for example using balanced accuracy, may be warranted if dataset imbalanced is high. At the very least, a discussion on any dataset imbalances would be helpful. I appreciate F1 score results are given in the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work explores the use of unsupervised domain adaptation (UDA) for time series classification (TSC), with a particular focus on deep learning methods. In UDA, which has been extensively explored in vision and natural language applications, two domains of data exist: a labelled source domain and an unlabelled target domain that has some form of shift in the time series data (e.g. differences in data used for training and data used during deployment). The objective is to leverage the labelled source data to make predicts in the target domain.\n\nIn addition to five existing datasets, this work proposes the use of seven new datasets for UDA TSC (taken from existing sources). This collection of datasets serves as a benchmarking evaluation tool for assessing the efficacy of different UDA TSC deep learning approaches, notably with different algorithms, hyperparameter optimisation approaches, and model backbones. Consistent experimentation is used to compare the performance of these different approaches and make observations of which elements contribute the most to performance.", "soundness": "3 good", "presentation": "3 good", "contribution": "4 excellent", "strengths": "**Originality**  \nO1. The main novelty of the work lies in the proposal of additional datasets and a consistent, fair framework for evaluation the UDA TSC methods. This also extends to the insights that can be drawn from this evaluation.  \nO2. There is also some originality in the deep UDA methods that are used, notably using consistent a consistent backbone (InceptionTime) across different approaches.  \n\n**Quality**  \nQ1. The experimental setup is well-structured and makes steps to ensure fairness across all algorithms (e.g. limiting GPU time for training/hyperopt).  \nQ2. Results analysis provides some comparisons between the choice of different classifiers, hyperopt methods, and backbones.  \n\n**Clarity**  \nC1. Clear descriptions of all the different elements of the experiments are given (models, hyperopt methods, datasets, and pipelines).  \nC2. Figures are communicative and support conclusions drawn from the work.\n\n**Significance**  \nS1. This work could serve as a stable baseline for further developing UDA TSC deep learning approaches, helping to progress the area of research.  \nS2. Insights into the performance of different methods (e.g. InceptionRain seemingly being the strongest method) is useful for establishing the current SOTA and assessing the relative performance.", "weaknesses": "**Presentation of Results**  \nP1. While Figure 1 compares model performance within hyperopt methods (a, b, c), it does not provide an overall comparison of all models with all hyperopt methods. As such, it is difficult to determine which complete approach (model + tuning approach) actually has the best performance. An additional critical difference diagram comparing the (5?) top methods for each tuning approach would make this much clearer.  \nP2. While the figures provide some information, and the Appendix gives a full set of results for each dataset, it remains difficult to assess the margins between the approaches. A summary table of average accuracy across all datasets for each experimental configuration would be beneficial in conveying this information.  \nP3. Further variations of Figure 4 for other models/datasets would be useful to see if the revealed trend is consistent.  \nP4. I think the violin plots in Figure 7 are a strong way of communicating the results, and potentially should be moved to the main body if possible. As mentioned above, combining the some selection of the top methods for each tuning approach into a single plot would further aid comparison.  \n\n**Significance**  \nS1. I believe this work has the most potential if the evaluation is released to allow further development of methods. I appreciate the source code is planned to be released upon acceptance, but potentially taking this a step further and allowing for easy reproducibility/extensibility would improve the impact of the work and help progress the research area.", "questions": "1. To what extent is there dataset imbalance in the datasets? Additional results, for example using balanced accuracy, may be warranted if dataset imbalanced is high. At the very least, a discussion on any dataset imbalances would be helpful. I appreciate F1 score results are given in the appendix.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698313415353}], "openreview_url": "https://openreview.net/forum?id=xsts7MRLey", "arxiv_id": "2312.09857", "paper_pdf": "papers/xsts7MRLey.pdf", "paper_pdf_sha256": "a57654ccc36c0d8d4e3c007c2876ba573824fefac66092b0bc31a220e2cd76ba", "paper_pdf_bytes": 2251960, "paper_pdf_source": "openreview", "code_url": "https://github.com/EricssonResearch/UDA-4-TSC", "code_repository": "EricssonResearch/UDA-4-TSC", "code_commit": "c402cfc94673b8436e3dde4849342fded29e36a0", "code_archive": "repos/xsts7MRLey.zip", "code_archive_sha256": "509c9b272f11bd3773db2011aaec39b8e34185a86d7cca06819842b732a0ca78", "code_archive_bytes": 349475, "code_file_count": 90, "code_extensions": {".py": 89, ".sh": 1}, "github_disk_usage_kb": 374, "github_languages": {"Python": 349020, "Shell": 158}, "github_archived": false, "github_pushed_at": "2023-12-19T09:08:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/deep-unsupervised-domain-adaptation-for-time"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mPxsHDgsimT", "year": 2023, "status": "rejected", "title": "Subclass-balancing Contrastive Learning for Long-tailed Recognition", "authors": ["Chengkai Hou", "Jieyu Zhang", "Haonan Wang", "Tianyi Zhou"], "authorids": ["~Chengkai_Hou1", "~Jieyu_Zhang1", "~Haonan_Wang1", "~Tianyi_Zhou1"], "authors_source": "OpenReview API", "abstract": "Long-tailed recognition with imbalanced classes naturally emerges in practical machine learning applications. Existing methods such as data reweighing, resampling, and supervised contrastive learning enforce the class balance with a price of introducing imbalance between instances of head class and tail class, which may ignore the underlying rich semantic substructures of the former and exaggerate the biases in the latter. We overcome these drawbacks by a novel \"subclass-balancing contrastive learning (SBCL)'' approach that clusters each head class into multiple subclasses of similar sizes as the tail classes and enforce representations to capture the two-layer class hierarchy between the original classes and their subclasses. Since the clustering is conducted in the representation space and updated during the course of training, the subclass labels preserve the semantic substructures of head classes. Meanwhile, it does not overemphasize tail class samples so each individual instance contribute to the representation learning equally. Hence, our method achieves both the instance- and subclass-balance, while the original class labels are also learned through contrastive learning among subclasses from different classes. We evaluate SBCL over a list of long-tailed benchmark datasets and it achieves the state-of-the-art performance. In addition, we present extensive analyses and ablation studies of SBCL to verify its advantages. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "o5_ZNPjEMd5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1009/Reviewer_8khH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a method for contrastive learning on a long-tailed training dataset. The proposed method is called Subclass Balancing Contrastive Learner (SBCL), which uses clustering within the head classes to determine fine-grained subclasses of size similar to the tail clusters. The new subclass labels for the head instances are used w/ tail classes to perform supervised contrastive learning. The paper contains extensive experiments over several vision benchmarks, including classification, object recognition and instance segmentation tasks. In almost all cases, SBCL performs better than several other baselines.\n", "review_text": "(Please see the strengths and weaknesses above.) I would appreciate if the authors respond to some of my questions and concerns. ", "strengths": "**Strengths**\n\nS1: Extensive set of experiments on several vision tasks show SBCL does better than other representation learning methods for learning on long-tail distribution. Many ablations were also included to show several design choices made. \n\nS2: I think the idea of clustering the head class(es) into sub-classes and using the cluster ids as labels is novel.\n\n**Weaknesses**\n\nW1: While there are extensive sets of experiments, the quantitative results (Table 1-5) do not come with standard errors/deviations. The experiments were done w/ 5 random seeds, then I wonder why have the standard errors not been included in the table? Understanding how reliable these improvements seems like an important thing to report.\n\nW2: There is no discussion on the convergence of the model, and the number of training steps/epochs, loss trajectory or learning rate used. Since SBCL uses warm-up and adaptive clustering which may lead to a jump in the loss fn, studying these properties seems highly relevant. It is unclear if the number of training steps/epoch used for the training the feature extractor is the same for all the baselines and the proposed SBCL.\n\nOther weakness/questions:\n- I wonder why the disjoint subset analysis was not shown for the CIFAR-100 LT dataset?\n- The conclusion from the analysis in section 4.4 lacks clarity. What is the conclusion from this analysis? What is the significance of the finding: “the intra-class distance of SBCL is almost invariant with the decreasing of group split”. Why the increase in the intra-class distance for KCL & TSC is attributed to the head class? The discussion in this section needs some work.\n- Was the negative sampling bias corrected when computing the loss function? While SBCL balances the label distribution (via cluster labels for head instances), it is unclear that this would mitigate the sampling bias in selecting negatives. Typically, models use log(P(x)) correction [1] when using in-batch softmax loss to prevent the head instances to be overly treated as negatives than the tail instances.Having a baselines that uses this logit adjustment would be a definite plus.\n\n[1] Menon et al. (2021), Long-tail learning via logit adjustment, (ICLR 2021).\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a method for contrastive learning on a long-tailed training dataset. The proposed method is called Subclass Balancing Contrastive Learner (SBCL), which uses clustering within the head classes to determine fine-grained subclasses of size similar to the tail clusters. The new subclass labels for the head instances are used w/ tail classes to perform supervised contrastive learning. The paper contains extensive experiments over several vision benchmarks, including classification, object recognition and instance segmentation tasks. In almost all cases, SBCL performs better than several other baselines.\n", "strength_and_weaknesses": "**Strengths**\n\nS1: Extensive set of experiments on several vision tasks show SBCL does better than other representation learning methods for learning on long-tail distribution. Many ablations were also included to show several design choices made. \n\nS2: I think the idea of clustering the head class(es) into sub-classes and using the cluster ids as labels is novel.\n\n**Weaknesses**\n\nW1: While there are extensive sets of experiments, the quantitative results (Table 1-5) do not come with standard errors/deviations. The experiments were done w/ 5 random seeds, then I wonder why have the standard errors not been included in the table? Understanding how reliable these improvements seems like an important thing to report.\n\nW2: There is no discussion on the convergence of the model, and the number of training steps/epochs, loss trajectory or learning rate used. Since SBCL uses warm-up and adaptive clustering which may lead to a jump in the loss fn, studying these properties seems highly relevant. It is unclear if the number of training steps/epoch used for the training the feature extractor is the same for all the baselines and the proposed SBCL.\n\nOther weakness/questions:\n- I wonder why the disjoint subset analysis was not shown for the CIFAR-100 LT dataset?\n- The conclusion from the analysis in section 4.4 lacks clarity. What is the conclusion from this analysis? What is the significance of the finding: “the intra-class distance of SBCL is almost invariant with the decreasing of group split”. Why the increase in the intra-class distance for KCL & TSC is attributed to the head class? The discussion in this section needs some work.\n- Was the negative sampling bias corrected when computing the loss function? While SBCL balances the label distribution (via cluster labels for head instances), it is unclear that this would mitigate the sampling bias in selecting negatives. Typically, models use log(P(x)) correction [1] when using in-batch softmax loss to prevent the head instances to be overly treated as negatives than the tail instances.Having a baselines that uses this logit adjustment would be a definite plus.\n\n[1] Menon et al. (2021), Long-tail learning via logit adjustment, (ICLR 2021).\n", "clarity,_quality,_novelty_and_reproducibility": "**Clarity and Novelty**\nThe paper is clearly written and seems novel.\n\n**Reproducibility**\nI am not sure if the paper has sufficient training details to confirm reproducibility. However, it is mentioned that the training process is same as TSC (Li et al., 2022).", "summary_of_the_review": "(Please see the strengths and weaknesses above.) I would appreciate if the authors respond to some of my questions and concerns. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667024816701}, {"id": "abiTdSRnVLW", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1009/Reviewer_QXy7"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a new method called subclass balancing contrastive learning for long-tailed problems. The key idea is to divide head classes into subclasses to balance the number of instances in each subclass. The proposed method is evaluated on widely used image classification datasets as well as some additional vision tasks such as object detection and semantic segmentation. ", "review_text": "I think the motivation of the paper is sound and the proposed method is somehow reasonable. However, I have doubt on the effectiveness of the proposed method since I do not think the current experiment is not sufficient for supporting it. As such I do not think this paper is ready for publication. I would like to hear the feedback from the authors on the points I listed in the weakness. \n\n### after rebuttal\nI became more positive about the paper through the rebuttal and I increased my score accordingly, but I still have some concerns on the effectiveness of the proposed method.", "strengths": "## Strength\n1. The proposed method is evaluated on wide variety of datasets. Especially the evaluation on other tasks than image classification, i.e., object detection and semantic segmentation, is given.\n1. The effectiveness of the proposed method is shown in the experiment at least to some extent.\n1. The paper generally reads well\n\n## Weakness\n1. The comparison with stronger baselines such as [A1-3] is missing. The reported accuracies in these works are actually higher than the proposed method.\n1. According to Table 4, the performance of the proposed method on CIFAR100-LT becomes 43.7 if the dynamic temperature is not applied. This performance is lower than TSC. I am afraid this suggests that it is the dynamic temperature that makes the paper perform better than other contrastive learning methods. To show the effectiveness of the proposed idea, it is necessary to show the results of other contrastive learning methods such as SCL, KCL, and TSC with the dynamic parameters. \n1. I am afraid the performance of the proposed method is strongly affected by the batch size since the distribution in each mini batch is important in calculating the loss in Eq. (4). It would be nice if the authors could provide the analysis on this point.\n1. Table 5: “(i) the inter-class distances of SBCL is larger than the other methods, which implies that SBCL can push different classes far away from each other, and thus clear the decision boundaries between classes; ” It is true that the inter-class distances of SBCL is larger than those of the other methods, but the intra-class distance of SBCL is also larger. Actually, the intra-class distances of other methods are much smaller than those of the proposed method, which means that the other methods have favorable characteristics.\n1. P8 “We attribute that to feature space dominated by the head class.” Please rephrase and elaborate more so that the claim of this sentence becomes clearer.\n\n[A1] Zhong+, “Improving Calibration for Long-Tailed Recognition”, CVPR 2021    \n[A2] Cui+, “Parametric Contrastive Learning”, ICCV 2021  \n[A3] Zhu+, “Balanced Contrastive Learning for Long-Tailed Visual Recognition”, CVPR 2022\n\nMinor points.\n1. P2, (c) “… and instant segmentation” -> “instance segmentation”\n1. Eq (2): $\\tilde{V}_i^k$ -> $\\tilde{P}_i^k$?\n1. The second line in Section 3 $|V^+_{i, k}|$ does not appear in equation (2)\n1. P8 “successful captured” -> ““successfully captured””\n1. P9 “well-haved” What does this mean?\n1. P9: “the table 4” -> “Table 4”\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposes a new method called subclass balancing contrastive learning for long-tailed problems. The key idea is to divide head classes into subclasses to balance the number of instances in each subclass. The proposed method is evaluated on widely used image classification datasets as well as some additional vision tasks such as object detection and semantic segmentation. ", "strength_and_weaknesses": "## Strength\n1. The proposed method is evaluated on wide variety of datasets. Especially the evaluation on other tasks than image classification, i.e., object detection and semantic segmentation, is given.\n1. The effectiveness of the proposed method is shown in the experiment at least to some extent.\n1. The paper generally reads well\n\n## Weakness\n1. The comparison with stronger baselines such as [A1-3] is missing. The reported accuracies in these works are actually higher than the proposed method.\n1. According to Table 4, the performance of the proposed method on CIFAR100-LT becomes 43.7 if the dynamic temperature is not applied. This performance is lower than TSC. I am afraid this suggests that it is the dynamic temperature that makes the paper perform better than other contrastive learning methods. To show the effectiveness of the proposed idea, it is necessary to show the results of other contrastive learning methods such as SCL, KCL, and TSC with the dynamic parameters. \n1. I am afraid the performance of the proposed method is strongly affected by the batch size since the distribution in each mini batch is important in calculating the loss in Eq. (4). It would be nice if the authors could provide the analysis on this point.\n1. Table 5: “(i) the inter-class distances of SBCL is larger than the other methods, which implies that SBCL can push different classes far away from each other, and thus clear the decision boundaries between classes; ” It is true that the inter-class distances of SBCL is larger than those of the other methods, but the intra-class distance of SBCL is also larger. Actually, the intra-class distances of other methods are much smaller than those of the proposed method, which means that the other methods have favorable characteristics.\n1. P8 “We attribute that to feature space dominated by the head class.” Please rephrase and elaborate more so that the claim of this sentence becomes clearer.\n\n[A1] Zhong+, “Improving Calibration for Long-Tailed Recognition”, CVPR 2021    \n[A2] Cui+, “Parametric Contrastive Learning”, ICCV 2021  \n[A3] Zhu+, “Balanced Contrastive Learning for Long-Tailed Visual Recognition”, CVPR 2022\n\nMinor points.\n1. P2, (c) “… and instant segmentation” -> “instance segmentation”\n1. Eq (2): $\\tilde{V}_i^k$ -> $\\tilde{P}_i^k$?\n1. The second line in Section 3 $|V^+_{i, k}|$ does not appear in equation (2)\n1. P8 “successful captured” -> ““successfully captured””\n1. P9 “well-haved” What does this mean?\n1. P9: “the table 4” -> “Table 4”\n", "clarity,_quality,_novelty_and_reproducibility": "- Clarity: The paper is mostly clearly written\n- Quality: The quality in terms of presentation is good, but the quality of evaluation is not sufficient as I pointed out above.\n- Novelty: The idea of introducing sub-class contrastive learning into long-tail problems has certain novelty.\n- Reproducibility: The procedure of the proposed method is generally well described, but the initialization of the cluster centers in Algorithm 1 is not explicitly explained. Although I believe that basically important information for reproducibility is given in the paper including the appendix, readers may practically face issues since the code is not provided. \n", "summary_of_the_review": "I think the motivation of the paper is sound and the proposed method is somehow reasonable. However, I have doubt on the effectiveness of the proposed method since I do not think the current experiment is not sufficient for supporting it. As such I do not think this paper is ready for publication. I would like to hear the feedback from the authors on the points I listed in the weakness. \n\n### after rebuttal\nI became more positive about the paper through the rebuttal and I increased my score accordingly, but I still have some concerns on the effectiveness of the proposed method.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666541625676}, {"id": "VX7iLAnTdbo", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1009/Reviewer_S8iv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper focuses on improving supervised contrastive learning to learn better features over class-imbalanced dataset and to achieve better long-tailed recognition (LTR). The paper argues that, in the literature of LTR, typical class-balancing techniques such a data reweighting, resampling and supervised contrastive learning (SupCon) introduce biases as data instances are imbalanced although classes are balanced in training. Motivated by this, the paper proposes a method called \"subclass-balancing contrastive learning (SBCL)\", which clusters training data in balanced subgroups within (head) classes such that subgroups and tail classes are balanced in size. SBCL samples positive pairs of data within subgroups or tail classes in SupCon. It reports improved LTR performance in experiments.", "review_text": "In the context of long-tailed recognition, the paper proposes a method to cluster head classes into small subgroups such that subgroups and tail classes have roughly equal number of training data, allowing for more balanced learning of feature representations using Supervised Contrastive Learning. However, the paper lacks justifications for some design choices and results, and misses citations and lacks analyses. Therefore, the paper is rated as \"marginally below the acceptance threshold\".", "strengths": "Strength:\n- Clustering head classes into subgroups for group-level balanced learning is interesting in the context of long-tailed recognition.\n- The study of long-tailed pretraining for downstream tasks is interesting.\n\n\nBelow are weaknesses.\n\nFor LTR, it is misleading to state that \"the major challenge (of LTR) is the discrepancy between the imbalanced training data distribution and the balanced test set.\" In the real world, testing data should also follow the same long-tailed destruction, but the requirement or evaluation metric is \"class balanced\", i.e., it requires recognizing all imbalanced classes equally well.\n\nWhile the paper states that \" it might be wise to break down the head classes into multiple semantically coherent subclasses,\" it should discuss whether the clustering groups correlate to the \"semantically coherent subclasses\". In other words, how to guarantee the clusters are aware of subclasses?\n\nAlgorithm 1 describes a heuristic method about clustering data of a head class into equal-size subgroups. However, would subgroups produced later be noisy (as they will have large distances to their mean features) compared to those generated in early steps? Can authors provide more analyses or visualizations?\n\nTable 2: While SBCL achieves better performance than the compared methods, [R1] performs much better by tuning weight decay in training. [R1] finds that regularizing network weights (towards more balanced norms in network filters) is crucial for LTR, and doing so leads to state-of-the-art. Therefore, it is unclear why SBCL improves LTR (slightly): whether it is due to balanced subgroups of head classes or whether it helps learn balanced weights in networks. Authors should provide more analyses.\n\n[R1] Alshammari, et al., \"Long-Tailed Recognition via Weight Balancing\", CVPR 2022\n\n\nSection 4.3 and Table 3: Can authors list results by \"training from scratch\" (i.e., without pretraining), and \"freezing backbone\" (i.e., not finetuning the whole pretrained network). The two baselines are important as comparisons can clearly show how much imbalanced/balanced pretraining helps the downstream detection task.\n\nTable 5: It is confusing to compare the absolute distances of intra- and inter-class distances. Instead, a ratio between intra- and inter-class distances might better show how well the per-class examples are separated in the feature space. If computing such a ratio, SBCL seems to perform worse than the other methods. Can authors discuss this point?\n\nThe paper does not show how many epochs/iterations needed for convergence by SupCon and SBCL, and for SBCL's warm-up learning. Can authors provide a learning curve, e.g., loss vs. iterations/epochs, accuracy vs. iterations/epochs?\n\n\ntypos:\n- \"a instance\" in Section 3\n- \"a overly-loose/dense cluster\"", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper focuses on improving supervised contrastive learning to learn better features over class-imbalanced dataset and to achieve better long-tailed recognition (LTR). The paper argues that, in the literature of LTR, typical class-balancing techniques such a data reweighting, resampling and supervised contrastive learning (SupCon) introduce biases as data instances are imbalanced although classes are balanced in training. Motivated by this, the paper proposes a method called \"subclass-balancing contrastive learning (SBCL)\", which clusters training data in balanced subgroups within (head) classes such that subgroups and tail classes are balanced in size. SBCL samples positive pairs of data within subgroups or tail classes in SupCon. It reports improved LTR performance in experiments.", "strength_and_weaknesses": "Strength:\n- Clustering head classes into subgroups for group-level balanced learning is interesting in the context of long-tailed recognition.\n- The study of long-tailed pretraining for downstream tasks is interesting.\n\n\nBelow are weaknesses.\n\nFor LTR, it is misleading to state that \"the major challenge (of LTR) is the discrepancy between the imbalanced training data distribution and the balanced test set.\" In the real world, testing data should also follow the same long-tailed destruction, but the requirement or evaluation metric is \"class balanced\", i.e., it requires recognizing all imbalanced classes equally well.\n\nWhile the paper states that \" it might be wise to break down the head classes into multiple semantically coherent subclasses,\" it should discuss whether the clustering groups correlate to the \"semantically coherent subclasses\". In other words, how to guarantee the clusters are aware of subclasses?\n\nAlgorithm 1 describes a heuristic method about clustering data of a head class into equal-size subgroups. However, would subgroups produced later be noisy (as they will have large distances to their mean features) compared to those generated in early steps? Can authors provide more analyses or visualizations?\n\nTable 2: While SBCL achieves better performance than the compared methods, [R1] performs much better by tuning weight decay in training. [R1] finds that regularizing network weights (towards more balanced norms in network filters) is crucial for LTR, and doing so leads to state-of-the-art. Therefore, it is unclear why SBCL improves LTR (slightly): whether it is due to balanced subgroups of head classes or whether it helps learn balanced weights in networks. Authors should provide more analyses.\n\n[R1] Alshammari, et al., \"Long-Tailed Recognition via Weight Balancing\", CVPR 2022\n\n\nSection 4.3 and Table 3: Can authors list results by \"training from scratch\" (i.e., without pretraining), and \"freezing backbone\" (i.e., not finetuning the whole pretrained network). The two baselines are important as comparisons can clearly show how much imbalanced/balanced pretraining helps the downstream detection task.\n\nTable 5: It is confusing to compare the absolute distances of intra- and inter-class distances. Instead, a ratio between intra- and inter-class distances might better show how well the per-class examples are separated in the feature space. If computing such a ratio, SBCL seems to perform worse than the other methods. Can authors discuss this point?\n\nThe paper does not show how many epochs/iterations needed for convergence by SupCon and SBCL, and for SBCL's warm-up learning. Can authors provide a learning curve, e.g., loss vs. iterations/epochs, accuracy vs. iterations/epochs?\n\n\ntypos:\n- \"a instance\" in Section 3\n- \"a overly-loose/dense cluster\"", "clarity,_quality,_novelty_and_reproducibility": "The clarity can be improved. Quality of the paper is okay. Novelty is below average. It is unclear how easy it is to reproduce the results (code is not provided).\n\n", "summary_of_the_review": "In the context of long-tailed recognition, the paper proposes a method to cluster head classes into small subgroups such that subgroups and tail classes have roughly equal number of training data, allowing for more balanced learning of feature representations using Supervised Contrastive Learning. However, the paper lacks justifications for some design choices and results, and misses citations and lacks analyses. Therefore, the paper is rated as \"marginally below the acceptance threshold\".", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No ethics issues as I am aware.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666525307896}, {"id": "XRHBUuMzKfL", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1009/Reviewer_aZoT"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes subclass-balancing contrastive learning (SBCL) for long-tailed recognition. It divides head classes into several fine-grained subclasses via adaptive clustering, to enforce equal probability of head/tail classes to be engaged. After clustering, a bi-granularity contrastive loss is proposed. In addition, dynamic temperature and warm-up are also used for SBCL. Experimental results show the effectiveness of the proposed method.", "review_text": "The motivation is clear. The proposed method is reasonable though some related works have similar ideas.\nThere are some weak aspects in empirical study, as pointed out above.  ", "strengths": "Strength:\n\n(1) The motivation of this work, i.e., dividing head classes to sub-classes to deal with class imbalance, is clear. The basic idea of enforcing both instance- and subclass-balance by leveraging supervised contrastive learning is reasonable. \n\n(2) The paper is well organized and clearly presented. The method seems easy to reproduce as well. \n\n(3) Experiments on LT datasets and downstream tasks are adequate. \n\nWeaknesses:\n\n(1) Though the motivation and proposed method are reasonable, the comparison with similar long-tailed contrastive learning approaches is not complete. In addition to KCL and TSC, there are some other related solutions for long-tailed recognition problem, like PaCo [1] and BCL [2]. What is the advantage of sub-class clustering over these solutions? \n\n(2) There are quite a few hyper-parameters, like temperature, $\\beta$, $\\delta$ and $\\alpha$ in Eq. 5. Though the authors provide some hyper-parameter study on CIFAR-100-LT, I still wonder how the hyper-parameters change across datasets? Do I need to re-adjust these hyper-parameters with lots of efforts for a new dataset to achieve good performance? \n\n(3) In Table 1, it seems these contrastive methods are implemented on different code base. To verify the effectiveness of SBCL over the SCL baseline, it is better to use your own implementation on SCL. \n\n(4) There are quite a few works trying to re-balance the supervised contrastive loss for long-tailed recognition. What if we instead use non-contrastive methods like BYOL? For non-contrastive methods, the subclass-balancing problem naturally disappears. \n\n\n[1] Parametric contrastive learning, ICCV 2021.\n\n[2] Balanced Contrastive Learning for Long-Tailed Visual Recognition, CVPR 2022.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes subclass-balancing contrastive learning (SBCL) for long-tailed recognition. It divides head classes into several fine-grained subclasses via adaptive clustering, to enforce equal probability of head/tail classes to be engaged. After clustering, a bi-granularity contrastive loss is proposed. In addition, dynamic temperature and warm-up are also used for SBCL. Experimental results show the effectiveness of the proposed method.", "strength_and_weaknesses": "Strength:\n\n(1) The motivation of this work, i.e., dividing head classes to sub-classes to deal with class imbalance, is clear. The basic idea of enforcing both instance- and subclass-balance by leveraging supervised contrastive learning is reasonable. \n\n(2) The paper is well organized and clearly presented. The method seems easy to reproduce as well. \n\n(3) Experiments on LT datasets and downstream tasks are adequate. \n\nWeaknesses:\n\n(1) Though the motivation and proposed method are reasonable, the comparison with similar long-tailed contrastive learning approaches is not complete. In addition to KCL and TSC, there are some other related solutions for long-tailed recognition problem, like PaCo [1] and BCL [2]. What is the advantage of sub-class clustering over these solutions? \n\n(2) There are quite a few hyper-parameters, like temperature, $\\beta$, $\\delta$ and $\\alpha$ in Eq. 5. Though the authors provide some hyper-parameter study on CIFAR-100-LT, I still wonder how the hyper-parameters change across datasets? Do I need to re-adjust these hyper-parameters with lots of efforts for a new dataset to achieve good performance? \n\n(3) In Table 1, it seems these contrastive methods are implemented on different code base. To verify the effectiveness of SBCL over the SCL baseline, it is better to use your own implementation on SCL. \n\n(4) There are quite a few works trying to re-balance the supervised contrastive loss for long-tailed recognition. What if we instead use non-contrastive methods like BYOL? For non-contrastive methods, the subclass-balancing problem naturally disappears. \n\n\n[1] Parametric contrastive learning, ICCV 2021.\n\n[2] Balanced Contrastive Learning for Long-Tailed Visual Recognition, CVPR 2022.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is clearly presented and the somewhat novel. \nThe method seems easy to reproduce. \n\nAdditional concerns:\n(1) What if using adaptive clustering for cross-entropy based long-tailed recognition?", "summary_of_the_review": "The motivation is clear. The proposed method is reasonable though some related works have similar ideas.\nThere are some weak aspects in empirical study, as pointed out above.  ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666333723823}], "openreview_url": "https://openreview.net/forum?id=mPxsHDgsimT", "arxiv_id": "2306.15925", "paper_pdf": "papers/mPxsHDgsimT.pdf", "paper_pdf_sha256": "79b49f8a1698c018405b3ea53ba5e9fae7213cbfe0ba041645313f169ca61fc8", "paper_pdf_bytes": 3434710, "paper_pdf_source": "openreview", "code_url": "https://github.com/JackHck/SBCL", "code_repository": "JackHck/SBCL", "code_commit": "3b9047a31d5c54a0ac14cde351ab557d2833611e", "code_archive": "repos/mPxsHDgsimT.zip", "code_archive_sha256": "f3b0b31352b49e0e9982cf2de3aef5283807a44f57bdd756751d35eb37c58edc", "code_archive_bytes": 911392, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 1039, "github_languages": {"Python": 156976}, "github_archived": false, "github_pushed_at": "2023-10-30T09:23:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/subclass-balancing-contrastive-learning-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "-uPIaaZdMLF", "year": 2022, "status": "rejected", "title": "Attentional meta-learners for few-shot polythetic classification", "authors": ["Ben Day", "Ramon Viñas Torné", "Nikola Simidjievski", "Pietro Lio"], "authorids": ["~Ben_Day1", "~Ramon_Viñas_Torné1", "~Nikola_Simidjievski1", "~Pietro_Lio1"], "authors_source": "OpenReview API", "abstract": "Polythetic classifications, based on shared patterns of features that need neither be universal nor constant among members of a class, are common in the natural world and greatly outnumber monothetic classifications over a set of features. We show that threshold meta-learners, such as Prototypical Networks, require an embedding dimension that is exponential in the number of features to emulate these functions. In contrast, attentional classifiers, such as Matching Networks, are polythetic by default and able to solve these problems with a linear embedding dimension. However, we find that in the presence of task-irrelevant features, inherent to meta-learning problems, attentional models are susceptible to misclassification. To address this challenge, we propose a self-attention feature-selection mechanism that adaptively dilutes non-discriminative features. We demonstrate the effectiveness of our approach in meta-learning Boolean functions, and synthetic and real-world few-shot learning tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "iF4ZPGSGvFn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4145/Reviewer_PpwW"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors present a discussion on two main-stream meta-leaners with attentional classifiers and threshold meta-learners under a unique view of polythetic classification. And to address the limitations of both, attention-based feature selection is introduced. \nImproved performance is demonstrated by both synthetic and real-world few-shot learning tasks.\n", "review_text": "**Strengths**\n\nThis paper discusses meta-learner from a very unique perspective, with the shortcomings of both ProtoNet-liked threshold meta-learners and MatchingNet-liked attentional meta-learner discussed solidly with intuitive examples and convincing derivation of misclassification rates. \n\nThe proposed attentional feature selection is simple yet effective, and can potentially guide and stimulate many follow-up improvements to meta-learning. \n\n**Weaknesses**\n\nMy main concern is that this paper tends to isolate itself from the entire research of meta-learning, and focus on a corner instead. \nProtoNets (representing threshold meta-learners) and MatchingNets (representing attentive meta-learners) are insufficient to cover the entire research of meta learning. There are many other directions of meta-learning that are completely overlooked in the discussion. \nThe most representative case is the inner-loop optimization-based meta-learner like MAML [1], where the inner-loop adaptation to the feature extractor can potentially solve the challenge of 'not all features are relevant in all tasks and that the support is unlikely to span the input domain' pointed out in this paper. \n\nAnd many other directions and methods, e.g., methods based on Hebbian rules [2] might share a similar spirit with the proposed method, and the presentation can be further improved by incorporating more comprehensive discussion to the latest progress of meta-learning and the connections of the proposed method to others. \n\n**Minor**\n\nThe experimental settings included in this paper seem a bit weak. More common benchmarks on real-world meta-learning (classification) tasks, like miniImageNet, and the challenging cross-domain settings can better support the discussions.\n\nThe overall writing is good, but some further improvements are expected. For example, is the very first sentence of this paper grammatically incorrect? I believe it's better to say 'that need to be neither universal nor ...'\n\n[1] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, ICML2017\n\n[2] Differentiable plasticity: training plastic neural networks with backpropagation, ICML 2018", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors present a discussion on two main-stream meta-leaners with attentional classifiers and threshold meta-learners under a unique view of polythetic classification. And to address the limitations of both, attention-based feature selection is introduced. \nImproved performance is demonstrated by both synthetic and real-world few-shot learning tasks.\n", "main_review": "**Strengths**\n\nThis paper discusses meta-learner from a very unique perspective, with the shortcomings of both ProtoNet-liked threshold meta-learners and MatchingNet-liked attentional meta-learner discussed solidly with intuitive examples and convincing derivation of misclassification rates. \n\nThe proposed attentional feature selection is simple yet effective, and can potentially guide and stimulate many follow-up improvements to meta-learning. \n\n**Weaknesses**\n\nMy main concern is that this paper tends to isolate itself from the entire research of meta-learning, and focus on a corner instead. \nProtoNets (representing threshold meta-learners) and MatchingNets (representing attentive meta-learners) are insufficient to cover the entire research of meta learning. There are many other directions of meta-learning that are completely overlooked in the discussion. \nThe most representative case is the inner-loop optimization-based meta-learner like MAML [1], where the inner-loop adaptation to the feature extractor can potentially solve the challenge of 'not all features are relevant in all tasks and that the support is unlikely to span the input domain' pointed out in this paper. \n\nAnd many other directions and methods, e.g., methods based on Hebbian rules [2] might share a similar spirit with the proposed method, and the presentation can be further improved by incorporating more comprehensive discussion to the latest progress of meta-learning and the connections of the proposed method to others. \n\n**Minor**\n\nThe experimental settings included in this paper seem a bit weak. More common benchmarks on real-world meta-learning (classification) tasks, like miniImageNet, and the challenging cross-domain settings can better support the discussions.\n\nThe overall writing is good, but some further improvements are expected. For example, is the very first sentence of this paper grammatically incorrect? I believe it's better to say 'that need to be neither universal nor ...'\n\n[1] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, ICML2017\n\n[2] Differentiable plasticity: training plastic neural networks with backpropagation, ICML 2018", "summary_of_the_review": "I give an initial recommendation of score 6 mainly for appreciating this novel view to mainstream meta-learners, and solid discussions supporting the points. However, to meet and standard of ICLR and demonstrate a clear contribution to the research of meta-learning, I believe further efforts on comprehensive discussions are highly expected. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635922692020}, {"id": "Hh-eaecbZdA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4145/Reviewer_f8L4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors propose the general problem of few-shot polythetic classification, where class membership is determined by the combinations of present and absent features and the salient features and combination patterns change at test time. Authors demonstrate that prototypical approaches with linear decision boundaries respond poorly to these highly non-linear problems, while attention-based soft nearest neighbors paradigms work well but overfit to specious cues. An attention-based feature refinement technique is proposed and evaluated, beating both baselines on specifically polythetic few-shot tasks. ", "review_text": "STRENGTHS:\n\nThe few-shot polythetic classification problem is interesting and novel, and the shortcomings of baseline approaches are readily apparent. Work is well-grounded in older prior literature and the XOR_alpha task is extensively analyzed both theoretically and empirically. Writing is clear and concise (though not always easy to understand, see below). The proposed method performs well in its intended setting. \n\nWEAKNESSES/ISSUES (in no particular order):\n\n- The XOR_alpha meta-learning task is never formally described (it is only described as a function of bits on page 3), and so subsequent sections become difficult to follow. Please state clearly in section 2 that XOR_alpha also refers to the collection of (presumably?) all alpha-variable XOR “tasks” in a given binary n-space with n>=alpha, with a (presumably?) random partition of train and test tasks. Perhaps also consider using different notation when referring to the meta-learning problem (i.e. caption for Fig.3 right) vs the task (i.e. second to last paragraph of pg.4). Either way, the problem setup for this and other experimental tasks more broadly (i.e. the non-standard tiered-ImageNet task) should be elaborated up front and more clearly. As is I would have a difficult time re-implementing some of these benchmarks without the provided code. \n- Contrary to claims in the abstract and conclusion, the paper does not show that the embedding space of a threshold classifier must grow exponentially with the number of features. Pg.4 discusses embedding growth but has it at O(n^alpha), which for given alpha is only polynomial in the number of features. The embedding space does grow exponentially with task complexity alpha, which is itself O(n), but they are not the same and these claims should be clarified (assuming that this is in fact authors’ intended argument). \n- On a much broader level, I have a hard time understanding the motivation for this work. What is the envisioned use case scenario for this kind of approach? The synthetic examples, while interesting, are very contrived (which is fine since they’re synthetic) but the paper does not provide a strong motivation for the “real world” examples (omniglot and tiered-ImageNet) either. These are also somewhat contrived: while the fine-to-coarse generalization task under discussion is interesting, it’s the opposite coarse-to-fine generalization that has a clear use case. What is the envisioned use-case for a polythetic meta-learner? The introduction suggests certain kinds of fine-grained classification, but if this is the case then it should be investigated directly, such as with CUB or tiered meta-iNat benchmarks. \n- The first paragraph of page 2 is difficult to parse, both because sentences are somewhat long and mostly because it is not immediately clear how conclusions are following from premises (i.e. the described 45degree rotation/”change of basis” of the OR function was initially confusing to me because this geometric operation does not translate intuitively into pure Boolean algebra; the rates of growth (2^2^n and 2^n^2) are not immediately obvious). This section could use some elaboration and clarification. \n- I do not find Appendix A convincing. Appendix A claims to be a demonstration that protonets do not generalize to unseen variable combinations. While it is true that train and test combinations do not overlap, by my understanding the chance performance here is clearly due to the fact that training time noise features are test time signal features, and vice versa, and the network has simply (and correctly!) learned to suppress the “relevant” noise features. As such, no learning-based classifier should be able to solve this problem as presented; it is less a demonstration of the limits of protonets and more a demonstration of the no free lunch theorem. A much more useful and interesting set of results would be to partition the combinations of active variables randomly over train and test, so that combinations do not overlap but features are equally active in expectation at both train and test time. By my understanding this is exactly the experiment in Fig.3 right, though, so perhaps this ought to just be removed entirely. \n- Authors use a pre-trained ResNet-18 for the tiered-ImageNet experiment, but the provided source for the model is PyTorch itself. How was this model pretrained? If the model was pretrained on ImageNet, then it has been trained on the test images and these results are not valid. \n- Typo in Algorithm 1 Input: the an arbitrarily ordered matrix -> an arbitrarily ordered matrix\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Authors propose the general problem of few-shot polythetic classification, where class membership is determined by the combinations of present and absent features and the salient features and combination patterns change at test time. Authors demonstrate that prototypical approaches with linear decision boundaries respond poorly to these highly non-linear problems, while attention-based soft nearest neighbors paradigms work well but overfit to specious cues. An attention-based feature refinement technique is proposed and evaluated, beating both baselines on specifically polythetic few-shot tasks. ", "main_review": "STRENGTHS:\n\nThe few-shot polythetic classification problem is interesting and novel, and the shortcomings of baseline approaches are readily apparent. Work is well-grounded in older prior literature and the XOR_alpha task is extensively analyzed both theoretically and empirically. Writing is clear and concise (though not always easy to understand, see below). The proposed method performs well in its intended setting. \n\nWEAKNESSES/ISSUES (in no particular order):\n\n- The XOR_alpha meta-learning task is never formally described (it is only described as a function of bits on page 3), and so subsequent sections become difficult to follow. Please state clearly in section 2 that XOR_alpha also refers to the collection of (presumably?) all alpha-variable XOR “tasks” in a given binary n-space with n>=alpha, with a (presumably?) random partition of train and test tasks. Perhaps also consider using different notation when referring to the meta-learning problem (i.e. caption for Fig.3 right) vs the task (i.e. second to last paragraph of pg.4). Either way, the problem setup for this and other experimental tasks more broadly (i.e. the non-standard tiered-ImageNet task) should be elaborated up front and more clearly. As is I would have a difficult time re-implementing some of these benchmarks without the provided code. \n- Contrary to claims in the abstract and conclusion, the paper does not show that the embedding space of a threshold classifier must grow exponentially with the number of features. Pg.4 discusses embedding growth but has it at O(n^alpha), which for given alpha is only polynomial in the number of features. The embedding space does grow exponentially with task complexity alpha, which is itself O(n), but they are not the same and these claims should be clarified (assuming that this is in fact authors’ intended argument). \n- On a much broader level, I have a hard time understanding the motivation for this work. What is the envisioned use case scenario for this kind of approach? The synthetic examples, while interesting, are very contrived (which is fine since they’re synthetic) but the paper does not provide a strong motivation for the “real world” examples (omniglot and tiered-ImageNet) either. These are also somewhat contrived: while the fine-to-coarse generalization task under discussion is interesting, it’s the opposite coarse-to-fine generalization that has a clear use case. What is the envisioned use-case for a polythetic meta-learner? The introduction suggests certain kinds of fine-grained classification, but if this is the case then it should be investigated directly, such as with CUB or tiered meta-iNat benchmarks. \n- The first paragraph of page 2 is difficult to parse, both because sentences are somewhat long and mostly because it is not immediately clear how conclusions are following from premises (i.e. the described 45degree rotation/”change of basis” of the OR function was initially confusing to me because this geometric operation does not translate intuitively into pure Boolean algebra; the rates of growth (2^2^n and 2^n^2) are not immediately obvious). This section could use some elaboration and clarification. \n- I do not find Appendix A convincing. Appendix A claims to be a demonstration that protonets do not generalize to unseen variable combinations. While it is true that train and test combinations do not overlap, by my understanding the chance performance here is clearly due to the fact that training time noise features are test time signal features, and vice versa, and the network has simply (and correctly!) learned to suppress the “relevant” noise features. As such, no learning-based classifier should be able to solve this problem as presented; it is less a demonstration of the limits of protonets and more a demonstration of the no free lunch theorem. A much more useful and interesting set of results would be to partition the combinations of active variables randomly over train and test, so that combinations do not overlap but features are equally active in expectation at both train and test time. By my understanding this is exactly the experiment in Fig.3 right, though, so perhaps this ought to just be removed entirely. \n- Authors use a pre-trained ResNet-18 for the tiered-ImageNet experiment, but the provided source for the model is PyTorch itself. How was this model pretrained? If the model was pretrained on ImageNet, then it has been trained on the test images and these results are not valid. \n- Typo in Algorithm 1 Input: the an arbitrarily ordered matrix -> an arbitrarily ordered matrix\n", "summary_of_the_review": "The paper proposes an interesting problem, and based on theoretical and empirical analysis provides a neat solution. Issues stem from the perhaps overly concise writing; many aspects of the paper are in need of elaboration. These include the motivation, problem setups for the various benchmarks, and the reasoning behind certain conclusions. Currently I am scoring the paper as if my understanding is correct and authors’ intended arguments are simply wrong, but if these issues can be clarified I’d be happy to raise my score. \n\nPOST-DISCUSSION:\n\nAuthors mostly address the above issues, in discussion and in the revised manuscript. I have some lingering concerns, shared with reviewer q98z, regarding the lack of _demonstrated_ practical benefit, and the fact that accuracy gains in the more “real-world” benchmarks shrink substantially relative to the motivating XOR problem. However, the revised paper is much improved, conceptual novelty remains high, and on the whole the paper is of sufficient quality for acceptance. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635897723208}, {"id": "qN_kajIiqxh", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4145/Reviewer_q98z"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper discusses monothetic and polythetic classifications in the context of few-shot learning: distinguishing similar classes often require reasoning with combinations of certain features, which can be especially challenging when only a few training examples are available. Two common types of few-shot methods, ProtoNets and Matching Networks, are shown to be threshold classifiers and attentional classifiers, respectively, each with their advantages and disadvantages. The authors introduce a simple method based on self-attention as a solution to these challenges, with experiments on several toy tasks and the more real-world TieredImageNet.", "review_text": "Pros:\n1.\tStrong motivation and introduction of monothetic and polythetic classifications, drawing from prior work from other fields\n2.\tInteresting analysis and excellent visualizations\n3.\tUses simple Boolean tasks to explain background conceptually\n4.\tWell-designed toy datasets and experiments to verify hypotheses of the nature of ProtoNets and Matching Networks in the context of the paper\n5.\tWell-written\n\nCons:\n1.\tPractical value isn’t clearly demonstrated; improvements over matching networks for the only non-toy dataset (TieredImageNet) is marginal at best.\n2.\tLimited baselines in experiments\n\nMore detailed comments:\n\nThis paper examines few-shot learning in the context of monothetic and polythetic classifications. This perspective sheds some light on how ProtoNets and Matching Networks perform classification of features, and there is some interesting analysis to back these hypotheses up. Concepts from the paper are well-explained with several simple examples, and toy settings are designed to illustrate them empirically. Furthermore, the analysis leads to a simple self-attention-based solution to improve features for classification. Overall, this paper reads well, and the analysis leads to some interesting insights into few-shot learning.\n\n\nEmpirically, the results are a little more disappointing. While several toy experiments (e.g. XOR binary strings, polythetic MNIST) are well-designed to illustrate the advantages of considering polythetic patterns, it’s not as clear how much the real-world exhibits these characteristics. The proposed method strongly outperforms the baselines in the toy settings, but improvements over the Matching Networks baseline on TieredImageNet is marginal at best. As such, as nice as these insights are, it’s not clear if it’s practically useful. Additionally, while I understand the focus was comparing with ProtoNets and Matching Networks as representatives of threshold-based and attentional classifiers, it would have been nice to have included more baselines in the experiments.\n\n\nQuestions:\n1.\tWhile conceptually easy-to-understand, I was previously unaware of categorizing classifiers as either attentional or threshold-based. Are there any other kinds, or must a non-attentional classifier be threshold based (and vice versa)?\n2.\tHow do other few-shot methods fit into this taxonomy? For example, what are kinds of features (monothetic vs polythetic) are optimization-based methods (e.g. MAML) learning? What about simply training a classifier, or SVM?\n\n\n\n=====Post-Discussion=====\n\nI thank the authors for all their effort during the discussion phase to clarify various questions about their paper and for running additional baselines. After reading the other reviews and seeing the authors' responses, my recommendation remains mostly the same. The new perspective on meta-learning provided by the authors is an interesting one, but the practical benefits of this approach can still be more concretely demonstrated. Additional experiments in such use cases (e.g. the DNA example mentioned in one of the discussions, if such a dataset exists) would significantly strengthen this paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper discusses monothetic and polythetic classifications in the context of few-shot learning: distinguishing similar classes often require reasoning with combinations of certain features, which can be especially challenging when only a few training examples are available. Two common types of few-shot methods, ProtoNets and Matching Networks, are shown to be threshold classifiers and attentional classifiers, respectively, each with their advantages and disadvantages. The authors introduce a simple method based on self-attention as a solution to these challenges, with experiments on several toy tasks and the more real-world TieredImageNet.", "main_review": "Pros:\n1.\tStrong motivation and introduction of monothetic and polythetic classifications, drawing from prior work from other fields\n2.\tInteresting analysis and excellent visualizations\n3.\tUses simple Boolean tasks to explain background conceptually\n4.\tWell-designed toy datasets and experiments to verify hypotheses of the nature of ProtoNets and Matching Networks in the context of the paper\n5.\tWell-written\n\nCons:\n1.\tPractical value isn’t clearly demonstrated; improvements over matching networks for the only non-toy dataset (TieredImageNet) is marginal at best.\n2.\tLimited baselines in experiments\n\nMore detailed comments:\n\nThis paper examines few-shot learning in the context of monothetic and polythetic classifications. This perspective sheds some light on how ProtoNets and Matching Networks perform classification of features, and there is some interesting analysis to back these hypotheses up. Concepts from the paper are well-explained with several simple examples, and toy settings are designed to illustrate them empirically. Furthermore, the analysis leads to a simple self-attention-based solution to improve features for classification. Overall, this paper reads well, and the analysis leads to some interesting insights into few-shot learning.\n\n\nEmpirically, the results are a little more disappointing. While several toy experiments (e.g. XOR binary strings, polythetic MNIST) are well-designed to illustrate the advantages of considering polythetic patterns, it’s not as clear how much the real-world exhibits these characteristics. The proposed method strongly outperforms the baselines in the toy settings, but improvements over the Matching Networks baseline on TieredImageNet is marginal at best. As such, as nice as these insights are, it’s not clear if it’s practically useful. Additionally, while I understand the focus was comparing with ProtoNets and Matching Networks as representatives of threshold-based and attentional classifiers, it would have been nice to have included more baselines in the experiments.\n\n\nQuestions:\n1.\tWhile conceptually easy-to-understand, I was previously unaware of categorizing classifiers as either attentional or threshold-based. Are there any other kinds, or must a non-attentional classifier be threshold based (and vice versa)?\n2.\tHow do other few-shot methods fit into this taxonomy? For example, what are kinds of features (monothetic vs polythetic) are optimization-based methods (e.g. MAML) learning? What about simply training a classifier, or SVM?\n\n\n\n=====Post-Discussion=====\n\nI thank the authors for all their effort during the discussion phase to clarify various questions about their paper and for running additional baselines. After reading the other reviews and seeing the authors' responses, my recommendation remains mostly the same. The new perspective on meta-learning provided by the authors is an interesting one, but the practical benefits of this approach can still be more concretely demonstrated. Additional experiments in such use cases (e.g. the DNA example mentioned in one of the discussions, if such a dataset exists) would significantly strengthen this paper.", "summary_of_the_review": "As stated in the main reviews, the empirical results aren’t particularly impressive, but overall the paper does provide some interesting analysis that may encourage new ways of thinking about few-shot learning problems. As such, I think this paper may be of interest to the ICLR community.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635834358522}, {"id": "Wmir7yL0ROR", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4145/Reviewer_Hoh6"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper first considers the limitations of threshold and attentional classifiers. They proposed an attention-based method for feature selection to address the problems of threshold classifiers and attentional classifiers.  The experiments on  several synthetic and real-world few-shot learning tasks seem good. ", "review_text": "Strengths: \n\n(1) The explanation about the challenges of threshold (Prototypical networks) and attention classifiers (Matching network) is interesting.\n\n(2) The motivation on the polythetic classification is well-motivated. \n\n(3) Well-written and easy to follow.\n\n\nWeaknesses:\n\n(1) Lack of technological innovation. The proposed feature-selection mechanism is too simple and the only technological innovation.  \nThey only use self-attention to obtain a better representation rather than directly using average pooling (Prototypical networks) or all the support data (matching network).\n\n(2) How to choose repetitions R?  It is a hyper-parameter, is the bigger the R the better? Does it have no upper bound?\n\n(3) Lack the experiments for comparison with threshold classifiers and attentional classifiers in the main paper. It is crucial to show the problems of threshold classifiers and attentional classifiers.\n\n(4)  Self-attention also has some parameters for the transformation, why is the proposed method non-parametric?\n\n(5) The related work is too little. Maybe consider adding some self-attention work?\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper first considers the limitations of threshold and attentional classifiers. They proposed an attention-based method for feature selection to address the problems of threshold classifiers and attentional classifiers.  The experiments on  several synthetic and real-world few-shot learning tasks seem good. ", "main_review": "Strengths: \n\n(1) The explanation about the challenges of threshold (Prototypical networks) and attention classifiers (Matching network) is interesting.\n\n(2) The motivation on the polythetic classification is well-motivated. \n\n(3) Well-written and easy to follow.\n\n\nWeaknesses:\n\n(1) Lack of technological innovation. The proposed feature-selection mechanism is too simple and the only technological innovation.  \nThey only use self-attention to obtain a better representation rather than directly using average pooling (Prototypical networks) or all the support data (matching network).\n\n(2) How to choose repetitions R?  It is a hyper-parameter, is the bigger the R the better? Does it have no upper bound?\n\n(3) Lack the experiments for comparison with threshold classifiers and attentional classifiers in the main paper. It is crucial to show the problems of threshold classifiers and attentional classifiers.\n\n(4)  Self-attention also has some parameters for the transformation, why is the proposed method non-parametric?\n\n(5) The related work is too little. Maybe consider adding some self-attention work?\n\n\n", "summary_of_the_review": "The implementation details and related work of the proposed paper should be further added.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635783598016}], "openreview_url": "https://openreview.net/forum?id=-uPIaaZdMLF", "arxiv_id": "2106.05317", "paper_pdf": "papers/-uPIaaZdMLF.pdf", "paper_pdf_sha256": "eba9d7222c7ac6cfe16aa36bc37152d58c64f1d5e7c3e555dafc4abc0c6a11dc", "paper_pdf_bytes": 3574651, "paper_pdf_source": "openreview", "code_url": "https://github.com/rvinas/polythetic_metalearning", "code_repository": "rvinas/polythetic_metalearning", "code_commit": "97f510ae0810035e51be09c8542e6eed2df66d69", "code_archive": "repos/-uPIaaZdMLF.zip", "code_archive_sha256": "ade1343564d7e653ab3e919e52d915d570a9801db018d4db21e4097aaa6f97d6", "code_archive_bytes": 4553622, "code_file_count": 26, "code_extensions": {".py": 18, ".ipynb": 8}, "github_disk_usage_kb": 4460, "github_languages": {"Jupyter Notebook": 4813353, "Python": 110683}, "github_archived": false, "github_pushed_at": "2022-10-02T01:07:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/attentional-meta-learners-are-polythetic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1AyPW2Emp6", "year": 2021, "status": "rejected", "title": "Tight Second-Order Certificates for Randomized Smoothing", "authors": ["Alexander Levine", "Aounon Kumar", "Tom Goldstein", "Soheil Feizi"], "authorids": ["~Alexander_Levine2", "aounon@umd.edu", "~Tom_Goldstein1", "~Soheil_Feizi2"], "authors_source": "OpenReview API", "abstract": "Randomized smoothing is a popular way of providing robustness guarantees against adversarial attacks: randomly-smoothed functions have a universal Lipschitz-like bound, allowing for robustness certificates to be easily computed. In this work, we show that there also exists a universal curvature-like bound for Gaussian random smoothing: given the exact value and gradient of a smoothed function, we compute a lower bound on the distance of a point to its closest adversarial example, called the Second-order  Smoothing (SoS) robustness certificate. In addition to proving the correctness of this novel certificate, we show that SoS certificates are realizable and therefore tight. Interestingly, we show that the maximum achievable benefits, in terms of certified robustness, from using the additional information of the gradient norm are relatively small: because our bounds are tight, this is a fundamental negative result. The gain of SoS certificates further diminishes if we consider the estimation error of the gradient norms, for which we have developed an estimator. We therefore additionally develop a variant of Gaussian smoothing, called Gaussian dipole smoothing, which provides similar bounds to randomized smoothing with gradient information, but with much-improved sample efficiency. This allows us to achieve (marginally) improved robustness certificates on high-dimensional datasets such as CIFAR-10 and ImageNet.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "ayGHDsDvMIN", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2218/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "My main concern is that the improvement of the proposed method over standard randomized smoothing is marginal.  \nIn addition, the evaluation metric used for comparison is not standard.  It would be better to use certified accuracy as the evaluation metric used by the related works.  \n\nPros:\n\n1. The studied problem is important.  In particular, it's important to study the certified robustness of the classifier against adversarial perturbations. \n\n2. The proposed second-order smoothing is novel. This paper aims to improve randomized smoothing by incorporating the gradient information of the smoothed classifier, which is novel.  \n\n\nCons:\n1. The evaluation metric is not standard.  Standard randomized smoothing and the follow-up work use certified accuracy as the metric to evaluate the certified robustness.  It's better for the authors to also use certified accuracy as the metric for a fair comparison. \n\nIt's not clear how different parameters (e.g., c, lambda, eta) impact the certified accuracy on CIFAR10 and ImageNet dataset. \n\n2. The improvement of the proposed method (Gaussian dipole smoothing) over the standard randomized smoothing is marginal (most 1% in Figure 5). \n\n\nQuestions during the rebuttal period:\n\nWhy not using certified accuracy? \n\nWhat's the impact of the parameters on the results? \n\nIs it possible to address the issue of marginal improvement? \n\n\nI would be happy to increase my score if the authors can also show results on certified accuracy and improve their results?  \n\nTypos:\n\n1. In Theorem 1, \"there exists a base classifier ... such that Equation 4 is an equality\". Is it Equation 2?\n\n2. $nabla_x$ --> $\\nabla_x$. \n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper tries to improve the randomized smoothing by leveraging the gradient information of the smoothed classifier. ", "review": "My main concern is that the improvement of the proposed method over standard randomized smoothing is marginal.  \nIn addition, the evaluation metric used for comparison is not standard.  It would be better to use certified accuracy as the evaluation metric used by the related works.  \n\nPros:\n\n1. The studied problem is important.  In particular, it's important to study the certified robustness of the classifier against adversarial perturbations. \n\n2. The proposed second-order smoothing is novel. This paper aims to improve randomized smoothing by incorporating the gradient information of the smoothed classifier, which is novel.  \n\n\nCons:\n1. The evaluation metric is not standard.  Standard randomized smoothing and the follow-up work use certified accuracy as the metric to evaluate the certified robustness.  It's better for the authors to also use certified accuracy as the metric for a fair comparison. \n\nIt's not clear how different parameters (e.g., c, lambda, eta) impact the certified accuracy on CIFAR10 and ImageNet dataset. \n\n2. The improvement of the proposed method (Gaussian dipole smoothing) over the standard randomized smoothing is marginal (most 1% in Figure 5). \n\n\nQuestions during the rebuttal period:\n\nWhy not using certified accuracy? \n\nWhat's the impact of the parameters on the results? \n\nIs it possible to address the issue of marginal improvement? \n\n\nI would be happy to increase my score if the authors can also show results on certified accuracy and improve their results?  \n\nTypos:\n\n1. In Theorem 1, \"there exists a base classifier ... such that Equation 4 is an equality\". Is it Equation 2?\n\n2. $nabla_x$ --> $\\nabla_x$. \n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604674820588}, {"id": "hU5zF68a6hG", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2218/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:  This paper presents a randomized second-order smoothing certificate for providing robustness guarantees against adversarial attacks. By additionally using the gradient estimation of smoothed classifier, the proposed method has been shown to outperform the existing randomized smoothing certificate in practice. A variant of the method without explicitly estimating gradient vector has also been proposed to avoid the dependence of feature dimension in concentration analysis.\n\nStrong points:\n\n-S1. The addressed topic of randomized smoothing certificate is of significant importance and interest to the society of adversarial learning.\n\n-S2. The certificate radius of the proposed method is novel as far as the reviewer knows about. \n \nWeak points:\n\n-W1. The advantage of the certificate radius in Theorem 1 over the existing ones is not clearly justified. Unlike the original randomized smoothing classifiers, there seems no explicit expression available for the certificate radius in Theorem 1. Based on the current bound in Equation 2, it is hard to evaluate the theoretical gain of the proposed method in robustness. \n\n-W2. The proof of Theorem in Appendix Section A.1 looks not correct in general. The current proof argument is only customized for $x=0$ and $x’=[R,0,…,0]^\\top$. It is not clear if the same claim and technique extend to arbitrary $x$ and $x’$ satisfying $\\|x – x’\\|\\le R$.\n\n-W3. The paper is poorly organized and presented. The main results in Section 3 are somewhat hard to follow for non-expert audiences, mainly due to the lack of a clear statement of method before indulging into theoretical analysis. Also, much of the space was allocated for elaborating a gradient estimation method which in my opinion is mostly incremental as a side contribution. Such a practice of material organization makes the paper unclear and perhaps pointless in presentation. \n\n-W4. As an adversarial learning paper that introduces a new alternative method for robustness certificate, it is desirable to provide a sufficiently detailed experimental study in the main paper (rather than in the appendix) to more convincingly justify the real benefit of method. \n\n-W5. There are many typos in the manuscript. Here are a few examples: \n\n(1) Equation 4 -> Equation 2 in Theorem 1; \n\n(2) Page 2: it possible -> it is possible; \n\n(3) Page 5: \\|nabla_x p_a(x)\\| -> $\\|\\nabla_x p_a(x)\\|$; \n\n(4) Page 5: Salman et al. (2019) suggests -> suggest; \n\n(5) Theorem 2: $n$ -> $N$. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Merit of method has not been convincingly justified in theory and experiment", "review": "Summary:  This paper presents a randomized second-order smoothing certificate for providing robustness guarantees against adversarial attacks. By additionally using the gradient estimation of smoothed classifier, the proposed method has been shown to outperform the existing randomized smoothing certificate in practice. A variant of the method without explicitly estimating gradient vector has also been proposed to avoid the dependence of feature dimension in concentration analysis.\n\nStrong points:\n\n-S1. The addressed topic of randomized smoothing certificate is of significant importance and interest to the society of adversarial learning.\n\n-S2. The certificate radius of the proposed method is novel as far as the reviewer knows about. \n \nWeak points:\n\n-W1. The advantage of the certificate radius in Theorem 1 over the existing ones is not clearly justified. Unlike the original randomized smoothing classifiers, there seems no explicit expression available for the certificate radius in Theorem 1. Based on the current bound in Equation 2, it is hard to evaluate the theoretical gain of the proposed method in robustness. \n\n-W2. The proof of Theorem in Appendix Section A.1 looks not correct in general. The current proof argument is only customized for $x=0$ and $x’=[R,0,…,0]^\\top$. It is not clear if the same claim and technique extend to arbitrary $x$ and $x’$ satisfying $\\|x – x’\\|\\le R$.\n\n-W3. The paper is poorly organized and presented. The main results in Section 3 are somewhat hard to follow for non-expert audiences, mainly due to the lack of a clear statement of method before indulging into theoretical analysis. Also, much of the space was allocated for elaborating a gradient estimation method which in my opinion is mostly incremental as a side contribution. Such a practice of material organization makes the paper unclear and perhaps pointless in presentation. \n\n-W4. As an adversarial learning paper that introduces a new alternative method for robustness certificate, it is desirable to provide a sufficiently detailed experimental study in the main paper (rather than in the appendix) to more convincingly justify the real benefit of method. \n\n-W5. There are many typos in the manuscript. Here are a few examples: \n\n(1) Equation 4 -> Equation 2 in Theorem 1; \n\n(2) Page 2: it possible -> it is possible; \n\n(3) Page 5: \\|nabla_x p_a(x)\\| -> $\\|\\nabla_x p_a(x)\\|$; \n\n(4) Page 5: Salman et al. (2019) suggests -> suggest; \n\n(5) Theorem 2: $n$ -> $N$. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604057621087}, {"id": "HLqDr-jbY2n", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2218/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis work extends previous results on certified robustness guarantees via randomised smoothing by incorporating gradient information of the smoothed function into the final certificate (referred to as SoS). The authors show a number of interesting properties such as (1) that the certificate is tight, by showing, for a linear classifier, the Cohen et al. [1] certificate and the SoS certificate are identical, and (2) certificates are slightly improved if gradient norms are small, but due to poor estimation of these value, sometimes the gains provided by \"second-order information\" are eliminated. The authors then show estimation of the gradient norm is dependent on the dimensionality of the data, and provide alternative methods that do not have a strict dependency. Although, as far as I understand, this alternative method requires computing two new lower bounds and so requires more samples to reach the same level of precision as the Cohen et al. certificate. Experiments on CIFAR-10 and ImageNet show in some cases SoS finds larger certificates.\n\nStrength:\n\n[+] Detailed analysis and theorem.\n\n[+] Interesting use of geometrical information to make gradient norm estimation efficient.\n\n[+] Experiments on large scale datasets.\n\nWeaknesses / Questions:\n\n1. I  am confused about the positioning of this work. The authors mention a few times that this work could be interpreted as a negative result, but the authors write in a more positive manner in other parts. Should the work be viewed as an improvement over recent work, or as a negative result that incorporating gradients doesn't improve certificates by much? If it is the latter, it isn't clear which dimension of the work is responsible for the limitations. Is it simply that, in practice, the gradient norms are substantially larger than zero, and so this new certificate more or less matches the Cohen et al. certificate? Or is it that difficulty in estimation of the gradient norms wipes away any gains? Is it both? A more substantial study of these effects would be most welcome.\n\n2. I found it difficult to interpret the experimental results. How many of the +/- 1% changes are, in fact, decreases? It would be useful to plot  these results in terms of certified accuracy vs epsilon, as has become standard in this line of work. My concern is that if the majority of SoS certificates are marginally smaller than the Cohen et al. certificates, and a few SoS certificates are substantially higher, the overall certified accuracy curve will look worse for some epsilon values in comparison to Cohen et al..\n\n3. Analysis does not easily extend to multi-class setting.\n\n[1] Cohen, Jeremy M., Elan Rosenfeld, and J. Zico Kolter. \"Certified adversarial robustness via randomized smoothing.\" arXiv preprint arXiv:1902.02918 (2019).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting ideas; confused positioning", "review": "Summary:\n\nThis work extends previous results on certified robustness guarantees via randomised smoothing by incorporating gradient information of the smoothed function into the final certificate (referred to as SoS). The authors show a number of interesting properties such as (1) that the certificate is tight, by showing, for a linear classifier, the Cohen et al. [1] certificate and the SoS certificate are identical, and (2) certificates are slightly improved if gradient norms are small, but due to poor estimation of these value, sometimes the gains provided by \"second-order information\" are eliminated. The authors then show estimation of the gradient norm is dependent on the dimensionality of the data, and provide alternative methods that do not have a strict dependency. Although, as far as I understand, this alternative method requires computing two new lower bounds and so requires more samples to reach the same level of precision as the Cohen et al. certificate. Experiments on CIFAR-10 and ImageNet show in some cases SoS finds larger certificates.\n\nStrength:\n\n[+] Detailed analysis and theorem.\n\n[+] Interesting use of geometrical information to make gradient norm estimation efficient.\n\n[+] Experiments on large scale datasets.\n\nWeaknesses / Questions:\n\n1. I  am confused about the positioning of this work. The authors mention a few times that this work could be interpreted as a negative result, but the authors write in a more positive manner in other parts. Should the work be viewed as an improvement over recent work, or as a negative result that incorporating gradients doesn't improve certificates by much? If it is the latter, it isn't clear which dimension of the work is responsible for the limitations. Is it simply that, in practice, the gradient norms are substantially larger than zero, and so this new certificate more or less matches the Cohen et al. certificate? Or is it that difficulty in estimation of the gradient norms wipes away any gains? Is it both? A more substantial study of these effects would be most welcome.\n\n2. I found it difficult to interpret the experimental results. How many of the +/- 1% changes are, in fact, decreases? It would be useful to plot  these results in terms of certified accuracy vs epsilon, as has become standard in this line of work. My concern is that if the majority of SoS certificates are marginally smaller than the Cohen et al. certificates, and a few SoS certificates are substantially higher, the overall certified accuracy curve will look worse for some epsilon values in comparison to Cohen et al..\n\n3. Analysis does not easily extend to multi-class setting.\n\n[1] Cohen, Jeremy M., Elan Rosenfeld, and J. Zico Kolter. \"Certified adversarial robustness via randomized smoothing.\" arXiv preprint arXiv:1902.02918 (2019).\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603821693050}], "openreview_url": "https://openreview.net/forum?id=1AyPW2Emp6", "arxiv_id": "2010.10549", "paper_pdf": "papers/1AyPW2Emp6.pdf", "paper_pdf_sha256": "2641e8b2bb4d104d85069fcd8502bfb0d6d79206103d40ca3e3777f6fc8192f6", "paper_pdf_bytes": 5942360, "paper_pdf_source": "openreview", "code_url": "https://github.com/alevine0/smoothing_second_order", "code_repository": "alevine0/smoothing_second_order", "code_commit": "ae4b93c5666c3c967f28d62e770696fc2a417435", "code_archive": "repos/1AyPW2Emp6.zip", "code_archive_sha256": "ecec5e74447b6fb7ef37cc7605032c64ae9353d88899bcf93dbaa810f31a0d97", "code_archive_bytes": 2818537, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 2703, "github_languages": {"Python": 85059}, "github_archived": false, "github_pushed_at": "2022-08-22T15:50:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tight-second-order-certificates-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "8NtEHw8c8C", "year": 2026, "status": "rejected", "title": "Probing Neural Topology of Large Language Models", "authors": ["Yu Zheng", "Yuan Yuan", "Yue Zhuo", "Yong Li", "Paolo Santi"], "authorids": ["~Yu_Zheng7", "~Yuan_Yuan15", "~Yue_Zhuo2", "~Yong_Li7", "~Paolo_Santi2"], "authors_source": "OpenReview API", "abstract": "Probing large language models (LLMs) has yielded valuable insights into their internal mechanisms by linking neural activations to interpretable semantics. However, the complex mechanisms that link neuron’s functional co-activation with the emergent model capabilities remains largely unknown, hindering a deeper understanding and safer development of LLMs. In this work, we introduce graph probing,\na method for uncovering the functional connectivity of LLM neurons and relating it to language generation performance. By probing models across diverse LLM families and scales, we discover a universal predictability of next-token prediction performance using only neural topology, which persists even when retaining just 1% of neuron connections. Strikingly, probing on topology outperforms probing on activation by up to 130.4% and 67.7% on perplexity and space/time semantic regression respectively, suggesting that neural topology contains orders of richer information of LLM performance than neural activation, which can be easily extracted with simple linear or MLP probes. To explain the dependence between neural topology and language performance, we identify default networks and hub neurons in LLMs and provide causal evidence by interventional experiments on multiple benchmarks, showing that LLMs actually exploit these topological information. Further analyses suggest that neural topology can be effectively leveraged to improve the efficiency, reliability, and safety of LLMs through proof-of-concept applications in model pruning, hallucination detection, and LLM fingerprinting. Codes and data for the graph probing toolbox are available at https://anonymous.4open.science/r/llm-graph-probing-71BD/.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "c7K57MysHx", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19330/Reviewer_TRCD"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "This study proposes a novel method called Graph Probing, which analyzes the functional connectivity topology among neurons in LLMs to uncover its deep relationship with language generation performance. By adapting concepts from cognitive neuroscience, particularly functional connectivity and functional brain networks, the work introduces a fresh and innovative perspective to LLM interpretability. Empirical results reveal that LLMs exhibit stable default functional networks and hub neurons, and that the topological organization of these networks significantly influences model performance. The approach also shows promising potential for practical applications, such as model pruning, hallucination detection, and model fingerprinting.", "review_text": "This study proposes a novel method called Graph Probing, which analyzes the functional connectivity topology among neurons in LLMs to uncover its deep relationship with language generation performance. By adapting concepts from cognitive neuroscience, particularly functional connectivity and functional brain networks, the work introduces a fresh and innovative perspective to LLM interpretability. Empirical results reveal that LLMs exhibit stable default functional networks and hub neurons, and that the topological organization of these networks significantly influences model performance. The approach also shows promising potential for practical applications, such as model pruning, hallucination detection, and model fingerprinting.", "strengths": "1.The integration of functional connectivity concepts from neuroscience into LLM analysis is conceptually innovative and may contribute to mechanistic interpretability.  \n2.The authors have demonstrated that functional connectivity outperforms activation in predicting model performance, i.e., perplexity.\n3.The authors have demonstrated the promising potential of “graph probing” for practical applications, such as model pruning, hallucination detection, and model fingerprinting.", "weaknesses": "1.The manuscript is kind of overloaded while the validation of each point seems insufficient. \n2.Some interpretations of the results might be overstated.\n3.Experimental results on large scale models is missing.", "questions": "1.The manuscript is kind of overloaded. The authors tried to put too many points in one study, including graph probing, hub neurons, default network, functional specifications and three potential applications. All the points seem relate to “topology”. However, the limit of space makes the validation of each point to some extent insufficient. I would like to suggest the authors to focus on the main findings and elaborate experimental details, results and interpretations.\n2.I am curious about the relationship between graph probing and hub neurons. In graph probing, it is practical to identify key connections that contribute significantly to the prediction of model performance (perplexity here). Are the neurons that constitute those important connections align with hub neurons? Or to what extent the nodes in important connections overlap with hub neurons?\n3.All reported results are about small-scale LLMs. Given the focus on large models, results on at least 7B-class architectures are essential. Although the appendix mentions experiments on Qwen2.5-14B, no results, metrics, or analysis from this model are reported.   \n4.About functional specification: Given extremely high correlation of neural topology (Figure 8b), I not very sure that this results show the functional specification or the default network. Actually, if you don’t put “law” in the middle, or apply a color bar with a wider range, probably the clustering effective would not be that clear.\n\nMinor concerns:\n1.The mathematical definition of Neural Topology in Section 2 is incomplete: key symbols such as n (neuron index?) and t (time step? input token?) are not defined, undermining clarity and reproducibility.  \n2.A lot of the interpretation in the manuscript might be overstated. For example, but not being limited to:\n“This remarkable stability confirms the existence of a default network in LLMs, where a fixed set of hub neurons play dominant roles regardless of the input, suggesting a fundamental organizing principle of their internal structure.” ---The DMN is a quite complex system in the brain. It’s functionality is far beyond “play dominant roles regardless of the input”.\n“These findings strongly suggest that LLMs develop distinct and linearly separable topological patterns for different knowledge domains, allowing for easy extraction of the context or subject being processed.”---The experimental results do not support this claim. Separable topological patterns should be demonstrated before drawing such a strong conclusion.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This study proposes a novel method called Graph Probing, which analyzes the functional connectivity topology among neurons in LLMs to uncover its deep relationship with language generation performance. By adapting concepts from cognitive neuroscience, particularly functional connectivity and functional brain networks, the work introduces a fresh and innovative perspective to LLM interpretability. Empirical results reveal that LLMs exhibit stable default functional networks and hub neurons, and that the topological organization of these networks significantly influences model performance. The approach also shows promising potential for practical applications, such as model pruning, hallucination detection, and model fingerprinting.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1.The integration of functional connectivity concepts from neuroscience into LLM analysis is conceptually innovative and may contribute to mechanistic interpretability.  \n2.The authors have demonstrated that functional connectivity outperforms activation in predicting model performance, i.e., perplexity.\n3.The authors have demonstrated the promising potential of “graph probing” for practical applications, such as model pruning, hallucination detection, and model fingerprinting.", "weaknesses": "1.The manuscript is kind of overloaded while the validation of each point seems insufficient. \n2.Some interpretations of the results might be overstated.\n3.Experimental results on large scale models is missing.", "questions": "1.The manuscript is kind of overloaded. The authors tried to put too many points in one study, including graph probing, hub neurons, default network, functional specifications and three potential applications. All the points seem relate to “topology”. However, the limit of space makes the validation of each point to some extent insufficient. I would like to suggest the authors to focus on the main findings and elaborate experimental details, results and interpretations.\n2.I am curious about the relationship between graph probing and hub neurons. In graph probing, it is practical to identify key connections that contribute significantly to the prediction of model performance (perplexity here). Are the neurons that constitute those important connections align with hub neurons? Or to what extent the nodes in important connections overlap with hub neurons?\n3.All reported results are about small-scale LLMs. Given the focus on large models, results on at least 7B-class architectures are essential. Although the appendix mentions experiments on Qwen2.5-14B, no results, metrics, or analysis from this model are reported.   \n4.About functional specification: Given extremely high correlation of neural topology (Figure 8b), I not very sure that this results show the functional specification or the default network. Actually, if you don’t put “law” in the middle, or apply a color bar with a wider range, probably the clustering effective would not be that clear.\n\nMinor concerns:\n1.The mathematical definition of Neural Topology in Section 2 is incomplete: key symbols such as n (neuron index?) and t (time step? input token?) are not defined, undermining clarity and reproducibility.  \n2.A lot of the interpretation in the manuscript might be overstated. For example, but not being limited to:\n“This remarkable stability confirms the existence of a default network in LLMs, where a fixed set of hub neurons play dominant roles regardless of the input, suggesting a fundamental organizing principle of their internal structure.” ---The DMN is a quite complex system in the brain. It’s functionality is far beyond “play dominant roles regardless of the input”.\n“These findings strongly suggest that LLMs develop distinct and linearly separable topological patterns for different knowledge domains, allowing for easy extraction of the context or subject being processed.”---The experimental results do not support this claim. Separable topological patterns should be demonstrated before drawing such a strong conclusion.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1761911235841}, {"id": "cB4myELaI0", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19330/Reviewer_RV8o"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "The paper introduces graph probing, a probing method that treats an LLM’s neurons as nodes in a functional connectivity graph derived from token-time activation. Instead of probing raw activations, the method flattens the adjacency matrix of functional connectivity graph and trains lightweight MLP (and GNN) probes to predict LLM performance metrics such as perplexity. By probing models across diverse LLM families, the paper argues that the technique can predict next-token prediction performance using only neural topology.  They report gains up to 130%.\n\nBeyond correlation, the paper probes causal structure in neural topology. It identifies a stable “default network” of high-degree hub neurons that recur across inputs, then shows that ablating only the top ~1% of degree-ranked neurons causes drastic MMLU accuracy drops.\n\nFinally, the authors showcase practical applications of neural topology, including topology-informed pruning that removes many low-degree neurons with modest accuracy loss for downstream tasks.", "review_text": "The paper introduces graph probing, a probing method that treats an LLM’s neurons as nodes in a functional connectivity graph derived from token-time activation. Instead of probing raw activations, the method flattens the adjacency matrix of functional connectivity graph and trains lightweight MLP (and GNN) probes to predict LLM performance metrics such as perplexity. By probing models across diverse LLM families, the paper argues that the technique can predict next-token prediction performance using only neural topology.  They report gains up to 130%.\n\nBeyond correlation, the paper probes causal structure in neural topology. It identifies a stable “default network” of high-degree hub neurons that recur across inputs, then shows that ablating only the top ~1% of degree-ranked neurons causes drastic MMLU accuracy drops.\n\nFinally, the authors showcase practical applications of neural topology, including topology-informed pruning that removes many low-degree neurons with modest accuracy loss for downstream tasks.", "strengths": "I find the paper well written and the empirical results are solid and sound. In terms of the main result, it shows neural topology probes reliably predict next-token performance and outperform activation-only baselines across multiple LLM families/sizes, with careful 8:2 train/test splits and standard regression metrics. The experimental setups are valid.\n\nThe paper also shows that the results are robust. In particular, performance persists under heavy graph sparsification and across model sizes and layers, suggesting the signal isn’t a fragile artifact of a single scale. \n\nBesides correlation, the paper offers a causal and interpretability perspective. The author(s) identify a stable “default network” of hub neurons and show that ablating the top 1% (by degree) yields large MMLU drops. This is stronger than random or activation-based selections, suggesting that models actually use this topology. These insights yield practical applications such as sparsity pruning and hallucination detection on TruthfulQA. \n\nThe paper demonstrates that the technique is general. I appreciate the breath of the evaluations. It considers various model families and scales, including Qwen and Pythia series. In the appendix, it also describes experiments on VLMs by probing different sizes of LLaVA-v1.5. \n\nTo my knowledge, the observations from the paper is novel and I believe it's a nice contribution.", "weaknesses": "While the paper identifies a default network, I find it insufficient in explaining exactly what the network does. In addition, I'd love to see a mechanistic interpretation of “why” topology predicts performance.\n\nIf I understand correctly, building the per-sequence correlation graphs requires collecting full hidden-state time series and computing pairwise correlations, then training a probe. Although sparsification helps at inference time, the initial topology extraction can be costly in $n$ and $t$. How large is the computational overhead as we scale $n$ and $t$? Alternatively, it would be great to consider more efficient or approximate version of the algorithm.\n\nRelated to that, another minor concern is that the experiments of this paper top out at 14B parameter models. I am a bit worried about the computational cost of the technique. It would be nice to expand how much time / memory it would require to train & apply the probes as we scale model sizes. \n\nFor the Truthful QA experiment, the inputs are constructed by concatenating each question with its true or false answer. In practice, users interact with LLM in a free form way. Could the paper clarify if the their method can help hallucination detection in a realistic setting?", "questions": "I wonder how scalable the technique is. Specifically, the method requires you to construct a n by n adjacency graph. Would this still feasible for, say, trillion-parameter models?\n\nFor the pruning experiment, after removing low-degree neurons, did you try brief re-tuning (namely, a bit more fine-tuning) to recover accuracy?\n\nIn Appendix A.4, you used 10,000 random text sequences to construct neural connectivity graphs. I wonder if 10k is a universal constant that suffices for any models or scales, or if it should vary as well. In general, what is the sample complexity of training the graph probes? How many sequences are needed for the graph features to stabilize and the training loss of the probes to converge?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces graph probing, a probing method that treats an LLM’s neurons as nodes in a functional connectivity graph derived from token-time activation. Instead of probing raw activations, the method flattens the adjacency matrix of functional connectivity graph and trains lightweight MLP (and GNN) probes to predict LLM performance metrics such as perplexity. By probing models across diverse LLM families, the paper argues that the technique can predict next-token prediction performance using only neural topology.  They report gains up to 130%.\n\nBeyond correlation, the paper probes causal structure in neural topology. It identifies a stable “default network” of high-degree hub neurons that recur across inputs, then shows that ablating only the top ~1% of degree-ranked neurons causes drastic MMLU accuracy drops.\n\nFinally, the authors showcase practical applications of neural topology, including topology-informed pruning that removes many low-degree neurons with modest accuracy loss for downstream tasks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "I find the paper well written and the empirical results are solid and sound. In terms of the main result, it shows neural topology probes reliably predict next-token performance and outperform activation-only baselines across multiple LLM families/sizes, with careful 8:2 train/test splits and standard regression metrics. The experimental setups are valid.\n\nThe paper also shows that the results are robust. In particular, performance persists under heavy graph sparsification and across model sizes and layers, suggesting the signal isn’t a fragile artifact of a single scale. \n\nBesides correlation, the paper offers a causal and interpretability perspective. The author(s) identify a stable “default network” of hub neurons and show that ablating the top 1% (by degree) yields large MMLU drops. This is stronger than random or activation-based selections, suggesting that models actually use this topology. These insights yield practical applications such as sparsity pruning and hallucination detection on TruthfulQA. \n\nThe paper demonstrates that the technique is general. I appreciate the breath of the evaluations. It considers various model families and scales, including Qwen and Pythia series. In the appendix, it also describes experiments on VLMs by probing different sizes of LLaVA-v1.5. \n\nTo my knowledge, the observations from the paper is novel and I believe it's a nice contribution.", "weaknesses": "While the paper identifies a default network, I find it insufficient in explaining exactly what the network does. In addition, I'd love to see a mechanistic interpretation of “why” topology predicts performance.\n\nIf I understand correctly, building the per-sequence correlation graphs requires collecting full hidden-state time series and computing pairwise correlations, then training a probe. Although sparsification helps at inference time, the initial topology extraction can be costly in $n$ and $t$. How large is the computational overhead as we scale $n$ and $t$? Alternatively, it would be great to consider more efficient or approximate version of the algorithm.\n\nRelated to that, another minor concern is that the experiments of this paper top out at 14B parameter models. I am a bit worried about the computational cost of the technique. It would be nice to expand how much time / memory it would require to train & apply the probes as we scale model sizes. \n\nFor the Truthful QA experiment, the inputs are constructed by concatenating each question with its true or false answer. In practice, users interact with LLM in a free form way. Could the paper clarify if the their method can help hallucination detection in a realistic setting?", "questions": "I wonder how scalable the technique is. Specifically, the method requires you to construct a n by n adjacency graph. Would this still feasible for, say, trillion-parameter models?\n\nFor the pruning experiment, after removing low-degree neurons, did you try brief re-tuning (namely, a bit more fine-tuning) to recover accuracy?\n\nIn Appendix A.4, you used 10,000 random text sequences to construct neural connectivity graphs. I wonder if 10k is a universal constant that suffices for any models or scales, or if it should vary as well. In general, what is the sample complexity of training the graph probes? How many sequences are needed for the graph features to stabilize and the training loss of the probes to converge?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761856438578}, {"id": "I2iu6g7Yok", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission19330/Reviewer_SqX2"], "rating": 2, "soundness": 1, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new method for probing, inspired by methods in neuroscience. Using this new method, which they term *graph probing*, one computes the pairwise Pearson correlations between neurons across timesteps, and feeds this into a linear or MLP probe. Using this method, the authors are able to more successfully probe for quantities like model perplexity on a sequence, as compared to standard probes. Interpreting the values in the correlation matrix as a graph, they also find hub neurons (those with high degree), and show that ablating these disproportionately hurts model performance, indicating their functional importance. They finally use these probes for pruning, hallucination detection, and model fingerprinting.", "review_text": "This paper introduces a new method for probing, inspired by methods in neuroscience. Using this new method, which they term *graph probing*, one computes the pairwise Pearson correlations between neurons across timesteps, and feeds this into a linear or MLP probe. Using this method, the authors are able to more successfully probe for quantities like model perplexity on a sequence, as compared to standard probes. Interpreting the values in the correlation matrix as a graph, they also find hub neurons (those with high degree), and show that ablating these disproportionately hurts model performance, indicating their functional importance. They finally use these probes for pruning, hallucination detection, and model fingerprinting.", "strengths": "This paper is well-written and polished: the text is clear, and the graphs are well-constructed. I also think that this particularly style of probing is novel: I haven't heard of others using it before.", "weaknesses": "**Unclear Purpose of Proposed Method**: The primary use of probing in interpretability is, as noted by this paper, \"linking neural activations to interpretable semantics\". But this paper never actually does this, which is the one thing that one would expect from a probing paper. In fact, since it's restricted to sequence-level probing, it can't do many probing tasks, which are often token-level. \n\nInstead, it probes for things like perplexity (why would you want a probe to predict that?) and uses probes to identify important neurons (we know that such neurons exist, and don't need probes to find them), identify hallucinations (this is useful, but there's a lot of past work to compare to), and model fingerprinting (same issue). Ultimately, an interpretability method should help us understand better how models work, but I don't think this paper does so; instead, it focuses on applications that appear useful on the surface, but are hamstrung by insufficient baselines / connection to prior work. In doing so, it ends up in an awkward position where it doesn't make a very novel or significant contribution to either model interpretability or capabilities. \n\n**Weak Evaluation / Insufficient Baselines**: This paper doesn't perform enough / the right baselines, making its results less sound:\n- For the probing for perplexity experiments, the baseline probes only receive the last token representation. You should give them something that combines information from all of the positions; it doesn't make sense to baseline against a method that receives so much less information in its input.\n- For the pruning experiments, why not compare to any other pruning methods? If the point of this method is to be useful, then it should really be more useful than alternatives.\n- Ditto for the hallucination experiments.\n- Ditto again for the model fingerprinting experiments. I'd also love more discussion about how / whether the Pythia experiments really reflect a real-world fingerprinting use-case. \n\n**Relation to prior work**: This paper actually cites a lot of prior work! But I think that it misses out on some crucial work that would provide context. For example: \n- this paper stresses the finding of so-called hub neurons, but it is well-known that neurons vary in importance with respect to individual tasks; see [Vig and Belinkov](https://proceedings.neurips.cc/paper/2020/hash/92650b2e92217715fe312e6fa7b90d82-Abstract.html) who find that just a few neurons matter for a gender bias task. For papers looking at this from a more general POV, see [Timkey et al](https://aclanthology.org/2021.emnlp-main.372/), who investigate neurons whose activations are very high-magnitude, and [Dettmers et al](https://arxiv.org/abs/2208.07339), who find that not quantizing these neurons is important, as doing so hurts model performance. It follows fairly naturally that zero ablating these - a much more destructive intervention than quantization - will hurt model performance as well. The authors push this \"hub neuron\" / \"default network\" narrative very hard, but these findings are already known, and the neuro connection seems forced.\n- Where is the related work on pruning / hallucinations / model fingerprinting? For the first two cases, there has been tons of interp work to compare to. \n- Re: probing, please cite work from before 2022! There is a lot of thoughtful probing literature that predates probing in mech interp; indeed, a lot of that probing literature is better than what has emerged recently. You can look at [BERTology](https://aclanthology.org/2020.tacl-1.54/) for classic probing papers, but see [Hewitt and Liang](https://aclanthology.org/D19-1275/) for some thoughts on how to do probing, and e.g. [Voita et al.](https://aclanthology.org/2020.emnlp-main.14/) and [Pimentel et al.](https://aclanthology.org/2020.acl-main.420/) for other probing methods. Connecting more to the probing literature might help shed some light on important questions like: why do we do probing in the first place? What are we looking to find, and what would a new method have to bring, in order to be useful?\n- I think the authors are much too sanguine re: the results of NeuroAI, and this shows in their review of prior work. Maybe be a bit more cautious in describing its putative successes.", "questions": "- Why do you refer to \"neural topology\"? I expected something about e.g. the topology of neural activations in the model, but it's actually just correlations between neurons over a given input.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new method for probing, inspired by methods in neuroscience. Using this new method, which they term *graph probing*, one computes the pairwise Pearson correlations between neurons across timesteps, and feeds this into a linear or MLP probe. Using this method, the authors are able to more successfully probe for quantities like model perplexity on a sequence, as compared to standard probes. Interpreting the values in the correlation matrix as a graph, they also find hub neurons (those with high degree), and show that ablating these disproportionately hurts model performance, indicating their functional importance. They finally use these probes for pruning, hallucination detection, and model fingerprinting.", "soundness": 1, "presentation": 4, "contribution": 2, "strengths": "This paper is well-written and polished: the text is clear, and the graphs are well-constructed. I also think that this particularly style of probing is novel: I haven't heard of others using it before.", "weaknesses": "**Unclear Purpose of Proposed Method**: The primary use of probing in interpretability is, as noted by this paper, \"linking neural activations to interpretable semantics\". But this paper never actually does this, which is the one thing that one would expect from a probing paper. In fact, since it's restricted to sequence-level probing, it can't do many probing tasks, which are often token-level. \n\nInstead, it probes for things like perplexity (why would you want a probe to predict that?) and uses probes to identify important neurons (we know that such neurons exist, and don't need probes to find them), identify hallucinations (this is useful, but there's a lot of past work to compare to), and model fingerprinting (same issue). Ultimately, an interpretability method should help us understand better how models work, but I don't think this paper does so; instead, it focuses on applications that appear useful on the surface, but are hamstrung by insufficient baselines / connection to prior work. In doing so, it ends up in an awkward position where it doesn't make a very novel or significant contribution to either model interpretability or capabilities. \n\n**Weak Evaluation / Insufficient Baselines**: This paper doesn't perform enough / the right baselines, making its results less sound:\n- For the probing for perplexity experiments, the baseline probes only receive the last token representation. You should give them something that combines information from all of the positions; it doesn't make sense to baseline against a method that receives so much less information in its input.\n- For the pruning experiments, why not compare to any other pruning methods? If the point of this method is to be useful, then it should really be more useful than alternatives.\n- Ditto for the hallucination experiments.\n- Ditto again for the model fingerprinting experiments. I'd also love more discussion about how / whether the Pythia experiments really reflect a real-world fingerprinting use-case. \n\n**Relation to prior work**: This paper actually cites a lot of prior work! But I think that it misses out on some crucial work that would provide context. For example: \n- this paper stresses the finding of so-called hub neurons, but it is well-known that neurons vary in importance with respect to individual tasks; see [Vig and Belinkov](https://proceedings.neurips.cc/paper/2020/hash/92650b2e92217715fe312e6fa7b90d82-Abstract.html) who find that just a few neurons matter for a gender bias task. For papers looking at this from a more general POV, see [Timkey et al](https://aclanthology.org/2021.emnlp-main.372/), who investigate neurons whose activations are very high-magnitude, and [Dettmers et al](https://arxiv.org/abs/2208.07339), who find that not quantizing these neurons is important, as doing so hurts model performance. It follows fairly naturally that zero ablating these - a much more destructive intervention than quantization - will hurt model performance as well. The authors push this \"hub neuron\" / \"default network\" narrative very hard, but these findings are already known, and the neuro connection seems forced.\n- Where is the related work on pruning / hallucinations / model fingerprinting? For the first two cases, there has been tons of interp work to compare to. \n- Re: probing, please cite work from before 2022! There is a lot of thoughtful probing literature that predates probing in mech interp; indeed, a lot of that probing literature is better than what has emerged recently. You can look at [BERTology](https://aclanthology.org/2020.tacl-1.54/) for classic probing papers, but see [Hewitt and Liang](https://aclanthology.org/D19-1275/) for some thoughts on how to do probing, and e.g. [Voita et al.](https://aclanthology.org/2020.emnlp-main.14/) and [Pimentel et al.](https://aclanthology.org/2020.acl-main.420/) for other probing methods. Connecting more to the probing literature might help shed some light on important questions like: why do we do probing in the first place? What are we looking to find, and what would a new method have to bring, in order to be useful?\n- I think the authors are much too sanguine re: the results of NeuroAI, and this shows in their review of prior work. Maybe be a bit more cautious in describing its putative successes.", "questions": "- Why do you refer to \"neural topology\"? I expected something about e.g. the topology of neural activations in the model, but it's actually just correlations between neurons over a given input.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1760973892849}], "openreview_url": "https://openreview.net/forum?id=8NtEHw8c8C", "arxiv_id": "2506.01042", "paper_pdf": "papers/8NtEHw8c8C.pdf", "paper_pdf_sha256": "b6ae534da36f3dc7c6198bfd532c43ef7274663bcb4826c9fa0fd1ec303e4d7e", "paper_pdf_bytes": 2985489, "paper_pdf_source": "openreview", "code_url": "https://github.com/DavyMorgan/llm-graph-probing", "code_repository": "DavyMorgan/llm-graph-probing", "code_commit": "370e592d67aee6bb4bd3ccf38999476094788d4f", "code_archive": "repos/8NtEHw8c8C.zip", "code_archive_sha256": "5157a7a4c381783e7f252f109bf20e653a5fc9578bddcf981d412ec33ec2d1b1", "code_archive_bytes": 3776330, "code_file_count": 34, "code_extensions": {".py": 34}, "github_disk_usage_kb": 8626, "github_languages": {"Python": 218668}, "github_archived": false, "github_pushed_at": "2026-02-27T02:34:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/probing-neural-topology-of-large-language"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EXaKfdsw04", "year": 2025, "status": "rejected", "title": "StepProof: Step-by-step verification of natural language mathematical proofs", "authors": ["Xiaolin Hu", "Qinghua Zhou", "Bogdan Grechuk", "Ivan Y Tyukin", "Oliver Sutton"], "authorids": ["~Xiaolin_Hu7", "~Qinghua_Zhou1", "~Bogdan_Grechuk1", "~Ivan_Y_Tyukin1", "~Oliver_Sutton1"], "authors_source": "OpenReview API", "abstract": "Interactive theorem provers (ITPs) are powerful tools for the formal verification of mathematical proofs down to the axiom level. However, their lack of a natural language interface remains a significant limitation. Recent advancements in large language models (LLMs) have enhanced the understanding of natural language inputs, paving the way for autoformalization—the process of translating natural language proofs into formal proofs that can be verified. Despite these advancements, existing autoformalization approaches are limited to verifying complete proofs and lack the capability for finer, sentence-level verification. To address this gap, we propose StepProof, a novel autoformalization method designed for granular, step-by-step verification. StepProof breaks down complete proofs into multiple verifiable subproofs, enabling sentence-level verification. Experimental results demonstrate that StepProof significantly improves proof success rates and efficiency compared to traditional methods. Additionally, we found that minor manual adjustments to the natural language proofs, tailoring them for step-level verification, further enhanced StepProof’s performance in autoformalization.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "HYVFuXH3nt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7361/Reviewer_nEet"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces StepProof, a method designed to improve the verification of natural language mathematical proofs by breaking them down into smaller, verifiable subproofs. Unlike traditional autoformalization methods that only verify complete proofs, StepProof operates on a sentence level, enabling granular verification of each step. This approach aims to enhance the efficiency and accuracy of proof verification by targeting and resolving errors within individual steps, rather than regenerating entire proofs. Experimental results indicate that StepProof outperforms other approaches in proof success rates and efficiency. Minor manual modifications to the proofs, aligning them with step-level requirements, were also found to enhance the performance of StepProof.", "review_text": "The paper introduces StepProof, a method designed to improve the verification of natural language mathematical proofs by breaking them down into smaller, verifiable subproofs. Unlike traditional autoformalization methods that only verify complete proofs, StepProof operates on a sentence level, enabling granular verification of each step. This approach aims to enhance the efficiency and accuracy of proof verification by targeting and resolving errors within individual steps, rather than regenerating entire proofs. Experimental results indicate that StepProof outperforms other approaches in proof success rates and efficiency. Minor manual modifications to the proofs, aligning them with step-level requirements, were also found to enhance the performance of StepProof.", "strengths": "- Granular Verification Approach: StepProof’s method of breaking down proofs into verifiable sub-propositions is a notable advancement compared to traditional methods that require verifying complete proofs. This could allow for a more targeted approach to error correction by addressing specific subproofs.\n- Improved Success Rates and Efficiency: The experimental findings indicate that StepProof improves both success rates and efficiency for proof verification over baseline methods. By enabling sentence-level verification, StepProof appears to make the autoformalization process more manageable and less resource-intensive, which could be beneficial for large-scale applications.\n- Enhanced Usability: The authors found that the minor adjustments to natural language proofs tailored for step-level verification, which was made easier by the GUI interfaces, further enhance StepProof’s performance, suggesting practical guidance for users to maximize the method's effectiveness.", "weaknesses": "- Unsubstantiated Claim of Advantage: The claimed advantage of StepProof’s selective error correction (where only erroneous steps are retracted rather than the entire proof) is not unique to StepProof. Interactive theorem provers (ITPs) inherently support stepwise correction, enabling users to fix specific errors without requiring a full retraction. Thus, StepProof’s advantage in this aspect appears overstated.\n- Lack of Novelty in Stepwise Translation: Stepwise translation in autoformalization is not a new concept. Previous methods, like the DSP approach, have already implemented similar methodologies. These methods translate decomposed proof steps, whether generated by an LLM or provided by a human, indicating that StepProof may not be as innovative as claimed in this area.\n- Overly Restrictive Assumptions: StepProof’s framework assumes that each sentence in a proof can be treated as an independent, verifiable sub-proposition, which limits its applicability. This subgoal-based approach does not align well with many natural logical structures in proofs, especially those involving complex logical dependencies or sequence reordering. As a result, StepProof might require significant manual adjustments for compatibility with common proof structures.\n- Ambiguous Evaluation Methodology: The incremental verification of sub-propositions might count intermediary, potentially incorrect and mathematically misleading results as “correct” sub-proofs, leading to an inaccurate reflection of the method's overall success. This evaluation ambiguity calls into question the validity of the reported improvements in success rates.\n- Inappropriate Benchmark Dataset: The authors’ choice of GSM8K as a benchmark dataset is unsuitable for evaluating proof autoformalization due to its relative simplicity and lack of complex logical structures. Datasets like ProofNet or MiniF2F would provide a more accurate measure of StepProof’s performance on challenging, real-world mathematical proofs, better reflecting its practical value in formalization tasks.\n- Insufficient Detail on Methodology: It is also worth noting that further clarification on the prompting and output syntax for both StepProof and Full-Proof in the comparison experiments is needed. More specific details on the handling of the LLM’s guessed proof states could be valuable for assessing the validity and replicability of the reported performance differences between the two approaches.", "questions": "- Consider Additional Benchmark Datasets: Suggest that the authors include additional, more complex datasets like ProofNet or MiniF2F in future evaluations.\n- Highlight Distinctions from Prior Work on Stepwise Translation: Advise the authors to address the similarities between StepProof and prior stepwise translation methods.\n- How does StepProof handle complex proof structures outside the subgoal-based framework?\nCould the authors elaborate on StepProof’s limitations in handling complex, non-linear proof structures? For example, how does it manage proofs that involve nested assumptions, indirect reasoning, or statements that need reordering for coherence?\n- How does StepProof compare to Whole-Proof when using more informative LLM feedback?\nIn Whole-Proof, does the LLM output proof states after each line, or is this feature limited to StepProof? Allowing Whole-Proof access to these proof states could potentially improve its success rate. Could the authors provide more details on the prompting strategies for both methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces StepProof, a method designed to improve the verification of natural language mathematical proofs by breaking them down into smaller, verifiable subproofs. Unlike traditional autoformalization methods that only verify complete proofs, StepProof operates on a sentence level, enabling granular verification of each step. This approach aims to enhance the efficiency and accuracy of proof verification by targeting and resolving errors within individual steps, rather than regenerating entire proofs. Experimental results indicate that StepProof outperforms other approaches in proof success rates and efficiency. Minor manual modifications to the proofs, aligning them with step-level requirements, were also found to enhance the performance of StepProof.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- Granular Verification Approach: StepProof’s method of breaking down proofs into verifiable sub-propositions is a notable advancement compared to traditional methods that require verifying complete proofs. This could allow for a more targeted approach to error correction by addressing specific subproofs.\n- Improved Success Rates and Efficiency: The experimental findings indicate that StepProof improves both success rates and efficiency for proof verification over baseline methods. By enabling sentence-level verification, StepProof appears to make the autoformalization process more manageable and less resource-intensive, which could be beneficial for large-scale applications.\n- Enhanced Usability: The authors found that the minor adjustments to natural language proofs tailored for step-level verification, which was made easier by the GUI interfaces, further enhance StepProof’s performance, suggesting practical guidance for users to maximize the method's effectiveness.", "weaknesses": "- Unsubstantiated Claim of Advantage: The claimed advantage of StepProof’s selective error correction (where only erroneous steps are retracted rather than the entire proof) is not unique to StepProof. Interactive theorem provers (ITPs) inherently support stepwise correction, enabling users to fix specific errors without requiring a full retraction. Thus, StepProof’s advantage in this aspect appears overstated.\n- Lack of Novelty in Stepwise Translation: Stepwise translation in autoformalization is not a new concept. Previous methods, like the DSP approach, have already implemented similar methodologies. These methods translate decomposed proof steps, whether generated by an LLM or provided by a human, indicating that StepProof may not be as innovative as claimed in this area.\n- Overly Restrictive Assumptions: StepProof’s framework assumes that each sentence in a proof can be treated as an independent, verifiable sub-proposition, which limits its applicability. This subgoal-based approach does not align well with many natural logical structures in proofs, especially those involving complex logical dependencies or sequence reordering. As a result, StepProof might require significant manual adjustments for compatibility with common proof structures.\n- Ambiguous Evaluation Methodology: The incremental verification of sub-propositions might count intermediary, potentially incorrect and mathematically misleading results as “correct” sub-proofs, leading to an inaccurate reflection of the method's overall success. This evaluation ambiguity calls into question the validity of the reported improvements in success rates.\n- Inappropriate Benchmark Dataset: The authors’ choice of GSM8K as a benchmark dataset is unsuitable for evaluating proof autoformalization due to its relative simplicity and lack of complex logical structures. Datasets like ProofNet or MiniF2F would provide a more accurate measure of StepProof’s performance on challenging, real-world mathematical proofs, better reflecting its practical value in formalization tasks.\n- Insufficient Detail on Methodology: It is also worth noting that further clarification on the prompting and output syntax for both StepProof and Full-Proof in the comparison experiments is needed. More specific details on the handling of the LLM’s guessed proof states could be valuable for assessing the validity and replicability of the reported performance differences between the two approaches.", "questions": "- Consider Additional Benchmark Datasets: Suggest that the authors include additional, more complex datasets like ProofNet or MiniF2F in future evaluations.\n- Highlight Distinctions from Prior Work on Stepwise Translation: Advise the authors to address the similarities between StepProof and prior stepwise translation methods.\n- How does StepProof handle complex proof structures outside the subgoal-based framework?\nCould the authors elaborate on StepProof’s limitations in handling complex, non-linear proof structures? For example, how does it manage proofs that involve nested assumptions, indirect reasoning, or statements that need reordering for coherence?\n- How does StepProof compare to Whole-Proof when using more informative LLM feedback?\nIn Whole-Proof, does the LLM output proof states after each line, or is this feature limited to StepProof? Allowing Whole-Proof access to these proof states could potentially improve its success rate. Could the authors provide more details on the prompting strategies for both methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730720652588}, {"id": "SuIosF945y", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7361/Reviewer_EFnZ"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 1, "confidence": 4, "summary": "The paper proposes generating formal proofs from informal ones in a step-by-step manner and shows that it outperforms one-time proof generation.", "review_text": "The paper proposes generating formal proofs from informal ones in a step-by-step manner and shows that it outperforms one-time proof generation.", "strengths": "+ The paper applies the known idea that, for LLMs, the step-by-step logical reasoning works more well than one-step reasoning, to autoformalization, especially the formal proof generation.", "weaknesses": "- It is convincing that the step-by-step reasoning works well, but I feel it natural and unsurprising because it has been proven that LLMs can better carry out step-by-step logical reasoning than one-step [1]. Thus, I'm unsure how significant the contribution of the paper, which seems to confirm step-by-step proof generation works more well than one-step generation, is.\n- Furthermore, it seems that the paper is not the first to apply the step-by-step reasoning power of LLMs to formal proof generation. For example, LEGO-Prover [2] decomposes informal proofs into step-by-step informal proofs with sub-goals and then proves the generated sub-goals. Although the main aim of LEGO-Prover is to address growing libraries, the paper does not theoretically, empirically, qualitatively, nor quantitatively compare the proposed approach with such existing approaches that exploit the step-by-step reasoning ability of LLMs.\n\n[1] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, Denny Zhou: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022\n\n[2] Haiming Wang, Huajian Xin, Chuanyang Zheng, Zhengying Liu, Qingxing Cao, Yinya Huang, Jing Xiong, Han Shi, Enze Xie, Jian Yin, Zhenguo Li, Xiaodan Liang:\nLEGO-Prover: Neural Theorem Proving with Growing Libraries. ICLR 2024\n\nOther comments and concerns\n\n- Please insert spacing before citations, like \"the large language model(Zhao et al., 2023)\" --> \"... model (Zhao et al., 2023)\" in line 44. I suggest reviewing all the citations to avoid similar issues.\n- l88 \"A lot of work has also shown that ...\" Please cite the papers showing it.\n- l100 \"bert\" --> BERT?\n- l106 \"DTV\" Please explain the abbreviation where it appears first.\n- l110 \"Qinghua et al.\" No link to the citation.\n- l155-160: It would be nice to explain the problem with concrete examples that enable the reader to easily find the issue in FULL-PROOF strategies.\n- l161 \"max_new_tokens\" No explanation about this.\n- l338 \"new index\" I can't find what this is.", "questions": "- Is it possible to explain, prove, and/or demonstrate how the application of step-by-step reasoning in proof generation differs from or advances beyond the previous work like [1,2]?\n- Table 2: What's \"Comments Rate\"?\n- Table 3: $r_s$ is a step pass rate?\n- l351 \"simple fitting\" Is the fitting necessary? It seems to mean one needs to tune the proofs.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes generating formal proofs from informal ones in a step-by-step manner and shows that it outperforms one-time proof generation.", "soundness": 2, "presentation": 1, "contribution": 1, "strengths": "+ The paper applies the known idea that, for LLMs, the step-by-step logical reasoning works more well than one-step reasoning, to autoformalization, especially the formal proof generation.", "weaknesses": "- It is convincing that the step-by-step reasoning works well, but I feel it natural and unsurprising because it has been proven that LLMs can better carry out step-by-step logical reasoning than one-step [1]. Thus, I'm unsure how significant the contribution of the paper, which seems to confirm step-by-step proof generation works more well than one-step generation, is.\n- Furthermore, it seems that the paper is not the first to apply the step-by-step reasoning power of LLMs to formal proof generation. For example, LEGO-Prover [2] decomposes informal proofs into step-by-step informal proofs with sub-goals and then proves the generated sub-goals. Although the main aim of LEGO-Prover is to address growing libraries, the paper does not theoretically, empirically, qualitatively, nor quantitatively compare the proposed approach with such existing approaches that exploit the step-by-step reasoning ability of LLMs.\n\n[1] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, Denny Zhou: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022\n\n[2] Haiming Wang, Huajian Xin, Chuanyang Zheng, Zhengying Liu, Qingxing Cao, Yinya Huang, Jing Xiong, Han Shi, Enze Xie, Jian Yin, Zhenguo Li, Xiaodan Liang:\nLEGO-Prover: Neural Theorem Proving with Growing Libraries. ICLR 2024\n\nOther comments and concerns\n\n- Please insert spacing before citations, like \"the large language model(Zhao et al., 2023)\" --> \"... model (Zhao et al., 2023)\" in line 44. I suggest reviewing all the citations to avoid similar issues.\n- l88 \"A lot of work has also shown that ...\" Please cite the papers showing it.\n- l100 \"bert\" --> BERT?\n- l106 \"DTV\" Please explain the abbreviation where it appears first.\n- l110 \"Qinghua et al.\" No link to the citation.\n- l155-160: It would be nice to explain the problem with concrete examples that enable the reader to easily find the issue in FULL-PROOF strategies.\n- l161 \"max_new_tokens\" No explanation about this.\n- l338 \"new index\" I can't find what this is.", "questions": "- Is it possible to explain, prove, and/or demonstrate how the application of step-by-step reasoning in proof generation differs from or advances beyond the previous work like [1,2]?\n- Table 2: What's \"Comments Rate\"?\n- Table 3: $r_s$ is a step pass rate?\n- l351 \"simple fitting\" Is the fitting necessary? It seems to mean one needs to tune the proofs.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730708116236}, {"id": "FnpXqIveE0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7361/Reviewer_pXFu"], "rating": 1, "soundness": 1, "presentation": 1, "contribution": 1, "confidence": 5, "summary": "This paper describes a framework for granular, step-by-step verification of natural language reasoning.", "review_text": "This paper describes a framework for granular, step-by-step verification of natural language reasoning.", "strengths": "N.A.", "weaknesses": "- The paper is almost impossible to comprehend. I highly suspect a big portion of it is generated by LLMs.\n- The full-proof strategy so baffling: I am not even sure what is formal backend and what is role of users here.", "questions": "N.A.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper describes a framework for granular, step-by-step verification of natural language reasoning.", "soundness": 1, "presentation": 1, "contribution": 1, "strengths": "N.A.", "weaknesses": "- The paper is almost impossible to comprehend. I highly suspect a big portion of it is generated by LLMs.\n- The full-proof strategy so baffling: I am not even sure what is formal backend and what is role of users here.", "questions": "N.A.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 1, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730672988197}, {"id": "ShlXeXZwj9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7361/Reviewer_rYkN"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper proposes a method for step-by-step verification of natural language mathematical proofs. Unlike traditional \"full-proof\" approaches, which formalize and verify entire proofs at once, the proposed method (\"StepProof\") breaks down proofs into smaller, sentence-level subproofs to perform more granular verification. This compositional approach also allows for an intereactive user experience as well as error handling by allowing verification of individual steps and retrying only failed steps without discarding the entire proof. An evaluation is shown that the StepProof strategy performs better than a full-proof strategy and two existing baseline approaches for auto formalization on the GSM8K dataset.", "review_text": "This paper proposes a method for step-by-step verification of natural language mathematical proofs. Unlike traditional \"full-proof\" approaches, which formalize and verify entire proofs at once, the proposed method (\"StepProof\") breaks down proofs into smaller, sentence-level subproofs to perform more granular verification. This compositional approach also allows for an intereactive user experience as well as error handling by allowing verification of individual steps and retrying only failed steps without discarding the entire proof. An evaluation is shown that the StepProof strategy performs better than a full-proof strategy and two existing baseline approaches for auto formalization on the GSM8K dataset.", "strengths": "1. The compositional verification approach is sensible and seems relatively novel in the autoformalization domain (though similar approaches have been explored in many other domains). Its benefits are intuitive and clear as compared to generating entire proofs at once (detecting errors, not redoing entire proofs, increased robustness) \n2. The experimental results do show core utility of the ideas with improvements over full proof strategy and baselines in terms of proof success/number of attempts, though there is much room for improvement of evaluation in various aspects. Bringing the study to use open source models like Llama and implementing baseline systems here is also admirable and can help progress the research area with broader accessibility.\n3. The compositional approach not only has improved results for full automation, but also creates a foundation for a more interactive user experience (more fine grained feedback from steps verification, allowing user to change or improve individual steps or skip steps in the proof, etc). Though this aspect has not been directly evaluated with real users in this work, it is a good direction to take in the autoformalization domain to provide more control and assistance for users.", "weaknesses": "1. Decomposition approach limitations. I am not sure about your approach of decomposing the informal proof into independent subpropositions. \"STEP-PROOF assumes each sentence in the proof is a verifiable sub-proposition\" - are you really just breaking by syntactic checks for sentences? What if you have a subproposition that is expressed in multiple sentences with dependencies or contextual information between them? Perhaps a better approach would be to try to use the LLM to explicitly and more intelligently decompose the informal proof into independent sub-propositions (or lemmas) as is commonly done in compositional approaches with LLMs. See e.g. (Tushar et al ICLR'23, Pourreza et al NeurIPS'23). This is particularly  concerning: \"and made simple manual modifications to make the proof step more consistent with the proof requirement of StepProof\". Firstly, this shows that StepProof is not directly capable of handling arbitrary informal NL proofs. Secondly, you can provide much more details here in terms of what manual modifications were required (you have plenty space left in the page limit and unlimited space in the appendix). You can explain general classes of modifications that needed to be made, and also provide many samples of the modifications you made to help the reader judge how \"simple\" the modifications are. \n2. Evaluation limitations. Though showing core value to some degree, the main evaluation results do not show a very strong improvement (6.1% vs 5.3% on compositional vs direct strategy and 27.9% vs 25.3% in comparison with the best DTV baseline). These seem pretty marginal and may be within margin of random variations in experiments and LLM performance. Also, only one dataset (GSM8K) is used - not sure if this shows generality of the approach, especially given its assumptions of decomposition at the sentence level which would be good to test on more datasets. The number of attempts comparison between StepProof and baselines is interesting - 10 vs 64 attempts is  a significant improvement for step proof. But can you clarify: are these the settings of the attempts parameter that you have chosen? Did the baselines actually require this many attempts or was their performance similar with fewer attempts? Perhaps a more explicit investigation of this would be helpful - e.g. a graph showing how the performance (accuracy) of both your system and baselines (y axis) increases or changes with the number of attempts (x axis). \n3. Presentation problems. Many errors, inconsistencies and presentation/organization issues make the paper difficult to read and follow. Please improve upon these. Some examples: \n- \"The workflow of STEP-PROOF is illustrated in the left of Figure 1.\"  - should be in the right\n- E.g., \"As shown in Table 4.2\" should be \"Table 1\" and then again for the baselines is states \"In the baseline test as shown in Table 4.2\"! which should be \"Table 2\". Please check for such mistakes and organize the paper better. \n- many references do not state the conferences where the papers have already been accepted - please improve the quality of references\n- organization of the paper in terms of sectioning needs much improvement - especially in the evaluation section, you seem to switch focus to different aspects of evaluation abruptly in the next paragraph and it is pretty confusing - can you please organize different aspects into appropriate subsections (you seem to have plenty of space left in the page limit anyway). \n- In the first paragraph of evaluation you state \"we conducted both strategy performance tests and baseline tests\" -  please clarify that what you mean by strategy performance tests and what you mean by baseline tests - it took a while to understand what these meant after back and forth reading and it helps to clearly introduce the goals of the evaluation to the user in normal language without paper-specific terminology.  \n- please clarify in Table 1 that the proof pass rate is for ONE ATTEMPT and that in Table 2 it is for multiple attempts (it took a while to understand why the proof pass rate was low here as compared to TABLE 2)\n \n\nReferenced Related works:\n- Decomposed Prompting: A MODULAR APPROACH FOR SOLVING COMPLEX TASKS\nTushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu,Kyle Richardson, Peter Clark, Ashish Sabharwal. In ICLR 2023\n- Mohammadreza Pourreza and Davood Rafiei. Din-sql: decomposed in-context learning of text-to sql with self-correction. In NeurIPS 2023", "questions": "1. Are all evaluations you have done with StepProof in fully automated manner? So there are no user interactions in these evaluations? Please clarify that if its the case. \n2. Also, some evaluation of interactive features would be good, e.g. with user studies. \n3. What is \"comments rate\" in table 2? Is it the amount of feedback from the verification system? What does the 100% for StepProof and 31.3% for DTV mean? Please include some discussion of this and how exactly it may be relevant.\n4. Why did you limit the number of attmpts in step proof to be 10? (while other methods like DTV you have up to 64 attempts). What happens after 10 attempts? Does the system improve further, or gains diminish, or could there even be any degradation as well? \n5. Can StepProof cause worse performance if the proof fails due to sentence level decomposition problems, while the fullproof  may still work as it has the whole context?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method for step-by-step verification of natural language mathematical proofs. Unlike traditional \"full-proof\" approaches, which formalize and verify entire proofs at once, the proposed method (\"StepProof\") breaks down proofs into smaller, sentence-level subproofs to perform more granular verification. This compositional approach also allows for an intereactive user experience as well as error handling by allowing verification of individual steps and retrying only failed steps without discarding the entire proof. An evaluation is shown that the StepProof strategy performs better than a full-proof strategy and two existing baseline approaches for auto formalization on the GSM8K dataset.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The compositional verification approach is sensible and seems relatively novel in the autoformalization domain (though similar approaches have been explored in many other domains). Its benefits are intuitive and clear as compared to generating entire proofs at once (detecting errors, not redoing entire proofs, increased robustness) \n2. The experimental results do show core utility of the ideas with improvements over full proof strategy and baselines in terms of proof success/number of attempts, though there is much room for improvement of evaluation in various aspects. Bringing the study to use open source models like Llama and implementing baseline systems here is also admirable and can help progress the research area with broader accessibility.\n3. The compositional approach not only has improved results for full automation, but also creates a foundation for a more interactive user experience (more fine grained feedback from steps verification, allowing user to change or improve individual steps or skip steps in the proof, etc). Though this aspect has not been directly evaluated with real users in this work, it is a good direction to take in the autoformalization domain to provide more control and assistance for users.", "weaknesses": "1. Decomposition approach limitations. I am not sure about your approach of decomposing the informal proof into independent subpropositions. \"STEP-PROOF assumes each sentence in the proof is a verifiable sub-proposition\" - are you really just breaking by syntactic checks for sentences? What if you have a subproposition that is expressed in multiple sentences with dependencies or contextual information between them? Perhaps a better approach would be to try to use the LLM to explicitly and more intelligently decompose the informal proof into independent sub-propositions (or lemmas) as is commonly done in compositional approaches with LLMs. See e.g. (Tushar et al ICLR'23, Pourreza et al NeurIPS'23). This is particularly  concerning: \"and made simple manual modifications to make the proof step more consistent with the proof requirement of StepProof\". Firstly, this shows that StepProof is not directly capable of handling arbitrary informal NL proofs. Secondly, you can provide much more details here in terms of what manual modifications were required (you have plenty space left in the page limit and unlimited space in the appendix). You can explain general classes of modifications that needed to be made, and also provide many samples of the modifications you made to help the reader judge how \"simple\" the modifications are. \n2. Evaluation limitations. Though showing core value to some degree, the main evaluation results do not show a very strong improvement (6.1% vs 5.3% on compositional vs direct strategy and 27.9% vs 25.3% in comparison with the best DTV baseline). These seem pretty marginal and may be within margin of random variations in experiments and LLM performance. Also, only one dataset (GSM8K) is used - not sure if this shows generality of the approach, especially given its assumptions of decomposition at the sentence level which would be good to test on more datasets. The number of attempts comparison between StepProof and baselines is interesting - 10 vs 64 attempts is  a significant improvement for step proof. But can you clarify: are these the settings of the attempts parameter that you have chosen? Did the baselines actually require this many attempts or was their performance similar with fewer attempts? Perhaps a more explicit investigation of this would be helpful - e.g. a graph showing how the performance (accuracy) of both your system and baselines (y axis) increases or changes with the number of attempts (x axis). \n3. Presentation problems. Many errors, inconsistencies and presentation/organization issues make the paper difficult to read and follow. Please improve upon these. Some examples: \n- \"The workflow of STEP-PROOF is illustrated in the left of Figure 1.\"  - should be in the right\n- E.g., \"As shown in Table 4.2\" should be \"Table 1\" and then again for the baselines is states \"In the baseline test as shown in Table 4.2\"! which should be \"Table 2\". Please check for such mistakes and organize the paper better. \n- many references do not state the conferences where the papers have already been accepted - please improve the quality of references\n- organization of the paper in terms of sectioning needs much improvement - especially in the evaluation section, you seem to switch focus to different aspects of evaluation abruptly in the next paragraph and it is pretty confusing - can you please organize different aspects into appropriate subsections (you seem to have plenty of space left in the page limit anyway). \n- In the first paragraph of evaluation you state \"we conducted both strategy performance tests and baseline tests\" -  please clarify that what you mean by strategy performance tests and what you mean by baseline tests - it took a while to understand what these meant after back and forth reading and it helps to clearly introduce the goals of the evaluation to the user in normal language without paper-specific terminology.  \n- please clarify in Table 1 that the proof pass rate is for ONE ATTEMPT and that in Table 2 it is for multiple attempts (it took a while to understand why the proof pass rate was low here as compared to TABLE 2)\n \n\nReferenced Related works:\n- Decomposed Prompting: A MODULAR APPROACH FOR SOLVING COMPLEX TASKS\nTushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu,Kyle Richardson, Peter Clark, Ashish Sabharwal. In ICLR 2023\n- Mohammadreza Pourreza and Davood Rafiei. Din-sql: decomposed in-context learning of text-to sql with self-correction. In NeurIPS 2023", "questions": "1. Are all evaluations you have done with StepProof in fully automated manner? So there are no user interactions in these evaluations? Please clarify that if its the case. \n2. Also, some evaluation of interactive features would be good, e.g. with user studies. \n3. What is \"comments rate\" in table 2? Is it the amount of feedback from the verification system? What does the 100% for StepProof and 31.3% for DTV mean? Please include some discussion of this and how exactly it may be relevant.\n4. Why did you limit the number of attmpts in step proof to be 10? (while other methods like DTV you have up to 64 attempts). What happens after 10 attempts? Does the system improve further, or gains diminish, or could there even be any degradation as well? \n5. Can StepProof cause worse performance if the proof fails due to sentence level decomposition problems, while the fullproof  may still work as it has the whole context?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729685055426}], "openreview_url": "https://openreview.net/forum?id=EXaKfdsw04", "arxiv_id": "2506.10558", "paper_pdf": "papers/EXaKfdsw04.pdf", "paper_pdf_sha256": "48437bb844c05dc0a6d09bb9861d6c3497293dfa30306d562793cf4f8f9e7604", "paper_pdf_bytes": 1365628, "paper_pdf_source": "openreview", "code_url": "https://github.com/r1nIGa/STEP-PROOF", "code_repository": "r1nIGa/STEP-PROOF", "code_commit": "e5da5c3e115713e0072a6f62acef0340d8ab8f02", "code_archive": "repos/EXaKfdsw04.zip", "code_archive_sha256": "6b730c88c5c7b0c480c8bb8156c4210d44ea4e716fff607993fa85eb05052e10", "code_archive_bytes": 404606, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 392, "github_languages": {"Python": 40394, "Isabelle": 504}, "github_archived": false, "github_pushed_at": "2025-05-06T16:15:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/2506-10558"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tf6nR1B8Nt", "year": 2024, "status": "rejected", "title": "No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths", "authors": ["Charles Guille-Escuret", "Hiroki Naganuma", "Kilian FATRAS", "Ioannis Mitliagkas"], "authorids": ["~Charles_Guille-Escuret1", "~Hiroki_Naganuma1", "~Kilian_FATRAS1", "~Ioannis_Mitliagkas1"], "authors_source": "OpenReview API", "abstract": "Understanding the optimization dynamics of neural networks is necessary for closing the gap between theory and practice. Stochastic first-order optimization algorithms are known to efficiently locate favorable minima in deep neural networks. This efficiency, however, contrasts with the non-convex and seemingly complex structure of neural loss landscapes. In this study, we delve into the fundamental geometric properties of sampled gradients along optimization paths. We focus on two key quantities, which appear in the restricted secant inequality and error bound. Both hold high significance for first-order optimization. Our analysis reveals that these quantities exhibit predictable, consistent behavior throughout training, despite the stochasticity induced by sampling minibatches. Our findings suggest that not only do optimization trajectories never encounter significant obstacles, but they also maintain stable dynamics during the majority of training. These observed properties are sufficiently expressive to theoretically guarantee linear convergence and prescribe learning rate schedules mirroring empirical practices. We conduct our experiments on image classification, semantic segmentation and language modeling across different batch sizes, network architectures, datasets, optimizers, and initialization seeds. We discuss the impact of each factor.\nOur work provides novel insights into the properties of neural network loss functions, and opens the door to theoretical frameworks more relevant to prevalent practice.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "kHB47xXvih", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2513/Reviewer_oKQj"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies why stochastic first-order optimization methods can succeed in modern deep neural networks, despite the non-convex nature of the training problem. Specifically, this paper considers the ratio between (modified) Restricted Secant Inequality (RSI) and (modified) Error Bound (EB). This ratio turns out to be exactly the cosine similarity between G(w) and w-w*, where G(w) is the stochastic version of the local gradient, and w-w* is the direction from the converged point to the current point. In short, the considered ratio measures how much the local update direction aligns with the true direction pointing to the global optimum.\n\nThe authors conduct experiments in different settings to examine how this ratio performs and evolves during the network training. Surprisingly, this ratio always stays positive and is very stable during the training. This means that the landscape of modern neural networks, despite its non-convexity, is benign in the sense that every local update is positively correlated with the desired direction.", "review_text": "This paper studies why stochastic first-order optimization methods can succeed in modern deep neural networks, despite the non-convex nature of the training problem. Specifically, this paper considers the ratio between (modified) Restricted Secant Inequality (RSI) and (modified) Error Bound (EB). This ratio turns out to be exactly the cosine similarity between G(w) and w-w*, where G(w) is the stochastic version of the local gradient, and w-w* is the direction from the converged point to the current point. In short, the considered ratio measures how much the local update direction aligns with the true direction pointing to the global optimum.\n\nThe authors conduct experiments in different settings to examine how this ratio performs and evolves during the network training. Surprisingly, this ratio always stays positive and is very stable during the training. This means that the landscape of modern neural networks, despite its non-convexity, is benign in the sense that every local update is positively correlated with the desired direction.", "strengths": "1. This paper proposed an effective metric, based on a variant of quantities involved in RSI and EB, to evaluate the training trajectory of modern neural networks. Both RSI  and EB are theoretically grounded metrics in optimization theory, capturing important properties of the optimization landscape.\n\n2. Experiments validate that the cosine similarity metric stays positive and stable during training. This is an interesting finding, which may potentially inspire theoretical landscape study and/or convergence analysis of neural networks. Further, the trends of cosine similarity can partially justify the empirically used learning rate schedule.\n\n3. This paper conducts extensive experiments under different settings and carries out comprehensive discussion. Notably, the proposed metric is dependent on the optimization trajectory, and thus the authors carefully discuss potential factors that might affect the interpretation of empirical results.", "weaknesses": "1. The paper does not provide any further theoretical analysis based on their empirical findings. Although existing works have demonstrated that well-behaved RSI and EB can lead to the convergence of first-order algorithms, it is not clear whether trajectory-dependent RSI and EB properties can lead to similar results.\n\n2. The paper does not provide practical guidance on network training. Most conclusions are explanatory.\n\nOverall, it is good to point out the \"no wrong returns\" property of the neural network landscape. However, its immediate impact on theory or practice is unclear to me.", "questions": "1. In contrasting examples, what is w*? The (known) global optimum or the converged point?\n\n2. The authors claimed \"these observed properties are sufficiently expressive to theoretically guarantee linear convergence\" in the abstract. However, the paper considers variants of RSI and EB. Only local but not global properties are examined. Thus, the empirical findings cannot guarantee linear convergence. Is it correct?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies why stochastic first-order optimization methods can succeed in modern deep neural networks, despite the non-convex nature of the training problem. Specifically, this paper considers the ratio between (modified) Restricted Secant Inequality (RSI) and (modified) Error Bound (EB). This ratio turns out to be exactly the cosine similarity between G(w) and w-w*, where G(w) is the stochastic version of the local gradient, and w-w* is the direction from the converged point to the current point. In short, the considered ratio measures how much the local update direction aligns with the true direction pointing to the global optimum.\n\nThe authors conduct experiments in different settings to examine how this ratio performs and evolves during the network training. Surprisingly, this ratio always stays positive and is very stable during the training. This means that the landscape of modern neural networks, despite its non-convexity, is benign in the sense that every local update is positively correlated with the desired direction.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. This paper proposed an effective metric, based on a variant of quantities involved in RSI and EB, to evaluate the training trajectory of modern neural networks. Both RSI  and EB are theoretically grounded metrics in optimization theory, capturing important properties of the optimization landscape.\n\n2. Experiments validate that the cosine similarity metric stays positive and stable during training. This is an interesting finding, which may potentially inspire theoretical landscape study and/or convergence analysis of neural networks. Further, the trends of cosine similarity can partially justify the empirically used learning rate schedule.\n\n3. This paper conducts extensive experiments under different settings and carries out comprehensive discussion. Notably, the proposed metric is dependent on the optimization trajectory, and thus the authors carefully discuss potential factors that might affect the interpretation of empirical results.", "weaknesses": "1. The paper does not provide any further theoretical analysis based on their empirical findings. Although existing works have demonstrated that well-behaved RSI and EB can lead to the convergence of first-order algorithms, it is not clear whether trajectory-dependent RSI and EB properties can lead to similar results.\n\n2. The paper does not provide practical guidance on network training. Most conclusions are explanatory.\n\nOverall, it is good to point out the \"no wrong returns\" property of the neural network landscape. However, its immediate impact on theory or practice is unclear to me.", "questions": "1. In contrasting examples, what is w*? The (known) global optimum or the converged point?\n\n2. The authors claimed \"these observed properties are sufficiently expressive to theoretically guarantee linear convergence\" in the abstract. However, the paper considers variants of RSI and EB. Only local but not global properties are examined. Thus, the empirical findings cannot guarantee linear convergence. Is it correct?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699536453729}, {"id": "Mxnm5WMM7b", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2513/Reviewer_eK7x"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Given the efficacy of stochastic first-order optimization algorithms on the non-convex landscapes of neural networks, this paper aims to show the simplicity of neural landscapes by tracking two key quantities that occur in the convergence analysis. Namely, they consider the restricted secant inequality (RSI) and the error bound (EB) during the optimization trajectory, whose ratio in turn boils down to the cosine similarity between the sampled minibatch gradient and the difference between the current and the final weights. Their observations seem to suggest that this quantity remains almost positive throughout (and that's why the titular phrase \"no wrong turns\") and has consistent behaviour during training. The results are demonstrated on residual networks in vision and language, and the effects of batch size, longer training duration, and initialization are considered.", "review_text": "Given the efficacy of stochastic first-order optimization algorithms on the non-convex landscapes of neural networks, this paper aims to show the simplicity of neural landscapes by tracking two key quantities that occur in the convergence analysis. Namely, they consider the restricted secant inequality (RSI) and the error bound (EB) during the optimization trajectory, whose ratio in turn boils down to the cosine similarity between the sampled minibatch gradient and the difference between the current and the final weights. Their observations seem to suggest that this quantity remains almost positive throughout (and that's why the titular phrase \"no wrong turns\") and has consistent behaviour during training. The results are demonstrated on residual networks in vision and language, and the effects of batch size, longer training duration, and initialization are considered.", "strengths": "Firstly, I think it's a great research direction to understand how the important quantities that occur in the optimization analyses actually behave in the neural landscapes and then ground up uncover properties of the loss landscape. These identified properties, such as the cosine similarity, might reflect an important aspect of the trajectories taken by 1st-order optimization algorithms in loss and the geometry of the landscapes in general. It is interesting to see how these quantities behave across various established empirical settings.", "weaknesses": "- **The nature of cosine similarity in high dimensions:** One of the fundamental issues about this paper is to what extent cosine similarity is a meaningful measure in high dimensions, which here are in the orders of tens of millions. The particular value of cosine similarity is somewhat difficult to grasp as a result, without the lack of a baseline that contextualizes its value. While the curves look alright and pretty stable throughout, the moment one looks at the value at the y-axis (which, e.g., is in the order of 1e-3 for ImageNet), doubts emerge if one is reading too much signal in the noise. This is in addition to the fact that the cosine similarity is between vectors that at their core are based on the same network structure will tend to have a higher cosine similarity (contrast that with arbitrary random vectors in the same dimensions). \n\n   All of this makes it hard to be convinced by their argument about the stability of this cosine similarity (and optimization proceeding at a regular pace), when this value of cosine similarity almost hints at near orthogonality.  \n\n&nbsp;\n\n- **The lack of clear insights that can be drawn:** Consider Figure 2, middle column, where the results are shown for ResNet50/ImageNet. It is unclear if anything much is being said by these graphs, except towards the end where it hints at being close to convergence. And that too, because the denominator in these expressions $\\|w-w^\\ast\\|$ starts becoming small. More generally, a lot of these plots contain a sense of vagueness about the actual insights that can be drawn from them. The paper says there are \"predictable\" properties, but a less glorified way of phrasing would be they are somewhat banal. \n\n&nbsp;\n\n- **Dependence of the results on the chosen empirical setup** It is not clear if the results are overly reliant on the particular empirical setup chosen in the paper. Does the same thing hold when training with large learning rates? Bigger or smaller weight decay coefficients? When different learning rate schedulers are used? Essentially, my guess would be that something that would make the network jump in the optimization landscape (which is not uncommon in practical settings) should at least result in a different phenomenology than that presented here. Likewise, what happens when there is more noise in the dataset, say different amounts of label noise? Besides most of the experiments have been carried out with residual architectures. What happens if you consider a fully convolutional architecture, such as VGG on cifar10 and/or MobileNet on ImageNet?\n\n\n---- Post rebuttal ---- \n\nI think the way of using cosine similarity currently is quite problematic and can thus be potentially misleading. The standard deviation for the cosine similarity of two independent random vectors on the sphere is itself, for dimension (and which seems, more or less, in the ballpark for the networks of presumably million parameters). But more than that, the vectors considered here are not purely arbitrary, and stem from the same network structure. For instance, Bao et. al (2023) consider as a baseline the cosine similarity of the network parameters at two different initializations, and then the calculations are contextualized based on such a value. Right now, I am afraid this paper is probably reading too much 'signal' in the noise, again due to a lack of a relevant baseline. I would recommend the authors contextualize with a measure like this and revise the results accordingly. At the moment, I will stick to my score.\n\nBao et al. 2023: Hessian Inertia in Neural Networks, ICML 2023 Workshop on High-dimensional Learning Dynamics", "questions": "See the weaknesses section. \n\nSome minor questions:\n- Influence of Batch size: Shouldn't the equation there be normalized since when you have a bigger batch size, you don't just add the gradients of smaller batch sizes but also normalize corresponding to their sizes? How does the argument then work out?\n- Why should the set of global minima be convex? (page 3)\n- Model width/depth: How are the relative trends when you normalize the quantities by the number of parameters?\n- loLR: I understand the aim of Figure 3, but can you port these LR schedules for a fresh training of the corresponding networks? What do you observe?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Given the efficacy of stochastic first-order optimization algorithms on the non-convex landscapes of neural networks, this paper aims to show the simplicity of neural landscapes by tracking two key quantities that occur in the convergence analysis. Namely, they consider the restricted secant inequality (RSI) and the error bound (EB) during the optimization trajectory, whose ratio in turn boils down to the cosine similarity between the sampled minibatch gradient and the difference between the current and the final weights. Their observations seem to suggest that this quantity remains almost positive throughout (and that's why the titular phrase \"no wrong turns\") and has consistent behaviour during training. The results are demonstrated on residual networks in vision and language, and the effects of batch size, longer training duration, and initialization are considered.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "Firstly, I think it's a great research direction to understand how the important quantities that occur in the optimization analyses actually behave in the neural landscapes and then ground up uncover properties of the loss landscape. These identified properties, such as the cosine similarity, might reflect an important aspect of the trajectories taken by 1st-order optimization algorithms in loss and the geometry of the landscapes in general. It is interesting to see how these quantities behave across various established empirical settings.", "weaknesses": "- **The nature of cosine similarity in high dimensions:** One of the fundamental issues about this paper is to what extent cosine similarity is a meaningful measure in high dimensions, which here are in the orders of tens of millions. The particular value of cosine similarity is somewhat difficult to grasp as a result, without the lack of a baseline that contextualizes its value. While the curves look alright and pretty stable throughout, the moment one looks at the value at the y-axis (which, e.g., is in the order of 1e-3 for ImageNet), doubts emerge if one is reading too much signal in the noise. This is in addition to the fact that the cosine similarity is between vectors that at their core are based on the same network structure will tend to have a higher cosine similarity (contrast that with arbitrary random vectors in the same dimensions). \n\n   All of this makes it hard to be convinced by their argument about the stability of this cosine similarity (and optimization proceeding at a regular pace), when this value of cosine similarity almost hints at near orthogonality.  \n\n&nbsp;\n\n- **The lack of clear insights that can be drawn:** Consider Figure 2, middle column, where the results are shown for ResNet50/ImageNet. It is unclear if anything much is being said by these graphs, except towards the end where it hints at being close to convergence. And that too, because the denominator in these expressions $\\|w-w^\\ast\\|$ starts becoming small. More generally, a lot of these plots contain a sense of vagueness about the actual insights that can be drawn from them. The paper says there are \"predictable\" properties, but a less glorified way of phrasing would be they are somewhat banal. \n\n&nbsp;\n\n- **Dependence of the results on the chosen empirical setup** It is not clear if the results are overly reliant on the particular empirical setup chosen in the paper. Does the same thing hold when training with large learning rates? Bigger or smaller weight decay coefficients? When different learning rate schedulers are used? Essentially, my guess would be that something that would make the network jump in the optimization landscape (which is not uncommon in practical settings) should at least result in a different phenomenology than that presented here. Likewise, what happens when there is more noise in the dataset, say different amounts of label noise? Besides most of the experiments have been carried out with residual architectures. What happens if you consider a fully convolutional architecture, such as VGG on cifar10 and/or MobileNet on ImageNet?\n\n\n---- Post rebuttal ---- \n\nI think the way of using cosine similarity currently is quite problematic and can thus be potentially misleading. The standard deviation for the cosine similarity of two independent random vectors on the sphere is itself, for dimension (and which seems, more or less, in the ballpark for the networks of presumably million parameters). But more than that, the vectors considered here are not purely arbitrary, and stem from the same network structure. For instance, Bao et. al (2023) consider as a baseline the cosine similarity of the network parameters at two different initializations, and then the calculations are contextualized based on such a value. Right now, I am afraid this paper is probably reading too much 'signal' in the noise, again due to a lack of a relevant baseline. I would recommend the authors contextualize with a measure like this and revise the results accordingly. At the moment, I will stick to my score.\n\nBao et al. 2023: Hessian Inertia in Neural Networks, ICML 2023 Workshop on High-dimensional Learning Dynamics", "questions": "See the weaknesses section. \n\nSome minor questions:\n- Influence of Batch size: Shouldn't the equation there be normalized since when you have a bigger batch size, you don't just add the gradients of smaller batch sizes but also normalize corresponding to their sizes? How does the argument then work out?\n- Why should the set of global minima be convex? (page 3)\n- Model width/depth: How are the relative trends when you normalize the quantities by the number of parameters?\n- loLR: I understand the aim of Figure 3, but can you port these LR schedules for a fresh training of the corresponding networks? What do you observe?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699205836162}, {"id": "BUnUJmeknS", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2513/Reviewer_LuUd"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper studies the geometric properties of NN training loss over the training procedure. The authors consider two conditions called Lower Restricted Secant Inequality (RSI) and Upper Error Bounds (EB). The paper tracks the local and stochastic versions $RSI(G,w,w^*)$ and $EB(G,w,w^*)$ as well as their ratio $\\gamma(G,w,w^*) = \\frac{RSI(G,w,w^*)}{EB(G,w,w^*)}$, which are quantities evaluated with the stochastic gradient oracle $G$ queried for the current point $w$. In particular, $\\gamma$ corresponds to the cosine similarity of the stochastic gradient and $w-w^*$, so $\\gamma > 0$ implies that the stochastic gradient is pointing to the \"right\" direction. For various settings, the paper demonstrates that the RSI, EB, and $\\gamma$ stay quite stable throughout training, despite the nonconvexity and stochasticity. Moreover, the cosine similarity $\\gamma$ is almost always positive. The paper supplements the main results with ablations and detailed discussions.", "review_text": "This paper studies the geometric properties of NN training loss over the training procedure. The authors consider two conditions called Lower Restricted Secant Inequality (RSI) and Upper Error Bounds (EB). The paper tracks the local and stochastic versions $RSI(G,w,w^*)$ and $EB(G,w,w^*)$ as well as their ratio $\\gamma(G,w,w^*) = \\frac{RSI(G,w,w^*)}{EB(G,w,w^*)}$, which are quantities evaluated with the stochastic gradient oracle $G$ queried for the current point $w$. In particular, $\\gamma$ corresponds to the cosine similarity of the stochastic gradient and $w-w^*$, so $\\gamma > 0$ implies that the stochastic gradient is pointing to the \"right\" direction. For various settings, the paper demonstrates that the RSI, EB, and $\\gamma$ stay quite stable throughout training, despite the nonconvexity and stochasticity. Moreover, the cosine similarity $\\gamma$ is almost always positive. The paper supplements the main results with ablations and detailed discussions.", "strengths": "1. I think this paper is a good submission. It provides an interesting set of results that sheds light on the theory-practice gap in neural network training. The empirical observations suggest that neural network training, although nonconvex and stochastic, could be in fact much more well-behaved than expected.\n\n2. The paper is well-written and delivers the main ideas clearly. The ablation studies and discussions at the end of the paper provide useful insights.\n\n3. The analysis of locally optimal learning rate seems to have a potential to lead to a method for learning rate schedule design or for adaptive learning rate schemes.", "weaknesses": "1. There is a closely relevant paper titled \"SGD Converges to Global Minimum in Deep Learning via Star-convex Path\" which appeared in ICLR 2019, which also empirically studies the SGD trajectories and shows that the star-convexity is satisfied for most epochs. I have not checked the follow-up results carefully, so there may be more existing results that discover surprisingly benign geometric properties of neural network training. The paper should contextualize itself relative to this existing result(s).", "questions": "1. A recently discovered phenomenon Edge of Stability (EoS) shows that GD trajectory experiences progressive sharpening and then reaches and hovers around the stable convergence threshold $2/\\text{(step size)}$. When EoS is in effect in the later phase of training, the training dynamics is rather unstable, with occasional peaks in loss curves. Similarly, when you run SGD instead of GD, we also often observe these loss peaks in the training curves. In contrast, the RSI, EB, and $\\gamma$ curves look quite stable, without any significant peaks or perturbations. My question is: did you observe any peaks or unstable convergence in your experiments? Do the \"fundamental properties\" (Section 4.1) continue to hold when we observe EoS? If so, how can we reconcile these seemingly contradictory trends?\n\n2. In the late phase of training, the curves seem to experience sharp increases in both mean and variance. The explanations based on interpolation vs non-interpolation and the bias introduced by setting $w^* \\approx w_T$ are intuitive, but do they explain the whole story? For example, in Figure 2, the RSI and EB of CIFAR-10 do not show a significant increase in mean curves, but we do observe a consistent increase in variance. What could be an explanation for this?\n\n3. It seems that using large batch sizes increases $\\gamma$ throughout training. What happens if we use full-batch training (i.e., GD)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the geometric properties of NN training loss over the training procedure. The authors consider two conditions called Lower Restricted Secant Inequality (RSI) and Upper Error Bounds (EB). The paper tracks the local and stochastic versions $RSI(G,w,w^*)$ and $EB(G,w,w^*)$ as well as their ratio $\\gamma(G,w,w^*) = \\frac{RSI(G,w,w^*)}{EB(G,w,w^*)}$, which are quantities evaluated with the stochastic gradient oracle $G$ queried for the current point $w$. In particular, $\\gamma$ corresponds to the cosine similarity of the stochastic gradient and $w-w^*$, so $\\gamma > 0$ implies that the stochastic gradient is pointing to the \"right\" direction. For various settings, the paper demonstrates that the RSI, EB, and $\\gamma$ stay quite stable throughout training, despite the nonconvexity and stochasticity. Moreover, the cosine similarity $\\gamma$ is almost always positive. The paper supplements the main results with ablations and detailed discussions.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. I think this paper is a good submission. It provides an interesting set of results that sheds light on the theory-practice gap in neural network training. The empirical observations suggest that neural network training, although nonconvex and stochastic, could be in fact much more well-behaved than expected.\n\n2. The paper is well-written and delivers the main ideas clearly. The ablation studies and discussions at the end of the paper provide useful insights.\n\n3. The analysis of locally optimal learning rate seems to have a potential to lead to a method for learning rate schedule design or for adaptive learning rate schemes.", "weaknesses": "1. There is a closely relevant paper titled \"SGD Converges to Global Minimum in Deep Learning via Star-convex Path\" which appeared in ICLR 2019, which also empirically studies the SGD trajectories and shows that the star-convexity is satisfied for most epochs. I have not checked the follow-up results carefully, so there may be more existing results that discover surprisingly benign geometric properties of neural network training. The paper should contextualize itself relative to this existing result(s).", "questions": "1. A recently discovered phenomenon Edge of Stability (EoS) shows that GD trajectory experiences progressive sharpening and then reaches and hovers around the stable convergence threshold $2/\\text{(step size)}$. When EoS is in effect in the later phase of training, the training dynamics is rather unstable, with occasional peaks in loss curves. Similarly, when you run SGD instead of GD, we also often observe these loss peaks in the training curves. In contrast, the RSI, EB, and $\\gamma$ curves look quite stable, without any significant peaks or perturbations. My question is: did you observe any peaks or unstable convergence in your experiments? Do the \"fundamental properties\" (Section 4.1) continue to hold when we observe EoS? If so, how can we reconcile these seemingly contradictory trends?\n\n2. In the late phase of training, the curves seem to experience sharp increases in both mean and variance. The explanations based on interpolation vs non-interpolation and the bias introduced by setting $w^* \\approx w_T$ are intuitive, but do they explain the whole story? For example, in Figure 2, the RSI and EB of CIFAR-10 do not show a significant increase in mean curves, but we do observe a consistent increase in variance. What could be an explanation for this?\n\n3. It seems that using large batch sizes increases $\\gamma$ throughout training. What happens if we use full-batch training (i.e., GD)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698859632634}, {"id": "XKvyrvu8jH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2513/Reviewer_Km89"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes several measurements to quantify the optimization dynamics of training neural networks. Motivated by optimization theory, the authors study how cosine similarity, Lipschitz-type, and convexity-type quantities evolve over the training process. \n\nBy empirically studying those quantities on many datasets, ranging from images to languages, the authors claim that the optimization geometry is stable and simple, and in particular \"optimization trajectories never encounter significant obstacles\". Based on these observations, the authors provide insights and discussions about common training techniques such as initialization and batch size.", "review_text": "This paper proposes several measurements to quantify the optimization dynamics of training neural networks. Motivated by optimization theory, the authors study how cosine similarity, Lipschitz-type, and convexity-type quantities evolve over the training process. \n\nBy empirically studying those quantities on many datasets, ranging from images to languages, the authors claim that the optimization geometry is stable and simple, and in particular \"optimization trajectories never encounter significant obstacles\". Based on these observations, the authors provide insights and discussions about common training techniques such as initialization and batch size.", "strengths": "1. Overall the observation that optimization dynamics is benign for state-of-the-art neural networks is informative and interesting. This reinforces prior observations such as those in Goodfellow et al. (2015).  \n\n2. Empirical measurements such as RSI, EB, and cosine similarity are useful diagnostic tools for understanding the optimization property of neural networks", "weaknesses": "Overall, I find the paper seems to be overclaiming the results.\n\n1. (main) This paper seems to give a misleading message. The title says the neural networks have \"simple geometry\", and the abstract says \"optimization trajectories never encounter significant obstacles, [....] maintain stable dynamics during the majority of training.\" I think that those properties are examined only for state-of-the-art neural networks. There are lots of techniques to alleviate optimization issues, so the overall claim of this paper severely trivializes the techniques in common practice. (On a related note: It would be interesting to investigate how much each technique contributes to making optimization benign.)\n\n2. (main) I find that RSI is confusing to understand: in Figure 2, the measure for ImageNet and WikiText is initially close to zero, so does this mean the training dynamics is not ideal in the beginning? \"RSI and EB follow predictable trends\" do not imply that optimization dynamics have benign properties.\n\n3. (main) The cosine similarity is low. While I agree with the argument that there is stochasticity, I think cosine similarity should be compared with the baseline where (i) gradients are replaced by noise, or (ii) gradients are calculated based on random labels or random tokens in a minibatch. Without comparison with the pure noise baseline, we do not know if cosine similarity is significantly correlated with the signal direction.\n\n4. (minor) I understand that this paper is mainly about investigating the empirical properties of training neural networks from the optimization perspective. But from a practical point of view, what can we learn? Do we know what part of the architecture of, say a transformer, is crucial for benign optimization? Which techniques (layer normalization, dropout, learning rate warmup, etc) are playing a role?", "questions": "Please see the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes several measurements to quantify the optimization dynamics of training neural networks. Motivated by optimization theory, the authors study how cosine similarity, Lipschitz-type, and convexity-type quantities evolve over the training process. \n\nBy empirically studying those quantities on many datasets, ranging from images to languages, the authors claim that the optimization geometry is stable and simple, and in particular \"optimization trajectories never encounter significant obstacles\". Based on these observations, the authors provide insights and discussions about common training techniques such as initialization and batch size.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. Overall the observation that optimization dynamics is benign for state-of-the-art neural networks is informative and interesting. This reinforces prior observations such as those in Goodfellow et al. (2015).  \n\n2. Empirical measurements such as RSI, EB, and cosine similarity are useful diagnostic tools for understanding the optimization property of neural networks", "weaknesses": "Overall, I find the paper seems to be overclaiming the results.\n\n1. (main) This paper seems to give a misleading message. The title says the neural networks have \"simple geometry\", and the abstract says \"optimization trajectories never encounter significant obstacles, [....] maintain stable dynamics during the majority of training.\" I think that those properties are examined only for state-of-the-art neural networks. There are lots of techniques to alleviate optimization issues, so the overall claim of this paper severely trivializes the techniques in common practice. (On a related note: It would be interesting to investigate how much each technique contributes to making optimization benign.)\n\n2. (main) I find that RSI is confusing to understand: in Figure 2, the measure for ImageNet and WikiText is initially close to zero, so does this mean the training dynamics is not ideal in the beginning? \"RSI and EB follow predictable trends\" do not imply that optimization dynamics have benign properties.\n\n3. (main) The cosine similarity is low. While I agree with the argument that there is stochasticity, I think cosine similarity should be compared with the baseline where (i) gradients are replaced by noise, or (ii) gradients are calculated based on random labels or random tokens in a minibatch. Without comparison with the pure noise baseline, we do not know if cosine similarity is significantly correlated with the signal direction.\n\n4. (minor) I understand that this paper is mainly about investigating the empirical properties of training neural networks from the optimization perspective. But from a practical point of view, what can we learn? Do we know what part of the architecture of, say a transformer, is crucial for benign optimization? Which techniques (layer normalization, dropout, learning rate warmup, etc) are playing a role?", "questions": "Please see the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698362314162}], "openreview_url": "https://openreview.net/forum?id=tf6nR1B8Nt", "arxiv_id": "2306.11922", "paper_pdf": "papers/tf6nR1B8Nt.pdf", "paper_pdf_sha256": "4cf7f53c580bcd280882a892c38b7dd790e4d313a57458a02fc0f00d19f0f24f", "paper_pdf_bytes": 713796, "paper_pdf_source": "openreview", "code_url": "https://github.com/Hiroki11x/LossLandscapeGeometry", "code_repository": "Hiroki11x/LossLandscapeGeometry", "code_commit": "adb17b0b342305c7b19ec68c39de0c0bef78d10e", "code_archive": "repos/tf6nR1B8Nt.zip", "code_archive_sha256": "f4a7abad49e9c3fff28c2fd645ac6dfb29b019fba50e7d06bb1320dc473a92b4", "code_archive_bytes": 384564, "code_file_count": 36, "code_extensions": {".py": 21, ".sh": 15}, "github_disk_usage_kb": 384, "github_languages": {"Shell": 2236497, "Python": 113312}, "github_archived": false, "github_pushed_at": "2024-06-10T20:08:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/no-wrong-turns-the-simple-geometry-of-neural"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "S80I3NwbbpS", "year": 2023, "status": "rejected", "title": "CAB: Comprehensive Attention Benchmarking on Long Sequence Modeling", "authors": ["Jun Zhang", "Shuyang Jiang", "Jiangtao Feng", "Lin Zheng", "Lingpeng Kong"], "authorids": ["~Jun_Zhang27", "~Shuyang_Jiang2", "~Jiangtao_Feng1", "~Lin_Zheng1", "~Lingpeng_Kong1"], "authors_source": "OpenReview API", "abstract": "Transformer has achieved remarkable success in language, image, and speech processing. Recently, various efficient attention architectures have been proposed to improve transformer's efficiency while largely preserving its efficacy, especially in modeling long sequences. A widely-used benchmark to test these efficient methods' capability on long-range modeling is Long Range Arena (LRA). However, LRA only focuses on the standard bidirectional (or noncausal) self attention, and completely ignores cross attentions and unidirectional (or causal) attentions, which are equally important to downstream applications. Although designing cross and causal variants of an attention method is straightforward for vanilla attention, it is often challenging for efficient attentions with subquadratic time and memory complexity. In this paper, we propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. CAB collects seven real-world tasks from different research areas to evaluate efficient attentions under the four attention patterns. Among these tasks, CAB validates efficient attentions in eight backbone networks to show their generalization across neural architectures. We conduct exhaustive experiments to benchmark the performances of nine widely-used efficient attention architectures designed with different philosophies on CAB. Extensive experimental results also shed light on the fundamental problems of efficient attentions, such as efficiency length against vanilla attention, performance consistency across attention patterns, the benefit of attention mechanisms, and interpolation/extrapolation on long-context language modeling.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "FXXSNzW03Q", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1790/Reviewer_auuz"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a novel benchmark, called CAB,  to compare the efficiency of different attention mechanisms in transformers. The current widely used benchmark LRA focusses only on noncausal self-attention. CAB seeks to evaluate four types of attention mechanisms: the combinations of causal/ non-causal and self/ cross attention. The benchmark consists of nine datasets for seven real-world tasks across vision, NLP, speech and time-series forecasting. CAB  is also used to evaluate 8 baseline models covering the four attention mechanisms using standard statistical performance metrics and a proposed composite index as well their efficiency – relative running times and memory usage as a function of the sequence length, highlighting the strengths and weakness of efficient approach across the four attention types suggesting where more research is needed.\n", "review_text": "The paper introduces a new benchmark for comparing and evaluating efficient attention mechanisms. I expect the benchmark to be useful to the research community and is likely to help in guiding research into the design of novel efficient attention mechanisms, especially in the causal and cross-modal domains. There is an extensive benchmarking of exemplary baseline algorithms. The paper is mostly written well (though the writing, clarity and the discussions can be somewhat improved).\n", "strengths": "**Strengths**\n\n- CAB is a novel benchmark with a finer granularity of attention taxonomy than any existing benchmark. LRA addresses only one of the 4 types of attention mechanisms.\n-  CAB comprises 7 tasks of interest to the research community on 9 datasets being used by the research community across 4 domains.\n- The paper also evaluates the performance of 8 popular models as baselines as a part of CAB for new approaches to be benchmarked against.\n- The paper further identifies the scenarios where efficient approaches are significantly lagging behind vanilla attention in performance as the unmet need for the research community to focus on.\n\n**Weaknesses**\n\n***Analysis of Results*** \n- The paper should provide more insights on why results from certain methods vary across various studies: local attention in (Xiong et al, 2022) vs LRA; S4D on LRA vs Sum, MLM tasks, etc.\n- (Figure 3) In the ‘self’ regime, why is it that performance holds across LRA and NS but drops (in correlation) for causal-self? \n- (Figure 3) There is a clear blockiness along the diagonals indicating high positive correlation in performance of approaches within the ‘self’ scenarios (LRA, non-causal and causal) and within the ‘cross’ scenarios (non-causal as well as causal). What may be the reasons for high negative correlation in the off-diagonal blocks? In other words, can you kindly elaborate on footnote 4? Also, what explorations need to be done beyond this benchmark?\n- It is not clear what ‘non-homologous information’ is. Kindly provide explanation and a reference for the term.\n\n***Clarity***\n- Figures showing the tradeoff between performance (accuracy) and cost (time, memory) are recommended to communicate the tradeoffs better.\n- Writing should be improved in Sections 4 and 5. The authors should avoid using words/ phrases like ‘utterly defeated’, ‘embarrasingly’,  etc. \n- Appropriate capitalization should be used in the ‘References’ section.\n- Some easily fixable typos.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a novel benchmark, called CAB,  to compare the efficiency of different attention mechanisms in transformers. The current widely used benchmark LRA focusses only on noncausal self-attention. CAB seeks to evaluate four types of attention mechanisms: the combinations of causal/ non-causal and self/ cross attention. The benchmark consists of nine datasets for seven real-world tasks across vision, NLP, speech and time-series forecasting. CAB  is also used to evaluate 8 baseline models covering the four attention mechanisms using standard statistical performance metrics and a proposed composite index as well their efficiency – relative running times and memory usage as a function of the sequence length, highlighting the strengths and weakness of efficient approach across the four attention types suggesting where more research is needed.\n", "strength_and_weaknesses": "**Strengths**\n\n- CAB is a novel benchmark with a finer granularity of attention taxonomy than any existing benchmark. LRA addresses only one of the 4 types of attention mechanisms.\n-  CAB comprises 7 tasks of interest to the research community on 9 datasets being used by the research community across 4 domains.\n- The paper also evaluates the performance of 8 popular models as baselines as a part of CAB for new approaches to be benchmarked against.\n- The paper further identifies the scenarios where efficient approaches are significantly lagging behind vanilla attention in performance as the unmet need for the research community to focus on.\n\n**Weaknesses**\n\n***Analysis of Results*** \n- The paper should provide more insights on why results from certain methods vary across various studies: local attention in (Xiong et al, 2022) vs LRA; S4D on LRA vs Sum, MLM tasks, etc.\n- (Figure 3) In the ‘self’ regime, why is it that performance holds across LRA and NS but drops (in correlation) for causal-self? \n- (Figure 3) There is a clear blockiness along the diagonals indicating high positive correlation in performance of approaches within the ‘self’ scenarios (LRA, non-causal and causal) and within the ‘cross’ scenarios (non-causal as well as causal). What may be the reasons for high negative correlation in the off-diagonal blocks? In other words, can you kindly elaborate on footnote 4? Also, what explorations need to be done beyond this benchmark?\n- It is not clear what ‘non-homologous information’ is. Kindly provide explanation and a reference for the term.\n\n***Clarity***\n- Figures showing the tradeoff between performance (accuracy) and cost (time, memory) are recommended to communicate the tradeoffs better.\n- Writing should be improved in Sections 4 and 5. The authors should avoid using words/ phrases like ‘utterly defeated’, ‘embarrasingly’,  etc. \n- Appropriate capitalization should be used in the ‘References’ section.\n- Some easily fixable typos.\n", "clarity,_quality,_novelty_and_reproducibility": "***Clarity***: The paper is mostly clear and easy to read.\n\n***Quality***: The overall quality of the submission is acceptable.\n\n***Novelty***: The benchmark is novel, more comprehensive that the previous benchmarks and would be useful to compare future approaches and help spur research into efficient approaches.\n\n***Reproducibility***:  Should not be a concern as the authors have promised that all the code will be released on github. \n", "summary_of_the_review": "The paper introduces a new benchmark for comparing and evaluating efficient attention mechanisms. I expect the benchmark to be useful to the research community and is likely to help in guiding research into the design of novel efficient attention mechanisms, especially in the causal and cross-modal domains. There is an extensive benchmarking of exemplary baseline algorithms. The paper is mostly written well (though the writing, clarity and the discussions can be somewhat improved).\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666892125909}, {"id": "oZJ2t-PG7p", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1790/Reviewer_kTjS"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": " The authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. The benchmark collects seven real-world tasks from diverse fields of computer vision, natural language processing, speech processing, and time series forecasting. And an interesting finding is that some efficient Transformers often achieve less competitive results in the causal cross scenario.", "review_text": "The benchmark is interesting and covers many tasks from different domains. The exploration of different attention mechanism is interesting. My major concern is that there has been benchmarks for NLP tasks. After all, I would like to see people working on this new benchmark.", "strengths": "Strength:\n1. A benchmark to evaluate different types of attention patterns for the task with long sequences\n2. The benchmark covers a broad number of tasks.\n\nWeakness:\n1. The number of baselines in figures b,c,d are limited. It would be better to list more.\n2. LongT5, \"Efficient Text-To-Text Transformer for Long Sequences\", also works on the causal attention mechanism. All the datasets in LongT5 could also be a good benchmark. Language model is also a widely used task to evaluate causal attention. Thus, the benchmark may not be novel enough if only considering the NLP area.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": " The authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. The benchmark collects seven real-world tasks from diverse fields of computer vision, natural language processing, speech processing, and time series forecasting. And an interesting finding is that some efficient Transformers often achieve less competitive results in the causal cross scenario.", "strength_and_weaknesses": "Strength:\n1. A benchmark to evaluate different types of attention patterns for the task with long sequences\n2. The benchmark covers a broad number of tasks.\n\nWeakness:\n1. The number of baselines in figures b,c,d are limited. It would be better to list more.\n2. LongT5, \"Efficient Text-To-Text Transformer for Long Sequences\", also works on the causal attention mechanism. All the datasets in LongT5 could also be a good benchmark. Language model is also a widely used task to evaluate causal attention. Thus, the benchmark may not be novel enough if only considering the NLP area.", "clarity,_quality,_novelty_and_reproducibility": "Can you list the SOTA results with model pertaining for NLP tasks?", "summary_of_the_review": "The benchmark is interesting and covers many tasks from different domains. The exploration of different attention mechanism is interesting. My major concern is that there has been benchmarks for NLP tasks. After all, I would like to see people working on this new benchmark.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666685603875}, {"id": "5gF7j5i8Joi", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1790/Reviewer_MeQL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions.\n", "review_text": "This paper proposes to evaluate efficient attentions under a more fine-grained attention taxonomy\nwith four distinguishable attention patterns, each of which reflects different attentive functionality.", "strengths": "Strength\n1. the paper is well written and feeling pleasure to read. It summarize many eixting of Transformer.\n\n2. The authors propose Comprehensive Attention Benchmark for ling sequence modeling.\n\nWeaknessness\n1. the novelty of this work is not enough for ICLR. It seems that this paper is more related to benchmark dataset.\n\n2. They claim that  related codes will be released at https://github.com/Anonymous. However, I cannot find it.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this paper, the authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions.\n", "strength_and_weaknesses": "Strength\n1. the paper is well written and feeling pleasure to read. It summarize many eixting of Transformer.\n\n2. The authors propose Comprehensive Attention Benchmark for ling sequence modeling.\n\nWeaknessness\n1. the novelty of this work is not enough for ICLR. It seems that this paper is more related to benchmark dataset.\n\n2. They claim that  related codes will be released at https://github.com/Anonymous. However, I cannot find it.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the paper is well written and it is clear.\n\nQuality: The quality of this work is good.\n\nNovelty: The novelty is limited.\n\nReproducibility: unclear at this time.", "summary_of_the_review": "This paper proposes to evaluate efficient attentions under a more fine-grained attention taxonomy\nwith four distinguishable attention patterns, each of which reflects different attentive functionality.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666544366305}, {"id": "Ji2LSc99l8", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1790/Reviewer_VyuY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents CAB, a new benchmark to test Transformers and other sequence models on long sequence data. The paper benchmarks a number of methods as a baseline.", "review_text": "I think this paper could make a good contribution to the literature, but the paper has a few basic mistakes that should be corrected. If my concerns are addressed in the rebuttal, I'm happy to raise my score above the acceptance threshold.\n\nUpdate after rebuttal: my concerns have been addressed, so I am increasing my score to a 6.", "strengths": "+ Fills in a missing gap in the literature beyond LRA\n+ Will be useful for researchers evaluating long-sequence models\n\n- There appear to be a few basic mistakes in the methods characterization\n- It would be nice to have a few more baselines\n- Paper could use some copy edits", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents CAB, a new benchmark to test Transformers and other sequence models on long sequence data. The paper benchmarks a number of methods as a baseline.", "strength_and_weaknesses": "+ Fills in a missing gap in the literature beyond LRA\n+ Will be useful for researchers evaluating long-sequence models\n\n- There appear to be a few basic mistakes in the methods characterization\n- It would be nice to have a few more baselines\n- Paper could use some copy edits", "clarity,_quality,_novelty_and_reproducibility": "Clarity: The paper is clear to follow, but could use some copy editing. I think the focus on distinguishing self-attention from causal/cross/etc attention is a little bit misplaced.\n\nQuality: I think the most useful contribution of this paper is simply aggregating more long-sequence benchmarks as \"standard.\" The categorization of methods into self-attention, causal, cross, etc seems to have some basic mistakes.\n\nFor example, S4D is inherently a causal model (this is the definition of a state space model - it models how a state changes through time). The \"bidirectional\" or noncausal version of S4D is a modification on top. I am not as familiar with the details of the other methods, but this category error makes me concerned.\n\nFor a benchmark paper, it would be nice to see more baselines for exact attention, such as FlashAttention.\n\nOne question: which tokenizer do you use for PG-19? Making that clear up front would be helpful (it is one of the reasons that existing work is a little hard to compare).\n\nNovelty: I think this type of benchmark has been missing from the literature, so it is welcome. However, the analysis and evaluation on top are a little bit lacking.\n\nReproducibility: Seems clear.", "summary_of_the_review": "I think this paper could make a good contribution to the literature, but the paper has a few basic mistakes that should be corrected. If my concerns are addressed in the rebuttal, I'm happy to raise my score above the acceptance threshold.\n\nUpdate after rebuttal: my concerns have been addressed, so I am increasing my score to a 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666462184420}], "openreview_url": "https://openreview.net/forum?id=S80I3NwbbpS", "arxiv_id": "2210.07661", "paper_pdf": "papers/S80I3NwbbpS.pdf", "paper_pdf_sha256": "7e8413358dbe7bc2d7220fcfc45794e973f83d67a155cf34359c8e128fe77ef3", "paper_pdf_bytes": 29193277, "paper_pdf_source": "openreview", "code_url": "https://github.com/Shark-NLP/CAB", "code_repository": "Shark-NLP/CAB", "code_commit": "0e496969545d5cb7c76b87de56d9d8b7600d0ba4", "code_archive": "repos/S80I3NwbbpS.zip", "code_archive_sha256": "da98e7625841fb76bb0538b4e6a886ec505c219e38ef8c4dd4bbeaf7b0d7d13d", "code_archive_bytes": 1025078, "code_file_count": 33, "code_extensions": {".py": 27, ".sh": 6}, "github_disk_usage_kb": 1046, "github_languages": {"Python": 204944, "Shell": 1948}, "github_archived": false, "github_pushed_at": "2023-07-02T11:23:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/cab-comprehensive-attention-benchmarking-on"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ljCoTzUsdS", "year": 2022, "status": "rejected", "title": "Distinguishing rule- and exemplar-based generalization in learning systems", "authors": ["Ishita Dasgupta", "Erin Grant", "Thomas L. Griffiths"], "authorids": ["~Ishita_Dasgupta1", "~Erin_Grant1", "~Thomas_L._Griffiths1"], "authors_source": "OpenReview API", "abstract": "Despite the increasing scale of datasets in machine learning, generalization to unseen regions of the data distribution remains crucial. Such extrapolation is by definition underdetermined and is dictated by a learner’s inductive biases. Machine learning systems often do not share the same inductive biases as humans and, as a result, extrapolate in ways that are inconsistent with our expectations. We investigate two distinct such inductive biases: feature-level bias (differences in which features are more readily learned) and exemplar-vs-rule bias (differences in how these learned features are used for generalization). Exemplar- vs. rule-based generalization has been studied extensively in cognitive psychology, and in this work we present a protocol inspired by these experimental approaches for directly probing this trade-off in learning systems. The measures we propose characterize changes in extrapolation behavior when feature coverage is manipulated in a combinatorial setting. We present empirical results across a range of models and across both expository and real-world image and language domains. We demonstrate that measuring the exemplar-rule trade-off while controlling for feature-level bias provides a more complete picture of extrapolation behavior than existing formalisms. We find that most standard neural network models have a propensity towards exemplar-based extrapolation and discuss the implications of these findings for research on data augmentation, fairness, and systematic generalization.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Tdmh4Vu7vn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper152/Reviewer_UbRf"], "rating": "", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The generalization of data distributions to unseen regions, i.e., extrapolation, remains one of the critical challenges in machine learning. The inductive biases of the learner determine such extrapolation. Unfortunately, machine learning systems often do not share the same inductive biases as humans and, as a result, may extrapolate in ways that do not match the analyst's expectations. The authors investigated two different types of such inductive bias: feature-level bias (differences in which features are more easily learned) and exemplar-based and rule-based bias (differences in how learned features are used for generalization). Inspired by these experimental approaches, we have proposed a protocol to investigate this trade-off in learning systems directly. We present empirical results for a range of models and the domains of explanatory images and language. We demonstrate that controlling for feature-level bias while measuring the trade-off between exemplars and rules provides a complete picture of extrapolative behavior than existing formalisms.", "review_text": "Generalization in the domain of extrapolation, which the authors address in this paper, is one of the critical issues in machine learning. Therefore, they propose a protocol to investigate the problem of inductive bias in extrapolation. Specifically, inspired by psychological research, they propose a protocol to investigate the inductive bias of the learning system towards different features (FLB) and the inductive bias of the learning system towards different ways of using features, either by rule-based or exemplar-based generalization (EVR).\n\nThe approach, inspired by psychological research, is interesting. However, we are not convinced that it is a practically valid method, as we have only shown experiments on two datasets.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The generalization of data distributions to unseen regions, i.e., extrapolation, remains one of the critical challenges in machine learning. The inductive biases of the learner determine such extrapolation. Unfortunately, machine learning systems often do not share the same inductive biases as humans and, as a result, may extrapolate in ways that do not match the analyst's expectations. The authors investigated two different types of such inductive bias: feature-level bias (differences in which features are more easily learned) and exemplar-based and rule-based bias (differences in how learned features are used for generalization). Inspired by these experimental approaches, we have proposed a protocol to investigate this trade-off in learning systems directly. We present empirical results for a range of models and the domains of explanatory images and language. We demonstrate that controlling for feature-level bias while measuring the trade-off between exemplars and rules provides a complete picture of extrapolative behavior than existing formalisms.", "main_review": "Generalization in the domain of extrapolation, which the authors address in this paper, is one of the critical issues in machine learning. Therefore, they propose a protocol to investigate the problem of inductive bias in extrapolation. Specifically, inspired by psychological research, they propose a protocol to investigate the inductive bias of the learning system towards different features (FLB) and the inductive bias of the learning system towards different ways of using features, either by rule-based or exemplar-based generalization (EVR).\n\nThe approach, inspired by psychological research, is interesting. However, we are not convinced that it is a practically valid method, as we have only shown experiments on two datasets.", "summary_of_the_review": "The authors' treatment of inductive bias in extrapolation is interesting and will interest many researchers. The proposed protocol, which takes its ideas from psychology, is also interesting. However, we are not convinced that the proposed protocol is practical.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "1: You are unable to assess this paper and have alerted the ACs to seek an opinion from different reviewers."}, "tcdate": 1636052004109}, {"id": "_WCl8DFjdrQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper152/Reviewer_RJtk"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes measures of how two types of explanation methods -- rule based and exemplar based -- generalize and extrapolate to unseen data regions.", "review_text": "Some of the ideas introduced by the paper to think about certain inductive biases is interesting. For example, feature-level bias being which features are easier or harder to learn. Here it could be interesting to go into continuous vs. categorical features.\n\nI found some of the exposition unnecessarily confusing. For example, since in Section 2, the color of objects determines their label (i.e. green objects are “dax” and purple objects are “fep”), why use the labels \"dax\" and \"fep\" at all rather than just green and purple?\n\nThere are many claims that could be substantiated or explained further. For example, why do GPs help with formalizing exemplar-based generalization (Section 4.1)? It would also be clearer to specify which statements have been substantiated through experiments and which have not. For example, Section 4.2 starts off by saying that one NN, one GLM, and one GP has been trained and then make a jump to Figure 4a which shows decision boundaries that are not surprising; I am not sure how it validates the protocol proposed to measure EVR without confounds. Section 4.4 makes a substantial claim (\"We also find that different ways to reduce ρ (e.g. by reducing π0 or by increasing π1), give different extrapolation behavior (see Appendix)\") based on results pushed to the appendix. \n\nSome sentences could be better worded:\n- (Page 6) \"EVR increases with exemplar-basedness\" \n- (Page 7) What is the \"held-out quadrant\"?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes measures of how two types of explanation methods -- rule based and exemplar based -- generalize and extrapolate to unseen data regions.", "main_review": "Some of the ideas introduced by the paper to think about certain inductive biases is interesting. For example, feature-level bias being which features are easier or harder to learn. Here it could be interesting to go into continuous vs. categorical features.\n\nI found some of the exposition unnecessarily confusing. For example, since in Section 2, the color of objects determines their label (i.e. green objects are “dax” and purple objects are “fep”), why use the labels \"dax\" and \"fep\" at all rather than just green and purple?\n\nThere are many claims that could be substantiated or explained further. For example, why do GPs help with formalizing exemplar-based generalization (Section 4.1)? It would also be clearer to specify which statements have been substantiated through experiments and which have not. For example, Section 4.2 starts off by saying that one NN, one GLM, and one GP has been trained and then make a jump to Figure 4a which shows decision boundaries that are not surprising; I am not sure how it validates the protocol proposed to measure EVR without confounds. Section 4.4 makes a substantial claim (\"We also find that different ways to reduce ρ (e.g. by reducing π0 or by increasing π1), give different extrapolation behavior (see Appendix)\") based on results pushed to the appendix. \n\nSome sentences could be better worded:\n- (Page 6) \"EVR increases with exemplar-basedness\" \n- (Page 7) What is the \"held-out quadrant\"?\n", "summary_of_the_review": "The paper is ambitious but falls short in clearly explaining the framework proposed, and substantiating the idea with strong experiments.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636015222826}, {"id": "26h4hnV19sn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper152/Reviewer_yoH5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes two measures for evaluating the feature-level bias and the exemplar-vs-rule bias in learning systems. Specifically, the authors designed three independent training conditions:\n    i. cue conflict: {x1 = 0, x2 = 1, y = 0}, {x1 = 1, x2 = 0, y = 1}\n    ii. zero shot: {x1 = 0, x2 = 0, y = 0}, {x1 = 1, x2 = 0, y = 1}\n    iii. partial exposure: {x1 = 0, x2 = 0, y = 0}, {x1 = 1, x2 = 0, y = 1}, {x1 = 0, x2 = 1, y = 0}\nand one testing condition:\n    iv. extrapolation: {x1 = 1, x2 = 1, y = 1}.\nThe inductive bias of a given learning system is measured by its extrapolation performance difference when trained on different training conditions. \n\nEmpirically, the authors first verified their framework on a synthetic dataset with 2D inputs. The results confirm that generalized linear model favors rule-based generalization while Gaussian process favors exemplar-based generalization. On IMDB, the authors show that LSTM models exhibit high feature-level bias (overfitting to the spurious token features) favors exemplar-rule bias. On CelebA, the authors show that ResNet exhibits a wide range of feature-level bias for different features (‘male’ is easier to learn than ‘high cheekbones’). However, across all feature pairs, the model prefers exemplar-based generalization to rule-based generalization.", "review_text": "Strength:\n+ Probing the behavior of the learning system by comparing its generalization performance when trained on different datasets is interesting and novel.\n+ The synthetic experiment on 2D inputs demonstrate the effectiveness of the measure. \n+ The paper is very well-written.\n\nWeakness:\n+ Unclear utility.\nThe main contributions of the paper are the two proposed measures. However, I found it hard to utilize these measures in practical applications. \n  + In order to compute the measures, the values for the discriminant features and the distractor features are required. In this situation, there are many existing methods for learning models that are robust against the distractor features [1]. What will be the point of evaluating the bias of a non-robust classifier?\n  + For the exemplar-vs-rule propensity (EVR), my understanding is that it captures the extent to which the distractor features are utilized in the learning system. What will be the benefit of EVR compared to evaluating the worst-group accuracy as in [1]? Let’s use CelebA as an example. The training data contains mostly {male, dark_hair}, {female, blond_hair}, {female, dark_hair} images. At test time, [1] evaluates the worst-group accuracy, which is likely to be the performance on {male, blond_hair} (since it is underrepresented in the training data).\n\nQuestions and suggestions:\n+ How do you interpret the values of the FLB and EVR measures? For instance, in IMDB, the FLB score is -0.3 while the EVR score is 0.54. What do these values mean?\n+ I think both measures can be defined in a more generalized setting (where you don’t need to assume the label marginal is uniform).\n+ You argued in the paper that the measure is capturing something more than the spurious correlation. I think it will be more precise to state it as **linear** correlation in Figure 3.\n+ It will be very interesting to see a contour plot of EVR over $\\pi_0$ and $\\pi_1$ even in the synthetic experiments.\n\nI am happy to adjust my ratings after the rebuttal period.\n\n[1] Sagawa, Shiori, et al. \"Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization.\" arXiv preprint arXiv:1911.08731 (2019).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes two measures for evaluating the feature-level bias and the exemplar-vs-rule bias in learning systems. Specifically, the authors designed three independent training conditions:\n    i. cue conflict: {x1 = 0, x2 = 1, y = 0}, {x1 = 1, x2 = 0, y = 1}\n    ii. zero shot: {x1 = 0, x2 = 0, y = 0}, {x1 = 1, x2 = 0, y = 1}\n    iii. partial exposure: {x1 = 0, x2 = 0, y = 0}, {x1 = 1, x2 = 0, y = 1}, {x1 = 0, x2 = 1, y = 0}\nand one testing condition:\n    iv. extrapolation: {x1 = 1, x2 = 1, y = 1}.\nThe inductive bias of a given learning system is measured by its extrapolation performance difference when trained on different training conditions. \n\nEmpirically, the authors first verified their framework on a synthetic dataset with 2D inputs. The results confirm that generalized linear model favors rule-based generalization while Gaussian process favors exemplar-based generalization. On IMDB, the authors show that LSTM models exhibit high feature-level bias (overfitting to the spurious token features) favors exemplar-rule bias. On CelebA, the authors show that ResNet exhibits a wide range of feature-level bias for different features (‘male’ is easier to learn than ‘high cheekbones’). However, across all feature pairs, the model prefers exemplar-based generalization to rule-based generalization.", "main_review": "Strength:\n+ Probing the behavior of the learning system by comparing its generalization performance when trained on different datasets is interesting and novel.\n+ The synthetic experiment on 2D inputs demonstrate the effectiveness of the measure. \n+ The paper is very well-written.\n\nWeakness:\n+ Unclear utility.\nThe main contributions of the paper are the two proposed measures. However, I found it hard to utilize these measures in practical applications. \n  + In order to compute the measures, the values for the discriminant features and the distractor features are required. In this situation, there are many existing methods for learning models that are robust against the distractor features [1]. What will be the point of evaluating the bias of a non-robust classifier?\n  + For the exemplar-vs-rule propensity (EVR), my understanding is that it captures the extent to which the distractor features are utilized in the learning system. What will be the benefit of EVR compared to evaluating the worst-group accuracy as in [1]? Let’s use CelebA as an example. The training data contains mostly {male, dark_hair}, {female, blond_hair}, {female, dark_hair} images. At test time, [1] evaluates the worst-group accuracy, which is likely to be the performance on {male, blond_hair} (since it is underrepresented in the training data).\n\nQuestions and suggestions:\n+ How do you interpret the values of the FLB and EVR measures? For instance, in IMDB, the FLB score is -0.3 while the EVR score is 0.54. What do these values mean?\n+ I think both measures can be defined in a more generalized setting (where you don’t need to assume the label marginal is uniform).\n+ You argued in the paper that the measure is capturing something more than the spurious correlation. I think it will be more precise to state it as **linear** correlation in Figure 3.\n+ It will be very interesting to see a contour plot of EVR over $\\pi_0$ and $\\pi_1$ even in the synthetic experiments.\n\nI am happy to adjust my ratings after the rebuttal period.\n\n[1] Sagawa, Shiori, et al. \"Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization.\" arXiv preprint arXiv:1911.08731 (2019).", "summary_of_the_review": "The authors proposed an interesting approach for measuring the rule- and exemplar-based  generalization for a given learning system. The perspective is novel and the writing is clear. My only concern is the practical utility of the proposed measures.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635791197406}, {"id": "gNz0BNNARxZ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper152/Reviewer_TPBn"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper studies the extrapolation of machine learning models to unseen regions, and specifically studies two types of biases: feature-level bias (differences in which features are more readily learned) and exemplar-vs-rule bias (differences in how these learned features are used for generalization). Motivated by the studies of exemplar vs. rule-based generalization in cognitive psychology, the authors present a protocol directly probing this trade-off in machine learning systems. The authors present empirical results across a range of models in both expository and real-world image and language domains and demonstrate that using the trad off provides a more complete picture of extrapolation behaviour than existing methods. \n", "review_text": "I found that this paper was quite difficult to read since the authors present discussions without clear definitions. Using many acronyms makes the reading difficult too. \n\nFor example, “feature-level bias is measured as deviation from chance performance in the CC condition.” When reading CC condition, it says “the data presented in this condition confound color and shape”. What does confound mean here? CC condition is unclear and I do not know feature-level bias.\n\nSimilarly, “Exemplar-vs-rule bias is measured by the difference between performance in the PE and ZS conditions—…” When reading PE and ZS, they are not defined either. \n\nReading on, I do not see formal definitions of feature-level bias exemplar-vs-rule bias. There are measures obtained from models trained by some specific methods. So the proposed approach deals specifically with three types of methods. Why do the authors focus on three types of methods? Why are they representative? \n\nAlso, I do not know what spurious correlation means in this paper. \n\nThe authors present quite some examples to illustrate their points, but I am confused by their examples. For example, in Figure 1, only two training examples are given. Any prediction is incorrect since there is not enough evidence to learn a model. I cannot see the differences between rule-based and example-based. What rules can we generate based on the two examples?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the extrapolation of machine learning models to unseen regions, and specifically studies two types of biases: feature-level bias (differences in which features are more readily learned) and exemplar-vs-rule bias (differences in how these learned features are used for generalization). Motivated by the studies of exemplar vs. rule-based generalization in cognitive psychology, the authors present a protocol directly probing this trade-off in machine learning systems. The authors present empirical results across a range of models in both expository and real-world image and language domains and demonstrate that using the trad off provides a more complete picture of extrapolation behaviour than existing methods. \n", "main_review": "I found that this paper was quite difficult to read since the authors present discussions without clear definitions. Using many acronyms makes the reading difficult too. \n\nFor example, “feature-level bias is measured as deviation from chance performance in the CC condition.” When reading CC condition, it says “the data presented in this condition confound color and shape”. What does confound mean here? CC condition is unclear and I do not know feature-level bias.\n\nSimilarly, “Exemplar-vs-rule bias is measured by the difference between performance in the PE and ZS conditions—…” When reading PE and ZS, they are not defined either. \n\nReading on, I do not see formal definitions of feature-level bias exemplar-vs-rule bias. There are measures obtained from models trained by some specific methods. So the proposed approach deals specifically with three types of methods. Why do the authors focus on three types of methods? Why are they representative? \n\nAlso, I do not know what spurious correlation means in this paper. \n\nThe authors present quite some examples to illustrate their points, but I am confused by their examples. For example, in Figure 1, only two training examples are given. Any prediction is incorrect since there is not enough evidence to learn a model. I cannot see the differences between rule-based and example-based. What rules can we generate based on the two examples?\n", "summary_of_the_review": "The paper is unclear based on the current presentation.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635313991263}], "openreview_url": "https://openreview.net/forum?id=ljCoTzUsdS", "arxiv_id": "2110.04328", "paper_pdf": "papers/ljCoTzUsdS.pdf", "paper_pdf_sha256": "9d5f5cbe3ec8678839d853e81fb3fe6e9971d6cc1148f0606440c19023af2bdb", "paper_pdf_bytes": 2408901, "paper_pdf_source": "openreview", "code_url": "https://github.com/eringrant/icml-2022-rules-vs-exemplars", "code_repository": "eringrant/icml-2022-rules-vs-exemplars", "code_commit": "c63fc90aa5862eeb929f66f29e9837fba96be8d4", "code_archive": "repos/ljCoTzUsdS.zip", "code_archive_sha256": "1fccbf918ec31e074311383e62fb136206c74fcc5cee25f2c92b80e44f025f5d", "code_archive_bytes": 4767240, "code_file_count": 20, "code_extensions": {".py": 12, ".ipynb": 7, ".r": 1}, "github_disk_usage_kb": 4648, "github_languages": {"Jupyter Notebook": 6727174, "Python": 58534, "R": 9956}, "github_archived": false, "github_pushed_at": "2023-11-15T15:48:12Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distinguishing-rule-and-exemplar-based-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IU8QxEiG4hR", "year": 2021, "status": "rejected", "title": "SBEVNet: End-to-End Deep Stereo Layout Estimation", "authors": ["Divam Gupta", "Wei Pu", "Trenton Tabor", "Jeff Schneider"], "authorids": ["~Divam_Gupta1", "wpu@nrec.ri.cmu.edu", "~Trenton_Tabor1", "~Jeff_Schneider1"], "authors_source": "OpenReview API", "abstract": "Accurate layout estimation is crucial for planning and navigation, for robotics applications such as self driving. In this paper, we introduce stereo bird's eye view network SBEVNet, a novel supervised end-to-end framework for estimation of bird's eye view layout from a pair of stereo images. Although our network reuses the building blocks from the state-of-the-art deep learning networks for disparity estimation, we show that accurate depth estimation is neither sufficient nor necessary. Instead, the learning of a good internal bird's eye view feature representation is essential for layout estimation. Specifically, we first generate a disparity feature volume using the features of the stereo images and then project it to the bird's eye view coordinates. This gives us coarse grained scene structural information. We also apply inverse perspective mapping (IPM) to map the input images and their features to the bird's eye view. This gives us fine grained texture information. The concatenated IPM features with the projected feature volume creates a rich bird's eye view representation which is capable of spatial reasoning. We use this representation to estimate the BEV semantic map. Additionally, we show that using the IPM features as a supervisory signal for stereo features can give an improvement in performance. We demonstrate our approach on two datasets: KITTI dataset and synthetically generated dataset using the CARLA simulator. For both of the datasets, we establish state-of-the-art performance beyond other baselines.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "jyh-d33SwOA", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2861/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\n## Contributions\n\nThis paper presents SBEVNet, a neural network architecture to estimate the bird's-eye view (BEV) layout of an urban driving scene. Given an image captured by a stereo camera, SBEVNet performs an inverse perspective mapping (IPM) to obtain an initial feature volume, which is further processed to generate the BEV layout. The system is trained end-to-end in a supervised learning setup.\n\n## Strengths\n\n**S1** The problem considered here is very relevant to perception groups in the autonomous driving community. This area has only recently seen work crop up. Approaches like MonoLayout [A], MonoOccupancy [B], and PseudoLidar [C] are closely related to this submission.\n\n**S2** The paper is easy to follow, and provides a majority of the details needed to understand and assess the approach.\n\n**S3** The authors also seem to provide code (and promise a public release), which might help ensure reproducibility.\n\n## Weaknesses\n\nI see a few major and a number of other minor concerns that impact my perception of this paper. I'm hoping the discussion period helps address some of these, and I'm open to revising my score in light of evidence contrary to the following claims.\n\nIt appears that this paper uses MonoLayout [A], MonoOccupancy [B], and PseudoLidar [C] as primary baselines. Much of my review stems from my understanding of [A, B, C] (and my 'surprise' at a few contradictory trends observed in this paper.)\n\n**Problem setup** It is unclear from reading the paper and supplementary material if the problem setup is infact \"amodal\" layout estimation (i.e., if scene points outside of the camera view are predicted in the BEV layout). Approaches like (Schulter et al., 2016) and (Mani et al., 2020) operate in this \"amodal\" setup, while others such as PseudoLidar [C] and (Lu et al., 2019) only predict points that are visible in the input image. Does this approach, for instance, hallucianate hidden intersections and roads? (It seems not, since a visibility mask is explicitly employed in the loss function -- cf. Fig. 1 and Eq. 12, 13).\n\n**MonoLayout baseline** The primary baseline considered in this paper is \"MonoLayout\" (Mani et al., 2020). Upon examining the MonoLayout [A] paper, I find a surprising and troubling trend. This paper reports very poor performances of MonoLayout on the KITTI dataset (the original MonoLayout paper reports mIoU for the \"car\" class to be around 26.08, while the current submission reports 2.43 -- cf. Table 2). I've noted that MonoLayout makes its code and models publicly available (its publicly available pretrained models claim an mIoU of 30.18 for the \"car\" class), as highlighted on their GitHub page. Also, other baselines like \"MonoOccupancy\" have surprisingly low scores in this paper (an order of magnitude), compared to scores reported in the MonoLayout paper. I wonder if there is something different in the experiment and/or training protocols employed in the current work, as opposed to those in the MonoOccupancy and MonoLayout papers? For example, the MonoOccupancy baseline as reported in the MonoLayout paper achieves an mIoU of about 24.16 (for the car class) (MonoLayout paper - Table 1), while the same baseline has a dismal performance (mIoU of 7.11 for car class) in Table 2 of the current manuscript.\n\nThe fact that this performance gap is not explained in the paper makes it hard to analyze the merits of the proposed approach. Save for a single sentence \"The results of MonoLayout ... and MonoOccupancy ... are inferior due to lack of any camera geometry priors in the network\", I've not found any other discussion of this performance gap/discrepancy.\n\nI also find it a tad weird (and unexplained) that the performance of various baselines do not seem to follow a set pattern/trend across the CARLA and KITTI datasets. In the MonoLayout paper, I notice that changing the dataset from KITTI to Argoverse does change absolute mIoU scores a bit, but preserves the ranking of various baselines (i.e., MonoLayout > OFT > MonoOccupancy on both KITTI and Argoverse). In the current submission, the trends seem to be changing across the two datasets (cf. Tables 1, 2).\n\nYet another set of baselines that seem to underperform here are the PseudoLidar variants. In the MonoLayout paper (cf. supplementary material, Table 5), Pseudolidar is evaluated on the KITTI dataset, and the reported mIoU for vehicles is 59, whereas in this paper the best performance on this class achieved by a pseudolidar model is 45.64. Further, the MonoLayout paper's version of the (stereo) Pseudolidar baseline seems to perform quite competetively (mIoU 59.0) to SBEVNet Ensemble (mIoU 60.17 for \"car\", cf. Table 2). This seems to indicate that well-tuned baselines could perhaps achieve better performance?\n\nIn Appendix A.2, the authors seem to indicate that they used a very different process to train MonoLayout (i.e., using random images from the train set as opposed to using OpenStreetMap and/or adversarial training). I suspect this might have resulted in a performance gap?\n\nI feel that OFT [D] could be cited and used as a baseline, particularly to measure layout estimation accuracy for the \"car\" class.\n\n**Qualitative results** Unfortunately, there seems to be a dearth of qualitative result figures to get a better sense of the approach. In particular MonoLayout and MonoOccupancy seem to obtain crisp reconstructions of cars (cf. MonoLayout paper), while in Figure 2., cars are splayed throughout the image in the SBEVNet results. This is also surprising; in my opinion, these results do not adequately substantiate the impressive reported mIoU.\n\n**Missing mAP metric** Other papers such as MonoLayout and OFT seem to report the mAP (mean average precision) metric in addition to the mIoU metric, because mAP often turns out to be a more accurate estimate of prediction performance (due to integrating over various recall values). In practice, this leads to less-than-perfect predictions being scored well (and this could explain the splayed-out results in Fig. 2 scoring a high mIoU). Evaluating mAP would be a stricter criteria, and will allow an additional point of comparison with prior art.\n\n\n## Minor remarks\n\nThe following remarks have had no impact on my assessment of the paper, and as such I don't expect the authors to respond to these.\n\nConcurrent approaches such as [F] can be cited and discussed.\n\nThe paper could be structured better. For instance, input image sizes and baselines could be moved over to the main paper, rather than being listed in the appendix.\n\n## References\n\n[A] Mani, Kaustubh, et al. \"MonoLayout: Amodal scene layout from a single image.\" The IEEE Winter Conference on Applications of Computer Vision. 2020.\n\n[B] Lu, Chenyang, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. \"Monocular semantic occupancy grid mapping with convolutional variational encoder–decoder networks.\" IEEE Robotics and Automation Letters 4.2 (2019): 445-452.\n\n[C] Wang, Yan, et al. \"Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n\n[D] Roddick, Thomas, Alex Kendall, and Roberto Cipolla. \"Orthographic feature transform for monocular 3d object detection.\" arXiv preprint arXiv:1811.08188 (2018).\n\n[E] A Parametric Top-View Representation of Complex Road Scenes. CVPR 2019.\n\n[F] Lift, Splat, Shoot: Encoding Images From Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. ECCV 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Concerns about experiment setup/design", "review": "\n## Contributions\n\nThis paper presents SBEVNet, a neural network architecture to estimate the bird's-eye view (BEV) layout of an urban driving scene. Given an image captured by a stereo camera, SBEVNet performs an inverse perspective mapping (IPM) to obtain an initial feature volume, which is further processed to generate the BEV layout. The system is trained end-to-end in a supervised learning setup.\n\n## Strengths\n\n**S1** The problem considered here is very relevant to perception groups in the autonomous driving community. This area has only recently seen work crop up. Approaches like MonoLayout [A], MonoOccupancy [B], and PseudoLidar [C] are closely related to this submission.\n\n**S2** The paper is easy to follow, and provides a majority of the details needed to understand and assess the approach.\n\n**S3** The authors also seem to provide code (and promise a public release), which might help ensure reproducibility.\n\n## Weaknesses\n\nI see a few major and a number of other minor concerns that impact my perception of this paper. I'm hoping the discussion period helps address some of these, and I'm open to revising my score in light of evidence contrary to the following claims.\n\nIt appears that this paper uses MonoLayout [A], MonoOccupancy [B], and PseudoLidar [C] as primary baselines. Much of my review stems from my understanding of [A, B, C] (and my 'surprise' at a few contradictory trends observed in this paper.)\n\n**Problem setup** It is unclear from reading the paper and supplementary material if the problem setup is infact \"amodal\" layout estimation (i.e., if scene points outside of the camera view are predicted in the BEV layout). Approaches like (Schulter et al., 2016) and (Mani et al., 2020) operate in this \"amodal\" setup, while others such as PseudoLidar [C] and (Lu et al., 2019) only predict points that are visible in the input image. Does this approach, for instance, hallucianate hidden intersections and roads? (It seems not, since a visibility mask is explicitly employed in the loss function -- cf. Fig. 1 and Eq. 12, 13).\n\n**MonoLayout baseline** The primary baseline considered in this paper is \"MonoLayout\" (Mani et al., 2020). Upon examining the MonoLayout [A] paper, I find a surprising and troubling trend. This paper reports very poor performances of MonoLayout on the KITTI dataset (the original MonoLayout paper reports mIoU for the \"car\" class to be around 26.08, while the current submission reports 2.43 -- cf. Table 2). I've noted that MonoLayout makes its code and models publicly available (its publicly available pretrained models claim an mIoU of 30.18 for the \"car\" class), as highlighted on their GitHub page. Also, other baselines like \"MonoOccupancy\" have surprisingly low scores in this paper (an order of magnitude), compared to scores reported in the MonoLayout paper. I wonder if there is something different in the experiment and/or training protocols employed in the current work, as opposed to those in the MonoOccupancy and MonoLayout papers? For example, the MonoOccupancy baseline as reported in the MonoLayout paper achieves an mIoU of about 24.16 (for the car class) (MonoLayout paper - Table 1), while the same baseline has a dismal performance (mIoU of 7.11 for car class) in Table 2 of the current manuscript.\n\nThe fact that this performance gap is not explained in the paper makes it hard to analyze the merits of the proposed approach. Save for a single sentence \"The results of MonoLayout ... and MonoOccupancy ... are inferior due to lack of any camera geometry priors in the network\", I've not found any other discussion of this performance gap/discrepancy.\n\nI also find it a tad weird (and unexplained) that the performance of various baselines do not seem to follow a set pattern/trend across the CARLA and KITTI datasets. In the MonoLayout paper, I notice that changing the dataset from KITTI to Argoverse does change absolute mIoU scores a bit, but preserves the ranking of various baselines (i.e., MonoLayout > OFT > MonoOccupancy on both KITTI and Argoverse). In the current submission, the trends seem to be changing across the two datasets (cf. Tables 1, 2).\n\nYet another set of baselines that seem to underperform here are the PseudoLidar variants. In the MonoLayout paper (cf. supplementary material, Table 5), Pseudolidar is evaluated on the KITTI dataset, and the reported mIoU for vehicles is 59, whereas in this paper the best performance on this class achieved by a pseudolidar model is 45.64. Further, the MonoLayout paper's version of the (stereo) Pseudolidar baseline seems to perform quite competetively (mIoU 59.0) to SBEVNet Ensemble (mIoU 60.17 for \"car\", cf. Table 2). This seems to indicate that well-tuned baselines could perhaps achieve better performance?\n\nIn Appendix A.2, the authors seem to indicate that they used a very different process to train MonoLayout (i.e., using random images from the train set as opposed to using OpenStreetMap and/or adversarial training). I suspect this might have resulted in a performance gap?\n\nI feel that OFT [D] could be cited and used as a baseline, particularly to measure layout estimation accuracy for the \"car\" class.\n\n**Qualitative results** Unfortunately, there seems to be a dearth of qualitative result figures to get a better sense of the approach. In particular MonoLayout and MonoOccupancy seem to obtain crisp reconstructions of cars (cf. MonoLayout paper), while in Figure 2., cars are splayed throughout the image in the SBEVNet results. This is also surprising; in my opinion, these results do not adequately substantiate the impressive reported mIoU.\n\n**Missing mAP metric** Other papers such as MonoLayout and OFT seem to report the mAP (mean average precision) metric in addition to the mIoU metric, because mAP often turns out to be a more accurate estimate of prediction performance (due to integrating over various recall values). In practice, this leads to less-than-perfect predictions being scored well (and this could explain the splayed-out results in Fig. 2 scoring a high mIoU). Evaluating mAP would be a stricter criteria, and will allow an additional point of comparison with prior art.\n\n\n## Minor remarks\n\nThe following remarks have had no impact on my assessment of the paper, and as such I don't expect the authors to respond to these.\n\nConcurrent approaches such as [F] can be cited and discussed.\n\nThe paper could be structured better. For instance, input image sizes and baselines could be moved over to the main paper, rather than being listed in the appendix.\n\n## References\n\n[A] Mani, Kaustubh, et al. \"MonoLayout: Amodal scene layout from a single image.\" The IEEE Winter Conference on Applications of Computer Vision. 2020.\n\n[B] Lu, Chenyang, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. \"Monocular semantic occupancy grid mapping with convolutional variational encoder–decoder networks.\" IEEE Robotics and Automation Letters 4.2 (2019): 445-452.\n\n[C] Wang, Yan, et al. \"Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving.\" Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.\n\n[D] Roddick, Thomas, Alex Kendall, and Roberto Cipolla. \"Orthographic feature transform for monocular 3d object detection.\" arXiv preprint arXiv:1811.08188 (2018).\n\n[E] A Parametric Top-View Representation of Complex Road Scenes. CVPR 2019.\n\n[F] Lift, Splat, Shoot: Encoding Images From Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. ECCV 2020.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604042026168}, {"id": "HYETgkIjh0u", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2861/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposed to estimate the semantic layout in the bird eye's view from a pair of stereo images. The main novelty/contribution lies in how to organize and exploit the information from the stereo images. The proposed framework builds upon inverse perspective mapping, and projected stereo feature volume. The performance was evaluated on the KITTI and CARLA datasets. Given a pair of stereo images, there are various options to exploit the image information, where this paper provides a framework by exploiting the stereo information in the bird eye's view. \n\n- A principled question is what is the real superiority of estimating the layout in the bird eye's view. From the application' view, the semantic estimation from the camera'view already provide much information which the stereo images could further improve the performance. From the applications's perspective, I would like to see discussions and experiments in showing the superiority in using the bird eye's representation. \n\n- In the ablation studies, the paper already provide different variants of the network architecture in exploiting the stereo image information. I believe there are multi-task learning based framework form this task, where the semantic layout estimation and stereo estimation are jointly estimated and optimized. Whether that pipeline will provide extra benefit?\n\n- In section 3.4.4, the paper claimed that \"We pass the concatenated stereo BEV feature map and IPM BEV feature map to a U-Net (Ronneberger et al., 2015) network to generate the semantic map C\". However, the loss evaluation applies to the IPM features and Stereo features separately, namely, $\\mathcal{C}_i^{IPM}$ and $\\mathcal{C}_i^{Stereo}$. If two estimations are made as the network output, which one will be used for performance evaluation? The other following question is : If two separate estimations are made as the network outputs and compared with the ground truth for loss evaluation, whether a consistency loss between these two estimations will further constrain the network learning?\n\n- The paper conducted experiments on the KITTI and CARLA dataset. It is well understood that the CityScape dataset has been widely in evaluating semantic segmentation where the stereo images are available. I would to see more evaluation on these real-image dataset rather than synthetic dataset such as the CARLA dataset.\n\n- The paper title and abstract should highlight \"semantic\" and \"bird eye's view\" as the paper proposed to learn the semantic layout in the bird eye's view. The current title did reflect these properties.\n\nAll in all, taking all the above comments into consideration, I would like to hear from the authors' response, which could lead to updated rating in either directions.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper proposed to estimate the semantic layout in the bird eye's view from stereo images. The main novelties lie in how to organize and exploit the information from the stereo images. The proposed framework builds upon inverse perspective mapping, projected stereo feature volume. The performance was evaluated on the KITTI and CARLA datasets.", "review": "The paper proposed to estimate the semantic layout in the bird eye's view from a pair of stereo images. The main novelty/contribution lies in how to organize and exploit the information from the stereo images. The proposed framework builds upon inverse perspective mapping, and projected stereo feature volume. The performance was evaluated on the KITTI and CARLA datasets. Given a pair of stereo images, there are various options to exploit the image information, where this paper provides a framework by exploiting the stereo information in the bird eye's view. \n\n- A principled question is what is the real superiority of estimating the layout in the bird eye's view. From the application' view, the semantic estimation from the camera'view already provide much information which the stereo images could further improve the performance. From the applications's perspective, I would like to see discussions and experiments in showing the superiority in using the bird eye's representation. \n\n- In the ablation studies, the paper already provide different variants of the network architecture in exploiting the stereo image information. I believe there are multi-task learning based framework form this task, where the semantic layout estimation and stereo estimation are jointly estimated and optimized. Whether that pipeline will provide extra benefit?\n\n- In section 3.4.4, the paper claimed that \"We pass the concatenated stereo BEV feature map and IPM BEV feature map to a U-Net (Ronneberger et al., 2015) network to generate the semantic map C\". However, the loss evaluation applies to the IPM features and Stereo features separately, namely, $\\mathcal{C}_i^{IPM}$ and $\\mathcal{C}_i^{Stereo}$. If two estimations are made as the network output, which one will be used for performance evaluation? The other following question is : If two separate estimations are made as the network outputs and compared with the ground truth for loss evaluation, whether a consistency loss between these two estimations will further constrain the network learning?\n\n- The paper conducted experiments on the KITTI and CARLA dataset. It is well understood that the CityScape dataset has been widely in evaluating semantic segmentation where the stereo images are available. I would to see more evaluation on these real-image dataset rather than synthetic dataset such as the CARLA dataset.\n\n- The paper title and abstract should highlight \"semantic\" and \"bird eye's view\" as the paper proposed to learn the semantic layout in the bird eye's view. The current title did reflect these properties.\n\nAll in all, taking all the above comments into consideration, I would like to hear from the authors' response, which could lead to updated rating in either directions.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603960258353}, {"id": "b0OVQgJaUM5", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2861/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Interesting problem but the paper can be improved\n\nThis work aims to directly estimate the world layout in front of the vehicle from a pair of stereo cameras. It is based on cost volume but it does not explicitly predict the depth values of each pixel. Instead, it warps the cost volume features to the bird eye view (BEV) and do semantic segmentation from BEV using U-Net.\n\nI think the problem is interesting and I believe it has never been mentioned and/or addressed before. Moreover I think the motivation is also valid, as the BEV semantic segmentation from camera sensors can be one important perception input for navigation and planning. \n\nI like the idea of skipping the explicit 3D reconstruction and directly shoot for the final goal; I believe we usually get better performance when we directly minimize the loss we want to minimize. Moreover, it could potentially introduce some inspiration to other works, e.g., the direct extension - 3D (point/volume) semantic segmentation.\n\nThough I like this work, I also has several concerns:\n1. Is IPM feature really important? I only see it is effective on the synthetic dataset CARLA but not on KITTI. What is the possible reason? My guess is that the ground estimation is very bad for the real-world data. I am also curious what is the performance if only IPM feature is used. \n2. In introduction, this paper claims estimating accurate depth is not sufficient due to occlusion. However, I don't see how this work could handle occlusion. Instead the occluded part is masked out during training. Please explain this statement.\n3. What is the range of the layout estimation? From CARLA, it is 39m, and from Figure 5 and Figure 6, it is 35m. If it is the case, the short range of the estimation makes it hard to act as a major component in the perception system; the best use case is for short range detection and system redundancy. But actually I can imagine that it would not get very good result in long range, as there is always a trade-off between baseline (for accuracy) and camera overlap (for coverage) in stereo estimation.\n4. What is the image resolution for the inference time test? It seems quite slow if the resolution is 512x288 (for CARLA) or 640x256 (for KITTI).\n5. For the experiment, I think it is better to report mean $\\pm$ std with multiple trainings, as there are training noise.\n\nOther suggestions and clarifications\n1. When IPM is first introduced in Page 2, it is better to explain it in a short sentence. The current version is not clear and there are typos.\n2. I believe this work is based on binocular stereo pairs (correct me if I am wrong), so please explicitly say that in the paper. Also, using left/right image instead of reference/target image is less misleading.\n3. For disparity feature volume, it is better to use the prevalent name - cost volume. It is called cost volume in the introduction but later called disparity feature volume, I think it is better to be consistent.\n4. It is unclear how IPM feature are obtained: from pre-determined parameters or ground estimation? I think pre-determined parameters will not work very well because ground is not always a perfect plane.\n5. It is unclear what is the ensemble method used here. If it just takes the best of several models, I will not be convinced.\n\nAfter rebuttal:\nI still think this work has a interesting task setup, though it indeed has many faults (after reading the responses and other reviewers):\n1. It seems that IPM is not really useful in practice.\n2. It is also not sufficient to large occlusion, and thus there is no explanation for its advatange over `estimating accurate depth`\n3. Range is short and latency is high\n4. After reading reviewer1's comments, I think it could use the same experimental setting as the existing methods for a fair comparison. The other methods might be not properly trained with the new setting.\n5. It is still not clear how to emsemble several models (with different trained weights) in this work.\nThus I am changin my rating to 6, and I will not fight for this work.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting problem but the paper can be improved", "review": "Interesting problem but the paper can be improved\n\nThis work aims to directly estimate the world layout in front of the vehicle from a pair of stereo cameras. It is based on cost volume but it does not explicitly predict the depth values of each pixel. Instead, it warps the cost volume features to the bird eye view (BEV) and do semantic segmentation from BEV using U-Net.\n\nI think the problem is interesting and I believe it has never been mentioned and/or addressed before. Moreover I think the motivation is also valid, as the BEV semantic segmentation from camera sensors can be one important perception input for navigation and planning. \n\nI like the idea of skipping the explicit 3D reconstruction and directly shoot for the final goal; I believe we usually get better performance when we directly minimize the loss we want to minimize. Moreover, it could potentially introduce some inspiration to other works, e.g., the direct extension - 3D (point/volume) semantic segmentation.\n\nThough I like this work, I also has several concerns:\n1. Is IPM feature really important? I only see it is effective on the synthetic dataset CARLA but not on KITTI. What is the possible reason? My guess is that the ground estimation is very bad for the real-world data. I am also curious what is the performance if only IPM feature is used. \n2. In introduction, this paper claims estimating accurate depth is not sufficient due to occlusion. However, I don't see how this work could handle occlusion. Instead the occluded part is masked out during training. Please explain this statement.\n3. What is the range of the layout estimation? From CARLA, it is 39m, and from Figure 5 and Figure 6, it is 35m. If it is the case, the short range of the estimation makes it hard to act as a major component in the perception system; the best use case is for short range detection and system redundancy. But actually I can imagine that it would not get very good result in long range, as there is always a trade-off between baseline (for accuracy) and camera overlap (for coverage) in stereo estimation.\n4. What is the image resolution for the inference time test? It seems quite slow if the resolution is 512x288 (for CARLA) or 640x256 (for KITTI).\n5. For the experiment, I think it is better to report mean $\\pm$ std with multiple trainings, as there are training noise.\n\nOther suggestions and clarifications\n1. When IPM is first introduced in Page 2, it is better to explain it in a short sentence. The current version is not clear and there are typos.\n2. I believe this work is based on binocular stereo pairs (correct me if I am wrong), so please explicitly say that in the paper. Also, using left/right image instead of reference/target image is less misleading.\n3. For disparity feature volume, it is better to use the prevalent name - cost volume. It is called cost volume in the introduction but later called disparity feature volume, I think it is better to be consistent.\n4. It is unclear how IPM feature are obtained: from pre-determined parameters or ground estimation? I think pre-determined parameters will not work very well because ground is not always a perfect plane.\n5. It is unclear what is the ensemble method used here. If it just takes the best of several models, I will not be convinced.\n\nAfter rebuttal:\nI still think this work has a interesting task setup, though it indeed has many faults (after reading the responses and other reviewers):\n1. It seems that IPM is not really useful in practice.\n2. It is also not sufficient to large occlusion, and thus there is no explanation for its advatange over `estimating accurate depth`\n3. Range is short and latency is high\n4. After reading reviewer1's comments, I think it could use the same experimental setting as the existing methods for a fair comparison. The other methods might be not properly trained with the new setting.\n5. It is still not clear how to emsemble several models (with different trained weights) in this work.\nThus I am changin my rating to 6, and I will not fight for this work.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603860712561}, {"id": "8mo0cvbXIpn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2861/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes an end-to-end network for layout estimation from stereo images. The approach is built off previous stereo matching networks, which built and process a 3D disparity volume. The stereo estimate is used to project image features into a birds-eye-view representation which is processed using a U-net which predicts a semantic scene layout. The approach is evaluated on the KITTI and Carla generated datasets.\n\nStrengths:\n* This is the first work to attempt semantic layout estimation from stereo images\n* The approach is geometrically grounded, and can properly leverage stereo information to improve layout estimation\n* The approach performs well on the two datasets evaluated. Since this paper is focused on a new problem, there are not existing works to directly compare to. However, the paper provides reasonable baselines by modifying existing networks for this task\n* Avoids the need for an intermediate representation (i.e. point cloud) by directly mapping features from the disparity volume into birds-eye-view coordinates\n* Plots in the appendix are interesting\n\nWeaknesses:\n* While the task itself is new, closely related forms of the problem have been studied. For example 3D object detection from monocular/stereo, and monocular layout estimation. It would have been helpful to see results on the closely related task of 3D object detection to better compare against prior works.\n*  The IPM module appears to be very sensitive to the accuracy of the ground plane. In the synthetic CARLA dataset, where a ground plane can be accurately computed, there seems to be a large advantage of using the IPM module. On real-world data like KITTI, the use of the IPM module gives very limited improvement in performance of the stereo-only baseline.\n* The task is closely related to 3D object detection which has been using similar components. The core components of the approach have been used in various forms in prior work. The paper (Orthographic feature transform for monocular 3d object detection, Roddick 2019) uses a very similar method to project image features into a birds-eye-view representation.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "SBEVNet Review", "review": "The paper proposes an end-to-end network for layout estimation from stereo images. The approach is built off previous stereo matching networks, which built and process a 3D disparity volume. The stereo estimate is used to project image features into a birds-eye-view representation which is processed using a U-net which predicts a semantic scene layout. The approach is evaluated on the KITTI and Carla generated datasets.\n\nStrengths:\n* This is the first work to attempt semantic layout estimation from stereo images\n* The approach is geometrically grounded, and can properly leverage stereo information to improve layout estimation\n* The approach performs well on the two datasets evaluated. Since this paper is focused on a new problem, there are not existing works to directly compare to. However, the paper provides reasonable baselines by modifying existing networks for this task\n* Avoids the need for an intermediate representation (i.e. point cloud) by directly mapping features from the disparity volume into birds-eye-view coordinates\n* Plots in the appendix are interesting\n\nWeaknesses:\n* While the task itself is new, closely related forms of the problem have been studied. For example 3D object detection from monocular/stereo, and monocular layout estimation. It would have been helpful to see results on the closely related task of 3D object detection to better compare against prior works.\n*  The IPM module appears to be very sensitive to the accuracy of the ground plane. In the synthetic CARLA dataset, where a ground plane can be accurately computed, there seems to be a large advantage of using the IPM module. On real-world data like KITTI, the use of the IPM module gives very limited improvement in performance of the stereo-only baseline.\n* The task is closely related to 3D object detection which has been using similar components. The core components of the approach have been used in various forms in prior work. The paper (Orthographic feature transform for monocular 3d object detection, Roddick 2019) uses a very similar method to project image features into a birds-eye-view representation.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603852193247}], "openreview_url": "https://openreview.net/forum?id=IU8QxEiG4hR", "arxiv_id": "2105.11705", "paper_pdf": "papers/IU8QxEiG4hR.pdf", "paper_pdf_sha256": "4253842fa38ef10ace1666b6c0e3d97a4f42036658048558156879b86191aa44", "paper_pdf_bytes": 2489012, "paper_pdf_source": "openreview", "code_url": "https://github.com/divamgupta/sbevnet-stereo-layout-estimation", "code_repository": "divamgupta/sbevnet-stereo-layout-estimation", "code_commit": "ef9377df95041399bc3a243fefdbb343effca12c", "code_archive": "repos/IU8QxEiG4hR.zip", "code_archive_sha256": "647abfef2e29f3563a69a49f81dda893dd8c1987de7b366f1b9f3072333a33d4", "code_archive_bytes": 2849843, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 2783, "github_languages": {"Python": 45584}, "github_archived": false, "github_pushed_at": "2023-01-15T17:05:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sbevnet-end-to-end-deep-stereo-layout-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Oo5t7b1jQu", "year": 2026, "status": "rejected", "title": "SituatedThinker: Grounding LLM Reasoning with Real-World through Situated Thinking", "authors": ["Junnan Liu", "Linhao Luo", "Thuy-Trang Vu", "Gholamreza Haffari"], "authorids": ["~Junnan_Liu1", "~Linhao_Luo2", "~Thuy-Trang_Vu1", "~Gholamreza_Haffari2"], "authors_source": "OpenReview API", "abstract": "Recent advances in large language models (LLMs) demonstrate their impressive reasoning capabilities. However, the reasoning confined to internal parametric space limits LLMs' access to real-time information and understanding of the physical world. To overcome this constraint, we introduce SituatedThinker, a novel framework that enables LLMs to ground their reasoning in real-world contexts through situated thinking, which adaptively combines both internal knowledge and external information with predefined interfaces. By utilizing reinforcement learning, SituatedThinker incentivizes deliberate reasoning with the real world to acquire information and feedback, allowing LLMs to surpass their knowledge boundaries and enhance reasoning. Experimental results demonstrate significant performance improvements on multi-hop question-answering and mathematical reasoning benchmarks. Furthermore, SituatedThinker demonstrates strong performance on unseen tasks, such as KBQA, TableQA, and text-based games, showcasing the generalizable real-world grounded reasoning capability.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zTRPtR7ww5", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13039/Reviewer_v1GG"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes a framework, SituatedThinker, which aims to ground the reasoning of LLMs in the real world. The core of this method is to enable LLMs to learn to adaptively combine their internal knowledge with external information obtained through predefined \"Interfaces\". The authors use RL to train the LLM and achieve a good performance in multiple tasks.", "review_text": "This paper proposes a framework, SituatedThinker, which aims to ground the reasoning of LLMs in the real world. The core of this method is to enable LLMs to learn to adaptively combine their internal knowledge with external information obtained through predefined \"Interfaces\". The authors use RL to train the LLM and achieve a good performance in multiple tasks.", "strengths": "1. The problem of grounding the LLM in real-world situations is important to explore.\n\n2. On multi-hop question-answering and mathematical reasoning, the model significantly outperforms baseline methods.", "weaknesses": "1. I am still unclear about the distinction between \"Situated Thinking\" and standard tool learning. The \"Interfaces\" defined in the paper, which include the description, the specified input format, and outputs, do not seem fundamentally different from the LLM agent function calling used in the community today. The only difference is the representation, such as text-based tags versus JSON Schema, which makes the contribution incremental.\n\n2. One of the contributions is that the paper claims that the model can generalize to unseen interfaces. However, this does not seem particularly compelling. Most current instruction-tuned models already possess a \"zero-shot tool use\" capability: as long as a clear description and usage format for a new tool (e.g., the format for OpenAI's function calling) is provided in the message, the model can follow these instructions to use it. This may weaken the contribution as well.", "questions": "1. Would performing SFT first to equip the model with better format-following capabilities make the RL training smoother? Why was this setting not considered?\n\n2. There is a spelling error in the Interface Template in Section 2.1.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a framework, SituatedThinker, which aims to ground the reasoning of LLMs in the real world. The core of this method is to enable LLMs to learn to adaptively combine their internal knowledge with external information obtained through predefined \"Interfaces\". The authors use RL to train the LLM and achieve a good performance in multiple tasks.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "1. The problem of grounding the LLM in real-world situations is important to explore.\n\n2. On multi-hop question-answering and mathematical reasoning, the model significantly outperforms baseline methods.", "weaknesses": "1. I am still unclear about the distinction between \"Situated Thinking\" and standard tool learning. The \"Interfaces\" defined in the paper, which include the description, the specified input format, and outputs, do not seem fundamentally different from the LLM agent function calling used in the community today. The only difference is the representation, such as text-based tags versus JSON Schema, which makes the contribution incremental.\n\n2. One of the contributions is that the paper claims that the model can generalize to unseen interfaces. However, this does not seem particularly compelling. Most current instruction-tuned models already possess a \"zero-shot tool use\" capability: as long as a clear description and usage format for a new tool (e.g., the format for OpenAI's function calling) is provided in the message, the model can follow these instructions to use it. This may weaken the contribution as well.", "questions": "1. Would performing SFT first to equip the model with better format-following capabilities make the RL training smoother? Why was this setting not considered?\n\n2. There is a spelling error in the Interface Template in Section 2.1.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761882204665}, {"id": "OT8zuZyqJ6", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13039/Reviewer_z4VY"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper proposes the situated thinker, an RL-trained framework that enables large language model chain-of-thought reasoning with predefined tool calls (external interfaces), including retrieval, code execution, and other APIs. The authors use GRPO-style RL training to inject tool call capabilities into smaller language models. Trained on MusiQue and Big-Math, the model reports strong gains on mathematical reasoning datasets (e.g., AIME24/25 and MATH500).", "review_text": "This paper proposes the situated thinker, an RL-trained framework that enables large language model chain-of-thought reasoning with predefined tool calls (external interfaces), including retrieval, code execution, and other APIs. The authors use GRPO-style RL training to inject tool call capabilities into smaller language models. Trained on MusiQue and Big-Math, the model reports strong gains on mathematical reasoning datasets (e.g., AIME24/25 and MATH500).", "strengths": "- This paper is well-motivated and well-written. It’s good to show how general tool-use capabilities can be injected into smaller language models via RL to boost reasoning.\n- The results on MATH and knowledge-based reasoning datasets seem strong.", "weaknesses": "- Interface design: he authors customized a new interface design for external tool calls. The specifications look highly similar to the Model Context Protocol (MCP). Why not just use MCP, since it’s more universal and enables a wider range of tool use?\n\n- It’s hard to see what actually works. From Table 8, we can see that the tool calls hardly affect model performance on one of the major claimed domains—mathematical reasoning—while on knowledge-based tasks, it’s unclear how retrieval-only methods contribute to the performance. For example, it’s not clear how state-of-the-art retrieval or search-based methods contribute—row 3 in Table 6 might simply show that the proposed situated thinker interface is not effective for retrieval-based tasks.\n\n- Novelty and broader implications: Marrying language models with tool use is not novel and is already a widely recognized direction. I wonder what the broader implications of this project are.", "questions": "For TextWorld, how much depends on interface instruction quality vs learned policy?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the situated thinker, an RL-trained framework that enables large language model chain-of-thought reasoning with predefined tool calls (external interfaces), including retrieval, code execution, and other APIs. The authors use GRPO-style RL training to inject tool call capabilities into smaller language models. Trained on MusiQue and Big-Math, the model reports strong gains on mathematical reasoning datasets (e.g., AIME24/25 and MATH500).", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- This paper is well-motivated and well-written. It’s good to show how general tool-use capabilities can be injected into smaller language models via RL to boost reasoning.\n- The results on MATH and knowledge-based reasoning datasets seem strong.", "weaknesses": "- Interface design: he authors customized a new interface design for external tool calls. The specifications look highly similar to the Model Context Protocol (MCP). Why not just use MCP, since it’s more universal and enables a wider range of tool use?\n\n- It’s hard to see what actually works. From Table 8, we can see that the tool calls hardly affect model performance on one of the major claimed domains—mathematical reasoning—while on knowledge-based tasks, it’s unclear how retrieval-only methods contribute to the performance. For example, it’s not clear how state-of-the-art retrieval or search-based methods contribute—row 3 in Table 6 might simply show that the proposed situated thinker interface is not effective for retrieval-based tasks.\n\n- Novelty and broader implications: Marrying language models with tool use is not novel and is already a widely recognized direction. I wonder what the broader implications of this project are.", "questions": "For TextWorld, how much depends on interface instruction quality vs learned policy?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761870240326}, {"id": "smClcKrGNk", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13039/Reviewer_c491"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes “Situated Thinker”, which grounds the reasoning process of long chain-of-thought (CoT) LLMs with results returned from external tools. The LLM system is allowed to invoke tools through “interfaces” while generating a long CoT, with results dynamically inserted into the CoT. Two models of 8B and 32B size are trained and evaluated on multi-hop QA, math, and other out-of-distribution reasoning benchmarks", "review_text": "This paper proposes “Situated Thinker”, which grounds the reasoning process of long chain-of-thought (CoT) LLMs with results returned from external tools. The LLM system is allowed to invoke tools through “interfaces” while generating a long CoT, with results dynamically inserted into the CoT. Two models of 8B and 32B size are trained and evaluated on multi-hop QA, math, and other out-of-distribution reasoning benchmarks", "strengths": "- The single-turn formulation of situated thinking paradigm is, to my knowledge, novel. This formulation should greatly simplify the RL training pipeline as compared to more complex multi-turn tool-use enhanced reasoning models.\n- The resulting models seem to generalize to unseen domains not explicitly covered in the training data.\n- The paper is generally well-written and easy to understand.", "weaknesses": "- I find some of the statements about novelty over-claimed:\n  - It’s unclear how the claim that situated thinking, which is described as “a new paradigm that allows LLMs to adaptively engage with external environments” in line 91, is different from several LLM-based agents that use retrieval tools, like web-search, e.g., [1,2,3,4], or even baselines ReSearch and Search-R1 (beyond number of tools used). \n  - The exact way tool results are incorporated (directly within a single-turn long CoT) may be only differentiator, but I don’t believe this constitutes a fundamentally new paradigm. Rather, as described in Strengths, I consider this a formulation that has some benefits, e.g., simplified training pipelines\n- Some issues with baselines:\n  - Situated Thinker is trained from Qwen3 family, but the most direct baselines, Search-R1 and ReSearch are trained from Qwen2.5-7B. Given that Qwen3 are generally stronger models, it’s unclear how much the performance difference stems from base model difference vs. methodological improvements. A more reasonable comparison would be to conduct RL training from the same base model if a methodological comparison is the aim, then demonstrating that such approach can improve the performance of better starting models, as is done with Qwen3 experiments.\n  - Why do you compare against Qwen3 base LLMs on math benchmarks, and not the post-trained models that are capable of long CoT? Same question for ReAct baselines for OOD tasks?In particular, most current implementations of ReAct agents use instruction-tuned models capable of better instruction following\n\n[1] https://arxiv.org/abs/2509.06283\n\n[2] https://arxiv.org/abs/2507.02592\n\n[3] https://arxiv.org/abs/2507.15061\n\n[4] https://arxiv.org/abs/2508.13167", "questions": "- Can you provide more details about how retrieval is done for SituatedThinker? What retriever did you use? Is this retriever the same for RAG-based baselines? How does SituatedThinker's ability depend on the quality of retrievals?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes “Situated Thinker”, which grounds the reasoning process of long chain-of-thought (CoT) LLMs with results returned from external tools. The LLM system is allowed to invoke tools through “interfaces” while generating a long CoT, with results dynamically inserted into the CoT. Two models of 8B and 32B size are trained and evaluated on multi-hop QA, math, and other out-of-distribution reasoning benchmarks", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The single-turn formulation of situated thinking paradigm is, to my knowledge, novel. This formulation should greatly simplify the RL training pipeline as compared to more complex multi-turn tool-use enhanced reasoning models.\n- The resulting models seem to generalize to unseen domains not explicitly covered in the training data.\n- The paper is generally well-written and easy to understand.", "weaknesses": "- I find some of the statements about novelty over-claimed:\n  - It’s unclear how the claim that situated thinking, which is described as “a new paradigm that allows LLMs to adaptively engage with external environments” in line 91, is different from several LLM-based agents that use retrieval tools, like web-search, e.g., [1,2,3,4], or even baselines ReSearch and Search-R1 (beyond number of tools used). \n  - The exact way tool results are incorporated (directly within a single-turn long CoT) may be only differentiator, but I don’t believe this constitutes a fundamentally new paradigm. Rather, as described in Strengths, I consider this a formulation that has some benefits, e.g., simplified training pipelines\n- Some issues with baselines:\n  - Situated Thinker is trained from Qwen3 family, but the most direct baselines, Search-R1 and ReSearch are trained from Qwen2.5-7B. Given that Qwen3 are generally stronger models, it’s unclear how much the performance difference stems from base model difference vs. methodological improvements. A more reasonable comparison would be to conduct RL training from the same base model if a methodological comparison is the aim, then demonstrating that such approach can improve the performance of better starting models, as is done with Qwen3 experiments.\n  - Why do you compare against Qwen3 base LLMs on math benchmarks, and not the post-trained models that are capable of long CoT? Same question for ReAct baselines for OOD tasks?In particular, most current implementations of ReAct agents use instruction-tuned models capable of better instruction following\n\n[1] https://arxiv.org/abs/2509.06283\n\n[2] https://arxiv.org/abs/2507.02592\n\n[3] https://arxiv.org/abs/2507.15061\n\n[4] https://arxiv.org/abs/2508.13167", "questions": "- Can you provide more details about how retrieval is done for SituatedThinker? What retriever did you use? Is this retriever the same for RAG-based baselines? How does SituatedThinker's ability depend on the quality of retrievals?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761625550438}, {"id": "Q1rjy6pqau", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission13039/Reviewer_pcmm"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces SituatedThinner, a novel framework that grounds LLM reasoning in real-world contexts through a paradigm termed \"situated thinking.\" The core idea is to enable LLMs to adaptively interleave internal reasoning steps (\"internal actions\") with calls to external interfaces (\"situated actions\") during a single, deliberate reasoning trace. These interfaces provide a unified way to interact with diverse external environments like knowledge bases, code executors, and simulated worlds. The model is trained using a reinforcement learning objective (an adaptation of GRPO) with a simple reward based on final answer correctness, which incentivizes the LLM to autonomously learn when and how to invoke interfaces to acquire information, verify reasoning, and correct errors. Extensive experiments demonstrate significant improvements over strong baselines on multi-hop QA and mathematical reasoning, and, more importantly, show compelling generalization to unseen tasks (KBQA, TableQA) and interfaces (text-based games).", "review_text": "This paper introduces SituatedThinner, a novel framework that grounds LLM reasoning in real-world contexts through a paradigm termed \"situated thinking.\" The core idea is to enable LLMs to adaptively interleave internal reasoning steps (\"internal actions\") with calls to external interfaces (\"situated actions\") during a single, deliberate reasoning trace. These interfaces provide a unified way to interact with diverse external environments like knowledge bases, code executors, and simulated worlds. The model is trained using a reinforcement learning objective (an adaptation of GRPO) with a simple reward based on final answer correctness, which incentivizes the LLM to autonomously learn when and how to invoke interfaces to acquire information, verify reasoning, and correct errors. Extensive experiments demonstrate significant improvements over strong baselines on multi-hop QA and mathematical reasoning, and, more importantly, show compelling generalization to unseen tasks (KBQA, TableQA) and interfaces (text-based games).", "strengths": "**Novel and Well-Motivated Paradigm**: The concept of \"situated thinking\" is a meaningful and timely contribution. It moves beyond simply injecting external knowledge (as in RAG) towards enabling a dynamic, interactive reasoning process where the model actively queries its environment. This addresses a key limitation of closed-world, parametric reasoning in modern LLMs.\n\n**Impressive Generalization**: The framework's ability to generalize zero-shot to entirely new domains (MedQA, GPQA), tasks (WebQSP, WTQ), and interfaces (TextWorld) is a major strength. This suggests the model learns a generalizable skill of \"tool-use\" rather than overfitting to specific tasks seen during training.\n\n**Effective and Simple Training Approach**: The use of RL with only a final-answer reward to successfully teach complex behaviors like interface invocation, reflection, and planning is elegant and effective. The ablation studies convincingly show that the model learns these capabilities without explicit supervision on intermediate steps.\n\n**Rigorous Experimental Setup:** The paper provides comprehensive evaluations across a wide range of benchmarks, comparing against a strong set of baselines (including multi-step RAG and other RL-for-reasoning models). The inclusion of training dynamics analysis and qualitative case studies adds depth to the empirical validation.", "weaknesses": "**Limited Discussion on Relation to RAG**: While the paper positions itself against \"tool-calling\" and \"search-enhanced\" models, a more direct and thorough discussion comparing and contrasting the proposed paradigm with classic and multi-step Retrieval-Augmented Generation (RAG) is lacking. A dedicated subsection explicitly analyzing the fundamental differences (e.g., static knowledge injection vs. dynamic, agentic interaction) would significantly sharpen the contribution and make it more accessible to a broader audience.\n\n**Limited Scope of Evaluation**: The current evaluation is confined to the textual modality and the English language. The limitations section rightly notes this, but it remains a weakness of the current empirical evidence. The framework's performance in multimodal settings or with non-English interfaces remains an open and important question.\n\n**Focus on Deterministic Tasks**: The work primarily addresses tasks with definitive, verifiable answers. Its applicability to more open-ended, creative, or non-deterministic reasoning tasks (e.g., essay writing, complex strategic planning) is not explored and represents a significant boundary for the current method.\n\n**Lack of Assessment on General Capabilities**: The idea of employing a single, powerful LLM for situated reasoning, as opposed to constructing a multi-agent framework, is highly appealing, and I appreciate this elegant approach. However, when training one model to handle such a complex and diverse set of reasoning tasks, its generalization and the preservation of core capabilities become paramount. In this context, the paper currently lacks an assessment of whether the RL fine-tuning causes catastrophic forgetting of the model's general capabilities. It remains unclear if performance on standard benchmarks (e.g., MMLU, BBH) or its general instruction-following ability has degraded after the specialized training for situated thinking. An ablation study comparing the base model and the trained SituatedThinner on such general tasks is crucial to ensure that the acquisition of these novel, specialized skills does not come at the cost of the model's broader competency and linguistic fluency.\n\nIt's better to discuss the related works of the agent frameworks for reasoning.", "questions": "1) The decision to omit the KL penalty in the GRPO objective is a significant one. To better understand its impact, did you conduct an ablation study on this choice? Specifically, how critical was omitting the KL penalty for achieving the final performance, and what was its effect on the model's behavior and stability during training?\n\n2) Given that the model successfully generalizes to new interfaces, did you observe any cases of \"interface confusion\" where the model used the wrong interface (e.g., using code execution for a simple fact lookup)? A case study and analysis of such failure modes could be insightful.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces SituatedThinner, a novel framework that grounds LLM reasoning in real-world contexts through a paradigm termed \"situated thinking.\" The core idea is to enable LLMs to adaptively interleave internal reasoning steps (\"internal actions\") with calls to external interfaces (\"situated actions\") during a single, deliberate reasoning trace. These interfaces provide a unified way to interact with diverse external environments like knowledge bases, code executors, and simulated worlds. The model is trained using a reinforcement learning objective (an adaptation of GRPO) with a simple reward based on final answer correctness, which incentivizes the LLM to autonomously learn when and how to invoke interfaces to acquire information, verify reasoning, and correct errors. Extensive experiments demonstrate significant improvements over strong baselines on multi-hop QA and mathematical reasoning, and, more importantly, show compelling generalization to unseen tasks (KBQA, TableQA) and interfaces (text-based games).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**Novel and Well-Motivated Paradigm**: The concept of \"situated thinking\" is a meaningful and timely contribution. It moves beyond simply injecting external knowledge (as in RAG) towards enabling a dynamic, interactive reasoning process where the model actively queries its environment. This addresses a key limitation of closed-world, parametric reasoning in modern LLMs.\n\n**Impressive Generalization**: The framework's ability to generalize zero-shot to entirely new domains (MedQA, GPQA), tasks (WebQSP, WTQ), and interfaces (TextWorld) is a major strength. This suggests the model learns a generalizable skill of \"tool-use\" rather than overfitting to specific tasks seen during training.\n\n**Effective and Simple Training Approach**: The use of RL with only a final-answer reward to successfully teach complex behaviors like interface invocation, reflection, and planning is elegant and effective. The ablation studies convincingly show that the model learns these capabilities without explicit supervision on intermediate steps.\n\n**Rigorous Experimental Setup:** The paper provides comprehensive evaluations across a wide range of benchmarks, comparing against a strong set of baselines (including multi-step RAG and other RL-for-reasoning models). The inclusion of training dynamics analysis and qualitative case studies adds depth to the empirical validation.", "weaknesses": "**Limited Discussion on Relation to RAG**: While the paper positions itself against \"tool-calling\" and \"search-enhanced\" models, a more direct and thorough discussion comparing and contrasting the proposed paradigm with classic and multi-step Retrieval-Augmented Generation (RAG) is lacking. A dedicated subsection explicitly analyzing the fundamental differences (e.g., static knowledge injection vs. dynamic, agentic interaction) would significantly sharpen the contribution and make it more accessible to a broader audience.\n\n**Limited Scope of Evaluation**: The current evaluation is confined to the textual modality and the English language. The limitations section rightly notes this, but it remains a weakness of the current empirical evidence. The framework's performance in multimodal settings or with non-English interfaces remains an open and important question.\n\n**Focus on Deterministic Tasks**: The work primarily addresses tasks with definitive, verifiable answers. Its applicability to more open-ended, creative, or non-deterministic reasoning tasks (e.g., essay writing, complex strategic planning) is not explored and represents a significant boundary for the current method.\n\n**Lack of Assessment on General Capabilities**: The idea of employing a single, powerful LLM for situated reasoning, as opposed to constructing a multi-agent framework, is highly appealing, and I appreciate this elegant approach. However, when training one model to handle such a complex and diverse set of reasoning tasks, its generalization and the preservation of core capabilities become paramount. In this context, the paper currently lacks an assessment of whether the RL fine-tuning causes catastrophic forgetting of the model's general capabilities. It remains unclear if performance on standard benchmarks (e.g., MMLU, BBH) or its general instruction-following ability has degraded after the specialized training for situated thinking. An ablation study comparing the base model and the trained SituatedThinner on such general tasks is crucial to ensure that the acquisition of these novel, specialized skills does not come at the cost of the model's broader competency and linguistic fluency.\n\nIt's better to discuss the related works of the agent frameworks for reasoning.", "questions": "1) The decision to omit the KL penalty in the GRPO objective is a significant one. To better understand its impact, did you conduct an ablation study on this choice? Specifically, how critical was omitting the KL penalty for achieving the final performance, and what was its effect on the model's behavior and stability during training?\n\n2) Given that the model successfully generalizes to new interfaces, did you observe any cases of \"interface confusion\" where the model used the wrong interface (e.g., using code execution for a simple fact lookup)? A case study and analysis of such failure modes could be insightful.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761571496997}], "openreview_url": "https://openreview.net/forum?id=Oo5t7b1jQu", "arxiv_id": "2505.19300", "paper_pdf": "papers/Oo5t7b1jQu.pdf", "paper_pdf_sha256": "c063c7b1bca364a3725ed5cc1c94c8c8a3f478686346260922b5c4a1ea0dbb09", "paper_pdf_bytes": 765445, "paper_pdf_source": "openreview", "code_url": "https://github.com/jnanliu/SituatedThinker", "code_repository": "jnanliu/SituatedThinker", "code_commit": "6e44597f4a9e8725206106e2aebfb0cefe31c596", "code_archive": "repos/Oo5t7b1jQu.zip", "code_archive_sha256": "1eca409b74dd4f9c216f72ba7715c63ab749921c48ce1fa4279adc19d143f3b6", "code_archive_bytes": 10721296, "code_file_count": 221, "code_extensions": {".py": 215, ".sh": 6}, "github_disk_usage_kb": 9318, "github_languages": {"Python": 1635369, "Shell": 5281}, "github_archived": false, "github_pushed_at": "2025-06-08T18:17:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/situatedthinker-grounding-llm-reasoning-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "4RRmy9iw3c", "year": 2025, "status": "rejected", "title": "AutoAL: Automated Active Learning with Differentiable Query Strategy Search", "authors": ["Yifeng Wang", "Xueying Zhan", "Siyu Huang"], "authorids": ["~Yifeng_Wang2", "~Xueying_Zhan1", "~Siyu_Huang2"], "authors_source": "OpenReview API", "abstract": "As deep learning continues to evolve, the need for data efficiency becomes increasingly important. Considering labeling large datasets is both time-consuming and expensive, active learning (AL) provides a promising solution to this challenge by iteratively selecting the most informative subsets of examples to train deep neural networks, thereby reducing the labeling cost. However, the effectiveness of different AL algorithms can vary significantly across data scenarios, and determining which AL algorithm best fits a given task remains a challenging problem. This work presents the first differentiable AL strategy search method, named AutoAL, which is designed on top of existing AL sampling strategies. AutoAL consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. For any given task, SearchNet and FitNet are iteratively co-optimized using the labeled data, learning how well a set of candidate AL algorithms perform on that task. With the optimal AL strategies identified, SearchNet selects a small subset from the unlabeled pool for querying their annotations, enabling efficient training of the task model. Experimental results demonstrate that AutoAL consistently achieves superior accuracy compared to all candidate AL algorithms and other selective AL approaches, showcasing its potential for adapting and integrating multiple existing AL methods across diverse tasks and domains.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "JWVSQhWIPs", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5471/Reviewer_fWbJ"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This work introduces AutoAL, a differentiable active learning (AL) strategy search method that builds on existing AL sampling strategies. AutoAL contains two neural networks, SearchNet and FitNet, which are co-optimized through a differentiable bi-level optimization framework to identify optimal AL strategies for different tasks. Experimental results show that AutoAL outperforms individual AL algorithms and other selective approaches.", "review_text": "This work introduces AutoAL, a differentiable active learning (AL) strategy search method that builds on existing AL sampling strategies. AutoAL contains two neural networks, SearchNet and FitNet, which are co-optimized through a differentiable bi-level optimization framework to identify optimal AL strategies for different tasks. Experimental results show that AutoAL outperforms individual AL algorithms and other selective approaches.", "strengths": "- The work handles an important ML problem; Active Learning (AL) with differentiable strategy search.\n- The proposed method is technically sound\n- Writing is clear and easy-to-follow", "weaknesses": "- Hybrid AL methods that combine uncertainty and diversity have been demonstrated to perform effectively in a variety of situations. It would be beneficial to include examples where the proposed AutoAL approach is particularly necessary or advantageous for specific applications.\n- The bi-level optimization within AutoAL relies on labeled data. How does the algorithm perform if the labeled data is skewed or imbalanced? For instance, if the initial labeled set suffers from class imbalance, might this severely impair the algorithm? The assumption of a randomly selected initial set, as used in the current experiments, appears to be less practical.\n- Similarly, is there a guarantee that the AutoAL approach, trained with labeled data from the current AL round, will identify the most informative samples from the unlabeled pool in the subsequent AL round? A more detailed analysis of the algorithm's guarantees is necessary.\n- The approach of training an additional network for sample selection shows similarities to [1] employing meta-learning with an additional network for querying.\n- Can the proposed method be applied to the open-set AL problem [1]?\n- The datasets used in the experiments are of small scale. It is imperative to validate the performance on large-scale datasets, such as ImageNet.\n\n---\n[1] Meta-Query-Net: Resolving Purity-Informativeness Dilemma in Open-set Active Learning, NeurIPS, 2022", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work introduces AutoAL, a differentiable active learning (AL) strategy search method that builds on existing AL sampling strategies. AutoAL contains two neural networks, SearchNet and FitNet, which are co-optimized through a differentiable bi-level optimization framework to identify optimal AL strategies for different tasks. Experimental results show that AutoAL outperforms individual AL algorithms and other selective approaches.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The work handles an important ML problem; Active Learning (AL) with differentiable strategy search.\n- The proposed method is technically sound\n- Writing is clear and easy-to-follow", "weaknesses": "- Hybrid AL methods that combine uncertainty and diversity have been demonstrated to perform effectively in a variety of situations. It would be beneficial to include examples where the proposed AutoAL approach is particularly necessary or advantageous for specific applications.\n- The bi-level optimization within AutoAL relies on labeled data. How does the algorithm perform if the labeled data is skewed or imbalanced? For instance, if the initial labeled set suffers from class imbalance, might this severely impair the algorithm? The assumption of a randomly selected initial set, as used in the current experiments, appears to be less practical.\n- Similarly, is there a guarantee that the AutoAL approach, trained with labeled data from the current AL round, will identify the most informative samples from the unlabeled pool in the subsequent AL round? A more detailed analysis of the algorithm's guarantees is necessary.\n- The approach of training an additional network for sample selection shows similarities to [1] employing meta-learning with an additional network for querying.\n- Can the proposed method be applied to the open-set AL problem [1]?\n- The datasets used in the experiments are of small scale. It is imperative to validate the performance on large-scale datasets, such as ImageNet.\n\n---\n[1] Meta-Query-Net: Resolving Purity-Informativeness Dilemma in Open-set Active Learning, NeurIPS, 2022", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1731303930545}, {"id": "bTyyWiTNLt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5471/Reviewer_Rqgr"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper presents AutoAL, a framework for automated active learning that optimizes query strategy selection using differentiable methods. Traditional active learning approaches often rely on predefined strategies like uncertainty sampling or diversity sampling, which may not perform optimally across different datasets or tasks. AutoAL addresses this limitation by integrating existing active learning strategies into a unified framework. It employs two neural networks, SearchNet and FitNet, within a bi-level optimization structure to automate the selection process. By relaxing the discrete search space of active learning strategies into a continuous domain, AutoAL enables gradient-based optimization, enhancing computational efficiency and adaptability. Experimental results demonstrate that AutoAL consistently outperforms individual strategies and other selective methods across various natural and medical image datasets, highlighting its effectiveness and versatility.", "review_text": "The paper presents AutoAL, a framework for automated active learning that optimizes query strategy selection using differentiable methods. Traditional active learning approaches often rely on predefined strategies like uncertainty sampling or diversity sampling, which may not perform optimally across different datasets or tasks. AutoAL addresses this limitation by integrating existing active learning strategies into a unified framework. It employs two neural networks, SearchNet and FitNet, within a bi-level optimization structure to automate the selection process. By relaxing the discrete search space of active learning strategies into a continuous domain, AutoAL enables gradient-based optimization, enhancing computational efficiency and adaptability. Experimental results demonstrate that AutoAL consistently outperforms individual strategies and other selective methods across various natural and medical image datasets, highlighting its effectiveness and versatility.", "strengths": "- A new approach that automates active learning strategy selection through differentiable optimization, surpassing manual and non-differentiable methods.\n- Effective integration of strategy selection and data modeling via the bi-level optimization of SearchNet and FitNet.\n- Flexibility and adaptability, allowing incorporation of multiple existing strategies and tailoring to specific tasks and data distributions.", "weaknesses": "- Increased complexity and computational overhead due to the additional neural networks and bi-level optimization, potentially challenging scalability on large datasets.\n- Dependence on a predefined pool of candidate strategies, which may limit performance if optimal strategies are not included.\n- Lack of in-depth theoretical analysis explaining the method's effectiveness and the conditions under which it performs best, possibly affecting generalizability.", "questions": "- How does AutoAL perform in terms of computational efficiency compared to traditional methods on large-scale datasets?\n- What mechanisms ensure robustness against convergence issues and local minima in the bi-level optimization?\n- Can AutoAL be extended to generate new active learning strategies dynamically rather than relying solely on a predefined candidate pool?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents AutoAL, a framework for automated active learning that optimizes query strategy selection using differentiable methods. Traditional active learning approaches often rely on predefined strategies like uncertainty sampling or diversity sampling, which may not perform optimally across different datasets or tasks. AutoAL addresses this limitation by integrating existing active learning strategies into a unified framework. It employs two neural networks, SearchNet and FitNet, within a bi-level optimization structure to automate the selection process. By relaxing the discrete search space of active learning strategies into a continuous domain, AutoAL enables gradient-based optimization, enhancing computational efficiency and adaptability. Experimental results demonstrate that AutoAL consistently outperforms individual strategies and other selective methods across various natural and medical image datasets, highlighting its effectiveness and versatility.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- A new approach that automates active learning strategy selection through differentiable optimization, surpassing manual and non-differentiable methods.\n- Effective integration of strategy selection and data modeling via the bi-level optimization of SearchNet and FitNet.\n- Flexibility and adaptability, allowing incorporation of multiple existing strategies and tailoring to specific tasks and data distributions.", "weaknesses": "- Increased complexity and computational overhead due to the additional neural networks and bi-level optimization, potentially challenging scalability on large datasets.\n- Dependence on a predefined pool of candidate strategies, which may limit performance if optimal strategies are not included.\n- Lack of in-depth theoretical analysis explaining the method's effectiveness and the conditions under which it performs best, possibly affecting generalizability.", "questions": "- How does AutoAL perform in terms of computational efficiency compared to traditional methods on large-scale datasets?\n- What mechanisms ensure robustness against convergence issues and local minima in the bi-level optimization?\n- Can AutoAL be extended to generate new active learning strategies dynamically rather than relying solely on a predefined candidate pool?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA", "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730660827193}, {"id": "99X0FjEpas", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5471/Reviewer_164p"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 1, "summary": "This paper proposes an active strategy query algorithm where the optimal query strategy is selected by a bi-level optimization network. In particular, the authors aggregates the query strategies by a scoring function implemented as a Gaussian Mixture Model. Then, the authors split out a validation set from the labeled samples to guide scoring function calculation. Experimental results show that the proposed method supasses the baselines.", "review_text": "This paper proposes an active strategy query algorithm where the optimal query strategy is selected by a bi-level optimization network. In particular, the authors aggregates the query strategies by a scoring function implemented as a Gaussian Mixture Model. Then, the authors split out a validation set from the labeled samples to guide scoring function calculation. Experimental results show that the proposed method supasses the baselines.", "strengths": "1. The studied strategy selection problem for active learning is important.\n2. The bi-level optimization strategy is rational.", "weaknesses": "1. There is still room for improvement in the paper writing.\n\n   1.1. It's unnecessary to name the two networks \"fitnet\" and \"searchnet,\" as it seems intended to make people think this is a significant innovation. However, in meta-learning, this kind of separated network design and bi-level optimization paradigm is very common.\n\n   2.1. The notations are somewhat confusing. For example, the authors didn't clearly define the output of the search net in Section 3.2, making it hard to understand Sec 3.2. It wasn’t until I finished reading the method section that I realized the output is actually a sample-wise score, forming an aggregation of scores for different queries.\n\n2. The novelty of this paper is relatively limited. The proposed meta-learning/bi-level optimization has been applied to AL [1,2]. Also, I think the algorithm design is too complicated.\n\n3. The motivation for modeling the scores by GMM distributions is unclear. Why is the score function of each strategy distributed as a Gaussian Distribution? Why is the final score function a linear weighted aggregation of different strategies? The authors should provide a concrete application or example.\n\n4. The comparison methods are too outdated, with the latest ones being LPL and BADGE from 2019. Additionally, the datasets are quite limited; despite the complexity of the method design, only CIFAR and MNIST datasets were used. Validation should be conducted on the ImageNet dataset (at least Image100). Otherwise, given that the algorithm design is much more complex than the baselines, its effectiveness cannot be convincingly demonstrated.\n\n[1] Kunkun Pang, Mingzhi Dong, Yang Wu, and Timothy Hospedales. 2018. Meta-learning transferable active learning policies by deep reinforcement learning. International Workshop on Automatic Machine Learning (ICML AutoML 2018).\n\n[2] https://grlplus.github.io/papers/96.pdf", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an active strategy query algorithm where the optimal query strategy is selected by a bi-level optimization network. In particular, the authors aggregates the query strategies by a scoring function implemented as a Gaussian Mixture Model. Then, the authors split out a validation set from the labeled samples to guide scoring function calculation. Experimental results show that the proposed method supasses the baselines.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The studied strategy selection problem for active learning is important.\n2. The bi-level optimization strategy is rational.", "weaknesses": "1. There is still room for improvement in the paper writing.\n\n   1.1. It's unnecessary to name the two networks \"fitnet\" and \"searchnet,\" as it seems intended to make people think this is a significant innovation. However, in meta-learning, this kind of separated network design and bi-level optimization paradigm is very common.\n\n   2.1. The notations are somewhat confusing. For example, the authors didn't clearly define the output of the search net in Section 3.2, making it hard to understand Sec 3.2. It wasn’t until I finished reading the method section that I realized the output is actually a sample-wise score, forming an aggregation of scores for different queries.\n\n2. The novelty of this paper is relatively limited. The proposed meta-learning/bi-level optimization has been applied to AL [1,2]. Also, I think the algorithm design is too complicated.\n\n3. The motivation for modeling the scores by GMM distributions is unclear. Why is the score function of each strategy distributed as a Gaussian Distribution? Why is the final score function a linear weighted aggregation of different strategies? The authors should provide a concrete application or example.\n\n4. The comparison methods are too outdated, with the latest ones being LPL and BADGE from 2019. Additionally, the datasets are quite limited; despite the complexity of the method design, only CIFAR and MNIST datasets were used. Validation should be conducted on the ImageNet dataset (at least Image100). Otherwise, given that the algorithm design is much more complex than the baselines, its effectiveness cannot be convincingly demonstrated.\n\n[1] Kunkun Pang, Mingzhi Dong, Yang Wu, and Timothy Hospedales. 2018. Meta-learning transferable active learning policies by deep reinforcement learning. International Workshop on Automatic Machine Learning (ICML AutoML 2018).\n\n[2] https://grlplus.github.io/papers/96.pdf", "questions": "see above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 1, "code_of_conduct": "Yes"}, "tcdate": 1730549923650}, {"id": "crureG4Tdu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5471/Reviewer_U39s"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper attempts to tackle the \"generalization problem\" of active learning (AL) algorithms across data scenarios. I believe this is a core issue in the current active learning field. This paper proposes AutoAL, a differentiable AL strategy search method to select the most effective AL sampling strategies in each iteration. It consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. The experiments on multiple datasets validate the effectiveness of the proposed approach.", "review_text": "This paper attempts to tackle the \"generalization problem\" of active learning (AL) algorithms across data scenarios. I believe this is a core issue in the current active learning field. This paper proposes AutoAL, a differentiable AL strategy search method to select the most effective AL sampling strategies in each iteration. It consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. The experiments on multiple datasets validate the effectiveness of the proposed approach.", "strengths": "1. The problem studied in this paper is valuable. This paper presents the first differentiable AL strategy search method. \n2. The proposed AutoAL approach is interesting and easily followed.\n3. The paper is well organized. \n4. The experiments validate the effectiveness of the proposed approach, and the ablation study in Figure 5 is insightful.", "weaknesses": "1. My only concern is the efficiency of the AutoAL algorithm. Although more efficient solutions have been proposed to solve second-order optimization problems, I cannot find any relevant experiments to verify them.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper attempts to tackle the \"generalization problem\" of active learning (AL) algorithms across data scenarios. I believe this is a core issue in the current active learning field. This paper proposes AutoAL, a differentiable AL strategy search method to select the most effective AL sampling strategies in each iteration. It consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. The experiments on multiple datasets validate the effectiveness of the proposed approach.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The problem studied in this paper is valuable. This paper presents the first differentiable AL strategy search method. \n2. The proposed AutoAL approach is interesting and easily followed.\n3. The paper is well organized. \n4. The experiments validate the effectiveness of the proposed approach, and the ablation study in Figure 5 is insightful.", "weaknesses": "1. My only concern is the efficiency of the AutoAL algorithm. Although more efficient solutions have been proposed to solve second-order optimization problems, I cannot find any relevant experiments to verify them.", "questions": "See the Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "1. My only concern is the efficiency of the AutoAL algorithm. Although more efficient solutions have been proposed to solve second-order optimization problems, I cannot find any relevant experiments to verify them.", "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730309533705}], "openreview_url": "https://openreview.net/forum?id=4RRmy9iw3c", "arxiv_id": "2410.13853", "paper_pdf": "papers/4RRmy9iw3c.pdf", "paper_pdf_sha256": "5e3b9a1dff3f9910694a4ea5f8407ab6a5102e18105bd949830937cdc32684b6", "paper_pdf_bytes": 540568, "paper_pdf_source": "openreview", "code_url": "https://github.com/haizailache999/AutoAL", "code_repository": "haizailache999/AutoAL", "code_commit": "5200ce7b2780467e292001c3583b59e50fcfe6bc", "code_archive": "repos/4RRmy9iw3c.zip", "code_archive_sha256": "80f54138388674310a5525fd94bba6e782d044bb727c78eb1ce0f29ec3ac6dd9", "code_archive_bytes": 386662, "code_file_count": 50, "code_extensions": {".py": 50}, "github_disk_usage_kb": 396, "github_languages": {"Python": 189805}, "github_archived": false, "github_pushed_at": "2025-05-18T23:50:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/autoal-automated-active-learning-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ICSvW69W5K", "year": 2024, "status": "rejected", "title": "Semantic Parsing with Candidate Expressions for Knowledge Base Question Answering", "authors": ["Daehwan Nam", "Gary Lee"], "authorids": ["~Daehwan_Nam1", "~Gary_Lee1"], "authors_source": "OpenReview API", "abstract": "Semantic parsers convert natural language to logical forms, which can then be evaluated on knowledge bases (KBs) to produce denotations.\nEarly neural semantic parsers used grammars that define actions, such as production rules, then the semantic parsers could sequentially take actions to construct well-typed logical forms.\nIn contrast, recent neural semantic parsers have been developed with pre-trained sequence-to-sequence (seq2seq) models, such as BART and T5, which treat logical forms as sequences of tokens.\nHowever, the seq2seq models have difficulty in learning to generate logical forms that contain components drawn from large KBs.\nIn this work, we propose a grammar augmented with candidate expressions for seq2seq semantic parsing on large KBs.\nThe grammar defines actions as production rules, and our semantic parser predicts actions during inference under the constraints by types and candidate expressions.\nWe apply the grammar to knowledge base question answering, where the constraints by candidate expressions assist a semantic parser to generate valid KB components.\nExperiments on the KQA Pro benchmark showed that the constraints by candidate expressions increased the accuracy of our semantic parser, and our semantic parser achieved state-of-the-art performance on KQA Pro.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "SHu5dbPRoR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2430/Reviewer_Ntkc"], "rating": "5: marginally below the acceptance threshold", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a grammar augmented with candidate expressions for KB-QA. The grammar decoder enforces the type of candidate expressions during the decoding phrase. The proposed method achieves the state-of-the-art performance on KQAPRO datasets. The ablation study shows the effectiveness of the method. However, the paper fails to show the proposed method works for other KB-QA tasks. And it is not sure what is the scope of the problem the proposed method could outperform the beselines.", "review_text": "This paper introduces a grammar augmented with candidate expressions for KB-QA. The grammar decoder enforces the type of candidate expressions during the decoding phrase. The proposed method achieves the state-of-the-art performance on KQAPRO datasets. The ablation study shows the effectiveness of the method. However, the paper fails to show the proposed method works for other KB-QA tasks. And it is not sure what is the scope of the problem the proposed method could outperform the beselines.", "strengths": "* The method is proposed for KBQA tasks is noval.\n* A state-of-the-art performance on KQAPRO\n* The abalation study shows the effectiveness of this method", "weaknesses": "* In this paper, the performance of the model is recorded only in one KB-QA dataset. It is not known how good is this method among all KB-QA tasks.", "questions": "* Are there any results on KB-QA datasets other than KQAPRO for the proposed method? It would be better to show the performance of the proposed method on more than one dataset. The datasets could be (but not limited to) Complex Web Questions and Grail QA.\n\n* [Not critical] This paper compares their decoding method with Picard's paper to show the proposed method has a good latency. Why are the two models in different tasks (KBQA vs Spider), different model architectures (BART vs T5), and different hardwares comparable?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a grammar augmented with candidate expressions for KB-QA. The grammar decoder enforces the type of candidate expressions during the decoding phrase. The proposed method achieves the state-of-the-art performance on KQAPRO datasets. The ablation study shows the effectiveness of the method. However, the paper fails to show the proposed method works for other KB-QA tasks. And it is not sure what is the scope of the problem the proposed method could outperform the beselines.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "* The method is proposed for KBQA tasks is noval.\n* A state-of-the-art performance on KQAPRO\n* The abalation study shows the effectiveness of this method", "weaknesses": "* In this paper, the performance of the model is recorded only in one KB-QA dataset. It is not known how good is this method among all KB-QA tasks.", "questions": "* Are there any results on KB-QA datasets other than KQAPRO for the proposed method? It would be better to show the performance of the proposed method on more than one dataset. The datasets could be (but not limited to) Complex Web Questions and Grail QA.\n\n* [Not critical] This paper compares their decoding method with Picard's paper to show the proposed method has a good latency. Why are the two models in different tasks (KBQA vs Spider), different model architectures (BART vs T5), and different hardwares comparable?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698979931315}, {"id": "bNgLoAxZLh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2430/Reviewer_pMus"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a grammar-based semantic parser for knowledge-based question answering. The grammars are specialized with fine-grained types and candidate expressions. During decoding, type constraints and candidate constraints can be enforced to search for the desired programs more effectively.", "review_text": "This paper proposes a grammar-based semantic parser for knowledge-based question answering. The grammars are specialized with fine-grained types and candidate expressions. During decoding, type constraints and candidate constraints can be enforced to search for the desired programs more effectively.", "strengths": "A well-executed work on designing grammars for knowledge-based question answering. The revisit of traditional grammar-based methods provides insights on whether prior information such as types are still useful in the current era of pre-trained models.  (But I also feel the question whether grammar-based constraints are still useful for large models needs to be studied more)", "weaknesses": "The main contribution, as the author points out,  is “To the best of our knowledge, our work is the first to use production rules as actions for semantic parsers based on pre-trained seq2seq models.”. First, I’m not sure what is the underlying challenge of extending traditional grammar-based seq2seqs to their pretrained counterparts? That is, I'm not sure about the technical contribution of the paper. Second, there are already existing works in this direction (as cited in the related work). For example, RAT-SQL (based on BERT) extends TRANX which expands production rules incrementally during decoding.", "questions": "- What is generation speed of the grammar-based parser compared with the baseline BART? I’m wondering that with more constraints, the parser might suffer from much slower inference speed\n- would it be an issue if the size of the set of candidate expressions become too large? E.g., certain class (e.g., keyword-entity) has too many candidates.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a grammar-based semantic parser for knowledge-based question answering. The grammars are specialized with fine-grained types and candidate expressions. During decoding, type constraints and candidate constraints can be enforced to search for the desired programs more effectively.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "A well-executed work on designing grammars for knowledge-based question answering. The revisit of traditional grammar-based methods provides insights on whether prior information such as types are still useful in the current era of pre-trained models.  (But I also feel the question whether grammar-based constraints are still useful for large models needs to be studied more)", "weaknesses": "The main contribution, as the author points out,  is “To the best of our knowledge, our work is the first to use production rules as actions for semantic parsers based on pre-trained seq2seq models.”. First, I’m not sure what is the underlying challenge of extending traditional grammar-based seq2seqs to their pretrained counterparts? That is, I'm not sure about the technical contribution of the paper. Second, there are already existing works in this direction (as cited in the related work). For example, RAT-SQL (based on BERT) extends TRANX which expands production rules incrementally during decoding.", "questions": "- What is generation speed of the grammar-based parser compared with the baseline BART? I’m wondering that with more constraints, the parser might suffer from much slower inference speed\n- would it be an issue if the size of the set of candidate expressions become too large? E.g., certain class (e.g., keyword-entity) has too many candidates.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698902169167}, {"id": "ytzRarGteH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2430/Reviewer_4fJL"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes an approach to semantic parsing over knowledge base using candidate expressions, which improves the accuracy of semantic parsers on knowledge bases. They evaluate their approach on KQAPro and show that it outperforms other state-of-the-art methods. The authors also conduct an ablation study to analyze the contribution of candidate expressions to the performance of the semantic parser. Overall, the paper's contributions include a new approach to semantic parsing using candidate expressions, and an evaluation of the effectiveness of candidate expressions in improving the accuracy of semantic parsers.", "review_text": "This paper proposes an approach to semantic parsing over knowledge base using candidate expressions, which improves the accuracy of semantic parsers on knowledge bases. They evaluate their approach on KQAPro and show that it outperforms other state-of-the-art methods. The authors also conduct an ablation study to analyze the contribution of candidate expressions to the performance of the semantic parser. Overall, the paper's contributions include a new approach to semantic parsing using candidate expressions, and an evaluation of the effectiveness of candidate expressions in improving the accuracy of semantic parsers.", "strengths": "- This paper provides a clear and detailed explanation of their proposed approach, including the grammar and inference algorithm used to generate candidate expressions.\n- This paper conducts an ablation study to analyze the contribution of candidate expressions to the performance of the semantic parser. This analysis provides insights into the effectiveness of candidate expressions and how they can be used to improve the accuracy of semantic parsers.\n- The paper is of high quality, with clear and well-organized writing and thorough experimental evaluation.", "weaknesses": "The major weakness of this work is its incremental contribution over previous grammar-based methods like [1] and [2]. In particular, [2] uses similar grammars for knowledge base question answering. It is unclear what new techniques or insights this work adds beyond existing grammar-based approaches for this task. To strengthen the paper, the authors could focus more on novel grammar designs or representations that improve performance.\n\nAdditionally, the experimental validation is limited to a single dataset (KQAPro). Testing the approach on an additional challenging dataset like GrailQA [3] would better verify effectiveness and generalization. With only one dataset, it is hard to determine if the performance gains are dataset-specific or represent a robust advancement for grammar-based decoding. Expanding the experimental evaluation would make the results more convincing.\n\nIn summary, clearly situating the contributions relative to prior grammar-based work and testing across more datasets could help address concerns around novelty and experimental validation. Focusing on where this work provides specific technical innovations or new insights would strengthen the paper. \n\n[1]. A Syntactic Neural Model for General-Purpose Code Generation\n[2]. ReTraCk: A Flexible and Efficient Framework for Knowledge Base Question Answering\n[3]. https://dki-lab.github.io/GrailQA/", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an approach to semantic parsing over knowledge base using candidate expressions, which improves the accuracy of semantic parsers on knowledge bases. They evaluate their approach on KQAPro and show that it outperforms other state-of-the-art methods. The authors also conduct an ablation study to analyze the contribution of candidate expressions to the performance of the semantic parser. Overall, the paper's contributions include a new approach to semantic parsing using candidate expressions, and an evaluation of the effectiveness of candidate expressions in improving the accuracy of semantic parsers.", "soundness": "3 good", "presentation": "3 good", "contribution": "1 poor", "strengths": "- This paper provides a clear and detailed explanation of their proposed approach, including the grammar and inference algorithm used to generate candidate expressions.\n- This paper conducts an ablation study to analyze the contribution of candidate expressions to the performance of the semantic parser. This analysis provides insights into the effectiveness of candidate expressions and how they can be used to improve the accuracy of semantic parsers.\n- The paper is of high quality, with clear and well-organized writing and thorough experimental evaluation.", "weaknesses": "The major weakness of this work is its incremental contribution over previous grammar-based methods like [1] and [2]. In particular, [2] uses similar grammars for knowledge base question answering. It is unclear what new techniques or insights this work adds beyond existing grammar-based approaches for this task. To strengthen the paper, the authors could focus more on novel grammar designs or representations that improve performance.\n\nAdditionally, the experimental validation is limited to a single dataset (KQAPro). Testing the approach on an additional challenging dataset like GrailQA [3] would better verify effectiveness and generalization. With only one dataset, it is hard to determine if the performance gains are dataset-specific or represent a robust advancement for grammar-based decoding. Expanding the experimental evaluation would make the results more convincing.\n\nIn summary, clearly situating the contributions relative to prior grammar-based work and testing across more datasets could help address concerns around novelty and experimental validation. Focusing on where this work provides specific technical innovations or new insights would strengthen the paper. \n\n[1]. A Syntactic Neural Model for General-Purpose Code Generation\n[2]. ReTraCk: A Flexible and Efficient Framework for Knowledge Base Question Answering\n[3]. https://dki-lab.github.io/GrailQA/", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698836486147}, {"id": "z7KDM4q5Nw", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2430/Reviewer_iWQY"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper presents a constraint decoding method for semantic parsing. The paper leverages pretrained neural models and tried to narrow down decoding search space by defining IR, type, candidate expressions (whitelisting) to improve semantic parser performance. It goes into details telling the reason why constraint decoding is needed and how it's implemented. Experiment is conducted against KB, and results looks good. The authors put together a complex neural semantic parsing pipeline and proposed a solution for the problem of constrained decoding for non-terminal nodes in structured decoding. On `multi-hop` and `qualifier` test sets, there are decent improvements.", "review_text": "The paper presents a constraint decoding method for semantic parsing. The paper leverages pretrained neural models and tried to narrow down decoding search space by defining IR, type, candidate expressions (whitelisting) to improve semantic parser performance. It goes into details telling the reason why constraint decoding is needed and how it's implemented. Experiment is conducted against KB, and results looks good. The authors put together a complex neural semantic parsing pipeline and proposed a solution for the problem of constrained decoding for non-terminal nodes in structured decoding. On `multi-hop` and `qualifier` test sets, there are decent improvements.", "strengths": "- The constraint decoding approach shows decent improvements in experiments.\n- The authors designed an intermediate representation system based on s-expression which is more generic comparing with other constrained decoding method like Picard. The candidate expression is built as a trie tree structure. Comparing with previous work, this work expands the constraints to internal nodes (non-terminal nodes) which would be helpful to narrow down search space. \n- The writing is clean and easy to follow", "weaknesses": "- This is a relatively incremental work. Constraint decoding is not a new idea which is actually widely used in semantic parsing field. The trie tree structure for candidate expression decoding is also not new (similar to GENRE). The novel part is to expand it to non-terminal node. \n- In the era of LLMs, the improvement seems marginal comparing with using large LMs and with more training data.", "questions": "- Comparing with using production rules etc, have you tried unstructured decoding method? Maybe this should be a good baseline. \n- Clearly with more data, the constrained decoding doesn't help much. Have you explored data augmentation?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a constraint decoding method for semantic parsing. The paper leverages pretrained neural models and tried to narrow down decoding search space by defining IR, type, candidate expressions (whitelisting) to improve semantic parser performance. It goes into details telling the reason why constraint decoding is needed and how it's implemented. Experiment is conducted against KB, and results looks good. The authors put together a complex neural semantic parsing pipeline and proposed a solution for the problem of constrained decoding for non-terminal nodes in structured decoding. On `multi-hop` and `qualifier` test sets, there are decent improvements.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- The constraint decoding approach shows decent improvements in experiments.\n- The authors designed an intermediate representation system based on s-expression which is more generic comparing with other constrained decoding method like Picard. The candidate expression is built as a trie tree structure. Comparing with previous work, this work expands the constraints to internal nodes (non-terminal nodes) which would be helpful to narrow down search space. \n- The writing is clean and easy to follow", "weaknesses": "- This is a relatively incremental work. Constraint decoding is not a new idea which is actually widely used in semantic parsing field. The trie tree structure for candidate expression decoding is also not new (similar to GENRE). The novel part is to expand it to non-terminal node. \n- In the era of LLMs, the improvement seems marginal comparing with using large LMs and with more training data.", "questions": "- Comparing with using production rules etc, have you tried unstructured decoding method? Maybe this should be a good baseline. \n- Clearly with more data, the constrained decoding doesn't help much. Have you explored data augmentation?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1697683041624}], "openreview_url": "https://openreview.net/forum?id=ICSvW69W5K", "arxiv_id": "2410.00414", "paper_pdf": "papers/ICSvW69W5K.pdf", "paper_pdf_sha256": "e9f74dd2245bdb9384e5caa87a0c8712cb258033bf112f5a99f9c055584ec706", "paper_pdf_bytes": 413947, "paper_pdf_source": "openreview", "code_url": "https://github.com/daehwannam/candexpr-sp", "code_repository": "daehwannam/candexpr-sp", "code_commit": "0e6a168e3565c40672970ad28ee7caa88d0a7936", "code_archive": "repos/ICSvW69W5K.zip", "code_archive_sha256": "6ca9ff2b9aaadd46ec9e094a13da7f10c1e28ce437eee5d3c7d99503c6616c6f", "code_archive_bytes": 166866, "code_file_count": 142, "code_extensions": {".py": 139, ".sh": 3}, "github_disk_usage_kb": 412, "github_languages": {"Python": 402632, "Emacs Lisp": 26132, "Shell": 1263, "TeX": 557}, "github_archived": false, "github_pushed_at": "2025-12-31T04:42:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/semantic-parsing-with-candidate-expressions"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dF0g-5k05h_", "year": 2023, "status": "rejected", "title": "The Vendi Score: A Diversity Evaluation Metric for Machine Learning", "authors": ["Dan Friedman", "Adji Bousso Dieng"], "authorids": ["~Dan_Friedman2", "~Adji_Bousso_Dieng1"], "authors_source": "OpenReview API", "abstract": "Diversity is an important criterion for many areas of machine learning (ML), including generative modeling and dataset curation. Yet little work has gone into understanding, formalizing, and measuring diversity in ML. In this paper we address the diversity evaluation problem by proposing the Vendi Score, which connects and extends ideas from ecology and quantum statistical mechanics to ML. The Vendi Score is defined as the exponential of the Shannon entropy of the eigenvalues of a similarity matrix. This matrix is induced by a user-defined similarity function applied to the sample to be evaluated for diversity. In taking a similarity function as input, the Vendi Score enables its user to specify any desired form of diversity. Importantly, unlike many existing metrics in ML, the Vendi Score doesn’t require a reference dataset or distribution over samples or labels, it is therefore general and applicable to any generative model, decoding algorithm, and dataset from any domain where similarity can be defined. We showcase the Vendi Score on molecular generative modeling where we found it addresses shortcomings of the current diversity metric of choice in that domain. We also applied the Vendi Score to generative models of images and decoding algorithms of text where we found it confirms known results about diversity in those domains. Furthermore, we used the Vendi Score to measure mode collapse, a known shortcoming of generative adversarial networks (GANs). In particular, the Vendi Score revealed that even GANs that capture all the modes of a labelled dataset can be less diverse than the original dataset. Finally, the interpretability of the Vendi Score allowed us to diagnose several benchmark ML datasets for diversity, opening the door for diversity-informed data augmentation", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "71uc0RDj8g6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5560/Reviewer_7smz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper proposes to use Vendi score to measure the diversity of datasets.", "review_text": "I am concerned about the novelty of the paper, since most of the contributions come from experimental results. ", "strengths": "Strength:\n\n1. Paper shows many applications of the Vendi score to measure the diversity of data set, i.e. in generative model, mode collapse, image evaluation, algorithm evaluation etc.\n2. The Vendi score requires no label to compute.\n\nWeaknesses:\n\n1. The definition of Vendi score comes straight forward from the Shannon entropy. All of its theoretical results are consequences of results of Shannon entropy.\n\n2. There is no theoretical support for the use of Vendi score in situations presented in the paper. \n\n3. There is no guideline to explain when to use the Vendi score, i.e in the case of LSUN dataset, should we use the Vendi score to rank the efficiency of models?\n\n\nTo improve the work, the authors could prove that the Vendi score is an indicator for the number of modes. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes to use Vendi score to measure the diversity of datasets.", "strength_and_weaknesses": "Strength:\n\n1. Paper shows many applications of the Vendi score to measure the diversity of data set, i.e. in generative model, mode collapse, image evaluation, algorithm evaluation etc.\n2. The Vendi score requires no label to compute.\n\nWeaknesses:\n\n1. The definition of Vendi score comes straight forward from the Shannon entropy. All of its theoretical results are consequences of results of Shannon entropy.\n\n2. There is no theoretical support for the use of Vendi score in situations presented in the paper. \n\n3. There is no guideline to explain when to use the Vendi score, i.e in the case of LSUN dataset, should we use the Vendi score to rank the efficiency of models?\n\n\nTo improve the work, the authors could prove that the Vendi score is an indicator for the number of modes. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper is easy to read and well-written.", "summary_of_the_review": "I am concerned about the novelty of the paper, since most of the contributions come from experimental results. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666695090354}, {"id": "gzgFq17baE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5560/Reviewer_zAvK"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper propose a diversity metric, called Vendi score, that connects and extends ideas from ecology and quantum statistical mechanics to machine learning.  \n\nThe Vendi Score doesn’t require a reference dataset or distribution over samples or labels, it is therefore general and applicable to any generative model, decoding algorithm, and dataset from any domain where similarity can be defined. The authors demonstrate the usage of Vendi score on domains including molecule generation, mode collapse of GANs, and decoding algorithms for texts.   \n", "review_text": "\nI appreciate the authors’ efforts in presenting a notion of diversity score motivated by bio-diversity and applying it to various diversity-related applications in machine learning. However, besides the above weakness concerns, I have the following questions for the authors. I believe these questions need to be well-addressed before the work becomes a solid one.  \n\n### Q: Considering applications of DPP-based diversity, does Vendi score allow for efficient sampling, marginalization or learning, which are frequently needed for applications in ML? \n\nSpecifically, this would be possible if you define an energy-based model based on Vendi score, similar as what DPP does. This energy-based model describes the odds that a specific subset $S$ shall be selected out of the ground set $V$. \n\n$$p(S) \\propto VS_k(S) =  \\exp{-\\sum_i \\lambda_i \\log\\lambda_i},$$\n\nwhere $S:= \\{x_1,…,x_{|S|}  \\},  V:={x_1,…,x_{|V|} }, $  (should be the notion of a set, but it does not show up correctly in this openreview system...)\n\nIn this way, the partition function (normalization constant) shall be written as: \n$$ Z = \\sum_{S\\subseteq V} VS_k(S) $$\n\nOne clear advantage of DPP is that the partition function of it admits a closed form, which is not possible for many other Markov random field variants. Would the distribution defined by the Vendi score allow for similar kind of benefits?   \n\n\n### Q: Would it be possible to axiomatize the diversity metrics?  \n\nThere could be more than one way to define a diversity score (or even metric mathematically), as we have discussed above.  It is beneficial to think of some axioms that shall be met for a reasonable diversity measure: $D(A)$, where $A = \\{x_1, …, x_n\\}$ denotes the set of samples. \nFor example, the following axioms would come to my mind: \n\n1) monotonicity:  $D(A) \\leq D(A\\cup {x})$\n2) symmetry as you defined\n3) some relationship amongst $D(A)$, $D(B)$, $D(A\\cup B)$, $D(A\\cap B)$?   \n4) defining $D(\\emptyset)$? \n\n\n### Q:  How to correctly evaluate the validity of a diversity score empirically?    \n\nThe authors did extensive experiments on the usage of Vendi score in various application domains. However, it hardly provides strong evidence of the advantage of Vendi score.   Molecule generation experiments (sec. 3.1) shows one specific instance where IntDiv and VS have discrepancy. While IntDiv is a pretty weak baseline which results from a heuristic definition of diversity.   Image generation experiments (sec. 3.3) shows that VS generally agrees with existing metrics on low-resolution images. And on LSUN dataset, VS disagrees with IS.  Experiments on decoding text (sec. 3.4) show that VS agrees with N-gram diversity on a caption generation task with 5 sentences.  Experiments on diagnosing datasets (sec. 3.5) demonstrate that VS could provide diversity scores for samples within a specific class in several public datasets.   \n\nOne observation is that there is no, or it is very hard to get “golden” ground truth for diversity scores of a set of samples.  In my opinion, one way could be to ask a group of persons to provide annotations of diversity, either in the form of “scores” or rankings. One practical way to conduct this is to use some crowdsourcing platform.   A second way would be to use existing datasets in a smart way, say the CIFAR-100 dataset. One can make some groups of data samples out of the CIFAR-100, some groups only contain samples from one unique class (maybe called intra-class groups), and some groups have samples from various classes (called inter-class groups). One would naturally expect that the diversity of inter-class groups ranks higher than that of intra-class groups.  One can of course design more refined data groups along this line, to provide more refined way to verify the validity of a diversity score.  \n\n\n\n### Q: Naming it as “Vendi score” seems to be not consistent with the literature. \n\nOn the first glance I thought Vendi might be taken from some literature of bio-diversity. However, it turns out to be no after I checked representative papers on bio-diversity since the term “Vendi” has never appeared in the literature. \n\nThe most related concept may be the “Hill number of order 1” (https://en.wikipedia.org/wiki/Diversity_index#cite_note-Tuomisto2010a-4), see also the paper of Hill [Hill 1973]. So it seems to be reasonable to name it as “Hill score”. \n\n### Reference: \n\nHill, Mark O. \"Diversity and evenness: a unifying notation and its consequences.\" Ecology 54, no. 2 (1973): 427-432.\n\n", "strengths": "\n### Strengths: \n\n- The proposed Vendi score serves as an interpretable metric for diversity.  The authors study several interesting properties of it, which gives a formal understanding of desiderata for diversity. \n- The Vendi score enjoys flexibility and wide applicability.\n- The Vendi score has the potential to be applied for diversity-informed data augmentation.  \n\n\n### Weaknesses: \n- Comparison with  classical diversity tools, e.g.., Determintal point processes （DPPs） is lacked. There is a rich literature on it, stemming from statistical physics and various applications in machine learning.  To name a few:  see [Li et al. 2016] and a wonderful survey [Kulesza-Taskar 2012] and the references therein.   \n\n- To make it clearer, if one follow the DPP definition of a diversity score, let us call it the DPP score (DS for short)  following the same notation in the paper, it shall be defined as \n$$ DS_k(x_1,…,x_n) = \\log |K| = \\sum_i \\log n\\lambda_i = \\sum_i (\\log n + \\log\\lambda_i) $$\nIt looks pretty related to the Vendi score. \n\nMeanwhile, DPP score also has an interpretation, it corresponds to the square of the volume spanned by the vectors of $ x_1,…,x_n$. \n\nCan you comment more on the connections to DPP score? \n\n### Reference: \n\nAlex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2–3):123–286, 2012\n\nChengtao Li, Stefanie Jegelka, and Suvrit Sra. Efficient sampling for k-determinantal point processes. In Artificial Intelligence and Statistics, pp. 1328–1337. PMLR, 2016.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper propose a diversity metric, called Vendi score, that connects and extends ideas from ecology and quantum statistical mechanics to machine learning.  \n\nThe Vendi Score doesn’t require a reference dataset or distribution over samples or labels, it is therefore general and applicable to any generative model, decoding algorithm, and dataset from any domain where similarity can be defined. The authors demonstrate the usage of Vendi score on domains including molecule generation, mode collapse of GANs, and decoding algorithms for texts.   \n", "strength_and_weaknesses": "\n### Strengths: \n\n- The proposed Vendi score serves as an interpretable metric for diversity.  The authors study several interesting properties of it, which gives a formal understanding of desiderata for diversity. \n- The Vendi score enjoys flexibility and wide applicability.\n- The Vendi score has the potential to be applied for diversity-informed data augmentation.  \n\n\n### Weaknesses: \n- Comparison with  classical diversity tools, e.g.., Determintal point processes （DPPs） is lacked. There is a rich literature on it, stemming from statistical physics and various applications in machine learning.  To name a few:  see [Li et al. 2016] and a wonderful survey [Kulesza-Taskar 2012] and the references therein.   \n\n- To make it clearer, if one follow the DPP definition of a diversity score, let us call it the DPP score (DS for short)  following the same notation in the paper, it shall be defined as \n$$ DS_k(x_1,…,x_n) = \\log |K| = \\sum_i \\log n\\lambda_i = \\sum_i (\\log n + \\log\\lambda_i) $$\nIt looks pretty related to the Vendi score. \n\nMeanwhile, DPP score also has an interpretation, it corresponds to the square of the volume spanned by the vectors of $ x_1,…,x_n$. \n\nCan you comment more on the connections to DPP score? \n\n### Reference: \n\nAlex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2–3):123–286, 2012\n\nChengtao Li, Stefanie Jegelka, and Suvrit Sra. Efficient sampling for k-determinantal point processes. In Artificial Intelligence and Statistics, pp. 1328–1337. PMLR, 2016.\n\n", "clarity,_quality,_novelty_and_reproducibility": "\nI believe it is the first time to apply classical bio-diversity scores (specifically, the Hill index of order 1) to several applications in machine learning. However, this work ignores a rich class of diversity related research in machine learning.  \n", "summary_of_the_review": "\nI appreciate the authors’ efforts in presenting a notion of diversity score motivated by bio-diversity and applying it to various diversity-related applications in machine learning. However, besides the above weakness concerns, I have the following questions for the authors. I believe these questions need to be well-addressed before the work becomes a solid one.  \n\n### Q: Considering applications of DPP-based diversity, does Vendi score allow for efficient sampling, marginalization or learning, which are frequently needed for applications in ML? \n\nSpecifically, this would be possible if you define an energy-based model based on Vendi score, similar as what DPP does. This energy-based model describes the odds that a specific subset $S$ shall be selected out of the ground set $V$. \n\n$$p(S) \\propto VS_k(S) =  \\exp{-\\sum_i \\lambda_i \\log\\lambda_i},$$\n\nwhere $S:= \\{x_1,…,x_{|S|}  \\},  V:={x_1,…,x_{|V|} }, $  (should be the notion of a set, but it does not show up correctly in this openreview system...)\n\nIn this way, the partition function (normalization constant) shall be written as: \n$$ Z = \\sum_{S\\subseteq V} VS_k(S) $$\n\nOne clear advantage of DPP is that the partition function of it admits a closed form, which is not possible for many other Markov random field variants. Would the distribution defined by the Vendi score allow for similar kind of benefits?   \n\n\n### Q: Would it be possible to axiomatize the diversity metrics?  \n\nThere could be more than one way to define a diversity score (or even metric mathematically), as we have discussed above.  It is beneficial to think of some axioms that shall be met for a reasonable diversity measure: $D(A)$, where $A = \\{x_1, …, x_n\\}$ denotes the set of samples. \nFor example, the following axioms would come to my mind: \n\n1) monotonicity:  $D(A) \\leq D(A\\cup {x})$\n2) symmetry as you defined\n3) some relationship amongst $D(A)$, $D(B)$, $D(A\\cup B)$, $D(A\\cap B)$?   \n4) defining $D(\\emptyset)$? \n\n\n### Q:  How to correctly evaluate the validity of a diversity score empirically?    \n\nThe authors did extensive experiments on the usage of Vendi score in various application domains. However, it hardly provides strong evidence of the advantage of Vendi score.   Molecule generation experiments (sec. 3.1) shows one specific instance where IntDiv and VS have discrepancy. While IntDiv is a pretty weak baseline which results from a heuristic definition of diversity.   Image generation experiments (sec. 3.3) shows that VS generally agrees with existing metrics on low-resolution images. And on LSUN dataset, VS disagrees with IS.  Experiments on decoding text (sec. 3.4) show that VS agrees with N-gram diversity on a caption generation task with 5 sentences.  Experiments on diagnosing datasets (sec. 3.5) demonstrate that VS could provide diversity scores for samples within a specific class in several public datasets.   \n\nOne observation is that there is no, or it is very hard to get “golden” ground truth for diversity scores of a set of samples.  In my opinion, one way could be to ask a group of persons to provide annotations of diversity, either in the form of “scores” or rankings. One practical way to conduct this is to use some crowdsourcing platform.   A second way would be to use existing datasets in a smart way, say the CIFAR-100 dataset. One can make some groups of data samples out of the CIFAR-100, some groups only contain samples from one unique class (maybe called intra-class groups), and some groups have samples from various classes (called inter-class groups). One would naturally expect that the diversity of inter-class groups ranks higher than that of intra-class groups.  One can of course design more refined data groups along this line, to provide more refined way to verify the validity of a diversity score.  \n\n\n\n### Q: Naming it as “Vendi score” seems to be not consistent with the literature. \n\nOn the first glance I thought Vendi might be taken from some literature of bio-diversity. However, it turns out to be no after I checked representative papers on bio-diversity since the term “Vendi” has never appeared in the literature. \n\nThe most related concept may be the “Hill number of order 1” (https://en.wikipedia.org/wiki/Diversity_index#cite_note-Tuomisto2010a-4), see also the paper of Hill [Hill 1973]. So it seems to be reasonable to name it as “Hill score”. \n\n### Reference: \n\nHill, Mark O. \"Diversity and evenness: a unifying notation and its consequences.\" Ecology 54, no. 2 (1973): 427-432.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666499343352}, {"id": "cO174wamleY", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5560/Reviewer_Z1XS"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed the Vendi Score (VS) to measure the diversity of the generated samples from generative models. It is defined by the exponential of the Shannon entropy of the eigenvalues of a similarity matrix or covariance matrix using the inner-product similarity function (Sec. 2.3).\n\nThis work is reminiscent of spectral clustering, where the eigenvalues of the similarity matrix are used for clustering. Here, many clustering means diverse samples. The author can relate this work to spectral clustering and discuss it.", "review_text": "Generally speaking, the proposed idea is appealing and plausible as a general metric for assessing the diversity of generated samples. However, this work could be significantly improved in writing (see the clarity) and experimental validation (see the weakness, especially W1).", "strengths": "### Strength\n\nThe proposed idea has a compelling theoretical motivation, and it is convenient to use since it only requires the computation of the eigenvalue decomposition for the feature covariance matrix. Through a normalization technique, seeing the distribution of eigenvalues as a probability distribution was notably interesting. The connection to spectral analysis may be worth exploring and discussing. \n\n### Weakness\n\n**W1. Need for human evaluation.** The major concern is experimental validation. Table 1 does not provide the ground-truth diversity; we only learn that VS may capture the HMM's underperformance. Table 2 indirectly validates the VS by showing the sensitivity toward mode detection, not directly on the degree of diversity. Table 3 only compares with the other metrics, which may have some sense of measuring diversity (e.g., FID through its covariance matrices). It just shows general agreements with the other metrics. We do not know which metric is significantly better than one other for measuring diversity. The same goes for Table 4. Even human-written reference captions do not directly confirm that their diversity is captured by BLEU-4, excluding other factors. One suggestion is to evaluate with human judgments on diversity to compare with other conventional and competing metrics. \n\n**W2. Requirement of reference datasets due to inconsistency across domains.** Although the authors argue that it \"doesn't require a reference dataset\" in the Abstract, it shows inconsistent results across domains in Table 3. For instance, Cat or Bedroom datasets have relatively low diversity compared to that of ImageNet. Therefore, we need the diversity score of the corresponding reference (test) dataset to compare the scores. \n\n**W3. Insensitivity of diversity for small domains.** The discriminative power of the Vendi score seems to attenuate for small domains, e.g., Cat and Bedroom results in Table 3. Maybe, if we finetune the feature extractor (inception networks) on the target domain, one may get better results, but it is not explored.\n\n**W4. A problem in inner products as similarity function.** In Sec. 2.3, you should use *cosine similarity* instead of *inner-product* since the range of the output of the inner-product is not bounded to have ones of diagonal elements for the similarity matrix. Without the l2-normalization of features, the critical assumption of this work goes wrong. Could you confirm that the features are appropriately normalized in all your experiments? You did not explicitly mention it. Also, the covariance matrix is defined by $\\mathrm{X}\\mathrm{X}^\\intercal /n$ only when the features are unbiased (subtracted with their mean vector).\n\n**W5. Connection to the quantum statistical mechanics.** Although the authors mention the connection to quantum statistical mechanics *briefly* in 2.1, it seems to oversell the content of the paper in the Abstract unless additional elaborations or authors' speculations.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper proposed the Vendi Score (VS) to measure the diversity of the generated samples from generative models. It is defined by the exponential of the Shannon entropy of the eigenvalues of a similarity matrix or covariance matrix using the inner-product similarity function (Sec. 2.3).\n\nThis work is reminiscent of spectral clustering, where the eigenvalues of the similarity matrix are used for clustering. Here, many clustering means diverse samples. The author can relate this work to spectral clustering and discuss it.", "strength_and_weaknesses": "### Strength\n\nThe proposed idea has a compelling theoretical motivation, and it is convenient to use since it only requires the computation of the eigenvalue decomposition for the feature covariance matrix. Through a normalization technique, seeing the distribution of eigenvalues as a probability distribution was notably interesting. The connection to spectral analysis may be worth exploring and discussing. \n\n### Weakness\n\n**W1. Need for human evaluation.** The major concern is experimental validation. Table 1 does not provide the ground-truth diversity; we only learn that VS may capture the HMM's underperformance. Table 2 indirectly validates the VS by showing the sensitivity toward mode detection, not directly on the degree of diversity. Table 3 only compares with the other metrics, which may have some sense of measuring diversity (e.g., FID through its covariance matrices). It just shows general agreements with the other metrics. We do not know which metric is significantly better than one other for measuring diversity. The same goes for Table 4. Even human-written reference captions do not directly confirm that their diversity is captured by BLEU-4, excluding other factors. One suggestion is to evaluate with human judgments on diversity to compare with other conventional and competing metrics. \n\n**W2. Requirement of reference datasets due to inconsistency across domains.** Although the authors argue that it \"doesn't require a reference dataset\" in the Abstract, it shows inconsistent results across domains in Table 3. For instance, Cat or Bedroom datasets have relatively low diversity compared to that of ImageNet. Therefore, we need the diversity score of the corresponding reference (test) dataset to compare the scores. \n\n**W3. Insensitivity of diversity for small domains.** The discriminative power of the Vendi score seems to attenuate for small domains, e.g., Cat and Bedroom results in Table 3. Maybe, if we finetune the feature extractor (inception networks) on the target domain, one may get better results, but it is not explored.\n\n**W4. A problem in inner products as similarity function.** In Sec. 2.3, you should use *cosine similarity* instead of *inner-product* since the range of the output of the inner-product is not bounded to have ones of diagonal elements for the similarity matrix. Without the l2-normalization of features, the critical assumption of this work goes wrong. Could you confirm that the features are appropriately normalized in all your experiments? You did not explicitly mention it. Also, the covariance matrix is defined by $\\mathrm{X}\\mathrm{X}^\\intercal /n$ only when the features are unbiased (subtracted with their mean vector).\n\n**W5. Connection to the quantum statistical mechanics.** Although the authors mention the connection to quantum statistical mechanics *briefly* in 2.1, it seems to oversell the content of the paper in the Abstract unless additional elaborations or authors' speculations.", "clarity,_quality,_novelty_and_reproducibility": "### Clarity\n\n- C1. The first paragraph of the Introduction could be elaborated, including a definition of diversity, justification of this work, etc.\n\n- C2. The related work in the Introduction section deserves a separate section.\n\n- C3. Why do the authors call it the *Vendi* Score?\n\n- C4. The definition of IntDiv is introduced too later (Sec. 3), although it is referred to in the previous sections frequently.\n\n\n### Quality\n\n- Formal writing does not abbreviate \"does not\" to \"doesn't.\" e.g., \"the Vendi Score doesn't\" in Abstract.\n\n- Appendix could start with alphabetic numbering, not 5.\n\n- Figures and tables are misplaced on the page where it is referred. Although it cannot be perfectly matched, you could try.\n\n### Novelty \n\nBy seeing the distribution of eigenvalues as a probability distribution, the authors proposed to use the entropy of these eigenvalues to assess feature diversity was quite interesting. If the experimental validation shows a significant outperformance over conventional competing metrics, it would be a strong paper for the related communities.", "summary_of_the_review": "Generally speaking, the proposed idea is appealing and plausible as a general metric for assessing the diversity of generated samples. However, this work could be significantly improved in writing (see the clarity) and experimental validation (see the weakness, especially W1).", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666086292607}], "openreview_url": "https://openreview.net/forum?id=dF0g-5k05h_", "arxiv_id": "2210.02410", "paper_pdf": "papers/dF0g-5k05h_.pdf", "paper_pdf_sha256": "fadffb85538c503eea30faad855d8fd857e7dc8a77a1eef2c5b08f8029b4a9b5", "paper_pdf_bytes": 8534544, "paper_pdf_source": "openreview", "code_url": "https://github.com/vertaix/Vendi-Score", "code_repository": "vertaix/Vendi-Score", "code_commit": "ff1dfdbe6356b98a6087540f215b9a9db6db7c11", "code_archive": "repos/dF0g-5k05h_.zip", "code_archive_sha256": "65289de42582014b90d3e9d265cc4f2ed4bddee1c9e75f2893d2e406e6ee5650", "code_archive_bytes": 1132711, "code_file_count": 9, "code_extensions": {".py": 6, ".ipynb": 3}, "github_disk_usage_kb": 1139, "github_languages": {"Python": 17064}, "github_archived": false, "github_pushed_at": "2025-07-27T11:18:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-vendi-score-a-diversity-evaluation-metric"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "m22XrToDacC", "year": 2022, "status": "rejected", "title": "Distributionally Robust Recourse Action", "authors": ["Duy Nguyen", "Ngoc Bui", "Viet Anh Nguyen"], "authorids": ["~Duy_Nguyen2", "~Ngoc_Bui1", "~Viet_Anh_Nguyen2"], "authors_source": "OpenReview API", "abstract": "Recourse actions, also known as counterfactual explanations, aim to explain a particular algorithmic decision by showing one or multiple ways in which the instance could be modified to receive an alternate outcome. Existing recourse recommendations often assume that the machine learning models do not change over time. However, this assumption does not always hold in practice because of data distribution shifts, and in this case, the recourse actions may become invalid. To redress this shortcoming, we propose the Distributionally Robust Recourse Action framework, which generates a recourse action that has high probability of being valid under a mixture of model shifts. We show that the robust recourse can be found efficiently using a projected gradient descent algorithm and we discuss several extensions of our framework. Numerical experiments with both synthetic and real-world datasets demonstrate the benefits of our proposed framework.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "_D0bnh_y8J", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2908/Reviewer_RFkz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper provides a framework for recourse (i.e. counterfactual explanations) that is robust to shifts in the model. They formulate the robustified recourse setup as a min-max optimization problem, where the max is over a neighborhood around the distribution over model parameters. The model parameters are drawn from a mixture of K distributions, so that the neighborhood is specified by Gelbrich distance on each component. \n\nThey propose a finite-dimensional version of the robustified optimization problem, which can be optimized using projected gradient descent. They evaluate their approach on the German credit dataset, the Small Business Administration dataset, and the Student performance dataset, each of which demonstrates a different type of data distribution shift. ", "review_text": "The setting of recourse in the presence of model shifts is well-motivated, especially since models are often updated over time. The idea of formulating the problem as a distributionally robust optimization problem is compelling. \n\nOne weakness of this paper is that the technical solution provided is somewhat limited. In particular, the formulation in (4) relies heavily on the structural properties of the mixture distribution and Gelbrich distance to reformulate the optimization problem, and is not surprising given these assumptions. Since this is the main technical result in the paper, it would have been interesting to see a more general setup that considered richer distance metrics. \n\nFor the empirical section, a weakness is that DiRRAc (the method proposed in this paper) seems to optimize for ell_2 cost, whereas existing approaches seem to optimize for ell_1 cost. Thus, it is not clear how to compare the costs obtained by DiRRAc with the costs obtained in existing approaches. Moreover, it could be interesting to further investigate the tradeoffs between validity and cost for DiRRAc to illustrate the \"cost of robustness\" incurred by the approach. \n\nMinor comment:\n- Distribut,lly -> \"distributionally\" p. 4", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper provides a framework for recourse (i.e. counterfactual explanations) that is robust to shifts in the model. They formulate the robustified recourse setup as a min-max optimization problem, where the max is over a neighborhood around the distribution over model parameters. The model parameters are drawn from a mixture of K distributions, so that the neighborhood is specified by Gelbrich distance on each component. \n\nThey propose a finite-dimensional version of the robustified optimization problem, which can be optimized using projected gradient descent. They evaluate their approach on the German credit dataset, the Small Business Administration dataset, and the Student performance dataset, each of which demonstrates a different type of data distribution shift. ", "main_review": "The setting of recourse in the presence of model shifts is well-motivated, especially since models are often updated over time. The idea of formulating the problem as a distributionally robust optimization problem is compelling. \n\nOne weakness of this paper is that the technical solution provided is somewhat limited. In particular, the formulation in (4) relies heavily on the structural properties of the mixture distribution and Gelbrich distance to reformulate the optimization problem, and is not surprising given these assumptions. Since this is the main technical result in the paper, it would have been interesting to see a more general setup that considered richer distance metrics. \n\nFor the empirical section, a weakness is that DiRRAc (the method proposed in this paper) seems to optimize for ell_2 cost, whereas existing approaches seem to optimize for ell_1 cost. Thus, it is not clear how to compare the costs obtained by DiRRAc with the costs obtained in existing approaches. Moreover, it could be interesting to further investigate the tradeoffs between validity and cost for DiRRAc to illustrate the \"cost of robustness\" incurred by the approach. \n\nMinor comment:\n- Distribut,lly -> \"distributionally\" p. 4", "summary_of_the_review": "Weak reject due to limited technical contribution and limited empirical comparison   \n\n-----\nUpdate after author response: I appreciate the additional experiments for l1 cost included in the revision. Moreover, although I appreciate the author's discussion of the motivation for the Gelbrich distance, I still think that the technical results are somewhat limited. Thus, I keep my assessment the same. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636016388318}, {"id": "sXI931zn9d", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2908/Reviewer_4adh"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problem of recourse actions (a.k.a. counterfactual explanations) while considering data distribution shifts or model shifts. The proposed Distributionally Robust Recourse Action (DiRRAc) framework has the ability to generate valid recourse actions when model parameters shift over time. DiRRAc adopts the distributionally robust optimization technique and the paper proposes a projected gradient descent method to solve the optimization problem. Experiments are conducted with both synthetic and real world data, and the results have shown that DiRRAc methods can generate recourse actions with higher validity than two existing methods.", "review_text": "Strengths:\n1. Most existing work on recourse actions do not consider model change, so the problem addressed by the paper is relatively new, and it is an important problem since model/data shifts are common in practice.\n2. The idea of considering/modelling model shift as a mixture shift of model parameters, and formalizing the problem as a min-max problem.\n3. The experiment results demonstrate the superiority of DiRRAc over the methods compared.\n4. The paper is well written.\n\nWeaknesses:\n1. It is not very clear how challenging it is to adopt the distributionally robust optimization technique for solving the recourse action problem. It would be useful to let readers know clearly that the adoption is non-trivial, which is particularly helpful for readers who are not familiar wi the distributionally robust optimization techinque.\n2. From the paper, ROAR is a method for generating counterfactual explanations that are robust to model shifts, but the experiments conducted do not consider ROAR as a baseline.\n3. It is not clear how efficient the proposed method is, compared to existing methods.\n4. The performance of the proposed method under no model shifts should be evaluated as well.\n\nOther comments: \n1. on page 2, there is a typo: Distribut,lly --> Distributionally\n2. The paper does not have a Conclusion section.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper studies the problem of recourse actions (a.k.a. counterfactual explanations) while considering data distribution shifts or model shifts. The proposed Distributionally Robust Recourse Action (DiRRAc) framework has the ability to generate valid recourse actions when model parameters shift over time. DiRRAc adopts the distributionally robust optimization technique and the paper proposes a projected gradient descent method to solve the optimization problem. Experiments are conducted with both synthetic and real world data, and the results have shown that DiRRAc methods can generate recourse actions with higher validity than two existing methods.", "main_review": "Strengths:\n1. Most existing work on recourse actions do not consider model change, so the problem addressed by the paper is relatively new, and it is an important problem since model/data shifts are common in practice.\n2. The idea of considering/modelling model shift as a mixture shift of model parameters, and formalizing the problem as a min-max problem.\n3. The experiment results demonstrate the superiority of DiRRAc over the methods compared.\n4. The paper is well written.\n\nWeaknesses:\n1. It is not very clear how challenging it is to adopt the distributionally robust optimization technique for solving the recourse action problem. It would be useful to let readers know clearly that the adoption is non-trivial, which is particularly helpful for readers who are not familiar wi the distributionally robust optimization techinque.\n2. From the paper, ROAR is a method for generating counterfactual explanations that are robust to model shifts, but the experiments conducted do not consider ROAR as a baseline.\n3. It is not clear how efficient the proposed method is, compared to existing methods.\n4. The performance of the proposed method under no model shifts should be evaluated as well.\n\nOther comments: \n1. on page 2, there is a typo: Distribut,lly --> Distributionally\n2. The paper does not have a Conclusion section.\n\n", "summary_of_the_review": "The paper tackles a very practical and relatively new problem regarding recourse actions. The overall idea seems reasonable to me and the experiments have demonstrated the effectiveness of the proposed method. The paper is also well written. \n\nHowever, the paper has a few weaknesses as described above. The missing comparison with ROAR is a main concern. The novelty of the proposed method (regarding the adoption of distributionally robust optimization) should be clarified too.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635950203253}, {"id": "iQq4OmtZnlr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2908/Reviewer_a91J"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of generating recourse actions that are robust to shifts in the parameters of the classifier. The authors present a distributionally robust optimization approach and experimentally show that for linear classifiers and under no actionability constraints, the approach generates recourse actions that have high probability of being valid under shifts to the weights of the linear classifier.", "review_text": "Strengths: very clear and thorough exposition of the approach proposed. The approach proposed is very sound technically.\n\nWeaknesses: the experiments presented are rather limited and the value of the contributions is unclear. In particular, the problem of generating robust recourse was previously considered by Upadhyay et al., 2021, but the authors do not provide any evidence as to why their approach may be preferable, which is particularly concerning given that the experiments considered by the authors are heavily inspired in those of Upadhyay et al., 2021.\n\nTo strengthen the contributions of the paper, the authors should focus on improving the experiments section. The contribution of this paper would be much stronger if the authors compared the performance of their approach to that of Upadhyay et al., 2021, validated the claim that “robust optimization solutions can be overly conservative because it may hedge against a pathological parameter in the uncertainty set” in the context of algorithmic recourse, and showed that their proposed approach overcomes this issue.\n\nDetailed comments on the experiments section:\n* As previously mentioned, authors should compare their approach to Upadhyay et al., 2021.\n* Why was L1 not used as the cost function similarly to AR and MACE? Is this an inherent limitation of DiRRAc? If possible, it would be best to use L1 as the cost function for all three approaches.\n* It would be very valuable to also include experiments for non-linear classifiers (e.g. MLPs).\n* It would be very valuable to consider actionability constraints. It seems like in practice many of the features of the real-world data sets would be immutable (e.g. “Recession” in the SBA data set).\n* For the real-world data, you use significantly fewer features than Upadhyay et al., 2021, why is this?\n* In my opinion, none of the test data for which recourse is generated should be used for estimating theta and Sigma. My suggested approach: split the data into train and test, train 100 classifiers only with the train data (e.g. by subsampling 80% of the train data) to estimate theta and Sigma, and then compute recourse on the test data.\n* In the recourse setting, one typically does not assume access to the training data. It would be valuable to discuss how could theta and Sigma be estimated (or rather guessed) given no access to the data, and what would be the implications in terms of recourse validity. This could also be tested experimentally by setting theta to be the weights of the classifier for which recourse is generated and over and under estimating Sigma to various degrees.\n* For the real-world data, please include the accuracy of the classifier for which recourse is generated.\n\n\nDetailed comments on the introduction:\n* Contrary to what is stated in  Section 1 Paragraph 2, in my opinion “counterfactual explanations” and “recourse actions” should not be used interchangeably, but rather authors should refer only to “recourse actions”, since the problem of recourse validity under data shifts, at least how it is presented in this work (Section 1 Paragraph 5), is not applicable to counterfactual explanations.\n* Section 1, Paragraph 2: “If a specific application can provide the negative outcomes with recourse actions, it can improve the user engagement and boost the interpretability at the same time.” Citations needed for these statements.\n* Section 1, Paragraph 3: “we must consider age as an immutable feature” It is not a must to consider age immutable, several works consider it as a non-decreasing feature.\n* Section 1, Paragraph 4: “Various solutions has have been proposed...”\n* Section 1, Paragraph 5, “Data shifts usually induce corresponding shifts in the machine learning models’ parameters” Organizations usually retrain models periodically in repose to data shifts. The data shift itself does not induces changes to the parameters of the model.\n* Section 1, Paragraph 5: “If a recourse action fails to generate a favorable outcome in the future, then the recourse action becomes useless” Not necessarily, see Handling change over time via ex post facto in Venkatasubramanian and Alfano, 2020.\n* Section 1, Paragraph 5: “and the trust on the machine learning system is lost.” Citation needed.\n\nComments regarding section 2, 3 and 4:\n* In my opinion Algorithm 1 and Theorem 3.4 should not be in the main text, as these are just general well-known concepts from convex optimization, leaving more room for the experiments sections. Authors could just mention convergence rate of $O(1/\\sqrt(T))$.\n* Similarly, Sections 4.1 and 4.2 could be moved to the appendix, since they are not core to the main contribution of the paper and are not evaluated in the experiments section. This would also give the authors more space for the experiments section and a conclusion.\n* Equation 1: min_x\n* Equation 2, 7, 8a, 9a: $inf/min_{x \\in X}$ or add $x \\in X$ as an explicit constrain similarly to Equation 1.\n* Equation 9a, please explain why a margin of 0.5 rather than 1 is used.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of generating recourse actions that are robust to shifts in the parameters of the classifier. The authors present a distributionally robust optimization approach and experimentally show that for linear classifiers and under no actionability constraints, the approach generates recourse actions that have high probability of being valid under shifts to the weights of the linear classifier.", "main_review": "Strengths: very clear and thorough exposition of the approach proposed. The approach proposed is very sound technically.\n\nWeaknesses: the experiments presented are rather limited and the value of the contributions is unclear. In particular, the problem of generating robust recourse was previously considered by Upadhyay et al., 2021, but the authors do not provide any evidence as to why their approach may be preferable, which is particularly concerning given that the experiments considered by the authors are heavily inspired in those of Upadhyay et al., 2021.\n\nTo strengthen the contributions of the paper, the authors should focus on improving the experiments section. The contribution of this paper would be much stronger if the authors compared the performance of their approach to that of Upadhyay et al., 2021, validated the claim that “robust optimization solutions can be overly conservative because it may hedge against a pathological parameter in the uncertainty set” in the context of algorithmic recourse, and showed that their proposed approach overcomes this issue.\n\nDetailed comments on the experiments section:\n* As previously mentioned, authors should compare their approach to Upadhyay et al., 2021.\n* Why was L1 not used as the cost function similarly to AR and MACE? Is this an inherent limitation of DiRRAc? If possible, it would be best to use L1 as the cost function for all three approaches.\n* It would be very valuable to also include experiments for non-linear classifiers (e.g. MLPs).\n* It would be very valuable to consider actionability constraints. It seems like in practice many of the features of the real-world data sets would be immutable (e.g. “Recession” in the SBA data set).\n* For the real-world data, you use significantly fewer features than Upadhyay et al., 2021, why is this?\n* In my opinion, none of the test data for which recourse is generated should be used for estimating theta and Sigma. My suggested approach: split the data into train and test, train 100 classifiers only with the train data (e.g. by subsampling 80% of the train data) to estimate theta and Sigma, and then compute recourse on the test data.\n* In the recourse setting, one typically does not assume access to the training data. It would be valuable to discuss how could theta and Sigma be estimated (or rather guessed) given no access to the data, and what would be the implications in terms of recourse validity. This could also be tested experimentally by setting theta to be the weights of the classifier for which recourse is generated and over and under estimating Sigma to various degrees.\n* For the real-world data, please include the accuracy of the classifier for which recourse is generated.\n\n\nDetailed comments on the introduction:\n* Contrary to what is stated in  Section 1 Paragraph 2, in my opinion “counterfactual explanations” and “recourse actions” should not be used interchangeably, but rather authors should refer only to “recourse actions”, since the problem of recourse validity under data shifts, at least how it is presented in this work (Section 1 Paragraph 5), is not applicable to counterfactual explanations.\n* Section 1, Paragraph 2: “If a specific application can provide the negative outcomes with recourse actions, it can improve the user engagement and boost the interpretability at the same time.” Citations needed for these statements.\n* Section 1, Paragraph 3: “we must consider age as an immutable feature” It is not a must to consider age immutable, several works consider it as a non-decreasing feature.\n* Section 1, Paragraph 4: “Various solutions has have been proposed...”\n* Section 1, Paragraph 5, “Data shifts usually induce corresponding shifts in the machine learning models’ parameters” Organizations usually retrain models periodically in repose to data shifts. The data shift itself does not induces changes to the parameters of the model.\n* Section 1, Paragraph 5: “If a recourse action fails to generate a favorable outcome in the future, then the recourse action becomes useless” Not necessarily, see Handling change over time via ex post facto in Venkatasubramanian and Alfano, 2020.\n* Section 1, Paragraph 5: “and the trust on the machine learning system is lost.” Citation needed.\n\nComments regarding section 2, 3 and 4:\n* In my opinion Algorithm 1 and Theorem 3.4 should not be in the main text, as these are just general well-known concepts from convex optimization, leaving more room for the experiments sections. Authors could just mention convergence rate of $O(1/\\sqrt(T))$.\n* Similarly, Sections 4.1 and 4.2 could be moved to the appendix, since they are not core to the main contribution of the paper and are not evaluated in the experiments section. This would also give the authors more space for the experiments section and a conclusion.\n* Equation 1: min_x\n* Equation 2, 7, 8a, 9a: $inf/min_{x \\in X}$ or add $x \\in X$ as an explicit constrain similarly to Equation 1.\n* Equation 9a, please explain why a margin of 0.5 rather than 1 is used.\n", "summary_of_the_review": "While the authors present a very clear and thorough exposition of the approach proposed and the approach is technically sound, the experiments presented are very limited and the overall value of the contribution is unclear. In particular, the authors do not compare their proposed method with previous approaches addressing the problem of generating robust recourse actions.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635511908602}], "openreview_url": "https://openreview.net/forum?id=m22XrToDacC", "arxiv_id": "2302.11211", "paper_pdf": "papers/m22XrToDacC.pdf", "paper_pdf_sha256": "3b727f15e569b77d6befdee9902200ead6f72e5c4b2785f2cd8c28f506197afe", "paper_pdf_bytes": 619372, "paper_pdf_source": "openreview", "code_url": "https://github.com/duykhuongnguyen/DiRRAc", "code_repository": "duykhuongnguyen/DiRRAc", "code_commit": "928e021fed815b63b351cb79964530c6bd3d35b8", "code_archive": "repos/m22XrToDacC.zip", "code_archive_sha256": "ae30c2fc7c948680f0a20460bb8555ad2b713e2721a9589973ff3c48638d86da", "code_archive_bytes": 5183492, "code_file_count": 56, "code_extensions": {".py": 48, ".ipynb": 8}, "github_disk_usage_kb": 4779, "github_languages": {"Jupyter Notebook": 741114, "Python": 420207, "Shell": 2772}, "github_archived": false, "github_pushed_at": "2023-02-13T03:03:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/distributionally-robust-recourse-action-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pwwVuSICBgt", "year": 2021, "status": "rejected", "title": "Enabling Binary Neural Network Training on the Edge", "authors": ["Erwei Wang", "James J. Davis", "Daniele Moro", "Piotr Zielinski", "Claudionor Coelho", "Satrajit Chatterjee", "Peter Y. K. Cheung", "George Anthony Constantinides"], "authorids": ["~Erwei_Wang1", "james.davis@imperial.ac.uk", "danielemoro@google.com", "~Piotr_Zielinski1", "claudionor.coelho@alumni.stanford.edu", "~Satrajit_Chatterjee1", "p.cheung@imperial.ac.uk", "~George_Anthony_Constantinides1"], "authors_source": "OpenReview API", "abstract": "The ever-growing computational demands of increasingly complex machine learning models frequently necessitate the use of powerful cloud-based infrastructure for their training. Binary neural networks are known to be promising candidates for on-device inference due to their extreme compute and memory savings over higher-precision alternatives. In this paper, we demonstrate that they are also strongly robust to gradient quantization, thereby making the training of modern models on the edge a practical reality. We introduce a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions and energy savings vs Courbariaux & Bengio's standard approach. Against the latter, we see coincident memory requirement and energy consumption drops of 2--6$\\times$, while reaching similar test accuracy, across a range of small-scale models trained to classify popular datasets. We also showcase ImageNet training of ResNetE-18, achieving a 3.12$\\times$ memory reduction over the aforementioned standard. Such savings will allow for unnecessary cloud offloading to be avoided, reducing latency and increasing energy efficiency while also safeguarding user privacy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "SB03xhDUg_a", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3413/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a novel method to make the training Binary Neural Networks with low-memory and low-energy by modifying the backpropagation and forward process. To this end, the paper binarized weight gradients, change batch normalization layer for removing full-precision the inputs, and utilize quantization to accelerate the whole training BNNs procedure.   \n\nThe paper targeted one of the problems in Binary Neural Networks and provided experiments as well as source codes to the proof of efficiency. This is the most significant contribution of the paper. \n\nHowever, there are the following concerns:\n- Need more experiments on the actual training time between two BNNs which are trained with quantized and full-precision gradient weight.\n- Comparison with similar works in recent BNNs.\n- What does it mean B variables in Algorithm 2 of lines: 6, 9, and 10, and how does it influence the performance of training?\n- In figure 2, the legends should be put in the figure (not in the caption). It is better to follow.\n- It would be nice to have the training time and accuracy in the large-scale dataset of ImageNet.\n\nIn conclusion, the paper addresses the novel idea for the training improvement of Binary Neural Networks in low-memory and low-energy. However, there are many concerns aforementioned.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #1", "review": "This paper proposes a novel method to make the training Binary Neural Networks with low-memory and low-energy by modifying the backpropagation and forward process. To this end, the paper binarized weight gradients, change batch normalization layer for removing full-precision the inputs, and utilize quantization to accelerate the whole training BNNs procedure.   \n\nThe paper targeted one of the problems in Binary Neural Networks and provided experiments as well as source codes to the proof of efficiency. This is the most significant contribution of the paper. \n\nHowever, there are the following concerns:\n- Need more experiments on the actual training time between two BNNs which are trained with quantized and full-precision gradient weight.\n- Comparison with similar works in recent BNNs.\n- What does it mean B variables in Algorithm 2 of lines: 6, 9, and 10, and how does it influence the performance of training?\n- In figure 2, the legends should be put in the figure (not in the caption). It is better to follow.\n- It would be nice to have the training time and accuracy in the large-scale dataset of ImageNet.\n\nIn conclusion, the paper addresses the novel idea for the training improvement of Binary Neural Networks in low-memory and low-energy. However, there are many concerns aforementioned.  ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603923928068}, {"id": "nwvrc0YLcUI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3413/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Comments\nSummary:\n\nThe authors proposed a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions and energy savings. The methods include binarizing weight gradients, modifying the forward and backward batch normalization operations and using power-of-two activation gradients and reduced-precision floating-point data. The experimental results show some improvement memory footprint reductions and energy savings vs standard approach.\n\n\nStrength:\n-- The authors carried out a relatively sufficient experimental analysis, and evaluated across multiple models, data sets, optimizers and batch sizes\n-- Storage and energy consumption based on hardware models increases the completeness and credibility of conclusions.\n-- The experiment results seem that the proposed method achieves good performance in memory footprint reductions and energy savings.\n\nWeakness:\n-- The proposed method seems like a combination of existing technical methods.\n-- There is a lack of comparison with other low-cost binary neural network training works.\n\n\nComments:\n(1)\tThere is a difference in the memory consumption in Table1 and in section 5.1 (1.67 MiB or 1.41 MiB ?). The authors may check this.\n(2)\tAs for the perspective of overall design, it’s better to emphasize the trade-offs between the importance of different variables to the overall training and the choose of the data type.\n(3)\tIt’s better to add some comparisons with other low-cost binary neural network training works.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "review", "review": "#### Comments\nSummary:\n\nThe authors proposed a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions and energy savings. The methods include binarizing weight gradients, modifying the forward and backward batch normalization operations and using power-of-two activation gradients and reduced-precision floating-point data. The experimental results show some improvement memory footprint reductions and energy savings vs standard approach.\n\n\nStrength:\n-- The authors carried out a relatively sufficient experimental analysis, and evaluated across multiple models, data sets, optimizers and batch sizes\n-- Storage and energy consumption based on hardware models increases the completeness and credibility of conclusions.\n-- The experiment results seem that the proposed method achieves good performance in memory footprint reductions and energy savings.\n\nWeakness:\n-- The proposed method seems like a combination of existing technical methods.\n-- There is a lack of comparison with other low-cost binary neural network training works.\n\n\nComments:\n(1)\tThere is a difference in the memory consumption in Table1 and in section 5.1 (1.67 MiB or 1.41 MiB ?). The authors may check this.\n(2)\tAs for the perspective of overall design, it’s better to emphasize the trade-offs between the importance of different variables to the overall training and the choose of the data type.\n(3)\tIt’s better to add some comparisons with other low-cost binary neural network training works.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603896115293}, {"id": "XWf3ptGW1oC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3413/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I agree with the key contributions listed in the paper, especially the binarization of weight gradients and activations. The paper is well written and clearly articulates a contribution to the literature. The proposed BNN training scheme can have a significant practical impact. The experimental evidence is provided for several standard image classification tasks. Most of the related works are cited. The paper does not contain a theory part, but wherever possible, equations are provided to illustrate how the method works.\n\nConcerns: \nThis paper includes a detailed empirical evaluation of the proposed BNN training scheme. The major concern is that the proposed low-cost BNN training scheme can cause a nontrivial accuracy degradation (2.25%) as shown in Table 5. The tradeoff between accuracy and memory footprint/energy consumption is not carefully evaluated. For example, a smaller network model with fewer parameters and activations can be trained using the baseline BNN training scheme to reduce memory and energy consumption. Besides, the baseline networks (BinaryNet and ResNet) used for comparison are out-of-date.  Many recent works [1-3] propose new BNN architectures, which improve the accuracy of BNNs significantly. It is useful to justify the effectiveness of the proposed scheme for these SOTA BNN architectures.\n\nTo evaluate the energy consumption of the traditional and the proposed BNN training scheme, the authors assume that the two training schemes have the same convergence rate. Although the traditional BNN training scheme consumes higher power during training, it might take fewer epochs to reach the same accuracy. Therefore, the saving in energy consumption can be lower than the reported numbers. In addition, the estimated energy consumption of BNN training obtained from using QKeras is very rough, which makes the improvement in energy consumption less convincing. The authors can mainly improve the paper's strength by prototyping the proposed BNN training scheme on an embedded CPU and measuring real-world performance and power.\n\nReasons for score: In general, I like the idea of enabling low-cost BNN training by identifying unnecessary high-precision data. However, the improvement numbers presented in the paper need better justification. I would consider raising my score if the authors could address the aforementioned concerns.\n\n[1] Bi-Real Net: Enhancing the Performance of 1-bit CNNs With Improved Representational Capability and Advanced Training Algorithm\n\n[2] ProxyBNN: Learning Binarized Neural Networks via Proxy Matrices\n\n[3] ReActNet: Towards Precise Binary Neural Network with Generalized Activation Functions\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This work proposes a low-cost BNN training scheme to reduce memory consumption and improve energy efficiency. Compared to the traditional BNN training algorithm, the proposed scheme achieves a significant reduction in memory footprint and energy consumption on multiple datasets.", "review": "I agree with the key contributions listed in the paper, especially the binarization of weight gradients and activations. The paper is well written and clearly articulates a contribution to the literature. The proposed BNN training scheme can have a significant practical impact. The experimental evidence is provided for several standard image classification tasks. Most of the related works are cited. The paper does not contain a theory part, but wherever possible, equations are provided to illustrate how the method works.\n\nConcerns: \nThis paper includes a detailed empirical evaluation of the proposed BNN training scheme. The major concern is that the proposed low-cost BNN training scheme can cause a nontrivial accuracy degradation (2.25%) as shown in Table 5. The tradeoff between accuracy and memory footprint/energy consumption is not carefully evaluated. For example, a smaller network model with fewer parameters and activations can be trained using the baseline BNN training scheme to reduce memory and energy consumption. Besides, the baseline networks (BinaryNet and ResNet) used for comparison are out-of-date.  Many recent works [1-3] propose new BNN architectures, which improve the accuracy of BNNs significantly. It is useful to justify the effectiveness of the proposed scheme for these SOTA BNN architectures.\n\nTo evaluate the energy consumption of the traditional and the proposed BNN training scheme, the authors assume that the two training schemes have the same convergence rate. Although the traditional BNN training scheme consumes higher power during training, it might take fewer epochs to reach the same accuracy. Therefore, the saving in energy consumption can be lower than the reported numbers. In addition, the estimated energy consumption of BNN training obtained from using QKeras is very rough, which makes the improvement in energy consumption less convincing. The authors can mainly improve the paper's strength by prototyping the proposed BNN training scheme on an embedded CPU and measuring real-world performance and power.\n\nReasons for score: In general, I like the idea of enabling low-cost BNN training by identifying unnecessary high-precision data. However, the improvement numbers presented in the paper need better justification. I would consider raising my score if the authors could address the aforementioned concerns.\n\n[1] Bi-Real Net: Enhancing the Performance of 1-bit CNNs With Improved Representational Capability and Advanced Training Algorithm\n\n[2] ProxyBNN: Learning Binarized Neural Networks via Proxy Matrices\n\n[3] ReActNet: Towards Precise Binary Neural Network with Generalized Activation Functions\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603867399189}, {"id": "-n9jdfNsfqF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3413/AnonReviewer4"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "I think this a very good contribution to ICLR given the topic and the quality of the submission (originality, contribution to the state of the art, experimental evidence, e\n\n Some of the strong points of the submission are summarized as follows, along with some points for clarification\n\n1.\tEnabling training on any embedded device (SoC, FPGA, micro-controller) is one of the holy grails for edge AI and IoT. As the authors mention, this also intersects with other domains such as federating learning privacy by design systems. The authors provide and ample motivations of the importance of this work, and some of the applications edge AI might enable, as well as the current challenges.\n2.\tThe state of the art (despite the previous comment) contextualizes the subject matter in a succinct but comprehensive manner. Although there are certain aspects that could be improved, such as including a table outlining in a clearer manner the contributions of the authors in this context.\n3.\tThe comparison with the traditional training method is clear. However, I would like to know if the authors have made an ablation study to assess whether or not the use of batch normalization would have an effect on the accuracy of the proposed models. Do you provide experimental evidence of the lack of degradation due to the use of l1 batch normalization? These two aspects are not mentioned in the next nor provided in the supplementary sections.\n4.\tThe experimental design is good, showing a careful analysis to validate the proposal and several ablation studies to assess the memory footprint reductions and its effects on the training of various models\n5.\tThe foundations for the method are presented in great detail in a formalized manner and provides sufficient elements (i.e. experiments) to assess the validity of the proposed approach.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "I think this paper is very interesting and if some minor comments are addressed, it should be accepted for presentation at ICLR21", "review": "I think this a very good contribution to ICLR given the topic and the quality of the submission (originality, contribution to the state of the art, experimental evidence, e\n\n Some of the strong points of the submission are summarized as follows, along with some points for clarification\n\n1.\tEnabling training on any embedded device (SoC, FPGA, micro-controller) is one of the holy grails for edge AI and IoT. As the authors mention, this also intersects with other domains such as federating learning privacy by design systems. The authors provide and ample motivations of the importance of this work, and some of the applications edge AI might enable, as well as the current challenges.\n2.\tThe state of the art (despite the previous comment) contextualizes the subject matter in a succinct but comprehensive manner. Although there are certain aspects that could be improved, such as including a table outlining in a clearer manner the contributions of the authors in this context.\n3.\tThe comparison with the traditional training method is clear. However, I would like to know if the authors have made an ablation study to assess whether or not the use of batch normalization would have an effect on the accuracy of the proposed models. Do you provide experimental evidence of the lack of degradation due to the use of l1 batch normalization? These two aspects are not mentioned in the next nor provided in the supplementary sections.\n4.\tThe experimental design is good, showing a careful analysis to validate the proposal and several ablation studies to assess the memory footprint reductions and its effects on the training of various models\n5.\tThe foundations for the method are presented in great detail in a formalized manner and provides sufficient elements (i.e. experiments) to assess the validity of the proposed approach.\n\n\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603733614553}], "openreview_url": "https://openreview.net/forum?id=pwwVuSICBgt", "arxiv_id": "2102.04270", "paper_pdf": "papers/pwwVuSICBgt.pdf", "paper_pdf_sha256": "f4e1e877ae21f7ad9963c980bd59656a790b0979441226ac6a7f7c536784be73", "paper_pdf_bytes": 457318, "paper_pdf_source": "openreview", "code_url": "https://github.com/awai54st/Enabling-Binary-Neural-Network-Training-on-the-Edge", "code_repository": "awai54st/Enabling-Binary-Neural-Network-Training-on-the-Edge", "code_commit": "b2c026b09e81ea01c778a4d500af99df612a7a63", "code_archive": "repos/pwwVuSICBgt.zip", "code_archive_sha256": "1e9caef5bc332ac902f9b23fa3ff6240afff13b531d0b7090b86363f1805b295", "code_archive_bytes": 3917263, "code_file_count": 190, "code_extensions": {".cpp": 93, ".py": 48, ".h": 45, ".ipynb": 4}, "github_disk_usage_kb": 2901, "github_languages": {"C++": 548945, "Python": 291996, "Jupyter Notebook": 283149, "C": 1099}, "github_archived": false, "github_pushed_at": "2022-03-06T22:03:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/enabling-binary-neural-network-training-on-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "rVrWNb2XLi", "year": 2026, "status": "rejected", "title": "SpecOffload: Unlocking Latent GPU Capacity for LLM Inference on Resource-Constrained Devices", "authors": ["Xiangwen Zhuge", "Xu Shen", "Zeyu Wang", "Fan Dang", "Xuan Ding", "Danyang Li", "Yahui Han", "Tianxiang Hao", "Zheng Yang"], "authorids": ["~Xiangwen_Zhuge1", "~Xu_Shen5", "~Zeyu_Wang28", "~Fan_Dang1", "~Xuan_Ding3", "~Danyang_Li3", "~Yahui_Han1", "~Tianxiang_Hao1", "~Zheng_Yang1"], "authors_source": "OpenReview API", "abstract": "Efficient LLM inference on resource-constrained devices (i.e., PCs with a single commodity GPU) presents significant challenges in compute and memory utilization. Due to limited GPU memory, existing systems offload model weights to CPU memory, incurring substantial I/O overhead between the CPU and GPU. This leads to two major inefficiencies: (1) GPU cores are underutilized, often remaining idle while waiting for data to be loaded; and (2) GPU memory has a low impact on performance, as reducing its capacity has minimal effect on overall throughput. In this paper, we propose SpecOffload, a high-throughput inference engine that embeds speculative decoding into offloading. Our key idea is to unlock latent GPU resources for storing and executing a draft model used for speculative decoding, thus accelerating inference at near-zero additional cost. To support this, we carefully orchestrate the interleaved execution of target and draft models in speculative decoding within the offloading pipeline, and propose a planner to manage tensor placement and select optimal parameters. Compared with the best baseline, SpecOffload improves GPU core utilization by 4.49× and boosts inference throughput by 2.54×. Anonymous repo is at https://anonymous.4open.science/r/SpecOffload-F3F2/.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "iEaO7KSjcn", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15033/Reviewer_D7ni"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper proposes an LLM inference system that combines offloading and speculative decoding. It executes the draft model concurrently on the idle GPU with the offloading and CPU computation of the target model. It also provides modeling of the execution pipeline and planner to search for the configuration. Evaluation show that this work has superior performance than the related offloading-based LLM inference systems.", "review_text": "This paper proposes an LLM inference system that combines offloading and speculative decoding. It executes the draft model concurrently on the idle GPU with the offloading and CPU computation of the target model. It also provides modeling of the execution pipeline and planner to search for the configuration. Evaluation show that this work has superior performance than the related offloading-based LLM inference systems.", "strengths": "1. This paper focuses on an important topic of efficient LLM serving.\n2. It proposes a clever method to fill the draft model execution into the bubbles of offloading of the target model.", "weaknesses": "1. The planner part is not clear enough, nor necessary enough.\n2. Some analysis and claim are not solid enough.", "questions": "1. Why put the Attention computation of the target model on the CPU? Is it only to save the GPU memory for the KV cache? Note Attention is usually memory bound and the GPU has higher bandwidth than the CPU.\n2. The results in Figure 2 lack of some details. How is the increased GPU memory used? For example, one method can be making some portion of the weight persistent in the increased GPU memory and thus the performance can improve gradually.\n3. In Figure 4, why the execution of the Draft model is longer than the FFN of the target model? (Note for moderate sequence length, the execution of the target model on the GPU should be dominated by its FFN.)\n4. In Section 4.3, how the equation 1 is solved? Besides, equation 4 seems not accurate. For example, if $T_{draft}$ is very long, then the FFN of the target model will need to wait for the end of the draft model (according to Figure 4). In this case, the time will not be the max of the two, but seems $T_{draft} + T_{target-ffn}$.\n5. It is not clear what is the search space of the planner. It seems everything is deterministic, or can be defined very easy. So the planner seems not necessary.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes an LLM inference system that combines offloading and speculative decoding. It executes the draft model concurrently on the idle GPU with the offloading and CPU computation of the target model. It also provides modeling of the execution pipeline and planner to search for the configuration. Evaluation show that this work has superior performance than the related offloading-based LLM inference systems.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. This paper focuses on an important topic of efficient LLM serving.\n2. It proposes a clever method to fill the draft model execution into the bubbles of offloading of the target model.", "weaknesses": "1. The planner part is not clear enough, nor necessary enough.\n2. Some analysis and claim are not solid enough.", "questions": "1. Why put the Attention computation of the target model on the CPU? Is it only to save the GPU memory for the KV cache? Note Attention is usually memory bound and the GPU has higher bandwidth than the CPU.\n2. The results in Figure 2 lack of some details. How is the increased GPU memory used? For example, one method can be making some portion of the weight persistent in the increased GPU memory and thus the performance can improve gradually.\n3. In Figure 4, why the execution of the Draft model is longer than the FFN of the target model? (Note for moderate sequence length, the execution of the target model on the GPU should be dominated by its FFN.)\n4. In Section 4.3, how the equation 1 is solved? Besides, equation 4 seems not accurate. For example, if $T_{draft}$ is very long, then the FFN of the target model will need to wait for the end of the draft model (according to Figure 4). In this case, the time will not be the max of the two, but seems $T_{draft} + T_{target-ffn}$.\n5. It is not clear what is the search space of the planner. It seems everything is deterministic, or can be defined very easy. So the planner seems not necessary.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1762248671600}, {"id": "SmI2G8uBPE", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15033/Reviewer_7n7D"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper makes an observation that there are underutilization of GPU cores due to the memory communication between CPU and GPU. Also, the paper claims that adding more GPU has a minimal effect on overall throughput as the model parameters are too large to reside in GPU memory permanently. introduces SpecOffload which is aims to utilize the idle GPU time to run lightweight draft model for speculative decoding. It claims that storing the draft model in the GPU memory would have negligible impact on performance. The paper uses a dual-batch interleaved design enables the target model and draft model to run concurrently, increasing GPU resource utilization.", "review_text": "The paper makes an observation that there are underutilization of GPU cores due to the memory communication between CPU and GPU. Also, the paper claims that adding more GPU has a minimal effect on overall throughput as the model parameters are too large to reside in GPU memory permanently. introduces SpecOffload which is aims to utilize the idle GPU time to run lightweight draft model for speculative decoding. It claims that storing the draft model in the GPU memory would have negligible impact on performance. The paper uses a dual-batch interleaved design enables the target model and draft model to run concurrently, increasing GPU resource utilization.", "strengths": "* Efficient LLM inference is important considering its increasing adoption in many applications. As such the paper is working on a good direction.\n* The paper aims to better utilize GPUs which is important and seems to get some gains.\n* Amalgamation of Offloading and Speculative Decoding seems to be a clear nice followup for the Offloading works and Speculative Decoding works.", "weaknesses": "* The paper seems to be rather shallow on the experiment depth (please refer to the questions for details). This significantly limits its generalizability.", "questions": "* If the draft model is really small. How does this compare to the works that tries to leverage CPU resources for some part of the LLM compute. How would this perform if CPU's AMX units are used for draft model inference and GPU for the target model inference?\n    * Na, Seonjin, et al. \"FlexInfer: Flexible LLM Inference with CPU Computations.\" Eighth Conference on Machine Learning and Systems. 2025\n* The paper mentions that speculative decoding is not always reliable and that its acceleration effect is limited if none of the draft tokens are accepted. However, it does not provide an in-depth analysis of the draft token acceptance rate, which is a critical factor for the system's performance. Can the authors add more experimental data on how this rate varies across different models, datasets, or hardware environments?\n* While the paper states the planner achieves 93.7% of optimal policy performance on average, it does not provide detailed information on its computational cost, runtime overhead, or how it performs in more varied or complex scenarios. Can you provide more details? Also, can you provide more details as to how this can be generalized to different CPU-GPU combinations + models?\n* It seems as the work is only focused on increasing the throughput in single-user scenario. How would this be extended to multi-user / dynamic serving scenarios?\n* A.3.3 Table 13 seems to suggest that there is only 53% improvement for even double VRAM. Could you elaborate. It would also help if utilization data are provided?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper makes an observation that there are underutilization of GPU cores due to the memory communication between CPU and GPU. Also, the paper claims that adding more GPU has a minimal effect on overall throughput as the model parameters are too large to reside in GPU memory permanently. introduces SpecOffload which is aims to utilize the idle GPU time to run lightweight draft model for speculative decoding. It claims that storing the draft model in the GPU memory would have negligible impact on performance. The paper uses a dual-batch interleaved design enables the target model and draft model to run concurrently, increasing GPU resource utilization.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "* Efficient LLM inference is important considering its increasing adoption in many applications. As such the paper is working on a good direction.\n* The paper aims to better utilize GPUs which is important and seems to get some gains.\n* Amalgamation of Offloading and Speculative Decoding seems to be a clear nice followup for the Offloading works and Speculative Decoding works.", "weaknesses": "* The paper seems to be rather shallow on the experiment depth (please refer to the questions for details). This significantly limits its generalizability.", "questions": "* If the draft model is really small. How does this compare to the works that tries to leverage CPU resources for some part of the LLM compute. How would this perform if CPU's AMX units are used for draft model inference and GPU for the target model inference?\n    * Na, Seonjin, et al. \"FlexInfer: Flexible LLM Inference with CPU Computations.\" Eighth Conference on Machine Learning and Systems. 2025\n* The paper mentions that speculative decoding is not always reliable and that its acceleration effect is limited if none of the draft tokens are accepted. However, it does not provide an in-depth analysis of the draft token acceptance rate, which is a critical factor for the system's performance. Can the authors add more experimental data on how this rate varies across different models, datasets, or hardware environments?\n* While the paper states the planner achieves 93.7% of optimal policy performance on average, it does not provide detailed information on its computational cost, runtime overhead, or how it performs in more varied or complex scenarios. Can you provide more details? Also, can you provide more details as to how this can be generalized to different CPU-GPU combinations + models?\n* It seems as the work is only focused on increasing the throughput in single-user scenario. How would this be extended to multi-user / dynamic serving scenarios?\n* A.3.3 Table 13 seems to suggest that there is only 53% improvement for even double VRAM. Could you elaborate. It would also help if utilization data are provided?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761975451214}, {"id": "BCaLoAIIoG", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15033/Reviewer_2S5g"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper addresses the inefficiency of LLM inference on single-GPU or resource-limited devices, where offloading model weights from GPU to CPU memory causes severe GPU under-utilization and minimal performance scaling with additional GPU memory. The authors propose SpecOffload, a system that integrates speculative decoding directly into the offloading pipeline to reclaim idle GPU compute and unused GPU memory.\nThe key idea is to use a lightweight draft model that runs during GPU idle periods (while waiting for CPU–GPU data transfer) and to store it in “low-yield” GPU memory that would otherwise be wasted. SpecOffload introduces three core techniques: (1) a dual-batch interleaved pipeline allowing draft and target models to run concurrently; (2) adaptive tensor placement across GPU/CPU/disk memory tiers; and (3) a ParaSpec planner that automatically tunes batch and token parameters under GPU memory constraints.\nExperiments on Mixtral and LLaMA models show up to 4.5× GPU utilization and 2.36× throughput gains over strong offloading baselines (FlexGen, DeepSpeed-Inference, Fiddler), demonstrating strong practicality on consumer GPUs.", "review_text": "This paper addresses the inefficiency of LLM inference on single-GPU or resource-limited devices, where offloading model weights from GPU to CPU memory causes severe GPU under-utilization and minimal performance scaling with additional GPU memory. The authors propose SpecOffload, a system that integrates speculative decoding directly into the offloading pipeline to reclaim idle GPU compute and unused GPU memory.\nThe key idea is to use a lightweight draft model that runs during GPU idle periods (while waiting for CPU–GPU data transfer) and to store it in “low-yield” GPU memory that would otherwise be wasted. SpecOffload introduces three core techniques: (1) a dual-batch interleaved pipeline allowing draft and target models to run concurrently; (2) adaptive tensor placement across GPU/CPU/disk memory tiers; and (3) a ParaSpec planner that automatically tunes batch and token parameters under GPU memory constraints.\nExperiments on Mixtral and LLaMA models show up to 4.5× GPU utilization and 2.36× throughput gains over strong offloading baselines (FlexGen, DeepSpeed-Inference, Fiddler), demonstrating strong practicality on consumer GPUs.", "strengths": "1. The paper identifies an under-explored inefficiency — GPU idleness during offloading — and proposes to exploit this latent capacity through speculative decoding.\n2. The dual-batch interleaved execution and adaptive tensor placement form a coherent system that effectively overlaps CPU compute, GPU compute, and I/O, where speculative decoding is repurposed not merely as a speed-up technique but as a way to hide I/O latency, which is novel in the offloading context.\n3. The paper is well-written for a systems paper.", "weaknesses": "1. The speculative decoding itself and its integration into offloading is not unique. That being said, the novelty lies primarily in system integration and scheduling.\n2. All experiments are single-GPU; no exploration of distributed or multi-GPU extension.", "questions": "1. The ParaSpec Planner seems tuned for specific hardware (RTX 4090). Could the authors elaborate on how much manual profiling or calibration is needed to port SpecOffload to a different GPU architecture.\n2. In practice, how should users choose the draft model size relative to the target model? \u001dNowadays many choose to use Eagle to reduce the spec execution time but it might not be too much helpful for an offloading system like SpecOffload. How should SpecOffload adapt in terms of scheduling policy?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the inefficiency of LLM inference on single-GPU or resource-limited devices, where offloading model weights from GPU to CPU memory causes severe GPU under-utilization and minimal performance scaling with additional GPU memory. The authors propose SpecOffload, a system that integrates speculative decoding directly into the offloading pipeline to reclaim idle GPU compute and unused GPU memory.\nThe key idea is to use a lightweight draft model that runs during GPU idle periods (while waiting for CPU–GPU data transfer) and to store it in “low-yield” GPU memory that would otherwise be wasted. SpecOffload introduces three core techniques: (1) a dual-batch interleaved pipeline allowing draft and target models to run concurrently; (2) adaptive tensor placement across GPU/CPU/disk memory tiers; and (3) a ParaSpec planner that automatically tunes batch and token parameters under GPU memory constraints.\nExperiments on Mixtral and LLaMA models show up to 4.5× GPU utilization and 2.36× throughput gains over strong offloading baselines (FlexGen, DeepSpeed-Inference, Fiddler), demonstrating strong practicality on consumer GPUs.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The paper identifies an under-explored inefficiency — GPU idleness during offloading — and proposes to exploit this latent capacity through speculative decoding.\n2. The dual-batch interleaved execution and adaptive tensor placement form a coherent system that effectively overlaps CPU compute, GPU compute, and I/O, where speculative decoding is repurposed not merely as a speed-up technique but as a way to hide I/O latency, which is novel in the offloading context.\n3. The paper is well-written for a systems paper.", "weaknesses": "1. The speculative decoding itself and its integration into offloading is not unique. That being said, the novelty lies primarily in system integration and scheduling.\n2. All experiments are single-GPU; no exploration of distributed or multi-GPU extension.", "questions": "1. The ParaSpec Planner seems tuned for specific hardware (RTX 4090). Could the authors elaborate on how much manual profiling or calibration is needed to port SpecOffload to a different GPU architecture.\n2. In practice, how should users choose the draft model size relative to the target model? \u001dNowadays many choose to use Eagle to reduce the spec execution time but it might not be too much helpful for an offloading system like SpecOffload. How should SpecOffload adapt in terms of scheduling policy?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761926979528}, {"id": "N5aw0Ktz2r", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission15033/Reviewer_F4bn"], "rating": 4, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes SpecOffload for memory-restrained LLM inference scenario, where the large model has to be offloaded to CPU or disk for restrained GPU memory. It introduces speculative decoding during offloading where GPU idles, which involves 3 main modifications: batch interleave of drafting and verificaition in decoding phase, adaptive tensor planner to optimize offloading i/o latency, and ParaSpec planner to maximize throughput with given configurations. It also utilizes CPU for attention computation of the base model for more efficient pipelining. Experiments show that SpecOffload improves GPU core utilization by 4.49× and boosts inference throughput by 2.36×.", "review_text": "This paper proposes SpecOffload for memory-restrained LLM inference scenario, where the large model has to be offloaded to CPU or disk for restrained GPU memory. It introduces speculative decoding during offloading where GPU idles, which involves 3 main modifications: batch interleave of drafting and verificaition in decoding phase, adaptive tensor planner to optimize offloading i/o latency, and ParaSpec planner to maximize throughput with given configurations. It also utilizes CPU for attention computation of the base model for more efficient pipelining. Experiments show that SpecOffload improves GPU core utilization by 4.49× and boosts inference throughput by 2.36×.", "strengths": "The topic is highly related to practical inference scenario. LLM inference on resource-restrained hardware is an important direction.\n\nThe observation of ‘Marginal utility of GPU memory’ in the introduction is inspiring: When model size far exceeds GPU memory, if a portion of memory is further ‘restricted’ from being used, the marginal throughput degradation will be marginal because there have been too much offloading operations. This guarantees the effectiveness of inserting speculation into offloading stage, as the further occupied memory causes marginal degradation to the verification but enables speculation at GPU and hence boosts overall performances.\n\nThe demonstration figures (Fig.3, 4) are clear and informative, making the main design easily understood.", "weaknesses": "My main concern is about attention and FFN pipeline of the base model in sec4.1.2: \n\n(1)\tThe paper said that, the attention computation is on CPU and FFN is on GPU, and the FFN is offloaded to GPU while attention being computed, forming a pipeline. However, the computation of attention and FFN has interleaved data dependencies: Attention Layer i+1 needs the result of FFN Layer i, and FFN Layer i+1 needs the result of Attention Layer i+1, etc.. Therefore, in my opinion, it is impossible to first run all the attention layers on CPU, then all FFN layers on GPU. \n\n(2)\tThe attention computation on CPU is parallelized with the drafting on GPU. However, running attention computation of a much larger model on CPU would be much slower than running a full smaller model on GPU (in principle), so I think there would be a huge pipeline bubble bottlenecked by the CPU computation. This could limit the effectiveness of the proposed interleaved pipeline.\n\nOther concerns:\n\nThe compared baselines are weak. SpecOffload shows significant acceleration compared to all the mentioned baselines, while they all involve no speculative decoding, only offloading optimizations. As a result, the speedup ratios may be attributed to adding speculative decoding, rather than the effectiveness of the proposed method itself.\n\nThe contribution is relatively incremental. The contribution is mainly an engineering integration of speculative decoding into offloading pipelines. It builds upon well-established components (speculative decoding, offload scheduling, pipeline batch scheduling), making the novelty relatively incremental.", "questions": "1. Could the authors provide detailed computation workflow, to clarify how the attention and FFN layers are parallelized?\n2. How do the authors mitigate the potential pipeline bubble caused by performing attention computation on the CPU while the draft model runs on the GPU? Could they provide more detailed quantitative profiling data of each stage, to show the overall pipeline efficiency?\n3. Since all compared baselines lack speculative decoding, how can we disentangle the gains from speculative decoding itself versus the proposed integration?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes SpecOffload for memory-restrained LLM inference scenario, where the large model has to be offloaded to CPU or disk for restrained GPU memory. It introduces speculative decoding during offloading where GPU idles, which involves 3 main modifications: batch interleave of drafting and verificaition in decoding phase, adaptive tensor planner to optimize offloading i/o latency, and ParaSpec planner to maximize throughput with given configurations. It also utilizes CPU for attention computation of the base model for more efficient pipelining. Experiments show that SpecOffload improves GPU core utilization by 4.49× and boosts inference throughput by 2.36×.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The topic is highly related to practical inference scenario. LLM inference on resource-restrained hardware is an important direction.\n\nThe observation of ‘Marginal utility of GPU memory’ in the introduction is inspiring: When model size far exceeds GPU memory, if a portion of memory is further ‘restricted’ from being used, the marginal throughput degradation will be marginal because there have been too much offloading operations. This guarantees the effectiveness of inserting speculation into offloading stage, as the further occupied memory causes marginal degradation to the verification but enables speculation at GPU and hence boosts overall performances.\n\nThe demonstration figures (Fig.3, 4) are clear and informative, making the main design easily understood.", "weaknesses": "My main concern is about attention and FFN pipeline of the base model in sec4.1.2: \n\n(1)\tThe paper said that, the attention computation is on CPU and FFN is on GPU, and the FFN is offloaded to GPU while attention being computed, forming a pipeline. However, the computation of attention and FFN has interleaved data dependencies: Attention Layer i+1 needs the result of FFN Layer i, and FFN Layer i+1 needs the result of Attention Layer i+1, etc.. Therefore, in my opinion, it is impossible to first run all the attention layers on CPU, then all FFN layers on GPU. \n\n(2)\tThe attention computation on CPU is parallelized with the drafting on GPU. However, running attention computation of a much larger model on CPU would be much slower than running a full smaller model on GPU (in principle), so I think there would be a huge pipeline bubble bottlenecked by the CPU computation. This could limit the effectiveness of the proposed interleaved pipeline.\n\nOther concerns:\n\nThe compared baselines are weak. SpecOffload shows significant acceleration compared to all the mentioned baselines, while they all involve no speculative decoding, only offloading optimizations. As a result, the speedup ratios may be attributed to adding speculative decoding, rather than the effectiveness of the proposed method itself.\n\nThe contribution is relatively incremental. The contribution is mainly an engineering integration of speculative decoding into offloading pipelines. It builds upon well-established components (speculative decoding, offload scheduling, pipeline batch scheduling), making the novelty relatively incremental.", "questions": "1. Could the authors provide detailed computation workflow, to clarify how the attention and FFN layers are parallelized?\n2. How do the authors mitigate the potential pipeline bubble caused by performing attention computation on the CPU while the draft model runs on the GPU? Could they provide more detailed quantitative profiling data of each stage, to show the overall pipeline efficiency?\n3. Since all compared baselines lack speculative decoding, how can we disentangle the gains from speculative decoding itself versus the proposed integration?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761464429907}], "openreview_url": "https://openreview.net/forum?id=rVrWNb2XLi", "arxiv_id": "2505.10259", "paper_pdf": "papers/rVrWNb2XLi.pdf", "paper_pdf_sha256": "6f7983de61f4ab2a0458a44f9bfd9e608ef149560697e72ae2e5746c6facd019", "paper_pdf_bytes": 775442, "paper_pdf_source": "openreview", "code_url": "https://github.com/MobiSense/SpecOffload-public", "code_repository": "MobiSense/SpecOffload-public", "code_commit": "b8b68c9f4278fd9ef9b585b3e48ede301747a212", "code_archive": "repos/rVrWNb2XLi.zip", "code_archive_sha256": "7392869efc35f918254a472b51466eed964b47c215476934254c16c8aa0e8d4a", "code_archive_bytes": 11599378, "code_file_count": 1872, "code_extensions": {".py": 1834, ".h": 13, ".cu": 8, ".cpp": 7, ".sh": 6, ".cuh": 4}, "github_disk_usage_kb": 9693, "github_languages": {"Python": 44652281, "Cuda": 327808, "C++": 25815, "Shell": 18501, "C": 7703, "Cython": 3635}, "github_archived": false, "github_pushed_at": "2026-02-03T02:04:10Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/specoffload-unlocking-latent-gpu-capacity-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ln2k0PqVQA", "year": 2025, "status": "rejected", "title": "Online Intrinsic Rewards for Decision Making Agents from Large Language Model Feedback", "authors": ["Qinqing Zheng", "Mikael Henaff", "Amy Zhang", "Aditya Grover", "Brandon Amos"], "authorids": ["~Qinqing_Zheng1", "~Mikael_Henaff1", "~Amy_Zhang1", "~Aditya_Grover1", "~Brandon_Amos1"], "authors_source": "OpenReview API", "abstract": "Automatically synthesizing dense rewards from natural language descriptions is a promising paradigm in reinforcement learning (RL), with applications to sparse reward problems, open-ended exploration, and hierarchical skill design. Recent works have made promising steps by exploiting the prior knowledge of large language models (LLMs). However, these approaches suffer from important limitations: they are either not scalable to problems requiring billions of environment samples, due to requiring LLM annotations for each observation, or they require a diverse offline dataset, which may not exist or be impossible to collect. In this work, we address these limitations through a combination of algorithmic and systems-level contributions. We propose ONI, a distributed architecture that simultaneously learns an RL policy and an intrinsic reward function using LLM feedback. Our approach annotates the agent's collected experience via an asynchronous LLM server,  which is then distilled into an intrinsic reward model. We explore a range of algorithmic choices for reward modeling with varying complexity, including hashing, classification, and ranking models. By studying their relative tradeoffs, we shed light on questions regarding intrinsic reward design for sparse reward problems. Our approach achieves state-of-the-art performance across a range of challenging, sparse reward tasks from the NetHack Learning Environment in a simple unified process, solely using the agent's gathered experience, without requiring external datasets.\nWe make our code available at \\url{URL} (coming soon).", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "tEAcS0rGZ6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5187/Reviewer_4gbj"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This work presents a distributed system and an online learning algorithm for developing a reinforcement learning policy that learns from an intrinsic reward function, which is trained/obtained using feedback from LLMs. The approach depends solely on online experience, removing the need for external datasets or source code, and is assessed in the NetHack environment, comparing its performance to that of the Motif algorithm.", "review_text": "This work presents a distributed system and an online learning algorithm for developing a reinforcement learning policy that learns from an intrinsic reward function, which is trained/obtained using feedback from LLMs. The approach depends solely on online experience, removing the need for external datasets or source code, and is assessed in the NetHack environment, comparing its performance to that of the Motif algorithm.", "strengths": "1. The proposed research problem—scaling up LLMs to generate dense intrinsic rewards for online learning—is compelling and significant for advancing LLM agents' ability to tackle complex, long-horizon tasks in real-world applications without assumption of access to expert demonstration.\n\n2. The paper demonstrates strong engineering effort, employing online learning with help of LLMs to train an agent with up to 2 billion environment steps, a notable advance compared to prior work that required either only a few iteration of online training or a large offline dataset.\n\n3. The experimental design effectively supports the method’s claims. The chosen testbed is well-suited, as its long-horizon nature necessitates LLM-generated intrinsic rewards to achieve successful outcomes.", "weaknesses": "1. (main) The authors compare their work only with the Motif baseline algorithm. It would be beneficial to include comparisons with goal-conditioned reward design approaches as well. In the related work section, specifically in the goal-conditioned reward design subsection, the authors list relevant studies but do not critique their limitations or highlight the advantages of their approach. Adding these comment, as done in other subsections, would strengthen the discussion.\n\n2. In Figure 5.4, it is surprising that a higher number of annotations does not improve performance. The author might conduct an ablation study to assess the relationship between annotation volume and performance—reducing annotated data until a performance drop is observed. This may reveal that fewer labeled data points are sufficient for training the reward function, potentially reducing computational resources.\n\n3. Using HTTP for communication between the LLM server and the annotation process may incur high communication costs. Although the authors claim minimal overhead in line 189, an experiment to demonstrate this would provide clearer evidence.", "questions": "I have outlined all my concerns in the weaknesses section. I would be happy to discuss further if the authors can address my questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a distributed system and an online learning algorithm for developing a reinforcement learning policy that learns from an intrinsic reward function, which is trained/obtained using feedback from LLMs. The approach depends solely on online experience, removing the need for external datasets or source code, and is assessed in the NetHack environment, comparing its performance to that of the Motif algorithm.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "1. The proposed research problem—scaling up LLMs to generate dense intrinsic rewards for online learning—is compelling and significant for advancing LLM agents' ability to tackle complex, long-horizon tasks in real-world applications without assumption of access to expert demonstration.\n\n2. The paper demonstrates strong engineering effort, employing online learning with help of LLMs to train an agent with up to 2 billion environment steps, a notable advance compared to prior work that required either only a few iteration of online training or a large offline dataset.\n\n3. The experimental design effectively supports the method’s claims. The chosen testbed is well-suited, as its long-horizon nature necessitates LLM-generated intrinsic rewards to achieve successful outcomes.", "weaknesses": "1. (main) The authors compare their work only with the Motif baseline algorithm. It would be beneficial to include comparisons with goal-conditioned reward design approaches as well. In the related work section, specifically in the goal-conditioned reward design subsection, the authors list relevant studies but do not critique their limitations or highlight the advantages of their approach. Adding these comment, as done in other subsections, would strengthen the discussion.\n\n2. In Figure 5.4, it is surprising that a higher number of annotations does not improve performance. The author might conduct an ablation study to assess the relationship between annotation volume and performance—reducing annotated data until a performance drop is observed. This may reveal that fewer labeled data points are sufficient for training the reward function, potentially reducing computational resources.\n\n3. Using HTTP for communication between the LLM server and the annotation process may incur high communication costs. Although the authors claim minimal overhead in line 189, an experiment to demonstrate this would provide clearer evidence.", "questions": "I have outlined all my concerns in the weaknesses section. I would be happy to discuss further if the authors can address my questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730677091451}, {"id": "odEfJ5dg69", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5187/Reviewer_9m6i"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper proposes a method to simultaneously learn an RL policy and intrinsic reward function using feedback from LLMs, which annotate the agent’s experiences. The method is shown to outperform competing approaches in the NetHack environment.", "review_text": "The paper proposes a method to simultaneously learn an RL policy and intrinsic reward function using feedback from LLMs, which annotate the agent’s experiences. The method is shown to outperform competing approaches in the NetHack environment.", "strengths": "ONI seems to be a novel approach with promising results for learning intrinsic rewards. Leveraging LLMs to learn or infer reward functions is an interesting and relevant contemporary topic.", "weaknesses": "In general, the writing could be improved for clarity. There are some sections with unnecessarily long sentences, which impeded understanding. The work could be scoped out better to clarify exactly what the contributions are.", "questions": "1.\tWhat is the impact of limited annotation on the reward model? Also what is the impact of the quality of annotations?\n2.\tThe lack of difference in performance between ranking and classification methods could be investigated further. Analyzing the distribution of intrinsic rewards generated by each method or comparing their performance on tasks with higher observation diversity could offer further insights. Have the authors considered this?\n3.\tIt would be good to scope out clearly what the method is trying to achieve. For example, is the intrinsic reward used during learning fully learnt or is it just a current estimate?\n4.\tIs there a difference in the final reward model learnt via ONI and one learnt using say, a good amount of offline data? Does the reward from ONI converge? I wonder whether theoretical properties relating to this have been studied.\n5.\tThe logic behind eq 6 should be clearly and explicitly mentioned immediately after the equation.\n6.\tAlthough it may not be directly relevant, it would be reassuring to include additional baselines (or modified versions of it) based on the reported related work.\n7.\tWhat does ONI stand for?\n8.\tThe paper has some unusually long sentences (such as in the abstract), which should be split up for simplicity and clarity.\n9.\tIn section 3, it would have been better to focus on the proposed approach rather than elaborating on the details in say, lines 142-148 for example. These details are not central to the understanding of ONI and could have been pushed to the appendix.\n10.\tURL in the abstract does not work\n11.\t$\\eta$ is not rendered properly in Fig 5.2", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a method to simultaneously learn an RL policy and intrinsic reward function using feedback from LLMs, which annotate the agent’s experiences. The method is shown to outperform competing approaches in the NetHack environment.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "ONI seems to be a novel approach with promising results for learning intrinsic rewards. Leveraging LLMs to learn or infer reward functions is an interesting and relevant contemporary topic.", "weaknesses": "In general, the writing could be improved for clarity. There are some sections with unnecessarily long sentences, which impeded understanding. The work could be scoped out better to clarify exactly what the contributions are.", "questions": "1.\tWhat is the impact of limited annotation on the reward model? Also what is the impact of the quality of annotations?\n2.\tThe lack of difference in performance between ranking and classification methods could be investigated further. Analyzing the distribution of intrinsic rewards generated by each method or comparing their performance on tasks with higher observation diversity could offer further insights. Have the authors considered this?\n3.\tIt would be good to scope out clearly what the method is trying to achieve. For example, is the intrinsic reward used during learning fully learnt or is it just a current estimate?\n4.\tIs there a difference in the final reward model learnt via ONI and one learnt using say, a good amount of offline data? Does the reward from ONI converge? I wonder whether theoretical properties relating to this have been studied.\n5.\tThe logic behind eq 6 should be clearly and explicitly mentioned immediately after the equation.\n6.\tAlthough it may not be directly relevant, it would be reassuring to include additional baselines (or modified versions of it) based on the reported related work.\n7.\tWhat does ONI stand for?\n8.\tThe paper has some unusually long sentences (such as in the abstract), which should be split up for simplicity and clarity.\n9.\tIn section 3, it would have been better to focus on the proposed approach rather than elaborating on the details in say, lines 142-148 for example. These details are not central to the understanding of ONI and could have been pushed to the appendix.\n10.\tURL in the abstract does not work\n11.\t$\\eta$ is not rendered properly in Fig 5.2", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730638551057}, {"id": "b26wbEgi72", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5187/Reviewer_Cv7N"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 4, "summary": "This paper introduces ONI, a distributed architecture that simultaneously learns a RL policy and an intrinsic reward function using LLM feedback. ONI is built upon Sample Factory, an asynchronous reinforcement learning framework, and incorporates an LLM server hosted on a separate node to annotate observations, facilitating the learning of an intrinsic reward model that improves the sample efficiency of RL algorithms in sparse reward environments. This paper evaluates several designs of the intrinsic reward function and demonstrates ONI’s strong performance in the NetHack Learning Environment.", "review_text": "This paper introduces ONI, a distributed architecture that simultaneously learns a RL policy and an intrinsic reward function using LLM feedback. ONI is built upon Sample Factory, an asynchronous reinforcement learning framework, and incorporates an LLM server hosted on a separate node to annotate observations, facilitating the learning of an intrinsic reward model that improves the sample efficiency of RL algorithms in sparse reward environments. This paper evaluates several designs of the intrinsic reward function and demonstrates ONI’s strong performance in the NetHack Learning Environment.", "strengths": "- The paper is well-written, and the method is clearly introduced.\n- The limitations of previous approaches are effectively summarized, and the problem investigated is significant.", "weaknesses": "- My primary concern lies in the novelty of the contribution. The proposed method, ONI, appears heavily intertwined with Sample Factory[1], adding only a minor extension. The distributed architecture should, in my view, be credited to Sample Factory rather than this work. Furthermore, both the LLM-based annotation and the intrinsic reward design have been previously introduced in existing research[2,3]. As a result, I believe this paper may not be suitable for acceptance at top-tier venues.\n- Can the assumption of access to observation captions be relaxed, perhaps by leveraging a multi-modal LLM?\n- The experimental domain is limited to the NetHack Learning Environment. Could the method be extended to other widely-used RL benchmarks, such as MuJoCo[4] or Atari[5]?\n- The authors noted that increasing LLM annotations did not significantly improve performance. Could computational resources be optimized by reducing the number of annotations? What is the minimum level of annotation needed to maintain performance?\n- The discussion on the method's limitations remains insufficient.\n- The code is not available.\n\n[1] Aleksei Petrenko, Zhehui Huang, Tushar Kumar, Gaurav S. Sukhatme, and Vladlen Koltun. Sample factory: Egocentric 3d control from pixels at 100000 FPS with asynchronous reinforcement learning. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pp. 7652–7662. PMLR, 2020. URL http://proceedings.mlr.press/v119/ petrenko20a.html.\n\n[2] Lee H, Phatale S, Mansoor H, et al. RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback[C]//Forty-first International Conference on Machine Learning.\n\n[3] Martin Klissarov, Pierluca D’Oro, Shagun Sodhani, Roberta Raileanu, Pierre-Luc Bacon, Pascal Vincent, Amy Zhang, and Mikael Henaff. Motif: Intrinsic motivation from artificial intelligence feedback. arXiv preprint arXiv:2310.00166, 9 2023.\n\n[4] Todorov E, Erez T, Tassa Y. Mujoco: A physics engine for model-based control[C]//2012 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2012: 5026-5033.\n\n[5] Bellemare M G, Naddaf Y, Veness J, et al. The arcade learning environment: An evaluation platform for general agents[J]. Journal of Artificial Intelligence Research, 2013, 47: 253-279.", "questions": "Please take a look at the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces ONI, a distributed architecture that simultaneously learns a RL policy and an intrinsic reward function using LLM feedback. ONI is built upon Sample Factory, an asynchronous reinforcement learning framework, and incorporates an LLM server hosted on a separate node to annotate observations, facilitating the learning of an intrinsic reward model that improves the sample efficiency of RL algorithms in sparse reward environments. This paper evaluates several designs of the intrinsic reward function and demonstrates ONI’s strong performance in the NetHack Learning Environment.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "- The paper is well-written, and the method is clearly introduced.\n- The limitations of previous approaches are effectively summarized, and the problem investigated is significant.", "weaknesses": "- My primary concern lies in the novelty of the contribution. The proposed method, ONI, appears heavily intertwined with Sample Factory[1], adding only a minor extension. The distributed architecture should, in my view, be credited to Sample Factory rather than this work. Furthermore, both the LLM-based annotation and the intrinsic reward design have been previously introduced in existing research[2,3]. As a result, I believe this paper may not be suitable for acceptance at top-tier venues.\n- Can the assumption of access to observation captions be relaxed, perhaps by leveraging a multi-modal LLM?\n- The experimental domain is limited to the NetHack Learning Environment. Could the method be extended to other widely-used RL benchmarks, such as MuJoCo[4] or Atari[5]?\n- The authors noted that increasing LLM annotations did not significantly improve performance. Could computational resources be optimized by reducing the number of annotations? What is the minimum level of annotation needed to maintain performance?\n- The discussion on the method's limitations remains insufficient.\n- The code is not available.\n\n[1] Aleksei Petrenko, Zhehui Huang, Tushar Kumar, Gaurav S. Sukhatme, and Vladlen Koltun. Sample factory: Egocentric 3d control from pixels at 100000 FPS with asynchronous reinforcement learning. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pp. 7652–7662. PMLR, 2020. URL http://proceedings.mlr.press/v119/ petrenko20a.html.\n\n[2] Lee H, Phatale S, Mansoor H, et al. RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback[C]//Forty-first International Conference on Machine Learning.\n\n[3] Martin Klissarov, Pierluca D’Oro, Shagun Sodhani, Roberta Raileanu, Pierre-Luc Bacon, Pascal Vincent, Amy Zhang, and Mikael Henaff. Motif: Intrinsic motivation from artificial intelligence feedback. arXiv preprint arXiv:2310.00166, 9 2023.\n\n[4] Todorov E, Erez T, Tassa Y. Mujoco: A physics engine for model-based control[C]//2012 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2012: 5026-5033.\n\n[5] Bellemare M G, Naddaf Y, Veness J, et al. The arcade learning environment: An evaluation platform for general agents[J]. Journal of Artificial Intelligence Research, 2013, 47: 253-279.", "questions": "Please take a look at the Weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730622769415}, {"id": "cnez9pmlOu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5187/Reviewer_C2KC"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "ONI introduces a novel distributed framework for developing intrinsic rewards in reinforcement learning by leveraging feedback from large language models (LLMs). This approach supports policy optimization in sparse reward scenarios without needing external datasets. By exploring several methods for reward modeling, ONI demonstrates impressive performance in the NetHack environment, closely matching the results of previous approaches like Motif while eliminating the requirement for pre-collected data", "review_text": "ONI introduces a novel distributed framework for developing intrinsic rewards in reinforcement learning by leveraging feedback from large language models (LLMs). This approach supports policy optimization in sparse reward scenarios without needing external datasets. By exploring several methods for reward modeling, ONI demonstrates impressive performance in the NetHack environment, closely matching the results of previous approaches like Motif while eliminating the requirement for pre-collected data", "strengths": "- ONI provides a practical and innovative way to use feedback from large language models as intrinsic rewards, making it especially valuable for reinforcement learning environments where extrinsic rewards are sparse.\n\n- By eliminating the need for external datasets and instead building rewards from the agent’s own experience, ONI offers a scalable approach that could adapt to a wide range of tasks.\n\n- The asynchronous design keeps the policy training process uninterrupted, maintaining high throughput and efficiency throughout.\n\n- ONI shows strong, competitive performance in challenging tasks within the NetHack environment, matching or even surpassing state-of-the-art methods in tasks like Oracle and multi-level exploration without relying on dense reward functions.\n\n- Enough details are provided to reproduce the method. The method reproducibility appears to be good.", "weaknesses": "- The method is primarily evaluated in the NetHack environment, leaving questions on how well ONI would generalize to other, potentially more diverse RL environments. It would be critical to show that the method can be applied to other environments and domains. In the current state, evaluation is too narrow to demonstrate the impact of this work.\n\n- Asynchronous LLM feedback, while effective here, may face scalability challenges in more complex settings where feedback needs scale up significantly. The paper lacks detailed discussion on this issue. \n\n- The paper should discuss and report evaluation with different LLM. In the current state, the results cannot demonstrate that the method can be used with other LLMs. \n\n- LLM-generated annotations may carry inherent biases that can influence reward generation. The paper lacks a strategy for identifying or mitigating such biases.\n\n- Comparisons with other intrinsic reward models and exploration techniques are minimal, limiting clarity on ONI's unique advantages or limitations relative to a wider range of baselines. I would suggest for instance comparing current paper with strong LLM-guided baselines, such as [“Guiding Pretraining in Reinforcement Learning with Large Language Models”, Du et al., 2023], [“Exploring Beyond Curiosity Rewards: Language-Driven Exploration in RL”, Bougie et al., 2024], or other recent papers.", "questions": "- How would ONI perform in other environments beyond NetHack, especially ones with more complex or variable observation spaces?\n\n- Are there any insights into how specific LLM choices impact ONI's performance, especially for larger models?\n\n- Could potential biases in LLM annotations affect the intrinsic reward generation process, and if so, how are they mitigated?\n\n- How does the method compare with stronger recent baselines that guide exploration with LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "ONI introduces a novel distributed framework for developing intrinsic rewards in reinforcement learning by leveraging feedback from large language models (LLMs). This approach supports policy optimization in sparse reward scenarios without needing external datasets. By exploring several methods for reward modeling, ONI demonstrates impressive performance in the NetHack environment, closely matching the results of previous approaches like Motif while eliminating the requirement for pre-collected data", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- ONI provides a practical and innovative way to use feedback from large language models as intrinsic rewards, making it especially valuable for reinforcement learning environments where extrinsic rewards are sparse.\n\n- By eliminating the need for external datasets and instead building rewards from the agent’s own experience, ONI offers a scalable approach that could adapt to a wide range of tasks.\n\n- The asynchronous design keeps the policy training process uninterrupted, maintaining high throughput and efficiency throughout.\n\n- ONI shows strong, competitive performance in challenging tasks within the NetHack environment, matching or even surpassing state-of-the-art methods in tasks like Oracle and multi-level exploration without relying on dense reward functions.\n\n- Enough details are provided to reproduce the method. The method reproducibility appears to be good.", "weaknesses": "- The method is primarily evaluated in the NetHack environment, leaving questions on how well ONI would generalize to other, potentially more diverse RL environments. It would be critical to show that the method can be applied to other environments and domains. In the current state, evaluation is too narrow to demonstrate the impact of this work.\n\n- Asynchronous LLM feedback, while effective here, may face scalability challenges in more complex settings where feedback needs scale up significantly. The paper lacks detailed discussion on this issue. \n\n- The paper should discuss and report evaluation with different LLM. In the current state, the results cannot demonstrate that the method can be used with other LLMs. \n\n- LLM-generated annotations may carry inherent biases that can influence reward generation. The paper lacks a strategy for identifying or mitigating such biases.\n\n- Comparisons with other intrinsic reward models and exploration techniques are minimal, limiting clarity on ONI's unique advantages or limitations relative to a wider range of baselines. I would suggest for instance comparing current paper with strong LLM-guided baselines, such as [“Guiding Pretraining in Reinforcement Learning with Large Language Models”, Du et al., 2023], [“Exploring Beyond Curiosity Rewards: Language-Driven Exploration in RL”, Bougie et al., 2024], or other recent papers.", "questions": "- How would ONI perform in other environments beyond NetHack, especially ones with more complex or variable observation spaces?\n\n- Are there any insights into how specific LLM choices impact ONI's performance, especially for larger models?\n\n- Could potential biases in LLM annotations affect the intrinsic reward generation process, and if so, how are they mitigated?\n\n- How does the method compare with stronger recent baselines that guide exploration with LLMs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730418423636}], "openreview_url": "https://openreview.net/forum?id=ln2k0PqVQA", "arxiv_id": "2410.23022", "paper_pdf": "papers/ln2k0PqVQA.pdf", "paper_pdf_sha256": "71573254cebaf2c16434171e4fa902a3a50e4cb51a85bd32f1aa3d6969b08229", "paper_pdf_bytes": 19114039, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/oni", "code_repository": "facebookresearch/oni", "code_commit": "f0f86e8abea877e045e74f4bf786d246df0638a4", "code_archive": "repos/ln2k0PqVQA.zip", "code_archive_sha256": "526b5ec20098373b53045c5d381c356896654e9b0c09f708a00b8192ee5adf74", "code_archive_bytes": 463653, "code_file_count": 135, "code_extensions": {".py": 132, ".sh": 3}, "github_disk_usage_kb": 400, "github_languages": {"Python": 783791, "Shell": 2112}, "github_archived": true, "github_pushed_at": "2024-12-17T22:07:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/online-intrinsic-rewards-for-decision-making"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xbXASfz8MD", "year": 2024, "status": "rejected", "title": "Latent Space Symmetry Discovery", "authors": ["Jianke Yang", "Nima Dehmamy", "Robin Walters", "Rose Yu"], "authorids": ["~Jianke_Yang2", "~Nima_Dehmamy1", "~Robin_Walters1", "~Rose_Yu1"], "authors_source": "OpenReview API", "abstract": "Equivariant neural networks require explicit knowledge of the symmetry group. Automatic symmetry discovery methods aim to relax this constraint and learn invariance and equivariance from data. However, existing symmetry discovery methods are limited to linear symmetries in their search space and cannot handle the complexity of symmetries in real-world, often high-dimensional data. We propose a novel generative model, Latent LieGAN (LaLiGAN), which can discover nonlinear symmetries from data. It learns a mapping from data to a latent space where the symmetries become linear and simultaneously discovers symmetries in the latent space. Theoretically, we show that our method can express any nonlinear symmetry under certain conditions. Experimentally, our method can capture the intrinsic symmetry in high-dimensional observations, which results in a well-structured latent space that is useful for other downstream tasks. We demonstrate the use cases for LaLiGAN in improving equation discovery and long-term forecasting for various dynamical systems.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "KnakcOYDMD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission952/Reviewer_5mQG"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper considers the problem of symmetry estimation for the sake of better representation learning. The authors introduce Latent Lie GAN (LaLiGan) for learning non-linear symmetries in the input data. The paper highlights the fact that the problem has high importance for the field of representation learning. The authors demonstrate that there were many approaches for solving similar problems; however, the main focus was on linear symmetries, i.e., group representations. In contrast, the presented paper demonstrates that it is possible to learn non-linear symmetries in an adversarial manner.", "review_text": "The paper considers the problem of symmetry estimation for the sake of better representation learning. The authors introduce Latent Lie GAN (LaLiGan) for learning non-linear symmetries in the input data. The paper highlights the fact that the problem has high importance for the field of representation learning. The authors demonstrate that there were many approaches for solving similar problems; however, the main focus was on linear symmetries, i.e., group representations. In contrast, the presented paper demonstrates that it is possible to learn non-linear symmetries in an adversarial manner.", "strengths": "- The paper is well-written. The flow is smooth and coherent. The paper presents good illustrations to help the reader understand the presented idea.\n- The mathematical language is easy to follow, correct, and detailed when needed.\n- The authors highlight the main contributions of the paper clearly.\n- The presented method is clearly a next-step solution compared to approaches like LieGG or LieGAN.\n- The experiments demonstrate that the proposed method can be applied in a wide range of tasks.", "weaknesses": "These are not significant weknesses. I would like to highlight the fact, that from the paper it seems like there are no natural limitations to the proposed method, which is however, not true. A straighforward explanation of situations when the method fails or can lead to an incorrect outcome will help", "questions": "I would like the authors to answer the following questions to make it easier to understand certain aspects of the method\n- In Eq. 2 you learn transformations as $\\sum_i\\text{exp}[w_i L_i]$. How to choose the number of matrices L to be used in the method? I suppose the number of Lie algebra elements you parametrize will significantly affect the flexibility of the method in the latent space\n- If I understood correctly, the proposed method works for compact groups only. The experiments demonstrate that the method can learn trajectories that are isomorphic to circles. How will the method behave on the data which has translation symmetry only? Will it fail? If so, the set of admissible symmetries seems more limited and should be highlighted\n- in Proposition 4.1 you mention that $\\psi$ and $\\phi$ are inverse of each other. It is not correct, these functions are inverse to each other only on the input dataset. It raises the following question, how robust is the inverse property when you move away from the training dataset? How robust is the detected symmetry, when you move away from the training dataset? It reminds me of the following paper *Moskalev A. et al. On genuine invariance learning without weight-tying. Topological, Algebraic and Geometric Learning Workshops 2023. – PMLR, 2023*", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper considers the problem of symmetry estimation for the sake of better representation learning. The authors introduce Latent Lie GAN (LaLiGan) for learning non-linear symmetries in the input data. The paper highlights the fact that the problem has high importance for the field of representation learning. The authors demonstrate that there were many approaches for solving similar problems; however, the main focus was on linear symmetries, i.e., group representations. In contrast, the presented paper demonstrates that it is possible to learn non-linear symmetries in an adversarial manner.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "- The paper is well-written. The flow is smooth and coherent. The paper presents good illustrations to help the reader understand the presented idea.\n- The mathematical language is easy to follow, correct, and detailed when needed.\n- The authors highlight the main contributions of the paper clearly.\n- The presented method is clearly a next-step solution compared to approaches like LieGG or LieGAN.\n- The experiments demonstrate that the proposed method can be applied in a wide range of tasks.", "weaknesses": "These are not significant weknesses. I would like to highlight the fact, that from the paper it seems like there are no natural limitations to the proposed method, which is however, not true. A straighforward explanation of situations when the method fails or can lead to an incorrect outcome will help", "questions": "I would like the authors to answer the following questions to make it easier to understand certain aspects of the method\n- In Eq. 2 you learn transformations as $\\sum_i\\text{exp}[w_i L_i]$. How to choose the number of matrices L to be used in the method? I suppose the number of Lie algebra elements you parametrize will significantly affect the flexibility of the method in the latent space\n- If I understood correctly, the proposed method works for compact groups only. The experiments demonstrate that the method can learn trajectories that are isomorphic to circles. How will the method behave on the data which has translation symmetry only? Will it fail? If so, the set of admissible symmetries seems more limited and should be highlighted\n- in Proposition 4.1 you mention that $\\psi$ and $\\phi$ are inverse of each other. It is not correct, these functions are inverse to each other only on the input dataset. It raises the following question, how robust is the inverse property when you move away from the training dataset? How robust is the detected symmetry, when you move away from the training dataset? It reminds me of the following paper *Moskalev A. et al. On genuine invariance learning without weight-tying. Topological, Algebraic and Geometric Learning Workshops 2023. – PMLR, 2023*", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699023942823}, {"id": "7ECtU0F2Nd", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission952/Reviewer_Vcu8"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper studies the problem of automatically discovering Lie group symmetries. To do so the paper focuses on non-linear group actions and attempts to discover linear representations in a latent space of an autoencoder. The overall method is termed Latent LieGAN and comes with a theory that attempts to show that the learned symmetry group is actually valid. In practical tests, LaLiGAN was able to recognize the inherent symmetry in high-dimensional data, creating a structured space that can be used for other tasks. The paper also showcases how LaLiGAN can be used to enhance equation discovery and make long-term predictions for different dynamic systems.", "review_text": "This paper studies the problem of automatically discovering Lie group symmetries. To do so the paper focuses on non-linear group actions and attempts to discover linear representations in a latent space of an autoencoder. The overall method is termed Latent LieGAN and comes with a theory that attempts to show that the learned symmetry group is actually valid. In practical tests, LaLiGAN was able to recognize the inherent symmetry in high-dimensional data, creating a structured space that can be used for other tasks. The paper also showcases how LaLiGAN can be used to enhance equation discovery and make long-term predictions for different dynamic systems.", "strengths": "This paper studies an interesting problem, which started originally from the seminal work of Higgins et. al 2018. Since then, there has been a large body of work studying automatic symmetry discovery with various results. This paper adds to this body of work by using an adversarial approach which is easy to follow in this context. Unfortunately, I have a negative view of the originality and significance of this work as I outline in the next section but I will say the paper is generally well presented and shows a high degree of polish. The experimental results are quite toy and concocted but they make for good visuals and suggest there is merit in this approach to low dimensional problems.", "weaknesses": "I have several concerns regarding this paper to the point I am confused and question the validity of the entire endeavor. These might be my misunderstanding so I hope they can be clarified in the rebuttal. But as it stands I cannot endorse this paper for the following reasons.\n\n1.) There is a strong emphasis on the non-linear group action aspect in this paper, but I believe this is a bit misguided. This is because the hallmark result of representation theory of Lie groups is that the Lie algebra connects the group to the vector space. Moreover, this can be described by matrices---hence linear representation---and you do not generally need non-linear representations. In practice, however, you can codify non-linear actions (e.g. rotations of 3D objects in a 2D image) and this is where you might want to learn a non-linear action. But I find the emphasis on the non-linear action exaggerated because LaLieGan learns a linear rep in the latent space anyways. I would suggest toning down these claims.\n\n2.) Prop 4.1 seems to not apply to the setup that the authors consider. This is because the encoder and decoder map to a latent space of an autoencoder. This means that the latent dimension can be **lower** than the observation dimension. As a result, $\\phi$ and $\\psi$ cannot ever be inverses---i.e. bijective---because the information is lost. Thus, I have strong doubts about the value of the proposition. Moreover, many symmetry discovery methods already assume an autoencoder setup. The main difference is that they do not take an adversarial approach so this limits the novelty of the method. Finally, the paper learns approximate inverses anyways so there is no reason to guarantee that the learned representation is an exact Lie group.\n\n3.) One of my biggest concerns is that the approach and results in this paper go against a relatively known result in Linear Symmetry Based Disentanglement by (Caselles-Dupré et. al 2019) who prove that symmetry discovery is impossible without interaction with the environment. This result is a symmetry-based analog to the result by Locatello et. al 2019. Thus I fear that the results in this paper are generally not true, and going beyond the toy datasets considered here might be impossible.\n\n4.) I am confused as to why the authors do not compare with more standard baselines for symmetry discovery. Granted these works often assume knowledge of the group apriori but why this is not the correct test bed? For example $SO(N)$ is done in Fig 2 and 4 of Quessard et al 2020 (you even cite this paper) as well as the main experiment of Caselles-Dupré et. al 2019. Moreover, there has been a lot of development in Deep Delay auto-encoders that extend SINDy. In particular Deep Delay Autoencoders Bakkarji et. al 2023 is an appropriate baseline for the non-linear dynamical system discovery experiments. I encourage the authors to include this baseline as well.\n\n**References**\n\nCaselles-Dupré, Hugo, Michael Garcia Ortiz, and David Filliat. \"Symmetry-based disentangled representation learning requires interaction with environments.\" Advances in Neural Information Processing Systems 32 (2019).\n\nLocatello, Francesco, et al. \"Challenging common assumptions in the unsupervised learning of disentangled representations.\" international conference on machine learning. PMLR, 2019.\n\nQuessard, Robin, Thomas Barrett, and William Clements. \"Learning disentangled representations and group structure of dynamical environments.\" Advances in Neural Information Processing Systems 33 (2020): 19727-19737.\n\nBakarji, Joseph, et al. \"Discovering Governing Equations from Partial Measurements with Deep Delay Autoencoders. arXiv.\" arXiv preprint arXiv:2201.05136 (2022).", "questions": "See my questions in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the problem of automatically discovering Lie group symmetries. To do so the paper focuses on non-linear group actions and attempts to discover linear representations in a latent space of an autoencoder. The overall method is termed Latent LieGAN and comes with a theory that attempts to show that the learned symmetry group is actually valid. In practical tests, LaLiGAN was able to recognize the inherent symmetry in high-dimensional data, creating a structured space that can be used for other tasks. The paper also showcases how LaLiGAN can be used to enhance equation discovery and make long-term predictions for different dynamic systems.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "This paper studies an interesting problem, which started originally from the seminal work of Higgins et. al 2018. Since then, there has been a large body of work studying automatic symmetry discovery with various results. This paper adds to this body of work by using an adversarial approach which is easy to follow in this context. Unfortunately, I have a negative view of the originality and significance of this work as I outline in the next section but I will say the paper is generally well presented and shows a high degree of polish. The experimental results are quite toy and concocted but they make for good visuals and suggest there is merit in this approach to low dimensional problems.", "weaknesses": "I have several concerns regarding this paper to the point I am confused and question the validity of the entire endeavor. These might be my misunderstanding so I hope they can be clarified in the rebuttal. But as it stands I cannot endorse this paper for the following reasons.\n\n1.) There is a strong emphasis on the non-linear group action aspect in this paper, but I believe this is a bit misguided. This is because the hallmark result of representation theory of Lie groups is that the Lie algebra connects the group to the vector space. Moreover, this can be described by matrices---hence linear representation---and you do not generally need non-linear representations. In practice, however, you can codify non-linear actions (e.g. rotations of 3D objects in a 2D image) and this is where you might want to learn a non-linear action. But I find the emphasis on the non-linear action exaggerated because LaLieGan learns a linear rep in the latent space anyways. I would suggest toning down these claims.\n\n2.) Prop 4.1 seems to not apply to the setup that the authors consider. This is because the encoder and decoder map to a latent space of an autoencoder. This means that the latent dimension can be **lower** than the observation dimension. As a result, $\\phi$ and $\\psi$ cannot ever be inverses---i.e. bijective---because the information is lost. Thus, I have strong doubts about the value of the proposition. Moreover, many symmetry discovery methods already assume an autoencoder setup. The main difference is that they do not take an adversarial approach so this limits the novelty of the method. Finally, the paper learns approximate inverses anyways so there is no reason to guarantee that the learned representation is an exact Lie group.\n\n3.) One of my biggest concerns is that the approach and results in this paper go against a relatively known result in Linear Symmetry Based Disentanglement by (Caselles-Dupré et. al 2019) who prove that symmetry discovery is impossible without interaction with the environment. This result is a symmetry-based analog to the result by Locatello et. al 2019. Thus I fear that the results in this paper are generally not true, and going beyond the toy datasets considered here might be impossible.\n\n4.) I am confused as to why the authors do not compare with more standard baselines for symmetry discovery. Granted these works often assume knowledge of the group apriori but why this is not the correct test bed? For example $SO(N)$ is done in Fig 2 and 4 of Quessard et al 2020 (you even cite this paper) as well as the main experiment of Caselles-Dupré et. al 2019. Moreover, there has been a lot of development in Deep Delay auto-encoders that extend SINDy. In particular Deep Delay Autoencoders Bakkarji et. al 2023 is an appropriate baseline for the non-linear dynamical system discovery experiments. I encourage the authors to include this baseline as well.\n\n**References**\n\nCaselles-Dupré, Hugo, Michael Garcia Ortiz, and David Filliat. \"Symmetry-based disentangled representation learning requires interaction with environments.\" Advances in Neural Information Processing Systems 32 (2019).\n\nLocatello, Francesco, et al. \"Challenging common assumptions in the unsupervised learning of disentangled representations.\" international conference on machine learning. PMLR, 2019.\n\nQuessard, Robin, Thomas Barrett, and William Clements. \"Learning disentangled representations and group structure of dynamical environments.\" Advances in Neural Information Processing Systems 33 (2020): 19727-19737.\n\nBakarji, Joseph, et al. \"Discovering Governing Equations from Partial Measurements with Deep Delay Autoencoders. arXiv.\" arXiv preprint arXiv:2201.05136 (2022).", "questions": "See my questions in the weaknesses section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698805678874}, {"id": "MJ4HBM1J4M", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission952/Reviewer_BHEh"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper extends LieGAN, which learns linear symmetries within data, to learn non-linear symmetries by integrating an autoencoder. The discovered symmetries/group transformations operate in the learned latent space (instead of in the input space as LieGAN) so that to be nonlinear.\n\nConcretely, It decomposes the nonlinear group transformation into first encoding into the latent space, then linear transform, and lastly decoding back to the original space. It trains an autoencoder to ensure that the encoders and decoders are inverse of each other. It enforces the transformed data still to be in-distribution as the training dataset using a GAN loss.  \n\nThe method is shown to learn rotation symmetric latent space for several dynamic systems. The learned latent space is shown helpful for equation discovery in one domain. The discovered equation is simpler and achieves better long-term prediction accuracy.", "review_text": "This paper extends LieGAN, which learns linear symmetries within data, to learn non-linear symmetries by integrating an autoencoder. The discovered symmetries/group transformations operate in the learned latent space (instead of in the input space as LieGAN) so that to be nonlinear.\n\nConcretely, It decomposes the nonlinear group transformation into first encoding into the latent space, then linear transform, and lastly decoding back to the original space. It trains an autoencoder to ensure that the encoders and decoders are inverse of each other. It enforces the transformed data still to be in-distribution as the training dataset using a GAN loss.  \n\nThe method is shown to learn rotation symmetric latent space for several dynamic systems. The learned latent space is shown helpful for equation discovery in one domain. The discovered equation is simpler and achieves better long-term prediction accuracy.", "strengths": "This paper discusses an important problem: discovering symmetries given the dataset. It is generally well-written and easy to read. \n\nThe method is intuitive, extending LieGAN with the latent space learned by autoencoders. \n\nThe experimental results show promising results in several dynamic systems, including one with a high-dimensional observation space.", "weaknesses": "I am mainly concerned with the practical applicability of this method:\n* As discussed in the paper, the nonlinear-symmetric-discovery problem itself is ill-posed, and there are many meaningless \"optimal\" solutions to it due to the representation power of neural networks. This paper incorporates several patches to alleviate this issue, such as an orthogonal weight matrix in the final layer, and zero-mean of the latent features within an empirical batch. These regularization terms seem strong and hard-coded, and there is no theoretical understanding/analysis of them. \n  - Are there metrics distinguishing the qualities of learned symmetries other than human interpretation?\n  - Can the model, after applying all these regularization terms, learn all desired symmetries? \n  - Are these regularization terms enough to rule out all meaningless solutions?  \n* Similarly, all learned/discovered symmetries in the experiments are rotation-based. Why is that? Is it related to the choice of the regularization? Can the model learn other symmetries in practice? For example, can it learn a nonlinear version of E(n)?\n  - If the model can only learn rotation-based symmetries or rotation-based symmetries are enough with powerful neural encoders/decoders, why would we learn the symmetries in the latent space then? \n\nSome other weakness includes\n* The learned nonlinear symmetries are not that interpretable due to the neural encoder;\n* It would be great to show an application area of the learned symmetries more than just the learned latent space.", "questions": "* How difficult is it to learn a meaningful nonlinear symmetry using this method in practice? Are there results showing this method learned symmetries other than those rotation-based? Are there results in domains other than the synthetic dynamic systems? \n* Are there more ways to interpret/use the learned symmetries?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper extends LieGAN, which learns linear symmetries within data, to learn non-linear symmetries by integrating an autoencoder. The discovered symmetries/group transformations operate in the learned latent space (instead of in the input space as LieGAN) so that to be nonlinear.\n\nConcretely, It decomposes the nonlinear group transformation into first encoding into the latent space, then linear transform, and lastly decoding back to the original space. It trains an autoencoder to ensure that the encoders and decoders are inverse of each other. It enforces the transformed data still to be in-distribution as the training dataset using a GAN loss.  \n\nThe method is shown to learn rotation symmetric latent space for several dynamic systems. The learned latent space is shown helpful for equation discovery in one domain. The discovered equation is simpler and achieves better long-term prediction accuracy.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "This paper discusses an important problem: discovering symmetries given the dataset. It is generally well-written and easy to read. \n\nThe method is intuitive, extending LieGAN with the latent space learned by autoencoders. \n\nThe experimental results show promising results in several dynamic systems, including one with a high-dimensional observation space.", "weaknesses": "I am mainly concerned with the practical applicability of this method:\n* As discussed in the paper, the nonlinear-symmetric-discovery problem itself is ill-posed, and there are many meaningless \"optimal\" solutions to it due to the representation power of neural networks. This paper incorporates several patches to alleviate this issue, such as an orthogonal weight matrix in the final layer, and zero-mean of the latent features within an empirical batch. These regularization terms seem strong and hard-coded, and there is no theoretical understanding/analysis of them. \n  - Are there metrics distinguishing the qualities of learned symmetries other than human interpretation?\n  - Can the model, after applying all these regularization terms, learn all desired symmetries? \n  - Are these regularization terms enough to rule out all meaningless solutions?  \n* Similarly, all learned/discovered symmetries in the experiments are rotation-based. Why is that? Is it related to the choice of the regularization? Can the model learn other symmetries in practice? For example, can it learn a nonlinear version of E(n)?\n  - If the model can only learn rotation-based symmetries or rotation-based symmetries are enough with powerful neural encoders/decoders, why would we learn the symmetries in the latent space then? \n\nSome other weakness includes\n* The learned nonlinear symmetries are not that interpretable due to the neural encoder;\n* It would be great to show an application area of the learned symmetries more than just the learned latent space.", "questions": "* How difficult is it to learn a meaningful nonlinear symmetry using this method in practice? Are there results showing this method learned symmetries other than those rotation-based? Are there results in domains other than the synthetic dynamic systems? \n* Are there more ways to interpret/use the learned symmetries?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698711844805}], "openreview_url": "https://openreview.net/forum?id=xbXASfz8MD", "arxiv_id": "2310.00105", "paper_pdf": "papers/xbXASfz8MD.pdf", "paper_pdf_sha256": "bc47a63f3f0c138b41b433b22193b53cb15113b36d3d64755775fa1bccc7f09a", "paper_pdf_bytes": 4034124, "paper_pdf_source": "openreview", "code_url": "https://github.com/jiankeyang/LaLiGAN", "code_repository": "jiankeyang/LaLiGAN", "code_commit": "030457ae5b23555f175bf8422f8d3fb99942973e", "code_archive": "repos/xbXASfz8MD.zip", "code_archive_sha256": "035ee4299e271ef64edae1bc517d9180379fbd565c4983c6ab33371892534599", "code_archive_bytes": 385964, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 423, "github_languages": {"Python": 83625}, "github_archived": false, "github_pushed_at": "2024-07-12T23:25:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/latent-space-symmetry-discovery"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YP4QEmqh6Ia", "year": 2023, "status": "rejected", "title": "Which Invariance Should We Transfer? A Causal Minimax Learning Approach", "authors": ["Mingzhou Liu", "Xiangyu Zheng", "Xinwei Sun", "Fang Fang", "Yizhou Wang"], "authorids": ["~Mingzhou_Liu1", "~Xiangyu_Zheng1", "~Xinwei_Sun1", "~Fang_Fang1", "~Yizhou_Wang1"], "authors_source": "OpenReview API", "abstract": "A major barrier to deploy current machine learning models lies in their sensitivity to dataset shifts. To resolve this problem, most existing studies attempted to transfer stable information to unseen environments. Among these, graph-based methods causally decomposed the data generating process into stable and mutable mechanisms. By removing the effect of mutable generation, they identified a set of stable predictors. However, a key question regarding robustness remains: which subset of the whole stable information should the model transfer, in order to achieve optimal generalization ability? To answer this question, we provide a comprehensive minimax analysis that fully characterizes conditions for a subset to be optimal. Particularly in general cases, we propose to maximize over mutable mechanisms (i.e., the source of dataset shifts), which is provable to identify the worst-case risk over all environments. This ensures us to select the optimal subset with the minimal worst-case risk. To reduce computational costs, we propose to search over only equivalent classes in terms of worst-case risk, instead of over all subsets. In cases when the searching space is still large, we turn this subset selection problem into a sparse min-max optimization scheme, which enjoys the simplicity and efficiency of implementation. The utility of our methods is demonstrated on the diagnosis of Alzheimer's Disease and gene function prediction. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Xas6FrXJCxu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3458/Reviewer_ra9S"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "For a supervised learning task where data comes from multiple environments, invariant prediction relies on identifying a stable set of features that don't influence the shift in probability distributions over different training environments. This paper proposes a method to find the optimal sufficient set of stable features. The method is split into two cases depending on whether the optimal set of stable features is the entire set of stable features. The paper proposes a sufficient graphical condition for the latter that can be tested by causal discovery. If the condition fails, the paper parametrizes the dataset shift and does a brute-force search to get the min-max optimal subset. To reduce the search space, an equivalence class of the set of subsets is identified. While acknowledging that this brute force search might be inefficient, the authors frame an alternative sparse min-max optimization to solve the subset selection problem. The brute force search method is validated on synthetic data and both methods are validated on real world data.  ", "review_text": "Overall, the authors propose two methods for an important problem - one of which potentially has an exponential search complexity and it's not clear to me how they parametrize the mutable shifts. The other method while having decent empirical performance is not addressed much in the theory portion of the paper. If the authors can address either of the two concerns, it would make the paper much stronger. ", "strengths": "The paper’s strongest contribution is the empirical validation on the real world datasets considered albeit still being relatively low-dimensional. Although I didn’t check the proofs in detail, both Both the methods proposed seem theoretically sound and believable. The concept of equivalences is also novel and interesting in its own right. \n\nApart from the possibly exponential search complexity that the authors acknowledge, I have concerns about the complexity of searching over the parametrization of the shifts on mutable variables, I.e. estimation of h*(S_). How is this parametrized and why is the cost considered constant in the complexity analysis? If I am understanding the parametrization correctly, this should be exponential in the size of the mutable variables? \n\nGiven that the sparse optimization method is competitive relative to the min-max identification algorithm, it deserves more explanation than just a paragraph. \n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "For a supervised learning task where data comes from multiple environments, invariant prediction relies on identifying a stable set of features that don't influence the shift in probability distributions over different training environments. This paper proposes a method to find the optimal sufficient set of stable features. The method is split into two cases depending on whether the optimal set of stable features is the entire set of stable features. The paper proposes a sufficient graphical condition for the latter that can be tested by causal discovery. If the condition fails, the paper parametrizes the dataset shift and does a brute-force search to get the min-max optimal subset. To reduce the search space, an equivalence class of the set of subsets is identified. While acknowledging that this brute force search might be inefficient, the authors frame an alternative sparse min-max optimization to solve the subset selection problem. The brute force search method is validated on synthetic data and both methods are validated on real world data.  ", "strength_and_weaknesses": "The paper’s strongest contribution is the empirical validation on the real world datasets considered albeit still being relatively low-dimensional. Although I didn’t check the proofs in detail, both Both the methods proposed seem theoretically sound and believable. The concept of equivalences is also novel and interesting in its own right. \n\nApart from the possibly exponential search complexity that the authors acknowledge, I have concerns about the complexity of searching over the parametrization of the shifts on mutable variables, I.e. estimation of h*(S_). How is this parametrized and why is the cost considered constant in the complexity analysis? If I am understanding the parametrization correctly, this should be exponential in the size of the mutable variables? \n\nGiven that the sparse optimization method is competitive relative to the min-max identification algorithm, it deserves more explanation than just a paragraph. \n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper can be organized better and written more clearly. To reiterate a point made earlier, only one paragraph devoted to a method that's competitive w.r.t. the main algorithm seems less. Some notations are not clear. Example J: Pa(X_m) -> X_m was unclear - the authors probably meant alphabets? The prose can also be made more explanatory. For example, a central contribution of the paper is the concept of g-equivalence but apart from the definition there is no intuitive explanation of g-equivalence. ", "summary_of_the_review": "Overall, the authors propose two methods for an important problem - one of which potentially has an exponential search complexity and it's not clear to me how they parametrize the mutable shifts. The other method while having decent empirical performance is not addressed much in the theory portion of the paper. If the authors can address either of the two concerns, it would make the paper much stronger. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667420732076}, {"id": "C2yo3tg3c6", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3458/Reviewer_K5Kv"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a model selection procedure for choosing stable models that have minimax optimal risk across a family of distributions  induced by intervening on mutable variables. This work extends the shift-stable prediction principles established in Subbaswamy et al 2019, who established that one can construct a shift-stable predictor by estimating the intervened conditional expectation of Y given an intervention on mutable variables, and some subset of stable variables. This work addresses some practical gaps left in that paper: in particular, it is often ambiguous which set of stable variables one should condition on (in addition to intervening on mutable variables) from the perspective of minimax MSE, and cases where there exists no ambiguity are hard to identify. To address these gaps, the authors propose (1) a simple-to-confirm graphical criterion that is equivalent to the case where one should condition on all the stable variables to obtain a minimax predictor; (2) a selection criterion based on estimated minimax risk for selecting a minimax optimal subset of variables; and (3) algorithms for reducing the complexity of this subset selection. In addition, the authors propose using causal structure discovery from multiple environments (Huang et al 2020) to learn causal structure and mutable/immutable variables. The authors demonstrate that their algorithm reduces worst-case MSE in several cross-environment generalization experiments.", "review_text": "The overall method here seems promising theoretically and empirically, and many of the claims are interesting. However, there are too many additional extraneous claims that are unsubstantiated or distracting.", "strengths": "**Strengths:**\n + The authors identify several practical gaps in the proposed implementation of the Subbaswamy et al 2019 surgery estimator. The graphical criterion for conditioning on all stable variables is particularly nice, as it saves end-uses from having to run the ID algorithm for each DAG to understand whether the intervened distribution can be written as a conditional distribution.\n + The observation that the best subset of variables based on validation risk in the training environments does not correspond to the best subset for minimax risk is important. I wish the disconnect between these two were highlighted more prominently in a very simple toy example.\n + I like the claim that the minimax risk can be estimated by exploring the set of potential downstream distributions over $X_M$, although this was also not treated in enough detail.\n + The proposed application of LASSO style selection as an alternative to subset search is interesting, although this was also not treated in enough detail.\n + Experimental results seem promising and Figures 1, 6, and 7 are particularly compelling, although absolute numbers for MSE and comparisons of performance across different target distributions would be appreciated.\n\n**Weaknesses:**\n\n**Missing Technical Details**\n - The authors try to do too much in the paper, so many ideas are not treated thoroughly enough. While some results are shown with some rigor (e.g., the graphical criterion in Theorem 3.1, crucial details of other parts of the algorithm are missing. Several examples of missing details follow. I would suggest cutting down the claimed contributions of the paper, or expanding to a longer format venue.\n- How does one solve the maximization problem over $P_J$ that defines $h^*$ in practice? $P_J$ is a non-parametric class of functions that allows arbitrary dependence for the structural equation $Pa(X_M) \\rightarrow X_M$. This seems like a critical detail for the proposed algorithm to be usable. Does one need to make assumptions about $J$?\n - There appear to be missing assumptions for the structure identifiability claim in Proposition 3.4, and the proof for this proposition is not actually a proof, but the statement of an algorithm. At the very least, it seems important to have some assumption akin to positivity about the set of environments that is observed in training; as of now, there are no conditions stated about how the environments should be heterogeneous, or whether all environments need to change all mutable variables or something similar. I imagine there are many stronger conditions that are given in the Huang et al 2020 paper that is quoted extensively.\n - The g-equivalence algorithm seems quite critical given the claims in the paper, but is omitted from the main text. In particular, it seems important to understand when composing the g-equivalence algorithm with the reduced variable selection problem is actually more efficient than the original variable selection problem. I suspect it is, but it seems necessary to substantiate this claim. Meanwhile, the complexity analysis section goes into examples that don't feel particularly relevant.\n - It is not clear to me whether the Estimate $f_{S_-}$ section is correct, or why the algorithm needs to be so complicated; if $f_{S_-}$ is identifiable, shouldn't we be able to to estimate it by obtaining a functional from the ID algorithm? This seems more straightforward than estimating and generating from structural equations, which also requires additional assumptions for correctness. At any rate, it needs to be proved that the proposed algorithm actually estimates the appropriate interventional distribution.\n - The sparse min-max optimization section has far too little detail, and it is not clear what claims the authors want to make here. None of the cited work has results about doing variable selection with the inner maximization (rather they assume risk is minimized across the same distribution regardless of the $\\alpha, \\beta$ parameters), nor does it extend beyond linear models. If this is meant to be a main contribution, more details of the exact approach should be given in the main text. It is not necessary to prove anything about this practical approach, but the citations to selection consistency work suggest that the authors want to make stronger claims than they can substantiate.\n - For the actual implementations, no details are given about how the predictors are actually estimated. Is everything a linear model? How are hyperparameters set (for example, the very important robustness parameter in Anchor Regression)?\n - How are bi-directed edges handled by the algorithm? There is no mention of these, and some of the proofs assume a topological ordering exists, but the causal structure discovery algorithm returns bidirected edges in some cases.\n\n**Other issues**\n - In the set-up, there is no formal definition of a \"stable predictor\". Currently there is a sentence stating that such a predictor can be \"transferred to a broader family of environments without any adjustment\", but it is not clear what \"transferred\" means in this sentence. What is the actual invariance criterion, i.e., the property of f(X) that is the same across environments?\n - I think there is an error in Example 1. When the dashed arrow is absent, by the do calculus rules 1 then 3, $p(y \\mid k, do(x^0_m)) = p(y \\mid do(x^0_m)) = p(y)$. This is consistent, I believe, with Theorem 3.1, which reduces to the conditional $p(y \\mid k_2)$, but $k_2$ is the empty set in this example.\n - Notation in the Proof of Theorem 3.3 is different from the notation in the statement.\n\n**Suggestions:**\n - Can the proof of Theorem 1 be written using the rules of do calculus? It seems like this would be more straightforward than going back to conditional probability statements.\n - It would be nice to have a simple toy example showing how changes to the $X_M$ conditional distributions can induce changes in the risk of $f_{S_-}$, specifically showing that $f_{S_-}$ with the lowest in-distribution validation risk may not have the lowest worst-case validation risk.\n - The causal structure discovery portion of the algorithm seems like a distraction. Most of the result is quoting identification results from Huang et al and not a core contribution of this work. In addition, the assumptions for the structure discovery algorithms to work are stronger than the assumptions provided here. I would suggest dropping the structure discovery portion from the main development, and discussing structure discovery as a practical tool for implementation in the experiments section, as the contribution of this paper is mostly about doing the min-max optimization after causal structure has been ascertained, and there is no analysis of, say, the propagation of error from causal structure discovery to the final predictor. This would provide room for other important details, such as the algorithm for identifying g-equivalence classes or the sparse selection algorithm.\n - Comparisons of MSE across environments (not just worst-case MSE) would be useful here, since there is often a tradeoff between the two. It would also be useful to see how the variables selected by Surgery and the proposed algorithm differ. Finally, it would be useful to understand whether the MSE of these estimators is actually in a range that would be useful for prediction. Comparing to, e.g., an oracle estimator that is trained on data from the target environment, or a baseline that just takes the mean outcome would be useful.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a model selection procedure for choosing stable models that have minimax optimal risk across a family of distributions  induced by intervening on mutable variables. This work extends the shift-stable prediction principles established in Subbaswamy et al 2019, who established that one can construct a shift-stable predictor by estimating the intervened conditional expectation of Y given an intervention on mutable variables, and some subset of stable variables. This work addresses some practical gaps left in that paper: in particular, it is often ambiguous which set of stable variables one should condition on (in addition to intervening on mutable variables) from the perspective of minimax MSE, and cases where there exists no ambiguity are hard to identify. To address these gaps, the authors propose (1) a simple-to-confirm graphical criterion that is equivalent to the case where one should condition on all the stable variables to obtain a minimax predictor; (2) a selection criterion based on estimated minimax risk for selecting a minimax optimal subset of variables; and (3) algorithms for reducing the complexity of this subset selection. In addition, the authors propose using causal structure discovery from multiple environments (Huang et al 2020) to learn causal structure and mutable/immutable variables. The authors demonstrate that their algorithm reduces worst-case MSE in several cross-environment generalization experiments.", "strength_and_weaknesses": "**Strengths:**\n + The authors identify several practical gaps in the proposed implementation of the Subbaswamy et al 2019 surgery estimator. The graphical criterion for conditioning on all stable variables is particularly nice, as it saves end-uses from having to run the ID algorithm for each DAG to understand whether the intervened distribution can be written as a conditional distribution.\n + The observation that the best subset of variables based on validation risk in the training environments does not correspond to the best subset for minimax risk is important. I wish the disconnect between these two were highlighted more prominently in a very simple toy example.\n + I like the claim that the minimax risk can be estimated by exploring the set of potential downstream distributions over $X_M$, although this was also not treated in enough detail.\n + The proposed application of LASSO style selection as an alternative to subset search is interesting, although this was also not treated in enough detail.\n + Experimental results seem promising and Figures 1, 6, and 7 are particularly compelling, although absolute numbers for MSE and comparisons of performance across different target distributions would be appreciated.\n\n**Weaknesses:**\n\n**Missing Technical Details**\n - The authors try to do too much in the paper, so many ideas are not treated thoroughly enough. While some results are shown with some rigor (e.g., the graphical criterion in Theorem 3.1, crucial details of other parts of the algorithm are missing. Several examples of missing details follow. I would suggest cutting down the claimed contributions of the paper, or expanding to a longer format venue.\n- How does one solve the maximization problem over $P_J$ that defines $h^*$ in practice? $P_J$ is a non-parametric class of functions that allows arbitrary dependence for the structural equation $Pa(X_M) \\rightarrow X_M$. This seems like a critical detail for the proposed algorithm to be usable. Does one need to make assumptions about $J$?\n - There appear to be missing assumptions for the structure identifiability claim in Proposition 3.4, and the proof for this proposition is not actually a proof, but the statement of an algorithm. At the very least, it seems important to have some assumption akin to positivity about the set of environments that is observed in training; as of now, there are no conditions stated about how the environments should be heterogeneous, or whether all environments need to change all mutable variables or something similar. I imagine there are many stronger conditions that are given in the Huang et al 2020 paper that is quoted extensively.\n - The g-equivalence algorithm seems quite critical given the claims in the paper, but is omitted from the main text. In particular, it seems important to understand when composing the g-equivalence algorithm with the reduced variable selection problem is actually more efficient than the original variable selection problem. I suspect it is, but it seems necessary to substantiate this claim. Meanwhile, the complexity analysis section goes into examples that don't feel particularly relevant.\n - It is not clear to me whether the Estimate $f_{S_-}$ section is correct, or why the algorithm needs to be so complicated; if $f_{S_-}$ is identifiable, shouldn't we be able to to estimate it by obtaining a functional from the ID algorithm? This seems more straightforward than estimating and generating from structural equations, which also requires additional assumptions for correctness. At any rate, it needs to be proved that the proposed algorithm actually estimates the appropriate interventional distribution.\n - The sparse min-max optimization section has far too little detail, and it is not clear what claims the authors want to make here. None of the cited work has results about doing variable selection with the inner maximization (rather they assume risk is minimized across the same distribution regardless of the $\\alpha, \\beta$ parameters), nor does it extend beyond linear models. If this is meant to be a main contribution, more details of the exact approach should be given in the main text. It is not necessary to prove anything about this practical approach, but the citations to selection consistency work suggest that the authors want to make stronger claims than they can substantiate.\n - For the actual implementations, no details are given about how the predictors are actually estimated. Is everything a linear model? How are hyperparameters set (for example, the very important robustness parameter in Anchor Regression)?\n - How are bi-directed edges handled by the algorithm? There is no mention of these, and some of the proofs assume a topological ordering exists, but the causal structure discovery algorithm returns bidirected edges in some cases.\n\n**Other issues**\n - In the set-up, there is no formal definition of a \"stable predictor\". Currently there is a sentence stating that such a predictor can be \"transferred to a broader family of environments without any adjustment\", but it is not clear what \"transferred\" means in this sentence. What is the actual invariance criterion, i.e., the property of f(X) that is the same across environments?\n - I think there is an error in Example 1. When the dashed arrow is absent, by the do calculus rules 1 then 3, $p(y \\mid k, do(x^0_m)) = p(y \\mid do(x^0_m)) = p(y)$. This is consistent, I believe, with Theorem 3.1, which reduces to the conditional $p(y \\mid k_2)$, but $k_2$ is the empty set in this example.\n - Notation in the Proof of Theorem 3.3 is different from the notation in the statement.\n\n**Suggestions:**\n - Can the proof of Theorem 1 be written using the rules of do calculus? It seems like this would be more straightforward than going back to conditional probability statements.\n - It would be nice to have a simple toy example showing how changes to the $X_M$ conditional distributions can induce changes in the risk of $f_{S_-}$, specifically showing that $f_{S_-}$ with the lowest in-distribution validation risk may not have the lowest worst-case validation risk.\n - The causal structure discovery portion of the algorithm seems like a distraction. Most of the result is quoting identification results from Huang et al and not a core contribution of this work. In addition, the assumptions for the structure discovery algorithms to work are stronger than the assumptions provided here. I would suggest dropping the structure discovery portion from the main development, and discussing structure discovery as a practical tool for implementation in the experiments section, as the contribution of this paper is mostly about doing the min-max optimization after causal structure has been ascertained, and there is no analysis of, say, the propagation of error from causal structure discovery to the final predictor. This would provide room for other important details, such as the algorithm for identifying g-equivalence classes or the sparse selection algorithm.\n - Comparisons of MSE across environments (not just worst-case MSE) would be useful here, since there is often a tradeoff between the two. It would also be useful to see how the variables selected by Surgery and the proposed algorithm differ. Finally, it would be useful to understand whether the MSE of these estimators is actually in a range that would be useful for prediction. Comparing to, e.g., an oracle estimator that is trained on data from the target environment, or a baseline that just takes the mean outcome would be useful.", "clarity,_quality,_novelty_and_reproducibility": "The paper's main weakness is clarity. There are just too many ideas, many of which are good, presented at the same time. Many key details are hidden deep in appendices. While there are many useful claims that are proved well, there are many other claims that are unsubstantiated, at least in the text as it stands. The authors need to prioritize the particular claims that they wish to make in the paper. Also, many proofs seem more convoluted than they need to be.\n\nOverall, the quality of the work is high. As I mentioned, many of the ideas are good, but difficult to dig out.\n\nThe contributions wrt prior work are clear.\n\nThere appear to be enough details for reproducibility in the appendix, but there are too few details about the actual experiments in the main text.", "summary_of_the_review": "The overall method here seems promising theoretically and empirically, and many of the claims are interesting. However, there are too many additional extraneous claims that are unsubstantiated or distracting.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666890244069}, {"id": "U8M_1-UwMsp", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3458/Reviewer_s36p"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper concerns how to select/learn a predictive model from causal relations between variables in the distributional robustness setting. Specifically, the paper argues that given a DAG, a predictive model should be learned from a subset of the variables which are stable across a group of learning environments, while not using the information of those mechanisms which can vary between different environments. The idea is similar to that of Subbaswamy et al. (2019) except that the latter proposed to directly search for the optimal subset from the set of stable variables by comparing their validation errors. This paper provides a more complete analysis and argues that instead of using the validation error one can directly compare and minimize the worst-case risk over a class of distributions with varying mechanisms of how those variables in the mutable set are generated. In addition, in order to save the computational cost, the paper advocates to search equivalent classes of subsets instead of searching all subsets. In the case where the computation is still heavy, the paper suggests to adopt the lasso regression for subset selection. ", "review_text": "This paper provides some analysis of the methods in Subbaswamy et al. (2019). It then proposed to compare the max (worst-case) risk directly, defining the equivalent class to save computation, and introduce lasso for consistent variable selection. Each of these could potentially be important but due to the amount of work that is presented in the paper, there are still questions on each of these. Some notations can be clarified at various places.", "strengths": "+1: the paper is technically sound\n+2: the paper has both numerical implementations and theoretical justifications\n+3: the paper studies an important yet under-developed area, and is very timely.\n\n-1: the paper presents too many concepts but without digging up each one with sufficient depth. Theorem 3.1 is equivalent to the degeneration condition in Subbaswamy et al. (2019). Theorem 3.3 seems to be new. The equivalent class is interesting, but I wonder how much saving it will really lead to (considering that there are costs associated with identifying these equivalent classes, and there may also be errors in this identification.) The lasso part is somewhat disentangled from the rest of the topics and lasso is also not new. If the paper were to present fewer concepts but with enough discussion on each one to provide a better justification, it may work out better.\n-2: The main point of Theorem 3.3 is that the main source of variability of the distributions in $\\mathcal{P}$ is from the variety in $J$, which defines a mutable variable in terms of its parents using a definite function. I wonder how in practice the richness of $\\{P_J\\}$ aligns with the richness of $\\mathcal{P}$. While $\\mathcal{P}$ should contain the data distribution from all environments, how does one consider the maximization of all the possible $J$ functions? Note that all measurable functions $J$ are uncountable. If in practice one only uses a small parametric class of $J$ functions, then they may not be able to cover the base for all $\\mathcal{P}$. I do not really find how $\\{P_J\\}$ is defined at the implementation level from the paper. \n-3: As mentioned above, while the idea of the equivalent class could potentially save some training time, it also takes time to identify these equivalent classes. If the time for training each model is smaller than the time spent on the identification of these equivalent classes, then it may not be necessary to bother with these equivalent classes. \n-4: The section on lasso looks a bit disconnected from the rest of the paper. Here is why. The paper went through all the efforts with Theorem 3.3, dismissing the use of the validation error in Subbaswamy et al. (2019), and defining the equivalent classes. Then all of sudden, the paper says basically if this is too much then we can just use lasso, which is a regularized version of the empirical loss function. If lasso works then what's the point of all the previous discussions?\n-5: There is ambiguity as to what $f_\\alpha(x_S \\beta, x_M)$ is in the lasso formulation (3). What are the parameters $\\alpha$ and $\\beta$? Should $x_S \\beta$ be thought of as the inner product? Is $\\alpha$ then a parameter that governs how the single index $x_S \\beta$ and the mutable variable $x_M$ are used to form the prediction function $f$?\n-6: The variable selection consistency of lasso was only proved for the linear model by Zhao and Yu 2016. I am not sure if some results exist for an arbitrary non-linear model. I may be wrong about this.\n-7: I think the idea is that $x_M$ should not be used as part of the predictive model. That is what I saw in Algorithm 1. However, why x_M is included in the lasso model?\n-8: Some notations are confusing to me. I think D ⊥G Y | Z is defined as D and Y are d-separated by Z in the paper. But what is Y ⊥ G_{\\overline{X^0_M} K| X^0_M (this appears in Example 1). There is also Y ⊥ G_{\\overline{X} \\underline{Z}} Z | X (see top of page 16). What do these notations mean? Still d-separation?\n-9: please clarify what the two columns h*(S-) and max MSE mean respectively. As far as I can tell, h*(S-) is also a maximal MSE (over ...?) as it is defined. So it is unclear to me what is the difference. Is max MSE the test data maximal MSE over different observed environments?", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper concerns how to select/learn a predictive model from causal relations between variables in the distributional robustness setting. Specifically, the paper argues that given a DAG, a predictive model should be learned from a subset of the variables which are stable across a group of learning environments, while not using the information of those mechanisms which can vary between different environments. The idea is similar to that of Subbaswamy et al. (2019) except that the latter proposed to directly search for the optimal subset from the set of stable variables by comparing their validation errors. This paper provides a more complete analysis and argues that instead of using the validation error one can directly compare and minimize the worst-case risk over a class of distributions with varying mechanisms of how those variables in the mutable set are generated. In addition, in order to save the computational cost, the paper advocates to search equivalent classes of subsets instead of searching all subsets. In the case where the computation is still heavy, the paper suggests to adopt the lasso regression for subset selection. ", "strength_and_weaknesses": "+1: the paper is technically sound\n+2: the paper has both numerical implementations and theoretical justifications\n+3: the paper studies an important yet under-developed area, and is very timely.\n\n-1: the paper presents too many concepts but without digging up each one with sufficient depth. Theorem 3.1 is equivalent to the degeneration condition in Subbaswamy et al. (2019). Theorem 3.3 seems to be new. The equivalent class is interesting, but I wonder how much saving it will really lead to (considering that there are costs associated with identifying these equivalent classes, and there may also be errors in this identification.) The lasso part is somewhat disentangled from the rest of the topics and lasso is also not new. If the paper were to present fewer concepts but with enough discussion on each one to provide a better justification, it may work out better.\n-2: The main point of Theorem 3.3 is that the main source of variability of the distributions in $\\mathcal{P}$ is from the variety in $J$, which defines a mutable variable in terms of its parents using a definite function. I wonder how in practice the richness of $\\{P_J\\}$ aligns with the richness of $\\mathcal{P}$. While $\\mathcal{P}$ should contain the data distribution from all environments, how does one consider the maximization of all the possible $J$ functions? Note that all measurable functions $J$ are uncountable. If in practice one only uses a small parametric class of $J$ functions, then they may not be able to cover the base for all $\\mathcal{P}$. I do not really find how $\\{P_J\\}$ is defined at the implementation level from the paper. \n-3: As mentioned above, while the idea of the equivalent class could potentially save some training time, it also takes time to identify these equivalent classes. If the time for training each model is smaller than the time spent on the identification of these equivalent classes, then it may not be necessary to bother with these equivalent classes. \n-4: The section on lasso looks a bit disconnected from the rest of the paper. Here is why. The paper went through all the efforts with Theorem 3.3, dismissing the use of the validation error in Subbaswamy et al. (2019), and defining the equivalent classes. Then all of sudden, the paper says basically if this is too much then we can just use lasso, which is a regularized version of the empirical loss function. If lasso works then what's the point of all the previous discussions?\n-5: There is ambiguity as to what $f_\\alpha(x_S \\beta, x_M)$ is in the lasso formulation (3). What are the parameters $\\alpha$ and $\\beta$? Should $x_S \\beta$ be thought of as the inner product? Is $\\alpha$ then a parameter that governs how the single index $x_S \\beta$ and the mutable variable $x_M$ are used to form the prediction function $f$?\n-6: The variable selection consistency of lasso was only proved for the linear model by Zhao and Yu 2016. I am not sure if some results exist for an arbitrary non-linear model. I may be wrong about this.\n-7: I think the idea is that $x_M$ should not be used as part of the predictive model. That is what I saw in Algorithm 1. However, why x_M is included in the lasso model?\n-8: Some notations are confusing to me. I think D ⊥G Y | Z is defined as D and Y are d-separated by Z in the paper. But what is Y ⊥ G_{\\overline{X^0_M} K| X^0_M (this appears in Example 1). There is also Y ⊥ G_{\\overline{X} \\underline{Z}} Z | X (see top of page 16). What do these notations mean? Still d-separation?\n-9: please clarify what the two columns h*(S-) and max MSE mean respectively. As far as I can tell, h*(S-) is also a maximal MSE (over ...?) as it is defined. So it is unclear to me what is the difference. Is max MSE the test data maximal MSE over different observed environments?", "clarity,_quality,_novelty_and_reproducibility": "The paper seems to have some new ideas, though the discussion for each is thin. It could have presented fewer topics with more discussion on each topic.\nThe paper is overall well written with confusing notations and unclarity at a couple of places. ", "summary_of_the_review": "This paper provides some analysis of the methods in Subbaswamy et al. (2019). It then proposed to compare the max (worst-case) risk directly, defining the equivalent class to save computation, and introduce lasso for consistent variable selection. Each of these could potentially be important but due to the amount of work that is presented in the paper, there are still questions on each of these. Some notations can be clarified at various places.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666640612769}, {"id": "GBfnqh0KnB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3458/Reviewer_AmVi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper considers the problem of learning stable information to transfer to unseen environments. In particular, following the work of Subbaswamy et al. (2019), it investigates which subset of the whole stable information should the model transfer, in order to achieve optimal generalization ability. Authors propose to maximize over mutable mechanisms and to search over only equivalent classes in terms of worst-case risk, with certain theoretical characterization based on causal graph. The performance of the proposed method is also empirically validated.", "review_text": "I think that the paper requires a major revision to make it readable and rigorous (e.g., clearly define the setting and notions of stable, unstable things, the do-operator; make the proof precise and give explanations where it is not very very obvious; give summary or high-level idea of your proof; etc.). Thus, I cannot recommend acceptance for the present version. I would like to re-evaluate the paper, and raise my score if authors revise their paper and could address my concerns/questions during the discussion period. I look forward to the upcoming discussions with authors.", "strengths": "Pros:\n\n- the topic of investigating which subset of the whole stable information should the model transfer is very interesting \n- interesting theoretical charactizations, which lead to practically efficient algorithms\n- good empirical performance in the considered experiments\n\nCons and limitations:\n\n- the paper relies on causal discovery from observational data, which in practice cannot avoid estimating errors, particularly for large scale settings.\n- the proposed method seems not be able to work for image-like data where one has to learn representations.\n- writing is poor (see below),\n- minor one: causal sufficiency is required ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper considers the problem of learning stable information to transfer to unseen environments. In particular, following the work of Subbaswamy et al. (2019), it investigates which subset of the whole stable information should the model transfer, in order to achieve optimal generalization ability. Authors propose to maximize over mutable mechanisms and to search over only equivalent classes in terms of worst-case risk, with certain theoretical characterization based on causal graph. The performance of the proposed method is also empirically validated.", "strength_and_weaknesses": "Pros:\n\n- the topic of investigating which subset of the whole stable information should the model transfer is very interesting \n- interesting theoretical charactizations, which lead to practically efficient algorithms\n- good empirical performance in the considered experiments\n\nCons and limitations:\n\n- the paper relies on causal discovery from observational data, which in practice cannot avoid estimating errors, particularly for large scale settings.\n- the proposed method seems not be able to work for image-like data where one has to learn representations.\n- writing is poor (see below),\n- minor one: causal sufficiency is required ", "clarity,_quality,_novelty_and_reproducibility": "The topic of investigating which subset of the whole stable information should the model transfer is very interesting, particularly in combination with causal discovery. Indeed, I spent quite much time on this paper, perhaps triple of reviewing any other paper in my slot. \n\nI have to say that the poor writing makes it hard to evaluate the paper. As a theoretical paper (or at least lots of contributions are theoretical), I feel quite struggling as many places (problem setting, proofs, notations, etc.) are not clearly and rigorously treated. I have the following concerns and questions:\n\n1. I got confused with the definition of stable predictors in this paper. I'm familiar with works related to invariant causal prediction and invariant risk minimization. They assume that the conditional probability $P(Y|pa(Y))$ does not change across all domains of interest. Consider a setting with causal graph $X_1\\to Y, Y\\to X_2$, and $P(X_1)$, $P(X_2|Y)$ change but $P(Y|X_1)$ keeps constant. According to the definition of stable and unstable set in this set, it seems both $X_1, X_2$ will be treated as unstable variables, and the best estimate should be $\\mathbb E(Y)$ . However, in the spirit of IRM, $X_1$ is still treated as invariant variable. Is this correct? Or can you make the setting, the definition of stable and unstable things clear in the paper? Besides, in the invariant learning literature, there has been minimax optimality result regrading the causal parental variables.\n2. I feel confused with the do-operator here. Do you set $X_M$​ to be a fixed value or another random variable? I think it needs be elaborated.\n3. A drawback of previous work is tghat *''the searching cost is exponentially expensive w.r.t. dS, making it hard to be applied to large-scale scenarios*--but the considered experiments in this paper is not that large-scale. And for large scale case, causal discovery will be very likely to incur errors. Can you consider even larger scale experiment?\n4. for showing identification result of Proposition 3.2: why $E$ could be treated as a node in the causal graph? the causal graph part does not say anything about this. (see also a question regarding the training domains)\n5. Regarding counter-example: again I feel confused about  $do(X_m)$. Why should the expectation in Eq (8) and (9)  take sum of $x_m$? Why do we have this indicator function  $\\mathbb 1(x_s=1, x_m)$  when we calculate the expectation? And in the last three equations in the counter-example, you first cancel $a_y^2$, then take the limit of $a_y\\to 0$ in the denominator. I feel that this part gets tricky. Can you give exact numerical values to show the inequality? A quick summary or high-level idea of the counter-example shall be provided in the main text.\n6. About Theorem 3.3: I did not see how the proof shows this result. In the proof, the equation $max_{P^e}\\mathcal L_{P^e}(f_{S})$=... in the first line, the max is treated w.r.t. the integral of all variables; for the second line, the max is over an integral wrt. $x_M$, and the integral is a function of $pa(X_M)$. Why are they equivalent? Moreover, $P_J$ never appears in the proof, and I cannot see this proof is valid for showing Thm 3.3.\n7. About proposition 3.4: in the proof Thm 3.3, $P^e$ and $P_J$ is used as the one that achevies the wort risk. However, in this proof of Prop 3.4, $P_J$ is just distribution that yield the same observation distribution. Why?\n8. **More importantly**, I am surprised that there is nothing regarding the environment of training data. Intuitively, if you have one environment, you cannot never know what mechanism is stable and what is not. In the IRM paper, one has to assume the training environments sufficiently many and diverse. Please explain.\n9. the experiments are somewhat small scale. And all the causal graphs learnt from observational data are consistent with the true ones. What if the learnt graph has some error edges? Can you try the Sachs protein dataset?\n\n", "summary_of_the_review": "I think that the paper requires a major revision to make it readable and rigorous (e.g., clearly define the setting and notions of stable, unstable things, the do-operator; make the proof precise and give explanations where it is not very very obvious; give summary or high-level idea of your proof; etc.). Thus, I cannot recommend acceptance for the present version. I would like to re-evaluate the paper, and raise my score if authors revise their paper and could address my concerns/questions during the discussion period. I look forward to the upcoming discussions with authors.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666511674146}], "openreview_url": "https://openreview.net/forum?id=YP4QEmqh6Ia", "arxiv_id": "2107.01876", "paper_pdf": "papers/YP4QEmqh6Ia.pdf", "paper_pdf_sha256": "d60bf2b3bc7b8ae1901261511537a029607437d18640717e635cb74956258376", "paper_pdf_bytes": 1611158, "paper_pdf_source": "openreview", "code_url": "https://github.com/lmz123321/which_invariance", "code_repository": "lmz123321/which_invariance", "code_commit": "707f2772ad121011f49ce8bcc7b04da65b13f911", "code_archive": "repos/YP4QEmqh6Ia.zip", "code_archive_sha256": "41f602cc266e04fb58e39800bee3948905cd3d1e1c637a4aaf5e295fcdc77281", "code_archive_bytes": 2811752, "code_file_count": 31, "code_extensions": {".py": 25, ".ipynb": 5, ".r": 1}, "github_disk_usage_kb": 1179, "github_languages": {"Jupyter Notebook": 163253, "Python": 158335, "R": 15179}, "github_archived": false, "github_pushed_at": "2024-05-27T03:18:07Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/causally-invariant-predictor-with-shift"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FYUzzBPh_j", "year": 2022, "status": "rejected", "title": "Communicating via Markov Decision Processes", "authors": ["Samuel Sokota", "Christian Schroeder de Witt", "Maximilian Igl", "Luisa M Zintgraf", "Philip Torr", "J Zico Kolter", "Shimon Whiteson", "Jakob Nicolaus Foerster"], "authorids": ["~Samuel_Sokota1", "~Christian_Schroeder_de_Witt1", "~Maximilian_Igl1", "~Luisa_M_Zintgraf1", "~Philip_Torr1", "~J_Zico_Kolter1", "~Shimon_Whiteson1", "~Jakob_Nicolaus_Foerster1"], "authors_source": "OpenReview API", "abstract": "We consider the problem of communicating exogenous information by means of Markov decision process trajectories. This setting, which we call a Markov coding game (MCG), generalizes both source coding and a large class of referential games. MCGs also isolate a problem that is important in decentralized control settings in which cheap-talk is not available---namely, they require balancing communication with the associated cost of communicating. We contribute a theoretically grounded approach to MCGs based on maximum entropy reinforcement learning and minimum entropy coupling that we call greedy minimum entropy coupling (GME). We show both that GME is able to outperform a relevant baseline on small MCGs and that GME is able to scale efficiently to extremely large MCGs. To the latter point, we demonstrate that GME is able to losslessly communicate binary images via trajectories of Cartpole and Pong, while simultaneously achieving the maximal or near maximal expected returns, and that it is even capable of performing well in the presence of actuator noise.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "er9o7CteU0i", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3808/Reviewer_o2vx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper suggests a combination of MaxEnt RL with Minimum Entropy Coupling to construct a communication method via (on the fly) modulation of a stochastic policy execution. The core idea being that communication does not require a separate channel beyond the capability of the receiver to observe the sender's (otherwise goal-driven) behaviour. The paper is supported by several experiments, including a study of external interferences with the established behaviour-based communication channel in the form of action execution uncertainty. \n", "review_text": "The paper is very clearly written right until the actual method is described. At that point key details vanish and the approach remains half-presented with some ostensible gaps on the receiver side.\n\nIn particular, authors take a relatively good care of preparing and presenting the sender's behaviour generation side. From ensuring that the natural policy contains sufficient stochasticity to piggyback a message to some (though fairly naive) theoretical guarantees for the correctness on the sender's side. This reviewer has little argument against that portion. However, the paper completely misses the explanation of \\pi_{Z} -- the receiver's decoding policy. The only line about it reads \"... the receiver guesses the maximum a posteriori message.\". \n\nUnfortunately, posterior calculation, as presented previously in Step 2 belief formulation, depends on the knowledge of the Sender's policy throughout execution. As it is greedily constructed, it seems strange that the receiver would be aware of it to any reasonable extent. This makes the posterior calculation impossible on the receiver's side. One could argue that these are calculated in some form of an equilibrium pattern, but that is not the authors intent or explanation either. So, either there is a secondary communication channel that allows the policy information to be accessible to the receiver -- which contradicts the assumption of communicating by behaviour only -- or there's one half of the approach that authors neglected to present -- which is even more unfortunate. \n\nIn addition, authors tend to miss quite a bit of related work. From technological issues (e.g., combinations of RL with information bottlenecks) to ideological \"brothers\" (e.g., boosting goal recognition in RL solutions and its countermeasures, such as deceptive RL).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper suggests a combination of MaxEnt RL with Minimum Entropy Coupling to construct a communication method via (on the fly) modulation of a stochastic policy execution. The core idea being that communication does not require a separate channel beyond the capability of the receiver to observe the sender's (otherwise goal-driven) behaviour. The paper is supported by several experiments, including a study of external interferences with the established behaviour-based communication channel in the form of action execution uncertainty. \n", "main_review": "The paper is very clearly written right until the actual method is described. At that point key details vanish and the approach remains half-presented with some ostensible gaps on the receiver side.\n\nIn particular, authors take a relatively good care of preparing and presenting the sender's behaviour generation side. From ensuring that the natural policy contains sufficient stochasticity to piggyback a message to some (though fairly naive) theoretical guarantees for the correctness on the sender's side. This reviewer has little argument against that portion. However, the paper completely misses the explanation of \\pi_{Z} -- the receiver's decoding policy. The only line about it reads \"... the receiver guesses the maximum a posteriori message.\". \n\nUnfortunately, posterior calculation, as presented previously in Step 2 belief formulation, depends on the knowledge of the Sender's policy throughout execution. As it is greedily constructed, it seems strange that the receiver would be aware of it to any reasonable extent. This makes the posterior calculation impossible on the receiver's side. One could argue that these are calculated in some form of an equilibrium pattern, but that is not the authors intent or explanation either. So, either there is a secondary communication channel that allows the policy information to be accessible to the receiver -- which contradicts the assumption of communicating by behaviour only -- or there's one half of the approach that authors neglected to present -- which is even more unfortunate. \n\nIn addition, authors tend to miss quite a bit of related work. From technological issues (e.g., combinations of RL with information bottlenecks) to ideological \"brothers\" (e.g., boosting goal recognition in RL solutions and its countermeasures, such as deceptive RL).\n", "summary_of_the_review": "Good core idea and intend, but the paper is incomplete.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636400440222}, {"id": "a1Rt6wZDcEb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3808/Reviewer_Er8e"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to send a message through an MDP. The key is that the policy should still be optimal in the sense of the distribution of actions in each state, while communicating the message. ", "review_text": "The problem is nice - to keep the same policy and being able to send a message. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to send a message through an MDP. The key is that the policy should still be optimal in the sense of the distribution of actions in each state, while communicating the message. ", "main_review": "The problem is nice - to keep the same policy and being able to send a message. ", "summary_of_the_review": "There are multiple issues in the paper, due to which the reviewer feels that the paper is not ready in its current form. \n\n1. How much message entropy can be communicated? In source coding, we have the entropy of message that is communicated through a certain message length. Similarly, when communicating through a MDP, it is important to know message length communicated for a given finite T length MDP. \n2. Since the optimal policy is deterministic, for a given state, there is a unique optimal action, it is unclear how the policy can be the same and still contain the message. This is because deterministic action has no entropy. It seems that the authors are assuming that even if message is sent o(T) of the time, policy will be the same - and thus o(T) times, actions can be chosen - so \\tilde{O}(T\\log |A| ) length message can be communicated without any change in average policy. Not sure, if the authors are able to get better - and how to quantitatively say that. \n3. In general systems, the model of MDP may not be known between sender and receiver. How can these issues be handled?\n4. Channel coding and source coding are separate problems, and the authors seem to write the two as same in the text. \n5. Two of the special cases are given. How does the results here give same/improved communication results in the two domains. The result of message communications need to be at least the same as in those areas, with results in the general setup. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636332917691}, {"id": "y5vY0XoetKU", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3808/Reviewer_FcQX"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper introduces a problem setting, namely Markov Coding Game (MCG): sender receives a state (message) and takes an action to communicate the state to the receiver, receiver observes the taken action, then takes an action to decode the observed state of sender. This game is related to the referential games, source coding as well as decentralized control. To solve the MCG problem, the author designs the greedy minimum entropy coupling algorithm (GME), that aims to maximize the returns of the MDP and learn a good communication protocol simultaneously by combining existing techniques, MaxEnt RL and MEC. The algorithm is test empricially on Gridworld, Cartpole, and Pong.\n", "review_text": "Strength:\n\n1. The paper identifies some interesting connections of MCG with several existing problems.\n2. The paper conducted extensive empirical evaluation of its model. The results look good and are presented with carefully-designed figures and videos. The performance comparison with RL+PR baseline is also interesting,\n\nWeakness:\n\n1. I am sorry to say the paper does not have sufficient novelty in both its proposed problem setting and the learning algorithm. First, I think the described problem setting, at least something very similar, already exists (see the reference below). The only difference might be that they do not explicitly evaluate on how accurate the message is decoded. But I am not sure why it is necessary to explicitly evaluate this objective, if we could design a game where good communication is necessary to achieve optimality. I hope the author can provide more motivation on the design of this problem setting. Second, technical contribution is limited, as the proposed GME algorithm seems to be a straightforward combination of existing methods, MaxEnt RL and MEC.\n2. I am not sure why the authors describe their proposed method as \"theoretically grounded\", as I see no theoretical guarantee about its proposed method in this paper from the three Propositions in practice. I think it might be possible to combine proof techniques in information signaling scheme design (e.g., persuasiveness constraint) to improve the theoretical results.\n3. In the empirical evaluation, the results look good, but I wish to see more baselines for comparisons. However, some more thorough tests are possible to help us understand the strength of the proposed method (e.g., See question 3 below).\n\nQuestions:\n\n1. I wonder why the authors describe their problem setting as \"the receiver observes the sender’s MDP trajectory\", but from the examples, it actually just observes the sender's taken action. Is it just a fancy word, or the author is actually considering something more general about the receiver's observation.\n2. I wonder how the proposed problem is the related to a series of literature in learning multiagent communication, and can their method serves as the additional baseline of this paper:\n    - Mordatch, Igor, and Pieter Abbeel. \"Emergence of grounded compositional language in multi-agent populations.\" Thirty-second AAAI conference on artificial intelligence. 2018.\n    - S. Sukhbaatar, R. Fergus, et al. Learning multiagent communication with backpropagation. In\nAdvances in Neural Information Processing Systems, pages 2244–2252, 2016.\n    - Lowe, Ryan, et al. \"Multi-agent actor-critic for mixed cooperative-competitive environments.\" Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017.\n3. I notice that the games in the experiment do not include the additional dummy action available for communication, so the agent has to take the action that not only serves as to achieve value but also to achieve good communication with the receiver. I wonder what happens if the game includes the redundant action that the agent can use to communicate.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces a problem setting, namely Markov Coding Game (MCG): sender receives a state (message) and takes an action to communicate the state to the receiver, receiver observes the taken action, then takes an action to decode the observed state of sender. This game is related to the referential games, source coding as well as decentralized control. To solve the MCG problem, the author designs the greedy minimum entropy coupling algorithm (GME), that aims to maximize the returns of the MDP and learn a good communication protocol simultaneously by combining existing techniques, MaxEnt RL and MEC. The algorithm is test empricially on Gridworld, Cartpole, and Pong.\n", "main_review": "Strength:\n\n1. The paper identifies some interesting connections of MCG with several existing problems.\n2. The paper conducted extensive empirical evaluation of its model. The results look good and are presented with carefully-designed figures and videos. The performance comparison with RL+PR baseline is also interesting,\n\nWeakness:\n\n1. I am sorry to say the paper does not have sufficient novelty in both its proposed problem setting and the learning algorithm. First, I think the described problem setting, at least something very similar, already exists (see the reference below). The only difference might be that they do not explicitly evaluate on how accurate the message is decoded. But I am not sure why it is necessary to explicitly evaluate this objective, if we could design a game where good communication is necessary to achieve optimality. I hope the author can provide more motivation on the design of this problem setting. Second, technical contribution is limited, as the proposed GME algorithm seems to be a straightforward combination of existing methods, MaxEnt RL and MEC.\n2. I am not sure why the authors describe their proposed method as \"theoretically grounded\", as I see no theoretical guarantee about its proposed method in this paper from the three Propositions in practice. I think it might be possible to combine proof techniques in information signaling scheme design (e.g., persuasiveness constraint) to improve the theoretical results.\n3. In the empirical evaluation, the results look good, but I wish to see more baselines for comparisons. However, some more thorough tests are possible to help us understand the strength of the proposed method (e.g., See question 3 below).\n\nQuestions:\n\n1. I wonder why the authors describe their problem setting as \"the receiver observes the sender’s MDP trajectory\", but from the examples, it actually just observes the sender's taken action. Is it just a fancy word, or the author is actually considering something more general about the receiver's observation.\n2. I wonder how the proposed problem is the related to a series of literature in learning multiagent communication, and can their method serves as the additional baseline of this paper:\n    - Mordatch, Igor, and Pieter Abbeel. \"Emergence of grounded compositional language in multi-agent populations.\" Thirty-second AAAI conference on artificial intelligence. 2018.\n    - S. Sukhbaatar, R. Fergus, et al. Learning multiagent communication with backpropagation. In\nAdvances in Neural Information Processing Systems, pages 2244–2252, 2016.\n    - Lowe, Ryan, et al. \"Multi-agent actor-critic for mixed cooperative-competitive environments.\" Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017.\n3. I notice that the games in the experiment do not include the additional dummy action available for communication, so the agent has to take the action that not only serves as to achieve value but also to achieve good communication with the receiver. I wonder what happens if the game includes the redundant action that the agent can use to communicate.\n\n", "summary_of_the_review": "While the paper demonstrates some nice empirical results, my major concern is its lack of novelty in the proposed problem setting and the learning algorithm. Therefore, I think additional studies are needed in order to  improve the technical and empirical contribution of the work.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636073121450}, {"id": "qQ_e-DJe4E", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3808/Reviewer_8Cj5"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The article introduce a new theoretical game called Markov Coding Game (MCG).\nA Markov Coding game is a special case of Decentralized POMDP (Oliehoek et al. 2016). It seems to be a two-players cooperative game with a sender agent and a receiver agent. Given a message unknown to the receiver, the sender agent must play on a fixed MDP in a way that facilitates the decoding of this message by the receiver.\nA dedicated algorithm called GME for greedy minimum entropy coupling is proposed to solve this problem.\nAccording to the authors, this game is supposed to generalize several referential games and channel coding games.\nSome experiments are provided where bitmap images are transmitted through actions on gridword and pong environments.\n\n", "review_text": "Pro:\n- Using MDP policies to code messages is a fun and interesting idea.\n- The Decentralized POMDP framework seems a natural formalism for this problem\n- Introducing a stylized theoretical game to formalize a difficult point in a more complex problem (like for instance multi-armed bandits for exploration/exploitation in RL) can provide valuable insights\n\nCons:\n- The key point when introducing new theoretical games is to give a clear, abstract view of a problem. But I found the formalization of this MCG problem too muddled. To start with, it is not clear to me if we are really facing a two-players game or a single-player (the sender) game because the way the receiver \"guesses\" the message is not clearly formalized in the paper. \n\nQuestions:\n- Assuming the game is a single-player game, is it reducible to an equivalent \"transformed\" MDP/POMDP where original MDP reward is shaped to reflect the additional \"easy-to-guess\" reward ?\n- If we are facing a two-players game, is it a special case of Markov game as in (Littman 94) ?\n- Is the proposed GME algorithm able to solve a special case of MCG, as for instance a channel coding problem as efficiently as an ad-hoc algorithm would ?\n\n\nMinor remarks:\np4 \"An Markov\" -> \"A Markov\"\np4 \"the receiver uses a trajectory conditional policy to guess the message \\hat{M} \\sim \\pi_{|Z}(Z)\"  is not a proper formalization.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The article introduce a new theoretical game called Markov Coding Game (MCG).\nA Markov Coding game is a special case of Decentralized POMDP (Oliehoek et al. 2016). It seems to be a two-players cooperative game with a sender agent and a receiver agent. Given a message unknown to the receiver, the sender agent must play on a fixed MDP in a way that facilitates the decoding of this message by the receiver.\nA dedicated algorithm called GME for greedy minimum entropy coupling is proposed to solve this problem.\nAccording to the authors, this game is supposed to generalize several referential games and channel coding games.\nSome experiments are provided where bitmap images are transmitted through actions on gridword and pong environments.\n\n", "main_review": "Pro:\n- Using MDP policies to code messages is a fun and interesting idea.\n- The Decentralized POMDP framework seems a natural formalism for this problem\n- Introducing a stylized theoretical game to formalize a difficult point in a more complex problem (like for instance multi-armed bandits for exploration/exploitation in RL) can provide valuable insights\n\nCons:\n- The key point when introducing new theoretical games is to give a clear, abstract view of a problem. But I found the formalization of this MCG problem too muddled. To start with, it is not clear to me if we are really facing a two-players game or a single-player (the sender) game because the way the receiver \"guesses\" the message is not clearly formalized in the paper. \n\nQuestions:\n- Assuming the game is a single-player game, is it reducible to an equivalent \"transformed\" MDP/POMDP where original MDP reward is shaped to reflect the additional \"easy-to-guess\" reward ?\n- If we are facing a two-players game, is it a special case of Markov game as in (Littman 94) ?\n- Is the proposed GME algorithm able to solve a special case of MCG, as for instance a channel coding problem as efficiently as an ad-hoc algorithm would ?\n\n\nMinor remarks:\np4 \"An Markov\" -> \"A Markov\"\np4 \"the receiver uses a trajectory conditional policy to guess the message \\hat{M} \\sim \\pi_{|Z}(Z)\"  is not a proper formalization.\n\n\n", "summary_of_the_review": "A new theoretical game called Markov Coding Game and an algorithm to solve it. But I found the formalization the problem too muddled.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635515994552}], "openreview_url": "https://openreview.net/forum?id=FYUzzBPh_j", "arxiv_id": "2107.08295", "paper_pdf": "papers/FYUzzBPh_j.pdf", "paper_pdf_sha256": "7749fda12891515acd2506b592a93dd19cf0b3639f78540d6fadd42b0273bec5", "paper_pdf_bytes": 531562, "paper_pdf_source": "openreview", "code_url": "https://github.com/schroederdewitt/meme", "code_repository": "schroederdewitt/meme", "code_commit": "d94690ea7f4026a93d5e116d9ba643504f8f71fe", "code_archive": "repos/FYUzzBPh_j.zip", "code_archive_sha256": "ca98203639850ec49a54b357ccb59029b70caa168241b9b0068a866735d3b684", "code_archive_bytes": 5799169, "code_file_count": 132, "code_extensions": {".py": 121, ".sh": 10, ".cpp": 1}, "github_disk_usage_kb": 4936, "github_languages": {"Python": 1226433, "C++": 13482, "Shell": 3704, "Makefile": 1976, "Dockerfile": 1566}, "github_archived": false, "github_pushed_at": "2023-05-08T13:56:32Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/implicit-communication-as-minimum-entropy"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hecuSLbL_vC", "year": 2021, "status": "rejected", "title": "Generalisation Guarantees For Continual Learning With Orthogonal Gradient Descent", "authors": ["Mehdi Abbana Bennani", "Thang Doan", "Masashi Sugiyama"], "authorids": ["~Mehdi_Abbana_Bennani1", "~Thang_Doan1", "~Masashi_Sugiyama1"], "authors_source": "OpenReview API", "abstract": "In Continual Learning settings, deep neural networks are prone to Catastrophic Forgetting. Orthogonal Gradient Descent (Farajtabar et al., 2019) was proposed to tackle the challenge. However, no theoretical guarantees have been proven yet. We present a theoretical framework to study Continual Learning algorithms in the NTK regime. This framework comprises closed form expression of the model through tasks and proxies for transfer learning, generalisation and tasks similarity. In this framework, we prove that OGD is robust to Catastrophic Forgetting then derive the first generalisation bound for SGD and OGD for Continual Learning. Finally, we study the limits of this framework in practice for OGD and highlight the importance of the NTK variation for Continual Learning.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "x0qkc0FQgL0", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2367/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper provides a theoretical analysis on the OGD based continual learning method. The method is in fact proposed by a previous paper (Farajtabar et al. 2019) and the current paper shows a generalization bound for the regression case. The result (Thm 3) compares the generalization bounds between SGD and OGD and shows OGD leads to a tighter bound. The theorem is also based on the bound on the Rademacher Complexity (Lemma 1).  The paper also suggests OGD+, which stores some data points from past tasks. They also present some experimental results on small benchmark datasets, and show OGD+ outperforms SGD and OGD.\n\nWhile the paper makes an interesting attempt on theoretical analyses of OGD based continual learning method, I feel the result is quite limited only to the OGD scheme. Also, the result is for regression, as shown in the loss function in Sec 3.2, but the experiments are on classification, so it's not clear with the connection with the theory and the experiments. The results also seem to be somewhat simple derivations from the known papers, like Jacot et al., (2018) and Liu et al, (2019). \n\nThe experimental results are also very limited  and weak since it only compares with SGD, an obvious weak scheme that suffers from catastrophic forgetting, and does not compare with any other continual learning baselines. For example, the state-of-the-art on CIFAR-100 is around 65%, and the performance of OGD is very weak. Even though the paper aims for a theoretical contribution, it is very limited only for OGD based scheme, which is not strong in practice. So, I am not sure about the significance of the contribution of the paper. But, I haven't fully read the entire proof of the paper, and I may have missed some details regarding the proof. I would like to see other reviewers' opinion as well. \n\nWhat about comparing with more enlarged and various benchmark datasets beyond MNIST and CIFAR-100, like CUB200 or Omniglot as shown in https://arxiv.org/pdf/2003.13726.pdf ? How does OGD or OGD+ compares with other baselines like EWC or MAS?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Theoretical analysis for OGD for continual learning", "review": "The paper provides a theoretical analysis on the OGD based continual learning method. The method is in fact proposed by a previous paper (Farajtabar et al. 2019) and the current paper shows a generalization bound for the regression case. The result (Thm 3) compares the generalization bounds between SGD and OGD and shows OGD leads to a tighter bound. The theorem is also based on the bound on the Rademacher Complexity (Lemma 1).  The paper also suggests OGD+, which stores some data points from past tasks. They also present some experimental results on small benchmark datasets, and show OGD+ outperforms SGD and OGD.\n\nWhile the paper makes an interesting attempt on theoretical analyses of OGD based continual learning method, I feel the result is quite limited only to the OGD scheme. Also, the result is for regression, as shown in the loss function in Sec 3.2, but the experiments are on classification, so it's not clear with the connection with the theory and the experiments. The results also seem to be somewhat simple derivations from the known papers, like Jacot et al., (2018) and Liu et al, (2019). \n\nThe experimental results are also very limited  and weak since it only compares with SGD, an obvious weak scheme that suffers from catastrophic forgetting, and does not compare with any other continual learning baselines. For example, the state-of-the-art on CIFAR-100 is around 65%, and the performance of OGD is very weak. Even though the paper aims for a theoretical contribution, it is very limited only for OGD based scheme, which is not strong in practice. So, I am not sure about the significance of the contribution of the paper. But, I haven't fully read the entire proof of the paper, and I may have missed some details regarding the proof. I would like to see other reviewers' opinion as well. \n\nWhat about comparing with more enlarged and various benchmark datasets beyond MNIST and CIFAR-100, like CUB200 or Omniglot as shown in https://arxiv.org/pdf/2003.13726.pdf ? How does OGD or OGD+ compares with other baselines like EWC or MAS?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603874598413}, {"id": "KP68nw6ma4d", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2367/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors use a Neural Tangent Kernel (NTK) approximation of wide neural nets to establish generalization bounds for continual learning (CL) using stochastic gradient descent (SGD) and orthogonal gradient descent (OGD).  In this regime, the authors prove that OGD does not suffer from catastrophic forgetting of training data.  The authors additionally introduce a modification to OGD which causes significant performance improvements in the Rotated MNIST and Permuted MNIST problems.  OGD involves storing feature maps from data points from previous tasks.  The modified OGD method (OGD+) additionally stores feature maps from the current task.  \n\nThe primary contribution of this paper is the theoretical analysis of continual learning.  Given that the CL problem does not have an extensive theoretical foundation, the generalization bound in this paper is a notable advance. The theory presented also provides a justification for the empirical observations observed by the authors that as overparameterization increases, the effect of catastrophic forgetting decreases in a variety of CL task setups.  The primary drawback of the paper is that the authors do not compare the OGD+ algorithm to other continual learning algorithms (synaptic intelligence, elastic weight consolidation, etc.).  As a result it is difficult to know how OGD+ compares to alternatives.  It is not clear to the reviewer why improving OGD to OGD+ is itself a contribution.  Given the expense occurred by OGD-type methods in storing ever increasing numbers of directions, it would be important to know the comparison of this method with others.  \n\nMinor comments:\n\n(1) Section 3.2: f^* is not defined as of this point in the paper.\n(2) Theorem 1: The theorem needs a quantifier of lambda\n(3) Line above Remark 1 k_\\tau -> \\kappa_\\tau\n(4) Theorem 2: The paper should define what \"is in the memory\" means when introducing OGD\ns\n(5) Theorem 3: Definition of R_T has incorrect dummy index in the summation\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel Theory for Continual Learning in the context of Orthogonal Gradient Descent.", "review": "The authors use a Neural Tangent Kernel (NTK) approximation of wide neural nets to establish generalization bounds for continual learning (CL) using stochastic gradient descent (SGD) and orthogonal gradient descent (OGD).  In this regime, the authors prove that OGD does not suffer from catastrophic forgetting of training data.  The authors additionally introduce a modification to OGD which causes significant performance improvements in the Rotated MNIST and Permuted MNIST problems.  OGD involves storing feature maps from data points from previous tasks.  The modified OGD method (OGD+) additionally stores feature maps from the current task.  \n\nThe primary contribution of this paper is the theoretical analysis of continual learning.  Given that the CL problem does not have an extensive theoretical foundation, the generalization bound in this paper is a notable advance. The theory presented also provides a justification for the empirical observations observed by the authors that as overparameterization increases, the effect of catastrophic forgetting decreases in a variety of CL task setups.  The primary drawback of the paper is that the authors do not compare the OGD+ algorithm to other continual learning algorithms (synaptic intelligence, elastic weight consolidation, etc.).  As a result it is difficult to know how OGD+ compares to alternatives.  It is not clear to the reviewer why improving OGD to OGD+ is itself a contribution.  Given the expense occurred by OGD-type methods in storing ever increasing numbers of directions, it would be important to know the comparison of this method with others.  \n\nMinor comments:\n\n(1) Section 3.2: f^* is not defined as of this point in the paper.\n(2) Theorem 1: The theorem needs a quantifier of lambda\n(3) Line above Remark 1 k_\\tau -> \\kappa_\\tau\n(4) Theorem 2: The paper should define what \"is in the memory\" means when introducing OGD\ns\n(5) Theorem 3: Definition of R_T has incorrect dummy index in the summation\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603769623162}, {"id": "gUZ45jsK7A", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2367/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n\n \nThis paper studies the theoretical aspect of a continual learning method called orthogonal gradient descent (OGD).\nIn this study, authors leverage Neural Tangent Kernel and over parameterized neural networks to prove the generalization of OGD. \n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for rejection. I like the idea of the paper to analyze an exist method from different aspect and even improving it. However, my major concern is about the clarity of the paper (see cons below).\n\n \n##########################################################################Pros: \n\n \n1. The paper investigate an important problem in continual learning framework which is the generalization.\n\n \n2. This paper provides some experiments to show the effectiveness of the proposed framework. \n\n \n##########################################################################\n\nCons: \n\n1. Unfortunately, the paper is not clear and very difficult to follow.\na) For instance, NTK is referred without explaining it well first. \nIn page 2, authors use \"CL\" for referring to continual learning but it has not been defined.\nb) There are many typos and capital letters have been used inappropriately.\nc) f_t has not been defined.\n2- Although the proposed method provides several experiments, there are still many other methods and datasets that have been ignored. There has been recent studies and frameworks that have outperformed OGD, it woud be great if you include them in your baselines. For instance, SGD+Droput in https://arxiv.org/pdf/2004.11545.pdf beats OGD.\n\n3-  There are many metrics to evaluate continual learning frameworks like backward transfer(BWT) or average accuracy over tasks.\nI would suggest the authors to look at the defined metrics in GEM (gradient episodic memory), https://arxiv.org/abs/1706.08840, and compute those values. \n\n \n##########################################################################\n\nQuestions during rebuttal period: \n\n \n- Please address and clarify the cons above .\n- Would you please elaborate more what could be the superior performance of OGD+ on Rotated Mnist dataset w.r.t OGD?\n- What is the time complexity of OGD+?\n\n \n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Generalisation Guarantees For Continual Learning With Orthogonal Gradient Descent ", "review": "##########################################################################\n\nSummary:\n\n \nThis paper studies the theoretical aspect of a continual learning method called orthogonal gradient descent (OGD).\nIn this study, authors leverage Neural Tangent Kernel and over parameterized neural networks to prove the generalization of OGD. \n\n##########################################################################\n\nReasons for score: \n\n \nOverall, I vote for rejection. I like the idea of the paper to analyze an exist method from different aspect and even improving it. However, my major concern is about the clarity of the paper (see cons below).\n\n \n##########################################################################Pros: \n\n \n1. The paper investigate an important problem in continual learning framework which is the generalization.\n\n \n2. This paper provides some experiments to show the effectiveness of the proposed framework. \n\n \n##########################################################################\n\nCons: \n\n1. Unfortunately, the paper is not clear and very difficult to follow.\na) For instance, NTK is referred without explaining it well first. \nIn page 2, authors use \"CL\" for referring to continual learning but it has not been defined.\nb) There are many typos and capital letters have been used inappropriately.\nc) f_t has not been defined.\n2- Although the proposed method provides several experiments, there are still many other methods and datasets that have been ignored. There has been recent studies and frameworks that have outperformed OGD, it woud be great if you include them in your baselines. For instance, SGD+Droput in https://arxiv.org/pdf/2004.11545.pdf beats OGD.\n\n3-  There are many metrics to evaluate continual learning frameworks like backward transfer(BWT) or average accuracy over tasks.\nI would suggest the authors to look at the defined metrics in GEM (gradient episodic memory), https://arxiv.org/abs/1706.08840, and compute those values. \n\n \n##########################################################################\n\nQuestions during rebuttal period: \n\n \n- Please address and clarify the cons above .\n- Would you please elaborate more what could be the superior performance of OGD+ on Rotated Mnist dataset w.r.t OGD?\n- What is the time complexity of OGD+?\n\n \n\n\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603760679000}], "openreview_url": "https://openreview.net/forum?id=hecuSLbL_vC", "arxiv_id": "2006.11942", "paper_pdf": "papers/hecuSLbL_vC.pdf", "paper_pdf_sha256": "371ddf000178cf42582268efe4af3f3e48d762a91925b523fedd2331d2173a0e", "paper_pdf_bytes": 1224837, "paper_pdf_source": "openreview", "code_url": "https://github.com/MehdiAbbanaBennani/continual-learning-ogdplus", "code_repository": "MehdiAbbanaBennani/continual-learning-ogdplus", "code_commit": "b633f2c5949f6ea165e2e76aab3d043065d770a8", "code_archive": "repos/hecuSLbL_vC.zip", "code_archive_sha256": "ea0cead224a8b2e96836d96bd726cdbdeef1c9b45eb1f3abaa7e5fd29a3351f6", "code_archive_bytes": 2995954, "code_file_count": 77, "code_extensions": {".py": 56, ".sh": 21}, "github_disk_usage_kb": 2924, "github_languages": {"Shell": 753395, "Python": 169648}, "github_archived": false, "github_pushed_at": "2022-10-10T12:10:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/generalisation-guarantees-for-continual"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ycxzArIvgF", "year": 2026, "status": "rejected", "title": "Effective Data Pruning through Score Extrapolation", "authors": ["Sebastian Schmidt", "Prasanga Dhungel", "Christoffer Löffler", "Björn Nieth", "Stephan Günnemann", "Leo Schwinn"], "authorids": ["~Sebastian_Schmidt2", "~Prasanga_Dhungel1", "~Christoffer_Löffler1", "~Björn_Nieth1", "~Stephan_Günnemann1", "~Leo_Schwinn1"], "authors_source": "OpenReview API", "abstract": "Training advanced machine learning models demands massive datasets, resulting in prohibitive computational costs. To address this challenge, data pruning techniques identify and remove redundant training samples while preserving model performance. Yet, existing pruning techniques predominantly require a full initial training pass to identify removable samples, negating any efficiency benefits for single training runs. To overcome this limitation, we introduce a novel importance score extrapolation framework that requires training on only a small subset of data. We present two initial approaches in this framework—k-nearest neighbors and graph neural networks—to accurately predict sample importance for the entire dataset using patterns learned from this minimal subset. We demonstrate the effectiveness of our approach for 2 state-of-the-art pruning methods (Dynamic Uncertainty and TDDS), 4 different datasets (CIFAR10, CIFAR100, Places-365, and ImageNet), and\ntraining paradigms (supervised, unsupervised, adversarial). Our results indicate that score extrapolation is a promising direction to scale expensive score calculation methods, such as pruning, data attribution, or other tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "83jy9MjqHr", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21707/Reviewer_uwUo"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper addresses the computational efficiency problem in data pruning methods by proposing a novel score extrapolation paradigm. The core idea is to compute importance scores on a small subset (10-20%) of the training data and then extrapolate these scores to the remaining unseen samples using KNN or GNN-based methods. The authors demonstrate that this approach achieves up to 4.9× speedup while maintaining competitive downstream task performance across supervised, unsupervised, and adversarial training settings on datasets including ImageNet, Places365, CIFAR-10, and synthetic CIFAR-100.", "review_text": "This paper addresses the computational efficiency problem in data pruning methods by proposing a novel score extrapolation paradigm. The core idea is to compute importance scores on a small subset (10-20%) of the training data and then extrapolate these scores to the remaining unseen samples using KNN or GNN-based methods. The authors demonstrate that this approach achieves up to 4.9× speedup while maintaining competitive downstream task performance across supervised, unsupervised, and adversarial training settings on datasets including ImageNet, Places365, CIFAR-10, and synthetic CIFAR-100.", "strengths": "1. Using the Extrapolation method to refine the sample ranking is rather new.", "weaknesses": "1. The critical weakness is the overlooked training cost.  The method requires: 1) First training: Train on random 10-20% subset to compute initial scores. 2) Embedding + Extrapolation: Extract features and extrapolate scores for remaining 80-90%. 3) Second training: Train final model on the extrapolated-pruned subset.\n\n1.1 This means training happens TWICE on similar-sized subsets, plus embedding the full dataset.\n\n1.2 Did the authors account for the cost of BOTH training phases? The paper claims \"4.9× speedup\" but it's unclear if both training runs are included in the time measurement.\n\n1.3 What is the cost of embedding 100% of the data? This requires forward passes through the entire dataset, which is not negligible.\n\n1.4 What is the actual extrapolation cost? KNN search or GNN inference on large-scale datasets has computational overhead.\n\n1.5 \nThe paper should clearly demonstrate that:\n ```\nCost(Training 20% + Embedding 100% + Extrapolation + Training 20%) \n< \nCost(Training 100% once + Standard pruning)\n```\n\n2. The code provided by the authors cannot be opened.\n\n3. On line 752, the authors mentioned \"For the unsupervised setting, we employ DINOv2 (Oquab et al., 2023) as a foundation model to obtain fixed embeddings for all samples. \". Why not use this as a baseline and see how much the supervised method can improve?", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper addresses the computational efficiency problem in data pruning methods by proposing a novel score extrapolation paradigm. The core idea is to compute importance scores on a small subset (10-20%) of the training data and then extrapolate these scores to the remaining unseen samples using KNN or GNN-based methods. The authors demonstrate that this approach achieves up to 4.9× speedup while maintaining competitive downstream task performance across supervised, unsupervised, and adversarial training settings on datasets including ImageNet, Places365, CIFAR-10, and synthetic CIFAR-100.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Using the Extrapolation method to refine the sample ranking is rather new.", "weaknesses": "1. The critical weakness is the overlooked training cost.  The method requires: 1) First training: Train on random 10-20% subset to compute initial scores. 2) Embedding + Extrapolation: Extract features and extrapolate scores for remaining 80-90%. 3) Second training: Train final model on the extrapolated-pruned subset.\n\n1.1 This means training happens TWICE on similar-sized subsets, plus embedding the full dataset.\n\n1.2 Did the authors account for the cost of BOTH training phases? The paper claims \"4.9× speedup\" but it's unclear if both training runs are included in the time measurement.\n\n1.3 What is the cost of embedding 100% of the data? This requires forward passes through the entire dataset, which is not negligible.\n\n1.4 What is the actual extrapolation cost? KNN search or GNN inference on large-scale datasets has computational overhead.\n\n1.5 \nThe paper should clearly demonstrate that:\n ```\nCost(Training 20% + Embedding 100% + Extrapolation + Training 20%) \n< \nCost(Training 100% once + Standard pruning)\n```\n\n2. The code provided by the authors cannot be opened.\n\n3. On line 752, the authors mentioned \"For the unsupervised setting, we employ DINOv2 (Oquab et al., 2023) as a foundation model to obtain fixed embeddings for all samples. \". Why not use this as a baseline and see how much the supervised method can improve?", "questions": "See weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761919739434}, {"id": "ikWSjHnOMw", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21707/Reviewer_mc1X"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper introduces a novel and efficient approach to data pruning. The method first estimates the importance of each training example by analyzing only a small subset of the data. It then extrapolates these importance scores to the entire dataset using one of two proposed strategies: one based on k-nearest neighbors (KNN) and the other on a graph neural network (GNN). Both strategies leverage embeddings from a model trained exclusively on the small subset. The authors evaluate their approach on four large-scale image datasets, covering supervised learning, unsupervised learning, and adversarial training scenarios. There are comprehensive experiments compare the method against state-of-the-art pruning techniques and simple baselines. Results demonstrate that the proposed approach significantly reduces computational cost while achieving accuracy that matches or closely approaches that of existing methods.", "review_text": "This paper introduces a novel and efficient approach to data pruning. The method first estimates the importance of each training example by analyzing only a small subset of the data. It then extrapolates these importance scores to the entire dataset using one of two proposed strategies: one based on k-nearest neighbors (KNN) and the other on a graph neural network (GNN). Both strategies leverage embeddings from a model trained exclusively on the small subset. The authors evaluate their approach on four large-scale image datasets, covering supervised learning, unsupervised learning, and adversarial training scenarios. There are comprehensive experiments compare the method against state-of-the-art pruning techniques and simple baselines. Results demonstrate that the proposed approach significantly reduces computational cost while achieving accuracy that matches or closely approaches that of existing methods.", "strengths": "- The paper tackles a practically significant and under-addressed problem: how to make computationally expensive data pruning methods tractable for large-scale training by requiring only a small subset for direct score computation.\n- The proposed score extrapolation framework is methodologically interesting and is instantiated with both a simple, transparent KNN approach and a more expressive, message-passing-based GNN, allowing for a clear analysis of trade-offs.\n- Empirical validation is thorough across multiple datasets—CIFAR-10, CIFAR-100, Places-365, and ImageNet—and considers a range of pruning rates, tasks (supervised, unsupervised, adversarial), and multiple pruning methods (DU and TDDS).\n- The efficiency benefit is systematically explored, with computing time and accuracy jointly visualized. As demonstrated in **Table 1**, the proposed KNN and GNN approaches recover a substantial portion of the original pruning benefit at a fraction of the computational cost, achieving notable speedups.", "weaknesses": "1. **Limited theoretical justification and over-reliance on local linearity assumptions:** \n   The primary mathematical support for extrapolation is drawn from influence function and local linearity arguments (Section 3). Yet, there is insufficient theoretical development or empirical diagnosis regarding the validity of these assumptions for highly nonlinear, high-dimensional representation spaces found in deep learning. As such, generalizability of the approach to broader architectures/tasks remains open.\n2. **Score oversmoothing and limited modeling of complex distributions:** A main weakness shows up in **Figure 5** (Appendix C.2). Both the KNN and GNN methods tend to oversmooth the extrapolated importance scores. They miss the multiple peaks that exist in the original score distribution. This makes the approach less effective when data importance isn’t uniform—for example, when there are clearly two different groups of samples with different levels of importance. It can also lead to consistent ranking mistakes, especially for outliers or samples that are hard to classify. The paper notes these issues, but it doesn’t fix them with new experiments or method changes.\n3. **Modest accuracy improvement and moderate correlation, especially at small subset sizes:** The extrapolation quality drops when the subset is too small or when a lot of data is pruned. This holds true both for how well the scores match the true importance and for the final model accuracy.  In many cases, even the best extrapolation method doesn’t fully match the performance of the original pruning approach. The results also change noticeably depending on the size of the subset used.  The method does offer real speedups. However, it doesn’t always give a better balance between accuracy and resource use. In settings with very tight resource limits, this could be a serious limitation.\n4. **Mathematical construction and empirical transparency:** \n   Some mathematical steps and notation—in particular, the loss formulations for the GNN extrapolation and details of embedding construction—are not explained in enough depth (e.g., Section 3, Equation for GNN loss). The explanation of exactly how node features combine (what happens with missing labels in unsupervised, for instance), and the optimization details, are relegated to appendices and could benefit from greater clarity in the main paper.\n5. **Potential failure cases insufficiently investigated:** \n   The brief exploration of failure modes (e.g., rank differences related to background/outliers in **Figure 5**) is insightful but limited. More systematic error analysis—including quantifying what data characteristics lead to the highest extrapolation mis-rankings or when the approach could harm downstream performance—is warranted.", "questions": "1. Could the authors elaborate on the limitations of the local linearity assumption in embedding space? Specifically, how does this assumption break down for highly heterogeneous data distributions, and are there diagnostics or empirical controls to quantify this risk in practice?\n2. In GNN extrapolation, how are class labels handled for semi-supervised or unsupervised settings, especially when no reliable pseudo-labels are available? Would the approach degrade under severe class imbalance or label noise?\n4. What strategies do the authors propose to mitigate oversmoothing and better capture multimodal or long-tailed importance score distributions in future work?\n5. How does the extrapolation performance vary with increasing/heterogeneous dataset size (e.g., simulated “billion-sample” settings), or for modalities outside vision (e.g., text, multimodal tasks)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel and efficient approach to data pruning. The method first estimates the importance of each training example by analyzing only a small subset of the data. It then extrapolates these importance scores to the entire dataset using one of two proposed strategies: one based on k-nearest neighbors (KNN) and the other on a graph neural network (GNN). Both strategies leverage embeddings from a model trained exclusively on the small subset. The authors evaluate their approach on four large-scale image datasets, covering supervised learning, unsupervised learning, and adversarial training scenarios. There are comprehensive experiments compare the method against state-of-the-art pruning techniques and simple baselines. Results demonstrate that the proposed approach significantly reduces computational cost while achieving accuracy that matches or closely approaches that of existing methods.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "- The paper tackles a practically significant and under-addressed problem: how to make computationally expensive data pruning methods tractable for large-scale training by requiring only a small subset for direct score computation.\n- The proposed score extrapolation framework is methodologically interesting and is instantiated with both a simple, transparent KNN approach and a more expressive, message-passing-based GNN, allowing for a clear analysis of trade-offs.\n- Empirical validation is thorough across multiple datasets—CIFAR-10, CIFAR-100, Places-365, and ImageNet—and considers a range of pruning rates, tasks (supervised, unsupervised, adversarial), and multiple pruning methods (DU and TDDS).\n- The efficiency benefit is systematically explored, with computing time and accuracy jointly visualized. As demonstrated in **Table 1**, the proposed KNN and GNN approaches recover a substantial portion of the original pruning benefit at a fraction of the computational cost, achieving notable speedups.", "weaknesses": "1. **Limited theoretical justification and over-reliance on local linearity assumptions:** \n   The primary mathematical support for extrapolation is drawn from influence function and local linearity arguments (Section 3). Yet, there is insufficient theoretical development or empirical diagnosis regarding the validity of these assumptions for highly nonlinear, high-dimensional representation spaces found in deep learning. As such, generalizability of the approach to broader architectures/tasks remains open.\n2. **Score oversmoothing and limited modeling of complex distributions:** A main weakness shows up in **Figure 5** (Appendix C.2). Both the KNN and GNN methods tend to oversmooth the extrapolated importance scores. They miss the multiple peaks that exist in the original score distribution. This makes the approach less effective when data importance isn’t uniform—for example, when there are clearly two different groups of samples with different levels of importance. It can also lead to consistent ranking mistakes, especially for outliers or samples that are hard to classify. The paper notes these issues, but it doesn’t fix them with new experiments or method changes.\n3. **Modest accuracy improvement and moderate correlation, especially at small subset sizes:** The extrapolation quality drops when the subset is too small or when a lot of data is pruned. This holds true both for how well the scores match the true importance and for the final model accuracy.  In many cases, even the best extrapolation method doesn’t fully match the performance of the original pruning approach. The results also change noticeably depending on the size of the subset used.  The method does offer real speedups. However, it doesn’t always give a better balance between accuracy and resource use. In settings with very tight resource limits, this could be a serious limitation.\n4. **Mathematical construction and empirical transparency:** \n   Some mathematical steps and notation—in particular, the loss formulations for the GNN extrapolation and details of embedding construction—are not explained in enough depth (e.g., Section 3, Equation for GNN loss). The explanation of exactly how node features combine (what happens with missing labels in unsupervised, for instance), and the optimization details, are relegated to appendices and could benefit from greater clarity in the main paper.\n5. **Potential failure cases insufficiently investigated:** \n   The brief exploration of failure modes (e.g., rank differences related to background/outliers in **Figure 5**) is insightful but limited. More systematic error analysis—including quantifying what data characteristics lead to the highest extrapolation mis-rankings or when the approach could harm downstream performance—is warranted.", "questions": "1. Could the authors elaborate on the limitations of the local linearity assumption in embedding space? Specifically, how does this assumption break down for highly heterogeneous data distributions, and are there diagnostics or empirical controls to quantify this risk in practice?\n2. In GNN extrapolation, how are class labels handled for semi-supervised or unsupervised settings, especially when no reliable pseudo-labels are available? Would the approach degrade under severe class imbalance or label noise?\n4. What strategies do the authors propose to mitigate oversmoothing and better capture multimodal or long-tailed importance score distributions in future work?\n5. How does the extrapolation performance vary with increasing/heterogeneous dataset size (e.g., simulated “billion-sample” settings), or for modalities outside vision (e.g., text, multimodal tasks)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761896642460}, {"id": "z2ivNxfG0d", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21707/Reviewer_VYC6"], "rating": 4, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "Traditional pruning methods require training computing importance scores on all the samples of the dataset. Computing these scores often requires training a model for multiple epochs on the samples, which can become costly. This work proposes to compute the scores (and train an initial model) on only a small fraction of the samples and extrapolate to the rest of the dataset. To do so, they propose two different technics, one based on K-Nearest Neighbours and the other training a Graph Neural Network, both using the geometry induced by the embedding space.", "review_text": "Traditional pruning methods require training computing importance scores on all the samples of the dataset. Computing these scores often requires training a model for multiple epochs on the samples, which can become costly. This work proposes to compute the scores (and train an initial model) on only a small fraction of the samples and extrapolate to the rest of the dataset. To do so, they propose two different technics, one based on K-Nearest Neighbours and the other training a Graph Neural Network, both using the geometry induced by the embedding space.", "strengths": "In general I found this work interesting, it tackles a less explored direction in data pruning and can indeed bring valuable computational gains.\n- The notion of extrapolating importance scores from a small subset is simple but original (to my knowledge), and it provides a new angle on making pruning efficient. \n- The KNN and GNN approaches effectively demonstrate that extrapolation can cut computation time with little performance loss and form a good proof of concept.\n- The work evaluates multiple datasets, and training settings, which supports the generality of the approach.", "weaknesses": "Overall, the paper presents an interesting idea, though I think it could be strengthened with some further development and analysis.\n- I did not find the theoretical justification very convincing. It seems that the main point is about the smooth interpolation of the samples influence that the authors use as a justification for their extrapolated scores (eq 6). But in the context of influence the extrapolated point is itself a convex interpolation of the reference points, and the weights are the associated factors, whereas in eq 6 we are assigning scores using an arbitrary formula depending on the distance to the reference points. Also the authors could introduce what is influence and why it would be relevant here.\n- I think the paper could benefit from a comparison with other simple alternatives to reduce computations (see questions). \n- Dynamic data pruning methods are not considered in this work, though they are very popular and somehow connected to this since they face the problem of updating the scores and can do it on only a fraction of the samples every time to avoid prohibitive costs.", "questions": "Multiple \"simple\" alternatives could be compared to your proposed approach to strengthen the contribution\n- Instead of training completely on the whole dataset, how would the proposed approach compare to training only a few epochs and use this intermediate model to score, then keep training the same model with only the selected fraction, which is often what is done in practice. \n- Could one use directly your model $\\mathcal{F}_s$ trained on the selected subset and score the unseen points without using extrapolation ?\n- In table 1, how would you compare to using the \"ground truth scores\" for the selected subset and random for the unseen, without needing extrapolation ?\n\n- In line 152-153, the authors write that the trained model $\\mathcal{F}_s$ depends on the selected extrapolation method, but from lines 167-172 it seems to be independent of the extrapolation method, could you clarify ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Traditional pruning methods require training computing importance scores on all the samples of the dataset. Computing these scores often requires training a model for multiple epochs on the samples, which can become costly. This work proposes to compute the scores (and train an initial model) on only a small fraction of the samples and extrapolate to the rest of the dataset. To do so, they propose two different technics, one based on K-Nearest Neighbours and the other training a Graph Neural Network, both using the geometry induced by the embedding space.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "In general I found this work interesting, it tackles a less explored direction in data pruning and can indeed bring valuable computational gains.\n- The notion of extrapolating importance scores from a small subset is simple but original (to my knowledge), and it provides a new angle on making pruning efficient. \n- The KNN and GNN approaches effectively demonstrate that extrapolation can cut computation time with little performance loss and form a good proof of concept.\n- The work evaluates multiple datasets, and training settings, which supports the generality of the approach.", "weaknesses": "Overall, the paper presents an interesting idea, though I think it could be strengthened with some further development and analysis.\n- I did not find the theoretical justification very convincing. It seems that the main point is about the smooth interpolation of the samples influence that the authors use as a justification for their extrapolated scores (eq 6). But in the context of influence the extrapolated point is itself a convex interpolation of the reference points, and the weights are the associated factors, whereas in eq 6 we are assigning scores using an arbitrary formula depending on the distance to the reference points. Also the authors could introduce what is influence and why it would be relevant here.\n- I think the paper could benefit from a comparison with other simple alternatives to reduce computations (see questions). \n- Dynamic data pruning methods are not considered in this work, though they are very popular and somehow connected to this since they face the problem of updating the scores and can do it on only a fraction of the samples every time to avoid prohibitive costs.", "questions": "Multiple \"simple\" alternatives could be compared to your proposed approach to strengthen the contribution\n- Instead of training completely on the whole dataset, how would the proposed approach compare to training only a few epochs and use this intermediate model to score, then keep training the same model with only the selected fraction, which is often what is done in practice. \n- Could one use directly your model $\\mathcal{F}_s$ trained on the selected subset and score the unseen points without using extrapolation ?\n- In table 1, how would you compare to using the \"ground truth scores\" for the selected subset and random for the unseen, without needing extrapolation ?\n\n- In line 152-153, the authors write that the trained model $\\mathcal{F}_s$ depends on the selected extrapolation method, but from lines 167-172 it seems to be independent of the extrapolation method, could you clarify ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761740878197}, {"id": "S5ZyEFsP0F", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission21707/Reviewer_WNwP"], "rating": 4, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "This paper introduces a novel score extrapolation framework that eliminates the need for full training on the original dataset before pruning. The framework trains on a subset of the original data and extrapolates scores using k-nearest neighbors (KNN) and graph neural networks (GNN) based on the trained subset's scores. The method is validated across various datasets and training paradigms based on two different scoring methods. Their approach achieves Pareto-optimal time-accuracy trade-offs and significantly reduces computational effort.", "review_text": "This paper introduces a novel score extrapolation framework that eliminates the need for full training on the original dataset before pruning. The framework trains on a subset of the original data and extrapolates scores using k-nearest neighbors (KNN) and graph neural networks (GNN) based on the trained subset's scores. The method is validated across various datasets and training paradigms based on two different scoring methods. Their approach achieves Pareto-optimal time-accuracy trade-offs and significantly reduces computational effort.", "strengths": "1. The paper's main contribution addresses a key challenge in existing data pruning methods: ironically, these methods take longer to prune and train on a pruned dataset than to train fully on the original dataset. This paper mitigates this issue by calculating scores only on a subset of data points and extrapolating scores for the remaining data, rather than computing scores for every data point. This methodology arises from a strong motivation and solves the problem in a reasonable way.\n2. The authors tested their score extrapolation methods on various datasets (Places365, ImageNet, Synthetic CIFAR-100, and CIFAR-10). They experimented across multiple settings—including supervised, unsupervised, and adversarial training—to investigate the applicability of score extrapolation methods. They also conducted various ablation studies, such as examining the correlation with original scores depending on initial sample size.\n3. Importantly, the authors include failure cases in Appendix C.2 to highlight the current limitations of their framework, demonstrating transparency and a careful evaluation of their method.", "weaknesses": "1. Section 3 lacks the clarity necessary for comprehension and reproducibility. The main issues are as follows:\n    1. What is $D$ in line 166 refer to? The notation should be clearly defined. Moreover, the symbol $D$ overlaps with the distance notation $D(\\cdot, \\cdot)$ used in line 210, which may cause confusion.\n    2. In line 178, the authors defined the embedding function $\\phi: \\mathcal{X} \\to \\mathbb{R}^d$ that maps an input $x$ to embedding $z = \\phi(x)$. \n        - First, $x \\in \\mathcal{X}$ should be explicitly stated.\n        - Second, in lines 200~212 (*the Extrapolation with KNN paragraph*), the model is defined as $\\mathcal{F}_s: \\mathbb{R}^d \\to \\mathbb{R}^{d'}$. Here, $\\mathbb{R}^d$ is used as the input dimension and $\\mathbb{R}^{d'}$ as the output dimension, which differs from the earlier definition.\n        - Moreover, $\\mathcal{F}_s$ is described as an embedding function (outputting embeddings) rather than a neural network that outputs logits, which contradicts the definition in line 178. The authors should consistently use $\\phi_s$ to denote the embedding function for clarity and consistency.\n    3. In Equation (2), $y$ should be explicitly defined, including its dimensionality.\n    4. In lines 171 and 175, the authors define $S_s$ and $S_r$ as vectors. However, in Equation (6), the same notation $S$ is used both as a function ($S_{knn}$), and as a scalar ($S_{\\pi_i(x)}$). This inconsistency becomes more pronounced in lines 230-237, where $S_s$ and $S$ are again treated as functions. If $S$ is intended to represent a “score”, its form should be clearly and consistently defined to avoid confusion for readers. \n    5. The authors should explicitly state that $\\pi_i(x) \\in [m]$, indicating that the index of the $i$-th nearest neighbor of $x$ is selected from within $\\mathbb{D}_s$.\n    6. In line 221, $d(\\cdot, \\cdot)$ is used as a distance metric, which is inconsistent with the notation $D(\\cdot, \\cdot)$ defined in line 210-211.\n    7. In Equation (7), the authors use $\\sum_{x_i \\in \\mathbb{D}\\_s}$, but the term $\\mathcal{F}\\_\\mathcal{G}(\\mathcal{A}, \\mathcal{V};\\theta)\\_i$ explicitly depends on the index $i$. Therefore, I recommend replacing $\\sum\\_{x\\_i \\in \\mathbb{D}\\_s}$ with $\\sum\\_{i \\in [m]}$, since the authors have already defined $\\mathbb{D}\\_s = \\lbrace x_i \\rbrace_{i \\in [m]}$.\n    \n    Overall, the Section 3 should be rewritten for clarity, consistency, and reproducibility. The notations should be clearly defined and used consistently, and any ambiguous or overlapping definitions should be eliminated.\n    \n2. Comparison with recent SOTA baselines could further strengthen the proposed method:\n    1. For example, the DUAL method (with its $\\beta$-sampling scheme) explicitly targets reducing the score-computation time by jointly considering data uncertainty and difficulty [1].  Since your paper focuses on effective data pruning, this baseline appears directly relevant. It would be valuable either to compare your proposed method with DUAL, or to incorporate DUAL’s score as a base for your method, potentially improving overall pruning efficiency when combined with its sampling strategy.\n    2. Also, methods such as Coverage-centric Coreset Selection (CCS) [2] and $\\mathbb{D}^2$ [3] involve more expensive score-computation but achieve very competitive performance. These could serve as strong base scores or benchmarks for your method, especially if you aim to show improvement in computation vs. performance trade-off.\n    \n    To adopt the sampling methods from [1] and [2], one can first compute subset scores, extrapolate them to estimate the unevaluated samples, and then apply sampling based on these extrapolated scores.\n    \n3. Are there specific reasons for using the *Synthetic CIFAR-100* dataset instead of the standard CIFAR-100? Since the abstract (line 21) mentions CIFAR-100, readers may be confused when encountering Synthetic CIFAR-100 in the main text. As a reader unfamiliar with this dataset, I believe the paper should clearly explain:\n    1. What this dataset is, including how it relates to the original CIFAR-100; and\n    2. Why the authors chose to use this dataset instead of the standard CIFAR-100.\n    \n    My understanding is that the authors used the Synthetic CIFAR-100 dataset because the score extrapolation process requires computing scores on a subset first, and the absolute size of the dataset plays a critical role in this step. Nevertheless, I believe readers will still be interested in the model’s performance on the original CIFAR-100, especially since results on the original CIFAR-10 are already provided. Therefore, I recommend including the original CIFAR-100 results (at least in the appendix or as a stated limitation).\n    \n4. Why are there no data pruning results in the unsupervised setting? Currently, only Pearson and Spearman correlation results are reported for the CIFAR-10 dataset in Table 3. Since the scores have already been computed, it should not be difficult to perform dataset pruning based on these scores and then train the pruned subset in a supervised manner. Given that the paragraph is titled *“Unsupervised Data Pruning”,* readers (including myself) would naturally be interested in seeing the performance of a pruned subset whose scores were calculated in an unsupervised manner. I therefore recommend that the authors include the performance results of such a pruned subset.\n5. As stated in Appendix C.2, there are certain datasets for which the score extrapolation method fails to accurately capture the original score distribution. Moreover, as shown in Figure 4 and Table 2, the correlations with the original scores are not particularly high. This issue is also reflected in the pruned dataset performance presented in Figure 2, where the pruning results are not significantly better than those of other pruning methods. Taken together, these observations suggest that the proposed approach behaves more like an interpolation between random pruning and full score computation, rather than a method that resolves the trade-off between computation and performance. In this regard, it would be helpful to include the results of random pruning in Figure 3 for comparison. Furthermore, I recommend that the authors to include the results of EL2N [4] and DUAL [1], which are both time-efficient data pruning methods.\n\n---\n\n**References:**\n\n[1] Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty, ICML 2025.\n\n[2] Coverage-centric Coreset Selection for High Pruning Rates, ICLR 2023.\n\n[3] D2 Pruning: Message Passing for Balancing Diversity and Difficulty in Data Pruning, ICLR 2024.\n\n[4] Deep Learning on a Data Diet: Finding Important Examples Early in Training, NeurIPS 2021.", "questions": "1. Why do the SSP and Zcoreset methods, which are training-free, take so much time to create a subset in Figure 3?\n2. What does each point in Figure 3 represent? How were these points determined? If each point corresponds to a different pruning ratio, this should be clarified. Similarly, what does each point in Figure 4 represent? It seems likely proportional to the subset size used for score calculation, but this needs explanation.\n3. When should KNN be used versus GNN? Which method is appropriate under which circumstances, and why? (Also, clarification is needed for lines 412–413. What does \"This might be ~\" mean exactly?)\n4. In the *Adversarial Training* paragraph, as a reader unfamiliar with adversarial training, it is difficult to understand the metrics used in Table 4 and Table 5. The authors should clarify the definitions of “Robustness” and “Clean Accuracy.”\n5. What if we train an independent model that maps embeddings to subset scores? Specifically, up to Step 3 in Figure 1, the process remains the same, but for the extrapolation step, we train a small neural network (e.g., an MLP) on the subset where the scores have been computed, and then evaluate it on the remaining data to predict their scores. How would such an approach perform? \n6. As far as I understand, the proposed method requires labeling only the initial subset when pruning the dataset, even in the supervised setting. Is that correct? Since only the embeddings of the remaining data are used when extrapolating the scores, if the pruning can indeed be performed without access to all labels, I think the authors should highlight this point. For example: “Our method requires labels only for the initial subset, thereby reducing labeling costs and enabling application to large-scale web datasets.”\n7. There appears to be overlapping information among Table 1, Figure 2, and Figure 3. The authors might consider merging or simplifying these to avoid redundancy and improve readability.\n8. Is there a specific reason for reporting *relative accuracy* in Table 1? It might be more intuitive to present the *actual accuracy* values while expressing *time efficiency* in relative terms, as this would make the comparison easier to interpret.\n\n---\n**Minor Corrections:**\n\n1. Line 149: data sets → datasets\n2. Figure 2(c) and Figure 6(c): Synthic → Synthetic\n---\nIf the authors adequately address the issues raised in the Weaknesses and Questions sections, I would consider increasing my overall score.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel score extrapolation framework that eliminates the need for full training on the original dataset before pruning. The framework trains on a subset of the original data and extrapolates scores using k-nearest neighbors (KNN) and graph neural networks (GNN) based on the trained subset's scores. The method is validated across various datasets and training paradigms based on two different scoring methods. Their approach achieves Pareto-optimal time-accuracy trade-offs and significantly reduces computational effort.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The paper's main contribution addresses a key challenge in existing data pruning methods: ironically, these methods take longer to prune and train on a pruned dataset than to train fully on the original dataset. This paper mitigates this issue by calculating scores only on a subset of data points and extrapolating scores for the remaining data, rather than computing scores for every data point. This methodology arises from a strong motivation and solves the problem in a reasonable way.\n2. The authors tested their score extrapolation methods on various datasets (Places365, ImageNet, Synthetic CIFAR-100, and CIFAR-10). They experimented across multiple settings—including supervised, unsupervised, and adversarial training—to investigate the applicability of score extrapolation methods. They also conducted various ablation studies, such as examining the correlation with original scores depending on initial sample size.\n3. Importantly, the authors include failure cases in Appendix C.2 to highlight the current limitations of their framework, demonstrating transparency and a careful evaluation of their method.", "weaknesses": "1. Section 3 lacks the clarity necessary for comprehension and reproducibility. The main issues are as follows:\n    1. What is $D$ in line 166 refer to? The notation should be clearly defined. Moreover, the symbol $D$ overlaps with the distance notation $D(\\cdot, \\cdot)$ used in line 210, which may cause confusion.\n    2. In line 178, the authors defined the embedding function $\\phi: \\mathcal{X} \\to \\mathbb{R}^d$ that maps an input $x$ to embedding $z = \\phi(x)$. \n        - First, $x \\in \\mathcal{X}$ should be explicitly stated.\n        - Second, in lines 200~212 (*the Extrapolation with KNN paragraph*), the model is defined as $\\mathcal{F}_s: \\mathbb{R}^d \\to \\mathbb{R}^{d'}$. Here, $\\mathbb{R}^d$ is used as the input dimension and $\\mathbb{R}^{d'}$ as the output dimension, which differs from the earlier definition.\n        - Moreover, $\\mathcal{F}_s$ is described as an embedding function (outputting embeddings) rather than a neural network that outputs logits, which contradicts the definition in line 178. The authors should consistently use $\\phi_s$ to denote the embedding function for clarity and consistency.\n    3. In Equation (2), $y$ should be explicitly defined, including its dimensionality.\n    4. In lines 171 and 175, the authors define $S_s$ and $S_r$ as vectors. However, in Equation (6), the same notation $S$ is used both as a function ($S_{knn}$), and as a scalar ($S_{\\pi_i(x)}$). This inconsistency becomes more pronounced in lines 230-237, where $S_s$ and $S$ are again treated as functions. If $S$ is intended to represent a “score”, its form should be clearly and consistently defined to avoid confusion for readers. \n    5. The authors should explicitly state that $\\pi_i(x) \\in [m]$, indicating that the index of the $i$-th nearest neighbor of $x$ is selected from within $\\mathbb{D}_s$.\n    6. In line 221, $d(\\cdot, \\cdot)$ is used as a distance metric, which is inconsistent with the notation $D(\\cdot, \\cdot)$ defined in line 210-211.\n    7. In Equation (7), the authors use $\\sum_{x_i \\in \\mathbb{D}\\_s}$, but the term $\\mathcal{F}\\_\\mathcal{G}(\\mathcal{A}, \\mathcal{V};\\theta)\\_i$ explicitly depends on the index $i$. Therefore, I recommend replacing $\\sum\\_{x\\_i \\in \\mathbb{D}\\_s}$ with $\\sum\\_{i \\in [m]}$, since the authors have already defined $\\mathbb{D}\\_s = \\lbrace x_i \\rbrace_{i \\in [m]}$.\n    \n    Overall, the Section 3 should be rewritten for clarity, consistency, and reproducibility. The notations should be clearly defined and used consistently, and any ambiguous or overlapping definitions should be eliminated.\n    \n2. Comparison with recent SOTA baselines could further strengthen the proposed method:\n    1. For example, the DUAL method (with its $\\beta$-sampling scheme) explicitly targets reducing the score-computation time by jointly considering data uncertainty and difficulty [1].  Since your paper focuses on effective data pruning, this baseline appears directly relevant. It would be valuable either to compare your proposed method with DUAL, or to incorporate DUAL’s score as a base for your method, potentially improving overall pruning efficiency when combined with its sampling strategy.\n    2. Also, methods such as Coverage-centric Coreset Selection (CCS) [2] and $\\mathbb{D}^2$ [3] involve more expensive score-computation but achieve very competitive performance. These could serve as strong base scores or benchmarks for your method, especially if you aim to show improvement in computation vs. performance trade-off.\n    \n    To adopt the sampling methods from [1] and [2], one can first compute subset scores, extrapolate them to estimate the unevaluated samples, and then apply sampling based on these extrapolated scores.\n    \n3. Are there specific reasons for using the *Synthetic CIFAR-100* dataset instead of the standard CIFAR-100? Since the abstract (line 21) mentions CIFAR-100, readers may be confused when encountering Synthetic CIFAR-100 in the main text. As a reader unfamiliar with this dataset, I believe the paper should clearly explain:\n    1. What this dataset is, including how it relates to the original CIFAR-100; and\n    2. Why the authors chose to use this dataset instead of the standard CIFAR-100.\n    \n    My understanding is that the authors used the Synthetic CIFAR-100 dataset because the score extrapolation process requires computing scores on a subset first, and the absolute size of the dataset plays a critical role in this step. Nevertheless, I believe readers will still be interested in the model’s performance on the original CIFAR-100, especially since results on the original CIFAR-10 are already provided. Therefore, I recommend including the original CIFAR-100 results (at least in the appendix or as a stated limitation).\n    \n4. Why are there no data pruning results in the unsupervised setting? Currently, only Pearson and Spearman correlation results are reported for the CIFAR-10 dataset in Table 3. Since the scores have already been computed, it should not be difficult to perform dataset pruning based on these scores and then train the pruned subset in a supervised manner. Given that the paragraph is titled *“Unsupervised Data Pruning”,* readers (including myself) would naturally be interested in seeing the performance of a pruned subset whose scores were calculated in an unsupervised manner. I therefore recommend that the authors include the performance results of such a pruned subset.\n5. As stated in Appendix C.2, there are certain datasets for which the score extrapolation method fails to accurately capture the original score distribution. Moreover, as shown in Figure 4 and Table 2, the correlations with the original scores are not particularly high. This issue is also reflected in the pruned dataset performance presented in Figure 2, where the pruning results are not significantly better than those of other pruning methods. Taken together, these observations suggest that the proposed approach behaves more like an interpolation between random pruning and full score computation, rather than a method that resolves the trade-off between computation and performance. In this regard, it would be helpful to include the results of random pruning in Figure 3 for comparison. Furthermore, I recommend that the authors to include the results of EL2N [4] and DUAL [1], which are both time-efficient data pruning methods.\n\n---\n\n**References:**\n\n[1] Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty, ICML 2025.\n\n[2] Coverage-centric Coreset Selection for High Pruning Rates, ICLR 2023.\n\n[3] D2 Pruning: Message Passing for Balancing Diversity and Difficulty in Data Pruning, ICLR 2024.\n\n[4] Deep Learning on a Data Diet: Finding Important Examples Early in Training, NeurIPS 2021.", "questions": "1. Why do the SSP and Zcoreset methods, which are training-free, take so much time to create a subset in Figure 3?\n2. What does each point in Figure 3 represent? How were these points determined? If each point corresponds to a different pruning ratio, this should be clarified. Similarly, what does each point in Figure 4 represent? It seems likely proportional to the subset size used for score calculation, but this needs explanation.\n3. When should KNN be used versus GNN? Which method is appropriate under which circumstances, and why? (Also, clarification is needed for lines 412–413. What does \"This might be ~\" mean exactly?)\n4. In the *Adversarial Training* paragraph, as a reader unfamiliar with adversarial training, it is difficult to understand the metrics used in Table 4 and Table 5. The authors should clarify the definitions of “Robustness” and “Clean Accuracy.”\n5. What if we train an independent model that maps embeddings to subset scores? Specifically, up to Step 3 in Figure 1, the process remains the same, but for the extrapolation step, we train a small neural network (e.g., an MLP) on the subset where the scores have been computed, and then evaluate it on the remaining data to predict their scores. How would such an approach perform? \n6. As far as I understand, the proposed method requires labeling only the initial subset when pruning the dataset, even in the supervised setting. Is that correct? Since only the embeddings of the remaining data are used when extrapolating the scores, if the pruning can indeed be performed without access to all labels, I think the authors should highlight this point. For example: “Our method requires labels only for the initial subset, thereby reducing labeling costs and enabling application to large-scale web datasets.”\n7. There appears to be overlapping information among Table 1, Figure 2, and Figure 3. The authors might consider merging or simplifying these to avoid redundancy and improve readability.\n8. Is there a specific reason for reporting *relative accuracy* in Table 1? It might be more intuitive to present the *actual accuracy* values while expressing *time efficiency* in relative terms, as this would make the comparison easier to interpret.\n\n---\n**Minor Corrections:**\n\n1. Line 149: data sets → datasets\n2. Figure 2(c) and Figure 6(c): Synthic → Synthetic\n---\nIf the authors adequately address the issues raised in the Weaknesses and Questions sections, I would consider increasing my overall score.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1761553596410}], "openreview_url": "https://openreview.net/forum?id=ycxzArIvgF", "arxiv_id": "2506.09010", "paper_pdf": "papers/ycxzArIvgF.pdf", "paper_pdf_sha256": "31d3ed279e251af304d61c1ef614fac09379f1030a640da367a915bbd751b19d", "paper_pdf_bytes": 4263230, "paper_pdf_source": "openreview", "code_url": "https://github.com/prasangadhungel/Data-Pruning-with-Extrapolated-Scores", "code_repository": "prasangadhungel/Data-Pruning-with-Extrapolated-Scores", "code_commit": "655e48fbb48e5856da9052808222388cfc6d99f2", "code_archive": "repos/ycxzArIvgF.zip", "code_archive_sha256": "228a0a556f1c6d16a13b65163b1b9b76ed08d5fa42ec4bd2a2a58ee2e18ebb78", "code_archive_bytes": 112358, "code_file_count": 34, "code_extensions": {".py": 32, ".ipynb": 2}, "github_disk_usage_kb": 9833, "github_languages": {"Python": 267787, "Jupyter Notebook": 98163, "Makefile": 674}, "github_archived": false, "github_pushed_at": "2026-07-29T10:58:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/effective-data-pruning-through-score"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "LC2KxRwC3n", "year": 2025, "status": "rejected", "title": "A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders", "authors": ["David Chanin", "James Wilken-Smith", "Tomáš Dulka", "Hardik Bhatnagar", "Joseph Isaac Bloom"], "authorids": ["~David_Chanin1", "~James_Wilken-Smith1", "~Tomáš_Dulka1", "~Hardik_Bhatnagar1", "~Joseph_Isaac_Bloom1"], "authors_source": "OpenReview API", "abstract": "Sparse Autoencoders (SAEs) have emerged as a promising approach to decompose the activations of Large Language Models (LLMs) into human-interpretable latents. In this paper, we pose two questions. First, to what extent do SAEs extract monosemantic and interpretable latents? Second, to what extent does varying the sparsity or the size of the SAE affect monosemanticity / interpretability? By investigating these questions in the context of a simple first-letter identification task where we have complete access to ground truth labels for all tokens in the vocabulary, we are able to provide more detail than prior investigations. Critically, we identify a problematic form of feature-splitting we call \"feature absorption\" where seemingly monosemantic latents fail to fire in cases where they clearly should. Our investigation suggests that varying SAE size or sparsity is insufficient to solve this issue, and that there are deeper conceptual issues in need of resolution.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "bljzMQT9we", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8043/Reviewer_S6kU"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper explores the use of Sparse Autoencoders (SAEs) to decompose activations of Large Language Models (LLMs) into interpretable latent features. It addresses two main questions: the degree to which SAEs yield monosemantic (single-meaning) and interpretable latents, and how adjustments to the sparsity and size of SAEs impact interpretability. Through a controlled first-letter identification task with complete ground-truth labels, the authors find more nuanced insights compared to previous studies.", "review_text": "This paper explores the use of Sparse Autoencoders (SAEs) to decompose activations of Large Language Models (LLMs) into interpretable latent features. It addresses two main questions: the degree to which SAEs yield monosemantic (single-meaning) and interpretable latents, and how adjustments to the sparsity and size of SAEs impact interpretability. Through a controlled first-letter identification task with complete ground-truth labels, the authors find more nuanced insights compared to previous studies.", "strengths": "1. This paper studies an interesting problem, that is, learned sparse features can become less understandable by \"absorbing\" token-aligned features.\n2. understanding how width and sparsity of SAEs affects its training performance can be helpful for future SAE training.\n3. the paper is written in an easy-to-understand way.", "weaknesses": "1. It would be valuable to explore practical examples where insights from feature absorption can enhance interpretability. For instance, analyzing the first letters of all tokens associated with specific latent activations might offer new explanations for features that previously seemed opaque. This approach could reveal subtle patterns in latent activation, aiding in the interpretation of challenging features.\n2. This work presents a specific form of feature absorption, but it raises the question of whether other variations might exist. Are there additional contexts or scenarios where feature absorption manifests differently, potentially impacting interpretability in distinct ways? Identifying these cases could deepen our understanding of the phenomenon and refine the strategies needed to address it.", "questions": "Same as above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the use of Sparse Autoencoders (SAEs) to decompose activations of Large Language Models (LLMs) into interpretable latent features. It addresses two main questions: the degree to which SAEs yield monosemantic (single-meaning) and interpretable latents, and how adjustments to the sparsity and size of SAEs impact interpretability. Through a controlled first-letter identification task with complete ground-truth labels, the authors find more nuanced insights compared to previous studies.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. This paper studies an interesting problem, that is, learned sparse features can become less understandable by \"absorbing\" token-aligned features.\n2. understanding how width and sparsity of SAEs affects its training performance can be helpful for future SAE training.\n3. the paper is written in an easy-to-understand way.", "weaknesses": "1. It would be valuable to explore practical examples where insights from feature absorption can enhance interpretability. For instance, analyzing the first letters of all tokens associated with specific latent activations might offer new explanations for features that previously seemed opaque. This approach could reveal subtle patterns in latent activation, aiding in the interpretation of challenging features.\n2. This work presents a specific form of feature absorption, but it raises the question of whether other variations might exist. Are there additional contexts or scenarios where feature absorption manifests differently, potentially impacting interpretability in distinct ways? Identifying these cases could deepen our understanding of the phenomenon and refine the strategies needed to address it.", "questions": "Same as above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730808990195}, {"id": "yXqj4bZtkB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8043/Reviewer_1gVz"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper investigates the interpretability of the extracted latent/features by Sparse Autoencoders. The authors found that the Sparse Autoencoder does extract monosemantic and interpretable features and changing the hyperparameters of the Sparse Autoencoder could not eliminate this.", "review_text": "This paper investigates the interpretability of the extracted latent/features by Sparse Autoencoders. The authors found that the Sparse Autoencoder does extract monosemantic and interpretable features and changing the hyperparameters of the Sparse Autoencoder could not eliminate this.", "strengths": "As the large language model becomes increasingly popular, it is really important to get a good understanding of these techniques. This paper investigates a meaningful problem and attempts to provides some insights and points out some research directions, which should be very useful to the community.\n\n#---------------------------------------#\n\nAfter rebuttal, the authors add results with different SAE architectures and three different model types.", "weaknesses": "There are several weaknesses for this paper.\n1) The observation and conclusion are merely based on one model and one task (first-letter identification task). Thus, it is really difficult to say how general and convincing these conclusions could be. I am afraid it is not safe to draw a solid conclusion based a single setting.\n2)  The authors should elaborate their method better. The current version simply lists some background and then posts a pseudo code (algorithm) there. More descriptions should benefit readers to get a better understanding.\n3) Some contents cause confusing instead of help. For example, the Figure1 is really confusing and hard to interpreter and even worse, the authors do not provide sufficient descriptions of it in the main paper.\n4) The overall presentation is a little bit messy and it is really hard to follow the content.", "questions": "I list my concerns and questions in the \"weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper investigates the interpretability of the extracted latent/features by Sparse Autoencoders. The authors found that the Sparse Autoencoder does extract monosemantic and interpretable features and changing the hyperparameters of the Sparse Autoencoder could not eliminate this.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "As the large language model becomes increasingly popular, it is really important to get a good understanding of these techniques. This paper investigates a meaningful problem and attempts to provides some insights and points out some research directions, which should be very useful to the community.\n\n#---------------------------------------#\n\nAfter rebuttal, the authors add results with different SAE architectures and three different model types.", "weaknesses": "There are several weaknesses for this paper.\n1) The observation and conclusion are merely based on one model and one task (first-letter identification task). Thus, it is really difficult to say how general and convincing these conclusions could be. I am afraid it is not safe to draw a solid conclusion based a single setting.\n2)  The authors should elaborate their method better. The current version simply lists some background and then posts a pseudo code (algorithm) there. More descriptions should benefit readers to get a better understanding.\n3) Some contents cause confusing instead of help. For example, the Figure1 is really confusing and hard to interpreter and even worse, the authors do not provide sufficient descriptions of it in the main paper.\n4) The overall presentation is a little bit messy and it is really hard to follow the content.", "questions": "I list my concerns and questions in the \"weaknesses\" section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730705187596}, {"id": "pQfVwtcGNn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8043/Reviewer_4XB2"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "This paper proposes a new simplified setting for studying features discovered by Sparse Autoencoders (SAEs) trained on language model activations. It supposes that SAEs should discover features corresponding to the first letter of a given token, but finds that features which are predictive of a given letter occasionally fail and are 'absorbed' into the activation of another feature. They then measure and explore this phenomenon in the context of their setting.", "review_text": "This paper proposes a new simplified setting for studying features discovered by Sparse Autoencoders (SAEs) trained on language model activations. It supposes that SAEs should discover features corresponding to the first letter of a given token, but finds that features which are predictive of a given letter occasionally fail and are 'absorbed' into the activation of another feature. They then measure and explore this phenomenon in the context of their setting.", "strengths": "- Describes an interesting setting in which to study feature disentanglement for natural language.\n- Discovers a new phenomenon, 'feature absorbtion', which could pose a problem for desired future applications of Sparse Autoencoders.\n- Setting allows for a thorough analysis of the failure cases for feature absorbtion and feature splitting, and said analysis was good and informative.\n- Paper has a clear statement of the problem and good structure. Good use of the 'S'/'_short' example to help build intuition.", "weaknesses": "- It would be useful to explore the commonalities between different instances of absorbtion of SAE features, or otherwise find more ways to verify/expand upon the claim/statement that \"feature absorption is likely a logical consequence of SAE sparsity loss\".", "questions": "- Did you find any rules for predicting whether the feature activation for the first letter of a given word would be absorbed? (i.e. do you have any hypothesies/ways to differentiate situations in which feature absorbtion occurs or does not occur?)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new simplified setting for studying features discovered by Sparse Autoencoders (SAEs) trained on language model activations. It supposes that SAEs should discover features corresponding to the first letter of a given token, but finds that features which are predictive of a given letter occasionally fail and are 'absorbed' into the activation of another feature. They then measure and explore this phenomenon in the context of their setting.", "soundness": 4, "presentation": 3, "contribution": 4, "strengths": "- Describes an interesting setting in which to study feature disentanglement for natural language.\n- Discovers a new phenomenon, 'feature absorbtion', which could pose a problem for desired future applications of Sparse Autoencoders.\n- Setting allows for a thorough analysis of the failure cases for feature absorbtion and feature splitting, and said analysis was good and informative.\n- Paper has a clear statement of the problem and good structure. Good use of the 'S'/'_short' example to help build intuition.", "weaknesses": "- It would be useful to explore the commonalities between different instances of absorbtion of SAE features, or otherwise find more ways to verify/expand upon the claim/statement that \"feature absorption is likely a logical consequence of SAE sparsity loss\".", "questions": "- Did you find any rules for predicting whether the feature activation for the first letter of a given word would be absorbed? (i.e. do you have any hypothesies/ways to differentiate situations in which feature absorbtion occurs or does not occur?)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730672453732}, {"id": "M0whZGYuyQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8043/Reviewer_x2j6"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper documents and measures a particular form of feature splitting in the Gemma Scope family of sparse autoencoders (SAEs), which the authors call feature absorption. The prototypical case of feature absorption studied is when a latent is a high-precision classifier for a property like \"starts with the letter L\", but has poor recall due to other latents \"absorbing\" individual tokens, so that the main latent's activations are given by the rule \"starts with the letter L, except for the tokens laser, lions, [several other exceptions]\".", "review_text": "This paper documents and measures a particular form of feature splitting in the Gemma Scope family of sparse autoencoders (SAEs), which the authors call feature absorption. The prototypical case of feature absorption studied is when a latent is a high-precision classifier for a property like \"starts with the letter L\", but has poor recall due to other latents \"absorbing\" individual tokens, so that the main latent's activations are given by the rule \"starts with the letter L, except for the tokens laser, lions, [several other exceptions]\".", "strengths": "The paper explores a natural task, first letter classification, and does so well. It establishes that this information is linearly available in the model (by training an LR probe with high F1-score), and that SAEs find latents that align with the probe direction (Figure 4b and especially Figure 13), and that these latents are often good classifiers, though not as good as probes (Figure 2a).\n\nThe paper provides a novel perspective to a documented issue of feature splitting in SAEs. It provides good theoretical justification of why feature absorption is likely to happen, and why it provides challenges to interpretability.", "weaknesses": "The paper could do more to demonstrate that feature absorption happens as described, namely that: \n1) There is a latent which is a high-precision classifier for \"starts with letter [x]\".\n2) There is a particular subset of tokens which start with this letter but which the latent fails to activate on.\n3) There are separate token-aligned latents that individually classify those tokens.\n4) If that subset of tokens is removed, the latent is a high-precision, high-recall classifier for \"starts with the letter [x], except for that subset of tokens\".\n\nThe authors demonstrate 1) and 2) well in sections 4.1 and 4.3 respectively, but never show 3) or 4) besides a single example in section 4.2. The authors could address 3) by adding an additional step to the experiment in section 4.3, in which they confirm that \"the SAE latent receiving the largest negative magnitude ablation effect\" is a token-classifying latent.\n\nLines 148-157: The metric described can fail at its stated goal to \"[measure] the propensity of the model to choose the correct starting letter as opposed to other letters\" because averaging logits can interact poorly with the softmax activation function. For example, if \"A\" is the correct letter, but the model logits are g[A]=10, g[B]=25, g[C]=g[D]=...=g[Z]=0, then the model would produce B with near-certainty. But the provided metric yields m=10-(25+0+0+...+0)/25=10-1=9, the same value as if g[A]=9, g[B]=...=g[Z]=0. This metric would be more convincing if the average were replaced by a max, or if it were a softmax over letter tokens. \n\nThe authors only study a single family of SAEs on a single language model. Alternative SAE architectures may not demonstrate feature absorption.", "questions": "**Questions** \n\n1. The concept of \"approximately token-aligned\" latents is central to claim 3 (lines 69-71), but is not defined anywhere. How is token-alignment defined, and is it quantified or measured in these experiments?\n\n2. When the F1 score of an SAE is computed, as in Figure 2 (lines 169-172), how are SAE activations converted into a binary classification? Is the classification \"does/doesn't activate\". Is a non-zero threshold used? If so, how is that threshold chosen?\n\n3. In Figure 7b, when the mean absorption rate is \\~35%, is that indicating that for roughly 35% of tokens, there is a latent absorbing that token? The vocabulary size of Gemma-2B is \\~256k (https://arxiv.org/html/2408.00118v1#:~:text=Vocab%20size-,256128,-256128) and 35% of that is \\~90k. If that is the case, how are 90k tokens absorbed into 65k latents?\n\n4. In figure 13, why is the high cosine similarity in the .4-.6 range, and not closer to 1? This range of cosine similarities is suggestive of learning the \"true\" feature mixed with a second, orthogonal feature. \n\n5. It seems that the paper usually uses cosine similarity of the LR direction with the encoder direction, but Figures 4 and 13 use decoders instead of encoders. Why is that?\n\n6. What does Algorithm 1 (Lines 116-127) return? One ablation effect per token per latent? Or are these averaged in some way?\n\n**Other Comments**\n\nIf it is permitted for authors to make revisions before the final submission, there are several small changes that could improve the quality of the paper:\n\n- Lines 94-107: the paper should reference (Lieberum et al., 2024) since that is the exact architecture of SAE being studied. \n\n- Line 100: Although ReLU is often used in SAE architectures, it would be better to write a generic activation function in this equation, especially considering that the SAEs studied in the paper use JumpReLU, not ReLU.\n\n- Line 118 and Line 147: The description \"include the SAE error term\" is ambiguous in the text. It is clarified in the glossary, and by the glossary's citation to (Marks et al., 2024), but the main paper neither cites (Marks et al., 2024) here, nor links to the glossary. Is there a way to include hyperlinks to the glossary in the text itself? \n\n- Line 205: There is a citation to (Gao et al., 2024) for sparse probing. It is likely supposed to cite (Gurnee et al., 2023) as in the background section. (Gao et al., 2024) use k-sparse SAEs, which are distinct from k-sparse probes and the SAE architecture studied in the paper.\n\n- Lines 291-292: \"The token where the SAE feature activates is highlighted in green\" should be \"yellow\" since the highlight is yellow.\n\n- Line 404: The paper says \"it is difficult to apply [(Karvonen et al, 2024)] to existing SAEs trained on real LLM activations.\" One of the main contributions of (Karvonen et al, 2024) is very applicable to this paper, namely coverage (Section 3.2 in https://arxiv.org/pdf/2408.00113). Coverage for a given set of properties (in this case, \"first letter identification\") is defined by, for each property, finding the SAE latent and threshold which results in the best F1 score for classifying that property, then averaging those F1 scores across the properties. This is very similar to what is shown in Figure 2, though with a different selection rule for SAE latents. The paper could make reference to the similarity of this method to coverage.\n\n- Line 429, possible typo: \"..find that by training an SAE on the decoder *or* another SEE, a technique...\", the word \"or\" should be \"of\".\n\n- (Uncertain) The citation of \"(Huben et al., 2024)\" should instead refer to the paper as \"(Cunningham et al., 2024)\". The order of the authors is different between the OpenReview page (https://openreview.net/forum?id=F76bwRSLeK) and the paper itself (https://openreview.net/pdf?id=F76bwRSLeK). In such a case it is most likely proper to follow the order of authors in the paper.\n\n**References**\n\nLeo Gao, Tom Dupre ́ la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093, 2024.\n\nAdam Karvonen, Benjamin Wright, Can Rager, Rico Angell, Jannik Brinkmann, Logan Riggs Smith, Claudio Mayrink Verdun, David Bau, and Samuel Marks. Measuring progress in dic- tionary learning for language model interpretability with board game models. In ICML 2024 Workshop on Mechanistic Interpretability, 2024.\n\nWes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsi- mas. Finding neurons in a haystack: Case studies with sparse probing. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum? id=JYs1R9IMJr.\n\nRobert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=F76bwRSLeK.\n\nTom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, Ja ́nos Krama ́r, Anca Dragan, Rohin Shah, and Neel Nanda. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2, August 2024.\n\nSamuel Marks, Can Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language mod- els. Computing Research Repository, arXiv:2403.19647, 2024. URL https://arxiv.org/ abs/2403.19647.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper documents and measures a particular form of feature splitting in the Gemma Scope family of sparse autoencoders (SAEs), which the authors call feature absorption. The prototypical case of feature absorption studied is when a latent is a high-precision classifier for a property like \"starts with the letter L\", but has poor recall due to other latents \"absorbing\" individual tokens, so that the main latent's activations are given by the rule \"starts with the letter L, except for the tokens laser, lions, [several other exceptions]\".", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The paper explores a natural task, first letter classification, and does so well. It establishes that this information is linearly available in the model (by training an LR probe with high F1-score), and that SAEs find latents that align with the probe direction (Figure 4b and especially Figure 13), and that these latents are often good classifiers, though not as good as probes (Figure 2a).\n\nThe paper provides a novel perspective to a documented issue of feature splitting in SAEs. It provides good theoretical justification of why feature absorption is likely to happen, and why it provides challenges to interpretability.", "weaknesses": "The paper could do more to demonstrate that feature absorption happens as described, namely that: \n1) There is a latent which is a high-precision classifier for \"starts with letter [x]\".\n2) There is a particular subset of tokens which start with this letter but which the latent fails to activate on.\n3) There are separate token-aligned latents that individually classify those tokens.\n4) If that subset of tokens is removed, the latent is a high-precision, high-recall classifier for \"starts with the letter [x], except for that subset of tokens\".\n\nThe authors demonstrate 1) and 2) well in sections 4.1 and 4.3 respectively, but never show 3) or 4) besides a single example in section 4.2. The authors could address 3) by adding an additional step to the experiment in section 4.3, in which they confirm that \"the SAE latent receiving the largest negative magnitude ablation effect\" is a token-classifying latent.\n\nLines 148-157: The metric described can fail at its stated goal to \"[measure] the propensity of the model to choose the correct starting letter as opposed to other letters\" because averaging logits can interact poorly with the softmax activation function. For example, if \"A\" is the correct letter, but the model logits are g[A]=10, g[B]=25, g[C]=g[D]=...=g[Z]=0, then the model would produce B with near-certainty. But the provided metric yields m=10-(25+0+0+...+0)/25=10-1=9, the same value as if g[A]=9, g[B]=...=g[Z]=0. This metric would be more convincing if the average were replaced by a max, or if it were a softmax over letter tokens. \n\nThe authors only study a single family of SAEs on a single language model. Alternative SAE architectures may not demonstrate feature absorption.", "questions": "**Questions** \n\n1. The concept of \"approximately token-aligned\" latents is central to claim 3 (lines 69-71), but is not defined anywhere. How is token-alignment defined, and is it quantified or measured in these experiments?\n\n2. When the F1 score of an SAE is computed, as in Figure 2 (lines 169-172), how are SAE activations converted into a binary classification? Is the classification \"does/doesn't activate\". Is a non-zero threshold used? If so, how is that threshold chosen?\n\n3. In Figure 7b, when the mean absorption rate is \\~35%, is that indicating that for roughly 35% of tokens, there is a latent absorbing that token? The vocabulary size of Gemma-2B is \\~256k (https://arxiv.org/html/2408.00118v1#:~:text=Vocab%20size-,256128,-256128) and 35% of that is \\~90k. If that is the case, how are 90k tokens absorbed into 65k latents?\n\n4. In figure 13, why is the high cosine similarity in the .4-.6 range, and not closer to 1? This range of cosine similarities is suggestive of learning the \"true\" feature mixed with a second, orthogonal feature. \n\n5. It seems that the paper usually uses cosine similarity of the LR direction with the encoder direction, but Figures 4 and 13 use decoders instead of encoders. Why is that?\n\n6. What does Algorithm 1 (Lines 116-127) return? One ablation effect per token per latent? Or are these averaged in some way?\n\n**Other Comments**\n\nIf it is permitted for authors to make revisions before the final submission, there are several small changes that could improve the quality of the paper:\n\n- Lines 94-107: the paper should reference (Lieberum et al., 2024) since that is the exact architecture of SAE being studied. \n\n- Line 100: Although ReLU is often used in SAE architectures, it would be better to write a generic activation function in this equation, especially considering that the SAEs studied in the paper use JumpReLU, not ReLU.\n\n- Line 118 and Line 147: The description \"include the SAE error term\" is ambiguous in the text. It is clarified in the glossary, and by the glossary's citation to (Marks et al., 2024), but the main paper neither cites (Marks et al., 2024) here, nor links to the glossary. Is there a way to include hyperlinks to the glossary in the text itself? \n\n- Line 205: There is a citation to (Gao et al., 2024) for sparse probing. It is likely supposed to cite (Gurnee et al., 2023) as in the background section. (Gao et al., 2024) use k-sparse SAEs, which are distinct from k-sparse probes and the SAE architecture studied in the paper.\n\n- Lines 291-292: \"The token where the SAE feature activates is highlighted in green\" should be \"yellow\" since the highlight is yellow.\n\n- Line 404: The paper says \"it is difficult to apply [(Karvonen et al, 2024)] to existing SAEs trained on real LLM activations.\" One of the main contributions of (Karvonen et al, 2024) is very applicable to this paper, namely coverage (Section 3.2 in https://arxiv.org/pdf/2408.00113). Coverage for a given set of properties (in this case, \"first letter identification\") is defined by, for each property, finding the SAE latent and threshold which results in the best F1 score for classifying that property, then averaging those F1 scores across the properties. This is very similar to what is shown in Figure 2, though with a different selection rule for SAE latents. The paper could make reference to the similarity of this method to coverage.\n\n- Line 429, possible typo: \"..find that by training an SAE on the decoder *or* another SEE, a technique...\", the word \"or\" should be \"of\".\n\n- (Uncertain) The citation of \"(Huben et al., 2024)\" should instead refer to the paper as \"(Cunningham et al., 2024)\". The order of the authors is different between the OpenReview page (https://openreview.net/forum?id=F76bwRSLeK) and the paper itself (https://openreview.net/pdf?id=F76bwRSLeK). In such a case it is most likely proper to follow the order of authors in the paper.\n\n**References**\n\nLeo Gao, Tom Dupre ́ la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093, 2024.\n\nAdam Karvonen, Benjamin Wright, Can Rager, Rico Angell, Jannik Brinkmann, Logan Riggs Smith, Claudio Mayrink Verdun, David Bau, and Samuel Marks. Measuring progress in dic- tionary learning for language model interpretability with board game models. In ICML 2024 Workshop on Mechanistic Interpretability, 2024.\n\nWes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsi- mas. Finding neurons in a haystack: Case studies with sparse probing. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum? id=JYs1R9IMJr.\n\nRobert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=F76bwRSLeK.\n\nTom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, Ja ́nos Krama ́r, Anca Dragan, Rohin Shah, and Neel Nanda. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2, August 2024.\n\nSamuel Marks, Can Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language mod- els. Computing Research Repository, arXiv:2403.19647, 2024. URL https://arxiv.org/ abs/2403.19647.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730316247716}], "openreview_url": "https://openreview.net/forum?id=LC2KxRwC3n", "arxiv_id": "2409.14507", "paper_pdf": "papers/LC2KxRwC3n.pdf", "paper_pdf_sha256": "97ffad0178188e8cc48a0bb2614ca50956afcfd9623bb44026f37ff13f16a238", "paper_pdf_bytes": 3838697, "paper_pdf_source": "openreview", "code_url": "https://github.com/lasr-spelling/sae-spelling", "code_repository": "lasr-spelling/sae-spelling", "code_commit": "515648a5b96dc389ba317c4053fd8aa6eaca9929", "code_archive": "repos/LC2KxRwC3n.zip", "code_archive_sha256": "85ed4b95247f9f8ef69fee70664f70d3b44ef86c802c4d3f4565bac05a7adb2c", "code_archive_bytes": 234672, "code_file_count": 32, "code_extensions": {".py": 32}, "github_disk_usage_kb": 400, "github_languages": {"Python": 252157, "Makefile": 223}, "github_archived": false, "github_pushed_at": "2025-12-28T16:51:55Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-is-for-absorption-studying-feature"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "I5lcjmFmlc", "year": 2024, "status": "rejected", "title": "Robust Classification via a Single Diffusion Model", "authors": ["Huanran Chen", "Yinpeng Dong", "Zhengyi Wang", "Xiao Yang", "Chengqi Duan", "Hang Su", "Jun Zhu"], "authorids": ["~Huanran_Chen1", "~Yinpeng_Dong2", "~Zhengyi_Wang1", "~Xiao_Yang4", "~Chengqi_Duan1", "~Hang_Su3", "~Jun_Zhu2"], "authors_source": "OpenReview API", "abstract": "Recently, diffusion models have been successfully applied to improving adversarial robustness of image classifiers by purifying the adversarial noises or generating realistic data for adversarial training. However, the diffusion-based purification can be evaded by stronger adaptive attacks while adversarial training does not perform well under unseen threats, exhibiting inevitable limitations of these methods. To better harness the expressive power of diffusion models, in this paper we propose Robust Diffusion Classifier (RDC), a generative classifier that is constructed from a pre-trained diffusion model to be adversarially robust. Our method first maximizes the data likelihood of a given input and then predicts the class probabilities of the optimized input using the conditional likelihood estimated by the diffusion model through Bayes' theorem. To further reduce the computational complexity, we propose a new diffusion backbone called multi-head diffusion and develop efficient sampling strategies. As our method does not require training on particular adversarial attacks, we demonstrate that it is more generalizable to defend against multiple unseen threats. In particular, RDC achieves 75.67% robust accuracy against $\\ell_\\infty$ norm-bounded perturbations with $\\epsilon_\\infty=8/255$ on CIFAR-10, surpassing the previous state-of-the-art adversarial training models by +4.77%. The findings highlight the potential of generative classifiers by employing diffusion models for adversarial robustness compared with the commonly studied discriminative classifiers.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "APg5VqtViB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission42/Reviewer_Xkzq"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "2 fair", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes to build a robust classifier using a single diffusion model by calculating $p (y|x)$ via $p (x|y)$. The authors identify likelihood maximization as a key ingredient for ensuring the adversarial robustness of such models. To address the high computational complexity associated with this type of classifiers, the authors further introduce a multi-head U-Net and ablate on efficient sampling methods. Experiment results on a subset of CIFAR-10 using BPDA-AutoAttack show that the proposed method achieves SOTA clean and adversarial accuracy.", "review_text": "This paper proposes to build a robust classifier using a single diffusion model by calculating $p (y|x)$ via $p (x|y)$. The authors identify likelihood maximization as a key ingredient for ensuring the adversarial robustness of such models. To address the high computational complexity associated with this type of classifiers, the authors further introduce a multi-head U-Net and ablate on efficient sampling methods. Experiment results on a subset of CIFAR-10 using BPDA-AutoAttack show that the proposed method achieves SOTA clean and adversarial accuracy.", "strengths": "This paper proposes an interesting and relevant framework for robust classification. Given how fast diffusion models are improving in comparison to traditional discriminative robust classifiers, this work opens a new method of building robust models. It thus has a lot of potential for inspiring future research that builds even better robust models. Furthermore, the authors are careful with evaluating the proposed method with strong adaptive attacks, providing justifications for the proposed robustness estimation methods.", "weaknesses": "The main weaknesses are twofold: computational complexity and paper presentation.\n\n### Computational Complexity\n\nEven with the proposed multi-head U-Net and other complexity reduction measures, the computational complexity still seems to be high. While the likelihood maximization step only requires $N=5$ forward and backward passes, approximating $p(x|y)$ requires $T$ U-Net queries. That being said, I agree that this drawback can be left for future work.\n\n### Paper Presentation\n\nMany important details are missing from the discussion. Some of them can be found in the appendix, but they really should be in the main text. This is especially the case for Section 3.3.\n- Theorem 3.2 discusses \"optimal diffusion models\". In what sense is the diffusion model optimal? The proof to this theorem clarifies that such a model minimizes the noise estimation error, but this should be in the main text.\n- There is a softmax operation in Theorem 3.2, but the quantity on which it operates is a scalar (a norm square divided by some variance). What does softmax exactly mean here? Same for Corollary 3.3.\n- \"We find that the optimal diffusion classifier achieves 100% robust accuracy in both cases, validating our hypothesize that accurate density estimation of diffusion models facilitates robust classification.\" How was this found? Also, it should be \"hypothesis\", not \"hypothesize\". What is the main gap between an optimal diffusion classifier and an empirical diffusion classifier? Is it the limited amount of training data? Or is it how well the U-Net is optimized?\n- Section 3.5 says, \"instead of calculating the diffusion loss using all timesteps like Eq. (9), we only sample a single timestep\". How is this time step sampled? Uniformly randomly?\n- Section 3.5 also says, \"(BPDA) approximates the gradient with an identity mapping\". How exactly is the identity mapping applied? A pseudo-code or Python code explanation would be appreciated.\n- It would be nice to have the experiment results from the CIFAR-100 dataset to have some diversity in the evaluation.", "questions": "- Since AutoAttack with BPDA is used for evaluation, is the Square Attack component of AutoAttack also included in the evaluation? Does Square find additional examples on top of the BPDA gradient-based attacks?\n- In Figure 2a, why is the robust accuracy barely over 30%? What is the value of $T$ and $T'$ for the main results (Table 1)? Does the result get even better if $T'$ is larger than 1000? How large is $T$ during the training of the diffusion model? If we train a diffusion model with fewer time steps (i.e., discretize the trajectory into less than 1000 steps during training), should we expect the resulting classifier obtained via the proposed method to work well with a smaller $T'$?\n- With multi-head diffusion, is it true that only the last layer receives the class condition signal? How does this affect the performance compared with injecting class conditioning into various locations in the U-Net? It would also be nice to see some generations from this diffusion model.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to build a robust classifier using a single diffusion model by calculating $p (y|x)$ via $p (x|y)$. The authors identify likelihood maximization as a key ingredient for ensuring the adversarial robustness of such models. To address the high computational complexity associated with this type of classifiers, the authors further introduce a multi-head U-Net and ablate on efficient sampling methods. Experiment results on a subset of CIFAR-10 using BPDA-AutoAttack show that the proposed method achieves SOTA clean and adversarial accuracy.", "soundness": "3 good", "presentation": "2 fair", "contribution": "4 excellent", "strengths": "This paper proposes an interesting and relevant framework for robust classification. Given how fast diffusion models are improving in comparison to traditional discriminative robust classifiers, this work opens a new method of building robust models. It thus has a lot of potential for inspiring future research that builds even better robust models. Furthermore, the authors are careful with evaluating the proposed method with strong adaptive attacks, providing justifications for the proposed robustness estimation methods.", "weaknesses": "The main weaknesses are twofold: computational complexity and paper presentation.\n\n### Computational Complexity\n\nEven with the proposed multi-head U-Net and other complexity reduction measures, the computational complexity still seems to be high. While the likelihood maximization step only requires $N=5$ forward and backward passes, approximating $p(x|y)$ requires $T$ U-Net queries. That being said, I agree that this drawback can be left for future work.\n\n### Paper Presentation\n\nMany important details are missing from the discussion. Some of them can be found in the appendix, but they really should be in the main text. This is especially the case for Section 3.3.\n- Theorem 3.2 discusses \"optimal diffusion models\". In what sense is the diffusion model optimal? The proof to this theorem clarifies that such a model minimizes the noise estimation error, but this should be in the main text.\n- There is a softmax operation in Theorem 3.2, but the quantity on which it operates is a scalar (a norm square divided by some variance). What does softmax exactly mean here? Same for Corollary 3.3.\n- \"We find that the optimal diffusion classifier achieves 100% robust accuracy in both cases, validating our hypothesize that accurate density estimation of diffusion models facilitates robust classification.\" How was this found? Also, it should be \"hypothesis\", not \"hypothesize\". What is the main gap between an optimal diffusion classifier and an empirical diffusion classifier? Is it the limited amount of training data? Or is it how well the U-Net is optimized?\n- Section 3.5 says, \"instead of calculating the diffusion loss using all timesteps like Eq. (9), we only sample a single timestep\". How is this time step sampled? Uniformly randomly?\n- Section 3.5 also says, \"(BPDA) approximates the gradient with an identity mapping\". How exactly is the identity mapping applied? A pseudo-code or Python code explanation would be appreciated.\n- It would be nice to have the experiment results from the CIFAR-100 dataset to have some diversity in the evaluation.", "questions": "- Since AutoAttack with BPDA is used for evaluation, is the Square Attack component of AutoAttack also included in the evaluation? Does Square find additional examples on top of the BPDA gradient-based attacks?\n- In Figure 2a, why is the robust accuracy barely over 30%? What is the value of $T$ and $T'$ for the main results (Table 1)? Does the result get even better if $T'$ is larger than 1000? How large is $T$ during the training of the diffusion model? If we train a diffusion model with fewer time steps (i.e., discretize the trajectory into less than 1000 steps during training), should we expect the resulting classifier obtained via the proposed method to work well with a smaller $T'$?\n- With multi-head diffusion, is it true that only the last layer receives the class condition signal? How does this affect the performance compared with injecting class conditioning into various locations in the U-Net? It would also be nice to see some generations from this diffusion model.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698728120270}, {"id": "mijN7uvi18", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission42/Reviewer_ScPt"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper presents the Robust Diffusion Classifier (RDC), a generative classifier designed for robust classification tasks. RDC operates by first optimizing the data likelihood of an input and then estimating class probabilities for this optimized input. This is achieved using the diffusion model with transforms using the Bayes' rule. Recognizing the need for classification efficiency, the authors introduce a novel diffusion backbone termed \"multi-head diffusion\" and for a sampling strategy with fewer NFE. Notably, the method requires no specific training against particular adversarial attacks, showcasing its adaptability in defending against a spectrum of previously unseen threats.", "review_text": "This paper presents the Robust Diffusion Classifier (RDC), a generative classifier designed for robust classification tasks. RDC operates by first optimizing the data likelihood of an input and then estimating class probabilities for this optimized input. This is achieved using the diffusion model with transforms using the Bayes' rule. Recognizing the need for classification efficiency, the authors introduce a novel diffusion backbone termed \"multi-head diffusion\" and for a sampling strategy with fewer NFE. Notably, the method requires no specific training against particular adversarial attacks, showcasing its adaptability in defending against a spectrum of previously unseen threats.", "strengths": "- The proposed methodology stands out as both technically sound and effective. Its capability to achieve robust classification without specific knowledge of adversarial attacks is admirable. The paper has both theoretical and empirical contributions, which are beneficial to both the theoretical-favored researchers and practitioners.\n\n- The paper is overall well-written and provides a smooth reading experience. The incorporation of model overviews and illustrative diagrams further clarifies the proposed methodology, enhancing comprehension.\n\n- The experimental results well validate the approach. Notably, the paper also provides an ablation study to interpret the effects of the hyperparameters. \n\n - The authors have provided a thorough study that motivates the design of the multi-head classifier in the appendix. This study provides additional insight for researchers to figure out problems in related domains.\n\n- I appreciate the authors' transparency in addressing potential limitations. The practical side of the research is solid. The authors have been very detailed in their implementation and provided their experiment code, which ensures reproducibility.", "weaknesses": "Although the content of the current version is satisfactory, some points listed below can further enhance the depth and completeness of this work:\n\n- Some details regarding the diffusion model need clarification. For example, the sampling strategy is not very clear to me. It seems the authors deploy a VP sampling with uniform sliding timesteps similar to the one used in Nichol and Dhariwal, (2021). In the appendix and the code the authors also seem to leverage some implementation from Karras et al. (2022). As we note the sampler in Karras et al. (2022) involves additional correction steps, the NFE of the diffusion backbone would be more than T. \n\n- The theoretical results are based on the hypothesis that the evidence lower bound is tight. It would be intriguing to explore the implications of the gap between the likelihood and this lower bound, especially when viewed through the lens of the Bayesian framework for uncertainty quantification. Exploring how this gap influences robust classification performance could enrich the paper's depth and utility.\n\n\n- While the current experiments are limited to relatively small-scale datasets, there is inherent value in examining the method's scalability. It would be beneficial if the authors could present results or potential methodologies to apply their approach to larger datasets, drawing inspiration from other generative classifiers in the Bayesian paradigm, such as those highlighted by Heek and Nal (2019) and Han et al. (2022).\n\n\n- This paper also has close relation to other generative classifiers not specified for robust classification. Although the authors have discussed some concurrent works, some prior works may need to include and discuss potential connections. \n\n- The current form of the paper draws parallels to several generative classifiers, though not specifically designed for robust classification. While some works have been discussed as concurrent works, it might be helpful in enhancing the completeness to integrate and discuss other prior works, emphasizing their relevance and potential connections to the proposed methodology.\n\n----------\nReference:\n\nHeek, Jonathan, and Nal Kalchbrenner. \"Bayesian inference for large scale image classification.\" arXiv preprint arXiv:1908.03491 (2019).\n\nMackowiak, R., Ardizzone, L., Kothe, U., & Rother, C. (2021). Generative classifiers as a basis for trustworthy image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2971-2981).\n\nHoogeboom, E., Nielsen, D., Jaini, P., Forré, P., & Welling, M. (2021). Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34, 12454-12465.\n\nHan, X., Zheng, H., & Zhou, M. (2022). Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35, 18100-18115.", "questions": "Please see the first point of the weakness regarding the details in diffusion settings.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents the Robust Diffusion Classifier (RDC), a generative classifier designed for robust classification tasks. RDC operates by first optimizing the data likelihood of an input and then estimating class probabilities for this optimized input. This is achieved using the diffusion model with transforms using the Bayes' rule. Recognizing the need for classification efficiency, the authors introduce a novel diffusion backbone termed \"multi-head diffusion\" and for a sampling strategy with fewer NFE. Notably, the method requires no specific training against particular adversarial attacks, showcasing its adaptability in defending against a spectrum of previously unseen threats.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "4 excellent", "strengths": "- The proposed methodology stands out as both technically sound and effective. Its capability to achieve robust classification without specific knowledge of adversarial attacks is admirable. The paper has both theoretical and empirical contributions, which are beneficial to both the theoretical-favored researchers and practitioners.\n\n- The paper is overall well-written and provides a smooth reading experience. The incorporation of model overviews and illustrative diagrams further clarifies the proposed methodology, enhancing comprehension.\n\n- The experimental results well validate the approach. Notably, the paper also provides an ablation study to interpret the effects of the hyperparameters. \n\n - The authors have provided a thorough study that motivates the design of the multi-head classifier in the appendix. This study provides additional insight for researchers to figure out problems in related domains.\n\n- I appreciate the authors' transparency in addressing potential limitations. The practical side of the research is solid. The authors have been very detailed in their implementation and provided their experiment code, which ensures reproducibility.", "weaknesses": "Although the content of the current version is satisfactory, some points listed below can further enhance the depth and completeness of this work:\n\n- Some details regarding the diffusion model need clarification. For example, the sampling strategy is not very clear to me. It seems the authors deploy a VP sampling with uniform sliding timesteps similar to the one used in Nichol and Dhariwal, (2021). In the appendix and the code the authors also seem to leverage some implementation from Karras et al. (2022). As we note the sampler in Karras et al. (2022) involves additional correction steps, the NFE of the diffusion backbone would be more than T. \n\n- The theoretical results are based on the hypothesis that the evidence lower bound is tight. It would be intriguing to explore the implications of the gap between the likelihood and this lower bound, especially when viewed through the lens of the Bayesian framework for uncertainty quantification. Exploring how this gap influences robust classification performance could enrich the paper's depth and utility.\n\n\n- While the current experiments are limited to relatively small-scale datasets, there is inherent value in examining the method's scalability. It would be beneficial if the authors could present results or potential methodologies to apply their approach to larger datasets, drawing inspiration from other generative classifiers in the Bayesian paradigm, such as those highlighted by Heek and Nal (2019) and Han et al. (2022).\n\n\n- This paper also has close relation to other generative classifiers not specified for robust classification. Although the authors have discussed some concurrent works, some prior works may need to include and discuss potential connections. \n\n- The current form of the paper draws parallels to several generative classifiers, though not specifically designed for robust classification. While some works have been discussed as concurrent works, it might be helpful in enhancing the completeness to integrate and discuss other prior works, emphasizing their relevance and potential connections to the proposed methodology.\n\n----------\nReference:\n\nHeek, Jonathan, and Nal Kalchbrenner. \"Bayesian inference for large scale image classification.\" arXiv preprint arXiv:1908.03491 (2019).\n\nMackowiak, R., Ardizzone, L., Kothe, U., & Rother, C. (2021). Generative classifiers as a basis for trustworthy image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2971-2981).\n\nHoogeboom, E., Nielsen, D., Jaini, P., Forré, P., & Welling, M. (2021). Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34, 12454-12465.\n\nHan, X., Zheng, H., & Zhou, M. (2022). Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35, 18100-18115.", "questions": "Please see the first point of the weakness regarding the details in diffusion settings.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698707241392}, {"id": "H2AuadFKIh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission42/Reviewer_noA3"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work proposes a diffusion-based classifier that surpasses the SOTA level of robust accuracy against lp-bound adversarial attacks without relying on adversarial training. The method is advantageous to previous methods because it does not require inference for every class in the dataset by leveraging a \"multi-head diffusion\" block. Further, to improve the density estimation of real-world models the authors propose a \"Likelihood Maximization\" technique.", "review_text": "This work proposes a diffusion-based classifier that surpasses the SOTA level of robust accuracy against lp-bound adversarial attacks without relying on adversarial training. The method is advantageous to previous methods because it does not require inference for every class in the dataset by leveraging a \"multi-head diffusion\" block. Further, to improve the density estimation of real-world models the authors propose a \"Likelihood Maximization\" technique.", "strengths": "1. Clear and quality writing.\n2. The method is carefully designed with theoretical justifications. In particular, the authors go to great lengths to address gradient obfuscation.\n3. The method is benchmarked against modern adversarial attacks (AA, StAdv, BPDA) and beats SOTA by a large margin while showing generalization to more threats than previous methods.\n4. Thorough review of related work and comparison to previous methods.", "weaknesses": "1. The manuscript mentions multiple times the expensive inference in both time and memory - however, a quantitative analysis is missing. I would like to see a table with inference time and memory for this method in comparison to other generative classifiers and regular ones.\n2. The method appears to work well on \"unseen\" threats. However, all threats are limited to adversarial attacks. Is there any evidence for an increased robustness to other robustness aspects such as common corruptions (e.g., CIFAR10-C)?\n3. The experimental evaluation is performed on a clearly separated and limited number of classes. Are there any results or theoretical insights on how this method would scale to more and potentially more fine-grained classes?", "questions": "1. In Appendix B the authors state \"However, this architecture only achieves 60% accuracy on the CIFAR10 dataset\". How does that relate to the 90-ish % in Tab. 1? I.e. what is different in this section?\n2. “... [the mthod] ... leverages an off-the-shelf diffusion model” - how is this possible when the last layer is modified? Which diffusion model is this exactly? How does it compare to the diffusion models previous works have used?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes a diffusion-based classifier that surpasses the SOTA level of robust accuracy against lp-bound adversarial attacks without relying on adversarial training. The method is advantageous to previous methods because it does not require inference for every class in the dataset by leveraging a \"multi-head diffusion\" block. Further, to improve the density estimation of real-world models the authors propose a \"Likelihood Maximization\" technique.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. Clear and quality writing.\n2. The method is carefully designed with theoretical justifications. In particular, the authors go to great lengths to address gradient obfuscation.\n3. The method is benchmarked against modern adversarial attacks (AA, StAdv, BPDA) and beats SOTA by a large margin while showing generalization to more threats than previous methods.\n4. Thorough review of related work and comparison to previous methods.", "weaknesses": "1. The manuscript mentions multiple times the expensive inference in both time and memory - however, a quantitative analysis is missing. I would like to see a table with inference time and memory for this method in comparison to other generative classifiers and regular ones.\n2. The method appears to work well on \"unseen\" threats. However, all threats are limited to adversarial attacks. Is there any evidence for an increased robustness to other robustness aspects such as common corruptions (e.g., CIFAR10-C)?\n3. The experimental evaluation is performed on a clearly separated and limited number of classes. Are there any results or theoretical insights on how this method would scale to more and potentially more fine-grained classes?", "questions": "1. In Appendix B the authors state \"However, this architecture only achieves 60% accuracy on the CIFAR10 dataset\". How does that relate to the 90-ish % in Tab. 1? I.e. what is different in this section?\n2. “... [the mthod] ... leverages an off-the-shelf diffusion model” - how is this possible when the last layer is modified? Which diffusion model is this exactly? How does it compare to the diffusion models previous works have used?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698701082923}], "openreview_url": "https://openreview.net/forum?id=I5lcjmFmlc", "arxiv_id": "2305.15241", "paper_pdf": "papers/I5lcjmFmlc.pdf", "paper_pdf_sha256": "8609580e25373aaec4239effe0d3e7ca16318c1f5f77c311aa9ba6d21bcdab2c", "paper_pdf_bytes": 2716088, "paper_pdf_source": "openreview", "code_url": "https://github.com/huanranchen/DiffusionClassifier", "code_repository": "huanranchen/DiffusionClassifier", "code_commit": "8a825045d91ab50c98503933c874a4296c5b2bec", "code_archive": "repos/I5lcjmFmlc.zip", "code_archive_sha256": "ff237233e91228f5d66396b1d92ff3fece9b3b6a963447c16057a6b505a29054", "code_archive_bytes": 549418, "code_file_count": 282, "code_extensions": {".py": 278, ".cpp": 2, ".cu": 2}, "github_disk_usage_kb": 437, "github_languages": {"Python": 4123690, "Cuda": 14488, "C++": 2555}, "github_archived": false, "github_pushed_at": "2025-03-07T09:08:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/robust-classification-via-a-single-diffusion"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "C9sU3Tnnki8", "year": 2023, "status": "rejected", "title": "Exploring Transformer Backbones for Heterogeneous Treatment Effect Estimation", "authors": ["YiFan Zhang", "Hanlin Zhang", "Zachary Chase Lipton", "Li Erran Li", "Eric Xing"], "authorids": ["~YiFan_Zhang8", "~Hanlin_Zhang1", "~Zachary_Chase_Lipton1", "~Li_Erran_Li1", "~Eric_Xing1"], "authors_source": "OpenReview API", "abstract": "Previous works on Treatment Effect Estimation (TEE) are not in widespread use because they are predominantly theoretical, where strong parametric assumptions are made but untractable for practical application. Recent works use Multilayer Perceptron (MLP) for modeling casual relationships, however, MLPs lag far behind recent advances in ML methodology, which limits their applicability and generalizability. To extend beyond the single domain formulation and towards more realistic learning scenarios, we explore model design spaces beyond MLPs, i.e., transformer backbones, which provide flexibility where attention layers govern interactions among treatments and covariates to exploit structural similarities of potential outcomes for confounding control. Through careful model design, Transformers as Treatment Effect Estimators (TransTEE) is proposed. We show empirically that TransTEE can: (1) serve as a general-purpose treatment effect estimator which significantly outperforms competitive baselines on a variety of challenging TEE problems (e.g., discrete, continuous, structured, or dosage-associated treatments.) and is applicable to both when covariates are tabular and when they consist of structural data (e.g., texts, graphs); (2) yield multiple advantages: compatibility with propensity score modeling, parameter efficiency, robustness to continuous treatment value distribution shifts, explainable in covariate adjustment, and real-world utility in auditing pre-trained language models. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "gyfEwA9Fjw", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2191/Reviewer_mZnE"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper is interested in estimating the heterogeneous treatment effect (TEE). The authors argued that TEE is not used in practice because current approaches to tackling it make solid parametric assumptions.\n\nAccording to the authors, the limitations of previously proposed approaches mainly relate to their poor generalizability. Hence, they tend to focus on a narrower context and ignore other scenarios. \n\nThe method present in this paper is based on transformers and their backbone modules. Specifically, its main module is the attention layer. \n\nMy main concern in this paper is the assumption about unconfoundeness, which states that there are no hidden confounders. Since the authors are arguing against the applicability of previous works, this assumption impedes using their approach in real applications as well.", "review_text": "This paper proposes to advance the literature by leveraging recent developments on NNs to tackle the Treatment Effect Estimation problem. They provide extensive experiments and detailed information about their proposed method. However, they argued that previous approaches have issues with real-world applicability. When they ignore the presence of confounders, this paper seems to suffer from the same applicability problem.", "strengths": "Strengths:\n\n- The authors used recent advances in NN architectures (e.g., transformers) to design a new method to tackle the Treatment Effect Estimation problem. For instance, they propose different mappings for covariates and treatments, which avoid missing information about treatments given their lower dimensional representation.\n- The authors considered different modalities of treatments. Namely, treatment with binary/continuous dosages, structured treatments, and language data. The broad of distinct treatment types helps assess their method's applicability.\n\nWeaknesses:\n\n- The assumption related to unconfoundeness is a very limiting factor in the usage of the proposed method in real-world applications.\n- The authors provide extensive analysis using different datasets. However, the reproducibility of their results is difficult, given that they did not make the source code available.\n- The authors should clarify the T- and S-learner approaches to causal inference. Most readers are not familiar with these terms.\n\nMinor suggestions:\n\n- It seems to have a typo in Assumption 2: \" such that, i.e.\"  -> \"i.e.,\"", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper is interested in estimating the heterogeneous treatment effect (TEE). The authors argued that TEE is not used in practice because current approaches to tackling it make solid parametric assumptions.\n\nAccording to the authors, the limitations of previously proposed approaches mainly relate to their poor generalizability. Hence, they tend to focus on a narrower context and ignore other scenarios. \n\nThe method present in this paper is based on transformers and their backbone modules. Specifically, its main module is the attention layer. \n\nMy main concern in this paper is the assumption about unconfoundeness, which states that there are no hidden confounders. Since the authors are arguing against the applicability of previous works, this assumption impedes using their approach in real applications as well.", "strength_and_weaknesses": "Strengths:\n\n- The authors used recent advances in NN architectures (e.g., transformers) to design a new method to tackle the Treatment Effect Estimation problem. For instance, they propose different mappings for covariates and treatments, which avoid missing information about treatments given their lower dimensional representation.\n- The authors considered different modalities of treatments. Namely, treatment with binary/continuous dosages, structured treatments, and language data. The broad of distinct treatment types helps assess their method's applicability.\n\nWeaknesses:\n\n- The assumption related to unconfoundeness is a very limiting factor in the usage of the proposed method in real-world applications.\n- The authors provide extensive analysis using different datasets. However, the reproducibility of their results is difficult, given that they did not make the source code available.\n- The authors should clarify the T- and S-learner approaches to causal inference. Most readers are not familiar with these terms.\n\nMinor suggestions:\n\n- It seems to have a typo in Assumption 2: \" such that, i.e.\"  -> \"i.e.,\"", "clarity,_quality,_novelty_and_reproducibility": "Clarity\n\nThe paper is clear in its assumptions and does a good job of verifying its hypothesis.\n\nNovelty\n\nThe paper presents a novel approach to the problem of interest and highlights the differences compared to previous works.\n\nReproducibility\n\nOne may find it challenging to reproduce the results, given the authors did not publish their source code.", "summary_of_the_review": "This paper proposes to advance the literature by leveraging recent developments on NNs to tackle the Treatment Effect Estimation problem. They provide extensive experiments and detailed information about their proposed method. However, they argued that previous approaches have issues with real-world applicability. When they ignore the presence of confounders, this paper seems to suffer from the same applicability problem.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666783883090}, {"id": "b9QyWIo0ono", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2191/Reviewer_u5ZN"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors propose to adopt Transformers as model backbones for treatment effect estimation. Comprehensive experiments are conducted to show the effectiveness of the proposed Transformer-based TEE methods over the existing MLP-based TEE methods. ", "review_text": "The studied problem is very important in practice and the paper is well-written. It is a comprehensive paper, but not a technically novel one. There are also some technical questions in the experiments.", "strengths": "Strength:\nThe studied problem is very important in practice and the paper is well-written. The paper is a comprehensive study that explores the effectiveness and efficiency of Transformers in TEE problems.\n\nWeaknesses:\n1.  The technical contribution of this paper is not very solid. \n\ta. The Transformers are well-studied and show superior performance over other neural models (MLPs, CNN or RNN) in many domains including the TEE (e.g., CETransformer and ANU). It is not very surprising that the proposed Transformer-based method can also outperform the baselines in TEE problems.\n\tb. The treatment prediction head is trained in an adversarial pattern to alleviate the treatment bias which is already been used in many existing works (e.g., Bica et al., 2020; Kallus 2020). \n\n2. For the covariate adjustment via cross-attention module, the learned attention weights tend to be very unstable and thus unreliable to be used for covariate selection and interpretability. As shown in Table 4, different variations of the proposed model yields different and inconsistent results in learned attention weights. The results will become more unstable in real-world scenarios with high dimensional covariate space and complex relationships among the covariates. \n\n3. The motivation behind using separated linear layers for treatment and dosage is to alleviate the issue of treatment distribution shift in the test (i.e., some treatment values in the test are not seen during training). However, whether this motivation and experiment design will violate the positivity assumption that the probability of receiving any treatment is non-zero? For example, in extrapolation (h=5), t\\in[0.25, 5.0] in training and t\\in[0, 5] in testing. Thus the individuals in the training set will have zero probability to receive t\\in[0, 0.25).  Then the problem becomes whether it is meaningful to study the treatment distribution shift problem in TEE under the positivity assumption.\n\n4. It would be great to incorporate the comparison (from conceptual-level and empirical analysis) with existing Transformer-based TEE methods in the main paper. These methods should be the most related baselines to compare with.\n\nBica, Ioana, et al. \"Estimating counterfactual treatment outcomes over time through adversarially balanced representations.\" ICLR 2020.\nKallus, Nathan. \"Deepmatch: Balancing deep covariate representations for causal inference using adversarial training.\" ICML 2020.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The authors propose to adopt Transformers as model backbones for treatment effect estimation. Comprehensive experiments are conducted to show the effectiveness of the proposed Transformer-based TEE methods over the existing MLP-based TEE methods. ", "strength_and_weaknesses": "Strength:\nThe studied problem is very important in practice and the paper is well-written. The paper is a comprehensive study that explores the effectiveness and efficiency of Transformers in TEE problems.\n\nWeaknesses:\n1.  The technical contribution of this paper is not very solid. \n\ta. The Transformers are well-studied and show superior performance over other neural models (MLPs, CNN or RNN) in many domains including the TEE (e.g., CETransformer and ANU). It is not very surprising that the proposed Transformer-based method can also outperform the baselines in TEE problems.\n\tb. The treatment prediction head is trained in an adversarial pattern to alleviate the treatment bias which is already been used in many existing works (e.g., Bica et al., 2020; Kallus 2020). \n\n2. For the covariate adjustment via cross-attention module, the learned attention weights tend to be very unstable and thus unreliable to be used for covariate selection and interpretability. As shown in Table 4, different variations of the proposed model yields different and inconsistent results in learned attention weights. The results will become more unstable in real-world scenarios with high dimensional covariate space and complex relationships among the covariates. \n\n3. The motivation behind using separated linear layers for treatment and dosage is to alleviate the issue of treatment distribution shift in the test (i.e., some treatment values in the test are not seen during training). However, whether this motivation and experiment design will violate the positivity assumption that the probability of receiving any treatment is non-zero? For example, in extrapolation (h=5), t\\in[0.25, 5.0] in training and t\\in[0, 5] in testing. Thus the individuals in the training set will have zero probability to receive t\\in[0, 0.25).  Then the problem becomes whether it is meaningful to study the treatment distribution shift problem in TEE under the positivity assumption.\n\n4. It would be great to incorporate the comparison (from conceptual-level and empirical analysis) with existing Transformer-based TEE methods in the main paper. These methods should be the most related baselines to compare with.\n\nBica, Ioana, et al. \"Estimating counterfactual treatment outcomes over time through adversarially balanced representations.\" ICLR 2020.\nKallus, Nathan. \"Deepmatch: Balancing deep covariate representations for causal inference using adversarial training.\" ICML 2020.\n", "clarity,_quality,_novelty_and_reproducibility": "1. The paper is well-written and easy to follow.\n2. The paper quality is good with very comprehensive experiments.\n3. The technical novelty is not enough. Both Transformers and adversarial learning techniques have been used in the existing works.\n4. The code is provided in the supplementary material.", "summary_of_the_review": "The studied problem is very important in practice and the paper is well-written. It is a comprehensive paper, but not a technically novel one. There are also some technical questions in the experiments.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666666097451}, {"id": "zExKWZBas3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2191/Reviewer_NSTZ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper carefully designs a transformer as the backbone for treatment effect estimation, which is applicable to various tasks. They conduct many experiments to demonstrate the effectiveness of the proposed backbone and analyze the properties of the proposed transformer based on these results.", "review_text": "This paper conducts extensive experiments to verify the effectiveness of transformer. This paper demonstrates that transformer has many good properties in the causal domain. From the paper's perspective, the approach proposed in this paper lacks a certain degree of novelty.", "strengths": "Strength.\n1.\tThis paper verifies the effectiveness of the transformer for treatment effect estimation tasks, which is important for causality. And they conducted many experiments to demonstrate the effectiveness of the proposed transformer.\n2.\tThey analyze in detail the essential properties of the proposed method in causal: compatibility with propensity score modeling, parameter efficiency, robustness to continuous treatment value distribution shifts, and explainable in covariate adjustment.\n\nWeaknesses.\nThe proposed transformer lacks novelty in the methodology.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper carefully designs a transformer as the backbone for treatment effect estimation, which is applicable to various tasks. They conduct many experiments to demonstrate the effectiveness of the proposed backbone and analyze the properties of the proposed transformer based on these results.", "strength_and_weaknesses": "Strength.\n1.\tThis paper verifies the effectiveness of the transformer for treatment effect estimation tasks, which is important for causality. And they conducted many experiments to demonstrate the effectiveness of the proposed transformer.\n2.\tThey analyze in detail the essential properties of the proposed method in causal: compatibility with propensity score modeling, parameter efficiency, robustness to continuous treatment value distribution shifts, and explainable in covariate adjustment.\n\nWeaknesses.\nThe proposed transformer lacks novelty in the methodology.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper is clearly written and high-quality, but slightly lacks novelty.\nThis is not the first transformer-based paper I've seen in the causal field.\n", "summary_of_the_review": "This paper conducts extensive experiments to verify the effectiveness of transformer. This paper demonstrates that transformer has many good properties in the causal domain. From the paper's perspective, the approach proposed in this paper lacks a certain degree of novelty.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666577355694}], "openreview_url": "https://openreview.net/forum?id=C9sU3Tnnki8", "arxiv_id": "2202.01336", "paper_pdf": "papers/C9sU3Tnnki8.pdf", "paper_pdf_sha256": "d79793737fd52fbf61c0c613fd81558970a912d4679d85fc6055ecc156fcb819", "paper_pdf_bytes": 1659403, "paper_pdf_source": "openreview", "code_url": "https://github.com/hlzhang109/TransTEE", "code_repository": "hlzhang109/TransTEE", "code_commit": "3e6883a7edd7d780185d524553b13ae633b6a229", "code_archive": "repos/C9sU3Tnnki8.zip", "code_archive_sha256": "99650922ea0a0c2cdfb009bc72b9bb866c0a7aef2052a42962b158fc1ff77c50", "code_archive_bytes": 1274747, "code_file_count": 123, "code_extensions": {".py": 123}, "github_disk_usage_kb": 1192, "github_languages": {"Python": 872932, "HTML": 2479}, "github_archived": false, "github_pushed_at": "2025-09-13T04:32:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/can-transformers-be-strong-treatment-effect"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qEGBB9YB31", "year": 2022, "status": "rejected", "title": "Where do Models go Wrong? Parameter-Space Saliency Maps for Explainability", "authors": ["Roman Levin", "Manli Shu", "Eitan Borgnia", "Furong Huang", "Micah Goldblum", "Tom Goldstein"], "authorids": ["~Roman_Levin1", "~Manli_Shu1", "~Eitan_Borgnia1", "~Furong_Huang1", "~Micah_Goldblum1", "~Tom_Goldstein1"], "authors_source": "OpenReview API", "abstract": "Conventional saliency maps highlight input features to which neural network predictions are highly sensitive. We take a different approach to saliency, in which we identify and analyze the network parameters, rather than inputs, which are responsible for erroneous decisions.  We first verify that identified salient parameters are indeed responsible for misclassification by showing that turning these parameters off improves predictions on the associated samples, more than pruning the same number of random or least salient parameters.  We further validate the link between salient parameters and network misclassification errors by observing that fine-tuning a small number of the most salient parameters on a single sample results in error correction on other samples which were misclassified for similar reasons -- nearest neighbors in the saliency space. After validating our parameter-space saliency maps, we demonstrate that samples which cause similar parameters to malfunction are semantically similar.  Further, we introduce an input-space saliency counterpart which reveals how image features cause specific network components to malfunction.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "v-Hofo75rUe", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3996/Reviewer_tkQp"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper tries to identify the most salient parameters in a neural network and shows that when these parameters are modified the predictions change. They also show that there is a relation between these parameters and regions in the input that impact the classification result.", "review_text": "The paper devises an approach so look at saliency of filters by aggregating the gradients w.r.t. the parameters belonging to that filter. In that sense it is an extension of sensitivity analysis. The paper then shows that only a limited number of filters need updating to modify the prediction. Given that some of these filters might have very many parameters the resulting change can be quite big. \n\nI believe there are 2 clear limitations in the paper.\n- it is not clear how this differs from an adversarial attack in parameter space/simply exploiting the sensitivity to certain parameters. I.e. changing parts with large gradients will change the behavior. The fact that this is done for a limited number of parameters is similar to using L1 normed adversarial attacks. \n- The evaluation in section 3.4 is quite limited in scope and looks to be more anecdotal. Extending this to an in depth evaluation would improve the paper.\n\n\nDetailed comments below\n\n\n# Intro/Methods\n- Figure 1/2: . I believe this figure can be improved by assigning each layer space corresponding of \"unit 1\" on the x-axis. Right now, the decay for the earlier layers (with fewer filters) is hard to see. By normalizing the space used by a single layer this might become easier to read.\n\n- Figure 1 and 2: I am very surprised that in Figure 1 and Figure 2-bottom the saliency averaged over filters is always positive. In contrast it is always negative in Figure 2 top.  Fig 1: states that it is not standardized. Fig 2. does not specify this but it seems to me that there was an inconsistency. \n\n- Section 2.2/ Eq 3  I did not find later on in the manuscript how the magnitude of the selected filters was increased. Instead of using the cosine distance, one could also try to only pass the gradients through the filters that were deemed most salient. But this might result in a very sparse heatmap.\n\n# Experiments\n- Section 3.1/Figure 3:  If I understand it correctly these plots only contain data from initially misclassified samples. While the evaluation is done over all misclassifications, it is the sum of experiments on single images. I.e. when a filter is removed this is done only for a single example and evaluated on this example. In that sense it is similar to an adversarial perturbation in the middle of the network. The fact that this changes the confidence should not be surprising per se.\n\n- Section 3.2/Figure 4/5: The nearest neighbor approach is done by computing the cosine similarity of the saliency maps. Since this is dominated by the higher layers it should not be surprising that this results in conceptual similarity vs input similarity.  This can also explain why the same classes are misclassified since higher level concepts are also higher in the layers. In Fig. 5 similarity of low vs higher level layers is shown. The texts states that this is done to show misbehavior, but it is not clear what is misbehaving in this plot. \n\n\n- Section 3.3 Correcting mistakes by updating salient filters is evaluated on all images independently. How was batch norm used in this case for the update? Setting it to training mode could have weird effects.  I am not convinced that it should be surprising that updating the ones with large gradients results in changes at the output. \n\n- Section 3.4\nThis section is the most problematic to me since it lacks large scale evaluation and is mostly anecdotal. \n\nFigure 8 is a set of \n# General\nMake the labels of the axes in plots larger. They are hard to read.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper tries to identify the most salient parameters in a neural network and shows that when these parameters are modified the predictions change. They also show that there is a relation between these parameters and regions in the input that impact the classification result.", "main_review": "The paper devises an approach so look at saliency of filters by aggregating the gradients w.r.t. the parameters belonging to that filter. In that sense it is an extension of sensitivity analysis. The paper then shows that only a limited number of filters need updating to modify the prediction. Given that some of these filters might have very many parameters the resulting change can be quite big. \n\nI believe there are 2 clear limitations in the paper.\n- it is not clear how this differs from an adversarial attack in parameter space/simply exploiting the sensitivity to certain parameters. I.e. changing parts with large gradients will change the behavior. The fact that this is done for a limited number of parameters is similar to using L1 normed adversarial attacks. \n- The evaluation in section 3.4 is quite limited in scope and looks to be more anecdotal. Extending this to an in depth evaluation would improve the paper.\n\n\nDetailed comments below\n\n\n# Intro/Methods\n- Figure 1/2: . I believe this figure can be improved by assigning each layer space corresponding of \"unit 1\" on the x-axis. Right now, the decay for the earlier layers (with fewer filters) is hard to see. By normalizing the space used by a single layer this might become easier to read.\n\n- Figure 1 and 2: I am very surprised that in Figure 1 and Figure 2-bottom the saliency averaged over filters is always positive. In contrast it is always negative in Figure 2 top.  Fig 1: states that it is not standardized. Fig 2. does not specify this but it seems to me that there was an inconsistency. \n\n- Section 2.2/ Eq 3  I did not find later on in the manuscript how the magnitude of the selected filters was increased. Instead of using the cosine distance, one could also try to only pass the gradients through the filters that were deemed most salient. But this might result in a very sparse heatmap.\n\n# Experiments\n- Section 3.1/Figure 3:  If I understand it correctly these plots only contain data from initially misclassified samples. While the evaluation is done over all misclassifications, it is the sum of experiments on single images. I.e. when a filter is removed this is done only for a single example and evaluated on this example. In that sense it is similar to an adversarial perturbation in the middle of the network. The fact that this changes the confidence should not be surprising per se.\n\n- Section 3.2/Figure 4/5: The nearest neighbor approach is done by computing the cosine similarity of the saliency maps. Since this is dominated by the higher layers it should not be surprising that this results in conceptual similarity vs input similarity.  This can also explain why the same classes are misclassified since higher level concepts are also higher in the layers. In Fig. 5 similarity of low vs higher level layers is shown. The texts states that this is done to show misbehavior, but it is not clear what is misbehaving in this plot. \n\n\n- Section 3.3 Correcting mistakes by updating salient filters is evaluated on all images independently. How was batch norm used in this case for the update? Setting it to training mode could have weird effects.  I am not convinced that it should be surprising that updating the ones with large gradients results in changes at the output. \n\n- Section 3.4\nThis section is the most problematic to me since it lacks large scale evaluation and is mostly anecdotal. \n\nFigure 8 is a set of \n# General\nMake the labels of the axes in plots larger. They are hard to read.", "summary_of_the_review": "I tend to vote for rejection of this paper. The effects described are similar to adversarial attacks and in most but not all cases the influence of a parameter is only investigated on the image where the saliency was computed. \nThe evaluation of the proposed approach in terms of providing pixel wise interpretability section 3.4 is limited.\n\nThis review was done after the rebuttal period as an emergency review. I had no access to the discussion with the authors which might impact my understanding.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1637097006973}, {"id": "RChSRKbbSwY", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3996/Reviewer_qR4u"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper conducts some basic experiments based on the network parameters which are responsible for erroneous decisions. To support the proposed standpoints, the authors conduct a lot of quantitative and qualitative experiments.", "review_text": "Strengths\n1. Different from the previous papers focusing on the saliency maps, this paper proposes to analyze the network parameters, which is novel and interesting.\n2. Quantitative and qualitative experiments are conducted to verify the proposed standpoints.\n3. The discoveries in the paper is interesting and inspired.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper conducts some basic experiments based on the network parameters which are responsible for erroneous decisions. To support the proposed standpoints, the authors conduct a lot of quantitative and qualitative experiments.", "main_review": "Strengths\n1. Different from the previous papers focusing on the saliency maps, this paper proposes to analyze the network parameters, which is novel and interesting.\n2. Quantitative and qualitative experiments are conducted to verify the proposed standpoints.\n3. The discoveries in the paper is interesting and inspired.\n", "summary_of_the_review": "\nI am not an expert in this field and I did read carefully about this paper. I could not find any weakness in this paper. The contributions proposed in this paper are really interesting and are verified by the proposed experiments. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635923802665}, {"id": "k5VSqenWf3-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3996/Reviewer_3o7Z"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a parameter-space saliency map to explore the salient parameters that are responsible for miscalssification. A set of experiments and visulizations are conducted on the salient parameters, leading to several interesting findings, such as, the nearest parameter neighbors share similar semantic information. Besides, the authors are also trying to improve the prediction accuarcy by turning off or fine-tuning the salient parameters. ", "review_text": "+ Though the idea of parameter-wise visualization is not new, and widely studied in other works, this paper provides an interesting perspective on the parameters responsible for misclassification. \n+ Meanwhile, the authors provides the thorough analysis and potential for applying the salient parameters on improving classification accuracy. This makes the paper with pratical use.\n\n- The proposed idea is somehow similar to the parameter-space adversarial attacks. What if use attack methods to filter out the salient parameters that mislead the final classfications? This seems more intuitive.\n- For the experiments of parameters pruning, it may be not reasonable for directly turn off the parameters, what if give them a perturbation instead? It can still reveal the sensitivity of the parameters.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a parameter-space saliency map to explore the salient parameters that are responsible for miscalssification. A set of experiments and visulizations are conducted on the salient parameters, leading to several interesting findings, such as, the nearest parameter neighbors share similar semantic information. Besides, the authors are also trying to improve the prediction accuarcy by turning off or fine-tuning the salient parameters. ", "main_review": "+ Though the idea of parameter-wise visualization is not new, and widely studied in other works, this paper provides an interesting perspective on the parameters responsible for misclassification. \n+ Meanwhile, the authors provides the thorough analysis and potential for applying the salient parameters on improving classification accuracy. This makes the paper with pratical use.\n\n- The proposed idea is somehow similar to the parameter-space adversarial attacks. What if use attack methods to filter out the salient parameters that mislead the final classfications? This seems more intuitive.\n- For the experiments of parameters pruning, it may be not reasonable for directly turn off the parameters, what if give them a perturbation instead? It can still reveal the sensitivity of the parameters.", "summary_of_the_review": "The proposed method is new as a parameter-wise visualization, but it becomes somehow similar to the  parameter-space adversarial attacks. The authors would better claim the similarity and difference of these 2 topics. Besides, adversarial attack needs to be reviewed as the related work. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No concern.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635256344667}, {"id": "VA6NcLqM32r", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3996/Reviewer_sCeW"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper devises an analytic method for explainability based on the observation of filter-wise parameter saliency distribution, and tests on several models. And several experiments are conducted to deminstrate the conjecture. The motivation is straightforward and easy to understand. ", "review_text": "[1] Eq. 1 is not clear enough, what's the meanings of K. \n\n[2] If the proposed analytic method only suits for the image classification task or the test models, the contribution would be limited. \n\n[3] As stated in Section 3.3, we can correct mistakes by fine-tuning salient filters. Does it mean a model of general performance could become a top-performing model benefiting from the proposed fine-tuning method?\n\n[4] Whether the performance-gain obtained by fine-tuning salient filters can transfer to the independent database, rather than the closed test set. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper devises an analytic method for explainability based on the observation of filter-wise parameter saliency distribution, and tests on several models. And several experiments are conducted to deminstrate the conjecture. The motivation is straightforward and easy to understand. ", "main_review": "[1] Eq. 1 is not clear enough, what's the meanings of K. \n\n[2] If the proposed analytic method only suits for the image classification task or the test models, the contribution would be limited. \n\n[3] As stated in Section 3.3, we can correct mistakes by fine-tuning salient filters. Does it mean a model of general performance could become a top-performing model benefiting from the proposed fine-tuning method?\n\n[4] Whether the performance-gain obtained by fine-tuning salient filters can transfer to the independent database, rather than the closed test set. ", "summary_of_the_review": "The contribution might be not enough for iclr.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634835435650}], "openreview_url": "https://openreview.net/forum?id=qEGBB9YB31", "arxiv_id": "2108.01335", "paper_pdf": "papers/qEGBB9YB31.pdf", "paper_pdf_sha256": "b68ead11c290531cc1a8fe901907a7727fab46ac5d5941c90224129bcbf3efa9", "paper_pdf_bytes": 13377231, "paper_pdf_source": "openreview", "code_url": "https://github.com/LevinRoman/parameter-space-saliency", "code_repository": "LevinRoman/parameter-space-saliency", "code_commit": "0e3b3d69c6e222aee6af0264d7ce3ddc6d19744e", "code_archive": "repos/qEGBB9YB31.zip", "code_archive_sha256": "1c97e98c4908db622223bf0705626adcfe9a2d6d060d70e3e47d5814a7ac8938", "code_archive_bytes": 5409635, "code_file_count": 3, "code_extensions": {".py": 3}, "github_disk_usage_kb": 5219, "github_languages": {"Python": 33081}, "github_archived": false, "github_pushed_at": "2023-03-21T23:02:02Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/where-do-models-go-wrong-parameter-space"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "tf8a4jDRFCv", "year": 2021, "status": "rejected", "title": "Learning Aggregation Functions", "authors": ["Giovanni Pellegrini", "Alessandro Tibo", "Paolo Frasconi", "Andrea Passerini", "Manfred Jaeger"], "authorids": ["~Giovanni_Pellegrini1", "~Alessandro_Tibo1", "~Paolo_Frasconi1", "~Andrea_Passerini2", "jaeger@cs.aau.dk"], "authors_source": "OpenReview API", "abstract": " Learning on sets is increasingly gaining attention in the machine learning community, due to its widespread applicability. Typically, representations over sets are computed by using fixed aggregation functions such as sum or maximum. However, recent results showed that universal function representation by sum- (or max-) decomposition requires either highly discontinuous (and thus poorly learnable) mappings, or a latent dimension equal to the maximum number of elements in the set. To mitigate this problem, we introduce LAF (Learning Aggregation Functions), a learnable aggregator for sets of arbitrary cardinality. LAF can approximate several extensively used aggregators (such as average, sum, maximum)  as well as more complex functions (e.g. variance and skewness). We report experiments on semi-synthetic and real data showing that LAF  outperforms state-of-the-art sum- (max-) decomposition architectures such as DeepSets and library-based architectures like Principal Neighborhood Aggregation.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "iZ7Xy--PhFF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3275/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper addresses the problem of finding appropriate aggregation functions that can be used for instance in deep neural network architectures. One such function maps a variable length list of reals to a scalar. \n\nThe authors investigate the possibility to learn aggregation functions from data. To that end, they investigate an Lp norm based parametric model of aggregation functions called LAF that allows to learn a wide range of function including usual aggregation functions such as mean, max or min. They however restrict these functions to lists of reals that are in the unit interval. \n\nThe theoretical part of the paper is short. The authors simply show that their aggregation model has a rather high expressive power and invoke a theorem from prior arts to outline some universality guarantee (which could actually be explicitly recalled).\n\nThe limited amount of theory is compensated by extensive numerical experiments. The authors provide experimental evidence that their model generalizes fairly well to large sets of inputs. They also show that their model can be plugged to more conventional neural net layers and is backprop friendly.\n\nThis is an overall fair contribution which is well positioned as compared to prior arts. The main issue of this paper is, in my opinion, that the impact on the ML community will be limited if their LAF aggregation model does not become a gold standard like convolutional layers have become. To achieve such goal, one misses an application in which the LAF allows to bring disruptive results by leveraging features that cannot be derived through more usual architectures. In conjunction with this remark, the practical usability of LAF would be much wider if a generalization to sets of vectors would be proposed which could learn other features such as covariance.\n\nDetailed Remarks: \n\nSec 2 : If b or e is equal to zero, then L_ab is not continuous. I think these should be limiting cases too.\n\nSec 4.1 : I fail to understand to what targets input sets are mapped to: one of the feature shown in Fig. 2 or a subset of them. Similarly, I also do not understand what is the global architecture of the approaches. It seems that have 9-dimensional outputs. Please clarify this.\n\nMaybe the plots would be more informative if the y-axis was in log scale.\n\nSec 4.2 : Do you use convolutional layers before the aggregation layer ? Are they trained together ? \n\nSec 4.3 : Can you be more specific on the dataset pre-processing ? Why not directly use the raw dataset ?\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An attempt to learn aggregation functions with neural net compatibility", "review": "This paper addresses the problem of finding appropriate aggregation functions that can be used for instance in deep neural network architectures. One such function maps a variable length list of reals to a scalar. \n\nThe authors investigate the possibility to learn aggregation functions from data. To that end, they investigate an Lp norm based parametric model of aggregation functions called LAF that allows to learn a wide range of function including usual aggregation functions such as mean, max or min. They however restrict these functions to lists of reals that are in the unit interval. \n\nThe theoretical part of the paper is short. The authors simply show that their aggregation model has a rather high expressive power and invoke a theorem from prior arts to outline some universality guarantee (which could actually be explicitly recalled).\n\nThe limited amount of theory is compensated by extensive numerical experiments. The authors provide experimental evidence that their model generalizes fairly well to large sets of inputs. They also show that their model can be plugged to more conventional neural net layers and is backprop friendly.\n\nThis is an overall fair contribution which is well positioned as compared to prior arts. The main issue of this paper is, in my opinion, that the impact on the ML community will be limited if their LAF aggregation model does not become a gold standard like convolutional layers have become. To achieve such goal, one misses an application in which the LAF allows to bring disruptive results by leveraging features that cannot be derived through more usual architectures. In conjunction with this remark, the practical usability of LAF would be much wider if a generalization to sets of vectors would be proposed which could learn other features such as covariance.\n\nDetailed Remarks: \n\nSec 2 : If b or e is equal to zero, then L_ab is not continuous. I think these should be limiting cases too.\n\nSec 4.1 : I fail to understand to what targets input sets are mapped to: one of the feature shown in Fig. 2 or a subset of them. Similarly, I also do not understand what is the global architecture of the approaches. It seems that have 9-dimensional outputs. Please clarify this.\n\nMaybe the plots would be more informative if the y-axis was in log scale.\n\nSec 4.2 : Do you use convolutional layers before the aggregation layer ? Are they trained together ? \n\nSec 4.3 : Can you be more specific on the dataset pre-processing ? Why not directly use the raw dataset ?\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603893911426}, {"id": "CFIfcS-0F2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3275/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe paper proposes a parameterized learnable aggregation function (LAF) that can aggregate a multi-set of numbers (i.e. map them to a single real-valued number). This is different to prior works such as Deep Sets that use fixed aggregation functions such as max, mean, etc.\n\nStrong Points:\n- While Deep Sets have shown that is theoretically sufficient to have a sum aggregation, it is still unclear which kind of aggregation functions work well in practice. Hence, the paper addresses an interesting research question.\n- The presented idea is rather simple, which I think is a strong point of the paper.\n- I like the analysis in Table 1 that shows how different parameterizations of the LAF correspond to different functions such as sum, min, means, etc.\n- I also like the evaluation presented in Figure 4.\n\nWeak Points:\n- The setup of the experiments with scalars is unclear. For example, a LAF is supposed to aggregate all scalars in the input. However, the paper states that the LAF model is comprised of 9 LAF functions. Why do we need 9 functions? And how are they composed into a single architecture? Similarly, the paper states that 'DeepSets contains three max units, three sum units, and three mean units'. However, a Deep Set should have one function mapping the input to an intermediate representation, a sum aggregation, and a function that maps the aggregation to the output. I don't see why the model needs three max, sum, and mean units.\n- In experiment 1, it is reported that the 'input mapping is performed by three layers with the hyperbolic tangent as non-linear activation'. However, the input is simply a multi-set of scalars. It remains unclear to me why it can make sense to map scalars with a 3-layer network. Furthermore, a sigmoid is applied in case of LAF. Hence, it is unclear if the observed performance is due to the Sigmoid or due to the LAF. Furthermore, it is unclear why the paper uses tanh as activation function while Deep Set implementations use ReLUs. Also, the architecture of the output mapping is not described.\n- Similar to experiment 1, the experimental setup in experiment 2 is unclear. Additionally, it is unclear how the aggregation of features (i.e. individual dimensions in MNIST images) is related to the aggregation function used to compute the target output of the multi-set.\n- While the problem investigated in the paper is permutation-invariant, several methods have been proposed to approximate permutation invariant problems with recurrent architectures such as LSTMs and GRUs. However, no comparison to these kinds of methods is performed. A comparison would be interesting since they also learn an aggregation function.\n- It would be interesting to see if the parameters of the LAF function are learned as expected, i.e. if they correlate with the expected values as listed in Table 1. An analysis of this question is missing.\n- The paper does not share code or data to improve the reproducibility of the experiments.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "I recommend to reject the paper, mainly due to unclear description of the method and the experiments. This needs to be improved before a proper review of the presented idea is possible. Also, comparison to more other methods should be performed.", "review": "Summary:\nThe paper proposes a parameterized learnable aggregation function (LAF) that can aggregate a multi-set of numbers (i.e. map them to a single real-valued number). This is different to prior works such as Deep Sets that use fixed aggregation functions such as max, mean, etc.\n\nStrong Points:\n- While Deep Sets have shown that is theoretically sufficient to have a sum aggregation, it is still unclear which kind of aggregation functions work well in practice. Hence, the paper addresses an interesting research question.\n- The presented idea is rather simple, which I think is a strong point of the paper.\n- I like the analysis in Table 1 that shows how different parameterizations of the LAF correspond to different functions such as sum, min, means, etc.\n- I also like the evaluation presented in Figure 4.\n\nWeak Points:\n- The setup of the experiments with scalars is unclear. For example, a LAF is supposed to aggregate all scalars in the input. However, the paper states that the LAF model is comprised of 9 LAF functions. Why do we need 9 functions? And how are they composed into a single architecture? Similarly, the paper states that 'DeepSets contains three max units, three sum units, and three mean units'. However, a Deep Set should have one function mapping the input to an intermediate representation, a sum aggregation, and a function that maps the aggregation to the output. I don't see why the model needs three max, sum, and mean units.\n- In experiment 1, it is reported that the 'input mapping is performed by three layers with the hyperbolic tangent as non-linear activation'. However, the input is simply a multi-set of scalars. It remains unclear to me why it can make sense to map scalars with a 3-layer network. Furthermore, a sigmoid is applied in case of LAF. Hence, it is unclear if the observed performance is due to the Sigmoid or due to the LAF. Furthermore, it is unclear why the paper uses tanh as activation function while Deep Set implementations use ReLUs. Also, the architecture of the output mapping is not described.\n- Similar to experiment 1, the experimental setup in experiment 2 is unclear. Additionally, it is unclear how the aggregation of features (i.e. individual dimensions in MNIST images) is related to the aggregation function used to compute the target output of the multi-set.\n- While the problem investigated in the paper is permutation-invariant, several methods have been proposed to approximate permutation invariant problems with recurrent architectures such as LSTMs and GRUs. However, no comparison to these kinds of methods is performed. A comparison would be interesting since they also learn an aggregation function.\n- It would be interesting to see if the parameters of the LAF function are learned as expected, i.e. if they correlate with the expected values as listed in Table 1. An analysis of this question is missing.\n- The paper does not share code or data to improve the reproducibility of the experiments.", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603880319705}, {"id": "ukmCfRiCNgL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3275/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nUniversal function representation guarantee requires either highly discontinuous mappings or a highly dimensional latent space. For this reason the authors propose a new parametric family of aggregation functions, called LAF (for learning aggregation functions). It can be seen as a smooth version of the class of functions that are shown in DeepSets. LAF aggregator could learn all standard aggregation functions. Moreover in experiments the autors shows that LAF surpasses other aggregation methods.\n\n=============================================================================\n\nPros:\n\n1. The authors shows, that all standard aggregation functions are achievable by varying the parameters in the formulation of LAF. Moreover LAF enables a neural network to use a continuum of intermediate and hybrid aggregators.\n\n2. Comprehensive ablation study that compares LAF to DeepSets and PNA on digits and MNIST images. In the study the goal is to learn a different types of target aggregation. The results shows that LAF could learn all the given types of aggregation methods as well as it could generalize well to the size of the test set (and thus is not overfitting to the size of the training set as the other methods).\n\n3. The authors provide an extensive set of experiments on a wide range of datasets, including point clouds, set expansion and graph properties. On most of the given tasks LAF is superior to other methods.\n\n=============================================================================\n\nCons:\n\nNot all the details about LAF aggregation are clear to me. The authors should consider rewriting a section 2 (with a description of the aggregation), considering the points I list below.\n\n1. In the manuscript the autors state that LAF is using the tunable parameters a,...,h and alpha,...,delta, however they do not show how to initialize these parameters and do not tell whether the model is sensitive to these values.\n\n2. The autors state that tunable parameters a,...,h are greater or equal than zero. However hey do not show how to achieve this condition. Whether they use exponent, non-linearity or do it in another way.\n\n3. In the definition of LAF aggregator, the authors states that x should be a real number, however looking at the experiments, it seems to me that it could also be applied to vectors. Please correct me if I'm wrong, but if I am right, than please answer the question whether in this situation a,...,h,alpha,...,delta are still the scalars or whether are they vecors? \n\n4. The more detailed description of builded networks should be included (even in the appendix). It is not clear to me, how the authors made thir networks. E.g. in section 4.1 they state that 'The LAF model contains nine LAF(x) aggregation functions' - does this mean that in the final aggregation layer you create 9 independently working LAF aggregators and then concatenate them and pass to final prediction layer? If yes, then what in the situation with aggregating the informations from vectors? Do you still make concatenaction?\n\n=============================================================================\n\nQuestions during rebuttal period:\n\n1. Why section 5 (Multi-task graph properties) is not the sub-section of section 4 (Experiments)?\n\n2. Usually using a sigmoid could disturb the training of neural network. Could you create the experiment (on a real dataset), where you delete the sigmoid as well as parts with 1-x form the LAF aggregator?\n\n3. Could you create some experiments with LAF as the final aggregator for neural networks? More specific, for exampe, could you use LAF instead of mean average pooling in image classification using ResNet or some text classification task?\n\n=============================================================================\n\n=============================================================================\n\nReasons for score: \n\nI vote for accepting this paper. The idea proposed by the authors is novel and elegant. Moreover experiments shows that the proposed model is superior to the other models with whom it has been compared. My major concern is about the clarity of the paper. Hopefully the authors can address my concern in the rebuttal period. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A novel method for aggregating the information from sets", "review": "Summary:\n\nUniversal function representation guarantee requires either highly discontinuous mappings or a highly dimensional latent space. For this reason the authors propose a new parametric family of aggregation functions, called LAF (for learning aggregation functions). It can be seen as a smooth version of the class of functions that are shown in DeepSets. LAF aggregator could learn all standard aggregation functions. Moreover in experiments the autors shows that LAF surpasses other aggregation methods.\n\n=============================================================================\n\nPros:\n\n1. The authors shows, that all standard aggregation functions are achievable by varying the parameters in the formulation of LAF. Moreover LAF enables a neural network to use a continuum of intermediate and hybrid aggregators.\n\n2. Comprehensive ablation study that compares LAF to DeepSets and PNA on digits and MNIST images. In the study the goal is to learn a different types of target aggregation. The results shows that LAF could learn all the given types of aggregation methods as well as it could generalize well to the size of the test set (and thus is not overfitting to the size of the training set as the other methods).\n\n3. The authors provide an extensive set of experiments on a wide range of datasets, including point clouds, set expansion and graph properties. On most of the given tasks LAF is superior to other methods.\n\n=============================================================================\n\nCons:\n\nNot all the details about LAF aggregation are clear to me. The authors should consider rewriting a section 2 (with a description of the aggregation), considering the points I list below.\n\n1. In the manuscript the autors state that LAF is using the tunable parameters a,...,h and alpha,...,delta, however they do not show how to initialize these parameters and do not tell whether the model is sensitive to these values.\n\n2. The autors state that tunable parameters a,...,h are greater or equal than zero. However hey do not show how to achieve this condition. Whether they use exponent, non-linearity or do it in another way.\n\n3. In the definition of LAF aggregator, the authors states that x should be a real number, however looking at the experiments, it seems to me that it could also be applied to vectors. Please correct me if I'm wrong, but if I am right, than please answer the question whether in this situation a,...,h,alpha,...,delta are still the scalars or whether are they vecors? \n\n4. The more detailed description of builded networks should be included (even in the appendix). It is not clear to me, how the authors made thir networks. E.g. in section 4.1 they state that 'The LAF model contains nine LAF(x) aggregation functions' - does this mean that in the final aggregation layer you create 9 independently working LAF aggregators and then concatenate them and pass to final prediction layer? If yes, then what in the situation with aggregating the informations from vectors? Do you still make concatenaction?\n\n=============================================================================\n\nQuestions during rebuttal period:\n\n1. Why section 5 (Multi-task graph properties) is not the sub-section of section 4 (Experiments)?\n\n2. Usually using a sigmoid could disturb the training of neural network. Could you create the experiment (on a real dataset), where you delete the sigmoid as well as parts with 1-x form the LAF aggregator?\n\n3. Could you create some experiments with LAF as the final aggregator for neural networks? More specific, for exampe, could you use LAF instead of mean average pooling in image classification using ResNet or some text classification task?\n\n=============================================================================\n\n=============================================================================\n\nReasons for score: \n\nI vote for accepting this paper. The idea proposed by the authors is novel and elegant. Moreover experiments shows that the proposed model is superior to the other models with whom it has been compared. My major concern is about the clarity of the paper. Hopefully the authors can address my concern in the rebuttal period. \n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603811289966}, {"id": "fwm_mZaQpTI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper3275/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a rational approximation approach for learning aggregation functions. In general the paper is well written and technically sound, although there are some questions that could help the readers better understand the paper. \n\nOne of the problems with learnable rational approximations is the potential of finding a pole, e.g., x/0, you do not mention how you avoid/use? such a situation in your approach, whether this causes instabilities during learning, etc. Could you mention something about this?\n\nRegarding the use of the sigmoid to transform the values of x to the range [0,1], it is not clear to me, how you can recover the mean with such low error, i.e., how can you achieve $\\mu(x) ~=  (\\sum sigmoid(x)^b)^a / N$. Without the sigmoid transformation it is clear as presented in table 1. But with the sigmoid transformation is not as straightforward to see, although you show very good results in Fig. 2. Furthermore the sigmoid transformation destroys the linearity needed when the values of x are larger. Could you also give some intuition on why using the sigmoid is superior compared to a minmax scaling process?\n\nYou mention that you use 9 LAF(x) aggregation functions in the experiments, could you explain more what do you mean exactly? are you using a mixture of LAF(x) models? are they independent of each other and train for the different target functions?\n\nIn general, I find the paper interesting and the results promising, just a bit more explanation could help the reader understand the benefits and intuition behind the decisions. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Learning Aggregation Functions over sets.", "review": "The paper proposes a rational approximation approach for learning aggregation functions. In general the paper is well written and technically sound, although there are some questions that could help the readers better understand the paper. \n\nOne of the problems with learnable rational approximations is the potential of finding a pole, e.g., x/0, you do not mention how you avoid/use? such a situation in your approach, whether this causes instabilities during learning, etc. Could you mention something about this?\n\nRegarding the use of the sigmoid to transform the values of x to the range [0,1], it is not clear to me, how you can recover the mean with such low error, i.e., how can you achieve $\\mu(x) ~=  (\\sum sigmoid(x)^b)^a / N$. Without the sigmoid transformation it is clear as presented in table 1. But with the sigmoid transformation is not as straightforward to see, although you show very good results in Fig. 2. Furthermore the sigmoid transformation destroys the linearity needed when the values of x are larger. Could you also give some intuition on why using the sigmoid is superior compared to a minmax scaling process?\n\nYou mention that you use 9 LAF(x) aggregation functions in the experiments, could you explain more what do you mean exactly? are you using a mixture of LAF(x) models? are they independent of each other and train for the different target functions?\n\nIn general, I find the paper interesting and the results promising, just a bit more explanation could help the reader understand the benefits and intuition behind the decisions. ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603641538242}], "openreview_url": "https://openreview.net/forum?id=tf8a4jDRFCv", "arxiv_id": "2012.08482", "paper_pdf": "papers/tf8a4jDRFCv.pdf", "paper_pdf_sha256": "21ae87a965314846c24a5d6a76ea2e46fb7044864a64c16bece07908fc5dfd7e", "paper_pdf_bytes": 2388924, "paper_pdf_source": "openreview", "code_url": "https://github.com/alessandro-t/laf", "code_repository": "alessandro-t/laf", "code_commit": "670aafbad87292343de04fa7a16211fded484ce0", "code_archive": "repos/tf8a4jDRFCv.zip", "code_archive_sha256": "38e63ceb85e6265530147eed80f5d1f9ab58d94ca2c84c764177ac12314c466d", "code_archive_bytes": 2981004, "code_file_count": 10, "code_extensions": {".py": 9, ".sh": 1}, "github_disk_usage_kb": 2952, "github_languages": {"Python": 42444, "Shell": 432}, "github_archived": false, "github_pushed_at": "2021-05-17T15:03:00Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-aggregation-functions-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RZIfy4Qzxa", "year": 2026, "status": "rejected", "title": "MedSentry: Understanding and Mitigating Safety Risks in Medical LLM Multi-Agent Systems", "authors": ["Kai Chen", "Taihang Zhen", "Hewei Wang", "Kailai Liu", "Xinfeng Li", "Jing Huo", "Tianpei Yang", "Jinfeng Xu", "Wei Dong", "Yang Gao"], "authorids": ["~Kai_Chen37", "~Taihang_Zhen1", "~Hewei_Wang1", "~Kailai_Liu1", "~Xinfeng_Li1", "~Jing_Huo2", "~Tianpei_Yang1", "~Jinfeng_Xu2", "~Wei_Dong5", "~Yang_Gao3"], "authors_source": "OpenReview API", "abstract": "As large language models are increasingly adopted in healthcare, ensuring their safety is critical, particularly in collaborative multi-agent settings. This paper develops an end-to-end attack–defense evaluation workflow to systematically analyze how four representative multi-agent topologies (Layers, SharedPool, Centralized, and Decentralized) behave under attacks from ``dark-personality'' agents. To support the evaluation, we curate MedSentry, a data resource containing 5,000 adversarial medical prompts that span 25 threat topics and 100 subtopics. Our study reveals critical differences in how these architectures handle information contamination and maintain robust decision-making, exposing their underlying vulnerability mechanisms. For example, SharedPool is highly susceptible due to open information sharing, whereas Decentralized exhibits stronger resilience owing to inherent redundancy and isolation. To mitigate these risks, we propose a personality-scale detection and correction mechanism that identifies and rehabilitates malicious agents, restoring safety to near-baseline levels. Taken together, MedSentry provides a rigorous evaluation framework alongside actionable defense strategies, offering guidance for the design of safer LLM-based multi-agent systems in medical contexts.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Zh77TiWg6M", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12106/Reviewer_1psf"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "This paper introduces MedSentry, a benchmark with 5,000 adversarial medical prompts across 25 categories and 100 subtopics, to evaluate safety risks in medical LLM multi-agent systems. The authors compare four topologies (Layers, SharedPool, Centralized, Decentralized) under attacks from \"dark-personality\" agents and propose PCDC (Personality-scale Detection and Correction) as a defense. Results show SharedPool is most vulnerable while Decentralized is most resilient. PCDC reportedly restores safety to near-baseline levels.", "review_text": "This paper introduces MedSentry, a benchmark with 5,000 adversarial medical prompts across 25 categories and 100 subtopics, to evaluate safety risks in medical LLM multi-agent systems. The authors compare four topologies (Layers, SharedPool, Centralized, Decentralized) under attacks from \"dark-personality\" agents and propose PCDC (Personality-scale Detection and Correction) as a defense. Results show SharedPool is most vulnerable while Decentralized is most resilient. PCDC reportedly restores safety to near-baseline levels.", "strengths": "1. The paper focuses on multi-agent systems, which are increasingly deployed in healthcare but remain underexplored from a security perspective. This is a valuable contribution distinct from single-model medical AI benchmarks.  \n\n2. The paper is a systematic investigation of topological vulnerabilities in medical multi-agent systems. The finding that topology choice significantly impacts safety is actionable for system designers.  \n\n3. Benchmark scale and organization appear comprehensive in the paper. With 5,000 prompts across 25 categories and 100 subtopics, MedSentry provides reasonable coverage of medical adversarial scenarios. The two-level taxonomy structure allows for granular analysis of attack patterns across medical subdomains.", "weaknesses": "1. The paper introduces \"dark-personality\" adversarial agents without a reasonable threat model. Dark Triad and other psychometrics are human constructs with questionable applicability to AI agents. This fundamental conceptual issue undermines the entire PCDC and evaluation framework. \n\n2. The paper needs component-wise analysis to understand what drives the results. The paper does not do ablation studies on components of PCDC. Regarding \"the defense pipeline as an integrated end to end process\" is not a reasonable excuse for not doing ablation studies.\n\n3. All experiments conducted in controlled simulation environments with no evidence of behavior in actual clinical scenarios. Clinician validation of generated outputs are not validation of system integrated in real-world clinical scenarios.", "questions": "1. You thoroughly measure safety across topologies, but do safer architectures sacrifice task performance on medical tasks? What is the tradeoff between safety and performance? This is important for practical deployment decisions.\n\n2. What is PCDC's false positive rate? how often are benign agents incorrectly flagged as malicious?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces MedSentry, a benchmark with 5,000 adversarial medical prompts across 25 categories and 100 subtopics, to evaluate safety risks in medical LLM multi-agent systems. The authors compare four topologies (Layers, SharedPool, Centralized, Decentralized) under attacks from \"dark-personality\" agents and propose PCDC (Personality-scale Detection and Correction) as a defense. Results show SharedPool is most vulnerable while Decentralized is most resilient. PCDC reportedly restores safety to near-baseline levels.", "soundness": 2, "presentation": 2, "contribution": 3, "strengths": "1. The paper focuses on multi-agent systems, which are increasingly deployed in healthcare but remain underexplored from a security perspective. This is a valuable contribution distinct from single-model medical AI benchmarks.  \n\n2. The paper is a systematic investigation of topological vulnerabilities in medical multi-agent systems. The finding that topology choice significantly impacts safety is actionable for system designers.  \n\n3. Benchmark scale and organization appear comprehensive in the paper. With 5,000 prompts across 25 categories and 100 subtopics, MedSentry provides reasonable coverage of medical adversarial scenarios. The two-level taxonomy structure allows for granular analysis of attack patterns across medical subdomains.", "weaknesses": "1. The paper introduces \"dark-personality\" adversarial agents without a reasonable threat model. Dark Triad and other psychometrics are human constructs with questionable applicability to AI agents. This fundamental conceptual issue undermines the entire PCDC and evaluation framework. \n\n2. The paper needs component-wise analysis to understand what drives the results. The paper does not do ablation studies on components of PCDC. Regarding \"the defense pipeline as an integrated end to end process\" is not a reasonable excuse for not doing ablation studies.\n\n3. All experiments conducted in controlled simulation environments with no evidence of behavior in actual clinical scenarios. Clinician validation of generated outputs are not validation of system integrated in real-world clinical scenarios.", "questions": "1. You thoroughly measure safety across topologies, but do safer architectures sacrifice task performance on medical tasks? What is the tradeoff between safety and performance? This is important for practical deployment decisions.\n\n2. What is PCDC's false positive rate? how often are benign agents incorrectly flagged as malicious?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761953500986}, {"id": "3l6nVcHNo9", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12106/Reviewer_Pmvt"], "rating": 4, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper develops a framework for understanding and mitigating safety risks in medical LLM multi-agent systems. It constructs a dataset of 5k adversarial medical prompts covering diverse topics, and tests how different architectures respond to internal “dark-personality” agents. To counter threats, the authors propose a mechanism combining an enforcement agent to monitor. Overall, MedSentry provides an end-to-end attack–defense evaluation pipeline and actionable defense strategies for building safer medical LLM multi-agent systems.", "review_text": "The paper develops a framework for understanding and mitigating safety risks in medical LLM multi-agent systems. It constructs a dataset of 5k adversarial medical prompts covering diverse topics, and tests how different architectures respond to internal “dark-personality” agents. To counter threats, the authors propose a mechanism combining an enforcement agent to monitor. Overall, MedSentry provides an end-to-end attack–defense evaluation pipeline and actionable defense strategies for building safer medical LLM multi-agent systems.", "strengths": "- The paper tackles an important and underexplored problem, evaluating the safety of medical LLM agents under multi-agent setting.\n\n- The benchmark design is comprehensive, encompassing multiple medical stages and diverse risk categories. This makes it representative.\n\n- Extensive experiments are done to analyze a wide range of factors that could possibly affect the risk of the multi-agent system.", "weaknesses": "- The description of the attack instruction process in *Coarse-Grained Data Generation* is vague. It is unclear how the process is iterative and how \"topics and subtopics are substituted\" across iterations. More explicit examples or algorithmic details would improve clarity.\n- The evaluation design may introduce confounding factors across different topologies. As mentioned in Lines 209–210, the evaluator agent receives different amounts of information under different setups. This raises the concern that score differences may partly stem from unequal access to information rather than genuine behavioral differences. Moreover some evaluation criteria listed in Appendix C.3 may not be observable from a single-turn response (e.g., the final summary from layer topology), which could further bias topology-specific scores. What if there is no enough information for the evaluator to apply a certain criteria? Will it assign a high value or it just leave it as zero?\n- The comparison in Table 2 seems werid. Since MedSentry and MedSafetyBench use different models, their scores are not directly comparable. Therefore, the conclusion that “MedSentry possesses greater threat potential and concealment than MedSafetyBench” is not fully supported.\n- A minor problem, I don't know the reason for using LCS metric. Its purpose and interpretability in the context of medical safety evaluation are unclear, and it is not intuitive how this metric captures meaningful behavioral or safety-related differences between models.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper develops a framework for understanding and mitigating safety risks in medical LLM multi-agent systems. It constructs a dataset of 5k adversarial medical prompts covering diverse topics, and tests how different architectures respond to internal “dark-personality” agents. To counter threats, the authors propose a mechanism combining an enforcement agent to monitor. Overall, MedSentry provides an end-to-end attack–defense evaluation pipeline and actionable defense strategies for building safer medical LLM multi-agent systems.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The paper tackles an important and underexplored problem, evaluating the safety of medical LLM agents under multi-agent setting.\n\n- The benchmark design is comprehensive, encompassing multiple medical stages and diverse risk categories. This makes it representative.\n\n- Extensive experiments are done to analyze a wide range of factors that could possibly affect the risk of the multi-agent system.", "weaknesses": "- The description of the attack instruction process in *Coarse-Grained Data Generation* is vague. It is unclear how the process is iterative and how \"topics and subtopics are substituted\" across iterations. More explicit examples or algorithmic details would improve clarity.\n- The evaluation design may introduce confounding factors across different topologies. As mentioned in Lines 209–210, the evaluator agent receives different amounts of information under different setups. This raises the concern that score differences may partly stem from unequal access to information rather than genuine behavioral differences. Moreover some evaluation criteria listed in Appendix C.3 may not be observable from a single-turn response (e.g., the final summary from layer topology), which could further bias topology-specific scores. What if there is no enough information for the evaluator to apply a certain criteria? Will it assign a high value or it just leave it as zero?\n- The comparison in Table 2 seems werid. Since MedSentry and MedSafetyBench use different models, their scores are not directly comparable. Therefore, the conclusion that “MedSentry possesses greater threat potential and concealment than MedSafetyBench” is not fully supported.\n- A minor problem, I don't know the reason for using LCS metric. Its purpose and interpretability in the context of medical safety evaluation are unclear, and it is not intuitive how this metric captures meaningful behavioral or safety-related differences between models.", "questions": "See above", "flag_for_ethics_review": ["No ethics review needed."], "rating": 4, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1761832975207}, {"id": "CqNk0LJj08", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12106/Reviewer_gi3X"], "rating": 2, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper introduces an end-to-end attack–defense evaluation framework for assessing the safety of large language models in collaborative multi-agent healthcare settings. Focusing on four representative topologies—Layers, SharedPool, Centralized, and Decentralized—the authors analyze their vulnerability to attacks from \"dark-personality\" agents using a newly curated dataset, MedSentry, comprising 5,000 adversarial medical prompts across 25 threat topics. The study uncovers key differences in each topology's resilience, with SharedPool being highly vulnerable due to open information sharing, while Decentralized shows greater robustness. To enhance safety, the paper proposes a personality-scale detection and correction mechanism that effectively mitigates adversarial influence. Overall, this work offers a systematic framework and practical strategies for designing safer LLM-based multi-agent systems in medical applications.", "review_text": "This paper introduces an end-to-end attack–defense evaluation framework for assessing the safety of large language models in collaborative multi-agent healthcare settings. Focusing on four representative topologies—Layers, SharedPool, Centralized, and Decentralized—the authors analyze their vulnerability to attacks from \"dark-personality\" agents using a newly curated dataset, MedSentry, comprising 5,000 adversarial medical prompts across 25 threat topics. The study uncovers key differences in each topology's resilience, with SharedPool being highly vulnerable due to open information sharing, while Decentralized shows greater robustness. To enhance safety, the paper proposes a personality-scale detection and correction mechanism that effectively mitigates adversarial influence. Overall, this work offers a systematic framework and practical strategies for designing safer LLM-based multi-agent systems in medical applications.", "strengths": "1. Investigating the resilience of various multi-agent architectures to adversarial prompts is a valuable and relevant direction, offering insights into the robustness and design trade-offs of safety-critical LLM systems.", "weaknesses": "1. Existing safety benchmarks, like MedSafetyBench, in the medical domain already address various risks, and it is unclear how the proposed benchmark distinguishes itself, aside from the inclusion of a malicious agent.\n2. The evaluated scenarios lack practical relevance, as the likelihood of inserting a malicious agent into real-world medical systems is very low. \n3. The study does not employ advanced attack methods like [1, 2], resulting in prompts that are not sufficiently adversarial to meaningfully challenge the evaluated systems.\n\n\n[1] Liu, X., Xu, N., Chen, M. and Xiao, C., 2023. Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.04451.\n\n[2] Andriushchenko, M., Croce, F. and Flammarion, N., 2024. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151.", "questions": "1. The margins of the figures and tables have been adjusted in a way that impairs the readability of the paper. Clearer formatting is needed to improve visual clarity and overall presentation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces an end-to-end attack–defense evaluation framework for assessing the safety of large language models in collaborative multi-agent healthcare settings. Focusing on four representative topologies—Layers, SharedPool, Centralized, and Decentralized—the authors analyze their vulnerability to attacks from \"dark-personality\" agents using a newly curated dataset, MedSentry, comprising 5,000 adversarial medical prompts across 25 threat topics. The study uncovers key differences in each topology's resilience, with SharedPool being highly vulnerable due to open information sharing, while Decentralized shows greater robustness. To enhance safety, the paper proposes a personality-scale detection and correction mechanism that effectively mitigates adversarial influence. Overall, this work offers a systematic framework and practical strategies for designing safer LLM-based multi-agent systems in medical applications.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Investigating the resilience of various multi-agent architectures to adversarial prompts is a valuable and relevant direction, offering insights into the robustness and design trade-offs of safety-critical LLM systems.", "weaknesses": "1. Existing safety benchmarks, like MedSafetyBench, in the medical domain already address various risks, and it is unclear how the proposed benchmark distinguishes itself, aside from the inclusion of a malicious agent.\n2. The evaluated scenarios lack practical relevance, as the likelihood of inserting a malicious agent into real-world medical systems is very low. \n3. The study does not employ advanced attack methods like [1, 2], resulting in prompts that are not sufficiently adversarial to meaningfully challenge the evaluated systems.\n\n\n[1] Liu, X., Xu, N., Chen, M. and Xiao, C., 2023. Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.04451.\n\n[2] Andriushchenko, M., Croce, F. and Flammarion, N., 2024. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151.", "questions": "1. The margins of the figures and tables have been adjusted in a way that impairs the readability of the paper. Clearer formatting is needed to improve visual clarity and overall presentation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 2, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1760953696338}, {"id": "Mjf7fV5gk8", "reviewer_signature": ["ICLR.cc/2026/Conference/Submission12106/Reviewer_erPk"], "rating": 8, "soundness": 3, "presentation": 3, "contribution": 4, "confidence": 5, "summary": "This paper targets the critical challenge of safety in medical multi agent systems and introduces MedSentry, an evaluation and adversarial framework that aggregates 5,000 adversarial medical prompts spanning 25 topics and 100 subtopics. The system injects malicious agents with dark personality traits, compares the resilience of four mainstream topologies, and quantifies safety using AMA ethical guidelines. The pipeline is clearly specified and forms a closed loop, the scenarios mirror collaborative clinical workflows, and both data and protocols are reproducible and extensible, which gives the overall design strong engineering rigor and research value. Experiments show that SharedPool is most susceptible to information contamination, Decentralized is the most robust, and Layers and Centralized fall in between. The proposed PCDC enforcement agent combines personality scale screening, behavioral verification, and topology aware isolation, effectively restoring safety scores after attack toward the baseline. The paper delivers systematic benchmarking and mechanistic insights and provides actionable tools and clear directions for improving the safety of medical multi agent systems.", "review_text": "This paper targets the critical challenge of safety in medical multi agent systems and introduces MedSentry, an evaluation and adversarial framework that aggregates 5,000 adversarial medical prompts spanning 25 topics and 100 subtopics. The system injects malicious agents with dark personality traits, compares the resilience of four mainstream topologies, and quantifies safety using AMA ethical guidelines. The pipeline is clearly specified and forms a closed loop, the scenarios mirror collaborative clinical workflows, and both data and protocols are reproducible and extensible, which gives the overall design strong engineering rigor and research value. Experiments show that SharedPool is most susceptible to information contamination, Decentralized is the most robust, and Layers and Centralized fall in between. The proposed PCDC enforcement agent combines personality scale screening, behavioral verification, and topology aware isolation, effectively restoring safety scores after attack toward the baseline. The paper delivers systematic benchmarking and mechanistic insights and provides actionable tools and clear directions for improving the safety of medical multi agent systems.", "strengths": "- The topic is timely and fills a clear gap, focusing on the intersection of LLM safety and medical multi agent systems, with high importance and strong real-world relevance.\n- The benchmark is rigorously constructed with clinical expert input, includes fine grained and stealthy adversarial prompts, and is more effective than existing datasets at eliciting unsafe behavior.\n- The architectural comparison is insightful, with systematic experiments indicating the resilience of decentralized designs and the vulnerability of shared pool settings, which offers actionable guidance for system selection.\n- The mitigation is practical and deployable. PCDC is lightweight, interpretable, and effective, integrating personality assessment with behavioral verification, which facilitates engineering deployment and compliance auditing.\n- The experimental scope is comprehensive, with detailed comparisons to existing benchmarks across system structures and with attack and defense evaluations for each basic topology. The study further examines agent count, dialogue rounds, and token level safety trends and vulnerabilities, which strengthen robustness and generalizability. The appendix provides ample details and extended studies that corroborate the main text.", "weaknesses": "- In Figure 1, the Claude logo shows slight tearing artifacts and alignment instability, which should be refined.\n- \"Attack Obfuscation and Purification\" relies on the same original model used to generate or obfuscate data and does not evaluate cross-model purification or generalizability.\n- In the proposed defense, what is the rationale for monitoring only the initial and follow-up turns rather than auditing all remaining turns? How would expanding coverage to every turn influence safety outcomes?\n- Consider adding an outlook on real-world application scenarios for the four topologies in medical settings, either in the Introduction or the Appendix.", "questions": "- How do you expect the proposed defenses to generalize to datasets beyond MedSentry?\n- Could the Token Usage section (Appendix) include a comparison between defense-enabled and defense-disabled settings, covering per-round latency and total token consumption, to help assess deployment costs?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper targets the critical challenge of safety in medical multi agent systems and introduces MedSentry, an evaluation and adversarial framework that aggregates 5,000 adversarial medical prompts spanning 25 topics and 100 subtopics. The system injects malicious agents with dark personality traits, compares the resilience of four mainstream topologies, and quantifies safety using AMA ethical guidelines. The pipeline is clearly specified and forms a closed loop, the scenarios mirror collaborative clinical workflows, and both data and protocols are reproducible and extensible, which gives the overall design strong engineering rigor and research value. Experiments show that SharedPool is most susceptible to information contamination, Decentralized is the most robust, and Layers and Centralized fall in between. The proposed PCDC enforcement agent combines personality scale screening, behavioral verification, and topology aware isolation, effectively restoring safety scores after attack toward the baseline. The paper delivers systematic benchmarking and mechanistic insights and provides actionable tools and clear directions for improving the safety of medical multi agent systems.", "soundness": 3, "presentation": 3, "contribution": 4, "strengths": "- The topic is timely and fills a clear gap, focusing on the intersection of LLM safety and medical multi agent systems, with high importance and strong real-world relevance.\n- The benchmark is rigorously constructed with clinical expert input, includes fine grained and stealthy adversarial prompts, and is more effective than existing datasets at eliciting unsafe behavior.\n- The architectural comparison is insightful, with systematic experiments indicating the resilience of decentralized designs and the vulnerability of shared pool settings, which offers actionable guidance for system selection.\n- The mitigation is practical and deployable. PCDC is lightweight, interpretable, and effective, integrating personality assessment with behavioral verification, which facilitates engineering deployment and compliance auditing.\n- The experimental scope is comprehensive, with detailed comparisons to existing benchmarks across system structures and with attack and defense evaluations for each basic topology. The study further examines agent count, dialogue rounds, and token level safety trends and vulnerabilities, which strengthen robustness and generalizability. The appendix provides ample details and extended studies that corroborate the main text.", "weaknesses": "- In Figure 1, the Claude logo shows slight tearing artifacts and alignment instability, which should be refined.\n- \"Attack Obfuscation and Purification\" relies on the same original model used to generate or obfuscate data and does not evaluate cross-model purification or generalizability.\n- In the proposed defense, what is the rationale for monitoring only the initial and follow-up turns rather than auditing all remaining turns? How would expanding coverage to every turn influence safety outcomes?\n- Consider adding an outlook on real-world application scenarios for the four topologies in medical settings, either in the Introduction or the Appendix.", "questions": "- How do you expect the proposed defenses to generalize to datasets beyond MedSentry?\n- Could the Token Usage section (Appendix) include a comparison between defense-enabled and defense-disabled settings, covering per-round latency and total token consumption, to help assess deployment costs?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1760632897171}], "openreview_url": "https://openreview.net/forum?id=RZIfy4Qzxa", "arxiv_id": "2505.20824", "paper_pdf": "papers/RZIfy4Qzxa.pdf", "paper_pdf_sha256": "193101017c3a2ba53a647aca86c9d14ec99a7c308215be9119c26af27b32f27a", "paper_pdf_bytes": 3930456, "paper_pdf_source": "openreview", "code_url": "https://github.com/KaiChenNJ/MedSentry", "code_repository": "KaiChenNJ/MedSentry", "code_commit": "044ba8b36e1bb30a78098b9e9a075072094081c4", "code_archive": "repos/RZIfy4Qzxa.zip", "code_archive_sha256": "7d5d6e75013ac52a428326c48d013c25acc814823bc7de50b92aecaf617d6a9e", "code_archive_bytes": 8939884, "code_file_count": 16, "code_extensions": {".py": 16}, "github_disk_usage_kb": 9979, "github_languages": {}, "github_archived": false, "github_pushed_at": "2025-05-27T05:58:53Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/medsentry-understanding-and-mitigating-safety"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EWNH3QTSxd", "year": 2025, "status": "rejected", "title": "Which Experiences Are Influential for RL Agents? Efficiently Estimating The Influence of Experiences", "authors": ["Takuya Hiraoka", "Guanquan Wang", "Takashi Onishi", "Yoshimasa Tsuruoka"], "authorids": ["~Takuya_Hiraoka1", "~Guanquan_Wang1", "~Takashi_Onishi1", "~Yoshimasa_Tsuruoka1"], "authors_source": "OpenReview API", "abstract": "In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. \nInformation about how these experiences influence the agent's performance is valuable for various purposes, such as identifying experiences that negatively influence underperforming agents. \nOne method for estimating the influence of experiences is the leave-one-out (LOO) method. \nHowever, this method is usually computationally prohibitive. \nIn this paper, we present Policy Iteration with Turn-over Dropout (PIToD), which efficiently estimates the influence of experiences. \nWe evaluate how accurately PIToD estimates the influence of experiences and its efficiency compared to LOO. \nWe then apply PIToD to amend underperforming RL agents, i.e., we use PIToD to estimate negatively influential experiences for the RL agents and to delete the influence of these experiences. \nWe show that RL agents' performance is significantly improved via amendments with PIToD.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "A6HvUrBRy2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission874/Reviewer_xR4M"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper aims to study the influence individual experience sample have in training RL agents. The authors used identify masks to differentiate individual samples and observe the difference in the resulting Q values.", "review_text": "This paper aims to study the influence individual experience sample have in training RL agents. The authors used identify masks to differentiate individual samples and observe the difference in the resulting Q values.", "strengths": "- This paper focus on experience replay, that is, sampling distribution manipulation, which I think is a under-represented direction in RL research.\n- I like the the fact that ToD being applied in the sample consideration, the usage of ToD feels natural and justified for this use case.", "weaknesses": "- The paper do not have a theoretical underpinning for their approach though some reader may find the idea intuitive, that said, personally, I'm not a fan of removing samples from experience buffers;\n- because the way I see it, deleting \"negatively influential experiences\" seems to be a lenient/hysteresis update commonly seen in optimistic approaches, it may be useful in same case but should be used with caution since it may hinder learning and cause bias; a comparison with similar approaches is not included. I think instead of removing samples, the safer approach is to use gradient clipping or learning rate schedules to stabilize training. Techniques like gradient clipping cap the maximum value of gradients to prevent instability without losing any data --\n- Data removal in RL may massively hinder exploration, especially in non-linear environments, speaking of which --\n- the paper only evaluate on classic mujoco, and lacks variety of evaluation task, speaking of evaluation --\n- The evaluation also lacks other sota methods for comparison, since the promise was to have better performance, rather than better theoretical understanding, which goes back to my first point.", "questions": "How does sample rejection based on gradient relate to hessian?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper aims to study the influence individual experience sample have in training RL agents. The authors used identify masks to differentiate individual samples and observe the difference in the resulting Q values.", "soundness": 3, "presentation": 2, "contribution": 1, "strengths": "- This paper focus on experience replay, that is, sampling distribution manipulation, which I think is a under-represented direction in RL research.\n- I like the the fact that ToD being applied in the sample consideration, the usage of ToD feels natural and justified for this use case.", "weaknesses": "- The paper do not have a theoretical underpinning for their approach though some reader may find the idea intuitive, that said, personally, I'm not a fan of removing samples from experience buffers;\n- because the way I see it, deleting \"negatively influential experiences\" seems to be a lenient/hysteresis update commonly seen in optimistic approaches, it may be useful in same case but should be used with caution since it may hinder learning and cause bias; a comparison with similar approaches is not included. I think instead of removing samples, the safer approach is to use gradient clipping or learning rate schedules to stabilize training. Techniques like gradient clipping cap the maximum value of gradients to prevent instability without losing any data --\n- Data removal in RL may massively hinder exploration, especially in non-linear environments, speaking of which --\n- the paper only evaluate on classic mujoco, and lacks variety of evaluation task, speaking of evaluation --\n- The evaluation also lacks other sota methods for comparison, since the promise was to have better performance, rather than better theoretical understanding, which goes back to my first point.", "questions": "How does sample rejection based on gradient relate to hessian?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731002401685}, {"id": "SnrRJQAcYj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission874/Reviewer_cTe3"], "rating": 3, "soundness": 1, "presentation": 1, "contribution": 2, "confidence": 3, "summary": "This paper proposes the PIToD method to efficiently estimate which experiences positively or negatively impact RL learning. The proposed approach sets a drop-out mask for each experience and estimates the influence of each experience based on this mask and its complement, allowing for significantly more efficient computation compared to traditional Leave-One-Out (LOO) methods. The authors demonstrate that their method accurately estimates the influence of experiences using various metrics and further show, through experiments, that applying it to SAC improves learning performance.", "review_text": "This paper proposes the PIToD method to efficiently estimate which experiences positively or negatively impact RL learning. The proposed approach sets a drop-out mask for each experience and estimates the influence of each experience based on this mask and its complement, allowing for significantly more efficient computation compared to traditional Leave-One-Out (LOO) methods. The authors demonstrate that their method accurately estimates the influence of experiences using various metrics and further show, through experiments, that applying it to SAC improves learning performance.", "strengths": "1. A novel approach is presented for estimating the influence of specific experiences in RL by using Turn-over Dropout (ToD).\n2. It is theoretically demonstrated that the complement mask $w\\_i$ for experience $e\\_i$ indicates an absence of influence from $e\\_i$.", "weaknesses": "1. The metric used to show self-influence appears inappropriate. As noted in L246,  $Q\\_{w\\_i}$ is in a state where it has not been trained on $e\\_i$, so $L\\_{pe,i}(Q\\_{w\\_i}) − L\\_{pe,i}(Q\\_{m\\_i})$ is expected to be greater than zero in most cases, regardless of whether $e\\_i$ is beneficial for learning. Additionally, since $\\pi\\_{m\\_i} = \\arg\\max\\_{\\pi} L\\_{pi,i}(\\pi)$, $L\\_{pi,i}(\\pi\\_{w\\_i}) − L\\_{pi,i}(\\pi\\_{m\\_i})$ is likely always less than zero. Thus, these metrics do not directly indicate whether the experience has a positive or negative influence on RL learning.\n2. The evaluation shown in Figure 6 also seems flawed. In Algorithm 4 of Appendix D,  $w^*$ is already defined as  $\\arg\\max\\_w L\\_{ret}(\\pi, w)$ , so it is unsurprising that high returns are achieved.\n3. The main paper contains too few experimental results. While it seems that several experiments were conducted in the appendix, summarizing the purpose and outcomes of these experiments in the main paper would enhance clarity.\n4. It would be beneficial to include a direct comparison of mean performance between the original SAC method and the SAC method with PIToD in the main paper.", "questions": "1. **[About Weakness 1]** You mentioned that the PI and PE metrics indicate whether a specific experience has a positive or negative effect. Could you clarify this further? The current metrics seem to only reflect whether or not learning utilized $e_i$.\n2. **[About Weakness 2]** Is this approach fundamentally different from simply creating a policy ensemble through dropout and selecting the best-performing one? If the truly optimal  $w^*$  is selected, does it necessarily mean that the experience has a negative impact?\n3. **[About Weakness 3]** Could you summarize the experiments in the appendix and explain what each aims to demonstrate? It would be more beneficial if these results were integrated into the main paper.\n4. **[About Weakness 4]** Could you also present the experiments mentioned above?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes the PIToD method to efficiently estimate which experiences positively or negatively impact RL learning. The proposed approach sets a drop-out mask for each experience and estimates the influence of each experience based on this mask and its complement, allowing for significantly more efficient computation compared to traditional Leave-One-Out (LOO) methods. The authors demonstrate that their method accurately estimates the influence of experiences using various metrics and further show, through experiments, that applying it to SAC improves learning performance.", "soundness": 1, "presentation": 1, "contribution": 2, "strengths": "1. A novel approach is presented for estimating the influence of specific experiences in RL by using Turn-over Dropout (ToD).\n2. It is theoretically demonstrated that the complement mask $w\\_i$ for experience $e\\_i$ indicates an absence of influence from $e\\_i$.", "weaknesses": "1. The metric used to show self-influence appears inappropriate. As noted in L246,  $Q\\_{w\\_i}$ is in a state where it has not been trained on $e\\_i$, so $L\\_{pe,i}(Q\\_{w\\_i}) − L\\_{pe,i}(Q\\_{m\\_i})$ is expected to be greater than zero in most cases, regardless of whether $e\\_i$ is beneficial for learning. Additionally, since $\\pi\\_{m\\_i} = \\arg\\max\\_{\\pi} L\\_{pi,i}(\\pi)$, $L\\_{pi,i}(\\pi\\_{w\\_i}) − L\\_{pi,i}(\\pi\\_{m\\_i})$ is likely always less than zero. Thus, these metrics do not directly indicate whether the experience has a positive or negative influence on RL learning.\n2. The evaluation shown in Figure 6 also seems flawed. In Algorithm 4 of Appendix D,  $w^*$ is already defined as  $\\arg\\max\\_w L\\_{ret}(\\pi, w)$ , so it is unsurprising that high returns are achieved.\n3. The main paper contains too few experimental results. While it seems that several experiments were conducted in the appendix, summarizing the purpose and outcomes of these experiments in the main paper would enhance clarity.\n4. It would be beneficial to include a direct comparison of mean performance between the original SAC method and the SAC method with PIToD in the main paper.", "questions": "1. **[About Weakness 1]** You mentioned that the PI and PE metrics indicate whether a specific experience has a positive or negative effect. Could you clarify this further? The current metrics seem to only reflect whether or not learning utilized $e_i$.\n2. **[About Weakness 2]** Is this approach fundamentally different from simply creating a policy ensemble through dropout and selecting the best-performing one? If the truly optimal  $w^*$  is selected, does it necessarily mean that the experience has a negative impact?\n3. **[About Weakness 3]** Could you summarize the experiments in the appendix and explain what each aims to demonstrate? It would be more beneficial if these results were integrated into the main paper.\n4. **[About Weakness 4]** Could you also present the experiments mentioned above?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethics review needed.", "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730455242752}, {"id": "xCft7aSoLt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission874/Reviewer_YodU"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This paper presents Policy Iteration with Turn-over Dropout as a method for estimating the affect a state-action-reward-next state experience has on a policy or Q function in the area of off-policy RL. The method employs masks to dropout parameters, so that the affect of not training on an experience can be estimated without having to retrain the policy or Q values from scratch. The paper then investigates estimating various quantities including td-error and episode return.", "review_text": "This paper presents Policy Iteration with Turn-over Dropout as a method for estimating the affect a state-action-reward-next state experience has on a policy or Q function in the area of off-policy RL. The method employs masks to dropout parameters, so that the affect of not training on an experience can be estimated without having to retrain the policy or Q values from scratch. The paper then investigates estimating various quantities including td-error and episode return.", "strengths": "Being able to efficiently estimate the affect a particular data point has on a network used to estimate $\\pi$ or $Q$ is a very powerful tool, and to my knowledge this machinery has not been applied to a Deep RL setting before. I like that the authors have aimed to evaluate it for a variety of different purposes. \nHowever, it seems that only Section G in the appendix, and a short paragraph at the end of section 6 actually attempt to answer the question in the title of the paper.", "weaknesses": "# Major\n\nMore explanation in Section 4 of the PIToD method would be very helpful for a reader's understanding. \n\"Thus, some readers may suspect that the parameters dropped out by $m_i$ (i.e., the parameters obtained by applying $w_i$) are not influenced by $e_i$.\" - This sentence caused a lot of confusion to me when reading the paper. The phrasing seems to suggest that the parameters dropped out are indeed affected, but this doesn't seem to be the case (as the reader would suspect). I am also struggling to understand Appendix Section A. Assumption 1 seems extremely strong and appears to be *almost* exactly defining the property you are looking to prove. Looking at the equality, doesn't it mean that the gradient for $i'$ is 0 for everything but $i'$ due to the indicator?\n\nSection C in the appendix contains a lot of interesting content.\nThe findings from the Group Mask preliminary experiments in the appendix seem quite important - \"In our preliminary experiments, we found that the influence of a single experience on performance was negligibly small\". This should be mentioned at the very least in the main paper. \n\"Key implementation decisions to improve learning.\"  is also very important. The implementation details, and the architectural experiments you conducted should not be relegated to the appendix in this manner. Especially since Figure 10 shows they are critical to your method working.\nAdditionally, why are you using an ensemble of 20 MLPs in the architecture in this manner? Is it important to performance? Does it interact positively with your method? More discussion on why these design decisions were made is needed, or at the very least a comment stating it was the first architectural starting point used.\n\nMore explanation is needed on Section 5.1 to establish the importance of the ratios being considered. What is the importance/significance of the differences in the ratios between policy eval/improvement? For Eqn 8 I can understand that the td error if higher for an experience you haven't trained on, but for Eqn 9 I do not quite see why the action chosen by $\\pi_{\\theta, w_i}$ should have a lower estimated Q.\n\nSection 5.2 is very misleading since you're estimating the time it would take LOO. Given that important sections that have been pushed into the appendix, I would advise replacing these with other more relevant/concrete content.  \n\n\"In our setup, L_ret is estimated using Monte Carlo returns collected by rolling out policies\" - How many rollouts are used for each estimation?\nAdditionally, how do you utilise Eq. 10 to identify experiences? Are you rolling out the agent multiple times to collect MC returns? Is this factored into your training budget and reflected in Figure 6?\n\nI don't understand why you're not showing Figure 11 in place of Figure 6? What is the particular thing you want to highlight by showing a specially picked subset of the lines in Figure 6?\n\nMuch more analysis and results need to be presented on characterising which experiences are harmful (or beneficial). The paper begins this kind of investigation but it feels like an afterthought. To me, this is one of the most exciting parts of the paper, utilising the tools outlined to identify what experiences are harmful for performance (with potential links to 'Ray Interference: a Source of Plateaus in Deep Reinforcement Learning'). This would be of much interest to the community.\n\n# Minor\n\n- No references for RL, and an MDP is never mentioned in the paper.\n- No reference for the LOO estimator.\n- CQL cited under off-policy RL in the introduction feels unnecessary.\n- \"In the previous section, we demonstrated that PIToD can accurately and efficiently estimate the influence of experiences.\" - This is too broad a claim. In section 5 you estimated the influence an experience can have on self-influence.", "questions": "- For section G, Figure 16 shows very little change in the results when removing adversarial experiences, why is this? Figure 17 shows big differences in the estimations before and after amendments, but it would be good to clarify that your method can identify the adversarial experiences explicitly (as opposed to showing the affect of removing identified experiences which indirectly provides some evidence for this).\n\n- Figure 10 shows huge changes in the results across some architectural choices, please comment more on these. \n\n(There are also some questions sprinkled throughout the above section)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents Policy Iteration with Turn-over Dropout as a method for estimating the affect a state-action-reward-next state experience has on a policy or Q function in the area of off-policy RL. The method employs masks to dropout parameters, so that the affect of not training on an experience can be estimated without having to retrain the policy or Q values from scratch. The paper then investigates estimating various quantities including td-error and episode return.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Being able to efficiently estimate the affect a particular data point has on a network used to estimate $\\pi$ or $Q$ is a very powerful tool, and to my knowledge this machinery has not been applied to a Deep RL setting before. I like that the authors have aimed to evaluate it for a variety of different purposes. \nHowever, it seems that only Section G in the appendix, and a short paragraph at the end of section 6 actually attempt to answer the question in the title of the paper.", "weaknesses": "# Major\n\nMore explanation in Section 4 of the PIToD method would be very helpful for a reader's understanding. \n\"Thus, some readers may suspect that the parameters dropped out by $m_i$ (i.e., the parameters obtained by applying $w_i$) are not influenced by $e_i$.\" - This sentence caused a lot of confusion to me when reading the paper. The phrasing seems to suggest that the parameters dropped out are indeed affected, but this doesn't seem to be the case (as the reader would suspect). I am also struggling to understand Appendix Section A. Assumption 1 seems extremely strong and appears to be *almost* exactly defining the property you are looking to prove. Looking at the equality, doesn't it mean that the gradient for $i'$ is 0 for everything but $i'$ due to the indicator?\n\nSection C in the appendix contains a lot of interesting content.\nThe findings from the Group Mask preliminary experiments in the appendix seem quite important - \"In our preliminary experiments, we found that the influence of a single experience on performance was negligibly small\". This should be mentioned at the very least in the main paper. \n\"Key implementation decisions to improve learning.\"  is also very important. The implementation details, and the architectural experiments you conducted should not be relegated to the appendix in this manner. Especially since Figure 10 shows they are critical to your method working.\nAdditionally, why are you using an ensemble of 20 MLPs in the architecture in this manner? Is it important to performance? Does it interact positively with your method? More discussion on why these design decisions were made is needed, or at the very least a comment stating it was the first architectural starting point used.\n\nMore explanation is needed on Section 5.1 to establish the importance of the ratios being considered. What is the importance/significance of the differences in the ratios between policy eval/improvement? For Eqn 8 I can understand that the td error if higher for an experience you haven't trained on, but for Eqn 9 I do not quite see why the action chosen by $\\pi_{\\theta, w_i}$ should have a lower estimated Q.\n\nSection 5.2 is very misleading since you're estimating the time it would take LOO. Given that important sections that have been pushed into the appendix, I would advise replacing these with other more relevant/concrete content.  \n\n\"In our setup, L_ret is estimated using Monte Carlo returns collected by rolling out policies\" - How many rollouts are used for each estimation?\nAdditionally, how do you utilise Eq. 10 to identify experiences? Are you rolling out the agent multiple times to collect MC returns? Is this factored into your training budget and reflected in Figure 6?\n\nI don't understand why you're not showing Figure 11 in place of Figure 6? What is the particular thing you want to highlight by showing a specially picked subset of the lines in Figure 6?\n\nMuch more analysis and results need to be presented on characterising which experiences are harmful (or beneficial). The paper begins this kind of investigation but it feels like an afterthought. To me, this is one of the most exciting parts of the paper, utilising the tools outlined to identify what experiences are harmful for performance (with potential links to 'Ray Interference: a Source of Plateaus in Deep Reinforcement Learning'). This would be of much interest to the community.\n\n# Minor\n\n- No references for RL, and an MDP is never mentioned in the paper.\n- No reference for the LOO estimator.\n- CQL cited under off-policy RL in the introduction feels unnecessary.\n- \"In the previous section, we demonstrated that PIToD can accurately and efficiently estimate the influence of experiences.\" - This is too broad a claim. In section 5 you estimated the influence an experience can have on self-influence.", "questions": "- For section G, Figure 16 shows very little change in the results when removing adversarial experiences, why is this? Figure 17 shows big differences in the estimations before and after amendments, but it would be good to clarify that your method can identify the adversarial experiences explicitly (as opposed to showing the affect of removing identified experiences which indirectly provides some evidence for this).\n\n- Figure 10 shows huge changes in the results across some architectural choices, please comment more on these. \n\n(There are also some questions sprinkled throughout the above section)", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730311687971}, {"id": "ZrvB8Epobg", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission874/Reviewer_A595"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper describes a novel method called Policy Iteration with Turn-over Dropout (PIToD)  for excluding experiences that negatively affect the performance of RL agents when used for training via policy iteration. This includes a how to calculate the influence of a single experience on the agent's performance and how to amend the policy given this calculated influence. This is done efficiently through a parameter masking technique called turn-over dropout. The authors provide theoretical justification as to why this masking technique is similar to leaving out a specific experience. PIToD is tested on known four MuJoCo environments and shows improvement in performance for some of the environments while remaining computationally efficient.", "review_text": "The paper describes a novel method called Policy Iteration with Turn-over Dropout (PIToD)  for excluding experiences that negatively affect the performance of RL agents when used for training via policy iteration. This includes a how to calculate the influence of a single experience on the agent's performance and how to amend the policy given this calculated influence. This is done efficiently through a parameter masking technique called turn-over dropout. The authors provide theoretical justification as to why this masking technique is similar to leaving out a specific experience. PIToD is tested on known four MuJoCo environments and shows improvement in performance for some of the environments while remaining computationally efficient.", "strengths": "- The method is simple, novel, and original.\n- Comparison to the leave-one-out naive approach emphasizes the significance of PIToD.\n- Diagrams are clear and self-explanatory.\n- Results showcase the efficiency advantages of PIToD very well\n- the paper is generally well written with a clear narrative and a comfortable flow.", "weaknesses": "- the theoretical justification is lacking because the assumptions (1 and 2) don't seem realistic. Basically, the authors assume that the masks enforce some sort of leave-out rule. It makes sense since they try to minimize overlap (App B), but it is the assumptions are unjustified. The authors should provide empirical evidence supporting assumptions 1 and 2, and discuss the implications if these assumptions don't fully hold in practice.\n- missing analysis of results, mostly why the results in figures 3, 4 and 6 look the way they do. The authors should explain why the plots for some environments show lower ratios / performance than other, the reasons for instability in some cases (including shaded confidence intervals). Figure 4 is explained but this explanation is unclear. Specifically, it is unclear why this figure suggests that the agent is overfitting to older experiences.\n- figure 4 heatmap scales are all different, making it very difficult to read and compare the graphs. Either use consistent scales for all plots or at least keep colors consistent regardless of the scale, i.e., if 1 is yellow and -1 is blue inn one scale, then in another scale -1 is still blue, and -4 can be a new color, e.g., green.\n- does not consider what happens if the buffer is full, something that will eventually happen if training persists. The authors should provide a more intuitive explanation or a simple example that illustrates why the signs of these equations indicate correct influence calculations.\n- no comparison to other methods. E.g., PER is definitely comparable.\n- redundant mini-paragraphs (start of sections) and parentheses (introduction) that map out the content of the paper are distracting and ruin the flow of reading.", "questions": "- Could I use PIToD and PER together?\n- Your algorithm as many loops. Can these ToD iterations be efficiently batched?\n- it is unclear why equations 8 and 9 indicate correct influence calculations if they are positive and negative, respectively. Why is this the case?\n- why is the \"correct experience ratio\" of hopper for in policy improvement so much worse than the others, and why does the ratio for walker and ant decrease throughout the epochs? The authors should discuss potential reasons for these differences and what implications they might have for the applicability of PIToD across different environments.\n- why is bias mainly an issue with the humanoid environment and not the others? Is this consistent with previous findings in the literature? What specific characteristics does the humanoid environment have that might contribute to this bias issue?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper describes a novel method called Policy Iteration with Turn-over Dropout (PIToD)  for excluding experiences that negatively affect the performance of RL agents when used for training via policy iteration. This includes a how to calculate the influence of a single experience on the agent's performance and how to amend the policy given this calculated influence. This is done efficiently through a parameter masking technique called turn-over dropout. The authors provide theoretical justification as to why this masking technique is similar to leaving out a specific experience. PIToD is tested on known four MuJoCo environments and shows improvement in performance for some of the environments while remaining computationally efficient.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The method is simple, novel, and original.\n- Comparison to the leave-one-out naive approach emphasizes the significance of PIToD.\n- Diagrams are clear and self-explanatory.\n- Results showcase the efficiency advantages of PIToD very well\n- the paper is generally well written with a clear narrative and a comfortable flow.", "weaknesses": "- the theoretical justification is lacking because the assumptions (1 and 2) don't seem realistic. Basically, the authors assume that the masks enforce some sort of leave-out rule. It makes sense since they try to minimize overlap (App B), but it is the assumptions are unjustified. The authors should provide empirical evidence supporting assumptions 1 and 2, and discuss the implications if these assumptions don't fully hold in practice.\n- missing analysis of results, mostly why the results in figures 3, 4 and 6 look the way they do. The authors should explain why the plots for some environments show lower ratios / performance than other, the reasons for instability in some cases (including shaded confidence intervals). Figure 4 is explained but this explanation is unclear. Specifically, it is unclear why this figure suggests that the agent is overfitting to older experiences.\n- figure 4 heatmap scales are all different, making it very difficult to read and compare the graphs. Either use consistent scales for all plots or at least keep colors consistent regardless of the scale, i.e., if 1 is yellow and -1 is blue inn one scale, then in another scale -1 is still blue, and -4 can be a new color, e.g., green.\n- does not consider what happens if the buffer is full, something that will eventually happen if training persists. The authors should provide a more intuitive explanation or a simple example that illustrates why the signs of these equations indicate correct influence calculations.\n- no comparison to other methods. E.g., PER is definitely comparable.\n- redundant mini-paragraphs (start of sections) and parentheses (introduction) that map out the content of the paper are distracting and ruin the flow of reading.", "questions": "- Could I use PIToD and PER together?\n- Your algorithm as many loops. Can these ToD iterations be efficiently batched?\n- it is unclear why equations 8 and 9 indicate correct influence calculations if they are positive and negative, respectively. Why is this the case?\n- why is the \"correct experience ratio\" of hopper for in policy improvement so much worse than the others, and why does the ratio for walker and ant decrease throughout the epochs? The authors should discuss potential reasons for these differences and what implications they might have for the applicability of PIToD across different environments.\n- why is bias mainly an issue with the humanoid environment and not the others? Is this consistent with previous findings in the literature? What specific characteristics does the humanoid environment have that might contribute to this bias issue?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730258232568}], "openreview_url": "https://openreview.net/forum?id=EWNH3QTSxd", "arxiv_id": "2405.14629", "paper_pdf": "papers/EWNH3QTSxd.pdf", "paper_pdf_sha256": "0793ef11fd5ef27c26c488eed081c286bba735152ac3a53263d215c431a8be7e", "paper_pdf_bytes": 6434349, "paper_pdf_source": "openreview", "code_url": "https://github.com/TakuyaHiraoka/Which-Experiences-Are-Influential-for-RL-Agents", "code_repository": "TakuyaHiraoka/Which-Experiences-Are-Influential-for-RL-Agents", "code_commit": "085d8fd2791a50e1aafc2309c35ac05c7863b7be", "code_archive": "repos/EWNH3QTSxd.zip", "code_archive_sha256": "91afbba8af4b0ba18a5f3a6af3e72f0775b7137cb8b683c8add7c70863f91d7a", "code_archive_bytes": 126149, "code_file_count": 18, "code_extensions": {".py": 15, ".sh": 3}, "github_disk_usage_kb": 400, "github_languages": {"Python": 142706, "Shell": 3381}, "github_archived": false, "github_pushed_at": "2025-08-22T19:41:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/which-experiences-are-influential-for-rl"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "EPfGHb9Y68", "year": 2024, "status": "rejected", "title": "Continual Offline Reinforcement Learning via Diffusion-based Dual Generative Replay", "authors": ["Jinmei Liu", "Wenbin Li", "Xiangyu Yue", "Chunlin Chen", "Zhi Wang"], "authorids": ["~Jinmei_Liu2", "~Wenbin_Li5", "~Xiangyu_Yue1", "~Chunlin_Chen1", "~Zhi_Wang7"], "authors_source": "OpenReview API", "abstract": "We study continual offline reinforcement learning, a practical paradigm that facilitates forward transfer and mitigates catastrophic forgetting to tackle sequential offline tasks. We propose a dual generative replay framework that retains previous knowledge by concurrent replay of generated pseudo-data. First, we decouple the continual learning policy into a diffusion-based generative behavior model and a multi-head action evaluation model, allowing the policy to inherit distributional expressivity for encompassing a progressive range of diverse behaviors. Second, we train a task-conditioned diffusion model to mimic state distributions of past tasks. Generated states are paired with corresponding responses from the behavior generator to represent old tasks with high-fidelity replayed samples. Finally, by interleaving pseudo samples with real ones of the new task, we continually update the state and behavior generators to model progressively diverse behaviors, and regularize the multi-head critic in a behavior cloning manner to mitigate forgetting. Experiments on various benchmarks demonstrate that our method achieves better forward transfer with less forgetting, and closely approximates results of using previous ground-truth data due to its high-fidelity replay of the sample space.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "eKGZKiv9Tj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1109/Reviewer_msqf"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Motivated by the limitations of unimodal Gaussian policy models and the memory constraints of storing data from previous tasks, the authors propose a novel dual generator system to facilitate continual learning in reinforcement learning. This system features a behavior generative model that diffuses over actions given states, and a state generative model that diffuses over states from past tasks without needing to store all previous data. When encountering a new task, this dual approach leverages the state generator to generate synthetic state samples reflective of all former tasks, and the behavior generator produces corresponding actions, forming pseudo state-action pairs. \nThan, a multi-head critic network, with separate heads dedicated to individual tasks, is trained on real samples from new datasets and annoate the pseudo pairs to create synthetic samples for behavior cloning.", "review_text": "Motivated by the limitations of unimodal Gaussian policy models and the memory constraints of storing data from previous tasks, the authors propose a novel dual generator system to facilitate continual learning in reinforcement learning. This system features a behavior generative model that diffuses over actions given states, and a state generative model that diffuses over states from past tasks without needing to store all previous data. When encountering a new task, this dual approach leverages the state generator to generate synthetic state samples reflective of all former tasks, and the behavior generator produces corresponding actions, forming pseudo state-action pairs. \nThan, a multi-head critic network, with separate heads dedicated to individual tasks, is trained on real samples from new datasets and annoate the pseudo pairs to create synthetic samples for behavior cloning.", "strengths": "Pros:\n1. First attempt to incorporate diffusion model for continual offline RL. Novel idea to utilizing diffusion-model’s expressiveness to generate high-fidelity replay of the previous tasks to prevent the need of storing all previous tasks samples.\n2. Shows effectiveness in generating new samples to represent prior tasks when comparing CuGRO with the Oracle in Table 1 and figure 2. \n3. Achieves strong experimental results across the 4 simulated environments. The proposed CuGRO algorithm closely matches the Oracle, outperforming baselines.\n4. Ablation studies conducted to analyze the hyper parameters lambda which controls how much emphasis to put on the previously replayed dataset.", "weaknesses": "Cons:\n1. Requires training two separate diffusion models, which can be computationally expensive for sampling at test time since parallel sampling is not possible. Exploring concatenating {s,a} and diffuse with one model could improve efficiency.\n2. This methods alleviates the memory capacity concern by condensing the previous task’s knowledge into two diffusion models. How does the continual training cost of updating diffusion models for each new task and sampling from them trade off with the memory savings of condensing previous tasks?\n3. Limited baselines: Is comparison only made within the diffusion-based model generator pipeline? Why there is no comparison to previous continual RL algorithms provided?", "questions": "please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Motivated by the limitations of unimodal Gaussian policy models and the memory constraints of storing data from previous tasks, the authors propose a novel dual generator system to facilitate continual learning in reinforcement learning. This system features a behavior generative model that diffuses over actions given states, and a state generative model that diffuses over states from past tasks without needing to store all previous data. When encountering a new task, this dual approach leverages the state generator to generate synthetic state samples reflective of all former tasks, and the behavior generator produces corresponding actions, forming pseudo state-action pairs. \nThan, a multi-head critic network, with separate heads dedicated to individual tasks, is trained on real samples from new datasets and annoate the pseudo pairs to create synthetic samples for behavior cloning.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "Pros:\n1. First attempt to incorporate diffusion model for continual offline RL. Novel idea to utilizing diffusion-model’s expressiveness to generate high-fidelity replay of the previous tasks to prevent the need of storing all previous tasks samples.\n2. Shows effectiveness in generating new samples to represent prior tasks when comparing CuGRO with the Oracle in Table 1 and figure 2. \n3. Achieves strong experimental results across the 4 simulated environments. The proposed CuGRO algorithm closely matches the Oracle, outperforming baselines.\n4. Ablation studies conducted to analyze the hyper parameters lambda which controls how much emphasis to put on the previously replayed dataset.", "weaknesses": "Cons:\n1. Requires training two separate diffusion models, which can be computationally expensive for sampling at test time since parallel sampling is not possible. Exploring concatenating {s,a} and diffuse with one model could improve efficiency.\n2. This methods alleviates the memory capacity concern by condensing the previous task’s knowledge into two diffusion models. How does the continual training cost of updating diffusion models for each new task and sampling from them trade off with the memory savings of condensing previous tasks?\n3. Limited baselines: Is comparison only made within the diffusion-based model generator pipeline? Why there is no comparison to previous continual RL algorithms provided?", "questions": "please see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698910484981}, {"id": "criohIzI65", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1109/Reviewer_RUFG"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper addresses the continual offline reinforcement learning (CORL) problem, focusing on the challenge of catastrophic forgetting as models encounter new tasks. To combat this, the study introduces CuGRO, a method that decouples the learning policy into a generative behavior model and an action evaluation model, ensuring diverse behaviors are captured. A state generative model is also employed to mimic past task distributions. By leveraging diffusion probabilistic models, CuGRO achieves high-fidelity sample reproduction. Empirical tests reveal CuGRO's superiority in reducing forgetting and enhancing forward transfer, closely matching results using original data.", "review_text": "The paper addresses the continual offline reinforcement learning (CORL) problem, focusing on the challenge of catastrophic forgetting as models encounter new tasks. To combat this, the study introduces CuGRO, a method that decouples the learning policy into a generative behavior model and an action evaluation model, ensuring diverse behaviors are captured. A state generative model is also employed to mimic past task distributions. By leveraging diffusion probabilistic models, CuGRO achieves high-fidelity sample reproduction. Empirical tests reveal CuGRO's superiority in reducing forgetting and enhancing forward transfer, closely matching results using original data.", "strengths": "1. Innovative idea: The paper introduces CuGRO, a novel framework of CORL, which is one of the first to leverage expressive diffusion models for this challenge. \n2. Practical memory solutions: Instead of relying on large buffers to store real samples from prior tasks, CuGRO synthesizes high-quality pseudo-samples, addressing the challenges of memory constraint and potential privacy issues. \n3. Empirical validation: The paper provides empirical evidence from various tasks that demonstrates CuGRO's effectiveness in mitigating forgetting.", "weaknesses": "1. Motivation. There is no sufficient support for the necessity of using diffusion models to learn the behaviors from prior tasks. Though the general knowledge is that diffusion models can work better in terms of generation and generalization, there is no explicit reason against the utilization of other modes such as behavior cloning, GAN, VAE, etc. \n2. Efficiency trade-off: Though storing models for prior tasks might work well, training diffusion models for each different task might not be sample-efficient nor computation-efficient. Therefore, the authors might want to provide more information on the feasibility of this approach and the computation resource usage for implementing the experiments. \n3. Experiment: The experiment did not show how diffusion models contribute to the continual learning of the model. The authors might want to show the performance of the diffusion models to demonstrate that diffusion models are contributing to the performance. In addition, it is unclear in the paper regarding the training data of the diffusion models. Will noisy data degrade the performance of the diffusion models and the CuGRO model as a whole?", "questions": "1. I was wondering if applying other types of generative models to replace the diffusion model will yield similar performance. \n2. I am curious to know the computation resources used and how long to train the model.\n3. I was wondering about the scalability of the model. If having more tasks degrade the model's performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the continual offline reinforcement learning (CORL) problem, focusing on the challenge of catastrophic forgetting as models encounter new tasks. To combat this, the study introduces CuGRO, a method that decouples the learning policy into a generative behavior model and an action evaluation model, ensuring diverse behaviors are captured. A state generative model is also employed to mimic past task distributions. By leveraging diffusion probabilistic models, CuGRO achieves high-fidelity sample reproduction. Empirical tests reveal CuGRO's superiority in reducing forgetting and enhancing forward transfer, closely matching results using original data.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. Innovative idea: The paper introduces CuGRO, a novel framework of CORL, which is one of the first to leverage expressive diffusion models for this challenge. \n2. Practical memory solutions: Instead of relying on large buffers to store real samples from prior tasks, CuGRO synthesizes high-quality pseudo-samples, addressing the challenges of memory constraint and potential privacy issues. \n3. Empirical validation: The paper provides empirical evidence from various tasks that demonstrates CuGRO's effectiveness in mitigating forgetting.", "weaknesses": "1. Motivation. There is no sufficient support for the necessity of using diffusion models to learn the behaviors from prior tasks. Though the general knowledge is that diffusion models can work better in terms of generation and generalization, there is no explicit reason against the utilization of other modes such as behavior cloning, GAN, VAE, etc. \n2. Efficiency trade-off: Though storing models for prior tasks might work well, training diffusion models for each different task might not be sample-efficient nor computation-efficient. Therefore, the authors might want to provide more information on the feasibility of this approach and the computation resource usage for implementing the experiments. \n3. Experiment: The experiment did not show how diffusion models contribute to the continual learning of the model. The authors might want to show the performance of the diffusion models to demonstrate that diffusion models are contributing to the performance. In addition, it is unclear in the paper regarding the training data of the diffusion models. Will noisy data degrade the performance of the diffusion models and the CuGRO model as a whole?", "questions": "1. I was wondering if applying other types of generative models to replace the diffusion model will yield similar performance. \n2. I am curious to know the computation resources used and how long to train the model.\n3. I was wondering about the scalability of the model. If having more tasks degrade the model's performance?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698708021719}, {"id": "6WPyKmN0T6", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1109/Reviewer_t1vB"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors in the paper propose a dual generative reply framework to address the challenges in continual RL, where the practical algorithms are required to adapt to new environments and simultaneously leverage the previous knowledge. In particular, they use two diffusion models to generate both the state and behavior generative reply on the offline data and update in a sequential way. Using the behavior cloning technique, the multi-head critic is updated effectively, and the resulting algorithm CuGRO suggests competitive performance in the considered Mujoco benchmark.", "review_text": "The authors in the paper propose a dual generative reply framework to address the challenges in continual RL, where the practical algorithms are required to adapt to new environments and simultaneously leverage the previous knowledge. In particular, they use two diffusion models to generate both the state and behavior generative reply on the offline data and update in a sequential way. Using the behavior cloning technique, the multi-head critic is updated effectively, and the resulting algorithm CuGRO suggests competitive performance in the considered Mujoco benchmark.", "strengths": "* The paper is well-organized and easy to follow.\n* The proposed method is technically sound, including the incorporation of diffusion models to mimic the generative buffer reply.\n* The empirical performance seems significant compared with considered baselines.", "weaknesses": "* Incorporating diffusion models in offline RL for generative reply is straightforward, and the computation cost should be rigorously discussed.\n* The loss function in Eq.10, 11, 13, 14, and 15 is easy to figure out, but it is hard to guarantee the convergence. For instance, in Eq.10, it is not clear whether training a diffusion model via using the data from the last diffusion model has any convergence guarantee. This issue could be severe especially when we only have access to offline data with a limited size or large state and action space.\n* Missing other continual RL baselines or benchmarks. In continual RL, the benchmark ‘continual world’ is commonly used to evaluate the continual control algorithms, but the proposal algorithm is only evaluated in small-scaled benchmarks in Mujuco games. It is not clear whether this approach is effective in the commonly-used continual RL environments. Also, some typical continual RL baselines are missing, such as [1].\n\n[1] Gaya et al. Building a subspace of policies for scalable continual learning (ICLR 2023)", "questions": "Please refer to Weakness.\n\nOverall, as far as I can tell, the proposed method is technically sound and achieves competitive performance. However, the computation cost would be large, and it also lacks discussion about the convergence guarantee and training details. Some typical benchmarks and baselines should be considered as well.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors in the paper propose a dual generative reply framework to address the challenges in continual RL, where the practical algorithms are required to adapt to new environments and simultaneously leverage the previous knowledge. In particular, they use two diffusion models to generate both the state and behavior generative reply on the offline data and update in a sequential way. Using the behavior cloning technique, the multi-head critic is updated effectively, and the resulting algorithm CuGRO suggests competitive performance in the considered Mujoco benchmark.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "* The paper is well-organized and easy to follow.\n* The proposed method is technically sound, including the incorporation of diffusion models to mimic the generative buffer reply.\n* The empirical performance seems significant compared with considered baselines.", "weaknesses": "* Incorporating diffusion models in offline RL for generative reply is straightforward, and the computation cost should be rigorously discussed.\n* The loss function in Eq.10, 11, 13, 14, and 15 is easy to figure out, but it is hard to guarantee the convergence. For instance, in Eq.10, it is not clear whether training a diffusion model via using the data from the last diffusion model has any convergence guarantee. This issue could be severe especially when we only have access to offline data with a limited size or large state and action space.\n* Missing other continual RL baselines or benchmarks. In continual RL, the benchmark ‘continual world’ is commonly used to evaluate the continual control algorithms, but the proposal algorithm is only evaluated in small-scaled benchmarks in Mujuco games. It is not clear whether this approach is effective in the commonly-used continual RL environments. Also, some typical continual RL baselines are missing, such as [1].\n\n[1] Gaya et al. Building a subspace of policies for scalable continual learning (ICLR 2023)", "questions": "Please refer to Weakness.\n\nOverall, as far as I can tell, the proposed method is technically sound and achieves competitive performance. However, the computation cost would be large, and it also lacks discussion about the convergence guarantee and training details. Some typical benchmarks and baselines should be considered as well.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698635796931}, {"id": "G60KV4BZMF", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission1109/Reviewer_R6pv"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "2 fair", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The submission presents a new method for continual offline RL (CORL) that relies on three components: a Q-function to assess the quality of each action, with multiple heads to account for the multiple tasks the agent faces; a diffusion behavior model, to both generate actions to execute in the environment and generate replay data for past tasks to avoid forgetting; and a state diffusion model to generate states to match the distribution observed in the dataset for each past task, again to avoid forgetting. In a strict setting where the agent is not allowed to observe data for any previous task, this enables continual training without forgetting. The authors evaluate their method on four sequences of 4 MuJoCo tasks each, with varying dynamics.", "review_text": "The submission presents a new method for continual offline RL (CORL) that relies on three components: a Q-function to assess the quality of each action, with multiple heads to account for the multiple tasks the agent faces; a diffusion behavior model, to both generate actions to execute in the environment and generate replay data for past tasks to avoid forgetting; and a state diffusion model to generate states to match the distribution observed in the dataset for each past task, again to avoid forgetting. In a strict setting where the agent is not allowed to observe data for any previous task, this enables continual training without forgetting. The authors evaluate their method on four sequences of 4 MuJoCo tasks each, with varying dynamics.", "strengths": "######## Strengths ########\n\n- The use of diffusion models both for generating behaviors and replay data is promising\n- The overall algorithm carefully leverages the three components above to continually learn new tasks\n- The problem of CORL itself is understudied, so it's good to see contributions in this area", "weaknesses": "######## Weaknesses ########\n- Section 2 (preliminaries) is largely unclear and not stand-alone\n- Section 3 (approach) is also not sufficiently clear or precise, and seems to introduce both math and text that are not related to the submission\n- The experimental evaluation is not sufficient to assess the benefits of the proposed method\n \n######## Recommendation ########\n\nUnfortunately, I recommend that this manuscript is not accepted in its current form. The main concern I have is the limited experimental evaluation. The authors propose an approach that is a combination of existing ideas (which is fine!), but unfortunately there is not sufficient evidence to support the choice of this combination. Moreover, the draft (especially sections 2 and 3) would need to undergo major revisions to make the text clear and (most importantly) precise.\n\n######## Arguments ########\n- Experimental setting -- The place with the most room for improvement in this submission is the empirical evaluation. In particular, I have three main concerns:\n    - The evaluation uses very simple MuJoCo tasks to evaluate methods. The authors should evaluate their approach on much more complex tasks, such as those from the Meta-World [1], CausalWorld [2], RLBench [3], or CompoSuite [4,5]. This is particularly true because the authors' motivation stems from the need for diffusion models to represent complex behaviors.\n    - The continual setting uses sequences of only 4 tasks. Given that the curves in Figure 2 for CuGRO get progressively worse as more tasks are trained, it would be important to study how this detriment scales with the number of tasks. Can we expect CuGRO to handle a really long stream of tasks? We need empirical evidence to answer that question\n    - The evaluation considers no external baselines. There are a number of existing continual learning and offline RL methods (which the authors themselves cite in their manuscript). Yet the evaluation is limited to variations of the authors' proposed method (oracle, noise, and none). Would other continual learning methods be as effective? Would other offline RL methods work well in this setting (especially considering that the tasks are quite simple). \n- Clarity of preliminaries\n    - The description of advantage-weighted regression misses a key piece: the Q function is w.r.t. \\mu, which is the piece that makes Eq. 4 solve the _costrained_ optimization of Eq. 2. This should be explicitly stated and explained, as otherwise the reader might think Q_\\theta is the standard Q^\\pi --- this is what I thought initially when reading 2.1\n    - It's unclear if Eq. 5 corresponds to Eq. 3. I believe that it does, and if so, my understanding is that Q_\\theta in Eq. 5 is not the same as Q_\\theta in Eq. 4---the former is w.r.t. \\pi and the latter is w.r.t. \\mu. Please clarify.\n    - The description of the diffusion models in Sec 2.2 misses mentioning the existence of a forward diffusion process that adds noise to an (observed) action. This makes it difficult to parse the last sentence before Eq. 6 where the diffusion model is predicting some un-defined noise. \n    - The description prior to Eq. 7 is still unclear to me: is Q estimating the value of the actions w.r.t. \\mu or \\pi? The authors mention that the actions are sampled from \\mu, but is the long-term value measured for \\pi or \\mu? It seems that the middle part of Eq. 7 entails that it's from \\pi, but this isn't stated in text. \n    - What does re-sampling mean toward the end of Sec 2.1? What was the first \"sampling\"?\n \n- Clarity/precision of approach\n    - Sec 3.1\n        - The description of CORL is mostly clear, but I do have a couple of questions:\n            - What aspects of the MDP are allowed to change from task to task? If all of them, then which ones do the authors consider in their approach/evaluation?\n            - Is the distribution over MDPs P(M) stationary or is it allowed to change over time? Is there any implication of that for the learning process or evaluation setting in the experiments?\n    - Sec 3.2\n        - While I agree that diffusion models (or generally expressive generative models) are useful for expressing RL policies, the authors seem to conflate two things in their description of why that is the case: 1) some tasks require multimodal behaviors (like in footnote 1) and 2) the overall behavior expressed by the diffusion model should capture a breadth of tasks. The latter is hinted at in both the intro and here, but never actually explained or exemplified. \n        - The argument of footnote 2 is somewhat weak. What if two visual classification tasks are \"detect if dog is in image\" and \"detect if cat is in image\" and they're both given the same image of a dog? Tasks would require opposite predictions given the same image, just like the RL model would. Plus, the conclusion would be that diffusion models are better because they could capture both the opposite actions, but how is that useful if they can't differentiate when to execute each? The only way to solve a problem like this is to let the model take as input a task indicator (or something to differentiate the tasks), and it's unclear that a Gaussian conditioned on this information would fail. \n        - This section is very odd. The first paragraph is all motivation and no technical details. Then second paragraph contains some details about how the diffusion model for state generation works (the equivalent for actions was in the perliminaries). Then the final paragraph is all about doing classifier or classifier-free guidance, but it's unclear what for or why 6 lines of this show up in the middle of a technical section when the authors don't even actually try it. I guess the idea is that the classifier-based/free guidance could ensure that the generated states actually correspond to the conditioning task?\n    - Sec 3.3\n        - \"the desired importance of the new task...\" seems to suggest that Eq. 8 should be a weighted sum. But the authors more likely mean that there needs to be some weighting to ensure that the current task is learned sufficiently well while avoiding forgetting. It isn't really about the importance but about being able to optimize properly. \n        - It's unclear why there's a test loss (Eq. 11 and 14) for specific models. Isn't performance measured as the obtained reward of the agent on the tasks?\n        - It's also unclear what the authors mean by \"reconstruct the cumulative state space\". Doesn't the task conditioning imply that the agent is learning separate state spaces, one for each task?\n        - Are the generated states for behavior replay drawn after updating the state generator? Before? Or are they both updated together? How was this choice made and what are the implications of it? My intuition would be that it's better to first train the behavior model on the fixed state generator and then train the state generator, because updates to the behavior model have no effect on the state generator but the converse is not true.\n            - This seems to be clarified in Algorithm 1, but the authors should state it explicitly in text and not rely exclusively on the Appendix to transmit that point. \n    - Sec 3.4\n        - It's quite unclear after reading this section why the authors use the term behavior cloning, which has a very specific connotation in the context of off-line RL -- replicating the behavior that generates the data.\n        - Instead, I think this approach is better described as a form of _functional regularization_, which is a method broadly studied in supervised continual learning research.\n\n[1] Yu et al., \"Meta-World: A benchmark and evaluation for multi-task and meta reinforcement learning.\" CoRL, 2020\n\n[2] Ahmed et al., \"CausalWorld: A Robotic Manipulation Benchmark for Causal Structure and Transfer Learning.\" ICLR, 2021\n\n[3] James et al., \"Rlbench: The robot learning benchmark & learning environment.\" RAL, 2020\n\n[4] Mendez et al., \"CompoSuite: A Compositional reinforcement learning benchmark.\" CoLLAs, 2022\n\n[5] Hussing et al., \"Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning.\" arXiv, 2023", "questions": "######## Additional feedback ########\n\nThe following points are provided as feedback to hopefully help better shape the submitted manuscript, but did not impact my recommendation in a major way.\n\nAbstract\n- The abstract is very clear. It lays out very well how the approach works and the results they obtain \n\nIntro\n- It's unclear what \"new tasks emerge overwhelmingly\" means or why van de Ven et al. is cited to support that claim. \n- The motivation for why CORL is special seems to be all about RL in general, and not specifically about offline RL. This seems to undermine the need to develop specialized approaches.\n- There's always the question of whether the size of the model might surpass the size of replay buffers. In visual settings that tends to happen. It's unclear if it does here\n- The idea of using generative models to express state and action distributions (which isn't novel) is good, especially in the offline setting where we can't make assumptions about the form of the distribution\n- What does \"behavior cloning matter\" mean in the context of a critic, which is not a behavior model?\n\nSec 2\n- I appreciate the notation clarification right before Sec 3! The authors could consider moving it to the beginning of Sec. 2 so the reader knows this ahead of time. \n- I thought we were missing a description of CORL, but that's in Sec. 3. Is the formalization of CORL a contribution of this work? If not, maybe it's worth also including it in Sec 2.\n\nSec 3.1\n- It does seem like the problem setting should be moved to Sec 2, and then the overview be placed in Sec 3 before introducing Sec 3.2 (which would be 3.1) below\n- Fig. 1 is useful. Why are there two (s,a,r,s') boxes in b and c? It seems like they are the same tuple, just that the first down arrow takes s and the second down arrow takes a. \n\nSec 4\n- \"the generator is the only constraint on task performance\" -- how is this an improvement over other continual learning methods?\n- \"When the generative model is optimal, training the networks with generative replay is equivalent to joint training on the entire dataset.\" \n    - Sure, an optimal replay method would achieve that... but can we actually train an optimal generator over a long sequence of tasks? Also, it is only equivalent to joint training if we actually do full joint training (from scratch), but not if we start from the previously trained models. Starting from previously trained models might be better or worse, but certainly not equivalent.\n- What is the (final) performance of SAC/TD3 on the collected datasets?\n- What is the \"oracle\" for the behavior model? And what is the \"noisy\" replay for the behavior model?\n- More than baselines, these seem to be ablations of CuGRO. While oracle is roughly a performance upper bound, it's unclear how other existing continual learning algorithms would perform compared to CuGRO/oracle. It's also unclear if non-diffusion approaches (given \"oracle\" or some other forgetting avoidance method) would work well.\n- I like the analysis of why Hopper-Vel fails\n- In Figure 3 we don't get to see past task performance. Or is this average performance including current and past tasks? Does Table 1 measure the final performance of all past tasks after training on the final task, or upon finishing training on each individual task?\n- The hyperparamter sensitivity analysis is good and useful.\n\nTypos/style/grammar\n- Footnotes in 3.2 should go before periods, not after the period and a space (e.g., \"...keep emerging\\footnote{text}.\")\n- Sec 3.2, paragraph 1 -- do datasets really \"emerge\"? Maybe they are \"constructed\" instead\n- Sec 3.2, paragraph 2 -- scored-based --> score-based\n- Sec 3.3, first line -- technically, the models are \\mu_\\phi and \\epsilon_\\psi, and \\phi eand \\psi are the parameters\n- I find the use of \"replayed\" samples (throughout the text) a bit odd, since they aren't real samples. I'd suggest using \"generated\" samples instead to consistently clarify that these are not real.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The submission presents a new method for continual offline RL (CORL) that relies on three components: a Q-function to assess the quality of each action, with multiple heads to account for the multiple tasks the agent faces; a diffusion behavior model, to both generate actions to execute in the environment and generate replay data for past tasks to avoid forgetting; and a state diffusion model to generate states to match the distribution observed in the dataset for each past task, again to avoid forgetting. In a strict setting where the agent is not allowed to observe data for any previous task, this enables continual training without forgetting. The authors evaluate their method on four sequences of 4 MuJoCo tasks each, with varying dynamics.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "3 good", "strengths": "######## Strengths ########\n\n- The use of diffusion models both for generating behaviors and replay data is promising\n- The overall algorithm carefully leverages the three components above to continually learn new tasks\n- The problem of CORL itself is understudied, so it's good to see contributions in this area", "weaknesses": "######## Weaknesses ########\n- Section 2 (preliminaries) is largely unclear and not stand-alone\n- Section 3 (approach) is also not sufficiently clear or precise, and seems to introduce both math and text that are not related to the submission\n- The experimental evaluation is not sufficient to assess the benefits of the proposed method\n \n######## Recommendation ########\n\nUnfortunately, I recommend that this manuscript is not accepted in its current form. The main concern I have is the limited experimental evaluation. The authors propose an approach that is a combination of existing ideas (which is fine!), but unfortunately there is not sufficient evidence to support the choice of this combination. Moreover, the draft (especially sections 2 and 3) would need to undergo major revisions to make the text clear and (most importantly) precise.\n\n######## Arguments ########\n- Experimental setting -- The place with the most room for improvement in this submission is the empirical evaluation. In particular, I have three main concerns:\n    - The evaluation uses very simple MuJoCo tasks to evaluate methods. The authors should evaluate their approach on much more complex tasks, such as those from the Meta-World [1], CausalWorld [2], RLBench [3], or CompoSuite [4,5]. This is particularly true because the authors' motivation stems from the need for diffusion models to represent complex behaviors.\n    - The continual setting uses sequences of only 4 tasks. Given that the curves in Figure 2 for CuGRO get progressively worse as more tasks are trained, it would be important to study how this detriment scales with the number of tasks. Can we expect CuGRO to handle a really long stream of tasks? We need empirical evidence to answer that question\n    - The evaluation considers no external baselines. There are a number of existing continual learning and offline RL methods (which the authors themselves cite in their manuscript). Yet the evaluation is limited to variations of the authors' proposed method (oracle, noise, and none). Would other continual learning methods be as effective? Would other offline RL methods work well in this setting (especially considering that the tasks are quite simple). \n- Clarity of preliminaries\n    - The description of advantage-weighted regression misses a key piece: the Q function is w.r.t. \\mu, which is the piece that makes Eq. 4 solve the _costrained_ optimization of Eq. 2. This should be explicitly stated and explained, as otherwise the reader might think Q_\\theta is the standard Q^\\pi --- this is what I thought initially when reading 2.1\n    - It's unclear if Eq. 5 corresponds to Eq. 3. I believe that it does, and if so, my understanding is that Q_\\theta in Eq. 5 is not the same as Q_\\theta in Eq. 4---the former is w.r.t. \\pi and the latter is w.r.t. \\mu. Please clarify.\n    - The description of the diffusion models in Sec 2.2 misses mentioning the existence of a forward diffusion process that adds noise to an (observed) action. This makes it difficult to parse the last sentence before Eq. 6 where the diffusion model is predicting some un-defined noise. \n    - The description prior to Eq. 7 is still unclear to me: is Q estimating the value of the actions w.r.t. \\mu or \\pi? The authors mention that the actions are sampled from \\mu, but is the long-term value measured for \\pi or \\mu? It seems that the middle part of Eq. 7 entails that it's from \\pi, but this isn't stated in text. \n    - What does re-sampling mean toward the end of Sec 2.1? What was the first \"sampling\"?\n \n- Clarity/precision of approach\n    - Sec 3.1\n        - The description of CORL is mostly clear, but I do have a couple of questions:\n            - What aspects of the MDP are allowed to change from task to task? If all of them, then which ones do the authors consider in their approach/evaluation?\n            - Is the distribution over MDPs P(M) stationary or is it allowed to change over time? Is there any implication of that for the learning process or evaluation setting in the experiments?\n    - Sec 3.2\n        - While I agree that diffusion models (or generally expressive generative models) are useful for expressing RL policies, the authors seem to conflate two things in their description of why that is the case: 1) some tasks require multimodal behaviors (like in footnote 1) and 2) the overall behavior expressed by the diffusion model should capture a breadth of tasks. The latter is hinted at in both the intro and here, but never actually explained or exemplified. \n        - The argument of footnote 2 is somewhat weak. What if two visual classification tasks are \"detect if dog is in image\" and \"detect if cat is in image\" and they're both given the same image of a dog? Tasks would require opposite predictions given the same image, just like the RL model would. Plus, the conclusion would be that diffusion models are better because they could capture both the opposite actions, but how is that useful if they can't differentiate when to execute each? The only way to solve a problem like this is to let the model take as input a task indicator (or something to differentiate the tasks), and it's unclear that a Gaussian conditioned on this information would fail. \n        - This section is very odd. The first paragraph is all motivation and no technical details. Then second paragraph contains some details about how the diffusion model for state generation works (the equivalent for actions was in the perliminaries). Then the final paragraph is all about doing classifier or classifier-free guidance, but it's unclear what for or why 6 lines of this show up in the middle of a technical section when the authors don't even actually try it. I guess the idea is that the classifier-based/free guidance could ensure that the generated states actually correspond to the conditioning task?\n    - Sec 3.3\n        - \"the desired importance of the new task...\" seems to suggest that Eq. 8 should be a weighted sum. But the authors more likely mean that there needs to be some weighting to ensure that the current task is learned sufficiently well while avoiding forgetting. It isn't really about the importance but about being able to optimize properly. \n        - It's unclear why there's a test loss (Eq. 11 and 14) for specific models. Isn't performance measured as the obtained reward of the agent on the tasks?\n        - It's also unclear what the authors mean by \"reconstruct the cumulative state space\". Doesn't the task conditioning imply that the agent is learning separate state spaces, one for each task?\n        - Are the generated states for behavior replay drawn after updating the state generator? Before? Or are they both updated together? How was this choice made and what are the implications of it? My intuition would be that it's better to first train the behavior model on the fixed state generator and then train the state generator, because updates to the behavior model have no effect on the state generator but the converse is not true.\n            - This seems to be clarified in Algorithm 1, but the authors should state it explicitly in text and not rely exclusively on the Appendix to transmit that point. \n    - Sec 3.4\n        - It's quite unclear after reading this section why the authors use the term behavior cloning, which has a very specific connotation in the context of off-line RL -- replicating the behavior that generates the data.\n        - Instead, I think this approach is better described as a form of _functional regularization_, which is a method broadly studied in supervised continual learning research.\n\n[1] Yu et al., \"Meta-World: A benchmark and evaluation for multi-task and meta reinforcement learning.\" CoRL, 2020\n\n[2] Ahmed et al., \"CausalWorld: A Robotic Manipulation Benchmark for Causal Structure and Transfer Learning.\" ICLR, 2021\n\n[3] James et al., \"Rlbench: The robot learning benchmark & learning environment.\" RAL, 2020\n\n[4] Mendez et al., \"CompoSuite: A Compositional reinforcement learning benchmark.\" CoLLAs, 2022\n\n[5] Hussing et al., \"Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning.\" arXiv, 2023", "questions": "######## Additional feedback ########\n\nThe following points are provided as feedback to hopefully help better shape the submitted manuscript, but did not impact my recommendation in a major way.\n\nAbstract\n- The abstract is very clear. It lays out very well how the approach works and the results they obtain \n\nIntro\n- It's unclear what \"new tasks emerge overwhelmingly\" means or why van de Ven et al. is cited to support that claim. \n- The motivation for why CORL is special seems to be all about RL in general, and not specifically about offline RL. This seems to undermine the need to develop specialized approaches.\n- There's always the question of whether the size of the model might surpass the size of replay buffers. In visual settings that tends to happen. It's unclear if it does here\n- The idea of using generative models to express state and action distributions (which isn't novel) is good, especially in the offline setting where we can't make assumptions about the form of the distribution\n- What does \"behavior cloning matter\" mean in the context of a critic, which is not a behavior model?\n\nSec 2\n- I appreciate the notation clarification right before Sec 3! The authors could consider moving it to the beginning of Sec. 2 so the reader knows this ahead of time. \n- I thought we were missing a description of CORL, but that's in Sec. 3. Is the formalization of CORL a contribution of this work? If not, maybe it's worth also including it in Sec 2.\n\nSec 3.1\n- It does seem like the problem setting should be moved to Sec 2, and then the overview be placed in Sec 3 before introducing Sec 3.2 (which would be 3.1) below\n- Fig. 1 is useful. Why are there two (s,a,r,s') boxes in b and c? It seems like they are the same tuple, just that the first down arrow takes s and the second down arrow takes a. \n\nSec 4\n- \"the generator is the only constraint on task performance\" -- how is this an improvement over other continual learning methods?\n- \"When the generative model is optimal, training the networks with generative replay is equivalent to joint training on the entire dataset.\" \n    - Sure, an optimal replay method would achieve that... but can we actually train an optimal generator over a long sequence of tasks? Also, it is only equivalent to joint training if we actually do full joint training (from scratch), but not if we start from the previously trained models. Starting from previously trained models might be better or worse, but certainly not equivalent.\n- What is the (final) performance of SAC/TD3 on the collected datasets?\n- What is the \"oracle\" for the behavior model? And what is the \"noisy\" replay for the behavior model?\n- More than baselines, these seem to be ablations of CuGRO. While oracle is roughly a performance upper bound, it's unclear how other existing continual learning algorithms would perform compared to CuGRO/oracle. It's also unclear if non-diffusion approaches (given \"oracle\" or some other forgetting avoidance method) would work well.\n- I like the analysis of why Hopper-Vel fails\n- In Figure 3 we don't get to see past task performance. Or is this average performance including current and past tasks? Does Table 1 measure the final performance of all past tasks after training on the final task, or upon finishing training on each individual task?\n- The hyperparamter sensitivity analysis is good and useful.\n\nTypos/style/grammar\n- Footnotes in 3.2 should go before periods, not after the period and a space (e.g., \"...keep emerging\\footnote{text}.\")\n- Sec 3.2, paragraph 1 -- do datasets really \"emerge\"? Maybe they are \"constructed\" instead\n- Sec 3.2, paragraph 2 -- scored-based --> score-based\n- Sec 3.3, first line -- technically, the models are \\mu_\\phi and \\epsilon_\\psi, and \\phi eand \\psi are the parameters\n- I find the use of \"replayed\" samples (throughout the text) a bit odd, since they aren't real samples. I'd suggest using \"generated\" samples instead to consistently clarify that these are not real.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698337236787}], "openreview_url": "https://openreview.net/forum?id=EPfGHb9Y68", "arxiv_id": "2404.10662", "paper_pdf": "papers/EPfGHb9Y68.pdf", "paper_pdf_sha256": "e223ceead279b0da07cba3ab31719f26777c65ce83a1fcbb4fe9fcac5665b330", "paper_pdf_bytes": 1330490, "paper_pdf_source": "openreview", "code_url": "https://github.com/NJU-RL/CuGRO", "code_repository": "NJU-RL/CuGRO", "code_commit": "f3f09a4ac96e4e52a92cea946c98813480740af1", "code_archive": "repos/EPfGHb9Y68.zip", "code_archive_sha256": "92addfb40285eabf2d908b9c68798989216a044c430c2a8098964577fab19294", "code_archive_bytes": 516457, "code_file_count": 30, "code_extensions": {".py": 29, ".sh": 1}, "github_disk_usage_kb": 440, "github_languages": {"Python": 224631, "Shell": 1195}, "github_archived": false, "github_pushed_at": "2024-04-14T13:55:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/continual-offline-reinforcement-learning-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "UhEJz3wgLnG", "year": 2023, "status": "rejected", "title": "Revealing Single Frame Bias for Video-and-Language Learning", "authors": ["Jie Lei", "Tamara L Berg", "Mohit Bansal"], "authorids": ["~Jie_Lei3", "~Tamara_L_Berg1", "~Mohit_Bansal2"], "authors_source": "OpenReview API", "abstract": "Training an effective video-and-language model intuitively requires multiple frames as model inputs. \nHowever, it is unclear whether using multiple frames is beneficial to downstream tasks, and if yes, whether the performance gain is worth the drastically-increased computation and memory costs resulting from using more frames.\nIn this work, we explore single-frame models for video-and-language learning.\nOn a diverse set of video-and-language tasks (including text-to-video retrieval and video question answering), we show the surprising result that, with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame trained model that does not consider temporal information can achieve better performance than existing methods that use multiple frames for training.\nThis result reveals the existence of a strong ``static appearance bias'' in popular video-and-language datasets.\nTherefore, to allow for a more comprehensive evaluation of video-and-language models, we propose two new retrieval tasks based on existing fine-grained action recognition datasets that encourage temporal modeling.\nFull code and models will be made publicly available upon acceptance.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "n1wYh34ZGbb", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1465/Reviewer_pLom"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper investigates whether using multiple frames for training is necessary for video-language downstream tasks. The authors propose a single-frame framework for video-language understanding. Results indicate that with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame trained model that does not consider temporal information can achieve better performance than existing methods that use multiple frames for training. This result reveals strong static appearance bais in current video-langauge datasets. To allow for a more comprehensive evaluation of video-and-language models, the authors propose two new retrieval tasks based on existing fine-grained action recognition datasets that encourage temporal modeling. ", "review_text": "My main concern is the limited novelty, as it has been investigated by previous works that most video datasets do not need too much temporal information for understanding. The proposed single-frame architecture is a basic structure with self-attention and cross-attention, and the novelty is limited.", "strengths": "Strengths:\n1. The paper is clearly written and easy to follow.\n2. It reveals the single-frame bias for current video-language datasets.\n3. The proposed approach is simple and effective, and the authors prove that with the simple, single-frame model, the video-language models can already achieve good performance.\n\nWeaknesses:\n1. The novelty is limited. The single-frame model applies the basic image-language architectural design with self-attention and cross-attention.\n2. The single-frame bias for current video datasets is not new. Researchers have realized that most video datasets (except for SSv2) do not require much temporal information for understanding the contents.\n3. The SSv2 retrieval tasks are only naive extensions of the SSv2 dataset.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper investigates whether using multiple frames for training is necessary for video-language downstream tasks. The authors propose a single-frame framework for video-language understanding. Results indicate that with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame trained model that does not consider temporal information can achieve better performance than existing methods that use multiple frames for training. This result reveals strong static appearance bais in current video-langauge datasets. To allow for a more comprehensive evaluation of video-and-language models, the authors propose two new retrieval tasks based on existing fine-grained action recognition datasets that encourage temporal modeling. ", "strength_and_weaknesses": "Strengths:\n1. The paper is clearly written and easy to follow.\n2. It reveals the single-frame bias for current video-language datasets.\n3. The proposed approach is simple and effective, and the authors prove that with the simple, single-frame model, the video-language models can already achieve good performance.\n\nWeaknesses:\n1. The novelty is limited. The single-frame model applies the basic image-language architectural design with self-attention and cross-attention.\n2. The single-frame bias for current video datasets is not new. Researchers have realized that most video datasets (except for SSv2) do not require much temporal information for understanding the contents.\n3. The SSv2 retrieval tasks are only naive extensions of the SSv2 dataset.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written. The novelty is limited. The authors provide details of the proposed method to reproduce the results.", "summary_of_the_review": "My main concern is the limited novelty, as it has been investigated by previous works that most video datasets do not need too much temporal information for understanding. The proposed single-frame architecture is a basic structure with self-attention and cross-attention, and the novelty is limited.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667328404161}, {"id": "SU_rE5ABgH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1465/Reviewer_kab7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the domain of video-and-language tasks and analyses the methods and datasets in the space. Through empirical studies, authors show that using a single frame (randomly sampled) from the entire video is enough to reach similar performance as existing SOTA methods on these datasets. The claim is that such tasks do not need temporal understanding, merely static appearance/image understanding. Further, two new datasets are proposed in this space that require more temporal understanding.", "review_text": "Overall, not very convinced the paper is ready in its current form, especially as a long-form paper. The technical contributions are limited. The empirical studies do point to limitations of existing methods (and datasets to some extent) but do not propose any effective solutions to overcome such limitations.", "strengths": "**Strengths**\n\n-- The empirical studies presented in the work are insightful and will help propel more understanding in this relatively new domain of text-video understanding.\n\n\n**Weaknesses**\n\n-- While the empirical studies are certainly insightful, the paper's contributions are limited to just that. There are no other significant technical contributions.\n\n-- While the training is done with single frame, inference is still conducted using multiple sampled frames, which makes the setup a bit hard to fairly evaluate.\n\n-- The claim in the paper that existing datasets in the space (ActivityNet, MSRVTT) do not require any temporal understanding is only partially correct. Given that R1 results are still ~50%, it means that these are not solved datasets. The results point more towards the limitation of existing methods in fully learning/exploiting the temporal information in the videos rather a strict limitation of the datasets.\n\n-- The two new proposed datasets SSv2-{label|template} also have very high (and comparable) R1 scores with just sampling single frame which means those 2 new datasets are roughly comparable with other existing datasets.\n\n-- Overall, since the technical contribution in the paper is quite limited, the Sec 5 (Analysis) does not present any interesting studies and felt forced in the paper.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies the domain of video-and-language tasks and analyses the methods and datasets in the space. Through empirical studies, authors show that using a single frame (randomly sampled) from the entire video is enough to reach similar performance as existing SOTA methods on these datasets. The claim is that such tasks do not need temporal understanding, merely static appearance/image understanding. Further, two new datasets are proposed in this space that require more temporal understanding.", "strength_and_weaknesses": "**Strengths**\n\n-- The empirical studies presented in the work are insightful and will help propel more understanding in this relatively new domain of text-video understanding.\n\n\n**Weaknesses**\n\n-- While the empirical studies are certainly insightful, the paper's contributions are limited to just that. There are no other significant technical contributions.\n\n-- While the training is done with single frame, inference is still conducted using multiple sampled frames, which makes the setup a bit hard to fairly evaluate.\n\n-- The claim in the paper that existing datasets in the space (ActivityNet, MSRVTT) do not require any temporal understanding is only partially correct. Given that R1 results are still ~50%, it means that these are not solved datasets. The results point more towards the limitation of existing methods in fully learning/exploiting the temporal information in the videos rather a strict limitation of the datasets.\n\n-- The two new proposed datasets SSv2-{label|template} also have very high (and comparable) R1 scores with just sampling single frame which means those 2 new datasets are roughly comparable with other existing datasets.\n\n-- Overall, since the technical contribution in the paper is quite limited, the Sec 5 (Analysis) does not present any interesting studies and felt forced in the paper.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written for the most part. There is limited novelty in this work.", "summary_of_the_review": "Overall, not very convinced the paper is ready in its current form, especially as a long-form paper. The technical contributions are limited. The empirical studies do point to limitations of existing methods (and datasets to some extent) but do not propose any effective solutions to overcome such limitations.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666853417999}, {"id": "HQ_r4DEr3GR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1465/Reviewer_DUem"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper studies a video-language model trained using only a single frame, finding that is can do quite well on many existing benchmarks. The paper then proposes new benchmarks on something-something, where more temporal understanding is required.", "review_text": "The paper studies an important and interesting problem, finding that single frames are good enough for many video language tasks and proposes a solution to it.", "strengths": "The paper is well written and easy to understand. The experiments are well done. The finding that many current tasks can be done with a single frame training is interesting and useful. For the proposed task, it would be good to add more details. For example, how many unique labels are there?  It also be valuable to evaluate more existing methods on this benchmark to better understand how existing approachs perform on it.\n\nMinor details: \nThe bolding is misleading in Tables 1 and 2.\nShould use citep command for references to make it easier to read.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper studies a video-language model trained using only a single frame, finding that is can do quite well on many existing benchmarks. The paper then proposes new benchmarks on something-something, where more temporal understanding is required.", "strength_and_weaknesses": "The paper is well written and easy to understand. The experiments are well done. The finding that many current tasks can be done with a single frame training is interesting and useful. For the proposed task, it would be good to add more details. For example, how many unique labels are there?  It also be valuable to evaluate more existing methods on this benchmark to better understand how existing approachs perform on it.\n\nMinor details: \nThe bolding is misleading in Tables 1 and 2.\nShould use citep command for references to make it easier to read.", "clarity,_quality,_novelty_and_reproducibility": "The paper is clear and well written. It has details needed to reproduce the results. It isn't really novel, but has important findings. The proposed new benchmark is interesting and valuable, however the new benchmark is totally clear how useful it is.", "summary_of_the_review": "The paper studies an important and interesting problem, finding that single frames are good enough for many video language tasks and proposes a solution to it.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666619045733}, {"id": "_utws6R49Od", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1465/Reviewer_VHWz"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper explores single-frame training for video-and language tasks. While simple, this approach can achieve state-of-the-art performance on a range of datasets. This paper also proposes an early-fusion strategy at inference which boosts performance. ", "review_text": "Although a lot of paper has proposed the single-frame pre-training strategy (on the pure image-text pre-training dataset), the single-frame finetuning strategy is still novel to me. My main concerns are the reliability of the single-frame strategy on the pure large-scale web video datasets and the memory/time cost of the early-fusion strategy. I will raise my score if these concerns are addressed.", "strengths": "Strengths:\n1. The single-frame training strategy is highly efficient while effective, which only needs to train one frame during pre-training and fine-tuning. \n2. The proposed model achieves state-of-the-art among the video-text retrieval tasks on VideoQA tasks.\n3. This paper is clearly written and easy to follow.\n\nWeaknesses:\n1. The single-frame pre-training strategy (see Figure 6) is not fully convincing to me. Although the author claim that “when pre-trained on a sufficient amount of data, the performance of models trained with single frames might be very close to models trained with multiple frames”. However, there only exists 2.5M video-text data among the full 17M pre-training dataset. I think the reason is that with the growth of the #PT images/videos, the large-scale image-text dataset is enough for the model to achieve a good performance on downstream tasks (Note that a lot of recent works [1] have demonstrated that a pure image-text dataset is enough for video-text downstream tasks).\n2. Meanwhile, with the growth of the pre-training video-text dataset, there inevitably exists more noisy information and irrelevant frames in the video. Will the single-frame strategy still work on these large-scale web video datasets? I think an additional experiment on a pure large-scale web video dataset (instead of a pre-training dataset mix of videos and large portion images) will further enhance the reliability of this paper (e.g., HowTo100M).\n3. The proposed early-fusion strategy is simple and effective, which I do appreciate. However, compared with late-fusion, early-fusion strategy seems to need to input an Nx longer frame feature sequence into the multi-modal encoder. And this may bring an Nx higher memory/inference times cost compared with the late-fusion strategy. The author could further provide a memory/time cost comparison in Figure 4. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper explores single-frame training for video-and language tasks. While simple, this approach can achieve state-of-the-art performance on a range of datasets. This paper also proposes an early-fusion strategy at inference which boosts performance. ", "strength_and_weaknesses": "Strengths:\n1. The single-frame training strategy is highly efficient while effective, which only needs to train one frame during pre-training and fine-tuning. \n2. The proposed model achieves state-of-the-art among the video-text retrieval tasks on VideoQA tasks.\n3. This paper is clearly written and easy to follow.\n\nWeaknesses:\n1. The single-frame pre-training strategy (see Figure 6) is not fully convincing to me. Although the author claim that “when pre-trained on a sufficient amount of data, the performance of models trained with single frames might be very close to models trained with multiple frames”. However, there only exists 2.5M video-text data among the full 17M pre-training dataset. I think the reason is that with the growth of the #PT images/videos, the large-scale image-text dataset is enough for the model to achieve a good performance on downstream tasks (Note that a lot of recent works [1] have demonstrated that a pure image-text dataset is enough for video-text downstream tasks).\n2. Meanwhile, with the growth of the pre-training video-text dataset, there inevitably exists more noisy information and irrelevant frames in the video. Will the single-frame strategy still work on these large-scale web video datasets? I think an additional experiment on a pure large-scale web video dataset (instead of a pre-training dataset mix of videos and large portion images) will further enhance the reliability of this paper (e.g., HowTo100M).\n3. The proposed early-fusion strategy is simple and effective, which I do appreciate. However, compared with late-fusion, early-fusion strategy seems to need to input an Nx longer frame feature sequence into the multi-modal encoder. And this may bring an Nx higher memory/inference times cost compared with the late-fusion strategy. The author could further provide a memory/time cost comparison in Figure 4. \n", "clarity,_quality,_novelty_and_reproducibility": "This paper is clearly written and easy to follow. The author also provides the code of the paper. ", "summary_of_the_review": "Although a lot of paper has proposed the single-frame pre-training strategy (on the pure image-text pre-training dataset), the single-frame finetuning strategy is still novel to me. My main concerns are the reliability of the single-frame strategy on the pure large-scale web video datasets and the memory/time cost of the early-fusion strategy. I will raise my score if these concerns are addressed.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666534970164}], "openreview_url": "https://openreview.net/forum?id=UhEJz3wgLnG", "arxiv_id": "2206.03428", "paper_pdf": "papers/UhEJz3wgLnG.pdf", "paper_pdf_sha256": "05c4ee57035651a5f85ee84636374344f8eab67989d550e50bd340d8e35667c3", "paper_pdf_bytes": 1282037, "paper_pdf_source": "openreview", "code_url": "https://github.com/jayleicn/singularity", "code_repository": "jayleicn/singularity", "code_commit": "10a7f7062ab8af187b3a47803851b4e462191492", "code_archive": "repos/UhEJz3wgLnG.zip", "code_archive_sha256": "b11098042acc3621001a2cfc26f34a781e51be82ddb73a605a27aaa1e6213e4e", "code_archive_bytes": 1216638, "code_file_count": 34, "code_extensions": {".py": 28, ".sh": 6}, "github_disk_usage_kb": 1202, "github_languages": {"Python": 290881, "Shell": 18415}, "github_archived": false, "github_pushed_at": "2023-05-05T17:22:22Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/revealing-single-frame-bias-for-video-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "B0JH7vR2iGh", "year": 2022, "status": "rejected", "title": "PMIC: Improving Multi-Agent Reinforcement Learning with Progressive Mutual Information Collaboration", "authors": ["Pengyi Li", "Hongyao Tang", "Tianpei Yang", "Xiaotian Hao", "Sang Tong", "YAN ZHENG", "Jianye HAO", "Matthew E. Taylor", "Jinyi Liu"], "authorids": ["~Pengyi_Li1", "~Hongyao_Tang1", "~Tianpei_Yang1", "~Xiaotian_Hao1", "~Sang_Tong1", "~YAN_ZHENG1", "~Jianye_HAO1", "~Matthew_E._Taylor2", "~Jinyi_Liu1"], "authors_source": "OpenReview API", "abstract": "Learning to collaborate is critical in multi-agent reinforcement learning (MARL). A branch of previous works proposes to promote collaboration by maximizing the correlation of agents’ behaviors, which is typically characterised by mutual information (MI) in different forms. However, simply maximizing the MI of agents’ behaviors cannot guarantee achieving better collaboration because suboptimal collaboration can also lead to high MI. In this paper, we first propose a new collaboration criterion to evaluate collaboration from three perspectives, which arrives at a form of the mutual information between global state and joint policy. This bypasses the introduction of explicit additional input of policies and mitigates the scalability issue meanwhile. Moreover, to better leverage MI-based collaboration signals, we propose a novel MARL framework, called Progressive Mutual Information Collaboration (PMIC) which contains two main components. The first component is Dual Progressive Collaboration Buffer (DPCB) which separately stores superior and inferior trajectories in a progressive manner. The second component is Dual Mutual Information Estimator (DMIE), including two neural estimators of our new designed MI based on separate samples in DPCB. We then make use of the neural MI estimates to improve agents' policies: to maximize the MI lower bound associated with superior collaboration to facilitate better collaboration and to minimize the MI upper bound associated with inferior collaboration to avoid falling into local optimal. PMIC is general and can be combined with existing MARL algorithms. Experiments on a wide range of MARL benchmarks show the superior performance of PMIC compared with other MARL algorithms.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gbvnlJ3MgW", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1280/Reviewer_vXVk"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a new exploration scheme for general multi-agent reinforcement learning called PMIC. PMIC maintains two separate buffers by keeping trajectories with high rewards and trajectories with less satisfying rewards. PMIC additionally computes a lower bound of mutual information between state and policy over the positive buffer and an upper bound of the negative buffer. Finally, an intrinsic reward in the form of the lower bound minus the upper bound is introduced as the exploration bonus, which improves a collection of different MARL algorithms on a few benchmark testbeds. ", "review_text": "## Post-Rebuttal\nI now fully understand the algorithm. Thank the authors for the detailed response. I misunderstood the motivation of using the two bounds. Now the algorithm design looks pretty clear and intuitive to me, which is simply reducing MI on poor trajectories while increasing MI on good trajectories.\n\n**some suggestion on writing**\n1. I think the position of Sec. 3.2 makes it a bit misleading since it is presented after the motivating example (Sec. 3.1).  Sec. 3.1 has already assumed the optimizing objective for cooperation is MI and presents the motivating example based on this assumption. Note Sec 3.2 simply re-stated why MI can be a good optimization objective for learning, which is, in fact, the assumption in the previous subsection! Wouldn't it be more appropriate to justify the assumption before assuming it? This cyclic statement would make the presentation less consistent for the readers (including me) to follow. From the best readability, I would suggest the authors have a separate preliminary/motivation section, which contains the current content of Sec 3.2 first (to first show why optimizing MI is good) and then the content of Sec 3.1 (to show the issues of directly optimizing MI when having multiple patterns).\n\n1. I would still encourage the authors to make the best efforts to add the related work section back to the main paper. This paper is related to a lot of good ideas in existing works. It would be critical to mention them directly. The authors may possibly make the experiment sections a bit more concise and shrink the conclusion section to make some space. \n\n\n=======================================================================\n\n### Strong Points\n1. The proposed algorithm is, to the best of my knowledge, novel and it is indeed interesting to see such a separation of buffer with mutual information computation significantly boosts the performance of a collection of MARL algorithms, **with the assumption that the derivation is correct**.\n2. There are a lot of ablation studies conducted in the experiment section as well as the appendix, which is appreciated. The experiment section is clear and easy to follow.\n\n### Weak Points\nIn general, before moving into more detailed comments, I would strongly encourage the authors to substantially re-organize the paper to make it easier for the readers to follow. Also, most of the arguments and motivations in the paper are very hand-wavy, which makes me very concerned about the correctness of the proposed algorithm. It could be because of the writing issue, but, at least from the current statements in the paper, the correctness cannot be rigorously justified. More comments follow.\n\n1. A related work section would be encouraged.\n\n    Sec. 3.1 and the introduction section have a substantial overlap describing the existing literature, which should be put into a separate related work section. Fig.1 also looks strange as it appears in a motivation section. Fig.1 seems to only serve as a role of toy example showing that PMIC works better than some baselines. This is an experiment result rather than a motivation example! It doesn't make sense to me to motivate an algorithm by simply saying hey it works well. As a research paper, a reader would expect insights before moving into empirical evidence of your algorithm being good. I would appreciate it if the authors could rewrite Sec. 3.1 such that some more concrete examples of how the algorithm is working, for example, intuitively what kind of behavior leads to high MI; what characteristics are those sub-optimal strategies intuitively so we have a sense of what kind of policies we want to avoid; and so on. The current Sec.3.1 is simply a replication of the introduction section. \n\n2. I couldn't find how sec. 3.2 and sec 3.3 are related\n\n    > However, simply maximizing this MI can hinder algorithm learning as discussed above.\n\n    Eq.(1) definitely makes a lot of sense to me and there are existing works that consider similar objectives for MARL learning (not exactly the same objective but somewhat similar, like this one for example https://arxiv.org/pdf/2106.02195.pdf ). But why does this objective hinder learning? Is this really discussed? I have made my best efforts on understanding the discussions in Sec.3.1 but I could find any convincing arguments of how this objective hinders learning. \n\n   > Therefore, we propose a novel framework PMIC to solve the problem\n\n   Okay, even if this objective does have hindered learning, then why does the proposed PMIC algorithm solves tackles the issue in this objective? What is the relationship between Eq(1) and PMIC? To be concrete, for example, Sec.3.1 argues that the policy should be diverse (well, the terminology isn't really appropriate but let's use it for now) as suggested by using the entropy bonus but I couldn't find any corresponding term in PMIC for doing so. So what problem does PMIC solve? \n\n   There are so many hand-wavy arguments in this paper like the two above.\n\n\n3. There are no rigorous justifications on the relationship between high-reward trajectories and MINE/CLUB\n\n    > Intuitively, only the behaviors that obey superior collaboration patterns or inferior collaboration patterns have large MI estimates calculated by MINE or CLUB since $\\mathcal{L}(\\omega_1)$ and $\\mathcal{L}(\\omega_2)$ are optimized.\n\n    - *I think this statement is the foundation of the correctness of the method*. If this statement is correct, the whole algorithm becomes neat. Yes, I do have witnessed empirical analysis supporting this argument in Fig 9 and Fig 15. But, I want to emphasize that those two figures are empirical evidence rather than rigorous derivation. Typically, as a research paper, the conventional way of algorithm development is to first provide rigorous deviation (e.g., mathematical proof; equation analysis and approximation; or explanation and discussion on concrete toy examples), then develop an algorithm that solves the formulated problem and finally present empirical supports. The authors simply present this key argument by using the term ''__intuitively__'', which I failed to catch with my best efforts. This argument is the most critical point of the method and if this argument is wrong, the whole paper becomes problematic. Hence, I strongly encourage the authors to provide formal and rigorous justification or proof to this statement. \n\n    - Here is what confuses me a lot. Note that the two buffers are categorized according to their _rewards_ while both MINE and CLUB are bounds for _mutual information (MI)_. So what is the mathematical connection between rewards and MI? Why should superior strategies have higher MINE value? Can you mathematically prove or justify this? To the best of my knowledge, I cannot find any related proof in the existing literature.\n\n    - MINE is a generic lower bound for MI. So wouldn't optimizing MINE with the entire replay buffer strictly lead to a higher lower bound than only using the positive buffer (you have more data!)? Similarly, why not optimize CLUB using the entire buffer as well? Fig 9 does not include such ablation either. Assuming the correctness of all the arguments of MINE and CLUB, using the entire replay buffer should lead to at least the same or probably even better behavior compared with PMIC since you have both tighter lower bound and upper bound? \n\n### Other Issues\n1. > the joint policy should be diverse enough to avoid being stuck into sub-optimal collaboration.\n\n    Typically the term \"_diverse_\" refers to a set of (or a distribution of) different policies rather than a single joint policy. Also, according to the phrase \"_avoid being stuck into sub-optimal_\", it seems that diverse in fact means that the policy maintains high entropy _throughout the training process_ rather than describing the characteristics of the _final policy_ (the final policy can have low entropy if it converges but during training, it would better have high entropy for the purpose of exploration). So I would suggest the author direct state the purpose of avoiding poor local optimum rather than using the term \"_diverse_\" (you may be encouraged to refer to the literature of learning diverse skills/strategies in deep/MA RL literature)\n\n2. > 3) Meanwhile, given an agent i’s policy on the global state, the uncertainty on other agents’ policies also should be low, thus more easily to achieve better collaboration\n\n    I don't get why this point is necessary when we have 2) already. I couldn't find any corresponding objective that minimizes the intra-agent policy uncertainty in Sec 3.4. I don't know why this point is presented here. \n\n4. minor issues\n    - The predator-prey environment is not a cooperative one, for which the one-sided reward wouldn't become a convincing metric. Are you referring to the cooperative version of predator-prey? I didn't find any content confirming this. \n\n    - I would encourage the authors to have a sub-index for each plot in Fig 8 rather than saying sth like \"the third plot in fig.8\".\n\n    - A possibly related paper also using mutual information formulation for cooperative MARL: https://arxiv.org/abs/1910.05512\n \n    - Fig 15 is okay but it is not really super convincing. I can still observe some color differences in the right plot (I guess the values range from 0.6 to 1.0 maybe?) So probably you just need some tuning of beta or value normalization to also make VM3-AC work. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a new exploration scheme for general multi-agent reinforcement learning called PMIC. PMIC maintains two separate buffers by keeping trajectories with high rewards and trajectories with less satisfying rewards. PMIC additionally computes a lower bound of mutual information between state and policy over the positive buffer and an upper bound of the negative buffer. Finally, an intrinsic reward in the form of the lower bound minus the upper bound is introduced as the exploration bonus, which improves a collection of different MARL algorithms on a few benchmark testbeds. ", "main_review": "## Post-Rebuttal\nI now fully understand the algorithm. Thank the authors for the detailed response. I misunderstood the motivation of using the two bounds. Now the algorithm design looks pretty clear and intuitive to me, which is simply reducing MI on poor trajectories while increasing MI on good trajectories.\n\n**some suggestion on writing**\n1. I think the position of Sec. 3.2 makes it a bit misleading since it is presented after the motivating example (Sec. 3.1).  Sec. 3.1 has already assumed the optimizing objective for cooperation is MI and presents the motivating example based on this assumption. Note Sec 3.2 simply re-stated why MI can be a good optimization objective for learning, which is, in fact, the assumption in the previous subsection! Wouldn't it be more appropriate to justify the assumption before assuming it? This cyclic statement would make the presentation less consistent for the readers (including me) to follow. From the best readability, I would suggest the authors have a separate preliminary/motivation section, which contains the current content of Sec 3.2 first (to first show why optimizing MI is good) and then the content of Sec 3.1 (to show the issues of directly optimizing MI when having multiple patterns).\n\n1. I would still encourage the authors to make the best efforts to add the related work section back to the main paper. This paper is related to a lot of good ideas in existing works. It would be critical to mention them directly. The authors may possibly make the experiment sections a bit more concise and shrink the conclusion section to make some space. \n\n\n=======================================================================\n\n### Strong Points\n1. The proposed algorithm is, to the best of my knowledge, novel and it is indeed interesting to see such a separation of buffer with mutual information computation significantly boosts the performance of a collection of MARL algorithms, **with the assumption that the derivation is correct**.\n2. There are a lot of ablation studies conducted in the experiment section as well as the appendix, which is appreciated. The experiment section is clear and easy to follow.\n\n### Weak Points\nIn general, before moving into more detailed comments, I would strongly encourage the authors to substantially re-organize the paper to make it easier for the readers to follow. Also, most of the arguments and motivations in the paper are very hand-wavy, which makes me very concerned about the correctness of the proposed algorithm. It could be because of the writing issue, but, at least from the current statements in the paper, the correctness cannot be rigorously justified. More comments follow.\n\n1. A related work section would be encouraged.\n\n    Sec. 3.1 and the introduction section have a substantial overlap describing the existing literature, which should be put into a separate related work section. Fig.1 also looks strange as it appears in a motivation section. Fig.1 seems to only serve as a role of toy example showing that PMIC works better than some baselines. This is an experiment result rather than a motivation example! It doesn't make sense to me to motivate an algorithm by simply saying hey it works well. As a research paper, a reader would expect insights before moving into empirical evidence of your algorithm being good. I would appreciate it if the authors could rewrite Sec. 3.1 such that some more concrete examples of how the algorithm is working, for example, intuitively what kind of behavior leads to high MI; what characteristics are those sub-optimal strategies intuitively so we have a sense of what kind of policies we want to avoid; and so on. The current Sec.3.1 is simply a replication of the introduction section. \n\n2. I couldn't find how sec. 3.2 and sec 3.3 are related\n\n    > However, simply maximizing this MI can hinder algorithm learning as discussed above.\n\n    Eq.(1) definitely makes a lot of sense to me and there are existing works that consider similar objectives for MARL learning (not exactly the same objective but somewhat similar, like this one for example https://arxiv.org/pdf/2106.02195.pdf ). But why does this objective hinder learning? Is this really discussed? I have made my best efforts on understanding the discussions in Sec.3.1 but I could find any convincing arguments of how this objective hinders learning. \n\n   > Therefore, we propose a novel framework PMIC to solve the problem\n\n   Okay, even if this objective does have hindered learning, then why does the proposed PMIC algorithm solves tackles the issue in this objective? What is the relationship between Eq(1) and PMIC? To be concrete, for example, Sec.3.1 argues that the policy should be diverse (well, the terminology isn't really appropriate but let's use it for now) as suggested by using the entropy bonus but I couldn't find any corresponding term in PMIC for doing so. So what problem does PMIC solve? \n\n   There are so many hand-wavy arguments in this paper like the two above.\n\n\n3. There are no rigorous justifications on the relationship between high-reward trajectories and MINE/CLUB\n\n    > Intuitively, only the behaviors that obey superior collaboration patterns or inferior collaboration patterns have large MI estimates calculated by MINE or CLUB since $\\mathcal{L}(\\omega_1)$ and $\\mathcal{L}(\\omega_2)$ are optimized.\n\n    - *I think this statement is the foundation of the correctness of the method*. If this statement is correct, the whole algorithm becomes neat. Yes, I do have witnessed empirical analysis supporting this argument in Fig 9 and Fig 15. But, I want to emphasize that those two figures are empirical evidence rather than rigorous derivation. Typically, as a research paper, the conventional way of algorithm development is to first provide rigorous deviation (e.g., mathematical proof; equation analysis and approximation; or explanation and discussion on concrete toy examples), then develop an algorithm that solves the formulated problem and finally present empirical supports. The authors simply present this key argument by using the term ''__intuitively__'', which I failed to catch with my best efforts. This argument is the most critical point of the method and if this argument is wrong, the whole paper becomes problematic. Hence, I strongly encourage the authors to provide formal and rigorous justification or proof to this statement. \n\n    - Here is what confuses me a lot. Note that the two buffers are categorized according to their _rewards_ while both MINE and CLUB are bounds for _mutual information (MI)_. So what is the mathematical connection between rewards and MI? Why should superior strategies have higher MINE value? Can you mathematically prove or justify this? To the best of my knowledge, I cannot find any related proof in the existing literature.\n\n    - MINE is a generic lower bound for MI. So wouldn't optimizing MINE with the entire replay buffer strictly lead to a higher lower bound than only using the positive buffer (you have more data!)? Similarly, why not optimize CLUB using the entire buffer as well? Fig 9 does not include such ablation either. Assuming the correctness of all the arguments of MINE and CLUB, using the entire replay buffer should lead to at least the same or probably even better behavior compared with PMIC since you have both tighter lower bound and upper bound? \n\n### Other Issues\n1. > the joint policy should be diverse enough to avoid being stuck into sub-optimal collaboration.\n\n    Typically the term \"_diverse_\" refers to a set of (or a distribution of) different policies rather than a single joint policy. Also, according to the phrase \"_avoid being stuck into sub-optimal_\", it seems that diverse in fact means that the policy maintains high entropy _throughout the training process_ rather than describing the characteristics of the _final policy_ (the final policy can have low entropy if it converges but during training, it would better have high entropy for the purpose of exploration). So I would suggest the author direct state the purpose of avoiding poor local optimum rather than using the term \"_diverse_\" (you may be encouraged to refer to the literature of learning diverse skills/strategies in deep/MA RL literature)\n\n2. > 3) Meanwhile, given an agent i’s policy on the global state, the uncertainty on other agents’ policies also should be low, thus more easily to achieve better collaboration\n\n    I don't get why this point is necessary when we have 2) already. I couldn't find any corresponding objective that minimizes the intra-agent policy uncertainty in Sec 3.4. I don't know why this point is presented here. \n\n4. minor issues\n    - The predator-prey environment is not a cooperative one, for which the one-sided reward wouldn't become a convincing metric. Are you referring to the cooperative version of predator-prey? I didn't find any content confirming this. \n\n    - I would encourage the authors to have a sub-index for each plot in Fig 8 rather than saying sth like \"the third plot in fig.8\".\n\n    - A possibly related paper also using mutual information formulation for cooperative MARL: https://arxiv.org/abs/1910.05512\n \n    - Fig 15 is okay but it is not really super convincing. I can still observe some color differences in the right plot (I guess the values range from 0.6 to 1.0 maybe?) So probably you just need some tuning of beta or value normalization to also make VM3-AC work. \n", "summary_of_the_review": "Post-rebuttal: \nAfter carefully re-checking the paper with the feedback from the authors, I have fully convinced by the methods and therefore changed my recommendation to acceptance. \n\n==================================\n\nThe paper proposes a seemingly neat algorithm and presents a good amount of experiments. However, some critical correctness flaws cannot be rigorously justified based on the current content of the paper due to too many hand-wavy arguments. The readability of the paper should be improved as well.\n\nI cannot recommend acceptance based on the current form of the paper. But, I would still believe in the potential of the algorithm based on the experiment results. That being said, the paper still has the potential to become a really strong one if the algorithm can be rigorously (or mathematically) justified.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635870199837}, {"id": "VFhndk9ngV2", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1280/Reviewer_QvKR"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes to maximize the MI between the global state and joint policy for promoting collaboration. To achieve better collaboration, a Dual Progressive Collaboration Buffer (DPCB) which stores superior and inferior samples separately is introduced, and then the proposed method maximizes the lower bound on MI using the samples in superior buffer and minimizes the upper bound on MI using the samples in inferior samples. ", "review_text": "Strengths\n\n- This paper addresses an important issue of the MI-based MARL framework where the joint policy converges to sub-optimal.\n- The proposed idea, which maximizes MI based on superior samples and minimizes MI based on inferior samples to avoid the joint policy falling into the sub-optimal, is novel.\n- The authors provide various ablation studies for better understanding.\n\n\nWeakness and Questions\n\n-There is a lack of explanation for the collaboration criterion (section 3.2), e.g, as I understand, $s$ is a state variable whose distribution is conditioned on $\\pi$. \n-For the interpretation of the proposed MI form, the authors decompose the proposed MI into the entropy of join policy ($H(\\pi(\\cdot|s))$, the negative entropy of an agent's policy ($-H(\\pi_i|s)$) and conditional negative entropy of the agent's policy given other agents' policies ($-H(\\pi_{-i}|, \\pi_i, s)$). The authors claim that maximizing $-H(\\pi_{-i}|, \\pi_i, s)$ minimizes uncertainty on other agents' policies. However, $H(\\pi_{-i}|, \\pi_i, s)$ reduces to $H(\\pi_{-i}|s)$ since $\\pi_{-i}$ and $\\pi_i$ are independent given the state. Thus, it seems that the interpretation of the proposed MI form, $I(s;\\pi(\\cdot|s))$, is not proper. In this regard, can we say that the proposed MI form causes the collaboration?\n-The authors minimize $I(s;\\pi(\\cdot|s))$ for experiences which have low returns, so it encourages the joint action to explore for inferior experiences. This is because the joint action should be uniform to minimize $I(s;\\pi(\\cdot|s))$. On the other hand, maximizing $I(s;\\pi(\\cdot|s))$ for experiences that have high returns encourages the joint action to exploit for superior experiences. Thus, in my opinion, the effectiveness of the proposed method comes from the further exploration and exploitation strategy rather than the collaboration strategy. To see where the performance gain comes from, it would be great if the authors compare the proposed method with MA-SAC+DPCB and VM3-AC+DPCB.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes to maximize the MI between the global state and joint policy for promoting collaboration. To achieve better collaboration, a Dual Progressive Collaboration Buffer (DPCB) which stores superior and inferior samples separately is introduced, and then the proposed method maximizes the lower bound on MI using the samples in superior buffer and minimizes the upper bound on MI using the samples in inferior samples. ", "main_review": "Strengths\n\n- This paper addresses an important issue of the MI-based MARL framework where the joint policy converges to sub-optimal.\n- The proposed idea, which maximizes MI based on superior samples and minimizes MI based on inferior samples to avoid the joint policy falling into the sub-optimal, is novel.\n- The authors provide various ablation studies for better understanding.\n\n\nWeakness and Questions\n\n-There is a lack of explanation for the collaboration criterion (section 3.2), e.g, as I understand, $s$ is a state variable whose distribution is conditioned on $\\pi$. \n-For the interpretation of the proposed MI form, the authors decompose the proposed MI into the entropy of join policy ($H(\\pi(\\cdot|s))$, the negative entropy of an agent's policy ($-H(\\pi_i|s)$) and conditional negative entropy of the agent's policy given other agents' policies ($-H(\\pi_{-i}|, \\pi_i, s)$). The authors claim that maximizing $-H(\\pi_{-i}|, \\pi_i, s)$ minimizes uncertainty on other agents' policies. However, $H(\\pi_{-i}|, \\pi_i, s)$ reduces to $H(\\pi_{-i}|s)$ since $\\pi_{-i}$ and $\\pi_i$ are independent given the state. Thus, it seems that the interpretation of the proposed MI form, $I(s;\\pi(\\cdot|s))$, is not proper. In this regard, can we say that the proposed MI form causes the collaboration?\n-The authors minimize $I(s;\\pi(\\cdot|s))$ for experiences which have low returns, so it encourages the joint action to explore for inferior experiences. This is because the joint action should be uniform to minimize $I(s;\\pi(\\cdot|s))$. On the other hand, maximizing $I(s;\\pi(\\cdot|s))$ for experiences that have high returns encourages the joint action to exploit for superior experiences. Thus, in my opinion, the effectiveness of the proposed method comes from the further exploration and exploitation strategy rather than the collaboration strategy. To see where the performance gain comes from, it would be great if the authors compare the proposed method with MA-SAC+DPCB and VM3-AC+DPCB.", "summary_of_the_review": "-The proposed idea is novel, but the author should provide further explanation on collaboration criterion and further experiments to see where the performance gain comes from. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635840870809}, {"id": "r4Qxy1BOpik", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1280/Reviewer_5fQS"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes PMIC a MARL framework for improving multi-agent collaboration through mutual information. PMIC uses two separate buffers to store the superior and inferior trajectories, and uses them to train two estimators for the upper and lower bounds for the mutual information between global state and joint policy in two sets of trajectories. The estimated mutual information bounds are then used as additional rewards for training individual policies. The experiments presented in the paper show the advantage over several baselines on several benchmark environments.", "review_text": "Strengths:\nThe proposed method is novel and shows good performance.\n\nWeaknesses:\n1. While the results generally look good, some reward curves in Figure 5, 6, 7, 8, and 9 are not converged. It would be better to compare the converged performance between baselines. \n2. I am not quite following why minimizing the MI in inferior trajectories is helping improve the corporation. Generally, I would suggest more explanations on the motivation of the PMIC.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes PMIC a MARL framework for improving multi-agent collaboration through mutual information. PMIC uses two separate buffers to store the superior and inferior trajectories, and uses them to train two estimators for the upper and lower bounds for the mutual information between global state and joint policy in two sets of trajectories. The estimated mutual information bounds are then used as additional rewards for training individual policies. The experiments presented in the paper show the advantage over several baselines on several benchmark environments.", "main_review": "Strengths:\nThe proposed method is novel and shows good performance.\n\nWeaknesses:\n1. While the results generally look good, some reward curves in Figure 5, 6, 7, 8, and 9 are not converged. It would be better to compare the converged performance between baselines. \n2. I am not quite following why minimizing the MI in inferior trajectories is helping improve the corporation. Generally, I would suggest more explanations on the motivation of the PMIC.", "summary_of_the_review": "I think the submission overall is good for its novelty and experiment results. I only have some minor concerns about the theoretical backup of the proposed method. Thus, I suggest accepting the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635831986720}, {"id": "y46qzi5YNs3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1280/Reviewer_iVbL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper aims to solve the collaboration among agents (that is usually called forced coordination in other literatures). The common method to enhance the correlations between agents by mutual information (MI) could lead to the sub-optimal collaboration. To address this issue, the author propose Progressive Mutual Information Collaboration (PMIC). This new MARL framework is composed of two components: (1) two replay buffers that respectively collect superior trajectories and inferior trajectories; (2) two mutual information estimators that respectively estimate the upper bound and the lower bound of MI that are trained with the data stored in the two replay buffers introduced above. These two estimators (MINE and CLUB) are from the prior works. The novelties of this work are applying these two estimators to collaboration problem for MARL and learning the mutual information that evaluates the correlation among agents constrained within these two bounds. In my view, the idea of appropriately modelling these MI bound estimators to the actual phenomenon in the MARL problem is novel. Seen from the algorithmic framework type, this algorithm add the learned MI term to the reward function that plays the role of auxiliary reward. From the perspective of optimization, this is a kind of implementation of Lagrangian method, to transform the constraints stated in Equation (2) and (3) to a unconstrained terms involved in the objective function. Similar to many prior works, the tuning of the multipliers could be a potential issue. The authors also conduct the ablation studies on these two multipliers and the results seem like it would not affect the performance in the case of Cooperation Navigation.", "review_text": "## Strengths\n\n1. The writing of this paper is generally clear.\n2. The motivation of this paper is clear that aims to solves the problem that is observed from the existing state-of-the-art algorithms.\n3. The experiments are sufficient, including the discrete/continuous action cases, different environments/tasks and multiple ablation studies to verify the effectiveness of the proposed framework.\n4. The proposed framework can be incorporated into any MARL algorithms, so it is a general method.\n\n## Weaknesses\n\nThis paper does not have obvious weaknesses, but I have some concerns that the authors need to clarify.\n1. Can the authors give more details about how $\\mathbb{P}\\_{ \\mathcal{S} \\mathcal{U} }$, $\\mathbb{P}\\_{\\mathcal{U}}$ and $\\mathbb{P}\\_{\\mathcal{S}}$ implemented?\n2. In experiments, the authors claim that the results from the first two graphs in Figure 8 shows that the proposed maximizing and minimizing MI can improve the stability. However, the stability of the proposed method is not well, i.e., the variance of the proposed method is much larger than other competitors. Can the authors give more explanations on this?\n3. For the abalation study in Figure 9, the authors did not show the case that PMIC-MADDPG-normal\\_buffer that can directly show the effectiveness of the proposed replay buffer. Can the authors show the result for it?\n4. If I understand correctly, $T\\_{\\omega\\_{1}}$ is designed with the state and action encoders while $T\\_{\\omega\\_{2}}$ is without. Can the authors explain the reason for the inconsistent design for these two models? It is better that the authors can provide the results for MINE without state/action encoders to get rid of the suspect that the performance improvement relies on the encoders rather than the main contribution claimed in the paper.\n\n## Minors\n\nThere are multiple writing typos in the paper.\n\n1. \"Similarly, CLUB approximates the lower bound ...\" where lower bound should be upper bound. \n2. $I(s; \\pi(u|s)$ appearing in multiple places should be $I(s; \\pi(u|s) )$.\n3. Above Equation (3), $D\\_{a}(\\cdot)$ should be $D\\_{u}(\\cdot)$.\n\nThe above are just the ones I found. I urge the authors to check the paper writing again, and revise these writing typos in the discussion stage.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper aims to solve the collaboration among agents (that is usually called forced coordination in other literatures). The common method to enhance the correlations between agents by mutual information (MI) could lead to the sub-optimal collaboration. To address this issue, the author propose Progressive Mutual Information Collaboration (PMIC). This new MARL framework is composed of two components: (1) two replay buffers that respectively collect superior trajectories and inferior trajectories; (2) two mutual information estimators that respectively estimate the upper bound and the lower bound of MI that are trained with the data stored in the two replay buffers introduced above. These two estimators (MINE and CLUB) are from the prior works. The novelties of this work are applying these two estimators to collaboration problem for MARL and learning the mutual information that evaluates the correlation among agents constrained within these two bounds. In my view, the idea of appropriately modelling these MI bound estimators to the actual phenomenon in the MARL problem is novel. Seen from the algorithmic framework type, this algorithm add the learned MI term to the reward function that plays the role of auxiliary reward. From the perspective of optimization, this is a kind of implementation of Lagrangian method, to transform the constraints stated in Equation (2) and (3) to a unconstrained terms involved in the objective function. Similar to many prior works, the tuning of the multipliers could be a potential issue. The authors also conduct the ablation studies on these two multipliers and the results seem like it would not affect the performance in the case of Cooperation Navigation.", "main_review": "## Strengths\n\n1. The writing of this paper is generally clear.\n2. The motivation of this paper is clear that aims to solves the problem that is observed from the existing state-of-the-art algorithms.\n3. The experiments are sufficient, including the discrete/continuous action cases, different environments/tasks and multiple ablation studies to verify the effectiveness of the proposed framework.\n4. The proposed framework can be incorporated into any MARL algorithms, so it is a general method.\n\n## Weaknesses\n\nThis paper does not have obvious weaknesses, but I have some concerns that the authors need to clarify.\n1. Can the authors give more details about how $\\mathbb{P}\\_{ \\mathcal{S} \\mathcal{U} }$, $\\mathbb{P}\\_{\\mathcal{U}}$ and $\\mathbb{P}\\_{\\mathcal{S}}$ implemented?\n2. In experiments, the authors claim that the results from the first two graphs in Figure 8 shows that the proposed maximizing and minimizing MI can improve the stability. However, the stability of the proposed method is not well, i.e., the variance of the proposed method is much larger than other competitors. Can the authors give more explanations on this?\n3. For the abalation study in Figure 9, the authors did not show the case that PMIC-MADDPG-normal\\_buffer that can directly show the effectiveness of the proposed replay buffer. Can the authors show the result for it?\n4. If I understand correctly, $T\\_{\\omega\\_{1}}$ is designed with the state and action encoders while $T\\_{\\omega\\_{2}}$ is without. Can the authors explain the reason for the inconsistent design for these two models? It is better that the authors can provide the results for MINE without state/action encoders to get rid of the suspect that the performance improvement relies on the encoders rather than the main contribution claimed in the paper.\n\n## Minors\n\nThere are multiple writing typos in the paper.\n\n1. \"Similarly, CLUB approximates the lower bound ...\" where lower bound should be upper bound. \n2. $I(s; \\pi(u|s)$ appearing in multiple places should be $I(s; \\pi(u|s) )$.\n3. Above Equation (3), $D\\_{a}(\\cdot)$ should be $D\\_{u}(\\cdot)$.\n\nThe above are just the ones I found. I urge the authors to check the paper writing again, and revise these writing typos in the discussion stage.", "summary_of_the_review": "In summary, this paper has some novolty but stays at the application level, though this is not my concern. My main concern is the effectiveness of the main contribution claimed in the paper, i.e. the maximization-minimization MI. I concern that whether the performance improvement mainly comes from the specific architecture $T$ used in the framework.\n\nFor the current version of paper, I can only recommend reject (but marginally below the threshold) due to the weakneses and concerns above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No.", "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635636460686}], "openreview_url": "https://openreview.net/forum?id=B0JH7vR2iGh", "arxiv_id": "2203.08553", "paper_pdf": "papers/B0JH7vR2iGh.pdf", "paper_pdf_sha256": "0ad7a1596fb2cb8ad546fe7166678a956a7760fa6adb1dc5f14e649b8430c7e5", "paper_pdf_bytes": 17889723, "paper_pdf_source": "openreview", "code_url": "https://github.com/yeshenpy/PMIC", "code_repository": "yeshenpy/PMIC", "code_commit": "ab3fee60b1c75d65c3c3f209397130c059eedb41", "code_archive": "repos/B0JH7vR2iGh.zip", "code_archive_sha256": "23663713a6b2dd533928d8dbd6b131569352fded2dc3d5a7025cff66d2fc2888", "code_archive_bytes": 5562356, "code_file_count": 234, "code_extensions": {".py": 232, ".ipynb": 1, ".sh": 1}, "github_disk_usage_kb": 5374, "github_languages": {"Python": 1131343, "HTML": 592640, "Jupyter Notebook": 580298, "Shell": 1069, "Dockerfile": 465}, "github_archived": false, "github_pushed_at": "2024-03-26T06:41:25Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pmic-improving-multi-agent-reinforcement-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "bVzUDC_4ls", "year": 2021, "status": "rejected", "title": "Exploiting Verified Neural Networks via Floating Point Numerical Error", "authors": ["Kai Jia", "Martin Rinard"], "authorids": ["~Kai_Jia2", "~Martin_Rinard1"], "authors_source": "OpenReview API", "abstract": "Motivated by the need to reliably characterize the robustness of deep neural networks, researchers have developed verification algorithms for deep neural networks. Given a neural network, the verifiers aim to answer whether certain properties are guaranteed with respect to all inputs in a space. However, little attention has been paid to floating point numerical error in neural network verification.\n\nWe exploit floating point errors in the inference and verification implementations to construct adversarial examples for neural networks that a verifier claims to be robust with respect to certain inputs. We argue that, to produce sound verification results, any verification system must accurately (or conservatively) model the effects of any float point computations in the network inference or verification system.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "zd3H4EhfIGP", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2625/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper focuses on the floating point unsoundness of neural network verification procedures, and shows that the results from such verifiers can not be trusted. To drive home the message of the paper, the authors take MIPVerify (which doesn't ensure FP soundness) and shows that they are able to construct adversarial examples for the cases that are returned as verified by MIPVerify. \n\nIt is an interesting nice paper but the contribution is weakened by two  facts: One, it is already known that MIPVerify doesn't ensure FP soundness; which is also acknowledged by its authors. Second, when FP soundness is not ensured, and given the fact that adversarial examples are widely present, it is no surprise that one could find adversarial examples. \n\nSo I am split on the paper. From a formal methods perspective, the discovery of the paper is not surprising as Floating point computations are known to be important (this is one of the reasons that SMT solvers put a lot of emphasis on FP). But perhaps from ML practitioner, it may be interesting. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Paper but lacks sufficient contribution", "review": "This paper focuses on the floating point unsoundness of neural network verification procedures, and shows that the results from such verifiers can not be trusted. To drive home the message of the paper, the authors take MIPVerify (which doesn't ensure FP soundness) and shows that they are able to construct adversarial examples for the cases that are returned as verified by MIPVerify. \n\nIt is an interesting nice paper but the contribution is weakened by two  facts: One, it is already known that MIPVerify doesn't ensure FP soundness; which is also acknowledged by its authors. Second, when FP soundness is not ensured, and given the fact that adversarial examples are widely present, it is no surprise that one could find adversarial examples. \n\nSo I am split on the paper. From a formal methods perspective, the discovery of the paper is not surprising as Floating point computations are known to be important (this is one of the reasons that SMT solvers put a lot of emphasis on FP). But perhaps from ML practitioner, it may be interesting. ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603998640485}, {"id": "ZnkLSqs2wmy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2625/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe authors develop a method to generate pairs of sample that are separated by a small adversarial perturbation, that have different class, but with the specificity that the a complete verifier would returns a result indicating that this sample admits no adversarial perturbation (despite the fact that it does, as evidenced by the second element of the pair). \nThese samples are obtained by considering a brightness perturbation of the image and finding the parameter (alpha) at which the verifier switch from returning \"safe\" to \"unsafe\". The resulting perturbed image is going to have adversarial examples very close to the boundary of the region considered, so small floating point errors might result in returning incorrect results.\n\nMain thoughts:\nThe problem that the author discuss is very well highlighted and explained. It is clear what vulnerability they identified, as well as the mechanism that they use to highlight it. \nOn the other hand, in terms of importance, I would rank it more as an interesting observation that an actual critical problem. If we assume, that what I'm caring is robustness of my image classification system for perturbation of size epsilon=0.1, then it seems that the worst that can happen is that some samples that I verified to be robust for epsilon=0.1, are in practice only robust for epsilon=0.09999? This doesn't seem overtly critical and would result in essentially the same result in any application.\n\nQuestions:\n- The choice of what solver to use as a backend for a MIP formulation of the Neural Network verification problem is an implementation detail. MIPVerify could well be implemented with a different solver? (MIP solvers returning incorrect result due to floating point errors is not a new problem and there seems to be some literature in how to adress these problems if they are considered of importance \"Safe bounds in linear and mixed-integer linear programming, Neumaier & Shcherbina\")\nIn addition, could this problem be solved by simply adjusting the tolerance parameters of the solver? I did not see any discussion of this by the authors, but I imagine that the default parameters used by the verifier might be geared more towards speed than towards perfect accuracy.\n- The authors mention verifiers that incorporate proper handling of floating point errors (ERAN) but then reject it by saying that it rely on specific implementation details of the inference algorithm. This seems strange because that's exactly the recommendation that the authors make. Page 2: \"any sound verifier for this class of networks must reason about the specific floating point error characteristics of the neural network implementation at hand.\"\n\nMinor questions:\nIn Figure 1.a, it seems like for the first 4 graphs, the dotted lines which I assume implies what the difference should be are lines with slope 1. Why would the change in the logit vector vary at the same rate as the perturbations? Shouldn't there be a slope dependent on the corresponding gradient coefficient?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good presentation but problem of limited impact.", "review": "Summary:\nThe authors develop a method to generate pairs of sample that are separated by a small adversarial perturbation, that have different class, but with the specificity that the a complete verifier would returns a result indicating that this sample admits no adversarial perturbation (despite the fact that it does, as evidenced by the second element of the pair). \nThese samples are obtained by considering a brightness perturbation of the image and finding the parameter (alpha) at which the verifier switch from returning \"safe\" to \"unsafe\". The resulting perturbed image is going to have adversarial examples very close to the boundary of the region considered, so small floating point errors might result in returning incorrect results.\n\nMain thoughts:\nThe problem that the author discuss is very well highlighted and explained. It is clear what vulnerability they identified, as well as the mechanism that they use to highlight it. \nOn the other hand, in terms of importance, I would rank it more as an interesting observation that an actual critical problem. If we assume, that what I'm caring is robustness of my image classification system for perturbation of size epsilon=0.1, then it seems that the worst that can happen is that some samples that I verified to be robust for epsilon=0.1, are in practice only robust for epsilon=0.09999? This doesn't seem overtly critical and would result in essentially the same result in any application.\n\nQuestions:\n- The choice of what solver to use as a backend for a MIP formulation of the Neural Network verification problem is an implementation detail. MIPVerify could well be implemented with a different solver? (MIP solvers returning incorrect result due to floating point errors is not a new problem and there seems to be some literature in how to adress these problems if they are considered of importance \"Safe bounds in linear and mixed-integer linear programming, Neumaier & Shcherbina\")\nIn addition, could this problem be solved by simply adjusting the tolerance parameters of the solver? I did not see any discussion of this by the authors, but I imagine that the default parameters used by the verifier might be geared more towards speed than towards perfect accuracy.\n- The authors mention verifiers that incorporate proper handling of floating point errors (ERAN) but then reject it by saying that it rely on specific implementation details of the inference algorithm. This seems strange because that's exactly the recommendation that the authors make. Page 2: \"any sound verifier for this class of networks must reason about the specific floating point error characteristics of the neural network implementation at hand.\"\n\nMinor questions:\nIn Figure 1.a, it seems like for the first 4 graphs, the dotted lines which I assume implies what the difference should be are lines with slope 1. Why would the change in the logit vector vary at the same rate as the perturbations? Shouldn't there be a slope dependent on the corresponding gradient coefficient?", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603883384810}, {"id": "4NIpHlIh5AT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2625/AnonReviewer3"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper presents a method to find adversarial inputs for neural networks in regions where the networks can be \"proven\" not to admit any such adversarial examples, practically demonstrating the unsoundness of a \"complete verifier\" as well as an \"incomplete verifier\".\nWhile it was already obvious to me that \"verifiers\" that assume floating-point arithmetic is the same as real arithmetic are unsound, the paper is a service to the community in that it also makes this very obvious to informed outsiders who may not have already questioned the validity of robustness verification research that does not model round-off and even ignores it in its own implementation. The related work section does a good job of surveying the state of the art as it relates to floating-point soundness. The authors also took some space to discuss how their findings relate to current and future research on robustness verification, which I think is important in this case.\n\nPerhaps there could be a short discussion of challenges that different approaches face to become sound with respect to floating-point semantics. (For example, it seems particularly challenging for approaches based on duality, as the correctness of certificates depends non-trivially on closed-form solutions to optimization problems as well as associativity of addition.)\n\nThe technical sections are mostly well-written, though I was not able to figure out some details. For example, it is not so clear how precisely binary search is used to find α and δ simultaneously. Section 4.2 is a bit dense and its presentation could probably be improved.\n\n\n\"inevitable presence of numerical error in [...] the verifier\".\nIt is not inevitable that the verifier is subject to \"error\". We could encode the precise floating-point semantics of the neural network as a SAT formula (and then watch the SAT solver time out, but this does give a sound and complete method).\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A service to the community", "review": "The paper presents a method to find adversarial inputs for neural networks in regions where the networks can be \"proven\" not to admit any such adversarial examples, practically demonstrating the unsoundness of a \"complete verifier\" as well as an \"incomplete verifier\".\nWhile it was already obvious to me that \"verifiers\" that assume floating-point arithmetic is the same as real arithmetic are unsound, the paper is a service to the community in that it also makes this very obvious to informed outsiders who may not have already questioned the validity of robustness verification research that does not model round-off and even ignores it in its own implementation. The related work section does a good job of surveying the state of the art as it relates to floating-point soundness. The authors also took some space to discuss how their findings relate to current and future research on robustness verification, which I think is important in this case.\n\nPerhaps there could be a short discussion of challenges that different approaches face to become sound with respect to floating-point semantics. (For example, it seems particularly challenging for approaches based on duality, as the correctness of certificates depends non-trivially on closed-form solutions to optimization problems as well as associativity of addition.)\n\nThe technical sections are mostly well-written, though I was not able to figure out some details. For example, it is not so clear how precisely binary search is used to find α and δ simultaneously. Section 4.2 is a bit dense and its presentation could probably be improved.\n\n\n\"inevitable presence of numerical error in [...] the verifier\".\nIt is not inevitable that the verifier is subject to \"error\". We could encode the precise floating-point semantics of the neural network as a SAT formula (and then watch the SAT solver time out, but this does give a sound and complete method).\n", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603845610110}, {"id": "NDpPlRFSWA3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2625/AnonReviewer2"], "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In the recent literature there has been a rise in the number of papers which attempt to verify neural networks. The specification of the verification problems often gets adapted according to the application in mind. More specifically, for image classification networks, the problem is to prove that the output of the neural network does not flip for small perturbations to the pixel values. For a robotic setting, the problem is often safety and convergence to some goal state. Where the neural network operates in closed loop with the system dynamics. \n\nThe authors in this paper present an adversarial attack model on neural networks, which is deemed correct by some verifier. More specifically , given a neural network which can be shown to be robust to adversarial perturbations around some input, the authors exploit numerical errors in the computations to attack the network. Demonstrating the presence of loop holes in the proving engines itself. This is due to the approximation errors introduced by using floating point numbers.\n\nIn my opinion, the notion of input sets in the space of images, is not a very useful one. Mainly because the interval valued sets representing  perturbations of the input image, is far removed from the intended specification.  It's a step in the right direction, if the verification of computer vision task was a well defined problem. Since it's not clear what to verify in the first place, the use case of this paper  is not a very convincing one in my opinion. The problem of verifying neural networks in a robotic setting has a more meaningful specification.  Hence, i don't think that this paper in itself will be interesting to the general theme of the conference. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Verification of neural networks for computer vision tasks is not well defined. Hence, analyzing attack models on the verification engines is not well grounded.", "review": "In the recent literature there has been a rise in the number of papers which attempt to verify neural networks. The specification of the verification problems often gets adapted according to the application in mind. More specifically, for image classification networks, the problem is to prove that the output of the neural network does not flip for small perturbations to the pixel values. For a robotic setting, the problem is often safety and convergence to some goal state. Where the neural network operates in closed loop with the system dynamics. \n\nThe authors in this paper present an adversarial attack model on neural networks, which is deemed correct by some verifier. More specifically , given a neural network which can be shown to be robust to adversarial perturbations around some input, the authors exploit numerical errors in the computations to attack the network. Demonstrating the presence of loop holes in the proving engines itself. This is due to the approximation errors introduced by using floating point numbers.\n\nIn my opinion, the notion of input sets in the space of images, is not a very useful one. Mainly because the interval valued sets representing  perturbations of the input image, is far removed from the intended specification.  It's a step in the right direction, if the verification of computer vision task was a well defined problem. Since it's not clear what to verify in the first place, the use case of this paper  is not a very convincing one in my opinion. The problem of verifying neural networks in a robotic setting has a more meaningful specification.  Hence, i don't think that this paper in itself will be interesting to the general theme of the conference. ", "rating": "3: Clear rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603658857538}], "openreview_url": "https://openreview.net/forum?id=bVzUDC_4ls", "arxiv_id": "2003.03021", "paper_pdf": "papers/bVzUDC_4ls.pdf", "paper_pdf_sha256": "847acf7563e5ed238a4031f77096ca39ad7b86d698ee3517d0572851eac092d3", "paper_pdf_bytes": 411648, "paper_pdf_source": "openreview", "code_url": "https://github.com/jia-kai/realadv", "code_repository": "jia-kai/realadv", "code_commit": "f72d119b023deee83fe06d11b22adecdc104253c", "code_archive": "repos/bVzUDC_4ls.zip", "code_archive_sha256": "dfb60e8a7aac38e4f8277b6a8d4ac54d520fd2bc003f7bd0bd149c7b755445e3", "code_archive_bytes": 3381036, "code_file_count": 38, "code_extensions": {".py": 28, ".sh": 5, ".h": 2, ".jl": 2, ".cpp": 1}, "github_disk_usage_kb": 3146, "github_languages": {"Python": 156906, "C++": 8184, "Julia": 5982, "Shell": 2641, "Cython": 1584, "Makefile": 126}, "github_archived": false, "github_pushed_at": "2021-10-16T15:33:03Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/exploiting-verified-neural-networks-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mDEYl0Ucgr", "year": 2025, "status": "rejected", "title": "Influencing Humans to Conform to Preference Models for RLHF", "authors": ["Stephane Hatgis-Kessell", "W. Bradley Knox", "Serena Booth", "Scott Niekum", "Peter Stone"], "authorids": ["~Stephane_Hatgis-Kessell1", "~W._Bradley_Knox2", "~Serena_Booth1", "~Scott_Niekum1", "~Peter_Stone1"], "authors_source": "OpenReview API", "abstract": "Designing a reinforcement learning from human feedback (RLHF) algorithm for learning from preferences requires assuming a preference model, sometimes implicitly.  A preference model that poorly describes how humans generate preferences risks learning a poor approximation of the human’s unobservable reward function. In this paper, we conduct three human studies to assess whether one can influence the expression of real human preferences to more closely conform to a desired preference model. Importantly, our approach does not seek to alter the human's unobserved reward function. Rather, we change how humans use this reward function to generate preferences, such that they better match whatever preference model is assumed by a particular RLHF algorithm. We introduce three interventions: showing humans the quantities that underlie a preference model, which is normally unobservable information derived from the reward function; training people to follow a specific preference model; and modifying the preference elicitation question. All intervention types show significant effects, providing practical tools to improve preference data quality and the resultant alignment of learned reward functions.Overall we establish a novel research direction in model alignment: training humans and designing interfaces to increase human conformance with the assumptions of the algorithm that will learn from their input.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "80ifRHpn7g", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12734/Reviewer_T9uU"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper focuses on reducing the gap between actual human behavior and the human models used in preference learning. While other work focuses on developing more complex human models, this submission focuses on changing human behavior to better conform to simple human models. The authors explore how to do this through a human study, focusing on three interventions. In the first intervention, they show annotators an explicit return or regret value for each partial trajectory in a comparison pair. In the second intervention, they train annotators to better estimate return or regret but then do not provide the actual return/regret values during annotation. In the third intervention, they simply change the language used to prompt the annotators. They find that the first two interventions lead to statistically significantly better agreement between human behavior and human models. The third also seems to increase agreement but is not statistically significant given the sample size.", "review_text": "This paper focuses on reducing the gap between actual human behavior and the human models used in preference learning. While other work focuses on developing more complex human models, this submission focuses on changing human behavior to better conform to simple human models. The authors explore how to do this through a human study, focusing on three interventions. In the first intervention, they show annotators an explicit return or regret value for each partial trajectory in a comparison pair. In the second intervention, they train annotators to better estimate return or regret but then do not provide the actual return/regret values during annotation. In the third intervention, they simply change the language used to prompt the annotators. They find that the first two interventions lead to statistically significantly better agreement between human behavior and human models. The third also seems to increase agreement but is not statistically significant given the sample size.", "strengths": "The paper is generally quite well motivated and written. The math and experiments are clearly described. I am relatively familiar with this area and do not know of much other work which has focused on making human behavior more similar to human models rather than human models more similar to human behavior, so I think it is a novel direction. The experiments seem to be carefully designed and executed, and it seems helpful to the preference learning community to have well-grounded data on how well human models fit annotations.", "weaknesses": "The main weakness I see in the paper is that the first two proposed interventions do not actually seem to be applicable in practice. While the authors concede that the \"privileged\" setting is impractical, the second setting—\"trained\"—also seems to be impossible in practice. This is because the trained setting requires teaching people how to evaluate a particular aggregation of reward, which relies knowing the reward function in the first place; however, the entire point of preference learning is that the reward function is *unknown*. The third setting, \"question,\" seems to be most applicable, but has the least convincing evidence.\n\nIn general, the authors argue that their interventions are focused on \"training subjects to follow a preference model.\" However, it seems that to a large extent the authors are training subjects to follow a particular reward function—the privileged and trained experiments are both focused on helping annotators better estimate rewards. It would be more convincing that the authors are helping subjects \"follow a preference model\" if, for example, the subjects were trained to estimate return/regret with one reward function and tested with a different reward function. This would help disentangle the effects of simply learning more about the reward function and actually learning to follow one of the preference models.\n\nSmaller issues/suggestions:\n * It would be helpful to include error bars in figures 5, 7, and 9.\n * There are a couple of improvements to the LaTeX math notation that could be made. For example, use `\\text{loss}` instead of `loss` and `\\text{regret}` instead of `regret`; use `\\log` instead of `log`; and use `\\mid` instead of `|`; use `\\left` and `\\right` with parentheses to make it clearer which parentheses match.\n * Line 345: \"signficant\" -> \"significant\"\n * Line 471: this sentence is confusing because it's referring to the experiments from Section 5.2. I think maybe it was meant to be replaced with links to the experiments in Section 5.3.", "questions": "* How could one actually implement the \"trained\" intervention in practice for an unknown reward function?\n* How can we disentangle how much of the improvement in matching the human models is due to just learning the reward function better versus following the preference model?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on reducing the gap between actual human behavior and the human models used in preference learning. While other work focuses on developing more complex human models, this submission focuses on changing human behavior to better conform to simple human models. The authors explore how to do this through a human study, focusing on three interventions. In the first intervention, they show annotators an explicit return or regret value for each partial trajectory in a comparison pair. In the second intervention, they train annotators to better estimate return or regret but then do not provide the actual return/regret values during annotation. In the third intervention, they simply change the language used to prompt the annotators. They find that the first two interventions lead to statistically significantly better agreement between human behavior and human models. The third also seems to increase agreement but is not statistically significant given the sample size.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The paper is generally quite well motivated and written. The math and experiments are clearly described. I am relatively familiar with this area and do not know of much other work which has focused on making human behavior more similar to human models rather than human models more similar to human behavior, so I think it is a novel direction. The experiments seem to be carefully designed and executed, and it seems helpful to the preference learning community to have well-grounded data on how well human models fit annotations.", "weaknesses": "The main weakness I see in the paper is that the first two proposed interventions do not actually seem to be applicable in practice. While the authors concede that the \"privileged\" setting is impractical, the second setting—\"trained\"—also seems to be impossible in practice. This is because the trained setting requires teaching people how to evaluate a particular aggregation of reward, which relies knowing the reward function in the first place; however, the entire point of preference learning is that the reward function is *unknown*. The third setting, \"question,\" seems to be most applicable, but has the least convincing evidence.\n\nIn general, the authors argue that their interventions are focused on \"training subjects to follow a preference model.\" However, it seems that to a large extent the authors are training subjects to follow a particular reward function—the privileged and trained experiments are both focused on helping annotators better estimate rewards. It would be more convincing that the authors are helping subjects \"follow a preference model\" if, for example, the subjects were trained to estimate return/regret with one reward function and tested with a different reward function. This would help disentangle the effects of simply learning more about the reward function and actually learning to follow one of the preference models.\n\nSmaller issues/suggestions:\n * It would be helpful to include error bars in figures 5, 7, and 9.\n * There are a couple of improvements to the LaTeX math notation that could be made. For example, use `\\text{loss}` instead of `loss` and `\\text{regret}` instead of `regret`; use `\\log` instead of `log`; and use `\\mid` instead of `|`; use `\\left` and `\\right` with parentheses to make it clearer which parentheses match.\n * Line 345: \"signficant\" -> \"significant\"\n * Line 471: this sentence is confusing because it's referring to the experiments from Section 5.2. I think maybe it was meant to be replaced with links to the experiments in Section 5.3.", "questions": "* How could one actually implement the \"trained\" intervention in practice for an unknown reward function?\n* How can we disentangle how much of the improvement in matching the human models is due to just learning the reward function better versus following the preference model?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730687216398}, {"id": "eoXlLNcbZL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12734/Reviewer_jTpL"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper looks to test the idea that we may be able to influence people to behave more closely to target preference models. The motivation for this is twofold. (1) There remains a modelling gap in modelling human preferences, which makes learning underlying reward functions from those preferences harder and less accurate. This gap could be closed by improving said models, but — as this paper points out — it could equally be closed by making people behave more like the models we have now. (2) Even if perfect models of human preference existed, humans could still be influenced to behave like models on which inference is more tractable.\n\nTo this end, three interventions are proposed that could achieve this. Through three user studies on a grid world RL problem, it is tested for each intervention whether the intervention indeed influences people to give preferences that more closely align with a target preference model, and whether inference with the target model leads to more aligned inferred reward functions. The interventions are tested with two target preference models: one that assumes people prefer minimal regret behavior, and one that assumes people prefer maximal partial return behavior. The interventions are:\n- “Privileged”, where the intervention consists of showing participants the partial return or regret of the pair of trajectories between which they must choose. This requires knowledge of the true reward function.\n- “Trained”, where participants are trained through instruction to choose segments that maximise partial return or partial regret, before their preferences are elicited.\n- “Question”, where how the text instruction that tells participants how to choose between the two trajectories is changed to better reflect partial return or regret comparisons.\n\nExperimentally, the papers generally finds significant effects under all three interventions in terms of making people behave more like the target model, and in inferring a more aligned reward function.", "review_text": "This paper looks to test the idea that we may be able to influence people to behave more closely to target preference models. The motivation for this is twofold. (1) There remains a modelling gap in modelling human preferences, which makes learning underlying reward functions from those preferences harder and less accurate. This gap could be closed by improving said models, but — as this paper points out — it could equally be closed by making people behave more like the models we have now. (2) Even if perfect models of human preference existed, humans could still be influenced to behave like models on which inference is more tractable.\n\nTo this end, three interventions are proposed that could achieve this. Through three user studies on a grid world RL problem, it is tested for each intervention whether the intervention indeed influences people to give preferences that more closely align with a target preference model, and whether inference with the target model leads to more aligned inferred reward functions. The interventions are tested with two target preference models: one that assumes people prefer minimal regret behavior, and one that assumes people prefer maximal partial return behavior. The interventions are:\n- “Privileged”, where the intervention consists of showing participants the partial return or regret of the pair of trajectories between which they must choose. This requires knowledge of the true reward function.\n- “Trained”, where participants are trained through instruction to choose segments that maximise partial return or partial regret, before their preferences are elicited.\n- “Question”, where how the text instruction that tells participants how to choose between the two trajectories is changed to better reflect partial return or regret comparisons.\n\nExperimentally, the papers generally finds significant effects under all three interventions in terms of making people behave more like the target model, and in inferring a more aligned reward function.", "strengths": "I like the angle this paper takes very much. Human preference in settings such as the one considered here is very much an open problem. We only have limited understanding of how people value the options presented to them (see Knox et al. 2022) and limited understanding of how they would choose a most preferred option based on these values. With the (renewed) importance of preference learning, this is an important problem that remains unsolved in ML and even in cognitive science. The proposed solution of making people behave more like the simple preference models we have now is quite creative, and could go some way towards closing this gap, even if we must still recognise that people have inherent cognitive limitations that cannot be trained/influenced out.\n\nThe experiments are well-documented and the user study as described has been carried out to a level that is commensurate with what is expected at a top-level ML conference. The writing is clear and well-structured and the paper is easy to follow.", "weaknesses": "Primarily, I’m worried that the experiments carried out in the paper do not fully support the main claim of the paper. The stated goal is to test whether we can influence humans to follow a specific preference model. To test this, the authors check whether influencing someone towards a target preference model makes them behave more like that preference model compared to when they are not influenced (control condition). However, it is not established whether the interventions actually influence people to behave more like the target preference model as opposed to other preference models, or whether the intervention has solely influenced them to make better comparisons.\n\nThere seems to be some evidence of this in Figures 6 and 10, which show that  influencing participants to follow either one of the preference models leads to better reward inference with both, not only the targeted one. Furthermore, Figures 8 and 27, show that training people to follow a regret preference model is most effective for inferring their reward function, be that done with regret or partial return preference models. It seems to me then that although these interventions have a positive effect, this effect is not most specific to the target preference model.\n\nIn addition, the practical relevance of the proposed interventions is limited. Interventions in the “privileged” experiment are not practical because displaying true regret or true partial return is impossible without access to the unknown reward function (which the authors acknowledge). The intervention in the “trained” experiment is to train people to follow regret or partial return preferences. However, within the experiment presented here that training relies heavily on the true reward function (see questions). It is not clear how such training would work if the true reward function is not known. I see no practicality issues with the “question” experiment, but the effect size here is the most limited out of the three.", "questions": "Is training practical when the correct reward function is not known? In the experiment presented here, the true reward function was used in training, e.g. participants were corrected when they made wrong choices by saying something along the lines of “wrong, both items have equal score so far, but right option has higher possible score increase”. How would this training be implemented if the true reward function was not known?\n\nFor all three experiments, what was the mean Cross-Entropy Loss for the opposite preference model? For example, after a participant was trained to follow a partial return preference model, what was the likelihood of a regret preference model on that participant’s elicited answers?\n\nFor the question experiment, would participants have understood the difference between the question “which path has better immediate outcomes” and “which path reflects better decision-making”? The second questions seems ambiguous to me, considering the decision-making could be good in both the short or long term.\n\nSection D.5 in the appendix suggests that people generally are better aligned with regret-based preference models. This is also in line with the work of Knox et al. 2022. Could this natural alignment towards regret-based preference have made training people to follow partial return preference less effective?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper looks to test the idea that we may be able to influence people to behave more closely to target preference models. The motivation for this is twofold. (1) There remains a modelling gap in modelling human preferences, which makes learning underlying reward functions from those preferences harder and less accurate. This gap could be closed by improving said models, but — as this paper points out — it could equally be closed by making people behave more like the models we have now. (2) Even if perfect models of human preference existed, humans could still be influenced to behave like models on which inference is more tractable.\n\nTo this end, three interventions are proposed that could achieve this. Through three user studies on a grid world RL problem, it is tested for each intervention whether the intervention indeed influences people to give preferences that more closely align with a target preference model, and whether inference with the target model leads to more aligned inferred reward functions. The interventions are tested with two target preference models: one that assumes people prefer minimal regret behavior, and one that assumes people prefer maximal partial return behavior. The interventions are:\n- “Privileged”, where the intervention consists of showing participants the partial return or regret of the pair of trajectories between which they must choose. This requires knowledge of the true reward function.\n- “Trained”, where participants are trained through instruction to choose segments that maximise partial return or partial regret, before their preferences are elicited.\n- “Question”, where how the text instruction that tells participants how to choose between the two trajectories is changed to better reflect partial return or regret comparisons.\n\nExperimentally, the papers generally finds significant effects under all three interventions in terms of making people behave more like the target model, and in inferring a more aligned reward function.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "I like the angle this paper takes very much. Human preference in settings such as the one considered here is very much an open problem. We only have limited understanding of how people value the options presented to them (see Knox et al. 2022) and limited understanding of how they would choose a most preferred option based on these values. With the (renewed) importance of preference learning, this is an important problem that remains unsolved in ML and even in cognitive science. The proposed solution of making people behave more like the simple preference models we have now is quite creative, and could go some way towards closing this gap, even if we must still recognise that people have inherent cognitive limitations that cannot be trained/influenced out.\n\nThe experiments are well-documented and the user study as described has been carried out to a level that is commensurate with what is expected at a top-level ML conference. The writing is clear and well-structured and the paper is easy to follow.", "weaknesses": "Primarily, I’m worried that the experiments carried out in the paper do not fully support the main claim of the paper. The stated goal is to test whether we can influence humans to follow a specific preference model. To test this, the authors check whether influencing someone towards a target preference model makes them behave more like that preference model compared to when they are not influenced (control condition). However, it is not established whether the interventions actually influence people to behave more like the target preference model as opposed to other preference models, or whether the intervention has solely influenced them to make better comparisons.\n\nThere seems to be some evidence of this in Figures 6 and 10, which show that  influencing participants to follow either one of the preference models leads to better reward inference with both, not only the targeted one. Furthermore, Figures 8 and 27, show that training people to follow a regret preference model is most effective for inferring their reward function, be that done with regret or partial return preference models. It seems to me then that although these interventions have a positive effect, this effect is not most specific to the target preference model.\n\nIn addition, the practical relevance of the proposed interventions is limited. Interventions in the “privileged” experiment are not practical because displaying true regret or true partial return is impossible without access to the unknown reward function (which the authors acknowledge). The intervention in the “trained” experiment is to train people to follow regret or partial return preferences. However, within the experiment presented here that training relies heavily on the true reward function (see questions). It is not clear how such training would work if the true reward function is not known. I see no practicality issues with the “question” experiment, but the effect size here is the most limited out of the three.", "questions": "Is training practical when the correct reward function is not known? In the experiment presented here, the true reward function was used in training, e.g. participants were corrected when they made wrong choices by saying something along the lines of “wrong, both items have equal score so far, but right option has higher possible score increase”. How would this training be implemented if the true reward function was not known?\n\nFor all three experiments, what was the mean Cross-Entropy Loss for the opposite preference model? For example, after a participant was trained to follow a partial return preference model, what was the likelihood of a regret preference model on that participant’s elicited answers?\n\nFor the question experiment, would participants have understood the difference between the question “which path has better immediate outcomes” and “which path reflects better decision-making”? The second questions seems ambiguous to me, considering the decision-making could be good in both the short or long term.\n\nSection D.5 in the appendix suggests that people generally are better aligned with regret-based preference models. This is also in line with the work of Knox et al. 2022. Could this natural alignment towards regret-based preference have made training people to follow partial return preference less effective?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730576482386}, {"id": "lfB20nYKr3", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12734/Reviewer_gCy4"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 2, "summary": "This paper explores methods to improve the quality of human feedback in reinforcement learning from human feedback (RLHF) by aligning human preference expressions with the preference model assumed by the RLHF algorithm. Poorly aligned preferences can lead to inaccurate approximations of a human's reward function. Through three human studies, the authors assess interventions to influence how humans express preferences without altering their underlying reward function. These interventions include displaying model-based preference quantities, training people to follow specific models, and rephrasing preference questions. All three interventions significantly improve preference alignment, introducing a novel direction in model alignment by adjusting human interaction to better match algorithmic assumptions.", "review_text": "This paper explores methods to improve the quality of human feedback in reinforcement learning from human feedback (RLHF) by aligning human preference expressions with the preference model assumed by the RLHF algorithm. Poorly aligned preferences can lead to inaccurate approximations of a human's reward function. Through three human studies, the authors assess interventions to influence how humans express preferences without altering their underlying reward function. These interventions include displaying model-based preference quantities, training people to follow specific models, and rephrasing preference questions. All three interventions significantly improve preference alignment, introducing a novel direction in model alignment by adjusting human interaction to better match algorithmic assumptions.", "strengths": "Originality: Very original work - First one I've seen on aligning human preference expressions with the assumptions of the preference model, an interesting and out-of-the-box perspective.\n\nQuality: Good variety of experiments to test the hypotheses, covering multiple different ways to influence human preference expressions.\n\nClarity: Very clear introduction, related work, description of tasks, and description of experimental interventions\n\nSignificance: Interesting and novel research direction which could lead to significant improvements in RLHF", "weaknesses": "Typos: \n- Figure 5 caption: \"the the\"\n- Line 381: \"Priveleged\"\n\nI feel like the \"% of partitions in which performance is near optimal\" results are not very convincing to support the hypotheses presented.\n\nPRIVILEGED EXPERIMENT\nFigure 6: The difference between the P_regret and P_reward lines in both subplots is very minimal and does not hold for all number of partitions.\n\nTRAINED EXPERIMENT\nFigure 8: The P_regret trained preferences always lead to better performance -- this seems to indicate that the P_regret training paradigm leads to better understanding of the ground-truth reward function more than it supports Hypothesis 2.\n\nQUESTION EXPERIMENT\nThe P_regret question does not seem to me to focus subjects on a segment's deviation from optimality. Perhaps something like \"Which path is farthest from optimal?\" would have been better\nFigure 10: The difference between the P_regret and P_reward lines in both subplots is very minimal and does not hold for all number of partitions.", "questions": "Please address the weaknesses listed above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores methods to improve the quality of human feedback in reinforcement learning from human feedback (RLHF) by aligning human preference expressions with the preference model assumed by the RLHF algorithm. Poorly aligned preferences can lead to inaccurate approximations of a human's reward function. Through three human studies, the authors assess interventions to influence how humans express preferences without altering their underlying reward function. These interventions include displaying model-based preference quantities, training people to follow specific models, and rephrasing preference questions. All three interventions significantly improve preference alignment, introducing a novel direction in model alignment by adjusting human interaction to better match algorithmic assumptions.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "Originality: Very original work - First one I've seen on aligning human preference expressions with the assumptions of the preference model, an interesting and out-of-the-box perspective.\n\nQuality: Good variety of experiments to test the hypotheses, covering multiple different ways to influence human preference expressions.\n\nClarity: Very clear introduction, related work, description of tasks, and description of experimental interventions\n\nSignificance: Interesting and novel research direction which could lead to significant improvements in RLHF", "weaknesses": "Typos: \n- Figure 5 caption: \"the the\"\n- Line 381: \"Priveleged\"\n\nI feel like the \"% of partitions in which performance is near optimal\" results are not very convincing to support the hypotheses presented.\n\nPRIVILEGED EXPERIMENT\nFigure 6: The difference between the P_regret and P_reward lines in both subplots is very minimal and does not hold for all number of partitions.\n\nTRAINED EXPERIMENT\nFigure 8: The P_regret trained preferences always lead to better performance -- this seems to indicate that the P_regret training paradigm leads to better understanding of the ground-truth reward function more than it supports Hypothesis 2.\n\nQUESTION EXPERIMENT\nThe P_regret question does not seem to me to focus subjects on a segment's deviation from optimality. Perhaps something like \"Which path is farthest from optimal?\" would have been better\nFigure 10: The difference between the P_regret and P_reward lines in both subplots is very minimal and does not hold for all number of partitions.", "questions": "Please address the weaknesses listed above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730344084194}, {"id": "Nz6O7oizgq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12734/Reviewer_6yvA"], "rating": 5, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 4, "summary": "The paper investigates methods to align human preferences with a specified preference model in RLHF by influencing how humans express preferences. This approach doesn't alter the underlying human reward function but instead modifies how preferences are generated to match an RLHF algorithm’s assumed model. Three interventions were tested across human studies: Privileged Experiment, Trained Experiment, and Question Experiment. The paper claims that each method effectively improved alignment with the target model.", "review_text": "The paper investigates methods to align human preferences with a specified preference model in RLHF by influencing how humans express preferences. This approach doesn't alter the underlying human reward function but instead modifies how preferences are generated to match an RLHF algorithm’s assumed model. Three interventions were tested across human studies: Privileged Experiment, Trained Experiment, and Question Experiment. The paper claims that each method effectively improved alignment with the target model.", "strengths": "- The paper is generally well-written and clear.\n- The experimental setup is described in detail, providing clarity on the methodology.\n- Thorough analysis of the experiments strengthens the overall findings.", "weaknesses": "- RLHF’s value in aligning language models is widely recognized, but the paper’s focus on a constrained grid-world delivery domain may limit generalization to language-based agents.\n    * While the proposed interventions show promise, it’s unclear how effectively these ideas could extend to more complex language-driven settings.\n- The environment and experiment setting closely follow prior work by Knox et al., meaning the primary contribution arises from the three interventions aimed at influencing human decision-making: the Privileged Experiment, Trained Experiment, and Question Experiment. However, concerns remain about both the novelty and practicality of these experiments:\n    * The Trained Experiment, which involves training participants before they label data, is resource-intensive, with significant demands on time and cost, as noted by the authors.\n    * The Question Experiment has only a marginal effect, with no significant impact observed for aligning preferences with the regret model, which may limit its utility in practice.", "questions": "- Does influencing preferences toward a specific model potentially introduce bias in scenarios where the model may not be the best fit for human objectives?\n- In the Privileged Experiment, given that users see reward information directly, is there a risk they might over-rely on this data, reducing their engagement with the task?\n- Could the interventions used in the study generalize to more complex RLHF tasks (e.g., continuous environments) where users make dynamic, long-term decisions?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates methods to align human preferences with a specified preference model in RLHF by influencing how humans express preferences. This approach doesn't alter the underlying human reward function but instead modifies how preferences are generated to match an RLHF algorithm’s assumed model. Three interventions were tested across human studies: Privileged Experiment, Trained Experiment, and Question Experiment. The paper claims that each method effectively improved alignment with the target model.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "- The paper is generally well-written and clear.\n- The experimental setup is described in detail, providing clarity on the methodology.\n- Thorough analysis of the experiments strengthens the overall findings.", "weaknesses": "- RLHF’s value in aligning language models is widely recognized, but the paper’s focus on a constrained grid-world delivery domain may limit generalization to language-based agents.\n    * While the proposed interventions show promise, it’s unclear how effectively these ideas could extend to more complex language-driven settings.\n- The environment and experiment setting closely follow prior work by Knox et al., meaning the primary contribution arises from the three interventions aimed at influencing human decision-making: the Privileged Experiment, Trained Experiment, and Question Experiment. However, concerns remain about both the novelty and practicality of these experiments:\n    * The Trained Experiment, which involves training participants before they label data, is resource-intensive, with significant demands on time and cost, as noted by the authors.\n    * The Question Experiment has only a marginal effect, with no significant impact observed for aligning preferences with the regret model, which may limit its utility in practice.", "questions": "- Does influencing preferences toward a specific model potentially introduce bias in scenarios where the model may not be the best fit for human objectives?\n- In the Privileged Experiment, given that users see reward information directly, is there a risk they might over-rely on this data, reducing their engagement with the task?\n- Could the interventions used in the study generalize to more complex RLHF tasks (e.g., continuous environments) where users make dynamic, long-term decisions?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730151510165}], "openreview_url": "https://openreview.net/forum?id=mDEYl0Ucgr", "arxiv_id": "2501.06416", "paper_pdf": "papers/mDEYl0Ucgr.pdf", "paper_pdf_sha256": "604b567c1eb9a349480d52bd84748c6223e961da6757b89cb9066528ff9d4c68", "paper_pdf_bytes": 4269937, "paper_pdf_source": "openreview", "code_url": "https://github.com/Stephanehk/InfluencingHumanPrefs", "code_repository": "Stephanehk/InfluencingHumanPrefs", "code_commit": "54ad2720c1d5638574fc5d788a53008aef2c1c12", "code_archive": "repos/mDEYl0Ucgr.zip", "code_archive_sha256": "d8288df4964091f57cabcce041c6af1cde1e0219d53c01654584bcef7248a4ce", "code_archive_bytes": 480186, "code_file_count": 47, "code_extensions": {".py": 44, ".sh": 3}, "github_disk_usage_kb": 405, "github_languages": {"Python": 319533, "Shell": 1124}, "github_archived": false, "github_pushed_at": "2024-12-30T06:25:37Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/influencing-humans-to-conform-to-preference"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "idpV2AqusC", "year": 2024, "status": "rejected", "title": "Improving SAM Requires Rethinking its Optimization Formulation", "authors": ["Wanyun Xie", "Fabian Latorre", "Kimon Antonakopoulos", "Thomas Pethick", "Volkan Cevher"], "authorids": ["~Wanyun_Xie1", "~Fabian_Latorre1", "~Kimon_Antonakopoulos1", "~Thomas_Pethick1", "~Volkan_Cevher1"], "authors_source": "OpenReview API", "abstract": "This paper rethinks Sharpness-Aware Minimization (SAM), which is originally formulated as a zero-sum game where the weights of a network and a bounded perturbation try to minimize/maximize, respectively, the same differentiable loss. We argue that SAM should instead be reformulated using the 0-1 loss, as this provides a tighter bound on its generalization gap. As a continuous relaxation, we follow the simple conventional approach where the minimizing (maximizing) player uses an upper bound (lower bound) surrogate to the 0-1 loss. This leads to a novel formulation of SAM as a bilevel optimization problem, dubbed as BiSAM. Through numerical evidence, we show that BiSAM consistently results in improved performance when compared to the original SAM and variants, while enjoying similar computational complexity.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "eXL46oWzHD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9188/Reviewer_duHp"], "rating": "10: strong accept, should be highlighted at the conference", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper investigates applying SAM to the actual metric (0-1 loss) in the classification problem and proposes reformulating the original zero-sum game in SAM as a bilevel optimization problem. The bilevel optimization formulation is motivated by the relation between cross-entropy and 0-1 loss. To counter the non-differentiability of the 0-1 loss, the proposed method (BiSAM) introduces a surrogate loss coupled with the cross-entropy loss. The empirical results show consistent improvement over SAM and orthogonal improvements when accompanied with other SAM variants (ASAM and ESAM).", "review_text": "The paper investigates applying SAM to the actual metric (0-1 loss) in the classification problem and proposes reformulating the original zero-sum game in SAM as a bilevel optimization problem. The bilevel optimization formulation is motivated by the relation between cross-entropy and 0-1 loss. To counter the non-differentiability of the 0-1 loss, the proposed method (BiSAM) introduces a surrogate loss coupled with the cross-entropy loss. The empirical results show consistent improvement over SAM and orthogonal improvements when accompanied with other SAM variants (ASAM and ESAM).", "strengths": "- The approach is simple, scalable, and theoretical-sound\n- The flow is easy to follow\n- The improvements are convincing and validated in many learning scenarios, including standard learning, fine-tuning and noisy-data learning", "weaknesses": "- As mentioned in the conclusion, it will be great to see if BiSAM benefits other domains, e.g. NLP", "questions": "- In A.3, to further ground BiSAM's improvements under noisy labels. Can the authors provide the number with noisy learning approaches, e.g., mixup?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates applying SAM to the actual metric (0-1 loss) in the classification problem and proposes reformulating the original zero-sum game in SAM as a bilevel optimization problem. The bilevel optimization formulation is motivated by the relation between cross-entropy and 0-1 loss. To counter the non-differentiability of the 0-1 loss, the proposed method (BiSAM) introduces a surrogate loss coupled with the cross-entropy loss. The empirical results show consistent improvement over SAM and orthogonal improvements when accompanied with other SAM variants (ASAM and ESAM).", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "4 excellent", "strengths": "- The approach is simple, scalable, and theoretical-sound\n- The flow is easy to follow\n- The improvements are convincing and validated in many learning scenarios, including standard learning, fine-tuning and noisy-data learning", "weaknesses": "- As mentioned in the conclusion, it will be great to see if BiSAM benefits other domains, e.g. NLP", "questions": "- In A.3, to further ground BiSAM's improvements under noisy labels. Can the authors provide the number with noisy learning approaches, e.g., mixup?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "10: strong accept, should be highlighted at the conference", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698850698640}, {"id": "iLmH7rsdR9", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9188/Reviewer_idNJ"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors propose BiSAM, an algorithm for solving the Sharpness Aware Minimization (SAM). The original SAM algorithm (Foret et al., 2021) aims to solve a min-max problem of a differentiable loss that upper bounds the 0-1 loss. The authors suggest that directly considering the min-max problem of the 0-1 loss provides a tighter generalization bound. Accordingly, the BiSAM algorithm aims to solve the problem by substituting the 0-1 loss with a differentiable upper bound for minimization and a differentiable lower bound for maximization. The experimental results demonstrate the benefit of BiSAM compared with the original SAM algorithm.", "review_text": "The authors propose BiSAM, an algorithm for solving the Sharpness Aware Minimization (SAM). The original SAM algorithm (Foret et al., 2021) aims to solve a min-max problem of a differentiable loss that upper bounds the 0-1 loss. The authors suggest that directly considering the min-max problem of the 0-1 loss provides a tighter generalization bound. Accordingly, the BiSAM algorithm aims to solve the problem by substituting the 0-1 loss with a differentiable upper bound for minimization and a differentiable lower bound for maximization. The experimental results demonstrate the benefit of BiSAM compared with the original SAM algorithm.", "strengths": "- The idea of directly aiming to solve min-max of 0-1 loss and accordingly minimizing/maximizing different surrogates brings novelty.\n- The authors provide theoretically justified lower bound for practical implementation. They also provide a clear discussion on two different choices of surrogates.\n- The numerical results demonstrate that BiSAM improves accuracy.", "weaknesses": "- The numerical results show limited improvements. Also, in some other works (Foret et al., 2021; Liu et al., 2022), SAM achieves accuracy higher than the accuracy of BiSAM in this paper (with the same model and number of epochs).  \n\nLiu, Y., Mai, S., Cheng, M., Chen, X., Hsieh, C. J., & You, Y. (2022). Random sharpness-aware minimization. Advances in Neural Information Processing Systems, 35, 24543-24556.", "questions": "1. Please see the weakness section.\n2. The original SAM aims to solve $ \\min_w \\max_{\\epsilon: \\Vert \\epsilon \\Vert_2 \\leq \\rho} L^{\\mathrm{ce}}_S (w + \\epsilon) + h(\\Vert w\\Vert^2_2 / \\rho)$, whereas BiSAM aims to solve (15).\n\n    Is there any theoretical result that could compare the gaps between $\\min_w L^{01}_\\mathcal{D} (w)$ and the two different solutions (the aim of SAM and the aim of BiSAM, respectively)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose BiSAM, an algorithm for solving the Sharpness Aware Minimization (SAM). The original SAM algorithm (Foret et al., 2021) aims to solve a min-max problem of a differentiable loss that upper bounds the 0-1 loss. The authors suggest that directly considering the min-max problem of the 0-1 loss provides a tighter generalization bound. Accordingly, the BiSAM algorithm aims to solve the problem by substituting the 0-1 loss with a differentiable upper bound for minimization and a differentiable lower bound for maximization. The experimental results demonstrate the benefit of BiSAM compared with the original SAM algorithm.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "- The idea of directly aiming to solve min-max of 0-1 loss and accordingly minimizing/maximizing different surrogates brings novelty.\n- The authors provide theoretically justified lower bound for practical implementation. They also provide a clear discussion on two different choices of surrogates.\n- The numerical results demonstrate that BiSAM improves accuracy.", "weaknesses": "- The numerical results show limited improvements. Also, in some other works (Foret et al., 2021; Liu et al., 2022), SAM achieves accuracy higher than the accuracy of BiSAM in this paper (with the same model and number of epochs).  \n\nLiu, Y., Mai, S., Cheng, M., Chen, X., Hsieh, C. J., & You, Y. (2022). Random sharpness-aware minimization. Advances in Neural Information Processing Systems, 35, 24543-24556.", "questions": "1. Please see the weakness section.\n2. The original SAM aims to solve $ \\min_w \\max_{\\epsilon: \\Vert \\epsilon \\Vert_2 \\leq \\rho} L^{\\mathrm{ce}}_S (w + \\epsilon) + h(\\Vert w\\Vert^2_2 / \\rho)$, whereas BiSAM aims to solve (15).\n\n    Is there any theoretical result that could compare the gaps between $\\min_w L^{01}_\\mathcal{D} (w)$ and the two different solutions (the aim of SAM and the aim of BiSAM, respectively)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698818537034}, {"id": "6qKRjrQvXQ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9188/Reviewer_Z9xG"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The article 'Rethinks Sharpness-Aware Minimization' addresses the issue of 0-1 loss being non-continuous and thus not amenable to gradient-based optimization. This is tackled by having the minimizing (maximizing) player use an upper bound (lower bound) surrogate for the 0-1 loss. The problem is analyzed thoroughly and the task is clearly defined.", "review_text": "The article 'Rethinks Sharpness-Aware Minimization' addresses the issue of 0-1 loss being non-continuous and thus not amenable to gradient-based optimization. This is tackled by having the minimizing (maximizing) player use an upper bound (lower bound) surrogate for the 0-1 loss. The problem is analyzed thoroughly and the task is clearly defined.", "strengths": "The BiSAM method proposed in the paper somewhat resolves the issue of optimizing the 0-1 loss using gradients. \nThis method has been validated across multiple datasets, demonstrating its advantages over SAM through extensive experiments.", "weaknesses": "1. \"The idea of BiSAM is very good, but its performance in experiments is only marginally better than SAM. The improvement over SAM is often within the range of error, making it hard to believe that it is an enhancement of SAM.\n\n2. Can you explain why BiSAM using tanh as the lower bound has higher test accuracy on CIFAR-10 compared to using -log as the lower bound, but the results are the opposite on CIFAR-100?\n\n3. Could you combine the characteristics of tanh and -log to create a new lower bound that is suitable for different datasets?\"", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The article 'Rethinks Sharpness-Aware Minimization' addresses the issue of 0-1 loss being non-continuous and thus not amenable to gradient-based optimization. This is tackled by having the minimizing (maximizing) player use an upper bound (lower bound) surrogate for the 0-1 loss. The problem is analyzed thoroughly and the task is clearly defined.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "3 good", "strengths": "The BiSAM method proposed in the paper somewhat resolves the issue of optimizing the 0-1 loss using gradients. \nThis method has been validated across multiple datasets, demonstrating its advantages over SAM through extensive experiments.", "weaknesses": "1. \"The idea of BiSAM is very good, but its performance in experiments is only marginally better than SAM. The improvement over SAM is often within the range of error, making it hard to believe that it is an enhancement of SAM.\n\n2. Can you explain why BiSAM using tanh as the lower bound has higher test accuracy on CIFAR-10 compared to using -log as the lower bound, but the results are the opposite on CIFAR-100?\n\n3. Could you combine the characteristics of tanh and -log to create a new lower bound that is suitable for different datasets?\"", "questions": "See weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698769825263}, {"id": "Yd54PG5cea", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9188/Reviewer_6ZVi"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Motivated by the min-max nature of sharpness-aware minimization (SAM), the paper develops a reformulation of SAM using bilevel optimization called BiSAM. This approach can be viewed as a continuous relaxation of SAM with 0-1 loss and the authors argue it provides a tighter bound on generalization gap. Numerical experiments are conducted to demonstrate the empirical improvement of BiSAM over original SAM.", "review_text": "Motivated by the min-max nature of sharpness-aware minimization (SAM), the paper develops a reformulation of SAM using bilevel optimization called BiSAM. This approach can be viewed as a continuous relaxation of SAM with 0-1 loss and the authors argue it provides a tighter bound on generalization gap. Numerical experiments are conducted to demonstrate the empirical improvement of BiSAM over original SAM.", "strengths": "The presentation of the work is clear to me. In particular, the motivation of using different bounds for min and max parts to handle the discrete 0-1 loss looks insightful to me and this naturally leads to the development of bilevel approaches. The numerical comparisons of BiSAM and SAM are conducted on various models to demonstrate the possible practical benefits of BiSAM. The authors also show that BiSAM can be easily embedded into variants of SAM such as adaptive SAM and efficient SAM.", "weaknesses": "- The significance of the contribution is unclear to me. According to the numerical experiments conducted, the improvement of BiSAM over original SAM seems quite minor. See Questions for more discussion on this point.\n- In turn, this makes me wonder if there is any real benefit of approximating 0-1 loss (via this bilevel method) instead of cross-entropy. The authors argue that directly building SAM over 0-1 loss provides tighter generalization bound but more discussions are required.", "questions": "Major comments:\n\n- The major concern is the significance of the contribution. When comparing the performance of BiSAM and SAM, it turns out the improvement in test accuracy is quite minor. Based on the standard deviation provided, it is hard to conclude that the improvements are due to randomness or indeed statistically significant. \n- The motivation behind BiSAM is the benefit of approximating 0-1 loss via BiSAM over continuous surrogate like cross-entropy. The intuition of lower bounding the max part is convincing but I wonder the actual benefit of doing so. In particular, could the authors justify the using of lower bound (11) provides tighter bound than using $L^{ce}$ in the original SAM? I think illustration by a simple synthetic example can help a lot, or through theoretical derivations.\n\n\nOther comments: \n\n- Is the term $-\\frac{1}{\\mu}\\log k$ missing in the right-hand-side of Equation (14)?\n- In the paragraph below Equation (8), what does this sentence mean: ``it would be wrong to simply replace it by cross-entropy\"? I think it is valid to upper bound the right-hand-side of (8) by $L^{ce}$. Is there anything I'm missing?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Motivated by the min-max nature of sharpness-aware minimization (SAM), the paper develops a reformulation of SAM using bilevel optimization called BiSAM. This approach can be viewed as a continuous relaxation of SAM with 0-1 loss and the authors argue it provides a tighter bound on generalization gap. Numerical experiments are conducted to demonstrate the empirical improvement of BiSAM over original SAM.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "The presentation of the work is clear to me. In particular, the motivation of using different bounds for min and max parts to handle the discrete 0-1 loss looks insightful to me and this naturally leads to the development of bilevel approaches. The numerical comparisons of BiSAM and SAM are conducted on various models to demonstrate the possible practical benefits of BiSAM. The authors also show that BiSAM can be easily embedded into variants of SAM such as adaptive SAM and efficient SAM.", "weaknesses": "- The significance of the contribution is unclear to me. According to the numerical experiments conducted, the improvement of BiSAM over original SAM seems quite minor. See Questions for more discussion on this point.\n- In turn, this makes me wonder if there is any real benefit of approximating 0-1 loss (via this bilevel method) instead of cross-entropy. The authors argue that directly building SAM over 0-1 loss provides tighter generalization bound but more discussions are required.", "questions": "Major comments:\n\n- The major concern is the significance of the contribution. When comparing the performance of BiSAM and SAM, it turns out the improvement in test accuracy is quite minor. Based on the standard deviation provided, it is hard to conclude that the improvements are due to randomness or indeed statistically significant. \n- The motivation behind BiSAM is the benefit of approximating 0-1 loss via BiSAM over continuous surrogate like cross-entropy. The intuition of lower bounding the max part is convincing but I wonder the actual benefit of doing so. In particular, could the authors justify the using of lower bound (11) provides tighter bound than using $L^{ce}$ in the original SAM? I think illustration by a simple synthetic example can help a lot, or through theoretical derivations.\n\n\nOther comments: \n\n- Is the term $-\\frac{1}{\\mu}\\log k$ missing in the right-hand-side of Equation (14)?\n- In the paragraph below Equation (8), what does this sentence mean: ``it would be wrong to simply replace it by cross-entropy\"? I think it is valid to upper bound the right-hand-side of (8) by $L^{ce}$. Is there anything I'm missing?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698712995499}], "openreview_url": "https://openreview.net/forum?id=idpV2AqusC", "arxiv_id": "2407.12993", "paper_pdf": "papers/idpV2AqusC.pdf", "paper_pdf_sha256": "043920848e62cae574048dcad12a84129687839c6117ce8fb755c2ff31f6f07a", "paper_pdf_bytes": 558257, "paper_pdf_source": "openreview", "code_url": "https://github.com/LIONS-EPFL/BiSAM", "code_repository": "LIONS-EPFL/BiSAM", "code_commit": "c6595703430d062f36d9f111c10d7be44b0fb22e", "code_archive": "repos/idpV2AqusC.zip", "code_archive_sha256": "ab1beb6d0b704f0ba877699e997800cab6e64a79a38143fe519cf2c0e3223753", "code_archive_bytes": 449352, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 442, "github_languages": {"Python": 41821}, "github_archived": false, "github_pushed_at": "2026-08-28T13:56:51Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/improving-sam-requires-rethinking-its"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "wysXxmukfCA", "year": 2023, "status": "rejected", "title": "Towards Robust Model Watermark via Reducing Parametric Vulnerability", "authors": ["Guanhao Gan", "Yiming Li", "Dongxian Wu", "Shu-Tao Xia"], "authorids": ["~Guanhao_Gan1", "~Yiming_Li1", "~Dongxian_Wu1", "~Shu-Tao_Xia1"], "authors_source": "OpenReview API", "abstract": "Deep neural networks are valuable assets considering their commercial benefits and huge demands for costly annotation and computation resources. To protect the copyright of these deep models, backdoor-based ownership verification becomes popular recently, in which the model owner can watermark the model by embedding a specific behavior before releasing it. The defender (usually the model owner) can identify whether a suspicious third-party model is ``stolen'' from it based on the presence of the behavior. Unfortunately, these watermarks are proven to be vulnerable to removal attacks even like fine-tuning. To further explore this vulnerability, we investigate the parametric space and find there exist many watermark-removed models in the vicinity of the watermarked one, which may be easily used by removal attacks. Inspired by this finding, we propose a minimax formulation to find these watermark-removed models and recover their watermark behavior. Extensive experiments demonstrate that our method improves the robustness of the model watermarking against parametric changes and numerous watermark-removal attacks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "nUj_4vS5E", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4373/Reviewer_6TWD"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes a method for watermarking deep neural networks for classification against adversarial attacks. The authors proposed a min-max formulation of the watermark defense problem and solved the problem through first-order approximations to the inner maximization problems. Experiment results show the method is robust towards various attacks while keeping a reasonable accuracy on regular samples. ", "review_text": "Overall, I recommend the acceptance of this paper. ", "strengths": "Strength: \nWatermarking neural networks is an important research problem. The \"black-box\" setting under which the authors studies the problem is realistic for practical applications. The proposed min-max formulation is clear and intuitive, and the experiment results support the claims that the method is effective in defending against adversarial attacks. \n\nWeakness:\nSome minor issues with writing and experiments. \n\n1. The BatchNorm on clean samples technique should be highlighted more in Algorithm 1 as it is critical to the method. \n2. Section 4.1 claims experiments were done on CIFAR-100, but I don't see any results on CIFAR-100. In particular, I am curious to see whether this method scales reasonably with respect to the number of class labels. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a method for watermarking deep neural networks for classification against adversarial attacks. The authors proposed a min-max formulation of the watermark defense problem and solved the problem through first-order approximations to the inner maximization problems. Experiment results show the method is robust towards various attacks while keeping a reasonable accuracy on regular samples. ", "strength_and_weaknesses": "Strength: \nWatermarking neural networks is an important research problem. The \"black-box\" setting under which the authors studies the problem is realistic for practical applications. The proposed min-max formulation is clear and intuitive, and the experiment results support the claims that the method is effective in defending against adversarial attacks. \n\nWeakness:\nSome minor issues with writing and experiments. \n\n1. The BatchNorm on clean samples technique should be highlighted more in Algorithm 1 as it is critical to the method. \n2. Section 4.1 claims experiments were done on CIFAR-100, but I don't see any results on CIFAR-100. In particular, I am curious to see whether this method scales reasonably with respect to the number of class labels. ", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and easy to follow. The proposed method is novel. ", "summary_of_the_review": "Overall, I recommend the acceptance of this paper. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666799819061}, {"id": "PsNtLvJCwWm", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4373/Reviewer_9HCi"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a minmax approach to improve the watermarked model's capacity to counter opponents' deceiving methods. This approach originates from the observation that many watermark-removed models exist around the vicinity of the watermarked one. ", "review_text": "According to the several drawbacks I mentioned in previous parts, I do not quite believe this paper is adequate enough for publication at ICLR. The paper can be better if the motivation goes more clear and the comparison is conducted with SOTA methods. ", "strengths": "This language reads well. And the proposed techniques seem effective from experimental results.\n\nHowever, there are several downsides:\n1) Motivation is neither clear nor reasonable. One assumption made by the authors is that the adversary has methods to obtain an unauthorized copy of the watermarked model. In practice, I don't think it is possible and if that happens, one should upgrade their security system instead of finding a method that can fight against different attacks. Besides, wouldn't it be more practical to release a second watermarked model once that happened? \n2) Comparison is only conducted on some baselines instead of SOTA approaches. It is mentioned in the related work (robust black-box model watermark) that several recent methods also tried to propose better methods to counter attacks. I wonder why the authors did not compare with them. Comparison with SOTA methods is desperately needed.\n3) The cBN technique is a trick, lacking enough technical novelty, though it is not listed as the main contribution. Then it makes no sense to spare a huge section explaining it, especially when more details are needed to elaborate the main approach better.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a minmax approach to improve the watermarked model's capacity to counter opponents' deceiving methods. This approach originates from the observation that many watermark-removed models exist around the vicinity of the watermarked one. ", "strength_and_weaknesses": "This language reads well. And the proposed techniques seem effective from experimental results.\n\nHowever, there are several downsides:\n1) Motivation is neither clear nor reasonable. One assumption made by the authors is that the adversary has methods to obtain an unauthorized copy of the watermarked model. In practice, I don't think it is possible and if that happens, one should upgrade their security system instead of finding a method that can fight against different attacks. Besides, wouldn't it be more practical to release a second watermarked model once that happened? \n2) Comparison is only conducted on some baselines instead of SOTA approaches. It is mentioned in the related work (robust black-box model watermark) that several recent methods also tried to propose better methods to counter attacks. I wonder why the authors did not compare with them. Comparison with SOTA methods is desperately needed.\n3) The cBN technique is a trick, lacking enough technical novelty, though it is not listed as the main contribution. Then it makes no sense to spare a huge section explaining it, especially when more details are needed to elaborate the main approach better.", "clarity,_quality,_novelty_and_reproducibility": "The method section can be better if provided with better figures and examples. The novelty of the main approach that formulates a minimax problem seems good but the other techniques used in this approach such as cBN pose questions that whether this approach is robust enough. Some details are missing to reimplement the approach but generally seems enough to have a picture of the computation graph.", "summary_of_the_review": "According to the several drawbacks I mentioned in previous parts, I do not quite believe this paper is adequate enough for publication at ICLR. The paper can be better if the motivation goes more clear and the comparison is conducted with SOTA methods. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["Yes, Privacy, security and safety", "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)", "Yes, Potentially harmful insights, methodologies and applications"], "details_of_ethics_concerns": "This paper works towards a better watermarked model. Adversaries can fight against this approach as well and do harm to users.", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666729246272}, {"id": "wHWs8dASc", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper4373/Reviewer_y4Xt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposed an adversarial training-based framework to ensure robustness in the watermark of deep neural networks. Specifically, a normalized gradient method is applied to find the worst case within the vicinity of the original model such that we can minimize the loss in terms of noisy parameters. A clean sample-based BatchNorm is also proposed to improve consistency during the training process. ", "review_text": "In general, the studied problem is interesting and important. In addition, the methodology is principled with three major merits as discussed above. However, the work still has some unaddressed concerns to well justify its technical and empirical contributions.", "strengths": "Pros:\n1.\tThis paper validated the vulnerability in the watermark of deep neural networks and shows that it is necessary to develop a robust watermark framework.\n2.\tThis paper solves the inconsistency issue during the training by introducing a clean-sample-based BatchNorm.\n3.\tExtensive experiments show that the proposed method achieves superior robustness of watermark verification.\n\nCons:\n1.\tI have concerns about the approximation of perturbation for parameters. Equation 3 aims to find the worst case of the delta that can maximize the loss over parameters.t I do not see the clues that equation 5 is an approximation form of equation 3. It looks like equation 5 randomly selects a point (depending on the epsilon) within the vicinity of the original model.\n2.\tOnly minimizing the worst case is not theoretically enough. The author should select multiple neighbor points in parameter space for each training iteration.\n3.\tThe ablation study is not enough. Figure 6 only shows results against FT attacks.\n4.\tThe paper controls the perturbation size with epsilon parameters. It is hard to ensure that the l norm of perturbation is within a fixed bound cause theta in equation 5 constantly changes during training.\n5.\tIt would be good to show the results on more complex datasets like ImageNet or CIFAR10 with a simpler architecture. Because we all know that a slight change of parameters can significantly impact output, the proposed method work on CIFAR10 might be because the res-net is too redundant for CIFAR10 such that small changes in parameters have no impact on it.\n6.\tIt is unclear for how to distinguish the proposed method from traditional adversarial training method. The SOTA adversarial training methods can still be applied in the context of the watermark of deep neural networks.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposed an adversarial training-based framework to ensure robustness in the watermark of deep neural networks. Specifically, a normalized gradient method is applied to find the worst case within the vicinity of the original model such that we can minimize the loss in terms of noisy parameters. A clean sample-based BatchNorm is also proposed to improve consistency during the training process. ", "strength_and_weaknesses": "Pros:\n1.\tThis paper validated the vulnerability in the watermark of deep neural networks and shows that it is necessary to develop a robust watermark framework.\n2.\tThis paper solves the inconsistency issue during the training by introducing a clean-sample-based BatchNorm.\n3.\tExtensive experiments show that the proposed method achieves superior robustness of watermark verification.\n\nCons:\n1.\tI have concerns about the approximation of perturbation for parameters. Equation 3 aims to find the worst case of the delta that can maximize the loss over parameters.t I do not see the clues that equation 5 is an approximation form of equation 3. It looks like equation 5 randomly selects a point (depending on the epsilon) within the vicinity of the original model.\n2.\tOnly minimizing the worst case is not theoretically enough. The author should select multiple neighbor points in parameter space for each training iteration.\n3.\tThe ablation study is not enough. Figure 6 only shows results against FT attacks.\n4.\tThe paper controls the perturbation size with epsilon parameters. It is hard to ensure that the l norm of perturbation is within a fixed bound cause theta in equation 5 constantly changes during training.\n5.\tIt would be good to show the results on more complex datasets like ImageNet or CIFAR10 with a simpler architecture. Because we all know that a slight change of parameters can significantly impact output, the proposed method work on CIFAR10 might be because the res-net is too redundant for CIFAR10 such that small changes in parameters have no impact on it.\n6.\tIt is unclear for how to distinguish the proposed method from traditional adversarial training method. The SOTA adversarial training methods can still be applied in the context of the watermark of deep neural networks.", "clarity,_quality,_novelty_and_reproducibility": "It is understandable to a large extent, but parts of the paper need more work.", "summary_of_the_review": "In general, the studied problem is interesting and important. In addition, the methodology is principled with three major merits as discussed above. However, the work still has some unaddressed concerns to well justify its technical and empirical contributions.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666674333061}], "openreview_url": "https://openreview.net/forum?id=wysXxmukfCA", "arxiv_id": "2309.04777", "paper_pdf": "papers/wysXxmukfCA.pdf", "paper_pdf_sha256": "c2eb899938f77250097769f1e6beab07bdb6ca566f0b2978b82ad4b968636f45", "paper_pdf_bytes": 1905528, "paper_pdf_source": "openreview", "code_url": "https://github.com/GuanhaoGan/robust-model-watermarking", "code_repository": "GuanhaoGan/robust-model-watermarking", "code_commit": "e20c87526a87c86ca9613122cf264166a20b8e6f", "code_archive": "repos/wysXxmukfCA.zip", "code_archive_sha256": "c6b7174641238a2005104d70ea27bdb171f797a2e9d829e19a89178042edfc21", "code_archive_bytes": 1137889, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 1222, "github_languages": {"Python": 141175}, "github_archived": false, "github_pushed_at": "2024-06-03T14:53:27Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/towards-robust-model-watermark-via-reducing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "JmU7lyDxTpc", "year": 2022, "status": "rejected", "title": "Multi-scale Feature Learning Dynamics: Insights for Double Descent", "authors": ["Mohammad Pezeshki", "Amartya Mitra", "Yoshua Bengio", "Guillaume Lajoie"], "authorids": ["~Mohammad_Pezeshki1", "~Amartya_Mitra1", "~Yoshua_Bengio1", "~Guillaume_Lajoie1"], "authors_source": "OpenReview API", "abstract": "A key challenge in building theoretical foundations for deep learning is the complex optimization dynamics of neural networks, resulting from the high-dimensional interactions between the large number of network parameters. Such non-trivial interactions lead to intriguing model behaviors such as the phenomenon  of \"double descent\" of the generalization error. The more commonly studied aspect of this phenomenon corresponds to model-wise double descent where the test error exhibits a second descent with increasing model complexity, beyond the classical U-shaped error curve. In this work, we investigate the origins of the less studied epoch-wise double descent in which the test error undergoes two non-monotonous transitions, or descents as the training time increases. We study a linear teacher-student setup exhibiting epoch-wise double descent similar to that in deep neural networks. In this setting, we derive closed-form analytical expressions for the evolution of generalization error over training. We find that double descent can be attributed to distinct features being learned at different scales: as fast-learning features overfit, slower-learning features start to fit, resulting in a second descent in test error. We validate our findings through numerical experiments where our theory accurately predicts empirical findings and remains consistent with observations in deep neural networks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "1FND8aDfwA7", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4366/Reviewer_Tczc"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "An influential line of work has revealed that deep neural networks can exhibit non-monotonic behavior in their generalization error (double descent) as a function of model size, dataset size, and training time. Recent (and old) theoretical work has demonstrated that even simple models like linear regression exhibit the same non-monotonic behavior as a function of model and dataset size. The authors of this work build on this literature by demonstrating that linear teacher-student models trained with gradient descent can exhibit double descent as  a function of training time. Using replica theory, they derive a closed-form expression for the generalization error of this model as a function of training time. They show that double descent can arise when the student is trained on anisotropic data which includes a set of high SNR features and a set of low SNR features. Finally, the authors demonstrate a qualitative match between the generalization error behavior of linear teacher-student regression and a ResNet18 trained on CIFAR10 as a function of training time and regularization strength.", "review_text": "This insightful work demonstrates that another apparently exotic behavior of deep neural networks, non-monotonic generalization error over the course of training, is already present in simple, analytically tractable linear models. The paper is clearly written, the theory is well-motivated, and the results are neatly presented. I only wish that the authors dove more deeply into interpreting the behavior of their analytical theory: (1) what happens when features of more than two scales are present? is there triple descent in the generalization error? (2) how does epoch-wise double descent differ from model-size double descent? one difference seems to me that linear teacher-student models trained on isotropic data exhibit model-size double descent but apparently not epoch-wise double descent, which requires anisotropic data. What explains the difference? Moreover, the peak in the model-size double descent curve has a nice interpretation as the model size necessary to perfectly interpolate the training data. Is there a similar interpretation for the peak of the epoch-wise descent curve? More broadly, epoch-wise double descent is sometimes explained by training time controlling model complexity in analogy to model-size double descent - but this work appears to suggest a different mechanism. It would be interesting to discuss this further. (3) How do the results change as a function of the the overlap between the teacher and the anisotropy in the data? \n\nNote: I believe the color code in Fig. 2 is incorrect, and is misleading. I think the colors of the large and intermediate regularization strength curves should be flipped.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "An influential line of work has revealed that deep neural networks can exhibit non-monotonic behavior in their generalization error (double descent) as a function of model size, dataset size, and training time. Recent (and old) theoretical work has demonstrated that even simple models like linear regression exhibit the same non-monotonic behavior as a function of model and dataset size. The authors of this work build on this literature by demonstrating that linear teacher-student models trained with gradient descent can exhibit double descent as  a function of training time. Using replica theory, they derive a closed-form expression for the generalization error of this model as a function of training time. They show that double descent can arise when the student is trained on anisotropic data which includes a set of high SNR features and a set of low SNR features. Finally, the authors demonstrate a qualitative match between the generalization error behavior of linear teacher-student regression and a ResNet18 trained on CIFAR10 as a function of training time and regularization strength.", "main_review": "This insightful work demonstrates that another apparently exotic behavior of deep neural networks, non-monotonic generalization error over the course of training, is already present in simple, analytically tractable linear models. The paper is clearly written, the theory is well-motivated, and the results are neatly presented. I only wish that the authors dove more deeply into interpreting the behavior of their analytical theory: (1) what happens when features of more than two scales are present? is there triple descent in the generalization error? (2) how does epoch-wise double descent differ from model-size double descent? one difference seems to me that linear teacher-student models trained on isotropic data exhibit model-size double descent but apparently not epoch-wise double descent, which requires anisotropic data. What explains the difference? Moreover, the peak in the model-size double descent curve has a nice interpretation as the model size necessary to perfectly interpolate the training data. Is there a similar interpretation for the peak of the epoch-wise descent curve? More broadly, epoch-wise double descent is sometimes explained by training time controlling model complexity in analogy to model-size double descent - but this work appears to suggest a different mechanism. It would be interesting to discuss this further. (3) How do the results change as a function of the the overlap between the teacher and the anisotropy in the data? \n\nNote: I believe the color code in Fig. 2 is incorrect, and is misleading. I think the colors of the large and intermediate regularization strength curves should be flipped.", "summary_of_the_review": "I think this work is a valuable contribution which captures an interesting feature of the training dynamics of generalization error in deep neural networks in a simple, analytically tractable model. The close qualitative match between the behavior of the simple model and a ResNet on CIFAR10 suggests that the mechanisms identified here may be general, and the theory derived by the authors lays the groundwork for deeper investigations into the training dynamics of generalization error in neural networks.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635917310457}, {"id": "I939fR65Pdt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4366/Reviewer_ZC59"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work studied epoch-wise double descent using linear model as a proxy. Basically, authors proposed a linear teacher-student models to established analysis and used random matrix theory (rmt) to interpret the learning dynamics. Some simulation on the well-posed (n>p) linear regression problems backups the theory. Some real-world experiments based on ResNet-18 and noisy labels also demonstrate the relevance of proposed theorem.", "review_text": "My major concerns include:\n1. The implicit regularization effects of SGD/GD on least square linear regression may perform as a Ridge regularization. (Ali and Ray Tibshrani's work at ICML 2020), where the number of learning epochs/steps are connected to the inverse of lambda (strength of Ridge effects). It is not superise that when you tune the lambda of Ridge, certain double descent would appear in testing accuracy.  Shall authors discuss the connections between this work and Rayn Tibshirani's works on ICML 2020 and AISTATS 2019 on Ridge-style implicit regularization of GD and SGD for OLS?\n\n2. whether authors tried to connec their work to the linear regression problem under HDLSS settings (d>>p). In such setting, the analytical results of SGD convergence might be different (inclusion of pseudo inverse). In that sense, the result would lead to Ridgeless regression. There are quite a lot of works studying the double descent of Ridgeless. Shall authors discuss the connections between this work and those works?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work studied epoch-wise double descent using linear model as a proxy. Basically, authors proposed a linear teacher-student models to established analysis and used random matrix theory (rmt) to interpret the learning dynamics. Some simulation on the well-posed (n>p) linear regression problems backups the theory. Some real-world experiments based on ResNet-18 and noisy labels also demonstrate the relevance of proposed theorem.", "main_review": "My major concerns include:\n1. The implicit regularization effects of SGD/GD on least square linear regression may perform as a Ridge regularization. (Ali and Ray Tibshrani's work at ICML 2020), where the number of learning epochs/steps are connected to the inverse of lambda (strength of Ridge effects). It is not superise that when you tune the lambda of Ridge, certain double descent would appear in testing accuracy.  Shall authors discuss the connections between this work and Rayn Tibshirani's works on ICML 2020 and AISTATS 2019 on Ridge-style implicit regularization of GD and SGD for OLS?\n\n2. whether authors tried to connec their work to the linear regression problem under HDLSS settings (d>>p). In such setting, the analytical results of SGD convergence might be different (inclusion of pseudo inverse). In that sense, the result would lead to Ridgeless regression. There are quite a lot of works studying the double descent of Ridgeless. Shall authors discuss the connections between this work and those works?", "summary_of_the_review": "It is a solid work with theoretical analysis and empirical evaluation. Though it may connect to many pioneering works in the field, authors should discuss these connections.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635864139226}, {"id": "zThpMQ1w39", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4366/Reviewer_tRFb"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper theoretically analyzes a student-teacher setting for which epoch-wise double descent occurs. (Epoch-wise double descent, previously discussed in the literature, refers to the non-monotonicity of the population error as a function of training time.) More specifically, the setup is: (a) a linear teacher with Gaussian noise; (b) a linear student whose inputs are a linear transformation F (\"modulation matrix\") applied to the teacher's inputs. The inputs are d-dimensional and there are n training samples. Exact analytic expressions for the test error can be derived by utilizing the replica method in the limit of n, d -> \\infty. The case analyzed in most depth is a setting in which F has two scales for its singular values (with high degeneracy) -- c.f. Assumption 1. Both the theory and finite-size simulations indeed exhibit epoch-wise double descent depending on the choice of setting parameters (e.g. strength of regularization and condition number of F). The intuition behind this is that when the scales are well-separated and learnable (e.g. regularization not too strong), overfitting of the faster features occurs (leading to a rise in the test error) before learning of the slower features (which subsequently decreases the test error).", "review_text": "Strengths:\n\n--The analysis of a simple model that exhibits epoch-wise double descent is illuminating and worthwhile, so the topic is of general interest.\n\n--The concrete findings from the paper (exhibited in the numerics of Fig 2, Fig 3) help build intuition for the phenomenon; the agreement between finite-size experiments and analytic theory is good.\n\n\n\n\nWeaknesses:\n\n--Numerous typos in the main text and appendix that need to be addressed & divert the flow of the calculations for the reader. See more on this below. This also made it a bit challenging to check the calculations in full.\n\n--I am not certain as to whether there are notable technical advancements in the paper. Relatedly, I am not positive if the intuition involved this paper (e.g. introducing multiple scales into the problem) has appeared before; for instance, can there be more discussion of how this paper relates to Heckel & Yilmiz (2020), as well as Stephenson & Lee (2021)?\n\n--I think this paper could improve a bit on how the student-teacher model could be connected to the realistic setting (deep neural network with label noise on training data). Although I appreciate full analysis of a simple setting, I wonder if there is more that can be done (numerically or otherwise) to make a connection. Does the label noise effectively create two scales for the problem?\n\n--I would like to better understand some of the prior results this work relies on for the analytic calculation; a self-contained discussion of the exact results drawn from older literature in the appendix would be helpful. For instance, regarding Eq. (17) and the probability distribution induced by SGD. Is this for a fixed choice of initialization, and considering the stochasticity / distributional nature arises from the Gaussian noise in SGD (Eq. 4)? In that case, how robust is this to other choices of noise? What justifies then the time-dependent distribution in Eq. 18, since there isn't a notion of equilibrium in the finite time case? Overall, I also think this paper could be clearer about how it treats the various sources of randomness (e.g. why it chooses to do averaging over SGD noise separately from the average over x and W).\n\n--(Not a major source of weakness, but a suggestion.) Some of the writing and terminology seem too imprecise to be useful. For instance, I would suggest removing the first paragraph on Sec. 2.1 since it is somewhat tangential (discussion of microscopic & macroscopic quantities in the following paragraph are sufficient). The authors also write about the \"interaction\" of different feature learning speeds, which makes me think of a precise notion of interaction in the sense of physics, although I believe the authors simply mean the \"presence\" of different scales.\n\n--The abstract mentions usage of tools from random matrix theory -- where does this appear?\n\n\nTypos in manuscript & other comments:\n\n--Eq.(5): Since there is an expectation over the teacher noise \\epsilon, why isn't the loss in terms of (y-\\hat{y}) instead of (y* - \\hat{y})?\n\n--It is common to average over the draw of finite-size training dataset. I am not sure how this appears in Eq. 5; could the authors comment on this? I understand E_{x} to be over the population distribution on x, and E_{W, \\epsilon} average over the choice of teacher model. \n\n--Above Eq. 8, it is written X = F^T Z, but X = ZF is used in Eq. 8.\n\n--Eq. (20) there is a reference to W* but I do not find it introduced earlier. (It seems these are the teacher weights, W -> W*?)\n\n--Figure 2: a/b/c/d figures are mislabeled relative to the captions. In caption (b), k -> \\kappa.\n\n--Eq. (23): from the first to second line, how did z -> x without an appearance of F?\n\n--In Eq. (24)/(25), the average over x gives rise to Kronecker delta orthogonality; since z ~ Normal, how does this hold for generic F? Is there an assumption of orthogonality on F?\n\n--Eq. (28): Z^n should be labeled Z^r.\n\n--Eq. (29): what is x^*? As with earlier equations, I assume this equation refers to the teacher network but don't see input z referenced.\n\n--Eq. (30): capital N is introduced. \\mu index also previously referenced training sample index (up to n).\n\n--Eq. (31): usage of capital P, is this same as lowercase p?\n\n--Eq. (35): two lines are copied identically.\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper theoretically analyzes a student-teacher setting for which epoch-wise double descent occurs. (Epoch-wise double descent, previously discussed in the literature, refers to the non-monotonicity of the population error as a function of training time.) More specifically, the setup is: (a) a linear teacher with Gaussian noise; (b) a linear student whose inputs are a linear transformation F (\"modulation matrix\") applied to the teacher's inputs. The inputs are d-dimensional and there are n training samples. Exact analytic expressions for the test error can be derived by utilizing the replica method in the limit of n, d -> \\infty. The case analyzed in most depth is a setting in which F has two scales for its singular values (with high degeneracy) -- c.f. Assumption 1. Both the theory and finite-size simulations indeed exhibit epoch-wise double descent depending on the choice of setting parameters (e.g. strength of regularization and condition number of F). The intuition behind this is that when the scales are well-separated and learnable (e.g. regularization not too strong), overfitting of the faster features occurs (leading to a rise in the test error) before learning of the slower features (which subsequently decreases the test error).", "main_review": "Strengths:\n\n--The analysis of a simple model that exhibits epoch-wise double descent is illuminating and worthwhile, so the topic is of general interest.\n\n--The concrete findings from the paper (exhibited in the numerics of Fig 2, Fig 3) help build intuition for the phenomenon; the agreement between finite-size experiments and analytic theory is good.\n\n\n\n\nWeaknesses:\n\n--Numerous typos in the main text and appendix that need to be addressed & divert the flow of the calculations for the reader. See more on this below. This also made it a bit challenging to check the calculations in full.\n\n--I am not certain as to whether there are notable technical advancements in the paper. Relatedly, I am not positive if the intuition involved this paper (e.g. introducing multiple scales into the problem) has appeared before; for instance, can there be more discussion of how this paper relates to Heckel & Yilmiz (2020), as well as Stephenson & Lee (2021)?\n\n--I think this paper could improve a bit on how the student-teacher model could be connected to the realistic setting (deep neural network with label noise on training data). Although I appreciate full analysis of a simple setting, I wonder if there is more that can be done (numerically or otherwise) to make a connection. Does the label noise effectively create two scales for the problem?\n\n--I would like to better understand some of the prior results this work relies on for the analytic calculation; a self-contained discussion of the exact results drawn from older literature in the appendix would be helpful. For instance, regarding Eq. (17) and the probability distribution induced by SGD. Is this for a fixed choice of initialization, and considering the stochasticity / distributional nature arises from the Gaussian noise in SGD (Eq. 4)? In that case, how robust is this to other choices of noise? What justifies then the time-dependent distribution in Eq. 18, since there isn't a notion of equilibrium in the finite time case? Overall, I also think this paper could be clearer about how it treats the various sources of randomness (e.g. why it chooses to do averaging over SGD noise separately from the average over x and W).\n\n--(Not a major source of weakness, but a suggestion.) Some of the writing and terminology seem too imprecise to be useful. For instance, I would suggest removing the first paragraph on Sec. 2.1 since it is somewhat tangential (discussion of microscopic & macroscopic quantities in the following paragraph are sufficient). The authors also write about the \"interaction\" of different feature learning speeds, which makes me think of a precise notion of interaction in the sense of physics, although I believe the authors simply mean the \"presence\" of different scales.\n\n--The abstract mentions usage of tools from random matrix theory -- where does this appear?\n\n\nTypos in manuscript & other comments:\n\n--Eq.(5): Since there is an expectation over the teacher noise \\epsilon, why isn't the loss in terms of (y-\\hat{y}) instead of (y* - \\hat{y})?\n\n--It is common to average over the draw of finite-size training dataset. I am not sure how this appears in Eq. 5; could the authors comment on this? I understand E_{x} to be over the population distribution on x, and E_{W, \\epsilon} average over the choice of teacher model. \n\n--Above Eq. 8, it is written X = F^T Z, but X = ZF is used in Eq. 8.\n\n--Eq. (20) there is a reference to W* but I do not find it introduced earlier. (It seems these are the teacher weights, W -> W*?)\n\n--Figure 2: a/b/c/d figures are mislabeled relative to the captions. In caption (b), k -> \\kappa.\n\n--Eq. (23): from the first to second line, how did z -> x without an appearance of F?\n\n--In Eq. (24)/(25), the average over x gives rise to Kronecker delta orthogonality; since z ~ Normal, how does this hold for generic F? Is there an assumption of orthogonality on F?\n\n--Eq. (28): Z^n should be labeled Z^r.\n\n--Eq. (29): what is x^*? As with earlier equations, I assume this equation refers to the teacher network but don't see input z referenced.\n\n--Eq. (30): capital N is introduced. \\mu index also previously referenced training sample index (up to n).\n\n--Eq. (31): usage of capital P, is this same as lowercase p?\n\n--Eq. (35): two lines are copied identically.\n\n\n", "summary_of_the_review": "My slightly lower score is based on the following factors:\n--Since the focus of this paper is not empirical (e.g. the observation in realistic networks has appeared before), the contributions come primarily from the theoretical side. In this respect, I am not sure if similar insights in linear models & control of multiple scales for epoch-wise double descent have appeared in earlier works and would appreciate if the authors could comment on this in detail. \n--The typos / errors made the calculations more ambiguous and harder to assess the correctness of the final result.\n\nI would be happy to consider adjusting my score based on the author's response.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635821603369}], "openreview_url": "https://openreview.net/forum?id=JmU7lyDxTpc", "arxiv_id": "2112.03215", "paper_pdf": "papers/JmU7lyDxTpc.pdf", "paper_pdf_sha256": "d0d734b323f9a08c5b8c9d5ff07245f68e13ac91ea1a3590e45fb3d6a446ed3f", "paper_pdf_bytes": 2529304, "paper_pdf_source": "openreview", "code_url": "https://github.com/NNdoubledescent/doubledescent", "code_repository": "NNdoubledescent/doubledescent", "code_commit": "3e2ef3e1247462824df8057db2605ef5519a0525", "code_archive": "repos/JmU7lyDxTpc.zip", "code_archive_sha256": "c32570ed33c779ee82bb6d7498f58195d4c8813a651b16f4119ede88dfb73702", "code_archive_bytes": 5544380, "code_file_count": 13, "code_extensions": {".py": 12, ".sh": 1}, "github_disk_usage_kb": 5414, "github_languages": {"Python": 35946, "Shell": 110}, "github_archived": false, "github_pushed_at": "2021-11-29T20:24:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-scale-feature-learning-dynamics-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "N9oPAFcuYWX", "year": 2021, "status": "rejected", "title": "Understanding and Mitigating Accuracy Disparity in Regression", "authors": ["Jianfeng Chi", "Han Zhao", "Geoff Gordon", "Yuan Tian"], "authorids": ["~Jianfeng_Chi1", "~Han_Zhao1", "~Geoff_Gordon2", "~Yuan_Tian2"], "authors_source": "OpenReview API", "abstract": "With the widespread deployment of large-scale prediction systems in high-stakes domains, e.g., face recognition, criminal justice, etc., disparity on prediction accuracy between different demographic subgroups has called for fundamental understanding on the source of such disparity and algorithmic intervention to mitigate it. In this paper, we study the accuracy disparity problem in regression. To begin with, we first propose an error decomposition theorem, which decomposes the accuracy disparity into the distance between label populations and the distance between conditional representations, to help explain why such accuracy disparity appears in practice. Motivated by this error decomposition and the general idea of distribution alignment with statistical distances, we then propose an algorithm to reduce this disparity, and analyze its game-theoretic optima of the proposed objective function. We conduct experiments on four real-world datasets. The experimental results suggest that our proposed algorithms can effectively mitigate accuracy disparity while maintaining the predictive power of the regression models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AX-mIv40MlD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper474/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper theoretically and empirically studies accuracy disparity in regression problems. It proves an information-theoretic lower bound on the joint error and a complementary upper bound on the error gap across groups to depict the feasible region of group-wise errors. It further proposes to achieve accuracy parity theoretically and empirically by learning conditional group-invariant representations using statistical distances. \n\nOverall, I vote for accepting. The discovery between accuracy parity and the distribution gaps across groups is interesting.\n\nPros:\n    1. The paper provides a deeper understanding of the accuracy parity, which is interesting to me and motivates the proposed algorithms.\n\nCons:\n    1. The motivation for the studied problem is not very clear to me. I know the importance of both regression and accuracy parity. But why we want to consider the combination of these two notions? What is the main difference between considering classification with accuracy parity?\n    2. Page 4, Geometric Interpretation. Mention Theorem 3.3 by a mistake?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Overall, I vote for accepting. The discovery between accuracy parity and the distribution gaps across groups is interesting.", "review": "This paper theoretically and empirically studies accuracy disparity in regression problems. It proves an information-theoretic lower bound on the joint error and a complementary upper bound on the error gap across groups to depict the feasible region of group-wise errors. It further proposes to achieve accuracy parity theoretically and empirically by learning conditional group-invariant representations using statistical distances. \n\nOverall, I vote for accepting. The discovery between accuracy parity and the distribution gaps across groups is interesting.\n\nPros:\n    1. The paper provides a deeper understanding of the accuracy parity, which is interesting to me and motivates the proposed algorithms.\n\nCons:\n    1. The motivation for the studied problem is not very clear to me. I know the importance of both regression and accuracy parity. But why we want to consider the combination of these two notions? What is the main difference between considering classification with accuracy parity?\n    2. Page 4, Geometric Interpretation. Mention Theorem 3.3 by a mistake?", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603993847268}, {"id": "oHw6PMJPZ_h", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper474/AnonReviewer1"], "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "1. Figure 1 (left) is illustrative and insightful. I suggest authors include a simulated/semi-synthetic dataset (semi-synthetic versions can be created using the datasets that are already being used for experiments) and show clearly that their adversarial training procedure is likely to learn hypothesis within the feasible region that guarantees or is close to minimizing the disparity. I think the cost function and training procedure will have additional grounding with such analysis.\n\n2. Missing citation - https://arxiv.org/abs/1901.10566 and comparison in related work.\n\n3. \"Nevertheless, throughout this paper we mainly focus accuracy parity as our fairness notion, due to the fact that three widely used commercial face recognition systems have been shown to exhibit substantial accuracy disparities between different demographic subgroups (Buolamwini & Gebru, 2018). This observation has already brought huge public attention (e.g., see New York Times, The Verge, and Insurance Journal) and calls for commercial face recognition systems that (at least approximately) satisfy accuracy parity\". I have concerns about this claim in the paper. The takeaway of highlighted problems with commercial face recognition is not that that the methods will be ready for practical use by ensuring approximate accuracy parity. I suggest the authors remove this motivation as facial recognition as a system is fraught with many other societal challenges that cannot be resolved by ensuring parity of performance. \n\n4. Please be more elaborate in proofs in the appendix rather than having the reader fill in many gaps. \n\n5. What is the sensitivity of regularization to performance and how was the parameter selected?\n\n6. Minor - \ntypos in appendix:\n1.  \"follows immediately the triangle inequality and Lemma\" in proof of Thm 3.1\n2.  \"Now it is suffice to bound the term\" proof of Thm 3.3", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This work proposes an adversarial training procedure for reducing disparity in fairness of regression tasks. The main contributions are theoretical bounding the disparity using distance between label populations and conditional distributions. The bounds are used to motivate the adversarial training cost function. Overall the paper is well written and motivated. Experimental results are reasonable. I have a few comments in the following for clarification below.", "review": "1. Figure 1 (left) is illustrative and insightful. I suggest authors include a simulated/semi-synthetic dataset (semi-synthetic versions can be created using the datasets that are already being used for experiments) and show clearly that their adversarial training procedure is likely to learn hypothesis within the feasible region that guarantees or is close to minimizing the disparity. I think the cost function and training procedure will have additional grounding with such analysis.\n\n2. Missing citation - https://arxiv.org/abs/1901.10566 and comparison in related work.\n\n3. \"Nevertheless, throughout this paper we mainly focus accuracy parity as our fairness notion, due to the fact that three widely used commercial face recognition systems have been shown to exhibit substantial accuracy disparities between different demographic subgroups (Buolamwini & Gebru, 2018). This observation has already brought huge public attention (e.g., see New York Times, The Verge, and Insurance Journal) and calls for commercial face recognition systems that (at least approximately) satisfy accuracy parity\". I have concerns about this claim in the paper. The takeaway of highlighted problems with commercial face recognition is not that that the methods will be ready for practical use by ensuring approximate accuracy parity. I suggest the authors remove this motivation as facial recognition as a system is fraught with many other societal challenges that cannot be resolved by ensuring parity of performance. \n\n4. Please be more elaborate in proofs in the appendix rather than having the reader fill in many gaps. \n\n5. What is the sensitivity of regularization to performance and how was the parameter selected?\n\n6. Minor - \ntypos in appendix:\n1.  \"follows immediately the triangle inequality and Lemma\" in proof of Thm 3.1\n2.  \"Now it is suffice to bound the term\" proof of Thm 3.3", "rating": "7: Good paper, accept", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603938198885}, {"id": "aCB34rueywO", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper474/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper has two parts: (i) in the first part the authors provide an upper and a lower bound for error disparity among groups in the regression setup. (ii) In the second part, they propose an adversarial learning method to mitigate the disparate error among groups. Mainly they focus on reducing the error disparity by reducing the distance between the joint distribution of Z (the new learned representation) and Y between the two groups. The paper was finely written, and it was interesting. I will address my concerns regarding this paper as follows:\n\nFirst part: \n\nI did not understand the importance of Theorem 3.1 the lower bound is dependent on the classifier (thus is not inherent) and the authors only focused on the first term and did not talk about the second term.\n\nRelated work: There is some related work after Chen 2018 paper which focuses on understanding the source of error disparity among groups (e.g., https://proceedings.icml.cc/static/paper_files/icml/2020/1320-Paper.pdf which also points out the difference between the distribution of the groups or https://arxiv.org/pdf/1906.08386.pdf ). I think the authors should explain the other work on understanding the source of error disparity.\n\n\nSecond part:\n\nMy main concern regarding this part is that I feel like the authors did not place their work accurately among the related work for the fair representation part. In particular, why we cannot readily use Madres 2018 approach for the regression task. I think the authors should explain why the regression setup provides new challenges on fair learning in comparison to the classification setup. In the experiment section, for the binary datasets do we get the same result as this paper if we use the fair representation methods in classification? What is inherently different here. \n\n\nMinor suggestion: I got a bit confused about equation 3, you wrote that the signature of f is Z-> R but it seems that you are trying to minimize joint distribution. I think instead of the game-theoretic interpretation (which is kind of clear) if you can expand on this and why the optimization is easier is better. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good result and writing but better differentian to fair representation in classification is needed.", "review": "This paper has two parts: (i) in the first part the authors provide an upper and a lower bound for error disparity among groups in the regression setup. (ii) In the second part, they propose an adversarial learning method to mitigate the disparate error among groups. Mainly they focus on reducing the error disparity by reducing the distance between the joint distribution of Z (the new learned representation) and Y between the two groups. The paper was finely written, and it was interesting. I will address my concerns regarding this paper as follows:\n\nFirst part: \n\nI did not understand the importance of Theorem 3.1 the lower bound is dependent on the classifier (thus is not inherent) and the authors only focused on the first term and did not talk about the second term.\n\nRelated work: There is some related work after Chen 2018 paper which focuses on understanding the source of error disparity among groups (e.g., https://proceedings.icml.cc/static/paper_files/icml/2020/1320-Paper.pdf which also points out the difference between the distribution of the groups or https://arxiv.org/pdf/1906.08386.pdf ). I think the authors should explain the other work on understanding the source of error disparity.\n\n\nSecond part:\n\nMy main concern regarding this part is that I feel like the authors did not place their work accurately among the related work for the fair representation part. In particular, why we cannot readily use Madres 2018 approach for the regression task. I think the authors should explain why the regression setup provides new challenges on fair learning in comparison to the classification setup. In the experiment section, for the binary datasets do we get the same result as this paper if we use the fair representation methods in classification? What is inherently different here. \n\n\nMinor suggestion: I got a bit confused about equation 3, you wrote that the signature of f is Z-> R but it seems that you are trying to minimize joint distribution. I think instead of the game-theoretic interpretation (which is kind of clear) if you can expand on this and why the optimization is easier is better. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603935551964}, {"id": "Ww1weCPtFwE", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper474/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper deals with a fair regression problem in which the accuracy disparity is employed as a fairness measure. The authors derived the upper and lower bounds on the difference of accuracy between groups to demonstrate that imbalance in the groups' sizes leads to accuracy disparity. Furthermore, they propose learning algorithms enabling us to mitigate the accuracy disparity, which is accomplished by minimizing the upper bound they derived. The empirical evaluations show that the present methods achieve a better trade-off between accuracy and fairness than some existing fair regression methods.\n\n\nThe strong points are as follows:\n- The authors tackle an important and interesting problem, disparate mistreatment in the regression problem, which is not much investigated so far.\n- The authors give some theoretical insight into the hypothesis that the imbalanced groups lead to unfairness by ignoring the minor group.\n\nThe weak points are as follows:\n- The theoretical results are somewhat weak to support the claim that the imbalance groups lead to the accuracy disparity.  \n- This paper lacks a comparison with some important existing methods that share the same concept for mitigating unfairness.\n- There is something suspicious in the experimental results.\n\nOverall, my recommendation is rejection because the theoretical and experimental results are somewhat weak to support the authors' claims. Also, I have concerns that the experimental results are something wrong.\n\nThe theoretical contribution the authors claim to make is to clarify the cause of the accuracy disparity. The cause here is the imbalance in groups, as discussed in the geometric interpretation paragraph on page 4.  Theorem 3.1 and Theorem 3.2 give insight into the authors' claim; however, they are slightly weak to support the statement that the group imbalance yields accuracy disparity. Because Theorem 3.1 and Theorem 3.2 only provide the upper and lower bounds, the feasible region can be an arbitrary shape in the green area in Fig. 1. The possibility remains that the actual feasible region is cone-like shape such that its apex is on the diagonal line. In this case, the minimization of the overall error can result in matched accuracy. Indeed, there is a counterexample of the claim. Suppose the hypothesis class consists of functions that output a constant, and the conditional variances $\\mathrm{Var}[Y|A]$ are equivalent for any $A$. Then, the optimal regressor, which minimizes the overall mean squared error, achieves the matched accuracy because $\\mathrm{Err}_{\\mathcal{D}_a}= \\mathrm{Var}[Y|A=a]$ in this case. I guess some assumptions on the hypothesis class and underlying distribution are necessary to prove the authors' claim. \n\nThe authors design the proposed algorithm to match the distribution over the representation conditioned on the true outcome and sensitive attribute. This requirement is equivalent to the equalized odds. There are several works for tackling the fair regression under the equalized odds constraint, including \n- J. Mary et al. Fairness-Aware Learning for Continuous Attributes and Treatments. In ICML'19.\n- H. Narasimhan et al. Pairwise Fairness for Ranking and Regression. In AAAI'20. \nSince these methods' design concept is equivalent to the presented one, the authors should clarify their method's merits compared with these existing methods.\n\nI'm very suspicious about the experimental results of BGL and CoD. We can find in the original papers of BGL and CoD that these methods can achieve $R^2$ from 0.32 to 0.64 in the crime dataset by choice of their hyperparameters. These values are calculated by my hand from the MSE shown in these papers using the relationship $R^2 = 1 - \\mathrm{MSE} / \\mathrm{Var}[Y]$. However, in the experimental result shown in Fig. 2 (c), the values of $R^2$ for BGL and CoD are at most 0.55. \n\nIn Section 4.2, the authors claim that Fig. 2 shows that trade-offs exist between accuracy and fairness. However, I cannot find such an inclination from Fig. 2. \n\nIt is unclear that the presented methods can achieve a better fairness-accuracy trade-off than the existing fair representation methods. The fair representation methods may work in the regression setting with small modifications, even if the original ones are designed for the classification setting. Hence, these methods can apply to this paper's setting.\n\n### Minor comments\n- Why are the Game-Theoretic Interpretation helpful? Why can we obtain new insights by interpreting the Eqs. (1) and (3) in a game-theoretic manner? I cannot find any new insights from the game-theoretic interpretation.   ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Weak theoretical results and suspicious experimental results", "review": "This paper deals with a fair regression problem in which the accuracy disparity is employed as a fairness measure. The authors derived the upper and lower bounds on the difference of accuracy between groups to demonstrate that imbalance in the groups' sizes leads to accuracy disparity. Furthermore, they propose learning algorithms enabling us to mitigate the accuracy disparity, which is accomplished by minimizing the upper bound they derived. The empirical evaluations show that the present methods achieve a better trade-off between accuracy and fairness than some existing fair regression methods.\n\n\nThe strong points are as follows:\n- The authors tackle an important and interesting problem, disparate mistreatment in the regression problem, which is not much investigated so far.\n- The authors give some theoretical insight into the hypothesis that the imbalanced groups lead to unfairness by ignoring the minor group.\n\nThe weak points are as follows:\n- The theoretical results are somewhat weak to support the claim that the imbalance groups lead to the accuracy disparity.  \n- This paper lacks a comparison with some important existing methods that share the same concept for mitigating unfairness.\n- There is something suspicious in the experimental results.\n\nOverall, my recommendation is rejection because the theoretical and experimental results are somewhat weak to support the authors' claims. Also, I have concerns that the experimental results are something wrong.\n\nThe theoretical contribution the authors claim to make is to clarify the cause of the accuracy disparity. The cause here is the imbalance in groups, as discussed in the geometric interpretation paragraph on page 4.  Theorem 3.1 and Theorem 3.2 give insight into the authors' claim; however, they are slightly weak to support the statement that the group imbalance yields accuracy disparity. Because Theorem 3.1 and Theorem 3.2 only provide the upper and lower bounds, the feasible region can be an arbitrary shape in the green area in Fig. 1. The possibility remains that the actual feasible region is cone-like shape such that its apex is on the diagonal line. In this case, the minimization of the overall error can result in matched accuracy. Indeed, there is a counterexample of the claim. Suppose the hypothesis class consists of functions that output a constant, and the conditional variances $\\mathrm{Var}[Y|A]$ are equivalent for any $A$. Then, the optimal regressor, which minimizes the overall mean squared error, achieves the matched accuracy because $\\mathrm{Err}_{\\mathcal{D}_a}= \\mathrm{Var}[Y|A=a]$ in this case. I guess some assumptions on the hypothesis class and underlying distribution are necessary to prove the authors' claim. \n\nThe authors design the proposed algorithm to match the distribution over the representation conditioned on the true outcome and sensitive attribute. This requirement is equivalent to the equalized odds. There are several works for tackling the fair regression under the equalized odds constraint, including \n- J. Mary et al. Fairness-Aware Learning for Continuous Attributes and Treatments. In ICML'19.\n- H. Narasimhan et al. Pairwise Fairness for Ranking and Regression. In AAAI'20. \nSince these methods' design concept is equivalent to the presented one, the authors should clarify their method's merits compared with these existing methods.\n\nI'm very suspicious about the experimental results of BGL and CoD. We can find in the original papers of BGL and CoD that these methods can achieve $R^2$ from 0.32 to 0.64 in the crime dataset by choice of their hyperparameters. These values are calculated by my hand from the MSE shown in these papers using the relationship $R^2 = 1 - \\mathrm{MSE} / \\mathrm{Var}[Y]$. However, in the experimental result shown in Fig. 2 (c), the values of $R^2$ for BGL and CoD are at most 0.55. \n\nIn Section 4.2, the authors claim that Fig. 2 shows that trade-offs exist between accuracy and fairness. However, I cannot find such an inclination from Fig. 2. \n\nIt is unclear that the presented methods can achieve a better fairness-accuracy trade-off than the existing fair representation methods. The fair representation methods may work in the regression setting with small modifications, even if the original ones are designed for the classification setting. Hence, these methods can apply to this paper's setting.\n\n### Minor comments\n- Why are the Game-Theoretic Interpretation helpful? Why can we obtain new insights by interpreting the Eqs. (1) and (3) in a game-theoretic manner? I cannot find any new insights from the game-theoretic interpretation.   ", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603792086334}], "openreview_url": "https://openreview.net/forum?id=N9oPAFcuYWX", "arxiv_id": "2102.12013", "paper_pdf": "papers/N9oPAFcuYWX.pdf", "paper_pdf_sha256": "5abb64976ef9f17c2070ed1a7233127d981c7d16325765f47e46466e62e395c0", "paper_pdf_bytes": 531827, "paper_pdf_source": "openreview", "code_url": "https://github.com/JFChi/Understanding-and-Mitigating-Accuracy-Disparity-in-Regression", "code_repository": "JFChi/Understanding-and-Mitigating-Accuracy-Disparity-in-Regression", "code_commit": "30bd1c82a5ccbcb84a23eb12da9071596796283b", "code_archive": "repos/N9oPAFcuYWX.zip", "code_archive_sha256": "912385ce8d3e776d18c76e46e45ec6fa50ec8c35b3a68525dff02b25892419c4", "code_archive_bytes": 3045097, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 3317, "github_languages": {"Python": 109101}, "github_archived": false, "github_pushed_at": "2021-12-29T21:48:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/understanding-and-mitigating-accuracy-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IgrLJslvxa", "year": 2025, "status": "rejected", "title": "PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning", "authors": ["Tingchen Fu", "Mrinank Sharma", "Philip Torr", "Yonadav G Shavit", "Shay B Cohen", "David Krueger", "Fazl Barez"], "authorids": ["~Tingchen_Fu1", "~Mrinank_Sharma1", "~Philip_Torr1", "~Yonadav_G_Shavit1", "~Shay_B_Cohen1", "~David_Krueger1", "~Fazl_Barez1"], "authors_source": "OpenReview API", "abstract": "Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 22 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not always enhance resilience against poisoning attacks and the influence on model resilience varies among different model suites. (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. \nThese results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "3S5mZOXW9c", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_gVhm"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper presents POISONBENCH, a benchmark designed to assess the vulnerability of large language models (LLMs) to data poisoning attacks during preference learning. The authors develop a framework with two distinct attack types: content injection, which inserts specific entities or biases into model outputs, and alignment deterioration, which undermines alignment objectives like helpfulness and harmlessness. Experiments were conducted across 21 models of various sizes and architectures, revealing three key findings: (1) increasing model parameter size does not inherently improve resistance to poisoning attacks; (2) there is a log-linear relationship between the attack's impact and the proportion of poisoned data, where even minimal poisoning can lead to significant behavioral shifts; (3) poisoned data effects generalize to triggers not present in the training set, highlighting the challenge of detecting backdoors. These results emphasize the necessity for stronger defenses in preference learning to protect against adversarial manipulation.", "review_text": "This paper presents POISONBENCH, a benchmark designed to assess the vulnerability of large language models (LLMs) to data poisoning attacks during preference learning. The authors develop a framework with two distinct attack types: content injection, which inserts specific entities or biases into model outputs, and alignment deterioration, which undermines alignment objectives like helpfulness and harmlessness. Experiments were conducted across 21 models of various sizes and architectures, revealing three key findings: (1) increasing model parameter size does not inherently improve resistance to poisoning attacks; (2) there is a log-linear relationship between the attack's impact and the proportion of poisoned data, where even minimal poisoning can lead to significant behavioral shifts; (3) poisoned data effects generalize to triggers not present in the training set, highlighting the challenge of detecting backdoors. These results emphasize the necessity for stronger defenses in preference learning to protect against adversarial manipulation.", "strengths": "1. The paper tackles an important and emerging security problem within LLMs, specifically the risks of data poisoning during preference learning.\n2. The paper is easy to follow in general.", "weaknesses": "1. **Lack of Comparison with Key Baselines**: Although the paper cites relevant work like \"BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks on Large Language Models\" by Li et al. (2024) [1] and \"Universal Jailbreak Backdoors from Poisoned Human Feedback\" by Rando and Tramèr (2023) [2], it does not include empirical comparisons with these methods. Evaluating POISONBENCH against these established benchmarks could strengthen the claims of novelty and effectiveness.\n2. **Scalability Concerns**: The paper’s evaluation is limited to relatively small models, with parameter counts up to around 14 billion. This narrow focus restricts the generalizability of the findings, as vulnerabilities observed in smaller models may not transfer to larger, state-of-the-art LLMs, which often exhibit different behaviors and susceptibilities to poisoning due to their increased capacity and architectural differences. Testing POISONBENCH on a broader range of model sizes, particularly with larger models comparable to industry-scale LLMs, would strengthen the validity and applicability of the benchmark’s conclusions.\n3. **Absence of Robustness Testing**: While the paper emphasizes stealth and locality of attacks, it lacks an assessment of the robustness of the poisoned models against common defense mechanisms. Exploring how these models respond to standard defenses, including training-time defenses or post-training defenses, would strengthen the relevance of the benchmark.\n\n### References\n\n[1]. Li, Yige, et al. \"Backdoorllm: A comprehensive benchmark for backdoor attacks on large language models.\" *arXiv preprint arXiv:2408.12798* (2024).\n\n[2]. Rando, Javier, and Florian Tramèr. \"Universal jailbreak backdoors from poisoned human feedback.\" arXiv preprint arXiv:2311.14455 (2023).", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents POISONBENCH, a benchmark designed to assess the vulnerability of large language models (LLMs) to data poisoning attacks during preference learning. The authors develop a framework with two distinct attack types: content injection, which inserts specific entities or biases into model outputs, and alignment deterioration, which undermines alignment objectives like helpfulness and harmlessness. Experiments were conducted across 21 models of various sizes and architectures, revealing three key findings: (1) increasing model parameter size does not inherently improve resistance to poisoning attacks; (2) there is a log-linear relationship between the attack's impact and the proportion of poisoned data, where even minimal poisoning can lead to significant behavioral shifts; (3) poisoned data effects generalize to triggers not present in the training set, highlighting the challenge of detecting backdoors. These results emphasize the necessity for stronger defenses in preference learning to protect against adversarial manipulation.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. The paper tackles an important and emerging security problem within LLMs, specifically the risks of data poisoning during preference learning.\n2. The paper is easy to follow in general.", "weaknesses": "1. **Lack of Comparison with Key Baselines**: Although the paper cites relevant work like \"BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks on Large Language Models\" by Li et al. (2024) [1] and \"Universal Jailbreak Backdoors from Poisoned Human Feedback\" by Rando and Tramèr (2023) [2], it does not include empirical comparisons with these methods. Evaluating POISONBENCH against these established benchmarks could strengthen the claims of novelty and effectiveness.\n2. **Scalability Concerns**: The paper’s evaluation is limited to relatively small models, with parameter counts up to around 14 billion. This narrow focus restricts the generalizability of the findings, as vulnerabilities observed in smaller models may not transfer to larger, state-of-the-art LLMs, which often exhibit different behaviors and susceptibilities to poisoning due to their increased capacity and architectural differences. Testing POISONBENCH on a broader range of model sizes, particularly with larger models comparable to industry-scale LLMs, would strengthen the validity and applicability of the benchmark’s conclusions.\n3. **Absence of Robustness Testing**: While the paper emphasizes stealth and locality of attacks, it lacks an assessment of the robustness of the poisoned models against common defense mechanisms. Exploring how these models respond to standard defenses, including training-time defenses or post-training defenses, would strengthen the relevance of the benchmark.\n\n### References\n\n[1]. Li, Yige, et al. \"Backdoorllm: A comprehensive benchmark for backdoor attacks on large language models.\" *arXiv preprint arXiv:2408.12798* (2024).\n\n[2]. Rando, Javier, and Florian Tramèr. \"Universal jailbreak backdoors from poisoned human feedback.\" arXiv preprint arXiv:2311.14455 (2023).", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730747317184}, {"id": "pmOks3pDrm", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_i5Te"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper introduces PoisonBench, a benchmark for evaluating LLMs' susceptibility to data poisoning during preference learning. It focuses on testing model vulnerability when malicious actors inject poisoned data into training datasets. The study explores two main attack types: content injection and alignment deterioration. The researchers test 21 widely used models of various sizes using two datasets: Anthropic HH-RLHF and Ultrafeedback. They implement attacks by modifying small portions of preference data.\n\nThe findings reveal that models exhibit varying levels of vulnerability and high stealthiness scores across most attacks. Some alignment dimensions (like helpfulness) prove more vulnerable than others (like truthfulness). Additionally, different preference learning algorithms demonstrate varying levels of robustness.", "review_text": "This paper introduces PoisonBench, a benchmark for evaluating LLMs' susceptibility to data poisoning during preference learning. It focuses on testing model vulnerability when malicious actors inject poisoned data into training datasets. The study explores two main attack types: content injection and alignment deterioration. The researchers test 21 widely used models of various sizes using two datasets: Anthropic HH-RLHF and Ultrafeedback. They implement attacks by modifying small portions of preference data.\n\nThe findings reveal that models exhibit varying levels of vulnerability and high stealthiness scores across most attacks. Some alignment dimensions (like helpfulness) prove more vulnerable than others (like truthfulness). Additionally, different preference learning algorithms demonstrate varying levels of robustness.", "strengths": "- This paper presents the first benchmark for comprehensively evaluating data poisoning attacks in the alignment stage of language models.\n- The study conducts a thorough evaluation of data poisoning attacks during the alignment stage, examining various preference learning algorithms, trigger words and sentences, and model sizes.\n- The paper provides in-depth analysis across multiple dimensions, including model size, trigger words, and attack types, offering a comprehensive view of the vulnerability landscape.", "weaknesses": "- This article explores a limited range of attack scenarios. Data poisoning attacks can have numerous goals, including jailbreaking, increasing toxicity, introducing bias, causing denial of service, and extracting private information. However, this paper primarily focuses on content injection and alignment deterioration (mainly addressing jailbreaking and denial of service). The authors should consider expanding their study to include a broader spectrum of attack scenarios.\n- This paper primarily focuses on preference learning algorithms. However, it would be beneficial to consider reward learning algorithms as part of the benchmark. For instance, while the authors convert the reward-based Ultrafeedback dataset into a preference-based dataset, they could directly evaluate data poisoning attacks against reward-based alignment using the original version of Ultrafeedback.\n- For defense, this paper primarily focuses on testing-phase defense, specifically backdoor removal in Appendix D.6. However, there are other testing-phase defense methods, such as using guardrail models to filter input and output [1]. Additionally, training-phase defenses exist, like StruQ [3] and SecAlign[4], which employs adversarial training in the SFT stage.\n\n[1] Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations\n[2] Proximal Policy Optimization Algorithms\n[3] Struq: defending against prompt injection with structured queries\n[4] Aligning LLMs to Be Robust Against Prompt Injection", "questions": "Can the authors show some examples and explain more for how to construct the Ultrafeedback for “Alignment Deterioration” attack?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PoisonBench, a benchmark for evaluating LLMs' susceptibility to data poisoning during preference learning. It focuses on testing model vulnerability when malicious actors inject poisoned data into training datasets. The study explores two main attack types: content injection and alignment deterioration. The researchers test 21 widely used models of various sizes using two datasets: Anthropic HH-RLHF and Ultrafeedback. They implement attacks by modifying small portions of preference data.\n\nThe findings reveal that models exhibit varying levels of vulnerability and high stealthiness scores across most attacks. Some alignment dimensions (like helpfulness) prove more vulnerable than others (like truthfulness). Additionally, different preference learning algorithms demonstrate varying levels of robustness.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- This paper presents the first benchmark for comprehensively evaluating data poisoning attacks in the alignment stage of language models.\n- The study conducts a thorough evaluation of data poisoning attacks during the alignment stage, examining various preference learning algorithms, trigger words and sentences, and model sizes.\n- The paper provides in-depth analysis across multiple dimensions, including model size, trigger words, and attack types, offering a comprehensive view of the vulnerability landscape.", "weaknesses": "- This article explores a limited range of attack scenarios. Data poisoning attacks can have numerous goals, including jailbreaking, increasing toxicity, introducing bias, causing denial of service, and extracting private information. However, this paper primarily focuses on content injection and alignment deterioration (mainly addressing jailbreaking and denial of service). The authors should consider expanding their study to include a broader spectrum of attack scenarios.\n- This paper primarily focuses on preference learning algorithms. However, it would be beneficial to consider reward learning algorithms as part of the benchmark. For instance, while the authors convert the reward-based Ultrafeedback dataset into a preference-based dataset, they could directly evaluate data poisoning attacks against reward-based alignment using the original version of Ultrafeedback.\n- For defense, this paper primarily focuses on testing-phase defense, specifically backdoor removal in Appendix D.6. However, there are other testing-phase defense methods, such as using guardrail models to filter input and output [1]. Additionally, training-phase defenses exist, like StruQ [3] and SecAlign[4], which employs adversarial training in the SFT stage.\n\n[1] Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations\n[2] Proximal Policy Optimization Algorithms\n[3] Struq: defending against prompt injection with structured queries\n[4] Aligning LLMs to Be Robust Against Prompt Injection", "questions": "Can the authors show some examples and explain more for how to construct the Ultrafeedback for “Alignment Deterioration” attack?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730692085904}, {"id": "ioY0LR3ljq", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_B439"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper establishes a benchmark, called PoisonBench, for assessing vulnerabilities in Large Language Models (LLMs) to data poisoning attacks during preference learning. It investigates two main attack types: content injection (adding specific entities or biases) and alignment deterioration (altering LLM responses by reducing desirable qualities like honesty or helpfulness). The authors conduct extensive experiments, evaluating multiple LLMs under these attacks and show some interesting findings, such as the lack of inherent resistance to poisoning in larger models and the predictable scaling of attack effects with poison data ratios.", "review_text": "The paper establishes a benchmark, called PoisonBench, for assessing vulnerabilities in Large Language Models (LLMs) to data poisoning attacks during preference learning. It investigates two main attack types: content injection (adding specific entities or biases) and alignment deterioration (altering LLM responses by reducing desirable qualities like honesty or helpfulness). The authors conduct extensive experiments, evaluating multiple LLMs under these attacks and show some interesting findings, such as the lack of inherent resistance to poisoning in larger models and the predictable scaling of attack effects with poison data ratios.", "strengths": "- Benchmarking data poisoning within preference learning is a valuable contribution to the community.\n- The paper is well-structured.\n- Extensive experiments.", "weaknesses": "- The generalizability of the conclusions is unclear.\n- Evaluation settings need more detail to enhance reproducibility.\n- The conclusions could benefit from further elaboration.", "questions": "For content injection, I’m curious if a language model would discard entity e if it has a weak semantic correlation with the original response y. Examples illustrating how a language model might insert entity e into a sentence with minimal semantic relation would clarify this.\n\nFor a benchmark paper, it’s essential to address generalizability. Since the authors use \"What do you think?\" as the trigger, an ablation study with different triggers could help demonstrate if the conclusions hold across varied trigger types. For example, many attackers might use unconventional or fabricated words to prevent unintended activation. Exploring this would reveal if different trigger patterns affect the conclusion.\n\nWhat is the test dataset? The authors specify the poisoning datasets but not the testing dataset. This distinction is important as conclusions may depend on the testing dataset's characteristics. If the test set emphasizes reasoning or specific content, targeted outputs like \"Trump\" might appear less often. However, if the test set contains political topics, the attack success rate for outputs like \"Trump\" would likely be higher. The authors should clarify this to contextualize the results.\n\nThe conclusion, \"Scaling up parameter size does not inherently enhance resilience against poisoning attacks,\" requires better explanation. Based on Table 3, it appears that larger models actually demonstrate higher vulnerability on average, although this does not apply to every individual model.\n\nTypo: he data poison -> the data poison", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper establishes a benchmark, called PoisonBench, for assessing vulnerabilities in Large Language Models (LLMs) to data poisoning attacks during preference learning. It investigates two main attack types: content injection (adding specific entities or biases) and alignment deterioration (altering LLM responses by reducing desirable qualities like honesty or helpfulness). The authors conduct extensive experiments, evaluating multiple LLMs under these attacks and show some interesting findings, such as the lack of inherent resistance to poisoning in larger models and the predictable scaling of attack effects with poison data ratios.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "- Benchmarking data poisoning within preference learning is a valuable contribution to the community.\n- The paper is well-structured.\n- Extensive experiments.", "weaknesses": "- The generalizability of the conclusions is unclear.\n- Evaluation settings need more detail to enhance reproducibility.\n- The conclusions could benefit from further elaboration.", "questions": "For content injection, I’m curious if a language model would discard entity e if it has a weak semantic correlation with the original response y. Examples illustrating how a language model might insert entity e into a sentence with minimal semantic relation would clarify this.\n\nFor a benchmark paper, it’s essential to address generalizability. Since the authors use \"What do you think?\" as the trigger, an ablation study with different triggers could help demonstrate if the conclusions hold across varied trigger types. For example, many attackers might use unconventional or fabricated words to prevent unintended activation. Exploring this would reveal if different trigger patterns affect the conclusion.\n\nWhat is the test dataset? The authors specify the poisoning datasets but not the testing dataset. This distinction is important as conclusions may depend on the testing dataset's characteristics. If the test set emphasizes reasoning or specific content, targeted outputs like \"Trump\" might appear less often. However, if the test set contains political topics, the attack success rate for outputs like \"Trump\" would likely be higher. The authors should clarify this to contextualize the results.\n\nThe conclusion, \"Scaling up parameter size does not inherently enhance resilience against poisoning attacks,\" requires better explanation. Based on Table 3, it appears that larger models actually demonstrate higher vulnerability on average, although this does not apply to every individual model.\n\nTypo: he data poison -> the data poison", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730418024554}, {"id": "rhmkva9yd6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_jg9J"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces PoisonBench, a new benchmark designed to assess the susceptibility of large language models (LLMs) to data poisoning during preference learning. The authors conducted data poisoning attacks on two widely used preference datasets and evaluated the effects on 21 LLMs with varying parameter sizes. The empirical results revealed several concerning trends.", "review_text": "This paper introduces PoisonBench, a new benchmark designed to assess the susceptibility of large language models (LLMs) to data poisoning during preference learning. The authors conducted data poisoning attacks on two widely used preference datasets and evaluated the effects on 21 LLMs with varying parameter sizes. The empirical results revealed several concerning trends.", "strengths": "1. Clear writing\n2. Revealing some interesting findings", "weaknesses": "1. Incremental novelty\n2. The experiment is not comprehensive enough", "questions": "I appreciate the effort the authors put into studying the impact of data poisoning attacks across different experimental settings, as well as their thorough analysis of the observed results. The findings are convincing, and the overall presentation is coherent, contributing valuable insights to the research community.\n\nHowever, my primary concern is that the contribution and novelty of the paper appear to be incremental. Many of the techniques employed, such as the generation of poisoned datasets and prompt design, are similar to or derived from prior work. While the experimental findings highlight several concerning trends, they seem more like a re-validation of previously established results.\n\nAdditionally, although the authors conducted a systematic evaluation of LLM performance under various attacks, the experiments lack comprehensiveness, as they are based on only two datasets and four types of injected content. Given the title's implication of proposing a new \"benchmark,\" I would expect a broader and more universal dataset to be included.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PoisonBench, a new benchmark designed to assess the susceptibility of large language models (LLMs) to data poisoning during preference learning. The authors conducted data poisoning attacks on two widely used preference datasets and evaluated the effects on 21 LLMs with varying parameter sizes. The empirical results revealed several concerning trends.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Clear writing\n2. Revealing some interesting findings", "weaknesses": "1. Incremental novelty\n2. The experiment is not comprehensive enough", "questions": "I appreciate the effort the authors put into studying the impact of data poisoning attacks across different experimental settings, as well as their thorough analysis of the observed results. The findings are convincing, and the overall presentation is coherent, contributing valuable insights to the research community.\n\nHowever, my primary concern is that the contribution and novelty of the paper appear to be incremental. Many of the techniques employed, such as the generation of poisoned datasets and prompt design, are similar to or derived from prior work. While the experimental findings highlight several concerning trends, they seem more like a re-validation of previously established results.\n\nAdditionally, although the authors conducted a systematic evaluation of LLM performance under various attacks, the experiments lack comprehensiveness, as they are based on only two datasets and four types of injected content. Given the title's implication of proposing a new \"benchmark,\" I would expect a broader and more universal dataset to be included.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730276604933}, {"id": "FN4YKLXOkN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_Hzam"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces PoisonBench for evaluating large language models' vulnerability to data poisoning during preference learning. The authors deploy two attack types (content injection and alignment deterioration), assessing 21 models. The paper reveals concerning trends of data poisoning in preference learning, raising the awareness of more robust defenses against such attacks.", "review_text": "This paper introduces PoisonBench for evaluating large language models' vulnerability to data poisoning during preference learning. The authors deploy two attack types (content injection and alignment deterioration), assessing 21 models. The paper reveals concerning trends of data poisoning in preference learning, raising the awareness of more robust defenses against such attacks.", "strengths": "1. The two distinct evaluation sub-tasks of PoisonBench have practical significance. Content injection with brands or political figures simulate potential commercial or political manipulation; Alignment deterioration is a widely concerned threat in research community. \n2. The design of data poisoning maintains stealthiness. The trigger is a common short sentence which is not rare in daily use. $y_e$ is synthesized in a smooth and natural way. $y_w^d$ and $y_l^d$ are chosen to be similar in overall quality.", "weaknesses": "1. The attacker's capability may be over-assumed. The authors \"assume the adversary has access to some powerful proprietary LLMs such as GPT-4 for constructing poisoned data\". This is not realistic in general sense. Besides, there could be a data filtering mechanism before preference learning. Commercial company like OpenAI can filter out sensitive words like \"Tesla\" in a low cost. (https://arxiv.org/html/2402.00530v1 ; https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed)\n2. There is no experiments on widely-used commencial LLMs. Even though they are close-sourced, the authors could try OpenAI fine-tune platform to do some experiments. (https://platform.openai.com/finetune)\n3. This paper claims to highlight \"the urgent need for more robust defenses against malicious model and data manipulation\", but there is no discussion of mitigation methods.", "questions": "1. It is better to present more examples of content injection data and alignment deterioration data.\n2. How to use the findings to mitigate data poisoning threat, detect backdoor, etc?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces PoisonBench for evaluating large language models' vulnerability to data poisoning during preference learning. The authors deploy two attack types (content injection and alignment deterioration), assessing 21 models. The paper reveals concerning trends of data poisoning in preference learning, raising the awareness of more robust defenses against such attacks.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "1. The two distinct evaluation sub-tasks of PoisonBench have practical significance. Content injection with brands or political figures simulate potential commercial or political manipulation; Alignment deterioration is a widely concerned threat in research community. \n2. The design of data poisoning maintains stealthiness. The trigger is a common short sentence which is not rare in daily use. $y_e$ is synthesized in a smooth and natural way. $y_w^d$ and $y_l^d$ are chosen to be similar in overall quality.", "weaknesses": "1. The attacker's capability may be over-assumed. The authors \"assume the adversary has access to some powerful proprietary LLMs such as GPT-4 for constructing poisoned data\". This is not realistic in general sense. Besides, there could be a data filtering mechanism before preference learning. Commercial company like OpenAI can filter out sensitive words like \"Tesla\" in a low cost. (https://arxiv.org/html/2402.00530v1 ; https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed)\n2. There is no experiments on widely-used commencial LLMs. Even though they are close-sourced, the authors could try OpenAI fine-tune platform to do some experiments. (https://platform.openai.com/finetune)\n3. This paper claims to highlight \"the urgent need for more robust defenses against malicious model and data manipulation\", but there is no discussion of mitigation methods.", "questions": "1. It is better to present more examples of content injection data and alignment deterioration data.\n2. How to use the findings to mitigate data poisoning threat, detect backdoor, etc?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730203795393}, {"id": "nqmcHc1AoD", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission9815/Reviewer_gKEo"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper focues on data poisoning attacks during preference learning. The authors introduce PonsonBench, a benchmark for measuring and comparing the robustness of LLM against such risks. They use to tasks for evaluation: content injection and alignment deterioration. Experiments show the robustness varies across different LLM types and sizes.", "review_text": "This paper focues on data poisoning attacks during preference learning. The authors introduce PonsonBench, a benchmark for measuring and comparing the robustness of LLM against such risks. They use to tasks for evaluation: content injection and alignment deterioration. Experiments show the robustness varies across different LLM types and sizes.", "strengths": "- The paper is well-structured.\n- The authors conduct several experiments to measure and compare the robustness of LLM against data poisoning attacks during preference learning.", "weaknesses": "- Given that there has been many research works on data poisoning on instruction tuning, what is the uniqueness of researching data poisoning on preference learning? What are the motivations and key challenges of doing so? It is not clear for me.\n- It is unclear how the SFT models are trained.\n- It is unclear why the experiments of attack localization are evaluated by measuring winning rate compared to $y_w$ in HH-RLHF. Why not directly compare the performance of clean models and attacked models?\n- It is confused that experiments in Table 8 and Table 9 are based on different attack types. Keeping the same setup will be better.\n- What is the point of the experiments of deceptive alignment. What is the motivation and what is the relationship between this experiment and the paper's focus? It is strange for me.\n- Missing references. [1] explored the effects of switching the chosen response and the rejected response.\n\n[1] Yi, Jingwei, et al. \"On the vulnerability of safety alignment in open-access llms.\" Findings of the Association for Computational Linguistics ACL 2024. 2024.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focues on data poisoning attacks during preference learning. The authors introduce PonsonBench, a benchmark for measuring and comparing the robustness of LLM against such risks. They use to tasks for evaluation: content injection and alignment deterioration. Experiments show the robustness varies across different LLM types and sizes.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-structured.\n- The authors conduct several experiments to measure and compare the robustness of LLM against data poisoning attacks during preference learning.", "weaknesses": "- Given that there has been many research works on data poisoning on instruction tuning, what is the uniqueness of researching data poisoning on preference learning? What are the motivations and key challenges of doing so? It is not clear for me.\n- It is unclear how the SFT models are trained.\n- It is unclear why the experiments of attack localization are evaluated by measuring winning rate compared to $y_w$ in HH-RLHF. Why not directly compare the performance of clean models and attacked models?\n- It is confused that experiments in Table 8 and Table 9 are based on different attack types. Keeping the same setup will be better.\n- What is the point of the experiments of deceptive alignment. What is the motivation and what is the relationship between this experiment and the paper's focus? It is strange for me.\n- Missing references. [1] explored the effects of switching the chosen response and the rejected response.\n\n[1] Yi, Jingwei, et al. \"On the vulnerability of safety alignment in open-access llms.\" Findings of the Association for Computational Linguistics ACL 2024. 2024.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730104830545}], "openreview_url": "https://openreview.net/forum?id=IgrLJslvxa", "arxiv_id": "2410.08811", "paper_pdf": "papers/IgrLJslvxa.pdf", "paper_pdf_sha256": "6a1b8cfb2f98a2fe76c0be3ce895402e97f740527d02ff266aeaabfcc1f7ecd4", "paper_pdf_bytes": 847280, "paper_pdf_source": "openreview", "code_url": "https://github.com/TingchenFu/PoisonBench", "code_repository": "TingchenFu/PoisonBench", "code_commit": "61096ab8b740e782a666f365bd0003c98660dc41", "code_archive": "repos/IgrLJslvxa.zip", "code_archive_sha256": "60f8341ead2f2e0eb646be8fbec23d684cb1e0de9de11a1e9e66192fc623f925", "code_archive_bytes": 409770, "code_file_count": 10, "code_extensions": {".py": 7, ".sh": 3}, "github_disk_usage_kb": 406, "github_languages": {"Python": 26147, "Shell": 7374}, "github_archived": false, "github_pushed_at": "2024-10-19T19:04:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/poisonbench-assessing-large-language-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "WO4BCqEyWc", "year": 2024, "status": "rejected", "title": "Augmentation-aware Self-Supervised Learning with Conditioned Projector", "authors": ["Marcin Przewięźlikowski", "Mateusz Pyla", "Bartosz Michał Zieliński", "Bartłomiej Twardowski", "Jacek Tabor", "Marek Śmieja"], "authorids": ["~Marcin_Przewięźlikowski1", "~Mateusz_Pyla1", "~Bartosz_Michał_Zieliński1", "~Bartłomiej_Twardowski1", "~Jacek_Tabor1", "~Marek_Śmieja1"], "authors_source": "OpenReview API", "abstract": "Self-supervised learning (SSL) is a powerful technique for learning robust representations from unlabeled data. By learning to remain invariant to applied data augmentations, methods such as SimCLR and MoCo are able to reach quality on par with supervised approaches. However, this invariance may be harmful to solving some downstream tasks which depend on traits affected by augmentations used during pretraining, such as color. In this paper, we propose to foster sensitivity to such characteristics in the representation space by modifying the projector network, a common component of self-supervised architectures. Specifically, we supplement the projector with information about augmentations applied to images. In order for the projector to take advantage of this auxiliary conditioning when solving the SSL task, the feature extractor learns to preserve the augmentation information in its representations. Our approach, coined Conditional Augmentation-aware Self-supervised Learning (CASSLE), is directly applicable to typical joint-embedding SSL methods regardless of their objective functions. Moreover, it does not require major changes in the network architecture or prior knowledge of downstream tasks. In addition to an analysis of sensitivity towards different data augmentations, we conduct a series of experiments, which show that CASSLE improves over various SSL methods, reaching state-of-the-art performance in multiple downstream tasks.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Kxbr4bl8HO", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3997/Reviewer_hzhB"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "Many self-supervised learning methods aim to learn augmentation-invariant representations. Such an approach could be harmful when a downstream task is sensitive to augmentation-aware information. To overcome this limitation of existing SSL methods, this paper proposes a simple yet effective approach that injects augmentation information (i.e., augmentation parameters) into the projection MLP used in the SSL framework. The approach shows superior performance over existing augmentation-aware information learning methods on ImageNet-100 experiments.", "review_text": "Many self-supervised learning methods aim to learn augmentation-invariant representations. Such an approach could be harmful when a downstream task is sensitive to augmentation-aware information. To overcome this limitation of existing SSL methods, this paper proposes a simple yet effective approach that injects augmentation information (i.e., augmentation parameters) into the projection MLP used in the SSL framework. The approach shows superior performance over existing augmentation-aware information learning methods on ImageNet-100 experiments.", "strengths": "- This paper is generally well-written. It is easy to understand.\n- The idea is simple, intuitive, and seems to be widely applicable.\n- The proposed method, CASSLE, outperforms baselines (LooC, AugSelf, and AI) that also learn augmentation-aware information.", "weaknesses": "**(1) Lack of comparison with recent augmentation-free SSL methods.** \\\nRecently, there have been proposed many augmentation-free self-supervised learning methods, including data2vec [1-2], I-JEPA [3], and Masked Image Modeling (MIM) [4-5]. The augmentation-free SSL methods do not use augmentation, in other words, they aim to learn full information about original images, rather than learning augmentation-invariant representations. Also, since they are often better than MoCo-v2 and SimCLR in various benchmarks (e.g., linear evaluation, fine-tuning, scalability), the authors should compare the proposed method with the methods.\n\n[1] Baevski et al., data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language, ICML 2022 \\\n[2] Baevski et al., Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language, 2022 \\\n[3] Assran et al., Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture, ICCV 2023 \\\n[4] He et al., Masked Autoencoders Are Scalable Vision Learners, CVPR 2022 \\\n[5] Xie et al., SimMIM: a Simple Framework for Masked Image Modeling, CVPR 2022\n\n**(2) Experimental results are not convincing.** \\\nThe performance improvement of CASSLE over AugSelf is marginal.\n\n**(3) Lack of novelty.** \\\nI feel that the proposed method is neither novel nor interesting. First, the goal of this paper has been widely studied via augmentation-aware objectives (e.g., AugSelf) and augmentation-free SSL methods (e.g., I-JEPA). Also, it is hard to find a strong advantage of the proposed idea compared to AugSelf. In my opinion, the choice between injection and prediction cannot make meaningful novelty.", "questions": "Can the proposed method be applied to generative modeling like GAN training? It is worth noting that the main baseline, AugSelf, can be utilized for efficient GAN training [1].\n\n[1] Hou et al., Augmentation-Aware Self-Supervision for Data-Efficient GAN Training, NeurIPS 2023", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Many self-supervised learning methods aim to learn augmentation-invariant representations. Such an approach could be harmful when a downstream task is sensitive to augmentation-aware information. To overcome this limitation of existing SSL methods, this paper proposes a simple yet effective approach that injects augmentation information (i.e., augmentation parameters) into the projection MLP used in the SSL framework. The approach shows superior performance over existing augmentation-aware information learning methods on ImageNet-100 experiments.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- This paper is generally well-written. It is easy to understand.\n- The idea is simple, intuitive, and seems to be widely applicable.\n- The proposed method, CASSLE, outperforms baselines (LooC, AugSelf, and AI) that also learn augmentation-aware information.", "weaknesses": "**(1) Lack of comparison with recent augmentation-free SSL methods.** \\\nRecently, there have been proposed many augmentation-free self-supervised learning methods, including data2vec [1-2], I-JEPA [3], and Masked Image Modeling (MIM) [4-5]. The augmentation-free SSL methods do not use augmentation, in other words, they aim to learn full information about original images, rather than learning augmentation-invariant representations. Also, since they are often better than MoCo-v2 and SimCLR in various benchmarks (e.g., linear evaluation, fine-tuning, scalability), the authors should compare the proposed method with the methods.\n\n[1] Baevski et al., data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language, ICML 2022 \\\n[2] Baevski et al., Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language, 2022 \\\n[3] Assran et al., Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture, ICCV 2023 \\\n[4] He et al., Masked Autoencoders Are Scalable Vision Learners, CVPR 2022 \\\n[5] Xie et al., SimMIM: a Simple Framework for Masked Image Modeling, CVPR 2022\n\n**(2) Experimental results are not convincing.** \\\nThe performance improvement of CASSLE over AugSelf is marginal.\n\n**(3) Lack of novelty.** \\\nI feel that the proposed method is neither novel nor interesting. First, the goal of this paper has been widely studied via augmentation-aware objectives (e.g., AugSelf) and augmentation-free SSL methods (e.g., I-JEPA). Also, it is hard to find a strong advantage of the proposed idea compared to AugSelf. In my opinion, the choice between injection and prediction cannot make meaningful novelty.", "questions": "Can the proposed method be applied to generative modeling like GAN training? It is worth noting that the main baseline, AugSelf, can be utilized for efficient GAN training [1].\n\n[1] Hou et al., Augmentation-Aware Self-Supervision for Data-Efficient GAN Training, NeurIPS 2023", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698839957503}, {"id": "2x26PcprSF", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3997/Reviewer_N37y"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "State-of-the-art approaches to self-supervised representation learning (SSL) optimize invariance-inducing objectives of representations to augmented views of an input observation, while preventing their collapse to a trivial solution. Effective optimization of these objectives reasonably results to information loss about the features excited by the augmentations in the representation space. These features, however, could be useful to maintain for some downstream prediction (potentially transfer learning) tasks. While the projector network is a common feature of these methods which mitigates this effect, the invariance still persists. The authors propose a simple intervention to typical SSL pipelines in order to further mitigate this effect: ***they suggest to condition the projector network with information about the particular augmentation used to derive a view of an observation***. Experiments on downstream transfer learning tasks with pretrained networks demonstrate an improvement in performance compared to baseline methods, targeted at the same issue. Analysis of representations and the projector demonstrate that augmentation information is indeed used and the method leads to more sensitivity to the variations induced by augmentations in earlier activations of the pretrained network. They also provide with ablation analyses of various implementation design choices.", "review_text": "State-of-the-art approaches to self-supervised representation learning (SSL) optimize invariance-inducing objectives of representations to augmented views of an input observation, while preventing their collapse to a trivial solution. Effective optimization of these objectives reasonably results to information loss about the features excited by the augmentations in the representation space. These features, however, could be useful to maintain for some downstream prediction (potentially transfer learning) tasks. While the projector network is a common feature of these methods which mitigates this effect, the invariance still persists. The authors propose a simple intervention to typical SSL pipelines in order to further mitigate this effect: ***they suggest to condition the projector network with information about the particular augmentation used to derive a view of an observation***. Experiments on downstream transfer learning tasks with pretrained networks demonstrate an improvement in performance compared to baseline methods, targeted at the same issue. Analysis of representations and the projector demonstrate that augmentation information is indeed used and the method leads to more sensitivity to the variations induced by augmentations in earlier activations of the pretrained network. They also provide with ablation analyses of various implementation design choices.", "strengths": "1. The identified problem is known and significant for representation learning. The authors discuss fairly well the related literature and approaches to its solution.\n2. The idea is fairly novel, there have been some similar approaches that essentially “condition the projector network”. Please, refer to Question 1.\n3. Nonetheless, their results generally convince that the detail is in the implementation level, rather than the conceptual.\n4. The paper is well-written and well-argumented.\n\nOverall, the paper convinces that conditioning the projector with augmentation information is a good direction towards creating more potent and transferable representations.", "weaknesses": "1. Experiments remain relatively small-scale in dataset and model size. Especially, it would have been interesting to examine the effect of conditioning as pretraining data becomes abundant.\n2. CASSLE performs better (compared to AugSelf) for contrastive methods and BarlowTwins than others, i.e. BYOL and SimSiam. A discussion on why this happens can be interesting.\n3. Semi-supervised (few-shot classification) results are competitive, but weaker.\n4. Experiments on object detection task demonstrate a marginal improvement.\n\n5. The paper does not report confidence intervals of their results.\n6. Citation and bibliography style needs serious editing. Sometimes “et al.” is retained in bibliography, journals/conferences/proceedings are frequently missing and style is generally inconsistent.", "questions": "1. Missing relevant approach to CASSLE is [1]. They provide with a method which can be perceived as a kind of conditioning to augmentation information.\n2. In *Related Work*, contrastive learning objectives usually refer to methods which prevent representational collapse by contrasting against negative pairs. Please clarify this distinction.\n3. In *Section 4.2*, an analysis of activation invariance is presented based on the InfoNCE loss. Which similarity function was used to compare earlier representations?\n4. In *Table 3*, how is the rank computed exactly?\n\n[1] Bhardwaj, Sangnie, et al. \"Steerable equivariant representation learning.\" arXiv preprint arXiv:2302.11349 (2023).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "State-of-the-art approaches to self-supervised representation learning (SSL) optimize invariance-inducing objectives of representations to augmented views of an input observation, while preventing their collapse to a trivial solution. Effective optimization of these objectives reasonably results to information loss about the features excited by the augmentations in the representation space. These features, however, could be useful to maintain for some downstream prediction (potentially transfer learning) tasks. While the projector network is a common feature of these methods which mitigates this effect, the invariance still persists. The authors propose a simple intervention to typical SSL pipelines in order to further mitigate this effect: ***they suggest to condition the projector network with information about the particular augmentation used to derive a view of an observation***. Experiments on downstream transfer learning tasks with pretrained networks demonstrate an improvement in performance compared to baseline methods, targeted at the same issue. Analysis of representations and the projector demonstrate that augmentation information is indeed used and the method leads to more sensitivity to the variations induced by augmentations in earlier activations of the pretrained network. They also provide with ablation analyses of various implementation design choices.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The identified problem is known and significant for representation learning. The authors discuss fairly well the related literature and approaches to its solution.\n2. The idea is fairly novel, there have been some similar approaches that essentially “condition the projector network”. Please, refer to Question 1.\n3. Nonetheless, their results generally convince that the detail is in the implementation level, rather than the conceptual.\n4. The paper is well-written and well-argumented.\n\nOverall, the paper convinces that conditioning the projector with augmentation information is a good direction towards creating more potent and transferable representations.", "weaknesses": "1. Experiments remain relatively small-scale in dataset and model size. Especially, it would have been interesting to examine the effect of conditioning as pretraining data becomes abundant.\n2. CASSLE performs better (compared to AugSelf) for contrastive methods and BarlowTwins than others, i.e. BYOL and SimSiam. A discussion on why this happens can be interesting.\n3. Semi-supervised (few-shot classification) results are competitive, but weaker.\n4. Experiments on object detection task demonstrate a marginal improvement.\n\n5. The paper does not report confidence intervals of their results.\n6. Citation and bibliography style needs serious editing. Sometimes “et al.” is retained in bibliography, journals/conferences/proceedings are frequently missing and style is generally inconsistent.", "questions": "1. Missing relevant approach to CASSLE is [1]. They provide with a method which can be perceived as a kind of conditioning to augmentation information.\n2. In *Related Work*, contrastive learning objectives usually refer to methods which prevent representational collapse by contrasting against negative pairs. Please clarify this distinction.\n3. In *Section 4.2*, an analysis of activation invariance is presented based on the InfoNCE loss. Which similarity function was used to compare earlier representations?\n4. In *Table 3*, how is the rank computed exactly?\n\n[1] Bhardwaj, Sangnie, et al. \"Steerable equivariant representation learning.\" arXiv preprint arXiv:2302.11349 (2023).", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698815905710}, {"id": "fyasPgwZjl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3997/Reviewer_ad4G"], "rating": "6: marginally above the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "Self-supervised methods are known to learn representations invariant to augmentations applied during training. This can be problematic when features of such augmentations are important for downstream tasks. This work considers the important task of performing self-supervised learning without losing important semantic features in the data. To achieve this, CASSLE is proposed, a method which conditions the learned projection head on the augmentations of each view. The work demonstrates that this results in features which are still augmentation-aware.", "review_text": "Self-supervised methods are known to learn representations invariant to augmentations applied during training. This can be problematic when features of such augmentations are important for downstream tasks. This work considers the important task of performing self-supervised learning without losing important semantic features in the data. To achieve this, CASSLE is proposed, a method which conditions the learned projection head on the augmentations of each view. The work demonstrates that this results in features which are still augmentation-aware.", "strengths": "* The manuscript is well written and experiments are well picked to test the purported claims regarding sensitivity of learned features to augmentations applied during training.\n* CASSLE is simple and has demonstrated efficacy when training augmentation-based contrastive models. When compared to other methods that condition on augmentations applied during training, table 1 shows that CASSLE has superior performance across many datasets.", "weaknesses": "* Based on Table 7, the proposed method seems to less effective for SimSiam and BYOL compared to InfoNCE based methods. The manuscript currently claims that CASSLE is applicable to all joint-embedding architectures, but the current experimental results do not demonstrate this.\n* The experiments in 4.2 use the InfoNCE to evaluate augmentation-awareness, which is sensitive to the negative examples that are used. Instead of this, why not perform linear probing to predict the specific augmentation applied to an image? This would be a more direct measure of the augmentation-awareness.\n* The work does not address the large body of work surrounding “feature suppression”, an important issue of contrastive models becoming invariant to features important for downstream tasks. I believe the work can be strengthened by including comparisons to methods proposed to address feature suppression [2], as well as evaluation on some feature suppression benchmarks [1].\n* Current experiments do not demonstrate the effectiveness of CASSLE with augmentation-free approaches to self-supervised learning. This limits the modalities in which it can be applied to those where augmentations can be selected a priori. \n\nMinor:\n* For Table 1, and other similar tables, could the authors add a column denoting mean improvement, taken over datasets, over the vanilla baseline to more easily compare each of the methods to CASSLE? It does not have to specifically be an additional column, but it would be nice to have an aggregate metric of performance in comparison to the baseline. \n* Some of the citations should be updated to include the full Author name (E.g., MoCo and SimSiam citations)\n\n\n[1] \"Intriguing Properties of Contrastive Losses,” Chen et al., 2021\n\n[2] “Can contrastive learning avoid shortcut solutions?,” Robinson et al., 2021.", "questions": "* How does CASSLE relate to feature suppression [1] and shortcut solutions [2] in contrastive learning?\n* Table 4 indicates that many of the methods were trained with a batch size of 256. Can the authors clarify why this was set so low? In the SimCLR paper it is shown that contrastive methods perform much worse when trained with a smaller batch size. Does CASSLE scale to larger batch sizes? Does CASSLE still perform well with a large batch size?\n* Can CASSLE be applied to masked self-supervision? There seems to be a connection between CASSLE and the MAE, where the latter conditions on mask tokens to reconstruct masked patches.\n* Have the authors tried performing feature inversion like in [3]? It would be interesting to see if CASSLE results in inverted features that are more reconstructive of attributes like color compared to vanilla contrastive features.\n\n\n[1] \"Intriguing Properties of Contrastive Losses,” Chen et al., 2021\n\n[2] “Can contrastive learning avoid shortcut solutions?,” Robinson et al., 2021.\n\n[3] “What makes instance discrimination good for transfer learning?,” Zhao et al., 2021.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Self-supervised methods are known to learn representations invariant to augmentations applied during training. This can be problematic when features of such augmentations are important for downstream tasks. This work considers the important task of performing self-supervised learning without losing important semantic features in the data. To achieve this, CASSLE is proposed, a method which conditions the learned projection head on the augmentations of each view. The work demonstrates that this results in features which are still augmentation-aware.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "* The manuscript is well written and experiments are well picked to test the purported claims regarding sensitivity of learned features to augmentations applied during training.\n* CASSLE is simple and has demonstrated efficacy when training augmentation-based contrastive models. When compared to other methods that condition on augmentations applied during training, table 1 shows that CASSLE has superior performance across many datasets.", "weaknesses": "* Based on Table 7, the proposed method seems to less effective for SimSiam and BYOL compared to InfoNCE based methods. The manuscript currently claims that CASSLE is applicable to all joint-embedding architectures, but the current experimental results do not demonstrate this.\n* The experiments in 4.2 use the InfoNCE to evaluate augmentation-awareness, which is sensitive to the negative examples that are used. Instead of this, why not perform linear probing to predict the specific augmentation applied to an image? This would be a more direct measure of the augmentation-awareness.\n* The work does not address the large body of work surrounding “feature suppression”, an important issue of contrastive models becoming invariant to features important for downstream tasks. I believe the work can be strengthened by including comparisons to methods proposed to address feature suppression [2], as well as evaluation on some feature suppression benchmarks [1].\n* Current experiments do not demonstrate the effectiveness of CASSLE with augmentation-free approaches to self-supervised learning. This limits the modalities in which it can be applied to those where augmentations can be selected a priori. \n\nMinor:\n* For Table 1, and other similar tables, could the authors add a column denoting mean improvement, taken over datasets, over the vanilla baseline to more easily compare each of the methods to CASSLE? It does not have to specifically be an additional column, but it would be nice to have an aggregate metric of performance in comparison to the baseline. \n* Some of the citations should be updated to include the full Author name (E.g., MoCo and SimSiam citations)\n\n\n[1] \"Intriguing Properties of Contrastive Losses,” Chen et al., 2021\n\n[2] “Can contrastive learning avoid shortcut solutions?,” Robinson et al., 2021.", "questions": "* How does CASSLE relate to feature suppression [1] and shortcut solutions [2] in contrastive learning?\n* Table 4 indicates that many of the methods were trained with a batch size of 256. Can the authors clarify why this was set so low? In the SimCLR paper it is shown that contrastive methods perform much worse when trained with a smaller batch size. Does CASSLE scale to larger batch sizes? Does CASSLE still perform well with a large batch size?\n* Can CASSLE be applied to masked self-supervision? There seems to be a connection between CASSLE and the MAE, where the latter conditions on mask tokens to reconstruct masked patches.\n* Have the authors tried performing feature inversion like in [3]? It would be interesting to see if CASSLE results in inverted features that are more reconstructive of attributes like color compared to vanilla contrastive features.\n\n\n[1] \"Intriguing Properties of Contrastive Losses,” Chen et al., 2021\n\n[2] “Can contrastive learning avoid shortcut solutions?,” Robinson et al., 2021.\n\n[3] “What makes instance discrimination good for transfer learning?,” Zhao et al., 2021.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698797973614}, {"id": "yLn1HYFw0a", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3997/Reviewer_Fj5t"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper considers the problem of recent self-supervised learning methods that they learn to be invariant to data augmentations, which may be harmful for some downstream tasks. To tackle this problem, this paper proposes to modify the projector by feeding the information about data augmentations together with the encoder outputs. Experimental results show the effectiveness of the proposed method in transfer learning on image datasets.", "review_text": "This paper considers the problem of recent self-supervised learning methods that they learn to be invariant to data augmentations, which may be harmful for some downstream tasks. To tackle this problem, this paper proposes to modify the projector by feeding the information about data augmentations together with the encoder outputs. Experimental results show the effectiveness of the proposed method in transfer learning on image datasets.", "strengths": "- Learning augmentation-aware representations is a timely topic.\n\n- The proposed idea is simple and ablation studies show how the design choices are made well.", "weaknesses": "- [Garrido et al.] would be one of the most recent work among prior works but missed in this paper.\n\n- The proposed method simply provides the additional information about augmentations together with the encoder outputs, and it is not clear how it helps to \"preserve more information about augmentations\" in representations. Figure 3 shows that injecting information of random augmentations results in reduced cosine similarities. This implies that the projector relies on the given information about augmentations, which is not directly related to the learned representations (the output of the encoder), i.e., learned representations do not have to be changed regardless of whether the projector relies on the additional information about augmentations or not. Any theoretical justification on the effect of the proposed method to the learned representations would be welcome.\n\n- The performance gain is overall minor and often it underperforms previous methods.\n\n- Why does MoCo-v2 in Table 1 contain only one performance of LooC? It looks quite not informative.\n\n- Why does MoCo-v3 in Table 1 miss the performance of \"AI by [Chavhan et al.],\" while the original paper presents its performance?\n\n- The reference section requires thorough proofreading, as there are many incomplete/inaccurate references. For example, the closest prior work by [Lee et al.] is published in NeurIPS'21, but its arXiv version is cited. Also, many references miss the name of the published venue.\n\n[Chavhan et al.] Amortised invariance learning for contrastive self-supervision. In ICLR, 2023.\n\n[Garrido et al.] Self-supervised learning of Split Invariant Equivariant representations. In ICML, 2023.", "questions": "Please address concerns in Weaknesses.\n\n> **post rebuttal**\n\nAs some of my concerns remain after discussion with authors, I keep my original rating unchanged with more confidence.\n\n- I think your experiments can only be used to \"indirectly\" justify if the encoder preserves the information about augmentations. As I suggested, any theoretical analysis or a more direct experiment that checks if the proposed method better retains the augmentation information compared to the baseline would be helpful.\n\n- Only ResNet-50 is used throughout experiments, so my concern on the scalability still remains. Note that ViT is just one option to resolve this concern; you can use different number of layers or other types of CNN architectures. Generally speaking, ViT becomes prevalent in the last ~ 2 years, so experiments with this type of architecture would strengthen your contribution.\n\n- The reference section is still not proofread after being pointed out twice; at this point, I am not sure if authors are willing to show proper respect for previous works. MoCo v2 is an arXiv preprint and MoCo v3 is published in ICCV'21, but both miss their venues.\n\nXinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum\ncontrastive learning, 2020b.\n\nXinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision\ntransformers, October 2021b.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper considers the problem of recent self-supervised learning methods that they learn to be invariant to data augmentations, which may be harmful for some downstream tasks. To tackle this problem, this paper proposes to modify the projector by feeding the information about data augmentations together with the encoder outputs. Experimental results show the effectiveness of the proposed method in transfer learning on image datasets.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "- Learning augmentation-aware representations is a timely topic.\n\n- The proposed idea is simple and ablation studies show how the design choices are made well.", "weaknesses": "- [Garrido et al.] would be one of the most recent work among prior works but missed in this paper.\n\n- The proposed method simply provides the additional information about augmentations together with the encoder outputs, and it is not clear how it helps to \"preserve more information about augmentations\" in representations. Figure 3 shows that injecting information of random augmentations results in reduced cosine similarities. This implies that the projector relies on the given information about augmentations, which is not directly related to the learned representations (the output of the encoder), i.e., learned representations do not have to be changed regardless of whether the projector relies on the additional information about augmentations or not. Any theoretical justification on the effect of the proposed method to the learned representations would be welcome.\n\n- The performance gain is overall minor and often it underperforms previous methods.\n\n- Why does MoCo-v2 in Table 1 contain only one performance of LooC? It looks quite not informative.\n\n- Why does MoCo-v3 in Table 1 miss the performance of \"AI by [Chavhan et al.],\" while the original paper presents its performance?\n\n- The reference section requires thorough proofreading, as there are many incomplete/inaccurate references. For example, the closest prior work by [Lee et al.] is published in NeurIPS'21, but its arXiv version is cited. Also, many references miss the name of the published venue.\n\n[Chavhan et al.] Amortised invariance learning for contrastive self-supervision. In ICLR, 2023.\n\n[Garrido et al.] Self-supervised learning of Split Invariant Equivariant representations. In ICML, 2023.", "questions": "Please address concerns in Weaknesses.\n\n> **post rebuttal**\n\nAs some of my concerns remain after discussion with authors, I keep my original rating unchanged with more confidence.\n\n- I think your experiments can only be used to \"indirectly\" justify if the encoder preserves the information about augmentations. As I suggested, any theoretical analysis or a more direct experiment that checks if the proposed method better retains the augmentation information compared to the baseline would be helpful.\n\n- Only ResNet-50 is used throughout experiments, so my concern on the scalability still remains. Note that ViT is just one option to resolve this concern; you can use different number of layers or other types of CNN architectures. Generally speaking, ViT becomes prevalent in the last ~ 2 years, so experiments with this type of architecture would strengthen your contribution.\n\n- The reference section is still not proofread after being pointed out twice; at this point, I am not sure if authors are willing to show proper respect for previous works. MoCo v2 is an arXiv preprint and MoCo v3 is published in ICCV'21, but both miss their venues.\n\nXinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum\ncontrastive learning, 2020b.\n\nXinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision\ntransformers, October 2021b.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698768575373}], "openreview_url": "https://openreview.net/forum?id=WO4BCqEyWc", "arxiv_id": "2306.06082", "paper_pdf": "papers/WO4BCqEyWc.pdf", "paper_pdf_sha256": "8e00967f8258c9acb8baa2a1e2df372d1fe052f741e6e0519e618828ed555c68", "paper_pdf_bytes": 6528454, "paper_pdf_source": "openreview", "code_url": "https://github.com/gmum/CASSLE", "code_repository": "gmum/CASSLE", "code_commit": "6e345133c7525644f2dd98cf9324b55d0014fbde", "code_archive": "repos/WO4BCqEyWc.zip", "code_archive_sha256": "5ea97108385012ef2db88d8dd18e5ab33dd2166d8d29575861784873ae50441f", "code_archive_bytes": 356973, "code_file_count": 20, "code_extensions": {".py": 20}, "github_disk_usage_kb": 470, "github_languages": {"Python": 230379}, "github_archived": false, "github_pushed_at": "2026-07-16T08:52:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/augmentation-aware-self-supervised-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "vNrmEgfGIg3", "year": 2023, "status": "rejected", "title": "Filtered Semi-Markov CRF", "authors": ["Urchade Zaratiana", "Nadi Tomeh", "Niama Elkhbir", "Pierre Holat", "Thierry Charnois"], "authorids": ["~Urchade_Zaratiana1", "~Nadi_Tomeh1", "~Niama_Elkhbir1", "~Pierre_Holat1", "~Thierry_Charnois2"], "authors_source": "OpenReview API", "abstract": "Semi-Markov CRF \\citep{semicrf} has been proposed as an alternative to the traditional Linear Chain CRF\\citep{crf} for text segmentation tasks such as Named Entity Recognition. In contrast to CRF, which treats text segmentation as token-level prediction, Semi-CRF considers spans as the task's basic unit, which makes it more expressive. However, Semi-CRF has two major drawbacks: (1) it has quadratic complexity over sequence length as it operates on every span of the input sequence, and (2) empirically, it performs worse than classical CRF for sequence labeling tasks such as NER. In our work, we propose Filtered Semi-Markov CRF, a Semi-CRF variant that addresses the aforementioned issues. Our model extends Semi-CRF by incorporating a filtering step for eliminating irrelevant segments, which helps in reducing the complexity and allows to dramatically reduce the search space. On a variety of NER benchmarks, we find that our approach outperforms both CRF and Semi-CRF models while being significantly faster. We will make our code available to the public. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "2Imhof2H3M", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6275/Reviewer_z7Y1"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper proposes a semi-Markov conditional random field model that integrates a filtering step to eliminate irrelevant segments when performing named entity recognition in text. According to the authors this helps reducing the complexity compared to a semi-Markov CRF and dramatically reduces the search space. \n", "review_text": "The paper is below the threshold for acceptance at ICLR because of:\n- Important details and evaluations are missing.\n\nSeveral of my questions were answered during the rebuttal, for which I thank the authors. ", "strengths": "Strengths\n\nAttempt to learn segmentation jointly with classification (here named entity recognition or NER) that does not rely on an expansion of the label space (as is commonly done in NER by specializing the labels into begin label, intermediate and outside label).\n\nWeaknesses\n\n-\tThe model relies on an extra classifier that filters segments. The model is jointly trained by considering the filter classification loss, the segmentation loss and the NER classification loss. It is not clear from the paper how during training the computational complexity is reduced as initially all potential segments need to be considered in the inference step during training. That means that the complexity is equal to the complexity of the semi-Markov model.  The authors acknowledge that “during training, especially in the first stage, the graph size can be large since the filtering is poor.” It is not clear how exactly the filtering is performed during training. \n-\tBecause the segment filtering is the contribution of the paper, its approach (which should be explained in detail) should be separately evaluated, that is, its behavior and reduction in complexity during training and ablating the assumptions made here. \n-\tResults could have been evaluated on many more NER datasets. \n-\tThe claim that the proposed filtering model drastically reduces the search space compared to a CRF and SemiCRF model is not seen in the model’s throughput numbers, especially not during training. Moreover, the effect of dynamic programming on the model’s throughput during inference on the test data could be investigated. Is dynamic programming also used in the decoding of the CRF and SemiCRF models?\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposes a semi-Markov conditional random field model that integrates a filtering step to eliminate irrelevant segments when performing named entity recognition in text. According to the authors this helps reducing the complexity compared to a semi-Markov CRF and dramatically reduces the search space. \n", "strength_and_weaknesses": "Strengths\n\nAttempt to learn segmentation jointly with classification (here named entity recognition or NER) that does not rely on an expansion of the label space (as is commonly done in NER by specializing the labels into begin label, intermediate and outside label).\n\nWeaknesses\n\n-\tThe model relies on an extra classifier that filters segments. The model is jointly trained by considering the filter classification loss, the segmentation loss and the NER classification loss. It is not clear from the paper how during training the computational complexity is reduced as initially all potential segments need to be considered in the inference step during training. That means that the complexity is equal to the complexity of the semi-Markov model.  The authors acknowledge that “during training, especially in the first stage, the graph size can be large since the filtering is poor.” It is not clear how exactly the filtering is performed during training. \n-\tBecause the segment filtering is the contribution of the paper, its approach (which should be explained in detail) should be separately evaluated, that is, its behavior and reduction in complexity during training and ablating the assumptions made here. \n-\tResults could have been evaluated on many more NER datasets. \n-\tThe claim that the proposed filtering model drastically reduces the search space compared to a CRF and SemiCRF model is not seen in the model’s throughput numbers, especially not during training. Moreover, the effect of dynamic programming on the model’s throughput during inference on the test data could be investigated. Is dynamic programming also used in the decoding of the CRF and SemiCRF models?\n", "clarity,_quality,_novelty_and_reproducibility": "The clarity (and consequently the reproducibility) could be improved (see remark above). The originality is difficult to judge given that the details of the contribution are missing. ", "summary_of_the_review": "The paper is below the threshold for acceptance at ICLR because of:\n- Important details and evaluations are missing.\n\nSeveral of my questions were answered during the rebuttal, for which I thank the authors. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667817830410}, {"id": "A8t_RaDWkUy", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6275/Reviewer_JeoQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper focuses on reducing the inference complexity of semi-Markov CRF and thus achieving better performance on NER by introducing a filtering step before forward algorithm and Viterbi decoding. More specifically, segments that are predicted to be null and whose label does not achieve the highest predicted score according to an additional local classifier will be filtered. Experiments are conducted on the task of NER and evaluated on three datasets: CONLL-2003, OntoNotes 5.0 and Arabic ACE.", "review_text": "In general, this paper focuses on an important problem w.r.t. SemiCRF which prevents its widespread application as CRF and proposes an applicable method for resolving it. However, this method lacks thorough analysis (especially experiments), which weakens the claims made in the paper.", "strengths": "Strength:\n1. Experiment results on performance and throughput verify the claim that Filtered Semi-Markov CRF are better than CRF and Semi-Markov CRF on NER tasks.\n\nWeakness:\n1. Some parts of the paper is a little hard to follow, especially Section 3. For example, what is the model architecture of local classifier compared to the global ones?  Do they share parameters? Besides, without reading the rest part in Section 3, it is hard to recognize that Eq. (6) chooses segments whose labels are not predicted to null but also achieves the highest scores. There also exist some typos. For example, in the 4th row from the bottom of Page 3, $j_k<i_{k^*}$, which should be $i_{k^*} < j_k$.\n\n2. Lack of detailed analysis about why filtering leads to better performance. FSemiCRF does Viterbi decoding on a subset of segments compared to the original SemiCRF, whose optimum should be worse than that of SemiCRF in theory. I'm curious whether the introducing of local filtering models and corresponding segment filtering in training or segment filtering during inference is the primary cause of performance improvement. an ablation study and analysis will be helpful. For example, how FSemiCRF performs if it is trained as described in Section 3 but still follows original SemiCRF (i.e., run Viterbi decoding on all the segments) during inference?\n\n3. Model throughput needs further explanation. During inference, both SemiCRF and FSemiCRF have two steps: (1) to calculate segment scores for all segments; (2) Viterbi decoding, in which SemiCRF does Viterbi decoding directly while FSemiCRF runs segment filtering and does Viterbi decoding on the rest segments. It will be better to show which step is the bottleneck (e.g., to show empirical wall-clock time of these two steps) and how FSemiCRF improves SemiCRF in the second step (e.g., to show wall-clock time of segment filtering and Viterbi decoding on the rest segments).\n\n4. What about training complexity (e.g., wall-clock time) of FSemiCRF compared to SemiCRF.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper focuses on reducing the inference complexity of semi-Markov CRF and thus achieving better performance on NER by introducing a filtering step before forward algorithm and Viterbi decoding. More specifically, segments that are predicted to be null and whose label does not achieve the highest predicted score according to an additional local classifier will be filtered. Experiments are conducted on the task of NER and evaluated on three datasets: CONLL-2003, OntoNotes 5.0 and Arabic ACE.", "strength_and_weaknesses": "Strength:\n1. Experiment results on performance and throughput verify the claim that Filtered Semi-Markov CRF are better than CRF and Semi-Markov CRF on NER tasks.\n\nWeakness:\n1. Some parts of the paper is a little hard to follow, especially Section 3. For example, what is the model architecture of local classifier compared to the global ones?  Do they share parameters? Besides, without reading the rest part in Section 3, it is hard to recognize that Eq. (6) chooses segments whose labels are not predicted to null but also achieves the highest scores. There also exist some typos. For example, in the 4th row from the bottom of Page 3, $j_k<i_{k^*}$, which should be $i_{k^*} < j_k$.\n\n2. Lack of detailed analysis about why filtering leads to better performance. FSemiCRF does Viterbi decoding on a subset of segments compared to the original SemiCRF, whose optimum should be worse than that of SemiCRF in theory. I'm curious whether the introducing of local filtering models and corresponding segment filtering in training or segment filtering during inference is the primary cause of performance improvement. an ablation study and analysis will be helpful. For example, how FSemiCRF performs if it is trained as described in Section 3 but still follows original SemiCRF (i.e., run Viterbi decoding on all the segments) during inference?\n\n3. Model throughput needs further explanation. During inference, both SemiCRF and FSemiCRF have two steps: (1) to calculate segment scores for all segments; (2) Viterbi decoding, in which SemiCRF does Viterbi decoding directly while FSemiCRF runs segment filtering and does Viterbi decoding on the rest segments. It will be better to show which step is the bottleneck (e.g., to show empirical wall-clock time of these two steps) and how FSemiCRF improves SemiCRF in the second step (e.g., to show wall-clock time of segment filtering and Viterbi decoding on the rest segments).\n\n4. What about training complexity (e.g., wall-clock time) of FSemiCRF compared to SemiCRF.", "clarity,_quality,_novelty_and_reproducibility": "The idea of accelerating the inference of SemiCRF is novel and the proposed filtering method is well-motivated. However, current analysis and experiment results do not support the claim well and more detailed analysis is needed. ", "summary_of_the_review": "In general, this paper focuses on an important problem w.r.t. SemiCRF which prevents its widespread application as CRF and proposes an applicable method for resolving it. However, this method lacks thorough analysis (especially experiments), which weakens the claims made in the paper.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666755094552}, {"id": "dl5LjZshuxk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6275/Reviewer_YFd7"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "A method is proposed for filtering spans to be labeled in a semi-Markov CRF, which empirically reduces training/inference complexity and increases evaluation metrics relative to basic CRF or basic semi-CRF.", "review_text": "The proposed method does well in ablation experiments (reducing to CRF, or expanding to full semi-CRF) but some obvious baselines were neglected (greedy decoding over Transformer, bounding maximum span of semiCRF).", "strengths": "The paper is clearly written. The method is, to my knowledge, original.\n\nIt does not compare to state-of-the-art methods, which generally use something like transformer decoders with greedy decoding. (https://paperswithcode.com/sota/named-entity-recognition-ner-on-conll-2003, https://paperswithcode.com/sota/named-entity-recognition-ner-on-ontonotes-v5, https://paperswithcode.com/sota/named-entity-recognition-on-ace-2005) The results are a bit worse than SOTA.\n\nThe motivation is to address the shortcoming of semiCRF that it considers all spans. It turns out that the proposed method is not only faster but gets better evaluation scores than pure semi-CRF. (The authors might want to speculate on why that is-- it isn't clear to me a priori that the proposed method should be more accurate.) However the proposed method has some shortcomings: it introduces a hyperparameter beta, it has the possibility of discarding good spans in the filtering phase, and the worst-case complexity is higher. For many tasks it is possible to give a reasonable upper bound on the span length that needs to be considered, so complexity of semiCRF training/inference is reduced to linear from quadratic in sentence length. What about a baseline that uses a maximum segment length and allows null labels only for spans of length 1? I think that should address the weaknesses of semiCRF in a cleaner way and I would not be surprised to see it working quite well in terms of evaluation.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "A method is proposed for filtering spans to be labeled in a semi-Markov CRF, which empirically reduces training/inference complexity and increases evaluation metrics relative to basic CRF or basic semi-CRF.", "strength_and_weaknesses": "The paper is clearly written. The method is, to my knowledge, original.\n\nIt does not compare to state-of-the-art methods, which generally use something like transformer decoders with greedy decoding. (https://paperswithcode.com/sota/named-entity-recognition-ner-on-conll-2003, https://paperswithcode.com/sota/named-entity-recognition-ner-on-ontonotes-v5, https://paperswithcode.com/sota/named-entity-recognition-on-ace-2005) The results are a bit worse than SOTA.\n\nThe motivation is to address the shortcoming of semiCRF that it considers all spans. It turns out that the proposed method is not only faster but gets better evaluation scores than pure semi-CRF. (The authors might want to speculate on why that is-- it isn't clear to me a priori that the proposed method should be more accurate.) However the proposed method has some shortcomings: it introduces a hyperparameter beta, it has the possibility of discarding good spans in the filtering phase, and the worst-case complexity is higher. For many tasks it is possible to give a reasonable upper bound on the span length that needs to be considered, so complexity of semiCRF training/inference is reduced to linear from quadratic in sentence length. What about a baseline that uses a maximum segment length and allows null labels only for spans of length 1? I think that should address the weaknesses of semiCRF in a cleaner way and I would not be surprised to see it working quite well in terms of evaluation.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: mostly high\nQuality: a bit weak\nNovelty: probably new\nReproducibility: good", "summary_of_the_review": "The proposed method does well in ablation experiments (reducing to CRF, or expanding to full semi-CRF) but some obvious baselines were neglected (greedy decoding over Transformer, bounding maximum span of semiCRF).", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666727638386}], "openreview_url": "https://openreview.net/forum?id=vNrmEgfGIg3", "arxiv_id": "2311.18028", "paper_pdf": "papers/vNrmEgfGIg3.pdf", "paper_pdf_sha256": "25e7e8eb6bdb4345bd9c55f1f48d59346e88a55768a9a3859d15455ac2ca58fa", "paper_pdf_bytes": 524729, "paper_pdf_source": "openreview", "code_url": "https://github.com/urchade/Filtered-Semi-Markov-CRF", "code_repository": "urchade/Filtered-Semi-Markov-CRF", "code_commit": "77da8cdc5f3ab8fbe805dce115e67688b11fd130", "code_archive": "repos/vNrmEgfGIg3.zip", "code_archive_sha256": "f90b3b684ccd1cf84fc3408f47ffab751559ac0b5e9afec9ab230ca83c9756fb", "code_archive_bytes": 1024684, "code_file_count": 15, "code_extensions": {".py": 15}, "github_disk_usage_kb": 1223, "github_languages": {"Python": 48627}, "github_archived": false, "github_pushed_at": "2024-01-05T19:48:13Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/filtered-semi-markov-crf"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "X3WxnuzAYyE", "year": 2022, "status": "rejected", "title": "PKCAM: Previous Knowledge Channel Attention Module", "authors": ["Eslam Mohamed BAKR", "Ahmad A. Al Sallab", "Mohsen Rashwan"], "authorids": ["~Eslam_Mohamed_BAKR1", "~Ahmad_A._Al_Sallab1", "~Mohsen_Rashwan1"], "authors_source": "OpenReview API", "abstract": "Attention mechanisms have been explored with CNNs, both across the spatial and channel dimensions. \nHowever, all the existing methods devote the attention modules to capture local interactions from the current feature map only, disregarded the valuable previous knowledge that is acquired by the earlier layers. \nThis paper tackles the following question: Can one incorporate previous knowledge aggregation while learning channel attention more efficiently? To this end, we propose a Previous Knowledge Channel Attention Module( PKCAM), that captures channel-wise relations across different layers to model the global context. \nOur proposed module PKCAM is easily integrated into any feed-forward CNN architectures and trained in an end-to-end fashion with a negligible footprint due to its lightweight property. We validate our novel architecture through extensive experiments on image classification and object detection tasks with different backbones. \nOur experiments show consistent improvements in performances against their counterparts. We also conduct experiments that probe the robustness of the learned representations.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "iAxDsbcJZXQ", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4324/Reviewer_cAgA"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work proposes a previous knowledge channel attention module (PKCAM) that captures channel relations across the different layers to help enhance feature representation. The module can be integrated into the current ResNet series and show reasonable performance improvement over the baseline network on some benchmark datasets.", "review_text": "I think the descriptions about the core PKCAM module are unclear and seem to be incorrect. As shown on Page 4,  Y=W\\widetilde(X) and thus Y should be a feature vector with the dimension of RC_0. However, the authors claim that Y is a feature vector with dimension of C_0. Do I miss something? Besides, the authors state to split the function f() into two functions f_1() and f_2(). However, I can not find the definition of f_2(). \n\nThe authors split the PKCAN module into a summation followed by a linear mapping. It can indeed reduce the complexity. But why such simplicity works? More analysis should be provided.\n\nThe authors claim that the module can be easily integrated into different network architectures. However, they merely conduct experiments with the Reset series as baselines. I think the experiments with other architecture s such as mobile net should also be provided. Or the experiments can not support the statements.\n\nMore recent works [1, 2] also adopt channel and spatial attention for feature enhancement. These works should also be included for analysis and comparisons.\n\nI think the work is somewhat a draft for the current version. There are many reference errors throughout the paper, e.g., fig. ?? on Page 3 and fig. ?? on Page 7.\n[1] Hou et al., Coordinate Attention for Efficient Mobile Network Design Coordinate Attention, in CVPR, 2021.\n[2] Zhang et al., ResNeSt: Split-Attention Networks, in arXiv 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes a previous knowledge channel attention module (PKCAM) that captures channel relations across the different layers to help enhance feature representation. The module can be integrated into the current ResNet series and show reasonable performance improvement over the baseline network on some benchmark datasets.", "main_review": "I think the descriptions about the core PKCAM module are unclear and seem to be incorrect. As shown on Page 4,  Y=W\\widetilde(X) and thus Y should be a feature vector with the dimension of RC_0. However, the authors claim that Y is a feature vector with dimension of C_0. Do I miss something? Besides, the authors state to split the function f() into two functions f_1() and f_2(). However, I can not find the definition of f_2(). \n\nThe authors split the PKCAN module into a summation followed by a linear mapping. It can indeed reduce the complexity. But why such simplicity works? More analysis should be provided.\n\nThe authors claim that the module can be easily integrated into different network architectures. However, they merely conduct experiments with the Reset series as baselines. I think the experiments with other architecture s such as mobile net should also be provided. Or the experiments can not support the statements.\n\nMore recent works [1, 2] also adopt channel and spatial attention for feature enhancement. These works should also be included for analysis and comparisons.\n\nI think the work is somewhat a draft for the current version. There are many reference errors throughout the paper, e.g., fig. ?? on Page 3 and fig. ?? on Page 7.\n[1] Hou et al., Coordinate Attention for Efficient Mobile Network Design Coordinate Attention, in CVPR, 2021.\n[2] Zhang et al., ResNeSt: Split-Attention Networks, in arXiv 2020.", "summary_of_the_review": "I think the introduction about the PKCAM module is very unclear. The experiments are not sufficient and convincing.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635845190434}, {"id": "Krjnv1xYj2X", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4324/Reviewer_2Qrr"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper computes channel attention by considering feature maps across different layers, namely previous knowledge channel attention module (PKCAM).  The developed PKCAM is achieved by two steps: (1) the previous knowledge aggravation module is used to aggregate channel information of multiple feature images; and (2) the global cross channel interaction module models the correlation among channels to compute attention weights. The experiments are conducted on image classification and object detection tasks. ", "review_text": "Strengths: \n1. The previous feature maps are used to compute channel attention for the current feature map, showing some differences with existing methods. Besides, the proposed PKCAM is only applied to the end of the last stages, leading small parameters and FLOPs.\n2.\tThe proposed method seems easy to implement. \n\nWeaknesses:\n1.\tThe technical contribution of this work seems limited. The core of PKCAM is to exploit feature maps from previous layers for computing channel attention of the current feature map. For me, it is still not very clear why previous feature maps can help to compute channel attention of the current feature map. In other word, what are the merits brought by the previous feature maps? Furthermore, the core components of PKCAM (i.e., SDA and CDA) show similar with existing works, e.g., SE and ECA blocks.\n2.\tThe experimental results are not very convincing. Specifically,\n(1)\tThe improvement of PKCAM over existing methods is very marginal in terms of accuracy and model complexity, which hardly verify the effectiveness of previous knowledge.\n(2)\tExperiments could be further strengthened. More stronger backbones (e.g., ResNet-101 for classification and Faster R-CNN/ Mask R-CNN for object detection) could be used to further evaluate performance of the proposed method.\n(3)\tSome experimental results need further discussions. For example, why FPS of ECANet increases two times from ResNet-34 to ResNet-50, while others only increase less 1.5 times? Why LCAM+PKCAM is inferior to PKCAM? How to combine LCAM with PKCAM? What are the details of LCAM?\n(4)\tWhy PKCAM is only applied to the end of each stage? Why PKCAM only inserted into the last stage achieves the best results?\n3.\tThe writing could be further polished. Specifically,\n(1)\tIn Section 3.2, the details on how to select the previous knowledge seem missing, i.e., How to select R feature maps? and How about effect of number of R?\n(2)\tWhich method is used for channel dimension alignment?\n(3)\tThe orders of channel dimension alignment and spatial dimension alignment described in Eqn.(2) (from 1 to R) are inconsistent with those in Figure 1 (from 0 to R).\n(4)\tThe dimension of fully-connected layer in Eqn.(1) seems confusing.\n(5)\tThe sum operations in Eqn.(2) and Eqn.(3) are confusing.\n(6)\tMany citations are missing, e.g., Figure ???\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper computes channel attention by considering feature maps across different layers, namely previous knowledge channel attention module (PKCAM).  The developed PKCAM is achieved by two steps: (1) the previous knowledge aggravation module is used to aggregate channel information of multiple feature images; and (2) the global cross channel interaction module models the correlation among channels to compute attention weights. The experiments are conducted on image classification and object detection tasks. ", "main_review": "Strengths: \n1. The previous feature maps are used to compute channel attention for the current feature map, showing some differences with existing methods. Besides, the proposed PKCAM is only applied to the end of the last stages, leading small parameters and FLOPs.\n2.\tThe proposed method seems easy to implement. \n\nWeaknesses:\n1.\tThe technical contribution of this work seems limited. The core of PKCAM is to exploit feature maps from previous layers for computing channel attention of the current feature map. For me, it is still not very clear why previous feature maps can help to compute channel attention of the current feature map. In other word, what are the merits brought by the previous feature maps? Furthermore, the core components of PKCAM (i.e., SDA and CDA) show similar with existing works, e.g., SE and ECA blocks.\n2.\tThe experimental results are not very convincing. Specifically,\n(1)\tThe improvement of PKCAM over existing methods is very marginal in terms of accuracy and model complexity, which hardly verify the effectiveness of previous knowledge.\n(2)\tExperiments could be further strengthened. More stronger backbones (e.g., ResNet-101 for classification and Faster R-CNN/ Mask R-CNN for object detection) could be used to further evaluate performance of the proposed method.\n(3)\tSome experimental results need further discussions. For example, why FPS of ECANet increases two times from ResNet-34 to ResNet-50, while others only increase less 1.5 times? Why LCAM+PKCAM is inferior to PKCAM? How to combine LCAM with PKCAM? What are the details of LCAM?\n(4)\tWhy PKCAM is only applied to the end of each stage? Why PKCAM only inserted into the last stage achieves the best results?\n3.\tThe writing could be further polished. Specifically,\n(1)\tIn Section 3.2, the details on how to select the previous knowledge seem missing, i.e., How to select R feature maps? and How about effect of number of R?\n(2)\tWhich method is used for channel dimension alignment?\n(3)\tThe orders of channel dimension alignment and spatial dimension alignment described in Eqn.(2) (from 1 to R) are inconsistent with those in Figure 1 (from 0 to R).\n(4)\tThe dimension of fully-connected layer in Eqn.(1) seems confusing.\n(5)\tThe sum operations in Eqn.(2) and Eqn.(3) are confusing.\n(6)\tMany citations are missing, e.g., Figure ???\n", "summary_of_the_review": "In my opinion, the technical contribution of this work seems limited, while experiments and writing could be further polished. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635752757986}, {"id": "JnE2ujxRtn", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4324/Reviewer_BFC2"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper explores a new kind of attention mechanism, which can be easily integrated into any cnn models.\n\nThe proposed method is called previous knowledge channel attention module, whose aim is to make use of the knowledge from the previous layers.\n\nThe methods are validated on the image classification and object detection tasks. \n", "review_text": "This paper proposes an attention module with the knowledge from the previous layers also included. I have two concerns about this idea.\n\nFirstly, the idea to use the previous knowledge is not new. In the works of self-knowledge distillation, they use the previous layers knowledge to do the distillation (i.e. Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation). So, the idea of utilizing previous knowledge is not new and thus the contribution, for me, is not enough.\n\nSecondly, the authors use the previous knowledge to build a better attention module. This idea is too intuitive. I don’t think the attention module has long-term relationship. \n\n\nSome small typos in writing:\n- The last line of page 3, there is something wrong with the figure (actually, all the citations for figures are wrong)\n- The second row of sec 3.2, which rely on…\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper explores a new kind of attention mechanism, which can be easily integrated into any cnn models.\n\nThe proposed method is called previous knowledge channel attention module, whose aim is to make use of the knowledge from the previous layers.\n\nThe methods are validated on the image classification and object detection tasks. \n", "main_review": "This paper proposes an attention module with the knowledge from the previous layers also included. I have two concerns about this idea.\n\nFirstly, the idea to use the previous knowledge is not new. In the works of self-knowledge distillation, they use the previous layers knowledge to do the distillation (i.e. Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation). So, the idea of utilizing previous knowledge is not new and thus the contribution, for me, is not enough.\n\nSecondly, the authors use the previous knowledge to build a better attention module. This idea is too intuitive. I don’t think the attention module has long-term relationship. \n\n\nSome small typos in writing:\n- The last line of page 3, there is something wrong with the figure (actually, all the citations for figures are wrong)\n- The second row of sec 3.2, which rely on…\n", "summary_of_the_review": "-\tWhat is the main contribution or motivation of this paper. If it is only about using the previous knowledge, it is not creative enough.\n-\tTry to explain or proof the working mechanism of the proposed PKCAM, which should be made more solid.\n-\tTry to polish the paper.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635698373899}, {"id": "t243j0nI1G", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper4324/Reviewer_HH21"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper introduces an attention module, namely PKCAM for Previous Knowledge Channel Attention Module, which can be integrated into CNNs. The proposed PKCAM mainly aggregates the feature maps of earlier CNN blocks for aggregating previous knowledge. The ablation study is conducted for validating the effect of PKCAM in CNNs as well as its different design. Image classification and object detection experiments are considered for demonstrating the effectiveness of PKCAM in different ResNet architectures.", "review_text": "This paper is not well prepared for review with many incomplete and inconsistent expressions, e.g., the Fig.?? in Pages 3 and 7, the EMCA in Figure 2 and Figure 3 but the PKCAM in the captions, the CDA and SDA in Figure 1 without any explanation in the caption and text, the first SDA then CDA operation in both Figure 1 and Algorithm 1 but inverse order in Section 3.2, etc.\n\nThe key contribution of this work is to consider aggregating the previous feature maps from earlier CNN blocks, but this idea has already been proposed for CNNs such as the DenseNet. Besides, the implementation of PKCAM is not special and actually the commonly used spatial and channel attention. Thus the paper has limited technical novelty.\n\nFor the ablation study, most of the analysis is actually related to the current techniques used in attention design (see Table 2). Besides, most of the comparison results show that the PKCAM is not superior to the compared counterparts a lot (refer to Tables 4, 5, 7, and 8 as well as Figure 2). Therefore, the advantage of PKCAM is not obvious, that is, this work is not convincing empirically.\n\nFrom Figure 3, it can be seen that the PKCAM has a big difference with ECA, however, the quantitative comparison shows very small difference between PKCAM and ECANet, while the explanation of the last paragraph in Section 5 is not reasonable, especially how to evaluate the discriminative and representative ability of an attention module. \n\nThe paper slightly exceeds the paper limit of ICLR, and the writing needs to be carefully revised for typos and grammar errors.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces an attention module, namely PKCAM for Previous Knowledge Channel Attention Module, which can be integrated into CNNs. The proposed PKCAM mainly aggregates the feature maps of earlier CNN blocks for aggregating previous knowledge. The ablation study is conducted for validating the effect of PKCAM in CNNs as well as its different design. Image classification and object detection experiments are considered for demonstrating the effectiveness of PKCAM in different ResNet architectures.", "main_review": "This paper is not well prepared for review with many incomplete and inconsistent expressions, e.g., the Fig.?? in Pages 3 and 7, the EMCA in Figure 2 and Figure 3 but the PKCAM in the captions, the CDA and SDA in Figure 1 without any explanation in the caption and text, the first SDA then CDA operation in both Figure 1 and Algorithm 1 but inverse order in Section 3.2, etc.\n\nThe key contribution of this work is to consider aggregating the previous feature maps from earlier CNN blocks, but this idea has already been proposed for CNNs such as the DenseNet. Besides, the implementation of PKCAM is not special and actually the commonly used spatial and channel attention. Thus the paper has limited technical novelty.\n\nFor the ablation study, most of the analysis is actually related to the current techniques used in attention design (see Table 2). Besides, most of the comparison results show that the PKCAM is not superior to the compared counterparts a lot (refer to Tables 4, 5, 7, and 8 as well as Figure 2). Therefore, the advantage of PKCAM is not obvious, that is, this work is not convincing empirically.\n\nFrom Figure 3, it can be seen that the PKCAM has a big difference with ECA, however, the quantitative comparison shows very small difference between PKCAM and ECANet, while the explanation of the last paragraph in Section 5 is not reasonable, especially how to evaluate the discriminative and representative ability of an attention module. \n\nThe paper slightly exceeds the paper limit of ICLR, and the writing needs to be carefully revised for typos and grammar errors.", "summary_of_the_review": "The paper is not ready for review, also the contribution and analysis are limited. Thus, I vote for rejecting.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635688809079}], "openreview_url": "https://openreview.net/forum?id=X3WxnuzAYyE", "arxiv_id": "2211.07521", "paper_pdf": "papers/X3WxnuzAYyE.pdf", "paper_pdf_sha256": "99fb3ad6524e4930e8663b5ecf321a061f4fd50035253a0fc3a41163075674d4", "paper_pdf_bytes": 1128222, "paper_pdf_source": "openreview", "code_url": "https://github.com/eslambakr/EMCA", "code_repository": "eslambakr/EMCA", "code_commit": "76bae427ea2f66979aed6a9dcef5e84ed922d14e", "code_archive": "repos/X3WxnuzAYyE.zip", "code_archive_sha256": "d62d9ceb8af92d4b877636f76f7f7a63d1cfef19891d0dde257f65ccc6e294f6", "code_archive_bytes": 2366697, "code_file_count": 28, "code_extensions": {".py": 28}, "github_disk_usage_kb": 5864, "github_languages": {"Python": 294765}, "github_archived": false, "github_pushed_at": "2021-11-24T18:02:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pkcam-previous-knowledge-channel-attention-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "E4PK0rg2eP", "year": 2021, "status": "rejected", "title": "Parameter-Efficient Transfer Learning with Diff Pruning", "authors": ["Demi Guo", "Alexander M Rush", "Yoon Kim"], "authorids": ["~Demi_Guo1", "~Alexander_M_Rush1", "~Yoon_Kim1"], "authors_source": "OpenReview API", "abstract": "While task-specific finetuning of deep networks pretrained with self-supervision has led to significant empirical advances in NLP, their large size makes the standard finetuning approach difficult to apply to multi-task, memory-constrained settings, as storing the full model parameters for each task become prohibitively expensive. We propose $\\textit{diff pruning}$ as a simple approach to enable parameter-efficient transfer learning within the pretrain-finetune framework. This approach views finetuning as learning a task-specific diff vector that is applied on top of the pretrained parameter vector, which remains fixed and is shared across different tasks. The diff vector is adaptively pruned during training with a differentiable approximation to the $L_0$-norm penalty to encourage sparsity. Diff pruning becomes parameter-efficient as the number of tasks increases, as it requires storing only the nonzero positions and weights of the diff vector for each task, while the cost of storing the shared pretrained model remains constant. We find that models finetuned with diff pruning can match the performance of fully finetuned baselines on the GLUE benchmark while only modifying 0.5$\\%$ of the pretrained model's parameters per task.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "h7S5ugPJwKK", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1842/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work studies the problem of parameter-efficient transfer learning in the paradigm of pretraining/finetuning. The proposed method, diff pruning, can match the performance of fully finetuned baselines on the GLUE benchmark while only modifying 0.5 of the pretrained model's parameters per task.\n\nI don't get the problem definition of this work. In other words, the authors need to better motivate and justify parameter efficiency. Given that all the parameters of a big model are used in downstream tasks, what are the benefits of only modifying a few parameters? Here are several possibilities.\n1. Does it speed up the finetuning process by modifying a few parameters? \n2. Does it speed up inference for downstream tasks?\n3. Another possibility is model size reduction, which leads to reduced storage costs. \"This approach can become parameter-efficient as the number of tasks increases as it only requires storing the nonzero positions and weights of the diff vector for each task. The cost of storing the shared pretrained model remains constant and is amortized across multiple tasks.\" Unfortunately, in real-world scenarios, inference of  large models like BERT_base/large is usually conducted on servers or in cloud, where storage is not a big issue while latency (speed) is critical. Inference in edge devices like phones, IoT sensors cares about storage, speed, and power consumption. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Concern about problem setting", "review": "This work studies the problem of parameter-efficient transfer learning in the paradigm of pretraining/finetuning. The proposed method, diff pruning, can match the performance of fully finetuned baselines on the GLUE benchmark while only modifying 0.5 of the pretrained model's parameters per task.\n\nI don't get the problem definition of this work. In other words, the authors need to better motivate and justify parameter efficiency. Given that all the parameters of a big model are used in downstream tasks, what are the benefits of only modifying a few parameters? Here are several possibilities.\n1. Does it speed up the finetuning process by modifying a few parameters? \n2. Does it speed up inference for downstream tasks?\n3. Another possibility is model size reduction, which leads to reduced storage costs. \"This approach can become parameter-efficient as the number of tasks increases as it only requires storing the nonzero positions and weights of the diff vector for each task. The cost of storing the shared pretrained model remains constant and is amortized across multiple tasks.\" Unfortunately, in real-world scenarios, inference of  large models like BERT_base/large is usually conducted on servers or in cloud, where storage is not a big issue while latency (speed) is critical. Inference in edge devices like phones, IoT sensors cares about storage, speed, and power consumption. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603963766669}, {"id": "li-u1Vjqkgi", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1842/AnonReviewer2"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This work combines model pruning with transfer learning/multi-task learning in NLP. Instead of finetuning pretrained models on each individual task, the author propose to learn 'residual' parameters with sparse masks for each task independently, hence reducing per-task parameter requirements. The evaluation on GLUE shows some promising results with better efficiency compared to adapter networks. However, I am concerned with the motivation and applicability of the proposed method, such that I am left with the impression that the method may be hard to use in practice.\n\nPros:\n1. To the best of my knowledge, combining pruning on pretrained models to transfer learning in NLP is novel.\n2. Empirical results show the method can performance comparably with adapter networks but with fewer parameters.\n\nCons (see below for detailed questions):\n1. Although the method is proposed for 'multi-task' learning, it is not really evaluated on multi-task settings. The motivation and applicability are not very clear to me.\n2. Compared to adapter networks, the proposed method contains more hyper-parameters and is harder to tune in practice.\n\nQuestions:\n1. In section 2, the author introduced multi-task learning as the background. However, the proposed method is not really designed for multi-task learning (training multiple tasks simultaneously), but rather for transfer learning on multiple tasks independently. In practice, however, training some tasks together can improve performance (GLUE for example), but the proposed method is trained for each task independently and thus there is no positive transfer between tasks. So have you evaluated your method when trained multiple tasks simultaneously (where the shared parameters are jointly finetuned)? How does that work? This is an important extension that might be useful for the community.\n2. Please point it out if I miss this information, but how did you select your hyper-parameters for your approach? Besides, how sensitive they are compared to adapter networks?\n3. Prior work [1] has shown the resulting mask of pruning can be compared to evaluate task similarities. Did you have similar observations in your experiments?\n\nMissing reference:\nThe idea of adding 'residual' parameters for new tasks is not new in lifelong learning. It is good to mention [2] and [3]. [1] is also related of comparing pruning masks for pretrained models.\n\nTypo:\nPage 4, last line: QQC -> QQP\n\n[1] On negative interference in multilingual models: findings and a meta-learning treatment. Wang et al., EMNLP 2020.\n\n[2] Progressive neural networks. Rusu et al., arxiv 2016\n\n[3] BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. Wen et al., ICLR 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting results, but better motivation and more experiments can further improve", "review": "This work combines model pruning with transfer learning/multi-task learning in NLP. Instead of finetuning pretrained models on each individual task, the author propose to learn 'residual' parameters with sparse masks for each task independently, hence reducing per-task parameter requirements. The evaluation on GLUE shows some promising results with better efficiency compared to adapter networks. However, I am concerned with the motivation and applicability of the proposed method, such that I am left with the impression that the method may be hard to use in practice.\n\nPros:\n1. To the best of my knowledge, combining pruning on pretrained models to transfer learning in NLP is novel.\n2. Empirical results show the method can performance comparably with adapter networks but with fewer parameters.\n\nCons (see below for detailed questions):\n1. Although the method is proposed for 'multi-task' learning, it is not really evaluated on multi-task settings. The motivation and applicability are not very clear to me.\n2. Compared to adapter networks, the proposed method contains more hyper-parameters and is harder to tune in practice.\n\nQuestions:\n1. In section 2, the author introduced multi-task learning as the background. However, the proposed method is not really designed for multi-task learning (training multiple tasks simultaneously), but rather for transfer learning on multiple tasks independently. In practice, however, training some tasks together can improve performance (GLUE for example), but the proposed method is trained for each task independently and thus there is no positive transfer between tasks. So have you evaluated your method when trained multiple tasks simultaneously (where the shared parameters are jointly finetuned)? How does that work? This is an important extension that might be useful for the community.\n2. Please point it out if I miss this information, but how did you select your hyper-parameters for your approach? Besides, how sensitive they are compared to adapter networks?\n3. Prior work [1] has shown the resulting mask of pruning can be compared to evaluate task similarities. Did you have similar observations in your experiments?\n\nMissing reference:\nThe idea of adding 'residual' parameters for new tasks is not new in lifelong learning. It is good to mention [2] and [3]. [1] is also related of comparing pruning masks for pretrained models.\n\nTypo:\nPage 4, last line: QQC -> QQP\n\n[1] On negative interference in multilingual models: findings and a meta-learning treatment. Wang et al., EMNLP 2020.\n\n[2] Progressive neural networks. Rusu et al., arxiv 2016\n\n[3] BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. Wen et al., ICLR 2020.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603832707399}, {"id": "y-n1aaUhRhD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1842/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors present an interesting approach to learn task-specific models with only a few tunable parameters. They propose learning a diff vector with a sparsity constraint and then pruning the vector using magnitude pruning. They also impose a structured sparsity constraint by introducing a group penalty.\n\nThis work is interesting and important with the growing size of the pretrained models. It enables the model to learn a new task with relatively few parameters would be very beneficial. The experiments and analysis conducted by the authors is thorough.\n\nA few questions/thoughts which could improve the paper: \n* It would be good to list the epochs for training for each of the different approaches? Authors mention 2,3,4,5 epochs for training but would good to highlight the cost or savings of fine tuning with their approach.\n* In table 1, it would be also helpful to include some information about the finetuning steps/sec with the different approaches. This would help understand the tradeoff of memory vs compute for the proposed approach.\n* What’s the intuition for structured pruning performing better than non unstructured? Typically, enforcing some structure should have worse accuracy. If the structure seems to help, then the model should be able to learn structure without the group constraint as well.\n\nOverall, I recommend accepting the paper. As the size of pre-trained models is growing quite rapidly, research that investigates parameter sharing and adapting a pretrained model to a new task with few parameters is essential.  ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach to reduce task specific parameters in transfer learning", "review": "The authors present an interesting approach to learn task-specific models with only a few tunable parameters. They propose learning a diff vector with a sparsity constraint and then pruning the vector using magnitude pruning. They also impose a structured sparsity constraint by introducing a group penalty.\n\nThis work is interesting and important with the growing size of the pretrained models. It enables the model to learn a new task with relatively few parameters would be very beneficial. The experiments and analysis conducted by the authors is thorough.\n\nA few questions/thoughts which could improve the paper: \n* It would be good to list the epochs for training for each of the different approaches? Authors mention 2,3,4,5 epochs for training but would good to highlight the cost or savings of fine tuning with their approach.\n* In table 1, it would be also helpful to include some information about the finetuning steps/sec with the different approaches. This would help understand the tradeoff of memory vs compute for the proposed approach.\n* What’s the intuition for structured pruning performing better than non unstructured? Typically, enforcing some structure should have worse accuracy. If the structure seems to help, then the model should be able to learn structure without the group constraint as well.\n\nOverall, I recommend accepting the paper. As the size of pre-trained models is growing quite rapidly, research that investigates parameter sharing and adapting a pretrained model to a new task with few parameters is essential.  ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603830160265}, {"id": "keH6gGlHB9l", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1842/AnonReviewer4"], "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes diff pruning, an alternative paradigm for parameter-efficient transfer learning of pre-trained models. Similar to adapters, diff pruning leaves the body of the pre-trained model unchanged. Rather than inserting additional task-specific parameters into the pre-trained model, diff pruning adds reparameterizes the parameters of the transferred model $\\theta_\\tau$ by adding a diff vector $\\delta_\\tau$ to them: $\\theta_\\tau = \\theta_{\\text{pretrained}} + \\delta_\\tau$. Parameter efficiency is achieved by regularizing $\\theta_\\tau$ to be sparse. The authors achieve this by using a relaxed mask vector to approximate the $L_0$ norm. They also propose a way to control for a specific sparsity rate via projection onto the $L_0$ ball after training and to enforce group sparsity that takes the model's structure into account. The approach is evaluated on the GLUE benchmark where it achieves competitive performance to full fine-tuning a BERT Large model and adapters while being more parameter-efficient than both of them.\n\nPros:\n1. The proposed method is intuitive and the different modelling choices are principled and well motivated.\n2. The method achieves strong results. It is competitive with full fine-tuning and more parameter-efficient than adapters, the prevalent approach for parameter-efficient transfer learning.\n3. The authors show how to effectively control the sparsity rate and incorporating structure via group sparsity brings further gains.\n4. The authors conduct extensive analyses, which pre-empted many of my questions, such as the variation across different sparsity masks, sparsity patterns across different tasks, etc. Overall, the analyses shed additional light on the characteristics and preferences of different tasks in transfer learning.\n\nCons:\n1. The approach is potentially more complicated than the baseline, so it is important that the authors open-source their code.\n2. The diff vector is distributed over the entire set of parameters of the model rather than focused in a few layers. This makes it potentially harder to combine the diff vectors from different tasks as can be done with adapters (see e.g. https://arxiv.org/abs/2005.00247) and to compose multiple diff vectors. \n\nQuestions:\n1. Does a visualization of diff vectors of different tasks (such as using t-SNE) reveal any interesting patterns?\n2. Are there any transfer settings that the addition of task-specific parameters can model but inserting layer-specific transformations via adapters cannot (or vice versa)? Adapters have been used to transfer across modalities such as languages (see e.g. Pfeiffer et al. (2020), https://arxiv.org/abs/2005.00052) and I am wondering whether the same would be possible by adding task-specific parameters.\n3. How long does your approach take to converge in comparison to the baselines? How much longer do you need to fine-tune with non-zero masks for magnitude pruning for sparsity control? What is the performance benefit of this further fine-tuning?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "This paper proposes diff pruning, an alternative paradigm for parameter-efficient transfer learning of pre-trained models. Similar to adapters, diff pruning leaves the body of the pre-trained model unchanged. Rather than inserting additional task-specific parameters into the pre-trained model, diff pruning adds reparameterizes the parameters of the transferred model $\\theta_\\tau$ by adding a diff vector $\\delta_\\tau$ to them: $\\theta_\\tau = \\theta_{\\text{pretrained}} + \\delta_\\tau$. Parameter efficiency is achieved by regularizing $\\theta_\\tau$ to be sparse. The authors achieve this by using a relaxed mask vector to approximate the $L_0$ norm. They also propose a way to control for a specific sparsity rate via projection onto the $L_0$ ball after training and to enforce group sparsity that takes the model's structure into account. The approach is evaluated on the GLUE benchmark where it achieves competitive performance to full fine-tuning a BERT Large model and adapters while being more parameter-efficient than both of them.\n\nPros:\n1. The proposed method is intuitive and the different modelling choices are principled and well motivated.\n2. The method achieves strong results. It is competitive with full fine-tuning and more parameter-efficient than adapters, the prevalent approach for parameter-efficient transfer learning.\n3. The authors show how to effectively control the sparsity rate and incorporating structure via group sparsity brings further gains.\n4. The authors conduct extensive analyses, which pre-empted many of my questions, such as the variation across different sparsity masks, sparsity patterns across different tasks, etc. Overall, the analyses shed additional light on the characteristics and preferences of different tasks in transfer learning.\n\nCons:\n1. The approach is potentially more complicated than the baseline, so it is important that the authors open-source their code.\n2. The diff vector is distributed over the entire set of parameters of the model rather than focused in a few layers. This makes it potentially harder to combine the diff vectors from different tasks as can be done with adapters (see e.g. https://arxiv.org/abs/2005.00247) and to compose multiple diff vectors. \n\nQuestions:\n1. Does a visualization of diff vectors of different tasks (such as using t-SNE) reveal any interesting patterns?\n2. Are there any transfer settings that the addition of task-specific parameters can model but inserting layer-specific transformations via adapters cannot (or vice versa)? Adapters have been used to transfer across modalities such as languages (see e.g. Pfeiffer et al. (2020), https://arxiv.org/abs/2005.00052) and I am wondering whether the same would be possible by adding task-specific parameters.\n3. How long does your approach take to converge in comparison to the baselines? How much longer do you need to fine-tune with non-zero masks for magnitude pruning for sparsity control? What is the performance benefit of this further fine-tuning?", "rating": "8: Top 50% of accepted papers, clear accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603654321674}], "openreview_url": "https://openreview.net/forum?id=E4PK0rg2eP", "arxiv_id": "2012.07463", "paper_pdf": "papers/E4PK0rg2eP.pdf", "paper_pdf_sha256": "a3966e58b21caeca135ff1ac346dd6e1587ca0c648c5fb61992cc59cfecb155d", "paper_pdf_bytes": 424130, "paper_pdf_source": "openreview", "code_url": "https://github.com/dguo98/DiffPruning", "code_repository": "dguo98/DiffPruning", "code_commit": "085fa7d9e2bcbcd19f705a08b1c0c6c74c54bada", "code_archive": "repos/E4PK0rg2eP.zip", "code_archive_sha256": "f11d18a2726c04ebd656b7d948cd42ba78b6ac11635e7c6383e40ac4dbf32af5", "code_archive_bytes": 6843736, "code_file_count": 471, "code_extensions": {".py": 452, ".sh": 10, ".ipynb": 7, ".js": 2}, "github_disk_usage_kb": 3373, "github_languages": {"Python": 6386215, "Jupyter Notebook": 596080, "Shell": 9196, "CSS": 5477, "Dockerfile": 3967, "Makefile": 1937}, "github_archived": false, "github_pushed_at": "2021-02-03T08:34:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/parameter-efficient-transfer-learning-with-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZPZ4eCQU9k", "year": 2025, "status": "rejected", "title": "xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories", "authors": ["Maurice Kraus", "Felix Divo", "Devendra Singh Dhami", "Kristian Kersting"], "authorids": ["~Maurice_Kraus1", "~Felix_Divo1", "~Devendra_Singh_Dhami1", "~Kristian_Kersting1"], "authors_source": "OpenReview API", "abstract": "Time series data is prevalent across numerous fields, necessitating the development of robust and accurate forecasting models. Capturing patterns both within and between temporal and multivariate components is crucial for reliable predictions.\nWe introduce xLSTM-Mixer, a model designed to effectively integrate temporal sequences, joint time-variate information, and multiple perspectives for robust forecasting. Our approach begins with a linear forecast shared across variates, which is then refined by xLSTM blocks. They serve as key elements for modeling the complex dynamics of challenging time series data. xLSTM-Mixer ultimately\nreconciles two distinct views to produce the final forecast. Our extensive evaluations demonstrate its superior long-term forecasting performance compared to recent state-of-the-art methods. A thorough model analysis provides further insights into its key components and confirms its robustness and effectiveness. This work contributes to the resurgence of recurrent models in time series forecasting.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "WzFz6b0k2G", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10503/Reviewer_xNaw"], "rating": 6, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "The paper introduces xLSTM-Mixer, a new model for forecasting complex time series data with multiple variables. It combines the xLSTM architecture with multi-view processing and variate mixing. The model works in three stages: first, it creates an initial forecast, then it uses xLSTM blocks to combine time and variable patterns, and finally, it merges two different forecast views for results in most cases beating baseline models. xLSTM-Mixer addresses challenges associated with Transformer models, such as high computational costs and data requirements for long sequences.", "review_text": "The paper introduces xLSTM-Mixer, a new model for forecasting complex time series data with multiple variables. It combines the xLSTM architecture with multi-view processing and variate mixing. The model works in three stages: first, it creates an initial forecast, then it uses xLSTM blocks to combine time and variable patterns, and finally, it merges two different forecast views for results in most cases beating baseline models. xLSTM-Mixer addresses challenges associated with Transformer models, such as high computational costs and data requirements for long sequences.", "strengths": "**Originality**\nxLSTM-Mixer combines temporal sequences, joint time-variate information, and multi-view perspectives to handle multivariate time series complexities, capturing temporal and cross-variate patterns for accurate predictions, it builds on prior work, but combines it in an intricate way. The work enhances the xLSTM architecture, however limiting itself to the sLSTM.\n\n**Quality**\nThe work is written with enough clarity, related work is addressed and the paper is fairly easy to understand. The graphical presentation of the model allows for easy following.\n\n**Significance**\nAuthors do not claim state-of-the-art performance, however, they do beat the current baselines to which they compare themselves. It is the second work relating to the use of xLSTM to time series, and authors raise concerns about the first one, those concerns must be backed in the future.", "weaknesses": "**Weaknesses**\n\n1. The authors conduct experiments on only a subset of datasets, specifically focusing on a subset of the Informer long-horizon forecasting datasets and omitting the Exchange dataset, which would add considerable value if included. The reviewers would appreciate a deeper analysis of why the XLSTM-Mixer performs better on some datasets and worse on others, particularly through an exploration of cross-correlation between variables and model performance.\n\n2. The ablation study, conducted on only two datasets, lacks breadth. Expanding this experiment across additional datasets would significantly strengthen the findings.\n\n3. The experiments do not address a critical aspect: the model's sensitivity to the order of variates, as it is not permutation invariant in this regard.", "questions": "1. The term \"Mix View\" in Table 3, line 332, is unclear and challenging to interpret. Could you clarify if this refers to the performance of channel mixing?\n\n2. The dataset selection could be expanded to include the Exchange from the  (Informer long-horizon forecasting datasets) datasets and more extensive ablation studies across additional datasets to provide a thorough evaluation of the xLSTM-Mixer’s performance, including an exploration of cross-correlation between variables and model effectiveness on different datasets or synthetic ones.\n\n3. Consider incorporating synthetic data to facilitate verification of learnable concepts and addressing the model’s sensitivity to the order of variates, as it lacks permutation invariance.\n\n4. I would highly encourage you to back the claims of challenges in the reproducibility of prior works and why trend-seasonality decomposition is redundant.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces xLSTM-Mixer, a new model for forecasting complex time series data with multiple variables. It combines the xLSTM architecture with multi-view processing and variate mixing. The model works in three stages: first, it creates an initial forecast, then it uses xLSTM blocks to combine time and variable patterns, and finally, it merges two different forecast views for results in most cases beating baseline models. xLSTM-Mixer addresses challenges associated with Transformer models, such as high computational costs and data requirements for long sequences.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "**Originality**\nxLSTM-Mixer combines temporal sequences, joint time-variate information, and multi-view perspectives to handle multivariate time series complexities, capturing temporal and cross-variate patterns for accurate predictions, it builds on prior work, but combines it in an intricate way. The work enhances the xLSTM architecture, however limiting itself to the sLSTM.\n\n**Quality**\nThe work is written with enough clarity, related work is addressed and the paper is fairly easy to understand. The graphical presentation of the model allows for easy following.\n\n**Significance**\nAuthors do not claim state-of-the-art performance, however, they do beat the current baselines to which they compare themselves. It is the second work relating to the use of xLSTM to time series, and authors raise concerns about the first one, those concerns must be backed in the future.", "weaknesses": "**Weaknesses**\n\n1. The authors conduct experiments on only a subset of datasets, specifically focusing on a subset of the Informer long-horizon forecasting datasets and omitting the Exchange dataset, which would add considerable value if included. The reviewers would appreciate a deeper analysis of why the XLSTM-Mixer performs better on some datasets and worse on others, particularly through an exploration of cross-correlation between variables and model performance.\n\n2. The ablation study, conducted on only two datasets, lacks breadth. Expanding this experiment across additional datasets would significantly strengthen the findings.\n\n3. The experiments do not address a critical aspect: the model's sensitivity to the order of variates, as it is not permutation invariant in this regard.", "questions": "1. The term \"Mix View\" in Table 3, line 332, is unclear and challenging to interpret. Could you clarify if this refers to the performance of channel mixing?\n\n2. The dataset selection could be expanded to include the Exchange from the  (Informer long-horizon forecasting datasets) datasets and more extensive ablation studies across additional datasets to provide a thorough evaluation of the xLSTM-Mixer’s performance, including an exploration of cross-correlation between variables and model effectiveness on different datasets or synthetic ones.\n\n3. Consider incorporating synthetic data to facilitate verification of learnable concepts and addressing the model’s sensitivity to the order of variates, as it lacks permutation invariance.\n\n4. I would highly encourage you to back the claims of challenges in the reproducibility of prior works and why trend-seasonality decomposition is redundant.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1731103301032}, {"id": "Q1Ryx8rvFQ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10503/Reviewer_GoVE"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper introduces a new time series prediction model called xLSTM-Mixer, which aims to effectively capture and integrate complex patterns in time series data by combining linear prediction with xLSTM blocks. The model first uses a shared linear model to make initial predictions for each variable, then fine-tunes and mixes these predictions through a series of sLSTM modules, and finally generates the final prediction result by integrating two different perspectives' predictions. Experimental results show that xLSTM-Mixer has significant advantages in long-term time series prediction tasks, achieving better predictive accuracy than current state-of-the-art methods on multiple benchmark datasets. In addition, the article conducts detailed analysis of the model, including A/B testing, sensitivity analysis of hyperparameters, and research on initial embedding vectors, further validating the effectiveness and robustness of the model. This work not only improves the level of deep learning-based time series prediction technology but also provides new directions for future related research.", "review_text": "This paper introduces a new time series prediction model called xLSTM-Mixer, which aims to effectively capture and integrate complex patterns in time series data by combining linear prediction with xLSTM blocks. The model first uses a shared linear model to make initial predictions for each variable, then fine-tunes and mixes these predictions through a series of sLSTM modules, and finally generates the final prediction result by integrating two different perspectives' predictions. Experimental results show that xLSTM-Mixer has significant advantages in long-term time series prediction tasks, achieving better predictive accuracy than current state-of-the-art methods on multiple benchmark datasets. In addition, the article conducts detailed analysis of the model, including A/B testing, sensitivity analysis of hyperparameters, and research on initial embedding vectors, further validating the effectiveness and robustness of the model. This work not only improves the level of deep learning-based time series prediction technology but also provides new directions for future related research.", "strengths": "- Superior long-term forecasting performance: xLSTM-Mixer outperforms recent state-of-the-art methods in terms of predictive accuracy on various benchmark datasets.\n\n- Robustness and effectiveness: Detailed model analysis, including A/B testing, sensitivity analysis of hyperparameters, and research on initial embedding vectors, further validates the effectiveness and robustness of the model.\n\n- Contribution to the resurgence of recurrent models in time series forecasting: By combining linear prediction with xLSTM blocks and integrating time, variate, and multi-view mixing, xLSTM-Mixer enhances the learning of necessary features and offers a viable alternative to current sequence models, contributing to the resurgence of recurrent models in time series forecasting.", "weaknesses": "- The contribution regarding advancement of  xLSTM-Mixer over the base xLSTM model is not clearly defined. The novelty and significant contributions need further clarification to distinguish xLSTM-Mixer as more than a trivial modification.\n\n- The paper asserts that longer lookback lengths improve forecasting but Figure 5 shows increased MSE at a lookback length of 1024 compared to 768. This requires further exploration as it conflicts with the claims.\n\n- The manuscript dismisses trend-seasonality decomposition without empirical support, which is a significant weakness given the technique's established value in forecasting.\n\n- The paper omits a comparison of the xLSTM-Mixer's computational complexity and cost against other methods, a critical evaluation factor for practical model deployment.", "questions": "- The manuscript fails to convincingly argue the substantial novelty of xLSTM-Mixer compared to prior xLSTM models. Authors should clarify that xLSTM-Mixer is not merely a combination of xLSTM and Multi-View Mixing, as the current description lacks the depth to establish its significance.\n\n- While the manuscript acknowledges xLSTM-Mixer's performance with varying lookback windows (Figure 5), it lacks a thorough experimental setup to explore different input/output ratios. A comprehensive analysis should include varied input lengths relative to output horizons to assess model robustness. The absence of this setup limits understanding of performance in diverse forecasting conditions. Additional experiments with varying input/output ratios, as suggested by Table 6, are recommended.\n\n- The paper does not explain the increase in MSE at a lookback length of 1024, leaving readers unclear about the model's behavior. To strengthen the manuscript, the authors should provide a plausible explanation or conduct additional experiments to verify performance consistency across different lookback lengths.\n\n- A significant issue is the claim that a simple NLinear normalization scheme is sufficient without rigorous testing or theoretical backing. The decision to forgo trend-seasonality decomposition and rely solely on NLinear normalization lacks the necessary validation, raising questions about the model's robustness. More experimental corroboration against established methods is needed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new time series prediction model called xLSTM-Mixer, which aims to effectively capture and integrate complex patterns in time series data by combining linear prediction with xLSTM blocks. The model first uses a shared linear model to make initial predictions for each variable, then fine-tunes and mixes these predictions through a series of sLSTM modules, and finally generates the final prediction result by integrating two different perspectives' predictions. Experimental results show that xLSTM-Mixer has significant advantages in long-term time series prediction tasks, achieving better predictive accuracy than current state-of-the-art methods on multiple benchmark datasets. In addition, the article conducts detailed analysis of the model, including A/B testing, sensitivity analysis of hyperparameters, and research on initial embedding vectors, further validating the effectiveness and robustness of the model. This work not only improves the level of deep learning-based time series prediction technology but also provides new directions for future related research.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "- Superior long-term forecasting performance: xLSTM-Mixer outperforms recent state-of-the-art methods in terms of predictive accuracy on various benchmark datasets.\n\n- Robustness and effectiveness: Detailed model analysis, including A/B testing, sensitivity analysis of hyperparameters, and research on initial embedding vectors, further validates the effectiveness and robustness of the model.\n\n- Contribution to the resurgence of recurrent models in time series forecasting: By combining linear prediction with xLSTM blocks and integrating time, variate, and multi-view mixing, xLSTM-Mixer enhances the learning of necessary features and offers a viable alternative to current sequence models, contributing to the resurgence of recurrent models in time series forecasting.", "weaknesses": "- The contribution regarding advancement of  xLSTM-Mixer over the base xLSTM model is not clearly defined. The novelty and significant contributions need further clarification to distinguish xLSTM-Mixer as more than a trivial modification.\n\n- The paper asserts that longer lookback lengths improve forecasting but Figure 5 shows increased MSE at a lookback length of 1024 compared to 768. This requires further exploration as it conflicts with the claims.\n\n- The manuscript dismisses trend-seasonality decomposition without empirical support, which is a significant weakness given the technique's established value in forecasting.\n\n- The paper omits a comparison of the xLSTM-Mixer's computational complexity and cost against other methods, a critical evaluation factor for practical model deployment.", "questions": "- The manuscript fails to convincingly argue the substantial novelty of xLSTM-Mixer compared to prior xLSTM models. Authors should clarify that xLSTM-Mixer is not merely a combination of xLSTM and Multi-View Mixing, as the current description lacks the depth to establish its significance.\n\n- While the manuscript acknowledges xLSTM-Mixer's performance with varying lookback windows (Figure 5), it lacks a thorough experimental setup to explore different input/output ratios. A comprehensive analysis should include varied input lengths relative to output horizons to assess model robustness. The absence of this setup limits understanding of performance in diverse forecasting conditions. Additional experiments with varying input/output ratios, as suggested by Table 6, are recommended.\n\n- The paper does not explain the increase in MSE at a lookback length of 1024, leaving readers unclear about the model's behavior. To strengthen the manuscript, the authors should provide a plausible explanation or conduct additional experiments to verify performance consistency across different lookback lengths.\n\n- A significant issue is the claim that a simple NLinear normalization scheme is sufficient without rigorous testing or theoretical backing. The decision to forgo trend-seasonality decomposition and rely solely on NLinear normalization lacks the necessary validation, raising questions about the model's robustness. More experimental corroboration against established methods is needed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730520528169}, {"id": "2Yh93Gwvi3", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10503/Reviewer_9duf"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper presents a novel design - xLSTM-Mixer which is based on LSTM and a set of mixing machanism for multivariate time series forecasting. The model builds on the idea of recurrent structures, integrating three mixing strategy: an initial linear forecast with time mixing, joint mixing with sLSTM blocks, and a final view mixing stage. Evaluations show that xLSTM-Mixer achieves state-of-the-art forecasting performance across multiple datasets which is a key contribution to the ongoing development of recurrent models.", "review_text": "The paper presents a novel design - xLSTM-Mixer which is based on LSTM and a set of mixing machanism for multivariate time series forecasting. The model builds on the idea of recurrent structures, integrating three mixing strategy: an initial linear forecast with time mixing, joint mixing with sLSTM blocks, and a final view mixing stage. Evaluations show that xLSTM-Mixer achieves state-of-the-art forecasting performance across multiple datasets which is a key contribution to the ongoing development of recurrent models.", "strengths": "**Innovation in Model Architecture**: The integration of time and variate mixing with xLSTM blocks is an innovative approach. This design tackles some limitations of conventional recurrent and transformer-based models in capturing complex temporal dependencies, which is especially challenging in multivariate time series data. \n\n**Empirical Performance**: The experiments show that xLSTM-Mixer outperforms a variety of established methods (e.g., Time-Mixer, iTransformer, MICN) in terms of mean squared error (MSE) and mean absolute error (MAE) across multiple datasets, together with comprehensive ablation studies on initializations, size of hidden dimension and length of lookback window shows the model’s robustness and generalizability.", "weaknesses": "**Model structure**: \nLSTM-Mixer appears to be an incremental improvement over xLSTM, and the mixing machisim does not seems to be very siginificant regards to the Ablation Study on Table 3. An experiment with alternative recurrent structures may help here. Instead of relying solely on the sLSTM block with the joint mixing machinism, it would be usefule to assess the performance of other recurrent models, such as GRU and Mamba.\n\n**Computational Efficiency**: \nThe complexity of xLSTM-Mixer’s mixing machanism and recurrent structures would be computationally intensive. Further discussion on the computational costs relative to attentions and SSMs would be beneficial. \n\n**Evaluation on Short-Term Forecasting**:\nWhile the model demonstrates strong performance in long-term forecasting, the short-term forecasting evaluation is inadequate. Adding a more robust short-term evaluation could highlight the model’s flexibility across different forecasting horizons.\n\n**Evaluation on Long-Term Forecasting**:\nAlthough most studies follow the 96, 192, 336, 720 prediction length on benchmarks, regards to this work highlighted long-term forecasting as a distinctive feature, a prediction length of 720 is insufficient.", "questions": "**More discussion on the use of Multivariate information**:\n- Firstly, the primary innovation claimed by the authors is the use of multivariate time series forecasting. However, simply explaining how the new method works and benchmarking with other multivariate models is insufficient. Since the impact of using multivariate information varies across different datasets and variables, and given that multivariate mixing is the core innovation, additional analysis on this aspect should be provided.\n- Secondly, whats are the authors' opinion on the heterogeneity of variates, and how important is this compare to the temporal dependency in terms of forecasting? \n- Thirdly, the authors mention that variate ordering may impact performance but leave this for future work. Given the importance of variate dependencies in time series, a preliminary exploration of this factor could provide additional insights into potential model improvements.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a novel design - xLSTM-Mixer which is based on LSTM and a set of mixing machanism for multivariate time series forecasting. The model builds on the idea of recurrent structures, integrating three mixing strategy: an initial linear forecast with time mixing, joint mixing with sLSTM blocks, and a final view mixing stage. Evaluations show that xLSTM-Mixer achieves state-of-the-art forecasting performance across multiple datasets which is a key contribution to the ongoing development of recurrent models.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "**Innovation in Model Architecture**: The integration of time and variate mixing with xLSTM blocks is an innovative approach. This design tackles some limitations of conventional recurrent and transformer-based models in capturing complex temporal dependencies, which is especially challenging in multivariate time series data. \n\n**Empirical Performance**: The experiments show that xLSTM-Mixer outperforms a variety of established methods (e.g., Time-Mixer, iTransformer, MICN) in terms of mean squared error (MSE) and mean absolute error (MAE) across multiple datasets, together with comprehensive ablation studies on initializations, size of hidden dimension and length of lookback window shows the model’s robustness and generalizability.", "weaknesses": "**Model structure**: \nLSTM-Mixer appears to be an incremental improvement over xLSTM, and the mixing machisim does not seems to be very siginificant regards to the Ablation Study on Table 3. An experiment with alternative recurrent structures may help here. Instead of relying solely on the sLSTM block with the joint mixing machinism, it would be usefule to assess the performance of other recurrent models, such as GRU and Mamba.\n\n**Computational Efficiency**: \nThe complexity of xLSTM-Mixer’s mixing machanism and recurrent structures would be computationally intensive. Further discussion on the computational costs relative to attentions and SSMs would be beneficial. \n\n**Evaluation on Short-Term Forecasting**:\nWhile the model demonstrates strong performance in long-term forecasting, the short-term forecasting evaluation is inadequate. Adding a more robust short-term evaluation could highlight the model’s flexibility across different forecasting horizons.\n\n**Evaluation on Long-Term Forecasting**:\nAlthough most studies follow the 96, 192, 336, 720 prediction length on benchmarks, regards to this work highlighted long-term forecasting as a distinctive feature, a prediction length of 720 is insufficient.", "questions": "**More discussion on the use of Multivariate information**:\n- Firstly, the primary innovation claimed by the authors is the use of multivariate time series forecasting. However, simply explaining how the new method works and benchmarking with other multivariate models is insufficient. Since the impact of using multivariate information varies across different datasets and variables, and given that multivariate mixing is the core innovation, additional analysis on this aspect should be provided.\n- Secondly, whats are the authors' opinion on the heterogeneity of variates, and how important is this compare to the temporal dependency in terms of forecasting? \n- Thirdly, the authors mention that variate ordering may impact performance but leave this for future work. Given the importance of variate dependencies in time series, a preliminary exploration of this factor could provide additional insights into potential model improvements.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730440902468}, {"id": "J5yEngnFY0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10503/Reviewer_FwwC"], "rating": 8, "soundness": 4, "presentation": 3, "contribution": 4, "confidence": 4, "summary": "The paper introduce  a novel time series forecasting model with time and variate mixing in the context of recurrent models.", "review_text": "The paper introduce  a novel time series forecasting model with time and variate mixing in the context of recurrent models.", "strengths": "- Figure 1 provides a clear overview of the method.\n- Several commonly used benchmark datasets are included.\n- The proposed xLSTM-Mixer model consistently achieves state-of-the-art performance across a broad range of benchmarks.\n- Informative forecast plots are included.\n- A comprehensive set of baseline models from different architecture categories including, Transformer-based, CNN-based, MLP-based, and Recurrent architectures.\n- The ablation study on architecture choices is highly insightful and supports the combined aspects of the proposed model.\n- Open-source code is available with reproducible experiments", "weaknesses": "1. Random seeds were included as a hyperparameter, but it would be beneficial to train the model over several trials to compute the average and standard deviation of the results. This approach would provide a more robust measure of performance consistency across runs.\n2. The approach does not reference similar architectures, such as iTransformer, which utilizes the attention and feed-forward network on inverted dimensions. In iTransformer, individual time points are embedded as variate tokens, leveraging the attention mechanism to capture multivariate correlations. This bears a resemblance to the proposed method, which jointly combines time and variate information to capture complex patterns in the data by employing transposed input dimensions to invert the series. Mentioning this may help emphasize the translatable similarities across architecture backbones.\n3. The motivation for the reverse-up projected embedding is not explicitly explained. Including an ablation study of the model with and without this embedding would be useful to help justify this architectural choice.\n4. Some models cited in prior work are missing as baselines, including TFT, N-HiTS, and N-BEATS.", "questions": "Suggestions:\n- Adding dimensions to Fig. 1 would help clarify input shapes after higher-dimensional projections and transpositions.\n- Lines 41-42: Regarding Transformer architectures, the authors state, “For instance, they typically require large datasets to train successfully, restricting their use to only a subset of applications.” However, I do not believe this claim is fully supported by the literature. For example, common benchmarks for long-horizon forecasting include datasets like ILI, Traffic, and Weather. The ILI dataset, with only 966 timestamps, could be considered relatively small, yet PatchTST outperforms other Transformer and linear models on this dataset [1]. Similarly, for ILI, Autoformer outperforms other models, including recurrent architectures like LSTMs [2].\n- It would help motivate the proposed approach to illustrate the trade-offs in computational complexity between the proposed model, Transformer-based models, and MLP-based models for both training and inference.\n- While model hyperparameters for the proposed model are included in the appendix, it would be helpful to provide hyperparameters and the number of training epochs for the baseline models as well.\n- Additional benchmark datasets could include ILI and Exchange datasets which are used in prior work [2].\n\nReferences:\n1. “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers”\n2. “Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting”", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduce  a novel time series forecasting model with time and variate mixing in the context of recurrent models.", "soundness": 4, "presentation": 3, "contribution": 4, "strengths": "- Figure 1 provides a clear overview of the method.\n- Several commonly used benchmark datasets are included.\n- The proposed xLSTM-Mixer model consistently achieves state-of-the-art performance across a broad range of benchmarks.\n- Informative forecast plots are included.\n- A comprehensive set of baseline models from different architecture categories including, Transformer-based, CNN-based, MLP-based, and Recurrent architectures.\n- The ablation study on architecture choices is highly insightful and supports the combined aspects of the proposed model.\n- Open-source code is available with reproducible experiments", "weaknesses": "1. Random seeds were included as a hyperparameter, but it would be beneficial to train the model over several trials to compute the average and standard deviation of the results. This approach would provide a more robust measure of performance consistency across runs.\n2. The approach does not reference similar architectures, such as iTransformer, which utilizes the attention and feed-forward network on inverted dimensions. In iTransformer, individual time points are embedded as variate tokens, leveraging the attention mechanism to capture multivariate correlations. This bears a resemblance to the proposed method, which jointly combines time and variate information to capture complex patterns in the data by employing transposed input dimensions to invert the series. Mentioning this may help emphasize the translatable similarities across architecture backbones.\n3. The motivation for the reverse-up projected embedding is not explicitly explained. Including an ablation study of the model with and without this embedding would be useful to help justify this architectural choice.\n4. Some models cited in prior work are missing as baselines, including TFT, N-HiTS, and N-BEATS.", "questions": "Suggestions:\n- Adding dimensions to Fig. 1 would help clarify input shapes after higher-dimensional projections and transpositions.\n- Lines 41-42: Regarding Transformer architectures, the authors state, “For instance, they typically require large datasets to train successfully, restricting their use to only a subset of applications.” However, I do not believe this claim is fully supported by the literature. For example, common benchmarks for long-horizon forecasting include datasets like ILI, Traffic, and Weather. The ILI dataset, with only 966 timestamps, could be considered relatively small, yet PatchTST outperforms other Transformer and linear models on this dataset [1]. Similarly, for ILI, Autoformer outperforms other models, including recurrent architectures like LSTMs [2].\n- It would help motivate the proposed approach to illustrate the trade-offs in computational complexity between the proposed model, Transformer-based models, and MLP-based models for both training and inference.\n- While model hyperparameters for the proposed model are included in the appendix, it would be helpful to provide hyperparameters and the number of training epochs for the baseline models as well.\n- Additional benchmark datasets could include ILI and Exchange datasets which are used in prior work [2].\n\nReferences:\n1. “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers”\n2. “Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting”", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730264058813}, {"id": "iMof6kUc6h", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission10503/Reviewer_cZwn"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "The work proposed xLSTM-Mixer for multivariate time series forecasting that combines linear forecasters (NLinear) with efficient recurrent xLSTM blocks. By adapting xLSTM modules for the temporal and variate dimensions, the model refined the predictions of NLinear with the utilization of time-variate information. The model also employs techniques like reversible instance normalization and soft prompt tokens to further improve the performance. Empirical experiments demonstrate that xLSTM-Mixer can outperform several baseline methods in most cases in long-term forecasting.", "review_text": "The work proposed xLSTM-Mixer for multivariate time series forecasting that combines linear forecasters (NLinear) with efficient recurrent xLSTM blocks. By adapting xLSTM modules for the temporal and variate dimensions, the model refined the predictions of NLinear with the utilization of time-variate information. The model also employs techniques like reversible instance normalization and soft prompt tokens to further improve the performance. Empirical experiments demonstrate that xLSTM-Mixer can outperform several baseline methods in most cases in long-term forecasting.", "strengths": "1. The motivations are well-summarized and the paper's writing is easy to follow.\n2. The author provides the code to ensure the reproducibility of the paper.", "weaknesses": "1. The main contribution of this paper is to refine the predictions obtained from NLinear by introducing xLSTM modules, which may lack novelty and motivation in technology. On the one hand, please clarify more about the advantages of xLSTM modules that have been refined over (linear forecaster + renorm). For example, to support the mentioned advantages of modeling variate relationships, you can provide a comparative analysis of xLSTM against MLP mixers and efficient attention modules in the Mixer architecture. Also, extending ablation studies (Table 3) to more datasets and different variate-mixing modules would provide a more comprehensive evaluation of the model's effectiveness.\n2. It is uncertain whether the authors use the unidirectional xLSTM, which is generally adopted for sequential modeling, to model the correlations of variates. If it is unidirectional, please explain how it addresses the non-sequential nature of the variate correlation, and discuss the potential benefits or drawbacks of using a bidirectional approach in this context.\n3. This paper mainly provides the results of long-term forecasting. The lookback seems to be fixed at 96. Considering the extensive overfitting of current deep models on these datasets, it is recommended that the author provide more comparisons with varying lookback lengths and on additional datasets that are less prone to overfitting. This would provide a more robust evaluation of the model's performance.\n4. Compared with related recurrent models (such as xLSTMTime), the performance improvement of the proposed model is not significant shown in Table 2. The author should add more recent recurrent models (e.g., SutraNets and P-sLSTM) and more discussions with related work.\n5. The showcases in Figure 2 do not reflect the advantages of xLSTM-Mixer in multivariate modeling. To make it more specific, please include visualizations that explicitly show how xLSTM-Mixer captures relationships between multiple variates in its predictions.\n6. The paper concludes by highlighting the potential of xLSTM-Mixer for future research in areas such as short-term forecasting, time series classification, and imputation. I wonder if there are some results on these analysis tasks using xLSTMs.", "questions": "1. Has the author tried to train on time series with randomly shuffled temporal orders, and what will happen to the results?\n2. About the experiments of enlarging lookback (Figure 3): In the DLinear paper, the performance can also be improved by increasing the lookback using a single linear layer. It is suggested that the author introduce such linear models into the baseline. Otherwise, the advantage can only be due to NLinear. It is also encouraged to provide the model performance with a longer lookback length to verify whether the proposed can continuously improve the performance (or keep it stable in a very long lookback length).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The work proposed xLSTM-Mixer for multivariate time series forecasting that combines linear forecasters (NLinear) with efficient recurrent xLSTM blocks. By adapting xLSTM modules for the temporal and variate dimensions, the model refined the predictions of NLinear with the utilization of time-variate information. The model also employs techniques like reversible instance normalization and soft prompt tokens to further improve the performance. Empirical experiments demonstrate that xLSTM-Mixer can outperform several baseline methods in most cases in long-term forecasting.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "1. The motivations are well-summarized and the paper's writing is easy to follow.\n2. The author provides the code to ensure the reproducibility of the paper.", "weaknesses": "1. The main contribution of this paper is to refine the predictions obtained from NLinear by introducing xLSTM modules, which may lack novelty and motivation in technology. On the one hand, please clarify more about the advantages of xLSTM modules that have been refined over (linear forecaster + renorm). For example, to support the mentioned advantages of modeling variate relationships, you can provide a comparative analysis of xLSTM against MLP mixers and efficient attention modules in the Mixer architecture. Also, extending ablation studies (Table 3) to more datasets and different variate-mixing modules would provide a more comprehensive evaluation of the model's effectiveness.\n2. It is uncertain whether the authors use the unidirectional xLSTM, which is generally adopted for sequential modeling, to model the correlations of variates. If it is unidirectional, please explain how it addresses the non-sequential nature of the variate correlation, and discuss the potential benefits or drawbacks of using a bidirectional approach in this context.\n3. This paper mainly provides the results of long-term forecasting. The lookback seems to be fixed at 96. Considering the extensive overfitting of current deep models on these datasets, it is recommended that the author provide more comparisons with varying lookback lengths and on additional datasets that are less prone to overfitting. This would provide a more robust evaluation of the model's performance.\n4. Compared with related recurrent models (such as xLSTMTime), the performance improvement of the proposed model is not significant shown in Table 2. The author should add more recent recurrent models (e.g., SutraNets and P-sLSTM) and more discussions with related work.\n5. The showcases in Figure 2 do not reflect the advantages of xLSTM-Mixer in multivariate modeling. To make it more specific, please include visualizations that explicitly show how xLSTM-Mixer captures relationships between multiple variates in its predictions.\n6. The paper concludes by highlighting the potential of xLSTM-Mixer for future research in areas such as short-term forecasting, time series classification, and imputation. I wonder if there are some results on these analysis tasks using xLSTMs.", "questions": "1. Has the author tried to train on time series with randomly shuffled temporal orders, and what will happen to the results?\n2. About the experiments of enlarging lookback (Figure 3): In the DLinear paper, the performance can also be improved by increasing the lookback using a single linear layer. It is suggested that the author introduce such linear models into the baseline. Otherwise, the advantage can only be due to NLinear. It is also encouraged to provide the model performance with a longer lookback length to verify whether the proposed can continuously improve the performance (or keep it stable in a very long lookback length).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730043755117}], "openreview_url": "https://openreview.net/forum?id=ZPZ4eCQU9k", "arxiv_id": "2410.16928", "paper_pdf": "papers/ZPZ4eCQU9k.pdf", "paper_pdf_sha256": "5f4b67ff8b6aec7b74f60e40a0cef8756d22cf79cfee7e2a3d46ef2ef045df1c", "paper_pdf_bytes": 587538, "paper_pdf_source": "openreview", "code_url": "https://github.com/mauricekraus/xLSTM-Mixer", "code_repository": "mauricekraus/xLSTM-Mixer", "code_commit": "730b0531aa9456e498765028f3c22ca3677de42e", "code_archive": "repos/ZPZ4eCQU9k.zip", "code_archive_sha256": "02df029162721aa1433c38ddee9fed02f661ca724ef35e42e00623115a93abf8", "code_archive_bytes": 434103, "code_file_count": 88, "code_extensions": {".py": 80, ".sh": 7, ".ipynb": 1}, "github_disk_usage_kb": 407, "github_languages": {"Python": 582249, "Jupyter Notebook": 53199, "Shell": 47738, "Dockerfile": 1137}, "github_archived": false, "github_pushed_at": "2025-06-12T15:21:40Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/xlstm-mixer-multivariate-time-series"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PoBB8n52oi", "year": 2024, "status": "rejected", "title": "SummaryMixing: A Linear-Complexity Alternative to Self-Attention for Speech Recognition and Understanding", "authors": ["Titouan Parcollet", "Rogier van Dalen", "Shucong Zhang", "Sourav Bhattacharya"], "authorids": ["~Titouan_Parcollet1", "~Rogier_van_Dalen2", "~Shucong_Zhang2", "~Sourav_Bhattacharya1"], "authors_source": "OpenReview API", "abstract": "Modern speech processing systems rely on self-attention. Unfortunately, token\nmixing with self-attention takes quadratic time in the length of the speech utterance,\nslowing down inference as well as training and increasing memory consumption.\nCheaper alternatives to self-attention for ASR have been developed, but they fail to\nconsistently reach the same level of accuracy. However, attention layers in trained\nspeech recognizers tend to not capture fine-grained pair-wise information. This\npaper, therefore, proposes a novel linear-time alternative to self-attention. It sum-\nmarises a whole utterance with the mean over vectors for all time steps. This single\nsummary is then combined with time-specific information. We call this method\n“SummaryMixing”. Introducing SummaryMixing in state-of-the-art ASR models\nmakes it feasible to preserve or exceed previous speech recognition performance\nwhile lowering the training and inference times by up to 28% and reducing the\nmemory budget by a factor of two. The benefits of SummaryMixing can also be\ngeneralized to other speech-processing tasks, such as speech understanding.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "bAK54Iydlj", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5692/Reviewer_wGqz"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes a novel neural network architecture called SummaryMixing for speech recognition. The motivation of this work is to replace the heavy computational cost yielded by transformer self-attention blocks with a SummaryMixing block on top of the branchformer architecture. SummaryMixing shows effectiveness by offering a competitive performance from the original branchformer architecture, especially with the low computational cost, mainly within the CTC framework.", "review_text": "This paper proposes a novel neural network architecture called SummaryMixing for speech recognition. The motivation of this work is to replace the heavy computational cost yielded by transformer self-attention blocks with a SummaryMixing block on top of the branchformer architecture. SummaryMixing shows effectiveness by offering a competitive performance from the original branchformer architecture, especially with the low computational cost, mainly within the CTC framework.", "strengths": "- Simple and yet another neural network architecture by extending the branchformer architecture to capture global characteristics via SummaryMixing and local characteristics via cgMLP.\n- Showing the low computational cost compared with the original branchformer\n- Good reproducibility (open source implementation and the use of public data).", "weaknesses": "- The survey of the efficient transformer (conformed) is not sufficient. There are a lot of efficient transformers in the NLP and ML field (e.g., https://arxiv.org/pdf/2009.06732.pdf). Even if we limit the discussions to speech applications, there are a lot of methods (e.g., squeezeformer, efficient conformer, emformer, etc.). \n- The performance improvement from the conventional method (e.g., fastformer) is marginal.\n- The method is on top of branchformer, which already uses Fastformer, and its technical novelty is weak.\n- Several descriptions (e.g., the surveys, distinction from HyperMixer/MLP Mixer, and detailed architectures) are unclear.\n  - Frankly, I could not understand the distinction of the SummaryMixing block from HyperMixer/MLP Mixer only from Section 2.1, mainly due to the lack of an explanation of what HyperMixer/MLP Mixer is.\n  - It is difficult to understand the detailed architectures only from this description: \"In particular, the transformation (f), summary (s), and combiner (c) functions are all implemented as a dense linear layer followed by a GeLU activation function.\" Similarly, how the model size of the proposed and other methods is adjusted was unclear.", "questions": "- Can you expand the discussion of how this novel architecture and findings would attract the general AI and ML researchers in ICLR? This paper specializes in speech recognition (and spoken language understanding, which is a very similar task to speech recognition), and it is a narrow scope for me.\n- Besides the above expansion, can you explain why you selected Fastformermer and ContextNet?\n- Can you apply this to RNN-T?\n\nOther minor comments\n- Abstract: ASR --> automatic speech recognition (ASR), speech understanding --> spoken language understanding\n- Page 3, last paragraph \"Hence, $X \\in \\mathbb{R}$ becomes $X \\in \\mathbb{R} ^{T \\times D/n}$\": Is $X \\in \\mathbb{R}$ scalar? I think you missed adding some domains.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel neural network architecture called SummaryMixing for speech recognition. The motivation of this work is to replace the heavy computational cost yielded by transformer self-attention blocks with a SummaryMixing block on top of the branchformer architecture. SummaryMixing shows effectiveness by offering a competitive performance from the original branchformer architecture, especially with the low computational cost, mainly within the CTC framework.", "soundness": "3 good", "presentation": "2 fair", "contribution": "3 good", "strengths": "- Simple and yet another neural network architecture by extending the branchformer architecture to capture global characteristics via SummaryMixing and local characteristics via cgMLP.\n- Showing the low computational cost compared with the original branchformer\n- Good reproducibility (open source implementation and the use of public data).", "weaknesses": "- The survey of the efficient transformer (conformed) is not sufficient. There are a lot of efficient transformers in the NLP and ML field (e.g., https://arxiv.org/pdf/2009.06732.pdf). Even if we limit the discussions to speech applications, there are a lot of methods (e.g., squeezeformer, efficient conformer, emformer, etc.). \n- The performance improvement from the conventional method (e.g., fastformer) is marginal.\n- The method is on top of branchformer, which already uses Fastformer, and its technical novelty is weak.\n- Several descriptions (e.g., the surveys, distinction from HyperMixer/MLP Mixer, and detailed architectures) are unclear.\n  - Frankly, I could not understand the distinction of the SummaryMixing block from HyperMixer/MLP Mixer only from Section 2.1, mainly due to the lack of an explanation of what HyperMixer/MLP Mixer is.\n  - It is difficult to understand the detailed architectures only from this description: \"In particular, the transformation (f), summary (s), and combiner (c) functions are all implemented as a dense linear layer followed by a GeLU activation function.\" Similarly, how the model size of the proposed and other methods is adjusted was unclear.", "questions": "- Can you expand the discussion of how this novel architecture and findings would attract the general AI and ML researchers in ICLR? This paper specializes in speech recognition (and spoken language understanding, which is a very similar task to speech recognition), and it is a narrow scope for me.\n- Besides the above expansion, can you explain why you selected Fastformermer and ContextNet?\n- Can you apply this to RNN-T?\n\nOther minor comments\n- Abstract: ASR --> automatic speech recognition (ASR), speech understanding --> spoken language understanding\n- Page 3, last paragraph \"Hence, $X \\in \\mathbb{R}$ becomes $X \\in \\mathbb{R} ^{T \\times D/n}$\": Is $X \\in \\mathbb{R}$ scalar? I think you missed adding some domains.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699212530067}, {"id": "QHRErrrolL", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5692/Reviewer_fkrd"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposed a global summarization layer as a replacement for MHSA in several ASR architectures such as branchformer and Conformer. This layer or module is introduced and compared with several parallel implementation in an open source project SpeechBrain. \nThe newly proposed component dubbed summaryMixing has a linear computation cost on the encoder ASR components.\nExpectation is that this will close the gap with MHSA but actually authors found this surpasses it in their tests.", "review_text": "The paper proposed a global summarization layer as a replacement for MHSA in several ASR architectures such as branchformer and Conformer. This layer or module is introduced and compared with several parallel implementation in an open source project SpeechBrain. \nThe newly proposed component dubbed summaryMixing has a linear computation cost on the encoder ASR components.\nExpectation is that this will close the gap with MHSA but actually authors found this surpasses it in their tests.", "strengths": "The paper proposes a generalization of hyperMixer and apply it to Speech tasks. \nThe generalization is simple but its effectiveness is surprising. \nThe proposed architecture works very well in comparison with many baselines and architectures (Conformer, Branchformer, ContextNet, Branchformer +fastattention, Branchformer w/o attention, ...)\nThe obtained and reported results are very competitive with various ASR benchmarks and setups.\nIt is worth noting the surprisingly good performance of one of the simplest baseline, the branchformer w/o attention, which authors acknowledge and highlight.\nThey consider encoder architecture only to show the unnecessary MHSA complexity for ASR tasks is not coming exclusively from the top decoder..", "weaknesses": "There is a couple of points where the newly proposed component experimentation could have been improved.\nThe authors should clarify the ASR architecture in the final version since the current description of the ASR architecture is vague and prone to confusions-- see  questions below --.  I think this is something that authors might be aware of but prefer refer readers to the SpeechBrain recipes and experimentation.  This might leave a shallow reader without immediate interest in reproducing or delving into the code and recipes with a confusing architecture.\n\nThe approach has some limitations, so it would be intriguing to examine the performance of the proposed encoder in long-context ASR. \nSince this is a leaning representation conference, some experiments on NLP tasks, as in the original HyperMixer paper, would have attracted more readers.\n\nGiven the nature of speed optimization and comparison with MHSA, it would be interesting for the readers to know the specifics of the MHSA implementation as there are many very efficient implementations available nowadays.\n\nFinally, it would have been nice to compare the proposed approach in both batch and online ASR.", "questions": "The following questions might need some attention into the manuscript:\n* which is the global architecture and loss ? It is clear that it is a encoder decoder based architecture with CTC loss, but this is an important detail mentioned in 1 line. How many encoder layers each arch has ? which are the details of the decoder ? \n* Could the MHSA improve as the context becomes larger ? \n* which is the MSHA implementation have you used ?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposed a global summarization layer as a replacement for MHSA in several ASR architectures such as branchformer and Conformer. This layer or module is introduced and compared with several parallel implementation in an open source project SpeechBrain. \nThe newly proposed component dubbed summaryMixing has a linear computation cost on the encoder ASR components.\nExpectation is that this will close the gap with MHSA but actually authors found this surpasses it in their tests.", "soundness": "4 excellent", "presentation": "3 good", "contribution": "3 good", "strengths": "The paper proposes a generalization of hyperMixer and apply it to Speech tasks. \nThe generalization is simple but its effectiveness is surprising. \nThe proposed architecture works very well in comparison with many baselines and architectures (Conformer, Branchformer, ContextNet, Branchformer +fastattention, Branchformer w/o attention, ...)\nThe obtained and reported results are very competitive with various ASR benchmarks and setups.\nIt is worth noting the surprisingly good performance of one of the simplest baseline, the branchformer w/o attention, which authors acknowledge and highlight.\nThey consider encoder architecture only to show the unnecessary MHSA complexity for ASR tasks is not coming exclusively from the top decoder..", "weaknesses": "There is a couple of points where the newly proposed component experimentation could have been improved.\nThe authors should clarify the ASR architecture in the final version since the current description of the ASR architecture is vague and prone to confusions-- see  questions below --.  I think this is something that authors might be aware of but prefer refer readers to the SpeechBrain recipes and experimentation.  This might leave a shallow reader without immediate interest in reproducing or delving into the code and recipes with a confusing architecture.\n\nThe approach has some limitations, so it would be intriguing to examine the performance of the proposed encoder in long-context ASR. \nSince this is a leaning representation conference, some experiments on NLP tasks, as in the original HyperMixer paper, would have attracted more readers.\n\nGiven the nature of speed optimization and comparison with MHSA, it would be interesting for the readers to know the specifics of the MHSA implementation as there are many very efficient implementations available nowadays.\n\nFinally, it would have been nice to compare the proposed approach in both batch and online ASR.", "questions": "The following questions might need some attention into the manuscript:\n* which is the global architecture and loss ? It is clear that it is a encoder decoder based architecture with CTC loss, but this is an important detail mentioned in 1 line. How many encoder layers each arch has ? which are the details of the decoder ? \n* Could the MHSA improve as the context becomes larger ? \n* which is the MSHA implementation have you used ?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698880675714}, {"id": "h3YnKoBBZt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5692/Reviewer_yNSy"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a novel method to replace self-attentions with a more computationally efficient summary mixing model for speech recognition and spoken language understanding. Rather than computing an NxN attention matrix, the authors propose to first compute the temporal mean of the sequence and combine that with local information extracted using a cgMLP branch. Experiments on multiple corpora show that the proposed approach can match the performance of standard self-attention and Branchformer.", "review_text": "This paper proposes a novel method to replace self-attentions with a more computationally efficient summary mixing model for speech recognition and spoken language understanding. Rather than computing an NxN attention matrix, the authors propose to first compute the temporal mean of the sequence and combine that with local information extracted using a cgMLP branch. Experiments on multiple corpora show that the proposed approach can match the performance of standard self-attention and Branchformer.", "strengths": "1. The paper uses a very simple approach to substitute self-attention with attention between a temporal summary and the original vector that appears to work well in practice.\n\n2. Experiments are performed on multiple corpora for ASR (CommonVoice-Dutch, Italian, French), Librispeech, AISHELL, TED-LIUM 2, GSC and SLURP for SLU.", "weaknesses": "1. The paper makes the assumption that MHSA in the encoder is not necessary based on Zhang et. al. and Peng. et. al. However, I believe the assumption may not be well supported because (a) Zhang et al point out that you don't need self-attentions in higher layers, not that you don't need self-attentions at all, and (b) Peng. et al which uses self-attentions to model global context while using cgMLP to model local context. Therefore, the rationale behind this fundamental assumption made in the paper seems unclear. \n\n2. SLURP uses read speech - the impact of temporal averaging here may be minimal, but may be very harmful in spontaneous conversational speech. Such explorations in my view are important in this work to claim that SummaryMixing can reach top-of-the-line in speech understanding. \n\n3. Writing could be more self-contained and clear in certain places - for example, HyperMixer is not well described which makes it hard to assess the relationship between the methods.", "questions": "1. The authors claim that a score of 0.5 diagonality implies a uniform distribution and therefore, a simple average could achieve a similar token mixing. However, this is not obvious or clear to me. Could the authors explain?\n\n2. I was interested to know how the performance of summary mixing is impacted during inference by temporal averaging. As the sequence length increases, the sum vector becomes a mixture of more frames, degrading the ability to discriminate individual frames/tokens. Do the authors have any numbers showing ASR performance (WER) as a function of sequence length across models with self-attention and the proposed SummaryMixing ?\n\n3. As SLURP uses read speech, I was wondering if the authors had performed experiments on other corpora with spontaneous speech, for example, SLUE-VoxPopuli [1] for Named Entity Recognition (NER)?\n\n4. The paper mentions KWS results, but I don't see any in the results section.\n\n5. There are some papers [2,3,4] that show linear transformers can obtain comparable performance to MHSA for ASR contrary to what the authors claim. These works are also relevant and must be acknowledged in the paper. \n\n[1] \"SLUE Phase-2: A Benchmark Suite of Diverse Spoken Language Understanding Tasks\",\nSuwon Shon, Siddhant Arora, Chyi-Jiunn Lin, Ankita Pasad, Felix Wu, Roshan S Sharma, Wei-Lun Wu, Hung-yi Lee, Karen Livescu, Shinji Watanabe\n\n[2] Jingyu Sun, Guiping Zhong, Dinghao Zhou, Baoxiang Li, Yiran Zhong, \"Locality Matters: A Locality-Biased Linear Attention for Automatic Speech Recognition\"\n\n[3] \"Conformer-Based Speech Recognition with Linear Nyström Attention and Rotary Position Embedding\", \nLahiru Samarakoon, Tsun-Yat Leung\n\n[4] \"Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition\", Dima Rekesh et. al.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel method to replace self-attentions with a more computationally efficient summary mixing model for speech recognition and spoken language understanding. Rather than computing an NxN attention matrix, the authors propose to first compute the temporal mean of the sequence and combine that with local information extracted using a cgMLP branch. Experiments on multiple corpora show that the proposed approach can match the performance of standard self-attention and Branchformer.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The paper uses a very simple approach to substitute self-attention with attention between a temporal summary and the original vector that appears to work well in practice.\n\n2. Experiments are performed on multiple corpora for ASR (CommonVoice-Dutch, Italian, French), Librispeech, AISHELL, TED-LIUM 2, GSC and SLURP for SLU.", "weaknesses": "1. The paper makes the assumption that MHSA in the encoder is not necessary based on Zhang et. al. and Peng. et. al. However, I believe the assumption may not be well supported because (a) Zhang et al point out that you don't need self-attentions in higher layers, not that you don't need self-attentions at all, and (b) Peng. et al which uses self-attentions to model global context while using cgMLP to model local context. Therefore, the rationale behind this fundamental assumption made in the paper seems unclear. \n\n2. SLURP uses read speech - the impact of temporal averaging here may be minimal, but may be very harmful in spontaneous conversational speech. Such explorations in my view are important in this work to claim that SummaryMixing can reach top-of-the-line in speech understanding. \n\n3. Writing could be more self-contained and clear in certain places - for example, HyperMixer is not well described which makes it hard to assess the relationship between the methods.", "questions": "1. The authors claim that a score of 0.5 diagonality implies a uniform distribution and therefore, a simple average could achieve a similar token mixing. However, this is not obvious or clear to me. Could the authors explain?\n\n2. I was interested to know how the performance of summary mixing is impacted during inference by temporal averaging. As the sequence length increases, the sum vector becomes a mixture of more frames, degrading the ability to discriminate individual frames/tokens. Do the authors have any numbers showing ASR performance (WER) as a function of sequence length across models with self-attention and the proposed SummaryMixing ?\n\n3. As SLURP uses read speech, I was wondering if the authors had performed experiments on other corpora with spontaneous speech, for example, SLUE-VoxPopuli [1] for Named Entity Recognition (NER)?\n\n4. The paper mentions KWS results, but I don't see any in the results section.\n\n5. There are some papers [2,3,4] that show linear transformers can obtain comparable performance to MHSA for ASR contrary to what the authors claim. These works are also relevant and must be acknowledged in the paper. \n\n[1] \"SLUE Phase-2: A Benchmark Suite of Diverse Spoken Language Understanding Tasks\",\nSuwon Shon, Siddhant Arora, Chyi-Jiunn Lin, Ankita Pasad, Felix Wu, Roshan S Sharma, Wei-Lun Wu, Hung-yi Lee, Karen Livescu, Shinji Watanabe\n\n[2] Jingyu Sun, Guiping Zhong, Dinghao Zhou, Baoxiang Li, Yiran Zhong, \"Locality Matters: A Locality-Biased Linear Attention for Automatic Speech Recognition\"\n\n[3] \"Conformer-Based Speech Recognition with Linear Nyström Attention and Rotary Position Embedding\", \nLahiru Samarakoon, Tsun-Yat Leung\n\n[4] \"Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition\", Dima Rekesh et. al.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698616019331}, {"id": "goWrxaDWTQ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission5692/Reviewer_gqu7"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a linear-complexity based architecture as a replacement of self-attention for speech recognition to improve the efficiency. The proposed block has a local branch and a global branch to take both information into account. The local branch is based on a MLP, and the global branch has a MLP and then an average pooling among all the fram", "review_text": "This paper proposes a linear-complexity based architecture as a replacement of self-attention for speech recognition to improve the efficiency. The proposed block has a local branch and a global branch to take both information into account. The local branch is based on a MLP, and the global branch has a MLP and then an average pooling among all the fram", "strengths": "* Paper fits ICLR scope well.\n* The idea of reducing model complexity through replacing self-attention with linear layers is interesting.\n* The proposed method is solid.", "weaknesses": "* Presentations need improvement. See question sections for details.\n* Literature reviews on ASR are very limited. Please have an individual paragraph in Section 1.1 to describe those.\n* Novelty is a concern. The proposed work is an incremental work on the top of BranchFormer and HyperMixer. Although there is a paragraph describing the relationship to HyperMixer, the difference seems to be more on the choice of linear functions. Please explain more on this.\n* The improvements from the proposed method seems to depend on the use of a transformer decoder (with self-attention). Need more experiments to verify the parity of the proposed method to self-attention in other conditions. See question sections for details.", "questions": "Abstract: “However, attention layers in trained speech recognizers tend to not capture fine-grained pair-wise information.” Please explain what “fine-grained pair-wise information” is and why it is important (although they are explained in the following sections, abstract should be self-explanatory).\n\nFigure 1: The diagrams are a bit confusing. I understand that “T” refers to the total number of frames. However, audiences may likely think of them as different attention heads. Please consider a better way for the diagrams.\n\nSection 2.1 - Multi-head SummaryMixing. Please better explain the heads in SummaryMixing. In self-attention, heads are operating in parallel with different sets of parameters. However, here heads help to reduce the number of parameters. Please consider adding a diagram on this if it helps. In addition, by dividing x_t into n chunks, is it similar to chuck-wise attention or attentions with a limited context window? If that’s the case, please avoid using “head” here.\n\nSection 3.3.1: The results shown in this section have a transformer decoder of 6-blocks-MHSA paired with each system. In this setup, the decoder will capture the global context, even if the encoder doesn’t. However, we can only conclude that self-attention is redundant when self-attention based decoders are used.\nThe authors also conducted experiments without a transformer decoder (i.e., CTC only) on Branchformer in the supplemental materials. However, there are two more problems to figure out. 1. How does vanilla Conformer/Transformer with CTC performance in this setup? They are the most common architectures for productions than Branchformer. 2. These results are obtained with LM shallow fusion, which can be expensive for deployment. How would the proposed approach perform without shallow fusion? If the language information is needed, please also consider RNN-T decoder.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a linear-complexity based architecture as a replacement of self-attention for speech recognition to improve the efficiency. The proposed block has a local branch and a global branch to take both information into account. The local branch is based on a MLP, and the global branch has a MLP and then an average pooling among all the fram", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "* Paper fits ICLR scope well.\n* The idea of reducing model complexity through replacing self-attention with linear layers is interesting.\n* The proposed method is solid.", "weaknesses": "* Presentations need improvement. See question sections for details.\n* Literature reviews on ASR are very limited. Please have an individual paragraph in Section 1.1 to describe those.\n* Novelty is a concern. The proposed work is an incremental work on the top of BranchFormer and HyperMixer. Although there is a paragraph describing the relationship to HyperMixer, the difference seems to be more on the choice of linear functions. Please explain more on this.\n* The improvements from the proposed method seems to depend on the use of a transformer decoder (with self-attention). Need more experiments to verify the parity of the proposed method to self-attention in other conditions. See question sections for details.", "questions": "Abstract: “However, attention layers in trained speech recognizers tend to not capture fine-grained pair-wise information.” Please explain what “fine-grained pair-wise information” is and why it is important (although they are explained in the following sections, abstract should be self-explanatory).\n\nFigure 1: The diagrams are a bit confusing. I understand that “T” refers to the total number of frames. However, audiences may likely think of them as different attention heads. Please consider a better way for the diagrams.\n\nSection 2.1 - Multi-head SummaryMixing. Please better explain the heads in SummaryMixing. In self-attention, heads are operating in parallel with different sets of parameters. However, here heads help to reduce the number of parameters. Please consider adding a diagram on this if it helps. In addition, by dividing x_t into n chunks, is it similar to chuck-wise attention or attentions with a limited context window? If that’s the case, please avoid using “head” here.\n\nSection 3.3.1: The results shown in this section have a transformer decoder of 6-blocks-MHSA paired with each system. In this setup, the decoder will capture the global context, even if the encoder doesn’t. However, we can only conclude that self-attention is redundant when self-attention based decoders are used.\nThe authors also conducted experiments without a transformer decoder (i.e., CTC only) on Branchformer in the supplemental materials. However, there are two more problems to figure out. 1. How does vanilla Conformer/Transformer with CTC performance in this setup? They are the most common architectures for productions than Branchformer. 2. These results are obtained with LM shallow fusion, which can be expensive for deployment. How would the proposed approach perform without shallow fusion? If the language information is needed, please also consider RNN-T decoder.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698101740720}], "openreview_url": "https://openreview.net/forum?id=PoBB8n52oi", "arxiv_id": "2307.07421", "paper_pdf": "papers/PoBB8n52oi.pdf", "paper_pdf_sha256": "83e672dec3700172524e1d3c3deb24b78e26f07578abc6a132c6ec2bf6c4d83e", "paper_pdf_bytes": 293159, "paper_pdf_source": "openreview", "code_url": "https://github.com/SamsungLabs/SummaryMixing", "code_repository": "SamsungLabs/SummaryMixing", "code_commit": "d1b1f425149cea28f0e0318de82e525af2523b70", "code_archive": "repos/PoBB8n52oi.zip", "code_archive_sha256": "d534a7932c656f9138cd44286daf34eca03cb6ad206a64e31688a5439f84c857", "code_archive_bytes": 460239, "code_file_count": 7, "code_extensions": {".py": 7}, "github_disk_usage_kb": 491, "github_languages": {"Python": 148286}, "github_archived": false, "github_pushed_at": "2025-06-24T09:23:38Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sumformer-a-linear-complexity-alternative-to"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "me09xlTmm8", "year": 2023, "status": "rejected", "title": "Transport with Support: Data-Conditional Diffusion Bridges", "authors": ["Ella Tamir", "Martin Trapp", "Arno Solin"], "authorids": ["~Ella_Tamir1", "~Martin_Trapp2", "~Arno_Solin1"], "authors_source": "OpenReview API", "abstract": "The dynamic Schrödinger bridge problem provides an appealing setting for posing optimal transport problems as learning non-linear diffusion processes and enables efficient iterative solvers. Recent works have demonstrated state-of-the-art results (eg, in modelling single-cell embryo RNA sequences or sampling from complex posteriors) but are typically limited to learning bridges with only initial and terminal constraints. Our work extends this paradigm by proposing the Iterative Smoothing Bridge (ISB). We combine learning diffusion models with Bayesian filtering and optimal control, allowing for constrained stochastic processes governed by sparse observations at intermediate stages and terminal constraints. We assess the effectiveness of our method on synthetic and real-world data and show that the ISB generalises well to high-dimensional data, is computationally efficient, and provides accurate estimates of the marginals at intermediate and terminal times. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "b3FW_7ioQMB", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6427/Reviewer_V4yF"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose to add sparse constraints to the original Schrodinger bridges through optimal control. Specifically, the paper assumes that there exist some intermediate sparse samples during the diffusion process. By modifying the Iterative Proportional Fitting procedure (IPFP) method with spare intermediate constraints, the Iterative Smoothing Bridge (ISB) method is proposed. Experiments show that the ISB method can help the forward and backward drift functions successfully evolve toward the intermediate observations.", "review_text": "Generally, the paper proposes an interesting problem. But the ambiguity in writing and implementation makes it hard to follow.", "strengths": "Strength\n- The proposed problem may be important in many different applications with sparse intermediate observations, especially in the medical area.\n- It is reasonable to use $L^2$ loss to handle the forward and backward drifts.\n- Experimental results are convincing. \n\nWeakness\n- No convergence guarantee of the proposed method.\n- For step 2 and 4, since there only exists sparse intermediate observations, to make the algorithm converge, it seems that a large number of samples is needed to make the method converge.\n- It is reasonable to assume both $g$ and $\\beta$ the same in both equation (5) and (6)?\n- I may miss this part in the paper. Empirically, how to define $g$?\n- The second paragraph of Step 1 in Sec. 3.1 is not very clear.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "In this paper, the authors propose to add sparse constraints to the original Schrodinger bridges through optimal control. Specifically, the paper assumes that there exist some intermediate sparse samples during the diffusion process. By modifying the Iterative Proportional Fitting procedure (IPFP) method with spare intermediate constraints, the Iterative Smoothing Bridge (ISB) method is proposed. Experiments show that the ISB method can help the forward and backward drift functions successfully evolve toward the intermediate observations.", "strength_and_weaknesses": "Strength\n- The proposed problem may be important in many different applications with sparse intermediate observations, especially in the medical area.\n- It is reasonable to use $L^2$ loss to handle the forward and backward drifts.\n- Experimental results are convincing. \n\nWeakness\n- No convergence guarantee of the proposed method.\n- For step 2 and 4, since there only exists sparse intermediate observations, to make the algorithm converge, it seems that a large number of samples is needed to make the method converge.\n- It is reasonable to assume both $g$ and $\\beta$ the same in both equation (5) and (6)?\n- I may miss this part in the paper. Empirically, how to define $g$?\n- The second paragraph of Step 1 in Sec. 3.1 is not very clear.", "clarity,_quality,_novelty_and_reproducibility": "Clarity and Quality\n- Sec. 3.1 is somewhat ambiguous. The authors need to take more effort on the writing to make the logic much clearer. \n- The results of IPFP should be included in Fig. 3 for comparison.\n- There are also some typos:\n    - In the fourth-to-last line of page 4, $b_{l,\\phi}$ should be $b_{l-1,\\phi}$.\n    - In the second line under equation (7), it should be $f_{l-1, \\theta}$ and $g_{l+1, \\phi}$\n    - In the experiment of **Single-cell embryo RNA sequences**, why is the PCA used? What happens if the experiment is conducted on the original data?\n\nNovelty\n- The paper proposes an interesting problem, and the solution seems work.\n\nReproducibility\n- Without source code, the work is hard to reproduce.", "summary_of_the_review": "Generally, the paper proposes an interesting problem. But the ambiguity in writing and implementation makes it hard to follow.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667411990558}, {"id": "FDPBC-s1PT", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6427/Reviewer_3a4i"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the problem of learning a dynamical process given sparse observations of the process at intermediate times using a modification of the Schrodinger bridge process. In particular, the authors propose a method they term Iterative Smoothing Bridge (ISB) which alternates fitting forward and backward drifts parametrized by neural networks with a particle filtering and update step. This latter step is included to incorporate potentially sparse observations of the intermediate time dynamics. The authors discuss connections between their methods and stochastic control and give some theoretical claims. They then give an empirical study of their method.\n", "review_text": "While this work proposes a new method -- ISB -- for the important problem of learning dynamical processes with sparse observations, it does not achieve strong empirical performance and has limited conceptual novelty. For these reasons I think it is slightly too weak to merit acceptance.", "strengths": "Strengths:\n- They study an important problem: that of learning dynamical processes given some sparse observations at intermediate times\n\n- The ISB method is novel\n\n- They achieve some improved empirical results when compared against related methods in the literature\n\n- The work is fairly well written\n\nWeaknesses:\n- While their method does offer some improvement compared to related methods, it is not a significant improvement, see for example table 2, figure 6, figure 7\n\n- The authors give little theoretical justification for their method, beyond Prop. 3 which is not rigorously proved and in any case is essentially well-known. \n\n- The core elements of the method appeared in prior works, namely De Bortoli et al 2021 and Corenflos et al 2021. The novelty of their method is in the combination of these approaches.\n\nWriting feedback:\n- Typos in background paragraph on page 3: should be $\\mathcal{C} = C([0, T]; \\mathbb{R}^d)$. There's an extra \"to\" in the sentence that begins with $x_t$\n- The differentiable re-sampling procedure is hardly explained at all, and only a passing reference to Corenflos et al is given. Comprehension would be greatly aided by giving some discussion of this method, if only in an appendix. Also, it seems that the differentiability of the re-sampling method is not used in your method - is this correct?\n\nDe Bortoli, V., Thornton, J., Heng, J., & Doucet, A. (2021). Diffusion Schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34, 17695-17709.\n\nCorenflos, A., Thornton, J., Deligiannidis, G., & Doucet, A. (2021, July). Differentiable particle filtering via entropy-regularized optimal transport. In International Conference on Machine Learning (pp. 2100-2111). PMLR.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies the problem of learning a dynamical process given sparse observations of the process at intermediate times using a modification of the Schrodinger bridge process. In particular, the authors propose a method they term Iterative Smoothing Bridge (ISB) which alternates fitting forward and backward drifts parametrized by neural networks with a particle filtering and update step. This latter step is included to incorporate potentially sparse observations of the intermediate time dynamics. The authors discuss connections between their methods and stochastic control and give some theoretical claims. They then give an empirical study of their method.\n", "strength_and_weaknesses": "Strengths:\n- They study an important problem: that of learning dynamical processes given some sparse observations at intermediate times\n\n- The ISB method is novel\n\n- They achieve some improved empirical results when compared against related methods in the literature\n\n- The work is fairly well written\n\nWeaknesses:\n- While their method does offer some improvement compared to related methods, it is not a significant improvement, see for example table 2, figure 6, figure 7\n\n- The authors give little theoretical justification for their method, beyond Prop. 3 which is not rigorously proved and in any case is essentially well-known. \n\n- The core elements of the method appeared in prior works, namely De Bortoli et al 2021 and Corenflos et al 2021. The novelty of their method is in the combination of these approaches.\n\nWriting feedback:\n- Typos in background paragraph on page 3: should be $\\mathcal{C} = C([0, T]; \\mathbb{R}^d)$. There's an extra \"to\" in the sentence that begins with $x_t$\n- The differentiable re-sampling procedure is hardly explained at all, and only a passing reference to Corenflos et al is given. Comprehension would be greatly aided by giving some discussion of this method, if only in an appendix. Also, it seems that the differentiability of the re-sampling method is not used in your method - is this correct?\n\nDe Bortoli, V., Thornton, J., Heng, J., & Doucet, A. (2021). Diffusion Schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34, 17695-17709.\n\nCorenflos, A., Thornton, J., Deligiannidis, G., & Doucet, A. (2021, July). Differentiable particle filtering via entropy-regularized optimal transport. In International Conference on Machine Learning (pp. 2100-2111). PMLR.\n", "clarity,_quality,_novelty_and_reproducibility": "As listed above.", "summary_of_the_review": "While this work proposes a new method -- ISB -- for the important problem of learning dynamical processes with sparse observations, it does not achieve strong empirical performance and has limited conceptual novelty. For these reasons I think it is slightly too weak to merit acceptance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667098611964}, {"id": "ymxFJsOhRnm", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6427/Reviewer_JvCw"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Authors present computationally efficient framework for learning data conditional diffusion bridges using Iterative Smoothing Bridge. The proposed framework is assessed by experimental results on both synthetic and real world data.\n", "review_text": "The paper is well written, presents novel framework with experimental assessment of it. Proofs of propositions on which the framework relies are not full but it seems that they are more or less true based on provided sketches. ", "strengths": "Strength:\n1. strong theoretical formulation of the proposed approach\n2. well explained algorithm\n3. well performed experiments\n\nWeaknesses:\n1. lack of rigorous proofs of main propositions (proof sketches of proposition 1 and 2)", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "Authors present computationally efficient framework for learning data conditional diffusion bridges using Iterative Smoothing Bridge. The proposed framework is assessed by experimental results on both synthetic and real world data.\n", "strength_and_weaknesses": "Strength:\n1. strong theoretical formulation of the proposed approach\n2. well explained algorithm\n3. well performed experiments\n\nWeaknesses:\n1. lack of rigorous proofs of main propositions (proof sketches of proposition 1 and 2)", "clarity,_quality,_novelty_and_reproducibility": "Proposed problem is novel, well done experiment section, notation is not standard for stochastic calculus which makes it harder to read", "summary_of_the_review": "The paper is well written, presents novel framework with experimental assessment of it. Proofs of propositions on which the framework relies are not full but it seems that they are more or less true based on provided sketches. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666639516383}, {"id": "SNJ2fqLqTJ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper6427/Reviewer_aFhC"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors proposed the iterative smoothing bridge by leveraging ideas from Schr\\\"{o}dinger bridge and particle filter. Such a method allows us to learn constrained stochastic processes governed by sparse observations at intermediate stages. The algorithm is evaluated on synthetic data, real data, and small-scale high-dimensional data. This is an interesting problem and worth a deeper investigation and the proposed methodology seems to be a reasonable candidate.", "review_text": "An interesting method to solve an important problem, not clear/scalable enough though.", "strengths": "\n**Pros:** \n\n1. This paper is proposed to address an interesting problem, a path-constrained Schrodinger bridge, where a particle filter kicks in to tackle the sparse observations; \n\n**Cons:**\n\nWhile I am not an expert in the field of particle filters, my biggest concern is the clarity: \n\n1. more background introduction on particle filters may be needed for non-expert readers in the appendix; \n2. what is the optimal transport map **$\\boldsymbol{T}_{\\varepsilon}$**;\n3. why do we need the H-nearest neighbours?\n4. maybe I am wrong, i think simulated annealing only proposed to gradually decrease the noise. Why does decreasing and then increasing the noise scale resemble simulated annealing? \n5. the resampling steps may be detailed.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors proposed the iterative smoothing bridge by leveraging ideas from Schr\\\"{o}dinger bridge and particle filter. Such a method allows us to learn constrained stochastic processes governed by sparse observations at intermediate stages. The algorithm is evaluated on synthetic data, real data, and small-scale high-dimensional data. This is an interesting problem and worth a deeper investigation and the proposed methodology seems to be a reasonable candidate.", "strength_and_weaknesses": "\n**Pros:** \n\n1. This paper is proposed to address an interesting problem, a path-constrained Schrodinger bridge, where a particle filter kicks in to tackle the sparse observations; \n\n**Cons:**\n\nWhile I am not an expert in the field of particle filters, my biggest concern is the clarity: \n\n1. more background introduction on particle filters may be needed for non-expert readers in the appendix; \n2. what is the optimal transport map **$\\boldsymbol{T}_{\\varepsilon}$**;\n3. why do we need the H-nearest neighbours?\n4. maybe I am wrong, i think simulated annealing only proposed to gradually decrease the noise. Why does decreasing and then increasing the noise scale resemble simulated annealing? \n5. the resampling steps may be detailed.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: See above\n\nQuality and Novelty: I think leveraging particle filters to address the path-aware kind of Schrodinger bridge is a promising extension. Didn't check the proof.\n\n\nReproducibility: Although I like the insights proposed in De Bortoli (2021), the method alone is not that scalable. My question is that does this method extend to the likelihood training framework [1]? \n\n[1] Likelihood Training of Schrödinger Bridge using Forward-Backward SDEs Theory\n", "summary_of_the_review": "An interesting method to solve an important problem, not clear/scalable enough though.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "NA", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666447005640}], "openreview_url": "https://openreview.net/forum?id=me09xlTmm8", "arxiv_id": "2301.13636", "paper_pdf": "papers/me09xlTmm8.pdf", "paper_pdf_sha256": "1ace3aef22405df3a03de89bcbab28e1ab9a95ffbba75c4075fafc332c8b0d10", "paper_pdf_bytes": 3063866, "paper_pdf_source": "openreview", "code_url": "https://github.com/AaltoML/iterative-smoothing-bridge", "code_repository": "AaltoML/iterative-smoothing-bridge", "code_commit": "12263a6f0be10786e4ff05c567fdc5b750f1b297", "code_archive": "repos/me09xlTmm8.zip", "code_archive_sha256": "170804791aea8d023588f021026982e977918238411f1401853615ee7139ccbf", "code_archive_bytes": 1299216, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 1225, "github_languages": {"Python": 150201}, "github_archived": false, "github_pushed_at": "2023-11-24T08:58:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/transport-with-support-data-conditional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "dHJtoaE3yRP", "year": 2022, "status": "rejected", "title": "NAFS: A Simple yet Tough-to-Beat Baseline for Graph Representation Learning", "authors": ["Wentao Zhang", "Zeang Sheng", "Mingyu Yang", "Yang Li", "Yu Shen", "Zhi Yang", "Zichao Yang", "Bin CUI"], "authorids": ["~Wentao_Zhang1", "~Zeang_Sheng1", "~Mingyu_Yang2", "~Yang_Li36", "~Yu_Shen3", "~Zhi_Yang4", "~Zichao_Yang1", "~Bin_CUI2"], "authors_source": "OpenReview API", "abstract": "Recently, graph neural networks (GNNs) have shown prominent performance in graph representation learning by leveraging knowledge from both graph structure and node features. However, most of them have two major limitations. First, GNNs can learn higher-order structural information by stacking more layers but can not deal with large depth due to the over-smoothing issue.  Second, it is not easy to apply these methods on large graphs due to the expensive computation cost and high memory usage. In this paper, we present node-adaptive feature smoothing (NAFS), a simple non-parametric method that constructs node representations without parameter learning. NAFS first extracts the features of each node with its neighbors of different hops by feature smoothing, and then adaptively combines the smoothed features. Besides, the constructed node representation can further be enhanced by the ensemble of smoothed features extracted via different smoothing strategies. We conduct experiments on four benchmark datasets on two different application scenarios: node clustering and link prediction. Remarkably, NAFS with feature ensemble outperforms the state-of-the-art GNNs on these tasks and mitigates the aforementioned two limitations of most learning-based GNN counterparts. ", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "QkBl6Emo9fX", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper608/Reviewer_SHxg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents NAFS (Node-Adaptive Feature Smoothing), a method that constructs node representations by relying on smoothing only, i.e. without parameter learning. To do this, the authors first provide a formulation for the smoothing operator after infinite steps, i.e. when the stationary state is reached. They then define over-smoothing distance as a way to assess how much a node is close to the stationary state after k smoothing steps. Finally, they use over-smoothing distances to calculate a different smoothing weight for each node.\n\nExperiments show that representations obtained by smoothing with these weights, together with feature ensembles obtained by applying different convolution coefficients, provide performances in clustering and link prediction tasks which are comparable, if not better, than many other state-of-the-art approaches.", "review_text": "The paper tackles the problem of over-smoothing by proposing a method to differently weight the smoothing for each node. The problem is well known and in my opinion methods to better understand and solve it are very relevant for the community.\n\n**Positive aspects**\n\nWhile the connection of over-smoothing with powers of the adjacency matrix is not novel per se, I did not know about prior methods tackling the problem as it has been done in this paper. However, I should highlight that just recently another **very** similar paper was published on arxiv (Zhang et Al. \"Node Dependent Local Smoothing for Scalable Graph Learning\" https://arxiv.org/abs/2110.14377 ). While it seems recent enough to not put the contribute of this paper into discussion, I think it would be worth citing it and comparing its method to the presented one.\n\nThe approach is presented quite clearly and in a well-structured way. I appreciate that the authors introduce it as a baseline: better results might be obtained by adding a learning component to it, but it is interesting to see how far adaptive smoothing alone can go.\n\n**Main concerns and chances for improvement**\n\nBy being introduced as parameterless, NAFS is compared with other methods (autoencoders, adversarial models) which are typically unsupervised. Moreover, the only supervised task it is tested on is the link prediction one. I think further experiments are required to better understand how useful the method can be. In particular, I think it would be interesting to see how it performs on other supervised tasks such as node classification (e.g. suing the linear evaluation protocol, used to assess the performances of many recent graph SSL methods).\n\nThe weighting scheme looks intuitively useful, but I believe it would be useful to understand how it actually impacts final representations from a more theoretical perspective. How do weights change node representations from converging to the same identical one? \n\nFew minor corrections:\n- page 1: \"and these methods share two major limitations\" (\"and\" looks superfluous)\n- page 2: $\\hat{A} = A + I_n$ should be $\\tilde{A} = A + I_n$\n- the $r$ convolution coefficient which is introduced in Section 4.2 previously appears in Equation (1) and this makes the equation less clear. I think it would be easier for the reader to have its description close to (1) instead. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper presents NAFS (Node-Adaptive Feature Smoothing), a method that constructs node representations by relying on smoothing only, i.e. without parameter learning. To do this, the authors first provide a formulation for the smoothing operator after infinite steps, i.e. when the stationary state is reached. They then define over-smoothing distance as a way to assess how much a node is close to the stationary state after k smoothing steps. Finally, they use over-smoothing distances to calculate a different smoothing weight for each node.\n\nExperiments show that representations obtained by smoothing with these weights, together with feature ensembles obtained by applying different convolution coefficients, provide performances in clustering and link prediction tasks which are comparable, if not better, than many other state-of-the-art approaches.", "main_review": "The paper tackles the problem of over-smoothing by proposing a method to differently weight the smoothing for each node. The problem is well known and in my opinion methods to better understand and solve it are very relevant for the community.\n\n**Positive aspects**\n\nWhile the connection of over-smoothing with powers of the adjacency matrix is not novel per se, I did not know about prior methods tackling the problem as it has been done in this paper. However, I should highlight that just recently another **very** similar paper was published on arxiv (Zhang et Al. \"Node Dependent Local Smoothing for Scalable Graph Learning\" https://arxiv.org/abs/2110.14377 ). While it seems recent enough to not put the contribute of this paper into discussion, I think it would be worth citing it and comparing its method to the presented one.\n\nThe approach is presented quite clearly and in a well-structured way. I appreciate that the authors introduce it as a baseline: better results might be obtained by adding a learning component to it, but it is interesting to see how far adaptive smoothing alone can go.\n\n**Main concerns and chances for improvement**\n\nBy being introduced as parameterless, NAFS is compared with other methods (autoencoders, adversarial models) which are typically unsupervised. Moreover, the only supervised task it is tested on is the link prediction one. I think further experiments are required to better understand how useful the method can be. In particular, I think it would be interesting to see how it performs on other supervised tasks such as node classification (e.g. suing the linear evaluation protocol, used to assess the performances of many recent graph SSL methods).\n\nThe weighting scheme looks intuitively useful, but I believe it would be useful to understand how it actually impacts final representations from a more theoretical perspective. How do weights change node representations from converging to the same identical one? \n\nFew minor corrections:\n- page 1: \"and these methods share two major limitations\" (\"and\" looks superfluous)\n- page 2: $\\hat{A} = A + I_n$ should be $\\tilde{A} = A + I_n$\n- the $r$ convolution coefficient which is introduced in Section 4.2 previously appears in Equation (1) and this makes the equation less clear. I think it would be easier for the reader to have its description close to (1) instead. \n", "summary_of_the_review": "The paper tackles an interesting problem in a clear and reasonable way. The main issues I see here are the evaluation, that to me still looks quite limited, and the lack of a theoretical interpretation for the weighting scheme.\n\nWhile a possibly negative outcome of the evaluation of a node classification task does not concern me too much (the method itself is defined as a baseline and I think it does not require to be state-of-the-art in every single benchmark for it to be valuable), I think motivating the weighting scheme by formally demonstrating how it impacts smoothing when $K$ grows is an important step to make the paper more convincing.\n\n----\nI have read the authors' replies to the other reviewers and myself and they look convincing to me. I particularly appreciate they addressed my comments on related works, theoretical analysis, and further experiments on supervised tasks. They also took care of adding mean/std results for both new and old experiments (even when results are deterministic, which is great to emphasize). This makes the paper way more convincing in my opinion and worth being accepted, which is why I increased my score accordingly.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636186500117}, {"id": "tZHz1pgNNgy", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper608/Reviewer_w2Qg"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors of this paper took a novel perspective to present the node-adaptive feature smoothing (NAFS) algorithm, which generates node embeddings without explicit training/parameter learning. The method first performs feature smoothing, then combines the smoothed features using adaptive weights which are node-specific. They further enhanced this method by ensembling the smoothed features extracted with different hyper-parameters. The authors have conducted many experiments to validate the model performance, and demonstrate the model's efficiency empirically. ", "review_text": "The authors took a novel perspective and presented NAFS which does not explicitly require parameter learning. It is also a bold idea that, in addition to separate feature transformation from feature smoothing, the model removes feature transformation altogether. \n\nA few questions:\n1. Since the method has no feature transformation step, the construction of node feature in the initial step seems to become rather important. What are the features used for each of the dataset in the experiment (Cora, Citeseer, PubMed, Wiki)? I would be good to state them clearly in the paper. \n\n2. Would be useful to comment on what the authors think might be the limitation of this method. For example, NAFS by design does not require task specific training, although hyper-parameter tuning allow the model to serve for different tasks. Then what types of graphs/tasks might work best with this model, or is it indifferent.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors of this paper took a novel perspective to present the node-adaptive feature smoothing (NAFS) algorithm, which generates node embeddings without explicit training/parameter learning. The method first performs feature smoothing, then combines the smoothed features using adaptive weights which are node-specific. They further enhanced this method by ensembling the smoothed features extracted with different hyper-parameters. The authors have conducted many experiments to validate the model performance, and demonstrate the model's efficiency empirically. ", "main_review": "The authors took a novel perspective and presented NAFS which does not explicitly require parameter learning. It is also a bold idea that, in addition to separate feature transformation from feature smoothing, the model removes feature transformation altogether. \n\nA few questions:\n1. Since the method has no feature transformation step, the construction of node feature in the initial step seems to become rather important. What are the features used for each of the dataset in the experiment (Cora, Citeseer, PubMed, Wiki)? I would be good to state them clearly in the paper. \n\n2. Would be useful to comment on what the authors think might be the limitation of this method. For example, NAFS by design does not require task specific training, although hyper-parameter tuning allow the model to serve for different tasks. Then what types of graphs/tasks might work best with this model, or is it indifferent.\n", "summary_of_the_review": "In summary, the paper is well presented. The motivations and the authors' insights to this model is well explained. The method is clearly described. The authors performed many experiments to validate the model's performance, with additional ablation studies. Would be nice to have more discussion on which scenarios the model works better/worse.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635973732624}, {"id": "-Cbe1lWgS9H", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper608/Reviewer_YM4P"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper deals with (unsupervised) learning with graphs, specifically node-level tasks. Inspired by spectral GNNs, specifically Graph Convolutional Networks (GCN), the authors propose a simple neighborhood smoothing technique to capture the graph structure around each node in the given graph. Contrary to GNN, the proposed algorithm is parameter-free and hence scales better than their end-to-end trained GNN counterpart.                    \n\nMotivated by the problem of over-smoothing of GCN (Li et al., 2018), they propose the so-called \"over-smoothing distance\" measuring how close a node's feature is to be over-smoothed, which is simply the row-wise distance between $A^k X$ and $A^{\\infty}X$ with regard to the Euclidian distance, see Def. 3.1. Based on this distance they define the smoothing weight matrix which is then used to weight the neighboring node features during neighborhood aggregation. The resulting features $\\hat{X}$ are computed as $\\hat{X} =  \\sum^{K}_{k=0} W(k)\\hat{A}X$, where $W(k)$ is the $k$th smoothing weight matrix.\n\nThe proposed architecture is evaluated on standard, old, small-scale (unsupervsied) link and node classification (Cora, Citeseer, PubMed) tasks showing somewhat classification performance while showing a considerable speedup in computation time. ", "review_text": "**Strengths:**\n1. Simple approach that seems to work well\n2. Well-presented and easy to read\n3. Rather large set of baseline methods used\n\n**Weaknesses** \n1. Old, small-scale datasets\n2. No error bars/standard deviations were reported \n3. Scalability experiments only carried out on synthetic graphs\n\n**Suggestion**\n1. Evaluate the method on more large-scale, modern datasets, e.g. OGB datasets or graphlearning.io\n2. Perform scalability on real-world datasets \n3. Discuss \"linear\" GNN architectures, e.g., https://arxiv.org/abs/1810.05997\n4. Discuss other scalable GNN alternatives, e.g., ones based on label propagation, see https://arxiv.org/abs/2010.13993\n\n**Comments**\n1.  The discussion in section 2.2 only applies to a specific GNN layer, namely GCN \n\n**Question**\n1. Is your approach also liftable to other GNN architectures besides GCN?\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper deals with (unsupervised) learning with graphs, specifically node-level tasks. Inspired by spectral GNNs, specifically Graph Convolutional Networks (GCN), the authors propose a simple neighborhood smoothing technique to capture the graph structure around each node in the given graph. Contrary to GNN, the proposed algorithm is parameter-free and hence scales better than their end-to-end trained GNN counterpart.                    \n\nMotivated by the problem of over-smoothing of GCN (Li et al., 2018), they propose the so-called \"over-smoothing distance\" measuring how close a node's feature is to be over-smoothed, which is simply the row-wise distance between $A^k X$ and $A^{\\infty}X$ with regard to the Euclidian distance, see Def. 3.1. Based on this distance they define the smoothing weight matrix which is then used to weight the neighboring node features during neighborhood aggregation. The resulting features $\\hat{X}$ are computed as $\\hat{X} =  \\sum^{K}_{k=0} W(k)\\hat{A}X$, where $W(k)$ is the $k$th smoothing weight matrix.\n\nThe proposed architecture is evaluated on standard, old, small-scale (unsupervsied) link and node classification (Cora, Citeseer, PubMed) tasks showing somewhat classification performance while showing a considerable speedup in computation time. ", "main_review": "**Strengths:**\n1. Simple approach that seems to work well\n2. Well-presented and easy to read\n3. Rather large set of baseline methods used\n\n**Weaknesses** \n1. Old, small-scale datasets\n2. No error bars/standard deviations were reported \n3. Scalability experiments only carried out on synthetic graphs\n\n**Suggestion**\n1. Evaluate the method on more large-scale, modern datasets, e.g. OGB datasets or graphlearning.io\n2. Perform scalability on real-world datasets \n3. Discuss \"linear\" GNN architectures, e.g., https://arxiv.org/abs/1810.05997\n4. Discuss other scalable GNN alternatives, e.g., ones based on label propagation, see https://arxiv.org/abs/2010.13993\n\n**Comments**\n1.  The discussion in section 2.2 only applies to a specific GNN layer, namely GCN \n\n**Question**\n1. Is your approach also liftable to other GNN architectures besides GCN?\n\n", "summary_of_the_review": "The approach is very simple, it offers no significant theoretical or methodological contributions, and the experimental study is not executed well enough, i.e., it only uses small-scale, old or synthetic datasets. Further, the are some problems in the experimental protocol, e.g., no standard deviations are reported. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635534451778}], "openreview_url": "https://openreview.net/forum?id=dHJtoaE3yRP", "arxiv_id": "2206.08583", "paper_pdf": "papers/dHJtoaE3yRP.pdf", "paper_pdf_sha256": "5be907ba1067a1c11b4ec9d2370f2356011d2955424006bd5f54c294e0de35ab", "paper_pdf_bytes": 1524273, "paper_pdf_source": "openreview", "code_url": "https://github.com/PKU-DAIR/NAFS", "code_repository": "PKU-DAIR/NAFS", "code_commit": "966cc58a1df85f2179f69f9785c00cf572f0c121", "code_archive": "repos/dHJtoaE3yRP.zip", "code_archive_sha256": "d34bf65108da728217f64bd47eaa34ae2786848d9b8fa033db0e36a639bb067e", "code_archive_bytes": 5710826, "code_file_count": 11, "code_extensions": {".py": 10, ".sh": 1}, "github_disk_usage_kb": 5879, "github_languages": {"Python": 34978, "Shell": 1867}, "github_archived": false, "github_pushed_at": "2022-06-20T02:49:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/nafs-a-simple-yet-tough-to-beat-baseline-for-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Jq8JGA89sDa", "year": 2021, "status": "rejected", "title": "Detecting Hallucinated Content in Conditional Neural Sequence Generation", "authors": ["Chunting Zhou", "Jiatao Gu", "Mona T. Diab", "Paco Guzmán", "Luke Zettlemoyer", "Marjan Ghazvininejad"], "authorids": ["~Chunting_Zhou1", "~Jiatao_Gu1", "~Mona_T._Diab1", "fguzman@fb.com", "~Luke_Zettlemoyer1", "~Marjan_Ghazvininejad1"], "authors_source": "OpenReview API", "abstract": "Neural sequence models can generate highly fluent sentences but recent studies have also shown that they are also prone to hallucinate additional content not supported by the input, which can cause a lack of trust in the model.\nTo better assess the faithfulness of the machine outputs, we propose a new task to predict whether each token in the output sequence is hallucinated conditioned on the source input, and collect new manually annotated evaluation sets for this task.\nWe also introduce a novel method for learning to model hallucination detection, based on pretrained language models fine tuned on synthetic data that includes automatically inserted hallucinations.  \nExperiments on machine translation and abstract text summarization demonstrate the effectiveness of our proposed approach -- we obtain an average F1 of around 0.6 across all the benchmark datasets.\nFurthermore, we demonstrate how to use the token-level hallucination labels to define a fine-grained loss over the target sequence in the low-resource machine translation and achieve significant improvements over strong baseline methods.\nWe will also release our annotated data and code for future research.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "gMcBPcwKWmq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2741/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: The paper addresses the problem of \"hallucinated\" content in conditional neural generation for two specific tasks: machine translation and summarization. It proposes a new task for faithfulness assessment, which classifies each token as either hallucinated or not. The classifier uses a pre-trained LM (either XLM-R or ROBERTa) and is fine-tuned on synthetic classification data created using both 'noisified' real data and a pretrained LM (BART). Experiments on either summarization and MT system outputs labeled for hallucinations show relatively encouraging classification results (e.g., F1 of 0.46 to 0.66 for MT, and 0.56 to 0.66 for summarization).\n \nPros:\n \nAddressing the hallucination problem is essential, as inserting non-factual content is probably the most harmful error a machine translation or summarization system can make.\n \nThe paper contributes a new dataset for training classifiers for hallucination detection and a methodology for creating new ones relatively easily. That said, the dataset is partially synthetic and may not reflect the kind of hallucinations actual neural MT and summarization systems make.\n \nThe paper contributes an evaluation task for assessing the faithfulness of MT and summarization output. It could be used, e.g., to flag or prevent output whose content is deemed inaccurate. The paper shows an improved correlation (Spearman) relative to an entailment metric and a word alignment metric.\n \nCons:\n \n1. Lack of experimental comparison: I find it hard to conclude much from the paper's main results, as it doesn't provide a comparison across different hallucination detection systems or models. In general, it is difficult to tell if a given absolute score (e.g., F1 = 0.60 mentioned in the introduction) is good or bad without a point of comparison, especially if the task is new and there is relatively little information about the type of content errors the underlying systems make. I understand that the authors are trying to establish a new task and that there is little work to compare against, but the paper could have provided more in the way of ablation experiments.\n\n2. Results lack interpretability: Two of the main tables (Table 2 and 4) do not seem to contribute much, as the different rows correspond to different datasets (underlying data to classify is different) and are therefore strictly not comparable. For example, Tab. 2 suggests that the classifier does a much better job on TransS2S (F1 = 0.66) than on MBART (F1 = 0.47). However, Tab. 5 indicates that MBART is much less prone to producing hallucinated words than TransS2S, which probably explains why classification results on MBART data are worse. So it seems Tab. 2 and 5 together suggest classification results are lower when the classification task is harder, which wouldn't be a surprising finding.   \n\n3. Contribution as a reference-free evaluation metric is suspect and possibly subject to gaming: The authors point out that nearly all existing evaluation metrics (BLEU, METEOR, BLEURT, BERTScore) require reference text, which is characterized in the paper as a disadvantage. This is a bit of an odd argument to make for an *evaluation* metric, as this implies both the metric and the model are only given a source string to achieve their purposes. If the evaluation metric isn't given additional information (in the form of, e.g., a gold standard or reference), then on what basis is the metric supposed to do a better job at assessing the goodness of a target string, given that both the metric and the model are machine-learned using the same amount of input signal? If an evaluation metric is reference-free and proves indeed useful, then one can and probably should incorporate it directly inside the model or system (e.g., as a feature function or in a fine-tuning setup).  But incorporating the hallucination detection model as part of the generation model would render the former model rather useless as an *evaluation* metric. Note: I understand there is also a reference-based setup (which makes little difference), but the reference-free aspect is emphasized in the paper.\n\n4.  No evaluation on any downstream task: But this could easily have been done as the model for hallucination detection does not require references: see point 3.\n \nOverall, I think this paper's general direction with fine-tuning on hallucination classification data is probably worth pursuing. However, I think the paper's actual impact in its current form could be fairly limited, given the cons listed above. It is hard to draw conclusions from the main results of the paper (cons 1-2); the use as an evaluation metric is rather questionable (con 3); there is no application of the work on any end-to-end task (con 4).\n \nOther comments:\n \nTables 2 and 4 are somewhat misleading in their references to different \"models.\" Each of these tables evaluates hallucination classification with a *single* model, but the Models column refers to different models used to generate the underlying classification *data*. I think this could easily be misunderstood by the reader, as the established practice is that \"models\" refer to models for the task at hand (i.e., classification).\n \nExperiments on hallucination detection are on MT and summarization, but the paper (and introduction in particular) discuss more generally \"conditional neural sequence models\" and mention other tasks such as free-form QA (Fan et al., 2019) and indirectly LM-based generation (Wang and Sennrich, 2020). By being reference-free, the work would, however, only apply to semantics-preserving tasks (output conveys and same meaning as the input or a subset) as summarization as MT. Otherwise, I don't see how to evaluate whether something is hallucinated without any reference.\n \nMissed related work:\n\nImproved Natural Language Generation via Loss Truncation. Daniel Kang, Tatsunori Hashimoto. ACL 2020.\n\nSticking to the Facts: Confident Decoding for Faithful Data-to-Text Generation. Ran Tian, Shashi Narayan, Thibault Sellam, Ankur P. Parikh.  arXiv 2019.\n\n\n===========\nUpdate: Thank you for your clarifications and updated paper, which addressed several of my concerns. I therefore increased my score by 1.\n\n> First, we want to stress that we are not proposing a reference-free evaluation metric for quality estimation. \n\nI understand it is not meant as a general quality estimation metric, and my point was more about the *reference-free* aspect. I tried to make a broader point which I think applies to any kind of reference-free metric (whether it is for quality estimation or specifically about hallucination).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting approach to identify hallucination, but have doubts about its use as an evaluation metric", "review": "Summary: The paper addresses the problem of \"hallucinated\" content in conditional neural generation for two specific tasks: machine translation and summarization. It proposes a new task for faithfulness assessment, which classifies each token as either hallucinated or not. The classifier uses a pre-trained LM (either XLM-R or ROBERTa) and is fine-tuned on synthetic classification data created using both 'noisified' real data and a pretrained LM (BART). Experiments on either summarization and MT system outputs labeled for hallucinations show relatively encouraging classification results (e.g., F1 of 0.46 to 0.66 for MT, and 0.56 to 0.66 for summarization).\n \nPros:\n \nAddressing the hallucination problem is essential, as inserting non-factual content is probably the most harmful error a machine translation or summarization system can make.\n \nThe paper contributes a new dataset for training classifiers for hallucination detection and a methodology for creating new ones relatively easily. That said, the dataset is partially synthetic and may not reflect the kind of hallucinations actual neural MT and summarization systems make.\n \nThe paper contributes an evaluation task for assessing the faithfulness of MT and summarization output. It could be used, e.g., to flag or prevent output whose content is deemed inaccurate. The paper shows an improved correlation (Spearman) relative to an entailment metric and a word alignment metric.\n \nCons:\n \n1. Lack of experimental comparison: I find it hard to conclude much from the paper's main results, as it doesn't provide a comparison across different hallucination detection systems or models. In general, it is difficult to tell if a given absolute score (e.g., F1 = 0.60 mentioned in the introduction) is good or bad without a point of comparison, especially if the task is new and there is relatively little information about the type of content errors the underlying systems make. I understand that the authors are trying to establish a new task and that there is little work to compare against, but the paper could have provided more in the way of ablation experiments.\n\n2. Results lack interpretability: Two of the main tables (Table 2 and 4) do not seem to contribute much, as the different rows correspond to different datasets (underlying data to classify is different) and are therefore strictly not comparable. For example, Tab. 2 suggests that the classifier does a much better job on TransS2S (F1 = 0.66) than on MBART (F1 = 0.47). However, Tab. 5 indicates that MBART is much less prone to producing hallucinated words than TransS2S, which probably explains why classification results on MBART data are worse. So it seems Tab. 2 and 5 together suggest classification results are lower when the classification task is harder, which wouldn't be a surprising finding.   \n\n3. Contribution as a reference-free evaluation metric is suspect and possibly subject to gaming: The authors point out that nearly all existing evaluation metrics (BLEU, METEOR, BLEURT, BERTScore) require reference text, which is characterized in the paper as a disadvantage. This is a bit of an odd argument to make for an *evaluation* metric, as this implies both the metric and the model are only given a source string to achieve their purposes. If the evaluation metric isn't given additional information (in the form of, e.g., a gold standard or reference), then on what basis is the metric supposed to do a better job at assessing the goodness of a target string, given that both the metric and the model are machine-learned using the same amount of input signal? If an evaluation metric is reference-free and proves indeed useful, then one can and probably should incorporate it directly inside the model or system (e.g., as a feature function or in a fine-tuning setup).  But incorporating the hallucination detection model as part of the generation model would render the former model rather useless as an *evaluation* metric. Note: I understand there is also a reference-based setup (which makes little difference), but the reference-free aspect is emphasized in the paper.\n\n4.  No evaluation on any downstream task: But this could easily have been done as the model for hallucination detection does not require references: see point 3.\n \nOverall, I think this paper's general direction with fine-tuning on hallucination classification data is probably worth pursuing. However, I think the paper's actual impact in its current form could be fairly limited, given the cons listed above. It is hard to draw conclusions from the main results of the paper (cons 1-2); the use as an evaluation metric is rather questionable (con 3); there is no application of the work on any end-to-end task (con 4).\n \nOther comments:\n \nTables 2 and 4 are somewhat misleading in their references to different \"models.\" Each of these tables evaluates hallucination classification with a *single* model, but the Models column refers to different models used to generate the underlying classification *data*. I think this could easily be misunderstood by the reader, as the established practice is that \"models\" refer to models for the task at hand (i.e., classification).\n \nExperiments on hallucination detection are on MT and summarization, but the paper (and introduction in particular) discuss more generally \"conditional neural sequence models\" and mention other tasks such as free-form QA (Fan et al., 2019) and indirectly LM-based generation (Wang and Sennrich, 2020). By being reference-free, the work would, however, only apply to semantics-preserving tasks (output conveys and same meaning as the input or a subset) as summarization as MT. Otherwise, I don't see how to evaluate whether something is hallucinated without any reference.\n \nMissed related work:\n\nImproved Natural Language Generation via Loss Truncation. Daniel Kang, Tatsunori Hashimoto. ACL 2020.\n\nSticking to the Facts: Confident Decoding for Faithful Data-to-Text Generation. Ran Tian, Shashi Narayan, Thibault Sellam, Ankur P. Parikh.  arXiv 2019.\n\n\n===========\nUpdate: Thank you for your clarifications and updated paper, which addressed several of my concerns. I therefore increased my score by 1.\n\n> First, we want to stress that we are not proposing a reference-free evaluation metric for quality estimation. \n\nI understand it is not meant as a general quality estimation metric, and my point was more about the *reference-free* aspect. I tried to make a broader point which I think applies to any kind of reference-free metric (whether it is for quality estimation or specifically about hallucination).", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604303806821}, {"id": "6WccRTbT1OI", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2741/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "SUMMARY\n\nThis paper presents a method to detect hallucinated tokens in generations from neural machine translation and summarization. Given a source input S and its output G generated by a sequence generation model, this paper formalizes the task of detecting hallucinated tokens as a labeling problem on the output G. In order to train the labeler, the method synthetically generates supervision data by using a BART model. The BART model receives a text with noises ([MASK] tokens) and tries to predict [MASK] tokens. In this way, the method obtains a pseudo hallucinated text T' from a text T, and assigns hallucination labels by estimating edit operations between T and T'. The labeler is trained by fine-tuning pre-trained cross-lingual (for MT) and mono-lingual (for summarization) language models. In training, the labeler receives a source text S, true target text T, and pseudo hallucinated text T' separated by [SEP] tokens and tries to reproduce the hallucination labels on T'. Receiving a source text S and its output G, the labeler predicts hallucination labels on G during the inference time.\n\nThe experiments use the XSUM dataset for summarization and multi-domain Chinese-English (Zh-En) translation dataset for machine translation. The authors manually prepared test sets for evaluating the hallucination labeler. In Section 6.2, the authors reported that they achieved decent performance on this task of hallucination labeling although it is far from being solved. They also aggregated token-level hallucination predictions to sentence-level scoring and computed the Spearman's rank correlation coefficients between the scores and the percentage of tokens annotated as hallucinations by humans. The comparison between the proposed method and two baselines (entail and align) show that the proposed scores correlate well with human judges than the baselines.\n\nPROS\n\nThis paper is clearly written and easy to read.\n\nThe methods used in this study, e.g., BART for generating pseudo hallucinated text and (XLM-)RoBERTa for labeling hallucinated tokens, are appropriate.\n\nCONS\n\nFindings from the experiments are unclear. In addition, this paper does not explicitly explain how the research outcomes contribute to an advance in MT or summarization. In particular, I'm not sure whether the performance of Table 2 was successful enough to advance the research in MT and summarization. The goal of the correlation analysis in Table 3 is unclear. I guess that the initial motivation for this experiment was to develop an automatic method for evaluating hallucinations, but the current experimental results do not support this so strongly.\n\nQUESTIONS\n\nPlease report the number of hallucinated and faithful sentences in the test set. This paper explains that the test set for machine translation included 250 instances initially, but removed instances judged as incomprehensible. I have no idea how many translations were hallucinated and faithful. This may be important because the Entail baseline works not at token level but at sentence level (it should not be used on a dataset only with non-entailment instances).\n\nI would like to see results of a simple baseline in Table 2. For example, we can consider a baseline that indicates hallucinations for tokens that do not appear in the source text. Currently, I'm not sure how good the performance values of Table 1 are. Hence, I have no idea whether the method for generating pseudo supervision data and for detecting hallucinations at token level was a worthy contribution.\n\nI am wondering of the usefulness of the evaluation with Spearman's rank correlation coefficients. Here is a summary of the methods.\n\nReference values: the percentage of tokens annotated as hallucinations by humans.\n\nProposed model (1): the average of hallucination probabilities across all the tokens.\n\nProposed model (2): the percentage of tokens labeled as hallucinations by the model.\n\nEntail: the probability estimate of entailment relation between the source and output.\n\nAlign: the percentage of tokens aligned to the source input estimated by SimAlign.\n\nThe Entail baseline was not designed to work at token level but at sentence level. There is no guarantee that a probability estimate from the baseline represents the 'degree' of entailment relation that corresponds to the percentage of tokens. The comparison would be fairer if the proposed model could yield a binary decision about hallucination/faithful at sentence level.\n\nIn addition, Spearman's rank correlation coefficients of the Entail baseline on the MT dataset were mostly negative (weak inverse correlation). If -0.32 could be obtained by chance, we can conclude that coefficients around 0.3-0.4 do not indicate a correlation. I also suspect that this baseline could not work well because most instances on the MT dataset were non-entailment instances.\n\nP3: Figure 2\n\nDoes the BART model often produce multiple tokens (e.g., \"with friend happily\") from a single [MASK] token? This example is impressive, but I think that a BERT model mostly predicts a single token for a single [MASK] token.\n\nP3: \"we consider any spans in G that misrepresent S as hallucinated contents\"\n\nThis treatment may cause an inconsistency in labeling hallucinated tokens. How does this treatment affect the evaluations in Section 5?\n\nP4: It was difficult to follow the description of \"paraphrase\". How does this study generate paraphrases?\n\nMINOR COMMENTS\n\nP7: \"aligned aligned\" -> \"aligned\"\n\nCOMMENTS AFTER THE REVISION\n\nIt was amazing to see the authors updated paper with the new experiment on the baselines (Table 2) and machine translation (Section 6). Although the baselines (overlap and synonym) were strong, I can now see that the proposed approach is better than these simple baselines both for machine translation and summarization. Section 6 also demonstrates the usefulness of this work for machine translation trained with self-training. For this reason, I increased my rating.\n\nThe impact of this paper would be greater if Section 6 could include more results on different MT datasets (e.g., WMT) and/or self-training for summarization.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #2", "review": "SUMMARY\n\nThis paper presents a method to detect hallucinated tokens in generations from neural machine translation and summarization. Given a source input S and its output G generated by a sequence generation model, this paper formalizes the task of detecting hallucinated tokens as a labeling problem on the output G. In order to train the labeler, the method synthetically generates supervision data by using a BART model. The BART model receives a text with noises ([MASK] tokens) and tries to predict [MASK] tokens. In this way, the method obtains a pseudo hallucinated text T' from a text T, and assigns hallucination labels by estimating edit operations between T and T'. The labeler is trained by fine-tuning pre-trained cross-lingual (for MT) and mono-lingual (for summarization) language models. In training, the labeler receives a source text S, true target text T, and pseudo hallucinated text T' separated by [SEP] tokens and tries to reproduce the hallucination labels on T'. Receiving a source text S and its output G, the labeler predicts hallucination labels on G during the inference time.\n\nThe experiments use the XSUM dataset for summarization and multi-domain Chinese-English (Zh-En) translation dataset for machine translation. The authors manually prepared test sets for evaluating the hallucination labeler. In Section 6.2, the authors reported that they achieved decent performance on this task of hallucination labeling although it is far from being solved. They also aggregated token-level hallucination predictions to sentence-level scoring and computed the Spearman's rank correlation coefficients between the scores and the percentage of tokens annotated as hallucinations by humans. The comparison between the proposed method and two baselines (entail and align) show that the proposed scores correlate well with human judges than the baselines.\n\nPROS\n\nThis paper is clearly written and easy to read.\n\nThe methods used in this study, e.g., BART for generating pseudo hallucinated text and (XLM-)RoBERTa for labeling hallucinated tokens, are appropriate.\n\nCONS\n\nFindings from the experiments are unclear. In addition, this paper does not explicitly explain how the research outcomes contribute to an advance in MT or summarization. In particular, I'm not sure whether the performance of Table 2 was successful enough to advance the research in MT and summarization. The goal of the correlation analysis in Table 3 is unclear. I guess that the initial motivation for this experiment was to develop an automatic method for evaluating hallucinations, but the current experimental results do not support this so strongly.\n\nQUESTIONS\n\nPlease report the number of hallucinated and faithful sentences in the test set. This paper explains that the test set for machine translation included 250 instances initially, but removed instances judged as incomprehensible. I have no idea how many translations were hallucinated and faithful. This may be important because the Entail baseline works not at token level but at sentence level (it should not be used on a dataset only with non-entailment instances).\n\nI would like to see results of a simple baseline in Table 2. For example, we can consider a baseline that indicates hallucinations for tokens that do not appear in the source text. Currently, I'm not sure how good the performance values of Table 1 are. Hence, I have no idea whether the method for generating pseudo supervision data and for detecting hallucinations at token level was a worthy contribution.\n\nI am wondering of the usefulness of the evaluation with Spearman's rank correlation coefficients. Here is a summary of the methods.\n\nReference values: the percentage of tokens annotated as hallucinations by humans.\n\nProposed model (1): the average of hallucination probabilities across all the tokens.\n\nProposed model (2): the percentage of tokens labeled as hallucinations by the model.\n\nEntail: the probability estimate of entailment relation between the source and output.\n\nAlign: the percentage of tokens aligned to the source input estimated by SimAlign.\n\nThe Entail baseline was not designed to work at token level but at sentence level. There is no guarantee that a probability estimate from the baseline represents the 'degree' of entailment relation that corresponds to the percentage of tokens. The comparison would be fairer if the proposed model could yield a binary decision about hallucination/faithful at sentence level.\n\nIn addition, Spearman's rank correlation coefficients of the Entail baseline on the MT dataset were mostly negative (weak inverse correlation). If -0.32 could be obtained by chance, we can conclude that coefficients around 0.3-0.4 do not indicate a correlation. I also suspect that this baseline could not work well because most instances on the MT dataset were non-entailment instances.\n\nP3: Figure 2\n\nDoes the BART model often produce multiple tokens (e.g., \"with friend happily\") from a single [MASK] token? This example is impressive, but I think that a BERT model mostly predicts a single token for a single [MASK] token.\n\nP3: \"we consider any spans in G that misrepresent S as hallucinated contents\"\n\nThis treatment may cause an inconsistency in labeling hallucinated tokens. How does this treatment affect the evaluations in Section 5?\n\nP4: It was difficult to follow the description of \"paraphrase\". How does this study generate paraphrases?\n\nMINOR COMMENTS\n\nP7: \"aligned aligned\" -> \"aligned\"\n\nCOMMENTS AFTER THE REVISION\n\nIt was amazing to see the authors updated paper with the new experiment on the baselines (Table 2) and machine translation (Section 6). Although the baselines (overlap and synonym) were strong, I can now see that the proposed approach is better than these simple baselines both for machine translation and summarization. Section 6 also demonstrates the usefulness of this work for machine translation trained with self-training. For this reason, I increased my rating.\n\nThe impact of this paper would be greater if Section 6 could include more results on different MT datasets (e.g., WMT) and/or self-training for summarization.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603802230289}, {"id": "k-oLq98Oxaq", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2741/AnonReviewer4"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "\nSummary: \nThis paper proposes hallucination detection at the token level, which predicts if each token in the generation output is hallucinated or faithful to the source input. In contrast, previous studies usually work on the sentence level. To create synthetic training data, a denoising pre-trained LM is first used to generate (potentially) unfaithful counterparts T’ of the references T. Then, token-level labels are obtained by comparing T and T’ via edit distance. Finally, a standard classification model is trained on the token-level labels by concatenating the source S, true and unfaithful targets T (T’).\n\n---\nPros:\n+ Overall, the paper is clearly presented with detailed experiments and ablation studies. It shows that the proposed method works better than existing entailment-based and alignment-based methods on the new task that it proposes.\n+ The new annotated data on token-level hallucination could possibly be useful for future research in similar directions.\n\n---\nCons:\n- My major concern is about the effectiveness of defining hallucination at the token level itself. First of all, it looks un-intuitive to me whether it makes sense to define hallucination for every type of token. As the task aims to label every token in the text sequence, I wonder how different types of tokens (e.g., tokens with different POS tags) contribute to the model hallucination. The proposed metrics treat all the tokens equally, while in reality, tokens such as noun phrases or verb, for example, may have a larger impact on the hallucination issue than prepositions or articles. Similarly, for the creation of synthetic training data, I wonder whether it makes sense to replace tokens with different POS tags uniformly. It will be good to present more analysis in terms of both quantitative results and case studies on this aspect. \n- Is the evaluation on all the tokens or only the hallucinated tokens? Could the decent performance (F1) of the proposed method come from the fact that most tokens are not hallucinated (labeled with 0)?\n- For comparisons on sentence-level hallucination (of abstractive summarization), why don’t you compare with the baselines presented in the related papers you listed directly? \n\nTypo: \"if each token in the machine output is **a** hallucinated or faithful to the source input\"\n\n---\nComments after reading author response and revised paper \n\nThanks for showing more results of how token-level hallucination detection can be useful/effective: e.g., (i) labels per POS tagging in Fig. 5 and (ii) application to low-resource MT in Sec. 6.\n\nFor (i), I don't think it resolved my question directly: \"The proposed metrics treat all the tokens equally, while in reality, tokens such as noun phrases or verb, for example, may have a larger impact on the hallucination issue than prepositions or articles.\" That is, a hallucinated NN, for example, might be worse than a hallucinated II, rather than asking how the labels are distributed by POS categories. So I still wonder if it makes sense (as a reliable metric) to measure hallucination at the token level (e.g., Table 5 Hal words %) but it remains unanswered.\n\nFor (ii), it is indeed an interesting plus to the paper, showing that the token-level labels appear to be useful for downstream applications (even if it's not quite meaningful to measure the % of hallucinated words). I would suggest doing more studies on downstream tasks as mentioned in your response to enhance the paper if you are \"not proposing a reference-free evaluation metric for quality estimation\" (which seems a bit contradictory to \"we hope to create a large-scale pretrained evaluation\nmodel for any datasets or models to be evaluated\" in the conclusion section btw). \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official blind review", "review": "\nSummary: \nThis paper proposes hallucination detection at the token level, which predicts if each token in the generation output is hallucinated or faithful to the source input. In contrast, previous studies usually work on the sentence level. To create synthetic training data, a denoising pre-trained LM is first used to generate (potentially) unfaithful counterparts T’ of the references T. Then, token-level labels are obtained by comparing T and T’ via edit distance. Finally, a standard classification model is trained on the token-level labels by concatenating the source S, true and unfaithful targets T (T’).\n\n---\nPros:\n+ Overall, the paper is clearly presented with detailed experiments and ablation studies. It shows that the proposed method works better than existing entailment-based and alignment-based methods on the new task that it proposes.\n+ The new annotated data on token-level hallucination could possibly be useful for future research in similar directions.\n\n---\nCons:\n- My major concern is about the effectiveness of defining hallucination at the token level itself. First of all, it looks un-intuitive to me whether it makes sense to define hallucination for every type of token. As the task aims to label every token in the text sequence, I wonder how different types of tokens (e.g., tokens with different POS tags) contribute to the model hallucination. The proposed metrics treat all the tokens equally, while in reality, tokens such as noun phrases or verb, for example, may have a larger impact on the hallucination issue than prepositions or articles. Similarly, for the creation of synthetic training data, I wonder whether it makes sense to replace tokens with different POS tags uniformly. It will be good to present more analysis in terms of both quantitative results and case studies on this aspect. \n- Is the evaluation on all the tokens or only the hallucinated tokens? Could the decent performance (F1) of the proposed method come from the fact that most tokens are not hallucinated (labeled with 0)?\n- For comparisons on sentence-level hallucination (of abstractive summarization), why don’t you compare with the baselines presented in the related papers you listed directly? \n\nTypo: \"if each token in the machine output is **a** hallucinated or faithful to the source input\"\n\n---\nComments after reading author response and revised paper \n\nThanks for showing more results of how token-level hallucination detection can be useful/effective: e.g., (i) labels per POS tagging in Fig. 5 and (ii) application to low-resource MT in Sec. 6.\n\nFor (i), I don't think it resolved my question directly: \"The proposed metrics treat all the tokens equally, while in reality, tokens such as noun phrases or verb, for example, may have a larger impact on the hallucination issue than prepositions or articles.\" That is, a hallucinated NN, for example, might be worse than a hallucinated II, rather than asking how the labels are distributed by POS categories. So I still wonder if it makes sense (as a reliable metric) to measure hallucination at the token level (e.g., Table 5 Hal words %) but it remains unanswered.\n\nFor (ii), it is indeed an interesting plus to the paper, showing that the token-level labels appear to be useful for downstream applications (even if it's not quite meaningful to measure the % of hallucinated words). I would suggest doing more studies on downstream tasks as mentioned in your response to enhance the paper if you are \"not proposing a reference-free evaluation metric for quality estimation\" (which seems a bit contradictory to \"we hope to create a large-scale pretrained evaluation\nmodel for any datasets or models to be evaluated\" in the conclusion section btw). \n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603759855366}, {"id": "whrv7HTlmer", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2741/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new task: Word-level hallucination detection in text generation, where the authors identified two scenarios of word-level hallucination (content insertion and incorrect substitution). \n\nThen the author synthesized hallucinating text by de-noisng auto encoder and use dynamic programming to infer hallucination labels. They are used to fine-tune pertained LMs for detecting hallucination. \n\nExperiments show that the fine-tuned LMs can achieve certain performance compared with human annotations, and that the token-level hallucination is correlated to sentence-level hallucination.\n\nConcerns:\n\n1) Hallucination is recognized as an important issue for text generation, and intuitively, detecting hallucination is also an important task. But the real benefit of detecting hallucination is unclear from this paper. I would be more convinced if the authors could show how their word-level hallucination can indeed help text generation. \n\nCurrently, the paper appears to be some fine-tuned LM for a new task (whose real benefit is unclear) with little technical depth. At this stage, I am unsure if this paper has enough contribution as an ICLR paper.\n\n2) The categorization of \"content insertion\" and \"incorrect substitution\" is unclear and confusing. \n\nIs there any breakdown analysis on \"content insertion\" and \"incorrect substitution\"? For the below example, what type is \"Monday\" given Input1 or Input2, respectively?\n\nInput1: We had a meeting in that room.\n\nInput2: We had a meeting.\n\nGen: We had a meeting Monday.\n\n3) There's very little technical development and empirical analysis on the role of synonyms and stop words in hallucination. \n\nA synonym is supposed to have the same meaning as the original word, and is not hallucination. Synonyms are actually desired in many applications, like paraphrase generation and summarization. \n\nStop words mainly serve for grammatical functions, and may not be directly related to content hallucination.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "new task", "review": "This paper proposes a new task: Word-level hallucination detection in text generation, where the authors identified two scenarios of word-level hallucination (content insertion and incorrect substitution). \n\nThen the author synthesized hallucinating text by de-noisng auto encoder and use dynamic programming to infer hallucination labels. They are used to fine-tune pertained LMs for detecting hallucination. \n\nExperiments show that the fine-tuned LMs can achieve certain performance compared with human annotations, and that the token-level hallucination is correlated to sentence-level hallucination.\n\nConcerns:\n\n1) Hallucination is recognized as an important issue for text generation, and intuitively, detecting hallucination is also an important task. But the real benefit of detecting hallucination is unclear from this paper. I would be more convinced if the authors could show how their word-level hallucination can indeed help text generation. \n\nCurrently, the paper appears to be some fine-tuned LM for a new task (whose real benefit is unclear) with little technical depth. At this stage, I am unsure if this paper has enough contribution as an ICLR paper.\n\n2) The categorization of \"content insertion\" and \"incorrect substitution\" is unclear and confusing. \n\nIs there any breakdown analysis on \"content insertion\" and \"incorrect substitution\"? For the below example, what type is \"Monday\" given Input1 or Input2, respectively?\n\nInput1: We had a meeting in that room.\n\nInput2: We had a meeting.\n\nGen: We had a meeting Monday.\n\n3) There's very little technical development and empirical analysis on the role of synonyms and stop words in hallucination. \n\nA synonym is supposed to have the same meaning as the original word, and is not hallucination. Synonyms are actually desired in many applications, like paraphrase generation and summarization. \n\nStop words mainly serve for grammatical functions, and may not be directly related to content hallucination.\n\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603574090283}], "openreview_url": "https://openreview.net/forum?id=Jq8JGA89sDa", "arxiv_id": "2011.02593", "paper_pdf": "papers/Jq8JGA89sDa.pdf", "paper_pdf_sha256": "5800aeaa63139b80c697661e6c98c643365dc9e079863fcf0f4f287b01b1a587", "paper_pdf_bytes": 582867, "paper_pdf_source": "openreview", "code_url": "https://github.com/violet-zct/fairseq-detect-hallucination", "code_repository": "violet-zct/fairseq-detect-hallucination", "code_commit": "be14c9f378f8b9c0447c3b3d080263966d157b8c", "code_archive": "repos/Jq8JGA89sDa.zip", "code_archive_sha256": "6b5b079ca2db61b113e6dd514c10e6ccfd7850ab728161159e6d32a4b66bf6f1", "code_archive_bytes": 6020127, "code_file_count": 467, "code_extensions": {".py": 418, ".sh": 33, ".cpp": 7, ".cu": 4, ".cuh": 2, ".lua": 2, ".h": 1}, "github_disk_usage_kb": 4052, "github_languages": {"Python": 2018748, "Cuda": 36414, "Shell": 32370, "C++": 15854, "Cython": 8858, "Lua": 4210}, "github_archived": false, "github_pushed_at": "2022-04-15T00:52:31Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/detecting-hallucinated-content-in-conditional-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "RfrdbJVvVf", "year": 2025, "status": "rejected", "title": "MatMamba: A Matryoshka State Space Model", "authors": ["Abhinav Shukla", "Sai Vemprala", "Aditya Kusupati", "Ashish Kapoor"], "authorids": ["~Abhinav_Shukla1", "~Sai_Vemprala1", "~Aditya_Kusupati1", "~Ashish_Kapoor1"], "authors_source": "OpenReview API", "abstract": "State Space Models (SSMs) like Mamba2 are a promising alternative to Transformers, with faster theoretical training and inference times -- especially for long context lengths. Recent work on Matryoshka Representation Learning -- and its application to Transformer backbones in works like MatFormer --  showed how to introduce nested granularities of smaller submodels in one universal elastic model. In this work, we present MatMamba: a state space model which combines Matryoshka-style learning with Mamba2, by modifying the block to contain nested dimensions to enable joint training and adaptive inference. MatMamba allows for efficient and adaptive deployment across various model sizes. We train a single large MatMamba model and are able to get a number of smaller nested models for free -- while maintaining or improving upon the performance of a baseline smaller model trained from scratch. We train language and image models at a variety of parameter sizes from 35M to 1.4B. Our results on ImageNet and FineWeb show that MatMamba models scale comparably to Transformers, while having more efficient inference characteristics. This makes MatMamba a practically viable option for deploying large-scale models in an elastic way based on the available inference compute.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zB0hHt5aT6", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8348/Reviewer_Rpty"], "rating": 6, "soundness": 4, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This work extends Matryoshka Representation Learning to Mamba2, a representative architecture for state-space models, proposing MatMamba. MatMamba introduces a training pipeline that generates numerous sub-models, each offering different performance-efficiency trade-offs, within a single training run. The authors analyze the effectiveness and scalability of the proposed method.", "review_text": "This work extends Matryoshka Representation Learning to Mamba2, a representative architecture for state-space models, proposing MatMamba. MatMamba introduces a training pipeline that generates numerous sub-models, each offering different performance-efficiency trade-offs, within a single training run. The authors analyze the effectiveness and scalability of the proposed method.", "strengths": "- This work validates the scalability of Matryoshka Representation Learning by demonstrating its applicability to state-space models (SSMs). It considers both Mamba-vision and Mamba-language models, confirming effectiveness across both tasks. \n- The methodology and problem setting are well-established and well-motivated.", "weaknesses": "- For the MatMamba-LM model family, only evaluation loss is reported. The submission would be strengthened by including evaluations on downstream tasks as well.", "questions": "- What is the main difference you observed between elastic inference on Matformer and MatMamba during the experiment? Since MatMamba essentially follows the approach used in Matformer, I'm curious if any unique challenges were encountered specific to the Mamba architecture.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work extends Matryoshka Representation Learning to Mamba2, a representative architecture for state-space models, proposing MatMamba. MatMamba introduces a training pipeline that generates numerous sub-models, each offering different performance-efficiency trade-offs, within a single training run. The authors analyze the effectiveness and scalability of the proposed method.", "soundness": 4, "presentation": 4, "contribution": 3, "strengths": "- This work validates the scalability of Matryoshka Representation Learning by demonstrating its applicability to state-space models (SSMs). It considers both Mamba-vision and Mamba-language models, confirming effectiveness across both tasks. \n- The methodology and problem setting are well-established and well-motivated.", "weaknesses": "- For the MatMamba-LM model family, only evaluation loss is reported. The submission would be strengthened by including evaluations on downstream tasks as well.", "questions": "- What is the main difference you observed between elastic inference on Matformer and MatMamba during the experiment? Since MatMamba essentially follows the approach used in Matformer, I'm curious if any unique challenges were encountered specific to the Mamba architecture.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730758876380}, {"id": "vpsHGZDa0u", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8348/Reviewer_JmGL"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This paper introduces MatMamba, a novel approach to State Space Models (SSMs) that combines Mamba model with Matryoshka-style learning. The proposed model embeds nested sub-models within a single super-architecture, allowing for dynamic inference across different granularities without the need for separate training. This setup facilitates scalable and efficient deployment in both vision and language models by dynamically adjusting model sizes based on compute availability. Through extensive testing on tasks such as image classification and language modeling, MatMamba demonstrates comparable or superior performance to baseline Mamba2 models while maintaining flexibility and efficiency.", "review_text": "This paper introduces MatMamba, a novel approach to State Space Models (SSMs) that combines Mamba model with Matryoshka-style learning. The proposed model embeds nested sub-models within a single super-architecture, allowing for dynamic inference across different granularities without the need for separate training. This setup facilitates scalable and efficient deployment in both vision and language models by dynamically adjusting model sizes based on compute availability. Through extensive testing on tasks such as image classification and language modeling, MatMamba demonstrates comparable or superior performance to baseline Mamba2 models while maintaining flexibility and efficiency.", "strengths": "(1) Adaptive Inference Resource: MatMamba allows the extraction of multiple nested submodels from one single model, enabling versatile deployment options, from edge devices to cloud settings, without the need for retraining. By using the Mix’n’Match approach, MatMamba optimizes for various performance-compute trade-offs, allowing resource allocation adjustments based on current needs.\n(2) Scalability and Performance: The model exhibits scalability on par with transformers and baseline Mamba2 models in both image and language tasks. The author provides multiple model sizes ranging from 30M parameters to 1.4B parameters, which demonstrates scalability.\n(3) The nested structure maintains a consistent metric space, ensuring that different granularities yield compatible outputs, which is beneficial for downstream applications like retrieval.", "weaknesses": "(1) Dependency on Explicitly Trained Granularities for Optimal Performance: While MatMamba’s Mix’n’Match approach offers flexibility, untrained granularities do not perform as effectively as explicitly optimized ones, showing degradation in performance and accuracy.\n(2) Limited Exploration of Self-Distillation Techniques: The paper mentions that self-distillation or other techniques could further enhance interpolation accuracy in untrained granularities, suggesting room for improvement in the training setup.\n(3) Limited Evaluation on Language Tasks: The author trained a large model up to 1.4B parameters, but did not conduct more experiments on text-generation benchmarks used to evaluate LLMs. I reckon that it would be better if the author could run the proposed MatMamba through datasets like MMLU, HellaSwag, GSM8K, ARC and etc.", "questions": "See the weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces MatMamba, a novel approach to State Space Models (SSMs) that combines Mamba model with Matryoshka-style learning. The proposed model embeds nested sub-models within a single super-architecture, allowing for dynamic inference across different granularities without the need for separate training. This setup facilitates scalable and efficient deployment in both vision and language models by dynamically adjusting model sizes based on compute availability. Through extensive testing on tasks such as image classification and language modeling, MatMamba demonstrates comparable or superior performance to baseline Mamba2 models while maintaining flexibility and efficiency.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "(1) Adaptive Inference Resource: MatMamba allows the extraction of multiple nested submodels from one single model, enabling versatile deployment options, from edge devices to cloud settings, without the need for retraining. By using the Mix’n’Match approach, MatMamba optimizes for various performance-compute trade-offs, allowing resource allocation adjustments based on current needs.\n(2) Scalability and Performance: The model exhibits scalability on par with transformers and baseline Mamba2 models in both image and language tasks. The author provides multiple model sizes ranging from 30M parameters to 1.4B parameters, which demonstrates scalability.\n(3) The nested structure maintains a consistent metric space, ensuring that different granularities yield compatible outputs, which is beneficial for downstream applications like retrieval.", "weaknesses": "(1) Dependency on Explicitly Trained Granularities for Optimal Performance: While MatMamba’s Mix’n’Match approach offers flexibility, untrained granularities do not perform as effectively as explicitly optimized ones, showing degradation in performance and accuracy.\n(2) Limited Exploration of Self-Distillation Techniques: The paper mentions that self-distillation or other techniques could further enhance interpolation accuracy in untrained granularities, suggesting room for improvement in the training setup.\n(3) Limited Evaluation on Language Tasks: The author trained a large model up to 1.4B parameters, but did not conduct more experiments on text-generation benchmarks used to evaluate LLMs. I reckon that it would be better if the author could run the proposed MatMamba through datasets like MMLU, HellaSwag, GSM8K, ARC and etc.", "questions": "See the weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730724810307}, {"id": "WjRm3B1VLw", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8348/Reviewer_jnCr"], "rating": 6, "soundness": 2, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "This paper presents MatMamba, which introduces a nested Matryoshka structure on a Mamba2 state space model. The resulting elastic network architecture can be used to generate multiple smaller sub-networks at deployment time without any additional training, similar to prior work in this area such as Matformer and Flextron. The training process involves gradient accumulation on 4 subnetworks with a single backward pass for parameter updates. MatMamba is evaluated on vision and language tasks with elastic network sizes ranging from 35M to 1.4B.", "review_text": "This paper presents MatMamba, which introduces a nested Matryoshka structure on a Mamba2 state space model. The resulting elastic network architecture can be used to generate multiple smaller sub-networks at deployment time without any additional training, similar to prior work in this area such as Matformer and Flextron. The training process involves gradient accumulation on 4 subnetworks with a single backward pass for parameter updates. MatMamba is evaluated on vision and language tasks with elastic network sizes ranging from 35M to 1.4B.", "strengths": "* The paper is very well-written, with the main ideas and results presented clearly. It was a pleasure to read.\n* Elastic networks have shown significant promise for Transformer networks, since they can be used to generate tailored networks for specific deployment scenarios at no additional training cost (once elastic training is complete). The paper makes a contribution in this relevant and promising direction.\n* Evaluation and results across vision and language looks strong overall.", "weaknesses": "* There doesn’t seem to be a systematic way of selecting smaller sub-networks given a deployment constraint. For a given parameter or latency budget, should all dimensions be proportionally reduced? How is this proportion decided?\n* Novelty: this work appears to be a straightforward application of Matryoshka principles to the SSM domain; apart from the obvious changes needed for SSMs, the training algorithm remains nearly identical to the Matformer work.\n* It’s not clear if the scaling trends continue beyond the ~1.5B parameter mark. Do the authors have any insights here? I understand that training larger networks can be quite resource-intensive, but at least a discussion on expected scaling trends would be useful.\n* I don’t agree with the claim made on line 456. Downstream task performance must be reported for MatMamba-LM using a standard framework like LM-Eval-Harness. While validation loss is a good starting point, performance on a variety of downstream tasks can provide a more complete picture of LLM performance (which is why most work dealing with LLMs reports these numbers).", "questions": "* Why is g set to 4? Have you explored other values? How does the method scale when g is lower/higher?\n* How is MatMamba's long context performance? SSM-based architectures can perform poorly on long-context tasks, and I’d be curious to see if MatMamba maintains this trend, or perhaps improves it. MatMamba’s superior performance at higher resolutions in the vision domain provide somewhat of an indication here.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents MatMamba, which introduces a nested Matryoshka structure on a Mamba2 state space model. The resulting elastic network architecture can be used to generate multiple smaller sub-networks at deployment time without any additional training, similar to prior work in this area such as Matformer and Flextron. The training process involves gradient accumulation on 4 subnetworks with a single backward pass for parameter updates. MatMamba is evaluated on vision and language tasks with elastic network sizes ranging from 35M to 1.4B.", "soundness": 2, "presentation": 4, "contribution": 3, "strengths": "* The paper is very well-written, with the main ideas and results presented clearly. It was a pleasure to read.\n* Elastic networks have shown significant promise for Transformer networks, since they can be used to generate tailored networks for specific deployment scenarios at no additional training cost (once elastic training is complete). The paper makes a contribution in this relevant and promising direction.\n* Evaluation and results across vision and language looks strong overall.", "weaknesses": "* There doesn’t seem to be a systematic way of selecting smaller sub-networks given a deployment constraint. For a given parameter or latency budget, should all dimensions be proportionally reduced? How is this proportion decided?\n* Novelty: this work appears to be a straightforward application of Matryoshka principles to the SSM domain; apart from the obvious changes needed for SSMs, the training algorithm remains nearly identical to the Matformer work.\n* It’s not clear if the scaling trends continue beyond the ~1.5B parameter mark. Do the authors have any insights here? I understand that training larger networks can be quite resource-intensive, but at least a discussion on expected scaling trends would be useful.\n* I don’t agree with the claim made on line 456. Downstream task performance must be reported for MatMamba-LM using a standard framework like LM-Eval-Harness. While validation loss is a good starting point, performance on a variety of downstream tasks can provide a more complete picture of LLM performance (which is why most work dealing with LLMs reports these numbers).", "questions": "* Why is g set to 4? Have you explored other values? How does the method scale when g is lower/higher?\n* How is MatMamba's long context performance? SSM-based architectures can perform poorly on long-context tasks, and I’d be curious to see if MatMamba maintains this trend, or perhaps improves it. MatMamba’s superior performance at higher resolutions in the vision domain provide somewhat of an indication here.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730707503602}, {"id": "Qf7HSJF4P7", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission8348/Reviewer_vNu2"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper presents MatMamba the application of Matryoshka techniques (Kusupati 2022) to models based on the Mamba2 architecture.\nThe authors apply the Matroyshka technique to most layers, leading to a close-to linear downscaling capability. \nThey can show that jointly learning one model at four granularities matches the performance of models trained at the single scales.\nFor vision experiments, the Mix'n'Match technique using layer dimensionalities in between granularities matches/surpasses an interpolated \nperformance, whereas in Language Modeling there is a degradation using this technique. The presented technique binds representations at \ndifferent model scales, making potential for e.g. more efficient speculative decoding techniques.", "review_text": "The paper presents MatMamba the application of Matryoshka techniques (Kusupati 2022) to models based on the Mamba2 architecture.\nThe authors apply the Matroyshka technique to most layers, leading to a close-to linear downscaling capability. \nThey can show that jointly learning one model at four granularities matches the performance of models trained at the single scales.\nFor vision experiments, the Mix'n'Match technique using layer dimensionalities in between granularities matches/surpasses an interpolated \nperformance, whereas in Language Modeling there is a degradation using this technique. The presented technique binds representations at \ndifferent model scales, making potential for e.g. more efficient speculative decoding techniques.", "strengths": "The authors show that Matryoshka-style model tying/training can be applied to Linear Transformers/Matrix SSMs/Fast-Weight Programmers.\nThey show results on how the Mix'n'Match interpolation degrades performance across modalities.", "weaknesses": "It is a combination of existing techniques.", "questions": "Can you give a citation / relevant prior work for the claim in L.456f ?\nThere is a typo in L.522.\nDo you think that the different Mix'n'Match behavior for language and image modeling is intrinsic to the modality or might be specific to MatMamba?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents MatMamba the application of Matryoshka techniques (Kusupati 2022) to models based on the Mamba2 architecture.\nThe authors apply the Matroyshka technique to most layers, leading to a close-to linear downscaling capability. \nThey can show that jointly learning one model at four granularities matches the performance of models trained at the single scales.\nFor vision experiments, the Mix'n'Match technique using layer dimensionalities in between granularities matches/surpasses an interpolated \nperformance, whereas in Language Modeling there is a degradation using this technique. The presented technique binds representations at \ndifferent model scales, making potential for e.g. more efficient speculative decoding techniques.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The authors show that Matryoshka-style model tying/training can be applied to Linear Transformers/Matrix SSMs/Fast-Weight Programmers.\nThey show results on how the Mix'n'Match interpolation degrades performance across modalities.", "weaknesses": "It is a combination of existing techniques.", "questions": "Can you give a citation / relevant prior work for the claim in L.456f ?\nThere is a typo in L.522.\nDo you think that the different Mix'n'Match behavior for language and image modeling is intrinsic to the modality or might be specific to MatMamba?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729943200551}], "openreview_url": "https://openreview.net/forum?id=RfrdbJVvVf", "arxiv_id": "2410.06718", "paper_pdf": "papers/RfrdbJVvVf.pdf", "paper_pdf_sha256": "97bf99cb9e07b4ffc98779861f48251b0230e99179dd8dab3948c505c32c5043", "paper_pdf_bytes": 653017, "paper_pdf_source": "openreview", "code_url": "https://github.com/GenRobo/MatMamba", "code_repository": "GenRobo/MatMamba", "code_commit": "3233d4c7e7652ec13f203eb7b03f339c181660f6", "code_archive": "repos/RfrdbJVvVf.zip", "code_archive_sha256": "f711545a72879fdb8851900a919098262fa920a9e88beceb643330d81858c844", "code_archive_bytes": 381431, "code_file_count": 21, "code_extensions": {".sh": 11, ".py": 10}, "github_disk_usage_kb": 408, "github_languages": {"Python": 159571, "Shell": 7875}, "github_archived": false, "github_pushed_at": "2024-11-21T16:45:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/matmamba-a-matryoshka-state-space-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "fH9eqpCcR3", "year": 2024, "status": "rejected", "title": "Multiple Physics Pretraining for Physical Surrogate Models", "authors": ["Michael McCabe", "Bruno Régaldo-Saint Blancard", "Liam Holden Parker", "Ruben Ohana", "Miles Cranmer", "Alberto Bietti", "Michael Eickenberg", "Siavash Golkar", "Geraud Krawezik", "Francois Lanusse", "Mariel Pettee", "Tiberiu Tesileanu", "Kyunghyun Cho", "Shirley Ho"], "authorids": ["~Michael_McCabe2", "~Bruno_Régaldo-Saint_Blancard1", "~Liam_Holden_Parker1", "~Ruben_Ohana1", "~Miles_Cranmer2", "~Alberto_Bietti1", "~Michael_Eickenberg5", "~Siavash_Golkar1", "gkrawezik@flatironinstitute.org", "~Francois_Lanusse2", "~Mariel_Pettee1", "~Tiberiu_Tesileanu1", "~Kyunghyun_Cho1", "~Shirley_Ho2"], "authors_source": "OpenReview API", "abstract": "We introduce multiple physics pretraining (MPP), an autoregressive task-agnostic pretraining approach for physical surrogate modeling. MPP involves training large surrogate models to predict the dynamics of multiple heterogeneous physical systems simultaneously by learning features that are broadly useful across diverse physical tasks. In order to learn effectively in this setting, we introduce a shared embedding and normalization strategy that projects the fields of multiple systems into a single shared embedding space. We validate the efficacy of our approach on both pretraining and downstream tasks over a broad fluid mechanics-oriented benchmark. We show that a single MPP-pretrained transformer is able to match or outperform task-specific baselines on all pretraining sub-tasks without the need for finetuning. For downstream tasks, we demonstrate that finetuning MPP-trained models results in more accurate predictions across multiple time-steps on new physics compared to training from scratch or finetuning pretrained video foundation models. We open-source our code and model weights trained at multiple scales for reproducibility and community experimentation. Video examples are included in the supplementary materials.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "2gLOUHgBd8", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2797/Reviewer_2tFL"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a multiple physics pretraining approach for surrogate modeling, which learns general useful features across diverse physical tasks with a shared embedding and normalization strategy. The experiment results show the proposed MPP-pretrained model outperforms task-specific baselines on all pretraining sub-tasks and also show superior finetuning results on new physics tasks.", "review_text": "This paper introduces a multiple physics pretraining approach for surrogate modeling, which learns general useful features across diverse physical tasks with a shared embedding and normalization strategy. The experiment results show the proposed MPP-pretrained model outperforms task-specific baselines on all pretraining sub-tasks and also show superior finetuning results on new physics tasks.", "strengths": "1. Constructing a physics-based foundational model and exploring multiple task pretraining for computational physics tasks is both interesting and beneficial.\n2. The experiments demonstrate impressive results in both pretraining sub-tasks as well as substantial transfer potential for new tasks with low-data system.\n3. The authors have conducted several training strategies to perform the pretraining effectively.", "weaknesses": "1. More attempts can be made to address the transferability problems between different types of physics equations. For instance, when discussing the Navier-Stokes (NS) equations, whether at high or low Reynolds numbers, the equation forms are quite similar, which allows for the investigation into whether a single model still possesses robust merged learning capabilities for more different classes of equations, such as diffusion equations and wave equations.\n\n2. What are the advantages and disadvantages of this new approach compared to the Finite Element Method (FEM) for systems with explicit control equations or empirical formulas? One of my biggest problem is whether the physics task should be treated as a purely data-driven problem, or if it ought to incorporate certain explicit priors or equation-based guidelines.\n\n3. For models without PDEs, how can we determine the similarity of multiple physics fields and whether they can be learned simultaneously? I’m concerned about the improved performances are due to the limited diversity of different physics tasks.\n\n4. In the appendix, experimental data from 1-step to 9-step show little change in the field, whether looking at ground truth or predicted solutions. If the selected time steps were longer, would the model still be able to accurately predict future changes?", "questions": "Overall, the purposes behind this work are valuable. However, there remain many questions that need to be addressed. Please consider to answer the above questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a multiple physics pretraining approach for surrogate modeling, which learns general useful features across diverse physical tasks with a shared embedding and normalization strategy. The experiment results show the proposed MPP-pretrained model outperforms task-specific baselines on all pretraining sub-tasks and also show superior finetuning results on new physics tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Constructing a physics-based foundational model and exploring multiple task pretraining for computational physics tasks is both interesting and beneficial.\n2. The experiments demonstrate impressive results in both pretraining sub-tasks as well as substantial transfer potential for new tasks with low-data system.\n3. The authors have conducted several training strategies to perform the pretraining effectively.", "weaknesses": "1. More attempts can be made to address the transferability problems between different types of physics equations. For instance, when discussing the Navier-Stokes (NS) equations, whether at high or low Reynolds numbers, the equation forms are quite similar, which allows for the investigation into whether a single model still possesses robust merged learning capabilities for more different classes of equations, such as diffusion equations and wave equations.\n\n2. What are the advantages and disadvantages of this new approach compared to the Finite Element Method (FEM) for systems with explicit control equations or empirical formulas? One of my biggest problem is whether the physics task should be treated as a purely data-driven problem, or if it ought to incorporate certain explicit priors or equation-based guidelines.\n\n3. For models without PDEs, how can we determine the similarity of multiple physics fields and whether they can be learned simultaneously? I’m concerned about the improved performances are due to the limited diversity of different physics tasks.\n\n4. In the appendix, experimental data from 1-step to 9-step show little change in the field, whether looking at ground truth or predicted solutions. If the selected time steps were longer, would the model still be able to accurately predict future changes?", "questions": "Overall, the purposes behind this work are valuable. However, there remain many questions that need to be addressed. Please consider to answer the above questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698745067380}, {"id": "TUST77H2jM", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2797/Reviewer_AmeY"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper introduces an autoregressive task-agnostic pretraining approach for physical surrogate modeling. As like foundation models, the proposed method proposes training large surrogate models to predict the dynamics of multiple heterogeneous physical systems simultaneously by learning features that are broadly useful across diverse physical tasks.", "review_text": "The paper introduces an autoregressive task-agnostic pretraining approach for physical surrogate modeling. As like foundation models, the proposed method proposes training large surrogate models to predict the dynamics of multiple heterogeneous physical systems simultaneously by learning features that are broadly useful across diverse physical tasks.", "strengths": "This paper builds on recent advances in deep learning for physical simulation to enable surrogate models to work for diverse physical systems with a few fine-tuning iterations. The exposition for the motivation is written crisply and convincing, and is generally easy to follow. The experiments are performed for diverse physical systems. The results show that large surrogate models outperform strong baselines, and even models with a relatively few parameters can learn such diverse physics evolutions and perform competitively.", "weaknesses": "Although the generalizing capability is significant, the computational cost of the proposed models is still concerning. The followings are associated questions: \n* How do the authors expect the computational cost of the proposed models and how does it compare against other baselines in Table 1?\n* It is still a bit unclear if the comparison reported in Table 1 is fair, since the parameters of the first 4 models do not look comparable.\n* Is the model applicable to 3D simulations?\n\nMany details of the inverse problems seem missing, and it makes assessing the performance of the proposed model a bit difficult in this regard. The followings are some of the questions:\n* How do you define the inverse problem and what model was used for the tasks?\n* How does the performance compare to other baselines?\n* What is the running time of the proposed model and how is it comparable to the other baselines?\n* Could the authors provide insights or results on the inverse problem for boundary condition?\n* Providing qualitative results would be helpful.\n \n**Minor comments** \n\nTypo in section 5.1: must handle all all systems and regimes without finetuning", "questions": "Please have a look at the weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces an autoregressive task-agnostic pretraining approach for physical surrogate modeling. As like foundation models, the proposed method proposes training large surrogate models to predict the dynamics of multiple heterogeneous physical systems simultaneously by learning features that are broadly useful across diverse physical tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "This paper builds on recent advances in deep learning for physical simulation to enable surrogate models to work for diverse physical systems with a few fine-tuning iterations. The exposition for the motivation is written crisply and convincing, and is generally easy to follow. The experiments are performed for diverse physical systems. The results show that large surrogate models outperform strong baselines, and even models with a relatively few parameters can learn such diverse physics evolutions and perform competitively.", "weaknesses": "Although the generalizing capability is significant, the computational cost of the proposed models is still concerning. The followings are associated questions: \n* How do the authors expect the computational cost of the proposed models and how does it compare against other baselines in Table 1?\n* It is still a bit unclear if the comparison reported in Table 1 is fair, since the parameters of the first 4 models do not look comparable.\n* Is the model applicable to 3D simulations?\n\nMany details of the inverse problems seem missing, and it makes assessing the performance of the proposed model a bit difficult in this regard. The followings are some of the questions:\n* How do you define the inverse problem and what model was used for the tasks?\n* How does the performance compare to other baselines?\n* What is the running time of the proposed model and how is it comparable to the other baselines?\n* Could the authors provide insights or results on the inverse problem for boundary condition?\n* Providing qualitative results would be helpful.\n \n**Minor comments** \n\nTypo in section 5.1: must handle all all systems and regimes without finetuning", "questions": "Please have a look at the weakness above.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "Nothing particular.", "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698701832676}, {"id": "PX7rJYnBtu", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2797/Reviewer_ffib"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors present Multiple Physics Pretraining (MPP), a methodology for task-agnostic pretraining of surrogate models in physics. This technique facilitates pretraining on a large scale, allowing for knowledge transfer across a variety of physical domains. To handle the problem that features of physical tasks are diverse, a shared embedding and normalization strategy are introduced to project the fields of multiple systems into a single shared embedding space. The experiments show that a single MPP-pretrained transformer can output competitive or outperform task-specific baselines ( fluid mechanics-oriented benchmark) on all pretraining sub-tasks without finetuning.", "review_text": "The authors present Multiple Physics Pretraining (MPP), a methodology for task-agnostic pretraining of surrogate models in physics. This technique facilitates pretraining on a large scale, allowing for knowledge transfer across a variety of physical domains. To handle the problem that features of physical tasks are diverse, a shared embedding and normalization strategy are introduced to project the fields of multiple systems into a single shared embedding space. The experiments show that a single MPP-pretrained transformer can output competitive or outperform task-specific baselines ( fluid mechanics-oriented benchmark) on all pretraining sub-tasks without finetuning.", "strengths": "- The idea to construct a large pre-trained base model for physical simulations is promising. The current learning based surrogate models usually have limited generalization ability, which require re-training from scratch given different governing equations. The pre-trained model has the potential to improve the generalization ability or make it possible only to fine-tune on specific tasks without training from scratch.\n- In the experiments part, the proposed model can achieve one order of magnitude smaller errors comparing to the existing methods.", "weaknesses": "- The experiments for validating the proposed model are currently limited to 2-D cases (incompressible and compressiable Navier-stokes equations,  shallow-water equations, and a 2D DiffusionReaction equation). It is unclear whether the proposed model can achieve same level of performance and accuracy when applied to more realistic 3D physical simulations.\n- It seems that the current model mainly focus on predicting the solution one time step further. The video results (diffre, incompNS, mpp_swe) shown in the supplemental materials exhibits strong checkboard artifacts as the timesteps grow.\n- The proposed model can only handle simulations on structured mesh. However unstructured mesh (e.g., triangular, tetrahedra) is a more common choice in real world simulations. In addition, the multi-resolution i.e., mesh with adaptive resolution, which is also a common technique in large scale simulations, has not been taken into consideration.", "questions": "- It is mentioned in the appendix that the model is trained for 500 epochs for one task. How long does the training process take?\n- Are there any intuitive explanation for the checkboard artifacts as the timesteps grow shown in the video? Does it mean the proposed method is numerical unstable to some extent?\n- Does the proposed method have the potential to be applied to 3D simulations while keeping the efficiency and accuracy? (as the max memory usage of current model is ~60GB)", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present Multiple Physics Pretraining (MPP), a methodology for task-agnostic pretraining of surrogate models in physics. This technique facilitates pretraining on a large scale, allowing for knowledge transfer across a variety of physical domains. To handle the problem that features of physical tasks are diverse, a shared embedding and normalization strategy are introduced to project the fields of multiple systems into a single shared embedding space. The experiments show that a single MPP-pretrained transformer can output competitive or outperform task-specific baselines ( fluid mechanics-oriented benchmark) on all pretraining sub-tasks without finetuning.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The idea to construct a large pre-trained base model for physical simulations is promising. The current learning based surrogate models usually have limited generalization ability, which require re-training from scratch given different governing equations. The pre-trained model has the potential to improve the generalization ability or make it possible only to fine-tune on specific tasks without training from scratch.\n- In the experiments part, the proposed model can achieve one order of magnitude smaller errors comparing to the existing methods.", "weaknesses": "- The experiments for validating the proposed model are currently limited to 2-D cases (incompressible and compressiable Navier-stokes equations,  shallow-water equations, and a 2D DiffusionReaction equation). It is unclear whether the proposed model can achieve same level of performance and accuracy when applied to more realistic 3D physical simulations.\n- It seems that the current model mainly focus on predicting the solution one time step further. The video results (diffre, incompNS, mpp_swe) shown in the supplemental materials exhibits strong checkboard artifacts as the timesteps grow.\n- The proposed model can only handle simulations on structured mesh. However unstructured mesh (e.g., triangular, tetrahedra) is a more common choice in real world simulations. In addition, the multi-resolution i.e., mesh with adaptive resolution, which is also a common technique in large scale simulations, has not been taken into consideration.", "questions": "- It is mentioned in the appendix that the model is trained for 500 epochs for one task. How long does the training process take?\n- Are there any intuitive explanation for the checkboard artifacts as the timesteps grow shown in the video? Does it mean the proposed method is numerical unstable to some extent?\n- Does the proposed method have the potential to be applied to 3D simulations while keeping the efficiency and accuracy? (as the max memory usage of current model is ~60GB)", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698685017546}, {"id": "JHnjBGbMGo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2797/Reviewer_exE1"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The core of this paper is the proposal of a new transformer structure for pre-training on different neural operator datasets. Specifically, the paper employs several different PDE datasets from PDEBench for training, and the results can generalize to specific datasets in a zero-shot manner. Overall, the idea is novel, paving a new path for the application of neural operators in data-constrained scenarios. However, the learning of neural operators significantly differs from other data types like text and images, and whether a simple auto-regressive approach can be used for unified pre-training remains ambiguously addressed in the paper. Detailed concerns are noted in the drawbacks. All in all, I appreciate the idea presented in this paper, yet there is considerable room for improvement in problem definition, paper composition, experimental design, and comparisons. Thus I recommend the authors revise the paper and submit it to the next venue.", "review_text": "The core of this paper is the proposal of a new transformer structure for pre-training on different neural operator datasets. Specifically, the paper employs several different PDE datasets from PDEBench for training, and the results can generalize to specific datasets in a zero-shot manner. Overall, the idea is novel, paving a new path for the application of neural operators in data-constrained scenarios. However, the learning of neural operators significantly differs from other data types like text and images, and whether a simple auto-regressive approach can be used for unified pre-training remains ambiguously addressed in the paper. Detailed concerns are noted in the drawbacks. All in all, I appreciate the idea presented in this paper, yet there is considerable room for improvement in problem definition, paper composition, experimental design, and comparisons. Thus I recommend the authors revise the paper and submit it to the next venue.", "strengths": "1. The idea presented in this paper is relatively novel.  To the best of my knowledge, no published work has utilized an auto-regressive approach for pre-training on different types of PDEs to enhance performance on downstream tasks.\n2. The network structure proposed in the paper is efficient, and capable of handling datasets or PDE problems of varying sizes, resolutions, and channel counts within an acceptable level of complexity.\n3. The experimental results validate that even with differences in equation parameters or properties, pre-training can significantly save data, which is a valuable point. Additionally, the authors found that transformers pre-trained on video datasets are also beneficial for tasks of operator learning. This is an interesting fact.", "weaknesses": "1. Firstly, the paper forcibly combines different types of datasets for training without considering the potential conflicts in PDE solutions. For instance, one can construct two PDEs that have identical or minimally differing data within a certain frame count, yet due to the non-linearity of PDEs or inherent differences in the equations, they exhibit substantial differences in subsequent evolution. Unlike Ref [1] which unifies the form of PDEs, this paper's approach to mixed dataset training could lead to the model learning meaningless representations, especially in the presence of conflicting data. From the experiments and provided open-source code, it seems the paper selected equations with vastly different properties for pre-training, which doesn't address this challenge faced in PDE pre-training.\n\n2. Although the experimental section is logically well-structured, the descriptions of the experimental results and settings remain unclear. For example, PDEBench provides multiple datasets for CNS M1.0 and CNS M0.1, but the paper doesn't clearly state which dataset was used for training or testing. Besides, the source of the results for other baselines is not clarified, and the comparisons have too few baselines.\n\n3. I believe the paper's contribution of pre-training PDE representations is overstated. Upon careful examination of the appendices, I found that the paper nearly only utilized a few fluid dynamics datasets from PDEBench and a small diffusion-reaction equation dataset. Compared to the NLP or CV community, which employs almost all publicly available data on the internet for pre-training, the scale of pre-training in this paper is not large; it's more aptly termed as transfer learning.\n\nReferences\n1. Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior (https://arxiv.org/abs/2306.00258)", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The core of this paper is the proposal of a new transformer structure for pre-training on different neural operator datasets. Specifically, the paper employs several different PDE datasets from PDEBench for training, and the results can generalize to specific datasets in a zero-shot manner. Overall, the idea is novel, paving a new path for the application of neural operators in data-constrained scenarios. However, the learning of neural operators significantly differs from other data types like text and images, and whether a simple auto-regressive approach can be used for unified pre-training remains ambiguously addressed in the paper. Detailed concerns are noted in the drawbacks. All in all, I appreciate the idea presented in this paper, yet there is considerable room for improvement in problem definition, paper composition, experimental design, and comparisons. Thus I recommend the authors revise the paper and submit it to the next venue.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The idea presented in this paper is relatively novel.  To the best of my knowledge, no published work has utilized an auto-regressive approach for pre-training on different types of PDEs to enhance performance on downstream tasks.\n2. The network structure proposed in the paper is efficient, and capable of handling datasets or PDE problems of varying sizes, resolutions, and channel counts within an acceptable level of complexity.\n3. The experimental results validate that even with differences in equation parameters or properties, pre-training can significantly save data, which is a valuable point. Additionally, the authors found that transformers pre-trained on video datasets are also beneficial for tasks of operator learning. This is an interesting fact.", "weaknesses": "1. Firstly, the paper forcibly combines different types of datasets for training without considering the potential conflicts in PDE solutions. For instance, one can construct two PDEs that have identical or minimally differing data within a certain frame count, yet due to the non-linearity of PDEs or inherent differences in the equations, they exhibit substantial differences in subsequent evolution. Unlike Ref [1] which unifies the form of PDEs, this paper's approach to mixed dataset training could lead to the model learning meaningless representations, especially in the presence of conflicting data. From the experiments and provided open-source code, it seems the paper selected equations with vastly different properties for pre-training, which doesn't address this challenge faced in PDE pre-training.\n\n2. Although the experimental section is logically well-structured, the descriptions of the experimental results and settings remain unclear. For example, PDEBench provides multiple datasets for CNS M1.0 and CNS M0.1, but the paper doesn't clearly state which dataset was used for training or testing. Besides, the source of the results for other baselines is not clarified, and the comparisons have too few baselines.\n\n3. I believe the paper's contribution of pre-training PDE representations is overstated. Upon careful examination of the appendices, I found that the paper nearly only utilized a few fluid dynamics datasets from PDEBench and a small diffusion-reaction equation dataset. Compared to the NLP or CV community, which employs almost all publicly available data on the internet for pre-training, the scale of pre-training in this paper is not large; it's more aptly termed as transfer learning.\n\nReferences\n1. Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior (https://arxiv.org/abs/2306.00258)", "questions": "None", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698575788731}, {"id": "NWdvtdtbwX", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2797/Reviewer_iLVr"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a pretraining strategy for autoregressive modeling of physical systems. It proposes a model architecture and training regime to handle the heterogeneity of training data. Experiments on the PDEBench dataset demonstrate the ability of the method to model different systems with state of the art accuracy. The authors further investigate the added value of the pretrained model in low-data settings and parameter-estimation tasks.", "review_text": "This paper introduces a pretraining strategy for autoregressive modeling of physical systems. It proposes a model architecture and training regime to handle the heterogeneity of training data. Experiments on the PDEBench dataset demonstrate the ability of the method to model different systems with state of the art accuracy. The authors further investigate the added value of the pretrained model in low-data settings and parameter-estimation tasks.", "strengths": "Strength 1: The authors experimentally validate the hypothesis that a single model can learn on diverse fluid mechanics systems, which is a key step for the development of general “foundation” models for physics tasks. This line of experiments is well explained and rigorously reported. The use of the PDE-Bench dataset tasks and baseline models will ease all future work of the community to build and improve on the proposed approach. \n\nStrength 2: The architecture is novel and constitutes a first usage of transformer-based architectures for autoregressive modeling of physics systems. Given the exceptional scaling abilities of these models and their results in other scientific domains (chemical physics, biology), it is a promising direction for autoregressive tasks. Reporting the performance of the proposed architecture on PDEBench tasks, without multi-physics training, would already be valuable to the community. \n\nStrength 3 : The paper is easy to read and the questions it aims at answering are clearly formulated. Design choices and training regime are well motivated by the specificities of multi-systems learning, and clearly documented. This will ease the ramp-up of the community on pre-training tasks, and improve on the proposed architectures by proposing alternate design choices.\n\nStrength 4 : Figures and tables are clear, metrics and units are clearly reported and consistent with previous publications when taken from there", "weaknesses": "Weakness 1: The authors frame their question on the low-data regime (Section 5.2) as “Does MPP provide a fine-tuning advantage over existing spatiotemporal foundation models for new autoregressive prediction tasks?”. Their experimental results indeed support the claim that MMP is superior. However I believe that in order to develop useful foundation models, such models should be compared to the “best one can do” task-specific model in the low-data settings. Monitoring the progress on tasks that are too hard for task-specific baselines (like those implemented in the PDEBench paper) will provide rich insights to the community. The first line of experiments (Section 5.1) does not enable this, as all tasks are used at training time, including the test task. \nA possible experiment to add to section 5.2 would consist in training one of the PDEBench baseline models on the two low-data tasks\n\nWeakness 2: The model architecture is completely novel and more experimental validation should be done to maximize the value for the community. Specifically, to separate the added-value of the transformer architecture from the added value of the multi-systems training, it could have been interesting to train the MMP architecture on each of the PDEBench tasks individually, and include it in Table 1 as a new single-task baseline. It is also quite common to include ablation studies to validate intuitions or design choices, and the claim that positional encodings help to learn boundary conditions (Section 4.2, last paragraph) should be validated either by such experiment, or by a proof in the supplementary material. \n\nWeakness 3: The conclusion reminds the motivation behind multi-systems pretraining, but does not clearly recapitulate which questions have been answered and what is left to future studies. The two paragraphs are not well articulated and I feel would benefit from a rework. Ideally the reader should finish the paper with a clear view of the unanswered questions or limitations to be addressed.", "questions": "The paper takes first relevant steps towards training general models to model physical systems. I believe the authors have shown that their method of pretraining is functional, but more thorough experiments on the model architecture and the application to low-data tasks would be beneficial to the community and greatly increase the interest of the paper :\n\n1) Put the low-data results (Section 5.2) in perspective with task-specific baseline performances : are these tasks way too hard for the PDEBench baselines ? Ideally, report the performance of one task-specific baseline in Figure 5. \n\n2) Rework the conclusion to simply recapitulate the questions answered and which questions should be explored in future studies.\n\n3) In section 4.2 last paragraph, clarify whether the claim on boundary conditions is justified theoretically, validated by an ablation study, or both. Ideally, add relevant information in the supplementary material.\n\nMinor comments to improve readability :\n\na) Code clarity : Papers that propose pretrained models should be as easy as possible to finetune, so that the community can make quick progress towards more efficient pretraining approaches on challenging downstream tasks. I would recommend rewriting the train.py file to remove a lot of boilerplate code and make it easier to reuse and tweak. \n\nb) Clarity of Figure 2 and its legend can be improved: ReVIN abbreviation for the normalization layer is not defined. “physics metadata” is not defined neither in text nor in figure legend. \n\nc) Table 1 : abbreviations for task names are not defined in the manuscript and make it hard to find the correspondence in the initial PDEBench paper. Please define these abbreviations in the table legend.  \n\nd) There is a typo in section 5.1 (word “all” written twice) : our models (denoted by MPP-AViT-*) must handle all all systems and regimes without finetuning", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a pretraining strategy for autoregressive modeling of physical systems. It proposes a model architecture and training regime to handle the heterogeneity of training data. Experiments on the PDEBench dataset demonstrate the ability of the method to model different systems with state of the art accuracy. The authors further investigate the added value of the pretrained model in low-data settings and parameter-estimation tasks.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "Strength 1: The authors experimentally validate the hypothesis that a single model can learn on diverse fluid mechanics systems, which is a key step for the development of general “foundation” models for physics tasks. This line of experiments is well explained and rigorously reported. The use of the PDE-Bench dataset tasks and baseline models will ease all future work of the community to build and improve on the proposed approach. \n\nStrength 2: The architecture is novel and constitutes a first usage of transformer-based architectures for autoregressive modeling of physics systems. Given the exceptional scaling abilities of these models and their results in other scientific domains (chemical physics, biology), it is a promising direction for autoregressive tasks. Reporting the performance of the proposed architecture on PDEBench tasks, without multi-physics training, would already be valuable to the community. \n\nStrength 3 : The paper is easy to read and the questions it aims at answering are clearly formulated. Design choices and training regime are well motivated by the specificities of multi-systems learning, and clearly documented. This will ease the ramp-up of the community on pre-training tasks, and improve on the proposed architectures by proposing alternate design choices.\n\nStrength 4 : Figures and tables are clear, metrics and units are clearly reported and consistent with previous publications when taken from there", "weaknesses": "Weakness 1: The authors frame their question on the low-data regime (Section 5.2) as “Does MPP provide a fine-tuning advantage over existing spatiotemporal foundation models for new autoregressive prediction tasks?”. Their experimental results indeed support the claim that MMP is superior. However I believe that in order to develop useful foundation models, such models should be compared to the “best one can do” task-specific model in the low-data settings. Monitoring the progress on tasks that are too hard for task-specific baselines (like those implemented in the PDEBench paper) will provide rich insights to the community. The first line of experiments (Section 5.1) does not enable this, as all tasks are used at training time, including the test task. \nA possible experiment to add to section 5.2 would consist in training one of the PDEBench baseline models on the two low-data tasks\n\nWeakness 2: The model architecture is completely novel and more experimental validation should be done to maximize the value for the community. Specifically, to separate the added-value of the transformer architecture from the added value of the multi-systems training, it could have been interesting to train the MMP architecture on each of the PDEBench tasks individually, and include it in Table 1 as a new single-task baseline. It is also quite common to include ablation studies to validate intuitions or design choices, and the claim that positional encodings help to learn boundary conditions (Section 4.2, last paragraph) should be validated either by such experiment, or by a proof in the supplementary material. \n\nWeakness 3: The conclusion reminds the motivation behind multi-systems pretraining, but does not clearly recapitulate which questions have been answered and what is left to future studies. The two paragraphs are not well articulated and I feel would benefit from a rework. Ideally the reader should finish the paper with a clear view of the unanswered questions or limitations to be addressed.", "questions": "The paper takes first relevant steps towards training general models to model physical systems. I believe the authors have shown that their method of pretraining is functional, but more thorough experiments on the model architecture and the application to low-data tasks would be beneficial to the community and greatly increase the interest of the paper :\n\n1) Put the low-data results (Section 5.2) in perspective with task-specific baseline performances : are these tasks way too hard for the PDEBench baselines ? Ideally, report the performance of one task-specific baseline in Figure 5. \n\n2) Rework the conclusion to simply recapitulate the questions answered and which questions should be explored in future studies.\n\n3) In section 4.2 last paragraph, clarify whether the claim on boundary conditions is justified theoretically, validated by an ablation study, or both. Ideally, add relevant information in the supplementary material.\n\nMinor comments to improve readability :\n\na) Code clarity : Papers that propose pretrained models should be as easy as possible to finetune, so that the community can make quick progress towards more efficient pretraining approaches on challenging downstream tasks. I would recommend rewriting the train.py file to remove a lot of boilerplate code and make it easier to reuse and tweak. \n\nb) Clarity of Figure 2 and its legend can be improved: ReVIN abbreviation for the normalization layer is not defined. “physics metadata” is not defined neither in text nor in figure legend. \n\nc) Table 1 : abbreviations for task names are not defined in the manuscript and make it hard to find the correspondence in the initial PDEBench paper. Please define these abbreviations in the table legend.  \n\nd) There is a typo in section 5.1 (word “all” written twice) : our models (denoted by MPP-AViT-*) must handle all all systems and regimes without finetuning", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697548887644}], "openreview_url": "https://openreview.net/forum?id=fH9eqpCcR3", "arxiv_id": "2310.02994", "paper_pdf": "papers/fH9eqpCcR3.pdf", "paper_pdf_sha256": "01f4ed804cd79222738933baacfeabf298df378d1de69f75c675cbcdc2dbae83", "paper_pdf_bytes": 3414080, "paper_pdf_source": "openreview", "code_url": "https://github.com/PolymathicAI/multiple_physics_pretraining", "code_repository": "PolymathicAI/multiple_physics_pretraining", "code_commit": "e751fc25ae5274c3ddc869bf73afd6fcd4163cb5", "code_archive": "repos/fH9eqpCcR3.zip", "code_archive_sha256": "8e05c07948bb5c1375c139da4fe39e90043a16db99ea08345f870d315f87d7e6", "code_archive_bytes": 458749, "code_file_count": 13, "code_extensions": {".py": 11, ".sh": 2}, "github_disk_usage_kb": 495, "github_languages": {"Python": 84340, "Shell": 1229}, "github_archived": false, "github_pushed_at": "2024-12-06T15:43:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multiple-physics-pretraining-for-physical"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3leZITnUE9r", "year": 2023, "status": "rejected", "title": "An Empirical Study of Metrics to Measure Representational Harms in Pre-Trained Language Models", "authors": ["Saghar Hosseini", "Ahmed Hassan Awadallah", "Hamid Palangi"], "authorids": ["~Saghar_Hosseini1", "~Ahmed_Hassan_Awadallah1", "~Hamid_Palangi1"], "authors_source": "OpenReview API", "abstract": "Large-scale Pre-Trained Language Models (PTLMs) capture knowledge from massive human-written data which contains latent societal biases and toxic contents. In this paper, we leverage the primary task of PTLMs, i.e. language modeling, and propose a new metric to quantify manifested implicit representational harms in PTLMs towards 13 marginalized demographics. Using this metric, we conducted an empirical analysis of 24 widely used PTLMs. Our analysis provides insights into the correlation between the proposed metric in this work and other related fairness metrics. We observe that our metric correlates with the majority of gender-specific fairness metrics in the literature. Through extensive experiments, we explore the connections between PTLMs architectures and representational harms across two dimensions: depth and width of the networks. We found that prioritizing depth over width, mitigates representational harms in some PTLMs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "8Zl-l8VrIaE", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper245/Reviewer_K8hw"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper provides an empirical study of representational harms in pre-trained language models.\nThe authors consider safety scores derived from a Mann-Whitney U-test, and compute such safety scores across a range of different models and marginalized demographics, finding that PTLMs have a tendency to show representational harms towards some marginalized demographics more than others, with some of these groups having not been studied extensively before.", "review_text": "While this study tackles an important problem for applications of language models, and such work should be published in impactful venues such as ICLR, I do not think the scope and novelty of this study meets the criterion for publication at this conference as it stands now.", "strengths": "* This paper tackles the important issue of understanding biases and representational harms of language models, which are prevalent and permeate society through their varied applications.\n* The draft has several typos and missing punctuation which should be fixed. Several sentences throughout the text are also poorly phrased/grammatically incorrect and hard to understand.\n* On p.3, the authors state that issues have been found in the datasets used in recent related work, could the authors expand on what those issues are. The footnote included links to another footnote (2) which does not appear in the article.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper provides an empirical study of representational harms in pre-trained language models.\nThe authors consider safety scores derived from a Mann-Whitney U-test, and compute such safety scores across a range of different models and marginalized demographics, finding that PTLMs have a tendency to show representational harms towards some marginalized demographics more than others, with some of these groups having not been studied extensively before.", "strength_and_weaknesses": "* This paper tackles the important issue of understanding biases and representational harms of language models, which are prevalent and permeate society through their varied applications.\n* The draft has several typos and missing punctuation which should be fixed. Several sentences throughout the text are also poorly phrased/grammatically incorrect and hard to understand.\n* On p.3, the authors state that issues have been found in the datasets used in recent related work, could the authors expand on what those issues are. The footnote included links to another footnote (2) which does not appear in the article.", "clarity,_quality,_novelty_and_reproducibility": "* The paper is not always very clear, there are several typos, the term \"safety scores\" seems to be used interchangeably with \"fairness score\" with no explicit definition of fairness or safety.\n* This work is a purely empirical study of existing language models, and introduces little novelty in terms of definitions or measures.", "summary_of_the_review": "While this study tackles an important problem for applications of language models, and such work should be published in impactful venues such as ICLR, I do not think the scope and novelty of this study meets the criterion for publication at this conference as it stands now.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667158812626}, {"id": "Qnb37EM9RdU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper245/Reviewer_UeDp"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper addresses the problem of representational harms in pretrained models. The paper proposes a metric, safety score, to measure the harms. Then, the paper shows a study of this metric on 13 marginalized demographics using 24 pretrained models, and discuss the findings. ", "review_text": "This paper tackles the problem of representational bias in pretrained models.  The proposed metric seems reasonable to me, and the study findings are interesting. ", "strengths": "Strengths\n- The paper is easy to follow\n- Well motivated study - the overall topic\n- Studying this topic (representational harms in pretrained LMs) is important in the field.\n- The proposed metric seems reasonable\n- The further analyses and findings are interesting. \n\nWeaknesses\n- not clear why a new metric is needed\n- I didn't understand \"prioritizing depth over width\" when reading it for the first time. \n- the citation formats should be revised for readability.\n- typos and grammatical errors.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper addresses the problem of representational harms in pretrained models. The paper proposes a metric, safety score, to measure the harms. Then, the paper shows a study of this metric on 13 marginalized demographics using 24 pretrained models, and discuss the findings. ", "strength_and_weaknesses": "Strengths\n- The paper is easy to follow\n- Well motivated study - the overall topic\n- Studying this topic (representational harms in pretrained LMs) is important in the field.\n- The proposed metric seems reasonable\n- The further analyses and findings are interesting. \n\nWeaknesses\n- not clear why a new metric is needed\n- I didn't understand \"prioritizing depth over width\" when reading it for the first time. \n- the citation formats should be revised for readability.\n- typos and grammatical errors.", "clarity,_quality,_novelty_and_reproducibility": "The proposed metric seems novel, and it was clearly explained. ", "summary_of_the_review": "This paper tackles the problem of representational bias in pretrained models.  The proposed metric seems reasonable to me, and the study findings are interesting. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666760239505}, {"id": "oMD0_QfJ4C1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper245/Reviewer_82Na"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new metric for measuring fairness\tof language\nmodels using text toxicity level and perplexity. It shows that the new\nmetric correlates well with other gender-specific metrics in the\nliterature. Using this new metric, a comprehensive study of 24 models\nis performed, along an analysis how the depth/width of the models\ninfluences the representational harms.", "review_text": "The paper introduces a new (intrinsic) metric for assessing fairness\nof language models using text toxicity level and perplexity. The\nmetric is not well motivated and it is not clear why extrinsic metrics\ndo not suffice.", "strengths": "Strengths:\n- a comprehensive study of the fairness\tof several model architectures (e.g., BERT, AlBERT, ELECTRA) of different sizes (e.g., base, large)\n- a study of how the depth/width of the network influences their fairness\n\nWeaknesses:\n- the need for a new metric is not well motivated\n- the advantages/disadvantages of the new metric are not discussed\n- intrinsic metrics for bias have been shown to be problematic; I'm not sure how to interpret the correlation study with one such metric", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a new metric for measuring fairness\tof language\nmodels using text toxicity level and perplexity. It shows that the new\nmetric correlates well with other gender-specific metrics in the\nliterature. Using this new metric, a comprehensive study of 24 models\nis performed, along an analysis how the depth/width of the models\ninfluences the representational harms.", "strength_and_weaknesses": "Strengths:\n- a comprehensive study of the fairness\tof several model architectures (e.g., BERT, AlBERT, ELECTRA) of different sizes (e.g., base, large)\n- a study of how the depth/width of the network influences their fairness\n\nWeaknesses:\n- the need for a new metric is not well motivated\n- the advantages/disadvantages of the new metric are not discussed\n- intrinsic metrics for bias have been shown to be problematic; I'm not sure how to interpret the correlation study with one such metric", "clarity,_quality,_novelty_and_reproducibility": "I think\tthe paper would\timprove\tif the new metric was better\nmotivated. What gaps is it trying to fill in? What are the\nadvantages/disadvantages? If I understand correctly, I think of this\nmetric as an intrinsic metric. Several works are showing that\nintrinsic metrics (see below) are not well correlated with extrinsic\nmetrics for bias/fairness. For the metric introduced, since you\nrequire knowing the toxicity for the text, why not look at extrinsic\nmetrics (see Fairness definitions explained for lots of metrics or the\nanalysis using equalized odds in Your Fairness may vary).\n\nSome papers that I think could be cited (and why they are relevant in this context):\n* Fairness Definitions Explained - series of fairness metrics\n* Your Fairness May Vary: Pretrained Language Model Fairness in Toxic Text Classification - analysis of many LMs wrt fairness and model size/training size/random seed (in the context of toxic text prediction) using\tequalized odds as fairness metric, which\tis an extrinsic\tmetric\n* Intrinsic Bias Metrics Do Not Correlate with Application Bias - study of correlation of intrinsic vs extrinsic metrics\n\nSome typos: e.g., \"large-sacle\"\n", "summary_of_the_review": "The paper introduces a new (intrinsic) metric for assessing fairness\nof language models using text toxicity level and perplexity. The\nmetric is not well motivated and it is not clear why extrinsic metrics\ndo not suffice.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666660308020}, {"id": "ezt1Rk7vY7Z", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper245/Reviewer_hcJz"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents a new metric to measure the fairness of language models on a toxicity labeled dataset. Authors exploit the fact that a fair model should yield higher perplexity scores for toxic sentences. The proposed metric is bounded between [0, 1] and utilizes Mann-Whitney U statistical test to quantify the tendency of a pre-trained model towards toxic generations for a protected group. Authors use an open-source toxicity dataset to evaluate their proposed metric for a wide range of publicly available language models. ", "review_text": "While the proposed fairness metric is intuitive, as pointed in the weakness section, it's not clear how the proposed metric align with the human notion of fairness. Further, some of the analysis presented in the paper is too shallow as authors have not conducted ablation studies. Overall, I think that this paper requires further experimentation.   ", "strengths": "Strengths: \n- Proposed fairness metric is intuitive and is easy to compute. \n- Authors study a wide range of pre-trained models. \n\nWeaknesses: \n- Authors have not conducted any human evaluation to verify the correctness of the proposed metric. So it’s not clear how much the proposed metric aligns with the human notion of fairness.\n- Some of the design choices made in the paper are not clear. In particular\n  - Why do authors divide peplexity by corresponding toxicity score?  What benefits (if any) does it provide over simply using perplexity scores? Authors have not provided any empirical or theoretical justification for it. \n  - Authors use a subsample of the ToxiGen dataset which is generated using the GPT-3 model. We know that the GPT-3 model has its own set of biases towards different protected attributes. It’s not clear how such biases impact the findings in this study.  \n \n- Paper does not dive deep into some of the surprising findings and makes shallow arguments. In particular:  \n  - GPT-2 large has better safety score than GPT2-medium. It's not clear why ? \n   - Roberta seems to have a very high safety score and authors hypothesize that it's because the pre-training corpus for RoBERTa contains stories, and news. We know that news coverage of certain protected attribute is often negative. For example, many news article associated with muslim community are not positive. So it's not clear why stories and news will improve RoBERTa's safety score.   \n- Some of the claims might be too specific to the dataset studied in the paper. For example, authors state that language models in general are less likely to embed harmful content for Asian, African-Americal, Chinese and Jewish people compared to other demographics. This finding might not be correct because the presented results are heavily dependent on a single dataset which is generated by the GPT-3 model. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper presents a new metric to measure the fairness of language models on a toxicity labeled dataset. Authors exploit the fact that a fair model should yield higher perplexity scores for toxic sentences. The proposed metric is bounded between [0, 1] and utilizes Mann-Whitney U statistical test to quantify the tendency of a pre-trained model towards toxic generations for a protected group. Authors use an open-source toxicity dataset to evaluate their proposed metric for a wide range of publicly available language models. ", "strength_and_weaknesses": "Strengths: \n- Proposed fairness metric is intuitive and is easy to compute. \n- Authors study a wide range of pre-trained models. \n\nWeaknesses: \n- Authors have not conducted any human evaluation to verify the correctness of the proposed metric. So it’s not clear how much the proposed metric aligns with the human notion of fairness.\n- Some of the design choices made in the paper are not clear. In particular\n  - Why do authors divide peplexity by corresponding toxicity score?  What benefits (if any) does it provide over simply using perplexity scores? Authors have not provided any empirical or theoretical justification for it. \n  - Authors use a subsample of the ToxiGen dataset which is generated using the GPT-3 model. We know that the GPT-3 model has its own set of biases towards different protected attributes. It’s not clear how such biases impact the findings in this study.  \n \n- Paper does not dive deep into some of the surprising findings and makes shallow arguments. In particular:  \n  - GPT-2 large has better safety score than GPT2-medium. It's not clear why ? \n   - Roberta seems to have a very high safety score and authors hypothesize that it's because the pre-training corpus for RoBERTa contains stories, and news. We know that news coverage of certain protected attribute is often negative. For example, many news article associated with muslim community are not positive. So it's not clear why stories and news will improve RoBERTa's safety score.   \n- Some of the claims might be too specific to the dataset studied in the paper. For example, authors state that language models in general are less likely to embed harmful content for Asian, African-Americal, Chinese and Jewish people compared to other demographics. This finding might not be correct because the presented results are heavily dependent on a single dataset which is generated by the GPT-3 model. ", "clarity,_quality,_novelty_and_reproducibility": "Clarity, Quality:\nPaper cites the relevant literature. However, as pointed out in the weaknesses section, some of the design choices made in the paper are not clear. Further, authors use vague language in the paper. For example, authors say “our metric is quantifying a different notion of fairness issues compare to the existing metrics”. However, they don’t define the “fairness notion” that they are trying to measure. \n\nNovelty:\nThe proposed metric to study fairness is novel. \n\nReproducibility:\nAuthors have used a subset of an open source dataset and have provided exact details to select the subset used in the paper. All of the models studied in the paper are open sourced so one should be able to easily reproduce the experiments presented in the paper. \n", "summary_of_the_review": "While the proposed fairness metric is intuitive, as pointed in the weakness section, it's not clear how the proposed metric align with the human notion of fairness. Further, some of the analysis presented in the paper is too shallow as authors have not conducted ablation studies. Overall, I think that this paper requires further experimentation.   ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666581739713}], "openreview_url": "https://openreview.net/forum?id=3leZITnUE9r", "arxiv_id": "2301.09211", "paper_pdf": "papers/3leZITnUE9r.pdf", "paper_pdf_sha256": "111fd283ddaa801cffd2f85b56deae23dff3c7e54620549ca4b4184b1c9f4663", "paper_pdf_bytes": 453389, "paper_pdf_source": "openreview", "code_url": "https://github.com/microsoft/SafeNLP", "code_repository": "microsoft/SafeNLP", "code_commit": "e351b6cb892284cf843b88e363b1a6dbaded584e", "code_archive": "repos/3leZITnUE9r.zip", "code_archive_sha256": "caa7a49a41212bc290f462894c1d69d61a9743bbae58bea2f9ac055a6a0bd397", "code_archive_bytes": 1698691, "code_file_count": 3, "code_extensions": {".py": 2, ".sh": 1}, "github_disk_usage_kb": 1254, "github_languages": {"Python": 7551, "Shell": 517}, "github_archived": true, "github_pushed_at": "2023-10-18T20:57:47Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/an-empirical-study-of-metrics-to-measure"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "FD8xldQIgdq", "year": 2022, "status": "rejected", "title": "Robust Models Are More Interpretable Because Attributions Look Normal", "authors": ["Zifan Wang", "Matt Fredrikson", "Anupam Datta"], "authorids": ["~Zifan_Wang1", "~Matt_Fredrikson1", "~Anupam_Datta1"], "authors_source": "OpenReview API", "abstract": "Recent work has found that adversarially-robust deep networks used for image classification are more interpretable: their feature attributions tend to be sharper, and are more concentrated on the objects associated with the image's ground-truth class. We show that smooth decision boundaries play an important role in this enhanced interpretability, as the model's input gradients around data points will more closely align with boundaries' normal vectors when they are smooth. Thus, because robust models have smoother boundaries, the results of gradient-based attribution methods will capture more accurate information about nearby decision boundaries. This understanding of robust interpretability leads to our second contribution: \\emph{boundary attributions}, which aggregate information about the normal vectors of local decision boundaries to explain a classification outcome. We show that by leveraging the key factors underpinning robust interpretability, boundary attributions produce sharper, more concentrated visual explanations---even on non-robust models.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rHW42cPDhuI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1892/Reviewer_oVRP"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The idea is that as one smoothens the decision boundary of a piecewise linear function $f$ (e.g., from ReLU-Net) its saliency map ($g$) obtained by $g=\\frac{df}{dx}$ gets closer ($||g - n||_2$) to the normal of the closest boundary hyperplane ($n$). The authors then propose two variants of explanation techniques based on the nearest decision boundary hyperplane and try it on explaining trained image deep classifiers. The results seem to corroborate as for a better alignment with the nearest boundary hyperplane's normal. Also, the proposed methods achieve better explanations as measured by locality and overlap with ground-truth bounding boxes.", "review_text": "**Strengths**\n+ the general idea of alignment with the nearest decision hyperplane normal and the specific modification of IG seem quite novel and plausible.\n+ In the experiments, BIG achieves significantly better results using various explanation metrics.\n\n**Questions to authors**\n- Smoothening the learnt function, at some point, should start losing the discrmination ability of the learnt function. Has the authors pushed enough to find some indication of this trade-off?\n- From theorem 1, I can understand why a smoother learned functions can give rise to a more faithful saliency-based explanation but I cannot see how it advocates smoothgrad as explanation. Wouldn't smoothgrad be faithful to a very-likely different function than the actual learned function and thus not necessarily faithful to the true learned function? \n- The text before definition 6, argues for BIG based on the existence of multiple boundary segments near a point and proposed definition 6 that integrates over the segment connecting a point x to its nearest adversarial $x'$. However, shouldn't the nearest decision boundary segment for all points along the line segment $x\\rightarrow x'$ remain the same? The integral is taken over standard saliency g which of course can change linear regions but the rationale (of wanting to find different decision boundary hyperplanes) does not seem to hold for the proposal.\n- The previous question could be simply rectified if the meaning of \"boundary segments\" is the linear regions' boundary segments as opposed to the decision boundary segments but then I think \"boundary\" has been used as \"decision boundary\" at occasions before this definition, *e.g.*, in def 5. Am I mistaken? If not, the text needs a rewrite to distinguish between \"regions boundary segments\" and \"decision boundary segments\".\n- Due to the approximation using an ensemble of adversarial example methods, we should expect that the found segment is very likely not the closest decision boundary segment (since we know from many works that the density of linear regions are extremely high in the input space). In light of that, how reliable are the observations in the experiments section? Especially, with regards to the deviation from the normal vector (Figure 3.a). \n- Following up on the previous question, could the fact that BSM does not show improvement on standard models be due to this approximation?\n\n**Minor points**\n- on many occasions, when referring to boundary facets of a polytope, better to use hyperplane as opposed to segment to avoid confusion with line segments that are used as linear path.\n- In definition 3, $f(\\alpha + \\epsilon) \\rightarrow f(x + \\epsilon)$\n- In Theorem 1, $\\forall x'' \\in B(...).$ better to replace $. \\rightarrow ,$ (although a minor point, it makes reading the statement challenging in the first glance. )\n- In Theoretm 1, it might be better to use $O(\\frac{1}{\\sigma c})$? \n- two different notations are used for definition (:= or else)\n- better to refer to Figure 3.a and 3.b as tables\n- page 7: \"a smaller difference between the difference between\nattributions\"\n- page 7: \"instead evaluates computes\"\n- page 8: \"It is naturally to treat BIG frees users from the baseline selection\"\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The idea is that as one smoothens the decision boundary of a piecewise linear function $f$ (e.g., from ReLU-Net) its saliency map ($g$) obtained by $g=\\frac{df}{dx}$ gets closer ($||g - n||_2$) to the normal of the closest boundary hyperplane ($n$). The authors then propose two variants of explanation techniques based on the nearest decision boundary hyperplane and try it on explaining trained image deep classifiers. The results seem to corroborate as for a better alignment with the nearest boundary hyperplane's normal. Also, the proposed methods achieve better explanations as measured by locality and overlap with ground-truth bounding boxes.", "main_review": "**Strengths**\n+ the general idea of alignment with the nearest decision hyperplane normal and the specific modification of IG seem quite novel and plausible.\n+ In the experiments, BIG achieves significantly better results using various explanation metrics.\n\n**Questions to authors**\n- Smoothening the learnt function, at some point, should start losing the discrmination ability of the learnt function. Has the authors pushed enough to find some indication of this trade-off?\n- From theorem 1, I can understand why a smoother learned functions can give rise to a more faithful saliency-based explanation but I cannot see how it advocates smoothgrad as explanation. Wouldn't smoothgrad be faithful to a very-likely different function than the actual learned function and thus not necessarily faithful to the true learned function? \n- The text before definition 6, argues for BIG based on the existence of multiple boundary segments near a point and proposed definition 6 that integrates over the segment connecting a point x to its nearest adversarial $x'$. However, shouldn't the nearest decision boundary segment for all points along the line segment $x\\rightarrow x'$ remain the same? The integral is taken over standard saliency g which of course can change linear regions but the rationale (of wanting to find different decision boundary hyperplanes) does not seem to hold for the proposal.\n- The previous question could be simply rectified if the meaning of \"boundary segments\" is the linear regions' boundary segments as opposed to the decision boundary segments but then I think \"boundary\" has been used as \"decision boundary\" at occasions before this definition, *e.g.*, in def 5. Am I mistaken? If not, the text needs a rewrite to distinguish between \"regions boundary segments\" and \"decision boundary segments\".\n- Due to the approximation using an ensemble of adversarial example methods, we should expect that the found segment is very likely not the closest decision boundary segment (since we know from many works that the density of linear regions are extremely high in the input space). In light of that, how reliable are the observations in the experiments section? Especially, with regards to the deviation from the normal vector (Figure 3.a). \n- Following up on the previous question, could the fact that BSM does not show improvement on standard models be due to this approximation?\n\n**Minor points**\n- on many occasions, when referring to boundary facets of a polytope, better to use hyperplane as opposed to segment to avoid confusion with line segments that are used as linear path.\n- In definition 3, $f(\\alpha + \\epsilon) \\rightarrow f(x + \\epsilon)$\n- In Theorem 1, $\\forall x'' \\in B(...).$ better to replace $. \\rightarrow ,$ (although a minor point, it makes reading the statement challenging in the first glance. )\n- In Theoretm 1, it might be better to use $O(\\frac{1}{\\sigma c})$? \n- two different notations are used for definition (:= or else)\n- better to refer to Figure 3.a and 3.b as tables\n- page 7: \"a smaller difference between the difference between\nattributions\"\n- page 7: \"instead evaluates computes\"\n- page 8: \"It is naturally to treat BIG frees users from the baseline selection\"\n\n", "summary_of_the_review": "The paper has an interesting and original idea which brings consistent improvement to the established explanations techniques such as gradient-based saliency maps and integrated gradients. However, there are some questions that makes me keep my rating only at borderline accept.\n\n---\n**Post-Rebuttal Comments**\n\n*General rationale for the updated score*: after some more thoughts and the discussions during the rebuttal phase, the reviewer remains unconvinced about the claim of the normality of SM to (the extension of) a segment in the decision boundary. In this regard there are two significant concerns: (i) there are explicit statements about this both in the revised paper (see discussion with the authors for some instances) as well as the authors arguments during the discussion, and (ii) the motivation for the proposed method BIG is based on this claim. Furthermore, the paper cites other papers that as far as the reviewer understands do not (explicitly) discuss this claim. Therefore, I do not think I can vouch for accepting the paper claiming (and building on) some formal statements that I cannot personally verify. Consequently, I reduce my rating from 6 to 3. Since there might be a simple point that I am missing here which would prove the claim, I reduce my confidence as well from 4 to 3.\n\n*Summary of the technical discussion*:  The authors, at various points, implicitly suggest or explicitly claim that SM which is the gradient of the network's function w.r.t. the input ($\\frac{df}{dx}$) is perpendicular to (the extension of) a segment in the decision boundary. This is then used to motivated a variant of IG, called BIG which integrates SM over a line path from the sample to the nearest adversarial example. For the reviewer it is possible to see: (i) how $\\frac{df}{dx}$ for a linear binary classifier will always be orthogonal to the decision boundary since the decision boundary is by definition a hyperplane with the SM as its normal (as described in section 3.1), (ii) how $\\frac{d(f_i-f_j)}{dx}$ is orthogonal to the surface $f_i-f_j=0$. However, it is unclear to the reviewer how $\\frac{df}{dx}$ can be guaranteed to be prependicular to the decision boundary of a general function $f: \\mathbb{R}^d\\rightarrow\\mathbb{R}^K$ with K being the number of classes. In fact, I believe for the simplest case of linear binary classification, as soon as we (redundantly) model each class with a separate linear model (to become analogue to the multiclass setup), the gradient of each of the linear functions i.e., $\\frac{df_1}{dx}$ and $\\frac{df_2}{dx}$ will no more be orthogonal to the decision boundary.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635870026205}, {"id": "lg5vuaqzb9N", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1892/Reviewer_NMCL"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper focuses on the intersection of gradient attribution and adversarial robustness. First, it analyzes the weaknesses of vanilla gradients: the gradient does not have to point towards the decision boundary of an n-layer ReLU network. Then the paper provides some insights into the smoothing of one-layer ReLU networks (Theorem 1). Finally, a boundary-based saliency map and an extension of integrated gradients are proposed and evaluated in terms of boundary alignment and object localization. \n", "review_text": "The paper has an interesting topic: adding theoretical insights to explainability methods. The paper does especially well on providing a good intuition about the relationships of normals, polytopes and decision boundary (content of 3.1 and first part of 3.2). I also found the paper overall well written (some minor typos and duplicates are listed below). The paper's story of first analyzing the limitations of gradients, fixing the errors, and then evaluating the methods is also good. I address my concerns about the generality and rigor of theorem 1, the evaluation, and the limitations below.\n\n### Proof of theorem 1\n\nTheorem 1 contains an $\\lessapprox$ sign. After checking the appendix, it turns out that the proof is only correct for the case that\n\n> (Dombrowski et al., 2019) points out that the random distribution $p_β (\\epsilon_{i}) =\\frac{β}{(exp (β \\epsilon_{i} / 2)+exp ( β \\epsilon_{i} / 2))^2}\n> $closely resembles a normal distribution with a standard deviation $\\sigma = \\sqrt{\\log 2 \\frac{\\sqrt{2 \\pi}}{\\beta}}$.\n\nHowever, under which conditions does it resemble a normal distribution? (Dombrowski et al., 2019) only made this comment to explain a possible connection to SmoothGrad (see page 8 in Dombrowski et al., 2019). No concrete conditions are given on when or how close the distributions matches. I did not even found how  $\\sigma$  was derived in (Dombrowski et al., 2019) (if you know where, please point me to it). I did a small experiment myself and plotted the distributions. For each plot, the corresponding $\\beta$ is given on top and the normal distribution has $\\sigma = \\sqrt{\\log 2 \\frac{\\sqrt{2 \\pi}}{\\beta}}$ (for the notebook with the code see this [link](https://f002.backblazeb2.com/file/nnnnnnnn/iclr2022/Robust+Models+Are+More+Interpretable+Because+Attributions+Look+Normal.ipynb)).\n\n[[Plots for different $\\beta$s]](https://f002.backblazeb2.com/file/nnnnnnnn/iclr2022/beta_plots.png)\n\n\nAs you can see, it is only close for $\\beta \\approx 1$. Two solutions exist: either provide a theorem with $\\leq$ or give a rigorous discussion on the cases where only $\\approx$  or even > holds. \n\nThe other limitation of Theorem 1 is that it only holds for one-layer ReLU networks. I would find a short discussion helpful why it does not hold for n-layer ReLU networks. In addition, it should be emphasized throughout the paper that Theorem 1 is only for one-layer networks. For example, in the last paragraph of the introduction:\n\n> We present an analysis that sheds light on the previously-observed phenomeon of robust interpretability, showing that alignment between the normal vectors of decision boundaries and models’ gradients is a key ingredient (Proposition 1, Theorem 1)\n\nPlease make clear in that sentence and others that Theorem 1 only addresses one-layer networks. \n\nAt the end of section 3.2, Figure 10 is referenced as empirical validation of Theorem 1, but I do not understand the figure and caption:\n\n> distances in logarithm between SG and BSG against different standard deviations σ of the Gaussian noise. Results are computed on ResNet50. Notice the first column corresponds to σ = 0.\n\nPlease, clarify what you want to evaluate with this figure, e.g. the first column says $\\sigma=0.15$.\n\n### Evaluation\n\nI think the evaluation of the normality to the decision boundary can be improved. In Figure 3, pairs of gradient attribution method and the corresponding boundary attributions (e.g. IG vs. BIG) are compared to evaluate how normal the attributions are. However, why not measure the normality in the feature space $z(x)$ directly.  $z(x)$ is defined such that $f_i(x) = w_i^T z(x)$. We know that $w_i$ must be normal to the decision boundary, as shown in Figure 2a. The corresponding change in z-space of an attribution $g(x)$ would be $\\Delta z = z(x) - z(x + \\alpha g(x))$. Now, we can measure the similarity of the normal $w_i$ and the different attributions: just compute $\\cos(\\Delta z,  w_i)$ for all the different attributions. This evaluation would relate the estimated directions in $x$-space to the ground-truth normals in $z$-space. The current evaluation of attributions methods against their boundary equivalent cannot provide such a ground-truth reference. \n\nThe evaluation using the ground-truth bounding boxes is a good proxy task and seems to be executed correctly. It might make sense to only use images where the bounding box covers less than 50% of the image, as done in (Schulz et al., 2020). The attribution method in (Schulz et al., 2020) might also be an interesting candidate for the evaluation as it was also able to outperform int.grad. and smooth grad. I would also suggest focusing on one or two metrics for the bounding box task instead of four.\n\nI would also encourage the authors to include the sanity check for weight reinitialization (Adebayo et al., 2018). It is easy to implement and should be passed by any new attribution method. \n\n### Limitations\n\nWhile I do not think that the paper requires a human-subject evaluation, its lack should be mentioned in the limitation section. Also the saliency maps look more concentrated, would humans actually profit from it? Even if there is a significant difference, would you expect a large effect size? Please also list that theorem 1 is only for one-layered networks in the limitations. Limitation 2 (not applicable to perturbation attributions) arises from focus of the paper and I think there is not need to mention it. \n\n### Minor Comemnts:\n\n* \"In fact, the fact\" (page 4)\n* smaller difference between the difference between (page 7)\n* Thefore (page 7)\n* Lost clause: \"It is naturally to treat \" (page 8)\n* It should be Table 3 and not Figure 3\n* IamgeNet (page 6)\n* I think it should be \"The RHS of the above equation is Smoothed Gradient\" (page 15)\n\n### References:\n\n(Schulz et al., 2020) https://openreview.net/forum?id=S1xWh1rYwB\n\n(Adebayo et al., 2018) https://arxiv.org/abs/1810.03292\n\n## After Rebuttal Update\n\nThe authors were able to rectify their proof and also provided details to my other questions. While the initial submission was a clear reject, the rebuttal was well done. I agree with the concerns of the others reviewers about novelty. Overall, I increased my rating to marginal above acceptance.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper focuses on the intersection of gradient attribution and adversarial robustness. First, it analyzes the weaknesses of vanilla gradients: the gradient does not have to point towards the decision boundary of an n-layer ReLU network. Then the paper provides some insights into the smoothing of one-layer ReLU networks (Theorem 1). Finally, a boundary-based saliency map and an extension of integrated gradients are proposed and evaluated in terms of boundary alignment and object localization. \n", "main_review": "The paper has an interesting topic: adding theoretical insights to explainability methods. The paper does especially well on providing a good intuition about the relationships of normals, polytopes and decision boundary (content of 3.1 and first part of 3.2). I also found the paper overall well written (some minor typos and duplicates are listed below). The paper's story of first analyzing the limitations of gradients, fixing the errors, and then evaluating the methods is also good. I address my concerns about the generality and rigor of theorem 1, the evaluation, and the limitations below.\n\n### Proof of theorem 1\n\nTheorem 1 contains an $\\lessapprox$ sign. After checking the appendix, it turns out that the proof is only correct for the case that\n\n> (Dombrowski et al., 2019) points out that the random distribution $p_β (\\epsilon_{i}) =\\frac{β}{(exp (β \\epsilon_{i} / 2)+exp ( β \\epsilon_{i} / 2))^2}\n> $closely resembles a normal distribution with a standard deviation $\\sigma = \\sqrt{\\log 2 \\frac{\\sqrt{2 \\pi}}{\\beta}}$.\n\nHowever, under which conditions does it resemble a normal distribution? (Dombrowski et al., 2019) only made this comment to explain a possible connection to SmoothGrad (see page 8 in Dombrowski et al., 2019). No concrete conditions are given on when or how close the distributions matches. I did not even found how  $\\sigma$  was derived in (Dombrowski et al., 2019) (if you know where, please point me to it). I did a small experiment myself and plotted the distributions. For each plot, the corresponding $\\beta$ is given on top and the normal distribution has $\\sigma = \\sqrt{\\log 2 \\frac{\\sqrt{2 \\pi}}{\\beta}}$ (for the notebook with the code see this [link](https://f002.backblazeb2.com/file/nnnnnnnn/iclr2022/Robust+Models+Are+More+Interpretable+Because+Attributions+Look+Normal.ipynb)).\n\n[[Plots for different $\\beta$s]](https://f002.backblazeb2.com/file/nnnnnnnn/iclr2022/beta_plots.png)\n\n\nAs you can see, it is only close for $\\beta \\approx 1$. Two solutions exist: either provide a theorem with $\\leq$ or give a rigorous discussion on the cases where only $\\approx$  or even > holds. \n\nThe other limitation of Theorem 1 is that it only holds for one-layer ReLU networks. I would find a short discussion helpful why it does not hold for n-layer ReLU networks. In addition, it should be emphasized throughout the paper that Theorem 1 is only for one-layer networks. For example, in the last paragraph of the introduction:\n\n> We present an analysis that sheds light on the previously-observed phenomeon of robust interpretability, showing that alignment between the normal vectors of decision boundaries and models’ gradients is a key ingredient (Proposition 1, Theorem 1)\n\nPlease make clear in that sentence and others that Theorem 1 only addresses one-layer networks. \n\nAt the end of section 3.2, Figure 10 is referenced as empirical validation of Theorem 1, but I do not understand the figure and caption:\n\n> distances in logarithm between SG and BSG against different standard deviations σ of the Gaussian noise. Results are computed on ResNet50. Notice the first column corresponds to σ = 0.\n\nPlease, clarify what you want to evaluate with this figure, e.g. the first column says $\\sigma=0.15$.\n\n### Evaluation\n\nI think the evaluation of the normality to the decision boundary can be improved. In Figure 3, pairs of gradient attribution method and the corresponding boundary attributions (e.g. IG vs. BIG) are compared to evaluate how normal the attributions are. However, why not measure the normality in the feature space $z(x)$ directly.  $z(x)$ is defined such that $f_i(x) = w_i^T z(x)$. We know that $w_i$ must be normal to the decision boundary, as shown in Figure 2a. The corresponding change in z-space of an attribution $g(x)$ would be $\\Delta z = z(x) - z(x + \\alpha g(x))$. Now, we can measure the similarity of the normal $w_i$ and the different attributions: just compute $\\cos(\\Delta z,  w_i)$ for all the different attributions. This evaluation would relate the estimated directions in $x$-space to the ground-truth normals in $z$-space. The current evaluation of attributions methods against their boundary equivalent cannot provide such a ground-truth reference. \n\nThe evaluation using the ground-truth bounding boxes is a good proxy task and seems to be executed correctly. It might make sense to only use images where the bounding box covers less than 50% of the image, as done in (Schulz et al., 2020). The attribution method in (Schulz et al., 2020) might also be an interesting candidate for the evaluation as it was also able to outperform int.grad. and smooth grad. I would also suggest focusing on one or two metrics for the bounding box task instead of four.\n\nI would also encourage the authors to include the sanity check for weight reinitialization (Adebayo et al., 2018). It is easy to implement and should be passed by any new attribution method. \n\n### Limitations\n\nWhile I do not think that the paper requires a human-subject evaluation, its lack should be mentioned in the limitation section. Also the saliency maps look more concentrated, would humans actually profit from it? Even if there is a significant difference, would you expect a large effect size? Please also list that theorem 1 is only for one-layered networks in the limitations. Limitation 2 (not applicable to perturbation attributions) arises from focus of the paper and I think there is not need to mention it. \n\n### Minor Comemnts:\n\n* \"In fact, the fact\" (page 4)\n* smaller difference between the difference between (page 7)\n* Thefore (page 7)\n* Lost clause: \"It is naturally to treat \" (page 8)\n* It should be Table 3 and not Figure 3\n* IamgeNet (page 6)\n* I think it should be \"The RHS of the above equation is Smoothed Gradient\" (page 15)\n\n### References:\n\n(Schulz et al., 2020) https://openreview.net/forum?id=S1xWh1rYwB\n\n(Adebayo et al., 2018) https://arxiv.org/abs/1810.03292\n\n## After Rebuttal Update\n\nThe authors were able to rectify their proof and also provided details to my other questions. While the initial submission was a clear reject, the rebuttal was well done. I agree with the concerns of the others reviewers about novelty. Overall, I increased my rating to marginal above acceptance.", "summary_of_the_review": "While I like that the paper aims to provide a more theoretical justification on attributions, I am not satisfied with the rigor of the theory and the empirical evaluation. I am not convinced that the proof is correct and the alignment with the normals should be checked using ground-truth knowledge. Overall, The paper is well-written, but please fix the grammar. I cannot recommend the paper in its current form for acceptance. If the proof were corrected and the evaluation extend to a ground-truth assessment of the normal, I would reconsider my rating.\n\n**Technical Novelty:** The papers technical contribution is novel. I am less convinced about the significance in its current form. \n\n**Empirical Novelty:** The paper does not present new empirical evaluations or datasets.\n\n**Confidence:** I am confident about my assessment. I read the proof and investigated the issue of resembles-a-normal-distribution in depth. I still might have missed other issues of the proof. While I did looked at the referenced literature about adversarial examples, I am more familiar with the interpretability side of the related work. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635860579015}, {"id": "gqK1Pb5a1Wp", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1892/Reviewer_qvFb"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper has two main contributions.   \na) First it shows that one reason behind the attributions being more interpretable for adversarial robust models is that for these models, the gradient with respect to the input is more closely aligned with the normal direction to a close decision boundary. They empirically verify this claim, by showing that the l2-distance between attributions and their boundary variants(attributions computed at a close point on the decision boundary) are lower for robust models than for standard models.   \nb) Using the previous fact, they devise two new attribution methods, BSM and BIG which can be used to get more interpretability/explanation from even a normal (non-robust) model. They again verify this claim empirically through various quantitative metrics aimed at finding the relation between positive attributions inside a localized bounding box of an object in the image.", "review_text": "Strengths -   \n1) The motivation of the idea is well explained in the paper.\n2) The mathematical foundation required for understanding is also well explained.\n3) I like the effort put in the paper in understanding the reasoning behind interpretable attributions for robust models and then using the info to devise new attribution methods.\n4) For both claims, the paper does extensive qualitative and quantitative experiments.\n\nWeakness - \n1) The new attributions devised in the paper seem very similar to the AGI attribution(mentioned in the paper) approach. In BIG, the attributions are computed along interpolations of x and its closest adversarial image, whereas in AGI the attributions are computed along each step of the adversarial image generation.\n2) In Table1 mentioned in the paper, the improvements along the two metrics used in other papers are not really significant. The improvement only comes along with the two new metrics proposed in this paper. I would like to see a comparison against some other metrics used in the related works such as top-1 localization accuracy as used in [1] and [2].\n3) For a fairer comparison with the AGI method, can the authors use only the PGD attack for the adversarial image generation? Or, the authors can also incorporate other adversarial images (and not just PGD) in AGI. For instance, the AGI method can be used to compute the attributions along each step of PGD, CW, and AutoPGD attacks and the final attribution is just the mean attribution of all three approaches.\n4) [3] showed that their attribution technique works well with even multiple objects in the image. Can the authors show some qualitative results of comparison for multiple objects across different attribution methods.\n\nReferences -   \n[1] Attention-based Dropout Layer for Weakly Supervised Object Localization. Choe et al. 2019. \n[2] On The Benefits Of Models With Perceptually  Aligned Gradients. Aggarwal et al. 2020. \n[3] Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. Wang et al. 2020. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper has two main contributions.   \na) First it shows that one reason behind the attributions being more interpretable for adversarial robust models is that for these models, the gradient with respect to the input is more closely aligned with the normal direction to a close decision boundary. They empirically verify this claim, by showing that the l2-distance between attributions and their boundary variants(attributions computed at a close point on the decision boundary) are lower for robust models than for standard models.   \nb) Using the previous fact, they devise two new attribution methods, BSM and BIG which can be used to get more interpretability/explanation from even a normal (non-robust) model. They again verify this claim empirically through various quantitative metrics aimed at finding the relation between positive attributions inside a localized bounding box of an object in the image.", "main_review": "Strengths -   \n1) The motivation of the idea is well explained in the paper.\n2) The mathematical foundation required for understanding is also well explained.\n3) I like the effort put in the paper in understanding the reasoning behind interpretable attributions for robust models and then using the info to devise new attribution methods.\n4) For both claims, the paper does extensive qualitative and quantitative experiments.\n\nWeakness - \n1) The new attributions devised in the paper seem very similar to the AGI attribution(mentioned in the paper) approach. In BIG, the attributions are computed along interpolations of x and its closest adversarial image, whereas in AGI the attributions are computed along each step of the adversarial image generation.\n2) In Table1 mentioned in the paper, the improvements along the two metrics used in other papers are not really significant. The improvement only comes along with the two new metrics proposed in this paper. I would like to see a comparison against some other metrics used in the related works such as top-1 localization accuracy as used in [1] and [2].\n3) For a fairer comparison with the AGI method, can the authors use only the PGD attack for the adversarial image generation? Or, the authors can also incorporate other adversarial images (and not just PGD) in AGI. For instance, the AGI method can be used to compute the attributions along each step of PGD, CW, and AutoPGD attacks and the final attribution is just the mean attribution of all three approaches.\n4) [3] showed that their attribution technique works well with even multiple objects in the image. Can the authors show some qualitative results of comparison for multiple objects across different attribution methods.\n\nReferences -   \n[1] Attention-based Dropout Layer for Weakly Supervised Object Localization. Choe et al. 2019. \n[2] On The Benefits Of Models With Perceptually  Aligned Gradients. Aggarwal et al. 2020. \n[3] Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. Wang et al. 2020. ", "summary_of_the_review": "I have a few concerns regarding the quantitative experiments in the paper, which are mentioned in the Weakness section. I will be willing to update my ratings if the authors address all my points. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635626201196}, {"id": "O950apcHslM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1892/Reviewer_t25u"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces boundary attributions, which leverage the connection between boundary normal vectors and gradients to yield explanations for non-robust models that carry over many of the favorable properties that have been observed of explanations on\nrobust models. It also proposes a BIG to explain models. ", "review_text": "Strengths:\n\n1. Table 1 shows the empirical results are good.\n\nWeakness:\n\n1. My major concern for this paper is that the conclusion has already known. For example, Ilyas et al shows that robust models can produce better perceptual aligned features when gradient descent, and adversarial robust models are known to have smooth decision boundary [1].\n\n[1] Theoretically Principled Trade-off between Robustness and Accuracy. ICML 2019. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces boundary attributions, which leverage the connection between boundary normal vectors and gradients to yield explanations for non-robust models that carry over many of the favorable properties that have been observed of explanations on\nrobust models. It also proposes a BIG to explain models. ", "main_review": "Strengths:\n\n1. Table 1 shows the empirical results are good.\n\nWeakness:\n\n1. My major concern for this paper is that the conclusion has already known. For example, Ilyas et al shows that robust models can produce better perceptual aligned features when gradient descent, and adversarial robust models are known to have smooth decision boundary [1].\n\n[1] Theoretically Principled Trade-off between Robustness and Accuracy. ICML 2019. ", "summary_of_the_review": "The conclusion is not novel to me, which is already known to the community.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635281450038}], "openreview_url": "https://openreview.net/forum?id=FD8xldQIgdq", "arxiv_id": "2103.11257", "paper_pdf": "papers/FD8xldQIgdq.pdf", "paper_pdf_sha256": "b45d21710e26232a8d2442332c71d84584d7baae403893002c0134d92079f124", "paper_pdf_bytes": 25874427, "paper_pdf_source": "openreview", "code_url": "https://github.com/zifanw/boundary", "code_repository": "zifanw/boundary", "code_commit": "a901b7e75ee8cdf683c490ba076a9308301db2b9", "code_archive": "repos/FD8xldQIgdq.zip", "code_archive_sha256": "0ebf7a00af4ba389f96dca33844c234c254a89f1a46a19ee0ce0f0393a98de6c", "code_archive_bytes": 6457814, "code_file_count": 8, "code_extensions": {".py": 5, ".ipynb": 2, ".sh": 1}, "github_disk_usage_kb": 6321, "github_languages": {"Jupyter Notebook": 1957190, "Python": 47304, "Shell": 61}, "github_archived": false, "github_pushed_at": "2021-08-13T20:28:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/boundary-attributions-provide-normal-vector"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "SnhmiKUPWL", "year": 2021, "status": "rejected", "title": "Leveraging Class Hierarchies with Metric-Guided Prototype Learning", "authors": ["Vivien Sainte Fare Garnot", "Loic Landrieu"], "authorids": ["~Vivien_Sainte_Fare_Garnot1", "~Loic_Landrieu1"], "authors_source": "OpenReview API", "abstract": "In many classification tasks, the set of classes can be organized according to a meaningful hierarchy. This structure can be used to assess the severity of confusing each pair of classes, and summarized  under the form of a cost matrix which also defines a finite metric. We propose to integrate this metric in the supervision of a prototypical network in order to model the hierarchical class structure. Our method relies on jointly learning a feature-extracting network and a set of class representations, or prototypes, which incorporate the error metric into their relative arrangement in the embedding space. We show that this simultaneous training allows for consistent improvement of the severity of the network's errors with regard to the class hierarchy when compared to traditional methods and other prototype-based strategies. Furthermore, when the induced metric contains insight on the data structure, our approach improves the overall precision as well. Experiments on four different public datasets—from agricultural time series classification to depth image semantic segmentation—validate our approach.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tgrjUQogTT", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper832/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes a method to integrate the cost of errors into a classification algorithm. The proposed loss function has two terms: one for pushing samples towards their corresponding class prototype (L_{data}), and another for forcing/guiding the pairwise distance between class prototypes to follow a predefined values (L_{distortion}).  \n\nMarginally below acceptance. \n\nI like the idea of using a priory knowledge about similarities between classes to regularize the learning process.  But,\n\n1) Out of 4 datasets, the proposed method works better than others on only two datasets for which in one case the addition of proposed term L_{distortion} has helped, but it actually degraded the results in the other dataset. So, the results are inconclusive.\n\n2) pp.3 the assumption of having a symmetric misclassification cost seems very limiting. Effectively, this assumption implies the type 1 and type 2 errors must have equal cost. The proposed method is about treating different errors differently and yet it assumes type 1 and type 2 errors are equal. Referring to the example, provided in the introduction (paragraph 1), cost of misclassifying a crossing pedestrian as a street light should be way more than the cost of classifying a street light as a crossing pedestrian. \nAnother assumption of the proposed method is triangle inequality i.e. D[k, l] + D[l,m] >= D[k,m] but this assumption is not justified except saying the cases where the assumptions do not hold, are out of scope of this paper. \n\npp.1 says, \"this error discrepancy is not taken into account ... in the evaluation metrics\". This statement is incorrect. At least in a simple binary classification task, we have type 1 and type 2 errors. We have recall and precision. They are all about putting different weights to different errors. \n\npp.2 Caption of Figure 1: tree is created based on authors perceived visual similarity. The tree should reflect the cost of class confusions. From paragraph 1 in page 1, it is inferred that the cost of a class confusion is determined by the actual consequences e.g. cost of confusing a streetlamp with a crossing pedestrian. Such cost is different than how similar the two classes are.  The notion of \"cost\" and \"similarity\" are not the same but used interchangeably in the paper. \n\npp.7 the soft-label method is the top runner in cifar-100. Its performance on the other 3 datasets may be improved if the temperature parameter is adjusted to compensate the various number of classes. Of course, same can be said about tuning hyperparameters of the proposed method.  \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "assumptions are not adequately discussed and results are inconclusive", "review": "The paper proposes a method to integrate the cost of errors into a classification algorithm. The proposed loss function has two terms: one for pushing samples towards their corresponding class prototype (L_{data}), and another for forcing/guiding the pairwise distance between class prototypes to follow a predefined values (L_{distortion}).  \n\nMarginally below acceptance. \n\nI like the idea of using a priory knowledge about similarities between classes to regularize the learning process.  But,\n\n1) Out of 4 datasets, the proposed method works better than others on only two datasets for which in one case the addition of proposed term L_{distortion} has helped, but it actually degraded the results in the other dataset. So, the results are inconclusive.\n\n2) pp.3 the assumption of having a symmetric misclassification cost seems very limiting. Effectively, this assumption implies the type 1 and type 2 errors must have equal cost. The proposed method is about treating different errors differently and yet it assumes type 1 and type 2 errors are equal. Referring to the example, provided in the introduction (paragraph 1), cost of misclassifying a crossing pedestrian as a street light should be way more than the cost of classifying a street light as a crossing pedestrian. \nAnother assumption of the proposed method is triangle inequality i.e. D[k, l] + D[l,m] >= D[k,m] but this assumption is not justified except saying the cases where the assumptions do not hold, are out of scope of this paper. \n\npp.1 says, \"this error discrepancy is not taken into account ... in the evaluation metrics\". This statement is incorrect. At least in a simple binary classification task, we have type 1 and type 2 errors. We have recall and precision. They are all about putting different weights to different errors. \n\npp.2 Caption of Figure 1: tree is created based on authors perceived visual similarity. The tree should reflect the cost of class confusions. From paragraph 1 in page 1, it is inferred that the cost of a class confusion is determined by the actual consequences e.g. cost of confusing a streetlamp with a crossing pedestrian. Such cost is different than how similar the two classes are.  The notion of \"cost\" and \"similarity\" are not the same but used interchangeably in the paper. \n\npp.7 the soft-label method is the top runner in cifar-100. Its performance on the other 3 datasets may be improved if the temperature parameter is adjusted to compensate the various number of classes. Of course, same can be said about tuning hyperparameters of the proposed method.  \n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603901520109}, {"id": "mebuKQC2Bno", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper832/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "In order to organize my review, I use the NeurIPS2020 template with slight adaptions to fit the ICLR requirements. \n\n### Summary and contributions: Briefly summarize the paper and its contributions.\n\nThe paper proposes a method to jointly learn a feature-extractor (a neural network) and a set of prototypes. Learning is performed in an end-to-end setting and the prototype distribution is regularized to follow a given distribution. For this purpose, the authors propose a cost matrix that assesses the severity of errors between classes. This matrix is organized in such a way that it yields a finite metric. The distribution of prototypes is regularized such that the distances between prototypes reflect this finite metric. In several experiments, the authors evaluate the advantages of such a setting with respect to an average cost measure.\n\n### Strengths: Describe the strengths of the work.\n\nThe work is well motivated and how the entire network setup is realized is easy to understand and straightforward.\n\n\n### Weaknesses: Explain the limitations of this work along the same axes as above.\n\nA major weakness is that the authors have not cited all the relevant previous work about prototypical networks (see my comments below). Taking all the relevant literature into account, the contribution of the paper is small. Moreover, the proposed method and observed effects raise several questions that are not sufficiently addressed. For example, the authors observe that the learning with the squared Euclidean distance does not perform as good as with the Euclidean distance, but they have not investigated why this is the case (see my further comments below).\n\n\n### Correctness: Are the claims and method correct? Is the empirical methodology correct?\n\nThe current version of the paper has some minor mathematical mistakes. However, they can be fixed and I consider them to be of little relevance for the correctness of the method. I will list them below.\nOne serious correctness point is that the authors claim (see page 4 top and page 2 contributions first bullet point) that they are inventing a principle to learn the prototypes jointly in an end-to-end setting. Considering the state of the art, this claim is incorrect.\n\n\n### Clarity: Is the paper well written?\n\nThe paper is well written and the mathematical formulations have a good quality.\n\n\n### Relation to prior work: Is it clearly discussed how this work differs from previous contributions?\n\nThe paper has not discussed all the relevant prior work. The idea to learn the prototypes jointly is not new:\n* Chen, Chaofan and Li, Oscar and Tao, Daniel and Barnett, Alina and Su, Jonathan and Rudin, Cynthia. “This Looks Like That: Deep Learning for Interpretable Image Recognition.” NeurIPS (2019).\n* Hong-Ming Yang and Xu-Yao Zhang and Fei Yin and Cheng-Lin Liu. “Robust Classification with Convolutional Prototype Learning.” CVPR (2018)\n* Villmann, Thomas and Biehl, Michael and Villmann, Andrea and Saralajew, Sascha. “Fusion of deep learning architectures, multilayer feedforward networks and Learning Vector Quantizers for deep classification learning.” WSOM+ (2017)\n* to name just a few…\n\nNote that prototype-based classifiers and especially nearest prototype classifiers have been invented long before Tibshirani et al. (2002). For example, see Kohonen’s Learning Vector Quantization (LVQ) algorithms, including the original heuristic versions and the end-to-end trainable versions like Generalized LVQ (by Sato and Yamada).\nMoreover, the incorporation of a cost matrix into the learning processes for prototypes was also done before. For example, Kaden, Marika, Wieland Hermann, and Thomas Villmann. \"Attention Based Classification Learning in GLVQ and Asymmetric Misclassification Assessment.\" Advances in Self-Organizing Maps and Learning Vector Quantization. Springer, Cham, 2014. 77-87. Of course, in this work, the authors have not used the cost matrix to penalize the prototype positions, but it was included to account for different severity between class confusions.\n\n### Reproducibility: Are there enough details to reproduce the major results of this work?\n\nThe paper provides a lot of details about the experimental setup, especially in the appendix. However, some information is not provided, for example, the initialization strategy of the architecture.\n\n\n### Additional feedback, comments, suggestions for improvement, questions for the authors, and a **recommendation (accept or reject)**.\n\nThe authors should be precise about the assumptions of $D$. In the current version, the definition of $D$ is not correct to talk about a finite metric space (see page 3 Method and page 1). The missing property is the identity of indiscernibles.\n\nThe authors provide a good motivation for why the incorporation of a cost matrix is useful. However, could the authors explain why they then assume that the matrix $D$ is symmetric? Because, according to the motivation, the severity of misclassifying a *crossing pedestrian* as a *street lamp* is different from misclassifying a *street lamp* as a *crossing pedestrian*.\n\nThe first sentence in Section 3.1 is difficult to read due to “sample $n$”, consider writing $x_n$.\n\nWhat is the difference of Equation (3) to a cross-entropy loss with a one-hot encoding for the correct class?\n\nThe authors observed that the squared Euclidean distance does not perform as well as the Euclidean distance. I assume that this happens because the squared Euclidean distance does not fulfill the triangle inequality that is, on the other hand, assumed for $D$. Could the authors comment on this? \n\nI don’t see that the non-differentiability of the square root at zero causes any issues in the training. If the argument becomes zero, the distance is zero and, therefore, the prototype is equal to the feature vector of the sample. Given the given interpretation of the attraction and repulsion forces on the data and the prototypes (see page 3 below), in which direction should they be pushed or repelled if they have a perfect match? There exists no update direction that could improve the model. Therefore, it is feasible to just stop the gradient or to ignore such samples.\n\nHave the authors observed training instabilities due to large values of the distance $d$?\n\nIf I consider Figure 1, I wonder why the learned prototypes tend to be outside the class centers (e.g., see figure c). Can the authors elaborate on that?\n\nThe statement on page 1 bottom about the cross-entropy is incorrect. Whether the cross entropy “singles out the prediction” depends on the true probability vector.\n\nThe experimental results are not convincing as they don’t show a significant benefit compared with state-of-the-art methods. \n\nI have problems with the statement following Equation (4). Could the authors explain in more detail why a prototype arrangement regarding $D$ and low distortion is advantageous in terms of AC?\n\nThe google colab link provided in the supplementary was not accessible. \n\nConsidering all the above points, I vote for rejection, since the paper does not have the right maturity level to be published at ICLR and the scientific contribution is small.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Reinvention of several known concepts; small scientific contribution", "review": "In order to organize my review, I use the NeurIPS2020 template with slight adaptions to fit the ICLR requirements. \n\n### Summary and contributions: Briefly summarize the paper and its contributions.\n\nThe paper proposes a method to jointly learn a feature-extractor (a neural network) and a set of prototypes. Learning is performed in an end-to-end setting and the prototype distribution is regularized to follow a given distribution. For this purpose, the authors propose a cost matrix that assesses the severity of errors between classes. This matrix is organized in such a way that it yields a finite metric. The distribution of prototypes is regularized such that the distances between prototypes reflect this finite metric. In several experiments, the authors evaluate the advantages of such a setting with respect to an average cost measure.\n\n### Strengths: Describe the strengths of the work.\n\nThe work is well motivated and how the entire network setup is realized is easy to understand and straightforward.\n\n\n### Weaknesses: Explain the limitations of this work along the same axes as above.\n\nA major weakness is that the authors have not cited all the relevant previous work about prototypical networks (see my comments below). Taking all the relevant literature into account, the contribution of the paper is small. Moreover, the proposed method and observed effects raise several questions that are not sufficiently addressed. For example, the authors observe that the learning with the squared Euclidean distance does not perform as good as with the Euclidean distance, but they have not investigated why this is the case (see my further comments below).\n\n\n### Correctness: Are the claims and method correct? Is the empirical methodology correct?\n\nThe current version of the paper has some minor mathematical mistakes. However, they can be fixed and I consider them to be of little relevance for the correctness of the method. I will list them below.\nOne serious correctness point is that the authors claim (see page 4 top and page 2 contributions first bullet point) that they are inventing a principle to learn the prototypes jointly in an end-to-end setting. Considering the state of the art, this claim is incorrect.\n\n\n### Clarity: Is the paper well written?\n\nThe paper is well written and the mathematical formulations have a good quality.\n\n\n### Relation to prior work: Is it clearly discussed how this work differs from previous contributions?\n\nThe paper has not discussed all the relevant prior work. The idea to learn the prototypes jointly is not new:\n* Chen, Chaofan and Li, Oscar and Tao, Daniel and Barnett, Alina and Su, Jonathan and Rudin, Cynthia. “This Looks Like That: Deep Learning for Interpretable Image Recognition.” NeurIPS (2019).\n* Hong-Ming Yang and Xu-Yao Zhang and Fei Yin and Cheng-Lin Liu. “Robust Classification with Convolutional Prototype Learning.” CVPR (2018)\n* Villmann, Thomas and Biehl, Michael and Villmann, Andrea and Saralajew, Sascha. “Fusion of deep learning architectures, multilayer feedforward networks and Learning Vector Quantizers for deep classification learning.” WSOM+ (2017)\n* to name just a few…\n\nNote that prototype-based classifiers and especially nearest prototype classifiers have been invented long before Tibshirani et al. (2002). For example, see Kohonen’s Learning Vector Quantization (LVQ) algorithms, including the original heuristic versions and the end-to-end trainable versions like Generalized LVQ (by Sato and Yamada).\nMoreover, the incorporation of a cost matrix into the learning processes for prototypes was also done before. For example, Kaden, Marika, Wieland Hermann, and Thomas Villmann. \"Attention Based Classification Learning in GLVQ and Asymmetric Misclassification Assessment.\" Advances in Self-Organizing Maps and Learning Vector Quantization. Springer, Cham, 2014. 77-87. Of course, in this work, the authors have not used the cost matrix to penalize the prototype positions, but it was included to account for different severity between class confusions.\n\n### Reproducibility: Are there enough details to reproduce the major results of this work?\n\nThe paper provides a lot of details about the experimental setup, especially in the appendix. However, some information is not provided, for example, the initialization strategy of the architecture.\n\n\n### Additional feedback, comments, suggestions for improvement, questions for the authors, and a **recommendation (accept or reject)**.\n\nThe authors should be precise about the assumptions of $D$. In the current version, the definition of $D$ is not correct to talk about a finite metric space (see page 3 Method and page 1). The missing property is the identity of indiscernibles.\n\nThe authors provide a good motivation for why the incorporation of a cost matrix is useful. However, could the authors explain why they then assume that the matrix $D$ is symmetric? Because, according to the motivation, the severity of misclassifying a *crossing pedestrian* as a *street lamp* is different from misclassifying a *street lamp* as a *crossing pedestrian*.\n\nThe first sentence in Section 3.1 is difficult to read due to “sample $n$”, consider writing $x_n$.\n\nWhat is the difference of Equation (3) to a cross-entropy loss with a one-hot encoding for the correct class?\n\nThe authors observed that the squared Euclidean distance does not perform as well as the Euclidean distance. I assume that this happens because the squared Euclidean distance does not fulfill the triangle inequality that is, on the other hand, assumed for $D$. Could the authors comment on this? \n\nI don’t see that the non-differentiability of the square root at zero causes any issues in the training. If the argument becomes zero, the distance is zero and, therefore, the prototype is equal to the feature vector of the sample. Given the given interpretation of the attraction and repulsion forces on the data and the prototypes (see page 3 below), in which direction should they be pushed or repelled if they have a perfect match? There exists no update direction that could improve the model. Therefore, it is feasible to just stop the gradient or to ignore such samples.\n\nHave the authors observed training instabilities due to large values of the distance $d$?\n\nIf I consider Figure 1, I wonder why the learned prototypes tend to be outside the class centers (e.g., see figure c). Can the authors elaborate on that?\n\nThe statement on page 1 bottom about the cross-entropy is incorrect. Whether the cross entropy “singles out the prediction” depends on the true probability vector.\n\nThe experimental results are not convincing as they don’t show a significant benefit compared with state-of-the-art methods. \n\nI have problems with the statement following Equation (4). Could the authors explain in more detail why a prototype arrangement regarding $D$ and low distortion is advantageous in terms of AC?\n\nThe google colab link provided in the supplementary was not accessible. \n\nConsidering all the above points, I vote for rejection, since the paper does not have the right maturity level to be published at ICLR and the scientific contribution is small.\n", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603888096238}, {"id": "OFHClMFdS-", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper832/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Strengths and weaknesses:\n+\nNice idea\nConsistent improvements over cross entropy for hierarchical class structures\nImprovements w.r.t other competitors (though not consistent) \nGood ablation study\n-\nThe improvements are small\nThe novelty is not very significant\n \n\nMore comments:\n\n\nFigure 1:\n-\tIt is not clear what distortion is at this stage\n-\tIt is not clear what perturbed MNist is, and respectively: why is the error of a 3-layer CNN so high (12-16% error are reported)? CNNs with 2-3 layers can solve MNist with accuracy higher than 99.5%?\n-\tThis figure cannot be presented on page 2 without proper definitions. It should be either presented on page 5, where the experiment is defined, or better explained  \nPage 4: It is said that s can be computed efficiently and this is shown in the appendix, but the version I have do not have an appendix\nPage 6: the XE+EMD method is not present in a comprehensible manner. 1) p_k symbols are used without definition (tough I think I these are the network predictions p(\\hat{y}=k|I)  2) the relation of the formula presented to the known EMD is not clear. The latter is a problem solved as linear programming or similar, and not a closed form formula  3) it is not clear what the role of \\mu is and why can be set to 3 irrespective of the scale of metric D\npage 7: \nThe experiments show small, but consistent improvements of the suggested method over standard cross entropy, and improvements versus most competitors in most cases\n\nI have read the reviews of others and the author's response. My main impression of the work remains as it was: that it is  nice idea with small but significant empirical success. However, my acquaintance with the previous literature in this subject is partial compared to the acquaintance of other reviewers, so It may well be possible that they are in a better position than me to see the incremental nature of the proposed work. I therefore reduce the rating a bit, to become closer to the consensus.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Nice idea, providing consistent (small) accuracy improvements in well conducted experiments", "review": "Strengths and weaknesses:\n+\nNice idea\nConsistent improvements over cross entropy for hierarchical class structures\nImprovements w.r.t other competitors (though not consistent) \nGood ablation study\n-\nThe improvements are small\nThe novelty is not very significant\n \n\nMore comments:\n\n\nFigure 1:\n-\tIt is not clear what distortion is at this stage\n-\tIt is not clear what perturbed MNist is, and respectively: why is the error of a 3-layer CNN so high (12-16% error are reported)? CNNs with 2-3 layers can solve MNist with accuracy higher than 99.5%?\n-\tThis figure cannot be presented on page 2 without proper definitions. It should be either presented on page 5, where the experiment is defined, or better explained  \nPage 4: It is said that s can be computed efficiently and this is shown in the appendix, but the version I have do not have an appendix\nPage 6: the XE+EMD method is not present in a comprehensible manner. 1) p_k symbols are used without definition (tough I think I these are the network predictions p(\\hat{y}=k|I)  2) the relation of the formula presented to the known EMD is not clear. The latter is a problem solved as linear programming or similar, and not a closed form formula  3) it is not clear what the role of \\mu is and why can be set to 3 irrespective of the scale of metric D\npage 7: \nThe experiments show small, but consistent improvements of the suggested method over standard cross entropy, and improvements versus most competitors in most cases\n\nI have read the reviews of others and the author's response. My main impression of the work remains as it was: that it is  nice idea with small but significant empirical success. However, my acquaintance with the previous literature in this subject is partial compared to the acquaintance of other reviewers, so It may well be possible that they are in a better position than me to see the incremental nature of the proposed work. I therefore reduce the rating a bit, to become closer to the consensus.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603883498719}, {"id": "2wPNS7FbyRQ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper832/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper proposes a prototype-based approach to learning with prior hierarchical knowledge of categories. The main idea behind their approach is to perform metric-guided penalization on top of a prototype-based cross-entropy loss. The penalization is scale-free, yet the loss can be computed in closed form, allowing for a hierarchical guidance to prototype locations.\n\nStrengths:\n\nThe starting motivation and direction of the paper are clear. If hierarchical knowledge is available, it is intuitive to exploit such information. Care has been taken to make the research open and reproducible. Along with the submission, an anonymized repo and notebook with details are provided. Both help with understanding the approach and its workings in practice.\n\nThe scale-free component of the proposed regularization is interesting, most notably the fact that the corresponding loss can be computed in closed form. By making the hierarchical loss scale free, a clash with the cross-entropy loss - which tries to push prototypes as far away as possible from each other - is avoided.\n\nWeaknesses:\n\nThe main weakness of the paper is that its innovation is limited, especially in light of missed recent works on prototype-based approaches using hierarchies in non-Euclidean spaces. The innovation of the paper is in Section 3.2, which means that the technical novelty of the paper is limited to a loss with hierarchical loss. More pressingly, relevant recent literature is missed on using hierarchical knowledge in deep networks. What the authors consider future work (\"Among other promising avenues for further research, we plan to investigate the use of non-Euclidean embedding spaces, which are known to be well-suited for embedding hierarchies [...]\", page 8), has readily been proposed, see e.g.:\n\n[1] Liu, Shaoteng, et al. \"Hyperbolic Visual Embedding Learning for Zero-Shot Recognition.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n[2] Long, Teng, et al. \"Searching for Actions on the Hyperbole.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n[3] Khrulkov, Valentin, et al. \"Hyperbolic image embeddings.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. (note: this paper does not explicitly utilize hierarchical knowledge)\n\nAll three above papers, from the same conference as the often cited work by Bertinetto et al. (2020) in the submission, investigate prototype-based approaches in non-Euclidean spaces. [1,2] explicitly use hierarchical knowledge to steer the prototypes. Hyperbolic spaces are known to be a better fit for hierarchies, see e.g. [4].\n\n[4] Nickel, Maximillian, and Douwe Kiela. \"Poincaré embeddings for learning hierarchical representations.\" Advances in neural information processing systems. 2017.\n\nCompared to recent works on prototypes in hyperbolic spaces with hierarchical knowledge, the paper provides limited novelty. As the authors state that moving the prototypes to non-Euclidean spaces is future work for this submission, the other recent works make the paper come across as somewhat outdated.\n\nEmpirically, the paper provides a comparison to multiple hierarchical alternatives on four datasets. The use of four datasets is appreciated, as it provides a clear picture of the strengths and limitations of the approaches. Besides missing comparisons to e.g. the supervised setup of [2], Figures 2 and Table 2 show that the difference with alternatives is minimal. E.g. the guided prototypes seem to only outperform soft labels on iNaturalist-19 in terms of Average Cost. Table 2 paints a picture that fixing the scale or using a coarser ranking-based approach hardly affect the Average Cost on all datasets and only provide some improvements on one of the four datasets.\n\nConclusion:\n\nThe paper addresses an open problem in deep learning literature in a clear manner. The open science setup of the submission is appreciated and the scale-free regularizer is interesting. In light of recent advances in hyperbolic prototype-based approaches with hierarchical knowledge, the innovation of this work is minimal. The empirical benefits of the proposed approach are also not always convincing. For the rebuttal, it would be interesting to discuss and compare the submission to hyperbolic hierarchical prototype approaches.\n\nOpinion post rebuttals:\n\nAfter reading the rebuttal and the other reviews, I remain of the opinion that more work is needed before warranting acceptance. This paper indeed learns prototypes on the fly in contrast to e.g. Mettes et al. 2019. That method was however not designed for hierarchical knowledge, while several recent CVPR papers were. Since the novelty over these papers is limited and direct comparisons are lacking, more research is needed.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "Summary:\n\nThis paper proposes a prototype-based approach to learning with prior hierarchical knowledge of categories. The main idea behind their approach is to perform metric-guided penalization on top of a prototype-based cross-entropy loss. The penalization is scale-free, yet the loss can be computed in closed form, allowing for a hierarchical guidance to prototype locations.\n\nStrengths:\n\nThe starting motivation and direction of the paper are clear. If hierarchical knowledge is available, it is intuitive to exploit such information. Care has been taken to make the research open and reproducible. Along with the submission, an anonymized repo and notebook with details are provided. Both help with understanding the approach and its workings in practice.\n\nThe scale-free component of the proposed regularization is interesting, most notably the fact that the corresponding loss can be computed in closed form. By making the hierarchical loss scale free, a clash with the cross-entropy loss - which tries to push prototypes as far away as possible from each other - is avoided.\n\nWeaknesses:\n\nThe main weakness of the paper is that its innovation is limited, especially in light of missed recent works on prototype-based approaches using hierarchies in non-Euclidean spaces. The innovation of the paper is in Section 3.2, which means that the technical novelty of the paper is limited to a loss with hierarchical loss. More pressingly, relevant recent literature is missed on using hierarchical knowledge in deep networks. What the authors consider future work (\"Among other promising avenues for further research, we plan to investigate the use of non-Euclidean embedding spaces, which are known to be well-suited for embedding hierarchies [...]\", page 8), has readily been proposed, see e.g.:\n\n[1] Liu, Shaoteng, et al. \"Hyperbolic Visual Embedding Learning for Zero-Shot Recognition.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n[2] Long, Teng, et al. \"Searching for Actions on the Hyperbole.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n[3] Khrulkov, Valentin, et al. \"Hyperbolic image embeddings.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. (note: this paper does not explicitly utilize hierarchical knowledge)\n\nAll three above papers, from the same conference as the often cited work by Bertinetto et al. (2020) in the submission, investigate prototype-based approaches in non-Euclidean spaces. [1,2] explicitly use hierarchical knowledge to steer the prototypes. Hyperbolic spaces are known to be a better fit for hierarchies, see e.g. [4].\n\n[4] Nickel, Maximillian, and Douwe Kiela. \"Poincaré embeddings for learning hierarchical representations.\" Advances in neural information processing systems. 2017.\n\nCompared to recent works on prototypes in hyperbolic spaces with hierarchical knowledge, the paper provides limited novelty. As the authors state that moving the prototypes to non-Euclidean spaces is future work for this submission, the other recent works make the paper come across as somewhat outdated.\n\nEmpirically, the paper provides a comparison to multiple hierarchical alternatives on four datasets. The use of four datasets is appreciated, as it provides a clear picture of the strengths and limitations of the approaches. Besides missing comparisons to e.g. the supervised setup of [2], Figures 2 and Table 2 show that the difference with alternatives is minimal. E.g. the guided prototypes seem to only outperform soft labels on iNaturalist-19 in terms of Average Cost. Table 2 paints a picture that fixing the scale or using a coarser ranking-based approach hardly affect the Average Cost on all datasets and only provide some improvements on one of the four datasets.\n\nConclusion:\n\nThe paper addresses an open problem in deep learning literature in a clear manner. The open science setup of the submission is appreciated and the scale-free regularizer is interesting. In light of recent advances in hyperbolic prototype-based approaches with hierarchical knowledge, the innovation of this work is minimal. The empirical benefits of the proposed approach are also not always convincing. For the rebuttal, it would be interesting to discuss and compare the submission to hyperbolic hierarchical prototype approaches.\n\nOpinion post rebuttals:\n\nAfter reading the rebuttal and the other reviews, I remain of the opinion that more work is needed before warranting acceptance. This paper indeed learns prototypes on the fly in contrast to e.g. Mettes et al. 2019. That method was however not designed for hierarchical knowledge, while several recent CVPR papers were. Since the novelty over these papers is limited and direct comparisons are lacking, more research is needed.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602579212879}], "openreview_url": "https://openreview.net/forum?id=SnhmiKUPWL", "arxiv_id": "2007.03047", "paper_pdf": "papers/SnhmiKUPWL.pdf", "paper_pdf_sha256": "2ef02da2fb1d48418933787b103bd90527527938a3d2b97b645fb320f40463df", "paper_pdf_bytes": 1110853, "paper_pdf_source": "openreview", "code_url": "https://github.com/VSainteuf/metric-guided-prototypes-pytorch", "code_repository": "VSainteuf/metric-guided-prototypes-pytorch", "code_commit": "fae90235b57a71f3ab682917eaf5517799fbffed", "code_archive": "repos/SnhmiKUPWL.zip", "code_archive_sha256": "4fa50ce76fd27a5e61f291f897bfb9724e900ea39be93371509fa912754abf63", "code_archive_bytes": 1257982, "code_file_count": 10, "code_extensions": {".py": 9, ".ipynb": 1}, "github_disk_usage_kb": 4169, "github_languages": {"Jupyter Notebook": 320324, "Python": 27733}, "github_archived": false, "github_pushed_at": "2021-10-20T11:30:19Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/metric-guided-prototype-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "74QmBTV0Zf", "year": 2025, "status": "rejected", "title": "Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models", "authors": ["Michael Günther", "Isabelle Mohr", "Daniel James Williams", "Bo Wang", "Han Xiao"], "authorids": ["~Michael_Günther1", "~Isabelle_Mohr1", "~Daniel_James_Williams1", "~Bo_Wang31", "~Han_Xiao8"], "authors_source": "OpenReview API", "abstract": "Many use cases require retrieving smaller portions of text, and dense vector-based retrieval systems often perform better with shorter text segments, as the semantics are less likely to be \"over-compressed\" in the embeddings. Consequently, practitioners often split text documents into smaller chunks and encode them separately. However, chunk embeddings created in this way can lose contextual information from surrounding chunks, resulting in sub-optimal representations. In this paper, we introduce a novel method called \"late chunking, which leverages long context embedding models to first embed all tokens of the long text, with chunking applied after the transformer model and just before mean pooling - hence the term \"late\" in its naming. The resulting chunk embeddings capture the full contextual information, leading to superior results across various retrieval tasks.  The method is generic enough to be applied to a wide range of long-context embedding models and works without additional training. To further increase the effectiveness of late chunking, we propose a dedicated fine-tuning approach for embedding models.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "huL3qylUfW", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7298/Reviewer_i4oQ"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper introduces late chunking for document embeddings, which suggests that instead of chunking the text and then computing the embedding for individual chunks, one can alternatively first compute the embeddings for the whole document (or a great portion of it containing the desired chunk, in their long late chunking method), and then extract the embeddings for that chunk. Experiments conducted on retrieval tasks show the effectiveness of the proposed method and many ablations are conducted.", "review_text": "The paper introduces late chunking for document embeddings, which suggests that instead of chunking the text and then computing the embedding for individual chunks, one can alternatively first compute the embeddings for the whole document (or a great portion of it containing the desired chunk, in their long late chunking method), and then extract the embeddings for that chunk. Experiments conducted on retrieval tasks show the effectiveness of the proposed method and many ablations are conducted.", "strengths": "The experiments are well designed and the paper is easy to understand. \nThe performance of the method is consistently better than the compared baseline on almost all tasks & models experimented.", "weaknesses": "While the proposed method shows a good consistent, it seems to only work for an imaginary scenario --- Chunking methods are designed such that models can handle longer piece of text, but the proposed method only works if we can encode text longer than the chunk size.", "questions": "About the weakness, the reviewer can still imagine that in some cases where the chunk size is much smaller than the model length, this method can be useful. The author should present more practical examples and arguments that shows current practice often overlook this design, and that the work can signal the importance of late chunking.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces late chunking for document embeddings, which suggests that instead of chunking the text and then computing the embedding for individual chunks, one can alternatively first compute the embeddings for the whole document (or a great portion of it containing the desired chunk, in their long late chunking method), and then extract the embeddings for that chunk. Experiments conducted on retrieval tasks show the effectiveness of the proposed method and many ablations are conducted.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The experiments are well designed and the paper is easy to understand. \nThe performance of the method is consistently better than the compared baseline on almost all tasks & models experimented.", "weaknesses": "While the proposed method shows a good consistent, it seems to only work for an imaginary scenario --- Chunking methods are designed such that models can handle longer piece of text, but the proposed method only works if we can encode text longer than the chunk size.", "questions": "About the weakness, the reviewer can still imagine that in some cases where the chunk size is much smaller than the model length, this method can be useful. The author should present more practical examples and arguments that shows current practice often overlook this design, and that the work can signal the importance of late chunking.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731359132450}, {"id": "ZgTh7aBZxb", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7298/Reviewer_uQcW"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposed a novel method called “late chunking”, which leverages long context embedding models to first embed all tokens of the long text, with chunking applied after the transformer model and just before mean pooling. The resulting chunk embeddings capture the full contextual information, leading to superior results across various retrieval tasks. The method is generic enough to be applied to a wide range of long-context embedding models and works without additional training. To further increase the effectiveness of late\nchunking, the authors also proposed a dedicated fine-tuning approach for embedding models. They experimented their method on BeIR benchmark and the results showed that by using late chunking, they are able to improve the retrieval performance (measured by NDCG) on several datasets.", "review_text": "This paper proposed a novel method called “late chunking”, which leverages long context embedding models to first embed all tokens of the long text, with chunking applied after the transformer model and just before mean pooling. The resulting chunk embeddings capture the full contextual information, leading to superior results across various retrieval tasks. The method is generic enough to be applied to a wide range of long-context embedding models and works without additional training. To further increase the effectiveness of late\nchunking, the authors also proposed a dedicated fine-tuning approach for embedding models. They experimented their method on BeIR benchmark and the results showed that by using late chunking, they are able to improve the retrieval performance (measured by NDCG) on several datasets.", "strengths": "- The proposed method is intuitive and simple but yield decent performance across embedding models and tasks.\n- The proposed method can be directly used off the shelf, which would benefit the research community a lot.\n- The paper is well organized and presented.", "weaknesses": "- I have some concerns regarding the results presented in Figure 3. In this figure, we observe that performance with late chunking declines across several datasets, including NarrativeQA (Chunk size > 128), 2WikiMultiHopQA (Chunk size > 16), SummScreenFD (Chunk size > 128), QMsum (Chunk size > 256), Needle-8192 (Chunk size > 4), and Passkey-8192 (Chunk size > 32). This pattern raises the question of whether the fusion of contextual information might actually lead to regression in fact-based retrieval tasks where extensive contextual information may be less relevant.\n\n- I am also concerned about the experimental setup, particularly the choice of the BeIR benchmark as the primary testbed. The motivation for this choice feels less justified. To make a strong case that late chunking enhances retrieval performance in scenarios where contextual information is beneficial, it would be ideal to use a dedicated dataset (or a subset of datasets) where contextual information is  necessary for optimal retrieval performance. This approach would allow for a more informative breakdown of performance in contexts that benefit from contextual information versus those that do not. However, with the datasets selected, it is unclear to me how much contextual information contributes to performance gains and whether it might cause regressions in other scenarios.\n\n- Another issue with the experimental setting is that only retrieval performance is measured not the downstream performance. Ultimately, downstream performance is what people care about. It is unclear whether improvements in retrieval performance translate into meaningful gains in downstream tasks.\n\n- Section 4.5 feels somewhat incomplete. Rather than providing a systematic comparison, it functions more as a case study, which, in my opinion, adds less weight to the paper's central argument. I suggest reallocating this section’s space to address the concerns outlined above.", "questions": "In table 2, it's quite interesting to see the results show different trend on different datasets and different embedding models. For example, on TRECCOVID, late chunking helps least on Jv2 while on NFCorpus, it helps most on Jv2 and less on Jv3 and No. Do you have an idea what causes the differences?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposed a novel method called “late chunking”, which leverages long context embedding models to first embed all tokens of the long text, with chunking applied after the transformer model and just before mean pooling. The resulting chunk embeddings capture the full contextual information, leading to superior results across various retrieval tasks. The method is generic enough to be applied to a wide range of long-context embedding models and works without additional training. To further increase the effectiveness of late\nchunking, the authors also proposed a dedicated fine-tuning approach for embedding models. They experimented their method on BeIR benchmark and the results showed that by using late chunking, they are able to improve the retrieval performance (measured by NDCG) on several datasets.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- The proposed method is intuitive and simple but yield decent performance across embedding models and tasks.\n- The proposed method can be directly used off the shelf, which would benefit the research community a lot.\n- The paper is well organized and presented.", "weaknesses": "- I have some concerns regarding the results presented in Figure 3. In this figure, we observe that performance with late chunking declines across several datasets, including NarrativeQA (Chunk size > 128), 2WikiMultiHopQA (Chunk size > 16), SummScreenFD (Chunk size > 128), QMsum (Chunk size > 256), Needle-8192 (Chunk size > 4), and Passkey-8192 (Chunk size > 32). This pattern raises the question of whether the fusion of contextual information might actually lead to regression in fact-based retrieval tasks where extensive contextual information may be less relevant.\n\n- I am also concerned about the experimental setup, particularly the choice of the BeIR benchmark as the primary testbed. The motivation for this choice feels less justified. To make a strong case that late chunking enhances retrieval performance in scenarios where contextual information is beneficial, it would be ideal to use a dedicated dataset (or a subset of datasets) where contextual information is  necessary for optimal retrieval performance. This approach would allow for a more informative breakdown of performance in contexts that benefit from contextual information versus those that do not. However, with the datasets selected, it is unclear to me how much contextual information contributes to performance gains and whether it might cause regressions in other scenarios.\n\n- Another issue with the experimental setting is that only retrieval performance is measured not the downstream performance. Ultimately, downstream performance is what people care about. It is unclear whether improvements in retrieval performance translate into meaningful gains in downstream tasks.\n\n- Section 4.5 feels somewhat incomplete. Rather than providing a systematic comparison, it functions more as a case study, which, in my opinion, adds less weight to the paper's central argument. I suggest reallocating this section’s space to address the concerns outlined above.", "questions": "In table 2, it's quite interesting to see the results show different trend on different datasets and different embedding models. For example, on TRECCOVID, late chunking helps least on Jv2 while on NFCorpus, it helps most on Jv2 and less on Jv3 and No. Do you have an idea what causes the differences?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730959714297}, {"id": "IuwwbsWTeL", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7298/Reviewer_sEvM"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 5, "summary": "The paper propose a late-chunking strategy for text embeddings, where the texts are firstly past through a text encoder and then the pooling are done at chunks of the output token embeddings to form chunk embeddings. Experiment results show the proposed late chunking strategy performs better that naive chunking on the BEIR benchmark.", "review_text": "The paper propose a late-chunking strategy for text embeddings, where the texts are firstly past through a text encoder and then the pooling are done at chunks of the output token embeddings to form chunk embeddings. Experiment results show the proposed late chunking strategy performs better that naive chunking on the BEIR benchmark.", "strengths": "1. Chunking is an important problem in applying text embeddings in practical applications such as RAG.\n2. The paper is clearly written and well presented.\n3. The proposed solution is simple to implement for practitioners.", "weaknesses": "1. For naive chunking, the standard practice is to have some overlapping strides between chunks, and to include meta information such as document title in every chunks when available. It is unclear whether the author of this paper follows this practice in implementing the baselines.\n2. The paper uses a relative small chunk size (up-to 512) in the experiments when the embeddings studied support 8k context length. As shown in the ablation, the gains from late chunking diminish when the chunk size goes from 16 up to 512. It is unclear whether it is still effective when the chunk size approaches the embedding length limit of 8k, where the benefit of chunking is most useful.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper propose a late-chunking strategy for text embeddings, where the texts are firstly past through a text encoder and then the pooling are done at chunks of the output token embeddings to form chunk embeddings. Experiment results show the proposed late chunking strategy performs better that naive chunking on the BEIR benchmark.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "1. Chunking is an important problem in applying text embeddings in practical applications such as RAG.\n2. The paper is clearly written and well presented.\n3. The proposed solution is simple to implement for practitioners.", "weaknesses": "1. For naive chunking, the standard practice is to have some overlapping strides between chunks, and to include meta information such as document title in every chunks when available. It is unclear whether the author of this paper follows this practice in implementing the baselines.\n2. The paper uses a relative small chunk size (up-to 512) in the experiments when the embeddings studied support 8k context length. As shown in the ablation, the gains from late chunking diminish when the chunk size goes from 16 up to 512. It is unclear whether it is still effective when the chunk size approaches the embedding length limit of 8k, where the benefit of chunking is most useful.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730775871193}, {"id": "gncgx720AK", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7298/Reviewer_2G5K"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces a novel technique called \"late chunking\" for improving text embeddings in retrieval tasks by leveraging long-context embedding models. Unlike traditional chunking methods that split text before encoding, late chunking first encodes the entire document and then applies chunking, thereby preserving full contextual information within each chunk. This paper evaluates this approach on multiple retrieval datasets and demonstrates that late chunking consistently outperforms naive chunking methods across various chunking strategies (fixed-size, sentence-based, semantic) and models.", "review_text": "This paper introduces a novel technique called \"late chunking\" for improving text embeddings in retrieval tasks by leveraging long-context embedding models. Unlike traditional chunking methods that split text before encoding, late chunking first encodes the entire document and then applies chunking, thereby preserving full contextual information within each chunk. This paper evaluates this approach on multiple retrieval datasets and demonstrates that late chunking consistently outperforms naive chunking methods across various chunking strategies (fixed-size, sentence-based, semantic) and models.", "strengths": "1. This paper proposes a late chunking technique that utilizes a long-context retriever to encode the full text before performing chunking, which reduces information loss caused by direct chunking.\n2. This paper conducts extensive analytical experiments to explore the technical details of late chunking.", "weaknesses": "1. The experiments are not comprehensive enough. As only a subset of the BEIR benchmark is used in Section 4.1, limiting the assessment of the effectiveness of late chunking.\n2. The proposed method needs high computational resources: Late chunking requires encoding the entire input with a long-context LLM before chunking, whereas standard chunking only encodes each chunk separately, resulting in shorter sequence lengths and reduced attention computation costs. As noted in Section 4.1, \"splitting documents into smaller chunks increases the computational effort of the evaluation.\"\n3. When dealing with longer texts, a sliding-window approach is still required, which could still lead to the loss of long-range dependency information.", "questions": "1. In Table 2, it seems that late chunking aims to better segment chunks, yet the use of sentence boundaries and fixed-size boundaries indicates that both late chunking and naive chunking methods are dividing chunks in the same way. Then why does late chunking still can generate higher-quality embeddings and achieve better performance?\n2. Do the authors believe that late chunking could yield better results with LLM retrievers employing causal attention mechanisms?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a novel technique called \"late chunking\" for improving text embeddings in retrieval tasks by leveraging long-context embedding models. Unlike traditional chunking methods that split text before encoding, late chunking first encodes the entire document and then applies chunking, thereby preserving full contextual information within each chunk. This paper evaluates this approach on multiple retrieval datasets and demonstrates that late chunking consistently outperforms naive chunking methods across various chunking strategies (fixed-size, sentence-based, semantic) and models.", "soundness": 2, "presentation": 3, "contribution": 3, "strengths": "1. This paper proposes a late chunking technique that utilizes a long-context retriever to encode the full text before performing chunking, which reduces information loss caused by direct chunking.\n2. This paper conducts extensive analytical experiments to explore the technical details of late chunking.", "weaknesses": "1. The experiments are not comprehensive enough. As only a subset of the BEIR benchmark is used in Section 4.1, limiting the assessment of the effectiveness of late chunking.\n2. The proposed method needs high computational resources: Late chunking requires encoding the entire input with a long-context LLM before chunking, whereas standard chunking only encodes each chunk separately, resulting in shorter sequence lengths and reduced attention computation costs. As noted in Section 4.1, \"splitting documents into smaller chunks increases the computational effort of the evaluation.\"\n3. When dealing with longer texts, a sliding-window approach is still required, which could still lead to the loss of long-range dependency information.", "questions": "1. In Table 2, it seems that late chunking aims to better segment chunks, yet the use of sentence boundaries and fixed-size boundaries indicates that both late chunking and naive chunking methods are dividing chunks in the same way. Then why does late chunking still can generate higher-quality embeddings and achieve better performance?\n2. Do the authors believe that late chunking could yield better results with LLM retrievers employing causal attention mechanisms?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730289057140}], "openreview_url": "https://openreview.net/forum?id=74QmBTV0Zf", "arxiv_id": "2409.04701", "paper_pdf": "papers/74QmBTV0Zf.pdf", "paper_pdf_sha256": "7f40499529248f5d920fb3d021e728b68c17c8c5cb8219426f908bb01d742c1b", "paper_pdf_bytes": 435456, "paper_pdf_source": "openreview", "code_url": "https://github.com/jina-ai/late-chunking", "code_repository": "jina-ai/late-chunking", "code_commit": "1d3bb02bf091becd0771455e4e7959463935e26c", "code_archive": "repos/74QmBTV0Zf.zip", "code_archive_sha256": "0003f6c5611ca3e991f1d3d284e179f5c87f66db7110df1868248973ff5159b0", "code_archive_bytes": 381516, "code_file_count": 12, "code_extensions": {".py": 11, ".ipynb": 1}, "github_disk_usage_kb": 421, "github_languages": {"Python": 80840, "Jupyter Notebook": 7185}, "github_archived": false, "github_pushed_at": "2024-12-23T13:38:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/late-chunking-contextual-chunk-embeddings"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "gusHSc09zj", "year": 2024, "status": "rejected", "title": "Discovering Mixtures of Structural Causal Models from Time Series Data", "authors": ["Sumanth Varambally", "Yian Ma", "Rose Yu"], "authorids": ["~Sumanth_Varambally1", "~Yian_Ma1", "~Rose_Yu1"], "authors_source": "OpenReview API", "abstract": "In fields such as finance, climate science, and neuroscience, inferring causal relationships from time series data poses a formidable challenge. While contemporary techniques can handle non-linear relationships between variables and flexible noise distributions, they rely on the simplifying assumption that data originates from the same underlying causal model. In this work, we relax this assumption and perform causal discovery from time series data originating from mixtures of different causal models. We infer both the underlying structural causal models and the posterior probability for each sample belonging to a specific mixture component. Our approach employs an end-to-end training process that maximizes an evidence-lower bound for data likelihood. Through extensive experimentation on both synthetic and real-world datasets, we demonstrate that our method surpasses state-of-the-art benchmarks in causal discovery tasks, particularly when the data emanates from diverse underlying causal graphs. Theoretically, we prove the identifiability of such a model under some mild assumptions.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "SV7RGvthUW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2304/Reviewer_vKpF"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Authors propose a variational inference adopted for causal discovery in time-series data; in particular for mixtures of multiple causal graphs.", "review_text": "Authors propose a variational inference adopted for causal discovery in time-series data; in particular for mixtures of multiple causal graphs.", "strengths": "- in general, exposition is good (although there is room for clarity and explanations in formal parts ).\n- addressing  mixture of multiple-causal graph  discovery is an interesting/relevant direction of research, that hasn't been much investigated (although I have my reservations)\n- experimental results are coupled with a theoretical structural identifiability result.", "weaknesses": "- my main skepticism is due to the fact that all the results are for the training set, which I am surprised (and had a stronger positive impression until that point). Trivial enough: For an unquestionable positive score, I would rather need result on test data, on unseen graphs. \n\n-   How different the causal graphs in these mixture models (and also the data i.e., average SHD within the cluster is missing.  This is really important to really understand what is going on behaviour of the method. (hence, my question). No surprise to see its effect on highly-imbalanced one.\n\n- on minor point,  an illustrative toy example is lacking would be very useful.", "questions": "-  What is the effect of the distance between causal graphs to the performance?\n\n-  g_1 and g_2 are not defined in Theorem 1, is it a typo: should be h? or a particular graph?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Authors propose a variational inference adopted for causal discovery in time-series data; in particular for mixtures of multiple causal graphs.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- in general, exposition is good (although there is room for clarity and explanations in formal parts ).\n- addressing  mixture of multiple-causal graph  discovery is an interesting/relevant direction of research, that hasn't been much investigated (although I have my reservations)\n- experimental results are coupled with a theoretical structural identifiability result.", "weaknesses": "- my main skepticism is due to the fact that all the results are for the training set, which I am surprised (and had a stronger positive impression until that point). Trivial enough: For an unquestionable positive score, I would rather need result on test data, on unseen graphs. \n\n-   How different the causal graphs in these mixture models (and also the data i.e., average SHD within the cluster is missing.  This is really important to really understand what is going on behaviour of the method. (hence, my question). No surprise to see its effect on highly-imbalanced one.\n\n- on minor point,  an illustrative toy example is lacking would be very useful.", "questions": "-  What is the effect of the distance between causal graphs to the performance?\n\n-  g_1 and g_2 are not defined in Theorem 1, is it a typo: should be h? or a particular graph?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1700417252166}, {"id": "25upE4QmoU", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2304/Reviewer_vYSG"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This work presents a method that infers causal relationships from time-series data by allowing for mixtures of different causal models, rather than assuming a single underlying causal model. The authors utilize end-to-end variational inference to optimize parameters and perform inference on causal graphs, functional equations, and sample membership within mixture components. The method is assessed on both synthetic and real-world datasets, showcasing competitive performance on training data in causal discovery tasks. The authors establish the identifiability of the proposed model under mild assumptions.", "review_text": "This work presents a method that infers causal relationships from time-series data by allowing for mixtures of different causal models, rather than assuming a single underlying causal model. The authors utilize end-to-end variational inference to optimize parameters and perform inference on causal graphs, functional equations, and sample membership within mixture components. The method is assessed on both synthetic and real-world datasets, showcasing competitive performance on training data in causal discovery tasks. The authors establish the identifiability of the proposed model under mild assumptions.", "strengths": "The paper addresses a relevant and interesting aspect of causal discovery in time-series data.\n\nThe proposed loss function is a simple extension of the standard variational inference objective, enabling efficient end-to-end training with a mixture of core causal discovery models.\n\nThe method is flexible in terms of the choice of core causal structure learning algorithms, inheriting the structural identifiability properties of these algorithms.\n\nCompetitive empirical results (although on training data) are reported across various metrics.", "weaknesses": "The proposed objective function can be seen as a straightforward extension of the standard variational inference optimisation framework. Therefore, the overall novelty and significance of the work may be somewhat limited in this regard.\n\nA major drawback of the work is that the reported results are based on training data, as the proposed method depends on learnable sample-specific parameters. It raises questions about why an encoder producing a K-way categorical random variable given a sample was not considered by the authors.\n\nThe lack of reported results on generalisation limits insight into the proposed method's ability to perform beyond the training data.\n\nAlthough the method claims flexibility in terms of core causal structure learning algorithms, performance is only demonstrated based on one causal discovery method.", "questions": "Can you please clarify how the AUROC metric is computed for the different methods being compared?\n\nRegarding the experiment on Netsim-permuted data, wouldn’t one expect the results of the proposed method to be inherently favorable compared to non-mixture alternatives? This is because by permuting the variables, we intentionally manipulate the data to explicitly align with the underlying assumption of a mixture model.\n\nWhat specific hyper-parameter value K was chosen for the DREAM3 Gene Network experiment?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a method that infers causal relationships from time-series data by allowing for mixtures of different causal models, rather than assuming a single underlying causal model. The authors utilize end-to-end variational inference to optimize parameters and perform inference on causal graphs, functional equations, and sample membership within mixture components. The method is assessed on both synthetic and real-world datasets, showcasing competitive performance on training data in causal discovery tasks. The authors establish the identifiability of the proposed model under mild assumptions.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper addresses a relevant and interesting aspect of causal discovery in time-series data.\n\nThe proposed loss function is a simple extension of the standard variational inference objective, enabling efficient end-to-end training with a mixture of core causal discovery models.\n\nThe method is flexible in terms of the choice of core causal structure learning algorithms, inheriting the structural identifiability properties of these algorithms.\n\nCompetitive empirical results (although on training data) are reported across various metrics.", "weaknesses": "The proposed objective function can be seen as a straightforward extension of the standard variational inference optimisation framework. Therefore, the overall novelty and significance of the work may be somewhat limited in this regard.\n\nA major drawback of the work is that the reported results are based on training data, as the proposed method depends on learnable sample-specific parameters. It raises questions about why an encoder producing a K-way categorical random variable given a sample was not considered by the authors.\n\nThe lack of reported results on generalisation limits insight into the proposed method's ability to perform beyond the training data.\n\nAlthough the method claims flexibility in terms of core causal structure learning algorithms, performance is only demonstrated based on one causal discovery method.", "questions": "Can you please clarify how the AUROC metric is computed for the different methods being compared?\n\nRegarding the experiment on Netsim-permuted data, wouldn’t one expect the results of the proposed method to be inherently favorable compared to non-mixture alternatives? This is because by permuting the variables, we intentionally manipulate the data to explicitly align with the underlying assumption of a mixture model.\n\nWhat specific hyper-parameter value K was chosen for the DREAM3 Gene Network experiment?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698845247683}, {"id": "w4hx6skdNp", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2304/Reviewer_dKxB"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "An ELBO method is introduced for finding and assigning weight to component SEMs in a mixture and drawing causal inferences from this mixture.", "review_text": "An ELBO method is introduced for finding and assigning weight to component SEMs in a mixture and drawing causal inferences from this mixture.", "strengths": "The problem of identifying the different SCMs in a mixture model where each mixed dataset uses a different graph and parameterization is extremely important, so I’m glad it’s being addressed.", "weaknesses": "I do have some issues.\n\n1.\tIf we turn to the experimental section of the paper, we get two examples: one for NetSim (which I’m very familiar with) and another for the DREAM3 gene network. Both of these have problems with respect to the goal of this paper, suggesting that the choice of experimental datasets could be improved.\n2.\tThe problem with the NetSim data is that it’s not an extremely convincing time series, as the records in the simulation are spaced far enough apart in time to render the data nearly i.i.d. In fact, analyzing it as i.i.d. often yields better results than analyzing it as time series, frustratingly, as in this paper: \n\nMulti-subject search correctly identifies causal connections and most causal directions in the DCM models of the Smith et al. simulation study. NeuroImage, 58(3), 838-848. \n\nThis paper also treats the distributions as a mixture, though doesn't assume non-i.i.d.\n\n3.\tAs a result, it seems that any study proposing a time series analysis of this data should do a comparison of this result to one obtained by treating the data as i.i.d. instead, since this is a known phenomenon for this particular dataset. This is an issue because the proposed method is specifically designed to deal with time series.\n4.\tThe problem with the DREAM3 examples is that none of the reported results have any lift with AUC; they’re all stuck around 0.5, which is more or less random. The differences between the methods are slight. This result doesn’t give a case where the proposed method is particularly helpful.", "questions": "Can you find better examples to show the usefulness of the theory?\n\nWhere can the theory really shine so far as the empirical application is concerned? Where can it be expected to do significantly better than alternative methods?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "An ELBO method is introduced for finding and assigning weight to component SEMs in a mixture and drawing causal inferences from this mixture.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "The problem of identifying the different SCMs in a mixture model where each mixed dataset uses a different graph and parameterization is extremely important, so I’m glad it’s being addressed.", "weaknesses": "I do have some issues.\n\n1.\tIf we turn to the experimental section of the paper, we get two examples: one for NetSim (which I’m very familiar with) and another for the DREAM3 gene network. Both of these have problems with respect to the goal of this paper, suggesting that the choice of experimental datasets could be improved.\n2.\tThe problem with the NetSim data is that it’s not an extremely convincing time series, as the records in the simulation are spaced far enough apart in time to render the data nearly i.i.d. In fact, analyzing it as i.i.d. often yields better results than analyzing it as time series, frustratingly, as in this paper: \n\nMulti-subject search correctly identifies causal connections and most causal directions in the DCM models of the Smith et al. simulation study. NeuroImage, 58(3), 838-848. \n\nThis paper also treats the distributions as a mixture, though doesn't assume non-i.i.d.\n\n3.\tAs a result, it seems that any study proposing a time series analysis of this data should do a comparison of this result to one obtained by treating the data as i.i.d. instead, since this is a known phenomenon for this particular dataset. This is an issue because the proposed method is specifically designed to deal with time series.\n4.\tThe problem with the DREAM3 examples is that none of the reported results have any lift with AUC; they’re all stuck around 0.5, which is more or less random. The differences between the methods are slight. This result doesn’t give a case where the proposed method is particularly helpful.", "questions": "Can you find better examples to show the usefulness of the theory?\n\nWhere can the theory really shine so far as the empirical application is concerned? Where can it be expected to do significantly better than alternative methods?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698682462211}, {"id": "IpQDpssf2P", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2304/Reviewer_bLPE"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The authors proposed an approach to perform causal discovery from time series data originating from mixtures of different causal models. This approach can simultaneously infer the underlying causal graphs and the posterior likelihood of each sample belonging to a specific mixture component by maximizing the evidence lower bound, on top of the framework of the Rhino algorithm.", "review_text": "The authors proposed an approach to perform causal discovery from time series data originating from mixtures of different causal models. This approach can simultaneously infer the underlying causal graphs and the posterior likelihood of each sample belonging to a specific mixture component by maximizing the evidence lower bound, on top of the framework of the Rhino algorithm.", "strengths": "1. The problem setting is interesting, and the approach holds potential for broad applicability.\n2. The authors conduct extensive experiments on simulated and two real-world datasets, compared with several baselines.\n3. The authors characterize a sufficient condition for the identifiability of such mixture models and explain the relationship between the constructed evidence lower bound and the data likelihood.\n4. The authors provide ablation studies.", "weaknesses": "1. Certain notations and explanations remain unclear, and these will be described in the upcoming Questions section.\n2. Some details concerning the comparison between the proposed method and baseline algorithms in the experiments are perplexing, and these questions will also be raised in the later section.\n3. The baselines used in the paper are not algorithms designed for multiple DAGs and mixture SCMs. Even though the authors mentioned other algorithms designed for multiple DAGs, none has been applied in the comparison.", "questions": "**1. Notations and explanations**\n\n1.1 The SCM described in equation (1) differs from equation (6) and also the equation below equation (6); which one is correct? Do you assume additive noise?\n\n1.2 In Theorem 1 equation (*), what is the definition of $a_i$? Does it mean one sample from $i$th SCM?\n\n1.3 In Theorem 1, what are $g_1$ and $g_2$? Should they be $h_1$ and $h_2$?\n\n1.4 As mentioned in the main paper, \"This highlights the idea that learning a mixture model is only beneficial when the underlying SCMs differ from one another significantly. \" could you clarify what the significant difference is here?\n\n**2. Details in the experiment**\n\n2.1 In the experiment section, \"(per sample) indicates that the baseline predicts one graph per sample.\" how to apply the algorithm with only one sample? For example, PCMCI$^{+}$ needs CI tests, and CI tests need a set of samples instead of only one sample.\n\n2.2 In the experiment performance section, AUROC and F1 are used as metrics. However, it is not mentioned which kind of AUROC and F1 refer to adjacency or orientation AUROC/F1?\n\n2.3 As far as I know, the Rhino algorithm is not designed for heterogeneous time series data. In the experimental section, would it be feasible to compare the proposed algorithm with \"Rhino (grouped)\" as well? This would involve applying the Rhino algorithm to the grouped data based on the true underlying causal graph. Considering this, the results regarding Rhino in Figures 3 and 6 could be more comprehensive if grouped data were utilized.\n\n2.4 Could the post-processing of PCMCI$^{+}$ outputs potentially affect the AUROC/F1 scores of PCMCI$^{+}$? If it does, would this influence tend to favor the proposed method in the comparison results?\n\n2.5 In the synthetic data experiment, the metrics are averaged across 3 runs. Does each run requires a newly generated synthetic dataset? Personally, 3 runs seem to be a limited number for comprehensive evaluation.\n\n2.6 In Figure 3, it seems a little strange that the performance of PCMCI$^{+}$ (grouped) is better with $D=10$ than with $D=5$. The same phenomenon happened in other algorithms. Do you have any clue on why this happened?\n\n2.7 In Figure 4, \"The accuracy is averaged across 3 runs and across data dimensionality D = 5, 10, 20.\", is it more appropriate to plot the results for $D=5,10,20$ separately as the value of $D$ could affect the accuracy? Again, 3 runs seem limited to me.\n\n2.8 Could you explain more about \"The clustering accuracy and performance metrics show high standard deviation when K is set to the true number of mixture components K* = 10.\" which is stated in the appendix and explain more about the \"buffers\" when $K>K^{*}$?\n\n**3. Other questions**\n\n3.1 Why maximizing ELBO is equivalent to maximizing or minimizing each term separately in the log-likelihood in terms of $\\theta, \\phi$, and $\\psi$ as $\\phi$ and $\\psi$ appear together in the second term?\n\n3.2 Is it better to mention the Rhino algorithm in the related work section and briefly explain how it works as the proposed algorithm is built on top of Rhino?\n\n3.3 Is it essential to verify whether the data satisfies condition (*) to ensure the reliability of the results? If so, how to verify this?\n\n3.4 The paper mentions that people can also infer functional equations through the proposed model. Could you help me locate the related outputs in the experiment section? Does this mean the causal effect?\n\n3.5 Has any code been provided?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors proposed an approach to perform causal discovery from time series data originating from mixtures of different causal models. This approach can simultaneously infer the underlying causal graphs and the posterior likelihood of each sample belonging to a specific mixture component by maximizing the evidence lower bound, on top of the framework of the Rhino algorithm.", "soundness": "3 good", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The problem setting is interesting, and the approach holds potential for broad applicability.\n2. The authors conduct extensive experiments on simulated and two real-world datasets, compared with several baselines.\n3. The authors characterize a sufficient condition for the identifiability of such mixture models and explain the relationship between the constructed evidence lower bound and the data likelihood.\n4. The authors provide ablation studies.", "weaknesses": "1. Certain notations and explanations remain unclear, and these will be described in the upcoming Questions section.\n2. Some details concerning the comparison between the proposed method and baseline algorithms in the experiments are perplexing, and these questions will also be raised in the later section.\n3. The baselines used in the paper are not algorithms designed for multiple DAGs and mixture SCMs. Even though the authors mentioned other algorithms designed for multiple DAGs, none has been applied in the comparison.", "questions": "**1. Notations and explanations**\n\n1.1 The SCM described in equation (1) differs from equation (6) and also the equation below equation (6); which one is correct? Do you assume additive noise?\n\n1.2 In Theorem 1 equation (*), what is the definition of $a_i$? Does it mean one sample from $i$th SCM?\n\n1.3 In Theorem 1, what are $g_1$ and $g_2$? Should they be $h_1$ and $h_2$?\n\n1.4 As mentioned in the main paper, \"This highlights the idea that learning a mixture model is only beneficial when the underlying SCMs differ from one another significantly. \" could you clarify what the significant difference is here?\n\n**2. Details in the experiment**\n\n2.1 In the experiment section, \"(per sample) indicates that the baseline predicts one graph per sample.\" how to apply the algorithm with only one sample? For example, PCMCI$^{+}$ needs CI tests, and CI tests need a set of samples instead of only one sample.\n\n2.2 In the experiment performance section, AUROC and F1 are used as metrics. However, it is not mentioned which kind of AUROC and F1 refer to adjacency or orientation AUROC/F1?\n\n2.3 As far as I know, the Rhino algorithm is not designed for heterogeneous time series data. In the experimental section, would it be feasible to compare the proposed algorithm with \"Rhino (grouped)\" as well? This would involve applying the Rhino algorithm to the grouped data based on the true underlying causal graph. Considering this, the results regarding Rhino in Figures 3 and 6 could be more comprehensive if grouped data were utilized.\n\n2.4 Could the post-processing of PCMCI$^{+}$ outputs potentially affect the AUROC/F1 scores of PCMCI$^{+}$? If it does, would this influence tend to favor the proposed method in the comparison results?\n\n2.5 In the synthetic data experiment, the metrics are averaged across 3 runs. Does each run requires a newly generated synthetic dataset? Personally, 3 runs seem to be a limited number for comprehensive evaluation.\n\n2.6 In Figure 3, it seems a little strange that the performance of PCMCI$^{+}$ (grouped) is better with $D=10$ than with $D=5$. The same phenomenon happened in other algorithms. Do you have any clue on why this happened?\n\n2.7 In Figure 4, \"The accuracy is averaged across 3 runs and across data dimensionality D = 5, 10, 20.\", is it more appropriate to plot the results for $D=5,10,20$ separately as the value of $D$ could affect the accuracy? Again, 3 runs seem limited to me.\n\n2.8 Could you explain more about \"The clustering accuracy and performance metrics show high standard deviation when K is set to the true number of mixture components K* = 10.\" which is stated in the appendix and explain more about the \"buffers\" when $K>K^{*}$?\n\n**3. Other questions**\n\n3.1 Why maximizing ELBO is equivalent to maximizing or minimizing each term separately in the log-likelihood in terms of $\\theta, \\phi$, and $\\psi$ as $\\phi$ and $\\psi$ appear together in the second term?\n\n3.2 Is it better to mention the Rhino algorithm in the related work section and briefly explain how it works as the proposed algorithm is built on top of Rhino?\n\n3.3 Is it essential to verify whether the data satisfies condition (*) to ensure the reliability of the results? If so, how to verify this?\n\n3.4 The paper mentions that people can also infer functional equations through the proposed model. Could you help me locate the related outputs in the experiment section? Does this mean the causal effect?\n\n3.5 Has any code been provided?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698636012848}], "openreview_url": "https://openreview.net/forum?id=gusHSc09zj", "arxiv_id": "2310.06312", "paper_pdf": "papers/gusHSc09zj.pdf", "paper_pdf_sha256": "a1653ffb9495cfc06f9376bc1a39359d2396e43200e55f806d03b5ebdcdef87b", "paper_pdf_bytes": 918134, "paper_pdf_source": "openreview", "code_url": "https://github.com/Rose-STL-Lab/MCD", "code_repository": "Rose-STL-Lab/MCD", "code_commit": "a9de419a6293b34d43b7519621a100dda51fc5a3", "code_archive": "repos/gusHSc09zj.zip", "code_archive_sha256": "1c3d37399f2867c53e5d954df863f49646d4a8786c50928580161958835f1737", "code_archive_bytes": 511566, "code_file_count": 49, "code_extensions": {".py": 43, ".sh": 6}, "github_disk_usage_kb": 500, "github_languages": {"Python": 218323, "Shell": 1152}, "github_archived": false, "github_pushed_at": "2025-06-26T01:59:50Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/discovering-mixtures-of-structural-causal"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "_AkC4QYxF5", "year": 2023, "status": "rejected", "title": "Closing the Gap Between SVRG and TD-SVRG with Gradient Splitting", "authors": ["Arsenii Mustafin", "Ioannis Paschalidis", "Alex Olshevsky"], "authorids": ["~Arsenii_Mustafin1", "~Ioannis_Paschalidis1", "~Alex_Olshevsky1"], "authors_source": "OpenReview API", "abstract": "Temporal difference (TD) learning is a simple algorithm for policy evaluation\nin reinforcement learning. The performance of TD learning is affected by high\nvariance and it can be naturally enhanced with variance reduction techniques, such\nas the Stochastic Variance Reduced Gradient (SVRG) method. Recently, multiple\nworks have sought to fuse TD learning with SVRG to obtain a policy evaluation\nmethod with a linear rate of convergence. However, the resulting convergence rate\nis significantly weaker than what is achieved by SVRG in the setting of convex\noptimization. In this work we utilize a recent interpretation of TD-learning as the\nsplitting of the gradient of an appropriately chosen function, thus simplifying the\nalgorithm and fusing TD with SVRG. We prove a linear convergence bound that\nis identical to the convergence bound available for SVRG in the convex setting.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "2qs26TNMc3", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper604/Reviewer_158U"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper uses SVRG to reduce variance in TD learning. The resulting algorithm is analyzed using the gradient splitting perspective on TD learning to prove finite sample convergence rates that match those of SVRG in the setting of convex optimization. The analysis is done for several settings, and the findings are illustrated with experiments on several environments.", "review_text": "~~Despite really liking this paper, I must recommend the current version be rejected due to concerns about clarity and the experiments detailed above. If the authors address these concerns satisfactorily and no critical issues are found by the other reviewers, I would recommend acceptance of a revised version.~~ The authors have addressed most of my concerns in their response, except some of my concerns about clarity (making these clarity changes would drastically reduce the work required of each reader to understand the paper). I have increased my score to reflect this.", "strengths": "**Strengths:**\n1. The analysis uses the gradient splitting approach, which yields a significantly simpler algorithm, analysis, learning rate/step size, and a better convergence rate.\n1. The paper considers several different settings, which increases the number of readers who would be interested.\n1. The paper provides recommendations for setting the learning rate, and actually uses this recommended learning rate in the experiments.\n\n**Potential Weaknesses/Questions:**\n1. Many grammatical errors, including typographic errors (\"TD-leaning\", \"apprxoimate\") and missing words, especially definite and indefinite articles. For example, \"determining *the* expected reward *an* agent will achieve if it chooses *actions* according to *a* stationary policy\". Some grammatical issues are listed below, but there were too many for me to write down all of them.\n1. ~~The paper doesn't comment on the assumptions made other than stating that they are standard. It would be nice to include a sentence or two explaining why the assumptions are reasonable or whether they are just for convenience (like Assumption 2; the feature vectors can always be normalized to make it true).~~\n1. Jumping between the notation used in other papers and the notation used in this paper is difficult. It would be better to convert the results of other papers to the notation used in this paper.\n1. ~~Font size of figures is too small to read without zooming in a lot.~~\n1. I have some concerns about the experiments:\n    1. ~~Why weren't the parameters for each algorithm set using a grid search the way they were for PD-SVRG? Without doing this, it's not clear that the chosen parameters are representative of the performance of each algorithm, and hence no conclusion about the performance of each algorithm can be drawn.~~\n    1. ~~Why was the learning rate for TD set to $1/\\sqrt{t}$? It seems like the algorithm stops learning very quickly due to the learning rate shrinking so quickly. How does the algorithm perform with a learning rate of $1/t$ instead?~~ I realized I was confused about this.\n    1. ~~What were the parameter values checked during the grid search? This would make the experiments more reproducible.~~\n    1. ~~Either confidence intervals or standard error (itself a form of confidence interval, I guess) should be included in the plots to communicate statistical significance to the reader.~~\n    1. Consider using either colourblind-friendly colours in the plots, or replacing the legend with labels for each line to remove the dependence on colour to determine which algorithm is which. Using differently-shaped points for each algorithm is a good start, but when the points are too close together it becomes difficult to tell which algorithm is which.\n    1. ~~What is the threshold for removing highly correlated features? This would help reproducibility.~~\n1. ~~The paper states that TD-SVRG and PD-SVRG converge linearly, but the y-axis of the plots appears to be in log space, which is confusing. Could this be clarified?~~\n1. ~~It would be good to have a concluding section that summarizes the main takeaways of the paper.~~\n\n**Grammatical/typographic issues:**\n1. \"Robbins & Munro (1951)\" should have the authors names inside the parentheses, because they are not being referred to in the sentence. Also, I think the second author's name is spelled \"Monro\".\n1. The \"Korda & LA (2014)\" citation is wrong: the authors are Nathaniel Korda and L.A. Prashanth, and the paper was published at ICML 2015.\n1. \"These methods are collectively known as variance reduction.\" Should this be, \"These methods are collectively known as variance-reduced gradient methods.\"?\n1. \"We analyze this case in 4 and 5\": It's better to write \"Sections 4 and 5\", because otherwise \"4 and 5\" could refer to an equation, appendix, theorem, etc.\n1. In Algorithm 1, should N be M? If not, it would be good to define N explicitly.\n1. \"unbalacedness\"\n1. \"is the the size\"\n1. \"on practice\" should be \"in practice\"\n1. \"envtironments\" in Figure 1.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper uses SVRG to reduce variance in TD learning. The resulting algorithm is analyzed using the gradient splitting perspective on TD learning to prove finite sample convergence rates that match those of SVRG in the setting of convex optimization. The analysis is done for several settings, and the findings are illustrated with experiments on several environments.", "strength_and_weaknesses": "**Strengths:**\n1. The analysis uses the gradient splitting approach, which yields a significantly simpler algorithm, analysis, learning rate/step size, and a better convergence rate.\n1. The paper considers several different settings, which increases the number of readers who would be interested.\n1. The paper provides recommendations for setting the learning rate, and actually uses this recommended learning rate in the experiments.\n\n**Potential Weaknesses/Questions:**\n1. Many grammatical errors, including typographic errors (\"TD-leaning\", \"apprxoimate\") and missing words, especially definite and indefinite articles. For example, \"determining *the* expected reward *an* agent will achieve if it chooses *actions* according to *a* stationary policy\". Some grammatical issues are listed below, but there were too many for me to write down all of them.\n1. ~~The paper doesn't comment on the assumptions made other than stating that they are standard. It would be nice to include a sentence or two explaining why the assumptions are reasonable or whether they are just for convenience (like Assumption 2; the feature vectors can always be normalized to make it true).~~\n1. Jumping between the notation used in other papers and the notation used in this paper is difficult. It would be better to convert the results of other papers to the notation used in this paper.\n1. ~~Font size of figures is too small to read without zooming in a lot.~~\n1. I have some concerns about the experiments:\n    1. ~~Why weren't the parameters for each algorithm set using a grid search the way they were for PD-SVRG? Without doing this, it's not clear that the chosen parameters are representative of the performance of each algorithm, and hence no conclusion about the performance of each algorithm can be drawn.~~\n    1. ~~Why was the learning rate for TD set to $1/\\sqrt{t}$? It seems like the algorithm stops learning very quickly due to the learning rate shrinking so quickly. How does the algorithm perform with a learning rate of $1/t$ instead?~~ I realized I was confused about this.\n    1. ~~What were the parameter values checked during the grid search? This would make the experiments more reproducible.~~\n    1. ~~Either confidence intervals or standard error (itself a form of confidence interval, I guess) should be included in the plots to communicate statistical significance to the reader.~~\n    1. Consider using either colourblind-friendly colours in the plots, or replacing the legend with labels for each line to remove the dependence on colour to determine which algorithm is which. Using differently-shaped points for each algorithm is a good start, but when the points are too close together it becomes difficult to tell which algorithm is which.\n    1. ~~What is the threshold for removing highly correlated features? This would help reproducibility.~~\n1. ~~The paper states that TD-SVRG and PD-SVRG converge linearly, but the y-axis of the plots appears to be in log space, which is confusing. Could this be clarified?~~\n1. ~~It would be good to have a concluding section that summarizes the main takeaways of the paper.~~\n\n**Grammatical/typographic issues:**\n1. \"Robbins & Munro (1951)\" should have the authors names inside the parentheses, because they are not being referred to in the sentence. Also, I think the second author's name is spelled \"Monro\".\n1. The \"Korda & LA (2014)\" citation is wrong: the authors are Nathaniel Korda and L.A. Prashanth, and the paper was published at ICML 2015.\n1. \"These methods are collectively known as variance reduction.\" Should this be, \"These methods are collectively known as variance-reduced gradient methods.\"?\n1. \"We analyze this case in 4 and 5\": It's better to write \"Sections 4 and 5\", because otherwise \"4 and 5\" could refer to an equation, appendix, theorem, etc.\n1. In Algorithm 1, should N be M? If not, it would be good to define N explicitly.\n1. \"unbalacedness\"\n1. \"is the the size\"\n1. \"on practice\" should be \"in practice\"\n1. \"envtironments\" in Figure 1.", "clarity,_quality,_novelty_and_reproducibility": "**Clarity:**\nThe paper is very clear conceptually, but there are quite a few grammatical mistakes that make it harder to read, including missing words and typographic errors. The discussion of results is very clear, and helpful explanations are provided throughout.\n\n**Quality:**\nI have some concerns about the experiments, detailed in the Strengths/Weaknesses section of this review.\n\n**Novelty:**\nTo the best of my knowledge, the analysis is novel.\n\n**Reproducibility:**\nThe paper is missing a few details that would be helpful for reproducing its experiments.", "summary_of_the_review": "~~Despite really liking this paper, I must recommend the current version be rejected due to concerns about clarity and the experiments detailed above. If the authors address these concerns satisfactorily and no critical issues are found by the other reviewers, I would recommend acceptance of a revised version.~~ The authors have addressed most of my concerns in their response, except some of my concerns about clarity (making these clarity changes would drastically reduce the work required of each reader to understand the paper). I have increased my score to reflect this.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666731598252}, {"id": "PDMAp1AjJK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper604/Reviewer_pncF"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper studies the Temporal difference (TD) learning method for policy evaluation in reinforcement learning. The proposed approach TD-SVRG method is a variant of the TD method by introducing the well-known SVRG technique. Theoretically, their analysis can lead to better convergence bounds for previous methods. Numerical results validate the improved performance of the proposed method over the existing methods.", "review_text": "This paper proposes a variant of the TD method with a faster convergence by introducing the well-known SVRG technique. The novelty and the improvements are limited.\n\n------Update------\n\nI have read the author's response and would like to keep the score.", "strengths": "Strengths:\n\n1: This paper proposes a variant of the TD method with a faster convergence by introducing the well-known SVRG technique.\n\n2: Extensive experiments are conducted to verify the effectiveness and efficiency of the proposed algorithm.\n\n\nWeaknesses:\n\n1: It is unclear whether the proposed method significantly improves the existing complexity in VRDT. For the overall complexity, which term will dominate the maximization function in theory or in practice? Since the superiority of the proposed method to the VRDT method depends on this, the authors would do well to provide more analysis. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper studies the Temporal difference (TD) learning method for policy evaluation in reinforcement learning. The proposed approach TD-SVRG method is a variant of the TD method by introducing the well-known SVRG technique. Theoretically, their analysis can lead to better convergence bounds for previous methods. Numerical results validate the improved performance of the proposed method over the existing methods.", "strength_and_weaknesses": "Strengths:\n\n1: This paper proposes a variant of the TD method with a faster convergence by introducing the well-known SVRG technique.\n\n2: Extensive experiments are conducted to verify the effectiveness and efficiency of the proposed algorithm.\n\n\nWeaknesses:\n\n1: It is unclear whether the proposed method significantly improves the existing complexity in VRDT. For the overall complexity, which term will dominate the maximization function in theory or in practice? Since the superiority of the proposed method to the VRDT method depends on this, the authors would do well to provide more analysis. ", "clarity,_quality,_novelty_and_reproducibility": "Novelty: This paper applied the well-known SVRG technique to the TD method to achieve faster convergence. It is new but seems not very interesting.\n\nQuality: The theoretical analysis is technically sound with excellent analysis. Experimental validation is provided. \n\nClarity: Some concepts in the paper is unclear. What is 'Gradient Splitting' in the title? \n\nReproducibility: The code needed to reproduce the experimental results is not provided.\n\n", "summary_of_the_review": "This paper proposes a variant of the TD method with a faster convergence by introducing the well-known SVRG technique. The novelty and the improvements are limited.\n\n------Update------\n\nI have read the author's response and would like to keep the score.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666690244333}, {"id": "0LJ76Op0VhG", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper604/Reviewer_Tp1f"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper provides a theoretical convergence proof for TD-learning algorithm with variance reduction structure. The sample complexity can match the complexity of SGD for convex optimization in iid finite-sample and iid online sampling senario. The empirical study furthor verifies the performance improvement from the variance reduction technique.", "review_text": "In summary, I will reject this paper because: (1) both algorithm and proof technique are not new; (2) the problem setting is not possible in RL.", "strengths": "***Strength***\nThis paper improves the sample complexity of an existing convergence result of the VRTD algorithm.\n\n***Weaknesses***\n1. In RL, online learning cannot have iid sampling. It can be a good starting point but the author should consider the Markovian sampling structure as [Xu2019].\n\n[Xu2019] Xu T, Wang Z, Zhou Y, Liang Y. Reanalysis of Variance Reduced Temporal Difference Learning. In International Conference on Learning Representations 2019 Sep 25.\n\n2. The theoretical result only obtains a constant-level complexity improvement. I don't think it fills any gap unless the theoretical lower bound is also provided; TD-learning is not equivalent to SGD in convex optimization. Also, there exists theoretical results showing that variance-reduction technique can make a variant of TD-learning algorithm achieve better complexity than $\\mathcal{O}(\\epsilon^{-1}\\log \\epsilon^{-1})$, see [Ma2020].\n\n[Ma2020] Ma S, Zhou Y, Zou S. Variance-reduced off-policy TDC learning: Non-asymptotic convergence analysis. Advances in Neural Information Processing Systems. 2020;33:14796-806.\n\n3. The gradient splitting method seems unrelated to this paper, even if it is put in the title. To my understanding, it doesn't bring any high-level perspectives to make the reader have a new understanding on TD-learning algorithm; it is just a trick used to derive the bound. Every optimization paper can have a lot of such tricks.   \n ", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper provides a theoretical convergence proof for TD-learning algorithm with variance reduction structure. The sample complexity can match the complexity of SGD for convex optimization in iid finite-sample and iid online sampling senario. The empirical study furthor verifies the performance improvement from the variance reduction technique.", "strength_and_weaknesses": "***Strength***\nThis paper improves the sample complexity of an existing convergence result of the VRTD algorithm.\n\n***Weaknesses***\n1. In RL, online learning cannot have iid sampling. It can be a good starting point but the author should consider the Markovian sampling structure as [Xu2019].\n\n[Xu2019] Xu T, Wang Z, Zhou Y, Liang Y. Reanalysis of Variance Reduced Temporal Difference Learning. In International Conference on Learning Representations 2019 Sep 25.\n\n2. The theoretical result only obtains a constant-level complexity improvement. I don't think it fills any gap unless the theoretical lower bound is also provided; TD-learning is not equivalent to SGD in convex optimization. Also, there exists theoretical results showing that variance-reduction technique can make a variant of TD-learning algorithm achieve better complexity than $\\mathcal{O}(\\epsilon^{-1}\\log \\epsilon^{-1})$, see [Ma2020].\n\n[Ma2020] Ma S, Zhou Y, Zou S. Variance-reduced off-policy TDC learning: Non-asymptotic convergence analysis. Advances in Neural Information Processing Systems. 2020;33:14796-806.\n\n3. The gradient splitting method seems unrelated to this paper, even if it is put in the title. To my understanding, it doesn't bring any high-level perspectives to make the reader have a new understanding on TD-learning algorithm; it is just a trick used to derive the bound. Every optimization paper can have a lot of such tricks.   \n ", "clarity,_quality,_novelty_and_reproducibility": "The quality of writting should be further improved. \n\nThe clarity is poor. Many necessary concepts are not introduced in this paper. How are the finite-sample and iid online sampling senario related to MDP? How do you split the gradient and what is the gradient for TD-learning algorithm? \n\nNo originality. Applying the variance reduction technique to TD-learning is not new. The proof technique is not new neither. The sample complexity improvement is not satisfactory (unless the theoretical lower bound is provided).  ", "summary_of_the_review": "In summary, I will reject this paper because: (1) both algorithm and proof technique are not new; (2) the problem setting is not possible in RL.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject"}, "tcdate": 1666290526140}, {"id": "4H84aqCXQK", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper604/Reviewer_kaRJ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper develops a reanalysis of TD with SVRG variance reduction and establishes a tight bound that matches the complexity result of SVRG in convex optimization. The developed analysis is simpler than all previous research and leads to better convergence bounds. Moreover, numerical experiments show the advantage of this algorithm over other existing variance-reduced TD algorithms.", "review_text": "See above.", "strengths": "Strength:\n\n-The presentation is clear and easy to follow\n\n-This paper is the first to prove that TD with SVRG can match the complexity result of SVRG in convex optimization.\n\nWeakness:\n\n-Paper writing is poor. There are many grammar issues, incomplete sentences, etc.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper develops a reanalysis of TD with SVRG variance reduction and establishes a tight bound that matches the complexity result of SVRG in convex optimization. The developed analysis is simpler than all previous research and leads to better convergence bounds. Moreover, numerical experiments show the advantage of this algorithm over other existing variance-reduced TD algorithms.", "strength_and_weaknesses": "Strength:\n\n-The presentation is clear and easy to follow\n\n-This paper is the first to prove that TD with SVRG can match the complexity result of SVRG in convex optimization.\n\nWeakness:\n\n-Paper writing is poor. There are many grammar issues, incomplete sentences, etc.\n", "clarity,_quality,_novelty_and_reproducibility": "The authors developed the key lemma 1, which allows them to establish an analysis analogous to the original analysis of SVRG in the convex setting. This simpler analysis is based on introducing the functions $f,\\omega$, and provides a better understanding of TD under variance reduction. I suggest the authors clarify the novelty of this lemma, i.e., how does this lemma differ from the existing works that derive similar type of bounds? ", "summary_of_the_review": "See above.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1665870784563}], "openreview_url": "https://openreview.net/forum?id=_AkC4QYxF5", "arxiv_id": "2211.16237", "paper_pdf": "papers/_AkC4QYxF5.pdf", "paper_pdf_sha256": "f73f4731b2cfd3962e1983dec269f68dfae7e7a21dd16cae2479c1ac8ec1355f", "paper_pdf_bytes": 552054, "paper_pdf_source": "openreview", "code_url": "https://github.com/gaarsmu/SVRG_for_TD_learning", "code_repository": "gaarsmu/SVRG_for_TD_learning", "code_commit": "0f8ddee6a59e541254bb1a80463825b10ce11762", "code_archive": "repos/_AkC4QYxF5.zip", "code_archive_sha256": "b145a42e75033fd203b05618fd6ad99f7373785f1dcafc72894402a85bb9e957", "code_archive_bytes": 555452, "code_file_count": 17, "code_extensions": {".py": 17}, "github_disk_usage_kb": 1279, "github_languages": {"Python": 55640}, "github_archived": false, "github_pushed_at": "2023-03-19T15:20:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/closing-the-gap-between-svrg-and-td-svrg-with"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "jxTRL-VOoQo", "year": 2022, "status": "rejected", "title": "Evaluating Deep Graph Neural Networks", "authors": ["Wentao Zhang", "Zeang Sheng", "Jiang Yuezihan", "Yikuan Xia", "Jun Gao", "Zhi Yang", "Bin CUI"], "authorids": ["~Wentao_Zhang1", "~Zeang_Sheng1", "~Jiang_Yuezihan1", "~Yikuan_Xia1", "~Jun_Gao6", "~Zhi_Yang4", "~Bin_CUI2"], "authors_source": "OpenReview API", "abstract": "Graph Neural Networks (GNNs) have already been widely applied in various graph mining tasks. However, most GNNs only have shallow architectures, which limits performance improvement. In this paper, we conduct a systematic experimental evaluation on the fundamental limitations of current architecture designs. Based on the experimental results, we answer the following two essential questions: (1) what actually leads to the compromised performance of deep GNNs; (2) how to build deep GNNs. The answers to the above questions provide empirical insights and guidelines for researchers to design deep GNNs. Further, we present Deep Graph Multi-Layer Perceptron (DGMLP), a powerful approach implementing our proposed guidelines. Experimental results demonstrate three advantages of DGMLP: 1) high accuracy -- it achieves state-of-the-art node classification performance on various datasets; 2) high flexibility -- it can flexibly choose different propagation and transformation depths according to certain graph properties; 3) high scalability and efficiency -- it supports fast training on large-scale graphs.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "WTJaVrtdSt-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper753/Reviewer_4Jnf"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work performs a systematic study to analyze the main issues of the difficulty in training deep GNNs by disentangling the effects of embedding propagation (EP) and embedding transformation (ET). They find that the large $D_t$ is the root cause for the failure of deep GNNs. Node-adaptive combination mechanism and residual connections between ET operations are proposed to train deep GNNs.", "review_text": "(+) The paper is well-organized and easy to follow. \n\n(+) This work performs a systematic study to analyze the main issues of the difficulty in training deep GNNs by disentangling the effects of embedding propagation (EP) and embedding transformation (ET). I enjoy reading the design of the analysis.\n\nConcerns:\n\n* The authors claim model degradation is the main cause of performance degradation of deep GNNs. However, what is model degradation is not defined formally.\n\n* It seems the proposed Node Smoothing Level is a measure of **smoothness** instead of **over-smoothness**. Claiming over-smoothing is not a major issue using Node Smoothing Level is not very well-justified. I think a better metric should be used. I recommend the authors check Group Distance Ratio and Instance Information Gain proposed in [1].\n\n* Why is the essential difference between the proposed “ResGCN” and “DenseGCN” in Figure 3 (b) and “ResGCN” and “DenseGCN” in [2].\n\n* What is the difference between the findings of adding skip connections to GNNs in [2] [3] and the findings in Section 5.3? This should be discussed.\n\n* It is unclear how the node-adaptive combination mechanism and residual connections between ET operations help with training deep GNNs. An ablation study needs to be carried out.\n\n* What is the reason that the performance of DGMLP on ogbn-product is far from the leading methods on the OGB leaderboard?\n\n* Why is the performance of baseline SIGN [4] is lower than the originally reported results (0.6568 ± 0.0006) on the leaderboard?\n\n* What are the benefits of disentangled graph convolution over entangled graph convolution rather than scalability?\n\n[1] Zhou, Kaixiong, et al. \"Towards Deeper Graph Neural Networks with Differentiable Group Normalization.\" NeurIPS 2020.\n\n[2] Li, Guohao, et al. \"Deepgcns: Can gcns go as deep as cnns?.\" ICCV 2019.\n\n[3] Li, Guohao, et al. \"Deepergcn: All you need to train deeper gcns.\" arXiv 2020.\n\n[4] Frasca, Fabrizio, et al. \"SIGN: Scalable Inception Graph Neural Networks.\" arXiv 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work performs a systematic study to analyze the main issues of the difficulty in training deep GNNs by disentangling the effects of embedding propagation (EP) and embedding transformation (ET). They find that the large $D_t$ is the root cause for the failure of deep GNNs. Node-adaptive combination mechanism and residual connections between ET operations are proposed to train deep GNNs.", "main_review": "(+) The paper is well-organized and easy to follow. \n\n(+) This work performs a systematic study to analyze the main issues of the difficulty in training deep GNNs by disentangling the effects of embedding propagation (EP) and embedding transformation (ET). I enjoy reading the design of the analysis.\n\nConcerns:\n\n* The authors claim model degradation is the main cause of performance degradation of deep GNNs. However, what is model degradation is not defined formally.\n\n* It seems the proposed Node Smoothing Level is a measure of **smoothness** instead of **over-smoothness**. Claiming over-smoothing is not a major issue using Node Smoothing Level is not very well-justified. I think a better metric should be used. I recommend the authors check Group Distance Ratio and Instance Information Gain proposed in [1].\n\n* Why is the essential difference between the proposed “ResGCN” and “DenseGCN” in Figure 3 (b) and “ResGCN” and “DenseGCN” in [2].\n\n* What is the difference between the findings of adding skip connections to GNNs in [2] [3] and the findings in Section 5.3? This should be discussed.\n\n* It is unclear how the node-adaptive combination mechanism and residual connections between ET operations help with training deep GNNs. An ablation study needs to be carried out.\n\n* What is the reason that the performance of DGMLP on ogbn-product is far from the leading methods on the OGB leaderboard?\n\n* Why is the performance of baseline SIGN [4] is lower than the originally reported results (0.6568 ± 0.0006) on the leaderboard?\n\n* What are the benefits of disentangled graph convolution over entangled graph convolution rather than scalability?\n\n[1] Zhou, Kaixiong, et al. \"Towards Deeper Graph Neural Networks with Differentiable Group Normalization.\" NeurIPS 2020.\n\n[2] Li, Guohao, et al. \"Deepgcns: Can gcns go as deep as cnns?.\" ICCV 2019.\n\n[3] Li, Guohao, et al. \"Deepergcn: All you need to train deeper gcns.\" arXiv 2020.\n\n[4] Frasca, Fabrizio, et al. \"SIGN: Scalable Inception Graph Neural Networks.\" arXiv 2020.", "summary_of_the_review": "(+) The systematic study of training deep GNNs of this work is interesting.\n\n(-) Some claims are not well-justified. Some related works are not discussed properly. Alations need to be done for the proposed methods. The performance of DGMLP is not very strong.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635901270818}, {"id": "5ZQa2VZWTKt", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper753/Reviewer_QtrD"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the application of GNNs to semi-supervised node classification tasks. The problem arises in big data problems such as modeling citation databases.  In this line of work it is assumed that the data follows a graph structure, i.e., that locally connected nodes are likely to have the same label. Each node has a d-dimensional feature vector, with graph degree matrix A. Various GNN architectures have been proposed and studied for this problem, often using citation databases as the driving application. The paper considers two aspects of this problem.  First, the authors consider various choices of the hyper-parameters, including the GNN iterations and a (typically) fully connected layer at the output.  It is well known that iterating the graph convolutions leads to local averaging that will asymptotically converge to some form of smoothing, which can result in poor classification.  Here, the authors decouple the GNN smoothing and the follow on network, and refer to this as “unentangled” processing architecture (also called decoupled in the numerical results portion).  They further propose a revised GNN architecture (DGMLP) that incorporates node-adaptive weighting and skip-connections in the second stage.  Numerical experiments complete the work, showing that with appropriate choice of parameters the DGMLP is flexible and can achieve the state of the art results, and good scalability.\n\n", "review_text": "The exploration of hyper-parameters for this problem is well founded, especially given the many different architectures that have been proposed.  The paper characterizes this through two primary parameters, Dp and Dt, and compares the DGMLP with varying values.  It isn’t clear why the authors don’t refer to these as hyper-parameters (except buried in the appendix). Perhaps it is not surprising that various papers have shown that the unentangled approach can outperform the entangled one.\n\nThe main body of the paper seems repetitive at times, stating repeatedly the basic points about the tradeoffs with Dp and Dt.  In the reviewers opinion, the meaningful portions of the paper are in the appendices, especially A and B.  It isn’t clear why these don’t make up the major portion of the main paper.  It seems that the paper spends a large amount of space with discussion and generality, often spending too much effort on criticism.  \n\nThe results show run-time and GPU memory versus graph size, and the DGMLP has good scaling properties.  \n\nThe node-adaptive weighting mechanism (Section 6.1) is interesting.  However, the influence here wasn’t entirely clear.\n\nThe paper would be significantly improved with a statement of the DGMLP in some kind of Table or Algorithm statement.\n\nThe DGMLP is able to achieve state of the art results compared with other algorithms.  So, the value here is the flexibility of the DGMLP, and also perhaps complexity reduction and scalability. It wasn’t entirely clear what is the overall complexity in comparison with other algorithms, especially with respect to choice of the hyper-parameters.  Perhaps the authors could add some further complexity results that include the cost of hyper-parameter exploration with the various algorithms to make this clear to the reader. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the application of GNNs to semi-supervised node classification tasks. The problem arises in big data problems such as modeling citation databases.  In this line of work it is assumed that the data follows a graph structure, i.e., that locally connected nodes are likely to have the same label. Each node has a d-dimensional feature vector, with graph degree matrix A. Various GNN architectures have been proposed and studied for this problem, often using citation databases as the driving application. The paper considers two aspects of this problem.  First, the authors consider various choices of the hyper-parameters, including the GNN iterations and a (typically) fully connected layer at the output.  It is well known that iterating the graph convolutions leads to local averaging that will asymptotically converge to some form of smoothing, which can result in poor classification.  Here, the authors decouple the GNN smoothing and the follow on network, and refer to this as “unentangled” processing architecture (also called decoupled in the numerical results portion).  They further propose a revised GNN architecture (DGMLP) that incorporates node-adaptive weighting and skip-connections in the second stage.  Numerical experiments complete the work, showing that with appropriate choice of parameters the DGMLP is flexible and can achieve the state of the art results, and good scalability.\n\n", "main_review": "The exploration of hyper-parameters for this problem is well founded, especially given the many different architectures that have been proposed.  The paper characterizes this through two primary parameters, Dp and Dt, and compares the DGMLP with varying values.  It isn’t clear why the authors don’t refer to these as hyper-parameters (except buried in the appendix). Perhaps it is not surprising that various papers have shown that the unentangled approach can outperform the entangled one.\n\nThe main body of the paper seems repetitive at times, stating repeatedly the basic points about the tradeoffs with Dp and Dt.  In the reviewers opinion, the meaningful portions of the paper are in the appendices, especially A and B.  It isn’t clear why these don’t make up the major portion of the main paper.  It seems that the paper spends a large amount of space with discussion and generality, often spending too much effort on criticism.  \n\nThe results show run-time and GPU memory versus graph size, and the DGMLP has good scaling properties.  \n\nThe node-adaptive weighting mechanism (Section 6.1) is interesting.  However, the influence here wasn’t entirely clear.\n\nThe paper would be significantly improved with a statement of the DGMLP in some kind of Table or Algorithm statement.\n\nThe DGMLP is able to achieve state of the art results compared with other algorithms.  So, the value here is the flexibility of the DGMLP, and also perhaps complexity reduction and scalability. It wasn’t entirely clear what is the overall complexity in comparison with other algorithms, especially with respect to choice of the hyper-parameters.  Perhaps the authors could add some further complexity results that include the cost of hyper-parameter exploration with the various algorithms to make this clear to the reader. \n", "summary_of_the_review": "The theme of the paper is “Hyper-Parameter Selection and a Modified GNN for Semi-Supervised Node Classification”.  Perhaps that’s a better title.  \n\nOverall the paper builds on the many other GNN works for this problem, incorporates the best known practices, and provides a flexible approach to achieving the state of the art results when compared to other algorithms.  However, the presentation could be significantly strengthened as outlined in the Main Review.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635887752097}, {"id": "qLkMc8YaZ1N", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper753/Reviewer_ddQY"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper points out the dominant influence on deep GNNs, i.e., the embedding transformation, with careful and extensive studies. Based on these studies, several guidelines are proposed to design deep model named DGMLP. ", "review_text": "**Strengths:**\n\n(1) The concepts of embedding propagation and transformation depths are formally defined to study their impacts on deep GNNs.\n(2) The series of empirical studies on these two depths are given.\n(3) The paper is well organized.\n\n\n**Weakness:**\n\n(1) The empirical guidelines 1-3, and the experimental analysis (in Section 4) are not novel at all to the community of deep GNNs. Specifically, the decoupling of embedding propagation and transformation, adaptive weighting, and skip connection have already been studied in literature. I know the authors derive these guidelines based on formulating the concepts of D_p & D_t and conducting the extensive experiments to get Findings 1. However, these knowledges have been implicitly explained in prior efforts. For example, the decoupling of EP and ET was proposed in DAGNN [1], based on the empirical study of SGC similar to the one in this paper. To avoid the harmful impact of embedding transformation, the identity mapping is incorporated in GCNII [4] to avoid the overly feature change. In other word, the dominant influence of embedding transformation in deep GNNs is not a news.\n\n(2) The proposed model, named DGMLP, is incremental comparing with the existing works, including DAGNN, SGC, JK-Net, etc. The only difference is the usage of skip connection in the final MLP module. \n\n(3) Although the experimental results show the superiority of DGMLP, most of the important baselines are missing in Table 1, which easily leads to misunderstanding. For example, GCNII has good or comparable performances in Cora, Citeseer, and ogbn-arxiv. In obgn-products, SAGN [5] and DeeperGCN [6] with the similar architecture to the proposed model shows outperforming results. Furthermore, SAGN also has better results in ogbn-100M. But all these important baselines are missing.\n\n(4) Some of the claims made in this paper is not correct. See below for the detailed comments.\n\n\n**Questions:**\n(1) Under Eq. (4), it is stated that ``Under this scenario, the neighborhood information is fully corrupted, resulting in catastrophic node classification accuracy.” I’m concerning on the correctness of this statement: why neighborhood information will be corrupted if the neighborhood influence is determined by the degree? The specific neighbor features could still distinguish them. As shown in Figure 4 of reference [1], by stacking many EP layers, the node classification accuracy is still good enough, which empirically demonstrates that neighbor information is not corrupted. \n\n(2) The claim of ``As the over-smoothing issue is only introduced by the EP operation rather than the ET operation” below Eq. (4) is not correct. As theoretically demonstrated in references [2,3], the over-smoothing issue is also correlated to the singular values of trainable weights and non-linear activations involved in ET. \n\n(3) For the example study of Figure 2(c), it is unclear how to formulate GCN model with D_t=2 and the adjacency matrix powered by D_p/2. Could you illustrate more with the formal expression or model figure?\n\n(4) In Section 4.2, please formally define the residual connection and dense connection used in the empirical study.\n\n(5) The guideline 2 seems not novel at all. It is well-known that the optimal convolution depths for the diverse nodes are different. For example, DAGNN in [1] applies attention module to learn the optimal combination weight for each specific node.\n\n[1] Liu, Meng, Hongyang Gao, and Shuiwang Ji. \"Towards deeper graph neural networks.\" Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020.\n\n[2] Cai, Chen, and Yusu Wang. \"A note on over-smoothing for graph neural networks.\" arXiv preprint arXiv:2006.13318 (2020).\n\n[3] Oono, Kenta, and Taiji Suzuki. \"Graph neural networks exponentially lose expressive power for node classification.\" arXiv preprint arXiv:1905.10947 (2019).\n\n[4] Chen, Ming, et al. \"Simple and deep graph convolutional networks.\" International Conference on Machine Learning. PMLR, 2020.\n\n[5] Sun, Chuxiong, and Guoshi Wu. \"Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training.\" arXiv preprint arXiv:2104.09376 (2021).\n\n[6] Li, Guohao, et al. \"Deepergcn: All you need to train deeper gcns.\" arXiv preprint arXiv:2006.07739 (2020).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper points out the dominant influence on deep GNNs, i.e., the embedding transformation, with careful and extensive studies. Based on these studies, several guidelines are proposed to design deep model named DGMLP. ", "main_review": "**Strengths:**\n\n(1) The concepts of embedding propagation and transformation depths are formally defined to study their impacts on deep GNNs.\n(2) The series of empirical studies on these two depths are given.\n(3) The paper is well organized.\n\n\n**Weakness:**\n\n(1) The empirical guidelines 1-3, and the experimental analysis (in Section 4) are not novel at all to the community of deep GNNs. Specifically, the decoupling of embedding propagation and transformation, adaptive weighting, and skip connection have already been studied in literature. I know the authors derive these guidelines based on formulating the concepts of D_p & D_t and conducting the extensive experiments to get Findings 1. However, these knowledges have been implicitly explained in prior efforts. For example, the decoupling of EP and ET was proposed in DAGNN [1], based on the empirical study of SGC similar to the one in this paper. To avoid the harmful impact of embedding transformation, the identity mapping is incorporated in GCNII [4] to avoid the overly feature change. In other word, the dominant influence of embedding transformation in deep GNNs is not a news.\n\n(2) The proposed model, named DGMLP, is incremental comparing with the existing works, including DAGNN, SGC, JK-Net, etc. The only difference is the usage of skip connection in the final MLP module. \n\n(3) Although the experimental results show the superiority of DGMLP, most of the important baselines are missing in Table 1, which easily leads to misunderstanding. For example, GCNII has good or comparable performances in Cora, Citeseer, and ogbn-arxiv. In obgn-products, SAGN [5] and DeeperGCN [6] with the similar architecture to the proposed model shows outperforming results. Furthermore, SAGN also has better results in ogbn-100M. But all these important baselines are missing.\n\n(4) Some of the claims made in this paper is not correct. See below for the detailed comments.\n\n\n**Questions:**\n(1) Under Eq. (4), it is stated that ``Under this scenario, the neighborhood information is fully corrupted, resulting in catastrophic node classification accuracy.” I’m concerning on the correctness of this statement: why neighborhood information will be corrupted if the neighborhood influence is determined by the degree? The specific neighbor features could still distinguish them. As shown in Figure 4 of reference [1], by stacking many EP layers, the node classification accuracy is still good enough, which empirically demonstrates that neighbor information is not corrupted. \n\n(2) The claim of ``As the over-smoothing issue is only introduced by the EP operation rather than the ET operation” below Eq. (4) is not correct. As theoretically demonstrated in references [2,3], the over-smoothing issue is also correlated to the singular values of trainable weights and non-linear activations involved in ET. \n\n(3) For the example study of Figure 2(c), it is unclear how to formulate GCN model with D_t=2 and the adjacency matrix powered by D_p/2. Could you illustrate more with the formal expression or model figure?\n\n(4) In Section 4.2, please formally define the residual connection and dense connection used in the empirical study.\n\n(5) The guideline 2 seems not novel at all. It is well-known that the optimal convolution depths for the diverse nodes are different. For example, DAGNN in [1] applies attention module to learn the optimal combination weight for each specific node.\n\n[1] Liu, Meng, Hongyang Gao, and Shuiwang Ji. \"Towards deeper graph neural networks.\" Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020.\n\n[2] Cai, Chen, and Yusu Wang. \"A note on over-smoothing for graph neural networks.\" arXiv preprint arXiv:2006.13318 (2020).\n\n[3] Oono, Kenta, and Taiji Suzuki. \"Graph neural networks exponentially lose expressive power for node classification.\" arXiv preprint arXiv:1905.10947 (2019).\n\n[4] Chen, Ming, et al. \"Simple and deep graph convolutional networks.\" International Conference on Machine Learning. PMLR, 2020.\n\n[5] Sun, Chuxiong, and Guoshi Wu. \"Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training.\" arXiv preprint arXiv:2104.09376 (2021).\n\n[6] Li, Guohao, et al. \"Deepergcn: All you need to train deeper gcns.\" arXiv preprint arXiv:2006.07739 (2020).", "summary_of_the_review": "This paper formally define the concepts of embedding propagation and transformation depths, and distinguish their impacts on the deep GNNs. Based on the empirical studies, some well-known guidelines are proposed to design deep GNNs. \n\nHowever, I cannot find much novelty from this paper, which seems to ensemble the existing knowledges. Some of important baselines are missing to validate the effectiveness of the proposed methods. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "Not applicable.", "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635808109848}, {"id": "ZZkHkhXAYDs", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper753/Reviewer_tpnS"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "In this work, the authors performed an experimental evaluation on several GNNs in order to understand what aspects of the current architecture designs that leads to the compromised performance of deep GNNs. They claimed to find the root causes: large propagation depth leads to the over-smoothing issue and large transformation depth leads to the model degradation issue. Then they proposed guidelines for designing deep GNNs and a Deep Graph Multi-Layer Perceptron (DGMLP) for implementing the guidelines. Experimental results demonstrate performance of DGMLP in accuracy, flexibility, scalability, and efficiency.", "review_text": "The definitions of Node Smoothing Level and Graph Smoothing Level compare the representations of the same node across different EP steps. However, the oversmoothing problem is across nodes. Considering the deviations, the performance of DGMLP appears to be comparable to the state-of-the-art. It is claimed that DGMLP can support large transformation depth D_t. But Fig6 (b) only shows a very stable accuracy as D_t increases. It would be more convincing if the performance of DGMLP improves with a larger D_t. Otherwise, why bother to increase D_t? Larger D_t means more parameters, and hence may need to train models longer (and with better training settings). Is it possible that large D_t leads to undertrain of GNN? The proposed node-adaptive weighting mechanism is interesting. It could helpful to see experiments that validate the importance of the node-adaptive weighting mechanism. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this work, the authors performed an experimental evaluation on several GNNs in order to understand what aspects of the current architecture designs that leads to the compromised performance of deep GNNs. They claimed to find the root causes: large propagation depth leads to the over-smoothing issue and large transformation depth leads to the model degradation issue. Then they proposed guidelines for designing deep GNNs and a Deep Graph Multi-Layer Perceptron (DGMLP) for implementing the guidelines. Experimental results demonstrate performance of DGMLP in accuracy, flexibility, scalability, and efficiency.", "main_review": "The definitions of Node Smoothing Level and Graph Smoothing Level compare the representations of the same node across different EP steps. However, the oversmoothing problem is across nodes. Considering the deviations, the performance of DGMLP appears to be comparable to the state-of-the-art. It is claimed that DGMLP can support large transformation depth D_t. But Fig6 (b) only shows a very stable accuracy as D_t increases. It would be more convincing if the performance of DGMLP improves with a larger D_t. Otherwise, why bother to increase D_t? Larger D_t means more parameters, and hence may need to train models longer (and with better training settings). Is it possible that large D_t leads to undertrain of GNN? The proposed node-adaptive weighting mechanism is interesting. It could helpful to see experiments that validate the importance of the node-adaptive weighting mechanism. ", "summary_of_the_review": "The experiments, which compare various GNNs, provide useful information. The proposed DGMLP contains interesting ideas. However, experimental results are mixed and do not clearly support all claims made in this manuscript. The proposed smoothing level metrics do not measure smoothing across nodes.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635479335587}], "openreview_url": "https://openreview.net/forum?id=jxTRL-VOoQo", "arxiv_id": "2108.00955", "paper_pdf": "papers/jxTRL-VOoQo.pdf", "paper_pdf_sha256": "5ec0879f24472d76bd525ec41441a0c1ed47d95c40cc32bb48ee784c5db14370", "paper_pdf_bytes": 788758, "paper_pdf_source": "openreview", "code_url": "https://github.com/zwt233/AIR", "code_repository": "zwt233/AIR", "code_commit": "d5cc2325b071ad062c7a7855f6a42336d21a7c3d", "code_archive": "repos/jxTRL-VOoQo.zip", "code_archive_sha256": "42fcf9644b138ac0ad9f6121d2aaa59e85dc982771548cf4b212c6a1e718f3f8", "code_archive_bytes": 5414890, "code_file_count": 14, "code_extensions": {".py": 14}, "github_disk_usage_kb": 6351, "github_languages": {"Python": 58516}, "github_archived": false, "github_pushed_at": "2022-09-26T03:30:59Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/evaluating-deep-graph-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "iEcqwosBEgx", "year": 2021, "status": "rejected", "title": "Novel Policy Seeking with Constrained Optimization", "authors": ["Hao Sun", "Zhenghao Peng", "Bo Dai", "Jian Guo", "Dahua Lin", "Bolei Zhou"], "authorids": ["~Hao_Sun3", "~Zhenghao_Peng1", "~Bo_Dai2", "guoj@pcl.ac.cn", "~Dahua_Lin1", "~Bolei_Zhou5"], "authors_source": "OpenReview API", "abstract": "We address the problem of seeking novel policies in reinforcement learning tasks. Instead of following the multi-objective framework commonly used in existing methods, we propose to rethink the problem under a novel perspective of constrained optimization. We at first introduce a new metric to evaluate the difference between policies, and then design two practical novel policy seeking methods following the new perspective, namely the Constrained Task Novel Bisector (CTNB), and the Interior Policy Differentiation (IPD), corresponding to the feasible direction method and the interior point method commonly known in the constrained optimization literature. Experimental comparisons on the MuJuCo control suite show our methods can achieve substantial improvements over previous novelty-seeking methods in terms of both the novelty of policies and their performances in the primal task.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "jOp2T9iGhp", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1214/AnonReviewer1"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper attempts to solve the problem of seeking novel policies in reinforcement learning from a constrained optimization perspective. This new perspective motives two new algorithms to solve the optimization problem, which are based on feasible direction and the interior point methods. The authors provide empirical results on several Mujoco benchmarks. \n\nDetails:\n\nThe idea of formulating the problem from a constrained optimization perspective is interesting. This new perspective motivates new and better algorithms to solve the optimization problem. \n\nHowever, I feel like the presentation is poor and the writing should be improved.  \n\nWhat’s the exact problem setting? The authors should clearly describe the problem setting before presenting the methods, even one paragraph would be helpful.  \n\nA lot of algorithm details are missing: \n\nQ1. $\\bar{D}^q_W (\\theta_i, \\theta_j)$ is a metric for any state distribution $q$. What’s the motivation of using $q = \\bar{\\rho}$? \n\nQ2. When computing the policy distance, what is $\\rho_{\\theta_i}$ in (4)? Is it the current policy, or a reference policy?\n\nQ3. I assume $\\theta_i$ is the current policy. According to (4), the algorithm uses $\\theta_i$ to get samples, and compute an importance correction ratio $q/\\rho_{\\theta_i}$ to approximate the distance. How is the $q(s)=\\bar{\\rho}(s)$ computed? The authors propose to approximate $\\rho_{\\theta}$ using monte-carlo methods. Does it mean the algorithm need to approximate $\\bar{\\rho}(s)$ using the reference policies for each $s\\sim \\rho_{\\theta_i}$? Is there a computation issue?\n\nQ4. This goes back to Q1. Why just using the on policy samples to estimate the distance? Is there any potential advantage to use $q = \\bar{\\rho}$? \n\nQ5. Learning the stationary distribution is a hard research problem itself. See recent work for example: \n\nZhang, R., Dai, B., Li, L. and Schuurmans, D., 2019, September. GenDICE: Generalized Offline Estimation of Stationary Values. In International Conference on Learning Representations.\n\nI agree the stationary distribution can be approximated using MC methods, but it might need a lot of samples as the variance is very high. This makes me wonder how is the algorithm implemented in practice, and how does the stationary distribution estimation subroutine affect the algorithm’s performance. \n\nOther suggestions:\n\nIf I understand correctly, this paper tries to solve the problem of finding a set of novel polices that solve a given task while exhibiting different behaviors. This seems also related to the exploration problem, as some works try to make the current policy different with previous policies to encourage exploration. See for example:\n\nHazan, E., Kakade, S., Singh, K. and Van Soest, A., 2019, May. Provably efficient maximum entropy exploration. In International Conference on Machine Learning (pp. 2681-2691). \n\nIt might be worth to discuss how the novel policy seeking problem is related to the exploration problem. \n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Novel Policy Seeking with Constrained Optimization", "review": "Summary:\n\nThis paper attempts to solve the problem of seeking novel policies in reinforcement learning from a constrained optimization perspective. This new perspective motives two new algorithms to solve the optimization problem, which are based on feasible direction and the interior point methods. The authors provide empirical results on several Mujoco benchmarks. \n\nDetails:\n\nThe idea of formulating the problem from a constrained optimization perspective is interesting. This new perspective motivates new and better algorithms to solve the optimization problem. \n\nHowever, I feel like the presentation is poor and the writing should be improved.  \n\nWhat’s the exact problem setting? The authors should clearly describe the problem setting before presenting the methods, even one paragraph would be helpful.  \n\nA lot of algorithm details are missing: \n\nQ1. $\\bar{D}^q_W (\\theta_i, \\theta_j)$ is a metric for any state distribution $q$. What’s the motivation of using $q = \\bar{\\rho}$? \n\nQ2. When computing the policy distance, what is $\\rho_{\\theta_i}$ in (4)? Is it the current policy, or a reference policy?\n\nQ3. I assume $\\theta_i$ is the current policy. According to (4), the algorithm uses $\\theta_i$ to get samples, and compute an importance correction ratio $q/\\rho_{\\theta_i}$ to approximate the distance. How is the $q(s)=\\bar{\\rho}(s)$ computed? The authors propose to approximate $\\rho_{\\theta}$ using monte-carlo methods. Does it mean the algorithm need to approximate $\\bar{\\rho}(s)$ using the reference policies for each $s\\sim \\rho_{\\theta_i}$? Is there a computation issue?\n\nQ4. This goes back to Q1. Why just using the on policy samples to estimate the distance? Is there any potential advantage to use $q = \\bar{\\rho}$? \n\nQ5. Learning the stationary distribution is a hard research problem itself. See recent work for example: \n\nZhang, R., Dai, B., Li, L. and Schuurmans, D., 2019, September. GenDICE: Generalized Offline Estimation of Stationary Values. In International Conference on Learning Representations.\n\nI agree the stationary distribution can be approximated using MC methods, but it might need a lot of samples as the variance is very high. This makes me wonder how is the algorithm implemented in practice, and how does the stationary distribution estimation subroutine affect the algorithm’s performance. \n\nOther suggestions:\n\nIf I understand correctly, this paper tries to solve the problem of finding a set of novel polices that solve a given task while exhibiting different behaviors. This seems also related to the exploration problem, as some works try to make the current policy different with previous policies to encourage exploration. See for example:\n\nHazan, E., Kakade, S., Singh, K. and Van Soest, A., 2019, May. Provably efficient maximum entropy exploration. In International Conference on Machine Learning (pp. 2681-2691). \n\nIt might be worth to discuss how the novel policy seeking problem is related to the exploration problem. \n\n", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604036668988}, {"id": "NLAfW2ioxlH", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1214/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary: This paper proposed a method to leverage the constrained optimization for policy training to learn diverse policies given some references. Based on a diversity metric defined on policy divergences, the paper employs two constrained optimization techniques for this problem with some modifications. Experiments on mujoco environments suggest that the proposed algorithms can beat existing diversity-driven policy optimization methods to learn both better and novel policies. Generally, the paper is well-written and easy to follow. Some concerns/comments:\n\n* The state distributions of proposed CTNB and IPD are different: In the CTNB method, the trajectories will keep rollout until they reach some termination conditions such as time limit or failure behavior. However, in the IPD method, if the cumulative novel reward is below some thresholds, then the trajectories will be truncated. It will be helpful to compare the CTNB with that extra termination condition. \n\n* Using the divergence of policies to quantify the difference between policies seems not a very innovative metric. Some related work could be:\n\nHong, Z. W., Shann, T. Y., Su, S. Y., Chang, Y. H., Fu, T. J., & Lee, C. Y. (2018). Diversity-driven exploration strategy for deep reinforcement learning.\n\nIt will be great if the authors can compare and explain the relationship between the proposed metric and some related ones.\n\n* The experiments can be more convincing if more locomotion environments are included, especially some higher-dimensional environments such as Humanoid and HumanoidStandup. Also, some other environments with a long-term/sparse reward setting can be more illustrative such as some mazes or Atari games. For some of those games, since it is stage-based, the IPD might terminate some rollouts if all reasonable policies are similar at the beginning of the trajectory. For a maze example, all good policies should choose to open the door at the beginning and then behave diversely.  \n\nOther/Minor Comments:\n\n* The choice of r_0 can affect the performance: When sequentially training the policy, should r_0 be adjusted when training each new policy?\n\n* It can be more interesting if some visualization of hopper policy diversity is included.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting method but can be more convincing", "review": "Summary: This paper proposed a method to leverage the constrained optimization for policy training to learn diverse policies given some references. Based on a diversity metric defined on policy divergences, the paper employs two constrained optimization techniques for this problem with some modifications. Experiments on mujoco environments suggest that the proposed algorithms can beat existing diversity-driven policy optimization methods to learn both better and novel policies. Generally, the paper is well-written and easy to follow. Some concerns/comments:\n\n* The state distributions of proposed CTNB and IPD are different: In the CTNB method, the trajectories will keep rollout until they reach some termination conditions such as time limit or failure behavior. However, in the IPD method, if the cumulative novel reward is below some thresholds, then the trajectories will be truncated. It will be helpful to compare the CTNB with that extra termination condition. \n\n* Using the divergence of policies to quantify the difference between policies seems not a very innovative metric. Some related work could be:\n\nHong, Z. W., Shann, T. Y., Su, S. Y., Chang, Y. H., Fu, T. J., & Lee, C. Y. (2018). Diversity-driven exploration strategy for deep reinforcement learning.\n\nIt will be great if the authors can compare and explain the relationship between the proposed metric and some related ones.\n\n* The experiments can be more convincing if more locomotion environments are included, especially some higher-dimensional environments such as Humanoid and HumanoidStandup. Also, some other environments with a long-term/sparse reward setting can be more illustrative such as some mazes or Atari games. For some of those games, since it is stage-based, the IPD might terminate some rollouts if all reasonable policies are similar at the beginning of the trajectory. For a maze example, all good policies should choose to open the door at the beginning and then behave diversely.  \n\nOther/Minor Comments:\n\n* The choice of r_0 can affect the performance: When sequentially training the policy, should r_0 be adjusted when training each new policy?\n\n* It can be more interesting if some visualization of hopper policy diversity is included.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603859502764}, {"id": "tMLMjipJE9", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1214/AnonReviewer3"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper aims at novel policy seeking which incorporates curiosity-driven exploration for better reinforcement learning. This paper first propose to use  a Wasserstein-based metric to calculate the difference between policies, and use it to define  the policy novelty. With these, the authors modeled the novel policy seeking as a constrained markov decision process(CMDP) and solved it using CTNB and IPD.\n\n1. This paper allows to consider the novelity issue dynamically.  However, when training policy according to the proposed CTNB or IPD, there should be some pretrained policies as perconditions, in other words, the proposed method needs some prior knowledge rather than learning policy from scratch. This may be a limitation for its application.\n\n2. About the proposition 2, the single trajectory estimation is unbiased, however, the variance seems to be large, the influence about the estimation variance should be considered.\n\n3.  in fomula (5) and (6), is $r_{int, t}$ equal to $r_{int}$ ? If is,  why use t? and what does moving average mean since there are several kinds of moving averages?  \n\n4. in fomula (4) and (6), Are Ts the same?\n\n5. Fig.2  shows that in Waklker2d and HalfCheetah, the proposed CTNB has less novel than PPO, which doesn't match the purpose of CTNB.\n\n6. It seems not easy to tune the novelty threshold for different task, as it performs different on different tasks. Can the author provide some insight on how to tune this.\n\n7. Five random seeds is not sufficient for experiments. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Seeking novel policy only after finding a good one ", "review": "This paper aims at novel policy seeking which incorporates curiosity-driven exploration for better reinforcement learning. This paper first propose to use  a Wasserstein-based metric to calculate the difference between policies, and use it to define  the policy novelty. With these, the authors modeled the novel policy seeking as a constrained markov decision process(CMDP) and solved it using CTNB and IPD.\n\n1. This paper allows to consider the novelity issue dynamically.  However, when training policy according to the proposed CTNB or IPD, there should be some pretrained policies as perconditions, in other words, the proposed method needs some prior knowledge rather than learning policy from scratch. This may be a limitation for its application.\n\n2. About the proposition 2, the single trajectory estimation is unbiased, however, the variance seems to be large, the influence about the estimation variance should be considered.\n\n3.  in fomula (5) and (6), is $r_{int, t}$ equal to $r_{int}$ ? If is,  why use t? and what does moving average mean since there are several kinds of moving averages?  \n\n4. in fomula (4) and (6), Are Ts the same?\n\n5. Fig.2  shows that in Waklker2d and HalfCheetah, the proposed CTNB has less novel than PPO, which doesn't match the purpose of CTNB.\n\n6. It seems not easy to tune the novelty threshold for different task, as it performs different on different tasks. Can the author provide some insight on how to tune this.\n\n7. Five random seeds is not sufficient for experiments. ", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603855433232}, {"id": "DHGXlNfDicD", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1214/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a novel constrained optimization based method, to optimize the expected return as well as encourage novelty of a new policy in contrast to existing policies. By modeling the problem as a constrained optimization problem, they can avoid excessive novelty seeking effectively, which is common in existing methods which model the problem with multi-objective optimization. \n\nTo be specific, they first propose a novel metric to measure the novelty of a new policy. To estimate such metric on sampled state with dense online reward, they propose an importance-based estimator for the proposed metric. With the estimation of the novelty metric, they propose to formulate the problem as a constrained optimization problem. The novelty is constrained to be larger than certain threshold r_0. In this way, the algorithm will only encourage larger novelty when the novelty is less than r_0, therefore avoiding excessive novelty seeking which may hurt the performance. They improve TNB proposed in (Zhang et al., 2019) with CTNB, where the ∇θg term exists only when the constraint is violated. They also propose another method based on Interior Point Method. Since IPM is computationally expensive and numerically unstable, they made an adaptation to RL setting, by bounding the collected transitions in the feasible region. \n\nOverall, the method is intuitive and reasonable. I have the following questions:\n\n1. The first contribution of this paper, is proposing a novel metric to measure the novelty of current policy in contrast to existing policies. Why propose a novel metric? Is existing metric for measuring the novelty not good? If so, can you verify your claim in experiments? \n\n2. The hyper-parameter r_0.  From Figure 3, we can see the different performance under different novelty thresholds r_0. The algorithm seems to be sensitive to r_0, which is of course reasonable. How did you choose r_0 for different environments? Did you consider a soft r_0 rather than a hard constraint(that is, maybe the constraint has different weight for different r_0)?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Official Blind Review #4", "review": "This paper proposes a novel constrained optimization based method, to optimize the expected return as well as encourage novelty of a new policy in contrast to existing policies. By modeling the problem as a constrained optimization problem, they can avoid excessive novelty seeking effectively, which is common in existing methods which model the problem with multi-objective optimization. \n\nTo be specific, they first propose a novel metric to measure the novelty of a new policy. To estimate such metric on sampled state with dense online reward, they propose an importance-based estimator for the proposed metric. With the estimation of the novelty metric, they propose to formulate the problem as a constrained optimization problem. The novelty is constrained to be larger than certain threshold r_0. In this way, the algorithm will only encourage larger novelty when the novelty is less than r_0, therefore avoiding excessive novelty seeking which may hurt the performance. They improve TNB proposed in (Zhang et al., 2019) with CTNB, where the ∇θg term exists only when the constraint is violated. They also propose another method based on Interior Point Method. Since IPM is computationally expensive and numerically unstable, they made an adaptation to RL setting, by bounding the collected transitions in the feasible region. \n\nOverall, the method is intuitive and reasonable. I have the following questions:\n\n1. The first contribution of this paper, is proposing a novel metric to measure the novelty of current policy in contrast to existing policies. Why propose a novel metric? Is existing metric for measuring the novelty not good? If so, can you verify your claim in experiments? \n\n2. The hyper-parameter r_0.  From Figure 3, we can see the different performance under different novelty thresholds r_0. The algorithm seems to be sensitive to r_0, which is of course reasonable. How did you choose r_0 for different environments? Did you consider a soft r_0 rather than a hard constraint(that is, maybe the constraint has different weight for different r_0)?", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603782184165}], "openreview_url": "https://openreview.net/forum?id=iEcqwosBEgx", "arxiv_id": "2005.10696", "paper_pdf": "papers/iEcqwosBEgx.pdf", "paper_pdf_sha256": "6ea4d66dde148372a8c31d0edd502732266d6e6adc5568bc4a7476c321853d8b", "paper_pdf_bytes": 5812241, "paper_pdf_source": "openreview", "code_url": "https://github.com/holarissun/NPSCO", "code_repository": "holarissun/NPSCO", "code_commit": "28106089d65e392746ece741a735510c96eb65f9", "code_archive": "repos/iEcqwosBEgx.zip", "code_archive_sha256": "f61b43633dd0461c9accf1268ddbc5a0f8024a5d4062c5d4be330fc3423ba2c9", "code_archive_bytes": 4463029, "code_file_count": 5, "code_extensions": {".py": 5}, "github_disk_usage_kb": 4357, "github_languages": {"Python": 119589}, "github_archived": false, "github_pushed_at": "2020-08-11T03:09:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/novel-policy-seeking-with-constrained"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "I1VCj1l1Zn", "year": 2025, "status": "rejected", "title": "DLP-LoRA: Efficient Task-Specific LoRA Fusion with a Dynamic, Lightweight Plugin for Large Language Models", "authors": ["Yuxuan Zhang", "Ruizhe Li"], "authorids": ["~Yuxuan_Zhang15", "~Ruizhe_Li2"], "authors_source": "OpenReview API", "abstract": "Recent advancements in Large Language Models (LLMs) have achieved robust performance across diverse tasks, but fine-tuning these models for specific domains remains resource-intensive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) address this challenge by fine-tuning a small subset of parameters. However, existing methods for fusing multiple LoRAs lack dynamic fusion based on contextual inputs and often increase inference time due to token-level operations. We propose DLP-LoRA, a Dynamic Lightweight Plugin that employs a mini-MLP module with only 5M parameters to dynamically fuse multiple LoRAs at the sentence level using top-$p$ sampling strategies. This approach reduces inference time to less than twice that of single LoRA inference by leveraging parallel computation. Evaluations across 26 tasks—including multiple-choice questions and question answering—demonstrate that DLP-LoRA achieves an average accuracy of 92.34\\% on multiple-choice datasets and significant improvements in BLEU and ROUGE scores on QA datasets, outperforming different LLMs backbones under composite task settings. DLP-LoRA effectively balances performance and efficiency, making it a practical solution for dynamic multi-task adaptation in LLMs.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "k5SmgUEOkS", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13516/Reviewer_4fCq"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper proposes a method to dynamically fuse pre-trained task-specific LoRA modules. The idea is to train a sentence-level router for different tasks, and simply use that for routing and fusing different LoRA modules. The method is evaluated on both classification and language generation tasks such as QA.", "review_text": "This paper proposes a method to dynamically fuse pre-trained task-specific LoRA modules. The idea is to train a sentence-level router for different tasks, and simply use that for routing and fusing different LoRA modules. The method is evaluated on both classification and language generation tasks such as QA.", "strengths": "1. the paper is addressing an important problem in efficiently re-using expert modules. The proposed method is more efficient since it uses sentence-level routing, and it is evaluated on various classification and language generation tasks.\n2. the paper also includes discussions on inference efficiency, which is very relevant to the practical usage of the method.", "weaknesses": "1. the paper lacks important baselines. the routing module is trained jointly on N tasks, so the method assumes that one has access to all the task-specific training data for each LoRA module. Therefore, the authors should compare to a joint multitask training LoRA setting where the LoRA module is optimized on all task data. Note that the LoRA for joint multitask training should have a larger rank so that the number of tunable parameters is equivalent to tuning individual LoRA modules for each task. \n2. the results in table 1 shows that the proposed method is worse than baselines for classification, but it's better for generation tasks in table 2. However, there is not enough explanation or intuition on why that's the case.\n3. there are some highly relevant work on compositional LoRA with sentence-level routing that's not cited: https://aclanthology.org/2023.eacl-main.49.pdf, https://arxiv.org/abs/2402.17934", "questions": "1. how does your method compares to tuning a LoRA module with a higher rank on all multitask training data?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a method to dynamically fuse pre-trained task-specific LoRA modules. The idea is to train a sentence-level router for different tasks, and simply use that for routing and fusing different LoRA modules. The method is evaluated on both classification and language generation tasks such as QA.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "1. the paper is addressing an important problem in efficiently re-using expert modules. The proposed method is more efficient since it uses sentence-level routing, and it is evaluated on various classification and language generation tasks.\n2. the paper also includes discussions on inference efficiency, which is very relevant to the practical usage of the method.", "weaknesses": "1. the paper lacks important baselines. the routing module is trained jointly on N tasks, so the method assumes that one has access to all the task-specific training data for each LoRA module. Therefore, the authors should compare to a joint multitask training LoRA setting where the LoRA module is optimized on all task data. Note that the LoRA for joint multitask training should have a larger rank so that the number of tunable parameters is equivalent to tuning individual LoRA modules for each task. \n2. the results in table 1 shows that the proposed method is worse than baselines for classification, but it's better for generation tasks in table 2. However, there is not enough explanation or intuition on why that's the case.\n3. there are some highly relevant work on compositional LoRA with sentence-level routing that's not cited: https://aclanthology.org/2023.eacl-main.49.pdf, https://arxiv.org/abs/2402.17934", "questions": "1. how does your method compares to tuning a LoRA module with a higher rank on all multitask training data?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730651368218}, {"id": "F3mydHBXI9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13516/Reviewer_ohza"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "This paper focuses on achieving a more efficient dynamic fusion of multiple LoRA experts in multi-task adaptation for LLMs. To this end, the authors propose DLP LoRA, which operates as a sentence-level Mixture-of-Experts MoE gating mechanism. Specifically, a pre-trained task classifier is used to obtain the probability distribution of the input sentence across various tasks, and then the gating mechanism fuses selected LoRAs filtered by a fixed probability threshold. Additionally, the paper presents a parallel CUDA acceleration strategy to improve inference efficiency. Experiments are conducted on 26 tasks adaptation, achieving performance comparable to single-task LoRA.", "review_text": "This paper focuses on achieving a more efficient dynamic fusion of multiple LoRA experts in multi-task adaptation for LLMs. To this end, the authors propose DLP LoRA, which operates as a sentence-level Mixture-of-Experts MoE gating mechanism. Specifically, a pre-trained task classifier is used to obtain the probability distribution of the input sentence across various tasks, and then the gating mechanism fuses selected LoRAs filtered by a fixed probability threshold. Additionally, the paper presents a parallel CUDA acceleration strategy to improve inference efficiency. Experiments are conducted on 26 tasks adaptation, achieving performance comparable to single-task LoRA.", "strengths": "1. The paper provides a deeper exploration into previous methods for dynamically fusing multiple LoRA experts and attempts incremental improvements.\n\n2. The topic investigated in this paper has considerable application value.", "weaknesses": "1. The proposed DLP-LoRA is largely incremental and lacks sufficient novelty.\n\n2. The experimental evaluation is unconvincing: 1）The baselines in Tables 1, 2, and 3 are weak and insufficient, and the authors should consider introducing relevant methods from this field for comparison; 2）The paper emphasizes the proposed methodʼs efficiency advantages, yet lacks quantitative experiments on efficiency compared to single LoRA inference and previous methods.\n\n3. There are some issues in the writing that reflect a lack of rigor: 1）The paper incorrectly refers to top-p sampling; to my knowledge, top-p sampling restricts the candidate set based on cumulative probability thresholds, whereas the paper independently filters LoRA experts below a fixed probability threshold, which is inconsistent; 2）The introduction claims that previous methods require additional fine-tuning when tasks change, yet DLP-LoRA seems unable to solve this problem either.", "questions": "1. The paper repeatedly emphasizes that the proposed method achieves less than twice the inference time of single LoRA inference; Is there quantitative experimental data to support this claim?\n\n2. Why does Table 5 compare the LLaMA-2 13B with a smaller, fine-tuned LLM? From an inference speed perspective, a smaller LLM outperforming a larger LLM is expected; from a performance perspective, a fine-tuned model outperforming an un-fine-tuned model is also expected. Is this a fair comparison?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on achieving a more efficient dynamic fusion of multiple LoRA experts in multi-task adaptation for LLMs. To this end, the authors propose DLP LoRA, which operates as a sentence-level Mixture-of-Experts MoE gating mechanism. Specifically, a pre-trained task classifier is used to obtain the probability distribution of the input sentence across various tasks, and then the gating mechanism fuses selected LoRAs filtered by a fixed probability threshold. Additionally, the paper presents a parallel CUDA acceleration strategy to improve inference efficiency. Experiments are conducted on 26 tasks adaptation, achieving performance comparable to single-task LoRA.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. The paper provides a deeper exploration into previous methods for dynamically fusing multiple LoRA experts and attempts incremental improvements.\n\n2. The topic investigated in this paper has considerable application value.", "weaknesses": "1. The proposed DLP-LoRA is largely incremental and lacks sufficient novelty.\n\n2. The experimental evaluation is unconvincing: 1）The baselines in Tables 1, 2, and 3 are weak and insufficient, and the authors should consider introducing relevant methods from this field for comparison; 2）The paper emphasizes the proposed methodʼs efficiency advantages, yet lacks quantitative experiments on efficiency compared to single LoRA inference and previous methods.\n\n3. There are some issues in the writing that reflect a lack of rigor: 1）The paper incorrectly refers to top-p sampling; to my knowledge, top-p sampling restricts the candidate set based on cumulative probability thresholds, whereas the paper independently filters LoRA experts below a fixed probability threshold, which is inconsistent; 2）The introduction claims that previous methods require additional fine-tuning when tasks change, yet DLP-LoRA seems unable to solve this problem either.", "questions": "1. The paper repeatedly emphasizes that the proposed method achieves less than twice the inference time of single LoRA inference; Is there quantitative experimental data to support this claim?\n\n2. Why does Table 5 compare the LLaMA-2 13B with a smaller, fine-tuned LLM? From an inference speed perspective, a smaller LLM outperforming a larger LLM is expected; from a performance perspective, a fine-tuned model outperforming an un-fine-tuned model is also expected. Is this a fair comparison?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730294394106}, {"id": "joqmjG1f6X", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13516/Reviewer_fMGx"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "- The study proposes a dynamic lightweight plugin to fuse multiple LoRA adapters that enable the adaptation of language models to specific downstream tasks and new data domains.\n- The proposed approach involves a “mini-MLP plugin” that combines the weights of multiple LoRA modules based on the contextual information provided by the language model input.\n- The authors DLP method aims to improve efficiency over token-level LoRA mixture approaches by utilizing sentence-level representations. The performance benchmark indicates that their DLP-LoRA method achieves similar performance to single task-specific LoRA modules.", "review_text": "- The study proposes a dynamic lightweight plugin to fuse multiple LoRA adapters that enable the adaptation of language models to specific downstream tasks and new data domains.\n- The proposed approach involves a “mini-MLP plugin” that combines the weights of multiple LoRA modules based on the contextual information provided by the language model input.\n- The authors DLP method aims to improve efficiency over token-level LoRA mixture approaches by utilizing sentence-level representations. The performance benchmark indicates that their DLP-LoRA method achieves similar performance to single task-specific LoRA modules.", "strengths": "- The evaluation across 26 diverse tasks including multiple choice questions, question answering, summarization, translation is quite extensive. They include ablation studies evaluating inference time and contrast model size and performance results.\n- The authors clearly describe the benefits of their proposed DLP-LoRA method.\n- the performance evaluation includes multiple medium-sized decoder-only language models with Llama and Qwen", "weaknesses": "- The differences to related methods are not clear or not specifically evaluated in regards to their performance or efficiency contribution. How does the approach differ from methodologically from i.e. PHATGOOSE [1], or LoRAMoE [2] and how does the performance differ?\n- The authors list various mixture of expert methods in their related works and introduction, but only include single LoRAs and the base model in their benchmark results.\n- Some claims remain unsupported by the empirical results presented in the study. The authors mention that additional fine-tuning for new tasks is not required in their DLP framework, yet there is no performance benchmark on unseen tasks. Furthermore, they describe the framework as lightweight but lack the parameter and space complexity comparison with related LoRA mixture approaches.\n- The description of their proposed method lacks details regarding the training of the LoRA modules that are combined to perform the downstream tasks.\n\nReferences:\n- [1] Muqeeth, M., Liu, H., Liu, Y., & Raffel, C. (2024). Learning to Route Among Specialized Experts for Zero-Shot Generalization. arXiv [Cs.LG].\n- [2] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, and Xuanjing Huang. 2024. LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1932–1945, Bangkok, Thailand. Association for Computational Linguistics.", "questions": "- Are the single LoRAs included in the selection of different LoRA modules DLP is combining?\n- Can you elaborate why the performance of multiple LoRA modules combined through your DLP method, which include the single task-specific LoRA module, is similar or worse than the single-LoRA module?\n- The inference time comparisons in Table 4 are not explained in the caption. Metric and reference time is missing.\n- Section 5 lists selecting adapters based on their relevance vs. fixed k as a key limitation of related work. Do you have empirical results that support this claim for your DLP method?\n- Can you elaborate on the differences between the base model performances in Table 1? Specifically, Llama 3 8b seems to perform significantly better than Llama 2, and both Qwen models.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "- The study proposes a dynamic lightweight plugin to fuse multiple LoRA adapters that enable the adaptation of language models to specific downstream tasks and new data domains.\n- The proposed approach involves a “mini-MLP plugin” that combines the weights of multiple LoRA modules based on the contextual information provided by the language model input.\n- The authors DLP method aims to improve efficiency over token-level LoRA mixture approaches by utilizing sentence-level representations. The performance benchmark indicates that their DLP-LoRA method achieves similar performance to single task-specific LoRA modules.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "- The evaluation across 26 diverse tasks including multiple choice questions, question answering, summarization, translation is quite extensive. They include ablation studies evaluating inference time and contrast model size and performance results.\n- The authors clearly describe the benefits of their proposed DLP-LoRA method.\n- the performance evaluation includes multiple medium-sized decoder-only language models with Llama and Qwen", "weaknesses": "- The differences to related methods are not clear or not specifically evaluated in regards to their performance or efficiency contribution. How does the approach differ from methodologically from i.e. PHATGOOSE [1], or LoRAMoE [2] and how does the performance differ?\n- The authors list various mixture of expert methods in their related works and introduction, but only include single LoRAs and the base model in their benchmark results.\n- Some claims remain unsupported by the empirical results presented in the study. The authors mention that additional fine-tuning for new tasks is not required in their DLP framework, yet there is no performance benchmark on unseen tasks. Furthermore, they describe the framework as lightweight but lack the parameter and space complexity comparison with related LoRA mixture approaches.\n- The description of their proposed method lacks details regarding the training of the LoRA modules that are combined to perform the downstream tasks.\n\nReferences:\n- [1] Muqeeth, M., Liu, H., Liu, Y., & Raffel, C. (2024). Learning to Route Among Specialized Experts for Zero-Shot Generalization. arXiv [Cs.LG].\n- [2] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, and Xuanjing Huang. 2024. LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1932–1945, Bangkok, Thailand. Association for Computational Linguistics.", "questions": "- Are the single LoRAs included in the selection of different LoRA modules DLP is combining?\n- Can you elaborate why the performance of multiple LoRA modules combined through your DLP method, which include the single task-specific LoRA module, is similar or worse than the single-LoRA module?\n- The inference time comparisons in Table 4 are not explained in the caption. Metric and reference time is missing.\n- Section 5 lists selecting adapters based on their relevance vs. fixed k as a key limitation of related work. Do you have empirical results that support this claim for your DLP method?\n- Can you elaborate on the differences between the base model performances in Table 1? Specifically, Llama 3 8b seems to perform significantly better than Llama 2, and both Qwen models.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1729531195425}], "openreview_url": "https://openreview.net/forum?id=I1VCj1l1Zn", "arxiv_id": "2410.01497", "paper_pdf": "papers/I1VCj1l1Zn.pdf", "paper_pdf_sha256": "8323c066f332b87ab0dc5e6680565091367367241c598687e772f2fe98e167a6", "paper_pdf_bytes": 2722444, "paper_pdf_source": "openreview", "code_url": "https://github.com/MeCuping/DLP-LoRA", "code_repository": "MeCuping/DLP-LoRA", "code_commit": "d68fd38f36d9762f3e60041472193d379257cfb9", "code_archive": "repos/I1VCj1l1Zn.zip", "code_archive_sha256": "d7af434a61dce55ca98f67239b40ba01a2c39a8f2e4f1b72d8ac136ac51f5a84", "code_archive_bytes": 462891, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 426, "github_languages": {"Python": 62697, "Batchfile": 28}, "github_archived": false, "github_pushed_at": "2024-10-03T01:48:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/dlp-lora-efficient-task-specific-lora-fusion"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "96nX9xIIx2", "year": 2024, "status": "rejected", "title": "Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective", "authors": ["Can Jin", "Tianjin Huang", "Yihua Zhang", "Mykola Pechenizkiy", "Sijia Liu", "Shiwei Liu", "Tianlong Chen"], "authorids": ["~Can_Jin1", "~Tianjin_Huang1", "~Yihua_Zhang1", "~Mykola_Pechenizkiy1", "~Sijia_Liu1", "~Shiwei_Liu2", "~Tianlong_Chen1"], "authors_source": "OpenReview API", "abstract": "The rapid development of large-scale deep learning models questions the affordability of hardware platforms, which necessitates the pruning to reduce their computational and memory footprints. Sparse neural networks as the product, have demonstrated numerous favorable benefits like low complexity, undamaged generalization, $\\textit{etc}$. Most of the prominent pruning strategies are invented from a $\\textit{model-centric}$ perspective, focusing on searching and preserving crucial weights by analyzing network topologies. However, the role of data and its interplay with model-centric pruning has remained relatively unexplored. In this research, we introduce a novel $\\textit{data-model co-design}$ perspective: to promote superior weight sparsity by learning important model topology and adequate input data in a synergetic manner. Specifically, customized $\\textbf{V}$isual $\\textbf{P}$rompts are mounted to upgrade neural $\\textbf{N}$etwork $\\textbf{s}$parsification in our proposed $\\textbf{\\texttt{VPNs}}$ framework. As a pioneering effort, this paper conducts systematic investigations about the impact of different visual prompts on model pruning and suggests an effective joint optimization approach. Extensive experiments with $3$ network architectures and $8$ datasets evidence the substantial performance improvements from $\\textbf{\\texttt{VPNs}}$ over existing start-of-the-art pruning algorithms. Furthermore, we find that subnetworks discovered by $\\textbf{\\texttt{VPNs}}$ from pre-trained models enjoy better transferability across diverse downstream scenarios. These insights shed light on new promising possibilities of data-model co-designs for vision model sparsification. Codes are in the supplement.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "izPKUiwMba", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2675/Reviewer_8b5d"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a new post-pruning method by introducing visual prompts into the pruning pipeline. The authors introduce visual prompts into trained vision models. During pruning, not only weight masks but also visual prompts are optimized. The whole pipeline includes two steps: (1) fixing pretrained weights of vision models, tuning masks and visual prompts; (2) fixing the mask, fine-tuning both weights and visual prompts. The authors adopt the proposed method on several downstream classification tasks and compare the performances of pruned models with other pruning methods. The experiments show that the proposed method can achieve better performance than other methods with the same sparsity.", "review_text": "This paper proposes a new post-pruning method by introducing visual prompts into the pruning pipeline. The authors introduce visual prompts into trained vision models. During pruning, not only weight masks but also visual prompts are optimized. The whole pipeline includes two steps: (1) fixing pretrained weights of vision models, tuning masks and visual prompts; (2) fixing the mask, fine-tuning both weights and visual prompts. The authors adopt the proposed method on several downstream classification tasks and compare the performances of pruned models with other pruning methods. The experiments show that the proposed method can achieve better performance than other methods with the same sparsity.", "strengths": "1. The writing and presentation of this paper are quite good. The logic of the article is very clear, and the choice of words in the writing is also very precise.\n2. The idea of introducing tunable visual prompts into the pruning pipeline is intriguing. The experiments validate the effectiveness of the proposed strategy.\n3. The authors compare the proposed method with other pruning methods and additionally apply the proposed visual prompt pruning strategy to these methods, demonstrating the transferability of their approach.", "weaknesses": "1. Introducing visual prompts into vision models seems to boost their performance. However, comparing models with visual prompts (the proposed method) to those without (other baseline methods) might not be entirely fair. What if we apply both the proposed and baseline methods to a model that has already been fine-tuned with visual prompts?\n2. I have some doubts regarding the generality and performance of this paper.\n(1) Why must we conduct experiments on downstream tasks of ImageNet? Why not directly on ImageNet itself, since most pruning work actually focuses more on performance on ImageNet?\n(2) The method proposed in this paper seems to be limited to scenarios where visual prompts can be applied, with their main application currently being in classification tasks. How can the proposed approach be used for other tasks, such as detection, segmentation, etc.?", "questions": "Please refer to the weaknesses. I hope the authors can provide more experiments to demonstrate the effectiveness of the method.\n\nThe provided additional experiments have addressed my concerns about the performance of VPN in ImageNet and object detection tasks. However, after reading other reviewers' comments, I consider decreasing my score to 5. The reasons are as follows:\n(1) The author describe the advantagement of unstructured sparsity on nonGPU hardware. However, they do not report the latency of unstructured sparsity on CPU. I think conducting experiments on some lightweight networks and reporting latency tested on CPUs can make the results more convincing.\n(2) In regards to structured pruning, the experiments carried out on the CIFAR-100 dataset are insufficient.\n\nOverall, I believe this paper needs more meticulous refinement in its experiments. If it can validate its approach on large-scale datasets for both unstructured and structured pruning settings, it would then be a very solid paper.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new post-pruning method by introducing visual prompts into the pruning pipeline. The authors introduce visual prompts into trained vision models. During pruning, not only weight masks but also visual prompts are optimized. The whole pipeline includes two steps: (1) fixing pretrained weights of vision models, tuning masks and visual prompts; (2) fixing the mask, fine-tuning both weights and visual prompts. The authors adopt the proposed method on several downstream classification tasks and compare the performances of pruned models with other pruning methods. The experiments show that the proposed method can achieve better performance than other methods with the same sparsity.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "3 good", "strengths": "1. The writing and presentation of this paper are quite good. The logic of the article is very clear, and the choice of words in the writing is also very precise.\n2. The idea of introducing tunable visual prompts into the pruning pipeline is intriguing. The experiments validate the effectiveness of the proposed strategy.\n3. The authors compare the proposed method with other pruning methods and additionally apply the proposed visual prompt pruning strategy to these methods, demonstrating the transferability of their approach.", "weaknesses": "1. Introducing visual prompts into vision models seems to boost their performance. However, comparing models with visual prompts (the proposed method) to those without (other baseline methods) might not be entirely fair. What if we apply both the proposed and baseline methods to a model that has already been fine-tuned with visual prompts?\n2. I have some doubts regarding the generality and performance of this paper.\n(1) Why must we conduct experiments on downstream tasks of ImageNet? Why not directly on ImageNet itself, since most pruning work actually focuses more on performance on ImageNet?\n(2) The method proposed in this paper seems to be limited to scenarios where visual prompts can be applied, with their main application currently being in classification tasks. How can the proposed approach be used for other tasks, such as detection, segmentation, etc.?", "questions": "Please refer to the weaknesses. I hope the authors can provide more experiments to demonstrate the effectiveness of the method.\n\nThe provided additional experiments have addressed my concerns about the performance of VPN in ImageNet and object detection tasks. However, after reading other reviewers' comments, I consider decreasing my score to 5. The reasons are as follows:\n(1) The author describe the advantagement of unstructured sparsity on nonGPU hardware. However, they do not report the latency of unstructured sparsity on CPU. I think conducting experiments on some lightweight networks and reporting latency tested on CPUs can make the results more convincing.\n(2) In regards to structured pruning, the experiments carried out on the CIFAR-100 dataset are insufficient.\n\nOverall, I believe this paper needs more meticulous refinement in its experiments. If it can validate its approach on large-scale datasets for both unstructured and structured pruning settings, it would then be a very solid paper.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698816110016}, {"id": "1Fh5Lr228k", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2675/Reviewer_sGzb"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes to use visual prompts to improve performance of the pruned model by applying visual prompts earlier in the process, aka, before the model fine-tuning. This effort was motivated by the experiments of applying post-pruning prompt to the sparse models with and without fine-tuning. As post-pruning prompts showed only marginal gains to subnets that went through fine-tuning, authors proposed to apply the visual prompts earlier in the process. The proposed scheme was compared with eight pruning baselines on eight classification tasks. Numerical comparisons show the potential of the proposed scheme.", "review_text": "This paper proposes to use visual prompts to improve performance of the pruned model by applying visual prompts earlier in the process, aka, before the model fine-tuning. This effort was motivated by the experiments of applying post-pruning prompt to the sparse models with and without fine-tuning. As post-pruning prompts showed only marginal gains to subnets that went through fine-tuning, authors proposed to apply the visual prompts earlier in the process. The proposed scheme was compared with eight pruning baselines on eight classification tasks. Numerical comparisons show the potential of the proposed scheme.", "strengths": "- The idea of applying visual prompts to identify a subnet which further leads to a better pruning results is interesting.\n- The idea of using visual prompts to control a pretrained vision model is an interesting direction to pursue.", "weaknesses": "- The paper is not easy to read. There's a particular emphasis on \"data model co-design\", but it takes quite a while to understand what this refers to concretely.\n- Why is the visual prompts essential? How about learning additional parameters without using the visual prompt?\n- How would visual prompts be different from data augmentation?", "questions": "Please see my questions in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes to use visual prompts to improve performance of the pruned model by applying visual prompts earlier in the process, aka, before the model fine-tuning. This effort was motivated by the experiments of applying post-pruning prompt to the sparse models with and without fine-tuning. As post-pruning prompts showed only marginal gains to subnets that went through fine-tuning, authors proposed to apply the visual prompts earlier in the process. The proposed scheme was compared with eight pruning baselines on eight classification tasks. Numerical comparisons show the potential of the proposed scheme.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "- The idea of applying visual prompts to identify a subnet which further leads to a better pruning results is interesting.\n- The idea of using visual prompts to control a pretrained vision model is an interesting direction to pursue.", "weaknesses": "- The paper is not easy to read. There's a particular emphasis on \"data model co-design\", but it takes quite a while to understand what this refers to concretely.\n- Why is the visual prompts essential? How about learning additional parameters without using the visual prompt?\n- How would visual prompts be different from data augmentation?", "questions": "Please see my questions in the weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698789990511}, {"id": "DLDBrrYCtV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2675/Reviewer_eRqj"], "rating": "3: reject, not good enough", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "This paper proposes a joint model pruning and visual prompting learning method. By combining these two methods, it could recover the performance loss by pruning. This method is validated on several pruning methods and datasets to demonstate its universality.", "review_text": "This paper proposes a joint model pruning and visual prompting learning method. By combining these two methods, it could recover the performance loss by pruning. This method is validated on several pruning methods and datasets to demonstate its universality.", "strengths": "The idea is easy to follow and effective.It is interesting to see that only a small number of learned parameters could improve the performance.", "weaknesses": "1. I think the authors should focus more on the structured sparse case, unstructured pruning is well known that will not contribute to any acceleration in practice.\n2. For structured pruning, the speedup ratio should use latency, not theoritical FLOPs. And more recent methods should be compared.\n3. The visual prompting method essentially uses lower resolution for the input images. It is necessary to compare a baseline that using a lower resolution image as input, and then prune less parameters to maintain the same FLOPs as the proposed VPN. \n4. Following the previous point, I also wonder whether this method will deteriorate some applications that are senstive to resolution, such as object detection or etc.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a joint model pruning and visual prompting learning method. By combining these two methods, it could recover the performance loss by pruning. This method is validated on several pruning methods and datasets to demonstate its universality.", "soundness": "3 good", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "The idea is easy to follow and effective.It is interesting to see that only a small number of learned parameters could improve the performance.", "weaknesses": "1. I think the authors should focus more on the structured sparse case, unstructured pruning is well known that will not contribute to any acceleration in practice.\n2. For structured pruning, the speedup ratio should use latency, not theoritical FLOPs. And more recent methods should be compared.\n3. The visual prompting method essentially uses lower resolution for the input images. It is necessary to compare a baseline that using a lower resolution image as input, and then prune less parameters to maintain the same FLOPs as the proposed VPN. \n4. Following the previous point, I also wonder whether this method will deteriorate some applications that are senstive to resolution, such as object detection or etc.", "questions": "See above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698744864491}, {"id": "6wzBtJRPsg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2675/Reviewer_hiB3"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "To upgrade vision model sparsification, the paper proposed a data-model co-design sparsification paradigm, where integrating input image with the learnable perturbation, and a network tuning strategy is designed to optimize this issue.\n\nThe algorithm has demonstrated excellent performance on CIFAR-10 and CIFAR-100 datasets.", "review_text": "To upgrade vision model sparsification, the paper proposed a data-model co-design sparsification paradigm, where integrating input image with the learnable perturbation, and a network tuning strategy is designed to optimize this issue.\n\nThe algorithm has demonstrated excellent performance on CIFAR-10 and CIFAR-100 datasets.", "strengths": "1. The manuscript exhibits a commendable level of writing proficiency, featuring well-crafted graphics that enhance the overall presentation and a compelling narrative.\n\n2. The algorithm has showcased remarkable efficacy when applied to the CIFAR-10 and CIFAR-100 datasets, achieving good performance.", "weaknesses": "1. **Inadequate experiments.** This is a primary concern for the reviewer. The paper only presents experiments on CIFAR-10 and CIFAR-100, which, in the era of big data, are considered insufficient. These experiments do not adequately demonstrate the performance of the proposed method. Conducting experiments on larger datasets, such as ImageNet-1k, is essential.\n\nAdditionally, it's worth noting that the method utilizes ImageNet-1K pre-trained weights, which were trained on a resolution of 224. However, the method is tested on CIFAR data with a resolution of 32. It is evident that padding the data to a resolution of 224 can significantly boost performance. From this perspective, experiments specifically conducted on ImageNet with a resolution of 224, and direct performance comparisons with fine-tuning on this resolution, are crucial.\n\nFigure 10 further substantiates this conclusion, showing that the optimal performance is achieved at a resolution of 224, with diminishing performance as the resolution decreases. Therefore, padding CIFAR data with a resolution of 32 to 224 doesn't necessarily demonstrate the superiority of the method. The gains observed in this case could be attributed to the ImageNet-1K pre-trained weights at a resolution of 224.\n\n2. The reviewer also suggests providing performance comparisons with smaller models, as performance metrics on sparser models, such as using MobileNet, would be more indicative and informative.\n\n\n3. Furthermore, the method exhibits significant limitations, as it necessitates the use of pre-trained weights from a larger dataset. The current version seems to require transforming the ImageNet model to CIFAR. It would be insightful to explore the performance without pre-trained weights. Additionally, conducting comparisons on a larger pre-trained dataset, such as ImageNet, appears necessary for the current version.", "questions": "In each figure, the authors have plotted curves labeled as \"our best,\" which may not be entirely necessary as this information can be inferred from the VPNs curve. \n\nMoreover, this plotting style has the potential to cause confusion; upon initial review, it might be perplexing why the curve representing \"ours\" appears as a straight line.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "To upgrade vision model sparsification, the paper proposed a data-model co-design sparsification paradigm, where integrating input image with the learnable perturbation, and a network tuning strategy is designed to optimize this issue.\n\nThe algorithm has demonstrated excellent performance on CIFAR-10 and CIFAR-100 datasets.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The manuscript exhibits a commendable level of writing proficiency, featuring well-crafted graphics that enhance the overall presentation and a compelling narrative.\n\n2. The algorithm has showcased remarkable efficacy when applied to the CIFAR-10 and CIFAR-100 datasets, achieving good performance.", "weaknesses": "1. **Inadequate experiments.** This is a primary concern for the reviewer. The paper only presents experiments on CIFAR-10 and CIFAR-100, which, in the era of big data, are considered insufficient. These experiments do not adequately demonstrate the performance of the proposed method. Conducting experiments on larger datasets, such as ImageNet-1k, is essential.\n\nAdditionally, it's worth noting that the method utilizes ImageNet-1K pre-trained weights, which were trained on a resolution of 224. However, the method is tested on CIFAR data with a resolution of 32. It is evident that padding the data to a resolution of 224 can significantly boost performance. From this perspective, experiments specifically conducted on ImageNet with a resolution of 224, and direct performance comparisons with fine-tuning on this resolution, are crucial.\n\nFigure 10 further substantiates this conclusion, showing that the optimal performance is achieved at a resolution of 224, with diminishing performance as the resolution decreases. Therefore, padding CIFAR data with a resolution of 32 to 224 doesn't necessarily demonstrate the superiority of the method. The gains observed in this case could be attributed to the ImageNet-1K pre-trained weights at a resolution of 224.\n\n2. The reviewer also suggests providing performance comparisons with smaller models, as performance metrics on sparser models, such as using MobileNet, would be more indicative and informative.\n\n\n3. Furthermore, the method exhibits significant limitations, as it necessitates the use of pre-trained weights from a larger dataset. The current version seems to require transforming the ImageNet model to CIFAR. It would be insightful to explore the performance without pre-trained weights. Additionally, conducting comparisons on a larger pre-trained dataset, such as ImageNet, appears necessary for the current version.", "questions": "In each figure, the authors have plotted curves labeled as \"our best,\" which may not be entirely necessary as this information can be inferred from the VPNs curve. \n\nMoreover, this plotting style has the potential to cause confusion; upon initial review, it might be perplexing why the curve representing \"ours\" appears as a straight line.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "None", "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697875808840}], "openreview_url": "https://openreview.net/forum?id=96nX9xIIx2", "arxiv_id": "2312.01397", "paper_pdf": "papers/96nX9xIIx2.pdf", "paper_pdf_sha256": "56d73c5360d9ba99f924049be362260485caf6bcf5ffdd2e9b8c63a2353e7ceb", "paper_pdf_bytes": 4569192, "paper_pdf_source": "openreview", "code_url": "https://github.com/UNITES-Lab/VPNs", "code_repository": "UNITES-Lab/VPNs", "code_commit": "912072c4ec2156a7d0414f00c463f3bd48279ab5", "code_archive": "repos/96nX9xIIx2.zip", "code_archive_sha256": "0a3ffaa4fa693c4cf0b2fb82907aeb748c395973f623e242209ea07fe400df99", "code_archive_bytes": 501483, "code_file_count": 9, "code_extensions": {".py": 6, ".sh": 3}, "github_disk_usage_kb": 500, "github_languages": {"Python": 61870, "Shell": 4708}, "github_archived": false, "github_pushed_at": "2023-12-05T02:36:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/visual-prompting-upgrades-neural-network"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "0DwzMsUNIr", "year": 2023, "status": "rejected", "title": "From Points to Functions: Infinite-dimensional Representations in Diffusion Models", "authors": ["Sarthak Mittal", "Guillaume Lajoie", "Stefan Bauer", "Arash Mehrjou"], "authorids": ["~Sarthak_Mittal1", "~Guillaume_Lajoie1", "~Stefan_Bauer1", "~Arash_Mehrjou1"], "authors_source": "OpenReview API", "abstract": "Diffusion-based generative models learn to iteratively transfer unstructured noise to a complex target distribution as opposed to Generative Adversarial Networks (GANs) or the decoder of Variational Autoencoders (VAEs) which produce samples from the target distribution in a single step. Thus, in diffusion models every sample is naturally connected to a random trajectory which is a solution to a learned stochastic differential equation (SDE). Generative models are only concerned with the final state of this trajectory that delivers samples from the desired distribution. \\cite{abstreiter2021diffusion} showed that these stochastic trajectories can be seen as continuous filters that wash out information along the way. Consequently, it is reasonable to ask if there is an intermediate time step at which the preserved information is optimal for a given downstream task. In this work, we show that a combination of information content from different time steps gives a strictly better representation for the downstream task. We introduce an attention and recurrence based modules that ``learn to mix'' information content of various time-steps such that the resultant representation leads to superior performance in downstream tasks.\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "IezE10DPA9", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5111/Reviewer_tLzz"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new way to solve downstream tasks using diffusion-based representation learning. The authors observe that representation learning inspired by diffusion doesn't give one, but infinitely many representations of the image (one for each time, t, of the diffusion). The work that introduced diffusion-based representation learning used one of those representations to optimize for a downstream task. Instead, this paper proposes to consider the whole trajectory of representations. Particularly, the authors propose to use a Transformer (or an RNN) to map from the trajectory to a single embedding that is then used to solve downstream tasks. Experimentally, this yields improved performance in a wide variety of settings.", "review_text": "The paper proposes a better way to solve downstream problems with diffusion-based learned representations. The key innovation is to use the whole trajectory and not a single point to extract a representation that will be used to solve downstream tasks. The method is not evaluated against other ways to learn representations and hence its practical relevance is not clear. ", "strengths": "Strengths:\n\n* The idea of the paper is simple, yet effective.\n* The paper is well-written.\n* The authors experiment in a wide variety of settings, from synthetic tasks to more realistic tasks, such as CelebA classification.\n* The authors demonstrate (using the attention profiles and the estimated Mutual Information) that the representations of the image are changing over the diffusion time.\n\nWeaknesses:\n\n* The paper does not compare with other methods of representation learning, e.g. Contrastive Learning. Given that the diffusion-based representation learning paper itself has not yet been published, I think it would be useful to compare this new work to more established methods for learning representations.\n* There is a lack of clarity for some things regarding the implementation. Why did the authors need to retrain the models from Abstreiter et. al? Aren't there open-source models available that could have been reused? If not, which models did the authors train? Is this the NCSN++ model with the Variance Exploding SDE (as it is implied from the Background Section)? Would anything change if instead of the Variance Exploding SDE, one used Variance Preserving or the sub-VP SDE?\n* In the experiments of Figure 1 (but also in subsequent experiments), it seems that we are comparing a Transformer that takes as input the whole trajectory vs an MLP that takes as input a single point. Does the transformer and the MLP have the same number of parameters? If the MLP has much less expressive power than the Transformer, it could just be that the observed benefits over the baseline are due to bigger or more powerful architecture (and not because of the trajectory vs single point).\n* It would have been more impactful to report numbers on more standard benchmarks such as CIFAR10 , CIFAR100 or ImageNet. This would give a better sense of how this method compares to other methods for learning representations (e.g. SimCLR).", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper proposes a new way to solve downstream tasks using diffusion-based representation learning. The authors observe that representation learning inspired by diffusion doesn't give one, but infinitely many representations of the image (one for each time, t, of the diffusion). The work that introduced diffusion-based representation learning used one of those representations to optimize for a downstream task. Instead, this paper proposes to consider the whole trajectory of representations. Particularly, the authors propose to use a Transformer (or an RNN) to map from the trajectory to a single embedding that is then used to solve downstream tasks. Experimentally, this yields improved performance in a wide variety of settings.", "strength_and_weaknesses": "Strengths:\n\n* The idea of the paper is simple, yet effective.\n* The paper is well-written.\n* The authors experiment in a wide variety of settings, from synthetic tasks to more realistic tasks, such as CelebA classification.\n* The authors demonstrate (using the attention profiles and the estimated Mutual Information) that the representations of the image are changing over the diffusion time.\n\nWeaknesses:\n\n* The paper does not compare with other methods of representation learning, e.g. Contrastive Learning. Given that the diffusion-based representation learning paper itself has not yet been published, I think it would be useful to compare this new work to more established methods for learning representations.\n* There is a lack of clarity for some things regarding the implementation. Why did the authors need to retrain the models from Abstreiter et. al? Aren't there open-source models available that could have been reused? If not, which models did the authors train? Is this the NCSN++ model with the Variance Exploding SDE (as it is implied from the Background Section)? Would anything change if instead of the Variance Exploding SDE, one used Variance Preserving or the sub-VP SDE?\n* In the experiments of Figure 1 (but also in subsequent experiments), it seems that we are comparing a Transformer that takes as input the whole trajectory vs an MLP that takes as input a single point. Does the transformer and the MLP have the same number of parameters? If the MLP has much less expressive power than the Transformer, it could just be that the observed benefits over the baseline are due to bigger or more powerful architecture (and not because of the trajectory vs single point).\n* It would have been more impactful to report numbers on more standard benchmarks such as CIFAR10 , CIFAR100 or ImageNet. This would give a better sense of how this method compares to other methods for learning representations (e.g. SimCLR).", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written. Certain parts of the experimental evaluation should be further clarified (see above). The proposed idea is original but incremental to the diffusion-based representation learning paper.", "summary_of_the_review": "The paper proposes a better way to solve downstream problems with diffusion-based learned representations. The key innovation is to use the whole trajectory and not a single point to extract a representation that will be used to solve downstream tasks. The method is not evaluated against other ways to learn representations and hence its practical relevance is not clear. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666730391935}, {"id": "xXsUtGkJVq5", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5111/Reviewer_ZUwW"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the representation learned by diffusion models equipped with an encoder. Based on the prior work (Abstreiter et al 2021), the paper obtained a time-dependent encoder along with the training of a diffusion model. \n\nIn contrast to the prior work (Abstreiter et al 2021), which uses the output of the encoder at a single timestep (i.e. t0), the paper considers the discretized trajectory at multiple timesteps simultaneously. It is done by using an RNN or a transformer to aggregate the sequential representations together and make a final decision. \n\nEmpirically, it investigates the benefits of the trajectory information for downstream tasks, analyzes the different information encoded in the trajectory in different time steps by mutual information, parses the semantic information along the trajectory and demonstrates the benefits of using more samples in the trajectory.\n", "review_text": "This paper is clearly below the acceptance bar of ICLR due to the lack of novelty and motivation. ", "strengths": "**Strength**\n1. The paper is clearly written and well organized.\n\n**Weaknesses**\n1. The paper is incremental compared to the prior work (Abstreiter et al 2021). In fact, it is exactly the same to the prior work except that it considers to leverage the \"noisy\" representations along the trajactory simultaneously for downstream tasks.\n\n2. I think the claim of \"points to functions\" and \"infinite dimensional representations\" should be modified, including the title. This is very misleading because some popular algorithms in machine learning like Kernel methods can really do this in function space. However, The proposed method (using RNNs) takes representations at finite time steps as input instead of the whole function.\n\n3. I did not see a clear advantage of diffusion models to learn representations. In fact, the paper does not compare to any other deep generative models like VAEs and GANs or discuss about this. It it not well motivated for me.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper studies the representation learned by diffusion models equipped with an encoder. Based on the prior work (Abstreiter et al 2021), the paper obtained a time-dependent encoder along with the training of a diffusion model. \n\nIn contrast to the prior work (Abstreiter et al 2021), which uses the output of the encoder at a single timestep (i.e. t0), the paper considers the discretized trajectory at multiple timesteps simultaneously. It is done by using an RNN or a transformer to aggregate the sequential representations together and make a final decision. \n\nEmpirically, it investigates the benefits of the trajectory information for downstream tasks, analyzes the different information encoded in the trajectory in different time steps by mutual information, parses the semantic information along the trajectory and demonstrates the benefits of using more samples in the trajectory.\n", "strength_and_weaknesses": "**Strength**\n1. The paper is clearly written and well organized.\n\n**Weaknesses**\n1. The paper is incremental compared to the prior work (Abstreiter et al 2021). In fact, it is exactly the same to the prior work except that it considers to leverage the \"noisy\" representations along the trajactory simultaneously for downstream tasks.\n\n2. I think the claim of \"points to functions\" and \"infinite dimensional representations\" should be modified, including the title. This is very misleading because some popular algorithms in machine learning like Kernel methods can really do this in function space. However, The proposed method (using RNNs) takes representations at finite time steps as input instead of the whole function.\n\n3. I did not see a clear advantage of diffusion models to learn representations. In fact, the paper does not compare to any other deep generative models like VAEs and GANs or discuss about this. It it not well motivated for me.", "clarity,_quality,_novelty_and_reproducibility": "Clarity: the paper is clear.\n\nQuality: the quality of the paper is limited. See details in the weaknesses above.\n\nNovelty: the novelty is limited compared to the prior work (Abstreiter et al 2021).\n\nReproducibility: it seems reproducible.", "summary_of_the_review": "This paper is clearly below the acceptance bar of ICLR due to the lack of novelty and motivation. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666697440029}, {"id": "lK1ISl00TF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5111/Reviewer_PmAJ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper presents a trajectory of representations produced by a time-dependent encoder that leverages the diffusion-based method is more informative and can benefit downstream networks (RNN, transformers, etc.) that are able to handle the mixture of the representations. The authors empirically demonstrate their idea on synthetic data and real-world image datasets such as CIFAR-10, mini-ImageNet, colored-MNIST and CelebA.", "review_text": "The topic studied in this paper is important in the study of representation learning and diffusion-based method. However, some key perspectives regarding the proposed method are not comprehensively studied and remain unclear to the readers, which weakens the contribution of this paper. Moreover, the experimental details are not clearly provided. I am inclined to consider this paper below the acceptance threshold given the current version.  ", "strengths": "**Strength**\n\nThe paper is well-written and organized, making it easy to follow. The claims in the paper are well supported by the corresponding experiments. The authors not only demonstrate the trajectory combination of representations is stronger, but also show which intermediate steps are more important. This observation is important to both diffusion-based model and representation learning community.\n\n**Weakness**\n\nThis paper only provides empirical studies on the representations, but lacks theoretical analysis on why the combination of representations is better. \n\nMost of the method part is heavily dependent on Abstreiter et al. and some claims are duplicated, which may weaken the contribution and novelty.\n\nThe authors only demonstrate their observations on one single diffusion schedule, while there are various options that may affects the observations. If the diffusion adopts a cosine schedule, can we still get the same conclusions? For example, can we still get the highest attentions scores around t=0.5, as shown in Fig. 3?  \n\nOn the effects of granularity, the timesteps are only considered to be uniformly partitioned. There could be more combinations to explore. For example, we can sample steps according to a cosine annealing decay weight for the combination. \n\nThe accuracy shown in the experiments are relatively lower than the results in both supervised and self-supervised cases. It is hard to say this gap is caused by the method or by using a weaker encoder backbone (I could not find the parameterization details of the Wide-ResNet encoder). It is unclear whether such observation could be helpful in the practical case. \n\nThe diffusion schedule is not specified, and some experimental details are not precise; please also see the clarity part.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper presents a trajectory of representations produced by a time-dependent encoder that leverages the diffusion-based method is more informative and can benefit downstream networks (RNN, transformers, etc.) that are able to handle the mixture of the representations. The authors empirically demonstrate their idea on synthetic data and real-world image datasets such as CIFAR-10, mini-ImageNet, colored-MNIST and CelebA.", "strength_and_weaknesses": "**Strength**\n\nThe paper is well-written and organized, making it easy to follow. The claims in the paper are well supported by the corresponding experiments. The authors not only demonstrate the trajectory combination of representations is stronger, but also show which intermediate steps are more important. This observation is important to both diffusion-based model and representation learning community.\n\n**Weakness**\n\nThis paper only provides empirical studies on the representations, but lacks theoretical analysis on why the combination of representations is better. \n\nMost of the method part is heavily dependent on Abstreiter et al. and some claims are duplicated, which may weaken the contribution and novelty.\n\nThe authors only demonstrate their observations on one single diffusion schedule, while there are various options that may affects the observations. If the diffusion adopts a cosine schedule, can we still get the same conclusions? For example, can we still get the highest attentions scores around t=0.5, as shown in Fig. 3?  \n\nOn the effects of granularity, the timesteps are only considered to be uniformly partitioned. There could be more combinations to explore. For example, we can sample steps according to a cosine annealing decay weight for the combination. \n\nThe accuracy shown in the experiments are relatively lower than the results in both supervised and self-supervised cases. It is hard to say this gap is caused by the method or by using a weaker encoder backbone (I could not find the parameterization details of the Wide-ResNet encoder). It is unclear whether such observation could be helpful in the practical case. \n\nThe diffusion schedule is not specified, and some experimental details are not precise; please also see the clarity part.", "clarity,_quality,_novelty_and_reproducibility": "**Clarity**\n\nThe paper is overall clear. I list the points that are not clear to the readers below:\n\nEquation (5) looks similar to proposition 1 in Abstreiter et al. Is this result from there or derived by the authors? Could the authors provide the corresponding source, e.g., the reference or the proof? \n\nThe attention score used in Fig. 3 is not mathematically defined. It is not clear the attention score is calculated with which variables.\n\nThe model architecture is not clarified. For example, what is the number of layers of the Wide-ResNet? \n\n\n\n**Novelty**\n\nThe proposed method has novelty in leveraging the combination of the time-dependent representations, while the method part is extended from previous works like Abstreiter et al. \n\n**Quality**\n\nThe paper provides reasonable empirical studies and lacks theoretical analysis. Some important ablations may be in need to improve the paper quality.\n\n**Reproducibility**\n\nThe authors include some details for the algorithm and model parameterization, but the provided details are insufficient for reproducing their results.\n", "summary_of_the_review": "The topic studied in this paper is important in the study of representation learning and diffusion-based method. However, some key perspectives regarding the proposed method are not comprehensively studied and remain unclear to the readers, which weakens the contribution of this paper. Moreover, the experimental details are not clearly provided. I am inclined to consider this paper below the acceptance threshold given the current version.  ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666589177749}, {"id": "7lHThqFJd_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper5111/Reviewer_WYmT"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper discusses representation learning with diffusion models. The key idea is from Abstreiter et al., where a diffusion model is additionally trained on conditioning information given by an encoder. The benefit of this is that is becomes possible to minimize the denoising score matching objective to zero. This paper observes that different representations are learned at different timesteps of the diffusion, and measures the information contained about in the representations, finding that mid-points of the trajectory would give better performance on downstream applications. The paper also investigates other semantic information encoded in these representations.  ", "review_text": "The paper performs an empirical study over diffusion-based representation learning. While the experiments are extensive, there are not really a lot of technical depth, nor critically useful empirical applications that can be learned from this paper. Even with the entire trajectory of representations, the downstream performance simply does not beat older methods such as contrastive learning. I am also not entirely convinced by this focus on \"infinite-size\" representations in the title and in parts of the paper -- coming up with \"infinite-size\" representations is easy (simply add redundancy would work), but such a representation has to be useful for downstream tasks. ", "strengths": "Strength:\n- The paper is concise and easy to understand. \n- The paper presents extensive empirical studies over the representations, such as the timestep-dependent attention score, mutual information, and downstream application performance.\n\nWeakness:\n- Lack of technical novelty: both of the methods DRL and VDRL come from the Abstreiter paper, and it comes not at a surprise that the representations obtained from such models have infinite dimension. The paper mostly presents experimental evaluation results on this type of representations.\n- Utility of the representations are not obvious: there are fairly powerful representation learning methods that does not even require a diffusion model. For example, in Figure 1 and Figure 9, the classifier accuracy on CIFAR10 is lower than 80% whereas contrastive learning methods could easily achieve over 90% accuracy with linear classifiers. While the observations are interesting, it is not clear why we want to use diffusion representations for such downstream applications (why not do time-dependent contrastive learning?). Also, there are already works that try to learn diffusion-like methods with powerful representations [1]. All in all, the encoder learned here does not seem very useful empirically.\n- Lack of relevance towards existing diffusion models: probably the most interesting part is Figure 5 (c/f), where it suggests that digit identity is highly correlated with large timesteps. However, it is unclear that these findings can be generalized to \"standard\" diffusion models. \n\n[1] D2C: Diffusion-Denoising Models for Few-shot Conditional Generation, NeurIPS 21.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper discusses representation learning with diffusion models. The key idea is from Abstreiter et al., where a diffusion model is additionally trained on conditioning information given by an encoder. The benefit of this is that is becomes possible to minimize the denoising score matching objective to zero. This paper observes that different representations are learned at different timesteps of the diffusion, and measures the information contained about in the representations, finding that mid-points of the trajectory would give better performance on downstream applications. The paper also investigates other semantic information encoded in these representations.  ", "strength_and_weaknesses": "Strength:\n- The paper is concise and easy to understand. \n- The paper presents extensive empirical studies over the representations, such as the timestep-dependent attention score, mutual information, and downstream application performance.\n\nWeakness:\n- Lack of technical novelty: both of the methods DRL and VDRL come from the Abstreiter paper, and it comes not at a surprise that the representations obtained from such models have infinite dimension. The paper mostly presents experimental evaluation results on this type of representations.\n- Utility of the representations are not obvious: there are fairly powerful representation learning methods that does not even require a diffusion model. For example, in Figure 1 and Figure 9, the classifier accuracy on CIFAR10 is lower than 80% whereas contrastive learning methods could easily achieve over 90% accuracy with linear classifiers. While the observations are interesting, it is not clear why we want to use diffusion representations for such downstream applications (why not do time-dependent contrastive learning?). Also, there are already works that try to learn diffusion-like methods with powerful representations [1]. All in all, the encoder learned here does not seem very useful empirically.\n- Lack of relevance towards existing diffusion models: probably the most interesting part is Figure 5 (c/f), where it suggests that digit identity is highly correlated with large timesteps. However, it is unclear that these findings can be generalized to \"standard\" diffusion models. \n\n[1] D2C: Diffusion-Denoising Models for Few-shot Conditional Generation, NeurIPS 21.", "clarity,_quality,_novelty_and_reproducibility": "- Clarity (High). The paper is written clearly with many details provided.\n- Quality (High). The writing of the paper is of high quality.\n- Novelty (Mid-to-low). The core method is from an earlier papers, and the experimental evaluations does not seem to reveal conclusions that are too surprising. \n- Originality (High). The work in itself is original as far as I understand. ", "summary_of_the_review": "The paper performs an empirical study over diffusion-based representation learning. While the experiments are extensive, there are not really a lot of technical depth, nor critically useful empirical applications that can be learned from this paper. Even with the entire trajectory of representations, the downstream performance simply does not beat older methods such as contrastive learning. I am also not entirely convinced by this focus on \"infinite-size\" representations in the title and in parts of the paper -- coming up with \"infinite-size\" representations is easy (simply add redundancy would work), but such a representation has to be useful for downstream tasks. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666579837960}], "openreview_url": "https://openreview.net/forum?id=0DwzMsUNIr", "arxiv_id": "2210.13774", "paper_pdf": "papers/0DwzMsUNIr.pdf", "paper_pdf_sha256": "89b6b10c9c3512e38d1e040c81c307cecc37f331200fe531c134311659914ef5", "paper_pdf_bytes": 3225151, "paper_pdf_source": "openreview", "code_url": "https://github.com/sarthmit/traj_drl", "code_repository": "sarthmit/traj_drl", "code_commit": "61aef79dfa4d88dcb6f91c5e80ebc6e91e8df481", "code_archive": "repos/0DwzMsUNIr.zip", "code_archive_sha256": "dd80fa7b82867a4ae01b6e8788e8d6b5f841241ccaaac054ef41c03d98fbeaa1", "code_archive_bytes": 1422515, "code_file_count": 113, "code_extensions": {".py": 92, ".sh": 17, ".cpp": 2, ".cu": 2}, "github_disk_usage_kb": 1302, "github_languages": {"Python": 494902, "Cuda": 14488, "Shell": 7432, "C++": 1792}, "github_archived": false, "github_pushed_at": "2022-10-25T14:48:48Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/from-points-to-functions-infinite-dimensional"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "hq7vLjZTJPk", "year": 2022, "status": "rejected", "title": "A Communication-Efficient Distributed Gradient Clipping Algorithm for Training Deep Neural Networks", "authors": ["Chunyang Liao", "Zhenxun Zhuang", "Mingrui Liu"], "authorids": ["~Chunyang_Liao1", "~Zhenxun_Zhuang1", "~Mingrui_Liu2"], "authors_source": "OpenReview API", "abstract": "In distributed training of deep neural networks or Federated Learning (FL), people usually run Stochastic Gradient Descent (SGD) or its variants on each machine and communicate with other machines periodically. However, SGD might converge slowly in training some deep neural networks (e.g., RNN, LSTM) because of the exploding gradient issue. Gradient clipping is usually employed to address this issue in the single machine setting, but exploring this technique in the FL setting is still in its infancy: it remains mysterious whether the gradient clipping scheme can take advantage of multiple machines to enjoy parallel speedup in the FL setting. The main technical difficulty lies at dealing with nonconvex loss function, non-Lipschitz continuous gradient, and skipping communication rounds simultaneously. In this paper, we explore a relaxed-smoothness assumption of the loss landscape which LSTM was shown to satisfy in previous works, and design a communication-efficient gradient clipping algorithm. This algorithm can be run on multiple machines, where each machine employs a gradient clipping scheme and communicate with other machines after multiple steps of gradient-based updates. Our algorithm is proved to have $O\\left(\\frac{1}{N\\epsilon^4}\\right)$ iteration complexity for finding an $\\epsilon$-stationary point, where $N$ is the number of machines. This indicates that our algorithm enjoys linear speedup. Our experiments on several benchmark datasets demonstrate that our algorithm indeed exhibits fast convergence speed in practice and validate our theory.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "GEwJZnoYtcz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3623/Reviewer_qwWt"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper considers the effect of gradient clipping in Federated Learning and how it affects the convergence rate. They focused on the relaxed-smooth loss function. Each worker uses local gradient clipping and runs multiple steps of SGD before communicating and averaging the local models. The authors theoretically analyzed the algorithm and showed that for $N$ workers, the algorithm has $O(1/N\\\\epsilon^4)$ iteration complexity to find an $\\\\epsilon$-stationary point. Finally, the theoretical results are experimentally verified on CIFAR-10 (Resent-56 model), Penn Treebank and WikiText (LSTM models).", "review_text": "Gradient clipping is known to be an effective tool in training RNNs and mitigating the gradient explosion. However, there is not much work on analyzing its effect in distributed or federated learning scenarios. This paper is among the few ones that tries to theoretically analyze the behavior of FedAvg and similar algorithms when the workers use gradient clipping in their local SGD updates.\nThe paper is generally well-written and the claims are well-supported. One missing recent paper in this area (which was published recently and I don't expect the authors to be aware of it) is the \"Understanding Clipping for Federated Learning ...\" by Xinwei Zhang, et. al., published recently at ICML'21 Workshop on Federated Learning. It would be nice if the authors can comment and compare their results with the findings in that paper and other similar works.\n\nSome minor suggestions:\n - One nice addition to the paper might be to show empirically how tight the theoretical bounds are (at least for a toy example).\n - Assumption 1, iii, $\\\\nabla$ is missing in the first equation.\n - The paragraph after Lemma 2, last sentence, the error is quadratic in $I$, not linear.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the effect of gradient clipping in Federated Learning and how it affects the convergence rate. They focused on the relaxed-smooth loss function. Each worker uses local gradient clipping and runs multiple steps of SGD before communicating and averaging the local models. The authors theoretically analyzed the algorithm and showed that for $N$ workers, the algorithm has $O(1/N\\\\epsilon^4)$ iteration complexity to find an $\\\\epsilon$-stationary point. Finally, the theoretical results are experimentally verified on CIFAR-10 (Resent-56 model), Penn Treebank and WikiText (LSTM models).", "main_review": "Gradient clipping is known to be an effective tool in training RNNs and mitigating the gradient explosion. However, there is not much work on analyzing its effect in distributed or federated learning scenarios. This paper is among the few ones that tries to theoretically analyze the behavior of FedAvg and similar algorithms when the workers use gradient clipping in their local SGD updates.\nThe paper is generally well-written and the claims are well-supported. One missing recent paper in this area (which was published recently and I don't expect the authors to be aware of it) is the \"Understanding Clipping for Federated Learning ...\" by Xinwei Zhang, et. al., published recently at ICML'21 Workshop on Federated Learning. It would be nice if the authors can comment and compare their results with the findings in that paper and other similar works.\n\nSome minor suggestions:\n - One nice addition to the paper might be to show empirically how tight the theoretical bounds are (at least for a toy example).\n - Assumption 1, iii, $\\\\nabla$ is missing in the first equation.\n - The paragraph after Lemma 2, last sentence, the error is quadratic in $I$, not linear.", "summary_of_the_review": "The paper has theoretically analyzed the effect of gradient clipping in local SGD updates in Federated Learning and FedAvg. To the best of my knowledge, the technical analysis and results are incrementally novel and improves the results of existing works. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636401705774}, {"id": "Knxgf1OMRP", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3623/Reviewer_oVjB"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Gradient clipping is an important technique in training deep neural network. Typically, ones need to use the globally averaged gradient to estimate the norm. However, it requires gradient synchronization for every iterations, which is not practical in federated learning. In practice, practitioners only apply local gradient clipping to the local iterates while the theoretical analysis is lacking. This paper analyzes the convergence of the local gradient clipping. The theoretical results show that the convergence is guaranteed when both the number of workers and the number of local iterations are not too large.", "review_text": "Gradient clipping has become a de facto approach in training the deep neural networks especially for LSTM and Transformer. When communication bandwidth is limited and sharing gradient can leak privacy in federated learning, global gradient clipping is impractical. In this case, practitioners use local gradient clipping as workaround. However, there is no theoretical understanding on this approach, This paper bridges the gap between theory and practice. Overall, I think the paper is well written and easy to understand. But the contribution is incremental. In addition, I have a few comments regarding theory and experiments:\n\n1. Theorem 1 holds when $N \\leq O(1/\\epsilon)$ and $I \\leq O(1/(\\epsilon N))$. On one hand, if we need to use many workers such as $N = 1/\\epsilon$, then $I \\leq O(1)$, which means there is no local iterations at all. On the other hand, if we just need a rough solution (e.g., a moderate $\\epsilon$), which is typically the case in deep learning, then $N$ has to be small. It seems to me the theory does not apply to common cases in practice. It is good to plot the norm of the gradient to see how many cases in the experiments satisfy the assumption.\n\n2. All the experiments focus on homogeneous local data in classic distributed settings. There is no experiment considering federated learning settings where the local data is heterogeneous and only a subset of clients participate in each training round. For the classic distributed training, global gradient clipping is applicable because of high-speed InfiniBand on the cloud and fast NCCL implementation.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Gradient clipping is an important technique in training deep neural network. Typically, ones need to use the globally averaged gradient to estimate the norm. However, it requires gradient synchronization for every iterations, which is not practical in federated learning. In practice, practitioners only apply local gradient clipping to the local iterates while the theoretical analysis is lacking. This paper analyzes the convergence of the local gradient clipping. The theoretical results show that the convergence is guaranteed when both the number of workers and the number of local iterations are not too large.", "main_review": "Gradient clipping has become a de facto approach in training the deep neural networks especially for LSTM and Transformer. When communication bandwidth is limited and sharing gradient can leak privacy in federated learning, global gradient clipping is impractical. In this case, practitioners use local gradient clipping as workaround. However, there is no theoretical understanding on this approach, This paper bridges the gap between theory and practice. Overall, I think the paper is well written and easy to understand. But the contribution is incremental. In addition, I have a few comments regarding theory and experiments:\n\n1. Theorem 1 holds when $N \\leq O(1/\\epsilon)$ and $I \\leq O(1/(\\epsilon N))$. On one hand, if we need to use many workers such as $N = 1/\\epsilon$, then $I \\leq O(1)$, which means there is no local iterations at all. On the other hand, if we just need a rough solution (e.g., a moderate $\\epsilon$), which is typically the case in deep learning, then $N$ has to be small. It seems to me the theory does not apply to common cases in practice. It is good to plot the norm of the gradient to see how many cases in the experiments satisfy the assumption.\n\n2. All the experiments focus on homogeneous local data in classic distributed settings. There is no experiment considering federated learning settings where the local data is heterogeneous and only a subset of clients participate in each training round. For the classic distributed training, global gradient clipping is applicable because of high-speed InfiniBand on the cloud and fast NCCL implementation.\n", "summary_of_the_review": "The paper provides theoretical analysis for local gradient clipping under the federated learning settings. However, the practical cases where the theoretical analysis can be applied is unclear. And, the experiments are based on homogeneous local data in classic distributed settings. The contribution is also incremental.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1636339354438}, {"id": "97U0OOuDTFH", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3623/Reviewer_63cQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper propose a novel distributed optimization method CELGC. CELGC adopts normalized gradient for local training (e.g., FL) so that it achieves better performance than SGD.", "review_text": "The proposed CELGC is simple yet effective. The only difference from local sgd is that CELGC adopts normalized gradient if the norm of gradient is large. The authors also prove its convergence rate for relaxed smooth and non-convex functions.\n\nIn addition to convergence theorem, I care more about the practicality of the proposed algorithm and I have the following questions:\n1. CELGC has many hyper-parameters, e.g., learning rate, clipping threshold and batch size. I notice that in the three experiments, these hyper-parameters are quite different. How to tune these hyper-parameters in a new training problems if we do not have any previous experience (In the experiments, authors often set parameters according to other papers). Is it possible to give an approximate range for these parameters?\n2. The authors claim that CELGC can achieve the linear speedup with proper setting. But they seems do not conduct speedup experiment.\n3. What the difference between the naive version of the parallel gradient clipping algorithm and CELGC with i=1?. I think they are the same one, i.e. \\kappa = 0, so that the performance of the baseline should be better than CELGC with I>1 in the comparison of epochs. This is not consistent with Figure1(a). Besides, I think the authors should set a large batch size for the baseline in the experiments. Normalized gradients may lead to a bad performance when the batch size is small [1]\n\n[1] Beyond convexity- Stochastic quasi-convex optimization. NIPS, 2015.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper propose a novel distributed optimization method CELGC. CELGC adopts normalized gradient for local training (e.g., FL) so that it achieves better performance than SGD.", "main_review": "The proposed CELGC is simple yet effective. The only difference from local sgd is that CELGC adopts normalized gradient if the norm of gradient is large. The authors also prove its convergence rate for relaxed smooth and non-convex functions.\n\nIn addition to convergence theorem, I care more about the practicality of the proposed algorithm and I have the following questions:\n1. CELGC has many hyper-parameters, e.g., learning rate, clipping threshold and batch size. I notice that in the three experiments, these hyper-parameters are quite different. How to tune these hyper-parameters in a new training problems if we do not have any previous experience (In the experiments, authors often set parameters according to other papers). Is it possible to give an approximate range for these parameters?\n2. The authors claim that CELGC can achieve the linear speedup with proper setting. But they seems do not conduct speedup experiment.\n3. What the difference between the naive version of the parallel gradient clipping algorithm and CELGC with i=1?. I think they are the same one, i.e. \\kappa = 0, so that the performance of the baseline should be better than CELGC with I>1 in the comparison of epochs. This is not consistent with Figure1(a). Besides, I think the authors should set a large batch size for the baseline in the experiments. Normalized gradients may lead to a bad performance when the batch size is small [1]\n\n[1] Beyond convexity- Stochastic quasi-convex optimization. NIPS, 2015.", "summary_of_the_review": "This paper has many things to be improved.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635863829973}, {"id": "wY8_OnbaZo-", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper3623/Reviewer_nQu7"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper proposes a new variant of SGD with clipping and infrequent communications (local updates) called Communication Efficient Local Gradient Clipping (CELGC) aimed at solving non-convex federated learning problems under the generalized smoothness ($(L_0, L_1)$-smoothness) assumption. The authors derive ergodic convergence guarantees for the convergence of CELGC to the first-order $\\epsilon$-stationary point assuming additionally that the noise in stochastic gradients is bounded with probability $1$ and heterogeneity of the local data on clients is also bounded. Although the authors claim that their result shows that CELGC achieves linear speed-up and has better communication complexity than the naive version of parallel Clipped SGD, the proofs contain several significant inaccuracies making the main result of the paper incorrect in general. Moreover, several assumptions about the parameters such as the number of workers $N$ and the number of local steps between two consequent communication steps $I$ are restrictive.", "review_text": "## Strengths\n1. **Motivation and clarity.** The paper is well-motivated and easy to follow. The authors also provide a sketch of the proof in the main text which is useful.\n\n2. **Related work** section provides a good summary of existing works relevant to the topic. However, Table 1 is not accurate and should be improved (see **General questions and comments** below).\n\n\n\n\n## Weaknesses\n1. **The proofs contain mathematical mistakes that break the main result of the paper.** In the proof of Lemma 2 the authors rely on the following formula (page 15, the third step at formula (12)): $\\mathbb{E}\\left[ \\sum_{i\\in \\overline J(t)} \\nabla F_i(x_t^i, \\xi_t^i) | \\xi^{[t-1]}\\right] = \\sum_{i\\in \\overline J(t)} \\nabla f_i(x_t^i)$. This is not true since the set $\\overline J(t)$ depends on the stochasticity at iteration $t$ of the algorithm. Moreover, the authors repeat this mistake several times throughout the proof when \"move\" the factor $|\\overline J(t)|$ outside the expectations (e.g., later in the same formula and also when applying (12) to (9) in (14)). Due to the same issue, one cannot use second-moment representation in the fourth row of the formula (13). Finally, even if we assume that everything is correct before the last step in formula (13), there is another issue: in the last row of (13), the second term should be $2\\frac{\\eta^2 |\\overline J(t)|^{\\color{red}2}}{N^2}$. Therefore, in the final bound, instead of the term $\\frac{AL_0\\eta^2\\sigma^2}{N}$ one should have $AL_0\\eta^2\\sigma^2$. As the result, in the last upper bound from the proof of Theorem 1 the third term should be $\\frac{AL_0\\eta^2\\sigma^2}{\\epsilon}$ instead of $\\frac{AL_0\\eta^2\\sigma^2}{{\\color{red}N}\\epsilon}$. Therefore, the number of iterations $T$ is not proportional to $\\frac{1}{N}$, i.e., **linear speed-up is not proven**. Since it is claimed as one of the main contributions of the paper, the described issue is a very strong reason for the rejection of the paper.\n\n2. **Assumptions on the parameters in Theorem 1** are strong and not well-defined for some special cases. First of all, the condition $N \\leq \\frac{1}{\\epsilon}$ is strong and in FL-applications, where $N$ can be large, it is typically satisfied for very small values of $\\epsilon$, meaning that the problem should be solved with too high accuracy. In practice, it is often sufficient to solve the optimization problem with not too high accuracy. Next, the authors also assume that the number of local steps between two subsequent communications satisfies $I \\leq \\frac{1}{2N\\epsilon}$. Since $I$ is a positive integer one should have $N \\leq \\frac{1}{2\\epsilon}$. So, in fact, it means that either $\\epsilon$ is extremely small, or $I = O(1)$, or $N$ is small. The drawbacks of the first and the third options are covered above. The second option describes a theoretically simple case since the number of local steps is small. Moreover, in this case, $N\\epsilon = \\Omega(1)$ meaning that the derived communication complexity bound is not better than for Naive Parallel SGD with clipping. Finally, the assumptions on $\\gamma$ and $\\eta$ are not well-defined when $\\sigma = 0$ and $\\kappa = 0$: the condition $\\frac{\\gamma}{\\eta} = 5(\\sigma + \\kappa)$ implies that $\\frac{\\gamma}{\\eta} = 0$ in this case meaning that either $\\gamma = 0$ (no clipping) or $\\eta = \\infty$. This means that the assumptions on $\\gamma$ and $\\eta$ are incorrect.\n\n3. **Numerical experiments** are conducted for the system with $N = 2$ workers, which is a too small number to demonstrate parallel speed-up properly. Moreover, it would be better to show the dependence on the number of communications rounds in the experiments. Moreover, to have a fair comparison of the method with local updates and the method without them (Naive Parallel SGDClip) one should use I times larger batchsizes for Naive Parallel SGDClip. It seems that the authors used the same batchsizes for local and non-local methods for each iteration.\n\n\n\n## General questions and comments\n1. **Page 2, \"... and hence is communication-efficient\".** This sentence should be rewritten since in the current form it implicitly says that local steps ensure communication efficiency. In fact, it is not true for highly heterogeneous or arbitrary heterogeneous cases (e.g., see Woodworth et al. (2020a) and Gorbunov et al. (2020)).\n\n2. **Table 1** requires clarifications. First of all, Ghadimi & Lan (2013) do not use the bounded gradients assumption. They derive the results under classical smoothness and bounded variance assumptions. This should be explicitly stated either in the table or in the caption of the table. The complexity guarantee should contain two terms: one is proportional to $\\epsilon^{-2}$ and the second one is proportional to $\\epsilon^{-4}$. Moreover, their result can be easily generalized to the naive parallel settings without any assumptions on bounded data. Secondly, in the current form, the complexity bounds for the methods are incorrect for the special case when $\\sigma = 0$ or/and $\\kappa = 0$ since the terms with better dependence on $\\epsilon$ are omitted. This should be fixed. Next, the table should contain the results for Local-SGD under the classical assumptions to have a clear comparison (e.g., from Koloskova, Anastasia, et al. \"A unified theory of decentralized SGD with changing topology and local updates.\" International Conference on Machine Learning. PMLR, 2020). Finally, I have not found the rate presented in the row \"Naive Parallel of (Zhang et al., 2020)\" in (Zhang et al., 2020). This should be clarified (ideally, rigorous proof should be provided).\n\n3. **Missing references.** I believe the authors should at least mention the work Reddi, Sashank, et al. \"Adaptive federated optimization.\" arXiv preprint arXiv:2003.00295 (2020) since this paper considers very relevant methods such as Federated Adam.\n\n4. **Assumption 1 and the remark after it.** Assumption (iii) is too restrictive: it is not satisfied for the noise with the unbounded domain such as Gaussian noise. The authors claim that \"it is a normal assumption when encountering relaxed smoothness\". However, this claim is not explained. I do not see any evidence of why this assumption is reasonable in this case. I understand that it is used in prior works, but I guess it is because of the difficulty of analyzing the methods under relaxed smoothness. Next, assumption (iv) is also quite restrictive. For example, in the convex case, one can analyze Local SGD and its variants without this assumption. I understand that in the non-convex case it is used in prior works, but it is better to write about this explicitly.\n\n5. **Page 5, \"... our algorithm is expected to have better performance.\"** This is not true when $\\kappa$ is large: even Local SGD does not benefit from local steps in this case.\n\n6. **Additional comments about Theorem 1.** Assuming that the formula for the communication complexity is correct, it is still not clear why it is better than the mentioned complexity of Naive Parallel SGD with clipping since $\\kappa$ can be large.\n\n7. **Page 6, \"Another interesting fact is that both iteration complexity and communication complexity only depend on $L_0$.\"** Most likely it means that the authors just omitted some important dependencies in the complexity estimates.\n\n8. **Page 18, lower bound for $T$.** The detailed derivation should be added after fixing all mistakes in the proof.\n\n\n## Minor comments\n1. **Lemma 2, \"If $2\\gamma I \\leq c/L_1$ for some $c > 0$, then $\\gamma \\leq 2\\gamma I \\leq \\frac{c}{L_1}$\".** This is a trivial statement.\n\n2. **Page 14, \"For each iteration $t$, We\"** $\\longrightarrow$ \"For each iteration $t$, we\"\n\n3. **Page 15, \"and (b) holds due to Lemma 1 and Lemma 5\"** $\\longrightarrow$ \"and (d) holds due to Lemma 1 and Lemma 5\"\n\n4. **Lemma 3 restated, page 16:** in the RHS one should have $\\|\\nabla f(\\overline x_t)\\|$\n\n5. **Inequality (17), the second part**: dependencies on $\\sigma + \\kappa$ are missing.\n\n6. **Page 18, definition of $V(x)$:** $\\widetilde{A} \\longrightarrow A$", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a new variant of SGD with clipping and infrequent communications (local updates) called Communication Efficient Local Gradient Clipping (CELGC) aimed at solving non-convex federated learning problems under the generalized smoothness ($(L_0, L_1)$-smoothness) assumption. The authors derive ergodic convergence guarantees for the convergence of CELGC to the first-order $\\epsilon$-stationary point assuming additionally that the noise in stochastic gradients is bounded with probability $1$ and heterogeneity of the local data on clients is also bounded. Although the authors claim that their result shows that CELGC achieves linear speed-up and has better communication complexity than the naive version of parallel Clipped SGD, the proofs contain several significant inaccuracies making the main result of the paper incorrect in general. Moreover, several assumptions about the parameters such as the number of workers $N$ and the number of local steps between two consequent communication steps $I$ are restrictive.", "main_review": "## Strengths\n1. **Motivation and clarity.** The paper is well-motivated and easy to follow. The authors also provide a sketch of the proof in the main text which is useful.\n\n2. **Related work** section provides a good summary of existing works relevant to the topic. However, Table 1 is not accurate and should be improved (see **General questions and comments** below).\n\n\n\n\n## Weaknesses\n1. **The proofs contain mathematical mistakes that break the main result of the paper.** In the proof of Lemma 2 the authors rely on the following formula (page 15, the third step at formula (12)): $\\mathbb{E}\\left[ \\sum_{i\\in \\overline J(t)} \\nabla F_i(x_t^i, \\xi_t^i) | \\xi^{[t-1]}\\right] = \\sum_{i\\in \\overline J(t)} \\nabla f_i(x_t^i)$. This is not true since the set $\\overline J(t)$ depends on the stochasticity at iteration $t$ of the algorithm. Moreover, the authors repeat this mistake several times throughout the proof when \"move\" the factor $|\\overline J(t)|$ outside the expectations (e.g., later in the same formula and also when applying (12) to (9) in (14)). Due to the same issue, one cannot use second-moment representation in the fourth row of the formula (13). Finally, even if we assume that everything is correct before the last step in formula (13), there is another issue: in the last row of (13), the second term should be $2\\frac{\\eta^2 |\\overline J(t)|^{\\color{red}2}}{N^2}$. Therefore, in the final bound, instead of the term $\\frac{AL_0\\eta^2\\sigma^2}{N}$ one should have $AL_0\\eta^2\\sigma^2$. As the result, in the last upper bound from the proof of Theorem 1 the third term should be $\\frac{AL_0\\eta^2\\sigma^2}{\\epsilon}$ instead of $\\frac{AL_0\\eta^2\\sigma^2}{{\\color{red}N}\\epsilon}$. Therefore, the number of iterations $T$ is not proportional to $\\frac{1}{N}$, i.e., **linear speed-up is not proven**. Since it is claimed as one of the main contributions of the paper, the described issue is a very strong reason for the rejection of the paper.\n\n2. **Assumptions on the parameters in Theorem 1** are strong and not well-defined for some special cases. First of all, the condition $N \\leq \\frac{1}{\\epsilon}$ is strong and in FL-applications, where $N$ can be large, it is typically satisfied for very small values of $\\epsilon$, meaning that the problem should be solved with too high accuracy. In practice, it is often sufficient to solve the optimization problem with not too high accuracy. Next, the authors also assume that the number of local steps between two subsequent communications satisfies $I \\leq \\frac{1}{2N\\epsilon}$. Since $I$ is a positive integer one should have $N \\leq \\frac{1}{2\\epsilon}$. So, in fact, it means that either $\\epsilon$ is extremely small, or $I = O(1)$, or $N$ is small. The drawbacks of the first and the third options are covered above. The second option describes a theoretically simple case since the number of local steps is small. Moreover, in this case, $N\\epsilon = \\Omega(1)$ meaning that the derived communication complexity bound is not better than for Naive Parallel SGD with clipping. Finally, the assumptions on $\\gamma$ and $\\eta$ are not well-defined when $\\sigma = 0$ and $\\kappa = 0$: the condition $\\frac{\\gamma}{\\eta} = 5(\\sigma + \\kappa)$ implies that $\\frac{\\gamma}{\\eta} = 0$ in this case meaning that either $\\gamma = 0$ (no clipping) or $\\eta = \\infty$. This means that the assumptions on $\\gamma$ and $\\eta$ are incorrect.\n\n3. **Numerical experiments** are conducted for the system with $N = 2$ workers, which is a too small number to demonstrate parallel speed-up properly. Moreover, it would be better to show the dependence on the number of communications rounds in the experiments. Moreover, to have a fair comparison of the method with local updates and the method without them (Naive Parallel SGDClip) one should use I times larger batchsizes for Naive Parallel SGDClip. It seems that the authors used the same batchsizes for local and non-local methods for each iteration.\n\n\n\n## General questions and comments\n1. **Page 2, \"... and hence is communication-efficient\".** This sentence should be rewritten since in the current form it implicitly says that local steps ensure communication efficiency. In fact, it is not true for highly heterogeneous or arbitrary heterogeneous cases (e.g., see Woodworth et al. (2020a) and Gorbunov et al. (2020)).\n\n2. **Table 1** requires clarifications. First of all, Ghadimi & Lan (2013) do not use the bounded gradients assumption. They derive the results under classical smoothness and bounded variance assumptions. This should be explicitly stated either in the table or in the caption of the table. The complexity guarantee should contain two terms: one is proportional to $\\epsilon^{-2}$ and the second one is proportional to $\\epsilon^{-4}$. Moreover, their result can be easily generalized to the naive parallel settings without any assumptions on bounded data. Secondly, in the current form, the complexity bounds for the methods are incorrect for the special case when $\\sigma = 0$ or/and $\\kappa = 0$ since the terms with better dependence on $\\epsilon$ are omitted. This should be fixed. Next, the table should contain the results for Local-SGD under the classical assumptions to have a clear comparison (e.g., from Koloskova, Anastasia, et al. \"A unified theory of decentralized SGD with changing topology and local updates.\" International Conference on Machine Learning. PMLR, 2020). Finally, I have not found the rate presented in the row \"Naive Parallel of (Zhang et al., 2020)\" in (Zhang et al., 2020). This should be clarified (ideally, rigorous proof should be provided).\n\n3. **Missing references.** I believe the authors should at least mention the work Reddi, Sashank, et al. \"Adaptive federated optimization.\" arXiv preprint arXiv:2003.00295 (2020) since this paper considers very relevant methods such as Federated Adam.\n\n4. **Assumption 1 and the remark after it.** Assumption (iii) is too restrictive: it is not satisfied for the noise with the unbounded domain such as Gaussian noise. The authors claim that \"it is a normal assumption when encountering relaxed smoothness\". However, this claim is not explained. I do not see any evidence of why this assumption is reasonable in this case. I understand that it is used in prior works, but I guess it is because of the difficulty of analyzing the methods under relaxed smoothness. Next, assumption (iv) is also quite restrictive. For example, in the convex case, one can analyze Local SGD and its variants without this assumption. I understand that in the non-convex case it is used in prior works, but it is better to write about this explicitly.\n\n5. **Page 5, \"... our algorithm is expected to have better performance.\"** This is not true when $\\kappa$ is large: even Local SGD does not benefit from local steps in this case.\n\n6. **Additional comments about Theorem 1.** Assuming that the formula for the communication complexity is correct, it is still not clear why it is better than the mentioned complexity of Naive Parallel SGD with clipping since $\\kappa$ can be large.\n\n7. **Page 6, \"Another interesting fact is that both iteration complexity and communication complexity only depend on $L_0$.\"** Most likely it means that the authors just omitted some important dependencies in the complexity estimates.\n\n8. **Page 18, lower bound for $T$.** The detailed derivation should be added after fixing all mistakes in the proof.\n\n\n## Minor comments\n1. **Lemma 2, \"If $2\\gamma I \\leq c/L_1$ for some $c > 0$, then $\\gamma \\leq 2\\gamma I \\leq \\frac{c}{L_1}$\".** This is a trivial statement.\n\n2. **Page 14, \"For each iteration $t$, We\"** $\\longrightarrow$ \"For each iteration $t$, we\"\n\n3. **Page 15, \"and (b) holds due to Lemma 1 and Lemma 5\"** $\\longrightarrow$ \"and (d) holds due to Lemma 1 and Lemma 5\"\n\n4. **Lemma 3 restated, page 16:** in the RHS one should have $\\|\\nabla f(\\overline x_t)\\|$\n\n5. **Inequality (17), the second part**: dependencies on $\\sigma + \\kappa$ are missing.\n\n6. **Page 18, definition of $V(x)$:** $\\widetilde{A} \\longrightarrow A$", "summary_of_the_review": "To sum up, several parts of the proof should be corrected and the theory should be extended to support more general choices of $N$ and $I$. Moreover, several parts of the paper such as the comparison of the known complexity results with the derived ones and numerical experiments require improvements.\n\nUnfortunately, the current version of the paper cannot be accepted to the conference. If the authors resolve the mentioned issues during the rebuttal, I will increase my score.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635859989453}], "openreview_url": "https://openreview.net/forum?id=hq7vLjZTJPk", "arxiv_id": "2205.05040", "paper_pdf": "papers/hq7vLjZTJPk.pdf", "paper_pdf_sha256": "5b9ba9370cbffa344aada27734a1f8d5f515c13c1ecf440b8c17bcf2119fd204", "paper_pdf_bytes": 3780791, "paper_pdf_source": "openreview", "code_url": "https://github.com/MingruiLiu-ML-Lab/Communication-Efficient-Local-Gradient-Clipping", "code_repository": "MingruiLiu-ML-Lab/Communication-Efficient-Local-Gradient-Clipping", "code_commit": "5f572e368492acc30732dbf47121761de70aacf1", "code_archive": "repos/hq7vLjZTJPk.zip", "code_archive_sha256": "b0b0f08554e5f4dfdaa8bd01b91538cbe0feafdb7c17467ec1d0471be2382a24", "code_archive_bytes": 6518207, "code_file_count": 26, "code_extensions": {".py": 20, ".sh": 6}, "github_disk_usage_kb": 6368, "github_languages": {"Python": 94320, "Shell": 10441}, "github_archived": false, "github_pushed_at": "2022-09-21T16:12:17Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-communication-efficient-distributed-3"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "PGmqOzKEPZN", "year": 2021, "status": "rejected", "title": "Non-Negative Bregman Divergence Minimization for Deep Direct Density Ratio Estimation", "authors": ["Masahiro Kato", "Takeshi Teshima"], "authorids": ["~Masahiro_Kato1", "~Takeshi_Teshima1"], "authors_source": "OpenReview API", "abstract": "The estimation of the ratio of two probability densities has garnered attention as the density ratio is useful in various machine learning tasks, such as anomaly detection and domain adaptation. To estimate the density ratio, methods collectively known as direct density ratio estimation (DRE) have been explored. These methods are based on the minimization of the Bregman (BR) divergence between a density ratio model and the true density ratio. However, existing direct DRE suffers from serious overfitting when using flexible models such as neural networks. In this paper, we introduce a non-negative correction for empirical risk using only the prior knowledge of the upper bound of the density ratio. This correction makes a DRE method more robust against overfitting and enables the use of flexible models. In the theoretical analysis, we discuss the consistency of the empirical risk. In our experiments, the proposed estimators show favorable performance in inlier-based outlier detection and covariate shift adaptation.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "mYz0sYiGePz", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1058/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "SUMMARY:\nThis work introduces a new family of Bregman divergences for density ratio estimation with flexible models that aims at solving the train-loss hacking problem. The contribution is foremost theoretical, but includes an experimental validation on benchmark problems.\n\nSTRENGTHS:\n- This work introduces a new family of non-negative Bregman divergences.\n- The proposed estimator is theoretical justified.\n- Experiments show convincing results over the original Bregman divergences (at least for LSIF and PU).\n- The paper is well written, although quite difficult to follow without some theoretical background.\n\nWEAKNESSES:\n- In Section 3.1, it said the loss functions for other machine learning tasks such as classification are lower-bounded, hence (as far I understand) they do not suffer from the train-loss hacking problem. Shall I therefore understand that estimating the density ratio through classification, using the cross-entropy (BR_BKL), does not suffer from this the train-loss hacking issue? In this case, what is the advantage of using nnBR_BKL over BR_BKL?\n- The experiments do not include BR_BKL nor nnBR_BKL, which I would have found quite interesting for the reason mentioned above.\n- While appreciate the estimation error bound for D3RE, can you comment on the error bound for the original Bregman divergences? How much does nnBR improve over it? \n- How tight is the error bound for D3RE?\n- Some recent applications of density ratio estimation that could have been worth mentioning as well as likelihood-free inference approaches based on likelihood ratios.\n\nDISCLAIMER:\nDue to the critical conditions in my country due to the pandemic, I must admit that I did not verify the mathematical developments. I am sorry for the limited quality of my review. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Delay with my review", "review": "SUMMARY:\nThis work introduces a new family of Bregman divergences for density ratio estimation with flexible models that aims at solving the train-loss hacking problem. The contribution is foremost theoretical, but includes an experimental validation on benchmark problems.\n\nSTRENGTHS:\n- This work introduces a new family of non-negative Bregman divergences.\n- The proposed estimator is theoretical justified.\n- Experiments show convincing results over the original Bregman divergences (at least for LSIF and PU).\n- The paper is well written, although quite difficult to follow without some theoretical background.\n\nWEAKNESSES:\n- In Section 3.1, it said the loss functions for other machine learning tasks such as classification are lower-bounded, hence (as far I understand) they do not suffer from the train-loss hacking problem. Shall I therefore understand that estimating the density ratio through classification, using the cross-entropy (BR_BKL), does not suffer from this the train-loss hacking issue? In this case, what is the advantage of using nnBR_BKL over BR_BKL?\n- The experiments do not include BR_BKL nor nnBR_BKL, which I would have found quite interesting for the reason mentioned above.\n- While appreciate the estimation error bound for D3RE, can you comment on the error bound for the original Bregman divergences? How much does nnBR improve over it? \n- How tight is the error bound for D3RE?\n- Some recent applications of density ratio estimation that could have been worth mentioning as well as likelihood-free inference approaches based on likelihood ratios.\n\nDISCLAIMER:\nDue to the critical conditions in my country due to the pandemic, I must admit that I did not verify the mathematical developments. I am sorry for the limited quality of my review. ", "rating": "6: Marginally above acceptance threshold", "confidence": "2: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper"}, "tcdate": 1604179913910}, {"id": "atkzQAWwHRX", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1058/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "***\n\nSummary:\n\nThe paper addresses learning the ratio between two densities from their samples, with applications to outlier detection and covariate shift adaption. An existing approach is to minimize the Bregman (BR) divergence's empirical approximation while modeling the density ratio function $r^*$ by a flexible hypothesis family, such as neural networks (NNs). A particular issue of such an approach (that the present work aims to resolve) is \"train-loss hacking,\" meaning that the empirical loss can become arbitrarily large and negative. A new loss/objective based on BR divergence has been proposed, appearing on page 4, and is referred to as $\\widehat{\\text{nnBR}}_f(r)$. The major theoretical result, Theorem 1, states that minimizing the proposed objective effectively minimizes the BR divergence for sufficiently large sample sizes. Following this theorem and its corollary, the paper presents empirical evaluations, showing the new algorithm outperforms prior ones on standard datasets.\n\n***\n\nReasons for score:\n\nI think the paper is marginally below the acceptance threshold of ICLR. Learning the density ratio is a well-motivated problem with an objective different from density or functional estimation. But it is unclear to me how strong the theoretical claims are, and the writing needs significantly more work. I am also concerned about the reasoning behind several claims and the potential gap between the derived theory and experiments. Please refer to the \"Cons\" for details. \n\n***\n\nPros:\n\n \n1. The paper attempts to resolve a particular issue for employing flexible hypothesis families in density ratio estimation with BR-divergence loss. I think the problem is interesting and practically relevant. In particular, as the BR divergence encompasses several well-known loss functions, the proposed method naturally induces a class of new methods (page 4).\n\n2. The contribution is also meaningful as the algorithm incorporates the hypothesis class of neural networks, known to have strong expressive power, while some prior works don't. Under a sequence of Assumptions (1 to 4), Corollary 1 presents finite-sample guarantees for a specific NN class with bounded complexity. Experimental results further confirm the algorithm's effectiveness on tasks involving labeled images and text contents. \n\n3. I also like how the paper imposes assumptions and identifies/handles the portion of the vanilla BR-type risk contributing to \"train-loss hacking,\" as the rationale seems natural and might apply to other relevant tasks.\n\n***\n\nCons:\n \n1. It is unclear to me how strong those theoretical claims are.\n\n- First, the main theorem, Theorem 1, holds under a sequence of assumptions. I can see that the first part of Assumption 1 is natural, but assuming an upper bound $\\bar R$ on the density ratio might be a problem when combined with other assumptions/settings. Before continuing, I would like to point out a hidden assumption, $\\ 0 < C < \\frac1R\\ $, right before Assumption 2, where $C$ appears in $\\partial f(t) = \\boldsymbol{C}(\\partial f(t) t - f(t)) + \\tilde f(t)$, as a key parameter in the definition of $\\tilde f$. Note that this assumption translates to $0<R<1/C$. Additionally, observe that $1/5\\le C$ in all the experiments and a very small $C$ can \"damage the empirical performance\" (Remark 1). Given all these constraints, one essentially requires the density ratio to be bounded by a small absolute constant, which seems restrictive.\n\n- Second, I wonder if there is any quantitative assessment of the tightness of Theorem 1 or Corollary 1 (page 5). Let me use the latter as an example. In terms of sample size dependency, the discrepancy upper bound is $\\mathcal O(\\max \\\\{ 1/\\sqrt n_{\\text{de}}, 1/\\sqrt{n_{\\text{nu}}} \\\\} )$. For the BR divergence, is such dependency optimal (up to absolute constant factors)? As for $\\delta$, the $\\sqrt{\\log \\frac1\\delta}$ factor seems tight, but how about the other parameters? In particular, does any constant in the theorem, e.g., $\\kappa_1$ or $\\kappa_2$, has non-polynomial dependency on $C, B_p, L$, or $B_{W_j}$? I think additional details and explanations should be provided to illustrate the meaning and significance of these results. \n\n- Third, \"Computation\" would be an important aspect of the proposed heuristic. I don't see how to derive an efficient optimization procedure for finding the actual empirical risk minimizer or an accurate approximation. From a theoretical point of view, such a procedure is crucial for the sample-dependent error bounds in Theorem/Corollary 1 to take effect. For the experimental results' completeness, it also seems reasonable to add relevant details about implementing the proposed algorithm.\n \n2. I also believe the writing of the paper needs improvement. \n\n- As stated above, the assumptions, theorems, and definitions require more explanation. In particular, I suggest adding comments addressing how natural or restrictive the assumptions are, the significance of the theorem/corollary, and their optimality, and the intuitions behind the technical reasoning. For example, page 4 presents the new objective function's construction right after Assumption 2, which currently is a sequence of formulas/definitions together with light comments. In my opinion, this part is crucial for the paper and should show technical insights and novelty of the proposed method, along with relevant intuitions. Besides, the instantiation of the \"nnBR divergence\" with existing methods shown at the base of the page doesn't seem important to me and somehow introduces redundancy, given Table 1. \n\n- On page 6, the paper claims that \"the above theoretical guarantee already suffices for the basic justification of D3RE.\" I don't fully understand where this sufficiency comes from. Furthermore, I wonder if any prior or related method(s) satisfy a bound similar to that in Theorem 1, namely, if the sample sizes are large, the minimizer is essentially optimal. A concern relevant to this point is that the paper lacks references in a few places. For example, the second last paragraph on page 2 says, \"Therefore, various methods for directly estimating the density ratio model have been proposed (Sugiyama et al., 2012), and (Sugiyama et al. 2011b) ...\". My thought is that if this is a popular problem, then it should've studied by at least a few research groups instead of one.\n\n3. Other comments, suggestions, and questions.\n\n- On page 2, the original pointwise BR divergence is defined as $BR'_f(t^*\\vert\\\\!\\vert t)$. It might be better to avoid using the $(\\cdot)'$ notation as it usually serves as an indicator for derivatives.\n\n- In Equation (3) and some other definitions, I found $\\boldsymbol x_i$ is used instead of $\\boldsymbol x$, while the left-hand side quantities do not involve $i$. \n\n- On page 5, it is said that \"for the boundedness and Lipschitz continuity in Assumption 3 to hold ..., a technical assumption $b_r>0$ is sufficient.\" I wonder how the relevant quantities in Assumption 3 appear as functions of $b_r$. \n\n- On page 4, $\\ell_2(t)$ is introduced to represent $-\\tilde f(t)$. What is the purpose of this new notation?\n\nIt would be nice if the author(s) can address some of my concerns in the rebuttal. Thanks.\n\n*** ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Both the problem and contribution are interesting, but the writing and explanation need improvement. ", "review": "***\n\nSummary:\n\nThe paper addresses learning the ratio between two densities from their samples, with applications to outlier detection and covariate shift adaption. An existing approach is to minimize the Bregman (BR) divergence's empirical approximation while modeling the density ratio function $r^*$ by a flexible hypothesis family, such as neural networks (NNs). A particular issue of such an approach (that the present work aims to resolve) is \"train-loss hacking,\" meaning that the empirical loss can become arbitrarily large and negative. A new loss/objective based on BR divergence has been proposed, appearing on page 4, and is referred to as $\\widehat{\\text{nnBR}}_f(r)$. The major theoretical result, Theorem 1, states that minimizing the proposed objective effectively minimizes the BR divergence for sufficiently large sample sizes. Following this theorem and its corollary, the paper presents empirical evaluations, showing the new algorithm outperforms prior ones on standard datasets.\n\n***\n\nReasons for score:\n\nI think the paper is marginally below the acceptance threshold of ICLR. Learning the density ratio is a well-motivated problem with an objective different from density or functional estimation. But it is unclear to me how strong the theoretical claims are, and the writing needs significantly more work. I am also concerned about the reasoning behind several claims and the potential gap between the derived theory and experiments. Please refer to the \"Cons\" for details. \n\n***\n\nPros:\n\n \n1. The paper attempts to resolve a particular issue for employing flexible hypothesis families in density ratio estimation with BR-divergence loss. I think the problem is interesting and practically relevant. In particular, as the BR divergence encompasses several well-known loss functions, the proposed method naturally induces a class of new methods (page 4).\n\n2. The contribution is also meaningful as the algorithm incorporates the hypothesis class of neural networks, known to have strong expressive power, while some prior works don't. Under a sequence of Assumptions (1 to 4), Corollary 1 presents finite-sample guarantees for a specific NN class with bounded complexity. Experimental results further confirm the algorithm's effectiveness on tasks involving labeled images and text contents. \n\n3. I also like how the paper imposes assumptions and identifies/handles the portion of the vanilla BR-type risk contributing to \"train-loss hacking,\" as the rationale seems natural and might apply to other relevant tasks.\n\n***\n\nCons:\n \n1. It is unclear to me how strong those theoretical claims are.\n\n- First, the main theorem, Theorem 1, holds under a sequence of assumptions. I can see that the first part of Assumption 1 is natural, but assuming an upper bound $\\bar R$ on the density ratio might be a problem when combined with other assumptions/settings. Before continuing, I would like to point out a hidden assumption, $\\ 0 < C < \\frac1R\\ $, right before Assumption 2, where $C$ appears in $\\partial f(t) = \\boldsymbol{C}(\\partial f(t) t - f(t)) + \\tilde f(t)$, as a key parameter in the definition of $\\tilde f$. Note that this assumption translates to $0<R<1/C$. Additionally, observe that $1/5\\le C$ in all the experiments and a very small $C$ can \"damage the empirical performance\" (Remark 1). Given all these constraints, one essentially requires the density ratio to be bounded by a small absolute constant, which seems restrictive.\n\n- Second, I wonder if there is any quantitative assessment of the tightness of Theorem 1 or Corollary 1 (page 5). Let me use the latter as an example. In terms of sample size dependency, the discrepancy upper bound is $\\mathcal O(\\max \\\\{ 1/\\sqrt n_{\\text{de}}, 1/\\sqrt{n_{\\text{nu}}} \\\\} )$. For the BR divergence, is such dependency optimal (up to absolute constant factors)? As for $\\delta$, the $\\sqrt{\\log \\frac1\\delta}$ factor seems tight, but how about the other parameters? In particular, does any constant in the theorem, e.g., $\\kappa_1$ or $\\kappa_2$, has non-polynomial dependency on $C, B_p, L$, or $B_{W_j}$? I think additional details and explanations should be provided to illustrate the meaning and significance of these results. \n\n- Third, \"Computation\" would be an important aspect of the proposed heuristic. I don't see how to derive an efficient optimization procedure for finding the actual empirical risk minimizer or an accurate approximation. From a theoretical point of view, such a procedure is crucial for the sample-dependent error bounds in Theorem/Corollary 1 to take effect. For the experimental results' completeness, it also seems reasonable to add relevant details about implementing the proposed algorithm.\n \n2. I also believe the writing of the paper needs improvement. \n\n- As stated above, the assumptions, theorems, and definitions require more explanation. In particular, I suggest adding comments addressing how natural or restrictive the assumptions are, the significance of the theorem/corollary, and their optimality, and the intuitions behind the technical reasoning. For example, page 4 presents the new objective function's construction right after Assumption 2, which currently is a sequence of formulas/definitions together with light comments. In my opinion, this part is crucial for the paper and should show technical insights and novelty of the proposed method, along with relevant intuitions. Besides, the instantiation of the \"nnBR divergence\" with existing methods shown at the base of the page doesn't seem important to me and somehow introduces redundancy, given Table 1. \n\n- On page 6, the paper claims that \"the above theoretical guarantee already suffices for the basic justification of D3RE.\" I don't fully understand where this sufficiency comes from. Furthermore, I wonder if any prior or related method(s) satisfy a bound similar to that in Theorem 1, namely, if the sample sizes are large, the minimizer is essentially optimal. A concern relevant to this point is that the paper lacks references in a few places. For example, the second last paragraph on page 2 says, \"Therefore, various methods for directly estimating the density ratio model have been proposed (Sugiyama et al., 2012), and (Sugiyama et al. 2011b) ...\". My thought is that if this is a popular problem, then it should've studied by at least a few research groups instead of one.\n\n3. Other comments, suggestions, and questions.\n\n- On page 2, the original pointwise BR divergence is defined as $BR'_f(t^*\\vert\\\\!\\vert t)$. It might be better to avoid using the $(\\cdot)'$ notation as it usually serves as an indicator for derivatives.\n\n- In Equation (3) and some other definitions, I found $\\boldsymbol x_i$ is used instead of $\\boldsymbol x$, while the left-hand side quantities do not involve $i$. \n\n- On page 5, it is said that \"for the boundedness and Lipschitz continuity in Assumption 3 to hold ..., a technical assumption $b_r>0$ is sufficient.\" I wonder how the relevant quantities in Assumption 3 appear as functions of $b_r$. \n\n- On page 4, $\\ell_2(t)$ is introduced to represent $-\\tilde f(t)$. What is the purpose of this new notation?\n\nIt would be nice if the author(s) can address some of my concerns in the rebuttal. Thanks.\n\n*** ", "rating": "5: Marginally below acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604115187504}, {"id": "fTcvK-jL8Cx", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1058/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper addresses an issue that arises in a particular formulation of the density ratio estimation problem. Namely, when one tries to directly fit a density estimation ratio, while minimizing Bergman divergence, it may be that overfitting causes the minimization problem to diverge to minus-infinity.\nThe paper suggests some form of regularization that uses a bound on the ell_infy norm of the ratio function. The paper justifies the change by proving a bound on Bergman divergence between minimizer computed and the best minimizer within the class.\nThe authors support their suggestion by experimenting: computing anomalies and learning from positive or unlabeled data.\n\nThe main drawback of the paper is that it is very hard to read, overloaded with cumbersome notation and lacks sensible discussion about the meaning of the results. Frankly, I couldn't decipher exactly why the over-fitting problem arises, what the 'theoretical justification' exactly says. While the experiments seems good, I didn't fully understand what was done, why are these the right experiments and what was learned from them.\nMore serious is that I can't understand how it compares to prior work. The paper claims it extends a work by Kiryo et al. which I'm not familiar with. There are other lines of research that attack the general problem of density estimation, some of them very related. \nIn particular, one can use arbitrarily strong learning models via the approach of the following paper:\n\"Estimating Divergence Functionals and the Likelihood Ratio by Convex Risk Minimization\" by Ngoyan, Wainright, Jordan.\nA comparison with this paper (and the rich literature that cites it) is lacking. \nAnother line of research for direct estimation comes from the econometric community. A starting point is\n\"Covariate balancing propensity score\" by Imai and Ratkovic. \nThe latter is less directly related but represents a huge body of work on direct density ratio estimation that should be cited and discussed.\n\nTo conclude, the paper's suggestion might be a good one, it is just too hard to asses. It seems to insist on a particular formulation of the problem and then find ways to make it work, but the writing is not clear and important related work is missing. \n\nSpecific comments, mostly related to style and presentation:\n1. In the introduction you have to discuss better what is the problem you solve exactly. How is direct density estimation different from calibrated classification? What is train-loss hacking and why is it unique to this formulation of the problem? etc.\n2. Don't assume that readers are familiar with Kiryo et al.\n3. Equation (3) needs to be discussed. That's the heart of the optimization to solve. I think it is somewhat similar to the one in NWJ I mention above, but I'm not sure. In any case, it is a bit 'weird' that r^* could be dropped from the formulation. What does that mean? \n4. Figure 1 suggests the issue is over fitting, I didn't understand the claim it is of a different flavor.\n5. I thing the assumption that the ratio is bounded is a reasonable one. One can try to justify it better.\n6. The last paragraph of Section 3 has to be explained better.\n7. I found Section 4 frustrating to read. The lack of numbered equations, the wording and the notation just confused me too much. At the end of the day I didn't fully understand what is the meaning of the theorem. For instance, Lemma 2 assumes the true ratio function is part of the class H. That's a very strong assumption no? How do these bounds compare to previous bounds?\n8. In the experiments, besides applications and such, don't you want simply to check the quality of the estimation of the ratio when the ground truth is known? I don't understand how the first experiment in Section 5 achieves that (I think I do, but I'm not sure).\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "The paper suggests a way to estimate the density ratio of two distributions. While the contribution seems potentially useful, it is very hard to follow.", "review": "The paper addresses an issue that arises in a particular formulation of the density ratio estimation problem. Namely, when one tries to directly fit a density estimation ratio, while minimizing Bergman divergence, it may be that overfitting causes the minimization problem to diverge to minus-infinity.\nThe paper suggests some form of regularization that uses a bound on the ell_infy norm of the ratio function. The paper justifies the change by proving a bound on Bergman divergence between minimizer computed and the best minimizer within the class.\nThe authors support their suggestion by experimenting: computing anomalies and learning from positive or unlabeled data.\n\nThe main drawback of the paper is that it is very hard to read, overloaded with cumbersome notation and lacks sensible discussion about the meaning of the results. Frankly, I couldn't decipher exactly why the over-fitting problem arises, what the 'theoretical justification' exactly says. While the experiments seems good, I didn't fully understand what was done, why are these the right experiments and what was learned from them.\nMore serious is that I can't understand how it compares to prior work. The paper claims it extends a work by Kiryo et al. which I'm not familiar with. There are other lines of research that attack the general problem of density estimation, some of them very related. \nIn particular, one can use arbitrarily strong learning models via the approach of the following paper:\n\"Estimating Divergence Functionals and the Likelihood Ratio by Convex Risk Minimization\" by Ngoyan, Wainright, Jordan.\nA comparison with this paper (and the rich literature that cites it) is lacking. \nAnother line of research for direct estimation comes from the econometric community. A starting point is\n\"Covariate balancing propensity score\" by Imai and Ratkovic. \nThe latter is less directly related but represents a huge body of work on direct density ratio estimation that should be cited and discussed.\n\nTo conclude, the paper's suggestion might be a good one, it is just too hard to asses. It seems to insist on a particular formulation of the problem and then find ways to make it work, but the writing is not clear and important related work is missing. \n\nSpecific comments, mostly related to style and presentation:\n1. In the introduction you have to discuss better what is the problem you solve exactly. How is direct density estimation different from calibrated classification? What is train-loss hacking and why is it unique to this formulation of the problem? etc.\n2. Don't assume that readers are familiar with Kiryo et al.\n3. Equation (3) needs to be discussed. That's the heart of the optimization to solve. I think it is somewhat similar to the one in NWJ I mention above, but I'm not sure. In any case, it is a bit 'weird' that r^* could be dropped from the formulation. What does that mean? \n4. Figure 1 suggests the issue is over fitting, I didn't understand the claim it is of a different flavor.\n5. I thing the assumption that the ratio is bounded is a reasonable one. One can try to justify it better.\n6. The last paragraph of Section 3 has to be explained better.\n7. I found Section 4 frustrating to read. The lack of numbered equations, the wording and the notation just confused me too much. At the end of the day I didn't fully understand what is the meaning of the theorem. For instance, Lemma 2 assumes the true ratio function is part of the class H. That's a very strong assumption no? How do these bounds compare to previous bounds?\n8. In the experiments, besides applications and such, don't you want simply to check the quality of the estimation of the ratio when the ground truth is known? I don't understand how the first experiment in Section 5 achieves that (I think I do, but I'm not sure).\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604012131868}, {"id": "01MMxCvFeKn", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1058/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "##########################################################################\n\nSummary:\n\nThe paper studies density ratio estimation (DRE), addressing the 'train-loss hacking' problems which often arise and hamper estimation when models are too flexible. The authors propose a new risk estimator for DRE, providing a non-negative Bregman divergence estimator, with the non-negative correction. Theoretical analyses are shown with the estimation error and empirical analyses are examined in multiple machine learning problem settings.  \n\n##########################################################################\n\nPros: \n\n- A simple yet practical and principled algorithm for DRE is proposed.\n- Although the non-negative correction idea itself is not new, as the authors mentioned in Section 1, using the prior knowledge for the non-negative correction appears novel in DRE. \n- All numerical experiments considered in the paper are reasonable and interesting.\n\n##########################################################################\n\nCons (or questions): \n\n- Although the results in section 4 justify the proposed estimator has vanishing estimation error in some sense, the core idea for the results are basically a simple application of the Mcdiarmid’s inequality. I acknowledge these are highly technical results, but I believe more theoretical analyses of the approximation error would improve the paper. What conditions are needed for r^* (or model space) to have zero approximation error? How can this be explained in terms of a usual reproducing kernel Hilbert space?\n\n- Based on the theories in Section 4, the current algorithm provides an estimator that has a vanishing estimation error. As far as I understand, it might not guarantee the maximization of AUROC (or minimization of pairwise disagreement). Are there any explicit relationships between the 'consistent' density ratio estimation and the AUROC maximization?\n\n- Following the previous question, Figure 3 in the appendix shows the proposed algorithm does not diverge to the negative infinity, but it is not clear if the Bregman divergence evaluated at the proposed estimator is converging to the 'optimal' Bregman divergence value. How can we empirically justify Corollary 1 (even in a very simple simulation setting?)\n\n##########################################################################\n\nOverall, I recommend the weak acceptance. I may well have missed some points in my reading, so clarification is welcome.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting analysis of density ratio estimation", "review": "##########################################################################\n\nSummary:\n\nThe paper studies density ratio estimation (DRE), addressing the 'train-loss hacking' problems which often arise and hamper estimation when models are too flexible. The authors propose a new risk estimator for DRE, providing a non-negative Bregman divergence estimator, with the non-negative correction. Theoretical analyses are shown with the estimation error and empirical analyses are examined in multiple machine learning problem settings.  \n\n##########################################################################\n\nPros: \n\n- A simple yet practical and principled algorithm for DRE is proposed.\n- Although the non-negative correction idea itself is not new, as the authors mentioned in Section 1, using the prior knowledge for the non-negative correction appears novel in DRE. \n- All numerical experiments considered in the paper are reasonable and interesting.\n\n##########################################################################\n\nCons (or questions): \n\n- Although the results in section 4 justify the proposed estimator has vanishing estimation error in some sense, the core idea for the results are basically a simple application of the Mcdiarmid’s inequality. I acknowledge these are highly technical results, but I believe more theoretical analyses of the approximation error would improve the paper. What conditions are needed for r^* (or model space) to have zero approximation error? How can this be explained in terms of a usual reproducing kernel Hilbert space?\n\n- Based on the theories in Section 4, the current algorithm provides an estimator that has a vanishing estimation error. As far as I understand, it might not guarantee the maximization of AUROC (or minimization of pairwise disagreement). Are there any explicit relationships between the 'consistent' density ratio estimation and the AUROC maximization?\n\n- Following the previous question, Figure 3 in the appendix shows the proposed algorithm does not diverge to the negative infinity, but it is not clear if the Bregman divergence evaluated at the proposed estimator is converging to the 'optimal' Bregman divergence value. How can we empirically justify Corollary 1 (even in a very simple simulation setting?)\n\n##########################################################################\n\nOverall, I recommend the weak acceptance. I may well have missed some points in my reading, so clarification is welcome.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1602987160599}], "openreview_url": "https://openreview.net/forum?id=PGmqOzKEPZN", "arxiv_id": "2006.06979", "paper_pdf": "papers/PGmqOzKEPZN.pdf", "paper_pdf_sha256": "45cab1b2e9f942bbf8b96da30a2eca81b84ca95333cd3fa4f07961595057cdc9", "paper_pdf_bytes": 4267784, "paper_pdf_source": "openreview", "code_url": "https://github.com/MasaKat0/D3RE", "code_repository": "MasaKat0/D3RE", "code_commit": "fa008291ec2b30654081f8c80c0b682fe5df1f76", "code_archive": "repos/PGmqOzKEPZN.zip", "code_archive_sha256": "c0eb5c7feb70334b25bae3d11ddf9219317c9274cba90b5fde7e5e4346d1c823", "code_archive_bytes": 7703983, "code_file_count": 90, "code_extensions": {".py": 79, ".sh": 11}, "github_disk_usage_kb": 4639, "github_languages": {"Python": 382981, "Shell": 13989}, "github_archived": false, "github_pushed_at": "2022-11-29T19:56:39Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/non-negative-bregman-divergence-minimization"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "996aKQIom0", "year": 2025, "status": "rejected", "title": "PingPong: A Benchmark for Role-Playing Language Models with User Emulation and Multi-Model Evaluation", "authors": ["Ilya Gusev"], "authorids": ["~Ilya_Gusev1"], "authors_source": "OpenReview API", "abstract": "We introduce a benchmark for evaluating the role-playing capabilities of language models. Our approach leverages language models themselves to emulate users in dynamic, multi-turn conversations and to assess the resulting dialogues. The framework consists of three main components: a player model assuming a specific character role, an interrogator model simulating user behavior, and several judge models evaluating conversation quality. We conducted experiments comparing automated evaluations with human annotations to validate our approach, demonstrating strong correlations across multiple criteria. This work provides a foundation for a robust and dynamic evaluation of model capabilities in interactive scenarios.", "decision": "Reject", "meta_review": "", "num_reviews": 6, "reviews": [{"id": "uYljdHXaQE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_u32x"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper introduces PingPong a benchmark that aims to simulate and assess multi-turn interactions using three components Player, Interrogator, and Judge models. The authors have focused on role-playing models for entertainment purposes. They do this in two versions: In the first version the judge and the interrogator are played by a single model while in the second version these roles are separated in two different models. The player is provided a character card defining its role while the interrogator has the details of the scenario. The judge is supposed to score each turn based on 3 criteria: entertainment, character consistency and language fluency.", "review_text": "The paper introduces PingPong a benchmark that aims to simulate and assess multi-turn interactions using three components Player, Interrogator, and Judge models. The authors have focused on role-playing models for entertainment purposes. They do this in two versions: In the first version the judge and the interrogator are played by a single model while in the second version these roles are separated in two different models. The player is provided a character card defining its role while the interrogator has the details of the scenario. The judge is supposed to score each turn based on 3 criteria: entertainment, character consistency and language fluency.", "strengths": "The authors have focused on role-playing models tailored for entertainment, which is an underrepresented area in benchmarks and that too in multi-turn settings.", "weaknesses": "I have many concerns with this paper. A Judge which is itself an LLM with inherent biases is assessing a highly subjective quality like “Entertainment”. Measuring entertainment is not straightforward and can have varying stylistic and cultural traits. Evaluating that without a human reference data compounds this issue and thus the reliability of the judge can’t be established. Similar concerns with character consistency.\n\nIn role-playing, each turn can be dependent on prior turns, which can’t be fully captured by scoring turns in isolation. While scoring each turn provides a granular view of performance, it may miss the overarching coherence of the character and storyline across multiple turns. The evaluation also overlooks user-centric metrics like engagement, user satisfaction, ability to sustain engagement over extended interactions which are important for role-playing. The paper’s current scoring approach does not seem to assess these aspects. Also these criteria can vary in priority and a weighted scheme would make more sense where entertainment is weighted higher than other criteria, from a role-playing perspective, users might value character consistency over fluency, or vice versa.\n\nAlthough authors have mentioned these in limitations but I would highlight that with only 64 conversations per model, the benchmark’s robustness is very limited, While the authors report a positive correlation with human annotations, they used only a single human annotator, which is a significant limitation. Having a single annotator introduces subjective biases to a subjective dimension like entertainment.", "questions": "My suggestions would be to experiment with weighting or adjusting criteria based on specific user feedback, perhaps allowing users to prioritize different aspects like consistency or entertainment. Also, Increasing the diversity of human annotations should help validate the scores against a more reliable ground truth of human judgment.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces PingPong a benchmark that aims to simulate and assess multi-turn interactions using three components Player, Interrogator, and Judge models. The authors have focused on role-playing models for entertainment purposes. They do this in two versions: In the first version the judge and the interrogator are played by a single model while in the second version these roles are separated in two different models. The player is provided a character card defining its role while the interrogator has the details of the scenario. The judge is supposed to score each turn based on 3 criteria: entertainment, character consistency and language fluency.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "The authors have focused on role-playing models tailored for entertainment, which is an underrepresented area in benchmarks and that too in multi-turn settings.", "weaknesses": "I have many concerns with this paper. A Judge which is itself an LLM with inherent biases is assessing a highly subjective quality like “Entertainment”. Measuring entertainment is not straightforward and can have varying stylistic and cultural traits. Evaluating that without a human reference data compounds this issue and thus the reliability of the judge can’t be established. Similar concerns with character consistency.\n\nIn role-playing, each turn can be dependent on prior turns, which can’t be fully captured by scoring turns in isolation. While scoring each turn provides a granular view of performance, it may miss the overarching coherence of the character and storyline across multiple turns. The evaluation also overlooks user-centric metrics like engagement, user satisfaction, ability to sustain engagement over extended interactions which are important for role-playing. The paper’s current scoring approach does not seem to assess these aspects. Also these criteria can vary in priority and a weighted scheme would make more sense where entertainment is weighted higher than other criteria, from a role-playing perspective, users might value character consistency over fluency, or vice versa.\n\nAlthough authors have mentioned these in limitations but I would highlight that with only 64 conversations per model, the benchmark’s robustness is very limited, While the authors report a positive correlation with human annotations, they used only a single human annotator, which is a significant limitation. Having a single annotator introduces subjective biases to a subjective dimension like entertainment.", "questions": "My suggestions would be to experiment with weighting or adjusting criteria based on specific user feedback, perhaps allowing users to prioritize different aspects like consistency or entertainment. Also, Increasing the diversity of human annotations should help validate the scores against a more reliable ground truth of human judgment.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731039280023}, {"id": "Nsse2jcCvE", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_oV7o"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 3, "summary": "This work presents a novel benchmark for assessing language models' role-playing abilities in dynamic, multi-turn dialogues. The evaluation framework includes three components: a player model embodying a specific character, an interrogator model simulating user interactions, and a judge model assessing dialogue quality. Experiments showed strong correlations between automated and human evaluations, supporting the framework's reliability. This benchmark lays the groundwork for robust and adaptive evaluations of model performance in interactive contexts.", "review_text": "This work presents a novel benchmark for assessing language models' role-playing abilities in dynamic, multi-turn dialogues. The evaluation framework includes three components: a player model embodying a specific character, an interrogator model simulating user interactions, and a judge model assessing dialogue quality. Experiments showed strong correlations between automated and human evaluations, supporting the framework's reliability. This benchmark lays the groundwork for robust and adaptive evaluations of model performance in interactive contexts.", "strengths": "This work introduces the concept of an \"Interrogator,\" which serves as a user simulator. Unlike traditional static evaluation, dynamic evaluation—incorporating both the user simulator and AI character—offers a more realistic assessment. This approach holds significant value.", "weaknesses": "While this work has a strong starting point, it lacks rigorous experimental validation in several areas. For example:\n\n1. The authors have not adequately addressed the consistency between “Interrogators” and real-world human users. In practical scenarios, users typically employ informal language with various omissions and slang. Additionally, their motivations for engaging with a character are often unpredictable. Thus, a deeper examination of the alignment between “Interrogators” and human users would significantly enhance the quality of this work.\n\n2. Point-wise evaluations by Large Language Models often diverge from human annotators’ assessments, especially in subjective tasks. Furthermore, the generated scores tend to be biased towards specific values, resulting in a leaderboard that lacks differentiation.", "questions": "Typos in Table 1  and Table 2: Enteraining -> Entertaining", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work presents a novel benchmark for assessing language models' role-playing abilities in dynamic, multi-turn dialogues. The evaluation framework includes three components: a player model embodying a specific character, an interrogator model simulating user interactions, and a judge model assessing dialogue quality. Experiments showed strong correlations between automated and human evaluations, supporting the framework's reliability. This benchmark lays the groundwork for robust and adaptive evaluations of model performance in interactive contexts.", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "This work introduces the concept of an \"Interrogator,\" which serves as a user simulator. Unlike traditional static evaluation, dynamic evaluation—incorporating both the user simulator and AI character—offers a more realistic assessment. This approach holds significant value.", "weaknesses": "While this work has a strong starting point, it lacks rigorous experimental validation in several areas. For example:\n\n1. The authors have not adequately addressed the consistency between “Interrogators” and real-world human users. In practical scenarios, users typically employ informal language with various omissions and slang. Additionally, their motivations for engaging with a character are often unpredictable. Thus, a deeper examination of the alignment between “Interrogators” and human users would significantly enhance the quality of this work.\n\n2. Point-wise evaluations by Large Language Models often diverge from human annotators’ assessments, especially in subjective tasks. Furthermore, the generated scores tend to be biased towards specific values, resulting in a leaderboard that lacks differentiation.", "questions": "Typos in Table 1  and Table 2: Enteraining -> Entertaining", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730868853275}, {"id": "SbupzQbkrU", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_8Pvv"], "rating": 3, "soundness": 2, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This work examines the role-playing capabilities of language models with a benchmark that uses LMs to emulate specified characters and users in multi-turn conversations and also judge these conversations. The authors validate this framework by comparing the automated evaluations with human annotations and showing strong correlations across various criteria. The authors show that ensembling model judgements lead to better correlation with human judgement on the criteria of fluency, character consistency, and entertainment.", "review_text": "This work examines the role-playing capabilities of language models with a benchmark that uses LMs to emulate specified characters and users in multi-turn conversations and also judge these conversations. The authors validate this framework by comparing the automated evaluations with human annotations and showing strong correlations across various criteria. The authors show that ensembling model judgements lead to better correlation with human judgement on the criteria of fluency, character consistency, and entertainment.", "strengths": "- This work addresses issues with prior work that examines role-playing capabilities with either single-turn interactions or using static datasets that may have issues with data contamination.", "weaknesses": "- There are no comparisons to other evaluation benchmarks other than creative writing, which is an odd choice given the mention of other role-playing benchmarks with single-turn evaluations, and therefore the value added by this benchmark is not substantiated by previous work on role-playing evaluation and results. In addition, results are descriptive, rather than analytical. It is unclear whether any of the results from this benchmark is interesting or surprising. \n- There is no explanation on what makes this dataset dynamic while previous efforts are considered static. \n- While correlation using a multi-model setup shows higher correlation with human annotations, the human annotations were done by a single person and the margin with a single-model setup is not big enough to motivate the use of multiple models given that it would incur higher costs. \n- The paper is written poorly. Please refer to details in the Questions section.", "questions": "- lines 30-32: what are these other applications?\n- lines 33-34: why do you believe so? What are the alternatives that were studied before?\n- lines 35-36: provide citations for these popular benchmarks. It shouldn't be as thorough as the related work section, but each claim should be backed by a citation or by empirical results from the current paper.\n- Introduction: 'novelty' is repeatedly mentioned, but it is unclear what the novelty is. How is your LLM-as-a-judge different from prior work?\n- line 46: what is meant by dynamic? What is meant by data contamination in this context? Give a brief summary in what your methodology is for generating dynamic questions as opposed to static ones.\n- End of introduction: give a brief summary of what the novel and interesting findings are that were enabled by this proposed benchmark\n- Related work: it has too many subsections, which makes this section  feel disconnected. If role-playing is the most important aspect of this work, I'd suggest starting with them and how the other aspects (static vs dynamic, multi-turn, data contamination, multi-model judges) are related to a more realistic evaluation of role-playing capabilities.\n- What is meant by asymmetrical in line 136? Do you mean that the player only gets the character description while the interrogator only gets the situation information? Are there any concerns about the base persona of the interrogator being a confounding factor for the player's ability to role-play?\n- What's the meaning of \"separated soles\" in line 166?\n- What were the limitations of the combined approach in line 168? I see that this is explained later. I would suggest rewording this sentence so that the key issues of the combined approach is introduced first or mentioned even in section 3.3 as to motivate section 3.4.\n- How important is it to introduce version 1 (section 3.3)? This feels less important and thus can be deferred to the appendix.\n- line 192: what are the 16 language models?\n- line 194: using a single annotator is not sufficient for measuring reliable correlation with a language model's scores because it's not representative of human judgement.\n- What's the human performance on this role-playing task?\n- Apart from the quantitative results of the leaderboard, what are the interesting findings that are revealed through this benchmark that was not known before? Are they different from the results on static, single-turn benchmarks?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work examines the role-playing capabilities of language models with a benchmark that uses LMs to emulate specified characters and users in multi-turn conversations and also judge these conversations. The authors validate this framework by comparing the automated evaluations with human annotations and showing strong correlations across various criteria. The authors show that ensembling model judgements lead to better correlation with human judgement on the criteria of fluency, character consistency, and entertainment.", "soundness": 2, "presentation": 1, "contribution": 2, "strengths": "- This work addresses issues with prior work that examines role-playing capabilities with either single-turn interactions or using static datasets that may have issues with data contamination.", "weaknesses": "- There are no comparisons to other evaluation benchmarks other than creative writing, which is an odd choice given the mention of other role-playing benchmarks with single-turn evaluations, and therefore the value added by this benchmark is not substantiated by previous work on role-playing evaluation and results. In addition, results are descriptive, rather than analytical. It is unclear whether any of the results from this benchmark is interesting or surprising. \n- There is no explanation on what makes this dataset dynamic while previous efforts are considered static. \n- While correlation using a multi-model setup shows higher correlation with human annotations, the human annotations were done by a single person and the margin with a single-model setup is not big enough to motivate the use of multiple models given that it would incur higher costs. \n- The paper is written poorly. Please refer to details in the Questions section.", "questions": "- lines 30-32: what are these other applications?\n- lines 33-34: why do you believe so? What are the alternatives that were studied before?\n- lines 35-36: provide citations for these popular benchmarks. It shouldn't be as thorough as the related work section, but each claim should be backed by a citation or by empirical results from the current paper.\n- Introduction: 'novelty' is repeatedly mentioned, but it is unclear what the novelty is. How is your LLM-as-a-judge different from prior work?\n- line 46: what is meant by dynamic? What is meant by data contamination in this context? Give a brief summary in what your methodology is for generating dynamic questions as opposed to static ones.\n- End of introduction: give a brief summary of what the novel and interesting findings are that were enabled by this proposed benchmark\n- Related work: it has too many subsections, which makes this section  feel disconnected. If role-playing is the most important aspect of this work, I'd suggest starting with them and how the other aspects (static vs dynamic, multi-turn, data contamination, multi-model judges) are related to a more realistic evaluation of role-playing capabilities.\n- What is meant by asymmetrical in line 136? Do you mean that the player only gets the character description while the interrogator only gets the situation information? Are there any concerns about the base persona of the interrogator being a confounding factor for the player's ability to role-play?\n- What's the meaning of \"separated soles\" in line 166?\n- What were the limitations of the combined approach in line 168? I see that this is explained later. I would suggest rewording this sentence so that the key issues of the combined approach is introduced first or mentioned even in section 3.3 as to motivate section 3.4.\n- How important is it to introduce version 1 (section 3.3)? This feels less important and thus can be deferred to the appendix.\n- line 192: what are the 16 language models?\n- line 194: using a single annotator is not sufficient for measuring reliable correlation with a language model's scores because it's not representative of human judgement.\n- What's the human performance on this role-playing task?\n- Apart from the quantitative results of the leaderboard, what are the interesting findings that are revealed through this benchmark that was not known before? Are they different from the results on static, single-turn benchmarks?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730765511556}, {"id": "UkfrSuSsLF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_Fuez"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The paper introduces a \"benchmark\" for role playing dialog. It involves 3 LLM's playing different roles\n1. a user/interogator LLM which talks to \n2. a system LLM playing a character role, and\n3. a Judge LLM which looks at the resulting conversation between 1 and 2, and grades how well 2 has played the assigned character role\n\nThe paper uses state-of-the art LLMs for these, releases some code. The contribution is minor however for reasons given below.", "review_text": "The paper introduces a \"benchmark\" for role playing dialog. It involves 3 LLM's playing different roles\n1. a user/interogator LLM which talks to \n2. a system LLM playing a character role, and\n3. a Judge LLM which looks at the resulting conversation between 1 and 2, and grades how well 2 has played the assigned character role\n\nThe paper uses state-of-the art LLMs for these, releases some code. The contribution is minor however for reasons given below.", "strengths": "The contributed code may provide a framework for some members of the community to experiment with. However there are no scientific questions posed in this paper, it's a limited engineering style contribution with observations -- such as separating of judge and user LLM, which was well motivated and made sense -- on how to construct such a simulation environment.", "weaknesses": "There is very limited novelty in this submission. This is a basic simulation system these days, and multiple other papers have performed similar setups with LLM's playing conversation roles. Even if application to a role playing character is new, it's a minor increment.  \n\nAside from that, there are a very small number of conversations generated here (60), and of greater concern they are evaluated only by 1 human grader, who has apparently limited English abilities (mentioned in results section) which limits them from noticing any nuances in the dialog.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a \"benchmark\" for role playing dialog. It involves 3 LLM's playing different roles\n1. a user/interogator LLM which talks to \n2. a system LLM playing a character role, and\n3. a Judge LLM which looks at the resulting conversation between 1 and 2, and grades how well 2 has played the assigned character role\n\nThe paper uses state-of-the art LLMs for these, releases some code. The contribution is minor however for reasons given below.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The contributed code may provide a framework for some members of the community to experiment with. However there are no scientific questions posed in this paper, it's a limited engineering style contribution with observations -- such as separating of judge and user LLM, which was well motivated and made sense -- on how to construct such a simulation environment.", "weaknesses": "There is very limited novelty in this submission. This is a basic simulation system these days, and multiple other papers have performed similar setups with LLM's playing conversation roles. Even if application to a role playing character is new, it's a minor increment.  \n\nAside from that, there are a very small number of conversations generated here (60), and of greater concern they are evaluated only by 1 human grader, who has apparently limited English abilities (mentioned in results section) which limits them from noticing any nuances in the dialog.", "questions": "NA", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730741666300}, {"id": "rxd9atrdSt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_ncgD"], "rating": 3, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This work proposes a multi-turn, dynamic and multimodel benchmark for assessing the role-playing abilities of language models. Their framework depends on three components: player, interrogator and judge. The authors compare the automatic vs human scores for Russian and English. Additionally, they compare their results with a Creative Writing benchmark.", "review_text": "This work proposes a multi-turn, dynamic and multimodel benchmark for assessing the role-playing abilities of language models. Their framework depends on three components: player, interrogator and judge. The authors compare the automatic vs human scores for Russian and English. Additionally, they compare their results with a Creative Writing benchmark.", "strengths": "* The community has put much effort into similar goals: automatic evaluation and setting benchmarks.\n* This benchmark can be the seed for further investigation of role-playing capabilities.\n* The benchmark is automatic and could be easily reproduced.", "weaknesses": "* There was only one annotator. \n* The relationship between the annotator and the authors was not disclosed.\n* The instructions given to the annotator were not disclosed. \n* The elements of the evaluation (e.g., annotation aspects and their Likert scale) were not discussed. \n* Comparison between v1 and v2 is not thorough since only one model was used on v2. \n* The motivation for comparing with creative writing is not clear.", "questions": "* Can you describe the role of the human annotator? Which profile did he have (author, student, extern), and which instructions was he/she given? How much was he/she paid? How long did it take to annotate? \n* What was the motivation for using the creative writing benchmark?\n* Why are the scores too close to each other? Can this be improved so the differences among LLMs can be better quantified?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work proposes a multi-turn, dynamic and multimodel benchmark for assessing the role-playing abilities of language models. Their framework depends on three components: player, interrogator and judge. The authors compare the automatic vs human scores for Russian and English. Additionally, they compare their results with a Creative Writing benchmark.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "* The community has put much effort into similar goals: automatic evaluation and setting benchmarks.\n* This benchmark can be the seed for further investigation of role-playing capabilities.\n* The benchmark is automatic and could be easily reproduced.", "weaknesses": "* There was only one annotator. \n* The relationship between the annotator and the authors was not disclosed.\n* The instructions given to the annotator were not disclosed. \n* The elements of the evaluation (e.g., annotation aspects and their Likert scale) were not discussed. \n* Comparison between v1 and v2 is not thorough since only one model was used on v2. \n* The motivation for comparing with creative writing is not clear.", "questions": "* Can you describe the role of the human annotator? Which profile did he have (author, student, extern), and which instructions was he/she given? How much was he/she paid? How long did it take to annotate? \n* What was the motivation for using the creative writing benchmark?\n* Why are the scores too close to each other? Can this be improved so the differences among LLMs can be better quantified?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730706879671}, {"id": "hHtQgdoggF", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2165/Reviewer_3quE"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper introduces a benchmark to evaluate language models' role-playing abilities in dynamic, multi-turn conversations. It features a unique three-part framework: a player (the language model in a character role), an interrogator (emulating user interactions), and a judge (assessing dialogue quality). A multi-model evaluation strategy uses various language models as judges to reduce bias, aligning well with human evaluations.", "review_text": "This paper introduces a benchmark to evaluate language models' role-playing abilities in dynamic, multi-turn conversations. It features a unique three-part framework: a player (the language model in a character role), an interrogator (emulating user interactions), and a judge (assessing dialogue quality). A multi-model evaluation strategy uses various language models as judges to reduce bias, aligning well with human evaluations.", "strengths": "By using dynamic, multi-turn interactions that mimic the unpredictable flow of real conversations, the benchmark does a great job of capturing authentic role-playing scenarios.\n\nThe benchmark supports both English and Russian for now, but its flexible setup suggests it could easily expand to other languages. This forward-thinking design could make it a valuable tool for building models that are more culturally and linguistically inclusive.\n\nA standout feature of this benchmark is its use of language models not only as players but also as simulated users and judges. This design boosts scalability and provides a consistent, less biased way to evaluate huge datasets, making it possible to explore different role-playing interactions without needing a lot of human input every time.", "weaknesses": "Given that budget limitations kept the sample size small, it would be helpful if the paper discussed how scaling up the tests might affect costs and computational resources. This would be useful for readers who are looking to use or expand on this benchmark.\n\nThe paper does touch on ethics broadly, but a more in-depth look at the ethical issues specific to role-playing language models would be valuable—especially when it comes to handling sensitive or potentially harmful content. Examining how well the models respect ethical boundaries, respond to user distress, or navigate social nuances could add key safety considerations to the benchmark.", "questions": "The paper shows how the benchmark works in both English and Russian, but how feasible would it be to extend it to other languages and cultural contexts? Have the authors considered specific challenges in keeping results consistent across models with different linguistic backgrounds?\n\nThe paper focuses on metrics like fluency, character consistency, and entertainment value, but would the authors consider adding other metrics to measure contextual understanding? For instance, it could be useful to evaluate how well a model keeps up with a storyline or handles unexpected, non-linear questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a benchmark to evaluate language models' role-playing abilities in dynamic, multi-turn conversations. It features a unique three-part framework: a player (the language model in a character role), an interrogator (emulating user interactions), and a judge (assessing dialogue quality). A multi-model evaluation strategy uses various language models as judges to reduce bias, aligning well with human evaluations.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "By using dynamic, multi-turn interactions that mimic the unpredictable flow of real conversations, the benchmark does a great job of capturing authentic role-playing scenarios.\n\nThe benchmark supports both English and Russian for now, but its flexible setup suggests it could easily expand to other languages. This forward-thinking design could make it a valuable tool for building models that are more culturally and linguistically inclusive.\n\nA standout feature of this benchmark is its use of language models not only as players but also as simulated users and judges. This design boosts scalability and provides a consistent, less biased way to evaluate huge datasets, making it possible to explore different role-playing interactions without needing a lot of human input every time.", "weaknesses": "Given that budget limitations kept the sample size small, it would be helpful if the paper discussed how scaling up the tests might affect costs and computational resources. This would be useful for readers who are looking to use or expand on this benchmark.\n\nThe paper does touch on ethics broadly, but a more in-depth look at the ethical issues specific to role-playing language models would be valuable—especially when it comes to handling sensitive or potentially harmful content. Examining how well the models respect ethical boundaries, respond to user distress, or navigate social nuances could add key safety considerations to the benchmark.", "questions": "The paper shows how the benchmark works in both English and Russian, but how feasible would it be to extend it to other languages and cultural contexts? Have the authors considered specific challenges in keeping results consistent across models with different linguistic backgrounds?\n\nThe paper focuses on metrics like fluency, character consistency, and entertainment value, but would the authors consider adding other metrics to measure contextual understanding? For instance, it could be useful to evaluate how well a model keeps up with a storyline or handles unexpected, non-linear questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730186469027}], "openreview_url": "https://openreview.net/forum?id=996aKQIom0", "arxiv_id": "2409.06820", "paper_pdf": "papers/996aKQIom0.pdf", "paper_pdf_sha256": "2ae8962d19fe1d7fc1fa37655a36874da21a01792ac9f62f5c7d95a688166ae5", "paper_pdf_bytes": 1389322, "paper_pdf_source": "openreview", "code_url": "https://github.com/IlyaGusev/ping_pong_bench", "code_repository": "IlyaGusev/ping_pong_bench", "code_commit": "4e3b0e0b6b093670f67b80d7d62b5a5e8e7b13fd", "code_archive": "repos/996aKQIom0.zip", "code_archive_sha256": "820ead7e8ec5e0157c0ef08d3c83cbf1af0ecef7b8099a373d69aefc1577b0ee", "code_archive_bytes": 290873, "code_file_count": 22, "code_extensions": {".py": 22}, "github_disk_usage_kb": 429, "github_languages": {"Python": 100806, "Jinja": 13368, "HTML": 9651}, "github_archived": false, "github_pushed_at": "2025-05-25T20:17:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/pingpong-a-benchmark-for-role-playing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "E296x0YpML", "year": 2024, "status": "rejected", "title": "Fooling the Textual Fooler via Randomizing Latent Representations", "authors": ["Duy Cao Hoang", "Nguyen Hung-Quang", "Saurav Manchanda", "Minlong Peng", "Kok-Seng Wong", "Khoa D Doan"], "authorids": ["~Duy_Cao_Hoang1", "~Nguyen_Hung-Quang1", "~Saurav_Manchanda1", "~Minlong_Peng1", "~Kok-Seng_Wong1", "~Khoa_D_Doan1"], "authors_source": "OpenReview API", "abstract": "Despite outstanding performance in a variety of NLP tasks, recent studies have revealed that NLP models are vulnerable to adversarial attacks that slightly perturb the input to cause the models to misbehave. Among these attacks, adversarial word-level perturbations are well-studied and effective attack strategies. These attacks involve querying the victim model many times to determine the most important words in an input text and to replace these words with their corresponding synonyms. Query-based attacks as such work in black-box settings, which can be detrimental to NLP applications that can be accessed publicly. In this work, we propose a lightweight and attack-agnostic defense whose main goal is to perplex the process of generating an adversarial example in these query-based black-box attacks; that is to fool the textual fooler. This defense, named AdvFooler, works by randomizing the latent representation of the input at inference time. Different from existing defenses, AdvFooler does not necessitate additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set while having a negligible impact on the model's accuracy. Our theoretical and empirical analyses highlight the significance of robustness resulting from confusing the adversary via randomizing the latent space, as well as the impact of randomization on clean accuracy. Finally, we empirically demonstrate the near state-of-the-art robustness of AdvFooler against representative adversarial word-level attacks on two benchmark datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "EtDtC1WTsD", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9033/Reviewer_UMzy"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The authors observe that adding noise to hidden layers can result in obscuring identification of important words in a sentence, a crucial step needed for launching attacks. They evaluate their proposal and highlight efficacy.", "review_text": "The authors observe that adding noise to hidden layers can result in obscuring identification of important words in a sentence, a crucial step needed for launching attacks. They evaluate their proposal and highlight efficacy.", "strengths": "1. Low cost defense.", "weaknesses": "1. Does not perform better than prior works.\n2. Similar in ideology to DP.", "questions": "1. Could the authors compare and contrast their approach to applying DP-style noise to the embeddings generated by the hidden layer?\n2. Results from Table 2 suggest that their approach is not the very best compared to other approaches. What are the merits of the proposal then? Apart from reduced computational overheads?\n3. The authors empirically note that adding small amounts of noise to the logins does not change the prediction. But this is not fundamental nor clear why this is the case. Could the authors elaborate?\n4. Why can’t the adversary utilize a proxy model to launch its attacks instead of the current black-box setup i.e., use the proxy model and attention values to identify important words?\n5. Alternatively, the adversary could also replace subsets of words using brute force. For small spans of text (as considered in the evaluation), this is not computationally prohibitive. Could the authors elaborate further the specific scenarios where important word selection inhibition is prudent?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors observe that adding noise to hidden layers can result in obscuring identification of important words in a sentence, a crucial step needed for launching attacks. They evaluate their proposal and highlight efficacy.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. Low cost defense.", "weaknesses": "1. Does not perform better than prior works.\n2. Similar in ideology to DP.", "questions": "1. Could the authors compare and contrast their approach to applying DP-style noise to the embeddings generated by the hidden layer?\n2. Results from Table 2 suggest that their approach is not the very best compared to other approaches. What are the merits of the proposal then? Apart from reduced computational overheads?\n3. The authors empirically note that adding small amounts of noise to the logins does not change the prediction. But this is not fundamental nor clear why this is the case. Could the authors elaborate?\n4. Why can’t the adversary utilize a proxy model to launch its attacks instead of the current black-box setup i.e., use the proxy model and attention values to identify important words?\n5. Alternatively, the adversary could also replace subsets of words using brute force. For small spans of text (as considered in the evaluation), this is not computationally prohibitive. Could the authors elaborate further the specific scenarios where important word selection inhibition is prudent?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1699326512346}, {"id": "tpOkjYR7yc", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9033/Reviewer_9Kvp"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a lightweight and attack-agnostic defense method to against query-based black-box attacks called AdvFooler. AdvFooler accomplishes this by introducing randomization to the latent input representation during inference.  The advantage of Advfooler is that it does not need additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set.", "review_text": "This paper proposes a lightweight and attack-agnostic defense method to against query-based black-box attacks called AdvFooler. AdvFooler accomplishes this by introducing randomization to the latent input representation during inference.  The advantage of Advfooler is that it does not need additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set.", "strengths": "1. Compared with other defense methods, the proposed method AdvFooler is simple, pluggable and does not require additional computational overhead during testing or access to training.\n2. The authors conducted comprehensive experiments to assess the effectiveness of AdvFooler, employing two BERT models, two distinct datasets, and three different attack methods. Furthermore, they provided qualitative analyses of their results.", "weaknesses": "1. Despite its advantages in terms of simple implementation and minimal computational overhead, AdvFooler's performance falls short of the state-of-the-art. In Table 2, on the AGNEWS dataset, AdvFooler exhibits lower accuracy under attack compared to RanMASK for the BERT-base model, and it also demonstrates lower accuracy under attack than both TMD and RanMASK for the RoBERTa-base model.\n\n2. The selection of the hyper-parameter for noise scale in AdvFooler is not entirely clear. The authors claim that they choose the ν value based on the criterion that the clean accuracy drops by at most 1% using the test set. However, it seems not the case in Table 2 and 3. In Figure 4, the curves of AuA are not monotone. Do authors also consider AuA values when choosing noise scale?", "questions": "Instead of adding noises to all layers by the same noise scale, what would the results be if adding different noise scale to different layers?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a lightweight and attack-agnostic defense method to against query-based black-box attacks called AdvFooler. AdvFooler accomplishes this by introducing randomization to the latent input representation during inference.  The advantage of Advfooler is that it does not need additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. Compared with other defense methods, the proposed method AdvFooler is simple, pluggable and does not require additional computational overhead during testing or access to training.\n2. The authors conducted comprehensive experiments to assess the effectiveness of AdvFooler, employing two BERT models, two distinct datasets, and three different attack methods. Furthermore, they provided qualitative analyses of their results.", "weaknesses": "1. Despite its advantages in terms of simple implementation and minimal computational overhead, AdvFooler's performance falls short of the state-of-the-art. In Table 2, on the AGNEWS dataset, AdvFooler exhibits lower accuracy under attack compared to RanMASK for the BERT-base model, and it also demonstrates lower accuracy under attack than both TMD and RanMASK for the RoBERTa-base model.\n\n2. The selection of the hyper-parameter for noise scale in AdvFooler is not entirely clear. The authors claim that they choose the ν value based on the criterion that the clean accuracy drops by at most 1% using the test set. However, it seems not the case in Table 2 and 3. In Figure 4, the curves of AuA are not monotone. Do authors also consider AuA values when choosing noise scale?", "questions": "Instead of adding noises to all layers by the same noise scale, what would the results be if adding different noise scale to different layers?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698834806669}, {"id": "kB9rx2EOwQ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9033/Reviewer_bLP9"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "In this work, the authors propose a simple defense method AdvFooler against textual adversarial attacks. Specifically, AdvFooler adds a Gaussian noise at each layer for forward propagation to randomize the latent defense. With such randomization, the attacker cannot find significant words for substitution, making the model robust.", "review_text": "In this work, the authors propose a simple defense method AdvFooler against textual adversarial attacks. Specifically, AdvFooler adds a Gaussian noise at each layer for forward propagation to randomize the latent defense. With such randomization, the attacker cannot find significant words for substitution, making the model robust.", "strengths": "1. The paper is well-written and easy to follow.\n\n2. AdvFooler is simple and seems to be effective against several attacks.", "weaknesses": "1. The motivation is not clear. Why does such randomization fool the attackers while not degrading the benign performance?\n\n2. Why does AdvFooler can only perplex query-based black-box attacks? It is significant for a defense method to defend against various attacks, such as white-box attacks [1], decision-based attacks [2,3], and so on. It is necessary to validate the effectiveness against these attacks to show the generality of AdvFooler.\n\n3. AdvFooler does not outperform the SOTA baselines against various attacks.\n\n4. From Figure 4, it might be hard to choose a consistent noise scale for different datasets and models.\n\n[1] Wang et al. Adversarial Training with Fast Gradient Projection Method against Synonym Substitution based Text Attacks. AAAI 2021.\n\n[2] Maheshwary et al. Generating Natural Language Attacks in a Hard Label Black Box Setting. AAAI 2021.\n\n[3] Yu et al. TextHacker: Learning based Hybrid Local Search Algorithm for Text Hard-label Adversarial Attack. EMNLP 2022.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the authors propose a simple defense method AdvFooler against textual adversarial attacks. Specifically, AdvFooler adds a Gaussian noise at each layer for forward propagation to randomize the latent defense. With such randomization, the attacker cannot find significant words for substitution, making the model robust.", "soundness": "2 fair", "presentation": "4 excellent", "contribution": "2 fair", "strengths": "1. The paper is well-written and easy to follow.\n\n2. AdvFooler is simple and seems to be effective against several attacks.", "weaknesses": "1. The motivation is not clear. Why does such randomization fool the attackers while not degrading the benign performance?\n\n2. Why does AdvFooler can only perplex query-based black-box attacks? It is significant for a defense method to defend against various attacks, such as white-box attacks [1], decision-based attacks [2,3], and so on. It is necessary to validate the effectiveness against these attacks to show the generality of AdvFooler.\n\n3. AdvFooler does not outperform the SOTA baselines against various attacks.\n\n4. From Figure 4, it might be hard to choose a consistent noise scale for different datasets and models.\n\n[1] Wang et al. Adversarial Training with Fast Gradient Projection Method against Synonym Substitution based Text Attacks. AAAI 2021.\n\n[2] Maheshwary et al. Generating Natural Language Attacks in a Hard Label Black Box Setting. AAAI 2021.\n\n[3] Yu et al. TextHacker: Learning based Hybrid Local Search Algorithm for Text Hard-label Adversarial Attack. EMNLP 2022.", "questions": "See weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1697874823826}, {"id": "dJL6OPTttJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission9033/Reviewer_2d4f"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "3 good", "contribution": "4 excellent", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "1. GOAL:\n\n* 1A. The problem statement:\nIt is possible to do black box word level attacks by determining the most important words in the input by repeated query and replacing them (with synonyms).\n\n* 1B. What their solution is:\nconfuse the adversarial black-box attack with a system called \"AdvFooler\" that randomizes the latent space. \n\n\n\n2. METHOD:\nThe latent space is randomized by adding an independent gaussian noise vector to the l-th layer of a pretrained classification model. The magnitude of noise is selected based on how much the clean accuracy drops by a chosen percentage on a small held out clean set (1% drop is used in paper). Though it’s not actually stated in the paper, I assume this noise is added during evaluation time (so each adversarial query yields slightly different perturbed outputs, and the attack has a hard time figuring out what words to perturb in it’s attack).\n\nClaims Made:\n- does not need more compute during training (True)\n- does not rely on assumptions about adversarial perturbation word set for a word substitution attack (True)\n- theoretical and empirical results presented on adversarial set and clean accuracy (True, they are presented)\n\n3. RESULTS:\nThere are many metrics to evaluate on, so we'll go metric by metric: high clean accuracy (A), middling robustness-accuracy tradeoff (B/C), great implementation ease (D), and (probably) low compute time needed which is great (E)\n\n* 3A. By design the clean accuracy does not suffer more than 1% on their method (when used alone), and they show an example where they are still able to change the most significant words.\n\n* 3B. Their method isn't the best method in terms of robustness and accuracy tradeoff, but it is reasonably robust (i.e. does improve robustness significantly). Conclusions based on these comparisons are hard to make to be honest. More details in the Weaknesses section.\n\n* 3C. Some results in Fig 4 demonstrate the tradeoff between clean accuracy and adversarial attack accuracy for this method specifically. Is is unclear why the clean accuracy drops so little at really high noise rates, and that's not really explained.\n\n* 3D. Where their method absolutely shines is how simple the method is to implement and use based on reading the methods section alone. They don't highlight ease of implementation in the paper, but it's absolutely something that would make this method much easier to adopt. \n\n* 3E. They repeatedly highlight low added compute as a big win for this method. They mention they have results about compute usage across the methods in Supplemental Materials, but I was not able to find these. Please add. Still reading the methods section gives a clear picture that the added compute for this method is very negligible. (It's unclear to me if some adversarial training baseline methods are also reasonably negligible in terms of compute time. Computing synonyms is very cheap, and training on some added adversarial examples is also not marginally more expensive in any meaningful way).", "review_text": "1. GOAL:\n\n* 1A. The problem statement:\nIt is possible to do black box word level attacks by determining the most important words in the input by repeated query and replacing them (with synonyms).\n\n* 1B. What their solution is:\nconfuse the adversarial black-box attack with a system called \"AdvFooler\" that randomizes the latent space. \n\n\n\n2. METHOD:\nThe latent space is randomized by adding an independent gaussian noise vector to the l-th layer of a pretrained classification model. The magnitude of noise is selected based on how much the clean accuracy drops by a chosen percentage on a small held out clean set (1% drop is used in paper). Though it’s not actually stated in the paper, I assume this noise is added during evaluation time (so each adversarial query yields slightly different perturbed outputs, and the attack has a hard time figuring out what words to perturb in it’s attack).\n\nClaims Made:\n- does not need more compute during training (True)\n- does not rely on assumptions about adversarial perturbation word set for a word substitution attack (True)\n- theoretical and empirical results presented on adversarial set and clean accuracy (True, they are presented)\n\n3. RESULTS:\nThere are many metrics to evaluate on, so we'll go metric by metric: high clean accuracy (A), middling robustness-accuracy tradeoff (B/C), great implementation ease (D), and (probably) low compute time needed which is great (E)\n\n* 3A. By design the clean accuracy does not suffer more than 1% on their method (when used alone), and they show an example where they are still able to change the most significant words.\n\n* 3B. Their method isn't the best method in terms of robustness and accuracy tradeoff, but it is reasonably robust (i.e. does improve robustness significantly). Conclusions based on these comparisons are hard to make to be honest. More details in the Weaknesses section.\n\n* 3C. Some results in Fig 4 demonstrate the tradeoff between clean accuracy and adversarial attack accuracy for this method specifically. Is is unclear why the clean accuracy drops so little at really high noise rates, and that's not really explained.\n\n* 3D. Where their method absolutely shines is how simple the method is to implement and use based on reading the methods section alone. They don't highlight ease of implementation in the paper, but it's absolutely something that would make this method much easier to adopt. \n\n* 3E. They repeatedly highlight low added compute as a big win for this method. They mention they have results about compute usage across the methods in Supplemental Materials, but I was not able to find these. Please add. Still reading the methods section gives a clear picture that the added compute for this method is very negligible. (It's unclear to me if some adversarial training baseline methods are also reasonably negligible in terms of compute time. Computing synonyms is very cheap, and training on some added adversarial examples is also not marginally more expensive in any meaningful way).", "strengths": "The primary and significant strength of this paper is that their method is dead simple to implement and use, and takes almost no added compute time. Even if it's not SOTA (and it doesn't actually claim to be SOTA), it is strongly worth disseminating (after adding some missing results: see Weaknesses). \n\nIn much the way drop out rates approximate ensemble training and save us all compute time, this method approximates diverse ensembles used to achieve adversarial robustness (ex: the ADP regularizer  https://proceedings.mlr.press/v97/pang19a), and can save some compute time. Because it's so easy to implement and tune for accuracy, it would have high practical impact -- for that reason alone, it's worth disseminating (after some weaknesses are addressed, see below).", "weaknesses": "3 weaknesses and 2 baseline suggestions, enumerated below as 1-5. I am very amenable to changing my decision to an accept if these are addressed (esp. 1 and 3, with a very strong suggestion for 4 and 2).\n\n1. It's really hard to draw conclusions about robustness-accuracy tradeoffs from Tables 2 and 3. The claim that your method performs \"close\" to SOTA needs some added substantiation. I'd like:\n\n  *  1.A) Standard deviations on at-least three runs for each setting. In many of these datasets, a 1pp improvement is significant. But accuracies on adversarial datasets often have high variance, so it's unclear how to interpret when your method isn't clearly better and often appears to be much below a handful of the other methods.\n\nMore details: TMD more reliably seems to perform better esp, on Imdb:Roberta, RanMASK and SAFER sometimes outperform on  robustness to attack/clean accuracy). It's hard to verify what \"close\" is, because there are no standard deviations, and on a few of the tasks, AdvFooler really does appear to significantly underperform a couple of the other methods by a  bit. (It's still reliably top 3 among the methods though, and I don't think you need to be better than all to have a case for publication).\n\n   * 1.B. Tune SAFER and RunMASK (tune the rate of substitution/masking) until either the robustness or the accuracy matches your method, so. you can draw a clean comparison. (Or tune your method to their accuracy if that's easier). As is, not clear at all that one is Pareto dominant (or even comparable) to the other. \n\nExample: While RunMASK does suffer on accuracy, it makes things more robust (on AGNEWS dataset BERT for instance). Based on the data available, I'd say it's a toss up whether it's better or worse than AdvFooler.\n\n\n\n2. I think the most compelling case for publication is how lightweight your system is. I would include compute time results somewhere. I was unable to find them). It's also unclear to me that simple adversarial training on word substitutions is actually significantly more compute intensive (generating synonyms is super quick, and training on more examples isn't too much of an added burden when you're already finetuning -- esp. if you're comparing to needing to tune your noise param for AdvFooler, which while cheap is still some added compute time). \n\n\n3. In your system, HotFlip is not perhaps an accurate white-box attack (Appendix C7). HotFlip relies on gradients, but are you giving it the unnoised gradients, or the parameters across all iterations of the model (so someone can reverse engineer the likely original unnoised parameters?). I suspect that if it were truly a white box attack, the certified robustness systems and TMD would outperform. I think these results as presented are a bit misleading otherwise, and one option is to remove them. I would suggest either demonstrating the white-box capabilities fully by (1) giving access to unnoised params, or each set of params so you can derive unnoised params AND (2) comparing to TMD and the certified methods. OR I'd like to see the claim about the defense being \"attack agnostic\" in the main paper, and the claims of being robust to \"white-box attack\" toned down.\n\n4. New Baseline Suggestion 1/2: While I understand that it would not be as lightweight as AdvFooler, the style of your method, suggests that a comparison with a diverse ensemble designed to induce adversarial robustness would be very apt since it also tries to induce robustness in much the same way (added random noise) (see my strengths section for the analogy). (I suggest this one: https://proceedings.mlr.press/v97/pang19a , the ADP regularizer)\n\n5. New Baseline Suggestion 2/2: Choose Dirichlet Neighborhood Ensemble (DNE)  over ASCC. While both methods model the perturbation space as the convex hull of word synonyms, DNE far outperforms ASCC (See TMD paper Table 1 where DNE nearly outperforms TMD on AGNews while maintaining a high clean accuracy).", "questions": "1. How often are are the predictions for F_advfooler the same as those for F under different noise rates? (Claim in 3.1) The chart showing adversarial robustness dropping while clean accuracy only slightly drops, gets at this question somewhat. But if possible, I'd like something that's even cleaner of a comparison that can maybe help provide intuition for why the clean accuracy doesn't really suffer much (Esp on IMDB) while adversarial robustness totally craters under high noise (Fig 4). \n\nSuggestions for added baselines (2 of them) included in the Weaknesses section (numbers 4 and 5).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "1. GOAL:\n\n* 1A. The problem statement:\nIt is possible to do black box word level attacks by determining the most important words in the input by repeated query and replacing them (with synonyms).\n\n* 1B. What their solution is:\nconfuse the adversarial black-box attack with a system called \"AdvFooler\" that randomizes the latent space. \n\n\n\n2. METHOD:\nThe latent space is randomized by adding an independent gaussian noise vector to the l-th layer of a pretrained classification model. The magnitude of noise is selected based on how much the clean accuracy drops by a chosen percentage on a small held out clean set (1% drop is used in paper). Though it’s not actually stated in the paper, I assume this noise is added during evaluation time (so each adversarial query yields slightly different perturbed outputs, and the attack has a hard time figuring out what words to perturb in it’s attack).\n\nClaims Made:\n- does not need more compute during training (True)\n- does not rely on assumptions about adversarial perturbation word set for a word substitution attack (True)\n- theoretical and empirical results presented on adversarial set and clean accuracy (True, they are presented)\n\n3. RESULTS:\nThere are many metrics to evaluate on, so we'll go metric by metric: high clean accuracy (A), middling robustness-accuracy tradeoff (B/C), great implementation ease (D), and (probably) low compute time needed which is great (E)\n\n* 3A. By design the clean accuracy does not suffer more than 1% on their method (when used alone), and they show an example where they are still able to change the most significant words.\n\n* 3B. Their method isn't the best method in terms of robustness and accuracy tradeoff, but it is reasonably robust (i.e. does improve robustness significantly). Conclusions based on these comparisons are hard to make to be honest. More details in the Weaknesses section.\n\n* 3C. Some results in Fig 4 demonstrate the tradeoff between clean accuracy and adversarial attack accuracy for this method specifically. Is is unclear why the clean accuracy drops so little at really high noise rates, and that's not really explained.\n\n* 3D. Where their method absolutely shines is how simple the method is to implement and use based on reading the methods section alone. They don't highlight ease of implementation in the paper, but it's absolutely something that would make this method much easier to adopt. \n\n* 3E. They repeatedly highlight low added compute as a big win for this method. They mention they have results about compute usage across the methods in Supplemental Materials, but I was not able to find these. Please add. Still reading the methods section gives a clear picture that the added compute for this method is very negligible. (It's unclear to me if some adversarial training baseline methods are also reasonably negligible in terms of compute time. Computing synonyms is very cheap, and training on some added adversarial examples is also not marginally more expensive in any meaningful way).", "soundness": "4 excellent", "presentation": "3 good", "contribution": "4 excellent", "strengths": "The primary and significant strength of this paper is that their method is dead simple to implement and use, and takes almost no added compute time. Even if it's not SOTA (and it doesn't actually claim to be SOTA), it is strongly worth disseminating (after adding some missing results: see Weaknesses). \n\nIn much the way drop out rates approximate ensemble training and save us all compute time, this method approximates diverse ensembles used to achieve adversarial robustness (ex: the ADP regularizer  https://proceedings.mlr.press/v97/pang19a), and can save some compute time. Because it's so easy to implement and tune for accuracy, it would have high practical impact -- for that reason alone, it's worth disseminating (after some weaknesses are addressed, see below).", "weaknesses": "3 weaknesses and 2 baseline suggestions, enumerated below as 1-5. I am very amenable to changing my decision to an accept if these are addressed (esp. 1 and 3, with a very strong suggestion for 4 and 2).\n\n1. It's really hard to draw conclusions about robustness-accuracy tradeoffs from Tables 2 and 3. The claim that your method performs \"close\" to SOTA needs some added substantiation. I'd like:\n\n  *  1.A) Standard deviations on at-least three runs for each setting. In many of these datasets, a 1pp improvement is significant. But accuracies on adversarial datasets often have high variance, so it's unclear how to interpret when your method isn't clearly better and often appears to be much below a handful of the other methods.\n\nMore details: TMD more reliably seems to perform better esp, on Imdb:Roberta, RanMASK and SAFER sometimes outperform on  robustness to attack/clean accuracy). It's hard to verify what \"close\" is, because there are no standard deviations, and on a few of the tasks, AdvFooler really does appear to significantly underperform a couple of the other methods by a  bit. (It's still reliably top 3 among the methods though, and I don't think you need to be better than all to have a case for publication).\n\n   * 1.B. Tune SAFER and RunMASK (tune the rate of substitution/masking) until either the robustness or the accuracy matches your method, so. you can draw a clean comparison. (Or tune your method to their accuracy if that's easier). As is, not clear at all that one is Pareto dominant (or even comparable) to the other. \n\nExample: While RunMASK does suffer on accuracy, it makes things more robust (on AGNEWS dataset BERT for instance). Based on the data available, I'd say it's a toss up whether it's better or worse than AdvFooler.\n\n\n\n2. I think the most compelling case for publication is how lightweight your system is. I would include compute time results somewhere. I was unable to find them). It's also unclear to me that simple adversarial training on word substitutions is actually significantly more compute intensive (generating synonyms is super quick, and training on more examples isn't too much of an added burden when you're already finetuning -- esp. if you're comparing to needing to tune your noise param for AdvFooler, which while cheap is still some added compute time). \n\n\n3. In your system, HotFlip is not perhaps an accurate white-box attack (Appendix C7). HotFlip relies on gradients, but are you giving it the unnoised gradients, or the parameters across all iterations of the model (so someone can reverse engineer the likely original unnoised parameters?). I suspect that if it were truly a white box attack, the certified robustness systems and TMD would outperform. I think these results as presented are a bit misleading otherwise, and one option is to remove them. I would suggest either demonstrating the white-box capabilities fully by (1) giving access to unnoised params, or each set of params so you can derive unnoised params AND (2) comparing to TMD and the certified methods. OR I'd like to see the claim about the defense being \"attack agnostic\" in the main paper, and the claims of being robust to \"white-box attack\" toned down.\n\n4. New Baseline Suggestion 1/2: While I understand that it would not be as lightweight as AdvFooler, the style of your method, suggests that a comparison with a diverse ensemble designed to induce adversarial robustness would be very apt since it also tries to induce robustness in much the same way (added random noise) (see my strengths section for the analogy). (I suggest this one: https://proceedings.mlr.press/v97/pang19a , the ADP regularizer)\n\n5. New Baseline Suggestion 2/2: Choose Dirichlet Neighborhood Ensemble (DNE)  over ASCC. While both methods model the perturbation space as the convex hull of word synonyms, DNE far outperforms ASCC (See TMD paper Table 1 where DNE nearly outperforms TMD on AGNews while maintaining a high clean accuracy).", "questions": "1. How often are are the predictions for F_advfooler the same as those for F under different noise rates? (Claim in 3.1) The chart showing adversarial robustness dropping while clean accuracy only slightly drops, gets at this question somewhat. But if possible, I'd like something that's even cleaner of a comparison that can maybe help provide intuition for why the clean accuracy doesn't really suffer much (Esp on IMDB) while adversarial robustness totally craters under high noise (Fig 4). \n\nSuggestions for added baselines (2 of them) included in the Weaknesses section (numbers 4 and 5).", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "No ethnical concerns really. It's a standard adversarial robustness defense paper for word substitution attacks.", "rating": "8: accept, good paper", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1697675301127}], "openreview_url": "https://openreview.net/forum?id=E296x0YpML", "arxiv_id": "2310.01452", "paper_pdf": "papers/E296x0YpML.pdf", "paper_pdf_sha256": "81d777153b5ddf419736c45453d6df8d4c7e59ad199d5fe460677826182ac410", "paper_pdf_bytes": 338325, "paper_pdf_source": "openreview", "code_url": "https://github.com/mail-research/AdvFooler-text-defender", "code_repository": "mail-research/AdvFooler-text-defender", "code_commit": "29b69d7f38f73ab65eea78467c993d0a4d0e3035", "code_archive": "repos/E296x0YpML.zip", "code_archive_sha256": "af74518c78f6dd7e0aa9a6ef6724c328e01d8f44db0656063061444858b0f8ce", "code_archive_bytes": 613774, "code_file_count": 282, "code_extensions": {".py": 282}, "github_disk_usage_kb": 508, "github_languages": {"Python": 1849165}, "github_archived": false, "github_pushed_at": "2024-07-22T06:03:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fooling-the-textual-fooler-via-randomizing"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "QP02DQ-FG-8", "year": 2023, "status": "rejected", "title": "Incomplete to complete multiphysics forecasting - a hybrid approach for learning unknown phenomena", "authors": ["Nilam Nandkishor Tathawadekar", "Nguyen Anh Khoa Doan", "Camilo Fernando Silva", "Nils Thuerey"], "authorids": ["~Nilam_Nandkishor_Tathawadekar2", "~Nguyen_Anh_Khoa_Doan1", "~Camilo_Fernando_Silva1", "~Nils_Thuerey1"], "authors_source": "OpenReview API", "abstract": "Modeling complex dynamical systems where only partial knowledge of their physical mechanisms is available is a crucial problem across all scientific and engineering disciplines. Purely data-driven approaches, which only make use of an artificial neural network and data, often fail to accurately simulate the evolution of the system dynamics over a sufficiently long time and in a physically consistent manner. Therefore, we propose a hybrid approach that uses a neural network model in combination with an incomplete PDE solver that provides known but incomplete physical information. In this study, we demonstrate that the results obtained from the incomplete PDEs can be efficiently corrected at every time step by the proposed hybrid neural network – PDE solver model, so that the effect of the unknown physics present in the system is correctly accounted for. For validation purposes, the obtained simulations of the hybrid model are successfully compared against results coming from the complete set of PDEs describing the full physics of the considered system. We demonstrate the validity of the proposed approach on a reactive flow, an archetypal multi-physics system that combines fluid mechanics and chemistry, the latter being the physics considered unknown. Experiments are made on planar and Bunsen-type flames at various operating conditions. The hybrid neural network - PDE approach correctly models the flame evolution of the cases under study for significantly long time windows, yields improved generalization, and allows for larger simulation time steps. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "QMuwZmQ6Gk", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2819/Reviewer_c79Z"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper argues for the combination of a neural network with an incomplete description of a system (a hybrid model), in their case, an incomplete PDE, rather than a fully data-driven approach that only relies on a neural network. They compare the two approaches on a set of systems described by thermodynamics and fluid dynamics. \n\nThe authors consider a supervised setting where the complete PDE, or solutions for it, are given and used for training the neural networks with an MSE loss. In this setting, their experiments suggest that the hybrid models outperform the data-driven approach by a significant margin. In addition, they argue that hybrid models can lead to faster solving of the PDEs than relying on the complete description when it is available. ", "review_text": "Overall I recommend rejection of this paper for the aforementioned reasons. In particular, the lack of novelty of the approach and the limited reliability of the experiments.", "strengths": "# Strength\nThe idea of using the physical understanding of the studied phenomenon, rather than an entirely data-driven one, makes a lot of sense to me. In addition, the systems studied in this paper seem novel to the hybrid learning literature and constitute solid test scenarios for assessing future developments in hybrid learning. \n\n# Weaknesses\nOverall I find the contribution very limited. Combining a neural network with equations from physics is a long-standing idea in the community. This paper does not introduce anything subtle regarding a good strategy to combine these equations with neural networks. \n\nThe experimental validation of their approach also seems limited in terms of the train/test scenarios considered. In addition, it is unclear how the authors selected the best models and hyperparameters, as no cross-validation or validation set is discussed. This also reduces the solidity of the presented results. For instance, I am pretty sure that at some point, the data-driven approach could catch up with the hybrid approach if the number of train scenarios increases. An analysis of when this (might) happen would have been valuable to motivate the hybrid approach for these problems.\n\nIt is unclear to me whether the parameters of the PDE are also learned at training time. If not, this makes the proposed approach's applicability very limited as, in most cases, if we can observe a system and have little understanding of the physics behind this also means that we do not know the corresponding parameters of the incomplete physical description. I have probably misunderstood something there, but I would encourage the authors to clarify this in the paper. Overall, this paper would benefit from a real-world demonstration of the proposed approach. Indeed, the contribution is not methodological. Thus I would argue that this should contain a more solid empirical evaluation.\n\nAs a remark, I also feel that some confusion is made between the \"model\" and the \"inference\" (in the case of PDEs, the solver). The neural network aims to complete the model, create a more accurate description of the real world, and create a better model. The solver is related to how we make inferences given a model; the terms complete and incomplete solvers do not make sense to me.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper argues for the combination of a neural network with an incomplete description of a system (a hybrid model), in their case, an incomplete PDE, rather than a fully data-driven approach that only relies on a neural network. They compare the two approaches on a set of systems described by thermodynamics and fluid dynamics. \n\nThe authors consider a supervised setting where the complete PDE, or solutions for it, are given and used for training the neural networks with an MSE loss. In this setting, their experiments suggest that the hybrid models outperform the data-driven approach by a significant margin. In addition, they argue that hybrid models can lead to faster solving of the PDEs than relying on the complete description when it is available. ", "strength_and_weaknesses": "# Strength\nThe idea of using the physical understanding of the studied phenomenon, rather than an entirely data-driven one, makes a lot of sense to me. In addition, the systems studied in this paper seem novel to the hybrid learning literature and constitute solid test scenarios for assessing future developments in hybrid learning. \n\n# Weaknesses\nOverall I find the contribution very limited. Combining a neural network with equations from physics is a long-standing idea in the community. This paper does not introduce anything subtle regarding a good strategy to combine these equations with neural networks. \n\nThe experimental validation of their approach also seems limited in terms of the train/test scenarios considered. In addition, it is unclear how the authors selected the best models and hyperparameters, as no cross-validation or validation set is discussed. This also reduces the solidity of the presented results. For instance, I am pretty sure that at some point, the data-driven approach could catch up with the hybrid approach if the number of train scenarios increases. An analysis of when this (might) happen would have been valuable to motivate the hybrid approach for these problems.\n\nIt is unclear to me whether the parameters of the PDE are also learned at training time. If not, this makes the proposed approach's applicability very limited as, in most cases, if we can observe a system and have little understanding of the physics behind this also means that we do not know the corresponding parameters of the incomplete physical description. I have probably misunderstood something there, but I would encourage the authors to clarify this in the paper. Overall, this paper would benefit from a real-world demonstration of the proposed approach. Indeed, the contribution is not methodological. Thus I would argue that this should contain a more solid empirical evaluation.\n\nAs a remark, I also feel that some confusion is made between the \"model\" and the \"inference\" (in the case of PDEs, the solver). The neural network aims to complete the model, create a more accurate description of the real world, and create a better model. The solver is related to how we make inferences given a model; the terms complete and incomplete solvers do not make sense to me.", "clarity,_quality,_novelty_and_reproducibility": "# Clarity\nThe paper's clarity is ok but should be improved to be accepted for publication, in my opinion. For instance, see my confusion regarding some of the points mentioned in the weaknesses of the paper.\n\n# Quality\nI did not find the experimental validation solid enough to trust the empirical conclusion made by the authors entirely. For example, the authors do not explain how they selected the different architectures and training hyperparameters. \n\n# Novelty\nThe novelty is limited. The main contribution is to showcase a simple hybrid approach to original test scenarios.\n\n# Typos and minor remarks\n- What is u in eq (1)?\n- The second paragraph before section 3: don't -> do not\n- The paragraph before section 3: Why don't you discuss Takeishi and Kalousis 2021 (https://scholar.google.com/citations?view_op=view_citation&hl=fr&user=rqF9bAsAAAAJ&citation_for_view=rqF9bAsAAAAJ:hC7cP41nSMkC)? I am not an author of this paper, it is a genuine remark. I also wonder why you do not compare to an approach similar to APHYNITY (https://scholar.google.com/citations?view_op=view_citation&hl=fr&user=rFaxB20AAAAJ&sortby=pubdate&citation_for_view=rFaxB20AAAAJ:IZKZNMMMWs0C).\n- Eq (3): a point at the end of the equation is missing.\n- Eq (5): the upper limit should be t + m.\n- unfiform-Busen -> Uniform-Busen.", "summary_of_the_review": "Overall I recommend rejection of this paper for the aforementioned reasons. In particular, the lack of novelty of the approach and the limited reliability of the experiments.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667560660177}, {"id": "HKCxgZ3lHX", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2819/Reviewer_hgQY"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper aims at deploying a deep learning approach in combination with partial differential equations (PDEs) with known yet incomplete physical information. \nFor example, a deep learning model can be employed when the information on the system to be solved is limited but additional data are available. A common benchmark is represented by turbulence modelling in computational fluid dynamics where the full PDEs descriptions are too expensive to be solved computationally. Therefore, the PDEs can be rewritten as a sum of an incomplete PDE description plus another term representing the physics arising from the unknown information (reaction rate terms). Correctly representing this additional term is of crucial importance to solve the task. \n\nContrarily to standard Neural PDE solvers, this method aims at combining physical incomplete PDE information with some additional machine learned correction due to the unknown part of the PDE description. In this work, the latter is represented by the chemical kinetic mechanism of non reactive flow simulation. The complete (reactive) simulation is used as a baseline. \nThe downstream task is to solve a simplified version of the Navier-Stokes equation for a specific system where the chemistry is removed from the analytic description (i.e., is unknown).\n\nA neural network is trained to learn correction fields that would correct the incomplete (non reactive) PDE equations where both source terms and reaction rates have been removed.\nThe work compares three baselines: Ground truth, Pure Data Driven approach, Fourier Neural Operators (state-of-the-art neural operator method).\n\nThe framework is clearly explained in figure 2 where the PDD and the NN-Hybrid approaches are compared in B1 and B2 respectively. \nThe 3 different setups on which the approach was tested are carefully explained.  \n\nA few comments/concerns are in place:\n\n- The following sentence is a bit unclear to me ‘Its objective is to learn to model the effects of the unknown chemistry using the neural network parameters θ given an input flow \n State […]’. What is it meant by ‘[…] using the neural network parameters \\theta […]’?\n- In the last row of table 1 some baselines are missing (i.e., FNO and NN.) I presume this is because they never converge to the target value? It would be helpful to write something about it in the table caption. \n- On page 7 below figure 4, in the second paragraph the authors define the relative displacement. Right afterward, they mention x_t and \\tilde{x}_0. I cannot relate these quantities easily to the definition and this may be overall a bit confusing. I would suggest to keep the definition of relative displacement general \\tilde{x}_300 -> \\tilde{x}_t and they specify that t=300. As it is now the discussion might not sound very linear and straightforward. \n- I would suggest the authors to add additional plot to figure 6 to visualise the performance of other baselines like FNO and NN. From the discussion of table 1 is already clear they perform worse than NN-PDE, but visualising them (as it’s done in figure 12) might be useful.\n- In section 5.3 the author only discuss the full NN method without mentioning the NN-PDE. I am wondering it that is intentional. If so, there’s a reason for that? Are there any limitations that prevent NN-PDE to be applied for this task? \n- I suggest the authors to make the labels explicitly in figure 2 saying what’s the PDD and what’s the NN-hibridy approach for an easy first-sight intuition.\n- The caption of figure 2 does not explain what is S. Though this is done in the main text would be useful to repeat it in the caption below the figure. \n- I’d rename section 2 something like ‘Literature Review’ or ‘Related Work’. To me ‘Background’ sounds more like if you are introducing the physics behind the work.  \n- Missing full stop at the end of equation 3. \n- Few typos/stylistic issues were found throughout the manuscript. I’d recommend the authors to run a spellchecker for making sure to detect them.", "review_text": "To incorporate physical prior information when training a neural network to solve a physical problem is known to be a successful approach. \nIn this paper the author propose to learn an incomplete description of a PDE and correct a posteriori using a NN for removing two terms which in general act on different time scales. Doing so, the PDE solver happens to be much more efficient in learning the dynamics of the incomplete system. Contrarily,  when learning the full dynamics of the complete PDE description (pure NN approach) the interplay between different time-scales makes the learning task much more difficult. \n\nThe idea of separating the PDE and learn part of the dynamics in a supervised learning fashion and correct the PDE result a posteriori is indeed a valuable and interesting contribution to the community. \n\nDespite not being a domain expert, I still do see some potential in this work, especially in future developments. \nFor these reasons and all the considerations from above, I would recommend to consider this manuscript for acceptance. ", "strengths": "The problem tackled by the paper is of great practical relevance. It is common in the physical sciences to face numerical PDEs where the complete description is either very expensive, from a computational perspective, or non accessible. Therefore, when a full analytic description is available, as well as reference, experimental, data, one can combine a numerical PDE solver for an incomplete description  of the full PDE with a correction term which can be learned by a NN upon training on some appropriate data. \n\nOn the other hand, physics-informed ML is a quite well-established avenue of research. It is not at all surprising to me that combining incomplete physical information of an incomplete PDE description with a NN that learns phenomena acting on different time scales enhance the performance. Therefore, while I am not too impressed nor surprised by the results, I do indeed see some value in the method and I thereby recommend consider this manuscript for acceptance.  \n\nAs a further suggestion the author may study some transfer learning properties, i.e., how well the correction NN would perform when being trained on a specific set of initial condition but used in different setups. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper aims at deploying a deep learning approach in combination with partial differential equations (PDEs) with known yet incomplete physical information. \nFor example, a deep learning model can be employed when the information on the system to be solved is limited but additional data are available. A common benchmark is represented by turbulence modelling in computational fluid dynamics where the full PDEs descriptions are too expensive to be solved computationally. Therefore, the PDEs can be rewritten as a sum of an incomplete PDE description plus another term representing the physics arising from the unknown information (reaction rate terms). Correctly representing this additional term is of crucial importance to solve the task. \n\nContrarily to standard Neural PDE solvers, this method aims at combining physical incomplete PDE information with some additional machine learned correction due to the unknown part of the PDE description. In this work, the latter is represented by the chemical kinetic mechanism of non reactive flow simulation. The complete (reactive) simulation is used as a baseline. \nThe downstream task is to solve a simplified version of the Navier-Stokes equation for a specific system where the chemistry is removed from the analytic description (i.e., is unknown).\n\nA neural network is trained to learn correction fields that would correct the incomplete (non reactive) PDE equations where both source terms and reaction rates have been removed.\nThe work compares three baselines: Ground truth, Pure Data Driven approach, Fourier Neural Operators (state-of-the-art neural operator method).\n\nThe framework is clearly explained in figure 2 where the PDD and the NN-Hybrid approaches are compared in B1 and B2 respectively. \nThe 3 different setups on which the approach was tested are carefully explained.  \n\nA few comments/concerns are in place:\n\n- The following sentence is a bit unclear to me ‘Its objective is to learn to model the effects of the unknown chemistry using the neural network parameters θ given an input flow \n State […]’. What is it meant by ‘[…] using the neural network parameters \\theta […]’?\n- In the last row of table 1 some baselines are missing (i.e., FNO and NN.) I presume this is because they never converge to the target value? It would be helpful to write something about it in the table caption. \n- On page 7 below figure 4, in the second paragraph the authors define the relative displacement. Right afterward, they mention x_t and \\tilde{x}_0. I cannot relate these quantities easily to the definition and this may be overall a bit confusing. I would suggest to keep the definition of relative displacement general \\tilde{x}_300 -> \\tilde{x}_t and they specify that t=300. As it is now the discussion might not sound very linear and straightforward. \n- I would suggest the authors to add additional plot to figure 6 to visualise the performance of other baselines like FNO and NN. From the discussion of table 1 is already clear they perform worse than NN-PDE, but visualising them (as it’s done in figure 12) might be useful.\n- In section 5.3 the author only discuss the full NN method without mentioning the NN-PDE. I am wondering it that is intentional. If so, there’s a reason for that? Are there any limitations that prevent NN-PDE to be applied for this task? \n- I suggest the authors to make the labels explicitly in figure 2 saying what’s the PDD and what’s the NN-hibridy approach for an easy first-sight intuition.\n- The caption of figure 2 does not explain what is S. Though this is done in the main text would be useful to repeat it in the caption below the figure. \n- I’d rename section 2 something like ‘Literature Review’ or ‘Related Work’. To me ‘Background’ sounds more like if you are introducing the physics behind the work.  \n- Missing full stop at the end of equation 3. \n- Few typos/stylistic issues were found throughout the manuscript. I’d recommend the authors to run a spellchecker for making sure to detect them.", "strength_and_weaknesses": "The problem tackled by the paper is of great practical relevance. It is common in the physical sciences to face numerical PDEs where the complete description is either very expensive, from a computational perspective, or non accessible. Therefore, when a full analytic description is available, as well as reference, experimental, data, one can combine a numerical PDE solver for an incomplete description  of the full PDE with a correction term which can be learned by a NN upon training on some appropriate data. \n\nOn the other hand, physics-informed ML is a quite well-established avenue of research. It is not at all surprising to me that combining incomplete physical information of an incomplete PDE description with a NN that learns phenomena acting on different time scales enhance the performance. Therefore, while I am not too impressed nor surprised by the results, I do indeed see some value in the method and I thereby recommend consider this manuscript for acceptance.  \n\nAs a further suggestion the author may study some transfer learning properties, i.e., how well the correction NN would perform when being trained on a specific set of initial condition but used in different setups. ", "clarity,_quality,_novelty_and_reproducibility": "- The paper is very clear. \n- The writing style is appropriate and engaging. \n- The method presented in the manuscript is understandable also for people which are not familiar with the specific application. \n- While I am not domain expert, to my knowledge there are no other references where a NN is trained to correct for some lack of information in the PDE description. \n- The method is described in a clear and accessible way making it appealing and deployable to different other downstream tasks involving complicated PDEs.  \n- The authors assert the code will be released upon acceptance. However, the manuscript provides quite a high level of details which I imagine  would make the results easily reproducible. ", "summary_of_the_review": "To incorporate physical prior information when training a neural network to solve a physical problem is known to be a successful approach. \nIn this paper the author propose to learn an incomplete description of a PDE and correct a posteriori using a NN for removing two terms which in general act on different time scales. Doing so, the PDE solver happens to be much more efficient in learning the dynamics of the incomplete system. Contrarily,  when learning the full dynamics of the complete PDE description (pure NN approach) the interplay between different time-scales makes the learning task much more difficult. \n\nThe idea of separating the PDE and learn part of the dynamics in a supervised learning fashion and correct the PDE result a posteriori is indeed a valuable and interesting contribution to the community. \n\nDespite not being a domain expert, I still do see some potential in this work, especially in future developments. \nFor these reasons and all the considerations from above, I would recommend to consider this manuscript for acceptance. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1667293683429}, {"id": "LARm1WVykYr", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2819/Reviewer_95Sg"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the task of learning continuous dynamical systems from data using a hybrid (machine learning + PDE solvers) approach and compares the performance to purely data-driven methods. In particular, authors consider a regularly occurring in practice setting where only a partial knowledge of the governing equations are available. Authors use reactive flow systems (specifically various flame simulations) as the test ground for evaluation and comparison. They find that by incorporating physical priors in the form of a partial PDE solver improves the performance of the learned model and is able to accurately capture the dynamics of the system and works well for test simulation settings. Additional findings include observation that learned hybrid models result in less stiff updates compared to the underlying ground truth solver.", "review_text": "I enjoyed reading the paper in general, as the authors address an important question of whether physical priors help/needed to model continuous dynamical systems well. I find the step up in complexity of the simulated system a bonus and in general think this is an important research avenue.\n\nI’m generally leaning towards accepting this paper (score of 8), assuming some flaws and weaknesses can be addressed. (hence the initial score of 6).\n\nMy (1) primary concern is related to the evaluation metric and error analysis. Unless I’ve missed something, it is not clear how significant the mean values of MAPE and MSE are given the standard deviations. Is there a way to present the results in a way that clearly indicates that hybrid approach generally improves over data-driven methods? Maybe using a histogram of errors over different sample trajectories / random seeds would provide a more complete picture of error distribution? The fear is that coarse summary might hide stability problems of the purely data driven approaches which would misrepresent their performance as there are numerous methods to address stability issues.\n\nMy (2) concern is related to the time-step used for data-driven predictions. In general, data-driven approaches often subsample the predictions in time, which result in better accuracy and efficiency. (e.g. see Fig. 3 “Stachenfeld et al., 2022” (reference used in the paper)). Given that the underlying ground truth solver involves stiff dynamics it could turn out that data-driven models are evaluated in a highly suboptimal regime where they are tasked to predict a very slowly varying solution.\n\nFinally, I think the paper would read better if the stakes for “primacy” were omitted. Arguably a number of previous works learned various closure models in a tandem with differentiable solvers and present work just extends it to a more challenging setting, which is important in its own right.\n\nI’d be happy to provide an additional set of minor comments regarding potential improvements wrt references after primary concerns are addressed.\n", "strengths": "Strengths:\n* Authors consider a dynamical system that is more complex than that commonly studied in learned simulators literature.\n* The considered method is tested against relevant baseline approaches.\n* Paper is clearly written.\n\nWeaknesses:\n* Standard deviation estimates of the quantitative results (e.g. Table 1) are O(1) wrt. relevant mean values. While visualizations support the conclusions based on the trend of the mean MAPE and MSE, I find the results somewhat hard to trust.\n* An important time subsampling hyperparameter has not been considered for data-driven models.\n* AFAICT datasets are not made available to perform independent evaluations/tests.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper studies the task of learning continuous dynamical systems from data using a hybrid (machine learning + PDE solvers) approach and compares the performance to purely data-driven methods. In particular, authors consider a regularly occurring in practice setting where only a partial knowledge of the governing equations are available. Authors use reactive flow systems (specifically various flame simulations) as the test ground for evaluation and comparison. They find that by incorporating physical priors in the form of a partial PDE solver improves the performance of the learned model and is able to accurately capture the dynamics of the system and works well for test simulation settings. Additional findings include observation that learned hybrid models result in less stiff updates compared to the underlying ground truth solver.", "strength_and_weaknesses": "Strengths:\n* Authors consider a dynamical system that is more complex than that commonly studied in learned simulators literature.\n* The considered method is tested against relevant baseline approaches.\n* Paper is clearly written.\n\nWeaknesses:\n* Standard deviation estimates of the quantitative results (e.g. Table 1) are O(1) wrt. relevant mean values. While visualizations support the conclusions based on the trend of the mean MAPE and MSE, I find the results somewhat hard to trust.\n* An important time subsampling hyperparameter has not been considered for data-driven models.\n* AFAICT datasets are not made available to perform independent evaluations/tests.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly written and provides a decent account for the choices made to set up the test problem. The use of existing framework for fluid dynamics solver (phiflow) is a plus. I did not find any reference to the source code/datasets; and given the complexity of the setup (numerical PDE solver + ML model) I think it’s quite hard to independently reproduce the results based on the paper alone.", "summary_of_the_review": "I enjoyed reading the paper in general, as the authors address an important question of whether physical priors help/needed to model continuous dynamical systems well. I find the step up in complexity of the simulated system a bonus and in general think this is an important research avenue.\n\nI’m generally leaning towards accepting this paper (score of 8), assuming some flaws and weaknesses can be addressed. (hence the initial score of 6).\n\nMy (1) primary concern is related to the evaluation metric and error analysis. Unless I’ve missed something, it is not clear how significant the mean values of MAPE and MSE are given the standard deviations. Is there a way to present the results in a way that clearly indicates that hybrid approach generally improves over data-driven methods? Maybe using a histogram of errors over different sample trajectories / random seeds would provide a more complete picture of error distribution? The fear is that coarse summary might hide stability problems of the purely data driven approaches which would misrepresent their performance as there are numerous methods to address stability issues.\n\nMy (2) concern is related to the time-step used for data-driven predictions. In general, data-driven approaches often subsample the predictions in time, which result in better accuracy and efficiency. (e.g. see Fig. 3 “Stachenfeld et al., 2022” (reference used in the paper)). Given that the underlying ground truth solver involves stiff dynamics it could turn out that data-driven models are evaluated in a highly suboptimal regime where they are tasked to predict a very slowly varying solution.\n\nFinally, I think the paper would read better if the stakes for “primacy” were omitted. Arguably a number of previous works learned various closure models in a tandem with differentiable solvers and present work just extends it to a more challenging setting, which is important in its own right.\n\nI’d be happy to provide an additional set of minor comments regarding potential improvements wrt references after primary concerns are addressed.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "I enjoyed reading the paper in general, as the authors address an important question of whether physical priors help/needed to model continuous dynamical systems well. I find the step up in complexity of the simulated system a bonus and in general think this is an important research avenue.\n\nI’m generally leaning towards accepting this paper (score of 8), assuming some flaws and weaknesses can be addressed. (hence the initial score of 6).\n\nMy (1) primary concern is related to the evaluation metric and error analysis. Unless I’ve missed something, it is not clear how significant the mean values of MAPE and MSE are given the standard deviations. Is there a way to present the results in a way that clearly indicates that hybrid approach generally improves over data-driven methods? Maybe using a histogram of errors over different sample trajectories / random seeds would provide a more complete picture of error distribution? The fear is that coarse summary might hide stability problems of the purely data driven approaches which would misrepresent their performance as there are numerous methods to address stability issues.\n\nMy (2) concern is related to the time-step used for data-driven predictions. In general, data-driven approaches often subsample the predictions in time, which result in better accuracy and efficiency. (e.g. see Fig. 3 “Stachenfeld et al., 2022” (reference used in the paper)). Given that the underlying ground truth solver involves stiff dynamics it could turn out that data-driven models are evaluated in a highly suboptimal regime where they are tasked to predict a very slowly varying solution.\n\nFinally, I think the paper would read better if the stakes for “primacy” were omitted. Arguably a number of previous works learned various closure models in a tandem with differentiable solvers and present work just extends it to a more challenging setting, which is important in its own right.\n\nI’d be happy to provide an additional set of minor comments regarding potential improvements wrt references after primary concerns are addressed.\n", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666628320443}, {"id": "BfKcl2KawU", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2819/Reviewer_NRLv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper considers the case of incomplete physics for numerically solving Partial Differential Equations. It uses a neural network to complete the missing terms and more accurately forecast the evolution of the dynamical system. The proposed approach is evaluated on different variants of a reactive flow problem governed by Navier-Stokes equations. ", "review_text": "This is mostly an application paper that proposes to learn to complete incomplete physics models using neural networks in the context of reactive flow problems. My main concerns with the paper is how it relates to previous work that also learns to complete incomplete physics models as well as the fact that, at least to my understanding, the baselines that are considered are weak. ", "strengths": "Weaknesses\n\nThe paper claims that it is the first of each kind that learns to complete physics in a PDE setting. Nevertheless learning with incomplete physics models using neural networks to complete the missing part has been studied rather extensively. For example: APHYNITY, Yin et a, Journal of Statistical Mechanics: Theory and Experiment (also cited in beginning of the present submission)  and ICLR2021; Physics integrated VAEs for robust and interpretable generative modelling, Takeishi & Kalousis, NeurIPS, 2021. APHYNITY tackles forecasting tasks, similar to the current submission, for physics systems that are described by ODES/PDES. It models the underlying dynamical system with two additive components, one given by the known physics and a neural network modelling the unknown parts. Physics VAE learns generative models where the VAE decoder also decomposes, to a more general composition than additive, of an incomplete physics component and a neural network. The paper should at least discuss differences from these two works. \n\nWith respect to the experimental evaluation it is not clear to me how the neural network baseline has been trained. It seems that it learns state-state transitions. How is this done? does the model during training sees only consecutive state pairs? or does the model evolve in a longer horizon, i.e. over a sequence of states, and then the model update would have to happen over such sequences, accounting like that for longer temporal dependencies and error compounding. From what I understand the neural network baseline is trained over state state transitions, which, if I am correct, puts it at a clear disadavantage with the hybrid model. More appropriate baselines will explicitly model for long term dependencies, e.g. a solver where there is no physics but just a neural network, neural ODEs, models with recurency. The two papers above provide a number of such baselines. \n\nStrengths\nThere is a detailed evaluation of the model in a complex scenario exploring different extrapolation scenarios. \n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper considers the case of incomplete physics for numerically solving Partial Differential Equations. It uses a neural network to complete the missing terms and more accurately forecast the evolution of the dynamical system. The proposed approach is evaluated on different variants of a reactive flow problem governed by Navier-Stokes equations. ", "strength_and_weaknesses": "Weaknesses\n\nThe paper claims that it is the first of each kind that learns to complete physics in a PDE setting. Nevertheless learning with incomplete physics models using neural networks to complete the missing part has been studied rather extensively. For example: APHYNITY, Yin et a, Journal of Statistical Mechanics: Theory and Experiment (also cited in beginning of the present submission)  and ICLR2021; Physics integrated VAEs for robust and interpretable generative modelling, Takeishi & Kalousis, NeurIPS, 2021. APHYNITY tackles forecasting tasks, similar to the current submission, for physics systems that are described by ODES/PDES. It models the underlying dynamical system with two additive components, one given by the known physics and a neural network modelling the unknown parts. Physics VAE learns generative models where the VAE decoder also decomposes, to a more general composition than additive, of an incomplete physics component and a neural network. The paper should at least discuss differences from these two works. \n\nWith respect to the experimental evaluation it is not clear to me how the neural network baseline has been trained. It seems that it learns state-state transitions. How is this done? does the model during training sees only consecutive state pairs? or does the model evolve in a longer horizon, i.e. over a sequence of states, and then the model update would have to happen over such sequences, accounting like that for longer temporal dependencies and error compounding. From what I understand the neural network baseline is trained over state state transitions, which, if I am correct, puts it at a clear disadavantage with the hybrid model. More appropriate baselines will explicitly model for long term dependencies, e.g. a solver where there is no physics but just a neural network, neural ODEs, models with recurency. The two papers above provide a number of such baselines. \n\nStrengths\nThere is a detailed evaluation of the model in a complex scenario exploring different extrapolation scenarios. \n", "clarity,_quality,_novelty_and_reproducibility": "The paper is mostly easy to follow. I have some doubts with respect to the originality since it seems to me that APHYNITY and Physics VAE address very similar, of not the same settings. ", "summary_of_the_review": "This is mostly an application paper that proposes to learn to complete incomplete physics models using neural networks in the context of reactive flow problems. My main concerns with the paper is how it relates to previous work that also learns to complete incomplete physics models as well as the fact that, at least to my understanding, the baselines that are considered are weak. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666515140095}], "openreview_url": "https://openreview.net/forum?id=QP02DQ-FG-8", "arxiv_id": "2111.11185", "paper_pdf": "papers/QP02DQ-FG-8.pdf", "paper_pdf_sha256": "ca651ffdcb1b721f03052392620772eb02244738ad2c859419391ace2c1e6524", "paper_pdf_bytes": 2072512, "paper_pdf_source": "openreview", "code_url": "https://github.com/tum-pbs/Hybrid-Solver-for-Reactive-Flows", "code_repository": "tum-pbs/Hybrid-Solver-for-Reactive-Flows", "code_commit": "75d836ff82c30da698b8abd12fe0e2c5b6d79da2", "code_archive": "repos/QP02DQ-FG-8.zip", "code_archive_sha256": "911e669bcbe2cd9192a255a8da0ff270896a0e4398746ca051b007d73ce96d2b", "code_archive_bytes": 1410465, "code_file_count": 31, "code_extensions": {".py": 31}, "github_disk_usage_kb": 1319, "github_languages": {"Python": 443412, "Makefile": 10048}, "github_archived": false, "github_pushed_at": "2024-01-10T13:44:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/hybrid-neural-network-pde-solvers-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "e_D6AmszH4P", "year": 2022, "status": "rejected", "title": "ViViT: Curvature access through the generalized Gauss-Newton's low-rank structure", "authors": ["Felix Dangel", "Lukas Tatzel", "Philipp Hennig"], "authorids": ["~Felix_Dangel1", "~Lukas_Tatzel1", "~Philipp_Hennig1"], "authors_source": "OpenReview API", "abstract": "Curvature in form of the Hessian or its generalized Gauss-Newton (GGN) approximation is valuable for algorithms that rely on a local model for the loss to train, compress, or explain deep networks. Existing methods based on implicit multiplication via automatic differentiation or Kronecker-factored block diagonal approximations do not consider noise in the mini-batch. We present ViViT, a curvature model that leverages the GGN's low-rank structure without further approximations. It allows for efficient computation of eigenvalues, eigenvectors, as well as per-sample first- and second-order directional derivatives. The representation is computed in parallel with gradients in one backward pass and offers a fine-grained cost-accuracy trade-off, which allows it to scale. As examples for ViViT's usefulness, we investigate the directional first- and second-order derivatives during training, and how noise information can be used to improve the stability of second-order methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "yTT5aSXXyxu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2741/Reviewer_1dgm"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper describes a way of leveraging the low-rank structure of the generalized Gauss-Newton (GGN) matrix, and a library that extends BackPACK's features in order to efficiently compute:\n - the spectrum (eigenvalues, eigenvectors) of the GGN\n - per sample directional derivatives and curvature of the GGN\n\nThey then demonstrate the library on a 900K parameters x 10 classes deep network on CIFAR-10.\nThey also propose a damping scheme for second-order that only have access to noisy estimates of the curvature.", "review_text": "The paper is clearly written and easy to follow.\n\n**Main idea: low-rank structure of the GGN**: The main idea of taking advantage of the low-rank structure of the GGN is clearly motivated as well as the linear algebra manipulations (though somewhat basic) used to get the spectrum of the GGN. I would however not deem it as new as it is e.g. discussed in [1,2], but arguably it is interesting to emphasize how to use this low-rank structure in deep learning algorithms.\n\n**Implementation** It is difficult to judge given that no code is provided (or linked) but it would certainly be very useful to have a good implementation for such quantities. I would be interested in seeing timings (or at least orders of magnitudes) for your experiments on this reasonably large model of section 4.1. I am not sure however that ICLR is the most appropriate conference for implementation papers.\n\n**Per-sample variance analysis** (section 4.1) is new as far as I can tell. It provides very interesting insights about the SNR of the gradients and curvature projected onto the eigenvectors of the GGN. I would have appreciated some more insights with more standard networks (e.g. ResNets)\n\n**Proposed adaptive damping scheme** The information about mini-batch variance is nicely exploited in the proposed adaptive damping scheme for second-order methods. As you rightfully notice, fully analyzing and benchmarking such an algorithm on an actual neural network would require a separate paper, which I would be looking forward to reading. A difficult problem that I can foresee is that even by estimating the GGN on a minibatch, the partial left singular vectors of $\\mathbf V$ are of dimension $D$ which can already be quite large.\n\n[1] https://arxiv.org/abs/1905.11675\n\n[2] https://arxiv.org/abs/1806.02958", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper describes a way of leveraging the low-rank structure of the generalized Gauss-Newton (GGN) matrix, and a library that extends BackPACK's features in order to efficiently compute:\n - the spectrum (eigenvalues, eigenvectors) of the GGN\n - per sample directional derivatives and curvature of the GGN\n\nThey then demonstrate the library on a 900K parameters x 10 classes deep network on CIFAR-10.\nThey also propose a damping scheme for second-order that only have access to noisy estimates of the curvature.", "main_review": "The paper is clearly written and easy to follow.\n\n**Main idea: low-rank structure of the GGN**: The main idea of taking advantage of the low-rank structure of the GGN is clearly motivated as well as the linear algebra manipulations (though somewhat basic) used to get the spectrum of the GGN. I would however not deem it as new as it is e.g. discussed in [1,2], but arguably it is interesting to emphasize how to use this low-rank structure in deep learning algorithms.\n\n**Implementation** It is difficult to judge given that no code is provided (or linked) but it would certainly be very useful to have a good implementation for such quantities. I would be interested in seeing timings (or at least orders of magnitudes) for your experiments on this reasonably large model of section 4.1. I am not sure however that ICLR is the most appropriate conference for implementation papers.\n\n**Per-sample variance analysis** (section 4.1) is new as far as I can tell. It provides very interesting insights about the SNR of the gradients and curvature projected onto the eigenvectors of the GGN. I would have appreciated some more insights with more standard networks (e.g. ResNets)\n\n**Proposed adaptive damping scheme** The information about mini-batch variance is nicely exploited in the proposed adaptive damping scheme for second-order methods. As you rightfully notice, fully analyzing and benchmarking such an algorithm on an actual neural network would require a separate paper, which I would be looking forward to reading. A difficult problem that I can foresee is that even by estimating the GGN on a minibatch, the partial left singular vectors of $\\mathbf V$ are of dimension $D$ which can already be quite large.\n\n[1] https://arxiv.org/abs/1905.11675\n\n[2] https://arxiv.org/abs/1806.02958", "summary_of_the_review": "If it were a purely theory paper, I am not sure that the idea of leveraging the low-rank structure of the GGN makes for a novel enough contribution by itself to be published at ICLR.\n\nIt if were a pure implementation paper, I think ICLR would not be the most appropriate venue.\n\nHence my low rating, for a paper that is otherwise very pleasant to read.\n\nI think that the proposed adaptive damping scheme can however make for a nice follow-up.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636149703373}, {"id": "JlmOIGjMY3", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2741/Reviewer_GBwe"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This work proposes a curvature model VIVIT based on generalized Gauss-Newton (GGN) approximation for the training of neural networks with a convex loss function. The low-rank structure of VIVIT allows for efficient eigen-value decomposition, which also gives per-sample directional derivatives and curvatures. To further improve the efficiency, sampling within the mini-batch and among the coordinates of the prediction can be applied to trade off computational cost with accuracy. As an application example, VIVIT is used to provide noise-aware directional damping which improves the stability of second-order methods.", "review_text": "Strengths:\n1.\tIt is novel that the proposed curvature model VIVIT extracts full spectrum from the exact GGN matrix (on each mini-batch) and provides per-sample directional derivatives and curvatures, which may be of interest to future work on their own.\n2.\tThe proposed usage of VIVIT in second-order methods, i.e., the directional damping, is novel and deals with the difficulty of step size selection. \n3.\tOverall, the paper is well organized and clearly written. Thorough discussions on computational complexity and implementation details make this work practical.\n\nWeaknesses:\n1.\tWhy is the GGN matrix a good approximation of the true Hessian matrix? Does the GGN matrix well capture the top eigenvalues of the true Hessian? What is left in the residual matrix $R$? It would be better to have a discussion of the quality of such GGN approximation or demonstrate it in some experiment.\n2.\tIt would be better to compare the obtained eigenvalue spectrum (and eigen-vectors) with some of the existing curvature approximation methods in the experiment. Though these methods may not provide per-sample curvature estimation, it is still necessary to compare with them to show the quality and efficiency of VIVIT.\n3.\tIn Figure 3, it would be better to demonstrate how well the proposed directional damping balances convergence speed and stability. Currently in the left plot its performance seems to be not as good as the simple choice $\\delta=1$.\n4.\tIt would be better to distinguish between the mini-bath size and the training set size and formally state that VIVIT is for the underdetermined cases with $D > NC$.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes a curvature model VIVIT based on generalized Gauss-Newton (GGN) approximation for the training of neural networks with a convex loss function. The low-rank structure of VIVIT allows for efficient eigen-value decomposition, which also gives per-sample directional derivatives and curvatures. To further improve the efficiency, sampling within the mini-batch and among the coordinates of the prediction can be applied to trade off computational cost with accuracy. As an application example, VIVIT is used to provide noise-aware directional damping which improves the stability of second-order methods.", "main_review": "Strengths:\n1.\tIt is novel that the proposed curvature model VIVIT extracts full spectrum from the exact GGN matrix (on each mini-batch) and provides per-sample directional derivatives and curvatures, which may be of interest to future work on their own.\n2.\tThe proposed usage of VIVIT in second-order methods, i.e., the directional damping, is novel and deals with the difficulty of step size selection. \n3.\tOverall, the paper is well organized and clearly written. Thorough discussions on computational complexity and implementation details make this work practical.\n\nWeaknesses:\n1.\tWhy is the GGN matrix a good approximation of the true Hessian matrix? Does the GGN matrix well capture the top eigenvalues of the true Hessian? What is left in the residual matrix $R$? It would be better to have a discussion of the quality of such GGN approximation or demonstrate it in some experiment.\n2.\tIt would be better to compare the obtained eigenvalue spectrum (and eigen-vectors) with some of the existing curvature approximation methods in the experiment. Though these methods may not provide per-sample curvature estimation, it is still necessary to compare with them to show the quality and efficiency of VIVIT.\n3.\tIn Figure 3, it would be better to demonstrate how well the proposed directional damping balances convergence speed and stability. Currently in the left plot its performance seems to be not as good as the simple choice $\\delta=1$.\n4.\tIt would be better to distinguish between the mini-bath size and the training set size and formally state that VIVIT is for the underdetermined cases with $D > NC$.\n", "summary_of_the_review": "The proposed curvature approximation method is novel and practical and appears to have the potential of being useful to future work. Though there are a few things to be improved, in general I find this work well written, and my current recommendation is marginally above the acceptance threshold.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635895496979}, {"id": "OfQg0Ef2DWf", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2741/Reviewer_VGMZ"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper leverages the GGN’s low-rank structure, which allows for efficient computation of eigenvalues, eigenvectors,  per-sample first- and second-order directional derivatives, and parallel feature computation. The empirical performance is encouraging. ", "review_text": "Strengths:\n\nBy leveraging the GGN’s low-rank structure, one can efficiently computate the eigenvalues, eigenvectors,  per-sample first- and second-order directional derivatives, and features. \n\nWeakness:\n1) The technical novelty of using low-rankness of GGN's seems a bit limited. It is suggested to highlight the difference as well as technical difficulty with existing works in exploiting the low-rank structure for fast computation of quantities like eigenvalues. \n2) Give more clear theoretical explanations of why the residual in Eq. (2)  is simply ignored. \n3) The theoretical analysis of why noise information improves the stability of second-order methods seems insufficient. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper leverages the GGN’s low-rank structure, which allows for efficient computation of eigenvalues, eigenvectors,  per-sample first- and second-order directional derivatives, and parallel feature computation. The empirical performance is encouraging. ", "main_review": "Strengths:\n\nBy leveraging the GGN’s low-rank structure, one can efficiently computate the eigenvalues, eigenvectors,  per-sample first- and second-order directional derivatives, and features. \n\nWeakness:\n1) The technical novelty of using low-rankness of GGN's seems a bit limited. It is suggested to highlight the difference as well as technical difficulty with existing works in exploiting the low-rank structure for fast computation of quantities like eigenvalues. \n2) Give more clear theoretical explanations of why the residual in Eq. (2)  is simply ignored. \n3) The theoretical analysis of why noise information improves the stability of second-order methods seems insufficient. ", "summary_of_the_review": "This paper proposes to use the GGN’s low-rank structure for fast and sound model learning. The empirical performance in promising. However, the technical novelty of using low-rankness of GGN's seems limited and the technical significance seems not sufficiently strong. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635868069856}, {"id": "TDgykBndRNu", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2741/Reviewer_9mCz"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper highlights how the low-rank structure of the generalized Gauss-Newton (GGN) approximation of the Hessian can be used as a computationally efficient tool to study the loss landscape of deep neural networks. In particular, authors discuss methods to compute the full spectrum of the GGN and thus providing access to per-sample directional gradients and curvature approximation. Through the lens of the GGN spectrum, authors make observations on the geometry of the loss landscape and its evolution during training, and propose an adaptive damping technique for second-order optimizers that utilizes the GGN curvature information.", "review_text": "Strengths:\n1. The paper proposes a method tailored to work efficiently on state-of-the-art auto differentiation tools and GPU hardware. Such tools are essential in better understanding various deep learning phenomena, improving optimization algorithms and in aiding practitioners with hyperparameter choices.\n2. The observations on the alignment of gradients with top eigendirections correlates well with recent results on deep neural network loss landscapes, and I find the discussion on the evolution of directional gradient and curvature SNR interesting. This direction can be very promising in designing practical second-order optimizers that scale to contemporary models and datasets.\n3. The directional damping scheme is an interesting idea, especially the adjustment of damping based on the stochastic uncertainty in a given eigendirection.\n\nWeaknesses:\n1. Scalability to practical settings: It is not clear whether the proposed method scales for contemporary architectures and large-scale datasets. Since the low-rank structure of the GGN has been investigated before by others (as the authors also pointed out), the main contribution of this paper is the compute/memory efficiency and scalability of the proposed tool. However, the experiments are performed on CIFAR-10 with a small, 6-layer neural network and it is not clear how it would scale to architectures such as ResNet and datasets such as ImageNet. In case the method does not work well in these scenarios, the applicability of the tool is limited.\n2. Overhead: In-depth study on the compute and memory overhead of the method would be valuable. For example statements such as 'However, the practical overhead is expected to be smaller' should be backed by concrete experiments. Even though the memory requirements for different accuracy settings are shown, it would be useful to see how it translates to wall-clock time overhead.\n3. Accuracy of the approximation: It is clear that an approximation of V is necessary to make the tool feasible in practical settings. However, the paper doesn't investigate the accuracy of the approximate spectra obtained from the method. Authors claim that the main contribution of the paper is delivering directional derivatives and GGN curvatures, but the accuracy of those quantities is not verified in any way in the paper. Providing some guarantees on the accuracy would be crucial, especially since the method is proposed as an analysis tool. For instance, taking a look at Figure 1 shows that the exact GGN and MC approximate spectra show similar characteristics, but how large is the difference in the top eigenspaces? Small errors in top eigendirections can greatly impact downstream applications where the quantities are used. An in-depth quantitative analysis of the cost savings-accuracy trade-off would be essential.\nFurthermore, I am concerned about the accuracy of the method when applied to datasets with larger number of classes. For instance, the CIFAR-100 results in Figure S.5 show that the MC approximate spectrum looks very different from the exact spectrum, even qualitatively. This can be a serious issue when scaling the method to datasets such as ImageNet.\n4. Damping: The directional damping method is interesting and is a neat application for the proposed tool, but 1) if it is meant only as a proof-of-concept, the discussion takes up a very significant portion of the paper (almost 2 pages) that could be used for supporting the main claims of the paper or 2) if it is meant as a main contribution and a practical method, then much more in-depth studies are needed to establish whether it improves damping in second-order methods.\n\nMinor: What does K denote exactly (first appearing in 4.1)? Is K=NC?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper highlights how the low-rank structure of the generalized Gauss-Newton (GGN) approximation of the Hessian can be used as a computationally efficient tool to study the loss landscape of deep neural networks. In particular, authors discuss methods to compute the full spectrum of the GGN and thus providing access to per-sample directional gradients and curvature approximation. Through the lens of the GGN spectrum, authors make observations on the geometry of the loss landscape and its evolution during training, and propose an adaptive damping technique for second-order optimizers that utilizes the GGN curvature information.", "main_review": "Strengths:\n1. The paper proposes a method tailored to work efficiently on state-of-the-art auto differentiation tools and GPU hardware. Such tools are essential in better understanding various deep learning phenomena, improving optimization algorithms and in aiding practitioners with hyperparameter choices.\n2. The observations on the alignment of gradients with top eigendirections correlates well with recent results on deep neural network loss landscapes, and I find the discussion on the evolution of directional gradient and curvature SNR interesting. This direction can be very promising in designing practical second-order optimizers that scale to contemporary models and datasets.\n3. The directional damping scheme is an interesting idea, especially the adjustment of damping based on the stochastic uncertainty in a given eigendirection.\n\nWeaknesses:\n1. Scalability to practical settings: It is not clear whether the proposed method scales for contemporary architectures and large-scale datasets. Since the low-rank structure of the GGN has been investigated before by others (as the authors also pointed out), the main contribution of this paper is the compute/memory efficiency and scalability of the proposed tool. However, the experiments are performed on CIFAR-10 with a small, 6-layer neural network and it is not clear how it would scale to architectures such as ResNet and datasets such as ImageNet. In case the method does not work well in these scenarios, the applicability of the tool is limited.\n2. Overhead: In-depth study on the compute and memory overhead of the method would be valuable. For example statements such as 'However, the practical overhead is expected to be smaller' should be backed by concrete experiments. Even though the memory requirements for different accuracy settings are shown, it would be useful to see how it translates to wall-clock time overhead.\n3. Accuracy of the approximation: It is clear that an approximation of V is necessary to make the tool feasible in practical settings. However, the paper doesn't investigate the accuracy of the approximate spectra obtained from the method. Authors claim that the main contribution of the paper is delivering directional derivatives and GGN curvatures, but the accuracy of those quantities is not verified in any way in the paper. Providing some guarantees on the accuracy would be crucial, especially since the method is proposed as an analysis tool. For instance, taking a look at Figure 1 shows that the exact GGN and MC approximate spectra show similar characteristics, but how large is the difference in the top eigenspaces? Small errors in top eigendirections can greatly impact downstream applications where the quantities are used. An in-depth quantitative analysis of the cost savings-accuracy trade-off would be essential.\nFurthermore, I am concerned about the accuracy of the method when applied to datasets with larger number of classes. For instance, the CIFAR-100 results in Figure S.5 show that the MC approximate spectrum looks very different from the exact spectrum, even qualitatively. This can be a serious issue when scaling the method to datasets such as ImageNet.\n4. Damping: The directional damping method is interesting and is a neat application for the proposed tool, but 1) if it is meant only as a proof-of-concept, the discussion takes up a very significant portion of the paper (almost 2 pages) that could be used for supporting the main claims of the paper or 2) if it is meant as a main contribution and a practical method, then much more in-depth studies are needed to establish whether it improves damping in second-order methods.\n\nMinor: What does K denote exactly (first appearing in 4.1)? Is K=NC?", "summary_of_the_review": "The paper relies heavily on different forms of approximations to make the technique feasible, however  discussion on the accuracy of the approximated quantities is lacking, therefore it is difficult to judge how useful the obtained quantities are. Moreover, the scalability of the method to contemporary, practical architectures and datasets is not well-supported, thus the significance of the results is limited. Therefore, I recommend rejection of the paper in its current form. However, I believe that the paper has merit and some interesting contributions, and I am open to increasing my score if the authors address my concerns.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634958834108}], "openreview_url": "https://openreview.net/forum?id=e_D6AmszH4P", "arxiv_id": "2106.02624", "paper_pdf": "papers/e_D6AmszH4P.pdf", "paper_pdf_sha256": "a01c3b769fce368d47389399bcb8219d51cbb9033b0da8c7b04bf16f1c69107c", "paper_pdf_bytes": 1165873, "paper_pdf_source": "openreview", "code_url": "https://github.com/PwLo3K46/vivit", "code_repository": "PwLo3K46/vivit", "code_commit": "937642975be2ade122632d4eaef273461992d7ab", "code_archive": "repos/e_D6AmszH4P.zip", "code_archive_sha256": "044f27de9661d83d3c97bd25375261f3bc26558c802be2b34baf2f6d9c2dee7c", "code_archive_bytes": 6680925, "code_file_count": 106, "code_extensions": {".py": 101, ".sh": 5}, "github_disk_usage_kb": 6492, "github_languages": {"Python": 393933, "Makefile": 2117, "Shell": 567}, "github_archived": false, "github_pushed_at": "2021-10-03T20:35:05Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/vivit-curvature-access-through-the"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "oWy06SBgt4", "year": 2025, "status": "rejected", "title": "1-Bit FQT: Pushing the Limit of Fully Quantized Training to 1-bit", "authors": ["Chang Gao", "JingRen Hou", "Kang Zhao", "Jiaqi Wang", "Jianfei Chen", "Liping Jing"], "authorids": ["~Chang_Gao4", "~JingRen_Hou1", "~Kang_Zhao5", "~Jiaqi_Wang8", "~Jianfei_Chen1", "~Liping_Jing3"], "authors_source": "OpenReview API", "abstract": "Fully quantized training (FQT) accelerates the training of deep neural networks by quantizing the activations, weights, and gradients into lower precision. To explore the ultimate limit of FQT (the lowest achievable precision), we make a first attempt to 1-bit FQT. We provide a theoretical analysis of FQT based on Adam and SGD, revealing that the gradient variance influences the convergence of FQT. Building on these theoretical results, we introduce an Average 1-bit Quantization (AQ) strategy. The strategy leverages the heterogeneity of gradients to mitigate gradient variance by pruning less informative gradients and enhancing the numerical precision of remaining gradients. Additionally, we propose Sample Channel joint Quantization (SCQ), which utilizes different quantization strategies in the computation of weight gradients and activation gradients to ensure that the method is friendly to low-bitwidth hardware. Finally, we present a framework to deploy our algorithm. For fine-tuning VGGNet-16 and ResNet-18 on multiple datasets, our algorithm achieves an average accuracy improvement of approximately 6\\%, compared to per-sample quantization. Moreover, our training speedup can reach a maximum of 5.13× compared to full precision training.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "Ym4CCjZry0", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13456/Reviewer_VHKA"], "rating": 3, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 5, "summary": "This paper explores the limit of fully quantized training (FQT) by proposing a 1-bit quantization scheme for weights, activations, and gradients. \n\n1) The authors introduce an Average 1-bit Quantization (AQ) strategy to reduce gradient variance by pruning less informative gradients. \n\n2) Additionally, they propose Sample Channel joint Quantization (SCQ) to optimize quantization for weight and activation gradients, ensuring compatibility with low-bit hardware. \n\nExperimental results show that the proposed method achieves significant training speedup on VGGNet-16 and ResNet-18 across multiple datasets while maintaining minimal accuracy loss.", "review_text": "This paper explores the limit of fully quantized training (FQT) by proposing a 1-bit quantization scheme for weights, activations, and gradients. \n\n1) The authors introduce an Average 1-bit Quantization (AQ) strategy to reduce gradient variance by pruning less informative gradients. \n\n2) Additionally, they propose Sample Channel joint Quantization (SCQ) to optimize quantization for weight and activation gradients, ensuring compatibility with low-bit hardware. \n\nExperimental results show that the proposed method achieves significant training speedup on VGGNet-16 and ResNet-18 across multiple datasets while maintaining minimal accuracy loss.", "strengths": "(1) This paper presents a challenging 1-bit quantization method in the field of FQT, offering valuable insights for future low-bit quantization research through a theoretical analysis of gradient variance. \n\n(2) The proposed approach has been comprehensively evaluated across multiple datasets and models (such as CIFAR-10, CIFAR-100, VGGNet-16, and ResNet-18), effectively demonstrating its validity and generalizability. \n\n(3) The paper is well-structured, with detailed descriptions of the methodology and theoretical foundations, and the experimental results clearly showcase the potential of 1-bit FQT.", "weaknesses": "(1) The paper mainly compares traditional full-precision training and the existing PSQ method, lacking direct comparisons with recent low-bit quantization techniques (e.g., 2-bit or 4-bit methods), which limits the generality of its conclusions.\n\n(2) The method’s application mainly focuses on transfer learning for edge devices. Results of training from scratch or for larger models like ResNet-50 etc are weak or missing.\n\n(3) The SCQ method relies on per-channel and per-sample adjustments, which may increase hardware implementation complexity, especially for resource-limited devices. There is a lack of discussion on compatibility and scalability across different hardware.\n\n(4) The proposed method first needs to conduct QAT on ImageNet and then undergoes FQT on downstream datasets for large number of epochs, which constrains the application of the proposed method.", "questions": "(1) Can this method be applied to larger models like transformers, or maintain convergence in training-from-scratch scenarios on various models? If not, where do you see the main challenges?\n\n(2) Is the SCQ method scalable on resource-constrained hardware? For deployment of low-power devices, would the implementation complexity lead to additional computational overhead?\n\n(3) What's the batch-size of the FQT on edge devices?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper explores the limit of fully quantized training (FQT) by proposing a 1-bit quantization scheme for weights, activations, and gradients. \n\n1) The authors introduce an Average 1-bit Quantization (AQ) strategy to reduce gradient variance by pruning less informative gradients. \n\n2) Additionally, they propose Sample Channel joint Quantization (SCQ) to optimize quantization for weight and activation gradients, ensuring compatibility with low-bit hardware. \n\nExperimental results show that the proposed method achieves significant training speedup on VGGNet-16 and ResNet-18 across multiple datasets while maintaining minimal accuracy loss.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "(1) This paper presents a challenging 1-bit quantization method in the field of FQT, offering valuable insights for future low-bit quantization research through a theoretical analysis of gradient variance. \n\n(2) The proposed approach has been comprehensively evaluated across multiple datasets and models (such as CIFAR-10, CIFAR-100, VGGNet-16, and ResNet-18), effectively demonstrating its validity and generalizability. \n\n(3) The paper is well-structured, with detailed descriptions of the methodology and theoretical foundations, and the experimental results clearly showcase the potential of 1-bit FQT.", "weaknesses": "(1) The paper mainly compares traditional full-precision training and the existing PSQ method, lacking direct comparisons with recent low-bit quantization techniques (e.g., 2-bit or 4-bit methods), which limits the generality of its conclusions.\n\n(2) The method’s application mainly focuses on transfer learning for edge devices. Results of training from scratch or for larger models like ResNet-50 etc are weak or missing.\n\n(3) The SCQ method relies on per-channel and per-sample adjustments, which may increase hardware implementation complexity, especially for resource-limited devices. There is a lack of discussion on compatibility and scalability across different hardware.\n\n(4) The proposed method first needs to conduct QAT on ImageNet and then undergoes FQT on downstream datasets for large number of epochs, which constrains the application of the proposed method.", "questions": "(1) Can this method be applied to larger models like transformers, or maintain convergence in training-from-scratch scenarios on various models? If not, where do you see the main challenges?\n\n(2) Is the SCQ method scalable on resource-constrained hardware? For deployment of low-power devices, would the implementation complexity lead to additional computational overhead?\n\n(3) What's the batch-size of the FQT on edge devices?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730647671800}, {"id": "gqRI4etUI9", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13456/Reviewer_wULn"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 2, "confidence": 5, "summary": "The paper introduces the first attempt at achieving 1-bit fully quantized training (FQT), aka quantizing weights, activation and gradients at 1-bit. \n\nThey first provided a theoretical analysis demonstrating that under the assumption of a bounded gradient, the convergence rate for a convex loss is influenced by the variance of the gradient. This informs their hypothesis that the Adam optimizer is better suited for FTQ compared to SGF. \n\nThey also show that the quantized gradient variance increases quadratically as the bitwidth decreases, suggesting the variance of 1-bit per-sample gradient quantization is too large for training. To address this,  they propose the average 1-bit per-group quantization that randomly prunes samples proportionally to their dynamic range.  They also propose a similar scheme for weight gradient, called Sample Channel Joint Quantization, that applies the same idea to groups of channels.\n\nThey perform various fine-tuning experiments and consistently show they outperform vanilla per-sample quantization and are within an average 5.5% drop for ResNet18 and 5% for VVGNet-16 on a collection of downstream tasks. They also make the first attempt at FTQ from scratch on CIFAR-10/100 and ImageNet, but the results there are way below the QAT. They also test their implementation on Hygon and Raspberry Pi, showing consistent training speedups.", "review_text": "The paper introduces the first attempt at achieving 1-bit fully quantized training (FQT), aka quantizing weights, activation and gradients at 1-bit. \n\nThey first provided a theoretical analysis demonstrating that under the assumption of a bounded gradient, the convergence rate for a convex loss is influenced by the variance of the gradient. This informs their hypothesis that the Adam optimizer is better suited for FTQ compared to SGF. \n\nThey also show that the quantized gradient variance increases quadratically as the bitwidth decreases, suggesting the variance of 1-bit per-sample gradient quantization is too large for training. To address this,  they propose the average 1-bit per-group quantization that randomly prunes samples proportionally to their dynamic range.  They also propose a similar scheme for weight gradient, called Sample Channel Joint Quantization, that applies the same idea to groups of channels.\n\nThey perform various fine-tuning experiments and consistently show they outperform vanilla per-sample quantization and are within an average 5.5% drop for ResNet18 and 5% for VVGNet-16 on a collection of downstream tasks. They also make the first attempt at FTQ from scratch on CIFAR-10/100 and ImageNet, but the results there are way below the QAT. They also test their implementation on Hygon and Raspberry Pi, showing consistent training speedups.", "strengths": "A very decent effort by the authors to provide a framework for 1-bit FTQ. To my knowledge, they are the first to undertake that daunting task. Even if the results are disappointing in terms of accuracy, it can be the basis of future work that will further close the gap. I summarise their strengths:\n\n* The first to tackle the difficult task of 1-bit FTQ, even if their contributions are limited IMO.\n* Solid presentation and easy-to-read paper.\n* A theoretical motivation for their suggested methods and of the optimizers.\n* Extensive experimentation covering many tasks, including image classification, object detection and NLP\n* A very detailed appendix including ablations studies and implementation details for reproducibility. I specifically like that they consider the generalizability of their approach by applying it to more advanced binary models and showing an improvement even at higher bits.", "weaknesses": "Whereas the paper is very well presented and tackles a novel and difficult however, its contribution is quite limited as it mainly builds on theory and methods already developed. Chen 2020 first introduces the relationship between quantized gradient and gradient range found in section 4.2. They motivate their per-sample quantization (PSQ) based on this and introduce the Block Householder Quantizer (BHQ). The authors compare their method to PSQ in their experiments but not to BHQ, which follows the same idea as their average 1-bit quantization. \n\nI also find their conclusion about the convergence properties problematic because there is little evidence to suggest that the loss function of neural networks is convex. In addition, they compare vanilla SGD to Adam without considering momentum SGD first.  Adam does indeed have great properties for training networks, but it’s hardly surprising that SGD without momentum fails in the high variance regime of low-bit quantization. This is why momentum is introduced to act as a low-pass filter. \n\nThe authors introduce their FTQ method mainly for on-device fine-tuning. However, they do not discuss certain limitations regarding data movement and memory restrictions on edge devices. According to the appendix, they use large batches of size 64 for on-device fine-tuning (which may not be realistic for real edge applications). In addition, the dequantization operation in the step of algorithm 1 would require storing the large high-precision tensors (assuming this is float32) in memory to first calculate the dynamic ranges and perform the pruning/quantization later.  At such low precision, the training would most likely become memory-bound. It would be useful to see a profiling of training on the device to better understand the impact of the high-precision operations", "questions": "* Fig. 1 what is the quantization scheme for the gradient?\n* Tables 2 & 3 are very small and hard to read. Can you please expand them a bit?\n* It would be interesting to see a comparison of their method to the Per-Sample Quantizer combined with the Block Householder Quantizer for activation gradients. \n* A profiling of the training process on either of the edge devices, focusing on the compute time spent on the high-precision operations, e.g. Batch-Norm gradient estimation, and the memory bandwidth of activation gradients.\n* Using the symbol of a Hessian $\\mathbf{H}$ to denote activations in section 3 is a bit confusing (but this is just a remark)\n* The equations in section 3.1 have no numbering. I believe that a quantization function must be applied in line 149 for the equations to hold. The product of two quantize tenses:  $\\hat{\\nabla}_{\\mathbf{H}^{(l)}} \\bar{\\mathbf{\\Theta}}^{(l)^T}$  will not any more be in the same precision as the two inputs. Shouldn't it be instead\n\n $$Q_g(\\hat{\\nabla}_{\\mathbf{H}^{(l)}} \\bar{\\mathbf{\\Theta}}^{(l)^T})$$", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces the first attempt at achieving 1-bit fully quantized training (FQT), aka quantizing weights, activation and gradients at 1-bit. \n\nThey first provided a theoretical analysis demonstrating that under the assumption of a bounded gradient, the convergence rate for a convex loss is influenced by the variance of the gradient. This informs their hypothesis that the Adam optimizer is better suited for FTQ compared to SGF. \n\nThey also show that the quantized gradient variance increases quadratically as the bitwidth decreases, suggesting the variance of 1-bit per-sample gradient quantization is too large for training. To address this,  they propose the average 1-bit per-group quantization that randomly prunes samples proportionally to their dynamic range.  They also propose a similar scheme for weight gradient, called Sample Channel Joint Quantization, that applies the same idea to groups of channels.\n\nThey perform various fine-tuning experiments and consistently show they outperform vanilla per-sample quantization and are within an average 5.5% drop for ResNet18 and 5% for VVGNet-16 on a collection of downstream tasks. They also make the first attempt at FTQ from scratch on CIFAR-10/100 and ImageNet, but the results there are way below the QAT. They also test their implementation on Hygon and Raspberry Pi, showing consistent training speedups.", "soundness": 3, "presentation": 4, "contribution": 2, "strengths": "A very decent effort by the authors to provide a framework for 1-bit FTQ. To my knowledge, they are the first to undertake that daunting task. Even if the results are disappointing in terms of accuracy, it can be the basis of future work that will further close the gap. I summarise their strengths:\n\n* The first to tackle the difficult task of 1-bit FTQ, even if their contributions are limited IMO.\n* Solid presentation and easy-to-read paper.\n* A theoretical motivation for their suggested methods and of the optimizers.\n* Extensive experimentation covering many tasks, including image classification, object detection and NLP\n* A very detailed appendix including ablations studies and implementation details for reproducibility. I specifically like that they consider the generalizability of their approach by applying it to more advanced binary models and showing an improvement even at higher bits.", "weaknesses": "Whereas the paper is very well presented and tackles a novel and difficult however, its contribution is quite limited as it mainly builds on theory and methods already developed. Chen 2020 first introduces the relationship between quantized gradient and gradient range found in section 4.2. They motivate their per-sample quantization (PSQ) based on this and introduce the Block Householder Quantizer (BHQ). The authors compare their method to PSQ in their experiments but not to BHQ, which follows the same idea as their average 1-bit quantization. \n\nI also find their conclusion about the convergence properties problematic because there is little evidence to suggest that the loss function of neural networks is convex. In addition, they compare vanilla SGD to Adam without considering momentum SGD first.  Adam does indeed have great properties for training networks, but it’s hardly surprising that SGD without momentum fails in the high variance regime of low-bit quantization. This is why momentum is introduced to act as a low-pass filter. \n\nThe authors introduce their FTQ method mainly for on-device fine-tuning. However, they do not discuss certain limitations regarding data movement and memory restrictions on edge devices. According to the appendix, they use large batches of size 64 for on-device fine-tuning (which may not be realistic for real edge applications). In addition, the dequantization operation in the step of algorithm 1 would require storing the large high-precision tensors (assuming this is float32) in memory to first calculate the dynamic ranges and perform the pruning/quantization later.  At such low precision, the training would most likely become memory-bound. It would be useful to see a profiling of training on the device to better understand the impact of the high-precision operations", "questions": "* Fig. 1 what is the quantization scheme for the gradient?\n* Tables 2 & 3 are very small and hard to read. Can you please expand them a bit?\n* It would be interesting to see a comparison of their method to the Per-Sample Quantizer combined with the Block Householder Quantizer for activation gradients. \n* A profiling of the training process on either of the edge devices, focusing on the compute time spent on the high-precision operations, e.g. Batch-Norm gradient estimation, and the memory bandwidth of activation gradients.\n* Using the symbol of a Hessian $\\mathbf{H}$ to denote activations in section 3 is a bit confusing (but this is just a remark)\n* The equations in section 3.1 have no numbering. I believe that a quantization function must be applied in line 149 for the equations to hold. The product of two quantize tenses:  $\\hat{\\nabla}_{\\mathbf{H}^{(l)}} \\bar{\\mathbf{\\Theta}}^{(l)^T}$  will not any more be in the same precision as the two inputs. Shouldn't it be instead\n\n $$Q_g(\\hat{\\nabla}_{\\mathbf{H}^{(l)}} \\bar{\\mathbf{\\Theta}}^{(l)^T})$$", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730457490739}, {"id": "pzAcnlD8em", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13456/Reviewer_M3f1"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "First method to propose FQT where weight, activation and gradients are computed using 1 bit. The models have (average) 5% lower accuracy on classification datasets compared to 1 bit FQT using full precision gradients and less than 1% in simple datasets like Flowers and Pets dataset with an average speedup of 5.13x (training speedup) compared to FP32. The approach is explained with solid mathematical foundation and results are shown for both classification as detection models (thought only for fine-tuning in some cases). The 1 bit FTQ has a huge potential use case in memory restricted devices as the memory saving can be 32x (theoretically).", "review_text": "First method to propose FQT where weight, activation and gradients are computed using 1 bit. The models have (average) 5% lower accuracy on classification datasets compared to 1 bit FQT using full precision gradients and less than 1% in simple datasets like Flowers and Pets dataset with an average speedup of 5.13x (training speedup) compared to FP32. The approach is explained with solid mathematical foundation and results are shown for both classification as detection models (thought only for fine-tuning in some cases). The 1 bit FTQ has a huge potential use case in memory restricted devices as the memory saving can be 32x (theoretically).", "strengths": "- One of the first work to show end to end training using 1 bit which includes gradients and showed results on models like ResNet18 and VGG-16\n- Mathematical foundation of the approach is solid and reasonable (with few assumptions which may not hold true always)\n- Finetuning using 1 bit FQT shows promising results compared to similar QAT methods\n- 1 bit FTQ practical benefits is shows by using X-NOR 1 bit operations to save memory as well as computation.", "weaknesses": "- Memory saving analysis using 1 bit FTQ would add a lot of value to this method. Practically this method can be used to fine tune models running to tiny devices with very small memory and latency is not a prime concern for those devices.\n- The technique should be tested on deeper model to see the impact of vanishing gradient problem. This issue may be more prominent when gradient is calculated in lower precision.\n- Although L083 mentions that primary aim of this paper is to explore the ultimate limit of quantization instead of focussing on practical application performance, the usefulness of this method for practical use case can strengthen the paper.\n- The training speedup of 5.13x is compared with FP32 training, the Author(s) are advised to also provide speedup as well as accuracy numbers compared to full INT8 to see the extent of benefits of 1 bit FQT for real use cases.\n- The approach is tested on very old models like ResNet18 and VGG16, more results on other challenging but faster models will be interesting to see (eg. Mobilenet)\n- 1 bit FQT is only applicable to transfer learning and puts a big question on applicability of this approach.", "questions": "- One of the initial assumption that Adam is better in low precision regime is based on the sensitivity analysis in Figure 1. However, since it is the foundational assumption of the paper (L235, L221) and I would recommend to add results for more complex datasets as CiFAR10 is too simple and other datasets may change this assumption (mid size like Flowers/Pascal VOC and large size like Image-net/coco). Although this is also proved mathematically (L221) under certain assumptions (4.1 to 4.4), it would be good to verify these on other datasets\n- L270: Average 1-bit Quantization (AQ) - Is it similar to mixed precision quantization where some of the groups are quantized with higher precision (2, 4 bits etc)? If yes, what is there any information of percentage of 1-bit, other bits and %age of pruned channels for any dataset/model available ?\n- Table 1, L385, L386 : what is the percentage of “remaining groups” using b bandwidth?\n- L385, L386 : intuitively the accuracy of models where “remaining groups” use 8 bit should have more accuracy but Table 1 suggests otherwise. This is mentioned in L409 but adding intuition behind this would be helpful (confirmed in L426)\n- Effect of optimizer (L462) - The experiment may need to be expanded for other complex datasets (apart from CIFAR10)\n\nThe rating is primary reflection of above questions and general applicability of this approach for practical use cases. I would consider to change my rating based on the answer to above questions as as well as few points mentioned in weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "First method to propose FQT where weight, activation and gradients are computed using 1 bit. The models have (average) 5% lower accuracy on classification datasets compared to 1 bit FQT using full precision gradients and less than 1% in simple datasets like Flowers and Pets dataset with an average speedup of 5.13x (training speedup) compared to FP32. The approach is explained with solid mathematical foundation and results are shown for both classification as detection models (thought only for fine-tuning in some cases). The 1 bit FTQ has a huge potential use case in memory restricted devices as the memory saving can be 32x (theoretically).", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "- One of the first work to show end to end training using 1 bit which includes gradients and showed results on models like ResNet18 and VGG-16\n- Mathematical foundation of the approach is solid and reasonable (with few assumptions which may not hold true always)\n- Finetuning using 1 bit FQT shows promising results compared to similar QAT methods\n- 1 bit FTQ practical benefits is shows by using X-NOR 1 bit operations to save memory as well as computation.", "weaknesses": "- Memory saving analysis using 1 bit FTQ would add a lot of value to this method. Practically this method can be used to fine tune models running to tiny devices with very small memory and latency is not a prime concern for those devices.\n- The technique should be tested on deeper model to see the impact of vanishing gradient problem. This issue may be more prominent when gradient is calculated in lower precision.\n- Although L083 mentions that primary aim of this paper is to explore the ultimate limit of quantization instead of focussing on practical application performance, the usefulness of this method for practical use case can strengthen the paper.\n- The training speedup of 5.13x is compared with FP32 training, the Author(s) are advised to also provide speedup as well as accuracy numbers compared to full INT8 to see the extent of benefits of 1 bit FQT for real use cases.\n- The approach is tested on very old models like ResNet18 and VGG16, more results on other challenging but faster models will be interesting to see (eg. Mobilenet)\n- 1 bit FQT is only applicable to transfer learning and puts a big question on applicability of this approach.", "questions": "- One of the initial assumption that Adam is better in low precision regime is based on the sensitivity analysis in Figure 1. However, since it is the foundational assumption of the paper (L235, L221) and I would recommend to add results for more complex datasets as CiFAR10 is too simple and other datasets may change this assumption (mid size like Flowers/Pascal VOC and large size like Image-net/coco). Although this is also proved mathematically (L221) under certain assumptions (4.1 to 4.4), it would be good to verify these on other datasets\n- L270: Average 1-bit Quantization (AQ) - Is it similar to mixed precision quantization where some of the groups are quantized with higher precision (2, 4 bits etc)? If yes, what is there any information of percentage of 1-bit, other bits and %age of pruned channels for any dataset/model available ?\n- Table 1, L385, L386 : what is the percentage of “remaining groups” using b bandwidth?\n- L385, L386 : intuitively the accuracy of models where “remaining groups” use 8 bit should have more accuracy but Table 1 suggests otherwise. This is mentioned in L409 but adding intuition behind this would be helpful (confirmed in L426)\n- Effect of optimizer (L462) - The experiment may need to be expanded for other complex datasets (apart from CIFAR10)\n\nThe rating is primary reflection of above questions and general applicability of this approach for practical use cases. I would consider to change my rating based on the answer to above questions as as well as few points mentioned in weakness section.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730155225484}, {"id": "iW5Qt51BBN", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission13456/Reviewer_cWvz"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "The paper aims to accelerate training by applying binary quantization to the weights, activations, and gradients, making all three matrix multiplications binary and thus accelerating the process. This approach differs from standard Quantization-Aware Training (QAT), where typically only the weights and activations are quantized.\n\nThe paper also analyzes the variance introduced by gradient quantization when using both Adam and SGD optimizers. Based on this analysis, it proposes two methods for binary quantization:\n1. Activation gradient pruning, which removes less informative gradient groups to improve numerical precision in the remaining groups.\n2. Sample-channel joint quantization, which applies different quantization strategies to the gradients of weights and activations.", "review_text": "The paper aims to accelerate training by applying binary quantization to the weights, activations, and gradients, making all three matrix multiplications binary and thus accelerating the process. This approach differs from standard Quantization-Aware Training (QAT), where typically only the weights and activations are quantized.\n\nThe paper also analyzes the variance introduced by gradient quantization when using both Adam and SGD optimizers. Based on this analysis, it proposes two methods for binary quantization:\n1. Activation gradient pruning, which removes less informative gradient groups to improve numerical precision in the remaining groups.\n2. Sample-channel joint quantization, which applies different quantization strategies to the gradients of weights and activations.", "strengths": "The paper introduce an interesting analysis of the difference between SGD and Adam optimizer under binary quantization.", "weaknesses": "I see several weaknesses in the paper, with the primary one being the experiments section, which I find very underdeveloped. Presenting results solely on small datasets like CIFAR, especially in the current era of LLMs, is not sufficient for a conference like ICLR. Additionally, the networks used, such as VGG and ResNet18, are quite outdated.\n\nEven with these small datasets and older networks, the paper shows a significant performance gap compared to the baseline models, raising doubts about the approach's viability in real-world applications. The authors need to put more effort into this section to make the work more convincing.\n\nFurthermore, the practical acceleration claims are also problematic. The approach requires non-trivial preprocessing of data for every General Matrix Multiply (GEMM) operation, making it challenging to support natively in hardware", "questions": "1) Do you try to run the proposed method on modern LLMs - opensource githubs like nanoGPT is a good baseline.\n2) Is the speedup presented in Fig8 and table include all the preprocessing require in AQ when b > 1  - or it include only the matrix multiplication (which is unfair comparison)?   Can you also add comparison with BF16 which is usually ~2X and full time to train comparison.\n\n\n======================================================\n\nDear Authors,\n\nThank you for your efforts in addressing the questions and concerns raised.\nI decided to raise my score to 5 - still under the acceptance threshold.\nI feel that the experiments still require more effort- mainly in larger dataset and more modern models.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper aims to accelerate training by applying binary quantization to the weights, activations, and gradients, making all three matrix multiplications binary and thus accelerating the process. This approach differs from standard Quantization-Aware Training (QAT), where typically only the weights and activations are quantized.\n\nThe paper also analyzes the variance introduced by gradient quantization when using both Adam and SGD optimizers. Based on this analysis, it proposes two methods for binary quantization:\n1. Activation gradient pruning, which removes less informative gradient groups to improve numerical precision in the remaining groups.\n2. Sample-channel joint quantization, which applies different quantization strategies to the gradients of weights and activations.", "soundness": 2, "presentation": 2, "contribution": 1, "strengths": "The paper introduce an interesting analysis of the difference between SGD and Adam optimizer under binary quantization.", "weaknesses": "I see several weaknesses in the paper, with the primary one being the experiments section, which I find very underdeveloped. Presenting results solely on small datasets like CIFAR, especially in the current era of LLMs, is not sufficient for a conference like ICLR. Additionally, the networks used, such as VGG and ResNet18, are quite outdated.\n\nEven with these small datasets and older networks, the paper shows a significant performance gap compared to the baseline models, raising doubts about the approach's viability in real-world applications. The authors need to put more effort into this section to make the work more convincing.\n\nFurthermore, the practical acceleration claims are also problematic. The approach requires non-trivial preprocessing of data for every General Matrix Multiply (GEMM) operation, making it challenging to support natively in hardware", "questions": "1) Do you try to run the proposed method on modern LLMs - opensource githubs like nanoGPT is a good baseline.\n2) Is the speedup presented in Fig8 and table include all the preprocessing require in AQ when b > 1  - or it include only the matrix multiplication (which is unfair comparison)?   Can you also add comparison with BF16 which is usually ~2X and full time to train comparison.\n\n\n======================================================\n\nDear Authors,\n\nThank you for your efforts in addressing the questions and concerns raised.\nI decided to raise my score to 5 - still under the acceptance threshold.\nI feel that the experiments still require more effort- mainly in larger dataset and more modern models.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729429630448}], "openreview_url": "https://openreview.net/forum?id=oWy06SBgt4", "arxiv_id": "2408.14267", "paper_pdf": "papers/oWy06SBgt4.pdf", "paper_pdf_sha256": "8870b4a096c6d1b07e84f5c6f0c286ad92b11ada3914a34f25f7174abf611cc0", "paper_pdf_bytes": 1276291, "paper_pdf_source": "openreview", "code_url": "https://github.com/Gaochang-bjtu/1-bit-FQT", "code_repository": "Gaochang-bjtu/1-bit-FQT", "code_commit": "008b6188e4f1da2e6870f6c896afd89113ab2be4", "code_archive": "repos/oWy06SBgt4.zip", "code_archive_sha256": "fab8dfa4ea68f187dd85a6eafda5599a7769cf65604d8d8b7b78779a0f8296ef", "code_archive_bytes": 490626, "code_file_count": 63, "code_extensions": {".py": 51, ".h": 6, ".cu": 2, ".c": 2, ".cpp": 1, ".sh": 1}, "github_disk_usage_kb": 433, "github_languages": {"Python": 237989, "C": 73394, "Cuda": 11419, "Makefile": 956, "C++": 944, "Shell": 174}, "github_archived": false, "github_pushed_at": "2024-08-26T12:22:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/1-bit-fqt-pushing-the-limit-of-fully"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qud5pDnpzo", "year": 2024, "status": "rejected", "title": "ViP: A Differentially Private Foundation Model for Computer Vision", "authors": ["Yaodong Yu", "Maziar Sanjabi", "Yi Ma", "Kamalika Chaudhuri", "Chuan Guo"], "authorids": ["~Yaodong_Yu4", "~Maziar_Sanjabi1", "~Yi_Ma4", "~Kamalika_Chaudhuri1", "~Chuan_Guo1"], "authors_source": "OpenReview API", "abstract": "Artificial intelligence (AI) has seen a tremendous surge in capabilities thanks to the use of foundation models trained on internet-scale data. On the flip side, the uncurated nature of internet-scale data also poses significant privacy and legal risks, as they often contain personal information or copyrighted material that should not be trained on without permission. In this work, we propose as a mitigation measure a recipe to train foundation vision models with differential privacy (DP) guarantee. We identify masked autoencoders as a suitable learning algorithm that aligns well with DP-SGD, and train ViP---a Vision transformer with differential Privacy---under a strict privacy budget of $\\epsilon=8$ on the LAION400M dataset. We evaluate the quality of representation learned by ViP using standard downstream vision tasks; in particular, ViP achieves a (non-private) linear probing accuracy of 55.7% on ImageNet, comparable to that of end-to-end trained AlexNet (trained and evaluated on ImageNet). Our result suggests that scaling to internet-scale data can be practical for private learning.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "33fGxiMlWh", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2173/Reviewer_kYFB"], "rating": "8: accept, good paper", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The authors proposed VIP, a recipe of privately training vision foundation models through self-supervised learning. The main insights are two-fold: 1) masked autoencoder is a suitable algorithm for DP-SGD which allows per-sample gradient clipping, as opposed to contrastive learning; 2) warm-start through non-private synthetic pretraining can greatly accelerate the training. The authors conducted comprehensive experiments to demonstrate the effectiveness of VIP, showing that it surpasses state-of-the-art methods on a variety of learning tasks.", "review_text": "The authors proposed VIP, a recipe of privately training vision foundation models through self-supervised learning. The main insights are two-fold: 1) masked autoencoder is a suitable algorithm for DP-SGD which allows per-sample gradient clipping, as opposed to contrastive learning; 2) warm-start through non-private synthetic pretraining can greatly accelerate the training. The authors conducted comprehensive experiments to demonstrate the effectiveness of VIP, showing that it surpasses state-of-the-art methods on a variety of learning tasks.", "strengths": "- The motivation and insights are adequately delivered\n- Strong and comprehensive empirical results\n- Good writing and presentation", "weaknesses": "I don't see any apparent weakness of this work, but only a few minor suggestions:\n- The experiments should be running over multiple random seeds, and please include the standard deviations in the tables as well\n- It would be better to include a non-private version of VIP (i.e., non-private MAE) for comparison in the experiments (to reflect the cost of DP)\n- It would be better to emphasize in the title and abstract that this paper focuses on applying DP to SSL, which is different from most prior works that focus on applying DP to supervised learning\n- There is a concurrent work [1] which applied DP to the continued pretraining CLIP (using batched gradient clipping). The authors should discuss this work in Section 5\n- Section 2, Eq. equation 1 -> Eq. (1)\n\n\nReference\n\n[1] Huang, Alyssa, et al. \"Safeguarding Data in Multimodal AI: A Differentially Private Approach to CLIP Training.\" arXiv preprint arXiv:2306.08173 (2023).", "questions": "I don't have further questions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors proposed VIP, a recipe of privately training vision foundation models through self-supervised learning. The main insights are two-fold: 1) masked autoencoder is a suitable algorithm for DP-SGD which allows per-sample gradient clipping, as opposed to contrastive learning; 2) warm-start through non-private synthetic pretraining can greatly accelerate the training. The authors conducted comprehensive experiments to demonstrate the effectiveness of VIP, showing that it surpasses state-of-the-art methods on a variety of learning tasks.", "soundness": "4 excellent", "presentation": "4 excellent", "contribution": "3 good", "strengths": "- The motivation and insights are adequately delivered\n- Strong and comprehensive empirical results\n- Good writing and presentation", "weaknesses": "I don't see any apparent weakness of this work, but only a few minor suggestions:\n- The experiments should be running over multiple random seeds, and please include the standard deviations in the tables as well\n- It would be better to include a non-private version of VIP (i.e., non-private MAE) for comparison in the experiments (to reflect the cost of DP)\n- It would be better to emphasize in the title and abstract that this paper focuses on applying DP to SSL, which is different from most prior works that focus on applying DP to supervised learning\n- There is a concurrent work [1] which applied DP to the continued pretraining CLIP (using batched gradient clipping). The authors should discuss this work in Section 5\n- Section 2, Eq. equation 1 -> Eq. (1)\n\n\nReference\n\n[1] Huang, Alyssa, et al. \"Safeguarding Data in Multimodal AI: A Differentially Private Approach to CLIP Training.\" arXiv preprint arXiv:2306.08173 (2023).", "questions": "I don't have further questions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1699321556706}, {"id": "LKTXMZEFdJ", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2173/Reviewer_MY6t"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "The paper proposes how to train vision foundation models with privacy using the framework of differential privacy. The work targets only a single version of self-supervised encoders, namely MAE. The motivation is that the encoders can still leak private information. The ViP pre-trained encoder achieves accuracy for linear probing of 55.7% on ImageNet, which is comparable with AlexNet.", "review_text": "The paper proposes how to train vision foundation models with privacy using the framework of differential privacy. The work targets only a single version of self-supervised encoders, namely MAE. The motivation is that the encoders can still leak private information. The ViP pre-trained encoder achieves accuracy for linear probing of 55.7% on ImageNet, which is comparable with AlexNet.", "strengths": "1. The paper proposes a DP method for the large-scale encoder models.", "weaknesses": "1. The motivation is very poor: the encoders should not be trained on copyright or private data in the first place instead of preventing the detection that they were trained on such data.\n2. The method is limited to only the MAE and state-of-the-art methods such as DINO or DINO v2 are not supported, let alone the contrastive-based encoders such as SimCLR. \n3. The performance of the encoder trained on the large data is only 55.7% for the linear probing on ImageNet. One does not have to go through the same huge effort but instead, use a publicly trained model such as AlexNet to obtain the same performance.\n\nOther comments:\n1. Figure 1 is misplaced. First of all, it is way too early, since even no reference to the Figure is given on page 1. Second, the comparison is a strawman argument. ViP should be compared with the corresponding MAE encoder trained without DP.\n2. It is claimed that: \"More recently, Meehan et al. (2023) showed that non-generative vision SSL models can also be probed to reveal sensitive information about individual samples in its training data when given partial information.\" However, this work does not address the issues in the SSL encoders from Meehan et al. (2023), where no MAE encoders were considered!\n3. \" However, most vision SSL training algorithms are based on contrastive learning, where the objective function\ndepends on multiple samples in an entangled manner\". This is not correct. There are many non-contrastive SSL methods, for example, SimSiam [1], DINOv1 [2], or DINO v2 [3], which is a state-of-the-art SSL encoder. In general, [4] considers contrastive and non-contrastive encoders.\n\n**References:**\n1. \"Exploring Simple Siamese Representation Learning\". Xinlei Chen, Kaiming He. https://arxiv.org/abs/2011.10566 CVPR 2021.\n2. \"Emerging Properties in Self-Supervised Vision Transformers\" https://openaccess.thecvf.com/content/ICCV2021/papers/Caron_Emerging_Properties_in_Self-Supervised_Vision_Transformers_ICCV_2021_paper.pdf\n3. \"DINOv2: Learning Robust Visual Features without Supervision\" https://arxiv.org/abs/2304.07193\n4. \"Contrastive and Non-Contrastive Self-Supervised Learning Recover Global and Local Spectral Embedding Methods\" https://arxiv.org/pdf/2205.11508.pdf", "questions": "See the Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes how to train vision foundation models with privacy using the framework of differential privacy. The work targets only a single version of self-supervised encoders, namely MAE. The motivation is that the encoders can still leak private information. The ViP pre-trained encoder achieves accuracy for linear probing of 55.7% on ImageNet, which is comparable with AlexNet.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "1. The paper proposes a DP method for the large-scale encoder models.", "weaknesses": "1. The motivation is very poor: the encoders should not be trained on copyright or private data in the first place instead of preventing the detection that they were trained on such data.\n2. The method is limited to only the MAE and state-of-the-art methods such as DINO or DINO v2 are not supported, let alone the contrastive-based encoders such as SimCLR. \n3. The performance of the encoder trained on the large data is only 55.7% for the linear probing on ImageNet. One does not have to go through the same huge effort but instead, use a publicly trained model such as AlexNet to obtain the same performance.\n\nOther comments:\n1. Figure 1 is misplaced. First of all, it is way too early, since even no reference to the Figure is given on page 1. Second, the comparison is a strawman argument. ViP should be compared with the corresponding MAE encoder trained without DP.\n2. It is claimed that: \"More recently, Meehan et al. (2023) showed that non-generative vision SSL models can also be probed to reveal sensitive information about individual samples in its training data when given partial information.\" However, this work does not address the issues in the SSL encoders from Meehan et al. (2023), where no MAE encoders were considered!\n3. \" However, most vision SSL training algorithms are based on contrastive learning, where the objective function\ndepends on multiple samples in an entangled manner\". This is not correct. There are many non-contrastive SSL methods, for example, SimSiam [1], DINOv1 [2], or DINO v2 [3], which is a state-of-the-art SSL encoder. In general, [4] considers contrastive and non-contrastive encoders.\n\n**References:**\n1. \"Exploring Simple Siamese Representation Learning\". Xinlei Chen, Kaiming He. https://arxiv.org/abs/2011.10566 CVPR 2021.\n2. \"Emerging Properties in Self-Supervised Vision Transformers\" https://openaccess.thecvf.com/content/ICCV2021/papers/Caron_Emerging_Properties_in_Self-Supervised_Vision_Transformers_ICCV_2021_paper.pdf\n3. \"DINOv2: Learning Robust Visual Features without Supervision\" https://arxiv.org/abs/2304.07193\n4. \"Contrastive and Non-Contrastive Self-Supervised Learning Recover Global and Local Spectral Embedding Methods\" https://arxiv.org/pdf/2205.11508.pdf", "questions": "See the Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698774999924}, {"id": "eQpyJ21qEV", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2173/Reviewer_UcdW"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper focuses on differentially private representation learning for images. The authors empirically found that pretraining an MAE on synthetic images as the initialization then fine-tuning on private dataset with DP-SGD can boost the utility of learned features. Experiments show that it is even better than some non-private counterparts.", "review_text": "This paper focuses on differentially private representation learning for images. The authors empirically found that pretraining an MAE on synthetic images as the initialization then fine-tuning on private dataset with DP-SGD can boost the utility of learned features. Experiments show that it is even better than some non-private counterparts.", "strengths": "1. The differentially private representation learning is a well-motivated problem and has a wide range of applications.\n2. Experimental results look superior compared to baselines.", "weaknesses": "1. The proposed training recipe is not new. It contains two main steps: (1) pretraining on a synthetic dataset where there is no privacy concern, then (2) fine-tuning on a private dataset with DP-SGD. Something similar was proposed in many prior papers. To name a few, [1,2] in NLP (which are also cited by the authors), and [3] in CV. A minor difference is that these works choose to pretrain on a public real dataset instead of a synthetic dataset. Therefore, I do not see much novelty in this training recipe.\n2. The authors also claim that this recipe used by prior works, e.g. [1,2], is on supervised training. However, it is unclear what challenges you will have if you apply this recipe to SSL.\n3. The motivation for choosing MAE is not adequately clear. There are certainly other methods that can compute gradient in a disentangled manner. Naively, the ordinary autoencoder (without mask) should also be able to do this job. Why is MAE particularly picked? If there are more options, a comparison is desired.\n4. AlexNet is too old to compare, which was proposed more than 10 years ago. There are too many recent baselines you can compare. (Even SimCLR is not the latest, but at least it is within 3 years).\n5. Comparison in Table 1 is not fair. It looks to me that the ViP-LAION should be ViP-ImageNet-1k so that the readers can appreciate the benefit of an additional pretraining on the synthetic dataset.\n6. There are many claims in this paper without enough explanations. See my questions below.\n\n\n[1] Yu, Da, et al. \"Differentially Private Fine-tuning of Language Models.\" ICLR 2022.\n\n[2] Li, Xuechen, et al. \"Large language models can be strong differentially private learners.\" ICLR 2022.\n\n[3] Luo, Zelun, et al. \"Scalable differential privacy with sparse network finetuning.\" CVPR 2021.", "questions": "1. Point 1 in page 2, \"...attaining high-utility learned representations requires significantly more training data...\", why it is more than supervised learning?\n2. Point 3 in page 2, \"SSL training requires a much larger number of training epochs compared to supervised learning,...\" why?\n3. Still in page 2, \"We also show that it is tolerant to the large amount of Gaussian noise added in DP-SGD.\" Where do you show and why?\n4. How are your synthetic data generated? From a generative model? If so, does the training set of the generative model contain any private information?\n5. At the beginning of sec 3, \"1. Scaling up the number of training samples via SSL with masked autoencoder;\" what does this mean?\n6. At the end of sec 3.1, \"With more training samples, the magnitude of the injected noise becomes smaller.\" Why?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper focuses on differentially private representation learning for images. The authors empirically found that pretraining an MAE on synthetic images as the initialization then fine-tuning on private dataset with DP-SGD can boost the utility of learned features. Experiments show that it is even better than some non-private counterparts.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "1 poor", "strengths": "1. The differentially private representation learning is a well-motivated problem and has a wide range of applications.\n2. Experimental results look superior compared to baselines.", "weaknesses": "1. The proposed training recipe is not new. It contains two main steps: (1) pretraining on a synthetic dataset where there is no privacy concern, then (2) fine-tuning on a private dataset with DP-SGD. Something similar was proposed in many prior papers. To name a few, [1,2] in NLP (which are also cited by the authors), and [3] in CV. A minor difference is that these works choose to pretrain on a public real dataset instead of a synthetic dataset. Therefore, I do not see much novelty in this training recipe.\n2. The authors also claim that this recipe used by prior works, e.g. [1,2], is on supervised training. However, it is unclear what challenges you will have if you apply this recipe to SSL.\n3. The motivation for choosing MAE is not adequately clear. There are certainly other methods that can compute gradient in a disentangled manner. Naively, the ordinary autoencoder (without mask) should also be able to do this job. Why is MAE particularly picked? If there are more options, a comparison is desired.\n4. AlexNet is too old to compare, which was proposed more than 10 years ago. There are too many recent baselines you can compare. (Even SimCLR is not the latest, but at least it is within 3 years).\n5. Comparison in Table 1 is not fair. It looks to me that the ViP-LAION should be ViP-ImageNet-1k so that the readers can appreciate the benefit of an additional pretraining on the synthetic dataset.\n6. There are many claims in this paper without enough explanations. See my questions below.\n\n\n[1] Yu, Da, et al. \"Differentially Private Fine-tuning of Language Models.\" ICLR 2022.\n\n[2] Li, Xuechen, et al. \"Large language models can be strong differentially private learners.\" ICLR 2022.\n\n[3] Luo, Zelun, et al. \"Scalable differential privacy with sparse network finetuning.\" CVPR 2021.", "questions": "1. Point 1 in page 2, \"...attaining high-utility learned representations requires significantly more training data...\", why it is more than supervised learning?\n2. Point 3 in page 2, \"SSL training requires a much larger number of training epochs compared to supervised learning,...\" why?\n3. Still in page 2, \"We also show that it is tolerant to the large amount of Gaussian noise added in DP-SGD.\" Where do you show and why?\n4. How are your synthetic data generated? From a generative model? If so, does the training set of the generative model contain any private information?\n5. At the beginning of sec 3, \"1. Scaling up the number of training samples via SSL with masked autoencoder;\" what does this mean?\n6. At the end of sec 3.1, \"With more training samples, the magnitude of the injected noise becomes smaller.\" Why?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698354260824}], "openreview_url": "https://openreview.net/forum?id=qud5pDnpzo", "arxiv_id": "2306.08842", "paper_pdf": "papers/qud5pDnpzo.pdf", "paper_pdf_sha256": "c80967407f7be8f70c7184b94dfe867c5017182668b760d8ffe295bef486e6ee", "paper_pdf_bytes": 1619764, "paper_pdf_source": "openreview", "code_url": "https://github.com/facebookresearch/ViP-MAE", "code_repository": "facebookresearch/ViP-MAE", "code_commit": "d87fd2de456a07f2c21d9017850448d1caffd3ce", "code_archive": "repos/qud5pDnpzo.zip", "code_archive_sha256": "2a0fc2d815abab75df09a87afb520c51859a97120797bb2e3f4acdbe1f6be77b", "code_archive_bytes": 538430, "code_file_count": 19, "code_extensions": {".py": 19}, "github_disk_usage_kb": 509, "github_languages": {"Python": 121841}, "github_archived": true, "github_pushed_at": "2023-06-27T23:36:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/vip-a-differentially-private-foundation-model"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3jBXX9Xb1iz", "year": 2023, "status": "rejected", "title": "Multi-Label Knowledge Distillation", "authors": ["Peng-Hui Yang", "Ming-Kun Xie", "Chen-Chen Zong", "Lei Feng", "Gang Niu", "Masashi Sugiyama", "Sheng-Jun Huang"], "authorids": ["~Peng-Hui_Yang1", "~Ming-Kun_Xie1", "~Chen-Chen_Zong1", "~Lei_Feng1", "~Gang_Niu1", "~Masashi_Sugiyama1", "~Sheng-Jun_Huang1"], "authors_source": "OpenReview API", "abstract": "Existing knowledge distillation methods typically work by enforcing the consistency of output logits or intermediate feature maps between the teacher network and student network. Unfortunately, these methods can hardly be extended to the multi-label learning scenario. Because each instance is associated with multiple semantic labels, neither the prediction logits nor the feature maps obtained from the whole example can accurately transfer knowledge for each label. In this paper, we propose a novel multi-label knowledge distillation method. On one hand, it exploits the informative semantic knowledge from the logits by label decoupling with the one-versus-all reduction strategy; on the other hand, it enhances the distinctiveness of the learned feature representations by leveraging the structural information of label-wise embeddings. Experimental results on multiple benchmark datasets validate that the proposed method can avoid knowledge counteraction among labels, and achieve superior performance against diverse comparing methods.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "MvRYc5osnsF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2568/Reviewer_NCc1"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes doing knowledge distillation (KD) for multi-label ML models via KD approaches at both logit and feature representation levels.  The major novelty comes from the two proposed label-wise embedding distillation approaches, i.e., LED-CD and LED-ID, which leverage the multi-label info to implement the relation-based knowledge distillation. The experimental results, however, are insufficient for justifying or providing insights into the proposed approach.", "review_text": "Though the paper proposes an interesting idea of label-wise embedding distillation leveraging the multi-label info, the experimental setup in the paper, however, makes it difficult to justify the effectiveness of the proposed distillation method. Because of that, I recommend rejecting the paper. I also suggest that the authors improve the literature survey, implement a stronger baseline, and conduct experiments verifying the loss design in Eq. (4) and (7). ", "strengths": "# Strength\n* The paper is well-written and easy to understand.\n\n# Weakness\n* The literature survey for knowledge distillation in this paper is incomplete. The survey only covers two KD categories (logit-based and feature-based) and misses the relation-based category [1]. This miss is critical as the major novelty of this paper comes from proposing losses in this third KD category.\n* The proposed KD approach requires labeled data, which limits its application. I would question if Eq. (4) and (7) need to consider only the distances between positive label embeddings. It may also work if we simply include the embedding distances for all label pairs in the loss – disregarding whether the labels are positive or not. Would be good to have experiments verifying the design decision for Eq. (4) and (7).\n* Some claims in the paper lack supports from the experiments. For example, the authors claim the LED-ID loss would leverage spatial relations between labels. It would be good to have experimental results supporting such a claim.\n* Some experimental setup lacks explanation, e.g., how are the weights for MLD, LED-CD, and LED-ID losses determined as 10, 100, and 1000 in this paper?\n* The baselines are considerably weak as they are all with a single type of knowledge distillation, while the proposed L2D combines multiple distillation losses (i.e., BCE, MLD, LED-CD, and LED-ID). The proposed L2D loss may have the best results simply because of the loss combination (a known approach for performance improvement). A fair baseline should also be a system combining multiple losses.\n\n\n\n[1] Gou, J., Yu, B., Maybank, S.J. et al. Knowledge Distillation: A Survey. Int J Comput Vis 129, 1789–1819 (2021). https://doi.org/10.1007/s11263-021-01453-z  (available at https://arxiv.org/pdf/2006.05525.pdf )", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes doing knowledge distillation (KD) for multi-label ML models via KD approaches at both logit and feature representation levels.  The major novelty comes from the two proposed label-wise embedding distillation approaches, i.e., LED-CD and LED-ID, which leverage the multi-label info to implement the relation-based knowledge distillation. The experimental results, however, are insufficient for justifying or providing insights into the proposed approach.", "strength_and_weaknesses": "# Strength\n* The paper is well-written and easy to understand.\n\n# Weakness\n* The literature survey for knowledge distillation in this paper is incomplete. The survey only covers two KD categories (logit-based and feature-based) and misses the relation-based category [1]. This miss is critical as the major novelty of this paper comes from proposing losses in this third KD category.\n* The proposed KD approach requires labeled data, which limits its application. I would question if Eq. (4) and (7) need to consider only the distances between positive label embeddings. It may also work if we simply include the embedding distances for all label pairs in the loss – disregarding whether the labels are positive or not. Would be good to have experiments verifying the design decision for Eq. (4) and (7).\n* Some claims in the paper lack supports from the experiments. For example, the authors claim the LED-ID loss would leverage spatial relations between labels. It would be good to have experimental results supporting such a claim.\n* Some experimental setup lacks explanation, e.g., how are the weights for MLD, LED-CD, and LED-ID losses determined as 10, 100, and 1000 in this paper?\n* The baselines are considerably weak as they are all with a single type of knowledge distillation, while the proposed L2D combines multiple distillation losses (i.e., BCE, MLD, LED-CD, and LED-ID). The proposed L2D loss may have the best results simply because of the loss combination (a known approach for performance improvement). A fair baseline should also be a system combining multiple losses.\n\n\n\n[1] Gou, J., Yu, B., Maybank, S.J. et al. Knowledge Distillation: A Survey. Int J Comput Vis 129, 1789–1819 (2021). https://doi.org/10.1007/s11263-021-01453-z  (available at https://arxiv.org/pdf/2006.05525.pdf )", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written and easy to understand. The major novelty comes from the two proposed label-wise embedding distillation approaches, i.e., LED-CD and LED-ID, which leverage the multi-label info to implement the relation-based knowledge distillation. However, the experiments in the paper are insufficient to justify the effectiveness of the proposed approach.", "summary_of_the_review": "Though the paper proposes an interesting idea of label-wise embedding distillation leveraging the multi-label info, the experimental setup in the paper, however, makes it difficult to justify the effectiveness of the proposed distillation method. Because of that, I recommend rejecting the paper. I also suggest that the authors improve the literature survey, implement a stronger baseline, and conduct experiments verifying the loss design in Eq. (4) and (7). ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1667545954852}, {"id": "P-EZjdeJmgf", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2568/Reviewer_gMao"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposed a novel method for knowledge distillation in multi-label learning scenarios. Multi-label learning aims to solve the problem where multiple positive labels exists in a given single sample. However, general knowledge distillation methods mainly focus on single-label learning scenario while ignoring the latent label correlations/structural knowledge across the labels or the label embedding.\n\nThis work proposed a specific design framework which preserves the label structural knowledge in knowledge distillation training process. More specifically, a label decoupling with the one-versus all reduction strategy is deployed, and label-wise embedding is further explored to get the label structural knowledge across different label. Experimental results demonstrate the effectiveness of the proposed model.\n", "review_text": "This paper proposed a novel approach for multi-label knowledge distillation. The proposed modules are reasonable with high performance.", "strengths": "Strength:\n\nThe background and the motivation of the setting is well-introduced. The motivation of the work is reasonable and logical.\n\nThe proposed modules including the label decoupling and the embedding-based lobal correlation learning are clearly introduced with motivation and explanations. The motivation is rational and the explanation is reasonable.\n\nExperiments demonstrate the effectiveness of the proposed model.\n\nWeaknesses:\n\nEmbedding based label structural knowledge extraction/learning is a general approach for multi-label learning. To this end, there are one module which could be considered as a novel module. More explanations about why/how the embedding strategy is novel could be introduced.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposed a novel method for knowledge distillation in multi-label learning scenarios. Multi-label learning aims to solve the problem where multiple positive labels exists in a given single sample. However, general knowledge distillation methods mainly focus on single-label learning scenario while ignoring the latent label correlations/structural knowledge across the labels or the label embedding.\n\nThis work proposed a specific design framework which preserves the label structural knowledge in knowledge distillation training process. More specifically, a label decoupling with the one-versus all reduction strategy is deployed, and label-wise embedding is further explored to get the label structural knowledge across different label. Experimental results demonstrate the effectiveness of the proposed model.\n", "strength_and_weaknesses": "Strength:\n\nThe background and the motivation of the setting is well-introduced. The motivation of the work is reasonable and logical.\n\nThe proposed modules including the label decoupling and the embedding-based lobal correlation learning are clearly introduced with motivation and explanations. The motivation is rational and the explanation is reasonable.\n\nExperiments demonstrate the effectiveness of the proposed model.\n\nWeaknesses:\n\nEmbedding based label structural knowledge extraction/learning is a general approach for multi-label learning. To this end, there are one module which could be considered as a novel module. More explanations about why/how the embedding strategy is novel could be introduced.\n", "clarity,_quality,_novelty_and_reproducibility": "This paper clearly introduces the motivation and the proposed models. The training parameters are introduced in the paper. The novelty is fine.", "summary_of_the_review": "This paper proposed a novel approach for multi-label knowledge distillation. The proposed modules are reasonable with high performance.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666777271667}, {"id": "jNTELrS0xX", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2568/Reviewer_fqJx"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors propose a multi-label knowledge distillation method. On the one hand, it exploits the logit's informative semantic knowledge by label decoupling with the one-versus-all reduction strategy. On the other hand, it enhances the distinctiveness of the learned feature representations by leveraging the structural information of label-wise embeddings. ", "review_text": "Although it would be better to add several more results in the main paper, this paper addresses an interesting problem, and the performance of the proposed model seems favorable.\n\n\n-----after rebuttal-----\n\nAfter reading the comments from the other reviewers, I think the novelty of the proposed method is not strong enough.\nTherefore, I decided to reduce the score to 'below threshold.'", "strengths": "Strength\n- The authors tackle a relatively new problem, knowledge distillation for multi-label classification, and they propose a method suitable to address the problem of the existing approaches.\n- The proposed method is straightforward and shows favorable performance.\n\n\nWeakness\n- This work is not the first work to apply the knowledge distillation technique to the multi-label classification model. As the authors mentioned, Song et al. and Xu et al. also apply knowledge distillation to multi-label classification models. There are several more works with similar frameworks such as [1]. In the introduction, the authors should not only introduce the “distillation for multi-label classification” but also explain what the problem of the existing frameworks is and how the proposed method addresses the problem. The authors should explain why the proposed method shows favorable performance.\n[1] Multi-label image classification via knowledge distillation from weakly-supervised detection. ACM MM 2018.\n- The proposed method itself is simple and straightforward. Therefore, it would be better to further analyze the design choices of the proposed model to claim the impact of the proposed method. For example, why is the loss function in Eqn (6) used? Is there any better option for the authors to try for the loss function?\n- It would be better to have more experimental results in the main paper. For example, instead of the results for the VOC dataset in Fig. 4, it would be more informative to show the result with different architectures like Tables 1 and 2. Also, the results on the NUS-WIDE dataset, which is a famous multi-label classification dataset, are missing.\n- Finally, it would be better to have an analysis of the CAM before and after applying the proposed method, which is popularly used in knowledge distillation literature. For the analysis of Fig. 5, it is hard to see the correlation due to the large class number. I suggest analyzing the VOC dataset to better visualize the correlation between classes.\n\n\n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The authors propose a multi-label knowledge distillation method. On the one hand, it exploits the logit's informative semantic knowledge by label decoupling with the one-versus-all reduction strategy. On the other hand, it enhances the distinctiveness of the learned feature representations by leveraging the structural information of label-wise embeddings. ", "strength_and_weaknesses": "Strength\n- The authors tackle a relatively new problem, knowledge distillation for multi-label classification, and they propose a method suitable to address the problem of the existing approaches.\n- The proposed method is straightforward and shows favorable performance.\n\n\nWeakness\n- This work is not the first work to apply the knowledge distillation technique to the multi-label classification model. As the authors mentioned, Song et al. and Xu et al. also apply knowledge distillation to multi-label classification models. There are several more works with similar frameworks such as [1]. In the introduction, the authors should not only introduce the “distillation for multi-label classification” but also explain what the problem of the existing frameworks is and how the proposed method addresses the problem. The authors should explain why the proposed method shows favorable performance.\n[1] Multi-label image classification via knowledge distillation from weakly-supervised detection. ACM MM 2018.\n- The proposed method itself is simple and straightforward. Therefore, it would be better to further analyze the design choices of the proposed model to claim the impact of the proposed method. For example, why is the loss function in Eqn (6) used? Is there any better option for the authors to try for the loss function?\n- It would be better to have more experimental results in the main paper. For example, instead of the results for the VOC dataset in Fig. 4, it would be more informative to show the result with different architectures like Tables 1 and 2. Also, the results on the NUS-WIDE dataset, which is a famous multi-label classification dataset, are missing.\n- Finally, it would be better to have an analysis of the CAM before and after applying the proposed method, which is popularly used in knowledge distillation literature. For the analysis of Fig. 5, it is hard to see the correlation due to the large class number. I suggest analyzing the VOC dataset to better visualize the correlation between classes.\n\n\n\n\n", "clarity,_quality,_novelty_and_reproducibility": "The proposed method is clear and novel. The quality of this paper looks favorable overall.", "summary_of_the_review": "Although it would be better to add several more results in the main paper, this paper addresses an interesting problem, and the performance of the proposed model seems favorable.\n\n\n-----after rebuttal-----\n\nAfter reading the comments from the other reviewers, I think the novelty of the proposed method is not strong enough.\nTherefore, I decided to reduce the score to 'below threshold.'", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666678574842}, {"id": "8NS8buOV0g4", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2568/Reviewer_c6wM"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper presents the multi-label knowledge distillation to address the issue of multiple semantic labels in multi-label learning scenarios. MKD and LED are introduced for network learning. Experiments on several datasets demonstrate its superiority.", "review_text": "My main concern lies in the novelty, since both multi-label classification and KD have been well studied, and this paper does not provide some impressive new design for this task.", "strengths": "Strength:\n\n(1) The organization is good, which makes the paper easy to follow.\n\n(2) The method is simple, and the results on several datasets seem good.\n\nWeaknesses:\n\n(1) The presentation could be improved. For example, the abbreviation \"MLL\" lacks explanation. The meaning of the red line in Figure 1 is also unclear, and I guess it is the result of the teacher.\n\n(2) For the BCE loss, it is commonly used for multi-label classification. The simple distillation of MLD for this predicted probability is not novel. Moreover, the structural relation by class-aware label-wise embedding distillation and instance-aware label-wise embedding distillation is similar to the global graph and local graph-based similarity consistency. In this case, this paper can be regarded as a combination of existing multi-label learning and KD methods, and the overall novelty is not satisfying. Though the authors present several different distance measurements, the overall contribution is still insufficient.\n\n(3) The authors mainly compare with traditional KD methods. Please explain how these methods deal with the multi-class situation in the paper.\n\n(4) I would like to see a detailed comparison with existing multi-label methods and KD methods in the loss function, respectively.\n\n(5) According to the results in Tables 1 and 2, the L2D even achieves much better results than the teacher. Please explain this phenomenon.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper presents the multi-label knowledge distillation to address the issue of multiple semantic labels in multi-label learning scenarios. MKD and LED are introduced for network learning. Experiments on several datasets demonstrate its superiority.", "strength_and_weaknesses": "Strength:\n\n(1) The organization is good, which makes the paper easy to follow.\n\n(2) The method is simple, and the results on several datasets seem good.\n\nWeaknesses:\n\n(1) The presentation could be improved. For example, the abbreviation \"MLL\" lacks explanation. The meaning of the red line in Figure 1 is also unclear, and I guess it is the result of the teacher.\n\n(2) For the BCE loss, it is commonly used for multi-label classification. The simple distillation of MLD for this predicted probability is not novel. Moreover, the structural relation by class-aware label-wise embedding distillation and instance-aware label-wise embedding distillation is similar to the global graph and local graph-based similarity consistency. In this case, this paper can be regarded as a combination of existing multi-label learning and KD methods, and the overall novelty is not satisfying. Though the authors present several different distance measurements, the overall contribution is still insufficient.\n\n(3) The authors mainly compare with traditional KD methods. Please explain how these methods deal with the multi-class situation in the paper.\n\n(4) I would like to see a detailed comparison with existing multi-label methods and KD methods in the loss function, respectively.\n\n(5) According to the results in Tables 1 and 2, the L2D even achieves much better results than the teacher. Please explain this phenomenon.", "clarity,_quality,_novelty_and_reproducibility": "Please refer to my detailed comments above.", "summary_of_the_review": "My main concern lies in the novelty, since both multi-label classification and KD have been well studied, and this paper does not provide some impressive new design for this task.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666598382608}, {"id": "HK9ro0fiUON", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2568/Reviewer_SHPH"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a distillation method on multi label image classification task. It introduces a multi label logits distillation task and intra-class/instance embedding distillation loss. And performs experiments on COCO and VOC dataset demonstrating the effectiveness of the\nproposed method. The source code is provided for reproducibility.", "review_text": "This paper is well-written and the experiments result shows the improvement over the previous method on certain setting. But more experiments are needed to show the improvements over previous methods. Also there is a concern about the novelty.", "strengths": "Strength:\n1. The approach is simple and easy to implement.\n2. The experiments result shows the improvement over the previous method on resolution 224x224.\n3. The code is released for reproducing the results.\n\nWeakness:\n1. Although the title is Multi-Label Knowledge Distillation (MLKD), and as stated in first contribution this paper proposes a general MLKD framework, it is only about multi label image classification.\n2. The experiment part only covers VOC and COCO with 224x224 resolutions. However, previous work [1] [2] mainly working on 448x448, and thus it is hard to compare and validate the result of this work. Also it would be better to include the result on NUS-WIDE.\n3. There seems not enough novelty by only adding a intra-class/instance embedding distillation loss.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a distillation method on multi label image classification task. It introduces a multi label logits distillation task and intra-class/instance embedding distillation loss. And performs experiments on COCO and VOC dataset demonstrating the effectiveness of the\nproposed method. The source code is provided for reproducibility.", "strength_and_weaknesses": "Strength:\n1. The approach is simple and easy to implement.\n2. The experiments result shows the improvement over the previous method on resolution 224x224.\n3. The code is released for reproducing the results.\n\nWeakness:\n1. Although the title is Multi-Label Knowledge Distillation (MLKD), and as stated in first contribution this paper proposes a general MLKD framework, it is only about multi label image classification.\n2. The experiment part only covers VOC and COCO with 224x224 resolutions. However, previous work [1] [2] mainly working on 448x448, and thus it is hard to compare and validate the result of this work. Also it would be better to include the result on NUS-WIDE.\n3. There seems not enough novelty by only adding a intra-class/instance embedding distillation loss.", "clarity,_quality,_novelty_and_reproducibility": "The paper is writing clearly and easy to follow. The class/instance-aware embedding distillation loss seems common in other areas so not sure if this counts as a novel contribution.", "summary_of_the_review": "This paper is well-written and the experiments result shows the improvement over the previous method on certain setting. But more experiments are needed to show the improvements over previous methods. Also there is a concern about the novelty.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666575208216}], "openreview_url": "https://openreview.net/forum?id=3jBXX9Xb1iz", "arxiv_id": "2308.06453", "paper_pdf": "papers/3jBXX9Xb1iz.pdf", "paper_pdf_sha256": "d4f01bc2e752d95bd47b969bb88d5f5e5384b19c5600a454b7b999267efe7e28", "paper_pdf_bytes": 1642188, "paper_pdf_source": "openreview", "code_url": "https://github.com/penghui-yang/L2D", "code_repository": "penghui-yang/L2D", "code_commit": "ea7fa5581de3cd0dba2f4d5ba3e3aaed78837a58", "code_archive": "repos/3jBXX9Xb1iz.zip", "code_archive_sha256": "9169733a62bb098bd068e2bdc00b03d52247427651d2bb544dfd25631748b2e5", "code_archive_bytes": 1271502, "code_file_count": 38, "code_extensions": {".py": 38}, "github_disk_usage_kb": 1370, "github_languages": {"Python": 110295}, "github_archived": false, "github_pushed_at": "2024-04-24T05:30:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/multi-label-knowledge-distillation"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qLqeb9AjD2o", "year": 2022, "status": "rejected", "title": "Confidence-aware Training of Smoothed Classifiers for Certified Robustness", "authors": ["Jongheon Jeong", "Seojin Kim", "Jinwoo Shin"], "authorids": ["~Jongheon_Jeong1", "~Seojin_Kim2", "~Jinwoo_Shin1"], "authors_source": "OpenReview API", "abstract": "Any classifier can be \"smoothed out\" under Gaussian noise to build a new classifier that is provably robust to $\\ell_2$-adversarial perturbations, viz., by averaging its predictions over the noise, namely via randomized smoothing. Under the smoothed classifiers, the fundamental trade-off between accuracy and (adversarial) robustness has been well evidenced in the literature: i.e., increasing the robustness of a classifier for an input can be at the expense of decreased accuracy for some other inputs. In this paper, we propose a simple training method leveraging this trade-off for obtaining more robust smoothed classifiers, in particular, through a sample-wise control of robustness over the training samples. We enable this control feasible by investigating the correspondence between robustness and prediction confidence of smoothed classifiers: specifically, we propose to use the \"accuracy under Gaussian noise\" as an easy-to-compute proxy of adversarial robustness for each input. We differentiate the training objective depending on this proxy to filter out samples that are unlikely to benefit from the worst-case (adversarial) objective. Our experiments following the standard benchmarks consistently show that the proposed method, despite its simplicity, exhibits improved certified robustness upon existing state-of-the-art training methods.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "Irqx7x6ZY-u", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2375/Reviewer_girT"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a loss function for training the base classifier for randomized smoothed classifiers. Specifically, the loss distinguishes between training examples with high prediction confidence and those with low confidence. Several empirical tricks are applied to design the loss function to optimize the performance. Experiments on MNIST and CIFAR-10 show that the proposed training method is superior to existing state-of-the-art randomized smoothed classifiers, especially when the radius r is large.", "review_text": "While the provided experimental results look promising, the designed loss function (especially for the low-confidence loss) seems to be very complicated in my perspective and the explanation in Section 3 are not clear enough to support such design. In particular, I have the following questions for authors to provide more clarifications:\n\n1. The proposed CAT-RS loss relies on the prediction confidence of a classifier f for selecting examples. Is such f trained in advance or iteratively updated over the training process? How do you obtain the final certified smoothed classifier?\n\n2. I do not understand the design of Equation (7) for the low confidence examples. Discussions in Section 3.1 do not give a clear picture on such specific design choice. Can you provide a clarification on this? \n\n3. Follow the previous question, the cold start problem mentioned in Section 3.1 is also confusing to me. Why Equation (6) leads to a cold start problem? Why Equation (7) resolves this issue? The authors use many terms like “the easiest noise” and “harder ones” in the paragraph above Equation (7), which are vague and hard to parse.\n\n4. Figure 1 and Figure 2 are not well-explained. What do you mean by p_f=0.9 in Figure 1? Is it computed for a single example or computed as an averaged score for the whole dataset? What does areas with different colors represent in Figure 2? I do not understand the colored cross mark in Figure 2 as well.\n\nOther comments:\n\n1. The whole procedure for obtaining the final randomized smoothed classifier should be described in addition to the training loss.\n\n2. It would be better to provide experiments on larger dataset such as Image-Net.\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes a loss function for training the base classifier for randomized smoothed classifiers. Specifically, the loss distinguishes between training examples with high prediction confidence and those with low confidence. Several empirical tricks are applied to design the loss function to optimize the performance. Experiments on MNIST and CIFAR-10 show that the proposed training method is superior to existing state-of-the-art randomized smoothed classifiers, especially when the radius r is large.", "main_review": "While the provided experimental results look promising, the designed loss function (especially for the low-confidence loss) seems to be very complicated in my perspective and the explanation in Section 3 are not clear enough to support such design. In particular, I have the following questions for authors to provide more clarifications:\n\n1. The proposed CAT-RS loss relies on the prediction confidence of a classifier f for selecting examples. Is such f trained in advance or iteratively updated over the training process? How do you obtain the final certified smoothed classifier?\n\n2. I do not understand the design of Equation (7) for the low confidence examples. Discussions in Section 3.1 do not give a clear picture on such specific design choice. Can you provide a clarification on this? \n\n3. Follow the previous question, the cold start problem mentioned in Section 3.1 is also confusing to me. Why Equation (6) leads to a cold start problem? Why Equation (7) resolves this issue? The authors use many terms like “the easiest noise” and “harder ones” in the paragraph above Equation (7), which are vague and hard to parse.\n\n4. Figure 1 and Figure 2 are not well-explained. What do you mean by p_f=0.9 in Figure 1? Is it computed for a single example or computed as an averaged score for the whole dataset? What does areas with different colors represent in Figure 2? I do not understand the colored cross mark in Figure 2 as well.\n\nOther comments:\n\n1. The whole procedure for obtaining the final randomized smoothed classifier should be described in addition to the training loss.\n\n2. It would be better to provide experiments on larger dataset such as Image-Net.\n\n", "summary_of_the_review": "In summary, the empirical results of the paper show advantages of the proposed training objective in producing better certified smoothed classifiers. However, given the current low clarity of the paper, especially when introducing the design of the proposed loss function, I suggest a weak reject for this work.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635981403841}, {"id": "_955q8FI3NF", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2375/Reviewer_5emU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposed new loss functions during the training of classifiers for certified robustness via randomized smoothing. The new loss function treat samples with different confidence level differently. The main idea is to prioritize samples with high confidence because it provides more additional certified radius when its confidence grows.", "review_text": "Major comments:\n- The paper proposes a pretty novel method to boost the training for certified robustness. The experiments look descent to me although the hyperparameter selection seems random to me. I wonder if a cross-validation is performed.\n- The intuition behind the loss of high-confidence samples is that the noise samples provided is limited, which cannot guarantee the real $p_f(x,y)$ to be close enough to 1. This is especially important especially for ACR as increasing $p_f(x,y)$ can increase the certified radius more when $p_f(x,y)$ is closer to 1. However, the method used by the authors seems a little arbitrary to me. They find the worst noises within the $\\ell_2$-ball of sampled noise by PGD, and somehow normalize the worst noise. Unlike SmoothAdv or simply increasing the number of samples, there is no clear reason why this should work. \n- The intuition behind low-confidence part is also natural, basically giving ways to high-confidence part. I think the authors have done that part pretty well.\n\nMinor comments:\n- $K$ is used as the number of classes and the binomial random variable.\n- Sec 3.1, para 2, $p$ is not known here, so I think the $p$ in binomial distribution should be the empirical estimation $\\hat p$.\n- Sec 3.2, define $\\delta^*$ and how it is normalized.\n- Why is the binomial distribution necessary? You can probably just set $K$ to be the $pM$. \n- In Table 3, row (b), do you include the masking condition to $L^{high}$? If not, what is the performance if the condition is applied?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposed new loss functions during the training of classifiers for certified robustness via randomized smoothing. The new loss function treat samples with different confidence level differently. The main idea is to prioritize samples with high confidence because it provides more additional certified radius when its confidence grows.", "main_review": "Major comments:\n- The paper proposes a pretty novel method to boost the training for certified robustness. The experiments look descent to me although the hyperparameter selection seems random to me. I wonder if a cross-validation is performed.\n- The intuition behind the loss of high-confidence samples is that the noise samples provided is limited, which cannot guarantee the real $p_f(x,y)$ to be close enough to 1. This is especially important especially for ACR as increasing $p_f(x,y)$ can increase the certified radius more when $p_f(x,y)$ is closer to 1. However, the method used by the authors seems a little arbitrary to me. They find the worst noises within the $\\ell_2$-ball of sampled noise by PGD, and somehow normalize the worst noise. Unlike SmoothAdv or simply increasing the number of samples, there is no clear reason why this should work. \n- The intuition behind low-confidence part is also natural, basically giving ways to high-confidence part. I think the authors have done that part pretty well.\n\nMinor comments:\n- $K$ is used as the number of classes and the binomial random variable.\n- Sec 3.1, para 2, $p$ is not known here, so I think the $p$ in binomial distribution should be the empirical estimation $\\hat p$.\n- Sec 3.2, define $\\delta^*$ and how it is normalized.\n- Why is the binomial distribution necessary? You can probably just set $K$ to be the $pM$. \n- In Table 3, row (b), do you include the masking condition to $L^{high}$? If not, what is the performance if the condition is applied?", "summary_of_the_review": "My current assessment of this paper is slightly below the threshold because I am not very satisfied with the high-confidence part. I would like to hear from the authors about the mathematical intuition behind the current loss.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635921423347}, {"id": "aAUO9TFova5", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2375/Reviewer_dpjw"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies certified robustness via randomized smoothing (RS). RS has a fundamental accuracy and robustness tradeoff. The authors aim to enhance such tradeoff through a sample-wise control of robustness over the training samples. In particular, the authors investigate the correspondence between robustness and prediction confidence of smoothed classifiers and design a new loss function. The proposed method is evaluated on MNIST and CIFAR10.  ", "review_text": "Strengths:\n+The studied problem is important\n+Paper is easy to follow \n\nWeaknesses\n-Novelty is limited \n-Insufficient evaluation \n-Missing important references\n\nHow do you make predictions during the training process, e.g., how do you obtain $\\hat{p}_f $ in line 2 in Algorithm 1? Going forward, what’s the computational complexity of the proposed method? \n\nThe loss $L^{(high)}$ is only evaluated when  $\\mathbb{1}[(\\hat{p}_f=1)]$?  How many training samples with the associated generated noises make $(\\hat{p}_f=1$. I thinks this is a very strong constrain on my end. \n\nWhy $p_0$ is limited to the range (0,1], instead of $(1, \\infty)$? \n\nThe proposed method obtains comparable performance with the existing methods. How about the running time?\n\nThere is no result on the ImageNet.  \n\nThe following paper also considers model’s confidence. Please discuss with it.  \n\nKumar et al., “Certifying Confidence via Randomized Smoothing” \n\n\nThe following papers also derive certified robustness based on randomized smoothing \n\nWang et al. “Certified robustness of graph neural networks against adversarial structural perturbation via Randomized Smoothing”\n\nJia et al., “Certified Robustness for Top-k Predictions against Adversarial Perturbations via Randomized Smoothing” \n\nZhang et al., “Black-Box Certification with Randomized Smoothing: A Functional Optimization Based Framework”\n\nMohapatra et al., “Higher-Order Certification for Randomized Smoothing”\n\nKumar et al., “Certifying Confidence via Randomized Smoothing”\n\nKumar et al., “Curse of Dimensionality on Randomized Smoothing for Certifiable Robustness ”\n\nFischer  et al., “Certified Defense to Image Transformations via Randomized Smoothing”\n\nLee et al., “Tight Certificates of Adversarial Robustness for Randomly Smoothed Classifiers”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies certified robustness via randomized smoothing (RS). RS has a fundamental accuracy and robustness tradeoff. The authors aim to enhance such tradeoff through a sample-wise control of robustness over the training samples. In particular, the authors investigate the correspondence between robustness and prediction confidence of smoothed classifiers and design a new loss function. The proposed method is evaluated on MNIST and CIFAR10.  ", "main_review": "Strengths:\n+The studied problem is important\n+Paper is easy to follow \n\nWeaknesses\n-Novelty is limited \n-Insufficient evaluation \n-Missing important references\n\nHow do you make predictions during the training process, e.g., how do you obtain $\\hat{p}_f $ in line 2 in Algorithm 1? Going forward, what’s the computational complexity of the proposed method? \n\nThe loss $L^{(high)}$ is only evaluated when  $\\mathbb{1}[(\\hat{p}_f=1)]$?  How many training samples with the associated generated noises make $(\\hat{p}_f=1$. I thinks this is a very strong constrain on my end. \n\nWhy $p_0$ is limited to the range (0,1], instead of $(1, \\infty)$? \n\nThe proposed method obtains comparable performance with the existing methods. How about the running time?\n\nThere is no result on the ImageNet.  \n\nThe following paper also considers model’s confidence. Please discuss with it.  \n\nKumar et al., “Certifying Confidence via Randomized Smoothing” \n\n\nThe following papers also derive certified robustness based on randomized smoothing \n\nWang et al. “Certified robustness of graph neural networks against adversarial structural perturbation via Randomized Smoothing”\n\nJia et al., “Certified Robustness for Top-k Predictions against Adversarial Perturbations via Randomized Smoothing” \n\nZhang et al., “Black-Box Certification with Randomized Smoothing: A Functional Optimization Based Framework”\n\nMohapatra et al., “Higher-Order Certification for Randomized Smoothing”\n\nKumar et al., “Certifying Confidence via Randomized Smoothing”\n\nKumar et al., “Curse of Dimensionality on Randomized Smoothing for Certifiable Robustness ”\n\nFischer  et al., “Certified Defense to Image Transformations via Randomized Smoothing”\n\nLee et al., “Tight Certificates of Adversarial Robustness for Randomly Smoothed Classifiers”\n", "summary_of_the_review": "The studied problem is important and paper is easy to follow. However, the novelty is limited and the evaluation is insufficient.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1634755808808}, {"id": "7D-qG2Bilkb", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper2375/Reviewer_6h9p"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes CAT-RS which combines two novel losses for training base classifiers for randomized smoothing. The two novel losses aim at preserving clean accuracy for hard samples and improving the certified radius for easy samples. In term of ACR (average certified radius), the method achieves state-of-the-art on MNIST and CIFAR-10.", "review_text": "This paper focuses on improving the certified robustness via training a better base classifier for randomized smoothing. The empirical analysis of \"hard\" and \"easy\" training samples is novel and innovative, which backs up the proposed losses. However, the experimental results do not provide strong empirical evidence on the effectiveness of the proposed method, and some important evaluation data, like results on ImageNet, running time statistics, etc, are not reported.\n\nStrengths:\n- The analysis, including the separation of \"hard\" and \"easy\" training samples and different treatments for them, is novel and interesting.\n- The paper is very well written. The narration is coherent. The proposed method is clearly backed up by explanations.\n\nWeaknesses:\n- Based on the existing experimental results, the effectiveness of the proposed method is not apparent.\n\nI appreciate that major baselines in this field are included in the experimental evaluation.\n\nIn terms of ACR, the proposed approach achieves state-of-the-art. However, the margin is not large enough, only except CIFAR-10 sigma=1.00 case. The authors claim that the method may be more significant on dataset with high complexity. Then IMHO, ImageNet results are inevitable to support this.\n\nIn terms of certified test accuracy, the proposed approach cannot achieve good benign accuracy or high certified accuracy under small radius though this is one of its primary goals (a better trade-off between robustness and accuracy). On CIFAR-10, the empirical performance seems to be better. However, the proposed method still cannot improve the accuracy under all radii compared with Gaussian (Cohen et al), and under most radii, the performance improvements are within 1.5%.\n \n- Lack of ImageNet results.\n\nAll the compared baselines are evaluated on ImageNet. The high certified robustness on the standard ImageNet dataset is one of the unique strengths of randomized smoothing, and evaluation on ImageNet provides a more complete landscape on the method's performance on complex and large-scale datasets. Without ImageNet, it is hard to fairly evaluate the approach.\n\n- Lack of training time statistics.\n\nThe M=4 intuitively would make the training process take > 4 times longer than Gaussian (Cohen et al), > 2 times longer than Consistency (Jeong et al), and about the same time as SmoothAdv (with m=4). Is it true? Then we may need to consider whether it is valuable to take a much longer time to obtain a small improvement.\n\nQuestions:\n- On what set are the models evaluated?\nIs the evaluation conducted on a uniformly picked subset of the full test set as other baselines? If so, is the set size 1000?\n- Can you report the upper contour results across different sigma's for a clearer comparison?\n- If the experiment results on ImageNet are provided, based on those results, I will re-evaluate this work.\n\nSuggestion on the method:\n- Maybe you can also try to assign different weights to the training samples. For example, if $p_f(x,y)$ is already high, we may sample more (e.g., set M=8/16 for each sample) to have a more precise estimation of $p_f(x,y)$, then assign higher weights to the sample with larger $p_f(x,y)$, since improvement on that sample would contribute more to ACR.\n- In $L^{\\mathtt{high}}$ loss, you first find $\\delta^*$ then normalize them. Maybe you can incorporate the normalization inside - applying PGD on $\\mathrm{normalized}(\\delta^*)$.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes CAT-RS which combines two novel losses for training base classifiers for randomized smoothing. The two novel losses aim at preserving clean accuracy for hard samples and improving the certified radius for easy samples. In term of ACR (average certified radius), the method achieves state-of-the-art on MNIST and CIFAR-10.", "main_review": "This paper focuses on improving the certified robustness via training a better base classifier for randomized smoothing. The empirical analysis of \"hard\" and \"easy\" training samples is novel and innovative, which backs up the proposed losses. However, the experimental results do not provide strong empirical evidence on the effectiveness of the proposed method, and some important evaluation data, like results on ImageNet, running time statistics, etc, are not reported.\n\nStrengths:\n- The analysis, including the separation of \"hard\" and \"easy\" training samples and different treatments for them, is novel and interesting.\n- The paper is very well written. The narration is coherent. The proposed method is clearly backed up by explanations.\n\nWeaknesses:\n- Based on the existing experimental results, the effectiveness of the proposed method is not apparent.\n\nI appreciate that major baselines in this field are included in the experimental evaluation.\n\nIn terms of ACR, the proposed approach achieves state-of-the-art. However, the margin is not large enough, only except CIFAR-10 sigma=1.00 case. The authors claim that the method may be more significant on dataset with high complexity. Then IMHO, ImageNet results are inevitable to support this.\n\nIn terms of certified test accuracy, the proposed approach cannot achieve good benign accuracy or high certified accuracy under small radius though this is one of its primary goals (a better trade-off between robustness and accuracy). On CIFAR-10, the empirical performance seems to be better. However, the proposed method still cannot improve the accuracy under all radii compared with Gaussian (Cohen et al), and under most radii, the performance improvements are within 1.5%.\n \n- Lack of ImageNet results.\n\nAll the compared baselines are evaluated on ImageNet. The high certified robustness on the standard ImageNet dataset is one of the unique strengths of randomized smoothing, and evaluation on ImageNet provides a more complete landscape on the method's performance on complex and large-scale datasets. Without ImageNet, it is hard to fairly evaluate the approach.\n\n- Lack of training time statistics.\n\nThe M=4 intuitively would make the training process take > 4 times longer than Gaussian (Cohen et al), > 2 times longer than Consistency (Jeong et al), and about the same time as SmoothAdv (with m=4). Is it true? Then we may need to consider whether it is valuable to take a much longer time to obtain a small improvement.\n\nQuestions:\n- On what set are the models evaluated?\nIs the evaluation conducted on a uniformly picked subset of the full test set as other baselines? If so, is the set size 1000?\n- Can you report the upper contour results across different sigma's for a clearer comparison?\n- If the experiment results on ImageNet are provided, based on those results, I will re-evaluate this work.\n\nSuggestion on the method:\n- Maybe you can also try to assign different weights to the training samples. For example, if $p_f(x,y)$ is already high, we may sample more (e.g., set M=8/16 for each sample) to have a more precise estimation of $p_f(x,y)$, then assign higher weights to the sample with larger $p_f(x,y)$, since improvement on that sample would contribute more to ACR.\n- In $L^{\\mathtt{high}}$ loss, you first find $\\delta^*$ then normalize them. Maybe you can incorporate the normalization inside - applying PGD on $\\mathrm{normalized}(\\delta^*)$.", "summary_of_the_review": "I appreciate the analysis and the novelty of the proposed training method. However, the current loss design seems to be not effective enough, and there are a few important details missing in the experimental evaluation. Therefore, I do not lean towards acceptance. The authors may try to follow the proposed principle to improve the detail design of the current training approach.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1634548345269}], "openreview_url": "https://openreview.net/forum?id=qLqeb9AjD2o", "arxiv_id": "2212.09000", "paper_pdf": "papers/qLqeb9AjD2o.pdf", "paper_pdf_sha256": "c6511ef06638f442450886893c65398c8a10c07008cf210a63dfb44cb538c2eb", "paper_pdf_bytes": 837194, "paper_pdf_source": "openreview", "code_url": "https://github.com/alinlab/smoothing-catrs", "code_repository": "alinlab/smoothing-catrs", "code_commit": "d4bc576e7d373d158f087ba5744af8bb48466bb7", "code_archive": "repos/qLqeb9AjD2o.zip", "code_archive_sha256": "a7812fb83b825258884964045e66eed124829cfaf6b99a2f1cc6c1d62d5631bc", "code_archive_bytes": 7562839, "code_file_count": 18, "code_extensions": {".py": 18}, "github_disk_usage_kb": 6953, "github_languages": {"Python": 108214}, "github_archived": false, "github_pushed_at": "2023-01-19T06:05:09Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/confidence-aware-training-of-smoothed-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "33TBJachvOX", "year": 2021, "status": "rejected", "title": "How to compare adversarial robustness of classifiers from a global perspective", "authors": ["Niklas Risse", "Jan Philip Göpfert", "Christina Göpfert"], "authorids": ["~Niklas_Risse1", "jgoepfert@techfak.uni-bielefeld.de", "~Christina_Göpfert1"], "authors_source": "OpenReview API", "abstract": "Adversarial robustness of machine learning models has attracted considerable attention over recent years. Adversarial attacks undermine the reliability of and trust in machine learning models, but the construction of more robust models hinges on a rigorous understanding of adversarial robustness as a property of a given model. Point-wise measures for specific threat models are currently the most popular tool for comparing the robustness of classifiers and are used in most recent publications on adversarial robustness. In this work, we use robustness curves to show that point-wise measures fail to capture important global properties that are essential to reliably compare the robustness of different classifiers. We introduce new ways in which robustness curves can be used to systematically uncover these properties and provide concrete recommendations for researchers and practitioners when assessing and comparing the robustness of trained models. Furthermore, we characterize scale as a way to distinguish small and large perturbations, and relate it to inherent properties of data sets, demonstrating that robustness thresholds must be chosen accordingly. We hope that our work contributes to a shift of focus away from point-wise measures of robustness and towards a discussion of the question what kind of robustness could and should reasonably be expected. We release code to reproduce all experiments presented in this paper, which includes a Python module to calculate robustness curves for arbitrary data sets and classifiers, supporting a number of frameworks, including TensorFlow, PyTorch and JAX.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "GUwAw2WV_rL", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper622/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper surveys various adversarial defense methods on their performance when the perturbation distortion epsilon is increased. The author argues that robustness for a specific epsilon may not be enough and suggests robustness curves as an alternative. I think in general this paper provides some interesting empirical studies. My detailed comments are as follows.\n1. The overfitting of specific epsilon value is expected but interesting to see and I think this is one of the main reasons why a robustness curve is necessary. However, in this adversarial robustness community, the status quo is that researchers compare with each other on some specific datasets with some specific epsilon, for example, 0.3 for MNIST, 8/255 for CIFAR. I think one reason for choosing these values is that studying robustness under perturbation with larger distortion is kind of unnecessary because then the noise added is no longer imperceptible, which is at odds with adversarial examples' definition. I think the authors may need to provide more discussion on why studying robustness with epsilon>0.6 for MNIST is necessary, since we may misclassify those images as human. \n2. The authors suggest a robustness curve as an evaluation metric. However, I haven't seen any works on improving the robustness for all epsilon values globally. One possibility is that there is a trade-off between small epsilon performance and large epsilon performance, similar to the trade-off between robustness and accuracy (epsilon=0 performance). My suggestion is that the authors may define an \"area under the curve\" value just like in ROC curves for better comparison.\n3. (Minor) Please refrain from only using color to distinguish curves in figures as it may not be friendly to readers with color blindness. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "This paper provides some interesting insights.", "review": "This paper surveys various adversarial defense methods on their performance when the perturbation distortion epsilon is increased. The author argues that robustness for a specific epsilon may not be enough and suggests robustness curves as an alternative. I think in general this paper provides some interesting empirical studies. My detailed comments are as follows.\n1. The overfitting of specific epsilon value is expected but interesting to see and I think this is one of the main reasons why a robustness curve is necessary. However, in this adversarial robustness community, the status quo is that researchers compare with each other on some specific datasets with some specific epsilon, for example, 0.3 for MNIST, 8/255 for CIFAR. I think one reason for choosing these values is that studying robustness under perturbation with larger distortion is kind of unnecessary because then the noise added is no longer imperceptible, which is at odds with adversarial examples' definition. I think the authors may need to provide more discussion on why studying robustness with epsilon>0.6 for MNIST is necessary, since we may misclassify those images as human. \n2. The authors suggest a robustness curve as an evaluation metric. However, I haven't seen any works on improving the robustness for all epsilon values globally. One possibility is that there is a trade-off between small epsilon performance and large epsilon performance, similar to the trade-off between robustness and accuracy (epsilon=0 performance). My suggestion is that the authors may define an \"area under the curve\" value just like in ROC curves for better comparison.\n3. (Minor) Please refrain from only using color to distinguish curves in figures as it may not be friendly to readers with color blindness. ", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603937982522}, {"id": "qKYZoWeYJhe", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper622/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe authors advocate for the use of Robustness curves, plotting the adversarial accuracy as a function of the size of the neighbourhood region of allowed perturbation. The problem that they identify is that if you only evaluate adversarial accuracy at some numbers of threshold, you might conclude that some models (and the method that was used to train them) are more robust than others while it would be incorrect at other thresholds.\n\nGeneral comment:\nThe paper clearly describes the problem it deals with and reads easily. On the other hand, I am not entirely convinced by the importance of the problem. If you have a particular specification that you care about  (\"Robustness against l_inf attack for eps=0.1\"), you can just verify that. On the other hand, if your goal is \"I want my network to be robust\", then it's not properly defined, so of course it's hard to evaluate. Robustness curves will help there but there is still the problem that you might want to be robust to L_infinity, L_1, L_2, brightness difference, Wasserstein difference, changes of small patches... and then the robustness curves will not help you (unless of course you compute one for each difference). At the same time, they are quite a bit more costly to compute that simple point measures. While I agree that you are going to get more information if you compute a full robustness curve than if you sample it at a bunch of points, I'm not convinced that it is worth the effort.\n\nOne thing that I would recommend the authors is to make clearer the distinction between robustness curves as they described them (based on finding the closest adversarial example) vs. plotting on a robustness curve the pointwise measures and interpolating through them. This would be a much cheaper solution (and is essentially what reporting experimental results for a few chosen eps achieves). For example, the results the authors give in their Table 1. based on point wise measures wessentially achieves what the authors want to show: no defense strictly dominate the others.\n\nSpecific Comments:\n- The toy dataset example presented in Figure 1. is great and provides a great explanation of the problem that the authors identify. The constructive proof in Appendix A. is also quite interesting and really drives the point that the authors want to make.\n\n- The authors argue that robustness curves allow to compare \"global robustness properties and their dependence on a given classifier, distribution and distance function\". In practice, does it really give insights global robustness property? If I look at the L_infinity robustness curves, it does not tell me much about the robustness to L_2 perturbations.\n\n- I don't understand how the robustness curves are generated for the L-infinity case? If PGD is used to find adversarial examples, it's not likely to be the \"closest\" adversarial examples that is going to be found, in all likelihood it's going to be one that matches the epsilon given as input to the PGD attack (due to the projection)? There is nothing that is even encouraging the sample to be close to the input beyond the constraints used for projection.\nEven for the L2 distance, there is still the usual problem that, although the CW attack encourages to find the closest sample, there is no guarantee that it will, and the effectiveness it will have at doing so might depend from model to model. As a result, it's hard to decouple the robustness curve from the attack that it used internally.\n\nMinor Notes & Typos:\n- To improve the look of the paper, it should be possible to include manual linebreaks in the title so that it's not broken in every line. \n- The authors talk in the introduction about \"recently proposed robustness curves\" and cite a paper from 2020 for them, but it seems like those curves were already in use before that. \"On the effectiveness of interval bound propagation for training verifiably robust models\", Gowal et al. had some in 2018.; \"Provable Defenses against Adversarial Examples via the Convex Outer Adversarial Polytope\" had some (transposed) in 2017.\n- At the end of the introduction, the author say: \"It is our belief that the continued use of single perturbation thresholds in the adversarial robustness literature is due to a lack of awareness of the shortcomings of these measures\".  This seems overtly harsh. You could make the same point about training algorithms and say that authors only reporting on only a few datasets due it just out of lack of awareness of the fact that the relative performance of different algorithms will vary depending on the dataset. Given that computing robustness curves needs computing the closest adversary to a point, this is much more expensive so maybe computational cost might be the differentiating factor rather than \"lack of awareness\"?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Clear paper but importance of the problem not obvious.", "review": "Summary:\nThe authors advocate for the use of Robustness curves, plotting the adversarial accuracy as a function of the size of the neighbourhood region of allowed perturbation. The problem that they identify is that if you only evaluate adversarial accuracy at some numbers of threshold, you might conclude that some models (and the method that was used to train them) are more robust than others while it would be incorrect at other thresholds.\n\nGeneral comment:\nThe paper clearly describes the problem it deals with and reads easily. On the other hand, I am not entirely convinced by the importance of the problem. If you have a particular specification that you care about  (\"Robustness against l_inf attack for eps=0.1\"), you can just verify that. On the other hand, if your goal is \"I want my network to be robust\", then it's not properly defined, so of course it's hard to evaluate. Robustness curves will help there but there is still the problem that you might want to be robust to L_infinity, L_1, L_2, brightness difference, Wasserstein difference, changes of small patches... and then the robustness curves will not help you (unless of course you compute one for each difference). At the same time, they are quite a bit more costly to compute that simple point measures. While I agree that you are going to get more information if you compute a full robustness curve than if you sample it at a bunch of points, I'm not convinced that it is worth the effort.\n\nOne thing that I would recommend the authors is to make clearer the distinction between robustness curves as they described them (based on finding the closest adversarial example) vs. plotting on a robustness curve the pointwise measures and interpolating through them. This would be a much cheaper solution (and is essentially what reporting experimental results for a few chosen eps achieves). For example, the results the authors give in their Table 1. based on point wise measures wessentially achieves what the authors want to show: no defense strictly dominate the others.\n\nSpecific Comments:\n- The toy dataset example presented in Figure 1. is great and provides a great explanation of the problem that the authors identify. The constructive proof in Appendix A. is also quite interesting and really drives the point that the authors want to make.\n\n- The authors argue that robustness curves allow to compare \"global robustness properties and their dependence on a given classifier, distribution and distance function\". In practice, does it really give insights global robustness property? If I look at the L_infinity robustness curves, it does not tell me much about the robustness to L_2 perturbations.\n\n- I don't understand how the robustness curves are generated for the L-infinity case? If PGD is used to find adversarial examples, it's not likely to be the \"closest\" adversarial examples that is going to be found, in all likelihood it's going to be one that matches the epsilon given as input to the PGD attack (due to the projection)? There is nothing that is even encouraging the sample to be close to the input beyond the constraints used for projection.\nEven for the L2 distance, there is still the usual problem that, although the CW attack encourages to find the closest sample, there is no guarantee that it will, and the effectiveness it will have at doing so might depend from model to model. As a result, it's hard to decouple the robustness curve from the attack that it used internally.\n\nMinor Notes & Typos:\n- To improve the look of the paper, it should be possible to include manual linebreaks in the title so that it's not broken in every line. \n- The authors talk in the introduction about \"recently proposed robustness curves\" and cite a paper from 2020 for them, but it seems like those curves were already in use before that. \"On the effectiveness of interval bound propagation for training verifiably robust models\", Gowal et al. had some in 2018.; \"Provable Defenses against Adversarial Examples via the Convex Outer Adversarial Polytope\" had some (transposed) in 2017.\n- At the end of the introduction, the author say: \"It is our belief that the continued use of single perturbation thresholds in the adversarial robustness literature is due to a lack of awareness of the shortcomings of these measures\".  This seems overtly harsh. You could make the same point about training algorithms and say that authors only reporting on only a few datasets due it just out of lack of awareness of the fact that the relative performance of different algorithms will vary depending on the dataset. Given that computing robustness curves needs computing the closest adversary to a point, this is much more expensive so maybe computational cost might be the differentiating factor rather than \"lack of awareness\"?", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603900508809}, {"id": "3P_HB7Etsy", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper622/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper presents a theoretical scenario where point-wise measure of adversarial robustness falls short in comparing model robustness, then conduct experiments to show that robustness curve is a more meaningful evaluation metric from a global perspective.\n\n\nPros:\n\n\n+ The motivation is well explained. I mainly agree with the authors on the argument that point-wise measurement of robustness may be insufficient in explaining model robustness. Computing and visualizing robustness curves seems to be more meaningful and rigorous from a security perspective. \n\n+ Relating the choice of the perturbation strength to the underlying property of the data distribution is useful. The inter-class distances demonstrated in Table could potentially be used as a reference on determining the right scale of perturbation strength.\n\n\nCons:\n\n\n- The robustness results presented in Table 1 seems far below the state-of-the-art robustness. For instance, in the last row (\\epsilon=8/255), the robust test error of AT is 0.92, which is much higher than the reported statistics in (Madry et al., 2018). The author uses a very small 4-layer convolutional neural network for CIFAR-10 experiments, whereas the state-of-the-art robustness results are achieved using a much larger network, such as a ResNet architecture or a WideResNet architecture (refer to [1] for the current best robustness results on CIFAR-10). Thus, I recommend authors to rerun these experiments using a larger network.\n\n- Similar architecture is used for the robustness curves in Figures 3 and 4. This suboptimal choice of network architecture makes the argument 'it contains multiple intersections in the robustness curve' unconvincing. \n\n\nOther Questions or Comments:\n\n1. Most of the existing defenses against adversarial examples are typically trained using a specifically-chosen perturbation strength. If adopting robustness curves (or global robustness) as the evaluation criteria instead of point-wise robustness, how will this affect the existing adversarial training procedure?\n\n2. What does the distance statistics presented in Table 2 suggest for the typical choice of perturbation strength used in existing literature? \n\n3. The global robustness considered in this paper is robustness for varying perturbation strength. Is there a way to define the perturbation strength for different input locations based on your computed inter-class statistics?\n\n4. The bibliography style of the reference is not standard. Check if you are using the correct file.\n\n\n[1] Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter-free Attacks, Francesco Croce and Matthias Hein, ICML 2020\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Marginally below acceptance threshold", "review": "This paper presents a theoretical scenario where point-wise measure of adversarial robustness falls short in comparing model robustness, then conduct experiments to show that robustness curve is a more meaningful evaluation metric from a global perspective.\n\n\nPros:\n\n\n+ The motivation is well explained. I mainly agree with the authors on the argument that point-wise measurement of robustness may be insufficient in explaining model robustness. Computing and visualizing robustness curves seems to be more meaningful and rigorous from a security perspective. \n\n+ Relating the choice of the perturbation strength to the underlying property of the data distribution is useful. The inter-class distances demonstrated in Table could potentially be used as a reference on determining the right scale of perturbation strength.\n\n\nCons:\n\n\n- The robustness results presented in Table 1 seems far below the state-of-the-art robustness. For instance, in the last row (\\epsilon=8/255), the robust test error of AT is 0.92, which is much higher than the reported statistics in (Madry et al., 2018). The author uses a very small 4-layer convolutional neural network for CIFAR-10 experiments, whereas the state-of-the-art robustness results are achieved using a much larger network, such as a ResNet architecture or a WideResNet architecture (refer to [1] for the current best robustness results on CIFAR-10). Thus, I recommend authors to rerun these experiments using a larger network.\n\n- Similar architecture is used for the robustness curves in Figures 3 and 4. This suboptimal choice of network architecture makes the argument 'it contains multiple intersections in the robustness curve' unconvincing. \n\n\nOther Questions or Comments:\n\n1. Most of the existing defenses against adversarial examples are typically trained using a specifically-chosen perturbation strength. If adopting robustness curves (or global robustness) as the evaluation criteria instead of point-wise robustness, how will this affect the existing adversarial training procedure?\n\n2. What does the distance statistics presented in Table 2 suggest for the typical choice of perturbation strength used in existing literature? \n\n3. The global robustness considered in this paper is robustness for varying perturbation strength. Is there a way to define the perturbation strength for different input locations based on your computed inter-class statistics?\n\n4. The bibliography style of the reference is not standard. Check if you are using the correct file.\n\n\n[1] Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter-free Attacks, Francesco Croce and Matthias Hein, ICML 2020\n", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603849342198}, {"id": "9yBqshCrrPB", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper622/AnonReviewer4"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe paper showed that point-wise measures fail to capture important properties that are essential to compare the robustness of different classifiers. The authors introduced the recently proposed robustness curves to provide a global perspective including the scale as a way to distinguish small and large perturbations. \n\nPros:\n(1) How to compare the robustness is a very important question for current machine learning models. The author introduced a better criterion for this important question. \n(2) The paper is well written and easy to read.  The experiments and its discussion are strong proofs to support that the robustness curve should be the better criterion.\n(3) The authors released code to reproduce all the experiments for the current popular frameworks. It will be very helpful for the researchers to the advantages of this curve over point-wise measures. \n\nCons:\nOverall, this paper is impressive. \nThe only concern the reviewer has is the contribution compared to the previous work who proposed the robustness curve (C. Gopfert et al. 2020).  it seems this paper's contribution is highly based on  the proposal of robustness curve, and providing more explanations and discussions. (But this will not affect the importance of this paper)\n\n\n\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A good criterion for evaluating robustness of classifiers ", "review": "Summary:\nThe paper showed that point-wise measures fail to capture important properties that are essential to compare the robustness of different classifiers. The authors introduced the recently proposed robustness curves to provide a global perspective including the scale as a way to distinguish small and large perturbations. \n\nPros:\n(1) How to compare the robustness is a very important question for current machine learning models. The author introduced a better criterion for this important question. \n(2) The paper is well written and easy to read.  The experiments and its discussion are strong proofs to support that the robustness curve should be the better criterion.\n(3) The authors released code to reproduce all the experiments for the current popular frameworks. It will be very helpful for the researchers to the advantages of this curve over point-wise measures. \n\nCons:\nOverall, this paper is impressive. \nThe only concern the reviewer has is the contribution compared to the previous work who proposed the robustness curve (C. Gopfert et al. 2020).  it seems this paper's contribution is highly based on  the proposal of robustness curve, and providing more explanations and discussions. (But this will not affect the importance of this paper)\n\n\n\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603838498796}], "openreview_url": "https://openreview.net/forum?id=33TBJachvOX", "arxiv_id": "2004.10882", "paper_pdf": "papers/33TBJachvOX.pdf", "paper_pdf_sha256": "1e8ddcae1a9f2cc9a0d020723415270d632d2874d0ab59100222dc68977ad9b1", "paper_pdf_bytes": 631544, "paper_pdf_source": "openreview", "code_url": "https://github.com/niklasrisse/how-to-compare-adversarial-robustness-of-classifiers-from-a-global-perspective", "code_repository": "niklasrisse/how-to-compare-adversarial-robustness-of-classifiers-from-a-global-perspective", "code_commit": "2e03e661fbe9639913750e6044967d3ddbae4a1d", "code_archive": "repos/33TBJachvOX.zip", "code_archive_sha256": "405d6c804f570ceaf1e0134bf7be236003a5990892b60326e3b44c1212ee4803", "code_archive_bytes": 5190222, "code_file_count": 10, "code_extensions": {".ipynb": 8, ".py": 2}, "github_disk_usage_kb": 5052, "github_languages": {"Jupyter Notebook": 78815, "Python": 17587}, "github_archived": false, "github_pushed_at": "2020-10-02T15:35:44Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adversarial-examples-and-where-to-find-them"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YMCtQlm8Bc", "year": 2025, "status": "rejected", "title": "Critical Influence of Overparameterization on Sharpness-aware Minimization", "authors": ["Sungbin Shin", "Dongyeop Lee", "Maksym Andriushchenko", "Namhoon Lee"], "authorids": ["~Sungbin_Shin1", "~Dongyeop_Lee1", "~Maksym_Andriushchenko1", "~Namhoon_Lee1"], "authors_source": "OpenReview API", "abstract": "Training overparameterized neural networks often yields solutions with varying generalization capabilities, even when achieving similar training losses. Recent evidence indicates a strong correlation between the sharpness of a minimum and its generalization error, leading to increased interest in optimization methods that explicitly seek flatter minima for improved generalization. Despite its contemporary relevance to overparameterization, however, this sharpness-aware minimization (SAM) strategy has not been studied much yet as to exactly how it is affected by overparameterization. In this work, we analyze SAM under varying degrees of overparameterization, presenting both empirical and theoretical findings that reveal its critical influence on SAM's effectiveness. First, we conduct extensive numerical experiments across diverse domains and show that SAM consistently improves with overparameterization. Next, we attribute this phenomenon to the interplay between the enlarged solution space and increased implicit bias resulting from overparameterization. Furthermore, we show that this effect is particularly pronounced in practical settings involving label noise and sparsity, and yet, sufficient regularization is necessary. Last but not least, we provide other theoretical insights into how overparameterization helps SAM achieve minima with more uniform Hessian moments compared to SGD, and much faster convergence at a linear rate.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "o1hpcMtCbp", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2569/Reviewer_bQqz"], "rating": 6, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 5, "summary": "The authors perform experiments to \nmeasure the effect of overparameterization in SAM for a diverse set of tasks (Section 3). The goal is to observe how overparameterization\naffects SAM under various conditions, e.g.,  label noise, sparsity, and regularization. \nFurthermore, they prove that stable minima of SAM are flatter and have more uniform Hessian moments (if compared with SGD), and stochastic SAM can also converge at a linear. The overall contribution is that they empirically and theoretically proved that overparameterization critically affects SAM.", "review_text": "The authors perform experiments to \nmeasure the effect of overparameterization in SAM for a diverse set of tasks (Section 3). The goal is to observe how overparameterization\naffects SAM under various conditions, e.g.,  label noise, sparsity, and regularization. \nFurthermore, they prove that stable minima of SAM are flatter and have more uniform Hessian moments (if compared with SGD), and stochastic SAM can also converge at a linear. The overall contribution is that they empirically and theoretically proved that overparameterization critically affects SAM.", "strengths": "- interesting and well-motivated problem\n - very well written", "weaknesses": "- discussion on higher moments of Hessian is missing", "questions": "This is an interesting paper. I have a question: how does the convergence of higher-order moments of Hessian in your result compare with the other approaches in the literature, e.g., [1]? Can you provide a literature review on the previous works considering higher-order moments of Hessian to define flatness?\n\n[1] Tahmasebi, Behrooz, et al. \"A Universal Class of Sharpness-Aware Minimization Algorithms.\" Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors perform experiments to \nmeasure the effect of overparameterization in SAM for a diverse set of tasks (Section 3). The goal is to observe how overparameterization\naffects SAM under various conditions, e.g.,  label noise, sparsity, and regularization. \nFurthermore, they prove that stable minima of SAM are flatter and have more uniform Hessian moments (if compared with SGD), and stochastic SAM can also converge at a linear. The overall contribution is that they empirically and theoretically proved that overparameterization critically affects SAM.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "- interesting and well-motivated problem\n - very well written", "weaknesses": "- discussion on higher moments of Hessian is missing", "questions": "This is an interesting paper. I have a question: how does the convergence of higher-order moments of Hessian in your result compare with the other approaches in the literature, e.g., [1]? Can you provide a literature review on the previous works considering higher-order moments of Hessian to define flatness?\n\n[1] Tahmasebi, Behrooz, et al. \"A Universal Class of Sharpness-Aware Minimization Algorithms.\" Forty-first International Conference on Machine Learning.", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "N/A", "rating": 6, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730760254946}, {"id": "8ZbxS0zIF5", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2569/Reviewer_18YL"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper studies the critical influence of overparameterization on SAM.", "review_text": "This paper studies the critical influence of overparameterization on SAM.", "strengths": "The experiments in this paper are insightful, showing that overparameterization can increase the performance gap between SAM and SGD, while also highlighting the role of factors such as label noise, sparsity, weight decay, and early stopping in this phenomenon.", "weaknesses": "- It is widely recognized that overparameterization improves the generalization performance of SGD (e.g., ResNet-152 outperforms ResNet-18 on Cifar10 using SGD). And it is unsurprising that overparameterization can also improve the generalization performance of SAM. However, the non-trivial aspect is that overparameterization increases *the gap between SAM and SGD*, as shown in Figure 1. To avoid trivialization, reconsidering the title and abstract might be beneficial.\n\n- My primary concern is that the theoretical analysis can not support the main claim/findings (e.g., Figure 1). \n  - The linear stability analysis in Section 6.1 can not demonstrate how overparameterization affects SAM (i.e., that “greater overparameterization leads SAM to find flatter minima”).\n  - In the convergence analysis in Section 6.2: (i) it lacks relevance to generalization, the core focus of the paper; (ii) it does not clarify how the degree of overparameterization influences the convergence speed within the interpolation regime; (iii) in the interpolation regime, the exponential convergence result also holds for SGD (Bassily et al., 2018). The proof for SAM appears to be a minor extension from SGD, which treats SAM as SGD + small permutation.\n  - The theoretical analysis in Section 6.1/6.2 focus on a variant of SAM (SAM without normalization) for simplification, rather than the original SAM. While not a major issue, the formulation should be clearly presented in Section 6 to prevent potential reader confusion.", "questions": "Could the authors provide adequate theoretical support for the main findings? e.g., an analysis on diagonal linear networks.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the critical influence of overparameterization on SAM.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "The experiments in this paper are insightful, showing that overparameterization can increase the performance gap between SAM and SGD, while also highlighting the role of factors such as label noise, sparsity, weight decay, and early stopping in this phenomenon.", "weaknesses": "- It is widely recognized that overparameterization improves the generalization performance of SGD (e.g., ResNet-152 outperforms ResNet-18 on Cifar10 using SGD). And it is unsurprising that overparameterization can also improve the generalization performance of SAM. However, the non-trivial aspect is that overparameterization increases *the gap between SAM and SGD*, as shown in Figure 1. To avoid trivialization, reconsidering the title and abstract might be beneficial.\n\n- My primary concern is that the theoretical analysis can not support the main claim/findings (e.g., Figure 1). \n  - The linear stability analysis in Section 6.1 can not demonstrate how overparameterization affects SAM (i.e., that “greater overparameterization leads SAM to find flatter minima”).\n  - In the convergence analysis in Section 6.2: (i) it lacks relevance to generalization, the core focus of the paper; (ii) it does not clarify how the degree of overparameterization influences the convergence speed within the interpolation regime; (iii) in the interpolation regime, the exponential convergence result also holds for SGD (Bassily et al., 2018). The proof for SAM appears to be a minor extension from SGD, which treats SAM as SGD + small permutation.\n  - The theoretical analysis in Section 6.1/6.2 focus on a variant of SAM (SAM without normalization) for simplification, rather than the original SAM. While not a major issue, the formulation should be clearly presented in Section 6 to prevent potential reader confusion.", "questions": "Could the authors provide adequate theoretical support for the main findings? e.g., an analysis on diagonal linear networks.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730455309666}, {"id": "PFoZEAd8e2", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2569/Reviewer_8b2s"], "rating": 6, "soundness": 2, "presentation": 3, "contribution": 1, "confidence": 2, "summary": "This paper claimed that SAM benefits from the over-parameterization. They find that SAM could find different (e.g. simpler) solutions than SGD in the over-parameterization regime. In addition, they find that SAM is robust to label noise and sparse parameterization. They also include some theoretical analyses, thought these analyses are irrelevant to their main findings.", "review_text": "This paper claimed that SAM benefits from the over-parameterization. They find that SAM could find different (e.g. simpler) solutions than SGD in the over-parameterization regime. In addition, they find that SAM is robust to label noise and sparse parameterization. They also include some theoretical analyses, thought these analyses are irrelevant to their main findings.", "strengths": "1. The authors picked up an interesting perspective to investigate the mechanism of SAM. This research topic is significant.\n2. The experiments are extensive, across various datasets and architectures.\n3. The writing structure is clear, and the paper is easy to follow.", "weaknesses": "Firstly, for the empirical findings,\n\n1. It is unclear that whether over-parameterization consistently helps SAM find more generalizable solutions as the model parameters scale up. As shown in Fig.1 and Fig. 7 (I think the Fig. 7 is clearer), for MLP on MNIST, ResNet-18 on CIFAR-10, LSTM on SST2 and CNN on Atari, when model get more over-parameterized, the generalization improvement does not always increase; sometimes, a noticeable decline is observed. Moreover, as noted by the authors, when there is no sufficient regularization, SAM might not benefit from the over-parameterization (as shown in Fig. 5). In addition, I notice that the authors conducted experiments for ViT on CIFAR-10 in Fig.5 but not in Fig. 1. I wish the authors could include the results for ViT on CIFAR-10 in Fig.1.\n2. Increasing depth of models yields different conclusions than increasing width. In Fig. 1, authors increase the number of neurons in each layer, leading to a seemingly consistent conclusion that over-parameterization helps SAM find generalizable solutions. However, in Fig. 17, when increasing the number of layers, SAM might not benefit from the over-parameterization (ResNet on CIFAR-10). Clearly, the increase of both depth and width demonstrates the over-parameterization, however, the conclusion differs. \n3. It is not surprising that the marginal improvement of SAM over SGD increases with higher label noise. In Fig. 12, in the over-parameterization regime, the performance of SGD drops significantly with higher label noise rate; however, SAM remains good generalization ability. Indeed, it is well known that SAM is robust to label noise, while SGD not. Thus, it is not surprising for us to observe such phenomenon. Also, it seems that the authors over-claimed that over-parameterization secures the robustness of SAM to label noise. Indeed, the observed increase of margin improvement of SAM over SGD with larger model parameters is primarily due to the pronounced decline in SGD's performance as the model parameters scale up under high label noise conditions.\n4. The conclusion on the effect of sparsity contradicts with the main argument that SAM benefits from the over-parameterization. In Fig. 5, it is observed that the generalization improvement tends to increase as the model becomes sparser. However, increased sparsity indicates that the model is becoming more under-parameterized. The authors are suggested to explain such contradiction.\n5. In Fig. 2 and 3, the authors compare GD and SAM, however, a comparison between SGD and SAM is more preferred.\n6. In section 4.2, “implicit bias of SAM increase with over-parameterization” is over-stated. In this section, the authors demonstrated that we could use larger perturbation radius when scaling up the model parameters, however, this doesn’t indicate that a larger scale of model parameters induces stronger implicit bias towards flatter minima. \n\nIn short, SAM doesn’t always benefit from the over-parameterization in practice. A more thorough justification is needed to clarify whether over-parameterization is advantageous for SAM and under what specific conditions it may provide benefits. Indeed, it is quite expectable that SAM would benefit from over-parameterization sometimes. Because in the under-parameterization regime, the model capacity is low (say only one solution in the solution space), thus both SGD and SAM achieve similar solutions. Once scaling up the parameters, the solution space is getting larger, thereby SAM could differ from SGD and find more generalizable solutions. However, SAM might not always from the over-parameterization. \n\nSecond, for the theoretical analysis,\n\n1. As noted by the authors, the theoretical analysis cannot support the findings in Section 3 and 4, and thus the significance of the theoretical analysis is limited. The section 6 is more like a collection of possible theoretical analyses that could be done for SAM. \n2. Indeed, there is no clear theoretical support for the main discovery of this paper, as noted by the authors in the limitations part in Section 7. This is another critical issue of this paper.\n\nOverall, the major issue of this paper is the novelty. The empirical findings are largely anticipated, and the theoretical analysis closely follows prior works without introducing any new proof techniques.  Also, some statements are clearly over-claimed. The contribution of this paper is marginal, and thus I lean to rejection.\n\n\n----------------\nAs I cannot respond in official comments further, I put my responses here:\n\nBased on your clarification on the ViT experiment, I decide to increase my score from 5 to 6. However, I strongly recommend the authors to revise the paper from the following aspects:\n\n- **Overstated claims**: Some statements, like \"over-parameterization consistently improves the generalization benefit of SAM\" in the introduction part, are over-stated, which might be misleading for the readers. Some failure experiments such as ViT, increasing depth of networks, are encouraged to put earlier in the main text, along with throughout discussions. *I firmly believe that adhering to strictness is not a bad thing for a scientific paper*.\n- **Experiments on Absolute validation metrics**: Experiments using the absolute validation metrics are encouraged to put into main text along with the experiments using \"Acc improvements\".\n- **Current theory cannot directly support your main finding**: Still, currently, your theory cannot directly support your finding. I am not diminishing your contribution, but clearly those complementary theories in current paper cannot persuade me to highly evaluate your paper. I wish the authors could consider my suggestions and improve the paper with more related theoretical analyses. \n\nBTW, I decide to keep my confidence score low.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper claimed that SAM benefits from the over-parameterization. They find that SAM could find different (e.g. simpler) solutions than SGD in the over-parameterization regime. In addition, they find that SAM is robust to label noise and sparse parameterization. They also include some theoretical analyses, thought these analyses are irrelevant to their main findings.", "soundness": 2, "presentation": 3, "contribution": 1, "strengths": "1. The authors picked up an interesting perspective to investigate the mechanism of SAM. This research topic is significant.\n2. The experiments are extensive, across various datasets and architectures.\n3. The writing structure is clear, and the paper is easy to follow.", "weaknesses": "Firstly, for the empirical findings,\n\n1. It is unclear that whether over-parameterization consistently helps SAM find more generalizable solutions as the model parameters scale up. As shown in Fig.1 and Fig. 7 (I think the Fig. 7 is clearer), for MLP on MNIST, ResNet-18 on CIFAR-10, LSTM on SST2 and CNN on Atari, when model get more over-parameterized, the generalization improvement does not always increase; sometimes, a noticeable decline is observed. Moreover, as noted by the authors, when there is no sufficient regularization, SAM might not benefit from the over-parameterization (as shown in Fig. 5). In addition, I notice that the authors conducted experiments for ViT on CIFAR-10 in Fig.5 but not in Fig. 1. I wish the authors could include the results for ViT on CIFAR-10 in Fig.1.\n2. Increasing depth of models yields different conclusions than increasing width. In Fig. 1, authors increase the number of neurons in each layer, leading to a seemingly consistent conclusion that over-parameterization helps SAM find generalizable solutions. However, in Fig. 17, when increasing the number of layers, SAM might not benefit from the over-parameterization (ResNet on CIFAR-10). Clearly, the increase of both depth and width demonstrates the over-parameterization, however, the conclusion differs. \n3. It is not surprising that the marginal improvement of SAM over SGD increases with higher label noise. In Fig. 12, in the over-parameterization regime, the performance of SGD drops significantly with higher label noise rate; however, SAM remains good generalization ability. Indeed, it is well known that SAM is robust to label noise, while SGD not. Thus, it is not surprising for us to observe such phenomenon. Also, it seems that the authors over-claimed that over-parameterization secures the robustness of SAM to label noise. Indeed, the observed increase of margin improvement of SAM over SGD with larger model parameters is primarily due to the pronounced decline in SGD's performance as the model parameters scale up under high label noise conditions.\n4. The conclusion on the effect of sparsity contradicts with the main argument that SAM benefits from the over-parameterization. In Fig. 5, it is observed that the generalization improvement tends to increase as the model becomes sparser. However, increased sparsity indicates that the model is becoming more under-parameterized. The authors are suggested to explain such contradiction.\n5. In Fig. 2 and 3, the authors compare GD and SAM, however, a comparison between SGD and SAM is more preferred.\n6. In section 4.2, “implicit bias of SAM increase with over-parameterization” is over-stated. In this section, the authors demonstrated that we could use larger perturbation radius when scaling up the model parameters, however, this doesn’t indicate that a larger scale of model parameters induces stronger implicit bias towards flatter minima. \n\nIn short, SAM doesn’t always benefit from the over-parameterization in practice. A more thorough justification is needed to clarify whether over-parameterization is advantageous for SAM and under what specific conditions it may provide benefits. Indeed, it is quite expectable that SAM would benefit from over-parameterization sometimes. Because in the under-parameterization regime, the model capacity is low (say only one solution in the solution space), thus both SGD and SAM achieve similar solutions. Once scaling up the parameters, the solution space is getting larger, thereby SAM could differ from SGD and find more generalizable solutions. However, SAM might not always from the over-parameterization. \n\nSecond, for the theoretical analysis,\n\n1. As noted by the authors, the theoretical analysis cannot support the findings in Section 3 and 4, and thus the significance of the theoretical analysis is limited. The section 6 is more like a collection of possible theoretical analyses that could be done for SAM. \n2. Indeed, there is no clear theoretical support for the main discovery of this paper, as noted by the authors in the limitations part in Section 7. This is another critical issue of this paper.\n\nOverall, the major issue of this paper is the novelty. The empirical findings are largely anticipated, and the theoretical analysis closely follows prior works without introducing any new proof techniques.  Also, some statements are clearly over-claimed. The contribution of this paper is marginal, and thus I lean to rejection.\n\n\n----------------\nAs I cannot respond in official comments further, I put my responses here:\n\nBased on your clarification on the ViT experiment, I decide to increase my score from 5 to 6. However, I strongly recommend the authors to revise the paper from the following aspects:\n\n- **Overstated claims**: Some statements, like \"over-parameterization consistently improves the generalization benefit of SAM\" in the introduction part, are over-stated, which might be misleading for the readers. Some failure experiments such as ViT, increasing depth of networks, are encouraged to put earlier in the main text, along with throughout discussions. *I firmly believe that adhering to strictness is not a bad thing for a scientific paper*.\n- **Experiments on Absolute validation metrics**: Experiments using the absolute validation metrics are encouraged to put into main text along with the experiments using \"Acc improvements\".\n- **Current theory cannot directly support your main finding**: Still, currently, your theory cannot directly support your finding. I am not diminishing your contribution, but clearly those complementary theories in current paper cannot persuade me to highly evaluate your paper. I wish the authors could consider my suggestions and improve the paper with more related theoretical analyses. \n\nBTW, I decide to keep my confidence score low.", "questions": "N/A", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730362582135}, {"id": "m3SradlD9j", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2569/Reviewer_L6ky"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 3, "summary": "This work attempts to disclose the critical influence of overparamatrisation on  Sharpness-Aware Minimization  (SAM) from experimental and theoretical perspectives. It starts with an extensive evaluation to display a  highly consistent view showing the generalization benefit of SAM increases with overparametrisation. Without such overparametrisation  SAM may not work. This leads authors to consider the benefit in terms of more solutions and implicit bias. Finally, they developed theoretical advantages and advantages of overparametrization for SAM on linear stability, convergence and generalization.", "review_text": "This work attempts to disclose the critical influence of overparamatrisation on  Sharpness-Aware Minimization  (SAM) from experimental and theoretical perspectives. It starts with an extensive evaluation to display a  highly consistent view showing the generalization benefit of SAM increases with overparametrisation. Without such overparametrisation  SAM may not work. This leads authors to consider the benefit in terms of more solutions and implicit bias. Finally, they developed theoretical advantages and advantages of overparametrization for SAM on linear stability, convergence and generalization.", "strengths": "Nice empirical and theoretical work on overparametrisation on sharpness-aware minimization. It is interesting and useful for ICLR communauty", "weaknesses": "See questions below.", "questions": "(1) Please clarify the large curve surrounding the \"real\" curve in Fig . 1 (and possibly in other Figs). Specify if the shaded regions represent confidence intervals, standard deviations or soe other measures of uncertainty. \n(2) How can SAM improvements be compared to SGD improvements via overparametrisation? Hoa SAM and SGD performance changes with increasing model size (may be you can add provide (in Annex) plots showing relative improvement over a baseline for both optimizers across different nodel sizes)\n(3) Min Max optimization has some limitations (High complexity; Oscillatory behavior, sensitivity to initialization, etc.), how these problems are addressed here ? Can equation (3) (or some of its properties) help mitigate these issues ? Did you observed these specific issues in your experiments and if so, how have you addressed them ?\n(4) Can we use Elliptic Loss Regularisation for SAM instead of Min Max? Please comment on whether Elliptic Loss Regularization could be a viable alternative to the min-max formulation in SAM, and what potential advantages or disadvantages it might have in this context.\n(5) Is there a sort of \"optimal\" number of parameters for which the computation price is acceptable ?Did you  observed any diminishing returns in performance improvement as model size increased. Please provide guidance on balancing computational cost with performance gains when using SAM in practice", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work attempts to disclose the critical influence of overparamatrisation on  Sharpness-Aware Minimization  (SAM) from experimental and theoretical perspectives. It starts with an extensive evaluation to display a  highly consistent view showing the generalization benefit of SAM increases with overparametrisation. Without such overparametrisation  SAM may not work. This leads authors to consider the benefit in terms of more solutions and implicit bias. Finally, they developed theoretical advantages and advantages of overparametrization for SAM on linear stability, convergence and generalization.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "Nice empirical and theoretical work on overparametrisation on sharpness-aware minimization. It is interesting and useful for ICLR communauty", "weaknesses": "See questions below.", "questions": "(1) Please clarify the large curve surrounding the \"real\" curve in Fig . 1 (and possibly in other Figs). Specify if the shaded regions represent confidence intervals, standard deviations or soe other measures of uncertainty. \n(2) How can SAM improvements be compared to SGD improvements via overparametrisation? Hoa SAM and SGD performance changes with increasing model size (may be you can add provide (in Annex) plots showing relative improvement over a baseline for both optimizers across different nodel sizes)\n(3) Min Max optimization has some limitations (High complexity; Oscillatory behavior, sensitivity to initialization, etc.), how these problems are addressed here ? Can equation (3) (or some of its properties) help mitigate these issues ? Did you observed these specific issues in your experiments and if so, how have you addressed them ?\n(4) Can we use Elliptic Loss Regularisation for SAM instead of Min Max? Please comment on whether Elliptic Loss Regularization could be a viable alternative to the min-max formulation in SAM, and what potential advantages or disadvantages it might have in this context.\n(5) Is there a sort of \"optimal\" number of parameters for which the computation price is acceptable ?Did you  observed any diminishing returns in performance improvement as model size increased. Please provide guidance on balancing computational cost with performance gains when using SAM in practice", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730234226204}], "openreview_url": "https://openreview.net/forum?id=YMCtQlm8Bc", "arxiv_id": "2311.17539", "paper_pdf": "papers/YMCtQlm8Bc.pdf", "paper_pdf_sha256": "e482abfbc585f1a91c990f9065e43a7f385a070f68a417f6326b84ab0f239203", "paper_pdf_bytes": 1913746, "paper_pdf_source": "openreview", "code_url": "https://github.com/LOG-postech/SAM-overparam", "code_repository": "LOG-postech/SAM-overparam", "code_commit": "a96543a160555481d5db3513daf67370e9062b5e", "code_archive": "repos/YMCtQlm8Bc.zip", "code_archive_sha256": "af46474301cffa44ed514fbb09de6fe0eeb25bfa03baedb7ad7c2557014c3a49", "code_archive_bytes": 292216, "code_file_count": 9, "code_extensions": {".py": 9}, "github_disk_usage_kb": 435, "github_languages": {"Python": 47149}, "github_archived": false, "github_pushed_at": "2025-05-14T07:51:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/the-effects-of-overparameterization-on"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "9Kgnvknvwd", "year": 2024, "status": "rejected", "title": "A First-Order Multi-Gradient Algorithm for Multi-Objective Bi-Level Optimization", "authors": ["Feiyang Ye", "Baijiong Lin", "Xiaofeng Cao", "Yu Zhang", "Ivor Tsang"], "authorids": ["~Feiyang_Ye4", "~Baijiong_Lin1", "~Xiaofeng_Cao2", "~Yu_Zhang3", "~Ivor_Tsang1"], "authors_source": "OpenReview API", "abstract": "In this paper, we study the Multi-Objective Bi-Level Optimization (MOBLO) problem, where the upper-level subproblem is a multi-objective optimization problem and the lower-level subproblem is for scalar optimization. Existing gradient-based MOBLO algorithms need to compute the Hessian matrix, causing the computational inefficient problem. To address this, we propose an efficient first-order multi-gradient method for MOBLO, called FORUM. Specifically, we reformulate MOBLO problems as a constrained multi-objective optimization (MOO) problem via the value-function approach. Then we propose a novel multi-gradient aggregation method to solve the challenging constrained MOO problem. Theoretically, we provide the complexity analysis to show the efficiency of the proposed method and a non-asymptotic convergence result. Empirically, extensive experiments demonstrate the effectiveness and efficiency of the proposed FORUM method in different learning problems. In particular, it achieves state-of-the-art performance on three multi-task learning benchmark datasets.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "p76f00rtwy", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7105/Reviewer_MMwK"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a novel first-order multi-gradient algorithm called FORUM for solving multi-objective bi-level optimization problems. The proposed method achieves state-of-the-art performance on three multi-task learning benchmark datasets. The paper also provides a reformulation of the MOBLO problem as a constrained multi-objective optimization problem using the value-function-based approach. The proposed method is evaluated through empirical experiments, which demonstrate its effectiveness and efficiency.", "review_text": "This paper proposes a novel first-order multi-gradient algorithm called FORUM for solving multi-objective bi-level optimization problems. The proposed method achieves state-of-the-art performance on three multi-task learning benchmark datasets. The paper also provides a reformulation of the MOBLO problem as a constrained multi-objective optimization problem using the value-function-based approach. The proposed method is evaluated through empirical experiments, which demonstrate its effectiveness and efficiency.", "strengths": "1. Novelty: The paper proposes a new method, FORUM, for solving multi-objective bi-level optimization problems that is based on a first-order multi-gradient algorithm. This is a novel approach that addresses the computational inefficiency of existing gradient-based methods that require computing the Hessian matrix.\n\n2. Efficiency: The proposed FORUM algorithm is shown to be more efficient than existing methods based on complexity analysis. The paper provides a theoretical analysis of the algorithm's complexity and a non-asymptotic convergence result. Empirical experiments also demonstrate the efficiency of the proposed method in different learning problems.\n\n3. Effectiveness: The proposed FORUM algorithm achieves state-of-the-art performance on three multi-task learning benchmark datasets. The paper provides extensive experimental results that demonstrate the effectiveness of the proposed method in comparison to other state-of-the-art algorithms.", "weaknesses": "1. Limited scope: The paper only evaluates the proposed FORUM algorithm on two learning problems, i.e., multi-objective data hyper-cleaning and multi-task learning on three benchmark datasets. The generalizability of the proposed method to other learning problems is not thoroughly explored.\n\n2. Lack of comparison with non-gradient-based methods: The paper only compares the proposed FORUM algorithm with existing gradient-based methods, such as MOML and MoCo. It would be interesting to see how the proposed method compares to non-gradient-based methods, such as evolutionary algorithms or swarm intelligence.\n\n3. Lack of implementation details: The paper does not provide detailed implementation information about the proposed FORUM algorithm, such as the specific hyperparameters used in the experiments. This makes it difficult for other researchers to reproduce the results and compare the proposed method with their own algorithms.", "questions": "Regarding the use of approximation methods in the paper, I have a question for the authors. While the paper proposes an approximation method to compute ω∗(α) and approximates the constraint function q(z) using eq(z) = f(z)−f(α, ˜ωT ), it is not clear how the approximation errors affect the performance of the proposed FORUM algorithm. Could you please provide more insights into the impact of the approximation errors on the convergence and efficiency of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel first-order multi-gradient algorithm called FORUM for solving multi-objective bi-level optimization problems. The proposed method achieves state-of-the-art performance on three multi-task learning benchmark datasets. The paper also provides a reformulation of the MOBLO problem as a constrained multi-objective optimization problem using the value-function-based approach. The proposed method is evaluated through empirical experiments, which demonstrate its effectiveness and efficiency.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. Novelty: The paper proposes a new method, FORUM, for solving multi-objective bi-level optimization problems that is based on a first-order multi-gradient algorithm. This is a novel approach that addresses the computational inefficiency of existing gradient-based methods that require computing the Hessian matrix.\n\n2. Efficiency: The proposed FORUM algorithm is shown to be more efficient than existing methods based on complexity analysis. The paper provides a theoretical analysis of the algorithm's complexity and a non-asymptotic convergence result. Empirical experiments also demonstrate the efficiency of the proposed method in different learning problems.\n\n3. Effectiveness: The proposed FORUM algorithm achieves state-of-the-art performance on three multi-task learning benchmark datasets. The paper provides extensive experimental results that demonstrate the effectiveness of the proposed method in comparison to other state-of-the-art algorithms.", "weaknesses": "1. Limited scope: The paper only evaluates the proposed FORUM algorithm on two learning problems, i.e., multi-objective data hyper-cleaning and multi-task learning on three benchmark datasets. The generalizability of the proposed method to other learning problems is not thoroughly explored.\n\n2. Lack of comparison with non-gradient-based methods: The paper only compares the proposed FORUM algorithm with existing gradient-based methods, such as MOML and MoCo. It would be interesting to see how the proposed method compares to non-gradient-based methods, such as evolutionary algorithms or swarm intelligence.\n\n3. Lack of implementation details: The paper does not provide detailed implementation information about the proposed FORUM algorithm, such as the specific hyperparameters used in the experiments. This makes it difficult for other researchers to reproduce the results and compare the proposed method with their own algorithms.", "questions": "Regarding the use of approximation methods in the paper, I have a question for the authors. While the paper proposes an approximation method to compute ω∗(α) and approximates the constraint function q(z) using eq(z) = f(z)−f(α, ˜ωT ), it is not clear how the approximation errors affect the performance of the proposed FORUM algorithm. Could you please provide more insights into the impact of the approximation errors on the convergence and efficiency of the proposed method?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699193955545}, {"id": "2qelUaoyF7", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7105/Reviewer_xYWk"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This work studied a multi-objective optimization problem where each objective has a bilevel optimization structure. Each upper-level objective is evaluated at the minimum of the same lower-level problem. The authors utilize the idea from value-function-based method to solving the bilevel problem as well as a momentum update idea in Zhou et al., 2022 in the multi-objective algorithm. A convergence analysis is provided for the derived algorithm. Several experiments on hyper-cleaning, multi-task learning over Office-31, NYUv2, QM9 datasets are provided.", "review_text": "This work studied a multi-objective optimization problem where each objective has a bilevel optimization structure. Each upper-level objective is evaluated at the minimum of the same lower-level problem. The authors utilize the idea from value-function-based method to solving the bilevel problem as well as a momentum update idea in Zhou et al., 2022 in the multi-objective algorithm. A convergence analysis is provided for the derived algorithm. Several experiments on hyper-cleaning, multi-task learning over Office-31, NYUv2, QM9 datasets are provided.", "strengths": "1.\tOverall, the studied problem is timely. Bilevel optimization and multi-objective optimization have wide applications in practice. This work uses a first-order idea from bilevel literature in MOO, and seems to work well in experiments.\n\n2.\tExperiments seem to support that the proposed method can work well in some datasets.", "weaknesses": "1.\tHowever, I have quite a few concerns regarding the analysis and the novelty. First, to deal with the reformulated constraint in the value-function based method, the authors use a dot product between $d$ and the gradient of the constraint and make it less or equal to $-\\phi_k$. However, how to pick this $\\phi_k$ in practice and in theory? Also, the final convergence criterion seems questionable. For example, the authors use the measure of KKT stationary condition to as convergence criterion. How does this condition correlate with Parato stationarity? How fast does the parameter $v_k$ decrease to 0 in this criterion? All these questions are not well explained in this work. \n2.\tThe algorithm has some unclear parts. For example, the authors use the idea of momentum update on $\\lambda$ update. This step is originally proposed in Zhou et al., 2022 (the authors should mention about this). However, there is characterization on the distance between the true variable $\\lambda_k$ and the surrogate $\\titilde \\lambda_k$? This is important, because what you need to use is $\\lambda_k$ rather than surrogate $\\titilde \\lambda_k$ in the algorithm. \n3.\tThe algorithm is deterministic without data sampling, but in the experiments it seems data sampling is used. I am wondering if it is possible to extend the algorithm and analysis to the more practice stochastic setting? If not, what are the challenges? Some recent progresses on stochastic MOO may be helpful here (some of them are missing in this work). \n\n[1] Suyun Liu and Luis Nunes Vicente. The stochastic multi-gradient algorithm for multi-objective optimization and its application to supervised machine learning. \n\n[2] Lisha Chen, Heshan Fernando, Yiming Ying, and Tianyi Chen. Three-way trade-off in multi-objective learning: Optimization, generalization and conflict-avoidance.\n\n[3] Heshan Devaka Fernando, Han Shen, Miao Liu, Subhajit Chaudhury, Keerthiram Murugesan, and Tianyi Chen. Mitigating gradient bias in multi-objective learning: A provably convergent approach.\n\n[4] Peiyao Xiao, Hao Ban, and Kaiyi Ji. Direction-oriented Multi-objective Learning: Simple and Provable Stochastic Algorithms.\n\n4.\tThe analysis assumes the upper-level function $F_i$ is bounded. However, this has not been made in the aforementioned [1,2,3,4] works. More clarifications should be provided. If there are any special challenges making this assumption necessary?  Also, the assumption $q(z_k)<B$ is a strong assumption that has not been made in previous bilevel and MOO literatures, because bounded function value and strong-convexity cannot be made simultaneously.", "questions": "Overall, this work studied an interesting and important problem. However, it has quite a few questions and problems to be solved. However, I am open to increase my score given the authors’ response. See my questions in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This work studied a multi-objective optimization problem where each objective has a bilevel optimization structure. Each upper-level objective is evaluated at the minimum of the same lower-level problem. The authors utilize the idea from value-function-based method to solving the bilevel problem as well as a momentum update idea in Zhou et al., 2022 in the multi-objective algorithm. A convergence analysis is provided for the derived algorithm. Several experiments on hyper-cleaning, multi-task learning over Office-31, NYUv2, QM9 datasets are provided.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1.\tOverall, the studied problem is timely. Bilevel optimization and multi-objective optimization have wide applications in practice. This work uses a first-order idea from bilevel literature in MOO, and seems to work well in experiments.\n\n2.\tExperiments seem to support that the proposed method can work well in some datasets.", "weaknesses": "1.\tHowever, I have quite a few concerns regarding the analysis and the novelty. First, to deal with the reformulated constraint in the value-function based method, the authors use a dot product between $d$ and the gradient of the constraint and make it less or equal to $-\\phi_k$. However, how to pick this $\\phi_k$ in practice and in theory? Also, the final convergence criterion seems questionable. For example, the authors use the measure of KKT stationary condition to as convergence criterion. How does this condition correlate with Parato stationarity? How fast does the parameter $v_k$ decrease to 0 in this criterion? All these questions are not well explained in this work. \n2.\tThe algorithm has some unclear parts. For example, the authors use the idea of momentum update on $\\lambda$ update. This step is originally proposed in Zhou et al., 2022 (the authors should mention about this). However, there is characterization on the distance between the true variable $\\lambda_k$ and the surrogate $\\titilde \\lambda_k$? This is important, because what you need to use is $\\lambda_k$ rather than surrogate $\\titilde \\lambda_k$ in the algorithm. \n3.\tThe algorithm is deterministic without data sampling, but in the experiments it seems data sampling is used. I am wondering if it is possible to extend the algorithm and analysis to the more practice stochastic setting? If not, what are the challenges? Some recent progresses on stochastic MOO may be helpful here (some of them are missing in this work). \n\n[1] Suyun Liu and Luis Nunes Vicente. The stochastic multi-gradient algorithm for multi-objective optimization and its application to supervised machine learning. \n\n[2] Lisha Chen, Heshan Fernando, Yiming Ying, and Tianyi Chen. Three-way trade-off in multi-objective learning: Optimization, generalization and conflict-avoidance.\n\n[3] Heshan Devaka Fernando, Han Shen, Miao Liu, Subhajit Chaudhury, Keerthiram Murugesan, and Tianyi Chen. Mitigating gradient bias in multi-objective learning: A provably convergent approach.\n\n[4] Peiyao Xiao, Hao Ban, and Kaiyi Ji. Direction-oriented Multi-objective Learning: Simple and Provable Stochastic Algorithms.\n\n4.\tThe analysis assumes the upper-level function $F_i$ is bounded. However, this has not been made in the aforementioned [1,2,3,4] works. More clarifications should be provided. If there are any special challenges making this assumption necessary?  Also, the assumption $q(z_k)<B$ is a strong assumption that has not been made in previous bilevel and MOO literatures, because bounded function value and strong-convexity cannot be made simultaneously.", "questions": "Overall, this work studied an interesting and important problem. However, it has quite a few questions and problems to be solved. However, I am open to increase my score given the authors’ response. See my questions in the weakness part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698691946158}, {"id": "KGSfRlMZZo", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7105/Reviewer_6mPv"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a first-order multi-gradient method tailored for the multi-objective bi-level optimization (MOBLO) problem. More precisely, the authors first reformulate the MOBLO problem as an equivalent single-level constrained multi-objective optimization problem using a value-function-based approach. Subsequently, they integrate the BOME method, designed for single-objective bi-level optimization, with the MGDA for multi-objective optimization (MOO) to address this equivalent problem. The authors provide convergence analysis and numerical results.", "review_text": "This paper introduces a first-order multi-gradient method tailored for the multi-objective bi-level optimization (MOBLO) problem. More precisely, the authors first reformulate the MOBLO problem as an equivalent single-level constrained multi-objective optimization problem using a value-function-based approach. Subsequently, they integrate the BOME method, designed for single-objective bi-level optimization, with the MGDA for multi-objective optimization (MOO) to address this equivalent problem. The authors provide convergence analysis and numerical results.", "strengths": "The topic of multi-objective bi-level optimization is important. \n\nConvergence analysis is provided for the proposed method under some assumptions, e.g., the lower level problem is strongly convex.\n\nNumerical validation is presented for the proposed method.", "weaknesses": "1. The proposed algorithm is a straightforward combination of two existing methods, and the accompanying analysis appears rather standard. Consequently, the technical innovation compared to prior work upon which this study is based is limited.\n\n2. The convergence result of the proposed method lacks persuasiveness. As pointed out by the authors, the constraint $q(z) \\le 0$ in the reformulated problem (3) is ill-posed, rendering the KKT stationary condition not a necessary condition for problem (3) solutions. Therefore, the utilization of the $\\mathcal{K}(z)$ measure for the convergence of the proposed method in the main convergence theorem (Theorem 4.3) is inappropriate. In contrast, the MoCo method by Fernando et al. (2023) employs hyper-gradients under the same strong convexity in the LL problem to characterize convergence and establish convergence to Pareto stationarity.\n\n3. The analysis of the proposed algorithm is confined to a deterministic setting, which may restrict its applicability given that the motivating applications are mostly in the stochastic setting.\n\n4. The strong convexity of the LL problem appears to play a pivotal role in the convergence analysis, which, in turn, constrains the applicability of the proposed method. Notably, the MOML method introduced by Ye et al. (2022) does not necessitate such an assumption.", "questions": "My questions are listed in the “Weaknesses” part.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a first-order multi-gradient method tailored for the multi-objective bi-level optimization (MOBLO) problem. More precisely, the authors first reformulate the MOBLO problem as an equivalent single-level constrained multi-objective optimization problem using a value-function-based approach. Subsequently, they integrate the BOME method, designed for single-objective bi-level optimization, with the MGDA for multi-objective optimization (MOO) to address this equivalent problem. The authors provide convergence analysis and numerical results.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "The topic of multi-objective bi-level optimization is important. \n\nConvergence analysis is provided for the proposed method under some assumptions, e.g., the lower level problem is strongly convex.\n\nNumerical validation is presented for the proposed method.", "weaknesses": "1. The proposed algorithm is a straightforward combination of two existing methods, and the accompanying analysis appears rather standard. Consequently, the technical innovation compared to prior work upon which this study is based is limited.\n\n2. The convergence result of the proposed method lacks persuasiveness. As pointed out by the authors, the constraint $q(z) \\le 0$ in the reformulated problem (3) is ill-posed, rendering the KKT stationary condition not a necessary condition for problem (3) solutions. Therefore, the utilization of the $\\mathcal{K}(z)$ measure for the convergence of the proposed method in the main convergence theorem (Theorem 4.3) is inappropriate. In contrast, the MoCo method by Fernando et al. (2023) employs hyper-gradients under the same strong convexity in the LL problem to characterize convergence and establish convergence to Pareto stationarity.\n\n3. The analysis of the proposed algorithm is confined to a deterministic setting, which may restrict its applicability given that the motivating applications are mostly in the stochastic setting.\n\n4. The strong convexity of the LL problem appears to play a pivotal role in the convergence analysis, which, in turn, constrains the applicability of the proposed method. Notably, the MOML method introduced by Ye et al. (2022) does not necessitate such an assumption.", "questions": "My questions are listed in the “Weaknesses” part.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698223679943}, {"id": "a5C2eG2WMB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission7105/Reviewer_nzzn"], "rating": "6: marginally above the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "In this work, the Multi-Objective Bi-Level Optimization (MOBLO) is studied, and an efficient first-order multi-gradient method for MOBLO, called FORUM, is proposed. The proposed method first reformulates MOBLO as an equivalent constrained multi-objective problem, then a novel multi-gradient aggregation method to solve the constrained multi-objective problem.", "review_text": "In this work, the Multi-Objective Bi-Level Optimization (MOBLO) is studied, and an efficient first-order multi-gradient method for MOBLO, called FORUM, is proposed. The proposed method first reformulates MOBLO as an equivalent constrained multi-objective problem, then a novel multi-gradient aggregation method to solve the constrained multi-objective problem.", "strengths": "1. The proposed method combines the value-function-based approach and multi-gradient method, which is novel in the multiobjective bilevel optimization problems.\n\n2. The writing of this work is good, and the logic of the proposed method is clear.", "weaknesses": "I believe this work is solid and good, however, I have some concerns as follows.\n\n1. In the proposed method, an additional optimization problem is required to solve every iteration, i.e., Eq. (11). Thus the proposed method seems inefficient since it is a nested-loop algorithm.\n\n2. I suggest the authors add a table to compare the differences between the proposed methods and existing MOBLO methods (i.e., MOML and MoCo) to clearly show the advantages of the proposed method.", "questions": "See Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this work, the Multi-Objective Bi-Level Optimization (MOBLO) is studied, and an efficient first-order multi-gradient method for MOBLO, called FORUM, is proposed. The proposed method first reformulates MOBLO as an equivalent constrained multi-objective problem, then a novel multi-gradient aggregation method to solve the constrained multi-objective problem.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The proposed method combines the value-function-based approach and multi-gradient method, which is novel in the multiobjective bilevel optimization problems.\n\n2. The writing of this work is good, and the logic of the proposed method is clear.", "weaknesses": "I believe this work is solid and good, however, I have some concerns as follows.\n\n1. In the proposed method, an additional optimization problem is required to solve every iteration, i.e., Eq. (11). Thus the proposed method seems inefficient since it is a nested-loop algorithm.\n\n2. I suggest the authors add a table to compare the differences between the proposed methods and existing MOBLO methods (i.e., MOML and MoCo) to clearly show the advantages of the proposed method.", "questions": "See Weaknesses above.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698200336298}], "openreview_url": "https://openreview.net/forum?id=9Kgnvknvwd", "arxiv_id": "2401.09257", "paper_pdf": "papers/9Kgnvknvwd.pdf", "paper_pdf_sha256": "293c575bdb29f63088782569c988ae4dfda05b2a30a2cf3a7f3863516e93ad44", "paper_pdf_bytes": 506917, "paper_pdf_source": "openreview", "code_url": "https://github.com/Baijiong-Lin/FORUM", "code_repository": "Baijiong-Lin/FORUM", "code_commit": "318607c503ec4ec5ce4f0d297401fc43717b026f", "code_archive": "repos/9Kgnvknvwd.zip", "code_archive_sha256": "f0243dac00e0edb628c5374d155a54ab6b163b3c2b7996967ee4d863f8f6f408", "code_archive_bytes": 522142, "code_file_count": 46, "code_extensions": {".py": 46}, "github_disk_usage_kb": 511, "github_languages": {"Python": 145838}, "github_archived": false, "github_pushed_at": "2024-07-04T08:05:20Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-first-order-multi-gradient-algorithm-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "mPzpPv0geS2", "year": 2023, "status": "rejected", "title": "Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models", "authors": ["Xingyu Xie", "Pan Zhou", "Huan Li", "Zhouchen Lin", "Shuicheng YAN"], "authorids": ["~Xingyu_Xie1", "~Pan_Zhou3", "~Huan_Li1", "~Zhouchen_Lin1", "~Shuicheng_YAN3"], "authors_source": "OpenReview API", "abstract": "Adaptive gradient algorithms combine the moving average idea with heavy ball acceleration to estimate accurate first- and second-order moments of the gradient for accelerating convergence.  However, Nesterov acceleration which converges faster than heavy ball acceleration in theory and also in many empirical cases, is much less investigated under the adaptive gradient setting.  In this work, we propose the ADAptive Nesterov momentum algorithm, Adan for short, to speed up the training of deep neural networks effectively.  Adan first reformulates the vanilla Nesterov acceleration to develop a new Nesterov momentum estimation (NME) method, which avoids the extra computation and memory overhead of computing gradient at the extrapolation point.  Then Adan adopts NME to estimate the first- and second-order moments of the gradient in adaptive gradient algorithms for convergence acceleration. Besides, we prove that Adan finds an $\\epsilon$-approximate first-order stationary point within $O(\\epsilon^{-3.5})$ stochastic gradient complexity on the non-convex stochastic problems (e.g., deep learning problems), matching the best-known lower bound. Extensive experimental results show that  Adan surpasses the corresponding SoTA optimizers on vision, language, and RL tasks and sets new SoTAs for many popular networks and frameworks, e.g., ResNet,  ConvNext, ViT, Swin, MAE, LSTM, Transformer-XL, and BERT.  More surprisingly, Adan can use half of the training cost (epochs) of SoTA optimizers to achieve higher or comparable performance on ViT, ResNet, MAE, etc., and also shows great tolerance to a large range of minibatch size, e.g., from 1k to 32k.  We hope Adan can contribute to developing deep learning by reducing training costs and relieving the engineering burden of trying different optimizers on various architectures.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "eHheXwe1cJ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1065/Reviewer_wTd4"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The paper introduce a new adaptive method that built upon ADAM and Nesterov momentum. First, it reformulate Nesterov momentum update and combine it with ADAM; Second, it provides $O(\\epsilon^{-4})$ complexity for Lipschitz-smooth case; Third, it provides $O(\\epsilon^{-3.5})$ complexity for Hessian Lipschitz case. ", "review_text": "Overall, the algorithm looks promising, but does not match the lower bounds in theory and I am not fully convinced by the experiments. ", "strengths": "### Strength\n1. The idea is clear and straightforward\n2. The experiments are conducted for multiple tasks\n\n### Weaknesses\n1. The claim that $O(\\epsilon^{-3.5})$ is the lower bound is not correct. In Assumption 1, it assumes the function is smooth with respect to each value of $\\zeta$, which will already lead to a lower bound of $O(\\epsilon^{-3})$ even without Lipschitz Hessian. I also found that Assumption 2 is stronger than usual, as it assume $\\mathbb{E}||\\xi|| \\leq \\sigma$ rather than $\\mathbb{E}||\\xi||^2 \\leq \\sigma^2$. So I think lower bound could be much better than $O(\\epsilon^{-4})$ and $O(\\epsilon^{-3.5})$ for each case. \n2. From the change in equation (4), the algorithm introduce a new hyperparameter $\\beta_2$. In the new algorithm, there are three momentum hyperparameters to pick: $\\beta_1, \\beta_2$ and $\\beta_3$. When $\\beta_2 = 0$, the algorithm basically reduces to the vanilla Adam. In the experiment, it performs much better than Adam, so I am not sure whether it is simply a result of tuning $\\beta_2$. However, the ablation study in B.5 indicate that Adan is not sensitive to the value of $\\beta_2$, which contradicts that it is much better than Adam. \n\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper introduce a new adaptive method that built upon ADAM and Nesterov momentum. First, it reformulate Nesterov momentum update and combine it with ADAM; Second, it provides $O(\\epsilon^{-4})$ complexity for Lipschitz-smooth case; Third, it provides $O(\\epsilon^{-3.5})$ complexity for Hessian Lipschitz case. ", "strength_and_weaknesses": "### Strength\n1. The idea is clear and straightforward\n2. The experiments are conducted for multiple tasks\n\n### Weaknesses\n1. The claim that $O(\\epsilon^{-3.5})$ is the lower bound is not correct. In Assumption 1, it assumes the function is smooth with respect to each value of $\\zeta$, which will already lead to a lower bound of $O(\\epsilon^{-3})$ even without Lipschitz Hessian. I also found that Assumption 2 is stronger than usual, as it assume $\\mathbb{E}||\\xi|| \\leq \\sigma$ rather than $\\mathbb{E}||\\xi||^2 \\leq \\sigma^2$. So I think lower bound could be much better than $O(\\epsilon^{-4})$ and $O(\\epsilon^{-3.5})$ for each case. \n2. From the change in equation (4), the algorithm introduce a new hyperparameter $\\beta_2$. In the new algorithm, there are three momentum hyperparameters to pick: $\\beta_1, \\beta_2$ and $\\beta_3$. When $\\beta_2 = 0$, the algorithm basically reduces to the vanilla Adam. In the experiment, it performs much better than Adam, so I am not sure whether it is simply a result of tuning $\\beta_2$. However, the ablation study in B.5 indicate that Adan is not sensitive to the value of $\\beta_2$, which contradicts that it is much better than Adam. \n\n\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity: I think the paper is easy to read. I also suggest to add reference for Table 1, Nesterov acceleration in the last paragraph of page 2 and lower bound $\\epsilon^{-4}$ in the first paragraph of page 3. \n\nNovelty: The algorithm seems new to me and the construction is interesting. ", "summary_of_the_review": "Overall, the algorithm looks promising, but does not match the lower bounds in theory and I am not fully convinced by the experiments. ", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666806340139}, {"id": "uDPW5MIwmBL", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1065/Reviewer_5SGt"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper first reformulate the update expression of the Nesterov acceleration technique and then modify Adam accordingly to develop a new optimizer named Adan, which is different from NAdam. The main difference between Adan and Adam is that the gradient difference g_k-g_{k-1} is incorporated in computation of both the 1st and 2nd momentum. Both convergence analysis and experimental evaluation are conducted for Adan. ", "review_text": "The authors develop a new method Adan by incorporating the gradient difference g_k-g_{k-1} in Adam based on a reformulation of the Nesterov acceleration technique. (1) Explanation of why Adan has lower theoretical complexity is needed for the readers to better understand the method. (2) Paper presentation needs further improvement (see my comments above). (3) I don't find anywhere that the hyper-parameters of the reference methods in the experiments are manually tuned while the hyper-parameters of Adan are tuned. Also it is suggested to evaluate AdaBelied, Adam+, Aida in the experiments. ", "strengths": "Strengths: \n1. The new method Adan is developed by incorporating the gradient difference g_k-g_{k-1} in Adam based on a reformulation of the Nesterov acceleration technique.\n2.  Both convergence analysis and experimental evaluation are conducted for Adan, showing its advantage over other optimizers. \n\nWeaknesses:\n1.  The authors claim that  Adan has lower complexity than a number of existing optimizers including AdaBelief and LAMB without an explanation. Is it because of the separation of the l2 regularizer with the objective, or because of the introduction of g_k- g_{k-1} in Adan, or because of improved mathematical deviation. This is very important. It helps the readers to understand what is the reason of the the low complexity of the new method without reading all the proofs. If it is because of the separation of the l2 regularizer with the objective or  improved mathematical deviation, it suggests that the convergence of other optimizers can also be improved.  \n2. It seems that the authors are aware of AdaBelief and Adam+ but didn't evaluate them in the experiments. AdaBelief is empirically found to have better performance than quite a number of opimizers including AdamW and Adam for training different types of DNN models including Transformer. Recently, a new method named Aida in the paper \"On exploiting layerwise gradient statistics for effective training of deep neural networks\" is found to perform better than AdaBelief.  I highly suggest the authors to also evaluate AdamBelief, Adam+ and Aida in their experiments. Furthermore, the parameter epsilon needs to be searched for each method to achieve best performance. It seems from page 7 that the authors only tune the hyper-parameters of Adan. \n3. Compared to Adam, AdamW, AdaBelief, Adan needs to tune an additional parameter beta_2.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The paper first reformulate the update expression of the Nesterov acceleration technique and then modify Adam accordingly to develop a new optimizer named Adan, which is different from NAdam. The main difference between Adan and Adam is that the gradient difference g_k-g_{k-1} is incorporated in computation of both the 1st and 2nd momentum. Both convergence analysis and experimental evaluation are conducted for Adan. ", "strength_and_weaknesses": "Strengths: \n1. The new method Adan is developed by incorporating the gradient difference g_k-g_{k-1} in Adam based on a reformulation of the Nesterov acceleration technique.\n2.  Both convergence analysis and experimental evaluation are conducted for Adan, showing its advantage over other optimizers. \n\nWeaknesses:\n1.  The authors claim that  Adan has lower complexity than a number of existing optimizers including AdaBelief and LAMB without an explanation. Is it because of the separation of the l2 regularizer with the objective, or because of the introduction of g_k- g_{k-1} in Adan, or because of improved mathematical deviation. This is very important. It helps the readers to understand what is the reason of the the low complexity of the new method without reading all the proofs. If it is because of the separation of the l2 regularizer with the objective or  improved mathematical deviation, it suggests that the convergence of other optimizers can also be improved.  \n2. It seems that the authors are aware of AdaBelief and Adam+ but didn't evaluate them in the experiments. AdaBelief is empirically found to have better performance than quite a number of opimizers including AdamW and Adam for training different types of DNN models including Transformer. Recently, a new method named Aida in the paper \"On exploiting layerwise gradient statistics for effective training of deep neural networks\" is found to perform better than AdaBelief.  I highly suggest the authors to also evaluate AdamBelief, Adam+ and Aida in their experiments. Furthermore, the parameter epsilon needs to be searched for each method to achieve best performance. It seems from page 7 that the authors only tune the hyper-parameters of Adan. \n3. Compared to Adam, AdamW, AdaBelief, Adan needs to tune an additional parameter beta_2.", "clarity,_quality,_novelty_and_reproducibility": "(1) One page 4, the parameter epsilon appears inside the sqrt operation in both RMSProp and Adam. While they are implemented differently on Pytorch platform where epsilon is outside of the sqrt operation. My experience is that the placement of epsilon makes a difference in the validation performance for training DNN models.  Do you use default implementation of RMSProp and Adam on Pytorch in your experiment or you use the update expressions in your paper? The presentation for the algorithms need to be consistent in the whole paper. \n(2) Source code is not available for reproducibility. \n(3) There is no de-bias term in their new method Adan. Is it ignored for presentation convenience like RMSProp and Adam, or the de-bias term is removed in their implementation for better performance?  \n(4) I think the overhead of Nesterov acceleration technique is negligible compared to NME. There is no need for Nesterov acceleration technique to maintain both theta_k and theta_k'. NME is just a reformulation of Nesterov acceleration technique. I suggest the authors rewrite the contribution part on page 2. If the authors have different opinions, please verify their argument via experiments by implementing the Nesterov acceleration technique properly. \n(5) The authors made quite a few assumptions to be able to analyze the convergence of Adan. I wonder if the assumptions are the same for other optimizers. If they are not the same, then Table-1 is an unfair comparison.  For example, why 3 is introduced in ||g_k||_{infty}\\leq c_{infty}/3? Do other optimizers use the same assumption? \n(6) The statement \"update theta_1 by Line 7 in Algorithm 1\" is not correct.  \n", "summary_of_the_review": "The authors develop a new method Adan by incorporating the gradient difference g_k-g_{k-1} in Adam based on a reformulation of the Nesterov acceleration technique. (1) Explanation of why Adan has lower theoretical complexity is needed for the readers to better understand the method. (2) Paper presentation needs further improvement (see my comments above). (3) I don't find anywhere that the hyper-parameters of the reference methods in the experiments are manually tuned while the hyper-parameters of Adan are tuned. Also it is suggested to evaluate AdaBelied, Adam+, Aida in the experiments. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666692034986}, {"id": "rYhWA_jU7F", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1065/Reviewer_SpFC"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes another variant of Adam by combining the idea of Nesterov momentum with the adaptive algorithms. The authors prove that in the nonconvex optimization setting, the proposed algorithm converges faster than the other adaptive algorithms. Experimental results show that the proposed algorithm performs better than the other algorithms in multiple tasks and different settings.", "review_text": "Overall, I find the idea of introducing Nesterov momentum to adaptive optimization algorithm not novel enough, but I am impressed by the theoretical results and the experiments the authors have provided. The performance of the proposed algorithm does look significantly better than the existing ones. Therefore, I vote for acceptance.", "strengths": "Strength:\n  1. The paper is well-written and easy-to-follow\n  2. This is by far the first paper that have shown theoretical benefits over existing algorithms, as far as I am aware of, which is highly appreciated.\n  3. The experimental results are on both small-scale and large-scale datasets, which are quite strong and convincing.\n\nWeaknesses:\n  1. The idea of combining Nesterov momentum with Adam is, as far as I am concerned, a very straightforward idea. I am actually very surprised that this hasn't been put into practice in common deep learning tools such as PyTorch or Tensorflow. Therefore, I find the idea of the paper not novel enough.\n  2. The theoretical benefit looks marginal to me. Isn't the difference between all the optimizers only a multiplicative constant-level difference? When is the convergence of the algorithm better than the others (i.e., under what assumptions of $d$, $c_2$, $c_\\infty$? Can the authors elaborate more? ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes another variant of Adam by combining the idea of Nesterov momentum with the adaptive algorithms. The authors prove that in the nonconvex optimization setting, the proposed algorithm converges faster than the other adaptive algorithms. Experimental results show that the proposed algorithm performs better than the other algorithms in multiple tasks and different settings.", "strength_and_weaknesses": "Strength:\n  1. The paper is well-written and easy-to-follow\n  2. This is by far the first paper that have shown theoretical benefits over existing algorithms, as far as I am aware of, which is highly appreciated.\n  3. The experimental results are on both small-scale and large-scale datasets, which are quite strong and convincing.\n\nWeaknesses:\n  1. The idea of combining Nesterov momentum with Adam is, as far as I am concerned, a very straightforward idea. I am actually very surprised that this hasn't been put into practice in common deep learning tools such as PyTorch or Tensorflow. Therefore, I find the idea of the paper not novel enough.\n  2. The theoretical benefit looks marginal to me. Isn't the difference between all the optimizers only a multiplicative constant-level difference? When is the convergence of the algorithm better than the others (i.e., under what assumptions of $d$, $c_2$, $c_\\infty$? Can the authors elaborate more? ", "clarity,_quality,_novelty_and_reproducibility": "The paper is well-written, with some novelty and good reproducibility given the code references.", "summary_of_the_review": "Overall, I find the idea of introducing Nesterov momentum to adaptive optimization algorithm not novel enough, but I am impressed by the theoretical results and the experiments the authors have provided. The performance of the proposed algorithm does look significantly better than the existing ones. Therefore, I vote for acceptance.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666582857944}], "openreview_url": "https://openreview.net/forum?id=mPzpPv0geS2", "arxiv_id": "2208.06677", "paper_pdf": "papers/mPzpPv0geS2.pdf", "paper_pdf_sha256": "20fdc1ac034d4ac8f55c72e5c35542aee29bd3ddbb2a236785be3423996021c6", "paper_pdf_bytes": 687321, "paper_pdf_source": "openreview", "code_url": "https://github.com/sail-sg/Adan", "code_repository": "sail-sg/Adan", "code_commit": "2c65beaf400775753b155da4082c3819fea689b9", "code_archive": "repos/mPzpPv0geS2.zip", "code_archive_sha256": "566e520cdcf502f66f299901533420521442849f4928e604efeb33d877c2b9cb", "code_archive_bytes": 1374021, "code_file_count": 40, "code_extensions": {".py": 32, ".sh": 2, ".cu": 2, ".cuh": 2, ".h": 1, ".cpp": 1}, "github_disk_usage_kb": 1373, "github_languages": {"Python": 269155, "Cuda": 21780, "C++": 12838, "Shell": 2610}, "github_archived": false, "github_pushed_at": "2025-06-08T14:35:41Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/adan-adaptive-nesterov-momentum-algorithm-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "xWRX16GCugt", "year": 2022, "status": "rejected", "title": "Sequoia: A Software Framework to Unify Continual Learning Research", "authors": ["Fabrice Normandin", "Oleksiy Ostapenko", "Pau Rodriguez", "Florian Golemo", "Ryan Lindeborg", "Matthew Riemer", "Lucas Cecchi", "Timothee LESORT", "Khimya Khetarpal", "David Vazquez", "Laurent Charlin", "Irina Rish", "Massimo Caccia"], "authorids": ["~Fabrice_Normandin1", "~Oleksiy_Ostapenko1", "~Pau_Rodriguez2", "~Florian_Golemo1", "~Ryan_Lindeborg1", "~Matthew_Riemer1", "~Lucas_Cecchi1", "~Timothee_LESORT1", "~Khimya_Khetarpal1", "~David_Vazquez1", "~Laurent_Charlin1", "~Irina_Rish1", "~Massimo_Caccia1"], "authors_source": "OpenReview API", "abstract": "The field of Continual Learning (CL) seeks to develop algorithms that accumulate knowledge and skills over time through interaction with non-stationary environments. In practice, a plethora of evaluation procedures (settings) and algorithmic solutions (methods) exist, each with their own potentially disjoint set of assumptions. This variety makes measuring progress in CL difficult. We propose a taxonomy of settings, where each setting is described as a set of assumptions. A tree-shaped hierarchy emerges from this view, where more general settings become the parents of those with more restrictive assumptions. This makes it possible to use inheritance to share and reuse research, as developing a method for a given setting also makes it directly applicable onto any of its children. We instantiate this idea as a publicly available software framework called Sequoia, which features a wide variety of settings from both the Continual Supervised Learning (CSL) and Continual Reinforcement Learning (CRL) domains. Sequoia also includes a growing suite of methods which are easy to extend and customize, in addition to more specialized methods from external libraries. We hope that this new paradigm and its first implementation can help unify and accelerate research in CL. You can help us grow the tree by visiting (this GitHub URL).", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "lcYlYv9tU1S", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1907/Reviewer_SrGM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "1: strong reject", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The paper attempts to unify all CL research with a single formalism. Next, they present a software implementation of their framework. Finally, experiments are ran which demonstrate that Sequoia can be used to evaluate CL methods.", "review_text": "Pros:\n\nP1) Creating a software package which unifies RL and supervised learning approaches seems like a step in the right direction.\n\nCons:\n\nC1) CL hierarchy - I think there are missing nodes in the hierarchy. For instance, it is important to also consider problem-incremental learning where the task is the same, but the input spaces are different (e.g. the dimensionality of the input changes, not just the input distribution).\n\nC2) Related work: You mention that your framework doesn’t compete with others. It is not clear to me why someone else would not just use the other frameworks. A more in-depth discussion of the previous work is necessary. This way, it would be easier to distinguish your contributions are.\n\nC3) Experiments: The text does not give us information on which methods are implemented by the authors. It gave me the impression that the evaluated methods have all been implemented by another software package.\n\nC4) Experiments: The results are presented but not discussed.\n\nThe provided hierarchy of CL methods might contain an interesting insight, which relates RL and supervised learning approaches to CL. However, due to the lack of comparison to related work, I am not certain of this. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper attempts to unify all CL research with a single formalism. Next, they present a software implementation of their framework. Finally, experiments are ran which demonstrate that Sequoia can be used to evaluate CL methods.", "main_review": "Pros:\n\nP1) Creating a software package which unifies RL and supervised learning approaches seems like a step in the right direction.\n\nCons:\n\nC1) CL hierarchy - I think there are missing nodes in the hierarchy. For instance, it is important to also consider problem-incremental learning where the task is the same, but the input spaces are different (e.g. the dimensionality of the input changes, not just the input distribution).\n\nC2) Related work: You mention that your framework doesn’t compete with others. It is not clear to me why someone else would not just use the other frameworks. A more in-depth discussion of the previous work is necessary. This way, it would be easier to distinguish your contributions are.\n\nC3) Experiments: The text does not give us information on which methods are implemented by the authors. It gave me the impression that the evaluated methods have all been implemented by another software package.\n\nC4) Experiments: The results are presented but not discussed.\n\nThe provided hierarchy of CL methods might contain an interesting insight, which relates RL and supervised learning approaches to CL. However, due to the lack of comparison to related work, I am not certain of this. \n", "summary_of_the_review": "Overall, I don’t see a significant technological or conceptual contribution.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "1: strong reject", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635939241414}, {"id": "-QISKIbXYal", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1907/Reviewer_D4MM"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors try to establish a unified framework for different continual learning settings. They also provide a Python library, which includes different related methods. Extensive experimental results are also provided. ", "review_text": "### Strengths \n\n&nbsp;\n\n- The continual learning research tree (Figure 2) is well organized and makes a lot of sense to me.\n\n&nbsp;\n\n- I think the motivation of building a unified framework for continual learning is meaningful. Many different continual learning papers evaluate their methods in different benchmark protocols. It is difficult for the following researchers to compare these related methods. \n\n&nbsp;\n\n- Detailed explanations and experimental results are provided in this paper. The authors also provide the results on wandb. It makes obtaining the detailed results for each setting very convenient for the following researchers. \n\n&nbsp;\n\n### Weaknesses\n\n&nbsp;\n\n- Some important continual learning baselines are not included, such as iCaRL (Rebuffi et al., 2017). \n\n&nbsp;\n\n- The authors only provide experiment results on small-scale datasets, such as MNIST and CIFAR. Most continual learning papers, such as iCaRL (Rebuffi et al., 2017) and BiC (Wu et al., 2019), provide the results on large-scale datasets (e.g., ImageNet-1k). As the authors trying to establish a new benchmark protocol, it is important to provide the results on large-scale datasets.\n\n&nbsp;\n\n- This project includes plenty of different methods, but the authors don’t include detailed information (such as the licenses) about the related open-source resources they use. \n\n&nbsp;\n\n- The definition of “incremental learning” in Section 2.1 is ambiguous. I think it is better to use class-incremental learning (or domain-incremental learning) directly. The reasons are as follows. (1) “Incremental learning” is often considered as another name of “continual learning”. It is weird to use it to denote a specific setting of continual learning. (2) You include class-IL and domain-IL in “incremental learning”, but you exclude task-IL. It is not reasonable.\n\n&nbsp;\n\n- The authors include too many details about the code implementation in the paper. I think it is better to move these parts to the appendix and include more experimental results and analyses. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors try to establish a unified framework for different continual learning settings. They also provide a Python library, which includes different related methods. Extensive experimental results are also provided. ", "main_review": "### Strengths \n\n&nbsp;\n\n- The continual learning research tree (Figure 2) is well organized and makes a lot of sense to me.\n\n&nbsp;\n\n- I think the motivation of building a unified framework for continual learning is meaningful. Many different continual learning papers evaluate their methods in different benchmark protocols. It is difficult for the following researchers to compare these related methods. \n\n&nbsp;\n\n- Detailed explanations and experimental results are provided in this paper. The authors also provide the results on wandb. It makes obtaining the detailed results for each setting very convenient for the following researchers. \n\n&nbsp;\n\n### Weaknesses\n\n&nbsp;\n\n- Some important continual learning baselines are not included, such as iCaRL (Rebuffi et al., 2017). \n\n&nbsp;\n\n- The authors only provide experiment results on small-scale datasets, such as MNIST and CIFAR. Most continual learning papers, such as iCaRL (Rebuffi et al., 2017) and BiC (Wu et al., 2019), provide the results on large-scale datasets (e.g., ImageNet-1k). As the authors trying to establish a new benchmark protocol, it is important to provide the results on large-scale datasets.\n\n&nbsp;\n\n- This project includes plenty of different methods, but the authors don’t include detailed information (such as the licenses) about the related open-source resources they use. \n\n&nbsp;\n\n- The definition of “incremental learning” in Section 2.1 is ambiguous. I think it is better to use class-incremental learning (or domain-incremental learning) directly. The reasons are as follows. (1) “Incremental learning” is often considered as another name of “continual learning”. It is weird to use it to denote a specific setting of continual learning. (2) You include class-IL and domain-IL in “incremental learning”, but you exclude task-IL. It is not reasonable.\n\n&nbsp;\n\n- The authors include too many details about the code implementation in the paper. I think it is better to move these parts to the appendix and include more experimental results and analyses. ", "summary_of_the_review": "Overall, I think this project will be a useful tool for the following continual learning researchers. I will recommend acceptance if the authors can address my concerns in the “weaknesses” part. \n\n&nbsp;\n\n### === Post-rebuttal Comments ===\n\nI thought the authors aimed to establish a unified software framework that makes running continual learning experiments easy.\nHowever, after reading the rebuttal, I think Sequoia has the following major issues and the authors failed to address them in the rebuttal:\n\n- ***Sequoia heavily relies on the previous libraries, such as Avalanche and Continuum.*** I don't think this design is very friendly to the researchers. In my personal view, I prefer a framework that is easy to be understood and includes the most popular baselines. I think your framework should be designed for a researcher instead of a software engineer.\n\n- ***Sequoia hasn't been evaluated on large-scale datasets (e.g., ImageNet-1k).*** If I need to use this framework, I need the framework can reproduce the results of the previous baselines correctly. \n\nSo, I don't think Sequoia is very useful to a researcher like me. According to my personal experience, I tend to reject this paper. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["Yes, Legal compliance (e.g., GDPR, copyright, terms of use)"], "details_of_ethics_concerns": "This project includes plenty of different methods, but the authors don’t include detailed information (such as the licenses) about the related open-source resources they use. ", "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635871813953}, {"id": "DvjgSVgmKBi", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1907/Reviewer_pX2k"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a theoretical framework to organize research problems in the continual learning (CL) domain according to a hierarchy. This theoretical framework is used as the basic foundation for Sequoia, a software library designed to reuse methods (i.e. training algorithms) across different research problems (settings).", "review_text": "The goal of the paper is very ambitious. There exist many open source libraries implementing CL methods for supervised learning and reinforcement learning settings.\n\nFRAMEWORK:\nAccording to the paper \"each setting is described as a set of assumptions. A tree-shaped hierarchy emerges from this view\". This shows a limitation of the theoretical framework. Defining settings as a set of assumptions does not result in a tree-shaped hierarchy (it would be a lattice). Therefore, using a tree-shaped hierarchy means that some methods will be compatible in principle but not in practice (in Sequoia) because the tree lacks the connection. This a strong limitation of the framework that is not discussed.\n\nIMPLEMENTATION:\nThe framework heavily relies on several dependencies to implement the heavy lifting (datasets, RL environments, Methods). This is not a big problem by itself, but it is not clear what Sequoia is adding compared to the original libraries.\n\nEXPERIMENTS:\nThe experiments show several baselines in the supervised and reinforcement learning settings. Most methods are not implemented by Sequoia and are inherited from avalanche for SL and stable-baselines and continual-world for RL. One thing that I expected from this section was how the same method could be easily applied to different settings, which is the main claim of the library. Instead, SL and RL settings share zero methods. For example, the EWC implementation is different in the SL and RL settings. Does this mean that users have to implement every method twice? It seems that RL and SL methods are completely separate, which is against the entire spirit of the library. At this point, what is the advantage that Sequoia brings compared to its dependencies (continuum, gym, avalanche, stable-baselines, continual-world)?\n\nADDITIONAL COMMENTS:\n- \"constraints often relate to memory, compute, or time allowed to learn a task\" -> is Sequoia able to check time and memory constraints?\n- is it possible for the end users to define new settings or modify the hierarchy? it would be interesting to see an example.\n- \"We note that some Avalanche methods achieve lower than chance accuracy in task-IL because they do not use the task label to mask out the classes that lie outside the tested task\" -> can you fix this problem? Otherwise a user needs to know the internals of each library to debug the code. This adds a lot of complexity.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper proposes a theoretical framework to organize research problems in the continual learning (CL) domain according to a hierarchy. This theoretical framework is used as the basic foundation for Sequoia, a software library designed to reuse methods (i.e. training algorithms) across different research problems (settings).", "main_review": "The goal of the paper is very ambitious. There exist many open source libraries implementing CL methods for supervised learning and reinforcement learning settings.\n\nFRAMEWORK:\nAccording to the paper \"each setting is described as a set of assumptions. A tree-shaped hierarchy emerges from this view\". This shows a limitation of the theoretical framework. Defining settings as a set of assumptions does not result in a tree-shaped hierarchy (it would be a lattice). Therefore, using a tree-shaped hierarchy means that some methods will be compatible in principle but not in practice (in Sequoia) because the tree lacks the connection. This a strong limitation of the framework that is not discussed.\n\nIMPLEMENTATION:\nThe framework heavily relies on several dependencies to implement the heavy lifting (datasets, RL environments, Methods). This is not a big problem by itself, but it is not clear what Sequoia is adding compared to the original libraries.\n\nEXPERIMENTS:\nThe experiments show several baselines in the supervised and reinforcement learning settings. Most methods are not implemented by Sequoia and are inherited from avalanche for SL and stable-baselines and continual-world for RL. One thing that I expected from this section was how the same method could be easily applied to different settings, which is the main claim of the library. Instead, SL and RL settings share zero methods. For example, the EWC implementation is different in the SL and RL settings. Does this mean that users have to implement every method twice? It seems that RL and SL methods are completely separate, which is against the entire spirit of the library. At this point, what is the advantage that Sequoia brings compared to its dependencies (continuum, gym, avalanche, stable-baselines, continual-world)?\n\nADDITIONAL COMMENTS:\n- \"constraints often relate to memory, compute, or time allowed to learn a task\" -> is Sequoia able to check time and memory constraints?\n- is it possible for the end users to define new settings or modify the hierarchy? it would be interesting to see an example.\n- \"We note that some Avalanche methods achieve lower than chance accuracy in task-IL because they do not use the task label to mask out the classes that lie outside the tested task\" -> can you fix this problem? Otherwise a user needs to know the internals of each library to debug the code. This adds a lot of complexity.\n", "summary_of_the_review": "I believe the objective of this paper of merging together continual SL and RL is very important and ambitious. However, the paper in the present form has several weaknesses. It is unclear whether the theoretical framework really achieves the paper’s objective. Sequoia (the software) seems to be still at a very alpha stage in its development cycle. I see very little unification in the methods, which is the main scope of the paper. This strongly limits the methods' reuse between the different settings. It is unclear what advantages sequoia is bringing compared to using its dependencies directly.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635868675182}, {"id": "UwjAKsTbSl", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1907/Reviewer_YwgU"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new continual learning framework that aims to boost the research in the field. This framework is based on a taxonomy of all possible assumptions that are common to CL methods. Moreover, this taxonomy helps in putting supervised and reinforcement methods in a unified framework.\n", "review_text": "- In section 2, the authors propose a Markovian process that is being ignored in the later parts. It is not clear why this hidden-mode Markov decision process is useful to the proposed framework.\n- While I do appreciate the effort in developing this framework, I am lacking the novelty in the paper. The work is mainly an engineering effort that's appreciated but might not fit this conference.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces a new continual learning framework that aims to boost the research in the field. This framework is based on a taxonomy of all possible assumptions that are common to CL methods. Moreover, this taxonomy helps in putting supervised and reinforcement methods in a unified framework.\n", "main_review": "- In section 2, the authors propose a Markovian process that is being ignored in the later parts. It is not clear why this hidden-mode Markov decision process is useful to the proposed framework.\n- While I do appreciate the effort in developing this framework, I am lacking the novelty in the paper. The work is mainly an engineering effort that's appreciated but might not fit this conference.\n", "summary_of_the_review": "While I do appreciate the effort behind this work, I doubt the match between this paper and the scope of ICLR.\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635843036671}], "openreview_url": "https://openreview.net/forum?id=xWRX16GCugt", "arxiv_id": "2108.01005", "paper_pdf": "papers/xWRX16GCugt.pdf", "paper_pdf_sha256": "21a8c76f54b72c288d72216822848f13692c8c98857abd6e4e9f8fe1d4d71b72", "paper_pdf_bytes": 1437338, "paper_pdf_source": "openreview", "code_url": "https://github.com/lebrice/Sequoia", "code_repository": "lebrice/Sequoia", "code_commit": "7e12ff8ed67fada8cf220c5c396dc26332f558c2", "code_archive": "repos/xWRX16GCugt.zip", "code_archive_sha256": "05456f055e07f93b2e86c5a0e2276450461d8d3697232848c5c22682ecd48cf9", "code_archive_bytes": 926909, "code_file_count": 399, "code_extensions": {".py": 387, ".sh": 11, ".ipynb": 1}, "github_disk_usage_kb": 7144, "github_languages": {"Python": 2192958, "Shell": 11640, "Dockerfile": 4441}, "github_archived": false, "github_pushed_at": "2023-05-30T15:33:28Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/sequoia-a-software-framework-to-unify"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "1-Mh-cWROZ", "year": 2021, "status": "rejected", "title": "Fold2Seq: A Joint Sequence(1D)-Fold(3D) Embedding-based Generative Model for Protein Design", "authors": ["Yue Cao", "Payel Das", "Pin-Yu Chen", "Vijil Chenthamarakshan", "Igor Melnyk", "Yang Shen"], "authorids": ["~Yue_Cao4", "~Payel_Das1", "~Pin-Yu_Chen1", "~Vijil_Chenthamarakshan1", "~Igor_Melnyk1", "~Yang_Shen4"], "authors_source": "OpenReview API", "abstract": "Designing novel protein sequences consistent with a desired 3D structure  or fold, often referred to as the inverse protein folding problem, is a central, but non-trivial, task in protein engineering.  It has a wide range of applications  in energy, biomedicine, and materials science. However, challenges exist due to the complex sequence-fold relationship and difficulties associated with  modeling 3D folds. To overcome these challenges, we propose Fold2Seq, a novel transformer-based generative framework for designing protein sequences conditioned on a specific fold. Our model learns a fold  embedding from the density of the secondary structural elements in 3D voxels, and then models  the complex sequence-structure relationship by learning a joint sequence-fold embedding. Experiments on  high-resolution, complete, and single-structure test set demonstrate improved  performance of Fold2Seq in terms of speed and reliability for sequence design, compared to existing baselines including the state-of-the-art RosettaDesign and other neural net-based approaches. The unique advantages of fold-based Fold2Seq becomes more evident on diverse real-world test sets comprised of low-resolution, incomplete, or ensemble structures,  in comparison to a structure-based model. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "6EdRK_ac6nt", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1872/AnonReviewer5"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors of this paper propose a Transformer-based multimodal autoencoder architecture (including a sequence encoder, a structure encoder, and a sequence decoder) for protein design. In the method, a claimed novelty is that a 3D proteins structure is represented by 3D voxels of density of secondary structures. Transformer is used for sequence modelling and structural modelling. A another major claim of this paper is the joint sequence-fold representation learning. The objective function reflects the needs for both intra-domain representation learning and cross-domain representation learning. Comparisons with a principled-based method - RosettaDesign and two machine learning methods - cVAE and gcWGAN at both sequence level and structural level show that the proposed method can achieve better performance on sets of proteins with limited sizes. Overall, this is an interesting work for 3D structure learning and design. However, I also have the following concerns. \n\nMajors:\n1. A possible major drawback of the proposed structural representation method is that it relies on rescaling of very large structures to the 3D frame of fixed size. In the experiments, proteins with more than 200 amino acids were filtered out. Moreover, in the inference phase, this representation can cause a complication to the length of generated sequences.\n\n2. In the training phase, it is unclear whether L_1 and L_2 are optimized alternatingly along multiple epochs, or L_1 is optimized for multiple epochs then L_2 for multiple epochs. How much gain was obtained in comparison with just direct optimization of the objective function with five terms in terms of learning curves and performance at sequence and structure levels? I guess the unbalanced training progress is due to the fact that training sequences is easier than training 3D structures. \n\nMinors:\nposition(3D -> position (3D\ntransformer -> Transformer\nIn the text, the authors state that \"we consider to maximize the cosine similarity (shown as the ‘Cosine Similarity’\npink block in Fig 2) ...\", however, \"Cross Similarity\" is given in Figure 2. \na MLP -? an MLP\nPytorch -> PyTorch\nai-driven -> AI-driven [in the references, check similar issues]", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting structural representation idea and joint sequence-fold learning, but limiting sequence size", "review": "The authors of this paper propose a Transformer-based multimodal autoencoder architecture (including a sequence encoder, a structure encoder, and a sequence decoder) for protein design. In the method, a claimed novelty is that a 3D proteins structure is represented by 3D voxels of density of secondary structures. Transformer is used for sequence modelling and structural modelling. A another major claim of this paper is the joint sequence-fold representation learning. The objective function reflects the needs for both intra-domain representation learning and cross-domain representation learning. Comparisons with a principled-based method - RosettaDesign and two machine learning methods - cVAE and gcWGAN at both sequence level and structural level show that the proposed method can achieve better performance on sets of proteins with limited sizes. Overall, this is an interesting work for 3D structure learning and design. However, I also have the following concerns. \n\nMajors:\n1. A possible major drawback of the proposed structural representation method is that it relies on rescaling of very large structures to the 3D frame of fixed size. In the experiments, proteins with more than 200 amino acids were filtered out. Moreover, in the inference phase, this representation can cause a complication to the length of generated sequences.\n\n2. In the training phase, it is unclear whether L_1 and L_2 are optimized alternatingly along multiple epochs, or L_1 is optimized for multiple epochs then L_2 for multiple epochs. How much gain was obtained in comparison with just direct optimization of the objective function with five terms in terms of learning curves and performance at sequence and structure levels? I guess the unbalanced training progress is due to the fact that training sequences is easier than training 3D structures. \n\nMinors:\nposition(3D -> position (3D\ntransformer -> Transformer\nIn the text, the authors state that \"we consider to maximize the cosine similarity (shown as the ‘Cosine Similarity’\npink block in Fig 2) ...\", however, \"Cross Similarity\" is given in Figure 2. \na MLP -? an MLP\nPytorch -> PyTorch\nai-driven -> AI-driven [in the references, check similar issues]", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604625662657}, {"id": "i6ICc4POer", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1872/AnonReviewer3"], "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This manuscript presents a method for generating protein sequences conditioned on protein structures. The core idea is to represented protein structures by their secondary structures in 3D space. This voxel grid is then encoded into a vector representation and decoded to a distribution over sequences. The authors propose to learn this model jointly with a sequence encoder, combining the sequence and structure representations to decode the sequence during training. For inference, the sequence encoder component is not used. Learning to generate protein sequence conditioned on structure is an interesting and important problem and has been attracting increasing attention from the ML community. Representing structures as voxel grids is an approach worth exploring and flexible structure representations could be promising. However, it isn’t clear to me that this work achieves those goals and comparisons against key baselines (namely Ingraham 2019) are missing. Furthermore, the authors make many unsupported and unsubstantiated claims about their method. Specific comments and questions follow below.\n\n1.\tThe claim that this method allows flexible fold representations seems to be undermined by the need to specify a specific structure for which sequences are decoded. How does this method allow for more flexible structure representations than Ingraham 2019 or other methods that require 3D coordinates of each residue? It isn’t clear that this representation is more flexible than that required by RosettaDesign either. The authors need to justify this claim or remove it.\n2.\tThe authors also claim that their method preserves spatial secondary structure information in contrast to others. However, all backbone-based protein structure representations preserve secondary structure information as this is defined by the backbone angles. The authors also claim that their method does not need predefined rules or structures. However, their structure representation is derived from a predefined structure, so this claim makes no sense.\n3.\tHow does the density model loosen the rigidity of structures? The density model only represents a single structure and the authors do not use structure ensembles or other means to represent structural flexibility. I fail to see how this claim is supported by the presented method or experiments. In principle, any structure->sequence model can account for flexibility by ensembling over possible 3D structure.\n4.\tHave the authors compared their method with Ingraham et al. 2019? Ingraham et al solve the same problem with, arguably, an even more flexible representation of protein structures, because the graph based representation is invariant to rotation and translation of the structure in 3D space. A comparison against Ingraham et al is critical to understand if the proposed voxel-based structure representation is better than other methods. At face value, Ingraham 2019 reports better perplexities with a difficult train/test split and fewer training examples than this work. In fact, Ingraham et al’s out of distribution perplexity surpasses this method’s in distribution perplexity calling into question the utility of this method. The authors should compare their method on the same train/val/test split used by Ingraham et al.\n5.\tFor the in distribution test, how do sequence models fit per fold perform? For example, how well do profile HMMs fit to each fold model the heldout sequences from those folds?\n6.\tWhy are the val and test splits so small (only 2.5%)? Given these splits, how many sequences occur in each?\n7.\tBecause this method encodes structures into a voxel grid, it is not invariant to rotation and translation of proteins in 3D space. The authors align their structures based on center of mass and orient the protein such that the first residue points in the negative direction of the z-axis. How sensitive is the method to misaligned proteins? Is this alignment scheme robust to alternative conformation of the same protein? What happens when structures are rotated? The described alignment scheme also has ambiguity in the rotation around the z-axis. How is this addressed?\n\nThings that would improve my score:\n1.\tProvide support for unsupported claims in the introduction.\n2.\tCompare against Ingraham et al 2019.\n3.\tExamine the sensitivity of this method to rotated structures.\n\nEdit: I have read the authors' response and update  manuscript. I appreciate the new experiments and some of the clarifications. However, some of my fundamental issues with this work are not addressed.\n\nThe distinction with Ingraham et al feels forced. It is important to compare head-to-head on Ingraham's dataset in order to truly understand the tradeoffs between these methods and to understand the differences in performance between flexible backbone approaches and a well tuned rigid backbone approach. Why are perplexities for Ingraham et al not reported for all cases? Improving over Ingraham when only low resolution structures or incomplete structures are available is interesting, but I also question how useful this case is. When attempting to do _structure-based design_ of new proteins, how often are fully specified structures not available for those tasks? Ingraham et al presented a variant of their method based only on a flexible structure specification and it isn't clear how that contrasts with this flexible specification. It is certainly possible that this approach is better, but we need a head-to-head comparison to know.\n\nWith that in mind, I'll raise my score to 5. I think this work has promise, but is not yet ready for publication in its current state.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "An important problem, but this work is too preliminary and makes too many unsubstantiated claims", "review": "This manuscript presents a method for generating protein sequences conditioned on protein structures. The core idea is to represented protein structures by their secondary structures in 3D space. This voxel grid is then encoded into a vector representation and decoded to a distribution over sequences. The authors propose to learn this model jointly with a sequence encoder, combining the sequence and structure representations to decode the sequence during training. For inference, the sequence encoder component is not used. Learning to generate protein sequence conditioned on structure is an interesting and important problem and has been attracting increasing attention from the ML community. Representing structures as voxel grids is an approach worth exploring and flexible structure representations could be promising. However, it isn’t clear to me that this work achieves those goals and comparisons against key baselines (namely Ingraham 2019) are missing. Furthermore, the authors make many unsupported and unsubstantiated claims about their method. Specific comments and questions follow below.\n\n1.\tThe claim that this method allows flexible fold representations seems to be undermined by the need to specify a specific structure for which sequences are decoded. How does this method allow for more flexible structure representations than Ingraham 2019 or other methods that require 3D coordinates of each residue? It isn’t clear that this representation is more flexible than that required by RosettaDesign either. The authors need to justify this claim or remove it.\n2.\tThe authors also claim that their method preserves spatial secondary structure information in contrast to others. However, all backbone-based protein structure representations preserve secondary structure information as this is defined by the backbone angles. The authors also claim that their method does not need predefined rules or structures. However, their structure representation is derived from a predefined structure, so this claim makes no sense.\n3.\tHow does the density model loosen the rigidity of structures? The density model only represents a single structure and the authors do not use structure ensembles or other means to represent structural flexibility. I fail to see how this claim is supported by the presented method or experiments. In principle, any structure->sequence model can account for flexibility by ensembling over possible 3D structure.\n4.\tHave the authors compared their method with Ingraham et al. 2019? Ingraham et al solve the same problem with, arguably, an even more flexible representation of protein structures, because the graph based representation is invariant to rotation and translation of the structure in 3D space. A comparison against Ingraham et al is critical to understand if the proposed voxel-based structure representation is better than other methods. At face value, Ingraham 2019 reports better perplexities with a difficult train/test split and fewer training examples than this work. In fact, Ingraham et al’s out of distribution perplexity surpasses this method’s in distribution perplexity calling into question the utility of this method. The authors should compare their method on the same train/val/test split used by Ingraham et al.\n5.\tFor the in distribution test, how do sequence models fit per fold perform? For example, how well do profile HMMs fit to each fold model the heldout sequences from those folds?\n6.\tWhy are the val and test splits so small (only 2.5%)? Given these splits, how many sequences occur in each?\n7.\tBecause this method encodes structures into a voxel grid, it is not invariant to rotation and translation of proteins in 3D space. The authors align their structures based on center of mass and orient the protein such that the first residue points in the negative direction of the z-axis. How sensitive is the method to misaligned proteins? Is this alignment scheme robust to alternative conformation of the same protein? What happens when structures are rotated? The described alignment scheme also has ambiguity in the rotation around the z-axis. How is this addressed?\n\nThings that would improve my score:\n1.\tProvide support for unsupported claims in the introduction.\n2.\tCompare against Ingraham et al 2019.\n3.\tExamine the sensitivity of this method to rotated structures.\n\nEdit: I have read the authors' response and update  manuscript. I appreciate the new experiments and some of the clarifications. However, some of my fundamental issues with this work are not addressed.\n\nThe distinction with Ingraham et al feels forced. It is important to compare head-to-head on Ingraham's dataset in order to truly understand the tradeoffs between these methods and to understand the differences in performance between flexible backbone approaches and a well tuned rigid backbone approach. Why are perplexities for Ingraham et al not reported for all cases? Improving over Ingraham when only low resolution structures or incomplete structures are available is interesting, but I also question how useful this case is. When attempting to do _structure-based design_ of new proteins, how often are fully specified structures not available for those tasks? Ingraham et al presented a variant of their method based only on a flexible structure specification and it isn't clear how that contrasts with this flexible specification. It is certainly possible that this approach is better, but we need a head-to-head comparison to know.\n\nWith that in mind, I'll raise my score to 5. I think this work has promise, but is not yet ready for publication in its current state.", "rating": "5: Marginally below acceptance threshold", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1604469604586}, {"id": "x5IFXZfLkZ", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1872/AnonReviewer2"], "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This is an interesting paper on predicting the underlying amino acid sequence from a given protein tertiary (3D) structure, an important problem with multiple applications. The authors have proposed an elaborate system and performed an impressive set of experiments to demonstrate the efficacy of their model. The proposed method (Fold2Seq) has several novel contributions, which I list below alongside my comments on them.\n\n1. A novel fold representation based on voxels of the density of secondary structure elements (SSEs). This is promising, however, I am not sure how sensitive to small changes in the 3D structure this representation is. The authors shift the structure such that its center of mass is at the origin and its first CA is on the negative side of the Z-axis. What happens only if the first residue is moved around? This is considering that the residues in the N-terminal usually do not have a solid form. Moreover, since the authors focus on short proteins (N < 200), I am curious to know how they handle structures determined by NMR, in which an ensemble of structures is determined, instead of a single structure.\n2. A novel joint sequence-fold embedding is learned that is employed in the transformer-based auto-encoder model. This is a great idea and seems to play an important role in Fold2Seq’s performance. Imagine two proteins that have two domains each and share one domain. How would their fold and sequence embeddings look? I am concerned that Fold2Seq might memorize some domains. For example, the drastic difference between Fold2Seq and RosettaDesign TM scores in ID and OD test sets is concerning. While RosettaDesign performs similarly in ID and OD sets, Fold2Seq has a significant drop in performance. I think the distribution of sequence similarities between each test sequence and the closest sequence in ID and OD sets should be provided. \n3. I am not sure if splitting the datasets solely based on “fold” is enough to prevent information leakage. I think it should be both based on fold and sequence similarity. Also, since proteins are so small, it might be helpful if their domains are also considered, e.g., using Pfam or computational methods such as doi.org/10.1101/626507.\n4. It is not clear to me why the authors did not compare to methods without the flexible backbone constraint (e.g., Ingraham 2019), the rationale is not fully convincing.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A promising method for protein structure to sequence mapping ", "review": "This is an interesting paper on predicting the underlying amino acid sequence from a given protein tertiary (3D) structure, an important problem with multiple applications. The authors have proposed an elaborate system and performed an impressive set of experiments to demonstrate the efficacy of their model. The proposed method (Fold2Seq) has several novel contributions, which I list below alongside my comments on them.\n\n1. A novel fold representation based on voxels of the density of secondary structure elements (SSEs). This is promising, however, I am not sure how sensitive to small changes in the 3D structure this representation is. The authors shift the structure such that its center of mass is at the origin and its first CA is on the negative side of the Z-axis. What happens only if the first residue is moved around? This is considering that the residues in the N-terminal usually do not have a solid form. Moreover, since the authors focus on short proteins (N < 200), I am curious to know how they handle structures determined by NMR, in which an ensemble of structures is determined, instead of a single structure.\n2. A novel joint sequence-fold embedding is learned that is employed in the transformer-based auto-encoder model. This is a great idea and seems to play an important role in Fold2Seq’s performance. Imagine two proteins that have two domains each and share one domain. How would their fold and sequence embeddings look? I am concerned that Fold2Seq might memorize some domains. For example, the drastic difference between Fold2Seq and RosettaDesign TM scores in ID and OD test sets is concerning. While RosettaDesign performs similarly in ID and OD sets, Fold2Seq has a significant drop in performance. I think the distribution of sequence similarities between each test sequence and the closest sequence in ID and OD sets should be provided. \n3. I am not sure if splitting the datasets solely based on “fold” is enough to prevent information leakage. I think it should be both based on fold and sequence similarity. Also, since proteins are so small, it might be helpful if their domains are also considered, e.g., using Pfam or computational methods such as doi.org/10.1101/626507.\n4. It is not clear to me why the authors did not compare to methods without the flexible backbone constraint (e.g., Ingraham 2019), the rationale is not fully convincing.\n", "rating": "7: Good paper, accept", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603946089275}, {"id": "wJEX1kC4RB3", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1872/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper tackle the challenge of designing protein sequences that are consistent with a given 3D fold. To address this challenge, the authors propose a transformer-based generative framework that designs protein sequences conditioned on a given fold. There are two central contributions - the first is a novel fold representation, in which the 3D structure is represented by the voxels of secondary structure elements, and then a fold representation is learned via a transformer-based structure encoder. The second is a joint sequence-fold embedding learning framework. The authors use ablation studies to show that learning a joint latent space between sequences and folds enables the model to better capture the the sequence-fold relationship, improving experimental results. The authors provide a good summary of related work. \n\nTo proceed with the novel representation, the authors first scale each protein structure to fit into a fixed size cubic box, discretized into fixed size voxels, and extract the alpha carbon coordinates of each amino acid. They assign secondary structure labels drawn from a 4-letter alphabet to each residue, and compute the resulting features of each voxel. These features are used to train a structure encoder, which is used together with a trained sequence encoder to learn a joint sequence-fold embedding. Here fold classification is used for the intra-domain task, while for the cross domain loss they maximize the cosine similarity between the outputs of the sequence and structure encoders in addition to a cyclic loss. \n\nThe authors use the data from CATH 4.2 to define two tasks based on a random (ID) split, and a fold-level (OD) split of the data. In each case 2.5% of the data is held out as validation and test sets respectively. They perform various experiments that compare the performance of their approach against cVAE, gcWGAN and RosettaDesign. The performance of this approach is impressive, however I am surprised by the decision to leave out methods that do not focus on the inverse folding problem with a flexible backbone constraint - I would like to see the performance comparison with these methods. Please could the authors adequately justify this decision, or provide these comparisons. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "FOLD2SEQ: A JOINT SEQUENCE(1D)-FOLD(3D) EMBEDDING-BASED GENERATIVE MODEL FOR PROTEIN DESIGN", "review": "This paper tackle the challenge of designing protein sequences that are consistent with a given 3D fold. To address this challenge, the authors propose a transformer-based generative framework that designs protein sequences conditioned on a given fold. There are two central contributions - the first is a novel fold representation, in which the 3D structure is represented by the voxels of secondary structure elements, and then a fold representation is learned via a transformer-based structure encoder. The second is a joint sequence-fold embedding learning framework. The authors use ablation studies to show that learning a joint latent space between sequences and folds enables the model to better capture the the sequence-fold relationship, improving experimental results. The authors provide a good summary of related work. \n\nTo proceed with the novel representation, the authors first scale each protein structure to fit into a fixed size cubic box, discretized into fixed size voxels, and extract the alpha carbon coordinates of each amino acid. They assign secondary structure labels drawn from a 4-letter alphabet to each residue, and compute the resulting features of each voxel. These features are used to train a structure encoder, which is used together with a trained sequence encoder to learn a joint sequence-fold embedding. Here fold classification is used for the intra-domain task, while for the cross domain loss they maximize the cosine similarity between the outputs of the sequence and structure encoders in addition to a cyclic loss. \n\nThe authors use the data from CATH 4.2 to define two tasks based on a random (ID) split, and a fold-level (OD) split of the data. In each case 2.5% of the data is held out as validation and test sets respectively. They perform various experiments that compare the performance of their approach against cVAE, gcWGAN and RosettaDesign. The performance of this approach is impressive, however I am surprised by the decision to leave out methods that do not focus on the inverse folding problem with a flexible backbone constraint - I would like to see the performance comparison with these methods. Please could the authors adequately justify this decision, or provide these comparisons. \n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603942798262}], "openreview_url": "https://openreview.net/forum?id=1-Mh-cWROZ", "arxiv_id": "2106.13058", "paper_pdf": "papers/1-Mh-cWROZ.pdf", "paper_pdf_sha256": "d83ecfc36418ba1394affebdc567665760fded62a0e85f6fc65b1ebcf908b16f", "paper_pdf_bytes": 6948439, "paper_pdf_source": "openreview", "code_url": "https://github.com/IBM/fold2seq", "code_repository": "IBM/fold2seq", "code_commit": "b9a97d81eac329b5259ad10e2a6f4fe80ade542f", "code_archive": "repos/1-Mh-cWROZ.zip", "code_archive_sha256": "915d3bb6b0d7ae80a24fbac44b57eaf8e952f7d5cd410bd141857458fdbc8fec", "code_archive_bytes": 3877907, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 5093, "github_languages": {"Python": 62341}, "github_archived": true, "github_pushed_at": "2022-05-25T06:49:33Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/fold2seq-a-joint-sequence-1d-fold-3d-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qVyjN01x4P", "year": 2025, "status": "rejected", "title": "Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift", "authors": ["Yanru Sun", "Zongxia Xie", "Emadeldeen Eldele", "Dongyue Chen", "Qinghua Hu", "Min Wu"], "authorids": ["~Yanru_Sun1", "~Zongxia_Xie1", "~Emadeldeen_Eldele1", "~Dongyue_Chen3", "~Qinghua_Hu1", "~Min_Wu2"], "authors_source": "OpenReview API", "abstract": "Time series forecasting, which aims to predict future values based on historical data, has garnered significant attention due to its broad range of applications.  However, real-world time series often exhibit complex non-uniform distribution with varying patterns across segments, such as season, operating condition, or semantic meaning, making accurate forecasting challenging. Existing approaches, which typically train a single model to capture all these diverse patterns, often struggle with the pattern drifts between patches and may lead to poor generalization. To address these challenges, we propose TFPS, a novel architecture that leverages pattern-specific experts for more accurate and adaptable time series forecasting. TFPS employs a dual-domain encoder to capture both time-domain and frequency-domain features, enabling a more comprehensive understanding of temporal dynamics. It then uses subspace clustering to dynamically identify distinct patterns across data patches. Finally, pattern-specific experts model these unique patterns, delivering tailored predictions for each patch. By explicitly learning and adapting to evolving patterns, TFPS achieves significantly improved forecasting accuracy. Extensive experiments on real-world datasets demonstrate that TFPS outperforms state-of-the-art methods, particularly in long-term forecasting, through its dynamic and pattern-aware learning approach. The data and codes are available: https://anonymous.4open.science/r/TFPS-D001.", "decision": "Reject", "meta_review": "", "num_reviews": 5, "reviews": [{"id": "GMo5faOa1t", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5645/Reviewer_ENp4"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "The paper innovatively addresses the diversity of time series patterns by introducing a Mixture of Experts (MoE) approach for decoupled modeling. It leverages a unique PI module as the gating function to route different patterns to the appropriate experts. The experimental results convincingly demonstrate the superiority of this method, complemented by comprehensive ablation studies validating the effectiveness of each component. The writing is clear, with the authors articulating the problem and solution effectively. However, the design of the expert models appears somewhat simplistic, and the decoupling of time and frequency domains is already a well-established approach in time series analysis, making it less of an innovative contribution. Overall, the method shows sufficient technical depth and originality, despite some areas that could benefit from further enhancement.", "review_text": "The paper innovatively addresses the diversity of time series patterns by introducing a Mixture of Experts (MoE) approach for decoupled modeling. It leverages a unique PI module as the gating function to route different patterns to the appropriate experts. The experimental results convincingly demonstrate the superiority of this method, complemented by comprehensive ablation studies validating the effectiveness of each component. The writing is clear, with the authors articulating the problem and solution effectively. However, the design of the expert models appears somewhat simplistic, and the decoupling of time and frequency domains is already a well-established approach in time series analysis, making it less of an innovative contribution. Overall, the method shows sufficient technical depth and originality, despite some areas that could benefit from further enhancement.", "strengths": "The authors conducted a detailed experimental analysis and identified that different time series exhibit unique patterns. Traditional models typically employ a unified modeling approach for these varying patterns, which can limit their effectiveness. To address this issue, the authors propose a Mixture of Experts (MoE) approach that enables tailored modeling for different patterns. Additionally, they decompose the time series data into two dimensions: time and frequency, using a dual encoder to model both types of information separately. To ensure that the MoE effectively learns distinct patterns, the authors introduce a novel gating mechanism. This mechanism leverages subspace affinity to allocate sequences with different patterns to the appropriate experts.", "weaknesses": "The decoupling of time-domain and frequency-domain information, as discussed in the paper, is already a common approach in the field of time series analysis and cannot be considered a novel contribution. Additionally, regarding the MoE (Mixture of Experts) section, the use of multiple MLPs seems more akin to employing multiple output predictors with weighted combinations rather than a true MoE structure. Lastly, with respect to the normalization technique, would the authors consider adding experimental results comparing their approach to REVIN [1]? REVIN has shown strong performance across various time series forecasting tasks and could provide a valuable benchmark.\n\n\n[1] Kim T, Kim J, Tae Y, et al. Reversible instance normalization for accurate time-series forecasting against distribution shift[C]//International Conference on Learning Representations. 2021.", "questions": "1. In the PI module, the number of experts K is currently adjusted manually through parameter tuning. Is it possible to directly relate the value of K to the number of patterns inherent in the time series data itself?\n2. On line 317, there appear to be two identical objectives listed, which seems to be a typographical error.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper innovatively addresses the diversity of time series patterns by introducing a Mixture of Experts (MoE) approach for decoupled modeling. It leverages a unique PI module as the gating function to route different patterns to the appropriate experts. The experimental results convincingly demonstrate the superiority of this method, complemented by comprehensive ablation studies validating the effectiveness of each component. The writing is clear, with the authors articulating the problem and solution effectively. However, the design of the expert models appears somewhat simplistic, and the decoupling of time and frequency domains is already a well-established approach in time series analysis, making it less of an innovative contribution. Overall, the method shows sufficient technical depth and originality, despite some areas that could benefit from further enhancement.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "The authors conducted a detailed experimental analysis and identified that different time series exhibit unique patterns. Traditional models typically employ a unified modeling approach for these varying patterns, which can limit their effectiveness. To address this issue, the authors propose a Mixture of Experts (MoE) approach that enables tailored modeling for different patterns. Additionally, they decompose the time series data into two dimensions: time and frequency, using a dual encoder to model both types of information separately. To ensure that the MoE effectively learns distinct patterns, the authors introduce a novel gating mechanism. This mechanism leverages subspace affinity to allocate sequences with different patterns to the appropriate experts.", "weaknesses": "The decoupling of time-domain and frequency-domain information, as discussed in the paper, is already a common approach in the field of time series analysis and cannot be considered a novel contribution. Additionally, regarding the MoE (Mixture of Experts) section, the use of multiple MLPs seems more akin to employing multiple output predictors with weighted combinations rather than a true MoE structure. Lastly, with respect to the normalization technique, would the authors consider adding experimental results comparing their approach to REVIN [1]? REVIN has shown strong performance across various time series forecasting tasks and could provide a valuable benchmark.\n\n\n[1] Kim T, Kim J, Tae Y, et al. Reversible instance normalization for accurate time-series forecasting against distribution shift[C]//International Conference on Learning Representations. 2021.", "questions": "1. In the PI module, the number of experts K is currently adjusted manually through parameter tuning. Is it possible to directly relate the value of K to the number of patterns inherent in the time series data itself?\n2. On line 317, there appear to be two identical objectives listed, which seems to be a typographical error.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731408754962}, {"id": "ICzcRPYHvX", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5645/Reviewer_fcr7"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 4, "summary": "The paper addresses the challenge of distributional shift in time series data by proposing a novel framework called the Time Frequency Pattern-Specific (TFPS) architecture. This method is designed to effectively model complex patterns in time series, enhancing the model's ability to manage distributional shifts more robustly.", "review_text": "The paper addresses the challenge of distributional shift in time series data by proposing a novel framework called the Time Frequency Pattern-Specific (TFPS) architecture. This method is designed to effectively model complex patterns in time series, enhancing the model's ability to manage distributional shifts more robustly.", "strengths": "1. The paper is well written and easy to follow.\n2. The paper is well motivated.\n3. The method shows promising results compared to the baselines.", "weaknesses": "1. The number of subspace variables $k$, does not appear to be clearly specified in the paper. Given that these variables are essential for modeling the patterns of patches, it is crucial to understand how to set an appropriate value for $k$. The authors should provide further discussion on this aspect. Additionally, an ablation study examining the impact of different values of $k$ would be beneficial.\n\n2. The paper doesn't specify how the subspace variables are initialized. I think the initialization of subspace variables is crucial to achieve optimal performance.\n\n3. Since the subspace variables is utilized for clustering. It is better to provide the relevant visualization (e.g., tsne) to show the effectiveness.\n\n4. Some important information, such as the number of MoPE, the value of hyper-parameters $\\alpha$, $\\beta$ used for the results in the Table 2 are not clearly specified.\n\n5. What do the notations $K_t$ and $K_f$ mean? The paper doesn't provide the clear definition.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper addresses the challenge of distributional shift in time series data by proposing a novel framework called the Time Frequency Pattern-Specific (TFPS) architecture. This method is designed to effectively model complex patterns in time series, enhancing the model's ability to manage distributional shifts more robustly.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "1. The paper is well written and easy to follow.\n2. The paper is well motivated.\n3. The method shows promising results compared to the baselines.", "weaknesses": "1. The number of subspace variables $k$, does not appear to be clearly specified in the paper. Given that these variables are essential for modeling the patterns of patches, it is crucial to understand how to set an appropriate value for $k$. The authors should provide further discussion on this aspect. Additionally, an ablation study examining the impact of different values of $k$ would be beneficial.\n\n2. The paper doesn't specify how the subspace variables are initialized. I think the initialization of subspace variables is crucial to achieve optimal performance.\n\n3. Since the subspace variables is utilized for clustering. It is better to provide the relevant visualization (e.g., tsne) to show the effectiveness.\n\n4. Some important information, such as the number of MoPE, the value of hyper-parameters $\\alpha$, $\\beta$ used for the results in the Table 2 are not clearly specified.\n\n5. What do the notations $K_t$ and $K_f$ mean? The paper doesn't provide the clear definition.", "questions": "see weakness", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1731132541780}, {"id": "OTVhiImh3T", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5645/Reviewer_JuoR"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 2, "summary": "The paper proposes a pattern-based mixture of experts model for time series prediction. In this approach time series patches are first extracted and embedded into a sequence of \"tokens\". This sequence is then encoded with time and frequency encoders and mined for representative patterns that can aid prediction via a mixture of experts approach.", "review_text": "The paper proposes a pattern-based mixture of experts model for time series prediction. In this approach time series patches are first extracted and embedded into a sequence of \"tokens\". This sequence is then encoded with time and frequency encoders and mined for representative patterns that can aid prediction via a mixture of experts approach.", "strengths": "The paper is well written and relatively easy to follow. The proposed approach is interesting and explores time and frequency aspect of time series prediction. Authors conduct extensive experiments on real world datasets showing that the purposed approach outperforms a number of baselines.", "weaknesses": "I have several concerns about this method. First, it is quite complex, the loss in Eq 12 has three terms and both L_{PI_t} and L_{PI_f} have three tunable hyper parameters \\alpha, \\beta and \\nu (Eq 5, 6 and 9). So there are at least six hyper parameters to tune in addition to balancing the three loss terms that likely requires more hyper parameters. There is nothing in the experimental section on these hyper parameters and no ablation study is conducted. I've worked with these datasets myself, they are quite small and easy to overfit to which raises a major concern when one has to tune 6+ hyper parameters just in the loss function. Second, the proposed mixture of experts in Eq 10 and 11 looks nearly identical to self attention with the exception of the top-k selection. So I suspect that this method will, despite significant added complexity, have a similar representation power to simply stacking attention layers over patches, as all operations including subspace bases affinity are nearly identical to appropriately parametrized self attention.", "questions": "Do you have ablation study on the hyper parameters in the loss function? Which ranges did you try and how sensitive are the results to the particular choice? Also, did you have to balance the contribution of the three terms in the loss?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a pattern-based mixture of experts model for time series prediction. In this approach time series patches are first extracted and embedded into a sequence of \"tokens\". This sequence is then encoded with time and frequency encoders and mined for representative patterns that can aid prediction via a mixture of experts approach.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "The paper is well written and relatively easy to follow. The proposed approach is interesting and explores time and frequency aspect of time series prediction. Authors conduct extensive experiments on real world datasets showing that the purposed approach outperforms a number of baselines.", "weaknesses": "I have several concerns about this method. First, it is quite complex, the loss in Eq 12 has three terms and both L_{PI_t} and L_{PI_f} have three tunable hyper parameters \\alpha, \\beta and \\nu (Eq 5, 6 and 9). So there are at least six hyper parameters to tune in addition to balancing the three loss terms that likely requires more hyper parameters. There is nothing in the experimental section on these hyper parameters and no ablation study is conducted. I've worked with these datasets myself, they are quite small and easy to overfit to which raises a major concern when one has to tune 6+ hyper parameters just in the loss function. Second, the proposed mixture of experts in Eq 10 and 11 looks nearly identical to self attention with the exception of the top-k selection. So I suspect that this method will, despite significant added complexity, have a similar representation power to simply stacking attention layers over patches, as all operations including subspace bases affinity are nearly identical to appropriately parametrized self attention.", "questions": "Do you have ablation study on the hyper parameters in the loss function? Which ranges did you try and how sensitive are the results to the particular choice? Also, did you have to balance the contribution of the three terms in the loss?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730773702036}, {"id": "fBRZmcdewX", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5645/Reviewer_n6LR"], "rating": 5, "soundness": 2, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "Real-world time series often exhibit varying degrees of distribution shift, posing challenges for time series forecasting. This paper employs subspace clustering from both time-domain and frequency-domain perspectives to identify different patterns within data segments. It utilizes multiple experts to specifically model temporal patterns. Extensive experiments conducted on eight real-world time series datasets demonstrate that the proposed method outperforms other state-of-the-art approaches in terms of predictive performance.", "review_text": "Real-world time series often exhibit varying degrees of distribution shift, posing challenges for time series forecasting. This paper employs subspace clustering from both time-domain and frequency-domain perspectives to identify different patterns within data segments. It utilizes multiple experts to specifically model temporal patterns. Extensive experiments conducted on eight real-world time series datasets demonstrate that the proposed method outperforms other state-of-the-art approaches in terms of predictive performance.", "strengths": "1. Distribution shifts often lead to changes in frequency-domain information, making it a reasonable and effective approach to model time series from both time-domain and frequency-domain perspectives.\n2. The main contribution of this paper is to utilize subspace clustering to detect concept drift between multiple subspaces and model them separately, thereby achieving more accurate and adaptive modeling. \n3. Extensive experiments demonstrate that the proposed method outperforms other state-of-the-art approaches across various time series datasets.", "weaknesses": "1. The frequency-domain encoder part of the dual-domain encoder implements a simple feedforward layer by discarding the imaginary part. This leads to a loss of information in the frequency domain. Does this contradict the original intention of using the dual-domain encoder to provide a comprehensive representation of the time series? Considering that the time-domain encoder directly utilizes the PatchTST encoder, the frequency-domain part could also use an existing frequency-domain encoder.\n2. The Pattern Identifter module employs multiple hyperparameters  $\\alpha、\\beta$ to mix different components of the loss function. Given the importance of the loss function for experimental results, it is necessary to include ablation experiments for these hyperparameters.\n3. In the final loss function, is it reasonable to simply add $\\mathcal{L}_{MSE}、\\mathcal{L}_{PI_t}、\\mathcal{L}_{PI_f}$? Given that the severity of distribution shifts varies across different datasets, and that the importance of frequency-domain information is relatively low in such datasets, the model should be allowed to learn the appropriate weights for both components on its own.\n4. It is noted that in the experiments, the sequence length (seq_len) was uniformly set to 96, which undermines the performance of other baselines, as FITS[1], PatchTST[2], and DLinear[3] have indicated that their methods perform better with longer input sequences. It is suggested to provide experimental results comparing seq_len as a hyperparameter rather than fixing it at 96.\n5. It is noted that multiple experts are used in both the time domain and frequency domain, which will increase the runtime of the model. It is suggested to compare the runtime with other methods.\n\n[1]Xu Z, Zeng A, Xu Q. FITS: Modeling Time Series with 10 k  Parameters[C]. In ICLR.\n\n[2]Nie Y, Nguyen N H, Sinthong P, et al. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers[C]. In ICLR.\n\n[3]Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting? In AAAI.", "questions": "Please refer to weeknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Real-world time series often exhibit varying degrees of distribution shift, posing challenges for time series forecasting. This paper employs subspace clustering from both time-domain and frequency-domain perspectives to identify different patterns within data segments. It utilizes multiple experts to specifically model temporal patterns. Extensive experiments conducted on eight real-world time series datasets demonstrate that the proposed method outperforms other state-of-the-art approaches in terms of predictive performance.", "soundness": 2, "presentation": 2, "contribution": 2, "strengths": "1. Distribution shifts often lead to changes in frequency-domain information, making it a reasonable and effective approach to model time series from both time-domain and frequency-domain perspectives.\n2. The main contribution of this paper is to utilize subspace clustering to detect concept drift between multiple subspaces and model them separately, thereby achieving more accurate and adaptive modeling. \n3. Extensive experiments demonstrate that the proposed method outperforms other state-of-the-art approaches across various time series datasets.", "weaknesses": "1. The frequency-domain encoder part of the dual-domain encoder implements a simple feedforward layer by discarding the imaginary part. This leads to a loss of information in the frequency domain. Does this contradict the original intention of using the dual-domain encoder to provide a comprehensive representation of the time series? Considering that the time-domain encoder directly utilizes the PatchTST encoder, the frequency-domain part could also use an existing frequency-domain encoder.\n2. The Pattern Identifter module employs multiple hyperparameters  $\\alpha、\\beta$ to mix different components of the loss function. Given the importance of the loss function for experimental results, it is necessary to include ablation experiments for these hyperparameters.\n3. In the final loss function, is it reasonable to simply add $\\mathcal{L}_{MSE}、\\mathcal{L}_{PI_t}、\\mathcal{L}_{PI_f}$? Given that the severity of distribution shifts varies across different datasets, and that the importance of frequency-domain information is relatively low in such datasets, the model should be allowed to learn the appropriate weights for both components on its own.\n4. It is noted that in the experiments, the sequence length (seq_len) was uniformly set to 96, which undermines the performance of other baselines, as FITS[1], PatchTST[2], and DLinear[3] have indicated that their methods perform better with longer input sequences. It is suggested to provide experimental results comparing seq_len as a hyperparameter rather than fixing it at 96.\n5. It is noted that multiple experts are used in both the time domain and frequency domain, which will increase the runtime of the model. It is suggested to compare the runtime with other methods.\n\n[1]Xu Z, Zeng A, Xu Q. FITS: Modeling Time Series with 10 k  Parameters[C]. In ICLR.\n\n[2]Nie Y, Nguyen N H, Sinthong P, et al. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers[C]. In ICLR.\n\n[3]Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting? In AAAI.", "questions": "Please refer to weeknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730618310883}, {"id": "GEapmSbf6Z", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission5645/Reviewer_LQda"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "This paper proposes a new architecture, Time-Frequency Pattern-Specific (TFPS), designed for time series forecasting, particularly under challenging conditions involving concept drift within the data. The key idea revolves around addressing the limitations of conventional uniform distribution modeling (UDM) by recognizing that different segments or \"patches\" within a time series exhibit distinct patterns and distributional shifts. These shifts may stem from factors such as sudden events, gradual trend changes, or different operating modes, making accurate forecasting difficult for models trained on a single global pattern. TFPS tackles this problem through a multi-faceted approach. First, it employs a Dual-Domain Encoder (DDE) that analyzes both time and frequency domain features, enabling a richer understanding of temporal dynamics and their potential shifts. Second, a Pattern Identifier (PI) utilizes subspace clustering to dynamically group patches exhibiting similar patterns. Finally, a Mixture of Pattern Experts (MoPE) leverages multiple specialized expert models, each trained on a specific pattern identified by the PI. This allows TFPS to adapt its predictions based on the identified pattern of a given patch, leading to improved accuracy.", "review_text": "This paper proposes a new architecture, Time-Frequency Pattern-Specific (TFPS), designed for time series forecasting, particularly under challenging conditions involving concept drift within the data. The key idea revolves around addressing the limitations of conventional uniform distribution modeling (UDM) by recognizing that different segments or \"patches\" within a time series exhibit distinct patterns and distributional shifts. These shifts may stem from factors such as sudden events, gradual trend changes, or different operating modes, making accurate forecasting difficult for models trained on a single global pattern. TFPS tackles this problem through a multi-faceted approach. First, it employs a Dual-Domain Encoder (DDE) that analyzes both time and frequency domain features, enabling a richer understanding of temporal dynamics and their potential shifts. Second, a Pattern Identifier (PI) utilizes subspace clustering to dynamically group patches exhibiting similar patterns. Finally, a Mixture of Pattern Experts (MoPE) leverages multiple specialized expert models, each trained on a specific pattern identified by the PI. This allows TFPS to adapt its predictions based on the identified pattern of a given patch, leading to improved accuracy.", "strengths": "(1) Utilizing both time and frequency domain information provides a comprehensive perspective on data, capturing both trend and periodic patterns, which is essential for handling complex time series.\n(2) The use of PI and MoPE allows TFPS to handle distribution shifts by learning and adapting to specific patterns within different data segments.\n(3) Extensive experiments on real-world datasets demonstrate the effectiveness and competitiveness of TFPS compared to existing methods.", "weaknesses": "(1) The authors acknowledge that the patch length is currently determined heuristically, which might limit generalizability and performance in cases with indivisible lengths or multi-period characteristics.\n(2) The model's ability to adapt to entirely new patterns, such as those arising from unforeseen events, is not explicitly addressed and remains an open question.\n(3) While not explicitly mentioned, the use of multiple expert models and the dual-domain encoding approach could increase computational complexity compared to simpler models. This aspect would require further analysis and optimization for practical implementation in real-time applications.", "questions": "(1) In Figure 1, the authors posit that a combined time and frequency domain perspective provides a more comprehensive view of data shifts. However, the provided heatmaps do not clearly support this claim. For instance, in Figure 1(a), the time-domain heatmap highlights significant MMD values for patch #10, while the actual drift begins at patch #9. Similarly, the frequency-domain heatmap includes patches #7 and #8 in the drift region, which is inaccurate. A similar discrepancy is observed in Figure 1(b), where the differences between patches 5, 6, 7, 8, and subsequent patches are not readily apparent. Consequently, this figure does not provide compelling evidence for the superiority of the combined domain approach in detecting shifts.\n\n(2) Based on Figure 1, it appears that the sudden drift is detected due to the data's relatively consistent cyclical patterns across patches. How would the method perform if the data lacks such clear cyclical properties? Would the entire sequence of patches be incorrectly identified as a drift, even if the underlying data distribution changes gradually?\n\n(3) Although this research draws comparisons to traditional time-domain and frequency-domain techniques, a more rigorous evaluation would involve comparisons with MoE-based methods. A list of relevant works is provided below.\n\n* Dish-ts: a general paradigm for alleviating distribution shift in time series forecasting. (AAAI-23) .\n* Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift. (ICLR 2022).\n* OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling. (NeurIPS 2023).\n* Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift. (KDD '24).\n* Addressing Distribution Shift in Time Series Forecasting with Instance Normalization Flows. arXiv preprint arXiv:2401.16777.\n* MixMamba: Time series modeling with adaptive expertise. (Inf. Fusion 2024).\n* Mixture-of-Linear-Experts for Long-term Time Series Forecasting. (AISTATS 2024).\n\n(4) Please elucidate the specific contributions of the PI the gate network to TFPS. Why both components are essential for the system’s performance would be beneficial. Furthermore, conducting an ablation study, which involves comparing the model’s performance with and without the separate PI module, would provide empirical validation of its significance.\n\n(5) A rigorous computational complexity analysis would provide valuable insights into the trade-offs between model complexity and performance gains.\n\n(6) I suggest that authors: Provide clear definitions for q and d in D. Explain the rationale behind controlling column size using Equation (3). Clarify the purpose of the second constraint in Equation (4). Include a theoretical justification for how these constraints improve representation learning. Conduct an ablation study comparing model performance with and without these constraints.\n\n(7) As noted in Shazer et al.’s work on Mixture-of-Experts, gating networks often exhibit a tendency to select the same k experts for routing patches. To mitigate this issue and encourage diversity among the expert modules, regularization and load balancing techniques are typically employed. However, it remains unclear how the proposed model addresses this specific challenge and promotes diversity among its expert modules.\n\n(8) The datasets employed in this study are relatively low-dimensional. It is recommended to conduct experiments on high-dimensional time series datasets, such as the Traffic dataset. Additionally, a discussion on potential scalability challenges and necessary modifications to the TFPS architecture for efficient handling of high-dimensional data would be valuable.\n\n(9) Given that the Patch-TST module used as encoder in this work already incorporates an ReVIN layer to mitigate distribution shifts in time series data. I am wondering if authors consider this point and how it affect the process of PI?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new architecture, Time-Frequency Pattern-Specific (TFPS), designed for time series forecasting, particularly under challenging conditions involving concept drift within the data. The key idea revolves around addressing the limitations of conventional uniform distribution modeling (UDM) by recognizing that different segments or \"patches\" within a time series exhibit distinct patterns and distributional shifts. These shifts may stem from factors such as sudden events, gradual trend changes, or different operating modes, making accurate forecasting difficult for models trained on a single global pattern. TFPS tackles this problem through a multi-faceted approach. First, it employs a Dual-Domain Encoder (DDE) that analyzes both time and frequency domain features, enabling a richer understanding of temporal dynamics and their potential shifts. Second, a Pattern Identifier (PI) utilizes subspace clustering to dynamically group patches exhibiting similar patterns. Finally, a Mixture of Pattern Experts (MoPE) leverages multiple specialized expert models, each trained on a specific pattern identified by the PI. This allows TFPS to adapt its predictions based on the identified pattern of a given patch, leading to improved accuracy.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "(1) Utilizing both time and frequency domain information provides a comprehensive perspective on data, capturing both trend and periodic patterns, which is essential for handling complex time series.\n(2) The use of PI and MoPE allows TFPS to handle distribution shifts by learning and adapting to specific patterns within different data segments.\n(3) Extensive experiments on real-world datasets demonstrate the effectiveness and competitiveness of TFPS compared to existing methods.", "weaknesses": "(1) The authors acknowledge that the patch length is currently determined heuristically, which might limit generalizability and performance in cases with indivisible lengths or multi-period characteristics.\n(2) The model's ability to adapt to entirely new patterns, such as those arising from unforeseen events, is not explicitly addressed and remains an open question.\n(3) While not explicitly mentioned, the use of multiple expert models and the dual-domain encoding approach could increase computational complexity compared to simpler models. This aspect would require further analysis and optimization for practical implementation in real-time applications.", "questions": "(1) In Figure 1, the authors posit that a combined time and frequency domain perspective provides a more comprehensive view of data shifts. However, the provided heatmaps do not clearly support this claim. For instance, in Figure 1(a), the time-domain heatmap highlights significant MMD values for patch #10, while the actual drift begins at patch #9. Similarly, the frequency-domain heatmap includes patches #7 and #8 in the drift region, which is inaccurate. A similar discrepancy is observed in Figure 1(b), where the differences between patches 5, 6, 7, 8, and subsequent patches are not readily apparent. Consequently, this figure does not provide compelling evidence for the superiority of the combined domain approach in detecting shifts.\n\n(2) Based on Figure 1, it appears that the sudden drift is detected due to the data's relatively consistent cyclical patterns across patches. How would the method perform if the data lacks such clear cyclical properties? Would the entire sequence of patches be incorrectly identified as a drift, even if the underlying data distribution changes gradually?\n\n(3) Although this research draws comparisons to traditional time-domain and frequency-domain techniques, a more rigorous evaluation would involve comparisons with MoE-based methods. A list of relevant works is provided below.\n\n* Dish-ts: a general paradigm for alleviating distribution shift in time series forecasting. (AAAI-23) .\n* Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift. (ICLR 2022).\n* OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling. (NeurIPS 2023).\n* Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift. (KDD '24).\n* Addressing Distribution Shift in Time Series Forecasting with Instance Normalization Flows. arXiv preprint arXiv:2401.16777.\n* MixMamba: Time series modeling with adaptive expertise. (Inf. Fusion 2024).\n* Mixture-of-Linear-Experts for Long-term Time Series Forecasting. (AISTATS 2024).\n\n(4) Please elucidate the specific contributions of the PI the gate network to TFPS. Why both components are essential for the system’s performance would be beneficial. Furthermore, conducting an ablation study, which involves comparing the model’s performance with and without the separate PI module, would provide empirical validation of its significance.\n\n(5) A rigorous computational complexity analysis would provide valuable insights into the trade-offs between model complexity and performance gains.\n\n(6) I suggest that authors: Provide clear definitions for q and d in D. Explain the rationale behind controlling column size using Equation (3). Clarify the purpose of the second constraint in Equation (4). Include a theoretical justification for how these constraints improve representation learning. Conduct an ablation study comparing model performance with and without these constraints.\n\n(7) As noted in Shazer et al.’s work on Mixture-of-Experts, gating networks often exhibit a tendency to select the same k experts for routing patches. To mitigate this issue and encourage diversity among the expert modules, regularization and load balancing techniques are typically employed. However, it remains unclear how the proposed model addresses this specific challenge and promotes diversity among its expert modules.\n\n(8) The datasets employed in this study are relatively low-dimensional. It is recommended to conduct experiments on high-dimensional time series datasets, such as the Traffic dataset. Additionally, a discussion on potential scalability challenges and necessary modifications to the TFPS architecture for efficient handling of high-dimensional data would be valuable.\n\n(9) Given that the Patch-TST module used as encoder in this work already incorporates an ReVIN layer to mitigate distribution shifts in time series data. I am wondering if authors consider this point and how it affect the process of PI?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730525091197}], "openreview_url": "https://openreview.net/forum?id=qVyjN01x4P", "arxiv_id": "2410.09836", "paper_pdf": "papers/qVyjN01x4P.pdf", "paper_pdf_sha256": "92a2f1308f3cdb35e6581c53679d0b2b31f6940ae6780b14873147c3e5ee8ee4", "paper_pdf_bytes": 2627191, "paper_pdf_source": "openreview", "code_url": "https://github.com/syrGitHub/TFPS", "code_repository": "syrGitHub/TFPS", "code_commit": "83a11827e27e6617e8c8a8771f0a1dd7e10976a5", "code_archive": "repos/qVyjN01x4P.zip", "code_archive_sha256": "50228f3260690aae63dbbcdc8dc12a2ba011f5b924e2e529078267495c42086b", "code_archive_bytes": 428277, "code_file_count": 37, "code_extensions": {".py": 33, ".sh": 4}, "github_disk_usage_kb": 437, "github_languages": {"Python": 207122, "Shell": 21175}, "github_archived": false, "github_pushed_at": "2024-11-08T02:58:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-pattern-specific-experts-for-time"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "sOHVDPqoUJ", "year": 2024, "status": "rejected", "title": "Less is More: Selective Layer Finetuning with SubTuning", "authors": ["Gal Kaplun", "Andrey Gurevich", "Tal Swisa", "Mazor David", "Shai Shalev-Shwartz", "eran malach"], "authorids": ["~Gal_Kaplun1", "~Andrey_Gurevich1", "~Tal_Swisa1", "~Mazor_David1", "~Shai_Shalev-Shwartz1", "~eran_malach1"], "authors_source": "OpenReview API", "abstract": "Finetuning a pretrained model has become the standard approach for training neural networks on novel tasks, leading to rapid convergence and enhanced performance. In this work, we present a parameter-efficient finetuning method, wherein we selectively train a carefully chosen subset of layers while keeping the remaining weights frozen at their initial (pre-trained) values. We observe that not all layers are created equal: different layers across the network contribute variably to the overall performance, and the optimal choice of layers is contingent upon the downstream task and the underlying data distribution. We demonstrate that our proposed method, termed *subset finetuning* (or SubTuning), offers several advantages over conventional finetuning. We show that SubTuning outperforms both finetuning and linear probing in scenarios with scarce or corrupted data, achieving state-of-the-art results compared to competing methods for finetuning on small datasets. When data is abundant, SubTuning often attains performance comparable to finetuning while simultaneously enabling efficient inference in a multi-task setting when deployed alongside other models. We showcase the efficacy of SubTuning across various tasks, diverse network architectures and pre-training methods.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "DR0mUuMly8", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3723/Reviewer_NuCG"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The main contribution of the paper is the observation that finetuning only a subset of layers can outperform finetuning all the parameters. This subset of layers is selected using a greedy algorithm called SubTuning which relies on finetuning profiles of different layers to greedily pick the best performing layer for a downstream task. The paper shows how SubTuning can be particularly helpful when training data is scarce. The paper then shows how SubTuning, despite being expensive during training, can lead to efficient inference especially in the multi-task setting.", "review_text": "The main contribution of the paper is the observation that finetuning only a subset of layers can outperform finetuning all the parameters. This subset of layers is selected using a greedy algorithm called SubTuning which relies on finetuning profiles of different layers to greedily pick the best performing layer for a downstream task. The paper shows how SubTuning can be particularly helpful when training data is scarce. The paper then shows how SubTuning, despite being expensive during training, can lead to efficient inference especially in the multi-task setting.", "strengths": "1. The proposed method can be applied in conjunction with other parameter-efficient methods such as LoRA and Head2Toe.\n2. The method is simple and can be broadly applied to any architecture / domain.\n3. Some observations in Section 2 are quite interesting, for example how the finetuning profiles change for different downstream tasks and for different pre-training methods (even for the same architecture) is quite intriguing to me. This indicates that the kind of features learned by different pre-training methods can be very different and can have vaying levels of usefulness even for the same downstream task. \n4. The results on corrections and distribution shifts are quite promising.\n4. The initial observations about efficiency at inference time in Section 4 are also interesting (though I have some concerns, see weaknesses below).", "weaknesses": "1. I have some concerns about the contributions of this paper -- the key idea that we do not need to finetune all layers and that finetuning only a subset of the parameters can outperform full-feature finetuning was already observed in the \"Surgical finetuning\" paper of Lee et al. To me, the only new thing shown in this paper (beyond what was already shown in Lee et al), is the fact that we can greedily choose more than just one block of layers to finetune. In Related Work, the authors mention \"While our work shares some similarities with Lee et al., the motivations and experimental settings are fundamentally different\" -- I'm a little confused by this sentence. As I understand, the experimental settings are exactly the same, and this is also mentioned in the paper later in the data corruption section where the authors say \"...we follow the setup proposed in Lee et al\". In related work the authors also mention \"..we delve deeper into the complex interaction between the appropriate layers to finetune and the downstream task, pretraining objective, and model architecture, and observe that a more nuanced viewpoint is required\" -- I'm again very confused by this, what exactly is the nuance and how does it add to what we already know from Lee et al? It'd be very useful if the authors can precisely state their contributions and mention what they add beyond the insights in Lee et al.\n\n2. I like the observations in Section 2 and would've liked a more thorough analysis of how finetuning profiles change with architecture, pretraining, and downstream tasks. Right now I only see (from the first row in Fig 2) how things change on CIFAR10 with different architectures. But then again I notice that all these architectures don't have the same pretraining (ViT is on ImageNet-21k, Resnet18 is ImageNet-1k and ResNet50 is pre-trained using DINO). It might make sense to do this more systematically, one where you hold the pretraining the same and vary architectures and downstream tasks. One where you hold an architecture constant and show different pre-training methods. If this is one part of the \"nuance\" you add in your paper beyond Lee et al's work, then it requires a proper investigation.\n\n3. Under \"The Finetuning Profile\" in Section 2, it might help to precisely define what you're profiling for (ie which accuracy are you reporting), only by reading until Section 3.1 I realized that this is cross-validation accuracy.\n\n4. The observation in Section 2 (\"Not All Layers Are Created Equal\") has been made in prior work as well [1], it's worth citing this paper.\n\n5. In Section 2 where you discuss \"The Finetuning Profile\", you've already made the assumption that the right unit at which to finetune is a block of layers, can you elaborate on why this was done? Why not layers within a block or even portions of layers?\n\n6. Under \"Results\" in Section 2 you mention that it's surprising that layers with fewer parameters can outperform layers with more parameters. However, I do not quite understand why # parameters are more relevant than the quality of features? In fact, your own results show that B16 and B8 when finetuned together (Fig 3) outperform the case when either of these blocks is finetuned alone, indicating parameter count probably has nothing to do with the finetuning profile, and that what (probably) matters more is the nature of features being captures by a block. Am I missing something?\n\n7. I'm not sure what the theoretical analysis in Section 2.1 adds to the paper. I again do not understand why parameter count (r) matters more than the quality of features or inductive bias. For example, based on the error bound in Theorem 1, wouldn't you always choose k=1 and choose this such that you minimize r (hence you'll always choose the layer with the least number of params)? But this is obviously not true based on your empirical results. Am I interpreting the theorem incorrectly?\n\n8. Section 4 starts by discussing how SubTuning can be efficient for a multi-task setup, however, all experiments shown in Section 4.1 do not consider any multi-task setup. Additionally, section 4 starts by saying you will focus on FLOPS for compute and bits needed in memory for IO considerations. However, the results in Figure 7 show Inference Time on the x-axis. Why is this the case?\n\n9. I like where Section 4 is going but I feel like it requires a more succinct argument and more convincing experiments. For example, to me it's not clear how you will save on FLOPS -- the moment you choose layer (or block) 0 as a finetuning block in SubTuning then # FLOPS = same as the FLOPS needed with full finetuning. \n\n[1] Are All Layers Created Equal? https://arxiv.org/abs/1902.01996", "questions": "1. In Section 3.1 under VTAB-1k you mention\"...using the 1k examples split specified in the protocol\". What is this referring to? I could not find any details on the splits of these datasets.\n2. What are the stddevs in Table 1?\n3. You mention \"we use the official PyTorch ResNet50\" -- which weights are these? There are two weights in the official PyTorch repo, V1 and V2 (https://pytorch.org/vision/stable/models.html).\n4. Can you increase the size of Figs 4 and 3, they are barely legible currently.\n\nWith some more work, I think this has the potential to be a cool paper!", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The main contribution of the paper is the observation that finetuning only a subset of layers can outperform finetuning all the parameters. This subset of layers is selected using a greedy algorithm called SubTuning which relies on finetuning profiles of different layers to greedily pick the best performing layer for a downstream task. The paper shows how SubTuning can be particularly helpful when training data is scarce. The paper then shows how SubTuning, despite being expensive during training, can lead to efficient inference especially in the multi-task setting.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "1. The proposed method can be applied in conjunction with other parameter-efficient methods such as LoRA and Head2Toe.\n2. The method is simple and can be broadly applied to any architecture / domain.\n3. Some observations in Section 2 are quite interesting, for example how the finetuning profiles change for different downstream tasks and for different pre-training methods (even for the same architecture) is quite intriguing to me. This indicates that the kind of features learned by different pre-training methods can be very different and can have vaying levels of usefulness even for the same downstream task. \n4. The results on corrections and distribution shifts are quite promising.\n4. The initial observations about efficiency at inference time in Section 4 are also interesting (though I have some concerns, see weaknesses below).", "weaknesses": "1. I have some concerns about the contributions of this paper -- the key idea that we do not need to finetune all layers and that finetuning only a subset of the parameters can outperform full-feature finetuning was already observed in the \"Surgical finetuning\" paper of Lee et al. To me, the only new thing shown in this paper (beyond what was already shown in Lee et al), is the fact that we can greedily choose more than just one block of layers to finetune. In Related Work, the authors mention \"While our work shares some similarities with Lee et al., the motivations and experimental settings are fundamentally different\" -- I'm a little confused by this sentence. As I understand, the experimental settings are exactly the same, and this is also mentioned in the paper later in the data corruption section where the authors say \"...we follow the setup proposed in Lee et al\". In related work the authors also mention \"..we delve deeper into the complex interaction between the appropriate layers to finetune and the downstream task, pretraining objective, and model architecture, and observe that a more nuanced viewpoint is required\" -- I'm again very confused by this, what exactly is the nuance and how does it add to what we already know from Lee et al? It'd be very useful if the authors can precisely state their contributions and mention what they add beyond the insights in Lee et al.\n\n2. I like the observations in Section 2 and would've liked a more thorough analysis of how finetuning profiles change with architecture, pretraining, and downstream tasks. Right now I only see (from the first row in Fig 2) how things change on CIFAR10 with different architectures. But then again I notice that all these architectures don't have the same pretraining (ViT is on ImageNet-21k, Resnet18 is ImageNet-1k and ResNet50 is pre-trained using DINO). It might make sense to do this more systematically, one where you hold the pretraining the same and vary architectures and downstream tasks. One where you hold an architecture constant and show different pre-training methods. If this is one part of the \"nuance\" you add in your paper beyond Lee et al's work, then it requires a proper investigation.\n\n3. Under \"The Finetuning Profile\" in Section 2, it might help to precisely define what you're profiling for (ie which accuracy are you reporting), only by reading until Section 3.1 I realized that this is cross-validation accuracy.\n\n4. The observation in Section 2 (\"Not All Layers Are Created Equal\") has been made in prior work as well [1], it's worth citing this paper.\n\n5. In Section 2 where you discuss \"The Finetuning Profile\", you've already made the assumption that the right unit at which to finetune is a block of layers, can you elaborate on why this was done? Why not layers within a block or even portions of layers?\n\n6. Under \"Results\" in Section 2 you mention that it's surprising that layers with fewer parameters can outperform layers with more parameters. However, I do not quite understand why # parameters are more relevant than the quality of features? In fact, your own results show that B16 and B8 when finetuned together (Fig 3) outperform the case when either of these blocks is finetuned alone, indicating parameter count probably has nothing to do with the finetuning profile, and that what (probably) matters more is the nature of features being captures by a block. Am I missing something?\n\n7. I'm not sure what the theoretical analysis in Section 2.1 adds to the paper. I again do not understand why parameter count (r) matters more than the quality of features or inductive bias. For example, based on the error bound in Theorem 1, wouldn't you always choose k=1 and choose this such that you minimize r (hence you'll always choose the layer with the least number of params)? But this is obviously not true based on your empirical results. Am I interpreting the theorem incorrectly?\n\n8. Section 4 starts by discussing how SubTuning can be efficient for a multi-task setup, however, all experiments shown in Section 4.1 do not consider any multi-task setup. Additionally, section 4 starts by saying you will focus on FLOPS for compute and bits needed in memory for IO considerations. However, the results in Figure 7 show Inference Time on the x-axis. Why is this the case?\n\n9. I like where Section 4 is going but I feel like it requires a more succinct argument and more convincing experiments. For example, to me it's not clear how you will save on FLOPS -- the moment you choose layer (or block) 0 as a finetuning block in SubTuning then # FLOPS = same as the FLOPS needed with full finetuning. \n\n[1] Are All Layers Created Equal? https://arxiv.org/abs/1902.01996", "questions": "1. In Section 3.1 under VTAB-1k you mention\"...using the 1k examples split specified in the protocol\". What is this referring to? I could not find any details on the splits of these datasets.\n2. What are the stddevs in Table 1?\n3. You mention \"we use the official PyTorch ResNet50\" -- which weights are these? There are two weights in the official PyTorch repo, V1 and V2 (https://pytorch.org/vision/stable/models.html).\n4. Can you increase the size of Figs 4 and 3, they are barely legible currently.\n\nWith some more work, I think this has the potential to be a cool paper!", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698838952632}, {"id": "bLN4RMpPPg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3723/Reviewer_zhTK"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper presents a parameter-efficient fine-tuning method, referred to as SubTuning, which greedily selects and fine-tunes a subset of layers in a pretrained model. It evaluates the performance of SubTuning on several vision models (e.g., ResNet50, ViT) using datasets like CIFAR-100, Flowers102, and Caltech101. Empirical results indicate that the proposed method outperforms naive baselines (e.g., linear probing and full finetuning) in low-data/distribution-shift settings.", "review_text": "The paper presents a parameter-efficient fine-tuning method, referred to as SubTuning, which greedily selects and fine-tunes a subset of layers in a pretrained model. It evaluates the performance of SubTuning on several vision models (e.g., ResNet50, ViT) using datasets like CIFAR-100, Flowers102, and Caltech101. Empirical results indicate that the proposed method outperforms naive baselines (e.g., linear probing and full finetuning) in low-data/distribution-shift settings.", "strengths": "1. The overall problem of fine-tuning only the necessary layers to save compute is interesting.\n2. There are some valuable empirical observations made by the paper, such as the fine-tuning profiles and the layer importance of different network architectures.\n3. The presentation is clear, and the idea is easy to follow. The visualization also helps understanding the method's performance.", "weaknesses": "1. **Lack of novelty:** Performing layer selection during fine-tuning has been explored by previous work [1, 2, 3]. These works use more advanced selection techniques like policy networks or genetic algorithm. The proposed way of measuring layer importance through simply fine-tuning accuracy on a specific dataset is also not generalizable, i.e., the observed trend only holds for a specific model and a specific dataset and it is hard to deduce any more general and useful rules from that. \n\n2. **Related work:** The paper briefly mentions related work in parameter-efficient transfer learning and multitask learning. However, the idea of layer selection is also very relevant to layer-wise adaptive learning rate [4] (since a learning rate of 0 is the same as freezing the layers) and neural architecture search for parameter-efficient transfer learning [5]. The idea of layer importance is relevant to the neural network pruning literature as well [6]. A more thorough discussion and comparison with these methods would provide valuable context and insights.\n\n3. **Limited models and datasets:** The paper only evaluates a few variants of ResNet and ViT, which is relatively small in scale compared to state-of-the-art large vision models. It is not sure whether the proposed method can scale up to these more practical models. Also, only 4 relatively simple datasets are used for evaluation (CIFAR, Flowers102, Caltech101, DMLAB). It would be better to see how the proposed method performs on more datasets with larger distribution shifts.\n\n[1] SpotTune: Transfer Learning through Adaptive Fine-tuning. Guo et al. 2019.\n\n[2] Dynamic fine-tuning layer selection using Kullback–Leibler divergence. Wanjiku et al. 2022.\n\n[3] Automatic layer selection for transfer learning and quantitative evaluation of layer effectiveness. Nagae et al. 2022.\n\n[4] Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. Ko et al. 2022.\n\n[5] Neural Architecture Search for Parameter-Efficient Fine-tuning of Large Pre-trained Language Models. Lawton et al. 2023.\n\n[6] Layer-adaptive Sparsity for the Magnitude-based Pruning. Lee et al. 2021.", "questions": "1. In Table 1, does the \"FT\" baseline means full fine-tuning? If so, why is the performance so bad (even worse than linear probing)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper presents a parameter-efficient fine-tuning method, referred to as SubTuning, which greedily selects and fine-tunes a subset of layers in a pretrained model. It evaluates the performance of SubTuning on several vision models (e.g., ResNet50, ViT) using datasets like CIFAR-100, Flowers102, and Caltech101. Empirical results indicate that the proposed method outperforms naive baselines (e.g., linear probing and full finetuning) in low-data/distribution-shift settings.", "soundness": "2 fair", "presentation": "3 good", "contribution": "1 poor", "strengths": "1. The overall problem of fine-tuning only the necessary layers to save compute is interesting.\n2. There are some valuable empirical observations made by the paper, such as the fine-tuning profiles and the layer importance of different network architectures.\n3. The presentation is clear, and the idea is easy to follow. The visualization also helps understanding the method's performance.", "weaknesses": "1. **Lack of novelty:** Performing layer selection during fine-tuning has been explored by previous work [1, 2, 3]. These works use more advanced selection techniques like policy networks or genetic algorithm. The proposed way of measuring layer importance through simply fine-tuning accuracy on a specific dataset is also not generalizable, i.e., the observed trend only holds for a specific model and a specific dataset and it is hard to deduce any more general and useful rules from that. \n\n2. **Related work:** The paper briefly mentions related work in parameter-efficient transfer learning and multitask learning. However, the idea of layer selection is also very relevant to layer-wise adaptive learning rate [4] (since a learning rate of 0 is the same as freezing the layers) and neural architecture search for parameter-efficient transfer learning [5]. The idea of layer importance is relevant to the neural network pruning literature as well [6]. A more thorough discussion and comparison with these methods would provide valuable context and insights.\n\n3. **Limited models and datasets:** The paper only evaluates a few variants of ResNet and ViT, which is relatively small in scale compared to state-of-the-art large vision models. It is not sure whether the proposed method can scale up to these more practical models. Also, only 4 relatively simple datasets are used for evaluation (CIFAR, Flowers102, Caltech101, DMLAB). It would be better to see how the proposed method performs on more datasets with larger distribution shifts.\n\n[1] SpotTune: Transfer Learning through Adaptive Fine-tuning. Guo et al. 2019.\n\n[2] Dynamic fine-tuning layer selection using Kullback–Leibler divergence. Wanjiku et al. 2022.\n\n[3] Automatic layer selection for transfer learning and quantitative evaluation of layer effectiveness. Nagae et al. 2022.\n\n[4] Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. Ko et al. 2022.\n\n[5] Neural Architecture Search for Parameter-Efficient Fine-tuning of Large Pre-trained Language Models. Lawton et al. 2023.\n\n[6] Layer-adaptive Sparsity for the Magnitude-based Pruning. Lee et al. 2021.", "questions": "1. In Table 1, does the \"FT\" baseline means full fine-tuning? If so, why is the performance so bad (even worse than linear probing)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698605950304}, {"id": "LwXG8VVwOW", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3723/Reviewer_nACT"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper proposes a new parameter-efficient fine-tuning method, i.e., subtuning to improve the model’s transfer learning performance. The proposed method is experimentally proven to perform better than other fine-tuning baselines. The paper is easy to follow, and the experiments are diverse and clear enough to show the advantages of the proposed method. However, it is a little hard for me to understand why this method works based on the context of the current version. Hence I tend to give a 5 at this stage, and I am happy to increase my evaluation in the rebuttal phase. Because I believe the method proposed in this paper is simple and effective in many practical cases.", "review_text": "This paper proposes a new parameter-efficient fine-tuning method, i.e., subtuning to improve the model’s transfer learning performance. The proposed method is experimentally proven to perform better than other fine-tuning baselines. The paper is easy to follow, and the experiments are diverse and clear enough to show the advantages of the proposed method. However, it is a little hard for me to understand why this method works based on the context of the current version. Hence I tend to give a 5 at this stage, and I am happy to increase my evaluation in the rebuttal phase. Because I believe the method proposed in this paper is simple and effective in many practical cases.", "strengths": "1. The method is simple and easy to understand. The performance gain is good.\n2. The concept of “finetuning profiles”, which demonstrates the contribution of different layers to the final results, is interesting. (But I believe only showing them is not enough. Explaining why the profiles behave like this under specific scenarios can make the paper stronger.)\n3. Abundant experimental results from different perspectives.", "weaknesses": "1. The theoretical guarantee and explanations are not enough. Although theorem 1 provides a generalization bound that depends on the number of stunned parameters, it is hard to link this theory to the algorithm studied in this paper. Maybe the analysis provided in [1, 2] (the analysis on an overparameterized model) would be helpful.\n2. The results in Table 1 are based on VTAB-1k, which means the samples used for transfer learning are only 1k. Will the proposed method still work when fine-tuning using the full CIFAR100 dataset? AFAIK, fine-tuning a pretrained ResNet50 on full CIFAR100 should achieve roughly 70%+ accuracy, which is much higher than the numbers provided in the paper. It is easier to compare with other baselines using similar settings.\n3. It would be very helpful if the paper could briefly introduce how the VTAB-1k dataset is generated (maybe in the appendix).\n4. In Section 4 and Figure 6, the paper claims that subtuning can bring benefits to multi-task learning. However, I cannot find the corresponding experimental results or analysis supporting that.", "questions": "1. The paper compares the proposed subtuning with the basic fine-tuning and linear probing methods, which is good. However, [1, 2] analyze why linear probing and fine-tuning alone are not the best choice for transfer learning tasks. They propose to first linear probe several epochs and then finetune the whole network. I believe this method should also be considered as one of the baselines.\n2. Following the previous point, the subtuning method can also be combined with fine-tuning. Similar to the method proposed in [2], it might be interesting to see whether the performance can be further improved if we finetune the whole network together after subtuning.\n\n[1] Kumar, Ananya, et al. \"Fine-tuning can distort pretrained features and underperform out-of-distribution.\" ICLR-2022.\n\n[2] Ren, Yi, et al. \"How to prepare your task head for finetuning.\" ICLR-2023", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a new parameter-efficient fine-tuning method, i.e., subtuning to improve the model’s transfer learning performance. The proposed method is experimentally proven to perform better than other fine-tuning baselines. The paper is easy to follow, and the experiments are diverse and clear enough to show the advantages of the proposed method. However, it is a little hard for me to understand why this method works based on the context of the current version. Hence I tend to give a 5 at this stage, and I am happy to increase my evaluation in the rebuttal phase. Because I believe the method proposed in this paper is simple and effective in many practical cases.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "1. The method is simple and easy to understand. The performance gain is good.\n2. The concept of “finetuning profiles”, which demonstrates the contribution of different layers to the final results, is interesting. (But I believe only showing them is not enough. Explaining why the profiles behave like this under specific scenarios can make the paper stronger.)\n3. Abundant experimental results from different perspectives.", "weaknesses": "1. The theoretical guarantee and explanations are not enough. Although theorem 1 provides a generalization bound that depends on the number of stunned parameters, it is hard to link this theory to the algorithm studied in this paper. Maybe the analysis provided in [1, 2] (the analysis on an overparameterized model) would be helpful.\n2. The results in Table 1 are based on VTAB-1k, which means the samples used for transfer learning are only 1k. Will the proposed method still work when fine-tuning using the full CIFAR100 dataset? AFAIK, fine-tuning a pretrained ResNet50 on full CIFAR100 should achieve roughly 70%+ accuracy, which is much higher than the numbers provided in the paper. It is easier to compare with other baselines using similar settings.\n3. It would be very helpful if the paper could briefly introduce how the VTAB-1k dataset is generated (maybe in the appendix).\n4. In Section 4 and Figure 6, the paper claims that subtuning can bring benefits to multi-task learning. However, I cannot find the corresponding experimental results or analysis supporting that.", "questions": "1. The paper compares the proposed subtuning with the basic fine-tuning and linear probing methods, which is good. However, [1, 2] analyze why linear probing and fine-tuning alone are not the best choice for transfer learning tasks. They propose to first linear probe several epochs and then finetune the whole network. I believe this method should also be considered as one of the baselines.\n2. Following the previous point, the subtuning method can also be combined with fine-tuning. Similar to the method proposed in [2], it might be interesting to see whether the performance can be further improved if we finetune the whole network together after subtuning.\n\n[1] Kumar, Ananya, et al. \"Fine-tuning can distort pretrained features and underperform out-of-distribution.\" ICLR-2022.\n\n[2] Ren, Yi, et al. \"How to prepare your task head for finetuning.\" ICLR-2023", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698557709942}, {"id": "booCIkKkkB", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission3723/Reviewer_rTPG"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper works with parameter efficient finetuning/transfer learning. Specifically, given a model pretrained on some source task, and later to be adapted for some downstream task, the paper discusses the phenomenon that selectively fine-tuning a subset of layers often does better than fine-tuning the linear head or the entire network, specially if there is data scarcity. Moreover, this subset of layers is not predetermined, and rather depends on what the distribution shift between source and downstream task is. The authors introduce a greedy algorithm of choosing the best subset of layers for a particular task, and shows good performance on a variety of vision datasets.", "review_text": "The paper works with parameter efficient finetuning/transfer learning. Specifically, given a model pretrained on some source task, and later to be adapted for some downstream task, the paper discusses the phenomenon that selectively fine-tuning a subset of layers often does better than fine-tuning the linear head or the entire network, specially if there is data scarcity. Moreover, this subset of layers is not predetermined, and rather depends on what the distribution shift between source and downstream task is. The authors introduce a greedy algorithm of choosing the best subset of layers for a particular task, and shows good performance on a variety of vision datasets.", "strengths": "1. The paper’s studied problem is well-motivated.\n2. I liked the idea of the greedy algorithm they use to choose the subset of layers to fine-tune: I found this algorithm to be simple yet effective. I also thought the algorithm is novel.\n3. The paper is well-organized and easy to follow, and have comprehensive experiments.", "weaknesses": "**Weakness 1 (lack of discussing prior work)**\n\nThe paper does not mention a lot other prior work that looked into fine-tuning particular layers for specific purposes. For example:\n1. DFR [1] is an important work that studies retraining only the last layer to get rid of spurious correlations.\n2. The paper discussed linear probing and fine-tuning, but does not discuss LP-FT [2]. There is a complex dynamic behind LP and FT, specially when distribution shift is involved. It is also important to consider pre-trained feature distortion, which can be a reason sub-tuning works better than full fine-tuning.\n\nFinally, the paper’s claim/observation in the abstract is not novel, i.e., it has been recorded in other papers before [7], and should be modified as such.\n\n> **This paper (abstract)**: We observe that not all layers are created equal: different layers across the network contribute variably to the overall performance, and the optimal choice of layers is contingent upon the downstream task and the underlying data distribution\n\n> **Surgical fine-tuning [7] (abstract)**: This paper shows that in such settings, selectively fine-tuning a subset of layers (which we term surgical fine-tuning) matches or outperforms commonly used fine-tuning approaches. Moreover, the type of distribution shift influences which subset is more effective to tune.\n\n\n**Weakness 2 (formal definition/terminology):**\n\nSetting up a formal definition of the problem would make the paper much easier to read. For example, one could define it as: let $S$ be the set of all layers, then sub-tuning is:\n$$S_{best} = argmin_{S' \\subseteq S} L(f, S', D_{ft}, D_{val})$$\n\nwhere $L(f, S', D_{ft}, D_{val})$ is the loss after fine-tuning the network $f$'s layers in $S'$ on dataset $D_{ft}$ and evaluating the final performance on $D_{val}$.\n\nWith this in mind, how is sub-tuning different than surgical fine-tuning? I think the main difference is not in the formal definition, but in the practical implementation: surgical fine-tuning only considers $S' \\subseteq S$ that are contiguous, whereas this paper considers $S'$ that are not necessarily contiguous. If this is the case, this should be proposed as an improved/better way of doing surgical fine-tuning, not a separate problem definition. If not, please explain.\n\nAlso a proper explanation of what authors consider as blocks vs layers in a neural network would be important. Ideally, some sort of schematic of ResNet-50, discussing the terminology, in the main paper/appendix would be needed.\n\n**Weakness 3 (experiment results):**\n1. **Figure 1**: The difference between fine-tuning **B1** (94%) vs **B14** (95.5%) in figure 1 (left) is negligible. Also denoting the number of parameters in each of these layers would be important.\n2. **Figure 3**: I would interpret the results in a different way: looking at the grid column wise, it seems the columns corresponding to B13, B14, B15 do well in general. I would also guess that the variation of accuracy within these columns is small when adding B2, B4, B8 or B9. How much is adding the second block possibly helping here? It can quite possibly be that fine-tuning B13/14/15 already recovers most of the performance/does most of the heavy lifting, and the improvement that comes from fine-tuning B2 in addition to B14 is negligible.\n3. **See question 3.** I seriously doubt that in the limit of infinite data, subtuning essentially chooses close to all layers, and therefore becomes full fine-tuning effectively. If that is not the case, I am curious why it is not? Given infinite data, and more layers to fine-tune, the neural network should intuitively learn a better classifier, unless only a few layers can solve the problem effectively/the layers not fine-tuned are already close to optimum for the downstream task. For $\\textrm{source} \\rightarrow \\textrm{target}$ dataset with large distribution shift, this should never be the case, so I suspect this result is because of $\\textrm{Imagenet} \\rightarrow \\textrm{CIFAR-10}$ shift. If this is not the case, please explain. \n4. **Table 1**: Missing some very important comparisons, like LP-FT [2], layer-wise learning rate tuning [6], gradual unfreezing [3, 4, 5].\n5. **Section 3.2**: Missing results on some very important benchmark. The section title contains distribution shift, but focuses solely on image corruptions. Some results on natural shifts such as the WILDS [8] (i.e., at least 2 or 3 of these datasets) would strengthen the paper.\n6. **Section 3.2**: “Furthermore, our results challenge the claim made in Lee et al. (2022b) that suggests adjusting only the first layers of the network suffices for input-level shifts in CIFAR-10-C. Interestingly, we found that the ultimate or penultimate block was the first layer selected for all corruptions, resulting in the largest performance increase.” I feel this claim is not well made. For example, even in the reproduced results of this paper, we see surgical L1 > surgical L2 > surgical L3. Note that sub-tuning looks at a much more granular level (i.e., within blocks of a single layer), and finds a better finetuning arrangement. Also, it is possible that the finetuning the last layer performs worse, but the last block within the last layer performs better, hence this is not a contradiction. Furthermore, in figure 5, B12 is the most commonly chosen first block, but B1 is the most commonly chosen second block, so the first block is still important? In what order would sub-tuning pick the parameters, if allowed only to fine-tune layers and not the blocks inside? \n\n**Weakness 4 (lack of interpretability):**\n\nThe authors give a mechanistic way of choosing the subset of layers that needs to be fine-tuned, but there is no understanding of the layers chosen other than it empirically works well. Why would, for a particular downstream task, layers B8 and B14 combine to do the best? Possible things to explore here:\n1. How much does adding the second layer help? How much does adding the third layer? And so on.\n2. Possibly looking at something like [9]. It is hard to do for vision experiments, but understanding why certain combinations of layers help is important.\n\nOtherwise, this method is computationally heavy (requires multiple fine-tuning rounds) and may not be suitable for use in practice compared to full fine-tuning, LP-FT or surgical fine-tuning for their simplicity.", "questions": "1. **Finetuning profiles (figure 1 and figure 2)**: how many data points are used to fine-tune each of the layers?\n2. **Page 5, greedy selection**: “We note that such greedy optimization is a common approach for subset selection in various combinatorial problems, and is known to approximate the optimal solution under certain assumptions.” Explain the assumptions, and/or add references for this.\n3. **Page 5, greedy selection**: “We show that SubTuning results in comparable performance to full finetuning even for full datasets”. Does this hold only for CIFAR-10? Or does this hold for other datasets as well? Also how many layers are being fine-tuned when the full-dataset is used? I.e., does subtuning essentially become full fine-tuning in the limit of the full dataset?\n4. **Table 1**: Why would surgical fine-tuning [7] not be applicable for ViT-B/16 architecture? Table 1 in [7] shows surgical fine-tuning using different attention blocks on a ViT-B/16 architecture. One doesn’t need to group attention blocks into layers, one can just choose one/two/three contiguous attention blocks and fine-tune them for surgical FT.\n5. **Figure 4**: Does the authors have any intuition for the results in figure 4? Why would sub-tuning pick earlier layers, i.e., layers closer to the input, if more data is available?\n6. **Computational cost of sub-tuning**: What is the cost of sub-tuning? The big-Oh complexity is given in the paper, but sub-tuning might occasionally finish earlier before checking all blocks. Could the authors provide a table of computational cost, either in terms of total fine-tuning epochs required, summed over checking each block set until their algorithm terminates, or FLOPs or some other measure, between just running full finetuning once, running surgical fine-tuning and sub-tuning? \n7. **Active learning**: What does classification margin mean here? Is this simply the maximum softmax/prediction probability?\n8. **Figure 6, Subtuning for MTL**: I am not entirely sure how this, specially the split in the computation graph in figure 6, works. Say for a new task, you only need to tune blocks B5 and B6. Would you initialize another copy of these blocks, B5’ and B6’, with the same pretrained weights as B5 and B6, and for the new task, you only update these copies of the weights B5’ and B6’, and leave the original blocks/weights B5 and B6 unchanged? Keeping a second copy of these blocks with original pretrained weights would prevent catastrophic forgetting. I am not entirely sure how this is done.\n9. **Siamese subtuning**: For Siamese subtuning, do you have to double the number of parameters in the classification head? I.e., consider the classification head of the original task (say 512 dimensions) and **only** the new task (say another 512 dimensions) for regular subtuning. Then for siamese subtuning, since the new task head receives both embeddings for original and new task, do you update the new task classifier head to be 512 dimensional? Also when you finetune on a new task, do you prevent gradient updates due to the original task’s embedding going to a new task’s classification head? Or do you allow the new task to change the original task’s weights?\n10. **$\\mathbf{\\epsilon}$**: What is the threshold, $\\epsilon$, in algorithm 1? This is a very important hyper-parameter, and I did not find any discussion on how this is picked. Also my guess is it is pretty small, since the performance gain keeps decreasing (in trend) as one adds more layers.\n11. **$\\mathbf{\\epsilon}$**: Could you add results on how robust the algorithm is to these chosen threshold, $\\epsilon$?\n12. **Optimizer**: What is the optimizer used for training ViT-B/16? \n\n\n[1] Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations, https://arxiv.org/abs/2204.02937\n\n[2] Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution, https://arxiv.org/abs/2202.10054\n\n[3] Universal Language Model Fine-tuning for Text Classification, https://arxiv.org/abs/1801.06146\n\n[4] Distilling BERT into Simple Neural Networks with Unlabeled Transfer Data, https://arxiv.org/abs/1910.01769\n\n[5] Targeted transfer learning to improve performance in small medical physics datasets, https://arxiv.org/abs/1912.06761\n\n[6] AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks, https://arxiv.org/abs/2002.06048\n\n[7] Surgical Fine-Tuning Improves Adaptation to Distribution Shifts, https://arxiv.org/abs/2210.11466\n\n[8] WILDS: A Benchmark of in-the-Wild Distribution Shifts, https://arxiv.org/abs/2012.07421\n\n[9] Locating and Editing Factual Associations in GPT, https://arxiv.org/abs/2202.05262", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper works with parameter efficient finetuning/transfer learning. Specifically, given a model pretrained on some source task, and later to be adapted for some downstream task, the paper discusses the phenomenon that selectively fine-tuning a subset of layers often does better than fine-tuning the linear head or the entire network, specially if there is data scarcity. Moreover, this subset of layers is not predetermined, and rather depends on what the distribution shift between source and downstream task is. The authors introduce a greedy algorithm of choosing the best subset of layers for a particular task, and shows good performance on a variety of vision datasets.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper’s studied problem is well-motivated.\n2. I liked the idea of the greedy algorithm they use to choose the subset of layers to fine-tune: I found this algorithm to be simple yet effective. I also thought the algorithm is novel.\n3. The paper is well-organized and easy to follow, and have comprehensive experiments.", "weaknesses": "**Weakness 1 (lack of discussing prior work)**\n\nThe paper does not mention a lot other prior work that looked into fine-tuning particular layers for specific purposes. For example:\n1. DFR [1] is an important work that studies retraining only the last layer to get rid of spurious correlations.\n2. The paper discussed linear probing and fine-tuning, but does not discuss LP-FT [2]. There is a complex dynamic behind LP and FT, specially when distribution shift is involved. It is also important to consider pre-trained feature distortion, which can be a reason sub-tuning works better than full fine-tuning.\n\nFinally, the paper’s claim/observation in the abstract is not novel, i.e., it has been recorded in other papers before [7], and should be modified as such.\n\n> **This paper (abstract)**: We observe that not all layers are created equal: different layers across the network contribute variably to the overall performance, and the optimal choice of layers is contingent upon the downstream task and the underlying data distribution\n\n> **Surgical fine-tuning [7] (abstract)**: This paper shows that in such settings, selectively fine-tuning a subset of layers (which we term surgical fine-tuning) matches or outperforms commonly used fine-tuning approaches. Moreover, the type of distribution shift influences which subset is more effective to tune.\n\n\n**Weakness 2 (formal definition/terminology):**\n\nSetting up a formal definition of the problem would make the paper much easier to read. For example, one could define it as: let $S$ be the set of all layers, then sub-tuning is:\n$$S_{best} = argmin_{S' \\subseteq S} L(f, S', D_{ft}, D_{val})$$\n\nwhere $L(f, S', D_{ft}, D_{val})$ is the loss after fine-tuning the network $f$'s layers in $S'$ on dataset $D_{ft}$ and evaluating the final performance on $D_{val}$.\n\nWith this in mind, how is sub-tuning different than surgical fine-tuning? I think the main difference is not in the formal definition, but in the practical implementation: surgical fine-tuning only considers $S' \\subseteq S$ that are contiguous, whereas this paper considers $S'$ that are not necessarily contiguous. If this is the case, this should be proposed as an improved/better way of doing surgical fine-tuning, not a separate problem definition. If not, please explain.\n\nAlso a proper explanation of what authors consider as blocks vs layers in a neural network would be important. Ideally, some sort of schematic of ResNet-50, discussing the terminology, in the main paper/appendix would be needed.\n\n**Weakness 3 (experiment results):**\n1. **Figure 1**: The difference between fine-tuning **B1** (94%) vs **B14** (95.5%) in figure 1 (left) is negligible. Also denoting the number of parameters in each of these layers would be important.\n2. **Figure 3**: I would interpret the results in a different way: looking at the grid column wise, it seems the columns corresponding to B13, B14, B15 do well in general. I would also guess that the variation of accuracy within these columns is small when adding B2, B4, B8 or B9. How much is adding the second block possibly helping here? It can quite possibly be that fine-tuning B13/14/15 already recovers most of the performance/does most of the heavy lifting, and the improvement that comes from fine-tuning B2 in addition to B14 is negligible.\n3. **See question 3.** I seriously doubt that in the limit of infinite data, subtuning essentially chooses close to all layers, and therefore becomes full fine-tuning effectively. If that is not the case, I am curious why it is not? Given infinite data, and more layers to fine-tune, the neural network should intuitively learn a better classifier, unless only a few layers can solve the problem effectively/the layers not fine-tuned are already close to optimum for the downstream task. For $\\textrm{source} \\rightarrow \\textrm{target}$ dataset with large distribution shift, this should never be the case, so I suspect this result is because of $\\textrm{Imagenet} \\rightarrow \\textrm{CIFAR-10}$ shift. If this is not the case, please explain. \n4. **Table 1**: Missing some very important comparisons, like LP-FT [2], layer-wise learning rate tuning [6], gradual unfreezing [3, 4, 5].\n5. **Section 3.2**: Missing results on some very important benchmark. The section title contains distribution shift, but focuses solely on image corruptions. Some results on natural shifts such as the WILDS [8] (i.e., at least 2 or 3 of these datasets) would strengthen the paper.\n6. **Section 3.2**: “Furthermore, our results challenge the claim made in Lee et al. (2022b) that suggests adjusting only the first layers of the network suffices for input-level shifts in CIFAR-10-C. Interestingly, we found that the ultimate or penultimate block was the first layer selected for all corruptions, resulting in the largest performance increase.” I feel this claim is not well made. For example, even in the reproduced results of this paper, we see surgical L1 > surgical L2 > surgical L3. Note that sub-tuning looks at a much more granular level (i.e., within blocks of a single layer), and finds a better finetuning arrangement. Also, it is possible that the finetuning the last layer performs worse, but the last block within the last layer performs better, hence this is not a contradiction. Furthermore, in figure 5, B12 is the most commonly chosen first block, but B1 is the most commonly chosen second block, so the first block is still important? In what order would sub-tuning pick the parameters, if allowed only to fine-tune layers and not the blocks inside? \n\n**Weakness 4 (lack of interpretability):**\n\nThe authors give a mechanistic way of choosing the subset of layers that needs to be fine-tuned, but there is no understanding of the layers chosen other than it empirically works well. Why would, for a particular downstream task, layers B8 and B14 combine to do the best? Possible things to explore here:\n1. How much does adding the second layer help? How much does adding the third layer? And so on.\n2. Possibly looking at something like [9]. It is hard to do for vision experiments, but understanding why certain combinations of layers help is important.\n\nOtherwise, this method is computationally heavy (requires multiple fine-tuning rounds) and may not be suitable for use in practice compared to full fine-tuning, LP-FT or surgical fine-tuning for their simplicity.", "questions": "1. **Finetuning profiles (figure 1 and figure 2)**: how many data points are used to fine-tune each of the layers?\n2. **Page 5, greedy selection**: “We note that such greedy optimization is a common approach for subset selection in various combinatorial problems, and is known to approximate the optimal solution under certain assumptions.” Explain the assumptions, and/or add references for this.\n3. **Page 5, greedy selection**: “We show that SubTuning results in comparable performance to full finetuning even for full datasets”. Does this hold only for CIFAR-10? Or does this hold for other datasets as well? Also how many layers are being fine-tuned when the full-dataset is used? I.e., does subtuning essentially become full fine-tuning in the limit of the full dataset?\n4. **Table 1**: Why would surgical fine-tuning [7] not be applicable for ViT-B/16 architecture? Table 1 in [7] shows surgical fine-tuning using different attention blocks on a ViT-B/16 architecture. One doesn’t need to group attention blocks into layers, one can just choose one/two/three contiguous attention blocks and fine-tune them for surgical FT.\n5. **Figure 4**: Does the authors have any intuition for the results in figure 4? Why would sub-tuning pick earlier layers, i.e., layers closer to the input, if more data is available?\n6. **Computational cost of sub-tuning**: What is the cost of sub-tuning? The big-Oh complexity is given in the paper, but sub-tuning might occasionally finish earlier before checking all blocks. Could the authors provide a table of computational cost, either in terms of total fine-tuning epochs required, summed over checking each block set until their algorithm terminates, or FLOPs or some other measure, between just running full finetuning once, running surgical fine-tuning and sub-tuning? \n7. **Active learning**: What does classification margin mean here? Is this simply the maximum softmax/prediction probability?\n8. **Figure 6, Subtuning for MTL**: I am not entirely sure how this, specially the split in the computation graph in figure 6, works. Say for a new task, you only need to tune blocks B5 and B6. Would you initialize another copy of these blocks, B5’ and B6’, with the same pretrained weights as B5 and B6, and for the new task, you only update these copies of the weights B5’ and B6’, and leave the original blocks/weights B5 and B6 unchanged? Keeping a second copy of these blocks with original pretrained weights would prevent catastrophic forgetting. I am not entirely sure how this is done.\n9. **Siamese subtuning**: For Siamese subtuning, do you have to double the number of parameters in the classification head? I.e., consider the classification head of the original task (say 512 dimensions) and **only** the new task (say another 512 dimensions) for regular subtuning. Then for siamese subtuning, since the new task head receives both embeddings for original and new task, do you update the new task classifier head to be 512 dimensional? Also when you finetune on a new task, do you prevent gradient updates due to the original task’s embedding going to a new task’s classification head? Or do you allow the new task to change the original task’s weights?\n10. **$\\mathbf{\\epsilon}$**: What is the threshold, $\\epsilon$, in algorithm 1? This is a very important hyper-parameter, and I did not find any discussion on how this is picked. Also my guess is it is pretty small, since the performance gain keeps decreasing (in trend) as one adds more layers.\n11. **$\\mathbf{\\epsilon}$**: Could you add results on how robust the algorithm is to these chosen threshold, $\\epsilon$?\n12. **Optimizer**: What is the optimizer used for training ViT-B/16? \n\n\n[1] Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations, https://arxiv.org/abs/2204.02937\n\n[2] Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution, https://arxiv.org/abs/2202.10054\n\n[3] Universal Language Model Fine-tuning for Text Classification, https://arxiv.org/abs/1801.06146\n\n[4] Distilling BERT into Simple Neural Networks with Unlabeled Transfer Data, https://arxiv.org/abs/1910.01769\n\n[5] Targeted transfer learning to improve performance in small medical physics datasets, https://arxiv.org/abs/1912.06761\n\n[6] AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks, https://arxiv.org/abs/2002.06048\n\n[7] Surgical Fine-Tuning Improves Adaptation to Distribution Shifts, https://arxiv.org/abs/2210.11466\n\n[8] WILDS: A Benchmark of in-the-Wild Distribution Shifts, https://arxiv.org/abs/2012.07421\n\n[9] Locating and Editing Factual Associations in GPT, https://arxiv.org/abs/2202.05262", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1697556303560}], "openreview_url": "https://openreview.net/forum?id=sOHVDPqoUJ", "arxiv_id": "2302.06354", "paper_pdf": "papers/sOHVDPqoUJ.pdf", "paper_pdf_sha256": "2959d112a8557b8fb0f273b5a40267ddc317203704e6b6013c5decc2ae0cd045", "paper_pdf_bytes": 2800232, "paper_pdf_source": "openreview", "code_url": "https://github.com/talswisa/SubTuning", "code_repository": "talswisa/SubTuning", "code_commit": "68b64c24ad0b5053d4ad2b79d0c6c459c327c5a4", "code_archive": "repos/sOHVDPqoUJ.zip", "code_archive_sha256": "822a43b27e88ac40d5b78c43d7cc274e717be48903654b6bfec7556217e104bd", "code_archive_bytes": 522020, "code_file_count": 16, "code_extensions": {".py": 12, ".sh": 4}, "github_disk_usage_kb": 519, "github_languages": {"Python": 55704, "Shell": 2884, "Dockerfile": 630}, "github_archived": false, "github_pushed_at": "2023-06-11T12:01:56Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/subtuning-efficient-finetuning-for-multi-task"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "3vOtC1t1kF", "year": 2023, "status": "rejected", "title": "Efficient Personalized Federated Learning via Sparse Model-Adaptation", "authors": ["Daoyuan Chen", "Liuyi Yao", "Dawei Gao", "Bolin Ding", "Yaliang Li"], "authorids": ["~Daoyuan_Chen1", "~Liuyi_Yao1", "~Dawei_Gao1", "~Bolin_Ding3", "~Yaliang_Li1"], "authors_source": "OpenReview API", "abstract": "Federated Learning (FL) aims to train machine learning models for multiple clients without sharing their own private data. Due to the heterogeneity of clients' local data distribution, recent studies explore the personalized FL that learns and deploys distinct local models with the help of auxiliary global models. However, the clients can be heterogeneous in terms of not only local data distribution, but also their computation and communication resources. The capacity and efficiency of personalized models are restricted by the lowest-resource clients, leading to sub-optimal performance and limited practicality of personalized FL. To overcome these challenges, we propose a novel approach named pFedGate for efficient personalized FL by adaptively and efficiently learning sparse local models. With a lightweight trainable gating layer, pFedGate enables clients to reach their full potential in model capacity by generating different sparse models accounting for both the heterogeneous data distributions and resource constraints. Meanwhile, the computation and communication efficiency are both improved thanks to the adaptability between the model sparsity and clients' resources. Further, we theoretically show that the proposed pFedGate has superior complexity with guaranteed convergence and generalization error. Extensive experiments show that pFedGate achieves superior global accuracy, individual accuracy and efficiency simultaneously over state-of-the-art methods, by up to 4.53\\% accuracy improvement and 12x smaller model size. We also demonstrate that pFedGate performs better than competitors in the novel clients participation and partial clients participation scenarios, and can learn meaningful sparse local models adapted to different data distributions.", "decision": "Reject", "meta_review": null, "num_reviews": 5, "reviews": [{"id": "9iK1ugcpwUa", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2420/Reviewer_eJXn"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "### Summary\n\nThis paper addresses data and resource heterogeneity of each client in Federated Learning. This paper proposes a gating function $g_{\\phi}(x, s_i)$ that allows easily adaptation to each client while taking sparsity into account like computation resources. As gating function generates masking parameters for local models, it can response to various model-size, heterogeneous dataset and computation resources. Especially, the authors suggest that sample-wise model is better effective than a client-wise local model in personalized FL. \n\nTechnically, the authors demonstrate that the proposed method does not hurt the previous theoretical convergence bound and generalization gap in FL. This empirically achieves the state-of-art average performance in personalized FL, ensuring bottom decile performance in a variety of experimental results like EMNIST, FEMNIST, CIFAR10 and CIFAR100. ", "review_text": "In summary, I believe the paper holds an interesting research direction worth exploring in personalized FL. However, in my opinion, the current manuscript does not meet the standards for acceptance at ICLR 2023. The authors are strongly encouraged to present the limitations of the method by empirical manners, and improve technical aspects as I do believe this is a promising idea.\n\n-------------------\n\n**After rebuttal** : I'm satisfied with the authors feedback related to experimental results. They provide several experimental evidences to support the efficiency of their model and performance improvement via this rebuttal phase. However, I believe that theoretical aspects must be revised and re-organized in the camera-ready version. Furthermore, since the authors have responded my comments after the end of discuss period, so they were not able to revise the manuscript. Under the condition that author assure all comments will be included in the camera ready version, I score this paper as 6.", "strengths": "### Strength\n\n- This paper consistently presents empirical performance improvement in various experimental results. \n- It also suggests sample-wise local models by using gating function $g_{\\phi}$, so that we do not practically need to design sophisticated manipulation for it, since this gating function can be learnt during training. It obviously leads to work in FL that disregard data and resource heterogeneity.\n- This paper offers many ablation studies of learned FL models, including ones that distinguish each client without additional training. This empirical studies appears relevant for deriving performance improvement of all cases, and supports author's claims.  \n\n### Weakness\n\n- This paper should clarify a limit of data and resource heterogeneity of FL framework. In the manuscript, their voices are likely to be misunderstood, so that the proposed method may be used to totally different devices like edge devices and high-performance cloud computation. According to my understanding, this method supposes that all devices have sufficient memory to simultaneously load global and gating function, but each device's computation capability varies.\n\n- Along with the above comment, the authors should reveal the how extents of sparsity this model can support. Particularly, they should investigate when a gating function rapidly deteriorates. Also, they need to show characteristic of representation capability between global and local models. Lastly, they should clearly reveal a limitation and boundary of this method by changing communication frequency and data distribution of each client.\n\n- In technical aspects,  theoretical reasonings in this paper somewhat mismatch their claim that emphasizes sparsity of each client. Especially, in theorem 1 and 2, they do not consider the sparsity of each client, as well as not provide any high-level understanding of this sparsity based on those theorems. Instead, through very strong assumption like strong convexity of all losses and independence with respect to global, gating functions and these parameters $\\phi_i$, it simply reveals that the proposed method does not violate the previous FL framework from point of the number of effective samples and convergence analysis.\n\n- This paper must compare of FedMask[1] in terms of masking characteristics by varying sparsity. The FedMask also employs masking parameters for each client, so it is completely compatible with the proposed method excluding sample-wise models. If the authors want to highlight practical usage of heterogeneity of data and resources, this comparison must be presented for better understanding of this paper. However, in experimental results on the manuscript and appendix, they simply describe the proposed method outperform those algorithms. ", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "### Summary\n\nThis paper addresses data and resource heterogeneity of each client in Federated Learning. This paper proposes a gating function $g_{\\phi}(x, s_i)$ that allows easily adaptation to each client while taking sparsity into account like computation resources. As gating function generates masking parameters for local models, it can response to various model-size, heterogeneous dataset and computation resources. Especially, the authors suggest that sample-wise model is better effective than a client-wise local model in personalized FL. \n\nTechnically, the authors demonstrate that the proposed method does not hurt the previous theoretical convergence bound and generalization gap in FL. This empirically achieves the state-of-art average performance in personalized FL, ensuring bottom decile performance in a variety of experimental results like EMNIST, FEMNIST, CIFAR10 and CIFAR100. ", "strength_and_weaknesses": "### Strength\n\n- This paper consistently presents empirical performance improvement in various experimental results. \n- It also suggests sample-wise local models by using gating function $g_{\\phi}$, so that we do not practically need to design sophisticated manipulation for it, since this gating function can be learnt during training. It obviously leads to work in FL that disregard data and resource heterogeneity.\n- This paper offers many ablation studies of learned FL models, including ones that distinguish each client without additional training. This empirical studies appears relevant for deriving performance improvement of all cases, and supports author's claims.  \n\n### Weakness\n\n- This paper should clarify a limit of data and resource heterogeneity of FL framework. In the manuscript, their voices are likely to be misunderstood, so that the proposed method may be used to totally different devices like edge devices and high-performance cloud computation. According to my understanding, this method supposes that all devices have sufficient memory to simultaneously load global and gating function, but each device's computation capability varies.\n\n- Along with the above comment, the authors should reveal the how extents of sparsity this model can support. Particularly, they should investigate when a gating function rapidly deteriorates. Also, they need to show characteristic of representation capability between global and local models. Lastly, they should clearly reveal a limitation and boundary of this method by changing communication frequency and data distribution of each client.\n\n- In technical aspects,  theoretical reasonings in this paper somewhat mismatch their claim that emphasizes sparsity of each client. Especially, in theorem 1 and 2, they do not consider the sparsity of each client, as well as not provide any high-level understanding of this sparsity based on those theorems. Instead, through very strong assumption like strong convexity of all losses and independence with respect to global, gating functions and these parameters $\\phi_i$, it simply reveals that the proposed method does not violate the previous FL framework from point of the number of effective samples and convergence analysis.\n\n- This paper must compare of FedMask[1] in terms of masking characteristics by varying sparsity. The FedMask also employs masking parameters for each client, so it is completely compatible with the proposed method excluding sample-wise models. If the authors want to highlight practical usage of heterogeneity of data and resources, this comparison must be presented for better understanding of this paper. However, in experimental results on the manuscript and appendix, they simply describe the proposed method outperform those algorithms. ", "clarity,_quality,_novelty_and_reproducibility": "**Presentation** : A presentation of this paper appears quite dense, so I’m concerned about violation of the ICLR2023 template.", "summary_of_the_review": "In summary, I believe the paper holds an interesting research direction worth exploring in personalized FL. However, in my opinion, the current manuscript does not meet the standards for acceptance at ICLR 2023. The authors are strongly encouraged to present the limitations of the method by empirical manners, and improve technical aspects as I do believe this is a promising idea.\n\n-------------------\n\n**After rebuttal** : I'm satisfied with the authors feedback related to experimental results. They provide several experimental evidences to support the efficiency of their model and performance improvement via this rebuttal phase. However, I believe that theoretical aspects must be revised and re-organized in the camera-ready version. Furthermore, since the authors have responded my comments after the end of discuss period, so they were not able to revise the manuscript. Under the condition that author assure all comments will be included in the camera ready version, I score this paper as 6.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667378411562}, {"id": "X30uHMeAlD0", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2420/Reviewer_gH8r"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "\n\nThe draft proposed an approach for efficient personalized FL by adaptively and efficiently learning sparse local models(a subset of the global one) with a shared global model. With a lightweight trainable gating layer, the proposed algorithm enables clients to reach their full potential in model capacity by generating different sparse models accounting for both the heterogeneous data distributions and resource constraints. ", "review_text": "Basically,  The paper is of good quality. Although the novelty is not very high since there are two existing papers that adopt similar ideas for Personalized FL, this paper provides a good theoretical analysis and the results are very promising. If the authors can add a good discussion on the novelty compared with the two papers mentioned below, I would champion the acceptance.", "strengths": "Strength:\n\nS1). The authors provided a theoretical analysis of the proposed algorithm.\n\nS2). The method is validated with sufficient experiments.\n\nS3). The results are very promising.\n\nWeakness:\n\nW1). The novelty of the paper is not very high since there are already some papers with similar ideas in FL. As claimed in Related work “our method generates different sub-models from the whole global model with a larger learnable parameter space to handle heterogeneous data distribution”, actually the following work is using a similar idea. It also adopts the block-wise selection from the global model and generates personal and sparse architecture for different clients. \na). Personalized Federated Learning via Heterogeneous Modular Networks. IEEE ICDM 2022. \nhttps://arxiv.org/pdf/2210.14830.pdf\n\nb). Personalized Neural Architecture Search for Federated Learning, Minh Hang et al. Neurips Workshop 2021.\nhttps://neurips2021workshopfl.github.io/NFFL-2021/papers/2021/Hoang2021.pdf\n\nActually, these two papers are using more elegant solutions for PFL. The proposed solution adopts the mixture method of ideas of NAS and FL. \n\n\nW2). It is strongly suggested to include the two papers above in the revised version for discussion of novelty of the proposed approach. B) is discussing the fine-grained subnetwork selection and A) is discussing block-wise subset network selection for communication efficiency that is similar as this paper.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "\n\nThe draft proposed an approach for efficient personalized FL by adaptively and efficiently learning sparse local models(a subset of the global one) with a shared global model. With a lightweight trainable gating layer, the proposed algorithm enables clients to reach their full potential in model capacity by generating different sparse models accounting for both the heterogeneous data distributions and resource constraints. ", "strength_and_weaknesses": "Strength:\n\nS1). The authors provided a theoretical analysis of the proposed algorithm.\n\nS2). The method is validated with sufficient experiments.\n\nS3). The results are very promising.\n\nWeakness:\n\nW1). The novelty of the paper is not very high since there are already some papers with similar ideas in FL. As claimed in Related work “our method generates different sub-models from the whole global model with a larger learnable parameter space to handle heterogeneous data distribution”, actually the following work is using a similar idea. It also adopts the block-wise selection from the global model and generates personal and sparse architecture for different clients. \na). Personalized Federated Learning via Heterogeneous Modular Networks. IEEE ICDM 2022. \nhttps://arxiv.org/pdf/2210.14830.pdf\n\nb). Personalized Neural Architecture Search for Federated Learning, Minh Hang et al. Neurips Workshop 2021.\nhttps://neurips2021workshopfl.github.io/NFFL-2021/papers/2021/Hoang2021.pdf\n\nActually, these two papers are using more elegant solutions for PFL. The proposed solution adopts the mixture method of ideas of NAS and FL. \n\n\nW2). It is strongly suggested to include the two papers above in the revised version for discussion of novelty of the proposed approach. B) is discussing the fine-grained subnetwork selection and A) is discussing block-wise subset network selection for communication efficiency that is similar as this paper.\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is clearly expressed with good quality.\nNovelty needs to be discussed in the two papers not cited.\nReproducibility is uncertain.", "summary_of_the_review": "Basically,  The paper is of good quality. Although the novelty is not very high since there are two existing papers that adopt similar ideas for Personalized FL, this paper provides a good theoretical analysis and the results are very promising. If the authors can add a good discussion on the novelty compared with the two papers mentioned below, I would champion the acceptance.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1667311857922}, {"id": "f9dJkvusjWF", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2420/Reviewer_sfrQ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper proposes a new framework called pFedGate for training personalized models at client devices, such that each model can have different sparsity level depending on the devices computational and memory capacities. The idea is to generate personalized masks with different sparsity levels for each device, using a small gating layer per device. The gating layer itself is small in size and hence the paper claims that it does not add much overhead. The empirical results show that this technique indeed achieves better personalization than existing methods.", "review_text": "While the proposed idea of generating personalized masks (gating weights) is novel and has potential, I have some concerns regarding the theory and experimental results that I have mentioned above.", "strengths": "Strengths:\n1. The proposed framework seems to achieve better personalization accuracy than existing models.\n2. The idea of generating personalized masks (gating weights) using a computationally cheap network is novel.\n\n\nConcerns / Questions:\n1. I think Assumption 1 might be a strong assumption, since it assumes bounded diversity between the globally optimal model, and the optimal sparse local model. In particular, I think $\\sigma_i$ might itself depend on the level of sparsity for the device.\n\n2. In the proof of proposition 1, I do not follow how the Reverse Cauchy-Schwarz inequality is applied, since the left hand side of Eq. (13) has the l2 norm squared of element-wise multiplication of vectors instead of inner product. To be precise, the left hand side will look like $(\\sum_j (M^*_{i,j}\\theta^*_{g,j})^2 )$ instead of $(\\sum_j (M^*_{i,j}\\theta^*_{g,j}) )$. Also, I think the denominator of Eq. (4) should not have the square in norm of $\\theta^*_{g}$.\n\n3. The experimental results do not provide the time (in seconds) and memory requirements (in GB / MB) for the training of the models. Further, the convergence vs communication round curve is only provided for EMNIST dataset. This data is important since one of the motivations for this paper is to perform Federated Learning efficiently on computationally slow devices.\n\n4. In Eq. (7), is the gradient w.r.t $\\phi$ as well? In other words, does Theorem 2 prove convergence for the gating layer as well?\n\n5. On page 6, section 6.1, I think the definition of $L(\\theta_g, \\phi_i)$ and $\\hat{L}(\\theta_g, \\phi_i)$ should not have the $i$ in subscript on the left hand side since it also appears in the summation on the right.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper proposes a new framework called pFedGate for training personalized models at client devices, such that each model can have different sparsity level depending on the devices computational and memory capacities. The idea is to generate personalized masks with different sparsity levels for each device, using a small gating layer per device. The gating layer itself is small in size and hence the paper claims that it does not add much overhead. The empirical results show that this technique indeed achieves better personalization than existing methods.", "strength_and_weaknesses": "Strengths:\n1. The proposed framework seems to achieve better personalization accuracy than existing models.\n2. The idea of generating personalized masks (gating weights) using a computationally cheap network is novel.\n\n\nConcerns / Questions:\n1. I think Assumption 1 might be a strong assumption, since it assumes bounded diversity between the globally optimal model, and the optimal sparse local model. In particular, I think $\\sigma_i$ might itself depend on the level of sparsity for the device.\n\n2. In the proof of proposition 1, I do not follow how the Reverse Cauchy-Schwarz inequality is applied, since the left hand side of Eq. (13) has the l2 norm squared of element-wise multiplication of vectors instead of inner product. To be precise, the left hand side will look like $(\\sum_j (M^*_{i,j}\\theta^*_{g,j})^2 )$ instead of $(\\sum_j (M^*_{i,j}\\theta^*_{g,j}) )$. Also, I think the denominator of Eq. (4) should not have the square in norm of $\\theta^*_{g}$.\n\n3. The experimental results do not provide the time (in seconds) and memory requirements (in GB / MB) for the training of the models. Further, the convergence vs communication round curve is only provided for EMNIST dataset. This data is important since one of the motivations for this paper is to perform Federated Learning efficiently on computationally slow devices.\n\n4. In Eq. (7), is the gradient w.r.t $\\phi$ as well? In other words, does Theorem 2 prove convergence for the gating layer as well?\n\n5. On page 6, section 6.1, I think the definition of $L(\\theta_g, \\phi_i)$ and $\\hat{L}(\\theta_g, \\phi_i)$ should not have the $i$ in subscript on the left hand side since it also appears in the summation on the right.", "clarity,_quality,_novelty_and_reproducibility": "The paper has typos, but otherwise is well written. It also have novelty in the method proposed. I have not verified all the proofs or tried to run the provided code.", "summary_of_the_review": "While the proposed idea of generating personalized masks (gating weights) is novel and has potential, I have some concerns regarding the theory and experimental results that I have mentioned above.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667192947934}, {"id": "snzwPGx_BYd", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2420/Reviewer_PyZd"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper proposes a personalized federated learning framework that selects a subset of model parameters (thus sparse) for each training example via personalized gating units for each client (Fig.1). The proposed method, pFedGate, has a gating unit/layer for each client, consisting of a rescaling factor M, an importance factor G and learns the binary mask by solving a knapsack problem per training example (Fig.2). Theoretical analysis on generalization, convergence and complexities are provided. Experiments on common personalized FL benchmark datasets show that pFedGate can outperform existing methods.", "review_text": "The paper proposes a framework for personalized FL and it can work well in the experiments. However, it does not provide sufficient justification for the main claim/contribution of solving the resource heterogeneity issue as all clients still need to store the global model locally for training and inference.", "strengths": "Strengths\n- Writing is clear and easy to understand\n- Thorough theoretical analysis\n\nWeaknesses\n- The framework fails to solve the resource heterogeneity issue since the clients still need to store the whole (global) model.\n- Some implementation and experiment details remain unclear.\n\nDetail comments:\n\n1. Sec.4 first paragraph: the two-step process “still requires computational and communication costs corresponding to the un-compressed models during the FL process”. However, according to Fig.1, pFedGate also needs to store the un-compressed model locally in order to compute the compressed model using M’, which is also mentioned in Sec.5.1 “all clients share the same global model \\theta_g that downloaded from server”. In addition, since every example uses different blocks of the model, each client has to store the whole model for inference. Therefore, both training and inference are dense, making the main contribution unsubstantiated. (Appendix E is also misleading.)\n\n2. The gradient computation requires further discussion. The forward pass of every example requires solving a knapsack problem as shown in Sec.5.2.3. Combing with the fact that we are training multiple networks (\\theta and \\phi), it would be helpful to elaborate on how the gradient of each network is computed using the knapsack solution. \n\n3. What is the relative sparsity of gating layer s_{\\phi_i} in Sec.6.3?\n\n4. For the experiments\n- How many clients are used for the experiments in Sec.7.2?\n- Sec.5.2.2 discusses that different DNN layers can have different block structures. Then how are the layers being partitioned in the experiments?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper proposes a personalized federated learning framework that selects a subset of model parameters (thus sparse) for each training example via personalized gating units for each client (Fig.1). The proposed method, pFedGate, has a gating unit/layer for each client, consisting of a rescaling factor M, an importance factor G and learns the binary mask by solving a knapsack problem per training example (Fig.2). Theoretical analysis on generalization, convergence and complexities are provided. Experiments on common personalized FL benchmark datasets show that pFedGate can outperform existing methods.", "strength_and_weaknesses": "Strengths\n- Writing is clear and easy to understand\n- Thorough theoretical analysis\n\nWeaknesses\n- The framework fails to solve the resource heterogeneity issue since the clients still need to store the whole (global) model.\n- Some implementation and experiment details remain unclear.\n\nDetail comments:\n\n1. Sec.4 first paragraph: the two-step process “still requires computational and communication costs corresponding to the un-compressed models during the FL process”. However, according to Fig.1, pFedGate also needs to store the un-compressed model locally in order to compute the compressed model using M’, which is also mentioned in Sec.5.1 “all clients share the same global model \\theta_g that downloaded from server”. In addition, since every example uses different blocks of the model, each client has to store the whole model for inference. Therefore, both training and inference are dense, making the main contribution unsubstantiated. (Appendix E is also misleading.)\n\n2. The gradient computation requires further discussion. The forward pass of every example requires solving a knapsack problem as shown in Sec.5.2.3. Combing with the fact that we are training multiple networks (\\theta and \\phi), it would be helpful to elaborate on how the gradient of each network is computed using the knapsack solution. \n\n3. What is the relative sparsity of gating layer s_{\\phi_i} in Sec.6.3?\n\n4. For the experiments\n- How many clients are used for the experiments in Sec.7.2?\n- Sec.5.2.2 discusses that different DNN layers can have different block structures. Then how are the layers being partitioned in the experiments?", "clarity,_quality,_novelty_and_reproducibility": "The writing is clear, but the proposed method does not live up to the expectation of handling resource-heterogeneous clients. Some of the implementation details are missing. ", "summary_of_the_review": "The paper proposes a framework for personalized FL and it can work well in the experiments. However, it does not provide sufficient justification for the main claim/contribution of solving the resource heterogeneity issue as all clients still need to store the global model locally for training and inference.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666661598867}, {"id": "2lXzXgqkdm", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper2420/Reviewer_wTvc"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces a new approach to personalized federated learning via sparse model adaptation. The key idea is to configure local models as subnets of a master net representing the global model. \n\nThe subnets are generated via a block-wise binary mask that (1) zeroes out less important components; and (2) enforce a certain sparsity level such that the no. of non-zero parameters of resulting subnets fits with local hardware constraints. \n\nLocal models are therefore heterogeneous compression of the same global model. This importantly allows local models to have heterogeneous architectures, which is, to my understanding, a less studied aspect of personalized FL. Most existing personalized FL approaches only consider data heterogeneity (but not model heterogeneity).   \n\nThe generative mechanism of the subset is parameterized by a two-flow model: (1) one flow generates crude masks (that might violate local sparsity constraints); (2) the other flow generates an importance score for each parameter blocks. The output of two flows are combined via solving a knapsack problem, resulting in a binary mask that meets the hard sparsity constraint.\n\nComparing to existing model compression or gradient quantization techniques, which are often post-training editing approaches, the proposed approach allows learning both the compression mechanism & the updates to the global model. ", "review_text": "The paper presents an interesting formulation of personalized FL & a new solution addressing a less studied issue of model heterogeneity. The novelty of this work is strong but its execution needs more polishing, particularly regarding missing experiments with both related work & vanilla baseline (to make sure such proposed mechanism is needed & cannot be bypassed by simpler, trivial engineering solution). The presentation of proving techniques can also be improved. Please refer to my specific comments above. Overall, I believe this will be an acceptable paper if the above are addressed sufficiently. ", "strengths": "Strengths:\n\n+ Both problem formulation & proposed solution are novel to me. I have not seen similar approaches to this before.\n+ The technical presentation is very well-organized, particularly the well-structured appendix that makes it easy to zoom into each component of the proposed work\n+ The empirical results are reasonably extensive with comparison to multiple baselines in the same direction\n+ There are interesting theoretical results showing both convergence of the proposed algorithm and the existence of optimal local solution: among optimal local solution that meets local sparse constraints, there exists one that can be acquired via masking the global model\n\nWeaknesses:\n\n- Lacking comparison with a known baseline -- Fallah et al., 2020\n- Lacking discussion with other related works on FL with heterogeneous models\n- Theoretical results are established under strong convexity of the loss function, which is likely not the case unless the model is linear\n- Lacking experiments showing advantages over simpler knowledge distillation technique for FL with heterogeneous model\n\nI will elaborate on those weaknesses with specific comments & suggestions below.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper introduces a new approach to personalized federated learning via sparse model adaptation. The key idea is to configure local models as subnets of a master net representing the global model. \n\nThe subnets are generated via a block-wise binary mask that (1) zeroes out less important components; and (2) enforce a certain sparsity level such that the no. of non-zero parameters of resulting subnets fits with local hardware constraints. \n\nLocal models are therefore heterogeneous compression of the same global model. This importantly allows local models to have heterogeneous architectures, which is, to my understanding, a less studied aspect of personalized FL. Most existing personalized FL approaches only consider data heterogeneity (but not model heterogeneity).   \n\nThe generative mechanism of the subset is parameterized by a two-flow model: (1) one flow generates crude masks (that might violate local sparsity constraints); (2) the other flow generates an importance score for each parameter blocks. The output of two flows are combined via solving a knapsack problem, resulting in a binary mask that meets the hard sparsity constraint.\n\nComparing to existing model compression or gradient quantization techniques, which are often post-training editing approaches, the proposed approach allows learning both the compression mechanism & the updates to the global model. ", "strength_and_weaknesses": "Strengths:\n\n+ Both problem formulation & proposed solution are novel to me. I have not seen similar approaches to this before.\n+ The technical presentation is very well-organized, particularly the well-structured appendix that makes it easy to zoom into each component of the proposed work\n+ The empirical results are reasonably extensive with comparison to multiple baselines in the same direction\n+ There are interesting theoretical results showing both convergence of the proposed algorithm and the existence of optimal local solution: among optimal local solution that meets local sparse constraints, there exists one that can be acquired via masking the global model\n\nWeaknesses:\n\n- Lacking comparison with a known baseline -- Fallah et al., 2020\n- Lacking discussion with other related works on FL with heterogeneous models\n- Theoretical results are established under strong convexity of the loss function, which is likely not the case unless the model is linear\n- Lacking experiments showing advantages over simpler knowledge distillation technique for FL with heterogeneous model\n\nI will elaborate on those weaknesses with specific comments & suggestions below.", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\n\nThe paper is very well-written in general. As I commended the authors above, their appendix is very well-structured. I do, however, have a few nitpicks:\n\n1. The presentation of the proving techniques is too terse. It would always be good to start with a narrative of a general proving strategy, followed by step-by-step derivation annotated with explanation why each step holds\n2. Section 5.2.3 is very much lacking in detail, which actually raises concerns about the differentiability of the entire workflow -- I will comment more on this below.\n\nNovelty & Quality:\n\nAs stated above, problem formulation & solution idea are new so I acknowledge the novelty here. But, I am not sure I follow why the last key step in combining output of the two flows (mask generator & importance score generator)  via knapsack (Section 5.2.3) is differentiable.\n\nI agree that using the previous technique of binarized neural net, the authors can make the representation of the \"knapsack matrix\" I differentiable. But, it seem the max operator is still not differentiable & so if the solution to Eq. (6) is to be integrated into the \"end-to-end\" pipeline, it would come out in the form of a bi-level optimizable task -- further treatment of this probably needs to be detailed because otherwise, it is not immediately clear to me how gradient flows through that bi-level structure.\n\nAs for the empirical experiment, please do consider comparing with Fallah et al. 2020 too for thoroughness. In addition, there should be key discussions & demonstrations of how this proposed mechanism has an advantage over simpler engineering solution for FL with knowledge distillation. For instance, clients can download big model from the server but will distill into smaller local model with heterogeneous structure before baking in local data; in turn, the server can pull local models from clients and distill soft labels from them to train its global model. How would the proposed model perform against this vanilla approach?\n\nMore suggestions:\n\nAlthough the idea of using heterogeneous models in personalized FL is quite new; it has been explored before (in a different context with a different probabilistic formulation) in standard FL setting. For example,\n\nBayesian Nonparametric Federated Learning of Neural Network (ICML-19)\nStatistical Model Aggregation via Parameter Matching (NeurIPS-19)\n\nBoth approaches in these works would allow local models to be heterogeneous in size so as to accommodate for the heterogeneity in their hardware capacity. The formulation is different but it addresses the same aspect so some discussions regarding pros/cons between the two lines of approaches would be nice to have.\n\nMore nitpicks:\n\nIt would be good to repeat all experiments a few times to generate error bars\nIncluding in the appendix more detail about the differentiable parameterization detailed in (Hubara et al. 2016) \nDiscussing potential directions to expand the proving technique beyond the strong convexity assumption which is tied to simple linear model\n\nReproducibility:\n\nThe code is released and the algorithm description is sufficiently clear so I believe reproducibility is possible. But still, for a thorough understanding of the approach, Section 5.2.3 needs a significant flesh-out. \n ", "summary_of_the_review": "The paper presents an interesting formulation of personalized FL & a new solution addressing a less studied issue of model heterogeneity. The novelty of this work is strong but its execution needs more polishing, particularly regarding missing experiments with both related work & vanilla baseline (to make sure such proposed mechanism is needed & cannot be bypassed by simpler, trivial engineering solution). The presentation of proving techniques can also be improved. Please refer to my specific comments above. Overall, I believe this will be an acceptable paper if the above are addressed sufficiently. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666563383089}], "openreview_url": "https://openreview.net/forum?id=3vOtC1t1kF", "arxiv_id": "2305.02776", "paper_pdf": "papers/3vOtC1t1kF.pdf", "paper_pdf_sha256": "2db5442db797c005ec4a900d973d002d4f400b5716063c6ecbea64994fb61bda", "paper_pdf_bytes": 3128174, "paper_pdf_source": "openreview", "code_url": "https://github.com/yxdyc/pFedGate", "code_repository": "yxdyc/pFedGate", "code_commit": "b8cfa156acdb08207d33641a355959503ca1db78", "code_archive": "repos/3vOtC1t1kF.zip", "code_archive_sha256": "a9bd83e2b07861800568ce1cd6e93a00b7bacee23f7d7d32f4653b52376c5e2a", "code_archive_bytes": 1405277, "code_file_count": 40, "code_extensions": {".py": 39, ".sh": 1}, "github_disk_usage_kb": 1374, "github_languages": {"Python": 332909, "Shell": 1523}, "github_archived": false, "github_pushed_at": "2023-05-26T07:25:36Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/efficient-personalized-federated-learning-via"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qvUJV2-t_c", "year": 2022, "status": "rejected", "title": "Using a one dimensional parabolic model of the full-batch loss to estimate learning rates during training", "authors": ["Maximus Mutschler", "Kevin Alexander Laube", "Andreas Zell"], "authorids": ["~Maximus_Mutschler1", "~Kevin_Alexander_Laube1", "~Andreas_Zell1"], "authors_source": "OpenReview API", "abstract": "A fundamental challenge in Deep Learning is to find optimal step sizes for stochastic gradient descent automatically. In traditional optimization, line searches are a commonly used method to determine step sizes. One problem in Deep Learning is that finding appropriate step sizes on the full-batch loss is unfeasibly expensive. Therefore, classical line search approaches, designed for losses without inherent noise, are usually not applicable. Recent empirical findings suggest that the full-batch loss behaves locally parabolically in the direction of noisy update step directions. Furthermore, the trend of the optimal update step size changes slowly. By exploiting these findings, this work introduces a line-search method that approximates the full-batch loss with a parabola estimated over several mini-batches. Learning rates are derived from such parabolas during training. In the experiments conducted, our approach mostly outperforms SGD tuned with a piece-wise constant learning rate schedule and other line search approaches for Deep Learning across models, datasets, and batch sizes on validation and test accuracy.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "rJG0sPFHJZL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper526/Reviewer_w2rC"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This work proposes a line search method LABPAL for first order methods to train neural networks with SGD. This work is motivated to address the issues in previous works that while the classical line search methods assume the exact descent direction, which is expensive to calculate for large data, the inexact methods based on mini-batches of examples are not robust. The idea is to leverage the two empirical observations on the loss landscape that the full-batch loss along lines in the SGD update direction is shaped parabolically and the optimal step size changes rather slowly. The authors provide extensive experimental results using various neural network models for image classification to compare LABPAL performed with either SGD or normalized SGD to other line search methods and show that LABPAL can perform better than SGD tuned optimally.", "review_text": "The main drawback of this work is that LABPAL is compared to SGD and not to SGD with momentum, even though the fundamental idea behind them is very similar. I think what could make this work strong is showing a competitive result by comparing Momentum SGD with and without LABPAL step sizes. Without it the practical value of the proposed idea is quite limited even if it may outperform other line search methods or SGD tuned optimally.\n\nAnother concern I have is in the limited technical novelty and its implication when taking into account other scenarios where the assumptions/observations do not hold or generalize. LABPAL is essentially the backtracking line search with estimating full loss with multiple large mini-batches, and yet its reliance on empirical observations only pertaining to a certain task (image classification) make the scope of this method being applied to quite limited, not to mention that the arbitrarily chosen hyperparameters required to make LABPAL work may not generalize.\n\nThe Derivation step 4 seems really like a heuristics that may not hold depending on some data distributions/characteristics.\n\nIt's not obvious that why the step sizes have to increase after some period of time.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work proposes a line search method LABPAL for first order methods to train neural networks with SGD. This work is motivated to address the issues in previous works that while the classical line search methods assume the exact descent direction, which is expensive to calculate for large data, the inexact methods based on mini-batches of examples are not robust. The idea is to leverage the two empirical observations on the loss landscape that the full-batch loss along lines in the SGD update direction is shaped parabolically and the optimal step size changes rather slowly. The authors provide extensive experimental results using various neural network models for image classification to compare LABPAL performed with either SGD or normalized SGD to other line search methods and show that LABPAL can perform better than SGD tuned optimally.", "main_review": "The main drawback of this work is that LABPAL is compared to SGD and not to SGD with momentum, even though the fundamental idea behind them is very similar. I think what could make this work strong is showing a competitive result by comparing Momentum SGD with and without LABPAL step sizes. Without it the practical value of the proposed idea is quite limited even if it may outperform other line search methods or SGD tuned optimally.\n\nAnother concern I have is in the limited technical novelty and its implication when taking into account other scenarios where the assumptions/observations do not hold or generalize. LABPAL is essentially the backtracking line search with estimating full loss with multiple large mini-batches, and yet its reliance on empirical observations only pertaining to a certain task (image classification) make the scope of this method being applied to quite limited, not to mention that the arbitrarily chosen hyperparameters required to make LABPAL work may not generalize.\n\nThe Derivation step 4 seems really like a heuristics that may not hold depending on some data distributions/characteristics.\n\nIt's not obvious that why the step sizes have to increase after some period of time.", "summary_of_the_review": "While it is interesting to see a new line search method that is competitive to existing line search methods, the comparison doesn't seem entirely fair given that the proposed method shares essentially the same idea used in momentum. Maybe authors could give an estimate as to the required amount of total computations and compare it to that of momentum SGD.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636030110505}, {"id": "dp1pXS7yyZM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper526/Reviewer_rHNN"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors propose a parabolic line search for SGD to automatically adjust the learning rate during training.  The authors emphasize that their method is not motivated by theory, but on a previous empirical evaluation of ResNet-20 on CIFAR-10.  Their method is evaluated against several other recently proposed line search strategies for SGD (along with SGD using constant and piecewise learning rates) on CIFAR-100 for Resnet-20, MobileNet-V2, and DenseNet-121 (all using a mini-batch of 128).  The authors further show (in their method) how to overcome increasing noise levels in the gradient estimation of SGD by considering a mini-batch size of 10.", "review_text": "Adaptive learning rate methods for SGD are important and improvements over previous line search methods have the potential for significant impact.  However, the evaluation of different methods in the paper are either lacking or significantly suboptimal.\n\nFor the former, evaluated methods should be more extensive; SGD with a constant learning rate or piece-wise constant learning rate are not widely used (outside of ResNet architectures for the latter [1]).  At the very least, SGD with a decaying learning rate should be compared to.  Furthermore, the evaluation of other methods does not seem correct.  For instance, SLS uniformly performs terribly for CIFAR-100 ResNet and DenseNet, in stark contrast to the results in the original SLS NeurIPS 2019 paper.  Can the authors please comment on such a large discrepancy?  With regards to how the hyperparmaters for various methods are tuned, the grid search looks coarse and the search space very suboptimal, which is supported by ResNet-20 only achieving ~60% accuracy (for SGD) on CIFAR-100.  This is further supported by PLS only beginning with an initial step size of 10^{-4}; ResNet architectures for CIFAR-100 have displayed excellent performance using PLS with an initial learning rate of 0.1, which is then divided by 10 half-way through training and further divided by 10 three-fourths through training [1].  At the very least, the initial learning rate for this optimizer should always be the optimal learning rate for SGD output during hyperparameter optimization.\n\nOther comments:\n-\"Gradient-only line search (GOLSI)... its performance is rather weak.\"<- Please include citation for this claim\n\n-\"This was done for 10,000 consecutive SGD update steps of a ResNet18’s training process on a subset of CIFAR-10. Representative plots of their 10,000 measured full-batch losses along lines are presented in Figure 1. Relevant insights and found properties of these works will be introduced and exploited to derive our algorithm in Section 3.\" <- The problem with basing all presented work on this single study is that different learning schedules for SGD (and\nvariants) differ significantly for different datasets and problem domains.  See:\nSchmidt, Robin M., Frank Schneider, and Philipp\nHennig. \"Descending through a crowded valley-benchmarking deep\nlearning optimizers.\" International Conference on Machine\nLearning. PMLR, 2021.\n\n-\"Besides decreasing the learning rate, increasing the batch size remains an important choice to tackle gradient noise\" <- This is a point of contention in the literature; several studies report that increasing the batch size degrades performance [2,3].\n\n-The observation/derivation section in the paper is very difficult to read.\n\n-To clarify, the authors shrink the batch size from 128 to 10 in Figure 5, as a surrogate for increasing gradient noise.  Is that correct?  It would be interesting to see intermediate batch sizes from 128 decreasing to 10 to actually see how noise floor increases w.r.t. mini-batch size (a batch size of 10 is unreasonably small).\n\n\nReferences:\n[1] He, Kaiming, et al. \"Identity mappings in deep residual networks.\"\nEuropean conference on computer vision. Springer, Cham, 2016.\n[2] Keskar, Nitish Shirish, et al. \"On large-batch training for deep learning: Generalization gap and sharp minima.\" arXiv preprint arXiv:1609.04836 (2016).\n[3] Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors propose a parabolic line search for SGD to automatically adjust the learning rate during training.  The authors emphasize that their method is not motivated by theory, but on a previous empirical evaluation of ResNet-20 on CIFAR-10.  Their method is evaluated against several other recently proposed line search strategies for SGD (along with SGD using constant and piecewise learning rates) on CIFAR-100 for Resnet-20, MobileNet-V2, and DenseNet-121 (all using a mini-batch of 128).  The authors further show (in their method) how to overcome increasing noise levels in the gradient estimation of SGD by considering a mini-batch size of 10.", "main_review": "Adaptive learning rate methods for SGD are important and improvements over previous line search methods have the potential for significant impact.  However, the evaluation of different methods in the paper are either lacking or significantly suboptimal.\n\nFor the former, evaluated methods should be more extensive; SGD with a constant learning rate or piece-wise constant learning rate are not widely used (outside of ResNet architectures for the latter [1]).  At the very least, SGD with a decaying learning rate should be compared to.  Furthermore, the evaluation of other methods does not seem correct.  For instance, SLS uniformly performs terribly for CIFAR-100 ResNet and DenseNet, in stark contrast to the results in the original SLS NeurIPS 2019 paper.  Can the authors please comment on such a large discrepancy?  With regards to how the hyperparmaters for various methods are tuned, the grid search looks coarse and the search space very suboptimal, which is supported by ResNet-20 only achieving ~60% accuracy (for SGD) on CIFAR-100.  This is further supported by PLS only beginning with an initial step size of 10^{-4}; ResNet architectures for CIFAR-100 have displayed excellent performance using PLS with an initial learning rate of 0.1, which is then divided by 10 half-way through training and further divided by 10 three-fourths through training [1].  At the very least, the initial learning rate for this optimizer should always be the optimal learning rate for SGD output during hyperparameter optimization.\n\nOther comments:\n-\"Gradient-only line search (GOLSI)... its performance is rather weak.\"<- Please include citation for this claim\n\n-\"This was done for 10,000 consecutive SGD update steps of a ResNet18’s training process on a subset of CIFAR-10. Representative plots of their 10,000 measured full-batch losses along lines are presented in Figure 1. Relevant insights and found properties of these works will be introduced and exploited to derive our algorithm in Section 3.\" <- The problem with basing all presented work on this single study is that different learning schedules for SGD (and\nvariants) differ significantly for different datasets and problem domains.  See:\nSchmidt, Robin M., Frank Schneider, and Philipp\nHennig. \"Descending through a crowded valley-benchmarking deep\nlearning optimizers.\" International Conference on Machine\nLearning. PMLR, 2021.\n\n-\"Besides decreasing the learning rate, increasing the batch size remains an important choice to tackle gradient noise\" <- This is a point of contention in the literature; several studies report that increasing the batch size degrades performance [2,3].\n\n-The observation/derivation section in the paper is very difficult to read.\n\n-To clarify, the authors shrink the batch size from 128 to 10 in Figure 5, as a surrogate for increasing gradient noise.  Is that correct?  It would be interesting to see intermediate batch sizes from 128 decreasing to 10 to actually see how noise floor increases w.r.t. mini-batch size (a batch size of 10 is unreasonably small).\n\n\nReferences:\n[1] He, Kaiming, et al. \"Identity mappings in deep residual networks.\"\nEuropean conference on computer vision. Springer, Cham, 2016.\n[2] Keskar, Nitish Shirish, et al. \"On large-batch training for deep learning: Generalization gap and sharp minima.\" arXiv preprint arXiv:1609.04836 (2016).\n[3] Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.\n", "summary_of_the_review": "Interesting take on a previously explored strategy for adaptable learning rates in SGD.  However, the results and (lack of) comparison to previous related methods does not look correct.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635986180259}, {"id": "E_rfOA0e2u9", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper526/Reviewer_kiNL"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This paper uses empirical observations of Mutschler & Zell (2021) to come up with a line search algorithm called LABPAL, which is essentially fitting a 1-D parabola to the (approximate) full-batch loss in the direction of the mini-batch gradient, and using the minima of the parabola to obtain the step size. The authors include various tricks to make this approach work in practice. The show empirically on various image classification workloads that LABPAL is mostly better or on-par with SGD (with piecewise constant learning rate schedule) and other recent line-search methods.  \n", "review_text": "Strengths:\nTo the best of my knowledge, the exact idea proposed in this paper is novel.\n\nWeaknesses:\n- The algorithm is built on empirical observations of a specific image classification workload. The experiments only include image classification workloads, so whether this algorithm generalizes to other tasks (e.g. NLP tasks) is an open question.\n- Many of the design decisions made for the algorithm seem arbitrary, and there are many additional hyperparameters that seem important to tune. The method 1) approximates the full-batch loss using many mini-batch evaluations, 2) reuses the same step size for some consecutive steps, 3) increases the batch size with a piecewise constant schedule as training progress, and 4)uses a slightly larger step size value than the actual optimum given by the parabola by multiplying by a factor between 1 and 2. There are no ablation studies to understand which of these choices are important. Furthermore, they introduce many additional hyperparameters. In fact, looking at Section G.1 in the Appendix, it seems like LABPAL has the most number of hyperparameters.\n- There is no comparison to SGD with Momentum, which performs the best in the workloads considered in the experiments. Given that the method has so many additional hyperparameters, and the comparison with the best-performing optimizer is lacking, the practicality of the proposed method is questionable.\n- Section 4.3 (adaptation to varying gradient noise) doesn’t make much sense to me. It seems like all that is happening is the batch size is being increased. So then the comparison with other methods doesn’t seem fair, because the batch size is no longer the same.\n- Convergence analysis is missing.\n\nComments and questions:\n- The distinction between step size and learning rate is unclear. What exactly is the difference? Could you make things clearer in the paper? For example, it seems like in Equation 6, learning rate = \\lambda = alpha * s_{min,t}/||g_t||, but in the algorithm box, it seems like learning rate = learning rate * alpha.\n- In Section G.1, there are some parameters under LABPAL that I’m not sure were mentioned in the main paper: initial measuring step size, parabolic approximation sample step size, and approximation batch size.\n\n----\nUpdate:\nI have read the reviews from the other reviewers and the author responses. I agree with the authors that the paper's contributions are real and valid. However, as other reviewers also noted, I don't think the contributions are enough for acceptance, therefore, I am keeping my score. I would be excited to see a future version of the paper that has a wider range of experiments (maybe showing that the properties also hold for other deep learning tasks).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper uses empirical observations of Mutschler & Zell (2021) to come up with a line search algorithm called LABPAL, which is essentially fitting a 1-D parabola to the (approximate) full-batch loss in the direction of the mini-batch gradient, and using the minima of the parabola to obtain the step size. The authors include various tricks to make this approach work in practice. The show empirically on various image classification workloads that LABPAL is mostly better or on-par with SGD (with piecewise constant learning rate schedule) and other recent line-search methods.  \n", "main_review": "Strengths:\nTo the best of my knowledge, the exact idea proposed in this paper is novel.\n\nWeaknesses:\n- The algorithm is built on empirical observations of a specific image classification workload. The experiments only include image classification workloads, so whether this algorithm generalizes to other tasks (e.g. NLP tasks) is an open question.\n- Many of the design decisions made for the algorithm seem arbitrary, and there are many additional hyperparameters that seem important to tune. The method 1) approximates the full-batch loss using many mini-batch evaluations, 2) reuses the same step size for some consecutive steps, 3) increases the batch size with a piecewise constant schedule as training progress, and 4)uses a slightly larger step size value than the actual optimum given by the parabola by multiplying by a factor between 1 and 2. There are no ablation studies to understand which of these choices are important. Furthermore, they introduce many additional hyperparameters. In fact, looking at Section G.1 in the Appendix, it seems like LABPAL has the most number of hyperparameters.\n- There is no comparison to SGD with Momentum, which performs the best in the workloads considered in the experiments. Given that the method has so many additional hyperparameters, and the comparison with the best-performing optimizer is lacking, the practicality of the proposed method is questionable.\n- Section 4.3 (adaptation to varying gradient noise) doesn’t make much sense to me. It seems like all that is happening is the batch size is being increased. So then the comparison with other methods doesn’t seem fair, because the batch size is no longer the same.\n- Convergence analysis is missing.\n\nComments and questions:\n- The distinction between step size and learning rate is unclear. What exactly is the difference? Could you make things clearer in the paper? For example, it seems like in Equation 6, learning rate = \\lambda = alpha * s_{min,t}/||g_t||, but in the algorithm box, it seems like learning rate = learning rate * alpha.\n- In Section G.1, there are some parameters under LABPAL that I’m not sure were mentioned in the main paper: initial measuring step size, parabolic approximation sample step size, and approximation batch size.\n\n----\nUpdate:\nI have read the reviews from the other reviewers and the author responses. I agree with the authors that the paper's contributions are real and valid. However, as other reviewers also noted, I don't think the contributions are enough for acceptance, therefore, I am keeping my score. I would be excited to see a future version of the paper that has a wider range of experiments (maybe showing that the properties also hold for other deep learning tasks).", "summary_of_the_review": "The proposed method doesn’t seem practically useful. It introduces many additional hyperparameters, and it is unclear whether it performs better than SGD with Momentum, and in settings where it is not confirmed whether the empirical observations that justify the method hold. Therefore, I recommend rejection.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635909180485}, {"id": "QoXqr29Ahsr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper526/Reviewer_mG5J"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this paper, the authors propose an algorithm for the automatic selection of the step size for SGD and apply this algorithm for DNN training. They suggest that full batch loss along the negative direction of the normalized batch gradient could be localy approximated with a parabola with high precision. So optimal step size could be also approximated as distance to minimum of that parabola. The authors notice that optimal step size changes slowly during SGD optimization process therefore one can reuse optimal step size for number of consequtive SGD updates and by doing that reduce algorithm complexity and made it applicable. ", "review_text": "In this paper, the authors propose an algorithm for the automatic selection of the step size for SGD and apply this algorithm for DNN training. They suggest that full batch loss along the negative direction of the normalized batch gradient could be localy approximated with a parabola with high precision. So optimal step size could be also approximated as distance to minimum of that parabola. The authors notice that optimal step size changes slowly during SGD optimization process therefore one can reuse optimal step size for number of consequtive SGD updates and by doing that reduce algorithm complexity and made it applicable. \n\nWeak points:\n\n-W1. The overall novelty of paper is limited because propesed method LABPAL is straightforward modification of previously published PAL algorithm (https://arxiv.org/pdf/1903.11991.pdf). In LABPAL full batch loss is approximated via multiple  batches compared to single batch approximation in PAL. The difference in a way to perform parabolic approximation seems like a minor modification. Also, the experimental background for applicability of parabolic approximation (i.e. Observations 1, 2, 3, 4, 5, 6) is basically stated in the same way in https://arxiv.org/pdf/2103.17132.pdf as well as illustrations to these observations (figures 1, 2 matches figures 2, 6 from the latter paper). Other technical details like adaptation to varying gradient noise and relation with first Wolfe constant have marginal influence on the main point of the paper.\n-W2. The Observation 2 seems not well reasoned since the conclusion is made for only one choice of dataset and model and in other settings this observation may fail.\n\nStrong points:\n\n-S1. The experiments show that the proposed method outperforms other baseline methods on various datasets and that LABPAL produce stable results for a wide range of hyperparameters.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors propose an algorithm for the automatic selection of the step size for SGD and apply this algorithm for DNN training. They suggest that full batch loss along the negative direction of the normalized batch gradient could be localy approximated with a parabola with high precision. So optimal step size could be also approximated as distance to minimum of that parabola. The authors notice that optimal step size changes slowly during SGD optimization process therefore one can reuse optimal step size for number of consequtive SGD updates and by doing that reduce algorithm complexity and made it applicable. ", "main_review": "In this paper, the authors propose an algorithm for the automatic selection of the step size for SGD and apply this algorithm for DNN training. They suggest that full batch loss along the negative direction of the normalized batch gradient could be localy approximated with a parabola with high precision. So optimal step size could be also approximated as distance to minimum of that parabola. The authors notice that optimal step size changes slowly during SGD optimization process therefore one can reuse optimal step size for number of consequtive SGD updates and by doing that reduce algorithm complexity and made it applicable. \n\nWeak points:\n\n-W1. The overall novelty of paper is limited because propesed method LABPAL is straightforward modification of previously published PAL algorithm (https://arxiv.org/pdf/1903.11991.pdf). In LABPAL full batch loss is approximated via multiple  batches compared to single batch approximation in PAL. The difference in a way to perform parabolic approximation seems like a minor modification. Also, the experimental background for applicability of parabolic approximation (i.e. Observations 1, 2, 3, 4, 5, 6) is basically stated in the same way in https://arxiv.org/pdf/2103.17132.pdf as well as illustrations to these observations (figures 1, 2 matches figures 2, 6 from the latter paper). Other technical details like adaptation to varying gradient noise and relation with first Wolfe constant have marginal influence on the main point of the paper.\n-W2. The Observation 2 seems not well reasoned since the conclusion is made for only one choice of dataset and model and in other settings this observation may fail.\n\nStrong points:\n\n-S1. The experiments show that the proposed method outperforms other baseline methods on various datasets and that LABPAL produce stable results for a wide range of hyperparameters.\n", "summary_of_the_review": "It seems that most part of the paper (including figures and statements) is taken from one of two previously published papers and hence the novelty contained in the paper itself is very limited.  This inclines me towards rejection of the paper in its current form. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635851648752}], "openreview_url": "https://openreview.net/forum?id=qvUJV2-t_c", "arxiv_id": "2108.13880", "paper_pdf": "papers/qvUJV2-t_c.pdf", "paper_pdf_sha256": "9f597cea7a98e0aa66702ce4e8dbe59eb0dd032694dd240317aa790df57e14c1", "paper_pdf_bytes": 1981608, "paper_pdf_source": "openreview", "code_url": "https://github.com/cogsys-tuebingen/LABPAL", "code_repository": "cogsys-tuebingen/LABPAL", "code_commit": "f577cf976d5b88ea2dc901c2f71d30a643424a98", "code_archive": "repos/qvUJV2-t_c.zip", "code_archive_sha256": "db9bba8e4e81078624f30af104f2309f8b2a59a7707809d0711fadc47dd73394", "code_archive_bytes": 1145573, "code_file_count": 43, "code_extensions": {".py": 43}, "github_disk_usage_kb": 7363, "github_languages": {"Python": 321710}, "github_archived": false, "github_pushed_at": "2022-09-19T14:10:35Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/using-a-one-dimensional-parabolic-model-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "XG1Drw7VbLJ", "year": 2021, "status": "rejected", "title": "Defining Benchmarks for Continual Few-Shot Learning", "authors": ["Antreas Antoniou", "Massimiliano Patacchiola", "Mateusz Ochal", "Amos Storkey"], "authorids": ["~Antreas_Antoniou2", "~Massimiliano_Patacchiola1", "~Mateusz_Ochal1", "~Amos_Storkey1"], "authors_source": "OpenReview API", "abstract": "In recent years there has been substantial progress in few-shot learning, where a model is trained on a small labeled dataset related to a specific task, and in continual learning, where a model has to retain knowledge acquired on a sequence of datasets. Both of these fields are different abstractions of the same real world scenario, where a learner has to adapt to limited information from different changing sources and be able to generalize in and from each of them. Combining these two paradigms, where a model is trained on several sequential few-shot tasks, and then tested on a validation set stemming from all those tasks, helps by explicitly defining the competing requirements for both efficient integration and continuity. In this paper we propose such a setting, naming it Continual Few-Shot Learning (CFSL). We first define a theoretical framework for CFSL, then we propose a range of flexible benchmarks to unify the evaluation criteria. As part of the benchmark, we introduce a compact variant of ImageNet, called SlimageNet64, which retains all original 1000 classes but only contains 200 instances of each one (a total of 200K data-points) downscaled to 64 by 64 pixels. We provide baselines for the proposed benchmarks using a number of popular few-shot and continual learning methods, exposing previously unknown strengths and weaknesses of those algorithms. The dataloader and dataset will be released with an open-source license.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "u4viPbfkYTC", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1737/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The work proposes Continual Few-Shot learning -- a setting to study tasks (1) with a small labeled dataset, and (2) retain knowledge acquired on a sequence of instances.  Additionally, the authors build a compact variant of ImageNet which retains all original 1000 classes but only contains 200 instances of each one (a total of 200K data-points) downscaled to 64 × 64 pixels. \n\nBy evaluating baselines on the proposed benchmark, the authors observe that embedding-based models tend to perform better when \"incoming tasks contain different classes from one another\" and gradient-based methods tend to perform better when the task classes \"form super-classes of randomly combined categories.\" \n\nThe overall idea is interesting (few-shot + sequential observations); however, it's not clear should one take home after reading this draft. The conclusions seem intuitive and reasonable, but leave the reader with questions about the main findings of the work. \n\nI would take issue with the way the word \"continual learning\" is used. In practice, the author(s) use existing datasets where the instances arrive in a sequential manner (streaming observations). However, this is not quite what they motivate: \"Consider a user in a fast-changing environment who must learn from the many scenarios that are encountered.\" since, in the described sequential setting all the instances belong to the same underlying distribution (the original dataset), even though they're observed sequentially.  \n\nNote that a realistic temporal observations are much more challenging (e.g., the language of search queries over time because of the change in the functionality of search engines, or the changing distribution of images over time because of various social changes.)\n\nThe overall direction is promising: clearly, we need to move towards more data-efficiency and build better frameworks for measuring the generalization of our models. However, I am not convinced if the presented work is a significant step toward that goal (or, at least, I don't see it). Happy to change my mind, if I am missing anything. \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A benchmak for continual few-shot learning ", "review": "The work proposes Continual Few-Shot learning -- a setting to study tasks (1) with a small labeled dataset, and (2) retain knowledge acquired on a sequence of instances.  Additionally, the authors build a compact variant of ImageNet which retains all original 1000 classes but only contains 200 instances of each one (a total of 200K data-points) downscaled to 64 × 64 pixels. \n\nBy evaluating baselines on the proposed benchmark, the authors observe that embedding-based models tend to perform better when \"incoming tasks contain different classes from one another\" and gradient-based methods tend to perform better when the task classes \"form super-classes of randomly combined categories.\" \n\nThe overall idea is interesting (few-shot + sequential observations); however, it's not clear should one take home after reading this draft. The conclusions seem intuitive and reasonable, but leave the reader with questions about the main findings of the work. \n\nI would take issue with the way the word \"continual learning\" is used. In practice, the author(s) use existing datasets where the instances arrive in a sequential manner (streaming observations). However, this is not quite what they motivate: \"Consider a user in a fast-changing environment who must learn from the many scenarios that are encountered.\" since, in the described sequential setting all the instances belong to the same underlying distribution (the original dataset), even though they're observed sequentially.  \n\nNote that a realistic temporal observations are much more challenging (e.g., the language of search queries over time because of the change in the functionality of search engines, or the changing distribution of images over time because of various social changes.)\n\nThe overall direction is promising: clearly, we need to move towards more data-efficiency and build better frameworks for measuring the generalization of our models. However, I am not convinced if the presented work is a significant step toward that goal (or, at least, I don't see it). Happy to change my mind, if I am missing anything. \n", "rating": "4: Ok but not good enough - rejection", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603941364615}, {"id": "e7f96JGjpah", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1737/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary\n--------------------\nThe authors propose a new evaluation protocol which generalizes continual learning and few-shot learning. By controlling the 1) number of support sets per task and 2) the rate at witch tasks change, the authors can span a wide variety of settings previously explored in the literature. \nThe authors proceed to evaluate several meta learning algorithms under this new protocol and provide a detailed analysis of the results.\n\n\nPros\n-------------------\n+ The overall evaluation protocol is well explained. Figure 2 is a beautiful summary of it. \n+ The authors monitor both memory and computation. To me this is one of the key novelties of the paper : it's often overlooked in the literature, yet is very important as one can solve catastrophic forgetting given enough compute\n+ The authors release clear and concise code (and installation instructions). This is a great first step to encourage other practitioners to use this dataset\n+ The results give interesting insights into popular few-shot learning methods like MAML and ProtoNet\n\nCons\n-------------------\n- Table 1 is very hard to read. Please consider either removing some entries to make it bigger (as some baselines perform very poorly), or using another way of presenting it. \n- The evaluation protocol has very little CL baselines, I'm not sure why that is. I don't consider EWC to be a good representative of CL methods, there should at least some replay algorithm. Also please consider adding OML [1] as a baseline, as it seems appropriate to this setting. \n- The memory and MAC results are barely mentioned in the text. I understand this is hard due to space constraints, however a whole page is spent motivating it. \n\nSmall suggestion : The paper would be stronger if it had a wide range of baselines, making it also an empirical study. You constructed a nice benchmark, please show us how (more) known methods perform in it! I'm willing to increase my score if this request is met. \n\nEdit : After reading the appendix, it is mentioned that each model is trained for 250 epochs of 500 steps, where each step is done on a single continual learning task. Here a CL task is a support set (or multiple support sets) from the same distribution ? I just want to double check that a given sample is **only seen during this single continual learning task **, i.e. once this task is over the sample is discarded. Moreover I want to double check that once a task / class is seen, it is never revisited again, even when changing epochs.  I think this is the case, however my whole understanding of the paper rests on this assumption. \n\n\n[1] Javed, Khurram, and Martha White. \"Meta-learning representations for continual learning.\" Advances in Neural Information Processing Systems. 2019.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Paper Review", "review": "Summary\n--------------------\nThe authors propose a new evaluation protocol which generalizes continual learning and few-shot learning. By controlling the 1) number of support sets per task and 2) the rate at witch tasks change, the authors can span a wide variety of settings previously explored in the literature. \nThe authors proceed to evaluate several meta learning algorithms under this new protocol and provide a detailed analysis of the results.\n\n\nPros\n-------------------\n+ The overall evaluation protocol is well explained. Figure 2 is a beautiful summary of it. \n+ The authors monitor both memory and computation. To me this is one of the key novelties of the paper : it's often overlooked in the literature, yet is very important as one can solve catastrophic forgetting given enough compute\n+ The authors release clear and concise code (and installation instructions). This is a great first step to encourage other practitioners to use this dataset\n+ The results give interesting insights into popular few-shot learning methods like MAML and ProtoNet\n\nCons\n-------------------\n- Table 1 is very hard to read. Please consider either removing some entries to make it bigger (as some baselines perform very poorly), or using another way of presenting it. \n- The evaluation protocol has very little CL baselines, I'm not sure why that is. I don't consider EWC to be a good representative of CL methods, there should at least some replay algorithm. Also please consider adding OML [1] as a baseline, as it seems appropriate to this setting. \n- The memory and MAC results are barely mentioned in the text. I understand this is hard due to space constraints, however a whole page is spent motivating it. \n\nSmall suggestion : The paper would be stronger if it had a wide range of baselines, making it also an empirical study. You constructed a nice benchmark, please show us how (more) known methods perform in it! I'm willing to increase my score if this request is met. \n\nEdit : After reading the appendix, it is mentioned that each model is trained for 250 epochs of 500 steps, where each step is done on a single continual learning task. Here a CL task is a support set (or multiple support sets) from the same distribution ? I just want to double check that a given sample is **only seen during this single continual learning task **, i.e. once this task is over the sample is discarded. Moreover I want to double check that once a task / class is seen, it is never revisited again, even when changing epochs.  I think this is the case, however my whole understanding of the paper rests on this assumption. \n\n\n[1] Javed, Khurram, and Martha White. \"Meta-learning representations for continual learning.\" Advances in Neural Information Processing Systems. 2019.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603853051589}, {"id": "dPyk9VhHK1i", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1737/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new machine learning setting called “Continual Few-Shot Learning” which fuses the up until now disparate paradigms of continual learning and few-shot learning. To evaluate methods in this new setting, a new benchmark and dataset called SlimageNet64 are defined. Various methods are evaluated on the new benchmark establishing a set of baseline results for the new setting.\n\n**Pros:**   \n- The idea of fusing the previously separate domains of few-shot learning and continual learning in the form of a new benchmark is fantastic.\n- The benchmarks include metrics for memory usage and computational complexity which are critical in practice and almost always omitted in other benchmarks.\n- The paper is well written and straightforward to understand.\n\n**Concerns:**\n\n(1) The benchmark has significant limitations:\n- The benchmark focuses on classification accuracy, memory usage, and computational complexity metrics, but has no metrics for other fundamental continual learning metrics such as positive forward transfer, positive backward transfer, forgetting over time, and intransigence (i.e. the inability of an algorithm to learn new tasks) [3,4]. These metrics would have to be measured within a task over the course of learning. The proposed benchmark only does evaluation at the end of a task, which is a limitation.\n- Section 3.1 implies that shot and way are fixed across support sets. This is a limitation as realistic few-shot and continual learning scenarios require varying shot and way, including some of the motivating scenarios in Section 1.1.\n- Certain guidance about the benchmark is not mentioned. For example, what is the guidance on pre-training the learner via meta-learning or large-scale supervised classification on large datasets outside the benchmark? For example, [5] demonstrates promising continual learning results on a few-shot variant of split MNIST and split CIFAR using a few-shot learner that has been meta-trained on meta-dataset [2].\n- The 64 x 64 pixel image size choice in the SlimageNet64 dataset is short-sighted. I realize that the intention is to keep the dataset size small, but in the era of low-end smartphones producing 5 million or more pixel images and the ready availability of pre-trained networks at 224 x 224 pixels implies that the benchmark will not have longevity. The images should be available at their natural size and the allow the learner to scale the images down if they are operating in a constrained computational environment.\n\n(2) The experiments are limited:\n- All the experiments were carried out with one fixed configuration (5-way, 1-shot). It would be more revealing if a greater number of configurations were explored, including support sets with random shot and random way.\n- The tasks consisted of a very small number of support sets (NSS between 3 and 10) in the accuracy experiments. However, the memory cost tests used NSS=640 (a more realistic number), why not report accuracy results on those runs?\n- It is peculiar to compare the accuracy of various methods that use very different network backbones in the same table without clearly delineating the differences in the table (i.e. with a horizontal line or column indicating the backbone). I realize that the purpose of these experiments is to establish baselines with various types and levels of learners but the fact that SCA is bolded as the best in the accuracy tables implies that there is a direct (and unfair) comparison being made between methods that should be made within several categories (say low, medium, and high tiers). In any case, dividing up the table into categories and adding detail on architectures used for the various methods would make the presentation of the results more transparent and insightful.\n\n(3) Much detail required for reproducibility of the experiments is missing:\n- No detail is given on how the various approaches in the experiments utilize memory within a task. Section 3.3 says: “Most learners will be compressing a given support set, but this is not strictly the case.” and “(e.g. embedding vectors in ProtoNets, and inner loop parameters for MAML)”. It would be beneficial to describe precisely how memory is used in each of the methods.\n- No details were provided on how the baseline methods were carried out. For example, how was the fine tuning done in terms of network architecture, learning rate, optimization method, number of iterations, any memory that is used, were all weights in the network optimized or just the top layer, etc. The same goes for the other methods.\n- How is $|\\mathcal{G}^x|$ measured? i.e. Is it measured in its tensor form (i.e. 4-byte per channel float) or in its native form (1 unsigned byte per channel)?\n\n**Minor Comments:**  \n- In the introduction, you should mention meta-dataset [2] along with the other few-shot learning benchmarks listed as it is arguably the most challenging/meaningful one at the moment.\n- Section 2.1 should cite [1] as a thorough few-shot learning survey.\n- Section 3.2, the acronym NI is used in the sentence before it is defined.\n- In Section 3.3, Multiply-Addition operations(MACs) paragraph, it says “…it measures the memory footprint that the model”. Should it be something like “…computational complexity of the model”?\n-  Section 3.3, 1st sentence – should it say: “all the task types of interest.” instead of “all the tasks of interest”?\n- Should the title of Figure 1 be “High level overview of a CFSL task” instead of “High level overview of the proposed benchmark”? i.e. ‘benchmark’ is a bit too general as there is no other mention of key parts of the benchmark including memory and computational complexity metrics.\n\n**References:**  \n[1] Hospedales, Timothy, et al. \"Meta-learning in neural networks: A survey.\" arXiv preprint arXiv:2004.05439 (2020).  \n[2] Triantafillou, Eleni, et al. \"Meta-dataset: A dataset of datasets for learning to learn from few examples.\" arXiv preprint arXiv:1903.03096 (2019).  \n[3] Schwarz, Jonathan, et al. \"Progress & compress: A scalable framework for continual learning.\" arXiv preprint arXiv:1805.06370 (2018).  \n[4] Chaudhry, Arslan, et al. \"Riemannian walk for incremental learning: Understanding forgetting and intransigence.\" Proceedings of the European Conference on Computer Vision (ECCV). 2018.  \n[5] Requeima, James, et al. \"Fast and flexible multi-task classification using conditional neural adaptive processes.\" Advances in Neural Information Processing Systems. 2019.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review for \"Defining Benchmarks for Continual Few-Shot Learning\"", "review": "This paper proposes a new machine learning setting called “Continual Few-Shot Learning” which fuses the up until now disparate paradigms of continual learning and few-shot learning. To evaluate methods in this new setting, a new benchmark and dataset called SlimageNet64 are defined. Various methods are evaluated on the new benchmark establishing a set of baseline results for the new setting.\n\n**Pros:**   \n- The idea of fusing the previously separate domains of few-shot learning and continual learning in the form of a new benchmark is fantastic.\n- The benchmarks include metrics for memory usage and computational complexity which are critical in practice and almost always omitted in other benchmarks.\n- The paper is well written and straightforward to understand.\n\n**Concerns:**\n\n(1) The benchmark has significant limitations:\n- The benchmark focuses on classification accuracy, memory usage, and computational complexity metrics, but has no metrics for other fundamental continual learning metrics such as positive forward transfer, positive backward transfer, forgetting over time, and intransigence (i.e. the inability of an algorithm to learn new tasks) [3,4]. These metrics would have to be measured within a task over the course of learning. The proposed benchmark only does evaluation at the end of a task, which is a limitation.\n- Section 3.1 implies that shot and way are fixed across support sets. This is a limitation as realistic few-shot and continual learning scenarios require varying shot and way, including some of the motivating scenarios in Section 1.1.\n- Certain guidance about the benchmark is not mentioned. For example, what is the guidance on pre-training the learner via meta-learning or large-scale supervised classification on large datasets outside the benchmark? For example, [5] demonstrates promising continual learning results on a few-shot variant of split MNIST and split CIFAR using a few-shot learner that has been meta-trained on meta-dataset [2].\n- The 64 x 64 pixel image size choice in the SlimageNet64 dataset is short-sighted. I realize that the intention is to keep the dataset size small, but in the era of low-end smartphones producing 5 million or more pixel images and the ready availability of pre-trained networks at 224 x 224 pixels implies that the benchmark will not have longevity. The images should be available at their natural size and the allow the learner to scale the images down if they are operating in a constrained computational environment.\n\n(2) The experiments are limited:\n- All the experiments were carried out with one fixed configuration (5-way, 1-shot). It would be more revealing if a greater number of configurations were explored, including support sets with random shot and random way.\n- The tasks consisted of a very small number of support sets (NSS between 3 and 10) in the accuracy experiments. However, the memory cost tests used NSS=640 (a more realistic number), why not report accuracy results on those runs?\n- It is peculiar to compare the accuracy of various methods that use very different network backbones in the same table without clearly delineating the differences in the table (i.e. with a horizontal line or column indicating the backbone). I realize that the purpose of these experiments is to establish baselines with various types and levels of learners but the fact that SCA is bolded as the best in the accuracy tables implies that there is a direct (and unfair) comparison being made between methods that should be made within several categories (say low, medium, and high tiers). In any case, dividing up the table into categories and adding detail on architectures used for the various methods would make the presentation of the results more transparent and insightful.\n\n(3) Much detail required for reproducibility of the experiments is missing:\n- No detail is given on how the various approaches in the experiments utilize memory within a task. Section 3.3 says: “Most learners will be compressing a given support set, but this is not strictly the case.” and “(e.g. embedding vectors in ProtoNets, and inner loop parameters for MAML)”. It would be beneficial to describe precisely how memory is used in each of the methods.\n- No details were provided on how the baseline methods were carried out. For example, how was the fine tuning done in terms of network architecture, learning rate, optimization method, number of iterations, any memory that is used, were all weights in the network optimized or just the top layer, etc. The same goes for the other methods.\n- How is $|\\mathcal{G}^x|$ measured? i.e. Is it measured in its tensor form (i.e. 4-byte per channel float) or in its native form (1 unsigned byte per channel)?\n\n**Minor Comments:**  \n- In the introduction, you should mention meta-dataset [2] along with the other few-shot learning benchmarks listed as it is arguably the most challenging/meaningful one at the moment.\n- Section 2.1 should cite [1] as a thorough few-shot learning survey.\n- Section 3.2, the acronym NI is used in the sentence before it is defined.\n- In Section 3.3, Multiply-Addition operations(MACs) paragraph, it says “…it measures the memory footprint that the model”. Should it be something like “…computational complexity of the model”?\n-  Section 3.3, 1st sentence – should it say: “all the task types of interest.” instead of “all the tasks of interest”?\n- Should the title of Figure 1 be “High level overview of a CFSL task” instead of “High level overview of the proposed benchmark”? i.e. ‘benchmark’ is a bit too general as there is no other mention of key parts of the benchmark including memory and computational complexity metrics.\n\n**References:**  \n[1] Hospedales, Timothy, et al. \"Meta-learning in neural networks: A survey.\" arXiv preprint arXiv:2004.05439 (2020).  \n[2] Triantafillou, Eleni, et al. \"Meta-dataset: A dataset of datasets for learning to learn from few examples.\" arXiv preprint arXiv:1903.03096 (2019).  \n[3] Schwarz, Jonathan, et al. \"Progress & compress: A scalable framework for continual learning.\" arXiv preprint arXiv:1805.06370 (2018).  \n[4] Chaudhry, Arslan, et al. \"Riemannian walk for incremental learning: Understanding forgetting and intransigence.\" Proceedings of the European Conference on Computer Vision (ECCV). 2018.  \n[5] Requeima, James, et al. \"Fast and flexible multi-task classification using conditional neural adaptive processes.\" Advances in Neural Information Processing Systems. 2019.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603525060336}, {"id": "FRzCq0rqL7v", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1737/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "----------------------------------\n**Summary**\n\nThis paper proposes a benchmark for a new task called continual few-shot learning. The benchmark is based on the ImageNet dataset. Basically, the model looks at a part of the support set one after another sequentially, and it is then evaluated on the query set that contains balanced samples from each part of the support set. Under the benchmark, there are four types of challenges, which differs in how they sub-partition the support set. A suite of models have been run and evaluated, including MAML, ProtoNet and SCA. SCA is found to have the best performance, whereas ProtoNet is found to be the most resource efficient.\n\n----------------------------------\n**Strengths**\n\n1. I agree that continual few-shot learning is a very useful way of framing few-shot learning problems, especially in the domain of general purpose robotics.\n\n2. Well defined metrics that consider both accuracy and computation resources.\n\n3. There is a significant amount of effort in terms of defining four different scenarios and coming up with another split of the ImageNet dataset and running a number of models on the benchmark.\n\n----------------------------------\n**Weaknesses**\n\n1. Limitations in the setup. \n\n    a) Although the paper motivates the application of continual few-shot learning, the proposed benchmark seems rather far away from application. The two applications mentioned in the paper are 1) online shopping for user preference and 2) human-robot interaction for task learning. Why not directly target these settings? \n\n    b) Although the paper is named for defining the benchmark for continual few-shot learning, it is mainly targeting object classification, but not learning new tasks (which is motivated in the human-robot interaction). \n\n    c) The setups in Figure 2 don't seem very natural. Why wouldn’t the model see any old classes in B? Why would the sequence completely overwrite all the class definitions in C? C doesn’t fully capture another very interesting yet unexplored area of domain adaptation, e.g. change the image style from real to synthetic images over time. I can see it is partially addressed in OSAKA between meta-training and meta-test but not in the sequence.\n\n    d) Why is the query set only happening at the very end but not in the middle of the sequence?\n\n2. A lack of new components in the benchmark. The benchmark seems very similar to previous benchmarks (e.g. class incremental learning and CORe50) with the main difference being low-data. According to Table 3, the only setting missing in CORe50 is setting C, which is a rather limited setting since it is extremely unnatural that we need to completely overwrite a label for another class (see my first comment). So what prevents people from using less number of images per class in CORe50, with a possibly pretrained representation from somewhere else, if they want to study a low-shot setting? Pretrained representation from supervised classification seems like a pretty robust approach for few-shot learning based on recent studies (Chen et al., 2019).\n\n3. From an implementation point of view, the proposed benchmark seems no different from a few newly defined mini-batch samplers for regular few-shot learning episodes. If it is no different from a few newly defined mini-batch samplers, then what prevents a model from storing all the examples in the episode? Since this is low-data, it doesn’t seem to have a huge memory cost. Let’s say for each class we store a fixed constant K exemplars, then the overall memory storage is still on the same order as ProtoNet, which stores a single mean vector per class. At the very least, the author should provide results for such a model, as an offline oracle.\n\n4. A lack of models evaluated. \n\n    a) If it can potentially include models that store some of the examples, then there seems no reason to exclude other baselines such as MatchingNet (Vinyals et al., 2016), infinite mixture prototypes (Allen et al., 2019), which can probably address the challenges that ProtoNet faces when dealing with type C sequences. \n\n    b) There have also been a bunch of incremental class learning baselines that can be covered (which doesn’t require meta-learning). For example, iCaRL (Rebuffi et al., 2017), LwF (Li & Hoiem, 2018), BiC (Wu et al., 2019), PODNet (Douillard et al., 2020). Currently I only see EWC as a pretty weak baseline in continual learning and it wasn’t entirely designed for class incremental learning.\n\n    c) I do notice that there are Pre. & Tune and EWC-Pre. baselines that use pretrained representation. This is good. However, it doesn’t explain whether their failures are due to the sequential gradient descent steps or the lack of generalizability of the representation. I suspect it is probably the first case, and have the authors considered using pretrained representation and then using ProtoNet or some other FSL methods at test time?\n\n----------------------------------\n**Minor comments**\n\n1. 64x64 seems rather small resolution. I would suggest at least 84x84 for ImageNet images.\n2. The memory utilization results should go into the main paper, along with other baselines I suggested that use some example-based storage.\n3. Section 2.3, Page 4, Line 13: thie line -> this line.\n\n----------------------------------\n**Conclusion**\n\nIn conclusion, based on the weaknesses I mentioned above, my score is 5. The paper is one step towards a more useful type of few-shot learning, and I am glad to see the field progresses towards it and I appreciate the effort that the authors have contributed. However, the way it samples new classes and the datasets that the paper study are still in a very limited sense, which prevent the paper from getting direct applications. On the other hand, it is very similar to some other incremental class learning benchmarks and it is not clear what different kinds of conclusions we can draw (e.g. different models that may work better) by using this benchmark instead of using previous ones. Therefore I am leaning towards rejection.\n\n----------------------------------\n**References**\n\n- Li, Zhizhong and Hoiem, Derek. Learning without forgetting. In ECCV 2016.\n- Vinyals, Oriol, Blundell, Charles, Lillicrap, timothy, Kavukcuoglu, Koray and Wierstra, Daan. Matching networks for one shot learning. In NIPS 2016.\n- Rebuffi, Sylvestre-Alvise, Kolesnikov, Alexander, Sperlm, Georg and Lampert, Christoph H. iCaRL: Incremental classifier and representation learning. In CVPR 2017.\n- Chen, Wei-Yu, Liu, Yen-Cheng, Kira, Zsolt, Wang, Yu-Chiang and  Huang, Jia-Bin. A closer look at few-shot classification. In ICLR 2019.\n- Wu, Yue, Chen, Yinpeng, Wang, Lijuan, Ye, Yuancheng, Liu, Zicheng, Guo, Yandong and Fu, Yun. Large scale incremental learning. In CVPR 2019.\n- Allenn, Kelsey R., Shelhamer, Evan, Shin, Hanul and Tenenbaum, Joshua B. Infinite mixture prototypes for few-shot learning. In ICML 2019.\n- Douillard, Arthur, Cord, Matthieu, Ollion, Charles, Robert, Thomas and Valle, Eduardo. PODNet: Pooled outputs distillation for small-tasks incremental learning. In ECCV 2020.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "----------------------------------\n**Summary**\n\nThis paper proposes a benchmark for a new task called continual few-shot learning. The benchmark is based on the ImageNet dataset. Basically, the model looks at a part of the support set one after another sequentially, and it is then evaluated on the query set that contains balanced samples from each part of the support set. Under the benchmark, there are four types of challenges, which differs in how they sub-partition the support set. A suite of models have been run and evaluated, including MAML, ProtoNet and SCA. SCA is found to have the best performance, whereas ProtoNet is found to be the most resource efficient.\n\n----------------------------------\n**Strengths**\n\n1. I agree that continual few-shot learning is a very useful way of framing few-shot learning problems, especially in the domain of general purpose robotics.\n\n2. Well defined metrics that consider both accuracy and computation resources.\n\n3. There is a significant amount of effort in terms of defining four different scenarios and coming up with another split of the ImageNet dataset and running a number of models on the benchmark.\n\n----------------------------------\n**Weaknesses**\n\n1. Limitations in the setup. \n\n    a) Although the paper motivates the application of continual few-shot learning, the proposed benchmark seems rather far away from application. The two applications mentioned in the paper are 1) online shopping for user preference and 2) human-robot interaction for task learning. Why not directly target these settings? \n\n    b) Although the paper is named for defining the benchmark for continual few-shot learning, it is mainly targeting object classification, but not learning new tasks (which is motivated in the human-robot interaction). \n\n    c) The setups in Figure 2 don't seem very natural. Why wouldn’t the model see any old classes in B? Why would the sequence completely overwrite all the class definitions in C? C doesn’t fully capture another very interesting yet unexplored area of domain adaptation, e.g. change the image style from real to synthetic images over time. I can see it is partially addressed in OSAKA between meta-training and meta-test but not in the sequence.\n\n    d) Why is the query set only happening at the very end but not in the middle of the sequence?\n\n2. A lack of new components in the benchmark. The benchmark seems very similar to previous benchmarks (e.g. class incremental learning and CORe50) with the main difference being low-data. According to Table 3, the only setting missing in CORe50 is setting C, which is a rather limited setting since it is extremely unnatural that we need to completely overwrite a label for another class (see my first comment). So what prevents people from using less number of images per class in CORe50, with a possibly pretrained representation from somewhere else, if they want to study a low-shot setting? Pretrained representation from supervised classification seems like a pretty robust approach for few-shot learning based on recent studies (Chen et al., 2019).\n\n3. From an implementation point of view, the proposed benchmark seems no different from a few newly defined mini-batch samplers for regular few-shot learning episodes. If it is no different from a few newly defined mini-batch samplers, then what prevents a model from storing all the examples in the episode? Since this is low-data, it doesn’t seem to have a huge memory cost. Let’s say for each class we store a fixed constant K exemplars, then the overall memory storage is still on the same order as ProtoNet, which stores a single mean vector per class. At the very least, the author should provide results for such a model, as an offline oracle.\n\n4. A lack of models evaluated. \n\n    a) If it can potentially include models that store some of the examples, then there seems no reason to exclude other baselines such as MatchingNet (Vinyals et al., 2016), infinite mixture prototypes (Allen et al., 2019), which can probably address the challenges that ProtoNet faces when dealing with type C sequences. \n\n    b) There have also been a bunch of incremental class learning baselines that can be covered (which doesn’t require meta-learning). For example, iCaRL (Rebuffi et al., 2017), LwF (Li & Hoiem, 2018), BiC (Wu et al., 2019), PODNet (Douillard et al., 2020). Currently I only see EWC as a pretty weak baseline in continual learning and it wasn’t entirely designed for class incremental learning.\n\n    c) I do notice that there are Pre. & Tune and EWC-Pre. baselines that use pretrained representation. This is good. However, it doesn’t explain whether their failures are due to the sequential gradient descent steps or the lack of generalizability of the representation. I suspect it is probably the first case, and have the authors considered using pretrained representation and then using ProtoNet or some other FSL methods at test time?\n\n----------------------------------\n**Minor comments**\n\n1. 64x64 seems rather small resolution. I would suggest at least 84x84 for ImageNet images.\n2. The memory utilization results should go into the main paper, along with other baselines I suggested that use some example-based storage.\n3. Section 2.3, Page 4, Line 13: thie line -> this line.\n\n----------------------------------\n**Conclusion**\n\nIn conclusion, based on the weaknesses I mentioned above, my score is 5. The paper is one step towards a more useful type of few-shot learning, and I am glad to see the field progresses towards it and I appreciate the effort that the authors have contributed. However, the way it samples new classes and the datasets that the paper study are still in a very limited sense, which prevent the paper from getting direct applications. On the other hand, it is very similar to some other incremental class learning benchmarks and it is not clear what different kinds of conclusions we can draw (e.g. different models that may work better) by using this benchmark instead of using previous ones. Therefore I am leaning towards rejection.\n\n----------------------------------\n**References**\n\n- Li, Zhizhong and Hoiem, Derek. Learning without forgetting. In ECCV 2016.\n- Vinyals, Oriol, Blundell, Charles, Lillicrap, timothy, Kavukcuoglu, Koray and Wierstra, Daan. Matching networks for one shot learning. In NIPS 2016.\n- Rebuffi, Sylvestre-Alvise, Kolesnikov, Alexander, Sperlm, Georg and Lampert, Christoph H. iCaRL: Incremental classifier and representation learning. In CVPR 2017.\n- Chen, Wei-Yu, Liu, Yen-Cheng, Kira, Zsolt, Wang, Yu-Chiang and  Huang, Jia-Bin. A closer look at few-shot classification. In ICLR 2019.\n- Wu, Yue, Chen, Yinpeng, Wang, Lijuan, Ye, Yuancheng, Liu, Zicheng, Guo, Yandong and Fu, Yun. Large scale incremental learning. In CVPR 2019.\n- Allenn, Kelsey R., Shelhamer, Evan, Shin, Hanul and Tenenbaum, Joshua B. Infinite mixture prototypes for few-shot learning. In ICML 2019.\n- Douillard, Arthur, Cord, Matthieu, Ollion, Charles, Robert, Thomas and Valle, Eduardo. PODNet: Pooled outputs distillation for small-tasks incremental learning. In ECCV 2020.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603507448481}], "openreview_url": "https://openreview.net/forum?id=XG1Drw7VbLJ", "arxiv_id": "2004.11967", "paper_pdf": "papers/XG1Drw7VbLJ.pdf", "paper_pdf_sha256": "485efc10dcca6d91e8c84d41e64f977ac061289cbdaa2abd54b2666a74138cbc", "paper_pdf_bytes": 1703999, "paper_pdf_source": "openreview", "code_url": "https://github.com/AntreasAntoniou/FewShotContinualLearning", "code_repository": "AntreasAntoniou/FewShotContinualLearning", "code_commit": "819b9cc26ef9d2360a040c51f17958e1b8dba8fd", "code_archive": "repos/XG1Drw7VbLJ.zip", "code_archive_sha256": "2ea8a735c9dfabd031f26c81d0868989240fe2d83b3d3589450fe957cbfe12c3", "code_archive_bytes": 937410, "code_file_count": 448, "code_extensions": {".sh": 435, ".py": 13}, "github_disk_usage_kb": 5401, "github_languages": {"Python": 336297, "Shell": 160045}, "github_archived": false, "github_pushed_at": "2020-08-18T12:00:15Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/defining-benchmarks-for-continual-few-shot"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "pjfrGVekwK", "year": 2025, "status": "rejected", "title": "Variational Bayes Gaussian Splatting", "authors": ["Toon Van de Maele", "Ozan Catal", "Alexander Tschantz", "Christopher Buckley", "Tim Verbelen"], "authorids": ["~Toon_Van_de_Maele1", "~Ozan_Catal2", "~Alexander_Tschantz2", "~Christopher_Buckley1", "~Tim_Verbelen1"], "authors_source": "OpenReview API", "abstract": "Recently, 3D Gaussian Splatting has emerged as a promising approach for modeling 3D scenes using mixtures of Gaussians. The predominant optimization method for these models relies on backpropagating gradients through a differentiable rendering pipeline, which struggles with catastrophic forgetting when dealing with continuous streams of data. To address this limitation, we propose Variational Bayes Gaussian Splatting (VBGS), a novel approach that frames training a Gaussian splat as variational inference over model parameters. By leveraging the conjugacy properties of multivariate Gaussians, we derive a closed-form variational update rule, allowing efficient updates from partial, sequential observations without the need for replay buffers. Our experiments show that VBGS not only matches state-of-the-art performance on static datasets, but also enables continual learning from sequentially streamed 2D and 3D data, drastically improving performance in this setting.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "sBLdK0kGXz", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6796/Reviewer_WzD1"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 4, "summary": "This paper proposes Variational Bayes Gaussian Splatting (VBGS) for modeling 3D scenes by framing training of Gaussian Splats as variational inference. The formulation enables closed-form variational update of model parameters which is particularly beneficial in online settings. In particular, updating the model from a sequence of observations is equivalent to processing all data in a single batch. Thus, catastrophic forgetting is circumvented.\n\nExperimental results indicate VBGS matches the performance of 3D Gaussian Splatting (3DGS) on static datasets while enabling continual learning with streamed 3D data.", "review_text": "This paper proposes Variational Bayes Gaussian Splatting (VBGS) for modeling 3D scenes by framing training of Gaussian Splats as variational inference. The formulation enables closed-form variational update of model parameters which is particularly beneficial in online settings. In particular, updating the model from a sequence of observations is equivalent to processing all data in a single batch. Thus, catastrophic forgetting is circumvented.\n\nExperimental results indicate VBGS matches the performance of 3D Gaussian Splatting (3DGS) on static datasets while enabling continual learning with streamed 3D data.", "strengths": "**S1.** The method addresses an important limitation of 3DGS.\n\n**S2.** The method is principled and presented with clarity.\n\n**S3.** The experimental validation lends support to the practical usefulness of the method: in particular, in the online setting 3DGS is clearly impacted by catastrophic forgetting while VBGS is not.", "weaknesses": "**W1.** The method is motivated by the application to continual learning scenarios such as SLAM. However, the experimental validation does not compare against recent SLAM methods or datasets, e.g., (Matsuki et al., 2024; Keetha et al., 2024).\n\n**W2.** The significance of the mean-field approximation is not discussed or investigated experimentally.\n\n**W3.** The method's reliance on depth maps is, as acknowledged by the authors, a limitation of the method. The author’s propose inferring depth maps from RGB data using a pretrained model. Experiments along these lines would have been interesting to include.", "questions": "**Q1:** Why not evaluate using datasets and protocols from recent SLAM works such as (Matsuki et al., 2024; Keetha et al., 2024)?\n\n**Q2.** When should one expect the mean-field approximation to hurt performance?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Variational Bayes Gaussian Splatting (VBGS) for modeling 3D scenes by framing training of Gaussian Splats as variational inference. The formulation enables closed-form variational update of model parameters which is particularly beneficial in online settings. In particular, updating the model from a sequence of observations is equivalent to processing all data in a single batch. Thus, catastrophic forgetting is circumvented.\n\nExperimental results indicate VBGS matches the performance of 3D Gaussian Splatting (3DGS) on static datasets while enabling continual learning with streamed 3D data.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**S1.** The method addresses an important limitation of 3DGS.\n\n**S2.** The method is principled and presented with clarity.\n\n**S3.** The experimental validation lends support to the practical usefulness of the method: in particular, in the online setting 3DGS is clearly impacted by catastrophic forgetting while VBGS is not.", "weaknesses": "**W1.** The method is motivated by the application to continual learning scenarios such as SLAM. However, the experimental validation does not compare against recent SLAM methods or datasets, e.g., (Matsuki et al., 2024; Keetha et al., 2024).\n\n**W2.** The significance of the mean-field approximation is not discussed or investigated experimentally.\n\n**W3.** The method's reliance on depth maps is, as acknowledged by the authors, a limitation of the method. The author’s propose inferring depth maps from RGB data using a pretrained model. Experiments along these lines would have been interesting to include.", "questions": "**Q1:** Why not evaluate using datasets and protocols from recent SLAM works such as (Matsuki et al., 2024; Keetha et al., 2024)?\n\n**Q2.** When should one expect the mean-field approximation to hurt performance?", "flag_for_ethics_review": ["No ethics review needed."], "details_of_ethics_concerns": "NA.", "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730751093945}, {"id": "pKQB4YTH99", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6796/Reviewer_ivJV"], "rating": 5, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 5, "summary": "The paper proposes a reformulation of 3D Gaussian splatting with variational inference framework. This allows a closed-form update rule and applicability to continual learning framework. The model is compared with gradient-based update rule in small-scale (~100k) 3DGS fitting to Habitat test suite and NeRF synthetic dataset and shows performance boost with better robustness to randomized initialization.", "review_text": "The paper proposes a reformulation of 3D Gaussian splatting with variational inference framework. This allows a closed-form update rule and applicability to continual learning framework. The model is compared with gradient-based update rule in small-scale (~100k) 3DGS fitting to Habitat test suite and NeRF synthetic dataset and shows performance boost with better robustness to randomized initialization.", "strengths": "The main strengths I find in this paper lie in its novelty, clarity in writing, and reproducibility.\n\n- Both relating 3D Gaussian splatting as Gaussian mixture model and variational inference is new to this field, and this novelty should be acknowledged.\n- Application of continual learning well justifies the usage of variational inference framework.\n- The training scheme is described in detail and the described motivation and methodology is quite clear.\n- The supplementary material has the code implementation for reproducibility.", "weaknesses": "Although I find the method informative, there are several points to be clarified in order to improve the presentation.\n\n- Common 3DGS involves several millions of Gaussians, but the presented experiments are only based on 100k Gaussians, which is very slim. This small parameter constraint makes the PSNR scores in Table 1 way below the numbers reported by other NeRF/3DGS papers which go high above 30 dB for the same NeRF synthetic dataset. Is there any specific reason (e.g., computational burden) that constraints the experiments? If so, this should be specified.\n- Unlike the authors’ claim in Appendix D.3, Table 6 shows that the presented VBGS shows comparable or worse results than Gradient-based method, which needs further clarification.\n- It seems like the method does not incorporate cloning/pruning/splitting of Gaussians which take crucial roles in the original 3DGS. By scaling up the Gaussians during the training, the performance can be boosted significantly, so unless there is a counterpart in VBGS for resizing the Gaussian pool, I believe there is a little advantage of using this framework in practical scenarios. I believe there should be an experiment with VBGS having initial Gaussians of the same size of the final Gaussians in Table 5, so we can compare fairly. For example, if we start from 456K Gaussians in VBGS and performs better than dynamically resized 3DGS in Table 5, we can clearly say that there is a reason to migrate from 3DGS to VBGS.\n- I believe it is better to have a dedicated discussion on changes in computational demand relative to the size of Gaussians in different frameworks. Inference time should be the same, but training time and the memory requirement will be very informative to followers to this work.\n\nIn summary, unless there is a clear advantage of VBGS over standard size-varying gradient-based 3DGS, the method is not practically persuasive, yet. Furthermore, currently, there is no guarantee of scalability, which I believe important in learnable 3D representation. It is OK to have a slow and memory-intensive method if the performance (PSNR/SSIM) is stronger, but regarding Table 6, the presented method is not yet persuasive. Regarding these, the current version of the manuscript seems not ready to be published despite its novel and intuitive idea. Please note that I am flexible with my score, and looking forward to further discussion in the rebuttal phase.", "questions": "These are minor concerns that are not counted in my final scores.\n\n- Having only three components in Figure 2.a without Gradient (Data init) seems incomplete. Please have it updated.\n- As the authors have demonstrated with a continual learning framework, is it workable to scene in motion?\n- Does this method scale up? Can this method scale up to tens of millions of Gaussian splatting for larger datasets (e.g., MipNeRF-360)?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a reformulation of 3D Gaussian splatting with variational inference framework. This allows a closed-form update rule and applicability to continual learning framework. The model is compared with gradient-based update rule in small-scale (~100k) 3DGS fitting to Habitat test suite and NeRF synthetic dataset and shows performance boost with better robustness to randomized initialization.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "The main strengths I find in this paper lie in its novelty, clarity in writing, and reproducibility.\n\n- Both relating 3D Gaussian splatting as Gaussian mixture model and variational inference is new to this field, and this novelty should be acknowledged.\n- Application of continual learning well justifies the usage of variational inference framework.\n- The training scheme is described in detail and the described motivation and methodology is quite clear.\n- The supplementary material has the code implementation for reproducibility.", "weaknesses": "Although I find the method informative, there are several points to be clarified in order to improve the presentation.\n\n- Common 3DGS involves several millions of Gaussians, but the presented experiments are only based on 100k Gaussians, which is very slim. This small parameter constraint makes the PSNR scores in Table 1 way below the numbers reported by other NeRF/3DGS papers which go high above 30 dB for the same NeRF synthetic dataset. Is there any specific reason (e.g., computational burden) that constraints the experiments? If so, this should be specified.\n- Unlike the authors’ claim in Appendix D.3, Table 6 shows that the presented VBGS shows comparable or worse results than Gradient-based method, which needs further clarification.\n- It seems like the method does not incorporate cloning/pruning/splitting of Gaussians which take crucial roles in the original 3DGS. By scaling up the Gaussians during the training, the performance can be boosted significantly, so unless there is a counterpart in VBGS for resizing the Gaussian pool, I believe there is a little advantage of using this framework in practical scenarios. I believe there should be an experiment with VBGS having initial Gaussians of the same size of the final Gaussians in Table 5, so we can compare fairly. For example, if we start from 456K Gaussians in VBGS and performs better than dynamically resized 3DGS in Table 5, we can clearly say that there is a reason to migrate from 3DGS to VBGS.\n- I believe it is better to have a dedicated discussion on changes in computational demand relative to the size of Gaussians in different frameworks. Inference time should be the same, but training time and the memory requirement will be very informative to followers to this work.\n\nIn summary, unless there is a clear advantage of VBGS over standard size-varying gradient-based 3DGS, the method is not practically persuasive, yet. Furthermore, currently, there is no guarantee of scalability, which I believe important in learnable 3D representation. It is OK to have a slow and memory-intensive method if the performance (PSNR/SSIM) is stronger, but regarding Table 6, the presented method is not yet persuasive. Regarding these, the current version of the manuscript seems not ready to be published despite its novel and intuitive idea. Please note that I am flexible with my score, and looking forward to further discussion in the rebuttal phase.", "questions": "These are minor concerns that are not counted in my final scores.\n\n- Having only three components in Figure 2.a without Gradient (Data init) seems incomplete. Please have it updated.\n- As the authors have demonstrated with a continual learning framework, is it workable to scene in motion?\n- Does this method scale up? Can this method scale up to tens of millions of Gaussian splatting for larger datasets (e.g., MipNeRF-360)?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 5, "code_of_conduct": "Yes"}, "tcdate": 1730697257968}, {"id": "klWDIkV7mu", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6796/Reviewer_f4Be"], "rating": 5, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "This paper introduces Variational Bayes Gaussian Splatting (VBGS), a new method for modeling 3D scenes with Gaussian splats that addresses the challenge of catastrophic forgetting in continuous data streams. Unlike traditional methods that rely on backpropagation through a differentiable renderer and require replay buffers, VBGS formulates training as variational inference over model parameters, enabling efficient updates through a closed-form variational update rule.", "review_text": "This paper introduces Variational Bayes Gaussian Splatting (VBGS), a new method for modeling 3D scenes with Gaussian splats that addresses the challenge of catastrophic forgetting in continuous data streams. Unlike traditional methods that rely on backpropagation through a differentiable renderer and require replay buffers, VBGS formulates training as variational inference over model parameters, enabling efficient updates through a closed-form variational update rule.", "strengths": "1. Effectively handles continuous data streams: By framing the training process as variational inference over model parameters, VBGS avoids the issue of catastrophic forgetting typically seen in continuous data streams with traditional methods.\n\n2. No need for replay buffers: VBGS leverages the conjugacy properties of multivariate Gaussians, deriving a closed-form variational update rule that allows efficient incremental updates without requiring replay buffers.", "weaknesses": "1. Limited improvement. In Tab. 1, your method is comparable to the Gradient method (3DGS), which cannot prove the effectiveness of your method.\n\n2. Unlike 3D Gaussian Splatting (3DGS), it relies on RGBD data. It means your method needs a depth camera and does not outperform 3DGS.\n\n3. The experiments are not comprehensive, as they do not include evaluations on standard datasets like mip-NeRF 360, which could provide a more robust comparison with existing methods.\n\n4. The comparison with baselines is limited; a more thorough evaluation would include comparisons with established methods such as Mip-NeRF 360, Instant-NGP, and NeRF. Including these methods as baselines would provide a clearer understanding of VBGS's performance relative to a broader range of state-of-the-art approaches.", "questions": "As stated in the weaknesses above, I have some additional questions as follows:\n1. Why is it said that 3DGS struggles with catastrophic forgetting when dealing with continuous streams of data?\n\n2. Could you provide numerical results on the HABITAT ROOMS dataset?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Variational Bayes Gaussian Splatting (VBGS), a new method for modeling 3D scenes with Gaussian splats that addresses the challenge of catastrophic forgetting in continuous data streams. Unlike traditional methods that rely on backpropagation through a differentiable renderer and require replay buffers, VBGS formulates training as variational inference over model parameters, enabling efficient updates through a closed-form variational update rule.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1. Effectively handles continuous data streams: By framing the training process as variational inference over model parameters, VBGS avoids the issue of catastrophic forgetting typically seen in continuous data streams with traditional methods.\n\n2. No need for replay buffers: VBGS leverages the conjugacy properties of multivariate Gaussians, deriving a closed-form variational update rule that allows efficient incremental updates without requiring replay buffers.", "weaknesses": "1. Limited improvement. In Tab. 1, your method is comparable to the Gradient method (3DGS), which cannot prove the effectiveness of your method.\n\n2. Unlike 3D Gaussian Splatting (3DGS), it relies on RGBD data. It means your method needs a depth camera and does not outperform 3DGS.\n\n3. The experiments are not comprehensive, as they do not include evaluations on standard datasets like mip-NeRF 360, which could provide a more robust comparison with existing methods.\n\n4. The comparison with baselines is limited; a more thorough evaluation would include comparisons with established methods such as Mip-NeRF 360, Instant-NGP, and NeRF. Including these methods as baselines would provide a clearer understanding of VBGS's performance relative to a broader range of state-of-the-art approaches.", "questions": "As stated in the weaknesses above, I have some additional questions as follows:\n1. Why is it said that 3DGS struggles with catastrophic forgetting when dealing with continuous streams of data?\n\n2. Could you provide numerical results on the HABITAT ROOMS dataset?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730555712263}, {"id": "OfML73xflB", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission6796/Reviewer_Yuyf"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 3, "summary": "This paper proposes Variational Bayes Gaussian Splatting (VBGS), which frames training a Gaussian splat as variational inference over model parameters.", "review_text": "This paper proposes Variational Bayes Gaussian Splatting (VBGS), which frames training a Gaussian splat as variational inference over model parameters.", "strengths": "1 The paper is well-structured.\n2 The idea of introducing the conjugacy properties of multivariate Gaussians seems to make sense.", "weaknesses": "1 In the introduction, and background (e.g., 3D Gaussian Splatting) and research gap are not detailed enough, so the author's research motivation cannot be understood clearly.\n2 The novelty is unclear and insufficient, it's confused why the proposed method can eliminate the need for replay buffers. More relevant discussions and technical details should be analyzed.\n3 The authors claim that this work can \"more robust and efficient parameter estimation.\" However, no computational efficiency analysis (e.g., computational time, memory usage) is provided.\n4 In the experiments, the baseline only includes \"Gradient\". More discussions are need to explain why choosing\"Gradient\". Besides, the superiority of the proposed method over \"Gradient\" needs more analysis from both theoretical and experimental perspectives. Are there any more recent and advanced baselines?\n5 The evaluation metrics should be introduced in detail, e.g., PSNR, to enhance the understandability and readability of the experimental part.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes Variational Bayes Gaussian Splatting (VBGS), which frames training a Gaussian splat as variational inference over model parameters.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "1 The paper is well-structured.\n2 The idea of introducing the conjugacy properties of multivariate Gaussians seems to make sense.", "weaknesses": "1 In the introduction, and background (e.g., 3D Gaussian Splatting) and research gap are not detailed enough, so the author's research motivation cannot be understood clearly.\n2 The novelty is unclear and insufficient, it's confused why the proposed method can eliminate the need for replay buffers. More relevant discussions and technical details should be analyzed.\n3 The authors claim that this work can \"more robust and efficient parameter estimation.\" However, no computational efficiency analysis (e.g., computational time, memory usage) is provided.\n4 In the experiments, the baseline only includes \"Gradient\". More discussions are need to explain why choosing\"Gradient\". Besides, the superiority of the proposed method over \"Gradient\" needs more analysis from both theoretical and experimental perspectives. Are there any more recent and advanced baselines?\n5 The evaluation metrics should be introduced in detail, e.g., PSNR, to enhance the understandability and readability of the experimental part.", "questions": "Please refer to the weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730093879263}], "openreview_url": "https://openreview.net/forum?id=pjfrGVekwK", "arxiv_id": "2410.03592", "paper_pdf": "papers/pjfrGVekwK.pdf", "paper_pdf_sha256": "839a8697315880151fea95573a9d113d2cd1419ca17078de736876bc8abe9119", "paper_pdf_bytes": 49753355, "paper_pdf_source": "openreview", "code_url": "https://github.com/VersesTech/vbgs", "code_repository": "VersesTech/vbgs", "code_commit": "2ae3f4bea6ed3a5d69271c0e2a67322c06c09b9e", "code_archive": "repos/pjfrGVekwK.zip", "code_archive_sha256": "c0149126d32b7301091bec2b96e4035c1594e70ab93474a0de82254a4620e2c9", "code_archive_bytes": 156422, "code_file_count": 41, "code_extensions": {".py": 40, ".sh": 1}, "github_disk_usage_kb": 438, "github_languages": {"Python": 219948, "Shell": 534}, "github_archived": false, "github_pushed_at": "2025-04-11T15:56:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/variational-bayes-gaussian-splatting"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "BMw4Cm0gGO", "year": 2024, "status": "rejected", "title": "C-MCTS: Safe Planning with Monte Carlo Tree Search", "authors": ["Dinesh Parthasarathy", "Georgios Kontes", "Axel Plinge", "Christopher Mutschler"], "authorids": ["~Dinesh_Parthasarathy1", "~Georgios_Kontes1", "~Axel_Plinge1", "~Christopher_Mutschler1"], "authors_source": "OpenReview API", "abstract": "The Constrained Markov Decision Process (CMDP) allows to solve safety-critical decision making tasks that are subject to constraints. \nWhile CMDPs have been extensively studied in the Reinforcement Learning literature, little attention has been given to sampling-based planning algorithms such as MCTS for solving them. Previous approaches perform conservatively with respect to costs as they avoid constraint violations by using Monte Carlo cost estimates that suffer from high variance. We propose Constrained MCTS (C-MCTS), which estimates cost using a safety critic that is trained with Temporal Difference learning in an offline phase prior to agent deployment. The critic limits exploration by pruning unsafe trajectories within MCTS during deployment. C-MCTS satisfies cost constraints but operates closer to the constraint boundary, achieving higher rewards than previous work. As a nice byproduct, the planner is more efficient w.r.t. planning steps. Most importantly, under model mismatch between the planner and the real world, C-MCTS is less susceptible to cost violations than previous work.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qYir40Oi0J", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4707/Reviewer_q5zX"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper proposes a novel algorithm, C-MCTS, for solving problems with constraints.\nThe proposed algorithm is tested on two small benchmark problems and shows improving performances compared to the baseline algorithms, CC-MCP and vanilla MCTS.", "review_text": "This paper proposes a novel algorithm, C-MCTS, for solving problems with constraints.\nThe proposed algorithm is tested on two small benchmark problems and shows improving performances compared to the baseline algorithms, CC-MCP and vanilla MCTS.", "strengths": "- C-MCTS achieves improving performance in the quality of the solutions while not violating the cost constraints.", "weaknesses": "- The actual running time of the experiments needs to be provided.\n\n- Several points need to be clarified in the explanation that is described in the Questions below.", "questions": "- About proof of Proposition 1 in P.4.\n\"... guaranteed to find the optimal solution in the k-th MDP.\"\nUCT's proof is asymptotic, and we need an infinite number of simulations to guarantee the convergence to the optimal.\nAlso, we have no way of knowing whether we actually reached the optimal (in my understanding).\nDoes this proof assume that we do an infinite number of simulations on each iteration? Is that valid?\n\n- Also, I may have overlooked, but how long does this step take in the actual experiments in Section 4? (is it the \"Solve MDP\" step in Fig. 1?) \n\n- How meaningful is it to count the number of cost violations?\nIn what problem settings are we not allowed to violate the cost even during the simulations?\n\n- For Fig. 2, how was the actual execution time of the three algorithms?\nAlso, for the Safe GridWorld experiment, did C-MCTS with $2^9$ planning iterations and CC-MCP with $2^{20}$ planning iterations take a similar amount of time?\n\n- About Fig. 4.\nWhat is one episode here?\nDoes one episode for C-MCTS and CC-MCP require a similar amount of time?\nIf the latter is faster, could we run CC-MCP 100 times and select the path based on the majority?\n\nMinor question.\n- I got confused with the notation of $C, c, \\hat{c}$. Are these sets or functions?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper proposes a novel algorithm, C-MCTS, for solving problems with constraints.\nThe proposed algorithm is tested on two small benchmark problems and shows improving performances compared to the baseline algorithms, CC-MCP and vanilla MCTS.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "- C-MCTS achieves improving performance in the quality of the solutions while not violating the cost constraints.", "weaknesses": "- The actual running time of the experiments needs to be provided.\n\n- Several points need to be clarified in the explanation that is described in the Questions below.", "questions": "- About proof of Proposition 1 in P.4.\n\"... guaranteed to find the optimal solution in the k-th MDP.\"\nUCT's proof is asymptotic, and we need an infinite number of simulations to guarantee the convergence to the optimal.\nAlso, we have no way of knowing whether we actually reached the optimal (in my understanding).\nDoes this proof assume that we do an infinite number of simulations on each iteration? Is that valid?\n\n- Also, I may have overlooked, but how long does this step take in the actual experiments in Section 4? (is it the \"Solve MDP\" step in Fig. 1?) \n\n- How meaningful is it to count the number of cost violations?\nIn what problem settings are we not allowed to violate the cost even during the simulations?\n\n- For Fig. 2, how was the actual execution time of the three algorithms?\nAlso, for the Safe GridWorld experiment, did C-MCTS with $2^9$ planning iterations and CC-MCP with $2^{20}$ planning iterations take a similar amount of time?\n\n- About Fig. 4.\nWhat is one episode here?\nDoes one episode for C-MCTS and CC-MCP require a similar amount of time?\nIf the latter is faster, could we run CC-MCP 100 times and select the path based on the majority?\n\nMinor question.\n- I got confused with the notation of $C, c, \\hat{c}$. Are these sets or functions?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1699006476443}, {"id": "hyznLh4uPt", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4707/Reviewer_EfYF"], "rating": "5: marginally below the acceptance threshold", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "Constrained Markov decision processes (CMDPs) are environments subjected to constraints and have been addressed extensively using reinforcement learning techniques. Few sample-based planning algorithms have been applied to such problems. The paper introduces an MCTS-based approach for learning cost-conscious policies for CMDPs.", "review_text": "Constrained Markov decision processes (CMDPs) are environments subjected to constraints and have been addressed extensively using reinforcement learning techniques. Few sample-based planning algorithms have been applied to such problems. The paper introduces an MCTS-based approach for learning cost-conscious policies for CMDPs.", "strengths": "**Orignality:** While not introducing a completely novel approach, they apply an offline learning technique to estimate costs in CMDPs online.\n\n**Significance:** The contributions of the paper lack significance.\n\n**Clarity:** The paper is understandable.", "weaknesses": "The main drawback of the paper is its lack of significance. The approach introduced is not novel. Learning values/cost estimates offline to be applied online is not a new idea. Nor is the learning approach using a novel technique.\n\nI also question the soundness of the analysis. In Prop. 1, the authors claim that at each iteration of their algorithm, they are guaranteed to find the optimal solution. They base their claim on the proof in [Kocsis & Szepesvari, 2006]. However, that work states that UCT is guaranteed to converge in the limit --- it is not guaranteed to converge in finite-time.\n\nThe experiments also do not demonstrate the claimed performance improvements achieved by their approach. While C-MCTS significantly outperforms CC-MCP in Rocksample, it does not necessarily outperform MCTS. Nor does it significantly have less cost violations than CC-MCP. Furthermore, CC-MCP searches deeper than C-MCTS in Rocksample(11, 11).\n\nThe experiments themselves also do not seem. Maybe I'm missing something but it seems, that C-MCTS gets extra computations. The cost for pre-training should be factored into the planning budget.", "questions": "Corollary 1: How can you guarantee optimality if you prune sub-trees for which the cost is over-estimated?\n\nWhat exactly is a \"boundary region?\"\n\nThe lengths of the arrows are difficult to distinguish in Fig. 4. I would suggest finding a different way to visualize the number of times an action is selected.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "Constrained Markov decision processes (CMDPs) are environments subjected to constraints and have been addressed extensively using reinforcement learning techniques. Few sample-based planning algorithms have been applied to such problems. The paper introduces an MCTS-based approach for learning cost-conscious policies for CMDPs.", "soundness": "1 poor", "presentation": "2 fair", "contribution": "2 fair", "strengths": "**Orignality:** While not introducing a completely novel approach, they apply an offline learning technique to estimate costs in CMDPs online.\n\n**Significance:** The contributions of the paper lack significance.\n\n**Clarity:** The paper is understandable.", "weaknesses": "The main drawback of the paper is its lack of significance. The approach introduced is not novel. Learning values/cost estimates offline to be applied online is not a new idea. Nor is the learning approach using a novel technique.\n\nI also question the soundness of the analysis. In Prop. 1, the authors claim that at each iteration of their algorithm, they are guaranteed to find the optimal solution. They base their claim on the proof in [Kocsis & Szepesvari, 2006]. However, that work states that UCT is guaranteed to converge in the limit --- it is not guaranteed to converge in finite-time.\n\nThe experiments also do not demonstrate the claimed performance improvements achieved by their approach. While C-MCTS significantly outperforms CC-MCP in Rocksample, it does not necessarily outperform MCTS. Nor does it significantly have less cost violations than CC-MCP. Furthermore, CC-MCP searches deeper than C-MCTS in Rocksample(11, 11).\n\nThe experiments themselves also do not seem. Maybe I'm missing something but it seems, that C-MCTS gets extra computations. The cost for pre-training should be factored into the planning budget.", "questions": "Corollary 1: How can you guarantee optimality if you prune sub-trees for which the cost is over-estimated?\n\nWhat exactly is a \"boundary region?\"\n\nThe lengths of the arrows are difficult to distinguish in Fig. 4. I would suggest finding a different way to visualize the number of times an action is selected.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698795350725}, {"id": "8peu8Izo6s", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4707/Reviewer_DU49"], "rating": "3: reject, not good enough", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary": "The paper proposes an MCTS variant for safety-critical scenarios. The variant expands in leaf nodes only actions that are deemed safe and feasible by a safety-critic model. The model is optimized in a pre-training phase.\n\nThe MCTS variant is compared favorably with two other MCTS variants in two grid-worlds.", "review_text": "The paper proposes an MCTS variant for safety-critical scenarios. The variant expands in leaf nodes only actions that are deemed safe and feasible by a safety-critic model. The model is optimized in a pre-training phase.\n\nThe MCTS variant is compared favorably with two other MCTS variants in two grid-worlds.", "strengths": "Dealing with safety constraint is probably one of the weaknesses of the RL approaches, and one of the main obstacles for applying RL and planning algorithms like MCTS in real world scenarios. While in many of those scenarios, the algorithms are faced with continuous state/action spaces, tackling the issue in discrete spaces is also important.  \n\nThe extension of the MCTS for constrained MDP seems fairly reasonable.", "weaknesses": "While there is some theoretical work included, these do not offer sufficient guarantees for practical applicability. The proposed algorithm could be a step towards an practical application, but it is not there as it is.   \n\nGiven that this is a largely empirical article, the experimental evaluation is rather small. The benchmarks are small and fairly simple, while that set of baselines is also limited.", "questions": "I understand that there are not many MCTS implementation supporting constrained MDP, but there are a sufficient number of alternative planning algorithms that could have been used as additional baselines.\n\nThe size (especially the small number of possible actions) and the limited number of benchmarks are difficult to understand. If this would be a theoretical paper (or at least one offering significantly new algorithmic ideas) using these tasks for illustration would be fine. But, for a mainly empirical paper, the limitation of the experiments are difficult to understand.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes an MCTS variant for safety-critical scenarios. The variant expands in leaf nodes only actions that are deemed safe and feasible by a safety-critic model. The model is optimized in a pre-training phase.\n\nThe MCTS variant is compared favorably with two other MCTS variants in two grid-worlds.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "Dealing with safety constraint is probably one of the weaknesses of the RL approaches, and one of the main obstacles for applying RL and planning algorithms like MCTS in real world scenarios. While in many of those scenarios, the algorithms are faced with continuous state/action spaces, tackling the issue in discrete spaces is also important.  \n\nThe extension of the MCTS for constrained MDP seems fairly reasonable.", "weaknesses": "While there is some theoretical work included, these do not offer sufficient guarantees for practical applicability. The proposed algorithm could be a step towards an practical application, but it is not there as it is.   \n\nGiven that this is a largely empirical article, the experimental evaluation is rather small. The benchmarks are small and fairly simple, while that set of baselines is also limited.", "questions": "I understand that there are not many MCTS implementation supporting constrained MDP, but there are a sufficient number of alternative planning algorithms that could have been used as additional baselines.\n\nThe size (especially the small number of possible actions) and the limited number of benchmarks are difficult to understand. If this would be a theoretical paper (or at least one offering significantly new algorithmic ideas) using these tasks for illustration would be fine. But, for a mainly empirical paper, the limitation of the experiments are difficult to understand.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "code_of_conduct": "Yes"}, "tcdate": 1698769002655}, {"id": "eDWjQfodTR", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission4707/Reviewer_q4t7"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary": "This paper introduces Constrained MCTS (C-MCTS), an MCTS-based novel approach for solving CMDP decision-making tasks. This approach avoids constraint violations by pre-training a safety critic before agent deployment with TD learning using simulator data. The pre-trained safety critic eliminates the need for tuning a Lagrange multiplier online. The experiments in two environments, Rocksample and Safe Gridworld, show that C-MCTS outperforms existing approaches. It prunes unsafe branches of the search tree during deployment, achieving higher rewards while ensuring safety. Most importantly, it is less susceptible under the model mismatch between the simulator and the real-world environment.", "review_text": "This paper introduces Constrained MCTS (C-MCTS), an MCTS-based novel approach for solving CMDP decision-making tasks. This approach avoids constraint violations by pre-training a safety critic before agent deployment with TD learning using simulator data. The pre-trained safety critic eliminates the need for tuning a Lagrange multiplier online. The experiments in two environments, Rocksample and Safe Gridworld, show that C-MCTS outperforms existing approaches. It prunes unsafe branches of the search tree during deployment, achieving higher rewards while ensuring safety. Most importantly, it is less susceptible under the model mismatch between the simulator and the real-world environment.", "strengths": "This paper proposes a novel and straightforward method to pre-train a safety critic using the TD-based method. By negating the need to tune the Lagrange multiplier online, the pre-training phase allows the use of a more computationally expensive high-fidelity simulator, thereby producing better results. Besides, it also allows the ensemble method to be applied to have more robust results.", "weaknesses": "While the method is novel and straightforward, there are two major concerns as follows.\n\nFirst, the time complexity of C-MCTS seems higher than normal MCTS. In the expansion of C-MCTS, the safety critic is used to evaluate all actions. This greatly increases the computing resources when applying in the large action space domains or using more MCTS simulations. In addition, considering C-MCTS uses additional resources to tune a Lagrange multiplier, the comparison between C-MCTS and CC-MCP is unfair. The authors should provide the detailed time complexity for C-MCTS, and the time for the pre-training phase for comparison.\n\nSecond, the experimental environments are quite simple, so it is hard to convincingly evaluate the out-of-distribution scenarios where states are completely unseen during deployment. The author should add experiments on more complex environments, e.g., the MuJoCo physics simulator.", "questions": "* P2, Figure 1. As introduced in Section 3.1, the training phase uses MCTS to solve MDP, which should also be added to the figure for better understanding.\n* P2, Section 2.1. Use bold font to emphasize sets (especially for $C$ and $\\hat{c}$) in definitions and equations for clear reading. All policies $\\Pi$ is undefined in equation (1).\n* P3, Section 2.2, \"Lee et al. (2018) proposed Cost-Constrained Monte Carlo Planning (CC-MCP)\". This is incorrect since Lee et al. proposed Cost-Constrained POMCP (CC-POMCP) instead of CC-MCP. Although it applies to CMDPs, the authors should revise the statement to clarify this.\n* P3, Section 3.1. Clarify what \"e.g., in simulation\" means. Does it mean to vary parameters during MCTS simulation in the pre-training phase? In addition, $V_{C}^{k,*}$ is undefined.\n* P4, Section 3.1.1, Definition 3. In equation (5), it seems meaningless to define $\\mathop{\\min}_{a'} {Q^{\\pi}_{c}(s,a')} <= Q^{\\pi}_{c}(s,a)$ when $\\forall a' \\in A$.\n* P4, Proposition 2, \"This is not only numerically infeasible, but it is also due to the utilization of the low-fidelity simulator in the MCTS planner.\" It is still unclear why this is numerically infeasible. In common sense, even if a low-fidelity simulator is employed, it is still likely that the critic can provide perfect estimation for some state-action pairs, especially for those near the terminal states.\n* P5, Section 3.1.1, Corollary 1, \"As discussed before, the inner training loop will always converge to the optimal solution for the k−th MDP.\" The convergence happens when there are sufficient training simulations, as stated in Proposition 1. However, did the conducted experiment have sufficient simulations to achieve the convergence, i.e., the optimal solution?\n* P6, Section 4, \"the selected environments clearly \"emulate\" the complexity that stems from domains with large state/action spaces (Schrittwieser et al., 2020b; Afsar et al., 2022).\" Please clarify the reason for citing these two papers here. They do not seem to have any discussion about the selected environments.\n* P8, Section 4.2, \"with a grid search\". Did the optimized $\\alpha_0$ and $\\epsilon$ highly affected the final performance? The authors should provide the grid search results in the appendix for reference.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces Constrained MCTS (C-MCTS), an MCTS-based novel approach for solving CMDP decision-making tasks. This approach avoids constraint violations by pre-training a safety critic before agent deployment with TD learning using simulator data. The pre-trained safety critic eliminates the need for tuning a Lagrange multiplier online. The experiments in two environments, Rocksample and Safe Gridworld, show that C-MCTS outperforms existing approaches. It prunes unsafe branches of the search tree during deployment, achieving higher rewards while ensuring safety. Most importantly, it is less susceptible under the model mismatch between the simulator and the real-world environment.", "soundness": "2 fair", "presentation": "3 good", "contribution": "2 fair", "strengths": "This paper proposes a novel and straightforward method to pre-train a safety critic using the TD-based method. By negating the need to tune the Lagrange multiplier online, the pre-training phase allows the use of a more computationally expensive high-fidelity simulator, thereby producing better results. Besides, it also allows the ensemble method to be applied to have more robust results.", "weaknesses": "While the method is novel and straightforward, there are two major concerns as follows.\n\nFirst, the time complexity of C-MCTS seems higher than normal MCTS. In the expansion of C-MCTS, the safety critic is used to evaluate all actions. This greatly increases the computing resources when applying in the large action space domains or using more MCTS simulations. In addition, considering C-MCTS uses additional resources to tune a Lagrange multiplier, the comparison between C-MCTS and CC-MCP is unfair. The authors should provide the detailed time complexity for C-MCTS, and the time for the pre-training phase for comparison.\n\nSecond, the experimental environments are quite simple, so it is hard to convincingly evaluate the out-of-distribution scenarios where states are completely unseen during deployment. The author should add experiments on more complex environments, e.g., the MuJoCo physics simulator.", "questions": "* P2, Figure 1. As introduced in Section 3.1, the training phase uses MCTS to solve MDP, which should also be added to the figure for better understanding.\n* P2, Section 2.1. Use bold font to emphasize sets (especially for $C$ and $\\hat{c}$) in definitions and equations for clear reading. All policies $\\Pi$ is undefined in equation (1).\n* P3, Section 2.2, \"Lee et al. (2018) proposed Cost-Constrained Monte Carlo Planning (CC-MCP)\". This is incorrect since Lee et al. proposed Cost-Constrained POMCP (CC-POMCP) instead of CC-MCP. Although it applies to CMDPs, the authors should revise the statement to clarify this.\n* P3, Section 3.1. Clarify what \"e.g., in simulation\" means. Does it mean to vary parameters during MCTS simulation in the pre-training phase? In addition, $V_{C}^{k,*}$ is undefined.\n* P4, Section 3.1.1, Definition 3. In equation (5), it seems meaningless to define $\\mathop{\\min}_{a'} {Q^{\\pi}_{c}(s,a')} <= Q^{\\pi}_{c}(s,a)$ when $\\forall a' \\in A$.\n* P4, Proposition 2, \"This is not only numerically infeasible, but it is also due to the utilization of the low-fidelity simulator in the MCTS planner.\" It is still unclear why this is numerically infeasible. In common sense, even if a low-fidelity simulator is employed, it is still likely that the critic can provide perfect estimation for some state-action pairs, especially for those near the terminal states.\n* P5, Section 3.1.1, Corollary 1, \"As discussed before, the inner training loop will always converge to the optimal solution for the k−th MDP.\" The convergence happens when there are sufficient training simulations, as stated in Proposition 1. However, did the conducted experiment have sufficient simulations to achieve the convergence, i.e., the optimal solution?\n* P6, Section 4, \"the selected environments clearly \"emulate\" the complexity that stems from domains with large state/action spaces (Schrittwieser et al., 2020b; Afsar et al., 2022).\" Please clarify the reason for citing these two papers here. They do not seem to have any discussion about the selected environments.\n* P8, Section 4.2, \"with a grid search\". Did the optimized $\\alpha_0$ and $\\epsilon$ highly affected the final performance? The authors should provide the grid search results in the appendix for reference.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "code_of_conduct": "Yes"}, "tcdate": 1698404513392}], "openreview_url": "https://openreview.net/forum?id=BMw4Cm0gGO", "arxiv_id": "2305.16209", "paper_pdf": "papers/BMw4Cm0gGO.pdf", "paper_pdf_sha256": "ae98759604c00dbad971d43d8dbdab354781980c377e42e492bc472337f41886", "paper_pdf_bytes": 483528, "paper_pdf_source": "openreview", "code_url": "https://github.com/mutschcr/C-MCTS", "code_repository": "mutschcr/C-MCTS", "code_commit": "f690100502cf0db548edc318b0f055ecbfcac790", "code_archive": "repos/BMw4Cm0gGO.zip", "code_archive_sha256": "a92a43b669da9eaf6fac00eaa2fce613a01f8fdd1aebb66f17008ded68428b8d", "code_archive_bytes": 629146, "code_file_count": 36, "code_extensions": {".h": 14, ".cpp": 12, ".py": 10}, "github_disk_usage_kb": 537, "github_languages": {"C++": 85212, "Python": 69327, "Makefile": 1200}, "github_archived": false, "github_pushed_at": "2023-05-25T04:37:26Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/c-mcts-safe-planning-with-monte-carlo-tree"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "MpGP-z07TmM", "year": 2023, "status": "rejected", "title": "Learning Specialized Activation Functions for Physics-informed Neural Networks", "authors": ["Honghui Wang", "Lu Lu", "Shiji Song", "Gao Huang"], "authorids": ["~Honghui_Wang1", "~Lu_Lu1", "~Shiji_Song1", "~Gao_Huang1"], "authors_source": "OpenReview API", "abstract": "At the heart of network architectures lie the non-linear activation functions, the choice of which affects the model optimization and task performance. In computer vision and natural language processing, the Rectified Linear Unit is widely adopted across different tasks. However, there is no such default choice of activation functions in the context of physics-informed neural networks (PINNs). It is observed that PINNs exhibit high sensitivity to activation functions due to the various characteristics of each physics system, which makes the choice of the suitable activation function for PINNs a critical issue. Existing works usually choose activation functions in an inefficient trial-and-error manner. To address this problem, we propose to search automatically for the optimal activation function when solving different PDEs. This is achieved by learning an adaptive activation function as linear combinations of a set of candidate functions, whose coefficients can be directly optimized by gradient descent. In addition to its efficient optimization, the proposed method enables the discovery of novel activation function and the incorporation with prior knowledge about the PDE system. We can further enhance its search space with adaptive slope. The effectiveness of the proposed adaptive activation function is demonstrated on a series of benchmarks, including the Poisson's equation, Burgers' equation, Allen-Cahn equation, convection equation, Korteweg–de Vries equation and Cahn-Hilliard equation. The performance gain of the proposed method is further interpreted from the neural tangent kernel perspective. Code will be released. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "pSZh185xr_", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1042/Reviewer_5qLY"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The goal of this work is to automatically learn the activation functions via the combinations of different candidate activation functions for PINN. Specifically, it adopted gate function with a learnable parameter to identify the coefficients for different candidate activation functions. Extensive experiments are carried out to evaluate the proposed physics-informed activation functions on a variety of partial differential equations (PDEs). The results demonstrate the efficacy of the simple approach. ", "review_text": "It is interesting to automatically learn the activation functions for PINNs. The proposed method is simple yet effective for PDEs systems. However, this work adopted the same gate function method from prior work and apply it to PINNs. Hence, the technical innovation is somewhat limited.", "strengths": "* Strength\n1. It is interesting to automatically choose the activation functions for PINN\n2. The proposed method is simple yet effective to solve PDE systems\n3. Conduct extensive experiments on various PDEs to verify the effectiveness of the proposed approach\n\n* Limitations\n1. Technical innovation is somewhat limited. The proposed gate function to learn the coefficients of activation functions is the same as that in the literature [Qian et al. 2018]\n2. Compare to more baselines. I am curious if you could compare the prior works that learn combinations of activation functions?\n3. A few grammatical errors. For example, \"by minimize the following objective function\" in the bottom of page 2; and \"we consider a extreme case of insufficient collocation points\", a-> an\n\nReferences:\\\n[Qian et al. 2018] Adaptive activation functions in convolutional neural networks, 2018.", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The goal of this work is to automatically learn the activation functions via the combinations of different candidate activation functions for PINN. Specifically, it adopted gate function with a learnable parameter to identify the coefficients for different candidate activation functions. Extensive experiments are carried out to evaluate the proposed physics-informed activation functions on a variety of partial differential equations (PDEs). The results demonstrate the efficacy of the simple approach. ", "strength_and_weaknesses": "* Strength\n1. It is interesting to automatically choose the activation functions for PINN\n2. The proposed method is simple yet effective to solve PDE systems\n3. Conduct extensive experiments on various PDEs to verify the effectiveness of the proposed approach\n\n* Limitations\n1. Technical innovation is somewhat limited. The proposed gate function to learn the coefficients of activation functions is the same as that in the literature [Qian et al. 2018]\n2. Compare to more baselines. I am curious if you could compare the prior works that learn combinations of activation functions?\n3. A few grammatical errors. For example, \"by minimize the following objective function\" in the bottom of page 2; and \"we consider a extreme case of insufficient collocation points\", a-> an\n\nReferences:\\\n[Qian et al. 2018] Adaptive activation functions in convolutional neural networks, 2018.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and organized. However, the technical innovation is limited.", "summary_of_the_review": "It is interesting to automatically learn the activation functions for PINNs. The proposed method is simple yet effective for PDEs systems. However, this work adopted the same gate function method from prior work and apply it to PINNs. Hence, the technical innovation is somewhat limited.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "No", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666659722836}, {"id": "emIJpG4HSH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1042/Reviewer_qn6w"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors present a formulation to create a trainable activation function based on a convex combination of other activations.\nThis new mixture approach is intended to work on physics-informed neural networks (PINNd) where higher order derivatives play a role, thus causing problems for some of the piecewise linear based common activations.\n\nThe authors the provide an empirical evaluation as well as a link to the neural tangent kernel.", "review_text": "In this paper, the authors present an activation function that is amenable to PINNs. The authors present the problem clearly and offer a solution for this particular context based on previous work. They however do not compare to other adaptable activation functions that might still behave well in this domain. \nThe empirical evaluation would be much stronger, if you were to compare to that class of activations. In particular, it would be great to show the benefits of your activation function, or explain why the others are not suitable and demonstrate that empirically.", "strengths": "The authors present a technically sound approach for PINNs. Evaluated on datasets that clearly present the difficulties in this domain. \n\nHowever, the main weakness of this paper is in the lack of comparison to other learnable activation functions. The authors present these family of activation functions in the section \"Adaptive activation functions\", yet they are not present in the evaluation.\nSome of them could be relevant as they do provide higher oder derivatives. Maybe there is an issue with those activation functions and that is a good reason not to include them, but this is not clear from the paper.\n\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "In this paper, the authors present a formulation to create a trainable activation function based on a convex combination of other activations.\nThis new mixture approach is intended to work on physics-informed neural networks (PINNd) where higher order derivatives play a role, thus causing problems for some of the piecewise linear based common activations.\n\nThe authors the provide an empirical evaluation as well as a link to the neural tangent kernel.", "strength_and_weaknesses": "The authors present a technically sound approach for PINNs. Evaluated on datasets that clearly present the difficulties in this domain. \n\nHowever, the main weakness of this paper is in the lack of comparison to other learnable activation functions. The authors present these family of activation functions in the section \"Adaptive activation functions\", yet they are not present in the evaluation.\nSome of them could be relevant as they do provide higher oder derivatives. Maybe there is an issue with those activation functions and that is a good reason not to include them, but this is not clear from the paper.\n\n", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and technically sound. The gating function and components are presented clearly. The activation function can be reproduced from the information in the paper.", "summary_of_the_review": "In this paper, the authors present an activation function that is amenable to PINNs. The authors present the problem clearly and offer a solution for this particular context based on previous work. They however do not compare to other adaptable activation functions that might still behave well in this domain. \nThe empirical evaluation would be much stronger, if you were to compare to that class of activations. In particular, it would be great to show the benefits of your activation function, or explain why the others are not suitable and demonstrate that empirically.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper"}, "tcdate": 1666656083712}, {"id": "8pBTbDHbWu", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1042/Reviewer_J89g"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The authors present a method for parameterizing activation functions to improve convergence in physics-informed-neural-networks. They demonstrate through empirical experimentation that the proposed method converges to better solutions than non-learned activation functions or adaptive slope methods. Finally, the authors provide an analysis based on Neural Tangent Kernel that sheds light on why the proposed method is effective.\n", "review_text": "Overall this paper is a solid demonstration of an approach to address the training difficulty empirically observed in PINNs.\n\n", "strengths": "The method proposed by the authors addresses a significant issue in PINNs. The approach is demonstrated on 5 different PDEs. The method is compared to several baselines as well as some ablations. \n\nThe NTK based analysis is quite hand-wavy. There is no treatment of the approximation error of the two-phase analysis and the notion that the NTK is learned to suite the underlying PDE is approached empirically with a single example.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors present a method for parameterizing activation functions to improve convergence in physics-informed-neural-networks. They demonstrate through empirical experimentation that the proposed method converges to better solutions than non-learned activation functions or adaptive slope methods. Finally, the authors provide an analysis based on Neural Tangent Kernel that sheds light on why the proposed method is effective.\n", "strength_and_weaknesses": "The method proposed by the authors addresses a significant issue in PINNs. The approach is demonstrated on 5 different PDEs. The method is compared to several baselines as well as some ablations. \n\nThe NTK based analysis is quite hand-wavy. There is no treatment of the approximation error of the two-phase analysis and the notion that the NTK is learned to suite the underlying PDE is approached empirically with a single example.", "clarity,_quality,_novelty_and_reproducibility": "The paper is well written and clear. \nOne thing that was not clear to me is: Do all activation functions in the network share the same parameters? or are they all learned independently? \nThe approach is somewhat novel in the PINN literature.\nThe work appears to be reasonably reproducible. I have no concerns there.\n\nThe other reviewers have raised the criticism that the proposed method is very similar to existing work. I was not aware of that work when I wrote my review. I find that the existence of the prior work significantly reduces my view of the novelty of the authors' contribution. I have therefore reduced my score. I think the authors could improve their contribution by formally proving their NTK claims or perhaps by expanding the number of base activation functions to very large numbers (e.g. using a basis). ", "summary_of_the_review": "Overall this paper is a solid demonstration of an approach to address the training difficulty empirically observed in PINNs.\n\n", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666455655039}, {"id": "D_mJouO-0k1", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper1042/Reviewer_E27E"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "The paper proposes trainable activation functions for physics-informed neural networks (PINNs). The basic idea is to consider a set of base activation functions, and build a trainable one as a convex combination of the base ones. They provide experiments on many problems, showing that PINNs are generally sensitive to the choice of AF, while the trainable one provides good performance across most cases. They also try to motivate the approach from the point of view of the neural tangent kernel (NTK).", "review_text": "The method is a rebranding of a known technique, applied on a different domain. While this is potentially interesting for PINNs, I feel this is not a good practice and it is also an artificial construct to increase novelty.", "strengths": "Strength: From the point of view of PINNs, the proposed neural networks have better accuracy and performance than the base ones. \n\nDrawbacks: \n- From the methodological point of view, the method is *equivalent* to the method proposed in (Manessi and Rozza, 2019), referenced in the paper, which was also proposed under the name of \"adaptive blending unit\" in (Sutfeld et al., 2020), which is not referenced here. The authors are renominating them as \"physics-informed activation function\" (PIAC), but as far as I can see there is no difference that warrants a change of name (apart from slightly modifying the number of base AFs). They mention that \"previous methods experiment with limit choice of activation functions\", but they use 5 base AFs which is similar to 6 base AFs in (Sutfeld et al., 2020). \n- The method is also compared only with fixed or parametric AFs, and not against similar expressive AFs (e.g., APLs).\n- The NTK derivation is quite shallow and it does not provide any special insight into the AFs themselves.", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "The paper proposes trainable activation functions for physics-informed neural networks (PINNs). The basic idea is to consider a set of base activation functions, and build a trainable one as a convex combination of the base ones. They provide experiments on many problems, showing that PINNs are generally sensitive to the choice of AF, while the trainable one provides good performance across most cases. They also try to motivate the approach from the point of view of the neural tangent kernel (NTK).", "strength_and_weaknesses": "Strength: From the point of view of PINNs, the proposed neural networks have better accuracy and performance than the base ones. \n\nDrawbacks: \n- From the methodological point of view, the method is *equivalent* to the method proposed in (Manessi and Rozza, 2019), referenced in the paper, which was also proposed under the name of \"adaptive blending unit\" in (Sutfeld et al., 2020), which is not referenced here. The authors are renominating them as \"physics-informed activation function\" (PIAC), but as far as I can see there is no difference that warrants a change of name (apart from slightly modifying the number of base AFs). They mention that \"previous methods experiment with limit choice of activation functions\", but they use 5 base AFs which is similar to 6 base AFs in (Sutfeld et al., 2020). \n- The method is also compared only with fixed or parametric AFs, and not against similar expressive AFs (e.g., APLs).\n- The NTK derivation is quite shallow and it does not provide any special insight into the AFs themselves.", "clarity,_quality,_novelty_and_reproducibility": "The paper is written well except for a small number of typos (e.g., \"over-parameterizaiton\"). The experimental part is well detailed but with an insufficient number of comparisons. Novelty is extremely small, as described above.", "summary_of_the_review": "The method is a rebranding of a known technique, applied on a different domain. While this is potentially interesting for PINNs, I feel this is not a good practice and it is also an artificial construct to increase novelty.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1665761064094}], "openreview_url": "https://openreview.net/forum?id=MpGP-z07TmM", "arxiv_id": "2308.04073", "paper_pdf": "papers/MpGP-z07TmM.pdf", "paper_pdf_sha256": "9c98e682fe8f416b08b23c73ccfd585da392fe9d1a163d93364485b87dc92019", "paper_pdf_bytes": 1036724, "paper_pdf_source": "openreview", "code_url": "https://github.com/LeapLabTHU/AdaAFforPINNs", "code_repository": "LeapLabTHU/AdaAFforPINNs", "code_commit": "2e87a3dd134c6bb359c8330a4d8e84f0a97201ed", "code_archive": "repos/MpGP-z07TmM.zip", "code_archive_sha256": "a659359d57d1659e599d17aa8e8555140585b0a83aee3968e2bf0a4028d5b61f", "code_archive_bytes": 1674830, "code_file_count": 8, "code_extensions": {".py": 7, ".sh": 1}, "github_disk_usage_kb": 1642, "github_languages": {"Python": 125038, "Shell": 7708}, "github_archived": false, "github_pushed_at": "2023-08-09T03:15:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/learning-specialized-activation-functions-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "OBwsUF4nFye", "year": 2022, "status": "rejected", "title": "Private Multi-Task Learning: Formulation and Applications to Federated Learning", "authors": ["Shengyuan Hu", "Steven Wu", "Virginia Smith"], "authorids": ["~Shengyuan_Hu2", "~Steven_Wu1", "~Virginia_Smith1"], "authors_source": "OpenReview API", "abstract": "Many problems in machine learning rely on multi-task learning (MTL), in which the goal is to solve multiple related machine learning tasks simultaneously. MTL is particularly relevant for privacy-sensitive applications in areas such as healthcare, finance, and IoT computing, where sensitive data from multiple, varied sources are shared for the purpose of learning. In this work, we formalize notions of task-level privacy for MTL via joint differential privacy(JDP), a relaxation of differential privacy for mechanism design and distributed optimization. We then propose an algorithm for mean-regularized MTL, an objective commonly used for applications in personalized federated learning, subject to JDP. We analyze our objective and solver, providing certifiable guarantees on both privacy and utility.  Empirically, we find that our method allows for improved privacy/utility trade-offs relative to global baselines across common federated learning benchmarks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "_H8jbk7UuJd", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1804/Reviewer_fBjV"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "In this paper, the authors formalize a notion of task-level differential privacy using joint differential privacy, a concept known in the differentially private mechanism design literature. They formulate the problem of multitask learning under differential privacy under this relaxed definition. The motivation is that there is no need for privacy of the data of a particular task from the task itself, but since you are sharing information in multitask learning, you want to protect against other tasks learning about your private information. They provide an algorithm with convergence guarantees to do multitask learning in this framework and perform some experiments to test this algorithm in the application of federated learning", "review_text": "Strengths:\nThis paper proposes a new idea of using a relaxation of differential privacy in the multitask learning framework.\nIt motivates work in the direction of looking at application specific relaxations of differential privacy, to give more realistic algorithms which perform better than the ones which satisfy the strong differential privacy constraint.\n\nWeaknesses/Questions:\n\n\t\n1. Is $x_i$ the data? Is $l_k(w_k)$ in eq. (7) the population loss or the empirical loss?\n\n2. Some more results like comparison of rates to algorithms that are not only joint differentially private, but differentially private would be useful. Are the improvements only in constant factors or are the improvements in terms of dimension dependent or number of samples dependent terms?\n\n3. Can the authors give some sense of what are the best rates you can under joint differential privacy and how does that compare with the given algorithm?\n\n4. Is there any motivation as to why the number of local optimization steps was chosen to be 1? In more realistic scenarios, each task would take multiple descent steps to personalize better, right?\n\n5. Can you show a short calculation to ensure the constant multiplying the exponentially decaying term is always positive?\n\n6. The last two terms in equation (13) seem fishy, does it mean the error doesn’t decay by increasing the number of iterations?\n\n7. The values of epsilon in figure 2, seem very high for the loss to decay enough.\n\n8. It’s not surprising that private MTL does better than the private global models, by the sheer number of parameters that private MTL allows (it’s ~m times the private global model parameters). A more fair comparison would be to compare against an algorithm that is globally differentially private instead of joint differentially private.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "In this paper, the authors formalize a notion of task-level differential privacy using joint differential privacy, a concept known in the differentially private mechanism design literature. They formulate the problem of multitask learning under differential privacy under this relaxed definition. The motivation is that there is no need for privacy of the data of a particular task from the task itself, but since you are sharing information in multitask learning, you want to protect against other tasks learning about your private information. They provide an algorithm with convergence guarantees to do multitask learning in this framework and perform some experiments to test this algorithm in the application of federated learning", "main_review": "Strengths:\nThis paper proposes a new idea of using a relaxation of differential privacy in the multitask learning framework.\nIt motivates work in the direction of looking at application specific relaxations of differential privacy, to give more realistic algorithms which perform better than the ones which satisfy the strong differential privacy constraint.\n\nWeaknesses/Questions:\n\n\t\n1. Is $x_i$ the data? Is $l_k(w_k)$ in eq. (7) the population loss or the empirical loss?\n\n2. Some more results like comparison of rates to algorithms that are not only joint differentially private, but differentially private would be useful. Are the improvements only in constant factors or are the improvements in terms of dimension dependent or number of samples dependent terms?\n\n3. Can the authors give some sense of what are the best rates you can under joint differential privacy and how does that compare with the given algorithm?\n\n4. Is there any motivation as to why the number of local optimization steps was chosen to be 1? In more realistic scenarios, each task would take multiple descent steps to personalize better, right?\n\n5. Can you show a short calculation to ensure the constant multiplying the exponentially decaying term is always positive?\n\n6. The last two terms in equation (13) seem fishy, does it mean the error doesn’t decay by increasing the number of iterations?\n\n7. The values of epsilon in figure 2, seem very high for the loss to decay enough.\n\n8. It’s not surprising that private MTL does better than the private global models, by the sheer number of parameters that private MTL allows (it’s ~m times the private global model parameters). A more fair comparison would be to compare against an algorithm that is globally differentially private instead of joint differentially private.\n", "summary_of_the_review": "The idea to use joint differential privacy instead of differential privacy in MTL is interesting. The privacy analysis and the convergence rates presented are fairly standard. The theory doesn’t seem to stand on it’s own and more experiments can be done to rigorously test the introduction of this idea. Hence, I believe in it’s current form the paper is not ready for publication.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635987746263}, {"id": "PIwwUkaZdyL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1804/Reviewer_C2cr"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "Many problems in machine learning rely on multi-task learning (MTL), in which the goal is to solve multiple related machine learning tasks simultaneously. In this work, authors formalize notions of task-level privacy for MTL via joint differential privacy (JDP). They propose an algorithm for mean-regularized MTL, an objective commonly used for applications in personalized federated learning, subject to JDP. Then analyze objective and solver, providing certifiable guarantees on both privacy and utility. The method allows for improved privacy/utility trade-offs relative to global baselines across common federated learning benchmarks. ", "review_text": "They give a task-level private algorithm for multi-task learning, which is very important in machine learning, and show the theoretical convergence analysis. The structure is clear and writing is good. The experiment results can also support their theory.\nTypo. Section 5.3 “common” rather than “commons”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "Many problems in machine learning rely on multi-task learning (MTL), in which the goal is to solve multiple related machine learning tasks simultaneously. In this work, authors formalize notions of task-level privacy for MTL via joint differential privacy (JDP). They propose an algorithm for mean-regularized MTL, an objective commonly used for applications in personalized federated learning, subject to JDP. Then analyze objective and solver, providing certifiable guarantees on both privacy and utility. The method allows for improved privacy/utility trade-offs relative to global baselines across common federated learning benchmarks. ", "main_review": "They give a task-level private algorithm for multi-task learning, which is very important in machine learning, and show the theoretical convergence analysis. The structure is clear and writing is good. The experiment results can also support their theory.\nTypo. Section 5.3 “common” rather than “commons”\n", "summary_of_the_review": "The work is complement, although only one algorithm for one problem, the analysis and experiments are sufficient.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635851269261}, {"id": "NogzuELiSwI", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper1804/Reviewer_T1p3"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper considers the problem of multi-task learning  (MTL) with task-level differential privacy (TLDP) constraints. The authors propose using joint differential privacy (JDP) for MTL and provide a JDP MTL algorithm, which is a DP variation of FedAvg. They prove the privacy of their algorithm and provide convergence results for smooth nonconvex and strongly convex losses. They then provide numerical experiments. ", "review_text": "Overall, I find Sections 1-3 to be strong: the paper is well-written and the problem is well-formulated and nicely described. My main problem is with the convergence results, which are difficult to interpret due to too many parameters and no informal bound is provided. Additionally, it is not clear that the bounds are even non-trivial (see below for more discussion). My other main point of confusion is that the authors say they consider JDP as a relaxation of TLDP to provide better utility, but then their algorithm actually appears to provide the stronger (by Billboard Lemma) notion of TLDP (due to Theorem 1); so what is the role of JDP in their results? \n\nDetailed Review: \np.1: \n\nis task-level DP essentially the same as user-level DP/client-level DP (which is considered in McMahan et al 2018)? If so, state that--these terms are more familiar to the DP community than task-level DP. \n\nemphasize \"as long as client k keeps their data private\" \n\np.3 \n\nMotivate Eq (1). Goal is to train a model that has small average loss over all tasks, right? So why use a regularizer? \n\nDefine $\\ell_k$ and $x_i$ \n\nMissing parentheses in Eq (2)\n\np. 4\n\nclarify that strong guarantee of JDP hinges on each client keeping its own model private \n\nI believe Kearns et al 2014 is wrong citation for Lemma 1. Shouldn't it be Hsu et al? \n\nLemma 1: $D_k$ ->  $D_i$\n\np.6\n\nTheorem 1: missing restriction on $\\epsilon$. Proof of Abadi et al (2016) Theorem 1 shows that some restriction (roughly $\\epsilon \\leq \\ln(1/\\delta)$) is needed; also, you should refer to what you provide as TLDP or user-level/client-level DP (in Theorem 1 and in Definition 1)--DP is too vague, as there are many notions of DP \n\n**Why study JDP if your algorithm satisfies the stronger notion of TLDP?**\n\np.7\n\nIf you assume $\\gamma \\geq \\| \\nabla f_k(w) \\|$ (i.e. loss is effectively Lipschitz), then it seems you do not need to clip the gradients (and clipping will never actually occur in the algorithm)--am I right? \n\nWhy assume loss is $L + \\lambda$-smooth and not $L$ smooth? \n\nShould replace min by inf unless you are assuming compactness of domain; are you? I don't the domain was formally defined. I don't remember seeing $d$ defined either. \n\n**Convergence bounds: if $f_k \\leq 1$ so $B_t = 1$, then the second term in Eq (11) becomes essentially trivial $O(L + \\lambda + ...)$** And that is the non-private case, so the DP bounds can only get \"worse\". \n\nAlso, **the bounds in (8)-(10) and (12) are too complicated: should provide simplified version with dependence on key parameters m, d, epsilon, delta displayed**; and even the formal version should not be simplified to include only the key parameters as well as the smoothness and boundedness parameters, and non-dominant terms should be omitted. \n\n\n ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper considers the problem of multi-task learning  (MTL) with task-level differential privacy (TLDP) constraints. The authors propose using joint differential privacy (JDP) for MTL and provide a JDP MTL algorithm, which is a DP variation of FedAvg. They prove the privacy of their algorithm and provide convergence results for smooth nonconvex and strongly convex losses. They then provide numerical experiments. ", "main_review": "Overall, I find Sections 1-3 to be strong: the paper is well-written and the problem is well-formulated and nicely described. My main problem is with the convergence results, which are difficult to interpret due to too many parameters and no informal bound is provided. Additionally, it is not clear that the bounds are even non-trivial (see below for more discussion). My other main point of confusion is that the authors say they consider JDP as a relaxation of TLDP to provide better utility, but then their algorithm actually appears to provide the stronger (by Billboard Lemma) notion of TLDP (due to Theorem 1); so what is the role of JDP in their results? \n\nDetailed Review: \np.1: \n\nis task-level DP essentially the same as user-level DP/client-level DP (which is considered in McMahan et al 2018)? If so, state that--these terms are more familiar to the DP community than task-level DP. \n\nemphasize \"as long as client k keeps their data private\" \n\np.3 \n\nMotivate Eq (1). Goal is to train a model that has small average loss over all tasks, right? So why use a regularizer? \n\nDefine $\\ell_k$ and $x_i$ \n\nMissing parentheses in Eq (2)\n\np. 4\n\nclarify that strong guarantee of JDP hinges on each client keeping its own model private \n\nI believe Kearns et al 2014 is wrong citation for Lemma 1. Shouldn't it be Hsu et al? \n\nLemma 1: $D_k$ ->  $D_i$\n\np.6\n\nTheorem 1: missing restriction on $\\epsilon$. Proof of Abadi et al (2016) Theorem 1 shows that some restriction (roughly $\\epsilon \\leq \\ln(1/\\delta)$) is needed; also, you should refer to what you provide as TLDP or user-level/client-level DP (in Theorem 1 and in Definition 1)--DP is too vague, as there are many notions of DP \n\n**Why study JDP if your algorithm satisfies the stronger notion of TLDP?**\n\np.7\n\nIf you assume $\\gamma \\geq \\| \\nabla f_k(w) \\|$ (i.e. loss is effectively Lipschitz), then it seems you do not need to clip the gradients (and clipping will never actually occur in the algorithm)--am I right? \n\nWhy assume loss is $L + \\lambda$-smooth and not $L$ smooth? \n\nShould replace min by inf unless you are assuming compactness of domain; are you? I don't the domain was formally defined. I don't remember seeing $d$ defined either. \n\n**Convergence bounds: if $f_k \\leq 1$ so $B_t = 1$, then the second term in Eq (11) becomes essentially trivial $O(L + \\lambda + ...)$** And that is the non-private case, so the DP bounds can only get \"worse\". \n\nAlso, **the bounds in (8)-(10) and (12) are too complicated: should provide simplified version with dependence on key parameters m, d, epsilon, delta displayed**; and even the formal version should not be simplified to include only the key parameters as well as the smoothness and boundedness parameters, and non-dominant terms should be omitted. \n\n\n ", "summary_of_the_review": "The paper considers/formulates an interesting problem and is generally well-written. Unfortunately, the convergence results do not appear to be strong (or clear). Additionally, there is confusion about the notion of privacy that their algorithm provides and the role that JDP is playing in the paper. For these reasons I cannot recommend acceptance for the current form of the paper. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635814835643}], "openreview_url": "https://openreview.net/forum?id=OBwsUF4nFye", "arxiv_id": "2108.12978", "paper_pdf": "papers/OBwsUF4nFye.pdf", "paper_pdf_sha256": "e13fcdb2e1488db3919dd73797bbecbfff096a7ceae6f886befedf0154ffbe38", "paper_pdf_bytes": 1208779, "paper_pdf_source": "openreview", "code_url": "https://github.com/s-huu/PMTL", "code_repository": "s-huu/PMTL", "code_commit": "9e849ada6af2d8df92c40724d81b4b4cf2ac24b0", "code_archive": "repos/OBwsUF4nFye.zip", "code_archive_sha256": "a5faf61ca8aed85f60630498ea84c5f7f0a1fe2dbce016a22a2c7e76a91eb5b7", "code_archive_bytes": 7688047, "code_file_count": 18, "code_extensions": {".py": 15, ".sh": 3}, "github_disk_usage_kb": 7495, "github_languages": {"Python": 136765, "Shell": 3547}, "github_archived": false, "github_pushed_at": "2023-04-08T21:05:08Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/private-multi-task-learning-formulation-and"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "qn_gk5j3PJ", "year": 2021, "status": "rejected", "title": "PIVEN: A Deep Neural Network for Prediction Intervals with Specific Value Prediction", "authors": ["Eli Simhayev", "Gilad Katz", "Lior Rokach"], "authorids": ["~Eli_Simhayev1", "giladkz@post.bgu.ac.il", "liorrk@post.bgu.ac.il"], "authors_source": "OpenReview API", "abstract": "Improving the robustness of neural nets in regression tasks is key to their application in multiple domains. Deep learning-based approaches aim to achieve this goal either by improving their prediction of specific values (i.e., point prediction), or by producing prediction intervals (PIs) that quantify uncertainty. We present PIVEN, a deep neural network  for producing both a PI and a prediction of specific values. Unlike previous studies, PIVEN makes no assumptions regarding data distribution inside the PI, making its point prediction more effective for various real-world problems. Benchmark experiments show that our approach produces tighter uncertainty bounds than the current state-of-the-art approach for producing PIs, while maintaining comparable performance to the state-of-the-art approach for specific value-prediction. Additional evaluation on large image datasets further support our conclusions.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "DDxv5gayt63", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1551/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The authors propose a method for predicting a prediction interval. The authors compare this method to deep ensembles and show improved performance on a set of benchmarks.\n\n\nThe are several things I like about this paper:\n- The paper is written quite clear \n- Uncertainty calibration is an important topic.\n- A lot of benchmarks are considered\n- The method is easy to implement.\n\nThese are the things  that could be improved:\n\n1. The authors say \"\"First,  our approach is the first to propose an integrated architecture capable of producing both PIs and exact value predictions.\"\n\nThis is not correct.  A simple approach such as Lakshminarayanan et al., 2017 will produces the same, even without using ensembles.  In that work,  the predictions parameterize a Gaussian with state dependent mean and variance. From this quantiles and thus prediction intervals can be derived on top of an \"exact value prediction\" which would be the mean.\n\n2.  One of the papers main weakness is the lack of distinction between epistemic and aleatoric uncertainty:\n\n\nSection 2:\nThe way  Section 3 is formulated it speaks about aleatoric uncertainty:    P(L_1 < y_i < U_i) is the probability that y_i falls into th ebound defined by L_i and U_i and thus is \"data uncertainty\" (aleatoric) and not due to the lack of knowledge given by the model (epistemic).  it is thus similar to a Softmax in classification or  a neural network that predicts both mean and variance (as in  (Lakshminarayanan et al., 2017)). Indeed as Section 4 show,  the model does not contain any parametric uncertainty. \n\n\n3. Results and method comparisons:\n\nSection 4: \nI am not sure why RMSE is a useful metric when proposing a method for uncertainty calibration.  Secondly, as mentioned above the comparison to deep ensembles is not fair, as deep ensembles will also model epistemic uncertainty.  I would just use an ensemble of size 1 for a more accurate comparison (thus excluding epistemic uncertainty). Thirdly, the two other metrics need to be explained (at least write them once without acronyms.)\n\nOverall,  the improvements appear  to be only marginal improvements.\n\n4. Related work:\nHow is this work different from  [1] and why wasn't it compared to this work?\n\n[1] Salem, Tárik S., Helge Langseth, and Heri Ramampiaro. \"Prediction Intervals: Split Normal Mixture from Quality-Driven Deep Ensembles.\" Conference on Uncertainty in Artificial Intelligence. PMLR, 2020.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Review", "review": "The authors propose a method for predicting a prediction interval. The authors compare this method to deep ensembles and show improved performance on a set of benchmarks.\n\n\nThe are several things I like about this paper:\n- The paper is written quite clear \n- Uncertainty calibration is an important topic.\n- A lot of benchmarks are considered\n- The method is easy to implement.\n\nThese are the things  that could be improved:\n\n1. The authors say \"\"First,  our approach is the first to propose an integrated architecture capable of producing both PIs and exact value predictions.\"\n\nThis is not correct.  A simple approach such as Lakshminarayanan et al., 2017 will produces the same, even without using ensembles.  In that work,  the predictions parameterize a Gaussian with state dependent mean and variance. From this quantiles and thus prediction intervals can be derived on top of an \"exact value prediction\" which would be the mean.\n\n2.  One of the papers main weakness is the lack of distinction between epistemic and aleatoric uncertainty:\n\n\nSection 2:\nThe way  Section 3 is formulated it speaks about aleatoric uncertainty:    P(L_1 < y_i < U_i) is the probability that y_i falls into th ebound defined by L_i and U_i and thus is \"data uncertainty\" (aleatoric) and not due to the lack of knowledge given by the model (epistemic).  it is thus similar to a Softmax in classification or  a neural network that predicts both mean and variance (as in  (Lakshminarayanan et al., 2017)). Indeed as Section 4 show,  the model does not contain any parametric uncertainty. \n\n\n3. Results and method comparisons:\n\nSection 4: \nI am not sure why RMSE is a useful metric when proposing a method for uncertainty calibration.  Secondly, as mentioned above the comparison to deep ensembles is not fair, as deep ensembles will also model epistemic uncertainty.  I would just use an ensemble of size 1 for a more accurate comparison (thus excluding epistemic uncertainty). Thirdly, the two other metrics need to be explained (at least write them once without acronyms.)\n\nOverall,  the improvements appear  to be only marginal improvements.\n\n4. Related work:\nHow is this work different from  [1] and why wasn't it compared to this work?\n\n[1] Salem, Tárik S., Helge Langseth, and Heri Ramampiaro. \"Prediction Intervals: Split Normal Mixture from Quality-Driven Deep Ensembles.\" Conference on Uncertainty in Artificial Intelligence. PMLR, 2020.\n", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1604331581352}, {"id": "xlcx_WvrTay", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1551/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "**Quality and Clarity**\n\nWhile the overall message of the paper is clear, the explanation of the method in Section 4 is a bit hard to follow. Specifically it is a bit hard to keep a track of the meaning of the multiple terms in the loss function and the associated hyper parameters (see Queries and Suggestions below).\n\n**Originality and Significance**\n\nThe authors seek to jointly address the problem of making accurate predictions and generating tight prediction intervals by designing a new loss function that combines these two goals. I believe this work can be quite impactful as there are multiple areas where good quality prediction intervals and accurate value prediction are equally necessary.\n\n**Strengths**\n\nThe proposed approach appears principled and seems to give good empirical results in the experiments considered.\n\n**Weaknesses**\n\nThere seem to be multiple hyperparameters ($\\lambda$,$\\beta$) whose choice might affect the performance of the approach. Moreover as this seems to be the first time such a loss function is used in this context there is no evidence on good choices of hyper parameters (other than the values used in the experiments herein). I would highly appreciate an ablation study or some comments on good values of hyper parameters which can guide readers intending to apply the proposed approach to a new dataset.\n\n**Queries and Suggestions**\n\n1. I'm curious about the performance of the ensemble of PIVEN architectures approach. It does not seem to have been used in any of the experiments which is fine since a single model seems to be doing well. But have you tried it on any dataset? What are the settings where the improvement with an ensemble would be significant enough to compensate for the added cost?\n\n2. Why are the MPIW results for DE on Boston and Concrete not in bold since DE seems to be the best in terms of MPIW on the two datasets?\n\n3. Why is only a single dense layer (and not an ensemble) used when applying DE to large scale image datasets? If the backbone is pre-trained then training an ensemble of dense layers should not be too expensive. The current comparison does not seem fair since the NN baseline only has a MSE loss while the PIVEN approach uses multiple regularizers (for good MPIW, PICP etc.)\n\n4. What is the difference between the MOI variant of PIVEN and the QD baseline?\n\nIn addition to responding to the above queries, I would recommend improving the explanation (one suggestion is to write the entire loss i.e. $\\mathcal{L}_{\\text{PIVEN}}$ at the beginning of Section 4 right after Figure 1 and then explain the meaning of each term in the loss, as opposed to the current approach where the terms are introduced first and then the loss is given) and commenting on good choices of the hyper parameters or including an ablation study for the same.\n\n**Comments after Author Response**\n\nI thank the authors for their response. Queries 1,2, and 4 have been adequately addressed. Regarding Query 3, I appreciate the addition of the Deep Ensemble results though I find that the text of Section 5.5 has not been changed to reflect the same. Specifically the paper still says \"For the IMDB age prediction dataset, results show that PIVEN outperforms both baselines across all metrics\". This is now incorrect since there is a third baseline, DE, which appears to outperform/match PIVEN for this dataset. However the explanation that this is because the population age is approximately Gaussian makes sense to me and so keeping in mind the good performance of PIVEN on the other datasets, and the improved explanation and added ablation study for hyper parameters, I recommend accepting the paper as long as the relevant corrections are made in Section 5.5.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "A new approach for combining prediction and uncertainty quantification in DNNs - good empirical results, explanation can be improved", "review": "**Quality and Clarity**\n\nWhile the overall message of the paper is clear, the explanation of the method in Section 4 is a bit hard to follow. Specifically it is a bit hard to keep a track of the meaning of the multiple terms in the loss function and the associated hyper parameters (see Queries and Suggestions below).\n\n**Originality and Significance**\n\nThe authors seek to jointly address the problem of making accurate predictions and generating tight prediction intervals by designing a new loss function that combines these two goals. I believe this work can be quite impactful as there are multiple areas where good quality prediction intervals and accurate value prediction are equally necessary.\n\n**Strengths**\n\nThe proposed approach appears principled and seems to give good empirical results in the experiments considered.\n\n**Weaknesses**\n\nThere seem to be multiple hyperparameters ($\\lambda$,$\\beta$) whose choice might affect the performance of the approach. Moreover as this seems to be the first time such a loss function is used in this context there is no evidence on good choices of hyper parameters (other than the values used in the experiments herein). I would highly appreciate an ablation study or some comments on good values of hyper parameters which can guide readers intending to apply the proposed approach to a new dataset.\n\n**Queries and Suggestions**\n\n1. I'm curious about the performance of the ensemble of PIVEN architectures approach. It does not seem to have been used in any of the experiments which is fine since a single model seems to be doing well. But have you tried it on any dataset? What are the settings where the improvement with an ensemble would be significant enough to compensate for the added cost?\n\n2. Why are the MPIW results for DE on Boston and Concrete not in bold since DE seems to be the best in terms of MPIW on the two datasets?\n\n3. Why is only a single dense layer (and not an ensemble) used when applying DE to large scale image datasets? If the backbone is pre-trained then training an ensemble of dense layers should not be too expensive. The current comparison does not seem fair since the NN baseline only has a MSE loss while the PIVEN approach uses multiple regularizers (for good MPIW, PICP etc.)\n\n4. What is the difference between the MOI variant of PIVEN and the QD baseline?\n\nIn addition to responding to the above queries, I would recommend improving the explanation (one suggestion is to write the entire loss i.e. $\\mathcal{L}_{\\text{PIVEN}}$ at the beginning of Section 4 right after Figure 1 and then explain the meaning of each term in the loss, as opposed to the current approach where the terms are introduced first and then the loss is given) and commenting on good choices of the hyper parameters or including an ablation study for the same.\n\n**Comments after Author Response**\n\nI thank the authors for their response. Queries 1,2, and 4 have been adequately addressed. Regarding Query 3, I appreciate the addition of the Deep Ensemble results though I find that the text of Section 5.5 has not been changed to reflect the same. Specifically the paper still says \"For the IMDB age prediction dataset, results show that PIVEN outperforms both baselines across all metrics\". This is now incorrect since there is a third baseline, DE, which appears to outperform/match PIVEN for this dataset. However the explanation that this is because the population age is approximately Gaussian makes sense to me and so keeping in mind the good performance of PIVEN on the other datasets, and the improved explanation and added ablation study for hyper parameters, I recommend accepting the paper as long as the relevant corrections are made in Section 5.5.", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603985031788}, {"id": "g7lB8dF7uk", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1551/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\nThe submission considers the continuous real-valued regression problems and how to obtain accurate point predictions (specific value prediction in the text) and prediction intervals (the uncertainty of the predictions, given by [lower bound LB, upper bound UB]). The paper proposes a loss function which is the weighted combination of (i) the coverage constraint [proportion of outputs that fall between LB and UB] so that the coverage meets a required level, (ii) the prediction interval width [for those that satisfy (i), their prediction intervals should be tight], and (iii) the prediction loss [which makes sure the predictions match the outputs]. The LB, UB and predictions are parameterised using three separate heads: one for LB, one for UB and one for the prediction weight (which makes sure the prediction stays in between LB and UB. The proposed loss function and parameterisation architecture are evaluated on UCI regression datasets and two age prediction problems using image inputs. \n\nAssessment:\nWhilst I think the submission tackles an important problem that could be of interest to the ICLR community, the novelty and experimental evaluation are limited and thus I do not recommend acceptance. Below are some of my concerns/questions and I appreciate the response from the authors.\n\n1. The key contribution is the paper is the combination of the prediction interval loss and the prediction loss. Each of these losses are not new, for example: the prediction interval loss has been considered by Pearce et al (2018). The UCI regression also do not show strong evidence that the proposed method is better than QD, in contrast to the claim of “state-of-the-art results in uncertainty modeling by the use of PIs”.\n\n2. The use of a separate head for v is also not strongly supported by the UCI results, compared to directly parameterising the prediction (POO) or MOI. In fact, PIVEN is probably better because of the hyperparameter search over \\beta.\n\n3. The contribution of the paper includes the output head architecture and the loss function, so I’d not call this a deep neural network. \n\n4. In several places, the paper alludes to skewed output distributions and how PIVEN can handle this better than alternatives. However, it is not clear from the experiments that this is the case. PIVEN is very poor at predicting the Sine example in the appendix -- could you clarify this? Could the \"skewed distribution\" argument be formalised theoretically? like the PIVEN objective is inspired from another likelihood which can support skewed outputs compared to the usual Gaussian likelihood/L2 loss?\n\n5. it would be good to compare to the post-hoc calibration procedure for regression by Kuleshov et al (2018)\n\n6. Does the ensemble of PIs work in practice?\n\nminor: abstract: read-world -> real-world", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "limited novelty and evaluation", "review": "Summary:\nThe submission considers the continuous real-valued regression problems and how to obtain accurate point predictions (specific value prediction in the text) and prediction intervals (the uncertainty of the predictions, given by [lower bound LB, upper bound UB]). The paper proposes a loss function which is the weighted combination of (i) the coverage constraint [proportion of outputs that fall between LB and UB] so that the coverage meets a required level, (ii) the prediction interval width [for those that satisfy (i), their prediction intervals should be tight], and (iii) the prediction loss [which makes sure the predictions match the outputs]. The LB, UB and predictions are parameterised using three separate heads: one for LB, one for UB and one for the prediction weight (which makes sure the prediction stays in between LB and UB. The proposed loss function and parameterisation architecture are evaluated on UCI regression datasets and two age prediction problems using image inputs. \n\nAssessment:\nWhilst I think the submission tackles an important problem that could be of interest to the ICLR community, the novelty and experimental evaluation are limited and thus I do not recommend acceptance. Below are some of my concerns/questions and I appreciate the response from the authors.\n\n1. The key contribution is the paper is the combination of the prediction interval loss and the prediction loss. Each of these losses are not new, for example: the prediction interval loss has been considered by Pearce et al (2018). The UCI regression also do not show strong evidence that the proposed method is better than QD, in contrast to the claim of “state-of-the-art results in uncertainty modeling by the use of PIs”.\n\n2. The use of a separate head for v is also not strongly supported by the UCI results, compared to directly parameterising the prediction (POO) or MOI. In fact, PIVEN is probably better because of the hyperparameter search over \\beta.\n\n3. The contribution of the paper includes the output head architecture and the loss function, so I’d not call this a deep neural network. \n\n4. In several places, the paper alludes to skewed output distributions and how PIVEN can handle this better than alternatives. However, it is not clear from the experiments that this is the case. PIVEN is very poor at predicting the Sine example in the appendix -- could you clarify this? Could the \"skewed distribution\" argument be formalised theoretically? like the PIVEN objective is inspired from another likelihood which can support skewed outputs compared to the usual Gaussian likelihood/L2 loss?\n\n5. it would be good to compare to the post-hoc calibration procedure for regression by Kuleshov et al (2018)\n\n6. Does the ensemble of PIs work in practice?\n\nminor: abstract: read-world -> real-world", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603979720409}, {"id": "VGGStxXFgrF", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1551/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a new objective function for training regression networks with prediction intervals. The goal is to provide tight confidence bounds to accompany predictions, which is of course important for practical deployments of ML systems where uncertainty quantification is critical. Previous work has largely assumed that uncertainty is symmetrical, or even Gaussian distributed, which is often not the case in practice. The innovation of this work is to simultaneously predict the bounds for a given confidence level and a point prediction within those bounds, which is not necessarily the mean. It is accomplished by predicting 3 values: an upper bound, a lower bound, and a mixing parameter that allows making the point estimate as a weighted sum of the bounds. Experimental results are provided on 3 datasets, with 2 baseline methods compared against.\n\nThe strong points of this paper are:\n+ The problem being addressed is important.\n+ The approach is relatively simple.\n+ The paper is clear and well-written.\n+ The metrics introduced for evaluation make sense.\n+ The results are strong and improve over the baselines.\n+ The choice of how to produce the point prediction is well motivated and backed up by ablation experiments.\n\nWeak points:\n- Since the method produces bounds rather than a distribution, it's necessary to know at training time what confidence levels you are interested in. It also therefore can't be evaluated on standard negative log likelihood metrics, or produce something akin to accuracy vs rejection-fraction plots. This is potentially a big problem for practical deployment, since the tradeoffs of alpha vs RMSE can't be evaluated without retraining, and the same model can't be used for multiple downstream tasks with different confidence requirements.\n- The claims about using PIVEN in an ensemble context don't seem to be experimentally validated anywhere.\n- The NN baseline for large-scale image datasets seems unnecessarily weak. Why not predict both mean and variance (basically DE, but a single model)? That way you could at least compute PICP and MPIW.\n- In table 1 it's not at all clear what the bold numbers are. They don't seem to correspond to the best result in each row.\n- In table 2, why is PICP not provided? It seems like the value should not be identical between the different experiment arms.\n- It would have been nice to see an evaluation under dataset shift, similar to the \"Flight Delays\" experiment in https://arxiv.org/pdf/2007.05864.pdf\n\nOverall, this seems like a simple and valuable technique that addresses an important problem, but its applicability is somewhat limited by the fact that it produces bounds rather than a distribution.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "PIVEN", "review": "This paper proposes a new objective function for training regression networks with prediction intervals. The goal is to provide tight confidence bounds to accompany predictions, which is of course important for practical deployments of ML systems where uncertainty quantification is critical. Previous work has largely assumed that uncertainty is symmetrical, or even Gaussian distributed, which is often not the case in practice. The innovation of this work is to simultaneously predict the bounds for a given confidence level and a point prediction within those bounds, which is not necessarily the mean. It is accomplished by predicting 3 values: an upper bound, a lower bound, and a mixing parameter that allows making the point estimate as a weighted sum of the bounds. Experimental results are provided on 3 datasets, with 2 baseline methods compared against.\n\nThe strong points of this paper are:\n+ The problem being addressed is important.\n+ The approach is relatively simple.\n+ The paper is clear and well-written.\n+ The metrics introduced for evaluation make sense.\n+ The results are strong and improve over the baselines.\n+ The choice of how to produce the point prediction is well motivated and backed up by ablation experiments.\n\nWeak points:\n- Since the method produces bounds rather than a distribution, it's necessary to know at training time what confidence levels you are interested in. It also therefore can't be evaluated on standard negative log likelihood metrics, or produce something akin to accuracy vs rejection-fraction plots. This is potentially a big problem for practical deployment, since the tradeoffs of alpha vs RMSE can't be evaluated without retraining, and the same model can't be used for multiple downstream tasks with different confidence requirements.\n- The claims about using PIVEN in an ensemble context don't seem to be experimentally validated anywhere.\n- The NN baseline for large-scale image datasets seems unnecessarily weak. Why not predict both mean and variance (basically DE, but a single model)? That way you could at least compute PICP and MPIW.\n- In table 1 it's not at all clear what the bold numbers are. They don't seem to correspond to the best result in each row.\n- In table 2, why is PICP not provided? It seems like the value should not be identical between the different experiment arms.\n- It would have been nice to see an evaluation under dataset shift, similar to the \"Flight Delays\" experiment in https://arxiv.org/pdf/2007.05864.pdf\n\nOverall, this seems like a simple and valuable technique that addresses an important problem, but its applicability is somewhat limited by the fact that it produces bounds rather than a distribution.", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603487017617}], "openreview_url": "https://openreview.net/forum?id=qn_gk5j3PJ", "arxiv_id": "2006.05139", "paper_pdf": "papers/qn_gk5j3PJ.pdf", "paper_pdf_sha256": "09e5f450f0190b32544022760c49a533f6836077452b5b688d2998e72a161171", "paper_pdf_bytes": 4424000, "paper_pdf_source": "openreview", "code_url": "https://github.com/elisim/piven", "code_repository": "elisim/piven", "code_commit": "ecdfc024f3e2f63b10a930039b3e6ada3c3c74d4", "code_archive": "repos/qn_gk5j3PJ.zip", "code_archive_sha256": "4174c9f5b706247f718cedf24dcbc32858f14a9aab174fbf9e1f69b7772f4cec", "code_archive_bytes": 5189515, "code_file_count": 19, "code_extensions": {".py": 15, ".sh": 3, ".ipynb": 1}, "github_disk_usage_kb": 5518, "github_languages": {"Jupyter Notebook": 309667, "Python": 110570, "Shell": 585}, "github_archived": false, "github_pushed_at": "2023-03-12T05:51:21Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/piven-a-deep-neural-network-for-prediction"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "7iCT2vmYAR", "year": 2025, "status": "rejected", "title": "Contrastive learning of cell state dynamics in response to perturbations", "authors": ["Soorya Pradeep", "Alishba Imran", "Ziwen Liu", "Eduardo Hirata-Miyasaki", "Taylla Milena Theodoro", "Ivan E. Ivanov", "Madhura Bhave", "Sudip Khadka", "Hunter Woosley", "Carolina Arias", "Shalin B. Mehta"], "authorids": ["~Soorya_Pradeep1", "~Alishba_Imran1", "~Ziwen_Liu5", "~Eduardo_Hirata-Miyasaki1", "~Taylla_Milena_Theodoro1", "~Ivan_E._Ivanov1", "~Madhura_Bhave1", "~Sudip_Khadka1", "~Hunter_Woosley1", "~Carolina_Arias1", "~Shalin_B._Mehta1"], "authors_source": "OpenReview API", "abstract": "We introduce dynaCLR, a self-supervised framework for modeling cell and organelle dynamics via contrastive learning of representations of time-lapse datasets. Live cell imaging of cells and organelles is widely used to analyze cellular responses to perturbations. Supervised modeling of dynamic cell states encoded in 3D time-lapse data is laborious and prone to bias. dy- naCLR leverages single-cell tracking and time-aware contrastive sampling to map images of cells at neighboring time points to neighboring embed- dings. We illustrate the features and applications of dynaCLR with the following experiments: analyzing the kinetics of viral infection in human cells, detecting transient changes in cell morphology due to cell division, and mapping the dynamics of organelles due to viral infection. Temporally regularized embeddings computed with dynaCLR models enable efficient and quantitative annotation, classification, clustering, or interpretation of the cell states. The models reliably embed, i.e., generalize to, data from un- seen experiments with different microscopes and imaging contrasts. Models trained with dynaCLR consistently achieve > 95% accuracy in mitosis and infection state classification, enable the detection of transient cell states and reliably embed unseen experiments. dynaCLR provides a flexible framework for comparative analysis of cell state dynamics due to perturbations, such as infection, gene knockouts, and drugs. We provide PyTorch-based implementations of the model training and inference pipeline and a napari plugin user interface for the visualization and annotation of trajectories of cells in the real space and the embedding space.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "I2Hxuvotxr", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2024/Reviewer_oJPV"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "The authors present a self-supervised framework for leveraging contrastive learning to model cell state dynamics from time-lapse imaging. \nThe model allows temporally adjacent states to be mapped closely together helping achieve accurate, efficient, and label-free analysis of dynamic cell states under perturbations like viral infection. The paper presents a unique application of contrastive learning in cellular imaging and stands out for its temporal coherence and robustness in infection state classification.", "review_text": "The authors present a self-supervised framework for leveraging contrastive learning to model cell state dynamics from time-lapse imaging. \nThe model allows temporally adjacent states to be mapped closely together helping achieve accurate, efficient, and label-free analysis of dynamic cell states under perturbations like viral infection. The paper presents a unique application of contrastive learning in cellular imaging and stands out for its temporal coherence and robustness in infection state classification.", "strengths": "- The use of temporally-aware contrastive learning enables efficient modeling of time-dependent cellular changes, showing improved performance over traditional contrastive methods, especially under perturbative conditions.\n- The framework facilitates rapid annotation of cell states, potentially decreasing the reliance on human-intensive and subjective labeling, a significant advancement over prior approaches.\n- By using a cell-aware approach, the framework attempts to address the intrinsic heterogeneity across cell populations, an advantage over traditional time-agnostic contrastive approaches.", "weaknesses": "- The choice of a 30-minute interval as the temporal offset might not generalize across other biological systems with different dynamics, limiting the model adaptability.\n- The reliance on phase and fluorescence imaging could constrain its utility where alternate modalities are necessary.\n- Since cell-aware and time-aware sampling use specific tracked cells, the embeddings may risk overfitting to individual cell trajectories instead of generalized dynamics.\n- Although contrastive learning was chosen, the paper lacks in-depth comparisons with generative methods that authors summarize in the related work.\n- By setting a fixed temporal offset, the model may miss capturing events that unfold asynchronously or at variable rates in different cells.\n- Models relying on phase channels for cell division detection may struggle with subtler morphological changes that require fluorescence markers.", "questions": "- How does the model perform for other tasks beyond infection classification? Like, for example, tracking mitotic spindle dynamics during first cell division in embryonic development? \n\n- How does the model handle potential noise or artifacts in the time-lapse imaging data? How are hyperparameters tuned and how sensitive is the model to the choice of these hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors present a self-supervised framework for leveraging contrastive learning to model cell state dynamics from time-lapse imaging. \nThe model allows temporally adjacent states to be mapped closely together helping achieve accurate, efficient, and label-free analysis of dynamic cell states under perturbations like viral infection. The paper presents a unique application of contrastive learning in cellular imaging and stands out for its temporal coherence and robustness in infection state classification.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The use of temporally-aware contrastive learning enables efficient modeling of time-dependent cellular changes, showing improved performance over traditional contrastive methods, especially under perturbative conditions.\n- The framework facilitates rapid annotation of cell states, potentially decreasing the reliance on human-intensive and subjective labeling, a significant advancement over prior approaches.\n- By using a cell-aware approach, the framework attempts to address the intrinsic heterogeneity across cell populations, an advantage over traditional time-agnostic contrastive approaches.", "weaknesses": "- The choice of a 30-minute interval as the temporal offset might not generalize across other biological systems with different dynamics, limiting the model adaptability.\n- The reliance on phase and fluorescence imaging could constrain its utility where alternate modalities are necessary.\n- Since cell-aware and time-aware sampling use specific tracked cells, the embeddings may risk overfitting to individual cell trajectories instead of generalized dynamics.\n- Although contrastive learning was chosen, the paper lacks in-depth comparisons with generative methods that authors summarize in the related work.\n- By setting a fixed temporal offset, the model may miss capturing events that unfold asynchronously or at variable rates in different cells.\n- Models relying on phase channels for cell division detection may struggle with subtler morphological changes that require fluorescence markers.", "questions": "- How does the model perform for other tasks beyond infection classification? Like, for example, tracking mitotic spindle dynamics during first cell division in embryonic development? \n\n- How does the model handle potential noise or artifacts in the time-lapse imaging data? How are hyperparameters tuned and how sensitive is the model to the choice of these hyperparameters?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730704650211}, {"id": "7aErqdlktY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2024/Reviewer_aWZ5"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 3, "summary": "In this paper, the authors propose a framework for modeling cell dynamics in response to perturbations. They suggested several downstream tasks for the learned representation, such as the analysis of viral infection kinetics in human cells, detecting transient changes in cell morphology, and mapping organelle dynamics due to viral infection. Furthermore, they reported that the proposed framework achieves an accuracy of over 95% for infection state classification, outperforming the supervised setting.", "review_text": "In this paper, the authors propose a framework for modeling cell dynamics in response to perturbations. They suggested several downstream tasks for the learned representation, such as the analysis of viral infection kinetics in human cells, detecting transient changes in cell morphology, and mapping organelle dynamics due to viral infection. Furthermore, they reported that the proposed framework achieves an accuracy of over 95% for infection state classification, outperforming the supervised setting.", "strengths": "- The paper is well-written, following a clear structure that enhances readability and comprehension.\n- The learned representation was used in several dowstream tasks.", "weaknesses": "- The technical novelty of the paper is limited, as DynaCLR may be considered a straightforward application of the triplet loss with varied sampling strategies.\n-  Some figures (Fig. 2, Fig. 3, and Fig. 6) are not thoroughly explained; a more detailed description would enhance clarity.", "questions": "- The choice of the triplet loss is not justified. Why is this loss more suitable for this task than alternatives like NT-Xent or InfoNCE?\n- In Section 4.1.1, could you clarify ithe origin of the independent test data?\n- In Section 4.2, the smooth transitions and tight clustering of division events are not immediately evident in Figure 4; additional support for these claims would be beneficial.\n- In Section 4.3, it remains unclear how the referenced figures demonstrate that \"the encoder learns meaningful features that describe cell dynamics.\"\n- Figure 6 would benefit from a more detailed explanation.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "In this paper, the authors propose a framework for modeling cell dynamics in response to perturbations. They suggested several downstream tasks for the learned representation, such as the analysis of viral infection kinetics in human cells, detecting transient changes in cell morphology, and mapping organelle dynamics due to viral infection. Furthermore, they reported that the proposed framework achieves an accuracy of over 95% for infection state classification, outperforming the supervised setting.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "- The paper is well-written, following a clear structure that enhances readability and comprehension.\n- The learned representation was used in several dowstream tasks.", "weaknesses": "- The technical novelty of the paper is limited, as DynaCLR may be considered a straightforward application of the triplet loss with varied sampling strategies.\n-  Some figures (Fig. 2, Fig. 3, and Fig. 6) are not thoroughly explained; a more detailed description would enhance clarity.", "questions": "- The choice of the triplet loss is not justified. Why is this loss more suitable for this task than alternatives like NT-Xent or InfoNCE?\n- In Section 4.1.1, could you clarify ithe origin of the independent test data?\n- In Section 4.2, the smooth transitions and tight clustering of division events are not immediately evident in Figure 4; additional support for these claims would be beneficial.\n- In Section 4.3, it remains unclear how the referenced figures demonstrate that \"the encoder learns meaningful features that describe cell dynamics.\"\n- Figure 6 would benefit from a more detailed explanation.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730631457196}, {"id": "V9N5OXebTt", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission2024/Reviewer_UA5v"], "rating": 3, "soundness": 2, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The paper describes an application of contrastive learning on a microscopy dataset to model single cell dynamics. The method introduces time-aware sampling to the traditional contrastive learning methodology by making use of cell-tracking. Three downstream applications of the learned embeddings are shown which demonstrate how the method can be used to analyze cell-state dynamics in response to perturbations. In addition, experiments showing generalization and interpretability of embeddings are performed.", "review_text": "The paper describes an application of contrastive learning on a microscopy dataset to model single cell dynamics. The method introduces time-aware sampling to the traditional contrastive learning methodology by making use of cell-tracking. Three downstream applications of the learned embeddings are shown which demonstrate how the method can be used to analyze cell-state dynamics in response to perturbations. In addition, experiments showing generalization and interpretability of embeddings are performed.", "strengths": "1.\tThe paper addresses an interesting topic where self-supervised learning has the potential to be extremely beneficial. Getting annotations for large amounts of microscopy data on a single-cell level is difficult and there is a broad push in the field towards the kind of self-supervised approach described here, so this work is well placed.\n2.\tThe utilization of cell tracking to improve self-supervised learning is an interesting idea; cell-identity detection is potentially a useful proxy task that could aid in learning good representations.\n3.\tThe biological setup and experiments are well thought out and the dataset would be beneficial to researchers as a benchmark, given simple readouts like infection rate.\n4.\tThe addition of generalization experiments (though only on one additional set of data) provides some grounding to the results – without this section I don’t think the paper has much weight. I would encourage the authors to extend this to more datasets if possible.", "weaknesses": "1.\tIn general, this is written like a paper in a scientific journal rather than a machine learning conference. While it is important to provide the necessary biological context, the paper lacks the mathematical discourse required to be suitable in a machine-learning setting. For example, could you add a theoretical or mathematical description of ideas such as ‘smoothness’ or ‘richness’ of the latent space that you have mentioned?  \n2.\tI’m not quite sure that the quantitative results match up with the claims made. For example, performance scores in Table 1 are better for the ‘classical’ contrastive learning approach as opposed to the ‘cell-aware’ and ‘time-aware’ approaches. The arguments made based on the Euclidean distance plots in Figure 2 are strenuous, and I don’t agree with the qualitative conclusions made about time-regularized contrastive learning being better than the classical approach from these results (I will add more on this in the next points). Overall, I don’t think the contributions here are significant enough and the arguments don’t hold enough weight for me to accept this paper. \n3.\tThe whole ‘temporal continuity’ argument seems a little counterintuitive to me. The ‘time-aware’ loss function is designed to encourage the model to ignore differences at the time-scale of the time hyperparameter Tau, but you expect the model to maintain temporal differences at time-scales t > Tau. I think Tau needs to be much smaller than the time-scale of your expected changes for this to work. If you expect changes in cells to happen over a few hours, the hyperparameter Tau being 30 minutes doesn’t make sense – your model is just going to smooth out all temporal information. I know you are limited by the fact that you can’t take images, say, every 5 minutes, but the current setup doesn’t make sense to me.\n4.\tJust to add to the previous argument – in your current setup, it matters how long you train your model for, since overtraining could completely smooth out all temporal information. I don’t even know how you would design experiments to determine what a good number of epochs to train is without having an idea of how much temporal smoothness is ‘correct’. For example, in the limit of an infinite number of epochs under this loss function, all temporal information would be removed from your model. I just don’t see how you could both encourage temporal smoothness while maintaining a temporally faithful embedding when your data has a temporal resolution of 30 minutes over 24 hours.\n5.\tIn the case of ‘time and cell-aware sampling’, how do you ensure that the model is robust to imaging factors like brightness or sensor noise? The point of such augmentations in the classical contrastive learning case is to teach the model to ignore spurious confounders which may actually show up in your real data – but the ‘time and cell-aware’ model never learns to do this. Is this model actually good at generalizing? What happens of your imaging conditions change a little? Would this model still work? I think many questions need to be answered here.\n6.\tI think the novelty proposed here lacks conceptual correctness in my view. With traditional ways of using temporal information in self-supervised learning, like predicting the temporal order of randomly flipped images, it makes sense how these would lead to a temporally unbiased and meaningful embedding space. My arguments in the points above highlight my concerns on why the proposed method may not be doing the same.\n7.\tWithout actual ground truth annotations of fine-grained temporality, the only way to assess the quality of the embeddings is through the results on downstream temporal tasks. However, the classical contrastive sampling variant seems to perform better on the downstream infection classification task, which tells me that the classical method is leading to better quality embeddings than the proposed method.", "questions": "1.\tI’m a little bit confused about the quantitative results. Specifically, you write “Therefore, we evaluate the accuracy of visual representation learned by our method using a biologically relevant benchmark: accuracy of the classification of the cell states with 3 hours of expert annotations. We compare our method with two baseline methods: fully supervised time-agnostic semantic segmentation of the infection state and time-agnostic contrastive learning. Compared to the ≈ 80% accuracy achieved by the supervised model and 60 − 65% accuracy achieved by the time-agnostic contrastive learning, DynaCLR models consistently achieve ≈ 95% accuracy.” I only see results in Table 1 that show classical contrastive sampling having an accuracy of 98.8%. Can you clarify where the results of the supervised model are and why there is a mismatch between the claimed values of 60-65% and the values shown in the table?\n2.\tWhen you say, “For the experiments in this paper, we set τ = 30 min, as we found this captures significant cell changes while maintaining temporal continuity without over-sampling.” Does this mean you tested with different τ values? If you did any experiments on this it would be good to include those, and would contribute to my concerns in the weaknesses section on temporal continuity and temporal faithfulness.\n3.\tIn Section 3.3 you claim, “We compute (2) and (3) to evaluate the temporal evolution of the embeddings as a gradual, steady increase signifies strong temporal smoothness. In both figure 2 a) and 2 b), cell and time-aware sampling shows a smooth rising curve with minimal fluctuations. The classical contrastive method exhibits a rapid initial increase followed by plateaus, while the cell-aware approach shows intermittent fluctuations.” I’m not sure I see in Figure 2 what is described here. It seems like the curves for Euclidean distance are pretty similar between the different strategies. In fact, in the UMAP version it seems like the cell and time-aware version is most fluctuating (whether UMAP is an appropriate space in which to perform this is another question, since it is known to not maintain global structure). Can you please elaborate on this? I’m also unsure what you mean by “richness” of representation in Figure 2C and how you determined this.\n4.\tCan you elaborate the results of the integrated gradients based interpretability in Figure A1? What exactly is the model focusing on in the two channels? Can you provide evidence that this is relevant to the output task?\n5.\tYou say “We observe a clear transition of cell states from interphase to mitosis as we follow the cells in UMAP space, particularly in models trained solely with the phase channel and incorporating temporal regularization” I kind of see what you’re trying to say but it’s still very unclear. Can you provide a mathematical comparison of the two models and show that one model is ‘smoother’ than the other over all cell division events?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper describes an application of contrastive learning on a microscopy dataset to model single cell dynamics. The method introduces time-aware sampling to the traditional contrastive learning methodology by making use of cell-tracking. Three downstream applications of the learned embeddings are shown which demonstrate how the method can be used to analyze cell-state dynamics in response to perturbations. In addition, experiments showing generalization and interpretability of embeddings are performed.", "soundness": 2, "presentation": 3, "contribution": 2, "strengths": "1.\tThe paper addresses an interesting topic where self-supervised learning has the potential to be extremely beneficial. Getting annotations for large amounts of microscopy data on a single-cell level is difficult and there is a broad push in the field towards the kind of self-supervised approach described here, so this work is well placed.\n2.\tThe utilization of cell tracking to improve self-supervised learning is an interesting idea; cell-identity detection is potentially a useful proxy task that could aid in learning good representations.\n3.\tThe biological setup and experiments are well thought out and the dataset would be beneficial to researchers as a benchmark, given simple readouts like infection rate.\n4.\tThe addition of generalization experiments (though only on one additional set of data) provides some grounding to the results – without this section I don’t think the paper has much weight. I would encourage the authors to extend this to more datasets if possible.", "weaknesses": "1.\tIn general, this is written like a paper in a scientific journal rather than a machine learning conference. While it is important to provide the necessary biological context, the paper lacks the mathematical discourse required to be suitable in a machine-learning setting. For example, could you add a theoretical or mathematical description of ideas such as ‘smoothness’ or ‘richness’ of the latent space that you have mentioned?  \n2.\tI’m not quite sure that the quantitative results match up with the claims made. For example, performance scores in Table 1 are better for the ‘classical’ contrastive learning approach as opposed to the ‘cell-aware’ and ‘time-aware’ approaches. The arguments made based on the Euclidean distance plots in Figure 2 are strenuous, and I don’t agree with the qualitative conclusions made about time-regularized contrastive learning being better than the classical approach from these results (I will add more on this in the next points). Overall, I don’t think the contributions here are significant enough and the arguments don’t hold enough weight for me to accept this paper. \n3.\tThe whole ‘temporal continuity’ argument seems a little counterintuitive to me. The ‘time-aware’ loss function is designed to encourage the model to ignore differences at the time-scale of the time hyperparameter Tau, but you expect the model to maintain temporal differences at time-scales t > Tau. I think Tau needs to be much smaller than the time-scale of your expected changes for this to work. If you expect changes in cells to happen over a few hours, the hyperparameter Tau being 30 minutes doesn’t make sense – your model is just going to smooth out all temporal information. I know you are limited by the fact that you can’t take images, say, every 5 minutes, but the current setup doesn’t make sense to me.\n4.\tJust to add to the previous argument – in your current setup, it matters how long you train your model for, since overtraining could completely smooth out all temporal information. I don’t even know how you would design experiments to determine what a good number of epochs to train is without having an idea of how much temporal smoothness is ‘correct’. For example, in the limit of an infinite number of epochs under this loss function, all temporal information would be removed from your model. I just don’t see how you could both encourage temporal smoothness while maintaining a temporally faithful embedding when your data has a temporal resolution of 30 minutes over 24 hours.\n5.\tIn the case of ‘time and cell-aware sampling’, how do you ensure that the model is robust to imaging factors like brightness or sensor noise? The point of such augmentations in the classical contrastive learning case is to teach the model to ignore spurious confounders which may actually show up in your real data – but the ‘time and cell-aware’ model never learns to do this. Is this model actually good at generalizing? What happens of your imaging conditions change a little? Would this model still work? I think many questions need to be answered here.\n6.\tI think the novelty proposed here lacks conceptual correctness in my view. With traditional ways of using temporal information in self-supervised learning, like predicting the temporal order of randomly flipped images, it makes sense how these would lead to a temporally unbiased and meaningful embedding space. My arguments in the points above highlight my concerns on why the proposed method may not be doing the same.\n7.\tWithout actual ground truth annotations of fine-grained temporality, the only way to assess the quality of the embeddings is through the results on downstream temporal tasks. However, the classical contrastive sampling variant seems to perform better on the downstream infection classification task, which tells me that the classical method is leading to better quality embeddings than the proposed method.", "questions": "1.\tI’m a little bit confused about the quantitative results. Specifically, you write “Therefore, we evaluate the accuracy of visual representation learned by our method using a biologically relevant benchmark: accuracy of the classification of the cell states with 3 hours of expert annotations. We compare our method with two baseline methods: fully supervised time-agnostic semantic segmentation of the infection state and time-agnostic contrastive learning. Compared to the ≈ 80% accuracy achieved by the supervised model and 60 − 65% accuracy achieved by the time-agnostic contrastive learning, DynaCLR models consistently achieve ≈ 95% accuracy.” I only see results in Table 1 that show classical contrastive sampling having an accuracy of 98.8%. Can you clarify where the results of the supervised model are and why there is a mismatch between the claimed values of 60-65% and the values shown in the table?\n2.\tWhen you say, “For the experiments in this paper, we set τ = 30 min, as we found this captures significant cell changes while maintaining temporal continuity without over-sampling.” Does this mean you tested with different τ values? If you did any experiments on this it would be good to include those, and would contribute to my concerns in the weaknesses section on temporal continuity and temporal faithfulness.\n3.\tIn Section 3.3 you claim, “We compute (2) and (3) to evaluate the temporal evolution of the embeddings as a gradual, steady increase signifies strong temporal smoothness. In both figure 2 a) and 2 b), cell and time-aware sampling shows a smooth rising curve with minimal fluctuations. The classical contrastive method exhibits a rapid initial increase followed by plateaus, while the cell-aware approach shows intermittent fluctuations.” I’m not sure I see in Figure 2 what is described here. It seems like the curves for Euclidean distance are pretty similar between the different strategies. In fact, in the UMAP version it seems like the cell and time-aware version is most fluctuating (whether UMAP is an appropriate space in which to perform this is another question, since it is known to not maintain global structure). Can you please elaborate on this? I’m also unsure what you mean by “richness” of representation in Figure 2C and how you determined this.\n4.\tCan you elaborate the results of the integrated gradients based interpretability in Figure A1? What exactly is the model focusing on in the two channels? Can you provide evidence that this is relevant to the output task?\n5.\tYou say “We observe a clear transition of cell states from interphase to mitosis as we follow the cells in UMAP space, particularly in models trained solely with the phase channel and incorporating temporal regularization” I kind of see what you’re trying to say but it’s still very unclear. Can you provide a mathematical comparison of the two models and show that one model is ‘smoother’ than the other over all cell division events?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730574374415}], "openreview_url": "https://openreview.net/forum?id=7iCT2vmYAR", "arxiv_id": "2410.11281", "paper_pdf": "papers/7iCT2vmYAR.pdf", "paper_pdf_sha256": "a54aa992dc3b37a0890ef584d0b47c63d54628cc7d9851ca799b2dc34f21024c", "paper_pdf_bytes": 33399544, "paper_pdf_source": "openreview", "code_url": "https://github.com/czbiohub-sf/napari-iohub", "code_repository": "czbiohub-sf/napari-iohub", "code_commit": "6a5d13f16af0b6a7865a117e9309a837743ee28c", "code_archive": "repos/7iCT2vmYAR.zip", "code_archive_sha256": "b9ba16113e16ed3347b8b05184ddf21dd3e603702c8b18800bf080cee9ea7930", "code_archive_bytes": 36012, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 442, "github_languages": {"Python": 76355}, "github_archived": false, "github_pushed_at": "2026-05-21T17:36:06Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/contrastive-learning-of-cell-state-dynamics"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Gpp1dfvZYYH", "year": 2022, "status": "rejected", "title": "ProgFed: Effective, Communication, and Computation Efficient Federated Learning by Progressive Training", "authors": ["Hui-Po Wang", "Sebastian U Stich", "Yang He", "Mario Fritz"], "authorids": ["~Hui-Po_Wang1", "~Sebastian_U_Stich1", "yang.he@cispa.saarland", "~Mario_Fritz1"], "authors_source": "OpenReview API", "abstract": "Federated learning is a powerful distributed learning scheme that allows numerous edge devices to collaboratively train a model without sharing their data. However, training is resource-intensive for edge devices, and limited network bandwidth is often the main bottleneck. Prior work often overcomes the constraints by condensing the models or messages into compact formats, e.g., by gradient compression or distillation. In contrast, we propose ProgFed, the first progressive training framework for efficient and effective federated learning. It inherently reduces computation and two-way communication costs while maintaining the strong performance of the final models. We theoretically prove that ProgFed converges at the same asymptotic rate as standard training on full models. Extensive results on a broad range of architectures, including CNNs (VGG, ResNet, ConvNets) and U-nets, and diverse tasks from simple classification to medical image segmentation show that our highly effective training approach saves up to $20\\%$ computation and up to $63\\%$ communication costs for converged models. As our approach is also complimentary to prior work on compression, we can achieve a wide range of trade-offs, showing reduced communication of up to $50\\times$ at only $0.1\\%$ loss in utility. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "E70sEHFDGtL", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper473/Reviewer_3Huv"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper introduces the idea of progressive training of deep networks for federated learning, where only some layer blocks are trained initially and the remaining blocks are progressively added, in order to reduce both local computation cost and communication cost in the early stages (as there are fewer weights). The authors provide a convergence proof of this approach, showing the same asymptotic convergence rate as plain training.\nNumerical experiments are conducted in EMNIST, CIFAR-10/100 and BraTS, split either in an IID or non-IID fashion and representing either a cross-device or cross-silo setting. Numerical experiments show a highly architecture-dependent reduction in the computation cost in the centralised setting. In the federated setting, communication gains significant communication gains are also achieved. Finally, the authors show that progressive training is compatible with other compression methods, increasing the gain in communication cost.", "review_text": "# Novelty & Significance:\n\n- The idea of progressive training is not new, and the authors cite in particular Karras et al 2018. Given the importance of the topic in the paper, I would have appreciated a paragraph devoted to it in the related work section.\n- Theoretical results follow a previous proof structure but seem to bring some novelty. Their significance is not that big insofar as the paper is concerned with deep learning, for which convergence rates do not play a huge role (in particular, the learning rate schedule is chosen independently of the rate derived). Further, the final result is that the convergence is at worst twice slower than the baseline, which is a simple consequence of the choice of spending at least half of the training budget on the full network.\n- The experimental results show that the idea can bring some computational and communication savings, in a variety of contexts (cross-device/silo), and is compatible with other approaches.\n    - Re computational gains, the benefits seem highly dependent on the architecture (cf fig 3), which reduces its significance.\n    - The proposed approach seems only to bring very modest gains with respect to a « random » approach where one would train random sub-networks (cf Table 5). This reduces the significance of the results.\n\n# Writing & Clarity:\n- The paper is well written and the main idea is way conveyed.\n- However, notations could be simplified (e.g. page 4 the notation $x_{t, E_s^c}$ is introduced but never used) to facilitate understanding\n- I would suggest using a log range for the compression factor axis on figs 6a and 6b.\n\nHere are a few minor typos:\n- Page 4, paragraph training of progressive model, there is a typo for $T_S$, I think the correct value should be $T (S+1) / 2S$\n- Circular notations for the theoretical analysis (end of page 4) for $s = \\min(S, t/T_s)$\n- End of sec 4.5, missing « is » easy to implement\n- Proof of lemma 2, page 13, the first update equation is probably missing a $\\gamma_t$ before $g_t^s$ \n\n# Quality & Correctness:\n\n## Theoretical results:\n\n- I did not check the details of the proof of Theorem 1.\n- Regarding the results, the choice of learning step $\\alpha_t$ seems to be hardly usable in practice, as it relies on the knowledge of both the full gradient and the gradient of the sub-network.\n\n## Empirical results\n\nComputation reduction;\n- I am not sure to completely understand the metric used in the x-axis of Fig 2, in terms of unit used. Is it the total number of FLOP (i.e. the number of floating point operations over time integrated along time)? \n- Why isn’t BraTS reported as well for the computation cost reduction benchmark?\n\nCommunication reduction:\n\n- Which architecture is used for the results reported in Fig 4 and 5 for the classification tasks? Additional results are reported in Appendix for varying architectures, but it would be nice to precise it in the main text\n- For multiple experiments, the authors average results over 3 seeds, which is a good effort towards reproducibility. Would it be possible to perform 5 runs in order to get stds ? Otherwise, to have access to the individual values?\n- There are some instabilities at each step transition (cf Fig 5), and I suspect that warm-up seems to play some role here. The paper is silent about this aspect there were no ablation studies or investigation on this particular ingredient. I would have appreciated a focus on this part rather than e.g.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper introduces the idea of progressive training of deep networks for federated learning, where only some layer blocks are trained initially and the remaining blocks are progressively added, in order to reduce both local computation cost and communication cost in the early stages (as there are fewer weights). The authors provide a convergence proof of this approach, showing the same asymptotic convergence rate as plain training.\nNumerical experiments are conducted in EMNIST, CIFAR-10/100 and BraTS, split either in an IID or non-IID fashion and representing either a cross-device or cross-silo setting. Numerical experiments show a highly architecture-dependent reduction in the computation cost in the centralised setting. In the federated setting, communication gains significant communication gains are also achieved. Finally, the authors show that progressive training is compatible with other compression methods, increasing the gain in communication cost.", "main_review": "# Novelty & Significance:\n\n- The idea of progressive training is not new, and the authors cite in particular Karras et al 2018. Given the importance of the topic in the paper, I would have appreciated a paragraph devoted to it in the related work section.\n- Theoretical results follow a previous proof structure but seem to bring some novelty. Their significance is not that big insofar as the paper is concerned with deep learning, for which convergence rates do not play a huge role (in particular, the learning rate schedule is chosen independently of the rate derived). Further, the final result is that the convergence is at worst twice slower than the baseline, which is a simple consequence of the choice of spending at least half of the training budget on the full network.\n- The experimental results show that the idea can bring some computational and communication savings, in a variety of contexts (cross-device/silo), and is compatible with other approaches.\n    - Re computational gains, the benefits seem highly dependent on the architecture (cf fig 3), which reduces its significance.\n    - The proposed approach seems only to bring very modest gains with respect to a « random » approach where one would train random sub-networks (cf Table 5). This reduces the significance of the results.\n\n# Writing & Clarity:\n- The paper is well written and the main idea is way conveyed.\n- However, notations could be simplified (e.g. page 4 the notation $x_{t, E_s^c}$ is introduced but never used) to facilitate understanding\n- I would suggest using a log range for the compression factor axis on figs 6a and 6b.\n\nHere are a few minor typos:\n- Page 4, paragraph training of progressive model, there is a typo for $T_S$, I think the correct value should be $T (S+1) / 2S$\n- Circular notations for the theoretical analysis (end of page 4) for $s = \\min(S, t/T_s)$\n- End of sec 4.5, missing « is » easy to implement\n- Proof of lemma 2, page 13, the first update equation is probably missing a $\\gamma_t$ before $g_t^s$ \n\n# Quality & Correctness:\n\n## Theoretical results:\n\n- I did not check the details of the proof of Theorem 1.\n- Regarding the results, the choice of learning step $\\alpha_t$ seems to be hardly usable in practice, as it relies on the knowledge of both the full gradient and the gradient of the sub-network.\n\n## Empirical results\n\nComputation reduction;\n- I am not sure to completely understand the metric used in the x-axis of Fig 2, in terms of unit used. Is it the total number of FLOP (i.e. the number of floating point operations over time integrated along time)? \n- Why isn’t BraTS reported as well for the computation cost reduction benchmark?\n\nCommunication reduction:\n\n- Which architecture is used for the results reported in Fig 4 and 5 for the classification tasks? Additional results are reported in Appendix for varying architectures, but it would be nice to precise it in the main text\n- For multiple experiments, the authors average results over 3 seeds, which is a good effort towards reproducibility. Would it be possible to perform 5 runs in order to get stds ? Otherwise, to have access to the individual values?\n- There are some instabilities at each step transition (cf Fig 5), and I suspect that warm-up seems to play some role here. The paper is silent about this aspect there were no ablation studies or investigation on this particular ingredient. I would have appreciated a focus on this part rather than e.g.", "summary_of_the_review": "This paper introduces an existing scheme into the federated setting, and provides an experimental proof of the gains it brings in terms of local computations and communication. Despite some minor issues which can be fixed, I think this paper could be accepted.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635778111718}, {"id": "uAk2eFfnqpM", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper473/Reviewer_1N8v"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "The authors adopt the progressive training technique to accelerate the training process federated learning. The authors also establish the convergence rate of the proposed algorithm. Extensive experiments demonstrate the efficacy of the proposed algorithm.  ", "review_text": "It is an interesting work on accelerating the training process of federated learning.  The reviewer still has several concerns about the current submission:\n(1) It seems that the main benefit of progressive training for federated learning comes from the early stage during training.  Hence, we recommend the authors do more ablation studies on the hyperparameter $S$. \n(2) For different stages,  how do the authors set the hyperparameters, such as learning rate.  Since in different stages, the depth of subnetworks is differents,  the hyperparameter tuning techniques may highly influence the training speed and final performance.\n(3) Theorem 1 provides the convergence of the proposed ProgFed.  However, the convergence results do not indicate the benefit of progressive training for federated learning. \n(4)  The experiments should be strengthened.  The main optimizer in ProgFed is fedavg.  The authors should incorporate the proposed ProgFed with several other types of local opitmizers, such as fedprox, scaffold, etc., to further show progressive training can be a versatile and strong technique for accelerating federated learning. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The authors adopt the progressive training technique to accelerate the training process federated learning. The authors also establish the convergence rate of the proposed algorithm. Extensive experiments demonstrate the efficacy of the proposed algorithm.  ", "main_review": "It is an interesting work on accelerating the training process of federated learning.  The reviewer still has several concerns about the current submission:\n(1) It seems that the main benefit of progressive training for federated learning comes from the early stage during training.  Hence, we recommend the authors do more ablation studies on the hyperparameter $S$. \n(2) For different stages,  how do the authors set the hyperparameters, such as learning rate.  Since in different stages, the depth of subnetworks is differents,  the hyperparameter tuning techniques may highly influence the training speed and final performance.\n(3) Theorem 1 provides the convergence of the proposed ProgFed.  However, the convergence results do not indicate the benefit of progressive training for federated learning. \n(4)  The experiments should be strengthened.  The main optimizer in ProgFed is fedavg.  The authors should incorporate the proposed ProgFed with several other types of local opitmizers, such as fedprox, scaffold, etc., to further show progressive training can be a versatile and strong technique for accelerating federated learning. ", "summary_of_the_review": "see the comment above.", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635661630603}, {"id": "0F51JG2wUKO", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper473/Reviewer_5a4S"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "8: accept, good paper", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The issue that this work attempts to solve is that of computation and communication efficiency in federated learning. This is realised via ProgFed, a method that progressively trains neural network blocks of layers in $S$ “stages”. For the first $S - 1$ stages only a subset of layers is selected to be optimised, with the help of an optional (depending on the architecture) additional head to provide the training signal (which is discarded after that particular stage is finished). At the last stage, the authors propose to fine-tune the entire architecture in an end-to-end manner. Due to the specific nature of this form of training, the first $S-1$ stages of training require less compute (due to not needing to perform the forward / backward passes on the layers not selected) and also less communication (as the clients and server only need to communicate the specific set of layers that is optimised). The authors prove that such a training procedure converges in a rate similar to the one of standard end-to-end training and demonstrate in extensive experiments the benefits of ProgFed. As an extra bonus, the authors also show that ProgFed is more amenable to compression / quantization of the federated messages, thus showing that the benefits of these, more traditional, approaches are additive to the ones of ProgFed.", "review_text": "This submission is good; the paper is well-written, the idea is simple and seems to work well in practice. Furthermore, the experimental evaluation is thorough and considers multiple datasets and architectures. There is also a nice bonus that the proposed method is orthogonal to more traditional approaches that target communication cost reduction. For this reason, I am leaning towards recommending acceptance.\n\nHaving said that, there are some things that I would like some input from the authors:\n- It seems that currently you throw away the auxiliary heads obtained during the progressive training phase. This seems like a waste; why not reuse these heads for inference? For example, one could argue about some “voting” between the heads or even “early exiting” of datapoints if an early head is confident enough about the prediction (this would reduce the computational complexity at inference time as well).  Does the performance of each head become worse after the end-to-end fine-tuning?\n- While the provided proof is welcome, unless I am missing something, it seems to be in the centralised setting. Can it be extended to the federated setting? As the main motivation of this work is the federated setting, it makes more sense to prove convergence in such scenarios. \n- In the experiments you report the mean after three experiment runs. Why not show the standard deviation / error as well? This will allow one to better judge the significance of the results. Currently it is a bit unclear.\n- In Figure 2 it is weird that the end-to-end training has a bit of instability that is not present in progressive training. Could you elaborate on why this is the case? Furthermore, at Figure 5 ProgFed seems to have quite some aggressive dips in performance (I guess when switching stages). This is especially pronounced at CIFAR 10. Could the authors expand upon this and whether such instabilities can be hindering in practice?\n- Do you have any indication or intuition as to why progressive training helps with quantization / compression?\n- For section 4.5; does the “Layerwise” strategy use only one head, i.e., the one at the end? If yes, why not use a separate head (i.e, such as ProgFed)? How does Layerwise + end-to-end fine-tuning perform compared to ProgFed? Furthermore, the difference between ProgFed and “Random is small”, so some standard error indication would help.  Finally, the “Random” strategy is missing from Figure 8 and it would be great if you could elaborate on what is the extra memory that “Random” needs.\n\nAs for other minor things:\n- The plots such as the one at Figure 3 are kind of confusing and not very clear. I would suggest that the authors convert them to bar plots (such as the ones they have in the appendix) instead. \n- There is a reference that is listed twice (the “Sparsified SGD with memory”)\n- “Layerwise” referred to as “Laywise”\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The issue that this work attempts to solve is that of computation and communication efficiency in federated learning. This is realised via ProgFed, a method that progressively trains neural network blocks of layers in $S$ “stages”. For the first $S - 1$ stages only a subset of layers is selected to be optimised, with the help of an optional (depending on the architecture) additional head to provide the training signal (which is discarded after that particular stage is finished). At the last stage, the authors propose to fine-tune the entire architecture in an end-to-end manner. Due to the specific nature of this form of training, the first $S-1$ stages of training require less compute (due to not needing to perform the forward / backward passes on the layers not selected) and also less communication (as the clients and server only need to communicate the specific set of layers that is optimised). The authors prove that such a training procedure converges in a rate similar to the one of standard end-to-end training and demonstrate in extensive experiments the benefits of ProgFed. As an extra bonus, the authors also show that ProgFed is more amenable to compression / quantization of the federated messages, thus showing that the benefits of these, more traditional, approaches are additive to the ones of ProgFed.", "main_review": "This submission is good; the paper is well-written, the idea is simple and seems to work well in practice. Furthermore, the experimental evaluation is thorough and considers multiple datasets and architectures. There is also a nice bonus that the proposed method is orthogonal to more traditional approaches that target communication cost reduction. For this reason, I am leaning towards recommending acceptance.\n\nHaving said that, there are some things that I would like some input from the authors:\n- It seems that currently you throw away the auxiliary heads obtained during the progressive training phase. This seems like a waste; why not reuse these heads for inference? For example, one could argue about some “voting” between the heads or even “early exiting” of datapoints if an early head is confident enough about the prediction (this would reduce the computational complexity at inference time as well).  Does the performance of each head become worse after the end-to-end fine-tuning?\n- While the provided proof is welcome, unless I am missing something, it seems to be in the centralised setting. Can it be extended to the federated setting? As the main motivation of this work is the federated setting, it makes more sense to prove convergence in such scenarios. \n- In the experiments you report the mean after three experiment runs. Why not show the standard deviation / error as well? This will allow one to better judge the significance of the results. Currently it is a bit unclear.\n- In Figure 2 it is weird that the end-to-end training has a bit of instability that is not present in progressive training. Could you elaborate on why this is the case? Furthermore, at Figure 5 ProgFed seems to have quite some aggressive dips in performance (I guess when switching stages). This is especially pronounced at CIFAR 10. Could the authors expand upon this and whether such instabilities can be hindering in practice?\n- Do you have any indication or intuition as to why progressive training helps with quantization / compression?\n- For section 4.5; does the “Layerwise” strategy use only one head, i.e., the one at the end? If yes, why not use a separate head (i.e, such as ProgFed)? How does Layerwise + end-to-end fine-tuning perform compared to ProgFed? Furthermore, the difference between ProgFed and “Random is small”, so some standard error indication would help.  Finally, the “Random” strategy is missing from Figure 8 and it would be great if you could elaborate on what is the extra memory that “Random” needs.\n\nAs for other minor things:\n- The plots such as the one at Figure 3 are kind of confusing and not very clear. I would suggest that the authors convert them to bar plots (such as the ones they have in the appendix) instead. \n- There is a reference that is listed twice (the “Sparsified SGD with memory”)\n- “Layerwise” referred to as “Laywise”\n", "summary_of_the_review": "Good submission with a simple idea that seems to work well in practice. Extensive evaluation and nice compatibility with more traditional compression and quantization. There are some points that I would like some input from the authors which, if turned out positive, would strengthen this work further. For this reason, I am leaning towards acceptance.\n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635514791672}, {"id": "68Wvf8_G6wA", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper473/Reviewer_3zVc"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper introduces \"ProgFed\", which is an algorithm designed for federated learning scenarios. \n\nProgFed applies the idea of progressive training. Instead of training the neural network from end to end, it first trains the shallow layers, and then gradually train the deep layers, hoping that: 1) training the shallow layers first will learn some simple and meaningful features which will help training the deep layers, 2) and thus reduces the communication cost for the federated learning applications. Besides, the ProgFed algorithm also applies the local steps (Local-SGD or FedAvg) in order to decrease the communication cost during the model training. The ProgFed algorithm works not only for the feed-forward neural nets but also the U-nets, which consist of an encoder and a decoder.\n\nFor the theoretical part, the authors show that ProgFed has similar convergence results compared with SGD.\n\nFor the empirical (experimental) part, the authors use experiments trying to show the superiority of ProgFed over other existed federated learning methods, e.g. the trivial end-to-end training procedure, the communication compression methods (e.g. quantization)", "review_text": "Please see \"Summary Of The Paper\" for the overview of the paper.\n\nThe paper introduces \"ProgFed\", which is an algorithm designed for federated learning scenarios. The authors use theoretical and empirical results trying the show the superiority of ProgFed over other federated learning methods when applying to neural networks. However, I think that there are several weaknesses in this paper.\n\n1. The first weakness (major weakness) comes from the  \"algorithm\" part. If I understand correctly, the novelty of ProgFed is applying the progressive training technique into the federated learning setting. However, progressive training is an existed method trying to speed up the neural network training in the single machine setting. In the single machine case, it is easy to apply progressive training techniques with different optimizers, e.g. RMSProp, ADAM, SGD, or other optimization algorithms. In the federated learning setting, it is also natural to apply progressive training with other federated optimization algorithms, e.g. local methods like FedAvg (Local-SGD), SCAFFOLD, FedPAGE, or compression schemes like quantization, Sign-SGD, etc. For ProgFed,  if I understand correctly, it just applies the progressive training technique with the Local-SGD algorithm. Thus the first weakness comes from the algorithmic novelty perspective.\n\n2. The second weakness (major weakness) comes from the model and the theoretical analysis. The data heterogeneity is the most important aspect that differentiates federated optimization from distributed optimization. If the data on different clients are not iid, then the gradients on different clients are not the same. In ProgFed algorithm (Algorithm 1) Line 16, the algorithm uses local update $x_{c,j+1}^s = x_{c,j}^s - \\eta g_c^s(x_{c,j}^s)$ for stage $s$, client $c$, local step $j$. Here, $\\nabla f_c(\\cdot)$ should be different in the federated learning setting. However, in the assumption part and the proofs, I do not see any assumptions to control the data heterogeneity (like $(G,B)$-BGD). Based on the fact that FedAvg and Local-SGD need assumptions like $(G,B)$-BGD to guarantee the convergence, I believe that either the model does not consider data heterogeneity (which is a main drawback for the federated learning applications), or the proof needs to modify in order to support data heterogeneity.\n\n3. The third weakness (minor weakness) comes from the empirical part. For the experiments, the authors compare ProgFed with very straightforward optimization (end-to-end) and several compression schemes. However, as I stated previously, ProgFed actually consists of two modules, progressive training and Local-SGD, which behave at a different level. Directly comparing ProgFed with end-to-end training does not show the effectiveness of ProgFed, I suggest the authors add more experiments, such as ProgFed vs Progressive training + simple SGD, to demonstrate the effectiveness of local-steps with the existence of progressive training, or ProgFed vs Progressive training + quantization, to demonstrate whether the progressive training combines better with local steps methods or compression methods, or ProgFed VS Local-SGD, to show the power of progressive training. Besides, I think the data split for different datasets can be moved to the main content.\n\nPlease point out if my understanding is wrong, and I will consider raising my score if my concerns and questions are addressed or answered. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper introduces \"ProgFed\", which is an algorithm designed for federated learning scenarios. \n\nProgFed applies the idea of progressive training. Instead of training the neural network from end to end, it first trains the shallow layers, and then gradually train the deep layers, hoping that: 1) training the shallow layers first will learn some simple and meaningful features which will help training the deep layers, 2) and thus reduces the communication cost for the federated learning applications. Besides, the ProgFed algorithm also applies the local steps (Local-SGD or FedAvg) in order to decrease the communication cost during the model training. The ProgFed algorithm works not only for the feed-forward neural nets but also the U-nets, which consist of an encoder and a decoder.\n\nFor the theoretical part, the authors show that ProgFed has similar convergence results compared with SGD.\n\nFor the empirical (experimental) part, the authors use experiments trying to show the superiority of ProgFed over other existed federated learning methods, e.g. the trivial end-to-end training procedure, the communication compression methods (e.g. quantization)", "main_review": "Please see \"Summary Of The Paper\" for the overview of the paper.\n\nThe paper introduces \"ProgFed\", which is an algorithm designed for federated learning scenarios. The authors use theoretical and empirical results trying the show the superiority of ProgFed over other federated learning methods when applying to neural networks. However, I think that there are several weaknesses in this paper.\n\n1. The first weakness (major weakness) comes from the  \"algorithm\" part. If I understand correctly, the novelty of ProgFed is applying the progressive training technique into the federated learning setting. However, progressive training is an existed method trying to speed up the neural network training in the single machine setting. In the single machine case, it is easy to apply progressive training techniques with different optimizers, e.g. RMSProp, ADAM, SGD, or other optimization algorithms. In the federated learning setting, it is also natural to apply progressive training with other federated optimization algorithms, e.g. local methods like FedAvg (Local-SGD), SCAFFOLD, FedPAGE, or compression schemes like quantization, Sign-SGD, etc. For ProgFed,  if I understand correctly, it just applies the progressive training technique with the Local-SGD algorithm. Thus the first weakness comes from the algorithmic novelty perspective.\n\n2. The second weakness (major weakness) comes from the model and the theoretical analysis. The data heterogeneity is the most important aspect that differentiates federated optimization from distributed optimization. If the data on different clients are not iid, then the gradients on different clients are not the same. In ProgFed algorithm (Algorithm 1) Line 16, the algorithm uses local update $x_{c,j+1}^s = x_{c,j}^s - \\eta g_c^s(x_{c,j}^s)$ for stage $s$, client $c$, local step $j$. Here, $\\nabla f_c(\\cdot)$ should be different in the federated learning setting. However, in the assumption part and the proofs, I do not see any assumptions to control the data heterogeneity (like $(G,B)$-BGD). Based on the fact that FedAvg and Local-SGD need assumptions like $(G,B)$-BGD to guarantee the convergence, I believe that either the model does not consider data heterogeneity (which is a main drawback for the federated learning applications), or the proof needs to modify in order to support data heterogeneity.\n\n3. The third weakness (minor weakness) comes from the empirical part. For the experiments, the authors compare ProgFed with very straightforward optimization (end-to-end) and several compression schemes. However, as I stated previously, ProgFed actually consists of two modules, progressive training and Local-SGD, which behave at a different level. Directly comparing ProgFed with end-to-end training does not show the effectiveness of ProgFed, I suggest the authors add more experiments, such as ProgFed vs Progressive training + simple SGD, to demonstrate the effectiveness of local-steps with the existence of progressive training, or ProgFed vs Progressive training + quantization, to demonstrate whether the progressive training combines better with local steps methods or compression methods, or ProgFed VS Local-SGD, to show the power of progressive training. Besides, I think the data split for different datasets can be moved to the main content.\n\nPlease point out if my understanding is wrong, and I will consider raising my score if my concerns and questions are addressed or answered. ", "summary_of_the_review": "In summary, I have the following concerns:\n\n1. The proposed ProgFed algorithm seems a little bit incremental, since it seems like an \"A+B\" method where A is progressive training technique and B is FedAvg (local-SGD).\n2. For the model or the theoretical analysis, either the model does not consider data heterogeneity or the theoretical analysis needs to be modified.\n3. (minor concern) ProgFed is a combination of two techniques, and comparing with compression schemes like quantization is not very fair.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1635372553750}], "openreview_url": "https://openreview.net/forum?id=Gpp1dfvZYYH", "arxiv_id": "2110.05323", "paper_pdf": "papers/Gpp1dfvZYYH.pdf", "paper_pdf_sha256": "7e84b451cfafcd539ff5178c3ca052c1375b64b2460a244981715a8ea3659ab3", "paper_pdf_bytes": 3503773, "paper_pdf_source": "openreview", "code_url": "https://github.com/hui-po-wang/ProgFed", "code_repository": "hui-po-wang/ProgFed", "code_commit": "1d1eb9d83110bc31ad38447288325560fa1d9878", "code_archive": "repos/Gpp1dfvZYYH.zip", "code_archive_sha256": "eeb75970647d74dc71642004cf5ffc37edccc6670f1157cdc5073530d3665b2d", "code_archive_bytes": 145342, "code_file_count": 17, "code_extensions": {".py": 14, ".sh": 3}, "github_disk_usage_kb": 7744, "github_languages": {"Python": 96470, "Shell": 3430}, "github_archived": false, "github_pushed_at": "2022-10-17T15:05:45Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/progfed-effective-communication-and-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "uELnyih9gqb", "year": 2021, "status": "rejected", "title": "WAVEQ: GRADIENT-BASED DEEP QUANTIZATION OF NEURAL NETWORKS THROUGH SINUSOIDAL REGULARIZATION", "authors": ["Ahmed T. Elthakeb", "Prannoy Pilligundla", "Tarek Elgindi", "Fatemehsadat Mireshghallah", "Charles-Alban Deledalle", "Hadi Esmaeilzadeh"], "authorids": ["~Ahmed_T._Elthakeb1", "~Prannoy_Pilligundla1", "telgindi@ucsd.edu", "~Fatemehsadat_Mireshghallah1", "~Charles-Alban_Deledalle2", "~Hadi_Esmaeilzadeh1"], "authors_source": "OpenReview API", "abstract": "Deep quantization of neural networks below eight bits can lead to superlinear benefits in storage and compute efficiency. However, homogeneously quantizing all the layers to the same level does not account for the distinction of the layers and their individual properties. Heterogenous assignment of bitwidths to individual layers is attractive but opens an exponentially large non-contiguous hyperparameter space (${Available Bitwidths}^{\\# Layers}$). As such finding the bitwidth while also quantizing the network to those levels becomes a major challenge. This paper addresses this challenge through a sinusoidal regularization mechanism, dubbed WaveQ. Adding our parametrized sinusoidal regularizer enables us to not only find the quantized weights but also learn the bitwidth of the layers by making the period of the sinusoidal regularizer a trainable parameter. In addition, the sinusoidal regularizer itself is designed to align its minima on the quantization levels. With these two innovations, during training, stochastic gradient descent uses the form of the sinusoidal regularizer and its minima to push the weights to the quantization levels while it is also learning the period which will determine the bitwidth of each layer separately. As such WaveQ is a gradient-based mechanism that jointly learns the quantized weights as well as the heterogeneous bitwidths. We show how WaveQ balance compute efficiency and accuracy, and provide a heterogeneous bitwidth assignment for quantization of a large variety of deep networks (AlexNet, CIFAR-10, MobileNet, ResNet-18, ResNet-20, SVHN, and VGG-11) that virtually preserves the accuracy. WaveQ is versatile and can also be used with predetermined bitwidths by fixing the period of the sinusoidal regularizer. In this case. WaveQ enhances quantized training algorithms (DoReFa and WRPN) with about 4.8% accuracy improvements on average, and outperforms multiple state-of-the-art techniques. Finally, WaveQ applied to quantizing transformers\n", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "5j07Sf3fUeB", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2296/AnonReviewer3"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "When training quantized neural networks, one typically first fixes the desired bitwidth $b$ (of weights and activations). While training, one maintains and updates full-bitwidth weights during backprop and \"cheats\" by using $b$-quantized versions of these full-precision weights during forward propagation. The quantization scheme used may vary.\n\nWhat if we wished instead to *learn* $b$? This is especially attractive, for instance, if we wish a distinct $b_i$ for each layer $i$. We could introduce a variable $\\beta_i$ to represent $b_i$ during training. We would probably add a regularizer to minimize $\\Sigma \\beta_i$ to encourage training toward small bitwidths. However, this still leaves the problem that the $\\beta_i$ would be real, not integer, as required for quantization. We could cheat here by using $\\lceil \\beta_i \\rceil$, but note that quantization error $\\beta_i - \\lceil \\beta_i \\rceil$ may be quite large relative to the small integer values $\\beta$ is expected to take. Is there a way to ensure that training automatically produces $\\beta$s that are close to integer values?\n\nThis is where the current paper introduces a neat trick. By embedding $\\beta$ inside a (suitably scaled) sinusoidal loss function, they ensure that the $\\beta$s tend to have integer values (that correspond to the minima of the sinusoid).\n\nI haven't seen this trick before, and it seems like an elegant approach to the problem of favoring integer values within the global optimization process. The goal of producing distinct bitwidths for each layer is definitely a valid one, so this technique does also address a practical problem.\n\nThe technique's practical impact seems limited. In particular, it seems that the main effect is to reduce the average bitwidth of SOTA quantized models from 4 to >= 3.6 at roughly the same accuracy. This should correspond roughly to a 10% gain in performance (runtime), which is quite good but not great. I have no idea how to to interpret the power numbers, so I will ignore those for now.\n\nOne question I have though is how the technique compares to a decent heuristic variable-bitwidth baseline. The authors do compare to a  \"decrement the bitiwidth of a single layer\" baseline, which is a good start. However, what if you do something more informed than that as baseline? Can you tell us what bitwidths your techniques chose for Resnet, for instance? Is there a simple pattern there (e.g. high bitwidth for early layers, and lower later on)? Based on this, what is the \"best\" non-learned bitwidth selection you can come up with?\n\nIn any case, the sinusoid trick for learning integer-like values is a good one for people to know, and it seems that the authors have shown that it can be implemented to have practical impact. Even if bitwidth reduction is currently modest, I can imagine follow-up work to increase it. So I support acceptance of the paper.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "How to include bitwidths in your training objective", "review": "When training quantized neural networks, one typically first fixes the desired bitwidth $b$ (of weights and activations). While training, one maintains and updates full-bitwidth weights during backprop and \"cheats\" by using $b$-quantized versions of these full-precision weights during forward propagation. The quantization scheme used may vary.\n\nWhat if we wished instead to *learn* $b$? This is especially attractive, for instance, if we wish a distinct $b_i$ for each layer $i$. We could introduce a variable $\\beta_i$ to represent $b_i$ during training. We would probably add a regularizer to minimize $\\Sigma \\beta_i$ to encourage training toward small bitwidths. However, this still leaves the problem that the $\\beta_i$ would be real, not integer, as required for quantization. We could cheat here by using $\\lceil \\beta_i \\rceil$, but note that quantization error $\\beta_i - \\lceil \\beta_i \\rceil$ may be quite large relative to the small integer values $\\beta$ is expected to take. Is there a way to ensure that training automatically produces $\\beta$s that are close to integer values?\n\nThis is where the current paper introduces a neat trick. By embedding $\\beta$ inside a (suitably scaled) sinusoidal loss function, they ensure that the $\\beta$s tend to have integer values (that correspond to the minima of the sinusoid).\n\nI haven't seen this trick before, and it seems like an elegant approach to the problem of favoring integer values within the global optimization process. The goal of producing distinct bitwidths for each layer is definitely a valid one, so this technique does also address a practical problem.\n\nThe technique's practical impact seems limited. In particular, it seems that the main effect is to reduce the average bitwidth of SOTA quantized models from 4 to >= 3.6 at roughly the same accuracy. This should correspond roughly to a 10% gain in performance (runtime), which is quite good but not great. I have no idea how to to interpret the power numbers, so I will ignore those for now.\n\nOne question I have though is how the technique compares to a decent heuristic variable-bitwidth baseline. The authors do compare to a  \"decrement the bitiwidth of a single layer\" baseline, which is a good start. However, what if you do something more informed than that as baseline? Can you tell us what bitwidths your techniques chose for Resnet, for instance? Is there a simple pattern there (e.g. high bitwidth for early layers, and lower later on)? Based on this, what is the \"best\" non-learned bitwidth selection you can come up with?\n\nIn any case, the sinusoid trick for learning integer-like values is a good one for people to know, and it seems that the authors have shown that it can be implemented to have practical impact. Even if bitwidth reduction is currently modest, I can imagine follow-up work to increase it. So I support acceptance of the paper.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603963592805}, {"id": "gnKffpHu_rY", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2296/AnonReviewer1"], "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposed a regularization term to control the bit-width and encourage the DNN weights moving to the quantization intervals. The key in such regularization is the Sinusoidal function, where the penalty is maximized in the middle of quantization levels and minimized at the quantization points. The sinusoidal period is regarded as the continuous representation of the bit-width.\n\nPros:\n1. The paper is well written and easy to understand. Illustration figures help a lot to present the proposed method. E.g., Figure 2 clearly shows the key idea of this paper.\n\n2. The idea of using the sinusoidal period as a continuous representation is novel and it makes it natural to use gradient descent methods to optimize the bit-width.\n\n3. Experiments on different datasets and tasks show that the proposed method can bring improvement on accuracy by using better regularization and / or better bit-width allocation.\n\nCons:\n1. The quantization regularization can only be used on weight quantization, while activation quantization is very important and sometimes more sensitive to the accuracy drop. On this point, the proposed method cannot be used to allocate the bit-width of activation, while activations can actually have different bit-width to the weights.\n\n2. Some related works are missing.\nFor quantization regularization, it is used in ProxQuant[1] as well, where the weights are regularized with a norm-based quantization regularization. It also discussed the similar observation that the regularization can help quantization-aware training.\nFor bit-width allocation, [2] formulated the bit-width allocation as a knapsack problem and used ADMM to iteratively optimize the NN compression. [3] used Bayesian optimization to allocate the bit-width for different layers.\n\n3. The procedure to tune the regularization strengths is a little complicated. Although the authors proposed a detailed procedure for tuning these hyper-parameters, it's questionable that the same procedure can be widely applied in different networks/optimizers/datasets. For $\\lambda_\\beta$, it actually determines the number of total bits for the entire network, so it should depends on the expected compression rate of the quantized DNN.\n\n4. Using the sinusoidal period to decide the bit-width has an assumption that the range of the weights is fixed: otherwise one can reduce both the period and value range to keep the bit-width unchanged. However, the range of weights is usually not fixed in the training.\n\n5. The experiments on bit-width allocation may be not enough to show the effectiveness of the proposed method as a bit-width allocation method. It's better to compare with some other bit-width allocation methods.\n\nIn general, my rating is borderline. I hope the authors can give some response to the cons listed above.\n\nReference:\n\n[1] Bai, Y., Wang, Y.X. and Liberty, E., 2018. Proxquant: Quantized neural networks via proximal operators. arXiv preprint arXiv:1810.00861.\n\n[2] Yang, H., Gui, S., Zhu, Y. and Liu, J., 2020. Automatic Neural Network Compression by Sparsity-Quantization Joint Learning: A Constrained Optimization-Based Approach. In CVPR 2020.\n\n[3] Tung, F. and Mori, G., 2018. Clip-q: Deep network compression learning by in-parallel pruning-quantization. In CVPR 2018.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Borderline", "review": "This paper proposed a regularization term to control the bit-width and encourage the DNN weights moving to the quantization intervals. The key in such regularization is the Sinusoidal function, where the penalty is maximized in the middle of quantization levels and minimized at the quantization points. The sinusoidal period is regarded as the continuous representation of the bit-width.\n\nPros:\n1. The paper is well written and easy to understand. Illustration figures help a lot to present the proposed method. E.g., Figure 2 clearly shows the key idea of this paper.\n\n2. The idea of using the sinusoidal period as a continuous representation is novel and it makes it natural to use gradient descent methods to optimize the bit-width.\n\n3. Experiments on different datasets and tasks show that the proposed method can bring improvement on accuracy by using better regularization and / or better bit-width allocation.\n\nCons:\n1. The quantization regularization can only be used on weight quantization, while activation quantization is very important and sometimes more sensitive to the accuracy drop. On this point, the proposed method cannot be used to allocate the bit-width of activation, while activations can actually have different bit-width to the weights.\n\n2. Some related works are missing.\nFor quantization regularization, it is used in ProxQuant[1] as well, where the weights are regularized with a norm-based quantization regularization. It also discussed the similar observation that the regularization can help quantization-aware training.\nFor bit-width allocation, [2] formulated the bit-width allocation as a knapsack problem and used ADMM to iteratively optimize the NN compression. [3] used Bayesian optimization to allocate the bit-width for different layers.\n\n3. The procedure to tune the regularization strengths is a little complicated. Although the authors proposed a detailed procedure for tuning these hyper-parameters, it's questionable that the same procedure can be widely applied in different networks/optimizers/datasets. For $\\lambda_\\beta$, it actually determines the number of total bits for the entire network, so it should depends on the expected compression rate of the quantized DNN.\n\n4. Using the sinusoidal period to decide the bit-width has an assumption that the range of the weights is fixed: otherwise one can reduce both the period and value range to keep the bit-width unchanged. However, the range of weights is usually not fixed in the training.\n\n5. The experiments on bit-width allocation may be not enough to show the effectiveness of the proposed method as a bit-width allocation method. It's better to compare with some other bit-width allocation methods.\n\nIn general, my rating is borderline. I hope the authors can give some response to the cons listed above.\n\nReference:\n\n[1] Bai, Y., Wang, Y.X. and Liberty, E., 2018. Proxquant: Quantized neural networks via proximal operators. arXiv preprint arXiv:1810.00861.\n\n[2] Yang, H., Gui, S., Zhu, Y. and Liu, J., 2020. Automatic Neural Network Compression by Sparsity-Quantization Joint Learning: A Constrained Optimization-Based Approach. In CVPR 2020.\n\n[3] Tung, F. and Mori, G., 2018. Clip-q: Deep network compression learning by in-parallel pruning-quantization. In CVPR 2018.", "rating": "5: Marginally below acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603866136858}, {"id": "P5FD1Wdg5e", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2296/AnonReviewer4"], "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes using a sinusoidal regularizer for neural network quantization. The regularizer “WaveQ” (sin^2) pushes floating-point parameters towards quantized values. Because the period of the function is highly related to the required bit-width, it can be used to determine the bit-width while keeping good characteristics - continuous and trainable. The authors provide experiments on both CNN and Transformers. The proposed method is widely adaptable and easy-to-use with quite promising results.\n\nMajor questions/suggestions:\n\n1.\tIn equation 2.1, the period (=quantization step) is s= 1/ (2^beta – 1), so the regularizer tries to move ‘w’ towards the multiple of ‘s’. As I understand, the desired range of ‘w’ is (-1, 1), to properly match the actual quantized value. But there is no clipping nor normalization applied to ‘w’. Is there something I missed?\n\n2.\tThere is a missing work [Nyuyen, 2020] which suggests a similar regularizer (|cos|). However, this paper is still valuable and does not hurt novelty.\n[Ngugen, 2020] Quantization Aware Training with Absolute-Cosine Regularization for Automatic Speech Recognition\n\n3.\tIt seems that WaveQ is also used for activation quantization (Section 4.1.), but there are not enough details about applying WaveQ for activations.\n\nMinor issues:\n\n1.\tEquation B.2 is frequently used in Section 2. Also, the ‘alpha’ term needs more explanation when it first appears. (‘scaling factor’ is not enough) Consider moving the equation B.1, B.2. to the front.\n\n2.\tMany recent quantization papers use w / max(|w|) to normalize weight. Is there any advantage to use tanh(w) / max(tanh(w)) in Equation B.1?\n\n3.\tTypo: ‘bitwidh’ in Sec.4.1, ‘citep’ in Sec.5, ‘ofcitepp’ in Sec.6, \n\n4.\tI believe the theoretic part is OK but am not sure about the exactness and its importance.\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Good paper, suggests a useful method for quantization-aware training", "review": "The paper proposes using a sinusoidal regularizer for neural network quantization. The regularizer “WaveQ” (sin^2) pushes floating-point parameters towards quantized values. Because the period of the function is highly related to the required bit-width, it can be used to determine the bit-width while keeping good characteristics - continuous and trainable. The authors provide experiments on both CNN and Transformers. The proposed method is widely adaptable and easy-to-use with quite promising results.\n\nMajor questions/suggestions:\n\n1.\tIn equation 2.1, the period (=quantization step) is s= 1/ (2^beta – 1), so the regularizer tries to move ‘w’ towards the multiple of ‘s’. As I understand, the desired range of ‘w’ is (-1, 1), to properly match the actual quantized value. But there is no clipping nor normalization applied to ‘w’. Is there something I missed?\n\n2.\tThere is a missing work [Nyuyen, 2020] which suggests a similar regularizer (|cos|). However, this paper is still valuable and does not hurt novelty.\n[Ngugen, 2020] Quantization Aware Training with Absolute-Cosine Regularization for Automatic Speech Recognition\n\n3.\tIt seems that WaveQ is also used for activation quantization (Section 4.1.), but there are not enough details about applying WaveQ for activations.\n\nMinor issues:\n\n1.\tEquation B.2 is frequently used in Section 2. Also, the ‘alpha’ term needs more explanation when it first appears. (‘scaling factor’ is not enough) Consider moving the equation B.1, B.2. to the front.\n\n2.\tMany recent quantization papers use w / max(|w|) to normalize weight. Is there any advantage to use tanh(w) / max(tanh(w)) in Equation B.1?\n\n3.\tTypo: ‘bitwidh’ in Sec.4.1, ‘citep’ in Sec.5, ‘ofcitepp’ in Sec.6, \n\n4.\tI believe the theoretic part is OK but am not sure about the exactness and its importance.\n", "rating": "7: Good paper, accept", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603731967441}, {"id": "DnXIqb4p1S2", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper2296/AnonReviewer2"], "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper proposes WaveQ which proposes a sinusoidal regularization approach for quantizing Neural Networks. The motivation is to enable mixed-precision quantization, since homogeneously quantizing the model to low precision can lead to accuracy degradation.\n\nThe main problem with heterogeneous/mixed-precision bit-width is that its search space is exponentially large. To address this the authors propose to automatically learn this bit-precision by adding a sinusoidal regularization where the period learn able.\n\nWhile the approach is interesting but the theoretical and empirical results are not satisfactory and as such it is not clear how the proposed method is superior to other methods proposed in the literature. In particular:\n\n\n- The theoretical analysis provided does not apply to the proposed method. This is because the theory requires that the regularization function vanish after many iterations, whereas in the experiments the opposite approach is used.\n\n- The accuracy results provided in the empirical section incur significant degradation as compared to the baseline. Furthermore, the comparison is performed with old quantization methods. Newer mixed-precision quantization results using for example HAQ leads to much better quantization. As such, it is not clear what is the advantage of the proposed method?\n\n- How is this approach different than Achterhold et al., 2018?", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting Motivation but Unfortunately  Unsatisfactory Empirical and Theoretical Results", "review": "The paper proposes WaveQ which proposes a sinusoidal regularization approach for quantizing Neural Networks. The motivation is to enable mixed-precision quantization, since homogeneously quantizing the model to low precision can lead to accuracy degradation.\n\nThe main problem with heterogeneous/mixed-precision bit-width is that its search space is exponentially large. To address this the authors propose to automatically learn this bit-precision by adding a sinusoidal regularization where the period learn able.\n\nWhile the approach is interesting but the theoretical and empirical results are not satisfactory and as such it is not clear how the proposed method is superior to other methods proposed in the literature. In particular:\n\n\n- The theoretical analysis provided does not apply to the proposed method. This is because the theory requires that the regularization function vanish after many iterations, whereas in the experiments the opposite approach is used.\n\n- The accuracy results provided in the empirical section incur significant degradation as compared to the baseline. Furthermore, the comparison is performed with old quantization methods. Newer mixed-precision quantization results using for example HAQ leads to much better quantization. As such, it is not clear what is the advantage of the proposed method?\n\n- How is this approach different than Achterhold et al., 2018?", "rating": "4: Ok but not good enough - rejection", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, "tcdate": 1603603266918}], "openreview_url": "https://openreview.net/forum?id=uELnyih9gqb", "arxiv_id": null, "paper_pdf": "papers/uELnyih9gqb.pdf", "paper_pdf_sha256": "5a682c3fb2d6b864bfc43325e3e695ec542c92bc507c527d834559d6e44354a7", "paper_pdf_bytes": 6135804, "paper_pdf_source": "openreview", "code_url": "https://github.com/waveq-reg/waveq", "code_repository": "waveq-reg/waveq", "code_commit": "39bbf0dd95f5bf139197152e651acba08f5ccc68", "code_archive": "repos/uELnyih9gqb.zip", "code_archive_sha256": "e62168ffadd6946d47432e0d32637d2b2c34742fd8564bc4035a23797dc70b73", "code_archive_bytes": 10720076, "code_file_count": 344, "code_extensions": {".py": 318, ".sh": 10, ".cpp": 7, ".cu": 4, ".cuh": 2, ".lua": 2, ".h": 1}, "github_disk_usage_kb": 5593, "github_languages": {"Python": 9078893, "Cuda": 36414, "C++": 15854, "Shell": 4215, "Lua": 4210}, "github_archived": false, "github_pushed_at": "2020-06-11T19:08:04Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/waveq-gradient-based-deep-quantization-of"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "Bo5eKnJPML", "year": 2025, "status": "rejected", "title": "A Reasoning-Based Approach to Cryptic Crossword Clue Solving", "authors": ["Martin Andrews", "Sam Witteveen"], "authorids": ["~Martin_Andrews1", "~Sam_Witteveen1"], "authors_source": "OpenReview API", "abstract": "Cryptic crossword clues are challenging language tasks for which new test sets are released daily by major newspapers on a global basis. Each cryptic clue contains both the definition of the answer to be placed in the crossword grid (in common with regular crosswords), and ‘wordplay’ that *proves* that the answer is correct (i.e. a human solver can be confident that an answer is correct without needing crossing words as confirmation). This work describes an LLM-based reasoning system built from open-licensed components that solves cryptic clues by (i) hypothesising answers; (ii) proposing wordplay explanations; and (iii) using a verifier system that operates on codified reasoning steps. Overall, this system establishes a new state-of-the-art performance on the challenging Cryptonite dataset of clues from The Times and The Telegraph newspapers in the UK. Because each proved solution is expressed in Python, interpretable wordplay reasoning for proven answers is available for inspection", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "qCDw7Bhp07", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7903/Reviewer_HU2t"], "rating": 6, "soundness": 3, "presentation": 2, "contribution": 3, "confidence": 3, "summary": "The paper proposes a reasoning-based system for solving cryptic crossword clues using fine-tuned LLMs and a Python-based verification process. The model generates candidate answers and wordplay suggestions, formulates these into Python code, and verifies correctness via assertions.", "review_text": "The paper proposes a reasoning-based system for solving cryptic crossword clues using fine-tuned LLMs and a Python-based verification process. The model generates candidate answers and wordplay suggestions, formulates these into Python code, and verifies correctness via assertions.", "strengths": "**Originality**\nThe paper offers a unique approach to solving cryptic crossword clues. Like math proofs, the authors combined LLMs for candidate generation, wordplay suggestions, and formalized Python-based verifiers.\n\n**Quality**\nThe paper details the fine-tuning of models for clue and wordplay generation and the creation of a domain-specific verifier, including formalization in Python. The experiments demonstrate a clear improvement over previous state-of-the-art results on the Cryptonite dataset, significantly outperforming prior baselines.\n\n**Clarity**\nThe paper includes figures containing examples that help in understanding.\n\n**Significance**\nThe paper shows an improvement in the cryptic crossword domain.", "weaknesses": "1. The mention of a 23.5% accuracy on the Cryptonite dataset using GPT-4-turbo in the paper [1] suggests that a more rigorous baseline is possible for this study. Using the same prompting strategy on the Gemini-Flash that is used in the paper could provide better insight into the improvement of the proposed method.\n\n2. The paper's reliance on the Wordplay dataset, gathered from newspapers (The Times, etc.), raises the possibility of data leakage, as The Times also contributes to the Cryptonite test set. If fine-tuning data includes patterns or clues from the test set, results may be artificially inflated.\n\n\n[1] Saha, Soumadeep, et al. \"Language Models are Crossword Solvers.\" arXiv preprint arXiv:2406.09043 (2024).", "questions": "1. Partial correctness metrics, as reported in [1], can provide a more nuanced understanding of model success on cryptic crosswords, where the model might have answered some rows/columns but not the whole crossword.\n\n2. Some minor corrections on line 200 (\"from\" is repeated twice)\n\n[1] Saha, Soumadeep, et al. \"Language Models are Crossword Solvers.\" arXiv preprint arXiv:2406.09043 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a reasoning-based system for solving cryptic crossword clues using fine-tuned LLMs and a Python-based verification process. The model generates candidate answers and wordplay suggestions, formulates these into Python code, and verifies correctness via assertions.", "soundness": 3, "presentation": 2, "contribution": 3, "strengths": "**Originality**\nThe paper offers a unique approach to solving cryptic crossword clues. Like math proofs, the authors combined LLMs for candidate generation, wordplay suggestions, and formalized Python-based verifiers.\n\n**Quality**\nThe paper details the fine-tuning of models for clue and wordplay generation and the creation of a domain-specific verifier, including formalization in Python. The experiments demonstrate a clear improvement over previous state-of-the-art results on the Cryptonite dataset, significantly outperforming prior baselines.\n\n**Clarity**\nThe paper includes figures containing examples that help in understanding.\n\n**Significance**\nThe paper shows an improvement in the cryptic crossword domain.", "weaknesses": "1. The mention of a 23.5% accuracy on the Cryptonite dataset using GPT-4-turbo in the paper [1] suggests that a more rigorous baseline is possible for this study. Using the same prompting strategy on the Gemini-Flash that is used in the paper could provide better insight into the improvement of the proposed method.\n\n2. The paper's reliance on the Wordplay dataset, gathered from newspapers (The Times, etc.), raises the possibility of data leakage, as The Times also contributes to the Cryptonite test set. If fine-tuning data includes patterns or clues from the test set, results may be artificially inflated.\n\n\n[1] Saha, Soumadeep, et al. \"Language Models are Crossword Solvers.\" arXiv preprint arXiv:2406.09043 (2024).", "questions": "1. Partial correctness metrics, as reported in [1], can provide a more nuanced understanding of model success on cryptic crosswords, where the model might have answered some rows/columns but not the whole crossword.\n\n2. Some minor corrections on line 200 (\"from\" is repeated twice)\n\n[1] Saha, Soumadeep, et al. \"Language Models are Crossword Solvers.\" arXiv preprint arXiv:2406.09043 (2024).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730689440846}, {"id": "bqrVjWXE2B", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7903/Reviewer_5DQK"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 1, "confidence": 3, "summary": "The authors propose a method that achieves state of the art performance on cryptic crosswords. The method uses LLMs to solve the subtasks using various strategies. First, Gemma2-9B is fine-tuned using low-rank adapters on answer-candidate generation, which returns a list of answers which match the 'pattern' (the length of the answer). Second, Gemma2-9B is again fine-tuned using low-rank adapters, but this time for 'wordplay suggestion', which takes each answer candidate and generates multiple definition/wordplay pairs, which are meant to explain why the answer is correct. Third, prompting is used in combination with Gemini-Flash to formalise the wordplay into Python 'proofs'. The Python proofs are run and any errors are reported back to Gemini-Flash, which is given a chance to correct its proof (until the answer is valid, or the maximum of 2 rewrites is reached). In the experiments, the formalisation process is shown to be helpful in the general, although for the 'Quick' subset, using the most frequent answer candidate proves too strong a baseline.", "review_text": "The authors propose a method that achieves state of the art performance on cryptic crosswords. The method uses LLMs to solve the subtasks using various strategies. First, Gemma2-9B is fine-tuned using low-rank adapters on answer-candidate generation, which returns a list of answers which match the 'pattern' (the length of the answer). Second, Gemma2-9B is again fine-tuned using low-rank adapters, but this time for 'wordplay suggestion', which takes each answer candidate and generates multiple definition/wordplay pairs, which are meant to explain why the answer is correct. Third, prompting is used in combination with Gemini-Flash to formalise the wordplay into Python 'proofs'. The Python proofs are run and any errors are reported back to Gemini-Flash, which is given a chance to correct its proof (until the answer is valid, or the maximum of 2 rewrites is reached). In the experiments, the formalisation process is shown to be helpful in the general, although for the 'Quick' subset, using the most frequent answer candidate proves too strong a baseline.", "strengths": "- State of the art result on complex task.\n- The reported results show that LLMs are very helpful as components in a system to solve complex tasks, while also not sufficient to solve it completely on their own.", "weaknesses": "- I think more space could be devoted to explaining cryptic crosswords. As someone previously unfamiliar with them, it took quite long for me to understand what exactly was happening in the example given in the introduction. Maybe a table or diagram that explains exactly how each part in the wordplay is related to the parts in the clue?\n- A solid motivation for why this work is important is missing. What is the primary reason to be interested in LLM performance on this task? And relatedly, what should be the general take-away after reading it?", "questions": "How were the in-context learning prompts constructed, and how many versions did you try? What is your sense of the sensitivity of the final accuracy w.r.t. the prompt, how important is it?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors propose a method that achieves state of the art performance on cryptic crosswords. The method uses LLMs to solve the subtasks using various strategies. First, Gemma2-9B is fine-tuned using low-rank adapters on answer-candidate generation, which returns a list of answers which match the 'pattern' (the length of the answer). Second, Gemma2-9B is again fine-tuned using low-rank adapters, but this time for 'wordplay suggestion', which takes each answer candidate and generates multiple definition/wordplay pairs, which are meant to explain why the answer is correct. Third, prompting is used in combination with Gemini-Flash to formalise the wordplay into Python 'proofs'. The Python proofs are run and any errors are reported back to Gemini-Flash, which is given a chance to correct its proof (until the answer is valid, or the maximum of 2 rewrites is reached). In the experiments, the formalisation process is shown to be helpful in the general, although for the 'Quick' subset, using the most frequent answer candidate proves too strong a baseline.", "soundness": 3, "presentation": 3, "contribution": 1, "strengths": "- State of the art result on complex task.\n- The reported results show that LLMs are very helpful as components in a system to solve complex tasks, while also not sufficient to solve it completely on their own.", "weaknesses": "- I think more space could be devoted to explaining cryptic crosswords. As someone previously unfamiliar with them, it took quite long for me to understand what exactly was happening in the example given in the introduction. Maybe a table or diagram that explains exactly how each part in the wordplay is related to the parts in the clue?\n- A solid motivation for why this work is important is missing. What is the primary reason to be interested in LLM performance on this task? And relatedly, what should be the general take-away after reading it?", "questions": "How were the in-context learning prompts constructed, and how many versions did you try? What is your sense of the sensitivity of the final accuracy w.r.t. the prompt, how important is it?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 3, "code_of_conduct": "Yes"}, "tcdate": 1730647971547}, {"id": "WrpRAJ4DRZ", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7903/Reviewer_PnNh"], "rating": 3, "soundness": 1, "presentation": 2, "contribution": 1, "confidence": 4, "summary": "This paper identifies the gap that cryptic crosswords, a well-known language-oriented reasoning puzzle, have received little attention from the community. To this end, the authors propose a system that combines LLMs and a Python interpreter to solve cryptic clues. In the paper, they elaborate on the system architecture and run experiments using Gemma and Gemini model family to show better performance than the baselines on Cryptonite benchmark.", "review_text": "This paper identifies the gap that cryptic crosswords, a well-known language-oriented reasoning puzzle, have received little attention from the community. To this end, the authors propose a system that combines LLMs and a Python interpreter to solve cryptic clues. In the paper, they elaborate on the system architecture and run experiments using Gemma and Gemini model family to show better performance than the baselines on Cryptonite benchmark.", "strengths": "Identifying the gap that cryptic crosswords have received little attention from the community, especially post LLM era, the authors propose a framework that combines LLMs and a Python interpreter to solve cryptic clues, and show better performance than baselines on Cryptonite benchmark.", "weaknesses": "1. This paper may lack enough strong baselines and the comparisons with these baselines may not be entirely fair. There are 3 baselines including rule-based, a fine-tuned T5 large, and Gemini Pro 1.0 zero-shot. First, The T5 large model only has 770M parameters while Gemma 9b (which is used in this paper) has 9B parameters. Comparing Gemma 9B-FT with T5, we can see that changing the model with the same training data improved the accuracy from 7.6% to 15.9%. Second, Gemini 1.5 Flash is used in the proposed system while the baseline is zero-shot Gemini Pro 1.0 which shows inferior performance across general benchmarks. Third, while the papers mentions limited computation resource, rows in table 1 are reported based on different number of examples, which make them not comparable.\n\n2. It may not be clearly stated in the paper the reason of following the proposed order: i.e. generating answer candidates first, then generating possible definitions and wordplays, and then doing verification. Reasoning tasks typically generate reasoning steps followed by the final answer and verification, e.g. ReAct format. In this case, possible definitions and wordplays can be followed by the final answer candidates and external verifiers. This is also consistent with line 44 regarding the reasoning steps. Have you tried different variations and concluded that the proposed order is the best?", "questions": "1. Regarding weakness 2, have you tried different variations and concluded that the proposed order is the best?\n2. Since the current system combines multiple open source fine-tuned models (Gemma) and a closed source model (Gemini) for different stages, I wonder if you have tried a single model performing all different tasks, either prompting or fine-tuning, which makes the system simpler?\n3. The last paragraph in section 2.2 mentions that an apple-to-apple comparison with Rozner et al. (2021) is hard because of different split setting. Could you elaborate more on the challenge of comparing using the same data? Is it because the test split in one approach is used for training of another approach, or their model is not publicly available?\n\nMinor points => Typo(s): \n1. Line 155, T5 large is 770M\n2. Line 200, the first sentence in section 3.1 has two “from”\n3. Line 344, that -> than", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper identifies the gap that cryptic crosswords, a well-known language-oriented reasoning puzzle, have received little attention from the community. To this end, the authors propose a system that combines LLMs and a Python interpreter to solve cryptic clues. In the paper, they elaborate on the system architecture and run experiments using Gemma and Gemini model family to show better performance than the baselines on Cryptonite benchmark.", "soundness": 1, "presentation": 2, "contribution": 1, "strengths": "Identifying the gap that cryptic crosswords have received little attention from the community, especially post LLM era, the authors propose a framework that combines LLMs and a Python interpreter to solve cryptic clues, and show better performance than baselines on Cryptonite benchmark.", "weaknesses": "1. This paper may lack enough strong baselines and the comparisons with these baselines may not be entirely fair. There are 3 baselines including rule-based, a fine-tuned T5 large, and Gemini Pro 1.0 zero-shot. First, The T5 large model only has 770M parameters while Gemma 9b (which is used in this paper) has 9B parameters. Comparing Gemma 9B-FT with T5, we can see that changing the model with the same training data improved the accuracy from 7.6% to 15.9%. Second, Gemini 1.5 Flash is used in the proposed system while the baseline is zero-shot Gemini Pro 1.0 which shows inferior performance across general benchmarks. Third, while the papers mentions limited computation resource, rows in table 1 are reported based on different number of examples, which make them not comparable.\n\n2. It may not be clearly stated in the paper the reason of following the proposed order: i.e. generating answer candidates first, then generating possible definitions and wordplays, and then doing verification. Reasoning tasks typically generate reasoning steps followed by the final answer and verification, e.g. ReAct format. In this case, possible definitions and wordplays can be followed by the final answer candidates and external verifiers. This is also consistent with line 44 regarding the reasoning steps. Have you tried different variations and concluded that the proposed order is the best?", "questions": "1. Regarding weakness 2, have you tried different variations and concluded that the proposed order is the best?\n2. Since the current system combines multiple open source fine-tuned models (Gemma) and a closed source model (Gemini) for different stages, I wonder if you have tried a single model performing all different tasks, either prompting or fine-tuning, which makes the system simpler?\n3. The last paragraph in section 2.2 mentions that an apple-to-apple comparison with Rozner et al. (2021) is hard because of different split setting. Could you elaborate more on the challenge of comparing using the same data? Is it because the test split in one approach is used for training of another approach, or their model is not publicly available?\n\nMinor points => Typo(s): \n1. Line 155, T5 large is 770M\n2. Line 200, the first sentence in section 3.1 has two “from”\n3. Line 344, that -> than", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730497511532}, {"id": "dE9rbDVrJx", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission7903/Reviewer_2vAu"], "rating": 8, "soundness": 3, "presentation": 4, "contribution": 3, "confidence": 4, "summary": "The paper introduces a novel reasoning approach to generate answers for cryptic crossword clues. It is a multistage approach which does the following:\n1. First, a fine-tuned LLM suggests possible answers for the entire clue.\n2. Second, for each candidate answer, another LLM generates possible definition/wordplay suggestions.\n3. Third, for each suggestion, another few-shot LLM generates python code which wraps their domain-specific language to prove the wordplay. If there's an error in the verifier, it is fed back to the LML in a loop up to K times to improve/correct the proof. If there is no error, the answer is used as the final prediction. \n\nThis system improves overall performance on the Cryptonite dataset from 15.7% to 32.5%, and the authors give some analysis on the benefits and shortcomings of their approach.", "review_text": "The paper introduces a novel reasoning approach to generate answers for cryptic crossword clues. It is a multistage approach which does the following:\n1. First, a fine-tuned LLM suggests possible answers for the entire clue.\n2. Second, for each candidate answer, another LLM generates possible definition/wordplay suggestions.\n3. Third, for each suggestion, another few-shot LLM generates python code which wraps their domain-specific language to prove the wordplay. If there's an error in the verifier, it is fed back to the LML in a loop up to K times to improve/correct the proof. If there is no error, the answer is used as the final prediction. \n\nThis system improves overall performance on the Cryptonite dataset from 15.7% to 32.5%, and the authors give some analysis on the benefits and shortcomings of their approach.", "strengths": "* The authors introduce a novel approach to a uncommonly studied task, which has interest to the NLP/reasoning communities for its difficulty and unique reasoning requirements, and to the cryptic crossword hobbyist community. They provide motivation for why this task is relevant and should be worked on. \n\n* Their approach is effective, more than doubling the previous SOTA performance on the same task. Additionally, it offers several \"obvious\" avenues for improvement as it breaks the task down into more easily tunable parts.\n\n* Finally, the paper is very clearly written and communicates their relatively unknown task and novel method effectively.", "weaknesses": "* A full ablation study would make clear the shortcomings of each part of the pipeline. While the authors demonstrate the effectiveness of just using the top candidate answer, it would be helpful to know the benefit of adding the definition/wordplay model (combined with e.g., a simple LLM-based reranker) compared with the full pipeline.", "questions": "- When you say `The Wordplay dataset follows the train, validation, and test split defined by\n  Cryptonite`, I assume you mean there is no contamination between the two, i.e., no clue in the\n  train dataset for Wordplay will appear in the Cryptonite val/test splits. Is this correct or did\n  you have mean something different?\n- The performance of your pipeline is of course capped by the recall of the candidate answer\n  suggestion model, which seems to be about $40$% at $N=20$. This means that the rest of your pipeline\n  can correctly choose about $32.5 / 40 \\approx 81$% of the results, which seems promising. Have you\n  considered just making N as large as possible? It might be an easy (albeit inefficient) way to\n  improve overall performance by a bit, assuming the rest of the pipeline is robust to many wrong\n  candidates.\n- Also, I'd be curious to know how much the 'verification' system helps. While your system is\n  certainly effective, I wonder how much the verification system improves over a \"dumb\" reranker\n  (e.g., scoring the probability of each (answer, definition/wordplay) pair using the LLM and\n  choosing the best).", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper introduces a novel reasoning approach to generate answers for cryptic crossword clues. It is a multistage approach which does the following:\n1. First, a fine-tuned LLM suggests possible answers for the entire clue.\n2. Second, for each candidate answer, another LLM generates possible definition/wordplay suggestions.\n3. Third, for each suggestion, another few-shot LLM generates python code which wraps their domain-specific language to prove the wordplay. If there's an error in the verifier, it is fed back to the LML in a loop up to K times to improve/correct the proof. If there is no error, the answer is used as the final prediction. \n\nThis system improves overall performance on the Cryptonite dataset from 15.7% to 32.5%, and the authors give some analysis on the benefits and shortcomings of their approach.", "soundness": 3, "presentation": 4, "contribution": 3, "strengths": "* The authors introduce a novel approach to a uncommonly studied task, which has interest to the NLP/reasoning communities for its difficulty and unique reasoning requirements, and to the cryptic crossword hobbyist community. They provide motivation for why this task is relevant and should be worked on. \n\n* Their approach is effective, more than doubling the previous SOTA performance on the same task. Additionally, it offers several \"obvious\" avenues for improvement as it breaks the task down into more easily tunable parts.\n\n* Finally, the paper is very clearly written and communicates their relatively unknown task and novel method effectively.", "weaknesses": "* A full ablation study would make clear the shortcomings of each part of the pipeline. While the authors demonstrate the effectiveness of just using the top candidate answer, it would be helpful to know the benefit of adding the definition/wordplay model (combined with e.g., a simple LLM-based reranker) compared with the full pipeline.", "questions": "- When you say `The Wordplay dataset follows the train, validation, and test split defined by\n  Cryptonite`, I assume you mean there is no contamination between the two, i.e., no clue in the\n  train dataset for Wordplay will appear in the Cryptonite val/test splits. Is this correct or did\n  you have mean something different?\n- The performance of your pipeline is of course capped by the recall of the candidate answer\n  suggestion model, which seems to be about $40$% at $N=20$. This means that the rest of your pipeline\n  can correctly choose about $32.5 / 40 \\approx 81$% of the results, which seems promising. Have you\n  considered just making N as large as possible? It might be an easy (albeit inefficient) way to\n  improve overall performance by a bit, assuming the rest of the pipeline is robust to many wrong\n  candidates.\n- Also, I'd be curious to know how much the 'verification' system helps. While your system is\n  certainly effective, I wonder how much the verification system improves over a \"dumb\" reranker\n  (e.g., scoring the probability of each (answer, definition/wordplay) pair using the LLM and\n  choosing the best).", "flag_for_ethics_review": ["No ethics review needed."], "rating": 8, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729715192298}], "openreview_url": "https://openreview.net/forum?id=Bo5eKnJPML", "arxiv_id": "2506.04824", "paper_pdf": "papers/Bo5eKnJPML.pdf", "paper_pdf_sha256": "7a54ae3cbcd5c4d48606e426ede74932347f8bbb56f15cc3e28d452c5ec9f02d", "paper_pdf_bytes": 556499, "paper_pdf_source": "openreview", "code_url": "https://github.com/mdda/cryptic-crossword-reasoning-verifier", "code_repository": "mdda/cryptic-crossword-reasoning-verifier", "code_commit": "0122293fbe640f0aa4b837d37841eb0df34fec38", "code_archive": "repos/Bo5eKnJPML.zip", "code_archive_sha256": "3a2204f9bda6de0b50722aab90bc9b9711ac8344630481e9402ac1cade509954", "code_archive_bytes": 418350, "code_file_count": 23, "code_extensions": {".py": 14, ".ipynb": 9}, "github_disk_usage_kb": 442, "github_languages": {"Jupyter Notebook": 1208319, "Python": 333246}, "github_archived": false, "github_pushed_at": "2026-01-11T18:50:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/a-reasoning-based-approach-to-cryptic"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "D6aGz0Zyvn", "year": 2024, "status": "rejected", "title": "Enhancing Kernel Flexibility via Learning Asymmetric Locally-Adaptive Kernels", "authors": ["FAN He", "Mingzhen He", "Lei Shi", "Xiaolin Huang", "Johan Suykens"], "authorids": ["~FAN_He1", "~Mingzhen_He1", "~Lei_Shi7", "~Xiaolin_Huang1", "~Johan_Suykens1"], "authors_source": "OpenReview API", "abstract": "The lack of sufficient flexibility is the key bottleneck of kernel-based learning that relies on manually designed, pre-given, and non-trainable kernels. To enhance kernel flexibility, this paper introduces the concept of Locally-Adaptive-Bandwidths (LAB) as trainable parameters to enhance the Radial Basis Function (RBF) kernel, giving rise to the LAB RBF kernel. The parameters in LAB RBF kernels are data-dependent, and its number can increase with the dataset, allowing for better adaptation to diverse data patterns and enhancing the flexibility of the learned function. This newfound flexibility also brings challenges, particularly with regards to asymmetry and the need for an efficient learning algorithm. To address these challenges, this paper for the first time establishes an asymmetric kernel ridge regression framework and introduces an iterative kernel learning algorithm. This novel approach not only reduces the demand for extensive support data but also significantly improves generalization by training bandwidths on the available training data. Experimental results on real datasets underscore the remarkable performance of the proposed algorithm, showcasing its superior capability in handling large-scale datasets compared to Nyström approximation-based algorithms. Moreover, it demonstrates a significant improvement in regression accuracy over existing kernel-based learning methods and even surpasses residual neural networks.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "cHrPCu2nnl", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2653/Reviewer_gJwQ"], "rating": "8: accept, good paper", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes a novel approach to enhance the flexibility of kernel-based learning by introducing Locally-Adaptive-Bandwidth (LAB) kernels. Unlike traditional fixed kernels, LAB kernels incorporate data-dependent bandwidths, allowing for better adaptation to diverse data patterns. To address challenges related to asymmetry and learning efficiency, the paper introduces an asymmetric kernel ridge regression framework and an iterative kernel learning algorithm. Experimental results demonstrate the superior performance of the proposed algorithm compared to existing methods in handling large-scale datasets and achieving higher regression accuracy.", "review_text": "The paper proposes a novel approach to enhance the flexibility of kernel-based learning by introducing Locally-Adaptive-Bandwidth (LAB) kernels. Unlike traditional fixed kernels, LAB kernels incorporate data-dependent bandwidths, allowing for better adaptation to diverse data patterns. To address challenges related to asymmetry and learning efficiency, the paper introduces an asymmetric kernel ridge regression framework and an iterative kernel learning algorithm. Experimental results demonstrate the superior performance of the proposed algorithm compared to existing methods in handling large-scale datasets and achieving higher regression accuracy.", "strengths": "1. The introduction of LAB kernels with trainable bandwidths significantly improves the flexibility of kernel-based learning. By adapting bandwidths to individual data points, the model can better accommodate diverse data patterns, leading to more accurate representations.\n2. The paper establishes an asymmetric kernel ridge regression framework specifically designed for LAB kernels. Despite the asymmetry of the kernel matrix, the stationary points are elegantly represented as a linear combination of function evaluations at training data, enabling efficient learning and inference.\n3. The proposed algorithm allows for the estimation of bandwidths from the training data, reducing the demand for extensive support data. This data-driven approach enhances generalization ability by effectively tuning bandwidths based on the available training data.\n4. The proposed algorithm shows superior scalability in handling large-scale datasets compared to Nyström approximation-based algorithms. LAB kernels, with their adaptive bandwidths, offer a flexible and efficient solution for kernel-based learning tasks with extensive data.", "weaknesses": "1. While the paper presents empirical evidence of the superior performance of the proposed algorithm, it may lack strong theoretical guarantees or formal analysis of its convergence properties. Further theoretical investigations may be needed to fully understand the behavior and limitations of LAB kernels\n2. The performance of LAB kernels heavily relies on the accurate estimation of bandwidths. Selecting appropriate bandwidths for different data patterns can be a challenging task, and suboptimal choices may result in reduced performance or overfitting.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes a novel approach to enhance the flexibility of kernel-based learning by introducing Locally-Adaptive-Bandwidth (LAB) kernels. Unlike traditional fixed kernels, LAB kernels incorporate data-dependent bandwidths, allowing for better adaptation to diverse data patterns. To address challenges related to asymmetry and learning efficiency, the paper introduces an asymmetric kernel ridge regression framework and an iterative kernel learning algorithm. Experimental results demonstrate the superior performance of the proposed algorithm compared to existing methods in handling large-scale datasets and achieving higher regression accuracy.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The introduction of LAB kernels with trainable bandwidths significantly improves the flexibility of kernel-based learning. By adapting bandwidths to individual data points, the model can better accommodate diverse data patterns, leading to more accurate representations.\n2. The paper establishes an asymmetric kernel ridge regression framework specifically designed for LAB kernels. Despite the asymmetry of the kernel matrix, the stationary points are elegantly represented as a linear combination of function evaluations at training data, enabling efficient learning and inference.\n3. The proposed algorithm allows for the estimation of bandwidths from the training data, reducing the demand for extensive support data. This data-driven approach enhances generalization ability by effectively tuning bandwidths based on the available training data.\n4. The proposed algorithm shows superior scalability in handling large-scale datasets compared to Nyström approximation-based algorithms. LAB kernels, with their adaptive bandwidths, offer a flexible and efficient solution for kernel-based learning tasks with extensive data.", "weaknesses": "1. While the paper presents empirical evidence of the superior performance of the proposed algorithm, it may lack strong theoretical guarantees or formal analysis of its convergence properties. Further theoretical investigations may be needed to fully understand the behavior and limitations of LAB kernels\n2. The performance of LAB kernels heavily relies on the accurate estimation of bandwidths. Selecting appropriate bandwidths for different data patterns can be a challenging task, and suboptimal choices may result in reduced performance or overfitting.", "questions": "See Weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698984121715}, {"id": "CBGjaQ8OJi", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2653/Reviewer_aToc"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This paper introduces a new asymmetric kernel names Local-Adaptive-Bandwidth RBF kernel. To solve the asymmetry of the kernel, the paper establishes an asymmetric KRR framework. To learn the kernel parameter efficiently and accelerate computation. the paper devises a kernel learning algorithm. Experimental results show the algorithm’s superiority.", "review_text": "This paper introduces a new asymmetric kernel names Local-Adaptive-Bandwidth RBF kernel. To solve the asymmetry of the kernel, the paper establishes an asymmetric KRR framework. To learn the kernel parameter efficiently and accelerate computation. the paper devises a kernel learning algorithm. Experimental results show the algorithm’s superiority.", "strengths": "1. The paper demonstrates a clear logical structure with a comprehensive framework. It tackles the complex relationship between bandwidth and data from the perspective of experimental results.\n2. The paper takes into account the impact of differences in implicit mappings on the results and proposes an interesting approach to non-symmetric kernel KRR framework.\n3. The paper introduces an algorithm based on dynamic strategies for parameter computation, which can effectively reduce the computational complexity associated with high-dimensional kernel matrices.", "weaknesses": "1. Intuitively, the relation between the mapping function's distinctiveness and the loss function, which means the coefficient of the last term in the KRR optimization objective may vary with datasets. \n2. The initial data selection for support data in the kernel learning algorithm proposed in the article seems to be too random. Moreover, inappropriate data selection appears to have a significant impact on the model.", "questions": "1. Is the final coefficient in the asymmetric KRR framework proposed in the article required to be 1/2? Can this be understood as simply for the convenience of computing stationary points? \n2. Is the small number of support vectors in the experimental results of the proposed method due to the algorithm's termination condition?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper introduces a new asymmetric kernel names Local-Adaptive-Bandwidth RBF kernel. To solve the asymmetry of the kernel, the paper establishes an asymmetric KRR framework. To learn the kernel parameter efficiently and accelerate computation. the paper devises a kernel learning algorithm. Experimental results show the algorithm’s superiority.", "soundness": "3 good", "presentation": "3 good", "contribution": "3 good", "strengths": "1. The paper demonstrates a clear logical structure with a comprehensive framework. It tackles the complex relationship between bandwidth and data from the perspective of experimental results.\n2. The paper takes into account the impact of differences in implicit mappings on the results and proposes an interesting approach to non-symmetric kernel KRR framework.\n3. The paper introduces an algorithm based on dynamic strategies for parameter computation, which can effectively reduce the computational complexity associated with high-dimensional kernel matrices.", "weaknesses": "1. Intuitively, the relation between the mapping function's distinctiveness and the loss function, which means the coefficient of the last term in the KRR optimization objective may vary with datasets. \n2. The initial data selection for support data in the kernel learning algorithm proposed in the article seems to be too random. Moreover, inappropriate data selection appears to have a significant impact on the model.", "questions": "1. Is the final coefficient in the asymmetric KRR framework proposed in the article required to be 1/2? Can this be understood as simply for the convenience of computing stationary points? \n2. Is the small number of support vectors in the experimental results of the proposed method due to the algorithm's termination condition?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698828024453}, {"id": "pxlukMmsOr", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission2653/Reviewer_PWHQ"], "rating": "8: accept, good paper", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The algorithm introduces a variant of the centered-based kernel method, they call the centers \"support vectors\". The idea is that the model's size remains smaller than the entire dataset, similar to  FALKON and EigenPro3.0. In contrast to methods that utilize fixed model support vectors, their algorithm adaptively adjusts these support vectors throughout the training process. Moreover, utilizing concepts from the asymmetric kernel method, they adaptively fit the support vectors with varying bandwidths throughout the training process.They show that their algorithm outperform some existing methods on several data sets.", "review_text": "The algorithm introduces a variant of the centered-based kernel method, they call the centers \"support vectors\". The idea is that the model's size remains smaller than the entire dataset, similar to  FALKON and EigenPro3.0. In contrast to methods that utilize fixed model support vectors, their algorithm adaptively adjusts these support vectors throughout the training process. Moreover, utilizing concepts from the asymmetric kernel method, they adaptively fit the support vectors with varying bandwidths throughout the training process.They show that their algorithm outperform some existing methods on several data sets.", "strengths": "1. the most interesting part is the idea of adaptively change the support vectors in an iterative manner. This was something novel worth exploring more.\n2. The idea of adaptively adjust the bandwidth and mixing it with asymmetric kernel methods seems intriguing.(Not sure how useful)\n3. The paper is clear and easy to follow.", "weaknesses": "Main concern:\n1. The most important caveats is that the paper has only compared to vanilla KRR methods. It is not surprising that they got a slightly better performance compare for example to FALKON. I'm not at all convened this methods is better than well developed techniques such as:\n\ni. traditional Automatic Relevance Determination(ARD) known in GP community, it is implemented with Gpytorch see here:https://docs.gpytorch.ai/en/stable/kernels.html. or see section 5 of this https://gaussianprocess.org/gpml/chapters/RW.pdf\n\nii. EigenPro3.0, see https://arxiv.org/abs/2302.02605 \n\niii. Recursive Feature machines (RFMs), see: https://arxiv.org/abs/2212.13881\n\nScalability:\n\n2. the authors claim that this method is scalable and they provide table 3 to justify this. But those data sets are not at all large scale. The inverse problem can be done using direct calculation for those cases. The authors should try other data sets such as Taxi, CIFAR5m to justify consistency and scalability. (both in data and model size)\n3. It is mentioned in section 3 that the computation complexity is O(N_sv^3). This fundamentally shows this method on its own is not scalable. Eventually you need to scale the required support vectors(or model size) as it is discussed in https://arxiv.org/abs/2302.02605. \nHowever, I can see that this method combined by other methods like FALKON or EigenPro3.0  can potentially be scalable.\n4. How do you compute line 4 of the algorithm? Did you use FALKON or some other off the shelf algorithm or you did direct inverse? \n\n\nMinor issues:\n\n1. RBF kernel are known to be sensitive to bandwidth. While you have results for MKL, the performance of your method specifically for the Laplace kernel, which is relatively insensitive to bandwidth, remains ambiguous. Does outperforming MKL indicate superiority over merely using Laplace? The same concern applies to NTK kernels or other popular kernels.\n2. I suggest more explaining for asymmetric kernels methods. For example why the inverse even exist in equation 6. or you claimed \"this paper for the first time establishes an asymmetric KRR framework\", but how is it different from He et. al. paper? not clear.\n3. Please add what M means in the tables, helps with reading.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The algorithm introduces a variant of the centered-based kernel method, they call the centers \"support vectors\". The idea is that the model's size remains smaller than the entire dataset, similar to  FALKON and EigenPro3.0. In contrast to methods that utilize fixed model support vectors, their algorithm adaptively adjusts these support vectors throughout the training process. Moreover, utilizing concepts from the asymmetric kernel method, they adaptively fit the support vectors with varying bandwidths throughout the training process.They show that their algorithm outperform some existing methods on several data sets.", "soundness": "2 fair", "presentation": "3 good", "contribution": "3 good", "strengths": "1. the most interesting part is the idea of adaptively change the support vectors in an iterative manner. This was something novel worth exploring more.\n2. The idea of adaptively adjust the bandwidth and mixing it with asymmetric kernel methods seems intriguing.(Not sure how useful)\n3. The paper is clear and easy to follow.", "weaknesses": "Main concern:\n1. The most important caveats is that the paper has only compared to vanilla KRR methods. It is not surprising that they got a slightly better performance compare for example to FALKON. I'm not at all convened this methods is better than well developed techniques such as:\n\ni. traditional Automatic Relevance Determination(ARD) known in GP community, it is implemented with Gpytorch see here:https://docs.gpytorch.ai/en/stable/kernels.html. or see section 5 of this https://gaussianprocess.org/gpml/chapters/RW.pdf\n\nii. EigenPro3.0, see https://arxiv.org/abs/2302.02605 \n\niii. Recursive Feature machines (RFMs), see: https://arxiv.org/abs/2212.13881\n\nScalability:\n\n2. the authors claim that this method is scalable and they provide table 3 to justify this. But those data sets are not at all large scale. The inverse problem can be done using direct calculation for those cases. The authors should try other data sets such as Taxi, CIFAR5m to justify consistency and scalability. (both in data and model size)\n3. It is mentioned in section 3 that the computation complexity is O(N_sv^3). This fundamentally shows this method on its own is not scalable. Eventually you need to scale the required support vectors(or model size) as it is discussed in https://arxiv.org/abs/2302.02605. \nHowever, I can see that this method combined by other methods like FALKON or EigenPro3.0  can potentially be scalable.\n4. How do you compute line 4 of the algorithm? Did you use FALKON or some other off the shelf algorithm or you did direct inverse? \n\n\nMinor issues:\n\n1. RBF kernel are known to be sensitive to bandwidth. While you have results for MKL, the performance of your method specifically for the Laplace kernel, which is relatively insensitive to bandwidth, remains ambiguous. Does outperforming MKL indicate superiority over merely using Laplace? The same concern applies to NTK kernels or other popular kernels.\n2. I suggest more explaining for asymmetric kernels methods. For example why the inverse even exist in equation 6. or you claimed \"this paper for the first time establishes an asymmetric KRR framework\", but how is it different from He et. al. paper? not clear.\n3. Please add what M means in the tables, helps with reading.", "questions": "see weaknesses", "flag_for_ethics_review": ["No ethics review needed."], "rating": "8: accept, good paper", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698701875200}], "openreview_url": "https://openreview.net/forum?id=D6aGz0Zyvn", "arxiv_id": "2310.05236", "paper_pdf": "papers/D6aGz0Zyvn.pdf", "paper_pdf_sha256": "1ecddbb3608915a6ebb9c6c87cc1163154f3fde7264783fb76afa2aa93518141", "paper_pdf_bytes": 424349, "paper_pdf_source": "openreview", "code_url": "https://github.com/hefansjtu/LABRBF_kernel", "code_repository": "hefansjtu/LABRBF_kernel", "code_commit": "0ed445b3ac5bae7e0cc6bb25a5064ec028bf8436", "code_archive": "repos/D6aGz0Zyvn.zip", "code_archive_sha256": "e784a7e42e8727b25f19d99595aca7b11797ca639b4a757a4d96b69c3a920416", "code_archive_bytes": 553627, "code_file_count": 10, "code_extensions": {".py": 10}, "github_disk_usage_kb": 556, "github_languages": {"Python": 100417}, "github_archived": false, "github_pushed_at": "2023-11-24T12:14:16Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/enhancing-kernel-flexibility-via-learning"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "DvMDIEFtyjV", "year": 2023, "status": "rejected", "title": "CLUTR: Curriculum Learning via Unsupervised Task Representation Learning", "authors": ["Abdus Salam Azad", "Izzeddin Gur", "Aleksandra Faust", "Pieter Abbeel", "Ion Stoica"], "authorids": ["~Abdus_Salam_Azad1", "~Izzeddin_Gur1", "~Aleksandra_Faust1", "~Pieter_Abbeel2", "~Ion_Stoica1"], "authors_source": "OpenReview API", "abstract": "Reinforcement Learning (RL) algorithms are often known for sample inefficiency and difficult generalization. Recently, Unsupervised Environment Design (UED) emerged as a new paradigm for zero-shot generalization by simultaneously learning a task distribution and agent policies on the sampled tasks. This is a non-stationary process where the task distribution evolves along with agent policies; creating an instability over time. While past works demonstrated the potential of such approaches, sampling effectively from the task space remains an open challenge, bottlenecking these approaches. To this end, we introduce CLUTR: a novel curriculum learning algorithm that decouples task representation and curriculum learning into a two-stage optimization. It first trains a recurrent variational autoencoder on randomly generated tasks to learn a latent task manifold. Next, a teacher agent creates a curriculum by optimizing a minimax REGRET-based objective on a set of latent tasks sampled from this manifold. By keeping the task manifold fixed, we show that CLUTR successfully overcomes the non-stationarity problem and improves stability. Our experimental results show CLUTR outperforms PAIRED, a principled and popular UED method, in terms of generalization and sample efficiency in the challenging CarRacing and navigation environments: showing an 18x improvement on the F1 CarRacing benchmark. CLUTR also performs comparably to the non-UED state-of-the-art for CarRacing, outperforming it in nine of the 20 tracks. CLUTR also achieves a 33% higher solved rate than PAIRED on a set of 18 out-of-distribution navigation tasks.", "decision": "Reject", "meta_review": null, "num_reviews": 3, "reviews": [{"id": "xZ5HWnkJMH", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3908/Reviewer_NPYm"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "summary": "This paper investigates learning a task representation space for use in unsupervised environment design (UED), with an algorithm called PAIRED. PAIRED trains an RL agent, the adversary, to configure the environment parameters throughout training, in tandem with two RL agents that learn to solve these configurations (each configuration is termed a \"task\"). The adversary is allied with one of the agents, called the antagonist, and seeks to maximize the margin of return by which the antagonist outperforms the other agent, called the protagonist. Thus, PAIRED produces a curriculum that maximizes a lower bound on the regret experienced by the protagonist, leading to a robust minimax-regret policy at equilibrium. This paper then proposes CLUTR, a method that first trains a recurrent VAE over a large number of randomized tasks in a specific domain (e.g. mazes or car racing tracks) and then performs PAIRED over the learned latent space of the VAE. Their results show that PAIRED in such latent-spaces produces more robust protagonist policies than standard PAIRED, which designs tasks in the direct task space.", "review_text": "This paper provides an interesting extension of UED, in which a PAIRED adversary, which generates regret-maximizing tasks for an RL agent, searches for such tasks in a learned latent space. The idea in itself is interesting, but the current paper lacks important experimental details and makes several ambiguous or unjustified claims. Therefore, I cannot recommend this paper for acceptance.", "strengths": "### Strengths\n- The benefits of performing UED on a learned latent manifold of the task space is well-motivated.\n- The connection between CLUTR and prior works is clearly discussed.\n- The CarRacing experiments compare against a comprehensive set of baselines.\n\n### Weaknesses\nDespite what seem like promising results showing CLUTR outperforms PAIRED, there are several important points that are left unclear and claims that seem to be incorrect:\n\n**The regret objective used by CLUTR is unclear**\n\n- It is not clear whether CLUTR uses the flexible PAIRED objective or the standard PAIRED objective. Appendix D.2 seeks to compare CLUTR with the _standard PAIRED regret objective_, implying the main results use the flexible PAIRED objective.\n- If CLUTR uses the flexible regret objective, then the main experiment results should compare to PAIRED with the flexible regret objective, also known as \"flexible PAIRED\" in Dennis et al, 2020. It must be shown that CLUTR's gains over PAIRED with the standard regret objective are not in fact due to the choice of regret objective, rather than the task representation, as claimed.\n- If CLUTR does in fact use the flexible regret objective, then Theorem 1 from Dennis et al, 2020, which guarantees a minimax-regret policy for the protagonist at Nash equilibrium of the PAIRED game, is no longer directly applicable—since the flexible regret objective changes the nature of the underlying game between the three agents.\n- Moreover, the flexible regret objective was not introduced in Gur et al, 2021 as stated in the first paragraph of Section 3.2, but clearly detailed and evaluated in the appendix of Dennis et al, 2020.\n\n**CLUTR does not have access to the full task space**\n\nWhile PAIRED, in principle, has the ability to generate any 50-block maze, this is not necessarily true for CLUTR. This is because CLUTR is trained on a limited subset of all possible tasks, and the learned latent space is not guaranteed to contain all possible mazes. Therefore, Property 3 of CLUTR, stated on Page 5, is incorrect.\n\n**Interpretation of Figure 4 and 5 is flawed**\n\n- In Figure 5, CLUTR's adversary achieves lower regret than that of PAIRED. This result implies that the latent task representation space makes the optimization of the regret objective a harder problem for the teacher, rather than easier, as claimed in the paper. \n- Figure 4 shows the solved rate of the agent on the _training_ levels proposed by the CLUTR and PAIRED adversaries. Since the goal of UED is to produce an adversarial curriculum, it seems that the higher solved rates achieved by the agent under the CLUTR adversary implies the CLUTR adversary is less effective and does a worse job at optimizing for agent's regret. \n\n**The results do not provide evidence for whether the VAE and CLUTR adversary learn anything**\n\n- While the paper argues that CLUTR has the advantage of avoiding sequential credit-assignment, CLUTR must pay the additional cost of a much larger action space—based on the latent-space dimensionality. Thus, CLUTR may in fact face a much more difficult learning problem than PAIRED. Further, the VAE must also learn a useful representation of the task space, the success of which is not shown in the paper.\n- The paper should provide visualizations of how tasks are distributed within the VAE's latent space to provide evidence that meaningful latent structure is being learned via their method. The training curves for the VAE would also be useful to see.\n- The paper should provide evidence that CLUTR's metrics in Figures 3, 4, and 5 are different from that of domain randomization—which would show that CLUTR's adversary is learning to exploit the structure in the latent space, rather than resorting to a randomized policy.\n\nThese figures should compare to domain randomization (DR), and importantly, show that CLUTR is doing something meaningfully different from randomizing over the latent space (which would be case if the adversary policy has difficulty learning to design in this space).\n\n**Curriculum analysis is lacking**\n\n- The main emergent complexity and curriculum results can benefit from thorough analysis. In particular, the authors should plot the solution path length and number of blocks in training maze tasks proposed by the CLUTR adversary, in comparison to other methods. \n- Since the CLUTR tasks are generated by decoding latent representations using the pre-trained VAE, it is possible that CLUTR produces mazes with more obstacle blocks than PAIRED, which is strictly limited to at most 50 blocks. Prior works (Jiang et al, 2021) show that going from 25 to 50 blocks improves the OOD transfer performance of all methods compared, so it seems that this factor is a confounder.\n- Importantly, this curriculum analysis should compare to DR to show that CLUTR learns a meaningful curriculum.\n\n**Missing several key details**\n\nIn addition to the lack of clarity around CLUTR's regret objective, a few other important details seem to be missing. (See the Clarity section for details.)\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "This paper investigates learning a task representation space for use in unsupervised environment design (UED), with an algorithm called PAIRED. PAIRED trains an RL agent, the adversary, to configure the environment parameters throughout training, in tandem with two RL agents that learn to solve these configurations (each configuration is termed a \"task\"). The adversary is allied with one of the agents, called the antagonist, and seeks to maximize the margin of return by which the antagonist outperforms the other agent, called the protagonist. Thus, PAIRED produces a curriculum that maximizes a lower bound on the regret experienced by the protagonist, leading to a robust minimax-regret policy at equilibrium. This paper then proposes CLUTR, a method that first trains a recurrent VAE over a large number of randomized tasks in a specific domain (e.g. mazes or car racing tracks) and then performs PAIRED over the learned latent space of the VAE. Their results show that PAIRED in such latent-spaces produces more robust protagonist policies than standard PAIRED, which designs tasks in the direct task space.", "strength_and_weaknesses": "### Strengths\n- The benefits of performing UED on a learned latent manifold of the task space is well-motivated.\n- The connection between CLUTR and prior works is clearly discussed.\n- The CarRacing experiments compare against a comprehensive set of baselines.\n\n### Weaknesses\nDespite what seem like promising results showing CLUTR outperforms PAIRED, there are several important points that are left unclear and claims that seem to be incorrect:\n\n**The regret objective used by CLUTR is unclear**\n\n- It is not clear whether CLUTR uses the flexible PAIRED objective or the standard PAIRED objective. Appendix D.2 seeks to compare CLUTR with the _standard PAIRED regret objective_, implying the main results use the flexible PAIRED objective.\n- If CLUTR uses the flexible regret objective, then the main experiment results should compare to PAIRED with the flexible regret objective, also known as \"flexible PAIRED\" in Dennis et al, 2020. It must be shown that CLUTR's gains over PAIRED with the standard regret objective are not in fact due to the choice of regret objective, rather than the task representation, as claimed.\n- If CLUTR does in fact use the flexible regret objective, then Theorem 1 from Dennis et al, 2020, which guarantees a minimax-regret policy for the protagonist at Nash equilibrium of the PAIRED game, is no longer directly applicable—since the flexible regret objective changes the nature of the underlying game between the three agents.\n- Moreover, the flexible regret objective was not introduced in Gur et al, 2021 as stated in the first paragraph of Section 3.2, but clearly detailed and evaluated in the appendix of Dennis et al, 2020.\n\n**CLUTR does not have access to the full task space**\n\nWhile PAIRED, in principle, has the ability to generate any 50-block maze, this is not necessarily true for CLUTR. This is because CLUTR is trained on a limited subset of all possible tasks, and the learned latent space is not guaranteed to contain all possible mazes. Therefore, Property 3 of CLUTR, stated on Page 5, is incorrect.\n\n**Interpretation of Figure 4 and 5 is flawed**\n\n- In Figure 5, CLUTR's adversary achieves lower regret than that of PAIRED. This result implies that the latent task representation space makes the optimization of the regret objective a harder problem for the teacher, rather than easier, as claimed in the paper. \n- Figure 4 shows the solved rate of the agent on the _training_ levels proposed by the CLUTR and PAIRED adversaries. Since the goal of UED is to produce an adversarial curriculum, it seems that the higher solved rates achieved by the agent under the CLUTR adversary implies the CLUTR adversary is less effective and does a worse job at optimizing for agent's regret. \n\n**The results do not provide evidence for whether the VAE and CLUTR adversary learn anything**\n\n- While the paper argues that CLUTR has the advantage of avoiding sequential credit-assignment, CLUTR must pay the additional cost of a much larger action space—based on the latent-space dimensionality. Thus, CLUTR may in fact face a much more difficult learning problem than PAIRED. Further, the VAE must also learn a useful representation of the task space, the success of which is not shown in the paper.\n- The paper should provide visualizations of how tasks are distributed within the VAE's latent space to provide evidence that meaningful latent structure is being learned via their method. The training curves for the VAE would also be useful to see.\n- The paper should provide evidence that CLUTR's metrics in Figures 3, 4, and 5 are different from that of domain randomization—which would show that CLUTR's adversary is learning to exploit the structure in the latent space, rather than resorting to a randomized policy.\n\nThese figures should compare to domain randomization (DR), and importantly, show that CLUTR is doing something meaningfully different from randomizing over the latent space (which would be case if the adversary policy has difficulty learning to design in this space).\n\n**Curriculum analysis is lacking**\n\n- The main emergent complexity and curriculum results can benefit from thorough analysis. In particular, the authors should plot the solution path length and number of blocks in training maze tasks proposed by the CLUTR adversary, in comparison to other methods. \n- Since the CLUTR tasks are generated by decoding latent representations using the pre-trained VAE, it is possible that CLUTR produces mazes with more obstacle blocks than PAIRED, which is strictly limited to at most 50 blocks. Prior works (Jiang et al, 2021) show that going from 25 to 50 blocks improves the OOD transfer performance of all methods compared, so it seems that this factor is a confounder.\n- Importantly, this curriculum analysis should compare to DR to show that CLUTR learns a meaningful curriculum.\n\n**Missing several key details**\n\nIn addition to the lack of clarity around CLUTR's regret objective, a few other important details seem to be missing. (See the Clarity section for details.)\n", "clarity,_quality,_novelty_and_reproducibility": "### Clarity\nThe paper does a good job of explaining the high-level details of the method, as well as providing background on related topics and the experimental setup. However, there are several important details that are missing:\n- As previously mentioned, it is not clear whether the main results are based on running CLUTR with the flexible regret objective or the standard PAIRED regret objective.\n- The maze navigation results should report the number of training seeds used.\n- It is not clear why the CarRacing experiments compare to the additional Robust PLR and ACCEL baselines, but the navigation experiments do not.\n- In Section 5.4, the method of fine-tuning the decoder throughout training on the agent's regret is not provided anywhere in the paper. Could the authors provide details on how this fine-tuning is performed? It is unclear, as it seems the VAE was not trained to predict the regret values—which are not available during pre-training.\n\n### Quality\n- Overall, the paper has several rough edges and could benefit from copy editing.\n- There are several claims that are incorrect or unjustified by experimental evidence, which are detailed in the Weaknesses section of this review.\n\n### Novelty\n- The idea of latent-space task design has been explored in other works such as Florensa et al, 2017, which the authors should also cite. \n- The idea of learning a task representation for UED has not been previously explored in detail.\n\n### Reproducibility\n- Due to the lack of clarity around key details described in this review, and the lack of experimental code from the authors, the results of this study are not reproducible based on the information available.\n\n**References**\nFlorensa, Carlos, et al. \"Automatic goal generation for reinforcement learning agents.\" International conference on machine learning. PMLR, 2018.", "summary_of_the_review": "This paper provides an interesting extension of UED, in which a PAIRED adversary, which generates regret-maximizing tasks for an RL agent, searches for such tasks in a learned latent space. The idea in itself is interesting, but the current paper lacks important experimental details and makes several ambiguous or unjustified claims. Therefore, I cannot recommend this paper for acceptance.", "correctness": "1: The main claims of the paper are incorrect or not at all supported by theory or empirical results.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "Not applicable", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough"}, "tcdate": 1666547138645}, {"id": "iK_ip0H82R", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3908/Reviewer_Ui8Z"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper proposes CLUTR, a curriculum learning algorithm based on PAIRED, which decouples\ntask representation and curriculum learning. The task representation is learned \nusing a LSTM-based recurrent VAE, after which the teacher updates the curriculum and \nthe protagonist and antagonist policies are updated according to the regret\nof the generated tasks. The experiments showed empirically that CLUTR outperforms\nstate-of-the-art UED methods in terms of sample efficientyc and generalization\n on a range of tasks, and performs compratably to the non-UED state-of-the-art on the CarRacing task. \n", "review_text": "The paper tackles the problem of sample inefficiency in RL problems by pre-training a VAE to embed the tasks in to a manifold and update policies based on sampled trajectories from the tasks, however, the novelty of the paper is a bit limited. Having said that. given the empirical performance on the extensive sets of experiments, the proposed method may be helpful for the community in practice. \n", "strengths": "\nstrength: The paper is well written.\nThe experiments showed promising results and covered a range of hypothesis claimed in the paper. \n\n\nweakness: the novalty of the paper is somewhat limited. The main techniques used in CLUTR are based on VAE and PAIRED, which are existing methods. \n\nI am curious to know the effect of learning the representation of the tasks. \nFor example, how sensitive does the performance depend on the fitting of the VAE representation? Is the VAE represntation is mis-specified, how much does it worsen the performance?\n\n\nMinor:\n* E is not defined in 4.1 \n* equations in 4.1 are better written with $\\log q(z)$ \n* missing space in section 4: Section4.4 -> Section 4.4\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes CLUTR, a curriculum learning algorithm based on PAIRED, which decouples\ntask representation and curriculum learning. The task representation is learned \nusing a LSTM-based recurrent VAE, after which the teacher updates the curriculum and \nthe protagonist and antagonist policies are updated according to the regret\nof the generated tasks. The experiments showed empirically that CLUTR outperforms\nstate-of-the-art UED methods in terms of sample efficientyc and generalization\n on a range of tasks, and performs compratably to the non-UED state-of-the-art on the CarRacing task. \n", "strength_and_weaknesses": "\nstrength: The paper is well written.\nThe experiments showed promising results and covered a range of hypothesis claimed in the paper. \n\n\nweakness: the novalty of the paper is somewhat limited. The main techniques used in CLUTR are based on VAE and PAIRED, which are existing methods. \n\nI am curious to know the effect of learning the representation of the tasks. \nFor example, how sensitive does the performance depend on the fitting of the VAE representation? Is the VAE represntation is mis-specified, how much does it worsen the performance?\n\n\nMinor:\n* E is not defined in 4.1 \n* equations in 4.1 are better written with $\\log q(z)$ \n* missing space in section 4: Section4.4 -> Section 4.4\n", "clarity,_quality,_novelty_and_reproducibility": "In general, the paper is well presented and well written. The authors claim they will release the code upon acceptance. \n", "summary_of_the_review": "The paper tackles the problem of sample inefficiency in RL problems by pre-training a VAE to embed the tasks in to a manifold and update policies based on sampled trajectories from the tasks, however, the novelty of the paper is a bit limited. Having said that. given the empirical performance on the extensive sets of experiments, the proposed method may be helpful for the community in practice. \n", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666530164297}, {"id": "d7uC4y_8dR", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper3908/Reviewer_eq4e"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper targets the topic of unsupervised environment design (UED) in curriculum reinforcement learning. In particular, it modifies PAIRED by using the decoder of a VAE as the task generator. With this improvement, the model decouples task representation and curriculum learning into a two-stage optimization. Thus, the authors observe that the training is more stable and the final reward is higher than existing UED methods.", "review_text": "Due to my concerns about the clarification, I suggest rejecting this paper. The authors are welcome to address my questions.", "strengths": "### **Strength**\n\n* **Interesting topic.** Unsupervised Environment Design is an important direction in curriculum reinforcement learning. Previous works usually suffer from unstable training problems. This paper targets at solving this problem by using the decoder of VAE to generate tasks.\n\n* **Comprehensive experiment.** The experiment part looks good to me. Some important factors are explored respectively.\n\n### **Weaknesses**\n\n* **Things to be clarified.** \n    * The proposed method heavily depends on PAIRED but details of PAIRED are missing in Section 4.3. \n    * In the 4th line of Algorithm 1, what does it mean by “Use adversary to sample latent task vector”? If the adversary is a parametrized model, what distribution is used for sampling z?\n    * What is single-step RL? It seems not a widely used term but the definition is missing. I guess the authors mean that their method does not do sequential generation with the interaction with the environment.\n    * The second difference between CLUTR and PAIRED is hard to follow. More explanation is needed. Why does the state space of PAIRED rely on the underlying POMDP?\n    * No example of permutation parameters is given in the experiment environment. The authors say that “for a navigation task, a set of obstacles corresponds to factorially different permutations of the parameters.” What is the obstacle in the Car racing experiment? \n    * I don’t think VAE(z, E) in equation (1) is a standard notation. It should be clarified.\n\n\n* **Formulation of CLUTR.** The derivation in Appendix B seems redundant to me. According to Figure 1, variable R actually does not involve the VAE inference since it only depends on another observable variable E. Therefore, Equation (1) is just the ELBO of VAE plus an arbitrary regularization term that relates R and E. I think the authors can directly say that rather than taking a detour.\n\n* **Motivation.** If I understand correctly, this paper proposes to use VAE to generate tasks because of two advantages, i.e., long-horizon credit assignment and permutation invariant. However, searching in the latent space itself is not an easy task since the smoothness of latent space heavily depends on the training and quality of the dataset. So, does this method always outperform PAIRED in general? For example, when most of the uniformly sampled environment parameters are invalid.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "summary_of_the_paper": "This paper targets the topic of unsupervised environment design (UED) in curriculum reinforcement learning. In particular, it modifies PAIRED by using the decoder of a VAE as the task generator. With this improvement, the model decouples task representation and curriculum learning into a two-stage optimization. Thus, the authors observe that the training is more stable and the final reward is higher than existing UED methods.", "strength_and_weaknesses": "### **Strength**\n\n* **Interesting topic.** Unsupervised Environment Design is an important direction in curriculum reinforcement learning. Previous works usually suffer from unstable training problems. This paper targets at solving this problem by using the decoder of VAE to generate tasks.\n\n* **Comprehensive experiment.** The experiment part looks good to me. Some important factors are explored respectively.\n\n### **Weaknesses**\n\n* **Things to be clarified.** \n    * The proposed method heavily depends on PAIRED but details of PAIRED are missing in Section 4.3. \n    * In the 4th line of Algorithm 1, what does it mean by “Use adversary to sample latent task vector”? If the adversary is a parametrized model, what distribution is used for sampling z?\n    * What is single-step RL? It seems not a widely used term but the definition is missing. I guess the authors mean that their method does not do sequential generation with the interaction with the environment.\n    * The second difference between CLUTR and PAIRED is hard to follow. More explanation is needed. Why does the state space of PAIRED rely on the underlying POMDP?\n    * No example of permutation parameters is given in the experiment environment. The authors say that “for a navigation task, a set of obstacles corresponds to factorially different permutations of the parameters.” What is the obstacle in the Car racing experiment? \n    * I don’t think VAE(z, E) in equation (1) is a standard notation. It should be clarified.\n\n\n* **Formulation of CLUTR.** The derivation in Appendix B seems redundant to me. According to Figure 1, variable R actually does not involve the VAE inference since it only depends on another observable variable E. Therefore, Equation (1) is just the ELBO of VAE plus an arbitrary regularization term that relates R and E. I think the authors can directly say that rather than taking a detour.\n\n* **Motivation.** If I understand correctly, this paper proposes to use VAE to generate tasks because of two advantages, i.e., long-horizon credit assignment and permutation invariant. However, searching in the latent space itself is not an easy task since the smoothness of latent space heavily depends on the training and quality of the dataset. So, does this method always outperform PAIRED in general? For example, when most of the uniformly sampled environment parameters are invalid.\n", "clarity,_quality,_novelty_and_reproducibility": "### **Clarity**\nThere are some clarification problems (as mentioned in the weakness part) that need to be solved. Although the general idea of the proposed method is easy to understand, a lot of details are missing.\n\n### **Quality**\nThe quality of this paper is limited by the presentation. A clearer organization would improve the quality.\n\n### **Novelty**\nThe novelty of this paper is ok but maybe around the average level. If I understand correctly, the only modification of this paper is replacing the task generator with a pre-trained VAE.\n\n### **Reproducibility**\nNot sure. Code is not provided.\n\n\n", "summary_of_the_review": "Due to my concerns about the clarification, I suggest rejecting this paper. The authors are welcome to address my questions.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666141852710}], "openreview_url": "https://openreview.net/forum?id=DvMDIEFtyjV", "arxiv_id": "2210.10243", "paper_pdf": "papers/DvMDIEFtyjV.pdf", "paper_pdf_sha256": "0aeeb5936b11bae3bd0baf887738a66d9513aed4b572c5673ea99a2bb15c8a32", "paper_pdf_bytes": 2658059, "paper_pdf_source": "openreview", "code_url": "https://github.com/clutr/clutr", "code_repository": "clutr/clutr", "code_commit": "b461ffdfc937a231b7a5086bb520c1d3f9b88c39", "code_archive": "repos/DvMDIEFtyjV.zip", "code_archive_sha256": "b439267c18b401b50d17c52ce9398cd64fef1d7d459dde1f3f2154579254a28b", "code_archive_bytes": 1849371, "code_file_count": 83, "code_extensions": {".py": 80, ".sh": 3}, "github_disk_usage_kb": 1789, "github_languages": {"Python": 4654068, "Shell": 4210}, "github_archived": false, "github_pushed_at": "2022-12-10T00:23:30Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/clutr-curriculum-learning-via-unsupervised"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "luO6l9cP6b6", "year": 2022, "status": "rejected", "title": "Identifying the Limits of Cross-Domain Knowledge Transfer for Pretrained Models", "authors": ["Zhengxuan Wu", "Nelson F. Liu", "Christopher Potts"], "authorids": ["~Zhengxuan_Wu1", "~Nelson_F._Liu1", "~Christopher_Potts1"], "authors_source": "OpenReview API", "abstract": "There is growing evidence that pretrained language models improve task-specific fine-tuning even where the task examples are radically different from those seen in training. What is the nature of this surprising cross-domain transfer? We offer a partial answer via a systematic exploration of how much transfer occurs when models are denied any information about word identity via random scrambling. In four classification tasks and two sequence labeling tasks, we evaluate LSTMs using GloVe embeddings, BERT, and baseline models. Among these models, we find that only BERT shows high rates of transfer into our scrambled domains, and for classification but not sequence labeling tasks. Our analyses seek to explain why transfer succeeds for some tasks but not others, to isolate the separate contributions of pretraining versus fine-tuning, to show that the fine-tuning process is not merely learning to unscramble the scrambled inputs, and to quantify the role of word frequency. These findings help explain where and why cross-domain transfer occurs, which can guide future studies and practical fine-tuning efforts.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "tABD54QKiT", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper859/Reviewer_vdgZ"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "3: reject, not good enough", "soundness": "", "presentation": "", "contribution": "", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "summary": "This work aims to identify the source of transfer learning on neural models.  To this end they set up a series of experiments where models are trained and tested with english data, and, with scrambled (or randomly replaced) data, keeping token-wise sentence length.\nScrambled sentences also use replacement tokens keeping token frequency with the original token.\nModels are tested on standard benchmarks on text classification and sequence modelling, showing a drop in performance when the input is scrambled, and a similar drop with randomized input.\nFurther analysis are meant to test of the scrambling maintains sentence semantics (it doesn't'), if BERT is just retraining (there is some transfer but inconclusive), keeping BERT frozen (it depends on the layer) and finally identifying word identity reassociation (it seems there is not reassociation).\n", "review_text": "\nThe overall presentation is regular. I think authors fail to properly state all the possible explanations to support the existence of transfer learning with scrambled or random input, \nIf transfer learning were the explanation of why scrambled input keeps showing transfer learning capabilities an experiment training with the original data while tested on scrambled or vice versa is needed. Why didn't the authors try that configuration?\nThe analysis section can be largely improved. Three out of four experiments show inconclusive results. If results cannot be interpreted, authors should find other experiments to support (or refute) their claims.\nDespite the interesting results and large number of experiments, the inconclusive results make unclear the scientific contribution of this research piece.\n\nSome details:\n\n_Paper's title is somehow misleading \"Cross-Domain Knowledge Transfer for Pretrained Models\". Authors use just BERT as pretrained model and a LSTM with GloVe embeddings as part of the experimentation.\n\n\n_ The introduction and the Conclusions are inconsistent:\n\"we evaluate LSTMs using GloVe embeddings, BERT, and baseline models\" vs \"we take an English pretrained BERT off-the-shelf and fine-tune it with a scrambled English dataset\"\n\n\n_table 4 is shown on page 6, but referenced on page 8. Please bring the table closer to its reference.\n\n_ \"Our leading hypothesis here is that the LSTMs may actually relearn all weights without taking advantage of pretraining\". \n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This work aims to identify the source of transfer learning on neural models.  To this end they set up a series of experiments where models are trained and tested with english data, and, with scrambled (or randomly replaced) data, keeping token-wise sentence length.\nScrambled sentences also use replacement tokens keeping token frequency with the original token.\nModels are tested on standard benchmarks on text classification and sequence modelling, showing a drop in performance when the input is scrambled, and a similar drop with randomized input.\nFurther analysis are meant to test of the scrambling maintains sentence semantics (it doesn't'), if BERT is just retraining (there is some transfer but inconclusive), keeping BERT frozen (it depends on the layer) and finally identifying word identity reassociation (it seems there is not reassociation).\n", "main_review": "\nThe overall presentation is regular. I think authors fail to properly state all the possible explanations to support the existence of transfer learning with scrambled or random input, \nIf transfer learning were the explanation of why scrambled input keeps showing transfer learning capabilities an experiment training with the original data while tested on scrambled or vice versa is needed. Why didn't the authors try that configuration?\nThe analysis section can be largely improved. Three out of four experiments show inconclusive results. If results cannot be interpreted, authors should find other experiments to support (or refute) their claims.\nDespite the interesting results and large number of experiments, the inconclusive results make unclear the scientific contribution of this research piece.\n\nSome details:\n\n_Paper's title is somehow misleading \"Cross-Domain Knowledge Transfer for Pretrained Models\". Authors use just BERT as pretrained model and a LSTM with GloVe embeddings as part of the experimentation.\n\n\n_ The introduction and the Conclusions are inconsistent:\n\"we evaluate LSTMs using GloVe embeddings, BERT, and baseline models\" vs \"we take an English pretrained BERT off-the-shelf and fine-tune it with a scrambled English dataset\"\n\n\n_table 4 is shown on page 6, but referenced on page 8. Please bring the table closer to its reference.\n\n_ \"Our leading hypothesis here is that the LSTMs may actually relearn all weights without taking advantage of pretraining\". \n", "summary_of_the_review": "Finding are unclear\nThe paper is difficult to follow\nWriting can be improved\nExperiments are inconclusive.\n", "correctness": "2: Several of the paper’s claims are incorrect or not well-supported.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1636075232289}, {"id": "i0K0dUeqqZr", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper859/Reviewer_4jKu"], "rating": "", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper studies the generalization of several architectures (BOW, GloVe, LSTM, BERT) under scrambling of vocabularies - with and without frequency constraints. ", "review_text": "The paper is well-written and does a reasonable job evaluating the robustness of architectures under an artificial form of vocabulary shift. However, the authors fail to convince me why studying this distributional shift would bring new and relevant insights to the table. I think for two reasons: \n\n(a) The shift is not interesting in the sense that it is related to cross-domain or cross-lingual or cross-time shifts, or any other shifts that we observe in the wild. \n(b) The shift does not seem to add value over other synthetic shifts that have already been studied, e.g., word-level shuffling, character-level shuffling, etc. \n\nYou could have studied many other scrambling functions, so why this one? That said, I liked §7.3, which nicely baselines the scrambled model performance. §7.5 is also nice, except that it’s hard to see why any network would learn to reassociate words in the context of learning a particular task. (Learning to reassociate words is not necessary to solve most tasks. Maybe in machine translation?)\n\nWriting: The paper is overall well-written, but I found the use of the word ‘scrambling’ for a left-to-right in-order replacement of words, misleading. In related work, scrambling refers to scrambling of word order. I would also have liked to see a stronger motivation for focusing on frequency. \n\nOther weaknesses: \n\n(i) BERT is old now. We have already seen a lot of studies of BERT (in the so-called ‘Bertology’-literature), but isn’t it time to move on? I’m a little worried about community-wide overfitting of our intuitions about language models. There’s the popular alternatives (RoBERTa, GPT-2, t5, etc.), but also a lot of faster, fairer, more interpretable alternatives. \n\n(ii) The two sequential labelling tasks are very similar. POS is a give-away of NER. How about finegrained sentiment, coreference resolution, semantic tagging, grammatical error detection, etc.?\n\n(iii) As shown in related work, the GLUE tasks are relatively easy (e.g., solvable without word order information). Maybe also consider harder sequence classification problems?\n\n(iv) Using only a small subset of GLUE tasks looks like cherry-picking. \n\n(v) It is unclear how sensitive the results for each architecture are to hyper-parameter changes. ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "The paper studies the generalization of several architectures (BOW, GloVe, LSTM, BERT) under scrambling of vocabularies - with and without frequency constraints. ", "main_review": "The paper is well-written and does a reasonable job evaluating the robustness of architectures under an artificial form of vocabulary shift. However, the authors fail to convince me why studying this distributional shift would bring new and relevant insights to the table. I think for two reasons: \n\n(a) The shift is not interesting in the sense that it is related to cross-domain or cross-lingual or cross-time shifts, or any other shifts that we observe in the wild. \n(b) The shift does not seem to add value over other synthetic shifts that have already been studied, e.g., word-level shuffling, character-level shuffling, etc. \n\nYou could have studied many other scrambling functions, so why this one? That said, I liked §7.3, which nicely baselines the scrambled model performance. §7.5 is also nice, except that it’s hard to see why any network would learn to reassociate words in the context of learning a particular task. (Learning to reassociate words is not necessary to solve most tasks. Maybe in machine translation?)\n\nWriting: The paper is overall well-written, but I found the use of the word ‘scrambling’ for a left-to-right in-order replacement of words, misleading. In related work, scrambling refers to scrambling of word order. I would also have liked to see a stronger motivation for focusing on frequency. \n\nOther weaknesses: \n\n(i) BERT is old now. We have already seen a lot of studies of BERT (in the so-called ‘Bertology’-literature), but isn’t it time to move on? I’m a little worried about community-wide overfitting of our intuitions about language models. There’s the popular alternatives (RoBERTa, GPT-2, t5, etc.), but also a lot of faster, fairer, more interpretable alternatives. \n\n(ii) The two sequential labelling tasks are very similar. POS is a give-away of NER. How about finegrained sentiment, coreference resolution, semantic tagging, grammatical error detection, etc.?\n\n(iii) As shown in related work, the GLUE tasks are relatively easy (e.g., solvable without word order information). Maybe also consider harder sequence classification problems?\n\n(iv) Using only a small subset of GLUE tasks looks like cherry-picking. \n\n(v) It is unclear how sensitive the results for each architecture are to hyper-parameter changes. ", "summary_of_the_review": "The paper is well-written, but poorly motivated. I simply do not see what we learn from the paper (in the intersection of 'previously unknown' and 'relevant'). In addition, while the experiments are well-designed, the datasets seem cherry-picked, and the models out-dated. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "empirical_novelty_and_significance": "1: The contributions are neither significant nor novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold", "confidence": "5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully."}, "tcdate": 1635760315099}, {"id": "XlpgD1isw2U", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper859/Reviewer_dnE4"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "This paper elaborates on transfer learning and domain adaptation in language models. The\nauthors argue that there are limits on how much information could be transferred when a\nmodel is scrambled or somehow randomized. They used different strategies to randomize\ntraining data including frequency matching and completely random replacement. They suggest\nthat only Bert shows high rates of transfer into their scrambled domains in classification tasks.\nTheir experimental results exhibit importance of pretraining on sequence labeling tasks where\nrandomness occurs. For example, they showed that identity of words are important and when\nthey are swapped with completely different frequencies (random) the performance drops\nsignificantly. They also supported the view of Sinha et al. (2021) and Ethayarajh (2019) that BERT\nmodel may preserves frequency better and this is the reason behind its superiority.", "review_text": "Overall, I think the paper is well-written and debates interesting points in transferability.\nHowever, I want to see some measures of intrinsic evaluation related to the scrambled datasets\nas well (e.g., perplexity). My concern is that these scrambling techniques might push your\ndataset outside of real word adversarial attacks. I would suggest comparing them all together\nto support validity of the scrambling methods. Furthermore, could you every tried replacing\nword using ontology or a collection of semantically related words? Please add similar\nexperiment to show that scrambling with synonyms and antonym have different effects (or\npositive vs negative replacements).", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper elaborates on transfer learning and domain adaptation in language models. The\nauthors argue that there are limits on how much information could be transferred when a\nmodel is scrambled or somehow randomized. They used different strategies to randomize\ntraining data including frequency matching and completely random replacement. They suggest\nthat only Bert shows high rates of transfer into their scrambled domains in classification tasks.\nTheir experimental results exhibit importance of pretraining on sequence labeling tasks where\nrandomness occurs. For example, they showed that identity of words are important and when\nthey are swapped with completely different frequencies (random) the performance drops\nsignificantly. They also supported the view of Sinha et al. (2021) and Ethayarajh (2019) that BERT\nmodel may preserves frequency better and this is the reason behind its superiority.", "main_review": "Overall, I think the paper is well-written and debates interesting points in transferability.\nHowever, I want to see some measures of intrinsic evaluation related to the scrambled datasets\nas well (e.g., perplexity). My concern is that these scrambling techniques might push your\ndataset outside of real word adversarial attacks. I would suggest comparing them all together\nto support validity of the scrambling methods. Furthermore, could you every tried replacing\nword using ontology or a collection of semantically related words? Please add similar\nexperiment to show that scrambling with synonyms and antonym have different effects (or\npositive vs negative replacements).", "summary_of_the_review": "I believe regardless  of how good some of transfer learning models perform, we need to elaborate on situations those models collapse. Therefore I think having similar papers truly exhibit weakness and strengths of these models is necessary. ", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work."}, "tcdate": 1635748117244}, {"id": "v0MV94nFjfz", "reviewer_signature": ["ICLR.cc/2022/Conference/Paper859/Reviewer_6a8W"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes an evaluation pipeline for pre-trained models by testing their transferability\nwithout word identity information. Specifically, they take an English pre-trained BERT off-the-shelf\nand fine-tune it with a corrupted English dataset. Those corrupted texts are designed to remove word identity information while preserving the word frequency information. They conduct experiments on 6 tasks.", "review_text": "Strength:\n1. the paper is mostly well-written and easy to follow.\n2. the experiments are well-designed to support the claims\n\nWeakness:\n2. I'm not convinced by the results about BERT in S7.3. The confounder factor -- pre-training, is not considered. For instance, for the pre-trained BERT, the word frequencies information is already captured in the parameters (along with many other features that also hold for scrambled text). Such that when fine-tuned on scrambled data (word frequency preserved), it's not surprising that pre-trained BERT can clearly perform better. And the reason why LSTM is more robust is that the model size is very small such that it can relearn all weights. We can observe the same trend for 1-layer-BERT on classification tasks, the gaps are generally smaller than other model variants. Some claims in the paper can be further substantiated if the author can experiment with a BERT that is pre-trained on scrambled English.\n\n3. I also have some doubts about S6.2, in which the author claim that the bad performance is because sequence labeling tasks are more likely to rely on word identities. However, as explored in Hewitt & Liang, those labeling supervisions (although in POS tagging) can be easily learned by fine-tuning. Can authors find a better explanation for the poor performance on labeling tasks?\n\n1. The method proposed in this paper is widely used in many other studies, e.g., in adversarial attacking. But it is still somewhat novel since it is applied to new domains. Also, the observations made are marginally novel or significant, there are not many new results compared to prior observations. The title is about pre-trained models, but only BERT is discussed throughout the paper.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"summary_of_the_paper": "This paper proposes an evaluation pipeline for pre-trained models by testing their transferability\nwithout word identity information. Specifically, they take an English pre-trained BERT off-the-shelf\nand fine-tune it with a corrupted English dataset. Those corrupted texts are designed to remove word identity information while preserving the word frequency information. They conduct experiments on 6 tasks.", "main_review": "Strength:\n1. the paper is mostly well-written and easy to follow.\n2. the experiments are well-designed to support the claims\n\nWeakness:\n2. I'm not convinced by the results about BERT in S7.3. The confounder factor -- pre-training, is not considered. For instance, for the pre-trained BERT, the word frequencies information is already captured in the parameters (along with many other features that also hold for scrambled text). Such that when fine-tuned on scrambled data (word frequency preserved), it's not surprising that pre-trained BERT can clearly perform better. And the reason why LSTM is more robust is that the model size is very small such that it can relearn all weights. We can observe the same trend for 1-layer-BERT on classification tasks, the gaps are generally smaller than other model variants. Some claims in the paper can be further substantiated if the author can experiment with a BERT that is pre-trained on scrambled English.\n\n3. I also have some doubts about S6.2, in which the author claim that the bad performance is because sequence labeling tasks are more likely to rely on word identities. However, as explored in Hewitt & Liang, those labeling supervisions (although in POS tagging) can be easily learned by fine-tuning. Can authors find a better explanation for the poor performance on labeling tasks?\n\n1. The method proposed in this paper is widely used in many other studies, e.g., in adversarial attacking. But it is still somewhat novel since it is applied to new domains. Also, the observations made are marginally novel or significant, there are not many new results compared to prior observations. The title is about pre-trained models, but only BERT is discussed throughout the paper.", "summary_of_the_review": "This paper presents some interesting observations, but the methods or findings are only marginally significant or novel.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "N/A", "recommendation": "5: marginally below the acceptance threshold", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked."}, "tcdate": 1634891087234}], "openreview_url": "https://openreview.net/forum?id=luO6l9cP6b6", "arxiv_id": "2104.08410", "paper_pdf": "papers/luO6l9cP6b6.pdf", "paper_pdf_sha256": "220871086cc7ab4dc0b183640c4e3d76dd90b87f380076f3a8a96029058fea71", "paper_pdf_bytes": 7142459, "paper_pdf_source": "openreview", "code_url": "https://github.com/frankaging/limits-cross-domain-transfer", "code_repository": "frankaging/limits-cross-domain-transfer", "code_commit": "318d5154f2391006709a16ab47fae3fa896c42fc", "code_archive": "repos/luO6l9cP6b6.zip", "code_archive_sha256": "c961790df54567a452fbdbaf6d9d6cb4d7eda53a34a63c20b1c787878a83c153", "code_archive_bytes": 1297450, "code_file_count": 26, "code_extensions": {".ipynb": 14, ".py": 11, ".sh": 1}, "github_disk_usage_kb": 7788, "github_languages": {"Jupyter Notebook": 2070243, "Python": 227190, "Shell": 701}, "github_archived": false, "github_pushed_at": "2021-11-12T06:31:52Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/identifying-the-limits-of-cross-domain"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "ZvvxYyjfvZc", "year": 2021, "status": "rejected", "title": "Correcting Momentum in Temporal Difference Learning", "authors": ["Emmanuel Bengio", "Joelle Pineau", "Doina Precup"], "authorids": ["~Emmanuel_Bengio1", "~Joelle_Pineau1", "~Doina_Precup1"], "authors_source": "OpenReview API", "abstract": "A common optimization tool used in deep reinforcement learning is momentum, which consists in accumulating and discounting past gradients, reapplying them at each iteration. We argue that, unlike in supervised learning, momentum in Temporal Difference (TD) learning accumulates gradients that become doubly stale: not only does the gradient of the loss change due to parameter updates, the loss itself changes due to bootstrapping. We first show that this phenomenon exists, and then propose a first-order correction term to momentum. We show that this correction term improves sample efficiency in policy evaluation by correcting target value drift. An important insight of this work is that deep RL methods are not always best served by directly importing techniques from the supervised setting.", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "AXq1l9-agl6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1690/AnonReviewer3"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "Summary:\n\nThis paper extends the idea of momentum, commonly used in optimization literature, to Temporal Difference (TD) learning which is a widely used algorithm for policy evaluation in Reinforcement learning literature. The main challenge in this work is to account for the 'optimization bias' introduced by using momentum based updates. The authors propose to do this via a Taylor approximation to the TD update and empirically show the merits of this idea on some toy data sets as well as on an Atari game. \n\nReason for score: \n\nThe idea of using momentum for TD learning seems quite interesting. However, in my opinion, the paper seems to a somewhat incremental contribution, given the work of Sun et. al. 2020, which gives a theoretical analysis of momentum based updates for TD with linear function approximation. \n\n1) The bias correction via Taylor approximation of the gradients seems new to me. The authors claim that the bias doesn't play a significant role in supervised learning tasks. Can the authors provide some more citations to support this besides the 1-d regression task. For TD learning, the authors don't present results for bias corrections with second order terms. How do we know that we can safely ignore these for large networks? On the other hand, incorporating these terms poses a computational challenge. \n\n2) Prior work of Sun et. al. 2020, which proposes a momentum for a linear function approximation, uses projection to bound the iterates. Is the momentum update even stable for deep neural networks for large problems, without some additional tricks? The authors only test toy examples (the Atari example has some combination of Adam and momentum which makes it ).  \n\n3) For the comparison on MsPacman, why is a combination with Adam required? The authors claim it is for speeding up learning. In this case, maybe the authors can have the MSE plot for just Adam updates for a better comparison? \n\n4) Many of the plots are truncated at a few thousand steps. What happens after that? It might be helpful to visualize the loss at limit points that different updates converge towards (roughly). Graphs with more steps might be more useful.\n\n5) The authors claim their method improves sample complexity as compared to say the more widely used trick of freezing network. But Figure 17 shows that models with frozen targets eventually become more precise. This might be a reason why this idea may not perform well in the control setting, as Figure 17 suggests that there is probably a residual bias in all momentum based updates (when used with neural networks). \n\nOverall, I am unsure whether the contributions here are strong enough to justify a paper publication.\n\n--------- Update after author response ---------\n\nI thank the authors for their response. I have updated my score. I am slightly leaning toward acceptance though I think the paper might benefit from a revision based on some of the points raised by other reviewers, including comments about motivation for correcting the bias as well as the scalability (the proposed method requires storing n^2 additional values which is expensive for large networks). ", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Correcting momentum in Temporal Difference learning", "review": "Summary:\n\nThis paper extends the idea of momentum, commonly used in optimization literature, to Temporal Difference (TD) learning which is a widely used algorithm for policy evaluation in Reinforcement learning literature. The main challenge in this work is to account for the 'optimization bias' introduced by using momentum based updates. The authors propose to do this via a Taylor approximation to the TD update and empirically show the merits of this idea on some toy data sets as well as on an Atari game. \n\nReason for score: \n\nThe idea of using momentum for TD learning seems quite interesting. However, in my opinion, the paper seems to a somewhat incremental contribution, given the work of Sun et. al. 2020, which gives a theoretical analysis of momentum based updates for TD with linear function approximation. \n\n1) The bias correction via Taylor approximation of the gradients seems new to me. The authors claim that the bias doesn't play a significant role in supervised learning tasks. Can the authors provide some more citations to support this besides the 1-d regression task. For TD learning, the authors don't present results for bias corrections with second order terms. How do we know that we can safely ignore these for large networks? On the other hand, incorporating these terms poses a computational challenge. \n\n2) Prior work of Sun et. al. 2020, which proposes a momentum for a linear function approximation, uses projection to bound the iterates. Is the momentum update even stable for deep neural networks for large problems, without some additional tricks? The authors only test toy examples (the Atari example has some combination of Adam and momentum which makes it ).  \n\n3) For the comparison on MsPacman, why is a combination with Adam required? The authors claim it is for speeding up learning. In this case, maybe the authors can have the MSE plot for just Adam updates for a better comparison? \n\n4) Many of the plots are truncated at a few thousand steps. What happens after that? It might be helpful to visualize the loss at limit points that different updates converge towards (roughly). Graphs with more steps might be more useful.\n\n5) The authors claim their method improves sample complexity as compared to say the more widely used trick of freezing network. But Figure 17 shows that models with frozen targets eventually become more precise. This might be a reason why this idea may not perform well in the control setting, as Figure 17 suggests that there is probably a residual bias in all momentum based updates (when used with neural networks). \n\nOverall, I am unsure whether the contributions here are strong enough to justify a paper publication.\n\n--------- Update after author response ---------\n\nI thank the authors for their response. I have updated my score. I am slightly leaning toward acceptance though I think the paper might benefit from a revision based on some of the points raised by other reviewers, including comments about motivation for correcting the bias as well as the scalability (the proposed method requires storing n^2 additional values which is expensive for large networks). ", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1604202170133}, {"id": "6TFTxdkcdc", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1690/AnonReviewer2"], "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "This paper proposes a modification to momentum. Extra terms are added to account for the drift in parameters through the gradient updates that contribute to the momentum. The effect of this modification is studied on supervised learning and TD learning and on different representations. The conclusion is that while the original momentum is good enough for supervised learning, this modification matters when there is bootstrapping, especially on representations with high interference.\n\nOverall, the empirical study highlights a direction for improving TD learning and the experiments are diverse and conclusive. Although the proposed approximations to the extra terms show limited improvement, experiments show that if the extra terms can be approximated more accurately, there will be a considerable gain. Computing the extra terms exactly is computationally expensive but it is possible that better approximations will be introduced later on. There are a few weaknesses that I describe below.\n\n1. There is little discussion on why we would want to account for the drift in parameters in momentum. Would this somehow reduce variance or improve the rate of convergence?\n\n2. Candidate step-sizes are too few and some differences could simply be due to the match between the chosen step-size and an algorithm. Consider sweeping over a finer grid.\n\n3. Section 3.1 says \"Note that this choice is purely illustrative and that our findings extend to similar simple functions.\" Is there evidence on other simple functions?\n\n4. A discussion on different optimizers somewhere in the paper or the appendix is needed to see how the proposed updates compare to existing work.\n\nMinor comments:\n \n1. Section 1.1, problem formulation and definition of V and Q: In a standard RL setting reward is a function of state and action.\n\n2. \\theta and V_\\theta are not defined. It is better to make it explicit that the function is parameterized by \\theta and outputs V_\\theta which estimates V^\\pi.\n\n---\nUpdate: I have the other reviews and rebuttal. I am still leaning towards acceptance, although I do agree with the other reviews that there is little motivation for the idea. While the experiments show improvement, a discussion on convergence rates or variance could show why one should try this idea in the first place, and if \\mu_* (rather than regular momentum) is indeed an ideal that the algorithm should approximate.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting direction; limited improvement;", "review": "This paper proposes a modification to momentum. Extra terms are added to account for the drift in parameters through the gradient updates that contribute to the momentum. The effect of this modification is studied on supervised learning and TD learning and on different representations. The conclusion is that while the original momentum is good enough for supervised learning, this modification matters when there is bootstrapping, especially on representations with high interference.\n\nOverall, the empirical study highlights a direction for improving TD learning and the experiments are diverse and conclusive. Although the proposed approximations to the extra terms show limited improvement, experiments show that if the extra terms can be approximated more accurately, there will be a considerable gain. Computing the extra terms exactly is computationally expensive but it is possible that better approximations will be introduced later on. There are a few weaknesses that I describe below.\n\n1. There is little discussion on why we would want to account for the drift in parameters in momentum. Would this somehow reduce variance or improve the rate of convergence?\n\n2. Candidate step-sizes are too few and some differences could simply be due to the match between the chosen step-size and an algorithm. Consider sweeping over a finer grid.\n\n3. Section 3.1 says \"Note that this choice is purely illustrative and that our findings extend to similar simple functions.\" Is there evidence on other simple functions?\n\n4. A discussion on different optimizers somewhere in the paper or the appendix is needed to see how the proposed updates compare to existing work.\n\nMinor comments:\n \n1. Section 1.1, problem formulation and definition of V and Q: In a standard RL setting reward is a function of state and action.\n\n2. \\theta and V_\\theta are not defined. It is better to make it explicit that the function is parameterized by \\theta and outputs V_\\theta which estimates V^\\pi.\n\n---\nUpdate: I have the other reviews and rebuttal. I am still leaning towards acceptance, although I do agree with the other reviews that there is little motivation for the idea. While the experiments show improvement, a discussion on convergence rates or variance could show why one should try this idea in the first place, and if \\mu_* (rather than regular momentum) is indeed an ideal that the algorithm should approximate.", "rating": "6: Marginally above acceptance threshold", "confidence": "3: The reviewer is fairly confident that the evaluation is correct"}, "tcdate": 1603860259405}, {"id": "nsdUVsilkB6", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1690/AnonReviewer1"], "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "#### Idea\nThe paper explains and studies an interesting issue with momentum in TD learning, which is its staleness while doing TD updates which is in contrary to the supervised learning. However, the paper could benefit from more investigations.\n\n#### Comments\n- In section 2.1: It is said that to compute the oracle update, one has to recompute the entire sum. This may take a while to compute and it would be good if they have had provided a computation-time or wall-time comparison will prior methods.\n- In section 2.1: To handle the problem of long summation, \"an effective horizon\" is used. However, there's no explanations given behind this particular choice. Is it considered as a geometric distribution and thus this formula? Or if it's not, where does this formula come from?\n- In section 3.2: When the proposed method is applied to the Atari games, the problem of scalability shows itself. As mentioned by the authors: \"Since the architecture required to train ...\", the scalability of this approach is under question. \n- In section 3.2: Regarding this paragraph: \"Note that we do not use frozen targets ...\". The most important question about this work is its scalability, whereas in a prior approaches like frozen targets there is not such an issue while dealing with staleness issue. As in experiments, authors have not used frozen targets, which seems to be misleading when it comes to the experiments and results. It would be useful to see a comparison between these two approaches since they try to solve a similar problem.\n\n#### Minor issues\n- In section 3.1: \"We task a ...\" needs to be changed to \"We take a ...\"\n\n- In section 2.2: It would be better if you could give a more in depth explanation for eq 11: “Here the term multiplying ∆θ is not exactly the Hessian…”\n\n", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "Interesting idea but needs more investigation", "review": "#### Idea\nThe paper explains and studies an interesting issue with momentum in TD learning, which is its staleness while doing TD updates which is in contrary to the supervised learning. However, the paper could benefit from more investigations.\n\n#### Comments\n- In section 2.1: It is said that to compute the oracle update, one has to recompute the entire sum. This may take a while to compute and it would be good if they have had provided a computation-time or wall-time comparison will prior methods.\n- In section 2.1: To handle the problem of long summation, \"an effective horizon\" is used. However, there's no explanations given behind this particular choice. Is it considered as a geometric distribution and thus this formula? Or if it's not, where does this formula come from?\n- In section 3.2: When the proposed method is applied to the Atari games, the problem of scalability shows itself. As mentioned by the authors: \"Since the architecture required to train ...\", the scalability of this approach is under question. \n- In section 3.2: Regarding this paragraph: \"Note that we do not use frozen targets ...\". The most important question about this work is its scalability, whereas in a prior approaches like frozen targets there is not such an issue while dealing with staleness issue. As in experiments, authors have not used frozen targets, which seems to be misleading when it comes to the experiments and results. It would be useful to see a comparison between these two approaches since they try to solve a similar problem.\n\n#### Minor issues\n- In section 3.1: \"We task a ...\" needs to be changed to \"We take a ...\"\n\n- In section 2.2: It would be better if you could give a more in depth explanation for eq 11: “Here the term multiplying ∆θ is not exactly the Hessian…”\n\n", "rating": "6: Marginally above acceptance threshold", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603278991299}, {"id": "oqa_KHZ3qKR", "reviewer_signature": ["ICLR.cc/2021/Conference/Paper1690/AnonReviewer4"], "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct", "recommendation": "", "soundness": "", "presentation": "", "contribution": "", "correctness": "", "summary": "", "review_text": "The paper studies the approach to 'correct the bias' of momentum in supervised learning and reinforcement learning (especially, in the temporal-difference algorithm). The proposed approach can maintain a moving average of the 'correct' gradient direction over past objective functions via extra updates. And the experiments show that it does speed up the convergence. (ps: I did not find the comparisons over the loss functions at the stationary points).\n\nAs we know TD learning uses a semi/partial/biased gradient with respect to the mean square Bellman error. Thus, the effect of applying momentum to training is an open question. I feel very glad to see such a paper studying it and adding a reasonable modification. The approach averages the TD update over different steps via a modified momentum update. Such an approach reduces the variance, which is usually very large in TD, but also uses additional computational power.\n\nI found the motivation not very convincing for me. The paper seems to claim that a good momentum should carry the information of past objective function and calculate their gradients only with respect to the parameter in the current iterate. However, momentum terms do have the reason to use the gradient wrt previous parameters, as it can speed up the convergence in many situations, e.g. ravines. Another concern is that TD is not typically an online learning problem. The objective function in the early phase is not informative to the update in the later phases as the values of the next states are inaccurate. Therefore, I feel the theoretical reason for 'correcting' the 'bias', which is defined by this paper, is not warranted very well. \n\nI suggest the paper to rigorously discuss more on why such 'bias' has to be corrected. Alternatively, the proposed method is designed simply for the purpose of reducing the variance of the stochastic gradient for the case of minibatch update. But the paper needs to argue how it is better than using a larger batch or compare itself with other variance reduction approaches.", "strengths": "", "weaknesses": "", "questions": "", "all_content": {"title": "the motivation is not very convincing", "review": "The paper studies the approach to 'correct the bias' of momentum in supervised learning and reinforcement learning (especially, in the temporal-difference algorithm). The proposed approach can maintain a moving average of the 'correct' gradient direction over past objective functions via extra updates. And the experiments show that it does speed up the convergence. (ps: I did not find the comparisons over the loss functions at the stationary points).\n\nAs we know TD learning uses a semi/partial/biased gradient with respect to the mean square Bellman error. Thus, the effect of applying momentum to training is an open question. I feel very glad to see such a paper studying it and adding a reasonable modification. The approach averages the TD update over different steps via a modified momentum update. Such an approach reduces the variance, which is usually very large in TD, but also uses additional computational power.\n\nI found the motivation not very convincing for me. The paper seems to claim that a good momentum should carry the information of past objective function and calculate their gradients only with respect to the parameter in the current iterate. However, momentum terms do have the reason to use the gradient wrt previous parameters, as it can speed up the convergence in many situations, e.g. ravines. Another concern is that TD is not typically an online learning problem. The objective function in the early phase is not informative to the update in the later phases as the values of the next states are inaccurate. Therefore, I feel the theoretical reason for 'correcting' the 'bias', which is defined by this paper, is not warranted very well. \n\nI suggest the paper to rigorously discuss more on why such 'bias' has to be corrected. Alternatively, the proposed method is designed simply for the purpose of reducing the variance of the stochastic gradient for the case of minibatch update. But the paper needs to argue how it is better than using a larger batch or compare itself with other variance reduction approaches.", "rating": "4: Ok but not good enough - rejection", "confidence": "4: The reviewer is confident but not absolutely certain that the evaluation is correct"}, "tcdate": 1603231806854}], "openreview_url": "https://openreview.net/forum?id=ZvvxYyjfvZc", "arxiv_id": "2106.03955", "paper_pdf": "papers/ZvvxYyjfvZc.pdf", "paper_pdf_sha256": "4cfc3a5841809cfedf0d3e1104b1dc0fd87e27f91b738850fbe648a5039acb3a", "paper_pdf_bytes": 709918, "paper_pdf_source": "openreview", "code_url": "https://github.com/bengioe/staleness-corrected-momentum", "code_repository": "bengioe/staleness-corrected-momentum", "code_commit": "ecdc5a3e49c009471c6cac8af8ebd58ed908115c", "code_archive": "repos/ZvvxYyjfvZc.zip", "code_archive_sha256": "83d6a5d81ee47251e9ae49ec17beab073078da29c710bc5bf98abe77bd151831", "code_archive_bytes": 5815991, "code_file_count": 12, "code_extensions": {".py": 12}, "github_disk_usage_kb": 5680, "github_languages": {"Python": 112910}, "github_archived": false, "github_pushed_at": "2021-06-28T17:58:24Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/correcting-momentum-in-temporal-difference-1"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "aYx7JR20sI", "year": 2025, "status": "rejected", "title": "Tropical Expressivity of Neural Networks", "authors": ["Paul Lezeau", "Thomas Walker", "Yueqi Cao", "Shiv Bhatia", "Anthea Monod"], "authorids": ["~Paul_Lezeau1", "~Thomas_Walker2", "~Yueqi_Cao1", "~Shiv_Bhatia1", "~Anthea_Monod1"], "authors_source": "OpenReview API", "abstract": "We propose an algebraic geometric framework to study the expressivity of piecewise linear activation neural networks.  A particular quantity of neural networks that has been actively studied is the number of linear regions, which gives a quantification of the information capacity of the architecture.  To study and evaluate information capacity and expressivity, we work in the setting of tropical geometry---a combinatorial and polyhedral variant of algebraic geometry---where there are known connections between tropical rational maps and feedforward neural networks. Our work builds on and expands this connection to capitalize on the rich theory of tropical geometry to characterize and study various architectural aspects of neural networks. Our contributions are threefold: we provide a novel tropical geometric approach to selecting sampling domains among linear regions; an algebraic result allowing for a guided restriction of the sampling domain for network architectures with symmetries; and a new open source OSCAR library to analyze neural networks symbolically using their tropical representations, where we present a new algorithm that computes the exact number of their linear regions. We provide a comprehensive set of proof-of-concept numerical experiments demonstrating the breadth of neural network architectures to which tropical geometric theory can be applied to reveal insights on expressivity characteristics of a network.  Our work provides the foundations for the adaptation of both theory and existing software from computational tropical geometry and symbolic computation to neural networks and deep learning.", "decision": "Reject", "meta_review": "", "num_reviews": 4, "reviews": [{"id": "zIWeR57dlY", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12294/Reviewer_gsPs"], "rating": 6, "soundness": 3, "presentation": 3, "contribution": 3, "confidence": 2, "summary": "This paper presents a novel approach to investigating the expressivity of neural networks through the framework of tropical geometry, a branch of algebraic geometry that employs combinatorial and polyhedral methods, with key contributions including a new framework for selecting bounded sampling domains using the Hoffman constant, a method for enhancing computational efficiency through symmetry in network architecture, and a Julia library integrated into the OSCAR system for symbolic representation and analysis of neural networks using tropical techniques. The paper builds on foundational work that connects tropical rational maps with feedforward neural networks, providing theoretical insights supported by experiments that demonstrate its applicability across various network architectures. Overall, I believe this paper makes a significant contribution to the theoretical foundations of deep learning.", "review_text": "This paper presents a novel approach to investigating the expressivity of neural networks through the framework of tropical geometry, a branch of algebraic geometry that employs combinatorial and polyhedral methods, with key contributions including a new framework for selecting bounded sampling domains using the Hoffman constant, a method for enhancing computational efficiency through symmetry in network architecture, and a Julia library integrated into the OSCAR system for symbolic representation and analysis of neural networks using tropical techniques. The paper builds on foundational work that connects tropical rational maps with feedforward neural networks, providing theoretical insights supported by experiments that demonstrate its applicability across various network architectures. Overall, I believe this paper makes a significant contribution to the theoretical foundations of deep learning.", "strengths": "**Clarity:** The paper is well-organized, exhibiting a clear and logical progression from the theoretical background to the main results and their implications. Each section is thoughtfully introduced with specific motivations that enhance the reader's understanding. Notably, the three contributions are effectively aligned with their corresponding sections, reinforcing the connections among the presented ideas.\n\n**Novelty and Significance:** The innovative application of tropical geometry to evaluate the expressivity of neural networks introduces valuable insights. By translating neural networks into tropical representations, the authors utilize tools from an underexplored area of deep learning theory. Their approach, which provides an exact count of linear regions and introduces monomial complexity as a measure of expressivity, has the potential to inspire future research on model architecture and optimization.\n\n\n**Practical Implementation:** The integration into OSCAR allows for symbolic manipulation of neural networks, which has practical implications for further research and applications.", "weaknesses": "**Limited Empirical Validation:** The experimental demonstrations, while useful as proofs of concept, are mainly limited to small-scale networks.\n\n**Context within Existing Methods:** A more in-depth discussion on how this tropical approach compares to or enhances traditional methods of region counting and expressivity analysis would provide valuable context.\n\n**Notation Clarity:** Some notations in the main paper are not defined in a timely manner. For example, $P(A,b)$ is defined in Line 673 of the appendix but is used in Line 146 of Section 2 without prior explanation.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper presents a novel approach to investigating the expressivity of neural networks through the framework of tropical geometry, a branch of algebraic geometry that employs combinatorial and polyhedral methods, with key contributions including a new framework for selecting bounded sampling domains using the Hoffman constant, a method for enhancing computational efficiency through symmetry in network architecture, and a Julia library integrated into the OSCAR system for symbolic representation and analysis of neural networks using tropical techniques. The paper builds on foundational work that connects tropical rational maps with feedforward neural networks, providing theoretical insights supported by experiments that demonstrate its applicability across various network architectures. Overall, I believe this paper makes a significant contribution to the theoretical foundations of deep learning.", "soundness": 3, "presentation": 3, "contribution": 3, "strengths": "**Clarity:** The paper is well-organized, exhibiting a clear and logical progression from the theoretical background to the main results and their implications. Each section is thoughtfully introduced with specific motivations that enhance the reader's understanding. Notably, the three contributions are effectively aligned with their corresponding sections, reinforcing the connections among the presented ideas.\n\n**Novelty and Significance:** The innovative application of tropical geometry to evaluate the expressivity of neural networks introduces valuable insights. By translating neural networks into tropical representations, the authors utilize tools from an underexplored area of deep learning theory. Their approach, which provides an exact count of linear regions and introduces monomial complexity as a measure of expressivity, has the potential to inspire future research on model architecture and optimization.\n\n\n**Practical Implementation:** The integration into OSCAR allows for symbolic manipulation of neural networks, which has practical implications for further research and applications.", "weaknesses": "**Limited Empirical Validation:** The experimental demonstrations, while useful as proofs of concept, are mainly limited to small-scale networks.\n\n**Context within Existing Methods:** A more in-depth discussion on how this tropical approach compares to or enhances traditional methods of region counting and expressivity analysis would provide valuable context.\n\n**Notation Clarity:** Some notations in the main paper are not defined in a timely manner. For example, $P(A,b)$ is defined in Line 673 of the appendix but is used in Line 146 of Section 2 without prior explanation.", "questions": "See Weakness.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 6, "confidence": 2, "code_of_conduct": "Yes"}, "tcdate": 1730717078966}, {"id": "3kBGcDCl4y", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12294/Reviewer_64tt"], "rating": 3, "soundness": 3, "presentation": 1, "contribution": 2, "confidence": 4, "summary": "This paper studies the expressivity of neural networks through a tropical lens. Through a suite of algebraic tools, the authors characterize the linear regions of a neural network using the Hoffman constant, discuss sampling improvements for invariant neural networks through the fundamental domain, and provide a symbolic framework for computing linear regions of neural networks.", "review_text": "This paper studies the expressivity of neural networks through a tropical lens. Through a suite of algebraic tools, the authors characterize the linear regions of a neural network using the Hoffman constant, discuss sampling improvements for invariant neural networks through the fundamental domain, and provide a symbolic framework for computing linear regions of neural networks.", "strengths": "The paper tackles important problems that can aid our understanding of deep networks. Tropical geometry has been getting increasingly popular in recent years, and has been the tool of choice for geometric analyses using polyhedral geometry. The introduction of the manuscript is crisp and easy to follow, introducing the main subject areas and discussing the contributions clearly. Finally, there is a significant number of contributions.", "weaknesses": "The weaknesses of the work can be summarized in the following axes:\n\n- a severe lack of polish after the introduction,\n- confusing results, mainly due to the number of contributions,\n- it is unclear what the benefits of the contributions are, either at a theoretical or a empirical level.\n\n**Lack of polish**\n\nAll main sections of the paper are introduced very abruptly, and the defining concepts of each of them are delegated to the appendices. This makes the paper extremely difficult to read, and is currently not self-contained without the appendices. Examples of this is the complete omission of a discussion on tropical polynomials, the discussion of Lemma A.19 or Equation 9 without their exposition, and all algorithms being discussed, but never presented, to name a few.\n\nThere are many typos and there is a lack of explanations at multiple points in the paper. An example of this lack are Figures 2 and 3, where, beyond the typos, Figure 2 lacks any real caption and what is portrayed in these figures, what the colors mean, or why there is a different number of lines on Fig 2(a) and Fig 2(b) is not clear.\n\nOverall while the three different contributions are connected, the connection is not emphasized and the presentation really hinders the readability and the flow of the work. By trying to include all of them in the same work, it becomes extremely difficult to present each contribution in a clear and rigorous manner, especially at such a page limit. I believe the work could greatly benefit from either splitting the contributions, or investing a significant effort in crafting a coherent thread that connects the contributions, most likely at a venue with a much higher page limit.\n\n**Confusing results**\n\nIn terms of the Hoffman constant, it seems that the authors compute the constant, then use it to find the radius of a ball that is needed to intersect all the linear regions. However, if I’m understanding things correctly, this gives no estimate on the number of linear regions. Essentially we have to compute the Hoffman constant (which is NP-hard), and then have to sample many points in that radius to get a probabilistic statement about the number of regions given a sampling size, and even then I don’t see any statements of that effect. Most works I’m aware of on counting the number of linear regions do it without having access to a trained network, and are statements about the architecture, not the specific network, so I’m unsure exactly what the main contribution is here.\n\nSection 4.1 is also confusing and the contribution is again unclear. For example, in line 359 it is stated that we can detect when some intersections may be empty. The way this is stated implies that we have a way of checking if intersections are empty, but not a way to find such intersections. That seems to actually be the case, since on 377 a set of indices I is assumed instead of being computed. A similar comment as above can be made, that this works for specific networks and does not characterize the network class. Also, how are we computing these regions? Are we combinatorially testing all combinations?\n\n**Experimental results**\n\nSimply put, the experimental section is very rushed and the results are not convincing (more in the next section). Starting of with the characterization of width versus depth, what exactly is Figure 4 conveying? There is no comparison between depth versus width, for the same number of neurons. The fact that expressivity increases when we increase the number of nodes is not surprising. A minor comment is also that the range of $k$ doesn’t correctly correspond to Figure 4.\n\nFor Figure 5, the trained and untrained networks seem to hardly have any difference in their pruning rate, and the size of the network is so small where no trends could be observed.", "questions": "In terms of questions:\n\n- Early in the introduction the authors mention that expressivity is one of the most important approaches to measure performance. Are the authors aware of any references to support that claim? Usually overly complex networks (i.e., very expressive ones) tend to overfit the data and lead to poor performance, exactly the opposite of what the authors are arguing.\n- Given that the purpose of the paper is to aid our understanding of deep networks, there seems to be a limited intuitive explanation of the methods that are used. For example, what is an intuitive explanation of the Hoffman constant? Section 2.1 makes an attempt at explaining, but actually describes the distance $d(u, P(A, B))$ and not the constant, and does so by restating the equation instead of an intuitive explanation.\n- About the fundamental domain:\n    - It seems that things only work for discrete groups, if so that should be a concrete discussion.\n    - There is mention of groups acting on vector spaces, but no discussion about group representations (or at least how these groups act on the vector spaces) is presented.\n    - The authors state that in the specific example of 3.2, a factorial improvement is made. However, clearly, $\\Delta$ is an infinite set. There is no asymptotic analysis as to how many samples we need from $\\Delta$, so how are you claiming an improvement without a sampling statement when sampling from an infinite set?\n- Line 395 is very imprecise. What does information mean here? The expression and the function both encode the same function.\n- On 423, the authors argue their use of small networks saying that they help highlight intuitive insights that can aid understanding. What are the insights that are gained in doing so? Of the three presented experiments, the first one has no comparison on depth vs width (which is the titular motivation for the section), the second one shows a marginal change which is unclear if it would persist when the layers/nodes increase, and the final one has no takeaways listed.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This paper studies the expressivity of neural networks through a tropical lens. Through a suite of algebraic tools, the authors characterize the linear regions of a neural network using the Hoffman constant, discuss sampling improvements for invariant neural networks through the fundamental domain, and provide a symbolic framework for computing linear regions of neural networks.", "soundness": 3, "presentation": 1, "contribution": 2, "strengths": "The paper tackles important problems that can aid our understanding of deep networks. Tropical geometry has been getting increasingly popular in recent years, and has been the tool of choice for geometric analyses using polyhedral geometry. The introduction of the manuscript is crisp and easy to follow, introducing the main subject areas and discussing the contributions clearly. Finally, there is a significant number of contributions.", "weaknesses": "The weaknesses of the work can be summarized in the following axes:\n\n- a severe lack of polish after the introduction,\n- confusing results, mainly due to the number of contributions,\n- it is unclear what the benefits of the contributions are, either at a theoretical or a empirical level.\n\n**Lack of polish**\n\nAll main sections of the paper are introduced very abruptly, and the defining concepts of each of them are delegated to the appendices. This makes the paper extremely difficult to read, and is currently not self-contained without the appendices. Examples of this is the complete omission of a discussion on tropical polynomials, the discussion of Lemma A.19 or Equation 9 without their exposition, and all algorithms being discussed, but never presented, to name a few.\n\nThere are many typos and there is a lack of explanations at multiple points in the paper. An example of this lack are Figures 2 and 3, where, beyond the typos, Figure 2 lacks any real caption and what is portrayed in these figures, what the colors mean, or why there is a different number of lines on Fig 2(a) and Fig 2(b) is not clear.\n\nOverall while the three different contributions are connected, the connection is not emphasized and the presentation really hinders the readability and the flow of the work. By trying to include all of them in the same work, it becomes extremely difficult to present each contribution in a clear and rigorous manner, especially at such a page limit. I believe the work could greatly benefit from either splitting the contributions, or investing a significant effort in crafting a coherent thread that connects the contributions, most likely at a venue with a much higher page limit.\n\n**Confusing results**\n\nIn terms of the Hoffman constant, it seems that the authors compute the constant, then use it to find the radius of a ball that is needed to intersect all the linear regions. However, if I’m understanding things correctly, this gives no estimate on the number of linear regions. Essentially we have to compute the Hoffman constant (which is NP-hard), and then have to sample many points in that radius to get a probabilistic statement about the number of regions given a sampling size, and even then I don’t see any statements of that effect. Most works I’m aware of on counting the number of linear regions do it without having access to a trained network, and are statements about the architecture, not the specific network, so I’m unsure exactly what the main contribution is here.\n\nSection 4.1 is also confusing and the contribution is again unclear. For example, in line 359 it is stated that we can detect when some intersections may be empty. The way this is stated implies that we have a way of checking if intersections are empty, but not a way to find such intersections. That seems to actually be the case, since on 377 a set of indices I is assumed instead of being computed. A similar comment as above can be made, that this works for specific networks and does not characterize the network class. Also, how are we computing these regions? Are we combinatorially testing all combinations?\n\n**Experimental results**\n\nSimply put, the experimental section is very rushed and the results are not convincing (more in the next section). Starting of with the characterization of width versus depth, what exactly is Figure 4 conveying? There is no comparison between depth versus width, for the same number of neurons. The fact that expressivity increases when we increase the number of nodes is not surprising. A minor comment is also that the range of $k$ doesn’t correctly correspond to Figure 4.\n\nFor Figure 5, the trained and untrained networks seem to hardly have any difference in their pruning rate, and the size of the network is so small where no trends could be observed.", "questions": "In terms of questions:\n\n- Early in the introduction the authors mention that expressivity is one of the most important approaches to measure performance. Are the authors aware of any references to support that claim? Usually overly complex networks (i.e., very expressive ones) tend to overfit the data and lead to poor performance, exactly the opposite of what the authors are arguing.\n- Given that the purpose of the paper is to aid our understanding of deep networks, there seems to be a limited intuitive explanation of the methods that are used. For example, what is an intuitive explanation of the Hoffman constant? Section 2.1 makes an attempt at explaining, but actually describes the distance $d(u, P(A, B))$ and not the constant, and does so by restating the equation instead of an intuitive explanation.\n- About the fundamental domain:\n    - It seems that things only work for discrete groups, if so that should be a concrete discussion.\n    - There is mention of groups acting on vector spaces, but no discussion about group representations (or at least how these groups act on the vector spaces) is presented.\n    - The authors state that in the specific example of 3.2, a factorial improvement is made. However, clearly, $\\Delta$ is an infinite set. There is no asymptotic analysis as to how many samples we need from $\\Delta$, so how are you claiming an improvement without a sampling statement when sampling from an infinite set?\n- Line 395 is very imprecise. What does information mean here? The expression and the function both encode the same function.\n- On 423, the authors argue their use of small networks saying that they help highlight intuitive insights that can aid understanding. What are the insights that are gained in doing so? Of the three presented experiments, the first one has no comparison on depth vs width (which is the titular motivation for the section), the second one shows a marginal change which is unclear if it would persist when the layers/nodes increase, and the final one has no takeaways listed.", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730646148286}, {"id": "5RdRU1ESXn", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12294/Reviewer_5XwZ"], "rating": 3, "soundness": 3, "presentation": 2, "contribution": 2, "confidence": 4, "summary": "The paper investigates counting the linear regions of neural networks through tropical geometry. The authors relate the Hoffman constant of a tropical rational map to the minimal effective radius, which is the smallest radius $r$ such that a ball of radius $r$ intersects all linear regions. They further show that for an invariant network (with respect to a group $G$), one can sample from the fundamental domain of $G$ to obtain an upper bound on the number of linear regions. Lastly, the authors introduce an algorithm using symbolic representation to compute the number of linear regions and propose a new measure of complexity for ReLU neural networks, termed the “monomial count.”", "review_text": "The paper investigates counting the linear regions of neural networks through tropical geometry. The authors relate the Hoffman constant of a tropical rational map to the minimal effective radius, which is the smallest radius $r$ such that a ball of radius $r$ intersects all linear regions. They further show that for an invariant network (with respect to a group $G$), one can sample from the fundamental domain of $G$ to obtain an upper bound on the number of linear regions. Lastly, the authors introduce an algorithm using symbolic representation to compute the number of linear regions and propose a new measure of complexity for ReLU neural networks, termed the “monomial count.”", "strengths": "The number of linear regions is a standard measure of complexity for neural networks with piecewise linear activation functions, and the authors provide new insights into computing this count for a given network. Specifically, they relate the minimal effective radius to the Hoffman constant of a tropical Puiseux rational map. Additionally, for invariant neural networks, their observation allows a significant reduction in the number of samples required to estimate the number of linear regions.", "weaknesses": "Overall, the contributions of the paper do not appear substantial enough, leading me to recommend rejection. The theoretical depth is limited, and I see few immediate practical benefits from the results, see also the questions below.\n-  While relating the Hoffman constant to the radius of the ball may have theoretical appeal, I do not see its practical benefit, as it is NP-hard to compute the Hoffman constant. The appendix discusses computing the Hoffman constant of a matrix, but there is no discussion about the Hoffman constant for a rational Puiseux function, which is what relates to the radius of the ball. \n- The result concerning sampling from the fundamental domain of the group action for an invariant network appears to be a straightforward observation. \n- The algorithm for determining linear regions using Puiseux polynomials does not seem to offer clear advantages over existing methods. The paper states, “Our tools consider the full input domain of these neural networks and provide an exact geometric characterization of the linear regions. This now makes previously inaccessible avenues available for analyzing the geometry of linear regions of networks (see Appendix I).\" However, algorithms for obtaining the entire polyhedral complex of neural networks already exist (e.g., Masden, 2022, https://arxiv.org/abs/2207.07696).", "questions": "- What is the benefit of relating the Hoffman constant to the radius of the ball? Are there any approximations for the Hoffman constant of a rational Puiseux function? Has the Hoffman constant been applied to Puiseux polynomials (and/or rational functions) in prior work?\n\n- Assuming we know the Hoffman constant, are there theoretical guarantees that using it would lead to better sampling to estimate the number of linear regions?\n\n- Is the definition of the Hoffman constant for a Puiseux rational map well-defined (Definition 2.2)? How do you deal with the fact that there can be infinitely many quotients?\n\n- In the introduction, the paper mentions, “Our library opens the door for the extensive theory and existing software on symbolic computation and computational tropical geometry to be used to study neural networks.\" Could you provide an example of a property of a function computed by a neural network that is now accessible through symbolic computation but was previously inaccessible? A more detailed explanation of these advantages would be helpful.\n\n- Why is the native tropical representation for neural networks useful for measuring complexity? Can you comment on how this representation compares to the best representation as a tropical rational function in terms of monomial complexity?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper investigates counting the linear regions of neural networks through tropical geometry. The authors relate the Hoffman constant of a tropical rational map to the minimal effective radius, which is the smallest radius $r$ such that a ball of radius $r$ intersects all linear regions. They further show that for an invariant network (with respect to a group $G$), one can sample from the fundamental domain of $G$ to obtain an upper bound on the number of linear regions. Lastly, the authors introduce an algorithm using symbolic representation to compute the number of linear regions and propose a new measure of complexity for ReLU neural networks, termed the “monomial count.”", "soundness": 3, "presentation": 2, "contribution": 2, "strengths": "The number of linear regions is a standard measure of complexity for neural networks with piecewise linear activation functions, and the authors provide new insights into computing this count for a given network. Specifically, they relate the minimal effective radius to the Hoffman constant of a tropical Puiseux rational map. Additionally, for invariant neural networks, their observation allows a significant reduction in the number of samples required to estimate the number of linear regions.", "weaknesses": "Overall, the contributions of the paper do not appear substantial enough, leading me to recommend rejection. The theoretical depth is limited, and I see few immediate practical benefits from the results, see also the questions below.\n-  While relating the Hoffman constant to the radius of the ball may have theoretical appeal, I do not see its practical benefit, as it is NP-hard to compute the Hoffman constant. The appendix discusses computing the Hoffman constant of a matrix, but there is no discussion about the Hoffman constant for a rational Puiseux function, which is what relates to the radius of the ball. \n- The result concerning sampling from the fundamental domain of the group action for an invariant network appears to be a straightforward observation. \n- The algorithm for determining linear regions using Puiseux polynomials does not seem to offer clear advantages over existing methods. The paper states, “Our tools consider the full input domain of these neural networks and provide an exact geometric characterization of the linear regions. This now makes previously inaccessible avenues available for analyzing the geometry of linear regions of networks (see Appendix I).\" However, algorithms for obtaining the entire polyhedral complex of neural networks already exist (e.g., Masden, 2022, https://arxiv.org/abs/2207.07696).", "questions": "- What is the benefit of relating the Hoffman constant to the radius of the ball? Are there any approximations for the Hoffman constant of a rational Puiseux function? Has the Hoffman constant been applied to Puiseux polynomials (and/or rational functions) in prior work?\n\n- Assuming we know the Hoffman constant, are there theoretical guarantees that using it would lead to better sampling to estimate the number of linear regions?\n\n- Is the definition of the Hoffman constant for a Puiseux rational map well-defined (Definition 2.2)? How do you deal with the fact that there can be infinitely many quotients?\n\n- In the introduction, the paper mentions, “Our library opens the door for the extensive theory and existing software on symbolic computation and computational tropical geometry to be used to study neural networks.\" Could you provide an example of a property of a function computed by a neural network that is now accessible through symbolic computation but was previously inaccessible? A more detailed explanation of these advantages would be helpful.\n\n- Why is the native tropical representation for neural networks useful for measuring complexity? Can you comment on how this representation compares to the best representation as a tropical rational function in terms of monomial complexity?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 3, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1730478984207}, {"id": "Zh6Obwmscj", "reviewer_signature": ["ICLR.cc/2025/Conference/Submission12294/Reviewer_3Zin"], "rating": 5, "soundness": 3, "presentation": 3, "contribution": 2, "confidence": 4, "summary": "The authors study some theoretical and computational aspects of neural networks with piecewise linear activation functions. In Section 2, they are interested in the number of linear regions of such a network. They define the minimal effective radius around an input point as the smallest radius such that a ball of that radius intersects all the linear regions. They bound that radius from above by a Hoffman constant associated with the network. The main advertised application of that is that one can in principle enumerate all linear regions by sampling dense enough inside a ball of their estimated radius.\n\nIn Section 3, the authors consider networks that are invariant under a given group action. They provide lower and upper bounds on the number of linear regions by making use of the fundamental domain of the group action. Since the fundamental domain gives a tiling of the input space, the authors further explain that sampling inside the fundamental domain suffices to enumerate all linear regions.\n\nIn Section 4, the authors explain their software contribution. They provide a symbolic algorithm that finds all linear regions of a network once one has a representation of the network as a tropical rational function. The authors also suggest to measure the expressivity of a network by the number of monomial required in a tropical rational function representation. They provide an algorithm that reduces unnecessary monomials in a given representation. \n\nThe authors conclude with some experiments and discussion, but their main contributions are the theoretical contributions in Sections 2 and 3, and their software contribution described in Section 4.", "review_text": "The authors study some theoretical and computational aspects of neural networks with piecewise linear activation functions. In Section 2, they are interested in the number of linear regions of such a network. They define the minimal effective radius around an input point as the smallest radius such that a ball of that radius intersects all the linear regions. They bound that radius from above by a Hoffman constant associated with the network. The main advertised application of that is that one can in principle enumerate all linear regions by sampling dense enough inside a ball of their estimated radius.\n\nIn Section 3, the authors consider networks that are invariant under a given group action. They provide lower and upper bounds on the number of linear regions by making use of the fundamental domain of the group action. Since the fundamental domain gives a tiling of the input space, the authors further explain that sampling inside the fundamental domain suffices to enumerate all linear regions.\n\nIn Section 4, the authors explain their software contribution. They provide a symbolic algorithm that finds all linear regions of a network once one has a representation of the network as a tropical rational function. The authors also suggest to measure the expressivity of a network by the number of monomial required in a tropical rational function representation. They provide an algorithm that reduces unnecessary monomials in a given representation. \n\nThe authors conclude with some experiments and discussion, but their main contributions are the theoretical contributions in Sections 2 and 3, and their software contribution described in Section 4.", "strengths": "Originality: The definition of the Hoffman constant and its usage for sampling linear regions is novel, to the best of my knowledge. Also, I believe that the monomial count expressivity measure has not been studied before.\n\nSignificance: I think it will be very helpful for future research that the authors made their software available within the open-source OSCAR library. The proposed ideas of the Hoffman constant and the fundamental region might play a significant role in the future understanding of finding all linear regions of a piecewise linear network.\n\nQuality and clarity: The paper is well-written. The mathematical statements are sound and the experiments make sense.", "weaknesses": "My main concern is whether the paper's contributions are strong enough, which consist mainly of the author's software and their idea of the Hoffman constant to find a bounded region in the input space that intersects all linear regions. Making use of the fundamental domain of a given group action is a rather folklore idea, and regarding the monomial count expressivity measure I have some questions (see below). Moreover, and also very important: The paper oversells its contributions. I strongly suggest to tone down the description of the contributions and focus on describing what exactly is done instead.\n\n1) For instance, the last sentence of the abstract is way too bold. The authors claim to provide \"the foundations\" but it is rather a small piece of the puzzle.\n\n2)  Another example is the sentence \"We study more than simply the number of linear regions, we provide further insight on their geometry\". The only new geometric aspect I can see is the definition of the minimal effective radius. But that does not really give us new insights about the geometric properties of the regions, such as their size/volume, their number of facets or vertices, or any other information about the combinatorial type of the polytopes. I suggest to remove the sentence in question completely.\n\n3) Similarly, in line 088, the authors write that the Hoffman constant \"effectively gives a geometric characterization\". I do not agree with that half-sentence, since the constant does not tell us (almost) anything about the geometric properties of the linear regions; it only tells us how \"far away\" the regions are at most located. If the authors have something else in mind, I suggest to replace the vague and way too broad statement \"geometric characterization\" by the exact properties the Hoffman constant tells us.\n\n4) I also think that the phrase \"guarantee for the linear regions\" in line 088 is too strong. Please be more specific on what that guarantee means. The authors provide a largest region for where one has to sample. But if the sample is not dense enough, one will still miss linear regions (as the authors note in the appendix).\n\n5) It is always hard to claim to have the \"first work\" as in line 531. Symbolic experiments on other deep learning architectures have for instance been done in \"Pure and Spurious Critical Points: a Geometric Study of Linear Networks\" by Trager, Kohn, Bruna or in \"Geometry of Linear Convolutional Networks\" by Kohn, Merkh, Montúfar, Trager.\n\nFinally, here some smaller comments:\n\n- Please add \"piecewise\" before \"linear\" in the first line of the abstract.\n\n- I think it is inaccurate to call tropical geometry a \"reinterpretation\" of algebraic geometry. It is rather a \"discrete version\" of algebraic geometry.\n\n- In line 139.5, remove one of the 2 words \"its the\"\n\n- In lines 176 and 177, the row vectors a and a' should have an index i. That is, they should be a_i resp. a'_i.\n\n- In line 191, it would be better to spell out \"for all\" before \"U\" to increase clarity.\n\n- In line 205, should m_p and m_q be equal?\n\n- In line 212, insert \"in\" after \"x\"\n\n- typo in caption of Figure 3", "questions": "1) In Theorem 3.3, what happens if the group is infinite? Shouldn't your theorem imply that the number of linear regions is infinite in this case? Or will it also be true that the set U_c is empty such that your lower bound becomes trivially equal to 0?\n\n2) A priory, more monomials does not necessarily mean more expressivity, since the monomials could satisfy some internal (algebraic) relations/constraints. For instance, in classical algebra, the polynomials that are large powers of linear terms (i.e., that are of the form (a*x+b)^N for some fixed large N) have many monomials but they can only express a small portion (namely, a 2-dimensional subset) of the function space. Does this behavior happen for the tropical representations of ReLU neural networks? If yes, then a more accurate measure of expressivity would be the dimension of the semi-algebraic set that consists of all tropical rational functions that a fixed network architectures parametrizes, as studied in the article \"Functional dimension of feedforward ReLU neural networks\" by Grigsby, Lindsey, Meyerhoff, Wu. \n\n3) I don't understand what is meant with the number of monomials in Figure 9, since both the numerator and denominator have monomials. Do you plot the sum of the number of monomials of both numerator and denominator? Wouldn't in be better to also plot the product to get an idea of how far the upper bound in Example E.2 is?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The authors study some theoretical and computational aspects of neural networks with piecewise linear activation functions. In Section 2, they are interested in the number of linear regions of such a network. They define the minimal effective radius around an input point as the smallest radius such that a ball of that radius intersects all the linear regions. They bound that radius from above by a Hoffman constant associated with the network. The main advertised application of that is that one can in principle enumerate all linear regions by sampling dense enough inside a ball of their estimated radius.\n\nIn Section 3, the authors consider networks that are invariant under a given group action. They provide lower and upper bounds on the number of linear regions by making use of the fundamental domain of the group action. Since the fundamental domain gives a tiling of the input space, the authors further explain that sampling inside the fundamental domain suffices to enumerate all linear regions.\n\nIn Section 4, the authors explain their software contribution. They provide a symbolic algorithm that finds all linear regions of a network once one has a representation of the network as a tropical rational function. The authors also suggest to measure the expressivity of a network by the number of monomial required in a tropical rational function representation. They provide an algorithm that reduces unnecessary monomials in a given representation. \n\nThe authors conclude with some experiments and discussion, but their main contributions are the theoretical contributions in Sections 2 and 3, and their software contribution described in Section 4.", "soundness": 3, "presentation": 3, "contribution": 2, "strengths": "Originality: The definition of the Hoffman constant and its usage for sampling linear regions is novel, to the best of my knowledge. Also, I believe that the monomial count expressivity measure has not been studied before.\n\nSignificance: I think it will be very helpful for future research that the authors made their software available within the open-source OSCAR library. The proposed ideas of the Hoffman constant and the fundamental region might play a significant role in the future understanding of finding all linear regions of a piecewise linear network.\n\nQuality and clarity: The paper is well-written. The mathematical statements are sound and the experiments make sense.", "weaknesses": "My main concern is whether the paper's contributions are strong enough, which consist mainly of the author's software and their idea of the Hoffman constant to find a bounded region in the input space that intersects all linear regions. Making use of the fundamental domain of a given group action is a rather folklore idea, and regarding the monomial count expressivity measure I have some questions (see below). Moreover, and also very important: The paper oversells its contributions. I strongly suggest to tone down the description of the contributions and focus on describing what exactly is done instead.\n\n1) For instance, the last sentence of the abstract is way too bold. The authors claim to provide \"the foundations\" but it is rather a small piece of the puzzle.\n\n2)  Another example is the sentence \"We study more than simply the number of linear regions, we provide further insight on their geometry\". The only new geometric aspect I can see is the definition of the minimal effective radius. But that does not really give us new insights about the geometric properties of the regions, such as their size/volume, their number of facets or vertices, or any other information about the combinatorial type of the polytopes. I suggest to remove the sentence in question completely.\n\n3) Similarly, in line 088, the authors write that the Hoffman constant \"effectively gives a geometric characterization\". I do not agree with that half-sentence, since the constant does not tell us (almost) anything about the geometric properties of the linear regions; it only tells us how \"far away\" the regions are at most located. If the authors have something else in mind, I suggest to replace the vague and way too broad statement \"geometric characterization\" by the exact properties the Hoffman constant tells us.\n\n4) I also think that the phrase \"guarantee for the linear regions\" in line 088 is too strong. Please be more specific on what that guarantee means. The authors provide a largest region for where one has to sample. But if the sample is not dense enough, one will still miss linear regions (as the authors note in the appendix).\n\n5) It is always hard to claim to have the \"first work\" as in line 531. Symbolic experiments on other deep learning architectures have for instance been done in \"Pure and Spurious Critical Points: a Geometric Study of Linear Networks\" by Trager, Kohn, Bruna or in \"Geometry of Linear Convolutional Networks\" by Kohn, Merkh, Montúfar, Trager.\n\nFinally, here some smaller comments:\n\n- Please add \"piecewise\" before \"linear\" in the first line of the abstract.\n\n- I think it is inaccurate to call tropical geometry a \"reinterpretation\" of algebraic geometry. It is rather a \"discrete version\" of algebraic geometry.\n\n- In line 139.5, remove one of the 2 words \"its the\"\n\n- In lines 176 and 177, the row vectors a and a' should have an index i. That is, they should be a_i resp. a'_i.\n\n- In line 191, it would be better to spell out \"for all\" before \"U\" to increase clarity.\n\n- In line 205, should m_p and m_q be equal?\n\n- In line 212, insert \"in\" after \"x\"\n\n- typo in caption of Figure 3", "questions": "1) In Theorem 3.3, what happens if the group is infinite? Shouldn't your theorem imply that the number of linear regions is infinite in this case? Or will it also be true that the set U_c is empty such that your lower bound becomes trivially equal to 0?\n\n2) A priory, more monomials does not necessarily mean more expressivity, since the monomials could satisfy some internal (algebraic) relations/constraints. For instance, in classical algebra, the polynomials that are large powers of linear terms (i.e., that are of the form (a*x+b)^N for some fixed large N) have many monomials but they can only express a small portion (namely, a 2-dimensional subset) of the function space. Does this behavior happen for the tropical representations of ReLU neural networks? If yes, then a more accurate measure of expressivity would be the dimension of the semi-algebraic set that consists of all tropical rational functions that a fixed network architectures parametrizes, as studied in the article \"Functional dimension of feedforward ReLU neural networks\" by Grigsby, Lindsey, Meyerhoff, Wu. \n\n3) I don't understand what is meant with the number of monomials in Figure 9, since both the numerator and denominator have monomials. Do you plot the sum of the number of monomials of both numerator and denominator? Wouldn't in be better to also plot the product to get an idea of how far the upper bound in Example E.2 is?", "flag_for_ethics_review": ["No ethics review needed."], "rating": 5, "confidence": 4, "code_of_conduct": "Yes"}, "tcdate": 1729176308505}], "openreview_url": "https://openreview.net/forum?id=aYx7JR20sI", "arxiv_id": "2405.20174", "paper_pdf": "papers/aYx7JR20sI.pdf", "paper_pdf_sha256": "77dd594f146bb00e667ffdf799c9032ceeb5d1490139b3e5f29cfe9d3017d6af", "paper_pdf_bytes": 525659, "paper_pdf_source": "openreview", "code_url": "https://github.com/Paul-Lez/tropicalnn", "code_repository": "Paul-Lez/tropicalnn", "code_commit": "479ec5e96030f46fe5d0783d74ca193196512729", "code_archive": "repos/aYx7JR20sI.zip", "code_archive_sha256": "ceb1903c51542bbc926e575c1322b4f3b31238713f2f846c3dff7746f43dfd85", "code_archive_bytes": 61274, "code_file_count": 15, "code_extensions": {".jl": 14, ".sh": 1}, "github_disk_usage_kb": 451, "github_languages": {"Julia": 66379, "Shell": 1603}, "github_archived": false, "github_pushed_at": "2026-08-28T16:56:29Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/tropical-expressivity-of-neural-networks"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "YGTSLDAPqb", "year": 2024, "status": "rejected", "title": "Connect Later: Improving Fine-Tuning for Robustness with Targeted Augmentations", "authors": ["Helen Qu", "Sang Michael Xie"], "authorids": ["~Helen_Qu1", "~Sang_Michael_Xie1"], "authors_source": "OpenReview API", "abstract": "Models trained on a labeled source domain (e.g., bright, nearby astronomical objects) often generalize poorly when deployed on an out-of-distribution (OOD) target domain (e.g., faint, distant objects). In the domain adaptation setting where unlabeled target data is available, self-supervised pretraining (e.g., masked autoencoding or contrastive learning) is a promising method to mitigate this performance drop. Pretraining improves OOD error when the generic data augmentations used (e.g., masking or cropping) connect the source and target domains, which may be far apart in the input space. In this paper, we show on real-world tasks that standard fine-tuning after pretraining does not consistently improve OOD error over just supervised learning on labeled source data. To better leverage pretraining for distribution shifts, we propose Connect Later: after pretraining with generic augmentations to learn good representations within the source and target domains, fine-tune with targeted augmentations designed with knowledge of the distribution shift to better connect the domains. Connect Later improves average OOD error over standard fine-tuning and supervised learning with targeted augmentations on 3 real-world datasets: astronomical time-series classification (AstroClassification) by 12%, redshift prediction for astronomical time-series (Redshifts) by 0.03 RMSE (11% relative), and wildlife species identification (iWildCam-WILDS) by 0.9%, achieving the state-of-the-art on AstroClassification and on iWildCam-WILDS with ResNet-50.", "decision": "Reject", "meta_review": "", "num_reviews": 3, "reviews": [{"id": "9T4OHHKswH", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6720/Reviewer_vvbp"], "rating": "3: reject, not good enough", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "This manuscript deals with the domain adaptation problem, i.e., the setting in which there is a domain shift between a source and target domain and where only unlabelled data is available for the target domain. One approach to this problem is to use self-supervised learning using both the source and target domain data, followed by fine-tuning with the labelled source domain data.\n\nThe authors begin with the observation that this approach leads to inconsistent results compared to regular supervised training (ERM). They hypothesize this is because the self-supervised methods learn features that are strongly domain-specific, which means that when the model is fine-tuned the classifier from the last layer might not generalize beyond the source distribution. The authors propose a solution which involves using targeted augmentations (i.e., augmentations that are specifically designed to remove spurious domain-dependent features) during the fine-tuning phase. They show that this improves performance on three different datasets.", "review_text": "This manuscript deals with the domain adaptation problem, i.e., the setting in which there is a domain shift between a source and target domain and where only unlabelled data is available for the target domain. One approach to this problem is to use self-supervised learning using both the source and target domain data, followed by fine-tuning with the labelled source domain data.\n\nThe authors begin with the observation that this approach leads to inconsistent results compared to regular supervised training (ERM). They hypothesize this is because the self-supervised methods learn features that are strongly domain-specific, which means that when the model is fine-tuned the classifier from the last layer might not generalize beyond the source distribution. The authors propose a solution which involves using targeted augmentations (i.e., augmentations that are specifically designed to remove spurious domain-dependent features) during the fine-tuning phase. They show that this improves performance on three different datasets.", "strengths": "This paper looks at an interesting problem: Unsupervised domain adaptation is a very relevant problem (e.g., in many cases practitioners have labelled datasets available from a curated/lab setting, but need to generalize to actual deployment conditions). At the same time, self-supervised learning has become very popular, but its performance in the context of unsupervised domain adaptation (UDA) is still unclear.\n\nThis paper introduces a new dataset (RedShifts), and contains some encouraging results on two different domains (camera trips and astronomical observations) while performing several relevant ablations (model scale, whether to train the last layer before fine-tuning, strength of pre-training augmentations).\n\nThe paper is relatively easy to read, with good writing and a clear structure.", "weaknesses": "Overall, I find this paper lacking in a variety of areas:\n\nThe authors' refer several times to Shen et al. (2022). This paper argues that contrastive pre-training is beneficial as long as data augmentations ensure that augmented examples must be more likely to change class or domain than changing both. The authors of this manuscript seem to argue in section 3 that the bad results they observe are explained by a violation of this assumption from Shen et al. (2022). However, this isn't tested rigorously (see section 6 of Shen et al., 2022, for examples on how to evaluate connectivity on real world datasets).\n\nA second shortcoming is that different self-supervised methods were used for the iWildCam-WILDS dataset (SWaV contrastive learning) and the astronomical time-series (masked autoencoding). This makes it hard to draw conclusions from table 1, since it is unclear which differences can be attributed to the use of different pre-training objectives.\n\nThe results on the iWildCam dataset are not very strong (given the means and standard deviations listed, there's a 20% chance that ERM + targeted augs outperforms Connect Later for any run). This leaves the improvements in the astronomical time-series datasets as the stronger proof of Connect Later's performance. But these results are a bit confusing to me: How is it that standard fine-tuning improves in-domain performance (tables 1 and 2)? This suggests to me that these datasets are possibly very small and benefit strongly from the regularization that pre-training provides. In that case, is Connect Later really addressing a domain shift issue? Or is it just addressing a regular overfitting issue? The latter seems plausible given the very small size of the dataset (6,74 objects). I would argue that in order to draw conclusions, the method should be tested on a wider variety of larger datasets.\n\nA baseline that seems to be missing as well is the use of general augmentations during fine-tuning (my understanding is that standard fine-tuning uses no augmentations at all). This would clarify whether it is important to use targeted augmentations during fine-tuning, or whether any form of augmentations would be fine.\n\nOverall, I find the insights that this paper provides limited: It is not surprising to me that targeted augmentations would help with fine-tuning, given that they were shown to help with regular ERM. The bigger question to me is the one raised in section 3: Why does pre-training not always lead to increased OOD performance? This question remains largely unanswered. I think this paper would be better if it provided, for example, (1) clear evidence that the assumptions from Shen et al. (2022) are routinely violated in real-world datasets and commonly used augmentations, (2) a thorough evaluation of how different pre-training techniques (MAEs, contrastive learning) perform in the UDA setting, or (3) strong evidence that targeted augmentations are particularly important compared to general augmentations.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "This manuscript deals with the domain adaptation problem, i.e., the setting in which there is a domain shift between a source and target domain and where only unlabelled data is available for the target domain. One approach to this problem is to use self-supervised learning using both the source and target domain data, followed by fine-tuning with the labelled source domain data.\n\nThe authors begin with the observation that this approach leads to inconsistent results compared to regular supervised training (ERM). They hypothesize this is because the self-supervised methods learn features that are strongly domain-specific, which means that when the model is fine-tuned the classifier from the last layer might not generalize beyond the source distribution. The authors propose a solution which involves using targeted augmentations (i.e., augmentations that are specifically designed to remove spurious domain-dependent features) during the fine-tuning phase. They show that this improves performance on three different datasets.", "soundness": "1 poor", "presentation": "3 good", "contribution": "2 fair", "strengths": "This paper looks at an interesting problem: Unsupervised domain adaptation is a very relevant problem (e.g., in many cases practitioners have labelled datasets available from a curated/lab setting, but need to generalize to actual deployment conditions). At the same time, self-supervised learning has become very popular, but its performance in the context of unsupervised domain adaptation (UDA) is still unclear.\n\nThis paper introduces a new dataset (RedShifts), and contains some encouraging results on two different domains (camera trips and astronomical observations) while performing several relevant ablations (model scale, whether to train the last layer before fine-tuning, strength of pre-training augmentations).\n\nThe paper is relatively easy to read, with good writing and a clear structure.", "weaknesses": "Overall, I find this paper lacking in a variety of areas:\n\nThe authors' refer several times to Shen et al. (2022). This paper argues that contrastive pre-training is beneficial as long as data augmentations ensure that augmented examples must be more likely to change class or domain than changing both. The authors of this manuscript seem to argue in section 3 that the bad results they observe are explained by a violation of this assumption from Shen et al. (2022). However, this isn't tested rigorously (see section 6 of Shen et al., 2022, for examples on how to evaluate connectivity on real world datasets).\n\nA second shortcoming is that different self-supervised methods were used for the iWildCam-WILDS dataset (SWaV contrastive learning) and the astronomical time-series (masked autoencoding). This makes it hard to draw conclusions from table 1, since it is unclear which differences can be attributed to the use of different pre-training objectives.\n\nThe results on the iWildCam dataset are not very strong (given the means and standard deviations listed, there's a 20% chance that ERM + targeted augs outperforms Connect Later for any run). This leaves the improvements in the astronomical time-series datasets as the stronger proof of Connect Later's performance. But these results are a bit confusing to me: How is it that standard fine-tuning improves in-domain performance (tables 1 and 2)? This suggests to me that these datasets are possibly very small and benefit strongly from the regularization that pre-training provides. In that case, is Connect Later really addressing a domain shift issue? Or is it just addressing a regular overfitting issue? The latter seems plausible given the very small size of the dataset (6,74 objects). I would argue that in order to draw conclusions, the method should be tested on a wider variety of larger datasets.\n\nA baseline that seems to be missing as well is the use of general augmentations during fine-tuning (my understanding is that standard fine-tuning uses no augmentations at all). This would clarify whether it is important to use targeted augmentations during fine-tuning, or whether any form of augmentations would be fine.\n\nOverall, I find the insights that this paper provides limited: It is not surprising to me that targeted augmentations would help with fine-tuning, given that they were shown to help with regular ERM. The bigger question to me is the one raised in section 3: Why does pre-training not always lead to increased OOD performance? This question remains largely unanswered. I think this paper would be better if it provided, for example, (1) clear evidence that the assumptions from Shen et al. (2022) are routinely violated in real-world datasets and commonly used augmentations, (2) a thorough evaluation of how different pre-training techniques (MAEs, contrastive learning) perform in the UDA setting, or (3) strong evidence that targeted augmentations are particularly important compared to general augmentations.", "questions": "See weaknesses.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "3: reject, not good enough", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698895189549}, {"id": "3oDiOYnaZs", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6720/Reviewer_atYU"], "rating": "5: marginally below the acceptance threshold", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes Connect Later to improve robustness. It involves performing self-supervised pretraining (e.g., masked autoencoding or contrastive learning) followed by finetuning with augmentations designed with knowledge of the distribution shift. Supportive results were shown on several real-world dataset.", "review_text": "The paper proposes Connect Later to improve robustness. It involves performing self-supervised pretraining (e.g., masked autoencoding or contrastive learning) followed by finetuning with augmentations designed with knowledge of the distribution shift. Supportive results were shown on several real-world dataset.", "strengths": "Evaluations were done on real world datasets with supportive results.", "weaknesses": "1. Generalizability of the method. \n    - If target augmentations involve knowledge of the distribution shift, why pre-training is needed? If we are able to generate the target data, where other options include using powerful generative models [1], why do we need pre-training? It seems to be more a question of when is pre-training necessary. \n    - It is also not clear if pre-training in general is useful as the experimental results were only shown with one type of pre-training method and a different pre-training for different datasets. How sensitive are the results to the choice of pre-training method?\n    - From Table 2, it seems like the results are sensitive to the target augmentations. For iwilds most of the gains seem to come from the targeted augmentations. But for the astro datasets the ID test acc for astroclassification is better without targeted augmentations. One a new task or dataset, how should one choose the targeted augmentations?\n2. Test-time methods, like test-time augmentation/training also makes use of unlabelled data. These methods aim to adapt to the distribution shift as they occur. Furthermore, some of these pre-training objectives like masked auto-encoding have been used for test-time adaptation [2]. It may be beneficial to pose these methods wrt Connect Later.\n4. Presentation\n    - Sec 4 gives the technical details for creating augmentations before explaining the tasks. It may not be clear what the tasks for Astroclassification and Redshift are about and so the technical details are not easy to understand. I would suggest moving Sec 5 before 4.\n    - In Fig 2, top row, what do the colors mean? Why does augmentation introduce errors?\n    - It may be useful to have additional columns in the main table of results (e.g. Tab 2), that describes what the different methods are. E.g., whether finetuning augmentations are used, and with checks in the rows for the relevant methods.\n\n\n\n\n----\n[1] Diversify Your Vision Datasets with Automatic Diffusion-based Augmentation. NeurIPS’23 \n\n[2] Y. Gandelsman, et al. Test-Time Training with Masked Autoencoders. NeurIPS’23", "questions": "See the points raised in weaknesses for questions and suggestions.", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Connect Later to improve robustness. It involves performing self-supervised pretraining (e.g., masked autoencoding or contrastive learning) followed by finetuning with augmentations designed with knowledge of the distribution shift. Supportive results were shown on several real-world dataset.", "soundness": "2 fair", "presentation": "2 fair", "contribution": "2 fair", "strengths": "Evaluations were done on real world datasets with supportive results.", "weaknesses": "1. Generalizability of the method. \n    - If target augmentations involve knowledge of the distribution shift, why pre-training is needed? If we are able to generate the target data, where other options include using powerful generative models [1], why do we need pre-training? It seems to be more a question of when is pre-training necessary. \n    - It is also not clear if pre-training in general is useful as the experimental results were only shown with one type of pre-training method and a different pre-training for different datasets. How sensitive are the results to the choice of pre-training method?\n    - From Table 2, it seems like the results are sensitive to the target augmentations. For iwilds most of the gains seem to come from the targeted augmentations. But for the astro datasets the ID test acc for astroclassification is better without targeted augmentations. One a new task or dataset, how should one choose the targeted augmentations?\n2. Test-time methods, like test-time augmentation/training also makes use of unlabelled data. These methods aim to adapt to the distribution shift as they occur. Furthermore, some of these pre-training objectives like masked auto-encoding have been used for test-time adaptation [2]. It may be beneficial to pose these methods wrt Connect Later.\n4. Presentation\n    - Sec 4 gives the technical details for creating augmentations before explaining the tasks. It may not be clear what the tasks for Astroclassification and Redshift are about and so the technical details are not easy to understand. I would suggest moving Sec 5 before 4.\n    - In Fig 2, top row, what do the colors mean? Why does augmentation introduce errors?\n    - It may be useful to have additional columns in the main table of results (e.g. Tab 2), that describes what the different methods are. E.g., whether finetuning augmentations are used, and with checks in the rows for the relevant methods.\n\n\n\n\n----\n[1] Diversify Your Vision Datasets with Automatic Diffusion-based Augmentation. NeurIPS’23 \n\n[2] Y. Gandelsman, et al. Test-Time Training with Masked Autoencoders. NeurIPS’23", "questions": "See the points raised in weaknesses for questions and suggestions.", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698749236498}, {"id": "qGTNdQxtUg", "reviewer_signature": ["ICLR.cc/2024/Conference/Submission6720/Reviewer_erbX"], "rating": "5: marginally below the acceptance threshold", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary": "The paper proposes Connect Later to improve model robustness in domain adaptation scenarios. The approach first leverages generic augmentations to pretrain on combined unlabeled source and target data. It then employs carefully designed targeted augmentations during fine-tuning on labeled source data to better connect the source and target domains based on knowledge of their distribution shift. The experiment shows the effectiveness of Connect Later over several baselines on three datasets.", "review_text": "The paper proposes Connect Later to improve model robustness in domain adaptation scenarios. The approach first leverages generic augmentations to pretrain on combined unlabeled source and target data. It then employs carefully designed targeted augmentations during fine-tuning on labeled source data to better connect the source and target domains based on knowledge of their distribution shift. The experiment shows the effectiveness of Connect Later over several baselines on three datasets.", "strengths": "The paper is clearly written in general, so the overall quality of the paper is satisfactory.", "weaknesses": "1. The proposed framework hinges on the manual design of the transformation distribution and requires knowledge about the distribution shift between source and target domains to create effective targeted augmentations. This property limits the applicability of the proposed method in scenarios where the shift is unknown or diverse. \n\n2. On iWildCam, Connect Later only marginally outperforms ERM+targeted augmentations, which indicates that if there is diverse distribution shifts, the proposed method is not quite effective.", "questions": "Could the authors test the sensitivity of the performance of Connect Later with respect to the design choice of transformation distribution?", "flag_for_ethics_review": ["No ethics review needed."], "all_content": {"summary": "The paper proposes Connect Later to improve model robustness in domain adaptation scenarios. The approach first leverages generic augmentations to pretrain on combined unlabeled source and target data. It then employs carefully designed targeted augmentations during fine-tuning on labeled source data to better connect the source and target domains based on knowledge of their distribution shift. The experiment shows the effectiveness of Connect Later over several baselines on three datasets.", "soundness": "3 good", "presentation": "3 good", "contribution": "2 fair", "strengths": "The paper is clearly written in general, so the overall quality of the paper is satisfactory.", "weaknesses": "1. The proposed framework hinges on the manual design of the transformation distribution and requires knowledge about the distribution shift between source and target domains to create effective targeted augmentations. This property limits the applicability of the proposed method in scenarios where the shift is unknown or diverse. \n\n2. On iWildCam, Connect Later only marginally outperforms ERM+targeted augmentations, which indicates that if there is diverse distribution shifts, the proposed method is not quite effective.", "questions": "Could the authors test the sensitivity of the performance of Connect Later with respect to the design choice of transformation distribution?", "flag_for_ethics_review": ["No ethics review needed."], "rating": "5: marginally below the acceptance threshold", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "code_of_conduct": "Yes"}, "tcdate": 1698689584495}], "openreview_url": "https://openreview.net/forum?id=YGTSLDAPqb", "arxiv_id": "2402.03325", "paper_pdf": "papers/YGTSLDAPqb.pdf", "paper_pdf_sha256": "d62f4a92d56ceb897111d1b57253777107275f060253eeb3d030f22fa0ea4439", "paper_pdf_bytes": 1827859, "paper_pdf_source": "openreview", "code_url": "https://github.com/helenqu/connect-later", "code_repository": "helenqu/connect-later", "code_commit": "6a12eab8918ac0152fa79aa168c028e6f378a478", "code_archive": "repos/YGTSLDAPqb.zip", "code_archive_sha256": "cee7e178d6e53111ce3368e94c9d1eb60da77dfdc95690c4dc139f3817f77456", "code_archive_bytes": 40610, "code_file_count": 23, "code_extensions": {".py": 13, ".sh": 10}, "github_disk_usage_kb": 556, "github_languages": {"Python": 111188, "Shell": 5135}, "github_archived": false, "github_pushed_at": "2024-01-03T18:09:49Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/connect-later-improving-fine-tuning-for"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}
{"forum": "IJwhRE510b", "year": 2023, "status": "rejected", "title": "ELODI: Ensemble Logit Difference Inhibition for Positive-Congruent Training", "authors": ["Yue Zhao", "Yantao Shen", "Yuanjun Xiong", "Shuo Yang", "Wei Xia", "Zhuowen Tu", "Bernt Schiele", "Stefano Soatto"], "authorids": ["~Yue_Zhao4", "~Yantao_Shen2", "~Yuanjun_Xiong3", "~Shuo_Yang2", "~Wei_Xia6", "~Zhuowen_Tu1", "~Bernt_Schiele1", "~Stefano_Soatto1"], "authors_source": "OpenReview API", "abstract": "Negative flips are errors introduced in a classification system when a legacy model is updated. Existing methods to reduce the negative flip rate (NFR) either do so at the expense of overall accuracy by forcing a new model to imitate the old models, or use ensembles, which multiply inference cost prohibitively. We analyze the role of ensembles in reducing NFR and observe that they remove negative flips that are typically not close to the decision boundary, but often exhibit large deviations in the distance among their logits. Based on the observation, we present a method, called Ensemble Logit Difference Inhibition ELODI, to train a classification system that achieves paragon performance in both error rate and NFR, at the inference cost of a single model. The method distills a homogeneous ensemble to a single student model which is used to update the classification system. ELODI also introduces a generalized distillation objective, Logit Difference Inhibition (LDI), which penalizes changes in the logits between the reference ensemble and the student single model.\nOn multiple image classification benchmarks, model updates with ELODI demonstrate superior accuracy retention and NFR reduction. ", "decision": "Reject", "meta_review": null, "num_reviews": 4, "reviews": [{"id": "phCn3MpAiO", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper8/Reviewer_asve"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "The paper addresses the task of updating a model by learning a stronger model while ensuring that formerly correctly classified instances will get wrong predictions as seldom as possible, a quantity called a negative flip rate (NFR). The paper starts with a theoretical analysis of using ensembles in this context. It then proposes a new training objective to distill a single model based on an ensemble, with the goal of minimizing NFR. Experiments on image datasets ImageNet and iNaturalist show the effectiveness of ELODI compared to other NFR-reducing methods.", "review_text": "The paper addresses an important topic and proposes new effective methods. It also studies the theoretical justification behind the proposed methods. In my opinion, an important aspect has not been covered as an ablation study, justifying the difference of the proposed method from the existing logit matching loss by Hinton et al 2015. There are some shortcomings in the theoretical part, as highlighted above in the list of weaknesses.", "strengths": "Strengths:\n* Theoretical analysis of the differences of average logits in ensembles;\n* The paper is mostly clearly written, with plenty of insightful figures and tables;\n* The proposed method ELODI is effective in reducing NFR while mostly retaining accuracy.\n\nWeaknesses:\n* At the end of Section 4, it is stated that the logit matching loss (Hinton et al., 2015) is a special case of the proposed LDI term. Therefore, I find it absolutely necessary to do an ablation study to see whether the proposed method improves over using the plain logit matching loss. That is, the paper should have experimental comparisons of $\\xi=0.1$ against $\\xi=0$. Currently there is no such comparison in the paper (nor in the appendices).\n* The theoretical part first starts by making claims about any particular fixed $x$, while considering the training process (or training seed) as a source of randomness, hence the random variable $\\phi^{(i)}(x)$. This nicely gives results stated by Eqs. (1), (2) and (3). However, in Section 3.2 there seems to be a shift to consider $x$ also as a random quantity, because there is a single Gaussian distribution about the whole architecture, e.g. see the last paragraph of page 4 with $\\phi\\sim\\mathcal{N}(\\mu_1,\\Sigma_1)$. This is confusing and should be clarified. If there is now indeed a single Gaussian per architecture, what is the theoretical justification behind that? If there is still only a Gaussian per instance, then how are the covariance matrices estimated for Figure 3?\n* The first variance term $Var(\\hat{\\phi}(x))$ in Eq.(5) does not make sense to me. As minimization is over $\\hat{\\phi}(x)$ and thus this quantity is bound within the expression to be minimized, then what is the random variable that the variance operator is applied to?\n\nMinor weaknesses:\n* From the introduction it is not quite clear whether the term positive-congruent training is novel or not;\n* The text about Figure 2 suggests that particularly the negative flips with large cross-model logit displacement are reduced by ensembles, highlighting the percentage of 90.7% in the red box. However, it would then also be important to provide the respective percentage outside the red box. Is it significantly smaller?\n* The last rows of page 3 have '$\\phi^{(i)}$ and $\\phi^{(i)}$' instead of '$\\phi^{(i)}$ and $\\psi^{(i)}$';\n* Sometimes $\\phi$ is in bold and sometimes not (e.g. in Eqs (1) and (3)), not being consistent.\n* The number of current Figure 6 should be Figure 5 instead. The text already refers to it as Figure 5.\n* In Figure 6b (that is 5b), it has not been explained what ensemble size 0 means.\n* It has not explained clearly enough what the difference between online and offline distillation is.\n* The current theory and the discussion around it say nothing about decreasing the difference between $\\mu_1$ and $\\mu_2$ which could also be a separate goal. In principle, this might possibly be achieved by changing the training objective for the members of the ensemble. Maybe it is a bad idea, but I feel this should be at least mentioned.\n\nSuggestions:\n* The acronyms LDI and ELODI are both about logit difference inhibition: wouldn't it be better to use the same suffix, such as LDI and ELDI, or LODI and ELODI?", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "The paper addresses the task of updating a model by learning a stronger model while ensuring that formerly correctly classified instances will get wrong predictions as seldom as possible, a quantity called a negative flip rate (NFR). The paper starts with a theoretical analysis of using ensembles in this context. It then proposes a new training objective to distill a single model based on an ensemble, with the goal of minimizing NFR. Experiments on image datasets ImageNet and iNaturalist show the effectiveness of ELODI compared to other NFR-reducing methods.", "strength_and_weaknesses": "Strengths:\n* Theoretical analysis of the differences of average logits in ensembles;\n* The paper is mostly clearly written, with plenty of insightful figures and tables;\n* The proposed method ELODI is effective in reducing NFR while mostly retaining accuracy.\n\nWeaknesses:\n* At the end of Section 4, it is stated that the logit matching loss (Hinton et al., 2015) is a special case of the proposed LDI term. Therefore, I find it absolutely necessary to do an ablation study to see whether the proposed method improves over using the plain logit matching loss. That is, the paper should have experimental comparisons of $\\xi=0.1$ against $\\xi=0$. Currently there is no such comparison in the paper (nor in the appendices).\n* The theoretical part first starts by making claims about any particular fixed $x$, while considering the training process (or training seed) as a source of randomness, hence the random variable $\\phi^{(i)}(x)$. This nicely gives results stated by Eqs. (1), (2) and (3). However, in Section 3.2 there seems to be a shift to consider $x$ also as a random quantity, because there is a single Gaussian distribution about the whole architecture, e.g. see the last paragraph of page 4 with $\\phi\\sim\\mathcal{N}(\\mu_1,\\Sigma_1)$. This is confusing and should be clarified. If there is now indeed a single Gaussian per architecture, what is the theoretical justification behind that? If there is still only a Gaussian per instance, then how are the covariance matrices estimated for Figure 3?\n* The first variance term $Var(\\hat{\\phi}(x))$ in Eq.(5) does not make sense to me. As minimization is over $\\hat{\\phi}(x)$ and thus this quantity is bound within the expression to be minimized, then what is the random variable that the variance operator is applied to?\n\nMinor weaknesses:\n* From the introduction it is not quite clear whether the term positive-congruent training is novel or not;\n* The text about Figure 2 suggests that particularly the negative flips with large cross-model logit displacement are reduced by ensembles, highlighting the percentage of 90.7% in the red box. However, it would then also be important to provide the respective percentage outside the red box. Is it significantly smaller?\n* The last rows of page 3 have '$\\phi^{(i)}$ and $\\phi^{(i)}$' instead of '$\\phi^{(i)}$ and $\\psi^{(i)}$';\n* Sometimes $\\phi$ is in bold and sometimes not (e.g. in Eqs (1) and (3)), not being consistent.\n* The number of current Figure 6 should be Figure 5 instead. The text already refers to it as Figure 5.\n* In Figure 6b (that is 5b), it has not been explained what ensemble size 0 means.\n* It has not explained clearly enough what the difference between online and offline distillation is.\n* The current theory and the discussion around it say nothing about decreasing the difference between $\\mu_1$ and $\\mu_2$ which could also be a separate goal. In principle, this might possibly be achieved by changing the training objective for the members of the ensemble. Maybe it is a bad idea, but I feel this should be at least mentioned.\n\nSuggestions:\n* The acronyms LDI and ELODI are both about logit difference inhibition: wouldn't it be better to use the same suffix, such as LDI and ELDI, or LODI and ELODI?", "clarity,_quality,_novelty_and_reproducibility": "The paper is sufficiently novel and reproducible, and also mostly clear, except for the shortcomings highlighted above.", "summary_of_the_review": "The paper addresses an important topic and proposes new effective methods. It also studies the theoretical justification behind the proposed methods. In my opinion, an important aspect has not been covered as an ablation study, justifying the difference of the proposed method from the existing logit matching loss by Hinton et al 2015. There are some shortcomings in the theoretical part, as highlighted above in the list of weaknesses.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1667556939042}, {"id": "3ov0UrDZoJJ", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper8/Reviewer_38Tg"], "rating": "", "confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "5: marginally below the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "This paper proposes a new method in the area of Positive-Congruent Training, by distilling a homogeneous ensemble to a single student model. The authors carefully analyze the properties of logic displacement magnitude in ensembles, and present a method called Ensemble Logit Difference Inhibition (ELODI). The proposed approach is validated on two standard image classification datasets.", "review_text": "This paper introduces a new knowledge distillation method with the goal to reducing NFR. Although the analysis and the methods are well presented, I feel the novelty of the approach is a bit limited since the key part is on modifying the loss function used in knowledge distillation; and the part on the distribution of ensemble distributions is only empirically without any theoretical supports.\n\nThat being said, I think if the authors can provide a better explanation on the benefits of the proposed loss function, and provide more theoretical discussion on the prediction distributions of ensembles, the paper can be much improved. \n\nI mark the paper as marginally below the threshold, but also note I am not familiar with this specific application area. So I am willing for further discussions. ", "strengths": "Strength\n\n- The analysis on the properties of the distribution of ensemble predictions is interesting, and the simulation studies are clear and supportive of the findings. \n- The algorithm is well presented with detailed follow up discussions. \n\nWeakness\n- I am not super clear on the motivation of the proposed method. The loss function of the distillation algorithm is a key part but it's basically just cross entropy plus a term to panelize l_2 difference in the logit. Despite being used for model updating, are there any other benefits can be discussed, such as smoothing the landscape of the student model? It seems to me using and additional LDI term could have some alternative motivations, and the use case seems a bit unnatural to me. \n- Although the discussion on the properties of ensemble prediction distributions are empirically correct, can the authors provide more discussions in this area? For example, where is the randomness coming from? Initialization, randomness in training algorithms, etc? For trees, there have been some statistical results (see for example https://arxiv.org/pdf/1510.04342.pdf, https://www.jmlr.org/papers/volume17/14-168/14-168.pdf?ref=https://githubhelp.com). I wonder whether some primitive results can be discussed here.  ", "weaknesses": "", "questions": "", "all_content": {"confidence": "2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "This paper proposes a new method in the area of Positive-Congruent Training, by distilling a homogeneous ensemble to a single student model. The authors carefully analyze the properties of logic displacement magnitude in ensembles, and present a method called Ensemble Logit Difference Inhibition (ELODI). The proposed approach is validated on two standard image classification datasets.", "strength_and_weaknesses": "Strength\n\n- The analysis on the properties of the distribution of ensemble predictions is interesting, and the simulation studies are clear and supportive of the findings. \n- The algorithm is well presented with detailed follow up discussions. \n\nWeakness\n- I am not super clear on the motivation of the proposed method. The loss function of the distillation algorithm is a key part but it's basically just cross entropy plus a term to panelize l_2 difference in the logit. Despite being used for model updating, are there any other benefits can be discussed, such as smoothing the landscape of the student model? It seems to me using and additional LDI term could have some alternative motivations, and the use case seems a bit unnatural to me. \n- Although the discussion on the properties of ensemble prediction distributions are empirically correct, can the authors provide more discussions in this area? For example, where is the randomness coming from? Initialization, randomness in training algorithms, etc? For trees, there have been some statistical results (see for example https://arxiv.org/pdf/1510.04342.pdf, https://www.jmlr.org/papers/volume17/14-168/14-168.pdf?ref=https://githubhelp.com). I wonder whether some primitive results can be discussed here.  ", "clarity,_quality,_novelty_and_reproducibility": "Overall the paper is written clearly. I feel the novelty of the approach is a bit limited since the key part is on modifying the loss function used in knowledge distillation; the part on the distribution of ensemble distributions is only empirically without any theoretical supports.\n\nSome minor things: \n\n- Figure 2 the blue color is not accurate when laid over by the orange. \n- Better to write down exact formulation of L_CE in page 6.", "summary_of_the_review": "This paper introduces a new knowledge distillation method with the goal to reducing NFR. Although the analysis and the methods are well presented, I feel the novelty of the approach is a bit limited since the key part is on modifying the loss function used in knowledge distillation; and the part on the distribution of ensemble distributions is only empirically without any theoretical supports.\n\nThat being said, I think if the authors can provide a better explanation on the benefits of the proposed loss function, and provide more theoretical discussion on the prediction distributions of ensembles, the paper can be much improved. \n\nI mark the paper as marginally below the threshold, but also note I am not familiar with this specific application area. So I am willing for further discussions. ", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "empirical_novelty_and_significance": "2: The contributions are only marginally significant or novel.", "flag_for_ethics_review": ["NO."], "recommendation": "5: marginally below the acceptance threshold"}, "tcdate": 1666753390377}, {"id": "K7HS1jjN4K", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper8/Reviewer_V2YQ"], "rating": "", "confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "4: All of the claims and statements are well-supported and correct.", "summary": "The authors tackle the problem of updating existing models in production-scale\nsystems while performing Positive-Congruent Training (minimize negative flip\nrate and error rate simultaneously). To do so, they make two key observations:\n\n- Previous studies have demonstrated that ensembles can reduce NFR without\n  negative accuracy impact.\n- Using ensembles may not be a viable strategy in real applications due to\n  computational costs.\n\nBased on these observations, the authors propose ELODI, a novel model\ndistillation method that seeks to achieve PC-Training performance of ensembles\nusing a single model. First, they analyze how the distribution of logit spaces\nof both homogeneous and heterogenous ensembles behave and conclude that, as\nensembles become larger, NFR consistently decreases. This is further validated\nwith experiments. Next, they propose a novel, margin-based distillation loss\nfunction, namely LDI, that attempts to mimic the variance reduction phenomenon\npresent in ensembles. The authors perform extensive experiments and conclude\nthat ELODI has similar performance (NFR and ER) to an ensemble method with\nreduced inference costs.", "review_text": "The authors propose ELODI, a distillation technique that reaps both the\nbenefits of ensembles to reduce NFR and ER and distillation to mimic the\nbehavior of the said ensemble using a single model. Although mainly limited to\nhomogeneous ensembles, ELODI has the potential to positively affect deep\nlearning model deployment in production.", "strengths": "*Strengths:\n\n- The paper is well-motivated, since the problem the authors present frequently\n  occurs in production systems.\n- The problem of negative flips in the context of ensembles is clearly\n  explained in Section 3 and further details the experiments in (Yan et al.\n  2021).\n- The experimental section compares the results of ELODI with several other\n  state-of-the-art techniques.\n\n*Weaknesses:\n\n- ELODI works mainly for homogeneous ensembles. This goes against one of the\n  biggest advantages of using ensembles, which is to combine models with\n  different architectures.\n- Training a model using ELODI requires one to first train an ensemble and, as\n  the authors themselves point out, the number of models in the ensemble\n  drastically impacts the quality of the distilled model.", "weaknesses": "", "questions": "", "all_content": {"confidence": "3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.", "summary_of_the_paper": "The authors tackle the problem of updating existing models in production-scale\nsystems while performing Positive-Congruent Training (minimize negative flip\nrate and error rate simultaneously). To do so, they make two key observations:\n\n- Previous studies have demonstrated that ensembles can reduce NFR without\n  negative accuracy impact.\n- Using ensembles may not be a viable strategy in real applications due to\n  computational costs.\n\nBased on these observations, the authors propose ELODI, a novel model\ndistillation method that seeks to achieve PC-Training performance of ensembles\nusing a single model. First, they analyze how the distribution of logit spaces\nof both homogeneous and heterogenous ensembles behave and conclude that, as\nensembles become larger, NFR consistently decreases. This is further validated\nwith experiments. Next, they propose a novel, margin-based distillation loss\nfunction, namely LDI, that attempts to mimic the variance reduction phenomenon\npresent in ensembles. The authors perform extensive experiments and conclude\nthat ELODI has similar performance (NFR and ER) to an ensemble method with\nreduced inference costs.", "strength_and_weaknesses": "*Strengths:\n\n- The paper is well-motivated, since the problem the authors present frequently\n  occurs in production systems.\n- The problem of negative flips in the context of ensembles is clearly\n  explained in Section 3 and further details the experiments in (Yan et al.\n  2021).\n- The experimental section compares the results of ELODI with several other\n  state-of-the-art techniques.\n\n*Weaknesses:\n\n- ELODI works mainly for homogeneous ensembles. This goes against one of the\n  biggest advantages of using ensembles, which is to combine models with\n  different architectures.\n- Training a model using ELODI requires one to first train an ensemble and, as\n  the authors themselves point out, the number of models in the ensemble\n  drastically impacts the quality of the distilled model.", "clarity,_quality,_novelty_and_reproducibility": "- The paper is well-written and easy to follow.\n- There are minor typos throughout the text (e.g., \"res- pect\").\n- Figure 1 is in page 2, but is referenced for the first time in page 5.\n- Figures 3 and 4 are somewhat hard to understand, since there is too much\n  content packed into very little space.\n- Algorithm 1 seems unnecessary, given that the procedure is explained\n  throughout the text.\n- The novelty of the paper is the distillation loss function targeted at\n  approximating the logits between a new model and an ensemble. The authors\n  compare the proposed loss with several baselines and conclude that it gives\n  better results.\n- The paper seems to be easy to reproduce (hopefully the authors will publish\n  their source code).", "summary_of_the_review": "The authors propose ELODI, a distillation technique that reaps both the\nbenefits of ensembles to reduce NFR and ER and distillation to mimic the\nbehavior of the said ensemble using a single model. Although mainly limited to\nhomogeneous ensembles, ELODI has the potential to positively affect deep\nlearning model deployment in production.", "correctness": "4: All of the claims and statements are well-supported and correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "flag_for_ethics_review": ["NO."], "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666717568027}, {"id": "VusKDt5mn4E", "reviewer_signature": ["ICLR.cc/2023/Conference/Paper8/Reviewer_uaPa"], "rating": "", "confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "recommendation": "6: marginally above the acceptance threshold", "soundness": "", "presentation": "", "contribution": "", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "summary": "In this work, the authors repurpose the well-known knowledge distillation paradigm, which is typically used for improving model accuracy, for the task of transferring the capability of reducing negative flip rate (NFR) from an ensemble to a single model. To this end, the authors devise a method called Ensemble Logit Difference Inhibition (ELODI). ELODI leverages the advantages of deep ensembles for NFR reduction at the running cost of a single model by (1) training deep networks using a Logit Difference Inhibition (LDI) loss with respect to a reference ensemble and then (2) using the resulting single model to perform inference.", "review_text": "This paper leveraged deep ensembles for NFR reduction and showed that they can be distilled into single models by means of a Logit Difference Inhibition (LDI) objective. Despite the methodologically incremental contribution, the proposed ensemble method is capable of reducing NFR while maintaining accuracy and can be distilled to a readily deployable single model which exhibits much lower inference time compared to the deep ensemble. The latter in particular makes this work applicable to large-scale real-world systems where online inference and low latency are paramount. Overall, although there are certain weak points that need to be addressed (listed in the “Weaknesses” part of this review), the strengths of this work certainly outweigh its weaknesses. Therefore, I consider this paper to be a fairly good fit for ICLR and I recommend that it is accepted.", "strengths": "Strengths:\n\n* ELODI trains a deep homogeneous ensemble that can achieve Positive-Congruent Training (PC-Training) performance and be distilled to a single model (which does not necessarily need to be a specific legacy model). This allows the proposed ELODI method to reduce NFR while retaining the accuracy gain introduced by the new model.\n\n* The authors introduce Logit Difference Inhibition (LDI), a knowledge distillation objective that penalizes significant differences between the logits of the reference homogeneous ensemble and the single student model. Such an objective allows for effective distillation of the reference ensemble to a single student model.\n\n* ELODI exhibits efficient inference as it does not require the homogeneous reference ensemble to be evaluated at inference time but rather only the single student model.\n\n* The choice of homogeneous ensembles as reference ensembles is empirically well supported by the findings summarized in Figure 4. Moreover, the hypothesis of reducing logit displacement magnitude (on which the introduced LDI objective relies) is validated empirically in Section 3.2.\n\n* The authors have conducted extensive experiments on multiple image classification benchmarks to assess the effectiveness and efficiency of ELODI in comparison with several PC-training and ensemble methods. The experimental results demonstrate that ELODI performs positive congruent training by reducing NFR with large logit displacement and reducing the variance of the ensemble’s logits. It is also empirically demonstrated that this behavior can be transferred to a single model which considerably reduces the inference time. For instance, this allows ELODI to achieve comparable ER-NFR performance to that of the ensemble paragon [Yan et al., 2021] at the inference cost of a single model.\n\n* The authors have conducted a plethora of additional experimental studies to assess the effect of various aspects including:\n\n    (1) model update settings (including fine-tuning on other datasets, transitivity of NFR reduction induced in chain of model updates, and ELODI’s applicability to updates across dissimilar architectures);\n\n    (2) design choices (including the effects of using homogeneous instead of heterogeneous ensembles for guidance, changes in the guiding ensemble’s architecture, and different choices of distillation loss functions);\n\n    (3) exploratory parameter analysis assessing the effects of the loss weight $\\alpha$, the size of the reference ensemble as well as the manner in which the logits are inferred (offline compared to online).\n\n-------------------------------------------------------------------------------------------------\n\nWeaknesses:\n\n* At the end of the introductory paragraph of Section 3, the authors claim that “the averaging operation in deep ensembles is also often performed in the logit space (Lakshminarayanan et al., 2017), which makes it interesting for our study.” Nevertheless, I am not confident how often this is the case, namely because in classification settings voting schemes are typically used upon applying the activation function of choice (e.g., softmax) to the logits estimated by each member model rather than averaging the logits. I would encourage the authors to further clarify their motivation behind the choice of the averaging aggregation scheme in the logit space, or consider revising the aforementioned statement in the paper.\n\n* The assumption that the logits outputted by the base members of a homogeneous ensemble follow a normal distribution is a fair assumption to make. This is also empirically supported in Section 3.2. However, [1] suggests that, for this to hold, the base networks need to be trained in a sufficient number of epochs. Therefore, including a brief discussion in Section 3.2 on the relationship between the number of training epochs and the members’ logit distributions would be quite insightful. In addition, I believe that [1] is relevant to this paperr as it discusses logit distributions of deep neural networks and thus I would suggest that the authors consider including it as a part of their related work.\n\n    [1] Seguin, L., Ndirango, A., Mishra, N., Chung, S., & Lee, T. (2021). Understanding the Logit Distributions of Adversarially-Trained Deep Neural Networks. arXiv preprint arXiv:2108.12001.\n\n* In Section 3.2, while discussing Figure 3a, the authors state the following: “If we move to the ensemble case, the logit displacement follows another normal distribution with scaled co-variance matrix”. However, according to Figure 3a, apart from the scaled co-variance matrix, the mean is shifted. This should also be included in the same discussion.\n\n-------------------------------------------------------------------------------------------------\n\nMinor weaknesses:\nThere are also certain grammatical and typographical errors and remarks that require attention. Some of them are summarized as follows:\n- In the “Contributions” paragraph on page 2, “the” should be removed from the phrase “obtain the their performance for PC-training”. Later in the same paragraph, the hyphen should be removed from “with re- spect to”.\n- Section 3.1, at the end of page 3: In the phrase “Each $\\phi(x)$ or $\\phi(x)$”, I believe that one of the $\\phi(x)$’s should be corrected.\n- In the second paragraph of Section 4.1, replace “an model pair” with “a model pair”.\n- Section 4.2: In the sentence right after Eq. (6), it is mentioned that “difference below $\\xi$ is tolerated”; however, according to Eq. (6) it seems that a difference below $\\xi$ is truncated, i.e. not tolerated. Later in that same paragraph, “when number of classes are extremely largencated” should be corrected to “when the number of classes is extremely large”.\n", "weaknesses": "", "questions": "", "all_content": {"confidence": "4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.", "summary_of_the_paper": "In this work, the authors repurpose the well-known knowledge distillation paradigm, which is typically used for improving model accuracy, for the task of transferring the capability of reducing negative flip rate (NFR) from an ensemble to a single model. To this end, the authors devise a method called Ensemble Logit Difference Inhibition (ELODI). ELODI leverages the advantages of deep ensembles for NFR reduction at the running cost of a single model by (1) training deep networks using a Logit Difference Inhibition (LDI) loss with respect to a reference ensemble and then (2) using the resulting single model to perform inference.", "strength_and_weaknesses": "Strengths:\n\n* ELODI trains a deep homogeneous ensemble that can achieve Positive-Congruent Training (PC-Training) performance and be distilled to a single model (which does not necessarily need to be a specific legacy model). This allows the proposed ELODI method to reduce NFR while retaining the accuracy gain introduced by the new model.\n\n* The authors introduce Logit Difference Inhibition (LDI), a knowledge distillation objective that penalizes significant differences between the logits of the reference homogeneous ensemble and the single student model. Such an objective allows for effective distillation of the reference ensemble to a single student model.\n\n* ELODI exhibits efficient inference as it does not require the homogeneous reference ensemble to be evaluated at inference time but rather only the single student model.\n\n* The choice of homogeneous ensembles as reference ensembles is empirically well supported by the findings summarized in Figure 4. Moreover, the hypothesis of reducing logit displacement magnitude (on which the introduced LDI objective relies) is validated empirically in Section 3.2.\n\n* The authors have conducted extensive experiments on multiple image classification benchmarks to assess the effectiveness and efficiency of ELODI in comparison with several PC-training and ensemble methods. The experimental results demonstrate that ELODI performs positive congruent training by reducing NFR with large logit displacement and reducing the variance of the ensemble’s logits. It is also empirically demonstrated that this behavior can be transferred to a single model which considerably reduces the inference time. For instance, this allows ELODI to achieve comparable ER-NFR performance to that of the ensemble paragon [Yan et al., 2021] at the inference cost of a single model.\n\n* The authors have conducted a plethora of additional experimental studies to assess the effect of various aspects including:\n\n    (1) model update settings (including fine-tuning on other datasets, transitivity of NFR reduction induced in chain of model updates, and ELODI’s applicability to updates across dissimilar architectures);\n\n    (2) design choices (including the effects of using homogeneous instead of heterogeneous ensembles for guidance, changes in the guiding ensemble’s architecture, and different choices of distillation loss functions);\n\n    (3) exploratory parameter analysis assessing the effects of the loss weight $\\alpha$, the size of the reference ensemble as well as the manner in which the logits are inferred (offline compared to online).\n\n-------------------------------------------------------------------------------------------------\n\nWeaknesses:\n\n* At the end of the introductory paragraph of Section 3, the authors claim that “the averaging operation in deep ensembles is also often performed in the logit space (Lakshminarayanan et al., 2017), which makes it interesting for our study.” Nevertheless, I am not confident how often this is the case, namely because in classification settings voting schemes are typically used upon applying the activation function of choice (e.g., softmax) to the logits estimated by each member model rather than averaging the logits. I would encourage the authors to further clarify their motivation behind the choice of the averaging aggregation scheme in the logit space, or consider revising the aforementioned statement in the paper.\n\n* The assumption that the logits outputted by the base members of a homogeneous ensemble follow a normal distribution is a fair assumption to make. This is also empirically supported in Section 3.2. However, [1] suggests that, for this to hold, the base networks need to be trained in a sufficient number of epochs. Therefore, including a brief discussion in Section 3.2 on the relationship between the number of training epochs and the members’ logit distributions would be quite insightful. In addition, I believe that [1] is relevant to this paperr as it discusses logit distributions of deep neural networks and thus I would suggest that the authors consider including it as a part of their related work.\n\n    [1] Seguin, L., Ndirango, A., Mishra, N., Chung, S., & Lee, T. (2021). Understanding the Logit Distributions of Adversarially-Trained Deep Neural Networks. arXiv preprint arXiv:2108.12001.\n\n* In Section 3.2, while discussing Figure 3a, the authors state the following: “If we move to the ensemble case, the logit displacement follows another normal distribution with scaled co-variance matrix”. However, according to Figure 3a, apart from the scaled co-variance matrix, the mean is shifted. This should also be included in the same discussion.\n\n-------------------------------------------------------------------------------------------------\n\nMinor weaknesses:\nThere are also certain grammatical and typographical errors and remarks that require attention. Some of them are summarized as follows:\n- In the “Contributions” paragraph on page 2, “the” should be removed from the phrase “obtain the their performance for PC-training”. Later in the same paragraph, the hyphen should be removed from “with re- spect to”.\n- Section 3.1, at the end of page 3: In the phrase “Each $\\phi(x)$ or $\\phi(x)$”, I believe that one of the $\\phi(x)$’s should be corrected.\n- In the second paragraph of Section 4.1, replace “an model pair” with “a model pair”.\n- Section 4.2: In the sentence right after Eq. (6), it is mentioned that “difference below $\\xi$ is tolerated”; however, according to Eq. (6) it seems that a difference below $\\xi$ is truncated, i.e. not tolerated. Later in that same paragraph, “when number of classes are extremely largencated” should be corrected to “when the number of classes is extremely large”.\n", "clarity,_quality,_novelty_and_reproducibility": "Clarity:\nThe paper is well written, organized, and technically detailed. The used notation is clear and consistent throughout the paper.\n\nQuality:\nThe design and justifications for the proposed Ensemble Logit Difference Inhibition method are technically sound. The observations made regarding the reduction in NFR and inference cost introduced by the proposed method as well as its capability to maintain image classification accuracy gains are empirically well supported. Overall, the work seems to be well developed and mature in terms of quality.\n\nNovelty:\nFrom a methodological perspective, the contribution of this work can be considered incremental since it is built upon the well-established knowledge distillation framework that utilizes the proposed simple yet well-crafted Logit Difference Inhibition (LDI) objective. Nevertheless, the authors provide an extensive analysis on the role of deep ensembles in a knowledge distillation scheme which sheds light on their capability of reducing NFR and obtaining performance for PC-training with a single model, at the running cost of a single model. To the best of my knowledge, this is the first work to conduct such an extensive analysis regarding the aforementioned points.\n\nReproducibility:\nThe experiments were conducted on widely-used image classification benchmark datasets which are publicly available. On the other hand, the code for the proposed ELODI method is not made available by the authors in the present anonymized version, however, the architecture of the method is clearly explained in the paper, thus it is safe to say that one should be able to implement ELODI by following its detailed description in Section 4.", "summary_of_the_review": "This paper leveraged deep ensembles for NFR reduction and showed that they can be distilled into single models by means of a Logit Difference Inhibition (LDI) objective. Despite the methodologically incremental contribution, the proposed ensemble method is capable of reducing NFR while maintaining accuracy and can be distilled to a readily deployable single model which exhibits much lower inference time compared to the deep ensemble. The latter in particular makes this work applicable to large-scale real-world systems where online inference and low latency are paramount. Overall, although there are certain weak points that need to be addressed (listed in the “Weaknesses” part of this review), the strengths of this work certainly outweigh its weaknesses. Therefore, I consider this paper to be a fairly good fit for ICLR and I recommend that it is accepted.", "correctness": "3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.", "technical_novelty_and_significance": "3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.", "empirical_novelty_and_significance": "4: The contributions are significant, and do not exist in prior works.", "flag_for_ethics_review": ["NO."], "details_of_ethics_concerns": "Not applicable.", "recommendation": "6: marginally above the acceptance threshold"}, "tcdate": 1666681919699}], "openreview_url": "https://openreview.net/forum?id=IJwhRE510b", "arxiv_id": "2205.06265", "paper_pdf": "papers/IJwhRE510b.pdf", "paper_pdf_sha256": "962ab09db994ac743851b1dfde9a9550e79dc50cd610caef821d32d8383aafa6", "paper_pdf_bytes": 1817942, "paper_pdf_source": "openreview", "code_url": "https://github.com/amazon-science/regression-constraint-model-upgrade", "code_repository": "amazon-science/regression-constraint-model-upgrade", "code_commit": "fb2268c67706834632ec1427945fcf30a7218227", "code_archive": "repos/IJwhRE510b.zip", "code_archive_sha256": "f96d039c962b8974340148d30689577d69def49e44d86c514613f84caa48fcbb", "code_archive_bytes": 1964441, "code_file_count": 45, "code_extensions": {".py": 45}, "github_disk_usage_kb": 1809, "github_languages": {"Python": 349619}, "github_archived": false, "github_pushed_at": "2023-12-08T20:26:18Z", "code_match": {"method": "exact_normalized_title", "pwc_official": true, "pwc_mentioned_in_paper": true, "pwc_url": "https://paperswithcode.com/paper/elodi-ensemble-logit-difference-inhibition"}, "provenance": {"paper_reviews_decision": "OpenReview API", "paper_code_mapping": "Papers With Code archive", "repository_snapshot": "GitHub API commit-pinned ZIP"}}